xref: /illumos-gate/usr/src/uts/common/vm/seg_vn.c (revision 04e56356520b98d5a93c496b10f02530bb6647e0)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 1986, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright 2018 Joyent, Inc.
24  * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
25  */
26 
27 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
28 /*	  All Rights Reserved  	*/
29 
30 /*
31  * University Copyright- Copyright (c) 1982, 1986, 1988
32  * The Regents of the University of California
33  * All Rights Reserved
34  *
35  * University Acknowledgment- Portions of this document are derived from
36  * software developed by the University of California, Berkeley, and its
37  * contributors.
38  */
39 
40 /*
41  * VM - shared or copy-on-write from a vnode/anonymous memory.
42  */
43 
44 #include <sys/types.h>
45 #include <sys/param.h>
46 #include <sys/t_lock.h>
47 #include <sys/errno.h>
48 #include <sys/systm.h>
49 #include <sys/mman.h>
50 #include <sys/debug.h>
51 #include <sys/cred.h>
52 #include <sys/vmsystm.h>
53 #include <sys/tuneable.h>
54 #include <sys/bitmap.h>
55 #include <sys/swap.h>
56 #include <sys/kmem.h>
57 #include <sys/sysmacros.h>
58 #include <sys/vtrace.h>
59 #include <sys/cmn_err.h>
60 #include <sys/callb.h>
61 #include <sys/vm.h>
62 #include <sys/dumphdr.h>
63 #include <sys/lgrp.h>
64 
65 #include <vm/hat.h>
66 #include <vm/as.h>
67 #include <vm/seg.h>
68 #include <vm/seg_vn.h>
69 #include <vm/pvn.h>
70 #include <vm/anon.h>
71 #include <vm/page.h>
72 #include <vm/vpage.h>
73 #include <sys/proc.h>
74 #include <sys/task.h>
75 #include <sys/project.h>
76 #include <sys/zone.h>
77 #include <sys/shm_impl.h>
78 
79 /*
80  * segvn_fault needs a temporary page list array.  To avoid calling kmem all
81  * the time, it creates a small (PVN_GETPAGE_NUM entry) array and uses it if
82  * it can.  In the rare case when this page list is not large enough, it
83  * goes and gets a large enough array from kmem.
84  *
85  * This small page list array covers either 8 pages or 64kB worth of pages -
86  * whichever is smaller.
87  */
88 #define	PVN_MAX_GETPAGE_SZ	0x10000
89 #define	PVN_MAX_GETPAGE_NUM	0x8
90 
91 #if PVN_MAX_GETPAGE_SZ > PVN_MAX_GETPAGE_NUM * PAGESIZE
92 #define	PVN_GETPAGE_SZ	ptob(PVN_MAX_GETPAGE_NUM)
93 #define	PVN_GETPAGE_NUM	PVN_MAX_GETPAGE_NUM
94 #else
95 #define	PVN_GETPAGE_SZ	PVN_MAX_GETPAGE_SZ
96 #define	PVN_GETPAGE_NUM	btop(PVN_MAX_GETPAGE_SZ)
97 #endif
98 
99 /*
100  * Private seg op routines.
101  */
102 static int	segvn_dup(struct seg *seg, struct seg *newseg);
103 static int	segvn_unmap(struct seg *seg, caddr_t addr, size_t len);
104 static void	segvn_free(struct seg *seg);
105 static faultcode_t segvn_fault(struct hat *hat, struct seg *seg,
106 		    caddr_t addr, size_t len, enum fault_type type,
107 		    enum seg_rw rw);
108 static faultcode_t segvn_faulta(struct seg *seg, caddr_t addr);
109 static int	segvn_setprot(struct seg *seg, caddr_t addr,
110 		    size_t len, uint_t prot);
111 static int	segvn_checkprot(struct seg *seg, caddr_t addr,
112 		    size_t len, uint_t prot);
113 static int	segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta);
114 static size_t	segvn_swapout(struct seg *seg);
115 static int	segvn_sync(struct seg *seg, caddr_t addr, size_t len,
116 		    int attr, uint_t flags);
117 static size_t	segvn_incore(struct seg *seg, caddr_t addr, size_t len,
118 		    char *vec);
119 static int	segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
120 		    int attr, int op, ulong_t *lockmap, size_t pos);
121 static int	segvn_getprot(struct seg *seg, caddr_t addr, size_t len,
122 		    uint_t *protv);
123 static u_offset_t	segvn_getoffset(struct seg *seg, caddr_t addr);
124 static int	segvn_gettype(struct seg *seg, caddr_t addr);
125 static int	segvn_getvp(struct seg *seg, caddr_t addr,
126 		    struct vnode **vpp);
127 static int	segvn_advise(struct seg *seg, caddr_t addr, size_t len,
128 		    uint_t behav);
129 static void	segvn_dump(struct seg *seg);
130 static int	segvn_pagelock(struct seg *seg, caddr_t addr, size_t len,
131 		    struct page ***ppp, enum lock_type type, enum seg_rw rw);
132 static int	segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len,
133 		    uint_t szc);
134 static int	segvn_getmemid(struct seg *seg, caddr_t addr,
135 		    memid_t *memidp);
136 static lgrp_mem_policy_info_t	*segvn_getpolicy(struct seg *, caddr_t);
137 static int	segvn_capable(struct seg *seg, segcapability_t capable);
138 static int	segvn_inherit(struct seg *, caddr_t, size_t, uint_t);
139 
140 struct	seg_ops segvn_ops = {
141 	segvn_dup,
142 	segvn_unmap,
143 	segvn_free,
144 	segvn_fault,
145 	segvn_faulta,
146 	segvn_setprot,
147 	segvn_checkprot,
148 	segvn_kluster,
149 	segvn_swapout,
150 	segvn_sync,
151 	segvn_incore,
152 	segvn_lockop,
153 	segvn_getprot,
154 	segvn_getoffset,
155 	segvn_gettype,
156 	segvn_getvp,
157 	segvn_advise,
158 	segvn_dump,
159 	segvn_pagelock,
160 	segvn_setpagesize,
161 	segvn_getmemid,
162 	segvn_getpolicy,
163 	segvn_capable,
164 	segvn_inherit
165 };
166 
167 /*
168  * Common zfod structures, provided as a shorthand for others to use.
169  */
170 static segvn_crargs_t zfod_segvn_crargs =
171 	SEGVN_ZFOD_ARGS(PROT_ZFOD, PROT_ALL);
172 static segvn_crargs_t kzfod_segvn_crargs =
173 	SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_USER,
174 	PROT_ALL & ~PROT_USER);
175 static segvn_crargs_t stack_noexec_crargs =
176 	SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_EXEC, PROT_ALL);
177 
178 caddr_t	zfod_argsp = (caddr_t)&zfod_segvn_crargs;	/* user zfod argsp */
179 caddr_t	kzfod_argsp = (caddr_t)&kzfod_segvn_crargs;	/* kernel zfod argsp */
180 caddr_t	stack_exec_argsp = (caddr_t)&zfod_segvn_crargs;	/* executable stack */
181 caddr_t	stack_noexec_argsp = (caddr_t)&stack_noexec_crargs; /* noexec stack */
182 
183 #define	vpgtob(n)	((n) * sizeof (struct vpage))	/* For brevity */
184 
185 size_t	segvn_comb_thrshld = UINT_MAX;	/* patchable -- see 1196681 */
186 
187 size_t	segvn_pglock_comb_thrshld = (1UL << 16);	/* 64K */
188 size_t	segvn_pglock_comb_balign = (1UL << 16);		/* 64K */
189 uint_t	segvn_pglock_comb_bshift;
190 size_t	segvn_pglock_comb_palign;
191 
192 static int	segvn_concat(struct seg *, struct seg *, int);
193 static int	segvn_extend_prev(struct seg *, struct seg *,
194 		    struct segvn_crargs *, size_t);
195 static int	segvn_extend_next(struct seg *, struct seg *,
196 		    struct segvn_crargs *, size_t);
197 static void	segvn_softunlock(struct seg *, caddr_t, size_t, enum seg_rw);
198 static void	segvn_pagelist_rele(page_t **);
199 static void	segvn_setvnode_mpss(vnode_t *);
200 static void	segvn_relocate_pages(page_t **, page_t *);
201 static int	segvn_full_szcpages(page_t **, uint_t, int *, uint_t *);
202 static int	segvn_fill_vp_pages(struct segvn_data *, vnode_t *, u_offset_t,
203     uint_t, page_t **, page_t **, uint_t *, int *);
204 static faultcode_t segvn_fault_vnodepages(struct hat *, struct seg *, caddr_t,
205     caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
206 static faultcode_t segvn_fault_anonpages(struct hat *, struct seg *, caddr_t,
207     caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
208 static faultcode_t segvn_faultpage(struct hat *, struct seg *, caddr_t,
209     u_offset_t, struct vpage *, page_t **, uint_t,
210     enum fault_type, enum seg_rw, int);
211 static void	segvn_vpage(struct seg *);
212 static size_t	segvn_count_swap_by_vpages(struct seg *);
213 
214 static void segvn_purge(struct seg *seg);
215 static int segvn_reclaim(void *, caddr_t, size_t, struct page **,
216     enum seg_rw, int);
217 static int shamp_reclaim(void *, caddr_t, size_t, struct page **,
218     enum seg_rw, int);
219 
220 static int sameprot(struct seg *, caddr_t, size_t);
221 
222 static int segvn_demote_range(struct seg *, caddr_t, size_t, int, uint_t);
223 static int segvn_clrszc(struct seg *);
224 static struct seg *segvn_split_seg(struct seg *, caddr_t);
225 static int segvn_claim_pages(struct seg *, struct vpage *, u_offset_t,
226     ulong_t, uint_t);
227 
228 static void segvn_hat_rgn_unload_callback(caddr_t, caddr_t, caddr_t,
229     size_t, void *, u_offset_t);
230 
231 static struct kmem_cache *segvn_cache;
232 static struct kmem_cache **segvn_szc_cache;
233 
234 #ifdef VM_STATS
235 static struct segvnvmstats_str {
236 	ulong_t	fill_vp_pages[31];
237 	ulong_t fltvnpages[49];
238 	ulong_t	fullszcpages[10];
239 	ulong_t	relocatepages[3];
240 	ulong_t	fltanpages[17];
241 	ulong_t pagelock[2];
242 	ulong_t	demoterange[3];
243 } segvnvmstats;
244 #endif /* VM_STATS */
245 
246 #define	SDR_RANGE	1		/* demote entire range */
247 #define	SDR_END		2		/* demote non aligned ends only */
248 
249 #define	CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr) {	    \
250 		if ((len) != 0) { 		      	      		      \
251 			lpgaddr = (caddr_t)P2ALIGN((uintptr_t)(addr), pgsz);  \
252 			ASSERT(lpgaddr >= (seg)->s_base);	      	      \
253 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)((addr) +    \
254 			    (len)), pgsz);				      \
255 			ASSERT(lpgeaddr > lpgaddr);		      	      \
256 			ASSERT(lpgeaddr <= (seg)->s_base + (seg)->s_size);    \
257 		} else {					      	      \
258 			lpgeaddr = lpgaddr = (addr);	      		      \
259 		}							      \
260 	}
261 
262 /*ARGSUSED*/
263 static int
264 segvn_cache_constructor(void *buf, void *cdrarg, int kmflags)
265 {
266 	struct segvn_data *svd = buf;
267 
268 	rw_init(&svd->lock, NULL, RW_DEFAULT, NULL);
269 	mutex_init(&svd->segfree_syncmtx, NULL, MUTEX_DEFAULT, NULL);
270 	svd->svn_trnext = svd->svn_trprev = NULL;
271 	return (0);
272 }
273 
274 /*ARGSUSED1*/
275 static void
276 segvn_cache_destructor(void *buf, void *cdrarg)
277 {
278 	struct segvn_data *svd = buf;
279 
280 	rw_destroy(&svd->lock);
281 	mutex_destroy(&svd->segfree_syncmtx);
282 }
283 
284 /*ARGSUSED*/
285 static int
286 svntr_cache_constructor(void *buf, void *cdrarg, int kmflags)
287 {
288 	bzero(buf, sizeof (svntr_t));
289 	return (0);
290 }
291 
292 /*
293  * Patching this variable to non-zero allows the system to run with
294  * stacks marked as "not executable".  It's a bit of a kludge, but is
295  * provided as a tweakable for platforms that export those ABIs
296  * (e.g. sparc V8) that have executable stacks enabled by default.
297  * There are also some restrictions for platforms that don't actually
298  * implement 'noexec' protections.
299  *
300  * Once enabled, the system is (therefore) unable to provide a fully
301  * ABI-compliant execution environment, though practically speaking,
302  * most everything works.  The exceptions are generally some interpreters
303  * and debuggers that create executable code on the stack and jump
304  * into it (without explicitly mprotecting the address range to include
305  * PROT_EXEC).
306  *
307  * One important class of applications that are disabled are those
308  * that have been transformed into malicious agents using one of the
309  * numerous "buffer overflow" attacks.  See 4007890.
310  */
311 int noexec_user_stack = 0;
312 int noexec_user_stack_log = 1;
313 
314 int segvn_lpg_disable = 0;
315 uint_t segvn_maxpgszc = 0;
316 
317 ulong_t segvn_vmpss_clrszc_cnt;
318 ulong_t segvn_vmpss_clrszc_err;
319 ulong_t segvn_fltvnpages_clrszc_cnt;
320 ulong_t segvn_fltvnpages_clrszc_err;
321 ulong_t segvn_setpgsz_align_err;
322 ulong_t segvn_setpgsz_anon_align_err;
323 ulong_t segvn_setpgsz_getattr_err;
324 ulong_t segvn_setpgsz_eof_err;
325 ulong_t segvn_faultvnmpss_align_err1;
326 ulong_t segvn_faultvnmpss_align_err2;
327 ulong_t segvn_faultvnmpss_align_err3;
328 ulong_t segvn_faultvnmpss_align_err4;
329 ulong_t segvn_faultvnmpss_align_err5;
330 ulong_t	segvn_vmpss_pageio_deadlk_err;
331 
332 int segvn_use_regions = 1;
333 
334 /*
335  * Segvn supports text replication optimization for NUMA platforms. Text
336  * replica's are represented by anon maps (amp). There's one amp per text file
337  * region per lgroup. A process chooses the amp for each of its text mappings
338  * based on the lgroup assignment of its main thread (t_tid = 1). All
339  * processes that want a replica on a particular lgroup for the same text file
340  * mapping share the same amp. amp's are looked up in svntr_hashtab hash table
341  * with vp,off,size,szc used as a key. Text replication segments are read only
342  * MAP_PRIVATE|MAP_TEXT segments that map vnode. Replication is achieved by
343  * forcing COW faults from vnode to amp and mapping amp pages instead of vnode
344  * pages. Replication amp is assigned to a segment when it gets its first
345  * pagefault. To handle main thread lgroup rehoming segvn_trasync_thread
346  * rechecks periodically if the process still maps an amp local to the main
347  * thread. If not async thread forces process to remap to an amp in the new
348  * home lgroup of the main thread. Current text replication implementation
349  * only provides the benefit to workloads that do most of their work in the
350  * main thread of a process or all the threads of a process run in the same
351  * lgroup. To extend text replication benefit to different types of
352  * multithreaded workloads further work would be needed in the hat layer to
353  * allow the same virtual address in the same hat to simultaneously map
354  * different physical addresses (i.e. page table replication would be needed
355  * for x86).
356  *
357  * amp pages are used instead of vnode pages as long as segment has a very
358  * simple life cycle.  It's created via segvn_create(), handles S_EXEC
359  * (S_READ) pagefaults and is fully unmapped.  If anything more complicated
360  * happens such as protection is changed, real COW fault happens, pagesize is
361  * changed, MC_LOCK is requested or segment is partially unmapped we turn off
362  * text replication by converting the segment back to vnode only segment
363  * (unmap segment's address range and set svd->amp to NULL).
364  *
365  * The original file can be changed after amp is inserted into
366  * svntr_hashtab. Processes that are launched after the file is already
367  * changed can't use the replica's created prior to the file change. To
368  * implement this functionality hash entries are timestamped. Replica's can
369  * only be used if current file modification time is the same as the timestamp
370  * saved when hash entry was created. However just timestamps alone are not
371  * sufficient to detect file modification via mmap(MAP_SHARED) mappings. We
372  * deal with file changes via MAP_SHARED mappings differently. When writable
373  * MAP_SHARED mappings are created to vnodes marked as executable we mark all
374  * existing replica's for this vnode as not usable for future text
375  * mappings. And we don't create new replica's for files that currently have
376  * potentially writable MAP_SHARED mappings (i.e. vn_is_mapped(V_WRITE) is
377  * true).
378  */
379 
380 #define	SEGVN_TEXTREPL_MAXBYTES_FACTOR	(20)
381 size_t	segvn_textrepl_max_bytes_factor = SEGVN_TEXTREPL_MAXBYTES_FACTOR;
382 
383 static ulong_t			svntr_hashtab_sz = 512;
384 static svntr_bucket_t		*svntr_hashtab = NULL;
385 static struct kmem_cache	*svntr_cache;
386 static svntr_stats_t		*segvn_textrepl_stats;
387 static ksema_t 			segvn_trasync_sem;
388 
389 int				segvn_disable_textrepl = 1;
390 size_t				textrepl_size_thresh = (size_t)-1;
391 size_t				segvn_textrepl_bytes = 0;
392 size_t				segvn_textrepl_max_bytes = 0;
393 clock_t				segvn_update_textrepl_interval = 0;
394 int				segvn_update_tr_time = 10;
395 int				segvn_disable_textrepl_update = 0;
396 
397 static void segvn_textrepl(struct seg *);
398 static void segvn_textunrepl(struct seg *, int);
399 static void segvn_inval_trcache(vnode_t *);
400 static void segvn_trasync_thread(void);
401 static void segvn_trupdate_wakeup(void *);
402 static void segvn_trupdate(void);
403 static void segvn_trupdate_seg(struct seg *, segvn_data_t *, svntr_t *,
404     ulong_t);
405 
406 /*
407  * Initialize segvn data structures
408  */
409 void
410 segvn_init(void)
411 {
412 	uint_t maxszc;
413 	uint_t szc;
414 	size_t pgsz;
415 
416 	segvn_cache = kmem_cache_create("segvn_cache",
417 	    sizeof (struct segvn_data), 0,
418 	    segvn_cache_constructor, segvn_cache_destructor, NULL,
419 	    NULL, NULL, 0);
420 
421 	if (segvn_lpg_disable == 0) {
422 		szc = maxszc = page_num_pagesizes() - 1;
423 		if (szc == 0) {
424 			segvn_lpg_disable = 1;
425 		}
426 		if (page_get_pagesize(0) != PAGESIZE) {
427 			panic("segvn_init: bad szc 0");
428 			/*NOTREACHED*/
429 		}
430 		while (szc != 0) {
431 			pgsz = page_get_pagesize(szc);
432 			if (pgsz <= PAGESIZE || !IS_P2ALIGNED(pgsz, pgsz)) {
433 				panic("segvn_init: bad szc %d", szc);
434 				/*NOTREACHED*/
435 			}
436 			szc--;
437 		}
438 		if (segvn_maxpgszc == 0 || segvn_maxpgszc > maxszc)
439 			segvn_maxpgszc = maxszc;
440 	}
441 
442 	if (segvn_maxpgszc) {
443 		segvn_szc_cache = (struct kmem_cache **)kmem_alloc(
444 		    (segvn_maxpgszc + 1) * sizeof (struct kmem_cache *),
445 		    KM_SLEEP);
446 	}
447 
448 	for (szc = 1; szc <= segvn_maxpgszc; szc++) {
449 		char	str[32];
450 
451 		(void) sprintf(str, "segvn_szc_cache%d", szc);
452 		segvn_szc_cache[szc] = kmem_cache_create(str,
453 		    page_get_pagecnt(szc) * sizeof (page_t *), 0,
454 		    NULL, NULL, NULL, NULL, NULL, KMC_NODEBUG);
455 	}
456 
457 
458 	if (segvn_use_regions && !hat_supported(HAT_SHARED_REGIONS, NULL))
459 		segvn_use_regions = 0;
460 
461 	/*
462 	 * For now shared regions and text replication segvn support
463 	 * are mutually exclusive. This is acceptable because
464 	 * currently significant benefit from text replication was
465 	 * only observed on AMD64 NUMA platforms (due to relatively
466 	 * small L2$ size) and currently we don't support shared
467 	 * regions on x86.
468 	 */
469 	if (segvn_use_regions && !segvn_disable_textrepl) {
470 		segvn_disable_textrepl = 1;
471 	}
472 
473 #if defined(_LP64)
474 	if (lgrp_optimizations() && textrepl_size_thresh != (size_t)-1 &&
475 	    !segvn_disable_textrepl) {
476 		ulong_t i;
477 		size_t hsz = svntr_hashtab_sz * sizeof (svntr_bucket_t);
478 
479 		svntr_cache = kmem_cache_create("svntr_cache",
480 		    sizeof (svntr_t), 0, svntr_cache_constructor, NULL,
481 		    NULL, NULL, NULL, 0);
482 		svntr_hashtab = kmem_zalloc(hsz, KM_SLEEP);
483 		for (i = 0; i < svntr_hashtab_sz; i++) {
484 			mutex_init(&svntr_hashtab[i].tr_lock,  NULL,
485 			    MUTEX_DEFAULT, NULL);
486 		}
487 		segvn_textrepl_max_bytes = ptob(physmem) /
488 		    segvn_textrepl_max_bytes_factor;
489 		segvn_textrepl_stats = kmem_zalloc(NCPU *
490 		    sizeof (svntr_stats_t), KM_SLEEP);
491 		sema_init(&segvn_trasync_sem, 0, NULL, SEMA_DEFAULT, NULL);
492 		(void) thread_create(NULL, 0, segvn_trasync_thread,
493 		    NULL, 0, &p0, TS_RUN, minclsyspri);
494 	}
495 #endif
496 
497 	if (!ISP2(segvn_pglock_comb_balign) ||
498 	    segvn_pglock_comb_balign < PAGESIZE) {
499 		segvn_pglock_comb_balign = 1UL << 16; /* 64K */
500 	}
501 	segvn_pglock_comb_bshift = highbit(segvn_pglock_comb_balign) - 1;
502 	segvn_pglock_comb_palign = btop(segvn_pglock_comb_balign);
503 }
504 
505 #define	SEGVN_PAGEIO	((void *)0x1)
506 #define	SEGVN_NOPAGEIO	((void *)0x2)
507 
508 static void
509 segvn_setvnode_mpss(vnode_t *vp)
510 {
511 	int err;
512 
513 	ASSERT(vp->v_mpssdata == NULL ||
514 	    vp->v_mpssdata == SEGVN_PAGEIO ||
515 	    vp->v_mpssdata == SEGVN_NOPAGEIO);
516 
517 	if (vp->v_mpssdata == NULL) {
518 		if (vn_vmpss_usepageio(vp)) {
519 			err = VOP_PAGEIO(vp, (page_t *)NULL,
520 			    (u_offset_t)0, 0, 0, CRED(), NULL);
521 		} else {
522 			err = ENOSYS;
523 		}
524 		/*
525 		 * set v_mpssdata just once per vnode life
526 		 * so that it never changes.
527 		 */
528 		mutex_enter(&vp->v_lock);
529 		if (vp->v_mpssdata == NULL) {
530 			if (err == EINVAL) {
531 				vp->v_mpssdata = SEGVN_PAGEIO;
532 			} else {
533 				vp->v_mpssdata = SEGVN_NOPAGEIO;
534 			}
535 		}
536 		mutex_exit(&vp->v_lock);
537 	}
538 }
539 
540 int
541 segvn_create(struct seg **segpp, void *argsp)
542 {
543 	struct seg *seg = *segpp;
544 	extern lgrp_mem_policy_t lgrp_mem_default_policy;
545 	struct segvn_crargs *a = (struct segvn_crargs *)argsp;
546 	struct segvn_data *svd;
547 	size_t swresv = 0;
548 	struct cred *cred;
549 	struct anon_map *amp;
550 	int error = 0;
551 	size_t pgsz;
552 	lgrp_mem_policy_t mpolicy = lgrp_mem_default_policy;
553 	int use_rgn = 0;
554 	int trok = 0;
555 
556 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
557 
558 	if (a->type != MAP_PRIVATE && a->type != MAP_SHARED) {
559 		panic("segvn_create type");
560 		/*NOTREACHED*/
561 	}
562 
563 	/*
564 	 * Check arguments.  If a shared anon structure is given then
565 	 * it is illegal to also specify a vp.
566 	 */
567 	if (a->amp != NULL && a->vp != NULL) {
568 		panic("segvn_create anon_map");
569 		/*NOTREACHED*/
570 	}
571 
572 	if (a->type == MAP_PRIVATE && (a->flags & MAP_TEXT) &&
573 	    a->vp != NULL && a->prot == (PROT_USER | PROT_READ | PROT_EXEC) &&
574 	    segvn_use_regions) {
575 		use_rgn = 1;
576 	}
577 
578 	/* MAP_NORESERVE on a MAP_SHARED segment is meaningless. */
579 	if (a->type == MAP_SHARED)
580 		a->flags &= ~MAP_NORESERVE;
581 
582 	if (a->szc != 0) {
583 		if (segvn_lpg_disable != 0 || (a->szc == AS_MAP_NO_LPOOB) ||
584 		    (a->amp != NULL && a->type == MAP_PRIVATE) ||
585 		    (a->flags & MAP_NORESERVE) || seg->s_as == &kas) {
586 			a->szc = 0;
587 		} else {
588 			if (a->szc > segvn_maxpgszc)
589 				a->szc = segvn_maxpgszc;
590 			pgsz = page_get_pagesize(a->szc);
591 			if (!IS_P2ALIGNED(seg->s_base, pgsz) ||
592 			    !IS_P2ALIGNED(seg->s_size, pgsz)) {
593 				a->szc = 0;
594 			} else if (a->vp != NULL) {
595 				if (IS_SWAPFSVP(a->vp) || VN_ISKAS(a->vp)) {
596 					/*
597 					 * paranoid check.
598 					 * hat_page_demote() is not supported
599 					 * on swapfs pages.
600 					 */
601 					a->szc = 0;
602 				} else if (map_addr_vacalign_check(seg->s_base,
603 				    a->offset & PAGEMASK)) {
604 					a->szc = 0;
605 				}
606 			} else if (a->amp != NULL) {
607 				pgcnt_t anum = btopr(a->offset);
608 				pgcnt_t pgcnt = page_get_pagecnt(a->szc);
609 				if (!IS_P2ALIGNED(anum, pgcnt)) {
610 					a->szc = 0;
611 				}
612 			}
613 		}
614 	}
615 
616 	/*
617 	 * If segment may need private pages, reserve them now.
618 	 */
619 	if (!(a->flags & MAP_NORESERVE) && ((a->vp == NULL && a->amp == NULL) ||
620 	    (a->type == MAP_PRIVATE && (a->prot & PROT_WRITE)))) {
621 		if (anon_resv_zone(seg->s_size,
622 		    seg->s_as->a_proc->p_zone) == 0)
623 			return (EAGAIN);
624 		swresv = seg->s_size;
625 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
626 		    seg, swresv, 1);
627 	}
628 
629 	/*
630 	 * Reserve any mapping structures that may be required.
631 	 *
632 	 * Don't do it for segments that may use regions. It's currently a
633 	 * noop in the hat implementations anyway.
634 	 */
635 	if (!use_rgn) {
636 		hat_map(seg->s_as->a_hat, seg->s_base, seg->s_size, HAT_MAP);
637 	}
638 
639 	if (a->cred) {
640 		cred = a->cred;
641 		crhold(cred);
642 	} else {
643 		crhold(cred = CRED());
644 	}
645 
646 	/* Inform the vnode of the new mapping */
647 	if (a->vp != NULL) {
648 		error = VOP_ADDMAP(a->vp, a->offset & PAGEMASK,
649 		    seg->s_as, seg->s_base, seg->s_size, a->prot,
650 		    a->maxprot, a->type, cred, NULL);
651 		if (error) {
652 			if (swresv != 0) {
653 				anon_unresv_zone(swresv,
654 				    seg->s_as->a_proc->p_zone);
655 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
656 				    "anon proc:%p %lu %u", seg, swresv, 0);
657 			}
658 			crfree(cred);
659 			if (!use_rgn) {
660 				hat_unload(seg->s_as->a_hat, seg->s_base,
661 				    seg->s_size, HAT_UNLOAD_UNMAP);
662 			}
663 			return (error);
664 		}
665 		/*
666 		 * svntr_hashtab will be NULL if we support shared regions.
667 		 */
668 		trok = ((a->flags & MAP_TEXT) &&
669 		    (seg->s_size > textrepl_size_thresh ||
670 		    (a->flags & _MAP_TEXTREPL)) &&
671 		    lgrp_optimizations() && svntr_hashtab != NULL &&
672 		    a->type == MAP_PRIVATE && swresv == 0 &&
673 		    !(a->flags & MAP_NORESERVE) &&
674 		    seg->s_as != &kas && a->vp->v_type == VREG);
675 
676 		ASSERT(!trok || !use_rgn);
677 	}
678 
679 	/*
680 	 * MAP_NORESERVE mappings don't count towards the VSZ of a process
681 	 * until we fault the pages in.
682 	 */
683 	if ((a->vp == NULL || a->vp->v_type != VREG) &&
684 	    a->flags & MAP_NORESERVE) {
685 		seg->s_as->a_resvsize -= seg->s_size;
686 	}
687 
688 	/*
689 	 * If more than one segment in the address space, and they're adjacent
690 	 * virtually, try to concatenate them.  Don't concatenate if an
691 	 * explicit anon_map structure was supplied (e.g., SystemV shared
692 	 * memory) or if we'll use text replication for this segment.
693 	 */
694 	if (a->amp == NULL && !use_rgn && !trok) {
695 		struct seg *pseg, *nseg;
696 		struct segvn_data *psvd, *nsvd;
697 		lgrp_mem_policy_t ppolicy, npolicy;
698 		uint_t	lgrp_mem_policy_flags = 0;
699 
700 		/*
701 		 * Memory policy flags (lgrp_mem_policy_flags) is valid when
702 		 * extending stack/heap segments.
703 		 */
704 		if ((a->vp == NULL) && (a->type == MAP_PRIVATE) &&
705 		    !(a->flags & MAP_NORESERVE) && (seg->s_as != &kas)) {
706 			lgrp_mem_policy_flags = a->lgrp_mem_policy_flags;
707 		} else {
708 			/*
709 			 * Get policy when not extending it from another segment
710 			 */
711 			mpolicy = lgrp_mem_policy_default(seg->s_size, a->type);
712 		}
713 
714 		/*
715 		 * First, try to concatenate the previous and new segments
716 		 */
717 		pseg = AS_SEGPREV(seg->s_as, seg);
718 		if (pseg != NULL &&
719 		    pseg->s_base + pseg->s_size == seg->s_base &&
720 		    pseg->s_ops == &segvn_ops) {
721 			/*
722 			 * Get memory allocation policy from previous segment.
723 			 * When extension is specified (e.g. for heap) apply
724 			 * this policy to the new segment regardless of the
725 			 * outcome of segment concatenation.  Extension occurs
726 			 * for non-default policy otherwise default policy is
727 			 * used and is based on extended segment size.
728 			 */
729 			psvd = (struct segvn_data *)pseg->s_data;
730 			ppolicy = psvd->policy_info.mem_policy;
731 			if (lgrp_mem_policy_flags ==
732 			    LGRP_MP_FLAG_EXTEND_UP) {
733 				if (ppolicy != lgrp_mem_default_policy) {
734 					mpolicy = ppolicy;
735 				} else {
736 					mpolicy = lgrp_mem_policy_default(
737 					    pseg->s_size + seg->s_size,
738 					    a->type);
739 				}
740 			}
741 
742 			if (mpolicy == ppolicy &&
743 			    (pseg->s_size + seg->s_size <=
744 			    segvn_comb_thrshld || psvd->amp == NULL) &&
745 			    segvn_extend_prev(pseg, seg, a, swresv) == 0) {
746 				/*
747 				 * success! now try to concatenate
748 				 * with following seg
749 				 */
750 				crfree(cred);
751 				nseg = AS_SEGNEXT(pseg->s_as, pseg);
752 				if (nseg != NULL &&
753 				    nseg != pseg &&
754 				    nseg->s_ops == &segvn_ops &&
755 				    pseg->s_base + pseg->s_size ==
756 				    nseg->s_base)
757 					(void) segvn_concat(pseg, nseg, 0);
758 				ASSERT(pseg->s_szc == 0 ||
759 				    (a->szc == pseg->s_szc &&
760 				    IS_P2ALIGNED(pseg->s_base, pgsz) &&
761 				    IS_P2ALIGNED(pseg->s_size, pgsz)));
762 				/*
763 				 * Communicate out the newly concatenated
764 				 * segment as part of the result.
765 				 */
766 				*segpp = pseg;
767 				return (0);
768 			}
769 		}
770 
771 		/*
772 		 * Failed, so try to concatenate with following seg
773 		 */
774 		nseg = AS_SEGNEXT(seg->s_as, seg);
775 		if (nseg != NULL &&
776 		    seg->s_base + seg->s_size == nseg->s_base &&
777 		    nseg->s_ops == &segvn_ops) {
778 			/*
779 			 * Get memory allocation policy from next segment.
780 			 * When extension is specified (e.g. for stack) apply
781 			 * this policy to the new segment regardless of the
782 			 * outcome of segment concatenation.  Extension occurs
783 			 * for non-default policy otherwise default policy is
784 			 * used and is based on extended segment size.
785 			 */
786 			nsvd = (struct segvn_data *)nseg->s_data;
787 			npolicy = nsvd->policy_info.mem_policy;
788 			if (lgrp_mem_policy_flags ==
789 			    LGRP_MP_FLAG_EXTEND_DOWN) {
790 				if (npolicy != lgrp_mem_default_policy) {
791 					mpolicy = npolicy;
792 				} else {
793 					mpolicy = lgrp_mem_policy_default(
794 					    nseg->s_size + seg->s_size,
795 					    a->type);
796 				}
797 			}
798 
799 			if (mpolicy == npolicy &&
800 			    segvn_extend_next(seg, nseg, a, swresv) == 0) {
801 				crfree(cred);
802 				ASSERT(nseg->s_szc == 0 ||
803 				    (a->szc == nseg->s_szc &&
804 				    IS_P2ALIGNED(nseg->s_base, pgsz) &&
805 				    IS_P2ALIGNED(nseg->s_size, pgsz)));
806 				/*
807 				 * Communicate out the newly concatenated
808 				 * segment as part of the result.
809 				 */
810 				*segpp = nseg;
811 				return (0);
812 			}
813 		}
814 	}
815 
816 	if (a->vp != NULL) {
817 		VN_HOLD(a->vp);
818 		if (a->type == MAP_SHARED)
819 			lgrp_shm_policy_init(NULL, a->vp);
820 	}
821 	svd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
822 
823 	seg->s_ops = &segvn_ops;
824 	seg->s_data = (void *)svd;
825 	seg->s_szc = a->szc;
826 
827 	svd->seg = seg;
828 	svd->vp = a->vp;
829 	/*
830 	 * Anonymous mappings have no backing file so the offset is meaningless.
831 	 */
832 	svd->offset = a->vp ? (a->offset & PAGEMASK) : 0;
833 	svd->prot = a->prot;
834 	svd->maxprot = a->maxprot;
835 	svd->pageprot = 0;
836 	svd->type = a->type;
837 	svd->vpage = NULL;
838 	svd->cred = cred;
839 	svd->advice = MADV_NORMAL;
840 	svd->pageadvice = 0;
841 	svd->flags = (ushort_t)a->flags;
842 	svd->softlockcnt = 0;
843 	svd->softlockcnt_sbase = 0;
844 	svd->softlockcnt_send = 0;
845 	svd->svn_inz = 0;
846 	svd->rcookie = HAT_INVALID_REGION_COOKIE;
847 	svd->pageswap = 0;
848 
849 	if (a->szc != 0 && a->vp != NULL) {
850 		segvn_setvnode_mpss(a->vp);
851 	}
852 	if (svd->type == MAP_SHARED && svd->vp != NULL &&
853 	    (svd->vp->v_flag & VVMEXEC) && (svd->prot & PROT_WRITE)) {
854 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
855 		segvn_inval_trcache(svd->vp);
856 	}
857 
858 	amp = a->amp;
859 	if ((svd->amp = amp) == NULL) {
860 		svd->anon_index = 0;
861 		if (svd->type == MAP_SHARED) {
862 			svd->swresv = 0;
863 			/*
864 			 * Shared mappings to a vp need no other setup.
865 			 * If we have a shared mapping to an anon_map object
866 			 * which hasn't been allocated yet,  allocate the
867 			 * struct now so that it will be properly shared
868 			 * by remembering the swap reservation there.
869 			 */
870 			if (a->vp == NULL) {
871 				svd->amp = anonmap_alloc(seg->s_size, swresv,
872 				    ANON_SLEEP);
873 				svd->amp->a_szc = seg->s_szc;
874 			}
875 		} else {
876 			/*
877 			 * Private mapping (with or without a vp).
878 			 * Allocate anon_map when needed.
879 			 */
880 			svd->swresv = swresv;
881 		}
882 	} else {
883 		pgcnt_t anon_num;
884 
885 		/*
886 		 * Mapping to an existing anon_map structure without a vp.
887 		 * For now we will insure that the segment size isn't larger
888 		 * than the size - offset gives us.  Later on we may wish to
889 		 * have the anon array dynamically allocated itself so that
890 		 * we don't always have to allocate all the anon pointer slots.
891 		 * This of course involves adding extra code to check that we
892 		 * aren't trying to use an anon pointer slot beyond the end
893 		 * of the currently allocated anon array.
894 		 */
895 		if ((amp->size - a->offset) < seg->s_size) {
896 			panic("segvn_create anon_map size");
897 			/*NOTREACHED*/
898 		}
899 
900 		anon_num = btopr(a->offset);
901 
902 		if (a->type == MAP_SHARED) {
903 			/*
904 			 * SHARED mapping to a given anon_map.
905 			 */
906 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
907 			amp->refcnt++;
908 			if (a->szc > amp->a_szc) {
909 				amp->a_szc = a->szc;
910 			}
911 			ANON_LOCK_EXIT(&amp->a_rwlock);
912 			svd->anon_index = anon_num;
913 			svd->swresv = 0;
914 		} else {
915 			/*
916 			 * PRIVATE mapping to a given anon_map.
917 			 * Make sure that all the needed anon
918 			 * structures are created (so that we will
919 			 * share the underlying pages if nothing
920 			 * is written by this mapping) and then
921 			 * duplicate the anon array as is done
922 			 * when a privately mapped segment is dup'ed.
923 			 */
924 			struct anon *ap;
925 			caddr_t addr;
926 			caddr_t eaddr;
927 			ulong_t	anon_idx;
928 			int hat_flag = HAT_LOAD;
929 
930 			if (svd->flags & MAP_TEXT) {
931 				hat_flag |= HAT_LOAD_TEXT;
932 			}
933 
934 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
935 			svd->amp->a_szc = seg->s_szc;
936 			svd->anon_index = 0;
937 			svd->swresv = swresv;
938 
939 			/*
940 			 * Prevent 2 threads from allocating anon
941 			 * slots simultaneously.
942 			 */
943 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
944 			eaddr = seg->s_base + seg->s_size;
945 
946 			for (anon_idx = anon_num, addr = seg->s_base;
947 			    addr < eaddr; addr += PAGESIZE, anon_idx++) {
948 				page_t *pp;
949 
950 				if ((ap = anon_get_ptr(amp->ahp,
951 				    anon_idx)) != NULL)
952 					continue;
953 
954 				/*
955 				 * Allocate the anon struct now.
956 				 * Might as well load up translation
957 				 * to the page while we're at it...
958 				 */
959 				pp = anon_zero(seg, addr, &ap, cred);
960 				if (ap == NULL || pp == NULL) {
961 					panic("segvn_create anon_zero");
962 					/*NOTREACHED*/
963 				}
964 
965 				/*
966 				 * Re-acquire the anon_map lock and
967 				 * initialize the anon array entry.
968 				 */
969 				ASSERT(anon_get_ptr(amp->ahp,
970 				    anon_idx) == NULL);
971 				(void) anon_set_ptr(amp->ahp, anon_idx, ap,
972 				    ANON_SLEEP);
973 
974 				ASSERT(seg->s_szc == 0);
975 				ASSERT(!IS_VMODSORT(pp->p_vnode));
976 
977 				ASSERT(use_rgn == 0);
978 				hat_memload(seg->s_as->a_hat, addr, pp,
979 				    svd->prot & ~PROT_WRITE, hat_flag);
980 
981 				page_unlock(pp);
982 			}
983 			ASSERT(seg->s_szc == 0);
984 			anon_dup(amp->ahp, anon_num, svd->amp->ahp,
985 			    0, seg->s_size);
986 			ANON_LOCK_EXIT(&amp->a_rwlock);
987 		}
988 	}
989 
990 	/*
991 	 * Set default memory allocation policy for segment
992 	 *
993 	 * Always set policy for private memory at least for initialization
994 	 * even if this is a shared memory segment
995 	 */
996 	(void) lgrp_privm_policy_set(mpolicy, &svd->policy_info, seg->s_size);
997 
998 	if (svd->type == MAP_SHARED)
999 		(void) lgrp_shm_policy_set(mpolicy, svd->amp, svd->anon_index,
1000 		    svd->vp, svd->offset, seg->s_size);
1001 
1002 	if (use_rgn) {
1003 		ASSERT(!trok);
1004 		ASSERT(svd->amp == NULL);
1005 		svd->rcookie = hat_join_region(seg->s_as->a_hat, seg->s_base,
1006 		    seg->s_size, (void *)svd->vp, svd->offset, svd->prot,
1007 		    (uchar_t)seg->s_szc, segvn_hat_rgn_unload_callback,
1008 		    HAT_REGION_TEXT);
1009 	}
1010 
1011 	ASSERT(!trok || !(svd->prot & PROT_WRITE));
1012 	svd->tr_state = trok ? SEGVN_TR_INIT : SEGVN_TR_OFF;
1013 
1014 	return (0);
1015 }
1016 
1017 /*
1018  * Concatenate two existing segments, if possible.
1019  * Return 0 on success, -1 if two segments are not compatible
1020  * or -2 on memory allocation failure.
1021  * If amp_cat == 1 then try and concat segments with anon maps
1022  */
1023 static int
1024 segvn_concat(struct seg *seg1, struct seg *seg2, int amp_cat)
1025 {
1026 	struct segvn_data *svd1 = seg1->s_data;
1027 	struct segvn_data *svd2 = seg2->s_data;
1028 	struct anon_map *amp1 = svd1->amp;
1029 	struct anon_map *amp2 = svd2->amp;
1030 	struct vpage *vpage1 = svd1->vpage;
1031 	struct vpage *vpage2 = svd2->vpage, *nvpage = NULL;
1032 	size_t size, nvpsize;
1033 	pgcnt_t npages1, npages2;
1034 
1035 	ASSERT(seg1->s_as && seg2->s_as && seg1->s_as == seg2->s_as);
1036 	ASSERT(AS_WRITE_HELD(seg1->s_as));
1037 	ASSERT(seg1->s_ops == seg2->s_ops);
1038 
1039 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie) ||
1040 	    HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1041 		return (-1);
1042 	}
1043 
1044 	/* both segments exist, try to merge them */
1045 #define	incompat(x)	(svd1->x != svd2->x)
1046 	if (incompat(vp) || incompat(maxprot) ||
1047 	    (!svd1->pageadvice && !svd2->pageadvice && incompat(advice)) ||
1048 	    (!svd1->pageprot && !svd2->pageprot && incompat(prot)) ||
1049 	    incompat(type) || incompat(cred) || incompat(flags) ||
1050 	    seg1->s_szc != seg2->s_szc || incompat(policy_info.mem_policy) ||
1051 	    (svd2->softlockcnt > 0) || svd1->softlockcnt_send > 0)
1052 		return (-1);
1053 #undef incompat
1054 
1055 	/*
1056 	 * vp == NULL implies zfod, offset doesn't matter
1057 	 */
1058 	if (svd1->vp != NULL &&
1059 	    svd1->offset + seg1->s_size != svd2->offset) {
1060 		return (-1);
1061 	}
1062 
1063 	/*
1064 	 * Don't concatenate if either segment uses text replication.
1065 	 */
1066 	if (svd1->tr_state != SEGVN_TR_OFF || svd2->tr_state != SEGVN_TR_OFF) {
1067 		return (-1);
1068 	}
1069 
1070 	/*
1071 	 * Fail early if we're not supposed to concatenate
1072 	 * segments with non NULL amp.
1073 	 */
1074 	if (amp_cat == 0 && (amp1 != NULL || amp2 != NULL)) {
1075 		return (-1);
1076 	}
1077 
1078 	if (svd1->vp == NULL && svd1->type == MAP_SHARED) {
1079 		if (amp1 != amp2) {
1080 			return (-1);
1081 		}
1082 		if (amp1 != NULL && svd1->anon_index + btop(seg1->s_size) !=
1083 		    svd2->anon_index) {
1084 			return (-1);
1085 		}
1086 		ASSERT(amp1 == NULL || amp1->refcnt >= 2);
1087 	}
1088 
1089 	/*
1090 	 * If either seg has vpages, create a new merged vpage array.
1091 	 */
1092 	if (vpage1 != NULL || vpage2 != NULL) {
1093 		struct vpage *vp, *evp;
1094 
1095 		npages1 = seg_pages(seg1);
1096 		npages2 = seg_pages(seg2);
1097 		nvpsize = vpgtob(npages1 + npages2);
1098 
1099 		if ((nvpage = kmem_zalloc(nvpsize, KM_NOSLEEP)) == NULL) {
1100 			return (-2);
1101 		}
1102 
1103 		if (vpage1 != NULL) {
1104 			bcopy(vpage1, nvpage, vpgtob(npages1));
1105 		} else {
1106 			evp = nvpage + npages1;
1107 			for (vp = nvpage; vp < evp; vp++) {
1108 				VPP_SETPROT(vp, svd1->prot);
1109 				VPP_SETADVICE(vp, svd1->advice);
1110 			}
1111 		}
1112 
1113 		if (vpage2 != NULL) {
1114 			bcopy(vpage2, nvpage + npages1, vpgtob(npages2));
1115 		} else {
1116 			evp = nvpage + npages1 + npages2;
1117 			for (vp = nvpage + npages1; vp < evp; vp++) {
1118 				VPP_SETPROT(vp, svd2->prot);
1119 				VPP_SETADVICE(vp, svd2->advice);
1120 			}
1121 		}
1122 
1123 		if (svd2->pageswap && (!svd1->pageswap && svd1->swresv)) {
1124 			ASSERT(svd1->swresv == seg1->s_size);
1125 			ASSERT(!(svd1->flags & MAP_NORESERVE));
1126 			ASSERT(!(svd2->flags & MAP_NORESERVE));
1127 			evp = nvpage + npages1;
1128 			for (vp = nvpage; vp < evp; vp++) {
1129 				VPP_SETSWAPRES(vp);
1130 			}
1131 		}
1132 
1133 		if (svd1->pageswap && (!svd2->pageswap && svd2->swresv)) {
1134 			ASSERT(svd2->swresv == seg2->s_size);
1135 			ASSERT(!(svd1->flags & MAP_NORESERVE));
1136 			ASSERT(!(svd2->flags & MAP_NORESERVE));
1137 			vp = nvpage + npages1;
1138 			evp = vp + npages2;
1139 			for (; vp < evp; vp++) {
1140 				VPP_SETSWAPRES(vp);
1141 			}
1142 		}
1143 	}
1144 	ASSERT((vpage1 != NULL || vpage2 != NULL) ||
1145 	    (svd1->pageswap == 0 && svd2->pageswap == 0));
1146 
1147 	/*
1148 	 * If either segment has private pages, create a new merged anon
1149 	 * array. If mergeing shared anon segments just decrement anon map's
1150 	 * refcnt.
1151 	 */
1152 	if (amp1 != NULL && svd1->type == MAP_SHARED) {
1153 		ASSERT(amp1 == amp2 && svd1->vp == NULL);
1154 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1155 		ASSERT(amp1->refcnt >= 2);
1156 		amp1->refcnt--;
1157 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1158 		svd2->amp = NULL;
1159 	} else if (amp1 != NULL || amp2 != NULL) {
1160 		struct anon_hdr *nahp;
1161 		struct anon_map *namp = NULL;
1162 		size_t asize;
1163 
1164 		ASSERT(svd1->type == MAP_PRIVATE);
1165 
1166 		asize = seg1->s_size + seg2->s_size;
1167 		if ((nahp = anon_create(btop(asize), ANON_NOSLEEP)) == NULL) {
1168 			if (nvpage != NULL) {
1169 				kmem_free(nvpage, nvpsize);
1170 			}
1171 			return (-2);
1172 		}
1173 		if (amp1 != NULL) {
1174 			/*
1175 			 * XXX anon rwlock is not really needed because
1176 			 * this is a private segment and we are writers.
1177 			 */
1178 			ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1179 			ASSERT(amp1->refcnt == 1);
1180 			if (anon_copy_ptr(amp1->ahp, svd1->anon_index,
1181 			    nahp, 0, btop(seg1->s_size), ANON_NOSLEEP)) {
1182 				anon_release(nahp, btop(asize));
1183 				ANON_LOCK_EXIT(&amp1->a_rwlock);
1184 				if (nvpage != NULL) {
1185 					kmem_free(nvpage, nvpsize);
1186 				}
1187 				return (-2);
1188 			}
1189 		}
1190 		if (amp2 != NULL) {
1191 			ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1192 			ASSERT(amp2->refcnt == 1);
1193 			if (anon_copy_ptr(amp2->ahp, svd2->anon_index,
1194 			    nahp, btop(seg1->s_size), btop(seg2->s_size),
1195 			    ANON_NOSLEEP)) {
1196 				anon_release(nahp, btop(asize));
1197 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1198 				if (amp1 != NULL) {
1199 					ANON_LOCK_EXIT(&amp1->a_rwlock);
1200 				}
1201 				if (nvpage != NULL) {
1202 					kmem_free(nvpage, nvpsize);
1203 				}
1204 				return (-2);
1205 			}
1206 		}
1207 		if (amp1 != NULL) {
1208 			namp = amp1;
1209 			anon_release(amp1->ahp, btop(amp1->size));
1210 		}
1211 		if (amp2 != NULL) {
1212 			if (namp == NULL) {
1213 				ASSERT(amp1 == NULL);
1214 				namp = amp2;
1215 				anon_release(amp2->ahp, btop(amp2->size));
1216 			} else {
1217 				amp2->refcnt--;
1218 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1219 				anonmap_free(amp2);
1220 			}
1221 			svd2->amp = NULL; /* needed for seg_free */
1222 		}
1223 		namp->ahp = nahp;
1224 		namp->size = asize;
1225 		svd1->amp = namp;
1226 		svd1->anon_index = 0;
1227 		ANON_LOCK_EXIT(&namp->a_rwlock);
1228 	}
1229 	/*
1230 	 * Now free the old vpage structures.
