xref: /titanic_50/usr/src/uts/common/vm/seg_vn.c (revision 1665390b481087ca2aaff06a502c62b3ee91d530)
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 2015, Joyent, Inc. All rights reserved.
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 *seg, void *argsp)
542 {
543 	struct segvn_crargs *a = (struct segvn_crargs *)argsp;
544 	struct segvn_data *svd;
545 	size_t swresv = 0;
546 	struct cred *cred;
547 	struct anon_map *amp;
548 	int error = 0;
549 	size_t pgsz;
550 	lgrp_mem_policy_t mpolicy = LGRP_MEM_POLICY_DEFAULT;
551 	int use_rgn = 0;
552 	int trok = 0;
553 
554 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
555 
556 	if (a->type != MAP_PRIVATE && a->type != MAP_SHARED) {
557 		panic("segvn_create type");
558 		/*NOTREACHED*/
559 	}
560 
561 	/*
562 	 * Check arguments.  If a shared anon structure is given then
563 	 * it is illegal to also specify a vp.
564 	 */
565 	if (a->amp != NULL && a->vp != NULL) {
566 		panic("segvn_create anon_map");
567 		/*NOTREACHED*/
568 	}
569 
570 	if (a->type == MAP_PRIVATE && (a->flags & MAP_TEXT) &&
571 	    a->vp != NULL && a->prot == (PROT_USER | PROT_READ | PROT_EXEC) &&
572 	    segvn_use_regions) {
573 		use_rgn = 1;
574 	}
575 
576 	/* MAP_NORESERVE on a MAP_SHARED segment is meaningless. */
577 	if (a->type == MAP_SHARED)
578 		a->flags &= ~MAP_NORESERVE;
579 
580 	if (a->szc != 0) {
581 		if (segvn_lpg_disable != 0 || (a->szc == AS_MAP_NO_LPOOB) ||
582 		    (a->amp != NULL && a->type == MAP_PRIVATE) ||
583 		    (a->flags & MAP_NORESERVE) || seg->s_as == &kas) {
584 			a->szc = 0;
585 		} else {
586 			if (a->szc > segvn_maxpgszc)
587 				a->szc = segvn_maxpgszc;
588 			pgsz = page_get_pagesize(a->szc);
589 			if (!IS_P2ALIGNED(seg->s_base, pgsz) ||
590 			    !IS_P2ALIGNED(seg->s_size, pgsz)) {
591 				a->szc = 0;
592 			} else if (a->vp != NULL) {
593 				if (IS_SWAPFSVP(a->vp) || VN_ISKAS(a->vp)) {
594 					/*
595 					 * paranoid check.
596 					 * hat_page_demote() is not supported
597 					 * on swapfs pages.
598 					 */
599 					a->szc = 0;
600 				} else if (map_addr_vacalign_check(seg->s_base,
601 				    a->offset & PAGEMASK)) {
602 					a->szc = 0;
603 				}
604 			} else if (a->amp != NULL) {
605 				pgcnt_t anum = btopr(a->offset);
606 				pgcnt_t pgcnt = page_get_pagecnt(a->szc);
607 				if (!IS_P2ALIGNED(anum, pgcnt)) {
608 					a->szc = 0;
609 				}
610 			}
611 		}
612 	}
613 
614 	/*
615 	 * If segment may need private pages, reserve them now.
616 	 */
617 	if (!(a->flags & MAP_NORESERVE) && ((a->vp == NULL && a->amp == NULL) ||
618 	    (a->type == MAP_PRIVATE && (a->prot & PROT_WRITE)))) {
619 		if (anon_resv_zone(seg->s_size,
620 		    seg->s_as->a_proc->p_zone) == 0)
621 			return (EAGAIN);
622 		swresv = seg->s_size;
623 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
624 		    seg, swresv, 1);
625 	}
626 
627 	/*
628 	 * Reserve any mapping structures that may be required.
629 	 *
630 	 * Don't do it for segments that may use regions. It's currently a
631 	 * noop in the hat implementations anyway.
632 	 */
633 	if (!use_rgn) {
634 		hat_map(seg->s_as->a_hat, seg->s_base, seg->s_size, HAT_MAP);
635 	}
636 
637 	if (a->cred) {
638 		cred = a->cred;
639 		crhold(cred);
640 	} else {
641 		crhold(cred = CRED());
642 	}
643 
644 	/* Inform the vnode of the new mapping */
645 	if (a->vp != NULL) {
646 		error = VOP_ADDMAP(a->vp, a->offset & PAGEMASK,
647 		    seg->s_as, seg->s_base, seg->s_size, a->prot,
648 		    a->maxprot, a->type, cred, NULL);
649 		if (error) {
650 			if (swresv != 0) {
651 				anon_unresv_zone(swresv,
652 				    seg->s_as->a_proc->p_zone);
653 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
654 				    "anon proc:%p %lu %u", seg, swresv, 0);
655 			}
656 			crfree(cred);
657 			if (!use_rgn) {
658 				hat_unload(seg->s_as->a_hat, seg->s_base,
659 				    seg->s_size, HAT_UNLOAD_UNMAP);
660 			}
661 			return (error);
662 		}
663 		/*
664 		 * svntr_hashtab will be NULL if we support shared regions.
665 		 */
666 		trok = ((a->flags & MAP_TEXT) &&
667 		    (seg->s_size > textrepl_size_thresh ||
668 		    (a->flags & _MAP_TEXTREPL)) &&
669 		    lgrp_optimizations() && svntr_hashtab != NULL &&
670 		    a->type == MAP_PRIVATE && swresv == 0 &&
671 		    !(a->flags & MAP_NORESERVE) &&
672 		    seg->s_as != &kas && a->vp->v_type == VREG);
673 
674 		ASSERT(!trok || !use_rgn);
675 	}
676 
677 	/*
678 	 * MAP_NORESERVE mappings don't count towards the VSZ of a process
679 	 * until we fault the pages in.
680 	 */
681 	if ((a->vp == NULL || a->vp->v_type != VREG) &&
682 	    a->flags & MAP_NORESERVE) {
683 		seg->s_as->a_resvsize -= seg->s_size;
684 	}
685 
686 	/*
687 	 * If more than one segment in the address space, and they're adjacent
688 	 * virtually, try to concatenate them.  Don't concatenate if an
689 	 * explicit anon_map structure was supplied (e.g., SystemV shared
690 	 * memory) or if we'll use text replication for this segment.
691 	 */
692 	if (a->amp == NULL && !use_rgn && !trok) {
693 		struct seg *pseg, *nseg;
694 		struct segvn_data *psvd, *nsvd;
695 		lgrp_mem_policy_t ppolicy, npolicy;
696 		uint_t	lgrp_mem_policy_flags = 0;
697 		extern lgrp_mem_policy_t lgrp_mem_default_policy;
698 
699 		/*
700 		 * Memory policy flags (lgrp_mem_policy_flags) is valid when
701 		 * extending stack/heap segments.
702 		 */
703 		if ((a->vp == NULL) && (a->type == MAP_PRIVATE) &&
704 		    !(a->flags & MAP_NORESERVE) && (seg->s_as != &kas)) {
705 			lgrp_mem_policy_flags = a->lgrp_mem_policy_flags;
706 		} else {
707 			/*
708 			 * Get policy when not extending it from another segment
709 			 */
710 			mpolicy = lgrp_mem_policy_default(seg->s_size, a->type);
711 		}
712 
713 		/*
714 		 * First, try to concatenate the previous and new segments
715 		 */
716 		pseg = AS_SEGPREV(seg->s_as, seg);
717 		if (pseg != NULL &&
718 		    pseg->s_base + pseg->s_size == seg->s_base &&
719 		    pseg->s_ops == &segvn_ops) {
720 			/*
721 			 * Get memory allocation policy from previous segment.
722 			 * When extension is specified (e.g. for heap) apply
723 			 * this policy to the new segment regardless of the
724 			 * outcome of segment concatenation.  Extension occurs
725 			 * for non-default policy otherwise default policy is
726 			 * used and is based on extended segment size.
727 			 */
728 			psvd = (struct segvn_data *)pseg->s_data;
729 			ppolicy = psvd->policy_info.mem_policy;
730 			if (lgrp_mem_policy_flags ==
731 			    LGRP_MP_FLAG_EXTEND_UP) {
732 				if (ppolicy != lgrp_mem_default_policy) {
733 					mpolicy = ppolicy;
734 				} else {
735 					mpolicy = lgrp_mem_policy_default(
736 					    pseg->s_size + seg->s_size,
737 					    a->type);
738 				}
739 			}
740 
741 			if (mpolicy == ppolicy &&
742 			    (pseg->s_size + seg->s_size <=
743 			    segvn_comb_thrshld || psvd->amp == NULL) &&
744 			    segvn_extend_prev(pseg, seg, a, swresv) == 0) {
745 				/*
746 				 * success! now try to concatenate
747 				 * with following seg
748 				 */
749 				crfree(cred);
750 				nseg = AS_SEGNEXT(pseg->s_as, pseg);
751 				if (nseg != NULL &&
752 				    nseg != pseg &&
753 				    nseg->s_ops == &segvn_ops &&
754 				    pseg->s_base + pseg->s_size ==
755 				    nseg->s_base)
756 					(void) segvn_concat(pseg, nseg, 0);
757 				ASSERT(pseg->s_szc == 0 ||
758 				    (a->szc == pseg->s_szc &&
759 				    IS_P2ALIGNED(pseg->s_base, pgsz) &&
760 				    IS_P2ALIGNED(pseg->s_size, pgsz)));
761 				return (0);
762 			}
763 		}
764 
765 		/*
766 		 * Failed, so try to concatenate with following seg
767 		 */
768 		nseg = AS_SEGNEXT(seg->s_as, seg);
769 		if (nseg != NULL &&
770 		    seg->s_base + seg->s_size == nseg->s_base &&
771 		    nseg->s_ops == &segvn_ops) {
772 			/*
773 			 * Get memory allocation policy from next segment.
774 			 * When extension is specified (e.g. for stack) apply
775 			 * this policy to the new segment regardless of the
776 			 * outcome of segment concatenation.  Extension occurs
777 			 * for non-default policy otherwise default policy is
778 			 * used and is based on extended segment size.
779 			 */
780 			nsvd = (struct segvn_data *)nseg->s_data;
781 			npolicy = nsvd->policy_info.mem_policy;
782 			if (lgrp_mem_policy_flags ==
783 			    LGRP_MP_FLAG_EXTEND_DOWN) {
784 				if (npolicy != lgrp_mem_default_policy) {
785 					mpolicy = npolicy;
786 				} else {
787 					mpolicy = lgrp_mem_policy_default(
788 					    nseg->s_size + seg->s_size,
789 					    a->type);
790 				}
791 			}
792 
793 			if (mpolicy == npolicy &&
794 			    segvn_extend_next(seg, nseg, a, swresv) == 0) {
795 				crfree(cred);
796 				ASSERT(nseg->s_szc == 0 ||
797 				    (a->szc == nseg->s_szc &&
798 				    IS_P2ALIGNED(nseg->s_base, pgsz) &&
799 				    IS_P2ALIGNED(nseg->s_size, pgsz)));
800 				return (0);
801 			}
802 		}
803 	}
804 
805 	if (a->vp != NULL) {
806 		VN_HOLD(a->vp);
807 		if (a->type == MAP_SHARED)
808 			lgrp_shm_policy_init(NULL, a->vp);
809 	}
810 	svd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
811 
812 	seg->s_ops = &segvn_ops;
813 	seg->s_data = (void *)svd;
814 	seg->s_szc = a->szc;
815 
816 	svd->seg = seg;
817 	svd->vp = a->vp;
818 	/*
819 	 * Anonymous mappings have no backing file so the offset is meaningless.
820 	 */
821 	svd->offset = a->vp ? (a->offset & PAGEMASK) : 0;
822 	svd->prot = a->prot;
823 	svd->maxprot = a->maxprot;
824 	svd->pageprot = 0;
825 	svd->type = a->type;
826 	svd->vpage = NULL;
827 	svd->cred = cred;
828 	svd->advice = MADV_NORMAL;
829 	svd->pageadvice = 0;
830 	svd->flags = (ushort_t)a->flags;
831 	svd->softlockcnt = 0;
832 	svd->softlockcnt_sbase = 0;
833 	svd->softlockcnt_send = 0;
834 	svd->svn_inz = 0;
835 	svd->rcookie = HAT_INVALID_REGION_COOKIE;
836 	svd->pageswap = 0;
837 
838 	if (a->szc != 0 && a->vp != NULL) {
839 		segvn_setvnode_mpss(a->vp);
840 	}
841 	if (svd->type == MAP_SHARED && svd->vp != NULL &&
842 	    (svd->vp->v_flag & VVMEXEC) && (svd->prot & PROT_WRITE)) {
843 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
844 		segvn_inval_trcache(svd->vp);
845 	}
846 
847 	amp = a->amp;
848 	if ((svd->amp = amp) == NULL) {
849 		svd->anon_index = 0;
850 		if (svd->type == MAP_SHARED) {
851 			svd->swresv = 0;
852 			/*
853 			 * Shared mappings to a vp need no other setup.
854 			 * If we have a shared mapping to an anon_map object
855 			 * which hasn't been allocated yet,  allocate the
856 			 * struct now so that it will be properly shared
857 			 * by remembering the swap reservation there.
858 			 */
859 			if (a->vp == NULL) {
860 				svd->amp = anonmap_alloc(seg->s_size, swresv,
861 				    ANON_SLEEP);
862 				svd->amp->a_szc = seg->s_szc;
863 			}
864 		} else {
865 			/*
866 			 * Private mapping (with or without a vp).
867 			 * Allocate anon_map when needed.
868 			 */
869 			svd->swresv = swresv;
870 		}
871 	} else {
872 		pgcnt_t anon_num;
873 
874 		/*
875 		 * Mapping to an existing anon_map structure without a vp.
876 		 * For now we will insure that the segment size isn't larger
877 		 * than the size - offset gives us.  Later on we may wish to
878 		 * have the anon array dynamically allocated itself so that
879 		 * we don't always have to allocate all the anon pointer slots.
880 		 * This of course involves adding extra code to check that we
881 		 * aren't trying to use an anon pointer slot beyond the end
882 		 * of the currently allocated anon array.
883 		 */
884 		if ((amp->size - a->offset) < seg->s_size) {
885 			panic("segvn_create anon_map size");
886 			/*NOTREACHED*/
887 		}
888 
889 		anon_num = btopr(a->offset);
890 
891 		if (a->type == MAP_SHARED) {
892 			/*
893 			 * SHARED mapping to a given anon_map.
894 			 */
895 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
896 			amp->refcnt++;
897 			if (a->szc > amp->a_szc) {
898 				amp->a_szc = a->szc;
899 			}
900 			ANON_LOCK_EXIT(&amp->a_rwlock);
901 			svd->anon_index = anon_num;
902 			svd->swresv = 0;
903 		} else {
904 			/*
905 			 * PRIVATE mapping to a given anon_map.
906 			 * Make sure that all the needed anon
907 			 * structures are created (so that we will
908 			 * share the underlying pages if nothing
909 			 * is written by this mapping) and then
910 			 * duplicate the anon array as is done
911 			 * when a privately mapped segment is dup'ed.
912 			 */
913 			struct anon *ap;
914 			caddr_t addr;
915 			caddr_t eaddr;
916 			ulong_t	anon_idx;
917 			int hat_flag = HAT_LOAD;
918 
919 			if (svd->flags & MAP_TEXT) {
920 				hat_flag |= HAT_LOAD_TEXT;
921 			}
922 
923 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
924 			svd->amp->a_szc = seg->s_szc;
925 			svd->anon_index = 0;
926 			svd->swresv = swresv;
927 
928 			/*
929 			 * Prevent 2 threads from allocating anon
930 			 * slots simultaneously.
931 			 */
932 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
933 			eaddr = seg->s_base + seg->s_size;
934 
935 			for (anon_idx = anon_num, addr = seg->s_base;
936 			    addr < eaddr; addr += PAGESIZE, anon_idx++) {
937 				page_t *pp;
938 
939 				if ((ap = anon_get_ptr(amp->ahp,
940 				    anon_idx)) != NULL)
941 					continue;
942 
943 				/*
944 				 * Allocate the anon struct now.
945 				 * Might as well load up translation
946 				 * to the page while we're at it...
947 				 */
948 				pp = anon_zero(seg, addr, &ap, cred);
949 				if (ap == NULL || pp == NULL) {
950 					panic("segvn_create anon_zero");
951 					/*NOTREACHED*/
952 				}
953 
954 				/*
955 				 * Re-acquire the anon_map lock and
956 				 * initialize the anon array entry.
957 				 */
958 				ASSERT(anon_get_ptr(amp->ahp,
959 				    anon_idx) == NULL);
960 				(void) anon_set_ptr(amp->ahp, anon_idx, ap,
961 				    ANON_SLEEP);
962 
963 				ASSERT(seg->s_szc == 0);
964 				ASSERT(!IS_VMODSORT(pp->p_vnode));
965 
966 				ASSERT(use_rgn == 0);
967 				hat_memload(seg->s_as->a_hat, addr, pp,
968 				    svd->prot & ~PROT_WRITE, hat_flag);
969 
970 				page_unlock(pp);
971 			}
972 			ASSERT(seg->s_szc == 0);
973 			anon_dup(amp->ahp, anon_num, svd->amp->ahp,
974 			    0, seg->s_size);
975 			ANON_LOCK_EXIT(&amp->a_rwlock);
976 		}
977 	}
978 
979 	/*
980 	 * Set default memory allocation policy for segment
981 	 *
982 	 * Always set policy for private memory at least for initialization
983 	 * even if this is a shared memory segment
984 	 */
985 	(void) lgrp_privm_policy_set(mpolicy, &svd->policy_info, seg->s_size);
986 
987 	if (svd->type == MAP_SHARED)
988 		(void) lgrp_shm_policy_set(mpolicy, svd->amp, svd->anon_index,
989 		    svd->vp, svd->offset, seg->s_size);
990 
991 	if (use_rgn) {
992 		ASSERT(!trok);
993 		ASSERT(svd->amp == NULL);
994 		svd->rcookie = hat_join_region(seg->s_as->a_hat, seg->s_base,
995 		    seg->s_size, (void *)svd->vp, svd->offset, svd->prot,
996 		    (uchar_t)seg->s_szc, segvn_hat_rgn_unload_callback,
997 		    HAT_REGION_TEXT);
998 	}
999 
1000 	ASSERT(!trok || !(svd->prot & PROT_WRITE));
1001 	svd->tr_state = trok ? SEGVN_TR_INIT : SEGVN_TR_OFF;
1002 
1003 	return (0);
1004 }
1005 
1006 /*
1007  * Concatenate two existing segments, if possible.
1008  * Return 0 on success, -1 if two segments are not compatible
1009  * or -2 on memory allocation failure.
1010  * If amp_cat == 1 then try and concat segments with anon maps
1011  */
1012 static int
1013 segvn_concat(struct seg *seg1, struct seg *seg2, int amp_cat)
1014 {
1015 	struct segvn_data *svd1 = seg1->s_data;
1016 	struct segvn_data *svd2 = seg2->s_data;
1017 	struct anon_map *amp1 = svd1->amp;
1018 	struct anon_map *amp2 = svd2->amp;
1019 	struct vpage *vpage1 = svd1->vpage;
1020 	struct vpage *vpage2 = svd2->vpage, *nvpage = NULL;
1021 	size_t size, nvpsize;
1022 	pgcnt_t npages1, npages2;
1023 
1024 	ASSERT(seg1->s_as && seg2->s_as && seg1->s_as == seg2->s_as);
1025 	ASSERT(AS_WRITE_HELD(seg1->s_as));
1026 	ASSERT(seg1->s_ops == seg2->s_ops);
1027 
1028 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie) ||
1029 	    HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1030 		return (-1);
1031 	}
1032 
1033 	/* both segments exist, try to merge them */
1034 #define	incompat(x)	(svd1->x != svd2->x)
1035 	if (incompat(vp) || incompat(maxprot) ||
1036 	    (!svd1->pageadvice && !svd2->pageadvice && incompat(advice)) ||
1037 	    (!svd1->pageprot && !svd2->pageprot && incompat(prot)) ||
1038 	    incompat(type) || incompat(cred) || incompat(flags) ||
1039 	    seg1->s_szc != seg2->s_szc || incompat(policy_info.mem_policy) ||
1040 	    (svd2->softlockcnt > 0) || svd1->softlockcnt_send > 0)
1041 		return (-1);
1042 #undef incompat
1043 
1044 	/*
1045 	 * vp == NULL implies zfod, offset doesn't matter
1046 	 */
1047 	if (svd1->vp != NULL &&
1048 	    svd1->offset + seg1->s_size != svd2->offset) {
1049 		return (-1);
1050 	}
1051 
1052 	/*
1053 	 * Don't concatenate if either segment uses text replication.
1054 	 */
1055 	if (svd1->tr_state != SEGVN_TR_OFF || svd2->tr_state != SEGVN_TR_OFF) {
1056 		return (-1);
1057 	}
1058 
1059 	/*
1060 	 * Fail early if we're not supposed to concatenate
1061 	 * segments with non NULL amp.
1062 	 */
1063 	if (amp_cat == 0 && (amp1 != NULL || amp2 != NULL)) {
1064 		return (-1);
1065 	}
1066 
1067 	if (svd1->vp == NULL && svd1->type == MAP_SHARED) {
1068 		if (amp1 != amp2) {
1069 			return (-1);
1070 		}
1071 		if (amp1 != NULL && svd1->anon_index + btop(seg1->s_size) !=
1072 		    svd2->anon_index) {
1073 			return (-1);
1074 		}
1075 		ASSERT(amp1 == NULL || amp1->refcnt >= 2);
1076 	}
1077 
1078 	/*
1079 	 * If either seg has vpages, create a new merged vpage array.
1080 	 */
1081 	if (vpage1 != NULL || vpage2 != NULL) {
1082 		struct vpage *vp, *evp;
1083 
1084 		npages1 = seg_pages(seg1);
1085 		npages2 = seg_pages(seg2);
1086 		nvpsize = vpgtob(npages1 + npages2);
1087 
1088 		if ((nvpage = kmem_zalloc(nvpsize, KM_NOSLEEP)) == NULL) {
1089 			return (-2);
1090 		}
1091 
1092 		if (vpage1 != NULL) {
1093 			bcopy(vpage1, nvpage, vpgtob(npages1));
1094 		} else {
1095 			evp = nvpage + npages1;
1096 			for (vp = nvpage; vp < evp; vp++) {
1097 				VPP_SETPROT(vp, svd1->prot);
1098 				VPP_SETADVICE(vp, svd1->advice);
1099 			}
1100 		}
1101 
1102 		if (vpage2 != NULL) {
1103 			bcopy(vpage2, nvpage + npages1, vpgtob(npages2));
1104 		} else {
1105 			evp = nvpage + npages1 + npages2;
1106 			for (vp = nvpage + npages1; vp < evp; vp++) {
1107 				VPP_SETPROT(vp, svd2->prot);
1108 				VPP_SETADVICE(vp, svd2->advice);
1109 			}
1110 		}
1111 
1112 		if (svd2->pageswap && (!svd1->pageswap && svd1->swresv)) {
1113 			ASSERT(svd1->swresv == seg1->s_size);
1114 			ASSERT(!(svd1->flags & MAP_NORESERVE));
1115 			ASSERT(!(svd2->flags & MAP_NORESERVE));
1116 			evp = nvpage + npages1;
1117 			for (vp = nvpage; vp < evp; vp++) {
1118 				VPP_SETSWAPRES(vp);
1119 			}
1120 		}
1121 
1122 		if (svd1->pageswap && (!svd2->pageswap && svd2->swresv)) {
1123 			ASSERT(svd2->swresv == seg2->s_size);
1124 			ASSERT(!(svd1->flags & MAP_NORESERVE));
1125 			ASSERT(!(svd2->flags & MAP_NORESERVE));
1126 			vp = nvpage + npages1;
1127 			evp = vp + npages2;
1128 			for (; vp < evp; vp++) {
1129 				VPP_SETSWAPRES(vp);
1130 			}
1131 		}
1132 	}
1133 	ASSERT((vpage1 != NULL || vpage2 != NULL) ||
1134 	    (svd1->pageswap == 0 && svd2->pageswap == 0));
1135 
1136 	/*
1137 	 * If either segment has private pages, create a new merged anon
1138 	 * array. If mergeing shared anon segments just decrement anon map's
1139 	 * refcnt.
1140 	 */
1141 	if (amp1 != NULL && svd1->type == MAP_SHARED) {
1142 		ASSERT(amp1 == amp2 && svd1->vp == NULL);
1143 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1144 		ASSERT(amp1->refcnt >= 2);
1145 		amp1->refcnt--;
1146 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1147 		svd2->amp = NULL;
1148 	} else if (amp1 != NULL || amp2 != NULL) {
1149 		struct anon_hdr *nahp;
1150 		struct anon_map *namp = NULL;
1151 		size_t asize;
1152 
1153 		ASSERT(svd1->type == MAP_PRIVATE);
1154 
1155 		asize = seg1->s_size + seg2->s_size;
1156 		if ((nahp = anon_create(btop(asize), ANON_NOSLEEP)) == NULL) {
1157 			if (nvpage != NULL) {
1158 				kmem_free(nvpage, nvpsize);
1159 			}
1160 			return (-2);
1161 		}
1162 		if (amp1 != NULL) {
1163 			/*
1164 			 * XXX anon rwlock is not really needed because
1165 			 * this is a private segment and we are writers.
1166 			 */
1167 			ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1168 			ASSERT(amp1->refcnt == 1);
1169 			if (anon_copy_ptr(amp1->ahp, svd1->anon_index,
1170 			    nahp, 0, btop(seg1->s_size), ANON_NOSLEEP)) {
1171 				anon_release(nahp, btop(asize));
1172 				ANON_LOCK_EXIT(&amp1->a_rwlock);
1173 				if (nvpage != NULL) {
1174 					kmem_free(nvpage, nvpsize);
1175 				}
1176 				return (-2);
1177 			}
1178 		}
1179 		if (amp2 != NULL) {
1180 			ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1181 			ASSERT(amp2->refcnt == 1);
1182 			if (anon_copy_ptr(amp2->ahp, svd2->anon_index,
1183 			    nahp, btop(seg1->s_size), btop(seg2->s_size),
1184 			    ANON_NOSLEEP)) {
1185 				anon_release(nahp, btop(asize));
1186 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1187 				if (amp1 != NULL) {
1188 					ANON_LOCK_EXIT(&amp1->a_rwlock);
1189 				}
1190 				if (nvpage != NULL) {
1191 					kmem_free(nvpage, nvpsize);
1192 				}
1193 				return (-2);
1194 			}
1195 		}
1196 		if (amp1 != NULL) {
1197 			namp = amp1;
1198 			anon_release(amp1->ahp, btop(amp1->size));
1199 		}
1200 		if (amp2 != NULL) {
1201 			if (namp == NULL) {
1202 				ASSERT(amp1 == NULL);
1203 				namp = amp2;
1204 				anon_release(amp2->ahp, btop(amp2->size));
1205 			} else {
1206 				amp2->refcnt--;
1207 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1208 				anonmap_free(amp2);
1209 			}
1210 			svd2->amp = NULL; /* needed for seg_free */
1211 		}
1212 		namp->ahp = nahp;
1213 		namp->size = asize;
1214 		svd1->amp = namp;
1215 		svd1->anon_index = 0;
1216 		ANON_LOCK_EXIT(&namp->a_rwlock);
1217 	}
1218 	/*
1219 	 * Now free the old vpage structures.
1220 	 */
1221 	if (nvpage != NULL) {
1222 		if (vpage1 != NULL) {
1223 			kmem_free(vpage1, vpgtob(npages1));
1224 		}
1225 		if (vpage2 != NULL) {
1226 			svd2->vpage = NULL;
1227 			kmem_free(vpage2, vpgtob(npages2));
1228 		}
1229 		if (svd2->pageprot) {
1230 			svd1->pageprot = 1;
1231 		}
1232 		if (svd2->pageadvice) {
1233 			svd1->pageadvice = 1;
1234 		}
1235 		if (svd2->pageswap) {
1236 			svd1->pageswap = 1;
1237 		}
1238 		svd1->vpage = nvpage;
1239 	}
1240 
1241 	/* all looks ok, merge segments */
1242 	svd1->swresv += svd2->swresv;
1243 	svd2->swresv = 0;  /* so seg_free doesn't release swap space */
1244 	size = seg2->s_size;
1245 	seg_free(seg2);
1246 	seg1->s_size += size;
1247 	return (0);
1248 }
1249 
1250 /*
1251  * Extend the previous segment (seg1) to include the
1252  * new segment (seg2 + a), if possible.
1253  * Return 0 on success.
1254  */
1255 static int
1256 segvn_extend_prev(seg1, seg2, a, swresv)
1257 	struct seg *seg1, *seg2;
1258 	struct segvn_crargs *a;
1259 	size_t swresv;
1260 {
1261 	struct segvn_data *svd1 = (struct segvn_data *)seg1->s_data;
1262 	size_t size;
1263 	struct anon_map *amp1;
1264 	struct vpage *new_vpage;
1265 
1266 	/*
1267 	 * We don't need any segment level locks for "segvn" data
1268 	 * since the address space is "write" locked.
1269 	 */
1270 	ASSERT(seg1->s_as && AS_WRITE_HELD(seg1->s_as));
1271 
1272 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie)) {
1273 		return (-1);
1274 	}
1275 
1276 	/* second segment is new, try to extend first */
1277 	/* XXX - should also check cred */
1278 	if (svd1->vp != a->vp || svd1->maxprot != a->maxprot ||
1279 	    (!svd1->pageprot && (svd1->prot != a->prot)) ||
1280 	    svd1->type != a->type || svd1->flags != a->flags ||
1281 	    seg1->s_szc != a->szc || svd1->softlockcnt_send > 0)
1282 		return (-1);
1283 
1284 	/* vp == NULL implies zfod, offset doesn't matter */
1285 	if (svd1->vp != NULL &&
1286 	    svd1->offset + seg1->s_size != (a->offset & PAGEMASK))
1287 		return (-1);
1288 
1289 	if (svd1->tr_state != SEGVN_TR_OFF) {
1290 		return (-1);
1291 	}
1292 
1293 	amp1 = svd1->amp;
1294 	if (amp1) {
1295 		pgcnt_t newpgs;
1296 
1297 		/*
1298 		 * Segment has private pages, can data structures
1299 		 * be expanded?
1300 		 *
1301 		 * Acquire the anon_map lock to prevent it from changing,
1302 		 * if it is shared.  This ensures that the anon_map
1303 		 * will not change while a thread which has a read/write
1304 		 * lock on an address space references it.
1305 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1306 		 * is 1.
1307 		 *
1308 		 * Can't grow a MAP_SHARED segment with an anonmap because
1309 		 * there may be existing anon slots where we want to extend
1310 		 * the segment and we wouldn't know what to do with them
1311 		 * (e.g., for tmpfs right thing is to just leave them there,
1312 		 * for /dev/zero they should be cleared out).
1313 		 */
1314 		if (svd1->type == MAP_SHARED)
1315 			return (-1);
1316 
1317 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1318 		if (amp1->refcnt > 1) {
1319 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1320 			return (-1);
1321 		}
1322 		newpgs = anon_grow(amp1->ahp, &svd1->anon_index,
1323 		    btop(seg1->s_size), btop(seg2->s_size), ANON_NOSLEEP);
1324 
1325 		if (newpgs == 0) {
1326 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1327 			return (-1);
1328 		}
1329 		amp1->size = ptob(newpgs);
1330 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1331 	}
1332 	if (svd1->vpage != NULL) {
1333 		struct vpage *vp, *evp;
1334 		new_vpage =
1335 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1336 			KM_NOSLEEP);
1337 		if (new_vpage == NULL)
1338 			return (-1);
1339 		bcopy(svd1->vpage, new_vpage, vpgtob(seg_pages(seg1)));
1340 		kmem_free(svd1->vpage, vpgtob(seg_pages(seg1)));
1341 		svd1->vpage = new_vpage;
1342 
1343 		vp = new_vpage + seg_pages(seg1);
1344 		evp = vp + seg_pages(seg2);
1345 		for (; vp < evp; vp++)
1346 			VPP_SETPROT(vp, a->prot);
1347 		if (svd1->pageswap && swresv) {
1348 			ASSERT(!(svd1->flags & MAP_NORESERVE));
1349 			ASSERT(swresv == seg2->s_size);
1350 			vp = new_vpage + seg_pages(seg1);
1351 			for (; vp < evp; vp++) {
1352 				VPP_SETSWAPRES(vp);
1353 			}
1354 		}
1355 	}
1356 	ASSERT(svd1->vpage != NULL || svd1->pageswap == 0);
1357 	size = seg2->s_size;
1358 	seg_free(seg2);
1359 	seg1->s_size += size;
1360 	svd1->swresv += swresv;
1361 	if (svd1->pageprot && (a->prot & PROT_WRITE) &&
1362 	    svd1->type == MAP_SHARED && svd1->vp != NULL &&
1363 	    (svd1->vp->v_flag & VVMEXEC)) {
1364 		ASSERT(vn_is_mapped(svd1->vp, V_WRITE));
1365 		segvn_inval_trcache(svd1->vp);
1366 	}
1367 	return (0);
1368 }
1369 
1370 /*
1371  * Extend the next segment (seg2) to include the
1372  * new segment (seg1 + a), if possible.
1373  * Return 0 on success.
1374  */
1375 static int
1376 segvn_extend_next(
1377 	struct seg *seg1,
1378 	struct seg *seg2,
1379 	struct segvn_crargs *a,
1380 	size_t swresv)
1381 {
1382 	struct segvn_data *svd2 = (struct segvn_data *)seg2->s_data;
1383 	size_t size;
1384 	struct anon_map *amp2;
1385 	struct vpage *new_vpage;
1386 
1387 	/*
1388 	 * We don't need any segment level locks for "segvn" data
1389 	 * since the address space is "write" locked.
1390 	 */
1391 	ASSERT(seg2->s_as && AS_WRITE_HELD(seg2->s_as));
1392 
1393 	if (HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1394 		return (-1);
1395 	}
1396 
1397 	/* first segment is new, try to extend second */
1398 	/* XXX - should also check cred */
1399 	if (svd2->vp != a->vp || svd2->maxprot != a->maxprot ||
1400 	    (!svd2->pageprot && (svd2->prot != a->prot)) ||
1401 	    svd2->type != a->type || svd2->flags != a->flags ||
1402 	    seg2->s_szc != a->szc || svd2->softlockcnt_sbase > 0)
1403 		return (-1);
1404 	/* vp == NULL implies zfod, offset doesn't matter */
1405 	if (svd2->vp != NULL &&
1406 	    (a->offset & PAGEMASK) + seg1->s_size != svd2->offset)
1407 		return (-1);
1408 
1409 	if (svd2->tr_state != SEGVN_TR_OFF) {
1410 		return (-1);
1411 	}
1412 
1413 	amp2 = svd2->amp;
1414 	if (amp2) {
1415 		pgcnt_t newpgs;
1416 
1417 		/*
1418 		 * Segment has private pages, can data structures
1419 		 * be expanded?
1420 		 *
1421 		 * Acquire the anon_map lock to prevent it from changing,
1422 		 * if it is shared.  This ensures that the anon_map
1423 		 * will not change while a thread which has a read/write
1424 		 * lock on an address space references it.
1425 		 *
1426 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1427 		 * is 1.
1428 		 */
1429 		if (svd2->type == MAP_SHARED)
1430 			return (-1);
1431 
1432 		ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1433 		if (amp2->refcnt > 1) {
1434 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1435 			return (-1);
1436 		}
1437 		newpgs = anon_grow(amp2->ahp, &svd2->anon_index,
1438 		    btop(seg2->s_size), btop(seg1->s_size),
1439 		    ANON_NOSLEEP | ANON_GROWDOWN);
1440 
1441 		if (newpgs == 0) {
1442 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1443 			return (-1);
1444 		}
1445 		amp2->size = ptob(newpgs);
1446 		ANON_LOCK_EXIT(&amp2->a_rwlock);
1447 	}
1448 	if (svd2->vpage != NULL) {
1449 		struct vpage *vp, *evp;
1450 		new_vpage =
1451 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1452 		    KM_NOSLEEP);
1453 		if (new_vpage == NULL) {
1454 			/* Not merging segments so adjust anon_index back */
1455 			if (amp2)
1456 				svd2->anon_index += seg_pages(seg1);
1457 			return (-1);
1458 		}
1459 		bcopy(svd2->vpage, new_vpage + seg_pages(seg1),
1460 		    vpgtob(seg_pages(seg2)));
1461 		kmem_free(svd2->vpage, vpgtob(seg_pages(seg2)));
1462 		svd2->vpage = new_vpage;
1463 
1464 		vp = new_vpage;
1465 		evp = vp + seg_pages(seg1);
1466 		for (; vp < evp; vp++)
1467 			VPP_SETPROT(vp, a->prot);
1468 		if (svd2->pageswap && swresv) {
1469 			ASSERT(!(svd2->flags & MAP_NORESERVE));
1470 			ASSERT(swresv == seg1->s_size);
1471 			vp = new_vpage;
1472 			for (; vp < evp; vp++) {
1473 				VPP_SETSWAPRES(vp);
1474 			}
1475 		}
1476 	}
1477 	ASSERT(svd2->vpage != NULL || svd2->pageswap == 0);
1478 	size = seg1->s_size;
1479 	seg_free(seg1);
1480 	seg2->s_size += size;
1481 	seg2->s_base -= size;
1482 	svd2->offset -= size;
1483 	svd2->swresv += swresv;
1484 	if (svd2->pageprot && (a->prot & PROT_WRITE) &&
1485 	    svd2->type == MAP_SHARED && svd2->vp != NULL &&
1486 	    (svd2->vp->v_flag & VVMEXEC)) {
1487 		ASSERT(vn_is_mapped(svd2->vp, V_WRITE));
1488 		segvn_inval_trcache(svd2->vp);
1489 	}
1490 	return (0);
1491 }
1492 
1493 /*
1494  * Duplicate all the pages in the segment. This may break COW sharing for a
1495  * given page. If the page is marked with inherit zero set, then instead of
1496  * duplicating the page, we zero the page.
1497  */
1498 static int
1499 segvn_dup_pages(struct seg *seg, struct seg *newseg)
1500 {
1501 	int error;
1502 	uint_t prot;
1503 	page_t *pp;
1504 	struct anon *ap, *newap;
1505 	size_t i;
1506 	caddr_t addr;
1507 
1508 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1509 	struct segvn_data *newsvd = (struct segvn_data *)newseg->s_data;
1510 	ulong_t old_idx = svd->anon_index;
1511 	ulong_t new_idx = 0;
1512 
1513 	i = btopr(seg->s_size);
1514 	addr = seg->s_base;
1515 
1516 	/*
1517 	 * XXX break cow sharing using PAGESIZE
1518 	 * pages. They will be relocated into larger
1519 	 * pages at fault time.
1520 	 */
1521 	while (i-- > 0) {
1522 		if ((ap = anon_get_ptr(svd->amp->ahp, old_idx)) != NULL) {
1523 			struct vpage *vpp;
1524 
1525 			vpp = &svd->vpage[seg_page(seg, addr)];
1526 
1527 			/*
1528 			 * prot need not be computed below 'cause anon_private
1529 			 * is going to ignore it anyway as child doesn't inherit
1530 			 * pagelock from parent.
1531 			 */
1532 			prot = svd->pageprot ? VPP_PROT(vpp) : svd->prot;
1533 
1534 			/*
1535 			 * Check whether we should zero this or dup it.
1536 			 */
1537 			if (svd->svn_inz == SEGVN_INZ_ALL ||
1538 			    (svd->svn_inz == SEGVN_INZ_VPP &&
1539 			    VPP_ISINHZERO(vpp))) {
1540 				pp = anon_zero(newseg, addr, &newap,
1541 				    newsvd->cred);
1542 			} else {
1543 				page_t *anon_pl[1+1];
1544 				uint_t vpprot;
1545 				error = anon_getpage(&ap, &vpprot, anon_pl,
1546 				    PAGESIZE, seg, addr, S_READ, svd->cred);
1547 				if (error != 0)
1548 					return (error);
1549 
1550 				pp = anon_private(&newap, newseg, addr, prot,
1551 				    anon_pl[0], 0, newsvd->cred);
1552 			}
1553 			if (pp == NULL) {
1554 				return (ENOMEM);
1555 			}
1556 			(void) anon_set_ptr(newsvd->amp->ahp, new_idx, newap,
1557 			    ANON_SLEEP);
1558 			page_unlock(pp);
1559 		}
1560 		addr += PAGESIZE;
1561 		old_idx++;
1562 		new_idx++;
1563 	}
1564 
1565 	return (0);
1566 }
1567 
1568 static int
1569 segvn_dup(struct seg *seg, struct seg *newseg)
1570 {
1571 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1572 	struct segvn_data *newsvd;
1573 	pgcnt_t npages = seg_pages(seg);
1574 	int error = 0;
1575 	size_t len;
1576 	struct anon_map *amp;
1577 
1578 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1579 	ASSERT(newseg->s_as->a_proc->p_parent == curproc);
1580 
1581 	/*
1582 	 * If segment has anon reserved, reserve more for the new seg.
1583 	 * For a MAP_NORESERVE segment swresv will be a count of all the
1584 	 * allocated anon slots; thus we reserve for the child as many slots
1585 	 * as the parent has allocated. This semantic prevents the child or
1586 	 * parent from dieing during a copy-on-write fault caused by trying
1587 	 * to write a shared pre-existing anon page.
1588 	 */
1589 	if ((len = svd->swresv) != 0) {
1590 		if (anon_resv(svd->swresv) == 0)
1591 			return (ENOMEM);
1592 
1593 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
1594 		    seg, len, 0);
1595 	}
1596 
1597 	newsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
1598 
1599 	newseg->s_ops = &segvn_ops;
1600 	newseg->s_data = (void *)newsvd;
1601 	newseg->s_szc = seg->s_szc;
1602 
1603 	newsvd->seg = newseg;
1604 	if ((newsvd->vp = svd->vp) != NULL) {
1605 		VN_HOLD(svd->vp);
1606 		if (svd->type == MAP_SHARED)
1607 			lgrp_shm_policy_init(NULL, svd->vp);
1608 	}
1609 	newsvd->offset = svd->offset;
1610 	newsvd->prot = svd->prot;
1611 	newsvd->maxprot = svd->maxprot;
1612 	newsvd->pageprot = svd->pageprot;
1613 	newsvd->type = svd->type;
1614 	newsvd->cred = svd->cred;
1615 	crhold(newsvd->cred);
1616 	newsvd->advice = svd->advice;
1617 	newsvd->pageadvice = svd->pageadvice;
1618 	newsvd->svn_inz = svd->svn_inz;
1619 	newsvd->swresv = svd->swresv;
1620 	newsvd->pageswap = svd->pageswap;
1621 	newsvd->flags = svd->flags;
1622 	newsvd->softlockcnt = 0;
1623 	newsvd->softlockcnt_sbase = 0;
1624 	newsvd->softlockcnt_send = 0;
1625 	newsvd->policy_info = svd->policy_info;
1626 	newsvd->rcookie = HAT_INVALID_REGION_COOKIE;
1627 
1628 	if ((amp = svd->amp) == NULL || svd->tr_state == SEGVN_TR_ON) {
1629 		/*
1630 		 * Not attaching to a shared anon object.
1631 		 */
1632 		ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie) ||
1633 		    svd->tr_state == SEGVN_TR_OFF);
1634 		if (svd->tr_state == SEGVN_TR_ON) {
1635 			ASSERT(newsvd->vp != NULL && amp != NULL);
1636 			newsvd->tr_state = SEGVN_TR_INIT;
1637 		} else {
1638 			newsvd->tr_state = svd->tr_state;
1639 		}
1640 		newsvd->amp = NULL;
1641 		newsvd->anon_index = 0;
1642 	} else {
1643 		/* regions for now are only used on pure vnode segments */
1644 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
1645 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1646 		newsvd->tr_state = SEGVN_TR_OFF;
1647 		if (svd->type == MAP_SHARED) {
1648 			ASSERT(svd->svn_inz == SEGVN_INZ_NONE);
1649 			newsvd->amp = amp;
1650 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1651 			amp->refcnt++;
1652 			ANON_LOCK_EXIT(&amp->a_rwlock);
1653 			newsvd->anon_index = svd->anon_index;
1654 		} else {
1655 			int reclaim = 1;
1656 
1657 			/*
1658 			 * Allocate and initialize new anon_map structure.
1659 			 */
1660 			newsvd->amp = anonmap_alloc(newseg->s_size, 0,
1661 			    ANON_SLEEP);
1662 			newsvd->amp->a_szc = newseg->s_szc;
1663 			newsvd->anon_index = 0;
1664 			ASSERT(svd->svn_inz == SEGVN_INZ_NONE ||
1665 			    svd->svn_inz == SEGVN_INZ_ALL ||
1666 			    svd->svn_inz == SEGVN_INZ_VPP);
1667 
1668 			/*
1669 			 * We don't have to acquire the anon_map lock
1670 			 * for the new segment (since it belongs to an
1671 			 * address space that is still not associated
1672 			 * with any process), or the segment in the old
1673 			 * address space (since all threads in it
1674 			 * are stopped while duplicating the address space).
1675 			 */
1676 
1677 			/*
1678 			 * The goal of the following code is to make sure that
1679 			 * softlocked pages do not end up as copy on write
1680 			 * pages.  This would cause problems where one
1681 			 * thread writes to a page that is COW and a different
1682 			 * thread in the same process has softlocked it.  The
1683 			 * softlock lock would move away from this process
1684 			 * because the write would cause this process to get
1685 			 * a copy (without the softlock).
1686 			 *
1687 			 * The strategy here is to just break the
1688 			 * sharing on pages that could possibly be
1689 			 * softlocked.
1690 			 *
1691 			 * In addition, if any pages have been marked that they
1692 			 * should be inherited as zero, then we immediately go
1693 			 * ahead and break COW and zero them. In the case of a
1694 			 * softlocked page that should be inherited zero, we
1695 			 * break COW and just get a zero page.
1696 			 */
1697 retry:
1698 			if (svd->softlockcnt ||
1699 			    svd->svn_inz != SEGVN_INZ_NONE) {
1700 				/*
1701 				 * The softlock count might be non zero
1702 				 * because some pages are still stuck in the
1703 				 * cache for lazy reclaim. Flush the cache
1704 				 * now. This should drop the count to zero.
1705 				 * [or there is really I/O going on to these
1706 				 * pages]. Note, we have the writers lock so
1707 				 * nothing gets inserted during the flush.
