xref: /illumos-gate/usr/src/uts/common/vm/seg_vn.c (revision a6bde1a23b60f140c7ed78df979c2e22b1ed9b2c)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 1986, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
26 /*	  All Rights Reserved  	*/
27 
28 /*
29  * University Copyright- Copyright (c) 1982, 1986, 1988
30  * The Regents of the University of California
31  * All Rights Reserved
32  *
33  * University Acknowledgment- Portions of this document are derived from
34  * software developed by the University of California, Berkeley, and its
35  * contributors.
36  */
37 
38 /*
39  * VM - shared or copy-on-write from a vnode/anonymous memory.
40  */
41 
42 #include <sys/types.h>
43 #include <sys/param.h>
44 #include <sys/t_lock.h>
45 #include <sys/errno.h>
46 #include <sys/systm.h>
47 #include <sys/mman.h>
48 #include <sys/debug.h>
49 #include <sys/cred.h>
50 #include <sys/vmsystm.h>
51 #include <sys/tuneable.h>
52 #include <sys/bitmap.h>
53 #include <sys/swap.h>
54 #include <sys/kmem.h>
55 #include <sys/sysmacros.h>
56 #include <sys/vtrace.h>
57 #include <sys/cmn_err.h>
58 #include <sys/callb.h>
59 #include <sys/vm.h>
60 #include <sys/dumphdr.h>
61 #include <sys/lgrp.h>
62 
63 #include <vm/hat.h>
64 #include <vm/as.h>
65 #include <vm/seg.h>
66 #include <vm/seg_vn.h>
67 #include <vm/pvn.h>
68 #include <vm/anon.h>
69 #include <vm/page.h>
70 #include <vm/vpage.h>
71 #include <sys/proc.h>
72 #include <sys/task.h>
73 #include <sys/project.h>
74 #include <sys/zone.h>
75 #include <sys/shm_impl.h>
76 /*
77  * Private seg op routines.
78  */
79 static int	segvn_dup(struct seg *seg, struct seg *newseg);
80 static int	segvn_unmap(struct seg *seg, caddr_t addr, size_t len);
81 static void	segvn_free(struct seg *seg);
82 static faultcode_t segvn_fault(struct hat *hat, struct seg *seg,
83 		    caddr_t addr, size_t len, enum fault_type type,
84 		    enum seg_rw rw);
85 static faultcode_t segvn_faulta(struct seg *seg, caddr_t addr);
86 static int	segvn_setprot(struct seg *seg, caddr_t addr,
87 		    size_t len, uint_t prot);
88 static int	segvn_checkprot(struct seg *seg, caddr_t addr,
89 		    size_t len, uint_t prot);
90 static int	segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta);
91 static size_t	segvn_swapout(struct seg *seg);
92 static int	segvn_sync(struct seg *seg, caddr_t addr, size_t len,
93 		    int attr, uint_t flags);
94 static size_t	segvn_incore(struct seg *seg, caddr_t addr, size_t len,
95 		    char *vec);
96 static int	segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
97 		    int attr, int op, ulong_t *lockmap, size_t pos);
98 static int	segvn_getprot(struct seg *seg, caddr_t addr, size_t len,
99 		    uint_t *protv);
100 static u_offset_t	segvn_getoffset(struct seg *seg, caddr_t addr);
101 static int	segvn_gettype(struct seg *seg, caddr_t addr);
102 static int	segvn_getvp(struct seg *seg, caddr_t addr,
103 		    struct vnode **vpp);
104 static int	segvn_advise(struct seg *seg, caddr_t addr, size_t len,
105 		    uint_t behav);
106 static void	segvn_dump(struct seg *seg);
107 static int	segvn_pagelock(struct seg *seg, caddr_t addr, size_t len,
108 		    struct page ***ppp, enum lock_type type, enum seg_rw rw);
109 static int	segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len,
110 		    uint_t szc);
111 static int	segvn_getmemid(struct seg *seg, caddr_t addr,
112 		    memid_t *memidp);
113 static lgrp_mem_policy_info_t	*segvn_getpolicy(struct seg *, caddr_t);
114 static int	segvn_capable(struct seg *seg, segcapability_t capable);
115 
116 struct	seg_ops segvn_ops = {
117 	segvn_dup,
118 	segvn_unmap,
119 	segvn_free,
120 	segvn_fault,
121 	segvn_faulta,
122 	segvn_setprot,
123 	segvn_checkprot,
124 	segvn_kluster,
125 	segvn_swapout,
126 	segvn_sync,
127 	segvn_incore,
128 	segvn_lockop,
129 	segvn_getprot,
130 	segvn_getoffset,
131 	segvn_gettype,
132 	segvn_getvp,
133 	segvn_advise,
134 	segvn_dump,
135 	segvn_pagelock,
136 	segvn_setpagesize,
137 	segvn_getmemid,
138 	segvn_getpolicy,
139 	segvn_capable,
140 };
141 
142 /*
143  * Common zfod structures, provided as a shorthand for others to use.
144  */
145 static segvn_crargs_t zfod_segvn_crargs =
146 	SEGVN_ZFOD_ARGS(PROT_ZFOD, PROT_ALL);
147 static segvn_crargs_t kzfod_segvn_crargs =
148 	SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_USER,
149 	PROT_ALL & ~PROT_USER);
150 static segvn_crargs_t stack_noexec_crargs =
151 	SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_EXEC, PROT_ALL);
152 
153 caddr_t	zfod_argsp = (caddr_t)&zfod_segvn_crargs;	/* user zfod argsp */
154 caddr_t	kzfod_argsp = (caddr_t)&kzfod_segvn_crargs;	/* kernel zfod argsp */
155 caddr_t	stack_exec_argsp = (caddr_t)&zfod_segvn_crargs;	/* executable stack */
156 caddr_t	stack_noexec_argsp = (caddr_t)&stack_noexec_crargs; /* noexec stack */
157 
158 #define	vpgtob(n)	((n) * sizeof (struct vpage))	/* For brevity */
159 
160 size_t	segvn_comb_thrshld = UINT_MAX;	/* patchable -- see 1196681 */
161 
162 size_t	segvn_pglock_comb_thrshld = (1UL << 16);	/* 64K */
163 size_t	segvn_pglock_comb_balign = (1UL << 16);		/* 64K */
164 uint_t	segvn_pglock_comb_bshift;
165 size_t	segvn_pglock_comb_palign;
166 
167 static int	segvn_concat(struct seg *, struct seg *, int);
168 static int	segvn_extend_prev(struct seg *, struct seg *,
169 		    struct segvn_crargs *, size_t);
170 static int	segvn_extend_next(struct seg *, struct seg *,
171 		    struct segvn_crargs *, size_t);
172 static void	segvn_softunlock(struct seg *, caddr_t, size_t, enum seg_rw);
173 static void	segvn_pagelist_rele(page_t **);
174 static void	segvn_setvnode_mpss(vnode_t *);
175 static void	segvn_relocate_pages(page_t **, page_t *);
176 static int	segvn_full_szcpages(page_t **, uint_t, int *, uint_t *);
177 static int	segvn_fill_vp_pages(struct segvn_data *, vnode_t *, u_offset_t,
178     uint_t, page_t **, page_t **, uint_t *, int *);
179 static faultcode_t segvn_fault_vnodepages(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_fault_anonpages(struct hat *, struct seg *, caddr_t,
182     caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
183 static faultcode_t segvn_faultpage(struct hat *, struct seg *, caddr_t,
184     u_offset_t, struct vpage *, page_t **, uint_t,
185     enum fault_type, enum seg_rw, int);
186 static void	segvn_vpage(struct seg *);
187 static size_t	segvn_count_swap_by_vpages(struct seg *);
188 
189 static void segvn_purge(struct seg *seg);
190 static int segvn_reclaim(void *, caddr_t, size_t, struct page **,
191     enum seg_rw, int);
192 static int shamp_reclaim(void *, caddr_t, size_t, struct page **,
193     enum seg_rw, int);
194 
195 static int sameprot(struct seg *, caddr_t, size_t);
196 
197 static int segvn_demote_range(struct seg *, caddr_t, size_t, int, uint_t);
198 static int segvn_clrszc(struct seg *);
199 static struct seg *segvn_split_seg(struct seg *, caddr_t);
200 static int segvn_claim_pages(struct seg *, struct vpage *, u_offset_t,
201     ulong_t, uint_t);
202 
203 static void segvn_hat_rgn_unload_callback(caddr_t, caddr_t, caddr_t,
204     size_t, void *, u_offset_t);
205 
206 static struct kmem_cache *segvn_cache;
207 static struct kmem_cache **segvn_szc_cache;
208 
209 #ifdef VM_STATS
210 static struct segvnvmstats_str {
211 	ulong_t	fill_vp_pages[31];
212 	ulong_t fltvnpages[49];
213 	ulong_t	fullszcpages[10];
214 	ulong_t	relocatepages[3];
215 	ulong_t	fltanpages[17];
216 	ulong_t pagelock[2];
217 	ulong_t	demoterange[3];
218 } segvnvmstats;
219 #endif /* VM_STATS */
220 
221 #define	SDR_RANGE	1		/* demote entire range */
222 #define	SDR_END		2		/* demote non aligned ends only */
223 
224 #define	CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr) {	    \
225 		if ((len) != 0) { 		      	      		      \
226 			lpgaddr = (caddr_t)P2ALIGN((uintptr_t)(addr), pgsz);  \
227 			ASSERT(lpgaddr >= (seg)->s_base);	      	      \
228 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)((addr) +    \
229 			    (len)), pgsz);				      \
230 			ASSERT(lpgeaddr > lpgaddr);		      	      \
231 			ASSERT(lpgeaddr <= (seg)->s_base + (seg)->s_size);    \
232 		} else {					      	      \
233 			lpgeaddr = lpgaddr = (addr);	      		      \
234 		}							      \
235 	}
236 
237 /*ARGSUSED*/
238 static int
239 segvn_cache_constructor(void *buf, void *cdrarg, int kmflags)
240 {
241 	struct segvn_data *svd = buf;
242 
243 	rw_init(&svd->lock, NULL, RW_DEFAULT, NULL);
244 	mutex_init(&svd->segfree_syncmtx, NULL, MUTEX_DEFAULT, NULL);
245 	svd->svn_trnext = svd->svn_trprev = NULL;
246 	return (0);
247 }
248 
249 /*ARGSUSED1*/
250 static void
251 segvn_cache_destructor(void *buf, void *cdrarg)
252 {
253 	struct segvn_data *svd = buf;
254 
255 	rw_destroy(&svd->lock);
256 	mutex_destroy(&svd->segfree_syncmtx);
257 }
258 
259 /*ARGSUSED*/
260 static int
261 svntr_cache_constructor(void *buf, void *cdrarg, int kmflags)
262 {
263 	bzero(buf, sizeof (svntr_t));
264 	return (0);
265 }
266 
267 /*
268  * Patching this variable to non-zero allows the system to run with
269  * stacks marked as "not executable".  It's a bit of a kludge, but is
270  * provided as a tweakable for platforms that export those ABIs
271  * (e.g. sparc V8) that have executable stacks enabled by default.
272  * There are also some restrictions for platforms that don't actually
273  * implement 'noexec' protections.
274  *
275  * Once enabled, the system is (therefore) unable to provide a fully
276  * ABI-compliant execution environment, though practically speaking,
277  * most everything works.  The exceptions are generally some interpreters
278  * and debuggers that create executable code on the stack and jump
279  * into it (without explicitly mprotecting the address range to include
280  * PROT_EXEC).
281  *
282  * One important class of applications that are disabled are those
283  * that have been transformed into malicious agents using one of the
284  * numerous "buffer overflow" attacks.  See 4007890.
285  */
286 int noexec_user_stack = 0;
287 int noexec_user_stack_log = 1;
288 
289 int segvn_lpg_disable = 0;
290 uint_t segvn_maxpgszc = 0;
291 
292 ulong_t segvn_vmpss_clrszc_cnt;
293 ulong_t segvn_vmpss_clrszc_err;
294 ulong_t segvn_fltvnpages_clrszc_cnt;
295 ulong_t segvn_fltvnpages_clrszc_err;
296 ulong_t segvn_setpgsz_align_err;
297 ulong_t segvn_setpgsz_anon_align_err;
298 ulong_t segvn_setpgsz_getattr_err;
299 ulong_t segvn_setpgsz_eof_err;
300 ulong_t segvn_faultvnmpss_align_err1;
301 ulong_t segvn_faultvnmpss_align_err2;
302 ulong_t segvn_faultvnmpss_align_err3;
303 ulong_t segvn_faultvnmpss_align_err4;
304 ulong_t segvn_faultvnmpss_align_err5;
305 ulong_t	segvn_vmpss_pageio_deadlk_err;
306 
307 int segvn_use_regions = 1;
308 
309 /*
310  * Segvn supports text replication optimization for NUMA platforms. Text
311  * replica's are represented by anon maps (amp). There's one amp per text file
312  * region per lgroup. A process chooses the amp for each of its text mappings
313  * based on the lgroup assignment of its main thread (t_tid = 1). All
314  * processes that want a replica on a particular lgroup for the same text file
315  * mapping share the same amp. amp's are looked up in svntr_hashtab hash table
316  * with vp,off,size,szc used as a key. Text replication segments are read only
317  * MAP_PRIVATE|MAP_TEXT segments that map vnode. Replication is achieved by
318  * forcing COW faults from vnode to amp and mapping amp pages instead of vnode
319  * pages. Replication amp is assigned to a segment when it gets its first
320  * pagefault. To handle main thread lgroup rehoming segvn_trasync_thread
321  * rechecks periodically if the process still maps an amp local to the main
322  * thread. If not async thread forces process to remap to an amp in the new
323  * home lgroup of the main thread. Current text replication implementation
324  * only provides the benefit to workloads that do most of their work in the
325  * main thread of a process or all the threads of a process run in the same
326  * lgroup. To extend text replication benefit to different types of
327  * multithreaded workloads further work would be needed in the hat layer to
328  * allow the same virtual address in the same hat to simultaneously map
329  * different physical addresses (i.e. page table replication would be needed
330  * for x86).
331  *
332  * amp pages are used instead of vnode pages as long as segment has a very
333  * simple life cycle.  It's created via segvn_create(), handles S_EXEC
334  * (S_READ) pagefaults and is fully unmapped.  If anything more complicated
335  * happens such as protection is changed, real COW fault happens, pagesize is
336  * changed, MC_LOCK is requested or segment is partially unmapped we turn off
337  * text replication by converting the segment back to vnode only segment
338  * (unmap segment's address range and set svd->amp to NULL).
339  *
340  * The original file can be changed after amp is inserted into
341  * svntr_hashtab. Processes that are launched after the file is already
342  * changed can't use the replica's created prior to the file change. To
343  * implement this functionality hash entries are timestamped. Replica's can
344  * only be used if current file modification time is the same as the timestamp
345  * saved when hash entry was created. However just timestamps alone are not
346  * sufficient to detect file modification via mmap(MAP_SHARED) mappings. We
347  * deal with file changes via MAP_SHARED mappings differently. When writable
348  * MAP_SHARED mappings are created to vnodes marked as executable we mark all
349  * existing replica's for this vnode as not usable for future text
350  * mappings. And we don't create new replica's for files that currently have
351  * potentially writable MAP_SHARED mappings (i.e. vn_is_mapped(V_WRITE) is
352  * true).
353  */
354 
355 #define	SEGVN_TEXTREPL_MAXBYTES_FACTOR	(20)
356 size_t	segvn_textrepl_max_bytes_factor = SEGVN_TEXTREPL_MAXBYTES_FACTOR;
357 
358 static ulong_t			svntr_hashtab_sz = 512;
359 static svntr_bucket_t		*svntr_hashtab = NULL;
360 static struct kmem_cache	*svntr_cache;
361 static svntr_stats_t		*segvn_textrepl_stats;
362 static ksema_t 			segvn_trasync_sem;
363 
364 int				segvn_disable_textrepl = 1;
365 size_t				textrepl_size_thresh = (size_t)-1;
366 size_t				segvn_textrepl_bytes = 0;
367 size_t				segvn_textrepl_max_bytes = 0;
368 clock_t				segvn_update_textrepl_interval = 0;
369 int				segvn_update_tr_time = 10;
370 int				segvn_disable_textrepl_update = 0;
371 
372 static void segvn_textrepl(struct seg *);
373 static void segvn_textunrepl(struct seg *, int);
374 static void segvn_inval_trcache(vnode_t *);
375 static void segvn_trasync_thread(void);
376 static void segvn_trupdate_wakeup(void *);
377 static void segvn_trupdate(void);
378 static void segvn_trupdate_seg(struct seg *, segvn_data_t *, svntr_t *,
379     ulong_t);
380 
381 /*
382  * Initialize segvn data structures
383  */
384 void
385 segvn_init(void)
386 {
387 	uint_t maxszc;
388 	uint_t szc;
389 	size_t pgsz;
390 
391 	segvn_cache = kmem_cache_create("segvn_cache",
392 	    sizeof (struct segvn_data), 0,
393 	    segvn_cache_constructor, segvn_cache_destructor, NULL,
394 	    NULL, NULL, 0);
395 
396 	if (segvn_lpg_disable == 0) {
397 		szc = maxszc = page_num_pagesizes() - 1;
398 		if (szc == 0) {
399 			segvn_lpg_disable = 1;
400 		}
401 		if (page_get_pagesize(0) != PAGESIZE) {
402 			panic("segvn_init: bad szc 0");
403 			/*NOTREACHED*/
404 		}
405 		while (szc != 0) {
406 			pgsz = page_get_pagesize(szc);
407 			if (pgsz <= PAGESIZE || !IS_P2ALIGNED(pgsz, pgsz)) {
408 				panic("segvn_init: bad szc %d", szc);
409 				/*NOTREACHED*/
410 			}
411 			szc--;
412 		}
413 		if (segvn_maxpgszc == 0 || segvn_maxpgszc > maxszc)
414 			segvn_maxpgszc = maxszc;
415 	}
416 
417 	if (segvn_maxpgszc) {
418 		segvn_szc_cache = (struct kmem_cache **)kmem_alloc(
419 		    (segvn_maxpgszc + 1) * sizeof (struct kmem_cache *),
420 		    KM_SLEEP);
421 	}
422 
423 	for (szc = 1; szc <= segvn_maxpgszc; szc++) {
424 		char	str[32];
425 
426 		(void) sprintf(str, "segvn_szc_cache%d", szc);
427 		segvn_szc_cache[szc] = kmem_cache_create(str,
428 		    page_get_pagecnt(szc) * sizeof (page_t *), 0,
429 		    NULL, NULL, NULL, NULL, NULL, KMC_NODEBUG);
430 	}
431 
432 
433 	if (segvn_use_regions && !hat_supported(HAT_SHARED_REGIONS, NULL))
434 		segvn_use_regions = 0;
435 
436 	/*
437 	 * For now shared regions and text replication segvn support
438 	 * are mutually exclusive. This is acceptable because
439 	 * currently significant benefit from text replication was
440 	 * only observed on AMD64 NUMA platforms (due to relatively
441 	 * small L2$ size) and currently we don't support shared
442 	 * regions on x86.
443 	 */
444 	if (segvn_use_regions && !segvn_disable_textrepl) {
445 		segvn_disable_textrepl = 1;
446 	}
447 
448 #if defined(_LP64)
449 	if (lgrp_optimizations() && textrepl_size_thresh != (size_t)-1 &&
450 	    !segvn_disable_textrepl) {
451 		ulong_t i;
452 		size_t hsz = svntr_hashtab_sz * sizeof (svntr_bucket_t);
453 
454 		svntr_cache = kmem_cache_create("svntr_cache",
455 		    sizeof (svntr_t), 0, svntr_cache_constructor, NULL,
456 		    NULL, NULL, NULL, 0);
457 		svntr_hashtab = kmem_zalloc(hsz, KM_SLEEP);
458 		for (i = 0; i < svntr_hashtab_sz; i++) {
459 			mutex_init(&svntr_hashtab[i].tr_lock,  NULL,
460 			    MUTEX_DEFAULT, NULL);
461 		}
462 		segvn_textrepl_max_bytes = ptob(physmem) /
463 		    segvn_textrepl_max_bytes_factor;
464 		segvn_textrepl_stats = kmem_zalloc(NCPU *
465 		    sizeof (svntr_stats_t), KM_SLEEP);
466 		sema_init(&segvn_trasync_sem, 0, NULL, SEMA_DEFAULT, NULL);
467 		(void) thread_create(NULL, 0, segvn_trasync_thread,
468 		    NULL, 0, &p0, TS_RUN, minclsyspri);
469 	}
470 #endif
471 
472 	if (!ISP2(segvn_pglock_comb_balign) ||
473 	    segvn_pglock_comb_balign < PAGESIZE) {
474 		segvn_pglock_comb_balign = 1UL << 16; /* 64K */
475 	}
476 	segvn_pglock_comb_bshift = highbit(segvn_pglock_comb_balign) - 1;
477 	segvn_pglock_comb_palign = btop(segvn_pglock_comb_balign);
478 }
479 
480 #define	SEGVN_PAGEIO	((void *)0x1)
481 #define	SEGVN_NOPAGEIO	((void *)0x2)
482 
483 static void
484 segvn_setvnode_mpss(vnode_t *vp)
485 {
486 	int err;
487 
488 	ASSERT(vp->v_mpssdata == NULL ||
489 	    vp->v_mpssdata == SEGVN_PAGEIO ||
490 	    vp->v_mpssdata == SEGVN_NOPAGEIO);
491 
492 	if (vp->v_mpssdata == NULL) {
493 		if (vn_vmpss_usepageio(vp)) {
494 			err = VOP_PAGEIO(vp, (page_t *)NULL,
495 			    (u_offset_t)0, 0, 0, CRED(), NULL);
496 		} else {
497 			err = ENOSYS;
498 		}
499 		/*
500 		 * set v_mpssdata just once per vnode life
501 		 * so that it never changes.
502 		 */
503 		mutex_enter(&vp->v_lock);
504 		if (vp->v_mpssdata == NULL) {
505 			if (err == EINVAL) {
506 				vp->v_mpssdata = SEGVN_PAGEIO;
507 			} else {
508 				vp->v_mpssdata = SEGVN_NOPAGEIO;
509 			}
510 		}
511 		mutex_exit(&vp->v_lock);
512 	}
513 }
514 
515 int
516 segvn_create(struct seg *seg, void *argsp)
517 {
518 	struct segvn_crargs *a = (struct segvn_crargs *)argsp;
519 	struct segvn_data *svd;
520 	size_t swresv = 0;
521 	struct cred *cred;
522 	struct anon_map *amp;
523 	int error = 0;
524 	size_t pgsz;
525 	lgrp_mem_policy_t mpolicy = LGRP_MEM_POLICY_DEFAULT;
526 	int use_rgn = 0;
527 	int trok = 0;
528 
529 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
530 
531 	if (a->type != MAP_PRIVATE && a->type != MAP_SHARED) {
532 		panic("segvn_create type");
533 		/*NOTREACHED*/
534 	}
535 
536 	/*
537 	 * Check arguments.  If a shared anon structure is given then
538 	 * it is illegal to also specify a vp.
