xref: /titanic_41/usr/src/uts/common/vm/seg_vn.c (revision 2bd43bc76740c0fef669e90cfcaa4e3f3cb2a2f5)
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 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
27 /*	  All Rights Reserved  	*/
28 
29 /*
30  * University Copyright- Copyright (c) 1982, 1986, 1988
31  * The Regents of the University of California
32  * All Rights Reserved
33  *
34  * University Acknowledgment- Portions of this document are derived from
35  * software developed by the University of California, Berkeley, and its
36  * contributors.
37  */
38 
39 #pragma ident	"%Z%%M%	%I%	%E% SMI"
40 
41 /*
42  * VM - shared or copy-on-write from a vnode/anonymous memory.
43  */
44 
45 #include <sys/types.h>
46 #include <sys/param.h>
47 #include <sys/t_lock.h>
48 #include <sys/errno.h>
49 #include <sys/systm.h>
50 #include <sys/mman.h>
51 #include <sys/debug.h>
52 #include <sys/cred.h>
53 #include <sys/vmsystm.h>
54 #include <sys/tuneable.h>
55 #include <sys/bitmap.h>
56 #include <sys/swap.h>
57 #include <sys/kmem.h>
58 #include <sys/sysmacros.h>
59 #include <sys/vtrace.h>
60 #include <sys/cmn_err.h>
61 #include <sys/callb.h>
62 #include <sys/vm.h>
63 #include <sys/dumphdr.h>
64 #include <sys/lgrp.h>
65 
66 #include <vm/hat.h>
67 #include <vm/as.h>
68 #include <vm/seg.h>
69 #include <vm/seg_vn.h>
70 #include <vm/pvn.h>
71 #include <vm/anon.h>
72 #include <vm/page.h>
73 #include <vm/vpage.h>
74 #include <sys/proc.h>
75 #include <sys/task.h>
76 #include <sys/project.h>
77 #include <sys/zone.h>
78 #include <sys/shm_impl.h>
79 /*
80  * Private seg op routines.
81  */
82 static int	segvn_dup(struct seg *seg, struct seg *newseg);
83 static int	segvn_unmap(struct seg *seg, caddr_t addr, size_t len);
84 static void	segvn_free(struct seg *seg);
85 static faultcode_t segvn_fault(struct hat *hat, struct seg *seg,
86 		    caddr_t addr, size_t len, enum fault_type type,
87 		    enum seg_rw rw);
88 static faultcode_t segvn_faulta(struct seg *seg, caddr_t addr);
89 static int	segvn_setprot(struct seg *seg, caddr_t addr,
90 		    size_t len, uint_t prot);
91 static int	segvn_checkprot(struct seg *seg, caddr_t addr,
92 		    size_t len, uint_t prot);
93 static int	segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta);
94 static size_t	segvn_swapout(struct seg *seg);
95 static int	segvn_sync(struct seg *seg, caddr_t addr, size_t len,
96 		    int attr, uint_t flags);
97 static size_t	segvn_incore(struct seg *seg, caddr_t addr, size_t len,
98 		    char *vec);
99 static int	segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
100 		    int attr, int op, ulong_t *lockmap, size_t pos);
101 static int	segvn_getprot(struct seg *seg, caddr_t addr, size_t len,
102 		    uint_t *protv);
103 static u_offset_t	segvn_getoffset(struct seg *seg, caddr_t addr);
104 static int	segvn_gettype(struct seg *seg, caddr_t addr);
105 static int	segvn_getvp(struct seg *seg, caddr_t addr,
106 		    struct vnode **vpp);
107 static int	segvn_advise(struct seg *seg, caddr_t addr, size_t len,
108 		    uint_t behav);
109 static void	segvn_dump(struct seg *seg);
110 static int	segvn_pagelock(struct seg *seg, caddr_t addr, size_t len,
111 		    struct page ***ppp, enum lock_type type, enum seg_rw rw);
112 static int	segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len,
113 		    uint_t szc);
114 static int	segvn_getmemid(struct seg *seg, caddr_t addr,
115 		    memid_t *memidp);
116 static lgrp_mem_policy_info_t	*segvn_getpolicy(struct seg *, caddr_t);
117 static int	segvn_capable(struct seg *seg, segcapability_t capable);
118 
119 struct	seg_ops segvn_ops = {
120 	segvn_dup,
121 	segvn_unmap,
122 	segvn_free,
123 	segvn_fault,
124 	segvn_faulta,
125 	segvn_setprot,
126 	segvn_checkprot,
127 	segvn_kluster,
128 	segvn_swapout,
129 	segvn_sync,
130 	segvn_incore,
131 	segvn_lockop,
132 	segvn_getprot,
133 	segvn_getoffset,
134 	segvn_gettype,
135 	segvn_getvp,
136 	segvn_advise,
137 	segvn_dump,
138 	segvn_pagelock,
139 	segvn_setpagesize,
140 	segvn_getmemid,
141 	segvn_getpolicy,
142 	segvn_capable,
143 };
144 
145 /*
146  * Common zfod structures, provided as a shorthand for others to use.
147  */
148 static segvn_crargs_t zfod_segvn_crargs =
149 	SEGVN_ZFOD_ARGS(PROT_ZFOD, PROT_ALL);
150 static segvn_crargs_t kzfod_segvn_crargs =
151 	SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_USER,
152 	PROT_ALL & ~PROT_USER);
153 static segvn_crargs_t stack_noexec_crargs =
154 	SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_EXEC, PROT_ALL);
155 
156 caddr_t	zfod_argsp = (caddr_t)&zfod_segvn_crargs;	/* user zfod argsp */
157 caddr_t	kzfod_argsp = (caddr_t)&kzfod_segvn_crargs;	/* kernel zfod argsp */
158 caddr_t	stack_exec_argsp = (caddr_t)&zfod_segvn_crargs;	/* executable stack */
159 caddr_t	stack_noexec_argsp = (caddr_t)&stack_noexec_crargs; /* noexec stack */
160 
161 #define	vpgtob(n)	((n) * sizeof (struct vpage))	/* For brevity */
162 
163 size_t	segvn_comb_thrshld = UINT_MAX;	/* patchable -- see 1196681 */
164 
165 static int	segvn_concat(struct seg *, struct seg *, int);
166 static int	segvn_extend_prev(struct seg *, struct seg *,
167 		    struct segvn_crargs *, size_t);
168 static int	segvn_extend_next(struct seg *, struct seg *,
169 		    struct segvn_crargs *, size_t);
170 static void	segvn_softunlock(struct seg *, caddr_t, size_t, enum seg_rw);
171 static void	segvn_pagelist_rele(page_t **);
172 static void	segvn_setvnode_mpss(vnode_t *);
173 static void	segvn_relocate_pages(page_t **, page_t *);
174 static int	segvn_full_szcpages(page_t **, uint_t, int *, uint_t *);
175 static int	segvn_fill_vp_pages(struct segvn_data *, vnode_t *, u_offset_t,
176     uint_t, page_t **, page_t **, uint_t *, int *);
177 static faultcode_t segvn_fault_vnodepages(struct hat *, struct seg *, caddr_t,
178     caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
179 static faultcode_t segvn_fault_anonpages(struct hat *, struct seg *, caddr_t,
180     caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
181 static faultcode_t segvn_faultpage(struct hat *, struct seg *, caddr_t,
182     u_offset_t, struct vpage *, page_t **, uint_t,
183     enum fault_type, enum seg_rw, int, int);
184 static void	segvn_vpage(struct seg *);
185 
186 static void segvn_purge(struct seg *seg);
187 static int segvn_reclaim(struct seg *, caddr_t, size_t, struct page **,
188     enum seg_rw);
189 
190 static int sameprot(struct seg *, caddr_t, size_t);
191 
192 static int segvn_demote_range(struct seg *, caddr_t, size_t, int, uint_t);
193 static int segvn_clrszc(struct seg *);
194 static struct seg *segvn_split_seg(struct seg *, caddr_t);
195 static int segvn_claim_pages(struct seg *, struct vpage *, u_offset_t,
196     ulong_t, uint_t);
197 
198 static void segvn_hat_rgn_unload_callback(caddr_t, caddr_t, caddr_t,
199     size_t, void *, u_offset_t);
200 
201 static int segvn_pp_lock_anonpages(page_t *, int);
202 static void segvn_pp_unlock_anonpages(page_t *, int);
203 
204 static struct kmem_cache *segvn_cache;
205 
206 #ifdef VM_STATS
207 static struct segvnvmstats_str {
208 	ulong_t	fill_vp_pages[31];
209 	ulong_t fltvnpages[49];
210 	ulong_t	fullszcpages[10];
211 	ulong_t	relocatepages[3];
212 	ulong_t	fltanpages[17];
213 	ulong_t pagelock[3];
214 	ulong_t	demoterange[3];
215 } segvnvmstats;
216 #endif /* VM_STATS */
217 
218 #define	SDR_RANGE	1		/* demote entire range */
219 #define	SDR_END		2		/* demote non aligned ends only */
220 
221 #define	CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr) {	    \
222 		if ((len) != 0) { 		      	      		      \
223 			lpgaddr = (caddr_t)P2ALIGN((uintptr_t)(addr), pgsz);  \
224 			ASSERT(lpgaddr >= (seg)->s_base);	      	      \
225 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)((addr) +    \
226 			    (len)), pgsz);				      \
227 			ASSERT(lpgeaddr > lpgaddr);		      	      \
228 			ASSERT(lpgeaddr <= (seg)->s_base + (seg)->s_size);    \
229 		} else {					      	      \
230 			lpgeaddr = lpgaddr = (addr);	      		      \
231 		}							      \
232 	}
233 
234 /*ARGSUSED*/
235 static int
236 segvn_cache_constructor(void *buf, void *cdrarg, int kmflags)
237 {
238 	struct segvn_data *svd = buf;
239 
240 	rw_init(&svd->lock, NULL, RW_DEFAULT, NULL);
241 	mutex_init(&svd->segp_slock, NULL, MUTEX_DEFAULT, NULL);
242 	svd->svn_trnext = svd->svn_trprev = NULL;
243 	return (0);
244 }
245 
246 /*ARGSUSED1*/
247 static void
248 segvn_cache_destructor(void *buf, void *cdrarg)
249 {
250 	struct segvn_data *svd = buf;
251 
252 	rw_destroy(&svd->lock);
253 	mutex_destroy(&svd->segp_slock);
254 }
255 
256 /*ARGSUSED*/
257 static int
258 svntr_cache_constructor(void *buf, void *cdrarg, int kmflags)
259 {
260 	bzero(buf, sizeof (svntr_t));
261 	return (0);
262 }
263 
264 /*
265  * Patching this variable to non-zero allows the system to run with
266  * stacks marked as "not executable".  It's a bit of a kludge, but is
267  * provided as a tweakable for platforms that export those ABIs
268  * (e.g. sparc V8) that have executable stacks enabled by default.
269  * There are also some restrictions for platforms that don't actually
270  * implement 'noexec' protections.
271  *
272  * Once enabled, the system is (therefore) unable to provide a fully
273  * ABI-compliant execution environment, though practically speaking,
274  * most everything works.  The exceptions are generally some interpreters
275  * and debuggers that create executable code on the stack and jump
276  * into it (without explicitly mprotecting the address range to include
277  * PROT_EXEC).
278  *
279  * One important class of applications that are disabled are those
280  * that have been transformed into malicious agents using one of the
281  * numerous "buffer overflow" attacks.  See 4007890.
282  */
283 int noexec_user_stack = 0;
284 int noexec_user_stack_log = 1;
285 
286 int segvn_lpg_disable = 0;
287 uint_t segvn_maxpgszc = 0;
288 
289 ulong_t segvn_vmpss_clrszc_cnt;
290 ulong_t segvn_vmpss_clrszc_err;
291 ulong_t segvn_fltvnpages_clrszc_cnt;
292 ulong_t segvn_fltvnpages_clrszc_err;
293 ulong_t segvn_setpgsz_align_err;
294 ulong_t segvn_setpgsz_anon_align_err;
295 ulong_t segvn_setpgsz_getattr_err;
296 ulong_t segvn_setpgsz_eof_err;
297 ulong_t segvn_faultvnmpss_align_err1;
298 ulong_t segvn_faultvnmpss_align_err2;
299 ulong_t segvn_faultvnmpss_align_err3;
300 ulong_t segvn_faultvnmpss_align_err4;
301 ulong_t segvn_faultvnmpss_align_err5;
302 ulong_t	segvn_vmpss_pageio_deadlk_err;
303 
304 int segvn_use_regions = 1;
305 
306 /*
307  * Segvn supports text replication optimization for NUMA platforms. Text
308  * replica's are represented by anon maps (amp). There's one amp per text file
309  * region per lgroup. A process chooses the amp for each of its text mappings
310  * based on the lgroup assignment of its main thread (t_tid = 1). All
311  * processes that want a replica on a particular lgroup for the same text file
312  * mapping share the same amp. amp's are looked up in svntr_hashtab hash table
313  * with vp,off,size,szc used as a key. Text replication segments are read only
314  * MAP_PRIVATE|MAP_TEXT segments that map vnode. Replication is achieved by
315  * forcing COW faults from vnode to amp and mapping amp pages instead of vnode
316  * pages. Replication amp is assigned to a segment when it gets its first
317  * pagefault. To handle main thread lgroup rehoming segvn_trasync_thread
318  * rechecks periodically if the process still maps an amp local to the main
319  * thread. If not async thread forces process to remap to an amp in the new
320  * home lgroup of the main thread. Current text replication implementation
321  * only provides the benefit to workloads that do most of their work in the
322  * main thread of a process or all the threads of a process run in the same
323  * lgroup. To extend text replication benefit to different types of
324  * multithreaded workloads further work would be needed in the hat layer to
325  * allow the same virtual address in the same hat to simultaneously map
326  * different physical addresses (i.e. page table replication would be needed
327  * for x86).
328  *
329  * amp pages are used instead of vnode pages as long as segment has a very
330  * simple life cycle.  It's created via segvn_create(), handles S_EXEC
331  * (S_READ) pagefaults and is fully unmapped.  If anything more complicated
332  * happens such as protection is changed, real COW fault happens, pagesize is
333  * changed, MC_LOCK is requested or segment is partially unmapped we turn off
334  * text replication by converting the segment back to vnode only segment
335  * (unmap segment's address range and set svd->amp to NULL).
336  *
337  * The original file can be changed after amp is inserted into
338  * svntr_hashtab. Processes that are launched after the file is already
339  * changed can't use the replica's created prior to the file change. To
340  * implement this functionality hash entries are timestamped. Replica's can
341  * only be used if current file modification time is the same as the timestamp
342  * saved when hash entry was created. However just timestamps alone are not
343  * sufficient to detect file modification via mmap(MAP_SHARED) mappings. We
344  * deal with file changes via MAP_SHARED mappings differently. When writable
345  * MAP_SHARED mappings are created to vnodes marked as executable we mark all
346  * existing replica's for this vnode as not usable for future text
347  * mappings. And we don't create new replica's for files that currently have
348  * potentially writable MAP_SHARED mappings (i.e. vn_is_mapped(V_WRITE) is
349  * true).
350  */
351 
352 #define	SEGVN_TEXTREPL_MAXBYTES_FACTOR	(20)
353 size_t	segvn_textrepl_max_bytes_factor = SEGVN_TEXTREPL_MAXBYTES_FACTOR;
354 
355 static ulong_t			svntr_hashtab_sz = 512;
356 static svntr_bucket_t		*svntr_hashtab = NULL;
357 static struct kmem_cache	*svntr_cache;
358 static svntr_stats_t		*segvn_textrepl_stats;
359 static ksema_t 			segvn_trasync_sem;
360 
361 int				segvn_disable_textrepl = 0;
362 size_t				textrepl_size_thresh = (size_t)-1;
363 size_t				segvn_textrepl_bytes = 0;
364 size_t				segvn_textrepl_max_bytes = 0;
365 clock_t				segvn_update_textrepl_interval = 0;
366 int				segvn_update_tr_time = 10;
367 int				segvn_disable_textrepl_update = 0;
368 
369 static void segvn_textrepl(struct seg *);
370 static void segvn_textunrepl(struct seg *, int);
371 static void segvn_inval_trcache(vnode_t *);
372 static void segvn_trasync_thread(void);
373 static void segvn_trupdate_wakeup(void *);
374 static void segvn_trupdate(void);
375 static void segvn_trupdate_seg(struct seg *, segvn_data_t *, svntr_t *,
376     ulong_t);
377 
378 /*
379  * Initialize segvn data structures
380  */
381 void
382 segvn_init(void)
383 {
384 	uint_t maxszc;
385 	uint_t szc;
386 	size_t pgsz;
387 
388 	segvn_cache = kmem_cache_create("segvn_cache",
389 		sizeof (struct segvn_data), 0,
390 		segvn_cache_constructor, segvn_cache_destructor, NULL,
391 		NULL, NULL, 0);
392 
393 	if (segvn_lpg_disable != 0)
394 		return;
395 	szc = maxszc = page_num_pagesizes() - 1;
396 	if (szc == 0) {
397 		segvn_lpg_disable = 1;
398 		return;
399 	}
400 	if (page_get_pagesize(0) != PAGESIZE) {
401 		panic("segvn_init: bad szc 0");
402 		/*NOTREACHED*/
403 	}
404 	while (szc != 0) {
405 		pgsz = page_get_pagesize(szc);
406 		if (pgsz <= PAGESIZE || !IS_P2ALIGNED(pgsz, pgsz)) {
407 			panic("segvn_init: bad szc %d", szc);
408 			/*NOTREACHED*/
409 		}
410 		szc--;
411 	}
412 	if (segvn_maxpgszc == 0 || segvn_maxpgszc > maxszc)
413 		segvn_maxpgszc = maxszc;
414 
415 	if (segvn_use_regions && !hat_supported(HAT_SHARED_REGIONS, NULL))
416 		segvn_use_regions = 0;
417 
418 	/*
419 	 * For now shared regions and text replication segvn support
420 	 * are mutually exclusive. This is acceptable because
421 	 * currently significant benefit from text replication was
422 	 * only observed on AMD64 NUMA platforms (due to relatively
423 	 * small L2$ size) and currently we don't support shared
424 	 * regions on x86.
425 	 */
426 	if (segvn_use_regions && !segvn_disable_textrepl) {
427 		segvn_disable_textrepl = 1;
428 	}
429 
430 	if (lgrp_optimizations() && textrepl_size_thresh != (size_t)-1 &&
431 	    !segvn_disable_textrepl) {
432 		ulong_t i;
433 		size_t hsz = svntr_hashtab_sz * sizeof (svntr_bucket_t);
434 
435 		svntr_cache = kmem_cache_create("svntr_cache",
436 		    sizeof (svntr_t), 0, svntr_cache_constructor, NULL,
437 		    NULL, NULL, NULL, 0);
438 		svntr_hashtab = kmem_zalloc(hsz, KM_SLEEP);
439 		for (i = 0; i < svntr_hashtab_sz; i++) {
440 			mutex_init(&svntr_hashtab[i].tr_lock,  NULL,
441 			    MUTEX_DEFAULT, NULL);
442 		}
443 		segvn_textrepl_max_bytes = ptob(physmem) /
444 		    segvn_textrepl_max_bytes_factor;
445 		segvn_textrepl_stats = kmem_zalloc(NCPU *
446 		    sizeof (svntr_stats_t), KM_SLEEP);
447 		sema_init(&segvn_trasync_sem, 0, NULL, SEMA_DEFAULT, NULL);
448 		(void) thread_create(NULL, 0, segvn_trasync_thread,
449 		    NULL, 0, &p0, TS_RUN, minclsyspri);
450 	}
451 }
452 
453 #define	SEGVN_PAGEIO	((void *)0x1)
454 #define	SEGVN_NOPAGEIO	((void *)0x2)
455 
456 static void
457 segvn_setvnode_mpss(vnode_t *vp)
458 {
459 	int err;
460 
461 	ASSERT(vp->v_mpssdata == NULL ||
462 	    vp->v_mpssdata == SEGVN_PAGEIO ||
463 	    vp->v_mpssdata == SEGVN_NOPAGEIO);
464 
465 	if (vp->v_mpssdata == NULL) {
466 		if (vn_vmpss_usepageio(vp)) {
467 			err = VOP_PAGEIO(vp, (page_t *)NULL,
468 			    (u_offset_t)0, 0, 0, CRED());
469 		} else {
470 			err = ENOSYS;
471 		}
472 		/*
473 		 * set v_mpssdata just once per vnode life
474 		 * so that it never changes.
475 		 */
476 		mutex_enter(&vp->v_lock);
477 		if (vp->v_mpssdata == NULL) {
478 			if (err == EINVAL) {
479 				vp->v_mpssdata = SEGVN_PAGEIO;
480 			} else {
481 				vp->v_mpssdata = SEGVN_NOPAGEIO;
482 			}
483 		}
484 		mutex_exit(&vp->v_lock);
485 	}
486 }
487 
488 int
489 segvn_create(struct seg *seg, void *argsp)
490 {
491 	struct segvn_crargs *a = (struct segvn_crargs *)argsp;
492 	struct segvn_data *svd;
493 	size_t swresv = 0;
494 	struct cred *cred;
495 	struct anon_map *amp;
496 	int error = 0;
497 	size_t pgsz;
498 	lgrp_mem_policy_t mpolicy = LGRP_MEM_POLICY_DEFAULT;
499 	int use_rgn = 0;
500 	int trok = 0;
501 
502 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
503 
504 	if (a->type != MAP_PRIVATE && a->type != MAP_SHARED) {
505 		panic("segvn_create type");
506 		/*NOTREACHED*/
507 	}
508 
509 	/*
510 	 * Check arguments.  If a shared anon structure is given then
511 	 * it is illegal to also specify a vp.
512 	 */
513 	if (a->amp != NULL && a->vp != NULL) {
514 		panic("segvn_create anon_map");
515 		/*NOTREACHED*/
516 	}
517 
518 	if (a->type == MAP_PRIVATE && (a->flags & MAP_TEXT) &&
519 	    a->vp != NULL && a->prot == (PROT_USER | PROT_READ | PROT_EXEC) &&
520 	    segvn_use_regions) {
521 		use_rgn = 1;
522 	}
523 
524 	/* MAP_NORESERVE on a MAP_SHARED segment is meaningless. */
525 	if (a->type == MAP_SHARED)
526 		a->flags &= ~MAP_NORESERVE;
527 
528 	if (a->szc != 0) {
529 		if (segvn_lpg_disable != 0 || (a->szc == AS_MAP_NO_LPOOB) ||
530 		    (a->amp != NULL && a->type == MAP_PRIVATE) ||
531 		    (a->flags & MAP_NORESERVE) || seg->s_as == &kas) {
532 			a->szc = 0;
533 		} else {
534 			if (a->szc > segvn_maxpgszc)
535 				a->szc = segvn_maxpgszc;
536 			pgsz = page_get_pagesize(a->szc);
537 			if (!IS_P2ALIGNED(seg->s_base, pgsz) ||
538 			    !IS_P2ALIGNED(seg->s_size, pgsz)) {
539 				a->szc = 0;
540 			} else if (a->vp != NULL) {
541 				extern struct vnode kvp;
542 				if (IS_SWAPFSVP(a->vp) || VN_ISKAS(a->vp)) {
543 					/*
544 					 * paranoid check.
545 					 * hat_page_demote() is not supported
546 					 * on swapfs pages.
547 					 */
548 					a->szc = 0;
549 				} else if (map_addr_vacalign_check(seg->s_base,
550 				    a->offset & PAGEMASK)) {
551 					a->szc = 0;
552 				}
553 			} else if (a->amp != NULL) {
554 				pgcnt_t anum = btopr(a->offset);
555 				pgcnt_t pgcnt = page_get_pagecnt(a->szc);
556 				if (!IS_P2ALIGNED(anum, pgcnt)) {
557 					a->szc = 0;
558 				}
559 			}
560 		}
561 	}
562 
563 	/*
564 	 * If segment may need private pages, reserve them now.
565 	 */
566 	if (!(a->flags & MAP_NORESERVE) && ((a->vp == NULL && a->amp == NULL) ||
567 	    (a->type == MAP_PRIVATE && (a->prot & PROT_WRITE)))) {
568 		if (anon_resv(seg->s_size) == 0)
569 			return (EAGAIN);
570 		swresv = seg->s_size;
571 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
572 			seg, swresv, 1);
573 	}
574 
575 	/*
576 	 * Reserve any mapping structures that may be required.
577 	 *
578 	 * Don't do it for segments that may use regions. It's currently a
579 	 * noop in the hat implementations anyway.
580 	 */
581 	if (!use_rgn) {
582 		hat_map(seg->s_as->a_hat, seg->s_base, seg->s_size, HAT_MAP);
583 	}
584 
585 	if (a->cred) {
586 		cred = a->cred;
587 		crhold(cred);
588 	} else {
589 		crhold(cred = CRED());
590 	}
591 
592 	/* Inform the vnode of the new mapping */
593 	if (a->vp != NULL) {
594 		error = VOP_ADDMAP(a->vp, a->offset & PAGEMASK,
595 		    seg->s_as, seg->s_base, seg->s_size, a->prot,
596 		    a->maxprot, a->type, cred);
597 		if (error) {
598 			if (swresv != 0) {
599 				anon_unresv(swresv);
600 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
601 					"anon proc:%p %lu %u",
602 					seg, swresv, 0);
603 			}
604 			crfree(cred);
605 			if (!use_rgn) {
606 				hat_unload(seg->s_as->a_hat, seg->s_base,
607 				    seg->s_size, HAT_UNLOAD_UNMAP);
608 			}
609 			return (error);
610 		}
611 		/*
612 		 * svntr_hashtab will be NULL if we support shared regions.
613 		 */
614 		trok = ((a->flags & MAP_TEXT) &&
615 		    (seg->s_size > textrepl_size_thresh ||
616 			(a->flags & _MAP_TEXTREPL)) &&
617 		    lgrp_optimizations() && svntr_hashtab != NULL &&
618 		    a->type == MAP_PRIVATE && swresv == 0 &&
619 		    !(a->flags & MAP_NORESERVE) &&
620 		    seg->s_as != &kas && a->vp->v_type == VREG);
621 
622 		ASSERT(!trok || !use_rgn);
623 	}
624 
625 	/*
626 	 * If more than one segment in the address space, and they're adjacent
627 	 * virtually, try to concatenate them.  Don't concatenate if an
628 	 * explicit anon_map structure was supplied (e.g., SystemV shared
629 	 * memory) or if we'll use text replication for this segment.
630 	 */
631 	if (a->amp == NULL && !use_rgn && !trok) {
632 		struct seg *pseg, *nseg;
633 		struct segvn_data *psvd, *nsvd;
634 		lgrp_mem_policy_t ppolicy, npolicy;
635 		uint_t	lgrp_mem_policy_flags = 0;
636 		extern lgrp_mem_policy_t lgrp_mem_default_policy;
637 
638 		/*
639 		 * Memory policy flags (lgrp_mem_policy_flags) is valid when
640 		 * extending stack/heap segments.
641 		 */
642 		if ((a->vp == NULL) && (a->type == MAP_PRIVATE) &&
643 		    !(a->flags & MAP_NORESERVE) && (seg->s_as != &kas)) {
644 			lgrp_mem_policy_flags = a->lgrp_mem_policy_flags;
645 		} else {
646 			/*
647 			 * Get policy when not extending it from another segment
648 			 */
649 			mpolicy = lgrp_mem_policy_default(seg->s_size, a->type);
650 		}
651 
652 		/*
653 		 * First, try to concatenate the previous and new segments
654 		 */
655 		pseg = AS_SEGPREV(seg->s_as, seg);
656 		if (pseg != NULL &&
657 		    pseg->s_base + pseg->s_size == seg->s_base &&
658 		    pseg->s_ops == &segvn_ops) {
659 			/*
660 			 * Get memory allocation policy from previous segment.
661 			 * When extension is specified (e.g. for heap) apply
662 			 * this policy to the new segment regardless of the
663 			 * outcome of segment concatenation.  Extension occurs
664 			 * for non-default policy otherwise default policy is
665 			 * used and is based on extended segment size.
666 			 */
667 			psvd = (struct segvn_data *)pseg->s_data;
668 			ppolicy = psvd->policy_info.mem_policy;
669 			if (lgrp_mem_policy_flags ==
670 			    LGRP_MP_FLAG_EXTEND_UP) {
671 				if (ppolicy != lgrp_mem_default_policy) {
672 					mpolicy = ppolicy;
673 				} else {
674 					mpolicy = lgrp_mem_policy_default(
675 					    pseg->s_size + seg->s_size,
676 					    a->type);
677 				}
678 			}
679 
680 			if (mpolicy == ppolicy &&
681 			    (pseg->s_size + seg->s_size <=
682 			    segvn_comb_thrshld || psvd->amp == NULL) &&
683 			    segvn_extend_prev(pseg, seg, a, swresv) == 0) {
684 				/*
685 				 * success! now try to concatenate
686 				 * with following seg
687 				 */
688 				crfree(cred);
689 				nseg = AS_SEGNEXT(pseg->s_as, pseg);
690 				if (nseg != NULL &&
691 				    nseg != pseg &&
692 				    nseg->s_ops == &segvn_ops &&
693 				    pseg->s_base + pseg->s_size ==
694 				    nseg->s_base)
695 					(void) segvn_concat(pseg, nseg, 0);
696 				ASSERT(pseg->s_szc == 0 ||
697 				    (a->szc == pseg->s_szc &&
698 				    IS_P2ALIGNED(pseg->s_base, pgsz) &&
699 				    IS_P2ALIGNED(pseg->s_size, pgsz)));
700 				return (0);
701 			}
702 		}
703 
704 		/*
705 		 * Failed, so try to concatenate with following seg
706 		 */
707 		nseg = AS_SEGNEXT(seg->s_as, seg);
708 		if (nseg != NULL &&
709 		    seg->s_base + seg->s_size == nseg->s_base &&
710 		    nseg->s_ops == &segvn_ops) {
711 			/*
712 			 * Get memory allocation policy from next segment.
713 			 * When extension is specified (e.g. for stack) apply
714 			 * this policy to the new segment regardless of the
715 			 * outcome of segment concatenation.  Extension occurs
716 			 * for non-default policy otherwise default policy is
717 			 * used and is based on extended segment size.
718 			 */
719 			nsvd = (struct segvn_data *)nseg->s_data;
720 			npolicy = nsvd->policy_info.mem_policy;
721 			if (lgrp_mem_policy_flags ==
722 			    LGRP_MP_FLAG_EXTEND_DOWN) {
723 				if (npolicy != lgrp_mem_default_policy) {
724 					mpolicy = npolicy;
725 				} else {
726 					mpolicy = lgrp_mem_policy_default(
727 					    nseg->s_size + seg->s_size,
728 					    a->type);
729 				}
730 			}
731 
732 			if (mpolicy == npolicy &&
733 			    segvn_extend_next(seg, nseg, a, swresv) == 0) {
734 				crfree(cred);
735 				ASSERT(nseg->s_szc == 0 ||
736 				    (a->szc == nseg->s_szc &&
737 				    IS_P2ALIGNED(nseg->s_base, pgsz) &&
738 				    IS_P2ALIGNED(nseg->s_size, pgsz)));
739 				return (0);
740 			}
741 		}
742 	}
743 
744 	if (a->vp != NULL) {
745 		VN_HOLD(a->vp);
746 		if (a->type == MAP_SHARED)
747 			lgrp_shm_policy_init(NULL, a->vp);
748 	}
749 	svd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
750 
751 	seg->s_ops = &segvn_ops;
752 	seg->s_data = (void *)svd;
753 	seg->s_szc = a->szc;
754 
755 	svd->seg = seg;
756 	svd->vp = a->vp;
757 	/*
758 	 * Anonymous mappings have no backing file so the offset is meaningless.
759 	 */
760 	svd->offset = a->vp ? (a->offset & PAGEMASK) : 0;
761 	svd->prot = a->prot;
762 	svd->maxprot = a->maxprot;
763 	svd->pageprot = 0;
764 	svd->type = a->type;
765 	svd->vpage = NULL;
766 	svd->cred = cred;
767 	svd->advice = MADV_NORMAL;
768 	svd->pageadvice = 0;
769 	svd->flags = (ushort_t)a->flags;
770 	svd->softlockcnt = 0;
771 	svd->rcookie = HAT_INVALID_REGION_COOKIE;
772 
773 	if (a->szc != 0 && a->vp != NULL) {
774 		segvn_setvnode_mpss(a->vp);
775 	}
776 	if (svd->type == MAP_SHARED && svd->vp != NULL &&
777 	    (svd->vp->v_flag & VVMEXEC) && (svd->prot & PROT_WRITE)) {
778 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
779 		segvn_inval_trcache(svd->vp);
780 	}
781 
782 	amp = a->amp;
783 	if ((svd->amp = amp) == NULL) {
784 		svd->anon_index = 0;
785 		if (svd->type == MAP_SHARED) {
786 			svd->swresv = 0;
787 			/*
788 			 * Shared mappings to a vp need no other setup.
789 			 * If we have a shared mapping to an anon_map object
790 			 * which hasn't been allocated yet,  allocate the
791 			 * struct now so that it will be properly shared
792 			 * by remembering the swap reservation there.
793 			 */
794 			if (a->vp == NULL) {
795 				svd->amp = anonmap_alloc(seg->s_size, swresv,
796 				    ANON_SLEEP);
797 				svd->amp->a_szc = seg->s_szc;
798 			}
799 		} else {
800 			/*
801 			 * Private mapping (with or without a vp).
802 			 * Allocate anon_map when needed.
803 			 */
804 			svd->swresv = swresv;
805 		}
806 	} else {
807 		pgcnt_t anon_num;
808 
809 		/*
810 		 * Mapping to an existing anon_map structure without a vp.
811 		 * For now we will insure that the segment size isn't larger
812 		 * than the size - offset gives us.  Later on we may wish to
813 		 * have the anon array dynamically allocated itself so that
814 		 * we don't always have to allocate all the anon pointer slots.
815 		 * This of course involves adding extra code to check that we
816 		 * aren't trying to use an anon pointer slot beyond the end
817 		 * of the currently allocated anon array.
818 		 */
819 		if ((amp->size - a->offset) < seg->s_size) {
820 			panic("segvn_create anon_map size");
821 			/*NOTREACHED*/
822 		}
823 
824 		anon_num = btopr(a->offset);
825 
826 		if (a->type == MAP_SHARED) {
827 			/*
828 			 * SHARED mapping to a given anon_map.
829 			 */
830 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
831 			amp->refcnt++;
832 			if (a->szc > amp->a_szc) {
833 				amp->a_szc = a->szc;
834 			}
835 			ANON_LOCK_EXIT(&amp->a_rwlock);
836 			svd->anon_index = anon_num;
837 			svd->swresv = 0;
838 		} else {
839 			/*
840 			 * PRIVATE mapping to a given anon_map.
841 			 * Make sure that all the needed anon
842 			 * structures are created (so that we will
843 			 * share the underlying pages if nothing
844 			 * is written by this mapping) and then
845 			 * duplicate the anon array as is done
846 			 * when a privately mapped segment is dup'ed.
847 			 */
848 			struct anon *ap;
849 			caddr_t addr;
850 			caddr_t eaddr;
851 			ulong_t	anon_idx;
852 			int hat_flag = HAT_LOAD;
853 
854 			if (svd->flags & MAP_TEXT) {
855 				hat_flag |= HAT_LOAD_TEXT;
856 			}
857 
858 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
859 			svd->amp->a_szc = seg->s_szc;
860 			svd->anon_index = 0;
861 			svd->swresv = swresv;
862 
863 			/*
864 			 * Prevent 2 threads from allocating anon
865 			 * slots simultaneously.
866 			 */
867 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
868 			eaddr = seg->s_base + seg->s_size;
869 
870 			for (anon_idx = anon_num, addr = seg->s_base;
871 			    addr < eaddr; addr += PAGESIZE, anon_idx++) {
872 				page_t *pp;
873 
874 				if ((ap = anon_get_ptr(amp->ahp,
875 				    anon_idx)) != NULL)
876 					continue;
877 
878 				/*
879 				 * Allocate the anon struct now.
880 				 * Might as well load up translation
881 				 * to the page while we're at it...
882 				 */
883 				pp = anon_zero(seg, addr, &ap, cred);
884 				if (ap == NULL || pp == NULL) {
885 					panic("segvn_create anon_zero");
886 					/*NOTREACHED*/
887 				}
888 
889 				/*
890 				 * Re-acquire the anon_map lock and
891 				 * initialize the anon array entry.
892 				 */
893 				ASSERT(anon_get_ptr(amp->ahp,
894 				    anon_idx) == NULL);
895 				(void) anon_set_ptr(amp->ahp, anon_idx, ap,
896 				    ANON_SLEEP);
897 
898 				ASSERT(seg->s_szc == 0);
899 				ASSERT(!IS_VMODSORT(pp->p_vnode));
900 
901 				ASSERT(use_rgn == 0);
902 				hat_memload(seg->s_as->a_hat, addr, pp,
903 					svd->prot & ~PROT_WRITE, hat_flag);
904 
905 				page_unlock(pp);
906 			}
907 			ASSERT(seg->s_szc == 0);
908 			anon_dup(amp->ahp, anon_num, svd->amp->ahp,
909 			    0, seg->s_size);
910 			ANON_LOCK_EXIT(&amp->a_rwlock);
911 		}
912 	}
913 
914 	/*
915 	 * Set default memory allocation policy for segment
916 	 *
917 	 * Always set policy for private memory at least for initialization
918 	 * even if this is a shared memory segment
919 	 */
920 	(void) lgrp_privm_policy_set(mpolicy, &svd->policy_info, seg->s_size);
921 
922 	if (svd->type == MAP_SHARED)
923 		(void) lgrp_shm_policy_set(mpolicy, svd->amp, svd->anon_index,
924 		    svd->vp, svd->offset, seg->s_size);
925 
926 	if (use_rgn) {
927 		ASSERT(!trok);
928 		ASSERT(svd->amp == NULL);
929 		svd->rcookie = hat_join_region(seg->s_as->a_hat, seg->s_base,
930 		    seg->s_size, (void *)svd->vp, svd->offset, svd->prot,
931 		    (uchar_t)seg->s_szc, segvn_hat_rgn_unload_callback,
932 		    HAT_REGION_TEXT);
933 	}
934 
935 	ASSERT(!trok || !(svd->prot & PROT_WRITE));
936 	svd->tr_state = trok ? SEGVN_TR_INIT : SEGVN_TR_OFF;
937 
938 	return (0);
939 }
940 
941 /*
942  * Concatenate two existing segments, if possible.
943  * Return 0 on success, -1 if two segments are not compatible
944  * or -2 on memory allocation failure.
945  * If amp_cat == 1 then try and concat segments with anon maps
946  */
947 static int
948 segvn_concat(struct seg *seg1, struct seg *seg2, int amp_cat)
949 {
950 	struct segvn_data *svd1 = seg1->s_data;
951 	struct segvn_data *svd2 = seg2->s_data;
952 	struct anon_map *amp1 = svd1->amp;
953 	struct anon_map *amp2 = svd2->amp;
954 	struct vpage *vpage1 = svd1->vpage;
955 	struct vpage *vpage2 = svd2->vpage, *nvpage = NULL;
956 	size_t size, nvpsize;
957 	pgcnt_t npages1, npages2;
958 
959 	ASSERT(seg1->s_as && seg2->s_as && seg1->s_as == seg2->s_as);
960 	ASSERT(AS_WRITE_HELD(seg1->s_as, &seg1->s_as->a_lock));
961 	ASSERT(seg1->s_ops == seg2->s_ops);
962 
963 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie) ||
964 	    HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
965 		return (-1);
966 	}
967 
968 	/* both segments exist, try to merge them */
969 #define	incompat(x)	(svd1->x != svd2->x)
970 	if (incompat(vp) || incompat(maxprot) ||
971 	    (!svd1->pageadvice && !svd2->pageadvice && incompat(advice)) ||
972 	    (!svd1->pageprot && !svd2->pageprot && incompat(prot)) ||
973 	    incompat(type) || incompat(cred) || incompat(flags) ||
974 	    seg1->s_szc != seg2->s_szc || incompat(policy_info.mem_policy) ||
975 	    (svd2->softlockcnt > 0))
976 		return (-1);
977 #undef incompat
978 
979 	/*
980 	 * vp == NULL implies zfod, offset doesn't matter
981 	 */
982 	if (svd1->vp != NULL &&
983 	    svd1->offset + seg1->s_size != svd2->offset) {
984 		return (-1);
985 	}
986 
987 	/*
988 	 * Don't concatenate if either segment uses text replication.
989 	 */
990 	if (svd1->tr_state != SEGVN_TR_OFF || svd2->tr_state != SEGVN_TR_OFF) {
991 		return (-1);
992 	}
993 
994 	/*
995 	 * Fail early if we're not supposed to concatenate
996 	 * segments with non NULL amp.
997 	 */
998 	if (amp_cat == 0 && (amp1 != NULL || amp2 != NULL)) {
999 		return (-1);
1000 	}
1001 
1002 	if (svd1->vp == NULL && svd1->type == MAP_SHARED) {
1003 		if (amp1 != amp2) {
1004 			return (-1);
1005 		}
1006 		if (amp1 != NULL && svd1->anon_index + btop(seg1->s_size) !=
1007 		    svd2->anon_index) {
1008 			return (-1);
1009 		}
1010 		ASSERT(amp1 == NULL || amp1->refcnt >= 2);
1011 	}
1012 
1013 	/*
1014 	 * If either seg has vpages, create a new merged vpage array.
1015 	 */
1016 	if (vpage1 != NULL || vpage2 != NULL) {
1017 		struct vpage *vp;
1018 
1019 		npages1 = seg_pages(seg1);
1020 		npages2 = seg_pages(seg2);
1021 		nvpsize = vpgtob(npages1 + npages2);
1022 
1023 		if ((nvpage = kmem_zalloc(nvpsize, KM_NOSLEEP)) == NULL) {
1024 			return (-2);
1025 		}
1026 
1027 		if (vpage1 != NULL) {
1028 			bcopy(vpage1, nvpage, vpgtob(npages1));
1029 		} else {
1030 			for (vp = nvpage; vp < nvpage + npages1; vp++) {
1031 				VPP_SETPROT(vp, svd1->prot);
1032 				VPP_SETADVICE(vp, svd1->advice);
1033 			}
1034 		}
1035 
1036 		if (vpage2 != NULL) {
1037 			bcopy(vpage2, nvpage + npages1, vpgtob(npages2));
1038 		} else {
1039 			for (vp = nvpage + npages1;
1040 			    vp < nvpage + npages1 + npages2; vp++) {
1041 				VPP_SETPROT(vp, svd2->prot);
1042 				VPP_SETADVICE(vp, svd2->advice);
1043 			}
1044 		}
1045 	}
1046 
1047 	/*
1048 	 * If either segment has private pages, create a new merged anon
1049 	 * array. If mergeing shared anon segments just decrement anon map's
1050 	 * refcnt.
1051 	 */
1052 	if (amp1 != NULL && svd1->type == MAP_SHARED) {
1053 		ASSERT(amp1 == amp2 && svd1->vp == NULL);
1054 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1055 		ASSERT(amp1->refcnt >= 2);
1056 		amp1->refcnt--;
1057 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1058 		svd2->amp = NULL;
1059 	} else if (amp1 != NULL || amp2 != NULL) {
1060 		struct anon_hdr *nahp;
1061 		struct anon_map *namp = NULL;
1062 		size_t asize;
1063 
1064 		ASSERT(svd1->type == MAP_PRIVATE);
1065 
1066 		asize = seg1->s_size + seg2->s_size;
1067 		if ((nahp = anon_create(btop(asize), ANON_NOSLEEP)) == NULL) {
1068 			if (nvpage != NULL) {
1069 				kmem_free(nvpage, nvpsize);
1070 			}
1071 			return (-2);
1072 		}
1073 		if (amp1 != NULL) {
1074 			/*
1075 			 * XXX anon rwlock is not really needed because
1076 			 * this is a private segment and we are writers.
1077 			 */
1078 			ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1079 			ASSERT(amp1->refcnt == 1);
1080 			if (anon_copy_ptr(amp1->ahp, svd1->anon_index,
1081 			    nahp, 0, btop(seg1->s_size), ANON_NOSLEEP)) {
1082 				anon_release(nahp, btop(asize));
1083 				ANON_LOCK_EXIT(&amp1->a_rwlock);
1084 				if (nvpage != NULL) {
1085 					kmem_free(nvpage, nvpsize);
1086 				}
1087 				return (-2);
1088 			}
1089 		}
1090 		if (amp2 != NULL) {
1091 			ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1092 			ASSERT(amp2->refcnt == 1);
1093 			if (anon_copy_ptr(amp2->ahp, svd2->anon_index,
1094 			    nahp, btop(seg1->s_size), btop(seg2->s_size),
1095 			    ANON_NOSLEEP)) {
1096 				anon_release(nahp, btop(asize));
1097 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1098 				if (amp1 != NULL) {
1099 					ANON_LOCK_EXIT(&amp1->a_rwlock);
1100 				}
1101 				if (nvpage != NULL) {
1102 					kmem_free(nvpage, nvpsize);
1103 				}
1104 				return (-2);
1105 			}
1106 		}
1107 		if (amp1 != NULL) {
1108 			namp = amp1;
1109 			anon_release(amp1->ahp, btop(amp1->size));
1110 		}
1111 		if (amp2 != NULL) {
1112 			if (namp == NULL) {
1113 				ASSERT(amp1 == NULL);
1114 				namp = amp2;
1115 				anon_release(amp2->ahp, btop(amp2->size));
1116 			} else {
1117 				amp2->refcnt--;
1118 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1119 				anonmap_free(amp2);
1120 			}
1121 			svd2->amp = NULL; /* needed for seg_free */
1122 		}
1123 		namp->ahp = nahp;
1124 		namp->size = asize;
1125 		svd1->amp = namp;
1126 		svd1->anon_index = 0;
1127 		ANON_LOCK_EXIT(&namp->a_rwlock);
1128 	}
1129 	/*
1130 	 * Now free the old vpage structures.