1708 				 */
1709 				if (svd->softlockcnt && reclaim == 1) {
1710 					segvn_purge(seg);
1711 					reclaim = 0;
1712 					goto retry;
1713 				}
1714 
1715 				error = segvn_dup_pages(seg, newseg);
1716 				if (error != 0) {
1717 					newsvd->vpage = NULL;
1718 					goto out;
1719 				}
1720 			} else {	/* common case */
1721 				if (seg->s_szc != 0) {
1722 					/*
1723 					 * If at least one of anon slots of a
1724 					 * large page exists then make sure
1725 					 * all anon slots of a large page
1726 					 * exist to avoid partial cow sharing
1727 					 * of a large page in the future.
1728 					 */
1729 					anon_dup_fill_holes(amp->ahp,
1730 					    svd->anon_index, newsvd->amp->ahp,
1731 					    0, seg->s_size, seg->s_szc,
1732 					    svd->vp != NULL);
1733 				} else {
1734 					anon_dup(amp->ahp, svd->anon_index,
1735 					    newsvd->amp->ahp, 0, seg->s_size);
1736 				}
1737 
1738 				hat_clrattr(seg->s_as->a_hat, seg->s_base,
1739 				    seg->s_size, PROT_WRITE);
1740 			}
1741 		}
1742 	}
1743 	/*
1744 	 * If necessary, create a vpage structure for the new segment.
1745 	 * Do not copy any page lock indications.
1746 	 */
1747 	if (svd->vpage != NULL) {
1748 		uint_t i;
1749 		struct vpage *ovp = svd->vpage;
1750 		struct vpage *nvp;
1751 
1752 		nvp = newsvd->vpage =
1753 		    kmem_alloc(vpgtob(npages), KM_SLEEP);
1754 		for (i = 0; i < npages; i++) {
1755 			*nvp = *ovp++;
1756 			VPP_CLRPPLOCK(nvp++);
1757 		}
1758 	} else
1759 		newsvd->vpage = NULL;
1760 
1761 	/* Inform the vnode of the new mapping */
1762 	if (newsvd->vp != NULL) {
1763 		error = VOP_ADDMAP(newsvd->vp, (offset_t)newsvd->offset,
1764 		    newseg->s_as, newseg->s_base, newseg->s_size, newsvd->prot,
1765 		    newsvd->maxprot, newsvd->type, newsvd->cred, NULL);
1766 	}
1767 out:
1768 	if (error == 0 && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1769 		ASSERT(newsvd->amp == NULL);
1770 		ASSERT(newsvd->tr_state == SEGVN_TR_OFF);
1771 		newsvd->rcookie = svd->rcookie;
1772 		hat_dup_region(newseg->s_as->a_hat, newsvd->rcookie);
1773 	}
1774 	return (error);
1775 }
1776 
1777 
1778 /*
1779  * callback function to invoke free_vp_pages() for only those pages actually
1780  * processed by the HAT when a shared region is destroyed.
1781  */
1782 extern int free_pages;
1783 
1784 static void
1785 segvn_hat_rgn_unload_callback(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
1786     size_t r_size, void *r_obj, u_offset_t r_objoff)
1787 {
1788 	u_offset_t off;
1789 	size_t len;
1790 	vnode_t *vp = (vnode_t *)r_obj;
1791 
1792 	ASSERT(eaddr > saddr);
1793 	ASSERT(saddr >= r_saddr);
1794 	ASSERT(saddr < r_saddr + r_size);
1795 	ASSERT(eaddr > r_saddr);
1796 	ASSERT(eaddr <= r_saddr + r_size);
1797 	ASSERT(vp != NULL);
1798 
1799 	if (!free_pages) {
1800 		return;
1801 	}
1802 
1803 	len = eaddr - saddr;
1804 	off = (saddr - r_saddr) + r_objoff;
1805 	free_vp_pages(vp, off, len);
1806 }
1807 
1808 /*
1809  * callback function used by segvn_unmap to invoke free_vp_pages() for only
1810  * those pages actually processed by the HAT
1811  */
1812 static void
1813 segvn_hat_unload_callback(hat_callback_t *cb)
1814 {
1815 	struct seg		*seg = cb->hcb_data;
1816 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
1817 	size_t			len;
1818 	u_offset_t		off;
1819 
1820 	ASSERT(svd->vp != NULL);
1821 	ASSERT(cb->hcb_end_addr > cb->hcb_start_addr);
1822 	ASSERT(cb->hcb_start_addr >= seg->s_base);
1823 
1824 	len = cb->hcb_end_addr - cb->hcb_start_addr;
1825 	off = cb->hcb_start_addr - seg->s_base;
1826 	free_vp_pages(svd->vp, svd->offset + off, len);
1827 }
1828 
1829 /*
1830  * This function determines the number of bytes of swap reserved by
1831  * a segment for which per-page accounting is present. It is used to
1832  * calculate the correct value of a segvn_data's swresv.
1833  */
1834 static size_t
1835 segvn_count_swap_by_vpages(struct seg *seg)
1836 {
1837 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1838 	struct vpage *vp, *evp;
1839 	size_t nswappages = 0;
1840 
1841 	ASSERT(svd->pageswap);
1842 	ASSERT(svd->vpage != NULL);
1843 
1844 	evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
1845 
1846 	for (vp = svd->vpage; vp < evp; vp++) {
1847 		if (VPP_ISSWAPRES(vp))
1848 			nswappages++;
1849 	}
1850 
1851 	return (nswappages << PAGESHIFT);
1852 }
1853 
1854 static int
1855 segvn_unmap(struct seg *seg, caddr_t addr, size_t len)
1856 {
1857 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1858 	struct segvn_data *nsvd;
1859 	struct seg *nseg;
1860 	struct anon_map *amp;
1861 	pgcnt_t	opages;		/* old segment size in pages */
1862 	pgcnt_t	npages;		/* new segment size in pages */
1863 	pgcnt_t	dpages;		/* pages being deleted (unmapped) */
1864 	hat_callback_t callback;	/* used for free_vp_pages() */
1865 	hat_callback_t *cbp = NULL;
1866 	caddr_t nbase;
1867 	size_t nsize;
1868 	size_t oswresv;
1869 	int reclaim = 1;
1870 
1871 	/*
1872 	 * We don't need any segment level locks for "segvn" data
1873 	 * since the address space is "write" locked.
1874 	 */
1875 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1876 
1877 	/*
1878 	 * Fail the unmap if pages are SOFTLOCKed through this mapping.
1879 	 * softlockcnt is protected from change by the as write lock.
1880 	 */
1881 retry:
1882 	if (svd->softlockcnt > 0) {
1883 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1884 
1885 		/*
1886 		 * If this is shared segment non 0 softlockcnt
1887 		 * means locked pages are still in use.
1888 		 */
1889 		if (svd->type == MAP_SHARED) {
1890 			return (EAGAIN);
1891 		}
1892 
1893 		/*
1894 		 * since we do have the writers lock nobody can fill
1895 		 * the cache during the purge. The flush either succeeds
1896 		 * or we still have pending I/Os.
1897 		 */
1898 		if (reclaim == 1) {
1899 			segvn_purge(seg);
1900 			reclaim = 0;
1901 			goto retry;
1902 		}
1903 		return (EAGAIN);
1904 	}
1905 
1906 	/*
1907 	 * Check for bad sizes
1908 	 */
1909 	if (addr < seg->s_base || addr + len > seg->s_base + seg->s_size ||
1910 	    (len & PAGEOFFSET) || ((uintptr_t)addr & PAGEOFFSET)) {
1911 		panic("segvn_unmap");
1912 		/*NOTREACHED*/
1913 	}
1914 
1915 	if (seg->s_szc != 0) {
1916 		size_t pgsz = page_get_pagesize(seg->s_szc);
1917 		int err;
1918 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
1919 			ASSERT(seg->s_base != addr || seg->s_size != len);
1920 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1921 				ASSERT(svd->amp == NULL);
1922 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1923 				hat_leave_region(seg->s_as->a_hat,
1924 				    svd->rcookie, HAT_REGION_TEXT);
1925 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
1926 				/*
1927 				 * could pass a flag to segvn_demote_range()
1928 				 * below to tell it not to do any unloads but
1929 				 * this case is rare enough to not bother for
1930 				 * now.
1931 				 */
1932 			} else if (svd->tr_state == SEGVN_TR_INIT) {
1933 				svd->tr_state = SEGVN_TR_OFF;
1934 			} else if (svd->tr_state == SEGVN_TR_ON) {
1935 				ASSERT(svd->amp != NULL);
1936 				segvn_textunrepl(seg, 1);
1937 				ASSERT(svd->amp == NULL);
1938 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1939 			}
1940 			VM_STAT_ADD(segvnvmstats.demoterange[0]);
1941 			err = segvn_demote_range(seg, addr, len, SDR_END, 0);
1942 			if (err == 0) {
1943 				return (IE_RETRY);
1944 			}
1945 			return (err);
1946 		}
1947 	}
1948 
1949 	/* Inform the vnode of the unmapping. */
1950 	if (svd->vp) {
1951 		int error;
1952 
1953 		error = VOP_DELMAP(svd->vp,
1954 		    (offset_t)svd->offset + (uintptr_t)(addr - seg->s_base),
1955 		    seg->s_as, addr, len, svd->prot, svd->maxprot,
1956 		    svd->type, svd->cred, NULL);
1957 
1958 		if (error == EAGAIN)
1959 			return (error);
1960 	}
1961 
1962 	/*
1963 	 * Remove any page locks set through this mapping.
1964 	 * If text replication is not off no page locks could have been
1965 	 * established via this mapping.
1966 	 */
1967 	if (svd->tr_state == SEGVN_TR_OFF) {
1968 		(void) segvn_lockop(seg, addr, len, 0, MC_UNLOCK, NULL, 0);
1969 	}
1970 
1971 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1972 		ASSERT(svd->amp == NULL);
1973 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1974 		ASSERT(svd->type == MAP_PRIVATE);
1975 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
1976 		    HAT_REGION_TEXT);
1977 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
1978 	} else if (svd->tr_state == SEGVN_TR_ON) {
1979 		ASSERT(svd->amp != NULL);
1980 		ASSERT(svd->pageprot == 0 && !(svd->prot & PROT_WRITE));
1981 		segvn_textunrepl(seg, 1);
1982 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
1983 	} else {
1984 		if (svd->tr_state != SEGVN_TR_OFF) {
1985 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
1986 			svd->tr_state = SEGVN_TR_OFF;
1987 		}
1988 		/*
1989 		 * Unload any hardware translations in the range to be taken
1990 		 * out. Use a callback to invoke free_vp_pages() effectively.
1991 		 */
1992 		if (svd->vp != NULL && free_pages != 0) {
1993 			callback.hcb_data = seg;
1994 			callback.hcb_function = segvn_hat_unload_callback;
1995 			cbp = &callback;
1996 		}
1997 		hat_unload_callback(seg->s_as->a_hat, addr, len,
1998 		    HAT_UNLOAD_UNMAP, cbp);
1999 
2000 		if (svd->type == MAP_SHARED && svd->vp != NULL &&
2001 		    (svd->vp->v_flag & VVMEXEC) &&
2002 		    ((svd->prot & PROT_WRITE) || svd->pageprot)) {
2003 			segvn_inval_trcache(svd->vp);
2004 		}
2005 	}
2006 
2007 	/*
2008 	 * Check for entire segment
2009 	 */
2010 	if (addr == seg->s_base && len == seg->s_size) {
2011 		seg_free(seg);
2012 		return (0);
2013 	}
2014 
2015 	opages = seg_pages(seg);
2016 	dpages = btop(len);
2017 	npages = opages - dpages;
2018 	amp = svd->amp;
2019 	ASSERT(amp == NULL || amp->a_szc >= seg->s_szc);
2020 
2021 	/*
2022 	 * Check for beginning of segment
2023 	 */
2024 	if (addr == seg->s_base) {
2025 		if (svd->vpage != NULL) {
2026 			size_t nbytes;
2027 			struct vpage *ovpage;
2028 
2029 			ovpage = svd->vpage;	/* keep pointer to vpage */
2030 
2031 			nbytes = vpgtob(npages);
2032 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2033 			bcopy(&ovpage[dpages], svd->vpage, nbytes);
2034 
2035 			/* free up old vpage */
2036 			kmem_free(ovpage, vpgtob(opages));
2037 		}
2038 		if (amp != NULL) {
2039 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2040 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2041 				/*
2042 				 * Shared anon map is no longer in use. Before
2043 				 * freeing its pages purge all entries from
2044 				 * pcache that belong to this amp.
2045 				 */
2046 				if (svd->type == MAP_SHARED) {
2047 					ASSERT(amp->refcnt == 1);
2048 					ASSERT(svd->softlockcnt == 0);
2049 					anonmap_purge(amp);
2050 				}
2051 				/*
2052 				 * Free up now unused parts of anon_map array.
2053 				 */
2054 				if (amp->a_szc == seg->s_szc) {
2055 					if (seg->s_szc != 0) {
2056 						anon_free_pages(amp->ahp,
2057 						    svd->anon_index, len,
2058 						    seg->s_szc);
2059 					} else {
2060 						anon_free(amp->ahp,
2061 						    svd->anon_index,
2062 						    len);
2063 					}
2064 				} else {
2065 					ASSERT(svd->type == MAP_SHARED);
2066 					ASSERT(amp->a_szc > seg->s_szc);
2067 					anon_shmap_free_pages(amp,
2068 					    svd->anon_index, len);
2069 				}
2070 
2071 				/*
2072 				 * Unreserve swap space for the
2073 				 * unmapped chunk of this segment in
2074 				 * case it's MAP_SHARED
2075 				 */
2076 				if (svd->type == MAP_SHARED) {
2077 					anon_unresv_zone(len,
2078 					    seg->s_as->a_proc->p_zone);
2079 					amp->swresv -= len;
2080 				}
2081 			}
2082 			ANON_LOCK_EXIT(&amp->a_rwlock);
2083 			svd->anon_index += dpages;
2084 		}
2085 		if (svd->vp != NULL)
2086 			svd->offset += len;
2087 
2088 		seg->s_base += len;
2089 		seg->s_size -= len;
2090 
2091 		if (svd->swresv) {
2092 			if (svd->flags & MAP_NORESERVE) {
2093 				ASSERT(amp);
2094 				oswresv = svd->swresv;
2095 
2096 				svd->swresv = ptob(anon_pages(amp->ahp,
2097 				    svd->anon_index, npages));
2098 				anon_unresv_zone(oswresv - svd->swresv,
2099 				    seg->s_as->a_proc->p_zone);
2100 				if (SEG_IS_PARTIAL_RESV(seg))
2101 					seg->s_as->a_resvsize -= oswresv -
2102 					    svd->swresv;
2103 			} else {
2104 				size_t unlen;
2105 
2106 				if (svd->pageswap) {
2107 					oswresv = svd->swresv;
2108 					svd->swresv =
2109 					    segvn_count_swap_by_vpages(seg);
2110 					ASSERT(oswresv >= svd->swresv);
2111 					unlen = oswresv - svd->swresv;
2112 				} else {
2113 					svd->swresv -= len;
2114 					ASSERT(svd->swresv == seg->s_size);
2115 					unlen = len;
2116 				}
2117 				anon_unresv_zone(unlen,
2118 				    seg->s_as->a_proc->p_zone);
2119 			}
2120 			TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2121 			    seg, len, 0);
2122 		}
2123 
2124 		return (0);
2125 	}
2126 
2127 	/*
2128 	 * Check for end of segment
2129 	 */
2130 	if (addr + len == seg->s_base + seg->s_size) {
2131 		if (svd->vpage != NULL) {
2132 			size_t nbytes;
2133 			struct vpage *ovpage;
2134 
2135 			ovpage = svd->vpage;	/* keep pointer to vpage */
2136 
2137 			nbytes = vpgtob(npages);
2138 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2139 			bcopy(ovpage, svd->vpage, nbytes);
2140 
2141 			/* free up old vpage */
2142 			kmem_free(ovpage, vpgtob(opages));
2143 
2144 		}
2145 		if (amp != NULL) {
2146 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2147 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2148 				/*
2149 				 * Free up now unused parts of anon_map array.
2150 				 */
2151 				ulong_t an_idx = svd->anon_index + npages;
2152 
2153 				/*
2154 				 * Shared anon map is no longer in use. Before
2155 				 * freeing its pages purge all entries from
2156 				 * pcache that belong to this amp.
2157 				 */
2158 				if (svd->type == MAP_SHARED) {
2159 					ASSERT(amp->refcnt == 1);
2160 					ASSERT(svd->softlockcnt == 0);
2161 					anonmap_purge(amp);
2162 				}
2163 
2164 				if (amp->a_szc == seg->s_szc) {
2165 					if (seg->s_szc != 0) {
2166 						anon_free_pages(amp->ahp,
2167 						    an_idx, len,
2168 						    seg->s_szc);
2169 					} else {
2170 						anon_free(amp->ahp, an_idx,
2171 						    len);
2172 					}
2173 				} else {
2174 					ASSERT(svd->type == MAP_SHARED);
2175 					ASSERT(amp->a_szc > seg->s_szc);
2176 					anon_shmap_free_pages(amp,
2177 					    an_idx, len);
2178 				}
2179 
2180 				/*
2181 				 * Unreserve swap space for the
2182 				 * unmapped chunk of this segment in
2183 				 * case it's MAP_SHARED
2184 				 */
2185 				if (svd->type == MAP_SHARED) {
2186 					anon_unresv_zone(len,
2187 					    seg->s_as->a_proc->p_zone);
2188 					amp->swresv -= len;
2189 				}
2190 			}
2191 			ANON_LOCK_EXIT(&amp->a_rwlock);
2192 		}
2193 
2194 		seg->s_size -= len;
2195 
2196 		if (svd->swresv) {
2197 			if (svd->flags & MAP_NORESERVE) {
2198 				ASSERT(amp);
2199 				oswresv = svd->swresv;
2200 				svd->swresv = ptob(anon_pages(amp->ahp,
2201 				    svd->anon_index, npages));
2202 				anon_unresv_zone(oswresv - svd->swresv,
2203 				    seg->s_as->a_proc->p_zone);
2204 				if (SEG_IS_PARTIAL_RESV(seg))
2205 					seg->s_as->a_resvsize -= oswresv -
2206 					    svd->swresv;
2207 			} else {
2208 				size_t unlen;
2209 
2210 				if (svd->pageswap) {
2211 					oswresv = svd->swresv;
2212 					svd->swresv =
2213 					    segvn_count_swap_by_vpages(seg);
2214 					ASSERT(oswresv >= svd->swresv);
2215 					unlen = oswresv - svd->swresv;
2216 				} else {
2217 					svd->swresv -= len;
2218 					ASSERT(svd->swresv == seg->s_size);
2219 					unlen = len;
2220 				}
2221 				anon_unresv_zone(unlen,
2222 				    seg->s_as->a_proc->p_zone);
2223 			}
2224 			TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2225 			    "anon proc:%p %lu %u", seg, len, 0);
2226 		}
2227 
2228 		return (0);
2229 	}
2230 
2231 	/*
2232 	 * The section to go is in the middle of the segment,
2233 	 * have to make it into two segments.  nseg is made for
2234 	 * the high end while seg is cut down at the low end.
2235 	 */
2236 	nbase = addr + len;				/* new seg base */
2237 	nsize = (seg->s_base + seg->s_size) - nbase;	/* new seg size */
2238 	seg->s_size = addr - seg->s_base;		/* shrink old seg */
2239 	nseg = seg_alloc(seg->s_as, nbase, nsize);
2240 	if (nseg == NULL) {
2241 		panic("segvn_unmap seg_alloc");
2242 		/*NOTREACHED*/
2243 	}
2244 	nseg->s_ops = seg->s_ops;
2245 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
2246 	nseg->s_data = (void *)nsvd;
2247 	nseg->s_szc = seg->s_szc;
2248 	*nsvd = *svd;
2249 	nsvd->seg = nseg;
2250 	nsvd->offset = svd->offset + (uintptr_t)(nseg->s_base - seg->s_base);
2251 	nsvd->swresv = 0;
2252 	nsvd->softlockcnt = 0;
2253 	nsvd->softlockcnt_sbase = 0;
2254 	nsvd->softlockcnt_send = 0;
2255 	nsvd->svn_inz = svd->svn_inz;
2256 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
2257 
2258 	if (svd->vp != NULL) {
2259 		VN_HOLD(nsvd->vp);
2260 		if (nsvd->type == MAP_SHARED)
2261 			lgrp_shm_policy_init(NULL, nsvd->vp);
2262 	}
2263 	crhold(svd->cred);
2264 
2265 	if (svd->vpage == NULL) {
2266 		nsvd->vpage = NULL;
2267 	} else {
2268 		/* need to split vpage into two arrays */
2269 		size_t nbytes;
2270 		struct vpage *ovpage;
2271 
2272 		ovpage = svd->vpage;		/* keep pointer to vpage */
2273 
2274 		npages = seg_pages(seg);	/* seg has shrunk */
2275 		nbytes = vpgtob(npages);
2276 		svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2277 
2278 		bcopy(ovpage, svd->vpage, nbytes);
2279 
2280 		npages = seg_pages(nseg);
2281 		nbytes = vpgtob(npages);
2282 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2283 
2284 		bcopy(&ovpage[opages - npages], nsvd->vpage, nbytes);
2285 
2286 		/* free up old vpage */
2287 		kmem_free(ovpage, vpgtob(opages));
2288 	}
2289 
2290 	if (amp == NULL) {
2291 		nsvd->amp = NULL;
2292 		nsvd->anon_index = 0;
2293 	} else {
2294 		/*
2295 		 * Need to create a new anon map for the new segment.
2296 		 * We'll also allocate a new smaller array for the old
2297 		 * smaller segment to save space.
2298 		 */
2299 		opages = btop((uintptr_t)(addr - seg->s_base));
2300 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2301 		if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2302 			/*
2303 			 * Free up now unused parts of anon_map array.
2304 			 */
2305 			ulong_t an_idx = svd->anon_index + opages;
2306 
2307 			/*
2308 			 * Shared anon map is no longer in use. Before
2309 			 * freeing its pages purge all entries from
2310 			 * pcache that belong to this amp.
2311 			 */
2312 			if (svd->type == MAP_SHARED) {
2313 				ASSERT(amp->refcnt == 1);
2314 				ASSERT(svd->softlockcnt == 0);
2315 				anonmap_purge(amp);
2316 			}
2317 
2318 			if (amp->a_szc == seg->s_szc) {
2319 				if (seg->s_szc != 0) {
2320 					anon_free_pages(amp->ahp, an_idx, len,
2321 					    seg->s_szc);
2322 				} else {
2323 					anon_free(amp->ahp, an_idx,
2324 					    len);
2325 				}
2326 			} else {
2327 				ASSERT(svd->type == MAP_SHARED);
2328 				ASSERT(amp->a_szc > seg->s_szc);
2329 				anon_shmap_free_pages(amp, an_idx, len);
2330 			}
2331 
2332 			/*
2333 			 * Unreserve swap space for the
2334 			 * unmapped chunk of this segment in
2335 			 * case it's MAP_SHARED
2336 			 */
2337 			if (svd->type == MAP_SHARED) {
2338 				anon_unresv_zone(len,
2339 				    seg->s_as->a_proc->p_zone);
2340 				amp->swresv -= len;
2341 			}
2342 		}
2343 		nsvd->anon_index = svd->anon_index +
2344 		    btop((uintptr_t)(nseg->s_base - seg->s_base));
2345 		if (svd->type == MAP_SHARED) {
2346 			amp->refcnt++;
2347 			nsvd->amp = amp;
2348 		} else {
2349 			struct anon_map *namp;
2350 			struct anon_hdr *nahp;
2351 
2352 			ASSERT(svd->type == MAP_PRIVATE);
2353 			nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
2354 			namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
2355 			namp->a_szc = seg->s_szc;
2356 			(void) anon_copy_ptr(amp->ahp, svd->anon_index, nahp,
2357 			    0, btop(seg->s_size), ANON_SLEEP);
2358 			(void) anon_copy_ptr(amp->ahp, nsvd->anon_index,
2359 			    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
2360 			anon_release(amp->ahp, btop(amp->size));
2361 			svd->anon_index = 0;
2362 			nsvd->anon_index = 0;
2363 			amp->ahp = nahp;
2364 			amp->size = seg->s_size;
2365 			nsvd->amp = namp;
2366 		}
2367 		ANON_LOCK_EXIT(&amp->a_rwlock);
2368 	}
2369 	if (svd->swresv) {
2370 		if (svd->flags & MAP_NORESERVE) {
2371 			ASSERT(amp);
2372 			oswresv = svd->swresv;
2373 			svd->swresv = ptob(anon_pages(amp->ahp,
2374 			    svd->anon_index, btop(seg->s_size)));
2375 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
2376 			    nsvd->anon_index, btop(nseg->s_size)));
2377 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2378 			anon_unresv_zone(oswresv - (svd->swresv + nsvd->swresv),
2379 			    seg->s_as->a_proc->p_zone);
2380 			if (SEG_IS_PARTIAL_RESV(seg))
2381 				seg->s_as->a_resvsize -= oswresv -
2382 				    (svd->swresv + nsvd->swresv);
2383 		} else {
2384 			size_t unlen;
2385 
2386 			if (svd->pageswap) {
2387 				oswresv = svd->swresv;
2388 				svd->swresv = segvn_count_swap_by_vpages(seg);
2389 				nsvd->swresv = segvn_count_swap_by_vpages(nseg);
2390 				ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2391 				unlen = oswresv - (svd->swresv + nsvd->swresv);
2392 			} else {
2393 				if (seg->s_size + nseg->s_size + len !=
2394 				    svd->swresv) {
2395 					panic("segvn_unmap: cannot split "
2396 					    "swap reservation");
2397 					/*NOTREACHED*/
2398 				}
2399 				svd->swresv = seg->s_size;
2400 				nsvd->swresv = nseg->s_size;
2401 				unlen = len;
2402 			}
2403 			anon_unresv_zone(unlen,
2404 			    seg->s_as->a_proc->p_zone);
2405 		}
2406 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2407 		    seg, len, 0);
2408 	}
2409 
2410 	return (0);			/* I'm glad that's all over with! */
2411 }
2412 
2413 static void
2414 segvn_free(struct seg *seg)
2415 {
2416 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2417 	pgcnt_t npages = seg_pages(seg);
2418 	struct anon_map *amp;
2419 	size_t len;
2420 
2421 	/*
2422 	 * We don't need any segment level locks for "segvn" data
2423 	 * since the address space is "write" locked.
2424 	 */
2425 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
2426 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
2427 
2428 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2429 
2430 	/*
2431 	 * Be sure to unlock pages. XXX Why do things get free'ed instead
2432 	 * of unmapped? XXX
2433 	 */
2434 	(void) segvn_lockop(seg, seg->s_base, seg->s_size,
2435 	    0, MC_UNLOCK, NULL, 0);
2436 
2437 	/*
2438 	 * Deallocate the vpage and anon pointers if necessary and possible.
2439 	 */
2440 	if (svd->vpage != NULL) {
2441 		kmem_free(svd->vpage, vpgtob(npages));
2442 		svd->vpage = NULL;
2443 	}
2444 	if ((amp = svd->amp) != NULL) {
2445 		/*
2446 		 * If there are no more references to this anon_map
2447 		 * structure, then deallocate the structure after freeing
2448 		 * up all the anon slot pointers that we can.
2449 		 */
2450 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2451 		ASSERT(amp->a_szc >= seg->s_szc);
2452 		if (--amp->refcnt == 0) {
2453 			if (svd->type == MAP_PRIVATE) {
2454 				/*
2455 				 * Private - we only need to anon_free
2456 				 * the part that this segment refers to.
2457 				 */
2458 				if (seg->s_szc != 0) {
2459 					anon_free_pages(amp->ahp,
2460 					    svd->anon_index, seg->s_size,
2461 					    seg->s_szc);
2462 				} else {
2463 					anon_free(amp->ahp, svd->anon_index,
2464 					    seg->s_size);
2465 				}
2466 			} else {
2467 
2468 				/*
2469 				 * Shared anon map is no longer in use. Before
2470 				 * freeing its pages purge all entries from
2471 				 * pcache that belong to this amp.
2472 				 */
2473 				ASSERT(svd->softlockcnt == 0);
2474 				anonmap_purge(amp);
2475 
2476 				/*
2477 				 * Shared - anon_free the entire
2478 				 * anon_map's worth of stuff and
2479 				 * release any swap reservation.
2480 				 */
2481 				if (amp->a_szc != 0) {
2482 					anon_shmap_free_pages(amp, 0,
2483 					    amp->size);
2484 				} else {
2485 					anon_free(amp->ahp, 0, amp->size);
2486 				}
2487 				if ((len = amp->swresv) != 0) {
2488 					anon_unresv_zone(len,
2489 					    seg->s_as->a_proc->p_zone);
2490 					TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2491 					    "anon proc:%p %lu %u", seg, len, 0);
2492 				}
2493 			}
2494 			svd->amp = NULL;
2495 			ANON_LOCK_EXIT(&amp->a_rwlock);
2496 			anonmap_free(amp);
2497 		} else if (svd->type == MAP_PRIVATE) {
2498 			/*
2499 			 * We had a private mapping which still has
2500 			 * a held anon_map so just free up all the
2501 			 * anon slot pointers that we were using.
2502 			 */
2503 			if (seg->s_szc != 0) {
2504 				anon_free_pages(amp->ahp, svd->anon_index,
2505 				    seg->s_size, seg->s_szc);
2506 			} else {
2507 				anon_free(amp->ahp, svd->anon_index,
2508 				    seg->s_size);
2509 			}
2510 			ANON_LOCK_EXIT(&amp->a_rwlock);
2511 		} else {
2512 			ANON_LOCK_EXIT(&amp->a_rwlock);
2513 		}
2514 	}
2515 
2516 	/*
2517 	 * Release swap reservation.
2518 	 */
2519 	if ((len = svd->swresv) != 0) {
2520 		anon_unresv_zone(svd->swresv,
2521 		    seg->s_as->a_proc->p_zone);
2522 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2523 		    seg, len, 0);
2524 		if (SEG_IS_PARTIAL_RESV(seg))
2525 			seg->s_as->a_resvsize -= svd->swresv;
2526 		svd->swresv = 0;
2527 	}
2528 	/*
2529 	 * Release claim on vnode, credentials, and finally free the
2530 	 * private data.
2531 	 */
2532 	if (svd->vp != NULL) {
2533 		if (svd->type == MAP_SHARED)
2534 			lgrp_shm_policy_fini(NULL, svd->vp);
2535 		VN_RELE(svd->vp);
2536 		svd->vp = NULL;
2537 	}
2538 	crfree(svd->cred);
2539 	svd->pageprot = 0;
2540 	svd->pageadvice = 0;
2541 	svd->pageswap = 0;
2542 	svd->cred = NULL;
2543 
2544 	/*
2545 	 * Take segfree_syncmtx lock to let segvn_reclaim() finish if it's
2546 	 * still working with this segment without holding as lock (in case
2547 	 * it's called by pcache async thread).
2548 	 */
2549 	ASSERT(svd->softlockcnt == 0);
2550 	mutex_enter(&svd->segfree_syncmtx);
2551 	mutex_exit(&svd->segfree_syncmtx);
2552 
2553 	seg->s_data = NULL;
2554 	kmem_cache_free(segvn_cache, svd);
2555 }
2556 
2557 /*
2558  * Do a F_SOFTUNLOCK call over the range requested.  The range must have
2559  * already been F_SOFTLOCK'ed.
2560  * Caller must always match addr and len of a softunlock with a previous
2561  * softlock with exactly the same addr and len.
2562  */
2563 static void
2564 segvn_softunlock(struct seg *seg, caddr_t addr, size_t len, enum seg_rw rw)
2565 {
2566 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2567 	page_t *pp;
2568 	caddr_t adr;
2569 	struct vnode *vp;
2570 	u_offset_t offset;
2571 	ulong_t anon_index;
2572 	struct anon_map *amp;
2573 	struct anon *ap = NULL;
2574 
2575 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2576 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
2577 
2578 	if ((amp = svd->amp) != NULL)
2579 		anon_index = svd->anon_index + seg_page(seg, addr);
2580 
2581 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
2582 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2583 		hat_unlock_region(seg->s_as->a_hat, addr, len, svd->rcookie);
2584 	} else {
2585 		hat_unlock(seg->s_as->a_hat, addr, len);
2586 	}
2587 	for (adr = addr; adr < addr + len; adr += PAGESIZE) {
2588 		if (amp != NULL) {
2589 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2590 			if ((ap = anon_get_ptr(amp->ahp, anon_index++))
2591 			    != NULL) {
2592 				swap_xlate(ap, &vp, &offset);
2593 			} else {
2594 				vp = svd->vp;
2595 				offset = svd->offset +
2596 				    (uintptr_t)(adr - seg->s_base);
2597 			}
2598 			ANON_LOCK_EXIT(&amp->a_rwlock);
2599 		} else {
2600 			vp = svd->vp;
2601 			offset = svd->offset +
2602 			    (uintptr_t)(adr - seg->s_base);
2603 		}
2604 
2605 		/*
2606 		 * Use page_find() instead of page_lookup() to
2607 		 * find the page since we know that it is locked.
2608 		 */
2609 		pp = page_find(vp, offset);
2610 		if (pp == NULL) {
2611 			panic(
2612 			    "segvn_softunlock: addr %p, ap %p, vp %p, off %llx",
2613 			    (void *)adr, (void *)ap, (void *)vp, offset);
2614 			/*NOTREACHED*/
2615 		}
2616 
2617 		if (rw == S_WRITE) {
2618 			hat_setrefmod(pp);
2619 			if (seg->s_as->a_vbits)
2620 				hat_setstat(seg->s_as, adr, PAGESIZE,
2621 				    P_REF | P_MOD);
2622 		} else if (rw != S_OTHER) {
2623 			hat_setref(pp);
2624 			if (seg->s_as->a_vbits)
2625 				hat_setstat(seg->s_as, adr, PAGESIZE, P_REF);
2626 		}
2627 		TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2628 		    "segvn_fault:pp %p vp %p offset %llx", pp, vp, offset);
2629 		page_unlock(pp);
2630 	}
2631 	ASSERT(svd->softlockcnt >= btop(len));
2632 	if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -btop(len))) {
2633 		/*
2634 		 * All SOFTLOCKS are gone. Wakeup any waiting
2635 		 * unmappers so they can try again to unmap.
2636 		 * Check for waiters first without the mutex
2637 		 * held so we don't always grab the mutex on
2638 		 * softunlocks.
2639 		 */
2640 		if (AS_ISUNMAPWAIT(seg->s_as)) {
2641 			mutex_enter(&seg->s_as->a_contents);
2642 			if (AS_ISUNMAPWAIT(seg->s_as)) {
2643 				AS_CLRUNMAPWAIT(seg->s_as);
2644 				cv_broadcast(&seg->s_as->a_cv);
2645 			}
2646 			mutex_exit(&seg->s_as->a_contents);
2647 		}
2648 	}
2649 }
2650 
2651 #define	PAGE_HANDLED	((page_t *)-1)
2652 
2653 /*
2654  * Release all the pages in the NULL terminated ppp list
2655  * which haven't already been converted to PAGE_HANDLED.
2656  */
2657 static void
2658 segvn_pagelist_rele(page_t **ppp)
2659 {
2660 	for (; *ppp != NULL; ppp++) {
2661 		if (*ppp != PAGE_HANDLED)
2662 			page_unlock(*ppp);
2663 	}
2664 }
2665 
2666 static int stealcow = 1;
2667 
2668 /*
2669  * Workaround for viking chip bug.  See bug id 1220902.
2670  * To fix this down in pagefault() would require importing so
2671  * much as and segvn code as to be unmaintainable.
2672  */
2673 int enable_mbit_wa = 0;
2674 
2675 /*
2676  * Handles all the dirty work of getting the right
2677  * anonymous pages and loading up the translations.
2678  * This routine is called only from segvn_fault()
2679  * when looping over the range of addresses requested.
2680  *
2681  * The basic algorithm here is:
2682  * 	If this is an anon_zero case
2683  *		Call anon_zero to allocate page
2684  *		Load up translation
2685  *		Return
2686  *	endif
2687  *	If this is an anon page
2688  *		Use anon_getpage to get the page
2689  *	else
2690  *		Find page in pl[] list passed in
2691  *	endif
2692  *	If not a cow
2693  *		Load up the translation to the page
2694  *		return
2695  *	endif
2696  *	Call anon_private to handle cow
2697  *	Load up (writable) translation to new page
2698  */
2699 static faultcode_t
2700 segvn_faultpage(
2701 	struct hat *hat,		/* the hat to use for mapping */
2702 	struct seg *seg,		/* seg_vn of interest */
2703 	caddr_t addr,			/* address in as */
2704 	u_offset_t off,			/* offset in vp */
2705 	struct vpage *vpage,		/* pointer to vpage for vp, off */
2706 	page_t *pl[],			/* object source page pointer */
2707 	uint_t vpprot,			/* access allowed to object pages */
2708 	enum fault_type type,		/* type of fault */
2709 	enum seg_rw rw,			/* type of access at fault */
2710 	int brkcow)			/* we may need to break cow */
2711 {
2712 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2713 	page_t *pp, **ppp;
2714 	uint_t pageflags = 0;
2715 	page_t *anon_pl[1 + 1];
2716 	page_t *opp = NULL;		/* original page */
2717 	uint_t prot;
2718 	int err;
2719 	int cow;
2720 	int claim;
2721 	int steal = 0;
2722 	ulong_t anon_index;
2723 	struct anon *ap, *oldap;
2724 	struct anon_map *amp;
2725 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
2726 	int anon_lock = 0;
2727 	anon_sync_obj_t cookie;
2728 
2729 	if (svd->flags & MAP_TEXT) {
2730 		hat_flag |= HAT_LOAD_TEXT;
2731 	}
2732 
2733 	ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
2734 	ASSERT(seg->s_szc == 0);
2735 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
2736 
2737 	/*
2738 	 * Initialize protection value for this page.
2739 	 * If we have per page protection values check it now.
2740 	 */
2741 	if (svd->pageprot) {
2742 		uint_t protchk;
2743 
2744 		switch (rw) {
2745 		case S_READ:
2746 			protchk = PROT_READ;
2747 			break;
2748 		case S_WRITE:
2749 			protchk = PROT_WRITE;
2750 			break;
2751 		case S_EXEC:
2752 			protchk = PROT_EXEC;
2753 			break;
2754 		case S_OTHER:
2755 		default:
2756 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
2757 			break;
2758 		}
2759 
2760 		prot = VPP_PROT(vpage);
2761 		if ((prot & protchk) == 0)
2762 			return (FC_PROT);	/* illegal access type */
2763 	} else {
2764 		prot = svd->prot;
2765 	}
2766 
2767 	if (type == F_SOFTLOCK) {
2768 		atomic_inc_ulong((ulong_t *)&svd->softlockcnt);
2769 	}
2770 
2771 	/*
2772 	 * Always acquire the anon array lock to prevent 2 threads from
2773 	 * allocating separate anon slots for the same "addr".
2774 	 */
2775 
2776 	if ((amp = svd->amp) != NULL) {
2777 		ASSERT(RW_READ_HELD(&amp->a_rwlock));
2778 		anon_index = svd->anon_index + seg_page(seg, addr);
2779 		anon_array_enter(amp, anon_index, &cookie);
2780 		anon_lock = 1;
2781 	}
2782 
2783 	if (svd->vp == NULL && amp != NULL) {
2784 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL) {
2785 			/*
2786 			 * Allocate a (normally) writable anonymous page of
2787 			 * zeroes. If no advance reservations, reserve now.
2788 			 */
2789 			if (svd->flags & MAP_NORESERVE) {
2790 				if (anon_resv_zone(ptob(1),
2791 				    seg->s_as->a_proc->p_zone)) {
2792 					atomic_add_long(&svd->swresv, ptob(1));
2793 					atomic_add_long(&seg->s_as->a_resvsize,
2794 					    ptob(1));
2795 				} else {
2796 					err = ENOMEM;
2797 					goto out;
2798 				}
2799 			}
2800 			if ((pp = anon_zero(seg, addr, &ap,
2801 			    svd->cred)) == NULL) {
2802 				err = ENOMEM;
2803 				goto out;	/* out of swap space */
2804 			}
2805 			/*
2806 			 * Re-acquire the anon_map lock and
2807 			 * initialize the anon array entry.
2808 			 */
2809 			(void) anon_set_ptr(amp->ahp, anon_index, ap,
2810 			    ANON_SLEEP);
2811 
2812 			ASSERT(pp->p_szc == 0);
2813 
2814 			/*
2815 			 * Handle pages that have been marked for migration
2816 			 */
2817 			if (lgrp_optimizations())
2818 				page_migrate(seg, addr, &pp, 1);
2819 
2820 			if (enable_mbit_wa) {
2821 				if (rw == S_WRITE)
2822 					hat_setmod(pp);
2823 				else if (!hat_ismod(pp))
2824 					prot &= ~PROT_WRITE;
2825 			}
2826 			/*
2827 			 * If AS_PAGLCK is set in a_flags (via memcntl(2)
2828 			 * with MC_LOCKAS, MCL_FUTURE) and this is a
2829 			 * MAP_NORESERVE segment, we may need to
2830 			 * permanently lock the page as it is being faulted
2831 			 * for the first time. The following text applies
2832 			 * only to MAP_NORESERVE segments:
2833 			 *
2834 			 * As per memcntl(2), if this segment was created
2835 			 * after MCL_FUTURE was applied (a "future"
2836 			 * segment), its pages must be locked.  If this
2837 			 * segment existed at MCL_FUTURE application (a
2838 			 * "past" segment), the interface is unclear.
2839 			 *
2840 			 * We decide to lock only if vpage is present:
2841 			 *
2842 			 * - "future" segments will have a vpage array (see
2843 			 *    as_map), and so will be locked as required
2844 			 *
2845 			 * - "past" segments may not have a vpage array,
2846 			 *    depending on whether events (such as
2847 			 *    mprotect) have occurred. Locking if vpage
2848 			 *    exists will preserve legacy behavior.  Not
2849 			 *    locking if vpage is absent, will not break
2850 			 *    the interface or legacy behavior.  Note that
2851 			 *    allocating vpage here if it's absent requires
2852 			 *    upgrading the segvn reader lock, the cost of
2853 			 *    which does not seem worthwhile.
2854 			 *
2855 			 * Usually testing and setting VPP_ISPPLOCK and
2856 			 * VPP_SETPPLOCK requires holding the segvn lock as
2857 			 * writer, but in this case all readers are
2858 			 * serializing on the anon array lock.
2859 			 */
2860 			if (AS_ISPGLCK(seg->s_as) && vpage != NULL &&
2861 			    (svd->flags & MAP_NORESERVE) &&
2862 			    !VPP_ISPPLOCK(vpage)) {
2863 				proc_t *p = seg->s_as->a_proc;
2864 				ASSERT(svd->type == MAP_PRIVATE);
2865 				mutex_enter(&p->p_lock);
2866 				if (rctl_incr_locked_mem(p, NULL, PAGESIZE,
2867 				    1) == 0) {
2868 					claim = VPP_PROT(vpage) & PROT_WRITE;
2869 					if (page_pp_lock(pp, claim, 0)) {
2870 						VPP_SETPPLOCK(vpage);
2871 					} else {
2872 						rctl_decr_locked_mem(p, NULL,
2873 						    PAGESIZE, 1);
2874 					}
2875 				}
2876 				mutex_exit(&p->p_lock);
2877 			}
2878 
2879 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2880 			hat_memload(hat, addr, pp, prot, hat_flag);
2881 
2882 			if (!(hat_flag & HAT_LOAD_LOCK))
2883 				page_unlock(pp);
2884 
2885 			anon_array_exit(&cookie);
2886 			return (0);
2887 		}
2888 	}
2889 
2890 	/*
2891 	 * Obtain the page structure via anon_getpage() if it is
2892 	 * a private copy of an object (the result of a previous
2893 	 * copy-on-write).
2894 	 */
2895 	if (amp != NULL) {
2896 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) != NULL) {
2897 			err = anon_getpage(&ap, &vpprot, anon_pl, PAGESIZE,
2898 			    seg, addr, rw, svd->cred);
2899 			if (err)
2900 				goto out;
2901 
2902 			if (svd->type == MAP_SHARED) {
2903 				/*
2904 				 * If this is a shared mapping to an
2905 				 * anon_map, then ignore the write
2906 				 * permissions returned by anon_getpage().
2907 				 * They apply to the private mappings
2908 				 * of this anon_map.
2909 				 */
2910 				vpprot |= PROT_WRITE;
2911 			}
2912 			opp = anon_pl[0];
2913 		}
2914 	}
2915 
2916 	/*
2917 	 * Search the pl[] list passed in if it is from the
2918 	 * original object (i.e., not a private copy).
2919 	 */
2920 	if (opp == NULL) {
2921 		/*
2922 		 * Find original page.  We must be bringing it in
2923 		 * from the list in pl[].
2924 		 */
2925 		for (ppp = pl; (opp = *ppp) != NULL; ppp++) {
2926 			if (opp == PAGE_HANDLED)
2927 				continue;
2928 			ASSERT(opp->p_vnode == svd->vp); /* XXX */
2929 			if (opp->p_offset == off)
2930 				break;
2931 		}
2932 		if (opp == NULL) {
2933 			panic("segvn_faultpage not found");
2934 			/*NOTREACHED*/
2935 		}
2936 		*ppp = PAGE_HANDLED;
2937 
2938 	}
2939 
2940 	ASSERT(PAGE_LOCKED(opp));
2941 
2942 	TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2943 	    "segvn_fault:pp %p vp %p offset %llx", opp, NULL, 0);
2944 
2945 	/*
2946 	 * The fault is treated as a copy-on-write fault if a
2947 	 * write occurs on a private segment and the object
2948 	 * page (i.e., mapping) is write protected.  We assume
2949 	 * that fatal protection checks have already been made.
2950 	 */
2951 
2952 	if (brkcow) {
2953 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2954 		cow = !(vpprot & PROT_WRITE);
2955 	} else if (svd->tr_state == SEGVN_TR_ON) {
2956 		/*
2957 		 * If we are doing text replication COW on first touch.
2958 		 */
2959 		ASSERT(amp != NULL);
2960 		ASSERT(svd->vp != NULL);
2961 		ASSERT(rw != S_WRITE);
2962 		cow = (ap == NULL);
2963 	} else {
2964 		cow = 0;
2965 	}
2966 
2967 	/*
2968 	 * If not a copy-on-write case load the translation
2969 	 * and return.