539 	 */
540 	if (a->amp != NULL && a->vp != NULL) {
541 		panic("segvn_create anon_map");
542 		/*NOTREACHED*/
543 	}
544 
545 	if (a->type == MAP_PRIVATE && (a->flags & MAP_TEXT) &&
546 	    a->vp != NULL && a->prot == (PROT_USER | PROT_READ | PROT_EXEC) &&
547 	    segvn_use_regions) {
548 		use_rgn = 1;
549 	}
550 
551 	/* MAP_NORESERVE on a MAP_SHARED segment is meaningless. */
552 	if (a->type == MAP_SHARED)
553 		a->flags &= ~MAP_NORESERVE;
554 
555 	if (a->szc != 0) {
556 		if (segvn_lpg_disable != 0 || (a->szc == AS_MAP_NO_LPOOB) ||
557 		    (a->amp != NULL && a->type == MAP_PRIVATE) ||
558 		    (a->flags & MAP_NORESERVE) || seg->s_as == &kas) {
559 			a->szc = 0;
560 		} else {
561 			if (a->szc > segvn_maxpgszc)
562 				a->szc = segvn_maxpgszc;
563 			pgsz = page_get_pagesize(a->szc);
564 			if (!IS_P2ALIGNED(seg->s_base, pgsz) ||
565 			    !IS_P2ALIGNED(seg->s_size, pgsz)) {
566 				a->szc = 0;
567 			} else if (a->vp != NULL) {
568 				if (IS_SWAPFSVP(a->vp) || VN_ISKAS(a->vp)) {
569 					/*
570 					 * paranoid check.
571 					 * hat_page_demote() is not supported
572 					 * on swapfs pages.
573 					 */
574 					a->szc = 0;
575 				} else if (map_addr_vacalign_check(seg->s_base,
576 				    a->offset & PAGEMASK)) {
577 					a->szc = 0;
578 				}
579 			} else if (a->amp != NULL) {
580 				pgcnt_t anum = btopr(a->offset);
581 				pgcnt_t pgcnt = page_get_pagecnt(a->szc);
582 				if (!IS_P2ALIGNED(anum, pgcnt)) {
583 					a->szc = 0;
584 				}
585 			}
586 		}
587 	}
588 
589 	/*
590 	 * If segment may need private pages, reserve them now.
591 	 */
592 	if (!(a->flags & MAP_NORESERVE) && ((a->vp == NULL && a->amp == NULL) ||
593 	    (a->type == MAP_PRIVATE && (a->prot & PROT_WRITE)))) {
594 		if (anon_resv_zone(seg->s_size,
595 		    seg->s_as->a_proc->p_zone) == 0)
596 			return (EAGAIN);
597 		swresv = seg->s_size;
598 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
599 		    seg, swresv, 1);
600 	}
601 
602 	/*
603 	 * Reserve any mapping structures that may be required.
604 	 *
605 	 * Don't do it for segments that may use regions. It's currently a
606 	 * noop in the hat implementations anyway.
607 	 */
608 	if (!use_rgn) {
609 		hat_map(seg->s_as->a_hat, seg->s_base, seg->s_size, HAT_MAP);
610 	}
611 
612 	if (a->cred) {
613 		cred = a->cred;
614 		crhold(cred);
615 	} else {
616 		crhold(cred = CRED());
617 	}
618 
619 	/* Inform the vnode of the new mapping */
620 	if (a->vp != NULL) {
621 		error = VOP_ADDMAP(a->vp, a->offset & PAGEMASK,
622 		    seg->s_as, seg->s_base, seg->s_size, a->prot,
623 		    a->maxprot, a->type, cred, NULL);
624 		if (error) {
625 			if (swresv != 0) {
626 				anon_unresv_zone(swresv,
627 				    seg->s_as->a_proc->p_zone);
628 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
629 				    "anon proc:%p %lu %u", seg, swresv, 0);
630 			}
631 			crfree(cred);
632 			if (!use_rgn) {
633 				hat_unload(seg->s_as->a_hat, seg->s_base,
634 				    seg->s_size, HAT_UNLOAD_UNMAP);
635 			}
636 			return (error);
637 		}
638 		/*
639 		 * svntr_hashtab will be NULL if we support shared regions.
640 		 */
641 		trok = ((a->flags & MAP_TEXT) &&
642 		    (seg->s_size > textrepl_size_thresh ||
643 		    (a->flags & _MAP_TEXTREPL)) &&
644 		    lgrp_optimizations() && svntr_hashtab != NULL &&
645 		    a->type == MAP_PRIVATE && swresv == 0 &&
646 		    !(a->flags & MAP_NORESERVE) &&
647 		    seg->s_as != &kas && a->vp->v_type == VREG);
648 
649 		ASSERT(!trok || !use_rgn);
650 	}
651 
652 	/*
653 	 * MAP_NORESERVE mappings don't count towards the VSZ of a process
654 	 * until we fault the pages in.
655 	 */
656 	if ((a->vp == NULL || a->vp->v_type != VREG) &&
657 	    a->flags & MAP_NORESERVE) {
658 		seg->s_as->a_resvsize -= seg->s_size;
659 	}
660 
661 	/*
662 	 * If more than one segment in the address space, and they're adjacent
663 	 * virtually, try to concatenate them.  Don't concatenate if an
664 	 * explicit anon_map structure was supplied (e.g., SystemV shared
665 	 * memory) or if we'll use text replication for this segment.
666 	 */
667 	if (a->amp == NULL && !use_rgn && !trok) {
668 		struct seg *pseg, *nseg;
669 		struct segvn_data *psvd, *nsvd;
670 		lgrp_mem_policy_t ppolicy, npolicy;
671 		uint_t	lgrp_mem_policy_flags = 0;
672 		extern lgrp_mem_policy_t lgrp_mem_default_policy;
673 
674 		/*
675 		 * Memory policy flags (lgrp_mem_policy_flags) is valid when
676 		 * extending stack/heap segments.
677 		 */
678 		if ((a->vp == NULL) && (a->type == MAP_PRIVATE) &&
679 		    !(a->flags & MAP_NORESERVE) && (seg->s_as != &kas)) {
680 			lgrp_mem_policy_flags = a->lgrp_mem_policy_flags;
681 		} else {
682 			/*
683 			 * Get policy when not extending it from another segment
684 			 */
685 			mpolicy = lgrp_mem_policy_default(seg->s_size, a->type);
686 		}
687 
688 		/*
689 		 * First, try to concatenate the previous and new segments
690 		 */
691 		pseg = AS_SEGPREV(seg->s_as, seg);
692 		if (pseg != NULL &&
693 		    pseg->s_base + pseg->s_size == seg->s_base &&
694 		    pseg->s_ops == &segvn_ops) {
695 			/*
696 			 * Get memory allocation policy from previous segment.
697 			 * When extension is specified (e.g. for heap) apply
698 			 * this policy to the new segment regardless of the
699 			 * outcome of segment concatenation.  Extension occurs
700 			 * for non-default policy otherwise default policy is
701 			 * used and is based on extended segment size.
702 			 */
703 			psvd = (struct segvn_data *)pseg->s_data;
704 			ppolicy = psvd->policy_info.mem_policy;
705 			if (lgrp_mem_policy_flags ==
706 			    LGRP_MP_FLAG_EXTEND_UP) {
707 				if (ppolicy != lgrp_mem_default_policy) {
708 					mpolicy = ppolicy;
709 				} else {
710 					mpolicy = lgrp_mem_policy_default(
711 					    pseg->s_size + seg->s_size,
712 					    a->type);
713 				}
714 			}
715 
716 			if (mpolicy == ppolicy &&
717 			    (pseg->s_size + seg->s_size <=
718 			    segvn_comb_thrshld || psvd->amp == NULL) &&
719 			    segvn_extend_prev(pseg, seg, a, swresv) == 0) {
720 				/*
721 				 * success! now try to concatenate
722 				 * with following seg
723 				 */
724 				crfree(cred);
725 				nseg = AS_SEGNEXT(pseg->s_as, pseg);
726 				if (nseg != NULL &&
727 				    nseg != pseg &&
728 				    nseg->s_ops == &segvn_ops &&
729 				    pseg->s_base + pseg->s_size ==
730 				    nseg->s_base)
731 					(void) segvn_concat(pseg, nseg, 0);
732 				ASSERT(pseg->s_szc == 0 ||
733 				    (a->szc == pseg->s_szc &&
734 				    IS_P2ALIGNED(pseg->s_base, pgsz) &&
735 				    IS_P2ALIGNED(pseg->s_size, pgsz)));
736 				return (0);
737 			}
738 		}
739 
740 		/*
741 		 * Failed, so try to concatenate with following seg
742 		 */
743 		nseg = AS_SEGNEXT(seg->s_as, seg);
744 		if (nseg != NULL &&
745 		    seg->s_base + seg->s_size == nseg->s_base &&
746 		    nseg->s_ops == &segvn_ops) {
747 			/*
748 			 * Get memory allocation policy from next segment.
749 			 * When extension is specified (e.g. for stack) apply
750 			 * this policy to the new segment regardless of the
751 			 * outcome of segment concatenation.  Extension occurs
752 			 * for non-default policy otherwise default policy is
753 			 * used and is based on extended segment size.
754 			 */
755 			nsvd = (struct segvn_data *)nseg->s_data;
756 			npolicy = nsvd->policy_info.mem_policy;
757 			if (lgrp_mem_policy_flags ==
758 			    LGRP_MP_FLAG_EXTEND_DOWN) {
759 				if (npolicy != lgrp_mem_default_policy) {
760 					mpolicy = npolicy;
761 				} else {
762 					mpolicy = lgrp_mem_policy_default(
763 					    nseg->s_size + seg->s_size,
764 					    a->type);
765 				}
766 			}
767 
768 			if (mpolicy == npolicy &&
769 			    segvn_extend_next(seg, nseg, a, swresv) == 0) {
770 				crfree(cred);
771 				ASSERT(nseg->s_szc == 0 ||
772 				    (a->szc == nseg->s_szc &&
773 				    IS_P2ALIGNED(nseg->s_base, pgsz) &&
774 				    IS_P2ALIGNED(nseg->s_size, pgsz)));
775 				return (0);
776 			}
777 		}
778 	}
779 
780 	if (a->vp != NULL) {
781 		VN_HOLD(a->vp);
782 		if (a->type == MAP_SHARED)
783 			lgrp_shm_policy_init(NULL, a->vp);
784 	}
785 	svd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
786 
787 	seg->s_ops = &segvn_ops;
788 	seg->s_data = (void *)svd;
789 	seg->s_szc = a->szc;
790 
791 	svd->seg = seg;
792 	svd->vp = a->vp;
793 	/*
794 	 * Anonymous mappings have no backing file so the offset is meaningless.
795 	 */
796 	svd->offset = a->vp ? (a->offset & PAGEMASK) : 0;
797 	svd->prot = a->prot;
798 	svd->maxprot = a->maxprot;
799 	svd->pageprot = 0;
800 	svd->type = a->type;
801 	svd->vpage = NULL;
802 	svd->cred = cred;
803 	svd->advice = MADV_NORMAL;
804 	svd->pageadvice = 0;
805 	svd->flags = (ushort_t)a->flags;
806 	svd->softlockcnt = 0;
807 	svd->softlockcnt_sbase = 0;
808 	svd->softlockcnt_send = 0;
809 	svd->rcookie = HAT_INVALID_REGION_COOKIE;
810 	svd->pageswap = 0;
811 
812 	if (a->szc != 0 && a->vp != NULL) {
813 		segvn_setvnode_mpss(a->vp);
814 	}
815 	if (svd->type == MAP_SHARED && svd->vp != NULL &&
816 	    (svd->vp->v_flag & VVMEXEC) && (svd->prot & PROT_WRITE)) {
817 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
818 		segvn_inval_trcache(svd->vp);
819 	}
820 
821 	amp = a->amp;
822 	if ((svd->amp = amp) == NULL) {
823 		svd->anon_index = 0;
824 		if (svd->type == MAP_SHARED) {
825 			svd->swresv = 0;
826 			/*
827 			 * Shared mappings to a vp need no other setup.
828 			 * If we have a shared mapping to an anon_map object
829 			 * which hasn't been allocated yet,  allocate the
830 			 * struct now so that it will be properly shared
831 			 * by remembering the swap reservation there.
832 			 */
833 			if (a->vp == NULL) {
834 				svd->amp = anonmap_alloc(seg->s_size, swresv,
835 				    ANON_SLEEP);
836 				svd->amp->a_szc = seg->s_szc;
837 			}
838 		} else {
839 			/*
840 			 * Private mapping (with or without a vp).
841 			 * Allocate anon_map when needed.
842 			 */
843 			svd->swresv = swresv;
844 		}
845 	} else {
846 		pgcnt_t anon_num;
847 
848 		/*
849 		 * Mapping to an existing anon_map structure without a vp.
850 		 * For now we will insure that the segment size isn't larger
851 		 * than the size - offset gives us.  Later on we may wish to
852 		 * have the anon array dynamically allocated itself so that
853 		 * we don't always have to allocate all the anon pointer slots.
854 		 * This of course involves adding extra code to check that we
855 		 * aren't trying to use an anon pointer slot beyond the end
856 		 * of the currently allocated anon array.
857 		 */
858 		if ((amp->size - a->offset) < seg->s_size) {
859 			panic("segvn_create anon_map size");
860 			/*NOTREACHED*/
861 		}
862 
863 		anon_num = btopr(a->offset);
864 
865 		if (a->type == MAP_SHARED) {
866 			/*
867 			 * SHARED mapping to a given anon_map.
868 			 */
869 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
870 			amp->refcnt++;
871 			if (a->szc > amp->a_szc) {
872 				amp->a_szc = a->szc;
873 			}
874 			ANON_LOCK_EXIT(&amp->a_rwlock);
875 			svd->anon_index = anon_num;
876 			svd->swresv = 0;
877 		} else {
878 			/*
879 			 * PRIVATE mapping to a given anon_map.
880 			 * Make sure that all the needed anon
881 			 * structures are created (so that we will
882 			 * share the underlying pages if nothing
883 			 * is written by this mapping) and then
884 			 * duplicate the anon array as is done
885 			 * when a privately mapped segment is dup'ed.
886 			 */
887 			struct anon *ap;
888 			caddr_t addr;
889 			caddr_t eaddr;
890 			ulong_t	anon_idx;
891 			int hat_flag = HAT_LOAD;
892 
893 			if (svd->flags & MAP_TEXT) {
894 				hat_flag |= HAT_LOAD_TEXT;
895 			}
896 
897 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
898 			svd->amp->a_szc = seg->s_szc;
899 			svd->anon_index = 0;
900 			svd->swresv = swresv;
901 
902 			/*
903 			 * Prevent 2 threads from allocating anon
904 			 * slots simultaneously.
905 			 */
906 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
907 			eaddr = seg->s_base + seg->s_size;
908 
909 			for (anon_idx = anon_num, addr = seg->s_base;
910 			    addr < eaddr; addr += PAGESIZE, anon_idx++) {
911 				page_t *pp;
912 
913 				if ((ap = anon_get_ptr(amp->ahp,
914 				    anon_idx)) != NULL)
915 					continue;
916 
917 				/*
918 				 * Allocate the anon struct now.
919 				 * Might as well load up translation
920 				 * to the page while we're at it...
921 				 */
922 				pp = anon_zero(seg, addr, &ap, cred);
923 				if (ap == NULL || pp == NULL) {
924 					panic("segvn_create anon_zero");
925 					/*NOTREACHED*/
926 				}
927 
928 				/*
929 				 * Re-acquire the anon_map lock and
930 				 * initialize the anon array entry.
931 				 */
932 				ASSERT(anon_get_ptr(amp->ahp,
933 				    anon_idx) == NULL);
934 				(void) anon_set_ptr(amp->ahp, anon_idx, ap,
935 				    ANON_SLEEP);
936 
937 				ASSERT(seg->s_szc == 0);
938 				ASSERT(!IS_VMODSORT(pp->p_vnode));
939 
940 				ASSERT(use_rgn == 0);
941 				hat_memload(seg->s_as->a_hat, addr, pp,
942 				    svd->prot & ~PROT_WRITE, hat_flag);
943 
944 				page_unlock(pp);
945 			}
946 			ASSERT(seg->s_szc == 0);
947 			anon_dup(amp->ahp, anon_num, svd->amp->ahp,
948 			    0, seg->s_size);
949 			ANON_LOCK_EXIT(&amp->a_rwlock);
950 		}
951 	}
952 
953 	/*
954 	 * Set default memory allocation policy for segment
955 	 *
956 	 * Always set policy for private memory at least for initialization
957 	 * even if this is a shared memory segment
958 	 */
959 	(void) lgrp_privm_policy_set(mpolicy, &svd->policy_info, seg->s_size);
960 
961 	if (svd->type == MAP_SHARED)
962 		(void) lgrp_shm_policy_set(mpolicy, svd->amp, svd->anon_index,
963 		    svd->vp, svd->offset, seg->s_size);
964 
965 	if (use_rgn) {
966 		ASSERT(!trok);
967 		ASSERT(svd->amp == NULL);
968 		svd->rcookie = hat_join_region(seg->s_as->a_hat, seg->s_base,
969 		    seg->s_size, (void *)svd->vp, svd->offset, svd->prot,
970 		    (uchar_t)seg->s_szc, segvn_hat_rgn_unload_callback,
971 		    HAT_REGION_TEXT);
972 	}
973 
974 	ASSERT(!trok || !(svd->prot & PROT_WRITE));
975 	svd->tr_state = trok ? SEGVN_TR_INIT : SEGVN_TR_OFF;
976 
977 	return (0);
978 }
979 
980 /*
981  * Concatenate two existing segments, if possible.
982  * Return 0 on success, -1 if two segments are not compatible
983  * or -2 on memory allocation failure.
984  * If amp_cat == 1 then try and concat segments with anon maps
985  */
986 static int
987 segvn_concat(struct seg *seg1, struct seg *seg2, int amp_cat)
988 {
989 	struct segvn_data *svd1 = seg1->s_data;
990 	struct segvn_data *svd2 = seg2->s_data;
991 	struct anon_map *amp1 = svd1->amp;
992 	struct anon_map *amp2 = svd2->amp;
993 	struct vpage *vpage1 = svd1->vpage;
994 	struct vpage *vpage2 = svd2->vpage, *nvpage = NULL;
995 	size_t size, nvpsize;
996 	pgcnt_t npages1, npages2;
997 
998 	ASSERT(seg1->s_as && seg2->s_as && seg1->s_as == seg2->s_as);
999 	ASSERT(AS_WRITE_HELD(seg1->s_as, &seg1->s_as->a_lock));
1000 	ASSERT(seg1->s_ops == seg2->s_ops);
1001 
1002 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie) ||
1003 	    HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1004 		return (-1);
1005 	}
1006 
1007 	/* both segments exist, try to merge them */
1008 #define	incompat(x)	(svd1->x != svd2->x)
1009 	if (incompat(vp) || incompat(maxprot) ||
1010 	    (!svd1->pageadvice && !svd2->pageadvice && incompat(advice)) ||
1011 	    (!svd1->pageprot && !svd2->pageprot && incompat(prot)) ||
1012 	    incompat(type) || incompat(cred) || incompat(flags) ||
1013 	    seg1->s_szc != seg2->s_szc || incompat(policy_info.mem_policy) ||
1014 	    (svd2->softlockcnt > 0) || svd1->softlockcnt_send > 0)
1015 		return (-1);
1016 #undef incompat
1017 
1018 	/*
1019 	 * vp == NULL implies zfod, offset doesn't matter
1020 	 */
1021 	if (svd1->vp != NULL &&
1022 	    svd1->offset + seg1->s_size != svd2->offset) {
1023 		return (-1);
1024 	}
1025 
1026 	/*
1027 	 * Don't concatenate if either segment uses text replication.
1028 	 */
1029 	if (svd1->tr_state != SEGVN_TR_OFF || svd2->tr_state != SEGVN_TR_OFF) {
1030 		return (-1);
1031 	}
1032 
1033 	/*
1034 	 * Fail early if we're not supposed to concatenate
1035 	 * segments with non NULL amp.
1036 	 */
1037 	if (amp_cat == 0 && (amp1 != NULL || amp2 != NULL)) {
1038 		return (-1);
1039 	}
1040 
1041 	if (svd1->vp == NULL && svd1->type == MAP_SHARED) {
1042 		if (amp1 != amp2) {
1043 			return (-1);
1044 		}
1045 		if (amp1 != NULL && svd1->anon_index + btop(seg1->s_size) !=
1046 		    svd2->anon_index) {
1047 			return (-1);
1048 		}
1049 		ASSERT(amp1 == NULL || amp1->refcnt >= 2);
1050 	}
1051 
1052 	/*
1053 	 * If either seg has vpages, create a new merged vpage array.
1054 	 */
1055 	if (vpage1 != NULL || vpage2 != NULL) {
1056 		struct vpage *vp, *evp;
1057 
1058 		npages1 = seg_pages(seg1);
1059 		npages2 = seg_pages(seg2);
1060 		nvpsize = vpgtob(npages1 + npages2);
1061 
1062 		if ((nvpage = kmem_zalloc(nvpsize, KM_NOSLEEP)) == NULL) {
1063 			return (-2);
1064 		}
1065 
1066 		if (vpage1 != NULL) {
1067 			bcopy(vpage1, nvpage, vpgtob(npages1));
1068 		} else {
1069 			evp = nvpage + npages1;
1070 			for (vp = nvpage; vp < evp; vp++) {
1071 				VPP_SETPROT(vp, svd1->prot);
1072 				VPP_SETADVICE(vp, svd1->advice);
1073 			}
1074 		}
1075 
1076 		if (vpage2 != NULL) {
1077 			bcopy(vpage2, nvpage + npages1, vpgtob(npages2));
1078 		} else {
1079 			evp = nvpage + npages1 + npages2;
1080 			for (vp = nvpage + npages1; vp < evp; vp++) {
1081 				VPP_SETPROT(vp, svd2->prot);
1082 				VPP_SETADVICE(vp, svd2->advice);
1083 			}
1084 		}
1085 
1086 		if (svd2->pageswap && (!svd1->pageswap && svd1->swresv)) {
1087 			ASSERT(svd1->swresv == seg1->s_size);
1088 			ASSERT(!(svd1->flags & MAP_NORESERVE));
1089 			ASSERT(!(svd2->flags & MAP_NORESERVE));
1090 			evp = nvpage + npages1;
1091 			for (vp = nvpage; vp < evp; vp++) {
1092 				VPP_SETSWAPRES(vp);
1093 			}
1094 		}
1095 
1096 		if (svd1->pageswap && (!svd2->pageswap && svd2->swresv)) {
1097 			ASSERT(svd2->swresv == seg2->s_size);
1098 			ASSERT(!(svd1->flags & MAP_NORESERVE));
1099 			ASSERT(!(svd2->flags & MAP_NORESERVE));
1100 			vp = nvpage + npages1;
1101 			evp = vp + npages2;
1102 			for (; vp < evp; vp++) {
1103 				VPP_SETSWAPRES(vp);
1104 			}
1105 		}
1106 	}
1107 	ASSERT((vpage1 != NULL || vpage2 != NULL) ||
1108 	    (svd1->pageswap == 0 && svd2->pageswap == 0));
1109 
1110 	/*
1111 	 * If either segment has private pages, create a new merged anon
1112 	 * array. If mergeing shared anon segments just decrement anon map's
1113 	 * refcnt.