1231 	 */
1232 	if (nvpage != NULL) {
1233 		if (vpage1 != NULL) {
1234 			kmem_free(vpage1, vpgtob(npages1));
1235 		}
1236 		if (vpage2 != NULL) {
1237 			svd2->vpage = NULL;
1238 			kmem_free(vpage2, vpgtob(npages2));
1239 		}
1240 		if (svd2->pageprot) {
1241 			svd1->pageprot = 1;
1242 		}
1243 		if (svd2->pageadvice) {
1244 			svd1->pageadvice = 1;
1245 		}
1246 		if (svd2->pageswap) {
1247 			svd1->pageswap = 1;
1248 		}
1249 		svd1->vpage = nvpage;
1250 	}
1251 
1252 	/* all looks ok, merge segments */
1253 	svd1->swresv += svd2->swresv;
1254 	svd2->swresv = 0;  /* so seg_free doesn't release swap space */
1255 	size = seg2->s_size;
1256 	seg_free(seg2);
1257 	seg1->s_size += size;
1258 	return (0);
1259 }
1260 
1261 /*
1262  * Extend the previous segment (seg1) to include the
1263  * new segment (seg2 + a), if possible.
1264  * Return 0 on success.
1265  */
1266 static int
1267 segvn_extend_prev(struct seg *seg1, struct seg *seg2, struct segvn_crargs *a,
1268     size_t swresv)
1269 {
1270 	struct segvn_data *svd1 = (struct segvn_data *)seg1->s_data;
1271 	size_t size;
1272 	struct anon_map *amp1;
1273 	struct vpage *new_vpage;
1274 
1275 	/*
1276 	 * We don't need any segment level locks for "segvn" data
1277 	 * since the address space is "write" locked.
1278 	 */
1279 	ASSERT(seg1->s_as && AS_WRITE_HELD(seg1->s_as));
1280 
1281 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie)) {
1282 		return (-1);
1283 	}
1284 
1285 	/* second segment is new, try to extend first */
1286 	/* XXX - should also check cred */
1287 	if (svd1->vp != a->vp || svd1->maxprot != a->maxprot ||
1288 	    (!svd1->pageprot && (svd1->prot != a->prot)) ||
1289 	    svd1->type != a->type || svd1->flags != a->flags ||
1290 	    seg1->s_szc != a->szc || svd1->softlockcnt_send > 0)
1291 		return (-1);
1292 
1293 	/* vp == NULL implies zfod, offset doesn't matter */
1294 	if (svd1->vp != NULL &&
1295 	    svd1->offset + seg1->s_size != (a->offset & PAGEMASK))
1296 		return (-1);
1297 
1298 	if (svd1->tr_state != SEGVN_TR_OFF) {
1299 		return (-1);
1300 	}
1301 
1302 	amp1 = svd1->amp;
1303 	if (amp1) {
1304 		pgcnt_t newpgs;
1305 
1306 		/*
1307 		 * Segment has private pages, can data structures
1308 		 * be expanded?
1309 		 *
1310 		 * Acquire the anon_map lock to prevent it from changing,
1311 		 * if it is shared.  This ensures that the anon_map
1312 		 * will not change while a thread which has a read/write
1313 		 * lock on an address space references it.
1314 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1315 		 * is 1.
1316 		 *
1317 		 * Can't grow a MAP_SHARED segment with an anonmap because
1318 		 * there may be existing anon slots where we want to extend
1319 		 * the segment and we wouldn't know what to do with them
1320 		 * (e.g., for tmpfs right thing is to just leave them there,
1321 		 * for /dev/zero they should be cleared out).
1322 		 */
1323 		if (svd1->type == MAP_SHARED)
1324 			return (-1);
1325 
1326 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1327 		if (amp1->refcnt > 1) {
1328 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1329 			return (-1);
1330 		}
1331 		newpgs = anon_grow(amp1->ahp, &svd1->anon_index,
1332 		    btop(seg1->s_size), btop(seg2->s_size), ANON_NOSLEEP);
1333 
1334 		if (newpgs == 0) {
1335 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1336 			return (-1);
1337 		}
1338 		amp1->size = ptob(newpgs);
1339 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1340 	}
1341 	if (svd1->vpage != NULL) {
1342 		struct vpage *vp, *evp;
1343 		new_vpage =
1344 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1345 		    KM_NOSLEEP);
1346 		if (new_vpage == NULL)
1347 			return (-1);
1348 		bcopy(svd1->vpage, new_vpage, vpgtob(seg_pages(seg1)));
1349 		kmem_free(svd1->vpage, vpgtob(seg_pages(seg1)));
1350 		svd1->vpage = new_vpage;
1351 
1352 		vp = new_vpage + seg_pages(seg1);
1353 		evp = vp + seg_pages(seg2);
1354 		for (; vp < evp; vp++)
1355 			VPP_SETPROT(vp, a->prot);
1356 		if (svd1->pageswap && swresv) {
1357 			ASSERT(!(svd1->flags & MAP_NORESERVE));
1358 			ASSERT(swresv == seg2->s_size);
1359 			vp = new_vpage + seg_pages(seg1);
1360 			for (; vp < evp; vp++) {
1361 				VPP_SETSWAPRES(vp);
1362 			}
1363 		}
1364 	}
1365 	ASSERT(svd1->vpage != NULL || svd1->pageswap == 0);
1366 	size = seg2->s_size;
1367 	seg_free(seg2);
1368 	seg1->s_size += size;
1369 	svd1->swresv += swresv;
1370 	if (svd1->pageprot && (a->prot & PROT_WRITE) &&
1371 	    svd1->type == MAP_SHARED && svd1->vp != NULL &&
1372 	    (svd1->vp->v_flag & VVMEXEC)) {
1373 		ASSERT(vn_is_mapped(svd1->vp, V_WRITE));
1374 		segvn_inval_trcache(svd1->vp);
1375 	}
1376 	return (0);
1377 }
1378 
1379 /*
1380  * Extend the next segment (seg2) to include the
1381  * new segment (seg1 + a), if possible.
1382  * Return 0 on success.
1383  */
1384 static int
1385 segvn_extend_next(struct seg *seg1, struct seg *seg2, struct segvn_crargs *a,
1386     size_t swresv)
1387 {
1388 	struct segvn_data *svd2 = (struct segvn_data *)seg2->s_data;
1389 	size_t size;
1390 	struct anon_map *amp2;
1391 	struct vpage *new_vpage;
1392 
1393 	/*
1394 	 * We don't need any segment level locks for "segvn" data
1395 	 * since the address space is "write" locked.
1396 	 */
1397 	ASSERT(seg2->s_as && AS_WRITE_HELD(seg2->s_as));
1398 
1399 	if (HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1400 		return (-1);
1401 	}
1402 
1403 	/* first segment is new, try to extend second */
1404 	/* XXX - should also check cred */
1405 	if (svd2->vp != a->vp || svd2->maxprot != a->maxprot ||
1406 	    (!svd2->pageprot && (svd2->prot != a->prot)) ||
1407 	    svd2->type != a->type || svd2->flags != a->flags ||
1408 	    seg2->s_szc != a->szc || svd2->softlockcnt_sbase > 0)
1409 		return (-1);
1410 	/* vp == NULL implies zfod, offset doesn't matter */
1411 	if (svd2->vp != NULL &&
1412 	    (a->offset & PAGEMASK) + seg1->s_size != svd2->offset)
1413 		return (-1);
1414 
1415 	if (svd2->tr_state != SEGVN_TR_OFF) {
1416 		return (-1);
1417 	}
1418 
1419 	amp2 = svd2->amp;
1420 	if (amp2) {
1421 		pgcnt_t newpgs;
1422 
1423 		/*
1424 		 * Segment has private pages, can data structures
1425 		 * be expanded?
1426 		 *
1427 		 * Acquire the anon_map lock to prevent it from changing,
1428 		 * if it is shared.  This ensures that the anon_map
1429 		 * will not change while a thread which has a read/write
1430 		 * lock on an address space references it.
1431 		 *
1432 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1433 		 * is 1.
1434 		 */
1435 		if (svd2->type == MAP_SHARED)
1436 			return (-1);
1437 
1438 		ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1439 		if (amp2->refcnt > 1) {
1440 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1441 			return (-1);
1442 		}
1443 		newpgs = anon_grow(amp2->ahp, &svd2->anon_index,
1444 		    btop(seg2->s_size), btop(seg1->s_size),
1445 		    ANON_NOSLEEP | ANON_GROWDOWN);
1446 
1447 		if (newpgs == 0) {
1448 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1449 			return (-1);
1450 		}
1451 		amp2->size = ptob(newpgs);
1452 		ANON_LOCK_EXIT(&amp2->a_rwlock);
1453 	}
1454 	if (svd2->vpage != NULL) {
1455 		struct vpage *vp, *evp;
1456 		new_vpage =
1457 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1458 		    KM_NOSLEEP);
1459 		if (new_vpage == NULL) {
1460 			/* Not merging segments so adjust anon_index back */
1461 			if (amp2)
1462 				svd2->anon_index += seg_pages(seg1);
1463 			return (-1);
1464 		}
1465 		bcopy(svd2->vpage, new_vpage + seg_pages(seg1),
1466 		    vpgtob(seg_pages(seg2)));
1467 		kmem_free(svd2->vpage, vpgtob(seg_pages(seg2)));
1468 		svd2->vpage = new_vpage;
1469 
1470 		vp = new_vpage;
1471 		evp = vp + seg_pages(seg1);
1472 		for (; vp < evp; vp++)
1473 			VPP_SETPROT(vp, a->prot);
1474 		if (svd2->pageswap && swresv) {
1475 			ASSERT(!(svd2->flags & MAP_NORESERVE));
1476 			ASSERT(swresv == seg1->s_size);
1477 			vp = new_vpage;
1478 			for (; vp < evp; vp++) {
1479 				VPP_SETSWAPRES(vp);
1480 			}
1481 		}
1482 	}
1483 	ASSERT(svd2->vpage != NULL || svd2->pageswap == 0);
1484 	size = seg1->s_size;
1485 	seg_free(seg1);
1486 	seg2->s_size += size;
1487 	seg2->s_base -= size;
1488 	svd2->offset -= size;
1489 	svd2->swresv += swresv;
1490 	if (svd2->pageprot && (a->prot & PROT_WRITE) &&
1491 	    svd2->type == MAP_SHARED && svd2->vp != NULL &&
1492 	    (svd2->vp->v_flag & VVMEXEC)) {
1493 		ASSERT(vn_is_mapped(svd2->vp, V_WRITE));
1494 		segvn_inval_trcache(svd2->vp);
1495 	}
1496 	return (0);
1497 }
1498 
1499 /*
1500  * Duplicate all the pages in the segment. This may break COW sharing for a
1501  * given page. If the page is marked with inherit zero set, then instead of
1502  * duplicating the page, we zero the page.
1503  */
1504 static int
1505 segvn_dup_pages(struct seg *seg, struct seg *newseg)
1506 {
1507 	int error;
1508 	uint_t prot;
1509 	page_t *pp;
1510 	struct anon *ap, *newap;
1511 	size_t i;
1512 	caddr_t addr;
1513 
1514 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1515 	struct segvn_data *newsvd = (struct segvn_data *)newseg->s_data;
1516 	ulong_t old_idx = svd->anon_index;
1517 	ulong_t new_idx = 0;
1518 
1519 	i = btopr(seg->s_size);
1520 	addr = seg->s_base;
1521 
1522 	/*
1523 	 * XXX break cow sharing using PAGESIZE
1524 	 * pages. They will be relocated into larger
1525 	 * pages at fault time.
1526 	 */
1527 	while (i-- > 0) {
1528 		if ((ap = anon_get_ptr(svd->amp->ahp, old_idx)) != NULL) {
1529 			struct vpage *vpp;
1530 
1531 			vpp = &svd->vpage[seg_page(seg, addr)];
1532 
1533 			/*
1534 			 * prot need not be computed below 'cause anon_private
1535 			 * is going to ignore it anyway as child doesn't inherit
1536 			 * pagelock from parent.
1537 			 */
1538 			prot = svd->pageprot ? VPP_PROT(vpp) : svd->prot;
1539 
1540 			/*
1541 			 * Check whether we should zero this or dup it.
1542 			 */
1543 			if (svd->svn_inz == SEGVN_INZ_ALL ||
1544 			    (svd->svn_inz == SEGVN_INZ_VPP &&
1545 			    VPP_ISINHZERO(vpp))) {
1546 				pp = anon_zero(newseg, addr, &newap,
1547 				    newsvd->cred);
1548 			} else {
1549 				page_t *anon_pl[1+1];
1550 				uint_t vpprot;
1551 				error = anon_getpage(&ap, &vpprot, anon_pl,
1552 				    PAGESIZE, seg, addr, S_READ, svd->cred);
1553 				if (error != 0)
1554 					return (error);
1555 
1556 				pp = anon_private(&newap, newseg, addr, prot,
1557 				    anon_pl[0], 0, newsvd->cred);
1558 			}
1559 			if (pp == NULL) {
1560 				return (ENOMEM);
1561 			}
1562 			(void) anon_set_ptr(newsvd->amp->ahp, new_idx, newap,
1563 			    ANON_SLEEP);
1564 			page_unlock(pp);
1565 		}
1566 		addr += PAGESIZE;
1567 		old_idx++;
1568 		new_idx++;
1569 	}
1570 
1571 	return (0);
1572 }
1573 
1574 static int
1575 segvn_dup(struct seg *seg, struct seg *newseg)
1576 {
1577 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1578 	struct segvn_data *newsvd;
1579 	pgcnt_t npages = seg_pages(seg);
1580 	int error = 0;
1581 	size_t len;
1582 	struct anon_map *amp;
1583 
1584 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1585 	ASSERT(newseg->s_as->a_proc->p_parent == curproc);
1586 
1587 	/*
1588 	 * If segment has anon reserved, reserve more for the new seg.
1589 	 * For a MAP_NORESERVE segment swresv will be a count of all the
1590 	 * allocated anon slots; thus we reserve for the child as many slots
1591 	 * as the parent has allocated. This semantic prevents the child or
1592 	 * parent from dieing during a copy-on-write fault caused by trying
1593 	 * to write a shared pre-existing anon page.
1594 	 */
1595 	if ((len = svd->swresv) != 0) {
1596 		if (anon_resv(svd->swresv) == 0)
1597 			return (ENOMEM);
1598 
1599 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
1600 		    seg, len, 0);
1601 	}
1602 
1603 	newsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
1604 
1605 	newseg->s_ops = &segvn_ops;
1606 	newseg->s_data = (void *)newsvd;
1607 	newseg->s_szc = seg->s_szc;
1608 
1609 	newsvd->seg = newseg;
1610 	if ((newsvd->vp = svd->vp) != NULL) {
1611 		VN_HOLD(svd->vp);
1612 		if (svd->type == MAP_SHARED)
1613 			lgrp_shm_policy_init(NULL, svd->vp);
1614 	}
1615 	newsvd->offset = svd->offset;
1616 	newsvd->prot = svd->prot;
1617 	newsvd->maxprot = svd->maxprot;
1618 	newsvd->pageprot = svd->pageprot;
1619 	newsvd->type = svd->type;
1620 	newsvd->cred = svd->cred;
1621 	crhold(newsvd->cred);
1622 	newsvd->advice = svd->advice;
1623 	newsvd->pageadvice = svd->pageadvice;
1624 	newsvd->svn_inz = svd->svn_inz;
1625 	newsvd->swresv = svd->swresv;
1626 	newsvd->pageswap = svd->pageswap;
1627 	newsvd->flags = svd->flags;
1628 	newsvd->softlockcnt = 0;
1629 	newsvd->softlockcnt_sbase = 0;
1630 	newsvd->softlockcnt_send = 0;
1631 	newsvd->policy_info = svd->policy_info;
1632 	newsvd->rcookie = HAT_INVALID_REGION_COOKIE;
1633 
1634 	if ((amp = svd->amp) == NULL || svd->tr_state == SEGVN_TR_ON) {
1635 		/*
1636 		 * Not attaching to a shared anon object.
1637 		 */
1638 		ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie) ||
1639 		    svd->tr_state == SEGVN_TR_OFF);
1640 		if (svd->tr_state == SEGVN_TR_ON) {
1641 			ASSERT(newsvd->vp != NULL && amp != NULL);
1642 			newsvd->tr_state = SEGVN_TR_INIT;
1643 		} else {
1644 			newsvd->tr_state = svd->tr_state;
1645 		}
1646 		newsvd->amp = NULL;
1647 		newsvd->anon_index = 0;
1648 	} else {
1649 		/* regions for now are only used on pure vnode segments */
1650 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
1651 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1652 		newsvd->tr_state = SEGVN_TR_OFF;
1653 		if (svd->type == MAP_SHARED) {
1654 			ASSERT(svd->svn_inz == SEGVN_INZ_NONE);
1655 			newsvd->amp = amp;
1656 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1657 			amp->refcnt++;
1658 			ANON_LOCK_EXIT(&amp->a_rwlock);
1659 			newsvd->anon_index = svd->anon_index;
1660 		} else {
1661 			int reclaim = 1;
1662 
1663 			/*
1664 			 * Allocate and initialize new anon_map structure.
1665 			 */
1666 			newsvd->amp = anonmap_alloc(newseg->s_size, 0,
1667 			    ANON_SLEEP);
1668 			newsvd->amp->a_szc = newseg->s_szc;
1669 			newsvd->anon_index = 0;
1670 			ASSERT(svd->svn_inz == SEGVN_INZ_NONE ||
1671 			    svd->svn_inz == SEGVN_INZ_ALL ||
1672 			    svd->svn_inz == SEGVN_INZ_VPP);
1673 
1674 			/*
1675 			 * We don't have to acquire the anon_map lock
1676 			 * for the new segment (since it belongs to an
1677 			 * address space that is still not associated
1678 			 * with any process), or the segment in the old
1679 			 * address space (since all threads in it
1680 			 * are stopped while duplicating the address space).
1681 			 */
1682 
1683 			/*
1684 			 * The goal of the following code is to make sure that
1685 			 * softlocked pages do not end up as copy on write
1686 			 * pages.  This would cause problems where one
1687 			 * thread writes to a page that is COW and a different
1688 			 * thread in the same process has softlocked it.  The
1689 			 * softlock lock would move away from this process
1690 			 * because the write would cause this process to get
1691 			 * a copy (without the softlock).
1692 			 *
1693 			 * The strategy here is to just break the
1694 			 * sharing on pages that could possibly be
1695 			 * softlocked.
1696 			 *
1697 			 * In addition, if any pages have been marked that they
1698 			 * should be inherited as zero, then we immediately go
1699 			 * ahead and break COW and zero them. In the case of a
1700 			 * softlocked page that should be inherited zero, we
1701 			 * break COW and just get a zero page.
1702 			 */
1703 retry:
1704 			if (svd->softlockcnt ||
1705 			    svd->svn_inz != SEGVN_INZ_NONE) {
1706 				/*
1707 				 * The softlock count might be non zero
1708 				 * because some pages are still stuck in the
1709 				 * cache for lazy reclaim. Flush the cache
1710 				 * now. This should drop the count to zero.
1711 				 * [or there is really I/O going on to these
1712 				 * pages]. Note, we have the writers lock so
1713 				 * nothing gets inserted during the flush.
1714 				 */
1715 				if (svd->softlockcnt && reclaim == 1) {
1716 					segvn_purge(seg);
1717 					reclaim = 0;
1718 					goto retry;
1719 				}
1720 
1721 				error = segvn_dup_pages(seg, newseg);
1722 				if (error != 0) {
1723 					newsvd->vpage = NULL;
1724 					goto out;
1725 				}
1726 			} else {	/* common case */
1727 				if (seg->s_szc != 0) {
1728 					/*
1729 					 * If at least one of anon slots of a
1730 					 * large page exists then make sure
1731 					 * all anon slots of a large page
1732 					 * exist to avoid partial cow sharing
1733 					 * of a large page in the future.
1734 					 */
1735 					anon_dup_fill_holes(amp->ahp,
1736 					    svd->anon_index, newsvd->amp->ahp,
1737 					    0, seg->s_size, seg->s_szc,
1738 					    svd->vp != NULL);
1739 				} else {
1740 					anon_dup(amp->ahp, svd->anon_index,
1741 					    newsvd->amp->ahp, 0, seg->s_size);
1742 				}
1743 
1744 				hat_clrattr(seg->s_as->a_hat, seg->s_base,
1745 				    seg->s_size, PROT_WRITE);
1746 			}
1747 		}
1748 	}
1749 	/*
1750 	 * If necessary, create a vpage structure for the new segment.
1751 	 * Do not copy any page lock indications.
1752 	 */
1753 	if (svd->vpage != NULL) {
1754 		uint_t i;
1755 		struct vpage *ovp = svd->vpage;
1756 		struct vpage *nvp;
1757 
1758 		nvp = newsvd->vpage =
1759 		    kmem_alloc(vpgtob(npages), KM_SLEEP);
1760 		for (i = 0; i < npages; i++) {
1761 			*nvp = *ovp++;
1762 			VPP_CLRPPLOCK(nvp++);
1763 		}
1764 	} else
1765 		newsvd->vpage = NULL;
1766 
1767 	/* Inform the vnode of the new mapping */
1768 	if (newsvd->vp != NULL) {
1769 		error = VOP_ADDMAP(newsvd->vp, (offset_t)newsvd->offset,
1770 		    newseg->s_as, newseg->s_base, newseg->s_size, newsvd->prot,
1771 		    newsvd->maxprot, newsvd->type, newsvd->cred, NULL);
1772 	}
1773 out:
1774 	if (error == 0 && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1775 		ASSERT(newsvd->amp == NULL);
1776 		ASSERT(newsvd->tr_state == SEGVN_TR_OFF);
1777 		newsvd->rcookie = svd->rcookie;
1778 		hat_dup_region(newseg->s_as->a_hat, newsvd->rcookie);
1779 	}
1780 	return (error);
1781 }
1782 
1783 
1784 /*
1785  * callback function to invoke free_vp_pages() for only those pages actually
1786  * processed by the HAT when a shared region is destroyed.
1787  */
1788 extern int free_pages;
1789 
1790 static void
1791 segvn_hat_rgn_unload_callback(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
1792     size_t r_size, void *r_obj, u_offset_t r_objoff)
1793 {
1794 	u_offset_t off;
1795 	size_t len;
1796 	vnode_t *vp = (vnode_t *)r_obj;
1797 
1798 	ASSERT(eaddr > saddr);
1799 	ASSERT(saddr >= r_saddr);
1800 	ASSERT(saddr < r_saddr + r_size);
1801 	ASSERT(eaddr > r_saddr);
1802 	ASSERT(eaddr <= r_saddr + r_size);
1803 	ASSERT(vp != NULL);
1804 
1805 	if (!free_pages) {
1806 		return;
1807 	}
1808 
1809 	len = eaddr - saddr;
1810 	off = (saddr - r_saddr) + r_objoff;
1811 	free_vp_pages(vp, off, len);
1812 }
1813 
1814 /*
1815  * callback function used by segvn_unmap to invoke free_vp_pages() for only
1816  * those pages actually processed by the HAT
1817  */
1818 static void
1819 segvn_hat_unload_callback(hat_callback_t *cb)
1820 {
1821 	struct seg		*seg = cb->hcb_data;
1822 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
1823 	size_t			len;
1824 	u_offset_t		off;
1825 
1826 	ASSERT(svd->vp != NULL);
1827 	ASSERT(cb->hcb_end_addr > cb->hcb_start_addr);
1828 	ASSERT(cb->hcb_start_addr >= seg->s_base);
1829 
1830 	len = cb->hcb_end_addr - cb->hcb_start_addr;
1831 	off = cb->hcb_start_addr - seg->s_base;
1832 	free_vp_pages(svd->vp, svd->offset + off, len);
1833 }
1834 
1835 /*
1836  * This function determines the number of bytes of swap reserved by
1837  * a segment for which per-page accounting is present. It is used to
1838  * calculate the correct value of a segvn_data's swresv.
1839  */
1840 static size_t
1841 segvn_count_swap_by_vpages(struct seg *seg)
1842 {
1843 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1844 	struct vpage *vp, *evp;
1845 	size_t nswappages = 0;
1846 
1847 	ASSERT(svd->pageswap);
1848 	ASSERT(svd->vpage != NULL);
1849 
1850 	evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
1851 
1852 	for (vp = svd->vpage; vp < evp; vp++) {
1853 		if (VPP_ISSWAPRES(vp))
1854 			nswappages++;
1855 	}
1856 
1857 	return (nswappages << PAGESHIFT);
1858 }
1859 
1860 static int
1861 segvn_unmap(struct seg *seg, caddr_t addr, size_t len)
1862 {
1863 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1864 	struct segvn_data *nsvd;
1865 	struct seg *nseg;
1866 	struct anon_map *amp;
1867 	pgcnt_t	opages;		/* old segment size in pages */
1868 	pgcnt_t	npages;		/* new segment size in pages */
1869 	pgcnt_t	dpages;		/* pages being deleted (unmapped) */
1870 	hat_callback_t callback;	/* used for free_vp_pages() */
1871 	hat_callback_t *cbp = NULL;
1872 	caddr_t nbase;
1873 	size_t nsize;
1874 	size_t oswresv;
1875 	int reclaim = 1;
1876 
1877 	/*
1878 	 * We don't need any segment level locks for "segvn" data
1879 	 * since the address space is "write" locked.
1880 	 */
1881 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1882 
1883 	/*
1884 	 * Fail the unmap if pages are SOFTLOCKed through this mapping.
1885 	 * softlockcnt is protected from change by the as write lock.
1886 	 */
1887 retry:
1888 	if (svd->softlockcnt > 0) {
1889 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1890 
1891 		/*
1892 		 * If this is shared segment non 0 softlockcnt
1893 		 * means locked pages are still in use.
1894 		 */
1895 		if (svd->type == MAP_SHARED) {
1896 			return (EAGAIN);
1897 		}
1898 
1899 		/*
1900 		 * since we do have the writers lock nobody can fill
1901 		 * the cache during the purge. The flush either succeeds
1902 		 * or we still have pending I/Os.
1903 		 */
1904 		if (reclaim == 1) {
1905 			segvn_purge(seg);
1906 			reclaim = 0;
1907 			goto retry;
1908 		}
1909 		return (EAGAIN);
1910 	}
1911 
1912 	/*
1913 	 * Check for bad sizes
1914 	 */
1915 	if (addr < seg->s_base || addr + len > seg->s_base + seg->s_size ||
1916 	    (len & PAGEOFFSET) || ((uintptr_t)addr & PAGEOFFSET)) {
1917 		panic("segvn_unmap");
1918 		/*NOTREACHED*/
1919 	}
1920 
1921 	if (seg->s_szc != 0) {
1922 		size_t pgsz = page_get_pagesize(seg->s_szc);
1923 		int err;
1924 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
1925 			ASSERT(seg->s_base != addr || seg->s_size != len);
1926 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1927 				ASSERT(svd->amp == NULL);
1928 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1929 				hat_leave_region(seg->s_as->a_hat,
1930 				    svd->rcookie, HAT_REGION_TEXT);
1931 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
1932 				/*
1933 				 * could pass a flag to segvn_demote_range()
1934 				 * below to tell it not to do any unloads but
1935 				 * this case is rare enough to not bother for
1936 				 * now.
1937 				 */
1938 			} else if (svd->tr_state == SEGVN_TR_INIT) {
1939 				svd->tr_state = SEGVN_TR_OFF;
1940 			} else if (svd->tr_state == SEGVN_TR_ON) {
1941 				ASSERT(svd->amp != NULL);
1942 				segvn_textunrepl(seg, 1);
1943 				ASSERT(svd->amp == NULL);
1944 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1945 			}
1946 			VM_STAT_ADD(segvnvmstats.demoterange[0]);
1947 			err = segvn_demote_range(seg, addr, len, SDR_END, 0);
1948 			if (err == 0) {
1949 				return (IE_RETRY);
1950 			}
1951 			return (err);
1952 		}
1953 	}
1954 
1955 	/* Inform the vnode of the unmapping. */
1956 	if (svd->vp) {
1957 		int error;
1958 
1959 		error = VOP_DELMAP(svd->vp,
1960 		    (offset_t)svd->offset + (uintptr_t)(addr - seg->s_base),
1961 		    seg->s_as, addr, len, svd->prot, svd->maxprot,
1962 		    svd->type, svd->cred, NULL);
1963 
1964 		if (error == EAGAIN)
1965 			return (error);
1966 	}
1967 
1968 	/*
1969 	 * Remove any page locks set through this mapping.
1970 	 * If text replication is not off no page locks could have been
1971 	 * established via this mapping.
1972 	 */
1973 	if (svd->tr_state == SEGVN_TR_OFF) {
1974 		(void) segvn_lockop(seg, addr, len, 0, MC_UNLOCK, NULL, 0);
1975 	}
1976 
1977 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1978 		ASSERT(svd->amp == NULL);
1979 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1980 		ASSERT(svd->type == MAP_PRIVATE);
1981 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
1982 		    HAT_REGION_TEXT);
1983 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
1984 	} else if (svd->tr_state == SEGVN_TR_ON) {
1985 		ASSERT(svd->amp != NULL);
1986 		ASSERT(svd->pageprot == 0 && !(svd->prot & PROT_WRITE));
1987 		segvn_textunrepl(seg, 1);
1988 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
1989 	} else {
1990 		if (svd->tr_state != SEGVN_TR_OFF) {
1991 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
1992 			svd->tr_state = SEGVN_TR_OFF;
1993 		}
1994 		/*
1995 		 * Unload any hardware translations in the range to be taken
1996 		 * out. Use a callback to invoke free_vp_pages() effectively.
1997 		 */
1998 		if (svd->vp != NULL && free_pages != 0) {
1999 			callback.hcb_data = seg;
2000 			callback.hcb_function = segvn_hat_unload_callback;
2001 			cbp = &callback;
2002 		}
2003 		hat_unload_callback(seg->s_as->a_hat, addr, len,
2004 		    HAT_UNLOAD_UNMAP, cbp);
2005 
2006 		if (svd->type == MAP_SHARED && svd->vp != NULL &&
2007 		    (svd->vp->v_flag & VVMEXEC) &&
2008 		    ((svd->prot & PROT_WRITE) || svd->pageprot)) {
2009 			segvn_inval_trcache(svd->vp);
2010 		}
2011 	}
2012 
2013 	/*
2014 	 * Check for entire segment
2015 	 */
2016 	if (addr == seg->s_base && len == seg->s_size) {
2017 		seg_free(seg);
2018 		return (0);
2019 	}
2020 
2021 	opages = seg_pages(seg);
2022 	dpages = btop(len);
2023 	npages = opages - dpages;
2024 	amp = svd->amp;
2025 	ASSERT(amp == NULL || amp->a_szc >= seg->s_szc);
2026 
2027 	/*
2028 	 * Check for beginning of segment
2029 	 */
2030 	if (addr == seg->s_base) {
2031 		if (svd->vpage != NULL) {
2032 			size_t nbytes;
2033 			struct vpage *ovpage;
2034 
2035 			ovpage = svd->vpage;	/* keep pointer to vpage */
2036 
2037 			nbytes = vpgtob(npages);
2038 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2039 			bcopy(&ovpage[dpages], svd->vpage, nbytes);
2040 
2041 			/* free up old vpage */
2042 			kmem_free(ovpage, vpgtob(opages));
2043 		}
2044 		if (amp != NULL) {
2045 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2046 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2047 				/*
2048 				 * Shared anon map is no longer in use. Before
2049 				 * freeing its pages purge all entries from
2050 				 * pcache that belong to this amp.
2051 				 */
2052 				if (svd->type == MAP_SHARED) {
2053 					ASSERT(amp->refcnt == 1);
2054 					ASSERT(svd->softlockcnt == 0);
2055 					anonmap_purge(amp);
2056 				}
2057 				/*
2058 				 * Free up now unused parts of anon_map array.
2059 				 */
2060 				if (amp->a_szc == seg->s_szc) {
2061 					if (seg->s_szc != 0) {
2062 						anon_free_pages(amp->ahp,
2063 						    svd->anon_index, len,
2064 						    seg->s_szc);
2065 					} else {
2066 						anon_free(amp->ahp,
2067 						    svd->anon_index,
2068 						    len);
2069 					}
2070 				} else {
2071 					ASSERT(svd->type == MAP_SHARED);
2072 					ASSERT(amp->a_szc > seg->s_szc);
2073 					anon_shmap_free_pages(amp,
2074 					    svd->anon_index, len);
2075 				}
2076 
2077 				/*
2078 				 * Unreserve swap space for the
2079 				 * unmapped chunk of this segment in
2080 				 * case it's MAP_SHARED
2081 				 */
2082 				if (svd->type == MAP_SHARED) {
2083 					anon_unresv_zone(len,
2084 					    seg->s_as->a_proc->p_zone);
2085 					amp->swresv -= len;
2086 				}
2087 			}
2088 			ANON_LOCK_EXIT(&amp->a_rwlock);
2089 			svd->anon_index += dpages;
2090 		}
2091 		if (svd->vp != NULL)
2092 			svd->offset += len;
2093 
2094 		seg->s_base += len;
2095 		seg->s_size -= len;
2096 
2097 		if (svd->swresv) {
2098 			if (svd->flags & MAP_NORESERVE) {
2099 				ASSERT(amp);
2100 				oswresv = svd->swresv;
2101 
2102 				svd->swresv = ptob(anon_pages(amp->ahp,
2103 				    svd->anon_index, npages));
2104 				anon_unresv_zone(oswresv - svd->swresv,
2105 				    seg->s_as->a_proc->p_zone);
2106 				if (SEG_IS_PARTIAL_RESV(seg))
2107 					seg->s_as->a_resvsize -= oswresv -
2108 					    svd->swresv;
2109 			} else {
2110 				size_t unlen;
2111 
2112 				if (svd->pageswap) {
2113 					oswresv = svd->swresv;
2114 					svd->swresv =
2115 					    segvn_count_swap_by_vpages(seg);
2116 					ASSERT(oswresv >= svd->swresv);
2117 					unlen = oswresv - svd->swresv;
2118 				} else {
2119 					svd->swresv -= len;
2120 					ASSERT(svd->swresv == seg->s_size);
2121 					unlen = len;
2122 				}
2123 				anon_unresv_zone(unlen,
2124 				    seg->s_as->a_proc->p_zone);
2125 			}
2126 			TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2127 			    seg, len, 0);
2128 		}
2129 
2130 		return (0);
2131 	}
2132 
2133 	/*
2134 	 * Check for end of segment
2135 	 */
2136 	if (addr + len == seg->s_base + seg->s_size) {
2137 		if (svd->vpage != NULL) {
2138 			size_t nbytes;
2139 			struct vpage *ovpage;
2140 
2141 			ovpage = svd->vpage;	/* keep pointer to vpage */
2142 
2143 			nbytes = vpgtob(npages);
2144 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2145 			bcopy(ovpage, svd->vpage, nbytes);
2146 
2147 			/* free up old vpage */
2148 			kmem_free(ovpage, vpgtob(opages));
2149 
2150 		}
2151 		if (amp != NULL) {
2152 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2153 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2154 				/*
2155 				 * Free up now unused parts of anon_map array.
2156 				 */
2157 				ulong_t an_idx = svd->anon_index + npages;
2158 
2159 				/*
2160 				 * Shared anon map is no longer in use. Before
2161 				 * freeing its pages purge all entries from
2162 				 * pcache that belong to this amp.
2163 				 */
2164 				if (svd->type == MAP_SHARED) {
2165 					ASSERT(amp->refcnt == 1);
2166 					ASSERT(svd->softlockcnt == 0);
2167 					anonmap_purge(amp);
2168 				}
2169 
2170 				if (amp->a_szc == seg->s_szc) {
2171 					if (seg->s_szc != 0) {
2172 						anon_free_pages(amp->ahp,
2173 						    an_idx, len,
2174 						    seg->s_szc);
2175 					} else {
2176 						anon_free(amp->ahp, an_idx,
2177 						    len);
2178 					}
2179 				} else {
2180 					ASSERT(svd->type == MAP_SHARED);
2181 					ASSERT(amp->a_szc > seg->s_szc);
2182 					anon_shmap_free_pages(amp,
2183 					    an_idx, len);
2184 				}
2185 
2186 				/*
2187 				 * Unreserve swap space for the
2188 				 * unmapped chunk of this segment in
2189 				 * case it's MAP_SHARED
2190 				 */
2191 				if (svd->type == MAP_SHARED) {
2192 					anon_unresv_zone(len,
2193 					    seg->s_as->a_proc->p_zone);
2194 					amp->swresv -= len;
2195 				}
2196 			}
2197 			ANON_LOCK_EXIT(&amp->a_rwlock);
2198 		}
2199 
2200 		seg->s_size -= len;
2201 
2202 		if (svd->swresv) {
2203 			if (svd->flags & MAP_NORESERVE) {
2204 				ASSERT(amp);
2205 				oswresv = svd->swresv;
2206 				svd->swresv = ptob(anon_pages(amp->ahp,
2207 				    svd->anon_index, npages));
2208 				anon_unresv_zone(oswresv - svd->swresv,
2209 				    seg->s_as->a_proc->p_zone);
2210 				if (SEG_IS_PARTIAL_RESV(seg))
2211 					seg->s_as->a_resvsize -= oswresv -
2212 					    svd->swresv;
2213 			} else {
2214 				size_t unlen;
2215 
2216 				if (svd->pageswap) {
2217 					oswresv = svd->swresv;
2218 					svd->swresv =
2219 					    segvn_count_swap_by_vpages(seg);
2220 					ASSERT(oswresv >= svd->swresv);
2221 					unlen = oswresv - svd->swresv;
2222 				} else {
2223 					svd->swresv -= len;
2224 					ASSERT(svd->swresv == seg->s_size);
2225 					unlen = len;
2226 				}
2227 				anon_unresv_zone(unlen,
2228 				    seg->s_as->a_proc->p_zone);
2229 			}
2230 			TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2231 			    "anon proc:%p %lu %u", seg, len, 0);
2232 		}
2233 
2234 		return (0);
2235 	}
2236 
2237 	/*
2238 	 * The section to go is in the middle of the segment,
2239 	 * have to make it into two segments.  nseg is made for
2240 	 * the high end while seg is cut down at the low end.
2241 	 */
2242 	nbase = addr + len;				/* new seg base */
2243 	nsize = (seg->s_base + seg->s_size) - nbase;	/* new seg size */
2244 	seg->s_size = addr - seg->s_base;		/* shrink old seg */
2245 	nseg = seg_alloc(seg->s_as, nbase, nsize);
2246 	if (nseg == NULL) {
2247 		panic("segvn_unmap seg_alloc");
2248 		/*NOTREACHED*/
2249 	}
2250 	nseg->s_ops = seg->s_ops;
2251 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
2252 	nseg->s_data = (void *)nsvd;
2253 	nseg->s_szc = seg->s_szc;
2254 	*nsvd = *svd;
2255 	nsvd->seg = nseg;
2256 	nsvd->offset = svd->offset + (uintptr_t)(nseg->s_base - seg->s_base);
2257 	nsvd->swresv = 0;
2258 	nsvd->softlockcnt = 0;
2259 	nsvd->softlockcnt_sbase = 0;
2260 	nsvd->softlockcnt_send = 0;
2261 	nsvd->svn_inz = svd->svn_inz;
2262 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
2263 
2264 	if (svd->vp != NULL) {
2265 		VN_HOLD(nsvd->vp);
2266 		if (nsvd->type == MAP_SHARED)
2267 			lgrp_shm_policy_init(NULL, nsvd->vp);
2268 	}
2269 	crhold(svd->cred);
2270 
2271 	if (svd->vpage == NULL) {
2272 		nsvd->vpage = NULL;
2273 	} else {
2274 		/* need to split vpage into two arrays */
2275 		size_t nbytes;
2276 		struct vpage *ovpage;
2277 
2278 		ovpage = svd->vpage;		/* keep pointer to vpage */
2279 
2280 		npages = seg_pages(seg);	/* seg has shrunk */
2281 		nbytes = vpgtob(npages);
2282 		svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2283 
2284 		bcopy(ovpage, svd->vpage, nbytes);
2285 
2286 		npages = seg_pages(nseg);
2287 		nbytes = vpgtob(npages);
2288 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2289 
2290 		bcopy(&ovpage[opages - npages], nsvd->vpage, nbytes);
2291 
2292 		/* free up old vpage */
2293 		kmem_free(ovpage, vpgtob(opages));
2294 	}
2295 
2296 	if (amp == NULL) {
2297 		nsvd->amp = NULL;
2298 		nsvd->anon_index = 0;
2299 	} else {
2300 		/*
2301 		 * Need to create a new anon map for the new segment.
2302 		 * We'll also allocate a new smaller array for the old
2303 		 * smaller segment to save space.
2304 		 */
2305 		opages = btop((uintptr_t)(addr - seg->s_base));
2306 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2307 		if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2308 			/*
2309 			 * Free up now unused parts of anon_map array.
2310 			 */
2311 			ulong_t an_idx = svd->anon_index + opages;
2312 
2313 			/*
2314 			 * Shared anon map is no longer in use. Before
2315 			 * freeing its pages purge all entries from
2316 			 * pcache that belong to this amp.
2317 			 */
2318 			if (svd->type == MAP_SHARED) {
2319 				ASSERT(amp->refcnt == 1);
2320 				ASSERT(svd->softlockcnt == 0);
2321 				anonmap_purge(amp);
2322 			}
2323 
2324 			if (amp->a_szc == seg->s_szc) {
2325 				if (seg->s_szc != 0) {
2326 					anon_free_pages(amp->ahp, an_idx, len,
2327 					    seg->s_szc);
2328 				} else {
2329 					anon_free(amp->ahp, an_idx,
2330 					    len);
2331 				}
2332 			} else {
2333 				ASSERT(svd->type == MAP_SHARED);
2334 				ASSERT(amp->a_szc > seg->s_szc);
2335 				anon_shmap_free_pages(amp, an_idx, len);
2336 			}
2337 
2338 			/*
2339 			 * Unreserve swap space for the
2340 			 * unmapped chunk of this segment in
2341 			 * case it's MAP_SHARED
2342 			 */
2343 			if (svd->type == MAP_SHARED) {
2344 				anon_unresv_zone(len,
2345 				    seg->s_as->a_proc->p_zone);
2346 				amp->swresv -= len;
2347 			}
2348 		}
2349 		nsvd->anon_index = svd->anon_index +
2350 		    btop((uintptr_t)(nseg->s_base - seg->s_base));
2351 		if (svd->type == MAP_SHARED) {
2352 			amp->refcnt++;
2353 			nsvd->amp = amp;
2354 		} else {
2355 			struct anon_map *namp;
2356 			struct anon_hdr *nahp;
2357 
2358 			ASSERT(svd->type == MAP_PRIVATE);
2359 			nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
2360 			namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
2361 			namp->a_szc = seg->s_szc;
2362 			(void) anon_copy_ptr(amp->ahp, svd->anon_index, nahp,
2363 			    0, btop(seg->s_size), ANON_SLEEP);
2364 			(void) anon_copy_ptr(amp->ahp, nsvd->anon_index,
2365 			    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
2366 			anon_release(amp->ahp, btop(amp->size));
2367 			svd->anon_index = 0;
2368 			nsvd->anon_index = 0;
2369 			amp->ahp = nahp;
2370 			amp->size = seg->s_size;
2371 			nsvd->amp = namp;
2372 		}
2373 		ANON_LOCK_EXIT(&amp->a_rwlock);
2374 	}
2375 	if (svd->swresv) {
2376 		if (svd->flags & MAP_NORESERVE) {
2377 			ASSERT(amp);
2378 			oswresv = svd->swresv;
2379 			svd->swresv = ptob(anon_pages(amp->ahp,
2380 			    svd->anon_index, btop(seg->s_size)));
2381 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
2382 			    nsvd->anon_index, btop(nseg->s_size)));
2383 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2384 			anon_unresv_zone(oswresv - (svd->swresv + nsvd->swresv),
2385 			    seg->s_as->a_proc->p_zone);
2386 			if (SEG_IS_PARTIAL_RESV(seg))
2387 				seg->s_as->a_resvsize -= oswresv -
2388 				    (svd->swresv + nsvd->swresv);
2389 		} else {
2390 			size_t unlen;
2391 
2392 			if (svd->pageswap) {
2393 				oswresv = svd->swresv;
2394 				svd->swresv = segvn_count_swap_by_vpages(seg);
2395 				nsvd->swresv = segvn_count_swap_by_vpages(nseg);
2396 				ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2397 				unlen = oswresv - (svd->swresv + nsvd->swresv);
2398 			} else {
2399 				if (seg->s_size + nseg->s_size + len !=
2400 				    svd->swresv) {
2401 					panic("segvn_unmap: cannot split "
2402 					    "swap reservation");
2403 					/*NOTREACHED*/
2404 				}
2405 				svd->swresv = seg->s_size;
2406 				nsvd->swresv = nseg->s_size;
2407 				unlen = len;
2408 			}
2409 			anon_unresv_zone(unlen,
2410 			    seg->s_as->a_proc->p_zone);
2411 		}
2412 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2413 		    seg, len, 0);
2414 	}
2415 
2416 	return (0);			/* I'm glad that's all over with! */
2417 }
2418 
2419 static void
2420 segvn_free(struct seg *seg)
2421 {
2422 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2423 	pgcnt_t npages = seg_pages(seg);
2424 	struct anon_map *amp;
2425 	size_t len;
2426 
2427 	/*
2428 	 * We don't need any segment level locks for "segvn" data
2429 	 * since the address space is "write" locked.
2430 	 */
2431 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
2432 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
2433 
2434 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2435 
2436 	/*
2437 	 * Be sure to unlock pages. XXX Why do things get free'ed instead
2438 	 * of unmapped? XXX
2439 	 */
2440 	(void) segvn_lockop(seg, seg->s_base, seg->s_size,
2441 	    0, MC_UNLOCK, NULL, 0);
2442 
2443 	/*
2444 	 * Deallocate the vpage and anon pointers if necessary and possible.
2445 	 */
2446 	if (svd->vpage != NULL) {
2447 		kmem_free(svd->vpage, vpgtob(npages));
2448 		svd->vpage = NULL;
2449 	}
2450 	if ((amp = svd->amp) != NULL) {
2451 		/*
2452 		 * If there are no more references to this anon_map
2453 		 * structure, then deallocate the structure after freeing
2454 		 * up all the anon slot pointers that we can.
2455 		 */
2456 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2457 		ASSERT(amp->a_szc >= seg->s_szc);
2458 		if (--amp->refcnt == 0) {
2459 			if (svd->type == MAP_PRIVATE) {
2460 				/*
2461 				 * Private - we only need to anon_free
2462 				 * the part that this segment refers to.
2463 				 */
2464 				if (seg->s_szc != 0) {
2465 					anon_free_pages(amp->ahp,
2466 					    svd->anon_index, seg->s_size,
2467 					    seg->s_szc);
2468 				} else {
2469 					anon_free(amp->ahp, svd->anon_index,
2470 					    seg->s_size);
2471 				}
2472 			} else {
2473 
2474 				/*
2475 				 * Shared anon map is no longer in use. Before
2476 				 * freeing its pages purge all entries from
2477 				 * pcache that belong to this amp.
2478 				 */
2479 				ASSERT(svd->softlockcnt == 0);
2480 				anonmap_purge(amp);
2481 
2482 				/*
2483 				 * Shared - anon_free the entire
2484 				 * anon_map's worth of stuff and
2485 				 * release any swap reservation.
2486 				 */
2487 				if (amp->a_szc != 0) {
2488 					anon_shmap_free_pages(amp, 0,
2489 					    amp->size);
2490 				} else {
2491 					anon_free(amp->ahp, 0, amp->size);
2492 				}
2493 				if ((len = amp->swresv) != 0) {
2494 					anon_unresv_zone(len,
2495 					    seg->s_as->a_proc->p_zone);
2496 					TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2497 					    "anon proc:%p %lu %u", seg, len, 0);
2498 				}
2499 			}
2500 			svd->amp = NULL;
2501 			ANON_LOCK_EXIT(&amp->a_rwlock);
2502 			anonmap_free(amp);
2503 		} else if (svd->type == MAP_PRIVATE) {
2504 			/*
2505 			 * We had a private mapping which still has
2506 			 * a held anon_map so just free up all the
2507 			 * anon slot pointers that we were using.
2508 			 */
2509 			if (seg->s_szc != 0) {
2510 				anon_free_pages(amp->ahp, svd->anon_index,
2511 				    seg->s_size, seg->s_szc);
2512 			} else {
2513 				anon_free(amp->ahp, svd->anon_index,
2514 				    seg->s_size);
2515 			}
2516 			ANON_LOCK_EXIT(&amp->a_rwlock);
2517 		} else {
2518 			ANON_LOCK_EXIT(&amp->a_rwlock);
2519 		}
2520 	}
2521 
2522 	/*
2523 	 * Release swap reservation.
2524 	 */
2525 	if ((len = svd->swresv) != 0) {
2526 		anon_unresv_zone(svd->swresv,
2527 		    seg->s_as->a_proc->p_zone);
2528 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2529 		    seg, len, 0);
2530 		if (SEG_IS_PARTIAL_RESV(seg))
2531 			seg->s_as->a_resvsize -= svd->swresv;
2532 		svd->swresv = 0;
2533 	}
2534 	/*
2535 	 * Release claim on vnode, credentials, and finally free the
2536 	 * private data.
