xref: /titanic_41/usr/src/uts/common/vm/seg_vn.c (revision f391a51a4e9639750045473dba1cc2831267c93e)
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 2014, 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 
139 struct	seg_ops segvn_ops = {
140 	segvn_dup,
141 	segvn_unmap,
142 	segvn_free,
143 	segvn_fault,
144 	segvn_faulta,
145 	segvn_setprot,
146 	segvn_checkprot,
147 	segvn_kluster,
148 	segvn_swapout,
149 	segvn_sync,
150 	segvn_incore,
151 	segvn_lockop,
152 	segvn_getprot,
153 	segvn_getoffset,
154 	segvn_gettype,
155 	segvn_getvp,
156 	segvn_advise,
157 	segvn_dump,
158 	segvn_pagelock,
159 	segvn_setpagesize,
160 	segvn_getmemid,
161 	segvn_getpolicy,
162 	segvn_capable,
163 };
164 
165 /*
166  * Common zfod structures, provided as a shorthand for others to use.
167  */
168 static segvn_crargs_t zfod_segvn_crargs =
169 	SEGVN_ZFOD_ARGS(PROT_ZFOD, PROT_ALL);
170 static segvn_crargs_t kzfod_segvn_crargs =
171 	SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_USER,
172 	PROT_ALL & ~PROT_USER);
173 static segvn_crargs_t stack_noexec_crargs =
174 	SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_EXEC, PROT_ALL);
175 
176 caddr_t	zfod_argsp = (caddr_t)&zfod_segvn_crargs;	/* user zfod argsp */
177 caddr_t	kzfod_argsp = (caddr_t)&kzfod_segvn_crargs;	/* kernel zfod argsp */
178 caddr_t	stack_exec_argsp = (caddr_t)&zfod_segvn_crargs;	/* executable stack */
179 caddr_t	stack_noexec_argsp = (caddr_t)&stack_noexec_crargs; /* noexec stack */
180 
181 #define	vpgtob(n)	((n) * sizeof (struct vpage))	/* For brevity */
182 
183 size_t	segvn_comb_thrshld = UINT_MAX;	/* patchable -- see 1196681 */
184 
185 size_t	segvn_pglock_comb_thrshld = (1UL << 16);	/* 64K */
186 size_t	segvn_pglock_comb_balign = (1UL << 16);		/* 64K */
187 uint_t	segvn_pglock_comb_bshift;
188 size_t	segvn_pglock_comb_palign;
189 
190 static int	segvn_concat(struct seg *, struct seg *, int);
191 static int	segvn_extend_prev(struct seg *, struct seg *,
192 		    struct segvn_crargs *, size_t);
193 static int	segvn_extend_next(struct seg *, struct seg *,
194 		    struct segvn_crargs *, size_t);
195 static void	segvn_softunlock(struct seg *, caddr_t, size_t, enum seg_rw);
196 static void	segvn_pagelist_rele(page_t **);
197 static void	segvn_setvnode_mpss(vnode_t *);
198 static void	segvn_relocate_pages(page_t **, page_t *);
199 static int	segvn_full_szcpages(page_t **, uint_t, int *, uint_t *);
200 static int	segvn_fill_vp_pages(struct segvn_data *, vnode_t *, u_offset_t,
201     uint_t, page_t **, page_t **, uint_t *, int *);
202 static faultcode_t segvn_fault_vnodepages(struct hat *, struct seg *, caddr_t,
203     caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
204 static faultcode_t segvn_fault_anonpages(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_faultpage(struct hat *, struct seg *, caddr_t,
207     u_offset_t, struct vpage *, page_t **, uint_t,
208     enum fault_type, enum seg_rw, int);
209 static void	segvn_vpage(struct seg *);
210 static size_t	segvn_count_swap_by_vpages(struct seg *);
211 
212 static void segvn_purge(struct seg *seg);
213 static int segvn_reclaim(void *, caddr_t, size_t, struct page **,
214     enum seg_rw, int);
215 static int shamp_reclaim(void *, caddr_t, size_t, struct page **,
216     enum seg_rw, int);
217 
218 static int sameprot(struct seg *, caddr_t, size_t);
219 
220 static int segvn_demote_range(struct seg *, caddr_t, size_t, int, uint_t);
221 static int segvn_clrszc(struct seg *);
222 static struct seg *segvn_split_seg(struct seg *, caddr_t);
223 static int segvn_claim_pages(struct seg *, struct vpage *, u_offset_t,
224     ulong_t, uint_t);
225 
226 static void segvn_hat_rgn_unload_callback(caddr_t, caddr_t, caddr_t,
227     size_t, void *, u_offset_t);
228 
229 static struct kmem_cache *segvn_cache;
230 static struct kmem_cache **segvn_szc_cache;
231 
232 #ifdef VM_STATS
233 static struct segvnvmstats_str {
234 	ulong_t	fill_vp_pages[31];
235 	ulong_t fltvnpages[49];
236 	ulong_t	fullszcpages[10];
237 	ulong_t	relocatepages[3];
238 	ulong_t	fltanpages[17];
239 	ulong_t pagelock[2];
240 	ulong_t	demoterange[3];
241 } segvnvmstats;
242 #endif /* VM_STATS */
243 
244 #define	SDR_RANGE	1		/* demote entire range */
245 #define	SDR_END		2		/* demote non aligned ends only */
246 
247 #define	CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr) {	    \
248 		if ((len) != 0) { 		      	      		      \
249 			lpgaddr = (caddr_t)P2ALIGN((uintptr_t)(addr), pgsz);  \
250 			ASSERT(lpgaddr >= (seg)->s_base);	      	      \
251 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)((addr) +    \
252 			    (len)), pgsz);				      \
253 			ASSERT(lpgeaddr > lpgaddr);		      	      \
254 			ASSERT(lpgeaddr <= (seg)->s_base + (seg)->s_size);    \
255 		} else {					      	      \
256 			lpgeaddr = lpgaddr = (addr);	      		      \
257 		}							      \
258 	}
259 
260 /*ARGSUSED*/
261 static int
262 segvn_cache_constructor(void *buf, void *cdrarg, int kmflags)
263 {
264 	struct segvn_data *svd = buf;
265 
266 	rw_init(&svd->lock, NULL, RW_DEFAULT, NULL);
267 	mutex_init(&svd->segfree_syncmtx, NULL, MUTEX_DEFAULT, NULL);
268 	svd->svn_trnext = svd->svn_trprev = NULL;
269 	return (0);
270 }
271 
272 /*ARGSUSED1*/
273 static void
274 segvn_cache_destructor(void *buf, void *cdrarg)
275 {
276 	struct segvn_data *svd = buf;
277 
278 	rw_destroy(&svd->lock);
279 	mutex_destroy(&svd->segfree_syncmtx);
280 }
281 
282 /*ARGSUSED*/
283 static int
284 svntr_cache_constructor(void *buf, void *cdrarg, int kmflags)
285 {
286 	bzero(buf, sizeof (svntr_t));
287 	return (0);
288 }
289 
290 /*
291  * Patching this variable to non-zero allows the system to run with
292  * stacks marked as "not executable".  It's a bit of a kludge, but is
293  * provided as a tweakable for platforms that export those ABIs
294  * (e.g. sparc V8) that have executable stacks enabled by default.
295  * There are also some restrictions for platforms that don't actually
296  * implement 'noexec' protections.
297  *
298  * Once enabled, the system is (therefore) unable to provide a fully
299  * ABI-compliant execution environment, though practically speaking,
300  * most everything works.  The exceptions are generally some interpreters
301  * and debuggers that create executable code on the stack and jump
302  * into it (without explicitly mprotecting the address range to include
303  * PROT_EXEC).
304  *
305  * One important class of applications that are disabled are those
306  * that have been transformed into malicious agents using one of the
307  * numerous "buffer overflow" attacks.  See 4007890.
308  */
309 int noexec_user_stack = 0;
310 int noexec_user_stack_log = 1;
311 
312 int segvn_lpg_disable = 0;
313 uint_t segvn_maxpgszc = 0;
314 
315 ulong_t segvn_vmpss_clrszc_cnt;
316 ulong_t segvn_vmpss_clrszc_err;
317 ulong_t segvn_fltvnpages_clrszc_cnt;
318 ulong_t segvn_fltvnpages_clrszc_err;
319 ulong_t segvn_setpgsz_align_err;
320 ulong_t segvn_setpgsz_anon_align_err;
321 ulong_t segvn_setpgsz_getattr_err;
322 ulong_t segvn_setpgsz_eof_err;
323 ulong_t segvn_faultvnmpss_align_err1;
324 ulong_t segvn_faultvnmpss_align_err2;
325 ulong_t segvn_faultvnmpss_align_err3;
326 ulong_t segvn_faultvnmpss_align_err4;
327 ulong_t segvn_faultvnmpss_align_err5;
328 ulong_t	segvn_vmpss_pageio_deadlk_err;
329 
330 int segvn_use_regions = 1;
331 
332 /*
333  * Segvn supports text replication optimization for NUMA platforms. Text
334  * replica's are represented by anon maps (amp). There's one amp per text file
335  * region per lgroup. A process chooses the amp for each of its text mappings
336  * based on the lgroup assignment of its main thread (t_tid = 1). All
337  * processes that want a replica on a particular lgroup for the same text file
338  * mapping share the same amp. amp's are looked up in svntr_hashtab hash table
339  * with vp,off,size,szc used as a key. Text replication segments are read only
340  * MAP_PRIVATE|MAP_TEXT segments that map vnode. Replication is achieved by
341  * forcing COW faults from vnode to amp and mapping amp pages instead of vnode
342  * pages. Replication amp is assigned to a segment when it gets its first
343  * pagefault. To handle main thread lgroup rehoming segvn_trasync_thread
344  * rechecks periodically if the process still maps an amp local to the main
345  * thread. If not async thread forces process to remap to an amp in the new
346  * home lgroup of the main thread. Current text replication implementation
347  * only provides the benefit to workloads that do most of their work in the
348  * main thread of a process or all the threads of a process run in the same
349  * lgroup. To extend text replication benefit to different types of
350  * multithreaded workloads further work would be needed in the hat layer to
351  * allow the same virtual address in the same hat to simultaneously map
352  * different physical addresses (i.e. page table replication would be needed
353  * for x86).
354  *
355  * amp pages are used instead of vnode pages as long as segment has a very
356  * simple life cycle.  It's created via segvn_create(), handles S_EXEC
357  * (S_READ) pagefaults and is fully unmapped.  If anything more complicated
358  * happens such as protection is changed, real COW fault happens, pagesize is
359  * changed, MC_LOCK is requested or segment is partially unmapped we turn off
360  * text replication by converting the segment back to vnode only segment
361  * (unmap segment's address range and set svd->amp to NULL).
362  *
363  * The original file can be changed after amp is inserted into
364  * svntr_hashtab. Processes that are launched after the file is already
365  * changed can't use the replica's created prior to the file change. To
366  * implement this functionality hash entries are timestamped. Replica's can
367  * only be used if current file modification time is the same as the timestamp
368  * saved when hash entry was created. However just timestamps alone are not
369  * sufficient to detect file modification via mmap(MAP_SHARED) mappings. We
370  * deal with file changes via MAP_SHARED mappings differently. When writable
371  * MAP_SHARED mappings are created to vnodes marked as executable we mark all
372  * existing replica's for this vnode as not usable for future text
373  * mappings. And we don't create new replica's for files that currently have
374  * potentially writable MAP_SHARED mappings (i.e. vn_is_mapped(V_WRITE) is
375  * true).
376  */
377 
378 #define	SEGVN_TEXTREPL_MAXBYTES_FACTOR	(20)
379 size_t	segvn_textrepl_max_bytes_factor = SEGVN_TEXTREPL_MAXBYTES_FACTOR;
380 
381 static ulong_t			svntr_hashtab_sz = 512;
382 static svntr_bucket_t		*svntr_hashtab = NULL;
383 static struct kmem_cache	*svntr_cache;
384 static svntr_stats_t		*segvn_textrepl_stats;
385 static ksema_t 			segvn_trasync_sem;
386 
387 int				segvn_disable_textrepl = 1;
388 size_t				textrepl_size_thresh = (size_t)-1;
389 size_t				segvn_textrepl_bytes = 0;
390 size_t				segvn_textrepl_max_bytes = 0;
391 clock_t				segvn_update_textrepl_interval = 0;
392 int				segvn_update_tr_time = 10;
393 int				segvn_disable_textrepl_update = 0;
394 
395 static void segvn_textrepl(struct seg *);
396 static void segvn_textunrepl(struct seg *, int);
397 static void segvn_inval_trcache(vnode_t *);
398 static void segvn_trasync_thread(void);
399 static void segvn_trupdate_wakeup(void *);
400 static void segvn_trupdate(void);
401 static void segvn_trupdate_seg(struct seg *, segvn_data_t *, svntr_t *,
402     ulong_t);
403 
404 /*
405  * Initialize segvn data structures
406  */
407 void
408 segvn_init(void)
409 {
410 	uint_t maxszc;
411 	uint_t szc;
412 	size_t pgsz;
413 
414 	segvn_cache = kmem_cache_create("segvn_cache",
415 	    sizeof (struct segvn_data), 0,
416 	    segvn_cache_constructor, segvn_cache_destructor, NULL,
417 	    NULL, NULL, 0);
418 
419 	if (segvn_lpg_disable == 0) {
420 		szc = maxszc = page_num_pagesizes() - 1;
421 		if (szc == 0) {
422 			segvn_lpg_disable = 1;
423 		}
424 		if (page_get_pagesize(0) != PAGESIZE) {
425 			panic("segvn_init: bad szc 0");
426 			/*NOTREACHED*/
427 		}
428 		while (szc != 0) {
429 			pgsz = page_get_pagesize(szc);
430 			if (pgsz <= PAGESIZE || !IS_P2ALIGNED(pgsz, pgsz)) {
431 				panic("segvn_init: bad szc %d", szc);
432 				/*NOTREACHED*/
433 			}
434 			szc--;
435 		}
436 		if (segvn_maxpgszc == 0 || segvn_maxpgszc > maxszc)
437 			segvn_maxpgszc = maxszc;
438 	}
439 
440 	if (segvn_maxpgszc) {
441 		segvn_szc_cache = (struct kmem_cache **)kmem_alloc(
442 		    (segvn_maxpgszc + 1) * sizeof (struct kmem_cache *),
443 		    KM_SLEEP);
444 	}
445 
446 	for (szc = 1; szc <= segvn_maxpgszc; szc++) {
447 		char	str[32];
448 
449 		(void) sprintf(str, "segvn_szc_cache%d", szc);
450 		segvn_szc_cache[szc] = kmem_cache_create(str,
451 		    page_get_pagecnt(szc) * sizeof (page_t *), 0,
452 		    NULL, NULL, NULL, NULL, NULL, KMC_NODEBUG);
453 	}
454 
455 
456 	if (segvn_use_regions && !hat_supported(HAT_SHARED_REGIONS, NULL))
457 		segvn_use_regions = 0;
458 
459 	/*
460 	 * For now shared regions and text replication segvn support
461 	 * are mutually exclusive. This is acceptable because
462 	 * currently significant benefit from text replication was
463 	 * only observed on AMD64 NUMA platforms (due to relatively
464 	 * small L2$ size) and currently we don't support shared
465 	 * regions on x86.
466 	 */
467 	if (segvn_use_regions && !segvn_disable_textrepl) {
468 		segvn_disable_textrepl = 1;
469 	}
470 
471 #if defined(_LP64)
472 	if (lgrp_optimizations() && textrepl_size_thresh != (size_t)-1 &&
473 	    !segvn_disable_textrepl) {
474 		ulong_t i;
475 		size_t hsz = svntr_hashtab_sz * sizeof (svntr_bucket_t);
476 
477 		svntr_cache = kmem_cache_create("svntr_cache",
478 		    sizeof (svntr_t), 0, svntr_cache_constructor, NULL,
479 		    NULL, NULL, NULL, 0);
480 		svntr_hashtab = kmem_zalloc(hsz, KM_SLEEP);
481 		for (i = 0; i < svntr_hashtab_sz; i++) {
482 			mutex_init(&svntr_hashtab[i].tr_lock,  NULL,
483 			    MUTEX_DEFAULT, NULL);
484 		}
485 		segvn_textrepl_max_bytes = ptob(physmem) /
486 		    segvn_textrepl_max_bytes_factor;
487 		segvn_textrepl_stats = kmem_zalloc(NCPU *
488 		    sizeof (svntr_stats_t), KM_SLEEP);
489 		sema_init(&segvn_trasync_sem, 0, NULL, SEMA_DEFAULT, NULL);
490 		(void) thread_create(NULL, 0, segvn_trasync_thread,
491 		    NULL, 0, &p0, TS_RUN, minclsyspri);
492 	}
493 #endif
494 
495 	if (!ISP2(segvn_pglock_comb_balign) ||
496 	    segvn_pglock_comb_balign < PAGESIZE) {
497 		segvn_pglock_comb_balign = 1UL << 16; /* 64K */
498 	}
499 	segvn_pglock_comb_bshift = highbit(segvn_pglock_comb_balign) - 1;
500 	segvn_pglock_comb_palign = btop(segvn_pglock_comb_balign);
501 }
502 
503 #define	SEGVN_PAGEIO	((void *)0x1)
504 #define	SEGVN_NOPAGEIO	((void *)0x2)
505 
506 static void
507 segvn_setvnode_mpss(vnode_t *vp)
508 {
509 	int err;
510 
511 	ASSERT(vp->v_mpssdata == NULL ||
512 	    vp->v_mpssdata == SEGVN_PAGEIO ||
513 	    vp->v_mpssdata == SEGVN_NOPAGEIO);
514 
515 	if (vp->v_mpssdata == NULL) {
516 		if (vn_vmpss_usepageio(vp)) {
517 			err = VOP_PAGEIO(vp, (page_t *)NULL,
518 			    (u_offset_t)0, 0, 0, CRED(), NULL);
519 		} else {
520 			err = ENOSYS;
521 		}
522 		/*
523 		 * set v_mpssdata just once per vnode life
524 		 * so that it never changes.
525 		 */
526 		mutex_enter(&vp->v_lock);
527 		if (vp->v_mpssdata == NULL) {
528 			if (err == EINVAL) {
529 				vp->v_mpssdata = SEGVN_PAGEIO;
530 			} else {
531 				vp->v_mpssdata = SEGVN_NOPAGEIO;
532 			}
533 		}
534 		mutex_exit(&vp->v_lock);
535 	}
536 }
537 
538 int
539 segvn_create(struct seg *seg, void *argsp)
540 {
541 	struct segvn_crargs *a = (struct segvn_crargs *)argsp;
542 	struct segvn_data *svd;
543 	size_t swresv = 0;
544 	struct cred *cred;
545 	struct anon_map *amp;
546 	int error = 0;
547 	size_t pgsz;
548 	lgrp_mem_policy_t mpolicy = LGRP_MEM_POLICY_DEFAULT;
549 	int use_rgn = 0;
550 	int trok = 0;
551 
552 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
553 
554 	if (a->type != MAP_PRIVATE && a->type != MAP_SHARED) {
555 		panic("segvn_create type");
556 		/*NOTREACHED*/
557 	}
558 
559 	/*
560 	 * Check arguments.  If a shared anon structure is given then
561 	 * it is illegal to also specify a vp.
562 	 */
563 	if (a->amp != NULL && a->vp != NULL) {
564 		panic("segvn_create anon_map");
565 		/*NOTREACHED*/
566 	}
567 
568 	if (a->type == MAP_PRIVATE && (a->flags & MAP_TEXT) &&
569 	    a->vp != NULL && a->prot == (PROT_USER | PROT_READ | PROT_EXEC) &&
570 	    segvn_use_regions) {
571 		use_rgn = 1;
572 	}
573 
574 	/* MAP_NORESERVE on a MAP_SHARED segment is meaningless. */
575 	if (a->type == MAP_SHARED)
576 		a->flags &= ~MAP_NORESERVE;
577 
578 	if (a->szc != 0) {
579 		if (segvn_lpg_disable != 0 || (a->szc == AS_MAP_NO_LPOOB) ||
580 		    (a->amp != NULL && a->type == MAP_PRIVATE) ||
581 		    (a->flags & MAP_NORESERVE) || seg->s_as == &kas) {
582 			a->szc = 0;
583 		} else {
584 			if (a->szc > segvn_maxpgszc)
585 				a->szc = segvn_maxpgszc;
586 			pgsz = page_get_pagesize(a->szc);
587 			if (!IS_P2ALIGNED(seg->s_base, pgsz) ||
588 			    !IS_P2ALIGNED(seg->s_size, pgsz)) {
589 				a->szc = 0;
590 			} else if (a->vp != NULL) {
591 				if (IS_SWAPFSVP(a->vp) || VN_ISKAS(a->vp)) {
592 					/*
593 					 * paranoid check.
594 					 * hat_page_demote() is not supported
595 					 * on swapfs pages.
596 					 */
597 					a->szc = 0;
598 				} else if (map_addr_vacalign_check(seg->s_base,
599 				    a->offset & PAGEMASK)) {
600 					a->szc = 0;
601 				}
602 			} else if (a->amp != NULL) {
603 				pgcnt_t anum = btopr(a->offset);
604 				pgcnt_t pgcnt = page_get_pagecnt(a->szc);
605 				if (!IS_P2ALIGNED(anum, pgcnt)) {
606 					a->szc = 0;
607 				}
608 			}
609 		}
610 	}
611 
612 	/*
613 	 * If segment may need private pages, reserve them now.
614 	 */
615 	if (!(a->flags & MAP_NORESERVE) && ((a->vp == NULL && a->amp == NULL) ||
616 	    (a->type == MAP_PRIVATE && (a->prot & PROT_WRITE)))) {
617 		if (anon_resv_zone(seg->s_size,
618 		    seg->s_as->a_proc->p_zone) == 0)
619 			return (EAGAIN);
620 		swresv = seg->s_size;
621 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
622 		    seg, swresv, 1);
623 	}
624 
625 	/*
626 	 * Reserve any mapping structures that may be required.
627 	 *
628 	 * Don't do it for segments that may use regions. It's currently a
629 	 * noop in the hat implementations anyway.
630 	 */
631 	if (!use_rgn) {
632 		hat_map(seg->s_as->a_hat, seg->s_base, seg->s_size, HAT_MAP);
633 	}
634 
635 	if (a->cred) {
636 		cred = a->cred;
637 		crhold(cred);
638 	} else {
639 		crhold(cred = CRED());
640 	}
641 
642 	/* Inform the vnode of the new mapping */
643 	if (a->vp != NULL) {
644 		error = VOP_ADDMAP(a->vp, a->offset & PAGEMASK,
645 		    seg->s_as, seg->s_base, seg->s_size, a->prot,
646 		    a->maxprot, a->type, cred, NULL);
647 		if (error) {
648 			if (swresv != 0) {
649 				anon_unresv_zone(swresv,
650 				    seg->s_as->a_proc->p_zone);
651 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
652 				    "anon proc:%p %lu %u", seg, swresv, 0);
653 			}
654 			crfree(cred);
655 			if (!use_rgn) {
656 				hat_unload(seg->s_as->a_hat, seg->s_base,
657 				    seg->s_size, HAT_UNLOAD_UNMAP);
658 			}
659 			return (error);
660 		}
661 		/*
662 		 * svntr_hashtab will be NULL if we support shared regions.
663 		 */
664 		trok = ((a->flags & MAP_TEXT) &&
665 		    (seg->s_size > textrepl_size_thresh ||
666 		    (a->flags & _MAP_TEXTREPL)) &&
667 		    lgrp_optimizations() && svntr_hashtab != NULL &&
668 		    a->type == MAP_PRIVATE && swresv == 0 &&
669 		    !(a->flags & MAP_NORESERVE) &&
670 		    seg->s_as != &kas && a->vp->v_type == VREG);
671 
672 		ASSERT(!trok || !use_rgn);
673 	}
674 
675 	/*
676 	 * MAP_NORESERVE mappings don't count towards the VSZ of a process
677 	 * until we fault the pages in.
678 	 */
679 	if ((a->vp == NULL || a->vp->v_type != VREG) &&
680 	    a->flags & MAP_NORESERVE) {
681 		seg->s_as->a_resvsize -= seg->s_size;
682 	}
683 
684 	/*
685 	 * If more than one segment in the address space, and they're adjacent
686 	 * virtually, try to concatenate them.  Don't concatenate if an
687 	 * explicit anon_map structure was supplied (e.g., SystemV shared
688 	 * memory) or if we'll use text replication for this segment.
689 	 */
690 	if (a->amp == NULL && !use_rgn && !trok) {
691 		struct seg *pseg, *nseg;
692 		struct segvn_data *psvd, *nsvd;
693 		lgrp_mem_policy_t ppolicy, npolicy;
694 		uint_t	lgrp_mem_policy_flags = 0;
695 		extern lgrp_mem_policy_t lgrp_mem_default_policy;
696 
697 		/*
698 		 * Memory policy flags (lgrp_mem_policy_flags) is valid when
699 		 * extending stack/heap segments.
700 		 */
701 		if ((a->vp == NULL) && (a->type == MAP_PRIVATE) &&
702 		    !(a->flags & MAP_NORESERVE) && (seg->s_as != &kas)) {
703 			lgrp_mem_policy_flags = a->lgrp_mem_policy_flags;
704 		} else {
705 			/*
706 			 * Get policy when not extending it from another segment
707 			 */
708 			mpolicy = lgrp_mem_policy_default(seg->s_size, a->type);
709 		}
710 
711 		/*
712 		 * First, try to concatenate the previous and new segments
713 		 */
714 		pseg = AS_SEGPREV(seg->s_as, seg);
715 		if (pseg != NULL &&
716 		    pseg->s_base + pseg->s_size == seg->s_base &&
717 		    pseg->s_ops == &segvn_ops) {
718 			/*
719 			 * Get memory allocation policy from previous segment.
720 			 * When extension is specified (e.g. for heap) apply
721 			 * this policy to the new segment regardless of the
722 			 * outcome of segment concatenation.  Extension occurs
723 			 * for non-default policy otherwise default policy is
724 			 * used and is based on extended segment size.
725 			 */
726 			psvd = (struct segvn_data *)pseg->s_data;
727 			ppolicy = psvd->policy_info.mem_policy;
728 			if (lgrp_mem_policy_flags ==
729 			    LGRP_MP_FLAG_EXTEND_UP) {
730 				if (ppolicy != lgrp_mem_default_policy) {
731 					mpolicy = ppolicy;
732 				} else {
733 					mpolicy = lgrp_mem_policy_default(
734 					    pseg->s_size + seg->s_size,
735 					    a->type);
736 				}
737 			}
738 
739 			if (mpolicy == ppolicy &&
740 			    (pseg->s_size + seg->s_size <=
741 			    segvn_comb_thrshld || psvd->amp == NULL) &&
742 			    segvn_extend_prev(pseg, seg, a, swresv) == 0) {
743 				/*
744 				 * success! now try to concatenate
745 				 * with following seg
746 				 */
747 				crfree(cred);
748 				nseg = AS_SEGNEXT(pseg->s_as, pseg);
749 				if (nseg != NULL &&
750 				    nseg != pseg &&
751 				    nseg->s_ops == &segvn_ops &&
752 				    pseg->s_base + pseg->s_size ==
753 				    nseg->s_base)
754 					(void) segvn_concat(pseg, nseg, 0);
755 				ASSERT(pseg->s_szc == 0 ||
756 				    (a->szc == pseg->s_szc &&
757 				    IS_P2ALIGNED(pseg->s_base, pgsz) &&
758 				    IS_P2ALIGNED(pseg->s_size, pgsz)));
759 				return (0);
760 			}
761 		}
762 
763 		/*
764 		 * Failed, so try to concatenate with following seg
765 		 */
766 		nseg = AS_SEGNEXT(seg->s_as, seg);
767 		if (nseg != NULL &&
768 		    seg->s_base + seg->s_size == nseg->s_base &&
769 		    nseg->s_ops == &segvn_ops) {
770 			/*
771 			 * Get memory allocation policy from next segment.
772 			 * When extension is specified (e.g. for stack) apply
773 			 * this policy to the new segment regardless of the
774 			 * outcome of segment concatenation.  Extension occurs
775 			 * for non-default policy otherwise default policy is
776 			 * used and is based on extended segment size.
777 			 */
778 			nsvd = (struct segvn_data *)nseg->s_data;
779 			npolicy = nsvd->policy_info.mem_policy;
780 			if (lgrp_mem_policy_flags ==
781 			    LGRP_MP_FLAG_EXTEND_DOWN) {
782 				if (npolicy != lgrp_mem_default_policy) {
783 					mpolicy = npolicy;
784 				} else {
785 					mpolicy = lgrp_mem_policy_default(
786 					    nseg->s_size + seg->s_size,
787 					    a->type);
788 				}
789 			}
790 
791 			if (mpolicy == npolicy &&
792 			    segvn_extend_next(seg, nseg, a, swresv) == 0) {
793 				crfree(cred);
794 				ASSERT(nseg->s_szc == 0 ||
795 				    (a->szc == nseg->s_szc &&
796 				    IS_P2ALIGNED(nseg->s_base, pgsz) &&
797 				    IS_P2ALIGNED(nseg->s_size, pgsz)));
798 				return (0);
799 			}
800 		}
801 	}
802 
803 	if (a->vp != NULL) {
804 		VN_HOLD(a->vp);
805 		if (a->type == MAP_SHARED)
806 			lgrp_shm_policy_init(NULL, a->vp);
807 	}
808 	svd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
809 
810 	seg->s_ops = &segvn_ops;
811 	seg->s_data = (void *)svd;
812 	seg->s_szc = a->szc;
813 
814 	svd->seg = seg;
815 	svd->vp = a->vp;
816 	/*
817 	 * Anonymous mappings have no backing file so the offset is meaningless.
818 	 */
819 	svd->offset = a->vp ? (a->offset & PAGEMASK) : 0;
820 	svd->prot = a->prot;
821 	svd->maxprot = a->maxprot;
822 	svd->pageprot = 0;
823 	svd->type = a->type;
824 	svd->vpage = NULL;
825 	svd->cred = cred;
826 	svd->advice = MADV_NORMAL;
827 	svd->pageadvice = 0;
828 	svd->flags = (ushort_t)a->flags;
829 	svd->softlockcnt = 0;
830 	svd->softlockcnt_sbase = 0;
831 	svd->softlockcnt_send = 0;
832 	svd->rcookie = HAT_INVALID_REGION_COOKIE;
833 	svd->pageswap = 0;
834 
835 	if (a->szc != 0 && a->vp != NULL) {
836 		segvn_setvnode_mpss(a->vp);
837 	}
838 	if (svd->type == MAP_SHARED && svd->vp != NULL &&
839 	    (svd->vp->v_flag & VVMEXEC) && (svd->prot & PROT_WRITE)) {
840 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
841 		segvn_inval_trcache(svd->vp);
842 	}
843 
844 	amp = a->amp;
845 	if ((svd->amp = amp) == NULL) {
846 		svd->anon_index = 0;
847 		if (svd->type == MAP_SHARED) {
848 			svd->swresv = 0;
849 			/*
850 			 * Shared mappings to a vp need no other setup.
851 			 * If we have a shared mapping to an anon_map object
852 			 * which hasn't been allocated yet,  allocate the
853 			 * struct now so that it will be properly shared
854 			 * by remembering the swap reservation there.
855 			 */
856 			if (a->vp == NULL) {
857 				svd->amp = anonmap_alloc(seg->s_size, swresv,
858 				    ANON_SLEEP);
859 				svd->amp->a_szc = seg->s_szc;
860 			}
861 		} else {
862 			/*
863 			 * Private mapping (with or without a vp).
864 			 * Allocate anon_map when needed.
865 			 */
866 			svd->swresv = swresv;
867 		}
868 	} else {
869 		pgcnt_t anon_num;
870 
871 		/*
872 		 * Mapping to an existing anon_map structure without a vp.
873 		 * For now we will insure that the segment size isn't larger
874 		 * than the size - offset gives us.  Later on we may wish to
875 		 * have the anon array dynamically allocated itself so that
876 		 * we don't always have to allocate all the anon pointer slots.
877 		 * This of course involves adding extra code to check that we
878 		 * aren't trying to use an anon pointer slot beyond the end
879 		 * of the currently allocated anon array.
880 		 */
881 		if ((amp->size - a->offset) < seg->s_size) {
882 			panic("segvn_create anon_map size");
883 			/*NOTREACHED*/
884 		}
885 
886 		anon_num = btopr(a->offset);
887 
888 		if (a->type == MAP_SHARED) {
889 			/*
890 			 * SHARED mapping to a given anon_map.
891 			 */
892 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
893 			amp->refcnt++;
894 			if (a->szc > amp->a_szc) {
895 				amp->a_szc = a->szc;
896 			}
897 			ANON_LOCK_EXIT(&amp->a_rwlock);
898 			svd->anon_index = anon_num;
899 			svd->swresv = 0;
900 		} else {
901 			/*
902 			 * PRIVATE mapping to a given anon_map.
903 			 * Make sure that all the needed anon
904 			 * structures are created (so that we will
905 			 * share the underlying pages if nothing
906 			 * is written by this mapping) and then
907 			 * duplicate the anon array as is done
908 			 * when a privately mapped segment is dup'ed.
909 			 */
910 			struct anon *ap;
911 			caddr_t addr;
912 			caddr_t eaddr;
913 			ulong_t	anon_idx;
914 			int hat_flag = HAT_LOAD;
915 
916 			if (svd->flags & MAP_TEXT) {
917 				hat_flag |= HAT_LOAD_TEXT;
918 			}
919 
920 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
921 			svd->amp->a_szc = seg->s_szc;
922 			svd->anon_index = 0;
923 			svd->swresv = swresv;
924 
925 			/*
926 			 * Prevent 2 threads from allocating anon
927 			 * slots simultaneously.
928 			 */
929 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
930 			eaddr = seg->s_base + seg->s_size;
931 
932 			for (anon_idx = anon_num, addr = seg->s_base;
933 			    addr < eaddr; addr += PAGESIZE, anon_idx++) {
934 				page_t *pp;
935 
936 				if ((ap = anon_get_ptr(amp->ahp,
937 				    anon_idx)) != NULL)
938 					continue;
939 
940 				/*
941 				 * Allocate the anon struct now.
942 				 * Might as well load up translation
943 				 * to the page while we're at it...
944 				 */
945 				pp = anon_zero(seg, addr, &ap, cred);
946 				if (ap == NULL || pp == NULL) {
947 					panic("segvn_create anon_zero");
948 					/*NOTREACHED*/
949 				}
950 
951 				/*
952 				 * Re-acquire the anon_map lock and
953 				 * initialize the anon array entry.
954 				 */
955 				ASSERT(anon_get_ptr(amp->ahp,
956 				    anon_idx) == NULL);
957 				(void) anon_set_ptr(amp->ahp, anon_idx, ap,
958 				    ANON_SLEEP);
959 
960 				ASSERT(seg->s_szc == 0);
961 				ASSERT(!IS_VMODSORT(pp->p_vnode));
962 
963 				ASSERT(use_rgn == 0);
964 				hat_memload(seg->s_as->a_hat, addr, pp,
965 				    svd->prot & ~PROT_WRITE, hat_flag);
966 
967 				page_unlock(pp);
968 			}
969 			ASSERT(seg->s_szc == 0);
970 			anon_dup(amp->ahp, anon_num, svd->amp->ahp,
971 			    0, seg->s_size);
972 			ANON_LOCK_EXIT(&amp->a_rwlock);
973 		}
974 	}
975 
976 	/*
977 	 * Set default memory allocation policy for segment
978 	 *
979 	 * Always set policy for private memory at least for initialization
980 	 * even if this is a shared memory segment
981 	 */
982 	(void) lgrp_privm_policy_set(mpolicy, &svd->policy_info, seg->s_size);
983 
984 	if (svd->type == MAP_SHARED)
985 		(void) lgrp_shm_policy_set(mpolicy, svd->amp, svd->anon_index,
986 		    svd->vp, svd->offset, seg->s_size);
987 
988 	if (use_rgn) {
989 		ASSERT(!trok);
990 		ASSERT(svd->amp == NULL);
991 		svd->rcookie = hat_join_region(seg->s_as->a_hat, seg->s_base,
992 		    seg->s_size, (void *)svd->vp, svd->offset, svd->prot,
993 		    (uchar_t)seg->s_szc, segvn_hat_rgn_unload_callback,
994 		    HAT_REGION_TEXT);
995 	}
996 
997 	ASSERT(!trok || !(svd->prot & PROT_WRITE));
998 	svd->tr_state = trok ? SEGVN_TR_INIT : SEGVN_TR_OFF;
999 
1000 	return (0);
1001 }
1002 
1003 /*
1004  * Concatenate two existing segments, if possible.
1005  * Return 0 on success, -1 if two segments are not compatible
1006  * or -2 on memory allocation failure.
1007  * If amp_cat == 1 then try and concat segments with anon maps
1008  */
1009 static int
1010 segvn_concat(struct seg *seg1, struct seg *seg2, int amp_cat)
1011 {
1012 	struct segvn_data *svd1 = seg1->s_data;
1013 	struct segvn_data *svd2 = seg2->s_data;
1014 	struct anon_map *amp1 = svd1->amp;
1015 	struct anon_map *amp2 = svd2->amp;
1016 	struct vpage *vpage1 = svd1->vpage;
1017 	struct vpage *vpage2 = svd2->vpage, *nvpage = NULL;
1018 	size_t size, nvpsize;
1019 	pgcnt_t npages1, npages2;
1020 
1021 	ASSERT(seg1->s_as && seg2->s_as && seg1->s_as == seg2->s_as);
1022 	ASSERT(AS_WRITE_HELD(seg1->s_as, &seg1->s_as->a_lock));
1023 	ASSERT(seg1->s_ops == seg2->s_ops);
1024 
1025 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie) ||
1026 	    HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1027 		return (-1);
1028 	}
1029 
1030 	/* both segments exist, try to merge them */
1031 #define	incompat(x)	(svd1->x != svd2->x)
1032 	if (incompat(vp) || incompat(maxprot) ||
1033 	    (!svd1->pageadvice && !svd2->pageadvice && incompat(advice)) ||
1034 	    (!svd1->pageprot && !svd2->pageprot && incompat(prot)) ||
1035 	    incompat(type) || incompat(cred) || incompat(flags) ||
1036 	    seg1->s_szc != seg2->s_szc || incompat(policy_info.mem_policy) ||
1037 	    (svd2->softlockcnt > 0) || svd1->softlockcnt_send > 0)
1038 		return (-1);
1039 #undef incompat
1040 
1041 	/*
1042 	 * vp == NULL implies zfod, offset doesn't matter
1043 	 */
1044 	if (svd1->vp != NULL &&
1045 	    svd1->offset + seg1->s_size != svd2->offset) {
1046 		return (-1);
1047 	}
1048 
1049 	/*
1050 	 * Don't concatenate if either segment uses text replication.
1051 	 */
1052 	if (svd1->tr_state != SEGVN_TR_OFF || svd2->tr_state != SEGVN_TR_OFF) {
1053 		return (-1);
1054 	}
1055 
1056 	/*
1057 	 * Fail early if we're not supposed to concatenate
1058 	 * segments with non NULL amp.
1059 	 */
1060 	if (amp_cat == 0 && (amp1 != NULL || amp2 != NULL)) {
1061 		return (-1);
1062 	}
1063 
1064 	if (svd1->vp == NULL && svd1->type == MAP_SHARED) {
1065 		if (amp1 != amp2) {
1066 			return (-1);
1067 		}
1068 		if (amp1 != NULL && svd1->anon_index + btop(seg1->s_size) !=
1069 		    svd2->anon_index) {
1070 			return (-1);
1071 		}
1072 		ASSERT(amp1 == NULL || amp1->refcnt >= 2);
1073 	}
1074 
1075 	/*
1076 	 * If either seg has vpages, create a new merged vpage array.
1077 	 */
1078 	if (vpage1 != NULL || vpage2 != NULL) {
1079 		struct vpage *vp, *evp;
1080 
1081 		npages1 = seg_pages(seg1);
1082 		npages2 = seg_pages(seg2);
1083 		nvpsize = vpgtob(npages1 + npages2);
1084 
1085 		if ((nvpage = kmem_zalloc(nvpsize, KM_NOSLEEP)) == NULL) {
1086 			return (-2);
1087 		}
1088 
1089 		if (vpage1 != NULL) {
1090 			bcopy(vpage1, nvpage, vpgtob(npages1));
1091 		} else {
1092 			evp = nvpage + npages1;
1093 			for (vp = nvpage; vp < evp; vp++) {
1094 				VPP_SETPROT(vp, svd1->prot);
1095 				VPP_SETADVICE(vp, svd1->advice);
1096 			}
1097 		}
1098 
1099 		if (vpage2 != NULL) {
1100 			bcopy(vpage2, nvpage + npages1, vpgtob(npages2));
1101 		} else {
1102 			evp = nvpage + npages1 + npages2;
1103 			for (vp = nvpage + npages1; vp < evp; vp++) {
1104 				VPP_SETPROT(vp, svd2->prot);
1105 				VPP_SETADVICE(vp, svd2->advice);
1106 			}
1107 		}
1108 
1109 		if (svd2->pageswap && (!svd1->pageswap && svd1->swresv)) {
1110 			ASSERT(svd1->swresv == seg1->s_size);
1111 			ASSERT(!(svd1->flags & MAP_NORESERVE));
1112 			ASSERT(!(svd2->flags & MAP_NORESERVE));
1113 			evp = nvpage + npages1;
1114 			for (vp = nvpage; vp < evp; vp++) {
1115 				VPP_SETSWAPRES(vp);
1116 			}
1117 		}
1118 
1119 		if (svd1->pageswap && (!svd2->pageswap && svd2->swresv)) {
1120 			ASSERT(svd2->swresv == seg2->s_size);
1121 			ASSERT(!(svd1->flags & MAP_NORESERVE));
1122 			ASSERT(!(svd2->flags & MAP_NORESERVE));
1123 			vp = nvpage + npages1;
1124 			evp = vp + npages2;
1125 			for (; vp < evp; vp++) {
1126 				VPP_SETSWAPRES(vp);
1127 			}
1128 		}
1129 	}
1130 	ASSERT((vpage1 != NULL || vpage2 != NULL) ||
1131 	    (svd1->pageswap == 0 && svd2->pageswap == 0));
1132 
1133 	/*
1134 	 * If either segment has private pages, create a new merged anon
1135 	 * array. If mergeing shared anon segments just decrement anon map's
1136 	 * refcnt.
1137 	 */
1138 	if (amp1 != NULL && svd1->type == MAP_SHARED) {
1139 		ASSERT(amp1 == amp2 && svd1->vp == NULL);
1140 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1141 		ASSERT(amp1->refcnt >= 2);
1142 		amp1->refcnt--;
1143 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1144 		svd2->amp = NULL;
1145 	} else if (amp1 != NULL || amp2 != NULL) {
1146 		struct anon_hdr *nahp;
1147 		struct anon_map *namp = NULL;
1148 		size_t asize;
1149 
1150 		ASSERT(svd1->type == MAP_PRIVATE);
1151 
1152 		asize = seg1->s_size + seg2->s_size;
1153 		if ((nahp = anon_create(btop(asize), ANON_NOSLEEP)) == NULL) {
1154 			if (nvpage != NULL) {
1155 				kmem_free(nvpage, nvpsize);
1156 			}
1157 			return (-2);
1158 		}
1159 		if (amp1 != NULL) {
1160 			/*
1161 			 * XXX anon rwlock is not really needed because
1162 			 * this is a private segment and we are writers.
1163 			 */
1164 			ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1165 			ASSERT(amp1->refcnt == 1);
1166 			if (anon_copy_ptr(amp1->ahp, svd1->anon_index,
1167 			    nahp, 0, btop(seg1->s_size), ANON_NOSLEEP)) {
1168 				anon_release(nahp, btop(asize));
1169 				ANON_LOCK_EXIT(&amp1->a_rwlock);
1170 				if (nvpage != NULL) {
1171 					kmem_free(nvpage, nvpsize);
1172 				}
1173 				return (-2);
1174 			}
1175 		}
1176 		if (amp2 != NULL) {
1177 			ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1178 			ASSERT(amp2->refcnt == 1);
1179 			if (anon_copy_ptr(amp2->ahp, svd2->anon_index,
1180 			    nahp, btop(seg1->s_size), btop(seg2->s_size),
1181 			    ANON_NOSLEEP)) {
1182 				anon_release(nahp, btop(asize));
1183 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1184 				if (amp1 != NULL) {
1185 					ANON_LOCK_EXIT(&amp1->a_rwlock);
1186 				}
1187 				if (nvpage != NULL) {
1188 					kmem_free(nvpage, nvpsize);
1189 				}
1190 				return (-2);
1191 			}
1192 		}
1193 		if (amp1 != NULL) {
1194 			namp = amp1;
1195 			anon_release(amp1->ahp, btop(amp1->size));
1196 		}
1197 		if (amp2 != NULL) {
1198 			if (namp == NULL) {
1199 				ASSERT(amp1 == NULL);
1200 				namp = amp2;
1201 				anon_release(amp2->ahp, btop(amp2->size));
1202 			} else {
1203 				amp2->refcnt--;
1204 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1205 				anonmap_free(amp2);
1206 			}
1207 			svd2->amp = NULL; /* needed for seg_free */
1208 		}
1209 		namp->ahp = nahp;
1210 		namp->size = asize;
1211 		svd1->amp = namp;
1212 		svd1->anon_index = 0;
1213 		ANON_LOCK_EXIT(&namp->a_rwlock);
1214 	}
1215 	/*
1216 	 * Now free the old vpage structures.
1217 	 */
1218 	if (nvpage != NULL) {
1219 		if (vpage1 != NULL) {
1220 			kmem_free(vpage1, vpgtob(npages1));
1221 		}
1222 		if (vpage2 != NULL) {
1223 			svd2->vpage = NULL;
1224 			kmem_free(vpage2, vpgtob(npages2));
1225 		}
1226 		if (svd2->pageprot) {
1227 			svd1->pageprot = 1;
1228 		}
1229 		if (svd2->pageadvice) {
1230 			svd1->pageadvice = 1;
1231 		}
1232 		if (svd2->pageswap) {
1233 			svd1->pageswap = 1;
1234 		}
1235 		svd1->vpage = nvpage;
1236 	}
1237 
1238 	/* all looks ok, merge segments */
1239 	svd1->swresv += svd2->swresv;
1240 	svd2->swresv = 0;  /* so seg_free doesn't release swap space */
1241 	size = seg2->s_size;
1242 	seg_free(seg2);
1243 	seg1->s_size += size;
1244 	return (0);
1245 }
1246 
1247 /*
1248  * Extend the previous segment (seg1) to include the
1249  * new segment (seg2 + a), if possible.
1250  * Return 0 on success.
1251  */
1252 static int
1253 segvn_extend_prev(seg1, seg2, a, swresv)
1254 	struct seg *seg1, *seg2;
1255 	struct segvn_crargs *a;
1256 	size_t swresv;
1257 {
1258 	struct segvn_data *svd1 = (struct segvn_data *)seg1->s_data;
1259 	size_t size;
1260 	struct anon_map *amp1;
1261 	struct vpage *new_vpage;
1262 
1263 	/*
1264 	 * We don't need any segment level locks for "segvn" data
1265 	 * since the address space is "write" locked.
1266 	 */
1267 	ASSERT(seg1->s_as && AS_WRITE_HELD(seg1->s_as, &seg1->s_as->a_lock));
1268 
1269 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie)) {
1270 		return (-1);
1271 	}
1272 
1273 	/* second segment is new, try to extend first */
1274 	/* XXX - should also check cred */
1275 	if (svd1->vp != a->vp || svd1->maxprot != a->maxprot ||
1276 	    (!svd1->pageprot && (svd1->prot != a->prot)) ||
1277 	    svd1->type != a->type || svd1->flags != a->flags ||
1278 	    seg1->s_szc != a->szc || svd1->softlockcnt_send > 0)
1279 		return (-1);
1280 
1281 	/* vp == NULL implies zfod, offset doesn't matter */
1282 	if (svd1->vp != NULL &&
1283 	    svd1->offset + seg1->s_size != (a->offset & PAGEMASK))
1284 		return (-1);
1285 
1286 	if (svd1->tr_state != SEGVN_TR_OFF) {
1287 		return (-1);
1288 	}
1289 
1290 	amp1 = svd1->amp;
1291 	if (amp1) {
1292 		pgcnt_t newpgs;
1293 
1294 		/*
1295 		 * Segment has private pages, can data structures
1296 		 * be expanded?
1297 		 *
1298 		 * Acquire the anon_map lock to prevent it from changing,
1299 		 * if it is shared.  This ensures that the anon_map
1300 		 * will not change while a thread which has a read/write
1301 		 * lock on an address space references it.