1131 	 */
1132 	if (nvpage != NULL) {
1133 		if (vpage1 != NULL) {
1134 			kmem_free(vpage1, vpgtob(npages1));
1135 		}
1136 		if (vpage2 != NULL) {
1137 			svd2->vpage = NULL;
1138 			kmem_free(vpage2, vpgtob(npages2));
1139 		}
1140 		if (svd2->pageprot) {
1141 			svd1->pageprot = 1;
1142 		}
1143 		if (svd2->pageadvice) {
1144 			svd1->pageadvice = 1;
1145 		}
1146 		svd1->vpage = nvpage;
1147 	}
1148 
1149 	/* all looks ok, merge segments */
1150 	svd1->swresv += svd2->swresv;
1151 	svd2->swresv = 0;  /* so seg_free doesn't release swap space */
1152 	size = seg2->s_size;
1153 	seg_free(seg2);
1154 	seg1->s_size += size;
1155 	return (0);
1156 }
1157 
1158 /*
1159  * Extend the previous segment (seg1) to include the
1160  * new segment (seg2 + a), if possible.
1161  * Return 0 on success.
1162  */
1163 static int
1164 segvn_extend_prev(seg1, seg2, a, swresv)
1165 	struct seg *seg1, *seg2;
1166 	struct segvn_crargs *a;
1167 	size_t swresv;
1168 {
1169 	struct segvn_data *svd1 = (struct segvn_data *)seg1->s_data;
1170 	size_t size;
1171 	struct anon_map *amp1;
1172 	struct vpage *new_vpage;
1173 
1174 	/*
1175 	 * We don't need any segment level locks for "segvn" data
1176 	 * since the address space is "write" locked.
1177 	 */
1178 	ASSERT(seg1->s_as && AS_WRITE_HELD(seg1->s_as, &seg1->s_as->a_lock));
1179 
1180 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie)) {
1181 		return (-1);
1182 	}
1183 
1184 	/* second segment is new, try to extend first */
1185 	/* XXX - should also check cred */
1186 	if (svd1->vp != a->vp || svd1->maxprot != a->maxprot ||
1187 	    (!svd1->pageprot && (svd1->prot != a->prot)) ||
1188 	    svd1->type != a->type || svd1->flags != a->flags ||
1189 	    seg1->s_szc != a->szc)
1190 		return (-1);
1191 
1192 	/* vp == NULL implies zfod, offset doesn't matter */
1193 	if (svd1->vp != NULL &&
1194 	    svd1->offset + seg1->s_size != (a->offset & PAGEMASK))
1195 		return (-1);
1196 
1197 	if (svd1->tr_state != SEGVN_TR_OFF) {
1198 		return (-1);
1199 	}
1200 
1201 	amp1 = svd1->amp;
1202 	if (amp1) {
1203 		pgcnt_t newpgs;
1204 
1205 		/*
1206 		 * Segment has private pages, can data structures
1207 		 * be expanded?
1208 		 *
1209 		 * Acquire the anon_map lock to prevent it from changing,
1210 		 * if it is shared.  This ensures that the anon_map
1211 		 * will not change while a thread which has a read/write
1212 		 * lock on an address space references it.
1213 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1214 		 * is 1.
1215 		 *
1216 		 * Can't grow a MAP_SHARED segment with an anonmap because
1217 		 * there may be existing anon slots where we want to extend
1218 		 * the segment and we wouldn't know what to do with them
1219 		 * (e.g., for tmpfs right thing is to just leave them there,
1220 		 * for /dev/zero they should be cleared out).
1221 		 */
1222 		if (svd1->type == MAP_SHARED)
1223 			return (-1);
1224 
1225 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1226 		if (amp1->refcnt > 1) {
1227 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1228 			return (-1);
1229 		}
1230 		newpgs = anon_grow(amp1->ahp, &svd1->anon_index,
1231 		    btop(seg1->s_size), btop(seg2->s_size), ANON_NOSLEEP);
1232 
1233 		if (newpgs == 0) {
1234 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1235 			return (-1);
1236 		}
1237 		amp1->size = ptob(newpgs);
1238 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1239 	}
1240 	if (svd1->vpage != NULL) {
1241 		struct vpage *vp, *evp;
1242 		new_vpage =
1243 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1244 			KM_NOSLEEP);
1245 		if (new_vpage == NULL)
1246 			return (-1);
1247 		bcopy(svd1->vpage, new_vpage, vpgtob(seg_pages(seg1)));
1248 		kmem_free(svd1->vpage, vpgtob(seg_pages(seg1)));
1249 		svd1->vpage = new_vpage;
1250 
1251 		vp = new_vpage + seg_pages(seg1);
1252 		evp = vp + seg_pages(seg2);
1253 		for (; vp < evp; vp++)
1254 			VPP_SETPROT(vp, a->prot);
1255 	}
1256 	size = seg2->s_size;
1257 	seg_free(seg2);
1258 	seg1->s_size += size;
1259 	svd1->swresv += swresv;
1260 	if (svd1->pageprot && (a->prot & PROT_WRITE) &&
1261 	    svd1->type == MAP_SHARED && svd1->vp != NULL &&
1262 	    (svd1->vp->v_flag & VVMEXEC)) {
1263 		ASSERT(vn_is_mapped(svd1->vp, V_WRITE));
1264 		segvn_inval_trcache(svd1->vp);
1265 	}
1266 	return (0);
1267 }
1268 
1269 /*
1270  * Extend the next segment (seg2) to include the
1271  * new segment (seg1 + a), if possible.
1272  * Return 0 on success.
1273  */
1274 static int
1275 segvn_extend_next(
1276 	struct seg *seg1,
1277 	struct seg *seg2,
1278 	struct segvn_crargs *a,
1279 	size_t swresv)
1280 {
1281 	struct segvn_data *svd2 = (struct segvn_data *)seg2->s_data;
1282 	size_t size;
1283 	struct anon_map *amp2;
1284 	struct vpage *new_vpage;
1285 
1286 	/*
1287 	 * We don't need any segment level locks for "segvn" data
1288 	 * since the address space is "write" locked.
1289 	 */
1290 	ASSERT(seg2->s_as && AS_WRITE_HELD(seg2->s_as, &seg2->s_as->a_lock));
1291 
1292 	if (HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1293 		return (-1);
1294 	}
1295 
1296 	/* first segment is new, try to extend second */
1297 	/* XXX - should also check cred */
1298 	if (svd2->vp != a->vp || svd2->maxprot != a->maxprot ||
1299 	    (!svd2->pageprot && (svd2->prot != a->prot)) ||
1300 	    svd2->type != a->type || svd2->flags != a->flags ||
1301 	    seg2->s_szc != a->szc)
1302 		return (-1);
1303 	/* vp == NULL implies zfod, offset doesn't matter */
1304 	if (svd2->vp != NULL &&
1305 	    (a->offset & PAGEMASK) + seg1->s_size != svd2->offset)
1306 		return (-1);
1307 
1308 	if (svd2->tr_state != SEGVN_TR_OFF) {
1309 		return (-1);
1310 	}
1311 
1312 	amp2 = svd2->amp;
1313 	if (amp2) {
1314 		pgcnt_t newpgs;
1315 
1316 		/*
1317 		 * Segment has private pages, can data structures
1318 		 * be expanded?
1319 		 *
1320 		 * Acquire the anon_map lock to prevent it from changing,
1321 		 * if it is shared.  This ensures that the anon_map
1322 		 * will not change while a thread which has a read/write
1323 		 * lock on an address space references it.
1324 		 *
1325 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1326 		 * is 1.
1327 		 */
1328 		if (svd2->type == MAP_SHARED)
1329 			return (-1);
1330 
1331 		ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1332 		if (amp2->refcnt > 1) {
1333 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1334 			return (-1);
1335 		}
1336 		newpgs = anon_grow(amp2->ahp, &svd2->anon_index,
1337 		    btop(seg2->s_size), btop(seg1->s_size),
1338 		    ANON_NOSLEEP | ANON_GROWDOWN);
1339 
1340 		if (newpgs == 0) {
1341 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1342 			return (-1);
1343 		}
1344 		amp2->size = ptob(newpgs);
1345 		ANON_LOCK_EXIT(&amp2->a_rwlock);
1346 	}
1347 	if (svd2->vpage != NULL) {
1348 		struct vpage *vp, *evp;
1349 		new_vpage =
1350 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1351 			KM_NOSLEEP);
1352 		if (new_vpage == NULL) {
1353 			/* Not merging segments so adjust anon_index back */
1354 			if (amp2)
1355 				svd2->anon_index += seg_pages(seg1);
1356 			return (-1);
1357 		}
1358 		bcopy(svd2->vpage, new_vpage + seg_pages(seg1),
1359 		    vpgtob(seg_pages(seg2)));
1360 		kmem_free(svd2->vpage, vpgtob(seg_pages(seg2)));
1361 		svd2->vpage = new_vpage;
1362 
1363 		vp = new_vpage;
1364 		evp = vp + seg_pages(seg1);
1365 		for (; vp < evp; vp++)
1366 			VPP_SETPROT(vp, a->prot);
1367 	}
1368 	size = seg1->s_size;
1369 	seg_free(seg1);
1370 	seg2->s_size += size;
1371 	seg2->s_base -= size;
1372 	svd2->offset -= size;
1373 	svd2->swresv += swresv;
1374 	if (svd2->pageprot && (a->prot & PROT_WRITE) &&
1375 	    svd2->type == MAP_SHARED && svd2->vp != NULL &&
1376 	    (svd2->vp->v_flag & VVMEXEC)) {
1377 		ASSERT(vn_is_mapped(svd2->vp, V_WRITE));
1378 		segvn_inval_trcache(svd2->vp);
1379 	}
1380 	return (0);
1381 }
1382 
1383 static int
1384 segvn_dup(struct seg *seg, struct seg *newseg)
1385 {
1386 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1387 	struct segvn_data *newsvd;
1388 	pgcnt_t npages = seg_pages(seg);
1389 	int error = 0;
1390 	uint_t prot;
1391 	size_t len;
1392 	struct anon_map *amp;
1393 
1394 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
1395 
1396 	/*
1397 	 * If segment has anon reserved, reserve more for the new seg.
1398 	 * For a MAP_NORESERVE segment swresv will be a count of all the
1399 	 * allocated anon slots; thus we reserve for the child as many slots
1400 	 * as the parent has allocated. This semantic prevents the child or
1401 	 * parent from dieing during a copy-on-write fault caused by trying
1402 	 * to write a shared pre-existing anon page.
1403 	 */
1404 	if ((len = svd->swresv) != 0) {
1405 		if (anon_resv(svd->swresv) == 0)
1406 			return (ENOMEM);
1407 
1408 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
1409 			seg, len, 0);
1410 	}
1411 
1412 	newsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
1413 
1414 	newseg->s_ops = &segvn_ops;
1415 	newseg->s_data = (void *)newsvd;
1416 	newseg->s_szc = seg->s_szc;
1417 
1418 	newsvd->seg = newseg;
1419 	if ((newsvd->vp = svd->vp) != NULL) {
1420 		VN_HOLD(svd->vp);
1421 		if (svd->type == MAP_SHARED)
1422 			lgrp_shm_policy_init(NULL, svd->vp);
1423 	}
1424 	newsvd->offset = svd->offset;
1425 	newsvd->prot = svd->prot;
1426 	newsvd->maxprot = svd->maxprot;
1427 	newsvd->pageprot = svd->pageprot;
1428 	newsvd->type = svd->type;
1429 	newsvd->cred = svd->cred;
1430 	crhold(newsvd->cred);
1431 	newsvd->advice = svd->advice;
1432 	newsvd->pageadvice = svd->pageadvice;
1433 	newsvd->swresv = svd->swresv;
1434 	newsvd->flags = svd->flags;
1435 	newsvd->softlockcnt = 0;
1436 	newsvd->policy_info = svd->policy_info;
1437 	newsvd->rcookie = HAT_INVALID_REGION_COOKIE;
1438 
1439 	if ((amp = svd->amp) == NULL || svd->tr_state == SEGVN_TR_ON) {
1440 		/*
1441 		 * Not attaching to a shared anon object.
1442 		 */
1443 		ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie) ||
1444 		    svd->tr_state == SEGVN_TR_OFF);
1445 		if (svd->tr_state == SEGVN_TR_ON) {
1446 			ASSERT(newsvd->vp != NULL && amp != NULL);
1447 			newsvd->tr_state = SEGVN_TR_INIT;
1448 		} else {
1449 			newsvd->tr_state = svd->tr_state;
1450 		}
1451 		newsvd->amp = NULL;
1452 		newsvd->anon_index = 0;
1453 	} else {
1454 		/* regions for now are only used on pure vnode segments */
1455 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
1456 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1457 		newsvd->tr_state = SEGVN_TR_OFF;
1458 		if (svd->type == MAP_SHARED) {
1459 			newsvd->amp = amp;
1460 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1461 			amp->refcnt++;
1462 			ANON_LOCK_EXIT(&amp->a_rwlock);
1463 			newsvd->anon_index = svd->anon_index;
1464 		} else {
1465 			int reclaim = 1;
1466 
1467 			/*
1468 			 * Allocate and initialize new anon_map structure.
1469 			 */
1470 			newsvd->amp = anonmap_alloc(newseg->s_size, 0,
1471 			    ANON_SLEEP);
1472 			newsvd->amp->a_szc = newseg->s_szc;
1473 			newsvd->anon_index = 0;
1474 
1475 			/*
1476 			 * We don't have to acquire the anon_map lock
1477 			 * for the new segment (since it belongs to an
1478 			 * address space that is still not associated
1479 			 * with any process), or the segment in the old
1480 			 * address space (since all threads in it
1481 			 * are stopped while duplicating the address space).
1482 			 */
1483 
1484 			/*
1485 			 * The goal of the following code is to make sure that
1486 			 * softlocked pages do not end up as copy on write
1487 			 * pages.  This would cause problems where one
1488 			 * thread writes to a page that is COW and a different
1489 			 * thread in the same process has softlocked it.  The
1490 			 * softlock lock would move away from this process
1491 			 * because the write would cause this process to get
1492 			 * a copy (without the softlock).
1493 			 *
1494 			 * The strategy here is to just break the
1495 			 * sharing on pages that could possibly be
1496 			 * softlocked.
1497 			 */
1498 retry:
1499 			if (svd->softlockcnt) {
1500 				struct anon *ap, *newap;
1501 				size_t i;
1502 				uint_t vpprot;
1503 				page_t *anon_pl[1+1], *pp;
1504 				caddr_t addr;
1505 				ulong_t old_idx = svd->anon_index;
1506 				ulong_t new_idx = 0;
1507 
1508 				/*
1509 				 * The softlock count might be non zero
1510 				 * because some pages are still stuck in the
1511 				 * cache for lazy reclaim. Flush the cache
1512 				 * now. This should drop the count to zero.
1513 				 * [or there is really I/O going on to these
1514 				 * pages]. Note, we have the writers lock so
1515 				 * nothing gets inserted during the flush.
1516 				 */
1517 				if (reclaim == 1) {
1518 					segvn_purge(seg);
1519 					reclaim = 0;
1520 					goto retry;
1521 				}
1522 				i = btopr(seg->s_size);
1523 				addr = seg->s_base;
1524 				/*
1525 				 * XXX break cow sharing using PAGESIZE
1526 				 * pages. They will be relocated into larger
1527 				 * pages at fault time.
1528 				 */
1529 				while (i-- > 0) {
1530 					if (ap = anon_get_ptr(amp->ahp,
1531 					    old_idx)) {
1532 						error = anon_getpage(&ap,
1533 						    &vpprot, anon_pl, PAGESIZE,
1534 						    seg, addr, S_READ,
1535 						    svd->cred);
1536 						if (error) {
1537 							newsvd->vpage = NULL;
1538 							goto out;
1539 						}
1540 						/*
1541 						 * prot need not be computed
1542 						 * below 'cause anon_private is
1543 						 * going to ignore it anyway
1544 						 * as child doesn't inherit
1545 						 * pagelock from parent.
1546 						 */
1547 						prot = svd->pageprot ?
1548 						    VPP_PROT(
1549 						    &svd->vpage[
1550 						    seg_page(seg, addr)])
1551 						    : svd->prot;
1552 						pp = anon_private(&newap,
1553 						    newseg, addr, prot,
1554 						    anon_pl[0],	0,
1555 						    newsvd->cred);
1556 						if (pp == NULL) {
1557 							/* no mem abort */
1558 							newsvd->vpage = NULL;
1559 							error = ENOMEM;
1560 							goto out;
1561 						}
1562 						(void) anon_set_ptr(
1563 						    newsvd->amp->ahp, new_idx,
1564 						    newap, ANON_SLEEP);
1565 						page_unlock(pp);
1566 					}
1567 					addr += PAGESIZE;
1568 					old_idx++;
1569 					new_idx++;
1570 				}
1571 			} else {	/* common case */
1572 				if (seg->s_szc != 0) {
1573 					/*
1574 					 * If at least one of anon slots of a
1575 					 * large page exists then make sure
1576 					 * all anon slots of a large page
1577 					 * exist to avoid partial cow sharing
1578 					 * of a large page in the future.
1579 					 */
1580 					anon_dup_fill_holes(amp->ahp,
1581 					    svd->anon_index, newsvd->amp->ahp,
1582 					    0, seg->s_size, seg->s_szc,
1583 					    svd->vp != NULL);
1584 				} else {
1585 					anon_dup(amp->ahp, svd->anon_index,
1586 					    newsvd->amp->ahp, 0, seg->s_size);
1587 				}
1588 
1589 				hat_clrattr(seg->s_as->a_hat, seg->s_base,
1590 				    seg->s_size, PROT_WRITE);
1591 			}
1592 		}
1593 	}
1594 	/*
1595 	 * If necessary, create a vpage structure for the new segment.
1596 	 * Do not copy any page lock indications.
1597 	 */
1598 	if (svd->vpage != NULL) {
1599 		uint_t i;
1600 		struct vpage *ovp = svd->vpage;
1601 		struct vpage *nvp;
1602 
1603 		nvp = newsvd->vpage =
1604 		    kmem_alloc(vpgtob(npages), KM_SLEEP);
1605 		for (i = 0; i < npages; i++) {
1606 			*nvp = *ovp++;
1607 			VPP_CLRPPLOCK(nvp++);
1608 		}
1609 	} else
1610 		newsvd->vpage = NULL;
1611 
1612 	/* Inform the vnode of the new mapping */
1613 	if (newsvd->vp != NULL) {
1614 		error = VOP_ADDMAP(newsvd->vp, (offset_t)newsvd->offset,
1615 		    newseg->s_as, newseg->s_base, newseg->s_size, newsvd->prot,
1616 		    newsvd->maxprot, newsvd->type, newsvd->cred);
1617 	}
1618 out:
1619 	if (error == 0 && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1620 		ASSERT(newsvd->amp == NULL);
1621 		ASSERT(newsvd->tr_state == SEGVN_TR_OFF);
1622 		newsvd->rcookie = svd->rcookie;
1623 		hat_dup_region(newseg->s_as->a_hat, newsvd->rcookie);
1624 	}
1625 	return (error);
1626 }
1627 
1628 
1629 /*
1630  * callback function used by segvn_unmap to invoke free_vp_pages() for only
1631  * those pages actually processed by the HAT
1632  */
1633 extern int free_pages;
1634 
1635 static void
1636 segvn_hat_rgn_unload_callback(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
1637     size_t r_size, void *r_obj, u_offset_t r_objoff)
1638 {
1639 	u_offset_t off;
1640 	size_t len;
1641 	vnode_t *vp = (vnode_t *)r_obj;
1642 
1643 	ASSERT(eaddr > saddr);
1644 	ASSERT(saddr >= r_saddr);
1645 	ASSERT(saddr < r_saddr + r_size);
1646 	ASSERT(eaddr > r_saddr);
1647 	ASSERT(eaddr <= r_saddr + r_size);
1648 	ASSERT(vp != NULL);
1649 
1650 	if (!free_pages) {
1651 		return;
1652 	}
1653 
1654 	len = eaddr - saddr;
1655 	off = (saddr - r_saddr) + r_objoff;
1656 	free_vp_pages(vp, off, len);
1657 }
1658 
1659 static void
1660 segvn_hat_unload_callback(hat_callback_t *cb)
1661 {
1662 	struct seg		*seg = cb->hcb_data;
1663 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
1664 	size_t			len;
1665 	u_offset_t		off;
1666 
1667 	ASSERT(svd->vp != NULL);
1668 	ASSERT(cb->hcb_end_addr > cb->hcb_start_addr);
1669 	ASSERT(cb->hcb_start_addr >= seg->s_base);
1670 
1671 	len = cb->hcb_end_addr - cb->hcb_start_addr;
1672 	off = cb->hcb_start_addr - seg->s_base;
1673 	free_vp_pages(svd->vp, svd->offset + off, len);
1674 }
1675 
1676 static int
1677 segvn_unmap(struct seg *seg, caddr_t addr, size_t len)
1678 {
1679 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1680 	struct segvn_data *nsvd;
1681 	struct seg *nseg;
1682 	struct anon_map *amp;
1683 	pgcnt_t	opages;		/* old segment size in pages */
1684 	pgcnt_t	npages;		/* new segment size in pages */
1685 	pgcnt_t	dpages;		/* pages being deleted (unmapped) */
1686 	hat_callback_t callback;	/* used for free_vp_pages() */
1687 	hat_callback_t *cbp = NULL;
1688 	caddr_t nbase;
1689 	size_t nsize;
1690 	size_t oswresv;
1691 	int reclaim = 1;
1692 
1693 	/*
1694 	 * We don't need any segment level locks for "segvn" data
1695 	 * since the address space is "write" locked.
1696 	 */
1697 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
1698 
1699 	/*
1700 	 * Fail the unmap if pages are SOFTLOCKed through this mapping.
1701 	 * softlockcnt is protected from change by the as write lock.
1702 	 */
1703 retry:
1704 	if (svd->softlockcnt > 0) {
1705 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1706 		/*
1707 		 * since we do have the writers lock nobody can fill
1708 		 * the cache during the purge. The flush either succeeds
1709 		 * or we still have pending I/Os.
1710 		 */
1711 		if (reclaim == 1) {
1712 			segvn_purge(seg);
1713 			reclaim = 0;
1714 			goto retry;
1715 		}
1716 		return (EAGAIN);
1717 	}
1718 
1719 	/*
1720 	 * Check for bad sizes
1721 	 */
1722 	if (addr < seg->s_base || addr + len > seg->s_base + seg->s_size ||
1723 	    (len & PAGEOFFSET) || ((uintptr_t)addr & PAGEOFFSET)) {
1724 		panic("segvn_unmap");
1725 		/*NOTREACHED*/
1726 	}
1727 
1728 	if (seg->s_szc != 0) {
1729 		size_t pgsz = page_get_pagesize(seg->s_szc);
1730 		int err;
1731 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
1732 			ASSERT(seg->s_base != addr || seg->s_size != len);
1733 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1734 				ASSERT(svd->amp == NULL);
1735 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1736 				hat_leave_region(seg->s_as->a_hat,
1737 				    svd->rcookie, HAT_REGION_TEXT);
1738 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
1739 				/*
1740 				 * could pass a flag to segvn_demote_range()
1741 				 * below to tell it not to do any unloads but
1742 				 * this case is rare enough to not bother for
1743 				 * now.
1744 				 */
1745 			} else if (svd->tr_state == SEGVN_TR_INIT) {
1746 				svd->tr_state = SEGVN_TR_OFF;
1747 			} else if (svd->tr_state == SEGVN_TR_ON) {
1748 				ASSERT(svd->amp != NULL);
1749 				segvn_textunrepl(seg, 1);
1750 				ASSERT(svd->amp == NULL);
1751 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1752 			}
1753 			VM_STAT_ADD(segvnvmstats.demoterange[0]);
1754 			err = segvn_demote_range(seg, addr, len, SDR_END, 0);
1755 			if (err == 0) {
1756 				return (IE_RETRY);
1757 			}
1758 			return (err);
1759 		}
1760 	}
1761 
1762 	/* Inform the vnode of the unmapping. */
1763 	if (svd->vp) {
1764 		int error;
1765 
1766 		error = VOP_DELMAP(svd->vp,
1767 			(offset_t)svd->offset + (uintptr_t)(addr - seg->s_base),
1768 			seg->s_as, addr, len, svd->prot, svd->maxprot,
1769 			svd->type, svd->cred);
1770 
1771 		if (error == EAGAIN)
1772 			return (error);
1773 	}
1774 
1775 	/*
1776 	 * Remove any page locks set through this mapping.
1777 	 * If text replication is not off no page locks could have been
1778 	 * established via this mapping.
1779 	 */
1780 	if (svd->tr_state == SEGVN_TR_OFF) {
1781 		(void) segvn_lockop(seg, addr, len, 0, MC_UNLOCK, NULL, 0);
1782 	}
1783 
1784 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1785 		ASSERT(svd->amp == NULL);
1786 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1787 		ASSERT(svd->type == MAP_PRIVATE);
1788 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
1789 		    HAT_REGION_TEXT);
1790 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
1791 	} else if (svd->tr_state == SEGVN_TR_ON) {
1792 		ASSERT(svd->amp != NULL);
1793 		ASSERT(svd->pageprot == 0 && !(svd->prot & PROT_WRITE));
1794 		segvn_textunrepl(seg, 1);
1795 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
1796 	} else {
1797 		if (svd->tr_state != SEGVN_TR_OFF) {
1798 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
1799 			svd->tr_state = SEGVN_TR_OFF;
1800 		}
1801 		/*
1802 		 * Unload any hardware translations in the range to be taken
1803 		 * out. Use a callback to invoke free_vp_pages() effectively.
1804 		 */
1805 		if (svd->vp != NULL && free_pages != 0) {
1806 			callback.hcb_data = seg;
1807 			callback.hcb_function = segvn_hat_unload_callback;
1808 			cbp = &callback;
1809 		}
1810 		hat_unload_callback(seg->s_as->a_hat, addr, len,
1811 		    HAT_UNLOAD_UNMAP, cbp);
1812 
1813 		if (svd->type == MAP_SHARED && svd->vp != NULL &&
1814 		    (svd->vp->v_flag & VVMEXEC) &&
1815 		    ((svd->prot & PROT_WRITE) || svd->pageprot)) {
1816 			segvn_inval_trcache(svd->vp);
1817 		}
1818 	}
1819 
1820 	/*
1821 	 * Check for entire segment
1822 	 */
1823 	if (addr == seg->s_base && len == seg->s_size) {
1824 		seg_free(seg);
1825 		return (0);
1826 	}
1827 
1828 	opages = seg_pages(seg);
1829 	dpages = btop(len);
1830 	npages = opages - dpages;
1831 	amp = svd->amp;
1832 	ASSERT(amp == NULL || amp->a_szc >= seg->s_szc);
1833 
1834 	/*
1835 	 * Check for beginning of segment
1836 	 */
1837 	if (addr == seg->s_base) {
1838 		if (svd->vpage != NULL) {
1839 			size_t nbytes;
1840 			struct vpage *ovpage;
1841 
1842 			ovpage = svd->vpage;	/* keep pointer to vpage */
1843 
1844 			nbytes = vpgtob(npages);
1845 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
1846 			bcopy(&ovpage[dpages], svd->vpage, nbytes);
1847 
1848 			/* free up old vpage */
1849 			kmem_free(ovpage, vpgtob(opages));
1850 		}
1851 		if (amp != NULL) {
1852 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1853 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
1854 				/*
1855 				 * Free up now unused parts of anon_map array.
1856 				 */
1857 				if (amp->a_szc == seg->s_szc) {
1858 					if (seg->s_szc != 0) {
1859 						anon_free_pages(amp->ahp,
1860 						    svd->anon_index, len,
1861 						    seg->s_szc);
1862 					} else {
1863 						anon_free(amp->ahp,
1864 						    svd->anon_index,
1865 						    len);
1866 					}
1867 				} else {
1868 					ASSERT(svd->type == MAP_SHARED);
1869 					ASSERT(amp->a_szc > seg->s_szc);
1870 					anon_shmap_free_pages(amp,
1871 					    svd->anon_index, len);
1872 				}
1873 
1874 				/*
1875 				 * Unreserve swap space for the
1876 				 * unmapped chunk of this segment in
1877 				 * case it's MAP_SHARED
1878 				 */
1879 				if (svd->type == MAP_SHARED) {
1880 					anon_unresv(len);
1881 					amp->swresv -= len;
1882 				}
1883 			}
1884 			ANON_LOCK_EXIT(&amp->a_rwlock);
1885 			svd->anon_index += dpages;
1886 		}
1887 		if (svd->vp != NULL)
1888 			svd->offset += len;
1889 
1890 		if (svd->swresv) {
1891 			if (svd->flags & MAP_NORESERVE) {
1892 				ASSERT(amp);
1893 				oswresv = svd->swresv;
1894 
1895 				svd->swresv = ptob(anon_pages(amp->ahp,
1896 				    svd->anon_index, npages));
1897 				anon_unresv(oswresv - svd->swresv);
1898 			} else {
1899 				anon_unresv(len);
1900 				svd->swresv -= len;
1901 			}
1902 			TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
1903 				seg, len, 0);
1904 		}
1905 
1906 		seg->s_base += len;
1907 		seg->s_size -= len;
1908 		return (0);
1909 	}
1910 
1911 	/*
1912 	 * Check for end of segment
1913 	 */
1914 	if (addr + len == seg->s_base + seg->s_size) {
1915 		if (svd->vpage != NULL) {
1916 			size_t nbytes;
1917 			struct vpage *ovpage;
1918 
1919 			ovpage = svd->vpage;	/* keep pointer to vpage */
1920 
1921 			nbytes = vpgtob(npages);
1922 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
1923 			bcopy(ovpage, svd->vpage, nbytes);
1924 
1925 			/* free up old vpage */
1926 			kmem_free(ovpage, vpgtob(opages));
1927 
1928 		}
1929 		if (amp != NULL) {
1930 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1931 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
1932 				/*
1933 				 * Free up now unused parts of anon_map array.
1934 				 */
1935 				ulong_t an_idx = svd->anon_index + npages;
1936 				if (amp->a_szc == seg->s_szc) {
1937 					if (seg->s_szc != 0) {
1938 						anon_free_pages(amp->ahp,
1939 						    an_idx, len,
1940 						    seg->s_szc);
1941 					} else {
1942 						anon_free(amp->ahp, an_idx,
1943 						    len);
1944 					}
1945 				} else {
1946 					ASSERT(svd->type == MAP_SHARED);
1947 					ASSERT(amp->a_szc > seg->s_szc);
1948 					anon_shmap_free_pages(amp,
1949 					    an_idx, len);
1950 				}
1951 
1952 				/*
1953 				 * Unreserve swap space for the
1954 				 * unmapped chunk of this segment in
1955 				 * case it's MAP_SHARED
1956 				 */
1957 				if (svd->type == MAP_SHARED) {
1958 					anon_unresv(len);
1959 					amp->swresv -= len;
1960 				}
1961 			}
1962 			ANON_LOCK_EXIT(&amp->a_rwlock);
1963 		}
1964 
1965 		if (svd->swresv) {
1966 			if (svd->flags & MAP_NORESERVE) {
1967 				ASSERT(amp);
1968 				oswresv = svd->swresv;
1969 				svd->swresv = ptob(anon_pages(amp->ahp,
1970 					svd->anon_index, npages));
1971 				anon_unresv(oswresv - svd->swresv);
1972 			} else {
1973 				anon_unresv(len);
1974 				svd->swresv -= len;
1975 			}
1976 			TRACE_3(TR_FAC_VM, TR_ANON_PROC,
1977 				"anon proc:%p %lu %u", seg, len, 0);
1978 		}
1979 
1980 		seg->s_size -= len;
1981 		return (0);
1982 	}
1983 
1984 	/*
1985 	 * The section to go is in the middle of the segment,
1986 	 * have to make it into two segments.  nseg is made for
1987 	 * the high end while seg is cut down at the low end.
1988 	 */
1989 	nbase = addr + len;				/* new seg base */
1990 	nsize = (seg->s_base + seg->s_size) - nbase;	/* new seg size */
1991 	seg->s_size = addr - seg->s_base;		/* shrink old seg */
1992 	nseg = seg_alloc(seg->s_as, nbase, nsize);
1993 	if (nseg == NULL) {
1994 		panic("segvn_unmap seg_alloc");
1995 		/*NOTREACHED*/
1996 	}
1997 	nseg->s_ops = seg->s_ops;
1998 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
1999 	nseg->s_data = (void *)nsvd;
2000 	nseg->s_szc = seg->s_szc;
2001 	*nsvd = *svd;
2002 	nsvd->seg = nseg;
2003 	nsvd->offset = svd->offset + (uintptr_t)(nseg->s_base - seg->s_base);
2004 	nsvd->swresv = 0;
2005 	nsvd->softlockcnt = 0;
2006 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
2007 
2008 	if (svd->vp != NULL) {
2009 		VN_HOLD(nsvd->vp);
2010 		if (nsvd->type == MAP_SHARED)
2011 			lgrp_shm_policy_init(NULL, nsvd->vp);
2012 	}
2013 	crhold(svd->cred);
2014 
2015 	if (svd->vpage == NULL) {
2016 		nsvd->vpage = NULL;
2017 	} else {
2018 		/* need to split vpage into two arrays */
2019 		size_t nbytes;
2020 		struct vpage *ovpage;
2021 
2022 		ovpage = svd->vpage;		/* keep pointer to vpage */
2023 
2024 		npages = seg_pages(seg);	/* seg has shrunk */
2025 		nbytes = vpgtob(npages);
2026 		svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2027 
2028 		bcopy(ovpage, svd->vpage, nbytes);
2029 
2030 		npages = seg_pages(nseg);
2031 		nbytes = vpgtob(npages);
2032 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2033 
2034 		bcopy(&ovpage[opages - npages], nsvd->vpage, nbytes);
2035 
2036 		/* free up old vpage */
2037 		kmem_free(ovpage, vpgtob(opages));
2038 	}
2039 
2040 	if (amp == NULL) {
2041 		nsvd->amp = NULL;
2042 		nsvd->anon_index = 0;
2043 	} else {
2044 		/*
2045 		 * Need to create a new anon map for the new segment.
2046 		 * We'll also allocate a new smaller array for the old
2047 		 * smaller segment to save space.
2048 		 */
2049 		opages = btop((uintptr_t)(addr - seg->s_base));
2050 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2051 		if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2052 			/*
2053 			 * Free up now unused parts of anon_map array.
2054 			 */
2055 			ulong_t an_idx = svd->anon_index + opages;
2056 			if (amp->a_szc == seg->s_szc) {
2057 				if (seg->s_szc != 0) {
2058 					anon_free_pages(amp->ahp, an_idx, len,
2059 					    seg->s_szc);
2060 				} else {
2061 					anon_free(amp->ahp, an_idx,
2062 					    len);
2063 				}
2064 			} else {
2065 				ASSERT(svd->type == MAP_SHARED);
2066 				ASSERT(amp->a_szc > seg->s_szc);
2067 				anon_shmap_free_pages(amp, an_idx, len);
2068 			}
2069 
2070 			/*
2071 			 * Unreserve swap space for the
2072 			 * unmapped chunk of this segment in
2073 			 * case it's MAP_SHARED
2074 			 */
2075 			if (svd->type == MAP_SHARED) {
2076 				anon_unresv(len);
2077 				amp->swresv -= len;
2078 			}
2079 		}
2080 		nsvd->anon_index = svd->anon_index +
2081 		    btop((uintptr_t)(nseg->s_base - seg->s_base));
2082 		if (svd->type == MAP_SHARED) {
2083 			amp->refcnt++;
2084 			nsvd->amp = amp;
2085 		} else {
2086 			struct anon_map *namp;
2087 			struct anon_hdr *nahp;
2088 
2089 			ASSERT(svd->type == MAP_PRIVATE);
2090 			nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
2091 			namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
2092 			namp->a_szc = seg->s_szc;
2093 			(void) anon_copy_ptr(amp->ahp, svd->anon_index, nahp,
2094 			    0, btop(seg->s_size), ANON_SLEEP);
2095 			(void) anon_copy_ptr(amp->ahp, nsvd->anon_index,
2096 			    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
2097 			anon_release(amp->ahp, btop(amp->size));
2098 			svd->anon_index = 0;
2099 			nsvd->anon_index = 0;
2100 			amp->ahp = nahp;
2101 			amp->size = seg->s_size;
2102 			nsvd->amp = namp;
2103 		}
2104 		ANON_LOCK_EXIT(&amp->a_rwlock);
2105 	}
2106 	if (svd->swresv) {
2107 		if (svd->flags & MAP_NORESERVE) {
2108 			ASSERT(amp);
2109 			oswresv = svd->swresv;
2110 			svd->swresv = ptob(anon_pages(amp->ahp,
2111 				svd->anon_index, btop(seg->s_size)));
2112 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
2113 				nsvd->anon_index, btop(nseg->s_size)));
2114 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2115 			anon_unresv(oswresv - (svd->swresv + nsvd->swresv));
2116 		} else {
2117 			if (seg->s_size + nseg->s_size + len != svd->swresv) {
2118 				panic("segvn_unmap: "
2119 				    "cannot split swap reservation");
2120 				/*NOTREACHED*/
2121 			}
2122 			anon_unresv(len);
2123 			svd->swresv = seg->s_size;
2124 			nsvd->swresv = nseg->s_size;
2125 		}
2126 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2127 			seg, len, 0);
2128 	}
2129 
2130 	return (0);			/* I'm glad that's all over with! */
2131 }
2132 
2133 static void
2134 segvn_free(struct seg *seg)
2135 {
2136 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2137 	pgcnt_t npages = seg_pages(seg);
2138 	struct anon_map *amp;
2139 	size_t len;
2140 
2141 	/*
2142 	 * We don't need any segment level locks for "segvn" data
2143 	 * since the address space is "write" locked.
2144 	 */
2145 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
2146 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
2147 
2148 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2149 
2150 	/*
2151 	 * Be sure to unlock pages. XXX Why do things get free'ed instead
2152 	 * of unmapped? XXX
2153 	 */
2154 	(void) segvn_lockop(seg, seg->s_base, seg->s_size,
2155 	    0, MC_UNLOCK, NULL, 0);
2156 
2157 	/*
2158 	 * Deallocate the vpage and anon pointers if necessary and possible.
2159 	 */
2160 	if (svd->vpage != NULL) {
2161 		kmem_free(svd->vpage, vpgtob(npages));
2162 		svd->vpage = NULL;
2163 	}
2164 	if ((amp = svd->amp) != NULL) {
2165 		/*
2166 		 * If there are no more references to this anon_map
2167 		 * structure, then deallocate the structure after freeing
2168 		 * up all the anon slot pointers that we can.
2169 		 */
2170 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2171 		ASSERT(amp->a_szc >= seg->s_szc);
2172 		if (--amp->refcnt == 0) {
2173 			if (svd->type == MAP_PRIVATE) {
2174 				/*
2175 				 * Private - we only need to anon_free
2176 				 * the part that this segment refers to.
2177 				 */
2178 				if (seg->s_szc != 0) {
2179 					anon_free_pages(amp->ahp,
2180 					    svd->anon_index, seg->s_size,
2181 					    seg->s_szc);
2182 				} else {
2183 					anon_free(amp->ahp, svd->anon_index,
2184 					    seg->s_size);
2185 				}
2186 			} else {
2187 				/*
2188 				 * Shared - anon_free the entire
2189 				 * anon_map's worth of stuff and
2190 				 * release any swap reservation.
2191 				 */
2192 				if (amp->a_szc != 0) {
2193 					anon_shmap_free_pages(amp, 0,
2194 					    amp->size);
2195 				} else {
2196 					anon_free(amp->ahp, 0, amp->size);
2197 				}
2198 				if ((len = amp->swresv) != 0) {
2199 					anon_unresv(len);
2200 					TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2201 						"anon proc:%p %lu %u",
2202 						seg, len, 0);
2203 				}
2204 			}
2205 			svd->amp = NULL;
2206 			ANON_LOCK_EXIT(&amp->a_rwlock);
2207 			anonmap_free(amp);
2208 		} else if (svd->type == MAP_PRIVATE) {
2209 			/*
2210 			 * We had a private mapping which still has
2211 			 * a held anon_map so just free up all the
2212 			 * anon slot pointers that we were using.
2213 			 */
2214 			if (seg->s_szc != 0) {
2215 				anon_free_pages(amp->ahp, svd->anon_index,
2216 				    seg->s_size, seg->s_szc);
2217 			} else {
2218 				anon_free(amp->ahp, svd->anon_index,
2219 				    seg->s_size);
2220 			}
2221 			ANON_LOCK_EXIT(&amp->a_rwlock);
2222 		} else {
2223 			ANON_LOCK_EXIT(&amp->a_rwlock);
2224 		}
2225 	}
2226 
2227 	/*
2228 	 * Release swap reservation.
2229 	 */
2230 	if ((len = svd->swresv) != 0) {
2231 		anon_unresv(svd->swresv);
2232 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2233 			seg, len, 0);
2234 		svd->swresv = 0;
2235 	}
2236 	/*
2237 	 * Release claim on vnode, credentials, and finally free the
2238 	 * private data.
2239 	 */
2240 	if (svd->vp != NULL) {
2241 		if (svd->type == MAP_SHARED)
2242 			lgrp_shm_policy_fini(NULL, svd->vp);
2243 		VN_RELE(svd->vp);
2244 		svd->vp = NULL;
2245 	}
2246 	crfree(svd->cred);
2247 	svd->cred = NULL;
2248 
2249 	seg->s_data = NULL;
2250 	kmem_cache_free(segvn_cache, svd);
2251 }
2252 
2253 ulong_t segvn_lpglck_limit = 0;
2254 /*
2255  * Support routines used by segvn_pagelock() and softlock faults for anonymous
2256  * pages to implement availrmem accounting in a way that makes sure the
2257  * same memory is accounted just once for all softlock/pagelock purposes.
2258  * This prevents a bug when availrmem is quickly incorrectly exausted from
2259  * several pagelocks to different parts of the same large page since each
2260  * pagelock has to decrement availrmem by the size of the entire large
2261  * page. Note those pages are not COW shared until softunlock/pageunlock so
2262  * we don't need to use cow style accounting here.  We also need to make sure
2263  * the entire large page is accounted even if softlock range is less than the
2264  * entire large page because large anon pages can't be demoted when any of
2265  * constituent pages is locked. The caller calls this routine for every page_t
2266  * it locks. The very first page in the range may not be the root page of a
2267  * large page. For all other pages it's guranteed we are going to visit the
2268  * root of a particular large page before any other constituent page as we are
2269  * locking sequential pages belonging to the same anon map. So we do all the
2270  * locking when the root is encountered except for the very first page.  Since
2271  * softlocking is not supported (except S_READ_NOCOW special case) for vmpss
2272  * segments and since vnode pages can be demoted without locking all
2273  * constituent pages vnode pages don't come here.  Unlocking relies on the
2274  * fact that pagesize can't change whenever any of constituent large pages is
2275  * locked at least SE_SHARED. This allows unlocking code to find the right
2276  * root and decrement availrmem by the same amount it was incremented when the
2277  * page was locked.
2278  */
2279 static int
2280 segvn_pp_lock_anonpages(page_t *pp, int first)
2281 {
2282 	pgcnt_t		pages;
2283 	pfn_t		pfn;
2284 	uchar_t		szc = pp->p_szc;
2285 
2286 	ASSERT(PAGE_LOCKED(pp));
2287 	ASSERT(pp->p_vnode != NULL);
2288 	ASSERT(IS_SWAPFSVP(pp->p_vnode));
2289 
2290 	/*
2291 	 * pagesize won't change as long as any constituent page is locked.
2292 	 */
2293 	pages = page_get_pagecnt(pp->p_szc);
2294 	pfn = page_pptonum(pp);
2295 
2296 	if (!first) {
2297 		if (!IS_P2ALIGNED(pfn, pages)) {
2298 #ifdef DEBUG
2299 			pp = &pp[-(spgcnt_t)(pfn & (pages - 1))];
2300 			pfn = page_pptonum(pp);
2301 			ASSERT(IS_P2ALIGNED(pfn, pages));
2302 			ASSERT(pp->p_szc == szc);
2303 			ASSERT(pp->p_vnode != NULL);
2304 			ASSERT(IS_SWAPFSVP(pp->p_vnode));
2305 			ASSERT(pp->p_slckcnt != 0);
2306 #endif /* DEBUG */
2307 			return (1);
2308 		}
2309 	} else if (!IS_P2ALIGNED(pfn, pages)) {
2310 		pp = &pp[-(spgcnt_t)(pfn & (pages - 1))];
2311 #ifdef DEBUG
2312 		pfn = page_pptonum(pp);
2313 		ASSERT(IS_P2ALIGNED(pfn, pages));
2314 		ASSERT(pp->p_szc == szc);
2315 		ASSERT(pp->p_vnode != NULL);
2316 		ASSERT(IS_SWAPFSVP(pp->p_vnode));
2317 #endif /* DEBUG */
2318 	}
2319 
2320 	/*
2321 	 * pp is a root page.
2322 	 * We haven't locked this large page yet.