1114 	 */
1115 	if (amp1 != NULL && svd1->type == MAP_SHARED) {
1116 		ASSERT(amp1 == amp2 && svd1->vp == NULL);
1117 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1118 		ASSERT(amp1->refcnt >= 2);
1119 		amp1->refcnt--;
1120 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1121 		svd2->amp = NULL;
1122 	} else if (amp1 != NULL || amp2 != NULL) {
1123 		struct anon_hdr *nahp;
1124 		struct anon_map *namp = NULL;
1125 		size_t asize;
1126 
1127 		ASSERT(svd1->type == MAP_PRIVATE);
1128 
1129 		asize = seg1->s_size + seg2->s_size;
1130 		if ((nahp = anon_create(btop(asize), ANON_NOSLEEP)) == NULL) {
1131 			if (nvpage != NULL) {
1132 				kmem_free(nvpage, nvpsize);
1133 			}
1134 			return (-2);
1135 		}
1136 		if (amp1 != NULL) {
1137 			/*
1138 			 * XXX anon rwlock is not really needed because
1139 			 * this is a private segment and we are writers.
1140 			 */
1141 			ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1142 			ASSERT(amp1->refcnt == 1);
1143 			if (anon_copy_ptr(amp1->ahp, svd1->anon_index,
1144 			    nahp, 0, btop(seg1->s_size), ANON_NOSLEEP)) {
1145 				anon_release(nahp, btop(asize));
1146 				ANON_LOCK_EXIT(&amp1->a_rwlock);
1147 				if (nvpage != NULL) {
1148 					kmem_free(nvpage, nvpsize);
1149 				}
1150 				return (-2);
1151 			}
1152 		}
1153 		if (amp2 != NULL) {
1154 			ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1155 			ASSERT(amp2->refcnt == 1);
1156 			if (anon_copy_ptr(amp2->ahp, svd2->anon_index,
1157 			    nahp, btop(seg1->s_size), btop(seg2->s_size),
1158 			    ANON_NOSLEEP)) {
1159 				anon_release(nahp, btop(asize));
1160 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1161 				if (amp1 != NULL) {
1162 					ANON_LOCK_EXIT(&amp1->a_rwlock);
1163 				}
1164 				if (nvpage != NULL) {
1165 					kmem_free(nvpage, nvpsize);
1166 				}
1167 				return (-2);
1168 			}
1169 		}
1170 		if (amp1 != NULL) {
1171 			namp = amp1;
1172 			anon_release(amp1->ahp, btop(amp1->size));
1173 		}
1174 		if (amp2 != NULL) {
1175 			if (namp == NULL) {
1176 				ASSERT(amp1 == NULL);
1177 				namp = amp2;
1178 				anon_release(amp2->ahp, btop(amp2->size));
1179 			} else {
1180 				amp2->refcnt--;
1181 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1182 				anonmap_free(amp2);
1183 			}
1184 			svd2->amp = NULL; /* needed for seg_free */
1185 		}
1186 		namp->ahp = nahp;
1187 		namp->size = asize;
1188 		svd1->amp = namp;
1189 		svd1->anon_index = 0;
1190 		ANON_LOCK_EXIT(&namp->a_rwlock);
1191 	}
1192 	/*
1193 	 * Now free the old vpage structures.
1194 	 */
1195 	if (nvpage != NULL) {
1196 		if (vpage1 != NULL) {
1197 			kmem_free(vpage1, vpgtob(npages1));
1198 		}
1199 		if (vpage2 != NULL) {
1200 			svd2->vpage = NULL;
1201 			kmem_free(vpage2, vpgtob(npages2));
1202 		}
1203 		if (svd2->pageprot) {
1204 			svd1->pageprot = 1;
1205 		}
1206 		if (svd2->pageadvice) {
1207 			svd1->pageadvice = 1;
1208 		}
1209 		if (svd2->pageswap) {
1210 			svd1->pageswap = 1;
1211 		}
1212 		svd1->vpage = nvpage;
1213 	}
1214 
1215 	/* all looks ok, merge segments */
1216 	svd1->swresv += svd2->swresv;
1217 	svd2->swresv = 0;  /* so seg_free doesn't release swap space */
1218 	size = seg2->s_size;
1219 	seg_free(seg2);
1220 	seg1->s_size += size;
1221 	return (0);
1222 }
1223 
1224 /*
1225  * Extend the previous segment (seg1) to include the
1226  * new segment (seg2 + a), if possible.
1227  * Return 0 on success.
1228  */
1229 static int
1230 segvn_extend_prev(seg1, seg2, a, swresv)
1231 	struct seg *seg1, *seg2;
1232 	struct segvn_crargs *a;
1233 	size_t swresv;
1234 {
1235 	struct segvn_data *svd1 = (struct segvn_data *)seg1->s_data;
1236 	size_t size;
1237 	struct anon_map *amp1;
1238 	struct vpage *new_vpage;
1239 
1240 	/*
1241 	 * We don't need any segment level locks for "segvn" data
1242 	 * since the address space is "write" locked.
1243 	 */
1244 	ASSERT(seg1->s_as && AS_WRITE_HELD(seg1->s_as, &seg1->s_as->a_lock));
1245 
1246 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie)) {
1247 		return (-1);
1248 	}
1249 
1250 	/* second segment is new, try to extend first */
1251 	/* XXX - should also check cred */
1252 	if (svd1->vp != a->vp || svd1->maxprot != a->maxprot ||
1253 	    (!svd1->pageprot && (svd1->prot != a->prot)) ||
1254 	    svd1->type != a->type || svd1->flags != a->flags ||
1255 	    seg1->s_szc != a->szc || svd1->softlockcnt_send > 0)
1256 		return (-1);
1257 
1258 	/* vp == NULL implies zfod, offset doesn't matter */
1259 	if (svd1->vp != NULL &&
1260 	    svd1->offset + seg1->s_size != (a->offset & PAGEMASK))
1261 		return (-1);
1262 
1263 	if (svd1->tr_state != SEGVN_TR_OFF) {
1264 		return (-1);
1265 	}
1266 
1267 	amp1 = svd1->amp;
1268 	if (amp1) {
1269 		pgcnt_t newpgs;
1270 
1271 		/*
1272 		 * Segment has private pages, can data structures
1273 		 * be expanded?
1274 		 *
1275 		 * Acquire the anon_map lock to prevent it from changing,
1276 		 * if it is shared.  This ensures that the anon_map
1277 		 * will not change while a thread which has a read/write
1278 		 * lock on an address space references it.
1279 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1280 		 * is 1.
1281 		 *
1282 		 * Can't grow a MAP_SHARED segment with an anonmap because
1283 		 * there may be existing anon slots where we want to extend
1284 		 * the segment and we wouldn't know what to do with them
1285 		 * (e.g., for tmpfs right thing is to just leave them there,
1286 		 * for /dev/zero they should be cleared out).
1287 		 */
1288 		if (svd1->type == MAP_SHARED)
1289 			return (-1);
1290 
1291 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1292 		if (amp1->refcnt > 1) {
1293 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1294 			return (-1);
1295 		}
1296 		newpgs = anon_grow(amp1->ahp, &svd1->anon_index,
1297 		    btop(seg1->s_size), btop(seg2->s_size), ANON_NOSLEEP);
1298 
1299 		if (newpgs == 0) {
1300 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1301 			return (-1);
1302 		}
1303 		amp1->size = ptob(newpgs);
1304 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1305 	}
1306 	if (svd1->vpage != NULL) {
1307 		struct vpage *vp, *evp;
1308 		new_vpage =
1309 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1310 			KM_NOSLEEP);
1311 		if (new_vpage == NULL)
1312 			return (-1);
1313 		bcopy(svd1->vpage, new_vpage, vpgtob(seg_pages(seg1)));
1314 		kmem_free(svd1->vpage, vpgtob(seg_pages(seg1)));
1315 		svd1->vpage = new_vpage;
1316 
1317 		vp = new_vpage + seg_pages(seg1);
1318 		evp = vp + seg_pages(seg2);
1319 		for (; vp < evp; vp++)
1320 			VPP_SETPROT(vp, a->prot);
1321 		if (svd1->pageswap && swresv) {
1322 			ASSERT(!(svd1->flags & MAP_NORESERVE));
1323 			ASSERT(swresv == seg2->s_size);
1324 			vp = new_vpage + seg_pages(seg1);
1325 			for (; vp < evp; vp++) {
1326 				VPP_SETSWAPRES(vp);
1327 			}
1328 		}
1329 	}
1330 	ASSERT(svd1->vpage != NULL || svd1->pageswap == 0);
1331 	size = seg2->s_size;
1332 	seg_free(seg2);
1333 	seg1->s_size += size;
1334 	svd1->swresv += swresv;
1335 	if (svd1->pageprot && (a->prot & PROT_WRITE) &&
1336 	    svd1->type == MAP_SHARED && svd1->vp != NULL &&
1337 	    (svd1->vp->v_flag & VVMEXEC)) {
1338 		ASSERT(vn_is_mapped(svd1->vp, V_WRITE));
1339 		segvn_inval_trcache(svd1->vp);
1340 	}
1341 	return (0);
1342 }
1343 
1344 /*
1345  * Extend the next segment (seg2) to include the
1346  * new segment (seg1 + a), if possible.
1347  * Return 0 on success.
1348  */
1349 static int
1350 segvn_extend_next(
1351 	struct seg *seg1,
1352 	struct seg *seg2,
1353 	struct segvn_crargs *a,
1354 	size_t swresv)
1355 {
1356 	struct segvn_data *svd2 = (struct segvn_data *)seg2->s_data;
1357 	size_t size;
1358 	struct anon_map *amp2;
1359 	struct vpage *new_vpage;
1360 
1361 	/*
1362 	 * We don't need any segment level locks for "segvn" data
1363 	 * since the address space is "write" locked.
1364 	 */
1365 	ASSERT(seg2->s_as && AS_WRITE_HELD(seg2->s_as, &seg2->s_as->a_lock));
1366 
1367 	if (HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1368 		return (-1);
1369 	}
1370 
1371 	/* first segment is new, try to extend second */
1372 	/* XXX - should also check cred */
1373 	if (svd2->vp != a->vp || svd2->maxprot != a->maxprot ||
1374 	    (!svd2->pageprot && (svd2->prot != a->prot)) ||
1375 	    svd2->type != a->type || svd2->flags != a->flags ||
1376 	    seg2->s_szc != a->szc || svd2->softlockcnt_sbase > 0)
1377 		return (-1);
1378 	/* vp == NULL implies zfod, offset doesn't matter */
1379 	if (svd2->vp != NULL &&
1380 	    (a->offset & PAGEMASK) + seg1->s_size != svd2->offset)
1381 		return (-1);
1382 
1383 	if (svd2->tr_state != SEGVN_TR_OFF) {
1384 		return (-1);
1385 	}
1386 
1387 	amp2 = svd2->amp;
1388 	if (amp2) {
1389 		pgcnt_t newpgs;
1390 
1391 		/*
1392 		 * Segment has private pages, can data structures
1393 		 * be expanded?
1394 		 *
1395 		 * Acquire the anon_map lock to prevent it from changing,
1396 		 * if it is shared.  This ensures that the anon_map
1397 		 * will not change while a thread which has a read/write
1398 		 * lock on an address space references it.
1399 		 *
1400 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1401 		 * is 1.
1402 		 */
1403 		if (svd2->type == MAP_SHARED)
1404 			return (-1);
1405 
1406 		ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1407 		if (amp2->refcnt > 1) {
1408 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1409 			return (-1);
1410 		}
1411 		newpgs = anon_grow(amp2->ahp, &svd2->anon_index,
1412 		    btop(seg2->s_size), btop(seg1->s_size),
1413 		    ANON_NOSLEEP | ANON_GROWDOWN);
1414 
1415 		if (newpgs == 0) {
1416 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1417 			return (-1);
1418 		}
1419 		amp2->size = ptob(newpgs);
1420 		ANON_LOCK_EXIT(&amp2->a_rwlock);
1421 	}
1422 	if (svd2->vpage != NULL) {
1423 		struct vpage *vp, *evp;
1424 		new_vpage =
1425 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1426 		    KM_NOSLEEP);
1427 		if (new_vpage == NULL) {
1428 			/* Not merging segments so adjust anon_index back */
1429 			if (amp2)
1430 				svd2->anon_index += seg_pages(seg1);
1431 			return (-1);
1432 		}
1433 		bcopy(svd2->vpage, new_vpage + seg_pages(seg1),
1434 		    vpgtob(seg_pages(seg2)));
1435 		kmem_free(svd2->vpage, vpgtob(seg_pages(seg2)));
1436 		svd2->vpage = new_vpage;
1437 
1438 		vp = new_vpage;
1439 		evp = vp + seg_pages(seg1);
1440 		for (; vp < evp; vp++)
1441 			VPP_SETPROT(vp, a->prot);
1442 		if (svd2->pageswap && swresv) {
1443 			ASSERT(!(svd2->flags & MAP_NORESERVE));
1444 			ASSERT(swresv == seg1->s_size);
1445 			vp = new_vpage;
1446 			for (; vp < evp; vp++) {
1447 				VPP_SETSWAPRES(vp);
1448 			}
1449 		}
1450 	}
1451 	ASSERT(svd2->vpage != NULL || svd2->pageswap == 0);
1452 	size = seg1->s_size;
1453 	seg_free(seg1);
1454 	seg2->s_size += size;
1455 	seg2->s_base -= size;
1456 	svd2->offset -= size;
1457 	svd2->swresv += swresv;
1458 	if (svd2->pageprot && (a->prot & PROT_WRITE) &&
1459 	    svd2->type == MAP_SHARED && svd2->vp != NULL &&
1460 	    (svd2->vp->v_flag & VVMEXEC)) {
1461 		ASSERT(vn_is_mapped(svd2->vp, V_WRITE));
1462 		segvn_inval_trcache(svd2->vp);
1463 	}
1464 	return (0);
1465 }
1466 
1467 static int
1468 segvn_dup(struct seg *seg, struct seg *newseg)
1469 {
1470 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1471 	struct segvn_data *newsvd;
1472 	pgcnt_t npages = seg_pages(seg);
1473 	int error = 0;
1474 	uint_t prot;
1475 	size_t len;
1476 	struct anon_map *amp;
1477 
1478 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
1479 	ASSERT(newseg->s_as->a_proc->p_parent == curproc);
1480 
1481 	/*
1482 	 * If segment has anon reserved, reserve more for the new seg.
1483 	 * For a MAP_NORESERVE segment swresv will be a count of all the
1484 	 * allocated anon slots; thus we reserve for the child as many slots
1485 	 * as the parent has allocated. This semantic prevents the child or
1486 	 * parent from dieing during a copy-on-write fault caused by trying
1487 	 * to write a shared pre-existing anon page.
1488 	 */
1489 	if ((len = svd->swresv) != 0) {
1490 		if (anon_resv(svd->swresv) == 0)
1491 			return (ENOMEM);
1492 
1493 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
1494 		    seg, len, 0);
1495 	}
1496 
1497 	newsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
1498 
1499 	newseg->s_ops = &segvn_ops;
1500 	newseg->s_data = (void *)newsvd;
1501 	newseg->s_szc = seg->s_szc;
1502 
1503 	newsvd->seg = newseg;
1504 	if ((newsvd->vp = svd->vp) != NULL) {
1505 		VN_HOLD(svd->vp);
1506 		if (svd->type == MAP_SHARED)
1507 			lgrp_shm_policy_init(NULL, svd->vp);
1508 	}
1509 	newsvd->offset = svd->offset;
1510 	newsvd->prot = svd->prot;
1511 	newsvd->maxprot = svd->maxprot;
1512 	newsvd->pageprot = svd->pageprot;
1513 	newsvd->type = svd->type;
1514 	newsvd->cred = svd->cred;
1515 	crhold(newsvd->cred);
1516 	newsvd->advice = svd->advice;
1517 	newsvd->pageadvice = svd->pageadvice;
1518 	newsvd->swresv = svd->swresv;
1519 	newsvd->pageswap = svd->pageswap;
1520 	newsvd->flags = svd->flags;
1521 	newsvd->softlockcnt = 0;
1522 	newsvd->softlockcnt_sbase = 0;
1523 	newsvd->softlockcnt_send = 0;
1524 	newsvd->policy_info = svd->policy_info;
1525 	newsvd->rcookie = HAT_INVALID_REGION_COOKIE;
1526 
1527 	if ((amp = svd->amp) == NULL || svd->tr_state == SEGVN_TR_ON) {
1528 		/*
1529 		 * Not attaching to a shared anon object.
1530 		 */
1531 		ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie) ||
1532 		    svd->tr_state == SEGVN_TR_OFF);
1533 		if (svd->tr_state == SEGVN_TR_ON) {
1534 			ASSERT(newsvd->vp != NULL && amp != NULL);
1535 			newsvd->tr_state = SEGVN_TR_INIT;
1536 		} else {
1537 			newsvd->tr_state = svd->tr_state;
1538 		}
1539 		newsvd->amp = NULL;
1540 		newsvd->anon_index = 0;
1541 	} else {
1542 		/* regions for now are only used on pure vnode segments */
1543 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
1544 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1545 		newsvd->tr_state = SEGVN_TR_OFF;
1546 		if (svd->type == MAP_SHARED) {
1547 			newsvd->amp = amp;
1548 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1549 			amp->refcnt++;
1550 			ANON_LOCK_EXIT(&amp->a_rwlock);
1551 			newsvd->anon_index = svd->anon_index;
1552 		} else {
1553 			int reclaim = 1;
1554 
1555 			/*
1556 			 * Allocate and initialize new anon_map structure.
1557 			 */
1558 			newsvd->amp = anonmap_alloc(newseg->s_size, 0,
1559 			    ANON_SLEEP);
1560 			newsvd->amp->a_szc = newseg->s_szc;
1561 			newsvd->anon_index = 0;
1562 
1563 			/*
1564 			 * We don't have to acquire the anon_map lock
1565 			 * for the new segment (since it belongs to an
1566 			 * address space that is still not associated
1567 			 * with any process), or the segment in the old
1568 			 * address space (since all threads in it
1569 			 * are stopped while duplicating the address space).
1570 			 */
1571 
1572 			/*
1573 			 * The goal of the following code is to make sure that
1574 			 * softlocked pages do not end up as copy on write
1575 			 * pages.  This would cause problems where one
1576 			 * thread writes to a page that is COW and a different
1577 			 * thread in the same process has softlocked it.  The
1578 			 * softlock lock would move away from this process
1579 			 * because the write would cause this process to get
1580 			 * a copy (without the softlock).
1581 			 *
1582 			 * The strategy here is to just break the
1583 			 * sharing on pages that could possibly be
1584 			 * softlocked.
1585 			 */
1586 retry:
1587 			if (svd->softlockcnt) {
1588 				struct anon *ap, *newap;
1589 				size_t i;
1590 				uint_t vpprot;
1591 				page_t *anon_pl[1+1], *pp;
1592 				caddr_t addr;
1593 				ulong_t old_idx = svd->anon_index;
1594 				ulong_t new_idx = 0;
1595 
1596 				/*
1597 				 * The softlock count might be non zero
1598 				 * because some pages are still stuck in the
1599 				 * cache for lazy reclaim. Flush the cache
1600 				 * now. This should drop the count to zero.
1601 				 * [or there is really I/O going on to these
1602 				 * pages]. Note, we have the writers lock so
1603 				 * nothing gets inserted during the flush.
1604 				 */
1605 				if (reclaim == 1) {
1606 					segvn_purge(seg);
1607 					reclaim = 0;
1608 					goto retry;
1609 				}
1610 				i = btopr(seg->s_size);
1611 				addr = seg->s_base;
1612 				/*
1613 				 * XXX break cow sharing using PAGESIZE
1614 				 * pages. They will be relocated into larger
1615 				 * pages at fault time.
1616 				 */
1617 				while (i-- > 0) {
1618 					if (ap = anon_get_ptr(amp->ahp,
1619 					    old_idx)) {
1620 						error = anon_getpage(&ap,
1621 						    &vpprot, anon_pl, PAGESIZE,
1622 						    seg, addr, S_READ,
1623 						    svd->cred);
1624 						if (error) {
1625 							newsvd->vpage = NULL;
1626 							goto out;
1627 						}
1628 						/*
1629 						 * prot need not be computed
1630 						 * below 'cause anon_private is
1631 						 * going to ignore it anyway
1632 						 * as child doesn't inherit
1633 						 * pagelock from parent.
1634 						 */
1635 						prot = svd->pageprot ?
1636 						    VPP_PROT(
1637 						    &svd->vpage[
1638 						    seg_page(seg, addr)])
1639 						    : svd->prot;
1640 						pp = anon_private(&newap,
1641 						    newseg, addr, prot,
1642 						    anon_pl[0],	0,
1643 						    newsvd->cred);
1644 						if (pp == NULL) {
1645 							/* no mem abort */
1646 							newsvd->vpage = NULL;
1647 							error = ENOMEM;
1648 							goto out;
1649 						}
1650 						(void) anon_set_ptr(
1651 						    newsvd->amp->ahp, new_idx,
1652 						    newap, ANON_SLEEP);
1653 						page_unlock(pp);
1654 					}
1655 					addr += PAGESIZE;
1656 					old_idx++;
1657 					new_idx++;
1658 				}
1659 			} else {	/* common case */
1660 				if (seg->s_szc != 0) {
1661 					/*
1662 					 * If at least one of anon slots of a
1663 					 * large page exists then make sure
1664 					 * all anon slots of a large page
1665 					 * exist to avoid partial cow sharing
1666 					 * of a large page in the future.
1667 					 */
1668 					anon_dup_fill_holes(amp->ahp,
1669 					    svd->anon_index, newsvd->amp->ahp,
1670 					    0, seg->s_size, seg->s_szc,
1671 					    svd->vp != NULL);
1672 				} else {
1673 					anon_dup(amp->ahp, svd->anon_index,
1674 					    newsvd->amp->ahp, 0, seg->s_size);
1675 				}
1676 
1677 				hat_clrattr(seg->s_as->a_hat, seg->s_base,
1678 				    seg->s_size, PROT_WRITE);
1679 			}
1680 		}
1681 	}
1682 	/*
1683 	 * If necessary, create a vpage structure for the new segment.
1684 	 * Do not copy any page lock indications.