1302 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1303 		 * is 1.
1304 		 *
1305 		 * Can't grow a MAP_SHARED segment with an anonmap because
1306 		 * there may be existing anon slots where we want to extend
1307 		 * the segment and we wouldn't know what to do with them
1308 		 * (e.g., for tmpfs right thing is to just leave them there,
1309 		 * for /dev/zero they should be cleared out).
1310 		 */
1311 		if (svd1->type == MAP_SHARED)
1312 			return (-1);
1313 
1314 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1315 		if (amp1->refcnt > 1) {
1316 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1317 			return (-1);
1318 		}
1319 		newpgs = anon_grow(amp1->ahp, &svd1->anon_index,
1320 		    btop(seg1->s_size), btop(seg2->s_size), ANON_NOSLEEP);
1321 
1322 		if (newpgs == 0) {
1323 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1324 			return (-1);
1325 		}
1326 		amp1->size = ptob(newpgs);
1327 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1328 	}
1329 	if (svd1->vpage != NULL) {
1330 		struct vpage *vp, *evp;
1331 		new_vpage =
1332 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1333 			KM_NOSLEEP);
1334 		if (new_vpage == NULL)
1335 			return (-1);
1336 		bcopy(svd1->vpage, new_vpage, vpgtob(seg_pages(seg1)));
1337 		kmem_free(svd1->vpage, vpgtob(seg_pages(seg1)));
1338 		svd1->vpage = new_vpage;
1339 
1340 		vp = new_vpage + seg_pages(seg1);
1341 		evp = vp + seg_pages(seg2);
1342 		for (; vp < evp; vp++)
1343 			VPP_SETPROT(vp, a->prot);
1344 		if (svd1->pageswap && swresv) {
1345 			ASSERT(!(svd1->flags & MAP_NORESERVE));
1346 			ASSERT(swresv == seg2->s_size);
1347 			vp = new_vpage + seg_pages(seg1);
1348 			for (; vp < evp; vp++) {
1349 				VPP_SETSWAPRES(vp);
1350 			}
1351 		}
1352 	}
1353 	ASSERT(svd1->vpage != NULL || svd1->pageswap == 0);
1354 	size = seg2->s_size;
1355 	seg_free(seg2);
1356 	seg1->s_size += size;
1357 	svd1->swresv += swresv;
1358 	if (svd1->pageprot && (a->prot & PROT_WRITE) &&
1359 	    svd1->type == MAP_SHARED && svd1->vp != NULL &&
1360 	    (svd1->vp->v_flag & VVMEXEC)) {
1361 		ASSERT(vn_is_mapped(svd1->vp, V_WRITE));
1362 		segvn_inval_trcache(svd1->vp);
1363 	}
1364 	return (0);
1365 }
1366 
1367 /*
1368  * Extend the next segment (seg2) to include the
1369  * new segment (seg1 + a), if possible.
1370  * Return 0 on success.
1371  */
1372 static int
1373 segvn_extend_next(
1374 	struct seg *seg1,
1375 	struct seg *seg2,
1376 	struct segvn_crargs *a,
1377 	size_t swresv)
1378 {
1379 	struct segvn_data *svd2 = (struct segvn_data *)seg2->s_data;
1380 	size_t size;
1381 	struct anon_map *amp2;
1382 	struct vpage *new_vpage;
1383 
1384 	/*
1385 	 * We don't need any segment level locks for "segvn" data
1386 	 * since the address space is "write" locked.
1387 	 */
1388 	ASSERT(seg2->s_as && AS_WRITE_HELD(seg2->s_as, &seg2->s_as->a_lock));
1389 
1390 	if (HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1391 		return (-1);
1392 	}
1393 
1394 	/* first segment is new, try to extend second */
1395 	/* XXX - should also check cred */
1396 	if (svd2->vp != a->vp || svd2->maxprot != a->maxprot ||
1397 	    (!svd2->pageprot && (svd2->prot != a->prot)) ||
1398 	    svd2->type != a->type || svd2->flags != a->flags ||
1399 	    seg2->s_szc != a->szc || svd2->softlockcnt_sbase > 0)
1400 		return (-1);
1401 	/* vp == NULL implies zfod, offset doesn't matter */
1402 	if (svd2->vp != NULL &&
1403 	    (a->offset & PAGEMASK) + seg1->s_size != svd2->offset)
1404 		return (-1);
1405 
1406 	if (svd2->tr_state != SEGVN_TR_OFF) {
1407 		return (-1);
1408 	}
1409 
1410 	amp2 = svd2->amp;
1411 	if (amp2) {
1412 		pgcnt_t newpgs;
1413 
1414 		/*
1415 		 * Segment has private pages, can data structures
1416 		 * be expanded?
1417 		 *
1418 		 * Acquire the anon_map lock to prevent it from changing,
1419 		 * if it is shared.  This ensures that the anon_map
1420 		 * will not change while a thread which has a read/write
1421 		 * lock on an address space references it.
1422 		 *
1423 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1424 		 * is 1.
1425 		 */
1426 		if (svd2->type == MAP_SHARED)
1427 			return (-1);
1428 
1429 		ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1430 		if (amp2->refcnt > 1) {
1431 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1432 			return (-1);
1433 		}
1434 		newpgs = anon_grow(amp2->ahp, &svd2->anon_index,
1435 		    btop(seg2->s_size), btop(seg1->s_size),
1436 		    ANON_NOSLEEP | ANON_GROWDOWN);
1437 
1438 		if (newpgs == 0) {
1439 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1440 			return (-1);
1441 		}
1442 		amp2->size = ptob(newpgs);
1443 		ANON_LOCK_EXIT(&amp2->a_rwlock);
1444 	}
1445 	if (svd2->vpage != NULL) {
1446 		struct vpage *vp, *evp;
1447 		new_vpage =
1448 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1449 		    KM_NOSLEEP);
1450 		if (new_vpage == NULL) {
1451 			/* Not merging segments so adjust anon_index back */
1452 			if (amp2)
1453 				svd2->anon_index += seg_pages(seg1);
1454 			return (-1);
1455 		}
1456 		bcopy(svd2->vpage, new_vpage + seg_pages(seg1),
1457 		    vpgtob(seg_pages(seg2)));
1458 		kmem_free(svd2->vpage, vpgtob(seg_pages(seg2)));
1459 		svd2->vpage = new_vpage;
1460 
1461 		vp = new_vpage;
1462 		evp = vp + seg_pages(seg1);
1463 		for (; vp < evp; vp++)
1464 			VPP_SETPROT(vp, a->prot);
1465 		if (svd2->pageswap && swresv) {
1466 			ASSERT(!(svd2->flags & MAP_NORESERVE));
1467 			ASSERT(swresv == seg1->s_size);
1468 			vp = new_vpage;
1469 			for (; vp < evp; vp++) {
1470 				VPP_SETSWAPRES(vp);
1471 			}
1472 		}
1473 	}
1474 	ASSERT(svd2->vpage != NULL || svd2->pageswap == 0);
1475 	size = seg1->s_size;
1476 	seg_free(seg1);
1477 	seg2->s_size += size;
1478 	seg2->s_base -= size;
1479 	svd2->offset -= size;
1480 	svd2->swresv += swresv;
1481 	if (svd2->pageprot && (a->prot & PROT_WRITE) &&
1482 	    svd2->type == MAP_SHARED && svd2->vp != NULL &&
1483 	    (svd2->vp->v_flag & VVMEXEC)) {
1484 		ASSERT(vn_is_mapped(svd2->vp, V_WRITE));
1485 		segvn_inval_trcache(svd2->vp);
1486 	}
1487 	return (0);
1488 }
1489 
1490 static int
1491 segvn_dup(struct seg *seg, struct seg *newseg)
1492 {
1493 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1494 	struct segvn_data *newsvd;
1495 	pgcnt_t npages = seg_pages(seg);
1496 	int error = 0;
1497 	uint_t prot;
1498 	size_t len;
1499 	struct anon_map *amp;
1500 
1501 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
1502 	ASSERT(newseg->s_as->a_proc->p_parent == curproc);
1503 
1504 	/*
1505 	 * If segment has anon reserved, reserve more for the new seg.
1506 	 * For a MAP_NORESERVE segment swresv will be a count of all the
1507 	 * allocated anon slots; thus we reserve for the child as many slots
1508 	 * as the parent has allocated. This semantic prevents the child or
1509 	 * parent from dieing during a copy-on-write fault caused by trying
1510 	 * to write a shared pre-existing anon page.
1511 	 */
1512 	if ((len = svd->swresv) != 0) {
1513 		if (anon_resv(svd->swresv) == 0)
1514 			return (ENOMEM);
1515 
1516 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
1517 		    seg, len, 0);
1518 	}
1519 
1520 	newsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
1521 
1522 	newseg->s_ops = &segvn_ops;
1523 	newseg->s_data = (void *)newsvd;
1524 	newseg->s_szc = seg->s_szc;
1525 
1526 	newsvd->seg = newseg;
1527 	if ((newsvd->vp = svd->vp) != NULL) {
1528 		VN_HOLD(svd->vp);
1529 		if (svd->type == MAP_SHARED)
1530 			lgrp_shm_policy_init(NULL, svd->vp);
1531 	}
1532 	newsvd->offset = svd->offset;
1533 	newsvd->prot = svd->prot;
1534 	newsvd->maxprot = svd->maxprot;
1535 	newsvd->pageprot = svd->pageprot;
1536 	newsvd->type = svd->type;
1537 	newsvd->cred = svd->cred;
1538 	crhold(newsvd->cred);
1539 	newsvd->advice = svd->advice;
1540 	newsvd->pageadvice = svd->pageadvice;
1541 	newsvd->swresv = svd->swresv;
1542 	newsvd->pageswap = svd->pageswap;
1543 	newsvd->flags = svd->flags;
1544 	newsvd->softlockcnt = 0;
1545 	newsvd->softlockcnt_sbase = 0;
1546 	newsvd->softlockcnt_send = 0;
1547 	newsvd->policy_info = svd->policy_info;
1548 	newsvd->rcookie = HAT_INVALID_REGION_COOKIE;
1549 
1550 	if ((amp = svd->amp) == NULL || svd->tr_state == SEGVN_TR_ON) {
1551 		/*
1552 		 * Not attaching to a shared anon object.
1553 		 */
1554 		ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie) ||
1555 		    svd->tr_state == SEGVN_TR_OFF);
1556 		if (svd->tr_state == SEGVN_TR_ON) {
1557 			ASSERT(newsvd->vp != NULL && amp != NULL);
1558 			newsvd->tr_state = SEGVN_TR_INIT;
1559 		} else {
1560 			newsvd->tr_state = svd->tr_state;
1561 		}
1562 		newsvd->amp = NULL;
1563 		newsvd->anon_index = 0;
1564 	} else {
1565 		/* regions for now are only used on pure vnode segments */
1566 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
1567 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1568 		newsvd->tr_state = SEGVN_TR_OFF;
1569 		if (svd->type == MAP_SHARED) {
1570 			newsvd->amp = amp;
1571 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1572 			amp->refcnt++;
1573 			ANON_LOCK_EXIT(&amp->a_rwlock);
1574 			newsvd->anon_index = svd->anon_index;
1575 		} else {
1576 			int reclaim = 1;
1577 
1578 			/*
1579 			 * Allocate and initialize new anon_map structure.
1580 			 */
1581 			newsvd->amp = anonmap_alloc(newseg->s_size, 0,
1582 			    ANON_SLEEP);
1583 			newsvd->amp->a_szc = newseg->s_szc;
1584 			newsvd->anon_index = 0;
1585 
1586 			/*
1587 			 * We don't have to acquire the anon_map lock
1588 			 * for the new segment (since it belongs to an
1589 			 * address space that is still not associated
1590 			 * with any process), or the segment in the old
1591 			 * address space (since all threads in it
1592 			 * are stopped while duplicating the address space).
1593 			 */
1594 
1595 			/*
1596 			 * The goal of the following code is to make sure that
1597 			 * softlocked pages do not end up as copy on write
1598 			 * pages.  This would cause problems where one
1599 			 * thread writes to a page that is COW and a different
1600 			 * thread in the same process has softlocked it.  The
1601 			 * softlock lock would move away from this process
1602 			 * because the write would cause this process to get
1603 			 * a copy (without the softlock).
1604 			 *
1605 			 * The strategy here is to just break the
1606 			 * sharing on pages that could possibly be
1607 			 * softlocked.
1608 			 */
1609 retry:
1610 			if (svd->softlockcnt) {
1611 				struct anon *ap, *newap;
1612 				size_t i;
1613 				uint_t vpprot;
1614 				page_t *anon_pl[1+1], *pp;
1615 				caddr_t addr;
1616 				ulong_t old_idx = svd->anon_index;
1617 				ulong_t new_idx = 0;
1618 
1619 				/*
1620 				 * The softlock count might be non zero
1621 				 * because some pages are still stuck in the
1622 				 * cache for lazy reclaim. Flush the cache
1623 				 * now. This should drop the count to zero.
1624 				 * [or there is really I/O going on to these
1625 				 * pages]. Note, we have the writers lock so
1626 				 * nothing gets inserted during the flush.
1627 				 */
1628 				if (reclaim == 1) {
1629 					segvn_purge(seg);
1630 					reclaim = 0;
1631 					goto retry;
1632 				}
1633 				i = btopr(seg->s_size);
1634 				addr = seg->s_base;
1635 				/*
1636 				 * XXX break cow sharing using PAGESIZE
1637 				 * pages. They will be relocated into larger
1638 				 * pages at fault time.
1639 				 */
1640 				while (i-- > 0) {
1641 					if (ap = anon_get_ptr(amp->ahp,
1642 					    old_idx)) {
1643 						error = anon_getpage(&ap,
1644 						    &vpprot, anon_pl, PAGESIZE,
1645 						    seg, addr, S_READ,
1646 						    svd->cred);
1647 						if (error) {
1648 							newsvd->vpage = NULL;
1649 							goto out;
1650 						}
1651 						/*
1652 						 * prot need not be computed
1653 						 * below 'cause anon_private is
1654 						 * going to ignore it anyway
1655 						 * as child doesn't inherit
1656 						 * pagelock from parent.
1657 						 */
1658 						prot = svd->pageprot ?
1659 						    VPP_PROT(
1660 						    &svd->vpage[
1661 						    seg_page(seg, addr)])
1662 						    : svd->prot;
1663 						pp = anon_private(&newap,
1664 						    newseg, addr, prot,
1665 						    anon_pl[0],	0,
1666 						    newsvd->cred);
1667 						if (pp == NULL) {
1668 							/* no mem abort */
1669 							newsvd->vpage = NULL;
1670 							error = ENOMEM;
1671 							goto out;
1672 						}
1673 						(void) anon_set_ptr(
1674 						    newsvd->amp->ahp, new_idx,
1675 						    newap, ANON_SLEEP);
1676 						page_unlock(pp);
1677 					}
1678 					addr += PAGESIZE;
1679 					old_idx++;
1680 					new_idx++;
1681 				}
1682 			} else {	/* common case */
1683 				if (seg->s_szc != 0) {
1684 					/*
1685 					 * If at least one of anon slots of a
1686 					 * large page exists then make sure
1687 					 * all anon slots of a large page
1688 					 * exist to avoid partial cow sharing
1689 					 * of a large page in the future.
1690 					 */
1691 					anon_dup_fill_holes(amp->ahp,
1692 					    svd->anon_index, newsvd->amp->ahp,
1693 					    0, seg->s_size, seg->s_szc,
1694 					    svd->vp != NULL);
1695 				} else {
1696 					anon_dup(amp->ahp, svd->anon_index,
1697 					    newsvd->amp->ahp, 0, seg->s_size);
1698 				}
1699 
1700 				hat_clrattr(seg->s_as->a_hat, seg->s_base,
1701 				    seg->s_size, PROT_WRITE);
1702 			}
1703 		}
1704 	}
1705 	/*
1706 	 * If necessary, create a vpage structure for the new segment.
1707 	 * Do not copy any page lock indications.
1708 	 */
1709 	if (svd->vpage != NULL) {
1710 		uint_t i;
1711 		struct vpage *ovp = svd->vpage;
1712 		struct vpage *nvp;
1713 
1714 		nvp = newsvd->vpage =
1715 		    kmem_alloc(vpgtob(npages), KM_SLEEP);
1716 		for (i = 0; i < npages; i++) {
1717 			*nvp = *ovp++;
1718 			VPP_CLRPPLOCK(nvp++);
1719 		}
1720 	} else
1721 		newsvd->vpage = NULL;
1722 
1723 	/* Inform the vnode of the new mapping */
1724 	if (newsvd->vp != NULL) {
1725 		error = VOP_ADDMAP(newsvd->vp, (offset_t)newsvd->offset,
1726 		    newseg->s_as, newseg->s_base, newseg->s_size, newsvd->prot,
1727 		    newsvd->maxprot, newsvd->type, newsvd->cred, NULL);
1728 	}
1729 out:
1730 	if (error == 0 && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1731 		ASSERT(newsvd->amp == NULL);
1732 		ASSERT(newsvd->tr_state == SEGVN_TR_OFF);
1733 		newsvd->rcookie = svd->rcookie;
1734 		hat_dup_region(newseg->s_as->a_hat, newsvd->rcookie);
1735 	}
1736 	return (error);
1737 }
1738 
1739 
1740 /*
1741  * callback function to invoke free_vp_pages() for only those pages actually
1742  * processed by the HAT when a shared region is destroyed.
1743  */
1744 extern int free_pages;
1745 
1746 static void
1747 segvn_hat_rgn_unload_callback(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
1748     size_t r_size, void *r_obj, u_offset_t r_objoff)
1749 {
1750 	u_offset_t off;
1751 	size_t len;
1752 	vnode_t *vp = (vnode_t *)r_obj;
1753 
1754 	ASSERT(eaddr > saddr);
1755 	ASSERT(saddr >= r_saddr);
1756 	ASSERT(saddr < r_saddr + r_size);
1757 	ASSERT(eaddr > r_saddr);
1758 	ASSERT(eaddr <= r_saddr + r_size);
1759 	ASSERT(vp != NULL);
1760 
1761 	if (!free_pages) {
1762 		return;
1763 	}
1764 
1765 	len = eaddr - saddr;
1766 	off = (saddr - r_saddr) + r_objoff;
1767 	free_vp_pages(vp, off, len);
1768 }
1769 
1770 /*
1771  * callback function used by segvn_unmap to invoke free_vp_pages() for only
1772  * those pages actually processed by the HAT
1773  */
1774 static void
1775 segvn_hat_unload_callback(hat_callback_t *cb)
1776 {
1777 	struct seg		*seg = cb->hcb_data;
1778 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
1779 	size_t			len;
1780 	u_offset_t		off;
1781 
1782 	ASSERT(svd->vp != NULL);
1783 	ASSERT(cb->hcb_end_addr > cb->hcb_start_addr);
1784 	ASSERT(cb->hcb_start_addr >= seg->s_base);
1785 
1786 	len = cb->hcb_end_addr - cb->hcb_start_addr;
1787 	off = cb->hcb_start_addr - seg->s_base;
1788 	free_vp_pages(svd->vp, svd->offset + off, len);
1789 }
1790 
1791 /*
1792  * This function determines the number of bytes of swap reserved by
1793  * a segment for which per-page accounting is present. It is used to
1794  * calculate the correct value of a segvn_data's swresv.
1795  */
1796 static size_t
1797 segvn_count_swap_by_vpages(struct seg *seg)
1798 {
1799 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1800 	struct vpage *vp, *evp;
1801 	size_t nswappages = 0;
1802 
1803 	ASSERT(svd->pageswap);
1804 	ASSERT(svd->vpage != NULL);
1805 
1806 	evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
1807 
1808 	for (vp = svd->vpage; vp < evp; vp++) {
1809 		if (VPP_ISSWAPRES(vp))
1810 			nswappages++;
1811 	}
1812 
1813 	return (nswappages << PAGESHIFT);
1814 }
1815 
1816 static int
1817 segvn_unmap(struct seg *seg, caddr_t addr, size_t len)
1818 {
1819 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1820 	struct segvn_data *nsvd;
1821 	struct seg *nseg;
1822 	struct anon_map *amp;
1823 	pgcnt_t	opages;		/* old segment size in pages */
1824 	pgcnt_t	npages;		/* new segment size in pages */
1825 	pgcnt_t	dpages;		/* pages being deleted (unmapped) */
1826 	hat_callback_t callback;	/* used for free_vp_pages() */
1827 	hat_callback_t *cbp = NULL;
1828 	caddr_t nbase;
1829 	size_t nsize;
1830 	size_t oswresv;
1831 	int reclaim = 1;
1832 
1833 	/*
1834 	 * We don't need any segment level locks for "segvn" data
1835 	 * since the address space is "write" locked.
1836 	 */
1837 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
1838 
1839 	/*
1840 	 * Fail the unmap if pages are SOFTLOCKed through this mapping.
1841 	 * softlockcnt is protected from change by the as write lock.
1842 	 */
1843 retry:
1844 	if (svd->softlockcnt > 0) {
1845 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1846 
1847 		/*
1848 		 * If this is shared segment non 0 softlockcnt
1849 		 * means locked pages are still in use.
1850 		 */
1851 		if (svd->type == MAP_SHARED) {
1852 			return (EAGAIN);
1853 		}
1854 
1855 		/*
1856 		 * since we do have the writers lock nobody can fill
1857 		 * the cache during the purge. The flush either succeeds
1858 		 * or we still have pending I/Os.
1859 		 */
1860 		if (reclaim == 1) {
1861 			segvn_purge(seg);
1862 			reclaim = 0;
1863 			goto retry;
1864 		}
1865 		return (EAGAIN);
1866 	}
1867 
1868 	/*
1869 	 * Check for bad sizes
1870 	 */
1871 	if (addr < seg->s_base || addr + len > seg->s_base + seg->s_size ||
1872 	    (len & PAGEOFFSET) || ((uintptr_t)addr & PAGEOFFSET)) {
1873 		panic("segvn_unmap");
1874 		/*NOTREACHED*/
1875 	}
1876 
1877 	if (seg->s_szc != 0) {
1878 		size_t pgsz = page_get_pagesize(seg->s_szc);
1879 		int err;
1880 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
1881 			ASSERT(seg->s_base != addr || seg->s_size != len);
1882 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1883 				ASSERT(svd->amp == NULL);
1884 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1885 				hat_leave_region(seg->s_as->a_hat,
1886 				    svd->rcookie, HAT_REGION_TEXT);
1887 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
1888 				/*
1889 				 * could pass a flag to segvn_demote_range()
1890 				 * below to tell it not to do any unloads but
1891 				 * this case is rare enough to not bother for
1892 				 * now.
1893 				 */
1894 			} else if (svd->tr_state == SEGVN_TR_INIT) {
1895 				svd->tr_state = SEGVN_TR_OFF;
1896 			} else if (svd->tr_state == SEGVN_TR_ON) {
1897 				ASSERT(svd->amp != NULL);
1898 				segvn_textunrepl(seg, 1);
1899 				ASSERT(svd->amp == NULL);
1900 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1901 			}
1902 			VM_STAT_ADD(segvnvmstats.demoterange[0]);
1903 			err = segvn_demote_range(seg, addr, len, SDR_END, 0);
1904 			if (err == 0) {
1905 				return (IE_RETRY);
1906 			}
1907 			return (err);
1908 		}
1909 	}
1910 
1911 	/* Inform the vnode of the unmapping. */
1912 	if (svd->vp) {
1913 		int error;
1914 
1915 		error = VOP_DELMAP(svd->vp,
1916 		    (offset_t)svd->offset + (uintptr_t)(addr - seg->s_base),
1917 		    seg->s_as, addr, len, svd->prot, svd->maxprot,
1918 		    svd->type, svd->cred, NULL);
1919 
1920 		if (error == EAGAIN)
1921 			return (error);
1922 	}
1923 
1924 	/*
1925 	 * Remove any page locks set through this mapping.
1926 	 * If text replication is not off no page locks could have been
1927 	 * established via this mapping.
1928 	 */
1929 	if (svd->tr_state == SEGVN_TR_OFF) {
1930 		(void) segvn_lockop(seg, addr, len, 0, MC_UNLOCK, NULL, 0);
1931 	}
1932 
1933 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1934 		ASSERT(svd->amp == NULL);
1935 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1936 		ASSERT(svd->type == MAP_PRIVATE);
1937 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
1938 		    HAT_REGION_TEXT);
1939 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
1940 	} else if (svd->tr_state == SEGVN_TR_ON) {
1941 		ASSERT(svd->amp != NULL);
1942 		ASSERT(svd->pageprot == 0 && !(svd->prot & PROT_WRITE));
1943 		segvn_textunrepl(seg, 1);
1944 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
1945 	} else {
1946 		if (svd->tr_state != SEGVN_TR_OFF) {
1947 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
1948 			svd->tr_state = SEGVN_TR_OFF;
1949 		}
1950 		/*
1951 		 * Unload any hardware translations in the range to be taken
1952 		 * out. Use a callback to invoke free_vp_pages() effectively.
1953 		 */
1954 		if (svd->vp != NULL && free_pages != 0) {
1955 			callback.hcb_data = seg;
1956 			callback.hcb_function = segvn_hat_unload_callback;
1957 			cbp = &callback;
1958 		}
1959 		hat_unload_callback(seg->s_as->a_hat, addr, len,
1960 		    HAT_UNLOAD_UNMAP, cbp);
1961 
1962 		if (svd->type == MAP_SHARED && svd->vp != NULL &&
1963 		    (svd->vp->v_flag & VVMEXEC) &&
1964 		    ((svd->prot & PROT_WRITE) || svd->pageprot)) {
1965 			segvn_inval_trcache(svd->vp);
1966 		}
1967 	}
1968 
1969 	/*
1970 	 * Check for entire segment
1971 	 */
1972 	if (addr == seg->s_base && len == seg->s_size) {
1973 		seg_free(seg);
1974 		return (0);
1975 	}
1976 
1977 	opages = seg_pages(seg);
1978 	dpages = btop(len);
1979 	npages = opages - dpages;
1980 	amp = svd->amp;
1981 	ASSERT(amp == NULL || amp->a_szc >= seg->s_szc);
1982 
1983 	/*
1984 	 * Check for beginning of segment
1985 	 */
1986 	if (addr == seg->s_base) {
1987 		if (svd->vpage != NULL) {
1988 			size_t nbytes;
1989 			struct vpage *ovpage;
1990 
1991 			ovpage = svd->vpage;	/* keep pointer to vpage */
1992 
1993 			nbytes = vpgtob(npages);
1994 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
1995 			bcopy(&ovpage[dpages], svd->vpage, nbytes);
1996 
1997 			/* free up old vpage */
1998 			kmem_free(ovpage, vpgtob(opages));
1999 		}
2000 		if (amp != NULL) {
2001 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2002 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2003 				/*
2004 				 * Shared anon map is no longer in use. Before
2005 				 * freeing its pages purge all entries from
2006 				 * pcache that belong to this amp.
2007 				 */
2008 				if (svd->type == MAP_SHARED) {
2009 					ASSERT(amp->refcnt == 1);
2010 					ASSERT(svd->softlockcnt == 0);
2011 					anonmap_purge(amp);
2012 				}
2013 				/*
2014 				 * Free up now unused parts of anon_map array.
2015 				 */
2016 				if (amp->a_szc == seg->s_szc) {
2017 					if (seg->s_szc != 0) {
2018 						anon_free_pages(amp->ahp,
2019 						    svd->anon_index, len,
2020 						    seg->s_szc);
2021 					} else {
2022 						anon_free(amp->ahp,
2023 						    svd->anon_index,
2024 						    len);
2025 					}
2026 				} else {
2027 					ASSERT(svd->type == MAP_SHARED);
2028 					ASSERT(amp->a_szc > seg->s_szc);
2029 					anon_shmap_free_pages(amp,
2030 					    svd->anon_index, len);
2031 				}
2032 
2033 				/*
2034 				 * Unreserve swap space for the
2035 				 * unmapped chunk of this segment in
2036 				 * case it's MAP_SHARED
2037 				 */
2038 				if (svd->type == MAP_SHARED) {
2039 					anon_unresv_zone(len,
2040 					    seg->s_as->a_proc->p_zone);
2041 					amp->swresv -= len;
2042 				}
2043 			}
2044 			ANON_LOCK_EXIT(&amp->a_rwlock);
2045 			svd->anon_index += dpages;
2046 		}
2047 		if (svd->vp != NULL)
2048 			svd->offset += len;
2049 
2050 		seg->s_base += len;
2051 		seg->s_size -= len;
2052 
2053 		if (svd->swresv) {
2054 			if (svd->flags & MAP_NORESERVE) {
2055 				ASSERT(amp);
2056 				oswresv = svd->swresv;
2057 
2058 				svd->swresv = ptob(anon_pages(amp->ahp,
2059 				    svd->anon_index, npages));
2060 				anon_unresv_zone(oswresv - svd->swresv,
2061 				    seg->s_as->a_proc->p_zone);
2062 				if (SEG_IS_PARTIAL_RESV(seg))
2063 					seg->s_as->a_resvsize -= oswresv -
2064 					    svd->swresv;
2065 			} else {
2066 				size_t unlen;
2067 
2068 				if (svd->pageswap) {
2069 					oswresv = svd->swresv;
2070 					svd->swresv =
2071 					    segvn_count_swap_by_vpages(seg);
2072 					ASSERT(oswresv >= svd->swresv);
2073 					unlen = oswresv - svd->swresv;
2074 				} else {
2075 					svd->swresv -= len;
2076 					ASSERT(svd->swresv == seg->s_size);
2077 					unlen = len;
2078 				}
2079 				anon_unresv_zone(unlen,
2080 				    seg->s_as->a_proc->p_zone);
2081 			}
2082 			TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2083 			    seg, len, 0);
2084 		}
2085 
2086 		return (0);
2087 	}
2088 
2089 	/*
2090 	 * Check for end of segment
2091 	 */
2092 	if (addr + len == seg->s_base + seg->s_size) {
2093 		if (svd->vpage != NULL) {
2094 			size_t nbytes;
2095 			struct vpage *ovpage;
2096 
2097 			ovpage = svd->vpage;	/* keep pointer to vpage */
2098 
2099 			nbytes = vpgtob(npages);
2100 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2101 			bcopy(ovpage, svd->vpage, nbytes);
2102 
2103 			/* free up old vpage */
2104 			kmem_free(ovpage, vpgtob(opages));
2105 
2106 		}
2107 		if (amp != NULL) {
2108 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2109 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2110 				/*
2111 				 * Free up now unused parts of anon_map array.
2112 				 */
2113 				ulong_t an_idx = svd->anon_index + npages;
2114 
2115 				/*
2116 				 * Shared anon map is no longer in use. Before
2117 				 * freeing its pages purge all entries from
2118 				 * pcache that belong to this amp.
2119 				 */
2120 				if (svd->type == MAP_SHARED) {
2121 					ASSERT(amp->refcnt == 1);
2122 					ASSERT(svd->softlockcnt == 0);
2123 					anonmap_purge(amp);
2124 				}
2125 
2126 				if (amp->a_szc == seg->s_szc) {
2127 					if (seg->s_szc != 0) {
2128 						anon_free_pages(amp->ahp,
2129 						    an_idx, len,
2130 						    seg->s_szc);
2131 					} else {
2132 						anon_free(amp->ahp, an_idx,
2133 						    len);
2134 					}
2135 				} else {
2136 					ASSERT(svd->type == MAP_SHARED);
2137 					ASSERT(amp->a_szc > seg->s_szc);
2138 					anon_shmap_free_pages(amp,
2139 					    an_idx, len);
2140 				}
2141 
2142 				/*
2143 				 * Unreserve swap space for the
2144 				 * unmapped chunk of this segment in
2145 				 * case it's MAP_SHARED
2146 				 */
2147 				if (svd->type == MAP_SHARED) {
2148 					anon_unresv_zone(len,
2149 					    seg->s_as->a_proc->p_zone);
2150 					amp->swresv -= len;
2151 				}
2152 			}
2153 			ANON_LOCK_EXIT(&amp->a_rwlock);
2154 		}
2155 
2156 		seg->s_size -= len;
2157 
2158 		if (svd->swresv) {
2159 			if (svd->flags & MAP_NORESERVE) {
2160 				ASSERT(amp);
2161 				oswresv = svd->swresv;
2162 				svd->swresv = ptob(anon_pages(amp->ahp,
2163 				    svd->anon_index, npages));
2164 				anon_unresv_zone(oswresv - svd->swresv,
2165 				    seg->s_as->a_proc->p_zone);
2166 				if (SEG_IS_PARTIAL_RESV(seg))
2167 					seg->s_as->a_resvsize -= oswresv -
2168 					    svd->swresv;
2169 			} else {
2170 				size_t unlen;
2171 
2172 				if (svd->pageswap) {
2173 					oswresv = svd->swresv;
2174 					svd->swresv =
2175 					    segvn_count_swap_by_vpages(seg);
2176 					ASSERT(oswresv >= svd->swresv);
2177 					unlen = oswresv - svd->swresv;
2178 				} else {
2179 					svd->swresv -= len;
2180 					ASSERT(svd->swresv == seg->s_size);
2181 					unlen = len;
2182 				}
2183 				anon_unresv_zone(unlen,
2184 				    seg->s_as->a_proc->p_zone);
2185 			}
2186 			TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2187 			    "anon proc:%p %lu %u", seg, len, 0);
2188 		}
2189 
2190 		return (0);
2191 	}
2192 
2193 	/*
2194 	 * The section to go is in the middle of the segment,
2195 	 * have to make it into two segments.  nseg is made for
2196 	 * the high end while seg is cut down at the low end.
2197 	 */
2198 	nbase = addr + len;				/* new seg base */
2199 	nsize = (seg->s_base + seg->s_size) - nbase;	/* new seg size */
2200 	seg->s_size = addr - seg->s_base;		/* shrink old seg */
2201 	nseg = seg_alloc(seg->s_as, nbase, nsize);
2202 	if (nseg == NULL) {
2203 		panic("segvn_unmap seg_alloc");
2204 		/*NOTREACHED*/
2205 	}
2206 	nseg->s_ops = seg->s_ops;
2207 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
2208 	nseg->s_data = (void *)nsvd;
2209 	nseg->s_szc = seg->s_szc;
2210 	*nsvd = *svd;
2211 	nsvd->seg = nseg;
2212 	nsvd->offset = svd->offset + (uintptr_t)(nseg->s_base - seg->s_base);
2213 	nsvd->swresv = 0;
2214 	nsvd->softlockcnt = 0;
2215 	nsvd->softlockcnt_sbase = 0;
2216 	nsvd->softlockcnt_send = 0;
2217 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
2218 
2219 	if (svd->vp != NULL) {
2220 		VN_HOLD(nsvd->vp);
2221 		if (nsvd->type == MAP_SHARED)
2222 			lgrp_shm_policy_init(NULL, nsvd->vp);
2223 	}
2224 	crhold(svd->cred);
2225 
2226 	if (svd->vpage == NULL) {
2227 		nsvd->vpage = NULL;
2228 	} else {
2229 		/* need to split vpage into two arrays */
2230 		size_t nbytes;
2231 		struct vpage *ovpage;
2232 
2233 		ovpage = svd->vpage;		/* keep pointer to vpage */
2234 
2235 		npages = seg_pages(seg);	/* seg has shrunk */
2236 		nbytes = vpgtob(npages);
2237 		svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2238 
2239 		bcopy(ovpage, svd->vpage, nbytes);
2240 
2241 		npages = seg_pages(nseg);
2242 		nbytes = vpgtob(npages);
2243 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2244 
2245 		bcopy(&ovpage[opages - npages], nsvd->vpage, nbytes);
2246 
2247 		/* free up old vpage */
2248 		kmem_free(ovpage, vpgtob(opages));
2249 	}
2250 
2251 	if (amp == NULL) {
2252 		nsvd->amp = NULL;
2253 		nsvd->anon_index = 0;
2254 	} else {
2255 		/*
2256 		 * Need to create a new anon map for the new segment.
2257 		 * We'll also allocate a new smaller array for the old
2258 		 * smaller segment to save space.
2259 		 */
2260 		opages = btop((uintptr_t)(addr - seg->s_base));
2261 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2262 		if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2263 			/*
2264 			 * Free up now unused parts of anon_map array.
2265 			 */
2266 			ulong_t an_idx = svd->anon_index + opages;
2267 
2268 			/*
2269 			 * Shared anon map is no longer in use. Before
2270 			 * freeing its pages purge all entries from
2271 			 * pcache that belong to this amp.
2272 			 */
2273 			if (svd->type == MAP_SHARED) {
2274 				ASSERT(amp->refcnt == 1);
2275 				ASSERT(svd->softlockcnt == 0);
2276 				anonmap_purge(amp);
2277 			}
2278 
2279 			if (amp->a_szc == seg->s_szc) {
2280 				if (seg->s_szc != 0) {
2281 					anon_free_pages(amp->ahp, an_idx, len,
2282 					    seg->s_szc);
2283 				} else {
2284 					anon_free(amp->ahp, an_idx,
2285 					    len);
2286 				}
2287 			} else {
2288 				ASSERT(svd->type == MAP_SHARED);
2289 				ASSERT(amp->a_szc > seg->s_szc);
2290 				anon_shmap_free_pages(amp, an_idx, len);
2291 			}
2292 
2293 			/*
2294 			 * Unreserve swap space for the
2295 			 * unmapped chunk of this segment in
2296 			 * case it's MAP_SHARED
2297 			 */
2298 			if (svd->type == MAP_SHARED) {
2299 				anon_unresv_zone(len,
2300 				    seg->s_as->a_proc->p_zone);
2301 				amp->swresv -= len;
2302 			}
2303 		}
2304 		nsvd->anon_index = svd->anon_index +
2305 		    btop((uintptr_t)(nseg->s_base - seg->s_base));
2306 		if (svd->type == MAP_SHARED) {
2307 			amp->refcnt++;
2308 			nsvd->amp = amp;
2309 		} else {
2310 			struct anon_map *namp;
2311 			struct anon_hdr *nahp;
2312 
2313 			ASSERT(svd->type == MAP_PRIVATE);
2314 			nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
2315 			namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
2316 			namp->a_szc = seg->s_szc;
2317 			(void) anon_copy_ptr(amp->ahp, svd->anon_index, nahp,
2318 			    0, btop(seg->s_size), ANON_SLEEP);
2319 			(void) anon_copy_ptr(amp->ahp, nsvd->anon_index,
2320 			    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
2321 			anon_release(amp->ahp, btop(amp->size));
2322 			svd->anon_index = 0;
2323 			nsvd->anon_index = 0;
2324 			amp->ahp = nahp;
2325 			amp->size = seg->s_size;
2326 			nsvd->amp = namp;
2327 		}
2328 		ANON_LOCK_EXIT(&amp->a_rwlock);
2329 	}
2330 	if (svd->swresv) {
2331 		if (svd->flags & MAP_NORESERVE) {
2332 			ASSERT(amp);
2333 			oswresv = svd->swresv;
2334 			svd->swresv = ptob(anon_pages(amp->ahp,
2335 			    svd->anon_index, btop(seg->s_size)));
2336 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
2337 			    nsvd->anon_index, btop(nseg->s_size)));
2338 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2339 			anon_unresv_zone(oswresv - (svd->swresv + nsvd->swresv),
2340 			    seg->s_as->a_proc->p_zone);
2341 			if (SEG_IS_PARTIAL_RESV(seg))
2342 				seg->s_as->a_resvsize -= oswresv -
2343 				    (svd->swresv + nsvd->swresv);
2344 		} else {
2345 			size_t unlen;
2346 
2347 			if (svd->pageswap) {
2348 				oswresv = svd->swresv;
2349 				svd->swresv = segvn_count_swap_by_vpages(seg);
2350 				nsvd->swresv = segvn_count_swap_by_vpages(nseg);
2351 				ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2352 				unlen = oswresv - (svd->swresv + nsvd->swresv);
2353 			} else {
2354 				if (seg->s_size + nseg->s_size + len !=
2355 				    svd->swresv) {
2356 					panic("segvn_unmap: cannot split "
2357 					    "swap reservation");
2358 					/*NOTREACHED*/
2359 				}
2360 				svd->swresv = seg->s_size;
2361 				nsvd->swresv = nseg->s_size;
2362 				unlen = len;
2363 			}
2364 			anon_unresv_zone(unlen,
2365 			    seg->s_as->a_proc->p_zone);
2366 		}
2367 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2368 		    seg, len, 0);
2369 	}
2370 
2371 	return (0);			/* I'm glad that's all over with! */
2372 }
2373 
2374 static void
2375 segvn_free(struct seg *seg)
2376 {
2377 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2378 	pgcnt_t npages = seg_pages(seg);
2379 	struct anon_map *amp;
2380 	size_t len;
2381 
2382 	/*
2383 	 * We don't need any segment level locks for "segvn" data
2384 	 * since the address space is "write" locked.
2385 	 */
2386 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
2387 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
2388 
2389 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2390 
2391 	/*
2392 	 * Be sure to unlock pages. XXX Why do things get free'ed instead
2393 	 * of unmapped? XXX
2394 	 */
2395 	(void) segvn_lockop(seg, seg->s_base, seg->s_size,
2396 	    0, MC_UNLOCK, NULL, 0);
2397 
2398 	/*
2399 	 * Deallocate the vpage and anon pointers if necessary and possible.
2400 	 */
2401 	if (svd->vpage != NULL) {
2402 		kmem_free(svd->vpage, vpgtob(npages));
2403 		svd->vpage = NULL;
2404 	}
2405 	if ((amp = svd->amp) != NULL) {
2406 		/*
2407 		 * If there are no more references to this anon_map
2408 		 * structure, then deallocate the structure after freeing
2409 		 * up all the anon slot pointers that we can.
2410 		 */
2411 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2412 		ASSERT(amp->a_szc >= seg->s_szc);
2413 		if (--amp->refcnt == 0) {
2414 			if (svd->type == MAP_PRIVATE) {
2415 				/*
2416 				 * Private - we only need to anon_free
2417 				 * the part that this segment refers to.
2418 				 */
2419 				if (seg->s_szc != 0) {
2420 					anon_free_pages(amp->ahp,
2421 					    svd->anon_index, seg->s_size,
2422 					    seg->s_szc);
2423 				} else {
2424 					anon_free(amp->ahp, svd->anon_index,
2425 					    seg->s_size);
2426 				}
2427 			} else {
2428 
2429 				/*
2430 				 * Shared anon map is no longer in use. Before
2431 				 * freeing its pages purge all entries from
2432 				 * pcache that belong to this amp.
2433 				 */
2434 				ASSERT(svd->softlockcnt == 0);
2435 				anonmap_purge(amp);
2436 
2437 				/*
2438 				 * Shared - anon_free the entire
2439 				 * anon_map's worth of stuff and
2440 				 * release any swap reservation.
2441 				 */
2442 				if (amp->a_szc != 0) {
2443 					anon_shmap_free_pages(amp, 0,
2444 					    amp->size);
2445 				} else {
2446 					anon_free(amp->ahp, 0, amp->size);
2447 				}
2448 				if ((len = amp->swresv) != 0) {
2449 					anon_unresv_zone(len,
2450 					    seg->s_as->a_proc->p_zone);
2451 					TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2452 					    "anon proc:%p %lu %u", seg, len, 0);
2453 				}
2454 			}
2455 			svd->amp = NULL;
2456 			ANON_LOCK_EXIT(&amp->a_rwlock);
2457 			anonmap_free(amp);
2458 		} else if (svd->type == MAP_PRIVATE) {
2459 			/*
2460 			 * We had a private mapping which still has
2461 			 * a held anon_map so just free up all the
2462 			 * anon slot pointers that we were using.
2463 			 */
2464 			if (seg->s_szc != 0) {
2465 				anon_free_pages(amp->ahp, svd->anon_index,
2466 				    seg->s_size, seg->s_szc);
2467 			} else {
2468 				anon_free(amp->ahp, svd->anon_index,
2469 				    seg->s_size);
2470 			}
2471 			ANON_LOCK_EXIT(&amp->a_rwlock);
2472 		} else {
2473 			ANON_LOCK_EXIT(&amp->a_rwlock);
2474 		}
2475 	}
2476 
2477 	/*
2478 	 * Release swap reservation.
2479 	 */
2480 	if ((len = svd->swresv) != 0) {
2481 		anon_unresv_zone(svd->swresv,
2482 		    seg->s_as->a_proc->p_zone);
2483 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2484 		    seg, len, 0);
2485 		if (SEG_IS_PARTIAL_RESV(seg))
2486 			seg->s_as->a_resvsize -= svd->swresv;
2487 		svd->swresv = 0;
2488 	}
2489 	/*
2490 	 * Release claim on vnode, credentials, and finally free the
2491 	 * private data.
2492 	 */
2493 	if (svd->vp != NULL) {
2494 		if (svd->type == MAP_SHARED)
2495 			lgrp_shm_policy_fini(NULL, svd->vp);
2496 		VN_RELE(svd->vp);
2497 		svd->vp = NULL;
2498 	}
2499 	crfree(svd->cred);
2500 	svd->pageprot = 0;
2501 	svd->pageadvice = 0;
2502 	svd->pageswap = 0;
2503 	svd->cred = NULL;
2504 
2505 	/*
2506 	 * Take segfree_syncmtx lock to let segvn_reclaim() finish if it's
2507 	 * still working with this segment without holding as lock (in case
2508 	 * it's called by pcache async thread).
2509 	 */
2510 	ASSERT(svd->softlockcnt == 0);
2511 	mutex_enter(&svd->segfree_syncmtx);
2512 	mutex_exit(&svd->segfree_syncmtx);
2513 
2514 	seg->s_data = NULL;
2515 	kmem_cache_free(segvn_cache, svd);
2516 }
2517 
2518 /*
2519  * Do a F_SOFTUNLOCK call over the range requested.  The range must have
2520  * already been F_SOFTLOCK'ed.
2521  * Caller must always match addr and len of a softunlock with a previous
2522  * softlock with exactly the same addr and len.
2523  */
2524 static void
2525 segvn_softunlock(struct seg *seg, caddr_t addr, size_t len, enum seg_rw rw)
2526 {
2527 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2528 	page_t *pp;
2529 	caddr_t adr;
2530 	struct vnode *vp;
2531 	u_offset_t offset;
2532 	ulong_t anon_index;
2533 	struct anon_map *amp;
2534 	struct anon *ap = NULL;
2535 
2536 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
2537 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
2538 
2539 	if ((amp = svd->amp) != NULL)
2540 		anon_index = svd->anon_index + seg_page(seg, addr);
2541 
2542 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
2543 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2544 		hat_unlock_region(seg->s_as->a_hat, addr, len, svd->rcookie);
2545 	} else {
2546 		hat_unlock(seg->s_as->a_hat, addr, len);
2547 	}
2548 	for (adr = addr; adr < addr + len; adr += PAGESIZE) {
2549 		if (amp != NULL) {
2550 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2551 			if ((ap = anon_get_ptr(amp->ahp, anon_index++))
2552 			    != NULL) {
2553 				swap_xlate(ap, &vp, &offset);
2554 			} else {
2555 				vp = svd->vp;
2556 				offset = svd->offset +
2557 				    (uintptr_t)(adr - seg->s_base);
2558 			}
2559 			ANON_LOCK_EXIT(&amp->a_rwlock);
2560 		} else {
2561 			vp = svd->vp;
2562 			offset = svd->offset +
2563 			    (uintptr_t)(adr - seg->s_base);
2564 		}
2565 
2566 		/*
2567 		 * Use page_find() instead of page_lookup() to
2568 		 * find the page since we know that it is locked.
2569 		 */
2570 		pp = page_find(vp, offset);
2571 		if (pp == NULL) {
2572 			panic(
2573 			    "segvn_softunlock: addr %p, ap %p, vp %p, off %llx",
2574 			    (void *)adr, (void *)ap, (void *)vp, offset);
2575 			/*NOTREACHED*/
2576 		}
2577 
2578 		if (rw == S_WRITE) {
2579 			hat_setrefmod(pp);
2580 			if (seg->s_as->a_vbits)
2581 				hat_setstat(seg->s_as, adr, PAGESIZE,
2582 				    P_REF | P_MOD);
2583 		} else if (rw != S_OTHER) {
2584 			hat_setref(pp);
2585 			if (seg->s_as->a_vbits)
2586 				hat_setstat(seg->s_as, adr, PAGESIZE, P_REF);
2587 		}
2588 		TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2589 		    "segvn_fault:pp %p vp %p offset %llx", pp, vp, offset);
2590 		page_unlock(pp);
2591 	}
2592 	ASSERT(svd->softlockcnt >= btop(len));
2593 	if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -btop(len))) {
2594 		/*
2595 		 * All SOFTLOCKS are gone. Wakeup any waiting
2596 		 * unmappers so they can try again to unmap.