2323 	 */
2324 	page_struct_lock(pp);
2325 	if (pp->p_slckcnt != 0) {
2326 		if (pp->p_slckcnt < PAGE_SLOCK_MAXIMUM) {
2327 			pp->p_slckcnt++;
2328 			page_struct_unlock(pp);
2329 			return (1);
2330 		}
2331 		page_struct_unlock(pp);
2332 		segvn_lpglck_limit++;
2333 		return (0);
2334 	}
2335 	mutex_enter(&freemem_lock);
2336 	if (availrmem < tune.t_minarmem + pages) {
2337 		mutex_exit(&freemem_lock);
2338 		page_struct_unlock(pp);
2339 		return (0);
2340 	}
2341 	pp->p_slckcnt++;
2342 	availrmem -= pages;
2343 	mutex_exit(&freemem_lock);
2344 	page_struct_unlock(pp);
2345 	return (1);
2346 }
2347 
2348 static void
2349 segvn_pp_unlock_anonpages(page_t *pp, int first)
2350 {
2351 	pgcnt_t		pages;
2352 	pfn_t		pfn;
2353 
2354 	ASSERT(PAGE_LOCKED(pp));
2355 	ASSERT(pp->p_vnode != NULL);
2356 	ASSERT(IS_SWAPFSVP(pp->p_vnode));
2357 
2358 	/*
2359 	 * pagesize won't change as long as any constituent page is locked.
2360 	 */
2361 	pages = page_get_pagecnt(pp->p_szc);
2362 	pfn = page_pptonum(pp);
2363 
2364 	if (!first) {
2365 		if (!IS_P2ALIGNED(pfn, pages)) {
2366 			return;
2367 		}
2368 	} else if (!IS_P2ALIGNED(pfn, pages)) {
2369 		pp = &pp[-(spgcnt_t)(pfn & (pages - 1))];
2370 #ifdef DEBUG
2371 		pfn = page_pptonum(pp);
2372 		ASSERT(IS_P2ALIGNED(pfn, pages));
2373 #endif /* DEBUG */
2374 	}
2375 	ASSERT(pp->p_vnode != NULL);
2376 	ASSERT(IS_SWAPFSVP(pp->p_vnode));
2377 	ASSERT(pp->p_slckcnt != 0);
2378 	page_struct_lock(pp);
2379 	if (--pp->p_slckcnt == 0) {
2380 		mutex_enter(&freemem_lock);
2381 		availrmem += pages;
2382 		mutex_exit(&freemem_lock);
2383 	}
2384 	page_struct_unlock(pp);
2385 }
2386 
2387 /*
2388  * Do a F_SOFTUNLOCK call over the range requested.  The range must have
2389  * already been F_SOFTLOCK'ed.
2390  * Caller must always match addr and len of a softunlock with a previous
2391  * softlock with exactly the same addr and len.
2392  */
2393 static void
2394 segvn_softunlock(struct seg *seg, caddr_t addr, size_t len, enum seg_rw rw)
2395 {
2396 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2397 	page_t *pp;
2398 	caddr_t adr;
2399 	struct vnode *vp;
2400 	u_offset_t offset;
2401 	ulong_t anon_index;
2402 	struct anon_map *amp;
2403 	struct anon *ap = NULL;
2404 
2405 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
2406 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
2407 
2408 	if ((amp = svd->amp) != NULL)
2409 		anon_index = svd->anon_index + seg_page(seg, addr);
2410 
2411 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
2412 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2413 		hat_unlock_region(seg->s_as->a_hat, addr, len, svd->rcookie);
2414 	} else {
2415 		hat_unlock(seg->s_as->a_hat, addr, len);
2416 	}
2417 	for (adr = addr; adr < addr + len; adr += PAGESIZE) {
2418 		if (amp != NULL) {
2419 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2420 			if ((ap = anon_get_ptr(amp->ahp, anon_index++))
2421 								!= NULL) {
2422 				swap_xlate(ap, &vp, &offset);
2423 			} else {
2424 				vp = svd->vp;
2425 				offset = svd->offset +
2426 				    (uintptr_t)(adr - seg->s_base);
2427 			}
2428 			ANON_LOCK_EXIT(&amp->a_rwlock);
2429 		} else {
2430 			vp = svd->vp;
2431 			offset = svd->offset +
2432 			    (uintptr_t)(adr - seg->s_base);
2433 		}
2434 
2435 		/*
2436 		 * Use page_find() instead of page_lookup() to
2437 		 * find the page since we know that it is locked.
2438 		 */
2439 		pp = page_find(vp, offset);
2440 		if (pp == NULL) {
2441 			panic(
2442 			    "segvn_softunlock: addr %p, ap %p, vp %p, off %llx",
2443 			    (void *)adr, (void *)ap, (void *)vp, offset);
2444 			/*NOTREACHED*/
2445 		}
2446 
2447 		if (rw == S_WRITE) {
2448 			hat_setrefmod(pp);
2449 			if (seg->s_as->a_vbits)
2450 				hat_setstat(seg->s_as, adr, PAGESIZE,
2451 				    P_REF | P_MOD);
2452 		} else if (rw != S_OTHER) {
2453 			hat_setref(pp);
2454 			if (seg->s_as->a_vbits)
2455 				hat_setstat(seg->s_as, adr, PAGESIZE, P_REF);
2456 		}
2457 		TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2458 			"segvn_fault:pp %p vp %p offset %llx", pp, vp, offset);
2459 		if (svd->vp == NULL) {
2460 			segvn_pp_unlock_anonpages(pp, adr == addr);
2461 		}
2462 		page_unlock(pp);
2463 	}
2464 	mutex_enter(&freemem_lock); /* for availrmem */
2465 	if (svd->vp != NULL) {
2466 		availrmem += btop(len);
2467 	}
2468 	segvn_pages_locked -= btop(len);
2469 	svd->softlockcnt -= btop(len);
2470 	mutex_exit(&freemem_lock);
2471 	if (svd->softlockcnt == 0) {
2472 		/*
2473 		 * All SOFTLOCKS are gone. Wakeup any waiting
2474 		 * unmappers so they can try again to unmap.
2475 		 * Check for waiters first without the mutex
2476 		 * held so we don't always grab the mutex on
2477 		 * softunlocks.
2478 		 */
2479 		if (AS_ISUNMAPWAIT(seg->s_as)) {
2480 			mutex_enter(&seg->s_as->a_contents);
2481 			if (AS_ISUNMAPWAIT(seg->s_as)) {
2482 				AS_CLRUNMAPWAIT(seg->s_as);
2483 				cv_broadcast(&seg->s_as->a_cv);
2484 			}
2485 			mutex_exit(&seg->s_as->a_contents);
2486 		}
2487 	}
2488 }
2489 
2490 #define	PAGE_HANDLED	((page_t *)-1)
2491 
2492 /*
2493  * Release all the pages in the NULL terminated ppp list
2494  * which haven't already been converted to PAGE_HANDLED.
2495  */
2496 static void
2497 segvn_pagelist_rele(page_t **ppp)
2498 {
2499 	for (; *ppp != NULL; ppp++) {
2500 		if (*ppp != PAGE_HANDLED)
2501 			page_unlock(*ppp);
2502 	}
2503 }
2504 
2505 static int stealcow = 1;
2506 
2507 /*
2508  * Workaround for viking chip bug.  See bug id 1220902.
2509  * To fix this down in pagefault() would require importing so
2510  * much as and segvn code as to be unmaintainable.
2511  */
2512 int enable_mbit_wa = 0;
2513 
2514 /*
2515  * Handles all the dirty work of getting the right
2516  * anonymous pages and loading up the translations.
2517  * This routine is called only from segvn_fault()
2518  * when looping over the range of addresses requested.
2519  *
2520  * The basic algorithm here is:
2521  * 	If this is an anon_zero case
2522  *		Call anon_zero to allocate page
2523  *		Load up translation
2524  *		Return
2525  *	endif
2526  *	If this is an anon page
2527  *		Use anon_getpage to get the page
2528  *	else
2529  *		Find page in pl[] list passed in
2530  *	endif
2531  *	If not a cow
2532  *		Load up the translation to the page
2533  *		return
2534  *	endif
2535  *	Call anon_private to handle cow
2536  *	Load up (writable) translation to new page
2537  */
2538 static faultcode_t
2539 segvn_faultpage(
2540 	struct hat *hat,		/* the hat to use for mapping */
2541 	struct seg *seg,		/* seg_vn of interest */
2542 	caddr_t addr,			/* address in as */
2543 	u_offset_t off,			/* offset in vp */
2544 	struct vpage *vpage,		/* pointer to vpage for vp, off */
2545 	page_t *pl[],			/* object source page pointer */
2546 	uint_t vpprot,			/* access allowed to object pages */
2547 	enum fault_type type,		/* type of fault */
2548 	enum seg_rw rw,			/* type of access at fault */
2549 	int brkcow,			/* we may need to break cow */
2550 	int first)			/* first page for this fault if 1 */
2551 {
2552 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2553 	page_t *pp, **ppp;
2554 	uint_t pageflags = 0;
2555 	page_t *anon_pl[1 + 1];
2556 	page_t *opp = NULL;		/* original page */
2557 	uint_t prot;
2558 	int err;
2559 	int cow;
2560 	int claim;
2561 	int steal = 0;
2562 	ulong_t anon_index;
2563 	struct anon *ap, *oldap;
2564 	struct anon_map *amp;
2565 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
2566 	int anon_lock = 0;
2567 	anon_sync_obj_t cookie;
2568 
2569 	if (svd->flags & MAP_TEXT) {
2570 		hat_flag |= HAT_LOAD_TEXT;
2571 	}
2572 
2573 	ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
2574 	ASSERT(seg->s_szc == 0);
2575 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
2576 
2577 	/*
2578 	 * Initialize protection value for this page.
2579 	 * If we have per page protection values check it now.
2580 	 */
2581 	if (svd->pageprot) {
2582 		uint_t protchk;
2583 
2584 		switch (rw) {
2585 		case S_READ:
2586 			protchk = PROT_READ;
2587 			break;
2588 		case S_WRITE:
2589 			protchk = PROT_WRITE;
2590 			break;
2591 		case S_EXEC:
2592 			protchk = PROT_EXEC;
2593 			break;
2594 		case S_OTHER:
2595 		default:
2596 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
2597 			break;
2598 		}
2599 
2600 		prot = VPP_PROT(vpage);
2601 		if ((prot & protchk) == 0)
2602 			return (FC_PROT);	/* illegal access type */
2603 	} else {
2604 		prot = svd->prot;
2605 	}
2606 
2607 	if (type == F_SOFTLOCK && svd->vp != NULL) {
2608 		mutex_enter(&freemem_lock);
2609 		if (availrmem <= tune.t_minarmem) {
2610 			mutex_exit(&freemem_lock);
2611 			return (FC_MAKE_ERR(ENOMEM));	/* out of real memory */
2612 		} else {
2613 			availrmem--;
2614 			svd->softlockcnt++;
2615 			segvn_pages_locked++;
2616 		}
2617 		mutex_exit(&freemem_lock);
2618 	}
2619 
2620 	/*
2621 	 * Always acquire the anon array lock to prevent 2 threads from
2622 	 * allocating separate anon slots for the same "addr".
2623 	 */
2624 
2625 	if ((amp = svd->amp) != NULL) {
2626 		ASSERT(RW_READ_HELD(&amp->a_rwlock));
2627 		anon_index = svd->anon_index + seg_page(seg, addr);
2628 		anon_array_enter(amp, anon_index, &cookie);
2629 		anon_lock = 1;
2630 	}
2631 
2632 	if (svd->vp == NULL && amp != NULL) {
2633 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL) {
2634 			/*
2635 			 * Allocate a (normally) writable anonymous page of
2636 			 * zeroes. If no advance reservations, reserve now.
2637 			 */
2638 			if (svd->flags & MAP_NORESERVE) {
2639 				if (anon_resv_zone(ptob(1),
2640 				    seg->s_as->a_proc->p_zone)) {
2641 					atomic_add_long(&svd->swresv, ptob(1));
2642 				} else {
2643 					err = ENOMEM;
2644 					goto out;
2645 				}
2646 			}
2647 			if ((pp = anon_zero(seg, addr, &ap,
2648 			    svd->cred)) == NULL) {
2649 				err = ENOMEM;
2650 				goto out;	/* out of swap space */
2651 			}
2652 			/*
2653 			 * Re-acquire the anon_map lock and
2654 			 * initialize the anon array entry.
2655 			 */
2656 			(void) anon_set_ptr(amp->ahp, anon_index, ap,
2657 				ANON_SLEEP);
2658 
2659 			ASSERT(pp->p_szc == 0);
2660 
2661 			/*
2662 			 * Handle pages that have been marked for migration
2663 			 */
2664 			if (lgrp_optimizations())
2665 				page_migrate(seg, addr, &pp, 1);
2666 
2667 			if (type == F_SOFTLOCK) {
2668 				if (!segvn_pp_lock_anonpages(pp, first)) {
2669 					page_unlock(pp);
2670 					err = ENOMEM;
2671 					goto out;
2672 				} else {
2673 					mutex_enter(&freemem_lock);
2674 					svd->softlockcnt++;
2675 					segvn_pages_locked++;
2676 					mutex_exit(&freemem_lock);
2677 				}
2678 			}
2679 
2680 			if (enable_mbit_wa) {
2681 				if (rw == S_WRITE)
2682 					hat_setmod(pp);
2683 				else if (!hat_ismod(pp))
2684 					prot &= ~PROT_WRITE;
2685 			}
2686 			/*
2687 			 * If AS_PAGLCK is set in a_flags (via memcntl(2)
2688 			 * with MC_LOCKAS, MCL_FUTURE) and this is a
2689 			 * MAP_NORESERVE segment, we may need to
2690 			 * permanently lock the page as it is being faulted
2691 			 * for the first time. The following text applies
2692 			 * only to MAP_NORESERVE segments:
2693 			 *
2694 			 * As per memcntl(2), if this segment was created
2695 			 * after MCL_FUTURE was applied (a "future"
2696 			 * segment), its pages must be locked.  If this
2697 			 * segment existed at MCL_FUTURE application (a
2698 			 * "past" segment), the interface is unclear.
2699 			 *
2700 			 * We decide to lock only if vpage is present:
2701 			 *
2702 			 * - "future" segments will have a vpage array (see
2703 			 *    as_map), and so will be locked as required
2704 			 *
2705 			 * - "past" segments may not have a vpage array,
2706 			 *    depending on whether events (such as
2707 			 *    mprotect) have occurred. Locking if vpage
2708 			 *    exists will preserve legacy behavior.  Not
2709 			 *    locking if vpage is absent, will not break
2710 			 *    the interface or legacy behavior.  Note that
2711 			 *    allocating vpage here if it's absent requires
2712 			 *    upgrading the segvn reader lock, the cost of
2713 			 *    which does not seem worthwhile.
2714 			 *
2715 			 * Usually testing and setting VPP_ISPPLOCK and
2716 			 * VPP_SETPPLOCK requires holding the segvn lock as
2717 			 * writer, but in this case all readers are
2718 			 * serializing on the anon array lock.
2719 			 */
2720 			if (AS_ISPGLCK(seg->s_as) && vpage != NULL &&
2721 			    (svd->flags & MAP_NORESERVE) &&
2722 			    !VPP_ISPPLOCK(vpage)) {
2723 				proc_t *p = seg->s_as->a_proc;
2724 				ASSERT(svd->type == MAP_PRIVATE);
2725 				mutex_enter(&p->p_lock);
2726 				if (rctl_incr_locked_mem(p, NULL, PAGESIZE,
2727 				    1) == 0) {
2728 					claim = VPP_PROT(vpage) & PROT_WRITE;
2729 					if (page_pp_lock(pp, claim, 0)) {
2730 						VPP_SETPPLOCK(vpage);
2731 					} else {
2732 						rctl_decr_locked_mem(p, NULL,
2733 						    PAGESIZE, 1);
2734 					}
2735 				}
2736 				mutex_exit(&p->p_lock);
2737 			}
2738 
2739 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2740 			hat_memload(hat, addr, pp, prot, hat_flag);
2741 
2742 			if (!(hat_flag & HAT_LOAD_LOCK))
2743 				page_unlock(pp);
2744 
2745 			anon_array_exit(&cookie);
2746 			return (0);
2747 		}
2748 	}
2749 
2750 	/*
2751 	 * Obtain the page structure via anon_getpage() if it is
2752 	 * a private copy of an object (the result of a previous
2753 	 * copy-on-write).
2754 	 */
2755 	if (amp != NULL) {
2756 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) != NULL) {
2757 			err = anon_getpage(&ap, &vpprot, anon_pl, PAGESIZE,
2758 			    seg, addr, rw, svd->cred);
2759 			if (err)
2760 				goto out;
2761 
2762 			if (svd->type == MAP_SHARED) {
2763 				/*
2764 				 * If this is a shared mapping to an
2765 				 * anon_map, then ignore the write
2766 				 * permissions returned by anon_getpage().
2767 				 * They apply to the private mappings
2768 				 * of this anon_map.
2769 				 */
2770 				vpprot |= PROT_WRITE;
2771 			}
2772 			opp = anon_pl[0];
2773 		}
2774 	}
2775 
2776 	/*
2777 	 * Search the pl[] list passed in if it is from the
2778 	 * original object (i.e., not a private copy).
2779 	 */
2780 	if (opp == NULL) {
2781 		/*
2782 		 * Find original page.  We must be bringing it in
2783 		 * from the list in pl[].
2784 		 */
2785 		for (ppp = pl; (opp = *ppp) != NULL; ppp++) {
2786 			if (opp == PAGE_HANDLED)
2787 				continue;
2788 			ASSERT(opp->p_vnode == svd->vp); /* XXX */
2789 			if (opp->p_offset == off)
2790 				break;
2791 		}
2792 		if (opp == NULL) {
2793 			panic("segvn_faultpage not found");
2794 			/*NOTREACHED*/
2795 		}
2796 		*ppp = PAGE_HANDLED;
2797 
2798 	}
2799 
2800 	ASSERT(PAGE_LOCKED(opp));
2801 
2802 	TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2803 		"segvn_fault:pp %p vp %p offset %llx",
2804 		opp, NULL, 0);
2805 
2806 	/*
2807 	 * The fault is treated as a copy-on-write fault if a
2808 	 * write occurs on a private segment and the object
2809 	 * page (i.e., mapping) is write protected.  We assume
2810 	 * that fatal protection checks have already been made.
2811 	 */
2812 
2813 	if (brkcow) {
2814 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2815 		cow = !(vpprot & PROT_WRITE);
2816 	} else if (svd->tr_state == SEGVN_TR_ON) {
2817 		/*
2818 		 * If we are doing text replication COW on first touch.
2819 		 */
2820 		ASSERT(amp != NULL);
2821 		ASSERT(svd->vp != NULL);
2822 		ASSERT(rw != S_WRITE);
2823 		cow = (ap == NULL);
2824 	} else {
2825 		cow = 0;
2826 	}
2827 
2828 	/*
2829 	 * If not a copy-on-write case load the translation
2830 	 * and return.
2831 	 */
2832 	if (cow == 0) {
2833 
2834 		/*
2835 		 * Handle pages that have been marked for migration
2836 		 */
2837 		if (lgrp_optimizations())
2838 			page_migrate(seg, addr, &opp, 1);
2839 
2840 		if (type == F_SOFTLOCK && svd->vp == NULL) {
2841 
2842 			ASSERT(opp->p_szc == 0 ||
2843 			    (svd->type == MAP_SHARED &&
2844 				amp != NULL && amp->a_szc != 0));
2845 
2846 			if (!segvn_pp_lock_anonpages(opp, first)) {
2847 				page_unlock(opp);
2848 				err = ENOMEM;
2849 				goto out;
2850 			} else {
2851 				mutex_enter(&freemem_lock);
2852 				svd->softlockcnt++;
2853 				segvn_pages_locked++;
2854 				mutex_exit(&freemem_lock);
2855 			}
2856 		}
2857 		if (IS_VMODSORT(opp->p_vnode) || enable_mbit_wa) {
2858 			if (rw == S_WRITE)
2859 				hat_setmod(opp);
2860 			else if (rw != S_OTHER && !hat_ismod(opp))
2861 				prot &= ~PROT_WRITE;
2862 		}
2863 
2864 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
2865 		    (!svd->pageprot && svd->prot == (prot & vpprot)));
2866 		ASSERT(amp == NULL ||
2867 		    svd->rcookie == HAT_INVALID_REGION_COOKIE);
2868 		hat_memload_region(hat, addr, opp, prot & vpprot, hat_flag,
2869 		    svd->rcookie);
2870 
2871 		if (!(hat_flag & HAT_LOAD_LOCK))
2872 			page_unlock(opp);
2873 
2874 		if (anon_lock) {
2875 			anon_array_exit(&cookie);
2876 		}
2877 		return (0);
2878 	}
2879 
2880 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2881 
2882 	hat_setref(opp);
2883 
2884 	ASSERT(amp != NULL && anon_lock);
2885 
2886 	/*
2887 	 * Steal the page only if it isn't a private page
2888 	 * since stealing a private page is not worth the effort.
2889 	 */
2890 	if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL)
2891 		steal = 1;
2892 
2893 	/*
2894 	 * Steal the original page if the following conditions are true:
2895 	 *
2896 	 * We are low on memory, the page is not private, page is not large,
2897 	 * not shared, not modified, not `locked' or if we have it `locked'
2898 	 * (i.e., p_cowcnt == 1 and p_lckcnt == 0, which also implies
2899 	 * that the page is not shared) and if it doesn't have any
2900 	 * translations. page_struct_lock isn't needed to look at p_cowcnt
2901 	 * and p_lckcnt because we first get exclusive lock on page.
2902 	 */
2903 	(void) hat_pagesync(opp, HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD);
2904 
2905 	if (stealcow && freemem < minfree && steal && opp->p_szc == 0 &&
2906 	    page_tryupgrade(opp) && !hat_ismod(opp) &&
2907 	    ((opp->p_lckcnt == 0 && opp->p_cowcnt == 0) ||
2908 	    (opp->p_lckcnt == 0 && opp->p_cowcnt == 1 &&
2909 	    vpage != NULL && VPP_ISPPLOCK(vpage)))) {
2910 		/*
2911 		 * Check if this page has other translations
2912 		 * after unloading our translation.
2913 		 */
2914 		if (hat_page_is_mapped(opp)) {
2915 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2916 			hat_unload(seg->s_as->a_hat, addr, PAGESIZE,
2917 				HAT_UNLOAD);
2918 		}
2919 
2920 		/*
2921 		 * hat_unload() might sync back someone else's recent
2922 		 * modification, so check again.
2923 		 */
2924 		if (!hat_ismod(opp) && !hat_page_is_mapped(opp))
2925 			pageflags |= STEAL_PAGE;
2926 	}
2927 
2928 	/*
2929 	 * If we have a vpage pointer, see if it indicates that we have
2930 	 * ``locked'' the page we map -- if so, tell anon_private to
2931 	 * transfer the locking resource to the new page.
2932 	 *
2933 	 * See Statement at the beginning of segvn_lockop regarding
2934 	 * the way lockcnts/cowcnts are handled during COW.
2935 	 *
2936 	 */
2937 	if (vpage != NULL && VPP_ISPPLOCK(vpage))
2938 		pageflags |= LOCK_PAGE;
2939 
2940 	/*
2941 	 * Allocate a private page and perform the copy.
2942 	 * For MAP_NORESERVE reserve swap space now, unless this
2943 	 * is a cow fault on an existing anon page in which case
2944 	 * MAP_NORESERVE will have made advance reservations.
2945 	 */
2946 	if ((svd->flags & MAP_NORESERVE) && (ap == NULL)) {
2947 		if (anon_resv_zone(ptob(1), seg->s_as->a_proc->p_zone)) {
2948 			atomic_add_long(&svd->swresv, ptob(1));
2949 		} else {
2950 			page_unlock(opp);
2951 			err = ENOMEM;
2952 			goto out;
2953 		}
2954 	}
2955 	oldap = ap;
2956 	pp = anon_private(&ap, seg, addr, prot, opp, pageflags, svd->cred);
2957 	if (pp == NULL) {
2958 		err = ENOMEM;	/* out of swap space */
2959 		goto out;
2960 	}
2961 
2962 	/*
2963 	 * If we copied away from an anonymous page, then
2964 	 * we are one step closer to freeing up an anon slot.
2965 	 *
2966 	 * NOTE:  The original anon slot must be released while
2967 	 * holding the "anon_map" lock.  This is necessary to prevent
2968 	 * other threads from obtaining a pointer to the anon slot
2969 	 * which may be freed if its "refcnt" is 1.
2970 	 */
2971 	if (oldap != NULL)
2972 		anon_decref(oldap);
2973 
2974 	(void) anon_set_ptr(amp->ahp, anon_index, ap, ANON_SLEEP);
2975 
2976 	/*
2977 	 * Handle pages that have been marked for migration
2978 	 */
2979 	if (lgrp_optimizations())
2980 		page_migrate(seg, addr, &pp, 1);
2981 
2982 	ASSERT(pp->p_szc == 0);
2983 	if (type == F_SOFTLOCK && svd->vp == NULL) {
2984 		if (!segvn_pp_lock_anonpages(pp, first)) {
2985 			page_unlock(pp);
2986 			err = ENOMEM;
2987 			goto out;
2988 		} else {
2989 			mutex_enter(&freemem_lock);
2990 			svd->softlockcnt++;
2991 			segvn_pages_locked++;
2992 			mutex_exit(&freemem_lock);
2993 		}
2994 	}
2995 
2996 	ASSERT(!IS_VMODSORT(pp->p_vnode));
2997 	if (enable_mbit_wa) {
2998 		if (rw == S_WRITE)
2999 			hat_setmod(pp);
3000 		else if (!hat_ismod(pp))
3001 			prot &= ~PROT_WRITE;
3002 	}
3003 
3004 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3005 	hat_memload(hat, addr, pp, prot, hat_flag);
3006 
3007 	if (!(hat_flag & HAT_LOAD_LOCK))
3008 		page_unlock(pp);
3009 
3010 	ASSERT(anon_lock);
3011 	anon_array_exit(&cookie);
3012 	return (0);
3013 out:
3014 	if (anon_lock)
3015 		anon_array_exit(&cookie);
3016 
3017 	if (type == F_SOFTLOCK && svd->vp != NULL) {
3018 		mutex_enter(&freemem_lock);
3019 		availrmem++;
3020 		segvn_pages_locked--;
3021 		svd->softlockcnt--;
3022 		mutex_exit(&freemem_lock);
3023 	}
3024 	return (FC_MAKE_ERR(err));
3025 }
3026 
3027 /*
3028  * relocate a bunch of smaller targ pages into one large repl page. all targ
3029  * pages must be complete pages smaller than replacement pages.
3030  * it's assumed that no page's szc can change since they are all PAGESIZE or
3031  * complete large pages locked SHARED.
3032  */
3033 static void
3034 segvn_relocate_pages(page_t **targ, page_t *replacement)
3035 {
3036 	page_t *pp;
3037 	pgcnt_t repl_npgs, curnpgs;
3038 	pgcnt_t i;
3039 	uint_t repl_szc = replacement->p_szc;
3040 	page_t *first_repl = replacement;
3041 	page_t *repl;
3042 	spgcnt_t npgs;
3043 
3044 	VM_STAT_ADD(segvnvmstats.relocatepages[0]);
3045 
3046 	ASSERT(repl_szc != 0);
3047 	npgs = repl_npgs = page_get_pagecnt(repl_szc);
3048 
3049 	i = 0;
3050 	while (repl_npgs) {
3051 		spgcnt_t nreloc;
3052 		int err;
3053 		ASSERT(replacement != NULL);
3054 		pp = targ[i];
3055 		ASSERT(pp->p_szc < repl_szc);
3056 		ASSERT(PAGE_EXCL(pp));
3057 		ASSERT(!PP_ISFREE(pp));
3058 		curnpgs = page_get_pagecnt(pp->p_szc);
3059 		if (curnpgs == 1) {
3060 			VM_STAT_ADD(segvnvmstats.relocatepages[1]);
3061 			repl = replacement;
3062 			page_sub(&replacement, repl);
3063 			ASSERT(PAGE_EXCL(repl));
3064 			ASSERT(!PP_ISFREE(repl));
3065 			ASSERT(repl->p_szc == repl_szc);
3066 		} else {
3067 			page_t *repl_savepp;
3068 			int j;
3069 			VM_STAT_ADD(segvnvmstats.relocatepages[2]);
3070 			repl_savepp = replacement;
3071 			for (j = 0; j < curnpgs; j++) {
3072 				repl = replacement;
3073 				page_sub(&replacement, repl);
3074 				ASSERT(PAGE_EXCL(repl));
3075 				ASSERT(!PP_ISFREE(repl));
3076 				ASSERT(repl->p_szc == repl_szc);
3077 				ASSERT(page_pptonum(targ[i + j]) ==
3078 				    page_pptonum(targ[i]) + j);
3079 			}
3080 			repl = repl_savepp;
3081 			ASSERT(IS_P2ALIGNED(page_pptonum(repl), curnpgs));
3082 		}
3083 		err = page_relocate(&pp, &repl, 0, 1, &nreloc, NULL);
3084 		if (err || nreloc != curnpgs) {
3085 			panic("segvn_relocate_pages: "
3086 			    "page_relocate failed err=%d curnpgs=%ld "
3087 			    "nreloc=%ld", err, curnpgs, nreloc);
3088 		}
3089 		ASSERT(curnpgs <= repl_npgs);
3090 		repl_npgs -= curnpgs;
3091 		i += curnpgs;
3092 	}
3093 	ASSERT(replacement == NULL);
3094 
3095 	repl = first_repl;
3096 	repl_npgs = npgs;
3097 	for (i = 0; i < repl_npgs; i++) {
3098 		ASSERT(PAGE_EXCL(repl));
3099 		ASSERT(!PP_ISFREE(repl));
3100 		targ[i] = repl;
3101 		page_downgrade(targ[i]);
3102 		repl++;
3103 	}
3104 }
3105 
3106 /*
3107  * Check if all pages in ppa array are complete smaller than szc pages and
3108  * their roots will still be aligned relative to their current size if the
3109  * entire ppa array is relocated into one szc page. If these conditions are
3110  * not met return 0.
3111  *
3112  * If all pages are properly aligned attempt to upgrade their locks
3113  * to exclusive mode. If it fails set *upgrdfail to 1 and return 0.
3114  * upgrdfail was set to 0 by caller.
3115  *
3116  * Return 1 if all pages are aligned and locked exclusively.
3117  *
3118  * If all pages in ppa array happen to be physically contiguous to make one
3119  * szc page and all exclusive locks are successfully obtained promote the page
3120  * size to szc and set *pszc to szc. Return 1 with pages locked shared.
3121  */
3122 static int
3123 segvn_full_szcpages(page_t **ppa, uint_t szc, int *upgrdfail, uint_t *pszc)
3124 {
3125 	page_t *pp;
3126 	pfn_t pfn;
3127 	pgcnt_t totnpgs = page_get_pagecnt(szc);
3128 	pfn_t first_pfn;
3129 	int contig = 1;
3130 	pgcnt_t i;
3131 	pgcnt_t j;
3132 	uint_t curszc;
3133 	pgcnt_t curnpgs;
3134 	int root = 0;
3135 
3136 	ASSERT(szc > 0);
3137 
3138 	VM_STAT_ADD(segvnvmstats.fullszcpages[0]);
3139 
3140 	for (i = 0; i < totnpgs; i++) {
3141 		pp = ppa[i];
3142 		ASSERT(PAGE_SHARED(pp));
3143 		ASSERT(!PP_ISFREE(pp));
3144 		pfn = page_pptonum(pp);
3145 		if (i == 0) {
3146 			if (!IS_P2ALIGNED(pfn, totnpgs)) {
3147 				contig = 0;
3148 			} else {
3149 				first_pfn = pfn;
3150 			}
3151 		} else if (contig && pfn != first_pfn + i) {
3152 			contig = 0;
3153 		}
3154 		if (pp->p_szc == 0) {
3155 			if (root) {
3156 				VM_STAT_ADD(segvnvmstats.fullszcpages[1]);
3157 				return (0);
3158 			}
3159 		} else if (!root) {
3160 			if ((curszc = pp->p_szc) >= szc) {
3161 				VM_STAT_ADD(segvnvmstats.fullszcpages[2]);
3162 				return (0);
3163 			}
3164 			if (curszc == 0) {
3165 				/*
3166 				 * p_szc changed means we don't have all pages
3167 				 * locked. return failure.
3168 				 */
3169 				VM_STAT_ADD(segvnvmstats.fullszcpages[3]);
3170 				return (0);
3171 			}
3172 			curnpgs = page_get_pagecnt(curszc);
3173 			if (!IS_P2ALIGNED(pfn, curnpgs) ||
3174 			    !IS_P2ALIGNED(i, curnpgs)) {
3175 				VM_STAT_ADD(segvnvmstats.fullszcpages[4]);
3176 				return (0);
3177 			}
3178 			root = 1;
3179 		} else {
3180 			ASSERT(i > 0);
3181 			VM_STAT_ADD(segvnvmstats.fullszcpages[5]);
3182 			if (pp->p_szc != curszc) {
3183 				VM_STAT_ADD(segvnvmstats.fullszcpages[6]);
3184 				return (0);
3185 			}
3186 			if (pfn - 1 != page_pptonum(ppa[i - 1])) {
3187 				panic("segvn_full_szcpages: "
3188 				    "large page not physically contiguous");
3189 			}
3190 			if (P2PHASE(pfn, curnpgs) == curnpgs - 1) {
3191 				root = 0;
3192 			}
3193 		}
3194 	}
3195 
3196 	for (i = 0; i < totnpgs; i++) {
3197 		ASSERT(ppa[i]->p_szc < szc);
3198 		if (!page_tryupgrade(ppa[i])) {
3199 			for (j = 0; j < i; j++) {
3200 				page_downgrade(ppa[j]);
3201 			}
3202 			*pszc = ppa[i]->p_szc;
3203 			*upgrdfail = 1;
3204 			VM_STAT_ADD(segvnvmstats.fullszcpages[7]);
3205 			return (0);
3206 		}
3207 	}
3208 
3209 	/*
3210 	 * When a page is put a free cachelist its szc is set to 0.  if file
3211 	 * system reclaimed pages from cachelist targ pages will be physically
3212 	 * contiguous with 0 p_szc.  in this case just upgrade szc of targ
3213 	 * pages without any relocations.
3214 	 * To avoid any hat issues with previous small mappings
3215 	 * hat_pageunload() the target pages first.
3216 	 */
3217 	if (contig) {
3218 		VM_STAT_ADD(segvnvmstats.fullszcpages[8]);
3219 		for (i = 0; i < totnpgs; i++) {
3220 			(void) hat_pageunload(ppa[i], HAT_FORCE_PGUNLOAD);
3221 		}
3222 		for (i = 0; i < totnpgs; i++) {
3223 			ppa[i]->p_szc = szc;
3224 		}
3225 		for (i = 0; i < totnpgs; i++) {
3226 			ASSERT(PAGE_EXCL(ppa[i]));
3227 			page_downgrade(ppa[i]);
3228 		}
3229 		if (pszc != NULL) {
3230 			*pszc = szc;
3231 		}
3232 	}
3233 	VM_STAT_ADD(segvnvmstats.fullszcpages[9]);
3234 	return (1);
3235 }
3236 
3237 /*
3238  * Create physically contiguous pages for [vp, off] - [vp, off +
3239  * page_size(szc)) range and for private segment return them in ppa array.
3240  * Pages are created either via IO or relocations.
3241  *
3242  * Return 1 on sucess and 0 on failure.
3243  *
3244  * If physically contiguos pages already exist for this range return 1 without
3245  * filling ppa array. Caller initializes ppa[0] as NULL to detect that ppa
3246  * array wasn't filled. In this case caller fills ppa array via VOP_GETPAGE().
3247  */
3248 
3249 static int
3250 segvn_fill_vp_pages(struct segvn_data *svd, vnode_t *vp, u_offset_t off,
3251     uint_t szc, page_t **ppa, page_t **ppplist, uint_t *ret_pszc,
3252     int *downsize)
3253 
3254 {
3255 	page_t *pplist = *ppplist;
3256 	size_t pgsz = page_get_pagesize(szc);
3257 	pgcnt_t pages = btop(pgsz);
3258 	ulong_t start_off = off;
3259 	u_offset_t eoff = off + pgsz;
3260 	spgcnt_t nreloc;
3261 	u_offset_t io_off = off;
3262 	size_t io_len;
3263 	page_t *io_pplist = NULL;
3264 	page_t *done_pplist = NULL;
3265 	pgcnt_t pgidx = 0;
3266 	page_t *pp;
3267 	page_t *newpp;
3268 	page_t *targpp;
3269 	int io_err = 0;
3270 	int i;
3271 	pfn_t pfn;
3272 	ulong_t ppages;
3273 	page_t *targ_pplist = NULL;
3274 	page_t *repl_pplist = NULL;
3275 	page_t *tmp_pplist;
3276 	int nios = 0;
3277 	uint_t pszc;
3278 	struct vattr va;
3279 
3280 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[0]);
3281 
3282 	ASSERT(szc != 0);
3283 	ASSERT(pplist->p_szc == szc);
3284 
3285 	/*
3286 	 * downsize will be set to 1 only if we fail to lock pages. this will
3287 	 * allow subsequent faults to try to relocate the page again. If we
3288 	 * fail due to misalignment don't downsize and let the caller map the
3289 	 * whole region with small mappings to avoid more faults into the area
3290 	 * where we can't get large pages anyway.
3291 	 */
3292 	*downsize = 0;
3293 
3294 	while (off < eoff) {
3295 		newpp = pplist;
3296 		ASSERT(newpp != NULL);
3297 		ASSERT(PAGE_EXCL(newpp));
3298 		ASSERT(!PP_ISFREE(newpp));
3299 		/*
3300 		 * we pass NULL for nrelocp to page_lookup_create()
3301 		 * so that it doesn't relocate. We relocate here
3302 		 * later only after we make sure we can lock all
3303 		 * pages in the range we handle and they are all
3304 		 * aligned.
3305 		 */
3306 		pp = page_lookup_create(vp, off, SE_SHARED, newpp, NULL, 0);
3307 		ASSERT(pp != NULL);
3308 		ASSERT(!PP_ISFREE(pp));
3309 		ASSERT(pp->p_vnode == vp);
3310 		ASSERT(pp->p_offset == off);
3311 		if (pp == newpp) {
3312 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[1]);
3313 			page_sub(&pplist, pp);
3314 			ASSERT(PAGE_EXCL(pp));
3315 			ASSERT(page_iolock_assert(pp));
3316 			page_list_concat(&io_pplist, &pp);
3317 			off += PAGESIZE;
3318 			continue;
3319 		}
3320 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[2]);
3321 		pfn = page_pptonum(pp);
3322 		pszc = pp->p_szc;
3323 		if (pszc >= szc && targ_pplist == NULL && io_pplist == NULL &&
3324 		    IS_P2ALIGNED(pfn, pages)) {
3325 			ASSERT(repl_pplist == NULL);
3326 			ASSERT(done_pplist == NULL);
3327 			ASSERT(pplist == *ppplist);
3328 			page_unlock(pp);
3329 			page_free_replacement_page(pplist);
3330 			page_create_putback(pages);
3331 			*ppplist = NULL;
3332 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[3]);
3333 			return (1);
3334 		}
3335 		if (pszc >= szc) {
3336 			page_unlock(pp);
3337 			segvn_faultvnmpss_align_err1++;
3338 			goto out;
3339 		}
3340 		ppages = page_get_pagecnt(pszc);
3341 		if (!IS_P2ALIGNED(pfn, ppages)) {
3342 			ASSERT(pszc > 0);
3343 			/*
3344 			 * sizing down to pszc won't help.
3345 			 */
3346 			page_unlock(pp);
3347 			segvn_faultvnmpss_align_err2++;
3348 			goto out;
3349 		}
3350 		pfn = page_pptonum(newpp);
3351 		if (!IS_P2ALIGNED(pfn, ppages)) {
3352 			ASSERT(pszc > 0);
3353 			/*
3354 			 * sizing down to pszc won't help.
3355 			 */
3356 			page_unlock(pp);
3357 			segvn_faultvnmpss_align_err3++;
3358 			goto out;
3359 		}
3360 		if (!PAGE_EXCL(pp)) {
3361 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[4]);
3362 			page_unlock(pp);
3363 			*downsize = 1;
3364 			*ret_pszc = pp->p_szc;
3365 			goto out;
3366 		}
3367 		targpp = pp;
3368 		if (io_pplist != NULL) {
3369 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[5]);
3370 			io_len = off - io_off;
3371 			/*
3372 			 * Some file systems like NFS don't check EOF
3373 			 * conditions in VOP_PAGEIO(). Check it here
3374 			 * now that pages are locked SE_EXCL. Any file
3375 			 * truncation will wait until the pages are
3376 			 * unlocked so no need to worry that file will
3377 			 * be truncated after we check its size here.
3378 			 * XXX fix NFS to remove this check.
3379 			 */
3380 			va.va_mask = AT_SIZE;
3381 			if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred) != 0) {
3382 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[6]);
3383 				page_unlock(targpp);
3384 				goto out;
3385 			}
3386 			if (btopr(va.va_size) < btopr(io_off + io_len)) {
3387 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[7]);
3388 				*downsize = 1;
3389 				*ret_pszc = 0;
3390 				page_unlock(targpp);
3391 				goto out;
3392 			}
3393 			io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3394 				B_READ, svd->cred);
3395 			if (io_err) {
3396 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[8]);
3397 				page_unlock(targpp);
3398 				if (io_err == EDEADLK) {
3399 					segvn_vmpss_pageio_deadlk_err++;
3400 				}
3401 				goto out;
3402 			}
3403 			nios++;
3404 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[9]);
3405 			while (io_pplist != NULL) {
3406 				pp = io_pplist;
3407 				page_sub(&io_pplist, pp);
3408 				ASSERT(page_iolock_assert(pp));
3409 				page_io_unlock(pp);
3410 				pgidx = (pp->p_offset - start_off) >>
3411 				    PAGESHIFT;
3412 				ASSERT(pgidx < pages);
3413 				ppa[pgidx] = pp;
3414 				page_list_concat(&done_pplist, &pp);
3415 			}
3416 		}
3417 		pp = targpp;
3418 		ASSERT(PAGE_EXCL(pp));
3419 		ASSERT(pp->p_szc <= pszc);
3420 		if (pszc != 0 && !group_page_trylock(pp, SE_EXCL)) {
3421 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[10]);
3422 			page_unlock(pp);
3423 			*downsize = 1;
3424 			*ret_pszc = pp->p_szc;
3425 			goto out;
3426 		}
3427 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[11]);
3428 		/*
3429 		 * page szc chould have changed before the entire group was
3430 		 * locked. reread page szc.
3431 		 */
3432 		pszc = pp->p_szc;
3433 		ppages = page_get_pagecnt(pszc);
3434 
3435 		/* link just the roots */
3436 		page_list_concat(&targ_pplist, &pp);
3437 		page_sub(&pplist, newpp);
3438 		page_list_concat(&repl_pplist, &newpp);
3439 		off += PAGESIZE;
3440 		while (--ppages != 0) {
3441 			newpp = pplist;
3442 			page_sub(&pplist, newpp);
3443 			off += PAGESIZE;
3444 		}
3445 		io_off = off;
3446 	}
3447 	if (io_pplist != NULL) {
3448 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[12]);
3449 		io_len = eoff - io_off;
3450 		va.va_mask = AT_SIZE;
3451 		if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred) != 0) {
3452 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[13]);
3453 			goto out;
3454 		}
3455 		if (btopr(va.va_size) < btopr(io_off + io_len)) {
3456 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[14]);
3457 			*downsize = 1;
3458 			*ret_pszc = 0;
3459 			goto out;
3460 		}
3461 		io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3462 		    B_READ, svd->cred);
3463 		if (io_err) {
3464 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[15]);
3465 			if (io_err == EDEADLK) {
3466 				segvn_vmpss_pageio_deadlk_err++;
3467 			}
3468 			goto out;
3469 		}
3470 		nios++;
3471 		while (io_pplist != NULL) {
3472 			pp = io_pplist;
3473 			page_sub(&io_pplist, pp);
3474 			ASSERT(page_iolock_assert(pp));
3475 			page_io_unlock(pp);
3476 			pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3477 			ASSERT(pgidx < pages);
3478 			ppa[pgidx] = pp;
3479 		}
3480 	}
3481 	/*
3482 	 * we're now bound to succeed or panic.
3483 	 * remove pages from done_pplist. it's not needed anymore.
3484 	 */
3485 	while (done_pplist != NULL) {
3486 		pp = done_pplist;
3487 		page_sub(&done_pplist, pp);
3488 	}
3489 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[16]);
3490 	ASSERT(pplist == NULL);
3491 	*ppplist = NULL;
3492 	while (targ_pplist != NULL) {
3493 		int ret;
3494 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[17]);
3495 		ASSERT(repl_pplist);
3496 		pp = targ_pplist;
3497 		page_sub(&targ_pplist, pp);
3498 		pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3499 		newpp = repl_pplist;
3500 		page_sub(&repl_pplist, newpp);
3501 #ifdef DEBUG
3502 		pfn = page_pptonum(pp);
3503 		pszc = pp->p_szc;
3504 		ppages = page_get_pagecnt(pszc);
3505 		ASSERT(IS_P2ALIGNED(pfn, ppages));
3506 		pfn = page_pptonum(newpp);
3507 		ASSERT(IS_P2ALIGNED(pfn, ppages));
3508 		ASSERT(P2PHASE(pfn, pages) == pgidx);
3509 #endif
3510 		nreloc = 0;
3511 		ret = page_relocate(&pp, &newpp, 0, 1, &nreloc, NULL);
3512 		if (ret != 0 || nreloc == 0) {
3513 			panic("segvn_fill_vp_pages: "
3514 			    "page_relocate failed");
3515 		}
3516 		pp = newpp;
3517 		while (nreloc-- != 0) {
3518 			ASSERT(PAGE_EXCL(pp));
3519 			ASSERT(pp->p_vnode == vp);
3520 			ASSERT(pgidx ==
3521 			    ((pp->p_offset - start_off) >> PAGESHIFT));
3522 			ppa[pgidx++] = pp;
3523 			pp++;
3524 		}
3525 	}
3526 
3527 	if (svd->type == MAP_PRIVATE) {
3528 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[18]);
3529 		for (i = 0; i < pages; i++) {
3530 			ASSERT(ppa[i] != NULL);
3531 			ASSERT(PAGE_EXCL(ppa[i]));
3532 			ASSERT(ppa[i]->p_vnode == vp);
3533 			ASSERT(ppa[i]->p_offset ==
3534 			    start_off + (i << PAGESHIFT));
3535 			page_downgrade(ppa[i]);
3536 		}
3537 		ppa[pages] = NULL;
3538 	} else {
3539 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[19]);
3540 		/*
3541 		 * the caller will still call VOP_GETPAGE() for shared segments
3542 		 * to check FS write permissions. For private segments we map
3543 		 * file read only anyway.  so no VOP_GETPAGE is needed.