2970 	 */
2971 	if (cow == 0) {
2972 
2973 		/*
2974 		 * Handle pages that have been marked for migration
2975 		 */
2976 		if (lgrp_optimizations())
2977 			page_migrate(seg, addr, &opp, 1);
2978 
2979 		if (IS_VMODSORT(opp->p_vnode) || enable_mbit_wa) {
2980 			if (rw == S_WRITE)
2981 				hat_setmod(opp);
2982 			else if (rw != S_OTHER && !hat_ismod(opp))
2983 				prot &= ~PROT_WRITE;
2984 		}
2985 
2986 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
2987 		    (!svd->pageprot && svd->prot == (prot & vpprot)));
2988 		ASSERT(amp == NULL ||
2989 		    svd->rcookie == HAT_INVALID_REGION_COOKIE);
2990 		hat_memload_region(hat, addr, opp, prot & vpprot, hat_flag,
2991 		    svd->rcookie);
2992 
2993 		if (!(hat_flag & HAT_LOAD_LOCK))
2994 			page_unlock(opp);
2995 
2996 		if (anon_lock) {
2997 			anon_array_exit(&cookie);
2998 		}
2999 		return (0);
3000 	}
3001 
3002 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3003 
3004 	hat_setref(opp);
3005 
3006 	ASSERT(amp != NULL && anon_lock);
3007 
3008 	/*
3009 	 * Steal the page only if it isn't a private page
3010 	 * since stealing a private page is not worth the effort.
3011 	 */
3012 	if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL)
3013 		steal = 1;
3014 
3015 	/*
3016 	 * Steal the original page if the following conditions are true:
3017 	 *
3018 	 * We are low on memory, the page is not private, page is not large,
3019 	 * not shared, not modified, not `locked' or if we have it `locked'
3020 	 * (i.e., p_cowcnt == 1 and p_lckcnt == 0, which also implies
3021 	 * that the page is not shared) and if it doesn't have any
3022 	 * translations. page_struct_lock isn't needed to look at p_cowcnt
3023 	 * and p_lckcnt because we first get exclusive lock on page.
3024 	 */
3025 	(void) hat_pagesync(opp, HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD);
3026 
3027 	if (stealcow && freemem < minfree && steal && opp->p_szc == 0 &&
3028 	    page_tryupgrade(opp) && !hat_ismod(opp) &&
3029 	    ((opp->p_lckcnt == 0 && opp->p_cowcnt == 0) ||
3030 	    (opp->p_lckcnt == 0 && opp->p_cowcnt == 1 &&
3031 	    vpage != NULL && VPP_ISPPLOCK(vpage)))) {
3032 		/*
3033 		 * Check if this page has other translations
3034 		 * after unloading our translation.
3035 		 */
3036 		if (hat_page_is_mapped(opp)) {
3037 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3038 			hat_unload(seg->s_as->a_hat, addr, PAGESIZE,
3039 			    HAT_UNLOAD);
3040 		}
3041 
3042 		/*
3043 		 * hat_unload() might sync back someone else's recent
3044 		 * modification, so check again.
3045 		 */
3046 		if (!hat_ismod(opp) && !hat_page_is_mapped(opp))
3047 			pageflags |= STEAL_PAGE;
3048 	}
3049 
3050 	/*
3051 	 * If we have a vpage pointer, see if it indicates that we have
3052 	 * ``locked'' the page we map -- if so, tell anon_private to
3053 	 * transfer the locking resource to the new page.
3054 	 *
3055 	 * See Statement at the beginning of segvn_lockop regarding
3056 	 * the way lockcnts/cowcnts are handled during COW.
3057 	 *
3058 	 */
3059 	if (vpage != NULL && VPP_ISPPLOCK(vpage))
3060 		pageflags |= LOCK_PAGE;
3061 
3062 	/*
3063 	 * Allocate a private page and perform the copy.
3064 	 * For MAP_NORESERVE reserve swap space now, unless this
3065 	 * is a cow fault on an existing anon page in which case
3066 	 * MAP_NORESERVE will have made advance reservations.
3067 	 */
3068 	if ((svd->flags & MAP_NORESERVE) && (ap == NULL)) {
3069 		if (anon_resv_zone(ptob(1), seg->s_as->a_proc->p_zone)) {
3070 			atomic_add_long(&svd->swresv, ptob(1));
3071 			atomic_add_long(&seg->s_as->a_resvsize, ptob(1));
3072 		} else {
3073 			page_unlock(opp);
3074 			err = ENOMEM;
3075 			goto out;
3076 		}
3077 	}
3078 	oldap = ap;
3079 	pp = anon_private(&ap, seg, addr, prot, opp, pageflags, svd->cred);
3080 	if (pp == NULL) {
3081 		err = ENOMEM;	/* out of swap space */
3082 		goto out;
3083 	}
3084 
3085 	/*
3086 	 * If we copied away from an anonymous page, then
3087 	 * we are one step closer to freeing up an anon slot.
3088 	 *
3089 	 * NOTE:  The original anon slot must be released while
3090 	 * holding the "anon_map" lock.  This is necessary to prevent
3091 	 * other threads from obtaining a pointer to the anon slot
3092 	 * which may be freed if its "refcnt" is 1.
3093 	 */
3094 	if (oldap != NULL)
3095 		anon_decref(oldap);
3096 
3097 	(void) anon_set_ptr(amp->ahp, anon_index, ap, ANON_SLEEP);
3098 
3099 	/*
3100 	 * Handle pages that have been marked for migration
3101 	 */
3102 	if (lgrp_optimizations())
3103 		page_migrate(seg, addr, &pp, 1);
3104 
3105 	ASSERT(pp->p_szc == 0);
3106 
3107 	ASSERT(!IS_VMODSORT(pp->p_vnode));
3108 	if (enable_mbit_wa) {
3109 		if (rw == S_WRITE)
3110 			hat_setmod(pp);
3111 		else if (!hat_ismod(pp))
3112 			prot &= ~PROT_WRITE;
3113 	}
3114 
3115 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3116 	hat_memload(hat, addr, pp, prot, hat_flag);
3117 
3118 	if (!(hat_flag & HAT_LOAD_LOCK))
3119 		page_unlock(pp);
3120 
3121 	ASSERT(anon_lock);
3122 	anon_array_exit(&cookie);
3123 	return (0);
3124 out:
3125 	if (anon_lock)
3126 		anon_array_exit(&cookie);
3127 
3128 	if (type == F_SOFTLOCK) {
3129 		atomic_dec_ulong((ulong_t *)&svd->softlockcnt);
3130 	}
3131 	return (FC_MAKE_ERR(err));
3132 }
3133 
3134 /*
3135  * relocate a bunch of smaller targ pages into one large repl page. all targ
3136  * pages must be complete pages smaller than replacement pages.
3137  * it's assumed that no page's szc can change since they are all PAGESIZE or
3138  * complete large pages locked SHARED.
3139  */
3140 static void
3141 segvn_relocate_pages(page_t **targ, page_t *replacement)
3142 {
3143 	page_t *pp;
3144 	pgcnt_t repl_npgs, curnpgs;
3145 	pgcnt_t i;
3146 	uint_t repl_szc = replacement->p_szc;
3147 	page_t *first_repl = replacement;
3148 	page_t *repl;
3149 	spgcnt_t npgs;
3150 
3151 	VM_STAT_ADD(segvnvmstats.relocatepages[0]);
3152 
3153 	ASSERT(repl_szc != 0);
3154 	npgs = repl_npgs = page_get_pagecnt(repl_szc);
3155 
3156 	i = 0;
3157 	while (repl_npgs) {
3158 		spgcnt_t nreloc;
3159 		int err;
3160 		ASSERT(replacement != NULL);
3161 		pp = targ[i];
3162 		ASSERT(pp->p_szc < repl_szc);
3163 		ASSERT(PAGE_EXCL(pp));
3164 		ASSERT(!PP_ISFREE(pp));
3165 		curnpgs = page_get_pagecnt(pp->p_szc);
3166 		if (curnpgs == 1) {
3167 			VM_STAT_ADD(segvnvmstats.relocatepages[1]);
3168 			repl = replacement;
3169 			page_sub(&replacement, repl);
3170 			ASSERT(PAGE_EXCL(repl));
3171 			ASSERT(!PP_ISFREE(repl));
3172 			ASSERT(repl->p_szc == repl_szc);
3173 		} else {
3174 			page_t *repl_savepp;
3175 			int j;
3176 			VM_STAT_ADD(segvnvmstats.relocatepages[2]);
3177 			repl_savepp = replacement;
3178 			for (j = 0; j < curnpgs; j++) {
3179 				repl = replacement;
3180 				page_sub(&replacement, repl);
3181 				ASSERT(PAGE_EXCL(repl));
3182 				ASSERT(!PP_ISFREE(repl));
3183 				ASSERT(repl->p_szc == repl_szc);
3184 				ASSERT(page_pptonum(targ[i + j]) ==
3185 				    page_pptonum(targ[i]) + j);
3186 			}
3187 			repl = repl_savepp;
3188 			ASSERT(IS_P2ALIGNED(page_pptonum(repl), curnpgs));
3189 		}
3190 		err = page_relocate(&pp, &repl, 0, 1, &nreloc, NULL);
3191 		if (err || nreloc != curnpgs) {
3192 			panic("segvn_relocate_pages: "
3193 			    "page_relocate failed err=%d curnpgs=%ld "
3194 			    "nreloc=%ld", err, curnpgs, nreloc);
3195 		}
3196 		ASSERT(curnpgs <= repl_npgs);
3197 		repl_npgs -= curnpgs;
3198 		i += curnpgs;
3199 	}
3200 	ASSERT(replacement == NULL);
3201 
3202 	repl = first_repl;
3203 	repl_npgs = npgs;
3204 	for (i = 0; i < repl_npgs; i++) {
3205 		ASSERT(PAGE_EXCL(repl));
3206 		ASSERT(!PP_ISFREE(repl));
3207 		targ[i] = repl;
3208 		page_downgrade(targ[i]);
3209 		repl++;
3210 	}
3211 }
3212 
3213 /*
3214  * Check if all pages in ppa array are complete smaller than szc pages and
3215  * their roots will still be aligned relative to their current size if the
3216  * entire ppa array is relocated into one szc page. If these conditions are
3217  * not met return 0.
3218  *
3219  * If all pages are properly aligned attempt to upgrade their locks
3220  * to exclusive mode. If it fails set *upgrdfail to 1 and return 0.
3221  * upgrdfail was set to 0 by caller.
3222  *
3223  * Return 1 if all pages are aligned and locked exclusively.
3224  *
3225  * If all pages in ppa array happen to be physically contiguous to make one
3226  * szc page and all exclusive locks are successfully obtained promote the page
3227  * size to szc and set *pszc to szc. Return 1 with pages locked shared.
3228  */
3229 static int
3230 segvn_full_szcpages(page_t **ppa, uint_t szc, int *upgrdfail, uint_t *pszc)
3231 {
3232 	page_t *pp;
3233 	pfn_t pfn;
3234 	pgcnt_t totnpgs = page_get_pagecnt(szc);
3235 	pfn_t first_pfn;
3236 	int contig = 1;
3237 	pgcnt_t i;
3238 	pgcnt_t j;
3239 	uint_t curszc;
3240 	pgcnt_t curnpgs;
3241 	int root = 0;
3242 
3243 	ASSERT(szc > 0);
3244 
3245 	VM_STAT_ADD(segvnvmstats.fullszcpages[0]);
3246 
3247 	for (i = 0; i < totnpgs; i++) {
3248 		pp = ppa[i];
3249 		ASSERT(PAGE_SHARED(pp));
3250 		ASSERT(!PP_ISFREE(pp));
3251 		pfn = page_pptonum(pp);
3252 		if (i == 0) {
3253 			if (!IS_P2ALIGNED(pfn, totnpgs)) {
3254 				contig = 0;
3255 			} else {
3256 				first_pfn = pfn;
3257 			}
3258 		} else if (contig && pfn != first_pfn + i) {
3259 			contig = 0;
3260 		}
3261 		if (pp->p_szc == 0) {
3262 			if (root) {
3263 				VM_STAT_ADD(segvnvmstats.fullszcpages[1]);
3264 				return (0);
3265 			}
3266 		} else if (!root) {
3267 			if ((curszc = pp->p_szc) >= szc) {
3268 				VM_STAT_ADD(segvnvmstats.fullszcpages[2]);
3269 				return (0);
3270 			}
3271 			if (curszc == 0) {
3272 				/*
3273 				 * p_szc changed means we don't have all pages
3274 				 * locked. return failure.
3275 				 */
3276 				VM_STAT_ADD(segvnvmstats.fullszcpages[3]);
3277 				return (0);
3278 			}
3279 			curnpgs = page_get_pagecnt(curszc);
3280 			if (!IS_P2ALIGNED(pfn, curnpgs) ||
3281 			    !IS_P2ALIGNED(i, curnpgs)) {
3282 				VM_STAT_ADD(segvnvmstats.fullszcpages[4]);
3283 				return (0);
3284 			}
3285 			root = 1;
3286 		} else {
3287 			ASSERT(i > 0);
3288 			VM_STAT_ADD(segvnvmstats.fullszcpages[5]);
3289 			if (pp->p_szc != curszc) {
3290 				VM_STAT_ADD(segvnvmstats.fullszcpages[6]);
3291 				return (0);
3292 			}
3293 			if (pfn - 1 != page_pptonum(ppa[i - 1])) {
3294 				panic("segvn_full_szcpages: "
3295 				    "large page not physically contiguous");
3296 			}
3297 			if (P2PHASE(pfn, curnpgs) == curnpgs - 1) {
3298 				root = 0;
3299 			}
3300 		}
3301 	}
3302 
3303 	for (i = 0; i < totnpgs; i++) {
3304 		ASSERT(ppa[i]->p_szc < szc);
3305 		if (!page_tryupgrade(ppa[i])) {
3306 			for (j = 0; j < i; j++) {
3307 				page_downgrade(ppa[j]);
3308 			}
3309 			*pszc = ppa[i]->p_szc;
3310 			*upgrdfail = 1;
3311 			VM_STAT_ADD(segvnvmstats.fullszcpages[7]);
3312 			return (0);
3313 		}
3314 	}
3315 
3316 	/*
3317 	 * When a page is put a free cachelist its szc is set to 0.  if file
3318 	 * system reclaimed pages from cachelist targ pages will be physically
3319 	 * contiguous with 0 p_szc.  in this case just upgrade szc of targ
3320 	 * pages without any relocations.
3321 	 * To avoid any hat issues with previous small mappings
3322 	 * hat_pageunload() the target pages first.
3323 	 */
3324 	if (contig) {
3325 		VM_STAT_ADD(segvnvmstats.fullszcpages[8]);
3326 		for (i = 0; i < totnpgs; i++) {
3327 			(void) hat_pageunload(ppa[i], HAT_FORCE_PGUNLOAD);
3328 		}
3329 		for (i = 0; i < totnpgs; i++) {
3330 			ppa[i]->p_szc = szc;
3331 		}
3332 		for (i = 0; i < totnpgs; i++) {
3333 			ASSERT(PAGE_EXCL(ppa[i]));
3334 			page_downgrade(ppa[i]);
3335 		}
3336 		if (pszc != NULL) {
3337 			*pszc = szc;
3338 		}
3339 	}
3340 	VM_STAT_ADD(segvnvmstats.fullszcpages[9]);
3341 	return (1);
3342 }
3343 
3344 /*
3345  * Create physically contiguous pages for [vp, off] - [vp, off +
3346  * page_size(szc)) range and for private segment return them in ppa array.
3347  * Pages are created either via IO or relocations.
3348  *
3349  * Return 1 on success and 0 on failure.
3350  *
3351  * If physically contiguous pages already exist for this range return 1 without
3352  * filling ppa array. Caller initializes ppa[0] as NULL to detect that ppa
3353  * array wasn't filled. In this case caller fills ppa array via VOP_GETPAGE().
3354  */
3355 
3356 static int
3357 segvn_fill_vp_pages(struct segvn_data *svd, vnode_t *vp, u_offset_t off,
3358     uint_t szc, page_t **ppa, page_t **ppplist, uint_t *ret_pszc,
3359     int *downsize)
3360 
3361 {
3362 	page_t *pplist = *ppplist;
3363 	size_t pgsz = page_get_pagesize(szc);
3364 	pgcnt_t pages = btop(pgsz);
3365 	ulong_t start_off = off;
3366 	u_offset_t eoff = off + pgsz;
3367 	spgcnt_t nreloc;
3368 	u_offset_t io_off = off;
3369 	size_t io_len;
3370 	page_t *io_pplist = NULL;
3371 	page_t *done_pplist = NULL;
3372 	pgcnt_t pgidx = 0;
3373 	page_t *pp;
3374 	page_t *newpp;
3375 	page_t *targpp;
3376 	int io_err = 0;
3377 	int i;
3378 	pfn_t pfn;
3379 	ulong_t ppages;
3380 	page_t *targ_pplist = NULL;
3381 	page_t *repl_pplist = NULL;
3382 	page_t *tmp_pplist;
3383 	int nios = 0;
3384 	uint_t pszc;
3385 	struct vattr va;
3386 
3387 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[0]);
3388 
3389 	ASSERT(szc != 0);
3390 	ASSERT(pplist->p_szc == szc);
3391 
3392 	/*
3393 	 * downsize will be set to 1 only if we fail to lock pages. this will
3394 	 * allow subsequent faults to try to relocate the page again. If we
3395 	 * fail due to misalignment don't downsize and let the caller map the
3396 	 * whole region with small mappings to avoid more faults into the area
3397 	 * where we can't get large pages anyway.
3398 	 */
3399 	*downsize = 0;
3400 
3401 	while (off < eoff) {
3402 		newpp = pplist;
3403 		ASSERT(newpp != NULL);
3404 		ASSERT(PAGE_EXCL(newpp));
3405 		ASSERT(!PP_ISFREE(newpp));
3406 		/*
3407 		 * we pass NULL for nrelocp to page_lookup_create()
3408 		 * so that it doesn't relocate. We relocate here
3409 		 * later only after we make sure we can lock all
3410 		 * pages in the range we handle and they are all
3411 		 * aligned.
3412 		 */
3413 		pp = page_lookup_create(vp, off, SE_SHARED, newpp, NULL, 0);
3414 		ASSERT(pp != NULL);
3415 		ASSERT(!PP_ISFREE(pp));
3416 		ASSERT(pp->p_vnode == vp);
3417 		ASSERT(pp->p_offset == off);
3418 		if (pp == newpp) {
3419 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[1]);
3420 			page_sub(&pplist, pp);
3421 			ASSERT(PAGE_EXCL(pp));
3422 			ASSERT(page_iolock_assert(pp));
3423 			page_list_concat(&io_pplist, &pp);
3424 			off += PAGESIZE;
3425 			continue;
3426 		}
3427 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[2]);
3428 		pfn = page_pptonum(pp);
3429 		pszc = pp->p_szc;
3430 		if (pszc >= szc && targ_pplist == NULL && io_pplist == NULL &&
3431 		    IS_P2ALIGNED(pfn, pages)) {
3432 			ASSERT(repl_pplist == NULL);
3433 			ASSERT(done_pplist == NULL);
3434 			ASSERT(pplist == *ppplist);
3435 			page_unlock(pp);
3436 			page_free_replacement_page(pplist);
3437 			page_create_putback(pages);
3438 			*ppplist = NULL;
3439 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[3]);
3440 			return (1);
3441 		}
3442 		if (pszc >= szc) {
3443 			page_unlock(pp);
3444 			segvn_faultvnmpss_align_err1++;
3445 			goto out;
3446 		}
3447 		ppages = page_get_pagecnt(pszc);
3448 		if (!IS_P2ALIGNED(pfn, ppages)) {
3449 			ASSERT(pszc > 0);
3450 			/*
3451 			 * sizing down to pszc won't help.
3452 			 */
3453 			page_unlock(pp);
3454 			segvn_faultvnmpss_align_err2++;
3455 			goto out;
3456 		}
3457 		pfn = page_pptonum(newpp);
3458 		if (!IS_P2ALIGNED(pfn, ppages)) {
3459 			ASSERT(pszc > 0);
3460 			/*
3461 			 * sizing down to pszc won't help.
3462 			 */
3463 			page_unlock(pp);
3464 			segvn_faultvnmpss_align_err3++;
3465 			goto out;
3466 		}
3467 		if (!PAGE_EXCL(pp)) {
3468 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[4]);
3469 			page_unlock(pp);
3470 			*downsize = 1;
3471 			*ret_pszc = pp->p_szc;
3472 			goto out;
3473 		}
3474 		targpp = pp;
3475 		if (io_pplist != NULL) {
3476 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[5]);
3477 			io_len = off - io_off;
3478 			/*
3479 			 * Some file systems like NFS don't check EOF
3480 			 * conditions in VOP_PAGEIO(). Check it here
3481 			 * now that pages are locked SE_EXCL. Any file
3482 			 * truncation will wait until the pages are
3483 			 * unlocked so no need to worry that file will
3484 			 * be truncated after we check its size here.
3485 			 * XXX fix NFS to remove this check.
3486 			 */
3487 			va.va_mask = AT_SIZE;
3488 			if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL)) {
3489 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[6]);
3490 				page_unlock(targpp);
3491 				goto out;
3492 			}
3493 			if (btopr(va.va_size) < btopr(io_off + io_len)) {
3494 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[7]);
3495 				*downsize = 1;
3496 				*ret_pszc = 0;
3497 				page_unlock(targpp);
3498 				goto out;
3499 			}
3500 			io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3501 				B_READ, svd->cred, NULL);
3502 			if (io_err) {
3503 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[8]);
3504 				page_unlock(targpp);
3505 				if (io_err == EDEADLK) {
3506 					segvn_vmpss_pageio_deadlk_err++;
3507 				}
3508 				goto out;
3509 			}
3510 			nios++;
3511 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[9]);
3512 			while (io_pplist != NULL) {
3513 				pp = io_pplist;
3514 				page_sub(&io_pplist, pp);
3515 				ASSERT(page_iolock_assert(pp));
3516 				page_io_unlock(pp);
3517 				pgidx = (pp->p_offset - start_off) >>
3518 				    PAGESHIFT;
3519 				ASSERT(pgidx < pages);
3520 				ppa[pgidx] = pp;
3521 				page_list_concat(&done_pplist, &pp);
3522 			}
3523 		}
3524 		pp = targpp;
3525 		ASSERT(PAGE_EXCL(pp));
3526 		ASSERT(pp->p_szc <= pszc);
3527 		if (pszc != 0 && !group_page_trylock(pp, SE_EXCL)) {
3528 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[10]);
3529 			page_unlock(pp);
3530 			*downsize = 1;
3531 			*ret_pszc = pp->p_szc;
3532 			goto out;
3533 		}
3534 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[11]);
3535 		/*
3536 		 * page szc chould have changed before the entire group was
3537 		 * locked. reread page szc.
3538 		 */
3539 		pszc = pp->p_szc;
3540 		ppages = page_get_pagecnt(pszc);
3541 
3542 		/* link just the roots */
3543 		page_list_concat(&targ_pplist, &pp);
3544 		page_sub(&pplist, newpp);
3545 		page_list_concat(&repl_pplist, &newpp);
3546 		off += PAGESIZE;
3547 		while (--ppages != 0) {
3548 			newpp = pplist;
3549 			page_sub(&pplist, newpp);
3550 			off += PAGESIZE;
3551 		}
3552 		io_off = off;
3553 	}
3554 	if (io_pplist != NULL) {
3555 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[12]);
3556 		io_len = eoff - io_off;
3557 		va.va_mask = AT_SIZE;
3558 		if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL) != 0) {
3559 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[13]);
3560 			goto out;
3561 		}
3562 		if (btopr(va.va_size) < btopr(io_off + io_len)) {
3563 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[14]);
3564 			*downsize = 1;
3565 			*ret_pszc = 0;
3566 			goto out;
3567 		}
3568 		io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3569 		    B_READ, svd->cred, NULL);
3570 		if (io_err) {
3571 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[15]);
3572 			if (io_err == EDEADLK) {
3573 				segvn_vmpss_pageio_deadlk_err++;
3574 			}
3575 			goto out;
3576 		}
3577 		nios++;
3578 		while (io_pplist != NULL) {
3579 			pp = io_pplist;
3580 			page_sub(&io_pplist, pp);
3581 			ASSERT(page_iolock_assert(pp));
3582 			page_io_unlock(pp);
3583 			pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3584 			ASSERT(pgidx < pages);
3585 			ppa[pgidx] = pp;
3586 		}
3587 	}
3588 	/*
3589 	 * we're now bound to succeed or panic.
3590 	 * remove pages from done_pplist. it's not needed anymore.
3591 	 */
3592 	while (done_pplist != NULL) {
3593 		pp = done_pplist;
3594 		page_sub(&done_pplist, pp);
3595 	}
3596 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[16]);
3597 	ASSERT(pplist == NULL);
3598 	*ppplist = NULL;
3599 	while (targ_pplist != NULL) {
3600 		int ret;
3601 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[17]);
3602 		ASSERT(repl_pplist);
3603 		pp = targ_pplist;
3604 		page_sub(&targ_pplist, pp);
3605 		pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3606 		newpp = repl_pplist;
3607 		page_sub(&repl_pplist, newpp);
3608 #ifdef DEBUG
3609 		pfn = page_pptonum(pp);
3610 		pszc = pp->p_szc;
3611 		ppages = page_get_pagecnt(pszc);
3612 		ASSERT(IS_P2ALIGNED(pfn, ppages));
3613 		pfn = page_pptonum(newpp);
3614 		ASSERT(IS_P2ALIGNED(pfn, ppages));
3615 		ASSERT(P2PHASE(pfn, pages) == pgidx);
3616 #endif
3617 		nreloc = 0;
3618 		ret = page_relocate(&pp, &newpp, 0, 1, &nreloc, NULL);
3619 		if (ret != 0 || nreloc == 0) {
3620 			panic("segvn_fill_vp_pages: "
3621 			    "page_relocate failed");
3622 		}
3623 		pp = newpp;
3624 		while (nreloc-- != 0) {
3625 			ASSERT(PAGE_EXCL(pp));
3626 			ASSERT(pp->p_vnode == vp);
3627 			ASSERT(pgidx ==
3628 			    ((pp->p_offset - start_off) >> PAGESHIFT));
3629 			ppa[pgidx++] = pp;
3630 			pp++;
3631 		}
3632 	}
3633 
3634 	if (svd->type == MAP_PRIVATE) {
3635 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[18]);
3636 		for (i = 0; i < pages; i++) {
3637 			ASSERT(ppa[i] != NULL);
3638 			ASSERT(PAGE_EXCL(ppa[i]));
3639 			ASSERT(ppa[i]->p_vnode == vp);
3640 			ASSERT(ppa[i]->p_offset ==
3641 			    start_off + (i << PAGESHIFT));
3642 			page_downgrade(ppa[i]);
3643 		}
3644 		ppa[pages] = NULL;
3645 	} else {
3646 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[19]);
3647 		/*
3648 		 * the caller will still call VOP_GETPAGE() for shared segments
3649 		 * to check FS write permissions. For private segments we map
3650 		 * file read only anyway.  so no VOP_GETPAGE is needed.
3651 		 */
3652 		for (i = 0; i < pages; i++) {
3653 			ASSERT(ppa[i] != NULL);
3654 			ASSERT(PAGE_EXCL(ppa[i]));
3655 			ASSERT(ppa[i]->p_vnode == vp);
3656 			ASSERT(ppa[i]->p_offset ==
3657 			    start_off + (i << PAGESHIFT));
3658 			page_unlock(ppa[i]);
3659 		}
3660 		ppa[0] = NULL;
3661 	}
3662 
3663 	return (1);
3664 out:
3665 	/*
3666 	 * Do the cleanup. Unlock target pages we didn't relocate. They are
3667 	 * linked on targ_pplist by root pages. reassemble unused replacement
3668 	 * and io pages back to pplist.
3669 	 */
3670 	if (io_pplist != NULL) {
3671 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[20]);
3672 		pp = io_pplist;
3673 		do {
3674 			ASSERT(pp->p_vnode == vp);
3675 			ASSERT(pp->p_offset == io_off);
3676 			ASSERT(page_iolock_assert(pp));
3677 			page_io_unlock(pp);
3678 			page_hashout(pp, NULL);
3679 			io_off += PAGESIZE;
3680 		} while ((pp = pp->p_next) != io_pplist);
3681 		page_list_concat(&io_pplist, &pplist);
3682 		pplist = io_pplist;
3683 	}
3684 	tmp_pplist = NULL;
3685 	while (targ_pplist != NULL) {
3686 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[21]);
3687 		pp = targ_pplist;
3688 		ASSERT(PAGE_EXCL(pp));
3689 		page_sub(&targ_pplist, pp);
3690 
3691 		pszc = pp->p_szc;
3692 		ppages = page_get_pagecnt(pszc);
3693 		ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3694 
3695 		if (pszc != 0) {
3696 			group_page_unlock(pp);
3697 		}
3698 		page_unlock(pp);
3699 
3700 		pp = repl_pplist;
3701 		ASSERT(pp != NULL);
3702 		ASSERT(PAGE_EXCL(pp));
3703 		ASSERT(pp->p_szc == szc);
3704 		page_sub(&repl_pplist, pp);
3705 
3706 		ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3707 
3708 		/* relink replacement page */
3709 		page_list_concat(&tmp_pplist, &pp);
3710 		while (--ppages != 0) {
3711 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[22]);
3712 			pp++;
3713 			ASSERT(PAGE_EXCL(pp));
3714 			ASSERT(pp->p_szc == szc);
3715 			page_list_concat(&tmp_pplist, &pp);
3716 		}
3717 	}
3718 	if (tmp_pplist != NULL) {
3719 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[23]);
3720 		page_list_concat(&tmp_pplist, &pplist);
3721 		pplist = tmp_pplist;
3722 	}
3723 	/*
3724 	 * at this point all pages are either on done_pplist or
3725 	 * pplist. They can't be all on done_pplist otherwise
3726 	 * we'd've been done.
3727 	 */
3728 	ASSERT(pplist != NULL);
3729 	if (nios != 0) {
3730 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[24]);
3731 		pp = pplist;
3732 		do {
3733 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[25]);
3734 			ASSERT(pp->p_szc == szc);
3735 			ASSERT(PAGE_EXCL(pp));
3736 			ASSERT(pp->p_vnode != vp);
3737 			pp->p_szc = 0;
3738 		} while ((pp = pp->p_next) != pplist);
3739 
3740 		pp = done_pplist;
3741 		do {
3742 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[26]);
3743 			ASSERT(pp->p_szc == szc);
3744 			ASSERT(PAGE_EXCL(pp));
3745 			ASSERT(pp->p_vnode == vp);
3746 			pp->p_szc = 0;
3747 		} while ((pp = pp->p_next) != done_pplist);
3748 
3749 		while (pplist != NULL) {
3750 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[27]);
3751 			pp = pplist;
3752 			page_sub(&pplist, pp);
3753 			page_free(pp, 0);
3754 		}
3755 
3756 		while (done_pplist != NULL) {
3757 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[28]);
3758 			pp = done_pplist;
3759 			page_sub(&done_pplist, pp);
3760 			page_unlock(pp);
3761 		}
3762 		*ppplist = NULL;
3763 		return (0);
3764 	}
3765 	ASSERT(pplist == *ppplist);
3766 	if (io_err) {
3767 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[29]);
3768 		/*
3769 		 * don't downsize on io error.
3770 		 * see if vop_getpage succeeds.
3771 		 * pplist may still be used in this case
3772 		 * for relocations.
3773 		 */
3774 		return (0);
3775 	}
3776 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[30]);
3777 	page_free_replacement_page(pplist);
3778 	page_create_putback(pages);
3779 	*ppplist = NULL;
3780 	return (0);
3781 }
3782 
3783 int segvn_anypgsz = 0;
3784 
3785 #define	SEGVN_RESTORE_SOFTLOCK_VP(type, pages) 				\
3786 		if ((type) == F_SOFTLOCK) {				\
3787 			atomic_add_long((ulong_t *)&(svd)->softlockcnt, \
3788 			    -(pages));					\
3789 		}
3790 
3791 #define	SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot)		\
3792 		if (IS_VMODSORT((ppa)[0]->p_vnode)) {			\
3793 			if ((rw) == S_WRITE) {				\
3794 				for (i = 0; i < (pages); i++) {		\
3795 					ASSERT((ppa)[i]->p_vnode ==	\
3796 					    (ppa)[0]->p_vnode);		\
3797 					hat_setmod((ppa)[i]);		\
3798 				}					\
3799 			} else if ((rw) != S_OTHER &&			\
3800 			    ((prot) & (vpprot) & PROT_WRITE)) {		\
3801 				for (i = 0; i < (pages); i++) {		\
3802 					ASSERT((ppa)[i]->p_vnode ==	\
3803 					    (ppa)[0]->p_vnode);		\
3804 					if (!hat_ismod((ppa)[i])) {	\
3805 						prot &= ~PROT_WRITE;	\
3806 						break;			\
3807 					}				\
3808 				}					\
3809 			}						\
3810 		}
3811 
3812 #ifdef  VM_STATS
3813 
3814 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)					\
3815 		VM_STAT_ADD(segvnvmstats.fltvnpages[(idx)]);
3816 
3817 #else /* VM_STATS */
3818 
3819 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)
3820 
3821 #endif
3822 
3823 static faultcode_t
3824 segvn_fault_vnodepages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
3825     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
3826     caddr_t eaddr, int brkcow)
3827 {
3828 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
3829 	struct anon_map *amp = svd->amp;
3830 	uchar_t segtype = svd->type;
3831 	uint_t szc = seg->s_szc;
3832 	size_t pgsz = page_get_pagesize(szc);
3833 	size_t maxpgsz = pgsz;
3834 	pgcnt_t pages = btop(pgsz);
3835 	pgcnt_t maxpages = pages;
3836 	size_t ppasize = (pages + 1) * sizeof (page_t *);
3837 	caddr_t a = lpgaddr;
3838 	caddr_t	maxlpgeaddr = lpgeaddr;
3839 	u_offset_t off = svd->offset + (uintptr_t)(a - seg->s_base);
3840 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
3841 	struct vpage *vpage = (svd->vpage != NULL) ?
3842 	    &svd->vpage[seg_page(seg, a)] : NULL;
3843 	vnode_t *vp = svd->vp;
3844 	page_t **ppa;
3845 	uint_t	pszc;
3846 	size_t	ppgsz;
3847 	pgcnt_t	ppages;
3848 	faultcode_t err = 0;
3849 	int ierr;
3850 	int vop_size_err = 0;
3851 	uint_t protchk, prot, vpprot;
3852 	ulong_t i;
3853 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
3854 	anon_sync_obj_t an_cookie;
3855 	enum seg_rw arw;
3856 	int alloc_failed = 0;
3857 	int adjszc_chk;
3858 	struct vattr va;
3859 	page_t *pplist;
3860 	pfn_t pfn;
3861 	int physcontig;
3862 	int upgrdfail;
3863 	int segvn_anypgsz_vnode = 0; /* for now map vnode with 2 page sizes */
3864 	int tron = (svd->tr_state == SEGVN_TR_ON);
3865 
3866 	ASSERT(szc != 0);
3867 	ASSERT(vp != NULL);
3868 	ASSERT(brkcow == 0 || amp != NULL);
3869 	ASSERT(tron == 0 || amp != NULL);
3870 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
3871 	ASSERT(!(svd->flags & MAP_NORESERVE));
3872 	ASSERT(type != F_SOFTUNLOCK);
3873 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
3874 	ASSERT(amp == NULL || IS_P2ALIGNED(aindx, maxpages));
3875 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
3876 	ASSERT(seg->s_szc < NBBY * sizeof (int));
3877 	ASSERT(type != F_SOFTLOCK || lpgeaddr - a == maxpgsz);
3878 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
3879 
3880 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltvnpages[0]);
3881 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltvnpages[1]);
3882 
3883 	if (svd->flags & MAP_TEXT) {
3884 		hat_flag |= HAT_LOAD_TEXT;
3885 	}
3886 
3887 	if (svd->pageprot) {
3888 		switch (rw) {
3889 		case S_READ:
3890 			protchk = PROT_READ;
3891 			break;
3892 		case S_WRITE:
3893 			protchk = PROT_WRITE;
3894 			break;
3895 		case S_EXEC:
3896 			protchk = PROT_EXEC;
3897 			break;
3898 		case S_OTHER:
3899 		default:
3900 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
3901 			break;
3902 		}
3903 	} else {
3904 		prot = svd->prot;
3905 		/* caller has already done segment level protection check. */
3906 	}
3907 
3908 	if (rw == S_WRITE && segtype == MAP_PRIVATE) {
3909 		SEGVN_VMSTAT_FLTVNPAGES(2);
3910 		arw = S_READ;
3911 	} else {
3912 		arw = rw;
3913 	}
3914 
3915 	ppa = kmem_alloc(ppasize, KM_SLEEP);
3916 
3917 	VM_STAT_COND_ADD(amp != NULL, segvnvmstats.fltvnpages[3]);
3918 
3919 	for (;;) {
3920 		adjszc_chk = 0;
3921 		for (; a < lpgeaddr; a += pgsz, off += pgsz, aindx += pages) {
3922 			if (adjszc_chk) {
3923 				while (szc < seg->s_szc) {
3924 					uintptr_t e;
3925 					uint_t tszc;
3926 					tszc = segvn_anypgsz_vnode ? szc + 1 :
3927 					    seg->s_szc;
3928 					ppgsz = page_get_pagesize(tszc);
3929 					if (!IS_P2ALIGNED(a, ppgsz) ||
3930 					    ((alloc_failed >> tszc) & 0x1)) {
3931 						break;
3932 					}
3933 					SEGVN_VMSTAT_FLTVNPAGES(4);
3934 					szc = tszc;
3935 					pgsz = ppgsz;
3936 					pages = btop(pgsz);
3937 					e = P2ROUNDUP((uintptr_t)eaddr, pgsz);
3938 					lpgeaddr = (caddr_t)e;
3939 				}
3940 			}
3941 
3942 		again:
3943 			if (IS_P2ALIGNED(a, maxpgsz) && amp != NULL) {
3944 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
3945 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
3946 				anon_array_enter(amp, aindx, &an_cookie);
3947 				if (anon_get_ptr(amp->ahp, aindx) != NULL) {
3948 					SEGVN_VMSTAT_FLTVNPAGES(5);
3949 					ASSERT(anon_pages(amp->ahp, aindx,
3950 					    maxpages) == maxpages);
3951 					anon_array_exit(&an_cookie);
3952 					ANON_LOCK_EXIT(&amp->a_rwlock);
3953 					err = segvn_fault_anonpages(hat, seg,
3954 					    a, a + maxpgsz, type, rw,
3955 					    MAX(a, addr),
3956 					    MIN(a + maxpgsz, eaddr), brkcow);
3957 					if (err != 0) {
3958 						SEGVN_VMSTAT_FLTVNPAGES(6);
3959 						goto out;
3960 					}
3961 					if (szc < seg->s_szc) {
3962 						szc = seg->s_szc;
3963 						pgsz = maxpgsz;
3964 						pages = maxpages;
3965 						lpgeaddr = maxlpgeaddr;
3966 					}
3967 					goto next;
3968 				} else {
3969 					ASSERT(anon_pages(amp->ahp, aindx,
3970 					    maxpages) == 0);
3971 					SEGVN_VMSTAT_FLTVNPAGES(7);
3972 					anon_array_exit(&an_cookie);
3973 					ANON_LOCK_EXIT(&amp->a_rwlock);
3974 				}
3975 			}
3976 			ASSERT(!brkcow || IS_P2ALIGNED(a, maxpgsz));
3977 			ASSERT(!tron || IS_P2ALIGNED(a, maxpgsz));
3978 
3979 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
3980 				ASSERT(vpage != NULL);
3981 				prot = VPP_PROT(vpage);
3982 				ASSERT(sameprot(seg, a, maxpgsz));
3983 				if ((prot & protchk) == 0) {
3984 					SEGVN_VMSTAT_FLTVNPAGES(8);
3985 					err = FC_PROT;
3986 					goto out;
3987 				}
3988 			}
3989 			if (type == F_SOFTLOCK) {
3990 				atomic_add_long((ulong_t *)&svd->softlockcnt,
3991 				    pages);
3992 			}
3993 
3994 			pplist = NULL;
3995 			physcontig = 0;
3996 			ppa[0] = NULL;
3997 			if (!brkcow && !tron && szc &&
3998 			    !page_exists_physcontig(vp, off, szc,
3999 			    segtype == MAP_PRIVATE ? ppa : NULL)) {
4000 				SEGVN_VMSTAT_FLTVNPAGES(9);
4001 				if (page_alloc_pages(vp, seg, a, &pplist, NULL,
4002 				    szc, 0, 0) && type != F_SOFTLOCK) {
4003 					SEGVN_VMSTAT_FLTVNPAGES(10);
4004 					pszc = 0;
4005 					ierr = -1;
4006 					alloc_failed |= (1 << szc);
4007 					break;
4008 				}
4009 				if (pplist != NULL &&
4010 				    vp->v_mpssdata == SEGVN_PAGEIO) {
4011 					int downsize;
4012 					SEGVN_VMSTAT_FLTVNPAGES(11);
4013 					physcontig = segvn_fill_vp_pages(svd,
4014 					    vp, off, szc, ppa, &pplist,
4015 					    &pszc, &downsize);
4016 					ASSERT(!physcontig || pplist == NULL);
4017 					if (!physcontig && downsize &&
4018 					    type != F_SOFTLOCK) {
4019 						ASSERT(pplist == NULL);
4020 						SEGVN_VMSTAT_FLTVNPAGES(12);
4021 						ierr = -1;
4022 						break;
4023 					}
4024 					ASSERT(!physcontig ||
4025 					    segtype == MAP_PRIVATE ||
4026 					    ppa[0] == NULL);
4027 					if (physcontig && ppa[0] == NULL) {
4028 						physcontig = 0;
4029 					}
4030 				}
4031 			} else if (!brkcow && !tron && szc && ppa[0] != NULL) {
4032 				SEGVN_VMSTAT_FLTVNPAGES(13);
4033 				ASSERT(segtype == MAP_PRIVATE);
4034 				physcontig = 1;
4035 			}
4036 
4037 			if (!physcontig) {
4038 				SEGVN_VMSTAT_FLTVNPAGES(14);
4039 				ppa[0] = NULL;
4040 				ierr = VOP_GETPAGE(vp, (offset_t)off, pgsz,
4041 				    &vpprot, ppa, pgsz, seg, a, arw,
4042 				    svd->cred, NULL);
4043 #ifdef DEBUG
4044 				if (ierr == 0) {
4045 					for (i = 0; i < pages; i++) {
4046 						ASSERT(PAGE_LOCKED(ppa[i]));
4047 						ASSERT(!PP_ISFREE(ppa[i]));
4048 						ASSERT(ppa[i]->p_vnode == vp);
4049 						ASSERT(ppa[i]->p_offset ==
4050 						    off + (i << PAGESHIFT));
4051 					}
4052 				}
4053 #endif /* DEBUG */
4054 				if (segtype == MAP_PRIVATE) {
4055 					SEGVN_VMSTAT_FLTVNPAGES(15);
4056 					vpprot &= ~PROT_WRITE;
4057 				}
4058 			} else {
4059 				ASSERT(segtype == MAP_PRIVATE);
4060 				SEGVN_VMSTAT_FLTVNPAGES(16);
4061 				vpprot = PROT_ALL & ~PROT_WRITE;
4062 				ierr = 0;
4063 			}
4064 
4065 			if (ierr != 0) {
4066 				SEGVN_VMSTAT_FLTVNPAGES(17);
4067 				if (pplist != NULL) {
4068 					SEGVN_VMSTAT_FLTVNPAGES(18);
4069 					page_free_replacement_page(pplist);
4070 					page_create_putback(pages);
4071 				}
4072 				SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4073 				if (a + pgsz <= eaddr) {
4074 					SEGVN_VMSTAT_FLTVNPAGES(19);
4075 					err = FC_MAKE_ERR(ierr);
4076 					goto out;
4077 				}
4078 				va.va_mask = AT_SIZE;
4079 				if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL)) {
4080 					SEGVN_VMSTAT_FLTVNPAGES(20);
4081 					err = FC_MAKE_ERR(EIO);
4082 					goto out;
4083 				}
4084 				if (btopr(va.va_size) >= btopr(off + pgsz)) {
4085 					SEGVN_VMSTAT_FLTVNPAGES(21);
4086 					err = FC_MAKE_ERR(ierr);
4087 					goto out;
4088 				}
4089 				if (btopr(va.va_size) <
4090 				    btopr(off + (eaddr - a))) {
4091 					SEGVN_VMSTAT_FLTVNPAGES(22);
4092 					err = FC_MAKE_ERR(ierr);
4093 					goto out;
4094 				}
4095 				if (brkcow || tron || type == F_SOFTLOCK) {
4096 					/* can't reduce map area */
4097 					SEGVN_VMSTAT_FLTVNPAGES(23);
4098 					vop_size_err = 1;
4099 					goto out;
4100 				}
4101 				SEGVN_VMSTAT_FLTVNPAGES(24);
4102 				ASSERT(szc != 0);
4103 				pszc = 0;
4104 				ierr = -1;
4105 				break;
4106 			}
4107 
4108 			if (amp != NULL) {
4109 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4110 				anon_array_enter(amp, aindx, &an_cookie);
4111 			}
4112 			if (amp != NULL &&
4113 			    anon_get_ptr(amp->ahp, aindx) != NULL) {
4114 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4115 
4116 				SEGVN_VMSTAT_FLTVNPAGES(25);
4117 				ASSERT(anon_pages(amp->ahp, taindx,
4118 				    maxpages) == maxpages);
4119 				for (i = 0; i < pages; i++) {
4120 					page_unlock(ppa[i]);
4121 				}
4122 				anon_array_exit(&an_cookie);
4123 				ANON_LOCK_EXIT(&amp->a_rwlock);
4124 				if (pplist != NULL) {
4125 					page_free_replacement_page(pplist);
4126 					page_create_putback(pages);
4127 				}
4128 				SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4129 				if (szc < seg->s_szc) {
4130 					SEGVN_VMSTAT_FLTVNPAGES(26);
4131 					/*
4132 					 * For private segments SOFTLOCK
4133 					 * either always breaks cow (any rw
4134 					 * type except S_READ_NOCOW) or
4135 					 * address space is locked as writer
4136 					 * (S_READ_NOCOW case) and anon slots
4137 					 * can't show up on second check.
4138 					 * Therefore if we are here for
4139 					 * SOFTLOCK case it must be a cow
4140 					 * break but cow break never reduces
4141 					 * szc. text replication (tron) in
4142 					 * this case works as cow break.
4143 					 * Thus the assert below.
4144 					 */
4145 					ASSERT(!brkcow && !tron &&
4146 					    type != F_SOFTLOCK);
4147 					pszc = seg->s_szc;
4148 					ierr = -2;
4149 					break;
4150 				}
4151 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4152 				goto again;
4153 			}
4154 #ifdef DEBUG
4155 			if (amp != NULL) {
4156 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4157 				ASSERT(!anon_pages(amp->ahp, taindx, maxpages));
4158 			}
4159 #endif /* DEBUG */
4160 
4161 			if (brkcow || tron) {
4162 				ASSERT(amp != NULL);
4163 				ASSERT(pplist == NULL);
4164 				ASSERT(szc == seg->s_szc);
4165 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4166 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
4167 				SEGVN_VMSTAT_FLTVNPAGES(27);
4168 				ierr = anon_map_privatepages(amp, aindx, szc,
4169 				    seg, a, prot, ppa, vpage, segvn_anypgsz,
4170 				    tron ? PG_LOCAL : 0, svd->cred);
4171 				if (ierr != 0) {
4172 					SEGVN_VMSTAT_FLTVNPAGES(28);
4173 					anon_array_exit(&an_cookie);
4174 					ANON_LOCK_EXIT(&amp->a_rwlock);
4175 					SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4176 					err = FC_MAKE_ERR(ierr);
4177 					goto out;
4178 				}
4179 
4180 				ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4181 				/*
4182 				 * p_szc can't be changed for locked
4183 				 * swapfs pages.