2537 	 */
2538 	if (svd->vp != NULL) {
2539 		if (svd->type == MAP_SHARED)
2540 			lgrp_shm_policy_fini(NULL, svd->vp);
2541 		VN_RELE(svd->vp);
2542 		svd->vp = NULL;
2543 	}
2544 	crfree(svd->cred);
2545 	svd->pageprot = 0;
2546 	svd->pageadvice = 0;
2547 	svd->pageswap = 0;
2548 	svd->cred = NULL;
2549 
2550 	/*
2551 	 * Take segfree_syncmtx lock to let segvn_reclaim() finish if it's
2552 	 * still working with this segment without holding as lock (in case
2553 	 * it's called by pcache async thread).
2554 	 */
2555 	ASSERT(svd->softlockcnt == 0);
2556 	mutex_enter(&svd->segfree_syncmtx);
2557 	mutex_exit(&svd->segfree_syncmtx);
2558 
2559 	seg->s_data = NULL;
2560 	kmem_cache_free(segvn_cache, svd);
2561 }
2562 
2563 /*
2564  * Do a F_SOFTUNLOCK call over the range requested.  The range must have
2565  * already been F_SOFTLOCK'ed.
2566  * Caller must always match addr and len of a softunlock with a previous
2567  * softlock with exactly the same addr and len.
2568  */
2569 static void
2570 segvn_softunlock(struct seg *seg, caddr_t addr, size_t len, enum seg_rw rw)
2571 {
2572 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2573 	page_t *pp;
2574 	caddr_t adr;
2575 	struct vnode *vp;
2576 	u_offset_t offset;
2577 	ulong_t anon_index;
2578 	struct anon_map *amp;
2579 	struct anon *ap = NULL;
2580 
2581 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2582 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
2583 
2584 	if ((amp = svd->amp) != NULL)
2585 		anon_index = svd->anon_index + seg_page(seg, addr);
2586 
2587 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
2588 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2589 		hat_unlock_region(seg->s_as->a_hat, addr, len, svd->rcookie);
2590 	} else {
2591 		hat_unlock(seg->s_as->a_hat, addr, len);
2592 	}
2593 	for (adr = addr; adr < addr + len; adr += PAGESIZE) {
2594 		if (amp != NULL) {
2595 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2596 			if ((ap = anon_get_ptr(amp->ahp, anon_index++))
2597 			    != NULL) {
2598 				swap_xlate(ap, &vp, &offset);
2599 			} else {
2600 				vp = svd->vp;
2601 				offset = svd->offset +
2602 				    (uintptr_t)(adr - seg->s_base);
2603 			}
2604 			ANON_LOCK_EXIT(&amp->a_rwlock);
2605 		} else {
2606 			vp = svd->vp;
2607 			offset = svd->offset +
2608 			    (uintptr_t)(adr - seg->s_base);
2609 		}
2610 
2611 		/*
2612 		 * Use page_find() instead of page_lookup() to
2613 		 * find the page since we know that it is locked.
2614 		 */
2615 		pp = page_find(vp, offset);
2616 		if (pp == NULL) {
2617 			panic(
2618 			    "segvn_softunlock: addr %p, ap %p, vp %p, off %llx",
2619 			    (void *)adr, (void *)ap, (void *)vp, offset);
2620 			/*NOTREACHED*/
2621 		}
2622 
2623 		if (rw == S_WRITE) {
2624 			hat_setrefmod(pp);
2625 			if (seg->s_as->a_vbits)
2626 				hat_setstat(seg->s_as, adr, PAGESIZE,
2627 				    P_REF | P_MOD);
2628 		} else if (rw != S_OTHER) {
2629 			hat_setref(pp);
2630 			if (seg->s_as->a_vbits)
2631 				hat_setstat(seg->s_as, adr, PAGESIZE, P_REF);
2632 		}
2633 		TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2634 		    "segvn_fault:pp %p vp %p offset %llx", pp, vp, offset);
2635 		page_unlock(pp);
2636 	}
2637 	ASSERT(svd->softlockcnt >= btop(len));
2638 	if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -btop(len))) {
2639 		/*
2640 		 * All SOFTLOCKS are gone. Wakeup any waiting
2641 		 * unmappers so they can try again to unmap.
2642 		 * Check for waiters first without the mutex
2643 		 * held so we don't always grab the mutex on
2644 		 * softunlocks.
2645 		 */
2646 		if (AS_ISUNMAPWAIT(seg->s_as)) {
2647 			mutex_enter(&seg->s_as->a_contents);
2648 			if (AS_ISUNMAPWAIT(seg->s_as)) {
2649 				AS_CLRUNMAPWAIT(seg->s_as);
2650 				cv_broadcast(&seg->s_as->a_cv);
2651 			}
2652 			mutex_exit(&seg->s_as->a_contents);
2653 		}
2654 	}
2655 }
2656 
2657 #define	PAGE_HANDLED	((page_t *)-1)
2658 
2659 /*
2660  * Release all the pages in the NULL terminated ppp list
2661  * which haven't already been converted to PAGE_HANDLED.
2662  */
2663 static void
2664 segvn_pagelist_rele(page_t **ppp)
2665 {
2666 	for (; *ppp != NULL; ppp++) {
2667 		if (*ppp != PAGE_HANDLED)
2668 			page_unlock(*ppp);
2669 	}
2670 }
2671 
2672 static int stealcow = 1;
2673 
2674 /*
2675  * Workaround for viking chip bug.  See bug id 1220902.
2676  * To fix this down in pagefault() would require importing so
2677  * much as and segvn code as to be unmaintainable.
2678  */
2679 int enable_mbit_wa = 0;
2680 
2681 /*
2682  * Handles all the dirty work of getting the right
2683  * anonymous pages and loading up the translations.
2684  * This routine is called only from segvn_fault()
2685  * when looping over the range of addresses requested.
2686  *
2687  * The basic algorithm here is:
2688  * 	If this is an anon_zero case
2689  *		Call anon_zero to allocate page
2690  *		Load up translation
2691  *		Return
2692  *	endif
2693  *	If this is an anon page
2694  *		Use anon_getpage to get the page
2695  *	else
2696  *		Find page in pl[] list passed in
2697  *	endif
2698  *	If not a cow
2699  *		Load up the translation to the page
2700  *		return
2701  *	endif
2702  *	Call anon_private to handle cow
2703  *	Load up (writable) translation to new page
2704  */
2705 static faultcode_t
2706 segvn_faultpage(
2707 	struct hat *hat,		/* the hat to use for mapping */
2708 	struct seg *seg,		/* seg_vn of interest */
2709 	caddr_t addr,			/* address in as */
2710 	u_offset_t off,			/* offset in vp */
2711 	struct vpage *vpage,		/* pointer to vpage for vp, off */
2712 	page_t *pl[],			/* object source page pointer */
2713 	uint_t vpprot,			/* access allowed to object pages */
2714 	enum fault_type type,		/* type of fault */
2715 	enum seg_rw rw,			/* type of access at fault */
2716 	int brkcow)			/* we may need to break cow */
2717 {
2718 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2719 	page_t *pp, **ppp;
2720 	uint_t pageflags = 0;
2721 	page_t *anon_pl[1 + 1];
2722 	page_t *opp = NULL;		/* original page */
2723 	uint_t prot;
2724 	int err;
2725 	int cow;
2726 	int claim;
2727 	int steal = 0;
2728 	ulong_t anon_index;
2729 	struct anon *ap, *oldap;
2730 	struct anon_map *amp;
2731 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
2732 	int anon_lock = 0;
2733 	anon_sync_obj_t cookie;
2734 
2735 	if (svd->flags & MAP_TEXT) {
2736 		hat_flag |= HAT_LOAD_TEXT;
2737 	}
2738 
2739 	ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
2740 	ASSERT(seg->s_szc == 0);
2741 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
2742 
2743 	/*
2744 	 * Initialize protection value for this page.
2745 	 * If we have per page protection values check it now.
2746 	 */
2747 	if (svd->pageprot) {
2748 		uint_t protchk;
2749 
2750 		switch (rw) {
2751 		case S_READ:
2752 			protchk = PROT_READ;
2753 			break;
2754 		case S_WRITE:
2755 			protchk = PROT_WRITE;
2756 			break;
2757 		case S_EXEC:
2758 			protchk = PROT_EXEC;
2759 			break;
2760 		case S_OTHER:
2761 		default:
2762 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
2763 			break;
2764 		}
2765 
2766 		prot = VPP_PROT(vpage);
2767 		if ((prot & protchk) == 0)
2768 			return (FC_PROT);	/* illegal access type */
2769 	} else {
2770 		prot = svd->prot;
2771 	}
2772 
2773 	if (type == F_SOFTLOCK) {
2774 		atomic_inc_ulong((ulong_t *)&svd->softlockcnt);
2775 	}
2776 
2777 	/*
2778 	 * Always acquire the anon array lock to prevent 2 threads from
2779 	 * allocating separate anon slots for the same "addr".
2780 	 */
2781 
2782 	if ((amp = svd->amp) != NULL) {
2783 		ASSERT(RW_READ_HELD(&amp->a_rwlock));
2784 		anon_index = svd->anon_index + seg_page(seg, addr);
2785 		anon_array_enter(amp, anon_index, &cookie);
2786 		anon_lock = 1;
2787 	}
2788 
2789 	if (svd->vp == NULL && amp != NULL) {
2790 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL) {
2791 			/*
2792 			 * Allocate a (normally) writable anonymous page of
2793 			 * zeroes. If no advance reservations, reserve now.
2794 			 */
2795 			if (svd->flags & MAP_NORESERVE) {
2796 				if (anon_resv_zone(ptob(1),
2797 				    seg->s_as->a_proc->p_zone)) {
2798 					atomic_add_long(&svd->swresv, ptob(1));
2799 					atomic_add_long(&seg->s_as->a_resvsize,
2800 					    ptob(1));
2801 				} else {
2802 					err = ENOMEM;
2803 					goto out;
2804 				}
2805 			}
2806 			if ((pp = anon_zero(seg, addr, &ap,
2807 			    svd->cred)) == NULL) {
2808 				err = ENOMEM;
2809 				goto out;	/* out of swap space */
2810 			}
2811 			/*
2812 			 * Re-acquire the anon_map lock and
2813 			 * initialize the anon array entry.
2814 			 */
2815 			(void) anon_set_ptr(amp->ahp, anon_index, ap,
2816 			    ANON_SLEEP);
2817 
2818 			ASSERT(pp->p_szc == 0);
2819 
2820 			/*
2821 			 * Handle pages that have been marked for migration
2822 			 */
2823 			if (lgrp_optimizations())
2824 				page_migrate(seg, addr, &pp, 1);
2825 
2826 			if (enable_mbit_wa) {
2827 				if (rw == S_WRITE)
2828 					hat_setmod(pp);
2829 				else if (!hat_ismod(pp))
2830 					prot &= ~PROT_WRITE;
2831 			}
2832 			/*
2833 			 * If AS_PAGLCK is set in a_flags (via memcntl(2)
2834 			 * with MC_LOCKAS, MCL_FUTURE) and this is a
2835 			 * MAP_NORESERVE segment, we may need to
2836 			 * permanently lock the page as it is being faulted
2837 			 * for the first time. The following text applies
2838 			 * only to MAP_NORESERVE segments:
2839 			 *
2840 			 * As per memcntl(2), if this segment was created
2841 			 * after MCL_FUTURE was applied (a "future"
2842 			 * segment), its pages must be locked.  If this
2843 			 * segment existed at MCL_FUTURE application (a
2844 			 * "past" segment), the interface is unclear.
2845 			 *
2846 			 * We decide to lock only if vpage is present:
2847 			 *
2848 			 * - "future" segments will have a vpage array (see
2849 			 *    as_map), and so will be locked as required
2850 			 *
2851 			 * - "past" segments may not have a vpage array,
2852 			 *    depending on whether events (such as
2853 			 *    mprotect) have occurred. Locking if vpage
2854 			 *    exists will preserve legacy behavior.  Not
2855 			 *    locking if vpage is absent, will not break
2856 			 *    the interface or legacy behavior.  Note that
2857 			 *    allocating vpage here if it's absent requires
2858 			 *    upgrading the segvn reader lock, the cost of
2859 			 *    which does not seem worthwhile.
2860 			 *
2861 			 * Usually testing and setting VPP_ISPPLOCK and
2862 			 * VPP_SETPPLOCK requires holding the segvn lock as
2863 			 * writer, but in this case all readers are
2864 			 * serializing on the anon array lock.
2865 			 */
2866 			if (AS_ISPGLCK(seg->s_as) && vpage != NULL &&
2867 			    (svd->flags & MAP_NORESERVE) &&
2868 			    !VPP_ISPPLOCK(vpage)) {
2869 				proc_t *p = seg->s_as->a_proc;
2870 				ASSERT(svd->type == MAP_PRIVATE);
2871 				mutex_enter(&p->p_lock);
2872 				if (rctl_incr_locked_mem(p, NULL, PAGESIZE,
2873 				    1) == 0) {
2874 					claim = VPP_PROT(vpage) & PROT_WRITE;
2875 					if (page_pp_lock(pp, claim, 0)) {
2876 						VPP_SETPPLOCK(vpage);
2877 					} else {
2878 						rctl_decr_locked_mem(p, NULL,
2879 						    PAGESIZE, 1);
2880 					}
2881 				}
2882 				mutex_exit(&p->p_lock);
2883 			}
2884 
2885 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2886 			hat_memload(hat, addr, pp, prot, hat_flag);
2887 
2888 			if (!(hat_flag & HAT_LOAD_LOCK))
2889 				page_unlock(pp);
2890 
2891 			anon_array_exit(&cookie);
2892 			return (0);
2893 		}
2894 	}
2895 
2896 	/*
2897 	 * Obtain the page structure via anon_getpage() if it is
2898 	 * a private copy of an object (the result of a previous
2899 	 * copy-on-write).
2900 	 */
2901 	if (amp != NULL) {
2902 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) != NULL) {
2903 			err = anon_getpage(&ap, &vpprot, anon_pl, PAGESIZE,
2904 			    seg, addr, rw, svd->cred);
2905 			if (err)
2906 				goto out;
2907 
2908 			if (svd->type == MAP_SHARED) {
2909 				/*
2910 				 * If this is a shared mapping to an
2911 				 * anon_map, then ignore the write
2912 				 * permissions returned by anon_getpage().
2913 				 * They apply to the private mappings
2914 				 * of this anon_map.
2915 				 */
2916 				vpprot |= PROT_WRITE;
2917 			}
2918 			opp = anon_pl[0];
2919 		}
2920 	}
2921 
2922 	/*
2923 	 * Search the pl[] list passed in if it is from the
2924 	 * original object (i.e., not a private copy).
2925 	 */
2926 	if (opp == NULL) {
2927 		/*
2928 		 * Find original page.  We must be bringing it in
2929 		 * from the list in pl[].
2930 		 */
2931 		for (ppp = pl; (opp = *ppp) != NULL; ppp++) {
2932 			if (opp == PAGE_HANDLED)
2933 				continue;
2934 			ASSERT(opp->p_vnode == svd->vp); /* XXX */
2935 			if (opp->p_offset == off)
2936 				break;
2937 		}
2938 		if (opp == NULL) {
2939 			panic("segvn_faultpage not found");
2940 			/*NOTREACHED*/
2941 		}
2942 		*ppp = PAGE_HANDLED;
2943 
2944 	}
2945 
2946 	ASSERT(PAGE_LOCKED(opp));
2947 
2948 	TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2949 	    "segvn_fault:pp %p vp %p offset %llx", opp, NULL, 0);
2950 
2951 	/*
2952 	 * The fault is treated as a copy-on-write fault if a
2953 	 * write occurs on a private segment and the object
2954 	 * page (i.e., mapping) is write protected.  We assume
2955 	 * that fatal protection checks have already been made.
2956 	 */
2957 
2958 	if (brkcow) {
2959 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2960 		cow = !(vpprot & PROT_WRITE);
2961 	} else if (svd->tr_state == SEGVN_TR_ON) {
2962 		/*
2963 		 * If we are doing text replication COW on first touch.
2964 		 */
2965 		ASSERT(amp != NULL);
2966 		ASSERT(svd->vp != NULL);
2967 		ASSERT(rw != S_WRITE);
2968 		cow = (ap == NULL);
2969 	} else {
2970 		cow = 0;
2971 	}
2972 
2973 	/*
2974 	 * If not a copy-on-write case load the translation
2975 	 * and return.
2976 	 */
2977 	if (cow == 0) {
2978 
2979 		/*
2980 		 * Handle pages that have been marked for migration
2981 		 */
2982 		if (lgrp_optimizations())
2983 			page_migrate(seg, addr, &opp, 1);
2984 
2985 		if (IS_VMODSORT(opp->p_vnode) || enable_mbit_wa) {
2986 			if (rw == S_WRITE)
2987 				hat_setmod(opp);
2988 			else if (rw != S_OTHER && !hat_ismod(opp))
2989 				prot &= ~PROT_WRITE;
2990 		}
2991 
2992 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
2993 		    (!svd->pageprot && svd->prot == (prot & vpprot)));
2994 		ASSERT(amp == NULL ||
2995 		    svd->rcookie == HAT_INVALID_REGION_COOKIE);
2996 		hat_memload_region(hat, addr, opp, prot & vpprot, hat_flag,
2997 		    svd->rcookie);
2998 
2999 		if (!(hat_flag & HAT_LOAD_LOCK))
3000 			page_unlock(opp);
3001 
3002 		if (anon_lock) {
3003 			anon_array_exit(&cookie);
3004 		}
3005 		return (0);
3006 	}
3007 
3008 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3009 
3010 	hat_setref(opp);
3011 
3012 	ASSERT(amp != NULL && anon_lock);
3013 
3014 	/*
3015 	 * Steal the page only if it isn't a private page
3016 	 * since stealing a private page is not worth the effort.
3017 	 */
3018 	if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL)
3019 		steal = 1;
3020 
3021 	/*
3022 	 * Steal the original page if the following conditions are true:
3023 	 *
3024 	 * We are low on memory, the page is not private, page is not large,
3025 	 * not shared, not modified, not `locked' or if we have it `locked'
3026 	 * (i.e., p_cowcnt == 1 and p_lckcnt == 0, which also implies
3027 	 * that the page is not shared) and if it doesn't have any
3028 	 * translations. page_struct_lock isn't needed to look at p_cowcnt
3029 	 * and p_lckcnt because we first get exclusive lock on page.
3030 	 */
3031 	(void) hat_pagesync(opp, HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD);
3032 
3033 	if (stealcow && freemem < minfree && steal && opp->p_szc == 0 &&
3034 	    page_tryupgrade(opp) && !hat_ismod(opp) &&
3035 	    ((opp->p_lckcnt == 0 && opp->p_cowcnt == 0) ||
3036 	    (opp->p_lckcnt == 0 && opp->p_cowcnt == 1 &&
3037 	    vpage != NULL && VPP_ISPPLOCK(vpage)))) {
3038 		/*
3039 		 * Check if this page has other translations
3040 		 * after unloading our translation.
3041 		 */
3042 		if (hat_page_is_mapped(opp)) {
3043 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3044 			hat_unload(seg->s_as->a_hat, addr, PAGESIZE,
3045 			    HAT_UNLOAD);
3046 		}
3047 
3048 		/*
3049 		 * hat_unload() might sync back someone else's recent
3050 		 * modification, so check again.
3051 		 */
3052 		if (!hat_ismod(opp) && !hat_page_is_mapped(opp))
3053 			pageflags |= STEAL_PAGE;
3054 	}
3055 
3056 	/*
3057 	 * If we have a vpage pointer, see if it indicates that we have
3058 	 * ``locked'' the page we map -- if so, tell anon_private to
3059 	 * transfer the locking resource to the new page.
3060 	 *
3061 	 * See Statement at the beginning of segvn_lockop regarding
3062 	 * the way lockcnts/cowcnts are handled during COW.
3063 	 *
3064 	 */
3065 	if (vpage != NULL && VPP_ISPPLOCK(vpage))
3066 		pageflags |= LOCK_PAGE;
3067 
3068 	/*
3069 	 * Allocate a private page and perform the copy.
3070 	 * For MAP_NORESERVE reserve swap space now, unless this
3071 	 * is a cow fault on an existing anon page in which case
3072 	 * MAP_NORESERVE will have made advance reservations.
3073 	 */
3074 	if ((svd->flags & MAP_NORESERVE) && (ap == NULL)) {
3075 		if (anon_resv_zone(ptob(1), seg->s_as->a_proc->p_zone)) {
3076 			atomic_add_long(&svd->swresv, ptob(1));
3077 			atomic_add_long(&seg->s_as->a_resvsize, ptob(1));
3078 		} else {
3079 			page_unlock(opp);
3080 			err = ENOMEM;
3081 			goto out;
3082 		}
3083 	}
3084 	oldap = ap;
3085 	pp = anon_private(&ap, seg, addr, prot, opp, pageflags, svd->cred);
3086 	if (pp == NULL) {
3087 		err = ENOMEM;	/* out of swap space */
3088 		goto out;
3089 	}
3090 
3091 	/*
3092 	 * If we copied away from an anonymous page, then
3093 	 * we are one step closer to freeing up an anon slot.
3094 	 *
3095 	 * NOTE:  The original anon slot must be released while
3096 	 * holding the "anon_map" lock.  This is necessary to prevent
3097 	 * other threads from obtaining a pointer to the anon slot
3098 	 * which may be freed if its "refcnt" is 1.
3099 	 */
3100 	if (oldap != NULL)
3101 		anon_decref(oldap);
3102 
3103 	(void) anon_set_ptr(amp->ahp, anon_index, ap, ANON_SLEEP);
3104 
3105 	/*
3106 	 * Handle pages that have been marked for migration
3107 	 */
3108 	if (lgrp_optimizations())
3109 		page_migrate(seg, addr, &pp, 1);
3110 
3111 	ASSERT(pp->p_szc == 0);
3112 
3113 	ASSERT(!IS_VMODSORT(pp->p_vnode));
3114 	if (enable_mbit_wa) {
3115 		if (rw == S_WRITE)
3116 			hat_setmod(pp);
3117 		else if (!hat_ismod(pp))
3118 			prot &= ~PROT_WRITE;
3119 	}
3120 
3121 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3122 	hat_memload(hat, addr, pp, prot, hat_flag);
3123 
3124 	if (!(hat_flag & HAT_LOAD_LOCK))
3125 		page_unlock(pp);
3126 
3127 	ASSERT(anon_lock);
3128 	anon_array_exit(&cookie);
3129 	return (0);
3130 out:
3131 	if (anon_lock)
3132 		anon_array_exit(&cookie);
3133 
3134 	if (type == F_SOFTLOCK) {
3135 		atomic_dec_ulong((ulong_t *)&svd->softlockcnt);
3136 	}
3137 	return (FC_MAKE_ERR(err));
3138 }
3139 
3140 /*
3141  * relocate a bunch of smaller targ pages into one large repl page. all targ
3142  * pages must be complete pages smaller than replacement pages.
3143  * it's assumed that no page's szc can change since they are all PAGESIZE or
3144  * complete large pages locked SHARED.
3145  */
3146 static void
3147 segvn_relocate_pages(page_t **targ, page_t *replacement)
3148 {
3149 	page_t *pp;
3150 	pgcnt_t repl_npgs, curnpgs;
3151 	pgcnt_t i;
3152 	uint_t repl_szc = replacement->p_szc;
3153 	page_t *first_repl = replacement;
3154 	page_t *repl;
3155 	spgcnt_t npgs;
3156 
3157 	VM_STAT_ADD(segvnvmstats.relocatepages[0]);
3158 
3159 	ASSERT(repl_szc != 0);
3160 	npgs = repl_npgs = page_get_pagecnt(repl_szc);
3161 
3162 	i = 0;
3163 	while (repl_npgs) {
3164 		spgcnt_t nreloc;
3165 		int err;
3166 		ASSERT(replacement != NULL);
3167 		pp = targ[i];
3168 		ASSERT(pp->p_szc < repl_szc);
3169 		ASSERT(PAGE_EXCL(pp));
3170 		ASSERT(!PP_ISFREE(pp));
3171 		curnpgs = page_get_pagecnt(pp->p_szc);
3172 		if (curnpgs == 1) {
3173 			VM_STAT_ADD(segvnvmstats.relocatepages[1]);
3174 			repl = replacement;
3175 			page_sub(&replacement, repl);
3176 			ASSERT(PAGE_EXCL(repl));
3177 			ASSERT(!PP_ISFREE(repl));
3178 			ASSERT(repl->p_szc == repl_szc);
3179 		} else {
3180 			page_t *repl_savepp;
3181 			int j;
3182 			VM_STAT_ADD(segvnvmstats.relocatepages[2]);
3183 			repl_savepp = replacement;
3184 			for (j = 0; j < curnpgs; j++) {
3185 				repl = replacement;
3186 				page_sub(&replacement, repl);
3187 				ASSERT(PAGE_EXCL(repl));
3188 				ASSERT(!PP_ISFREE(repl));
3189 				ASSERT(repl->p_szc == repl_szc);
3190 				ASSERT(page_pptonum(targ[i + j]) ==
3191 				    page_pptonum(targ[i]) + j);
3192 			}
3193 			repl = repl_savepp;
3194 			ASSERT(IS_P2ALIGNED(page_pptonum(repl), curnpgs));
3195 		}
3196 		err = page_relocate(&pp, &repl, 0, 1, &nreloc, NULL);
3197 		if (err || nreloc != curnpgs) {
3198 			panic("segvn_relocate_pages: "
3199 			    "page_relocate failed err=%d curnpgs=%ld "
3200 			    "nreloc=%ld", err, curnpgs, nreloc);
3201 		}
3202 		ASSERT(curnpgs <= repl_npgs);
3203 		repl_npgs -= curnpgs;
3204 		i += curnpgs;
3205 	}
3206 	ASSERT(replacement == NULL);
3207 
3208 	repl = first_repl;
3209 	repl_npgs = npgs;
3210 	for (i = 0; i < repl_npgs; i++) {
3211 		ASSERT(PAGE_EXCL(repl));
3212 		ASSERT(!PP_ISFREE(repl));
3213 		targ[i] = repl;
3214 		page_downgrade(targ[i]);
3215 		repl++;
3216 	}
3217 }
3218 
3219 /*
3220  * Check if all pages in ppa array are complete smaller than szc pages and
3221  * their roots will still be aligned relative to their current size if the
3222  * entire ppa array is relocated into one szc page. If these conditions are
3223  * not met return 0.
3224  *
3225  * If all pages are properly aligned attempt to upgrade their locks
3226  * to exclusive mode. If it fails set *upgrdfail to 1 and return 0.
3227  * upgrdfail was set to 0 by caller.
3228  *
3229  * Return 1 if all pages are aligned and locked exclusively.
3230  *
3231  * If all pages in ppa array happen to be physically contiguous to make one
3232  * szc page and all exclusive locks are successfully obtained promote the page
3233  * size to szc and set *pszc to szc. Return 1 with pages locked shared.
3234  */
3235 static int
3236 segvn_full_szcpages(page_t **ppa, uint_t szc, int *upgrdfail, uint_t *pszc)
3237 {
3238 	page_t *pp;
3239 	pfn_t pfn;
3240 	pgcnt_t totnpgs = page_get_pagecnt(szc);
3241 	pfn_t first_pfn;
3242 	int contig = 1;
3243 	pgcnt_t i;
3244 	pgcnt_t j;
3245 	uint_t curszc;
3246 	pgcnt_t curnpgs;
3247 	int root = 0;
3248 
3249 	ASSERT(szc > 0);
3250 
3251 	VM_STAT_ADD(segvnvmstats.fullszcpages[0]);
3252 
3253 	for (i = 0; i < totnpgs; i++) {
3254 		pp = ppa[i];
3255 		ASSERT(PAGE_SHARED(pp));
3256 		ASSERT(!PP_ISFREE(pp));
3257 		pfn = page_pptonum(pp);
3258 		if (i == 0) {
3259 			if (!IS_P2ALIGNED(pfn, totnpgs)) {
3260 				contig = 0;
3261 			} else {
3262 				first_pfn = pfn;
3263 			}
3264 		} else if (contig && pfn != first_pfn + i) {
3265 			contig = 0;
3266 		}
3267 		if (pp->p_szc == 0) {
3268 			if (root) {
3269 				VM_STAT_ADD(segvnvmstats.fullszcpages[1]);
3270 				return (0);
3271 			}
3272 		} else if (!root) {
3273 			if ((curszc = pp->p_szc) >= szc) {
3274 				VM_STAT_ADD(segvnvmstats.fullszcpages[2]);
3275 				return (0);
3276 			}
3277 			if (curszc == 0) {
3278 				/*
3279 				 * p_szc changed means we don't have all pages
3280 				 * locked. return failure.
3281 				 */
3282 				VM_STAT_ADD(segvnvmstats.fullszcpages[3]);
3283 				return (0);
3284 			}
3285 			curnpgs = page_get_pagecnt(curszc);
3286 			if (!IS_P2ALIGNED(pfn, curnpgs) ||
3287 			    !IS_P2ALIGNED(i, curnpgs)) {
3288 				VM_STAT_ADD(segvnvmstats.fullszcpages[4]);
3289 				return (0);
3290 			}
3291 			root = 1;
3292 		} else {
3293 			ASSERT(i > 0);
3294 			VM_STAT_ADD(segvnvmstats.fullszcpages[5]);
3295 			if (pp->p_szc != curszc) {
3296 				VM_STAT_ADD(segvnvmstats.fullszcpages[6]);
3297 				return (0);
3298 			}
3299 			if (pfn - 1 != page_pptonum(ppa[i - 1])) {
3300 				panic("segvn_full_szcpages: "
3301 				    "large page not physically contiguous");
3302 			}
3303 			if (P2PHASE(pfn, curnpgs) == curnpgs - 1) {
3304 				root = 0;
3305 			}
3306 		}
3307 	}
3308 
3309 	for (i = 0; i < totnpgs; i++) {
3310 		ASSERT(ppa[i]->p_szc < szc);
3311 		if (!page_tryupgrade(ppa[i])) {
3312 			for (j = 0; j < i; j++) {
3313 				page_downgrade(ppa[j]);
3314 			}
3315 			*pszc = ppa[i]->p_szc;
3316 			*upgrdfail = 1;
3317 			VM_STAT_ADD(segvnvmstats.fullszcpages[7]);
3318 			return (0);
3319 		}
3320 	}
3321 
3322 	/*
3323 	 * When a page is put a free cachelist its szc is set to 0.  if file
3324 	 * system reclaimed pages from cachelist targ pages will be physically
3325 	 * contiguous with 0 p_szc.  in this case just upgrade szc of targ
3326 	 * pages without any relocations.
3327 	 * To avoid any hat issues with previous small mappings
3328 	 * hat_pageunload() the target pages first.
3329 	 */
3330 	if (contig) {
3331 		VM_STAT_ADD(segvnvmstats.fullszcpages[8]);
3332 		for (i = 0; i < totnpgs; i++) {
3333 			(void) hat_pageunload(ppa[i], HAT_FORCE_PGUNLOAD);
3334 		}
3335 		for (i = 0; i < totnpgs; i++) {
3336 			ppa[i]->p_szc = szc;
3337 		}
3338 		for (i = 0; i < totnpgs; i++) {
3339 			ASSERT(PAGE_EXCL(ppa[i]));
3340 			page_downgrade(ppa[i]);
3341 		}
3342 		if (pszc != NULL) {
3343 			*pszc = szc;
3344 		}
3345 	}
3346 	VM_STAT_ADD(segvnvmstats.fullszcpages[9]);
3347 	return (1);
3348 }
3349 
3350 /*
3351  * Create physically contiguous pages for [vp, off] - [vp, off +
3352  * page_size(szc)) range and for private segment return them in ppa array.
3353  * Pages are created either via IO or relocations.
3354  *
3355  * Return 1 on success and 0 on failure.
3356  *
3357  * If physically contiguous pages already exist for this range return 1 without
3358  * filling ppa array. Caller initializes ppa[0] as NULL to detect that ppa
3359  * array wasn't filled. In this case caller fills ppa array via VOP_GETPAGE().
3360  */
3361 
3362 static int
3363 segvn_fill_vp_pages(struct segvn_data *svd, vnode_t *vp, u_offset_t off,
3364     uint_t szc, page_t **ppa, page_t **ppplist, uint_t *ret_pszc,
3365     int *downsize)
3366 {
3367 	page_t *pplist = *ppplist;
3368 	size_t pgsz = page_get_pagesize(szc);
3369 	pgcnt_t pages = btop(pgsz);
3370 	ulong_t start_off = off;
3371 	u_offset_t eoff = off + pgsz;
3372 	spgcnt_t nreloc;
3373 	u_offset_t io_off = off;
3374 	size_t io_len;
3375 	page_t *io_pplist = NULL;
3376 	page_t *done_pplist = NULL;
3377 	pgcnt_t pgidx = 0;
3378 	page_t *pp;
3379 	page_t *newpp;
3380 	page_t *targpp;
3381 	int io_err = 0;
3382 	int i;
3383 	pfn_t pfn;
3384 	ulong_t ppages;
3385 	page_t *targ_pplist = NULL;
3386 	page_t *repl_pplist = NULL;
3387 	page_t *tmp_pplist;
3388 	int nios = 0;
3389 	uint_t pszc;
3390 	struct vattr va;
3391 
3392 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[0]);
3393 
3394 	ASSERT(szc != 0);
3395 	ASSERT(pplist->p_szc == szc);
3396 
3397 	/*
3398 	 * downsize will be set to 1 only if we fail to lock pages. this will
3399 	 * allow subsequent faults to try to relocate the page again. If we
3400 	 * fail due to misalignment don't downsize and let the caller map the
3401 	 * whole region with small mappings to avoid more faults into the area
3402 	 * where we can't get large pages anyway.
3403 	 */
3404 	*downsize = 0;
3405 
3406 	while (off < eoff) {
3407 		newpp = pplist;
3408 		ASSERT(newpp != NULL);
3409 		ASSERT(PAGE_EXCL(newpp));
3410 		ASSERT(!PP_ISFREE(newpp));
3411 		/*
3412 		 * we pass NULL for nrelocp to page_lookup_create()
3413 		 * so that it doesn't relocate. We relocate here
3414 		 * later only after we make sure we can lock all
3415 		 * pages in the range we handle and they are all
3416 		 * aligned.
3417 		 */
3418 		pp = page_lookup_create(vp, off, SE_SHARED, newpp, NULL, 0);
3419 		ASSERT(pp != NULL);
3420 		ASSERT(!PP_ISFREE(pp));
3421 		ASSERT(pp->p_vnode == vp);
3422 		ASSERT(pp->p_offset == off);
3423 		if (pp == newpp) {
3424 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[1]);
3425 			page_sub(&pplist, pp);
3426 			ASSERT(PAGE_EXCL(pp));
3427 			ASSERT(page_iolock_assert(pp));
3428 			page_list_concat(&io_pplist, &pp);
3429 			off += PAGESIZE;
3430 			continue;
3431 		}
3432 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[2]);
3433 		pfn = page_pptonum(pp);
3434 		pszc = pp->p_szc;
3435 		if (pszc >= szc && targ_pplist == NULL && io_pplist == NULL &&
3436 		    IS_P2ALIGNED(pfn, pages)) {
3437 			ASSERT(repl_pplist == NULL);
3438 			ASSERT(done_pplist == NULL);
3439 			ASSERT(pplist == *ppplist);
3440 			page_unlock(pp);
3441 			page_free_replacement_page(pplist);
3442 			page_create_putback(pages);
3443 			*ppplist = NULL;
3444 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[3]);
3445 			return (1);
3446 		}
3447 		if (pszc >= szc) {
3448 			page_unlock(pp);
3449 			segvn_faultvnmpss_align_err1++;
3450 			goto out;
3451 		}
3452 		ppages = page_get_pagecnt(pszc);
3453 		if (!IS_P2ALIGNED(pfn, ppages)) {
3454 			ASSERT(pszc > 0);
3455 			/*
3456 			 * sizing down to pszc won't help.
3457 			 */
3458 			page_unlock(pp);
3459 			segvn_faultvnmpss_align_err2++;
3460 			goto out;
3461 		}
3462 		pfn = page_pptonum(newpp);
3463 		if (!IS_P2ALIGNED(pfn, ppages)) {
3464 			ASSERT(pszc > 0);
3465 			/*
3466 			 * sizing down to pszc won't help.
3467 			 */
3468 			page_unlock(pp);
3469 			segvn_faultvnmpss_align_err3++;
3470 			goto out;
3471 		}
3472 		if (!PAGE_EXCL(pp)) {
3473 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[4]);
3474 			page_unlock(pp);
3475 			*downsize = 1;
3476 			*ret_pszc = pp->p_szc;
3477 			goto out;
3478 		}
3479 		targpp = pp;
3480 		if (io_pplist != NULL) {
3481 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[5]);
3482 			io_len = off - io_off;
3483 			/*
3484 			 * Some file systems like NFS don't check EOF
3485 			 * conditions in VOP_PAGEIO(). Check it here
3486 			 * now that pages are locked SE_EXCL. Any file
3487 			 * truncation will wait until the pages are
3488 			 * unlocked so no need to worry that file will
3489 			 * be truncated after we check its size here.
3490 			 * XXX fix NFS to remove this check.
3491 			 */
3492 			va.va_mask = AT_SIZE;
3493 			if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL)) {
3494 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[6]);
3495 				page_unlock(targpp);
3496 				goto out;
3497 			}
3498 			if (btopr(va.va_size) < btopr(io_off + io_len)) {
3499 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[7]);
3500 				*downsize = 1;
3501 				*ret_pszc = 0;
3502 				page_unlock(targpp);
3503 				goto out;
3504 			}
3505 			io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3506 			    B_READ, svd->cred, NULL);
3507 			if (io_err) {
3508 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[8]);
3509 				page_unlock(targpp);
3510 				if (io_err == EDEADLK) {
3511 					segvn_vmpss_pageio_deadlk_err++;
3512 				}
3513 				goto out;
3514 			}
3515 			nios++;
3516 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[9]);
3517 			while (io_pplist != NULL) {
3518 				pp = io_pplist;
3519 				page_sub(&io_pplist, pp);
3520 				ASSERT(page_iolock_assert(pp));
3521 				page_io_unlock(pp);
3522 				pgidx = (pp->p_offset - start_off) >>
3523 				    PAGESHIFT;
3524 				ASSERT(pgidx < pages);
3525 				ppa[pgidx] = pp;
3526 				page_list_concat(&done_pplist, &pp);
3527 			}
3528 		}
3529 		pp = targpp;
3530 		ASSERT(PAGE_EXCL(pp));
3531 		ASSERT(pp->p_szc <= pszc);
3532 		if (pszc != 0 && !group_page_trylock(pp, SE_EXCL)) {
3533 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[10]);
3534 			page_unlock(pp);
3535 			*downsize = 1;
3536 			*ret_pszc = pp->p_szc;
3537 			goto out;
3538 		}
3539 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[11]);
3540 		/*
3541 		 * page szc chould have changed before the entire group was
3542 		 * locked. reread page szc.
3543 		 */
3544 		pszc = pp->p_szc;
3545 		ppages = page_get_pagecnt(pszc);
3546 
3547 		/* link just the roots */
3548 		page_list_concat(&targ_pplist, &pp);
3549 		page_sub(&pplist, newpp);
3550 		page_list_concat(&repl_pplist, &newpp);
3551 		off += PAGESIZE;
3552 		while (--ppages != 0) {
3553 			newpp = pplist;
3554 			page_sub(&pplist, newpp);
3555 			off += PAGESIZE;
3556 		}
3557 		io_off = off;
3558 	}
3559 	if (io_pplist != NULL) {
3560 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[12]);
3561 		io_len = eoff - io_off;
3562 		va.va_mask = AT_SIZE;
3563 		if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL) != 0) {
3564 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[13]);
3565 			goto out;
3566 		}
3567 		if (btopr(va.va_size) < btopr(io_off + io_len)) {
3568 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[14]);
3569 			*downsize = 1;
3570 			*ret_pszc = 0;
3571 			goto out;
3572 		}
3573 		io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3574 		    B_READ, svd->cred, NULL);
3575 		if (io_err) {
3576 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[15]);
3577 			if (io_err == EDEADLK) {
3578 				segvn_vmpss_pageio_deadlk_err++;
3579 			}
3580 			goto out;
3581 		}
3582 		nios++;
3583 		while (io_pplist != NULL) {
3584 			pp = io_pplist;
3585 			page_sub(&io_pplist, pp);
3586 			ASSERT(page_iolock_assert(pp));
3587 			page_io_unlock(pp);
3588 			pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3589 			ASSERT(pgidx < pages);
3590 			ppa[pgidx] = pp;
3591 		}
3592 	}
3593 	/*
3594 	 * we're now bound to succeed or panic.
3595 	 * remove pages from done_pplist. it's not needed anymore.
3596 	 */
3597 	while (done_pplist != NULL) {
3598 		pp = done_pplist;
3599 		page_sub(&done_pplist, pp);
3600 	}
3601 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[16]);
3602 	ASSERT(pplist == NULL);
3603 	*ppplist = NULL;
3604 	while (targ_pplist != NULL) {
3605 		int ret;
3606 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[17]);
3607 		ASSERT(repl_pplist);
3608 		pp = targ_pplist;
3609 		page_sub(&targ_pplist, pp);
3610 		pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3611 		newpp = repl_pplist;
3612 		page_sub(&repl_pplist, newpp);
3613 #ifdef DEBUG
3614 		pfn = page_pptonum(pp);
3615 		pszc = pp->p_szc;
3616 		ppages = page_get_pagecnt(pszc);
3617 		ASSERT(IS_P2ALIGNED(pfn, ppages));
3618 		pfn = page_pptonum(newpp);
3619 		ASSERT(IS_P2ALIGNED(pfn, ppages));
3620 		ASSERT(P2PHASE(pfn, pages) == pgidx);
3621 #endif
3622 		nreloc = 0;
3623 		ret = page_relocate(&pp, &newpp, 0, 1, &nreloc, NULL);
3624 		if (ret != 0 || nreloc == 0) {
3625 			panic("segvn_fill_vp_pages: "
3626 			    "page_relocate failed");
3627 		}
3628 		pp = newpp;
3629 		while (nreloc-- != 0) {
3630 			ASSERT(PAGE_EXCL(pp));
3631 			ASSERT(pp->p_vnode == vp);
3632 			ASSERT(pgidx ==
3633 			    ((pp->p_offset - start_off) >> PAGESHIFT));
3634 			ppa[pgidx++] = pp;
3635 			pp++;
3636 		}
3637 	}
3638 
3639 	if (svd->type == MAP_PRIVATE) {
3640 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[18]);
3641 		for (i = 0; i < pages; i++) {
3642 			ASSERT(ppa[i] != NULL);
3643 			ASSERT(PAGE_EXCL(ppa[i]));
3644 			ASSERT(ppa[i]->p_vnode == vp);
3645 			ASSERT(ppa[i]->p_offset ==
3646 			    start_off + (i << PAGESHIFT));
3647 			page_downgrade(ppa[i]);
3648 		}
3649 		ppa[pages] = NULL;
3650 	} else {
3651 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[19]);
3652 		/*
3653 		 * the caller will still call VOP_GETPAGE() for shared segments
3654 		 * to check FS write permissions. For private segments we map
3655 		 * file read only anyway.  so no VOP_GETPAGE is needed.
3656 		 */
3657 		for (i = 0; i < pages; i++) {
3658 			ASSERT(ppa[i] != NULL);
3659 			ASSERT(PAGE_EXCL(ppa[i]));
3660 			ASSERT(ppa[i]->p_vnode == vp);
3661 			ASSERT(ppa[i]->p_offset ==
3662 			    start_off + (i << PAGESHIFT));
3663 			page_unlock(ppa[i]);
3664 		}
3665 		ppa[0] = NULL;
3666 	}
3667 
3668 	return (1);
3669 out:
3670 	/*
3671 	 * Do the cleanup. Unlock target pages we didn't relocate. They are
3672 	 * linked on targ_pplist by root pages. reassemble unused replacement
3673 	 * and io pages back to pplist.
3674 	 */
3675 	if (io_pplist != NULL) {
3676 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[20]);
3677 		pp = io_pplist;
3678 		do {
3679 			ASSERT(pp->p_vnode == vp);
3680 			ASSERT(pp->p_offset == io_off);
3681 			ASSERT(page_iolock_assert(pp));
3682 			page_io_unlock(pp);
3683 			page_hashout(pp, NULL);
3684 			io_off += PAGESIZE;
3685 		} while ((pp = pp->p_next) != io_pplist);
3686 		page_list_concat(&io_pplist, &pplist);
3687 		pplist = io_pplist;
3688 	}
3689 	tmp_pplist = NULL;
3690 	while (targ_pplist != NULL) {
3691 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[21]);
3692 		pp = targ_pplist;
3693 		ASSERT(PAGE_EXCL(pp));
3694 		page_sub(&targ_pplist, pp);
3695 
3696 		pszc = pp->p_szc;
3697 		ppages = page_get_pagecnt(pszc);
3698 		ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3699 
3700 		if (pszc != 0) {
3701 			group_page_unlock(pp);
3702 		}
3703 		page_unlock(pp);
3704 
3705 		pp = repl_pplist;
3706 		ASSERT(pp != NULL);
3707 		ASSERT(PAGE_EXCL(pp));
3708 		ASSERT(pp->p_szc == szc);
3709 		page_sub(&repl_pplist, pp);
3710 
3711 		ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3712 
3713 		/* relink replacement page */
3714 		page_list_concat(&tmp_pplist, &pp);
3715 		while (--ppages != 0) {
3716 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[22]);
3717 			pp++;
3718 			ASSERT(PAGE_EXCL(pp));
3719 			ASSERT(pp->p_szc == szc);
3720 			page_list_concat(&tmp_pplist, &pp);
3721 		}
3722 	}
3723 	if (tmp_pplist != NULL) {
3724 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[23]);
3725 		page_list_concat(&tmp_pplist, &pplist);
3726 		pplist = tmp_pplist;
3727 	}
3728 	/*
3729 	 * at this point all pages are either on done_pplist or
3730 	 * pplist. They can't be all on done_pplist otherwise
3731 	 * we'd've been done.
3732 	 */
3733 	ASSERT(pplist != NULL);
3734 	if (nios != 0) {
3735 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[24]);
3736 		pp = pplist;
3737 		do {
3738 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[25]);
3739 			ASSERT(pp->p_szc == szc);
3740 			ASSERT(PAGE_EXCL(pp));
3741 			ASSERT(pp->p_vnode != vp);
3742 			pp->p_szc = 0;
3743 		} while ((pp = pp->p_next) != pplist);
3744 
3745 		pp = done_pplist;
3746 		do {
3747 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[26]);
3748 			ASSERT(pp->p_szc == szc);
3749 			ASSERT(PAGE_EXCL(pp));
3750 			ASSERT(pp->p_vnode == vp);
3751 			pp->p_szc = 0;
3752 		} while ((pp = pp->p_next) != done_pplist);
3753 
3754 		while (pplist != NULL) {
3755 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[27]);
3756 			pp = pplist;
3757 			page_sub(&pplist, pp);
3758 			page_free(pp, 0);
3759 		}
3760 
3761 		while (done_pplist != NULL) {
3762 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[28]);
3763 			pp = done_pplist;
3764 			page_sub(&done_pplist, pp);
3765 			page_unlock(pp);
3766 		}
3767 		*ppplist = NULL;
3768 		return (0);
3769 	}
3770 	ASSERT(pplist == *ppplist);
3771 	if (io_err) {
3772 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[29]);
3773 		/*
3774 		 * don't downsize on io error.
3775 		 * see if vop_getpage succeeds.
3776 		 * pplist may still be used in this case
3777 		 * for relocations.
3778 		 */
3779 		return (0);
3780 	}
3781 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[30]);
3782 	page_free_replacement_page(pplist);
3783 	page_create_putback(pages);
3784 	*ppplist = NULL;
3785 	return (0);
3786 }
3787 
3788 int segvn_anypgsz = 0;
3789 
3790 #define	SEGVN_RESTORE_SOFTLOCK_VP(type, pages) 				\
3791 		if ((type) == F_SOFTLOCK) {				\
3792 			atomic_add_long((ulong_t *)&(svd)->softlockcnt, \
3793 			    -(pages));					\
3794 		}
3795 
3796 #define	SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot)		\
3797 		if (IS_VMODSORT((ppa)[0]->p_vnode)) {			\
3798 			if ((rw) == S_WRITE) {				\
3799 				for (i = 0; i < (pages); i++) {		\
3800 					ASSERT((ppa)[i]->p_vnode ==	\
3801 					    (ppa)[0]->p_vnode);		\
3802 					hat_setmod((ppa)[i]);		\
3803 				}					\
3804 			} else if ((rw) != S_OTHER &&			\
3805 			    ((prot) & (vpprot) & PROT_WRITE)) {		\
3806 				for (i = 0; i < (pages); i++) {		\
3807 					ASSERT((ppa)[i]->p_vnode ==	\
3808 					    (ppa)[0]->p_vnode);		\
3809 					if (!hat_ismod((ppa)[i])) {	\
3810 						prot &= ~PROT_WRITE;	\
3811 						break;			\
3812 					}				\
3813 				}					\
3814 			}						\
3815 		}
3816 
3817 #ifdef  VM_STATS
3818 
3819 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)					\
3820 		VM_STAT_ADD(segvnvmstats.fltvnpages[(idx)]);
3821 
3822 #else /* VM_STATS */
3823 
3824 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)
3825 
3826 #endif
3827 
3828 static faultcode_t
3829 segvn_fault_vnodepages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
3830     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
3831     caddr_t eaddr, int brkcow)
3832 {
3833 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
3834 	struct anon_map *amp = svd->amp;
3835 	uchar_t segtype = svd->type;
3836 	uint_t szc = seg->s_szc;
3837 	size_t pgsz = page_get_pagesize(szc);
3838 	size_t maxpgsz = pgsz;
3839 	pgcnt_t pages = btop(pgsz);
3840 	pgcnt_t maxpages = pages;
3841 	size_t ppasize = (pages + 1) * sizeof (page_t *);
3842 	caddr_t a = lpgaddr;
3843 	caddr_t	maxlpgeaddr = lpgeaddr;
3844 	u_offset_t off = svd->offset + (uintptr_t)(a - seg->s_base);
3845 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
3846 	struct vpage *vpage = (svd->vpage != NULL) ?