3544 		 */
3545 		for (i = 0; i < pages; i++) {
3546 			ASSERT(ppa[i] != NULL);
3547 			ASSERT(PAGE_EXCL(ppa[i]));
3548 			ASSERT(ppa[i]->p_vnode == vp);
3549 			ASSERT(ppa[i]->p_offset ==
3550 			    start_off + (i << PAGESHIFT));
3551 			page_unlock(ppa[i]);
3552 		}
3553 		ppa[0] = NULL;
3554 	}
3555 
3556 	return (1);
3557 out:
3558 	/*
3559 	 * Do the cleanup. Unlock target pages we didn't relocate. They are
3560 	 * linked on targ_pplist by root pages. reassemble unused replacement
3561 	 * and io pages back to pplist.
3562 	 */
3563 	if (io_pplist != NULL) {
3564 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[20]);
3565 		pp = io_pplist;
3566 		do {
3567 			ASSERT(pp->p_vnode == vp);
3568 			ASSERT(pp->p_offset == io_off);
3569 			ASSERT(page_iolock_assert(pp));
3570 			page_io_unlock(pp);
3571 			page_hashout(pp, NULL);
3572 			io_off += PAGESIZE;
3573 		} while ((pp = pp->p_next) != io_pplist);
3574 		page_list_concat(&io_pplist, &pplist);
3575 		pplist = io_pplist;
3576 	}
3577 	tmp_pplist = NULL;
3578 	while (targ_pplist != NULL) {
3579 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[21]);
3580 		pp = targ_pplist;
3581 		ASSERT(PAGE_EXCL(pp));
3582 		page_sub(&targ_pplist, pp);
3583 
3584 		pszc = pp->p_szc;
3585 		ppages = page_get_pagecnt(pszc);
3586 		ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3587 
3588 		if (pszc != 0) {
3589 			group_page_unlock(pp);
3590 		}
3591 		page_unlock(pp);
3592 
3593 		pp = repl_pplist;
3594 		ASSERT(pp != NULL);
3595 		ASSERT(PAGE_EXCL(pp));
3596 		ASSERT(pp->p_szc == szc);
3597 		page_sub(&repl_pplist, pp);
3598 
3599 		ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3600 
3601 		/* relink replacement page */
3602 		page_list_concat(&tmp_pplist, &pp);
3603 		while (--ppages != 0) {
3604 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[22]);
3605 			pp++;
3606 			ASSERT(PAGE_EXCL(pp));
3607 			ASSERT(pp->p_szc == szc);
3608 			page_list_concat(&tmp_pplist, &pp);
3609 		}
3610 	}
3611 	if (tmp_pplist != NULL) {
3612 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[23]);
3613 		page_list_concat(&tmp_pplist, &pplist);
3614 		pplist = tmp_pplist;
3615 	}
3616 	/*
3617 	 * at this point all pages are either on done_pplist or
3618 	 * pplist. They can't be all on done_pplist otherwise
3619 	 * we'd've been done.
3620 	 */
3621 	ASSERT(pplist != NULL);
3622 	if (nios != 0) {
3623 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[24]);
3624 		pp = pplist;
3625 		do {
3626 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[25]);
3627 			ASSERT(pp->p_szc == szc);
3628 			ASSERT(PAGE_EXCL(pp));
3629 			ASSERT(pp->p_vnode != vp);
3630 			pp->p_szc = 0;
3631 		} while ((pp = pp->p_next) != pplist);
3632 
3633 		pp = done_pplist;
3634 		do {
3635 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[26]);
3636 			ASSERT(pp->p_szc == szc);
3637 			ASSERT(PAGE_EXCL(pp));
3638 			ASSERT(pp->p_vnode == vp);
3639 			pp->p_szc = 0;
3640 		} while ((pp = pp->p_next) != done_pplist);
3641 
3642 		while (pplist != NULL) {
3643 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[27]);
3644 			pp = pplist;
3645 			page_sub(&pplist, pp);
3646 			page_free(pp, 0);
3647 		}
3648 
3649 		while (done_pplist != NULL) {
3650 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[28]);
3651 			pp = done_pplist;
3652 			page_sub(&done_pplist, pp);
3653 			page_unlock(pp);
3654 		}
3655 		*ppplist = NULL;
3656 		return (0);
3657 	}
3658 	ASSERT(pplist == *ppplist);
3659 	if (io_err) {
3660 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[29]);
3661 		/*
3662 		 * don't downsize on io error.
3663 		 * see if vop_getpage succeeds.
3664 		 * pplist may still be used in this case
3665 		 * for relocations.
3666 		 */
3667 		return (0);
3668 	}
3669 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[30]);
3670 	page_free_replacement_page(pplist);
3671 	page_create_putback(pages);
3672 	*ppplist = NULL;
3673 	return (0);
3674 }
3675 
3676 int segvn_anypgsz = 0;
3677 
3678 #define	SEGVN_RESTORE_SOFTLOCK(type, pages) 		\
3679 		if ((type) == F_SOFTLOCK) {		\
3680 			mutex_enter(&freemem_lock);	\
3681 			availrmem += (pages);		\
3682 			segvn_pages_locked -= (pages);	\
3683 			svd->softlockcnt -= (pages);	\
3684 			mutex_exit(&freemem_lock);	\
3685 		}
3686 
3687 #define	SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot)		\
3688 		if (IS_VMODSORT((ppa)[0]->p_vnode)) {			\
3689 			if ((rw) == S_WRITE) {				\
3690 				for (i = 0; i < (pages); i++) {		\
3691 					ASSERT((ppa)[i]->p_vnode ==	\
3692 					    (ppa)[0]->p_vnode);		\
3693 					hat_setmod((ppa)[i]);		\
3694 				}					\
3695 			} else if ((rw) != S_OTHER &&			\
3696 			    ((prot) & (vpprot) & PROT_WRITE)) {		\
3697 				for (i = 0; i < (pages); i++) {		\
3698 					ASSERT((ppa)[i]->p_vnode ==	\
3699 					    (ppa)[0]->p_vnode);		\
3700 					if (!hat_ismod((ppa)[i])) {	\
3701 						prot &= ~PROT_WRITE;	\
3702 						break;			\
3703 					}				\
3704 				}					\
3705 			}						\
3706 		}
3707 
3708 #ifdef  VM_STATS
3709 
3710 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)					\
3711 		VM_STAT_ADD(segvnvmstats.fltvnpages[(idx)]);
3712 
3713 #else /* VM_STATS */
3714 
3715 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)
3716 
3717 #endif
3718 
3719 static faultcode_t
3720 segvn_fault_vnodepages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
3721     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
3722     caddr_t eaddr, int brkcow)
3723 {
3724 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
3725 	struct anon_map *amp = svd->amp;
3726 	uchar_t segtype = svd->type;
3727 	uint_t szc = seg->s_szc;
3728 	size_t pgsz = page_get_pagesize(szc);
3729 	size_t maxpgsz = pgsz;
3730 	pgcnt_t pages = btop(pgsz);
3731 	pgcnt_t maxpages = pages;
3732 	size_t ppasize = (pages + 1) * sizeof (page_t *);
3733 	caddr_t a = lpgaddr;
3734 	caddr_t	maxlpgeaddr = lpgeaddr;
3735 	u_offset_t off = svd->offset + (uintptr_t)(a - seg->s_base);
3736 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
3737 	struct vpage *vpage = (svd->vpage != NULL) ?
3738 	    &svd->vpage[seg_page(seg, a)] : NULL;
3739 	vnode_t *vp = svd->vp;
3740 	page_t **ppa;
3741 	uint_t	pszc;
3742 	size_t	ppgsz;
3743 	pgcnt_t	ppages;
3744 	faultcode_t err = 0;
3745 	int ierr;
3746 	int vop_size_err = 0;
3747 	uint_t protchk, prot, vpprot;
3748 	ulong_t i;
3749 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
3750 	anon_sync_obj_t an_cookie;
3751 	enum seg_rw arw;
3752 	int alloc_failed = 0;
3753 	int adjszc_chk;
3754 	struct vattr va;
3755 	int xhat = 0;
3756 	page_t *pplist;
3757 	pfn_t pfn;
3758 	int physcontig;
3759 	int upgrdfail;
3760 	int segvn_anypgsz_vnode = 0; /* for now map vnode with 2 page sizes */
3761 	int tron = (svd->tr_state == SEGVN_TR_ON);
3762 
3763 	ASSERT(szc != 0);
3764 	ASSERT(vp != NULL);
3765 	ASSERT(brkcow == 0 || amp != NULL);
3766 	ASSERT(tron == 0 || amp != NULL);
3767 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
3768 	ASSERT(!(svd->flags & MAP_NORESERVE));
3769 	ASSERT(type != F_SOFTUNLOCK);
3770 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
3771 	ASSERT(amp == NULL || IS_P2ALIGNED(aindx, maxpages));
3772 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
3773 	ASSERT(seg->s_szc < NBBY * sizeof (int));
3774 	ASSERT(type != F_SOFTLOCK || lpgeaddr - a == maxpgsz);
3775 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
3776 
3777 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltvnpages[0]);
3778 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltvnpages[1]);
3779 
3780 	if (svd->flags & MAP_TEXT) {
3781 		hat_flag |= HAT_LOAD_TEXT;
3782 	}
3783 
3784 	if (svd->pageprot) {
3785 		switch (rw) {
3786 		case S_READ:
3787 			protchk = PROT_READ;
3788 			break;
3789 		case S_WRITE:
3790 			protchk = PROT_WRITE;
3791 			break;
3792 		case S_EXEC:
3793 			protchk = PROT_EXEC;
3794 			break;
3795 		case S_OTHER:
3796 		default:
3797 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
3798 			break;
3799 		}
3800 	} else {
3801 		prot = svd->prot;
3802 		/* caller has already done segment level protection check. */
3803 	}
3804 
3805 	if (seg->s_as->a_hat != hat) {
3806 		xhat = 1;
3807 	}
3808 
3809 	if (rw == S_WRITE && segtype == MAP_PRIVATE) {
3810 		SEGVN_VMSTAT_FLTVNPAGES(2);
3811 		arw = S_READ;
3812 	} else {
3813 		arw = rw;
3814 	}
3815 
3816 	ppa = kmem_alloc(ppasize, KM_SLEEP);
3817 
3818 	VM_STAT_COND_ADD(amp != NULL, segvnvmstats.fltvnpages[3]);
3819 
3820 	for (;;) {
3821 		adjszc_chk = 0;
3822 		for (; a < lpgeaddr; a += pgsz, off += pgsz, aindx += pages) {
3823 			if (adjszc_chk) {
3824 				while (szc < seg->s_szc) {
3825 					uintptr_t e;
3826 					uint_t tszc;
3827 					tszc = segvn_anypgsz_vnode ? szc + 1 :
3828 					    seg->s_szc;
3829 					ppgsz = page_get_pagesize(tszc);
3830 					if (!IS_P2ALIGNED(a, ppgsz) ||
3831 					    ((alloc_failed >> tszc) &
3832 						0x1)) {
3833 						break;
3834 					}
3835 					SEGVN_VMSTAT_FLTVNPAGES(4);
3836 					szc = tszc;
3837 					pgsz = ppgsz;
3838 					pages = btop(pgsz);
3839 					e = P2ROUNDUP((uintptr_t)eaddr, pgsz);
3840 					lpgeaddr = (caddr_t)e;
3841 				}
3842 			}
3843 
3844 		again:
3845 			if (IS_P2ALIGNED(a, maxpgsz) && amp != NULL) {
3846 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
3847 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
3848 				anon_array_enter(amp, aindx, &an_cookie);
3849 				if (anon_get_ptr(amp->ahp, aindx) != NULL) {
3850 					SEGVN_VMSTAT_FLTVNPAGES(5);
3851 					ASSERT(anon_pages(amp->ahp, aindx,
3852 					    maxpages) == maxpages);
3853 					anon_array_exit(&an_cookie);
3854 					ANON_LOCK_EXIT(&amp->a_rwlock);
3855 					err = segvn_fault_anonpages(hat, seg,
3856 					    a, a + maxpgsz, type, rw,
3857 					    MAX(a, addr),
3858 					    MIN(a + maxpgsz, eaddr), brkcow);
3859 					if (err != 0) {
3860 						SEGVN_VMSTAT_FLTVNPAGES(6);
3861 						goto out;
3862 					}
3863 					if (szc < seg->s_szc) {
3864 						szc = seg->s_szc;
3865 						pgsz = maxpgsz;
3866 						pages = maxpages;
3867 						lpgeaddr = maxlpgeaddr;
3868 					}
3869 					goto next;
3870 				} else {
3871 					ASSERT(anon_pages(amp->ahp, aindx,
3872 					    maxpages) == 0);
3873 					SEGVN_VMSTAT_FLTVNPAGES(7);
3874 					anon_array_exit(&an_cookie);
3875 					ANON_LOCK_EXIT(&amp->a_rwlock);
3876 				}
3877 			}
3878 			ASSERT(!brkcow || IS_P2ALIGNED(a, maxpgsz));
3879 			ASSERT(!tron || IS_P2ALIGNED(a, maxpgsz));
3880 
3881 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
3882 				ASSERT(vpage != NULL);
3883 				prot = VPP_PROT(vpage);
3884 				ASSERT(sameprot(seg, a, maxpgsz));
3885 				if ((prot & protchk) == 0) {
3886 					SEGVN_VMSTAT_FLTVNPAGES(8);
3887 					err = FC_PROT;
3888 					goto out;
3889 				}
3890 			}
3891 			if (type == F_SOFTLOCK) {
3892 				mutex_enter(&freemem_lock);
3893 				if (availrmem < tune.t_minarmem + pages) {
3894 					mutex_exit(&freemem_lock);
3895 					err = FC_MAKE_ERR(ENOMEM);
3896 					goto out;
3897 				} else {
3898 					availrmem -= pages;
3899 					segvn_pages_locked += pages;
3900 					svd->softlockcnt += pages;
3901 				}
3902 				mutex_exit(&freemem_lock);
3903 			}
3904 
3905 			pplist = NULL;
3906 			physcontig = 0;
3907 			ppa[0] = NULL;
3908 			if (!brkcow && !tron && szc &&
3909 			    !page_exists_physcontig(vp, off, szc,
3910 				segtype == MAP_PRIVATE ? ppa : NULL)) {
3911 				SEGVN_VMSTAT_FLTVNPAGES(9);
3912 				if (page_alloc_pages(vp, seg, a, &pplist, NULL,
3913 				    szc, 0, 0) && type != F_SOFTLOCK) {
3914 					SEGVN_VMSTAT_FLTVNPAGES(10);
3915 					pszc = 0;
3916 					ierr = -1;
3917 					alloc_failed |= (1 << szc);
3918 					break;
3919 				}
3920 				if (pplist != NULL &&
3921 				    vp->v_mpssdata == SEGVN_PAGEIO) {
3922 					int downsize;
3923 					SEGVN_VMSTAT_FLTVNPAGES(11);
3924 					physcontig = segvn_fill_vp_pages(svd,
3925 					    vp, off, szc, ppa, &pplist,
3926 					    &pszc, &downsize);
3927 					ASSERT(!physcontig || pplist == NULL);
3928 					if (!physcontig && downsize &&
3929 					    type != F_SOFTLOCK) {
3930 						ASSERT(pplist == NULL);
3931 						SEGVN_VMSTAT_FLTVNPAGES(12);
3932 						ierr = -1;
3933 						break;
3934 					}
3935 					ASSERT(!physcontig ||
3936 					    segtype == MAP_PRIVATE ||
3937 					    ppa[0] == NULL);
3938 					if (physcontig && ppa[0] == NULL) {
3939 						physcontig = 0;
3940 					}
3941 				}
3942 			} else if (!brkcow && !tron && szc && ppa[0] != NULL) {
3943 				SEGVN_VMSTAT_FLTVNPAGES(13);
3944 				ASSERT(segtype == MAP_PRIVATE);
3945 				physcontig = 1;
3946 			}
3947 
3948 			if (!physcontig) {
3949 				SEGVN_VMSTAT_FLTVNPAGES(14);
3950 				ppa[0] = NULL;
3951 				ierr = VOP_GETPAGE(vp, (offset_t)off, pgsz,
3952 				    &vpprot, ppa, pgsz, seg, a, arw,
3953 				    svd->cred);
3954 #ifdef DEBUG
3955 				if (ierr == 0) {
3956 					for (i = 0; i < pages; i++) {
3957 						ASSERT(PAGE_LOCKED(ppa[i]));
3958 						ASSERT(!PP_ISFREE(ppa[i]));
3959 						ASSERT(ppa[i]->p_vnode == vp);
3960 						ASSERT(ppa[i]->p_offset ==
3961 						    off + (i << PAGESHIFT));
3962 					}
3963 				}
3964 #endif /* DEBUG */
3965 				if (segtype == MAP_PRIVATE) {
3966 					SEGVN_VMSTAT_FLTVNPAGES(15);
3967 					vpprot &= ~PROT_WRITE;
3968 				}
3969 			} else {
3970 				ASSERT(segtype == MAP_PRIVATE);
3971 				SEGVN_VMSTAT_FLTVNPAGES(16);
3972 				vpprot = PROT_ALL & ~PROT_WRITE;
3973 				ierr = 0;
3974 			}
3975 
3976 			if (ierr != 0) {
3977 				SEGVN_VMSTAT_FLTVNPAGES(17);
3978 				if (pplist != NULL) {
3979 					SEGVN_VMSTAT_FLTVNPAGES(18);
3980 					page_free_replacement_page(pplist);
3981 					page_create_putback(pages);
3982 				}
3983 				SEGVN_RESTORE_SOFTLOCK(type, pages);
3984 				if (a + pgsz <= eaddr) {
3985 					SEGVN_VMSTAT_FLTVNPAGES(19);
3986 					err = FC_MAKE_ERR(ierr);
3987 					goto out;
3988 				}
3989 				va.va_mask = AT_SIZE;
3990 				if (VOP_GETATTR(vp, &va, 0, svd->cred) != 0) {
3991 					SEGVN_VMSTAT_FLTVNPAGES(20);
3992 					err = FC_MAKE_ERR(EIO);
3993 					goto out;
3994 				}
3995 				if (btopr(va.va_size) >= btopr(off + pgsz)) {
3996 					SEGVN_VMSTAT_FLTVNPAGES(21);
3997 					err = FC_MAKE_ERR(ierr);
3998 					goto out;
3999 				}
4000 				if (btopr(va.va_size) <
4001 				    btopr(off + (eaddr - a))) {
4002 					SEGVN_VMSTAT_FLTVNPAGES(22);
4003 					err = FC_MAKE_ERR(ierr);
4004 					goto out;
4005 				}
4006 				if (brkcow || tron || type == F_SOFTLOCK) {
4007 					/* can't reduce map area */
4008 					SEGVN_VMSTAT_FLTVNPAGES(23);
4009 					vop_size_err = 1;
4010 					goto out;
4011 				}
4012 				SEGVN_VMSTAT_FLTVNPAGES(24);
4013 				ASSERT(szc != 0);
4014 				pszc = 0;
4015 				ierr = -1;
4016 				break;
4017 			}
4018 
4019 			if (amp != NULL) {
4020 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4021 				anon_array_enter(amp, aindx, &an_cookie);
4022 			}
4023 			if (amp != NULL &&
4024 			    anon_get_ptr(amp->ahp, aindx) != NULL) {
4025 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4026 
4027 				SEGVN_VMSTAT_FLTVNPAGES(25);
4028 				ASSERT(anon_pages(amp->ahp, taindx,
4029 				    maxpages) == maxpages);
4030 				for (i = 0; i < pages; i++) {
4031 					page_unlock(ppa[i]);
4032 				}
4033 				anon_array_exit(&an_cookie);
4034 				ANON_LOCK_EXIT(&amp->a_rwlock);
4035 				if (pplist != NULL) {
4036 					page_free_replacement_page(pplist);
4037 					page_create_putback(pages);
4038 				}
4039 				SEGVN_RESTORE_SOFTLOCK(type, pages);
4040 				if (szc < seg->s_szc) {
4041 					SEGVN_VMSTAT_FLTVNPAGES(26);
4042 					/*
4043 					 * For private segments SOFTLOCK
4044 					 * either always breaks cow (any rw
4045 					 * type except S_READ_NOCOW) or
4046 					 * address space is locked as writer
4047 					 * (S_READ_NOCOW case) and anon slots
4048 					 * can't show up on second check.
4049 					 * Therefore if we are here for
4050 					 * SOFTLOCK case it must be a cow
4051 					 * break but cow break never reduces
4052 					 * szc. text replication (tron) in
4053 					 * this case works as cow break.
4054 					 * Thus the assert below.
4055 					 */
4056 					ASSERT(!brkcow && !tron &&
4057 					    type != F_SOFTLOCK);
4058 					pszc = seg->s_szc;
4059 					ierr = -2;
4060 					break;
4061 				}
4062 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4063 				goto again;
4064 			}
4065 #ifdef DEBUG
4066 			if (amp != NULL) {
4067 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4068 				ASSERT(!anon_pages(amp->ahp, taindx, maxpages));
4069 			}
4070 #endif /* DEBUG */
4071 
4072 			if (brkcow || tron) {
4073 				ASSERT(amp != NULL);
4074 				ASSERT(pplist == NULL);
4075 				ASSERT(szc == seg->s_szc);
4076 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4077 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
4078 				SEGVN_VMSTAT_FLTVNPAGES(27);
4079 				ierr = anon_map_privatepages(amp, aindx, szc,
4080 				    seg, a, prot, ppa, vpage, segvn_anypgsz,
4081 				    tron ? PG_LOCAL : 0, svd->cred);
4082 				if (ierr != 0) {
4083 					SEGVN_VMSTAT_FLTVNPAGES(28);
4084 					anon_array_exit(&an_cookie);
4085 					ANON_LOCK_EXIT(&amp->a_rwlock);
4086 					SEGVN_RESTORE_SOFTLOCK(type, pages);
4087 					err = FC_MAKE_ERR(ierr);
4088 					goto out;
4089 				}
4090 
4091 				ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4092 				/*
4093 				 * p_szc can't be changed for locked
4094 				 * swapfs pages.
4095 				 */
4096 				ASSERT(svd->rcookie ==
4097 				    HAT_INVALID_REGION_COOKIE);
4098 				hat_memload_array(hat, a, pgsz, ppa, prot,
4099 				    hat_flag);
4100 
4101 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4102 					SEGVN_VMSTAT_FLTVNPAGES(29);
4103 					for (i = 0; i < pages; i++) {
4104 						page_unlock(ppa[i]);
4105 					}
4106 				}
4107 				anon_array_exit(&an_cookie);
4108 				ANON_LOCK_EXIT(&amp->a_rwlock);
4109 				goto next;
4110 			}
4111 
4112 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
4113 			    (!svd->pageprot && svd->prot == (prot & vpprot)));
4114 
4115 			pfn = page_pptonum(ppa[0]);
4116 			/*
4117 			 * hat_page_demote() needs an SE_EXCL lock on one of
4118 			 * constituent page_t's and it decreases root's p_szc
4119 			 * last. This means if root's p_szc is equal szc and
4120 			 * all its constituent pages are locked
4121 			 * hat_page_demote() that could have changed p_szc to
4122 			 * szc is already done and no new have page_demote()
4123 			 * can start for this large page.
4124 			 */
4125 
4126 			/*
4127 			 * we need to make sure same mapping size is used for
4128 			 * the same address range if there's a possibility the
4129 			 * adddress is already mapped because hat layer panics
4130 			 * when translation is loaded for the range already
4131 			 * mapped with a different page size.  We achieve it
4132 			 * by always using largest page size possible subject
4133 			 * to the constraints of page size, segment page size
4134 			 * and page alignment.  Since mappings are invalidated
4135 			 * when those constraints change and make it
4136 			 * impossible to use previously used mapping size no
4137 			 * mapping size conflicts should happen.
4138 			 */
4139 
4140 		chkszc:
4141 			if ((pszc = ppa[0]->p_szc) == szc &&
4142 			    IS_P2ALIGNED(pfn, pages)) {
4143 
4144 				SEGVN_VMSTAT_FLTVNPAGES(30);
4145 #ifdef DEBUG
4146 				for (i = 0; i < pages; i++) {
4147 					ASSERT(PAGE_LOCKED(ppa[i]));
4148 					ASSERT(!PP_ISFREE(ppa[i]));
4149 					ASSERT(page_pptonum(ppa[i]) ==
4150 					    pfn + i);
4151 					ASSERT(ppa[i]->p_szc == szc);
4152 					ASSERT(ppa[i]->p_vnode == vp);
4153 					ASSERT(ppa[i]->p_offset ==
4154 					    off + (i << PAGESHIFT));
4155 				}
4156 #endif /* DEBUG */
4157 				/*
4158 				 * All pages are of szc we need and they are
4159 				 * all locked so they can't change szc. load
4160 				 * translations.
4161 				 *
4162 				 * if page got promoted since last check
4163 				 * we don't need pplist.
4164 				 */
4165 				if (pplist != NULL) {
4166 					page_free_replacement_page(pplist);
4167 					page_create_putback(pages);
4168 				}
4169 				if (PP_ISMIGRATE(ppa[0])) {
4170 					page_migrate(seg, a, ppa, pages);
4171 				}
4172 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4173 				    prot, vpprot);
4174 				if (!xhat) {
4175 					hat_memload_array_region(hat, a, pgsz,
4176 					    ppa, prot & vpprot, hat_flag,
4177 					    svd->rcookie);
4178 				} else {
4179 					/*
4180 					 * avoid large xhat mappings to FS
4181 					 * pages so that hat_page_demote()
4182 					 * doesn't need to check for xhat
4183 					 * large mappings.
4184 					 * Don't use regions with xhats.
4185 					 */
4186 					for (i = 0; i < pages; i++) {
4187 						hat_memload(hat,
4188 						    a + (i << PAGESHIFT),
4189 						    ppa[i], prot & vpprot,
4190 						    hat_flag);
4191 					}
4192 				}
4193 
4194 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4195 					for (i = 0; i < pages; i++) {
4196 						page_unlock(ppa[i]);
4197 					}
4198 				}
4199 				if (amp != NULL) {
4200 					anon_array_exit(&an_cookie);
4201 					ANON_LOCK_EXIT(&amp->a_rwlock);
4202 				}
4203 				goto next;
4204 			}
4205 
4206 			/*
4207 			 * See if upsize is possible.
4208 			 */
4209 			if (pszc > szc && szc < seg->s_szc &&
4210 			    (segvn_anypgsz_vnode || pszc >= seg->s_szc)) {
4211 				pgcnt_t aphase;
4212 				uint_t pszc1 = MIN(pszc, seg->s_szc);
4213 				ppgsz = page_get_pagesize(pszc1);
4214 				ppages = btop(ppgsz);
4215 				aphase = btop(P2PHASE((uintptr_t)a, ppgsz));
4216 
4217 				ASSERT(type != F_SOFTLOCK);
4218 
4219 				SEGVN_VMSTAT_FLTVNPAGES(31);
4220 				if (aphase != P2PHASE(pfn, ppages)) {
4221 					segvn_faultvnmpss_align_err4++;
4222 				} else {
4223 					SEGVN_VMSTAT_FLTVNPAGES(32);
4224 					if (pplist != NULL) {
4225 						page_t *pl = pplist;
4226 						page_free_replacement_page(pl);
4227 						page_create_putback(pages);
4228 					}
4229 					for (i = 0; i < pages; i++) {
4230 						page_unlock(ppa[i]);
4231 					}
4232 					if (amp != NULL) {
4233 						anon_array_exit(&an_cookie);
4234 						ANON_LOCK_EXIT(&amp->a_rwlock);
4235 					}
4236 					pszc = pszc1;
4237 					ierr = -2;
4238 					break;
4239 				}
4240 			}
4241 
4242 			/*
4243 			 * check if we should use smallest mapping size.
4244 			 */
4245 			upgrdfail = 0;
4246 			if (szc == 0 || xhat ||
4247 			    (pszc >= szc &&
4248 			    !IS_P2ALIGNED(pfn, pages)) ||
4249 			    (pszc < szc &&
4250 			    !segvn_full_szcpages(ppa, szc, &upgrdfail,
4251 				&pszc))) {
4252 
4253 				if (upgrdfail && type != F_SOFTLOCK) {
4254 					/*
4255 					 * segvn_full_szcpages failed to lock
4256 					 * all pages EXCL. Size down.
4257 					 */
4258 					ASSERT(pszc < szc);
4259 
4260 					SEGVN_VMSTAT_FLTVNPAGES(33);
4261 
4262 					if (pplist != NULL) {
4263 						page_t *pl = pplist;
4264 						page_free_replacement_page(pl);
4265 						page_create_putback(pages);
4266 					}
4267 
4268 					for (i = 0; i < pages; i++) {
4269 						page_unlock(ppa[i]);
4270 					}
4271 					if (amp != NULL) {
4272 						anon_array_exit(&an_cookie);
4273 						ANON_LOCK_EXIT(&amp->a_rwlock);
4274 					}
4275 					ierr = -1;
4276 					break;
4277 				}
4278 				if (szc != 0 && !xhat && !upgrdfail) {
4279 					segvn_faultvnmpss_align_err5++;
4280 				}
4281 				SEGVN_VMSTAT_FLTVNPAGES(34);
4282 				if (pplist != NULL) {
4283 					page_free_replacement_page(pplist);
4284 					page_create_putback(pages);
4285 				}
4286 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4287 				    prot, vpprot);
4288 				if (upgrdfail && segvn_anypgsz_vnode) {
4289 					/* SOFTLOCK case */
4290 					hat_memload_array_region(hat, a, pgsz,
4291 					    ppa, prot & vpprot, hat_flag,
4292 					    svd->rcookie);
4293 				} else {
4294 					for (i = 0; i < pages; i++) {
4295 						hat_memload_region(hat,
4296 						    a + (i << PAGESHIFT),
4297 						    ppa[i], prot & vpprot,
4298 						    hat_flag, svd->rcookie);
4299 					}
4300 				}
4301 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4302 					for (i = 0; i < pages; i++) {
4303 						page_unlock(ppa[i]);
4304 					}
4305 				}
4306 				if (amp != NULL) {
4307 					anon_array_exit(&an_cookie);
4308 					ANON_LOCK_EXIT(&amp->a_rwlock);
4309 				}
4310 				goto next;
4311 			}
4312 
4313 			if (pszc == szc) {
4314 				/*
4315 				 * segvn_full_szcpages() upgraded pages szc.
4316 				 */
4317 				ASSERT(pszc == ppa[0]->p_szc);
4318 				ASSERT(IS_P2ALIGNED(pfn, pages));
4319 				goto chkszc;
4320 			}
4321 
4322 			if (pszc > szc) {
4323 				kmutex_t *szcmtx;
4324 				SEGVN_VMSTAT_FLTVNPAGES(35);
4325 				/*
4326 				 * p_szc of ppa[0] can change since we haven't
4327 				 * locked all constituent pages. Call
4328 				 * page_lock_szc() to prevent szc changes.
4329 				 * This should be a rare case that happens when
4330 				 * multiple segments use a different page size
4331 				 * to map the same file offsets.
4332 				 */
4333 				szcmtx = page_szc_lock(ppa[0]);
4334 				pszc = ppa[0]->p_szc;
4335 				ASSERT(szcmtx != NULL || pszc == 0);
4336 				ASSERT(ppa[0]->p_szc <= pszc);
4337 				if (pszc <= szc) {
4338 					SEGVN_VMSTAT_FLTVNPAGES(36);
4339 					if (szcmtx != NULL) {
4340 						mutex_exit(szcmtx);
4341 					}
4342 					goto chkszc;
4343 				}
4344 				if (pplist != NULL) {
4345 					/*
4346 					 * page got promoted since last check.
4347 					 * we don't need preaalocated large
4348 					 * page.
4349 					 */
4350 					SEGVN_VMSTAT_FLTVNPAGES(37);
4351 					page_free_replacement_page(pplist);
4352 					page_create_putback(pages);
4353 				}
4354 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4355 				    prot, vpprot);
4356 				hat_memload_array_region(hat, a, pgsz, ppa,
4357 				    prot & vpprot, hat_flag, svd->rcookie);
4358 				mutex_exit(szcmtx);
4359 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4360 					for (i = 0; i < pages; i++) {
4361 						page_unlock(ppa[i]);
4362 					}
4363 				}
4364 				if (amp != NULL) {
4365 					anon_array_exit(&an_cookie);
4366 					ANON_LOCK_EXIT(&amp->a_rwlock);
4367 				}
4368 				goto next;
4369 			}
4370 
4371 			/*
4372 			 * if page got demoted since last check
4373 			 * we could have not allocated larger page.
4374 			 * allocate now.
4375 			 */
4376 			if (pplist == NULL &&
4377 			    page_alloc_pages(vp, seg, a, &pplist, NULL,
4378 				szc, 0, 0) && type != F_SOFTLOCK) {
4379 				SEGVN_VMSTAT_FLTVNPAGES(38);
4380 				for (i = 0; i < pages; i++) {
4381 					page_unlock(ppa[i]);
4382 				}
4383 				if (amp != NULL) {
4384 					anon_array_exit(&an_cookie);
4385 					ANON_LOCK_EXIT(&amp->a_rwlock);
4386 				}
4387 				ierr = -1;
4388 				alloc_failed |= (1 << szc);
4389 				break;
4390 			}
4391 
4392 			SEGVN_VMSTAT_FLTVNPAGES(39);
4393 
4394 			if (pplist != NULL) {
4395 				segvn_relocate_pages(ppa, pplist);
4396 #ifdef DEBUG
4397 			} else {
4398 				ASSERT(type == F_SOFTLOCK);
4399 				SEGVN_VMSTAT_FLTVNPAGES(40);
4400 #endif /* DEBUG */
4401 			}
4402 
4403 			SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot);
4404 
4405 			if (pplist == NULL && segvn_anypgsz_vnode == 0) {
4406 				ASSERT(type == F_SOFTLOCK);
4407 				for (i = 0; i < pages; i++) {
4408 					ASSERT(ppa[i]->p_szc < szc);
4409 					hat_memload_region(hat,
4410 					    a + (i << PAGESHIFT),
4411 					    ppa[i], prot & vpprot, hat_flag,
4412 					    svd->rcookie);
4413 				}
4414 			} else {
4415 				ASSERT(pplist != NULL || type == F_SOFTLOCK);
4416 				hat_memload_array_region(hat, a, pgsz, ppa,
4417 				    prot & vpprot, hat_flag, svd->rcookie);
4418 			}
4419 			if (!(hat_flag & HAT_LOAD_LOCK)) {
4420 				for (i = 0; i < pages; i++) {
4421 					ASSERT(PAGE_SHARED(ppa[i]));
4422 					page_unlock(ppa[i]);
4423 				}
4424 			}
4425 			if (amp != NULL) {
4426 				anon_array_exit(&an_cookie);
4427 				ANON_LOCK_EXIT(&amp->a_rwlock);
4428 			}
4429 
4430 		next:
4431 			if (vpage != NULL) {
4432 				vpage += pages;
4433 			}
4434 			adjszc_chk = 1;
4435 		}
4436 		if (a == lpgeaddr)
4437 			break;
4438 		ASSERT(a < lpgeaddr);
4439 
4440 		ASSERT(!brkcow && !tron && type != F_SOFTLOCK);
4441 
4442 		/*
4443 		 * ierr == -1 means we failed to map with a large page.
4444 		 * (either due to allocation/relocation failures or
4445 		 * misalignment with other mappings to this file.
4446 		 *
4447 		 * ierr == -2 means some other thread allocated a large page
4448 		 * after we gave up tp map with a large page.  retry with
4449 		 * larger mapping.
4450 		 */
4451 		ASSERT(ierr == -1 || ierr == -2);
4452 		ASSERT(ierr == -2 || szc != 0);
4453 		ASSERT(ierr == -1 || szc < seg->s_szc);
4454 		if (ierr == -2) {
4455 			SEGVN_VMSTAT_FLTVNPAGES(41);
4456 			ASSERT(pszc > szc && pszc <= seg->s_szc);
4457 			szc = pszc;
4458 		} else if (segvn_anypgsz_vnode) {
4459 			SEGVN_VMSTAT_FLTVNPAGES(42);
4460 			szc--;
4461 		} else {
4462 			SEGVN_VMSTAT_FLTVNPAGES(43);
4463 			ASSERT(pszc < szc);
4464 			/*
4465 			 * other process created pszc large page.
4466 			 * but we still have to drop to 0 szc.
4467 			 */
4468 			szc = 0;
4469 		}
4470 
4471 		pgsz = page_get_pagesize(szc);
4472 		pages = btop(pgsz);
4473 		if (ierr == -2) {
4474 			/*
4475 			 * Size up case. Note lpgaddr may only be needed for
4476 			 * softlock case so we don't adjust it here.
4477 			 */
4478 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4479 			ASSERT(a >= lpgaddr);
4480 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4481 			off = svd->offset + (uintptr_t)(a - seg->s_base);
4482 			aindx = svd->anon_index + seg_page(seg, a);
4483 			vpage = (svd->vpage != NULL) ?
4484 			    &svd->vpage[seg_page(seg, a)] : NULL;
4485 		} else {
4486 			/*
4487 			 * Size down case. Note lpgaddr may only be needed for
4488 			 * softlock case so we don't adjust it here.
4489 			 */
4490 			ASSERT(IS_P2ALIGNED(a, pgsz));
4491 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4492 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4493 			ASSERT(a < lpgeaddr);
4494 			if (a < addr) {
4495 				SEGVN_VMSTAT_FLTVNPAGES(44);
4496 				/*
4497 				 * The beginning of the large page region can
4498 				 * be pulled to the right to make a smaller
4499 				 * region. We haven't yet faulted a single
4500 				 * page.
4501 				 */
4502 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4503 				ASSERT(a >= lpgaddr);
4504 				off = svd->offset +
4505 				    (uintptr_t)(a - seg->s_base);
4506 				aindx = svd->anon_index + seg_page(seg, a);
4507 				vpage = (svd->vpage != NULL) ?
4508 				    &svd->vpage[seg_page(seg, a)] : NULL;
4509 			}
4510 		}
4511 	}
4512 out:
4513 	kmem_free(ppa, ppasize);
4514 	if (!err && !vop_size_err) {
4515 		SEGVN_VMSTAT_FLTVNPAGES(45);
4516 		return (0);
4517 	}
4518 	if (type == F_SOFTLOCK && a > lpgaddr) {
4519 		SEGVN_VMSTAT_FLTVNPAGES(46);
4520 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4521 	}
4522 	if (!vop_size_err) {
4523 		SEGVN_VMSTAT_FLTVNPAGES(47);
4524 		return (err);
4525 	}
4526 	ASSERT(brkcow || tron || type == F_SOFTLOCK);
4527 	/*
4528 	 * Large page end is mapped beyond the end of file and it's a cow
4529 	 * fault (can be a text replication induced cow) or softlock so we can't
4530 	 * reduce the map area.  For now just demote the segment. This should
4531 	 * really only happen if the end of the file changed after the mapping
4532 	 * was established since when large page segments are created we make
4533 	 * sure they don't extend beyond the end of the file.
4534 	 */
4535 	SEGVN_VMSTAT_FLTVNPAGES(48);
4536 
4537 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4538 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4539 	err = 0;
4540 	if (seg->s_szc != 0) {
4541 		segvn_fltvnpages_clrszc_cnt++;
4542 		ASSERT(svd->softlockcnt == 0);
4543 		err = segvn_clrszc(seg);
4544 		if (err != 0) {
4545 			segvn_fltvnpages_clrszc_err++;
4546 		}
4547 	}
4548 	ASSERT(err || seg->s_szc == 0);
4549 	SEGVN_LOCK_DOWNGRADE(seg->s_as, &svd->lock);
4550 	/* segvn_fault will do its job as if szc had been zero to begin with */
4551 	return (err == 0 ? IE_RETRY : FC_MAKE_ERR(err));
4552 }
4553 
4554 /*
4555  * This routine will attempt to fault in one large page.
4556  * it will use smaller pages if that fails.
4557  * It should only be called for pure anonymous segments.
4558  */
4559 static faultcode_t
4560 segvn_fault_anonpages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
4561     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
4562     caddr_t eaddr, int brkcow)
4563 {
4564 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4565 	struct anon_map *amp = svd->amp;
4566 	uchar_t segtype = svd->type;
4567 	uint_t szc = seg->s_szc;
4568 	size_t pgsz = page_get_pagesize(szc);
4569 	size_t maxpgsz = pgsz;
4570 	pgcnt_t pages = btop(pgsz);
4571 	size_t ppasize = pages * sizeof (page_t *);
4572 	caddr_t a = lpgaddr;
4573 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
4574 	struct vpage *vpage = (svd->vpage != NULL) ?
4575 	    &svd->vpage[seg_page(seg, a)] : NULL;
4576 	page_t **ppa;
4577 	uint_t	ppa_szc;
4578 	faultcode_t err;
4579 	int ierr;
4580 	uint_t protchk, prot, vpprot;
4581 	ulong_t i;
4582 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
4583 	anon_sync_obj_t cookie;
4584 	int first = 1;
4585 	int adjszc_chk;
4586 	int purged = 0;
4587 	int pgflags = (svd->tr_state == SEGVN_TR_ON) ? PG_LOCAL : 0;
4588 
4589 	ASSERT(szc != 0);
4590 	ASSERT(amp != NULL);
4591 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
4592 	ASSERT(!(svd->flags & MAP_NORESERVE));
4593 	ASSERT(type != F_SOFTUNLOCK);
4594 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
4595 	ASSERT(!brkcow || svd->tr_state == SEGVN_TR_OFF);
4596 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
4597 
4598 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
4599 
4600 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltanpages[0]);
4601 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltanpages[1]);
4602 
4603 	if (svd->flags & MAP_TEXT) {
4604 		hat_flag |= HAT_LOAD_TEXT;
4605 	}
4606 
4607 	if (svd->pageprot) {
4608 		switch (rw) {
4609 		case S_READ:
4610 			protchk = PROT_READ;
4611 			break;
4612 		case S_WRITE:
4613 			protchk = PROT_WRITE;
4614 			break;
4615 		case S_EXEC:
4616 			protchk = PROT_EXEC;
4617 			break;
4618 		case S_OTHER:
4619 		default:
4620 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4621 			break;
4622 		}
4623 		VM_STAT_ADD(segvnvmstats.fltanpages[2]);
4624 	} else {
4625 		prot = svd->prot;
4626 		/* caller has already done segment level protection check. */
4627 	}
4628 
4629 	ppa = kmem_alloc(ppasize, KM_SLEEP);
4630 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4631 	for (;;) {
4632 		adjszc_chk = 0;
4633 		for (; a < lpgeaddr; a += pgsz, aindx += pages) {
4634 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
4635 				VM_STAT_ADD(segvnvmstats.fltanpages[3]);
4636 				ASSERT(vpage != NULL);
4637 				prot = VPP_PROT(vpage);
4638 				ASSERT(sameprot(seg, a, maxpgsz));
4639 				if ((prot & protchk) == 0) {
4640 					err = FC_PROT;
4641 					goto error;
4642 				}
4643 			}
4644 			if (adjszc_chk && IS_P2ALIGNED(a, maxpgsz) &&
4645 			    pgsz < maxpgsz) {
4646 				ASSERT(a > lpgaddr);
4647 				szc = seg->s_szc;
4648 				pgsz = maxpgsz;
4649 				pages = btop(pgsz);
4650 				ASSERT(IS_P2ALIGNED(aindx, pages));
4651 				lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr,
4652 				    pgsz);
4653 			}
4654 			if (type == F_SOFTLOCK && svd->vp != NULL) {
4655 				mutex_enter(&freemem_lock);
4656 				if (availrmem < tune.t_minarmem + pages) {
4657 					mutex_exit(&freemem_lock);
4658 					err = FC_MAKE_ERR(ENOMEM);
4659 					goto error;
4660 				} else {
4661 					availrmem -= pages;
4662 					segvn_pages_locked += pages;
4663 					svd->softlockcnt += pages;
4664 				}
4665 				mutex_exit(&freemem_lock);
4666 			}
4667 			anon_array_enter(amp, aindx, &cookie);
4668 			ppa_szc = (uint_t)-1;
4669 			ierr = anon_map_getpages(amp, aindx, szc, seg, a,
4670 				prot, &vpprot, ppa, &ppa_szc, vpage, rw, brkcow,
4671 				segvn_anypgsz, pgflags, svd->cred);
4672 			if (ierr != 0) {
4673 				anon_array_exit(&cookie);
4674 				VM_STAT_ADD(segvnvmstats.fltanpages[4]);
4675 				if (type == F_SOFTLOCK && svd->vp != NULL) {
4676 					VM_STAT_ADD(segvnvmstats.fltanpages[5]);
4677 					mutex_enter(&freemem_lock);
4678 					availrmem += pages;
4679 					segvn_pages_locked -= pages;
4680 					svd->softlockcnt -= pages;
4681 					mutex_exit(&freemem_lock);
4682 				}
4683 				if (ierr > 0) {
4684 					VM_STAT_ADD(segvnvmstats.fltanpages[6]);
4685 					err = FC_MAKE_ERR(ierr);
4686 					goto error;
4687 				}
4688 				break;
4689 			}
4690 
4691 			ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4692 
4693 			ASSERT(segtype == MAP_SHARED ||
4694 			    ppa[0]->p_szc <= szc);
4695 			ASSERT(segtype == MAP_PRIVATE ||
4696 			    ppa[0]->p_szc >= szc);
4697 
4698 			/*
4699 			 * Handle pages that have been marked for migration
4700 			 */
4701 			if (lgrp_optimizations())
4702 				page_migrate(seg, a, ppa, pages);
4703 
4704 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
4705 			if (type == F_SOFTLOCK && svd->vp == NULL) {
4706 				/*
4707 				 * All pages in ppa array belong to the same
4708 				 * large page. This means it's ok to call
4709 				 * segvn_pp_lock_anonpages just for ppa[0].