4184 				 */
4185 				ASSERT(svd->rcookie ==
4186 				    HAT_INVALID_REGION_COOKIE);
4187 				hat_memload_array(hat, a, pgsz, ppa, prot,
4188 				    hat_flag);
4189 
4190 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4191 					SEGVN_VMSTAT_FLTVNPAGES(29);
4192 					for (i = 0; i < pages; i++) {
4193 						page_unlock(ppa[i]);
4194 					}
4195 				}
4196 				anon_array_exit(&an_cookie);
4197 				ANON_LOCK_EXIT(&amp->a_rwlock);
4198 				goto next;
4199 			}
4200 
4201 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
4202 			    (!svd->pageprot && svd->prot == (prot & vpprot)));
4203 
4204 			pfn = page_pptonum(ppa[0]);
4205 			/*
4206 			 * hat_page_demote() needs an SE_EXCL lock on one of
4207 			 * constituent page_t's and it decreases root's p_szc
4208 			 * last. This means if root's p_szc is equal szc and
4209 			 * all its constituent pages are locked
4210 			 * hat_page_demote() that could have changed p_szc to
4211 			 * szc is already done and no new have page_demote()
4212 			 * can start for this large page.
4213 			 */
4214 
4215 			/*
4216 			 * we need to make sure same mapping size is used for
4217 			 * the same address range if there's a possibility the
4218 			 * adddress is already mapped because hat layer panics
4219 			 * when translation is loaded for the range already
4220 			 * mapped with a different page size.  We achieve it
4221 			 * by always using largest page size possible subject
4222 			 * to the constraints of page size, segment page size
4223 			 * and page alignment.  Since mappings are invalidated
4224 			 * when those constraints change and make it
4225 			 * impossible to use previously used mapping size no
4226 			 * mapping size conflicts should happen.
4227 			 */
4228 
4229 		chkszc:
4230 			if ((pszc = ppa[0]->p_szc) == szc &&
4231 			    IS_P2ALIGNED(pfn, pages)) {
4232 
4233 				SEGVN_VMSTAT_FLTVNPAGES(30);
4234 #ifdef DEBUG
4235 				for (i = 0; i < pages; i++) {
4236 					ASSERT(PAGE_LOCKED(ppa[i]));
4237 					ASSERT(!PP_ISFREE(ppa[i]));
4238 					ASSERT(page_pptonum(ppa[i]) ==
4239 					    pfn + i);
4240 					ASSERT(ppa[i]->p_szc == szc);
4241 					ASSERT(ppa[i]->p_vnode == vp);
4242 					ASSERT(ppa[i]->p_offset ==
4243 					    off + (i << PAGESHIFT));
4244 				}
4245 #endif /* DEBUG */
4246 				/*
4247 				 * All pages are of szc we need and they are
4248 				 * all locked so they can't change szc. load
4249 				 * translations.
4250 				 *
4251 				 * if page got promoted since last check
4252 				 * we don't need pplist.
4253 				 */
4254 				if (pplist != NULL) {
4255 					page_free_replacement_page(pplist);
4256 					page_create_putback(pages);
4257 				}
4258 				if (PP_ISMIGRATE(ppa[0])) {
4259 					page_migrate(seg, a, ppa, pages);
4260 				}
4261 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4262 				    prot, vpprot);
4263 				hat_memload_array_region(hat, a, pgsz,
4264 				    ppa, prot & vpprot, hat_flag,
4265 				    svd->rcookie);
4266 
4267 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4268 					for (i = 0; i < pages; i++) {
4269 						page_unlock(ppa[i]);
4270 					}
4271 				}
4272 				if (amp != NULL) {
4273 					anon_array_exit(&an_cookie);
4274 					ANON_LOCK_EXIT(&amp->a_rwlock);
4275 				}
4276 				goto next;
4277 			}
4278 
4279 			/*
4280 			 * See if upsize is possible.
4281 			 */
4282 			if (pszc > szc && szc < seg->s_szc &&
4283 			    (segvn_anypgsz_vnode || pszc >= seg->s_szc)) {
4284 				pgcnt_t aphase;
4285 				uint_t pszc1 = MIN(pszc, seg->s_szc);
4286 				ppgsz = page_get_pagesize(pszc1);
4287 				ppages = btop(ppgsz);
4288 				aphase = btop(P2PHASE((uintptr_t)a, ppgsz));
4289 
4290 				ASSERT(type != F_SOFTLOCK);
4291 
4292 				SEGVN_VMSTAT_FLTVNPAGES(31);
4293 				if (aphase != P2PHASE(pfn, ppages)) {
4294 					segvn_faultvnmpss_align_err4++;
4295 				} else {
4296 					SEGVN_VMSTAT_FLTVNPAGES(32);
4297 					if (pplist != NULL) {
4298 						page_t *pl = pplist;
4299 						page_free_replacement_page(pl);
4300 						page_create_putback(pages);
4301 					}
4302 					for (i = 0; i < pages; i++) {
4303 						page_unlock(ppa[i]);
4304 					}
4305 					if (amp != NULL) {
4306 						anon_array_exit(&an_cookie);
4307 						ANON_LOCK_EXIT(&amp->a_rwlock);
4308 					}
4309 					pszc = pszc1;
4310 					ierr = -2;
4311 					break;
4312 				}
4313 			}
4314 
4315 			/*
4316 			 * check if we should use smallest mapping size.
4317 			 */
4318 			upgrdfail = 0;
4319 			if (szc == 0 ||
4320 			    (pszc >= szc &&
4321 			    !IS_P2ALIGNED(pfn, pages)) ||
4322 			    (pszc < szc &&
4323 			    !segvn_full_szcpages(ppa, szc, &upgrdfail,
4324 			    &pszc))) {
4325 
4326 				if (upgrdfail && type != F_SOFTLOCK) {
4327 					/*
4328 					 * segvn_full_szcpages failed to lock
4329 					 * all pages EXCL. Size down.
4330 					 */
4331 					ASSERT(pszc < szc);
4332 
4333 					SEGVN_VMSTAT_FLTVNPAGES(33);
4334 
4335 					if (pplist != NULL) {
4336 						page_t *pl = pplist;
4337 						page_free_replacement_page(pl);
4338 						page_create_putback(pages);
4339 					}
4340 
4341 					for (i = 0; i < pages; i++) {
4342 						page_unlock(ppa[i]);
4343 					}
4344 					if (amp != NULL) {
4345 						anon_array_exit(&an_cookie);
4346 						ANON_LOCK_EXIT(&amp->a_rwlock);
4347 					}
4348 					ierr = -1;
4349 					break;
4350 				}
4351 				if (szc != 0 && !upgrdfail) {
4352 					segvn_faultvnmpss_align_err5++;
4353 				}
4354 				SEGVN_VMSTAT_FLTVNPAGES(34);
4355 				if (pplist != NULL) {
4356 					page_free_replacement_page(pplist);
4357 					page_create_putback(pages);
4358 				}
4359 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4360 				    prot, vpprot);
4361 				if (upgrdfail && segvn_anypgsz_vnode) {
4362 					/* SOFTLOCK case */
4363 					hat_memload_array_region(hat, a, pgsz,
4364 					    ppa, prot & vpprot, hat_flag,
4365 					    svd->rcookie);
4366 				} else {
4367 					for (i = 0; i < pages; i++) {
4368 						hat_memload_region(hat,
4369 						    a + (i << PAGESHIFT),
4370 						    ppa[i], prot & vpprot,
4371 						    hat_flag, svd->rcookie);
4372 					}
4373 				}
4374 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4375 					for (i = 0; i < pages; i++) {
4376 						page_unlock(ppa[i]);
4377 					}
4378 				}
4379 				if (amp != NULL) {
4380 					anon_array_exit(&an_cookie);
4381 					ANON_LOCK_EXIT(&amp->a_rwlock);
4382 				}
4383 				goto next;
4384 			}
4385 
4386 			if (pszc == szc) {
4387 				/*
4388 				 * segvn_full_szcpages() upgraded pages szc.
4389 				 */
4390 				ASSERT(pszc == ppa[0]->p_szc);
4391 				ASSERT(IS_P2ALIGNED(pfn, pages));
4392 				goto chkszc;
4393 			}
4394 
4395 			if (pszc > szc) {
4396 				kmutex_t *szcmtx;
4397 				SEGVN_VMSTAT_FLTVNPAGES(35);
4398 				/*
4399 				 * p_szc of ppa[0] can change since we haven't
4400 				 * locked all constituent pages. Call
4401 				 * page_lock_szc() to prevent szc changes.
4402 				 * This should be a rare case that happens when
4403 				 * multiple segments use a different page size
4404 				 * to map the same file offsets.
4405 				 */
4406 				szcmtx = page_szc_lock(ppa[0]);
4407 				pszc = ppa[0]->p_szc;
4408 				ASSERT(szcmtx != NULL || pszc == 0);
4409 				ASSERT(ppa[0]->p_szc <= pszc);
4410 				if (pszc <= szc) {
4411 					SEGVN_VMSTAT_FLTVNPAGES(36);
4412 					if (szcmtx != NULL) {
4413 						mutex_exit(szcmtx);
4414 					}
4415 					goto chkszc;
4416 				}
4417 				if (pplist != NULL) {
4418 					/*
4419 					 * page got promoted since last check.
4420 					 * we don't need preaalocated large
4421 					 * page.
4422 					 */
4423 					SEGVN_VMSTAT_FLTVNPAGES(37);
4424 					page_free_replacement_page(pplist);
4425 					page_create_putback(pages);
4426 				}
4427 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4428 				    prot, vpprot);
4429 				hat_memload_array_region(hat, a, pgsz, ppa,
4430 				    prot & vpprot, hat_flag, svd->rcookie);
4431 				mutex_exit(szcmtx);
4432 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4433 					for (i = 0; i < pages; i++) {
4434 						page_unlock(ppa[i]);
4435 					}
4436 				}
4437 				if (amp != NULL) {
4438 					anon_array_exit(&an_cookie);
4439 					ANON_LOCK_EXIT(&amp->a_rwlock);
4440 				}
4441 				goto next;
4442 			}
4443 
4444 			/*
4445 			 * if page got demoted since last check
4446 			 * we could have not allocated larger page.
4447 			 * allocate now.
4448 			 */
4449 			if (pplist == NULL &&
4450 			    page_alloc_pages(vp, seg, a, &pplist, NULL,
4451 			    szc, 0, 0) && type != F_SOFTLOCK) {
4452 				SEGVN_VMSTAT_FLTVNPAGES(38);
4453 				for (i = 0; i < pages; i++) {
4454 					page_unlock(ppa[i]);
4455 				}
4456 				if (amp != NULL) {
4457 					anon_array_exit(&an_cookie);
4458 					ANON_LOCK_EXIT(&amp->a_rwlock);
4459 				}
4460 				ierr = -1;
4461 				alloc_failed |= (1 << szc);
4462 				break;
4463 			}
4464 
4465 			SEGVN_VMSTAT_FLTVNPAGES(39);
4466 
4467 			if (pplist != NULL) {
4468 				segvn_relocate_pages(ppa, pplist);
4469 #ifdef DEBUG
4470 			} else {
4471 				ASSERT(type == F_SOFTLOCK);
4472 				SEGVN_VMSTAT_FLTVNPAGES(40);
4473 #endif /* DEBUG */
4474 			}
4475 
4476 			SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot);
4477 
4478 			if (pplist == NULL && segvn_anypgsz_vnode == 0) {
4479 				ASSERT(type == F_SOFTLOCK);
4480 				for (i = 0; i < pages; i++) {
4481 					ASSERT(ppa[i]->p_szc < szc);
4482 					hat_memload_region(hat,
4483 					    a + (i << PAGESHIFT),
4484 					    ppa[i], prot & vpprot, hat_flag,
4485 					    svd->rcookie);
4486 				}
4487 			} else {
4488 				ASSERT(pplist != NULL || type == F_SOFTLOCK);
4489 				hat_memload_array_region(hat, a, pgsz, ppa,
4490 				    prot & vpprot, hat_flag, svd->rcookie);
4491 			}
4492 			if (!(hat_flag & HAT_LOAD_LOCK)) {
4493 				for (i = 0; i < pages; i++) {
4494 					ASSERT(PAGE_SHARED(ppa[i]));
4495 					page_unlock(ppa[i]);
4496 				}
4497 			}
4498 			if (amp != NULL) {
4499 				anon_array_exit(&an_cookie);
4500 				ANON_LOCK_EXIT(&amp->a_rwlock);
4501 			}
4502 
4503 		next:
4504 			if (vpage != NULL) {
4505 				vpage += pages;
4506 			}
4507 			adjszc_chk = 1;
4508 		}
4509 		if (a == lpgeaddr)
4510 			break;
4511 		ASSERT(a < lpgeaddr);
4512 
4513 		ASSERT(!brkcow && !tron && type != F_SOFTLOCK);
4514 
4515 		/*
4516 		 * ierr == -1 means we failed to map with a large page.
4517 		 * (either due to allocation/relocation failures or
4518 		 * misalignment with other mappings to this file.
4519 		 *
4520 		 * ierr == -2 means some other thread allocated a large page
4521 		 * after we gave up tp map with a large page.  retry with
4522 		 * larger mapping.
4523 		 */
4524 		ASSERT(ierr == -1 || ierr == -2);
4525 		ASSERT(ierr == -2 || szc != 0);
4526 		ASSERT(ierr == -1 || szc < seg->s_szc);
4527 		if (ierr == -2) {
4528 			SEGVN_VMSTAT_FLTVNPAGES(41);
4529 			ASSERT(pszc > szc && pszc <= seg->s_szc);
4530 			szc = pszc;
4531 		} else if (segvn_anypgsz_vnode) {
4532 			SEGVN_VMSTAT_FLTVNPAGES(42);
4533 			szc--;
4534 		} else {
4535 			SEGVN_VMSTAT_FLTVNPAGES(43);
4536 			ASSERT(pszc < szc);
4537 			/*
4538 			 * other process created pszc large page.
4539 			 * but we still have to drop to 0 szc.
4540 			 */
4541 			szc = 0;
4542 		}
4543 
4544 		pgsz = page_get_pagesize(szc);
4545 		pages = btop(pgsz);
4546 		if (ierr == -2) {
4547 			/*
4548 			 * Size up case. Note lpgaddr may only be needed for
4549 			 * softlock case so we don't adjust it here.
4550 			 */
4551 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4552 			ASSERT(a >= lpgaddr);
4553 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4554 			off = svd->offset + (uintptr_t)(a - seg->s_base);
4555 			aindx = svd->anon_index + seg_page(seg, a);
4556 			vpage = (svd->vpage != NULL) ?
4557 			    &svd->vpage[seg_page(seg, a)] : NULL;
4558 		} else {
4559 			/*
4560 			 * Size down case. Note lpgaddr may only be needed for
4561 			 * softlock case so we don't adjust it here.
4562 			 */
4563 			ASSERT(IS_P2ALIGNED(a, pgsz));
4564 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4565 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4566 			ASSERT(a < lpgeaddr);
4567 			if (a < addr) {
4568 				SEGVN_VMSTAT_FLTVNPAGES(44);
4569 				/*
4570 				 * The beginning of the large page region can
4571 				 * be pulled to the right to make a smaller
4572 				 * region. We haven't yet faulted a single
4573 				 * page.
4574 				 */
4575 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4576 				ASSERT(a >= lpgaddr);
4577 				off = svd->offset +
4578 				    (uintptr_t)(a - seg->s_base);
4579 				aindx = svd->anon_index + seg_page(seg, a);
4580 				vpage = (svd->vpage != NULL) ?
4581 				    &svd->vpage[seg_page(seg, a)] : NULL;
4582 			}
4583 		}
4584 	}
4585 out:
4586 	kmem_free(ppa, ppasize);
4587 	if (!err && !vop_size_err) {
4588 		SEGVN_VMSTAT_FLTVNPAGES(45);
4589 		return (0);
4590 	}
4591 	if (type == F_SOFTLOCK && a > lpgaddr) {
4592 		SEGVN_VMSTAT_FLTVNPAGES(46);
4593 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4594 	}
4595 	if (!vop_size_err) {
4596 		SEGVN_VMSTAT_FLTVNPAGES(47);
4597 		return (err);
4598 	}
4599 	ASSERT(brkcow || tron || type == F_SOFTLOCK);
4600 	/*
4601 	 * Large page end is mapped beyond the end of file and it's a cow
4602 	 * fault (can be a text replication induced cow) or softlock so we can't
4603 	 * reduce the map area.  For now just demote the segment. This should
4604 	 * really only happen if the end of the file changed after the mapping
4605 	 * was established since when large page segments are created we make
4606 	 * sure they don't extend beyond the end of the file.
4607 	 */
4608 	SEGVN_VMSTAT_FLTVNPAGES(48);
4609 
4610 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4611 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4612 	err = 0;
4613 	if (seg->s_szc != 0) {
4614 		segvn_fltvnpages_clrszc_cnt++;
4615 		ASSERT(svd->softlockcnt == 0);
4616 		err = segvn_clrszc(seg);
4617 		if (err != 0) {
4618 			segvn_fltvnpages_clrszc_err++;
4619 		}
4620 	}
4621 	ASSERT(err || seg->s_szc == 0);
4622 	SEGVN_LOCK_DOWNGRADE(seg->s_as, &svd->lock);
4623 	/* segvn_fault will do its job as if szc had been zero to begin with */
4624 	return (err == 0 ? IE_RETRY : FC_MAKE_ERR(err));
4625 }
4626 
4627 /*
4628  * This routine will attempt to fault in one large page.
4629  * it will use smaller pages if that fails.
4630  * It should only be called for pure anonymous segments.
4631  */
4632 static faultcode_t
4633 segvn_fault_anonpages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
4634     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
4635     caddr_t eaddr, int brkcow)
4636 {
4637 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4638 	struct anon_map *amp = svd->amp;
4639 	uchar_t segtype = svd->type;
4640 	uint_t szc = seg->s_szc;
4641 	size_t pgsz = page_get_pagesize(szc);
4642 	size_t maxpgsz = pgsz;
4643 	pgcnt_t pages = btop(pgsz);
4644 	uint_t ppaszc = szc;
4645 	caddr_t a = lpgaddr;
4646 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
4647 	struct vpage *vpage = (svd->vpage != NULL) ?
4648 	    &svd->vpage[seg_page(seg, a)] : NULL;
4649 	page_t **ppa;
4650 	uint_t	ppa_szc;
4651 	faultcode_t err;
4652 	int ierr;
4653 	uint_t protchk, prot, vpprot;
4654 	ulong_t i;
4655 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
4656 	anon_sync_obj_t cookie;
4657 	int adjszc_chk;
4658 	int pgflags = (svd->tr_state == SEGVN_TR_ON) ? PG_LOCAL : 0;
4659 
4660 	ASSERT(szc != 0);
4661 	ASSERT(amp != NULL);
4662 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
4663 	ASSERT(!(svd->flags & MAP_NORESERVE));
4664 	ASSERT(type != F_SOFTUNLOCK);
4665 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
4666 	ASSERT(!brkcow || svd->tr_state == SEGVN_TR_OFF);
4667 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
4668 
4669 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
4670 
4671 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltanpages[0]);
4672 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltanpages[1]);
4673 
4674 	if (svd->flags & MAP_TEXT) {
4675 		hat_flag |= HAT_LOAD_TEXT;
4676 	}
4677 
4678 	if (svd->pageprot) {
4679 		switch (rw) {
4680 		case S_READ:
4681 			protchk = PROT_READ;
4682 			break;
4683 		case S_WRITE:
4684 			protchk = PROT_WRITE;
4685 			break;
4686 		case S_EXEC:
4687 			protchk = PROT_EXEC;
4688 			break;
4689 		case S_OTHER:
4690 		default:
4691 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4692 			break;
4693 		}
4694 		VM_STAT_ADD(segvnvmstats.fltanpages[2]);
4695 	} else {
4696 		prot = svd->prot;
4697 		/* caller has already done segment level protection check. */
4698 	}
4699 
4700 	ppa = kmem_cache_alloc(segvn_szc_cache[ppaszc], KM_SLEEP);
4701 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4702 	for (;;) {
4703 		adjszc_chk = 0;
4704 		for (; a < lpgeaddr; a += pgsz, aindx += pages) {
4705 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
4706 				VM_STAT_ADD(segvnvmstats.fltanpages[3]);
4707 				ASSERT(vpage != NULL);
4708 				prot = VPP_PROT(vpage);
4709 				ASSERT(sameprot(seg, a, maxpgsz));
4710 				if ((prot & protchk) == 0) {
4711 					err = FC_PROT;
4712 					goto error;
4713 				}
4714 			}
4715 			if (adjszc_chk && IS_P2ALIGNED(a, maxpgsz) &&
4716 			    pgsz < maxpgsz) {
4717 				ASSERT(a > lpgaddr);
4718 				szc = seg->s_szc;
4719 				pgsz = maxpgsz;
4720 				pages = btop(pgsz);
4721 				ASSERT(IS_P2ALIGNED(aindx, pages));
4722 				lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr,
4723 				    pgsz);
4724 			}
4725 			if (type == F_SOFTLOCK) {
4726 				atomic_add_long((ulong_t *)&svd->softlockcnt,
4727 				    pages);
4728 			}
4729 			anon_array_enter(amp, aindx, &cookie);
4730 			ppa_szc = (uint_t)-1;
4731 			ierr = anon_map_getpages(amp, aindx, szc, seg, a,
4732 			    prot, &vpprot, ppa, &ppa_szc, vpage, rw, brkcow,
4733 			    segvn_anypgsz, pgflags, svd->cred);
4734 			if (ierr != 0) {
4735 				anon_array_exit(&cookie);
4736 				VM_STAT_ADD(segvnvmstats.fltanpages[4]);
4737 				if (type == F_SOFTLOCK) {
4738 					atomic_add_long(
4739 					    (ulong_t *)&svd->softlockcnt,
4740 					    -pages);
4741 				}
4742 				if (ierr > 0) {
4743 					VM_STAT_ADD(segvnvmstats.fltanpages[6]);
4744 					err = FC_MAKE_ERR(ierr);
4745 					goto error;
4746 				}
4747 				break;
4748 			}
4749 
4750 			ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4751 
4752 			ASSERT(segtype == MAP_SHARED ||
4753 			    ppa[0]->p_szc <= szc);
4754 			ASSERT(segtype == MAP_PRIVATE ||
4755 			    ppa[0]->p_szc >= szc);
4756 
4757 			/*
4758 			 * Handle pages that have been marked for migration
4759 			 */
4760 			if (lgrp_optimizations())
4761 				page_migrate(seg, a, ppa, pages);
4762 
4763 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
4764 
4765 			if (segtype == MAP_SHARED) {
4766 				vpprot |= PROT_WRITE;
4767 			}
4768 
4769 			hat_memload_array(hat, a, pgsz, ppa,
4770 			    prot & vpprot, hat_flag);
4771 
4772 			if (hat_flag & HAT_LOAD_LOCK) {
4773 				VM_STAT_ADD(segvnvmstats.fltanpages[7]);
4774 			} else {
4775 				VM_STAT_ADD(segvnvmstats.fltanpages[8]);
4776 				for (i = 0; i < pages; i++)
4777 					page_unlock(ppa[i]);
4778 			}
4779 			if (vpage != NULL)
4780 				vpage += pages;
4781 
4782 			anon_array_exit(&cookie);
4783 			adjszc_chk = 1;
4784 		}
4785 		if (a == lpgeaddr)
4786 			break;
4787 		ASSERT(a < lpgeaddr);
4788 		/*
4789 		 * ierr == -1 means we failed to allocate a large page.
4790 		 * so do a size down operation.
4791 		 *
4792 		 * ierr == -2 means some other process that privately shares
4793 		 * pages with this process has allocated a larger page and we
4794 		 * need to retry with larger pages. So do a size up
4795 		 * operation. This relies on the fact that large pages are
4796 		 * never partially shared i.e. if we share any constituent
4797 		 * page of a large page with another process we must share the
4798 		 * entire large page. Note this cannot happen for SOFTLOCK
4799 		 * case, unless current address (a) is at the beginning of the
4800 		 * next page size boundary because the other process couldn't
4801 		 * have relocated locked pages.
4802 		 */
4803 		ASSERT(ierr == -1 || ierr == -2);
4804 
4805 		if (segvn_anypgsz) {
4806 			ASSERT(ierr == -2 || szc != 0);
4807 			ASSERT(ierr == -1 || szc < seg->s_szc);
4808 			szc = (ierr == -1) ? szc - 1 : szc + 1;
4809 		} else {
4810 			/*
4811 			 * For non COW faults and segvn_anypgsz == 0
4812 			 * we need to be careful not to loop forever
4813 			 * if existing page is found with szc other
4814 			 * than 0 or seg->s_szc. This could be due
4815 			 * to page relocations on behalf of DR or
4816 			 * more likely large page creation. For this
4817 			 * case simply re-size to existing page's szc
4818 			 * if returned by anon_map_getpages().
4819 			 */
4820 			if (ppa_szc == (uint_t)-1) {
4821 				szc = (ierr == -1) ? 0 : seg->s_szc;
4822 			} else {
4823 				ASSERT(ppa_szc <= seg->s_szc);
4824 				ASSERT(ierr == -2 || ppa_szc < szc);
4825 				ASSERT(ierr == -1 || ppa_szc > szc);
4826 				szc = ppa_szc;
4827 			}
4828 		}
4829 
4830 		pgsz = page_get_pagesize(szc);
4831 		pages = btop(pgsz);
4832 		ASSERT(type != F_SOFTLOCK || ierr == -1 ||
4833 		    (IS_P2ALIGNED(a, pgsz) && IS_P2ALIGNED(lpgeaddr, pgsz)));
4834 		if (type == F_SOFTLOCK) {
4835 			/*
4836 			 * For softlocks we cannot reduce the fault area
4837 			 * (calculated based on the largest page size for this
4838 			 * segment) for size down and a is already next
4839 			 * page size aligned as assertted above for size
4840 			 * ups. Therefore just continue in case of softlock.
4841 			 */
4842 			VM_STAT_ADD(segvnvmstats.fltanpages[9]);
4843 			continue; /* keep lint happy */
4844 		} else if (ierr == -2) {
4845 
4846 			/*
4847 			 * Size up case. Note lpgaddr may only be needed for
4848 			 * softlock case so we don't adjust it here.
4849 			 */
4850 			VM_STAT_ADD(segvnvmstats.fltanpages[10]);
4851 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4852 			ASSERT(a >= lpgaddr);
4853 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4854 			aindx = svd->anon_index + seg_page(seg, a);
4855 			vpage = (svd->vpage != NULL) ?
4856 			    &svd->vpage[seg_page(seg, a)] : NULL;
4857 		} else {
4858 			/*
4859 			 * Size down case. Note lpgaddr may only be needed for
4860 			 * softlock case so we don't adjust it here.
4861 			 */
4862 			VM_STAT_ADD(segvnvmstats.fltanpages[11]);
4863 			ASSERT(IS_P2ALIGNED(a, pgsz));
4864 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4865 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4866 			ASSERT(a < lpgeaddr);
4867 			if (a < addr) {
4868 				/*
4869 				 * The beginning of the large page region can
4870 				 * be pulled to the right to make a smaller
4871 				 * region. We haven't yet faulted a single
4872 				 * page.
4873 				 */
4874 				VM_STAT_ADD(segvnvmstats.fltanpages[12]);
4875 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4876 				ASSERT(a >= lpgaddr);
4877 				aindx = svd->anon_index + seg_page(seg, a);
4878 				vpage = (svd->vpage != NULL) ?
4879 				    &svd->vpage[seg_page(seg, a)] : NULL;
4880 			}
4881 		}
4882 	}
4883 	VM_STAT_ADD(segvnvmstats.fltanpages[13]);
4884 	ANON_LOCK_EXIT(&amp->a_rwlock);
4885 	kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4886 	return (0);
4887 error:
4888 	VM_STAT_ADD(segvnvmstats.fltanpages[14]);
4889 	ANON_LOCK_EXIT(&amp->a_rwlock);
4890 	kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4891 	if (type == F_SOFTLOCK && a > lpgaddr) {
4892 		VM_STAT_ADD(segvnvmstats.fltanpages[15]);
4893 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4894 	}
4895 	return (err);
4896 }
4897 
4898 int fltadvice = 1;	/* set to free behind pages for sequential access */
4899 
4900 /*
4901  * This routine is called via a machine specific fault handling routine.
4902  * It is also called by software routines wishing to lock or unlock
4903  * a range of addresses.
4904  *
4905  * Here is the basic algorithm:
4906  *	If unlocking
4907  *		Call segvn_softunlock
4908  *		Return
4909  *	endif
4910  *	Checking and set up work
4911  *	If we will need some non-anonymous pages
4912  *		Call VOP_GETPAGE over the range of non-anonymous pages
4913  *	endif
4914  *	Loop over all addresses requested
4915  *		Call segvn_faultpage passing in page list
4916  *		    to load up translations and handle anonymous pages
4917  *	endloop
4918  *	Load up translation to any additional pages in page list not
4919  *	    already handled that fit into this segment
4920  */
4921 static faultcode_t
4922 segvn_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
4923     enum fault_type type, enum seg_rw rw)
4924 {
4925 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4926 	page_t **plp, **ppp, *pp;
4927 	u_offset_t off;
4928 	caddr_t a;
4929 	struct vpage *vpage;
4930 	uint_t vpprot, prot;
4931 	int err;
4932 	page_t *pl[PVN_GETPAGE_NUM + 1];
4933 	size_t plsz, pl_alloc_sz;
4934 	size_t page;
4935 	ulong_t anon_index;
4936 	struct anon_map *amp;
4937 	int dogetpage = 0;
4938 	caddr_t	lpgaddr, lpgeaddr;
4939 	size_t pgsz;
4940 	anon_sync_obj_t cookie;
4941 	int brkcow = BREAK_COW_SHARE(rw, type, svd->type);
4942 
4943 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
4944 	ASSERT(svd->amp == NULL || svd->rcookie == HAT_INVALID_REGION_COOKIE);
4945 
4946 	/*
4947 	 * First handle the easy stuff
4948 	 */
4949 	if (type == F_SOFTUNLOCK) {
4950 		if (rw == S_READ_NOCOW) {
4951 			rw = S_READ;
4952 			ASSERT(AS_WRITE_HELD(seg->s_as));
4953 		}
4954 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4955 		pgsz = (seg->s_szc == 0) ? PAGESIZE :
4956 		    page_get_pagesize(seg->s_szc);
4957 		VM_STAT_COND_ADD(pgsz > PAGESIZE, segvnvmstats.fltanpages[16]);
4958 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
4959 		segvn_softunlock(seg, lpgaddr, lpgeaddr - lpgaddr, rw);
4960 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4961 		return (0);
4962 	}
4963 
4964 	ASSERT(svd->tr_state == SEGVN_TR_OFF ||
4965 	    !HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
4966 	if (brkcow == 0) {
4967 		if (svd->tr_state == SEGVN_TR_INIT) {
4968 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4969 			if (svd->tr_state == SEGVN_TR_INIT) {
4970 				ASSERT(svd->vp != NULL && svd->amp == NULL);
4971 				ASSERT(svd->flags & MAP_TEXT);
4972 				ASSERT(svd->type == MAP_PRIVATE);
4973 				segvn_textrepl(seg);
4974 				ASSERT(svd->tr_state != SEGVN_TR_INIT);
4975 				ASSERT(svd->tr_state != SEGVN_TR_ON ||
4976 				    svd->amp != NULL);
4977 			}
4978 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4979 		}
4980 	} else if (svd->tr_state != SEGVN_TR_OFF) {
4981 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4982 
4983 		if (rw == S_WRITE && svd->tr_state != SEGVN_TR_OFF) {
4984 			ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
4985 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4986 			return (FC_PROT);
4987 		}
4988 
4989 		if (svd->tr_state == SEGVN_TR_ON) {
4990 			ASSERT(svd->vp != NULL && svd->amp != NULL);
4991 			segvn_textunrepl(seg, 0);
4992 			ASSERT(svd->amp == NULL &&
4993 			    svd->tr_state == SEGVN_TR_OFF);
4994 		} else if (svd->tr_state != SEGVN_TR_OFF) {
4995 			svd->tr_state = SEGVN_TR_OFF;
4996 		}
4997 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
4998 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4999 	}
5000 
5001 top:
5002 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5003 
5004 	/*
5005 	 * If we have the same protections for the entire segment,
5006 	 * insure that the access being attempted is legitimate.
5007 	 */
5008 
5009 	if (svd->pageprot == 0) {
5010 		uint_t protchk;
5011 
5012 		switch (rw) {
5013 		case S_READ:
5014 		case S_READ_NOCOW:
5015 			protchk = PROT_READ;
5016 			break;
5017 		case S_WRITE:
5018 			protchk = PROT_WRITE;
5019 			break;
5020 		case S_EXEC:
5021 			protchk = PROT_EXEC;
5022 			break;
5023 		case S_OTHER:
5024 		default:
5025 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
5026 			break;
5027 		}
5028 
5029 		if ((svd->prot & protchk) == 0) {
5030 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5031 			return (FC_PROT);	/* illegal access type */
5032 		}
5033 	}
5034 
5035 	if (brkcow && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5036 		/* this must be SOFTLOCK S_READ fault */
5037 		ASSERT(svd->amp == NULL);
5038 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5039 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5040 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5041 		if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5042 			/*
5043 			 * this must be the first ever non S_READ_NOCOW
5044 			 * softlock for this segment.
5045 			 */
5046 			ASSERT(svd->softlockcnt == 0);
5047 			hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5048 			    HAT_REGION_TEXT);
5049 			svd->rcookie = HAT_INVALID_REGION_COOKIE;
5050 		}
5051 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5052 		goto top;
5053 	}
5054 
5055 	/*
5056 	 * We can't allow the long term use of softlocks for vmpss segments,
5057 	 * because in some file truncation cases we should be able to demote
5058 	 * the segment, which requires that there are no softlocks.  The
5059 	 * only case where it's ok to allow a SOFTLOCK fault against a vmpss
5060 	 * segment is S_READ_NOCOW, where the caller holds the address space
5061 	 * locked as writer and calls softunlock before dropping the as lock.
5062 	 * S_READ_NOCOW is used by /proc to read memory from another user.
5063 	 *
5064 	 * Another deadlock between SOFTLOCK and file truncation can happen
5065 	 * because segvn_fault_vnodepages() calls the FS one pagesize at
5066 	 * a time. A second VOP_GETPAGE() call by segvn_fault_vnodepages()
5067 	 * can cause a deadlock because the first set of page_t's remain
5068 	 * locked SE_SHARED.  To avoid this, we demote segments on a first
5069 	 * SOFTLOCK if they have a length greater than the segment's
5070 	 * page size.
5071 	 *
5072 	 * So for now, we only avoid demoting a segment on a SOFTLOCK when
5073 	 * the access type is S_READ_NOCOW and the fault length is less than
5074 	 * or equal to the segment's page size. While this is quite restrictive,
5075 	 * it should be the most common case of SOFTLOCK against a vmpss
5076 	 * segment.
5077 	 *
5078 	 * For S_READ_NOCOW, it's safe not to do a copy on write because the
5079 	 * caller makes sure no COW will be caused by another thread for a
5080 	 * softlocked page.
5081 	 */
5082 	if (type == F_SOFTLOCK && svd->vp != NULL && seg->s_szc != 0) {
5083 		int demote = 0;
5084 
5085 		if (rw != S_READ_NOCOW) {
5086 			demote = 1;
5087 		}
5088 		if (!demote && len > PAGESIZE) {
5089 			pgsz = page_get_pagesize(seg->s_szc);
5090 			CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr,
5091 			    lpgeaddr);
5092 			if (lpgeaddr - lpgaddr > pgsz) {
5093 				demote = 1;
5094 			}
5095 		}
5096 
5097 		ASSERT(demote || AS_WRITE_HELD(seg->s_as));
5098 
5099 		if (demote) {
5100 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5101 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5102 			if (seg->s_szc != 0) {
5103 				segvn_vmpss_clrszc_cnt++;
5104 				ASSERT(svd->softlockcnt == 0);
5105 				err = segvn_clrszc(seg);
5106 				if (err) {
5107 					segvn_vmpss_clrszc_err++;
5108 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5109 					return (FC_MAKE_ERR(err));
5110 				}
5111 			}
5112 			ASSERT(seg->s_szc == 0);
5113 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5114 			goto top;
5115 		}
5116 	}
5117 
5118 	/*
5119 	 * Check to see if we need to allocate an anon_map structure.
5120 	 */
5121 	if (svd->amp == NULL && (svd->vp == NULL || brkcow)) {
5122 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5123 		/*
5124 		 * Drop the "read" lock on the segment and acquire
5125 		 * the "write" version since we have to allocate the
5126 		 * anon_map.
5127 		 */
5128 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5129 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5130 
5131 		if (svd->amp == NULL) {
5132 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
5133 			svd->amp->a_szc = seg->s_szc;
5134 		}
5135 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5136 
5137 		/*
5138 		 * Start all over again since segment protections
5139 		 * may have changed after we dropped the "read" lock.
5140 		 */
5141 		goto top;
5142 	}
5143 
5144 	/*
5145 	 * S_READ_NOCOW vs S_READ distinction was
5146 	 * only needed for the code above. After
5147 	 * that we treat it as S_READ.
5148 	 */
5149 	if (rw == S_READ_NOCOW) {
5150 		ASSERT(type == F_SOFTLOCK);
5151 		ASSERT(AS_WRITE_HELD(seg->s_as));
5152 		rw = S_READ;
5153 	}
5154 
5155 	amp = svd->amp;
5156 
5157 	/*
5158 	 * MADV_SEQUENTIAL work is ignored for large page segments.
5159 	 */
5160 	if (seg->s_szc != 0) {
5161 		pgsz = page_get_pagesize(seg->s_szc);
5162 		ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
5163 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
5164 		if (svd->vp == NULL) {
5165 			err = segvn_fault_anonpages(hat, seg, lpgaddr,
5166 			    lpgeaddr, type, rw, addr, addr + len, brkcow);
5167 		} else {
5168 			err = segvn_fault_vnodepages(hat, seg, lpgaddr,
5169 			    lpgeaddr, type, rw, addr, addr + len, brkcow);
5170 			if (err == IE_RETRY) {
5171 				ASSERT(seg->s_szc == 0);
5172 				ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
5173 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5174 				goto top;
5175 			}
5176 		}
5177 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5178 		return (err);
5179 	}
5180 
5181 	page = seg_page(seg, addr);
5182 	if (amp != NULL) {
5183 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5184 		anon_index = svd->anon_index + page;
5185 
5186 		if (type == F_PROT && rw == S_READ &&
5187 		    svd->tr_state == SEGVN_TR_OFF &&
5188 		    svd->type == MAP_PRIVATE && svd->pageprot == 0) {
5189 			size_t index = anon_index;
5190 			struct anon *ap;
5191 
5192 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5193 			/*
5194 			 * The fast path could apply to S_WRITE also, except
5195 			 * that the protection fault could be caused by lazy
5196 			 * tlb flush when ro->rw. In this case, the pte is
5197 			 * RW already. But RO in the other cpu's tlb causes
5198 			 * the fault. Since hat_chgprot won't do anything if
5199 			 * pte doesn't change, we may end up faulting
5200 			 * indefinitely until the RO tlb entry gets replaced.
5201 			 */
5202 			for (a = addr; a < addr + len; a += PAGESIZE, index++) {
5203 				anon_array_enter(amp, index, &cookie);
5204 				ap = anon_get_ptr(amp->ahp, index);
5205 				anon_array_exit(&cookie);
5206 				if ((ap == NULL) || (ap->an_refcnt != 1)) {
5207 					ANON_LOCK_EXIT(&amp->a_rwlock);
5208 					goto slow;
5209 				}
5210 			}
5211 			hat_chgprot(seg->s_as->a_hat, addr, len, svd->prot);
5212 			ANON_LOCK_EXIT(&amp->a_rwlock);
5213 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5214 			return (0);
5215 		}
5216 	}
5217 slow:
5218 
5219 	if (svd->vpage == NULL)
5220 		vpage = NULL;
5221 	else
5222 		vpage = &svd->vpage[page];
5223 
5224 	off = svd->offset + (uintptr_t)(addr - seg->s_base);
5225 
5226 	/*
5227 	 * If MADV_SEQUENTIAL has been set for the particular page we
5228 	 * are faulting on, free behind all pages in the segment and put
5229 	 * them on the free list.
5230 	 */
5231 
5232 	if ((page != 0) && fltadvice && svd->tr_state != SEGVN_TR_ON) {
5233 		struct vpage *vpp;
5234 		ulong_t fanon_index;
5235 		size_t fpage;
5236 		u_offset_t pgoff, fpgoff;
5237 		struct vnode *fvp;
5238 		struct anon *fap = NULL;
5239 
5240 		if (svd->advice == MADV_SEQUENTIAL ||
5241 		    (svd->pageadvice &&
5242 		    VPP_ADVICE(vpage) == MADV_SEQUENTIAL)) {
5243 			pgoff = off - PAGESIZE;
5244 			fpage = page - 1;
5245 			if (vpage != NULL)
5246 				vpp = &svd->vpage[fpage];
5247 			if (amp != NULL)
5248 				fanon_index = svd->anon_index + fpage;
5249 
5250 			while (pgoff > svd->offset) {
5251 				if (svd->advice != MADV_SEQUENTIAL &&
5252 				    (!svd->pageadvice || (vpage &&
5253 				    VPP_ADVICE(vpp) != MADV_SEQUENTIAL)))
5254 					break;
5255 
5256 				/*
5257 				 * If this is an anon page, we must find the
5258 				 * correct <vp, offset> for it
5259 				 */
5260 				fap = NULL;
5261 				if (amp != NULL) {
5262 					ANON_LOCK_ENTER(&amp->a_rwlock,
5263 					    RW_READER);
5264 					anon_array_enter(amp, fanon_index,
5265 					    &cookie);
5266 					fap = anon_get_ptr(amp->ahp,
5267 					    fanon_index);
5268 					if (fap != NULL) {
5269 						swap_xlate(fap, &fvp, &fpgoff);
5270 					} else {
5271 						fpgoff = pgoff;
5272 						fvp = svd->vp;
5273 					}
5274 					anon_array_exit(&cookie);
5275 					ANON_LOCK_EXIT(&amp->a_rwlock);
5276 				} else {
5277 					fpgoff = pgoff;
5278 					fvp = svd->vp;
5279 				}
5280 				if (fvp == NULL)
5281 					break;	/* XXX */
5282 				/*
5283 				 * Skip pages that are free or have an
5284 				 * "exclusive" lock.
5285 				 */
5286 				pp = page_lookup_nowait(fvp, fpgoff, SE_SHARED);
5287 				if (pp == NULL)
5288 					break;
5289 				/*
5290 				 * We don't need the page_struct_lock to test
5291 				 * as this is only advisory; even if we
5292 				 * acquire it someone might race in and lock
5293 				 * the page after we unlock and before the
5294 				 * PUTPAGE, then VOP_PUTPAGE will do nothing.
5295 				 */
5296 				if (pp->p_lckcnt == 0 && pp->p_cowcnt == 0) {
5297 					/*
5298 					 * Hold the vnode before releasing
5299 					 * the page lock to prevent it from
5300 					 * being freed and re-used by some
5301 					 * other thread.
5302 					 */
5303 					VN_HOLD(fvp);
5304 					page_unlock(pp);
5305 					/*
5306 					 * We should build a page list
5307 					 * to kluster putpages XXX
5308 					 */
5309 					(void) VOP_PUTPAGE(fvp,
5310 					    (offset_t)fpgoff, PAGESIZE,
5311 					    (B_DONTNEED|B_FREE|B_ASYNC),
5312 					    svd->cred, NULL);
5313 					VN_RELE(fvp);
5314 				} else {
5315 					/*
5316 					 * XXX - Should the loop terminate if
5317 					 * the page is `locked'?
5318 					 */
5319 					page_unlock(pp);
5320 				}
5321 				--vpp;
5322 				--fanon_index;
5323 				pgoff -= PAGESIZE;
5324 			}
5325 		}
5326 	}
5327 
5328 	plp = pl;
5329 	*plp = NULL;
5330 	pl_alloc_sz = 0;
5331 
5332 	/*
5333 	 * See if we need to call VOP_GETPAGE for
5334 	 * *any* of the range being faulted on.
5335 	 * We can skip all of this work if there
5336 	 * was no original vnode.
5337 	 */
5338 	if (svd->vp != NULL) {
5339 		u_offset_t vp_off;
5340 		size_t vp_len;
5341 		struct anon *ap;
5342 		vnode_t *vp;
5343 
5344 		vp_off = off;
5345 		vp_len = len;
5346 
5347 		if (amp == NULL)
5348 			dogetpage = 1;
5349 		else {
5350 			/*
5351 			 * Only acquire reader lock to prevent amp->ahp
5352 			 * from being changed.  It's ok to miss pages,
5353 			 * hence we don't do anon_array_enter
5354 			 */
5355 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5356 			ap = anon_get_ptr(amp->ahp, anon_index);
5357 
5358 			if (len <= PAGESIZE)
5359 				/* inline non_anon() */
5360 				dogetpage = (ap == NULL);
5361 			else
5362 				dogetpage = non_anon(amp->ahp, anon_index,
5363 				    &vp_off, &vp_len);
5364 			ANON_LOCK_EXIT(&amp->a_rwlock);
5365 		}
5366 
5367 		if (dogetpage) {
5368 			enum seg_rw arw;
5369 			struct as *as = seg->s_as;
5370 
5371 			if (len > ptob((sizeof (pl) / sizeof (pl[0])) - 1)) {
5372 				/*
5373 				 * Page list won't fit in local array,
5374 				 * allocate one of the needed size.
5375 				 */
5376 				pl_alloc_sz =
5377 				    (btop(len) + 1) * sizeof (page_t *);
5378 				plp = kmem_alloc(pl_alloc_sz, KM_SLEEP);
5379 				plp[0] = NULL;
5380 				plsz = len;
5381 			} else if (rw == S_WRITE && svd->type == MAP_PRIVATE ||
5382 			    svd->tr_state == SEGVN_TR_ON || rw == S_OTHER ||
5383 			    (((size_t)(addr + PAGESIZE) <
5384 			    (size_t)(seg->s_base + seg->s_size)) &&
5385 			    hat_probe(as->a_hat, addr + PAGESIZE))) {
5386 				/*
5387 				 * Ask VOP_GETPAGE to return the exact number
5388 				 * of pages if
5389 				 * (a) this is a COW fault, or
5390 				 * (b) this is a software fault, or
5391 				 * (c) next page is already mapped.