2597 		 * Check for waiters first without the mutex
2598 		 * held so we don't always grab the mutex on
2599 		 * softunlocks.
2600 		 */
2601 		if (AS_ISUNMAPWAIT(seg->s_as)) {
2602 			mutex_enter(&seg->s_as->a_contents);
2603 			if (AS_ISUNMAPWAIT(seg->s_as)) {
2604 				AS_CLRUNMAPWAIT(seg->s_as);
2605 				cv_broadcast(&seg->s_as->a_cv);
2606 			}
2607 			mutex_exit(&seg->s_as->a_contents);
2608 		}
2609 	}
2610 }
2611 
2612 #define	PAGE_HANDLED	((page_t *)-1)
2613 
2614 /*
2615  * Release all the pages in the NULL terminated ppp list
2616  * which haven't already been converted to PAGE_HANDLED.
2617  */
2618 static void
2619 segvn_pagelist_rele(page_t **ppp)
2620 {
2621 	for (; *ppp != NULL; ppp++) {
2622 		if (*ppp != PAGE_HANDLED)
2623 			page_unlock(*ppp);
2624 	}
2625 }
2626 
2627 static int stealcow = 1;
2628 
2629 /*
2630  * Workaround for viking chip bug.  See bug id 1220902.
2631  * To fix this down in pagefault() would require importing so
2632  * much as and segvn code as to be unmaintainable.
2633  */
2634 int enable_mbit_wa = 0;
2635 
2636 /*
2637  * Handles all the dirty work of getting the right
2638  * anonymous pages and loading up the translations.
2639  * This routine is called only from segvn_fault()
2640  * when looping over the range of addresses requested.
2641  *
2642  * The basic algorithm here is:
2643  * 	If this is an anon_zero case
2644  *		Call anon_zero to allocate page
2645  *		Load up translation
2646  *		Return
2647  *	endif
2648  *	If this is an anon page
2649  *		Use anon_getpage to get the page
2650  *	else
2651  *		Find page in pl[] list passed in
2652  *	endif
2653  *	If not a cow
2654  *		Load up the translation to the page
2655  *		return
2656  *	endif
2657  *	Call anon_private to handle cow
2658  *	Load up (writable) translation to new page
2659  */
2660 static faultcode_t
2661 segvn_faultpage(
2662 	struct hat *hat,		/* the hat to use for mapping */
2663 	struct seg *seg,		/* seg_vn of interest */
2664 	caddr_t addr,			/* address in as */
2665 	u_offset_t off,			/* offset in vp */
2666 	struct vpage *vpage,		/* pointer to vpage for vp, off */
2667 	page_t *pl[],			/* object source page pointer */
2668 	uint_t vpprot,			/* access allowed to object pages */
2669 	enum fault_type type,		/* type of fault */
2670 	enum seg_rw rw,			/* type of access at fault */
2671 	int brkcow)			/* we may need to break cow */
2672 {
2673 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2674 	page_t *pp, **ppp;
2675 	uint_t pageflags = 0;
2676 	page_t *anon_pl[1 + 1];
2677 	page_t *opp = NULL;		/* original page */
2678 	uint_t prot;
2679 	int err;
2680 	int cow;
2681 	int claim;
2682 	int steal = 0;
2683 	ulong_t anon_index;
2684 	struct anon *ap, *oldap;
2685 	struct anon_map *amp;
2686 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
2687 	int anon_lock = 0;
2688 	anon_sync_obj_t cookie;
2689 
2690 	if (svd->flags & MAP_TEXT) {
2691 		hat_flag |= HAT_LOAD_TEXT;
2692 	}
2693 
2694 	ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
2695 	ASSERT(seg->s_szc == 0);
2696 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
2697 
2698 	/*
2699 	 * Initialize protection value for this page.
2700 	 * If we have per page protection values check it now.
2701 	 */
2702 	if (svd->pageprot) {
2703 		uint_t protchk;
2704 
2705 		switch (rw) {
2706 		case S_READ:
2707 			protchk = PROT_READ;
2708 			break;
2709 		case S_WRITE:
2710 			protchk = PROT_WRITE;
2711 			break;
2712 		case S_EXEC:
2713 			protchk = PROT_EXEC;
2714 			break;
2715 		case S_OTHER:
2716 		default:
2717 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
2718 			break;
2719 		}
2720 
2721 		prot = VPP_PROT(vpage);
2722 		if ((prot & protchk) == 0)
2723 			return (FC_PROT);	/* illegal access type */
2724 	} else {
2725 		prot = svd->prot;
2726 	}
2727 
2728 	if (type == F_SOFTLOCK) {
2729 		atomic_inc_ulong((ulong_t *)&svd->softlockcnt);
2730 	}
2731 
2732 	/*
2733 	 * Always acquire the anon array lock to prevent 2 threads from
2734 	 * allocating separate anon slots for the same "addr".
2735 	 */
2736 
2737 	if ((amp = svd->amp) != NULL) {
2738 		ASSERT(RW_READ_HELD(&amp->a_rwlock));
2739 		anon_index = svd->anon_index + seg_page(seg, addr);
2740 		anon_array_enter(amp, anon_index, &cookie);
2741 		anon_lock = 1;
2742 	}
2743 
2744 	if (svd->vp == NULL && amp != NULL) {
2745 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL) {
2746 			/*
2747 			 * Allocate a (normally) writable anonymous page of
2748 			 * zeroes. If no advance reservations, reserve now.
2749 			 */
2750 			if (svd->flags & MAP_NORESERVE) {
2751 				if (anon_resv_zone(ptob(1),
2752 				    seg->s_as->a_proc->p_zone)) {
2753 					atomic_add_long(&svd->swresv, ptob(1));
2754 					atomic_add_long(&seg->s_as->a_resvsize,
2755 					    ptob(1));
2756 				} else {
2757 					err = ENOMEM;
2758 					goto out;
2759 				}
2760 			}
2761 			if ((pp = anon_zero(seg, addr, &ap,
2762 			    svd->cred)) == NULL) {
2763 				err = ENOMEM;
2764 				goto out;	/* out of swap space */
2765 			}
2766 			/*
2767 			 * Re-acquire the anon_map lock and
2768 			 * initialize the anon array entry.
2769 			 */
2770 			(void) anon_set_ptr(amp->ahp, anon_index, ap,
2771 			    ANON_SLEEP);
2772 
2773 			ASSERT(pp->p_szc == 0);
2774 
2775 			/*
2776 			 * Handle pages that have been marked for migration
2777 			 */
2778 			if (lgrp_optimizations())
2779 				page_migrate(seg, addr, &pp, 1);
2780 
2781 			if (enable_mbit_wa) {
2782 				if (rw == S_WRITE)
2783 					hat_setmod(pp);
2784 				else if (!hat_ismod(pp))
2785 					prot &= ~PROT_WRITE;
2786 			}
2787 			/*
2788 			 * If AS_PAGLCK is set in a_flags (via memcntl(2)
2789 			 * with MC_LOCKAS, MCL_FUTURE) and this is a
2790 			 * MAP_NORESERVE segment, we may need to
2791 			 * permanently lock the page as it is being faulted
2792 			 * for the first time. The following text applies
2793 			 * only to MAP_NORESERVE segments:
2794 			 *
2795 			 * As per memcntl(2), if this segment was created
2796 			 * after MCL_FUTURE was applied (a "future"
2797 			 * segment), its pages must be locked.  If this
2798 			 * segment existed at MCL_FUTURE application (a
2799 			 * "past" segment), the interface is unclear.
2800 			 *
2801 			 * We decide to lock only if vpage is present:
2802 			 *
2803 			 * - "future" segments will have a vpage array (see
2804 			 *    as_map), and so will be locked as required
2805 			 *
2806 			 * - "past" segments may not have a vpage array,
2807 			 *    depending on whether events (such as
2808 			 *    mprotect) have occurred. Locking if vpage
2809 			 *    exists will preserve legacy behavior.  Not
2810 			 *    locking if vpage is absent, will not break
2811 			 *    the interface or legacy behavior.  Note that
2812 			 *    allocating vpage here if it's absent requires
2813 			 *    upgrading the segvn reader lock, the cost of
2814 			 *    which does not seem worthwhile.
2815 			 *
2816 			 * Usually testing and setting VPP_ISPPLOCK and
2817 			 * VPP_SETPPLOCK requires holding the segvn lock as
2818 			 * writer, but in this case all readers are
2819 			 * serializing on the anon array lock.
2820 			 */
2821 			if (AS_ISPGLCK(seg->s_as) && vpage != NULL &&
2822 			    (svd->flags & MAP_NORESERVE) &&
2823 			    !VPP_ISPPLOCK(vpage)) {
2824 				proc_t *p = seg->s_as->a_proc;
2825 				ASSERT(svd->type == MAP_PRIVATE);
2826 				mutex_enter(&p->p_lock);
2827 				if (rctl_incr_locked_mem(p, NULL, PAGESIZE,
2828 				    1) == 0) {
2829 					claim = VPP_PROT(vpage) & PROT_WRITE;
2830 					if (page_pp_lock(pp, claim, 0)) {
2831 						VPP_SETPPLOCK(vpage);
2832 					} else {
2833 						rctl_decr_locked_mem(p, NULL,
2834 						    PAGESIZE, 1);
2835 					}
2836 				}
2837 				mutex_exit(&p->p_lock);
2838 			}
2839 
2840 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2841 			hat_memload(hat, addr, pp, prot, hat_flag);
2842 
2843 			if (!(hat_flag & HAT_LOAD_LOCK))
2844 				page_unlock(pp);
2845 
2846 			anon_array_exit(&cookie);
2847 			return (0);
2848 		}
2849 	}
2850 
2851 	/*
2852 	 * Obtain the page structure via anon_getpage() if it is
2853 	 * a private copy of an object (the result of a previous
2854 	 * copy-on-write).
2855 	 */
2856 	if (amp != NULL) {
2857 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) != NULL) {
2858 			err = anon_getpage(&ap, &vpprot, anon_pl, PAGESIZE,
2859 			    seg, addr, rw, svd->cred);
2860 			if (err)
2861 				goto out;
2862 
2863 			if (svd->type == MAP_SHARED) {
2864 				/*
2865 				 * If this is a shared mapping to an
2866 				 * anon_map, then ignore the write
2867 				 * permissions returned by anon_getpage().
2868 				 * They apply to the private mappings
2869 				 * of this anon_map.
2870 				 */
2871 				vpprot |= PROT_WRITE;
2872 			}
2873 			opp = anon_pl[0];
2874 		}
2875 	}
2876 
2877 	/*
2878 	 * Search the pl[] list passed in if it is from the
2879 	 * original object (i.e., not a private copy).
2880 	 */
2881 	if (opp == NULL) {
2882 		/*
2883 		 * Find original page.  We must be bringing it in
2884 		 * from the list in pl[].
2885 		 */
2886 		for (ppp = pl; (opp = *ppp) != NULL; ppp++) {
2887 			if (opp == PAGE_HANDLED)
2888 				continue;
2889 			ASSERT(opp->p_vnode == svd->vp); /* XXX */
2890 			if (opp->p_offset == off)
2891 				break;
2892 		}
2893 		if (opp == NULL) {
2894 			panic("segvn_faultpage not found");
2895 			/*NOTREACHED*/
2896 		}
2897 		*ppp = PAGE_HANDLED;
2898 
2899 	}
2900 
2901 	ASSERT(PAGE_LOCKED(opp));
2902 
2903 	TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2904 	    "segvn_fault:pp %p vp %p offset %llx", opp, NULL, 0);
2905 
2906 	/*
2907 	 * The fault is treated as a copy-on-write fault if a
2908 	 * write occurs on a private segment and the object
2909 	 * page (i.e., mapping) is write protected.  We assume
2910 	 * that fatal protection checks have already been made.
2911 	 */
2912 
2913 	if (brkcow) {
2914 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2915 		cow = !(vpprot & PROT_WRITE);
2916 	} else if (svd->tr_state == SEGVN_TR_ON) {
2917 		/*
2918 		 * If we are doing text replication COW on first touch.
2919 		 */
2920 		ASSERT(amp != NULL);
2921 		ASSERT(svd->vp != NULL);
2922 		ASSERT(rw != S_WRITE);
2923 		cow = (ap == NULL);
2924 	} else {
2925 		cow = 0;
2926 	}
2927 
2928 	/*
2929 	 * If not a copy-on-write case load the translation
2930 	 * and return.
2931 	 */
2932 	if (cow == 0) {
2933 
2934 		/*
2935 		 * Handle pages that have been marked for migration
2936 		 */
2937 		if (lgrp_optimizations())
2938 			page_migrate(seg, addr, &opp, 1);
2939 
2940 		if (IS_VMODSORT(opp->p_vnode) || enable_mbit_wa) {
2941 			if (rw == S_WRITE)
2942 				hat_setmod(opp);
2943 			else if (rw != S_OTHER && !hat_ismod(opp))
2944 				prot &= ~PROT_WRITE;
2945 		}
2946 
2947 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
2948 		    (!svd->pageprot && svd->prot == (prot & vpprot)));
2949 		ASSERT(amp == NULL ||
2950 		    svd->rcookie == HAT_INVALID_REGION_COOKIE);
2951 		hat_memload_region(hat, addr, opp, prot & vpprot, hat_flag,
2952 		    svd->rcookie);
2953 
2954 		if (!(hat_flag & HAT_LOAD_LOCK))
2955 			page_unlock(opp);
2956 
2957 		if (anon_lock) {
2958 			anon_array_exit(&cookie);
2959 		}
2960 		return (0);
2961 	}
2962 
2963 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2964 
2965 	hat_setref(opp);
2966 
2967 	ASSERT(amp != NULL && anon_lock);
2968 
2969 	/*
2970 	 * Steal the page only if it isn't a private page
2971 	 * since stealing a private page is not worth the effort.
2972 	 */
2973 	if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL)
2974 		steal = 1;
2975 
2976 	/*
2977 	 * Steal the original page if the following conditions are true:
2978 	 *
2979 	 * We are low on memory, the page is not private, page is not large,
2980 	 * not shared, not modified, not `locked' or if we have it `locked'
2981 	 * (i.e., p_cowcnt == 1 and p_lckcnt == 0, which also implies
2982 	 * that the page is not shared) and if it doesn't have any
2983 	 * translations. page_struct_lock isn't needed to look at p_cowcnt
2984 	 * and p_lckcnt because we first get exclusive lock on page.
2985 	 */
2986 	(void) hat_pagesync(opp, HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD);
2987 
2988 	if (stealcow && freemem < minfree && steal && opp->p_szc == 0 &&
2989 	    page_tryupgrade(opp) && !hat_ismod(opp) &&
2990 	    ((opp->p_lckcnt == 0 && opp->p_cowcnt == 0) ||
2991 	    (opp->p_lckcnt == 0 && opp->p_cowcnt == 1 &&
2992 	    vpage != NULL && VPP_ISPPLOCK(vpage)))) {
2993 		/*
2994 		 * Check if this page has other translations
2995 		 * after unloading our translation.
2996 		 */
2997 		if (hat_page_is_mapped(opp)) {
2998 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2999 			hat_unload(seg->s_as->a_hat, addr, PAGESIZE,
3000 			    HAT_UNLOAD);
3001 		}
3002 
3003 		/*
3004 		 * hat_unload() might sync back someone else's recent
3005 		 * modification, so check again.
3006 		 */
3007 		if (!hat_ismod(opp) && !hat_page_is_mapped(opp))
3008 			pageflags |= STEAL_PAGE;
3009 	}
3010 
3011 	/*
3012 	 * If we have a vpage pointer, see if it indicates that we have
3013 	 * ``locked'' the page we map -- if so, tell anon_private to
3014 	 * transfer the locking resource to the new page.
3015 	 *
3016 	 * See Statement at the beginning of segvn_lockop regarding
3017 	 * the way lockcnts/cowcnts are handled during COW.
3018 	 *
3019 	 */
3020 	if (vpage != NULL && VPP_ISPPLOCK(vpage))
3021 		pageflags |= LOCK_PAGE;
3022 
3023 	/*
3024 	 * Allocate a private page and perform the copy.
3025 	 * For MAP_NORESERVE reserve swap space now, unless this
3026 	 * is a cow fault on an existing anon page in which case
3027 	 * MAP_NORESERVE will have made advance reservations.
3028 	 */
3029 	if ((svd->flags & MAP_NORESERVE) && (ap == NULL)) {
3030 		if (anon_resv_zone(ptob(1), seg->s_as->a_proc->p_zone)) {
3031 			atomic_add_long(&svd->swresv, ptob(1));
3032 			atomic_add_long(&seg->s_as->a_resvsize, ptob(1));
3033 		} else {
3034 			page_unlock(opp);
3035 			err = ENOMEM;
3036 			goto out;
3037 		}
3038 	}
3039 	oldap = ap;
3040 	pp = anon_private(&ap, seg, addr, prot, opp, pageflags, svd->cred);
3041 	if (pp == NULL) {
3042 		err = ENOMEM;	/* out of swap space */
3043 		goto out;
3044 	}
3045 
3046 	/*
3047 	 * If we copied away from an anonymous page, then
3048 	 * we are one step closer to freeing up an anon slot.
3049 	 *
3050 	 * NOTE:  The original anon slot must be released while
3051 	 * holding the "anon_map" lock.  This is necessary to prevent
3052 	 * other threads from obtaining a pointer to the anon slot
3053 	 * which may be freed if its "refcnt" is 1.
3054 	 */
3055 	if (oldap != NULL)
3056 		anon_decref(oldap);
3057 
3058 	(void) anon_set_ptr(amp->ahp, anon_index, ap, ANON_SLEEP);
3059 
3060 	/*
3061 	 * Handle pages that have been marked for migration
3062 	 */
3063 	if (lgrp_optimizations())
3064 		page_migrate(seg, addr, &pp, 1);
3065 
3066 	ASSERT(pp->p_szc == 0);
3067 
3068 	ASSERT(!IS_VMODSORT(pp->p_vnode));
3069 	if (enable_mbit_wa) {
3070 		if (rw == S_WRITE)
3071 			hat_setmod(pp);
3072 		else if (!hat_ismod(pp))
3073 			prot &= ~PROT_WRITE;
3074 	}
3075 
3076 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3077 	hat_memload(hat, addr, pp, prot, hat_flag);
3078 
3079 	if (!(hat_flag & HAT_LOAD_LOCK))
3080 		page_unlock(pp);
3081 
3082 	ASSERT(anon_lock);
3083 	anon_array_exit(&cookie);
3084 	return (0);
3085 out:
3086 	if (anon_lock)
3087 		anon_array_exit(&cookie);
3088 
3089 	if (type == F_SOFTLOCK) {
3090 		atomic_dec_ulong((ulong_t *)&svd->softlockcnt);
3091 	}
3092 	return (FC_MAKE_ERR(err));
3093 }
3094 
3095 /*
3096  * relocate a bunch of smaller targ pages into one large repl page. all targ
3097  * pages must be complete pages smaller than replacement pages.
3098  * it's assumed that no page's szc can change since they are all PAGESIZE or
3099  * complete large pages locked SHARED.
3100  */
3101 static void
3102 segvn_relocate_pages(page_t **targ, page_t *replacement)
3103 {
3104 	page_t *pp;
3105 	pgcnt_t repl_npgs, curnpgs;
3106 	pgcnt_t i;
3107 	uint_t repl_szc = replacement->p_szc;
3108 	page_t *first_repl = replacement;
3109 	page_t *repl;
3110 	spgcnt_t npgs;
3111 
3112 	VM_STAT_ADD(segvnvmstats.relocatepages[0]);
3113 
3114 	ASSERT(repl_szc != 0);
3115 	npgs = repl_npgs = page_get_pagecnt(repl_szc);
3116 
3117 	i = 0;
3118 	while (repl_npgs) {
3119 		spgcnt_t nreloc;
3120 		int err;
3121 		ASSERT(replacement != NULL);
3122 		pp = targ[i];
3123 		ASSERT(pp->p_szc < repl_szc);
3124 		ASSERT(PAGE_EXCL(pp));
3125 		ASSERT(!PP_ISFREE(pp));
3126 		curnpgs = page_get_pagecnt(pp->p_szc);
3127 		if (curnpgs == 1) {
3128 			VM_STAT_ADD(segvnvmstats.relocatepages[1]);
3129 			repl = replacement;
3130 			page_sub(&replacement, repl);
3131 			ASSERT(PAGE_EXCL(repl));
3132 			ASSERT(!PP_ISFREE(repl));
3133 			ASSERT(repl->p_szc == repl_szc);
3134 		} else {
3135 			page_t *repl_savepp;
3136 			int j;
3137 			VM_STAT_ADD(segvnvmstats.relocatepages[2]);
3138 			repl_savepp = replacement;
3139 			for (j = 0; j < curnpgs; j++) {
3140 				repl = replacement;
3141 				page_sub(&replacement, repl);
3142 				ASSERT(PAGE_EXCL(repl));
3143 				ASSERT(!PP_ISFREE(repl));
3144 				ASSERT(repl->p_szc == repl_szc);
3145 				ASSERT(page_pptonum(targ[i + j]) ==
3146 				    page_pptonum(targ[i]) + j);
3147 			}
3148 			repl = repl_savepp;
3149 			ASSERT(IS_P2ALIGNED(page_pptonum(repl), curnpgs));
3150 		}
3151 		err = page_relocate(&pp, &repl, 0, 1, &nreloc, NULL);
3152 		if (err || nreloc != curnpgs) {
3153 			panic("segvn_relocate_pages: "
3154 			    "page_relocate failed err=%d curnpgs=%ld "
3155 			    "nreloc=%ld", err, curnpgs, nreloc);
3156 		}
3157 		ASSERT(curnpgs <= repl_npgs);
3158 		repl_npgs -= curnpgs;
3159 		i += curnpgs;
3160 	}
3161 	ASSERT(replacement == NULL);
3162 
3163 	repl = first_repl;
3164 	repl_npgs = npgs;
3165 	for (i = 0; i < repl_npgs; i++) {
3166 		ASSERT(PAGE_EXCL(repl));
3167 		ASSERT(!PP_ISFREE(repl));
3168 		targ[i] = repl;
3169 		page_downgrade(targ[i]);
3170 		repl++;
3171 	}
3172 }
3173 
3174 /*
3175  * Check if all pages in ppa array are complete smaller than szc pages and
3176  * their roots will still be aligned relative to their current size if the
3177  * entire ppa array is relocated into one szc page. If these conditions are
3178  * not met return 0.
3179  *
3180  * If all pages are properly aligned attempt to upgrade their locks
3181  * to exclusive mode. If it fails set *upgrdfail to 1 and return 0.
3182  * upgrdfail was set to 0 by caller.
3183  *
3184  * Return 1 if all pages are aligned and locked exclusively.
3185  *
3186  * If all pages in ppa array happen to be physically contiguous to make one
3187  * szc page and all exclusive locks are successfully obtained promote the page
3188  * size to szc and set *pszc to szc. Return 1 with pages locked shared.
3189  */
3190 static int
3191 segvn_full_szcpages(page_t **ppa, uint_t szc, int *upgrdfail, uint_t *pszc)
3192 {
3193 	page_t *pp;
3194 	pfn_t pfn;
3195 	pgcnt_t totnpgs = page_get_pagecnt(szc);
3196 	pfn_t first_pfn;
3197 	int contig = 1;
3198 	pgcnt_t i;
3199 	pgcnt_t j;
3200 	uint_t curszc;
3201 	pgcnt_t curnpgs;
3202 	int root = 0;
3203 
3204 	ASSERT(szc > 0);
3205 
3206 	VM_STAT_ADD(segvnvmstats.fullszcpages[0]);
3207 
3208 	for (i = 0; i < totnpgs; i++) {
3209 		pp = ppa[i];
3210 		ASSERT(PAGE_SHARED(pp));
3211 		ASSERT(!PP_ISFREE(pp));
3212 		pfn = page_pptonum(pp);
3213 		if (i == 0) {
3214 			if (!IS_P2ALIGNED(pfn, totnpgs)) {
3215 				contig = 0;
3216 			} else {
3217 				first_pfn = pfn;
3218 			}
3219 		} else if (contig && pfn != first_pfn + i) {
3220 			contig = 0;
3221 		}
3222 		if (pp->p_szc == 0) {
3223 			if (root) {
3224 				VM_STAT_ADD(segvnvmstats.fullszcpages[1]);
3225 				return (0);
3226 			}
3227 		} else if (!root) {
3228 			if ((curszc = pp->p_szc) >= szc) {
3229 				VM_STAT_ADD(segvnvmstats.fullszcpages[2]);
3230 				return (0);
3231 			}
3232 			if (curszc == 0) {
3233 				/*
3234 				 * p_szc changed means we don't have all pages
3235 				 * locked. return failure.
3236 				 */
3237 				VM_STAT_ADD(segvnvmstats.fullszcpages[3]);
3238 				return (0);
3239 			}
3240 			curnpgs = page_get_pagecnt(curszc);
3241 			if (!IS_P2ALIGNED(pfn, curnpgs) ||
3242 			    !IS_P2ALIGNED(i, curnpgs)) {
3243 				VM_STAT_ADD(segvnvmstats.fullszcpages[4]);
3244 				return (0);
3245 			}
3246 			root = 1;
3247 		} else {
3248 			ASSERT(i > 0);
3249 			VM_STAT_ADD(segvnvmstats.fullszcpages[5]);
3250 			if (pp->p_szc != curszc) {
3251 				VM_STAT_ADD(segvnvmstats.fullszcpages[6]);
3252 				return (0);
3253 			}
3254 			if (pfn - 1 != page_pptonum(ppa[i - 1])) {
3255 				panic("segvn_full_szcpages: "
3256 				    "large page not physically contiguous");
3257 			}
3258 			if (P2PHASE(pfn, curnpgs) == curnpgs - 1) {
3259 				root = 0;
3260 			}
3261 		}
3262 	}
3263 
3264 	for (i = 0; i < totnpgs; i++) {
3265 		ASSERT(ppa[i]->p_szc < szc);
3266 		if (!page_tryupgrade(ppa[i])) {
3267 			for (j = 0; j < i; j++) {
3268 				page_downgrade(ppa[j]);
3269 			}
3270 			*pszc = ppa[i]->p_szc;
3271 			*upgrdfail = 1;
3272 			VM_STAT_ADD(segvnvmstats.fullszcpages[7]);
3273 			return (0);
3274 		}
3275 	}
3276 
3277 	/*
3278 	 * When a page is put a free cachelist its szc is set to 0.  if file
3279 	 * system reclaimed pages from cachelist targ pages will be physically
3280 	 * contiguous with 0 p_szc.  in this case just upgrade szc of targ
3281 	 * pages without any relocations.
3282 	 * To avoid any hat issues with previous small mappings
3283 	 * hat_pageunload() the target pages first.
3284 	 */
3285 	if (contig) {
3286 		VM_STAT_ADD(segvnvmstats.fullszcpages[8]);
3287 		for (i = 0; i < totnpgs; i++) {
3288 			(void) hat_pageunload(ppa[i], HAT_FORCE_PGUNLOAD);
3289 		}
3290 		for (i = 0; i < totnpgs; i++) {
3291 			ppa[i]->p_szc = szc;
3292 		}
3293 		for (i = 0; i < totnpgs; i++) {
3294 			ASSERT(PAGE_EXCL(ppa[i]));
3295 			page_downgrade(ppa[i]);
3296 		}
3297 		if (pszc != NULL) {
3298 			*pszc = szc;
3299 		}
3300 	}
3301 	VM_STAT_ADD(segvnvmstats.fullszcpages[9]);
3302 	return (1);
3303 }
3304 
3305 /*
3306  * Create physically contiguous pages for [vp, off] - [vp, off +
3307  * page_size(szc)) range and for private segment return them in ppa array.
3308  * Pages are created either via IO or relocations.
3309  *
3310  * Return 1 on success and 0 on failure.
3311  *
3312  * If physically contiguous pages already exist for this range return 1 without
3313  * filling ppa array. Caller initializes ppa[0] as NULL to detect that ppa
3314  * array wasn't filled. In this case caller fills ppa array via VOP_GETPAGE().
3315  */
3316 
3317 static int
3318 segvn_fill_vp_pages(struct segvn_data *svd, vnode_t *vp, u_offset_t off,
3319     uint_t szc, page_t **ppa, page_t **ppplist, uint_t *ret_pszc,
3320     int *downsize)
3321 
3322 {
3323 	page_t *pplist = *ppplist;
3324 	size_t pgsz = page_get_pagesize(szc);
3325 	pgcnt_t pages = btop(pgsz);
3326 	ulong_t start_off = off;
3327 	u_offset_t eoff = off + pgsz;
3328 	spgcnt_t nreloc;
3329 	u_offset_t io_off = off;
3330 	size_t io_len;
3331 	page_t *io_pplist = NULL;
3332 	page_t *done_pplist = NULL;
3333 	pgcnt_t pgidx = 0;
3334 	page_t *pp;
3335 	page_t *newpp;
3336 	page_t *targpp;
3337 	int io_err = 0;
3338 	int i;
3339 	pfn_t pfn;
3340 	ulong_t ppages;
3341 	page_t *targ_pplist = NULL;
3342 	page_t *repl_pplist = NULL;
3343 	page_t *tmp_pplist;
3344 	int nios = 0;
3345 	uint_t pszc;
3346 	struct vattr va;
3347 
3348 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[0]);
3349 
3350 	ASSERT(szc != 0);
3351 	ASSERT(pplist->p_szc == szc);
3352 
3353 	/*
3354 	 * downsize will be set to 1 only if we fail to lock pages. this will
3355 	 * allow subsequent faults to try to relocate the page again. If we
3356 	 * fail due to misalignment don't downsize and let the caller map the
3357 	 * whole region with small mappings to avoid more faults into the area
3358 	 * where we can't get large pages anyway.
3359 	 */
3360 	*downsize = 0;
3361 
3362 	while (off < eoff) {
3363 		newpp = pplist;
3364 		ASSERT(newpp != NULL);
3365 		ASSERT(PAGE_EXCL(newpp));
3366 		ASSERT(!PP_ISFREE(newpp));
3367 		/*
3368 		 * we pass NULL for nrelocp to page_lookup_create()
3369 		 * so that it doesn't relocate. We relocate here
3370 		 * later only after we make sure we can lock all
3371 		 * pages in the range we handle and they are all
3372 		 * aligned.
3373 		 */
3374 		pp = page_lookup_create(vp, off, SE_SHARED, newpp, NULL, 0);
3375 		ASSERT(pp != NULL);
3376 		ASSERT(!PP_ISFREE(pp));
3377 		ASSERT(pp->p_vnode == vp);
3378 		ASSERT(pp->p_offset == off);
3379 		if (pp == newpp) {
3380 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[1]);
3381 			page_sub(&pplist, pp);
3382 			ASSERT(PAGE_EXCL(pp));
3383 			ASSERT(page_iolock_assert(pp));
3384 			page_list_concat(&io_pplist, &pp);
3385 			off += PAGESIZE;
3386 			continue;
3387 		}
3388 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[2]);
3389 		pfn = page_pptonum(pp);
3390 		pszc = pp->p_szc;
3391 		if (pszc >= szc && targ_pplist == NULL && io_pplist == NULL &&
3392 		    IS_P2ALIGNED(pfn, pages)) {
3393 			ASSERT(repl_pplist == NULL);
3394 			ASSERT(done_pplist == NULL);
3395 			ASSERT(pplist == *ppplist);
3396 			page_unlock(pp);
3397 			page_free_replacement_page(pplist);
3398 			page_create_putback(pages);
3399 			*ppplist = NULL;
3400 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[3]);
3401 			return (1);
3402 		}
3403 		if (pszc >= szc) {
3404 			page_unlock(pp);
3405 			segvn_faultvnmpss_align_err1++;
3406 			goto out;
3407 		}
3408 		ppages = page_get_pagecnt(pszc);
3409 		if (!IS_P2ALIGNED(pfn, ppages)) {
3410 			ASSERT(pszc > 0);
3411 			/*
3412 			 * sizing down to pszc won't help.
3413 			 */
3414 			page_unlock(pp);
3415 			segvn_faultvnmpss_align_err2++;
3416 			goto out;
3417 		}
3418 		pfn = page_pptonum(newpp);
3419 		if (!IS_P2ALIGNED(pfn, ppages)) {
3420 			ASSERT(pszc > 0);
3421 			/*
3422 			 * sizing down to pszc won't help.
3423 			 */
3424 			page_unlock(pp);
3425 			segvn_faultvnmpss_align_err3++;
3426 			goto out;
3427 		}
3428 		if (!PAGE_EXCL(pp)) {
3429 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[4]);
3430 			page_unlock(pp);
3431 			*downsize = 1;
3432 			*ret_pszc = pp->p_szc;
3433 			goto out;
3434 		}
3435 		targpp = pp;
3436 		if (io_pplist != NULL) {
3437 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[5]);
3438 			io_len = off - io_off;
3439 			/*
3440 			 * Some file systems like NFS don't check EOF
3441 			 * conditions in VOP_PAGEIO(). Check it here
3442 			 * now that pages are locked SE_EXCL. Any file
3443 			 * truncation will wait until the pages are
3444 			 * unlocked so no need to worry that file will
3445 			 * be truncated after we check its size here.
3446 			 * XXX fix NFS to remove this check.
3447 			 */
3448 			va.va_mask = AT_SIZE;
3449 			if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL)) {
3450 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[6]);
3451 				page_unlock(targpp);
3452 				goto out;
3453 			}
3454 			if (btopr(va.va_size) < btopr(io_off + io_len)) {
3455 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[7]);
3456 				*downsize = 1;
3457 				*ret_pszc = 0;
3458 				page_unlock(targpp);
3459 				goto out;
3460 			}
3461 			io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3462 				B_READ, svd->cred, NULL);
3463 			if (io_err) {
3464 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[8]);
3465 				page_unlock(targpp);
3466 				if (io_err == EDEADLK) {
3467 					segvn_vmpss_pageio_deadlk_err++;
3468 				}
3469 				goto out;
3470 			}
3471 			nios++;
3472 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[9]);
3473 			while (io_pplist != NULL) {
3474 				pp = io_pplist;
3475 				page_sub(&io_pplist, pp);
3476 				ASSERT(page_iolock_assert(pp));
3477 				page_io_unlock(pp);
3478 				pgidx = (pp->p_offset - start_off) >>
3479 				    PAGESHIFT;
3480 				ASSERT(pgidx < pages);
3481 				ppa[pgidx] = pp;
3482 				page_list_concat(&done_pplist, &pp);
3483 			}
3484 		}
3485 		pp = targpp;
3486 		ASSERT(PAGE_EXCL(pp));
3487 		ASSERT(pp->p_szc <= pszc);
3488 		if (pszc != 0 && !group_page_trylock(pp, SE_EXCL)) {
3489 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[10]);
3490 			page_unlock(pp);
3491 			*downsize = 1;
3492 			*ret_pszc = pp->p_szc;
3493 			goto out;
3494 		}
3495 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[11]);
3496 		/*
3497 		 * page szc chould have changed before the entire group was
3498 		 * locked. reread page szc.
3499 		 */
3500 		pszc = pp->p_szc;
3501 		ppages = page_get_pagecnt(pszc);
3502 
3503 		/* link just the roots */
3504 		page_list_concat(&targ_pplist, &pp);
3505 		page_sub(&pplist, newpp);
3506 		page_list_concat(&repl_pplist, &newpp);
3507 		off += PAGESIZE;
3508 		while (--ppages != 0) {
3509 			newpp = pplist;
3510 			page_sub(&pplist, newpp);
3511 			off += PAGESIZE;
3512 		}
3513 		io_off = off;
3514 	}
3515 	if (io_pplist != NULL) {
3516 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[12]);
3517 		io_len = eoff - io_off;
3518 		va.va_mask = AT_SIZE;
3519 		if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL) != 0) {
3520 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[13]);
3521 			goto out;
3522 		}
3523 		if (btopr(va.va_size) < btopr(io_off + io_len)) {
3524 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[14]);
3525 			*downsize = 1;
3526 			*ret_pszc = 0;
3527 			goto out;
3528 		}
3529 		io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3530 		    B_READ, svd->cred, NULL);
3531 		if (io_err) {
3532 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[15]);
3533 			if (io_err == EDEADLK) {
3534 				segvn_vmpss_pageio_deadlk_err++;
3535 			}
3536 			goto out;
3537 		}
3538 		nios++;
3539 		while (io_pplist != NULL) {
3540 			pp = io_pplist;
3541 			page_sub(&io_pplist, pp);
3542 			ASSERT(page_iolock_assert(pp));
3543 			page_io_unlock(pp);
3544 			pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3545 			ASSERT(pgidx < pages);
3546 			ppa[pgidx] = pp;
3547 		}
3548 	}
3549 	/*
3550 	 * we're now bound to succeed or panic.
3551 	 * remove pages from done_pplist. it's not needed anymore.
3552 	 */
3553 	while (done_pplist != NULL) {
3554 		pp = done_pplist;
3555 		page_sub(&done_pplist, pp);
3556 	}
3557 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[16]);
3558 	ASSERT(pplist == NULL);
3559 	*ppplist = NULL;
3560 	while (targ_pplist != NULL) {
3561 		int ret;
3562 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[17]);
3563 		ASSERT(repl_pplist);
3564 		pp = targ_pplist;
3565 		page_sub(&targ_pplist, pp);
3566 		pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3567 		newpp = repl_pplist;
3568 		page_sub(&repl_pplist, newpp);
3569 #ifdef DEBUG
3570 		pfn = page_pptonum(pp);
3571 		pszc = pp->p_szc;
3572 		ppages = page_get_pagecnt(pszc);
3573 		ASSERT(IS_P2ALIGNED(pfn, ppages));
3574 		pfn = page_pptonum(newpp);
3575 		ASSERT(IS_P2ALIGNED(pfn, ppages));
3576 		ASSERT(P2PHASE(pfn, pages) == pgidx);
3577 #endif
3578 		nreloc = 0;
3579 		ret = page_relocate(&pp, &newpp, 0, 1, &nreloc, NULL);
3580 		if (ret != 0 || nreloc == 0) {
3581 			panic("segvn_fill_vp_pages: "
3582 			    "page_relocate failed");
3583 		}
3584 		pp = newpp;
3585 		while (nreloc-- != 0) {
3586 			ASSERT(PAGE_EXCL(pp));
3587 			ASSERT(pp->p_vnode == vp);
3588 			ASSERT(pgidx ==
3589 			    ((pp->p_offset - start_off) >> PAGESHIFT));
3590 			ppa[pgidx++] = pp;
3591 			pp++;
3592 		}
3593 	}
3594 
3595 	if (svd->type == MAP_PRIVATE) {
3596 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[18]);
3597 		for (i = 0; i < pages; i++) {
3598 			ASSERT(ppa[i] != NULL);
3599 			ASSERT(PAGE_EXCL(ppa[i]));
3600 			ASSERT(ppa[i]->p_vnode == vp);
3601 			ASSERT(ppa[i]->p_offset ==
3602 			    start_off + (i << PAGESHIFT));
3603 			page_downgrade(ppa[i]);
3604 		}
3605 		ppa[pages] = NULL;
3606 	} else {
3607 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[19]);
3608 		/*
3609 		 * the caller will still call VOP_GETPAGE() for shared segments
3610 		 * to check FS write permissions. For private segments we map
3611 		 * file read only anyway.  so no VOP_GETPAGE is needed.
3612 		 */
3613 		for (i = 0; i < pages; i++) {
3614 			ASSERT(ppa[i] != NULL);
3615 			ASSERT(PAGE_EXCL(ppa[i]));
3616 			ASSERT(ppa[i]->p_vnode == vp);
3617 			ASSERT(ppa[i]->p_offset ==
3618 			    start_off + (i << PAGESHIFT));
3619 			page_unlock(ppa[i]);
3620 		}
3621 		ppa[0] = NULL;
3622 	}
3623 
3624 	return (1);
3625 out:
3626 	/*
3627 	 * Do the cleanup. Unlock target pages we didn't relocate. They are
3628 	 * linked on targ_pplist by root pages. reassemble unused replacement
3629 	 * and io pages back to pplist.
3630 	 */
3631 	if (io_pplist != NULL) {
3632 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[20]);
3633 		pp = io_pplist;
3634 		do {
3635 			ASSERT(pp->p_vnode == vp);
3636 			ASSERT(pp->p_offset == io_off);
3637 			ASSERT(page_iolock_assert(pp));
3638 			page_io_unlock(pp);
3639 			page_hashout(pp, NULL);
3640 			io_off += PAGESIZE;
3641 		} while ((pp = pp->p_next) != io_pplist);
3642 		page_list_concat(&io_pplist, &pplist);
3643 		pplist = io_pplist;
3644 	}
3645 	tmp_pplist = NULL;
3646 	while (targ_pplist != NULL) {
3647 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[21]);
3648 		pp = targ_pplist;
3649 		ASSERT(PAGE_EXCL(pp));
3650 		page_sub(&targ_pplist, pp);
3651 
3652 		pszc = pp->p_szc;
3653 		ppages = page_get_pagecnt(pszc);
3654 		ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3655 
3656 		if (pszc != 0) {
3657 			group_page_unlock(pp);
3658 		}
3659 		page_unlock(pp);
3660 
3661 		pp = repl_pplist;
3662 		ASSERT(pp != NULL);
3663 		ASSERT(PAGE_EXCL(pp));
3664 		ASSERT(pp->p_szc == szc);
3665 		page_sub(&repl_pplist, pp);
3666 
3667 		ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3668 
3669 		/* relink replacement page */
3670 		page_list_concat(&tmp_pplist, &pp);
3671 		while (--ppages != 0) {
3672 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[22]);
3673 			pp++;
3674 			ASSERT(PAGE_EXCL(pp));
3675 			ASSERT(pp->p_szc == szc);
3676 			page_list_concat(&tmp_pplist, &pp);
3677 		}
3678 	}
3679 	if (tmp_pplist != NULL) {
3680 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[23]);
3681 		page_list_concat(&tmp_pplist, &pplist);
3682 		pplist = tmp_pplist;
3683 	}
3684 	/*
3685 	 * at this point all pages are either on done_pplist or
3686 	 * pplist. They can't be all on done_pplist otherwise
3687 	 * we'd've been done.
3688 	 */
3689 	ASSERT(pplist != NULL);
3690 	if (nios != 0) {
3691 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[24]);
3692 		pp = pplist;
3693 		do {
3694 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[25]);
3695 			ASSERT(pp->p_szc == szc);
3696 			ASSERT(PAGE_EXCL(pp));
3697 			ASSERT(pp->p_vnode != vp);
3698 			pp->p_szc = 0;
3699 		} while ((pp = pp->p_next) != pplist);
3700 
3701 		pp = done_pplist;
3702 		do {
3703 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[26]);
3704 			ASSERT(pp->p_szc == szc);
3705 			ASSERT(PAGE_EXCL(pp));
3706 			ASSERT(pp->p_vnode == vp);
3707 			pp->p_szc = 0;
3708 		} while ((pp = pp->p_next) != done_pplist);
3709 
3710 		while (pplist != NULL) {
3711 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[27]);
3712 			pp = pplist;
3713 			page_sub(&pplist, pp);
3714 			page_free(pp, 0);
3715 		}
3716 
3717 		while (done_pplist != NULL) {
3718 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[28]);
3719 			pp = done_pplist;
3720 			page_sub(&done_pplist, pp);
3721 			page_unlock(pp);
3722 		}
3723 		*ppplist = NULL;
3724 		return (0);
3725 	}
3726 	ASSERT(pplist == *ppplist);
3727 	if (io_err) {
3728 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[29]);
3729 		/*
3730 		 * don't downsize on io error.
3731 		 * see if vop_getpage succeeds.
3732 		 * pplist may still be used in this case
3733 		 * for relocations.
3734 		 */
3735 		return (0);
3736 	}
3737 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[30]);
3738 	page_free_replacement_page(pplist);
3739 	page_create_putback(pages);
3740 	*ppplist = NULL;
3741 	return (0);
3742 }
3743 
3744 int segvn_anypgsz = 0;
3745 
3746 #define	SEGVN_RESTORE_SOFTLOCK_VP(type, pages) 				\
3747 		if ((type) == F_SOFTLOCK) {				\
3748 			atomic_add_long((ulong_t *)&(svd)->softlockcnt, \
3749 			    -(pages));					\
3750 		}
3751 
3752 #define	SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot)		\
3753 		if (IS_VMODSORT((ppa)[0]->p_vnode)) {			\
3754 			if ((rw) == S_WRITE) {				\
3755 				for (i = 0; i < (pages); i++) {		\
3756 					ASSERT((ppa)[i]->p_vnode ==	\
3757 					    (ppa)[0]->p_vnode);		\
3758 					hat_setmod((ppa)[i]);		\
3759 				}					\
3760 			} else if ((rw) != S_OTHER &&			\
3761 			    ((prot) & (vpprot) & PROT_WRITE)) {		\
3762 				for (i = 0; i < (pages); i++) {		\
3763 					ASSERT((ppa)[i]->p_vnode ==	\
3764 					    (ppa)[0]->p_vnode);		\
3765 					if (!hat_ismod((ppa)[i])) {	\
3766 						prot &= ~PROT_WRITE;	\
3767 						break;			\
3768 					}				\
3769 				}					\
3770 			}						\
3771 		}
3772 
3773 #ifdef  VM_STATS
3774 
3775 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)					\
3776 		VM_STAT_ADD(segvnvmstats.fltvnpages[(idx)]);
3777 
3778 #else /* VM_STATS */
3779 
3780 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)
3781 
3782 #endif
3783 
3784 static faultcode_t
3785 segvn_fault_vnodepages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
3786     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
3787     caddr_t eaddr, int brkcow)
3788 {
3789 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
3790 	struct anon_map *amp = svd->amp;
3791 	uchar_t segtype = svd->type;
3792 	uint_t szc = seg->s_szc;
3793 	size_t pgsz = page_get_pagesize(szc);
3794 	size_t maxpgsz = pgsz;
3795 	pgcnt_t pages = btop(pgsz);
3796 	pgcnt_t maxpages = pages;
3797 	size_t ppasize = (pages + 1) * sizeof (page_t *);
3798 	caddr_t a = lpgaddr;
3799 	caddr_t	maxlpgeaddr = lpgeaddr;
3800 	u_offset_t off = svd->offset + (uintptr_t)(a - seg->s_base);
3801 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
3802 	struct vpage *vpage = (svd->vpage != NULL) ?