3847 	    &svd->vpage[seg_page(seg, a)] : NULL;
3848 	vnode_t *vp = svd->vp;
3849 	page_t **ppa;
3850 	uint_t	pszc;
3851 	size_t	ppgsz;
3852 	pgcnt_t	ppages;
3853 	faultcode_t err = 0;
3854 	int ierr;
3855 	int vop_size_err = 0;
3856 	uint_t protchk, prot, vpprot;
3857 	ulong_t i;
3858 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
3859 	anon_sync_obj_t an_cookie;
3860 	enum seg_rw arw;
3861 	int alloc_failed = 0;
3862 	int adjszc_chk;
3863 	struct vattr va;
3864 	page_t *pplist;
3865 	pfn_t pfn;
3866 	int physcontig;
3867 	int upgrdfail;
3868 	int segvn_anypgsz_vnode = 0; /* for now map vnode with 2 page sizes */
3869 	int tron = (svd->tr_state == SEGVN_TR_ON);
3870 
3871 	ASSERT(szc != 0);
3872 	ASSERT(vp != NULL);
3873 	ASSERT(brkcow == 0 || amp != NULL);
3874 	ASSERT(tron == 0 || amp != NULL);
3875 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
3876 	ASSERT(!(svd->flags & MAP_NORESERVE));
3877 	ASSERT(type != F_SOFTUNLOCK);
3878 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
3879 	ASSERT(amp == NULL || IS_P2ALIGNED(aindx, maxpages));
3880 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
3881 	ASSERT(seg->s_szc < NBBY * sizeof (int));
3882 	ASSERT(type != F_SOFTLOCK || lpgeaddr - a == maxpgsz);
3883 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
3884 
3885 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltvnpages[0]);
3886 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltvnpages[1]);
3887 
3888 	if (svd->flags & MAP_TEXT) {
3889 		hat_flag |= HAT_LOAD_TEXT;
3890 	}
3891 
3892 	if (svd->pageprot) {
3893 		switch (rw) {
3894 		case S_READ:
3895 			protchk = PROT_READ;
3896 			break;
3897 		case S_WRITE:
3898 			protchk = PROT_WRITE;
3899 			break;
3900 		case S_EXEC:
3901 			protchk = PROT_EXEC;
3902 			break;
3903 		case S_OTHER:
3904 		default:
3905 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
3906 			break;
3907 		}
3908 	} else {
3909 		prot = svd->prot;
3910 		/* caller has already done segment level protection check. */
3911 	}
3912 
3913 	if (rw == S_WRITE && segtype == MAP_PRIVATE) {
3914 		SEGVN_VMSTAT_FLTVNPAGES(2);
3915 		arw = S_READ;
3916 	} else {
3917 		arw = rw;
3918 	}
3919 
3920 	ppa = kmem_alloc(ppasize, KM_SLEEP);
3921 
3922 	VM_STAT_COND_ADD(amp != NULL, segvnvmstats.fltvnpages[3]);
3923 
3924 	for (;;) {
3925 		adjszc_chk = 0;
3926 		for (; a < lpgeaddr; a += pgsz, off += pgsz, aindx += pages) {
3927 			if (adjszc_chk) {
3928 				while (szc < seg->s_szc) {
3929 					uintptr_t e;
3930 					uint_t tszc;
3931 					tszc = segvn_anypgsz_vnode ? szc + 1 :
3932 					    seg->s_szc;
3933 					ppgsz = page_get_pagesize(tszc);
3934 					if (!IS_P2ALIGNED(a, ppgsz) ||
3935 					    ((alloc_failed >> tszc) & 0x1)) {
3936 						break;
3937 					}
3938 					SEGVN_VMSTAT_FLTVNPAGES(4);
3939 					szc = tszc;
3940 					pgsz = ppgsz;
3941 					pages = btop(pgsz);
3942 					e = P2ROUNDUP((uintptr_t)eaddr, pgsz);
3943 					lpgeaddr = (caddr_t)e;
3944 				}
3945 			}
3946 
3947 		again:
3948 			if (IS_P2ALIGNED(a, maxpgsz) && amp != NULL) {
3949 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
3950 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
3951 				anon_array_enter(amp, aindx, &an_cookie);
3952 				if (anon_get_ptr(amp->ahp, aindx) != NULL) {
3953 					SEGVN_VMSTAT_FLTVNPAGES(5);
3954 					ASSERT(anon_pages(amp->ahp, aindx,
3955 					    maxpages) == maxpages);
3956 					anon_array_exit(&an_cookie);
3957 					ANON_LOCK_EXIT(&amp->a_rwlock);
3958 					err = segvn_fault_anonpages(hat, seg,
3959 					    a, a + maxpgsz, type, rw,
3960 					    MAX(a, addr),
3961 					    MIN(a + maxpgsz, eaddr), brkcow);
3962 					if (err != 0) {
3963 						SEGVN_VMSTAT_FLTVNPAGES(6);
3964 						goto out;
3965 					}
3966 					if (szc < seg->s_szc) {
3967 						szc = seg->s_szc;
3968 						pgsz = maxpgsz;
3969 						pages = maxpages;
3970 						lpgeaddr = maxlpgeaddr;
3971 					}
3972 					goto next;
3973 				} else {
3974 					ASSERT(anon_pages(amp->ahp, aindx,
3975 					    maxpages) == 0);
3976 					SEGVN_VMSTAT_FLTVNPAGES(7);
3977 					anon_array_exit(&an_cookie);
3978 					ANON_LOCK_EXIT(&amp->a_rwlock);
3979 				}
3980 			}
3981 			ASSERT(!brkcow || IS_P2ALIGNED(a, maxpgsz));
3982 			ASSERT(!tron || IS_P2ALIGNED(a, maxpgsz));
3983 
3984 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
3985 				ASSERT(vpage != NULL);
3986 				prot = VPP_PROT(vpage);
3987 				ASSERT(sameprot(seg, a, maxpgsz));
3988 				if ((prot & protchk) == 0) {
3989 					SEGVN_VMSTAT_FLTVNPAGES(8);
3990 					err = FC_PROT;
3991 					goto out;
3992 				}
3993 			}
3994 			if (type == F_SOFTLOCK) {
3995 				atomic_add_long((ulong_t *)&svd->softlockcnt,
3996 				    pages);
3997 			}
3998 
3999 			pplist = NULL;
4000 			physcontig = 0;
4001 			ppa[0] = NULL;
4002 			if (!brkcow && !tron && szc &&
4003 			    !page_exists_physcontig(vp, off, szc,
4004 			    segtype == MAP_PRIVATE ? ppa : NULL)) {
4005 				SEGVN_VMSTAT_FLTVNPAGES(9);
4006 				if (page_alloc_pages(vp, seg, a, &pplist, NULL,
4007 				    szc, 0, 0) && type != F_SOFTLOCK) {
4008 					SEGVN_VMSTAT_FLTVNPAGES(10);
4009 					pszc = 0;
4010 					ierr = -1;
4011 					alloc_failed |= (1 << szc);
4012 					break;
4013 				}
4014 				if (pplist != NULL &&
4015 				    vp->v_mpssdata == SEGVN_PAGEIO) {
4016 					int downsize;
4017 					SEGVN_VMSTAT_FLTVNPAGES(11);
4018 					physcontig = segvn_fill_vp_pages(svd,
4019 					    vp, off, szc, ppa, &pplist,
4020 					    &pszc, &downsize);
4021 					ASSERT(!physcontig || pplist == NULL);
4022 					if (!physcontig && downsize &&
4023 					    type != F_SOFTLOCK) {
4024 						ASSERT(pplist == NULL);
4025 						SEGVN_VMSTAT_FLTVNPAGES(12);
4026 						ierr = -1;
4027 						break;
4028 					}
4029 					ASSERT(!physcontig ||
4030 					    segtype == MAP_PRIVATE ||
4031 					    ppa[0] == NULL);
4032 					if (physcontig && ppa[0] == NULL) {
4033 						physcontig = 0;
4034 					}
4035 				}
4036 			} else if (!brkcow && !tron && szc && ppa[0] != NULL) {
4037 				SEGVN_VMSTAT_FLTVNPAGES(13);
4038 				ASSERT(segtype == MAP_PRIVATE);
4039 				physcontig = 1;
4040 			}
4041 
4042 			if (!physcontig) {
4043 				SEGVN_VMSTAT_FLTVNPAGES(14);
4044 				ppa[0] = NULL;
4045 				ierr = VOP_GETPAGE(vp, (offset_t)off, pgsz,
4046 				    &vpprot, ppa, pgsz, seg, a, arw,
4047 				    svd->cred, NULL);
4048 #ifdef DEBUG
4049 				if (ierr == 0) {
4050 					for (i = 0; i < pages; i++) {
4051 						ASSERT(PAGE_LOCKED(ppa[i]));
4052 						ASSERT(!PP_ISFREE(ppa[i]));
4053 						ASSERT(ppa[i]->p_vnode == vp);
4054 						ASSERT(ppa[i]->p_offset ==
4055 						    off + (i << PAGESHIFT));
4056 					}
4057 				}
4058 #endif /* DEBUG */
4059 				if (segtype == MAP_PRIVATE) {
4060 					SEGVN_VMSTAT_FLTVNPAGES(15);
4061 					vpprot &= ~PROT_WRITE;
4062 				}
4063 			} else {
4064 				ASSERT(segtype == MAP_PRIVATE);
4065 				SEGVN_VMSTAT_FLTVNPAGES(16);
4066 				vpprot = PROT_ALL & ~PROT_WRITE;
4067 				ierr = 0;
4068 			}
4069 
4070 			if (ierr != 0) {
4071 				SEGVN_VMSTAT_FLTVNPAGES(17);
4072 				if (pplist != NULL) {
4073 					SEGVN_VMSTAT_FLTVNPAGES(18);
4074 					page_free_replacement_page(pplist);
4075 					page_create_putback(pages);
4076 				}
4077 				SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4078 				if (a + pgsz <= eaddr) {
4079 					SEGVN_VMSTAT_FLTVNPAGES(19);
4080 					err = FC_MAKE_ERR(ierr);
4081 					goto out;
4082 				}
4083 				va.va_mask = AT_SIZE;
4084 				if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL)) {
4085 					SEGVN_VMSTAT_FLTVNPAGES(20);
4086 					err = FC_MAKE_ERR(EIO);
4087 					goto out;
4088 				}
4089 				if (btopr(va.va_size) >= btopr(off + pgsz)) {
4090 					SEGVN_VMSTAT_FLTVNPAGES(21);
4091 					err = FC_MAKE_ERR(ierr);
4092 					goto out;
4093 				}
4094 				if (btopr(va.va_size) <
4095 				    btopr(off + (eaddr - a))) {
4096 					SEGVN_VMSTAT_FLTVNPAGES(22);
4097 					err = FC_MAKE_ERR(ierr);
4098 					goto out;
4099 				}
4100 				if (brkcow || tron || type == F_SOFTLOCK) {
4101 					/* can't reduce map area */
4102 					SEGVN_VMSTAT_FLTVNPAGES(23);
4103 					vop_size_err = 1;
4104 					goto out;
4105 				}
4106 				SEGVN_VMSTAT_FLTVNPAGES(24);
4107 				ASSERT(szc != 0);
4108 				pszc = 0;
4109 				ierr = -1;
4110 				break;
4111 			}
4112 
4113 			if (amp != NULL) {
4114 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4115 				anon_array_enter(amp, aindx, &an_cookie);
4116 			}
4117 			if (amp != NULL &&
4118 			    anon_get_ptr(amp->ahp, aindx) != NULL) {
4119 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4120 
4121 				SEGVN_VMSTAT_FLTVNPAGES(25);
4122 				ASSERT(anon_pages(amp->ahp, taindx,
4123 				    maxpages) == maxpages);
4124 				for (i = 0; i < pages; i++) {
4125 					page_unlock(ppa[i]);
4126 				}
4127 				anon_array_exit(&an_cookie);
4128 				ANON_LOCK_EXIT(&amp->a_rwlock);
4129 				if (pplist != NULL) {
4130 					page_free_replacement_page(pplist);
4131 					page_create_putback(pages);
4132 				}
4133 				SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4134 				if (szc < seg->s_szc) {
4135 					SEGVN_VMSTAT_FLTVNPAGES(26);
4136 					/*
4137 					 * For private segments SOFTLOCK
4138 					 * either always breaks cow (any rw
4139 					 * type except S_READ_NOCOW) or
4140 					 * address space is locked as writer
4141 					 * (S_READ_NOCOW case) and anon slots
4142 					 * can't show up on second check.
4143 					 * Therefore if we are here for
4144 					 * SOFTLOCK case it must be a cow
4145 					 * break but cow break never reduces
4146 					 * szc. text replication (tron) in
4147 					 * this case works as cow break.
4148 					 * Thus the assert below.
4149 					 */
4150 					ASSERT(!brkcow && !tron &&
4151 					    type != F_SOFTLOCK);
4152 					pszc = seg->s_szc;
4153 					ierr = -2;
4154 					break;
4155 				}
4156 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4157 				goto again;
4158 			}
4159 #ifdef DEBUG
4160 			if (amp != NULL) {
4161 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4162 				ASSERT(!anon_pages(amp->ahp, taindx, maxpages));
4163 			}
4164 #endif /* DEBUG */
4165 
4166 			if (brkcow || tron) {
4167 				ASSERT(amp != NULL);
4168 				ASSERT(pplist == NULL);
4169 				ASSERT(szc == seg->s_szc);
4170 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4171 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
4172 				SEGVN_VMSTAT_FLTVNPAGES(27);
4173 				ierr = anon_map_privatepages(amp, aindx, szc,
4174 				    seg, a, prot, ppa, vpage, segvn_anypgsz,
4175 				    tron ? PG_LOCAL : 0, svd->cred);
4176 				if (ierr != 0) {
4177 					SEGVN_VMSTAT_FLTVNPAGES(28);
4178 					anon_array_exit(&an_cookie);
4179 					ANON_LOCK_EXIT(&amp->a_rwlock);
4180 					SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4181 					err = FC_MAKE_ERR(ierr);
4182 					goto out;
4183 				}
4184 
4185 				ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4186 				/*
4187 				 * p_szc can't be changed for locked
4188 				 * swapfs pages.
4189 				 */
4190 				ASSERT(svd->rcookie ==
4191 				    HAT_INVALID_REGION_COOKIE);
4192 				hat_memload_array(hat, a, pgsz, ppa, prot,
4193 				    hat_flag);
4194 
4195 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4196 					SEGVN_VMSTAT_FLTVNPAGES(29);
4197 					for (i = 0; i < pages; i++) {
4198 						page_unlock(ppa[i]);
4199 					}
4200 				}
4201 				anon_array_exit(&an_cookie);
4202 				ANON_LOCK_EXIT(&amp->a_rwlock);
4203 				goto next;
4204 			}
4205 
4206 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
4207 			    (!svd->pageprot && svd->prot == (prot & vpprot)));
4208 
4209 			pfn = page_pptonum(ppa[0]);
4210 			/*
4211 			 * hat_page_demote() needs an SE_EXCL lock on one of
4212 			 * constituent page_t's and it decreases root's p_szc
4213 			 * last. This means if root's p_szc is equal szc and
4214 			 * all its constituent pages are locked
4215 			 * hat_page_demote() that could have changed p_szc to
4216 			 * szc is already done and no new have page_demote()
4217 			 * can start for this large page.
4218 			 */
4219 
4220 			/*
4221 			 * we need to make sure same mapping size is used for
4222 			 * the same address range if there's a possibility the
4223 			 * adddress is already mapped because hat layer panics
4224 			 * when translation is loaded for the range already
4225 			 * mapped with a different page size.  We achieve it
4226 			 * by always using largest page size possible subject
4227 			 * to the constraints of page size, segment page size
4228 			 * and page alignment.  Since mappings are invalidated
4229 			 * when those constraints change and make it
4230 			 * impossible to use previously used mapping size no
4231 			 * mapping size conflicts should happen.
4232 			 */
4233 
4234 		chkszc:
4235 			if ((pszc = ppa[0]->p_szc) == szc &&
4236 			    IS_P2ALIGNED(pfn, pages)) {
4237 
4238 				SEGVN_VMSTAT_FLTVNPAGES(30);
4239 #ifdef DEBUG
4240 				for (i = 0; i < pages; i++) {
4241 					ASSERT(PAGE_LOCKED(ppa[i]));
4242 					ASSERT(!PP_ISFREE(ppa[i]));
4243 					ASSERT(page_pptonum(ppa[i]) ==
4244 					    pfn + i);
4245 					ASSERT(ppa[i]->p_szc == szc);
4246 					ASSERT(ppa[i]->p_vnode == vp);
4247 					ASSERT(ppa[i]->p_offset ==
4248 					    off + (i << PAGESHIFT));
4249 				}
4250 #endif /* DEBUG */
4251 				/*
4252 				 * All pages are of szc we need and they are
4253 				 * all locked so they can't change szc. load
4254 				 * translations.
4255 				 *
4256 				 * if page got promoted since last check
4257 				 * we don't need pplist.
4258 				 */
4259 				if (pplist != NULL) {
4260 					page_free_replacement_page(pplist);
4261 					page_create_putback(pages);
4262 				}
4263 				if (PP_ISMIGRATE(ppa[0])) {
4264 					page_migrate(seg, a, ppa, pages);
4265 				}
4266 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4267 				    prot, vpprot);
4268 				hat_memload_array_region(hat, a, pgsz,
4269 				    ppa, prot & vpprot, hat_flag,
4270 				    svd->rcookie);
4271 
4272 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4273 					for (i = 0; i < pages; i++) {
4274 						page_unlock(ppa[i]);
4275 					}
4276 				}
4277 				if (amp != NULL) {
4278 					anon_array_exit(&an_cookie);
4279 					ANON_LOCK_EXIT(&amp->a_rwlock);
4280 				}
4281 				goto next;
4282 			}
4283 
4284 			/*
4285 			 * See if upsize is possible.
4286 			 */
4287 			if (pszc > szc && szc < seg->s_szc &&
4288 			    (segvn_anypgsz_vnode || pszc >= seg->s_szc)) {
4289 				pgcnt_t aphase;
4290 				uint_t pszc1 = MIN(pszc, seg->s_szc);
4291 				ppgsz = page_get_pagesize(pszc1);
4292 				ppages = btop(ppgsz);
4293 				aphase = btop(P2PHASE((uintptr_t)a, ppgsz));
4294 
4295 				ASSERT(type != F_SOFTLOCK);
4296 
4297 				SEGVN_VMSTAT_FLTVNPAGES(31);
4298 				if (aphase != P2PHASE(pfn, ppages)) {
4299 					segvn_faultvnmpss_align_err4++;
4300 				} else {
4301 					SEGVN_VMSTAT_FLTVNPAGES(32);
4302 					if (pplist != NULL) {
4303 						page_t *pl = pplist;
4304 						page_free_replacement_page(pl);
4305 						page_create_putback(pages);
4306 					}
4307 					for (i = 0; i < pages; i++) {
4308 						page_unlock(ppa[i]);
4309 					}
4310 					if (amp != NULL) {
4311 						anon_array_exit(&an_cookie);
4312 						ANON_LOCK_EXIT(&amp->a_rwlock);
4313 					}
4314 					pszc = pszc1;
4315 					ierr = -2;
4316 					break;
4317 				}
4318 			}
4319 
4320 			/*
4321 			 * check if we should use smallest mapping size.
4322 			 */
4323 			upgrdfail = 0;
4324 			if (szc == 0 ||
4325 			    (pszc >= szc &&
4326 			    !IS_P2ALIGNED(pfn, pages)) ||
4327 			    (pszc < szc &&
4328 			    !segvn_full_szcpages(ppa, szc, &upgrdfail,
4329 			    &pszc))) {
4330 
4331 				if (upgrdfail && type != F_SOFTLOCK) {
4332 					/*
4333 					 * segvn_full_szcpages failed to lock
4334 					 * all pages EXCL. Size down.
4335 					 */
4336 					ASSERT(pszc < szc);
4337 
4338 					SEGVN_VMSTAT_FLTVNPAGES(33);
4339 
4340 					if (pplist != NULL) {
4341 						page_t *pl = pplist;
4342 						page_free_replacement_page(pl);
4343 						page_create_putback(pages);
4344 					}
4345 
4346 					for (i = 0; i < pages; i++) {
4347 						page_unlock(ppa[i]);
4348 					}
4349 					if (amp != NULL) {
4350 						anon_array_exit(&an_cookie);
4351 						ANON_LOCK_EXIT(&amp->a_rwlock);
4352 					}
4353 					ierr = -1;
4354 					break;
4355 				}
4356 				if (szc != 0 && !upgrdfail) {
4357 					segvn_faultvnmpss_align_err5++;
4358 				}
4359 				SEGVN_VMSTAT_FLTVNPAGES(34);
4360 				if (pplist != NULL) {
4361 					page_free_replacement_page(pplist);
4362 					page_create_putback(pages);
4363 				}
4364 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4365 				    prot, vpprot);
4366 				if (upgrdfail && segvn_anypgsz_vnode) {
4367 					/* SOFTLOCK case */
4368 					hat_memload_array_region(hat, a, pgsz,
4369 					    ppa, prot & vpprot, hat_flag,
4370 					    svd->rcookie);
4371 				} else {
4372 					for (i = 0; i < pages; i++) {
4373 						hat_memload_region(hat,
4374 						    a + (i << PAGESHIFT),
4375 						    ppa[i], prot & vpprot,
4376 						    hat_flag, svd->rcookie);
4377 					}
4378 				}
4379 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4380 					for (i = 0; i < pages; i++) {
4381 						page_unlock(ppa[i]);
4382 					}
4383 				}
4384 				if (amp != NULL) {
4385 					anon_array_exit(&an_cookie);
4386 					ANON_LOCK_EXIT(&amp->a_rwlock);
4387 				}
4388 				goto next;
4389 			}
4390 
4391 			if (pszc == szc) {
4392 				/*
4393 				 * segvn_full_szcpages() upgraded pages szc.
4394 				 */
4395 				ASSERT(pszc == ppa[0]->p_szc);
4396 				ASSERT(IS_P2ALIGNED(pfn, pages));
4397 				goto chkszc;
4398 			}
4399 
4400 			if (pszc > szc) {
4401 				kmutex_t *szcmtx;
4402 				SEGVN_VMSTAT_FLTVNPAGES(35);
4403 				/*
4404 				 * p_szc of ppa[0] can change since we haven't
4405 				 * locked all constituent pages. Call
4406 				 * page_lock_szc() to prevent szc changes.
4407 				 * This should be a rare case that happens when
4408 				 * multiple segments use a different page size
4409 				 * to map the same file offsets.
4410 				 */
4411 				szcmtx = page_szc_lock(ppa[0]);
4412 				pszc = ppa[0]->p_szc;
4413 				ASSERT(szcmtx != NULL || pszc == 0);
4414 				ASSERT(ppa[0]->p_szc <= pszc);
4415 				if (pszc <= szc) {
4416 					SEGVN_VMSTAT_FLTVNPAGES(36);
4417 					if (szcmtx != NULL) {
4418 						mutex_exit(szcmtx);
4419 					}
4420 					goto chkszc;
4421 				}
4422 				if (pplist != NULL) {
4423 					/*
4424 					 * page got promoted since last check.
4425 					 * we don't need preaalocated large
4426 					 * page.
4427 					 */
4428 					SEGVN_VMSTAT_FLTVNPAGES(37);
4429 					page_free_replacement_page(pplist);
4430 					page_create_putback(pages);
4431 				}
4432 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4433 				    prot, vpprot);
4434 				hat_memload_array_region(hat, a, pgsz, ppa,
4435 				    prot & vpprot, hat_flag, svd->rcookie);
4436 				mutex_exit(szcmtx);
4437 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4438 					for (i = 0; i < pages; i++) {
4439 						page_unlock(ppa[i]);
4440 					}
4441 				}
4442 				if (amp != NULL) {
4443 					anon_array_exit(&an_cookie);
4444 					ANON_LOCK_EXIT(&amp->a_rwlock);
4445 				}
4446 				goto next;
4447 			}
4448 
4449 			/*
4450 			 * if page got demoted since last check
4451 			 * we could have not allocated larger page.
4452 			 * allocate now.
4453 			 */
4454 			if (pplist == NULL &&
4455 			    page_alloc_pages(vp, seg, a, &pplist, NULL,
4456 			    szc, 0, 0) && type != F_SOFTLOCK) {
4457 				SEGVN_VMSTAT_FLTVNPAGES(38);
4458 				for (i = 0; i < pages; i++) {
4459 					page_unlock(ppa[i]);
4460 				}
4461 				if (amp != NULL) {
4462 					anon_array_exit(&an_cookie);
4463 					ANON_LOCK_EXIT(&amp->a_rwlock);
4464 				}
4465 				ierr = -1;
4466 				alloc_failed |= (1 << szc);
4467 				break;
4468 			}
4469 
4470 			SEGVN_VMSTAT_FLTVNPAGES(39);
4471 
4472 			if (pplist != NULL) {
4473 				segvn_relocate_pages(ppa, pplist);
4474 #ifdef DEBUG
4475 			} else {
4476 				ASSERT(type == F_SOFTLOCK);
4477 				SEGVN_VMSTAT_FLTVNPAGES(40);
4478 #endif /* DEBUG */
4479 			}
4480 
4481 			SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot);
4482 
4483 			if (pplist == NULL && segvn_anypgsz_vnode == 0) {
4484 				ASSERT(type == F_SOFTLOCK);
4485 				for (i = 0; i < pages; i++) {
4486 					ASSERT(ppa[i]->p_szc < szc);
4487 					hat_memload_region(hat,
4488 					    a + (i << PAGESHIFT),
4489 					    ppa[i], prot & vpprot, hat_flag,
4490 					    svd->rcookie);
4491 				}
4492 			} else {
4493 				ASSERT(pplist != NULL || type == F_SOFTLOCK);
4494 				hat_memload_array_region(hat, a, pgsz, ppa,
4495 				    prot & vpprot, hat_flag, svd->rcookie);
4496 			}
4497 			if (!(hat_flag & HAT_LOAD_LOCK)) {
4498 				for (i = 0; i < pages; i++) {
4499 					ASSERT(PAGE_SHARED(ppa[i]));
4500 					page_unlock(ppa[i]);
4501 				}
4502 			}
4503 			if (amp != NULL) {
4504 				anon_array_exit(&an_cookie);
4505 				ANON_LOCK_EXIT(&amp->a_rwlock);
4506 			}
4507 
4508 		next:
4509 			if (vpage != NULL) {
4510 				vpage += pages;
4511 			}
4512 			adjszc_chk = 1;
4513 		}
4514 		if (a == lpgeaddr)
4515 			break;
4516 		ASSERT(a < lpgeaddr);
4517 
4518 		ASSERT(!brkcow && !tron && type != F_SOFTLOCK);
4519 
4520 		/*
4521 		 * ierr == -1 means we failed to map with a large page.
4522 		 * (either due to allocation/relocation failures or
4523 		 * misalignment with other mappings to this file.
4524 		 *
4525 		 * ierr == -2 means some other thread allocated a large page
4526 		 * after we gave up tp map with a large page.  retry with
4527 		 * larger mapping.
4528 		 */
4529 		ASSERT(ierr == -1 || ierr == -2);
4530 		ASSERT(ierr == -2 || szc != 0);
4531 		ASSERT(ierr == -1 || szc < seg->s_szc);
4532 		if (ierr == -2) {
4533 			SEGVN_VMSTAT_FLTVNPAGES(41);
4534 			ASSERT(pszc > szc && pszc <= seg->s_szc);
4535 			szc = pszc;
4536 		} else if (segvn_anypgsz_vnode) {
4537 			SEGVN_VMSTAT_FLTVNPAGES(42);
4538 			szc--;
4539 		} else {
4540 			SEGVN_VMSTAT_FLTVNPAGES(43);
4541 			ASSERT(pszc < szc);
4542 			/*
4543 			 * other process created pszc large page.
4544 			 * but we still have to drop to 0 szc.
4545 			 */
4546 			szc = 0;
4547 		}
4548 
4549 		pgsz = page_get_pagesize(szc);
4550 		pages = btop(pgsz);
4551 		if (ierr == -2) {
4552 			/*
4553 			 * Size up case. Note lpgaddr may only be needed for
4554 			 * softlock case so we don't adjust it here.
4555 			 */
4556 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4557 			ASSERT(a >= lpgaddr);
4558 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4559 			off = svd->offset + (uintptr_t)(a - seg->s_base);
4560 			aindx = svd->anon_index + seg_page(seg, a);
4561 			vpage = (svd->vpage != NULL) ?
4562 			    &svd->vpage[seg_page(seg, a)] : NULL;
4563 		} else {
4564 			/*
4565 			 * Size down case. Note lpgaddr may only be needed for
4566 			 * softlock case so we don't adjust it here.
4567 			 */
4568 			ASSERT(IS_P2ALIGNED(a, pgsz));
4569 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4570 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4571 			ASSERT(a < lpgeaddr);
4572 			if (a < addr) {
4573 				SEGVN_VMSTAT_FLTVNPAGES(44);
4574 				/*
4575 				 * The beginning of the large page region can
4576 				 * be pulled to the right to make a smaller
4577 				 * region. We haven't yet faulted a single
4578 				 * page.
4579 				 */
4580 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4581 				ASSERT(a >= lpgaddr);
4582 				off = svd->offset +
4583 				    (uintptr_t)(a - seg->s_base);
4584 				aindx = svd->anon_index + seg_page(seg, a);
4585 				vpage = (svd->vpage != NULL) ?
4586 				    &svd->vpage[seg_page(seg, a)] : NULL;
4587 			}
4588 		}
4589 	}
4590 out:
4591 	kmem_free(ppa, ppasize);
4592 	if (!err && !vop_size_err) {
4593 		SEGVN_VMSTAT_FLTVNPAGES(45);
4594 		return (0);
4595 	}
4596 	if (type == F_SOFTLOCK && a > lpgaddr) {
4597 		SEGVN_VMSTAT_FLTVNPAGES(46);
4598 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4599 	}
4600 	if (!vop_size_err) {
4601 		SEGVN_VMSTAT_FLTVNPAGES(47);
4602 		return (err);
4603 	}
4604 	ASSERT(brkcow || tron || type == F_SOFTLOCK);
4605 	/*
4606 	 * Large page end is mapped beyond the end of file and it's a cow
4607 	 * fault (can be a text replication induced cow) or softlock so we can't
4608 	 * reduce the map area.  For now just demote the segment. This should
4609 	 * really only happen if the end of the file changed after the mapping
4610 	 * was established since when large page segments are created we make
4611 	 * sure they don't extend beyond the end of the file.
4612 	 */
4613 	SEGVN_VMSTAT_FLTVNPAGES(48);
4614 
4615 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4616 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4617 	err = 0;
4618 	if (seg->s_szc != 0) {
4619 		segvn_fltvnpages_clrszc_cnt++;
4620 		ASSERT(svd->softlockcnt == 0);
4621 		err = segvn_clrszc(seg);
4622 		if (err != 0) {
4623 			segvn_fltvnpages_clrszc_err++;
4624 		}
4625 	}
4626 	ASSERT(err || seg->s_szc == 0);
4627 	SEGVN_LOCK_DOWNGRADE(seg->s_as, &svd->lock);
4628 	/* segvn_fault will do its job as if szc had been zero to begin with */
4629 	return (err == 0 ? IE_RETRY : FC_MAKE_ERR(err));
4630 }
4631 
4632 /*
4633  * This routine will attempt to fault in one large page.
4634  * it will use smaller pages if that fails.
4635  * It should only be called for pure anonymous segments.
4636  */
4637 static faultcode_t
4638 segvn_fault_anonpages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
4639     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
4640     caddr_t eaddr, int brkcow)
4641 {
4642 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4643 	struct anon_map *amp = svd->amp;
4644 	uchar_t segtype = svd->type;
4645 	uint_t szc = seg->s_szc;
4646 	size_t pgsz = page_get_pagesize(szc);
4647 	size_t maxpgsz = pgsz;
4648 	pgcnt_t pages = btop(pgsz);
4649 	uint_t ppaszc = szc;
4650 	caddr_t a = lpgaddr;
4651 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
4652 	struct vpage *vpage = (svd->vpage != NULL) ?
4653 	    &svd->vpage[seg_page(seg, a)] : NULL;
4654 	page_t **ppa;
4655 	uint_t	ppa_szc;
4656 	faultcode_t err;
4657 	int ierr;
4658 	uint_t protchk, prot, vpprot;
4659 	ulong_t i;
4660 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
4661 	anon_sync_obj_t cookie;
4662 	int adjszc_chk;
4663 	int pgflags = (svd->tr_state == SEGVN_TR_ON) ? PG_LOCAL : 0;
4664 
4665 	ASSERT(szc != 0);
4666 	ASSERT(amp != NULL);
4667 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
4668 	ASSERT(!(svd->flags & MAP_NORESERVE));
4669 	ASSERT(type != F_SOFTUNLOCK);
4670 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
4671 	ASSERT(!brkcow || svd->tr_state == SEGVN_TR_OFF);
4672 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
4673 
4674 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
4675 
4676 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltanpages[0]);
4677 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltanpages[1]);
4678 
4679 	if (svd->flags & MAP_TEXT) {
4680 		hat_flag |= HAT_LOAD_TEXT;
4681 	}
4682 
4683 	if (svd->pageprot) {
4684 		switch (rw) {
4685 		case S_READ:
4686 			protchk = PROT_READ;
4687 			break;
4688 		case S_WRITE:
4689 			protchk = PROT_WRITE;
4690 			break;
4691 		case S_EXEC:
4692 			protchk = PROT_EXEC;
4693 			break;
4694 		case S_OTHER:
4695 		default:
4696 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4697 			break;
4698 		}
4699 		VM_STAT_ADD(segvnvmstats.fltanpages[2]);
4700 	} else {
4701 		prot = svd->prot;
4702 		/* caller has already done segment level protection check. */
4703 	}
4704 
4705 	ppa = kmem_cache_alloc(segvn_szc_cache[ppaszc], KM_SLEEP);
4706 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4707 	for (;;) {
4708 		adjszc_chk = 0;
4709 		for (; a < lpgeaddr; a += pgsz, aindx += pages) {
4710 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
4711 				VM_STAT_ADD(segvnvmstats.fltanpages[3]);
4712 				ASSERT(vpage != NULL);
4713 				prot = VPP_PROT(vpage);
4714 				ASSERT(sameprot(seg, a, maxpgsz));
4715 				if ((prot & protchk) == 0) {
4716 					err = FC_PROT;
4717 					goto error;
4718 				}
4719 			}
4720 			if (adjszc_chk && IS_P2ALIGNED(a, maxpgsz) &&
4721 			    pgsz < maxpgsz) {
4722 				ASSERT(a > lpgaddr);
4723 				szc = seg->s_szc;
4724 				pgsz = maxpgsz;
4725 				pages = btop(pgsz);
4726 				ASSERT(IS_P2ALIGNED(aindx, pages));
4727 				lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr,
4728 				    pgsz);
4729 			}
4730 			if (type == F_SOFTLOCK) {
4731 				atomic_add_long((ulong_t *)&svd->softlockcnt,
4732 				    pages);
4733 			}
4734 			anon_array_enter(amp, aindx, &cookie);
4735 			ppa_szc = (uint_t)-1;
4736 			ierr = anon_map_getpages(amp, aindx, szc, seg, a,
4737 			    prot, &vpprot, ppa, &ppa_szc, vpage, rw, brkcow,
4738 			    segvn_anypgsz, pgflags, svd->cred);
4739 			if (ierr != 0) {
4740 				anon_array_exit(&cookie);
4741 				VM_STAT_ADD(segvnvmstats.fltanpages[4]);
4742 				if (type == F_SOFTLOCK) {
4743 					atomic_add_long(
4744 					    (ulong_t *)&svd->softlockcnt,
4745 					    -pages);
4746 				}
4747 				if (ierr > 0) {
4748 					VM_STAT_ADD(segvnvmstats.fltanpages[6]);
4749 					err = FC_MAKE_ERR(ierr);
4750 					goto error;
4751 				}
4752 				break;
4753 			}
4754 
4755 			ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4756 
4757 			ASSERT(segtype == MAP_SHARED ||
4758 			    ppa[0]->p_szc <= szc);
4759 			ASSERT(segtype == MAP_PRIVATE ||
4760 			    ppa[0]->p_szc >= szc);
4761 
4762 			/*
4763 			 * Handle pages that have been marked for migration
4764 			 */
4765 			if (lgrp_optimizations())
4766 				page_migrate(seg, a, ppa, pages);
4767 
4768 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
4769 
4770 			if (segtype == MAP_SHARED) {
4771 				vpprot |= PROT_WRITE;
4772 			}
4773 
4774 			hat_memload_array(hat, a, pgsz, ppa,
4775 			    prot & vpprot, hat_flag);
4776 
4777 			if (hat_flag & HAT_LOAD_LOCK) {
4778 				VM_STAT_ADD(segvnvmstats.fltanpages[7]);
4779 			} else {
4780 				VM_STAT_ADD(segvnvmstats.fltanpages[8]);
4781 				for (i = 0; i < pages; i++)
4782 					page_unlock(ppa[i]);
4783 			}
4784 			if (vpage != NULL)
4785 				vpage += pages;
4786 
4787 			anon_array_exit(&cookie);
4788 			adjszc_chk = 1;
4789 		}
4790 		if (a == lpgeaddr)
4791 			break;
4792 		ASSERT(a < lpgeaddr);
4793 		/*
4794 		 * ierr == -1 means we failed to allocate a large page.
4795 		 * so do a size down operation.
4796 		 *
4797 		 * ierr == -2 means some other process that privately shares
4798 		 * pages with this process has allocated a larger page and we
4799 		 * need to retry with larger pages. So do a size up
4800 		 * operation. This relies on the fact that large pages are
4801 		 * never partially shared i.e. if we share any constituent
4802 		 * page of a large page with another process we must share the
4803 		 * entire large page. Note this cannot happen for SOFTLOCK
4804 		 * case, unless current address (a) is at the beginning of the
4805 		 * next page size boundary because the other process couldn't
4806 		 * have relocated locked pages.
4807 		 */
4808 		ASSERT(ierr == -1 || ierr == -2);
4809 
4810 		if (segvn_anypgsz) {
4811 			ASSERT(ierr == -2 || szc != 0);
4812 			ASSERT(ierr == -1 || szc < seg->s_szc);
4813 			szc = (ierr == -1) ? szc - 1 : szc + 1;
4814 		} else {
4815 			/*
4816 			 * For non COW faults and segvn_anypgsz == 0
4817 			 * we need to be careful not to loop forever
4818 			 * if existing page is found with szc other
4819 			 * than 0 or seg->s_szc. This could be due
4820 			 * to page relocations on behalf of DR or
4821 			 * more likely large page creation. For this
4822 			 * case simply re-size to existing page's szc
4823 			 * if returned by anon_map_getpages().
4824 			 */
4825 			if (ppa_szc == (uint_t)-1) {
4826 				szc = (ierr == -1) ? 0 : seg->s_szc;
4827 			} else {
4828 				ASSERT(ppa_szc <= seg->s_szc);
4829 				ASSERT(ierr == -2 || ppa_szc < szc);
4830 				ASSERT(ierr == -1 || ppa_szc > szc);
4831 				szc = ppa_szc;
4832 			}
4833 		}
4834 
4835 		pgsz = page_get_pagesize(szc);
4836 		pages = btop(pgsz);
4837 		ASSERT(type != F_SOFTLOCK || ierr == -1 ||
4838 		    (IS_P2ALIGNED(a, pgsz) && IS_P2ALIGNED(lpgeaddr, pgsz)));
4839 		if (type == F_SOFTLOCK) {
4840 			/*
4841 			 * For softlocks we cannot reduce the fault area
4842 			 * (calculated based on the largest page size for this
4843 			 * segment) for size down and a is already next
4844 			 * page size aligned as assertted above for size
4845 			 * ups. Therefore just continue in case of softlock.
4846 			 */
4847 			VM_STAT_ADD(segvnvmstats.fltanpages[9]);
4848 			continue; /* keep lint happy */
4849 		} else if (ierr == -2) {
4850 
4851 			/*
4852 			 * Size up case. Note lpgaddr may only be needed for
4853 			 * softlock case so we don't adjust it here.
4854 			 */
4855 			VM_STAT_ADD(segvnvmstats.fltanpages[10]);
4856 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4857 			ASSERT(a >= lpgaddr);
4858 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4859 			aindx = svd->anon_index + seg_page(seg, a);
4860 			vpage = (svd->vpage != NULL) ?
4861 			    &svd->vpage[seg_page(seg, a)] : NULL;
4862 		} else {
4863 			/*
4864 			 * Size down case. Note lpgaddr may only be needed for
4865 			 * softlock case so we don't adjust it here.
4866 			 */
4867 			VM_STAT_ADD(segvnvmstats.fltanpages[11]);
4868 			ASSERT(IS_P2ALIGNED(a, pgsz));
4869 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4870 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4871 			ASSERT(a < lpgeaddr);
4872 			if (a < addr) {
4873 				/*
4874 				 * The beginning of the large page region can
4875 				 * be pulled to the right to make a smaller
4876 				 * region. We haven't yet faulted a single
4877 				 * page.
4878 				 */
4879 				VM_STAT_ADD(segvnvmstats.fltanpages[12]);
4880 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4881 				ASSERT(a >= lpgaddr);
4882 				aindx = svd->anon_index + seg_page(seg, a);
4883 				vpage = (svd->vpage != NULL) ?
4884 				    &svd->vpage[seg_page(seg, a)] : NULL;
4885 			}
4886 		}
4887 	}
4888 	VM_STAT_ADD(segvnvmstats.fltanpages[13]);
4889 	ANON_LOCK_EXIT(&amp->a_rwlock);
4890 	kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4891 	return (0);
4892 error:
4893 	VM_STAT_ADD(segvnvmstats.fltanpages[14]);
4894 	ANON_LOCK_EXIT(&amp->a_rwlock);
4895 	kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4896 	if (type == F_SOFTLOCK && a > lpgaddr) {
4897 		VM_STAT_ADD(segvnvmstats.fltanpages[15]);
4898 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4899 	}
4900 	return (err);
4901 }
4902 
4903 int fltadvice = 1;	/* set to free behind pages for sequential access */
4904 
4905 /*
4906  * This routine is called via a machine specific fault handling routine.
4907  * It is also called by software routines wishing to lock or unlock
4908  * a range of addresses.
4909  *
4910  * Here is the basic algorithm:
4911  *	If unlocking
4912  *		Call segvn_softunlock
4913  *		Return
4914  *	endif
4915  *	Checking and set up work
4916  *	If we will need some non-anonymous pages
4917  *		Call VOP_GETPAGE over the range of non-anonymous pages
4918  *	endif
4919  *	Loop over all addresses requested
4920  *		Call segvn_faultpage passing in page list
4921  *		    to load up translations and handle anonymous pages
4922  *	endloop
4923  *	Load up translation to any additional pages in page list not
4924  *	    already handled that fit into this segment
4925  */
4926 static faultcode_t
4927 segvn_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
4928     enum fault_type type, enum seg_rw rw)
4929 {
4930 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4931 	page_t **plp, **ppp, *pp;
4932 	u_offset_t off;
4933 	caddr_t a;
4934 	struct vpage *vpage;
4935 	uint_t vpprot, prot;
4936 	int err;
4937 	page_t *pl[PVN_GETPAGE_NUM + 1];
4938 	size_t plsz, pl_alloc_sz;
4939 	size_t page;
4940 	ulong_t anon_index;
4941 	struct anon_map *amp;
4942 	int dogetpage = 0;
4943 	caddr_t	lpgaddr, lpgeaddr;
4944 	size_t pgsz;
4945 	anon_sync_obj_t cookie;
4946 	int brkcow = BREAK_COW_SHARE(rw, type, svd->type);
4947 
4948 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
4949 	ASSERT(svd->amp == NULL || svd->rcookie == HAT_INVALID_REGION_COOKIE);
4950 
4951 	/*
4952 	 * First handle the easy stuff
4953 	 */
4954 	if (type == F_SOFTUNLOCK) {
4955 		if (rw == S_READ_NOCOW) {
4956 			rw = S_READ;
4957 			ASSERT(AS_WRITE_HELD(seg->s_as));
4958 		}
4959 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4960 		pgsz = (seg->s_szc == 0) ? PAGESIZE :
4961 		    page_get_pagesize(seg->s_szc);
4962 		VM_STAT_COND_ADD(pgsz > PAGESIZE, segvnvmstats.fltanpages[16]);
4963 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
4964 		segvn_softunlock(seg, lpgaddr, lpgeaddr - lpgaddr, rw);
4965 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4966 		return (0);
4967 	}
4968 
4969 	ASSERT(svd->tr_state == SEGVN_TR_OFF ||
4970 	    !HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
4971 	if (brkcow == 0) {
4972 		if (svd->tr_state == SEGVN_TR_INIT) {
4973 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4974 			if (svd->tr_state == SEGVN_TR_INIT) {
4975 				ASSERT(svd->vp != NULL && svd->amp == NULL);
4976 				ASSERT(svd->flags & MAP_TEXT);
4977 				ASSERT(svd->type == MAP_PRIVATE);
4978 				segvn_textrepl(seg);
4979 				ASSERT(svd->tr_state != SEGVN_TR_INIT);
4980 				ASSERT(svd->tr_state != SEGVN_TR_ON ||
4981 				    svd->amp != NULL);
4982 			}
4983 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4984 		}
4985 	} else if (svd->tr_state != SEGVN_TR_OFF) {
4986 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4987 
4988 		if (rw == S_WRITE && svd->tr_state != SEGVN_TR_OFF) {
4989 			ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
4990 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4991 			return (FC_PROT);
4992 		}
4993 
4994 		if (svd->tr_state == SEGVN_TR_ON) {
4995 			ASSERT(svd->vp != NULL && svd->amp != NULL);
4996 			segvn_textunrepl(seg, 0);
4997 			ASSERT(svd->amp == NULL &&
4998 			    svd->tr_state == SEGVN_TR_OFF);
4999 		} else if (svd->tr_state != SEGVN_TR_OFF) {
5000 			svd->tr_state = SEGVN_TR_OFF;
5001 		}
5002 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
5003 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5004 	}
5005 
5006 top:
5007 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5008 
5009 	/*
5010 	 * If we have the same protections for the entire segment,
5011 	 * insure that the access being attempted is legitimate.
5012 	 */
5013 
5014 	if (svd->pageprot == 0) {
5015 		uint_t protchk;
5016 
5017 		switch (rw) {
5018 		case S_READ:
5019 		case S_READ_NOCOW:
5020 			protchk = PROT_READ;
5021 			break;
5022 		case S_WRITE:
5023 			protchk = PROT_WRITE;
5024 			break;
5025 		case S_EXEC:
5026 			protchk = PROT_EXEC;
5027 			break;
5028 		case S_OTHER:
5029 		default:
5030 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
5031 			break;
5032 		}
5033 
5034 		if ((svd->prot & protchk) == 0) {
5035 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5036 			return (FC_PROT);	/* illegal access type */
5037 		}
5038 	}
5039 
5040 	if (brkcow && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5041 		/* this must be SOFTLOCK S_READ fault */
5042 		ASSERT(svd->amp == NULL);
5043 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5044 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5045 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5046 		if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5047 			/*
5048 			 * this must be the first ever non S_READ_NOCOW
5049 			 * softlock for this segment.
5050 			 */
5051 			ASSERT(svd->softlockcnt == 0);
5052 			hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5053 			    HAT_REGION_TEXT);
5054 			svd->rcookie = HAT_INVALID_REGION_COOKIE;
5055 		}
5056 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5057 		goto top;
5058 	}
5059 
5060 	/*
5061 	 * We can't allow the long term use of softlocks for vmpss segments,
5062 	 * because in some file truncation cases we should be able to demote
5063 	 * the segment, which requires that there are no softlocks.  The
5064 	 * only case where it's ok to allow a SOFTLOCK fault against a vmpss
5065 	 * segment is S_READ_NOCOW, where the caller holds the address space
5066 	 * locked as writer and calls softunlock before dropping the as lock.
5067 	 * S_READ_NOCOW is used by /proc to read memory from another user.
5068 	 *
5069 	 * Another deadlock between SOFTLOCK and file truncation can happen
5070 	 * because segvn_fault_vnodepages() calls the FS one pagesize at
5071 	 * a time. A second VOP_GETPAGE() call by segvn_fault_vnodepages()
5072 	 * can cause a deadlock because the first set of page_t's remain
5073 	 * locked SE_SHARED.  To avoid this, we demote segments on a first
5074 	 * SOFTLOCK if they have a length greater than the segment's
5075 	 * page size.
5076 	 *
5077 	 * So for now, we only avoid demoting a segment on a SOFTLOCK when
5078 	 * the access type is S_READ_NOCOW and the fault length is less than
5079 	 * or equal to the segment's page size. While this is quite restrictive,
5080 	 * it should be the most common case of SOFTLOCK against a vmpss
5081 	 * segment.
5082 	 *
5083 	 * For S_READ_NOCOW, it's safe not to do a copy on write because the
5084 	 * caller makes sure no COW will be caused by another thread for a
5085 	 * softlocked page.