5392 				 */
5393 				plsz = len;
5394 			} else {
5395 				/*
5396 				 * Ask VOP_GETPAGE to return adjacent pages
5397 				 * within the segment.
5398 				 */
5399 				plsz = MIN((size_t)PVN_GETPAGE_SZ, (size_t)
5400 				    ((seg->s_base + seg->s_size) - addr));
5401 				ASSERT((addr + plsz) <=
5402 				    (seg->s_base + seg->s_size));
5403 			}
5404 
5405 			/*
5406 			 * Need to get some non-anonymous pages.
5407 			 * We need to make only one call to GETPAGE to do
5408 			 * this to prevent certain deadlocking conditions
5409 			 * when we are doing locking.  In this case
5410 			 * non_anon() should have picked up the smallest
5411 			 * range which includes all the non-anonymous
5412 			 * pages in the requested range.  We have to
5413 			 * be careful regarding which rw flag to pass in
5414 			 * because on a private mapping, the underlying
5415 			 * object is never allowed to be written.
5416 			 */
5417 			if (rw == S_WRITE && svd->type == MAP_PRIVATE) {
5418 				arw = S_READ;
5419 			} else {
5420 				arw = rw;
5421 			}
5422 			vp = svd->vp;
5423 			TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5424 			    "segvn_getpage:seg %p addr %p vp %p",
5425 			    seg, addr, vp);
5426 			err = VOP_GETPAGE(vp, (offset_t)vp_off, vp_len,
5427 			    &vpprot, plp, plsz, seg, addr + (vp_off - off), arw,
5428 			    svd->cred, NULL);
5429 			if (err) {
5430 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5431 				segvn_pagelist_rele(plp);
5432 				if (pl_alloc_sz)
5433 					kmem_free(plp, pl_alloc_sz);
5434 				return (FC_MAKE_ERR(err));
5435 			}
5436 			if (svd->type == MAP_PRIVATE)
5437 				vpprot &= ~PROT_WRITE;
5438 		}
5439 	}
5440 
5441 	/*
5442 	 * N.B. at this time the plp array has all the needed non-anon
5443 	 * pages in addition to (possibly) having some adjacent pages.
5444 	 */
5445 
5446 	/*
5447 	 * Always acquire the anon_array_lock to prevent
5448 	 * 2 threads from allocating separate anon slots for
5449 	 * the same "addr".
5450 	 *
5451 	 * If this is a copy-on-write fault and we don't already
5452 	 * have the anon_array_lock, acquire it to prevent the
5453 	 * fault routine from handling multiple copy-on-write faults
5454 	 * on the same "addr" in the same address space.
5455 	 *
5456 	 * Only one thread should deal with the fault since after
5457 	 * it is handled, the other threads can acquire a translation
5458 	 * to the newly created private page.  This prevents two or
5459 	 * more threads from creating different private pages for the
5460 	 * same fault.
5461 	 *
5462 	 * We grab "serialization" lock here if this is a MAP_PRIVATE segment
5463 	 * to prevent deadlock between this thread and another thread
5464 	 * which has soft-locked this page and wants to acquire serial_lock.
5465 	 * ( bug 4026339 )
5466 	 *
5467 	 * The fix for bug 4026339 becomes unnecessary when using the
5468 	 * locking scheme with per amp rwlock and a global set of hash
5469 	 * lock, anon_array_lock.  If we steal a vnode page when low
5470 	 * on memory and upgrad the page lock through page_rename,
5471 	 * then the page is PAGE_HANDLED, nothing needs to be done
5472 	 * for this page after returning from segvn_faultpage.
5473 	 *
5474 	 * But really, the page lock should be downgraded after
5475 	 * the stolen page is page_rename'd.
5476 	 */
5477 
5478 	if (amp != NULL)
5479 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5480 
5481 	/*
5482 	 * Ok, now loop over the address range and handle faults
5483 	 */
5484 	for (a = addr; a < addr + len; a += PAGESIZE, off += PAGESIZE) {
5485 		err = segvn_faultpage(hat, seg, a, off, vpage, plp, vpprot,
5486 		    type, rw, brkcow);
5487 		if (err) {
5488 			if (amp != NULL)
5489 				ANON_LOCK_EXIT(&amp->a_rwlock);
5490 			if (type == F_SOFTLOCK && a > addr) {
5491 				segvn_softunlock(seg, addr, (a - addr),
5492 				    S_OTHER);
5493 			}
5494 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5495 			segvn_pagelist_rele(plp);
5496 			if (pl_alloc_sz)
5497 				kmem_free(plp, pl_alloc_sz);
5498 			return (err);
5499 		}
5500 		if (vpage) {
5501 			vpage++;
5502 		} else if (svd->vpage) {
5503 			page = seg_page(seg, addr);
5504 			vpage = &svd->vpage[++page];
5505 		}
5506 	}
5507 
5508 	/* Didn't get pages from the underlying fs so we're done */
5509 	if (!dogetpage)
5510 		goto done;
5511 
5512 	/*
5513 	 * Now handle any other pages in the list returned.
5514 	 * If the page can be used, load up the translations now.
5515 	 * Note that the for loop will only be entered if "plp"
5516 	 * is pointing to a non-NULL page pointer which means that
5517 	 * VOP_GETPAGE() was called and vpprot has been initialized.
5518 	 */
5519 	if (svd->pageprot == 0)
5520 		prot = svd->prot & vpprot;
5521 
5522 
5523 	/*
5524 	 * Large Files: diff should be unsigned value because we started
5525 	 * supporting > 2GB segment sizes from 2.5.1 and when a
5526 	 * large file of size > 2GB gets mapped to address space
5527 	 * the diff value can be > 2GB.
5528 	 */
5529 
5530 	for (ppp = plp; (pp = *ppp) != NULL; ppp++) {
5531 		size_t diff;
5532 		struct anon *ap;
5533 		int anon_index;
5534 		anon_sync_obj_t cookie;
5535 		int hat_flag = HAT_LOAD_ADV;
5536 
5537 		if (svd->flags & MAP_TEXT) {
5538 			hat_flag |= HAT_LOAD_TEXT;
5539 		}
5540 
5541 		if (pp == PAGE_HANDLED)
5542 			continue;
5543 
5544 		if (svd->tr_state != SEGVN_TR_ON &&
5545 		    pp->p_offset >=  svd->offset &&
5546 		    pp->p_offset < svd->offset + seg->s_size) {
5547 
5548 			diff = pp->p_offset - svd->offset;
5549 
5550 			/*
5551 			 * Large Files: Following is the assertion
5552 			 * validating the above cast.
5553 			 */
5554 			ASSERT(svd->vp == pp->p_vnode);
5555 
5556 			page = btop(diff);
5557 			if (svd->pageprot)
5558 				prot = VPP_PROT(&svd->vpage[page]) & vpprot;
5559 
5560 			/*
5561 			 * Prevent other threads in the address space from
5562 			 * creating private pages (i.e., allocating anon slots)
5563 			 * while we are in the process of loading translations
5564 			 * to additional pages returned by the underlying
5565 			 * object.
5566 			 */
5567 			if (amp != NULL) {
5568 				anon_index = svd->anon_index + page;
5569 				anon_array_enter(amp, anon_index, &cookie);
5570 				ap = anon_get_ptr(amp->ahp, anon_index);
5571 			}
5572 			if ((amp == NULL) || (ap == NULL)) {
5573 				if (IS_VMODSORT(pp->p_vnode) ||
5574 				    enable_mbit_wa) {
5575 					if (rw == S_WRITE)
5576 						hat_setmod(pp);
5577 					else if (rw != S_OTHER &&
5578 					    !hat_ismod(pp))
5579 						prot &= ~PROT_WRITE;
5580 				}
5581 				/*
5582 				 * Skip mapping read ahead pages marked
5583 				 * for migration, so they will get migrated
5584 				 * properly on fault
5585 				 */
5586 				ASSERT(amp == NULL ||
5587 				    svd->rcookie == HAT_INVALID_REGION_COOKIE);
5588 				if ((prot & PROT_READ) && !PP_ISMIGRATE(pp)) {
5589 					hat_memload_region(hat,
5590 					    seg->s_base + diff,
5591 					    pp, prot, hat_flag,
5592 					    svd->rcookie);
5593 				}
5594 			}
5595 			if (amp != NULL)
5596 				anon_array_exit(&cookie);
5597 		}
5598 		page_unlock(pp);
5599 	}
5600 done:
5601 	if (amp != NULL)
5602 		ANON_LOCK_EXIT(&amp->a_rwlock);
5603 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5604 	if (pl_alloc_sz)
5605 		kmem_free(plp, pl_alloc_sz);
5606 	return (0);
5607 }
5608 
5609 /*
5610  * This routine is used to start I/O on pages asynchronously.  XXX it will
5611  * only create PAGESIZE pages. At fault time they will be relocated into
5612  * larger pages.
5613  */
5614 static faultcode_t
5615 segvn_faulta(struct seg *seg, caddr_t addr)
5616 {
5617 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5618 	int err;
5619 	struct anon_map *amp;
5620 	vnode_t *vp;
5621 
5622 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
5623 
5624 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5625 	if ((amp = svd->amp) != NULL) {
5626 		struct anon *ap;
5627 
5628 		/*
5629 		 * Reader lock to prevent amp->ahp from being changed.
5630 		 * This is advisory, it's ok to miss a page, so
5631 		 * we don't do anon_array_enter lock.
5632 		 */
5633 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5634 		if ((ap = anon_get_ptr(amp->ahp,
5635 		    svd->anon_index + seg_page(seg, addr))) != NULL) {
5636 
5637 			err = anon_getpage(&ap, NULL, NULL,
5638 			    0, seg, addr, S_READ, svd->cred);
5639 
5640 			ANON_LOCK_EXIT(&amp->a_rwlock);
5641 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5642 			if (err)
5643 				return (FC_MAKE_ERR(err));
5644 			return (0);
5645 		}
5646 		ANON_LOCK_EXIT(&amp->a_rwlock);
5647 	}
5648 
5649 	if (svd->vp == NULL) {
5650 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5651 		return (0);			/* zfod page - do nothing now */
5652 	}
5653 
5654 	vp = svd->vp;
5655 	TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5656 	    "segvn_getpage:seg %p addr %p vp %p", seg, addr, vp);
5657 	err = VOP_GETPAGE(vp,
5658 	    (offset_t)(svd->offset + (uintptr_t)(addr - seg->s_base)),
5659 	    PAGESIZE, NULL, NULL, 0, seg, addr,
5660 	    S_OTHER, svd->cred, NULL);
5661 
5662 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5663 	if (err)
5664 		return (FC_MAKE_ERR(err));
5665 	return (0);
5666 }
5667 
5668 static int
5669 segvn_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
5670 {
5671 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5672 	struct vpage *cvp, *svp, *evp;
5673 	struct vnode *vp;
5674 	size_t pgsz;
5675 	pgcnt_t pgcnt;
5676 	anon_sync_obj_t cookie;
5677 	int unload_done = 0;
5678 
5679 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
5680 
5681 	if ((svd->maxprot & prot) != prot)
5682 		return (EACCES);			/* violated maxprot */
5683 
5684 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5685 
5686 	/* return if prot is the same */
5687 	if (!svd->pageprot && svd->prot == prot) {
5688 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5689 		return (0);
5690 	}
5691 
5692 	/*
5693 	 * Since we change protections we first have to flush the cache.
5694 	 * This makes sure all the pagelock calls have to recheck
5695 	 * protections.
5696 	 */
5697 	if (svd->softlockcnt > 0) {
5698 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5699 
5700 		/*
5701 		 * If this is shared segment non 0 softlockcnt
5702 		 * means locked pages are still in use.
5703 		 */
5704 		if (svd->type == MAP_SHARED) {
5705 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5706 			return (EAGAIN);
5707 		}
5708 
5709 		/*
5710 		 * Since we do have the segvn writers lock nobody can fill
5711 		 * the cache with entries belonging to this seg during
5712 		 * the purge. The flush either succeeds or we still have
5713 		 * pending I/Os.
5714 		 */
5715 		segvn_purge(seg);
5716 		if (svd->softlockcnt > 0) {
5717 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5718 			return (EAGAIN);
5719 		}
5720 	}
5721 
5722 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5723 		ASSERT(svd->amp == NULL);
5724 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5725 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5726 		    HAT_REGION_TEXT);
5727 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
5728 		unload_done = 1;
5729 	} else if (svd->tr_state == SEGVN_TR_INIT) {
5730 		svd->tr_state = SEGVN_TR_OFF;
5731 	} else if (svd->tr_state == SEGVN_TR_ON) {
5732 		ASSERT(svd->amp != NULL);
5733 		segvn_textunrepl(seg, 0);
5734 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
5735 		unload_done = 1;
5736 	}
5737 
5738 	if ((prot & PROT_WRITE) && svd->type == MAP_SHARED &&
5739 	    svd->vp != NULL && (svd->vp->v_flag & VVMEXEC)) {
5740 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
5741 		segvn_inval_trcache(svd->vp);
5742 	}
5743 	if (seg->s_szc != 0) {
5744 		int err;
5745 		pgsz = page_get_pagesize(seg->s_szc);
5746 		pgcnt = pgsz >> PAGESHIFT;
5747 		ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
5748 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
5749 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5750 			ASSERT(seg->s_base != addr || seg->s_size != len);
5751 			/*
5752 			 * If we are holding the as lock as a reader then
5753 			 * we need to return IE_RETRY and let the as
5754 			 * layer drop and re-acquire the lock as a writer.
5755 			 */
5756 			if (AS_READ_HELD(seg->s_as))
5757 				return (IE_RETRY);
5758 			VM_STAT_ADD(segvnvmstats.demoterange[1]);
5759 			if (svd->type == MAP_PRIVATE || svd->vp != NULL) {
5760 				err = segvn_demote_range(seg, addr, len,
5761 				    SDR_END, 0);
5762 			} else {
5763 				uint_t szcvec = map_pgszcvec(seg->s_base,
5764 				    pgsz, (uintptr_t)seg->s_base,
5765 				    (svd->flags & MAP_TEXT), MAPPGSZC_SHM, 0);
5766 				err = segvn_demote_range(seg, addr, len,
5767 				    SDR_END, szcvec);
5768 			}
5769 			if (err == 0)
5770 				return (IE_RETRY);
5771 			if (err == ENOMEM)
5772 				return (IE_NOMEM);
5773 			return (err);
5774 		}
5775 	}
5776 
5777 
5778 	/*
5779 	 * If it's a private mapping and we're making it writable then we
5780 	 * may have to reserve the additional swap space now. If we are
5781 	 * making writable only a part of the segment then we use its vpage
5782 	 * array to keep a record of the pages for which we have reserved
5783 	 * swap. In this case we set the pageswap field in the segment's
5784 	 * segvn structure to record this.
5785 	 *
5786 	 * If it's a private mapping to a file (i.e., vp != NULL) and we're
5787 	 * removing write permission on the entire segment and we haven't
5788 	 * modified any pages, we can release the swap space.
5789 	 */
5790 	if (svd->type == MAP_PRIVATE) {
5791 		if (prot & PROT_WRITE) {
5792 			if (!(svd->flags & MAP_NORESERVE) &&
5793 			    !(svd->swresv && svd->pageswap == 0)) {
5794 				size_t sz = 0;
5795 
5796 				/*
5797 				 * Start by determining how much swap
5798 				 * space is required.
5799 				 */
5800 				if (addr == seg->s_base &&
5801 				    len == seg->s_size &&
5802 				    svd->pageswap == 0) {
5803 					/* The whole segment */
5804 					sz = seg->s_size;
5805 				} else {
5806 					/*
5807 					 * Make sure that the vpage array
5808 					 * exists, and make a note of the
5809 					 * range of elements corresponding
5810 					 * to len.
5811 					 */
5812 					segvn_vpage(seg);
5813 					if (svd->vpage == NULL) {
5814 						SEGVN_LOCK_EXIT(seg->s_as,
5815 						    &svd->lock);
5816 						return (ENOMEM);
5817 					}
5818 					svp = &svd->vpage[seg_page(seg, addr)];
5819 					evp = &svd->vpage[seg_page(seg,
5820 					    addr + len)];
5821 
5822 					if (svd->pageswap == 0) {
5823 						/*
5824 						 * This is the first time we've
5825 						 * asked for a part of this
5826 						 * segment, so we need to
5827 						 * reserve everything we've
5828 						 * been asked for.
5829 						 */
5830 						sz = len;
5831 					} else {
5832 						/*
5833 						 * We have to count the number
5834 						 * of pages required.
5835 						 */
5836 						for (cvp = svp;  cvp < evp;
5837 						    cvp++) {
5838 							if (!VPP_ISSWAPRES(cvp))
5839 								sz++;
5840 						}
5841 						sz <<= PAGESHIFT;
5842 					}
5843 				}
5844 
5845 				/* Try to reserve the necessary swap. */
5846 				if (anon_resv_zone(sz,
5847 				    seg->s_as->a_proc->p_zone) == 0) {
5848 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5849 					return (IE_NOMEM);
5850 				}
5851 
5852 				/*
5853 				 * Make a note of how much swap space
5854 				 * we've reserved.
5855 				 */
5856 				if (svd->pageswap == 0 && sz == seg->s_size) {
5857 					svd->swresv = sz;
5858 				} else {
5859 					ASSERT(svd->vpage != NULL);
5860 					svd->swresv += sz;
5861 					svd->pageswap = 1;
5862 					for (cvp = svp; cvp < evp; cvp++) {
5863 						if (!VPP_ISSWAPRES(cvp))
5864 							VPP_SETSWAPRES(cvp);
5865 					}
5866 				}
5867 			}
5868 		} else {
5869 			/*
5870 			 * Swap space is released only if this segment
5871 			 * does not map anonymous memory, since read faults
5872 			 * on such segments still need an anon slot to read
5873 			 * in the data.
5874 			 */
5875 			if (svd->swresv != 0 && svd->vp != NULL &&
5876 			    svd->amp == NULL && addr == seg->s_base &&
5877 			    len == seg->s_size && svd->pageprot == 0) {
5878 				ASSERT(svd->pageswap == 0);
5879 				anon_unresv_zone(svd->swresv,
5880 				    seg->s_as->a_proc->p_zone);
5881 				svd->swresv = 0;
5882 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
5883 				    "anon proc:%p %lu %u", seg, 0, 0);
5884 			}
5885 		}
5886 	}
5887 
5888 	if (addr == seg->s_base && len == seg->s_size && svd->vpage == NULL) {
5889 		if (svd->prot == prot) {
5890 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5891 			return (0);			/* all done */
5892 		}
5893 		svd->prot = (uchar_t)prot;
5894 	} else if (svd->type == MAP_PRIVATE) {
5895 		struct anon *ap = NULL;
5896 		page_t *pp;
5897 		u_offset_t offset, off;
5898 		struct anon_map *amp;
5899 		ulong_t anon_idx = 0;
5900 
5901 		/*
5902 		 * A vpage structure exists or else the change does not
5903 		 * involve the entire segment.  Establish a vpage structure
5904 		 * if none is there.  Then, for each page in the range,
5905 		 * adjust its individual permissions.  Note that write-
5906 		 * enabling a MAP_PRIVATE page can affect the claims for
5907 		 * locked down memory.  Overcommitting memory terminates
5908 		 * the operation.
5909 		 */
5910 		segvn_vpage(seg);
5911 		if (svd->vpage == NULL) {
5912 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5913 			return (ENOMEM);
5914 		}
5915 		svd->pageprot = 1;
5916 		if ((amp = svd->amp) != NULL) {
5917 			anon_idx = svd->anon_index + seg_page(seg, addr);
5918 			ASSERT(seg->s_szc == 0 ||
5919 			    IS_P2ALIGNED(anon_idx, pgcnt));
5920 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5921 		}
5922 
5923 		offset = svd->offset + (uintptr_t)(addr - seg->s_base);
5924 		evp = &svd->vpage[seg_page(seg, addr + len)];
5925 
5926 		/*
5927 		 * See Statement at the beginning of segvn_lockop regarding
5928 		 * the way cowcnts and lckcnts are handled.
5929 		 */
5930 		for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5931 
5932 			if (seg->s_szc != 0) {
5933 				if (amp != NULL) {
5934 					anon_array_enter(amp, anon_idx,
5935 					    &cookie);
5936 				}
5937 				if (IS_P2ALIGNED(anon_idx, pgcnt) &&
5938 				    !segvn_claim_pages(seg, svp, offset,
5939 				    anon_idx, prot)) {
5940 					if (amp != NULL) {
5941 						anon_array_exit(&cookie);
5942 					}
5943 					break;
5944 				}
5945 				if (amp != NULL) {
5946 					anon_array_exit(&cookie);
5947 				}
5948 				anon_idx++;
5949 			} else {
5950 				if (amp != NULL) {
5951 					anon_array_enter(amp, anon_idx,
5952 					    &cookie);
5953 					ap = anon_get_ptr(amp->ahp, anon_idx++);
5954 				}
5955 
5956 				if (VPP_ISPPLOCK(svp) &&
5957 				    VPP_PROT(svp) != prot) {
5958 
5959 					if (amp == NULL || ap == NULL) {
5960 						vp = svd->vp;
5961 						off = offset;
5962 					} else
5963 						swap_xlate(ap, &vp, &off);
5964 					if (amp != NULL)
5965 						anon_array_exit(&cookie);
5966 
5967 					if ((pp = page_lookup(vp, off,
5968 					    SE_SHARED)) == NULL) {
5969 						panic("segvn_setprot: no page");
5970 						/*NOTREACHED*/
5971 					}
5972 					ASSERT(seg->s_szc == 0);
5973 					if ((VPP_PROT(svp) ^ prot) &
5974 					    PROT_WRITE) {
5975 						if (prot & PROT_WRITE) {
5976 							if (!page_addclaim(
5977 							    pp)) {
5978 								page_unlock(pp);
5979 								break;
5980 							}
5981 						} else {
5982 							if (!page_subclaim(
5983 							    pp)) {
5984 								page_unlock(pp);
5985 								break;
5986 							}
5987 						}
5988 					}
5989 					page_unlock(pp);
5990 				} else if (amp != NULL)
5991 					anon_array_exit(&cookie);
5992 			}
5993 			VPP_SETPROT(svp, prot);
5994 			offset += PAGESIZE;
5995 		}
5996 		if (amp != NULL)
5997 			ANON_LOCK_EXIT(&amp->a_rwlock);
5998 
5999 		/*
6000 		 * Did we terminate prematurely?  If so, simply unload
6001 		 * the translations to the things we've updated so far.
6002 		 */
6003 		if (svp != evp) {
6004 			if (unload_done) {
6005 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6006 				return (IE_NOMEM);
6007 			}
6008 			len = (svp - &svd->vpage[seg_page(seg, addr)]) *
6009 			    PAGESIZE;
6010 			ASSERT(seg->s_szc == 0 || IS_P2ALIGNED(len, pgsz));
6011 			if (len != 0)
6012 				hat_unload(seg->s_as->a_hat, addr,
6013 				    len, HAT_UNLOAD);
6014 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6015 			return (IE_NOMEM);
6016 		}
6017 	} else {
6018 		segvn_vpage(seg);
6019 		if (svd->vpage == NULL) {
6020 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6021 			return (ENOMEM);
6022 		}
6023 		svd->pageprot = 1;
6024 		evp = &svd->vpage[seg_page(seg, addr + len)];
6025 		for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
6026 			VPP_SETPROT(svp, prot);
6027 		}
6028 	}
6029 
6030 	if (unload_done) {
6031 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6032 		return (0);
6033 	}
6034 
6035 	if (((prot & PROT_WRITE) != 0 &&
6036 	    (svd->vp != NULL || svd->type == MAP_PRIVATE)) ||
6037 	    (prot & ~PROT_USER) == PROT_NONE) {
6038 		/*
6039 		 * Either private or shared data with write access (in
6040 		 * which case we need to throw out all former translations
6041 		 * so that we get the right translations set up on fault
6042 		 * and we don't allow write access to any copy-on-write pages
6043 		 * that might be around or to prevent write access to pages
6044 		 * representing holes in a file), or we don't have permission
6045 		 * to access the memory at all (in which case we have to
6046 		 * unload any current translations that might exist).
6047 		 */
6048 		hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
6049 	} else {
6050 		/*
6051 		 * A shared mapping or a private mapping in which write
6052 		 * protection is going to be denied - just change all the
6053 		 * protections over the range of addresses in question.
6054 		 * segvn does not support any other attributes other
6055 		 * than prot so we can use hat_chgattr.
6056 		 */
6057 		hat_chgattr(seg->s_as->a_hat, addr, len, prot);
6058 	}
6059 
6060 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6061 
6062 	return (0);
6063 }
6064 
6065 /*
6066  * segvn_setpagesize is called via SEGOP_SETPAGESIZE from as_setpagesize,
6067  * to determine if the seg is capable of mapping the requested szc.
6068  */
6069 static int
6070 segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len, uint_t szc)
6071 {
6072 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6073 	struct segvn_data *nsvd;
6074 	struct anon_map *amp = svd->amp;
6075 	struct seg *nseg;
6076 	caddr_t eaddr = addr + len, a;
6077 	size_t pgsz = page_get_pagesize(szc);
6078 	pgcnt_t pgcnt = page_get_pagecnt(szc);
6079 	int err;
6080 	u_offset_t off = svd->offset + (uintptr_t)(addr - seg->s_base);
6081 
6082 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
6083 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6084 
6085 	if (seg->s_szc == szc || segvn_lpg_disable != 0) {
6086 		return (0);
6087 	}
6088 
6089 	/*
6090 	 * addr should always be pgsz aligned but eaddr may be misaligned if
6091 	 * it's at the end of the segment.
6092 	 *
6093 	 * XXX we should assert this condition since as_setpagesize() logic
6094 	 * guarantees it.
6095 	 */
6096 	if (!IS_P2ALIGNED(addr, pgsz) ||
6097 	    (!IS_P2ALIGNED(eaddr, pgsz) &&
6098 	    eaddr != seg->s_base + seg->s_size)) {
6099 
6100 		segvn_setpgsz_align_err++;
6101 		return (EINVAL);
6102 	}
6103 
6104 	if (amp != NULL && svd->type == MAP_SHARED) {
6105 		ulong_t an_idx = svd->anon_index + seg_page(seg, addr);
6106 		if (!IS_P2ALIGNED(an_idx, pgcnt)) {
6107 
6108 			segvn_setpgsz_anon_align_err++;
6109 			return (EINVAL);
6110 		}
6111 	}
6112 
6113 	if ((svd->flags & MAP_NORESERVE) || seg->s_as == &kas ||
6114 	    szc > segvn_maxpgszc) {
6115 		return (EINVAL);
6116 	}
6117 
6118 	/* paranoid check */
6119 	if (svd->vp != NULL &&
6120 	    (IS_SWAPFSVP(svd->vp) || VN_ISKAS(svd->vp))) {
6121 		return (EINVAL);
6122 	}
6123 
6124 	if (seg->s_szc == 0 && svd->vp != NULL &&
6125 	    map_addr_vacalign_check(addr, off)) {
6126 		return (EINVAL);
6127 	}
6128 
6129 	/*
6130 	 * Check that protections are the same within new page
6131 	 * size boundaries.
6132 	 */
6133 	if (svd->pageprot) {
6134 		for (a = addr; a < eaddr; a += pgsz) {
6135 			if ((a + pgsz) > eaddr) {
6136 				if (!sameprot(seg, a, eaddr - a)) {
6137 					return (EINVAL);
6138 				}
6139 			} else {
6140 				if (!sameprot(seg, a, pgsz)) {
6141 					return (EINVAL);
6142 				}
6143 			}
6144 		}
6145 	}
6146 
6147 	/*
6148 	 * Since we are changing page size we first have to flush
6149 	 * the cache. This makes sure all the pagelock calls have
6150 	 * to recheck protections.
6151 	 */
6152 	if (svd->softlockcnt > 0) {
6153 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6154 
6155 		/*
6156 		 * If this is shared segment non 0 softlockcnt
6157 		 * means locked pages are still in use.
6158 		 */
6159 		if (svd->type == MAP_SHARED) {
6160 			return (EAGAIN);
6161 		}
6162 
6163 		/*
6164 		 * Since we do have the segvn writers lock nobody can fill
6165 		 * the cache with entries belonging to this seg during
6166 		 * the purge. The flush either succeeds or we still have
6167 		 * pending I/Os.
6168 		 */
6169 		segvn_purge(seg);
6170 		if (svd->softlockcnt > 0) {
6171 			return (EAGAIN);
6172 		}
6173 	}
6174 
6175 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6176 		ASSERT(svd->amp == NULL);
6177 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6178 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6179 		    HAT_REGION_TEXT);
6180 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6181 	} else if (svd->tr_state == SEGVN_TR_INIT) {
6182 		svd->tr_state = SEGVN_TR_OFF;
6183 	} else if (svd->tr_state == SEGVN_TR_ON) {
6184 		ASSERT(svd->amp != NULL);
6185 		segvn_textunrepl(seg, 1);
6186 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6187 		amp = NULL;
6188 	}
6189 
6190 	/*
6191 	 * Operation for sub range of existing segment.
6192 	 */
6193 	if (addr != seg->s_base || eaddr != (seg->s_base + seg->s_size)) {
6194 		if (szc < seg->s_szc) {
6195 			VM_STAT_ADD(segvnvmstats.demoterange[2]);
6196 			err = segvn_demote_range(seg, addr, len, SDR_RANGE, 0);
6197 			if (err == 0) {
6198 				return (IE_RETRY);
6199 			}
6200 			if (err == ENOMEM) {
6201 				return (IE_NOMEM);
6202 			}
6203 			return (err);
6204 		}
6205 		if (addr != seg->s_base) {
6206 			nseg = segvn_split_seg(seg, addr);
6207 			if (eaddr != (nseg->s_base + nseg->s_size)) {
6208 				/* eaddr is szc aligned */
6209 				(void) segvn_split_seg(nseg, eaddr);
6210 			}
6211 			return (IE_RETRY);
6212 		}
6213 		if (eaddr != (seg->s_base + seg->s_size)) {
6214 			/* eaddr is szc aligned */
6215 			(void) segvn_split_seg(seg, eaddr);
6216 		}
6217 		return (IE_RETRY);
6218 	}
6219 
6220 	/*
6221 	 * Break any low level sharing and reset seg->s_szc to 0.
6222 	 */
6223 	if ((err = segvn_clrszc(seg)) != 0) {
6224 		if (err == ENOMEM) {
6225 			err = IE_NOMEM;
6226 		}
6227 		return (err);
6228 	}
6229 	ASSERT(seg->s_szc == 0);
6230 
6231 	/*
6232 	 * If the end of the current segment is not pgsz aligned
6233 	 * then attempt to concatenate with the next segment.
6234 	 */
6235 	if (!IS_P2ALIGNED(eaddr, pgsz)) {
6236 		nseg = AS_SEGNEXT(seg->s_as, seg);
6237 		if (nseg == NULL || nseg == seg || eaddr != nseg->s_base) {
6238 			return (ENOMEM);
6239 		}
6240 		if (nseg->s_ops != &segvn_ops) {
6241 			return (EINVAL);
6242 		}
6243 		nsvd = (struct segvn_data *)nseg->s_data;
6244 		if (nsvd->softlockcnt > 0) {
6245 			/*
6246 			 * If this is shared segment non 0 softlockcnt
6247 			 * means locked pages are still in use.
6248 			 */
6249 			if (nsvd->type == MAP_SHARED) {
6250 				return (EAGAIN);
6251 			}
6252 			segvn_purge(nseg);
6253 			if (nsvd->softlockcnt > 0) {
6254 				return (EAGAIN);
6255 			}
6256 		}
6257 		err = segvn_clrszc(nseg);
6258 		if (err == ENOMEM) {
6259 			err = IE_NOMEM;
6260 		}
6261 		if (err != 0) {
6262 			return (err);
6263 		}
6264 		ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6265 		err = segvn_concat(seg, nseg, 1);
6266 		if (err == -1) {
6267 			return (EINVAL);
6268 		}
6269 		if (err == -2) {
6270 			return (IE_NOMEM);
6271 		}
6272 		return (IE_RETRY);
6273 	}
6274 
6275 	/*
6276 	 * May need to re-align anon array to
6277 	 * new szc.
6278 	 */
6279 	if (amp != NULL) {
6280 		if (!IS_P2ALIGNED(svd->anon_index, pgcnt)) {
6281 			struct anon_hdr *nahp;
6282 
6283 			ASSERT(svd->type == MAP_PRIVATE);
6284 
6285 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6286 			ASSERT(amp->refcnt == 1);
6287 			nahp = anon_create(btop(amp->size), ANON_NOSLEEP);
6288 			if (nahp == NULL) {
6289 				ANON_LOCK_EXIT(&amp->a_rwlock);
6290 				return (IE_NOMEM);
6291 			}
6292 			if (anon_copy_ptr(amp->ahp, svd->anon_index,
6293 			    nahp, 0, btop(seg->s_size), ANON_NOSLEEP)) {
6294 				anon_release(nahp, btop(amp->size));
6295 				ANON_LOCK_EXIT(&amp->a_rwlock);
6296 				return (IE_NOMEM);
6297 			}
6298 			anon_release(amp->ahp, btop(amp->size));
6299 			amp->ahp = nahp;
6300 			svd->anon_index = 0;
6301 			ANON_LOCK_EXIT(&amp->a_rwlock);
6302 		}
6303 	}
6304 	if (svd->vp != NULL && szc != 0) {
6305 		struct vattr va;
6306 		u_offset_t eoffpage = svd->offset;
6307 		va.va_mask = AT_SIZE;
6308 		eoffpage += seg->s_size;
6309 		eoffpage = btopr(eoffpage);
6310 		if (VOP_GETATTR(svd->vp, &va, 0, svd->cred, NULL) != 0) {
6311 			segvn_setpgsz_getattr_err++;
6312 			return (EINVAL);
6313 		}
6314 		if (btopr(va.va_size) < eoffpage) {
6315 			segvn_setpgsz_eof_err++;
6316 			return (EINVAL);
6317 		}
6318 		if (amp != NULL) {
6319 			/*
6320 			 * anon_fill_cow_holes() may call VOP_GETPAGE().
6321 			 * don't take anon map lock here to avoid holding it
6322 			 * across VOP_GETPAGE() calls that may call back into
6323 			 * segvn for klsutering checks. We don't really need
6324 			 * anon map lock here since it's a private segment and
6325 			 * we hold as level lock as writers.
6326 			 */
6327 			if ((err = anon_fill_cow_holes(seg, seg->s_base,
6328 			    amp->ahp, svd->anon_index, svd->vp, svd->offset,
6329 			    seg->s_size, szc, svd->prot, svd->vpage,
6330 			    svd->cred)) != 0) {
6331 				return (EINVAL);
6332 			}
6333 		}
6334 		segvn_setvnode_mpss(svd->vp);
6335 	}
6336 
6337 	if (amp != NULL) {
6338 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6339 		if (svd->type == MAP_PRIVATE) {
6340 			amp->a_szc = szc;
6341 		} else if (szc > amp->a_szc) {
6342 			amp->a_szc = szc;
6343 		}
6344 		ANON_LOCK_EXIT(&amp->a_rwlock);
6345 	}
6346 
6347 	seg->s_szc = szc;
6348 
6349 	return (0);
6350 }
6351 
6352 static int
6353 segvn_clrszc(struct seg *seg)
6354 {
6355 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6356 	struct anon_map *amp = svd->amp;
6357 	size_t pgsz;
6358 	pgcnt_t pages;
6359 	int err = 0;
6360 	caddr_t a = seg->s_base;
6361 	caddr_t ea = a + seg->s_size;
6362 	ulong_t an_idx = svd->anon_index;
6363 	vnode_t *vp = svd->vp;
6364 	struct vpage *vpage = svd->vpage;
6365 	page_t *anon_pl[1 + 1], *pp;
6366 	struct anon *ap, *oldap;
6367 	uint_t prot = svd->prot, vpprot;
6368 	int pageflag = 0;
6369 
6370 	ASSERT(AS_WRITE_HELD(seg->s_as) ||
6371 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
6372 	ASSERT(svd->softlockcnt == 0);
6373 
6374 	if (vp == NULL && amp == NULL) {
6375 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6376 		seg->s_szc = 0;
6377 		return (0);
6378 	}
6379 
6380 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6381 		ASSERT(svd->amp == NULL);
6382 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6383 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6384 		    HAT_REGION_TEXT);
6385 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6386 	} else if (svd->tr_state == SEGVN_TR_ON) {
6387 		ASSERT(svd->amp != NULL);
6388 		segvn_textunrepl(seg, 1);
6389 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6390 		amp = NULL;
6391 	} else {
6392 		if (svd->tr_state != SEGVN_TR_OFF) {
6393 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
6394 			svd->tr_state = SEGVN_TR_OFF;
6395 		}
6396 
6397 		/*
6398 		 * do HAT_UNLOAD_UNMAP since we are changing the pagesize.
6399 		 * unload argument is 0 when we are freeing the segment
6400 		 * and unload was already done.
6401 		 */
6402 		hat_unload(seg->s_as->a_hat, seg->s_base, seg->s_size,
6403 		    HAT_UNLOAD_UNMAP);
6404 	}
6405 
6406 	if (amp == NULL || svd->type == MAP_SHARED) {
6407 		seg->s_szc = 0;
6408 		return (0);
6409 	}
6410 
6411 	pgsz = page_get_pagesize(seg->s_szc);
6412 	pages = btop(pgsz);
6413 
6414 	/*
6415 	 * XXX anon rwlock is not really needed because this is a
6416 	 * private segment and we are writers.
6417 	 */
6418 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6419 
6420 	for (; a < ea; a += pgsz, an_idx += pages) {
6421 		if ((oldap = anon_get_ptr(amp->ahp, an_idx)) != NULL) {
6422 			ASSERT(vpage != NULL || svd->pageprot == 0);
6423 			if (vpage != NULL) {
6424 				ASSERT(sameprot(seg, a, pgsz));
6425 				prot = VPP_PROT(vpage);
6426 				pageflag = VPP_ISPPLOCK(vpage) ? LOCK_PAGE : 0;
6427 			}
6428 			if (seg->s_szc != 0) {
6429 				ASSERT(vp == NULL || anon_pages(amp->ahp,
6430 				    an_idx, pages) == pages);
6431 				if ((err = anon_map_demotepages(amp, an_idx,
6432 				    seg, a, prot, vpage, svd->cred)) != 0) {
6433 					goto out;
6434 				}
6435 			} else {
6436 				if (oldap->an_refcnt == 1) {
6437 					continue;
6438 				}
6439 				if ((err = anon_getpage(&oldap, &vpprot,
6440 				    anon_pl, PAGESIZE, seg, a, S_READ,
6441 				    svd->cred))) {
6442 					goto out;
6443 				}
6444 				if ((pp = anon_private(&ap, seg, a, prot,
6445 				    anon_pl[0], pageflag, svd->cred)) == NULL) {
6446 					err = ENOMEM;
6447 					goto out;
6448 				}
6449 				anon_decref(oldap);
6450 				(void) anon_set_ptr(amp->ahp, an_idx, ap,
6451 				    ANON_SLEEP);
6452 				page_unlock(pp);
6453 			}
6454 		}
6455 		vpage = (vpage == NULL) ? NULL : vpage + pages;
6456 	}
6457 
6458 	amp->a_szc = 0;
6459 	seg->s_szc = 0;
6460 out:
6461 	ANON_LOCK_EXIT(&amp->a_rwlock);
6462 	return (err);
6463 }
6464 
6465 static int
6466 segvn_claim_pages(
6467 	struct seg *seg,
6468 	struct vpage *svp,
6469 	u_offset_t off,
6470 	ulong_t anon_idx,
6471 	uint_t prot)
6472 {
6473 	pgcnt_t	pgcnt = page_get_pagecnt(seg->s_szc);
6474 	size_t ppasize = (pgcnt + 1) * sizeof (page_t *);
6475 	page_t	**ppa;
6476 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6477 	struct anon_map *amp = svd->amp;
6478 	struct vpage *evp = svp + pgcnt;
6479 	caddr_t addr = ((uintptr_t)(svp - svd->vpage) << PAGESHIFT)
6480 	    + seg->s_base;
6481 	struct anon *ap;
6482 	struct vnode *vp = svd->vp;
6483 	page_t *pp;
6484 	pgcnt_t pg_idx, i;
6485 	int err = 0;
6486 	anoff_t aoff;
6487 	int anon = (amp != NULL) ? 1 : 0;
6488 
6489 	ASSERT(svd->type == MAP_PRIVATE);
6490 	ASSERT(svd->vpage != NULL);
6491 	ASSERT(seg->s_szc != 0);
6492 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
6493 	ASSERT(amp == NULL || IS_P2ALIGNED(anon_idx, pgcnt));
6494 	ASSERT(sameprot(seg, addr, pgcnt << PAGESHIFT));
6495 
6496 	if (VPP_PROT(svp) == prot)
6497 		return (1);
6498 	if (!((VPP_PROT(svp) ^ prot) & PROT_WRITE))
6499 		return (1);
6500 
6501 	ppa = kmem_alloc(ppasize, KM_SLEEP);
6502 	if (anon && vp != NULL) {
6503 		if (anon_get_ptr(amp->ahp, anon_idx) == NULL) {
6504 			anon = 0;
6505 			ASSERT(!anon_pages(amp->ahp, anon_idx, pgcnt));
6506 		}
6507 		ASSERT(!anon ||
6508 		    anon_pages(amp->ahp, anon_idx, pgcnt) == pgcnt);
6509 	}
6510 
6511 	for (*ppa = NULL, pg_idx = 0; svp < evp; svp++, anon_idx++) {
6512 		if (!VPP_ISPPLOCK(svp))
6513 			continue;
6514 		if (anon) {
6515 			ap = anon_get_ptr(amp->ahp, anon_idx);
6516 			if (ap == NULL) {
6517 				panic("segvn_claim_pages: no anon slot");
6518 			}
6519 			swap_xlate(ap, &vp, &aoff);
6520 			off = (u_offset_t)aoff;
6521 		}
6522 		ASSERT(vp != NULL);
6523 		if ((pp = page_lookup(vp,
6524 		    (u_offset_t)off, SE_SHARED)) == NULL) {
6525 			panic("segvn_claim_pages: no page");
6526 		}
6527 		ppa[pg_idx++] = pp;
6528 		off += PAGESIZE;
6529 	}
6530 
6531 	if (ppa[0] == NULL) {
6532 		kmem_free(ppa, ppasize);
6533 		return (1);
6534 	}
6535 
6536 	ASSERT(pg_idx <= pgcnt);
6537 	ppa[pg_idx] = NULL;
6538 
6539 
6540 	/* Find each large page within ppa, and adjust its claim */
6541 
6542 	/* Does ppa cover a single large page? */
6543 	if (ppa[0]->p_szc == seg->s_szc) {
6544 		if (prot & PROT_WRITE)
6545 			err = page_addclaim_pages(ppa);
6546 		else
6547 			err = page_subclaim_pages(ppa);
6548 	} else {
6549 		for (i = 0; ppa[i]; i += pgcnt) {
6550 			ASSERT(IS_P2ALIGNED(page_pptonum(ppa[i]), pgcnt));
6551 			if (prot & PROT_WRITE)
6552 				err = page_addclaim_pages(&ppa[i]);
6553 			else
6554 				err = page_subclaim_pages(&ppa[i]);
6555 			if (err == 0)
6556 				break;
6557 		}
6558 	}
6559 
6560 	for (i = 0; i < pg_idx; i++) {
6561 		ASSERT(ppa[i] != NULL);
6562 		page_unlock(ppa[i]);
6563 	}
6564 
6565 	kmem_free(ppa, ppasize);
6566 	return (err);
6567 }
6568 
6569 /*
6570  * Returns right (upper address) segment if split occurred.
6571  * If the address is equal to the beginning or end of its segment it returns
6572  * the current segment.
6573  */
6574 static struct seg *
6575 segvn_split_seg(struct seg *seg, caddr_t addr)
6576 {
6577 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6578 	struct seg *nseg;
6579 	size_t nsize;
6580 	struct segvn_data *nsvd;
6581 
6582 	ASSERT(AS_WRITE_HELD(seg->s_as));
6583 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6584 
6585 	ASSERT(addr >= seg->s_base);
6586 	ASSERT(addr <= seg->s_base + seg->s_size);
6587 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6588 
6589 	if (addr == seg->s_base || addr == seg->s_base + seg->s_size)
6590 		return (seg);
6591 
6592 	nsize = seg->s_base + seg->s_size - addr;
6593 	seg->s_size = addr - seg->s_base;
6594 	nseg = seg_alloc(seg->s_as, addr, nsize);
6595 	ASSERT(nseg != NULL);
6596 	nseg->s_ops = seg->s_ops;
6597 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
6598 	nseg->s_data = (void *)nsvd;
6599 	nseg->s_szc = seg->s_szc;
6600 	*nsvd = *svd;
6601 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6602 	nsvd->seg = nseg;
6603 	rw_init(&nsvd->lock, NULL, RW_DEFAULT, NULL);
6604 
6605 	if (nsvd->vp != NULL) {
6606 		VN_HOLD(nsvd->vp);
6607 		nsvd->offset = svd->offset +
6608 		    (uintptr_t)(nseg->s_base - seg->s_base);
6609 		if (nsvd->type == MAP_SHARED)
6610 			lgrp_shm_policy_init(NULL, nsvd->vp);
6611 	} else {
6612 		/*
6613 		 * The offset for an anonymous segment has no signifigance in
6614 		 * terms of an offset into a file. If we were to use the above
6615 		 * calculation instead, the structures read out of
6616 		 * /proc/<pid>/xmap would be more difficult to decipher since
6617 		 * it would be unclear whether two seemingly contiguous
6618 		 * prxmap_t structures represented different segments or a
6619 		 * single segment that had been split up into multiple prxmap_t
6620 		 * structures (e.g. if some part of the segment had not yet
6621 		 * been faulted in).