3803 	    &svd->vpage[seg_page(seg, a)] : NULL;
3804 	vnode_t *vp = svd->vp;
3805 	page_t **ppa;
3806 	uint_t	pszc;
3807 	size_t	ppgsz;
3808 	pgcnt_t	ppages;
3809 	faultcode_t err = 0;
3810 	int ierr;
3811 	int vop_size_err = 0;
3812 	uint_t protchk, prot, vpprot;
3813 	ulong_t i;
3814 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
3815 	anon_sync_obj_t an_cookie;
3816 	enum seg_rw arw;
3817 	int alloc_failed = 0;
3818 	int adjszc_chk;
3819 	struct vattr va;
3820 	int xhat = 0;
3821 	page_t *pplist;
3822 	pfn_t pfn;
3823 	int physcontig;
3824 	int upgrdfail;
3825 	int segvn_anypgsz_vnode = 0; /* for now map vnode with 2 page sizes */
3826 	int tron = (svd->tr_state == SEGVN_TR_ON);
3827 
3828 	ASSERT(szc != 0);
3829 	ASSERT(vp != NULL);
3830 	ASSERT(brkcow == 0 || amp != NULL);
3831 	ASSERT(tron == 0 || amp != NULL);
3832 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
3833 	ASSERT(!(svd->flags & MAP_NORESERVE));
3834 	ASSERT(type != F_SOFTUNLOCK);
3835 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
3836 	ASSERT(amp == NULL || IS_P2ALIGNED(aindx, maxpages));
3837 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
3838 	ASSERT(seg->s_szc < NBBY * sizeof (int));
3839 	ASSERT(type != F_SOFTLOCK || lpgeaddr - a == maxpgsz);
3840 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
3841 
3842 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltvnpages[0]);
3843 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltvnpages[1]);
3844 
3845 	if (svd->flags & MAP_TEXT) {
3846 		hat_flag |= HAT_LOAD_TEXT;
3847 	}
3848 
3849 	if (svd->pageprot) {
3850 		switch (rw) {
3851 		case S_READ:
3852 			protchk = PROT_READ;
3853 			break;
3854 		case S_WRITE:
3855 			protchk = PROT_WRITE;
3856 			break;
3857 		case S_EXEC:
3858 			protchk = PROT_EXEC;
3859 			break;
3860 		case S_OTHER:
3861 		default:
3862 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
3863 			break;
3864 		}
3865 	} else {
3866 		prot = svd->prot;
3867 		/* caller has already done segment level protection check. */
3868 	}
3869 
3870 	if (seg->s_as->a_hat != hat) {
3871 		xhat = 1;
3872 	}
3873 
3874 	if (rw == S_WRITE && segtype == MAP_PRIVATE) {
3875 		SEGVN_VMSTAT_FLTVNPAGES(2);
3876 		arw = S_READ;
3877 	} else {
3878 		arw = rw;
3879 	}
3880 
3881 	ppa = kmem_alloc(ppasize, KM_SLEEP);
3882 
3883 	VM_STAT_COND_ADD(amp != NULL, segvnvmstats.fltvnpages[3]);
3884 
3885 	for (;;) {
3886 		adjszc_chk = 0;
3887 		for (; a < lpgeaddr; a += pgsz, off += pgsz, aindx += pages) {
3888 			if (adjszc_chk) {
3889 				while (szc < seg->s_szc) {
3890 					uintptr_t e;
3891 					uint_t tszc;
3892 					tszc = segvn_anypgsz_vnode ? szc + 1 :
3893 					    seg->s_szc;
3894 					ppgsz = page_get_pagesize(tszc);
3895 					if (!IS_P2ALIGNED(a, ppgsz) ||
3896 					    ((alloc_failed >> tszc) & 0x1)) {
3897 						break;
3898 					}
3899 					SEGVN_VMSTAT_FLTVNPAGES(4);
3900 					szc = tszc;
3901 					pgsz = ppgsz;
3902 					pages = btop(pgsz);
3903 					e = P2ROUNDUP((uintptr_t)eaddr, pgsz);
3904 					lpgeaddr = (caddr_t)e;
3905 				}
3906 			}
3907 
3908 		again:
3909 			if (IS_P2ALIGNED(a, maxpgsz) && amp != NULL) {
3910 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
3911 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
3912 				anon_array_enter(amp, aindx, &an_cookie);
3913 				if (anon_get_ptr(amp->ahp, aindx) != NULL) {
3914 					SEGVN_VMSTAT_FLTVNPAGES(5);
3915 					ASSERT(anon_pages(amp->ahp, aindx,
3916 					    maxpages) == maxpages);
3917 					anon_array_exit(&an_cookie);
3918 					ANON_LOCK_EXIT(&amp->a_rwlock);
3919 					err = segvn_fault_anonpages(hat, seg,
3920 					    a, a + maxpgsz, type, rw,
3921 					    MAX(a, addr),
3922 					    MIN(a + maxpgsz, eaddr), brkcow);
3923 					if (err != 0) {
3924 						SEGVN_VMSTAT_FLTVNPAGES(6);
3925 						goto out;
3926 					}
3927 					if (szc < seg->s_szc) {
3928 						szc = seg->s_szc;
3929 						pgsz = maxpgsz;
3930 						pages = maxpages;
3931 						lpgeaddr = maxlpgeaddr;
3932 					}
3933 					goto next;
3934 				} else {
3935 					ASSERT(anon_pages(amp->ahp, aindx,
3936 					    maxpages) == 0);
3937 					SEGVN_VMSTAT_FLTVNPAGES(7);
3938 					anon_array_exit(&an_cookie);
3939 					ANON_LOCK_EXIT(&amp->a_rwlock);
3940 				}
3941 			}
3942 			ASSERT(!brkcow || IS_P2ALIGNED(a, maxpgsz));
3943 			ASSERT(!tron || IS_P2ALIGNED(a, maxpgsz));
3944 
3945 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
3946 				ASSERT(vpage != NULL);
3947 				prot = VPP_PROT(vpage);
3948 				ASSERT(sameprot(seg, a, maxpgsz));
3949 				if ((prot & protchk) == 0) {
3950 					SEGVN_VMSTAT_FLTVNPAGES(8);
3951 					err = FC_PROT;
3952 					goto out;
3953 				}
3954 			}
3955 			if (type == F_SOFTLOCK) {
3956 				atomic_add_long((ulong_t *)&svd->softlockcnt,
3957 				    pages);
3958 			}
3959 
3960 			pplist = NULL;
3961 			physcontig = 0;
3962 			ppa[0] = NULL;
3963 			if (!brkcow && !tron && szc &&
3964 			    !page_exists_physcontig(vp, off, szc,
3965 			    segtype == MAP_PRIVATE ? ppa : NULL)) {
3966 				SEGVN_VMSTAT_FLTVNPAGES(9);
3967 				if (page_alloc_pages(vp, seg, a, &pplist, NULL,
3968 				    szc, 0, 0) && type != F_SOFTLOCK) {
3969 					SEGVN_VMSTAT_FLTVNPAGES(10);
3970 					pszc = 0;
3971 					ierr = -1;
3972 					alloc_failed |= (1 << szc);
3973 					break;
3974 				}
3975 				if (pplist != NULL &&
3976 				    vp->v_mpssdata == SEGVN_PAGEIO) {
3977 					int downsize;
3978 					SEGVN_VMSTAT_FLTVNPAGES(11);
3979 					physcontig = segvn_fill_vp_pages(svd,
3980 					    vp, off, szc, ppa, &pplist,
3981 					    &pszc, &downsize);
3982 					ASSERT(!physcontig || pplist == NULL);
3983 					if (!physcontig && downsize &&
3984 					    type != F_SOFTLOCK) {
3985 						ASSERT(pplist == NULL);
3986 						SEGVN_VMSTAT_FLTVNPAGES(12);
3987 						ierr = -1;
3988 						break;
3989 					}
3990 					ASSERT(!physcontig ||
3991 					    segtype == MAP_PRIVATE ||
3992 					    ppa[0] == NULL);
3993 					if (physcontig && ppa[0] == NULL) {
3994 						physcontig = 0;
3995 					}
3996 				}
3997 			} else if (!brkcow && !tron && szc && ppa[0] != NULL) {
3998 				SEGVN_VMSTAT_FLTVNPAGES(13);
3999 				ASSERT(segtype == MAP_PRIVATE);
4000 				physcontig = 1;
4001 			}
4002 
4003 			if (!physcontig) {
4004 				SEGVN_VMSTAT_FLTVNPAGES(14);
4005 				ppa[0] = NULL;
4006 				ierr = VOP_GETPAGE(vp, (offset_t)off, pgsz,
4007 				    &vpprot, ppa, pgsz, seg, a, arw,
4008 				    svd->cred, NULL);
4009 #ifdef DEBUG
4010 				if (ierr == 0) {
4011 					for (i = 0; i < pages; i++) {
4012 						ASSERT(PAGE_LOCKED(ppa[i]));
4013 						ASSERT(!PP_ISFREE(ppa[i]));
4014 						ASSERT(ppa[i]->p_vnode == vp);
4015 						ASSERT(ppa[i]->p_offset ==
4016 						    off + (i << PAGESHIFT));
4017 					}
4018 				}
4019 #endif /* DEBUG */
4020 				if (segtype == MAP_PRIVATE) {
4021 					SEGVN_VMSTAT_FLTVNPAGES(15);
4022 					vpprot &= ~PROT_WRITE;
4023 				}
4024 			} else {
4025 				ASSERT(segtype == MAP_PRIVATE);
4026 				SEGVN_VMSTAT_FLTVNPAGES(16);
4027 				vpprot = PROT_ALL & ~PROT_WRITE;
4028 				ierr = 0;
4029 			}
4030 
4031 			if (ierr != 0) {
4032 				SEGVN_VMSTAT_FLTVNPAGES(17);
4033 				if (pplist != NULL) {
4034 					SEGVN_VMSTAT_FLTVNPAGES(18);
4035 					page_free_replacement_page(pplist);
4036 					page_create_putback(pages);
4037 				}
4038 				SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4039 				if (a + pgsz <= eaddr) {
4040 					SEGVN_VMSTAT_FLTVNPAGES(19);
4041 					err = FC_MAKE_ERR(ierr);
4042 					goto out;
4043 				}
4044 				va.va_mask = AT_SIZE;
4045 				if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL)) {
4046 					SEGVN_VMSTAT_FLTVNPAGES(20);
4047 					err = FC_MAKE_ERR(EIO);
4048 					goto out;
4049 				}
4050 				if (btopr(va.va_size) >= btopr(off + pgsz)) {
4051 					SEGVN_VMSTAT_FLTVNPAGES(21);
4052 					err = FC_MAKE_ERR(ierr);
4053 					goto out;
4054 				}
4055 				if (btopr(va.va_size) <
4056 				    btopr(off + (eaddr - a))) {
4057 					SEGVN_VMSTAT_FLTVNPAGES(22);
4058 					err = FC_MAKE_ERR(ierr);
4059 					goto out;
4060 				}
4061 				if (brkcow || tron || type == F_SOFTLOCK) {
4062 					/* can't reduce map area */
4063 					SEGVN_VMSTAT_FLTVNPAGES(23);
4064 					vop_size_err = 1;
4065 					goto out;
4066 				}
4067 				SEGVN_VMSTAT_FLTVNPAGES(24);
4068 				ASSERT(szc != 0);
4069 				pszc = 0;
4070 				ierr = -1;
4071 				break;
4072 			}
4073 
4074 			if (amp != NULL) {
4075 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4076 				anon_array_enter(amp, aindx, &an_cookie);
4077 			}
4078 			if (amp != NULL &&
4079 			    anon_get_ptr(amp->ahp, aindx) != NULL) {
4080 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4081 
4082 				SEGVN_VMSTAT_FLTVNPAGES(25);
4083 				ASSERT(anon_pages(amp->ahp, taindx,
4084 				    maxpages) == maxpages);
4085 				for (i = 0; i < pages; i++) {
4086 					page_unlock(ppa[i]);
4087 				}
4088 				anon_array_exit(&an_cookie);
4089 				ANON_LOCK_EXIT(&amp->a_rwlock);
4090 				if (pplist != NULL) {
4091 					page_free_replacement_page(pplist);
4092 					page_create_putback(pages);
4093 				}
4094 				SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4095 				if (szc < seg->s_szc) {
4096 					SEGVN_VMSTAT_FLTVNPAGES(26);
4097 					/*
4098 					 * For private segments SOFTLOCK
4099 					 * either always breaks cow (any rw
4100 					 * type except S_READ_NOCOW) or
4101 					 * address space is locked as writer
4102 					 * (S_READ_NOCOW case) and anon slots
4103 					 * can't show up on second check.
4104 					 * Therefore if we are here for
4105 					 * SOFTLOCK case it must be a cow
4106 					 * break but cow break never reduces
4107 					 * szc. text replication (tron) in
4108 					 * this case works as cow break.
4109 					 * Thus the assert below.
4110 					 */
4111 					ASSERT(!brkcow && !tron &&
4112 					    type != F_SOFTLOCK);
4113 					pszc = seg->s_szc;
4114 					ierr = -2;
4115 					break;
4116 				}
4117 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4118 				goto again;
4119 			}
4120 #ifdef DEBUG
4121 			if (amp != NULL) {
4122 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4123 				ASSERT(!anon_pages(amp->ahp, taindx, maxpages));
4124 			}
4125 #endif /* DEBUG */
4126 
4127 			if (brkcow || tron) {
4128 				ASSERT(amp != NULL);
4129 				ASSERT(pplist == NULL);
4130 				ASSERT(szc == seg->s_szc);
4131 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4132 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
4133 				SEGVN_VMSTAT_FLTVNPAGES(27);
4134 				ierr = anon_map_privatepages(amp, aindx, szc,
4135 				    seg, a, prot, ppa, vpage, segvn_anypgsz,
4136 				    tron ? PG_LOCAL : 0, svd->cred);
4137 				if (ierr != 0) {
4138 					SEGVN_VMSTAT_FLTVNPAGES(28);
4139 					anon_array_exit(&an_cookie);
4140 					ANON_LOCK_EXIT(&amp->a_rwlock);
4141 					SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4142 					err = FC_MAKE_ERR(ierr);
4143 					goto out;
4144 				}
4145 
4146 				ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4147 				/*
4148 				 * p_szc can't be changed for locked
4149 				 * swapfs pages.
4150 				 */
4151 				ASSERT(svd->rcookie ==
4152 				    HAT_INVALID_REGION_COOKIE);
4153 				hat_memload_array(hat, a, pgsz, ppa, prot,
4154 				    hat_flag);
4155 
4156 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4157 					SEGVN_VMSTAT_FLTVNPAGES(29);
4158 					for (i = 0; i < pages; i++) {
4159 						page_unlock(ppa[i]);
4160 					}
4161 				}
4162 				anon_array_exit(&an_cookie);
4163 				ANON_LOCK_EXIT(&amp->a_rwlock);
4164 				goto next;
4165 			}
4166 
4167 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
4168 			    (!svd->pageprot && svd->prot == (prot & vpprot)));
4169 
4170 			pfn = page_pptonum(ppa[0]);
4171 			/*
4172 			 * hat_page_demote() needs an SE_EXCL lock on one of
4173 			 * constituent page_t's and it decreases root's p_szc
4174 			 * last. This means if root's p_szc is equal szc and
4175 			 * all its constituent pages are locked
4176 			 * hat_page_demote() that could have changed p_szc to
4177 			 * szc is already done and no new have page_demote()
4178 			 * can start for this large page.
4179 			 */
4180 
4181 			/*
4182 			 * we need to make sure same mapping size is used for
4183 			 * the same address range if there's a possibility the
4184 			 * adddress is already mapped because hat layer panics
4185 			 * when translation is loaded for the range already
4186 			 * mapped with a different page size.  We achieve it
4187 			 * by always using largest page size possible subject
4188 			 * to the constraints of page size, segment page size
4189 			 * and page alignment.  Since mappings are invalidated
4190 			 * when those constraints change and make it
4191 			 * impossible to use previously used mapping size no
4192 			 * mapping size conflicts should happen.
4193 			 */
4194 
4195 		chkszc:
4196 			if ((pszc = ppa[0]->p_szc) == szc &&
4197 			    IS_P2ALIGNED(pfn, pages)) {
4198 
4199 				SEGVN_VMSTAT_FLTVNPAGES(30);
4200 #ifdef DEBUG
4201 				for (i = 0; i < pages; i++) {
4202 					ASSERT(PAGE_LOCKED(ppa[i]));
4203 					ASSERT(!PP_ISFREE(ppa[i]));
4204 					ASSERT(page_pptonum(ppa[i]) ==
4205 					    pfn + i);
4206 					ASSERT(ppa[i]->p_szc == szc);
4207 					ASSERT(ppa[i]->p_vnode == vp);
4208 					ASSERT(ppa[i]->p_offset ==
4209 					    off + (i << PAGESHIFT));
4210 				}
4211 #endif /* DEBUG */
4212 				/*
4213 				 * All pages are of szc we need and they are
4214 				 * all locked so they can't change szc. load
4215 				 * translations.
4216 				 *
4217 				 * if page got promoted since last check
4218 				 * we don't need pplist.
4219 				 */
4220 				if (pplist != NULL) {
4221 					page_free_replacement_page(pplist);
4222 					page_create_putback(pages);
4223 				}
4224 				if (PP_ISMIGRATE(ppa[0])) {
4225 					page_migrate(seg, a, ppa, pages);
4226 				}
4227 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4228 				    prot, vpprot);
4229 				if (!xhat) {
4230 					hat_memload_array_region(hat, a, pgsz,
4231 					    ppa, prot & vpprot, hat_flag,
4232 					    svd->rcookie);
4233 				} else {
4234 					/*
4235 					 * avoid large xhat mappings to FS
4236 					 * pages so that hat_page_demote()
4237 					 * doesn't need to check for xhat
4238 					 * large mappings.
4239 					 * Don't use regions with xhats.
4240 					 */
4241 					for (i = 0; i < pages; i++) {
4242 						hat_memload(hat,
4243 						    a + (i << PAGESHIFT),
4244 						    ppa[i], prot & vpprot,
4245 						    hat_flag);
4246 					}
4247 				}
4248 
4249 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4250 					for (i = 0; i < pages; i++) {
4251 						page_unlock(ppa[i]);
4252 					}
4253 				}
4254 				if (amp != NULL) {
4255 					anon_array_exit(&an_cookie);
4256 					ANON_LOCK_EXIT(&amp->a_rwlock);
4257 				}
4258 				goto next;
4259 			}
4260 
4261 			/*
4262 			 * See if upsize is possible.
4263 			 */
4264 			if (pszc > szc && szc < seg->s_szc &&
4265 			    (segvn_anypgsz_vnode || pszc >= seg->s_szc)) {
4266 				pgcnt_t aphase;
4267 				uint_t pszc1 = MIN(pszc, seg->s_szc);
4268 				ppgsz = page_get_pagesize(pszc1);
4269 				ppages = btop(ppgsz);
4270 				aphase = btop(P2PHASE((uintptr_t)a, ppgsz));
4271 
4272 				ASSERT(type != F_SOFTLOCK);
4273 
4274 				SEGVN_VMSTAT_FLTVNPAGES(31);
4275 				if (aphase != P2PHASE(pfn, ppages)) {
4276 					segvn_faultvnmpss_align_err4++;
4277 				} else {
4278 					SEGVN_VMSTAT_FLTVNPAGES(32);
4279 					if (pplist != NULL) {
4280 						page_t *pl = pplist;
4281 						page_free_replacement_page(pl);
4282 						page_create_putback(pages);
4283 					}
4284 					for (i = 0; i < pages; i++) {
4285 						page_unlock(ppa[i]);
4286 					}
4287 					if (amp != NULL) {
4288 						anon_array_exit(&an_cookie);
4289 						ANON_LOCK_EXIT(&amp->a_rwlock);
4290 					}
4291 					pszc = pszc1;
4292 					ierr = -2;
4293 					break;
4294 				}
4295 			}
4296 
4297 			/*
4298 			 * check if we should use smallest mapping size.
4299 			 */
4300 			upgrdfail = 0;
4301 			if (szc == 0 || xhat ||
4302 			    (pszc >= szc &&
4303 			    !IS_P2ALIGNED(pfn, pages)) ||
4304 			    (pszc < szc &&
4305 			    !segvn_full_szcpages(ppa, szc, &upgrdfail,
4306 			    &pszc))) {
4307 
4308 				if (upgrdfail && type != F_SOFTLOCK) {
4309 					/*
4310 					 * segvn_full_szcpages failed to lock
4311 					 * all pages EXCL. Size down.
4312 					 */
4313 					ASSERT(pszc < szc);
4314 
4315 					SEGVN_VMSTAT_FLTVNPAGES(33);
4316 
4317 					if (pplist != NULL) {
4318 						page_t *pl = pplist;
4319 						page_free_replacement_page(pl);
4320 						page_create_putback(pages);
4321 					}
4322 
4323 					for (i = 0; i < pages; i++) {
4324 						page_unlock(ppa[i]);
4325 					}
4326 					if (amp != NULL) {
4327 						anon_array_exit(&an_cookie);
4328 						ANON_LOCK_EXIT(&amp->a_rwlock);
4329 					}
4330 					ierr = -1;
4331 					break;
4332 				}
4333 				if (szc != 0 && !xhat && !upgrdfail) {
4334 					segvn_faultvnmpss_align_err5++;
4335 				}
4336 				SEGVN_VMSTAT_FLTVNPAGES(34);
4337 				if (pplist != NULL) {
4338 					page_free_replacement_page(pplist);
4339 					page_create_putback(pages);
4340 				}
4341 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4342 				    prot, vpprot);
4343 				if (upgrdfail && segvn_anypgsz_vnode) {
4344 					/* SOFTLOCK case */
4345 					hat_memload_array_region(hat, a, pgsz,
4346 					    ppa, prot & vpprot, hat_flag,
4347 					    svd->rcookie);
4348 				} else {
4349 					for (i = 0; i < pages; i++) {
4350 						hat_memload_region(hat,
4351 						    a + (i << PAGESHIFT),
4352 						    ppa[i], prot & vpprot,
4353 						    hat_flag, svd->rcookie);
4354 					}
4355 				}
4356 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4357 					for (i = 0; i < pages; i++) {
4358 						page_unlock(ppa[i]);
4359 					}
4360 				}
4361 				if (amp != NULL) {
4362 					anon_array_exit(&an_cookie);
4363 					ANON_LOCK_EXIT(&amp->a_rwlock);
4364 				}
4365 				goto next;
4366 			}
4367 
4368 			if (pszc == szc) {
4369 				/*
4370 				 * segvn_full_szcpages() upgraded pages szc.
4371 				 */
4372 				ASSERT(pszc == ppa[0]->p_szc);
4373 				ASSERT(IS_P2ALIGNED(pfn, pages));
4374 				goto chkszc;
4375 			}
4376 
4377 			if (pszc > szc) {
4378 				kmutex_t *szcmtx;
4379 				SEGVN_VMSTAT_FLTVNPAGES(35);
4380 				/*
4381 				 * p_szc of ppa[0] can change since we haven't
4382 				 * locked all constituent pages. Call
4383 				 * page_lock_szc() to prevent szc changes.
4384 				 * This should be a rare case that happens when
4385 				 * multiple segments use a different page size
4386 				 * to map the same file offsets.
4387 				 */
4388 				szcmtx = page_szc_lock(ppa[0]);
4389 				pszc = ppa[0]->p_szc;
4390 				ASSERT(szcmtx != NULL || pszc == 0);
4391 				ASSERT(ppa[0]->p_szc <= pszc);
4392 				if (pszc <= szc) {
4393 					SEGVN_VMSTAT_FLTVNPAGES(36);
4394 					if (szcmtx != NULL) {
4395 						mutex_exit(szcmtx);
4396 					}
4397 					goto chkszc;
4398 				}
4399 				if (pplist != NULL) {
4400 					/*
4401 					 * page got promoted since last check.
4402 					 * we don't need preaalocated large
4403 					 * page.
4404 					 */
4405 					SEGVN_VMSTAT_FLTVNPAGES(37);
4406 					page_free_replacement_page(pplist);
4407 					page_create_putback(pages);
4408 				}
4409 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4410 				    prot, vpprot);
4411 				hat_memload_array_region(hat, a, pgsz, ppa,
4412 				    prot & vpprot, hat_flag, svd->rcookie);
4413 				mutex_exit(szcmtx);
4414 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4415 					for (i = 0; i < pages; i++) {
4416 						page_unlock(ppa[i]);
4417 					}
4418 				}
4419 				if (amp != NULL) {
4420 					anon_array_exit(&an_cookie);
4421 					ANON_LOCK_EXIT(&amp->a_rwlock);
4422 				}
4423 				goto next;
4424 			}
4425 
4426 			/*
4427 			 * if page got demoted since last check
4428 			 * we could have not allocated larger page.
4429 			 * allocate now.
4430 			 */
4431 			if (pplist == NULL &&
4432 			    page_alloc_pages(vp, seg, a, &pplist, NULL,
4433 			    szc, 0, 0) && type != F_SOFTLOCK) {
4434 				SEGVN_VMSTAT_FLTVNPAGES(38);
4435 				for (i = 0; i < pages; i++) {
4436 					page_unlock(ppa[i]);
4437 				}
4438 				if (amp != NULL) {
4439 					anon_array_exit(&an_cookie);
4440 					ANON_LOCK_EXIT(&amp->a_rwlock);
4441 				}
4442 				ierr = -1;
4443 				alloc_failed |= (1 << szc);
4444 				break;
4445 			}
4446 
4447 			SEGVN_VMSTAT_FLTVNPAGES(39);
4448 
4449 			if (pplist != NULL) {
4450 				segvn_relocate_pages(ppa, pplist);
4451 #ifdef DEBUG
4452 			} else {
4453 				ASSERT(type == F_SOFTLOCK);
4454 				SEGVN_VMSTAT_FLTVNPAGES(40);
4455 #endif /* DEBUG */
4456 			}
4457 
4458 			SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot);
4459 
4460 			if (pplist == NULL && segvn_anypgsz_vnode == 0) {
4461 				ASSERT(type == F_SOFTLOCK);
4462 				for (i = 0; i < pages; i++) {
4463 					ASSERT(ppa[i]->p_szc < szc);
4464 					hat_memload_region(hat,
4465 					    a + (i << PAGESHIFT),
4466 					    ppa[i], prot & vpprot, hat_flag,
4467 					    svd->rcookie);
4468 				}
4469 			} else {
4470 				ASSERT(pplist != NULL || type == F_SOFTLOCK);
4471 				hat_memload_array_region(hat, a, pgsz, ppa,
4472 				    prot & vpprot, hat_flag, svd->rcookie);
4473 			}
4474 			if (!(hat_flag & HAT_LOAD_LOCK)) {
4475 				for (i = 0; i < pages; i++) {
4476 					ASSERT(PAGE_SHARED(ppa[i]));
4477 					page_unlock(ppa[i]);
4478 				}
4479 			}
4480 			if (amp != NULL) {
4481 				anon_array_exit(&an_cookie);
4482 				ANON_LOCK_EXIT(&amp->a_rwlock);
4483 			}
4484 
4485 		next:
4486 			if (vpage != NULL) {
4487 				vpage += pages;
4488 			}
4489 			adjszc_chk = 1;
4490 		}
4491 		if (a == lpgeaddr)
4492 			break;
4493 		ASSERT(a < lpgeaddr);
4494 
4495 		ASSERT(!brkcow && !tron && type != F_SOFTLOCK);
4496 
4497 		/*
4498 		 * ierr == -1 means we failed to map with a large page.
4499 		 * (either due to allocation/relocation failures or
4500 		 * misalignment with other mappings to this file.
4501 		 *
4502 		 * ierr == -2 means some other thread allocated a large page
4503 		 * after we gave up tp map with a large page.  retry with
4504 		 * larger mapping.
4505 		 */
4506 		ASSERT(ierr == -1 || ierr == -2);
4507 		ASSERT(ierr == -2 || szc != 0);
4508 		ASSERT(ierr == -1 || szc < seg->s_szc);
4509 		if (ierr == -2) {
4510 			SEGVN_VMSTAT_FLTVNPAGES(41);
4511 			ASSERT(pszc > szc && pszc <= seg->s_szc);
4512 			szc = pszc;
4513 		} else if (segvn_anypgsz_vnode) {
4514 			SEGVN_VMSTAT_FLTVNPAGES(42);
4515 			szc--;
4516 		} else {
4517 			SEGVN_VMSTAT_FLTVNPAGES(43);
4518 			ASSERT(pszc < szc);
4519 			/*
4520 			 * other process created pszc large page.
4521 			 * but we still have to drop to 0 szc.
4522 			 */
4523 			szc = 0;
4524 		}
4525 
4526 		pgsz = page_get_pagesize(szc);
4527 		pages = btop(pgsz);
4528 		if (ierr == -2) {
4529 			/*
4530 			 * Size up case. Note lpgaddr may only be needed for
4531 			 * softlock case so we don't adjust it here.
4532 			 */
4533 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4534 			ASSERT(a >= lpgaddr);
4535 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4536 			off = svd->offset + (uintptr_t)(a - seg->s_base);
4537 			aindx = svd->anon_index + seg_page(seg, a);
4538 			vpage = (svd->vpage != NULL) ?
4539 			    &svd->vpage[seg_page(seg, a)] : NULL;
4540 		} else {
4541 			/*
4542 			 * Size down case. Note lpgaddr may only be needed for
4543 			 * softlock case so we don't adjust it here.
4544 			 */
4545 			ASSERT(IS_P2ALIGNED(a, pgsz));
4546 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4547 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4548 			ASSERT(a < lpgeaddr);
4549 			if (a < addr) {
4550 				SEGVN_VMSTAT_FLTVNPAGES(44);
4551 				/*
4552 				 * The beginning of the large page region can
4553 				 * be pulled to the right to make a smaller
4554 				 * region. We haven't yet faulted a single
4555 				 * page.
4556 				 */
4557 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4558 				ASSERT(a >= lpgaddr);
4559 				off = svd->offset +
4560 				    (uintptr_t)(a - seg->s_base);
4561 				aindx = svd->anon_index + seg_page(seg, a);
4562 				vpage = (svd->vpage != NULL) ?
4563 				    &svd->vpage[seg_page(seg, a)] : NULL;
4564 			}
4565 		}
4566 	}
4567 out:
4568 	kmem_free(ppa, ppasize);
4569 	if (!err && !vop_size_err) {
4570 		SEGVN_VMSTAT_FLTVNPAGES(45);
4571 		return (0);
4572 	}
4573 	if (type == F_SOFTLOCK && a > lpgaddr) {
4574 		SEGVN_VMSTAT_FLTVNPAGES(46);
4575 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4576 	}
4577 	if (!vop_size_err) {
4578 		SEGVN_VMSTAT_FLTVNPAGES(47);
4579 		return (err);
4580 	}
4581 	ASSERT(brkcow || tron || type == F_SOFTLOCK);
4582 	/*
4583 	 * Large page end is mapped beyond the end of file and it's a cow
4584 	 * fault (can be a text replication induced cow) or softlock so we can't
4585 	 * reduce the map area.  For now just demote the segment. This should
4586 	 * really only happen if the end of the file changed after the mapping
4587 	 * was established since when large page segments are created we make
4588 	 * sure they don't extend beyond the end of the file.
4589 	 */
4590 	SEGVN_VMSTAT_FLTVNPAGES(48);
4591 
4592 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4593 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4594 	err = 0;
4595 	if (seg->s_szc != 0) {
4596 		segvn_fltvnpages_clrszc_cnt++;
4597 		ASSERT(svd->softlockcnt == 0);
4598 		err = segvn_clrszc(seg);
4599 		if (err != 0) {
4600 			segvn_fltvnpages_clrszc_err++;
4601 		}
4602 	}
4603 	ASSERT(err || seg->s_szc == 0);
4604 	SEGVN_LOCK_DOWNGRADE(seg->s_as, &svd->lock);
4605 	/* segvn_fault will do its job as if szc had been zero to begin with */
4606 	return (err == 0 ? IE_RETRY : FC_MAKE_ERR(err));
4607 }
4608 
4609 /*
4610  * This routine will attempt to fault in one large page.
4611  * it will use smaller pages if that fails.
4612  * It should only be called for pure anonymous segments.
4613  */
4614 static faultcode_t
4615 segvn_fault_anonpages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
4616     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
4617     caddr_t eaddr, int brkcow)
4618 {
4619 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4620 	struct anon_map *amp = svd->amp;
4621 	uchar_t segtype = svd->type;
4622 	uint_t szc = seg->s_szc;
4623 	size_t pgsz = page_get_pagesize(szc);
4624 	size_t maxpgsz = pgsz;
4625 	pgcnt_t pages = btop(pgsz);
4626 	uint_t ppaszc = szc;
4627 	caddr_t a = lpgaddr;
4628 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
4629 	struct vpage *vpage = (svd->vpage != NULL) ?
4630 	    &svd->vpage[seg_page(seg, a)] : NULL;
4631 	page_t **ppa;
4632 	uint_t	ppa_szc;
4633 	faultcode_t err;
4634 	int ierr;
4635 	uint_t protchk, prot, vpprot;
4636 	ulong_t i;
4637 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
4638 	anon_sync_obj_t cookie;
4639 	int adjszc_chk;
4640 	int pgflags = (svd->tr_state == SEGVN_TR_ON) ? PG_LOCAL : 0;
4641 
4642 	ASSERT(szc != 0);
4643 	ASSERT(amp != NULL);
4644 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
4645 	ASSERT(!(svd->flags & MAP_NORESERVE));
4646 	ASSERT(type != F_SOFTUNLOCK);
4647 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
4648 	ASSERT(!brkcow || svd->tr_state == SEGVN_TR_OFF);
4649 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
4650 
4651 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
4652 
4653 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltanpages[0]);
4654 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltanpages[1]);
4655 
4656 	if (svd->flags & MAP_TEXT) {
4657 		hat_flag |= HAT_LOAD_TEXT;
4658 	}
4659 
4660 	if (svd->pageprot) {
4661 		switch (rw) {
4662 		case S_READ:
4663 			protchk = PROT_READ;
4664 			break;
4665 		case S_WRITE:
4666 			protchk = PROT_WRITE;
4667 			break;
4668 		case S_EXEC:
4669 			protchk = PROT_EXEC;
4670 			break;
4671 		case S_OTHER:
4672 		default:
4673 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4674 			break;
4675 		}
4676 		VM_STAT_ADD(segvnvmstats.fltanpages[2]);
4677 	} else {
4678 		prot = svd->prot;
4679 		/* caller has already done segment level protection check. */
4680 	}
4681 
4682 	ppa = kmem_cache_alloc(segvn_szc_cache[ppaszc], KM_SLEEP);
4683 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4684 	for (;;) {
4685 		adjszc_chk = 0;
4686 		for (; a < lpgeaddr; a += pgsz, aindx += pages) {
4687 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
4688 				VM_STAT_ADD(segvnvmstats.fltanpages[3]);
4689 				ASSERT(vpage != NULL);
4690 				prot = VPP_PROT(vpage);
4691 				ASSERT(sameprot(seg, a, maxpgsz));
4692 				if ((prot & protchk) == 0) {
4693 					err = FC_PROT;
4694 					goto error;
4695 				}
4696 			}
4697 			if (adjszc_chk && IS_P2ALIGNED(a, maxpgsz) &&
4698 			    pgsz < maxpgsz) {
4699 				ASSERT(a > lpgaddr);
4700 				szc = seg->s_szc;
4701 				pgsz = maxpgsz;
4702 				pages = btop(pgsz);
4703 				ASSERT(IS_P2ALIGNED(aindx, pages));
4704 				lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr,
4705 				    pgsz);
4706 			}
4707 			if (type == F_SOFTLOCK) {
4708 				atomic_add_long((ulong_t *)&svd->softlockcnt,
4709 				    pages);
4710 			}
4711 			anon_array_enter(amp, aindx, &cookie);
4712 			ppa_szc = (uint_t)-1;
4713 			ierr = anon_map_getpages(amp, aindx, szc, seg, a,
4714 			    prot, &vpprot, ppa, &ppa_szc, vpage, rw, brkcow,
4715 			    segvn_anypgsz, pgflags, svd->cred);
4716 			if (ierr != 0) {
4717 				anon_array_exit(&cookie);
4718 				VM_STAT_ADD(segvnvmstats.fltanpages[4]);
4719 				if (type == F_SOFTLOCK) {
4720 					atomic_add_long(
4721 					    (ulong_t *)&svd->softlockcnt,
4722 					    -pages);
4723 				}
4724 				if (ierr > 0) {
4725 					VM_STAT_ADD(segvnvmstats.fltanpages[6]);
4726 					err = FC_MAKE_ERR(ierr);
4727 					goto error;
4728 				}
4729 				break;
4730 			}
4731 
4732 			ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4733 
4734 			ASSERT(segtype == MAP_SHARED ||
4735 			    ppa[0]->p_szc <= szc);
4736 			ASSERT(segtype == MAP_PRIVATE ||
4737 			    ppa[0]->p_szc >= szc);
4738 
4739 			/*
4740 			 * Handle pages that have been marked for migration
4741 			 */
4742 			if (lgrp_optimizations())
4743 				page_migrate(seg, a, ppa, pages);
4744 
4745 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
4746 
4747 			if (segtype == MAP_SHARED) {
4748 				vpprot |= PROT_WRITE;
4749 			}
4750 
4751 			hat_memload_array(hat, a, pgsz, ppa,
4752 			    prot & vpprot, hat_flag);
4753 
4754 			if (hat_flag & HAT_LOAD_LOCK) {
4755 				VM_STAT_ADD(segvnvmstats.fltanpages[7]);
4756 			} else {
4757 				VM_STAT_ADD(segvnvmstats.fltanpages[8]);
4758 				for (i = 0; i < pages; i++)
4759 					page_unlock(ppa[i]);
4760 			}
4761 			if (vpage != NULL)
4762 				vpage += pages;
4763 
4764 			anon_array_exit(&cookie);
4765 			adjszc_chk = 1;
4766 		}
4767 		if (a == lpgeaddr)
4768 			break;
4769 		ASSERT(a < lpgeaddr);
4770 		/*
4771 		 * ierr == -1 means we failed to allocate a large page.
4772 		 * so do a size down operation.
4773 		 *
4774 		 * ierr == -2 means some other process that privately shares
4775 		 * pages with this process has allocated a larger page and we
4776 		 * need to retry with larger pages. So do a size up
4777 		 * operation. This relies on the fact that large pages are
4778 		 * never partially shared i.e. if we share any constituent
4779 		 * page of a large page with another process we must share the
4780 		 * entire large page. Note this cannot happen for SOFTLOCK
4781 		 * case, unless current address (a) is at the beginning of the
4782 		 * next page size boundary because the other process couldn't
4783 		 * have relocated locked pages.
4784 		 */
4785 		ASSERT(ierr == -1 || ierr == -2);
4786 
4787 		if (segvn_anypgsz) {
4788 			ASSERT(ierr == -2 || szc != 0);
4789 			ASSERT(ierr == -1 || szc < seg->s_szc);
4790 			szc = (ierr == -1) ? szc - 1 : szc + 1;
4791 		} else {
4792 			/*
4793 			 * For non COW faults and segvn_anypgsz == 0
4794 			 * we need to be careful not to loop forever
4795 			 * if existing page is found with szc other
4796 			 * than 0 or seg->s_szc. This could be due
4797 			 * to page relocations on behalf of DR or
4798 			 * more likely large page creation. For this
4799 			 * case simply re-size to existing page's szc
4800 			 * if returned by anon_map_getpages().
4801 			 */
4802 			if (ppa_szc == (uint_t)-1) {
4803 				szc = (ierr == -1) ? 0 : seg->s_szc;
4804 			} else {
4805 				ASSERT(ppa_szc <= seg->s_szc);
4806 				ASSERT(ierr == -2 || ppa_szc < szc);
4807 				ASSERT(ierr == -1 || ppa_szc > szc);
4808 				szc = ppa_szc;
4809 			}
4810 		}
4811 
4812 		pgsz = page_get_pagesize(szc);
4813 		pages = btop(pgsz);
4814 		ASSERT(type != F_SOFTLOCK || ierr == -1 ||
4815 		    (IS_P2ALIGNED(a, pgsz) && IS_P2ALIGNED(lpgeaddr, pgsz)));
4816 		if (type == F_SOFTLOCK) {
4817 			/*
4818 			 * For softlocks we cannot reduce the fault area
4819 			 * (calculated based on the largest page size for this
4820 			 * segment) for size down and a is already next
4821 			 * page size aligned as assertted above for size
4822 			 * ups. Therefore just continue in case of softlock.
4823 			 */
4824 			VM_STAT_ADD(segvnvmstats.fltanpages[9]);
4825 			continue; /* keep lint happy */
4826 		} else if (ierr == -2) {
4827 
4828 			/*
4829 			 * Size up case. Note lpgaddr may only be needed for
4830 			 * softlock case so we don't adjust it here.
4831 			 */
4832 			VM_STAT_ADD(segvnvmstats.fltanpages[10]);
4833 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4834 			ASSERT(a >= lpgaddr);
4835 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4836 			aindx = svd->anon_index + seg_page(seg, a);
4837 			vpage = (svd->vpage != NULL) ?
4838 			    &svd->vpage[seg_page(seg, a)] : NULL;
4839 		} else {
4840 			/*
4841 			 * Size down case. Note lpgaddr may only be needed for
4842 			 * softlock case so we don't adjust it here.
4843 			 */
4844 			VM_STAT_ADD(segvnvmstats.fltanpages[11]);
4845 			ASSERT(IS_P2ALIGNED(a, pgsz));
4846 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4847 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4848 			ASSERT(a < lpgeaddr);
4849 			if (a < addr) {
4850 				/*
4851 				 * The beginning of the large page region can
4852 				 * be pulled to the right to make a smaller
4853 				 * region. We haven't yet faulted a single
4854 				 * page.
4855 				 */
4856 				VM_STAT_ADD(segvnvmstats.fltanpages[12]);
4857 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4858 				ASSERT(a >= lpgaddr);
4859 				aindx = svd->anon_index + seg_page(seg, a);
4860 				vpage = (svd->vpage != NULL) ?
4861 				    &svd->vpage[seg_page(seg, a)] : NULL;
4862 			}
4863 		}
4864 	}
4865 	VM_STAT_ADD(segvnvmstats.fltanpages[13]);
4866 	ANON_LOCK_EXIT(&amp->a_rwlock);
4867 	kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4868 	return (0);
4869 error:
4870 	VM_STAT_ADD(segvnvmstats.fltanpages[14]);
4871 	ANON_LOCK_EXIT(&amp->a_rwlock);
4872 	kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4873 	if (type == F_SOFTLOCK && a > lpgaddr) {
4874 		VM_STAT_ADD(segvnvmstats.fltanpages[15]);
4875 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4876 	}
4877 	return (err);
4878 }
4879 
4880 int fltadvice = 1;	/* set to free behind pages for sequential access */
4881 
4882 /*
4883  * This routine is called via a machine specific fault handling routine.
4884  * It is also called by software routines wishing to lock or unlock
4885  * a range of addresses.
4886  *
4887  * Here is the basic algorithm:
4888  *	If unlocking
4889  *		Call segvn_softunlock
4890  *		Return
4891  *	endif
4892  *	Checking and set up work
4893  *	If we will need some non-anonymous pages
4894  *		Call VOP_GETPAGE over the range of non-anonymous pages
4895  *	endif
4896  *	Loop over all addresses requested
4897  *		Call segvn_faultpage passing in page list
4898  *		    to load up translations and handle anonymous pages
4899  *	endloop
4900  *	Load up translation to any additional pages in page list not
4901  *	    already handled that fit into this segment
4902  */
4903 static faultcode_t
4904 segvn_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
4905     enum fault_type type, enum seg_rw rw)
4906 {
4907 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4908 	page_t **plp, **ppp, *pp;
4909 	u_offset_t off;
4910 	caddr_t a;
4911 	struct vpage *vpage;
4912 	uint_t vpprot, prot;
4913 	int err;
4914 	page_t *pl[PVN_GETPAGE_NUM + 1];
4915 	size_t plsz, pl_alloc_sz;
4916 	size_t page;
4917 	ulong_t anon_index;
4918 	struct anon_map *amp;
4919 	int dogetpage = 0;
4920 	caddr_t	lpgaddr, lpgeaddr;
4921 	size_t pgsz;
4922 	anon_sync_obj_t cookie;
4923 	int brkcow = BREAK_COW_SHARE(rw, type, svd->type);
4924 
4925 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
4926 	ASSERT(svd->amp == NULL || svd->rcookie == HAT_INVALID_REGION_COOKIE);
4927 
4928 	/*
4929 	 * First handle the easy stuff
4930 	 */
4931 	if (type == F_SOFTUNLOCK) {
4932 		if (rw == S_READ_NOCOW) {
4933 			rw = S_READ;
4934 			ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
4935 		}
4936 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4937 		pgsz = (seg->s_szc == 0) ? PAGESIZE :
4938 		    page_get_pagesize(seg->s_szc);
4939 		VM_STAT_COND_ADD(pgsz > PAGESIZE, segvnvmstats.fltanpages[16]);
4940 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
4941 		segvn_softunlock(seg, lpgaddr, lpgeaddr - lpgaddr, rw);
4942 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4943 		return (0);
4944 	}
4945 
4946 	ASSERT(svd->tr_state == SEGVN_TR_OFF ||
4947 	    !HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
4948 	if (brkcow == 0) {
4949 		if (svd->tr_state == SEGVN_TR_INIT) {
4950 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4951 			if (svd->tr_state == SEGVN_TR_INIT) {
4952 				ASSERT(svd->vp != NULL && svd->amp == NULL);
4953 				ASSERT(svd->flags & MAP_TEXT);
4954 				ASSERT(svd->type == MAP_PRIVATE);
4955 				segvn_textrepl(seg);
4956 				ASSERT(svd->tr_state != SEGVN_TR_INIT);
4957 				ASSERT(svd->tr_state != SEGVN_TR_ON ||
4958 				    svd->amp != NULL);
4959 			}
4960 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4961 		}
4962 	} else if (svd->tr_state != SEGVN_TR_OFF) {
4963 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4964 
4965 		if (rw == S_WRITE && svd->tr_state != SEGVN_TR_OFF) {
4966 			ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
4967 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4968 			return (FC_PROT);
4969 		}
4970 
4971 		if (svd->tr_state == SEGVN_TR_ON) {
4972 			ASSERT(svd->vp != NULL && svd->amp != NULL);
4973 			segvn_textunrepl(seg, 0);
4974 			ASSERT(svd->amp == NULL &&
4975 			    svd->tr_state == SEGVN_TR_OFF);
4976 		} else if (svd->tr_state != SEGVN_TR_OFF) {
4977 			svd->tr_state = SEGVN_TR_OFF;
4978 		}
4979 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
4980 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4981 	}
4982 
4983 top:
4984 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4985 
4986 	/*
4987 	 * If we have the same protections for the entire segment,
4988 	 * insure that the access being attempted is legitimate.
4989 	 */
4990 
4991 	if (svd->pageprot == 0) {
4992 		uint_t protchk;
4993 
4994 		switch (rw) {
4995 		case S_READ:
4996 		case S_READ_NOCOW:
4997 			protchk = PROT_READ;
4998 			break;
4999 		case S_WRITE:
5000 			protchk = PROT_WRITE;
5001 			break;
5002 		case S_EXEC:
5003 			protchk = PROT_EXEC;
5004 			break;
5005 		case S_OTHER:
5006 		default:
5007 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
5008 			break;
5009 		}
5010 
5011 		if ((svd->prot & protchk) == 0) {
5012 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5013 			return (FC_PROT);	/* illegal access type */
5014 		}
5015 	}
5016 
5017 	if (brkcow && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5018 		/* this must be SOFTLOCK S_READ fault */
5019 		ASSERT(svd->amp == NULL);
5020 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5021 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5022 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5023 		if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5024 			/*
5025 			 * this must be the first ever non S_READ_NOCOW
5026 			 * softlock for this segment.
5027 			 */
5028 			ASSERT(svd->softlockcnt == 0);
5029 			hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5030 			    HAT_REGION_TEXT);
5031 			svd->rcookie = HAT_INVALID_REGION_COOKIE;
5032 		}
5033 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5034 		goto top;
5035 	}
5036 
5037 	/*
5038 	 * We can't allow the long term use of softlocks for vmpss segments,
5039 	 * because in some file truncation cases we should be able to demote
5040 	 * the segment, which requires that there are no softlocks.  The
5041 	 * only case where it's ok to allow a SOFTLOCK fault against a vmpss
5042 	 * segment is S_READ_NOCOW, where the caller holds the address space
5043 	 * locked as writer and calls softunlock before dropping the as lock.
5044 	 * S_READ_NOCOW is used by /proc to read memory from another user.
5045 	 *
5046 	 * Another deadlock between SOFTLOCK and file truncation can happen
5047 	 * because segvn_fault_vnodepages() calls the FS one pagesize at
5048 	 * a time. A second VOP_GETPAGE() call by segvn_fault_vnodepages()
5049 	 * can cause a deadlock because the first set of page_t's remain
5050 	 * locked SE_SHARED.  To avoid this, we demote segments on a first
5051 	 * SOFTLOCK if they have a length greater than the segment's
5052 	 * page size.
5053 	 *
5054 	 * So for now, we only avoid demoting a segment on a SOFTLOCK when
5055 	 * the access type is S_READ_NOCOW and the fault length is less than
5056 	 * or equal to the segment's page size. While this is quite restrictive,
5057 	 * it should be the most common case of SOFTLOCK against a vmpss
5058 	 * segment.
5059 	 *
5060 	 * For S_READ_NOCOW, it's safe not to do a copy on write because the
5061 	 * caller makes sure no COW will be caused by another thread for a
5062 	 * softlocked page.
5063 	 */
5064 	if (type == F_SOFTLOCK && svd->vp != NULL && seg->s_szc != 0) {
5065 		int demote = 0;
5066 
5067 		if (rw != S_READ_NOCOW) {
5068 			demote = 1;
5069 		}
5070 		if (!demote && len > PAGESIZE) {
5071 			pgsz = page_get_pagesize(seg->s_szc);
5072 			CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr,
5073 			    lpgeaddr);
5074 			if (lpgeaddr - lpgaddr > pgsz) {
5075 				demote = 1;
5076 			}
5077 		}
5078 
5079 		ASSERT(demote || AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5080 
5081 		if (demote) {
5082 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5083 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5084 			if (seg->s_szc != 0) {
5085 				segvn_vmpss_clrszc_cnt++;
5086 				ASSERT(svd->softlockcnt == 0);
5087 				err = segvn_clrszc(seg);
5088 				if (err) {
5089 					segvn_vmpss_clrszc_err++;
5090 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5091 					return (FC_MAKE_ERR(err));
5092 				}
5093 			}
5094 			ASSERT(seg->s_szc == 0);
5095 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5096 			goto top;
5097 		}
5098 	}
5099 
5100 	/*
5101 	 * Check to see if we need to allocate an anon_map structure.