1685 	 */
1686 	if (svd->vpage != NULL) {
1687 		uint_t i;
1688 		struct vpage *ovp = svd->vpage;
1689 		struct vpage *nvp;
1690 
1691 		nvp = newsvd->vpage =
1692 		    kmem_alloc(vpgtob(npages), KM_SLEEP);
1693 		for (i = 0; i < npages; i++) {
1694 			*nvp = *ovp++;
1695 			VPP_CLRPPLOCK(nvp++);
1696 		}
1697 	} else
1698 		newsvd->vpage = NULL;
1699 
1700 	/* Inform the vnode of the new mapping */
1701 	if (newsvd->vp != NULL) {
1702 		error = VOP_ADDMAP(newsvd->vp, (offset_t)newsvd->offset,
1703 		    newseg->s_as, newseg->s_base, newseg->s_size, newsvd->prot,
1704 		    newsvd->maxprot, newsvd->type, newsvd->cred, NULL);
1705 	}
1706 out:
1707 	if (error == 0 && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1708 		ASSERT(newsvd->amp == NULL);
1709 		ASSERT(newsvd->tr_state == SEGVN_TR_OFF);
1710 		newsvd->rcookie = svd->rcookie;
1711 		hat_dup_region(newseg->s_as->a_hat, newsvd->rcookie);
1712 	}
1713 	return (error);
1714 }
1715 
1716 
1717 /*
1718  * callback function to invoke free_vp_pages() for only those pages actually
1719  * processed by the HAT when a shared region is destroyed.
1720  */
1721 extern int free_pages;
1722 
1723 static void
1724 segvn_hat_rgn_unload_callback(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
1725     size_t r_size, void *r_obj, u_offset_t r_objoff)
1726 {
1727 	u_offset_t off;
1728 	size_t len;
1729 	vnode_t *vp = (vnode_t *)r_obj;
1730 
1731 	ASSERT(eaddr > saddr);
1732 	ASSERT(saddr >= r_saddr);
1733 	ASSERT(saddr < r_saddr + r_size);
1734 	ASSERT(eaddr > r_saddr);
1735 	ASSERT(eaddr <= r_saddr + r_size);
1736 	ASSERT(vp != NULL);
1737 
1738 	if (!free_pages) {
1739 		return;
1740 	}
1741 
1742 	len = eaddr - saddr;
1743 	off = (saddr - r_saddr) + r_objoff;
1744 	free_vp_pages(vp, off, len);
1745 }
1746 
1747 /*
1748  * callback function used by segvn_unmap to invoke free_vp_pages() for only
1749  * those pages actually processed by the HAT
1750  */
1751 static void
1752 segvn_hat_unload_callback(hat_callback_t *cb)
1753 {
1754 	struct seg		*seg = cb->hcb_data;
1755 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
1756 	size_t			len;
1757 	u_offset_t		off;
1758 
1759 	ASSERT(svd->vp != NULL);
1760 	ASSERT(cb->hcb_end_addr > cb->hcb_start_addr);
1761 	ASSERT(cb->hcb_start_addr >= seg->s_base);
1762 
1763 	len = cb->hcb_end_addr - cb->hcb_start_addr;
1764 	off = cb->hcb_start_addr - seg->s_base;
1765 	free_vp_pages(svd->vp, svd->offset + off, len);
1766 }
1767 
1768 /*
1769  * This function determines the number of bytes of swap reserved by
1770  * a segment for which per-page accounting is present. It is used to
1771  * calculate the correct value of a segvn_data's swresv.
1772  */
1773 static size_t
1774 segvn_count_swap_by_vpages(struct seg *seg)
1775 {
1776 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1777 	struct vpage *vp, *evp;
1778 	size_t nswappages = 0;
1779 
1780 	ASSERT(svd->pageswap);
1781 	ASSERT(svd->vpage != NULL);
1782 
1783 	evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
1784 
1785 	for (vp = svd->vpage; vp < evp; vp++) {
1786 		if (VPP_ISSWAPRES(vp))
1787 			nswappages++;
1788 	}
1789 
1790 	return (nswappages << PAGESHIFT);
1791 }
1792 
1793 static int
1794 segvn_unmap(struct seg *seg, caddr_t addr, size_t len)
1795 {
1796 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1797 	struct segvn_data *nsvd;
1798 	struct seg *nseg;
1799 	struct anon_map *amp;
1800 	pgcnt_t	opages;		/* old segment size in pages */
1801 	pgcnt_t	npages;		/* new segment size in pages */
1802 	pgcnt_t	dpages;		/* pages being deleted (unmapped) */
1803 	hat_callback_t callback;	/* used for free_vp_pages() */
1804 	hat_callback_t *cbp = NULL;
1805 	caddr_t nbase;
1806 	size_t nsize;
1807 	size_t oswresv;
1808 	int reclaim = 1;
1809 
1810 	/*
1811 	 * We don't need any segment level locks for "segvn" data
1812 	 * since the address space is "write" locked.
1813 	 */
1814 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
1815 
1816 	/*
1817 	 * Fail the unmap if pages are SOFTLOCKed through this mapping.
1818 	 * softlockcnt is protected from change by the as write lock.
1819 	 */
1820 retry:
1821 	if (svd->softlockcnt > 0) {
1822 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1823 
1824 		/*
1825 		 * If this is shared segment non 0 softlockcnt
1826 		 * means locked pages are still in use.
1827 		 */
1828 		if (svd->type == MAP_SHARED) {
1829 			return (EAGAIN);
1830 		}
1831 
1832 		/*
1833 		 * since we do have the writers lock nobody can fill
1834 		 * the cache during the purge. The flush either succeeds
1835 		 * or we still have pending I/Os.
1836 		 */
1837 		if (reclaim == 1) {
1838 			segvn_purge(seg);
1839 			reclaim = 0;
1840 			goto retry;
1841 		}
1842 		return (EAGAIN);
1843 	}
1844 
1845 	/*
1846 	 * Check for bad sizes
1847 	 */
1848 	if (addr < seg->s_base || addr + len > seg->s_base + seg->s_size ||
1849 	    (len & PAGEOFFSET) || ((uintptr_t)addr & PAGEOFFSET)) {
1850 		panic("segvn_unmap");
1851 		/*NOTREACHED*/
1852 	}
1853 
1854 	if (seg->s_szc != 0) {
1855 		size_t pgsz = page_get_pagesize(seg->s_szc);
1856 		int err;
1857 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
1858 			ASSERT(seg->s_base != addr || seg->s_size != len);
1859 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1860 				ASSERT(svd->amp == NULL);
1861 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1862 				hat_leave_region(seg->s_as->a_hat,
1863 				    svd->rcookie, HAT_REGION_TEXT);
1864 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
1865 				/*
1866 				 * could pass a flag to segvn_demote_range()
1867 				 * below to tell it not to do any unloads but
1868 				 * this case is rare enough to not bother for
1869 				 * now.
1870 				 */
1871 			} else if (svd->tr_state == SEGVN_TR_INIT) {
1872 				svd->tr_state = SEGVN_TR_OFF;
1873 			} else if (svd->tr_state == SEGVN_TR_ON) {
1874 				ASSERT(svd->amp != NULL);
1875 				segvn_textunrepl(seg, 1);
1876 				ASSERT(svd->amp == NULL);
1877 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1878 			}
1879 			VM_STAT_ADD(segvnvmstats.demoterange[0]);
1880 			err = segvn_demote_range(seg, addr, len, SDR_END, 0);
1881 			if (err == 0) {
1882 				return (IE_RETRY);
1883 			}
1884 			return (err);
1885 		}
1886 	}
1887 
1888 	/* Inform the vnode of the unmapping. */
1889 	if (svd->vp) {
1890 		int error;
1891 
1892 		error = VOP_DELMAP(svd->vp,
1893 		    (offset_t)svd->offset + (uintptr_t)(addr - seg->s_base),
1894 		    seg->s_as, addr, len, svd->prot, svd->maxprot,
1895 		    svd->type, svd->cred, NULL);
1896 
1897 		if (error == EAGAIN)
1898 			return (error);
1899 	}
1900 
1901 	/*
1902 	 * Remove any page locks set through this mapping.
1903 	 * If text replication is not off no page locks could have been
1904 	 * established via this mapping.
1905 	 */
1906 	if (svd->tr_state == SEGVN_TR_OFF) {
1907 		(void) segvn_lockop(seg, addr, len, 0, MC_UNLOCK, NULL, 0);
1908 	}
1909 
1910 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1911 		ASSERT(svd->amp == NULL);
1912 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1913 		ASSERT(svd->type == MAP_PRIVATE);
1914 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
1915 		    HAT_REGION_TEXT);
1916 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
1917 	} else if (svd->tr_state == SEGVN_TR_ON) {
1918 		ASSERT(svd->amp != NULL);
1919 		ASSERT(svd->pageprot == 0 && !(svd->prot & PROT_WRITE));
1920 		segvn_textunrepl(seg, 1);
1921 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
1922 	} else {
1923 		if (svd->tr_state != SEGVN_TR_OFF) {
1924 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
1925 			svd->tr_state = SEGVN_TR_OFF;
1926 		}
1927 		/*
1928 		 * Unload any hardware translations in the range to be taken
1929 		 * out. Use a callback to invoke free_vp_pages() effectively.
1930 		 */
1931 		if (svd->vp != NULL && free_pages != 0) {
1932 			callback.hcb_data = seg;
1933 			callback.hcb_function = segvn_hat_unload_callback;
1934 			cbp = &callback;
1935 		}
1936 		hat_unload_callback(seg->s_as->a_hat, addr, len,
1937 		    HAT_UNLOAD_UNMAP, cbp);
1938 
1939 		if (svd->type == MAP_SHARED && svd->vp != NULL &&
1940 		    (svd->vp->v_flag & VVMEXEC) &&
1941 		    ((svd->prot & PROT_WRITE) || svd->pageprot)) {
1942 			segvn_inval_trcache(svd->vp);
1943 		}
1944 	}
1945 
1946 	/*
1947 	 * Check for entire segment
1948 	 */
1949 	if (addr == seg->s_base && len == seg->s_size) {
1950 		seg_free(seg);
1951 		return (0);
1952 	}
1953 
1954 	opages = seg_pages(seg);
1955 	dpages = btop(len);
1956 	npages = opages - dpages;
1957 	amp = svd->amp;
1958 	ASSERT(amp == NULL || amp->a_szc >= seg->s_szc);
1959 
1960 	/*
1961 	 * Check for beginning of segment
1962 	 */
1963 	if (addr == seg->s_base) {
1964 		if (svd->vpage != NULL) {
1965 			size_t nbytes;
1966 			struct vpage *ovpage;
1967 
1968 			ovpage = svd->vpage;	/* keep pointer to vpage */
1969 
1970 			nbytes = vpgtob(npages);
1971 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
1972 			bcopy(&ovpage[dpages], svd->vpage, nbytes);
1973 
1974 			/* free up old vpage */
1975 			kmem_free(ovpage, vpgtob(opages));
1976 		}
1977 		if (amp != NULL) {
1978 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1979 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
1980 				/*
1981 				 * Shared anon map is no longer in use. Before
1982 				 * freeing its pages purge all entries from
1983 				 * pcache that belong to this amp.
1984 				 */
1985 				if (svd->type == MAP_SHARED) {
1986 					ASSERT(amp->refcnt == 1);
1987 					ASSERT(svd->softlockcnt == 0);
1988 					anonmap_purge(amp);
1989 				}
1990 				/*
1991 				 * Free up now unused parts of anon_map array.
1992 				 */
1993 				if (amp->a_szc == seg->s_szc) {
1994 					if (seg->s_szc != 0) {
1995 						anon_free_pages(amp->ahp,
1996 						    svd->anon_index, len,
1997 						    seg->s_szc);
1998 					} else {
1999 						anon_free(amp->ahp,
2000 						    svd->anon_index,
2001 						    len);
2002 					}
2003 				} else {
2004 					ASSERT(svd->type == MAP_SHARED);
2005 					ASSERT(amp->a_szc > seg->s_szc);
2006 					anon_shmap_free_pages(amp,
2007 					    svd->anon_index, len);
2008 				}
2009 
2010 				/*
2011 				 * Unreserve swap space for the
2012 				 * unmapped chunk of this segment in
2013 				 * case it's MAP_SHARED
2014 				 */
2015 				if (svd->type == MAP_SHARED) {
2016 					anon_unresv_zone(len,
2017 					    seg->s_as->a_proc->p_zone);
2018 					amp->swresv -= len;
2019 				}
2020 			}
2021 			ANON_LOCK_EXIT(&amp->a_rwlock);
2022 			svd->anon_index += dpages;
2023 		}
2024 		if (svd->vp != NULL)
2025 			svd->offset += len;
2026 
2027 		seg->s_base += len;
2028 		seg->s_size -= len;
2029 
2030 		if (svd->swresv) {
2031 			if (svd->flags & MAP_NORESERVE) {
2032 				ASSERT(amp);
2033 				oswresv = svd->swresv;
2034 
2035 				svd->swresv = ptob(anon_pages(amp->ahp,
2036 				    svd->anon_index, npages));
2037 				anon_unresv_zone(oswresv - svd->swresv,
2038 				    seg->s_as->a_proc->p_zone);
2039 				if (SEG_IS_PARTIAL_RESV(seg))
2040 					seg->s_as->a_resvsize -= oswresv -
2041 					    svd->swresv;
2042 			} else {
2043 				size_t unlen;
2044 
2045 				if (svd->pageswap) {
2046 					oswresv = svd->swresv;
2047 					svd->swresv =
2048 					    segvn_count_swap_by_vpages(seg);
2049 					ASSERT(oswresv >= svd->swresv);
2050 					unlen = oswresv - svd->swresv;
2051 				} else {
2052 					svd->swresv -= len;
2053 					ASSERT(svd->swresv == seg->s_size);
2054 					unlen = len;
2055 				}
2056 				anon_unresv_zone(unlen,
2057 				    seg->s_as->a_proc->p_zone);
2058 			}
2059 			TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2060 			    seg, len, 0);
2061 		}
2062 
2063 		return (0);
2064 	}
2065 
2066 	/*
2067 	 * Check for end of segment
2068 	 */
2069 	if (addr + len == seg->s_base + seg->s_size) {
2070 		if (svd->vpage != NULL) {
2071 			size_t nbytes;
2072 			struct vpage *ovpage;
2073 
2074 			ovpage = svd->vpage;	/* keep pointer to vpage */
2075 
2076 			nbytes = vpgtob(npages);
2077 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2078 			bcopy(ovpage, svd->vpage, nbytes);
2079 
2080 			/* free up old vpage */
2081 			kmem_free(ovpage, vpgtob(opages));
2082 
2083 		}
2084 		if (amp != NULL) {
2085 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2086 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2087 				/*
2088 				 * Free up now unused parts of anon_map array.
2089 				 */
2090 				ulong_t an_idx = svd->anon_index + npages;
2091 
2092 				/*
2093 				 * Shared anon map is no longer in use. Before
2094 				 * freeing its pages purge all entries from
2095 				 * pcache that belong to this amp.
2096 				 */
2097 				if (svd->type == MAP_SHARED) {
2098 					ASSERT(amp->refcnt == 1);
2099 					ASSERT(svd->softlockcnt == 0);
2100 					anonmap_purge(amp);
2101 				}
2102 
2103 				if (amp->a_szc == seg->s_szc) {
2104 					if (seg->s_szc != 0) {
2105 						anon_free_pages(amp->ahp,
2106 						    an_idx, len,
2107 						    seg->s_szc);
2108 					} else {
2109 						anon_free(amp->ahp, an_idx,
2110 						    len);
2111 					}
2112 				} else {
2113 					ASSERT(svd->type == MAP_SHARED);
2114 					ASSERT(amp->a_szc > seg->s_szc);
2115 					anon_shmap_free_pages(amp,
2116 					    an_idx, len);
2117 				}
2118 
2119 				/*
2120 				 * Unreserve swap space for the
2121 				 * unmapped chunk of this segment in
2122 				 * case it's MAP_SHARED
2123 				 */
2124 				if (svd->type == MAP_SHARED) {
2125 					anon_unresv_zone(len,
2126 					    seg->s_as->a_proc->p_zone);
2127 					amp->swresv -= len;
2128 				}
2129 			}
2130 			ANON_LOCK_EXIT(&amp->a_rwlock);
2131 		}
2132 
2133 		seg->s_size -= len;
2134 
2135 		if (svd->swresv) {
2136 			if (svd->flags & MAP_NORESERVE) {
2137 				ASSERT(amp);
2138 				oswresv = svd->swresv;
2139 				svd->swresv = ptob(anon_pages(amp->ahp,
2140 				    svd->anon_index, npages));
2141 				anon_unresv_zone(oswresv - svd->swresv,
2142 				    seg->s_as->a_proc->p_zone);
2143 				if (SEG_IS_PARTIAL_RESV(seg))
2144 					seg->s_as->a_resvsize -= oswresv -
2145 					    svd->swresv;
2146 			} else {
2147 				size_t unlen;
2148 
2149 				if (svd->pageswap) {
2150 					oswresv = svd->swresv;
2151 					svd->swresv =
2152 					    segvn_count_swap_by_vpages(seg);
2153 					ASSERT(oswresv >= svd->swresv);
2154 					unlen = oswresv - svd->swresv;
2155 				} else {
2156 					svd->swresv -= len;
2157 					ASSERT(svd->swresv == seg->s_size);
2158 					unlen = len;
2159 				}
2160 				anon_unresv_zone(unlen,
2161 				    seg->s_as->a_proc->p_zone);
2162 			}
2163 			TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2164 			    "anon proc:%p %lu %u", seg, len, 0);
2165 		}
2166 
2167 		return (0);
2168 	}
2169 
2170 	/*
2171 	 * The section to go is in the middle of the segment,
2172 	 * have to make it into two segments.  nseg is made for
2173 	 * the high end while seg is cut down at the low end.
2174 	 */
2175 	nbase = addr + len;				/* new seg base */
2176 	nsize = (seg->s_base + seg->s_size) - nbase;	/* new seg size */
2177 	seg->s_size = addr - seg->s_base;		/* shrink old seg */
2178 	nseg = seg_alloc(seg->s_as, nbase, nsize);
2179 	if (nseg == NULL) {
2180 		panic("segvn_unmap seg_alloc");
2181 		/*NOTREACHED*/
2182 	}
2183 	nseg->s_ops = seg->s_ops;
2184 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
2185 	nseg->s_data = (void *)nsvd;
2186 	nseg->s_szc = seg->s_szc;
2187 	*nsvd = *svd;
2188 	nsvd->seg = nseg;
2189 	nsvd->offset = svd->offset + (uintptr_t)(nseg->s_base - seg->s_base);
2190 	nsvd->swresv = 0;
2191 	nsvd->softlockcnt = 0;
2192 	nsvd->softlockcnt_sbase = 0;
2193 	nsvd->softlockcnt_send = 0;
2194 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
2195 
2196 	if (svd->vp != NULL) {
2197 		VN_HOLD(nsvd->vp);
2198 		if (nsvd->type == MAP_SHARED)
2199 			lgrp_shm_policy_init(NULL, nsvd->vp);
2200 	}
2201 	crhold(svd->cred);
2202 
2203 	if (svd->vpage == NULL) {
2204 		nsvd->vpage = NULL;
2205 	} else {
2206 		/* need to split vpage into two arrays */
2207 		size_t nbytes;
2208 		struct vpage *ovpage;
2209 
2210 		ovpage = svd->vpage;		/* keep pointer to vpage */
2211 
2212 		npages = seg_pages(seg);	/* seg has shrunk */
2213 		nbytes = vpgtob(npages);
2214 		svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2215 
2216 		bcopy(ovpage, svd->vpage, nbytes);
2217 
2218 		npages = seg_pages(nseg);
2219 		nbytes = vpgtob(npages);
2220 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2221 
2222 		bcopy(&ovpage[opages - npages], nsvd->vpage, nbytes);
2223 
2224 		/* free up old vpage */
2225 		kmem_free(ovpage, vpgtob(opages));
2226 	}
2227 
2228 	if (amp == NULL) {
2229 		nsvd->amp = NULL;
2230 		nsvd->anon_index = 0;
2231 	} else {
2232 		/*
2233 		 * Need to create a new anon map for the new segment.
2234 		 * We'll also allocate a new smaller array for the old
2235 		 * smaller segment to save space.
2236 		 */
2237 		opages = btop((uintptr_t)(addr - seg->s_base));
2238 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2239 		if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2240 			/*
2241 			 * Free up now unused parts of anon_map array.
2242 			 */
2243 			ulong_t an_idx = svd->anon_index + opages;
2244 
2245 			/*
2246 			 * Shared anon map is no longer in use. Before
2247 			 * freeing its pages purge all entries from
2248 			 * pcache that belong to this amp.
2249 			 */
2250 			if (svd->type == MAP_SHARED) {
2251 				ASSERT(amp->refcnt == 1);
2252 				ASSERT(svd->softlockcnt == 0);
2253 				anonmap_purge(amp);
2254 			}
2255 
2256 			if (amp->a_szc == seg->s_szc) {
2257 				if (seg->s_szc != 0) {
2258 					anon_free_pages(amp->ahp, an_idx, len,
2259 					    seg->s_szc);
2260 				} else {
2261 					anon_free(amp->ahp, an_idx,
2262 					    len);
2263 				}
2264 			} else {
2265 				ASSERT(svd->type == MAP_SHARED);
2266 				ASSERT(amp->a_szc > seg->s_szc);
2267 				anon_shmap_free_pages(amp, an_idx, len);
2268 			}
2269 
2270 			/*
2271 			 * Unreserve swap space for the
2272 			 * unmapped chunk of this segment in
2273 			 * case it's MAP_SHARED
2274 			 */
2275 			if (svd->type == MAP_SHARED) {
2276 				anon_unresv_zone(len,
2277 				    seg->s_as->a_proc->p_zone);
2278 				amp->swresv -= len;
2279 			}
2280 		}
2281 		nsvd->anon_index = svd->anon_index +
2282 		    btop((uintptr_t)(nseg->s_base - seg->s_base));
2283 		if (svd->type == MAP_SHARED) {
2284 			amp->refcnt++;
2285 			nsvd->amp = amp;
2286 		} else {
2287 			struct anon_map *namp;
2288 			struct anon_hdr *nahp;
2289 
2290 			ASSERT(svd->type == MAP_PRIVATE);
2291 			nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
2292 			namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
2293 			namp->a_szc = seg->s_szc;
2294 			(void) anon_copy_ptr(amp->ahp, svd->anon_index, nahp,
2295 			    0, btop(seg->s_size), ANON_SLEEP);
2296 			(void) anon_copy_ptr(amp->ahp, nsvd->anon_index,
2297 			    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
2298 			anon_release(amp->ahp, btop(amp->size));
2299 			svd->anon_index = 0;
2300 			nsvd->anon_index = 0;
2301 			amp->ahp = nahp;
2302 			amp->size = seg->s_size;
2303 			nsvd->amp = namp;
2304 		}
2305 		ANON_LOCK_EXIT(&amp->a_rwlock);
2306 	}
2307 	if (svd->swresv) {
2308 		if (svd->flags & MAP_NORESERVE) {
2309 			ASSERT(amp);
2310 			oswresv = svd->swresv;
2311 			svd->swresv = ptob(anon_pages(amp->ahp,
2312 			    svd->anon_index, btop(seg->s_size)));
2313 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
2314 			    nsvd->anon_index, btop(nseg->s_size)));
2315 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2316 			anon_unresv_zone(oswresv - (svd->swresv + nsvd->swresv),
2317 			    seg->s_as->a_proc->p_zone);
2318 			if (SEG_IS_PARTIAL_RESV(seg))
2319 				seg->s_as->a_resvsize -= oswresv -
2320 				    (svd->swresv + nsvd->swresv);
2321 		} else {
2322 			size_t unlen;
2323 
2324 			if (svd->pageswap) {
2325 				oswresv = svd->swresv;
2326 				svd->swresv = segvn_count_swap_by_vpages(seg);
2327 				nsvd->swresv = segvn_count_swap_by_vpages(nseg);
2328 				ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2329 				unlen = oswresv - (svd->swresv + nsvd->swresv);
2330 			} else {
2331 				if (seg->s_size + nseg->s_size + len !=
2332 				    svd->swresv) {
2333 					panic("segvn_unmap: cannot split "
2334 					    "swap reservation");
2335 					/*NOTREACHED*/
2336 				}
2337 				svd->swresv = seg->s_size;
2338 				nsvd->swresv = nseg->s_size;
2339 				unlen = len;
2340 			}
2341 			anon_unresv_zone(unlen,
2342 			    seg->s_as->a_proc->p_zone);
2343 		}
2344 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2345 		    seg, len, 0);
2346 	}
2347 
2348 	return (0);			/* I'm glad that's all over with! */
2349 }
2350 
2351 static void
2352 segvn_free(struct seg *seg)
2353 {
2354 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2355 	pgcnt_t npages = seg_pages(seg);
2356 	struct anon_map *amp;
2357 	size_t len;
2358 
2359 	/*
2360 	 * We don't need any segment level locks for "segvn" data
2361 	 * since the address space is "write" locked.