6622 		 */
6623 		nsvd->offset = 0;
6624 	}
6625 
6626 	ASSERT(svd->softlockcnt == 0);
6627 	ASSERT(svd->softlockcnt_sbase == 0);
6628 	ASSERT(svd->softlockcnt_send == 0);
6629 	crhold(svd->cred);
6630 
6631 	if (svd->vpage != NULL) {
6632 		size_t bytes = vpgtob(seg_pages(seg));
6633 		size_t nbytes = vpgtob(seg_pages(nseg));
6634 		struct vpage *ovpage = svd->vpage;
6635 
6636 		svd->vpage = kmem_alloc(bytes, KM_SLEEP);
6637 		bcopy(ovpage, svd->vpage, bytes);
6638 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
6639 		bcopy(ovpage + seg_pages(seg), nsvd->vpage, nbytes);
6640 		kmem_free(ovpage, bytes + nbytes);
6641 	}
6642 	if (svd->amp != NULL && svd->type == MAP_PRIVATE) {
6643 		struct anon_map *oamp = svd->amp, *namp;
6644 		struct anon_hdr *nahp;
6645 
6646 		ANON_LOCK_ENTER(&oamp->a_rwlock, RW_WRITER);
6647 		ASSERT(oamp->refcnt == 1);
6648 		nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
6649 		(void) anon_copy_ptr(oamp->ahp, svd->anon_index,
6650 		    nahp, 0, btop(seg->s_size), ANON_SLEEP);
6651 
6652 		namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
6653 		namp->a_szc = nseg->s_szc;
6654 		(void) anon_copy_ptr(oamp->ahp,
6655 		    svd->anon_index + btop(seg->s_size),
6656 		    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
6657 		anon_release(oamp->ahp, btop(oamp->size));
6658 		oamp->ahp = nahp;
6659 		oamp->size = seg->s_size;
6660 		svd->anon_index = 0;
6661 		nsvd->amp = namp;
6662 		nsvd->anon_index = 0;
6663 		ANON_LOCK_EXIT(&oamp->a_rwlock);
6664 	} else if (svd->amp != NULL) {
6665 		pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6666 		ASSERT(svd->amp == nsvd->amp);
6667 		ASSERT(seg->s_szc <= svd->amp->a_szc);
6668 		nsvd->anon_index = svd->anon_index + seg_pages(seg);
6669 		ASSERT(IS_P2ALIGNED(nsvd->anon_index, pgcnt));
6670 		ANON_LOCK_ENTER(&svd->amp->a_rwlock, RW_WRITER);
6671 		svd->amp->refcnt++;
6672 		ANON_LOCK_EXIT(&svd->amp->a_rwlock);
6673 	}
6674 
6675 	/*
6676 	 * Split the amount of swap reserved.
6677 	 */
6678 	if (svd->swresv) {
6679 		/*
6680 		 * For MAP_NORESERVE, only allocate swap reserve for pages
6681 		 * being used.  Other segments get enough to cover whole
6682 		 * segment.
6683 		 */
6684 		if (svd->flags & MAP_NORESERVE) {
6685 			size_t	oswresv;
6686 
6687 			ASSERT(svd->amp);
6688 			oswresv = svd->swresv;
6689 			svd->swresv = ptob(anon_pages(svd->amp->ahp,
6690 			    svd->anon_index, btop(seg->s_size)));
6691 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
6692 			    nsvd->anon_index, btop(nseg->s_size)));
6693 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
6694 		} else {
6695 			if (svd->pageswap) {
6696 				svd->swresv = segvn_count_swap_by_vpages(seg);
6697 				ASSERT(nsvd->swresv >= svd->swresv);
6698 				nsvd->swresv -= svd->swresv;
6699 			} else {
6700 				ASSERT(svd->swresv == seg->s_size +
6701 				    nseg->s_size);
6702 				svd->swresv = seg->s_size;
6703 				nsvd->swresv = nseg->s_size;
6704 			}
6705 		}
6706 	}
6707 
6708 	return (nseg);
6709 }
6710 
6711 /*
6712  * called on memory operations (unmap, setprot, setpagesize) for a subset
6713  * of a large page segment to either demote the memory range (SDR_RANGE)
6714  * or the ends (SDR_END) by addr/len.
6715  *
6716  * returns 0 on success. returns errno, including ENOMEM, on failure.
6717  */
6718 static int
6719 segvn_demote_range(
6720 	struct seg *seg,
6721 	caddr_t addr,
6722 	size_t len,
6723 	int flag,
6724 	uint_t szcvec)
6725 {
6726 	caddr_t eaddr = addr + len;
6727 	caddr_t lpgaddr, lpgeaddr;
6728 	struct seg *nseg;
6729 	struct seg *badseg1 = NULL;
6730 	struct seg *badseg2 = NULL;
6731 	size_t pgsz;
6732 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6733 	int err;
6734 	uint_t szc = seg->s_szc;
6735 	uint_t tszcvec;
6736 
6737 	ASSERT(AS_WRITE_HELD(seg->s_as));
6738 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6739 	ASSERT(szc != 0);
6740 	pgsz = page_get_pagesize(szc);
6741 	ASSERT(seg->s_base != addr || seg->s_size != len);
6742 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6743 	ASSERT(svd->softlockcnt == 0);
6744 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6745 	ASSERT(szcvec == 0 || (flag == SDR_END && svd->type == MAP_SHARED));
6746 
6747 	CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
6748 	ASSERT(flag == SDR_RANGE || eaddr < lpgeaddr || addr > lpgaddr);
6749 	if (flag == SDR_RANGE) {
6750 		/* demote entire range */
6751 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6752 		(void) segvn_split_seg(nseg, lpgeaddr);
6753 		ASSERT(badseg1->s_base == lpgaddr);
6754 		ASSERT(badseg1->s_size == lpgeaddr - lpgaddr);
6755 	} else if (addr != lpgaddr) {
6756 		ASSERT(flag == SDR_END);
6757 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6758 		if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz &&
6759 		    eaddr < lpgaddr + 2 * pgsz) {
6760 			(void) segvn_split_seg(nseg, lpgeaddr);
6761 			ASSERT(badseg1->s_base == lpgaddr);
6762 			ASSERT(badseg1->s_size == 2 * pgsz);
6763 		} else {
6764 			nseg = segvn_split_seg(nseg, lpgaddr + pgsz);
6765 			ASSERT(badseg1->s_base == lpgaddr);
6766 			ASSERT(badseg1->s_size == pgsz);
6767 			if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz) {
6768 				ASSERT(lpgeaddr - lpgaddr > 2 * pgsz);
6769 				nseg = segvn_split_seg(nseg, lpgeaddr - pgsz);
6770 				badseg2 = nseg;
6771 				(void) segvn_split_seg(nseg, lpgeaddr);
6772 				ASSERT(badseg2->s_base == lpgeaddr - pgsz);
6773 				ASSERT(badseg2->s_size == pgsz);
6774 			}
6775 		}
6776 	} else {
6777 		ASSERT(flag == SDR_END);
6778 		ASSERT(eaddr < lpgeaddr);
6779 		badseg1 = nseg = segvn_split_seg(seg, lpgeaddr - pgsz);
6780 		(void) segvn_split_seg(nseg, lpgeaddr);
6781 		ASSERT(badseg1->s_base == lpgeaddr - pgsz);
6782 		ASSERT(badseg1->s_size == pgsz);
6783 	}
6784 
6785 	ASSERT(badseg1 != NULL);
6786 	ASSERT(badseg1->s_szc == szc);
6787 	ASSERT(flag == SDR_RANGE || badseg1->s_size == pgsz ||
6788 	    badseg1->s_size == 2 * pgsz);
6789 	ASSERT(sameprot(badseg1, badseg1->s_base, pgsz));
6790 	ASSERT(badseg1->s_size == pgsz ||
6791 	    sameprot(badseg1, badseg1->s_base + pgsz, pgsz));
6792 	if (err = segvn_clrszc(badseg1)) {
6793 		return (err);
6794 	}
6795 	ASSERT(badseg1->s_szc == 0);
6796 
6797 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6798 		uint_t tszc = highbit(tszcvec) - 1;
6799 		caddr_t ta = MAX(addr, badseg1->s_base);
6800 		caddr_t te;
6801 		size_t tpgsz = page_get_pagesize(tszc);
6802 
6803 		ASSERT(svd->type == MAP_SHARED);
6804 		ASSERT(flag == SDR_END);
6805 		ASSERT(tszc < szc && tszc > 0);
6806 
6807 		if (eaddr > badseg1->s_base + badseg1->s_size) {
6808 			te = badseg1->s_base + badseg1->s_size;
6809 		} else {
6810 			te = eaddr;
6811 		}
6812 
6813 		ASSERT(ta <= te);
6814 		badseg1->s_szc = tszc;
6815 		if (!IS_P2ALIGNED(ta, tpgsz) || !IS_P2ALIGNED(te, tpgsz)) {
6816 			if (badseg2 != NULL) {
6817 				err = segvn_demote_range(badseg1, ta, te - ta,
6818 				    SDR_END, tszcvec);
6819 				if (err != 0) {
6820 					return (err);
6821 				}
6822 			} else {
6823 				return (segvn_demote_range(badseg1, ta,
6824 				    te - ta, SDR_END, tszcvec));
6825 			}
6826 		}
6827 	}
6828 
6829 	if (badseg2 == NULL)
6830 		return (0);
6831 	ASSERT(badseg2->s_szc == szc);
6832 	ASSERT(badseg2->s_size == pgsz);
6833 	ASSERT(sameprot(badseg2, badseg2->s_base, badseg2->s_size));
6834 	if (err = segvn_clrszc(badseg2)) {
6835 		return (err);
6836 	}
6837 	ASSERT(badseg2->s_szc == 0);
6838 
6839 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6840 		uint_t tszc = highbit(tszcvec) - 1;
6841 		size_t tpgsz = page_get_pagesize(tszc);
6842 
6843 		ASSERT(svd->type == MAP_SHARED);
6844 		ASSERT(flag == SDR_END);
6845 		ASSERT(tszc < szc && tszc > 0);
6846 		ASSERT(badseg2->s_base > addr);
6847 		ASSERT(eaddr > badseg2->s_base);
6848 		ASSERT(eaddr < badseg2->s_base + badseg2->s_size);
6849 
6850 		badseg2->s_szc = tszc;
6851 		if (!IS_P2ALIGNED(eaddr, tpgsz)) {
6852 			return (segvn_demote_range(badseg2, badseg2->s_base,
6853 			    eaddr - badseg2->s_base, SDR_END, tszcvec));
6854 		}
6855 	}
6856 
6857 	return (0);
6858 }
6859 
6860 static int
6861 segvn_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
6862 {
6863 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6864 	struct vpage *vp, *evp;
6865 
6866 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6867 
6868 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6869 	/*
6870 	 * If segment protection can be used, simply check against them.
6871 	 */
6872 	if (svd->pageprot == 0) {
6873 		int err;
6874 
6875 		err = ((svd->prot & prot) != prot) ? EACCES : 0;
6876 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6877 		return (err);
6878 	}
6879 
6880 	/*
6881 	 * Have to check down to the vpage level.
6882 	 */
6883 	evp = &svd->vpage[seg_page(seg, addr + len)];
6884 	for (vp = &svd->vpage[seg_page(seg, addr)]; vp < evp; vp++) {
6885 		if ((VPP_PROT(vp) & prot) != prot) {
6886 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6887 			return (EACCES);
6888 		}
6889 	}
6890 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6891 	return (0);
6892 }
6893 
6894 static int
6895 segvn_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
6896 {
6897 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6898 	size_t pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
6899 
6900 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6901 
6902 	if (pgno != 0) {
6903 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6904 		if (svd->pageprot == 0) {
6905 			do {
6906 				protv[--pgno] = svd->prot;
6907 			} while (pgno != 0);
6908 		} else {
6909 			size_t pgoff = seg_page(seg, addr);
6910 
6911 			do {
6912 				pgno--;
6913 				protv[pgno] = VPP_PROT(&svd->vpage[pgno+pgoff]);
6914 			} while (pgno != 0);
6915 		}
6916 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6917 	}
6918 	return (0);
6919 }
6920 
6921 static u_offset_t
6922 segvn_getoffset(struct seg *seg, caddr_t addr)
6923 {
6924 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6925 
6926 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6927 
6928 	return (svd->offset + (uintptr_t)(addr - seg->s_base));
6929 }
6930 
6931 /*ARGSUSED*/
6932 static int
6933 segvn_gettype(struct seg *seg, caddr_t addr)
6934 {
6935 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6936 
6937 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6938 
6939 	return (svd->type | (svd->flags & (MAP_NORESERVE | MAP_TEXT |
6940 	    MAP_INITDATA)));
6941 }
6942 
6943 /*ARGSUSED*/
6944 static int
6945 segvn_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
6946 {
6947 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6948 
6949 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6950 
6951 	*vpp = svd->vp;
6952 	return (0);
6953 }
6954 
6955 /*
6956  * Check to see if it makes sense to do kluster/read ahead to
6957  * addr + delta relative to the mapping at addr.  We assume here
6958  * that delta is a signed PAGESIZE'd multiple (which can be negative).
6959  *
6960  * For segvn, we currently "approve" of the action if we are
6961  * still in the segment and it maps from the same vp/off,
6962  * or if the advice stored in segvn_data or vpages allows it.
6963  * Currently, klustering is not allowed only if MADV_RANDOM is set.
6964  */
6965 static int
6966 segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
6967 {
6968 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6969 	struct anon *oap, *ap;
6970 	ssize_t pd;
6971 	size_t page;
6972 	struct vnode *vp1, *vp2;
6973 	u_offset_t off1, off2;
6974 	struct anon_map *amp;
6975 
6976 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6977 	ASSERT(AS_WRITE_HELD(seg->s_as) ||
6978 	    SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
6979 
6980 	if (addr + delta < seg->s_base ||
6981 	    addr + delta >= (seg->s_base + seg->s_size))
6982 		return (-1);		/* exceeded segment bounds */
6983 
6984 	pd = delta / (ssize_t)PAGESIZE;	/* divide to preserve sign bit */
6985 	page = seg_page(seg, addr);
6986 
6987 	/*
6988 	 * Check to see if either of the pages addr or addr + delta
6989 	 * have advice set that prevents klustering (if MADV_RANDOM advice
6990 	 * is set for entire segment, or MADV_SEQUENTIAL is set and delta
6991 	 * is negative).
6992 	 */
6993 	if (svd->advice == MADV_RANDOM ||
6994 	    svd->advice == MADV_SEQUENTIAL && delta < 0)
6995 		return (-1);
6996 	else if (svd->pageadvice && svd->vpage) {
6997 		struct vpage *bvpp, *evpp;
6998 
6999 		bvpp = &svd->vpage[page];
7000 		evpp = &svd->vpage[page + pd];
7001 		if (VPP_ADVICE(bvpp) == MADV_RANDOM ||
7002 		    VPP_ADVICE(evpp) == MADV_SEQUENTIAL && delta < 0)
7003 			return (-1);
7004 		if (VPP_ADVICE(bvpp) != VPP_ADVICE(evpp) &&
7005 		    VPP_ADVICE(evpp) == MADV_RANDOM)
7006 			return (-1);
7007 	}
7008 
7009 	if (svd->type == MAP_SHARED)
7010 		return (0);		/* shared mapping - all ok */
7011 
7012 	if ((amp = svd->amp) == NULL)
7013 		return (0);		/* off original vnode */
7014 
7015 	page += svd->anon_index;
7016 
7017 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7018 
7019 	oap = anon_get_ptr(amp->ahp, page);
7020 	ap = anon_get_ptr(amp->ahp, page + pd);
7021 
7022 	ANON_LOCK_EXIT(&amp->a_rwlock);
7023 
7024 	if ((oap == NULL && ap != NULL) || (oap != NULL && ap == NULL)) {
7025 		return (-1);		/* one with and one without an anon */
7026 	}
7027 
7028 	if (oap == NULL) {		/* implies that ap == NULL */
7029 		return (0);		/* off original vnode */
7030 	}
7031 
7032 	/*
7033 	 * Now we know we have two anon pointers - check to
7034 	 * see if they happen to be properly allocated.
7035 	 */
7036 
7037 	/*
7038 	 * XXX We cheat here and don't lock the anon slots. We can't because
7039 	 * we may have been called from the anon layer which might already
7040 	 * have locked them. We are holding a refcnt on the slots so they
7041 	 * can't disappear. The worst that will happen is we'll get the wrong
7042 	 * names (vp, off) for the slots and make a poor klustering decision.
7043 	 */
7044 	swap_xlate(ap, &vp1, &off1);
7045 	swap_xlate(oap, &vp2, &off2);
7046 
7047 
7048 	if (!VOP_CMP(vp1, vp2, NULL) || off1 - off2 != delta)
7049 		return (-1);
7050 	return (0);
7051 }
7052 
7053 /*
7054  * Swap the pages of seg out to secondary storage, returning the
7055  * number of bytes of storage freed.
7056  *
7057  * The basic idea is first to unload all translations and then to call
7058  * VOP_PUTPAGE() for all newly-unmapped pages, to push them out to the
7059  * swap device.  Pages to which other segments have mappings will remain
7060  * mapped and won't be swapped.  Our caller (as_swapout) has already
7061  * performed the unloading step.
7062  *
7063  * The value returned is intended to correlate well with the process's
7064  * memory requirements.  However, there are some caveats:
7065  * 1)	When given a shared segment as argument, this routine will
7066  *	only succeed in swapping out pages for the last sharer of the
7067  *	segment.  (Previous callers will only have decremented mapping
7068  *	reference counts.)
7069  * 2)	We assume that the hat layer maintains a large enough translation
7070  *	cache to capture process reference patterns.
7071  */
7072 static size_t
7073 segvn_swapout(struct seg *seg)
7074 {
7075 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7076 	struct anon_map *amp;
7077 	pgcnt_t pgcnt = 0;
7078 	pgcnt_t npages;
7079 	pgcnt_t page;
7080 	ulong_t anon_index;
7081 
7082 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7083 
7084 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7085 	/*
7086 	 * Find pages unmapped by our caller and force them
7087 	 * out to the virtual swap device.
7088 	 */
7089 	if ((amp = svd->amp) != NULL)
7090 		anon_index = svd->anon_index;
7091 	npages = seg->s_size >> PAGESHIFT;
7092 	for (page = 0; page < npages; page++) {
7093 		page_t *pp;
7094 		struct anon *ap;
7095 		struct vnode *vp;
7096 		u_offset_t off;
7097 		anon_sync_obj_t cookie;
7098 
7099 		/*
7100 		 * Obtain <vp, off> pair for the page, then look it up.
7101 		 *
7102 		 * Note that this code is willing to consider regular
7103 		 * pages as well as anon pages.  Is this appropriate here?
7104 		 */
7105 		ap = NULL;
7106 		if (amp != NULL) {
7107 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7108 			if (anon_array_try_enter(amp, anon_index + page,
7109 			    &cookie)) {
7110 				ANON_LOCK_EXIT(&amp->a_rwlock);
7111 				continue;
7112 			}
7113 			ap = anon_get_ptr(amp->ahp, anon_index + page);
7114 			if (ap != NULL) {
7115 				swap_xlate(ap, &vp, &off);
7116 			} else {
7117 				vp = svd->vp;
7118 				off = svd->offset + ptob(page);
7119 			}
7120 			anon_array_exit(&cookie);
7121 			ANON_LOCK_EXIT(&amp->a_rwlock);
7122 		} else {
7123 			vp = svd->vp;
7124 			off = svd->offset + ptob(page);
7125 		}
7126 		if (vp == NULL) {		/* untouched zfod page */
7127 			ASSERT(ap == NULL);
7128 			continue;
7129 		}
7130 
7131 		pp = page_lookup_nowait(vp, off, SE_SHARED);
7132 		if (pp == NULL)
7133 			continue;
7134 
7135 
7136 		/*
7137 		 * Examine the page to see whether it can be tossed out,
7138 		 * keeping track of how many we've found.
7139 		 */
7140 		if (!page_tryupgrade(pp)) {
7141 			/*
7142 			 * If the page has an i/o lock and no mappings,
7143 			 * it's very likely that the page is being
7144 			 * written out as a result of klustering.
7145 			 * Assume this is so and take credit for it here.
7146 			 */
7147 			if (!page_io_trylock(pp)) {
7148 				if (!hat_page_is_mapped(pp))
7149 					pgcnt++;
7150 			} else {
7151 				page_io_unlock(pp);
7152 			}
7153 			page_unlock(pp);
7154 			continue;
7155 		}
7156 		ASSERT(!page_iolock_assert(pp));
7157 
7158 
7159 		/*
7160 		 * Skip if page is locked or has mappings.
7161 		 * We don't need the page_struct_lock to look at lckcnt
7162 		 * and cowcnt because the page is exclusive locked.
7163 		 */
7164 		if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0 ||
7165 		    hat_page_is_mapped(pp)) {
7166 			page_unlock(pp);
7167 			continue;
7168 		}
7169 
7170 		/*
7171 		 * dispose skips large pages so try to demote first.
7172 		 */
7173 		if (pp->p_szc != 0 && !page_try_demote_pages(pp)) {
7174 			page_unlock(pp);
7175 			/*
7176 			 * XXX should skip the remaining page_t's of this
7177 			 * large page.
7178 			 */
7179 			continue;
7180 		}
7181 
7182 		ASSERT(pp->p_szc == 0);
7183 
7184 		/*
7185 		 * No longer mapped -- we can toss it out.  How
7186 		 * we do so depends on whether or not it's dirty.
7187 		 */
7188 		if (hat_ismod(pp) && pp->p_vnode) {
7189 			/*
7190 			 * We must clean the page before it can be
7191 			 * freed.  Setting B_FREE will cause pvn_done
7192 			 * to free the page when the i/o completes.
7193 			 * XXX:	This also causes it to be accounted
7194 			 *	as a pageout instead of a swap: need
7195 			 *	B_SWAPOUT bit to use instead of B_FREE.
7196 			 *
7197 			 * Hold the vnode before releasing the page lock
7198 			 * to prevent it from being freed and re-used by
7199 			 * some other thread.
7200 			 */
7201 			VN_HOLD(vp);
7202 			page_unlock(pp);
7203 
7204 			/*
7205 			 * Queue all i/o requests for the pageout thread
7206 			 * to avoid saturating the pageout devices.
7207 			 */
7208 			if (!queue_io_request(vp, off))
7209 				VN_RELE(vp);
7210 		} else {
7211 			/*
7212 			 * The page was clean, free it.
7213 			 *
7214 			 * XXX:	Can we ever encounter modified pages
7215 			 *	with no associated vnode here?
7216 			 */
7217 			ASSERT(pp->p_vnode != NULL);
7218 			/*LINTED: constant in conditional context*/
7219 			VN_DISPOSE(pp, B_FREE, 0, kcred);
7220 		}
7221 
7222 		/*
7223 		 * Credit now even if i/o is in progress.
7224 		 */
7225 		pgcnt++;
7226 	}
7227 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7228 
7229 	/*
7230 	 * Wakeup pageout to initiate i/o on all queued requests.
7231 	 */
7232 	cv_signal_pageout();
7233 	return (ptob(pgcnt));
7234 }
7235 
7236 /*
7237  * Synchronize primary storage cache with real object in virtual memory.
7238  *
7239  * XXX - Anonymous pages should not be sync'ed out at all.
7240  */
7241 static int
7242 segvn_sync(struct seg *seg, caddr_t addr, size_t len, int attr, uint_t flags)
7243 {
7244 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7245 	struct vpage *vpp;
7246 	page_t *pp;
7247 	u_offset_t offset;
7248 	struct vnode *vp;
7249 	u_offset_t off;
7250 	caddr_t eaddr;
7251 	int bflags;
7252 	int err = 0;
7253 	int segtype;
7254 	int pageprot;
7255 	int prot;
7256 	ulong_t anon_index;
7257 	struct anon_map *amp;
7258 	struct anon *ap;
7259 	anon_sync_obj_t cookie;
7260 
7261 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7262 
7263 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7264 
7265 	if (svd->softlockcnt > 0) {
7266 		/*
7267 		 * If this is shared segment non 0 softlockcnt
7268 		 * means locked pages are still in use.
7269 		 */
7270 		if (svd->type == MAP_SHARED) {
7271 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7272 			return (EAGAIN);
7273 		}
7274 
7275 		/*
7276 		 * flush all pages from seg cache
7277 		 * otherwise we may deadlock in swap_putpage
7278 		 * for B_INVAL page (4175402).
7279 		 *
7280 		 * Even if we grab segvn WRITER's lock
7281 		 * here, there might be another thread which could've
7282 		 * successfully performed lookup/insert just before
7283 		 * we acquired the lock here.  So, grabbing either
7284 		 * lock here is of not much use.  Until we devise
7285 		 * a strategy at upper layers to solve the
7286 		 * synchronization issues completely, we expect
7287 		 * applications to handle this appropriately.
7288 		 */
7289 		segvn_purge(seg);
7290 		if (svd->softlockcnt > 0) {
7291 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7292 			return (EAGAIN);
7293 		}
7294 	} else if (svd->type == MAP_SHARED && svd->amp != NULL &&
7295 	    svd->amp->a_softlockcnt > 0) {
7296 		/*
7297 		 * Try to purge this amp's entries from pcache. It will
7298 		 * succeed only if other segments that share the amp have no
7299 		 * outstanding softlock's.
7300 		 */
7301 		segvn_purge(seg);
7302 		if (svd->amp->a_softlockcnt > 0 || svd->softlockcnt > 0) {
7303 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7304 			return (EAGAIN);
7305 		}
7306 	}
7307 
7308 	vpp = svd->vpage;
7309 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7310 	bflags = ((flags & MS_ASYNC) ? B_ASYNC : 0) |
7311 	    ((flags & MS_INVALIDATE) ? B_INVAL : 0);
7312 
7313 	if (attr) {
7314 		pageprot = attr & ~(SHARED|PRIVATE);
7315 		segtype = (attr & SHARED) ? MAP_SHARED : MAP_PRIVATE;
7316 
7317 		/*
7318 		 * We are done if the segment types don't match
7319 		 * or if we have segment level protections and
7320 		 * they don't match.
7321 		 */
7322 		if (svd->type != segtype) {
7323 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7324 			return (0);
7325 		}
7326 		if (vpp == NULL) {
7327 			if (svd->prot != pageprot) {
7328 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7329 				return (0);
7330 			}
7331 			prot = svd->prot;
7332 		} else
7333 			vpp = &svd->vpage[seg_page(seg, addr)];
7334 
7335 	} else if (svd->vp && svd->amp == NULL &&
7336 	    (flags & MS_INVALIDATE) == 0) {
7337 
7338 		/*
7339 		 * No attributes, no anonymous pages and MS_INVALIDATE flag
7340 		 * is not on, just use one big request.
7341 		 */
7342 		err = VOP_PUTPAGE(svd->vp, (offset_t)offset, len,
7343 		    bflags, svd->cred, NULL);
7344 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7345 		return (err);
7346 	}
7347 
7348 	if ((amp = svd->amp) != NULL)
7349 		anon_index = svd->anon_index + seg_page(seg, addr);
7350 
7351 	for (eaddr = addr + len; addr < eaddr; addr += PAGESIZE) {
7352 		ap = NULL;
7353 		if (amp != NULL) {
7354 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7355 			anon_array_enter(amp, anon_index, &cookie);
7356 			ap = anon_get_ptr(amp->ahp, anon_index++);
7357 			if (ap != NULL) {
7358 				swap_xlate(ap, &vp, &off);
7359 			} else {
7360 				vp = svd->vp;
7361 				off = offset;
7362 			}
7363 			anon_array_exit(&cookie);
7364 			ANON_LOCK_EXIT(&amp->a_rwlock);
7365 		} else {
7366 			vp = svd->vp;
7367 			off = offset;
7368 		}
7369 		offset += PAGESIZE;
7370 
7371 		if (vp == NULL)		/* untouched zfod page */
7372 			continue;
7373 
7374 		if (attr) {
7375 			if (vpp) {
7376 				prot = VPP_PROT(vpp);
7377 				vpp++;
7378 			}
7379 			if (prot != pageprot) {
7380 				continue;
7381 			}
7382 		}
7383 
7384 		/*
7385 		 * See if any of these pages are locked --  if so, then we
7386 		 * will have to truncate an invalidate request at the first
7387 		 * locked one. We don't need the page_struct_lock to test
7388 		 * as this is only advisory; even if we acquire it someone
7389 		 * might race in and lock the page after we unlock and before
7390 		 * we do the PUTPAGE, then PUTPAGE simply does nothing.
7391 		 */
7392 		if (flags & MS_INVALIDATE) {
7393 			if ((pp = page_lookup(vp, off, SE_SHARED)) != NULL) {
7394 				if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
7395 					page_unlock(pp);
7396 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7397 					return (EBUSY);
7398 				}
7399 				if (ap != NULL && pp->p_szc != 0 &&
7400 				    page_tryupgrade(pp)) {
7401 					if (pp->p_lckcnt == 0 &&
7402 					    pp->p_cowcnt == 0) {
7403 						/*
7404 						 * swapfs VN_DISPOSE() won't
7405 						 * invalidate large pages.
7406 						 * Attempt to demote.
7407 						 * XXX can't help it if it
7408 						 * fails. But for swapfs
7409 						 * pages it is no big deal.
7410 						 */
7411 						(void) page_try_demote_pages(
7412 						    pp);
7413 					}
7414 				}
7415 				page_unlock(pp);
7416 			}
7417 		} else if (svd->type == MAP_SHARED && amp != NULL) {
7418 			/*
7419 			 * Avoid writing out to disk ISM's large pages
7420 			 * because segspt_free_pages() relies on NULL an_pvp
7421 			 * of anon slots of such pages.
7422 			 */
7423 
7424 			ASSERT(svd->vp == NULL);
7425 			/*
7426 			 * swapfs uses page_lookup_nowait if not freeing or
7427 			 * invalidating and skips a page if
7428 			 * page_lookup_nowait returns NULL.
7429 			 */
7430 			pp = page_lookup_nowait(vp, off, SE_SHARED);
7431 			if (pp == NULL) {
7432 				continue;
7433 			}
7434 			if (pp->p_szc != 0) {
7435 				page_unlock(pp);
7436 				continue;
7437 			}
7438 
7439 			/*
7440 			 * Note ISM pages are created large so (vp, off)'s
7441 			 * page cannot suddenly become large after we unlock
7442 			 * pp.
7443 			 */
7444 			page_unlock(pp);
7445 		}
7446 		/*
7447 		 * XXX - Should ultimately try to kluster
7448 		 * calls to VOP_PUTPAGE() for performance.
7449 		 */
7450 		VN_HOLD(vp);
7451 		err = VOP_PUTPAGE(vp, (offset_t)off, PAGESIZE,
7452 		    (bflags | (IS_SWAPFSVP(vp) ? B_PAGE_NOWAIT : 0)),
7453 		    svd->cred, NULL);
7454 
7455 		VN_RELE(vp);
7456 		if (err)
7457 			break;
7458 	}
7459 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7460 	return (err);
7461 }
7462 
7463 /*
7464  * Determine if we have data corresponding to pages in the
7465  * primary storage virtual memory cache (i.e., "in core").
7466  */
7467 static size_t
7468 segvn_incore(struct seg *seg, caddr_t addr, size_t len, char *vec)
7469 {
7470 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7471 	struct vnode *vp, *avp;
7472 	u_offset_t offset, aoffset;
7473 	size_t p, ep;
7474 	int ret;
7475 	struct vpage *vpp;
7476 	page_t *pp;
7477 	uint_t start;
7478 	struct anon_map *amp;		/* XXX - for locknest */
7479 	struct anon *ap;
7480 	uint_t attr;
7481 	anon_sync_obj_t cookie;
7482 
7483 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7484 
7485 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7486 	if (svd->amp == NULL && svd->vp == NULL) {
7487 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7488 		bzero(vec, btopr(len));
7489 		return (len);	/* no anonymous pages created yet */
7490 	}
7491 
7492 	p = seg_page(seg, addr);
7493 	ep = seg_page(seg, addr + len);
7494 	start = svd->vp ? SEG_PAGE_VNODEBACKED : 0;
7495 
7496 	amp = svd->amp;
7497 	for (; p < ep; p++, addr += PAGESIZE) {
7498 		vpp = (svd->vpage) ? &svd->vpage[p]: NULL;
7499 		ret = start;
7500 		ap = NULL;
7501 		avp = NULL;
7502 		/* Grab the vnode/offset for the anon slot */
7503 		if (amp != NULL) {
7504 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7505 			anon_array_enter(amp, svd->anon_index + p, &cookie);
7506 			ap = anon_get_ptr(amp->ahp, svd->anon_index + p);
7507 			if (ap != NULL) {
7508 				swap_xlate(ap, &avp, &aoffset);
7509 			}
7510 			anon_array_exit(&cookie);
7511 			ANON_LOCK_EXIT(&amp->a_rwlock);
7512 		}
7513 		if ((avp != NULL) && page_exists(avp, aoffset)) {
7514 			/* A page exists for the anon slot */
7515 			ret |= SEG_PAGE_INCORE;
7516 
7517 			/*
7518 			 * If page is mapped and writable
7519 			 */
7520 			attr = (uint_t)0;
7521 			if ((hat_getattr(seg->s_as->a_hat, addr,
7522 			    &attr) != -1) && (attr & PROT_WRITE)) {
7523 				ret |= SEG_PAGE_ANON;
7524 			}
7525 			/*
7526 			 * Don't get page_struct lock for lckcnt and cowcnt,
7527 			 * since this is purely advisory.
7528 			 */
7529 			if ((pp = page_lookup_nowait(avp, aoffset,
7530 			    SE_SHARED)) != NULL) {
7531 				if (pp->p_lckcnt)
7532 					ret |= SEG_PAGE_SOFTLOCK;
7533 				if (pp->p_cowcnt)
7534 					ret |= SEG_PAGE_HASCOW;
7535 				page_unlock(pp);
7536 			}
7537 		}
7538 
7539 		/* Gather vnode statistics */
7540 		vp = svd->vp;
7541 		offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7542 
7543 		if (vp != NULL) {
7544 			/*
7545 			 * Try to obtain a "shared" lock on the page
7546 			 * without blocking.  If this fails, determine
7547 			 * if the page is in memory.
7548 			 */
7549 			pp = page_lookup_nowait(vp, offset, SE_SHARED);
7550 			if ((pp == NULL) && (page_exists(vp, offset))) {
7551 				/* Page is incore, and is named */
7552 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7553 			}
7554 			/*
7555 			 * Don't get page_struct lock for lckcnt and cowcnt,
7556 			 * since this is purely advisory.
7557 			 */
7558 			if (pp != NULL) {
7559 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7560 				if (pp->p_lckcnt)
7561 					ret |= SEG_PAGE_SOFTLOCK;
7562 				if (pp->p_cowcnt)
7563 					ret |= SEG_PAGE_HASCOW;
7564 				page_unlock(pp);
7565 			}
7566 		}
7567 
7568 		/* Gather virtual page information */
7569 		if (vpp) {
7570 			if (VPP_ISPPLOCK(vpp))
7571 				ret |= SEG_PAGE_LOCKED;
7572 			vpp++;
7573 		}
7574 
7575 		*vec++ = (char)ret;
7576 	}
7577 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7578 	return (len);
7579 }
7580 
7581 /*
7582  * Statement for p_cowcnts/p_lckcnts.
7583  *
7584  * p_cowcnt is updated while mlock/munlocking MAP_PRIVATE and PROT_WRITE region
7585  * irrespective of the following factors or anything else:
7586  *
7587  *	(1) anon slots are populated or not
7588  *	(2) cow is broken or not
7589  *	(3) refcnt on ap is 1 or greater than 1
7590  *
7591  * If it's not MAP_PRIVATE and PROT_WRITE, p_lckcnt is updated during mlock
7592  * and munlock.
7593  *
7594  *
7595  * Handling p_cowcnts/p_lckcnts during copy-on-write fault:
7596  *
7597  *	if vpage has PROT_WRITE
7598  *		transfer cowcnt on the oldpage -> cowcnt on the newpage
7599  *	else
7600  *		transfer lckcnt on the oldpage -> lckcnt on the newpage
7601  *
7602  *	During copy-on-write, decrement p_cowcnt on the oldpage and increment
7603  *	p_cowcnt on the newpage *if* the corresponding vpage has PROT_WRITE.
7604  *
7605  *	We may also break COW if softlocking on read access in the physio case.
7606  *	In this case, vpage may not have PROT_WRITE. So, we need to decrement
7607  *	p_lckcnt on the oldpage and increment p_lckcnt on the newpage *if* the
7608  *	vpage doesn't have PROT_WRITE.
7609  *
7610  *
7611  * Handling p_cowcnts/p_lckcnts during mprotect on mlocked region:
7612  *
7613  * 	If a MAP_PRIVATE region loses PROT_WRITE, we decrement p_cowcnt and
7614  *	increment p_lckcnt by calling page_subclaim() which takes care of
7615  * 	availrmem accounting and p_lckcnt overflow.
7616  *
7617  *	If a MAP_PRIVATE region gains PROT_WRITE, we decrement p_lckcnt and
7618  *	increment p_cowcnt by calling page_addclaim() which takes care of
7619  *	availrmem availability and p_cowcnt overflow.
7620  */
7621 
7622 /*
7623  * Lock down (or unlock) pages mapped by this segment.
7624  *
7625  * XXX only creates PAGESIZE pages if anon slots are not initialized.
7626  * At fault time they will be relocated into larger pages.
7627  */
7628 static int
7629 segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
7630     int attr, int op, ulong_t *lockmap, size_t pos)
7631 {
7632 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7633 	struct vpage *vpp;
7634 	struct vpage *evp;
7635 	page_t *pp;
7636 	u_offset_t offset;
7637 	u_offset_t off;
7638 	int segtype;
7639 	int pageprot;
7640 	int claim;
7641 	struct vnode *vp;
7642 	ulong_t anon_index;
7643 	struct anon_map *amp;
7644 	struct anon *ap;
7645 	struct vattr va;
7646 	anon_sync_obj_t cookie;
7647 	struct kshmid *sp = NULL;
7648 	struct proc	*p = curproc;
7649 	kproject_t	*proj = NULL;
7650 	int chargeproc = 1;
7651 	size_t locked_bytes = 0;
7652 	size_t unlocked_bytes = 0;
7653 	int err = 0;
7654 
7655 	/*
7656 	 * Hold write lock on address space because may split or concatenate
7657 	 * segments
7658 	 */
7659 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7660 
7661 	/*
7662 	 * If this is a shm, use shm's project and zone, else use
7663 	 * project and zone of calling process
7664 	 */
7665 
7666 	/* Determine if this segment backs a sysV shm */
7667 	if (svd->amp != NULL && svd->amp->a_sp != NULL) {
7668 		ASSERT(svd->type == MAP_SHARED);
7669 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
7670 		sp = svd->amp->a_sp;
7671 		proj = sp->shm_perm.ipc_proj;
7672 		chargeproc = 0;
7673 	}
7674 
7675 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7676 	if (attr) {
7677 		pageprot = attr & ~(SHARED|PRIVATE);
7678 		segtype = attr & SHARED ? MAP_SHARED : MAP_PRIVATE;
7679 
7680 		/*
7681 		 * We are done if the segment types don't match
7682 		 * or if we have segment level protections and
7683 		 * they don't match.
7684 		 */
7685 		if (svd->type != segtype) {
7686 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7687 			return (0);
7688 		}
7689 		if (svd->pageprot == 0 && svd->prot != pageprot) {
7690 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7691 			return (0);
7692 		}
7693 	}
7694 
7695 	if (op == MC_LOCK) {
7696 		if (svd->tr_state == SEGVN_TR_INIT) {
7697 			svd->tr_state = SEGVN_TR_OFF;
7698 		} else if (svd->tr_state == SEGVN_TR_ON) {
7699 			ASSERT(svd->amp != NULL);
7700 			segvn_textunrepl(seg, 0);
7701 			ASSERT(svd->amp == NULL &&
7702 			    svd->tr_state == SEGVN_TR_OFF);
7703 		}
7704 	}
7705 
7706 	/*
7707 	 * If we're locking, then we must create a vpage structure if
7708 	 * none exists.  If we're unlocking, then check to see if there
7709 	 * is a vpage --  if not, then we could not have locked anything.
7710 	 */
7711 
7712 	if ((vpp = svd->vpage) == NULL) {
7713 		if (op == MC_LOCK) {
7714 			segvn_vpage(seg);
7715 			if (svd->vpage == NULL) {
7716 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7717 				return (ENOMEM);
7718 			}
7719 		} else {
7720 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7721 			return (0);
7722 		}
7723 	}
7724 
7725 	/*
7726 	 * The anonymous data vector (i.e., previously
7727 	 * unreferenced mapping to swap space) can be allocated
7728 	 * by lazily testing for its existence.
7729 	 */
7730 	if (op == MC_LOCK && svd->amp == NULL && svd->vp == NULL) {
7731 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
7732 		svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
7733 		svd->amp->a_szc = seg->s_szc;
7734 	}
7735 
7736 	if ((amp = svd->amp) != NULL) {
7737 		anon_index = svd->anon_index + seg_page(seg, addr);
7738 	}
7739 
7740 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7741 	evp = &svd->vpage[seg_page(seg, addr + len)];
7742 
7743 	if (sp != NULL)
7744 		mutex_enter(&sp->shm_mlock);
7745 
7746 	/* determine number of unlocked bytes in range for lock operation */
7747 	if (op == MC_LOCK) {
7748 
7749 		if (sp == NULL) {
7750 			for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7751 			    vpp++) {
7752 				if (!VPP_ISPPLOCK(vpp))
7753 					unlocked_bytes += PAGESIZE;
7754 			}
7755 		} else {
7756 			ulong_t		i_idx, i_edx;
7757 			anon_sync_obj_t	i_cookie;
7758 			struct anon	*i_ap;
7759 			struct vnode	*i_vp;
7760 			u_offset_t	i_off;
7761 
7762 			/* Only count sysV pages once for locked memory */
7763 			i_edx = svd->anon_index + seg_page(seg, addr + len);
7764 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7765 			for (i_idx = anon_index; i_idx < i_edx; i_idx++) {
7766 				anon_array_enter(amp, i_idx, &i_cookie);
7767 				i_ap = anon_get_ptr(amp->ahp, i_idx);
7768 				if (i_ap == NULL) {
7769 					unlocked_bytes += PAGESIZE;
7770 					anon_array_exit(&i_cookie);
7771 					continue;
7772 				}
7773 				swap_xlate(i_ap, &i_vp, &i_off);
7774 				anon_array_exit(&i_cookie);
7775 				pp = page_lookup(i_vp, i_off, SE_SHARED);
7776 				if (pp == NULL) {
7777 					unlocked_bytes += PAGESIZE;
7778 					continue;
7779 				} else if (pp->p_lckcnt == 0)
7780 					unlocked_bytes += PAGESIZE;
7781 				page_unlock(pp);
7782 			}
7783 			ANON_LOCK_EXIT(&amp->a_rwlock);
7784 		}
7785 
7786 		mutex_enter(&p->p_lock);
7787 		err = rctl_incr_locked_mem(p, proj, unlocked_bytes,
7788 		    chargeproc);
7789 		mutex_exit(&p->p_lock);
7790 
7791 		if (err) {
7792 			if (sp != NULL)
7793 				mutex_exit(&sp->shm_mlock);
7794 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7795 			return (err);
7796 		}
7797 	}
7798 	/*
7799 	 * Loop over all pages in the range.  Process if we're locking and
7800 	 * page has not already been locked in this mapping; or if we're
7801 	 * unlocking and the page has been locked.
7802 	 */
7803 	for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7804 	    vpp++, pos++, addr += PAGESIZE, offset += PAGESIZE, anon_index++) {
7805 		if ((attr == 0 || VPP_PROT(vpp) == pageprot) &&
7806 		    ((op == MC_LOCK && !VPP_ISPPLOCK(vpp)) ||
7807 		    (op == MC_UNLOCK && VPP_ISPPLOCK(vpp)))) {
7808 
7809 			if (amp != NULL)
7810 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7811 			/*
7812 			 * If this isn't a MAP_NORESERVE segment and
7813 			 * we're locking, allocate anon slots if they
7814 			 * don't exist.  The page is brought in later on.
7815 			 */
7816 			if (op == MC_LOCK && svd->vp == NULL &&
7817 			    ((svd->flags & MAP_NORESERVE) == 0) &&
7818 			    amp != NULL &&
7819 			    ((ap = anon_get_ptr(amp->ahp, anon_index))
7820 			    == NULL)) {
7821 				anon_array_enter(amp, anon_index, &cookie);
7822 
7823 				if ((ap = anon_get_ptr(amp->ahp,
7824 				    anon_index)) == NULL) {
7825 					pp = anon_zero(seg, addr, &ap,
7826 					    svd->cred);
7827 					if (pp == NULL) {
7828 						anon_array_exit(&cookie);
7829 						ANON_LOCK_EXIT(&amp->a_rwlock);
7830 						err = ENOMEM;
7831 						goto out;
7832 					}
7833 					ASSERT(anon_get_ptr(amp->ahp,
7834 					    anon_index) == NULL);
7835 					(void) anon_set_ptr(amp->ahp,
7836 					    anon_index, ap, ANON_SLEEP);
7837 					page_unlock(pp);
7838 				}
7839 				anon_array_exit(&cookie);
7840 			}
7841 
7842 			/*
7843 			 * Get name for page, accounting for
7844 			 * existence of private copy.
7845 			 */
7846 			ap = NULL;
7847 			if (amp != NULL) {
7848 				anon_array_enter(amp, anon_index, &cookie);
7849 				ap = anon_get_ptr(amp->ahp, anon_index);
7850 				if (ap != NULL) {
7851 					swap_xlate(ap, &vp, &off);
7852 				} else {
7853 					if (svd->vp == NULL &&
7854 					    (svd->flags & MAP_NORESERVE)) {
7855 						anon_array_exit(&cookie);
7856 						ANON_LOCK_EXIT(&amp->a_rwlock);
7857 						continue;
7858 					}
7859 					vp = svd->vp;
7860 					off = offset;
7861 				}
7862 				if (op != MC_LOCK || ap == NULL) {
7863 					anon_array_exit(&cookie);
7864 					ANON_LOCK_EXIT(&amp->a_rwlock);
7865 				}
7866 			} else {
7867 				vp = svd->vp;
7868 				off = offset;
7869 			}
7870 
7871 			/*
7872 			 * Get page frame.  It's ok if the page is
7873 			 * not available when we're unlocking, as this
7874 			 * may simply mean that a page we locked got
7875 			 * truncated out of existence after we locked it.
7876 			 *
7877 			 * Invoke VOP_GETPAGE() to obtain the page struct
7878 			 * since we may need to read it from disk if its
7879 			 * been paged out.
7880 			 */
7881 			if (op != MC_LOCK)
7882 				pp = page_lookup(vp, off, SE_SHARED);
7883 			else {
7884 				page_t *pl[1 + 1];
7885 				int error;
7886 
7887 				ASSERT(vp != NULL);
7888 
7889 				error = VOP_GETPAGE(vp, (offset_t)off, PAGESIZE,
7890 				    (uint_t *)NULL, pl, PAGESIZE, seg, addr,
7891 				    S_OTHER, svd->cred, NULL);
7892 
7893 				if (error && ap != NULL) {
7894 					anon_array_exit(&cookie);
7895 					ANON_LOCK_EXIT(&amp->a_rwlock);
7896 				}
7897 
7898 				/*
7899 				 * If the error is EDEADLK then we must bounce
7900 				 * up and drop all vm subsystem locks and then
7901 				 * retry the operation later
7902 				 * This behavior is a temporary measure because
7903 				 * ufs/sds logging is badly designed and will
7904 				 * deadlock if we don't allow this bounce to
7905 				 * happen.  The real solution is to re-design
7906 				 * the logging code to work properly.  See bug
7907 				 * 4125102 for details of the problem.