4710 				 */
4711 				if (!segvn_pp_lock_anonpages(ppa[0], first)) {
4712 					for (i = 0; i < pages; i++) {
4713 						page_unlock(ppa[i]);
4714 					}
4715 					err = FC_MAKE_ERR(ENOMEM);
4716 					goto error;
4717 				}
4718 				first = 0;
4719 				mutex_enter(&freemem_lock);
4720 				svd->softlockcnt += pages;
4721 				segvn_pages_locked += pages;
4722 				mutex_exit(&freemem_lock);
4723 			}
4724 
4725 			if (segtype == MAP_SHARED) {
4726 				vpprot |= PROT_WRITE;
4727 			}
4728 
4729 			hat_memload_array(hat, a, pgsz, ppa,
4730 			    prot & vpprot, hat_flag);
4731 
4732 			if (hat_flag & HAT_LOAD_LOCK) {
4733 				VM_STAT_ADD(segvnvmstats.fltanpages[7]);
4734 			} else {
4735 				VM_STAT_ADD(segvnvmstats.fltanpages[8]);
4736 				for (i = 0; i < pages; i++)
4737 					page_unlock(ppa[i]);
4738 			}
4739 			if (vpage != NULL)
4740 				vpage += pages;
4741 
4742 			anon_array_exit(&cookie);
4743 			adjszc_chk = 1;
4744 		}
4745 		if (a == lpgeaddr)
4746 			break;
4747 		ASSERT(a < lpgeaddr);
4748 		/*
4749 		 * ierr == -1 means we failed to allocate a large page.
4750 		 * so do a size down operation.
4751 		 *
4752 		 * ierr == -2 means some other process that privately shares
4753 		 * pages with this process has allocated a larger page and we
4754 		 * need to retry with larger pages. So do a size up
4755 		 * operation. This relies on the fact that large pages are
4756 		 * never partially shared i.e. if we share any constituent
4757 		 * page of a large page with another process we must share the
4758 		 * entire large page. Note this cannot happen for SOFTLOCK
4759 		 * case, unless current address (a) is at the beginning of the
4760 		 * next page size boundary because the other process couldn't
4761 		 * have relocated locked pages.
4762 		 */
4763 		ASSERT(ierr == -1 || ierr == -2);
4764 		/*
4765 		 * For the very first relocation failure try to purge this
4766 		 * segment's cache so that the relocator can obtain an
4767 		 * exclusive lock on pages we want to relocate.
4768 		 */
4769 		if (!purged && ierr == -1 && ppa_szc != (uint_t)-1 &&
4770 		    svd->softlockcnt != 0) {
4771 			purged = 1;
4772 			segvn_purge(seg);
4773 			continue;
4774 		}
4775 
4776 		if (segvn_anypgsz) {
4777 			ASSERT(ierr == -2 || szc != 0);
4778 			ASSERT(ierr == -1 || szc < seg->s_szc);
4779 			szc = (ierr == -1) ? szc - 1 : szc + 1;
4780 		} else {
4781 			/*
4782 			 * For non COW faults and segvn_anypgsz == 0
4783 			 * we need to be careful not to loop forever
4784 			 * if existing page is found with szc other
4785 			 * than 0 or seg->s_szc. This could be due
4786 			 * to page relocations on behalf of DR or
4787 			 * more likely large page creation. For this
4788 			 * case simply re-size to existing page's szc
4789 			 * if returned by anon_map_getpages().
4790 			 */
4791 			if (ppa_szc == (uint_t)-1) {
4792 				szc = (ierr == -1) ? 0 : seg->s_szc;
4793 			} else {
4794 				ASSERT(ppa_szc <= seg->s_szc);
4795 				ASSERT(ierr == -2 || ppa_szc < szc);
4796 				ASSERT(ierr == -1 || ppa_szc > szc);
4797 				szc = ppa_szc;
4798 			}
4799 		}
4800 
4801 		pgsz = page_get_pagesize(szc);
4802 		pages = btop(pgsz);
4803 		ASSERT(type != F_SOFTLOCK || ierr == -1 ||
4804 		    (IS_P2ALIGNED(a, pgsz) && IS_P2ALIGNED(lpgeaddr, pgsz)));
4805 		if (type == F_SOFTLOCK) {
4806 			/*
4807 			 * For softlocks we cannot reduce the fault area
4808 			 * (calculated based on the largest page size for this
4809 			 * segment) for size down and a is already next
4810 			 * page size aligned as assertted above for size
4811 			 * ups. Therefore just continue in case of softlock.
4812 			 */
4813 			VM_STAT_ADD(segvnvmstats.fltanpages[9]);
4814 			continue; /* keep lint happy */
4815 		} else if (ierr == -2) {
4816 
4817 			/*
4818 			 * Size up case. Note lpgaddr may only be needed for
4819 			 * softlock case so we don't adjust it here.
4820 			 */
4821 			VM_STAT_ADD(segvnvmstats.fltanpages[10]);
4822 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4823 			ASSERT(a >= lpgaddr);
4824 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4825 			aindx = svd->anon_index + seg_page(seg, a);
4826 			vpage = (svd->vpage != NULL) ?
4827 			    &svd->vpage[seg_page(seg, a)] : NULL;
4828 		} else {
4829 			/*
4830 			 * Size down case. Note lpgaddr may only be needed for
4831 			 * softlock case so we don't adjust it here.
4832 			 */
4833 			VM_STAT_ADD(segvnvmstats.fltanpages[11]);
4834 			ASSERT(IS_P2ALIGNED(a, pgsz));
4835 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4836 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4837 			ASSERT(a < lpgeaddr);
4838 			if (a < addr) {
4839 				/*
4840 				 * The beginning of the large page region can
4841 				 * be pulled to the right to make a smaller
4842 				 * region. We haven't yet faulted a single
4843 				 * page.
4844 				 */
4845 				VM_STAT_ADD(segvnvmstats.fltanpages[12]);
4846 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4847 				ASSERT(a >= lpgaddr);
4848 				aindx = svd->anon_index + seg_page(seg, a);
4849 				vpage = (svd->vpage != NULL) ?
4850 				    &svd->vpage[seg_page(seg, a)] : NULL;
4851 			}
4852 		}
4853 	}
4854 	VM_STAT_ADD(segvnvmstats.fltanpages[13]);
4855 	ANON_LOCK_EXIT(&amp->a_rwlock);
4856 	kmem_free(ppa, ppasize);
4857 	return (0);
4858 error:
4859 	VM_STAT_ADD(segvnvmstats.fltanpages[14]);
4860 	ANON_LOCK_EXIT(&amp->a_rwlock);
4861 	kmem_free(ppa, ppasize);
4862 	if (type == F_SOFTLOCK && a > lpgaddr) {
4863 		VM_STAT_ADD(segvnvmstats.fltanpages[15]);
4864 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4865 	}
4866 	return (err);
4867 }
4868 
4869 int fltadvice = 1;	/* set to free behind pages for sequential access */
4870 
4871 /*
4872  * This routine is called via a machine specific fault handling routine.
4873  * It is also called by software routines wishing to lock or unlock
4874  * a range of addresses.
4875  *
4876  * Here is the basic algorithm:
4877  *	If unlocking
4878  *		Call segvn_softunlock
4879  *		Return
4880  *	endif
4881  *	Checking and set up work
4882  *	If we will need some non-anonymous pages
4883  *		Call VOP_GETPAGE over the range of non-anonymous pages
4884  *	endif
4885  *	Loop over all addresses requested
4886  *		Call segvn_faultpage passing in page list
4887  *		    to load up translations and handle anonymous pages
4888  *	endloop
4889  *	Load up translation to any additional pages in page list not
4890  *	    already handled that fit into this segment
4891  */
4892 static faultcode_t
4893 segvn_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
4894     enum fault_type type, enum seg_rw rw)
4895 {
4896 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4897 	page_t **plp, **ppp, *pp;
4898 	u_offset_t off;
4899 	caddr_t a;
4900 	struct vpage *vpage;
4901 	uint_t vpprot, prot;
4902 	int err;
4903 	page_t *pl[PVN_GETPAGE_NUM + 1];
4904 	size_t plsz, pl_alloc_sz;
4905 	size_t page;
4906 	ulong_t anon_index;
4907 	struct anon_map *amp;
4908 	int dogetpage = 0;
4909 	caddr_t	lpgaddr, lpgeaddr;
4910 	size_t pgsz;
4911 	anon_sync_obj_t cookie;
4912 	int brkcow = BREAK_COW_SHARE(rw, type, svd->type);
4913 
4914 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
4915 	ASSERT(svd->amp == NULL || svd->rcookie == HAT_INVALID_REGION_COOKIE);
4916 
4917 	/*
4918 	 * First handle the easy stuff
4919 	 */
4920 	if (type == F_SOFTUNLOCK) {
4921 		if (rw == S_READ_NOCOW) {
4922 			rw = S_READ;
4923 			ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
4924 		}
4925 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4926 		pgsz = (seg->s_szc == 0) ? PAGESIZE :
4927 		    page_get_pagesize(seg->s_szc);
4928 		VM_STAT_COND_ADD(pgsz > PAGESIZE, segvnvmstats.fltanpages[16]);
4929 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
4930 		segvn_softunlock(seg, lpgaddr, lpgeaddr - lpgaddr, rw);
4931 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4932 		return (0);
4933 	}
4934 
4935 	ASSERT(svd->tr_state == SEGVN_TR_OFF ||
4936 	    !HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
4937 	if (brkcow == 0) {
4938 		if (svd->tr_state == SEGVN_TR_INIT) {
4939 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4940 			if (svd->tr_state == SEGVN_TR_INIT) {
4941 				ASSERT(svd->vp != NULL && svd->amp == NULL);
4942 				ASSERT(svd->flags & MAP_TEXT);
4943 				ASSERT(svd->type == MAP_PRIVATE);
4944 				segvn_textrepl(seg);
4945 				ASSERT(svd->tr_state != SEGVN_TR_INIT);
4946 				ASSERT(svd->tr_state != SEGVN_TR_ON ||
4947 				    svd->amp != NULL);
4948 			}
4949 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4950 		}
4951 	} else if (svd->tr_state != SEGVN_TR_OFF) {
4952 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4953 
4954 		if (rw == S_WRITE && svd->tr_state != SEGVN_TR_OFF) {
4955 			ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
4956 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4957 			return (FC_PROT);
4958 		}
4959 
4960 		if (svd->tr_state == SEGVN_TR_ON) {
4961 			ASSERT(svd->vp != NULL && svd->amp != NULL);
4962 			segvn_textunrepl(seg, 0);
4963 			ASSERT(svd->amp == NULL &&
4964 			    svd->tr_state == SEGVN_TR_OFF);
4965 		} else if (svd->tr_state != SEGVN_TR_OFF) {
4966 			svd->tr_state = SEGVN_TR_OFF;
4967 		}
4968 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
4969 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4970 	}
4971 
4972 top:
4973 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4974 
4975 	/*
4976 	 * If we have the same protections for the entire segment,
4977 	 * insure that the access being attempted is legitimate.
4978 	 */
4979 
4980 	if (svd->pageprot == 0) {
4981 		uint_t protchk;
4982 
4983 		switch (rw) {
4984 		case S_READ:
4985 		case S_READ_NOCOW:
4986 			protchk = PROT_READ;
4987 			break;
4988 		case S_WRITE:
4989 			protchk = PROT_WRITE;
4990 			break;
4991 		case S_EXEC:
4992 			protchk = PROT_EXEC;
4993 			break;
4994 		case S_OTHER:
4995 		default:
4996 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4997 			break;
4998 		}
4999 
5000 		if ((svd->prot & protchk) == 0) {
5001 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5002 			return (FC_PROT);	/* illegal access type */
5003 		}
5004 	}
5005 
5006 	if (brkcow && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5007 		/* this must be SOFTLOCK S_READ fault */
5008 		ASSERT(svd->amp == NULL);
5009 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5010 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5011 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5012 		if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5013 			/*
5014 			 * this must be the first ever non S_READ_NOCOW
5015 			 * softlock for this segment.
5016 			 */
5017 			ASSERT(svd->softlockcnt == 0);
5018 			hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5019 			    HAT_REGION_TEXT);
5020 			svd->rcookie = HAT_INVALID_REGION_COOKIE;
5021 		}
5022 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5023 		goto top;
5024 	}
5025 
5026 	/*
5027 	 * We can't allow the long term use of softlocks for vmpss segments,
5028 	 * because in some file truncation cases we should be able to demote
5029 	 * the segment, which requires that there are no softlocks.  The
5030 	 * only case where it's ok to allow a SOFTLOCK fault against a vmpss
5031 	 * segment is S_READ_NOCOW, where the caller holds the address space
5032 	 * locked as writer and calls softunlock before dropping the as lock.
5033 	 * S_READ_NOCOW is used by /proc to read memory from another user.
5034 	 *
5035 	 * Another deadlock between SOFTLOCK and file truncation can happen
5036 	 * because segvn_fault_vnodepages() calls the FS one pagesize at
5037 	 * a time. A second VOP_GETPAGE() call by segvn_fault_vnodepages()
5038 	 * can cause a deadlock because the first set of page_t's remain
5039 	 * locked SE_SHARED.  To avoid this, we demote segments on a first
5040 	 * SOFTLOCK if they have a length greater than the segment's
5041 	 * page size.
5042 	 *
5043 	 * So for now, we only avoid demoting a segment on a SOFTLOCK when
5044 	 * the access type is S_READ_NOCOW and the fault length is less than
5045 	 * or equal to the segment's page size. While this is quite restrictive,
5046 	 * it should be the most common case of SOFTLOCK against a vmpss
5047 	 * segment.
5048 	 *
5049 	 * For S_READ_NOCOW, it's safe not to do a copy on write because the
5050 	 * caller makes sure no COW will be caused by another thread for a
5051 	 * softlocked page.
5052 	 */
5053 	if (type == F_SOFTLOCK && svd->vp != NULL && seg->s_szc != 0) {
5054 		int demote = 0;
5055 
5056 		if (rw != S_READ_NOCOW) {
5057 			demote = 1;
5058 		}
5059 		if (!demote && len > PAGESIZE) {
5060 			pgsz = page_get_pagesize(seg->s_szc);
5061 			CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr,
5062 			    lpgeaddr);
5063 			if (lpgeaddr - lpgaddr > pgsz) {
5064 				demote = 1;
5065 			}
5066 		}
5067 
5068 		ASSERT(demote || AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5069 
5070 		if (demote) {
5071 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5072 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5073 			if (seg->s_szc != 0) {
5074 				segvn_vmpss_clrszc_cnt++;
5075 				ASSERT(svd->softlockcnt == 0);
5076 				err = segvn_clrszc(seg);
5077 				if (err) {
5078 					segvn_vmpss_clrszc_err++;
5079 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5080 					return (FC_MAKE_ERR(err));
5081 				}
5082 			}
5083 			ASSERT(seg->s_szc == 0);
5084 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5085 			goto top;
5086 		}
5087 	}
5088 
5089 	/*
5090 	 * Check to see if we need to allocate an anon_map structure.
5091 	 */
5092 	if (svd->amp == NULL && (svd->vp == NULL || brkcow)) {
5093 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5094 		/*
5095 		 * Drop the "read" lock on the segment and acquire
5096 		 * the "write" version since we have to allocate the
5097 		 * anon_map.
5098 		 */
5099 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5100 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5101 
5102 		if (svd->amp == NULL) {
5103 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
5104 			svd->amp->a_szc = seg->s_szc;
5105 		}
5106 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5107 
5108 		/*
5109 		 * Start all over again since segment protections
5110 		 * may have changed after we dropped the "read" lock.
5111 		 */
5112 		goto top;
5113 	}
5114 
5115 	/*
5116 	 * S_READ_NOCOW vs S_READ distinction was
5117 	 * only needed for the code above. After
5118 	 * that we treat it as S_READ.
5119 	 */
5120 	if (rw == S_READ_NOCOW) {
5121 		ASSERT(type == F_SOFTLOCK);
5122 		ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5123 		rw = S_READ;
5124 	}
5125 
5126 	amp = svd->amp;
5127 
5128 	/*
5129 	 * MADV_SEQUENTIAL work is ignored for large page segments.
5130 	 */
5131 	if (seg->s_szc != 0) {
5132 		pgsz = page_get_pagesize(seg->s_szc);
5133 		ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
5134 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
5135 		if (svd->vp == NULL) {
5136 			err = segvn_fault_anonpages(hat, seg, lpgaddr,
5137 			    lpgeaddr, type, rw, addr, addr + len, brkcow);
5138 		} else {
5139 			err = segvn_fault_vnodepages(hat, seg, lpgaddr,
5140 				lpgeaddr, type, rw, addr, addr + len, brkcow);
5141 			if (err == IE_RETRY) {
5142 				ASSERT(seg->s_szc == 0);
5143 				ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
5144 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5145 				goto top;
5146 			}
5147 		}
5148 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5149 		return (err);
5150 	}
5151 
5152 	page = seg_page(seg, addr);
5153 	if (amp != NULL) {
5154 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5155 		anon_index = svd->anon_index + page;
5156 
5157 		if (type == F_PROT && rw == S_READ &&
5158 		    svd->tr_state == SEGVN_TR_OFF &&
5159 		    svd->type == MAP_PRIVATE && svd->pageprot == 0) {
5160 			size_t index = anon_index;
5161 			struct anon *ap;
5162 
5163 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5164 			/*
5165 			 * The fast path could apply to S_WRITE also, except
5166 			 * that the protection fault could be caused by lazy
5167 			 * tlb flush when ro->rw. In this case, the pte is
5168 			 * RW already. But RO in the other cpu's tlb causes
5169 			 * the fault. Since hat_chgprot won't do anything if
5170 			 * pte doesn't change, we may end up faulting
5171 			 * indefinitely until the RO tlb entry gets replaced.
5172 			 */
5173 			for (a = addr; a < addr + len; a += PAGESIZE, index++) {
5174 				anon_array_enter(amp, index, &cookie);
5175 				ap = anon_get_ptr(amp->ahp, index);
5176 				anon_array_exit(&cookie);
5177 				if ((ap == NULL) || (ap->an_refcnt != 1)) {
5178 					ANON_LOCK_EXIT(&amp->a_rwlock);
5179 					goto slow;
5180 				}
5181 			}
5182 			hat_chgprot(seg->s_as->a_hat, addr, len, svd->prot);
5183 			ANON_LOCK_EXIT(&amp->a_rwlock);
5184 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5185 			return (0);
5186 		}
5187 	}
5188 slow:
5189 
5190 	if (svd->vpage == NULL)
5191 		vpage = NULL;
5192 	else
5193 		vpage = &svd->vpage[page];
5194 
5195 	off = svd->offset + (uintptr_t)(addr - seg->s_base);
5196 
5197 	/*
5198 	 * If MADV_SEQUENTIAL has been set for the particular page we
5199 	 * are faulting on, free behind all pages in the segment and put
5200 	 * them on the free list.
5201 	 */
5202 
5203 	if ((page != 0) && fltadvice && svd->tr_state != SEGVN_TR_ON) {
5204 		struct vpage *vpp;
5205 		ulong_t fanon_index;
5206 		size_t fpage;
5207 		u_offset_t pgoff, fpgoff;
5208 		struct vnode *fvp;
5209 		struct anon *fap = NULL;
5210 
5211 		if (svd->advice == MADV_SEQUENTIAL ||
5212 		    (svd->pageadvice &&
5213 		    VPP_ADVICE(vpage) == MADV_SEQUENTIAL)) {
5214 			pgoff = off - PAGESIZE;
5215 			fpage = page - 1;
5216 			if (vpage != NULL)
5217 				vpp = &svd->vpage[fpage];
5218 			if (amp != NULL)
5219 				fanon_index = svd->anon_index + fpage;
5220 
5221 			while (pgoff > svd->offset) {
5222 				if (svd->advice != MADV_SEQUENTIAL &&
5223 				    (!svd->pageadvice || (vpage &&
5224 				    VPP_ADVICE(vpp) != MADV_SEQUENTIAL)))
5225 					break;
5226 
5227 				/*
5228 				 * If this is an anon page, we must find the
5229 				 * correct <vp, offset> for it
5230 				 */
5231 				fap = NULL;
5232 				if (amp != NULL) {
5233 					ANON_LOCK_ENTER(&amp->a_rwlock,
5234 						RW_READER);
5235 					anon_array_enter(amp, fanon_index,
5236 						&cookie);
5237 					fap = anon_get_ptr(amp->ahp,
5238 					    fanon_index);
5239 					if (fap != NULL) {
5240 						swap_xlate(fap, &fvp, &fpgoff);
5241 					} else {
5242 						fpgoff = pgoff;
5243 						fvp = svd->vp;
5244 					}
5245 					anon_array_exit(&cookie);
5246 					ANON_LOCK_EXIT(&amp->a_rwlock);
5247 				} else {
5248 					fpgoff = pgoff;
5249 					fvp = svd->vp;
5250 				}
5251 				if (fvp == NULL)
5252 					break;	/* XXX */
5253 				/*
5254 				 * Skip pages that are free or have an
5255 				 * "exclusive" lock.
5256 				 */
5257 				pp = page_lookup_nowait(fvp, fpgoff, SE_SHARED);
5258 				if (pp == NULL)
5259 					break;
5260 				/*
5261 				 * We don't need the page_struct_lock to test
5262 				 * as this is only advisory; even if we
5263 				 * acquire it someone might race in and lock
5264 				 * the page after we unlock and before the
5265 				 * PUTPAGE, then VOP_PUTPAGE will do nothing.
5266 				 */
5267 				if (pp->p_lckcnt == 0 && pp->p_cowcnt == 0) {
5268 					/*
5269 					 * Hold the vnode before releasing
5270 					 * the page lock to prevent it from
5271 					 * being freed and re-used by some
5272 					 * other thread.
5273 					 */
5274 					VN_HOLD(fvp);
5275 					page_unlock(pp);
5276 					/*
5277 					 * We should build a page list
5278 					 * to kluster putpages XXX
5279 					 */
5280 					(void) VOP_PUTPAGE(fvp,
5281 					    (offset_t)fpgoff, PAGESIZE,
5282 					    (B_DONTNEED|B_FREE|B_ASYNC),
5283 					    svd->cred);
5284 					VN_RELE(fvp);
5285 				} else {
5286 					/*
5287 					 * XXX - Should the loop terminate if
5288 					 * the page is `locked'?
5289 					 */
5290 					page_unlock(pp);
5291 				}
5292 				--vpp;
5293 				--fanon_index;
5294 				pgoff -= PAGESIZE;
5295 			}
5296 		}
5297 	}
5298 
5299 	plp = pl;
5300 	*plp = NULL;
5301 	pl_alloc_sz = 0;
5302 
5303 	/*
5304 	 * See if we need to call VOP_GETPAGE for
5305 	 * *any* of the range being faulted on.
5306 	 * We can skip all of this work if there
5307 	 * was no original vnode.
5308 	 */
5309 	if (svd->vp != NULL) {
5310 		u_offset_t vp_off;
5311 		size_t vp_len;
5312 		struct anon *ap;
5313 		vnode_t *vp;
5314 
5315 		vp_off = off;
5316 		vp_len = len;
5317 
5318 		if (amp == NULL)
5319 			dogetpage = 1;
5320 		else {
5321 			/*
5322 			 * Only acquire reader lock to prevent amp->ahp
5323 			 * from being changed.  It's ok to miss pages,
5324 			 * hence we don't do anon_array_enter
5325 			 */
5326 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5327 			ap = anon_get_ptr(amp->ahp, anon_index);
5328 
5329 			if (len <= PAGESIZE)
5330 				/* inline non_anon() */
5331 				dogetpage = (ap == NULL);
5332 			else
5333 				dogetpage = non_anon(amp->ahp, anon_index,
5334 				    &vp_off, &vp_len);
5335 			ANON_LOCK_EXIT(&amp->a_rwlock);
5336 		}
5337 
5338 		if (dogetpage) {
5339 			enum seg_rw arw;
5340 			struct as *as = seg->s_as;
5341 
5342 			if (len > ptob((sizeof (pl) / sizeof (pl[0])) - 1)) {
5343 				/*
5344 				 * Page list won't fit in local array,
5345 				 * allocate one of the needed size.
5346 				 */
5347 				pl_alloc_sz =
5348 				    (btop(len) + 1) * sizeof (page_t *);
5349 				plp = kmem_alloc(pl_alloc_sz, KM_SLEEP);
5350 				plp[0] = NULL;
5351 				plsz = len;
5352 			} else if (rw == S_WRITE && svd->type == MAP_PRIVATE ||
5353 			    svd->tr_state == SEGVN_TR_ON || rw == S_OTHER ||
5354 			    (((size_t)(addr + PAGESIZE) <
5355 			    (size_t)(seg->s_base + seg->s_size)) &&
5356 			    hat_probe(as->a_hat, addr + PAGESIZE))) {
5357 				/*
5358 				 * Ask VOP_GETPAGE to return the exact number
5359 				 * of pages if
5360 				 * (a) this is a COW fault, or
5361 				 * (b) this is a software fault, or
5362 				 * (c) next page is already mapped.
5363 				 */
5364 				plsz = len;
5365 			} else {
5366 				/*
5367 				 * Ask VOP_GETPAGE to return adjacent pages
5368 				 * within the segment.
5369 				 */
5370 				plsz = MIN((size_t)PVN_GETPAGE_SZ, (size_t)
5371 					((seg->s_base + seg->s_size) - addr));
5372 				ASSERT((addr + plsz) <=
5373 				    (seg->s_base + seg->s_size));
5374 			}
5375 
5376 			/*
5377 			 * Need to get some non-anonymous pages.
5378 			 * We need to make only one call to GETPAGE to do
5379 			 * this to prevent certain deadlocking conditions
5380 			 * when we are doing locking.  In this case
5381 			 * non_anon() should have picked up the smallest
5382 			 * range which includes all the non-anonymous
5383 			 * pages in the requested range.  We have to
5384 			 * be careful regarding which rw flag to pass in
5385 			 * because on a private mapping, the underlying
5386 			 * object is never allowed to be written.
5387 			 */
5388 			if (rw == S_WRITE && svd->type == MAP_PRIVATE) {
5389 				arw = S_READ;
5390 			} else {
5391 				arw = rw;
5392 			}
5393 			vp = svd->vp;
5394 			TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5395 				"segvn_getpage:seg %p addr %p vp %p",
5396 				seg, addr, vp);
5397 			err = VOP_GETPAGE(vp, (offset_t)vp_off, vp_len,
5398 			    &vpprot, plp, plsz, seg, addr + (vp_off - off), arw,
5399 			    svd->cred);
5400 			if (err) {
5401 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5402 				segvn_pagelist_rele(plp);
5403 				if (pl_alloc_sz)
5404 					kmem_free(plp, pl_alloc_sz);
5405 				return (FC_MAKE_ERR(err));
5406 			}
5407 			if (svd->type == MAP_PRIVATE)
5408 				vpprot &= ~PROT_WRITE;
5409 		}
5410 	}
5411 
5412 	/*
5413 	 * N.B. at this time the plp array has all the needed non-anon
5414 	 * pages in addition to (possibly) having some adjacent pages.
5415 	 */
5416 
5417 	/*
5418 	 * Always acquire the anon_array_lock to prevent
5419 	 * 2 threads from allocating separate anon slots for
5420 	 * the same "addr".
5421 	 *
5422 	 * If this is a copy-on-write fault and we don't already
5423 	 * have the anon_array_lock, acquire it to prevent the
5424 	 * fault routine from handling multiple copy-on-write faults
5425 	 * on the same "addr" in the same address space.
5426 	 *
5427 	 * Only one thread should deal with the fault since after
5428 	 * it is handled, the other threads can acquire a translation
5429 	 * to the newly created private page.  This prevents two or
5430 	 * more threads from creating different private pages for the
5431 	 * same fault.
5432 	 *
5433 	 * We grab "serialization" lock here if this is a MAP_PRIVATE segment
5434 	 * to prevent deadlock between this thread and another thread
5435 	 * which has soft-locked this page and wants to acquire serial_lock.
5436 	 * ( bug 4026339 )
5437 	 *
5438 	 * The fix for bug 4026339 becomes unnecessary when using the
5439 	 * locking scheme with per amp rwlock and a global set of hash
5440 	 * lock, anon_array_lock.  If we steal a vnode page when low
5441 	 * on memory and upgrad the page lock through page_rename,
5442 	 * then the page is PAGE_HANDLED, nothing needs to be done
5443 	 * for this page after returning from segvn_faultpage.
5444 	 *
5445 	 * But really, the page lock should be downgraded after
5446 	 * the stolen page is page_rename'd.
5447 	 */
5448 
5449 	if (amp != NULL)
5450 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5451 
5452 	/*
5453 	 * Ok, now loop over the address range and handle faults
5454 	 */
5455 	for (a = addr; a < addr + len; a += PAGESIZE, off += PAGESIZE) {
5456 		err = segvn_faultpage(hat, seg, a, off, vpage, plp, vpprot,
5457 		    type, rw, brkcow, a == addr);
5458 		if (err) {
5459 			if (amp != NULL)
5460 				ANON_LOCK_EXIT(&amp->a_rwlock);
5461 			if (type == F_SOFTLOCK && a > addr) {
5462 				segvn_softunlock(seg, addr, (a - addr),
5463 				    S_OTHER);
5464 			}
5465 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5466 			segvn_pagelist_rele(plp);
5467 			if (pl_alloc_sz)
5468 				kmem_free(plp, pl_alloc_sz);
5469 			return (err);
5470 		}
5471 		if (vpage) {
5472 			vpage++;
5473 		} else if (svd->vpage) {
5474 			page = seg_page(seg, addr);
5475 			vpage = &svd->vpage[++page];
5476 		}
5477 	}
5478 
5479 	/* Didn't get pages from the underlying fs so we're done */
5480 	if (!dogetpage)
5481 		goto done;
5482 
5483 	/*
5484 	 * Now handle any other pages in the list returned.
5485 	 * If the page can be used, load up the translations now.
5486 	 * Note that the for loop will only be entered if "plp"
5487 	 * is pointing to a non-NULL page pointer which means that
5488 	 * VOP_GETPAGE() was called and vpprot has been initialized.
5489 	 */
5490 	if (svd->pageprot == 0)
5491 		prot = svd->prot & vpprot;
5492 
5493 
5494 	/*
5495 	 * Large Files: diff should be unsigned value because we started
5496 	 * supporting > 2GB segment sizes from 2.5.1 and when a
5497 	 * large file of size > 2GB gets mapped to address space
5498 	 * the diff value can be > 2GB.
5499 	 */
5500 
5501 	for (ppp = plp; (pp = *ppp) != NULL; ppp++) {
5502 		size_t diff;
5503 		struct anon *ap;
5504 		int anon_index;
5505 		anon_sync_obj_t cookie;
5506 		int hat_flag = HAT_LOAD_ADV;
5507 
5508 		if (svd->flags & MAP_TEXT) {
5509 			hat_flag |= HAT_LOAD_TEXT;
5510 		}
5511 
5512 		if (pp == PAGE_HANDLED)
5513 			continue;
5514 
5515 		if (svd->tr_state != SEGVN_TR_ON &&
5516 		    pp->p_offset >=  svd->offset &&
5517 		    pp->p_offset < svd->offset + seg->s_size) {
5518 
5519 			diff = pp->p_offset - svd->offset;
5520 
5521 			/*
5522 			 * Large Files: Following is the assertion
5523 			 * validating the above cast.
5524 			 */
5525 			ASSERT(svd->vp == pp->p_vnode);
5526 
5527 			page = btop(diff);
5528 			if (svd->pageprot)
5529 				prot = VPP_PROT(&svd->vpage[page]) & vpprot;
5530 
5531 			/*
5532 			 * Prevent other threads in the address space from
5533 			 * creating private pages (i.e., allocating anon slots)
5534 			 * while we are in the process of loading translations
5535 			 * to additional pages returned by the underlying
5536 			 * object.
5537 			 */
5538 			if (amp != NULL) {
5539 				anon_index = svd->anon_index + page;
5540 				anon_array_enter(amp, anon_index, &cookie);
5541 				ap = anon_get_ptr(amp->ahp, anon_index);
5542 			}
5543 			if ((amp == NULL) || (ap == NULL)) {
5544 				if (IS_VMODSORT(pp->p_vnode) ||
5545 				    enable_mbit_wa) {
5546 					if (rw == S_WRITE)
5547 						hat_setmod(pp);
5548 					else if (rw != S_OTHER &&
5549 					    !hat_ismod(pp))
5550 						prot &= ~PROT_WRITE;
5551 				}
5552 				/*
5553 				 * Skip mapping read ahead pages marked
5554 				 * for migration, so they will get migrated
5555 				 * properly on fault
5556 				 */
5557 				ASSERT(amp == NULL ||
5558 				    svd->rcookie == HAT_INVALID_REGION_COOKIE);
5559 				if ((prot & PROT_READ) && !PP_ISMIGRATE(pp)) {
5560 					hat_memload_region(hat,
5561 					    seg->s_base + diff,
5562 					    pp, prot, hat_flag,
5563 					    svd->rcookie);
5564 				}
5565 			}
5566 			if (amp != NULL)
5567 				anon_array_exit(&cookie);
5568 		}
5569 		page_unlock(pp);
5570 	}
5571 done:
5572 	if (amp != NULL)
5573 		ANON_LOCK_EXIT(&amp->a_rwlock);
5574 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5575 	if (pl_alloc_sz)
5576 		kmem_free(plp, pl_alloc_sz);
5577 	return (0);
5578 }
5579 
5580 /*
5581  * This routine is used to start I/O on pages asynchronously.  XXX it will
5582  * only create PAGESIZE pages. At fault time they will be relocated into
5583  * larger pages.
5584  */
5585 static faultcode_t
5586 segvn_faulta(struct seg *seg, caddr_t addr)
5587 {
5588 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5589 	int err;
5590 	struct anon_map *amp;
5591 	vnode_t *vp;
5592 
5593 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5594 
5595 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5596 	if ((amp = svd->amp) != NULL) {
5597 		struct anon *ap;
5598 
5599 		/*
5600 		 * Reader lock to prevent amp->ahp from being changed.
5601 		 * This is advisory, it's ok to miss a page, so
5602 		 * we don't do anon_array_enter lock.
5603 		 */
5604 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5605 		if ((ap = anon_get_ptr(amp->ahp,
5606 			svd->anon_index + seg_page(seg, addr))) != NULL) {
5607 
5608 			err = anon_getpage(&ap, NULL, NULL,
5609 			    0, seg, addr, S_READ, svd->cred);
5610 
5611 			ANON_LOCK_EXIT(&amp->a_rwlock);
5612 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5613 			if (err)
5614 				return (FC_MAKE_ERR(err));
5615 			return (0);
5616 		}
5617 		ANON_LOCK_EXIT(&amp->a_rwlock);
5618 	}
5619 
5620 	if (svd->vp == NULL) {
5621 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5622 		return (0);			/* zfod page - do nothing now */
5623 	}
5624 
5625 	vp = svd->vp;
5626 	TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5627 		"segvn_getpage:seg %p addr %p vp %p", seg, addr, vp);
5628 	err = VOP_GETPAGE(vp,
5629 	    (offset_t)(svd->offset + (uintptr_t)(addr - seg->s_base)),
5630 	    PAGESIZE, NULL, NULL, 0, seg, addr,
5631 	    S_OTHER, svd->cred);
5632 
5633 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5634 	if (err)
5635 		return (FC_MAKE_ERR(err));
5636 	return (0);
5637 }
5638 
5639 static int
5640 segvn_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
5641 {
5642 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5643 	struct vpage *svp, *evp;
5644 	struct vnode *vp;
5645 	size_t pgsz;
5646 	pgcnt_t pgcnt;
5647 	anon_sync_obj_t cookie;
5648 	int unload_done = 0;
5649 
5650 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5651 
5652 	if ((svd->maxprot & prot) != prot)
5653 		return (EACCES);			/* violated maxprot */
5654 
5655 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5656 
5657 	/* return if prot is the same */
5658 	if (!svd->pageprot && svd->prot == prot) {
5659 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5660 		return (0);
5661 	}
5662 
5663 	/*
5664 	 * Since we change protections we first have to flush the cache.
5665 	 * This makes sure all the pagelock calls have to recheck
5666 	 * protections.
5667 	 */
5668 	if (svd->softlockcnt > 0) {
5669 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5670 		/*
5671 		 * Since we do have the segvn writers lock nobody can fill
5672 		 * the cache with entries belonging to this seg during
5673 		 * the purge. The flush either succeeds or we still have
5674 		 * pending I/Os.
5675 		 */
5676 		segvn_purge(seg);
5677 		if (svd->softlockcnt > 0) {
5678 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5679 			return (EAGAIN);
5680 		}
5681 	}
5682 
5683 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5684 		ASSERT(svd->amp == NULL);
5685 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5686 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5687 		    HAT_REGION_TEXT);
5688 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
5689 		unload_done = 1;
5690 	} else if (svd->tr_state == SEGVN_TR_INIT) {
5691 		svd->tr_state = SEGVN_TR_OFF;
5692 	} else if (svd->tr_state == SEGVN_TR_ON) {
5693 		ASSERT(svd->amp != NULL);
5694 		segvn_textunrepl(seg, 0);
5695 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
5696 		unload_done = 1;
5697 	}
5698 
5699 	if ((prot & PROT_WRITE) && svd->type == MAP_SHARED &&
5700 	    svd->vp != NULL && (svd->vp->v_flag & VVMEXEC)) {
5701 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
5702 		segvn_inval_trcache(svd->vp);
5703 	}
5704 	if (seg->s_szc != 0) {
5705 		int err;
5706 		pgsz = page_get_pagesize(seg->s_szc);
5707 		pgcnt = pgsz >> PAGESHIFT;
5708 		ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
5709 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
5710 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5711 			ASSERT(seg->s_base != addr || seg->s_size != len);
5712 			/*
5713 			 * If we are holding the as lock as a reader then
5714 			 * we need to return IE_RETRY and let the as
5715 			 * layer drop and re-aquire the lock as a writer.
5716 			 */
5717 			if (AS_READ_HELD(seg->s_as, &seg->s_as->a_lock))
5718 				return (IE_RETRY);
5719 			VM_STAT_ADD(segvnvmstats.demoterange[1]);
5720 			if (svd->type == MAP_PRIVATE || svd->vp != NULL) {
5721 				err = segvn_demote_range(seg, addr, len,
5722 				    SDR_END, 0);
5723 			} else {
5724 				uint_t szcvec = map_pgszcvec(seg->s_base,
5725 				    pgsz, (uintptr_t)seg->s_base,
5726 				    (svd->flags & MAP_TEXT), MAPPGSZC_SHM, 0);
5727 				err = segvn_demote_range(seg, addr, len,
5728 				    SDR_END, szcvec);
5729 			}
5730 			if (err == 0)
5731 				return (IE_RETRY);
5732 			if (err == ENOMEM)
5733 				return (IE_NOMEM);
5734 			return (err);
5735 		}
5736 	}
5737 
5738 
5739 	/*
5740 	 * If it's a private mapping and we're making it writable
5741 	 * and no swap space has been reserved, have to reserve
5742 	 * it all now.  If it's a private mapping to a file (i.e., vp != NULL)
5743 	 * and we're removing write permission on the entire segment and
5744 	 * we haven't modified any pages, we can release the swap space.
5745 	 */
5746 	if (svd->type == MAP_PRIVATE) {
5747 		if (prot & PROT_WRITE) {
5748 			size_t sz;
5749 			if (svd->swresv == 0 && !(svd->flags & MAP_NORESERVE)) {
5750 				if (anon_resv_zone(seg->s_size,
5751 				    seg->s_as->a_proc->p_zone) == 0) {
5752 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5753 					return (IE_NOMEM);
5754 				}
5755 				sz = svd->swresv = seg->s_size;
5756 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
5757 					"anon proc:%p %lu %u",
5758 					seg, sz, 1);
5759 			}
5760 		} else {
5761 			/*
5762 			 * Swap space is released only if this segment
5763 			 * does not map anonymous memory, since read faults
5764 			 * on such segments still need an anon slot to read
5765 			 * in the data.
5766 			 */
5767 			if (svd->swresv != 0 && svd->vp != NULL &&
5768 			    svd->amp == NULL && addr == seg->s_base &&
5769 			    len == seg->s_size && svd->pageprot == 0) {
5770 				anon_unresv_zone(svd->swresv,
5771 				    seg->s_as->a_proc->p_zone);
5772 				svd->swresv = 0;
5773 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
5774 					"anon proc:%p %lu %u",
5775 					seg, 0, 0);
5776 			}
5777 		}
5778 	}
5779 
5780 	if (addr == seg->s_base && len == seg->s_size && svd->vpage == NULL) {
5781 		if (svd->prot == prot) {
5782 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5783 			return (0);			/* all done */
5784 		}
5785 		svd->prot = (uchar_t)prot;
5786 	} else if (svd->type == MAP_PRIVATE) {
5787 		struct anon *ap = NULL;
5788 		page_t *pp;
5789 		u_offset_t offset, off;
5790 		struct anon_map *amp;
5791 		ulong_t anon_idx = 0;
5792 
5793 		/*
5794 		 * A vpage structure exists or else the change does not
5795 		 * involve the entire segment.  Establish a vpage structure
5796 		 * if none is there.  Then, for each page in the range,
5797 		 * adjust its individual permissions.  Note that write-
5798 		 * enabling a MAP_PRIVATE page can affect the claims for
5799 		 * locked down memory.  Overcommitting memory terminates
5800 		 * the operation.
5801 		 */
5802 		segvn_vpage(seg);
5803 		svd->pageprot = 1;
5804 		if ((amp = svd->amp) != NULL) {
5805 			anon_idx = svd->anon_index + seg_page(seg, addr);
5806 			ASSERT(seg->s_szc == 0 ||
5807 			    IS_P2ALIGNED(anon_idx, pgcnt));
5808 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5809 		}
5810 
5811 		offset = svd->offset + (uintptr_t)(addr - seg->s_base);
5812 		evp = &svd->vpage[seg_page(seg, addr + len)];
5813 
5814 		/*
5815 		 * See Statement at the beginning of segvn_lockop regarding
5816 		 * the way cowcnts and lckcnts are handled.
5817 		 */
5818 		for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5819 
5820 			if (seg->s_szc != 0) {
5821 				if (amp != NULL) {
5822 					anon_array_enter(amp, anon_idx,
5823 					    &cookie);
5824 				}
5825 				if (IS_P2ALIGNED(anon_idx, pgcnt) &&
5826 				    !segvn_claim_pages(seg, svp, offset,
5827 					anon_idx, prot)) {
5828 					if (amp != NULL) {
5829 						anon_array_exit(&cookie);
5830 					}
5831 					break;
5832 				}
5833 				if (amp != NULL) {
5834 					anon_array_exit(&cookie);
5835 				}
5836 				anon_idx++;
5837 			} else {
5838 				if (amp != NULL) {
5839 					anon_array_enter(amp, anon_idx,
5840 						&cookie);
5841 					ap = anon_get_ptr(amp->ahp, anon_idx++);
5842 				}
5843 
5844 				if (VPP_ISPPLOCK(svp) &&
5845 				    VPP_PROT(svp) != prot) {
5846 
5847 					if (amp == NULL || ap == NULL) {
5848 						vp = svd->vp;
5849 						off = offset;
5850 					} else
5851 						swap_xlate(ap, &vp, &off);
5852 					if (amp != NULL)
5853 						anon_array_exit(&cookie);
5854 
5855 					if ((pp = page_lookup(vp, off,
5856 					    SE_SHARED)) == NULL) {
5857 						panic("segvn_setprot: no page");
5858 						/*NOTREACHED*/
5859 					}
5860 					ASSERT(seg->s_szc == 0);
5861 					if ((VPP_PROT(svp) ^ prot) &
5862 					    PROT_WRITE) {
5863 						if (prot & PROT_WRITE) {
5864 						    if (!page_addclaim(pp)) {
5865 							page_unlock(pp);
5866 							break;
5867 						    }
5868 						} else {
5869 						    if (!page_subclaim(pp)) {
5870 							page_unlock(pp);
5871 							break;
5872 						    }
5873 						}
5874 					}
5875 					page_unlock(pp);
5876 				} else if (amp != NULL)
5877 					anon_array_exit(&cookie);
5878 			}
5879 			VPP_SETPROT(svp, prot);
5880 			offset += PAGESIZE;
5881 		}
5882 		if (amp != NULL)
5883 			ANON_LOCK_EXIT(&amp->a_rwlock);
5884 
5885 		/*
5886 		 * Did we terminate prematurely?  If so, simply unload
5887 		 * the translations to the things we've updated so far.