2362 	 */
2363 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
2364 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
2365 
2366 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2367 
2368 	/*
2369 	 * Be sure to unlock pages. XXX Why do things get free'ed instead
2370 	 * of unmapped? XXX
2371 	 */
2372 	(void) segvn_lockop(seg, seg->s_base, seg->s_size,
2373 	    0, MC_UNLOCK, NULL, 0);
2374 
2375 	/*
2376 	 * Deallocate the vpage and anon pointers if necessary and possible.
2377 	 */
2378 	if (svd->vpage != NULL) {
2379 		kmem_free(svd->vpage, vpgtob(npages));
2380 		svd->vpage = NULL;
2381 	}
2382 	if ((amp = svd->amp) != NULL) {
2383 		/*
2384 		 * If there are no more references to this anon_map
2385 		 * structure, then deallocate the structure after freeing
2386 		 * up all the anon slot pointers that we can.
2387 		 */
2388 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2389 		ASSERT(amp->a_szc >= seg->s_szc);
2390 		if (--amp->refcnt == 0) {
2391 			if (svd->type == MAP_PRIVATE) {
2392 				/*
2393 				 * Private - we only need to anon_free
2394 				 * the part that this segment refers to.
2395 				 */
2396 				if (seg->s_szc != 0) {
2397 					anon_free_pages(amp->ahp,
2398 					    svd->anon_index, seg->s_size,
2399 					    seg->s_szc);
2400 				} else {
2401 					anon_free(amp->ahp, svd->anon_index,
2402 					    seg->s_size);
2403 				}
2404 			} else {
2405 
2406 				/*
2407 				 * Shared anon map is no longer in use. Before
2408 				 * freeing its pages purge all entries from
2409 				 * pcache that belong to this amp.
2410 				 */
2411 				ASSERT(svd->softlockcnt == 0);
2412 				anonmap_purge(amp);
2413 
2414 				/*
2415 				 * Shared - anon_free the entire
2416 				 * anon_map's worth of stuff and
2417 				 * release any swap reservation.
2418 				 */
2419 				if (amp->a_szc != 0) {
2420 					anon_shmap_free_pages(amp, 0,
2421 					    amp->size);
2422 				} else {
2423 					anon_free(amp->ahp, 0, amp->size);
2424 				}
2425 				if ((len = amp->swresv) != 0) {
2426 					anon_unresv_zone(len,
2427 					    seg->s_as->a_proc->p_zone);
2428 					TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2429 					    "anon proc:%p %lu %u", seg, len, 0);
2430 				}
2431 			}
2432 			svd->amp = NULL;
2433 			ANON_LOCK_EXIT(&amp->a_rwlock);
2434 			anonmap_free(amp);
2435 		} else if (svd->type == MAP_PRIVATE) {
2436 			/*
2437 			 * We had a private mapping which still has
2438 			 * a held anon_map so just free up all the
2439 			 * anon slot pointers that we were using.
2440 			 */
2441 			if (seg->s_szc != 0) {
2442 				anon_free_pages(amp->ahp, svd->anon_index,
2443 				    seg->s_size, seg->s_szc);
2444 			} else {
2445 				anon_free(amp->ahp, svd->anon_index,
2446 				    seg->s_size);
2447 			}
2448 			ANON_LOCK_EXIT(&amp->a_rwlock);
2449 		} else {
2450 			ANON_LOCK_EXIT(&amp->a_rwlock);
2451 		}
2452 	}
2453 
2454 	/*
2455 	 * Release swap reservation.
2456 	 */
2457 	if ((len = svd->swresv) != 0) {
2458 		anon_unresv_zone(svd->swresv,
2459 		    seg->s_as->a_proc->p_zone);
2460 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2461 		    seg, len, 0);
2462 		if (SEG_IS_PARTIAL_RESV(seg))
2463 			seg->s_as->a_resvsize -= svd->swresv;
2464 		svd->swresv = 0;
2465 	}
2466 	/*
2467 	 * Release claim on vnode, credentials, and finally free the
2468 	 * private data.
2469 	 */
2470 	if (svd->vp != NULL) {
2471 		if (svd->type == MAP_SHARED)
2472 			lgrp_shm_policy_fini(NULL, svd->vp);
2473 		VN_RELE(svd->vp);
2474 		svd->vp = NULL;
2475 	}
2476 	crfree(svd->cred);
2477 	svd->pageprot = 0;
2478 	svd->pageadvice = 0;
2479 	svd->pageswap = 0;
2480 	svd->cred = NULL;
2481 
2482 	/*
2483 	 * Take segfree_syncmtx lock to let segvn_reclaim() finish if it's
2484 	 * still working with this segment without holding as lock (in case
2485 	 * it's called by pcache async thread).
2486 	 */
2487 	ASSERT(svd->softlockcnt == 0);
2488 	mutex_enter(&svd->segfree_syncmtx);
2489 	mutex_exit(&svd->segfree_syncmtx);
2490 
2491 	seg->s_data = NULL;
2492 	kmem_cache_free(segvn_cache, svd);
2493 }
2494 
2495 /*
2496  * Do a F_SOFTUNLOCK call over the range requested.  The range must have
2497  * already been F_SOFTLOCK'ed.
2498  * Caller must always match addr and len of a softunlock with a previous
2499  * softlock with exactly the same addr and len.
2500  */
2501 static void
2502 segvn_softunlock(struct seg *seg, caddr_t addr, size_t len, enum seg_rw rw)
2503 {
2504 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2505 	page_t *pp;
2506 	caddr_t adr;
2507 	struct vnode *vp;
2508 	u_offset_t offset;
2509 	ulong_t anon_index;
2510 	struct anon_map *amp;
2511 	struct anon *ap = NULL;
2512 
2513 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
2514 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
2515 
2516 	if ((amp = svd->amp) != NULL)
2517 		anon_index = svd->anon_index + seg_page(seg, addr);
2518 
2519 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
2520 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2521 		hat_unlock_region(seg->s_as->a_hat, addr, len, svd->rcookie);
2522 	} else {
2523 		hat_unlock(seg->s_as->a_hat, addr, len);
2524 	}
2525 	for (adr = addr; adr < addr + len; adr += PAGESIZE) {
2526 		if (amp != NULL) {
2527 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2528 			if ((ap = anon_get_ptr(amp->ahp, anon_index++))
2529 			    != NULL) {
2530 				swap_xlate(ap, &vp, &offset);
2531 			} else {
2532 				vp = svd->vp;
2533 				offset = svd->offset +
2534 				    (uintptr_t)(adr - seg->s_base);
2535 			}
2536 			ANON_LOCK_EXIT(&amp->a_rwlock);
2537 		} else {
2538 			vp = svd->vp;
2539 			offset = svd->offset +
2540 			    (uintptr_t)(adr - seg->s_base);
2541 		}
2542 
2543 		/*
2544 		 * Use page_find() instead of page_lookup() to
2545 		 * find the page since we know that it is locked.
2546 		 */
2547 		pp = page_find(vp, offset);
2548 		if (pp == NULL) {
2549 			panic(
2550 			    "segvn_softunlock: addr %p, ap %p, vp %p, off %llx",
2551 			    (void *)adr, (void *)ap, (void *)vp, offset);
2552 			/*NOTREACHED*/
2553 		}
2554 
2555 		if (rw == S_WRITE) {
2556 			hat_setrefmod(pp);
2557 			if (seg->s_as->a_vbits)
2558 				hat_setstat(seg->s_as, adr, PAGESIZE,
2559 				    P_REF | P_MOD);
2560 		} else if (rw != S_OTHER) {
2561 			hat_setref(pp);
2562 			if (seg->s_as->a_vbits)
2563 				hat_setstat(seg->s_as, adr, PAGESIZE, P_REF);
2564 		}
2565 		TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2566 		    "segvn_fault:pp %p vp %p offset %llx", pp, vp, offset);
2567 		page_unlock(pp);
2568 	}
2569 	ASSERT(svd->softlockcnt >= btop(len));
2570 	if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -btop(len))) {
2571 		/*
2572 		 * All SOFTLOCKS are gone. Wakeup any waiting
2573 		 * unmappers so they can try again to unmap.
2574 		 * Check for waiters first without the mutex
2575 		 * held so we don't always grab the mutex on
2576 		 * softunlocks.
2577 		 */
2578 		if (AS_ISUNMAPWAIT(seg->s_as)) {
2579 			mutex_enter(&seg->s_as->a_contents);
2580 			if (AS_ISUNMAPWAIT(seg->s_as)) {
2581 				AS_CLRUNMAPWAIT(seg->s_as);
2582 				cv_broadcast(&seg->s_as->a_cv);
2583 			}
2584 			mutex_exit(&seg->s_as->a_contents);
2585 		}
2586 	}
2587 }
2588 
2589 #define	PAGE_HANDLED	((page_t *)-1)
2590 
2591 /*
2592  * Release all the pages in the NULL terminated ppp list
2593  * which haven't already been converted to PAGE_HANDLED.
2594  */
2595 static void
2596 segvn_pagelist_rele(page_t **ppp)
2597 {
2598 	for (; *ppp != NULL; ppp++) {
2599 		if (*ppp != PAGE_HANDLED)
2600 			page_unlock(*ppp);
2601 	}
2602 }
2603 
2604 static int stealcow = 1;
2605 
2606 /*
2607  * Workaround for viking chip bug.  See bug id 1220902.
2608  * To fix this down in pagefault() would require importing so
2609  * much as and segvn code as to be unmaintainable.
2610  */
2611 int enable_mbit_wa = 0;
2612 
2613 /*
2614  * Handles all the dirty work of getting the right
2615  * anonymous pages and loading up the translations.
2616  * This routine is called only from segvn_fault()
2617  * when looping over the range of addresses requested.
2618  *
2619  * The basic algorithm here is:
2620  * 	If this is an anon_zero case
2621  *		Call anon_zero to allocate page
2622  *		Load up translation
2623  *		Return
2624  *	endif
2625  *	If this is an anon page
2626  *		Use anon_getpage to get the page
2627  *	else
2628  *		Find page in pl[] list passed in
2629  *	endif
2630  *	If not a cow
2631  *		Load up the translation to the page
2632  *		return
2633  *	endif
2634  *	Call anon_private to handle cow
2635  *	Load up (writable) translation to new page
2636  */
2637 static faultcode_t
2638 segvn_faultpage(
2639 	struct hat *hat,		/* the hat to use for mapping */
2640 	struct seg *seg,		/* seg_vn of interest */
2641 	caddr_t addr,			/* address in as */
2642 	u_offset_t off,			/* offset in vp */
2643 	struct vpage *vpage,		/* pointer to vpage for vp, off */
2644 	page_t *pl[],			/* object source page pointer */
2645 	uint_t vpprot,			/* access allowed to object pages */
2646 	enum fault_type type,		/* type of fault */
2647 	enum seg_rw rw,			/* type of access at fault */
2648 	int brkcow)			/* we may need to break cow */
2649 {
2650 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2651 	page_t *pp, **ppp;
2652 	uint_t pageflags = 0;
2653 	page_t *anon_pl[1 + 1];
2654 	page_t *opp = NULL;		/* original page */
2655 	uint_t prot;
2656 	int err;
2657 	int cow;
2658 	int claim;
2659 	int steal = 0;
2660 	ulong_t anon_index;
2661 	struct anon *ap, *oldap;
2662 	struct anon_map *amp;
2663 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
2664 	int anon_lock = 0;
2665 	anon_sync_obj_t cookie;
2666 
2667 	if (svd->flags & MAP_TEXT) {
2668 		hat_flag |= HAT_LOAD_TEXT;
2669 	}
2670 
2671 	ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
2672 	ASSERT(seg->s_szc == 0);
2673 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
2674 
2675 	/*
2676 	 * Initialize protection value for this page.
2677 	 * If we have per page protection values check it now.
2678 	 */
2679 	if (svd->pageprot) {
2680 		uint_t protchk;
2681 
2682 		switch (rw) {
2683 		case S_READ:
2684 			protchk = PROT_READ;
2685 			break;
2686 		case S_WRITE:
2687 			protchk = PROT_WRITE;
2688 			break;
2689 		case S_EXEC:
2690 			protchk = PROT_EXEC;
2691 			break;
2692 		case S_OTHER:
2693 		default:
2694 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
2695 			break;
2696 		}
2697 
2698 		prot = VPP_PROT(vpage);
2699 		if ((prot & protchk) == 0)
2700 			return (FC_PROT);	/* illegal access type */
2701 	} else {
2702 		prot = svd->prot;
2703 	}
2704 
2705 	if (type == F_SOFTLOCK) {
2706 		atomic_inc_ulong((ulong_t *)&svd->softlockcnt);
2707 	}
2708 
2709 	/*
2710 	 * Always acquire the anon array lock to prevent 2 threads from
2711 	 * allocating separate anon slots for the same "addr".
2712 	 */
2713 
2714 	if ((amp = svd->amp) != NULL) {
2715 		ASSERT(RW_READ_HELD(&amp->a_rwlock));
2716 		anon_index = svd->anon_index + seg_page(seg, addr);
2717 		anon_array_enter(amp, anon_index, &cookie);
2718 		anon_lock = 1;
2719 	}
2720 
2721 	if (svd->vp == NULL && amp != NULL) {
2722 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL) {
2723 			/*
2724 			 * Allocate a (normally) writable anonymous page of
2725 			 * zeroes. If no advance reservations, reserve now.
2726 			 */
2727 			if (svd->flags & MAP_NORESERVE) {
2728 				if (anon_resv_zone(ptob(1),
2729 				    seg->s_as->a_proc->p_zone)) {
2730 					atomic_add_long(&svd->swresv, ptob(1));
2731 					atomic_add_long(&seg->s_as->a_resvsize,
2732 					    ptob(1));
2733 				} else {
2734 					err = ENOMEM;
2735 					goto out;
2736 				}
2737 			}
2738 			if ((pp = anon_zero(seg, addr, &ap,
2739 			    svd->cred)) == NULL) {
2740 				err = ENOMEM;
2741 				goto out;	/* out of swap space */
2742 			}
2743 			/*
2744 			 * Re-acquire the anon_map lock and
2745 			 * initialize the anon array entry.
2746 			 */
2747 			(void) anon_set_ptr(amp->ahp, anon_index, ap,
2748 			    ANON_SLEEP);
2749 
2750 			ASSERT(pp->p_szc == 0);
2751 
2752 			/*
2753 			 * Handle pages that have been marked for migration
2754 			 */
2755 			if (lgrp_optimizations())
2756 				page_migrate(seg, addr, &pp, 1);
2757 
2758 			if (enable_mbit_wa) {
2759 				if (rw == S_WRITE)
2760 					hat_setmod(pp);
2761 				else if (!hat_ismod(pp))
2762 					prot &= ~PROT_WRITE;
2763 			}
2764 			/*
2765 			 * If AS_PAGLCK is set in a_flags (via memcntl(2)
2766 			 * with MC_LOCKAS, MCL_FUTURE) and this is a
2767 			 * MAP_NORESERVE segment, we may need to
2768 			 * permanently lock the page as it is being faulted
2769 			 * for the first time. The following text applies
2770 			 * only to MAP_NORESERVE segments:
2771 			 *
2772 			 * As per memcntl(2), if this segment was created
2773 			 * after MCL_FUTURE was applied (a "future"
2774 			 * segment), its pages must be locked.  If this
2775 			 * segment existed at MCL_FUTURE application (a
2776 			 * "past" segment), the interface is unclear.
2777 			 *
2778 			 * We decide to lock only if vpage is present:
2779 			 *
2780 			 * - "future" segments will have a vpage array (see
2781 			 *    as_map), and so will be locked as required
2782 			 *
2783 			 * - "past" segments may not have a vpage array,
2784 			 *    depending on whether events (such as
2785 			 *    mprotect) have occurred. Locking if vpage
2786 			 *    exists will preserve legacy behavior.  Not
2787 			 *    locking if vpage is absent, will not break
2788 			 *    the interface or legacy behavior.  Note that
2789 			 *    allocating vpage here if it's absent requires
2790 			 *    upgrading the segvn reader lock, the cost of
2791 			 *    which does not seem worthwhile.
2792 			 *
2793 			 * Usually testing and setting VPP_ISPPLOCK and
2794 			 * VPP_SETPPLOCK requires holding the segvn lock as
2795 			 * writer, but in this case all readers are
2796 			 * serializing on the anon array lock.
2797 			 */
2798 			if (AS_ISPGLCK(seg->s_as) && vpage != NULL &&
2799 			    (svd->flags & MAP_NORESERVE) &&
2800 			    !VPP_ISPPLOCK(vpage)) {
2801 				proc_t *p = seg->s_as->a_proc;
2802 				ASSERT(svd->type == MAP_PRIVATE);
2803 				mutex_enter(&p->p_lock);
2804 				if (rctl_incr_locked_mem(p, NULL, PAGESIZE,
2805 				    1) == 0) {
2806 					claim = VPP_PROT(vpage) & PROT_WRITE;
2807 					if (page_pp_lock(pp, claim, 0)) {
2808 						VPP_SETPPLOCK(vpage);
2809 					} else {
2810 						rctl_decr_locked_mem(p, NULL,
2811 						    PAGESIZE, 1);
2812 					}
2813 				}
2814 				mutex_exit(&p->p_lock);
2815 			}
2816 
2817 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2818 			hat_memload(hat, addr, pp, prot, hat_flag);
2819 
2820 			if (!(hat_flag & HAT_LOAD_LOCK))
2821 				page_unlock(pp);
2822 
2823 			anon_array_exit(&cookie);
2824 			return (0);
2825 		}
2826 	}
2827 
2828 	/*
2829 	 * Obtain the page structure via anon_getpage() if it is
2830 	 * a private copy of an object (the result of a previous
2831 	 * copy-on-write).
2832 	 */
2833 	if (amp != NULL) {
2834 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) != NULL) {
2835 			err = anon_getpage(&ap, &vpprot, anon_pl, PAGESIZE,
2836 			    seg, addr, rw, svd->cred);
2837 			if (err)
2838 				goto out;
2839 
2840 			if (svd->type == MAP_SHARED) {
2841 				/*
2842 				 * If this is a shared mapping to an
2843 				 * anon_map, then ignore the write
2844 				 * permissions returned by anon_getpage().
2845 				 * They apply to the private mappings
2846 				 * of this anon_map.
2847 				 */
2848 				vpprot |= PROT_WRITE;
2849 			}
2850 			opp = anon_pl[0];
2851 		}
2852 	}
2853 
2854 	/*
2855 	 * Search the pl[] list passed in if it is from the
2856 	 * original object (i.e., not a private copy).
2857 	 */
2858 	if (opp == NULL) {
2859 		/*
2860 		 * Find original page.  We must be bringing it in
2861 		 * from the list in pl[].
2862 		 */
2863 		for (ppp = pl; (opp = *ppp) != NULL; ppp++) {
2864 			if (opp == PAGE_HANDLED)
2865 				continue;
2866 			ASSERT(opp->p_vnode == svd->vp); /* XXX */
2867 			if (opp->p_offset == off)
2868 				break;
2869 		}
2870 		if (opp == NULL) {
2871 			panic("segvn_faultpage not found");
2872 			/*NOTREACHED*/
2873 		}
2874 		*ppp = PAGE_HANDLED;
2875 
2876 	}
2877 
2878 	ASSERT(PAGE_LOCKED(opp));
2879 
2880 	TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2881 	    "segvn_fault:pp %p vp %p offset %llx", opp, NULL, 0);
2882 
2883 	/*
2884 	 * The fault is treated as a copy-on-write fault if a
2885 	 * write occurs on a private segment and the object
2886 	 * page (i.e., mapping) is write protected.  We assume
2887 	 * that fatal protection checks have already been made.
2888 	 */
2889 
2890 	if (brkcow) {
2891 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2892 		cow = !(vpprot & PROT_WRITE);
2893 	} else if (svd->tr_state == SEGVN_TR_ON) {
2894 		/*
2895 		 * If we are doing text replication COW on first touch.
2896 		 */
2897 		ASSERT(amp != NULL);
2898 		ASSERT(svd->vp != NULL);
2899 		ASSERT(rw != S_WRITE);
2900 		cow = (ap == NULL);
2901 	} else {
2902 		cow = 0;
2903 	}
2904 
2905 	/*
2906 	 * If not a copy-on-write case load the translation
2907 	 * and return.
2908 	 */
2909 	if (cow == 0) {
2910 
2911 		/*
2912 		 * Handle pages that have been marked for migration
2913 		 */
2914 		if (lgrp_optimizations())
2915 			page_migrate(seg, addr, &opp, 1);
2916 
2917 		if (IS_VMODSORT(opp->p_vnode) || enable_mbit_wa) {
2918 			if (rw == S_WRITE)
2919 				hat_setmod(opp);
2920 			else if (rw != S_OTHER && !hat_ismod(opp))
2921 				prot &= ~PROT_WRITE;
2922 		}
2923 
2924 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
2925 		    (!svd->pageprot && svd->prot == (prot & vpprot)));
2926 		ASSERT(amp == NULL ||
2927 		    svd->rcookie == HAT_INVALID_REGION_COOKIE);
2928 		hat_memload_region(hat, addr, opp, prot & vpprot, hat_flag,
2929 		    svd->rcookie);
2930 
2931 		if (!(hat_flag & HAT_LOAD_LOCK))
2932 			page_unlock(opp);
2933 
2934 		if (anon_lock) {
2935 			anon_array_exit(&cookie);
2936 		}
2937 		return (0);
2938 	}
2939 
2940 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2941 
2942 	hat_setref(opp);
2943 
2944 	ASSERT(amp != NULL && anon_lock);
2945 
2946 	/*
2947 	 * Steal the page only if it isn't a private page
2948 	 * since stealing a private page is not worth the effort.