5086 	 */
5087 	if (type == F_SOFTLOCK && svd->vp != NULL && seg->s_szc != 0) {
5088 		int demote = 0;
5089 
5090 		if (rw != S_READ_NOCOW) {
5091 			demote = 1;
5092 		}
5093 		if (!demote && len > PAGESIZE) {
5094 			pgsz = page_get_pagesize(seg->s_szc);
5095 			CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr,
5096 			    lpgeaddr);
5097 			if (lpgeaddr - lpgaddr > pgsz) {
5098 				demote = 1;
5099 			}
5100 		}
5101 
5102 		ASSERT(demote || AS_WRITE_HELD(seg->s_as));
5103 
5104 		if (demote) {
5105 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5106 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5107 			if (seg->s_szc != 0) {
5108 				segvn_vmpss_clrszc_cnt++;
5109 				ASSERT(svd->softlockcnt == 0);
5110 				err = segvn_clrszc(seg);
5111 				if (err) {
5112 					segvn_vmpss_clrszc_err++;
5113 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5114 					return (FC_MAKE_ERR(err));
5115 				}
5116 			}
5117 			ASSERT(seg->s_szc == 0);
5118 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5119 			goto top;
5120 		}
5121 	}
5122 
5123 	/*
5124 	 * Check to see if we need to allocate an anon_map structure.
5125 	 */
5126 	if (svd->amp == NULL && (svd->vp == NULL || brkcow)) {
5127 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5128 		/*
5129 		 * Drop the "read" lock on the segment and acquire
5130 		 * the "write" version since we have to allocate the
5131 		 * anon_map.
5132 		 */
5133 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5134 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5135 
5136 		if (svd->amp == NULL) {
5137 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
5138 			svd->amp->a_szc = seg->s_szc;
5139 		}
5140 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5141 
5142 		/*
5143 		 * Start all over again since segment protections
5144 		 * may have changed after we dropped the "read" lock.
5145 		 */
5146 		goto top;
5147 	}
5148 
5149 	/*
5150 	 * S_READ_NOCOW vs S_READ distinction was
5151 	 * only needed for the code above. After
5152 	 * that we treat it as S_READ.
5153 	 */
5154 	if (rw == S_READ_NOCOW) {
5155 		ASSERT(type == F_SOFTLOCK);
5156 		ASSERT(AS_WRITE_HELD(seg->s_as));
5157 		rw = S_READ;
5158 	}
5159 
5160 	amp = svd->amp;
5161 
5162 	/*
5163 	 * MADV_SEQUENTIAL work is ignored for large page segments.
5164 	 */
5165 	if (seg->s_szc != 0) {
5166 		pgsz = page_get_pagesize(seg->s_szc);
5167 		ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
5168 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
5169 		if (svd->vp == NULL) {
5170 			err = segvn_fault_anonpages(hat, seg, lpgaddr,
5171 			    lpgeaddr, type, rw, addr, addr + len, brkcow);
5172 		} else {
5173 			err = segvn_fault_vnodepages(hat, seg, lpgaddr,
5174 			    lpgeaddr, type, rw, addr, addr + len, brkcow);
5175 			if (err == IE_RETRY) {
5176 				ASSERT(seg->s_szc == 0);
5177 				ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
5178 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5179 				goto top;
5180 			}
5181 		}
5182 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5183 		return (err);
5184 	}
5185 
5186 	page = seg_page(seg, addr);
5187 	if (amp != NULL) {
5188 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5189 		anon_index = svd->anon_index + page;
5190 
5191 		if (type == F_PROT && rw == S_READ &&
5192 		    svd->tr_state == SEGVN_TR_OFF &&
5193 		    svd->type == MAP_PRIVATE && svd->pageprot == 0) {
5194 			size_t index = anon_index;
5195 			struct anon *ap;
5196 
5197 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5198 			/*
5199 			 * The fast path could apply to S_WRITE also, except
5200 			 * that the protection fault could be caused by lazy
5201 			 * tlb flush when ro->rw. In this case, the pte is
5202 			 * RW already. But RO in the other cpu's tlb causes
5203 			 * the fault. Since hat_chgprot won't do anything if
5204 			 * pte doesn't change, we may end up faulting
5205 			 * indefinitely until the RO tlb entry gets replaced.
5206 			 */
5207 			for (a = addr; a < addr + len; a += PAGESIZE, index++) {
5208 				anon_array_enter(amp, index, &cookie);
5209 				ap = anon_get_ptr(amp->ahp, index);
5210 				anon_array_exit(&cookie);
5211 				if ((ap == NULL) || (ap->an_refcnt != 1)) {
5212 					ANON_LOCK_EXIT(&amp->a_rwlock);
5213 					goto slow;
5214 				}
5215 			}
5216 			hat_chgprot(seg->s_as->a_hat, addr, len, svd->prot);
5217 			ANON_LOCK_EXIT(&amp->a_rwlock);
5218 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5219 			return (0);
5220 		}
5221 	}
5222 slow:
5223 
5224 	if (svd->vpage == NULL)
5225 		vpage = NULL;
5226 	else
5227 		vpage = &svd->vpage[page];
5228 
5229 	off = svd->offset + (uintptr_t)(addr - seg->s_base);
5230 
5231 	/*
5232 	 * If MADV_SEQUENTIAL has been set for the particular page we
5233 	 * are faulting on, free behind all pages in the segment and put
5234 	 * them on the free list.
5235 	 */
5236 
5237 	if ((page != 0) && fltadvice && svd->tr_state != SEGVN_TR_ON) {
5238 		struct vpage *vpp;
5239 		ulong_t fanon_index;
5240 		size_t fpage;
5241 		u_offset_t pgoff, fpgoff;
5242 		struct vnode *fvp;
5243 		struct anon *fap = NULL;
5244 
5245 		if (svd->advice == MADV_SEQUENTIAL ||
5246 		    (svd->pageadvice &&
5247 		    VPP_ADVICE(vpage) == MADV_SEQUENTIAL)) {
5248 			pgoff = off - PAGESIZE;
5249 			fpage = page - 1;
5250 			if (vpage != NULL)
5251 				vpp = &svd->vpage[fpage];
5252 			if (amp != NULL)
5253 				fanon_index = svd->anon_index + fpage;
5254 
5255 			while (pgoff > svd->offset) {
5256 				if (svd->advice != MADV_SEQUENTIAL &&
5257 				    (!svd->pageadvice || (vpage &&
5258 				    VPP_ADVICE(vpp) != MADV_SEQUENTIAL)))
5259 					break;
5260 
5261 				/*
5262 				 * If this is an anon page, we must find the
5263 				 * correct <vp, offset> for it
5264 				 */
5265 				fap = NULL;
5266 				if (amp != NULL) {
5267 					ANON_LOCK_ENTER(&amp->a_rwlock,
5268 					    RW_READER);
5269 					anon_array_enter(amp, fanon_index,
5270 					    &cookie);
5271 					fap = anon_get_ptr(amp->ahp,
5272 					    fanon_index);
5273 					if (fap != NULL) {
5274 						swap_xlate(fap, &fvp, &fpgoff);
5275 					} else {
5276 						fpgoff = pgoff;
5277 						fvp = svd->vp;
5278 					}
5279 					anon_array_exit(&cookie);
5280 					ANON_LOCK_EXIT(&amp->a_rwlock);
5281 				} else {
5282 					fpgoff = pgoff;
5283 					fvp = svd->vp;
5284 				}
5285 				if (fvp == NULL)
5286 					break;	/* XXX */
5287 				/*
5288 				 * Skip pages that are free or have an
5289 				 * "exclusive" lock.
5290 				 */
5291 				pp = page_lookup_nowait(fvp, fpgoff, SE_SHARED);
5292 				if (pp == NULL)
5293 					break;
5294 				/*
5295 				 * We don't need the page_struct_lock to test
5296 				 * as this is only advisory; even if we
5297 				 * acquire it someone might race in and lock
5298 				 * the page after we unlock and before the
5299 				 * PUTPAGE, then VOP_PUTPAGE will do nothing.
5300 				 */
5301 				if (pp->p_lckcnt == 0 && pp->p_cowcnt == 0) {
5302 					/*
5303 					 * Hold the vnode before releasing
5304 					 * the page lock to prevent it from
5305 					 * being freed and re-used by some
5306 					 * other thread.
5307 					 */
5308 					VN_HOLD(fvp);
5309 					page_unlock(pp);
5310 					/*
5311 					 * We should build a page list
5312 					 * to kluster putpages XXX
5313 					 */
5314 					(void) VOP_PUTPAGE(fvp,
5315 					    (offset_t)fpgoff, PAGESIZE,
5316 					    (B_DONTNEED|B_FREE|B_ASYNC),
5317 					    svd->cred, NULL);
5318 					VN_RELE(fvp);
5319 				} else {
5320 					/*
5321 					 * XXX - Should the loop terminate if
5322 					 * the page is `locked'?
5323 					 */
5324 					page_unlock(pp);
5325 				}
5326 				--vpp;
5327 				--fanon_index;
5328 				pgoff -= PAGESIZE;
5329 			}
5330 		}
5331 	}
5332 
5333 	plp = pl;
5334 	*plp = NULL;
5335 	pl_alloc_sz = 0;
5336 
5337 	/*
5338 	 * See if we need to call VOP_GETPAGE for
5339 	 * *any* of the range being faulted on.
5340 	 * We can skip all of this work if there
5341 	 * was no original vnode.
5342 	 */
5343 	if (svd->vp != NULL) {
5344 		u_offset_t vp_off;
5345 		size_t vp_len;
5346 		struct anon *ap;
5347 		vnode_t *vp;
5348 
5349 		vp_off = off;
5350 		vp_len = len;
5351 
5352 		if (amp == NULL)
5353 			dogetpage = 1;
5354 		else {
5355 			/*
5356 			 * Only acquire reader lock to prevent amp->ahp
5357 			 * from being changed.  It's ok to miss pages,
5358 			 * hence we don't do anon_array_enter
5359 			 */
5360 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5361 			ap = anon_get_ptr(amp->ahp, anon_index);
5362 
5363 			if (len <= PAGESIZE)
5364 				/* inline non_anon() */
5365 				dogetpage = (ap == NULL);
5366 			else
5367 				dogetpage = non_anon(amp->ahp, anon_index,
5368 				    &vp_off, &vp_len);
5369 			ANON_LOCK_EXIT(&amp->a_rwlock);
5370 		}
5371 
5372 		if (dogetpage) {
5373 			enum seg_rw arw;
5374 			struct as *as = seg->s_as;
5375 
5376 			if (len > ptob((sizeof (pl) / sizeof (pl[0])) - 1)) {
5377 				/*
5378 				 * Page list won't fit in local array,
5379 				 * allocate one of the needed size.
5380 				 */
5381 				pl_alloc_sz =
5382 				    (btop(len) + 1) * sizeof (page_t *);
5383 				plp = kmem_alloc(pl_alloc_sz, KM_SLEEP);
5384 				plp[0] = NULL;
5385 				plsz = len;
5386 			} else if (rw == S_WRITE && svd->type == MAP_PRIVATE ||
5387 			    svd->tr_state == SEGVN_TR_ON || rw == S_OTHER ||
5388 			    (((size_t)(addr + PAGESIZE) <
5389 			    (size_t)(seg->s_base + seg->s_size)) &&
5390 			    hat_probe(as->a_hat, addr + PAGESIZE))) {
5391 				/*
5392 				 * Ask VOP_GETPAGE to return the exact number
5393 				 * of pages if
5394 				 * (a) this is a COW fault, or
5395 				 * (b) this is a software fault, or
5396 				 * (c) next page is already mapped.
5397 				 */
5398 				plsz = len;
5399 			} else {
5400 				/*
5401 				 * Ask VOP_GETPAGE to return adjacent pages
5402 				 * within the segment.
5403 				 */
5404 				plsz = MIN((size_t)PVN_GETPAGE_SZ, (size_t)
5405 				    ((seg->s_base + seg->s_size) - addr));
5406 				ASSERT((addr + plsz) <=
5407 				    (seg->s_base + seg->s_size));
5408 			}
5409 
5410 			/*
5411 			 * Need to get some non-anonymous pages.
5412 			 * We need to make only one call to GETPAGE to do
5413 			 * this to prevent certain deadlocking conditions
5414 			 * when we are doing locking.  In this case
5415 			 * non_anon() should have picked up the smallest
5416 			 * range which includes all the non-anonymous
5417 			 * pages in the requested range.  We have to
5418 			 * be careful regarding which rw flag to pass in
5419 			 * because on a private mapping, the underlying
5420 			 * object is never allowed to be written.
5421 			 */
5422 			if (rw == S_WRITE && svd->type == MAP_PRIVATE) {
5423 				arw = S_READ;
5424 			} else {
5425 				arw = rw;
5426 			}
5427 			vp = svd->vp;
5428 			TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5429 			    "segvn_getpage:seg %p addr %p vp %p",
5430 			    seg, addr, vp);
5431 			err = VOP_GETPAGE(vp, (offset_t)vp_off, vp_len,
5432 			    &vpprot, plp, plsz, seg, addr + (vp_off - off), arw,
5433 			    svd->cred, NULL);
5434 			if (err) {
5435 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5436 				segvn_pagelist_rele(plp);
5437 				if (pl_alloc_sz)
5438 					kmem_free(plp, pl_alloc_sz);
5439 				return (FC_MAKE_ERR(err));
5440 			}
5441 			if (svd->type == MAP_PRIVATE)
5442 				vpprot &= ~PROT_WRITE;
5443 		}
5444 	}
5445 
5446 	/*
5447 	 * N.B. at this time the plp array has all the needed non-anon
5448 	 * pages in addition to (possibly) having some adjacent pages.
5449 	 */
5450 
5451 	/*
5452 	 * Always acquire the anon_array_lock to prevent
5453 	 * 2 threads from allocating separate anon slots for
5454 	 * the same "addr".
5455 	 *
5456 	 * If this is a copy-on-write fault and we don't already
5457 	 * have the anon_array_lock, acquire it to prevent the
5458 	 * fault routine from handling multiple copy-on-write faults
5459 	 * on the same "addr" in the same address space.
5460 	 *
5461 	 * Only one thread should deal with the fault since after
5462 	 * it is handled, the other threads can acquire a translation
5463 	 * to the newly created private page.  This prevents two or
5464 	 * more threads from creating different private pages for the
5465 	 * same fault.
5466 	 *
5467 	 * We grab "serialization" lock here if this is a MAP_PRIVATE segment
5468 	 * to prevent deadlock between this thread and another thread
5469 	 * which has soft-locked this page and wants to acquire serial_lock.
5470 	 * ( bug 4026339 )
5471 	 *
5472 	 * The fix for bug 4026339 becomes unnecessary when using the
5473 	 * locking scheme with per amp rwlock and a global set of hash
5474 	 * lock, anon_array_lock.  If we steal a vnode page when low
5475 	 * on memory and upgrad the page lock through page_rename,
5476 	 * then the page is PAGE_HANDLED, nothing needs to be done
5477 	 * for this page after returning from segvn_faultpage.
5478 	 *
5479 	 * But really, the page lock should be downgraded after
5480 	 * the stolen page is page_rename'd.
5481 	 */
5482 
5483 	if (amp != NULL)
5484 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5485 
5486 	/*
5487 	 * Ok, now loop over the address range and handle faults
5488 	 */
5489 	for (a = addr; a < addr + len; a += PAGESIZE, off += PAGESIZE) {
5490 		err = segvn_faultpage(hat, seg, a, off, vpage, plp, vpprot,
5491 		    type, rw, brkcow);
5492 		if (err) {
5493 			if (amp != NULL)
5494 				ANON_LOCK_EXIT(&amp->a_rwlock);
5495 			if (type == F_SOFTLOCK && a > addr) {
5496 				segvn_softunlock(seg, addr, (a - addr),
5497 				    S_OTHER);
5498 			}
5499 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5500 			segvn_pagelist_rele(plp);
5501 			if (pl_alloc_sz)
5502 				kmem_free(plp, pl_alloc_sz);
5503 			return (err);
5504 		}
5505 		if (vpage) {
5506 			vpage++;
5507 		} else if (svd->vpage) {
5508 			page = seg_page(seg, addr);
5509 			vpage = &svd->vpage[++page];
5510 		}
5511 	}
5512 
5513 	/* Didn't get pages from the underlying fs so we're done */
5514 	if (!dogetpage)
5515 		goto done;
5516 
5517 	/*
5518 	 * Now handle any other pages in the list returned.
5519 	 * If the page can be used, load up the translations now.
5520 	 * Note that the for loop will only be entered if "plp"
5521 	 * is pointing to a non-NULL page pointer which means that
5522 	 * VOP_GETPAGE() was called and vpprot has been initialized.
5523 	 */
5524 	if (svd->pageprot == 0)
5525 		prot = svd->prot & vpprot;
5526 
5527 
5528 	/*
5529 	 * Large Files: diff should be unsigned value because we started
5530 	 * supporting > 2GB segment sizes from 2.5.1 and when a
5531 	 * large file of size > 2GB gets mapped to address space
5532 	 * the diff value can be > 2GB.
5533 	 */
5534 
5535 	for (ppp = plp; (pp = *ppp) != NULL; ppp++) {
5536 		size_t diff;
5537 		struct anon *ap;
5538 		int anon_index;
5539 		anon_sync_obj_t cookie;
5540 		int hat_flag = HAT_LOAD_ADV;
5541 
5542 		if (svd->flags & MAP_TEXT) {
5543 			hat_flag |= HAT_LOAD_TEXT;
5544 		}
5545 
5546 		if (pp == PAGE_HANDLED)
5547 			continue;
5548 
5549 		if (svd->tr_state != SEGVN_TR_ON &&
5550 		    pp->p_offset >=  svd->offset &&
5551 		    pp->p_offset < svd->offset + seg->s_size) {
5552 
5553 			diff = pp->p_offset - svd->offset;
5554 
5555 			/*
5556 			 * Large Files: Following is the assertion
5557 			 * validating the above cast.
5558 			 */
5559 			ASSERT(svd->vp == pp->p_vnode);
5560 
5561 			page = btop(diff);
5562 			if (svd->pageprot)
5563 				prot = VPP_PROT(&svd->vpage[page]) & vpprot;
5564 
5565 			/*
5566 			 * Prevent other threads in the address space from
5567 			 * creating private pages (i.e., allocating anon slots)
5568 			 * while we are in the process of loading translations
5569 			 * to additional pages returned by the underlying
5570 			 * object.
5571 			 */
5572 			if (amp != NULL) {
5573 				anon_index = svd->anon_index + page;
5574 				anon_array_enter(amp, anon_index, &cookie);
5575 				ap = anon_get_ptr(amp->ahp, anon_index);
5576 			}
5577 			if ((amp == NULL) || (ap == NULL)) {
5578 				if (IS_VMODSORT(pp->p_vnode) ||
5579 				    enable_mbit_wa) {
5580 					if (rw == S_WRITE)
5581 						hat_setmod(pp);
5582 					else if (rw != S_OTHER &&
5583 					    !hat_ismod(pp))
5584 						prot &= ~PROT_WRITE;
5585 				}
5586 				/*
5587 				 * Skip mapping read ahead pages marked
5588 				 * for migration, so they will get migrated
5589 				 * properly on fault
5590 				 */
5591 				ASSERT(amp == NULL ||
5592 				    svd->rcookie == HAT_INVALID_REGION_COOKIE);
5593 				if ((prot & PROT_READ) && !PP_ISMIGRATE(pp)) {
5594 					hat_memload_region(hat,
5595 					    seg->s_base + diff,
5596 					    pp, prot, hat_flag,
5597 					    svd->rcookie);
5598 				}
5599 			}
5600 			if (amp != NULL)
5601 				anon_array_exit(&cookie);
5602 		}
5603 		page_unlock(pp);
5604 	}
5605 done:
5606 	if (amp != NULL)
5607 		ANON_LOCK_EXIT(&amp->a_rwlock);
5608 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5609 	if (pl_alloc_sz)
5610 		kmem_free(plp, pl_alloc_sz);
5611 	return (0);
5612 }
5613 
5614 /*
5615  * This routine is used to start I/O on pages asynchronously.  XXX it will
5616  * only create PAGESIZE pages. At fault time they will be relocated into
5617  * larger pages.
5618  */
5619 static faultcode_t
5620 segvn_faulta(struct seg *seg, caddr_t addr)
5621 {
5622 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5623 	int err;
5624 	struct anon_map *amp;
5625 	vnode_t *vp;
5626 
5627 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
5628 
5629 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5630 	if ((amp = svd->amp) != NULL) {
5631 		struct anon *ap;
5632 
5633 		/*
5634 		 * Reader lock to prevent amp->ahp from being changed.
5635 		 * This is advisory, it's ok to miss a page, so
5636 		 * we don't do anon_array_enter lock.
5637 		 */
5638 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5639 		if ((ap = anon_get_ptr(amp->ahp,
5640 		    svd->anon_index + seg_page(seg, addr))) != NULL) {
5641 
5642 			err = anon_getpage(&ap, NULL, NULL,
5643 			    0, seg, addr, S_READ, svd->cred);
5644 
5645 			ANON_LOCK_EXIT(&amp->a_rwlock);
5646 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5647 			if (err)
5648 				return (FC_MAKE_ERR(err));
5649 			return (0);
5650 		}
5651 		ANON_LOCK_EXIT(&amp->a_rwlock);
5652 	}
5653 
5654 	if (svd->vp == NULL) {
5655 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5656 		return (0);			/* zfod page - do nothing now */
5657 	}
5658 
5659 	vp = svd->vp;
5660 	TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5661 	    "segvn_getpage:seg %p addr %p vp %p", seg, addr, vp);
5662 	err = VOP_GETPAGE(vp,
5663 	    (offset_t)(svd->offset + (uintptr_t)(addr - seg->s_base)),
5664 	    PAGESIZE, NULL, NULL, 0, seg, addr,
5665 	    S_OTHER, svd->cred, NULL);
5666 
5667 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5668 	if (err)
5669 		return (FC_MAKE_ERR(err));
5670 	return (0);
5671 }
5672 
5673 static int
5674 segvn_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
5675 {
5676 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5677 	struct vpage *cvp, *svp, *evp;
5678 	struct vnode *vp;
5679 	size_t pgsz;
5680 	pgcnt_t pgcnt;
5681 	anon_sync_obj_t cookie;
5682 	int unload_done = 0;
5683 
5684 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
5685 
5686 	if ((svd->maxprot & prot) != prot)
5687 		return (EACCES);			/* violated maxprot */
5688 
5689 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5690 
5691 	/* return if prot is the same */
5692 	if (!svd->pageprot && svd->prot == prot) {
5693 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5694 		return (0);
5695 	}
5696 
5697 	/*
5698 	 * Since we change protections we first have to flush the cache.
5699 	 * This makes sure all the pagelock calls have to recheck
5700 	 * protections.
5701 	 */
5702 	if (svd->softlockcnt > 0) {
5703 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5704 
5705 		/*
5706 		 * If this is shared segment non 0 softlockcnt
5707 		 * means locked pages are still in use.
5708 		 */
5709 		if (svd->type == MAP_SHARED) {
5710 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5711 			return (EAGAIN);
5712 		}
5713 
5714 		/*
5715 		 * Since we do have the segvn writers lock nobody can fill
5716 		 * the cache with entries belonging to this seg during
5717 		 * the purge. The flush either succeeds or we still have
5718 		 * pending I/Os.
5719 		 */
5720 		segvn_purge(seg);
5721 		if (svd->softlockcnt > 0) {
5722 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5723 			return (EAGAIN);
5724 		}
5725 	}
5726 
5727 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5728 		ASSERT(svd->amp == NULL);
5729 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5730 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5731 		    HAT_REGION_TEXT);
5732 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
5733 		unload_done = 1;
5734 	} else if (svd->tr_state == SEGVN_TR_INIT) {
5735 		svd->tr_state = SEGVN_TR_OFF;
5736 	} else if (svd->tr_state == SEGVN_TR_ON) {
5737 		ASSERT(svd->amp != NULL);
5738 		segvn_textunrepl(seg, 0);
5739 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
5740 		unload_done = 1;
5741 	}
5742 
5743 	if ((prot & PROT_WRITE) && svd->type == MAP_SHARED &&
5744 	    svd->vp != NULL && (svd->vp->v_flag & VVMEXEC)) {
5745 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
5746 		segvn_inval_trcache(svd->vp);
5747 	}
5748 	if (seg->s_szc != 0) {
5749 		int err;
5750 		pgsz = page_get_pagesize(seg->s_szc);
5751 		pgcnt = pgsz >> PAGESHIFT;
5752 		ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
5753 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
5754 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5755 			ASSERT(seg->s_base != addr || seg->s_size != len);
5756 			/*
5757 			 * If we are holding the as lock as a reader then
5758 			 * we need to return IE_RETRY and let the as
5759 			 * layer drop and re-acquire the lock as a writer.
5760 			 */
5761 			if (AS_READ_HELD(seg->s_as))
5762 				return (IE_RETRY);
5763 			VM_STAT_ADD(segvnvmstats.demoterange[1]);
5764 			if (svd->type == MAP_PRIVATE || svd->vp != NULL) {
5765 				err = segvn_demote_range(seg, addr, len,
5766 				    SDR_END, 0);
5767 			} else {
5768 				uint_t szcvec = map_pgszcvec(seg->s_base,
5769 				    pgsz, (uintptr_t)seg->s_base,
5770 				    (svd->flags & MAP_TEXT), MAPPGSZC_SHM, 0);
5771 				err = segvn_demote_range(seg, addr, len,
5772 				    SDR_END, szcvec);
5773 			}
5774 			if (err == 0)
5775 				return (IE_RETRY);
5776 			if (err == ENOMEM)
5777 				return (IE_NOMEM);
5778 			return (err);
5779 		}
5780 	}
5781 
5782 
5783 	/*
5784 	 * If it's a private mapping and we're making it writable then we
5785 	 * may have to reserve the additional swap space now. If we are
5786 	 * making writable only a part of the segment then we use its vpage
5787 	 * array to keep a record of the pages for which we have reserved
5788 	 * swap. In this case we set the pageswap field in the segment's
5789 	 * segvn structure to record this.
5790 	 *
5791 	 * If it's a private mapping to a file (i.e., vp != NULL) and we're
5792 	 * removing write permission on the entire segment and we haven't
5793 	 * modified any pages, we can release the swap space.
5794 	 */
5795 	if (svd->type == MAP_PRIVATE) {
5796 		if (prot & PROT_WRITE) {
5797 			if (!(svd->flags & MAP_NORESERVE) &&
5798 			    !(svd->swresv && svd->pageswap == 0)) {
5799 				size_t sz = 0;
5800 
5801 				/*
5802 				 * Start by determining how much swap
5803 				 * space is required.
5804 				 */
5805 				if (addr == seg->s_base &&
5806 				    len == seg->s_size &&
5807 				    svd->pageswap == 0) {
5808 					/* The whole segment */
5809 					sz = seg->s_size;
5810 				} else {
5811 					/*
5812 					 * Make sure that the vpage array
5813 					 * exists, and make a note of the
5814 					 * range of elements corresponding
5815 					 * to len.
5816 					 */
5817 					segvn_vpage(seg);
5818 					if (svd->vpage == NULL) {
5819 						SEGVN_LOCK_EXIT(seg->s_as,
5820 						    &svd->lock);
5821 						return (ENOMEM);
5822 					}
5823 					svp = &svd->vpage[seg_page(seg, addr)];
5824 					evp = &svd->vpage[seg_page(seg,
5825 					    addr + len)];
5826 
5827 					if (svd->pageswap == 0) {
5828 						/*
5829 						 * This is the first time we've
5830 						 * asked for a part of this
5831 						 * segment, so we need to
5832 						 * reserve everything we've
5833 						 * been asked for.
5834 						 */
5835 						sz = len;
5836 					} else {
5837 						/*
5838 						 * We have to count the number
5839 						 * of pages required.
5840 						 */
5841 						for (cvp = svp;  cvp < evp;
5842 						    cvp++) {
5843 							if (!VPP_ISSWAPRES(cvp))
5844 								sz++;
5845 						}
5846 						sz <<= PAGESHIFT;
5847 					}
5848 				}
5849 
5850 				/* Try to reserve the necessary swap. */
5851 				if (anon_resv_zone(sz,
5852 				    seg->s_as->a_proc->p_zone) == 0) {
5853 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5854 					return (IE_NOMEM);
5855 				}
5856 
5857 				/*
5858 				 * Make a note of how much swap space
5859 				 * we've reserved.
5860 				 */
5861 				if (svd->pageswap == 0 && sz == seg->s_size) {
5862 					svd->swresv = sz;
5863 				} else {
5864 					ASSERT(svd->vpage != NULL);
5865 					svd->swresv += sz;
5866 					svd->pageswap = 1;
5867 					for (cvp = svp; cvp < evp; cvp++) {
5868 						if (!VPP_ISSWAPRES(cvp))
5869 							VPP_SETSWAPRES(cvp);
5870 					}
5871 				}
5872 			}
5873 		} else {
5874 			/*
5875 			 * Swap space is released only if this segment
5876 			 * does not map anonymous memory, since read faults
5877 			 * on such segments still need an anon slot to read
5878 			 * in the data.
5879 			 */
5880 			if (svd->swresv != 0 && svd->vp != NULL &&
5881 			    svd->amp == NULL && addr == seg->s_base &&
5882 			    len == seg->s_size && svd->pageprot == 0) {
5883 				ASSERT(svd->pageswap == 0);
5884 				anon_unresv_zone(svd->swresv,
5885 				    seg->s_as->a_proc->p_zone);
5886 				svd->swresv = 0;
5887 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
5888 				    "anon proc:%p %lu %u", seg, 0, 0);
5889 			}
5890 		}
5891 	}
5892 
5893 	if (addr == seg->s_base && len == seg->s_size && svd->vpage == NULL) {
5894 		if (svd->prot == prot) {
5895 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5896 			return (0);			/* all done */
5897 		}
5898 		svd->prot = (uchar_t)prot;
5899 	} else if (svd->type == MAP_PRIVATE) {
5900 		struct anon *ap = NULL;
5901 		page_t *pp;
5902 		u_offset_t offset, off;
5903 		struct anon_map *amp;
5904 		ulong_t anon_idx = 0;
5905 
5906 		/*
5907 		 * A vpage structure exists or else the change does not
5908 		 * involve the entire segment.  Establish a vpage structure
5909 		 * if none is there.  Then, for each page in the range,
5910 		 * adjust its individual permissions.  Note that write-
5911 		 * enabling a MAP_PRIVATE page can affect the claims for
5912 		 * locked down memory.  Overcommitting memory terminates
5913 		 * the operation.
5914 		 */
5915 		segvn_vpage(seg);
5916 		if (svd->vpage == NULL) {
5917 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5918 			return (ENOMEM);
5919 		}
5920 		svd->pageprot = 1;
5921 		if ((amp = svd->amp) != NULL) {
5922 			anon_idx = svd->anon_index + seg_page(seg, addr);
5923 			ASSERT(seg->s_szc == 0 ||
5924 			    IS_P2ALIGNED(anon_idx, pgcnt));
5925 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5926 		}
5927 
5928 		offset = svd->offset + (uintptr_t)(addr - seg->s_base);
5929 		evp = &svd->vpage[seg_page(seg, addr + len)];
5930 
5931 		/*
5932 		 * See Statement at the beginning of segvn_lockop regarding
5933 		 * the way cowcnts and lckcnts are handled.
5934 		 */
5935 		for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5936 
5937 			if (seg->s_szc != 0) {
5938 				if (amp != NULL) {
5939 					anon_array_enter(amp, anon_idx,
5940 					    &cookie);
5941 				}
5942 				if (IS_P2ALIGNED(anon_idx, pgcnt) &&
5943 				    !segvn_claim_pages(seg, svp, offset,
5944 				    anon_idx, prot)) {
5945 					if (amp != NULL) {
5946 						anon_array_exit(&cookie);
5947 					}
5948 					break;
5949 				}
5950 				if (amp != NULL) {
5951 					anon_array_exit(&cookie);
5952 				}
5953 				anon_idx++;
5954 			} else {
5955 				if (amp != NULL) {
5956 					anon_array_enter(amp, anon_idx,
5957 					    &cookie);
5958 					ap = anon_get_ptr(amp->ahp, anon_idx++);
5959 				}
5960 
5961 				if (VPP_ISPPLOCK(svp) &&
5962 				    VPP_PROT(svp) != prot) {
5963 
5964 					if (amp == NULL || ap == NULL) {
5965 						vp = svd->vp;
5966 						off = offset;
5967 					} else
5968 						swap_xlate(ap, &vp, &off);
5969 					if (amp != NULL)
5970 						anon_array_exit(&cookie);
5971 
5972 					if ((pp = page_lookup(vp, off,
5973 					    SE_SHARED)) == NULL) {
5974 						panic("segvn_setprot: no page");
5975 						/*NOTREACHED*/
5976 					}
5977 					ASSERT(seg->s_szc == 0);
5978 					if ((VPP_PROT(svp) ^ prot) &
5979 					    PROT_WRITE) {
5980 						if (prot & PROT_WRITE) {
5981 							if (!page_addclaim(
5982 							    pp)) {
5983 								page_unlock(pp);
5984 								break;
5985 							}
5986 						} else {
5987 							if (!page_subclaim(
5988 							    pp)) {
5989 								page_unlock(pp);
5990 								break;
5991 							}
5992 						}
5993 					}
5994 					page_unlock(pp);
5995 				} else if (amp != NULL)
5996 					anon_array_exit(&cookie);
5997 			}
5998 			VPP_SETPROT(svp, prot);
5999 			offset += PAGESIZE;
6000 		}
6001 		if (amp != NULL)
6002 			ANON_LOCK_EXIT(&amp->a_rwlock);
6003 
6004 		/*
6005 		 * Did we terminate prematurely?  If so, simply unload
6006 		 * the translations to the things we've updated so far.
6007 		 */
6008 		if (svp != evp) {
6009 			if (unload_done) {
6010 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6011 				return (IE_NOMEM);
6012 			}
6013 			len = (svp - &svd->vpage[seg_page(seg, addr)]) *
6014 			    PAGESIZE;
6015 			ASSERT(seg->s_szc == 0 || IS_P2ALIGNED(len, pgsz));
6016 			if (len != 0)
6017 				hat_unload(seg->s_as->a_hat, addr,
6018 				    len, HAT_UNLOAD);
6019 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6020 			return (IE_NOMEM);
6021 		}
6022 	} else {
6023 		segvn_vpage(seg);
6024 		if (svd->vpage == NULL) {
6025 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6026 			return (ENOMEM);
6027 		}
6028 		svd->pageprot = 1;
6029 		evp = &svd->vpage[seg_page(seg, addr + len)];
6030 		for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
6031 			VPP_SETPROT(svp, prot);
6032 		}
6033 	}
6034 
6035 	if (unload_done) {
6036 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6037 		return (0);
6038 	}
6039 
6040 	if (((prot & PROT_WRITE) != 0 &&
6041 	    (svd->vp != NULL || svd->type == MAP_PRIVATE)) ||
6042 	    (prot & ~PROT_USER) == PROT_NONE) {
6043 		/*
6044 		 * Either private or shared data with write access (in
6045 		 * which case we need to throw out all former translations
6046 		 * so that we get the right translations set up on fault
6047 		 * and we don't allow write access to any copy-on-write pages
6048 		 * that might be around or to prevent write access to pages
6049 		 * representing holes in a file), or we don't have permission
6050 		 * to access the memory at all (in which case we have to
6051 		 * unload any current translations that might exist).
6052 		 */
6053 		hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
6054 	} else {
6055 		/*
6056 		 * A shared mapping or a private mapping in which write
6057 		 * protection is going to be denied - just change all the
6058 		 * protections over the range of addresses in question.
6059 		 * segvn does not support any other attributes other
6060 		 * than prot so we can use hat_chgattr.
6061 		 */
6062 		hat_chgattr(seg->s_as->a_hat, addr, len, prot);
6063 	}
6064 
6065 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6066 
6067 	return (0);
6068 }
6069 
6070 /*
6071  * segvn_setpagesize is called via SEGOP_SETPAGESIZE from as_setpagesize,
6072  * to determine if the seg is capable of mapping the requested szc.
6073  */
6074 static int
6075 segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len, uint_t szc)
6076 {
6077 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6078 	struct segvn_data *nsvd;
6079 	struct anon_map *amp = svd->amp;
6080 	struct seg *nseg;
6081 	caddr_t eaddr = addr + len, a;
6082 	size_t pgsz = page_get_pagesize(szc);
6083 	pgcnt_t pgcnt = page_get_pagecnt(szc);
6084 	int err;
6085 	u_offset_t off = svd->offset + (uintptr_t)(addr - seg->s_base);
6086 
6087 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
6088 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6089 
6090 	if (seg->s_szc == szc || segvn_lpg_disable != 0) {
6091 		return (0);
6092 	}
6093 
6094 	/*
6095 	 * addr should always be pgsz aligned but eaddr may be misaligned if
6096 	 * it's at the end of the segment.
6097 	 *
6098 	 * XXX we should assert this condition since as_setpagesize() logic
6099 	 * guarantees it.
6100 	 */
6101 	if (!IS_P2ALIGNED(addr, pgsz) ||
6102 	    (!IS_P2ALIGNED(eaddr, pgsz) &&
6103 	    eaddr != seg->s_base + seg->s_size)) {
6104 
6105 		segvn_setpgsz_align_err++;
6106 		return (EINVAL);
6107 	}
6108 
6109 	if (amp != NULL && svd->type == MAP_SHARED) {
6110 		ulong_t an_idx = svd->anon_index + seg_page(seg, addr);
6111 		if (!IS_P2ALIGNED(an_idx, pgcnt)) {
6112 
6113 			segvn_setpgsz_anon_align_err++;
6114 			return (EINVAL);
6115 		}
6116 	}
6117 
6118 	if ((svd->flags & MAP_NORESERVE) || seg->s_as == &kas ||
6119 	    szc > segvn_maxpgszc) {
6120 		return (EINVAL);
6121 	}
6122 
6123 	/* paranoid check */
6124 	if (svd->vp != NULL &&
6125 	    (IS_SWAPFSVP(svd->vp) || VN_ISKAS(svd->vp))) {
6126 		return (EINVAL);
6127 	}
6128 
6129 	if (seg->s_szc == 0 && svd->vp != NULL &&
6130 	    map_addr_vacalign_check(addr, off)) {
6131 		return (EINVAL);
6132 	}
6133 
6134 	/*
6135 	 * Check that protections are the same within new page
6136 	 * size boundaries.
6137 	 */
6138 	if (svd->pageprot) {
6139 		for (a = addr; a < eaddr; a += pgsz) {
6140 			if ((a + pgsz) > eaddr) {
6141 				if (!sameprot(seg, a, eaddr - a)) {
6142 					return (EINVAL);
6143 				}
6144 			} else {
6145 				if (!sameprot(seg, a, pgsz)) {
6146 					return (EINVAL);
6147 				}
6148 			}
6149 		}
6150 	}
6151 
6152 	/*
6153 	 * Since we are changing page size we first have to flush
6154 	 * the cache. This makes sure all the pagelock calls have
6155 	 * to recheck protections.
6156 	 */
6157 	if (svd->softlockcnt > 0) {
6158 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6159 
6160 		/*
6161 		 * If this is shared segment non 0 softlockcnt
6162 		 * means locked pages are still in use.
6163 		 */
6164 		if (svd->type == MAP_SHARED) {
6165 			return (EAGAIN);
6166 		}
6167 
6168 		/*
6169 		 * Since we do have the segvn writers lock nobody can fill
6170 		 * the cache with entries belonging to this seg during
6171 		 * the purge. The flush either succeeds or we still have
6172 		 * pending I/Os.
6173 		 */
6174 		segvn_purge(seg);
6175 		if (svd->softlockcnt > 0) {
6176 			return (EAGAIN);
6177 		}
6178 	}
6179 
6180 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6181 		ASSERT(svd->amp == NULL);
6182 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6183 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6184 		    HAT_REGION_TEXT);
6185 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6186 	} else if (svd->tr_state == SEGVN_TR_INIT) {
6187 		svd->tr_state = SEGVN_TR_OFF;
6188 	} else if (svd->tr_state == SEGVN_TR_ON) {
6189 		ASSERT(svd->amp != NULL);
6190 		segvn_textunrepl(seg, 1);
6191 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6192 		amp = NULL;
6193 	}
6194 
6195 	/*
6196 	 * Operation for sub range of existing segment.
6197 	 */
6198 	if (addr != seg->s_base || eaddr != (seg->s_base + seg->s_size)) {
6199 		if (szc < seg->s_szc) {
6200 			VM_STAT_ADD(segvnvmstats.demoterange[2]);
6201 			err = segvn_demote_range(seg, addr, len, SDR_RANGE, 0);
6202 			if (err == 0) {
6203 				return (IE_RETRY);
6204 			}
6205 			if (err == ENOMEM) {
6206 				return (IE_NOMEM);
6207 			}
6208 			return (err);
6209 		}
6210 		if (addr != seg->s_base) {
6211 			nseg = segvn_split_seg(seg, addr);
6212 			if (eaddr != (nseg->s_base + nseg->s_size)) {
6213 				/* eaddr is szc aligned */
6214 				(void) segvn_split_seg(nseg, eaddr);
6215 			}
6216 			return (IE_RETRY);
6217 		}
6218 		if (eaddr != (seg->s_base + seg->s_size)) {
6219 			/* eaddr is szc aligned */
6220 			(void) segvn_split_seg(seg, eaddr);
6221 		}
6222 		return (IE_RETRY);
6223 	}
6224 
6225 	/*
6226 	 * Break any low level sharing and reset seg->s_szc to 0.
6227 	 */
6228 	if ((err = segvn_clrszc(seg)) != 0) {
6229 		if (err == ENOMEM) {
6230 			err = IE_NOMEM;
6231 		}
6232 		return (err);
6233 	}
6234 	ASSERT(seg->s_szc == 0);
6235 
6236 	/*
6237 	 * If the end of the current segment is not pgsz aligned
6238 	 * then attempt to concatenate with the next segment.
6239 	 */
6240 	if (!IS_P2ALIGNED(eaddr, pgsz)) {
6241 		nseg = AS_SEGNEXT(seg->s_as, seg);
6242 		if (nseg == NULL || nseg == seg || eaddr != nseg->s_base) {
6243 			return (ENOMEM);
6244 		}
6245 		if (nseg->s_ops != &segvn_ops) {
6246 			return (EINVAL);
6247 		}
6248 		nsvd = (struct segvn_data *)nseg->s_data;
6249 		if (nsvd->softlockcnt > 0) {
6250 			/*
6251 			 * If this is shared segment non 0 softlockcnt
6252 			 * means locked pages are still in use.
6253 			 */
6254 			if (nsvd->type == MAP_SHARED) {
6255 				return (EAGAIN);
6256 			}
6257 			segvn_purge(nseg);
6258 			if (nsvd->softlockcnt > 0) {
6259 				return (EAGAIN);
6260 			}
6261 		}
6262 		err = segvn_clrszc(nseg);
6263 		if (err == ENOMEM) {
6264 			err = IE_NOMEM;
6265 		}
6266 		if (err != 0) {
6267 			return (err);
6268 		}
6269 		ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6270 		err = segvn_concat(seg, nseg, 1);
6271 		if (err == -1) {
6272 			return (EINVAL);
6273 		}
6274 		if (err == -2) {
6275 			return (IE_NOMEM);
6276 		}
6277 		return (IE_RETRY);
6278 	}
6279 
6280 	/*
6281 	 * May need to re-align anon array to
6282 	 * new szc.
6283 	 */
6284 	if (amp != NULL) {
6285 		if (!IS_P2ALIGNED(svd->anon_index, pgcnt)) {
6286 			struct anon_hdr *nahp;
6287 
6288 			ASSERT(svd->type == MAP_PRIVATE);
6289 
6290 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6291 			ASSERT(amp->refcnt == 1);
6292 			nahp = anon_create(btop(amp->size), ANON_NOSLEEP);
6293 			if (nahp == NULL) {
6294 				ANON_LOCK_EXIT(&amp->a_rwlock);
6295 				return (IE_NOMEM);
6296 			}
6297 			if (anon_copy_ptr(amp->ahp, svd->anon_index,
6298 			    nahp, 0, btop(seg->s_size), ANON_NOSLEEP)) {
6299 				anon_release(nahp, btop(amp->size));
6300 				ANON_LOCK_EXIT(&amp->a_rwlock);
6301 				return (IE_NOMEM);
6302 			}
6303 			anon_release(amp->ahp, btop(amp->size));
6304 			amp->ahp = nahp;
6305 			svd->anon_index = 0;
6306 			ANON_LOCK_EXIT(&amp->a_rwlock);
6307 		}
6308 	}
6309 	if (svd->vp != NULL && szc != 0) {
6310 		struct vattr va;
6311 		u_offset_t eoffpage = svd->offset;
6312 		va.va_mask = AT_SIZE;
6313 		eoffpage += seg->s_size;
6314 		eoffpage = btopr(eoffpage);
6315 		if (VOP_GETATTR(svd->vp, &va, 0, svd->cred, NULL) != 0) {
6316 			segvn_setpgsz_getattr_err++;
6317 			return (EINVAL);
6318 		}
6319 		if (btopr(va.va_size) < eoffpage) {
6320 			segvn_setpgsz_eof_err++;
6321 			return (EINVAL);
6322 		}
6323 		if (amp != NULL) {
6324 			/*
6325 			 * anon_fill_cow_holes() may call VOP_GETPAGE().
6326 			 * don't take anon map lock here to avoid holding it
6327 			 * across VOP_GETPAGE() calls that may call back into
6328 			 * segvn for klsutering checks. We don't really need
6329 			 * anon map lock here since it's a private segment and
6330 			 * we hold as level lock as writers.
6331 			 */
6332 			if ((err = anon_fill_cow_holes(seg, seg->s_base,
6333 			    amp->ahp, svd->anon_index, svd->vp, svd->offset,
6334 			    seg->s_size, szc, svd->prot, svd->vpage,
6335 			    svd->cred)) != 0) {
6336 				return (EINVAL);
6337 			}
6338 		}
6339 		segvn_setvnode_mpss(svd->vp);
6340 	}
6341 
6342 	if (amp != NULL) {
6343 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6344 		if (svd->type == MAP_PRIVATE) {
6345 			amp->a_szc = szc;
6346 		} else if (szc > amp->a_szc) {
6347 			amp->a_szc = szc;
6348 		}
6349 		ANON_LOCK_EXIT(&amp->a_rwlock);
6350 	}
6351 
6352 	seg->s_szc = szc;
6353 
6354 	return (0);
6355 }
6356 
6357 static int
6358 segvn_clrszc(struct seg *seg)
6359 {
6360 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6361 	struct anon_map *amp = svd->amp;
6362 	size_t pgsz;
6363 	pgcnt_t pages;
6364 	int err = 0;
6365 	caddr_t a = seg->s_base;
6366 	caddr_t ea = a + seg->s_size;
6367 	ulong_t an_idx = svd->anon_index;
6368 	vnode_t *vp = svd->vp;
6369 	struct vpage *vpage = svd->vpage;
6370 	page_t *anon_pl[1 + 1], *pp;
6371 	struct anon *ap, *oldap;
6372 	uint_t prot = svd->prot, vpprot;
6373 	int pageflag = 0;
6374 
6375 	ASSERT(AS_WRITE_HELD(seg->s_as) ||
6376 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
6377 	ASSERT(svd->softlockcnt == 0);
6378 
6379 	if (vp == NULL && amp == NULL) {
6380 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6381 		seg->s_szc = 0;
6382 		return (0);
6383 	}
6384 
6385 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6386 		ASSERT(svd->amp == NULL);
6387 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6388 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6389 		    HAT_REGION_TEXT);
6390 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6391 	} else if (svd->tr_state == SEGVN_TR_ON) {
6392 		ASSERT(svd->amp != NULL);
6393 		segvn_textunrepl(seg, 1);
6394 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6395 		amp = NULL;
6396 	} else {
6397 		if (svd->tr_state != SEGVN_TR_OFF) {
6398 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
6399 			svd->tr_state = SEGVN_TR_OFF;
6400 		}
6401 
6402 		/*
6403 		 * do HAT_UNLOAD_UNMAP since we are changing the pagesize.
6404 		 * unload argument is 0 when we are freeing the segment
6405 		 * and unload was already done.
6406 		 */
6407 		hat_unload(seg->s_as->a_hat, seg->s_base, seg->s_size,
6408 		    HAT_UNLOAD_UNMAP);
6409 	}
6410 
6411 	if (amp == NULL || svd->type == MAP_SHARED) {
6412 		seg->s_szc = 0;
6413 		return (0);
6414 	}
6415 
6416 	pgsz = page_get_pagesize(seg->s_szc);
6417 	pages = btop(pgsz);
6418 
6419 	/*
6420 	 * XXX anon rwlock is not really needed because this is a
6421 	 * private segment and we are writers.