5888 		 */
5889 		if (svp != evp) {
5890 			if (unload_done) {
5891 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5892 				return (IE_NOMEM);
5893 			}
5894 			len = (svp - &svd->vpage[seg_page(seg, addr)]) *
5895 			    PAGESIZE;
5896 			ASSERT(seg->s_szc == 0 || IS_P2ALIGNED(len, pgsz));
5897 			if (len != 0)
5898 				hat_unload(seg->s_as->a_hat, addr,
5899 				    len, HAT_UNLOAD);
5900 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5901 			return (IE_NOMEM);
5902 		}
5903 	} else {
5904 		segvn_vpage(seg);
5905 		svd->pageprot = 1;
5906 		evp = &svd->vpage[seg_page(seg, addr + len)];
5907 		for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5908 			VPP_SETPROT(svp, prot);
5909 		}
5910 	}
5911 
5912 	if (unload_done) {
5913 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5914 		return (0);
5915 	}
5916 
5917 	if (((prot & PROT_WRITE) != 0 &&
5918 	    (svd->vp != NULL || svd->type == MAP_PRIVATE)) ||
5919 	    (prot & ~PROT_USER) == PROT_NONE) {
5920 		/*
5921 		 * Either private or shared data with write access (in
5922 		 * which case we need to throw out all former translations
5923 		 * so that we get the right translations set up on fault
5924 		 * and we don't allow write access to any copy-on-write pages
5925 		 * that might be around or to prevent write access to pages
5926 		 * representing holes in a file), or we don't have permission
5927 		 * to access the memory at all (in which case we have to
5928 		 * unload any current translations that might exist).
5929 		 */
5930 		hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
5931 	} else {
5932 		/*
5933 		 * A shared mapping or a private mapping in which write
5934 		 * protection is going to be denied - just change all the
5935 		 * protections over the range of addresses in question.
5936 		 * segvn does not support any other attributes other
5937 		 * than prot so we can use hat_chgattr.
5938 		 */
5939 		hat_chgattr(seg->s_as->a_hat, addr, len, prot);
5940 	}
5941 
5942 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5943 
5944 	return (0);
5945 }
5946 
5947 /*
5948  * segvn_setpagesize is called via SEGOP_SETPAGESIZE from as_setpagesize,
5949  * to determine if the seg is capable of mapping the requested szc.
5950  */
5951 static int
5952 segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len, uint_t szc)
5953 {
5954 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5955 	struct segvn_data *nsvd;
5956 	struct anon_map *amp = svd->amp;
5957 	struct seg *nseg;
5958 	caddr_t eaddr = addr + len, a;
5959 	size_t pgsz = page_get_pagesize(szc);
5960 	pgcnt_t pgcnt = page_get_pagecnt(szc);
5961 	int err;
5962 	u_offset_t off = svd->offset + (uintptr_t)(addr - seg->s_base);
5963 	extern struct vnode kvp;
5964 
5965 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5966 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
5967 
5968 	if (seg->s_szc == szc || segvn_lpg_disable != 0) {
5969 		return (0);
5970 	}
5971 
5972 	/*
5973 	 * addr should always be pgsz aligned but eaddr may be misaligned if
5974 	 * it's at the end of the segment.
5975 	 *
5976 	 * XXX we should assert this condition since as_setpagesize() logic
5977 	 * guarantees it.
5978 	 */
5979 	if (!IS_P2ALIGNED(addr, pgsz) ||
5980 	    (!IS_P2ALIGNED(eaddr, pgsz) &&
5981 		eaddr != seg->s_base + seg->s_size)) {
5982 
5983 		segvn_setpgsz_align_err++;
5984 		return (EINVAL);
5985 	}
5986 
5987 	if (amp != NULL && svd->type == MAP_SHARED) {
5988 		ulong_t an_idx = svd->anon_index + seg_page(seg, addr);
5989 		if (!IS_P2ALIGNED(an_idx, pgcnt)) {
5990 
5991 			segvn_setpgsz_anon_align_err++;
5992 			return (EINVAL);
5993 		}
5994 	}
5995 
5996 	if ((svd->flags & MAP_NORESERVE) || seg->s_as == &kas ||
5997 	    szc > segvn_maxpgszc) {
5998 		return (EINVAL);
5999 	}
6000 
6001 	/* paranoid check */
6002 	if (svd->vp != NULL &&
6003 	    (IS_SWAPFSVP(svd->vp) || VN_ISKAS(svd->vp))) {
6004 		    return (EINVAL);
6005 	}
6006 
6007 	if (seg->s_szc == 0 && svd->vp != NULL &&
6008 	    map_addr_vacalign_check(addr, off)) {
6009 		return (EINVAL);
6010 	}
6011 
6012 	/*
6013 	 * Check that protections are the same within new page
6014 	 * size boundaries.
6015 	 */
6016 	if (svd->pageprot) {
6017 		for (a = addr; a < eaddr; a += pgsz) {
6018 			if ((a + pgsz) > eaddr) {
6019 				if (!sameprot(seg, a, eaddr - a)) {
6020 					return (EINVAL);
6021 				}
6022 			} else {
6023 				if (!sameprot(seg, a, pgsz)) {
6024 					return (EINVAL);
6025 				}
6026 			}
6027 		}
6028 	}
6029 
6030 	/*
6031 	 * Since we are changing page size we first have to flush
6032 	 * the cache. This makes sure all the pagelock calls have
6033 	 * to recheck protections.
6034 	 */
6035 	if (svd->softlockcnt > 0) {
6036 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6037 		/*
6038 		 * Since we do have the segvn writers lock nobody can fill
6039 		 * the cache with entries belonging to this seg during
6040 		 * the purge. The flush either succeeds or we still have
6041 		 * pending I/Os.
6042 		 */
6043 		segvn_purge(seg);
6044 		if (svd->softlockcnt > 0) {
6045 			return (EAGAIN);
6046 		}
6047 	}
6048 
6049 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6050 		ASSERT(svd->amp == NULL);
6051 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6052 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6053 		    HAT_REGION_TEXT);
6054 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6055 	} else if (svd->tr_state == SEGVN_TR_INIT) {
6056 		svd->tr_state = SEGVN_TR_OFF;
6057 	} else if (svd->tr_state == SEGVN_TR_ON) {
6058 		ASSERT(svd->amp != NULL);
6059 		segvn_textunrepl(seg, 1);
6060 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6061 		amp = NULL;
6062 	}
6063 
6064 	/*
6065 	 * Operation for sub range of existing segment.
6066 	 */
6067 	if (addr != seg->s_base || eaddr != (seg->s_base + seg->s_size)) {
6068 		if (szc < seg->s_szc) {
6069 			VM_STAT_ADD(segvnvmstats.demoterange[2]);
6070 			err = segvn_demote_range(seg, addr, len, SDR_RANGE, 0);
6071 			if (err == 0) {
6072 				return (IE_RETRY);
6073 			}
6074 			if (err == ENOMEM) {
6075 				return (IE_NOMEM);
6076 			}
6077 			return (err);
6078 		}
6079 		if (addr != seg->s_base) {
6080 			nseg = segvn_split_seg(seg, addr);
6081 			if (eaddr != (nseg->s_base + nseg->s_size)) {
6082 				/* eaddr is szc aligned */
6083 				(void) segvn_split_seg(nseg, eaddr);
6084 			}
6085 			return (IE_RETRY);
6086 		}
6087 		if (eaddr != (seg->s_base + seg->s_size)) {
6088 			/* eaddr is szc aligned */
6089 			(void) segvn_split_seg(seg, eaddr);
6090 		}
6091 		return (IE_RETRY);
6092 	}
6093 
6094 	/*
6095 	 * Break any low level sharing and reset seg->s_szc to 0.
6096 	 */
6097 	if ((err = segvn_clrszc(seg)) != 0) {
6098 		if (err == ENOMEM) {
6099 			err = IE_NOMEM;
6100 		}
6101 		return (err);
6102 	}
6103 	ASSERT(seg->s_szc == 0);
6104 
6105 	/*
6106 	 * If the end of the current segment is not pgsz aligned
6107 	 * then attempt to concatenate with the next segment.
6108 	 */
6109 	if (!IS_P2ALIGNED(eaddr, pgsz)) {
6110 		nseg = AS_SEGNEXT(seg->s_as, seg);
6111 		if (nseg == NULL || nseg == seg || eaddr != nseg->s_base) {
6112 			return (ENOMEM);
6113 		}
6114 		if (nseg->s_ops != &segvn_ops) {
6115 			return (EINVAL);
6116 		}
6117 		nsvd = (struct segvn_data *)nseg->s_data;
6118 		if (nsvd->softlockcnt > 0) {
6119 			segvn_purge(nseg);
6120 			if (nsvd->softlockcnt > 0) {
6121 				return (EAGAIN);
6122 			}
6123 		}
6124 		err = segvn_clrszc(nseg);
6125 		if (err == ENOMEM) {
6126 			err = IE_NOMEM;
6127 		}
6128 		if (err != 0) {
6129 			return (err);
6130 		}
6131 		ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6132 		err = segvn_concat(seg, nseg, 1);
6133 		if (err == -1) {
6134 			return (EINVAL);
6135 		}
6136 		if (err == -2) {
6137 			return (IE_NOMEM);
6138 		}
6139 		return (IE_RETRY);
6140 	}
6141 
6142 	/*
6143 	 * May need to re-align anon array to
6144 	 * new szc.
6145 	 */
6146 	if (amp != NULL) {
6147 		if (!IS_P2ALIGNED(svd->anon_index, pgcnt)) {
6148 			struct anon_hdr *nahp;
6149 
6150 			ASSERT(svd->type == MAP_PRIVATE);
6151 
6152 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6153 			ASSERT(amp->refcnt == 1);
6154 			nahp = anon_create(btop(amp->size), ANON_NOSLEEP);
6155 			if (nahp == NULL) {
6156 				ANON_LOCK_EXIT(&amp->a_rwlock);
6157 				return (IE_NOMEM);
6158 			}
6159 			if (anon_copy_ptr(amp->ahp, svd->anon_index,
6160 				nahp, 0, btop(seg->s_size), ANON_NOSLEEP)) {
6161 				anon_release(nahp, btop(amp->size));
6162 				ANON_LOCK_EXIT(&amp->a_rwlock);
6163 				return (IE_NOMEM);
6164 			}
6165 			anon_release(amp->ahp, btop(amp->size));
6166 			amp->ahp = nahp;
6167 			svd->anon_index = 0;
6168 			ANON_LOCK_EXIT(&amp->a_rwlock);
6169 		}
6170 	}
6171 	if (svd->vp != NULL && szc != 0) {
6172 		struct vattr va;
6173 		u_offset_t eoffpage = svd->offset;
6174 		va.va_mask = AT_SIZE;
6175 		eoffpage += seg->s_size;
6176 		eoffpage = btopr(eoffpage);
6177 		if (VOP_GETATTR(svd->vp, &va, 0, svd->cred) != 0) {
6178 			segvn_setpgsz_getattr_err++;
6179 			return (EINVAL);
6180 		}
6181 		if (btopr(va.va_size) < eoffpage) {
6182 			segvn_setpgsz_eof_err++;
6183 			return (EINVAL);
6184 		}
6185 		if (amp != NULL) {
6186 			/*
6187 			 * anon_fill_cow_holes() may call VOP_GETPAGE().
6188 			 * don't take anon map lock here to avoid holding it
6189 			 * across VOP_GETPAGE() calls that may call back into
6190 			 * segvn for klsutering checks. We don't really need
6191 			 * anon map lock here since it's a private segment and
6192 			 * we hold as level lock as writers.
6193 			 */
6194 			if ((err = anon_fill_cow_holes(seg, seg->s_base,
6195 			    amp->ahp, svd->anon_index, svd->vp, svd->offset,
6196 			    seg->s_size, szc, svd->prot, svd->vpage,
6197 			    svd->cred)) != 0) {
6198 				return (EINVAL);
6199 			}
6200 		}
6201 		segvn_setvnode_mpss(svd->vp);
6202 	}
6203 
6204 	if (amp != NULL) {
6205 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6206 		if (svd->type == MAP_PRIVATE) {
6207 			amp->a_szc = szc;
6208 		} else if (szc > amp->a_szc) {
6209 			amp->a_szc = szc;
6210 		}
6211 		ANON_LOCK_EXIT(&amp->a_rwlock);
6212 	}
6213 
6214 	seg->s_szc = szc;
6215 
6216 	return (0);
6217 }
6218 
6219 static int
6220 segvn_clrszc(struct seg *seg)
6221 {
6222 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6223 	struct anon_map *amp = svd->amp;
6224 	size_t pgsz;
6225 	pgcnt_t pages;
6226 	int err = 0;
6227 	caddr_t a = seg->s_base;
6228 	caddr_t ea = a + seg->s_size;
6229 	ulong_t an_idx = svd->anon_index;
6230 	vnode_t *vp = svd->vp;
6231 	struct vpage *vpage = svd->vpage;
6232 	page_t *anon_pl[1 + 1], *pp;
6233 	struct anon *ap, *oldap;
6234 	uint_t prot = svd->prot, vpprot;
6235 	int pageflag = 0;
6236 
6237 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
6238 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
6239 	ASSERT(svd->softlockcnt == 0);
6240 
6241 	if (vp == NULL && amp == NULL) {
6242 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6243 		seg->s_szc = 0;
6244 		return (0);
6245 	}
6246 
6247 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6248 		ASSERT(svd->amp == NULL);
6249 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6250 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6251 		    HAT_REGION_TEXT);
6252 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6253 	} else if (svd->tr_state == SEGVN_TR_ON) {
6254 		ASSERT(svd->amp != NULL);
6255 		segvn_textunrepl(seg, 1);
6256 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6257 		amp = NULL;
6258 	} else {
6259 		if (svd->tr_state != SEGVN_TR_OFF) {
6260 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
6261 			svd->tr_state = SEGVN_TR_OFF;
6262 		}
6263 
6264 		/*
6265 		 * do HAT_UNLOAD_UNMAP since we are changing the pagesize.
6266 		 * unload argument is 0 when we are freeing the segment
6267 		 * and unload was already done.
6268 		 */
6269 		hat_unload(seg->s_as->a_hat, seg->s_base, seg->s_size,
6270 		    HAT_UNLOAD_UNMAP);
6271 	}
6272 
6273 	if (amp == NULL || svd->type == MAP_SHARED) {
6274 		seg->s_szc = 0;
6275 		return (0);
6276 	}
6277 
6278 	pgsz = page_get_pagesize(seg->s_szc);
6279 	pages = btop(pgsz);
6280 
6281 	/*
6282 	 * XXX anon rwlock is not really needed because this is a
6283 	 * private segment and we are writers.
6284 	 */
6285 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6286 
6287 	for (; a < ea; a += pgsz, an_idx += pages) {
6288 		if ((oldap = anon_get_ptr(amp->ahp, an_idx)) != NULL) {
6289 			ASSERT(vpage != NULL || svd->pageprot == 0);
6290 			if (vpage != NULL) {
6291 				ASSERT(sameprot(seg, a, pgsz));
6292 				prot = VPP_PROT(vpage);
6293 				pageflag = VPP_ISPPLOCK(vpage) ? LOCK_PAGE : 0;
6294 			}
6295 			if (seg->s_szc != 0) {
6296 				ASSERT(vp == NULL || anon_pages(amp->ahp,
6297 				    an_idx, pages) == pages);
6298 				if ((err = anon_map_demotepages(amp, an_idx,
6299 				    seg, a, prot, vpage, svd->cred)) != 0) {
6300 					goto out;
6301 				}
6302 			} else {
6303 				if (oldap->an_refcnt == 1) {
6304 					continue;
6305 				}
6306 				if ((err = anon_getpage(&oldap, &vpprot,
6307 				    anon_pl, PAGESIZE, seg, a, S_READ,
6308 				    svd->cred))) {
6309 					goto out;
6310 				}
6311 				if ((pp = anon_private(&ap, seg, a, prot,
6312 				    anon_pl[0], pageflag, svd->cred)) == NULL) {
6313 					err = ENOMEM;
6314 					goto out;
6315 				}
6316 				anon_decref(oldap);
6317 				(void) anon_set_ptr(amp->ahp, an_idx, ap,
6318 				    ANON_SLEEP);
6319 				page_unlock(pp);
6320 			}
6321 		}
6322 		vpage = (vpage == NULL) ? NULL : vpage + pages;
6323 	}
6324 
6325 	amp->a_szc = 0;
6326 	seg->s_szc = 0;
6327 out:
6328 	ANON_LOCK_EXIT(&amp->a_rwlock);
6329 	return (err);
6330 }
6331 
6332 static int
6333 segvn_claim_pages(
6334 	struct seg *seg,
6335 	struct vpage *svp,
6336 	u_offset_t off,
6337 	ulong_t anon_idx,
6338 	uint_t prot)
6339 {
6340 	pgcnt_t	pgcnt = page_get_pagecnt(seg->s_szc);
6341 	size_t ppasize = (pgcnt + 1) * sizeof (page_t *);
6342 	page_t	**ppa;
6343 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6344 	struct anon_map *amp = svd->amp;
6345 	struct vpage *evp = svp + pgcnt;
6346 	caddr_t addr = ((uintptr_t)(svp - svd->vpage) << PAGESHIFT)
6347 	    + seg->s_base;
6348 	struct anon *ap;
6349 	struct vnode *vp = svd->vp;
6350 	page_t *pp;
6351 	pgcnt_t pg_idx, i;
6352 	int err = 0;
6353 	anoff_t aoff;
6354 	int anon = (amp != NULL) ? 1 : 0;
6355 
6356 	ASSERT(svd->type == MAP_PRIVATE);
6357 	ASSERT(svd->vpage != NULL);
6358 	ASSERT(seg->s_szc != 0);
6359 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
6360 	ASSERT(amp == NULL || IS_P2ALIGNED(anon_idx, pgcnt));
6361 	ASSERT(sameprot(seg, addr, pgcnt << PAGESHIFT));
6362 
6363 	if (VPP_PROT(svp) == prot)
6364 		return (1);
6365 	if (!((VPP_PROT(svp) ^ prot) & PROT_WRITE))
6366 		return (1);
6367 
6368 	ppa = kmem_alloc(ppasize, KM_SLEEP);
6369 	if (anon && vp != NULL) {
6370 		if (anon_get_ptr(amp->ahp, anon_idx) == NULL) {
6371 			anon = 0;
6372 			ASSERT(!anon_pages(amp->ahp, anon_idx, pgcnt));
6373 		}
6374 		ASSERT(!anon ||
6375 		    anon_pages(amp->ahp, anon_idx, pgcnt) == pgcnt);
6376 	}
6377 
6378 	for (*ppa = NULL, pg_idx = 0; svp < evp; svp++, anon_idx++) {
6379 		if (!VPP_ISPPLOCK(svp))
6380 			continue;
6381 		if (anon) {
6382 			ap = anon_get_ptr(amp->ahp, anon_idx);
6383 			if (ap == NULL) {
6384 				panic("segvn_claim_pages: no anon slot");
6385 			}
6386 			swap_xlate(ap, &vp, &aoff);
6387 			off = (u_offset_t)aoff;
6388 		}
6389 		ASSERT(vp != NULL);
6390 		if ((pp = page_lookup(vp,
6391 		    (u_offset_t)off, SE_SHARED)) == NULL) {
6392 			panic("segvn_claim_pages: no page");
6393 		}
6394 		ppa[pg_idx++] = pp;
6395 		off += PAGESIZE;
6396 	}
6397 
6398 	if (ppa[0] == NULL) {
6399 		kmem_free(ppa, ppasize);
6400 		return (1);
6401 	}
6402 
6403 	ASSERT(pg_idx <= pgcnt);
6404 	ppa[pg_idx] = NULL;
6405 
6406 	if (prot & PROT_WRITE)
6407 		err = page_addclaim_pages(ppa);
6408 	else
6409 		err = page_subclaim_pages(ppa);
6410 
6411 	for (i = 0; i < pg_idx; i++) {
6412 		ASSERT(ppa[i] != NULL);
6413 		page_unlock(ppa[i]);
6414 	}
6415 
6416 	kmem_free(ppa, ppasize);
6417 	return (err);
6418 }
6419 
6420 /*
6421  * Returns right (upper address) segment if split occured.
6422  * If the address is equal to the beginning or end of its segment it returns
6423  * the current segment.
6424  */
6425 static struct seg *
6426 segvn_split_seg(struct seg *seg, caddr_t addr)
6427 {
6428 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6429 	struct seg *nseg;
6430 	size_t nsize;
6431 	struct segvn_data *nsvd;
6432 
6433 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6434 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6435 
6436 	ASSERT(addr >= seg->s_base);
6437 	ASSERT(addr <= seg->s_base + seg->s_size);
6438 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6439 
6440 	if (addr == seg->s_base || addr == seg->s_base + seg->s_size)
6441 		return (seg);
6442 
6443 	nsize = seg->s_base + seg->s_size - addr;
6444 	seg->s_size = addr - seg->s_base;
6445 	nseg = seg_alloc(seg->s_as, addr, nsize);
6446 	ASSERT(nseg != NULL);
6447 	nseg->s_ops = seg->s_ops;
6448 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
6449 	nseg->s_data = (void *)nsvd;
6450 	nseg->s_szc = seg->s_szc;
6451 	*nsvd = *svd;
6452 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6453 	nsvd->seg = nseg;
6454 	rw_init(&nsvd->lock, NULL, RW_DEFAULT, NULL);
6455 
6456 	if (nsvd->vp != NULL) {
6457 		VN_HOLD(nsvd->vp);
6458 		nsvd->offset = svd->offset +
6459 		    (uintptr_t)(nseg->s_base - seg->s_base);
6460 		if (nsvd->type == MAP_SHARED)
6461 			lgrp_shm_policy_init(NULL, nsvd->vp);
6462 	} else {
6463 		/*
6464 		 * The offset for an anonymous segment has no signifigance in
6465 		 * terms of an offset into a file. If we were to use the above
6466 		 * calculation instead, the structures read out of
6467 		 * /proc/<pid>/xmap would be more difficult to decipher since
6468 		 * it would be unclear whether two seemingly contiguous
6469 		 * prxmap_t structures represented different segments or a
6470 		 * single segment that had been split up into multiple prxmap_t
6471 		 * structures (e.g. if some part of the segment had not yet
6472 		 * been faulted in).
6473 		 */
6474 		nsvd->offset = 0;
6475 	}
6476 
6477 	ASSERT(svd->softlockcnt == 0);
6478 	crhold(svd->cred);
6479 
6480 	if (svd->vpage != NULL) {
6481 		size_t bytes = vpgtob(seg_pages(seg));
6482 		size_t nbytes = vpgtob(seg_pages(nseg));
6483 		struct vpage *ovpage = svd->vpage;
6484 
6485 		svd->vpage = kmem_alloc(bytes, KM_SLEEP);
6486 		bcopy(ovpage, svd->vpage, bytes);
6487 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
6488 		bcopy(ovpage + seg_pages(seg), nsvd->vpage, nbytes);
6489 		kmem_free(ovpage, bytes + nbytes);
6490 	}
6491 	if (svd->amp != NULL && svd->type == MAP_PRIVATE) {
6492 		struct anon_map *oamp = svd->amp, *namp;
6493 		struct anon_hdr *nahp;
6494 
6495 		ANON_LOCK_ENTER(&oamp->a_rwlock, RW_WRITER);
6496 		ASSERT(oamp->refcnt == 1);
6497 		nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
6498 		(void) anon_copy_ptr(oamp->ahp, svd->anon_index,
6499 		    nahp, 0, btop(seg->s_size), ANON_SLEEP);
6500 
6501 		namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
6502 		namp->a_szc = nseg->s_szc;
6503 		(void) anon_copy_ptr(oamp->ahp,
6504 		    svd->anon_index + btop(seg->s_size),
6505 		    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
6506 		anon_release(oamp->ahp, btop(oamp->size));
6507 		oamp->ahp = nahp;
6508 		oamp->size = seg->s_size;
6509 		svd->anon_index = 0;
6510 		nsvd->amp = namp;
6511 		nsvd->anon_index = 0;
6512 		ANON_LOCK_EXIT(&oamp->a_rwlock);
6513 	} else if (svd->amp != NULL) {
6514 		pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6515 		ASSERT(svd->amp == nsvd->amp);
6516 		ASSERT(seg->s_szc <= svd->amp->a_szc);
6517 		nsvd->anon_index = svd->anon_index + seg_pages(seg);
6518 		ASSERT(IS_P2ALIGNED(nsvd->anon_index, pgcnt));
6519 		ANON_LOCK_ENTER(&svd->amp->a_rwlock, RW_WRITER);
6520 		svd->amp->refcnt++;
6521 		ANON_LOCK_EXIT(&svd->amp->a_rwlock);
6522 	}
6523 
6524 	/*
6525 	 * Split amount of swap reserve
6526 	 */
6527 	if (svd->swresv) {
6528 		/*
6529 		 * For MAP_NORESERVE, only allocate swap reserve for pages
6530 		 * being used.  Other segments get enough to cover whole
6531 		 * segment.
6532 		 */
6533 		if (svd->flags & MAP_NORESERVE) {
6534 			size_t	oswresv;
6535 
6536 			ASSERT(svd->amp);
6537 			oswresv = svd->swresv;
6538 			svd->swresv = ptob(anon_pages(svd->amp->ahp,
6539 				svd->anon_index, btop(seg->s_size)));
6540 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
6541 				nsvd->anon_index, btop(nseg->s_size)));
6542 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
6543 		} else {
6544 			ASSERT(svd->swresv == seg->s_size + nseg->s_size);
6545 			svd->swresv = seg->s_size;
6546 			nsvd->swresv = nseg->s_size;
6547 		}
6548 	}
6549 
6550 	return (nseg);
6551 }
6552 
6553 /*
6554  * called on memory operations (unmap, setprot, setpagesize) for a subset
6555  * of a large page segment to either demote the memory range (SDR_RANGE)
6556  * or the ends (SDR_END) by addr/len.
6557  *
6558  * returns 0 on success. returns errno, including ENOMEM, on failure.
6559  */
6560 static int
6561 segvn_demote_range(
6562 	struct seg *seg,
6563 	caddr_t addr,
6564 	size_t len,
6565 	int flag,
6566 	uint_t szcvec)
6567 {
6568 	caddr_t eaddr = addr + len;
6569 	caddr_t lpgaddr, lpgeaddr;
6570 	struct seg *nseg;
6571 	struct seg *badseg1 = NULL;
6572 	struct seg *badseg2 = NULL;
6573 	size_t pgsz;
6574 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6575 	int err;
6576 	uint_t szc = seg->s_szc;
6577 	uint_t tszcvec;
6578 
6579 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6580 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6581 	ASSERT(szc != 0);
6582 	pgsz = page_get_pagesize(szc);
6583 	ASSERT(seg->s_base != addr || seg->s_size != len);
6584 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6585 	ASSERT(svd->softlockcnt == 0);
6586 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6587 	ASSERT(szcvec == 0 || (flag == SDR_END && svd->type == MAP_SHARED));
6588 
6589 	CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
6590 	ASSERT(flag == SDR_RANGE || eaddr < lpgeaddr || addr > lpgaddr);
6591 	if (flag == SDR_RANGE) {
6592 		/* demote entire range */
6593 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6594 		(void) segvn_split_seg(nseg, lpgeaddr);
6595 		ASSERT(badseg1->s_base == lpgaddr);
6596 		ASSERT(badseg1->s_size == lpgeaddr - lpgaddr);
6597 	} else if (addr != lpgaddr) {
6598 		ASSERT(flag == SDR_END);
6599 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6600 		if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz &&
6601 		    eaddr < lpgaddr + 2 * pgsz) {
6602 			(void) segvn_split_seg(nseg, lpgeaddr);
6603 			ASSERT(badseg1->s_base == lpgaddr);
6604 			ASSERT(badseg1->s_size == 2 * pgsz);
6605 		} else {
6606 			nseg = segvn_split_seg(nseg, lpgaddr + pgsz);
6607 			ASSERT(badseg1->s_base == lpgaddr);
6608 			ASSERT(badseg1->s_size == pgsz);
6609 			if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz) {
6610 				ASSERT(lpgeaddr - lpgaddr > 2 * pgsz);
6611 				nseg = segvn_split_seg(nseg, lpgeaddr - pgsz);
6612 				badseg2 = nseg;
6613 				(void) segvn_split_seg(nseg, lpgeaddr);
6614 				ASSERT(badseg2->s_base == lpgeaddr - pgsz);
6615 				ASSERT(badseg2->s_size == pgsz);
6616 			}
6617 		}
6618 	} else {
6619 		ASSERT(flag == SDR_END);
6620 		ASSERT(eaddr < lpgeaddr);
6621 		badseg1 = nseg = segvn_split_seg(seg, lpgeaddr - pgsz);
6622 		(void) segvn_split_seg(nseg, lpgeaddr);
6623 		ASSERT(badseg1->s_base == lpgeaddr - pgsz);
6624 		ASSERT(badseg1->s_size == pgsz);
6625 	}
6626 
6627 	ASSERT(badseg1 != NULL);
6628 	ASSERT(badseg1->s_szc == szc);
6629 	ASSERT(flag == SDR_RANGE || badseg1->s_size == pgsz ||
6630 	    badseg1->s_size == 2 * pgsz);
6631 	ASSERT(sameprot(badseg1, badseg1->s_base, pgsz));
6632 	ASSERT(badseg1->s_size == pgsz ||
6633 	    sameprot(badseg1, badseg1->s_base + pgsz, pgsz));
6634 	if (err = segvn_clrszc(badseg1)) {
6635 		return (err);
6636 	}
6637 	ASSERT(badseg1->s_szc == 0);
6638 
6639 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6640 		uint_t tszc = highbit(tszcvec) - 1;
6641 		caddr_t ta = MAX(addr, badseg1->s_base);
6642 		caddr_t te;
6643 		size_t tpgsz = page_get_pagesize(tszc);
6644 
6645 		ASSERT(svd->type == MAP_SHARED);
6646 		ASSERT(flag == SDR_END);
6647 		ASSERT(tszc < szc && tszc > 0);
6648 
6649 		if (eaddr > badseg1->s_base + badseg1->s_size) {
6650 			te = badseg1->s_base + badseg1->s_size;
6651 		} else {
6652 			te = eaddr;
6653 		}
6654 
6655 		ASSERT(ta <= te);
6656 		badseg1->s_szc = tszc;
6657 		if (!IS_P2ALIGNED(ta, tpgsz) || !IS_P2ALIGNED(te, tpgsz)) {
6658 			if (badseg2 != NULL) {
6659 				err = segvn_demote_range(badseg1, ta, te - ta,
6660 				    SDR_END, tszcvec);
6661 				if (err != 0) {
6662 					return (err);
6663 				}
6664 			} else {
6665 				return (segvn_demote_range(badseg1, ta,
6666 				    te - ta, SDR_END, tszcvec));
6667 			}
6668 		}
6669 	}
6670 
6671 	if (badseg2 == NULL)
6672 		return (0);
6673 	ASSERT(badseg2->s_szc == szc);
6674 	ASSERT(badseg2->s_size == pgsz);
6675 	ASSERT(sameprot(badseg2, badseg2->s_base, badseg2->s_size));
6676 	if (err = segvn_clrszc(badseg2)) {
6677 		return (err);
6678 	}
6679 	ASSERT(badseg2->s_szc == 0);
6680 
6681 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6682 		uint_t tszc = highbit(tszcvec) - 1;
6683 		size_t tpgsz = page_get_pagesize(tszc);
6684 
6685 		ASSERT(svd->type == MAP_SHARED);
6686 		ASSERT(flag == SDR_END);
6687 		ASSERT(tszc < szc && tszc > 0);
6688 		ASSERT(badseg2->s_base > addr);
6689 		ASSERT(eaddr > badseg2->s_base);
6690 		ASSERT(eaddr < badseg2->s_base + badseg2->s_size);
6691 
6692 		badseg2->s_szc = tszc;
6693 		if (!IS_P2ALIGNED(eaddr, tpgsz)) {
6694 			return (segvn_demote_range(badseg2, badseg2->s_base,
6695 			    eaddr - badseg2->s_base, SDR_END, tszcvec));
6696 		}
6697 	}
6698 
6699 	return (0);
6700 }
6701 
6702 static int
6703 segvn_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
6704 {
6705 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6706 	struct vpage *vp, *evp;
6707 
6708 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6709 
6710 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6711 	/*
6712 	 * If segment protection can be used, simply check against them.
6713 	 */
6714 	if (svd->pageprot == 0) {
6715 		int err;
6716 
6717 		err = ((svd->prot & prot) != prot) ? EACCES : 0;
6718 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6719 		return (err);
6720 	}
6721 
6722 	/*
6723 	 * Have to check down to the vpage level.
6724 	 */
6725 	evp = &svd->vpage[seg_page(seg, addr + len)];
6726 	for (vp = &svd->vpage[seg_page(seg, addr)]; vp < evp; vp++) {
6727 		if ((VPP_PROT(vp) & prot) != prot) {
6728 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6729 			return (EACCES);
6730 		}
6731 	}
6732 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6733 	return (0);
6734 }
6735 
6736 static int
6737 segvn_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
6738 {
6739 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6740 	size_t pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
6741 
6742 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6743 
6744 	if (pgno != 0) {
6745 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6746 		if (svd->pageprot == 0) {
6747 			do
6748 				protv[--pgno] = svd->prot;
6749 			while (pgno != 0);
6750 		} else {
6751 			size_t pgoff = seg_page(seg, addr);
6752 
6753 			do {
6754 				pgno--;
6755 				protv[pgno] = VPP_PROT(&svd->vpage[pgno+pgoff]);
6756 			} while (pgno != 0);
6757 		}
6758 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6759 	}
6760 	return (0);
6761 }
6762 
6763 static u_offset_t
6764 segvn_getoffset(struct seg *seg, caddr_t addr)
6765 {
6766 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6767 
6768 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6769 
6770 	return (svd->offset + (uintptr_t)(addr - seg->s_base));
6771 }
6772 
6773 /*ARGSUSED*/
6774 static int
6775 segvn_gettype(struct seg *seg, caddr_t addr)
6776 {
6777 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6778 
6779 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6780 
6781 	return (svd->type | (svd->flags & (MAP_NORESERVE | MAP_TEXT |
6782 	    MAP_INITDATA)));
6783 }
6784 
6785 /*ARGSUSED*/
6786 static int
6787 segvn_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
6788 {
6789 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6790 
6791 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6792 
6793 	*vpp = svd->vp;
6794 	return (0);
6795 }
6796 
6797 /*
6798  * Check to see if it makes sense to do kluster/read ahead to
6799  * addr + delta relative to the mapping at addr.  We assume here
6800  * that delta is a signed PAGESIZE'd multiple (which can be negative).
6801  *
6802  * For segvn, we currently "approve" of the action if we are
6803  * still in the segment and it maps from the same vp/off,
6804  * or if the advice stored in segvn_data or vpages allows it.
6805  * Currently, klustering is not allowed only if MADV_RANDOM is set.
6806  */
6807 static int
6808 segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
6809 {
6810 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6811 	struct anon *oap, *ap;
6812 	ssize_t pd;
6813 	size_t page;
6814 	struct vnode *vp1, *vp2;
6815 	u_offset_t off1, off2;
6816 	struct anon_map *amp;
6817 
6818 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6819 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
6820 	    SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
6821 
6822 	if (addr + delta < seg->s_base ||
6823 	    addr + delta >= (seg->s_base + seg->s_size))
6824 		return (-1);		/* exceeded segment bounds */
6825 
6826 	pd = delta / (ssize_t)PAGESIZE;	/* divide to preserve sign bit */
6827 	page = seg_page(seg, addr);
6828 
6829 	/*
6830 	 * Check to see if either of the pages addr or addr + delta
6831 	 * have advice set that prevents klustering (if MADV_RANDOM advice
6832 	 * is set for entire segment, or MADV_SEQUENTIAL is set and delta
6833 	 * is negative).
6834 	 */
6835 	if (svd->advice == MADV_RANDOM ||
6836 	    svd->advice == MADV_SEQUENTIAL && delta < 0)
6837 		return (-1);
6838 	else if (svd->pageadvice && svd->vpage) {
6839 		struct vpage *bvpp, *evpp;
6840 
6841 		bvpp = &svd->vpage[page];
6842 		evpp = &svd->vpage[page + pd];
6843 		if (VPP_ADVICE(bvpp) == MADV_RANDOM ||
6844 		    VPP_ADVICE(evpp) == MADV_SEQUENTIAL && delta < 0)
6845 			return (-1);
6846 		if (VPP_ADVICE(bvpp) != VPP_ADVICE(evpp) &&
6847 		    VPP_ADVICE(evpp) == MADV_RANDOM)
6848 			return (-1);
6849 	}
6850 
6851 	if (svd->type == MAP_SHARED)
6852 		return (0);		/* shared mapping - all ok */
6853 
6854 	if ((amp = svd->amp) == NULL)
6855 		return (0);		/* off original vnode */
6856 
6857 	page += svd->anon_index;
6858 
6859 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
6860 
6861 	oap = anon_get_ptr(amp->ahp, page);
6862 	ap = anon_get_ptr(amp->ahp, page + pd);
6863 
6864 	ANON_LOCK_EXIT(&amp->a_rwlock);
6865 
6866 	if ((oap == NULL && ap != NULL) || (oap != NULL && ap == NULL)) {
6867 		return (-1);		/* one with and one without an anon */
6868 	}
6869 
6870 	if (oap == NULL) {		/* implies that ap == NULL */
6871 		return (0);		/* off original vnode */
6872 	}
6873 
6874 	/*
6875 	 * Now we know we have two anon pointers - check to
6876 	 * see if they happen to be properly allocated.
6877 	 */
6878 
6879 	/*
6880 	 * XXX We cheat here and don't lock the anon slots. We can't because
6881 	 * we may have been called from the anon layer which might already
6882 	 * have locked them. We are holding a refcnt on the slots so they
6883 	 * can't disappear. The worst that will happen is we'll get the wrong
6884 	 * names (vp, off) for the slots and make a poor klustering decision.
6885 	 */
6886 	swap_xlate(ap, &vp1, &off1);
6887 	swap_xlate(oap, &vp2, &off2);
6888 
6889 
6890 	if (!VOP_CMP(vp1, vp2) || off1 - off2 != delta)
6891 		return (-1);
6892 	return (0);
6893 }
6894 
6895 /*
6896  * Swap the pages of seg out to secondary storage, returning the
6897  * number of bytes of storage freed.
6898  *
6899  * The basic idea is first to unload all translations and then to call
6900  * VOP_PUTPAGE() for all newly-unmapped pages, to push them out to the
6901  * swap device.  Pages to which other segments have mappings will remain
6902  * mapped and won't be swapped.  Our caller (as_swapout) has already
6903  * performed the unloading step.
6904  *
6905  * The value returned is intended to correlate well with the process's
6906  * memory requirements.  However, there are some caveats:
6907  * 1)	When given a shared segment as argument, this routine will
6908  *	only succeed in swapping out pages for the last sharer of the
6909  *	segment.  (Previous callers will only have decremented mapping
6910  *	reference counts.)
6911  * 2)	We assume that the hat layer maintains a large enough translation
6912  *	cache to capture process reference patterns.
6913  */
6914 static size_t
6915 segvn_swapout(struct seg *seg)
6916 {
6917 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6918 	struct anon_map *amp;
6919 	pgcnt_t pgcnt = 0;
6920 	pgcnt_t npages;
6921 	pgcnt_t page;
6922 	ulong_t anon_index;
6923 
6924 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6925 
6926 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6927 	/*
6928 	 * Find pages unmapped by our caller and force them
6929 	 * out to the virtual swap device.
6930 	 */
6931 	if ((amp = svd->amp) != NULL)
6932 		anon_index = svd->anon_index;
6933 	npages = seg->s_size >> PAGESHIFT;
6934 	for (page = 0; page < npages; page++) {
6935 		page_t *pp;
6936 		struct anon *ap;
6937 		struct vnode *vp;
6938 		u_offset_t off;
6939 		anon_sync_obj_t cookie;
6940 
6941 		/*
6942 		 * Obtain <vp, off> pair for the page, then look it up.
6943 		 *
6944 		 * Note that this code is willing to consider regular
6945 		 * pages as well as anon pages.  Is this appropriate here?
6946 		 */
6947 		ap = NULL;
6948 		if (amp != NULL) {
6949 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
6950 			if (anon_array_try_enter(amp, anon_index + page,
6951 						&cookie)) {
6952 				ANON_LOCK_EXIT(&amp->a_rwlock);
6953 				continue;
6954 			}
6955 			ap = anon_get_ptr(amp->ahp, anon_index + page);
6956 			if (ap != NULL) {
6957 				swap_xlate(ap, &vp, &off);
6958 			} else {
6959 				vp = svd->vp;
6960 				off = svd->offset + ptob(page);
6961 			}
6962 			anon_array_exit(&cookie);
6963 			ANON_LOCK_EXIT(&amp->a_rwlock);
6964 		} else {
6965 			vp = svd->vp;
6966 			off = svd->offset + ptob(page);
6967 		}
6968 		if (vp == NULL) {		/* untouched zfod page */
6969 			ASSERT(ap == NULL);
6970 			continue;
6971 		}
6972 
6973 		pp = page_lookup_nowait(vp, off, SE_SHARED);
6974 		if (pp == NULL)
6975 			continue;
6976 
6977 
6978 		/*
6979 		 * Examine the page to see whether it can be tossed out,
6980 		 * keeping track of how many we've found.
6981 		 */
6982 		if (!page_tryupgrade(pp)) {
6983 			/*
6984 			 * If the page has an i/o lock and no mappings,
6985 			 * it's very likely that the page is being
6986 			 * written out as a result of klustering.
6987 			 * Assume this is so and take credit for it here.
6988 			 */
6989 			if (!page_io_trylock(pp)) {
6990 				if (!hat_page_is_mapped(pp))
6991 					pgcnt++;
6992 			} else {
6993 				page_io_unlock(pp);
6994 			}
6995 			page_unlock(pp);
6996 			continue;
6997 		}
6998 		ASSERT(!page_iolock_assert(pp));
6999 
7000 
7001 		/*
7002 		 * Skip if page is locked or has mappings.
7003 		 * We don't need the page_struct_lock to look at lckcnt
7004 		 * and cowcnt because the page is exclusive locked.
7005 		 */
7006 		if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0 ||
7007 		    hat_page_is_mapped(pp)) {
7008 			page_unlock(pp);
7009 			continue;
7010 		}
7011 
7012 		/*
7013 		 * dispose skips large pages so try to demote first.
7014 		 */
7015 		if (pp->p_szc != 0 && !page_try_demote_pages(pp)) {
7016 			page_unlock(pp);
7017 			/*
7018 			 * XXX should skip the remaining page_t's of this
7019 			 * large page.
7020 			 */
7021 			continue;
7022 		}
7023 
7024 		ASSERT(pp->p_szc == 0);
7025 
7026 		/*
7027 		 * No longer mapped -- we can toss it out.  How
7028 		 * we do so depends on whether or not it's dirty.
7029 		 */
7030 		if (hat_ismod(pp) && pp->p_vnode) {
7031 			/*
7032 			 * We must clean the page before it can be
7033 			 * freed.  Setting B_FREE will cause pvn_done
7034 			 * to free the page when the i/o completes.
7035 			 * XXX:	This also causes it to be accounted
7036 			 *	as a pageout instead of a swap: need
7037 			 *	B_SWAPOUT bit to use instead of B_FREE.
7038 			 *
7039 			 * Hold the vnode before releasing the page lock
7040 			 * to prevent it from being freed and re-used by
7041 			 * some other thread.
7042 			 */
7043 			VN_HOLD(vp);
7044 			page_unlock(pp);
7045 
7046 			/*
7047 			 * Queue all i/o requests for the pageout thread
7048 			 * to avoid saturating the pageout devices.
7049 			 */
7050 			if (!queue_io_request(vp, off))
7051 				VN_RELE(vp);
7052 		} else {
7053 			/*
7054 			 * The page was clean, free it.
7055 			 *
7056 			 * XXX:	Can we ever encounter modified pages
7057 			 *	with no associated vnode here?
7058 			 */
7059 			ASSERT(pp->p_vnode != NULL);
7060 			/*LINTED: constant in conditional context*/
7061 			VN_DISPOSE(pp, B_FREE, 0, kcred);
7062 		}
7063 
7064 		/*
7065 		 * Credit now even if i/o is in progress.
7066 		 */
7067 		pgcnt++;
7068 	}
7069 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7070 
7071 	/*
7072 	 * Wakeup pageout to initiate i/o on all queued requests.
7073 	 */
7074 	cv_signal_pageout();
7075 	return (ptob(pgcnt));
7076 }
7077 
7078 /*
7079  * Synchronize primary storage cache with real object in virtual memory.
7080  *
7081  * XXX - Anonymous pages should not be sync'ed out at all.
7082  */
7083 static int
7084 segvn_sync(struct seg *seg, caddr_t addr, size_t len, int attr, uint_t flags)
7085 {
7086 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7087 	struct vpage *vpp;
7088 	page_t *pp;
7089 	u_offset_t offset;
7090 	struct vnode *vp;
7091 	u_offset_t off;
7092 	caddr_t eaddr;
7093 	int bflags;
7094 	int err = 0;
7095 	int segtype;
7096 	int pageprot;
7097 	int prot;
7098 	ulong_t anon_index;
7099 	struct anon_map *amp;
7100 	struct anon *ap;
7101 	anon_sync_obj_t cookie;
7102 
7103 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7104 
7105 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7106 
7107 	if (svd->softlockcnt > 0) {
7108 		/*
7109 		 * flush all pages from seg cache
7110 		 * otherwise we may deadlock in swap_putpage
7111 		 * for B_INVAL page (4175402).