2949 	 */
2950 	if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL)
2951 		steal = 1;
2952 
2953 	/*
2954 	 * Steal the original page if the following conditions are true:
2955 	 *
2956 	 * We are low on memory, the page is not private, page is not large,
2957 	 * not shared, not modified, not `locked' or if we have it `locked'
2958 	 * (i.e., p_cowcnt == 1 and p_lckcnt == 0, which also implies
2959 	 * that the page is not shared) and if it doesn't have any
2960 	 * translations. page_struct_lock isn't needed to look at p_cowcnt
2961 	 * and p_lckcnt because we first get exclusive lock on page.
2962 	 */
2963 	(void) hat_pagesync(opp, HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD);
2964 
2965 	if (stealcow && freemem < minfree && steal && opp->p_szc == 0 &&
2966 	    page_tryupgrade(opp) && !hat_ismod(opp) &&
2967 	    ((opp->p_lckcnt == 0 && opp->p_cowcnt == 0) ||
2968 	    (opp->p_lckcnt == 0 && opp->p_cowcnt == 1 &&
2969 	    vpage != NULL && VPP_ISPPLOCK(vpage)))) {
2970 		/*
2971 		 * Check if this page has other translations
2972 		 * after unloading our translation.
2973 		 */
2974 		if (hat_page_is_mapped(opp)) {
2975 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2976 			hat_unload(seg->s_as->a_hat, addr, PAGESIZE,
2977 			    HAT_UNLOAD);
2978 		}
2979 
2980 		/*
2981 		 * hat_unload() might sync back someone else's recent
2982 		 * modification, so check again.
2983 		 */
2984 		if (!hat_ismod(opp) && !hat_page_is_mapped(opp))
2985 			pageflags |= STEAL_PAGE;
2986 	}
2987 
2988 	/*
2989 	 * If we have a vpage pointer, see if it indicates that we have
2990 	 * ``locked'' the page we map -- if so, tell anon_private to
2991 	 * transfer the locking resource to the new page.
2992 	 *
2993 	 * See Statement at the beginning of segvn_lockop regarding
2994 	 * the way lockcnts/cowcnts are handled during COW.
2995 	 *
2996 	 */
2997 	if (vpage != NULL && VPP_ISPPLOCK(vpage))
2998 		pageflags |= LOCK_PAGE;
2999 
3000 	/*
3001 	 * Allocate a private page and perform the copy.
3002 	 * For MAP_NORESERVE reserve swap space now, unless this
3003 	 * is a cow fault on an existing anon page in which case
3004 	 * MAP_NORESERVE will have made advance reservations.
3005 	 */
3006 	if ((svd->flags & MAP_NORESERVE) && (ap == NULL)) {
3007 		if (anon_resv_zone(ptob(1), seg->s_as->a_proc->p_zone)) {
3008 			atomic_add_long(&svd->swresv, ptob(1));
3009 			atomic_add_long(&seg->s_as->a_resvsize, ptob(1));
3010 		} else {
3011 			page_unlock(opp);
3012 			err = ENOMEM;
3013 			goto out;
3014 		}
3015 	}
3016 	oldap = ap;
3017 	pp = anon_private(&ap, seg, addr, prot, opp, pageflags, svd->cred);
3018 	if (pp == NULL) {
3019 		err = ENOMEM;	/* out of swap space */
3020 		goto out;
3021 	}
3022 
3023 	/*
3024 	 * If we copied away from an anonymous page, then
3025 	 * we are one step closer to freeing up an anon slot.
3026 	 *
3027 	 * NOTE:  The original anon slot must be released while
3028 	 * holding the "anon_map" lock.  This is necessary to prevent
3029 	 * other threads from obtaining a pointer to the anon slot
3030 	 * which may be freed if its "refcnt" is 1.
3031 	 */
3032 	if (oldap != NULL)
3033 		anon_decref(oldap);
3034 
3035 	(void) anon_set_ptr(amp->ahp, anon_index, ap, ANON_SLEEP);
3036 
3037 	/*
3038 	 * Handle pages that have been marked for migration
3039 	 */
3040 	if (lgrp_optimizations())
3041 		page_migrate(seg, addr, &pp, 1);
3042 
3043 	ASSERT(pp->p_szc == 0);
3044 
3045 	ASSERT(!IS_VMODSORT(pp->p_vnode));
3046 	if (enable_mbit_wa) {
3047 		if (rw == S_WRITE)
3048 			hat_setmod(pp);
3049 		else if (!hat_ismod(pp))
3050 			prot &= ~PROT_WRITE;
3051 	}
3052 
3053 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3054 	hat_memload(hat, addr, pp, prot, hat_flag);
3055 
3056 	if (!(hat_flag & HAT_LOAD_LOCK))
3057 		page_unlock(pp);
3058 
3059 	ASSERT(anon_lock);
3060 	anon_array_exit(&cookie);
3061 	return (0);
3062 out:
3063 	if (anon_lock)
3064 		anon_array_exit(&cookie);
3065 
3066 	if (type == F_SOFTLOCK) {
3067 		atomic_dec_ulong((ulong_t *)&svd->softlockcnt);
3068 	}
3069 	return (FC_MAKE_ERR(err));
3070 }
3071 
3072 /*
3073  * relocate a bunch of smaller targ pages into one large repl page. all targ
3074  * pages must be complete pages smaller than replacement pages.
3075  * it's assumed that no page's szc can change since they are all PAGESIZE or
3076  * complete large pages locked SHARED.
3077  */
3078 static void
3079 segvn_relocate_pages(page_t **targ, page_t *replacement)
3080 {
3081 	page_t *pp;
3082 	pgcnt_t repl_npgs, curnpgs;
3083 	pgcnt_t i;
3084 	uint_t repl_szc = replacement->p_szc;
3085 	page_t *first_repl = replacement;
3086 	page_t *repl;
3087 	spgcnt_t npgs;
3088 
3089 	VM_STAT_ADD(segvnvmstats.relocatepages[0]);
3090 
3091 	ASSERT(repl_szc != 0);
3092 	npgs = repl_npgs = page_get_pagecnt(repl_szc);
3093 
3094 	i = 0;
3095 	while (repl_npgs) {
3096 		spgcnt_t nreloc;
3097 		int err;
3098 		ASSERT(replacement != NULL);
3099 		pp = targ[i];
3100 		ASSERT(pp->p_szc < repl_szc);
3101 		ASSERT(PAGE_EXCL(pp));
3102 		ASSERT(!PP_ISFREE(pp));
3103 		curnpgs = page_get_pagecnt(pp->p_szc);
3104 		if (curnpgs == 1) {
3105 			VM_STAT_ADD(segvnvmstats.relocatepages[1]);
3106 			repl = replacement;
3107 			page_sub(&replacement, repl);
3108 			ASSERT(PAGE_EXCL(repl));
3109 			ASSERT(!PP_ISFREE(repl));
3110 			ASSERT(repl->p_szc == repl_szc);
3111 		} else {
3112 			page_t *repl_savepp;
3113 			int j;
3114 			VM_STAT_ADD(segvnvmstats.relocatepages[2]);
3115 			repl_savepp = replacement;
3116 			for (j = 0; j < curnpgs; j++) {
3117 				repl = replacement;
3118 				page_sub(&replacement, repl);
3119 				ASSERT(PAGE_EXCL(repl));
3120 				ASSERT(!PP_ISFREE(repl));
3121 				ASSERT(repl->p_szc == repl_szc);
3122 				ASSERT(page_pptonum(targ[i + j]) ==
3123 				    page_pptonum(targ[i]) + j);
3124 			}
3125 			repl = repl_savepp;
3126 			ASSERT(IS_P2ALIGNED(page_pptonum(repl), curnpgs));
3127 		}
3128 		err = page_relocate(&pp, &repl, 0, 1, &nreloc, NULL);
3129 		if (err || nreloc != curnpgs) {
3130 			panic("segvn_relocate_pages: "
3131 			    "page_relocate failed err=%d curnpgs=%ld "
3132 			    "nreloc=%ld", err, curnpgs, nreloc);
3133 		}
3134 		ASSERT(curnpgs <= repl_npgs);
3135 		repl_npgs -= curnpgs;
3136 		i += curnpgs;
3137 	}
3138 	ASSERT(replacement == NULL);
3139 
3140 	repl = first_repl;
3141 	repl_npgs = npgs;
3142 	for (i = 0; i < repl_npgs; i++) {
3143 		ASSERT(PAGE_EXCL(repl));
3144 		ASSERT(!PP_ISFREE(repl));
3145 		targ[i] = repl;
3146 		page_downgrade(targ[i]);
3147 		repl++;
3148 	}
3149 }
3150 
3151 /*
3152  * Check if all pages in ppa array are complete smaller than szc pages and
3153  * their roots will still be aligned relative to their current size if the
3154  * entire ppa array is relocated into one szc page. If these conditions are
3155  * not met return 0.
3156  *
3157  * If all pages are properly aligned attempt to upgrade their locks
3158  * to exclusive mode. If it fails set *upgrdfail to 1 and return 0.
3159  * upgrdfail was set to 0 by caller.
3160  *
3161  * Return 1 if all pages are aligned and locked exclusively.
3162  *
3163  * If all pages in ppa array happen to be physically contiguous to make one
3164  * szc page and all exclusive locks are successfully obtained promote the page
3165  * size to szc and set *pszc to szc. Return 1 with pages locked shared.
3166  */
3167 static int
3168 segvn_full_szcpages(page_t **ppa, uint_t szc, int *upgrdfail, uint_t *pszc)
3169 {
3170 	page_t *pp;
3171 	pfn_t pfn;
3172 	pgcnt_t totnpgs = page_get_pagecnt(szc);
3173 	pfn_t first_pfn;
3174 	int contig = 1;
3175 	pgcnt_t i;
3176 	pgcnt_t j;
3177 	uint_t curszc;
3178 	pgcnt_t curnpgs;
3179 	int root = 0;
3180 
3181 	ASSERT(szc > 0);
3182 
3183 	VM_STAT_ADD(segvnvmstats.fullszcpages[0]);
3184 
3185 	for (i = 0; i < totnpgs; i++) {
3186 		pp = ppa[i];
3187 		ASSERT(PAGE_SHARED(pp));
3188 		ASSERT(!PP_ISFREE(pp));
3189 		pfn = page_pptonum(pp);
3190 		if (i == 0) {
3191 			if (!IS_P2ALIGNED(pfn, totnpgs)) {
3192 				contig = 0;
3193 			} else {
3194 				first_pfn = pfn;
3195 			}
3196 		} else if (contig && pfn != first_pfn + i) {
3197 			contig = 0;
3198 		}
3199 		if (pp->p_szc == 0) {
3200 			if (root) {
3201 				VM_STAT_ADD(segvnvmstats.fullszcpages[1]);
3202 				return (0);
3203 			}
3204 		} else if (!root) {
3205 			if ((curszc = pp->p_szc) >= szc) {
3206 				VM_STAT_ADD(segvnvmstats.fullszcpages[2]);
3207 				return (0);
3208 			}
3209 			if (curszc == 0) {
3210 				/*
3211 				 * p_szc changed means we don't have all pages
3212 				 * locked. return failure.
3213 				 */
3214 				VM_STAT_ADD(segvnvmstats.fullszcpages[3]);
3215 				return (0);
3216 			}
3217 			curnpgs = page_get_pagecnt(curszc);
3218 			if (!IS_P2ALIGNED(pfn, curnpgs) ||
3219 			    !IS_P2ALIGNED(i, curnpgs)) {
3220 				VM_STAT_ADD(segvnvmstats.fullszcpages[4]);
3221 				return (0);
3222 			}
3223 			root = 1;
3224 		} else {
3225 			ASSERT(i > 0);
3226 			VM_STAT_ADD(segvnvmstats.fullszcpages[5]);
3227 			if (pp->p_szc != curszc) {
3228 				VM_STAT_ADD(segvnvmstats.fullszcpages[6]);
3229 				return (0);
3230 			}
3231 			if (pfn - 1 != page_pptonum(ppa[i - 1])) {
3232 				panic("segvn_full_szcpages: "
3233 				    "large page not physically contiguous");
3234 			}
3235 			if (P2PHASE(pfn, curnpgs) == curnpgs - 1) {
3236 				root = 0;
3237 			}
3238 		}
3239 	}
3240 
3241 	for (i = 0; i < totnpgs; i++) {
3242 		ASSERT(ppa[i]->p_szc < szc);
3243 		if (!page_tryupgrade(ppa[i])) {
3244 			for (j = 0; j < i; j++) {
3245 				page_downgrade(ppa[j]);
3246 			}
3247 			*pszc = ppa[i]->p_szc;
3248 			*upgrdfail = 1;
3249 			VM_STAT_ADD(segvnvmstats.fullszcpages[7]);
3250 			return (0);
3251 		}
3252 	}
3253 
3254 	/*
3255 	 * When a page is put a free cachelist its szc is set to 0.  if file
3256 	 * system reclaimed pages from cachelist targ pages will be physically
3257 	 * contiguous with 0 p_szc.  in this case just upgrade szc of targ
3258 	 * pages without any relocations.
3259 	 * To avoid any hat issues with previous small mappings
3260 	 * hat_pageunload() the target pages first.
3261 	 */
3262 	if (contig) {
3263 		VM_STAT_ADD(segvnvmstats.fullszcpages[8]);
3264 		for (i = 0; i < totnpgs; i++) {
3265 			(void) hat_pageunload(ppa[i], HAT_FORCE_PGUNLOAD);
3266 		}
3267 		for (i = 0; i < totnpgs; i++) {
3268 			ppa[i]->p_szc = szc;
3269 		}
3270 		for (i = 0; i < totnpgs; i++) {
3271 			ASSERT(PAGE_EXCL(ppa[i]));
3272 			page_downgrade(ppa[i]);
3273 		}
3274 		if (pszc != NULL) {
3275 			*pszc = szc;
3276 		}
3277 	}
3278 	VM_STAT_ADD(segvnvmstats.fullszcpages[9]);
3279 	return (1);
3280 }
3281 
3282 /*
3283  * Create physically contiguous pages for [vp, off] - [vp, off +
3284  * page_size(szc)) range and for private segment return them in ppa array.
3285  * Pages are created either via IO or relocations.
3286  *
3287  * Return 1 on success and 0 on failure.
3288  *
3289  * If physically contiguous pages already exist for this range return 1 without
3290  * filling ppa array. Caller initializes ppa[0] as NULL to detect that ppa
3291  * array wasn't filled. In this case caller fills ppa array via VOP_GETPAGE().
3292  */
3293 
3294 static int
3295 segvn_fill_vp_pages(struct segvn_data *svd, vnode_t *vp, u_offset_t off,
3296     uint_t szc, page_t **ppa, page_t **ppplist, uint_t *ret_pszc,
3297     int *downsize)
3298 
3299 {
3300 	page_t *pplist = *ppplist;
3301 	size_t pgsz = page_get_pagesize(szc);
3302 	pgcnt_t pages = btop(pgsz);
3303 	ulong_t start_off = off;
3304 	u_offset_t eoff = off + pgsz;
3305 	spgcnt_t nreloc;
3306 	u_offset_t io_off = off;
3307 	size_t io_len;
3308 	page_t *io_pplist = NULL;
3309 	page_t *done_pplist = NULL;
3310 	pgcnt_t pgidx = 0;
3311 	page_t *pp;
3312 	page_t *newpp;
3313 	page_t *targpp;
3314 	int io_err = 0;
3315 	int i;
3316 	pfn_t pfn;
3317 	ulong_t ppages;
3318 	page_t *targ_pplist = NULL;
3319 	page_t *repl_pplist = NULL;
3320 	page_t *tmp_pplist;
3321 	int nios = 0;
3322 	uint_t pszc;
3323 	struct vattr va;
3324 
3325 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[0]);
3326 
3327 	ASSERT(szc != 0);
3328 	ASSERT(pplist->p_szc == szc);
3329 
3330 	/*
3331 	 * downsize will be set to 1 only if we fail to lock pages. this will
3332 	 * allow subsequent faults to try to relocate the page again. If we
3333 	 * fail due to misalignment don't downsize and let the caller map the
3334 	 * whole region with small mappings to avoid more faults into the area
3335 	 * where we can't get large pages anyway.
3336 	 */
3337 	*downsize = 0;
3338 
3339 	while (off < eoff) {
3340 		newpp = pplist;
3341 		ASSERT(newpp != NULL);
3342 		ASSERT(PAGE_EXCL(newpp));
3343 		ASSERT(!PP_ISFREE(newpp));
3344 		/*
3345 		 * we pass NULL for nrelocp to page_lookup_create()
3346 		 * so that it doesn't relocate. We relocate here
3347 		 * later only after we make sure we can lock all
3348 		 * pages in the range we handle and they are all
3349 		 * aligned.
3350 		 */
3351 		pp = page_lookup_create(vp, off, SE_SHARED, newpp, NULL, 0);
3352 		ASSERT(pp != NULL);
3353 		ASSERT(!PP_ISFREE(pp));
3354 		ASSERT(pp->p_vnode == vp);
3355 		ASSERT(pp->p_offset == off);
3356 		if (pp == newpp) {
3357 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[1]);
3358 			page_sub(&pplist, pp);
3359 			ASSERT(PAGE_EXCL(pp));
3360 			ASSERT(page_iolock_assert(pp));
3361 			page_list_concat(&io_pplist, &pp);
3362 			off += PAGESIZE;
3363 			continue;
3364 		}
3365 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[2]);
3366 		pfn = page_pptonum(pp);
3367 		pszc = pp->p_szc;
3368 		if (pszc >= szc && targ_pplist == NULL && io_pplist == NULL &&
3369 		    IS_P2ALIGNED(pfn, pages)) {
3370 			ASSERT(repl_pplist == NULL);
3371 			ASSERT(done_pplist == NULL);
3372 			ASSERT(pplist == *ppplist);
3373 			page_unlock(pp);
3374 			page_free_replacement_page(pplist);
3375 			page_create_putback(pages);
3376 			*ppplist = NULL;
3377 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[3]);
3378 			return (1);
3379 		}
3380 		if (pszc >= szc) {
3381 			page_unlock(pp);
3382 			segvn_faultvnmpss_align_err1++;
3383 			goto out;
3384 		}
3385 		ppages = page_get_pagecnt(pszc);
3386 		if (!IS_P2ALIGNED(pfn, ppages)) {
3387 			ASSERT(pszc > 0);
3388 			/*
3389 			 * sizing down to pszc won't help.
3390 			 */
3391 			page_unlock(pp);
3392 			segvn_faultvnmpss_align_err2++;
3393 			goto out;
3394 		}
3395 		pfn = page_pptonum(newpp);
3396 		if (!IS_P2ALIGNED(pfn, ppages)) {
3397 			ASSERT(pszc > 0);
3398 			/*
3399 			 * sizing down to pszc won't help.
3400 			 */
3401 			page_unlock(pp);
3402 			segvn_faultvnmpss_align_err3++;
3403 			goto out;
3404 		}
3405 		if (!PAGE_EXCL(pp)) {
3406 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[4]);
3407 			page_unlock(pp);
3408 			*downsize = 1;
3409 			*ret_pszc = pp->p_szc;
3410 			goto out;
3411 		}
3412 		targpp = pp;
3413 		if (io_pplist != NULL) {
3414 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[5]);
3415 			io_len = off - io_off;
3416 			/*
3417 			 * Some file systems like NFS don't check EOF
3418 			 * conditions in VOP_PAGEIO(). Check it here
3419 			 * now that pages are locked SE_EXCL. Any file
3420 			 * truncation will wait until the pages are
3421 			 * unlocked so no need to worry that file will
3422 			 * be truncated after we check its size here.
3423 			 * XXX fix NFS to remove this check.
3424 			 */
3425 			va.va_mask = AT_SIZE;
3426 			if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL)) {
3427 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[6]);
3428 				page_unlock(targpp);
3429 				goto out;
3430 			}
3431 			if (btopr(va.va_size) < btopr(io_off + io_len)) {
3432 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[7]);
3433 				*downsize = 1;
3434 				*ret_pszc = 0;
3435 				page_unlock(targpp);
3436 				goto out;
3437 			}
3438 			io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3439 				B_READ, svd->cred, NULL);
3440 			if (io_err) {
3441 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[8]);
3442 				page_unlock(targpp);
3443 				if (io_err == EDEADLK) {
3444 					segvn_vmpss_pageio_deadlk_err++;
3445 				}
3446 				goto out;
3447 			}
3448 			nios++;
3449 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[9]);
3450 			while (io_pplist != NULL) {
3451 				pp = io_pplist;
3452 				page_sub(&io_pplist, pp);
3453 				ASSERT(page_iolock_assert(pp));
3454 				page_io_unlock(pp);
3455 				pgidx = (pp->p_offset - start_off) >>
3456 				    PAGESHIFT;
3457 				ASSERT(pgidx < pages);
3458 				ppa[pgidx] = pp;
3459 				page_list_concat(&done_pplist, &pp);
3460 			}
3461 		}
3462 		pp = targpp;
3463 		ASSERT(PAGE_EXCL(pp));
3464 		ASSERT(pp->p_szc <= pszc);
3465 		if (pszc != 0 && !group_page_trylock(pp, SE_EXCL)) {
3466 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[10]);
3467 			page_unlock(pp);
3468 			*downsize = 1;
3469 			*ret_pszc = pp->p_szc;
3470 			goto out;
3471 		}
3472 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[11]);
3473 		/*
3474 		 * page szc chould have changed before the entire group was
3475 		 * locked. reread page szc.
3476 		 */
3477 		pszc = pp->p_szc;
3478 		ppages = page_get_pagecnt(pszc);
3479 
3480 		/* link just the roots */
3481 		page_list_concat(&targ_pplist, &pp);
3482 		page_sub(&pplist, newpp);
3483 		page_list_concat(&repl_pplist, &newpp);
3484 		off += PAGESIZE;
3485 		while (--ppages != 0) {
3486 			newpp = pplist;
3487 			page_sub(&pplist, newpp);
3488 			off += PAGESIZE;
3489 		}
3490 		io_off = off;
3491 	}
3492 	if (io_pplist != NULL) {
3493 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[12]);
3494 		io_len = eoff - io_off;
3495 		va.va_mask = AT_SIZE;
3496 		if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL) != 0) {
3497 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[13]);
3498 			goto out;
3499 		}
3500 		if (btopr(va.va_size) < btopr(io_off + io_len)) {
3501 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[14]);
3502 			*downsize = 1;
3503 			*ret_pszc = 0;
3504 			goto out;
3505 		}
3506 		io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3507 		    B_READ, svd->cred, NULL);
3508 		if (io_err) {
3509 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[15]);
3510 			if (io_err == EDEADLK) {
3511 				segvn_vmpss_pageio_deadlk_err++;
3512 			}
3513 			goto out;
3514 		}
3515 		nios++;
3516 		while (io_pplist != NULL) {
3517 			pp = io_pplist;
3518 			page_sub(&io_pplist, pp);
3519 			ASSERT(page_iolock_assert(pp));
3520 			page_io_unlock(pp);
3521 			pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3522 			ASSERT(pgidx < pages);
3523 			ppa[pgidx] = pp;
3524 		}
3525 	}
3526 	/*
3527 	 * we're now bound to succeed or panic.