7908 				 */
7909 				if (error == EDEADLK) {
7910 					err = error;
7911 					goto out;
7912 				}
7913 				/*
7914 				 * Quit if we fail to fault in the page.  Treat
7915 				 * the failure as an error, unless the addr
7916 				 * is mapped beyond the end of a file.
7917 				 */
7918 				if (error && svd->vp) {
7919 					va.va_mask = AT_SIZE;
7920 					if (VOP_GETATTR(svd->vp, &va, 0,
7921 					    svd->cred, NULL) != 0) {
7922 						err = EIO;
7923 						goto out;
7924 					}
7925 					if (btopr(va.va_size) >=
7926 					    btopr(off + 1)) {
7927 						err = EIO;
7928 						goto out;
7929 					}
7930 					goto out;
7931 
7932 				} else if (error) {
7933 					err = EIO;
7934 					goto out;
7935 				}
7936 				pp = pl[0];
7937 				ASSERT(pp != NULL);
7938 			}
7939 
7940 			/*
7941 			 * See Statement at the beginning of this routine.
7942 			 *
7943 			 * claim is always set if MAP_PRIVATE and PROT_WRITE
7944 			 * irrespective of following factors:
7945 			 *
7946 			 * (1) anon slots are populated or not
7947 			 * (2) cow is broken or not
7948 			 * (3) refcnt on ap is 1 or greater than 1
7949 			 *
7950 			 * See 4140683 for details
7951 			 */
7952 			claim = ((VPP_PROT(vpp) & PROT_WRITE) &&
7953 			    (svd->type == MAP_PRIVATE));
7954 
7955 			/*
7956 			 * Perform page-level operation appropriate to
7957 			 * operation.  If locking, undo the SOFTLOCK
7958 			 * performed to bring the page into memory
7959 			 * after setting the lock.  If unlocking,
7960 			 * and no page was found, account for the claim
7961 			 * separately.
7962 			 */
7963 			if (op == MC_LOCK) {
7964 				int ret = 1;	/* Assume success */
7965 
7966 				ASSERT(!VPP_ISPPLOCK(vpp));
7967 
7968 				ret = page_pp_lock(pp, claim, 0);
7969 				if (ap != NULL) {
7970 					if (ap->an_pvp != NULL) {
7971 						anon_swap_free(ap, pp);
7972 					}
7973 					anon_array_exit(&cookie);
7974 					ANON_LOCK_EXIT(&amp->a_rwlock);
7975 				}
7976 				if (ret == 0) {
7977 					/* locking page failed */
7978 					page_unlock(pp);
7979 					err = EAGAIN;
7980 					goto out;
7981 				}
7982 				VPP_SETPPLOCK(vpp);
7983 				if (sp != NULL) {
7984 					if (pp->p_lckcnt == 1)
7985 						locked_bytes += PAGESIZE;
7986 				} else
7987 					locked_bytes += PAGESIZE;
7988 
7989 				if (lockmap != (ulong_t *)NULL)
7990 					BT_SET(lockmap, pos);
7991 
7992 				page_unlock(pp);
7993 			} else {
7994 				ASSERT(VPP_ISPPLOCK(vpp));
7995 				if (pp != NULL) {
7996 					/* sysV pages should be locked */
7997 					ASSERT(sp == NULL || pp->p_lckcnt > 0);
7998 					page_pp_unlock(pp, claim, 0);
7999 					if (sp != NULL) {
8000 						if (pp->p_lckcnt == 0)
8001 							unlocked_bytes
8002 							    += PAGESIZE;
8003 					} else
8004 						unlocked_bytes += PAGESIZE;
8005 					page_unlock(pp);
8006 				} else {
8007 					ASSERT(sp == NULL);
8008 					unlocked_bytes += PAGESIZE;
8009 				}
8010 				VPP_CLRPPLOCK(vpp);
8011 			}
8012 		}
8013 	}
8014 out:
8015 	if (op == MC_LOCK) {
8016 		/* Credit back bytes that did not get locked */
8017 		if ((unlocked_bytes - locked_bytes) > 0) {
8018 			if (proj == NULL)
8019 				mutex_enter(&p->p_lock);
8020 			rctl_decr_locked_mem(p, proj,
8021 			    (unlocked_bytes - locked_bytes), chargeproc);
8022 			if (proj == NULL)
8023 				mutex_exit(&p->p_lock);
8024 		}
8025 
8026 	} else {
8027 		/* Account bytes that were unlocked */
8028 		if (unlocked_bytes > 0) {
8029 			if (proj == NULL)
8030 				mutex_enter(&p->p_lock);
8031 			rctl_decr_locked_mem(p, proj, unlocked_bytes,
8032 			    chargeproc);
8033 			if (proj == NULL)
8034 				mutex_exit(&p->p_lock);
8035 		}
8036 	}
8037 	if (sp != NULL)
8038 		mutex_exit(&sp->shm_mlock);
8039 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8040 
8041 	return (err);
8042 }
8043 
8044 /*
8045  * Set advice from user for specified pages
8046  * There are 9 types of advice:
8047  *	MADV_NORMAL	- Normal (default) behavior (whatever that is)
8048  *	MADV_RANDOM	- Random page references
8049  *				do not allow readahead or 'klustering'
8050  *	MADV_SEQUENTIAL	- Sequential page references
8051  *				Pages previous to the one currently being
8052  *				accessed (determined by fault) are 'not needed'
8053  *				and are freed immediately
8054  *	MADV_WILLNEED	- Pages are likely to be used (fault ahead in mctl)
8055  *	MADV_DONTNEED	- Pages are not needed (synced out in mctl)
8056  *	MADV_FREE	- Contents can be discarded
8057  *	MADV_ACCESS_DEFAULT- Default access
8058  *	MADV_ACCESS_LWP	- Next LWP will access heavily
8059  *	MADV_ACCESS_MANY- Many LWPs or processes will access heavily
8060  */
8061 static int
8062 segvn_advise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
8063 {
8064 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8065 	size_t page;
8066 	int err = 0;
8067 	int already_set;
8068 	struct anon_map *amp;
8069 	ulong_t anon_index;
8070 	struct seg *next;
8071 	lgrp_mem_policy_t policy;
8072 	struct seg *prev;
8073 	struct vnode *vp;
8074 
8075 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
8076 
8077 	/*
8078 	 * In case of MADV_FREE, we won't be modifying any segment private
8079 	 * data structures; so, we only need to grab READER's lock
8080 	 */
8081 	if (behav != MADV_FREE) {
8082 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
8083 		if (svd->tr_state != SEGVN_TR_OFF) {
8084 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8085 			return (0);
8086 		}
8087 	} else {
8088 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8089 	}
8090 
8091 	/*
8092 	 * Large pages are assumed to be only turned on when accesses to the
8093 	 * segment's address range have spatial and temporal locality. That
8094 	 * justifies ignoring MADV_SEQUENTIAL for large page segments.
8095 	 * Also, ignore advice affecting lgroup memory allocation
8096 	 * if don't need to do lgroup optimizations on this system
8097 	 */
8098 
8099 	if ((behav == MADV_SEQUENTIAL &&
8100 	    (seg->s_szc != 0 || HAT_IS_REGION_COOKIE_VALID(svd->rcookie))) ||
8101 	    (!lgrp_optimizations() && (behav == MADV_ACCESS_DEFAULT ||
8102 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY))) {
8103 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8104 		return (0);
8105 	}
8106 
8107 	if (behav == MADV_SEQUENTIAL || behav == MADV_ACCESS_DEFAULT ||
8108 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY) {
8109 		/*
8110 		 * Since we are going to unload hat mappings
8111 		 * we first have to flush the cache. Otherwise
8112 		 * this might lead to system panic if another
8113 		 * thread is doing physio on the range whose
8114 		 * mappings are unloaded by madvise(3C).
8115 		 */
8116 		if (svd->softlockcnt > 0) {
8117 			/*
8118 			 * If this is shared segment non 0 softlockcnt
8119 			 * means locked pages are still in use.
8120 			 */
8121 			if (svd->type == MAP_SHARED) {
8122 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8123 				return (EAGAIN);
8124 			}
8125 			/*
8126 			 * Since we do have the segvn writers lock
8127 			 * nobody can fill the cache with entries
8128 			 * belonging to this seg during the purge.
8129 			 * The flush either succeeds or we still
8130 			 * have pending I/Os. In the later case,
8131 			 * madvise(3C) fails.
8132 			 */
8133 			segvn_purge(seg);
8134 			if (svd->softlockcnt > 0) {
8135 				/*
8136 				 * Since madvise(3C) is advisory and
8137 				 * it's not part of UNIX98, madvise(3C)
8138 				 * failure here doesn't cause any hardship.
8139 				 * Note that we don't block in "as" layer.
8140 				 */
8141 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8142 				return (EAGAIN);
8143 			}
8144 		} else if (svd->type == MAP_SHARED && svd->amp != NULL &&
8145 		    svd->amp->a_softlockcnt > 0) {
8146 			/*
8147 			 * Try to purge this amp's entries from pcache. It
8148 			 * will succeed only if other segments that share the
8149 			 * amp have no outstanding softlock's.
8150 			 */
8151 			segvn_purge(seg);
8152 		}
8153 	}
8154 
8155 	amp = svd->amp;
8156 	vp = svd->vp;
8157 	if (behav == MADV_FREE) {
8158 		/*
8159 		 * MADV_FREE is not supported for segments with
8160 		 * underlying object; if anonmap is NULL, anon slots
8161 		 * are not yet populated and there is nothing for
8162 		 * us to do. As MADV_FREE is advisory, we don't
8163 		 * return error in either case.
8164 		 */
8165 		if (vp != NULL || amp == NULL) {
8166 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8167 			return (0);
8168 		}
8169 
8170 		segvn_purge(seg);
8171 
8172 		page = seg_page(seg, addr);
8173 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8174 		anon_disclaim(amp, svd->anon_index + page, len);
8175 		ANON_LOCK_EXIT(&amp->a_rwlock);
8176 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8177 		return (0);
8178 	}
8179 
8180 	/*
8181 	 * If advice is to be applied to entire segment,
8182 	 * use advice field in seg_data structure
8183 	 * otherwise use appropriate vpage entry.
8184 	 */
8185 	if ((addr == seg->s_base) && (len == seg->s_size)) {
8186 		switch (behav) {
8187 		case MADV_ACCESS_LWP:
8188 		case MADV_ACCESS_MANY:
8189 		case MADV_ACCESS_DEFAULT:
8190 			/*
8191 			 * Set memory allocation policy for this segment
8192 			 */
8193 			policy = lgrp_madv_to_policy(behav, len, svd->type);
8194 			if (svd->type == MAP_SHARED)
8195 				already_set = lgrp_shm_policy_set(policy, amp,
8196 				    svd->anon_index, vp, svd->offset, len);
8197 			else {
8198 				/*
8199 				 * For private memory, need writers lock on
8200 				 * address space because the segment may be
8201 				 * split or concatenated when changing policy
8202 				 */
8203 				if (AS_READ_HELD(seg->s_as)) {
8204 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8205 					return (IE_RETRY);
8206 				}
8207 
8208 				already_set = lgrp_privm_policy_set(policy,
8209 				    &svd->policy_info, len);
8210 			}
8211 
8212 			/*
8213 			 * If policy set already and it shouldn't be reapplied,
8214 			 * don't do anything.
8215 			 */
8216 			if (already_set &&
8217 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8218 				break;
8219 
8220 			/*
8221 			 * Mark any existing pages in given range for
8222 			 * migration
8223 			 */
8224 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8225 			    vp, svd->offset, 1);
8226 
8227 			/*
8228 			 * If same policy set already or this is a shared
8229 			 * memory segment, don't need to try to concatenate
8230 			 * segment with adjacent ones.
8231 			 */
8232 			if (already_set || svd->type == MAP_SHARED)
8233 				break;
8234 
8235 			/*
8236 			 * Try to concatenate this segment with previous
8237 			 * one and next one, since we changed policy for
8238 			 * this one and it may be compatible with adjacent
8239 			 * ones now.
8240 			 */
8241 			prev = AS_SEGPREV(seg->s_as, seg);
8242 			next = AS_SEGNEXT(seg->s_as, seg);
8243 
8244 			if (next && next->s_ops == &segvn_ops &&
8245 			    addr + len == next->s_base)
8246 				(void) segvn_concat(seg, next, 1);
8247 
8248 			if (prev && prev->s_ops == &segvn_ops &&
8249 			    addr == prev->s_base + prev->s_size) {
8250 				/*
8251 				 * Drop lock for private data of current
8252 				 * segment before concatenating (deleting) it
8253 				 * and return IE_REATTACH to tell as_ctl() that
8254 				 * current segment has changed
8255 				 */
8256 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8257 				if (!segvn_concat(prev, seg, 1))
8258 					err = IE_REATTACH;
8259 
8260 				return (err);
8261 			}
8262 			break;
8263 
8264 		case MADV_SEQUENTIAL:
8265 			/*
8266 			 * unloading mapping guarantees
8267 			 * detection in segvn_fault
8268 			 */
8269 			ASSERT(seg->s_szc == 0);
8270 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8271 			hat_unload(seg->s_as->a_hat, addr, len,
8272 			    HAT_UNLOAD);
8273 			/* FALLTHROUGH */
8274 		case MADV_NORMAL:
8275 		case MADV_RANDOM:
8276 			svd->advice = (uchar_t)behav;
8277 			svd->pageadvice = 0;
8278 			break;
8279 		case MADV_WILLNEED:	/* handled in memcntl */
8280 		case MADV_DONTNEED:	/* handled in memcntl */
8281 		case MADV_FREE:		/* handled above */
8282 			break;
8283 		default:
8284 			err = EINVAL;
8285 		}
8286 	} else {
8287 		caddr_t			eaddr;
8288 		struct seg		*new_seg;
8289 		struct segvn_data	*new_svd;
8290 		u_offset_t		off;
8291 		caddr_t			oldeaddr;
8292 
8293 		page = seg_page(seg, addr);
8294 
8295 		segvn_vpage(seg);
8296 		if (svd->vpage == NULL) {
8297 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8298 			return (ENOMEM);
8299 		}
8300 
8301 		switch (behav) {
8302 			struct vpage *bvpp, *evpp;
8303 
8304 		case MADV_ACCESS_LWP:
8305 		case MADV_ACCESS_MANY:
8306 		case MADV_ACCESS_DEFAULT:
8307 			/*
8308 			 * Set memory allocation policy for portion of this
8309 			 * segment
8310 			 */
8311 
8312 			/*
8313 			 * Align address and length of advice to page
8314 			 * boundaries for large pages
8315 			 */
8316 			if (seg->s_szc != 0) {
8317 				size_t	pgsz;
8318 
8319 				pgsz = page_get_pagesize(seg->s_szc);
8320 				addr = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
8321 				len = P2ROUNDUP(len, pgsz);
8322 			}
8323 
8324 			/*
8325 			 * Check to see whether policy is set already
8326 			 */
8327 			policy = lgrp_madv_to_policy(behav, len, svd->type);
8328 
8329 			anon_index = svd->anon_index + page;
8330 			off = svd->offset + (uintptr_t)(addr - seg->s_base);
8331 
8332 			if (svd->type == MAP_SHARED)
8333 				already_set = lgrp_shm_policy_set(policy, amp,
8334 				    anon_index, vp, off, len);
8335 			else
8336 				already_set =
8337 				    (policy == svd->policy_info.mem_policy);
8338 
8339 			/*
8340 			 * If policy set already and it shouldn't be reapplied,
8341 			 * don't do anything.
8342 			 */
8343 			if (already_set &&
8344 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8345 				break;
8346 
8347 			/*
8348 			 * For private memory, need writers lock on
8349 			 * address space because the segment may be
8350 			 * split or concatenated when changing policy
8351 			 */
8352 			if (svd->type == MAP_PRIVATE &&
8353 			    AS_READ_HELD(seg->s_as)) {
8354 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8355 				return (IE_RETRY);
8356 			}
8357 
8358 			/*
8359 			 * Mark any existing pages in given range for
8360 			 * migration
8361 			 */
8362 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8363 			    vp, svd->offset, 1);
8364 
8365 			/*
8366 			 * Don't need to try to split or concatenate
8367 			 * segments, since policy is same or this is a shared
8368 			 * memory segment
8369 			 */
8370 			if (already_set || svd->type == MAP_SHARED)
8371 				break;
8372 
8373 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
8374 				ASSERT(svd->amp == NULL);
8375 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
8376 				ASSERT(svd->softlockcnt == 0);
8377 				hat_leave_region(seg->s_as->a_hat, svd->rcookie,
8378 				    HAT_REGION_TEXT);
8379 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
8380 			}
8381 
8382 			/*
8383 			 * Split off new segment if advice only applies to a
8384 			 * portion of existing segment starting in middle
8385 			 */
8386 			new_seg = NULL;
8387 			eaddr = addr + len;
8388 			oldeaddr = seg->s_base + seg->s_size;
8389 			if (addr > seg->s_base) {
8390 				/*
8391 				 * Must flush I/O page cache
8392 				 * before splitting segment
8393 				 */
8394 				if (svd->softlockcnt > 0)
8395 					segvn_purge(seg);
8396 
8397 				/*
8398 				 * Split segment and return IE_REATTACH to tell
8399 				 * as_ctl() that current segment changed
8400 				 */
8401 				new_seg = segvn_split_seg(seg, addr);
8402 				new_svd = (struct segvn_data *)new_seg->s_data;
8403 				err = IE_REATTACH;
8404 
8405 				/*
8406 				 * If new segment ends where old one
8407 				 * did, try to concatenate the new
8408 				 * segment with next one.
8409 				 */
8410 				if (eaddr == oldeaddr) {
8411 					/*
8412 					 * Set policy for new segment
8413 					 */
8414 					(void) lgrp_privm_policy_set(policy,
8415 					    &new_svd->policy_info,
8416 					    new_seg->s_size);
8417 
8418 					next = AS_SEGNEXT(new_seg->s_as,
8419 					    new_seg);
8420 
8421 					if (next &&
8422 					    next->s_ops == &segvn_ops &&
8423 					    eaddr == next->s_base)
8424 						(void) segvn_concat(new_seg,
8425 						    next, 1);
8426 				}
8427 			}
8428 
8429 			/*
8430 			 * Split off end of existing segment if advice only
8431 			 * applies to a portion of segment ending before
8432 			 * end of the existing segment
8433 			 */
8434 			if (eaddr < oldeaddr) {
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 				 * If beginning of old segment was already
8444 				 * split off, use new segment to split end off
8445 				 * from.
8446 				 */
8447 				if (new_seg != NULL && new_seg != seg) {
8448 					/*
8449 					 * Split segment
8450 					 */
8451 					(void) segvn_split_seg(new_seg, eaddr);
8452 
8453 					/*
8454 					 * Set policy for new segment
8455 					 */
8456 					(void) lgrp_privm_policy_set(policy,
8457 					    &new_svd->policy_info,
8458 					    new_seg->s_size);
8459 				} else {
8460 					/*
8461 					 * Split segment and return IE_REATTACH
8462 					 * to tell as_ctl() that current
8463 					 * segment changed
8464 					 */
8465 					(void) segvn_split_seg(seg, eaddr);
8466 					err = IE_REATTACH;
8467 
8468 					(void) lgrp_privm_policy_set(policy,
8469 					    &svd->policy_info, seg->s_size);
8470 
8471 					/*
8472 					 * If new segment starts where old one
8473 					 * did, try to concatenate it with
8474 					 * previous segment.
8475 					 */
8476 					if (addr == seg->s_base) {
8477 						prev = AS_SEGPREV(seg->s_as,
8478 						    seg);
8479 
8480 						/*
8481 						 * Drop lock for private data
8482 						 * of current segment before
8483 						 * concatenating (deleting) it
8484 						 */
8485 						if (prev &&
8486 						    prev->s_ops ==
8487 						    &segvn_ops &&
8488 						    addr == prev->s_base +
8489 						    prev->s_size) {
8490 							SEGVN_LOCK_EXIT(
8491 							    seg->s_as,
8492 							    &svd->lock);
8493 							(void) segvn_concat(
8494 							    prev, seg, 1);
8495 							return (err);
8496 						}
8497 					}
8498 				}
8499 			}
8500 			break;
8501 		case MADV_SEQUENTIAL:
8502 			ASSERT(seg->s_szc == 0);
8503 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8504 			hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
8505 			/* FALLTHROUGH */
8506 		case MADV_NORMAL:
8507 		case MADV_RANDOM:
8508 			bvpp = &svd->vpage[page];
8509 			evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8510 			for (; bvpp < evpp; bvpp++)
8511 				VPP_SETADVICE(bvpp, behav);
8512 			svd->advice = MADV_NORMAL;
8513 			break;
8514 		case MADV_WILLNEED:	/* handled in memcntl */
8515 		case MADV_DONTNEED:	/* handled in memcntl */
8516 		case MADV_FREE:		/* handled above */
8517 			break;
8518 		default:
8519 			err = EINVAL;
8520 		}
8521 	}
8522 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8523 	return (err);
8524 }
8525 
8526 /*
8527  * There is one kind of inheritance that can be specified for pages:
8528  *
8529  *     SEGP_INH_ZERO - Pages should be zeroed in the child
8530  */
8531 static int
8532 segvn_inherit(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
8533 {
8534 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8535 	struct vpage *bvpp, *evpp;
8536 	size_t page;
8537 	int ret = 0;
8538 
8539 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
8540 
8541 	/* Can't support something we don't know about */
8542 	if (behav != SEGP_INH_ZERO)
8543 		return (ENOTSUP);
8544 
8545 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
8546 
8547 	/*
8548 	 * This must be a straightforward anonymous segment that is mapped
8549 	 * privately and is not backed by a vnode.
8550 	 */
8551 	if (svd->tr_state != SEGVN_TR_OFF ||
8552 	    svd->type != MAP_PRIVATE ||
8553 	    svd->vp != NULL) {
8554 		ret = EINVAL;
8555 		goto out;
8556 	}
8557 
8558 	/*
8559 	 * If the entire segment has been marked as inherit zero, then no reason
8560 	 * to do anything else.
8561 	 */
8562 	if (svd->svn_inz == SEGVN_INZ_ALL) {
8563 		ret = 0;
8564 		goto out;
8565 	}
8566 
8567 	/*
8568 	 * If this applies to the entire segment, simply mark it and we're done.
8569 	 */
8570 	if ((addr == seg->s_base) && (len == seg->s_size)) {
8571 		svd->svn_inz = SEGVN_INZ_ALL;
8572 		ret = 0;
8573 		goto out;
8574 	}
8575 
8576 	/*
8577 	 * We've been asked to mark a subset of this segment as inherit zero,
8578 	 * therefore we need to mainpulate its vpages.
8579 	 */
8580 	if (svd->vpage == NULL) {
8581 		segvn_vpage(seg);
8582 		if (svd->vpage == NULL) {
8583 			ret = ENOMEM;
8584 			goto out;
8585 		}
8586 	}
8587 
8588 	svd->svn_inz = SEGVN_INZ_VPP;
8589 	page = seg_page(seg, addr);
8590 	bvpp = &svd->vpage[page];
8591 	evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8592 	for (; bvpp < evpp; bvpp++)
8593 		VPP_SETINHZERO(bvpp);
8594 	ret = 0;
8595 
8596 out:
8597 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8598 	return (ret);
8599 }
8600 
8601 /*
8602  * Create a vpage structure for this seg.
8603  */
8604 static void
8605 segvn_vpage(struct seg *seg)
8606 {
8607 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8608 	struct vpage *vp, *evp;
8609 	static pgcnt_t page_limit = 0;
8610 
8611 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
8612 
8613 	/*
8614 	 * If no vpage structure exists, allocate one.  Copy the protections
8615 	 * and the advice from the segment itself to the individual pages.
8616 	 */
8617 	if (svd->vpage == NULL) {
8618 		/*
8619 		 * Start by calculating the number of pages we must allocate to
8620 		 * track the per-page vpage structs needs for this entire
8621 		 * segment. If we know now that it will require more than our
8622 		 * heuristic for the maximum amount of kmem we can consume then
8623 		 * fail. We do this here, instead of trying to detect this deep
8624 		 * in page_resv and propagating the error up, since the entire
8625 		 * memory allocation stack is not amenable to passing this
8626 		 * back. Instead, it wants to keep trying.
8627 		 *
8628 		 * As a heuristic we set a page limit of 5/8s of total_pages
8629 		 * for this allocation. We use shifts so that no floating
8630 		 * point conversion takes place and only need to do the
8631 		 * calculation once.
8632 		 */
8633 		ulong_t mem_needed = seg_pages(seg) * sizeof (struct vpage);
8634 		pgcnt_t npages = mem_needed >> PAGESHIFT;
8635 
8636 		if (page_limit == 0)
8637 			page_limit = (total_pages >> 1) + (total_pages >> 3);
8638 
8639 		if (npages > page_limit)
8640 			return;
8641 
8642 		svd->pageadvice = 1;
8643 		svd->vpage = kmem_zalloc(mem_needed, KM_SLEEP);
8644 		evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
8645 		for (vp = svd->vpage; vp < evp; vp++) {
8646 			VPP_SETPROT(vp, svd->prot);
8647 			VPP_SETADVICE(vp, svd->advice);
8648 		}
8649 	}
8650 }
8651 
8652 /*
8653  * Dump the pages belonging to this segvn segment.
8654  */
8655 static void
8656 segvn_dump(struct seg *seg)
8657 {
8658 	struct segvn_data *svd;
8659 	page_t *pp;
8660 	struct anon_map *amp;
8661 	ulong_t	anon_index;
8662 	struct vnode *vp;
8663 	u_offset_t off, offset;
8664 	pfn_t pfn;
8665 	pgcnt_t page, npages;
8666 	caddr_t addr;
8667 
8668 	npages = seg_pages(seg);
8669 	svd = (struct segvn_data *)seg->s_data;
8670 	vp = svd->vp;
8671 	off = offset = svd->offset;
8672 	addr = seg->s_base;
8673 
8674 	if ((amp = svd->amp) != NULL) {
8675 		anon_index = svd->anon_index;
8676 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8677 	}
8678 
8679 	for (page = 0; page < npages; page++, offset += PAGESIZE) {
8680 		struct anon *ap;
8681 		int we_own_it = 0;
8682 
8683 		if (amp && (ap = anon_get_ptr(svd->amp->ahp, anon_index++))) {
8684 			swap_xlate_nopanic(ap, &vp, &off);
8685 		} else {
8686 			vp = svd->vp;
8687 			off = offset;
8688 		}
8689 
8690 		/*
8691 		 * If pp == NULL, the page either does not exist
8692 		 * or is exclusively locked.  So determine if it
8693 		 * exists before searching for it.
8694 		 */
8695 
8696 		if ((pp = page_lookup_nowait(vp, off, SE_SHARED)))
8697 			we_own_it = 1;
8698 		else
8699 			pp = page_exists(vp, off);
8700 
8701 		if (pp) {
8702 			pfn = page_pptonum(pp);
8703 			dump_addpage(seg->s_as, addr, pfn);
8704 			if (we_own_it)
8705 				page_unlock(pp);
8706 		}
8707 		addr += PAGESIZE;
8708 		dump_timeleft = dump_timeout;
8709 	}
8710 
8711 	if (amp != NULL)
8712 		ANON_LOCK_EXIT(&amp->a_rwlock);
8713 }
8714 
8715 #ifdef DEBUG
8716 static uint32_t segvn_pglock_mtbf = 0;
8717 #endif
8718 
8719 #define	PCACHE_SHWLIST		((page_t *)-2)
8720 #define	NOPCACHE_SHWLIST	((page_t *)-1)
8721 
8722 /*
8723  * Lock/Unlock anon pages over a given range. Return shadow list. This routine
8724  * uses global segment pcache to cache shadow lists (i.e. pp arrays) of pages
8725  * to avoid the overhead of per page locking, unlocking for subsequent IOs to
8726  * the same parts of the segment. Currently shadow list creation is only
8727  * supported for pure anon segments. MAP_PRIVATE segment pcache entries are
8728  * tagged with segment pointer, starting virtual address and length. This
8729  * approach for MAP_SHARED segments may add many pcache entries for the same
8730  * set of pages and lead to long hash chains that decrease pcache lookup
8731  * performance. To avoid this issue for shared segments shared anon map and
8732  * starting anon index are used for pcache entry tagging. This allows all
8733  * segments to share pcache entries for the same anon range and reduces pcache
8734  * chain's length as well as memory overhead from duplicate shadow lists and
8735  * pcache entries.
8736  *
8737  * softlockcnt field in segvn_data structure counts the number of F_SOFTLOCK'd
8738  * pages via segvn_fault() and pagelock'd pages via this routine. But pagelock
8739  * part of softlockcnt accounting is done differently for private and shared
8740  * segments. In private segment case softlock is only incremented when a new
8741  * shadow list is created but not when an existing one is found via
8742  * seg_plookup(). pcache entries have reference count incremented/decremented
8743  * by each seg_plookup()/seg_pinactive() operation. Only entries that have 0
8744  * reference count can be purged (and purging is needed before segment can be
8745  * freed). When a private segment pcache entry is purged segvn_reclaim() will
8746  * decrement softlockcnt. Since in private segment case each of its pcache
8747  * entries only belongs to this segment we can expect that when
8748  * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8749  * segment purge will succeed and softlockcnt will drop to 0. In shared
8750  * segment case reference count in pcache entry counts active locks from many
8751  * different segments so we can't expect segment purging to succeed even when
8752  * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8753  * segment. To be able to determine when there're no pending pagelocks in
8754  * shared segment case we don't rely on purging to make softlockcnt drop to 0
8755  * but instead softlockcnt is incremented and decremented for every
8756  * segvn_pagelock(L_PAGELOCK/L_PAGEUNLOCK) call regardless if a new shadow
8757  * list was created or an existing one was found. When softlockcnt drops to 0
8758  * this segment no longer has any claims for pcached shadow lists and the
8759  * segment can be freed even if there're still active pcache entries
8760  * shared by this segment anon map. Shared segment pcache entries belong to
8761  * anon map and are typically removed when anon map is freed after all
8762  * processes destroy the segments that use this anon map.
8763  */
8764 static int
8765 segvn_pagelock(struct seg *seg, caddr_t addr, size_t len, struct page ***ppp,
8766     enum lock_type type, enum seg_rw rw)
8767 {
8768 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8769 	size_t np;
8770 	pgcnt_t adjustpages;
8771 	pgcnt_t npages;
8772 	ulong_t anon_index;
8773 	uint_t protchk = (rw == S_READ) ? PROT_READ : PROT_WRITE;
8774 	uint_t error;
8775 	struct anon_map *amp;
8776 	pgcnt_t anpgcnt;
8777 	struct page **pplist, **pl, *pp;
8778 	caddr_t a;
8779 	size_t page;
8780 	caddr_t lpgaddr, lpgeaddr;
8781 	anon_sync_obj_t cookie;
8782 	int anlock;
8783 	struct anon_map *pamp;
8784 	caddr_t paddr;
8785 	seg_preclaim_cbfunc_t preclaim_callback;
8786 	size_t pgsz;
8787 	int use_pcache;
8788 	size_t wlen;
8789 	uint_t pflags = 0;
8790 	int sftlck_sbase = 0;
8791 	int sftlck_send = 0;
8792 
8793 #ifdef DEBUG
8794 	if (type == L_PAGELOCK && segvn_pglock_mtbf) {
8795 		hrtime_t ts = gethrtime();
8796 		if ((ts % segvn_pglock_mtbf) == 0) {
8797 			return (ENOTSUP);
8798 		}
8799 		if ((ts % segvn_pglock_mtbf) == 1) {
8800 			return (EFAULT);
8801 		}
8802 	}
8803 #endif
8804 
8805 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_START,
8806 	    "segvn_pagelock: start seg %p addr %p", seg, addr);
8807 
8808 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
8809 	ASSERT(type == L_PAGELOCK || type == L_PAGEUNLOCK);
8810 
8811 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8812 
8813 	/*
8814 	 * for now we only support pagelock to anon memory. We would have to
8815 	 * check protections for vnode objects and call into the vnode driver.
8816 	 * That's too much for a fast path. Let the fault entry point handle
8817 	 * it.
8818 	 */
8819 	if (svd->vp != NULL) {
8820 		if (type == L_PAGELOCK) {
8821 			error = ENOTSUP;
8822 			goto out;
8823 		}
8824 		panic("segvn_pagelock(L_PAGEUNLOCK): vp != NULL");
8825 	}
8826 	if ((amp = svd->amp) == NULL) {
8827 		if (type == L_PAGELOCK) {
8828 			error = EFAULT;
8829 			goto out;
8830 		}
8831 		panic("segvn_pagelock(L_PAGEUNLOCK): amp == NULL");
8832 	}
8833 	if (rw != S_READ && rw != S_WRITE) {
8834 		if (type == L_PAGELOCK) {
8835 			error = ENOTSUP;
8836 			goto out;
8837 		}
8838 		panic("segvn_pagelock(L_PAGEUNLOCK): bad rw");
8839 	}
8840 
8841 	if (seg->s_szc != 0) {
8842 		/*
8843 		 * We are adjusting the pagelock region to the large page size
8844 		 * boundary because the unlocked part of a large page cannot
8845 		 * be freed anyway unless all constituent pages of a large
8846 		 * page are locked. Bigger regions reduce pcache chain length
8847 		 * and improve lookup performance. The tradeoff is that the
8848 		 * very first segvn_pagelock() call for a given page is more
8849 		 * expensive if only 1 page_t is needed for IO. This is only
8850 		 * an issue if pcache entry doesn't get reused by several
8851 		 * subsequent calls. We optimize here for the case when pcache
8852 		 * is heavily used by repeated IOs to the same address range.
8853 		 *
8854 		 * Note segment's page size cannot change while we are holding
8855 		 * as lock.  And then it cannot change while softlockcnt is
8856 		 * not 0. This will allow us to correctly recalculate large
8857 		 * page size region for the matching pageunlock/reclaim call
8858 		 * since as_pageunlock() caller must always match
8859 		 * as_pagelock() call's addr and len.
8860 		 *
8861 		 * For pageunlock *ppp points to the pointer of page_t that
8862 		 * corresponds to the real unadjusted start address. Similar
8863 		 * for pagelock *ppp must point to the pointer of page_t that
8864 		 * corresponds to the real unadjusted start address.
8865 		 */
8866 		pgsz = page_get_pagesize(seg->s_szc);
8867 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
8868 		adjustpages = btop((uintptr_t)(addr - lpgaddr));
8869 	} else if (len < segvn_pglock_comb_thrshld) {
8870 		lpgaddr = addr;
8871 		lpgeaddr = addr + len;
8872 		adjustpages = 0;
8873 		pgsz = PAGESIZE;
8874 	} else {
8875 		/*
8876 		 * Align the address range of large enough requests to allow
8877 		 * combining of different shadow lists into 1 to reduce memory
8878 		 * overhead from potentially overlapping large shadow lists
8879 		 * (worst case is we have a 1MB IO into buffers with start
8880 		 * addresses separated by 4K).  Alignment is only possible if
8881 		 * padded chunks have sufficient access permissions. Note
8882 		 * permissions won't change between L_PAGELOCK and
8883 		 * L_PAGEUNLOCK calls since non 0 softlockcnt will force
8884 		 * segvn_setprot() to wait until softlockcnt drops to 0. This
8885 		 * allows us to determine in L_PAGEUNLOCK the same range we
8886 		 * computed in L_PAGELOCK.
8887 		 *
8888 		 * If alignment is limited by segment ends set
8889 		 * sftlck_sbase/sftlck_send flags. In L_PAGELOCK case when
8890 		 * these flags are set bump softlockcnt_sbase/softlockcnt_send
8891 		 * per segment counters. In L_PAGEUNLOCK case decrease
8892 		 * softlockcnt_sbase/softlockcnt_send counters if
8893 		 * sftlck_sbase/sftlck_send flags are set.  When
8894 		 * softlockcnt_sbase/softlockcnt_send are non 0
8895 		 * segvn_concat()/segvn_extend_prev()/segvn_extend_next()
8896 		 * won't merge the segments. This restriction combined with
8897 		 * restriction on segment unmapping and splitting for segments
8898 		 * that have non 0 softlockcnt allows L_PAGEUNLOCK to
8899 		 * correctly determine the same range that was previously
8900 		 * locked by matching L_PAGELOCK.
8901 		 */
8902 		pflags = SEGP_PSHIFT | (segvn_pglock_comb_bshift << 16);
8903 		pgsz = PAGESIZE;
8904 		if (svd->type == MAP_PRIVATE) {
8905 			lpgaddr = (caddr_t)P2ALIGN((uintptr_t)addr,
8906 			    segvn_pglock_comb_balign);
8907 			if (lpgaddr < seg->s_base) {
8908 				lpgaddr = seg->s_base;
8909 				sftlck_sbase = 1;
8910 			}
8911 		} else {
8912 			ulong_t aix = svd->anon_index + seg_page(seg, addr);
8913 			ulong_t aaix = P2ALIGN(aix, segvn_pglock_comb_palign);
8914 			if (aaix < svd->anon_index) {
8915 				lpgaddr = seg->s_base;
8916 				sftlck_sbase = 1;
8917 			} else {
8918 				lpgaddr = addr - ptob(aix - aaix);
8919 				ASSERT(lpgaddr >= seg->s_base);
8920 			}
8921 		}
8922 		if (svd->pageprot && lpgaddr != addr) {
8923 			struct vpage *vp = &svd->vpage[seg_page(seg, lpgaddr)];
8924 			struct vpage *evp = &svd->vpage[seg_page(seg, addr)];
8925 			while (vp < evp) {
8926 				if ((VPP_PROT(vp) & protchk) == 0) {
8927 					break;
8928 				}
8929 				vp++;
8930 			}
8931 			if (vp < evp) {
8932 				lpgaddr = addr;
8933 				pflags = 0;
8934 			}
8935 		}
8936 		lpgeaddr = addr + len;
8937 		if (pflags) {
8938 			if (svd->type == MAP_PRIVATE) {
8939 				lpgeaddr = (caddr_t)P2ROUNDUP(
8940 				    (uintptr_t)lpgeaddr,
8941 				    segvn_pglock_comb_balign);
8942 			} else {
8943 				ulong_t aix = svd->anon_index +
8944 				    seg_page(seg, lpgeaddr);
8945 				ulong_t aaix = P2ROUNDUP(aix,
8946 				    segvn_pglock_comb_palign);
8947 				if (aaix < aix) {
8948 					lpgeaddr = 0;
8949 				} else {
8950 					lpgeaddr += ptob(aaix - aix);
8951 				}
8952 			}
8953 			if (lpgeaddr == 0 ||
8954 			    lpgeaddr > seg->s_base + seg->s_size) {
8955 				lpgeaddr = seg->s_base + seg->s_size;
8956 				sftlck_send = 1;
8957 			}
8958 		}
8959 		if (svd->pageprot && lpgeaddr != addr + len) {
8960 			struct vpage *vp;
8961 			struct vpage *evp;
8962 
8963 			vp = &svd->vpage[seg_page(seg, addr + len)];
8964 			evp = &svd->vpage[seg_page(seg, lpgeaddr)];
8965 
8966 			while (vp < evp) {
8967 				if ((VPP_PROT(vp) & protchk) == 0) {
8968 					break;
8969 				}
8970 				vp++;
8971 			}
8972 			if (vp < evp) {
8973 				lpgeaddr = addr + len;
8974 			}
8975 		}
8976 		adjustpages = btop((uintptr_t)(addr - lpgaddr));
8977 	}
8978 
8979 	/*
8980 	 * For MAP_SHARED segments we create pcache entries tagged by amp and
8981 	 * anon index so that we can share pcache entries with other segments
8982 	 * that map this amp.  For private segments pcache entries are tagged
8983 	 * with segment and virtual address.
8984 	 */
8985 	if (svd->type == MAP_SHARED) {
8986 		pamp = amp;
8987 		paddr = (caddr_t)((lpgaddr - seg->s_base) +
8988 		    ptob(svd->anon_index));
8989 		preclaim_callback = shamp_reclaim;
8990 	} else {
8991 		pamp = NULL;
8992 		paddr = lpgaddr;
8993 		preclaim_callback = segvn_reclaim;
8994 	}
8995 
8996 	if (type == L_PAGEUNLOCK) {
8997 		VM_STAT_ADD(segvnvmstats.pagelock[0]);
8998 
8999 		/*
9000 		 * update hat ref bits for /proc. We need to make sure
9001 		 * that threads tracing the ref and mod bits of the
9002 		 * address space get the right data.
9003 		 * Note: page ref and mod bits are updated at reclaim time
9004 		 */
9005 		if (seg->s_as->a_vbits) {
9006 			for (a = addr; a < addr + len; a += PAGESIZE) {
9007 				if (rw == S_WRITE) {
9008 					hat_setstat(seg->s_as, a,
9009 					    PAGESIZE, P_REF | P_MOD);
9010 				} else {
9011 					hat_setstat(seg->s_as, a,
9012 					    PAGESIZE, P_REF);
9013 				}
9014 			}
9015 		}
9016 
9017 		/*
9018 		 * Check the shadow list entry after the last page used in
9019 		 * this IO request. If it's NOPCACHE_SHWLIST the shadow list
9020 		 * was not inserted into pcache and is not large page
9021 		 * adjusted.  In this case call reclaim callback directly and
9022 		 * don't adjust the shadow list start and size for large
9023 		 * pages.
9024 		 */
9025 		npages = btop(len);
9026 		if ((*ppp)[npages] == NOPCACHE_SHWLIST) {
9027 			void *ptag;
9028 			if (pamp != NULL) {
9029 				ASSERT(svd->type == MAP_SHARED);
9030 				ptag = (void *)pamp;
9031 				paddr = (caddr_t)((addr - seg->s_base) +
9032 				    ptob(svd->anon_index));
9033 			} else {
9034 				ptag = (void *)seg;
9035 				paddr = addr;
9036 			}
9037 			(*preclaim_callback)(ptag, paddr, len, *ppp, rw, 0);
9038 		} else {
9039 			ASSERT((*ppp)[npages] == PCACHE_SHWLIST ||
9040 			    IS_SWAPFSVP((*ppp)[npages]->p_vnode));
9041 			len = lpgeaddr - lpgaddr;
9042 			npages = btop(len);
9043 			seg_pinactive(seg, pamp, paddr, len,
9044 			    *ppp - adjustpages, rw, pflags, preclaim_callback);
9045 		}
9046 
9047 		if (pamp != NULL) {
9048 			ASSERT(svd->type == MAP_SHARED);
9049 			ASSERT(svd->softlockcnt >= npages);
9050 			atomic_add_long((ulong_t *)&svd->softlockcnt, -npages);
9051 		}
9052 
9053 		if (sftlck_sbase) {
9054 			ASSERT(svd->softlockcnt_sbase > 0);
9055 			atomic_dec_ulong((ulong_t *)&svd->softlockcnt_sbase);
9056 		}
9057 		if (sftlck_send) {
9058 			ASSERT(svd->softlockcnt_send > 0);
9059 			atomic_dec_ulong((ulong_t *)&svd->softlockcnt_send);
9060 		}
9061 
9062 		/*
9063 		 * If someone is blocked while unmapping, we purge
9064 		 * segment page cache and thus reclaim pplist synchronously
9065 		 * without waiting for seg_pasync_thread. This speeds up
9066 		 * unmapping in cases where munmap(2) is called, while
9067 		 * raw async i/o is still in progress or where a thread
9068 		 * exits on data fault in a multithreaded application.
9069 		 */
9070 		if (AS_ISUNMAPWAIT(seg->s_as)) {
9071 			if (svd->softlockcnt == 0) {
9072 				mutex_enter(&seg->s_as->a_contents);
9073 				if (AS_ISUNMAPWAIT(seg->s_as)) {
9074 					AS_CLRUNMAPWAIT(seg->s_as);
9075 					cv_broadcast(&seg->s_as->a_cv);
9076 				}
9077 				mutex_exit(&seg->s_as->a_contents);
9078 			} else if (pamp == NULL) {
9079 				/*
9080 				 * softlockcnt is not 0 and this is a
9081 				 * MAP_PRIVATE segment. Try to purge its
9082 				 * pcache entries to reduce softlockcnt.
9083 				 * If it drops to 0 segvn_reclaim()
9084 				 * will wake up a thread waiting on
9085 				 * unmapwait flag.
9086 				 *
9087 				 * We don't purge MAP_SHARED segments with non
9088 				 * 0 softlockcnt since IO is still in progress
9089 				 * for such segments.
9090 				 */
9091 				ASSERT(svd->type == MAP_PRIVATE);
9092 				segvn_purge(seg);
9093 			}
9094 		}
9095 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9096 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_UNLOCK_END,
9097 		    "segvn_pagelock: unlock seg %p addr %p", seg, addr);
9098 		return (0);
9099 	}
9100 
9101 	/* The L_PAGELOCK case ... */
9102 
9103 	VM_STAT_ADD(segvnvmstats.pagelock[1]);
9104 
9105 	/*
9106 	 * For MAP_SHARED segments we have to check protections before
9107 	 * seg_plookup() since pcache entries may be shared by many segments
9108 	 * with potentially different page protections.
9109 	 */
9110 	if (pamp != NULL) {
9111 		ASSERT(svd->type == MAP_SHARED);
9112 		if (svd->pageprot == 0) {
9113 			if ((svd->prot & protchk) == 0) {
9114 				error = EACCES;
9115 				goto out;
9116 			}
9117 		} else {
9118 			/*
9119 			 * check page protections
9120 			 */
9121 			caddr_t ea;
9122 
9123 			if (seg->s_szc) {
9124 				a = lpgaddr;
9125 				ea = lpgeaddr;
9126 			} else {
9127 				a = addr;
9128 				ea = addr + len;
9129 			}
9130 			for (; a < ea; a += pgsz) {
9131 				struct vpage *vp;
9132 
9133 				ASSERT(seg->s_szc == 0 ||
9134 				    sameprot(seg, a, pgsz));
9135 				vp = &svd->vpage[seg_page(seg, a)];
9136 				if ((VPP_PROT(vp) & protchk) == 0) {
9137 					error = EACCES;
9138 					goto out;
9139 				}
9140 			}
9141 		}
9142 	}
9143 
9144 	/*
9145 	 * try to find pages in segment page cache
9146 	 */
9147 	pplist = seg_plookup(seg, pamp, paddr, lpgeaddr - lpgaddr, rw, pflags);
9148 	if (pplist != NULL) {
9149 		if (pamp != NULL) {
9150 			npages = btop((uintptr_t)(lpgeaddr - lpgaddr));
9151 			ASSERT(svd->type == MAP_SHARED);
9152 			atomic_add_long((ulong_t *)&svd->softlockcnt,
9153 			    npages);
9154 		}
9155 		if (sftlck_sbase) {
9156 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_sbase);
9157 		}
9158 		if (sftlck_send) {
9159 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_send);
9160 		}
9161 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9162 		*ppp = pplist + adjustpages;
9163 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_HIT_END,
9164 		    "segvn_pagelock: cache hit seg %p addr %p", seg, addr);
9165 		return (0);
9166 	}
9167 
9168 	/*
9169 	 * For MAP_SHARED segments we already verified above that segment
9170 	 * protections allow this pagelock operation.