7112 		 *
7113 		 * Even if we grab segvn WRITER's lock or segp_slock
7114 		 * here, there might be another thread which could've
7115 		 * successfully performed lookup/insert just before
7116 		 * we acquired the lock here.  So, grabbing either
7117 		 * lock here is of not much use.  Until we devise
7118 		 * a strategy at upper layers to solve the
7119 		 * synchronization issues completely, we expect
7120 		 * applications to handle this appropriately.
7121 		 */
7122 		segvn_purge(seg);
7123 		if (svd->softlockcnt > 0) {
7124 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7125 			return (EAGAIN);
7126 		}
7127 	}
7128 
7129 	vpp = svd->vpage;
7130 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7131 	bflags = ((flags & MS_ASYNC) ? B_ASYNC : 0) |
7132 	    ((flags & MS_INVALIDATE) ? B_INVAL : 0);
7133 
7134 	if (attr) {
7135 		pageprot = attr & ~(SHARED|PRIVATE);
7136 		segtype = (attr & SHARED) ? MAP_SHARED : MAP_PRIVATE;
7137 
7138 		/*
7139 		 * We are done if the segment types don't match
7140 		 * or if we have segment level protections and
7141 		 * they don't match.
7142 		 */
7143 		if (svd->type != segtype) {
7144 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7145 			return (0);
7146 		}
7147 		if (vpp == NULL) {
7148 			if (svd->prot != pageprot) {
7149 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7150 				return (0);
7151 			}
7152 			prot = svd->prot;
7153 		} else
7154 			vpp = &svd->vpage[seg_page(seg, addr)];
7155 
7156 	} else if (svd->vp && svd->amp == NULL &&
7157 	    (flags & MS_INVALIDATE) == 0) {
7158 
7159 		/*
7160 		 * No attributes, no anonymous pages and MS_INVALIDATE flag
7161 		 * is not on, just use one big request.
7162 		 */
7163 		err = VOP_PUTPAGE(svd->vp, (offset_t)offset, len,
7164 		    bflags, svd->cred);
7165 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7166 		return (err);
7167 	}
7168 
7169 	if ((amp = svd->amp) != NULL)
7170 		anon_index = svd->anon_index + seg_page(seg, addr);
7171 
7172 	for (eaddr = addr + len; addr < eaddr; addr += PAGESIZE) {
7173 		ap = NULL;
7174 		if (amp != NULL) {
7175 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7176 			anon_array_enter(amp, anon_index, &cookie);
7177 			ap = anon_get_ptr(amp->ahp, anon_index++);
7178 			if (ap != NULL) {
7179 				swap_xlate(ap, &vp, &off);
7180 			} else {
7181 				vp = svd->vp;
7182 				off = offset;
7183 			}
7184 			anon_array_exit(&cookie);
7185 			ANON_LOCK_EXIT(&amp->a_rwlock);
7186 		} else {
7187 			vp = svd->vp;
7188 			off = offset;
7189 		}
7190 		offset += PAGESIZE;
7191 
7192 		if (vp == NULL)		/* untouched zfod page */
7193 			continue;
7194 
7195 		if (attr) {
7196 			if (vpp) {
7197 				prot = VPP_PROT(vpp);
7198 				vpp++;
7199 			}
7200 			if (prot != pageprot) {
7201 				continue;
7202 			}
7203 		}
7204 
7205 		/*
7206 		 * See if any of these pages are locked --  if so, then we
7207 		 * will have to truncate an invalidate request at the first
7208 		 * locked one. We don't need the page_struct_lock to test
7209 		 * as this is only advisory; even if we acquire it someone
7210 		 * might race in and lock the page after we unlock and before
7211 		 * we do the PUTPAGE, then PUTPAGE simply does nothing.
7212 		 */
7213 		if (flags & MS_INVALIDATE) {
7214 			if ((pp = page_lookup(vp, off, SE_SHARED)) != NULL) {
7215 				if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
7216 					page_unlock(pp);
7217 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7218 					return (EBUSY);
7219 				}
7220 				if (ap != NULL && pp->p_szc != 0 &&
7221 				    page_tryupgrade(pp)) {
7222 					if (pp->p_lckcnt == 0 &&
7223 					    pp->p_cowcnt == 0) {
7224 						/*
7225 						 * swapfs VN_DISPOSE() won't
7226 						 * invalidate large pages.
7227 						 * Attempt to demote.
7228 						 * XXX can't help it if it
7229 						 * fails. But for swapfs
7230 						 * pages it is no big deal.
7231 						 */
7232 						(void) page_try_demote_pages(
7233 						    pp);
7234 				    }
7235 				}
7236 				page_unlock(pp);
7237 			}
7238 		} else if (svd->type == MAP_SHARED && amp != NULL) {
7239 			/*
7240 			 * Avoid writting out to disk ISM's large pages
7241 			 * because segspt_free_pages() relies on NULL an_pvp
7242 			 * of anon slots of such pages.
7243 			 */
7244 
7245 			ASSERT(svd->vp == NULL);
7246 			/*
7247 			 * swapfs uses page_lookup_nowait if not freeing or
7248 			 * invalidating and skips a page if
7249 			 * page_lookup_nowait returns NULL.
7250 			 */
7251 			pp = page_lookup_nowait(vp, off, SE_SHARED);
7252 			if (pp == NULL) {
7253 				continue;
7254 			}
7255 			if (pp->p_szc != 0) {
7256 				page_unlock(pp);
7257 				continue;
7258 			}
7259 
7260 			/*
7261 			 * Note ISM pages are created large so (vp, off)'s
7262 			 * page cannot suddenly become large after we unlock
7263 			 * pp.
7264 			 */
7265 			page_unlock(pp);
7266 		}
7267 		/*
7268 		 * XXX - Should ultimately try to kluster
7269 		 * calls to VOP_PUTPAGE() for performance.
7270 		 */
7271 		VN_HOLD(vp);
7272 		err = VOP_PUTPAGE(vp, (offset_t)off, PAGESIZE,
7273 		    bflags, svd->cred);
7274 		VN_RELE(vp);
7275 		if (err)
7276 			break;
7277 	}
7278 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7279 	return (err);
7280 }
7281 
7282 /*
7283  * Determine if we have data corresponding to pages in the
7284  * primary storage virtual memory cache (i.e., "in core").
7285  */
7286 static size_t
7287 segvn_incore(struct seg *seg, caddr_t addr, size_t len, char *vec)
7288 {
7289 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7290 	struct vnode *vp, *avp;
7291 	u_offset_t offset, aoffset;
7292 	size_t p, ep;
7293 	int ret;
7294 	struct vpage *vpp;
7295 	page_t *pp;
7296 	uint_t start;
7297 	struct anon_map *amp;		/* XXX - for locknest */
7298 	struct anon *ap;
7299 	uint_t attr;
7300 	anon_sync_obj_t cookie;
7301 
7302 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7303 
7304 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7305 	if (svd->amp == NULL && svd->vp == NULL) {
7306 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7307 		bzero(vec, btopr(len));
7308 		return (len);	/* no anonymous pages created yet */
7309 	}
7310 
7311 	p = seg_page(seg, addr);
7312 	ep = seg_page(seg, addr + len);
7313 	start = svd->vp ? SEG_PAGE_VNODEBACKED : 0;
7314 
7315 	amp = svd->amp;
7316 	for (; p < ep; p++, addr += PAGESIZE) {
7317 		vpp = (svd->vpage) ? &svd->vpage[p]: NULL;
7318 		ret = start;
7319 		ap = NULL;
7320 		avp = NULL;
7321 		/* Grab the vnode/offset for the anon slot */
7322 		if (amp != NULL) {
7323 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7324 			anon_array_enter(amp, svd->anon_index + p, &cookie);
7325 			ap = anon_get_ptr(amp->ahp, svd->anon_index + p);
7326 			if (ap != NULL) {
7327 				swap_xlate(ap, &avp, &aoffset);
7328 			}
7329 			anon_array_exit(&cookie);
7330 			ANON_LOCK_EXIT(&amp->a_rwlock);
7331 		}
7332 		if ((avp != NULL) && page_exists(avp, aoffset)) {
7333 			/* A page exists for the anon slot */
7334 			ret |= SEG_PAGE_INCORE;
7335 
7336 			/*
7337 			 * If page is mapped and writable
7338 			 */
7339 			attr = (uint_t)0;
7340 			if ((hat_getattr(seg->s_as->a_hat, addr,
7341 			    &attr) != -1) && (attr & PROT_WRITE)) {
7342 				ret |= SEG_PAGE_ANON;
7343 			}
7344 			/*
7345 			 * Don't get page_struct lock for lckcnt and cowcnt,
7346 			 * since this is purely advisory.
7347 			 */
7348 			if ((pp = page_lookup_nowait(avp, aoffset,
7349 			    SE_SHARED)) != NULL) {
7350 				if (pp->p_lckcnt)
7351 					ret |= SEG_PAGE_SOFTLOCK;
7352 				if (pp->p_cowcnt)
7353 					ret |= SEG_PAGE_HASCOW;
7354 				page_unlock(pp);
7355 			}
7356 		}
7357 
7358 		/* Gather vnode statistics */
7359 		vp = svd->vp;
7360 		offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7361 
7362 		if (vp != NULL) {
7363 			/*
7364 			 * Try to obtain a "shared" lock on the page
7365 			 * without blocking.  If this fails, determine
7366 			 * if the page is in memory.
7367 			 */
7368 			pp = page_lookup_nowait(vp, offset, SE_SHARED);
7369 			if ((pp == NULL) && (page_exists(vp, offset))) {
7370 				/* Page is incore, and is named */
7371 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7372 			}
7373 			/*
7374 			 * Don't get page_struct lock for lckcnt and cowcnt,
7375 			 * since this is purely advisory.
7376 			 */
7377 			if (pp != NULL) {
7378 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7379 				if (pp->p_lckcnt)
7380 					ret |= SEG_PAGE_SOFTLOCK;
7381 				if (pp->p_cowcnt)
7382 					ret |= SEG_PAGE_HASCOW;
7383 				page_unlock(pp);
7384 			}
7385 		}
7386 
7387 		/* Gather virtual page information */
7388 		if (vpp) {
7389 			if (VPP_ISPPLOCK(vpp))
7390 				ret |= SEG_PAGE_LOCKED;
7391 			vpp++;
7392 		}
7393 
7394 		*vec++ = (char)ret;
7395 	}
7396 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7397 	return (len);
7398 }
7399 
7400 /*
7401  * Statement for p_cowcnts/p_lckcnts.
7402  *
7403  * p_cowcnt is updated while mlock/munlocking MAP_PRIVATE and PROT_WRITE region
7404  * irrespective of the following factors or anything else:
7405  *
7406  *	(1) anon slots are populated or not
7407  *	(2) cow is broken or not
7408  *	(3) refcnt on ap is 1 or greater than 1
7409  *
7410  * If it's not MAP_PRIVATE and PROT_WRITE, p_lckcnt is updated during mlock
7411  * and munlock.
7412  *
7413  *
7414  * Handling p_cowcnts/p_lckcnts during copy-on-write fault:
7415  *
7416  *	if vpage has PROT_WRITE
7417  *		transfer cowcnt on the oldpage -> cowcnt on the newpage
7418  *	else
7419  *		transfer lckcnt on the oldpage -> lckcnt on the newpage
7420  *
7421  *	During copy-on-write, decrement p_cowcnt on the oldpage and increment
7422  *	p_cowcnt on the newpage *if* the corresponding vpage has PROT_WRITE.
7423  *
7424  *	We may also break COW if softlocking on read access in the physio case.
7425  *	In this case, vpage may not have PROT_WRITE. So, we need to decrement
7426  *	p_lckcnt on the oldpage and increment p_lckcnt on the newpage *if* the
7427  *	vpage doesn't have PROT_WRITE.
7428  *
7429  *
7430  * Handling p_cowcnts/p_lckcnts during mprotect on mlocked region:
7431  *
7432  * 	If a MAP_PRIVATE region loses PROT_WRITE, we decrement p_cowcnt and
7433  *	increment p_lckcnt by calling page_subclaim() which takes care of
7434  * 	availrmem accounting and p_lckcnt overflow.
7435  *
7436  *	If a MAP_PRIVATE region gains PROT_WRITE, we decrement p_lckcnt and
7437  *	increment p_cowcnt by calling page_addclaim() which takes care of
7438  *	availrmem availability and p_cowcnt overflow.
7439  */
7440 
7441 /*
7442  * Lock down (or unlock) pages mapped by this segment.
7443  *
7444  * XXX only creates PAGESIZE pages if anon slots are not initialized.
7445  * At fault time they will be relocated into larger pages.
7446  */
7447 static int
7448 segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
7449     int attr, int op, ulong_t *lockmap, size_t pos)
7450 {
7451 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7452 	struct vpage *vpp;
7453 	struct vpage *evp;
7454 	page_t *pp;
7455 	u_offset_t offset;
7456 	u_offset_t off;
7457 	int segtype;
7458 	int pageprot;
7459 	int claim;
7460 	struct vnode *vp;
7461 	ulong_t anon_index;
7462 	struct anon_map *amp;
7463 	struct anon *ap;
7464 	struct vattr va;
7465 	anon_sync_obj_t cookie;
7466 	struct kshmid *sp = NULL;
7467 	struct proc	*p = curproc;
7468 	kproject_t	*proj = NULL;
7469 	int chargeproc = 1;
7470 	size_t locked_bytes = 0;
7471 	size_t unlocked_bytes = 0;
7472 	int err = 0;
7473 
7474 	/*
7475 	 * Hold write lock on address space because may split or concatenate
7476 	 * segments
7477 	 */
7478 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7479 
7480 	/*
7481 	 * If this is a shm, use shm's project and zone, else use
7482 	 * project and zone of calling process
7483 	 */
7484 
7485 	/* Determine if this segment backs a sysV shm */
7486 	if (svd->amp != NULL && svd->amp->a_sp != NULL) {
7487 		ASSERT(svd->type == MAP_SHARED);
7488 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
7489 		sp = svd->amp->a_sp;
7490 		proj = sp->shm_perm.ipc_proj;
7491 		chargeproc = 0;
7492 	}
7493 
7494 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7495 	if (attr) {
7496 		pageprot = attr & ~(SHARED|PRIVATE);
7497 		segtype = attr & SHARED ? MAP_SHARED : MAP_PRIVATE;
7498 
7499 		/*
7500 		 * We are done if the segment types don't match
7501 		 * or if we have segment level protections and
7502 		 * they don't match.
7503 		 */
7504 		if (svd->type != segtype) {
7505 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7506 			return (0);
7507 		}
7508 		if (svd->pageprot == 0 && svd->prot != pageprot) {
7509 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7510 			return (0);
7511 		}
7512 	}
7513 
7514 	if (op == MC_LOCK) {
7515 		if (svd->tr_state == SEGVN_TR_INIT) {
7516 			svd->tr_state = SEGVN_TR_OFF;
7517 		} else if (svd->tr_state == SEGVN_TR_ON) {
7518 			ASSERT(svd->amp != NULL);
7519 			segvn_textunrepl(seg, 0);
7520 			ASSERT(svd->amp == NULL &&
7521 			    svd->tr_state == SEGVN_TR_OFF);
7522 		}
7523 	}
7524 
7525 	/*
7526 	 * If we're locking, then we must create a vpage structure if
7527 	 * none exists.  If we're unlocking, then check to see if there
7528 	 * is a vpage --  if not, then we could not have locked anything.
7529 	 */
7530 
7531 	if ((vpp = svd->vpage) == NULL) {
7532 		if (op == MC_LOCK)
7533 			segvn_vpage(seg);
7534 		else {
7535 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7536 			return (0);
7537 		}
7538 	}
7539 
7540 	/*
7541 	 * The anonymous data vector (i.e., previously
7542 	 * unreferenced mapping to swap space) can be allocated
7543 	 * by lazily testing for its existence.
7544 	 */
7545 	if (op == MC_LOCK && svd->amp == NULL && svd->vp == NULL) {
7546 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
7547 		svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
7548 		svd->amp->a_szc = seg->s_szc;
7549 	}
7550 
7551 	if ((amp = svd->amp) != NULL) {
7552 		anon_index = svd->anon_index + seg_page(seg, addr);
7553 	}
7554 
7555 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7556 	evp = &svd->vpage[seg_page(seg, addr + len)];
7557 
7558 	if (sp != NULL)
7559 		mutex_enter(&sp->shm_mlock);
7560 
7561 	/* determine number of unlocked bytes in range for lock operation */
7562 	if (op == MC_LOCK) {
7563 
7564 		if (sp == NULL) {
7565 			for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7566 			    vpp++) {
7567 				if (!VPP_ISPPLOCK(vpp))
7568 					unlocked_bytes += PAGESIZE;
7569 			}
7570 		} else {
7571 			ulong_t		i_idx, i_edx;
7572 			anon_sync_obj_t	i_cookie;
7573 			struct anon	*i_ap;
7574 			struct vnode	*i_vp;
7575 			u_offset_t	i_off;
7576 
7577 			/* Only count sysV pages once for locked memory */
7578 			i_edx = svd->anon_index + seg_page(seg, addr + len);
7579 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7580 			for (i_idx = anon_index; i_idx < i_edx; i_idx++) {
7581 				anon_array_enter(amp, i_idx, &i_cookie);
7582 				i_ap = anon_get_ptr(amp->ahp, i_idx);
7583 				if (i_ap == NULL) {
7584 					unlocked_bytes += PAGESIZE;
7585 					anon_array_exit(&i_cookie);
7586 					continue;
7587 				}
7588 				swap_xlate(i_ap, &i_vp, &i_off);
7589 				anon_array_exit(&i_cookie);
7590 				pp = page_lookup(i_vp, i_off, SE_SHARED);
7591 				if (pp == NULL) {
7592 					unlocked_bytes += PAGESIZE;
7593 					continue;
7594 				} else if (pp->p_lckcnt == 0)
7595 					unlocked_bytes += PAGESIZE;
7596 				page_unlock(pp);
7597 			}
7598 			ANON_LOCK_EXIT(&amp->a_rwlock);
7599 		}
7600 
7601 		mutex_enter(&p->p_lock);
7602 		err = rctl_incr_locked_mem(p, proj, unlocked_bytes,
7603 		    chargeproc);
7604 		mutex_exit(&p->p_lock);
7605 
7606 		if (err) {
7607 			if (sp != NULL)
7608 				mutex_exit(&sp->shm_mlock);
7609 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7610 			return (err);
7611 		}
7612 	}
7613 	/*
7614 	 * Loop over all pages in the range.  Process if we're locking and
7615 	 * page has not already been locked in this mapping; or if we're
7616 	 * unlocking and the page has been locked.
7617 	 */
7618 	for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7619 	    vpp++, pos++, addr += PAGESIZE, offset += PAGESIZE, anon_index++) {
7620 		if ((attr == 0 || VPP_PROT(vpp) == pageprot) &&
7621 		    ((op == MC_LOCK && !VPP_ISPPLOCK(vpp)) ||
7622 		    (op == MC_UNLOCK && VPP_ISPPLOCK(vpp)))) {
7623 
7624 			if (amp != NULL)
7625 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7626 			/*
7627 			 * If this isn't a MAP_NORESERVE segment and
7628 			 * we're locking, allocate anon slots if they
7629 			 * don't exist.  The page is brought in later on.
7630 			 */
7631 			if (op == MC_LOCK && svd->vp == NULL &&
7632 			    ((svd->flags & MAP_NORESERVE) == 0) &&
7633 			    amp != NULL &&
7634 			    ((ap = anon_get_ptr(amp->ahp, anon_index))
7635 								== NULL)) {
7636 				anon_array_enter(amp, anon_index, &cookie);
7637 
7638 				if ((ap = anon_get_ptr(amp->ahp,
7639 						anon_index)) == NULL) {
7640 					pp = anon_zero(seg, addr, &ap,
7641 					    svd->cred);
7642 					if (pp == NULL) {
7643 						anon_array_exit(&cookie);
7644 						ANON_LOCK_EXIT(&amp->a_rwlock);
7645 						err = ENOMEM;
7646 						goto out;
7647 					}
7648 					ASSERT(anon_get_ptr(amp->ahp,
7649 						anon_index) == NULL);
7650 					(void) anon_set_ptr(amp->ahp,
7651 						anon_index, ap, ANON_SLEEP);
7652 					page_unlock(pp);
7653 				}
7654 				anon_array_exit(&cookie);
7655 			}
7656 
7657 			/*
7658 			 * Get name for page, accounting for
7659 			 * existence of private copy.
7660 			 */
7661 			ap = NULL;
7662 			if (amp != NULL) {
7663 				anon_array_enter(amp, anon_index, &cookie);
7664 				ap = anon_get_ptr(amp->ahp, anon_index);
7665 				if (ap != NULL) {
7666 					swap_xlate(ap, &vp, &off);
7667 				} else {
7668 					if (svd->vp == NULL &&
7669 					    (svd->flags & MAP_NORESERVE)) {
7670 						anon_array_exit(&cookie);
7671 						ANON_LOCK_EXIT(&amp->a_rwlock);
7672 						continue;
7673 					}
7674 					vp = svd->vp;
7675 					off = offset;
7676 				}
7677 				anon_array_exit(&cookie);
7678 				ANON_LOCK_EXIT(&amp->a_rwlock);
7679 			} else {
7680 				vp = svd->vp;
7681 				off = offset;
7682 			}
7683 
7684 			/*
7685 			 * Get page frame.  It's ok if the page is
7686 			 * not available when we're unlocking, as this
7687 			 * may simply mean that a page we locked got
7688 			 * truncated out of existence after we locked it.
7689 			 *
7690 			 * Invoke VOP_GETPAGE() to obtain the page struct
7691 			 * since we may need to read it from disk if its
7692 			 * been paged out.
7693 			 */
7694 			if (op != MC_LOCK)
7695 				pp = page_lookup(vp, off, SE_SHARED);
7696 			else {
7697 				page_t *pl[1 + 1];
7698 				int error;
7699 
7700 				ASSERT(vp != NULL);
7701 
7702 				error = VOP_GETPAGE(vp, (offset_t)off, PAGESIZE,
7703 				    (uint_t *)NULL, pl, PAGESIZE, seg, addr,
7704 				    S_OTHER, svd->cred);
7705 
7706 				/*
7707 				 * If the error is EDEADLK then we must bounce
7708 				 * up and drop all vm subsystem locks and then
7709 				 * retry the operation later
7710 				 * This behavior is a temporary measure because
7711 				 * ufs/sds logging is badly designed and will
7712 				 * deadlock if we don't allow this bounce to
7713 				 * happen.  The real solution is to re-design
7714 				 * the logging code to work properly.  See bug
7715 				 * 4125102 for details of the problem.
7716 				 */
7717 				if (error == EDEADLK) {
7718 					err = error;
7719 					goto out;
7720 				}
7721 				/*
7722 				 * Quit if we fail to fault in the page.  Treat
7723 				 * the failure as an error, unless the addr
7724 				 * is mapped beyond the end of a file.
7725 				 */
7726 				if (error && svd->vp) {
7727 					va.va_mask = AT_SIZE;
7728 					if (VOP_GETATTR(svd->vp, &va, 0,
7729 					    svd->cred) != 0) {
7730 						err = EIO;
7731 						goto out;
7732 					}
7733 					if (btopr(va.va_size) >=
7734 					    btopr(off + 1)) {
7735 						err = EIO;
7736 						goto out;
7737 					}
7738 					goto out;
7739 
7740 				} else if (error) {
7741 					err = EIO;
7742 					goto out;
7743 				}
7744 				pp = pl[0];
7745 				ASSERT(pp != NULL);
7746 			}
7747 
7748 			/*
7749 			 * See Statement at the beginning of this routine.
7750 			 *
7751 			 * claim is always set if MAP_PRIVATE and PROT_WRITE
7752 			 * irrespective of following factors:
7753 			 *
7754 			 * (1) anon slots are populated or not
7755 			 * (2) cow is broken or not
7756 			 * (3) refcnt on ap is 1 or greater than 1
7757 			 *
7758 			 * See 4140683 for details
7759 			 */
7760 			claim = ((VPP_PROT(vpp) & PROT_WRITE) &&
7761 				(svd->type == MAP_PRIVATE));
7762 
7763 			/*
7764 			 * Perform page-level operation appropriate to
7765 			 * operation.  If locking, undo the SOFTLOCK
7766 			 * performed to bring the page into memory
7767 			 * after setting the lock.  If unlocking,
7768 			 * and no page was found, account for the claim
7769 			 * separately.
7770 			 */
7771 			if (op == MC_LOCK) {
7772 				int ret = 1;	/* Assume success */
7773 
7774 				ASSERT(!VPP_ISPPLOCK(vpp));
7775 
7776 				ret = page_pp_lock(pp, claim, 0);
7777 				if (ret == 0) {
7778 					/* locking page failed */
7779 					page_unlock(pp);
7780 					err = EAGAIN;
7781 					goto out;
7782 				}
7783 				VPP_SETPPLOCK(vpp);
7784 				if (sp != NULL) {
7785 					if (pp->p_lckcnt == 1)
7786 						locked_bytes += PAGESIZE;
7787 				} else
7788 					locked_bytes += PAGESIZE;
7789 
7790 				if (lockmap != (ulong_t *)NULL)
7791 					BT_SET(lockmap, pos);
7792 
7793 				page_unlock(pp);
7794 			} else {
7795 				ASSERT(VPP_ISPPLOCK(vpp));
7796 				if (pp != NULL) {
7797 					/* sysV pages should be locked */
7798 					ASSERT(sp == NULL || pp->p_lckcnt > 0);
7799 					page_pp_unlock(pp, claim, 0);
7800 					if (sp != NULL) {
7801 						if (pp->p_lckcnt == 0)
7802 							unlocked_bytes
7803 							    += PAGESIZE;
7804 					} else
7805 						unlocked_bytes += PAGESIZE;
7806 					page_unlock(pp);
7807 				} else {
7808 					ASSERT(sp == NULL);
7809 					unlocked_bytes += PAGESIZE;
7810 				}
7811 				VPP_CLRPPLOCK(vpp);
7812 			}
7813 		}
7814 	}
7815 out:
7816 	if (op == MC_LOCK) {
7817 		/* Credit back bytes that did not get locked */
7818 		if ((unlocked_bytes - locked_bytes) > 0) {
7819 			if (proj == NULL)
7820 				mutex_enter(&p->p_lock);
7821 			rctl_decr_locked_mem(p, proj,
7822 			    (unlocked_bytes - locked_bytes), chargeproc);
7823 			if (proj == NULL)
7824 				mutex_exit(&p->p_lock);
7825 		}
7826 
7827 	} else {
7828 		/* Account bytes that were unlocked */
7829 		if (unlocked_bytes > 0) {
7830 			if (proj == NULL)
7831 				mutex_enter(&p->p_lock);
7832 			rctl_decr_locked_mem(p, proj, unlocked_bytes,
7833 			    chargeproc);
7834 			if (proj == NULL)
7835 				mutex_exit(&p->p_lock);
7836 		}
7837 	}
7838 	if (sp != NULL)
7839 		mutex_exit(&sp->shm_mlock);
7840 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7841 
7842 	return (err);
7843 }
7844 
7845 /*
7846  * Set advice from user for specified pages
7847  * There are 5 types of advice:
7848  *	MADV_NORMAL	- Normal (default) behavior (whatever that is)
7849  *	MADV_RANDOM	- Random page references
7850  *				do not allow readahead or 'klustering'
7851  *	MADV_SEQUENTIAL	- Sequential page references
7852  *				Pages previous to the one currently being
7853  *				accessed (determined by fault) are 'not needed'
7854  *				and are freed immediately
7855  *	MADV_WILLNEED	- Pages are likely to be used (fault ahead in mctl)
7856  *	MADV_DONTNEED	- Pages are not needed (synced out in mctl)
7857  *	MADV_FREE	- Contents can be discarded
7858  *	MADV_ACCESS_DEFAULT- Default access
7859  *	MADV_ACCESS_LWP	- Next LWP will access heavily
7860  *	MADV_ACCESS_MANY- Many LWPs or processes will access heavily
7861  */
7862 static int
7863 segvn_advise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
7864 {
7865 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7866 	size_t page;
7867 	int err = 0;
7868 	int already_set;
7869 	struct anon_map *amp;
7870 	ulong_t anon_index;
7871 	struct seg *next;
7872 	lgrp_mem_policy_t policy;
7873 	struct seg *prev;
7874 	struct vnode *vp;
7875 
7876 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7877 
7878 	/*
7879 	 * In case of MADV_FREE, we won't be modifying any segment private
7880 	 * data structures; so, we only need to grab READER's lock
7881 	 */
7882 	if (behav != MADV_FREE) {
7883 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7884 		if (svd->tr_state != SEGVN_TR_OFF) {
7885 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7886 			return (0);
7887 		}
7888 	} else {
7889 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7890 	}
7891 
7892 	/*
7893 	 * Large pages are assumed to be only turned on when accesses to the
7894 	 * segment's address range have spatial and temporal locality. That
7895 	 * justifies ignoring MADV_SEQUENTIAL for large page segments.
7896 	 * Also, ignore advice affecting lgroup memory allocation
7897 	 * if don't need to do lgroup optimizations on this system
7898 	 */
7899 
7900 	if ((behav == MADV_SEQUENTIAL &&
7901 	    (seg->s_szc != 0 || HAT_IS_REGION_COOKIE_VALID(svd->rcookie))) ||
7902 	    (!lgrp_optimizations() && (behav == MADV_ACCESS_DEFAULT ||
7903 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY))) {
7904 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7905 		return (0);
7906 	}
7907 
7908 	if (behav == MADV_SEQUENTIAL || behav == MADV_ACCESS_DEFAULT ||
7909 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY) {
7910 		/*
7911 		 * Since we are going to unload hat mappings
7912 		 * we first have to flush the cache. Otherwise
7913 		 * this might lead to system panic if another
7914 		 * thread is doing physio on the range whose
7915 		 * mappings are unloaded by madvise(3C).
7916 		 */
7917 		if (svd->softlockcnt > 0) {
7918 			/*
7919 			 * Since we do have the segvn writers lock
7920 			 * nobody can fill the cache with entries
7921 			 * belonging to this seg during the purge.
7922 			 * The flush either succeeds or we still
7923 			 * have pending I/Os. In the later case,
7924 			 * madvise(3C) fails.
7925 			 */
7926 			segvn_purge(seg);
7927 			if (svd->softlockcnt > 0) {
7928 				/*
7929 				 * Since madvise(3C) is advisory and
7930 				 * it's not part of UNIX98, madvise(3C)
7931 				 * failure here doesn't cause any hardship.
7932 				 * Note that we don't block in "as" layer.
7933 				 */
7934 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7935 				return (EAGAIN);
7936 			}
7937 		}
7938 	}
7939 
7940 	amp = svd->amp;
7941 	vp = svd->vp;
7942 	if (behav == MADV_FREE) {
7943 		/*
7944 		 * MADV_FREE is not supported for segments with
7945 		 * underlying object; if anonmap is NULL, anon slots
7946 		 * are not yet populated and there is nothing for
7947 		 * us to do. As MADV_FREE is advisory, we don't
7948 		 * return error in either case.
7949 		 */
7950 		if (vp != NULL || amp == NULL) {
7951 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7952 			return (0);
7953 		}
7954 
7955 		page = seg_page(seg, addr);
7956 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7957 		anon_disclaim(amp, svd->anon_index + page, len, 0);
7958 		ANON_LOCK_EXIT(&amp->a_rwlock);
7959 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7960 		return (0);
7961 	}
7962 
7963 	/*
7964 	 * If advice is to be applied to entire segment,
7965 	 * use advice field in seg_data structure
7966 	 * otherwise use appropriate vpage entry.
7967 	 */
7968 	if ((addr == seg->s_base) && (len == seg->s_size)) {
7969 		switch (behav) {
7970 		case MADV_ACCESS_LWP:
7971 		case MADV_ACCESS_MANY:
7972 		case MADV_ACCESS_DEFAULT:
7973 			/*
7974 			 * Set memory allocation policy for this segment
7975 			 */
7976 			policy = lgrp_madv_to_policy(behav, len, svd->type);
7977 			if (svd->type == MAP_SHARED)
7978 				already_set = lgrp_shm_policy_set(policy, amp,
7979 				    svd->anon_index, vp, svd->offset, len);
7980 			else {
7981 				/*
7982 				 * For private memory, need writers lock on
7983 				 * address space because the segment may be
7984 				 * split or concatenated when changing policy
7985 				 */
7986 				if (AS_READ_HELD(seg->s_as,
7987 				    &seg->s_as->a_lock)) {
7988 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7989 					return (IE_RETRY);
7990 				}
7991 
7992 				already_set = lgrp_privm_policy_set(policy,
7993 				    &svd->policy_info, len);
7994 			}
7995 
7996 			/*
7997 			 * If policy set already and it shouldn't be reapplied,
7998 			 * don't do anything.
7999 			 */
8000 			if (already_set &&
8001 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8002 				break;
8003 
8004 			/*
8005 			 * Mark any existing pages in given range for
8006 			 * migration
8007 			 */
8008 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8009 			    vp, svd->offset, 1);
8010 
8011 			/*
8012 			 * If same policy set already or this is a shared
8013 			 * memory segment, don't need to try to concatenate
8014 			 * segment with adjacent ones.
8015 			 */
8016 			if (already_set || svd->type == MAP_SHARED)
8017 				break;
8018 
8019 			/*
8020 			 * Try to concatenate this segment with previous
8021 			 * one and next one, since we changed policy for
8022 			 * this one and it may be compatible with adjacent
8023 			 * ones now.
8024 			 */
8025 			prev = AS_SEGPREV(seg->s_as, seg);
8026 			next = AS_SEGNEXT(seg->s_as, seg);
8027 
8028 			if (next && next->s_ops == &segvn_ops &&
8029 			    addr + len == next->s_base)
8030 				(void) segvn_concat(seg, next, 1);
8031 
8032 			if (prev && prev->s_ops == &segvn_ops &&
8033 			    addr == prev->s_base + prev->s_size) {
8034 				/*
8035 				 * Drop lock for private data of current
8036 				 * segment before concatenating (deleting) it
8037 				 * and return IE_REATTACH to tell as_ctl() that
8038 				 * current segment has changed
8039 				 */
8040 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8041 				if (!segvn_concat(prev, seg, 1))
8042 					err = IE_REATTACH;
8043 
8044 				return (err);
8045 			}
8046 			break;
8047 
8048 		case MADV_SEQUENTIAL:
8049 			/*
8050 			 * unloading mapping guarantees
8051 			 * detection in segvn_fault
8052 			 */
8053 			ASSERT(seg->s_szc == 0);
8054 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8055 			hat_unload(seg->s_as->a_hat, addr, len,
8056 				HAT_UNLOAD);
8057 			/* FALLTHROUGH */
8058 		case MADV_NORMAL:
8059 		case MADV_RANDOM:
8060 			svd->advice = (uchar_t)behav;
8061 			svd->pageadvice = 0;
8062 			break;
8063 		case MADV_WILLNEED:	/* handled in memcntl */
8064 		case MADV_DONTNEED:	/* handled in memcntl */
8065 		case MADV_FREE:		/* handled above */
8066 			break;
8067 		default:
8068 			err = EINVAL;
8069 		}
8070 	} else {
8071 		caddr_t			eaddr;
8072 		struct seg		*new_seg;
8073 		struct segvn_data	*new_svd;
8074 		u_offset_t		off;
8075 		caddr_t			oldeaddr;
8076 
8077 		page = seg_page(seg, addr);
8078 
8079 		segvn_vpage(seg);
8080 
8081 		switch (behav) {
8082 			struct vpage *bvpp, *evpp;
8083 
8084 		case MADV_ACCESS_LWP:
8085 		case MADV_ACCESS_MANY:
8086 		case MADV_ACCESS_DEFAULT:
8087 			/*
8088 			 * Set memory allocation policy for portion of this
8089 			 * segment
8090 			 */
8091 
8092 			/*
8093 			 * Align address and length of advice to page
8094 			 * boundaries for large pages
8095 			 */
8096 			if (seg->s_szc != 0) {
8097 				size_t	pgsz;
8098 
8099 				pgsz = page_get_pagesize(seg->s_szc);
8100 				addr = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
8101 				len = P2ROUNDUP(len, pgsz);
8102 			}
8103 
8104 			/*
8105 			 * Check to see whether policy is set already
8106 			 */
8107 			policy = lgrp_madv_to_policy(behav, len, svd->type);
8108 
8109 			anon_index = svd->anon_index + page;
8110 			off = svd->offset + (uintptr_t)(addr - seg->s_base);
8111 
8112 			if (svd->type == MAP_SHARED)
8113 				already_set = lgrp_shm_policy_set(policy, amp,
8114 				    anon_index, vp, off, len);
8115 			else
8116 				already_set =
8117 				    (policy == svd->policy_info.mem_policy);
8118 
8119 			/*
8120 			 * If policy set already and it shouldn't be reapplied,
8121 			 * don't do anything.
8122 			 */
8123 			if (already_set &&
8124 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8125 				break;
8126 
8127 			/*
8128 			 * For private memory, need writers lock on
8129 			 * address space because the segment may be
8130 			 * split or concatenated when changing policy
8131 			 */
8132 			if (svd->type == MAP_PRIVATE &&
8133 			    AS_READ_HELD(seg->s_as, &seg->s_as->a_lock)) {
8134 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8135 				return (IE_RETRY);
8136 			}
8137 
8138 			/*
8139 			 * Mark any existing pages in given range for
8140 			 * migration
8141 			 */
8142 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8143 			    vp, svd->offset, 1);
8144 
8145 			/*
8146 			 * Don't need to try to split or concatenate
8147 			 * segments, since policy is same or this is a shared
8148 			 * memory segment
8149 			 */
8150 			if (already_set || svd->type == MAP_SHARED)
8151 				break;
8152 
8153 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
8154 				ASSERT(svd->amp == NULL);
8155 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
8156 				ASSERT(svd->softlockcnt == 0);
8157 				hat_leave_region(seg->s_as->a_hat, svd->rcookie,
8158 				    HAT_REGION_TEXT);
8159 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
8160 			}
8161 
8162 			/*
8163 			 * Split off new segment if advice only applies to a
8164 			 * portion of existing segment starting in middle
8165 			 */
8166 			new_seg = NULL;
8167 			eaddr = addr + len;
8168 			oldeaddr = seg->s_base + seg->s_size;
8169 			if (addr > seg->s_base) {
8170 				/*
8171 				 * Must flush I/O page cache
8172 				 * before splitting segment
8173 				 */
8174 				if (svd->softlockcnt > 0)
8175 					segvn_purge(seg);
8176 
8177 				/*
8178 				 * Split segment and return IE_REATTACH to tell
8179 				 * as_ctl() that current segment changed
8180 				 */
8181 				new_seg = segvn_split_seg(seg, addr);
8182 				new_svd = (struct segvn_data *)new_seg->s_data;
8183 				err = IE_REATTACH;
8184 
8185 				/*
8186 				 * If new segment ends where old one
8187 				 * did, try to concatenate the new
8188 				 * segment with next one.
8189 				 */
8190 				if (eaddr == oldeaddr) {
8191 					/*
8192 					 * Set policy for new segment
8193 					 */
8194 					(void) lgrp_privm_policy_set(policy,
8195 					    &new_svd->policy_info,
8196 					    new_seg->s_size);
8197 
8198 					next = AS_SEGNEXT(new_seg->s_as,
8199 					    new_seg);
8200 
8201 					if (next &&
8202 					    next->s_ops == &segvn_ops &&
8203 					    eaddr == next->s_base)
8204 						(void) segvn_concat(new_seg,
8205 						    next, 1);
8206 				}
8207 			}
8208 
8209 			/*
8210 			 * Split off end of existing segment if advice only
8211 			 * applies to a portion of segment ending before
8212 			 * end of the existing segment
8213 			 */
8214 			if (eaddr < oldeaddr) {
8215 				/*
8216 				 * Must flush I/O page cache
8217 				 * before splitting segment
8218 				 */
8219 				if (svd->softlockcnt > 0)
8220 					segvn_purge(seg);
8221 
8222 				/*
8223 				 * If beginning of old segment was already
8224 				 * split off, use new segment to split end off
8225 				 * from.
8226 				 */
8227 				if (new_seg != NULL && new_seg != seg) {
8228 					/*
8229 					 * Split segment
8230 					 */
8231 					(void) segvn_split_seg(new_seg, eaddr);
8232 
8233 					/*
8234 					 * Set policy for new segment
8235 					 */
8236 					(void) lgrp_privm_policy_set(policy,
8237 					    &new_svd->policy_info,
8238 					    new_seg->s_size);
8239 				} else {
8240 					/*
8241 					 * Split segment and return IE_REATTACH
8242 					 * to tell as_ctl() that current
8243 					 * segment changed
8244 					 */
8245 					(void) segvn_split_seg(seg, eaddr);
8246 					err = IE_REATTACH;
8247 
8248 					(void) lgrp_privm_policy_set(policy,
8249 					    &svd->policy_info, seg->s_size);
8250 
8251 					/*
8252 					 * If new segment starts where old one
8253 					 * did, try to concatenate it with
8254 					 * previous segment.
8255 					 */
8256 					if (addr == seg->s_base) {
8257 						prev = AS_SEGPREV(seg->s_as,
8258 						    seg);
8259 
8260 						/*
8261 						 * Drop lock for private data
8262 						 * of current segment before
8263 						 * concatenating (deleting) it
8264 						 */
8265 						if (prev &&
8266 						    prev->s_ops ==
8267 						    &segvn_ops &&
8268 						    addr == prev->s_base +
8269 						    prev->s_size) {
8270 							SEGVN_LOCK_EXIT(
8271 							    seg->s_as,
8272 							    &svd->lock);
8273 							(void) segvn_concat(
8274 							    prev, seg, 1);
8275 							return (err);
8276 						}
8277 					}
8278 				}
8279 			}
8280 			break;
8281 		case MADV_SEQUENTIAL:
8282 			ASSERT(seg->s_szc == 0);
8283 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8284 			hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
8285 			/* FALLTHROUGH */
8286 		case MADV_NORMAL:
8287 		case MADV_RANDOM:
8288 			bvpp = &svd->vpage[page];
8289 			evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8290 			for (; bvpp < evpp; bvpp++)
8291 				VPP_SETADVICE(bvpp, behav);
8292 			svd->advice = MADV_NORMAL;
8293 			break;
8294 		case MADV_WILLNEED:	/* handled in memcntl */
8295 		case MADV_DONTNEED:	/* handled in memcntl */
8296 		case MADV_FREE:		/* handled above */
8297 			break;
8298 		default:
8299 			err = EINVAL;
8300 		}
8301 	}
8302 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8303 	return (err);
8304 }
8305 
8306 /*
8307  * Create a vpage structure for this seg.
8308  */
8309 static void
8310 segvn_vpage(struct seg *seg)
8311 {
8312 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8313 	struct vpage *vp, *evp;
8314 
8315 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
8316 
8317 	/*
8318 	 * If no vpage structure exists, allocate one.  Copy the protections
8319 	 * and the advice from the segment itself to the individual pages.
8320 	 */
8321 	if (svd->vpage == NULL) {
8322 		svd->pageadvice = 1;
8323 		svd->vpage = kmem_zalloc(seg_pages(seg) * sizeof (struct vpage),
8324 		    KM_SLEEP);
8325 		evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
8326 		for (vp = svd->vpage; vp < evp; vp++) {
8327 			VPP_SETPROT(vp, svd->prot);
8328 			VPP_SETADVICE(vp, svd->advice);
8329 		}
8330 	}
8331 }
8332 
8333 /*
8334  * Dump the pages belonging to this segvn segment.
8335  */
8336 static void
8337 segvn_dump(struct seg *seg)
8338 {
8339 	struct segvn_data *svd;
8340 	page_t *pp;
8341 	struct anon_map *amp;
8342 	ulong_t	anon_index;
8343 	struct vnode *vp;
8344 	u_offset_t off, offset;
8345 	pfn_t pfn;
8346 	pgcnt_t page, npages;
8347 	caddr_t addr;
8348 
8349 	npages = seg_pages(seg);
8350 	svd = (struct segvn_data *)seg->s_data;
8351 	vp = svd->vp;
8352 	off = offset = svd->offset;
8353 	addr = seg->s_base;
8354 
8355 	if ((amp = svd->amp) != NULL) {
8356 		anon_index = svd->anon_index;
8357 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8358 	}
8359 
8360 	for (page = 0; page < npages; page++, offset += PAGESIZE) {
8361 		struct anon *ap;
8362 		int we_own_it = 0;
8363 
8364 		if (amp && (ap = anon_get_ptr(svd->amp->ahp, anon_index++))) {
8365 			swap_xlate_nopanic(ap, &vp, &off);
8366 		} else {
8367 			vp = svd->vp;
8368 			off = offset;
8369 		}
8370 
8371 		/*
8372 		 * If pp == NULL, the page either does not exist
8373 		 * or is exclusively locked.  So determine if it
8374 		 * exists before searching for it.