6422 	 */
6423 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6424 
6425 	for (; a < ea; a += pgsz, an_idx += pages) {
6426 		if ((oldap = anon_get_ptr(amp->ahp, an_idx)) != NULL) {
6427 			ASSERT(vpage != NULL || svd->pageprot == 0);
6428 			if (vpage != NULL) {
6429 				ASSERT(sameprot(seg, a, pgsz));
6430 				prot = VPP_PROT(vpage);
6431 				pageflag = VPP_ISPPLOCK(vpage) ? LOCK_PAGE : 0;
6432 			}
6433 			if (seg->s_szc != 0) {
6434 				ASSERT(vp == NULL || anon_pages(amp->ahp,
6435 				    an_idx, pages) == pages);
6436 				if ((err = anon_map_demotepages(amp, an_idx,
6437 				    seg, a, prot, vpage, svd->cred)) != 0) {
6438 					goto out;
6439 				}
6440 			} else {
6441 				if (oldap->an_refcnt == 1) {
6442 					continue;
6443 				}
6444 				if ((err = anon_getpage(&oldap, &vpprot,
6445 				    anon_pl, PAGESIZE, seg, a, S_READ,
6446 				    svd->cred))) {
6447 					goto out;
6448 				}
6449 				if ((pp = anon_private(&ap, seg, a, prot,
6450 				    anon_pl[0], pageflag, svd->cred)) == NULL) {
6451 					err = ENOMEM;
6452 					goto out;
6453 				}
6454 				anon_decref(oldap);
6455 				(void) anon_set_ptr(amp->ahp, an_idx, ap,
6456 				    ANON_SLEEP);
6457 				page_unlock(pp);
6458 			}
6459 		}
6460 		vpage = (vpage == NULL) ? NULL : vpage + pages;
6461 	}
6462 
6463 	amp->a_szc = 0;
6464 	seg->s_szc = 0;
6465 out:
6466 	ANON_LOCK_EXIT(&amp->a_rwlock);
6467 	return (err);
6468 }
6469 
6470 static int
6471 segvn_claim_pages(
6472 	struct seg *seg,
6473 	struct vpage *svp,
6474 	u_offset_t off,
6475 	ulong_t anon_idx,
6476 	uint_t prot)
6477 {
6478 	pgcnt_t	pgcnt = page_get_pagecnt(seg->s_szc);
6479 	size_t ppasize = (pgcnt + 1) * sizeof (page_t *);
6480 	page_t	**ppa;
6481 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6482 	struct anon_map *amp = svd->amp;
6483 	struct vpage *evp = svp + pgcnt;
6484 	caddr_t addr = ((uintptr_t)(svp - svd->vpage) << PAGESHIFT)
6485 	    + seg->s_base;
6486 	struct anon *ap;
6487 	struct vnode *vp = svd->vp;
6488 	page_t *pp;
6489 	pgcnt_t pg_idx, i;
6490 	int err = 0;
6491 	anoff_t aoff;
6492 	int anon = (amp != NULL) ? 1 : 0;
6493 
6494 	ASSERT(svd->type == MAP_PRIVATE);
6495 	ASSERT(svd->vpage != NULL);
6496 	ASSERT(seg->s_szc != 0);
6497 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
6498 	ASSERT(amp == NULL || IS_P2ALIGNED(anon_idx, pgcnt));
6499 	ASSERT(sameprot(seg, addr, pgcnt << PAGESHIFT));
6500 
6501 	if (VPP_PROT(svp) == prot)
6502 		return (1);
6503 	if (!((VPP_PROT(svp) ^ prot) & PROT_WRITE))
6504 		return (1);
6505 
6506 	ppa = kmem_alloc(ppasize, KM_SLEEP);
6507 	if (anon && vp != NULL) {
6508 		if (anon_get_ptr(amp->ahp, anon_idx) == NULL) {
6509 			anon = 0;
6510 			ASSERT(!anon_pages(amp->ahp, anon_idx, pgcnt));
6511 		}
6512 		ASSERT(!anon ||
6513 		    anon_pages(amp->ahp, anon_idx, pgcnt) == pgcnt);
6514 	}
6515 
6516 	for (*ppa = NULL, pg_idx = 0; svp < evp; svp++, anon_idx++) {
6517 		if (!VPP_ISPPLOCK(svp))
6518 			continue;
6519 		if (anon) {
6520 			ap = anon_get_ptr(amp->ahp, anon_idx);
6521 			if (ap == NULL) {
6522 				panic("segvn_claim_pages: no anon slot");
6523 			}
6524 			swap_xlate(ap, &vp, &aoff);
6525 			off = (u_offset_t)aoff;
6526 		}
6527 		ASSERT(vp != NULL);
6528 		if ((pp = page_lookup(vp,
6529 		    (u_offset_t)off, SE_SHARED)) == NULL) {
6530 			panic("segvn_claim_pages: no page");
6531 		}
6532 		ppa[pg_idx++] = pp;
6533 		off += PAGESIZE;
6534 	}
6535 
6536 	if (ppa[0] == NULL) {
6537 		kmem_free(ppa, ppasize);
6538 		return (1);
6539 	}
6540 
6541 	ASSERT(pg_idx <= pgcnt);
6542 	ppa[pg_idx] = NULL;
6543 
6544 
6545 	/* Find each large page within ppa, and adjust its claim */
6546 
6547 	/* Does ppa cover a single large page? */
6548 	if (ppa[0]->p_szc == seg->s_szc) {
6549 		if (prot & PROT_WRITE)
6550 			err = page_addclaim_pages(ppa);
6551 		else
6552 			err = page_subclaim_pages(ppa);
6553 	} else {
6554 		for (i = 0; ppa[i]; i += pgcnt) {
6555 			ASSERT(IS_P2ALIGNED(page_pptonum(ppa[i]), pgcnt));
6556 			if (prot & PROT_WRITE)
6557 				err = page_addclaim_pages(&ppa[i]);
6558 			else
6559 				err = page_subclaim_pages(&ppa[i]);
6560 			if (err == 0)
6561 				break;
6562 		}
6563 	}
6564 
6565 	for (i = 0; i < pg_idx; i++) {
6566 		ASSERT(ppa[i] != NULL);
6567 		page_unlock(ppa[i]);
6568 	}
6569 
6570 	kmem_free(ppa, ppasize);
6571 	return (err);
6572 }
6573 
6574 /*
6575  * Returns right (upper address) segment if split occurred.
6576  * If the address is equal to the beginning or end of its segment it returns
6577  * the current segment.
6578  */
6579 static struct seg *
6580 segvn_split_seg(struct seg *seg, caddr_t addr)
6581 {
6582 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6583 	struct seg *nseg;
6584 	size_t nsize;
6585 	struct segvn_data *nsvd;
6586 
6587 	ASSERT(AS_WRITE_HELD(seg->s_as));
6588 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6589 
6590 	ASSERT(addr >= seg->s_base);
6591 	ASSERT(addr <= seg->s_base + seg->s_size);
6592 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6593 
6594 	if (addr == seg->s_base || addr == seg->s_base + seg->s_size)
6595 		return (seg);
6596 
6597 	nsize = seg->s_base + seg->s_size - addr;
6598 	seg->s_size = addr - seg->s_base;
6599 	nseg = seg_alloc(seg->s_as, addr, nsize);
6600 	ASSERT(nseg != NULL);
6601 	nseg->s_ops = seg->s_ops;
6602 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
6603 	nseg->s_data = (void *)nsvd;
6604 	nseg->s_szc = seg->s_szc;
6605 	*nsvd = *svd;
6606 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6607 	nsvd->seg = nseg;
6608 	rw_init(&nsvd->lock, NULL, RW_DEFAULT, NULL);
6609 
6610 	if (nsvd->vp != NULL) {
6611 		VN_HOLD(nsvd->vp);
6612 		nsvd->offset = svd->offset +
6613 		    (uintptr_t)(nseg->s_base - seg->s_base);
6614 		if (nsvd->type == MAP_SHARED)
6615 			lgrp_shm_policy_init(NULL, nsvd->vp);
6616 	} else {
6617 		/*
6618 		 * The offset for an anonymous segment has no signifigance in
6619 		 * terms of an offset into a file. If we were to use the above
6620 		 * calculation instead, the structures read out of
6621 		 * /proc/<pid>/xmap would be more difficult to decipher since
6622 		 * it would be unclear whether two seemingly contiguous
6623 		 * prxmap_t structures represented different segments or a
6624 		 * single segment that had been split up into multiple prxmap_t
6625 		 * structures (e.g. if some part of the segment had not yet
6626 		 * been faulted in).
6627 		 */
6628 		nsvd->offset = 0;
6629 	}
6630 
6631 	ASSERT(svd->softlockcnt == 0);
6632 	ASSERT(svd->softlockcnt_sbase == 0);
6633 	ASSERT(svd->softlockcnt_send == 0);
6634 	crhold(svd->cred);
6635 
6636 	if (svd->vpage != NULL) {
6637 		size_t bytes = vpgtob(seg_pages(seg));
6638 		size_t nbytes = vpgtob(seg_pages(nseg));
6639 		struct vpage *ovpage = svd->vpage;
6640 
6641 		svd->vpage = kmem_alloc(bytes, KM_SLEEP);
6642 		bcopy(ovpage, svd->vpage, bytes);
6643 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
6644 		bcopy(ovpage + seg_pages(seg), nsvd->vpage, nbytes);
6645 		kmem_free(ovpage, bytes + nbytes);
6646 	}
6647 	if (svd->amp != NULL && svd->type == MAP_PRIVATE) {
6648 		struct anon_map *oamp = svd->amp, *namp;
6649 		struct anon_hdr *nahp;
6650 
6651 		ANON_LOCK_ENTER(&oamp->a_rwlock, RW_WRITER);
6652 		ASSERT(oamp->refcnt == 1);
6653 		nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
6654 		(void) anon_copy_ptr(oamp->ahp, svd->anon_index,
6655 		    nahp, 0, btop(seg->s_size), ANON_SLEEP);
6656 
6657 		namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
6658 		namp->a_szc = nseg->s_szc;
6659 		(void) anon_copy_ptr(oamp->ahp,
6660 		    svd->anon_index + btop(seg->s_size),
6661 		    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
6662 		anon_release(oamp->ahp, btop(oamp->size));
6663 		oamp->ahp = nahp;
6664 		oamp->size = seg->s_size;
6665 		svd->anon_index = 0;
6666 		nsvd->amp = namp;
6667 		nsvd->anon_index = 0;
6668 		ANON_LOCK_EXIT(&oamp->a_rwlock);
6669 	} else if (svd->amp != NULL) {
6670 		pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6671 		ASSERT(svd->amp == nsvd->amp);
6672 		ASSERT(seg->s_szc <= svd->amp->a_szc);
6673 		nsvd->anon_index = svd->anon_index + seg_pages(seg);
6674 		ASSERT(IS_P2ALIGNED(nsvd->anon_index, pgcnt));
6675 		ANON_LOCK_ENTER(&svd->amp->a_rwlock, RW_WRITER);
6676 		svd->amp->refcnt++;
6677 		ANON_LOCK_EXIT(&svd->amp->a_rwlock);
6678 	}
6679 
6680 	/*
6681 	 * Split the amount of swap reserved.
6682 	 */
6683 	if (svd->swresv) {
6684 		/*
6685 		 * For MAP_NORESERVE, only allocate swap reserve for pages
6686 		 * being used.  Other segments get enough to cover whole
6687 		 * segment.
6688 		 */
6689 		if (svd->flags & MAP_NORESERVE) {
6690 			size_t	oswresv;
6691 
6692 			ASSERT(svd->amp);
6693 			oswresv = svd->swresv;
6694 			svd->swresv = ptob(anon_pages(svd->amp->ahp,
6695 			    svd->anon_index, btop(seg->s_size)));
6696 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
6697 			    nsvd->anon_index, btop(nseg->s_size)));
6698 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
6699 		} else {
6700 			if (svd->pageswap) {
6701 				svd->swresv = segvn_count_swap_by_vpages(seg);
6702 				ASSERT(nsvd->swresv >= svd->swresv);
6703 				nsvd->swresv -= svd->swresv;
6704 			} else {
6705 				ASSERT(svd->swresv == seg->s_size +
6706 				    nseg->s_size);
6707 				svd->swresv = seg->s_size;
6708 				nsvd->swresv = nseg->s_size;
6709 			}
6710 		}
6711 	}
6712 
6713 	return (nseg);
6714 }
6715 
6716 /*
6717  * called on memory operations (unmap, setprot, setpagesize) for a subset
6718  * of a large page segment to either demote the memory range (SDR_RANGE)
6719  * or the ends (SDR_END) by addr/len.
6720  *
6721  * returns 0 on success. returns errno, including ENOMEM, on failure.
6722  */
6723 static int
6724 segvn_demote_range(
6725 	struct seg *seg,
6726 	caddr_t addr,
6727 	size_t len,
6728 	int flag,
6729 	uint_t szcvec)
6730 {
6731 	caddr_t eaddr = addr + len;
6732 	caddr_t lpgaddr, lpgeaddr;
6733 	struct seg *nseg;
6734 	struct seg *badseg1 = NULL;
6735 	struct seg *badseg2 = NULL;
6736 	size_t pgsz;
6737 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6738 	int err;
6739 	uint_t szc = seg->s_szc;
6740 	uint_t tszcvec;
6741 
6742 	ASSERT(AS_WRITE_HELD(seg->s_as));
6743 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6744 	ASSERT(szc != 0);
6745 	pgsz = page_get_pagesize(szc);
6746 	ASSERT(seg->s_base != addr || seg->s_size != len);
6747 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6748 	ASSERT(svd->softlockcnt == 0);
6749 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6750 	ASSERT(szcvec == 0 || (flag == SDR_END && svd->type == MAP_SHARED));
6751 
6752 	CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
6753 	ASSERT(flag == SDR_RANGE || eaddr < lpgeaddr || addr > lpgaddr);
6754 	if (flag == SDR_RANGE) {
6755 		/* demote entire range */
6756 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6757 		(void) segvn_split_seg(nseg, lpgeaddr);
6758 		ASSERT(badseg1->s_base == lpgaddr);
6759 		ASSERT(badseg1->s_size == lpgeaddr - lpgaddr);
6760 	} else if (addr != lpgaddr) {
6761 		ASSERT(flag == SDR_END);
6762 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6763 		if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz &&
6764 		    eaddr < lpgaddr + 2 * pgsz) {
6765 			(void) segvn_split_seg(nseg, lpgeaddr);
6766 			ASSERT(badseg1->s_base == lpgaddr);
6767 			ASSERT(badseg1->s_size == 2 * pgsz);
6768 		} else {
6769 			nseg = segvn_split_seg(nseg, lpgaddr + pgsz);
6770 			ASSERT(badseg1->s_base == lpgaddr);
6771 			ASSERT(badseg1->s_size == pgsz);
6772 			if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz) {
6773 				ASSERT(lpgeaddr - lpgaddr > 2 * pgsz);
6774 				nseg = segvn_split_seg(nseg, lpgeaddr - pgsz);
6775 				badseg2 = nseg;
6776 				(void) segvn_split_seg(nseg, lpgeaddr);
6777 				ASSERT(badseg2->s_base == lpgeaddr - pgsz);
6778 				ASSERT(badseg2->s_size == pgsz);
6779 			}
6780 		}
6781 	} else {
6782 		ASSERT(flag == SDR_END);
6783 		ASSERT(eaddr < lpgeaddr);
6784 		badseg1 = nseg = segvn_split_seg(seg, lpgeaddr - pgsz);
6785 		(void) segvn_split_seg(nseg, lpgeaddr);
6786 		ASSERT(badseg1->s_base == lpgeaddr - pgsz);
6787 		ASSERT(badseg1->s_size == pgsz);
6788 	}
6789 
6790 	ASSERT(badseg1 != NULL);
6791 	ASSERT(badseg1->s_szc == szc);
6792 	ASSERT(flag == SDR_RANGE || badseg1->s_size == pgsz ||
6793 	    badseg1->s_size == 2 * pgsz);
6794 	ASSERT(sameprot(badseg1, badseg1->s_base, pgsz));
6795 	ASSERT(badseg1->s_size == pgsz ||
6796 	    sameprot(badseg1, badseg1->s_base + pgsz, pgsz));
6797 	if (err = segvn_clrszc(badseg1)) {
6798 		return (err);
6799 	}
6800 	ASSERT(badseg1->s_szc == 0);
6801 
6802 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6803 		uint_t tszc = highbit(tszcvec) - 1;
6804 		caddr_t ta = MAX(addr, badseg1->s_base);
6805 		caddr_t te;
6806 		size_t tpgsz = page_get_pagesize(tszc);
6807 
6808 		ASSERT(svd->type == MAP_SHARED);
6809 		ASSERT(flag == SDR_END);
6810 		ASSERT(tszc < szc && tszc > 0);
6811 
6812 		if (eaddr > badseg1->s_base + badseg1->s_size) {
6813 			te = badseg1->s_base + badseg1->s_size;
6814 		} else {
6815 			te = eaddr;
6816 		}
6817 
6818 		ASSERT(ta <= te);
6819 		badseg1->s_szc = tszc;
6820 		if (!IS_P2ALIGNED(ta, tpgsz) || !IS_P2ALIGNED(te, tpgsz)) {
6821 			if (badseg2 != NULL) {
6822 				err = segvn_demote_range(badseg1, ta, te - ta,
6823 				    SDR_END, tszcvec);
6824 				if (err != 0) {
6825 					return (err);
6826 				}
6827 			} else {
6828 				return (segvn_demote_range(badseg1, ta,
6829 				    te - ta, SDR_END, tszcvec));
6830 			}
6831 		}
6832 	}
6833 
6834 	if (badseg2 == NULL)
6835 		return (0);
6836 	ASSERT(badseg2->s_szc == szc);
6837 	ASSERT(badseg2->s_size == pgsz);
6838 	ASSERT(sameprot(badseg2, badseg2->s_base, badseg2->s_size));
6839 	if (err = segvn_clrszc(badseg2)) {
6840 		return (err);
6841 	}
6842 	ASSERT(badseg2->s_szc == 0);
6843 
6844 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6845 		uint_t tszc = highbit(tszcvec) - 1;
6846 		size_t tpgsz = page_get_pagesize(tszc);
6847 
6848 		ASSERT(svd->type == MAP_SHARED);
6849 		ASSERT(flag == SDR_END);
6850 		ASSERT(tszc < szc && tszc > 0);
6851 		ASSERT(badseg2->s_base > addr);
6852 		ASSERT(eaddr > badseg2->s_base);
6853 		ASSERT(eaddr < badseg2->s_base + badseg2->s_size);
6854 
6855 		badseg2->s_szc = tszc;
6856 		if (!IS_P2ALIGNED(eaddr, tpgsz)) {
6857 			return (segvn_demote_range(badseg2, badseg2->s_base,
6858 			    eaddr - badseg2->s_base, SDR_END, tszcvec));
6859 		}
6860 	}
6861 
6862 	return (0);
6863 }
6864 
6865 static int
6866 segvn_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
6867 {
6868 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6869 	struct vpage *vp, *evp;
6870 
6871 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6872 
6873 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6874 	/*
6875 	 * If segment protection can be used, simply check against them.
6876 	 */
6877 	if (svd->pageprot == 0) {
6878 		int err;
6879 
6880 		err = ((svd->prot & prot) != prot) ? EACCES : 0;
6881 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6882 		return (err);
6883 	}
6884 
6885 	/*
6886 	 * Have to check down to the vpage level.
6887 	 */
6888 	evp = &svd->vpage[seg_page(seg, addr + len)];
6889 	for (vp = &svd->vpage[seg_page(seg, addr)]; vp < evp; vp++) {
6890 		if ((VPP_PROT(vp) & prot) != prot) {
6891 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6892 			return (EACCES);
6893 		}
6894 	}
6895 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6896 	return (0);
6897 }
6898 
6899 static int
6900 segvn_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
6901 {
6902 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6903 	size_t pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
6904 
6905 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6906 
6907 	if (pgno != 0) {
6908 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6909 		if (svd->pageprot == 0) {
6910 			do {
6911 				protv[--pgno] = svd->prot;
6912 			} while (pgno != 0);
6913 		} else {
6914 			size_t pgoff = seg_page(seg, addr);
6915 
6916 			do {
6917 				pgno--;
6918 				protv[pgno] = VPP_PROT(&svd->vpage[pgno+pgoff]);
6919 			} while (pgno != 0);
6920 		}
6921 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6922 	}
6923 	return (0);
6924 }
6925 
6926 static u_offset_t
6927 segvn_getoffset(struct seg *seg, caddr_t addr)
6928 {
6929 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6930 
6931 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6932 
6933 	return (svd->offset + (uintptr_t)(addr - seg->s_base));
6934 }
6935 
6936 /*ARGSUSED*/
6937 static int
6938 segvn_gettype(struct seg *seg, caddr_t addr)
6939 {
6940 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6941 
6942 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6943 
6944 	return (svd->type | (svd->flags & (MAP_NORESERVE | MAP_TEXT |
6945 	    MAP_INITDATA)));
6946 }
6947 
6948 /*ARGSUSED*/
6949 static int
6950 segvn_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
6951 {
6952 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6953 
6954 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6955 
6956 	*vpp = svd->vp;
6957 	return (0);
6958 }
6959 
6960 /*
6961  * Check to see if it makes sense to do kluster/read ahead to
6962  * addr + delta relative to the mapping at addr.  We assume here
6963  * that delta is a signed PAGESIZE'd multiple (which can be negative).
6964  *
6965  * For segvn, we currently "approve" of the action if we are
6966  * still in the segment and it maps from the same vp/off,
6967  * or if the advice stored in segvn_data or vpages allows it.
6968  * Currently, klustering is not allowed only if MADV_RANDOM is set.
6969  */
6970 static int
6971 segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
6972 {
6973 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6974 	struct anon *oap, *ap;
6975 	ssize_t pd;
6976 	size_t page;
6977 	struct vnode *vp1, *vp2;
6978 	u_offset_t off1, off2;
6979 	struct anon_map *amp;
6980 
6981 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6982 	ASSERT(AS_WRITE_HELD(seg->s_as) ||
6983 	    SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
6984 
6985 	if (addr + delta < seg->s_base ||
6986 	    addr + delta >= (seg->s_base + seg->s_size))
6987 		return (-1);		/* exceeded segment bounds */
6988 
6989 	pd = delta / (ssize_t)PAGESIZE;	/* divide to preserve sign bit */
6990 	page = seg_page(seg, addr);
6991 
6992 	/*
6993 	 * Check to see if either of the pages addr or addr + delta
6994 	 * have advice set that prevents klustering (if MADV_RANDOM advice
6995 	 * is set for entire segment, or MADV_SEQUENTIAL is set and delta
6996 	 * is negative).
6997 	 */
6998 	if (svd->advice == MADV_RANDOM ||
6999 	    svd->advice == MADV_SEQUENTIAL && delta < 0)
7000 		return (-1);
7001 	else if (svd->pageadvice && svd->vpage) {
7002 		struct vpage *bvpp, *evpp;
7003 
7004 		bvpp = &svd->vpage[page];
7005 		evpp = &svd->vpage[page + pd];
7006 		if (VPP_ADVICE(bvpp) == MADV_RANDOM ||
7007 		    VPP_ADVICE(evpp) == MADV_SEQUENTIAL && delta < 0)
7008 			return (-1);
7009 		if (VPP_ADVICE(bvpp) != VPP_ADVICE(evpp) &&
7010 		    VPP_ADVICE(evpp) == MADV_RANDOM)
7011 			return (-1);
7012 	}
7013 
7014 	if (svd->type == MAP_SHARED)
7015 		return (0);		/* shared mapping - all ok */
7016 
7017 	if ((amp = svd->amp) == NULL)
7018 		return (0);		/* off original vnode */
7019 
7020 	page += svd->anon_index;
7021 
7022 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7023 
7024 	oap = anon_get_ptr(amp->ahp, page);
7025 	ap = anon_get_ptr(amp->ahp, page + pd);
7026 
7027 	ANON_LOCK_EXIT(&amp->a_rwlock);
7028 
7029 	if ((oap == NULL && ap != NULL) || (oap != NULL && ap == NULL)) {
7030 		return (-1);		/* one with and one without an anon */
7031 	}
7032 
7033 	if (oap == NULL) {		/* implies that ap == NULL */
7034 		return (0);		/* off original vnode */
7035 	}
7036 
7037 	/*
7038 	 * Now we know we have two anon pointers - check to
7039 	 * see if they happen to be properly allocated.
7040 	 */
7041 
7042 	/*
7043 	 * XXX We cheat here and don't lock the anon slots. We can't because
7044 	 * we may have been called from the anon layer which might already
7045 	 * have locked them. We are holding a refcnt on the slots so they
7046 	 * can't disappear. The worst that will happen is we'll get the wrong
7047 	 * names (vp, off) for the slots and make a poor klustering decision.
7048 	 */
7049 	swap_xlate(ap, &vp1, &off1);
7050 	swap_xlate(oap, &vp2, &off2);
7051 
7052 
7053 	if (!VOP_CMP(vp1, vp2, NULL) || off1 - off2 != delta)
7054 		return (-1);
7055 	return (0);
7056 }
7057 
7058 /*
7059  * Swap the pages of seg out to secondary storage, returning the
7060  * number of bytes of storage freed.
7061  *
7062  * The basic idea is first to unload all translations and then to call
7063  * VOP_PUTPAGE() for all newly-unmapped pages, to push them out to the
7064  * swap device.  Pages to which other segments have mappings will remain
7065  * mapped and won't be swapped.  Our caller (as_swapout) has already
7066  * performed the unloading step.
7067  *
7068  * The value returned is intended to correlate well with the process's
7069  * memory requirements.  However, there are some caveats:
7070  * 1)	When given a shared segment as argument, this routine will
7071  *	only succeed in swapping out pages for the last sharer of the
7072  *	segment.  (Previous callers will only have decremented mapping
7073  *	reference counts.)
7074  * 2)	We assume that the hat layer maintains a large enough translation
7075  *	cache to capture process reference patterns.
7076  */
7077 static size_t
7078 segvn_swapout(struct seg *seg)
7079 {
7080 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7081 	struct anon_map *amp;
7082 	pgcnt_t pgcnt = 0;
7083 	pgcnt_t npages;
7084 	pgcnt_t page;
7085 	ulong_t anon_index;
7086 
7087 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7088 
7089 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7090 	/*
7091 	 * Find pages unmapped by our caller and force them
7092 	 * out to the virtual swap device.
7093 	 */
7094 	if ((amp = svd->amp) != NULL)
7095 		anon_index = svd->anon_index;
7096 	npages = seg->s_size >> PAGESHIFT;
7097 	for (page = 0; page < npages; page++) {
7098 		page_t *pp;
7099 		struct anon *ap;
7100 		struct vnode *vp;
7101 		u_offset_t off;
7102 		anon_sync_obj_t cookie;
7103 
7104 		/*
7105 		 * Obtain <vp, off> pair for the page, then look it up.
7106 		 *
7107 		 * Note that this code is willing to consider regular
7108 		 * pages as well as anon pages.  Is this appropriate here?
7109 		 */
7110 		ap = NULL;
7111 		if (amp != NULL) {
7112 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7113 			if (anon_array_try_enter(amp, anon_index + page,
7114 			    &cookie)) {
7115 				ANON_LOCK_EXIT(&amp->a_rwlock);
7116 				continue;
7117 			}
7118 			ap = anon_get_ptr(amp->ahp, anon_index + page);
7119 			if (ap != NULL) {
7120 				swap_xlate(ap, &vp, &off);
7121 			} else {
7122 				vp = svd->vp;
7123 				off = svd->offset + ptob(page);
7124 			}
7125 			anon_array_exit(&cookie);
7126 			ANON_LOCK_EXIT(&amp->a_rwlock);
7127 		} else {
7128 			vp = svd->vp;
7129 			off = svd->offset + ptob(page);
7130 		}
7131 		if (vp == NULL) {		/* untouched zfod page */
7132 			ASSERT(ap == NULL);
7133 			continue;
7134 		}
7135 
7136 		pp = page_lookup_nowait(vp, off, SE_SHARED);
7137 		if (pp == NULL)
7138 			continue;
7139 
7140 
7141 		/*
7142 		 * Examine the page to see whether it can be tossed out,
7143 		 * keeping track of how many we've found.
7144 		 */
7145 		if (!page_tryupgrade(pp)) {
7146 			/*
7147 			 * If the page has an i/o lock and no mappings,
7148 			 * it's very likely that the page is being
7149 			 * written out as a result of klustering.
7150 			 * Assume this is so and take credit for it here.
7151 			 */
7152 			if (!page_io_trylock(pp)) {
7153 				if (!hat_page_is_mapped(pp))
7154 					pgcnt++;
7155 			} else {
7156 				page_io_unlock(pp);
7157 			}
7158 			page_unlock(pp);
7159 			continue;
7160 		}
7161 		ASSERT(!page_iolock_assert(pp));
7162 
7163 
7164 		/*
7165 		 * Skip if page is locked or has mappings.
7166 		 * We don't need the page_struct_lock to look at lckcnt
7167 		 * and cowcnt because the page is exclusive locked.
7168 		 */
7169 		if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0 ||
7170 		    hat_page_is_mapped(pp)) {
7171 			page_unlock(pp);
7172 			continue;
7173 		}
7174 
7175 		/*
7176 		 * dispose skips large pages so try to demote first.
7177 		 */
7178 		if (pp->p_szc != 0 && !page_try_demote_pages(pp)) {
7179 			page_unlock(pp);
7180 			/*
7181 			 * XXX should skip the remaining page_t's of this
7182 			 * large page.
7183 			 */
7184 			continue;
7185 		}
7186 
7187 		ASSERT(pp->p_szc == 0);
7188 
7189 		/*
7190 		 * No longer mapped -- we can toss it out.  How
7191 		 * we do so depends on whether or not it's dirty.
7192 		 */
7193 		if (hat_ismod(pp) && pp->p_vnode) {
7194 			/*
7195 			 * We must clean the page before it can be
7196 			 * freed.  Setting B_FREE will cause pvn_done
7197 			 * to free the page when the i/o completes.
7198 			 * XXX:	This also causes it to be accounted
7199 			 *	as a pageout instead of a swap: need
7200 			 *	B_SWAPOUT bit to use instead of B_FREE.
7201 			 *
7202 			 * Hold the vnode before releasing the page lock
7203 			 * to prevent it from being freed and re-used by
7204 			 * some other thread.
7205 			 */
7206 			VN_HOLD(vp);
7207 			page_unlock(pp);
7208 
7209 			/*
7210 			 * Queue all i/o requests for the pageout thread
7211 			 * to avoid saturating the pageout devices.
7212 			 */
7213 			if (!queue_io_request(vp, off))
7214 				VN_RELE(vp);
7215 		} else {
7216 			/*
7217 			 * The page was clean, free it.
7218 			 *
7219 			 * XXX:	Can we ever encounter modified pages
7220 			 *	with no associated vnode here?
7221 			 */
7222 			ASSERT(pp->p_vnode != NULL);
7223 			/*LINTED: constant in conditional context*/
7224 			VN_DISPOSE(pp, B_FREE, 0, kcred);
7225 		}
7226 
7227 		/*
7228 		 * Credit now even if i/o is in progress.
7229 		 */
7230 		pgcnt++;
7231 	}
7232 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7233 
7234 	/*
7235 	 * Wakeup pageout to initiate i/o on all queued requests.
7236 	 */
7237 	cv_signal_pageout();
7238 	return (ptob(pgcnt));
7239 }
7240 
7241 /*
7242  * Synchronize primary storage cache with real object in virtual memory.
7243  *
7244  * XXX - Anonymous pages should not be sync'ed out at all.
7245  */
7246 static int
7247 segvn_sync(struct seg *seg, caddr_t addr, size_t len, int attr, uint_t flags)
7248 {
7249 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7250 	struct vpage *vpp;
7251 	page_t *pp;
7252 	u_offset_t offset;
7253 	struct vnode *vp;
7254 	u_offset_t off;
7255 	caddr_t eaddr;
7256 	int bflags;
7257 	int err = 0;
7258 	int segtype;
7259 	int pageprot;
7260 	int prot;
7261 	ulong_t anon_index;
7262 	struct anon_map *amp;
7263 	struct anon *ap;
7264 	anon_sync_obj_t cookie;
7265 
7266 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7267 
7268 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7269 
7270 	if (svd->softlockcnt > 0) {
7271 		/*
7272 		 * If this is shared segment non 0 softlockcnt
7273 		 * means locked pages are still in use.
7274 		 */
7275 		if (svd->type == MAP_SHARED) {
7276 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7277 			return (EAGAIN);
7278 		}
7279 
7280 		/*
7281 		 * flush all pages from seg cache
7282 		 * otherwise we may deadlock in swap_putpage
7283 		 * for B_INVAL page (4175402).
7284 		 *
7285 		 * Even if we grab segvn WRITER's lock
7286 		 * here, there might be another thread which could've
7287 		 * successfully performed lookup/insert just before
7288 		 * we acquired the lock here.  So, grabbing either
7289 		 * lock here is of not much use.  Until we devise
7290 		 * a strategy at upper layers to solve the
7291 		 * synchronization issues completely, we expect
7292 		 * applications to handle this appropriately.
7293 		 */
7294 		segvn_purge(seg);
7295 		if (svd->softlockcnt > 0) {
7296 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7297 			return (EAGAIN);
7298 		}
7299 	} else if (svd->type == MAP_SHARED && svd->amp != NULL &&
7300 	    svd->amp->a_softlockcnt > 0) {
7301 		/*
7302 		 * Try to purge this amp's entries from pcache. It will
7303 		 * succeed only if other segments that share the amp have no
7304 		 * outstanding softlock's.
7305 		 */
7306 		segvn_purge(seg);
7307 		if (svd->amp->a_softlockcnt > 0 || svd->softlockcnt > 0) {
7308 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7309 			return (EAGAIN);
7310 		}
7311 	}
7312 
7313 	vpp = svd->vpage;
7314 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7315 	bflags = ((flags & MS_ASYNC) ? B_ASYNC : 0) |
7316 	    ((flags & MS_INVALIDATE) ? B_INVAL : 0);
7317 
7318 	if (attr) {
7319 		pageprot = attr & ~(SHARED|PRIVATE);
7320 		segtype = (attr & SHARED) ? MAP_SHARED : MAP_PRIVATE;
7321 
7322 		/*
7323 		 * We are done if the segment types don't match
7324 		 * or if we have segment level protections and
7325 		 * they don't match.
7326 		 */
7327 		if (svd->type != segtype) {
7328 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7329 			return (0);
7330 		}
7331 		if (vpp == NULL) {
7332 			if (svd->prot != pageprot) {
7333 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7334 				return (0);
7335 			}
7336 			prot = svd->prot;
7337 		} else
7338 			vpp = &svd->vpage[seg_page(seg, addr)];
7339 
7340 	} else if (svd->vp && svd->amp == NULL &&
7341 	    (flags & MS_INVALIDATE) == 0) {
7342 
7343 		/*
7344 		 * No attributes, no anonymous pages and MS_INVALIDATE flag
7345 		 * is not on, just use one big request.
7346 		 */
7347 		err = VOP_PUTPAGE(svd->vp, (offset_t)offset, len,
7348 		    bflags, svd->cred, NULL);
7349 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7350 		return (err);
7351 	}
7352 
7353 	if ((amp = svd->amp) != NULL)
7354 		anon_index = svd->anon_index + seg_page(seg, addr);
7355 
7356 	for (eaddr = addr + len; addr < eaddr; addr += PAGESIZE) {
7357 		ap = NULL;
7358 		if (amp != NULL) {
7359 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7360 			anon_array_enter(amp, anon_index, &cookie);
7361 			ap = anon_get_ptr(amp->ahp, anon_index++);
7362 			if (ap != NULL) {
7363 				swap_xlate(ap, &vp, &off);
7364 			} else {
7365 				vp = svd->vp;
7366 				off = offset;
7367 			}
7368 			anon_array_exit(&cookie);
7369 			ANON_LOCK_EXIT(&amp->a_rwlock);
7370 		} else {
7371 			vp = svd->vp;
7372 			off = offset;
7373 		}
7374 		offset += PAGESIZE;
7375 
7376 		if (vp == NULL)		/* untouched zfod page */
7377 			continue;
7378 
7379 		if (attr) {
7380 			if (vpp) {
7381 				prot = VPP_PROT(vpp);
7382 				vpp++;
7383 			}
7384 			if (prot != pageprot) {
7385 				continue;
7386 			}
7387 		}
7388 
7389 		/*
7390 		 * See if any of these pages are locked --  if so, then we
7391 		 * will have to truncate an invalidate request at the first
7392 		 * locked one. We don't need the page_struct_lock to test
7393 		 * as this is only advisory; even if we acquire it someone
7394 		 * might race in and lock the page after we unlock and before
7395 		 * we do the PUTPAGE, then PUTPAGE simply does nothing.
7396 		 */
7397 		if (flags & MS_INVALIDATE) {
7398 			if ((pp = page_lookup(vp, off, SE_SHARED)) != NULL) {
7399 				if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
7400 					page_unlock(pp);
7401 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7402 					return (EBUSY);
7403 				}
7404 				if (ap != NULL && pp->p_szc != 0 &&
7405 				    page_tryupgrade(pp)) {
7406 					if (pp->p_lckcnt == 0 &&
7407 					    pp->p_cowcnt == 0) {
7408 						/*
7409 						 * swapfs VN_DISPOSE() won't
7410 						 * invalidate large pages.
7411 						 * Attempt to demote.
7412 						 * XXX can't help it if it
7413 						 * fails. But for swapfs
7414 						 * pages it is no big deal.
7415 						 */
7416 						(void) page_try_demote_pages(
7417 						    pp);
7418 					}
7419 				}
7420 				page_unlock(pp);
7421 			}
7422 		} else if (svd->type == MAP_SHARED && amp != NULL) {
7423 			/*
7424 			 * Avoid writing out to disk ISM's large pages
7425 			 * because segspt_free_pages() relies on NULL an_pvp
7426 			 * of anon slots of such pages.
7427 			 */
7428 
7429 			ASSERT(svd->vp == NULL);
7430 			/*
7431 			 * swapfs uses page_lookup_nowait if not freeing or
7432 			 * invalidating and skips a page if
7433 			 * page_lookup_nowait returns NULL.
7434 			 */
7435 			pp = page_lookup_nowait(vp, off, SE_SHARED);
7436 			if (pp == NULL) {
7437 				continue;
7438 			}
7439 			if (pp->p_szc != 0) {
7440 				page_unlock(pp);
7441 				continue;
7442 			}
7443 
7444 			/*
7445 			 * Note ISM pages are created large so (vp, off)'s
7446 			 * page cannot suddenly become large after we unlock
7447 			 * pp.
7448 			 */
7449 			page_unlock(pp);
7450 		}
7451 		/*
7452 		 * XXX - Should ultimately try to kluster
7453 		 * calls to VOP_PUTPAGE() for performance.
7454 		 */
7455 		VN_HOLD(vp);
7456 		err = VOP_PUTPAGE(vp, (offset_t)off, PAGESIZE,
7457 		    (bflags | (IS_SWAPFSVP(vp) ? B_PAGE_NOWAIT : 0)),
7458 		    svd->cred, NULL);
7459 
7460 		VN_RELE(vp);
7461 		if (err)
7462 			break;
7463 	}
7464 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7465 	return (err);
7466 }
7467 
7468 /*
7469  * Determine if we have data corresponding to pages in the
7470  * primary storage virtual memory cache (i.e., "in core").
7471  */
7472 static size_t
7473 segvn_incore(struct seg *seg, caddr_t addr, size_t len, char *vec)
7474 {
7475 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7476 	struct vnode *vp, *avp;
7477 	u_offset_t offset, aoffset;
7478 	size_t p, ep;
7479 	int ret;
7480 	struct vpage *vpp;
7481 	page_t *pp;
7482 	uint_t start;
7483 	struct anon_map *amp;		/* XXX - for locknest */
7484 	struct anon *ap;
7485 	uint_t attr;
7486 	anon_sync_obj_t cookie;
7487 
7488 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7489 
7490 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7491 	if (svd->amp == NULL && svd->vp == NULL) {
7492 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7493 		bzero(vec, btopr(len));
7494 		return (len);	/* no anonymous pages created yet */
7495 	}
7496 
7497 	p = seg_page(seg, addr);
7498 	ep = seg_page(seg, addr + len);
7499 	start = svd->vp ? SEG_PAGE_VNODEBACKED : 0;
7500 
7501 	amp = svd->amp;
7502 	for (; p < ep; p++, addr += PAGESIZE) {
7503 		vpp = (svd->vpage) ? &svd->vpage[p]: NULL;
7504 		ret = start;
7505 		ap = NULL;
7506 		avp = NULL;
7507 		/* Grab the vnode/offset for the anon slot */
7508 		if (amp != NULL) {
7509 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7510 			anon_array_enter(amp, svd->anon_index + p, &cookie);
7511 			ap = anon_get_ptr(amp->ahp, svd->anon_index + p);
7512 			if (ap != NULL) {
7513 				swap_xlate(ap, &avp, &aoffset);
7514 			}
7515 			anon_array_exit(&cookie);
7516 			ANON_LOCK_EXIT(&amp->a_rwlock);
7517 		}
7518 		if ((avp != NULL) && page_exists(avp, aoffset)) {
7519 			/* A page exists for the anon slot */
7520 			ret |= SEG_PAGE_INCORE;
7521 
7522 			/*
7523 			 * If page is mapped and writable
7524 			 */
7525 			attr = (uint_t)0;
7526 			if ((hat_getattr(seg->s_as->a_hat, addr,
7527 			    &attr) != -1) && (attr & PROT_WRITE)) {
7528 				ret |= SEG_PAGE_ANON;
7529 			}
7530 			/*
7531 			 * Don't get page_struct lock for lckcnt and cowcnt,
7532 			 * since this is purely advisory.
7533 			 */
7534 			if ((pp = page_lookup_nowait(avp, aoffset,
7535 			    SE_SHARED)) != NULL) {
7536 				if (pp->p_lckcnt)
7537 					ret |= SEG_PAGE_SOFTLOCK;
7538 				if (pp->p_cowcnt)
7539 					ret |= SEG_PAGE_HASCOW;
7540 				page_unlock(pp);
7541 			}
7542 		}
7543 
7544 		/* Gather vnode statistics */
7545 		vp = svd->vp;
7546 		offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7547 
7548 		if (vp != NULL) {
7549 			/*
7550 			 * Try to obtain a "shared" lock on the page
7551 			 * without blocking.  If this fails, determine
7552 			 * if the page is in memory.
7553 			 */
7554 			pp = page_lookup_nowait(vp, offset, SE_SHARED);
7555 			if ((pp == NULL) && (page_exists(vp, offset))) {
7556 				/* Page is incore, and is named */
7557 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7558 			}
7559 			/*
7560 			 * Don't get page_struct lock for lckcnt and cowcnt,
7561 			 * since this is purely advisory.
7562 			 */
7563 			if (pp != NULL) {
7564 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7565 				if (pp->p_lckcnt)
7566 					ret |= SEG_PAGE_SOFTLOCK;
7567 				if (pp->p_cowcnt)
7568 					ret |= SEG_PAGE_HASCOW;
7569 				page_unlock(pp);
7570 			}
7571 		}
7572 
7573 		/* Gather virtual page information */
7574 		if (vpp) {
7575 			if (VPP_ISPPLOCK(vpp))
7576 				ret |= SEG_PAGE_LOCKED;
7577 			vpp++;
7578 		}
7579 
7580 		*vec++ = (char)ret;
7581 	}
7582 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7583 	return (len);
7584 }
7585 
7586 /*
7587  * Statement for p_cowcnts/p_lckcnts.
7588  *
7589  * p_cowcnt is updated while mlock/munlocking MAP_PRIVATE and PROT_WRITE region
7590  * irrespective of the following factors or anything else:
7591  *
7592  *	(1) anon slots are populated or not
7593  *	(2) cow is broken or not
7594  *	(3) refcnt on ap is 1 or greater than 1
7595  *
7596  * If it's not MAP_PRIVATE and PROT_WRITE, p_lckcnt is updated during mlock
7597  * and munlock.
7598  *
7599  *
7600  * Handling p_cowcnts/p_lckcnts during copy-on-write fault:
7601  *
7602  *	if vpage has PROT_WRITE
7603  *		transfer cowcnt on the oldpage -> cowcnt on the newpage
7604  *	else
7605  *		transfer lckcnt on the oldpage -> lckcnt on the newpage
7606  *
7607  *	During copy-on-write, decrement p_cowcnt on the oldpage and increment
7608  *	p_cowcnt on the newpage *if* the corresponding vpage has PROT_WRITE.
7609  *
7610  *	We may also break COW if softlocking on read access in the physio case.
7611  *	In this case, vpage may not have PROT_WRITE. So, we need to decrement
7612  *	p_lckcnt on the oldpage and increment p_lckcnt on the newpage *if* the
7613  *	vpage doesn't have PROT_WRITE.
7614  *
7615  *
7616  * Handling p_cowcnts/p_lckcnts during mprotect on mlocked region:
7617  *
7618  * 	If a MAP_PRIVATE region loses PROT_WRITE, we decrement p_cowcnt and
7619  *	increment p_lckcnt by calling page_subclaim() which takes care of
7620  * 	availrmem accounting and p_lckcnt overflow.
7621  *
7622  *	If a MAP_PRIVATE region gains PROT_WRITE, we decrement p_lckcnt and
7623  *	increment p_cowcnt by calling page_addclaim() which takes care of
7624  *	availrmem availability and p_cowcnt overflow.
7625  */
7626 
7627 /*
7628  * Lock down (or unlock) pages mapped by this segment.
7629  *
7630  * XXX only creates PAGESIZE pages if anon slots are not initialized.
7631  * At fault time they will be relocated into larger pages.
7632  */
7633 static int
7634 segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
7635     int attr, int op, ulong_t *lockmap, size_t pos)
7636 {
7637 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7638 	struct vpage *vpp;
7639 	struct vpage *evp;
7640 	page_t *pp;
7641 	u_offset_t offset;
7642 	u_offset_t off;
7643 	int segtype;
7644 	int pageprot;
7645 	int claim;
7646 	struct vnode *vp;
7647 	ulong_t anon_index;
7648 	struct anon_map *amp;
7649 	struct anon *ap;
7650 	struct vattr va;
7651 	anon_sync_obj_t cookie;
7652 	struct kshmid *sp = NULL;
7653 	struct proc	*p = curproc;
7654 	kproject_t	*proj = NULL;
7655 	int chargeproc = 1;
7656 	size_t locked_bytes = 0;
7657 	size_t unlocked_bytes = 0;
7658 	int err = 0;
7659 
7660 	/*
7661 	 * Hold write lock on address space because may split or concatenate
7662 	 * segments
7663 	 */
7664 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7665 
7666 	/*
7667 	 * If this is a shm, use shm's project and zone, else use
7668 	 * project and zone of calling process
7669 	 */
7670 
7671 	/* Determine if this segment backs a sysV shm */
7672 	if (svd->amp != NULL && svd->amp->a_sp != NULL) {
7673 		ASSERT(svd->type == MAP_SHARED);
7674 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
7675 		sp = svd->amp->a_sp;
7676 		proj = sp->shm_perm.ipc_proj;
7677 		chargeproc = 0;
7678 	}
7679 
7680 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7681 	if (attr) {
7682 		pageprot = attr & ~(SHARED|PRIVATE);
7683 		segtype = attr & SHARED ? MAP_SHARED : MAP_PRIVATE;
7684 
7685 		/*
7686 		 * We are done if the segment types don't match
7687 		 * or if we have segment level protections and
7688 		 * they don't match.
7689 		 */
7690 		if (svd->type != segtype) {
7691 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7692 			return (0);
7693 		}
7694 		if (svd->pageprot == 0 && svd->prot != pageprot) {
7695 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7696 			return (0);
7697 		}
7698 	}
7699 
7700 	if (op == MC_LOCK) {
7701 		if (svd->tr_state == SEGVN_TR_INIT) {
7702 			svd->tr_state = SEGVN_TR_OFF;
7703 		} else if (svd->tr_state == SEGVN_TR_ON) {
7704 			ASSERT(svd->amp != NULL);
7705 			segvn_textunrepl(seg, 0);
7706 			ASSERT(svd->amp == NULL &&
7707 			    svd->tr_state == SEGVN_TR_OFF);
7708 		}
7709 	}
7710 
7711 	/*
7712 	 * If we're locking, then we must create a vpage structure if
7713 	 * none exists.  If we're unlocking, then check to see if there
7714 	 * is a vpage --  if not, then we could not have locked anything.
7715 	 */
7716 
7717 	if ((vpp = svd->vpage) == NULL) {
7718 		if (op == MC_LOCK) {
7719 			segvn_vpage(seg);
7720 			if (svd->vpage == NULL) {
7721 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7722 				return (ENOMEM);
7723 			}
7724 		} else {
7725 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7726 			return (0);
7727 		}
7728 	}
7729 
7730 	/*
7731 	 * The anonymous data vector (i.e., previously
7732 	 * unreferenced mapping to swap space) can be allocated
7733 	 * by lazily testing for its existence.
7734 	 */
7735 	if (op == MC_LOCK && svd->amp == NULL && svd->vp == NULL) {
7736 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
7737 		svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
7738 		svd->amp->a_szc = seg->s_szc;
7739 	}
7740 
7741 	if ((amp = svd->amp) != NULL) {
7742 		anon_index = svd->anon_index + seg_page(seg, addr);
7743 	}
7744 
7745 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7746 	evp = &svd->vpage[seg_page(seg, addr + len)];
7747 
7748 	if (sp != NULL)
7749 		mutex_enter(&sp->shm_mlock);
7750 
7751 	/* determine number of unlocked bytes in range for lock operation */
7752 	if (op == MC_LOCK) {
7753 
7754 		if (sp == NULL) {
7755 			for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7756 			    vpp++) {
7757 				if (!VPP_ISPPLOCK(vpp))
7758 					unlocked_bytes += PAGESIZE;
7759 			}
7760 		} else {
7761 			ulong_t		i_idx, i_edx;
7762 			anon_sync_obj_t	i_cookie;
7763 			struct anon	*i_ap;
7764 			struct vnode	*i_vp;
7765 			u_offset_t	i_off;
7766 
7767 			/* Only count sysV pages once for locked memory */
7768 			i_edx = svd->anon_index + seg_page(seg, addr + len);
7769 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7770 			for (i_idx = anon_index; i_idx < i_edx; i_idx++) {
7771 				anon_array_enter(amp, i_idx, &i_cookie);
7772 				i_ap = anon_get_ptr(amp->ahp, i_idx);
7773 				if (i_ap == NULL) {
7774 					unlocked_bytes += PAGESIZE;
7775 					anon_array_exit(&i_cookie);
7776 					continue;
7777 				}
7778 				swap_xlate(i_ap, &i_vp, &i_off);
7779 				anon_array_exit(&i_cookie);
7780 				pp = page_lookup(i_vp, i_off, SE_SHARED);
7781 				if (pp == NULL) {
7782 					unlocked_bytes += PAGESIZE;
7783 					continue;
7784 				} else if (pp->p_lckcnt == 0)
7785 					unlocked_bytes += PAGESIZE;
7786 				page_unlock(pp);
7787 			}
7788 			ANON_LOCK_EXIT(&amp->a_rwlock);
7789 		}
7790 
7791 		mutex_enter(&p->p_lock);
7792 		err = rctl_incr_locked_mem(p, proj, unlocked_bytes,
7793 		    chargeproc);
7794 		mutex_exit(&p->p_lock);
7795 
7796 		if (err) {
7797 			if (sp != NULL)
7798 				mutex_exit(&sp->shm_mlock);
7799 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7800 			return (err);
7801 		}
7802 	}
7803 	/*
7804 	 * Loop over all pages in the range.  Process if we're locking and
7805 	 * page has not already been locked in this mapping; or if we're
7806 	 * unlocking and the page has been locked.