5102 	 */
5103 	if (svd->amp == NULL && (svd->vp == NULL || brkcow)) {
5104 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5105 		/*
5106 		 * Drop the "read" lock on the segment and acquire
5107 		 * the "write" version since we have to allocate the
5108 		 * anon_map.
5109 		 */
5110 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5111 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5112 
5113 		if (svd->amp == NULL) {
5114 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
5115 			svd->amp->a_szc = seg->s_szc;
5116 		}
5117 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5118 
5119 		/*
5120 		 * Start all over again since segment protections
5121 		 * may have changed after we dropped the "read" lock.
5122 		 */
5123 		goto top;
5124 	}
5125 
5126 	/*
5127 	 * S_READ_NOCOW vs S_READ distinction was
5128 	 * only needed for the code above. After
5129 	 * that we treat it as S_READ.
5130 	 */
5131 	if (rw == S_READ_NOCOW) {
5132 		ASSERT(type == F_SOFTLOCK);
5133 		ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5134 		rw = S_READ;
5135 	}
5136 
5137 	amp = svd->amp;
5138 
5139 	/*
5140 	 * MADV_SEQUENTIAL work is ignored for large page segments.
5141 	 */
5142 	if (seg->s_szc != 0) {
5143 		pgsz = page_get_pagesize(seg->s_szc);
5144 		ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
5145 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
5146 		if (svd->vp == NULL) {
5147 			err = segvn_fault_anonpages(hat, seg, lpgaddr,
5148 			    lpgeaddr, type, rw, addr, addr + len, brkcow);
5149 		} else {
5150 			err = segvn_fault_vnodepages(hat, seg, lpgaddr,
5151 			    lpgeaddr, type, rw, addr, addr + len, brkcow);
5152 			if (err == IE_RETRY) {
5153 				ASSERT(seg->s_szc == 0);
5154 				ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
5155 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5156 				goto top;
5157 			}
5158 		}
5159 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5160 		return (err);
5161 	}
5162 
5163 	page = seg_page(seg, addr);
5164 	if (amp != NULL) {
5165 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5166 		anon_index = svd->anon_index + page;
5167 
5168 		if (type == F_PROT && rw == S_READ &&
5169 		    svd->tr_state == SEGVN_TR_OFF &&
5170 		    svd->type == MAP_PRIVATE && svd->pageprot == 0) {
5171 			size_t index = anon_index;
5172 			struct anon *ap;
5173 
5174 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5175 			/*
5176 			 * The fast path could apply to S_WRITE also, except
5177 			 * that the protection fault could be caused by lazy
5178 			 * tlb flush when ro->rw. In this case, the pte is
5179 			 * RW already. But RO in the other cpu's tlb causes
5180 			 * the fault. Since hat_chgprot won't do anything if
5181 			 * pte doesn't change, we may end up faulting
5182 			 * indefinitely until the RO tlb entry gets replaced.
5183 			 */
5184 			for (a = addr; a < addr + len; a += PAGESIZE, index++) {
5185 				anon_array_enter(amp, index, &cookie);
5186 				ap = anon_get_ptr(amp->ahp, index);
5187 				anon_array_exit(&cookie);
5188 				if ((ap == NULL) || (ap->an_refcnt != 1)) {
5189 					ANON_LOCK_EXIT(&amp->a_rwlock);
5190 					goto slow;
5191 				}
5192 			}
5193 			hat_chgprot(seg->s_as->a_hat, addr, len, svd->prot);
5194 			ANON_LOCK_EXIT(&amp->a_rwlock);
5195 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5196 			return (0);
5197 		}
5198 	}
5199 slow:
5200 
5201 	if (svd->vpage == NULL)
5202 		vpage = NULL;
5203 	else
5204 		vpage = &svd->vpage[page];
5205 
5206 	off = svd->offset + (uintptr_t)(addr - seg->s_base);
5207 
5208 	/*
5209 	 * If MADV_SEQUENTIAL has been set for the particular page we
5210 	 * are faulting on, free behind all pages in the segment and put
5211 	 * them on the free list.
5212 	 */
5213 
5214 	if ((page != 0) && fltadvice && svd->tr_state != SEGVN_TR_ON) {
5215 		struct vpage *vpp;
5216 		ulong_t fanon_index;
5217 		size_t fpage;
5218 		u_offset_t pgoff, fpgoff;
5219 		struct vnode *fvp;
5220 		struct anon *fap = NULL;
5221 
5222 		if (svd->advice == MADV_SEQUENTIAL ||
5223 		    (svd->pageadvice &&
5224 		    VPP_ADVICE(vpage) == MADV_SEQUENTIAL)) {
5225 			pgoff = off - PAGESIZE;
5226 			fpage = page - 1;
5227 			if (vpage != NULL)
5228 				vpp = &svd->vpage[fpage];
5229 			if (amp != NULL)
5230 				fanon_index = svd->anon_index + fpage;
5231 
5232 			while (pgoff > svd->offset) {
5233 				if (svd->advice != MADV_SEQUENTIAL &&
5234 				    (!svd->pageadvice || (vpage &&
5235 				    VPP_ADVICE(vpp) != MADV_SEQUENTIAL)))
5236 					break;
5237 
5238 				/*
5239 				 * If this is an anon page, we must find the
5240 				 * correct <vp, offset> for it
5241 				 */
5242 				fap = NULL;
5243 				if (amp != NULL) {
5244 					ANON_LOCK_ENTER(&amp->a_rwlock,
5245 					    RW_READER);
5246 					anon_array_enter(amp, fanon_index,
5247 					    &cookie);
5248 					fap = anon_get_ptr(amp->ahp,
5249 					    fanon_index);
5250 					if (fap != NULL) {
5251 						swap_xlate(fap, &fvp, &fpgoff);
5252 					} else {
5253 						fpgoff = pgoff;
5254 						fvp = svd->vp;
5255 					}
5256 					anon_array_exit(&cookie);
5257 					ANON_LOCK_EXIT(&amp->a_rwlock);
5258 				} else {
5259 					fpgoff = pgoff;
5260 					fvp = svd->vp;
5261 				}
5262 				if (fvp == NULL)
5263 					break;	/* XXX */
5264 				/*
5265 				 * Skip pages that are free or have an
5266 				 * "exclusive" lock.
5267 				 */
5268 				pp = page_lookup_nowait(fvp, fpgoff, SE_SHARED);
5269 				if (pp == NULL)
5270 					break;
5271 				/*
5272 				 * We don't need the page_struct_lock to test
5273 				 * as this is only advisory; even if we
5274 				 * acquire it someone might race in and lock
5275 				 * the page after we unlock and before the
5276 				 * PUTPAGE, then VOP_PUTPAGE will do nothing.
5277 				 */
5278 				if (pp->p_lckcnt == 0 && pp->p_cowcnt == 0) {
5279 					/*
5280 					 * Hold the vnode before releasing
5281 					 * the page lock to prevent it from
5282 					 * being freed and re-used by some
5283 					 * other thread.
5284 					 */
5285 					VN_HOLD(fvp);
5286 					page_unlock(pp);
5287 					/*
5288 					 * We should build a page list
5289 					 * to kluster putpages XXX
5290 					 */
5291 					(void) VOP_PUTPAGE(fvp,
5292 					    (offset_t)fpgoff, PAGESIZE,
5293 					    (B_DONTNEED|B_FREE|B_ASYNC),
5294 					    svd->cred, NULL);
5295 					VN_RELE(fvp);
5296 				} else {
5297 					/*
5298 					 * XXX - Should the loop terminate if
5299 					 * the page is `locked'?
5300 					 */
5301 					page_unlock(pp);
5302 				}
5303 				--vpp;
5304 				--fanon_index;
5305 				pgoff -= PAGESIZE;
5306 			}
5307 		}
5308 	}
5309 
5310 	plp = pl;
5311 	*plp = NULL;
5312 	pl_alloc_sz = 0;
5313 
5314 	/*
5315 	 * See if we need to call VOP_GETPAGE for
5316 	 * *any* of the range being faulted on.
5317 	 * We can skip all of this work if there
5318 	 * was no original vnode.
5319 	 */
5320 	if (svd->vp != NULL) {
5321 		u_offset_t vp_off;
5322 		size_t vp_len;
5323 		struct anon *ap;
5324 		vnode_t *vp;
5325 
5326 		vp_off = off;
5327 		vp_len = len;
5328 
5329 		if (amp == NULL)
5330 			dogetpage = 1;
5331 		else {
5332 			/*
5333 			 * Only acquire reader lock to prevent amp->ahp
5334 			 * from being changed.  It's ok to miss pages,
5335 			 * hence we don't do anon_array_enter
5336 			 */
5337 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5338 			ap = anon_get_ptr(amp->ahp, anon_index);
5339 
5340 			if (len <= PAGESIZE)
5341 				/* inline non_anon() */
5342 				dogetpage = (ap == NULL);
5343 			else
5344 				dogetpage = non_anon(amp->ahp, anon_index,
5345 				    &vp_off, &vp_len);
5346 			ANON_LOCK_EXIT(&amp->a_rwlock);
5347 		}
5348 
5349 		if (dogetpage) {
5350 			enum seg_rw arw;
5351 			struct as *as = seg->s_as;
5352 
5353 			if (len > ptob((sizeof (pl) / sizeof (pl[0])) - 1)) {
5354 				/*
5355 				 * Page list won't fit in local array,
5356 				 * allocate one of the needed size.
5357 				 */
5358 				pl_alloc_sz =
5359 				    (btop(len) + 1) * sizeof (page_t *);
5360 				plp = kmem_alloc(pl_alloc_sz, KM_SLEEP);
5361 				plp[0] = NULL;
5362 				plsz = len;
5363 			} else if (rw == S_WRITE && svd->type == MAP_PRIVATE ||
5364 			    svd->tr_state == SEGVN_TR_ON || rw == S_OTHER ||
5365 			    (((size_t)(addr + PAGESIZE) <
5366 			    (size_t)(seg->s_base + seg->s_size)) &&
5367 			    hat_probe(as->a_hat, addr + PAGESIZE))) {
5368 				/*
5369 				 * Ask VOP_GETPAGE to return the exact number
5370 				 * of pages if
5371 				 * (a) this is a COW fault, or
5372 				 * (b) this is a software fault, or
5373 				 * (c) next page is already mapped.
5374 				 */
5375 				plsz = len;
5376 			} else {
5377 				/*
5378 				 * Ask VOP_GETPAGE to return adjacent pages
5379 				 * within the segment.
5380 				 */
5381 				plsz = MIN((size_t)PVN_GETPAGE_SZ, (size_t)
5382 				    ((seg->s_base + seg->s_size) - addr));
5383 				ASSERT((addr + plsz) <=
5384 				    (seg->s_base + seg->s_size));
5385 			}
5386 
5387 			/*
5388 			 * Need to get some non-anonymous pages.
5389 			 * We need to make only one call to GETPAGE to do
5390 			 * this to prevent certain deadlocking conditions
5391 			 * when we are doing locking.  In this case
5392 			 * non_anon() should have picked up the smallest
5393 			 * range which includes all the non-anonymous
5394 			 * pages in the requested range.  We have to
5395 			 * be careful regarding which rw flag to pass in
5396 			 * because on a private mapping, the underlying
5397 			 * object is never allowed to be written.
5398 			 */
5399 			if (rw == S_WRITE && svd->type == MAP_PRIVATE) {
5400 				arw = S_READ;
5401 			} else {
5402 				arw = rw;
5403 			}
5404 			vp = svd->vp;
5405 			TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5406 			    "segvn_getpage:seg %p addr %p vp %p",
5407 			    seg, addr, vp);
5408 			err = VOP_GETPAGE(vp, (offset_t)vp_off, vp_len,
5409 			    &vpprot, plp, plsz, seg, addr + (vp_off - off), arw,
5410 			    svd->cred, NULL);
5411 			if (err) {
5412 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5413 				segvn_pagelist_rele(plp);
5414 				if (pl_alloc_sz)
5415 					kmem_free(plp, pl_alloc_sz);
5416 				return (FC_MAKE_ERR(err));
5417 			}
5418 			if (svd->type == MAP_PRIVATE)
5419 				vpprot &= ~PROT_WRITE;
5420 		}
5421 	}
5422 
5423 	/*
5424 	 * N.B. at this time the plp array has all the needed non-anon
5425 	 * pages in addition to (possibly) having some adjacent pages.
5426 	 */
5427 
5428 	/*
5429 	 * Always acquire the anon_array_lock to prevent
5430 	 * 2 threads from allocating separate anon slots for
5431 	 * the same "addr".
5432 	 *
5433 	 * If this is a copy-on-write fault and we don't already
5434 	 * have the anon_array_lock, acquire it to prevent the
5435 	 * fault routine from handling multiple copy-on-write faults
5436 	 * on the same "addr" in the same address space.
5437 	 *
5438 	 * Only one thread should deal with the fault since after
5439 	 * it is handled, the other threads can acquire a translation
5440 	 * to the newly created private page.  This prevents two or
5441 	 * more threads from creating different private pages for the
5442 	 * same fault.
5443 	 *
5444 	 * We grab "serialization" lock here if this is a MAP_PRIVATE segment
5445 	 * to prevent deadlock between this thread and another thread
5446 	 * which has soft-locked this page and wants to acquire serial_lock.
5447 	 * ( bug 4026339 )
5448 	 *
5449 	 * The fix for bug 4026339 becomes unnecessary when using the
5450 	 * locking scheme with per amp rwlock and a global set of hash
5451 	 * lock, anon_array_lock.  If we steal a vnode page when low
5452 	 * on memory and upgrad the page lock through page_rename,
5453 	 * then the page is PAGE_HANDLED, nothing needs to be done
5454 	 * for this page after returning from segvn_faultpage.
5455 	 *
5456 	 * But really, the page lock should be downgraded after
5457 	 * the stolen page is page_rename'd.
5458 	 */
5459 
5460 	if (amp != NULL)
5461 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5462 
5463 	/*
5464 	 * Ok, now loop over the address range and handle faults
5465 	 */
5466 	for (a = addr; a < addr + len; a += PAGESIZE, off += PAGESIZE) {
5467 		err = segvn_faultpage(hat, seg, a, off, vpage, plp, vpprot,
5468 		    type, rw, brkcow);
5469 		if (err) {
5470 			if (amp != NULL)
5471 				ANON_LOCK_EXIT(&amp->a_rwlock);
5472 			if (type == F_SOFTLOCK && a > addr) {
5473 				segvn_softunlock(seg, addr, (a - addr),
5474 				    S_OTHER);
5475 			}
5476 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5477 			segvn_pagelist_rele(plp);
5478 			if (pl_alloc_sz)
5479 				kmem_free(plp, pl_alloc_sz);
5480 			return (err);
5481 		}
5482 		if (vpage) {
5483 			vpage++;
5484 		} else if (svd->vpage) {
5485 			page = seg_page(seg, addr);
5486 			vpage = &svd->vpage[++page];
5487 		}
5488 	}
5489 
5490 	/* Didn't get pages from the underlying fs so we're done */
5491 	if (!dogetpage)
5492 		goto done;
5493 
5494 	/*
5495 	 * Now handle any other pages in the list returned.
5496 	 * If the page can be used, load up the translations now.
5497 	 * Note that the for loop will only be entered if "plp"
5498 	 * is pointing to a non-NULL page pointer which means that
5499 	 * VOP_GETPAGE() was called and vpprot has been initialized.
5500 	 */
5501 	if (svd->pageprot == 0)
5502 		prot = svd->prot & vpprot;
5503 
5504 
5505 	/*
5506 	 * Large Files: diff should be unsigned value because we started
5507 	 * supporting > 2GB segment sizes from 2.5.1 and when a
5508 	 * large file of size > 2GB gets mapped to address space
5509 	 * the diff value can be > 2GB.
5510 	 */
5511 
5512 	for (ppp = plp; (pp = *ppp) != NULL; ppp++) {
5513 		size_t diff;
5514 		struct anon *ap;
5515 		int anon_index;
5516 		anon_sync_obj_t cookie;
5517 		int hat_flag = HAT_LOAD_ADV;
5518 
5519 		if (svd->flags & MAP_TEXT) {
5520 			hat_flag |= HAT_LOAD_TEXT;
5521 		}
5522 
5523 		if (pp == PAGE_HANDLED)
5524 			continue;
5525 
5526 		if (svd->tr_state != SEGVN_TR_ON &&
5527 		    pp->p_offset >=  svd->offset &&
5528 		    pp->p_offset < svd->offset + seg->s_size) {
5529 
5530 			diff = pp->p_offset - svd->offset;
5531 
5532 			/*
5533 			 * Large Files: Following is the assertion
5534 			 * validating the above cast.
5535 			 */
5536 			ASSERT(svd->vp == pp->p_vnode);
5537 
5538 			page = btop(diff);
5539 			if (svd->pageprot)
5540 				prot = VPP_PROT(&svd->vpage[page]) & vpprot;
5541 
5542 			/*
5543 			 * Prevent other threads in the address space from
5544 			 * creating private pages (i.e., allocating anon slots)
5545 			 * while we are in the process of loading translations
5546 			 * to additional pages returned by the underlying
5547 			 * object.
5548 			 */
5549 			if (amp != NULL) {
5550 				anon_index = svd->anon_index + page;
5551 				anon_array_enter(amp, anon_index, &cookie);
5552 				ap = anon_get_ptr(amp->ahp, anon_index);
5553 			}
5554 			if ((amp == NULL) || (ap == NULL)) {
5555 				if (IS_VMODSORT(pp->p_vnode) ||
5556 				    enable_mbit_wa) {
5557 					if (rw == S_WRITE)
5558 						hat_setmod(pp);
5559 					else if (rw != S_OTHER &&
5560 					    !hat_ismod(pp))
5561 						prot &= ~PROT_WRITE;
5562 				}
5563 				/*
5564 				 * Skip mapping read ahead pages marked
5565 				 * for migration, so they will get migrated
5566 				 * properly on fault
5567 				 */
5568 				ASSERT(amp == NULL ||
5569 				    svd->rcookie == HAT_INVALID_REGION_COOKIE);
5570 				if ((prot & PROT_READ) && !PP_ISMIGRATE(pp)) {
5571 					hat_memload_region(hat,
5572 					    seg->s_base + diff,
5573 					    pp, prot, hat_flag,
5574 					    svd->rcookie);
5575 				}
5576 			}
5577 			if (amp != NULL)
5578 				anon_array_exit(&cookie);
5579 		}
5580 		page_unlock(pp);
5581 	}
5582 done:
5583 	if (amp != NULL)
5584 		ANON_LOCK_EXIT(&amp->a_rwlock);
5585 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5586 	if (pl_alloc_sz)
5587 		kmem_free(plp, pl_alloc_sz);
5588 	return (0);
5589 }
5590 
5591 /*
5592  * This routine is used to start I/O on pages asynchronously.  XXX it will
5593  * only create PAGESIZE pages. At fault time they will be relocated into
5594  * larger pages.
5595  */
5596 static faultcode_t
5597 segvn_faulta(struct seg *seg, caddr_t addr)
5598 {
5599 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5600 	int err;
5601 	struct anon_map *amp;
5602 	vnode_t *vp;
5603 
5604 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5605 
5606 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5607 	if ((amp = svd->amp) != NULL) {
5608 		struct anon *ap;
5609 
5610 		/*
5611 		 * Reader lock to prevent amp->ahp from being changed.
5612 		 * This is advisory, it's ok to miss a page, so
5613 		 * we don't do anon_array_enter lock.
5614 		 */
5615 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5616 		if ((ap = anon_get_ptr(amp->ahp,
5617 		    svd->anon_index + seg_page(seg, addr))) != NULL) {
5618 
5619 			err = anon_getpage(&ap, NULL, NULL,
5620 			    0, seg, addr, S_READ, svd->cred);
5621 
5622 			ANON_LOCK_EXIT(&amp->a_rwlock);
5623 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5624 			if (err)
5625 				return (FC_MAKE_ERR(err));
5626 			return (0);
5627 		}
5628 		ANON_LOCK_EXIT(&amp->a_rwlock);
5629 	}
5630 
5631 	if (svd->vp == NULL) {
5632 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5633 		return (0);			/* zfod page - do nothing now */
5634 	}
5635 
5636 	vp = svd->vp;
5637 	TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5638 	    "segvn_getpage:seg %p addr %p vp %p", seg, addr, vp);
5639 	err = VOP_GETPAGE(vp,
5640 	    (offset_t)(svd->offset + (uintptr_t)(addr - seg->s_base)),
5641 	    PAGESIZE, NULL, NULL, 0, seg, addr,
5642 	    S_OTHER, svd->cred, NULL);
5643 
5644 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5645 	if (err)
5646 		return (FC_MAKE_ERR(err));
5647 	return (0);
5648 }
5649 
5650 static int
5651 segvn_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
5652 {
5653 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5654 	struct vpage *cvp, *svp, *evp;
5655 	struct vnode *vp;
5656 	size_t pgsz;
5657 	pgcnt_t pgcnt;
5658 	anon_sync_obj_t cookie;
5659 	int unload_done = 0;
5660 
5661 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5662 
5663 	if ((svd->maxprot & prot) != prot)
5664 		return (EACCES);			/* violated maxprot */
5665 
5666 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5667 
5668 	/* return if prot is the same */
5669 	if (!svd->pageprot && svd->prot == prot) {
5670 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5671 		return (0);
5672 	}
5673 
5674 	/*
5675 	 * Since we change protections we first have to flush the cache.
5676 	 * This makes sure all the pagelock calls have to recheck
5677 	 * protections.
5678 	 */
5679 	if (svd->softlockcnt > 0) {
5680 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5681 
5682 		/*
5683 		 * If this is shared segment non 0 softlockcnt
5684 		 * means locked pages are still in use.
5685 		 */
5686 		if (svd->type == MAP_SHARED) {
5687 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5688 			return (EAGAIN);
5689 		}
5690 
5691 		/*
5692 		 * Since we do have the segvn writers lock nobody can fill
5693 		 * the cache with entries belonging to this seg during
5694 		 * the purge. The flush either succeeds or we still have
5695 		 * pending I/Os.
5696 		 */
5697 		segvn_purge(seg);
5698 		if (svd->softlockcnt > 0) {
5699 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5700 			return (EAGAIN);
5701 		}
5702 	}
5703 
5704 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5705 		ASSERT(svd->amp == NULL);
5706 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5707 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5708 		    HAT_REGION_TEXT);
5709 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
5710 		unload_done = 1;
5711 	} else if (svd->tr_state == SEGVN_TR_INIT) {
5712 		svd->tr_state = SEGVN_TR_OFF;
5713 	} else if (svd->tr_state == SEGVN_TR_ON) {
5714 		ASSERT(svd->amp != NULL);
5715 		segvn_textunrepl(seg, 0);
5716 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
5717 		unload_done = 1;
5718 	}
5719 
5720 	if ((prot & PROT_WRITE) && svd->type == MAP_SHARED &&
5721 	    svd->vp != NULL && (svd->vp->v_flag & VVMEXEC)) {
5722 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
5723 		segvn_inval_trcache(svd->vp);
5724 	}
5725 	if (seg->s_szc != 0) {
5726 		int err;
5727 		pgsz = page_get_pagesize(seg->s_szc);
5728 		pgcnt = pgsz >> PAGESHIFT;
5729 		ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
5730 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
5731 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5732 			ASSERT(seg->s_base != addr || seg->s_size != len);
5733 			/*
5734 			 * If we are holding the as lock as a reader then
5735 			 * we need to return IE_RETRY and let the as
5736 			 * layer drop and re-acquire the lock as a writer.
5737 			 */
5738 			if (AS_READ_HELD(seg->s_as, &seg->s_as->a_lock))
5739 				return (IE_RETRY);
5740 			VM_STAT_ADD(segvnvmstats.demoterange[1]);
5741 			if (svd->type == MAP_PRIVATE || svd->vp != NULL) {
5742 				err = segvn_demote_range(seg, addr, len,
5743 				    SDR_END, 0);
5744 			} else {
5745 				uint_t szcvec = map_pgszcvec(seg->s_base,
5746 				    pgsz, (uintptr_t)seg->s_base,
5747 				    (svd->flags & MAP_TEXT), MAPPGSZC_SHM, 0);
5748 				err = segvn_demote_range(seg, addr, len,
5749 				    SDR_END, szcvec);
5750 			}
5751 			if (err == 0)
5752 				return (IE_RETRY);
5753 			if (err == ENOMEM)
5754 				return (IE_NOMEM);
5755 			return (err);
5756 		}
5757 	}
5758 
5759 
5760 	/*
5761 	 * If it's a private mapping and we're making it writable then we
5762 	 * may have to reserve the additional swap space now. If we are
5763 	 * making writable only a part of the segment then we use its vpage
5764 	 * array to keep a record of the pages for which we have reserved
5765 	 * swap. In this case we set the pageswap field in the segment's
5766 	 * segvn structure to record this.
5767 	 *
5768 	 * If it's a private mapping to a file (i.e., vp != NULL) and we're
5769 	 * removing write permission on the entire segment and we haven't
5770 	 * modified any pages, we can release the swap space.
5771 	 */
5772 	if (svd->type == MAP_PRIVATE) {
5773 		if (prot & PROT_WRITE) {
5774 			if (!(svd->flags & MAP_NORESERVE) &&
5775 			    !(svd->swresv && svd->pageswap == 0)) {
5776 				size_t sz = 0;
5777 
5778 				/*
5779 				 * Start by determining how much swap
5780 				 * space is required.
5781 				 */
5782 				if (addr == seg->s_base &&
5783 				    len == seg->s_size &&
5784 				    svd->pageswap == 0) {
5785 					/* The whole segment */
5786 					sz = seg->s_size;
5787 				} else {
5788 					/*
5789 					 * Make sure that the vpage array
5790 					 * exists, and make a note of the
5791 					 * range of elements corresponding
5792 					 * to len.
5793 					 */
5794 					segvn_vpage(seg);
5795 					if (svd->vpage == NULL) {
5796 						SEGVN_LOCK_EXIT(seg->s_as,
5797 						    &svd->lock);
5798 						return (ENOMEM);
5799 					}
5800 					svp = &svd->vpage[seg_page(seg, addr)];
5801 					evp = &svd->vpage[seg_page(seg,
5802 					    addr + len)];
5803 
5804 					if (svd->pageswap == 0) {
5805 						/*
5806 						 * This is the first time we've
5807 						 * asked for a part of this
5808 						 * segment, so we need to
5809 						 * reserve everything we've
5810 						 * been asked for.
5811 						 */
5812 						sz = len;
5813 					} else {
5814 						/*
5815 						 * We have to count the number
5816 						 * of pages required.
5817 						 */
5818 						for (cvp = svp;  cvp < evp;
5819 						    cvp++) {
5820 							if (!VPP_ISSWAPRES(cvp))
5821 								sz++;
5822 						}
5823 						sz <<= PAGESHIFT;
5824 					}
5825 				}
5826 
5827 				/* Try to reserve the necessary swap. */
5828 				if (anon_resv_zone(sz,
5829 				    seg->s_as->a_proc->p_zone) == 0) {
5830 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5831 					return (IE_NOMEM);
5832 				}
5833 
5834 				/*
5835 				 * Make a note of how much swap space
5836 				 * we've reserved.
5837 				 */
5838 				if (svd->pageswap == 0 && sz == seg->s_size) {
5839 					svd->swresv = sz;
5840 				} else {
5841 					ASSERT(svd->vpage != NULL);
5842 					svd->swresv += sz;
5843 					svd->pageswap = 1;
5844 					for (cvp = svp; cvp < evp; cvp++) {
5845 						if (!VPP_ISSWAPRES(cvp))
5846 							VPP_SETSWAPRES(cvp);
5847 					}
5848 				}
5849 			}
5850 		} else {
5851 			/*
5852 			 * Swap space is released only if this segment
5853 			 * does not map anonymous memory, since read faults
5854 			 * on such segments still need an anon slot to read
5855 			 * in the data.
5856 			 */
5857 			if (svd->swresv != 0 && svd->vp != NULL &&
5858 			    svd->amp == NULL && addr == seg->s_base &&
5859 			    len == seg->s_size && svd->pageprot == 0) {
5860 				ASSERT(svd->pageswap == 0);
5861 				anon_unresv_zone(svd->swresv,
5862 				    seg->s_as->a_proc->p_zone);
5863 				svd->swresv = 0;
5864 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
5865 				    "anon proc:%p %lu %u", seg, 0, 0);
5866 			}
5867 		}
5868 	}
5869 
5870 	if (addr == seg->s_base && len == seg->s_size && svd->vpage == NULL) {
5871 		if (svd->prot == prot) {
5872 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5873 			return (0);			/* all done */
5874 		}
5875 		svd->prot = (uchar_t)prot;
5876 	} else if (svd->type == MAP_PRIVATE) {
5877 		struct anon *ap = NULL;
5878 		page_t *pp;
5879 		u_offset_t offset, off;
5880 		struct anon_map *amp;
5881 		ulong_t anon_idx = 0;
5882 
5883 		/*
5884 		 * A vpage structure exists or else the change does not
5885 		 * involve the entire segment.  Establish a vpage structure
5886 		 * if none is there.  Then, for each page in the range,
5887 		 * adjust its individual permissions.  Note that write-
5888 		 * enabling a MAP_PRIVATE page can affect the claims for
5889 		 * locked down memory.  Overcommitting memory terminates
5890 		 * the operation.
5891 		 */
5892 		segvn_vpage(seg);
5893 		if (svd->vpage == NULL) {
5894 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5895 			return (ENOMEM);
5896 		}
5897 		svd->pageprot = 1;
5898 		if ((amp = svd->amp) != NULL) {
5899 			anon_idx = svd->anon_index + seg_page(seg, addr);
5900 			ASSERT(seg->s_szc == 0 ||
5901 			    IS_P2ALIGNED(anon_idx, pgcnt));
5902 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5903 		}
5904 
5905 		offset = svd->offset + (uintptr_t)(addr - seg->s_base);
5906 		evp = &svd->vpage[seg_page(seg, addr + len)];
5907 
5908 		/*
5909 		 * See Statement at the beginning of segvn_lockop regarding
5910 		 * the way cowcnts and lckcnts are handled.
5911 		 */
5912 		for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5913 
5914 			if (seg->s_szc != 0) {
5915 				if (amp != NULL) {
5916 					anon_array_enter(amp, anon_idx,
5917 					    &cookie);
5918 				}
5919 				if (IS_P2ALIGNED(anon_idx, pgcnt) &&
5920 				    !segvn_claim_pages(seg, svp, offset,
5921 				    anon_idx, prot)) {
5922 					if (amp != NULL) {
5923 						anon_array_exit(&cookie);
5924 					}
5925 					break;
5926 				}
5927 				if (amp != NULL) {
5928 					anon_array_exit(&cookie);
5929 				}
5930 				anon_idx++;
5931 			} else {
5932 				if (amp != NULL) {
5933 					anon_array_enter(amp, anon_idx,
5934 					    &cookie);
5935 					ap = anon_get_ptr(amp->ahp, anon_idx++);
5936 				}
5937 
5938 				if (VPP_ISPPLOCK(svp) &&
5939 				    VPP_PROT(svp) != prot) {
5940 
5941 					if (amp == NULL || ap == NULL) {
5942 						vp = svd->vp;
5943 						off = offset;
5944 					} else
5945 						swap_xlate(ap, &vp, &off);
5946 					if (amp != NULL)
5947 						anon_array_exit(&cookie);
5948 
5949 					if ((pp = page_lookup(vp, off,
5950 					    SE_SHARED)) == NULL) {
5951 						panic("segvn_setprot: no page");
5952 						/*NOTREACHED*/
5953 					}
5954 					ASSERT(seg->s_szc == 0);
5955 					if ((VPP_PROT(svp) ^ prot) &
5956 					    PROT_WRITE) {
5957 						if (prot & PROT_WRITE) {
5958 							if (!page_addclaim(
5959 							    pp)) {
5960 								page_unlock(pp);
5961 								break;
5962 							}
5963 						} else {
5964 							if (!page_subclaim(
5965 							    pp)) {
5966 								page_unlock(pp);
5967 								break;
5968 							}
5969 						}
5970 					}
5971 					page_unlock(pp);
5972 				} else if (amp != NULL)
5973 					anon_array_exit(&cookie);
5974 			}
5975 			VPP_SETPROT(svp, prot);
5976 			offset += PAGESIZE;
5977 		}
5978 		if (amp != NULL)
5979 			ANON_LOCK_EXIT(&amp->a_rwlock);
5980 
5981 		/*
5982 		 * Did we terminate prematurely?  If so, simply unload
5983 		 * the translations to the things we've updated so far.
5984 		 */
5985 		if (svp != evp) {
5986 			if (unload_done) {
5987 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5988 				return (IE_NOMEM);
5989 			}
5990 			len = (svp - &svd->vpage[seg_page(seg, addr)]) *
5991 			    PAGESIZE;
5992 			ASSERT(seg->s_szc == 0 || IS_P2ALIGNED(len, pgsz));
5993 			if (len != 0)
5994 				hat_unload(seg->s_as->a_hat, addr,
5995 				    len, HAT_UNLOAD);
5996 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5997 			return (IE_NOMEM);
5998 		}
5999 	} else {
6000 		segvn_vpage(seg);
6001 		if (svd->vpage == NULL) {
6002 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6003 			return (ENOMEM);
6004 		}
6005 		svd->pageprot = 1;
6006 		evp = &svd->vpage[seg_page(seg, addr + len)];
6007 		for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
6008 			VPP_SETPROT(svp, prot);
6009 		}
6010 	}
6011 
6012 	if (unload_done) {
6013 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6014 		return (0);
6015 	}
6016 
6017 	if (((prot & PROT_WRITE) != 0 &&
6018 	    (svd->vp != NULL || svd->type == MAP_PRIVATE)) ||
6019 	    (prot & ~PROT_USER) == PROT_NONE) {
6020 		/*
6021 		 * Either private or shared data with write access (in
6022 		 * which case we need to throw out all former translations
6023 		 * so that we get the right translations set up on fault
6024 		 * and we don't allow write access to any copy-on-write pages
6025 		 * that might be around or to prevent write access to pages
6026 		 * representing holes in a file), or we don't have permission
6027 		 * to access the memory at all (in which case we have to
6028 		 * unload any current translations that might exist).
6029 		 */
6030 		hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
6031 	} else {
6032 		/*
6033 		 * A shared mapping or a private mapping in which write
6034 		 * protection is going to be denied - just change all the
6035 		 * protections over the range of addresses in question.
6036 		 * segvn does not support any other attributes other
6037 		 * than prot so we can use hat_chgattr.
6038 		 */
6039 		hat_chgattr(seg->s_as->a_hat, addr, len, prot);
6040 	}
6041 
6042 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6043 
6044 	return (0);
6045 }
6046 
6047 /*
6048  * segvn_setpagesize is called via SEGOP_SETPAGESIZE from as_setpagesize,
6049  * to determine if the seg is capable of mapping the requested szc.
6050  */
6051 static int
6052 segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len, uint_t szc)
6053 {
6054 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6055 	struct segvn_data *nsvd;
6056 	struct anon_map *amp = svd->amp;
6057 	struct seg *nseg;
6058 	caddr_t eaddr = addr + len, a;
6059 	size_t pgsz = page_get_pagesize(szc);
6060 	pgcnt_t pgcnt = page_get_pagecnt(szc);
6061 	int err;
6062 	u_offset_t off = svd->offset + (uintptr_t)(addr - seg->s_base);
6063 
6064 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6065 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6066 
6067 	if (seg->s_szc == szc || segvn_lpg_disable != 0) {
6068 		return (0);
6069 	}
6070 
6071 	/*
6072 	 * addr should always be pgsz aligned but eaddr may be misaligned if
6073 	 * it's at the end of the segment.
6074 	 *
6075 	 * XXX we should assert this condition since as_setpagesize() logic
6076 	 * guarantees it.
6077 	 */
6078 	if (!IS_P2ALIGNED(addr, pgsz) ||
6079 	    (!IS_P2ALIGNED(eaddr, pgsz) &&
6080 	    eaddr != seg->s_base + seg->s_size)) {
6081 
6082 		segvn_setpgsz_align_err++;
6083 		return (EINVAL);
6084 	}
6085 
6086 	if (amp != NULL && svd->type == MAP_SHARED) {
6087 		ulong_t an_idx = svd->anon_index + seg_page(seg, addr);
6088 		if (!IS_P2ALIGNED(an_idx, pgcnt)) {
6089 
6090 			segvn_setpgsz_anon_align_err++;
6091 			return (EINVAL);
6092 		}
6093 	}
6094 
6095 	if ((svd->flags & MAP_NORESERVE) || seg->s_as == &kas ||
6096 	    szc > segvn_maxpgszc) {
6097 		return (EINVAL);
6098 	}
6099 
6100 	/* paranoid check */
6101 	if (svd->vp != NULL &&
6102 	    (IS_SWAPFSVP(svd->vp) || VN_ISKAS(svd->vp))) {
6103 		return (EINVAL);
6104 	}
6105 
6106 	if (seg->s_szc == 0 && svd->vp != NULL &&
6107 	    map_addr_vacalign_check(addr, off)) {
6108 		return (EINVAL);
6109 	}
6110 
6111 	/*
6112 	 * Check that protections are the same within new page
6113 	 * size boundaries.
6114 	 */
6115 	if (svd->pageprot) {
6116 		for (a = addr; a < eaddr; a += pgsz) {
6117 			if ((a + pgsz) > eaddr) {
6118 				if (!sameprot(seg, a, eaddr - a)) {
6119 					return (EINVAL);
6120 				}
6121 			} else {
6122 				if (!sameprot(seg, a, pgsz)) {
6123 					return (EINVAL);
6124 				}
6125 			}
6126 		}
6127 	}
6128 
6129 	/*
6130 	 * Since we are changing page size we first have to flush
6131 	 * the cache. This makes sure all the pagelock calls have
6132 	 * to recheck protections.
6133 	 */
6134 	if (svd->softlockcnt > 0) {
6135 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6136 
6137 		/*
6138 		 * If this is shared segment non 0 softlockcnt
6139 		 * means locked pages are still in use.
6140 		 */
6141 		if (svd->type == MAP_SHARED) {
6142 			return (EAGAIN);
6143 		}
6144 
6145 		/*
6146 		 * Since we do have the segvn writers lock nobody can fill
6147 		 * the cache with entries belonging to this seg during
6148 		 * the purge. The flush either succeeds or we still have
6149 		 * pending I/Os.
6150 		 */
6151 		segvn_purge(seg);
6152 		if (svd->softlockcnt > 0) {
6153 			return (EAGAIN);
6154 		}
6155 	}
6156 
6157 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6158 		ASSERT(svd->amp == NULL);
6159 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6160 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6161 		    HAT_REGION_TEXT);
6162 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6163 	} else if (svd->tr_state == SEGVN_TR_INIT) {
6164 		svd->tr_state = SEGVN_TR_OFF;
6165 	} else if (svd->tr_state == SEGVN_TR_ON) {
6166 		ASSERT(svd->amp != NULL);
6167 		segvn_textunrepl(seg, 1);
6168 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6169 		amp = NULL;
6170 	}
6171 
6172 	/*
6173 	 * Operation for sub range of existing segment.
6174 	 */
6175 	if (addr != seg->s_base || eaddr != (seg->s_base + seg->s_size)) {
6176 		if (szc < seg->s_szc) {
6177 			VM_STAT_ADD(segvnvmstats.demoterange[2]);
6178 			err = segvn_demote_range(seg, addr, len, SDR_RANGE, 0);
6179 			if (err == 0) {
6180 				return (IE_RETRY);
6181 			}
6182 			if (err == ENOMEM) {
6183 				return (IE_NOMEM);
6184 			}
6185 			return (err);
6186 		}
6187 		if (addr != seg->s_base) {
6188 			nseg = segvn_split_seg(seg, addr);
6189 			if (eaddr != (nseg->s_base + nseg->s_size)) {
6190 				/* eaddr is szc aligned */
6191 				(void) segvn_split_seg(nseg, eaddr);
6192 			}
6193 			return (IE_RETRY);
6194 		}
6195 		if (eaddr != (seg->s_base + seg->s_size)) {
6196 			/* eaddr is szc aligned */
6197 			(void) segvn_split_seg(seg, eaddr);
6198 		}
6199 		return (IE_RETRY);
6200 	}
6201 
6202 	/*
6203 	 * Break any low level sharing and reset seg->s_szc to 0.
6204 	 */
6205 	if ((err = segvn_clrszc(seg)) != 0) {
6206 		if (err == ENOMEM) {
6207 			err = IE_NOMEM;
6208 		}
6209 		return (err);
6210 	}
6211 	ASSERT(seg->s_szc == 0);
6212 
6213 	/*
6214 	 * If the end of the current segment is not pgsz aligned
6215 	 * then attempt to concatenate with the next segment.
6216 	 */
6217 	if (!IS_P2ALIGNED(eaddr, pgsz)) {
6218 		nseg = AS_SEGNEXT(seg->s_as, seg);
6219 		if (nseg == NULL || nseg == seg || eaddr != nseg->s_base) {
6220 			return (ENOMEM);
6221 		}
6222 		if (nseg->s_ops != &segvn_ops) {
6223 			return (EINVAL);
6224 		}
6225 		nsvd = (struct segvn_data *)nseg->s_data;
6226 		if (nsvd->softlockcnt > 0) {
6227 			/*
6228 			 * If this is shared segment non 0 softlockcnt
6229 			 * means locked pages are still in use.
6230 			 */
6231 			if (nsvd->type == MAP_SHARED) {
6232 				return (EAGAIN);
6233 			}
6234 			segvn_purge(nseg);
6235 			if (nsvd->softlockcnt > 0) {
6236 				return (EAGAIN);
6237 			}
6238 		}
6239 		err = segvn_clrszc(nseg);
6240 		if (err == ENOMEM) {
6241 			err = IE_NOMEM;
6242 		}
6243 		if (err != 0) {
6244 			return (err);
6245 		}
6246 		ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6247 		err = segvn_concat(seg, nseg, 1);
6248 		if (err == -1) {
6249 			return (EINVAL);
6250 		}
6251 		if (err == -2) {
6252 			return (IE_NOMEM);
6253 		}
6254 		return (IE_RETRY);
6255 	}
6256 
6257 	/*
6258 	 * May need to re-align anon array to
6259 	 * new szc.
6260 	 */
6261 	if (amp != NULL) {
6262 		if (!IS_P2ALIGNED(svd->anon_index, pgcnt)) {
6263 			struct anon_hdr *nahp;
6264 
6265 			ASSERT(svd->type == MAP_PRIVATE);
6266 
6267 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6268 			ASSERT(amp->refcnt == 1);
6269 			nahp = anon_create(btop(amp->size), ANON_NOSLEEP);
6270 			if (nahp == NULL) {
6271 				ANON_LOCK_EXIT(&amp->a_rwlock);
6272 				return (IE_NOMEM);
6273 			}
6274 			if (anon_copy_ptr(amp->ahp, svd->anon_index,
6275 			    nahp, 0, btop(seg->s_size), ANON_NOSLEEP)) {
6276 				anon_release(nahp, btop(amp->size));
6277 				ANON_LOCK_EXIT(&amp->a_rwlock);
6278 				return (IE_NOMEM);
6279 			}
6280 			anon_release(amp->ahp, btop(amp->size));
6281 			amp->ahp = nahp;
6282 			svd->anon_index = 0;
6283 			ANON_LOCK_EXIT(&amp->a_rwlock);
6284 		}
6285 	}
6286 	if (svd->vp != NULL && szc != 0) {
6287 		struct vattr va;
6288 		u_offset_t eoffpage = svd->offset;
6289 		va.va_mask = AT_SIZE;
6290 		eoffpage += seg->s_size;
6291 		eoffpage = btopr(eoffpage);
6292 		if (VOP_GETATTR(svd->vp, &va, 0, svd->cred, NULL) != 0) {
6293 			segvn_setpgsz_getattr_err++;
6294 			return (EINVAL);
6295 		}
6296 		if (btopr(va.va_size) < eoffpage) {
6297 			segvn_setpgsz_eof_err++;
6298 			return (EINVAL);
6299 		}
6300 		if (amp != NULL) {
6301 			/*
6302 			 * anon_fill_cow_holes() may call VOP_GETPAGE().
6303 			 * don't take anon map lock here to avoid holding it
6304 			 * across VOP_GETPAGE() calls that may call back into
6305 			 * segvn for klsutering checks. We don't really need
6306 			 * anon map lock here since it's a private segment and
6307 			 * we hold as level lock as writers.
6308 			 */
6309 			if ((err = anon_fill_cow_holes(seg, seg->s_base,
6310 			    amp->ahp, svd->anon_index, svd->vp, svd->offset,
6311 			    seg->s_size, szc, svd->prot, svd->vpage,
6312 			    svd->cred)) != 0) {
6313 				return (EINVAL);
6314 			}
6315 		}
6316 		segvn_setvnode_mpss(svd->vp);
6317 	}
6318 
6319 	if (amp != NULL) {
6320 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6321 		if (svd->type == MAP_PRIVATE) {
6322 			amp->a_szc = szc;
6323 		} else if (szc > amp->a_szc) {
6324 			amp->a_szc = szc;
6325 		}
6326 		ANON_LOCK_EXIT(&amp->a_rwlock);
6327 	}
6328 
6329 	seg->s_szc = szc;
6330 
6331 	return (0);
6332 }
6333 
6334 static int
6335 segvn_clrszc(struct seg *seg)
6336 {
6337 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6338 	struct anon_map *amp = svd->amp;
6339 	size_t pgsz;
6340 	pgcnt_t pages;
6341 	int err = 0;
6342 	caddr_t a = seg->s_base;
6343 	caddr_t ea = a + seg->s_size;
6344 	ulong_t an_idx = svd->anon_index;
6345 	vnode_t *vp = svd->vp;
6346 	struct vpage *vpage = svd->vpage;
6347 	page_t *anon_pl[1 + 1], *pp;
6348 	struct anon *ap, *oldap;
6349 	uint_t prot = svd->prot, vpprot;
6350 	int pageflag = 0;
6351 
6352 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
6353 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
6354 	ASSERT(svd->softlockcnt == 0);
6355 
6356 	if (vp == NULL && amp == NULL) {
6357 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6358 		seg->s_szc = 0;
6359 		return (0);
6360 	}
6361 
6362 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6363 		ASSERT(svd->amp == NULL);
6364 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6365 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6366 		    HAT_REGION_TEXT);
6367 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6368 	} else if (svd->tr_state == SEGVN_TR_ON) {
6369 		ASSERT(svd->amp != NULL);
6370 		segvn_textunrepl(seg, 1);
6371 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6372 		amp = NULL;
6373 	} else {
6374 		if (svd->tr_state != SEGVN_TR_OFF) {
6375 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
6376 			svd->tr_state = SEGVN_TR_OFF;
6377 		}
6378 
6379 		/*
6380 		 * do HAT_UNLOAD_UNMAP since we are changing the pagesize.
6381 		 * unload argument is 0 when we are freeing the segment
6382 		 * and unload was already done.
6383 		 */
6384 		hat_unload(seg->s_as->a_hat, seg->s_base, seg->s_size,
6385 		    HAT_UNLOAD_UNMAP);
6386 	}
6387 
6388 	if (amp == NULL || svd->type == MAP_SHARED) {
6389 		seg->s_szc = 0;
6390 		return (0);
6391 	}
6392 
6393 	pgsz = page_get_pagesize(seg->s_szc);
6394 	pages = btop(pgsz);
6395 
6396 	/*
6397 	 * XXX anon rwlock is not really needed because this is a
6398 	 * private segment and we are writers.