3528 	 * remove pages from done_pplist. it's not needed anymore.
3529 	 */
3530 	while (done_pplist != NULL) {
3531 		pp = done_pplist;
3532 		page_sub(&done_pplist, pp);
3533 	}
3534 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[16]);
3535 	ASSERT(pplist == NULL);
3536 	*ppplist = NULL;
3537 	while (targ_pplist != NULL) {
3538 		int ret;
3539 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[17]);
3540 		ASSERT(repl_pplist);
3541 		pp = targ_pplist;
3542 		page_sub(&targ_pplist, pp);
3543 		pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3544 		newpp = repl_pplist;
3545 		page_sub(&repl_pplist, newpp);
3546 #ifdef DEBUG
3547 		pfn = page_pptonum(pp);
3548 		pszc = pp->p_szc;
3549 		ppages = page_get_pagecnt(pszc);
3550 		ASSERT(IS_P2ALIGNED(pfn, ppages));
3551 		pfn = page_pptonum(newpp);
3552 		ASSERT(IS_P2ALIGNED(pfn, ppages));
3553 		ASSERT(P2PHASE(pfn, pages) == pgidx);
3554 #endif
3555 		nreloc = 0;
3556 		ret = page_relocate(&pp, &newpp, 0, 1, &nreloc, NULL);
3557 		if (ret != 0 || nreloc == 0) {
3558 			panic("segvn_fill_vp_pages: "
3559 			    "page_relocate failed");
3560 		}
3561 		pp = newpp;
3562 		while (nreloc-- != 0) {
3563 			ASSERT(PAGE_EXCL(pp));
3564 			ASSERT(pp->p_vnode == vp);
3565 			ASSERT(pgidx ==
3566 			    ((pp->p_offset - start_off) >> PAGESHIFT));
3567 			ppa[pgidx++] = pp;
3568 			pp++;
3569 		}
3570 	}
3571 
3572 	if (svd->type == MAP_PRIVATE) {
3573 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[18]);
3574 		for (i = 0; i < pages; i++) {
3575 			ASSERT(ppa[i] != NULL);
3576 			ASSERT(PAGE_EXCL(ppa[i]));
3577 			ASSERT(ppa[i]->p_vnode == vp);
3578 			ASSERT(ppa[i]->p_offset ==
3579 			    start_off + (i << PAGESHIFT));
3580 			page_downgrade(ppa[i]);
3581 		}
3582 		ppa[pages] = NULL;
3583 	} else {
3584 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[19]);
3585 		/*
3586 		 * the caller will still call VOP_GETPAGE() for shared segments
3587 		 * to check FS write permissions. For private segments we map
3588 		 * file read only anyway.  so no VOP_GETPAGE is needed.
3589 		 */
3590 		for (i = 0; i < pages; i++) {
3591 			ASSERT(ppa[i] != NULL);
3592 			ASSERT(PAGE_EXCL(ppa[i]));
3593 			ASSERT(ppa[i]->p_vnode == vp);
3594 			ASSERT(ppa[i]->p_offset ==
3595 			    start_off + (i << PAGESHIFT));
3596 			page_unlock(ppa[i]);
3597 		}
3598 		ppa[0] = NULL;
3599 	}
3600 
3601 	return (1);
3602 out:
3603 	/*
3604 	 * Do the cleanup. Unlock target pages we didn't relocate. They are
3605 	 * linked on targ_pplist by root pages. reassemble unused replacement
3606 	 * and io pages back to pplist.
3607 	 */
3608 	if (io_pplist != NULL) {
3609 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[20]);
3610 		pp = io_pplist;
3611 		do {
3612 			ASSERT(pp->p_vnode == vp);
3613 			ASSERT(pp->p_offset == io_off);
3614 			ASSERT(page_iolock_assert(pp));
3615 			page_io_unlock(pp);
3616 			page_hashout(pp, NULL);
3617 			io_off += PAGESIZE;
3618 		} while ((pp = pp->p_next) != io_pplist);
3619 		page_list_concat(&io_pplist, &pplist);
3620 		pplist = io_pplist;
3621 	}
3622 	tmp_pplist = NULL;
3623 	while (targ_pplist != NULL) {
3624 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[21]);
3625 		pp = targ_pplist;
3626 		ASSERT(PAGE_EXCL(pp));
3627 		page_sub(&targ_pplist, pp);
3628 
3629 		pszc = pp->p_szc;
3630 		ppages = page_get_pagecnt(pszc);
3631 		ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3632 
3633 		if (pszc != 0) {
3634 			group_page_unlock(pp);
3635 		}
3636 		page_unlock(pp);
3637 
3638 		pp = repl_pplist;
3639 		ASSERT(pp != NULL);
3640 		ASSERT(PAGE_EXCL(pp));
3641 		ASSERT(pp->p_szc == szc);
3642 		page_sub(&repl_pplist, pp);
3643 
3644 		ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3645 
3646 		/* relink replacement page */
3647 		page_list_concat(&tmp_pplist, &pp);
3648 		while (--ppages != 0) {
3649 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[22]);
3650 			pp++;
3651 			ASSERT(PAGE_EXCL(pp));
3652 			ASSERT(pp->p_szc == szc);
3653 			page_list_concat(&tmp_pplist, &pp);
3654 		}
3655 	}
3656 	if (tmp_pplist != NULL) {
3657 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[23]);
3658 		page_list_concat(&tmp_pplist, &pplist);
3659 		pplist = tmp_pplist;
3660 	}
3661 	/*
3662 	 * at this point all pages are either on done_pplist or
3663 	 * pplist. They can't be all on done_pplist otherwise
3664 	 * we'd've been done.
3665 	 */
3666 	ASSERT(pplist != NULL);
3667 	if (nios != 0) {
3668 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[24]);
3669 		pp = pplist;
3670 		do {
3671 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[25]);
3672 			ASSERT(pp->p_szc == szc);
3673 			ASSERT(PAGE_EXCL(pp));
3674 			ASSERT(pp->p_vnode != vp);
3675 			pp->p_szc = 0;
3676 		} while ((pp = pp->p_next) != pplist);
3677 
3678 		pp = done_pplist;
3679 		do {
3680 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[26]);
3681 			ASSERT(pp->p_szc == szc);
3682 			ASSERT(PAGE_EXCL(pp));
3683 			ASSERT(pp->p_vnode == vp);
3684 			pp->p_szc = 0;
3685 		} while ((pp = pp->p_next) != done_pplist);
3686 
3687 		while (pplist != NULL) {
3688 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[27]);
3689 			pp = pplist;
3690 			page_sub(&pplist, pp);
3691 			page_free(pp, 0);
3692 		}
3693 
3694 		while (done_pplist != NULL) {
3695 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[28]);
3696 			pp = done_pplist;
3697 			page_sub(&done_pplist, pp);
3698 			page_unlock(pp);
3699 		}
3700 		*ppplist = NULL;
3701 		return (0);
3702 	}
3703 	ASSERT(pplist == *ppplist);
3704 	if (io_err) {
3705 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[29]);
3706 		/*
3707 		 * don't downsize on io error.
3708 		 * see if vop_getpage succeeds.
3709 		 * pplist may still be used in this case
3710 		 * for relocations.
3711 		 */
3712 		return (0);
3713 	}
3714 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[30]);
3715 	page_free_replacement_page(pplist);
3716 	page_create_putback(pages);
3717 	*ppplist = NULL;
3718 	return (0);
3719 }
3720 
3721 int segvn_anypgsz = 0;
3722 
3723 #define	SEGVN_RESTORE_SOFTLOCK_VP(type, pages) 				\
3724 		if ((type) == F_SOFTLOCK) {				\
3725 			atomic_add_long((ulong_t *)&(svd)->softlockcnt, \
3726 			    -(pages));					\
3727 		}
3728 
3729 #define	SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot)		\
3730 		if (IS_VMODSORT((ppa)[0]->p_vnode)) {			\
3731 			if ((rw) == S_WRITE) {				\
3732 				for (i = 0; i < (pages); i++) {		\
3733 					ASSERT((ppa)[i]->p_vnode ==	\
3734 					    (ppa)[0]->p_vnode);		\
3735 					hat_setmod((ppa)[i]);		\
3736 				}					\
3737 			} else if ((rw) != S_OTHER &&			\
3738 			    ((prot) & (vpprot) & PROT_WRITE)) {		\
3739 				for (i = 0; i < (pages); i++) {		\
3740 					ASSERT((ppa)[i]->p_vnode ==	\
3741 					    (ppa)[0]->p_vnode);		\
3742 					if (!hat_ismod((ppa)[i])) {	\
3743 						prot &= ~PROT_WRITE;	\
3744 						break;			\
3745 					}				\
3746 				}					\
3747 			}						\
3748 		}
3749 
3750 #ifdef  VM_STATS
3751 
3752 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)					\
3753 		VM_STAT_ADD(segvnvmstats.fltvnpages[(idx)]);
3754 
3755 #else /* VM_STATS */
3756 
3757 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)
3758 
3759 #endif
3760 
3761 static faultcode_t
3762 segvn_fault_vnodepages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
3763     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
3764     caddr_t eaddr, int brkcow)
3765 {
3766 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
3767 	struct anon_map *amp = svd->amp;
3768 	uchar_t segtype = svd->type;
3769 	uint_t szc = seg->s_szc;
3770 	size_t pgsz = page_get_pagesize(szc);
3771 	size_t maxpgsz = pgsz;
3772 	pgcnt_t pages = btop(pgsz);
3773 	pgcnt_t maxpages = pages;
3774 	size_t ppasize = (pages + 1) * sizeof (page_t *);
3775 	caddr_t a = lpgaddr;
3776 	caddr_t	maxlpgeaddr = lpgeaddr;
3777 	u_offset_t off = svd->offset + (uintptr_t)(a - seg->s_base);
3778 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
3779 	struct vpage *vpage = (svd->vpage != NULL) ?
3780 	    &svd->vpage[seg_page(seg, a)] : NULL;
3781 	vnode_t *vp = svd->vp;
3782 	page_t **ppa;
3783 	uint_t	pszc;
3784 	size_t	ppgsz;
3785 	pgcnt_t	ppages;
3786 	faultcode_t err = 0;
3787 	int ierr;
3788 	int vop_size_err = 0;
3789 	uint_t protchk, prot, vpprot;
3790 	ulong_t i;
3791 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
3792 	anon_sync_obj_t an_cookie;
3793 	enum seg_rw arw;
3794 	int alloc_failed = 0;
3795 	int adjszc_chk;
3796 	struct vattr va;
3797 	int xhat = 0;
3798 	page_t *pplist;
3799 	pfn_t pfn;
3800 	int physcontig;
3801 	int upgrdfail;
3802 	int segvn_anypgsz_vnode = 0; /* for now map vnode with 2 page sizes */
3803 	int tron = (svd->tr_state == SEGVN_TR_ON);
3804 
3805 	ASSERT(szc != 0);
3806 	ASSERT(vp != NULL);
3807 	ASSERT(brkcow == 0 || amp != NULL);
3808 	ASSERT(tron == 0 || amp != NULL);
3809 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
3810 	ASSERT(!(svd->flags & MAP_NORESERVE));
3811 	ASSERT(type != F_SOFTUNLOCK);
3812 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
3813 	ASSERT(amp == NULL || IS_P2ALIGNED(aindx, maxpages));
3814 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
3815 	ASSERT(seg->s_szc < NBBY * sizeof (int));
3816 	ASSERT(type != F_SOFTLOCK || lpgeaddr - a == maxpgsz);
3817 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
3818 
3819 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltvnpages[0]);
3820 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltvnpages[1]);
3821 
3822 	if (svd->flags & MAP_TEXT) {
3823 		hat_flag |= HAT_LOAD_TEXT;
3824 	}
3825 
3826 	if (svd->pageprot) {
3827 		switch (rw) {
3828 		case S_READ:
3829 			protchk = PROT_READ;
3830 			break;
3831 		case S_WRITE:
3832 			protchk = PROT_WRITE;
3833 			break;
3834 		case S_EXEC:
3835 			protchk = PROT_EXEC;
3836 			break;
3837 		case S_OTHER:
3838 		default:
3839 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
3840 			break;
3841 		}
3842 	} else {
3843 		prot = svd->prot;
3844 		/* caller has already done segment level protection check. */
3845 	}
3846 
3847 	if (seg->s_as->a_hat != hat) {
3848 		xhat = 1;
3849 	}
3850 
3851 	if (rw == S_WRITE && segtype == MAP_PRIVATE) {
3852 		SEGVN_VMSTAT_FLTVNPAGES(2);
3853 		arw = S_READ;
3854 	} else {
3855 		arw = rw;
3856 	}
3857 
3858 	ppa = kmem_alloc(ppasize, KM_SLEEP);
3859 
3860 	VM_STAT_COND_ADD(amp != NULL, segvnvmstats.fltvnpages[3]);
3861 
3862 	for (;;) {
3863 		adjszc_chk = 0;
3864 		for (; a < lpgeaddr; a += pgsz, off += pgsz, aindx += pages) {
3865 			if (adjszc_chk) {
3866 				while (szc < seg->s_szc) {
3867 					uintptr_t e;
3868 					uint_t tszc;
3869 					tszc = segvn_anypgsz_vnode ? szc + 1 :
3870 					    seg->s_szc;
3871 					ppgsz = page_get_pagesize(tszc);
3872 					if (!IS_P2ALIGNED(a, ppgsz) ||
3873 					    ((alloc_failed >> tszc) & 0x1)) {
3874 						break;
3875 					}
3876 					SEGVN_VMSTAT_FLTVNPAGES(4);
3877 					szc = tszc;
3878 					pgsz = ppgsz;
3879 					pages = btop(pgsz);
3880 					e = P2ROUNDUP((uintptr_t)eaddr, pgsz);
3881 					lpgeaddr = (caddr_t)e;
3882 				}
3883 			}
3884 
3885 		again:
3886 			if (IS_P2ALIGNED(a, maxpgsz) && amp != NULL) {
3887 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
3888 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
3889 				anon_array_enter(amp, aindx, &an_cookie);
3890 				if (anon_get_ptr(amp->ahp, aindx) != NULL) {
3891 					SEGVN_VMSTAT_FLTVNPAGES(5);
3892 					ASSERT(anon_pages(amp->ahp, aindx,
3893 					    maxpages) == maxpages);
3894 					anon_array_exit(&an_cookie);
3895 					ANON_LOCK_EXIT(&amp->a_rwlock);
3896 					err = segvn_fault_anonpages(hat, seg,
3897 					    a, a + maxpgsz, type, rw,
3898 					    MAX(a, addr),
3899 					    MIN(a + maxpgsz, eaddr), brkcow);
3900 					if (err != 0) {
3901 						SEGVN_VMSTAT_FLTVNPAGES(6);
3902 						goto out;
3903 					}
3904 					if (szc < seg->s_szc) {
3905 						szc = seg->s_szc;
3906 						pgsz = maxpgsz;
3907 						pages = maxpages;
3908 						lpgeaddr = maxlpgeaddr;
3909 					}
3910 					goto next;
3911 				} else {
3912 					ASSERT(anon_pages(amp->ahp, aindx,
3913 					    maxpages) == 0);
3914 					SEGVN_VMSTAT_FLTVNPAGES(7);
3915 					anon_array_exit(&an_cookie);
3916 					ANON_LOCK_EXIT(&amp->a_rwlock);
3917 				}
3918 			}
3919 			ASSERT(!brkcow || IS_P2ALIGNED(a, maxpgsz));
3920 			ASSERT(!tron || IS_P2ALIGNED(a, maxpgsz));
3921 
3922 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
3923 				ASSERT(vpage != NULL);
3924 				prot = VPP_PROT(vpage);
3925 				ASSERT(sameprot(seg, a, maxpgsz));
3926 				if ((prot & protchk) == 0) {
3927 					SEGVN_VMSTAT_FLTVNPAGES(8);
3928 					err = FC_PROT;
3929 					goto out;
3930 				}
3931 			}
3932 			if (type == F_SOFTLOCK) {
3933 				atomic_add_long((ulong_t *)&svd->softlockcnt,
3934 				    pages);
3935 			}
3936 
3937 			pplist = NULL;
3938 			physcontig = 0;
3939 			ppa[0] = NULL;
3940 			if (!brkcow && !tron && szc &&
3941 			    !page_exists_physcontig(vp, off, szc,
3942 			    segtype == MAP_PRIVATE ? ppa : NULL)) {
3943 				SEGVN_VMSTAT_FLTVNPAGES(9);
3944 				if (page_alloc_pages(vp, seg, a, &pplist, NULL,
3945 				    szc, 0, 0) && type != F_SOFTLOCK) {
3946 					SEGVN_VMSTAT_FLTVNPAGES(10);
3947 					pszc = 0;
3948 					ierr = -1;
3949 					alloc_failed |= (1 << szc);
3950 					break;
3951 				}
3952 				if (pplist != NULL &&
3953 				    vp->v_mpssdata == SEGVN_PAGEIO) {
3954 					int downsize;
3955 					SEGVN_VMSTAT_FLTVNPAGES(11);
3956 					physcontig = segvn_fill_vp_pages(svd,
3957 					    vp, off, szc, ppa, &pplist,
3958 					    &pszc, &downsize);
3959 					ASSERT(!physcontig || pplist == NULL);
3960 					if (!physcontig && downsize &&
3961 					    type != F_SOFTLOCK) {
3962 						ASSERT(pplist == NULL);
3963 						SEGVN_VMSTAT_FLTVNPAGES(12);
3964 						ierr = -1;
3965 						break;
3966 					}
3967 					ASSERT(!physcontig ||
3968 					    segtype == MAP_PRIVATE ||
3969 					    ppa[0] == NULL);
3970 					if (physcontig && ppa[0] == NULL) {
3971 						physcontig = 0;
3972 					}
3973 				}
3974 			} else if (!brkcow && !tron && szc && ppa[0] != NULL) {
3975 				SEGVN_VMSTAT_FLTVNPAGES(13);
3976 				ASSERT(segtype == MAP_PRIVATE);
3977 				physcontig = 1;
3978 			}
3979 
3980 			if (!physcontig) {
3981 				SEGVN_VMSTAT_FLTVNPAGES(14);
3982 				ppa[0] = NULL;
3983 				ierr = VOP_GETPAGE(vp, (offset_t)off, pgsz,
3984 				    &vpprot, ppa, pgsz, seg, a, arw,
3985 				    svd->cred, NULL);
3986 #ifdef DEBUG
3987 				if (ierr == 0) {
3988 					for (i = 0; i < pages; i++) {
3989 						ASSERT(PAGE_LOCKED(ppa[i]));
3990 						ASSERT(!PP_ISFREE(ppa[i]));
3991 						ASSERT(ppa[i]->p_vnode == vp);
3992 						ASSERT(ppa[i]->p_offset ==
3993 						    off + (i << PAGESHIFT));
3994 					}
3995 				}
3996 #endif /* DEBUG */
3997 				if (segtype == MAP_PRIVATE) {
3998 					SEGVN_VMSTAT_FLTVNPAGES(15);
3999 					vpprot &= ~PROT_WRITE;
4000 				}
4001 			} else {
4002 				ASSERT(segtype == MAP_PRIVATE);
4003 				SEGVN_VMSTAT_FLTVNPAGES(16);
4004 				vpprot = PROT_ALL & ~PROT_WRITE;
4005 				ierr = 0;
4006 			}
4007 
4008 			if (ierr != 0) {
4009 				SEGVN_VMSTAT_FLTVNPAGES(17);
4010 				if (pplist != NULL) {
4011 					SEGVN_VMSTAT_FLTVNPAGES(18);
4012 					page_free_replacement_page(pplist);
4013 					page_create_putback(pages);
4014 				}
4015 				SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4016 				if (a + pgsz <= eaddr) {
4017 					SEGVN_VMSTAT_FLTVNPAGES(19);
4018 					err = FC_MAKE_ERR(ierr);
4019 					goto out;
4020 				}
4021 				va.va_mask = AT_SIZE;
4022 				if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL)) {
4023 					SEGVN_VMSTAT_FLTVNPAGES(20);
4024 					err = FC_MAKE_ERR(EIO);
4025 					goto out;
4026 				}
4027 				if (btopr(va.va_size) >= btopr(off + pgsz)) {
4028 					SEGVN_VMSTAT_FLTVNPAGES(21);
4029 					err = FC_MAKE_ERR(ierr);
4030 					goto out;
4031 				}
4032 				if (btopr(va.va_size) <
4033 				    btopr(off + (eaddr - a))) {
4034 					SEGVN_VMSTAT_FLTVNPAGES(22);
4035 					err = FC_MAKE_ERR(ierr);
4036 					goto out;
4037 				}
4038 				if (brkcow || tron || type == F_SOFTLOCK) {
4039 					/* can't reduce map area */
4040 					SEGVN_VMSTAT_FLTVNPAGES(23);
4041 					vop_size_err = 1;
4042 					goto out;
4043 				}
4044 				SEGVN_VMSTAT_FLTVNPAGES(24);
4045 				ASSERT(szc != 0);
4046 				pszc = 0;
4047 				ierr = -1;
4048 				break;
4049 			}
4050 
4051 			if (amp != NULL) {
4052 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4053 				anon_array_enter(amp, aindx, &an_cookie);
4054 			}
4055 			if (amp != NULL &&
4056 			    anon_get_ptr(amp->ahp, aindx) != NULL) {
4057 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4058 
4059 				SEGVN_VMSTAT_FLTVNPAGES(25);
4060 				ASSERT(anon_pages(amp->ahp, taindx,
4061 				    maxpages) == maxpages);
4062 				for (i = 0; i < pages; i++) {
4063 					page_unlock(ppa[i]);
4064 				}
4065 				anon_array_exit(&an_cookie);
4066 				ANON_LOCK_EXIT(&amp->a_rwlock);
4067 				if (pplist != NULL) {
4068 					page_free_replacement_page(pplist);
4069 					page_create_putback(pages);
4070 				}
4071 				SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4072 				if (szc < seg->s_szc) {
4073 					SEGVN_VMSTAT_FLTVNPAGES(26);
4074 					/*
4075 					 * For private segments SOFTLOCK
4076 					 * either always breaks cow (any rw
4077 					 * type except S_READ_NOCOW) or
4078 					 * address space is locked as writer
4079 					 * (S_READ_NOCOW case) and anon slots
4080 					 * can't show up on second check.
4081 					 * Therefore if we are here for
4082 					 * SOFTLOCK case it must be a cow
4083 					 * break but cow break never reduces
4084 					 * szc. text replication (tron) in
4085 					 * this case works as cow break.
4086 					 * Thus the assert below.