7807 	 */
7808 	for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7809 	    vpp++, pos++, addr += PAGESIZE, offset += PAGESIZE, anon_index++) {
7810 		if ((attr == 0 || VPP_PROT(vpp) == pageprot) &&
7811 		    ((op == MC_LOCK && !VPP_ISPPLOCK(vpp)) ||
7812 		    (op == MC_UNLOCK && VPP_ISPPLOCK(vpp)))) {
7813 
7814 			if (amp != NULL)
7815 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7816 			/*
7817 			 * If this isn't a MAP_NORESERVE segment and
7818 			 * we're locking, allocate anon slots if they
7819 			 * don't exist.  The page is brought in later on.
7820 			 */
7821 			if (op == MC_LOCK && svd->vp == NULL &&
7822 			    ((svd->flags & MAP_NORESERVE) == 0) &&
7823 			    amp != NULL &&
7824 			    ((ap = anon_get_ptr(amp->ahp, anon_index))
7825 			    == NULL)) {
7826 				anon_array_enter(amp, anon_index, &cookie);
7827 
7828 				if ((ap = anon_get_ptr(amp->ahp,
7829 				    anon_index)) == NULL) {
7830 					pp = anon_zero(seg, addr, &ap,
7831 					    svd->cred);
7832 					if (pp == NULL) {
7833 						anon_array_exit(&cookie);
7834 						ANON_LOCK_EXIT(&amp->a_rwlock);
7835 						err = ENOMEM;
7836 						goto out;
7837 					}
7838 					ASSERT(anon_get_ptr(amp->ahp,
7839 					    anon_index) == NULL);
7840 					(void) anon_set_ptr(amp->ahp,
7841 					    anon_index, ap, ANON_SLEEP);
7842 					page_unlock(pp);
7843 				}
7844 				anon_array_exit(&cookie);
7845 			}
7846 
7847 			/*
7848 			 * Get name for page, accounting for
7849 			 * existence of private copy.
7850 			 */
7851 			ap = NULL;
7852 			if (amp != NULL) {
7853 				anon_array_enter(amp, anon_index, &cookie);
7854 				ap = anon_get_ptr(amp->ahp, anon_index);
7855 				if (ap != NULL) {
7856 					swap_xlate(ap, &vp, &off);
7857 				} else {
7858 					if (svd->vp == NULL &&
7859 					    (svd->flags & MAP_NORESERVE)) {
7860 						anon_array_exit(&cookie);
7861 						ANON_LOCK_EXIT(&amp->a_rwlock);
7862 						continue;
7863 					}
7864 					vp = svd->vp;
7865 					off = offset;
7866 				}
7867 				if (op != MC_LOCK || ap == NULL) {
7868 					anon_array_exit(&cookie);
7869 					ANON_LOCK_EXIT(&amp->a_rwlock);
7870 				}
7871 			} else {
7872 				vp = svd->vp;
7873 				off = offset;
7874 			}
7875 
7876 			/*
7877 			 * Get page frame.  It's ok if the page is
7878 			 * not available when we're unlocking, as this
7879 			 * may simply mean that a page we locked got
7880 			 * truncated out of existence after we locked it.
7881 			 *
7882 			 * Invoke VOP_GETPAGE() to obtain the page struct
7883 			 * since we may need to read it from disk if its
7884 			 * been paged out.
7885 			 */
7886 			if (op != MC_LOCK)
7887 				pp = page_lookup(vp, off, SE_SHARED);
7888 			else {
7889 				page_t *pl[1 + 1];
7890 				int error;
7891 
7892 				ASSERT(vp != NULL);
7893 
7894 				error = VOP_GETPAGE(vp, (offset_t)off, PAGESIZE,
7895 				    (uint_t *)NULL, pl, PAGESIZE, seg, addr,
7896 				    S_OTHER, svd->cred, NULL);
7897 
7898 				if (error && ap != NULL) {
7899 					anon_array_exit(&cookie);
7900 					ANON_LOCK_EXIT(&amp->a_rwlock);
7901 				}
7902 
7903 				/*
7904 				 * If the error is EDEADLK then we must bounce
7905 				 * up and drop all vm subsystem locks and then
7906 				 * retry the operation later
7907 				 * This behavior is a temporary measure because
7908 				 * ufs/sds logging is badly designed and will
7909 				 * deadlock if we don't allow this bounce to
7910 				 * happen.  The real solution is to re-design
7911 				 * the logging code to work properly.  See bug
7912 				 * 4125102 for details of the problem.
7913 				 */
7914 				if (error == EDEADLK) {
7915 					err = error;
7916 					goto out;
7917 				}
7918 				/*
7919 				 * Quit if we fail to fault in the page.  Treat
7920 				 * the failure as an error, unless the addr
7921 				 * is mapped beyond the end of a file.
7922 				 */
7923 				if (error && svd->vp) {
7924 					va.va_mask = AT_SIZE;
7925 					if (VOP_GETATTR(svd->vp, &va, 0,
7926 					    svd->cred, NULL) != 0) {
7927 						err = EIO;
7928 						goto out;
7929 					}
7930 					if (btopr(va.va_size) >=
7931 					    btopr(off + 1)) {
7932 						err = EIO;
7933 						goto out;
7934 					}
7935 					goto out;
7936 
7937 				} else if (error) {
7938 					err = EIO;
7939 					goto out;
7940 				}
7941 				pp = pl[0];
7942 				ASSERT(pp != NULL);
7943 			}
7944 
7945 			/*
7946 			 * See Statement at the beginning of this routine.
7947 			 *
7948 			 * claim is always set if MAP_PRIVATE and PROT_WRITE
7949 			 * irrespective of following factors:
7950 			 *
7951 			 * (1) anon slots are populated or not
7952 			 * (2) cow is broken or not
7953 			 * (3) refcnt on ap is 1 or greater than 1
7954 			 *
7955 			 * See 4140683 for details
7956 			 */
7957 			claim = ((VPP_PROT(vpp) & PROT_WRITE) &&
7958 			    (svd->type == MAP_PRIVATE));
7959 
7960 			/*
7961 			 * Perform page-level operation appropriate to
7962 			 * operation.  If locking, undo the SOFTLOCK
7963 			 * performed to bring the page into memory
7964 			 * after setting the lock.  If unlocking,
7965 			 * and no page was found, account for the claim
7966 			 * separately.
7967 			 */
7968 			if (op == MC_LOCK) {
7969 				int ret = 1;	/* Assume success */
7970 
7971 				ASSERT(!VPP_ISPPLOCK(vpp));
7972 
7973 				ret = page_pp_lock(pp, claim, 0);
7974 				if (ap != NULL) {
7975 					if (ap->an_pvp != NULL) {
7976 						anon_swap_free(ap, pp);
7977 					}
7978 					anon_array_exit(&cookie);
7979 					ANON_LOCK_EXIT(&amp->a_rwlock);
7980 				}
7981 				if (ret == 0) {
7982 					/* locking page failed */
7983 					page_unlock(pp);
7984 					err = EAGAIN;
7985 					goto out;
7986 				}
7987 				VPP_SETPPLOCK(vpp);
7988 				if (sp != NULL) {
7989 					if (pp->p_lckcnt == 1)
7990 						locked_bytes += PAGESIZE;
7991 				} else
7992 					locked_bytes += PAGESIZE;
7993 
7994 				if (lockmap != (ulong_t *)NULL)
7995 					BT_SET(lockmap, pos);
7996 
7997 				page_unlock(pp);
7998 			} else {
7999 				ASSERT(VPP_ISPPLOCK(vpp));
8000 				if (pp != NULL) {
8001 					/* sysV pages should be locked */
8002 					ASSERT(sp == NULL || pp->p_lckcnt > 0);
8003 					page_pp_unlock(pp, claim, 0);
8004 					if (sp != NULL) {
8005 						if (pp->p_lckcnt == 0)
8006 							unlocked_bytes
8007 							    += PAGESIZE;
8008 					} else
8009 						unlocked_bytes += PAGESIZE;
8010 					page_unlock(pp);
8011 				} else {
8012 					ASSERT(sp == NULL);
8013 					unlocked_bytes += PAGESIZE;
8014 				}
8015 				VPP_CLRPPLOCK(vpp);
8016 			}
8017 		}
8018 	}
8019 out:
8020 	if (op == MC_LOCK) {
8021 		/* Credit back bytes that did not get locked */
8022 		if ((unlocked_bytes - locked_bytes) > 0) {
8023 			if (proj == NULL)
8024 				mutex_enter(&p->p_lock);
8025 			rctl_decr_locked_mem(p, proj,
8026 			    (unlocked_bytes - locked_bytes), chargeproc);
8027 			if (proj == NULL)
8028 				mutex_exit(&p->p_lock);
8029 		}
8030 
8031 	} else {
8032 		/* Account bytes that were unlocked */
8033 		if (unlocked_bytes > 0) {
8034 			if (proj == NULL)
8035 				mutex_enter(&p->p_lock);
8036 			rctl_decr_locked_mem(p, proj, unlocked_bytes,
8037 			    chargeproc);
8038 			if (proj == NULL)
8039 				mutex_exit(&p->p_lock);
8040 		}
8041 	}
8042 	if (sp != NULL)
8043 		mutex_exit(&sp->shm_mlock);
8044 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8045 
8046 	return (err);
8047 }
8048 
8049 /*
8050  * Set advice from user for specified pages
8051  * There are 10 types of advice:
8052  *	MADV_NORMAL	- Normal (default) behavior (whatever that is)
8053  *	MADV_RANDOM	- Random page references
8054  *				do not allow readahead or 'klustering'
8055  *	MADV_SEQUENTIAL	- Sequential page references
8056  *				Pages previous to the one currently being
8057  *				accessed (determined by fault) are 'not needed'
8058  *				and are freed immediately
8059  *	MADV_WILLNEED	- Pages are likely to be used (fault ahead in mctl)
8060  *	MADV_DONTNEED	- Pages are not needed (synced out in mctl)
8061  *	MADV_FREE	- Contents can be discarded
8062  *	MADV_ACCESS_DEFAULT- Default access
8063  *	MADV_ACCESS_LWP	- Next LWP will access heavily
8064  *	MADV_ACCESS_MANY- Many LWPs or processes will access heavily
8065  *	MADV_PURGE	- Contents will be immediately discarded
8066  */
8067 static int
8068 segvn_advise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
8069 {
8070 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8071 	size_t page;
8072 	int err = 0;
8073 	int already_set;
8074 	struct anon_map *amp;
8075 	ulong_t anon_index;
8076 	struct seg *next;
8077 	lgrp_mem_policy_t policy;
8078 	struct seg *prev;
8079 	struct vnode *vp;
8080 
8081 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
8082 
8083 	/*
8084 	 * In case of MADV_FREE/MADV_PURGE, we won't be modifying any segment
8085 	 * private data structures; so, we only need to grab READER's lock
8086 	 */
8087 	if (behav != MADV_FREE && behav != MADV_PURGE) {
8088 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
8089 		if (svd->tr_state != SEGVN_TR_OFF) {
8090 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8091 			return (0);
8092 		}
8093 	} else {
8094 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8095 	}
8096 
8097 	/*
8098 	 * Large pages are assumed to be only turned on when accesses to the
8099 	 * segment's address range have spatial and temporal locality. That
8100 	 * justifies ignoring MADV_SEQUENTIAL for large page segments.
8101 	 * Also, ignore advice affecting lgroup memory allocation
8102 	 * if don't need to do lgroup optimizations on this system
8103 	 */
8104 
8105 	if ((behav == MADV_SEQUENTIAL &&
8106 	    (seg->s_szc != 0 || HAT_IS_REGION_COOKIE_VALID(svd->rcookie))) ||
8107 	    (!lgrp_optimizations() && (behav == MADV_ACCESS_DEFAULT ||
8108 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY))) {
8109 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8110 		return (0);
8111 	}
8112 
8113 	if (behav == MADV_SEQUENTIAL || behav == MADV_ACCESS_DEFAULT ||
8114 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY) {
8115 		/*
8116 		 * Since we are going to unload hat mappings
8117 		 * we first have to flush the cache. Otherwise
8118 		 * this might lead to system panic if another
8119 		 * thread is doing physio on the range whose
8120 		 * mappings are unloaded by madvise(3C).
8121 		 */
8122 		if (svd->softlockcnt > 0) {
8123 			/*
8124 			 * If this is shared segment non 0 softlockcnt
8125 			 * means locked pages are still in use.
8126 			 */
8127 			if (svd->type == MAP_SHARED) {
8128 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8129 				return (EAGAIN);
8130 			}
8131 			/*
8132 			 * Since we do have the segvn writers lock
8133 			 * nobody can fill the cache with entries
8134 			 * belonging to this seg during the purge.
8135 			 * The flush either succeeds or we still
8136 			 * have pending I/Os. In the later case,
8137 			 * madvise(3C) fails.
8138 			 */
8139 			segvn_purge(seg);
8140 			if (svd->softlockcnt > 0) {
8141 				/*
8142 				 * Since madvise(3C) is advisory and
8143 				 * it's not part of UNIX98, madvise(3C)
8144 				 * failure here doesn't cause any hardship.
8145 				 * Note that we don't block in "as" layer.
8146 				 */
8147 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8148 				return (EAGAIN);
8149 			}
8150 		} else if (svd->type == MAP_SHARED && svd->amp != NULL &&
8151 		    svd->amp->a_softlockcnt > 0) {
8152 			/*
8153 			 * Try to purge this amp's entries from pcache. It
8154 			 * will succeed only if other segments that share the
8155 			 * amp have no outstanding softlock's.
8156 			 */
8157 			segvn_purge(seg);
8158 		}
8159 	}
8160 
8161 	amp = svd->amp;
8162 	vp = svd->vp;
8163 	if (behav == MADV_FREE || behav == MADV_PURGE) {
8164 		pgcnt_t purged;
8165 
8166 		if (behav == MADV_FREE && (vp != NULL || amp == NULL)) {
8167 			/*
8168 			 * MADV_FREE is not supported for segments with an
8169 			 * underlying object; if anonmap is NULL, anon slots
8170 			 * are not yet populated and there is nothing for us
8171 			 * to do. As MADV_FREE is advisory, we don't return an
8172 			 * error in either case.
8173 			 */
8174 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8175 			return (0);
8176 		}
8177 
8178 		if (amp == NULL) {
8179 			/*
8180 			 * If we're here with a NULL anonmap, it's because we
8181 			 * are doing a MADV_PURGE.  We have nothing to do, but
8182 			 * because MADV_PURGE isn't merely advisory, we return
8183 			 * an error in this case.
8184 			 */
8185 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8186 			return (EBUSY);
8187 		}
8188 
8189 		segvn_purge(seg);
8190 
8191 		page = seg_page(seg, addr);
8192 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8193 		err = anon_disclaim(amp,
8194 		    svd->anon_index + page, len, behav, &purged);
8195 
8196 		if (purged != 0 && (svd->flags & MAP_NORESERVE)) {
8197 			/*
8198 			 * If we purged pages on a MAP_NORESERVE mapping, we
8199 			 * need to be sure to now unreserve our reserved swap.
8200 			 * (We use the atomic operations to manipulate our
8201 			 * segment and address space counters because we only
8202 			 * have the corresponding locks held as reader, not
8203 			 * writer.)
8204 			 */
8205 			ssize_t bytes = ptob(purged);
8206 
8207 			anon_unresv_zone(bytes, seg->s_as->a_proc->p_zone);
8208 			atomic_add_long(&svd->swresv, -bytes);
8209 			atomic_add_long(&seg->s_as->a_resvsize, -bytes);
8210 		}
8211 
8212 		ANON_LOCK_EXIT(&amp->a_rwlock);
8213 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8214 
8215 		/*
8216 		 * MADV_PURGE and MADV_FREE differ in their return semantics:
8217 		 * because MADV_PURGE is designed to be bug-for-bug compatible
8218 		 * with its clumsy Linux forebear, it will fail where MADV_FREE
8219 		 * does not.
8220 		 */
8221 		return (behav == MADV_PURGE ? err : 0);
8222 	}
8223 
8224 	/*
8225 	 * If advice is to be applied to entire segment,
8226 	 * use advice field in seg_data structure
8227 	 * otherwise use appropriate vpage entry.
8228 	 */
8229 	if ((addr == seg->s_base) && (len == seg->s_size)) {
8230 		switch (behav) {
8231 		case MADV_ACCESS_LWP:
8232 		case MADV_ACCESS_MANY:
8233 		case MADV_ACCESS_DEFAULT:
8234 			/*
8235 			 * Set memory allocation policy for this segment
8236 			 */
8237 			policy = lgrp_madv_to_policy(behav, len, svd->type);
8238 			if (svd->type == MAP_SHARED)
8239 				already_set = lgrp_shm_policy_set(policy, amp,
8240 				    svd->anon_index, vp, svd->offset, len);
8241 			else {
8242 				/*
8243 				 * For private memory, need writers lock on
8244 				 * address space because the segment may be
8245 				 * split or concatenated when changing policy
8246 				 */
8247 				if (AS_READ_HELD(seg->s_as)) {
8248 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8249 					return (IE_RETRY);
8250 				}
8251 
8252 				already_set = lgrp_privm_policy_set(policy,
8253 				    &svd->policy_info, len);
8254 			}
8255 
8256 			/*
8257 			 * If policy set already and it shouldn't be reapplied,
8258 			 * don't do anything.
8259 			 */
8260 			if (already_set &&
8261 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8262 				break;
8263 
8264 			/*
8265 			 * Mark any existing pages in given range for
8266 			 * migration
8267 			 */
8268 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8269 			    vp, svd->offset, 1);
8270 
8271 			/*
8272 			 * If same policy set already or this is a shared
8273 			 * memory segment, don't need to try to concatenate
8274 			 * segment with adjacent ones.
8275 			 */
8276 			if (already_set || svd->type == MAP_SHARED)
8277 				break;
8278 
8279 			/*
8280 			 * Try to concatenate this segment with previous
8281 			 * one and next one, since we changed policy for
8282 			 * this one and it may be compatible with adjacent
8283 			 * ones now.
8284 			 */
8285 			prev = AS_SEGPREV(seg->s_as, seg);
8286 			next = AS_SEGNEXT(seg->s_as, seg);
8287 
8288 			if (next && next->s_ops == &segvn_ops &&
8289 			    addr + len == next->s_base)
8290 				(void) segvn_concat(seg, next, 1);
8291 
8292 			if (prev && prev->s_ops == &segvn_ops &&
8293 			    addr == prev->s_base + prev->s_size) {
8294 				/*
8295 				 * Drop lock for private data of current
8296 				 * segment before concatenating (deleting) it
8297 				 * and return IE_REATTACH to tell as_ctl() that
8298 				 * current segment has changed
8299 				 */
8300 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8301 				if (!segvn_concat(prev, seg, 1))
8302 					err = IE_REATTACH;
8303 
8304 				return (err);
8305 			}
8306 			break;
8307 
8308 		case MADV_SEQUENTIAL:
8309 			/*
8310 			 * unloading mapping guarantees
8311 			 * detection in segvn_fault
8312 			 */
8313 			ASSERT(seg->s_szc == 0);
8314 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8315 			hat_unload(seg->s_as->a_hat, addr, len,
8316 			    HAT_UNLOAD);
8317 			/* FALLTHROUGH */
8318 		case MADV_NORMAL:
8319 		case MADV_RANDOM:
8320 			svd->advice = (uchar_t)behav;
8321 			svd->pageadvice = 0;
8322 			break;
8323 		case MADV_WILLNEED:	/* handled in memcntl */
8324 		case MADV_DONTNEED:	/* handled in memcntl */
8325 		case MADV_FREE:		/* handled above */
8326 		case MADV_PURGE:	/* handled above */
8327 			break;
8328 		default:
8329 			err = EINVAL;
8330 		}
8331 	} else {
8332 		caddr_t			eaddr;
8333 		struct seg		*new_seg;
8334 		struct segvn_data	*new_svd;
8335 		u_offset_t		off;
8336 		caddr_t			oldeaddr;
8337 
8338 		page = seg_page(seg, addr);
8339 
8340 		segvn_vpage(seg);
8341 		if (svd->vpage == NULL) {
8342 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8343 			return (ENOMEM);
8344 		}
8345 
8346 		switch (behav) {
8347 			struct vpage *bvpp, *evpp;
8348 
8349 		case MADV_ACCESS_LWP:
8350 		case MADV_ACCESS_MANY:
8351 		case MADV_ACCESS_DEFAULT:
8352 			/*
8353 			 * Set memory allocation policy for portion of this
8354 			 * segment
8355 			 */
8356 
8357 			/*
8358 			 * Align address and length of advice to page
8359 			 * boundaries for large pages
8360 			 */
8361 			if (seg->s_szc != 0) {
8362 				size_t	pgsz;
8363 
8364 				pgsz = page_get_pagesize(seg->s_szc);
8365 				addr = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
8366 				len = P2ROUNDUP(len, pgsz);
8367 			}
8368 
8369 			/*
8370 			 * Check to see whether policy is set already
8371 			 */
8372 			policy = lgrp_madv_to_policy(behav, len, svd->type);
8373 
8374 			anon_index = svd->anon_index + page;
8375 			off = svd->offset + (uintptr_t)(addr - seg->s_base);
8376 
8377 			if (svd->type == MAP_SHARED)
8378 				already_set = lgrp_shm_policy_set(policy, amp,
8379 				    anon_index, vp, off, len);
8380 			else
8381 				already_set =
8382 				    (policy == svd->policy_info.mem_policy);
8383 
8384 			/*
8385 			 * If policy set already and it shouldn't be reapplied,
8386 			 * don't do anything.
8387 			 */
8388 			if (already_set &&
8389 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8390 				break;
8391 
8392 			/*
8393 			 * For private memory, need writers lock on
8394 			 * address space because the segment may be
8395 			 * split or concatenated when changing policy
8396 			 */
8397 			if (svd->type == MAP_PRIVATE &&
8398 			    AS_READ_HELD(seg->s_as)) {
8399 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8400 				return (IE_RETRY);
8401 			}
8402 
8403 			/*
8404 			 * Mark any existing pages in given range for
8405 			 * migration
8406 			 */
8407 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8408 			    vp, svd->offset, 1);
8409 
8410 			/*
8411 			 * Don't need to try to split or concatenate
8412 			 * segments, since policy is same or this is a shared
8413 			 * memory segment
8414 			 */
8415 			if (already_set || svd->type == MAP_SHARED)
8416 				break;
8417 
8418 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
8419 				ASSERT(svd->amp == NULL);
8420 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
8421 				ASSERT(svd->softlockcnt == 0);
8422 				hat_leave_region(seg->s_as->a_hat, svd->rcookie,
8423 				    HAT_REGION_TEXT);
8424 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
8425 			}
8426 
8427 			/*
8428 			 * Split off new segment if advice only applies to a
8429 			 * portion of existing segment starting in middle
8430 			 */
8431 			new_seg = NULL;
8432 			eaddr = addr + len;
8433 			oldeaddr = seg->s_base + seg->s_size;
8434 			if (addr > seg->s_base) {
8435 				/*
8436 				 * Must flush I/O page cache
8437 				 * before splitting segment
8438 				 */
8439 				if (svd->softlockcnt > 0)
8440 					segvn_purge(seg);
8441 
8442 				/*
8443 				 * Split segment and return IE_REATTACH to tell
8444 				 * as_ctl() that current segment changed
8445 				 */
8446 				new_seg = segvn_split_seg(seg, addr);
8447 				new_svd = (struct segvn_data *)new_seg->s_data;
8448 				err = IE_REATTACH;
8449 
8450 				/*
8451 				 * If new segment ends where old one
8452 				 * did, try to concatenate the new
8453 				 * segment with next one.
8454 				 */
8455 				if (eaddr == oldeaddr) {
8456 					/*
8457 					 * Set policy for new segment
8458 					 */
8459 					(void) lgrp_privm_policy_set(policy,
8460 					    &new_svd->policy_info,
8461 					    new_seg->s_size);
8462 
8463 					next = AS_SEGNEXT(new_seg->s_as,
8464 					    new_seg);
8465 
8466 					if (next &&
8467 					    next->s_ops == &segvn_ops &&
8468 					    eaddr == next->s_base)
8469 						(void) segvn_concat(new_seg,
8470 						    next, 1);
8471 				}
8472 			}
8473 
8474 			/*
8475 			 * Split off end of existing segment if advice only
8476 			 * applies to a portion of segment ending before
8477 			 * end of the existing segment
8478 			 */
8479 			if (eaddr < oldeaddr) {
8480 				/*
8481 				 * Must flush I/O page cache
8482 				 * before splitting segment
8483 				 */
8484 				if (svd->softlockcnt > 0)
8485 					segvn_purge(seg);
8486 
8487 				/*
8488 				 * If beginning of old segment was already
8489 				 * split off, use new segment to split end off
8490 				 * from.
8491 				 */
8492 				if (new_seg != NULL && new_seg != seg) {
8493 					/*
8494 					 * Split segment
8495 					 */
8496 					(void) segvn_split_seg(new_seg, eaddr);
8497 
8498 					/*
8499 					 * Set policy for new segment
8500 					 */
8501 					(void) lgrp_privm_policy_set(policy,
8502 					    &new_svd->policy_info,
8503 					    new_seg->s_size);
8504 				} else {
8505 					/*
8506 					 * Split segment and return IE_REATTACH
8507 					 * to tell as_ctl() that current
8508 					 * segment changed
8509 					 */
8510 					(void) segvn_split_seg(seg, eaddr);
8511 					err = IE_REATTACH;
8512 
8513 					(void) lgrp_privm_policy_set(policy,
8514 					    &svd->policy_info, seg->s_size);
8515 
8516 					/*
8517 					 * If new segment starts where old one
8518 					 * did, try to concatenate it with
8519 					 * previous segment.
8520 					 */
8521 					if (addr == seg->s_base) {
8522 						prev = AS_SEGPREV(seg->s_as,
8523 						    seg);
8524 
8525 						/*
8526 						 * Drop lock for private data
8527 						 * of current segment before
8528 						 * concatenating (deleting) it
8529 						 */
8530 						if (prev &&
8531 						    prev->s_ops ==
8532 						    &segvn_ops &&
8533 						    addr == prev->s_base +
8534 						    prev->s_size) {
8535 							SEGVN_LOCK_EXIT(
8536 							    seg->s_as,
8537 							    &svd->lock);
8538 							(void) segvn_concat(
8539 							    prev, seg, 1);
8540 							return (err);
8541 						}
8542 					}
8543 				}
8544 			}
8545 			break;
8546 		case MADV_SEQUENTIAL:
8547 			ASSERT(seg->s_szc == 0);
8548 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8549 			hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
8550 			/* FALLTHROUGH */
8551 		case MADV_NORMAL:
8552 		case MADV_RANDOM:
8553 			bvpp = &svd->vpage[page];
8554 			evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8555 			for (; bvpp < evpp; bvpp++)
8556 				VPP_SETADVICE(bvpp, behav);
8557 			svd->advice = MADV_NORMAL;
8558 			break;
8559 		case MADV_WILLNEED:	/* handled in memcntl */
8560 		case MADV_DONTNEED:	/* handled in memcntl */
8561 		case MADV_FREE:		/* handled above */
8562 		case MADV_PURGE:	/* handled above */
8563 			break;
8564 		default:
8565 			err = EINVAL;
8566 		}
8567 	}
8568 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8569 	return (err);
8570 }
8571 
8572 /*
8573  * There is one kind of inheritance that can be specified for pages:
8574  *
8575  *     SEGP_INH_ZERO - Pages should be zeroed in the child
8576  */
8577 static int
8578 segvn_inherit(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
8579 {
8580 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8581 	struct vpage *bvpp, *evpp;
8582 	size_t page;
8583 	int ret = 0;
8584 
8585 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
8586 
8587 	/* Can't support something we don't know about */
8588 	if (behav != SEGP_INH_ZERO)
8589 		return (ENOTSUP);
8590 
8591 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
8592 
8593 	/*
8594 	 * This must be a straightforward anonymous segment that is mapped
8595 	 * privately and is not backed by a vnode.
8596 	 */
8597 	if (svd->tr_state != SEGVN_TR_OFF ||
8598 	    svd->type != MAP_PRIVATE ||
8599 	    svd->vp != NULL) {
8600 		ret = EINVAL;
8601 		goto out;
8602 	}
8603 
8604 	/*
8605 	 * If the entire segment has been marked as inherit zero, then no reason
8606 	 * to do anything else.
8607 	 */
8608 	if (svd->svn_inz == SEGVN_INZ_ALL) {
8609 		ret = 0;
8610 		goto out;
8611 	}
8612 
8613 	/*
8614 	 * If this applies to the entire segment, simply mark it and we're done.
8615 	 */
8616 	if ((addr == seg->s_base) && (len == seg->s_size)) {
8617 		svd->svn_inz = SEGVN_INZ_ALL;
8618 		ret = 0;
8619 		goto out;
8620 	}
8621 
8622 	/*
8623 	 * We've been asked to mark a subset of this segment as inherit zero,
8624 	 * therefore we need to mainpulate its vpages.
8625 	 */
8626 	if (svd->vpage == NULL) {
8627 		segvn_vpage(seg);
8628 		if (svd->vpage == NULL) {
8629 			ret = ENOMEM;
8630 			goto out;
8631 		}
8632 	}
8633 
8634 	svd->svn_inz = SEGVN_INZ_VPP;
8635 	page = seg_page(seg, addr);
8636 	bvpp = &svd->vpage[page];
8637 	evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8638 	for (; bvpp < evpp; bvpp++)
8639 		VPP_SETINHZERO(bvpp);
8640 	ret = 0;
8641 
8642 out:
8643 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8644 	return (ret);
8645 }
8646 
8647 /*
8648  * Create a vpage structure for this seg.
8649  */
8650 static void
8651 segvn_vpage(struct seg *seg)
8652 {
8653 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8654 	struct vpage *vp, *evp;
8655 	static pgcnt_t page_limit = 0;
8656 
8657 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
8658 
8659 	/*
8660 	 * If no vpage structure exists, allocate one.  Copy the protections
8661 	 * and the advice from the segment itself to the individual pages.
8662 	 */
8663 	if (svd->vpage == NULL) {
8664 		/*
8665 		 * Start by calculating the number of pages we must allocate to
8666 		 * track the per-page vpage structs needs for this entire
8667 		 * segment. If we know now that it will require more than our
8668 		 * heuristic for the maximum amount of kmem we can consume then
8669 		 * fail. We do this here, instead of trying to detect this deep
8670 		 * in page_resv and propagating the error up, since the entire
8671 		 * memory allocation stack is not amenable to passing this
8672 		 * back. Instead, it wants to keep trying.
8673 		 *
8674 		 * As a heuristic we set a page limit of 5/8s of total_pages
8675 		 * for this allocation. We use shifts so that no floating
8676 		 * point conversion takes place and only need to do the
8677 		 * calculation once.
8678 		 */
8679 		ulong_t mem_needed = seg_pages(seg) * sizeof (struct vpage);
8680 		pgcnt_t npages = mem_needed >> PAGESHIFT;
8681 
8682 		if (page_limit == 0)
8683 			page_limit = (total_pages >> 1) + (total_pages >> 3);
8684 
8685 		if (npages > page_limit)
8686 			return;
8687 
8688 		svd->pageadvice = 1;
8689 		svd->vpage = kmem_zalloc(mem_needed, KM_SLEEP);
8690 		evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
8691 		for (vp = svd->vpage; vp < evp; vp++) {
8692 			VPP_SETPROT(vp, svd->prot);
8693 			VPP_SETADVICE(vp, svd->advice);
8694 		}
8695 	}
8696 }
8697 
8698 /*
8699  * Dump the pages belonging to this segvn segment.
8700  */
8701 static void
8702 segvn_dump(struct seg *seg)
8703 {
8704 	struct segvn_data *svd;
8705 	page_t *pp;
8706 	struct anon_map *amp;
8707 	ulong_t	anon_index;
8708 	struct vnode *vp;
8709 	u_offset_t off, offset;
8710 	pfn_t pfn;
8711 	pgcnt_t page, npages;
8712 	caddr_t addr;
8713 
8714 	npages = seg_pages(seg);
8715 	svd = (struct segvn_data *)seg->s_data;
8716 	vp = svd->vp;
8717 	off = offset = svd->offset;
8718 	addr = seg->s_base;
8719 
8720 	if ((amp = svd->amp) != NULL) {
8721 		anon_index = svd->anon_index;
8722 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8723 	}
8724 
8725 	for (page = 0; page < npages; page++, offset += PAGESIZE) {
8726 		struct anon *ap;
8727 		int we_own_it = 0;
8728 
8729 		if (amp && (ap = anon_get_ptr(svd->amp->ahp, anon_index++))) {
8730 			swap_xlate_nopanic(ap, &vp, &off);
8731 		} else {
8732 			vp = svd->vp;
8733 			off = offset;
8734 		}
8735 
8736 		/*
8737 		 * If pp == NULL, the page either does not exist
8738 		 * or is exclusively locked.  So determine if it
8739 		 * exists before searching for it.
8740 		 */
8741 
8742 		if ((pp = page_lookup_nowait(vp, off, SE_SHARED)))
8743 			we_own_it = 1;
8744 		else
8745 			pp = page_exists(vp, off);
8746 
8747 		if (pp) {
8748 			pfn = page_pptonum(pp);
8749 			dump_addpage(seg->s_as, addr, pfn);
8750 			if (we_own_it)
8751 				page_unlock(pp);
8752 		}
8753 		addr += PAGESIZE;
8754 		dump_timeleft = dump_timeout;
8755 	}
8756 
8757 	if (amp != NULL)
8758 		ANON_LOCK_EXIT(&amp->a_rwlock);
8759 }
8760 
8761 #ifdef DEBUG
8762 static uint32_t segvn_pglock_mtbf = 0;
8763 #endif
8764 
8765 #define	PCACHE_SHWLIST		((page_t *)-2)
8766 #define	NOPCACHE_SHWLIST	((page_t *)-1)
8767 
8768 /*
8769  * Lock/Unlock anon pages over a given range. Return shadow list. This routine
8770  * uses global segment pcache to cache shadow lists (i.e. pp arrays) of pages
8771  * to avoid the overhead of per page locking, unlocking for subsequent IOs to
8772  * the same parts of the segment. Currently shadow list creation is only
8773  * supported for pure anon segments. MAP_PRIVATE segment pcache entries are
8774  * tagged with segment pointer, starting virtual address and length. This
8775  * approach for MAP_SHARED segments may add many pcache entries for the same
8776  * set of pages and lead to long hash chains that decrease pcache lookup
8777  * performance. To avoid this issue for shared segments shared anon map and
8778  * starting anon index are used for pcache entry tagging. This allows all
8779  * segments to share pcache entries for the same anon range and reduces pcache
8780  * chain's length as well as memory overhead from duplicate shadow lists and
8781  * pcache entries.
8782  *
8783  * softlockcnt field in segvn_data structure counts the number of F_SOFTLOCK'd
8784  * pages via segvn_fault() and pagelock'd pages via this routine. But pagelock
8785  * part of softlockcnt accounting is done differently for private and shared
8786  * segments. In private segment case softlock is only incremented when a new
8787  * shadow list is created but not when an existing one is found via
8788  * seg_plookup(). pcache entries have reference count incremented/decremented
8789  * by each seg_plookup()/seg_pinactive() operation. Only entries that have 0
8790  * reference count can be purged (and purging is needed before segment can be
8791  * freed). When a private segment pcache entry is purged segvn_reclaim() will
8792  * decrement softlockcnt. Since in private segment case each of its pcache
8793  * entries only belongs to this segment we can expect that when
8794  * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8795  * segment purge will succeed and softlockcnt will drop to 0. In shared
8796  * segment case reference count in pcache entry counts active locks from many
8797  * different segments so we can't expect segment purging to succeed even when
8798  * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8799  * segment. To be able to determine when there're no pending pagelocks in
8800  * shared segment case we don't rely on purging to make softlockcnt drop to 0
8801  * but instead softlockcnt is incremented and decremented for every
8802  * segvn_pagelock(L_PAGELOCK/L_PAGEUNLOCK) call regardless if a new shadow
8803  * list was created or an existing one was found. When softlockcnt drops to 0
8804  * this segment no longer has any claims for pcached shadow lists and the
8805  * segment can be freed even if there're still active pcache entries
8806  * shared by this segment anon map. Shared segment pcache entries belong to
8807  * anon map and are typically removed when anon map is freed after all
8808  * processes destroy the segments that use this anon map.
8809  */
8810 static int
8811 segvn_pagelock(struct seg *seg, caddr_t addr, size_t len, struct page ***ppp,
8812     enum lock_type type, enum seg_rw rw)
8813 {
8814 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8815 	size_t np;
8816 	pgcnt_t adjustpages;
8817 	pgcnt_t npages;
8818 	ulong_t anon_index;
8819 	uint_t protchk = (rw == S_READ) ? PROT_READ : PROT_WRITE;
8820 	uint_t error;
8821 	struct anon_map *amp;
8822 	pgcnt_t anpgcnt;
8823 	struct page **pplist, **pl, *pp;
8824 	caddr_t a;
8825 	size_t page;
8826 	caddr_t lpgaddr, lpgeaddr;
8827 	anon_sync_obj_t cookie;
8828 	int anlock;
8829 	struct anon_map *pamp;
8830 	caddr_t paddr;
8831 	seg_preclaim_cbfunc_t preclaim_callback;
8832 	size_t pgsz;
8833 	int use_pcache;
8834 	size_t wlen;
8835 	uint_t pflags = 0;
8836 	int sftlck_sbase = 0;
8837 	int sftlck_send = 0;
8838 
8839 #ifdef DEBUG
8840 	if (type == L_PAGELOCK && segvn_pglock_mtbf) {
8841 		hrtime_t ts = gethrtime();
8842 		if ((ts % segvn_pglock_mtbf) == 0) {
8843 			return (ENOTSUP);
8844 		}
8845 		if ((ts % segvn_pglock_mtbf) == 1) {
8846 			return (EFAULT);
8847 		}
8848 	}
8849 #endif
8850 
8851 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_START,
8852 	    "segvn_pagelock: start seg %p addr %p", seg, addr);
8853 
8854 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
8855 	ASSERT(type == L_PAGELOCK || type == L_PAGEUNLOCK);
8856 
8857 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8858 
8859 	/*
8860 	 * for now we only support pagelock to anon memory. We would have to
8861 	 * check protections for vnode objects and call into the vnode driver.
8862 	 * That's too much for a fast path. Let the fault entry point handle
8863 	 * it.
8864 	 */
8865 	if (svd->vp != NULL) {
8866 		if (type == L_PAGELOCK) {
8867 			error = ENOTSUP;
8868 			goto out;
8869 		}
8870 		panic("segvn_pagelock(L_PAGEUNLOCK): vp != NULL");
8871 	}
8872 	if ((amp = svd->amp) == NULL) {
8873 		if (type == L_PAGELOCK) {
8874 			error = EFAULT;
8875 			goto out;
8876 		}
8877 		panic("segvn_pagelock(L_PAGEUNLOCK): amp == NULL");
8878 	}
8879 	if (rw != S_READ && rw != S_WRITE) {
8880 		if (type == L_PAGELOCK) {
8881 			error = ENOTSUP;
8882 			goto out;
8883 		}
8884 		panic("segvn_pagelock(L_PAGEUNLOCK): bad rw");
8885 	}
8886 
8887 	if (seg->s_szc != 0) {
8888 		/*
8889 		 * We are adjusting the pagelock region to the large page size
8890 		 * boundary because the unlocked part of a large page cannot
8891 		 * be freed anyway unless all constituent pages of a large
8892 		 * page are locked. Bigger regions reduce pcache chain length
8893 		 * and improve lookup performance. The tradeoff is that the
8894 		 * very first segvn_pagelock() call for a given page is more
8895 		 * expensive if only 1 page_t is needed for IO. This is only
8896 		 * an issue if pcache entry doesn't get reused by several
8897 		 * subsequent calls. We optimize here for the case when pcache
8898 		 * is heavily used by repeated IOs to the same address range.
8899 		 *
8900 		 * Note segment's page size cannot change while we are holding
8901 		 * as lock.  And then it cannot change while softlockcnt is
8902 		 * not 0. This will allow us to correctly recalculate large
8903 		 * page size region for the matching pageunlock/reclaim call
8904 		 * since as_pageunlock() caller must always match
8905 		 * as_pagelock() call's addr and len.
8906 		 *
8907 		 * For pageunlock *ppp points to the pointer of page_t that
8908 		 * corresponds to the real unadjusted start address. Similar
8909 		 * for pagelock *ppp must point to the pointer of page_t that
8910 		 * corresponds to the real unadjusted start address.
8911 		 */
8912 		pgsz = page_get_pagesize(seg->s_szc);
8913 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
8914 		adjustpages = btop((uintptr_t)(addr - lpgaddr));
8915 	} else if (len < segvn_pglock_comb_thrshld) {
8916 		lpgaddr = addr;
8917 		lpgeaddr = addr + len;
8918 		adjustpages = 0;
8919 		pgsz = PAGESIZE;
8920 	} else {
8921 		/*
8922 		 * Align the address range of large enough requests to allow
8923 		 * combining of different shadow lists into 1 to reduce memory
8924 		 * overhead from potentially overlapping large shadow lists
8925 		 * (worst case is we have a 1MB IO into buffers with start
8926 		 * addresses separated by 4K).  Alignment is only possible if
8927 		 * padded chunks have sufficient access permissions. Note
8928 		 * permissions won't change between L_PAGELOCK and
8929 		 * L_PAGEUNLOCK calls since non 0 softlockcnt will force
8930 		 * segvn_setprot() to wait until softlockcnt drops to 0. This
8931 		 * allows us to determine in L_PAGEUNLOCK the same range we
8932 		 * computed in L_PAGELOCK.
8933 		 *
8934 		 * If alignment is limited by segment ends set
8935 		 * sftlck_sbase/sftlck_send flags. In L_PAGELOCK case when
8936 		 * these flags are set bump softlockcnt_sbase/softlockcnt_send
8937 		 * per segment counters. In L_PAGEUNLOCK case decrease
8938 		 * softlockcnt_sbase/softlockcnt_send counters if
8939 		 * sftlck_sbase/sftlck_send flags are set.  When
8940 		 * softlockcnt_sbase/softlockcnt_send are non 0
8941 		 * segvn_concat()/segvn_extend_prev()/segvn_extend_next()
8942 		 * won't merge the segments. This restriction combined with
8943 		 * restriction on segment unmapping and splitting for segments
8944 		 * that have non 0 softlockcnt allows L_PAGEUNLOCK to
8945 		 * correctly determine the same range that was previously
8946 		 * locked by matching L_PAGELOCK.
8947 		 */
8948 		pflags = SEGP_PSHIFT | (segvn_pglock_comb_bshift << 16);
8949 		pgsz = PAGESIZE;
8950 		if (svd->type == MAP_PRIVATE) {
8951 			lpgaddr = (caddr_t)P2ALIGN((uintptr_t)addr,
8952 			    segvn_pglock_comb_balign);
8953 			if (lpgaddr < seg->s_base) {
8954 				lpgaddr = seg->s_base;
8955 				sftlck_sbase = 1;
8956 			}
8957 		} else {
8958 			ulong_t aix = svd->anon_index + seg_page(seg, addr);
8959 			ulong_t aaix = P2ALIGN(aix, segvn_pglock_comb_palign);
8960 			if (aaix < svd->anon_index) {
8961 				lpgaddr = seg->s_base;
8962 				sftlck_sbase = 1;
8963 			} else {
8964 				lpgaddr = addr - ptob(aix - aaix);
8965 				ASSERT(lpgaddr >= seg->s_base);
8966 			}
8967 		}
8968 		if (svd->pageprot && lpgaddr != addr) {
8969 			struct vpage *vp = &svd->vpage[seg_page(seg, lpgaddr)];
8970 			struct vpage *evp = &svd->vpage[seg_page(seg, addr)];
8971 			while (vp < evp) {
8972 				if ((VPP_PROT(vp) & protchk) == 0) {
8973 					break;
8974 				}
8975 				vp++;
8976 			}
8977 			if (vp < evp) {
8978 				lpgaddr = addr;
8979 				pflags = 0;
8980 			}
8981 		}
8982 		lpgeaddr = addr + len;
8983 		if (pflags) {
8984 			if (svd->type == MAP_PRIVATE) {
8985 				lpgeaddr = (caddr_t)P2ROUNDUP(
8986 				    (uintptr_t)lpgeaddr,
8987 				    segvn_pglock_comb_balign);
8988 			} else {
8989 				ulong_t aix = svd->anon_index +
8990 				    seg_page(seg, lpgeaddr);
8991 				ulong_t aaix = P2ROUNDUP(aix,
8992 				    segvn_pglock_comb_palign);
8993 				if (aaix < aix) {
8994 					lpgeaddr = 0;
8995 				} else {
8996 					lpgeaddr += ptob(aaix - aix);
8997 				}
8998 			}
8999 			if (lpgeaddr == 0 ||
9000 			    lpgeaddr > seg->s_base + seg->s_size) {
9001 				lpgeaddr = seg->s_base + seg->s_size;
9002 				sftlck_send = 1;
9003 			}
9004 		}
9005 		if (svd->pageprot && lpgeaddr != addr + len) {
9006 			struct vpage *vp;
9007 			struct vpage *evp;
9008 
9009 			vp = &svd->vpage[seg_page(seg, addr + len)];
9010 			evp = &svd->vpage[seg_page(seg, lpgeaddr)];
9011 
9012 			while (vp < evp) {
9013 				if ((VPP_PROT(vp) & protchk) == 0) {
9014 					break;
9015 				}
9016 				vp++;
9017 			}
9018 			if (vp < evp) {
9019 				lpgeaddr = addr + len;
9020 			}
9021 		}
9022 		adjustpages = btop((uintptr_t)(addr - lpgaddr));
9023 	}
9024 
9025 	/*
9026 	 * For MAP_SHARED segments we create pcache entries tagged by amp and
9027 	 * anon index so that we can share pcache entries with other segments
9028 	 * that map this amp.  For private segments pcache entries are tagged
9029 	 * with segment and virtual address.
9030 	 */
9031 	if (svd->type == MAP_SHARED) {
9032 		pamp = amp;
9033 		paddr = (caddr_t)((lpgaddr - seg->s_base) +
9034 		    ptob(svd->anon_index));
9035 		preclaim_callback = shamp_reclaim;
9036 	} else {
9037 		pamp = NULL;
9038 		paddr = lpgaddr;
9039 		preclaim_callback = segvn_reclaim;
9040 	}
9041 
9042 	if (type == L_PAGEUNLOCK) {
9043 		VM_STAT_ADD(segvnvmstats.pagelock[0]);
9044 
9045 		/*
9046 		 * update hat ref bits for /proc. We need to make sure
9047 		 * that threads tracing the ref and mod bits of the
9048 		 * address space get the right data.
9049 		 * Note: page ref and mod bits are updated at reclaim time
9050 		 */
9051 		if (seg->s_as->a_vbits) {
9052 			for (a = addr; a < addr + len; a += PAGESIZE) {
9053 				if (rw == S_WRITE) {
9054 					hat_setstat(seg->s_as, a,
9055 					    PAGESIZE, P_REF | P_MOD);
9056 				} else {
9057 					hat_setstat(seg->s_as, a,
9058 					    PAGESIZE, P_REF);
9059 				}
9060 			}
9061 		}
9062 
9063 		/*
9064 		 * Check the shadow list entry after the last page used in
9065 		 * this IO request. If it's NOPCACHE_SHWLIST the shadow list
9066 		 * was not inserted into pcache and is not large page
9067 		 * adjusted.  In this case call reclaim callback directly and
9068 		 * don't adjust the shadow list start and size for large
9069 		 * pages.
9070 		 */
9071 		npages = btop(len);
9072 		if ((*ppp)[npages] == NOPCACHE_SHWLIST) {
9073 			void *ptag;
9074 			if (pamp != NULL) {
9075 				ASSERT(svd->type == MAP_SHARED);
9076 				ptag = (void *)pamp;
9077 				paddr = (caddr_t)((addr - seg->s_base) +
9078 				    ptob(svd->anon_index));
9079 			} else {
9080 				ptag = (void *)seg;
9081 				paddr = addr;
9082 			}
9083 			(*preclaim_callback)(ptag, paddr, len, *ppp, rw, 0);
9084 		} else {
9085 			ASSERT((*ppp)[npages] == PCACHE_SHWLIST ||
9086 			    IS_SWAPFSVP((*ppp)[npages]->p_vnode));
9087 			len = lpgeaddr - lpgaddr;
9088 			npages = btop(len);
9089 			seg_pinactive(seg, pamp, paddr, len,
9090 			    *ppp - adjustpages, rw, pflags, preclaim_callback);
9091 		}
9092 
9093 		if (pamp != NULL) {
9094 			ASSERT(svd->type == MAP_SHARED);
9095 			ASSERT(svd->softlockcnt >= npages);
9096 			atomic_add_long((ulong_t *)&svd->softlockcnt, -npages);
9097 		}
9098 
9099 		if (sftlck_sbase) {
9100 			ASSERT(svd->softlockcnt_sbase > 0);
9101 			atomic_dec_ulong((ulong_t *)&svd->softlockcnt_sbase);
9102 		}
9103 		if (sftlck_send) {
9104 			ASSERT(svd->softlockcnt_send > 0);
9105 			atomic_dec_ulong((ulong_t *)&svd->softlockcnt_send);
9106 		}
9107 
9108 		/*
9109 		 * If someone is blocked while unmapping, we purge
9110 		 * segment page cache and thus reclaim pplist synchronously
9111 		 * without waiting for seg_pasync_thread. This speeds up
9112 		 * unmapping in cases where munmap(2) is called, while
9113 		 * raw async i/o is still in progress or where a thread
9114 		 * exits on data fault in a multithreaded application.
9115 		 */
9116 		if (AS_ISUNMAPWAIT(seg->s_as)) {
9117 			if (svd->softlockcnt == 0) {
9118 				mutex_enter(&seg->s_as->a_contents);
9119 				if (AS_ISUNMAPWAIT(seg->s_as)) {
9120 					AS_CLRUNMAPWAIT(seg->s_as);
9121 					cv_broadcast(&seg->s_as->a_cv);
9122 				}
9123 				mutex_exit(&seg->s_as->a_contents);
9124 			} else if (pamp == NULL) {
9125 				/*
9126 				 * softlockcnt is not 0 and this is a
9127 				 * MAP_PRIVATE segment. Try to purge its
9128 				 * pcache entries to reduce softlockcnt.