6399 	 */
6400 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6401 
6402 	for (; a < ea; a += pgsz, an_idx += pages) {
6403 		if ((oldap = anon_get_ptr(amp->ahp, an_idx)) != NULL) {
6404 			ASSERT(vpage != NULL || svd->pageprot == 0);
6405 			if (vpage != NULL) {
6406 				ASSERT(sameprot(seg, a, pgsz));
6407 				prot = VPP_PROT(vpage);
6408 				pageflag = VPP_ISPPLOCK(vpage) ? LOCK_PAGE : 0;
6409 			}
6410 			if (seg->s_szc != 0) {
6411 				ASSERT(vp == NULL || anon_pages(amp->ahp,
6412 				    an_idx, pages) == pages);
6413 				if ((err = anon_map_demotepages(amp, an_idx,
6414 				    seg, a, prot, vpage, svd->cred)) != 0) {
6415 					goto out;
6416 				}
6417 			} else {
6418 				if (oldap->an_refcnt == 1) {
6419 					continue;
6420 				}
6421 				if ((err = anon_getpage(&oldap, &vpprot,
6422 				    anon_pl, PAGESIZE, seg, a, S_READ,
6423 				    svd->cred))) {
6424 					goto out;
6425 				}
6426 				if ((pp = anon_private(&ap, seg, a, prot,
6427 				    anon_pl[0], pageflag, svd->cred)) == NULL) {
6428 					err = ENOMEM;
6429 					goto out;
6430 				}
6431 				anon_decref(oldap);
6432 				(void) anon_set_ptr(amp->ahp, an_idx, ap,
6433 				    ANON_SLEEP);
6434 				page_unlock(pp);
6435 			}
6436 		}
6437 		vpage = (vpage == NULL) ? NULL : vpage + pages;
6438 	}
6439 
6440 	amp->a_szc = 0;
6441 	seg->s_szc = 0;
6442 out:
6443 	ANON_LOCK_EXIT(&amp->a_rwlock);
6444 	return (err);
6445 }
6446 
6447 static int
6448 segvn_claim_pages(
6449 	struct seg *seg,
6450 	struct vpage *svp,
6451 	u_offset_t off,
6452 	ulong_t anon_idx,
6453 	uint_t prot)
6454 {
6455 	pgcnt_t	pgcnt = page_get_pagecnt(seg->s_szc);
6456 	size_t ppasize = (pgcnt + 1) * sizeof (page_t *);
6457 	page_t	**ppa;
6458 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6459 	struct anon_map *amp = svd->amp;
6460 	struct vpage *evp = svp + pgcnt;
6461 	caddr_t addr = ((uintptr_t)(svp - svd->vpage) << PAGESHIFT)
6462 	    + seg->s_base;
6463 	struct anon *ap;
6464 	struct vnode *vp = svd->vp;
6465 	page_t *pp;
6466 	pgcnt_t pg_idx, i;
6467 	int err = 0;
6468 	anoff_t aoff;
6469 	int anon = (amp != NULL) ? 1 : 0;
6470 
6471 	ASSERT(svd->type == MAP_PRIVATE);
6472 	ASSERT(svd->vpage != NULL);
6473 	ASSERT(seg->s_szc != 0);
6474 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
6475 	ASSERT(amp == NULL || IS_P2ALIGNED(anon_idx, pgcnt));
6476 	ASSERT(sameprot(seg, addr, pgcnt << PAGESHIFT));
6477 
6478 	if (VPP_PROT(svp) == prot)
6479 		return (1);
6480 	if (!((VPP_PROT(svp) ^ prot) & PROT_WRITE))
6481 		return (1);
6482 
6483 	ppa = kmem_alloc(ppasize, KM_SLEEP);
6484 	if (anon && vp != NULL) {
6485 		if (anon_get_ptr(amp->ahp, anon_idx) == NULL) {
6486 			anon = 0;
6487 			ASSERT(!anon_pages(amp->ahp, anon_idx, pgcnt));
6488 		}
6489 		ASSERT(!anon ||
6490 		    anon_pages(amp->ahp, anon_idx, pgcnt) == pgcnt);
6491 	}
6492 
6493 	for (*ppa = NULL, pg_idx = 0; svp < evp; svp++, anon_idx++) {
6494 		if (!VPP_ISPPLOCK(svp))
6495 			continue;
6496 		if (anon) {
6497 			ap = anon_get_ptr(amp->ahp, anon_idx);
6498 			if (ap == NULL) {
6499 				panic("segvn_claim_pages: no anon slot");
6500 			}
6501 			swap_xlate(ap, &vp, &aoff);
6502 			off = (u_offset_t)aoff;
6503 		}
6504 		ASSERT(vp != NULL);
6505 		if ((pp = page_lookup(vp,
6506 		    (u_offset_t)off, SE_SHARED)) == NULL) {
6507 			panic("segvn_claim_pages: no page");
6508 		}
6509 		ppa[pg_idx++] = pp;
6510 		off += PAGESIZE;
6511 	}
6512 
6513 	if (ppa[0] == NULL) {
6514 		kmem_free(ppa, ppasize);
6515 		return (1);
6516 	}
6517 
6518 	ASSERT(pg_idx <= pgcnt);
6519 	ppa[pg_idx] = NULL;
6520 
6521 
6522 	/* Find each large page within ppa, and adjust its claim */
6523 
6524 	/* Does ppa cover a single large page? */
6525 	if (ppa[0]->p_szc == seg->s_szc) {
6526 		if (prot & PROT_WRITE)
6527 			err = page_addclaim_pages(ppa);
6528 		else
6529 			err = page_subclaim_pages(ppa);
6530 	} else {
6531 		for (i = 0; ppa[i]; i += pgcnt) {
6532 			ASSERT(IS_P2ALIGNED(page_pptonum(ppa[i]), pgcnt));
6533 			if (prot & PROT_WRITE)
6534 				err = page_addclaim_pages(&ppa[i]);
6535 			else
6536 				err = page_subclaim_pages(&ppa[i]);
6537 			if (err == 0)
6538 				break;
6539 		}
6540 	}
6541 
6542 	for (i = 0; i < pg_idx; i++) {
6543 		ASSERT(ppa[i] != NULL);
6544 		page_unlock(ppa[i]);
6545 	}
6546 
6547 	kmem_free(ppa, ppasize);
6548 	return (err);
6549 }
6550 
6551 /*
6552  * Returns right (upper address) segment if split occurred.
6553  * If the address is equal to the beginning or end of its segment it returns
6554  * the current segment.
6555  */
6556 static struct seg *
6557 segvn_split_seg(struct seg *seg, caddr_t addr)
6558 {
6559 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6560 	struct seg *nseg;
6561 	size_t nsize;
6562 	struct segvn_data *nsvd;
6563 
6564 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6565 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6566 
6567 	ASSERT(addr >= seg->s_base);
6568 	ASSERT(addr <= seg->s_base + seg->s_size);
6569 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6570 
6571 	if (addr == seg->s_base || addr == seg->s_base + seg->s_size)
6572 		return (seg);
6573 
6574 	nsize = seg->s_base + seg->s_size - addr;
6575 	seg->s_size = addr - seg->s_base;
6576 	nseg = seg_alloc(seg->s_as, addr, nsize);
6577 	ASSERT(nseg != NULL);
6578 	nseg->s_ops = seg->s_ops;
6579 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
6580 	nseg->s_data = (void *)nsvd;
6581 	nseg->s_szc = seg->s_szc;
6582 	*nsvd = *svd;
6583 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6584 	nsvd->seg = nseg;
6585 	rw_init(&nsvd->lock, NULL, RW_DEFAULT, NULL);
6586 
6587 	if (nsvd->vp != NULL) {
6588 		VN_HOLD(nsvd->vp);
6589 		nsvd->offset = svd->offset +
6590 		    (uintptr_t)(nseg->s_base - seg->s_base);
6591 		if (nsvd->type == MAP_SHARED)
6592 			lgrp_shm_policy_init(NULL, nsvd->vp);
6593 	} else {
6594 		/*
6595 		 * The offset for an anonymous segment has no signifigance in
6596 		 * terms of an offset into a file. If we were to use the above
6597 		 * calculation instead, the structures read out of
6598 		 * /proc/<pid>/xmap would be more difficult to decipher since
6599 		 * it would be unclear whether two seemingly contiguous
6600 		 * prxmap_t structures represented different segments or a
6601 		 * single segment that had been split up into multiple prxmap_t
6602 		 * structures (e.g. if some part of the segment had not yet
6603 		 * been faulted in).
6604 		 */
6605 		nsvd->offset = 0;
6606 	}
6607 
6608 	ASSERT(svd->softlockcnt == 0);
6609 	ASSERT(svd->softlockcnt_sbase == 0);
6610 	ASSERT(svd->softlockcnt_send == 0);
6611 	crhold(svd->cred);
6612 
6613 	if (svd->vpage != NULL) {
6614 		size_t bytes = vpgtob(seg_pages(seg));
6615 		size_t nbytes = vpgtob(seg_pages(nseg));
6616 		struct vpage *ovpage = svd->vpage;
6617 
6618 		svd->vpage = kmem_alloc(bytes, KM_SLEEP);
6619 		bcopy(ovpage, svd->vpage, bytes);
6620 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
6621 		bcopy(ovpage + seg_pages(seg), nsvd->vpage, nbytes);
6622 		kmem_free(ovpage, bytes + nbytes);
6623 	}
6624 	if (svd->amp != NULL && svd->type == MAP_PRIVATE) {
6625 		struct anon_map *oamp = svd->amp, *namp;
6626 		struct anon_hdr *nahp;
6627 
6628 		ANON_LOCK_ENTER(&oamp->a_rwlock, RW_WRITER);
6629 		ASSERT(oamp->refcnt == 1);
6630 		nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
6631 		(void) anon_copy_ptr(oamp->ahp, svd->anon_index,
6632 		    nahp, 0, btop(seg->s_size), ANON_SLEEP);
6633 
6634 		namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
6635 		namp->a_szc = nseg->s_szc;
6636 		(void) anon_copy_ptr(oamp->ahp,
6637 		    svd->anon_index + btop(seg->s_size),
6638 		    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
6639 		anon_release(oamp->ahp, btop(oamp->size));
6640 		oamp->ahp = nahp;
6641 		oamp->size = seg->s_size;
6642 		svd->anon_index = 0;
6643 		nsvd->amp = namp;
6644 		nsvd->anon_index = 0;
6645 		ANON_LOCK_EXIT(&oamp->a_rwlock);
6646 	} else if (svd->amp != NULL) {
6647 		pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6648 		ASSERT(svd->amp == nsvd->amp);
6649 		ASSERT(seg->s_szc <= svd->amp->a_szc);
6650 		nsvd->anon_index = svd->anon_index + seg_pages(seg);
6651 		ASSERT(IS_P2ALIGNED(nsvd->anon_index, pgcnt));
6652 		ANON_LOCK_ENTER(&svd->amp->a_rwlock, RW_WRITER);
6653 		svd->amp->refcnt++;
6654 		ANON_LOCK_EXIT(&svd->amp->a_rwlock);
6655 	}
6656 
6657 	/*
6658 	 * Split the amount of swap reserved.
6659 	 */
6660 	if (svd->swresv) {
6661 		/*
6662 		 * For MAP_NORESERVE, only allocate swap reserve for pages
6663 		 * being used.  Other segments get enough to cover whole
6664 		 * segment.
6665 		 */
6666 		if (svd->flags & MAP_NORESERVE) {
6667 			size_t	oswresv;
6668 
6669 			ASSERT(svd->amp);
6670 			oswresv = svd->swresv;
6671 			svd->swresv = ptob(anon_pages(svd->amp->ahp,
6672 			    svd->anon_index, btop(seg->s_size)));
6673 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
6674 			    nsvd->anon_index, btop(nseg->s_size)));
6675 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
6676 		} else {
6677 			if (svd->pageswap) {
6678 				svd->swresv = segvn_count_swap_by_vpages(seg);
6679 				ASSERT(nsvd->swresv >= svd->swresv);
6680 				nsvd->swresv -= svd->swresv;
6681 			} else {
6682 				ASSERT(svd->swresv == seg->s_size +
6683 				    nseg->s_size);
6684 				svd->swresv = seg->s_size;
6685 				nsvd->swresv = nseg->s_size;
6686 			}
6687 		}
6688 	}
6689 
6690 	return (nseg);
6691 }
6692 
6693 /*
6694  * called on memory operations (unmap, setprot, setpagesize) for a subset
6695  * of a large page segment to either demote the memory range (SDR_RANGE)
6696  * or the ends (SDR_END) by addr/len.
6697  *
6698  * returns 0 on success. returns errno, including ENOMEM, on failure.
6699  */
6700 static int
6701 segvn_demote_range(
6702 	struct seg *seg,
6703 	caddr_t addr,
6704 	size_t len,
6705 	int flag,
6706 	uint_t szcvec)
6707 {
6708 	caddr_t eaddr = addr + len;
6709 	caddr_t lpgaddr, lpgeaddr;
6710 	struct seg *nseg;
6711 	struct seg *badseg1 = NULL;
6712 	struct seg *badseg2 = NULL;
6713 	size_t pgsz;
6714 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6715 	int err;
6716 	uint_t szc = seg->s_szc;
6717 	uint_t tszcvec;
6718 
6719 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6720 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6721 	ASSERT(szc != 0);
6722 	pgsz = page_get_pagesize(szc);
6723 	ASSERT(seg->s_base != addr || seg->s_size != len);
6724 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6725 	ASSERT(svd->softlockcnt == 0);
6726 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6727 	ASSERT(szcvec == 0 || (flag == SDR_END && svd->type == MAP_SHARED));
6728 
6729 	CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
6730 	ASSERT(flag == SDR_RANGE || eaddr < lpgeaddr || addr > lpgaddr);
6731 	if (flag == SDR_RANGE) {
6732 		/* demote entire range */
6733 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6734 		(void) segvn_split_seg(nseg, lpgeaddr);
6735 		ASSERT(badseg1->s_base == lpgaddr);
6736 		ASSERT(badseg1->s_size == lpgeaddr - lpgaddr);
6737 	} else if (addr != lpgaddr) {
6738 		ASSERT(flag == SDR_END);
6739 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6740 		if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz &&
6741 		    eaddr < lpgaddr + 2 * pgsz) {
6742 			(void) segvn_split_seg(nseg, lpgeaddr);
6743 			ASSERT(badseg1->s_base == lpgaddr);
6744 			ASSERT(badseg1->s_size == 2 * pgsz);
6745 		} else {
6746 			nseg = segvn_split_seg(nseg, lpgaddr + pgsz);
6747 			ASSERT(badseg1->s_base == lpgaddr);
6748 			ASSERT(badseg1->s_size == pgsz);
6749 			if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz) {
6750 				ASSERT(lpgeaddr - lpgaddr > 2 * pgsz);
6751 				nseg = segvn_split_seg(nseg, lpgeaddr - pgsz);
6752 				badseg2 = nseg;
6753 				(void) segvn_split_seg(nseg, lpgeaddr);
6754 				ASSERT(badseg2->s_base == lpgeaddr - pgsz);
6755 				ASSERT(badseg2->s_size == pgsz);
6756 			}
6757 		}
6758 	} else {
6759 		ASSERT(flag == SDR_END);
6760 		ASSERT(eaddr < lpgeaddr);
6761 		badseg1 = nseg = segvn_split_seg(seg, lpgeaddr - pgsz);
6762 		(void) segvn_split_seg(nseg, lpgeaddr);
6763 		ASSERT(badseg1->s_base == lpgeaddr - pgsz);
6764 		ASSERT(badseg1->s_size == pgsz);
6765 	}
6766 
6767 	ASSERT(badseg1 != NULL);
6768 	ASSERT(badseg1->s_szc == szc);
6769 	ASSERT(flag == SDR_RANGE || badseg1->s_size == pgsz ||
6770 	    badseg1->s_size == 2 * pgsz);
6771 	ASSERT(sameprot(badseg1, badseg1->s_base, pgsz));
6772 	ASSERT(badseg1->s_size == pgsz ||
6773 	    sameprot(badseg1, badseg1->s_base + pgsz, pgsz));
6774 	if (err = segvn_clrszc(badseg1)) {
6775 		return (err);
6776 	}
6777 	ASSERT(badseg1->s_szc == 0);
6778 
6779 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6780 		uint_t tszc = highbit(tszcvec) - 1;
6781 		caddr_t ta = MAX(addr, badseg1->s_base);
6782 		caddr_t te;
6783 		size_t tpgsz = page_get_pagesize(tszc);
6784 
6785 		ASSERT(svd->type == MAP_SHARED);
6786 		ASSERT(flag == SDR_END);
6787 		ASSERT(tszc < szc && tszc > 0);
6788 
6789 		if (eaddr > badseg1->s_base + badseg1->s_size) {
6790 			te = badseg1->s_base + badseg1->s_size;
6791 		} else {
6792 			te = eaddr;
6793 		}
6794 
6795 		ASSERT(ta <= te);
6796 		badseg1->s_szc = tszc;
6797 		if (!IS_P2ALIGNED(ta, tpgsz) || !IS_P2ALIGNED(te, tpgsz)) {
6798 			if (badseg2 != NULL) {
6799 				err = segvn_demote_range(badseg1, ta, te - ta,
6800 				    SDR_END, tszcvec);
6801 				if (err != 0) {
6802 					return (err);
6803 				}
6804 			} else {
6805 				return (segvn_demote_range(badseg1, ta,
6806 				    te - ta, SDR_END, tszcvec));
6807 			}
6808 		}
6809 	}
6810 
6811 	if (badseg2 == NULL)
6812 		return (0);
6813 	ASSERT(badseg2->s_szc == szc);
6814 	ASSERT(badseg2->s_size == pgsz);
6815 	ASSERT(sameprot(badseg2, badseg2->s_base, badseg2->s_size));
6816 	if (err = segvn_clrszc(badseg2)) {
6817 		return (err);
6818 	}
6819 	ASSERT(badseg2->s_szc == 0);
6820 
6821 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6822 		uint_t tszc = highbit(tszcvec) - 1;
6823 		size_t tpgsz = page_get_pagesize(tszc);
6824 
6825 		ASSERT(svd->type == MAP_SHARED);
6826 		ASSERT(flag == SDR_END);
6827 		ASSERT(tszc < szc && tszc > 0);
6828 		ASSERT(badseg2->s_base > addr);
6829 		ASSERT(eaddr > badseg2->s_base);
6830 		ASSERT(eaddr < badseg2->s_base + badseg2->s_size);
6831 
6832 		badseg2->s_szc = tszc;
6833 		if (!IS_P2ALIGNED(eaddr, tpgsz)) {
6834 			return (segvn_demote_range(badseg2, badseg2->s_base,
6835 			    eaddr - badseg2->s_base, SDR_END, tszcvec));
6836 		}
6837 	}
6838 
6839 	return (0);
6840 }
6841 
6842 static int
6843 segvn_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
6844 {
6845 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6846 	struct vpage *vp, *evp;
6847 
6848 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6849 
6850 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6851 	/*
6852 	 * If segment protection can be used, simply check against them.
6853 	 */
6854 	if (svd->pageprot == 0) {
6855 		int err;
6856 
6857 		err = ((svd->prot & prot) != prot) ? EACCES : 0;
6858 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6859 		return (err);
6860 	}
6861 
6862 	/*
6863 	 * Have to check down to the vpage level.
6864 	 */
6865 	evp = &svd->vpage[seg_page(seg, addr + len)];
6866 	for (vp = &svd->vpage[seg_page(seg, addr)]; vp < evp; vp++) {
6867 		if ((VPP_PROT(vp) & prot) != prot) {
6868 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6869 			return (EACCES);
6870 		}
6871 	}
6872 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6873 	return (0);
6874 }
6875 
6876 static int
6877 segvn_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
6878 {
6879 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6880 	size_t pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
6881 
6882 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6883 
6884 	if (pgno != 0) {
6885 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6886 		if (svd->pageprot == 0) {
6887 			do {
6888 				protv[--pgno] = svd->prot;
6889 			} while (pgno != 0);
6890 		} else {
6891 			size_t pgoff = seg_page(seg, addr);
6892 
6893 			do {
6894 				pgno--;
6895 				protv[pgno] = VPP_PROT(&svd->vpage[pgno+pgoff]);
6896 			} while (pgno != 0);
6897 		}
6898 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6899 	}
6900 	return (0);
6901 }
6902 
6903 static u_offset_t
6904 segvn_getoffset(struct seg *seg, caddr_t addr)
6905 {
6906 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6907 
6908 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6909 
6910 	return (svd->offset + (uintptr_t)(addr - seg->s_base));
6911 }
6912 
6913 /*ARGSUSED*/
6914 static int
6915 segvn_gettype(struct seg *seg, caddr_t addr)
6916 {
6917 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6918 
6919 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6920 
6921 	return (svd->type | (svd->flags & (MAP_NORESERVE | MAP_TEXT |
6922 	    MAP_INITDATA)));
6923 }
6924 
6925 /*ARGSUSED*/
6926 static int
6927 segvn_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
6928 {
6929 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6930 
6931 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6932 
6933 	*vpp = svd->vp;
6934 	return (0);
6935 }
6936 
6937 /*
6938  * Check to see if it makes sense to do kluster/read ahead to
6939  * addr + delta relative to the mapping at addr.  We assume here
6940  * that delta is a signed PAGESIZE'd multiple (which can be negative).
6941  *
6942  * For segvn, we currently "approve" of the action if we are
6943  * still in the segment and it maps from the same vp/off,
6944  * or if the advice stored in segvn_data or vpages allows it.
6945  * Currently, klustering is not allowed only if MADV_RANDOM is set.
6946  */
6947 static int
6948 segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
6949 {
6950 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6951 	struct anon *oap, *ap;
6952 	ssize_t pd;
6953 	size_t page;
6954 	struct vnode *vp1, *vp2;
6955 	u_offset_t off1, off2;
6956 	struct anon_map *amp;
6957 
6958 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6959 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
6960 	    SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
6961 
6962 	if (addr + delta < seg->s_base ||
6963 	    addr + delta >= (seg->s_base + seg->s_size))
6964 		return (-1);		/* exceeded segment bounds */
6965 
6966 	pd = delta / (ssize_t)PAGESIZE;	/* divide to preserve sign bit */
6967 	page = seg_page(seg, addr);
6968 
6969 	/*
6970 	 * Check to see if either of the pages addr or addr + delta
6971 	 * have advice set that prevents klustering (if MADV_RANDOM advice
6972 	 * is set for entire segment, or MADV_SEQUENTIAL is set and delta
6973 	 * is negative).
6974 	 */
6975 	if (svd->advice == MADV_RANDOM ||
6976 	    svd->advice == MADV_SEQUENTIAL && delta < 0)
6977 		return (-1);
6978 	else if (svd->pageadvice && svd->vpage) {
6979 		struct vpage *bvpp, *evpp;
6980 
6981 		bvpp = &svd->vpage[page];
6982 		evpp = &svd->vpage[page + pd];
6983 		if (VPP_ADVICE(bvpp) == MADV_RANDOM ||
6984 		    VPP_ADVICE(evpp) == MADV_SEQUENTIAL && delta < 0)
6985 			return (-1);
6986 		if (VPP_ADVICE(bvpp) != VPP_ADVICE(evpp) &&
6987 		    VPP_ADVICE(evpp) == MADV_RANDOM)
6988 			return (-1);
6989 	}
6990 
6991 	if (svd->type == MAP_SHARED)
6992 		return (0);		/* shared mapping - all ok */
6993 
6994 	if ((amp = svd->amp) == NULL)
6995 		return (0);		/* off original vnode */
6996 
6997 	page += svd->anon_index;
6998 
6999 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7000 
7001 	oap = anon_get_ptr(amp->ahp, page);
7002 	ap = anon_get_ptr(amp->ahp, page + pd);
7003 
7004 	ANON_LOCK_EXIT(&amp->a_rwlock);
7005 
7006 	if ((oap == NULL && ap != NULL) || (oap != NULL && ap == NULL)) {
7007 		return (-1);		/* one with and one without an anon */
7008 	}
7009 
7010 	if (oap == NULL) {		/* implies that ap == NULL */
7011 		return (0);		/* off original vnode */
7012 	}
7013 
7014 	/*
7015 	 * Now we know we have two anon pointers - check to
7016 	 * see if they happen to be properly allocated.
7017 	 */
7018 
7019 	/*
7020 	 * XXX We cheat here and don't lock the anon slots. We can't because
7021 	 * we may have been called from the anon layer which might already
7022 	 * have locked them. We are holding a refcnt on the slots so they
7023 	 * can't disappear. The worst that will happen is we'll get the wrong
7024 	 * names (vp, off) for the slots and make a poor klustering decision.
7025 	 */
7026 	swap_xlate(ap, &vp1, &off1);
7027 	swap_xlate(oap, &vp2, &off2);
7028 
7029 
7030 	if (!VOP_CMP(vp1, vp2, NULL) || off1 - off2 != delta)
7031 		return (-1);
7032 	return (0);
7033 }
7034 
7035 /*
7036  * Swap the pages of seg out to secondary storage, returning the
7037  * number of bytes of storage freed.
7038  *
7039  * The basic idea is first to unload all translations and then to call
7040  * VOP_PUTPAGE() for all newly-unmapped pages, to push them out to the
7041  * swap device.  Pages to which other segments have mappings will remain
7042  * mapped and won't be swapped.  Our caller (as_swapout) has already
7043  * performed the unloading step.
7044  *
7045  * The value returned is intended to correlate well with the process's
7046  * memory requirements.  However, there are some caveats:
7047  * 1)	When given a shared segment as argument, this routine will
7048  *	only succeed in swapping out pages for the last sharer of the
7049  *	segment.  (Previous callers will only have decremented mapping
7050  *	reference counts.)
7051  * 2)	We assume that the hat layer maintains a large enough translation
7052  *	cache to capture process reference patterns.
7053  */
7054 static size_t
7055 segvn_swapout(struct seg *seg)
7056 {
7057 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7058 	struct anon_map *amp;
7059 	pgcnt_t pgcnt = 0;
7060 	pgcnt_t npages;
7061 	pgcnt_t page;
7062 	ulong_t anon_index;
7063 
7064 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7065 
7066 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7067 	/*
7068 	 * Find pages unmapped by our caller and force them
7069 	 * out to the virtual swap device.
7070 	 */
7071 	if ((amp = svd->amp) != NULL)
7072 		anon_index = svd->anon_index;
7073 	npages = seg->s_size >> PAGESHIFT;
7074 	for (page = 0; page < npages; page++) {
7075 		page_t *pp;
7076 		struct anon *ap;
7077 		struct vnode *vp;
7078 		u_offset_t off;
7079 		anon_sync_obj_t cookie;
7080 
7081 		/*
7082 		 * Obtain <vp, off> pair for the page, then look it up.
7083 		 *
7084 		 * Note that this code is willing to consider regular
7085 		 * pages as well as anon pages.  Is this appropriate here?
7086 		 */
7087 		ap = NULL;
7088 		if (amp != NULL) {
7089 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7090 			if (anon_array_try_enter(amp, anon_index + page,
7091 			    &cookie)) {
7092 				ANON_LOCK_EXIT(&amp->a_rwlock);
7093 				continue;
7094 			}
7095 			ap = anon_get_ptr(amp->ahp, anon_index + page);
7096 			if (ap != NULL) {
7097 				swap_xlate(ap, &vp, &off);
7098 			} else {
7099 				vp = svd->vp;
7100 				off = svd->offset + ptob(page);
7101 			}
7102 			anon_array_exit(&cookie);
7103 			ANON_LOCK_EXIT(&amp->a_rwlock);
7104 		} else {
7105 			vp = svd->vp;
7106 			off = svd->offset + ptob(page);
7107 		}
7108 		if (vp == NULL) {		/* untouched zfod page */
7109 			ASSERT(ap == NULL);
7110 			continue;
7111 		}
7112 
7113 		pp = page_lookup_nowait(vp, off, SE_SHARED);
7114 		if (pp == NULL)
7115 			continue;
7116 
7117 
7118 		/*
7119 		 * Examine the page to see whether it can be tossed out,
7120 		 * keeping track of how many we've found.
7121 		 */
7122 		if (!page_tryupgrade(pp)) {
7123 			/*
7124 			 * If the page has an i/o lock and no mappings,
7125 			 * it's very likely that the page is being
7126 			 * written out as a result of klustering.
7127 			 * Assume this is so and take credit for it here.
7128 			 */
7129 			if (!page_io_trylock(pp)) {
7130 				if (!hat_page_is_mapped(pp))
7131 					pgcnt++;
7132 			} else {
7133 				page_io_unlock(pp);
7134 			}
7135 			page_unlock(pp);
7136 			continue;
7137 		}
7138 		ASSERT(!page_iolock_assert(pp));
7139 
7140 
7141 		/*
7142 		 * Skip if page is locked or has mappings.
7143 		 * We don't need the page_struct_lock to look at lckcnt
7144 		 * and cowcnt because the page is exclusive locked.
7145 		 */
7146 		if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0 ||
7147 		    hat_page_is_mapped(pp)) {
7148 			page_unlock(pp);
7149 			continue;
7150 		}
7151 
7152 		/*
7153 		 * dispose skips large pages so try to demote first.
7154 		 */
7155 		if (pp->p_szc != 0 && !page_try_demote_pages(pp)) {
7156 			page_unlock(pp);
7157 			/*
7158 			 * XXX should skip the remaining page_t's of this
7159 			 * large page.
7160 			 */
7161 			continue;
7162 		}
7163 
7164 		ASSERT(pp->p_szc == 0);
7165 
7166 		/*
7167 		 * No longer mapped -- we can toss it out.  How
7168 		 * we do so depends on whether or not it's dirty.
7169 		 */
7170 		if (hat_ismod(pp) && pp->p_vnode) {
7171 			/*
7172 			 * We must clean the page before it can be
7173 			 * freed.  Setting B_FREE will cause pvn_done
7174 			 * to free the page when the i/o completes.
7175 			 * XXX:	This also causes it to be accounted
7176 			 *	as a pageout instead of a swap: need
7177 			 *	B_SWAPOUT bit to use instead of B_FREE.
7178 			 *
7179 			 * Hold the vnode before releasing the page lock
7180 			 * to prevent it from being freed and re-used by
7181 			 * some other thread.
7182 			 */
7183 			VN_HOLD(vp);
7184 			page_unlock(pp);
7185 
7186 			/*
7187 			 * Queue all i/o requests for the pageout thread
7188 			 * to avoid saturating the pageout devices.
7189 			 */
7190 			if (!queue_io_request(vp, off))
7191 				VN_RELE(vp);
7192 		} else {
7193 			/*
7194 			 * The page was clean, free it.
7195 			 *
7196 			 * XXX:	Can we ever encounter modified pages
7197 			 *	with no associated vnode here?
7198 			 */
7199 			ASSERT(pp->p_vnode != NULL);
7200 			/*LINTED: constant in conditional context*/
7201 			VN_DISPOSE(pp, B_FREE, 0, kcred);
7202 		}
7203 
7204 		/*
7205 		 * Credit now even if i/o is in progress.
7206 		 */
7207 		pgcnt++;
7208 	}
7209 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7210 
7211 	/*
7212 	 * Wakeup pageout to initiate i/o on all queued requests.
7213 	 */
7214 	cv_signal_pageout();
7215 	return (ptob(pgcnt));
7216 }
7217 
7218 /*
7219  * Synchronize primary storage cache with real object in virtual memory.
7220  *
7221  * XXX - Anonymous pages should not be sync'ed out at all.
7222  */
7223 static int
7224 segvn_sync(struct seg *seg, caddr_t addr, size_t len, int attr, uint_t flags)
7225 {
7226 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7227 	struct vpage *vpp;
7228 	page_t *pp;
7229 	u_offset_t offset;
7230 	struct vnode *vp;
7231 	u_offset_t off;
7232 	caddr_t eaddr;
7233 	int bflags;
7234 	int err = 0;
7235 	int segtype;
7236 	int pageprot;
7237 	int prot;
7238 	ulong_t anon_index;
7239 	struct anon_map *amp;
7240 	struct anon *ap;
7241 	anon_sync_obj_t cookie;
7242 
7243 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7244 
7245 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7246 
7247 	if (svd->softlockcnt > 0) {
7248 		/*
7249 		 * If this is shared segment non 0 softlockcnt
7250 		 * means locked pages are still in use.
7251 		 */
7252 		if (svd->type == MAP_SHARED) {
7253 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7254 			return (EAGAIN);
7255 		}
7256 
7257 		/*
7258 		 * flush all pages from seg cache
7259 		 * otherwise we may deadlock in swap_putpage
7260 		 * for B_INVAL page (4175402).
7261 		 *
7262 		 * Even if we grab segvn WRITER's lock
7263 		 * here, there might be another thread which could've
7264 		 * successfully performed lookup/insert just before
7265 		 * we acquired the lock here.  So, grabbing either
7266 		 * lock here is of not much use.  Until we devise
7267 		 * a strategy at upper layers to solve the
7268 		 * synchronization issues completely, we expect
7269 		 * applications to handle this appropriately.
7270 		 */
7271 		segvn_purge(seg);
7272 		if (svd->softlockcnt > 0) {
7273 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7274 			return (EAGAIN);
7275 		}
7276 	} else if (svd->type == MAP_SHARED && svd->amp != NULL &&
7277 	    svd->amp->a_softlockcnt > 0) {
7278 		/*
7279 		 * Try to purge this amp's entries from pcache. It will
7280 		 * succeed only if other segments that share the amp have no
7281 		 * outstanding softlock's.
7282 		 */
7283 		segvn_purge(seg);
7284 		if (svd->amp->a_softlockcnt > 0 || svd->softlockcnt > 0) {
7285 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7286 			return (EAGAIN);
7287 		}
7288 	}
7289 
7290 	vpp = svd->vpage;
7291 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7292 	bflags = ((flags & MS_ASYNC) ? B_ASYNC : 0) |
7293 	    ((flags & MS_INVALIDATE) ? B_INVAL : 0);
7294 
7295 	if (attr) {
7296 		pageprot = attr & ~(SHARED|PRIVATE);
7297 		segtype = (attr & SHARED) ? MAP_SHARED : MAP_PRIVATE;
7298 
7299 		/*
7300 		 * We are done if the segment types don't match
7301 		 * or if we have segment level protections and
7302 		 * they don't match.
7303 		 */
7304 		if (svd->type != segtype) {
7305 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7306 			return (0);
7307 		}
7308 		if (vpp == NULL) {
7309 			if (svd->prot != pageprot) {
7310 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7311 				return (0);
7312 			}
7313 			prot = svd->prot;
7314 		} else
7315 			vpp = &svd->vpage[seg_page(seg, addr)];
7316 
7317 	} else if (svd->vp && svd->amp == NULL &&
7318 	    (flags & MS_INVALIDATE) == 0) {
7319 
7320 		/*
7321 		 * No attributes, no anonymous pages and MS_INVALIDATE flag
7322 		 * is not on, just use one big request.
7323 		 */
7324 		err = VOP_PUTPAGE(svd->vp, (offset_t)offset, len,
7325 		    bflags, svd->cred, NULL);
7326 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7327 		return (err);
7328 	}
7329 
7330 	if ((amp = svd->amp) != NULL)
7331 		anon_index = svd->anon_index + seg_page(seg, addr);
7332 
7333 	for (eaddr = addr + len; addr < eaddr; addr += PAGESIZE) {
7334 		ap = NULL;
7335 		if (amp != NULL) {
7336 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7337 			anon_array_enter(amp, anon_index, &cookie);
7338 			ap = anon_get_ptr(amp->ahp, anon_index++);
7339 			if (ap != NULL) {
7340 				swap_xlate(ap, &vp, &off);
7341 			} else {
7342 				vp = svd->vp;
7343 				off = offset;
7344 			}
7345 			anon_array_exit(&cookie);
7346 			ANON_LOCK_EXIT(&amp->a_rwlock);
7347 		} else {
7348 			vp = svd->vp;
7349 			off = offset;
7350 		}
7351 		offset += PAGESIZE;
7352 
7353 		if (vp == NULL)		/* untouched zfod page */
7354 			continue;
7355 
7356 		if (attr) {
7357 			if (vpp) {
7358 				prot = VPP_PROT(vpp);
7359 				vpp++;
7360 			}
7361 			if (prot != pageprot) {
7362 				continue;
7363 			}
7364 		}
7365 
7366 		/*
7367 		 * See if any of these pages are locked --  if so, then we
7368 		 * will have to truncate an invalidate request at the first
7369 		 * locked one. We don't need the page_struct_lock to test
7370 		 * as this is only advisory; even if we acquire it someone
7371 		 * might race in and lock the page after we unlock and before
7372 		 * we do the PUTPAGE, then PUTPAGE simply does nothing.
7373 		 */
7374 		if (flags & MS_INVALIDATE) {
7375 			if ((pp = page_lookup(vp, off, SE_SHARED)) != NULL) {
7376 				if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
7377 					page_unlock(pp);
7378 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7379 					return (EBUSY);
7380 				}
7381 				if (ap != NULL && pp->p_szc != 0 &&
7382 				    page_tryupgrade(pp)) {
7383 					if (pp->p_lckcnt == 0 &&
7384 					    pp->p_cowcnt == 0) {
7385 						/*
7386 						 * swapfs VN_DISPOSE() won't
7387 						 * invalidate large pages.
7388 						 * Attempt to demote.
7389 						 * XXX can't help it if it
7390 						 * fails. But for swapfs
7391 						 * pages it is no big deal.
7392 						 */
7393 						(void) page_try_demote_pages(
7394 						    pp);
7395 					}
7396 				}
7397 				page_unlock(pp);
7398 			}
7399 		} else if (svd->type == MAP_SHARED && amp != NULL) {
7400 			/*
7401 			 * Avoid writing out to disk ISM's large pages
7402 			 * because segspt_free_pages() relies on NULL an_pvp
7403 			 * of anon slots of such pages.
7404 			 */
7405 
7406 			ASSERT(svd->vp == NULL);
7407 			/*
7408 			 * swapfs uses page_lookup_nowait if not freeing or
7409 			 * invalidating and skips a page if
7410 			 * page_lookup_nowait returns NULL.
7411 			 */
7412 			pp = page_lookup_nowait(vp, off, SE_SHARED);
7413 			if (pp == NULL) {
7414 				continue;
7415 			}
7416 			if (pp->p_szc != 0) {
7417 				page_unlock(pp);
7418 				continue;
7419 			}
7420 
7421 			/*
7422 			 * Note ISM pages are created large so (vp, off)'s
7423 			 * page cannot suddenly become large after we unlock
7424 			 * pp.
7425 			 */
7426 			page_unlock(pp);
7427 		}
7428 		/*
7429 		 * XXX - Should ultimately try to kluster
7430 		 * calls to VOP_PUTPAGE() for performance.
7431 		 */
7432 		VN_HOLD(vp);
7433 		err = VOP_PUTPAGE(vp, (offset_t)off, PAGESIZE,
7434 		    (bflags | (IS_SWAPFSVP(vp) ? B_PAGE_NOWAIT : 0)),
7435 		    svd->cred, NULL);
7436 
7437 		VN_RELE(vp);
7438 		if (err)
7439 			break;
7440 	}
7441 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7442 	return (err);
7443 }
7444 
7445 /*
7446  * Determine if we have data corresponding to pages in the
7447  * primary storage virtual memory cache (i.e., "in core").
7448  */
7449 static size_t
7450 segvn_incore(struct seg *seg, caddr_t addr, size_t len, char *vec)
7451 {
7452 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7453 	struct vnode *vp, *avp;
7454 	u_offset_t offset, aoffset;
7455 	size_t p, ep;
7456 	int ret;
7457 	struct vpage *vpp;
7458 	page_t *pp;
7459 	uint_t start;
7460 	struct anon_map *amp;		/* XXX - for locknest */
7461 	struct anon *ap;
7462 	uint_t attr;
7463 	anon_sync_obj_t cookie;
7464 
7465 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7466 
7467 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7468 	if (svd->amp == NULL && svd->vp == NULL) {
7469 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7470 		bzero(vec, btopr(len));
7471 		return (len);	/* no anonymous pages created yet */
7472 	}
7473 
7474 	p = seg_page(seg, addr);
7475 	ep = seg_page(seg, addr + len);
7476 	start = svd->vp ? SEG_PAGE_VNODEBACKED : 0;
7477 
7478 	amp = svd->amp;
7479 	for (; p < ep; p++, addr += PAGESIZE) {
7480 		vpp = (svd->vpage) ? &svd->vpage[p]: NULL;
7481 		ret = start;
7482 		ap = NULL;
7483 		avp = NULL;
7484 		/* Grab the vnode/offset for the anon slot */
7485 		if (amp != NULL) {
7486 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7487 			anon_array_enter(amp, svd->anon_index + p, &cookie);
7488 			ap = anon_get_ptr(amp->ahp, svd->anon_index + p);
7489 			if (ap != NULL) {
7490 				swap_xlate(ap, &avp, &aoffset);
7491 			}
7492 			anon_array_exit(&cookie);
7493 			ANON_LOCK_EXIT(&amp->a_rwlock);
7494 		}
7495 		if ((avp != NULL) && page_exists(avp, aoffset)) {
7496 			/* A page exists for the anon slot */
7497 			ret |= SEG_PAGE_INCORE;
7498 
7499 			/*
7500 			 * If page is mapped and writable
7501 			 */
7502 			attr = (uint_t)0;
7503 			if ((hat_getattr(seg->s_as->a_hat, addr,
7504 			    &attr) != -1) && (attr & PROT_WRITE)) {
7505 				ret |= SEG_PAGE_ANON;
7506 			}
7507 			/*
7508 			 * Don't get page_struct lock for lckcnt and cowcnt,
7509 			 * since this is purely advisory.
7510 			 */
7511 			if ((pp = page_lookup_nowait(avp, aoffset,
7512 			    SE_SHARED)) != NULL) {
7513 				if (pp->p_lckcnt)
7514 					ret |= SEG_PAGE_SOFTLOCK;
7515 				if (pp->p_cowcnt)
7516 					ret |= SEG_PAGE_HASCOW;
7517 				page_unlock(pp);
7518 			}
7519 		}
7520 
7521 		/* Gather vnode statistics */
7522 		vp = svd->vp;
7523 		offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7524 
7525 		if (vp != NULL) {
7526 			/*
7527 			 * Try to obtain a "shared" lock on the page
7528 			 * without blocking.  If this fails, determine
7529 			 * if the page is in memory.
7530 			 */
7531 			pp = page_lookup_nowait(vp, offset, SE_SHARED);
7532 			if ((pp == NULL) && (page_exists(vp, offset))) {
7533 				/* Page is incore, and is named */
7534 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7535 			}
7536 			/*
7537 			 * Don't get page_struct lock for lckcnt and cowcnt,
7538 			 * since this is purely advisory.
7539 			 */
7540 			if (pp != NULL) {
7541 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7542 				if (pp->p_lckcnt)
7543 					ret |= SEG_PAGE_SOFTLOCK;
7544 				if (pp->p_cowcnt)
7545 					ret |= SEG_PAGE_HASCOW;
7546 				page_unlock(pp);
7547 			}
7548 		}
7549 
7550 		/* Gather virtual page information */
7551 		if (vpp) {
7552 			if (VPP_ISPPLOCK(vpp))
7553 				ret |= SEG_PAGE_LOCKED;
7554 			vpp++;
7555 		}
7556 
7557 		*vec++ = (char)ret;
7558 	}
7559 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7560 	return (len);
7561 }
7562 
7563 /*
7564  * Statement for p_cowcnts/p_lckcnts.
7565  *
7566  * p_cowcnt is updated while mlock/munlocking MAP_PRIVATE and PROT_WRITE region
7567  * irrespective of the following factors or anything else:
7568  *
7569  *	(1) anon slots are populated or not
7570  *	(2) cow is broken or not
7571  *	(3) refcnt on ap is 1 or greater than 1
7572  *
7573  * If it's not MAP_PRIVATE and PROT_WRITE, p_lckcnt is updated during mlock
7574  * and munlock.
7575  *
7576  *
7577  * Handling p_cowcnts/p_lckcnts during copy-on-write fault:
7578  *
7579  *	if vpage has PROT_WRITE
7580  *		transfer cowcnt on the oldpage -> cowcnt on the newpage
7581  *	else
7582  *		transfer lckcnt on the oldpage -> lckcnt on the newpage
7583  *
7584  *	During copy-on-write, decrement p_cowcnt on the oldpage and increment
7585  *	p_cowcnt on the newpage *if* the corresponding vpage has PROT_WRITE.
7586  *
7587  *	We may also break COW if softlocking on read access in the physio case.
7588  *	In this case, vpage may not have PROT_WRITE. So, we need to decrement
7589  *	p_lckcnt on the oldpage and increment p_lckcnt on the newpage *if* the
7590  *	vpage doesn't have PROT_WRITE.
7591  *
7592  *
7593  * Handling p_cowcnts/p_lckcnts during mprotect on mlocked region:
7594  *
7595  * 	If a MAP_PRIVATE region loses PROT_WRITE, we decrement p_cowcnt and
7596  *	increment p_lckcnt by calling page_subclaim() which takes care of
7597  * 	availrmem accounting and p_lckcnt overflow.
7598  *
7599  *	If a MAP_PRIVATE region gains PROT_WRITE, we decrement p_lckcnt and
7600  *	increment p_cowcnt by calling page_addclaim() which takes care of
7601  *	availrmem availability and p_cowcnt overflow.
7602  */
7603 
7604 /*
7605  * Lock down (or unlock) pages mapped by this segment.
7606  *
7607  * XXX only creates PAGESIZE pages if anon slots are not initialized.
7608  * At fault time they will be relocated into larger pages.
7609  */
7610 static int
7611 segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
7612     int attr, int op, ulong_t *lockmap, size_t pos)
7613 {
7614 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7615 	struct vpage *vpp;
7616 	struct vpage *evp;
7617 	page_t *pp;
7618 	u_offset_t offset;
7619 	u_offset_t off;
7620 	int segtype;
7621 	int pageprot;
7622 	int claim;
7623 	struct vnode *vp;
7624 	ulong_t anon_index;
7625 	struct anon_map *amp;
7626 	struct anon *ap;
7627 	struct vattr va;
7628 	anon_sync_obj_t cookie;
7629 	struct kshmid *sp = NULL;
7630 	struct proc	*p = curproc;
7631 	kproject_t	*proj = NULL;
7632 	int chargeproc = 1;
7633 	size_t locked_bytes = 0;
7634 	size_t unlocked_bytes = 0;
7635 	int err = 0;
7636 
7637 	/*
7638 	 * Hold write lock on address space because may split or concatenate
7639 	 * segments
7640 	 */
7641 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7642 
7643 	/*
7644 	 * If this is a shm, use shm's project and zone, else use
7645 	 * project and zone of calling process
7646 	 */
7647 
7648 	/* Determine if this segment backs a sysV shm */
7649 	if (svd->amp != NULL && svd->amp->a_sp != NULL) {
7650 		ASSERT(svd->type == MAP_SHARED);
7651 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
7652 		sp = svd->amp->a_sp;
7653 		proj = sp->shm_perm.ipc_proj;
7654 		chargeproc = 0;
7655 	}
7656 
7657 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7658 	if (attr) {
7659 		pageprot = attr & ~(SHARED|PRIVATE);
7660 		segtype = attr & SHARED ? MAP_SHARED : MAP_PRIVATE;
7661 
7662 		/*
7663 		 * We are done if the segment types don't match
7664 		 * or if we have segment level protections and
7665 		 * they don't match.
7666 		 */
7667 		if (svd->type != segtype) {
7668 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7669 			return (0);
7670 		}
7671 		if (svd->pageprot == 0 && svd->prot != pageprot) {
7672 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7673 			return (0);
7674 		}
7675 	}
7676 
7677 	if (op == MC_LOCK) {
7678 		if (svd->tr_state == SEGVN_TR_INIT) {
7679 			svd->tr_state = SEGVN_TR_OFF;
7680 		} else if (svd->tr_state == SEGVN_TR_ON) {
7681 			ASSERT(svd->amp != NULL);
7682 			segvn_textunrepl(seg, 0);
7683 			ASSERT(svd->amp == NULL &&
7684 			    svd->tr_state == SEGVN_TR_OFF);
7685 		}
7686 	}
7687 
7688 	/*
7689 	 * If we're locking, then we must create a vpage structure if
7690 	 * none exists.  If we're unlocking, then check to see if there
7691 	 * is a vpage --  if not, then we could not have locked anything.
7692 	 */
7693 
7694 	if ((vpp = svd->vpage) == NULL) {
7695 		if (op == MC_LOCK) {
7696 			segvn_vpage(seg);
7697 			if (svd->vpage == NULL) {
7698 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7699 				return (ENOMEM);
7700 			}
7701 		} else {
7702 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7703 			return (0);
7704 		}
7705 	}
7706 
7707 	/*
7708 	 * The anonymous data vector (i.e., previously
7709 	 * unreferenced mapping to swap space) can be allocated
7710 	 * by lazily testing for its existence.