9171 	 */
9172 	if (pamp == NULL) {
9173 		ASSERT(svd->type == MAP_PRIVATE);
9174 		if (svd->pageprot == 0) {
9175 			if ((svd->prot & protchk) == 0) {
9176 				error = EACCES;
9177 				goto out;
9178 			}
9179 			if (svd->prot & PROT_WRITE) {
9180 				wlen = lpgeaddr - lpgaddr;
9181 			} else {
9182 				wlen = 0;
9183 				ASSERT(rw == S_READ);
9184 			}
9185 		} else {
9186 			int wcont = 1;
9187 			/*
9188 			 * check page protections
9189 			 */
9190 			for (a = lpgaddr, wlen = 0; a < lpgeaddr; a += pgsz) {
9191 				struct vpage *vp;
9192 
9193 				ASSERT(seg->s_szc == 0 ||
9194 				    sameprot(seg, a, pgsz));
9195 				vp = &svd->vpage[seg_page(seg, a)];
9196 				if ((VPP_PROT(vp) & protchk) == 0) {
9197 					error = EACCES;
9198 					goto out;
9199 				}
9200 				if (wcont && (VPP_PROT(vp) & PROT_WRITE)) {
9201 					wlen += pgsz;
9202 				} else {
9203 					wcont = 0;
9204 					ASSERT(rw == S_READ);
9205 				}
9206 			}
9207 		}
9208 		ASSERT(rw == S_READ || wlen == lpgeaddr - lpgaddr);
9209 		ASSERT(rw == S_WRITE || wlen <= lpgeaddr - lpgaddr);
9210 	}
9211 
9212 	/*
9213 	 * Only build large page adjusted shadow list if we expect to insert
9214 	 * it into pcache. For large enough pages it's a big overhead to
9215 	 * create a shadow list of the entire large page. But this overhead
9216 	 * should be amortized over repeated pcache hits on subsequent reuse
9217 	 * of this shadow list (IO into any range within this shadow list will
9218 	 * find it in pcache since we large page align the request for pcache
9219 	 * lookups). pcache performance is improved with bigger shadow lists
9220 	 * as it reduces the time to pcache the entire big segment and reduces
9221 	 * pcache chain length.
9222 	 */
9223 	if (seg_pinsert_check(seg, pamp, paddr,
9224 	    lpgeaddr - lpgaddr, pflags) == SEGP_SUCCESS) {
9225 		addr = lpgaddr;
9226 		len = lpgeaddr - lpgaddr;
9227 		use_pcache = 1;
9228 	} else {
9229 		use_pcache = 0;
9230 		/*
9231 		 * Since this entry will not be inserted into the pcache, we
9232 		 * will not do any adjustments to the starting address or
9233 		 * size of the memory to be locked.
9234 		 */
9235 		adjustpages = 0;
9236 	}
9237 	npages = btop(len);
9238 
9239 	pplist = kmem_alloc(sizeof (page_t *) * (npages + 1), KM_SLEEP);
9240 	pl = pplist;
9241 	*ppp = pplist + adjustpages;
9242 	/*
9243 	 * If use_pcache is 0 this shadow list is not large page adjusted.
9244 	 * Record this info in the last entry of shadow array so that
9245 	 * L_PAGEUNLOCK can determine if it should large page adjust the
9246 	 * address range to find the real range that was locked.
9247 	 */
9248 	pl[npages] = use_pcache ? PCACHE_SHWLIST : NOPCACHE_SHWLIST;
9249 
9250 	page = seg_page(seg, addr);
9251 	anon_index = svd->anon_index + page;
9252 
9253 	anlock = 0;
9254 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
9255 	ASSERT(amp->a_szc >= seg->s_szc);
9256 	anpgcnt = page_get_pagecnt(amp->a_szc);
9257 	for (a = addr; a < addr + len; a += PAGESIZE, anon_index++) {
9258 		struct anon *ap;
9259 		struct vnode *vp;
9260 		u_offset_t off;
9261 
9262 		/*
9263 		 * Lock and unlock anon array only once per large page.
9264 		 * anon_array_enter() locks the root anon slot according to
9265 		 * a_szc which can't change while anon map is locked.  We lock
9266 		 * anon the first time through this loop and each time we
9267 		 * reach anon index that corresponds to a root of a large
9268 		 * page.
9269 		 */
9270 		if (a == addr || P2PHASE(anon_index, anpgcnt) == 0) {
9271 			ASSERT(anlock == 0);
9272 			anon_array_enter(amp, anon_index, &cookie);
9273 			anlock = 1;
9274 		}
9275 		ap = anon_get_ptr(amp->ahp, anon_index);
9276 
9277 		/*
9278 		 * We must never use seg_pcache for COW pages
9279 		 * because we might end up with original page still
9280 		 * lying in seg_pcache even after private page is
9281 		 * created. This leads to data corruption as
9282 		 * aio_write refers to the page still in cache
9283 		 * while all other accesses refer to the private
9284 		 * page.
9285 		 */
9286 		if (ap == NULL || ap->an_refcnt != 1) {
9287 			struct vpage *vpage;
9288 
9289 			if (seg->s_szc) {
9290 				error = EFAULT;
9291 				break;
9292 			}
9293 			if (svd->vpage != NULL) {
9294 				vpage = &svd->vpage[seg_page(seg, a)];
9295 			} else {
9296 				vpage = NULL;
9297 			}
9298 			ASSERT(anlock);
9299 			anon_array_exit(&cookie);
9300 			anlock = 0;
9301 			pp = NULL;
9302 			error = segvn_faultpage(seg->s_as->a_hat, seg, a, 0,
9303 			    vpage, &pp, 0, F_INVAL, rw, 1);
9304 			if (error) {
9305 				error = fc_decode(error);
9306 				break;
9307 			}
9308 			anon_array_enter(amp, anon_index, &cookie);
9309 			anlock = 1;
9310 			ap = anon_get_ptr(amp->ahp, anon_index);
9311 			if (ap == NULL || ap->an_refcnt != 1) {
9312 				error = EFAULT;
9313 				break;
9314 			}
9315 		}
9316 		swap_xlate(ap, &vp, &off);
9317 		pp = page_lookup_nowait(vp, off, SE_SHARED);
9318 		if (pp == NULL) {
9319 			error = EFAULT;
9320 			break;
9321 		}
9322 		if (ap->an_pvp != NULL) {
9323 			anon_swap_free(ap, pp);
9324 		}
9325 		/*
9326 		 * Unlock anon if this is the last slot in a large page.
9327 		 */
9328 		if (P2PHASE(anon_index, anpgcnt) == anpgcnt - 1) {
9329 			ASSERT(anlock);
9330 			anon_array_exit(&cookie);
9331 			anlock = 0;
9332 		}
9333 		*pplist++ = pp;
9334 	}
9335 	if (anlock) {		/* Ensure the lock is dropped */
9336 		anon_array_exit(&cookie);
9337 	}
9338 	ANON_LOCK_EXIT(&amp->a_rwlock);
9339 
9340 	if (a >= addr + len) {
9341 		atomic_add_long((ulong_t *)&svd->softlockcnt, npages);
9342 		if (pamp != NULL) {
9343 			ASSERT(svd->type == MAP_SHARED);
9344 			atomic_add_long((ulong_t *)&pamp->a_softlockcnt,
9345 			    npages);
9346 			wlen = len;
9347 		}
9348 		if (sftlck_sbase) {
9349 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_sbase);
9350 		}
9351 		if (sftlck_send) {
9352 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_send);
9353 		}
9354 		if (use_pcache) {
9355 			(void) seg_pinsert(seg, pamp, paddr, len, wlen, pl,
9356 			    rw, pflags, preclaim_callback);
9357 		}
9358 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9359 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_FILL_END,
9360 		    "segvn_pagelock: cache fill seg %p addr %p", seg, addr);
9361 		return (0);
9362 	}
9363 
9364 	pplist = pl;
9365 	np = ((uintptr_t)(a - addr)) >> PAGESHIFT;
9366 	while (np > (uint_t)0) {
9367 		ASSERT(PAGE_LOCKED(*pplist));
9368 		page_unlock(*pplist);
9369 		np--;
9370 		pplist++;
9371 	}
9372 	kmem_free(pl, sizeof (page_t *) * (npages + 1));
9373 out:
9374 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9375 	*ppp = NULL;
9376 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
9377 	    "segvn_pagelock: cache miss seg %p addr %p", seg, addr);
9378 	return (error);
9379 }
9380 
9381 /*
9382  * purge any cached pages in the I/O page cache
9383  */
9384 static void
9385 segvn_purge(struct seg *seg)
9386 {
9387 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9388 
9389 	/*
9390 	 * pcache is only used by pure anon segments.
9391 	 */
9392 	if (svd->amp == NULL || svd->vp != NULL) {
9393 		return;
9394 	}
9395 
9396 	/*
9397 	 * For MAP_SHARED segments non 0 segment's softlockcnt means
9398 	 * active IO is still in progress via this segment. So we only
9399 	 * purge MAP_SHARED segments when their softlockcnt is 0.
9400 	 */
9401 	if (svd->type == MAP_PRIVATE) {
9402 		if (svd->softlockcnt) {
9403 			seg_ppurge(seg, NULL, 0);
9404 		}
9405 	} else if (svd->softlockcnt == 0 && svd->amp->a_softlockcnt != 0) {
9406 		seg_ppurge(seg, svd->amp, 0);
9407 	}
9408 }
9409 
9410 /*
9411  * If async argument is not 0 we are called from pcache async thread and don't
9412  * hold AS lock.
9413  */
9414 
9415 /*ARGSUSED*/
9416 static int
9417 segvn_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9418 	enum seg_rw rw, int async)
9419 {
9420 	struct seg *seg = (struct seg *)ptag;
9421 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9422 	pgcnt_t np, npages;
9423 	struct page **pl;
9424 
9425 	npages = np = btop(len);
9426 	ASSERT(npages);
9427 
9428 	ASSERT(svd->vp == NULL && svd->amp != NULL);
9429 	ASSERT(svd->softlockcnt >= npages);
9430 	ASSERT(async || AS_LOCK_HELD(seg->s_as));
9431 
9432 	pl = pplist;
9433 
9434 	ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9435 	ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9436 
9437 	while (np > (uint_t)0) {
9438 		if (rw == S_WRITE) {
9439 			hat_setrefmod(*pplist);
9440 		} else {
9441 			hat_setref(*pplist);
9442 		}
9443 		page_unlock(*pplist);
9444 		np--;
9445 		pplist++;
9446 	}
9447 
9448 	kmem_free(pl, sizeof (page_t *) * (npages + 1));
9449 
9450 	/*
9451 	 * If we are pcache async thread we don't hold AS lock. This means if
9452 	 * softlockcnt drops to 0 after the decrement below address space may
9453 	 * get freed. We can't allow it since after softlock derement to 0 we
9454 	 * still need to access as structure for possible wakeup of unmap
9455 	 * waiters. To prevent the disappearance of as we take this segment
9456 	 * segfree_syncmtx. segvn_free() also takes this mutex as a barrier to
9457 	 * make sure this routine completes before segment is freed.
9458 	 *
9459 	 * The second complication we have to deal with in async case is a
9460 	 * possibility of missed wake up of unmap wait thread. When we don't
9461 	 * hold as lock here we may take a_contents lock before unmap wait
9462 	 * thread that was first to see softlockcnt was still not 0. As a
9463 	 * result we'll fail to wake up an unmap wait thread. To avoid this
9464 	 * race we set nounmapwait flag in as structure if we drop softlockcnt
9465 	 * to 0 when we were called by pcache async thread.  unmapwait thread
9466 	 * will not block if this flag is set.
9467 	 */
9468 	if (async) {
9469 		mutex_enter(&svd->segfree_syncmtx);
9470 	}
9471 
9472 	if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -npages)) {
9473 		if (async || AS_ISUNMAPWAIT(seg->s_as)) {
9474 			mutex_enter(&seg->s_as->a_contents);
9475 			if (async) {
9476 				AS_SETNOUNMAPWAIT(seg->s_as);
9477 			}
9478 			if (AS_ISUNMAPWAIT(seg->s_as)) {
9479 				AS_CLRUNMAPWAIT(seg->s_as);
9480 				cv_broadcast(&seg->s_as->a_cv);
9481 			}
9482 			mutex_exit(&seg->s_as->a_contents);
9483 		}
9484 	}
9485 
9486 	if (async) {
9487 		mutex_exit(&svd->segfree_syncmtx);
9488 	}
9489 	return (0);
9490 }
9491 
9492 /*ARGSUSED*/
9493 static int
9494 shamp_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9495 	enum seg_rw rw, int async)
9496 {
9497 	amp_t *amp = (amp_t *)ptag;
9498 	pgcnt_t np, npages;
9499 	struct page **pl;
9500 
9501 	npages = np = btop(len);
9502 	ASSERT(npages);
9503 	ASSERT(amp->a_softlockcnt >= npages);
9504 
9505 	pl = pplist;
9506 
9507 	ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9508 	ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9509 
9510 	while (np > (uint_t)0) {
9511 		if (rw == S_WRITE) {
9512 			hat_setrefmod(*pplist);
9513 		} else {
9514 			hat_setref(*pplist);
9515 		}
9516 		page_unlock(*pplist);
9517 		np--;
9518 		pplist++;
9519 	}
9520 
9521 	kmem_free(pl, sizeof (page_t *) * (npages + 1));
9522 
9523 	/*
9524 	 * If somebody sleeps in anonmap_purge() wake them up if a_softlockcnt
9525 	 * drops to 0. anon map can't be freed until a_softlockcnt drops to 0
9526 	 * and anonmap_purge() acquires a_purgemtx.
9527 	 */
9528 	mutex_enter(&amp->a_purgemtx);
9529 	if (!atomic_add_long_nv((ulong_t *)&amp->a_softlockcnt, -npages) &&
9530 	    amp->a_purgewait) {
9531 		amp->a_purgewait = 0;
9532 		cv_broadcast(&amp->a_purgecv);
9533 	}
9534 	mutex_exit(&amp->a_purgemtx);
9535 	return (0);
9536 }
9537 
9538 /*
9539  * get a memory ID for an addr in a given segment
9540  *
9541  * XXX only creates PAGESIZE pages if anon slots are not initialized.
9542  * At fault time they will be relocated into larger pages.
9543  */
9544 static int
9545 segvn_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
9546 {
9547 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9548 	struct anon 	*ap = NULL;
9549 	ulong_t		anon_index;
9550 	struct anon_map	*amp;
9551 	anon_sync_obj_t cookie;
9552 
9553 	if (svd->type == MAP_PRIVATE) {
9554 		memidp->val[0] = (uintptr_t)seg->s_as;
9555 		memidp->val[1] = (uintptr_t)addr;
9556 		return (0);
9557 	}
9558 
9559 	if (svd->type == MAP_SHARED) {
9560 		if (svd->vp) {
9561 			memidp->val[0] = (uintptr_t)svd->vp;
9562 			memidp->val[1] = (u_longlong_t)svd->offset +
9563 			    (uintptr_t)(addr - seg->s_base);
9564 			return (0);
9565 		} else {
9566 
9567 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
9568 			if ((amp = svd->amp) != NULL) {
9569 				anon_index = svd->anon_index +
9570 				    seg_page(seg, addr);
9571 			}
9572 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9573 
9574 			ASSERT(amp != NULL);
9575 
9576 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
9577 			anon_array_enter(amp, anon_index, &cookie);
9578 			ap = anon_get_ptr(amp->ahp, anon_index);
9579 			if (ap == NULL) {
9580 				page_t		*pp;
9581 
9582 				pp = anon_zero(seg, addr, &ap, svd->cred);
9583 				if (pp == NULL) {
9584 					anon_array_exit(&cookie);
9585 					ANON_LOCK_EXIT(&amp->a_rwlock);
9586 					return (ENOMEM);
9587 				}
9588 				ASSERT(anon_get_ptr(amp->ahp, anon_index)
9589 				    == NULL);
9590 				(void) anon_set_ptr(amp->ahp, anon_index,
9591 				    ap, ANON_SLEEP);
9592 				page_unlock(pp);
9593 			}
9594 
9595 			anon_array_exit(&cookie);
9596 			ANON_LOCK_EXIT(&amp->a_rwlock);
9597 
9598 			memidp->val[0] = (uintptr_t)ap;
9599 			memidp->val[1] = (uintptr_t)addr & PAGEOFFSET;
9600 			return (0);
9601 		}
9602 	}
9603 	return (EINVAL);
9604 }
9605 
9606 static int
9607 sameprot(struct seg *seg, caddr_t a, size_t len)
9608 {
9609 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9610 	struct vpage *vpage;
9611 	spgcnt_t pages = btop(len);
9612 	uint_t prot;
9613 
9614 	if (svd->pageprot == 0)
9615 		return (1);
9616 
9617 	ASSERT(svd->vpage != NULL);
9618 
9619 	vpage = &svd->vpage[seg_page(seg, a)];
9620 	prot = VPP_PROT(vpage);
9621 	vpage++;
9622 	pages--;
9623 	while (pages-- > 0) {
9624 		if (prot != VPP_PROT(vpage))
9625 			return (0);
9626 		vpage++;
9627 	}
9628 	return (1);
9629 }
9630 
9631 /*
9632  * Get memory allocation policy info for specified address in given segment
9633  */
9634 static lgrp_mem_policy_info_t *
9635 segvn_getpolicy(struct seg *seg, caddr_t addr)
9636 {
9637 	struct anon_map		*amp;
9638 	ulong_t			anon_index;
9639 	lgrp_mem_policy_info_t	*policy_info;
9640 	struct segvn_data	*svn_data;
9641 	u_offset_t		vn_off;
9642 	vnode_t			*vp;
9643 
9644 	ASSERT(seg != NULL);
9645 
9646 	svn_data = (struct segvn_data *)seg->s_data;
9647 	if (svn_data == NULL)
9648 		return (NULL);
9649 
9650 	/*
9651 	 * Get policy info for private or shared memory
9652 	 */
9653 	if (svn_data->type != MAP_SHARED) {
9654 		if (svn_data->tr_state != SEGVN_TR_ON) {
9655 			policy_info = &svn_data->policy_info;
9656 		} else {
9657 			policy_info = &svn_data->tr_policy_info;
9658 			ASSERT(policy_info->mem_policy ==
9659 			    LGRP_MEM_POLICY_NEXT_SEG);
9660 		}
9661 	} else {
9662 		amp = svn_data->amp;
9663 		anon_index = svn_data->anon_index + seg_page(seg, addr);
9664 		vp = svn_data->vp;
9665 		vn_off = svn_data->offset + (uintptr_t)(addr - seg->s_base);
9666 		policy_info = lgrp_shm_policy_get(amp, anon_index, vp, vn_off);
9667 	}
9668 
9669 	return (policy_info);
9670 }
9671 
9672 /*ARGSUSED*/
9673 static int
9674 segvn_capable(struct seg *seg, segcapability_t capability)
9675 {
9676 	return (0);
9677 }
9678 
9679 /*
9680  * Bind text vnode segment to an amp. If we bind successfully mappings will be
9681  * established to per vnode mapping per lgroup amp pages instead of to vnode
9682  * pages. There's one amp per vnode text mapping per lgroup. Many processes
9683  * may share the same text replication amp. If a suitable amp doesn't already
9684  * exist in svntr hash table create a new one.  We may fail to bind to amp if
9685  * segment is not eligible for text replication.  Code below first checks for
9686  * these conditions. If binding is successful segment tr_state is set to on
9687  * and svd->amp points to the amp to use. Otherwise tr_state is set to off and
9688  * svd->amp remains as NULL.
9689  */
9690 static void
9691 segvn_textrepl(struct seg *seg)
9692 {
9693 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
9694 	vnode_t			*vp = svd->vp;
9695 	u_offset_t		off = svd->offset;
9696 	size_t			size = seg->s_size;
9697 	u_offset_t		eoff = off + size;
9698 	uint_t			szc = seg->s_szc;
9699 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9700 	svntr_t			*svntrp;
9701 	struct vattr		va;
9702 	proc_t			*p = seg->s_as->a_proc;
9703 	lgrp_id_t		lgrp_id;
9704 	lgrp_id_t		olid;
9705 	int			first;
9706 	struct anon_map		*amp;
9707 
9708 	ASSERT(AS_LOCK_HELD(seg->s_as));
9709 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9710 	ASSERT(p != NULL);
9711 	ASSERT(svd->tr_state == SEGVN_TR_INIT);
9712 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9713 	ASSERT(svd->flags & MAP_TEXT);
9714 	ASSERT(svd->type == MAP_PRIVATE);
9715 	ASSERT(vp != NULL && svd->amp == NULL);
9716 	ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
9717 	ASSERT(!(svd->flags & MAP_NORESERVE) && svd->swresv == 0);
9718 	ASSERT(seg->s_as != &kas);
9719 	ASSERT(off < eoff);
9720 	ASSERT(svntr_hashtab != NULL);
9721 
9722 	/*
9723 	 * If numa optimizations are no longer desired bail out.
9724 	 */
9725 	if (!lgrp_optimizations()) {
9726 		svd->tr_state = SEGVN_TR_OFF;
9727 		return;
9728 	}
9729 
9730 	/*
9731 	 * Avoid creating anon maps with size bigger than the file size.
9732 	 * If VOP_GETATTR() call fails bail out.
9733 	 */
9734 	va.va_mask = AT_SIZE | AT_MTIME | AT_CTIME;
9735 	if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL) != 0) {
9736 		svd->tr_state = SEGVN_TR_OFF;
9737 		SEGVN_TR_ADDSTAT(gaerr);
9738 		return;
9739 	}
9740 	if (btopr(va.va_size) < btopr(eoff)) {
9741 		svd->tr_state = SEGVN_TR_OFF;
9742 		SEGVN_TR_ADDSTAT(overmap);
9743 		return;
9744 	}
9745 
9746 	/*
9747 	 * VVMEXEC may not be set yet if exec() prefaults text segment. Set
9748 	 * this flag now before vn_is_mapped(V_WRITE) so that MAP_SHARED
9749 	 * mapping that checks if trcache for this vnode needs to be
9750 	 * invalidated can't miss us.
9751 	 */
9752 	if (!(vp->v_flag & VVMEXEC)) {
9753 		mutex_enter(&vp->v_lock);
9754 		vp->v_flag |= VVMEXEC;
9755 		mutex_exit(&vp->v_lock);
9756 	}
9757 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9758 	/*
9759 	 * Bail out if potentially MAP_SHARED writable mappings exist to this
9760 	 * vnode.  We don't want to use old file contents from existing
9761 	 * replicas if this mapping was established after the original file
9762 	 * was changed.
9763 	 */
9764 	if (vn_is_mapped(vp, V_WRITE)) {
9765 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9766 		svd->tr_state = SEGVN_TR_OFF;
9767 		SEGVN_TR_ADDSTAT(wrcnt);
9768 		return;
9769 	}
9770 	svntrp = svntr_hashtab[hash].tr_head;
9771 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9772 		ASSERT(svntrp->tr_refcnt != 0);
9773 		if (svntrp->tr_vp != vp) {
9774 			continue;
9775 		}
9776 
9777 		/*
9778 		 * Bail out if the file or its attributes were changed after
9779 		 * this replication entry was created since we need to use the
9780 		 * latest file contents. Note that mtime test alone is not
9781 		 * sufficient because a user can explicitly change mtime via
9782 		 * utimes(2) interfaces back to the old value after modifiying
9783 		 * the file contents. To detect this case we also have to test
9784 		 * ctime which among other things records the time of the last
9785 		 * mtime change by utimes(2). ctime is not changed when the file
9786 		 * is only read or executed so we expect that typically existing
9787 		 * replication amp's can be used most of the time.
9788 		 */
9789 		if (!svntrp->tr_valid ||
9790 		    svntrp->tr_mtime.tv_sec != va.va_mtime.tv_sec ||
9791 		    svntrp->tr_mtime.tv_nsec != va.va_mtime.tv_nsec ||
9792 		    svntrp->tr_ctime.tv_sec != va.va_ctime.tv_sec ||
9793 		    svntrp->tr_ctime.tv_nsec != va.va_ctime.tv_nsec) {
9794 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9795 			svd->tr_state = SEGVN_TR_OFF;
9796 			SEGVN_TR_ADDSTAT(stale);
9797 			return;
9798 		}
9799 		/*
9800 		 * if off, eoff and szc match current segment we found the
9801 		 * existing entry we can use.
9802 		 */
9803 		if (svntrp->tr_off == off && svntrp->tr_eoff == eoff &&
9804 		    svntrp->tr_szc == szc) {
9805 			break;
9806 		}
9807 		/*
9808 		 * Don't create different but overlapping in file offsets
9809 		 * entries to avoid replication of the same file pages more
9810 		 * than once per lgroup.
9811 		 */
9812 		if ((off >= svntrp->tr_off && off < svntrp->tr_eoff) ||
9813 		    (eoff > svntrp->tr_off && eoff <= svntrp->tr_eoff)) {
9814 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9815 			svd->tr_state = SEGVN_TR_OFF;
9816 			SEGVN_TR_ADDSTAT(overlap);
9817 			return;
9818 		}
9819 	}
9820 	/*
9821 	 * If we didn't find existing entry create a new one.
9822 	 */
9823 	if (svntrp == NULL) {
9824 		svntrp = kmem_cache_alloc(svntr_cache, KM_NOSLEEP);
9825 		if (svntrp == NULL) {
9826 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9827 			svd->tr_state = SEGVN_TR_OFF;
9828 			SEGVN_TR_ADDSTAT(nokmem);
9829 			return;
9830 		}
9831 #ifdef DEBUG
9832 		{
9833 			lgrp_id_t i;
9834 			for (i = 0; i < NLGRPS_MAX; i++) {
9835 				ASSERT(svntrp->tr_amp[i] == NULL);
9836 			}
9837 		}
9838 #endif /* DEBUG */
9839 		svntrp->tr_vp = vp;
9840 		svntrp->tr_off = off;
9841 		svntrp->tr_eoff = eoff;
9842 		svntrp->tr_szc = szc;
9843 		svntrp->tr_valid = 1;
9844 		svntrp->tr_mtime = va.va_mtime;
9845 		svntrp->tr_ctime = va.va_ctime;
9846 		svntrp->tr_refcnt = 0;
9847 		svntrp->tr_next = svntr_hashtab[hash].tr_head;
9848 		svntr_hashtab[hash].tr_head = svntrp;
9849 	}
9850 	first = 1;
9851 again:
9852 	/*
9853 	 * We want to pick a replica with pages on main thread's (t_tid = 1,
9854 	 * aka T1) lgrp. Currently text replication is only optimized for
9855 	 * workloads that either have all threads of a process on the same
9856 	 * lgrp or execute their large text primarily on main thread.
9857 	 */
9858 	lgrp_id = p->p_t1_lgrpid;
9859 	if (lgrp_id == LGRP_NONE) {
9860 		/*
9861 		 * In case exec() prefaults text on non main thread use
9862 		 * current thread lgrpid.  It will become main thread anyway
9863 		 * soon.
9864 		 */
9865 		lgrp_id = lgrp_home_id(curthread);
9866 	}
9867 	/*
9868 	 * Set p_tr_lgrpid to lgrpid if it hasn't been set yet.  Otherwise
9869 	 * just set it to NLGRPS_MAX if it's different from current process T1
9870 	 * home lgrp.  p_tr_lgrpid is used to detect if process uses text
9871 	 * replication and T1 new home is different from lgrp used for text
9872 	 * replication. When this happens asyncronous segvn thread rechecks if
9873 	 * segments should change lgrps used for text replication.  If we fail
9874 	 * to set p_tr_lgrpid with atomic_cas_32 then set it to NLGRPS_MAX
9875 	 * without cas if it's not already NLGRPS_MAX and not equal lgrp_id
9876 	 * we want to use.  We don't need to use cas in this case because
9877 	 * another thread that races in between our non atomic check and set
9878 	 * may only change p_tr_lgrpid to NLGRPS_MAX at this point.
9879 	 */
9880 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9881 	olid = p->p_tr_lgrpid;
9882 	if (lgrp_id != olid && olid != NLGRPS_MAX) {
9883 		lgrp_id_t nlid = (olid == LGRP_NONE) ? lgrp_id : NLGRPS_MAX;
9884 		if (atomic_cas_32((uint32_t *)&p->p_tr_lgrpid, olid, nlid) !=
9885 		    olid) {
9886 			olid = p->p_tr_lgrpid;
9887 			ASSERT(olid != LGRP_NONE);
9888 			if (olid != lgrp_id && olid != NLGRPS_MAX) {
9889 				p->p_tr_lgrpid = NLGRPS_MAX;
9890 			}
9891 		}
9892 		ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9893 		membar_producer();
9894 		/*
9895 		 * lgrp_move_thread() won't schedule async recheck after
9896 		 * p->p_t1_lgrpid update unless p->p_tr_lgrpid is not
9897 		 * LGRP_NONE. Recheck p_t1_lgrpid once now that p->p_tr_lgrpid
9898 		 * is not LGRP_NONE.
9899 		 */
9900 		if (first && p->p_t1_lgrpid != LGRP_NONE &&
9901 		    p->p_t1_lgrpid != lgrp_id) {
9902 			first = 0;
9903 			goto again;
9904 		}
9905 	}
9906 	/*
9907 	 * If no amp was created yet for lgrp_id create a new one as long as
9908 	 * we have enough memory to afford it.
9909 	 */
9910 	if ((amp = svntrp->tr_amp[lgrp_id]) == NULL) {
9911 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9912 		if (trmem > segvn_textrepl_max_bytes) {
9913 			SEGVN_TR_ADDSTAT(normem);
9914 			goto fail;
9915 		}
9916 		if (anon_try_resv_zone(size, NULL) == 0) {
9917 			SEGVN_TR_ADDSTAT(noanon);
9918 			goto fail;
9919 		}
9920 		amp = anonmap_alloc(size, size, ANON_NOSLEEP);
9921 		if (amp == NULL) {
9922 			anon_unresv_zone(size, NULL);
9923 			SEGVN_TR_ADDSTAT(nokmem);
9924 			goto fail;
9925 		}
9926 		ASSERT(amp->refcnt == 1);
9927 		amp->a_szc = szc;
9928 		svntrp->tr_amp[lgrp_id] = amp;
9929 		SEGVN_TR_ADDSTAT(newamp);
9930 	}
9931 	svntrp->tr_refcnt++;
9932 	ASSERT(svd->svn_trnext == NULL);
9933 	ASSERT(svd->svn_trprev == NULL);
9934 	svd->svn_trnext = svntrp->tr_svnhead;
9935 	svd->svn_trprev = NULL;
9936 	if (svntrp->tr_svnhead != NULL) {
9937 		svntrp->tr_svnhead->svn_trprev = svd;
9938 	}
9939 	svntrp->tr_svnhead = svd;
9940 	ASSERT(amp->a_szc == szc && amp->size == size && amp->swresv == size);
9941 	ASSERT(amp->refcnt >= 1);
9942 	svd->amp = amp;
9943 	svd->anon_index = 0;
9944 	svd->tr_policy_info.mem_policy = LGRP_MEM_POLICY_NEXT_SEG;
9945 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
9946 	svd->tr_state = SEGVN_TR_ON;
9947 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9948 	SEGVN_TR_ADDSTAT(repl);
9949 	return;
9950 fail:
9951 	ASSERT(segvn_textrepl_bytes >= size);
9952 	atomic_add_long(&segvn_textrepl_bytes, -size);
9953 	ASSERT(svntrp != NULL);
9954 	ASSERT(svntrp->tr_amp[lgrp_id] == NULL);
9955 	if (svntrp->tr_refcnt == 0) {
9956 		ASSERT(svntrp == svntr_hashtab[hash].tr_head);
9957 		svntr_hashtab[hash].tr_head = svntrp->tr_next;
9958 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9959 		kmem_cache_free(svntr_cache, svntrp);
9960 	} else {
9961 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9962 	}
9963 	svd->tr_state = SEGVN_TR_OFF;
9964 }
9965 
9966 /*
9967  * Convert seg back to regular vnode mapping seg by unbinding it from its text
9968  * replication amp.  This routine is most typically called when segment is
9969  * unmapped but can also be called when segment no longer qualifies for text
9970  * replication (e.g. due to protection changes). If unload_unmap is set use
9971  * HAT_UNLOAD_UNMAP flag in hat_unload_callback().  If we are the last user of
9972  * svntr free all its anon maps and remove it from the hash table.
9973  */
9974 static void
9975 segvn_textunrepl(struct seg *seg, int unload_unmap)
9976 {
9977 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
9978 	vnode_t			*vp = svd->vp;
9979 	u_offset_t		off = svd->offset;
9980 	size_t			size = seg->s_size;
9981 	u_offset_t		eoff = off + size;
9982 	uint_t			szc = seg->s_szc;
9983 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9984 	svntr_t			*svntrp;
9985 	svntr_t			**prv_svntrp;
9986 	lgrp_id_t		lgrp_id = svd->tr_policy_info.mem_lgrpid;
9987 	lgrp_id_t		i;
9988 
9989 	ASSERT(AS_LOCK_HELD(seg->s_as));
9990 	ASSERT(AS_WRITE_HELD(seg->s_as) ||
9991 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9992 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9993 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9994 	ASSERT(svd->amp != NULL);
9995 	ASSERT(svd->amp->refcnt >= 1);
9996 	ASSERT(svd->anon_index == 0);
9997 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9998 	ASSERT(svntr_hashtab != NULL);
9999 
10000 	mutex_enter(&svntr_hashtab[hash].tr_lock);
10001 	prv_svntrp = &svntr_hashtab[hash].tr_head;
10002 	for (; (svntrp = *prv_svntrp) != NULL; prv_svntrp = &svntrp->tr_next) {
10003 		ASSERT(svntrp->tr_refcnt != 0);
10004 		if (svntrp->tr_vp == vp && svntrp->tr_off == off &&
10005 		    svntrp->tr_eoff == eoff && svntrp->tr_szc == szc) {
10006 			break;
10007 		}
10008 	}
10009 	if (svntrp == NULL) {
10010 		panic("segvn_textunrepl: svntr record not found");
10011 	}
10012 	if (svntrp->tr_amp[lgrp_id] != svd->amp) {
10013 		panic("segvn_textunrepl: amp mismatch");
10014 	}
10015 	svd->tr_state = SEGVN_TR_OFF;
10016 	svd->amp = NULL;
10017 	if (svd->svn_trprev == NULL) {
10018 		ASSERT(svntrp->tr_svnhead == svd);
10019 		svntrp->tr_svnhead = svd->svn_trnext;
10020 		if (svntrp->tr_svnhead != NULL) {
10021 			svntrp->tr_svnhead->svn_trprev = NULL;
10022 		}
10023 		svd->svn_trnext = NULL;
10024 	} else {
10025 		svd->svn_trprev->svn_trnext = svd->svn_trnext;
10026 		if (svd->svn_trnext != NULL) {
10027 			svd->svn_trnext->svn_trprev = svd->svn_trprev;
10028 			svd->svn_trnext = NULL;
10029 		}
10030 		svd->svn_trprev = NULL;
10031 	}
10032 	if (--svntrp->tr_refcnt) {
10033 		mutex_exit(&svntr_hashtab[hash].tr_lock);
10034 		goto done;
10035 	}
10036 	*prv_svntrp = svntrp->tr_next;
10037 	mutex_exit(&svntr_hashtab[hash].tr_lock);
10038 	for (i = 0; i < NLGRPS_MAX; i++) {
10039 		struct anon_map *amp = svntrp->tr_amp[i];
10040 		if (amp == NULL) {
10041 			continue;
10042 		}
10043 		ASSERT(amp->refcnt == 1);
10044 		ASSERT(amp->swresv == size);
10045 		ASSERT(amp->size == size);
10046 		ASSERT(amp->a_szc == szc);
10047 		if (amp->a_szc != 0) {
10048 			anon_free_pages(amp->ahp, 0, size, szc);
10049 		} else {
10050 			anon_free(amp->ahp, 0, size);
10051 		}
10052 		svntrp->tr_amp[i] = NULL;
10053 		ASSERT(segvn_textrepl_bytes >= size);
10054 		atomic_add_long(&segvn_textrepl_bytes, -size);
10055 		anon_unresv_zone(amp->swresv, NULL);
10056 		amp->refcnt = 0;
10057 		anonmap_free(amp);
10058 	}
10059 	kmem_cache_free(svntr_cache, svntrp);
10060 done:
10061 	hat_unload_callback(seg->s_as->a_hat, seg->s_base, size,
10062 	    unload_unmap ? HAT_UNLOAD_UNMAP : 0, NULL);
10063 }
10064 
10065 /*
10066  * This is called when a MAP_SHARED writable mapping is created to a vnode
10067  * that is currently used for execution (VVMEXEC flag is set). In this case we
10068  * need to prevent further use of existing replicas.
10069  */
10070 static void
10071 segvn_inval_trcache(vnode_t *vp)
10072 {
10073 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
10074 	svntr_t			*svntrp;
10075 
10076 	ASSERT(vp->v_flag & VVMEXEC);
10077 
10078 	if (svntr_hashtab == NULL) {
10079 		return;
10080 	}
10081 
10082 	mutex_enter(&svntr_hashtab[hash].tr_lock);
10083 	svntrp = svntr_hashtab[hash].tr_head;
10084 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
10085 		ASSERT(svntrp->tr_refcnt != 0);
10086 		if (svntrp->tr_vp == vp && svntrp->tr_valid) {
10087 			svntrp->tr_valid = 0;
10088 		}
10089 	}
10090 	mutex_exit(&svntr_hashtab[hash].tr_lock);
10091 }
10092 
10093 static void
10094 segvn_trasync_thread(void)
10095 {
10096 	callb_cpr_t cpr_info;
10097 	kmutex_t cpr_lock;	/* just for CPR stuff */
10098 
10099 	mutex_init(&cpr_lock, NULL, MUTEX_DEFAULT, NULL);
10100 
10101 	CALLB_CPR_INIT(&cpr_info, &cpr_lock,
10102 	    callb_generic_cpr, "segvn_async");
10103 
10104 	if (segvn_update_textrepl_interval == 0) {
10105 		segvn_update_textrepl_interval = segvn_update_tr_time * hz;
10106 	} else {
10107 		segvn_update_textrepl_interval *= hz;
10108 	}
10109 	(void) timeout(segvn_trupdate_wakeup, NULL,
10110 	    segvn_update_textrepl_interval);
10111 
10112 	for (;;) {
10113 		mutex_enter(&cpr_lock);
10114 		CALLB_CPR_SAFE_BEGIN(&cpr_info);
10115 		mutex_exit(&cpr_lock);
10116 		sema_p(&segvn_trasync_sem);
10117 		mutex_enter(&cpr_lock);
10118 		CALLB_CPR_SAFE_END(&cpr_info, &cpr_lock);
10119 		mutex_exit(&cpr_lock);
10120 		segvn_trupdate();
10121 	}
10122 }
10123 
10124 static uint64_t segvn_lgrp_trthr_migrs_snpsht = 0;
10125 
10126 static void
10127 segvn_trupdate_wakeup(void *dummy)
10128 {
10129 	uint64_t cur_lgrp_trthr_migrs = lgrp_get_trthr_migrations();
10130 
10131 	if (cur_lgrp_trthr_migrs != segvn_lgrp_trthr_migrs_snpsht) {
10132 		segvn_lgrp_trthr_migrs_snpsht = cur_lgrp_trthr_migrs;
10133 		sema_v(&segvn_trasync_sem);
10134 	}
10135 
10136 	if (!segvn_disable_textrepl_update &&
10137 	    segvn_update_textrepl_interval != 0) {
10138 		(void) timeout(segvn_trupdate_wakeup, dummy,
10139 		    segvn_update_textrepl_interval);
10140 	}
10141 }
10142 
10143 static void
10144 segvn_trupdate(void)
10145 {
10146 	ulong_t		hash;
10147 	svntr_t		*svntrp;
10148 	segvn_data_t	*svd;
10149 
10150 	ASSERT(svntr_hashtab != NULL);
10151 
10152 	for (hash = 0; hash < svntr_hashtab_sz; hash++) {
10153 		mutex_enter(&svntr_hashtab[hash].tr_lock);
10154 		svntrp = svntr_hashtab[hash].tr_head;
10155 		for (; svntrp != NULL; svntrp = svntrp->tr_next) {
10156 			ASSERT(svntrp->tr_refcnt != 0);
10157 			svd = svntrp->tr_svnhead;
10158 			for (; svd != NULL; svd = svd->svn_trnext) {
10159 				segvn_trupdate_seg(svd->seg, svd, svntrp,
10160 				    hash);
10161 			}
10162 		}
10163 		mutex_exit(&svntr_hashtab[hash].tr_lock);
10164 	}
10165 }
10166 
10167 static void
10168 segvn_trupdate_seg(struct seg *seg,
10169 	segvn_data_t *svd,
10170 	svntr_t *svntrp,
10171 	ulong_t hash)
10172 {
10173 	proc_t			*p;
10174 	lgrp_id_t		lgrp_id;
10175 	struct as		*as;
10176 	size_t			size;
10177 	struct anon_map		*amp;
10178 
10179 	ASSERT(svd->vp != NULL);
10180 	ASSERT(svd->vp == svntrp->tr_vp);
10181 	ASSERT(svd->offset == svntrp->tr_off);
10182 	ASSERT(svd->offset + seg->s_size == svntrp->tr_eoff);
10183 	ASSERT(seg != NULL);
10184 	ASSERT(svd->seg == seg);
10185 	ASSERT(seg->s_data == (void *)svd);
10186 	ASSERT(seg->s_szc == svntrp->tr_szc);
10187 	ASSERT(svd->tr_state == SEGVN_TR_ON);
10188 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
10189 	ASSERT(svd->amp != NULL);
10190 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
10191 	ASSERT(svd->tr_policy_info.mem_lgrpid != LGRP_NONE);
10192 	ASSERT(svd->tr_policy_info.mem_lgrpid < NLGRPS_MAX);
10193 	ASSERT(svntrp->tr_amp[svd->tr_policy_info.mem_lgrpid] == svd->amp);
10194 	ASSERT(svntrp->tr_refcnt != 0);
10195 	ASSERT(mutex_owned(&svntr_hashtab[hash].tr_lock));
10196 
10197 	as = seg->s_as;
10198 	ASSERT(as != NULL && as != &kas);
10199 	p = as->a_proc;
10200 	ASSERT(p != NULL);
10201 	ASSERT(p->p_tr_lgrpid != LGRP_NONE);
10202 	lgrp_id = p->p_t1_lgrpid;
10203 	if (lgrp_id == LGRP_NONE) {
10204 		return;
10205 	}
10206 	ASSERT(lgrp_id < NLGRPS_MAX);
10207 	if (svd->tr_policy_info.mem_lgrpid == lgrp_id) {
10208 		return;
10209 	}
10210 
10211 	/*
10212 	 * Use tryenter locking since we are locking as/seg and svntr hash
10213 	 * lock in reverse from syncrounous thread order.
10214 	 */
10215 	if (!AS_LOCK_TRYENTER(as, RW_READER)) {
10216 		SEGVN_TR_ADDSTAT(nolock);
10217 		if (segvn_lgrp_trthr_migrs_snpsht) {
10218 			segvn_lgrp_trthr_migrs_snpsht = 0;
10219 		}
10220 		return;
10221 	}
10222 	if (!SEGVN_LOCK_TRYENTER(seg->s_as, &svd->lock, RW_WRITER)) {
10223 		AS_LOCK_EXIT(as);
10224 		SEGVN_TR_ADDSTAT(nolock);
10225 		if (segvn_lgrp_trthr_migrs_snpsht) {
10226 			segvn_lgrp_trthr_migrs_snpsht = 0;
10227 		}
10228 		return;
10229 	}
10230 	size = seg->s_size;
10231 	if (svntrp->tr_amp[lgrp_id] == NULL) {
10232 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
10233 		if (trmem > segvn_textrepl_max_bytes) {
10234 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10235 			AS_LOCK_EXIT(as);
10236 			atomic_add_long(&segvn_textrepl_bytes, -size);
10237 			SEGVN_TR_ADDSTAT(normem);
10238 			return;
10239 		}
10240 		if (anon_try_resv_zone(size, NULL) == 0) {
10241 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10242 			AS_LOCK_EXIT(as);
10243 			atomic_add_long(&segvn_textrepl_bytes, -size);
10244 			SEGVN_TR_ADDSTAT(noanon);
10245 			return;
10246 		}
10247 		amp = anonmap_alloc(size, size, KM_NOSLEEP);
10248 		if (amp == NULL) {
10249 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10250 			AS_LOCK_EXIT(as);
10251 			atomic_add_long(&segvn_textrepl_bytes, -size);
10252 			anon_unresv_zone(size, NULL);
10253 			SEGVN_TR_ADDSTAT(nokmem);
10254 			return;
10255 		}
10256 		ASSERT(amp->refcnt == 1);
10257 		amp->a_szc = seg->s_szc;
10258 		svntrp->tr_amp[lgrp_id] = amp;
10259 	}
10260 	/*
10261 	 * We don't need to drop the bucket lock but here we give other
10262 	 * threads a chance.  svntr and svd can't be unlinked as long as
10263 	 * segment lock is held as a writer and AS held as well.  After we
10264 	 * retake bucket lock we'll continue from where we left. We'll be able
10265 	 * to reach the end of either list since new entries are always added
10266 	 * to the beginning of the lists.
10267 	 */
10268 	mutex_exit(&svntr_hashtab[hash].tr_lock);
10269 	hat_unload_callback(as->a_hat, seg->s_base, size, 0, NULL);
10270 	mutex_enter(&svntr_hashtab[hash].tr_lock);
10271 
10272 	ASSERT(svd->tr_state == SEGVN_TR_ON);
10273 	ASSERT(svd->amp != NULL);
10274 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
10275 	ASSERT(svd->tr_policy_info.mem_lgrpid != lgrp_id);
10276 	ASSERT(svd->amp != svntrp->tr_amp[lgrp_id]);
10277 
10278 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
10279 	svd->amp = svntrp->tr_amp[lgrp_id];
10280 	p->p_tr_lgrpid = NLGRPS_MAX;
10281 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10282 	AS_LOCK_EXIT(as);
10283 
10284 	ASSERT(svntrp->tr_refcnt != 0);
10285 	ASSERT(svd->vp == svntrp->tr_vp);
10286 	ASSERT(svd->tr_policy_info.mem_lgrpid == lgrp_id);
10287 	ASSERT(svd->amp != NULL && svd->amp == svntrp->tr_amp[lgrp_id]);
10288 	ASSERT(svd->seg == seg);
10289 	ASSERT(svd->tr_state == SEGVN_TR_ON);
10290 
10291 	SEGVN_TR_ADDSTAT(asyncrepl);
10292 }
10293