8375 		 */
8376 
8377 		if ((pp = page_lookup_nowait(vp, off, SE_SHARED)))
8378 			we_own_it = 1;
8379 		else
8380 			pp = page_exists(vp, off);
8381 
8382 		if (pp) {
8383 			pfn = page_pptonum(pp);
8384 			dump_addpage(seg->s_as, addr, pfn);
8385 			if (we_own_it)
8386 				page_unlock(pp);
8387 		}
8388 		addr += PAGESIZE;
8389 		dump_timeleft = dump_timeout;
8390 	}
8391 
8392 	if (amp != NULL)
8393 		ANON_LOCK_EXIT(&amp->a_rwlock);
8394 }
8395 
8396 /*
8397  * lock/unlock anon pages over a given range. Return shadow list
8398  */
8399 static int
8400 segvn_pagelock(struct seg *seg, caddr_t addr, size_t len, struct page ***ppp,
8401     enum lock_type type, enum seg_rw rw)
8402 {
8403 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8404 	size_t np, adjustpages = 0, npages = (len >> PAGESHIFT);
8405 	ulong_t anon_index;
8406 	uint_t protchk;
8407 	uint_t error;
8408 	struct anon_map *amp;
8409 	struct page **pplist, **pl, *pp;
8410 	caddr_t a;
8411 	size_t page;
8412 	caddr_t lpgaddr, lpgeaddr;
8413 	pgcnt_t szc0_npages = 0;
8414 
8415 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_START,
8416 		"segvn_pagelock: start seg %p addr %p", seg, addr);
8417 
8418 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
8419 	if (seg->s_szc != 0 && (type == L_PAGELOCK || type == L_PAGEUNLOCK)) {
8420 		/*
8421 		 * We are adjusting the pagelock region to the large page size
8422 		 * boundary because the unlocked part of a large page cannot
8423 		 * be freed anyway unless all constituent pages of a large
8424 		 * page are locked. Therefore this adjustment allows us to
8425 		 * decrement availrmem by the right value (note we don't want
8426 		 * to just decrement availrem by the large page size without
8427 		 * adjusting addr and len because then we may end up
8428 		 * decrementing availrmem by large page size for every
8429 		 * constituent page locked by a new as_pagelock call).
8430 		 * as_pageunlock caller must always match as_pagelock call's
8431 		 * addr and len.
8432 		 *
8433 		 * Note segment's page size cannot change while we are holding
8434 		 * as lock.  And then it cannot change while softlockcnt is
8435 		 * not 0. This will allow us to correctly recalculate large
8436 		 * page size region for the matching pageunlock/reclaim call.
8437 		 *
8438 		 * for pageunlock *ppp points to the pointer of page_t that
8439 		 * corresponds to the real unadjusted start address. Similar
8440 		 * for pagelock *ppp must point to the pointer of page_t that
8441 		 * corresponds to the real unadjusted start address.
8442 		 */
8443 		size_t pgsz = page_get_pagesize(seg->s_szc);
8444 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
8445 		adjustpages = ((uintptr_t)(addr - lpgaddr)) >> PAGESHIFT;
8446 	}
8447 
8448 	if (type == L_PAGEUNLOCK) {
8449 
8450 		/*
8451 		 * update hat ref bits for /proc. We need to make sure
8452 		 * that threads tracing the ref and mod bits of the
8453 		 * address space get the right data.
8454 		 * Note: page ref and mod bits are updated at reclaim time
8455 		 */
8456 		if (seg->s_as->a_vbits) {
8457 			for (a = addr; a < addr + len; a += PAGESIZE) {
8458 				if (rw == S_WRITE) {
8459 					hat_setstat(seg->s_as, a,
8460 					    PAGESIZE, P_REF | P_MOD);
8461 				} else {
8462 					hat_setstat(seg->s_as, a,
8463 					    PAGESIZE, P_REF);
8464 				}
8465 			}
8466 		}
8467 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8468 		if (seg->s_szc != 0) {
8469 			VM_STAT_ADD(segvnvmstats.pagelock[0]);
8470 			seg_pinactive(seg, lpgaddr, lpgeaddr - lpgaddr,
8471 			    *ppp - adjustpages, rw, segvn_reclaim);
8472 		} else {
8473 			seg_pinactive(seg, addr, len, *ppp, rw, segvn_reclaim);
8474 		}
8475 
8476 		/*
8477 		 * If someone is blocked while unmapping, we purge
8478 		 * segment page cache and thus reclaim pplist synchronously
8479 		 * without waiting for seg_pasync_thread. This speeds up
8480 		 * unmapping in cases where munmap(2) is called, while
8481 		 * raw async i/o is still in progress or where a thread
8482 		 * exits on data fault in a multithreaded application.
8483 		 */
8484 		if (AS_ISUNMAPWAIT(seg->s_as) && (svd->softlockcnt > 0)) {
8485 			/*
8486 			 * Even if we grab segvn WRITER's lock or segp_slock
8487 			 * here, there might be another thread which could've
8488 			 * successfully performed lookup/insert just before
8489 			 * we acquired the lock here.  So, grabbing either
8490 			 * lock here is of not much use.  Until we devise
8491 			 * a strategy at upper layers to solve the
8492 			 * synchronization issues completely, we expect
8493 			 * applications to handle this appropriately.
8494 			 */
8495 			segvn_purge(seg);
8496 		}
8497 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8498 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_UNLOCK_END,
8499 			"segvn_pagelock: unlock seg %p addr %p", seg, addr);
8500 		return (0);
8501 	} else if (type == L_PAGERECLAIM) {
8502 		VM_STAT_COND_ADD(seg->s_szc != 0, segvnvmstats.pagelock[1]);
8503 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8504 		(void) segvn_reclaim(seg, addr, len, *ppp, rw);
8505 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8506 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_UNLOCK_END,
8507 			"segvn_pagelock: reclaim seg %p addr %p", seg, addr);
8508 		return (0);
8509 	}
8510 
8511 	if (seg->s_szc != 0) {
8512 		VM_STAT_ADD(segvnvmstats.pagelock[2]);
8513 		addr = lpgaddr;
8514 		len = lpgeaddr - lpgaddr;
8515 		npages = (len >> PAGESHIFT);
8516 	}
8517 
8518 	/*
8519 	 * for now we only support pagelock to anon memory. We've to check
8520 	 * protections for vnode objects and call into the vnode driver.
8521 	 * That's too much for a fast path. Let the fault entry point handle it.
8522 	 */
8523 	if (svd->vp != NULL) {
8524 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
8525 		    "segvn_pagelock: mapped vnode seg %p addr %p", seg, addr);
8526 		*ppp = NULL;
8527 		return (ENOTSUP);
8528 	}
8529 
8530 	/*
8531 	 * if anonmap is not yet created, let the fault entry point populate it
8532 	 * with anon ptrs.
8533 	 */
8534 	if ((amp = svd->amp) == NULL) {
8535 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
8536 		    "segvn_pagelock: anonmap null seg %p addr %p", seg, addr);
8537 		*ppp = NULL;
8538 		return (EFAULT);
8539 	}
8540 
8541 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8542 
8543 	/*
8544 	 * we acquire segp_slock to prevent duplicate entries
8545 	 * in seg_pcache
8546 	 */
8547 	mutex_enter(&svd->segp_slock);
8548 
8549 	/*
8550 	 * try to find pages in segment page cache
8551 	 */
8552 	pplist = seg_plookup(seg, addr, len, rw);
8553 	if (pplist != NULL) {
8554 		mutex_exit(&svd->segp_slock);
8555 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8556 		*ppp = pplist + adjustpages;
8557 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_HIT_END,
8558 			"segvn_pagelock: cache hit seg %p addr %p", seg, addr);
8559 		return (0);
8560 	}
8561 
8562 	if (rw == S_READ) {
8563 		protchk = PROT_READ;
8564 	} else {
8565 		protchk = PROT_WRITE;
8566 	}
8567 
8568 	if (svd->pageprot == 0) {
8569 		if ((svd->prot & protchk) == 0) {
8570 			mutex_exit(&svd->segp_slock);
8571 			error = EFAULT;
8572 			goto out;
8573 		}
8574 	} else {
8575 		/*
8576 		 * check page protections
8577 		 */
8578 		for (a = addr; a < addr + len; a += PAGESIZE) {
8579 			struct vpage *vp;
8580 
8581 			vp = &svd->vpage[seg_page(seg, a)];
8582 			if ((VPP_PROT(vp) & protchk) == 0) {
8583 				mutex_exit(&svd->segp_slock);
8584 				error = EFAULT;
8585 				goto out;
8586 			}
8587 		}
8588 	}
8589 
8590 	/*
8591 	 * Avoid per page overhead of segvn_pp_lock_anonpages() for small
8592 	 * pages. For large pages segvn_pp_lock_anonpages() only does real
8593 	 * work once per large page.  The tradeoff is that we may decrement
8594 	 * availrmem more than once for the same page but this is ok
8595 	 * for small pages.
8596 	 */
8597 	if (seg->s_szc == 0) {
8598 		mutex_enter(&freemem_lock);
8599 		if (availrmem < tune.t_minarmem + npages) {
8600 			mutex_exit(&freemem_lock);
8601 			mutex_exit(&svd->segp_slock);
8602 			error = ENOMEM;
8603 			goto out;
8604 		}
8605 		availrmem -= npages;
8606 		mutex_exit(&freemem_lock);
8607 	}
8608 
8609 	pplist = kmem_alloc(sizeof (page_t *) * npages, KM_SLEEP);
8610 	pl = pplist;
8611 	*ppp = pplist + adjustpages;
8612 
8613 	page = seg_page(seg, addr);
8614 	anon_index = svd->anon_index + page;
8615 
8616 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8617 	for (a = addr; a < addr + len; a += PAGESIZE, anon_index++) {
8618 		struct anon *ap;
8619 		struct vnode *vp;
8620 		u_offset_t off;
8621 		anon_sync_obj_t cookie;
8622 
8623 		anon_array_enter(amp, anon_index, &cookie);
8624 		ap = anon_get_ptr(amp->ahp, anon_index);
8625 		if (ap == NULL) {
8626 			anon_array_exit(&cookie);
8627 			break;
8628 		} else {
8629 			/*
8630 			 * We must never use seg_pcache for COW pages
8631 			 * because we might end up with original page still
8632 			 * lying in seg_pcache even after private page is
8633 			 * created. This leads to data corruption as
8634 			 * aio_write refers to the page still in cache
8635 			 * while all other accesses refer to the private
8636 			 * page.
8637 			 */
8638 			if (ap->an_refcnt != 1) {
8639 				anon_array_exit(&cookie);
8640 				break;
8641 			}
8642 		}
8643 		swap_xlate(ap, &vp, &off);
8644 		anon_array_exit(&cookie);
8645 
8646 		pp = page_lookup_nowait(vp, off, SE_SHARED);
8647 		if (pp == NULL) {
8648 			break;
8649 		}
8650 		if (seg->s_szc != 0 || pp->p_szc != 0) {
8651 			if (!segvn_pp_lock_anonpages(pp, a == addr)) {
8652 				page_unlock(pp);
8653 				break;
8654 			}
8655 		} else {
8656 			szc0_npages++;
8657 		}
8658 		*pplist++ = pp;
8659 	}
8660 	ANON_LOCK_EXIT(&amp->a_rwlock);
8661 
8662 	ASSERT(npages >= szc0_npages);
8663 
8664 	if (a >= addr + len) {
8665 		mutex_enter(&freemem_lock);
8666 		if (seg->s_szc == 0 && npages != szc0_npages) {
8667 			ASSERT(svd->type == MAP_SHARED && amp->a_szc > 0);
8668 			availrmem += (npages - szc0_npages);
8669 		}
8670 		svd->softlockcnt += npages;
8671 		segvn_pages_locked += npages;
8672 		mutex_exit(&freemem_lock);
8673 		(void) seg_pinsert(seg, addr, len, pl, rw, SEGP_ASYNC_FLUSH,
8674 			segvn_reclaim);
8675 		mutex_exit(&svd->segp_slock);
8676 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8677 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_FILL_END,
8678 		    "segvn_pagelock: cache fill seg %p addr %p", seg, addr);
8679 		return (0);
8680 	}
8681 
8682 	mutex_exit(&svd->segp_slock);
8683 	if (seg->s_szc == 0) {
8684 		mutex_enter(&freemem_lock);
8685 		availrmem += npages;
8686 		mutex_exit(&freemem_lock);
8687 	}
8688 	error = EFAULT;
8689 	pplist = pl;
8690 	np = ((uintptr_t)(a - addr)) >> PAGESHIFT;
8691 	while (np > (uint_t)0) {
8692 		ASSERT(PAGE_LOCKED(*pplist));
8693 		if (seg->s_szc != 0 || (*pplist)->p_szc != 0) {
8694 			segvn_pp_unlock_anonpages(*pplist, pplist == pl);
8695 		}
8696 		page_unlock(*pplist);
8697 		np--;
8698 		pplist++;
8699 	}
8700 	kmem_free(pl, sizeof (page_t *) * npages);
8701 out:
8702 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8703 	*ppp = NULL;
8704 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
8705 		"segvn_pagelock: cache miss seg %p addr %p", seg, addr);
8706 	return (error);
8707 }
8708 
8709 /*
8710  * purge any cached pages in the I/O page cache
8711  */
8712 static void
8713 segvn_purge(struct seg *seg)
8714 {
8715 	seg_ppurge(seg);
8716 }
8717 
8718 static int
8719 segvn_reclaim(struct seg *seg, caddr_t addr, size_t len, struct page **pplist,
8720 	enum seg_rw rw)
8721 {
8722 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8723 	pgcnt_t np, npages;
8724 	struct page **pl;
8725 	pgcnt_t szc0_npages = 0;
8726 
8727 #ifdef lint
8728 	addr = addr;
8729 #endif
8730 
8731 	npages = np = (len >> PAGESHIFT);
8732 	ASSERT(npages);
8733 	pl = pplist;
8734 	if (seg->s_szc != 0) {
8735 		size_t pgsz = page_get_pagesize(seg->s_szc);
8736 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
8737 			panic("segvn_reclaim: unaligned addr or len");
8738 			/*NOTREACHED*/
8739 		}
8740 	}
8741 
8742 	ASSERT(svd->vp == NULL && svd->amp != NULL);
8743 
8744 	while (np > (uint_t)0) {
8745 		if (rw == S_WRITE) {
8746 			hat_setrefmod(*pplist);
8747 		} else {
8748 			hat_setref(*pplist);
8749 		}
8750 		if (seg->s_szc != 0 || (*pplist)->p_szc != 0) {
8751 			segvn_pp_unlock_anonpages(*pplist, pplist == pl);
8752 		} else {
8753 			szc0_npages++;
8754 		}
8755 		page_unlock(*pplist);
8756 		np--;
8757 		pplist++;
8758 	}
8759 	kmem_free(pl, sizeof (page_t *) * npages);
8760 
8761 	mutex_enter(&freemem_lock);
8762 	segvn_pages_locked -= npages;
8763 	svd->softlockcnt -= npages;
8764 	if (szc0_npages != 0) {
8765 		availrmem += szc0_npages;
8766 	}
8767 	mutex_exit(&freemem_lock);
8768 	if (svd->softlockcnt <= 0) {
8769 		if (AS_ISUNMAPWAIT(seg->s_as)) {
8770 			mutex_enter(&seg->s_as->a_contents);
8771 			if (AS_ISUNMAPWAIT(seg->s_as)) {
8772 				AS_CLRUNMAPWAIT(seg->s_as);
8773 				cv_broadcast(&seg->s_as->a_cv);
8774 			}
8775 			mutex_exit(&seg->s_as->a_contents);
8776 		}
8777 	}
8778 	return (0);
8779 }
8780 /*
8781  * get a memory ID for an addr in a given segment
8782  *
8783  * XXX only creates PAGESIZE pages if anon slots are not initialized.
8784  * At fault time they will be relocated into larger pages.
8785  */
8786 static int
8787 segvn_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
8788 {
8789 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8790 	struct anon 	*ap = NULL;
8791 	ulong_t		anon_index;
8792 	struct anon_map	*amp;
8793 	anon_sync_obj_t cookie;
8794 
8795 	if (svd->type == MAP_PRIVATE) {
8796 		memidp->val[0] = (uintptr_t)seg->s_as;
8797 		memidp->val[1] = (uintptr_t)addr;
8798 		return (0);
8799 	}
8800 
8801 	if (svd->type == MAP_SHARED) {
8802 		if (svd->vp) {
8803 			memidp->val[0] = (uintptr_t)svd->vp;
8804 			memidp->val[1] = (u_longlong_t)svd->offset +
8805 			    (uintptr_t)(addr - seg->s_base);
8806 			return (0);
8807 		} else {
8808 
8809 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8810 			if ((amp = svd->amp) != NULL) {
8811 				anon_index = svd->anon_index +
8812 				    seg_page(seg, addr);
8813 			}
8814 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8815 
8816 			ASSERT(amp != NULL);
8817 
8818 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8819 			anon_array_enter(amp, anon_index, &cookie);
8820 			ap = anon_get_ptr(amp->ahp, anon_index);
8821 			if (ap == NULL) {
8822 				page_t		*pp;
8823 
8824 				pp = anon_zero(seg, addr, &ap, svd->cred);
8825 				if (pp == NULL) {
8826 					anon_array_exit(&cookie);
8827 					ANON_LOCK_EXIT(&amp->a_rwlock);
8828 					return (ENOMEM);
8829 				}
8830 				ASSERT(anon_get_ptr(amp->ahp, anon_index)
8831 								== NULL);
8832 				(void) anon_set_ptr(amp->ahp, anon_index,
8833 				    ap, ANON_SLEEP);
8834 				page_unlock(pp);
8835 			}
8836 
8837 			anon_array_exit(&cookie);
8838 			ANON_LOCK_EXIT(&amp->a_rwlock);
8839 
8840 			memidp->val[0] = (uintptr_t)ap;
8841 			memidp->val[1] = (uintptr_t)addr & PAGEOFFSET;
8842 			return (0);
8843 		}
8844 	}
8845 	return (EINVAL);
8846 }
8847 
8848 static int
8849 sameprot(struct seg *seg, caddr_t a, size_t len)
8850 {
8851 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8852 	struct vpage *vpage;
8853 	spgcnt_t pages = btop(len);
8854 	uint_t prot;
8855 
8856 	if (svd->pageprot == 0)
8857 		return (1);
8858 
8859 	ASSERT(svd->vpage != NULL);
8860 
8861 	vpage = &svd->vpage[seg_page(seg, a)];
8862 	prot = VPP_PROT(vpage);
8863 	vpage++;
8864 	pages--;
8865 	while (pages-- > 0) {
8866 		if (prot != VPP_PROT(vpage))
8867 			return (0);
8868 		vpage++;
8869 	}
8870 	return (1);
8871 }
8872 
8873 /*
8874  * Get memory allocation policy info for specified address in given segment
8875  */
8876 static lgrp_mem_policy_info_t *
8877 segvn_getpolicy(struct seg *seg, caddr_t addr)
8878 {
8879 	struct anon_map		*amp;
8880 	ulong_t			anon_index;
8881 	lgrp_mem_policy_info_t	*policy_info;
8882 	struct segvn_data	*svn_data;
8883 	u_offset_t		vn_off;
8884 	vnode_t			*vp;
8885 
8886 	ASSERT(seg != NULL);
8887 
8888 	svn_data = (struct segvn_data *)seg->s_data;
8889 	if (svn_data == NULL)
8890 		return (NULL);
8891 
8892 	/*
8893 	 * Get policy info for private or shared memory
8894 	 */
8895 	if (svn_data->type != MAP_SHARED) {
8896 		if (svn_data->tr_state != SEGVN_TR_ON) {
8897 			policy_info = &svn_data->policy_info;
8898 		} else {
8899 			policy_info = &svn_data->tr_policy_info;
8900 			ASSERT(policy_info->mem_policy ==
8901 			    LGRP_MEM_POLICY_NEXT_SEG);
8902 		}
8903 	} else {
8904 		amp = svn_data->amp;
8905 		anon_index = svn_data->anon_index + seg_page(seg, addr);
8906 		vp = svn_data->vp;
8907 		vn_off = svn_data->offset + (uintptr_t)(addr - seg->s_base);
8908 		policy_info = lgrp_shm_policy_get(amp, anon_index, vp, vn_off);
8909 	}
8910 
8911 	return (policy_info);
8912 }
8913 
8914 /*ARGSUSED*/
8915 static int
8916 segvn_capable(struct seg *seg, segcapability_t capability)
8917 {
8918 	return (0);
8919 }
8920 
8921 /*
8922  * Bind text vnode segment to an amp. If we bind successfully mappings will be
8923  * established to per vnode mapping per lgroup amp pages instead of to vnode
8924  * pages. There's one amp per vnode text mapping per lgroup. Many processes
8925  * may share the same text replication amp. If a suitable amp doesn't already
8926  * exist in svntr hash table create a new one.  We may fail to bind to amp if
8927  * segment is not eligible for text replication.  Code below first checks for
8928  * these conditions. If binding is successful segment tr_state is set to on
8929  * and svd->amp points to the amp to use. Otherwise tr_state is set to off and
8930  * svd->amp remains as NULL.
8931  */
8932 static void
8933 segvn_textrepl(struct seg *seg)
8934 {
8935 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
8936 	vnode_t			*vp = svd->vp;
8937 	u_offset_t		off = svd->offset;
8938 	size_t			size = seg->s_size;
8939 	u_offset_t		eoff = off + size;
8940 	uint_t			szc = seg->s_szc;
8941 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
8942 	svntr_t			*svntrp;
8943 	struct vattr		va;
8944 	proc_t			*p = seg->s_as->a_proc;
8945 	lgrp_id_t		lgrp_id;
8946 	lgrp_id_t		olid;
8947 	int			first;
8948 	struct anon_map		*amp;
8949 
8950 	ASSERT(AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
8951 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
8952 	ASSERT(p != NULL);
8953 	ASSERT(svd->tr_state == SEGVN_TR_INIT);
8954 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
8955 	ASSERT(svd->flags & MAP_TEXT);
8956 	ASSERT(svd->type == MAP_PRIVATE);
8957 	ASSERT(vp != NULL && svd->amp == NULL);
8958 	ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
8959 	ASSERT(!(svd->flags & MAP_NORESERVE) && svd->swresv == 0);
8960 	ASSERT(seg->s_as != &kas);
8961 	ASSERT(off < eoff);
8962 	ASSERT(svntr_hashtab != NULL);
8963 
8964 	/*
8965 	 * If numa optimizations are no longer desired bail out.
8966 	 */
8967 	if (!lgrp_optimizations()) {
8968 		svd->tr_state = SEGVN_TR_OFF;
8969 		return;
8970 	}
8971 
8972 	/*
8973 	 * Avoid creating anon maps with size bigger than the file size.
8974 	 * If VOP_GETATTR() call fails bail out.
8975 	 */
8976 	va.va_mask = AT_SIZE | AT_MTIME;
8977 	if (VOP_GETATTR(vp, &va, 0, svd->cred) != 0) {
8978 		svd->tr_state = SEGVN_TR_OFF;
8979 		SEGVN_TR_ADDSTAT(gaerr);
8980 		return;
8981 	}
8982 	if (btopr(va.va_size) < btopr(eoff)) {
8983 		svd->tr_state = SEGVN_TR_OFF;
8984 		SEGVN_TR_ADDSTAT(overmap);
8985 		return;
8986 	}
8987 
8988 	/*
8989 	 * VVMEXEC may not be set yet if exec() prefaults text segment. Set
8990 	 * this flag now before vn_is_mapped(V_WRITE) so that MAP_SHARED
8991 	 * mapping that checks if trcache for this vnode needs to be
8992 	 * invalidated can't miss us.
8993 	 */
8994 	if (!(vp->v_flag & VVMEXEC)) {
8995 		mutex_enter(&vp->v_lock);
8996 		vp->v_flag |= VVMEXEC;
8997 		mutex_exit(&vp->v_lock);
8998 	}
8999 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9000 	/*
9001 	 * Bail out if potentially MAP_SHARED writable mappings exist to this
9002 	 * vnode.  We don't want to use old file contents from existing
9003 	 * replicas if this mapping was established after the original file
9004 	 * was changed.
9005 	 */
9006 	if (vn_is_mapped(vp, V_WRITE)) {
9007 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9008 		svd->tr_state = SEGVN_TR_OFF;
9009 		SEGVN_TR_ADDSTAT(wrcnt);
9010 		return;
9011 	}
9012 	svntrp = svntr_hashtab[hash].tr_head;
9013 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9014 		ASSERT(svntrp->tr_refcnt != 0);
9015 		if (svntrp->tr_vp != vp) {
9016 			continue;
9017 		}
9018 		/*
9019 		 * Bail out if file was changed after this replication entry
9020 		 * was created since we need to use the latest file contents.
9021 		 */
9022 		if (!svntrp->tr_valid ||
9023 		    svntrp->tr_mtime.tv_sec != va.va_mtime.tv_sec ||
9024 		    svntrp->tr_mtime.tv_nsec != va.va_mtime.tv_nsec) {
9025 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9026 			svd->tr_state = SEGVN_TR_OFF;
9027 			SEGVN_TR_ADDSTAT(stale);
9028 			return;
9029 		}
9030 		/*
9031 		 * if off, eoff and szc match current segment we found the
9032 		 * existing entry we can use.
9033 		 */
9034 		if (svntrp->tr_off == off && svntrp->tr_eoff == eoff &&
9035 		    svntrp->tr_szc == szc) {
9036 			break;
9037 		}
9038 		/*
9039 		 * Don't create different but overlapping in file offsets
9040 		 * entries to avoid replication of the same file pages more
9041 		 * than once per lgroup.
9042 		 */
9043 		if ((off >= svntrp->tr_off && off < svntrp->tr_eoff) ||
9044 		    (eoff > svntrp->tr_off && eoff <= svntrp->tr_eoff)) {
9045 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9046 			svd->tr_state = SEGVN_TR_OFF;
9047 			SEGVN_TR_ADDSTAT(overlap);
9048 			return;
9049 		}
9050 	}
9051 	/*
9052 	 * If we didn't find existing entry create a new one.
9053 	 */
9054 	if (svntrp == NULL) {
9055 		svntrp = kmem_cache_alloc(svntr_cache, KM_NOSLEEP);
9056 		if (svntrp == NULL) {
9057 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9058 			svd->tr_state = SEGVN_TR_OFF;
9059 			SEGVN_TR_ADDSTAT(nokmem);
9060 			return;
9061 		}
9062 #ifdef DEBUG
9063 		{
9064 			lgrp_id_t i;
9065 			for (i = 0; i < NLGRPS_MAX; i++) {
9066 				ASSERT(svntrp->tr_amp[i] == NULL);
9067 			}
9068 		}
9069 #endif /* DEBUG */
9070 		svntrp->tr_vp = vp;
9071 		svntrp->tr_off = off;
9072 		svntrp->tr_eoff = eoff;
9073 		svntrp->tr_szc = szc;
9074 		svntrp->tr_valid = 1;
9075 		svntrp->tr_mtime = va.va_mtime;
9076 		svntrp->tr_refcnt = 0;
9077 		svntrp->tr_next = svntr_hashtab[hash].tr_head;
9078 		svntr_hashtab[hash].tr_head = svntrp;
9079 	}
9080 	first = 1;
9081 again:
9082 	/*
9083 	 * We want to pick a replica with pages on main thread's (t_tid = 1,
9084 	 * aka T1) lgrp. Currently text replication is only optimized for
9085 	 * workloads that either have all threads of a process on the same
9086 	 * lgrp or execute their large text primarily on main thread.
9087 	 */
9088 	lgrp_id = p->p_t1_lgrpid;
9089 	if (lgrp_id == LGRP_NONE) {
9090 		/*
9091 		 * In case exec() prefaults text on non main thread use
9092 		 * current thread lgrpid.  It will become main thread anyway
9093 		 * soon.
9094 		 */
9095 		lgrp_id = lgrp_home_id(curthread);
9096 	}
9097 	/*
9098 	 * Set p_tr_lgrpid to lgrpid if it hasn't been set yet.  Otherwise
9099 	 * just set it to NLGRPS_MAX if it's different from current process T1
9100 	 * home lgrp.  p_tr_lgrpid is used to detect if process uses text
9101 	 * replication and T1 new home is different from lgrp used for text
9102 	 * replication. When this happens asyncronous segvn thread rechecks if
9103 	 * segments should change lgrps used for text replication.  If we fail
9104 	 * to set p_tr_lgrpid with cas32 then set it to NLGRPS_MAX without cas
9105 	 * if it's not already NLGRPS_MAX and not equal lgrp_id we want to
9106 	 * use.  We don't need to use cas in this case because another thread
9107 	 * that races in between our non atomic check and set may only change
9108 	 * p_tr_lgrpid to NLGRPS_MAX at this point.
9109 	 */
9110 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9111 	olid = p->p_tr_lgrpid;
9112 	if (lgrp_id != olid && olid != NLGRPS_MAX) {
9113 		lgrp_id_t nlid = (olid == LGRP_NONE) ? lgrp_id : NLGRPS_MAX;
9114 		if (cas32((uint32_t *)&p->p_tr_lgrpid, olid, nlid) != olid) {
9115 			olid = p->p_tr_lgrpid;
9116 			ASSERT(olid != LGRP_NONE);
9117 			if (olid != lgrp_id && olid != NLGRPS_MAX) {
9118 				p->p_tr_lgrpid = NLGRPS_MAX;
9119 			}
9120 		}
9121 		ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9122 		membar_producer();
9123 		/*
9124 		 * lgrp_move_thread() won't schedule async recheck after
9125 		 * p->p_t1_lgrpid update unless p->p_tr_lgrpid is not
9126 		 * LGRP_NONE. Recheck p_t1_lgrpid once now that p->p_tr_lgrpid
9127 		 * is not LGRP_NONE.
9128 		 */
9129 		if (first && p->p_t1_lgrpid != LGRP_NONE &&
9130 		    p->p_t1_lgrpid != lgrp_id) {
9131 			first = 0;
9132 			goto again;
9133 		}
9134 	}
9135 	/*
9136 	 * If no amp was created yet for lgrp_id create a new one as long as
9137 	 * we have enough memory to afford it.
9138 	 */
9139 	if ((amp = svntrp->tr_amp[lgrp_id]) == NULL) {
9140 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9141 		if (trmem > segvn_textrepl_max_bytes) {
9142 			SEGVN_TR_ADDSTAT(normem);
9143 			goto fail;
9144 		}
9145 		if (anon_try_resv_zone(size, NULL) == 0) {
9146 			SEGVN_TR_ADDSTAT(noanon);
9147 			goto fail;
9148 		}
9149 		amp = anonmap_alloc(size, size, ANON_NOSLEEP);
9150 		if (amp == NULL) {
9151 			anon_unresv_zone(size, NULL);
9152 			SEGVN_TR_ADDSTAT(nokmem);
9153 			goto fail;
9154 		}
9155 		ASSERT(amp->refcnt == 1);
9156 		amp->a_szc = szc;
9157 		svntrp->tr_amp[lgrp_id] = amp;
9158 		SEGVN_TR_ADDSTAT(newamp);
9159 	}
9160 	svntrp->tr_refcnt++;
9161 	ASSERT(svd->svn_trnext == NULL);
9162 	ASSERT(svd->svn_trprev == NULL);
9163 	svd->svn_trnext = svntrp->tr_svnhead;
9164 	svd->svn_trprev = NULL;
9165 	if (svntrp->tr_svnhead != NULL) {
9166 		svntrp->tr_svnhead->svn_trprev = svd;
9167 	}
9168 	svntrp->tr_svnhead = svd;
9169 	ASSERT(amp->a_szc == szc && amp->size == size && amp->swresv == size);
9170 	ASSERT(amp->refcnt >= 1);
9171 	svd->amp = amp;
9172 	svd->anon_index = 0;
9173 	svd->tr_policy_info.mem_policy = LGRP_MEM_POLICY_NEXT_SEG;
9174 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
9175 	svd->tr_state = SEGVN_TR_ON;
9176 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9177 	SEGVN_TR_ADDSTAT(repl);
9178 	return;
9179 fail:
9180 	ASSERT(segvn_textrepl_bytes >= size);
9181 	atomic_add_long(&segvn_textrepl_bytes, -size);
9182 	ASSERT(svntrp != NULL);
9183 	ASSERT(svntrp->tr_amp[lgrp_id] == NULL);
9184 	if (svntrp->tr_refcnt == 0) {
9185 		ASSERT(svntrp == svntr_hashtab[hash].tr_head);
9186 		svntr_hashtab[hash].tr_head = svntrp->tr_next;
9187 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9188 		kmem_cache_free(svntr_cache, svntrp);
9189 	} else {
9190 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9191 	}
9192 	svd->tr_state = SEGVN_TR_OFF;
9193 }
9194 
9195 /*
9196  * Convert seg back to regular vnode mapping seg by unbinding it from its text
9197  * replication amp.  This routine is most typically called when segment is
9198  * unmapped but can also be called when segment no longer qualifies for text
9199  * replication (e.g. due to protection changes). If unload_unmap is set use
9200  * HAT_UNLOAD_UNMAP flag in hat_unload_callback().  If we are the last user of
9201  * svntr free all its anon maps and remove it from the hash table.
9202  */
9203 static void
9204 segvn_textunrepl(struct seg *seg, int unload_unmap)
9205 {
9206 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
9207 	vnode_t			*vp = svd->vp;
9208 	u_offset_t		off = svd->offset;
9209 	size_t			size = seg->s_size;
9210 	u_offset_t		eoff = off + size;
9211 	uint_t			szc = seg->s_szc;
9212 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9213 	svntr_t			*svntrp;
9214 	svntr_t			**prv_svntrp;
9215 	lgrp_id_t		lgrp_id = svd->tr_policy_info.mem_lgrpid;
9216 	lgrp_id_t		i;
9217 
9218 	ASSERT(AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
9219 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
9220 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9221 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9222 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9223 	ASSERT(svd->amp != NULL);
9224 	ASSERT(svd->amp->refcnt >= 1);
9225 	ASSERT(svd->anon_index == 0);
9226 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9227 	ASSERT(svntr_hashtab != NULL);
9228 
9229 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9230 	prv_svntrp = &svntr_hashtab[hash].tr_head;
9231 	for (; (svntrp = *prv_svntrp) != NULL; prv_svntrp = &svntrp->tr_next) {
9232 		ASSERT(svntrp->tr_refcnt != 0);
9233 		if (svntrp->tr_vp == vp && svntrp->tr_off == off &&
9234 		    svntrp->tr_eoff == eoff && svntrp->tr_szc == szc) {
9235 			break;
9236 		}
9237 	}
9238 	if (svntrp == NULL) {
9239 		panic("segvn_textunrepl: svntr record not found");
9240 	}
9241 	if (svntrp->tr_amp[lgrp_id] != svd->amp) {
9242 		panic("segvn_textunrepl: amp mismatch");
9243 	}
9244 	svd->tr_state = SEGVN_TR_OFF;
9245 	svd->amp = NULL;
9246 	if (svd->svn_trprev == NULL) {
9247 		ASSERT(svntrp->tr_svnhead == svd);
9248 		svntrp->tr_svnhead = svd->svn_trnext;
9249 		if (svntrp->tr_svnhead != NULL) {
9250 			svntrp->tr_svnhead->svn_trprev = NULL;
9251 		}
9252 		svd->svn_trnext = NULL;
9253 	} else {
9254 		svd->svn_trprev->svn_trnext = svd->svn_trnext;
9255 		if (svd->svn_trnext != NULL) {
9256 			svd->svn_trnext->svn_trprev = svd->svn_trprev;
9257 			svd->svn_trnext = NULL;
9258 		}
9259 		svd->svn_trprev = NULL;
9260 	}
9261 	if (--svntrp->tr_refcnt) {
9262 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9263 		goto done;
9264 	}
9265 	*prv_svntrp = svntrp->tr_next;
9266 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9267 	for (i = 0; i < NLGRPS_MAX; i++) {
9268 		struct anon_map *amp = svntrp->tr_amp[i];
9269 		if (amp == NULL) {
9270 			continue;
9271 		}
9272 		ASSERT(amp->refcnt == 1);
9273 		ASSERT(amp->swresv == size);
9274 		ASSERT(amp->size == size);
9275 		ASSERT(amp->a_szc == szc);
9276 		if (amp->a_szc != 0) {
9277 			anon_free_pages(amp->ahp, 0, size, szc);
9278 		} else {
9279 			anon_free(amp->ahp, 0, size);
9280 		}
9281 		svntrp->tr_amp[i] = NULL;
9282 		ASSERT(segvn_textrepl_bytes >= size);
9283 		atomic_add_long(&segvn_textrepl_bytes, -size);
9284 		anon_unresv_zone(amp->swresv, NULL);
9285 		amp->refcnt = 0;
9286 		anonmap_free(amp);
9287 	}
9288 	kmem_cache_free(svntr_cache, svntrp);
9289 done:
9290 	hat_unload_callback(seg->s_as->a_hat, seg->s_base, size,
9291 	    unload_unmap ? HAT_UNLOAD_UNMAP : 0, NULL);
9292 }
9293 
9294 /*
9295  * This is called when a MAP_SHARED writabble mapping is created to a vnode
9296  * that is currently used for execution (VVMEXEC flag is set). In this case we
9297  * need to prevent further use of existing replicas.
9298  */
9299 static void
9300 segvn_inval_trcache(vnode_t *vp)
9301 {
9302 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9303 	svntr_t			*svntrp;
9304 
9305 	ASSERT(vp->v_flag & VVMEXEC);
9306 
9307 	if (svntr_hashtab == NULL) {
9308 		return;
9309 	}
9310 
9311 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9312 	svntrp = svntr_hashtab[hash].tr_head;
9313 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9314 		ASSERT(svntrp->tr_refcnt != 0);
9315 		if (svntrp->tr_vp == vp && svntrp->tr_valid) {
9316 			svntrp->tr_valid = 0;
9317 		}
9318 	}
9319 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9320 }
9321 
9322 static void
9323 segvn_trasync_thread(void)
9324 {
9325 	callb_cpr_t cpr_info;
9326 	kmutex_t cpr_lock;	/* just for CPR stuff */
9327 
9328 	mutex_init(&cpr_lock, NULL, MUTEX_DEFAULT, NULL);
9329 
9330 	CALLB_CPR_INIT(&cpr_info, &cpr_lock,
9331 	    callb_generic_cpr, "segvn_async");
9332 
9333 	if (segvn_update_textrepl_interval == 0) {
9334 		segvn_update_textrepl_interval = segvn_update_tr_time * hz;
9335 	} else {
9336 		segvn_update_textrepl_interval *= hz;
9337 	}
9338 	(void) timeout(segvn_trupdate_wakeup, NULL,
9339 	    segvn_update_textrepl_interval);
9340 
9341 	for (;;) {
9342 		mutex_enter(&cpr_lock);
9343 		CALLB_CPR_SAFE_BEGIN(&cpr_info);
9344 		mutex_exit(&cpr_lock);
9345 		sema_p(&segvn_trasync_sem);
9346 		mutex_enter(&cpr_lock);
9347 		CALLB_CPR_SAFE_END(&cpr_info, &cpr_lock);
9348 		mutex_exit(&cpr_lock);
9349 		segvn_trupdate();
9350 	}
9351 }
9352 
9353 static uint64_t segvn_lgrp_trthr_migrs_snpsht = 0;
9354 
9355 static void
9356 segvn_trupdate_wakeup(void *dummy)
9357 {
9358 	uint64_t cur_lgrp_trthr_migrs = lgrp_get_trthr_migrations();
9359 
9360 	if (cur_lgrp_trthr_migrs != segvn_lgrp_trthr_migrs_snpsht) {
9361 		segvn_lgrp_trthr_migrs_snpsht = cur_lgrp_trthr_migrs;
9362 		sema_v(&segvn_trasync_sem);
9363 	}
9364 
9365 	if (!segvn_disable_textrepl_update &&
9366 	    segvn_update_textrepl_interval != 0) {
9367 		(void) timeout(segvn_trupdate_wakeup, dummy,
9368 		    segvn_update_textrepl_interval);
9369 	}
9370 }
9371 
9372 static void
9373 segvn_trupdate(void)
9374 {
9375 	ulong_t		hash;
9376 	svntr_t		*svntrp;
9377 	segvn_data_t	*svd;
9378 
9379 	ASSERT(svntr_hashtab != NULL);
9380 
9381 	for (hash = 0; hash < svntr_hashtab_sz; hash++) {
9382 		mutex_enter(&svntr_hashtab[hash].tr_lock);
9383 		svntrp = svntr_hashtab[hash].tr_head;
9384 		for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9385 			ASSERT(svntrp->tr_refcnt != 0);
9386 			svd = svntrp->tr_svnhead;
9387 			for (; svd != NULL; svd = svd->svn_trnext) {
9388 				segvn_trupdate_seg(svd->seg, svd, svntrp,
9389 				    hash);
9390 			}
9391 		}
9392 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9393 	}
9394 }
9395 
9396 static void
9397 segvn_trupdate_seg(struct seg *seg,
9398 	segvn_data_t *svd,
9399 	svntr_t *svntrp,
9400 	ulong_t hash)
9401 {
9402 	proc_t			*p;
9403 	lgrp_id_t		lgrp_id;
9404 	struct as		*as;
9405 	size_t			size;
9406 	struct anon_map		*amp;
9407 
9408 	ASSERT(svd->vp != NULL);
9409 	ASSERT(svd->vp == svntrp->tr_vp);
9410 	ASSERT(svd->offset == svntrp->tr_off);
9411 	ASSERT(svd->offset + seg->s_size == svntrp->tr_eoff);
9412 	ASSERT(seg != NULL);
9413 	ASSERT(svd->seg == seg);
9414 	ASSERT(seg->s_data == (void *)svd);
9415 	ASSERT(seg->s_szc == svntrp->tr_szc);
9416 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9417 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9418 	ASSERT(svd->amp != NULL);
9419 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
9420 	ASSERT(svd->tr_policy_info.mem_lgrpid != LGRP_NONE);
9421 	ASSERT(svd->tr_policy_info.mem_lgrpid < NLGRPS_MAX);
9422 	ASSERT(svntrp->tr_amp[svd->tr_policy_info.mem_lgrpid] == svd->amp);
9423 	ASSERT(svntrp->tr_refcnt != 0);
9424 	ASSERT(mutex_owned(&svntr_hashtab[hash].tr_lock));
9425 
9426 	as = seg->s_as;
9427 	ASSERT(as != NULL && as != &kas);
9428 	p = as->a_proc;
9429 	ASSERT(p != NULL);
9430 	ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9431 	lgrp_id = p->p_t1_lgrpid;
9432 	if (lgrp_id == LGRP_NONE) {
9433 		return;
9434 	}
9435 	ASSERT(lgrp_id < NLGRPS_MAX);
9436 	if (svd->tr_policy_info.mem_lgrpid == lgrp_id) {
9437 		return;
9438 	}
9439 
9440 	/*
9441 	 * Use tryenter locking since we are locking as/seg and svntr hash
9442 	 * lock in reverse from syncrounous thread order.
9443 	 */
9444 	if (!AS_LOCK_TRYENTER(as, &as->a_lock, RW_READER)) {
9445 		SEGVN_TR_ADDSTAT(nolock);
9446 		if (segvn_lgrp_trthr_migrs_snpsht) {
9447 			segvn_lgrp_trthr_migrs_snpsht = 0;
9448 		}
9449 		return;
9450 	}
9451 	if (!SEGVN_LOCK_TRYENTER(seg->s_as, &svd->lock, RW_WRITER)) {
9452 		AS_LOCK_EXIT(as, &as->a_lock);
9453 		SEGVN_TR_ADDSTAT(nolock);
9454 		if (segvn_lgrp_trthr_migrs_snpsht) {
9455 			segvn_lgrp_trthr_migrs_snpsht = 0;
9456 		}
9457 		return;
9458 	}
9459 	size = seg->s_size;
9460 	if (svntrp->tr_amp[lgrp_id] == NULL) {
9461 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9462 		if (trmem > segvn_textrepl_max_bytes) {
9463 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9464 			AS_LOCK_EXIT(as, &as->a_lock);
9465 			atomic_add_long(&segvn_textrepl_bytes, -size);
9466 			SEGVN_TR_ADDSTAT(normem);
9467 			return;
9468 		}
9469 		if (anon_try_resv_zone(size, NULL) == 0) {
9470 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9471 			AS_LOCK_EXIT(as, &as->a_lock);
9472 			atomic_add_long(&segvn_textrepl_bytes, -size);
9473 			SEGVN_TR_ADDSTAT(noanon);
9474 			return;
9475 		}
9476 		amp = anonmap_alloc(size, size, KM_NOSLEEP);
9477 		if (amp == NULL) {
9478 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9479 			AS_LOCK_EXIT(as, &as->a_lock);
9480 			atomic_add_long(&segvn_textrepl_bytes, -size);
9481 			anon_unresv_zone(size, NULL);
9482 			SEGVN_TR_ADDSTAT(nokmem);
9483 			return;
9484 		}
9485 		ASSERT(amp->refcnt == 1);
9486 		amp->a_szc = seg->s_szc;
9487 		svntrp->tr_amp[lgrp_id] = amp;
9488 	}
9489 	/*
9490 	 * We don't need to drop the bucket lock but here we give other
9491 	 * threads a chance.  svntr and svd can't be unlinked as long as
9492 	 * segment lock is held as a writer and AS held as well.  After we
9493 	 * retake bucket lock we'll continue from where we left. We'll be able
9494 	 * to reach the end of either list since new entries are always added
9495 	 * to the beginning of the lists.
9496 	 */
9497 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9498 	hat_unload_callback(as->a_hat, seg->s_base, size, 0, NULL);
9499 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9500 
9501 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9502 	ASSERT(svd->amp != NULL);
9503 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
9504 	ASSERT(svd->tr_policy_info.mem_lgrpid != lgrp_id);
9505 	ASSERT(svd->amp != svntrp->tr_amp[lgrp_id]);
9506 
9507 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
9508 	svd->amp = svntrp->tr_amp[lgrp_id];
9509 	p->p_tr_lgrpid = NLGRPS_MAX;
9510 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9511 	AS_LOCK_EXIT(as, &as->a_lock);
9512 
9513 	ASSERT(svntrp->tr_refcnt != 0);
9514 	ASSERT(svd->vp == svntrp->tr_vp);
9515 	ASSERT(svd->tr_policy_info.mem_lgrpid == lgrp_id);
9516 	ASSERT(svd->amp != NULL && svd->amp == svntrp->tr_amp[lgrp_id]);
9517 	ASSERT(svd->seg == seg);
9518 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9519 
9520 	SEGVN_TR_ADDSTAT(asyncrepl);
9521 }
9522