7711 	 */
7712 	if (op == MC_LOCK && svd->amp == NULL && svd->vp == NULL) {
7713 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
7714 		svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
7715 		svd->amp->a_szc = seg->s_szc;
7716 	}
7717 
7718 	if ((amp = svd->amp) != NULL) {
7719 		anon_index = svd->anon_index + seg_page(seg, addr);
7720 	}
7721 
7722 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7723 	evp = &svd->vpage[seg_page(seg, addr + len)];
7724 
7725 	if (sp != NULL)
7726 		mutex_enter(&sp->shm_mlock);
7727 
7728 	/* determine number of unlocked bytes in range for lock operation */
7729 	if (op == MC_LOCK) {
7730 
7731 		if (sp == NULL) {
7732 			for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7733 			    vpp++) {
7734 				if (!VPP_ISPPLOCK(vpp))
7735 					unlocked_bytes += PAGESIZE;
7736 			}
7737 		} else {
7738 			ulong_t		i_idx, i_edx;
7739 			anon_sync_obj_t	i_cookie;
7740 			struct anon	*i_ap;
7741 			struct vnode	*i_vp;
7742 			u_offset_t	i_off;
7743 
7744 			/* Only count sysV pages once for locked memory */
7745 			i_edx = svd->anon_index + seg_page(seg, addr + len);
7746 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7747 			for (i_idx = anon_index; i_idx < i_edx; i_idx++) {
7748 				anon_array_enter(amp, i_idx, &i_cookie);
7749 				i_ap = anon_get_ptr(amp->ahp, i_idx);
7750 				if (i_ap == NULL) {
7751 					unlocked_bytes += PAGESIZE;
7752 					anon_array_exit(&i_cookie);
7753 					continue;
7754 				}
7755 				swap_xlate(i_ap, &i_vp, &i_off);
7756 				anon_array_exit(&i_cookie);
7757 				pp = page_lookup(i_vp, i_off, SE_SHARED);
7758 				if (pp == NULL) {
7759 					unlocked_bytes += PAGESIZE;
7760 					continue;
7761 				} else if (pp->p_lckcnt == 0)
7762 					unlocked_bytes += PAGESIZE;
7763 				page_unlock(pp);
7764 			}
7765 			ANON_LOCK_EXIT(&amp->a_rwlock);
7766 		}
7767 
7768 		mutex_enter(&p->p_lock);
7769 		err = rctl_incr_locked_mem(p, proj, unlocked_bytes,
7770 		    chargeproc);
7771 		mutex_exit(&p->p_lock);
7772 
7773 		if (err) {
7774 			if (sp != NULL)
7775 				mutex_exit(&sp->shm_mlock);
7776 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7777 			return (err);
7778 		}
7779 	}
7780 	/*
7781 	 * Loop over all pages in the range.  Process if we're locking and
7782 	 * page has not already been locked in this mapping; or if we're
7783 	 * unlocking and the page has been locked.
7784 	 */
7785 	for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7786 	    vpp++, pos++, addr += PAGESIZE, offset += PAGESIZE, anon_index++) {
7787 		if ((attr == 0 || VPP_PROT(vpp) == pageprot) &&
7788 		    ((op == MC_LOCK && !VPP_ISPPLOCK(vpp)) ||
7789 		    (op == MC_UNLOCK && VPP_ISPPLOCK(vpp)))) {
7790 
7791 			if (amp != NULL)
7792 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7793 			/*
7794 			 * If this isn't a MAP_NORESERVE segment and
7795 			 * we're locking, allocate anon slots if they
7796 			 * don't exist.  The page is brought in later on.
7797 			 */
7798 			if (op == MC_LOCK && svd->vp == NULL &&
7799 			    ((svd->flags & MAP_NORESERVE) == 0) &&
7800 			    amp != NULL &&
7801 			    ((ap = anon_get_ptr(amp->ahp, anon_index))
7802 			    == NULL)) {
7803 				anon_array_enter(amp, anon_index, &cookie);
7804 
7805 				if ((ap = anon_get_ptr(amp->ahp,
7806 				    anon_index)) == NULL) {
7807 					pp = anon_zero(seg, addr, &ap,
7808 					    svd->cred);
7809 					if (pp == NULL) {
7810 						anon_array_exit(&cookie);
7811 						ANON_LOCK_EXIT(&amp->a_rwlock);
7812 						err = ENOMEM;
7813 						goto out;
7814 					}
7815 					ASSERT(anon_get_ptr(amp->ahp,
7816 					    anon_index) == NULL);
7817 					(void) anon_set_ptr(amp->ahp,
7818 					    anon_index, ap, ANON_SLEEP);
7819 					page_unlock(pp);
7820 				}
7821 				anon_array_exit(&cookie);
7822 			}
7823 
7824 			/*
7825 			 * Get name for page, accounting for
7826 			 * existence of private copy.
7827 			 */
7828 			ap = NULL;
7829 			if (amp != NULL) {
7830 				anon_array_enter(amp, anon_index, &cookie);
7831 				ap = anon_get_ptr(amp->ahp, anon_index);
7832 				if (ap != NULL) {
7833 					swap_xlate(ap, &vp, &off);
7834 				} else {
7835 					if (svd->vp == NULL &&
7836 					    (svd->flags & MAP_NORESERVE)) {
7837 						anon_array_exit(&cookie);
7838 						ANON_LOCK_EXIT(&amp->a_rwlock);
7839 						continue;
7840 					}
7841 					vp = svd->vp;
7842 					off = offset;
7843 				}
7844 				if (op != MC_LOCK || ap == NULL) {
7845 					anon_array_exit(&cookie);
7846 					ANON_LOCK_EXIT(&amp->a_rwlock);
7847 				}
7848 			} else {
7849 				vp = svd->vp;
7850 				off = offset;
7851 			}
7852 
7853 			/*
7854 			 * Get page frame.  It's ok if the page is
7855 			 * not available when we're unlocking, as this
7856 			 * may simply mean that a page we locked got
7857 			 * truncated out of existence after we locked it.
7858 			 *
7859 			 * Invoke VOP_GETPAGE() to obtain the page struct
7860 			 * since we may need to read it from disk if its
7861 			 * been paged out.
7862 			 */
7863 			if (op != MC_LOCK)
7864 				pp = page_lookup(vp, off, SE_SHARED);
7865 			else {
7866 				page_t *pl[1 + 1];
7867 				int error;
7868 
7869 				ASSERT(vp != NULL);
7870 
7871 				error = VOP_GETPAGE(vp, (offset_t)off, PAGESIZE,
7872 				    (uint_t *)NULL, pl, PAGESIZE, seg, addr,
7873 				    S_OTHER, svd->cred, NULL);
7874 
7875 				if (error && ap != NULL) {
7876 					anon_array_exit(&cookie);
7877 					ANON_LOCK_EXIT(&amp->a_rwlock);
7878 				}
7879 
7880 				/*
7881 				 * If the error is EDEADLK then we must bounce
7882 				 * up and drop all vm subsystem locks and then
7883 				 * retry the operation later
7884 				 * This behavior is a temporary measure because
7885 				 * ufs/sds logging is badly designed and will
7886 				 * deadlock if we don't allow this bounce to
7887 				 * happen.  The real solution is to re-design
7888 				 * the logging code to work properly.  See bug
7889 				 * 4125102 for details of the problem.
7890 				 */
7891 				if (error == EDEADLK) {
7892 					err = error;
7893 					goto out;
7894 				}
7895 				/*
7896 				 * Quit if we fail to fault in the page.  Treat
7897 				 * the failure as an error, unless the addr
7898 				 * is mapped beyond the end of a file.
7899 				 */
7900 				if (error && svd->vp) {
7901 					va.va_mask = AT_SIZE;
7902 					if (VOP_GETATTR(svd->vp, &va, 0,
7903 					    svd->cred, NULL) != 0) {
7904 						err = EIO;
7905 						goto out;
7906 					}
7907 					if (btopr(va.va_size) >=
7908 					    btopr(off + 1)) {
7909 						err = EIO;
7910 						goto out;
7911 					}
7912 					goto out;
7913 
7914 				} else if (error) {
7915 					err = EIO;
7916 					goto out;
7917 				}
7918 				pp = pl[0];
7919 				ASSERT(pp != NULL);
7920 			}
7921 
7922 			/*
7923 			 * See Statement at the beginning of this routine.
7924 			 *
7925 			 * claim is always set if MAP_PRIVATE and PROT_WRITE
7926 			 * irrespective of following factors:
7927 			 *
7928 			 * (1) anon slots are populated or not
7929 			 * (2) cow is broken or not
7930 			 * (3) refcnt on ap is 1 or greater than 1
7931 			 *
7932 			 * See 4140683 for details
7933 			 */
7934 			claim = ((VPP_PROT(vpp) & PROT_WRITE) &&
7935 			    (svd->type == MAP_PRIVATE));
7936 
7937 			/*
7938 			 * Perform page-level operation appropriate to
7939 			 * operation.  If locking, undo the SOFTLOCK
7940 			 * performed to bring the page into memory
7941 			 * after setting the lock.  If unlocking,
7942 			 * and no page was found, account for the claim
7943 			 * separately.
7944 			 */
7945 			if (op == MC_LOCK) {
7946 				int ret = 1;	/* Assume success */
7947 
7948 				ASSERT(!VPP_ISPPLOCK(vpp));
7949 
7950 				ret = page_pp_lock(pp, claim, 0);
7951 				if (ap != NULL) {
7952 					if (ap->an_pvp != NULL) {
7953 						anon_swap_free(ap, pp);
7954 					}
7955 					anon_array_exit(&cookie);
7956 					ANON_LOCK_EXIT(&amp->a_rwlock);
7957 				}
7958 				if (ret == 0) {
7959 					/* locking page failed */
7960 					page_unlock(pp);
7961 					err = EAGAIN;
7962 					goto out;
7963 				}
7964 				VPP_SETPPLOCK(vpp);
7965 				if (sp != NULL) {
7966 					if (pp->p_lckcnt == 1)
7967 						locked_bytes += PAGESIZE;
7968 				} else
7969 					locked_bytes += PAGESIZE;
7970 
7971 				if (lockmap != (ulong_t *)NULL)
7972 					BT_SET(lockmap, pos);
7973 
7974 				page_unlock(pp);
7975 			} else {
7976 				ASSERT(VPP_ISPPLOCK(vpp));
7977 				if (pp != NULL) {
7978 					/* sysV pages should be locked */
7979 					ASSERT(sp == NULL || pp->p_lckcnt > 0);
7980 					page_pp_unlock(pp, claim, 0);
7981 					if (sp != NULL) {
7982 						if (pp->p_lckcnt == 0)
7983 							unlocked_bytes
7984 							    += PAGESIZE;
7985 					} else
7986 						unlocked_bytes += PAGESIZE;
7987 					page_unlock(pp);
7988 				} else {
7989 					ASSERT(sp == NULL);
7990 					unlocked_bytes += PAGESIZE;
7991 				}
7992 				VPP_CLRPPLOCK(vpp);
7993 			}
7994 		}
7995 	}
7996 out:
7997 	if (op == MC_LOCK) {
7998 		/* Credit back bytes that did not get locked */
7999 		if ((unlocked_bytes - locked_bytes) > 0) {
8000 			if (proj == NULL)
8001 				mutex_enter(&p->p_lock);
8002 			rctl_decr_locked_mem(p, proj,
8003 			    (unlocked_bytes - locked_bytes), chargeproc);
8004 			if (proj == NULL)
8005 				mutex_exit(&p->p_lock);
8006 		}
8007 
8008 	} else {
8009 		/* Account bytes that were unlocked */
8010 		if (unlocked_bytes > 0) {
8011 			if (proj == NULL)
8012 				mutex_enter(&p->p_lock);
8013 			rctl_decr_locked_mem(p, proj, unlocked_bytes,
8014 			    chargeproc);
8015 			if (proj == NULL)
8016 				mutex_exit(&p->p_lock);
8017 		}
8018 	}
8019 	if (sp != NULL)
8020 		mutex_exit(&sp->shm_mlock);
8021 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8022 
8023 	return (err);
8024 }
8025 
8026 /*
8027  * Set advice from user for specified pages
8028  * There are 5 types of advice:
8029  *	MADV_NORMAL	- Normal (default) behavior (whatever that is)
8030  *	MADV_RANDOM	- Random page references
8031  *				do not allow readahead or 'klustering'
8032  *	MADV_SEQUENTIAL	- Sequential page references
8033  *				Pages previous to the one currently being
8034  *				accessed (determined by fault) are 'not needed'
8035  *				and are freed immediately
8036  *	MADV_WILLNEED	- Pages are likely to be used (fault ahead in mctl)
8037  *	MADV_DONTNEED	- Pages are not needed (synced out in mctl)
8038  *	MADV_FREE	- Contents can be discarded
8039  *	MADV_ACCESS_DEFAULT- Default access
8040  *	MADV_ACCESS_LWP	- Next LWP will access heavily
8041  *	MADV_ACCESS_MANY- Many LWPs or processes will access heavily
8042  */
8043 static int
8044 segvn_advise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
8045 {
8046 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8047 	size_t page;
8048 	int err = 0;
8049 	int already_set;
8050 	struct anon_map *amp;
8051 	ulong_t anon_index;
8052 	struct seg *next;
8053 	lgrp_mem_policy_t policy;
8054 	struct seg *prev;
8055 	struct vnode *vp;
8056 
8057 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
8058 
8059 	/*
8060 	 * In case of MADV_FREE, we won't be modifying any segment private
8061 	 * data structures; so, we only need to grab READER's lock
8062 	 */
8063 	if (behav != MADV_FREE) {
8064 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
8065 		if (svd->tr_state != SEGVN_TR_OFF) {
8066 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8067 			return (0);
8068 		}
8069 	} else {
8070 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8071 	}
8072 
8073 	/*
8074 	 * Large pages are assumed to be only turned on when accesses to the
8075 	 * segment's address range have spatial and temporal locality. That
8076 	 * justifies ignoring MADV_SEQUENTIAL for large page segments.
8077 	 * Also, ignore advice affecting lgroup memory allocation
8078 	 * if don't need to do lgroup optimizations on this system
8079 	 */
8080 
8081 	if ((behav == MADV_SEQUENTIAL &&
8082 	    (seg->s_szc != 0 || HAT_IS_REGION_COOKIE_VALID(svd->rcookie))) ||
8083 	    (!lgrp_optimizations() && (behav == MADV_ACCESS_DEFAULT ||
8084 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY))) {
8085 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8086 		return (0);
8087 	}
8088 
8089 	if (behav == MADV_SEQUENTIAL || behav == MADV_ACCESS_DEFAULT ||
8090 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY) {
8091 		/*
8092 		 * Since we are going to unload hat mappings
8093 		 * we first have to flush the cache. Otherwise
8094 		 * this might lead to system panic if another
8095 		 * thread is doing physio on the range whose
8096 		 * mappings are unloaded by madvise(3C).
8097 		 */
8098 		if (svd->softlockcnt > 0) {
8099 			/*
8100 			 * If this is shared segment non 0 softlockcnt
8101 			 * means locked pages are still in use.
8102 			 */
8103 			if (svd->type == MAP_SHARED) {
8104 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8105 				return (EAGAIN);
8106 			}
8107 			/*
8108 			 * Since we do have the segvn writers lock
8109 			 * nobody can fill the cache with entries
8110 			 * belonging to this seg during the purge.
8111 			 * The flush either succeeds or we still
8112 			 * have pending I/Os. In the later case,
8113 			 * madvise(3C) fails.
8114 			 */
8115 			segvn_purge(seg);
8116 			if (svd->softlockcnt > 0) {
8117 				/*
8118 				 * Since madvise(3C) is advisory and
8119 				 * it's not part of UNIX98, madvise(3C)
8120 				 * failure here doesn't cause any hardship.
8121 				 * Note that we don't block in "as" layer.
8122 				 */
8123 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8124 				return (EAGAIN);
8125 			}
8126 		} else if (svd->type == MAP_SHARED && svd->amp != NULL &&
8127 		    svd->amp->a_softlockcnt > 0) {
8128 			/*
8129 			 * Try to purge this amp's entries from pcache. It
8130 			 * will succeed only if other segments that share the
8131 			 * amp have no outstanding softlock's.
8132 			 */
8133 			segvn_purge(seg);
8134 		}
8135 	}
8136 
8137 	amp = svd->amp;
8138 	vp = svd->vp;
8139 	if (behav == MADV_FREE) {
8140 		/*
8141 		 * MADV_FREE is not supported for segments with
8142 		 * underlying object; if anonmap is NULL, anon slots
8143 		 * are not yet populated and there is nothing for
8144 		 * us to do. As MADV_FREE is advisory, we don't
8145 		 * return error in either case.
8146 		 */
8147 		if (vp != NULL || amp == NULL) {
8148 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8149 			return (0);
8150 		}
8151 
8152 		segvn_purge(seg);
8153 
8154 		page = seg_page(seg, addr);
8155 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8156 		anon_disclaim(amp, svd->anon_index + page, len);
8157 		ANON_LOCK_EXIT(&amp->a_rwlock);
8158 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8159 		return (0);
8160 	}
8161 
8162 	/*
8163 	 * If advice is to be applied to entire segment,
8164 	 * use advice field in seg_data structure
8165 	 * otherwise use appropriate vpage entry.
8166 	 */
8167 	if ((addr == seg->s_base) && (len == seg->s_size)) {
8168 		switch (behav) {
8169 		case MADV_ACCESS_LWP:
8170 		case MADV_ACCESS_MANY:
8171 		case MADV_ACCESS_DEFAULT:
8172 			/*
8173 			 * Set memory allocation policy for this segment
8174 			 */
8175 			policy = lgrp_madv_to_policy(behav, len, svd->type);
8176 			if (svd->type == MAP_SHARED)
8177 				already_set = lgrp_shm_policy_set(policy, amp,
8178 				    svd->anon_index, vp, svd->offset, len);
8179 			else {
8180 				/*
8181 				 * For private memory, need writers lock on
8182 				 * address space because the segment may be
8183 				 * split or concatenated when changing policy
8184 				 */
8185 				if (AS_READ_HELD(seg->s_as,
8186 				    &seg->s_as->a_lock)) {
8187 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8188 					return (IE_RETRY);
8189 				}
8190 
8191 				already_set = lgrp_privm_policy_set(policy,
8192 				    &svd->policy_info, len);
8193 			}
8194 
8195 			/*
8196 			 * If policy set already and it shouldn't be reapplied,
8197 			 * don't do anything.
8198 			 */
8199 			if (already_set &&
8200 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8201 				break;
8202 
8203 			/*
8204 			 * Mark any existing pages in given range for
8205 			 * migration
8206 			 */
8207 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8208 			    vp, svd->offset, 1);
8209 
8210 			/*
8211 			 * If same policy set already or this is a shared
8212 			 * memory segment, don't need to try to concatenate
8213 			 * segment with adjacent ones.
8214 			 */
8215 			if (already_set || svd->type == MAP_SHARED)
8216 				break;
8217 
8218 			/*
8219 			 * Try to concatenate this segment with previous
8220 			 * one and next one, since we changed policy for
8221 			 * this one and it may be compatible with adjacent
8222 			 * ones now.
8223 			 */
8224 			prev = AS_SEGPREV(seg->s_as, seg);
8225 			next = AS_SEGNEXT(seg->s_as, seg);
8226 
8227 			if (next && next->s_ops == &segvn_ops &&
8228 			    addr + len == next->s_base)
8229 				(void) segvn_concat(seg, next, 1);
8230 
8231 			if (prev && prev->s_ops == &segvn_ops &&
8232 			    addr == prev->s_base + prev->s_size) {
8233 				/*
8234 				 * Drop lock for private data of current
8235 				 * segment before concatenating (deleting) it
8236 				 * and return IE_REATTACH to tell as_ctl() that
8237 				 * current segment has changed
8238 				 */
8239 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8240 				if (!segvn_concat(prev, seg, 1))
8241 					err = IE_REATTACH;
8242 
8243 				return (err);
8244 			}
8245 			break;
8246 
8247 		case MADV_SEQUENTIAL:
8248 			/*
8249 			 * unloading mapping guarantees
8250 			 * detection in segvn_fault
8251 			 */
8252 			ASSERT(seg->s_szc == 0);
8253 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8254 			hat_unload(seg->s_as->a_hat, addr, len,
8255 			    HAT_UNLOAD);
8256 			/* FALLTHROUGH */
8257 		case MADV_NORMAL:
8258 		case MADV_RANDOM:
8259 			svd->advice = (uchar_t)behav;
8260 			svd->pageadvice = 0;
8261 			break;
8262 		case MADV_WILLNEED:	/* handled in memcntl */
8263 		case MADV_DONTNEED:	/* handled in memcntl */
8264 		case MADV_FREE:		/* handled above */
8265 			break;
8266 		default:
8267 			err = EINVAL;
8268 		}
8269 	} else {
8270 		caddr_t			eaddr;
8271 		struct seg		*new_seg;
8272 		struct segvn_data	*new_svd;
8273 		u_offset_t		off;
8274 		caddr_t			oldeaddr;
8275 
8276 		page = seg_page(seg, addr);
8277 
8278 		segvn_vpage(seg);
8279 		if (svd->vpage == NULL) {
8280 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8281 			return (ENOMEM);
8282 		}
8283 
8284 		switch (behav) {
8285 			struct vpage *bvpp, *evpp;
8286 
8287 		case MADV_ACCESS_LWP:
8288 		case MADV_ACCESS_MANY:
8289 		case MADV_ACCESS_DEFAULT:
8290 			/*
8291 			 * Set memory allocation policy for portion of this
8292 			 * segment
8293 			 */
8294 
8295 			/*
8296 			 * Align address and length of advice to page
8297 			 * boundaries for large pages
8298 			 */
8299 			if (seg->s_szc != 0) {
8300 				size_t	pgsz;
8301 
8302 				pgsz = page_get_pagesize(seg->s_szc);
8303 				addr = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
8304 				len = P2ROUNDUP(len, pgsz);
8305 			}
8306 
8307 			/*
8308 			 * Check to see whether policy is set already
8309 			 */
8310 			policy = lgrp_madv_to_policy(behav, len, svd->type);
8311 
8312 			anon_index = svd->anon_index + page;
8313 			off = svd->offset + (uintptr_t)(addr - seg->s_base);
8314 
8315 			if (svd->type == MAP_SHARED)
8316 				already_set = lgrp_shm_policy_set(policy, amp,
8317 				    anon_index, vp, off, len);
8318 			else
8319 				already_set =
8320 				    (policy == svd->policy_info.mem_policy);
8321 
8322 			/*
8323 			 * If policy set already and it shouldn't be reapplied,
8324 			 * don't do anything.
8325 			 */
8326 			if (already_set &&
8327 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8328 				break;
8329 
8330 			/*
8331 			 * For private memory, need writers lock on
8332 			 * address space because the segment may be
8333 			 * split or concatenated when changing policy
8334 			 */
8335 			if (svd->type == MAP_PRIVATE &&
8336 			    AS_READ_HELD(seg->s_as, &seg->s_as->a_lock)) {
8337 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8338 				return (IE_RETRY);
8339 			}
8340 
8341 			/*
8342 			 * Mark any existing pages in given range for
8343 			 * migration
8344 			 */
8345 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8346 			    vp, svd->offset, 1);
8347 
8348 			/*
8349 			 * Don't need to try to split or concatenate
8350 			 * segments, since policy is same or this is a shared
8351 			 * memory segment
8352 			 */
8353 			if (already_set || svd->type == MAP_SHARED)
8354 				break;
8355 
8356 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
8357 				ASSERT(svd->amp == NULL);
8358 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
8359 				ASSERT(svd->softlockcnt == 0);
8360 				hat_leave_region(seg->s_as->a_hat, svd->rcookie,
8361 				    HAT_REGION_TEXT);
8362 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
8363 			}
8364 
8365 			/*
8366 			 * Split off new segment if advice only applies to a
8367 			 * portion of existing segment starting in middle
8368 			 */
8369 			new_seg = NULL;
8370 			eaddr = addr + len;
8371 			oldeaddr = seg->s_base + seg->s_size;
8372 			if (addr > seg->s_base) {
8373 				/*
8374 				 * Must flush I/O page cache
8375 				 * before splitting segment
8376 				 */
8377 				if (svd->softlockcnt > 0)
8378 					segvn_purge(seg);
8379 
8380 				/*
8381 				 * Split segment and return IE_REATTACH to tell
8382 				 * as_ctl() that current segment changed
8383 				 */
8384 				new_seg = segvn_split_seg(seg, addr);
8385 				new_svd = (struct segvn_data *)new_seg->s_data;
8386 				err = IE_REATTACH;
8387 
8388 				/*
8389 				 * If new segment ends where old one
8390 				 * did, try to concatenate the new
8391 				 * segment with next one.
8392 				 */
8393 				if (eaddr == oldeaddr) {
8394 					/*
8395 					 * Set policy for new segment
8396 					 */
8397 					(void) lgrp_privm_policy_set(policy,
8398 					    &new_svd->policy_info,
8399 					    new_seg->s_size);
8400 
8401 					next = AS_SEGNEXT(new_seg->s_as,
8402 					    new_seg);
8403 
8404 					if (next &&
8405 					    next->s_ops == &segvn_ops &&
8406 					    eaddr == next->s_base)
8407 						(void) segvn_concat(new_seg,
8408 						    next, 1);
8409 				}
8410 			}
8411 
8412 			/*
8413 			 * Split off end of existing segment if advice only
8414 			 * applies to a portion of segment ending before
8415 			 * end of the existing segment
8416 			 */
8417 			if (eaddr < oldeaddr) {
8418 				/*
8419 				 * Must flush I/O page cache
8420 				 * before splitting segment
8421 				 */
8422 				if (svd->softlockcnt > 0)
8423 					segvn_purge(seg);
8424 
8425 				/*
8426 				 * If beginning of old segment was already
8427 				 * split off, use new segment to split end off
8428 				 * from.
8429 				 */
8430 				if (new_seg != NULL && new_seg != seg) {
8431 					/*
8432 					 * Split segment
8433 					 */
8434 					(void) segvn_split_seg(new_seg, eaddr);
8435 
8436 					/*
8437 					 * Set policy for new segment
8438 					 */
8439 					(void) lgrp_privm_policy_set(policy,
8440 					    &new_svd->policy_info,
8441 					    new_seg->s_size);
8442 				} else {
8443 					/*
8444 					 * Split segment and return IE_REATTACH
8445 					 * to tell as_ctl() that current
8446 					 * segment changed
8447 					 */
8448 					(void) segvn_split_seg(seg, eaddr);
8449 					err = IE_REATTACH;
8450 
8451 					(void) lgrp_privm_policy_set(policy,
8452 					    &svd->policy_info, seg->s_size);
8453 
8454 					/*
8455 					 * If new segment starts where old one
8456 					 * did, try to concatenate it with
8457 					 * previous segment.
8458 					 */
8459 					if (addr == seg->s_base) {
8460 						prev = AS_SEGPREV(seg->s_as,
8461 						    seg);
8462 
8463 						/*
8464 						 * Drop lock for private data
8465 						 * of current segment before
8466 						 * concatenating (deleting) it
8467 						 */
8468 						if (prev &&
8469 						    prev->s_ops ==
8470 						    &segvn_ops &&
8471 						    addr == prev->s_base +
8472 						    prev->s_size) {
8473 							SEGVN_LOCK_EXIT(
8474 							    seg->s_as,
8475 							    &svd->lock);
8476 							(void) segvn_concat(
8477 							    prev, seg, 1);
8478 							return (err);
8479 						}
8480 					}
8481 				}
8482 			}
8483 			break;
8484 		case MADV_SEQUENTIAL:
8485 			ASSERT(seg->s_szc == 0);
8486 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8487 			hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
8488 			/* FALLTHROUGH */
8489 		case MADV_NORMAL:
8490 		case MADV_RANDOM:
8491 			bvpp = &svd->vpage[page];
8492 			evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8493 			for (; bvpp < evpp; bvpp++)
8494 				VPP_SETADVICE(bvpp, behav);
8495 			svd->advice = MADV_NORMAL;
8496 			break;
8497 		case MADV_WILLNEED:	/* handled in memcntl */
8498 		case MADV_DONTNEED:	/* handled in memcntl */
8499 		case MADV_FREE:		/* handled above */
8500 			break;
8501 		default:
8502 			err = EINVAL;
8503 		}
8504 	}
8505 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8506 	return (err);
8507 }
8508 
8509 /*
8510  * Create a vpage structure for this seg.
8511  */
8512 static void
8513 segvn_vpage(struct seg *seg)
8514 {
8515 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8516 	struct vpage *vp, *evp;
8517 	static pgcnt_t page_limit = 0;
8518 
8519 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
8520 
8521 	/*
8522 	 * If no vpage structure exists, allocate one.  Copy the protections
8523 	 * and the advice from the segment itself to the individual pages.
8524 	 */
8525 	if (svd->vpage == NULL) {
8526 		/*
8527 		 * Start by calculating the number of pages we must allocate to
8528 		 * track the per-page vpage structs needs for this entire
8529 		 * segment. If we know now that it will require more than our
8530 		 * heuristic for the maximum amount of kmem we can consume then
8531 		 * fail. We do this here, instead of trying to detect this deep
8532 		 * in page_resv and propagating the error up, since the entire
8533 		 * memory allocation stack is not amenable to passing this
8534 		 * back. Instead, it wants to keep trying.
8535 		 *
8536 		 * As a heuristic we set a page limit of 5/8s of total_pages
8537 		 * for this allocation. We use shifts so that no floating
8538 		 * point conversion takes place and only need to do the
8539 		 * calculation once.
8540 		 */
8541 		ulong_t mem_needed = seg_pages(seg) * sizeof (struct vpage);
8542 		pgcnt_t npages = mem_needed >> PAGESHIFT;
8543 
8544 		if (page_limit == 0)
8545 			page_limit = (total_pages >> 1) + (total_pages >> 3);
8546 
8547 		if (npages > page_limit)
8548 			return;
8549 
8550 		svd->pageadvice = 1;
8551 		svd->vpage = kmem_zalloc(mem_needed, KM_SLEEP);
8552 		evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
8553 		for (vp = svd->vpage; vp < evp; vp++) {
8554 			VPP_SETPROT(vp, svd->prot);
8555 			VPP_SETADVICE(vp, svd->advice);
8556 		}
8557 	}
8558 }
8559 
8560 /*
8561  * Dump the pages belonging to this segvn segment.
8562  */
8563 static void
8564 segvn_dump(struct seg *seg)
8565 {
8566 	struct segvn_data *svd;
8567 	page_t *pp;
8568 	struct anon_map *amp;
8569 	ulong_t	anon_index;
8570 	struct vnode *vp;
8571 	u_offset_t off, offset;
8572 	pfn_t pfn;
8573 	pgcnt_t page, npages;
8574 	caddr_t addr;
8575 
8576 	npages = seg_pages(seg);
8577 	svd = (struct segvn_data *)seg->s_data;
8578 	vp = svd->vp;
8579 	off = offset = svd->offset;
8580 	addr = seg->s_base;
8581 
8582 	if ((amp = svd->amp) != NULL) {
8583 		anon_index = svd->anon_index;
8584 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8585 	}
8586 
8587 	for (page = 0; page < npages; page++, offset += PAGESIZE) {
8588 		struct anon *ap;
8589 		int we_own_it = 0;
8590 
8591 		if (amp && (ap = anon_get_ptr(svd->amp->ahp, anon_index++))) {
8592 			swap_xlate_nopanic(ap, &vp, &off);
8593 		} else {
8594 			vp = svd->vp;
8595 			off = offset;
8596 		}
8597 
8598 		/*
8599 		 * If pp == NULL, the page either does not exist
8600 		 * or is exclusively locked.  So determine if it
8601 		 * exists before searching for it.
8602 		 */
8603 
8604 		if ((pp = page_lookup_nowait(vp, off, SE_SHARED)))
8605 			we_own_it = 1;
8606 		else
8607 			pp = page_exists(vp, off);
8608 
8609 		if (pp) {
8610 			pfn = page_pptonum(pp);
8611 			dump_addpage(seg->s_as, addr, pfn);
8612 			if (we_own_it)
8613 				page_unlock(pp);
8614 		}
8615 		addr += PAGESIZE;
8616 		dump_timeleft = dump_timeout;
8617 	}
8618 
8619 	if (amp != NULL)
8620 		ANON_LOCK_EXIT(&amp->a_rwlock);
8621 }
8622 
8623 #ifdef DEBUG
8624 static uint32_t segvn_pglock_mtbf = 0;
8625 #endif
8626 
8627 #define	PCACHE_SHWLIST		((page_t *)-2)
8628 #define	NOPCACHE_SHWLIST	((page_t *)-1)
8629 
8630 /*
8631  * Lock/Unlock anon pages over a given range. Return shadow list. This routine
8632  * uses global segment pcache to cache shadow lists (i.e. pp arrays) of pages
8633  * to avoid the overhead of per page locking, unlocking for subsequent IOs to
8634  * the same parts of the segment. Currently shadow list creation is only
8635  * supported for pure anon segments. MAP_PRIVATE segment pcache entries are
8636  * tagged with segment pointer, starting virtual address and length. This
8637  * approach for MAP_SHARED segments may add many pcache entries for the same
8638  * set of pages and lead to long hash chains that decrease pcache lookup
8639  * performance. To avoid this issue for shared segments shared anon map and
8640  * starting anon index are used for pcache entry tagging. This allows all
8641  * segments to share pcache entries for the same anon range and reduces pcache
8642  * chain's length as well as memory overhead from duplicate shadow lists and
8643  * pcache entries.
8644  *
8645  * softlockcnt field in segvn_data structure counts the number of F_SOFTLOCK'd
8646  * pages via segvn_fault() and pagelock'd pages via this routine. But pagelock
8647  * part of softlockcnt accounting is done differently for private and shared
8648  * segments. In private segment case softlock is only incremented when a new
8649  * shadow list is created but not when an existing one is found via
8650  * seg_plookup(). pcache entries have reference count incremented/decremented
8651  * by each seg_plookup()/seg_pinactive() operation. Only entries that have 0
8652  * reference count can be purged (and purging is needed before segment can be
8653  * freed). When a private segment pcache entry is purged segvn_reclaim() will
8654  * decrement softlockcnt. Since in private segment case each of its pcache
8655  * entries only belongs to this segment we can expect that when
8656  * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8657  * segment purge will succeed and softlockcnt will drop to 0. In shared
8658  * segment case reference count in pcache entry counts active locks from many
8659  * different segments so we can't expect segment purging to succeed even when
8660  * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8661  * segment. To be able to determine when there're no pending pagelocks in
8662  * shared segment case we don't rely on purging to make softlockcnt drop to 0
8663  * but instead softlockcnt is incremented and decremented for every
8664  * segvn_pagelock(L_PAGELOCK/L_PAGEUNLOCK) call regardless if a new shadow
8665  * list was created or an existing one was found. When softlockcnt drops to 0
8666  * this segment no longer has any claims for pcached shadow lists and the
8667  * segment can be freed even if there're still active pcache entries
8668  * shared by this segment anon map. Shared segment pcache entries belong to
8669  * anon map and are typically removed when anon map is freed after all
8670  * processes destroy the segments that use this anon map.
8671  */
8672 static int
8673 segvn_pagelock(struct seg *seg, caddr_t addr, size_t len, struct page ***ppp,
8674     enum lock_type type, enum seg_rw rw)
8675 {
8676 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8677 	size_t np;
8678 	pgcnt_t adjustpages;
8679 	pgcnt_t npages;
8680 	ulong_t anon_index;
8681 	uint_t protchk = (rw == S_READ) ? PROT_READ : PROT_WRITE;
8682 	uint_t error;
8683 	struct anon_map *amp;
8684 	pgcnt_t anpgcnt;
8685 	struct page **pplist, **pl, *pp;
8686 	caddr_t a;
8687 	size_t page;
8688 	caddr_t lpgaddr, lpgeaddr;
8689 	anon_sync_obj_t cookie;
8690 	int anlock;
8691 	struct anon_map *pamp;
8692 	caddr_t paddr;
8693 	seg_preclaim_cbfunc_t preclaim_callback;
8694 	size_t pgsz;
8695 	int use_pcache;
8696 	size_t wlen;
8697 	uint_t pflags = 0;
8698 	int sftlck_sbase = 0;
8699 	int sftlck_send = 0;
8700 
8701 #ifdef DEBUG
8702 	if (type == L_PAGELOCK && segvn_pglock_mtbf) {
8703 		hrtime_t ts = gethrtime();
8704 		if ((ts % segvn_pglock_mtbf) == 0) {
8705 			return (ENOTSUP);
8706 		}
8707 		if ((ts % segvn_pglock_mtbf) == 1) {
8708 			return (EFAULT);
8709 		}
8710 	}
8711 #endif
8712 
8713 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_START,
8714 	    "segvn_pagelock: start seg %p addr %p", seg, addr);
8715 
8716 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
8717 	ASSERT(type == L_PAGELOCK || type == L_PAGEUNLOCK);
8718 
8719 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8720 
8721 	/*
8722 	 * for now we only support pagelock to anon memory. We would have to
8723 	 * check protections for vnode objects and call into the vnode driver.
8724 	 * That's too much for a fast path. Let the fault entry point handle
8725 	 * it.
8726 	 */
8727 	if (svd->vp != NULL) {
8728 		if (type == L_PAGELOCK) {
8729 			error = ENOTSUP;
8730 			goto out;
8731 		}
8732 		panic("segvn_pagelock(L_PAGEUNLOCK): vp != NULL");
8733 	}
8734 	if ((amp = svd->amp) == NULL) {
8735 		if (type == L_PAGELOCK) {
8736 			error = EFAULT;
8737 			goto out;
8738 		}
8739 		panic("segvn_pagelock(L_PAGEUNLOCK): amp == NULL");
8740 	}
8741 	if (rw != S_READ && rw != S_WRITE) {
8742 		if (type == L_PAGELOCK) {
8743 			error = ENOTSUP;
8744 			goto out;
8745 		}
8746 		panic("segvn_pagelock(L_PAGEUNLOCK): bad rw");
8747 	}
8748 
8749 	if (seg->s_szc != 0) {
8750 		/*
8751 		 * We are adjusting the pagelock region to the large page size
8752 		 * boundary because the unlocked part of a large page cannot
8753 		 * be freed anyway unless all constituent pages of a large
8754 		 * page are locked. Bigger regions reduce pcache chain length
8755 		 * and improve lookup performance. The tradeoff is that the
8756 		 * very first segvn_pagelock() call for a given page is more
8757 		 * expensive if only 1 page_t is needed for IO. This is only
8758 		 * an issue if pcache entry doesn't get reused by several
8759 		 * subsequent calls. We optimize here for the case when pcache
8760 		 * is heavily used by repeated IOs to the same address range.
8761 		 *
8762 		 * Note segment's page size cannot change while we are holding
8763 		 * as lock.  And then it cannot change while softlockcnt is
8764 		 * not 0. This will allow us to correctly recalculate large
8765 		 * page size region for the matching pageunlock/reclaim call
8766 		 * since as_pageunlock() caller must always match
8767 		 * as_pagelock() call's addr and len.
8768 		 *
8769 		 * For pageunlock *ppp points to the pointer of page_t that
8770 		 * corresponds to the real unadjusted start address. Similar
8771 		 * for pagelock *ppp must point to the pointer of page_t that
8772 		 * corresponds to the real unadjusted start address.
8773 		 */
8774 		pgsz = page_get_pagesize(seg->s_szc);
8775 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
8776 		adjustpages = btop((uintptr_t)(addr - lpgaddr));
8777 	} else if (len < segvn_pglock_comb_thrshld) {
8778 		lpgaddr = addr;
8779 		lpgeaddr = addr + len;
8780 		adjustpages = 0;
8781 		pgsz = PAGESIZE;
8782 	} else {
8783 		/*
8784 		 * Align the address range of large enough requests to allow
8785 		 * combining of different shadow lists into 1 to reduce memory
8786 		 * overhead from potentially overlapping large shadow lists
8787 		 * (worst case is we have a 1MB IO into buffers with start
8788 		 * addresses separated by 4K).  Alignment is only possible if
8789 		 * padded chunks have sufficient access permissions. Note
8790 		 * permissions won't change between L_PAGELOCK and
8791 		 * L_PAGEUNLOCK calls since non 0 softlockcnt will force
8792 		 * segvn_setprot() to wait until softlockcnt drops to 0. This
8793 		 * allows us to determine in L_PAGEUNLOCK the same range we
8794 		 * computed in L_PAGELOCK.
8795 		 *
8796 		 * If alignment is limited by segment ends set
8797 		 * sftlck_sbase/sftlck_send flags. In L_PAGELOCK case when
8798 		 * these flags are set bump softlockcnt_sbase/softlockcnt_send
8799 		 * per segment counters. In L_PAGEUNLOCK case decrease
8800 		 * softlockcnt_sbase/softlockcnt_send counters if
8801 		 * sftlck_sbase/sftlck_send flags are set.  When
8802 		 * softlockcnt_sbase/softlockcnt_send are non 0
8803 		 * segvn_concat()/segvn_extend_prev()/segvn_extend_next()
8804 		 * won't merge the segments. This restriction combined with
8805 		 * restriction on segment unmapping and splitting for segments
8806 		 * that have non 0 softlockcnt allows L_PAGEUNLOCK to
8807 		 * correctly determine the same range that was previously
8808 		 * locked by matching L_PAGELOCK.
8809 		 */
8810 		pflags = SEGP_PSHIFT | (segvn_pglock_comb_bshift << 16);
8811 		pgsz = PAGESIZE;
8812 		if (svd->type == MAP_PRIVATE) {
8813 			lpgaddr = (caddr_t)P2ALIGN((uintptr_t)addr,
8814 			    segvn_pglock_comb_balign);
8815 			if (lpgaddr < seg->s_base) {
8816 				lpgaddr = seg->s_base;
8817 				sftlck_sbase = 1;
8818 			}
8819 		} else {
8820 			ulong_t aix = svd->anon_index + seg_page(seg, addr);
8821 			ulong_t aaix = P2ALIGN(aix, segvn_pglock_comb_palign);
8822 			if (aaix < svd->anon_index) {
8823 				lpgaddr = seg->s_base;
8824 				sftlck_sbase = 1;
8825 			} else {
8826 				lpgaddr = addr - ptob(aix - aaix);
8827 				ASSERT(lpgaddr >= seg->s_base);
8828 			}
8829 		}
8830 		if (svd->pageprot && lpgaddr != addr) {
8831 			struct vpage *vp = &svd->vpage[seg_page(seg, lpgaddr)];
8832 			struct vpage *evp = &svd->vpage[seg_page(seg, addr)];
8833 			while (vp < evp) {
8834 				if ((VPP_PROT(vp) & protchk) == 0) {
8835 					break;
8836 				}
8837 				vp++;
8838 			}
8839 			if (vp < evp) {
8840 				lpgaddr = addr;
8841 				pflags = 0;
8842 			}
8843 		}
8844 		lpgeaddr = addr + len;
8845 		if (pflags) {
8846 			if (svd->type == MAP_PRIVATE) {
8847 				lpgeaddr = (caddr_t)P2ROUNDUP(
8848 				    (uintptr_t)lpgeaddr,
8849 				    segvn_pglock_comb_balign);
8850 			} else {
8851 				ulong_t aix = svd->anon_index +
8852 				    seg_page(seg, lpgeaddr);
8853 				ulong_t aaix = P2ROUNDUP(aix,
8854 				    segvn_pglock_comb_palign);
8855 				if (aaix < aix) {
8856 					lpgeaddr = 0;
8857 				} else {
8858 					lpgeaddr += ptob(aaix - aix);
8859 				}
8860 			}
8861 			if (lpgeaddr == 0 ||
8862 			    lpgeaddr > seg->s_base + seg->s_size) {
8863 				lpgeaddr = seg->s_base + seg->s_size;
8864 				sftlck_send = 1;
8865 			}
8866 		}
8867 		if (svd->pageprot && lpgeaddr != addr + len) {
8868 			struct vpage *vp;
8869 			struct vpage *evp;
8870 
8871 			vp = &svd->vpage[seg_page(seg, addr + len)];
8872 			evp = &svd->vpage[seg_page(seg, lpgeaddr)];
8873 
8874 			while (vp < evp) {
8875 				if ((VPP_PROT(vp) & protchk) == 0) {
8876 					break;
8877 				}
8878 				vp++;
8879 			}
8880 			if (vp < evp) {
8881 				lpgeaddr = addr + len;
8882 			}
8883 		}
8884 		adjustpages = btop((uintptr_t)(addr - lpgaddr));
8885 	}
8886 
8887 	/*
8888 	 * For MAP_SHARED segments we create pcache entries tagged by amp and
8889 	 * anon index so that we can share pcache entries with other segments
8890 	 * that map this amp.  For private segments pcache entries are tagged
8891 	 * with segment and virtual address.
8892 	 */
8893 	if (svd->type == MAP_SHARED) {
8894 		pamp = amp;
8895 		paddr = (caddr_t)((lpgaddr - seg->s_base) +
8896 		    ptob(svd->anon_index));
8897 		preclaim_callback = shamp_reclaim;
8898 	} else {
8899 		pamp = NULL;
8900 		paddr = lpgaddr;
8901 		preclaim_callback = segvn_reclaim;
8902 	}
8903 
8904 	if (type == L_PAGEUNLOCK) {
8905 		VM_STAT_ADD(segvnvmstats.pagelock[0]);
8906 
8907 		/*
8908 		 * update hat ref bits for /proc. We need to make sure
8909 		 * that threads tracing the ref and mod bits of the
8910 		 * address space get the right data.
8911 		 * Note: page ref and mod bits are updated at reclaim time
8912 		 */
8913 		if (seg->s_as->a_vbits) {
8914 			for (a = addr; a < addr + len; a += PAGESIZE) {
8915 				if (rw == S_WRITE) {
8916 					hat_setstat(seg->s_as, a,
8917 					    PAGESIZE, P_REF | P_MOD);
8918 				} else {
8919 					hat_setstat(seg->s_as, a,
8920 					    PAGESIZE, P_REF);
8921 				}
8922 			}
8923 		}
8924 
8925 		/*
8926 		 * Check the shadow list entry after the last page used in
8927 		 * this IO request. If it's NOPCACHE_SHWLIST the shadow list
8928 		 * was not inserted into pcache and is not large page
8929 		 * adjusted.  In this case call reclaim callback directly and
8930 		 * don't adjust the shadow list start and size for large
8931 		 * pages.
8932 		 */
8933 		npages = btop(len);
8934 		if ((*ppp)[npages] == NOPCACHE_SHWLIST) {
8935 			void *ptag;
8936 			if (pamp != NULL) {
8937 				ASSERT(svd->type == MAP_SHARED);
8938 				ptag = (void *)pamp;
8939 				paddr = (caddr_t)((addr - seg->s_base) +
8940 				    ptob(svd->anon_index));
8941 			} else {
8942 				ptag = (void *)seg;
8943 				paddr = addr;
8944 			}
8945 			(*preclaim_callback)(ptag, paddr, len, *ppp, rw, 0);
8946 		} else {
8947 			ASSERT((*ppp)[npages] == PCACHE_SHWLIST ||
8948 			    IS_SWAPFSVP((*ppp)[npages]->p_vnode));
8949 			len = lpgeaddr - lpgaddr;
8950 			npages = btop(len);
8951 			seg_pinactive(seg, pamp, paddr, len,
8952 			    *ppp - adjustpages, rw, pflags, preclaim_callback);
8953 		}
8954 
8955 		if (pamp != NULL) {
8956 			ASSERT(svd->type == MAP_SHARED);
8957 			ASSERT(svd->softlockcnt >= npages);
8958 			atomic_add_long((ulong_t *)&svd->softlockcnt, -npages);
8959 		}
8960 
8961 		if (sftlck_sbase) {
8962 			ASSERT(svd->softlockcnt_sbase > 0);
8963 			atomic_dec_ulong((ulong_t *)&svd->softlockcnt_sbase);
8964 		}
8965 		if (sftlck_send) {
8966 			ASSERT(svd->softlockcnt_send > 0);
8967 			atomic_dec_ulong((ulong_t *)&svd->softlockcnt_send);
8968 		}
8969 
8970 		/*
8971 		 * If someone is blocked while unmapping, we purge
8972 		 * segment page cache and thus reclaim pplist synchronously
8973 		 * without waiting for seg_pasync_thread. This speeds up
8974 		 * unmapping in cases where munmap(2) is called, while
8975 		 * raw async i/o is still in progress or where a thread
8976 		 * exits on data fault in a multithreaded application.
8977 		 */
8978 		if (AS_ISUNMAPWAIT(seg->s_as)) {
8979 			if (svd->softlockcnt == 0) {
8980 				mutex_enter(&seg->s_as->a_contents);
8981 				if (AS_ISUNMAPWAIT(seg->s_as)) {
8982 					AS_CLRUNMAPWAIT(seg->s_as);
8983 					cv_broadcast(&seg->s_as->a_cv);
8984 				}
8985 				mutex_exit(&seg->s_as->a_contents);
8986 			} else if (pamp == NULL) {
8987 				/*
8988 				 * softlockcnt is not 0 and this is a
8989 				 * MAP_PRIVATE segment. Try to purge its
8990 				 * pcache entries to reduce softlockcnt.
8991 				 * If it drops to 0 segvn_reclaim()
8992 				 * will wake up a thread waiting on
8993 				 * unmapwait flag.
8994 				 *
8995 				 * We don't purge MAP_SHARED segments with non
8996 				 * 0 softlockcnt since IO is still in progress
8997 				 * for such segments.