9129 				 * If it drops to 0 segvn_reclaim()
9130 				 * will wake up a thread waiting on
9131 				 * unmapwait flag.
9132 				 *
9133 				 * We don't purge MAP_SHARED segments with non
9134 				 * 0 softlockcnt since IO is still in progress
9135 				 * for such segments.
9136 				 */
9137 				ASSERT(svd->type == MAP_PRIVATE);
9138 				segvn_purge(seg);
9139 			}
9140 		}
9141 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9142 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_UNLOCK_END,
9143 		    "segvn_pagelock: unlock seg %p addr %p", seg, addr);
9144 		return (0);
9145 	}
9146 
9147 	/* The L_PAGELOCK case ... */
9148 
9149 	VM_STAT_ADD(segvnvmstats.pagelock[1]);
9150 
9151 	/*
9152 	 * For MAP_SHARED segments we have to check protections before
9153 	 * seg_plookup() since pcache entries may be shared by many segments
9154 	 * with potentially different page protections.
9155 	 */
9156 	if (pamp != NULL) {
9157 		ASSERT(svd->type == MAP_SHARED);
9158 		if (svd->pageprot == 0) {
9159 			if ((svd->prot & protchk) == 0) {
9160 				error = EACCES;
9161 				goto out;
9162 			}
9163 		} else {
9164 			/*
9165 			 * check page protections
9166 			 */
9167 			caddr_t ea;
9168 
9169 			if (seg->s_szc) {
9170 				a = lpgaddr;
9171 				ea = lpgeaddr;
9172 			} else {
9173 				a = addr;
9174 				ea = addr + len;
9175 			}
9176 			for (; a < ea; a += pgsz) {
9177 				struct vpage *vp;
9178 
9179 				ASSERT(seg->s_szc == 0 ||
9180 				    sameprot(seg, a, pgsz));
9181 				vp = &svd->vpage[seg_page(seg, a)];
9182 				if ((VPP_PROT(vp) & protchk) == 0) {
9183 					error = EACCES;
9184 					goto out;
9185 				}
9186 			}
9187 		}
9188 	}
9189 
9190 	/*
9191 	 * try to find pages in segment page cache
9192 	 */
9193 	pplist = seg_plookup(seg, pamp, paddr, lpgeaddr - lpgaddr, rw, pflags);
9194 	if (pplist != NULL) {
9195 		if (pamp != NULL) {
9196 			npages = btop((uintptr_t)(lpgeaddr - lpgaddr));
9197 			ASSERT(svd->type == MAP_SHARED);
9198 			atomic_add_long((ulong_t *)&svd->softlockcnt,
9199 			    npages);
9200 		}
9201 		if (sftlck_sbase) {
9202 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_sbase);
9203 		}
9204 		if (sftlck_send) {
9205 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_send);
9206 		}
9207 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9208 		*ppp = pplist + adjustpages;
9209 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_HIT_END,
9210 		    "segvn_pagelock: cache hit seg %p addr %p", seg, addr);
9211 		return (0);
9212 	}
9213 
9214 	/*
9215 	 * For MAP_SHARED segments we already verified above that segment
9216 	 * protections allow this pagelock operation.
9217 	 */
9218 	if (pamp == NULL) {
9219 		ASSERT(svd->type == MAP_PRIVATE);
9220 		if (svd->pageprot == 0) {
9221 			if ((svd->prot & protchk) == 0) {
9222 				error = EACCES;
9223 				goto out;
9224 			}
9225 			if (svd->prot & PROT_WRITE) {
9226 				wlen = lpgeaddr - lpgaddr;
9227 			} else {
9228 				wlen = 0;
9229 				ASSERT(rw == S_READ);
9230 			}
9231 		} else {
9232 			int wcont = 1;
9233 			/*
9234 			 * check page protections
9235 			 */
9236 			for (a = lpgaddr, wlen = 0; a < lpgeaddr; a += pgsz) {
9237 				struct vpage *vp;
9238 
9239 				ASSERT(seg->s_szc == 0 ||
9240 				    sameprot(seg, a, pgsz));
9241 				vp = &svd->vpage[seg_page(seg, a)];
9242 				if ((VPP_PROT(vp) & protchk) == 0) {
9243 					error = EACCES;
9244 					goto out;
9245 				}
9246 				if (wcont && (VPP_PROT(vp) & PROT_WRITE)) {
9247 					wlen += pgsz;
9248 				} else {
9249 					wcont = 0;
9250 					ASSERT(rw == S_READ);
9251 				}
9252 			}
9253 		}
9254 		ASSERT(rw == S_READ || wlen == lpgeaddr - lpgaddr);
9255 		ASSERT(rw == S_WRITE || wlen <= lpgeaddr - lpgaddr);
9256 	}
9257 
9258 	/*
9259 	 * Only build large page adjusted shadow list if we expect to insert
9260 	 * it into pcache. For large enough pages it's a big overhead to
9261 	 * create a shadow list of the entire large page. But this overhead
9262 	 * should be amortized over repeated pcache hits on subsequent reuse
9263 	 * of this shadow list (IO into any range within this shadow list will
9264 	 * find it in pcache since we large page align the request for pcache
9265 	 * lookups). pcache performance is improved with bigger shadow lists
9266 	 * as it reduces the time to pcache the entire big segment and reduces
9267 	 * pcache chain length.
9268 	 */
9269 	if (seg_pinsert_check(seg, pamp, paddr,
9270 	    lpgeaddr - lpgaddr, pflags) == SEGP_SUCCESS) {
9271 		addr = lpgaddr;
9272 		len = lpgeaddr - lpgaddr;
9273 		use_pcache = 1;
9274 	} else {
9275 		use_pcache = 0;
9276 		/*
9277 		 * Since this entry will not be inserted into the pcache, we
9278 		 * will not do any adjustments to the starting address or
9279 		 * size of the memory to be locked.
9280 		 */
9281 		adjustpages = 0;
9282 	}
9283 	npages = btop(len);
9284 
9285 	pplist = kmem_alloc(sizeof (page_t *) * (npages + 1), KM_SLEEP);
9286 	pl = pplist;
9287 	*ppp = pplist + adjustpages;
9288 	/*
9289 	 * If use_pcache is 0 this shadow list is not large page adjusted.
9290 	 * Record this info in the last entry of shadow array so that
9291 	 * L_PAGEUNLOCK can determine if it should large page adjust the
9292 	 * address range to find the real range that was locked.
9293 	 */
9294 	pl[npages] = use_pcache ? PCACHE_SHWLIST : NOPCACHE_SHWLIST;
9295 
9296 	page = seg_page(seg, addr);
9297 	anon_index = svd->anon_index + page;
9298 
9299 	anlock = 0;
9300 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
9301 	ASSERT(amp->a_szc >= seg->s_szc);
9302 	anpgcnt = page_get_pagecnt(amp->a_szc);
9303 	for (a = addr; a < addr + len; a += PAGESIZE, anon_index++) {
9304 		struct anon *ap;
9305 		struct vnode *vp;
9306 		u_offset_t off;
9307 
9308 		/*
9309 		 * Lock and unlock anon array only once per large page.
9310 		 * anon_array_enter() locks the root anon slot according to
9311 		 * a_szc which can't change while anon map is locked.  We lock
9312 		 * anon the first time through this loop and each time we
9313 		 * reach anon index that corresponds to a root of a large
9314 		 * page.
9315 		 */
9316 		if (a == addr || P2PHASE(anon_index, anpgcnt) == 0) {
9317 			ASSERT(anlock == 0);
9318 			anon_array_enter(amp, anon_index, &cookie);
9319 			anlock = 1;
9320 		}
9321 		ap = anon_get_ptr(amp->ahp, anon_index);
9322 
9323 		/*
9324 		 * We must never use seg_pcache for COW pages
9325 		 * because we might end up with original page still
9326 		 * lying in seg_pcache even after private page is
9327 		 * created. This leads to data corruption as
9328 		 * aio_write refers to the page still in cache
9329 		 * while all other accesses refer to the private
9330 		 * page.
9331 		 */
9332 		if (ap == NULL || ap->an_refcnt != 1) {
9333 			struct vpage *vpage;
9334 
9335 			if (seg->s_szc) {
9336 				error = EFAULT;
9337 				break;
9338 			}
9339 			if (svd->vpage != NULL) {
9340 				vpage = &svd->vpage[seg_page(seg, a)];
9341 			} else {
9342 				vpage = NULL;
9343 			}
9344 			ASSERT(anlock);
9345 			anon_array_exit(&cookie);
9346 			anlock = 0;
9347 			pp = NULL;
9348 			error = segvn_faultpage(seg->s_as->a_hat, seg, a, 0,
9349 			    vpage, &pp, 0, F_INVAL, rw, 1);
9350 			if (error) {
9351 				error = fc_decode(error);
9352 				break;
9353 			}
9354 			anon_array_enter(amp, anon_index, &cookie);
9355 			anlock = 1;
9356 			ap = anon_get_ptr(amp->ahp, anon_index);
9357 			if (ap == NULL || ap->an_refcnt != 1) {
9358 				error = EFAULT;
9359 				break;
9360 			}
9361 		}
9362 		swap_xlate(ap, &vp, &off);
9363 		pp = page_lookup_nowait(vp, off, SE_SHARED);
9364 		if (pp == NULL) {
9365 			error = EFAULT;
9366 			break;
9367 		}
9368 		if (ap->an_pvp != NULL) {
9369 			anon_swap_free(ap, pp);
9370 		}
9371 		/*
9372 		 * Unlock anon if this is the last slot in a large page.
9373 		 */
9374 		if (P2PHASE(anon_index, anpgcnt) == anpgcnt - 1) {
9375 			ASSERT(anlock);
9376 			anon_array_exit(&cookie);
9377 			anlock = 0;
9378 		}
9379 		*pplist++ = pp;
9380 	}
9381 	if (anlock) {		/* Ensure the lock is dropped */
9382 		anon_array_exit(&cookie);
9383 	}
9384 	ANON_LOCK_EXIT(&amp->a_rwlock);
9385 
9386 	if (a >= addr + len) {
9387 		atomic_add_long((ulong_t *)&svd->softlockcnt, npages);
9388 		if (pamp != NULL) {
9389 			ASSERT(svd->type == MAP_SHARED);
9390 			atomic_add_long((ulong_t *)&pamp->a_softlockcnt,
9391 			    npages);
9392 			wlen = len;
9393 		}
9394 		if (sftlck_sbase) {
9395 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_sbase);
9396 		}
9397 		if (sftlck_send) {
9398 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_send);
9399 		}
9400 		if (use_pcache) {
9401 			(void) seg_pinsert(seg, pamp, paddr, len, wlen, pl,
9402 			    rw, pflags, preclaim_callback);
9403 		}
9404 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9405 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_FILL_END,
9406 		    "segvn_pagelock: cache fill seg %p addr %p", seg, addr);
9407 		return (0);
9408 	}
9409 
9410 	pplist = pl;
9411 	np = ((uintptr_t)(a - addr)) >> PAGESHIFT;
9412 	while (np > (uint_t)0) {
9413 		ASSERT(PAGE_LOCKED(*pplist));
9414 		page_unlock(*pplist);
9415 		np--;
9416 		pplist++;
9417 	}
9418 	kmem_free(pl, sizeof (page_t *) * (npages + 1));
9419 out:
9420 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9421 	*ppp = NULL;
9422 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
9423 	    "segvn_pagelock: cache miss seg %p addr %p", seg, addr);
9424 	return (error);
9425 }
9426 
9427 /*
9428  * purge any cached pages in the I/O page cache
9429  */
9430 static void
9431 segvn_purge(struct seg *seg)
9432 {
9433 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9434 
9435 	/*
9436 	 * pcache is only used by pure anon segments.
9437 	 */
9438 	if (svd->amp == NULL || svd->vp != NULL) {
9439 		return;
9440 	}
9441 
9442 	/*
9443 	 * For MAP_SHARED segments non 0 segment's softlockcnt means
9444 	 * active IO is still in progress via this segment. So we only
9445 	 * purge MAP_SHARED segments when their softlockcnt is 0.
9446 	 */
9447 	if (svd->type == MAP_PRIVATE) {
9448 		if (svd->softlockcnt) {
9449 			seg_ppurge(seg, NULL, 0);
9450 		}
9451 	} else if (svd->softlockcnt == 0 && svd->amp->a_softlockcnt != 0) {
9452 		seg_ppurge(seg, svd->amp, 0);
9453 	}
9454 }
9455 
9456 /*
9457  * If async argument is not 0 we are called from pcache async thread and don't
9458  * hold AS lock.
9459  */
9460 
9461 /*ARGSUSED*/
9462 static int
9463 segvn_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9464     enum seg_rw rw, int async)
9465 {
9466 	struct seg *seg = (struct seg *)ptag;
9467 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9468 	pgcnt_t np, npages;
9469 	struct page **pl;
9470 
9471 	npages = np = btop(len);
9472 	ASSERT(npages);
9473 
9474 	ASSERT(svd->vp == NULL && svd->amp != NULL);
9475 	ASSERT(svd->softlockcnt >= npages);
9476 	ASSERT(async || AS_LOCK_HELD(seg->s_as));
9477 
9478 	pl = pplist;
9479 
9480 	ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9481 	ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9482 
9483 	while (np > (uint_t)0) {
9484 		if (rw == S_WRITE) {
9485 			hat_setrefmod(*pplist);
9486 		} else {
9487 			hat_setref(*pplist);
9488 		}
9489 		page_unlock(*pplist);
9490 		np--;
9491 		pplist++;
9492 	}
9493 
9494 	kmem_free(pl, sizeof (page_t *) * (npages + 1));
9495 
9496 	/*
9497 	 * If we are pcache async thread we don't hold AS lock. This means if
9498 	 * softlockcnt drops to 0 after the decrement below address space may
9499 	 * get freed. We can't allow it since after softlock derement to 0 we
9500 	 * still need to access as structure for possible wakeup of unmap
9501 	 * waiters. To prevent the disappearance of as we take this segment
9502 	 * segfree_syncmtx. segvn_free() also takes this mutex as a barrier to
9503 	 * make sure this routine completes before segment is freed.
9504 	 *
9505 	 * The second complication we have to deal with in async case is a
9506 	 * possibility of missed wake up of unmap wait thread. When we don't
9507 	 * hold as lock here we may take a_contents lock before unmap wait
9508 	 * thread that was first to see softlockcnt was still not 0. As a
9509 	 * result we'll fail to wake up an unmap wait thread. To avoid this
9510 	 * race we set nounmapwait flag in as structure if we drop softlockcnt
9511 	 * to 0 when we were called by pcache async thread.  unmapwait thread
9512 	 * will not block if this flag is set.
9513 	 */
9514 	if (async) {
9515 		mutex_enter(&svd->segfree_syncmtx);
9516 	}
9517 
9518 	if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -npages)) {
9519 		if (async || AS_ISUNMAPWAIT(seg->s_as)) {
9520 			mutex_enter(&seg->s_as->a_contents);
9521 			if (async) {
9522 				AS_SETNOUNMAPWAIT(seg->s_as);
9523 			}
9524 			if (AS_ISUNMAPWAIT(seg->s_as)) {
9525 				AS_CLRUNMAPWAIT(seg->s_as);
9526 				cv_broadcast(&seg->s_as->a_cv);
9527 			}
9528 			mutex_exit(&seg->s_as->a_contents);
9529 		}
9530 	}
9531 
9532 	if (async) {
9533 		mutex_exit(&svd->segfree_syncmtx);
9534 	}
9535 	return (0);
9536 }
9537 
9538 /*ARGSUSED*/
9539 static int
9540 shamp_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9541     enum seg_rw rw, int async)
9542 {
9543 	amp_t *amp = (amp_t *)ptag;
9544 	pgcnt_t np, npages;
9545 	struct page **pl;
9546 
9547 	npages = np = btop(len);
9548 	ASSERT(npages);
9549 	ASSERT(amp->a_softlockcnt >= npages);
9550 
9551 	pl = pplist;
9552 
9553 	ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9554 	ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9555 
9556 	while (np > (uint_t)0) {
9557 		if (rw == S_WRITE) {
9558 			hat_setrefmod(*pplist);
9559 		} else {
9560 			hat_setref(*pplist);
9561 		}
9562 		page_unlock(*pplist);
9563 		np--;
9564 		pplist++;
9565 	}
9566 
9567 	kmem_free(pl, sizeof (page_t *) * (npages + 1));
9568 
9569 	/*
9570 	 * If somebody sleeps in anonmap_purge() wake them up if a_softlockcnt
9571 	 * drops to 0. anon map can't be freed until a_softlockcnt drops to 0
9572 	 * and anonmap_purge() acquires a_purgemtx.
9573 	 */
9574 	mutex_enter(&amp->a_purgemtx);
9575 	if (!atomic_add_long_nv((ulong_t *)&amp->a_softlockcnt, -npages) &&
9576 	    amp->a_purgewait) {
9577 		amp->a_purgewait = 0;
9578 		cv_broadcast(&amp->a_purgecv);
9579 	}
9580 	mutex_exit(&amp->a_purgemtx);
9581 	return (0);
9582 }
9583 
9584 /*
9585  * get a memory ID for an addr in a given segment
9586  *
9587  * XXX only creates PAGESIZE pages if anon slots are not initialized.
9588  * At fault time they will be relocated into larger pages.
9589  */
9590 static int
9591 segvn_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
9592 {
9593 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9594 	struct anon 	*ap = NULL;
9595 	ulong_t		anon_index;
9596 	struct anon_map	*amp;
9597 	anon_sync_obj_t cookie;
9598 
9599 	if (svd->type == MAP_PRIVATE) {
9600 		memidp->val[0] = (uintptr_t)seg->s_as;
9601 		memidp->val[1] = (uintptr_t)addr;
9602 		return (0);
9603 	}
9604 
9605 	if (svd->type == MAP_SHARED) {
9606 		if (svd->vp) {
9607 			memidp->val[0] = (uintptr_t)svd->vp;
9608 			memidp->val[1] = (u_longlong_t)svd->offset +
9609 			    (uintptr_t)(addr - seg->s_base);
9610 			return (0);
9611 		} else {
9612 
9613 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
9614 			if ((amp = svd->amp) != NULL) {
9615 				anon_index = svd->anon_index +
9616 				    seg_page(seg, addr);
9617 			}
9618 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9619 
9620 			ASSERT(amp != NULL);
9621 
9622 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
9623 			anon_array_enter(amp, anon_index, &cookie);
9624 			ap = anon_get_ptr(amp->ahp, anon_index);
9625 			if (ap == NULL) {
9626 				page_t		*pp;
9627 
9628 				pp = anon_zero(seg, addr, &ap, svd->cred);
9629 				if (pp == NULL) {
9630 					anon_array_exit(&cookie);
9631 					ANON_LOCK_EXIT(&amp->a_rwlock);
9632 					return (ENOMEM);
9633 				}
9634 				ASSERT(anon_get_ptr(amp->ahp, anon_index)
9635 				    == NULL);
9636 				(void) anon_set_ptr(amp->ahp, anon_index,
9637 				    ap, ANON_SLEEP);
9638 				page_unlock(pp);
9639 			}
9640 
9641 			anon_array_exit(&cookie);
9642 			ANON_LOCK_EXIT(&amp->a_rwlock);
9643 
9644 			memidp->val[0] = (uintptr_t)ap;
9645 			memidp->val[1] = (uintptr_t)addr & PAGEOFFSET;
9646 			return (0);
9647 		}
9648 	}
9649 	return (EINVAL);
9650 }
9651 
9652 static int
9653 sameprot(struct seg *seg, caddr_t a, size_t len)
9654 {
9655 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9656 	struct vpage *vpage;
9657 	spgcnt_t pages = btop(len);
9658 	uint_t prot;
9659 
9660 	if (svd->pageprot == 0)
9661 		return (1);
9662 
9663 	ASSERT(svd->vpage != NULL);
9664 
9665 	vpage = &svd->vpage[seg_page(seg, a)];
9666 	prot = VPP_PROT(vpage);
9667 	vpage++;
9668 	pages--;
9669 	while (pages-- > 0) {
9670 		if (prot != VPP_PROT(vpage))
9671 			return (0);
9672 		vpage++;
9673 	}
9674 	return (1);
9675 }
9676 
9677 /*
9678  * Get memory allocation policy info for specified address in given segment
9679  */
9680 static lgrp_mem_policy_info_t *
9681 segvn_getpolicy(struct seg *seg, caddr_t addr)
9682 {
9683 	struct anon_map		*amp;
9684 	ulong_t			anon_index;
9685 	lgrp_mem_policy_info_t	*policy_info;
9686 	struct segvn_data	*svn_data;
9687 	u_offset_t		vn_off;
9688 	vnode_t			*vp;
9689 
9690 	ASSERT(seg != NULL);
9691 
9692 	svn_data = (struct segvn_data *)seg->s_data;
9693 	if (svn_data == NULL)
9694 		return (NULL);
9695 
9696 	/*
9697 	 * Get policy info for private or shared memory
9698 	 */
9699 	if (svn_data->type != MAP_SHARED) {
9700 		if (svn_data->tr_state != SEGVN_TR_ON) {
9701 			policy_info = &svn_data->policy_info;
9702 		} else {
9703 			policy_info = &svn_data->tr_policy_info;
9704 			ASSERT(policy_info->mem_policy ==
9705 			    LGRP_MEM_POLICY_NEXT_SEG);
9706 		}
9707 	} else {
9708 		amp = svn_data->amp;
9709 		anon_index = svn_data->anon_index + seg_page(seg, addr);
9710 		vp = svn_data->vp;
9711 		vn_off = svn_data->offset + (uintptr_t)(addr - seg->s_base);
9712 		policy_info = lgrp_shm_policy_get(amp, anon_index, vp, vn_off);
9713 	}
9714 
9715 	return (policy_info);
9716 }
9717 
9718 /*ARGSUSED*/
9719 static int
9720 segvn_capable(struct seg *seg, segcapability_t capability)
9721 {
9722 	return (0);
9723 }
9724 
9725 /*
9726  * Bind text vnode segment to an amp. If we bind successfully mappings will be
9727  * established to per vnode mapping per lgroup amp pages instead of to vnode
9728  * pages. There's one amp per vnode text mapping per lgroup. Many processes
9729  * may share the same text replication amp. If a suitable amp doesn't already
9730  * exist in svntr hash table create a new one.  We may fail to bind to amp if
9731  * segment is not eligible for text replication.  Code below first checks for
9732  * these conditions. If binding is successful segment tr_state is set to on
9733  * and svd->amp points to the amp to use. Otherwise tr_state is set to off and
9734  * svd->amp remains as NULL.
9735  */
9736 static void
9737 segvn_textrepl(struct seg *seg)
9738 {
9739 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
9740 	vnode_t			*vp = svd->vp;
9741 	u_offset_t		off = svd->offset;
9742 	size_t			size = seg->s_size;
9743 	u_offset_t		eoff = off + size;
9744 	uint_t			szc = seg->s_szc;
9745 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9746 	svntr_t			*svntrp;
9747 	struct vattr		va;
9748 	proc_t			*p = seg->s_as->a_proc;
9749 	lgrp_id_t		lgrp_id;
9750 	lgrp_id_t		olid;
9751 	int			first;
9752 	struct anon_map		*amp;
9753 
9754 	ASSERT(AS_LOCK_HELD(seg->s_as));
9755 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9756 	ASSERT(p != NULL);
9757 	ASSERT(svd->tr_state == SEGVN_TR_INIT);
9758 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9759 	ASSERT(svd->flags & MAP_TEXT);
9760 	ASSERT(svd->type == MAP_PRIVATE);
9761 	ASSERT(vp != NULL && svd->amp == NULL);
9762 	ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
9763 	ASSERT(!(svd->flags & MAP_NORESERVE) && svd->swresv == 0);
9764 	ASSERT(seg->s_as != &kas);
9765 	ASSERT(off < eoff);
9766 	ASSERT(svntr_hashtab != NULL);
9767 
9768 	/*
9769 	 * If numa optimizations are no longer desired bail out.
9770 	 */
9771 	if (!lgrp_optimizations()) {
9772 		svd->tr_state = SEGVN_TR_OFF;
9773 		return;
9774 	}
9775 
9776 	/*
9777 	 * Avoid creating anon maps with size bigger than the file size.
9778 	 * If VOP_GETATTR() call fails bail out.
9779 	 */
9780 	va.va_mask = AT_SIZE | AT_MTIME | AT_CTIME;
9781 	if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL) != 0) {
9782 		svd->tr_state = SEGVN_TR_OFF;
9783 		SEGVN_TR_ADDSTAT(gaerr);
9784 		return;
9785 	}
9786 	if (btopr(va.va_size) < btopr(eoff)) {
9787 		svd->tr_state = SEGVN_TR_OFF;
9788 		SEGVN_TR_ADDSTAT(overmap);
9789 		return;
9790 	}
9791 
9792 	/*
9793 	 * VVMEXEC may not be set yet if exec() prefaults text segment. Set
9794 	 * this flag now before vn_is_mapped(V_WRITE) so that MAP_SHARED
9795 	 * mapping that checks if trcache for this vnode needs to be
9796 	 * invalidated can't miss us.
9797 	 */
9798 	if (!(vp->v_flag & VVMEXEC)) {
9799 		mutex_enter(&vp->v_lock);
9800 		vp->v_flag |= VVMEXEC;
9801 		mutex_exit(&vp->v_lock);
9802 	}
9803 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9804 	/*
9805 	 * Bail out if potentially MAP_SHARED writable mappings exist to this
9806 	 * vnode.  We don't want to use old file contents from existing
9807 	 * replicas if this mapping was established after the original file
9808 	 * was changed.
9809 	 */
9810 	if (vn_is_mapped(vp, V_WRITE)) {
9811 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9812 		svd->tr_state = SEGVN_TR_OFF;
9813 		SEGVN_TR_ADDSTAT(wrcnt);
9814 		return;
9815 	}
9816 	svntrp = svntr_hashtab[hash].tr_head;
9817 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9818 		ASSERT(svntrp->tr_refcnt != 0);
9819 		if (svntrp->tr_vp != vp) {
9820 			continue;
9821 		}
9822 
9823 		/*
9824 		 * Bail out if the file or its attributes were changed after
9825 		 * this replication entry was created since we need to use the
9826 		 * latest file contents. Note that mtime test alone is not
9827 		 * sufficient because a user can explicitly change mtime via
9828 		 * utimes(2) interfaces back to the old value after modifiying
9829 		 * the file contents. To detect this case we also have to test
9830 		 * ctime which among other things records the time of the last
9831 		 * mtime change by utimes(2). ctime is not changed when the file
9832 		 * is only read or executed so we expect that typically existing
9833 		 * replication amp's can be used most of the time.
9834 		 */
9835 		if (!svntrp->tr_valid ||
9836 		    svntrp->tr_mtime.tv_sec != va.va_mtime.tv_sec ||
9837 		    svntrp->tr_mtime.tv_nsec != va.va_mtime.tv_nsec ||
9838 		    svntrp->tr_ctime.tv_sec != va.va_ctime.tv_sec ||
9839 		    svntrp->tr_ctime.tv_nsec != va.va_ctime.tv_nsec) {
9840 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9841 			svd->tr_state = SEGVN_TR_OFF;
9842 			SEGVN_TR_ADDSTAT(stale);
9843 			return;
9844 		}
9845 		/*
9846 		 * if off, eoff and szc match current segment we found the
9847 		 * existing entry we can use.
9848 		 */
9849 		if (svntrp->tr_off == off && svntrp->tr_eoff == eoff &&
9850 		    svntrp->tr_szc == szc) {
9851 			break;
9852 		}
9853 		/*
9854 		 * Don't create different but overlapping in file offsets
9855 		 * entries to avoid replication of the same file pages more
9856 		 * than once per lgroup.
9857 		 */
9858 		if ((off >= svntrp->tr_off && off < svntrp->tr_eoff) ||
9859 		    (eoff > svntrp->tr_off && eoff <= svntrp->tr_eoff)) {
9860 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9861 			svd->tr_state = SEGVN_TR_OFF;
9862 			SEGVN_TR_ADDSTAT(overlap);
9863 			return;
9864 		}
9865 	}
9866 	/*
9867 	 * If we didn't find existing entry create a new one.
9868 	 */
9869 	if (svntrp == NULL) {
9870 		svntrp = kmem_cache_alloc(svntr_cache, KM_NOSLEEP);
9871 		if (svntrp == NULL) {
9872 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9873 			svd->tr_state = SEGVN_TR_OFF;
9874 			SEGVN_TR_ADDSTAT(nokmem);
9875 			return;
9876 		}
9877 #ifdef DEBUG
9878 		{
9879 			lgrp_id_t i;
9880 			for (i = 0; i < NLGRPS_MAX; i++) {
9881 				ASSERT(svntrp->tr_amp[i] == NULL);
9882 			}
9883 		}
9884 #endif /* DEBUG */
9885 		svntrp->tr_vp = vp;
9886 		svntrp->tr_off = off;
9887 		svntrp->tr_eoff = eoff;
9888 		svntrp->tr_szc = szc;
9889 		svntrp->tr_valid = 1;
9890 		svntrp->tr_mtime = va.va_mtime;
9891 		svntrp->tr_ctime = va.va_ctime;
9892 		svntrp->tr_refcnt = 0;
9893 		svntrp->tr_next = svntr_hashtab[hash].tr_head;
9894 		svntr_hashtab[hash].tr_head = svntrp;
9895 	}
9896 	first = 1;
9897 again:
9898 	/*
9899 	 * We want to pick a replica with pages on main thread's (t_tid = 1,
9900 	 * aka T1) lgrp. Currently text replication is only optimized for
9901 	 * workloads that either have all threads of a process on the same
9902 	 * lgrp or execute their large text primarily on main thread.
9903 	 */
9904 	lgrp_id = p->p_t1_lgrpid;
9905 	if (lgrp_id == LGRP_NONE) {
9906 		/*
9907 		 * In case exec() prefaults text on non main thread use
9908 		 * current thread lgrpid.  It will become main thread anyway
9909 		 * soon.
9910 		 */
9911 		lgrp_id = lgrp_home_id(curthread);
9912 	}
9913 	/*
9914 	 * Set p_tr_lgrpid to lgrpid if it hasn't been set yet.  Otherwise
9915 	 * just set it to NLGRPS_MAX if it's different from current process T1
9916 	 * home lgrp.  p_tr_lgrpid is used to detect if process uses text
9917 	 * replication and T1 new home is different from lgrp used for text
9918 	 * replication. When this happens asyncronous segvn thread rechecks if
9919 	 * segments should change lgrps used for text replication.  If we fail
9920 	 * to set p_tr_lgrpid with atomic_cas_32 then set it to NLGRPS_MAX
9921 	 * without cas if it's not already NLGRPS_MAX and not equal lgrp_id
9922 	 * we want to use.  We don't need to use cas in this case because
9923 	 * another thread that races in between our non atomic check and set
9924 	 * may only change p_tr_lgrpid to NLGRPS_MAX at this point.
9925 	 */
9926 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9927 	olid = p->p_tr_lgrpid;
9928 	if (lgrp_id != olid && olid != NLGRPS_MAX) {
9929 		lgrp_id_t nlid = (olid == LGRP_NONE) ? lgrp_id : NLGRPS_MAX;
9930 		if (atomic_cas_32((uint32_t *)&p->p_tr_lgrpid, olid, nlid) !=
9931 		    olid) {
9932 			olid = p->p_tr_lgrpid;
9933 			ASSERT(olid != LGRP_NONE);
9934 			if (olid != lgrp_id && olid != NLGRPS_MAX) {
9935 				p->p_tr_lgrpid = NLGRPS_MAX;
9936 			}
9937 		}
9938 		ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9939 		membar_producer();
9940 		/*
9941 		 * lgrp_move_thread() won't schedule async recheck after
9942 		 * p->p_t1_lgrpid update unless p->p_tr_lgrpid is not
9943 		 * LGRP_NONE. Recheck p_t1_lgrpid once now that p->p_tr_lgrpid
9944 		 * is not LGRP_NONE.
9945 		 */
9946 		if (first && p->p_t1_lgrpid != LGRP_NONE &&
9947 		    p->p_t1_lgrpid != lgrp_id) {
9948 			first = 0;
9949 			goto again;
9950 		}
9951 	}
9952 	/*
9953 	 * If no amp was created yet for lgrp_id create a new one as long as
9954 	 * we have enough memory to afford it.
9955 	 */
9956 	if ((amp = svntrp->tr_amp[lgrp_id]) == NULL) {
9957 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9958 		if (trmem > segvn_textrepl_max_bytes) {
9959 			SEGVN_TR_ADDSTAT(normem);
9960 			goto fail;
9961 		}
9962 		if (anon_try_resv_zone(size, NULL) == 0) {
9963 			SEGVN_TR_ADDSTAT(noanon);
9964 			goto fail;
9965 		}
9966 		amp = anonmap_alloc(size, size, ANON_NOSLEEP);
9967 		if (amp == NULL) {
9968 			anon_unresv_zone(size, NULL);
9969 			SEGVN_TR_ADDSTAT(nokmem);
9970 			goto fail;
9971 		}
9972 		ASSERT(amp->refcnt == 1);
9973 		amp->a_szc = szc;
9974 		svntrp->tr_amp[lgrp_id] = amp;
9975 		SEGVN_TR_ADDSTAT(newamp);
9976 	}
9977 	svntrp->tr_refcnt++;
9978 	ASSERT(svd->svn_trnext == NULL);
9979 	ASSERT(svd->svn_trprev == NULL);
9980 	svd->svn_trnext = svntrp->tr_svnhead;
9981 	svd->svn_trprev = NULL;
9982 	if (svntrp->tr_svnhead != NULL) {
9983 		svntrp->tr_svnhead->svn_trprev = svd;
9984 	}
9985 	svntrp->tr_svnhead = svd;
9986 	ASSERT(amp->a_szc == szc && amp->size == size && amp->swresv == size);
9987 	ASSERT(amp->refcnt >= 1);
9988 	svd->amp = amp;
9989 	svd->anon_index = 0;
9990 	svd->tr_policy_info.mem_policy = LGRP_MEM_POLICY_NEXT_SEG;
9991 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
9992 	svd->tr_state = SEGVN_TR_ON;
9993 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9994 	SEGVN_TR_ADDSTAT(repl);
9995 	return;
9996 fail:
9997 	ASSERT(segvn_textrepl_bytes >= size);
9998 	atomic_add_long(&segvn_textrepl_bytes, -size);
9999 	ASSERT(svntrp != NULL);
10000 	ASSERT(svntrp->tr_amp[lgrp_id] == NULL);
10001 	if (svntrp->tr_refcnt == 0) {
10002 		ASSERT(svntrp == svntr_hashtab[hash].tr_head);
10003 		svntr_hashtab[hash].tr_head = svntrp->tr_next;
10004 		mutex_exit(&svntr_hashtab[hash].tr_lock);
10005 		kmem_cache_free(svntr_cache, svntrp);
10006 	} else {
10007 		mutex_exit(&svntr_hashtab[hash].tr_lock);
10008 	}
10009 	svd->tr_state = SEGVN_TR_OFF;
10010 }
10011 
10012 /*
10013  * Convert seg back to regular vnode mapping seg by unbinding it from its text
10014  * replication amp.  This routine is most typically called when segment is
10015  * unmapped but can also be called when segment no longer qualifies for text
10016  * replication (e.g. due to protection changes). If unload_unmap is set use
10017  * HAT_UNLOAD_UNMAP flag in hat_unload_callback().  If we are the last user of
10018  * svntr free all its anon maps and remove it from the hash table.
10019  */
10020 static void
10021 segvn_textunrepl(struct seg *seg, int unload_unmap)
10022 {
10023 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
10024 	vnode_t			*vp = svd->vp;
10025 	u_offset_t		off = svd->offset;
10026 	size_t			size = seg->s_size;
10027 	u_offset_t		eoff = off + size;
10028 	uint_t			szc = seg->s_szc;
10029 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
10030 	svntr_t			*svntrp;
10031 	svntr_t			**prv_svntrp;
10032 	lgrp_id_t		lgrp_id = svd->tr_policy_info.mem_lgrpid;
10033 	lgrp_id_t		i;
10034 
10035 	ASSERT(AS_LOCK_HELD(seg->s_as));
10036 	ASSERT(AS_WRITE_HELD(seg->s_as) ||
10037 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
10038 	ASSERT(svd->tr_state == SEGVN_TR_ON);
10039 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
10040 	ASSERT(svd->amp != NULL);
10041 	ASSERT(svd->amp->refcnt >= 1);
10042 	ASSERT(svd->anon_index == 0);
10043 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
10044 	ASSERT(svntr_hashtab != NULL);
10045 
10046 	mutex_enter(&svntr_hashtab[hash].tr_lock);
10047 	prv_svntrp = &svntr_hashtab[hash].tr_head;
10048 	for (; (svntrp = *prv_svntrp) != NULL; prv_svntrp = &svntrp->tr_next) {
10049 		ASSERT(svntrp->tr_refcnt != 0);
10050 		if (svntrp->tr_vp == vp && svntrp->tr_off == off &&
10051 		    svntrp->tr_eoff == eoff && svntrp->tr_szc == szc) {
10052 			break;
10053 		}
10054 	}
10055 	if (svntrp == NULL) {
10056 		panic("segvn_textunrepl: svntr record not found");
10057 	}
10058 	if (svntrp->tr_amp[lgrp_id] != svd->amp) {
10059 		panic("segvn_textunrepl: amp mismatch");
10060 	}
10061 	svd->tr_state = SEGVN_TR_OFF;
10062 	svd->amp = NULL;
10063 	if (svd->svn_trprev == NULL) {
10064 		ASSERT(svntrp->tr_svnhead == svd);
10065 		svntrp->tr_svnhead = svd->svn_trnext;
10066 		if (svntrp->tr_svnhead != NULL) {
10067 			svntrp->tr_svnhead->svn_trprev = NULL;
10068 		}
10069 		svd->svn_trnext = NULL;
10070 	} else {
10071 		svd->svn_trprev->svn_trnext = svd->svn_trnext;
10072 		if (svd->svn_trnext != NULL) {
10073 			svd->svn_trnext->svn_trprev = svd->svn_trprev;
10074 			svd->svn_trnext = NULL;
10075 		}
10076 		svd->svn_trprev = NULL;
10077 	}
10078 	if (--svntrp->tr_refcnt) {
10079 		mutex_exit(&svntr_hashtab[hash].tr_lock);
10080 		goto done;
10081 	}
10082 	*prv_svntrp = svntrp->tr_next;
10083 	mutex_exit(&svntr_hashtab[hash].tr_lock);
10084 	for (i = 0; i < NLGRPS_MAX; i++) {
10085 		struct anon_map *amp = svntrp->tr_amp[i];
10086 		if (amp == NULL) {
10087 			continue;
10088 		}
10089 		ASSERT(amp->refcnt == 1);
10090 		ASSERT(amp->swresv == size);
10091 		ASSERT(amp->size == size);
10092 		ASSERT(amp->a_szc == szc);
10093 		if (amp->a_szc != 0) {
10094 			anon_free_pages(amp->ahp, 0, size, szc);
10095 		} else {
10096 			anon_free(amp->ahp, 0, size);
10097 		}
10098 		svntrp->tr_amp[i] = NULL;
10099 		ASSERT(segvn_textrepl_bytes >= size);
10100 		atomic_add_long(&segvn_textrepl_bytes, -size);
10101 		anon_unresv_zone(amp->swresv, NULL);
10102 		amp->refcnt = 0;
10103 		anonmap_free(amp);
10104 	}
10105 	kmem_cache_free(svntr_cache, svntrp);
10106 done:
10107 	hat_unload_callback(seg->s_as->a_hat, seg->s_base, size,
10108 	    unload_unmap ? HAT_UNLOAD_UNMAP : 0, NULL);
10109 }
10110 
10111 /*
10112  * This is called when a MAP_SHARED writable mapping is created to a vnode
10113  * that is currently used for execution (VVMEXEC flag is set). In this case we
10114  * need to prevent further use of existing replicas.
10115  */
10116 static void
10117 segvn_inval_trcache(vnode_t *vp)
10118 {
10119 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
10120 	svntr_t			*svntrp;
10121 
10122 	ASSERT(vp->v_flag & VVMEXEC);
10123 
10124 	if (svntr_hashtab == NULL) {
10125 		return;
10126 	}
10127 
10128 	mutex_enter(&svntr_hashtab[hash].tr_lock);
10129 	svntrp = svntr_hashtab[hash].tr_head;
10130 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
10131 		ASSERT(svntrp->tr_refcnt != 0);
10132 		if (svntrp->tr_vp == vp && svntrp->tr_valid) {
10133 			svntrp->tr_valid = 0;
10134 		}
10135 	}
10136 	mutex_exit(&svntr_hashtab[hash].tr_lock);
10137 }
10138 
10139 static void
10140 segvn_trasync_thread(void)
10141 {
10142 	callb_cpr_t cpr_info;
10143 	kmutex_t cpr_lock;	/* just for CPR stuff */
10144 
10145 	mutex_init(&cpr_lock, NULL, MUTEX_DEFAULT, NULL);
10146 
10147 	CALLB_CPR_INIT(&cpr_info, &cpr_lock,
10148 	    callb_generic_cpr, "segvn_async");
10149 
10150 	if (segvn_update_textrepl_interval == 0) {
10151 		segvn_update_textrepl_interval = segvn_update_tr_time * hz;
10152 	} else {
10153 		segvn_update_textrepl_interval *= hz;
10154 	}
10155 	(void) timeout(segvn_trupdate_wakeup, NULL,
10156 	    segvn_update_textrepl_interval);
10157 
10158 	for (;;) {
10159 		mutex_enter(&cpr_lock);
10160 		CALLB_CPR_SAFE_BEGIN(&cpr_info);
10161 		mutex_exit(&cpr_lock);
10162 		sema_p(&segvn_trasync_sem);
10163 		mutex_enter(&cpr_lock);
10164 		CALLB_CPR_SAFE_END(&cpr_info, &cpr_lock);
10165 		mutex_exit(&cpr_lock);
10166 		segvn_trupdate();
10167 	}
10168 }
10169 
10170 static uint64_t segvn_lgrp_trthr_migrs_snpsht = 0;
10171 
10172 static void
10173 segvn_trupdate_wakeup(void *dummy)
10174 {
10175 	uint64_t cur_lgrp_trthr_migrs = lgrp_get_trthr_migrations();
10176 
10177 	if (cur_lgrp_trthr_migrs != segvn_lgrp_trthr_migrs_snpsht) {
10178 		segvn_lgrp_trthr_migrs_snpsht = cur_lgrp_trthr_migrs;
10179 		sema_v(&segvn_trasync_sem);
10180 	}
10181 
10182 	if (!segvn_disable_textrepl_update &&
10183 	    segvn_update_textrepl_interval != 0) {
10184 		(void) timeout(segvn_trupdate_wakeup, dummy,
10185 		    segvn_update_textrepl_interval);
10186 	}
10187 }
10188 
10189 static void
10190 segvn_trupdate(void)
10191 {
10192 	ulong_t		hash;
10193 	svntr_t		*svntrp;
10194 	segvn_data_t	*svd;
10195 
10196 	ASSERT(svntr_hashtab != NULL);
10197 
10198 	for (hash = 0; hash < svntr_hashtab_sz; hash++) {
10199 		mutex_enter(&svntr_hashtab[hash].tr_lock);
10200 		svntrp = svntr_hashtab[hash].tr_head;
10201 		for (; svntrp != NULL; svntrp = svntrp->tr_next) {
10202 			ASSERT(svntrp->tr_refcnt != 0);
10203 			svd = svntrp->tr_svnhead;
10204 			for (; svd != NULL; svd = svd->svn_trnext) {
10205 				segvn_trupdate_seg(svd->seg, svd, svntrp,
10206 				    hash);
10207 			}
10208 		}
10209 		mutex_exit(&svntr_hashtab[hash].tr_lock);
10210 	}
10211 }
10212 
10213 static void
10214 segvn_trupdate_seg(struct seg *seg, segvn_data_t *svd, svntr_t *svntrp,
10215     ulong_t hash)
10216 {
10217 	proc_t			*p;
10218 	lgrp_id_t		lgrp_id;
10219 	struct as		*as;
10220 	size_t			size;
10221 	struct anon_map		*amp;
10222 
10223 	ASSERT(svd->vp != NULL);
10224 	ASSERT(svd->vp == svntrp->tr_vp);
10225 	ASSERT(svd->offset == svntrp->tr_off);
10226 	ASSERT(svd->offset + seg->s_size == svntrp->tr_eoff);
10227 	ASSERT(seg != NULL);
10228 	ASSERT(svd->seg == seg);
10229 	ASSERT(seg->s_data == (void *)svd);
10230 	ASSERT(seg->s_szc == svntrp->tr_szc);
10231 	ASSERT(svd->tr_state == SEGVN_TR_ON);
10232 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
10233 	ASSERT(svd->amp != NULL);
10234 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
10235 	ASSERT(svd->tr_policy_info.mem_lgrpid != LGRP_NONE);
10236 	ASSERT(svd->tr_policy_info.mem_lgrpid < NLGRPS_MAX);
10237 	ASSERT(svntrp->tr_amp[svd->tr_policy_info.mem_lgrpid] == svd->amp);
10238 	ASSERT(svntrp->tr_refcnt != 0);
10239 	ASSERT(mutex_owned(&svntr_hashtab[hash].tr_lock));
10240 
10241 	as = seg->s_as;
10242 	ASSERT(as != NULL && as != &kas);
10243 	p = as->a_proc;
10244 	ASSERT(p != NULL);
10245 	ASSERT(p->p_tr_lgrpid != LGRP_NONE);
10246 	lgrp_id = p->p_t1_lgrpid;
10247 	if (lgrp_id == LGRP_NONE) {
10248 		return;
10249 	}
10250 	ASSERT(lgrp_id < NLGRPS_MAX);
10251 	if (svd->tr_policy_info.mem_lgrpid == lgrp_id) {
10252 		return;
10253 	}
10254 
10255 	/*
10256 	 * Use tryenter locking since we are locking as/seg and svntr hash
10257 	 * lock in reverse from syncrounous thread order.
10258 	 */
10259 	if (!AS_LOCK_TRYENTER(as, RW_READER)) {
10260 		SEGVN_TR_ADDSTAT(nolock);
10261 		if (segvn_lgrp_trthr_migrs_snpsht) {
10262 			segvn_lgrp_trthr_migrs_snpsht = 0;
10263 		}
10264 		return;
10265 	}
10266 	if (!SEGVN_LOCK_TRYENTER(seg->s_as, &svd->lock, RW_WRITER)) {
10267 		AS_LOCK_EXIT(as);
10268 		SEGVN_TR_ADDSTAT(nolock);
10269 		if (segvn_lgrp_trthr_migrs_snpsht) {
10270 			segvn_lgrp_trthr_migrs_snpsht = 0;
10271 		}
10272 		return;
10273 	}
10274 	size = seg->s_size;
10275 	if (svntrp->tr_amp[lgrp_id] == NULL) {
10276 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
10277 		if (trmem > segvn_textrepl_max_bytes) {
10278 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10279 			AS_LOCK_EXIT(as);
10280 			atomic_add_long(&segvn_textrepl_bytes, -size);
10281 			SEGVN_TR_ADDSTAT(normem);
10282 			return;
10283 		}
10284 		if (anon_try_resv_zone(size, NULL) == 0) {
10285 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10286 			AS_LOCK_EXIT(as);
10287 			atomic_add_long(&segvn_textrepl_bytes, -size);
10288 			SEGVN_TR_ADDSTAT(noanon);
10289 			return;
10290 		}
10291 		amp = anonmap_alloc(size, size, KM_NOSLEEP);
10292 		if (amp == NULL) {
10293 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10294 			AS_LOCK_EXIT(as);
10295 			atomic_add_long(&segvn_textrepl_bytes, -size);
10296 			anon_unresv_zone(size, NULL);
10297 			SEGVN_TR_ADDSTAT(nokmem);
10298 			return;
10299 		}
10300 		ASSERT(amp->refcnt == 1);
10301 		amp->a_szc = seg->s_szc;
10302 		svntrp->tr_amp[lgrp_id] = amp;
10303 	}
10304 	/*
10305 	 * We don't need to drop the bucket lock but here we give other
10306 	 * threads a chance.  svntr and svd can't be unlinked as long as
10307 	 * segment lock is held as a writer and AS held as well.  After we
10308 	 * retake bucket lock we'll continue from where we left. We'll be able
10309 	 * to reach the end of either list since new entries are always added
10310 	 * to the beginning of the lists.
10311 	 */
10312 	mutex_exit(&svntr_hashtab[hash].tr_lock);
10313 	hat_unload_callback(as->a_hat, seg->s_base, size, 0, NULL);
10314 	mutex_enter(&svntr_hashtab[hash].tr_lock);
10315 
10316 	ASSERT(svd->tr_state == SEGVN_TR_ON);
10317 	ASSERT(svd->amp != NULL);
10318 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
10319 	ASSERT(svd->tr_policy_info.mem_lgrpid != lgrp_id);
10320 	ASSERT(svd->amp != svntrp->tr_amp[lgrp_id]);
10321 
10322 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
10323 	svd->amp = svntrp->tr_amp[lgrp_id];
10324 	p->p_tr_lgrpid = NLGRPS_MAX;
10325 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10326 	AS_LOCK_EXIT(as);
10327 
10328 	ASSERT(svntrp->tr_refcnt != 0);
10329 	ASSERT(svd->vp == svntrp->tr_vp);
10330 	ASSERT(svd->tr_policy_info.mem_lgrpid == lgrp_id);
10331 	ASSERT(svd->amp != NULL && svd->amp == svntrp->tr_amp[lgrp_id]);
10332 	ASSERT(svd->seg == seg);
10333 	ASSERT(svd->tr_state == SEGVN_TR_ON);
10334 
10335 	SEGVN_TR_ADDSTAT(asyncrepl);
10336 }
10337