8998 				 */
8999 				ASSERT(svd->type == MAP_PRIVATE);
9000 				segvn_purge(seg);
9001 			}
9002 		}
9003 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9004 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_UNLOCK_END,
9005 		    "segvn_pagelock: unlock seg %p addr %p", seg, addr);
9006 		return (0);
9007 	}
9008 
9009 	/* The L_PAGELOCK case ... */
9010 
9011 	VM_STAT_ADD(segvnvmstats.pagelock[1]);
9012 
9013 	/*
9014 	 * For MAP_SHARED segments we have to check protections before
9015 	 * seg_plookup() since pcache entries may be shared by many segments
9016 	 * with potentially different page protections.
9017 	 */
9018 	if (pamp != NULL) {
9019 		ASSERT(svd->type == MAP_SHARED);
9020 		if (svd->pageprot == 0) {
9021 			if ((svd->prot & protchk) == 0) {
9022 				error = EACCES;
9023 				goto out;
9024 			}
9025 		} else {
9026 			/*
9027 			 * check page protections
9028 			 */
9029 			caddr_t ea;
9030 
9031 			if (seg->s_szc) {
9032 				a = lpgaddr;
9033 				ea = lpgeaddr;
9034 			} else {
9035 				a = addr;
9036 				ea = addr + len;
9037 			}
9038 			for (; a < ea; a += pgsz) {
9039 				struct vpage *vp;
9040 
9041 				ASSERT(seg->s_szc == 0 ||
9042 				    sameprot(seg, a, pgsz));
9043 				vp = &svd->vpage[seg_page(seg, a)];
9044 				if ((VPP_PROT(vp) & protchk) == 0) {
9045 					error = EACCES;
9046 					goto out;
9047 				}
9048 			}
9049 		}
9050 	}
9051 
9052 	/*
9053 	 * try to find pages in segment page cache
9054 	 */
9055 	pplist = seg_plookup(seg, pamp, paddr, lpgeaddr - lpgaddr, rw, pflags);
9056 	if (pplist != NULL) {
9057 		if (pamp != NULL) {
9058 			npages = btop((uintptr_t)(lpgeaddr - lpgaddr));
9059 			ASSERT(svd->type == MAP_SHARED);
9060 			atomic_add_long((ulong_t *)&svd->softlockcnt,
9061 			    npages);
9062 		}
9063 		if (sftlck_sbase) {
9064 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_sbase);
9065 		}
9066 		if (sftlck_send) {
9067 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_send);
9068 		}
9069 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9070 		*ppp = pplist + adjustpages;
9071 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_HIT_END,
9072 		    "segvn_pagelock: cache hit seg %p addr %p", seg, addr);
9073 		return (0);
9074 	}
9075 
9076 	/*
9077 	 * For MAP_SHARED segments we already verified above that segment
9078 	 * protections allow this pagelock operation.
9079 	 */
9080 	if (pamp == NULL) {
9081 		ASSERT(svd->type == MAP_PRIVATE);
9082 		if (svd->pageprot == 0) {
9083 			if ((svd->prot & protchk) == 0) {
9084 				error = EACCES;
9085 				goto out;
9086 			}
9087 			if (svd->prot & PROT_WRITE) {
9088 				wlen = lpgeaddr - lpgaddr;
9089 			} else {
9090 				wlen = 0;
9091 				ASSERT(rw == S_READ);
9092 			}
9093 		} else {
9094 			int wcont = 1;
9095 			/*
9096 			 * check page protections
9097 			 */
9098 			for (a = lpgaddr, wlen = 0; a < lpgeaddr; a += pgsz) {
9099 				struct vpage *vp;
9100 
9101 				ASSERT(seg->s_szc == 0 ||
9102 				    sameprot(seg, a, pgsz));
9103 				vp = &svd->vpage[seg_page(seg, a)];
9104 				if ((VPP_PROT(vp) & protchk) == 0) {
9105 					error = EACCES;
9106 					goto out;
9107 				}
9108 				if (wcont && (VPP_PROT(vp) & PROT_WRITE)) {
9109 					wlen += pgsz;
9110 				} else {
9111 					wcont = 0;
9112 					ASSERT(rw == S_READ);
9113 				}
9114 			}
9115 		}
9116 		ASSERT(rw == S_READ || wlen == lpgeaddr - lpgaddr);
9117 		ASSERT(rw == S_WRITE || wlen <= lpgeaddr - lpgaddr);
9118 	}
9119 
9120 	/*
9121 	 * Only build large page adjusted shadow list if we expect to insert
9122 	 * it into pcache. For large enough pages it's a big overhead to
9123 	 * create a shadow list of the entire large page. But this overhead
9124 	 * should be amortized over repeated pcache hits on subsequent reuse
9125 	 * of this shadow list (IO into any range within this shadow list will
9126 	 * find it in pcache since we large page align the request for pcache
9127 	 * lookups). pcache performance is improved with bigger shadow lists
9128 	 * as it reduces the time to pcache the entire big segment and reduces
9129 	 * pcache chain length.
9130 	 */
9131 	if (seg_pinsert_check(seg, pamp, paddr,
9132 	    lpgeaddr - lpgaddr, pflags) == SEGP_SUCCESS) {
9133 		addr = lpgaddr;
9134 		len = lpgeaddr - lpgaddr;
9135 		use_pcache = 1;
9136 	} else {
9137 		use_pcache = 0;
9138 		/*
9139 		 * Since this entry will not be inserted into the pcache, we
9140 		 * will not do any adjustments to the starting address or
9141 		 * size of the memory to be locked.
9142 		 */
9143 		adjustpages = 0;
9144 	}
9145 	npages = btop(len);
9146 
9147 	pplist = kmem_alloc(sizeof (page_t *) * (npages + 1), KM_SLEEP);
9148 	pl = pplist;
9149 	*ppp = pplist + adjustpages;
9150 	/*
9151 	 * If use_pcache is 0 this shadow list is not large page adjusted.
9152 	 * Record this info in the last entry of shadow array so that
9153 	 * L_PAGEUNLOCK can determine if it should large page adjust the
9154 	 * address range to find the real range that was locked.
9155 	 */
9156 	pl[npages] = use_pcache ? PCACHE_SHWLIST : NOPCACHE_SHWLIST;
9157 
9158 	page = seg_page(seg, addr);
9159 	anon_index = svd->anon_index + page;
9160 
9161 	anlock = 0;
9162 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
9163 	ASSERT(amp->a_szc >= seg->s_szc);
9164 	anpgcnt = page_get_pagecnt(amp->a_szc);
9165 	for (a = addr; a < addr + len; a += PAGESIZE, anon_index++) {
9166 		struct anon *ap;
9167 		struct vnode *vp;
9168 		u_offset_t off;
9169 
9170 		/*
9171 		 * Lock and unlock anon array only once per large page.
9172 		 * anon_array_enter() locks the root anon slot according to
9173 		 * a_szc which can't change while anon map is locked.  We lock
9174 		 * anon the first time through this loop and each time we
9175 		 * reach anon index that corresponds to a root of a large
9176 		 * page.
9177 		 */
9178 		if (a == addr || P2PHASE(anon_index, anpgcnt) == 0) {
9179 			ASSERT(anlock == 0);
9180 			anon_array_enter(amp, anon_index, &cookie);
9181 			anlock = 1;
9182 		}
9183 		ap = anon_get_ptr(amp->ahp, anon_index);
9184 
9185 		/*
9186 		 * We must never use seg_pcache for COW pages
9187 		 * because we might end up with original page still
9188 		 * lying in seg_pcache even after private page is
9189 		 * created. This leads to data corruption as
9190 		 * aio_write refers to the page still in cache
9191 		 * while all other accesses refer to the private
9192 		 * page.
9193 		 */
9194 		if (ap == NULL || ap->an_refcnt != 1) {
9195 			struct vpage *vpage;
9196 
9197 			if (seg->s_szc) {
9198 				error = EFAULT;
9199 				break;
9200 			}
9201 			if (svd->vpage != NULL) {
9202 				vpage = &svd->vpage[seg_page(seg, a)];
9203 			} else {
9204 				vpage = NULL;
9205 			}
9206 			ASSERT(anlock);
9207 			anon_array_exit(&cookie);
9208 			anlock = 0;
9209 			pp = NULL;
9210 			error = segvn_faultpage(seg->s_as->a_hat, seg, a, 0,
9211 			    vpage, &pp, 0, F_INVAL, rw, 1);
9212 			if (error) {
9213 				error = fc_decode(error);
9214 				break;
9215 			}
9216 			anon_array_enter(amp, anon_index, &cookie);
9217 			anlock = 1;
9218 			ap = anon_get_ptr(amp->ahp, anon_index);
9219 			if (ap == NULL || ap->an_refcnt != 1) {
9220 				error = EFAULT;
9221 				break;
9222 			}
9223 		}
9224 		swap_xlate(ap, &vp, &off);
9225 		pp = page_lookup_nowait(vp, off, SE_SHARED);
9226 		if (pp == NULL) {
9227 			error = EFAULT;
9228 			break;
9229 		}
9230 		if (ap->an_pvp != NULL) {
9231 			anon_swap_free(ap, pp);
9232 		}
9233 		/*
9234 		 * Unlock anon if this is the last slot in a large page.
9235 		 */
9236 		if (P2PHASE(anon_index, anpgcnt) == anpgcnt - 1) {
9237 			ASSERT(anlock);
9238 			anon_array_exit(&cookie);
9239 			anlock = 0;
9240 		}
9241 		*pplist++ = pp;
9242 	}
9243 	if (anlock) {		/* Ensure the lock is dropped */
9244 		anon_array_exit(&cookie);
9245 	}
9246 	ANON_LOCK_EXIT(&amp->a_rwlock);
9247 
9248 	if (a >= addr + len) {
9249 		atomic_add_long((ulong_t *)&svd->softlockcnt, npages);
9250 		if (pamp != NULL) {
9251 			ASSERT(svd->type == MAP_SHARED);
9252 			atomic_add_long((ulong_t *)&pamp->a_softlockcnt,
9253 			    npages);
9254 			wlen = len;
9255 		}
9256 		if (sftlck_sbase) {
9257 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_sbase);
9258 		}
9259 		if (sftlck_send) {
9260 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_send);
9261 		}
9262 		if (use_pcache) {
9263 			(void) seg_pinsert(seg, pamp, paddr, len, wlen, pl,
9264 			    rw, pflags, preclaim_callback);
9265 		}
9266 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9267 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_FILL_END,
9268 		    "segvn_pagelock: cache fill seg %p addr %p", seg, addr);
9269 		return (0);
9270 	}
9271 
9272 	pplist = pl;
9273 	np = ((uintptr_t)(a - addr)) >> PAGESHIFT;
9274 	while (np > (uint_t)0) {
9275 		ASSERT(PAGE_LOCKED(*pplist));
9276 		page_unlock(*pplist);
9277 		np--;
9278 		pplist++;
9279 	}
9280 	kmem_free(pl, sizeof (page_t *) * (npages + 1));
9281 out:
9282 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9283 	*ppp = NULL;
9284 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
9285 	    "segvn_pagelock: cache miss seg %p addr %p", seg, addr);
9286 	return (error);
9287 }
9288 
9289 /*
9290  * purge any cached pages in the I/O page cache
9291  */
9292 static void
9293 segvn_purge(struct seg *seg)
9294 {
9295 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9296 
9297 	/*
9298 	 * pcache is only used by pure anon segments.
9299 	 */
9300 	if (svd->amp == NULL || svd->vp != NULL) {
9301 		return;
9302 	}
9303 
9304 	/*
9305 	 * For MAP_SHARED segments non 0 segment's softlockcnt means
9306 	 * active IO is still in progress via this segment. So we only
9307 	 * purge MAP_SHARED segments when their softlockcnt is 0.
9308 	 */
9309 	if (svd->type == MAP_PRIVATE) {
9310 		if (svd->softlockcnt) {
9311 			seg_ppurge(seg, NULL, 0);
9312 		}
9313 	} else if (svd->softlockcnt == 0 && svd->amp->a_softlockcnt != 0) {
9314 		seg_ppurge(seg, svd->amp, 0);
9315 	}
9316 }
9317 
9318 /*
9319  * If async argument is not 0 we are called from pcache async thread and don't
9320  * hold AS lock.
9321  */
9322 
9323 /*ARGSUSED*/
9324 static int
9325 segvn_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9326 	enum seg_rw rw, int async)
9327 {
9328 	struct seg *seg = (struct seg *)ptag;
9329 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9330 	pgcnt_t np, npages;
9331 	struct page **pl;
9332 
9333 	npages = np = btop(len);
9334 	ASSERT(npages);
9335 
9336 	ASSERT(svd->vp == NULL && svd->amp != NULL);
9337 	ASSERT(svd->softlockcnt >= npages);
9338 	ASSERT(async || AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
9339 
9340 	pl = pplist;
9341 
9342 	ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9343 	ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9344 
9345 	while (np > (uint_t)0) {
9346 		if (rw == S_WRITE) {
9347 			hat_setrefmod(*pplist);
9348 		} else {
9349 			hat_setref(*pplist);
9350 		}
9351 		page_unlock(*pplist);
9352 		np--;
9353 		pplist++;
9354 	}
9355 
9356 	kmem_free(pl, sizeof (page_t *) * (npages + 1));
9357 
9358 	/*
9359 	 * If we are pcache async thread we don't hold AS lock. This means if
9360 	 * softlockcnt drops to 0 after the decrement below address space may
9361 	 * get freed. We can't allow it since after softlock derement to 0 we
9362 	 * still need to access as structure for possible wakeup of unmap
9363 	 * waiters. To prevent the disappearance of as we take this segment
9364 	 * segfree_syncmtx. segvn_free() also takes this mutex as a barrier to
9365 	 * make sure this routine completes before segment is freed.
9366 	 *
9367 	 * The second complication we have to deal with in async case is a
9368 	 * possibility of missed wake up of unmap wait thread. When we don't
9369 	 * hold as lock here we may take a_contents lock before unmap wait
9370 	 * thread that was first to see softlockcnt was still not 0. As a
9371 	 * result we'll fail to wake up an unmap wait thread. To avoid this
9372 	 * race we set nounmapwait flag in as structure if we drop softlockcnt
9373 	 * to 0 when we were called by pcache async thread.  unmapwait thread
9374 	 * will not block if this flag is set.
9375 	 */
9376 	if (async) {
9377 		mutex_enter(&svd->segfree_syncmtx);
9378 	}
9379 
9380 	if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -npages)) {
9381 		if (async || AS_ISUNMAPWAIT(seg->s_as)) {
9382 			mutex_enter(&seg->s_as->a_contents);
9383 			if (async) {
9384 				AS_SETNOUNMAPWAIT(seg->s_as);
9385 			}
9386 			if (AS_ISUNMAPWAIT(seg->s_as)) {
9387 				AS_CLRUNMAPWAIT(seg->s_as);
9388 				cv_broadcast(&seg->s_as->a_cv);
9389 			}
9390 			mutex_exit(&seg->s_as->a_contents);
9391 		}
9392 	}
9393 
9394 	if (async) {
9395 		mutex_exit(&svd->segfree_syncmtx);
9396 	}
9397 	return (0);
9398 }
9399 
9400 /*ARGSUSED*/
9401 static int
9402 shamp_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9403 	enum seg_rw rw, int async)
9404 {
9405 	amp_t *amp = (amp_t *)ptag;
9406 	pgcnt_t np, npages;
9407 	struct page **pl;
9408 
9409 	npages = np = btop(len);
9410 	ASSERT(npages);
9411 	ASSERT(amp->a_softlockcnt >= npages);
9412 
9413 	pl = pplist;
9414 
9415 	ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9416 	ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9417 
9418 	while (np > (uint_t)0) {
9419 		if (rw == S_WRITE) {
9420 			hat_setrefmod(*pplist);
9421 		} else {
9422 			hat_setref(*pplist);
9423 		}
9424 		page_unlock(*pplist);
9425 		np--;
9426 		pplist++;
9427 	}
9428 
9429 	kmem_free(pl, sizeof (page_t *) * (npages + 1));
9430 
9431 	/*
9432 	 * If somebody sleeps in anonmap_purge() wake them up if a_softlockcnt
9433 	 * drops to 0. anon map can't be freed until a_softlockcnt drops to 0
9434 	 * and anonmap_purge() acquires a_purgemtx.
9435 	 */
9436 	mutex_enter(&amp->a_purgemtx);
9437 	if (!atomic_add_long_nv((ulong_t *)&amp->a_softlockcnt, -npages) &&
9438 	    amp->a_purgewait) {
9439 		amp->a_purgewait = 0;
9440 		cv_broadcast(&amp->a_purgecv);
9441 	}
9442 	mutex_exit(&amp->a_purgemtx);
9443 	return (0);
9444 }
9445 
9446 /*
9447  * get a memory ID for an addr in a given segment
9448  *
9449  * XXX only creates PAGESIZE pages if anon slots are not initialized.
9450  * At fault time they will be relocated into larger pages.
9451  */
9452 static int
9453 segvn_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
9454 {
9455 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9456 	struct anon 	*ap = NULL;
9457 	ulong_t		anon_index;
9458 	struct anon_map	*amp;
9459 	anon_sync_obj_t cookie;
9460 
9461 	if (svd->type == MAP_PRIVATE) {
9462 		memidp->val[0] = (uintptr_t)seg->s_as;
9463 		memidp->val[1] = (uintptr_t)addr;
9464 		return (0);
9465 	}
9466 
9467 	if (svd->type == MAP_SHARED) {
9468 		if (svd->vp) {
9469 			memidp->val[0] = (uintptr_t)svd->vp;
9470 			memidp->val[1] = (u_longlong_t)svd->offset +
9471 			    (uintptr_t)(addr - seg->s_base);
9472 			return (0);
9473 		} else {
9474 
9475 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
9476 			if ((amp = svd->amp) != NULL) {
9477 				anon_index = svd->anon_index +
9478 				    seg_page(seg, addr);
9479 			}
9480 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9481 
9482 			ASSERT(amp != NULL);
9483 
9484 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
9485 			anon_array_enter(amp, anon_index, &cookie);
9486 			ap = anon_get_ptr(amp->ahp, anon_index);
9487 			if (ap == NULL) {
9488 				page_t		*pp;
9489 
9490 				pp = anon_zero(seg, addr, &ap, svd->cred);
9491 				if (pp == NULL) {
9492 					anon_array_exit(&cookie);
9493 					ANON_LOCK_EXIT(&amp->a_rwlock);
9494 					return (ENOMEM);
9495 				}
9496 				ASSERT(anon_get_ptr(amp->ahp, anon_index)
9497 				    == NULL);
9498 				(void) anon_set_ptr(amp->ahp, anon_index,
9499 				    ap, ANON_SLEEP);
9500 				page_unlock(pp);
9501 			}
9502 
9503 			anon_array_exit(&cookie);
9504 			ANON_LOCK_EXIT(&amp->a_rwlock);
9505 
9506 			memidp->val[0] = (uintptr_t)ap;
9507 			memidp->val[1] = (uintptr_t)addr & PAGEOFFSET;
9508 			return (0);
9509 		}
9510 	}
9511 	return (EINVAL);
9512 }
9513 
9514 static int
9515 sameprot(struct seg *seg, caddr_t a, size_t len)
9516 {
9517 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9518 	struct vpage *vpage;
9519 	spgcnt_t pages = btop(len);
9520 	uint_t prot;
9521 
9522 	if (svd->pageprot == 0)
9523 		return (1);
9524 
9525 	ASSERT(svd->vpage != NULL);
9526 
9527 	vpage = &svd->vpage[seg_page(seg, a)];
9528 	prot = VPP_PROT(vpage);
9529 	vpage++;
9530 	pages--;
9531 	while (pages-- > 0) {
9532 		if (prot != VPP_PROT(vpage))
9533 			return (0);
9534 		vpage++;
9535 	}
9536 	return (1);
9537 }
9538 
9539 /*
9540  * Get memory allocation policy info for specified address in given segment
9541  */
9542 static lgrp_mem_policy_info_t *
9543 segvn_getpolicy(struct seg *seg, caddr_t addr)
9544 {
9545 	struct anon_map		*amp;
9546 	ulong_t			anon_index;
9547 	lgrp_mem_policy_info_t	*policy_info;
9548 	struct segvn_data	*svn_data;
9549 	u_offset_t		vn_off;
9550 	vnode_t			*vp;
9551 
9552 	ASSERT(seg != NULL);
9553 
9554 	svn_data = (struct segvn_data *)seg->s_data;
9555 	if (svn_data == NULL)
9556 		return (NULL);
9557 
9558 	/*
9559 	 * Get policy info for private or shared memory
9560 	 */
9561 	if (svn_data->type != MAP_SHARED) {
9562 		if (svn_data->tr_state != SEGVN_TR_ON) {
9563 			policy_info = &svn_data->policy_info;
9564 		} else {
9565 			policy_info = &svn_data->tr_policy_info;
9566 			ASSERT(policy_info->mem_policy ==
9567 			    LGRP_MEM_POLICY_NEXT_SEG);
9568 		}
9569 	} else {
9570 		amp = svn_data->amp;
9571 		anon_index = svn_data->anon_index + seg_page(seg, addr);
9572 		vp = svn_data->vp;
9573 		vn_off = svn_data->offset + (uintptr_t)(addr - seg->s_base);
9574 		policy_info = lgrp_shm_policy_get(amp, anon_index, vp, vn_off);
9575 	}
9576 
9577 	return (policy_info);
9578 }
9579 
9580 /*ARGSUSED*/
9581 static int
9582 segvn_capable(struct seg *seg, segcapability_t capability)
9583 {
9584 	return (0);
9585 }
9586 
9587 /*
9588  * Bind text vnode segment to an amp. If we bind successfully mappings will be
9589  * established to per vnode mapping per lgroup amp pages instead of to vnode
9590  * pages. There's one amp per vnode text mapping per lgroup. Many processes
9591  * may share the same text replication amp. If a suitable amp doesn't already
9592  * exist in svntr hash table create a new one.  We may fail to bind to amp if
9593  * segment is not eligible for text replication.  Code below first checks for
9594  * these conditions. If binding is successful segment tr_state is set to on
9595  * and svd->amp points to the amp to use. Otherwise tr_state is set to off and
9596  * svd->amp remains as NULL.
9597  */
9598 static void
9599 segvn_textrepl(struct seg *seg)
9600 {
9601 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
9602 	vnode_t			*vp = svd->vp;
9603 	u_offset_t		off = svd->offset;
9604 	size_t			size = seg->s_size;
9605 	u_offset_t		eoff = off + size;
9606 	uint_t			szc = seg->s_szc;
9607 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9608 	svntr_t			*svntrp;
9609 	struct vattr		va;
9610 	proc_t			*p = seg->s_as->a_proc;
9611 	lgrp_id_t		lgrp_id;
9612 	lgrp_id_t		olid;
9613 	int			first;
9614 	struct anon_map		*amp;
9615 
9616 	ASSERT(AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
9617 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9618 	ASSERT(p != NULL);
9619 	ASSERT(svd->tr_state == SEGVN_TR_INIT);
9620 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9621 	ASSERT(svd->flags & MAP_TEXT);
9622 	ASSERT(svd->type == MAP_PRIVATE);
9623 	ASSERT(vp != NULL && svd->amp == NULL);
9624 	ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
9625 	ASSERT(!(svd->flags & MAP_NORESERVE) && svd->swresv == 0);
9626 	ASSERT(seg->s_as != &kas);
9627 	ASSERT(off < eoff);
9628 	ASSERT(svntr_hashtab != NULL);
9629 
9630 	/*
9631 	 * If numa optimizations are no longer desired bail out.
9632 	 */
9633 	if (!lgrp_optimizations()) {
9634 		svd->tr_state = SEGVN_TR_OFF;
9635 		return;
9636 	}
9637 
9638 	/*
9639 	 * Avoid creating anon maps with size bigger than the file size.
9640 	 * If VOP_GETATTR() call fails bail out.
9641 	 */
9642 	va.va_mask = AT_SIZE | AT_MTIME | AT_CTIME;
9643 	if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL) != 0) {
9644 		svd->tr_state = SEGVN_TR_OFF;
9645 		SEGVN_TR_ADDSTAT(gaerr);
9646 		return;
9647 	}
9648 	if (btopr(va.va_size) < btopr(eoff)) {
9649 		svd->tr_state = SEGVN_TR_OFF;
9650 		SEGVN_TR_ADDSTAT(overmap);
9651 		return;
9652 	}
9653 
9654 	/*
9655 	 * VVMEXEC may not be set yet if exec() prefaults text segment. Set
9656 	 * this flag now before vn_is_mapped(V_WRITE) so that MAP_SHARED
9657 	 * mapping that checks if trcache for this vnode needs to be
9658 	 * invalidated can't miss us.
9659 	 */
9660 	if (!(vp->v_flag & VVMEXEC)) {
9661 		mutex_enter(&vp->v_lock);
9662 		vp->v_flag |= VVMEXEC;
9663 		mutex_exit(&vp->v_lock);
9664 	}
9665 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9666 	/*
9667 	 * Bail out if potentially MAP_SHARED writable mappings exist to this
9668 	 * vnode.  We don't want to use old file contents from existing
9669 	 * replicas if this mapping was established after the original file
9670 	 * was changed.
9671 	 */
9672 	if (vn_is_mapped(vp, V_WRITE)) {
9673 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9674 		svd->tr_state = SEGVN_TR_OFF;
9675 		SEGVN_TR_ADDSTAT(wrcnt);
9676 		return;
9677 	}
9678 	svntrp = svntr_hashtab[hash].tr_head;
9679 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9680 		ASSERT(svntrp->tr_refcnt != 0);
9681 		if (svntrp->tr_vp != vp) {
9682 			continue;
9683 		}
9684 
9685 		/*
9686 		 * Bail out if the file or its attributes were changed after
9687 		 * this replication entry was created since we need to use the
9688 		 * latest file contents. Note that mtime test alone is not
9689 		 * sufficient because a user can explicitly change mtime via
9690 		 * utimes(2) interfaces back to the old value after modifiying
9691 		 * the file contents. To detect this case we also have to test
9692 		 * ctime which among other things records the time of the last
9693 		 * mtime change by utimes(2). ctime is not changed when the file
9694 		 * is only read or executed so we expect that typically existing
9695 		 * replication amp's can be used most of the time.
9696 		 */
9697 		if (!svntrp->tr_valid ||
9698 		    svntrp->tr_mtime.tv_sec != va.va_mtime.tv_sec ||
9699 		    svntrp->tr_mtime.tv_nsec != va.va_mtime.tv_nsec ||
9700 		    svntrp->tr_ctime.tv_sec != va.va_ctime.tv_sec ||
9701 		    svntrp->tr_ctime.tv_nsec != va.va_ctime.tv_nsec) {
9702 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9703 			svd->tr_state = SEGVN_TR_OFF;
9704 			SEGVN_TR_ADDSTAT(stale);
9705 			return;
9706 		}
9707 		/*
9708 		 * if off, eoff and szc match current segment we found the
9709 		 * existing entry we can use.
9710 		 */
9711 		if (svntrp->tr_off == off && svntrp->tr_eoff == eoff &&
9712 		    svntrp->tr_szc == szc) {
9713 			break;
9714 		}
9715 		/*
9716 		 * Don't create different but overlapping in file offsets
9717 		 * entries to avoid replication of the same file pages more
9718 		 * than once per lgroup.
9719 		 */
9720 		if ((off >= svntrp->tr_off && off < svntrp->tr_eoff) ||
9721 		    (eoff > svntrp->tr_off && eoff <= svntrp->tr_eoff)) {
9722 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9723 			svd->tr_state = SEGVN_TR_OFF;
9724 			SEGVN_TR_ADDSTAT(overlap);
9725 			return;
9726 		}
9727 	}
9728 	/*
9729 	 * If we didn't find existing entry create a new one.
9730 	 */
9731 	if (svntrp == NULL) {
9732 		svntrp = kmem_cache_alloc(svntr_cache, KM_NOSLEEP);
9733 		if (svntrp == NULL) {
9734 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9735 			svd->tr_state = SEGVN_TR_OFF;
9736 			SEGVN_TR_ADDSTAT(nokmem);
9737 			return;
9738 		}
9739 #ifdef DEBUG
9740 		{
9741 			lgrp_id_t i;
9742 			for (i = 0; i < NLGRPS_MAX; i++) {
9743 				ASSERT(svntrp->tr_amp[i] == NULL);
9744 			}
9745 		}
9746 #endif /* DEBUG */
9747 		svntrp->tr_vp = vp;
9748 		svntrp->tr_off = off;
9749 		svntrp->tr_eoff = eoff;
9750 		svntrp->tr_szc = szc;
9751 		svntrp->tr_valid = 1;
9752 		svntrp->tr_mtime = va.va_mtime;
9753 		svntrp->tr_ctime = va.va_ctime;
9754 		svntrp->tr_refcnt = 0;
9755 		svntrp->tr_next = svntr_hashtab[hash].tr_head;
9756 		svntr_hashtab[hash].tr_head = svntrp;
9757 	}
9758 	first = 1;
9759 again:
9760 	/*
9761 	 * We want to pick a replica with pages on main thread's (t_tid = 1,
9762 	 * aka T1) lgrp. Currently text replication is only optimized for
9763 	 * workloads that either have all threads of a process on the same
9764 	 * lgrp or execute their large text primarily on main thread.
9765 	 */
9766 	lgrp_id = p->p_t1_lgrpid;
9767 	if (lgrp_id == LGRP_NONE) {
9768 		/*
9769 		 * In case exec() prefaults text on non main thread use
9770 		 * current thread lgrpid.  It will become main thread anyway
9771 		 * soon.
9772 		 */
9773 		lgrp_id = lgrp_home_id(curthread);
9774 	}
9775 	/*
9776 	 * Set p_tr_lgrpid to lgrpid if it hasn't been set yet.  Otherwise
9777 	 * just set it to NLGRPS_MAX if it's different from current process T1
9778 	 * home lgrp.  p_tr_lgrpid is used to detect if process uses text
9779 	 * replication and T1 new home is different from lgrp used for text
9780 	 * replication. When this happens asyncronous segvn thread rechecks if
9781 	 * segments should change lgrps used for text replication.  If we fail
9782 	 * to set p_tr_lgrpid with atomic_cas_32 then set it to NLGRPS_MAX
9783 	 * without cas if it's not already NLGRPS_MAX and not equal lgrp_id
9784 	 * we want to use.  We don't need to use cas in this case because
9785 	 * another thread that races in between our non atomic check and set
9786 	 * may only change p_tr_lgrpid to NLGRPS_MAX at this point.
9787 	 */
9788 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9789 	olid = p->p_tr_lgrpid;
9790 	if (lgrp_id != olid && olid != NLGRPS_MAX) {
9791 		lgrp_id_t nlid = (olid == LGRP_NONE) ? lgrp_id : NLGRPS_MAX;
9792 		if (atomic_cas_32((uint32_t *)&p->p_tr_lgrpid, olid, nlid) !=
9793 		    olid) {
9794 			olid = p->p_tr_lgrpid;
9795 			ASSERT(olid != LGRP_NONE);
9796 			if (olid != lgrp_id && olid != NLGRPS_MAX) {
9797 				p->p_tr_lgrpid = NLGRPS_MAX;
9798 			}
9799 		}
9800 		ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9801 		membar_producer();
9802 		/*
9803 		 * lgrp_move_thread() won't schedule async recheck after
9804 		 * p->p_t1_lgrpid update unless p->p_tr_lgrpid is not
9805 		 * LGRP_NONE. Recheck p_t1_lgrpid once now that p->p_tr_lgrpid
9806 		 * is not LGRP_NONE.
9807 		 */
9808 		if (first && p->p_t1_lgrpid != LGRP_NONE &&
9809 		    p->p_t1_lgrpid != lgrp_id) {
9810 			first = 0;
9811 			goto again;
9812 		}
9813 	}
9814 	/*
9815 	 * If no amp was created yet for lgrp_id create a new one as long as
9816 	 * we have enough memory to afford it.
9817 	 */
9818 	if ((amp = svntrp->tr_amp[lgrp_id]) == NULL) {
9819 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9820 		if (trmem > segvn_textrepl_max_bytes) {
9821 			SEGVN_TR_ADDSTAT(normem);
9822 			goto fail;
9823 		}
9824 		if (anon_try_resv_zone(size, NULL) == 0) {
9825 			SEGVN_TR_ADDSTAT(noanon);
9826 			goto fail;
9827 		}
9828 		amp = anonmap_alloc(size, size, ANON_NOSLEEP);
9829 		if (amp == NULL) {
9830 			anon_unresv_zone(size, NULL);
9831 			SEGVN_TR_ADDSTAT(nokmem);
9832 			goto fail;
9833 		}
9834 		ASSERT(amp->refcnt == 1);
9835 		amp->a_szc = szc;
9836 		svntrp->tr_amp[lgrp_id] = amp;
9837 		SEGVN_TR_ADDSTAT(newamp);
9838 	}
9839 	svntrp->tr_refcnt++;
9840 	ASSERT(svd->svn_trnext == NULL);
9841 	ASSERT(svd->svn_trprev == NULL);
9842 	svd->svn_trnext = svntrp->tr_svnhead;
9843 	svd->svn_trprev = NULL;
9844 	if (svntrp->tr_svnhead != NULL) {
9845 		svntrp->tr_svnhead->svn_trprev = svd;
9846 	}
9847 	svntrp->tr_svnhead = svd;
9848 	ASSERT(amp->a_szc == szc && amp->size == size && amp->swresv == size);
9849 	ASSERT(amp->refcnt >= 1);
9850 	svd->amp = amp;
9851 	svd->anon_index = 0;
9852 	svd->tr_policy_info.mem_policy = LGRP_MEM_POLICY_NEXT_SEG;
9853 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
9854 	svd->tr_state = SEGVN_TR_ON;
9855 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9856 	SEGVN_TR_ADDSTAT(repl);
9857 	return;
9858 fail:
9859 	ASSERT(segvn_textrepl_bytes >= size);
9860 	atomic_add_long(&segvn_textrepl_bytes, -size);
9861 	ASSERT(svntrp != NULL);
9862 	ASSERT(svntrp->tr_amp[lgrp_id] == NULL);
9863 	if (svntrp->tr_refcnt == 0) {
9864 		ASSERT(svntrp == svntr_hashtab[hash].tr_head);
9865 		svntr_hashtab[hash].tr_head = svntrp->tr_next;
9866 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9867 		kmem_cache_free(svntr_cache, svntrp);
9868 	} else {
9869 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9870 	}
9871 	svd->tr_state = SEGVN_TR_OFF;
9872 }
9873 
9874 /*
9875  * Convert seg back to regular vnode mapping seg by unbinding it from its text
9876  * replication amp.  This routine is most typically called when segment is
9877  * unmapped but can also be called when segment no longer qualifies for text
9878  * replication (e.g. due to protection changes). If unload_unmap is set use
9879  * HAT_UNLOAD_UNMAP flag in hat_unload_callback().  If we are the last user of
9880  * svntr free all its anon maps and remove it from the hash table.
9881  */
9882 static void
9883 segvn_textunrepl(struct seg *seg, int unload_unmap)
9884 {
9885 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
9886 	vnode_t			*vp = svd->vp;
9887 	u_offset_t		off = svd->offset;
9888 	size_t			size = seg->s_size;
9889 	u_offset_t		eoff = off + size;
9890 	uint_t			szc = seg->s_szc;
9891 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9892 	svntr_t			*svntrp;
9893 	svntr_t			**prv_svntrp;
9894 	lgrp_id_t		lgrp_id = svd->tr_policy_info.mem_lgrpid;
9895 	lgrp_id_t		i;
9896 
9897 	ASSERT(AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
9898 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
9899 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9900 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9901 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9902 	ASSERT(svd->amp != NULL);
9903 	ASSERT(svd->amp->refcnt >= 1);
9904 	ASSERT(svd->anon_index == 0);
9905 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9906 	ASSERT(svntr_hashtab != NULL);
9907 
9908 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9909 	prv_svntrp = &svntr_hashtab[hash].tr_head;
9910 	for (; (svntrp = *prv_svntrp) != NULL; prv_svntrp = &svntrp->tr_next) {
9911 		ASSERT(svntrp->tr_refcnt != 0);
9912 		if (svntrp->tr_vp == vp && svntrp->tr_off == off &&
9913 		    svntrp->tr_eoff == eoff && svntrp->tr_szc == szc) {
9914 			break;
9915 		}
9916 	}
9917 	if (svntrp == NULL) {
9918 		panic("segvn_textunrepl: svntr record not found");
9919 	}
9920 	if (svntrp->tr_amp[lgrp_id] != svd->amp) {
9921 		panic("segvn_textunrepl: amp mismatch");
9922 	}
9923 	svd->tr_state = SEGVN_TR_OFF;
9924 	svd->amp = NULL;
9925 	if (svd->svn_trprev == NULL) {
9926 		ASSERT(svntrp->tr_svnhead == svd);
9927 		svntrp->tr_svnhead = svd->svn_trnext;
9928 		if (svntrp->tr_svnhead != NULL) {
9929 			svntrp->tr_svnhead->svn_trprev = NULL;
9930 		}
9931 		svd->svn_trnext = NULL;
9932 	} else {
9933 		svd->svn_trprev->svn_trnext = svd->svn_trnext;
9934 		if (svd->svn_trnext != NULL) {
9935 			svd->svn_trnext->svn_trprev = svd->svn_trprev;
9936 			svd->svn_trnext = NULL;
9937 		}
9938 		svd->svn_trprev = NULL;
9939 	}
9940 	if (--svntrp->tr_refcnt) {
9941 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9942 		goto done;
9943 	}
9944 	*prv_svntrp = svntrp->tr_next;
9945 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9946 	for (i = 0; i < NLGRPS_MAX; i++) {
9947 		struct anon_map *amp = svntrp->tr_amp[i];
9948 		if (amp == NULL) {
9949 			continue;
9950 		}
9951 		ASSERT(amp->refcnt == 1);
9952 		ASSERT(amp->swresv == size);
9953 		ASSERT(amp->size == size);
9954 		ASSERT(amp->a_szc == szc);
9955 		if (amp->a_szc != 0) {
9956 			anon_free_pages(amp->ahp, 0, size, szc);
9957 		} else {
9958 			anon_free(amp->ahp, 0, size);
9959 		}
9960 		svntrp->tr_amp[i] = NULL;
9961 		ASSERT(segvn_textrepl_bytes >= size);
9962 		atomic_add_long(&segvn_textrepl_bytes, -size);
9963 		anon_unresv_zone(amp->swresv, NULL);
9964 		amp->refcnt = 0;
9965 		anonmap_free(amp);
9966 	}
9967 	kmem_cache_free(svntr_cache, svntrp);
9968 done:
9969 	hat_unload_callback(seg->s_as->a_hat, seg->s_base, size,
9970 	    unload_unmap ? HAT_UNLOAD_UNMAP : 0, NULL);
9971 }
9972 
9973 /*
9974  * This is called when a MAP_SHARED writable mapping is created to a vnode
9975  * that is currently used for execution (VVMEXEC flag is set). In this case we
9976  * need to prevent further use of existing replicas.
9977  */
9978 static void
9979 segvn_inval_trcache(vnode_t *vp)
9980 {
9981 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9982 	svntr_t			*svntrp;
9983 
9984 	ASSERT(vp->v_flag & VVMEXEC);
9985 
9986 	if (svntr_hashtab == NULL) {
9987 		return;
9988 	}
9989 
9990 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9991 	svntrp = svntr_hashtab[hash].tr_head;
9992 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9993 		ASSERT(svntrp->tr_refcnt != 0);
9994 		if (svntrp->tr_vp == vp && svntrp->tr_valid) {
9995 			svntrp->tr_valid = 0;
9996 		}
9997 	}
9998 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9999 }
10000 
10001 static void
10002 segvn_trasync_thread(void)
10003 {
10004 	callb_cpr_t cpr_info;
10005 	kmutex_t cpr_lock;	/* just for CPR stuff */
10006 
10007 	mutex_init(&cpr_lock, NULL, MUTEX_DEFAULT, NULL);
10008 
10009 	CALLB_CPR_INIT(&cpr_info, &cpr_lock,
10010 	    callb_generic_cpr, "segvn_async");
10011 
10012 	if (segvn_update_textrepl_interval == 0) {
10013 		segvn_update_textrepl_interval = segvn_update_tr_time * hz;
10014 	} else {
10015 		segvn_update_textrepl_interval *= hz;
10016 	}
10017 	(void) timeout(segvn_trupdate_wakeup, NULL,
10018 	    segvn_update_textrepl_interval);
10019 
10020 	for (;;) {
10021 		mutex_enter(&cpr_lock);
10022 		CALLB_CPR_SAFE_BEGIN(&cpr_info);
10023 		mutex_exit(&cpr_lock);
10024 		sema_p(&segvn_trasync_sem);
10025 		mutex_enter(&cpr_lock);
10026 		CALLB_CPR_SAFE_END(&cpr_info, &cpr_lock);
10027 		mutex_exit(&cpr_lock);
10028 		segvn_trupdate();
10029 	}
10030 }
10031 
10032 static uint64_t segvn_lgrp_trthr_migrs_snpsht = 0;
10033 
10034 static void
10035 segvn_trupdate_wakeup(void *dummy)
10036 {
10037 	uint64_t cur_lgrp_trthr_migrs = lgrp_get_trthr_migrations();
10038 
10039 	if (cur_lgrp_trthr_migrs != segvn_lgrp_trthr_migrs_snpsht) {
10040 		segvn_lgrp_trthr_migrs_snpsht = cur_lgrp_trthr_migrs;
10041 		sema_v(&segvn_trasync_sem);
10042 	}
10043 
10044 	if (!segvn_disable_textrepl_update &&
10045 	    segvn_update_textrepl_interval != 0) {
10046 		(void) timeout(segvn_trupdate_wakeup, dummy,
10047 		    segvn_update_textrepl_interval);
10048 	}
10049 }
10050 
10051 static void
10052 segvn_trupdate(void)
10053 {
10054 	ulong_t		hash;
10055 	svntr_t		*svntrp;
10056 	segvn_data_t	*svd;
10057 
10058 	ASSERT(svntr_hashtab != NULL);
10059 
10060 	for (hash = 0; hash < svntr_hashtab_sz; hash++) {
10061 		mutex_enter(&svntr_hashtab[hash].tr_lock);
10062 		svntrp = svntr_hashtab[hash].tr_head;
10063 		for (; svntrp != NULL; svntrp = svntrp->tr_next) {
10064 			ASSERT(svntrp->tr_refcnt != 0);
10065 			svd = svntrp->tr_svnhead;
10066 			for (; svd != NULL; svd = svd->svn_trnext) {
10067 				segvn_trupdate_seg(svd->seg, svd, svntrp,
10068 				    hash);
10069 			}
10070 		}
10071 		mutex_exit(&svntr_hashtab[hash].tr_lock);
10072 	}
10073 }
10074 
10075 static void
10076 segvn_trupdate_seg(struct seg *seg,
10077 	segvn_data_t *svd,
10078 	svntr_t *svntrp,
10079 	ulong_t hash)
10080 {
10081 	proc_t			*p;
10082 	lgrp_id_t		lgrp_id;
10083 	struct as		*as;
10084 	size_t			size;
10085 	struct anon_map		*amp;
10086 
10087 	ASSERT(svd->vp != NULL);
10088 	ASSERT(svd->vp == svntrp->tr_vp);
10089 	ASSERT(svd->offset == svntrp->tr_off);
10090 	ASSERT(svd->offset + seg->s_size == svntrp->tr_eoff);
10091 	ASSERT(seg != NULL);
10092 	ASSERT(svd->seg == seg);
10093 	ASSERT(seg->s_data == (void *)svd);
10094 	ASSERT(seg->s_szc == svntrp->tr_szc);
10095 	ASSERT(svd->tr_state == SEGVN_TR_ON);
10096 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
10097 	ASSERT(svd->amp != NULL);
10098 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
10099 	ASSERT(svd->tr_policy_info.mem_lgrpid != LGRP_NONE);
10100 	ASSERT(svd->tr_policy_info.mem_lgrpid < NLGRPS_MAX);
10101 	ASSERT(svntrp->tr_amp[svd->tr_policy_info.mem_lgrpid] == svd->amp);
10102 	ASSERT(svntrp->tr_refcnt != 0);
10103 	ASSERT(mutex_owned(&svntr_hashtab[hash].tr_lock));
10104 
10105 	as = seg->s_as;
10106 	ASSERT(as != NULL && as != &kas);
10107 	p = as->a_proc;
10108 	ASSERT(p != NULL);
10109 	ASSERT(p->p_tr_lgrpid != LGRP_NONE);
10110 	lgrp_id = p->p_t1_lgrpid;
10111 	if (lgrp_id == LGRP_NONE) {
10112 		return;
10113 	}
10114 	ASSERT(lgrp_id < NLGRPS_MAX);
10115 	if (svd->tr_policy_info.mem_lgrpid == lgrp_id) {
10116 		return;
10117 	}
10118 
10119 	/*
10120 	 * Use tryenter locking since we are locking as/seg and svntr hash
10121 	 * lock in reverse from syncrounous thread order.
10122 	 */
10123 	if (!AS_LOCK_TRYENTER(as, &as->a_lock, RW_READER)) {
10124 		SEGVN_TR_ADDSTAT(nolock);
10125 		if (segvn_lgrp_trthr_migrs_snpsht) {
10126 			segvn_lgrp_trthr_migrs_snpsht = 0;
10127 		}
10128 		return;
10129 	}
10130 	if (!SEGVN_LOCK_TRYENTER(seg->s_as, &svd->lock, RW_WRITER)) {
10131 		AS_LOCK_EXIT(as, &as->a_lock);
10132 		SEGVN_TR_ADDSTAT(nolock);
10133 		if (segvn_lgrp_trthr_migrs_snpsht) {
10134 			segvn_lgrp_trthr_migrs_snpsht = 0;
10135 		}
10136 		return;
10137 	}
10138 	size = seg->s_size;
10139 	if (svntrp->tr_amp[lgrp_id] == NULL) {
10140 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
10141 		if (trmem > segvn_textrepl_max_bytes) {
10142 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10143 			AS_LOCK_EXIT(as, &as->a_lock);
10144 			atomic_add_long(&segvn_textrepl_bytes, -size);
10145 			SEGVN_TR_ADDSTAT(normem);
10146 			return;
10147 		}
10148 		if (anon_try_resv_zone(size, NULL) == 0) {
10149 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10150 			AS_LOCK_EXIT(as, &as->a_lock);
10151 			atomic_add_long(&segvn_textrepl_bytes, -size);
10152 			SEGVN_TR_ADDSTAT(noanon);
10153 			return;
10154 		}
10155 		amp = anonmap_alloc(size, size, KM_NOSLEEP);
10156 		if (amp == NULL) {
10157 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10158 			AS_LOCK_EXIT(as, &as->a_lock);
10159 			atomic_add_long(&segvn_textrepl_bytes, -size);
10160 			anon_unresv_zone(size, NULL);
10161 			SEGVN_TR_ADDSTAT(nokmem);
10162 			return;
10163 		}
10164 		ASSERT(amp->refcnt == 1);
10165 		amp->a_szc = seg->s_szc;
10166 		svntrp->tr_amp[lgrp_id] = amp;
10167 	}
10168 	/*
10169 	 * We don't need to drop the bucket lock but here we give other
10170 	 * threads a chance.  svntr and svd can't be unlinked as long as
10171 	 * segment lock is held as a writer and AS held as well.  After we
10172 	 * retake bucket lock we'll continue from where we left. We'll be able
10173 	 * to reach the end of either list since new entries are always added
10174 	 * to the beginning of the lists.
10175 	 */
10176 	mutex_exit(&svntr_hashtab[hash].tr_lock);
10177 	hat_unload_callback(as->a_hat, seg->s_base, size, 0, NULL);
10178 	mutex_enter(&svntr_hashtab[hash].tr_lock);
10179 
10180 	ASSERT(svd->tr_state == SEGVN_TR_ON);
10181 	ASSERT(svd->amp != NULL);
10182 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
10183 	ASSERT(svd->tr_policy_info.mem_lgrpid != lgrp_id);
10184 	ASSERT(svd->amp != svntrp->tr_amp[lgrp_id]);
10185 
10186 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
10187 	svd->amp = svntrp->tr_amp[lgrp_id];
10188 	p->p_tr_lgrpid = NLGRPS_MAX;
10189 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10190 	AS_LOCK_EXIT(as, &as->a_lock);
10191 
10192 	ASSERT(svntrp->tr_refcnt != 0);
10193 	ASSERT(svd->vp == svntrp->tr_vp);
10194 	ASSERT(svd->tr_policy_info.mem_lgrpid == lgrp_id);
10195 	ASSERT(svd->amp != NULL && svd->amp == svntrp->tr_amp[lgrp_id]);
10196 	ASSERT(svd->seg == seg);
10197 	ASSERT(svd->tr_state == SEGVN_TR_ON);
10198 
10199 	SEGVN_TR_ADDSTAT(asyncrepl);
10200 }
10201