4087 					 */
4088 					ASSERT(!brkcow && !tron &&
4089 					    type != F_SOFTLOCK);
4090 					pszc = seg->s_szc;
4091 					ierr = -2;
4092 					break;
4093 				}
4094 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4095 				goto again;
4096 			}
4097 #ifdef DEBUG
4098 			if (amp != NULL) {
4099 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4100 				ASSERT(!anon_pages(amp->ahp, taindx, maxpages));
4101 			}
4102 #endif /* DEBUG */
4103 
4104 			if (brkcow || tron) {
4105 				ASSERT(amp != NULL);
4106 				ASSERT(pplist == NULL);
4107 				ASSERT(szc == seg->s_szc);
4108 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4109 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
4110 				SEGVN_VMSTAT_FLTVNPAGES(27);
4111 				ierr = anon_map_privatepages(amp, aindx, szc,
4112 				    seg, a, prot, ppa, vpage, segvn_anypgsz,
4113 				    tron ? PG_LOCAL : 0, svd->cred);
4114 				if (ierr != 0) {
4115 					SEGVN_VMSTAT_FLTVNPAGES(28);
4116 					anon_array_exit(&an_cookie);
4117 					ANON_LOCK_EXIT(&amp->a_rwlock);
4118 					SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4119 					err = FC_MAKE_ERR(ierr);
4120 					goto out;
4121 				}
4122 
4123 				ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4124 				/*
4125 				 * p_szc can't be changed for locked
4126 				 * swapfs pages.
4127 				 */
4128 				ASSERT(svd->rcookie ==
4129 				    HAT_INVALID_REGION_COOKIE);
4130 				hat_memload_array(hat, a, pgsz, ppa, prot,
4131 				    hat_flag);
4132 
4133 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4134 					SEGVN_VMSTAT_FLTVNPAGES(29);
4135 					for (i = 0; i < pages; i++) {
4136 						page_unlock(ppa[i]);
4137 					}
4138 				}
4139 				anon_array_exit(&an_cookie);
4140 				ANON_LOCK_EXIT(&amp->a_rwlock);
4141 				goto next;
4142 			}
4143 
4144 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
4145 			    (!svd->pageprot && svd->prot == (prot & vpprot)));
4146 
4147 			pfn = page_pptonum(ppa[0]);
4148 			/*
4149 			 * hat_page_demote() needs an SE_EXCL lock on one of
4150 			 * constituent page_t's and it decreases root's p_szc
4151 			 * last. This means if root's p_szc is equal szc and
4152 			 * all its constituent pages are locked
4153 			 * hat_page_demote() that could have changed p_szc to
4154 			 * szc is already done and no new have page_demote()
4155 			 * can start for this large page.
4156 			 */
4157 
4158 			/*
4159 			 * we need to make sure same mapping size is used for
4160 			 * the same address range if there's a possibility the
4161 			 * adddress is already mapped because hat layer panics
4162 			 * when translation is loaded for the range already
4163 			 * mapped with a different page size.  We achieve it
4164 			 * by always using largest page size possible subject
4165 			 * to the constraints of page size, segment page size
4166 			 * and page alignment.  Since mappings are invalidated
4167 			 * when those constraints change and make it
4168 			 * impossible to use previously used mapping size no
4169 			 * mapping size conflicts should happen.
4170 			 */
4171 
4172 		chkszc:
4173 			if ((pszc = ppa[0]->p_szc) == szc &&
4174 			    IS_P2ALIGNED(pfn, pages)) {
4175 
4176 				SEGVN_VMSTAT_FLTVNPAGES(30);
4177 #ifdef DEBUG
4178 				for (i = 0; i < pages; i++) {
4179 					ASSERT(PAGE_LOCKED(ppa[i]));
4180 					ASSERT(!PP_ISFREE(ppa[i]));
4181 					ASSERT(page_pptonum(ppa[i]) ==
4182 					    pfn + i);
4183 					ASSERT(ppa[i]->p_szc == szc);
4184 					ASSERT(ppa[i]->p_vnode == vp);
4185 					ASSERT(ppa[i]->p_offset ==
4186 					    off + (i << PAGESHIFT));
4187 				}
4188 #endif /* DEBUG */
4189 				/*
4190 				 * All pages are of szc we need and they are
4191 				 * all locked so they can't change szc. load
4192 				 * translations.
4193 				 *
4194 				 * if page got promoted since last check
4195 				 * we don't need pplist.
4196 				 */
4197 				if (pplist != NULL) {
4198 					page_free_replacement_page(pplist);
4199 					page_create_putback(pages);
4200 				}
4201 				if (PP_ISMIGRATE(ppa[0])) {
4202 					page_migrate(seg, a, ppa, pages);
4203 				}
4204 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4205 				    prot, vpprot);
4206 				if (!xhat) {
4207 					hat_memload_array_region(hat, a, pgsz,
4208 					    ppa, prot & vpprot, hat_flag,
4209 					    svd->rcookie);
4210 				} else {
4211 					/*
4212 					 * avoid large xhat mappings to FS
4213 					 * pages so that hat_page_demote()
4214 					 * doesn't need to check for xhat
4215 					 * large mappings.
4216 					 * Don't use regions with xhats.
4217 					 */
4218 					for (i = 0; i < pages; i++) {
4219 						hat_memload(hat,
4220 						    a + (i << PAGESHIFT),
4221 						    ppa[i], prot & vpprot,
4222 						    hat_flag);
4223 					}
4224 				}
4225 
4226 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4227 					for (i = 0; i < pages; i++) {
4228 						page_unlock(ppa[i]);
4229 					}
4230 				}
4231 				if (amp != NULL) {
4232 					anon_array_exit(&an_cookie);
4233 					ANON_LOCK_EXIT(&amp->a_rwlock);
4234 				}
4235 				goto next;
4236 			}
4237 
4238 			/*
4239 			 * See if upsize is possible.
4240 			 */
4241 			if (pszc > szc && szc < seg->s_szc &&
4242 			    (segvn_anypgsz_vnode || pszc >= seg->s_szc)) {
4243 				pgcnt_t aphase;
4244 				uint_t pszc1 = MIN(pszc, seg->s_szc);
4245 				ppgsz = page_get_pagesize(pszc1);
4246 				ppages = btop(ppgsz);
4247 				aphase = btop(P2PHASE((uintptr_t)a, ppgsz));
4248 
4249 				ASSERT(type != F_SOFTLOCK);
4250 
4251 				SEGVN_VMSTAT_FLTVNPAGES(31);
4252 				if (aphase != P2PHASE(pfn, ppages)) {
4253 					segvn_faultvnmpss_align_err4++;
4254 				} else {
4255 					SEGVN_VMSTAT_FLTVNPAGES(32);
4256 					if (pplist != NULL) {
4257 						page_t *pl = pplist;
4258 						page_free_replacement_page(pl);
4259 						page_create_putback(pages);
4260 					}
4261 					for (i = 0; i < pages; i++) {
4262 						page_unlock(ppa[i]);
4263 					}
4264 					if (amp != NULL) {
4265 						anon_array_exit(&an_cookie);
4266 						ANON_LOCK_EXIT(&amp->a_rwlock);
4267 					}
4268 					pszc = pszc1;
4269 					ierr = -2;
4270 					break;
4271 				}
4272 			}
4273 
4274 			/*
4275 			 * check if we should use smallest mapping size.
4276 			 */
4277 			upgrdfail = 0;
4278 			if (szc == 0 || xhat ||
4279 			    (pszc >= szc &&
4280 			    !IS_P2ALIGNED(pfn, pages)) ||
4281 			    (pszc < szc &&
4282 			    !segvn_full_szcpages(ppa, szc, &upgrdfail,
4283 			    &pszc))) {
4284 
4285 				if (upgrdfail && type != F_SOFTLOCK) {
4286 					/*
4287 					 * segvn_full_szcpages failed to lock
4288 					 * all pages EXCL. Size down.
4289 					 */
4290 					ASSERT(pszc < szc);
4291 
4292 					SEGVN_VMSTAT_FLTVNPAGES(33);
4293 
4294 					if (pplist != NULL) {
4295 						page_t *pl = pplist;
4296 						page_free_replacement_page(pl);
4297 						page_create_putback(pages);
4298 					}
4299 
4300 					for (i = 0; i < pages; i++) {
4301 						page_unlock(ppa[i]);
4302 					}
4303 					if (amp != NULL) {
4304 						anon_array_exit(&an_cookie);
4305 						ANON_LOCK_EXIT(&amp->a_rwlock);
4306 					}
4307 					ierr = -1;
4308 					break;
4309 				}
4310 				if (szc != 0 && !xhat && !upgrdfail) {
4311 					segvn_faultvnmpss_align_err5++;
4312 				}
4313 				SEGVN_VMSTAT_FLTVNPAGES(34);
4314 				if (pplist != NULL) {
4315 					page_free_replacement_page(pplist);
4316 					page_create_putback(pages);
4317 				}
4318 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4319 				    prot, vpprot);
4320 				if (upgrdfail && segvn_anypgsz_vnode) {
4321 					/* SOFTLOCK case */
4322 					hat_memload_array_region(hat, a, pgsz,
4323 					    ppa, prot & vpprot, hat_flag,
4324 					    svd->rcookie);
4325 				} else {
4326 					for (i = 0; i < pages; i++) {
4327 						hat_memload_region(hat,
4328 						    a + (i << PAGESHIFT),
4329 						    ppa[i], prot & vpprot,
4330 						    hat_flag, svd->rcookie);
4331 					}
4332 				}
4333 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4334 					for (i = 0; i < pages; i++) {
4335 						page_unlock(ppa[i]);
4336 					}
4337 				}
4338 				if (amp != NULL) {
4339 					anon_array_exit(&an_cookie);
4340 					ANON_LOCK_EXIT(&amp->a_rwlock);
4341 				}
4342 				goto next;
4343 			}
4344 
4345 			if (pszc == szc) {
4346 				/*
4347 				 * segvn_full_szcpages() upgraded pages szc.
4348 				 */
4349 				ASSERT(pszc == ppa[0]->p_szc);
4350 				ASSERT(IS_P2ALIGNED(pfn, pages));
4351 				goto chkszc;
4352 			}
4353 
4354 			if (pszc > szc) {
4355 				kmutex_t *szcmtx;
4356 				SEGVN_VMSTAT_FLTVNPAGES(35);
4357 				/*
4358 				 * p_szc of ppa[0] can change since we haven't
4359 				 * locked all constituent pages. Call
4360 				 * page_lock_szc() to prevent szc changes.
4361 				 * This should be a rare case that happens when
4362 				 * multiple segments use a different page size
4363 				 * to map the same file offsets.
4364 				 */
4365 				szcmtx = page_szc_lock(ppa[0]);
4366 				pszc = ppa[0]->p_szc;
4367 				ASSERT(szcmtx != NULL || pszc == 0);
4368 				ASSERT(ppa[0]->p_szc <= pszc);
4369 				if (pszc <= szc) {
4370 					SEGVN_VMSTAT_FLTVNPAGES(36);
4371 					if (szcmtx != NULL) {
4372 						mutex_exit(szcmtx);
4373 					}
4374 					goto chkszc;
4375 				}
4376 				if (pplist != NULL) {
4377 					/*
4378 					 * page got promoted since last check.
4379 					 * we don't need preaalocated large
4380 					 * page.
4381 					 */
4382 					SEGVN_VMSTAT_FLTVNPAGES(37);
4383 					page_free_replacement_page(pplist);
4384 					page_create_putback(pages);
4385 				}
4386 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4387 				    prot, vpprot);
4388 				hat_memload_array_region(hat, a, pgsz, ppa,
4389 				    prot & vpprot, hat_flag, svd->rcookie);
4390 				mutex_exit(szcmtx);
4391 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4392 					for (i = 0; i < pages; i++) {
4393 						page_unlock(ppa[i]);
4394 					}
4395 				}
4396 				if (amp != NULL) {
4397 					anon_array_exit(&an_cookie);
4398 					ANON_LOCK_EXIT(&amp->a_rwlock);
4399 				}
4400 				goto next;
4401 			}
4402 
4403 			/*
4404 			 * if page got demoted since last check
4405 			 * we could have not allocated larger page.
4406 			 * allocate now.
4407 			 */
4408 			if (pplist == NULL &&
4409 			    page_alloc_pages(vp, seg, a, &pplist, NULL,
4410 			    szc, 0, 0) && type != F_SOFTLOCK) {
4411 				SEGVN_VMSTAT_FLTVNPAGES(38);
4412 				for (i = 0; i < pages; i++) {
4413 					page_unlock(ppa[i]);
4414 				}
4415 				if (amp != NULL) {
4416 					anon_array_exit(&an_cookie);
4417 					ANON_LOCK_EXIT(&amp->a_rwlock);
4418 				}
4419 				ierr = -1;
4420 				alloc_failed |= (1 << szc);
4421 				break;
4422 			}
4423 
4424 			SEGVN_VMSTAT_FLTVNPAGES(39);
4425 
4426 			if (pplist != NULL) {
4427 				segvn_relocate_pages(ppa, pplist);
4428 #ifdef DEBUG
4429 			} else {
4430 				ASSERT(type == F_SOFTLOCK);
4431 				SEGVN_VMSTAT_FLTVNPAGES(40);
4432 #endif /* DEBUG */
4433 			}
4434 
4435 			SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot);
4436 
4437 			if (pplist == NULL && segvn_anypgsz_vnode == 0) {
4438 				ASSERT(type == F_SOFTLOCK);
4439 				for (i = 0; i < pages; i++) {
4440 					ASSERT(ppa[i]->p_szc < szc);
4441 					hat_memload_region(hat,
4442 					    a + (i << PAGESHIFT),
4443 					    ppa[i], prot & vpprot, hat_flag,
4444 					    svd->rcookie);
4445 				}
4446 			} else {
4447 				ASSERT(pplist != NULL || type == F_SOFTLOCK);
4448 				hat_memload_array_region(hat, a, pgsz, ppa,
4449 				    prot & vpprot, hat_flag, svd->rcookie);
4450 			}
4451 			if (!(hat_flag & HAT_LOAD_LOCK)) {
4452 				for (i = 0; i < pages; i++) {
4453 					ASSERT(PAGE_SHARED(ppa[i]));
4454 					page_unlock(ppa[i]);
4455 				}
4456 			}
4457 			if (amp != NULL) {
4458 				anon_array_exit(&an_cookie);
4459 				ANON_LOCK_EXIT(&amp->a_rwlock);
4460 			}
4461 
4462 		next:
4463 			if (vpage != NULL) {
4464 				vpage += pages;
4465 			}
4466 			adjszc_chk = 1;
4467 		}
4468 		if (a == lpgeaddr)
4469 			break;
4470 		ASSERT(a < lpgeaddr);
4471 
4472 		ASSERT(!brkcow && !tron && type != F_SOFTLOCK);
4473 
4474 		/*
4475 		 * ierr == -1 means we failed to map with a large page.
4476 		 * (either due to allocation/relocation failures or
4477 		 * misalignment with other mappings to this file.
4478 		 *
4479 		 * ierr == -2 means some other thread allocated a large page
4480 		 * after we gave up tp map with a large page.  retry with
4481 		 * larger mapping.
4482 		 */
4483 		ASSERT(ierr == -1 || ierr == -2);
4484 		ASSERT(ierr == -2 || szc != 0);
4485 		ASSERT(ierr == -1 || szc < seg->s_szc);
4486 		if (ierr == -2) {
4487 			SEGVN_VMSTAT_FLTVNPAGES(41);
4488 			ASSERT(pszc > szc && pszc <= seg->s_szc);
4489 			szc = pszc;
4490 		} else if (segvn_anypgsz_vnode) {
4491 			SEGVN_VMSTAT_FLTVNPAGES(42);
4492 			szc--;
4493 		} else {
4494 			SEGVN_VMSTAT_FLTVNPAGES(43);
4495 			ASSERT(pszc < szc);
4496 			/*
4497 			 * other process created pszc large page.
4498 			 * but we still have to drop to 0 szc.
4499 			 */
4500 			szc = 0;
4501 		}
4502 
4503 		pgsz = page_get_pagesize(szc);
4504 		pages = btop(pgsz);
4505 		if (ierr == -2) {
4506 			/*
4507 			 * Size up case. Note lpgaddr may only be needed for
4508 			 * softlock case so we don't adjust it here.
4509 			 */
4510 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4511 			ASSERT(a >= lpgaddr);
4512 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4513 			off = svd->offset + (uintptr_t)(a - seg->s_base);
4514 			aindx = svd->anon_index + seg_page(seg, a);
4515 			vpage = (svd->vpage != NULL) ?
4516 			    &svd->vpage[seg_page(seg, a)] : NULL;
4517 		} else {
4518 			/*
4519 			 * Size down case. Note lpgaddr may only be needed for
4520 			 * softlock case so we don't adjust it here.
4521 			 */
4522 			ASSERT(IS_P2ALIGNED(a, pgsz));
4523 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4524 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4525 			ASSERT(a < lpgeaddr);
4526 			if (a < addr) {
4527 				SEGVN_VMSTAT_FLTVNPAGES(44);
4528 				/*
4529 				 * The beginning of the large page region can
4530 				 * be pulled to the right to make a smaller
4531 				 * region. We haven't yet faulted a single
4532 				 * page.
4533 				 */
4534 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4535 				ASSERT(a >= lpgaddr);
4536 				off = svd->offset +
4537 				    (uintptr_t)(a - seg->s_base);
4538 				aindx = svd->anon_index + seg_page(seg, a);
4539 				vpage = (svd->vpage != NULL) ?
4540 				    &svd->vpage[seg_page(seg, a)] : NULL;
4541 			}
4542 		}
4543 	}
4544 out:
4545 	kmem_free(ppa, ppasize);
4546 	if (!err && !vop_size_err) {
4547 		SEGVN_VMSTAT_FLTVNPAGES(45);
4548 		return (0);
4549 	}
4550 	if (type == F_SOFTLOCK && a > lpgaddr) {
4551 		SEGVN_VMSTAT_FLTVNPAGES(46);
4552 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4553 	}
4554 	if (!vop_size_err) {
4555 		SEGVN_VMSTAT_FLTVNPAGES(47);
4556 		return (err);
4557 	}
4558 	ASSERT(brkcow || tron || type == F_SOFTLOCK);
4559 	/*
4560 	 * Large page end is mapped beyond the end of file and it's a cow
4561 	 * fault (can be a text replication induced cow) or softlock so we can't
4562 	 * reduce the map area.  For now just demote the segment. This should
4563 	 * really only happen if the end of the file changed after the mapping
4564 	 * was established since when large page segments are created we make
4565 	 * sure they don't extend beyond the end of the file.
4566 	 */
4567 	SEGVN_VMSTAT_FLTVNPAGES(48);
4568 
4569 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4570 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4571 	err = 0;
4572 	if (seg->s_szc != 0) {
4573 		segvn_fltvnpages_clrszc_cnt++;
4574 		ASSERT(svd->softlockcnt == 0);
4575 		err = segvn_clrszc(seg);
4576 		if (err != 0) {
4577 			segvn_fltvnpages_clrszc_err++;
4578 		}
4579 	}
4580 	ASSERT(err || seg->s_szc == 0);
4581 	SEGVN_LOCK_DOWNGRADE(seg->s_as, &svd->lock);
4582 	/* segvn_fault will do its job as if szc had been zero to begin with */
4583 	return (err == 0 ? IE_RETRY : FC_MAKE_ERR(err));
4584 }
4585 
4586 /*
4587  * This routine will attempt to fault in one large page.
4588  * it will use smaller pages if that fails.
4589  * It should only be called for pure anonymous segments.
4590  */
4591 static faultcode_t
4592 segvn_fault_anonpages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
4593     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
4594     caddr_t eaddr, int brkcow)
4595 {
4596 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4597 	struct anon_map *amp = svd->amp;
4598 	uchar_t segtype = svd->type;
4599 	uint_t szc = seg->s_szc;
4600 	size_t pgsz = page_get_pagesize(szc);
4601 	size_t maxpgsz = pgsz;
4602 	pgcnt_t pages = btop(pgsz);
4603 	uint_t ppaszc = szc;
4604 	caddr_t a = lpgaddr;
4605 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
4606 	struct vpage *vpage = (svd->vpage != NULL) ?
4607 	    &svd->vpage[seg_page(seg, a)] : NULL;
4608 	page_t **ppa;
4609 	uint_t	ppa_szc;
4610 	faultcode_t err;
4611 	int ierr;
4612 	uint_t protchk, prot, vpprot;
4613 	ulong_t i;
4614 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
4615 	anon_sync_obj_t cookie;
4616 	int adjszc_chk;
4617 	int pgflags = (svd->tr_state == SEGVN_TR_ON) ? PG_LOCAL : 0;
4618 
4619 	ASSERT(szc != 0);
4620 	ASSERT(amp != NULL);
4621 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
4622 	ASSERT(!(svd->flags & MAP_NORESERVE));
4623 	ASSERT(type != F_SOFTUNLOCK);
4624 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
4625 	ASSERT(!brkcow || svd->tr_state == SEGVN_TR_OFF);
4626 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
4627 
4628 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
4629 
4630 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltanpages[0]);
4631 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltanpages[1]);
4632 
4633 	if (svd->flags & MAP_TEXT) {
4634 		hat_flag |= HAT_LOAD_TEXT;
4635 	}
4636 
4637 	if (svd->pageprot) {
4638 		switch (rw) {
4639 		case S_READ:
4640 			protchk = PROT_READ;
4641 			break;
4642 		case S_WRITE:
4643 			protchk = PROT_WRITE;
4644 			break;
4645 		case S_EXEC:
4646 			protchk = PROT_EXEC;
4647 			break;
4648 		case S_OTHER:
4649 		default:
4650 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4651 			break;
4652 		}
4653 		VM_STAT_ADD(segvnvmstats.fltanpages[2]);
4654 	} else {
4655 		prot = svd->prot;
4656 		/* caller has already done segment level protection check. */
4657 	}
4658 
4659 	ppa = kmem_cache_alloc(segvn_szc_cache[ppaszc], KM_SLEEP);
4660 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4661 	for (;;) {
4662 		adjszc_chk = 0;
4663 		for (; a < lpgeaddr; a += pgsz, aindx += pages) {
4664 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
4665 				VM_STAT_ADD(segvnvmstats.fltanpages[3]);
4666 				ASSERT(vpage != NULL);
4667 				prot = VPP_PROT(vpage);
4668 				ASSERT(sameprot(seg, a, maxpgsz));
4669 				if ((prot & protchk) == 0) {
4670 					err = FC_PROT;
4671 					goto error;
4672 				}
4673 			}
4674 			if (adjszc_chk && IS_P2ALIGNED(a, maxpgsz) &&
4675 			    pgsz < maxpgsz) {
4676 				ASSERT(a > lpgaddr);
4677 				szc = seg->s_szc;
4678 				pgsz = maxpgsz;
4679 				pages = btop(pgsz);
4680 				ASSERT(IS_P2ALIGNED(aindx, pages));
4681 				lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr,
4682 				    pgsz);
4683 			}
4684 			if (type == F_SOFTLOCK) {
4685 				atomic_add_long((ulong_t *)&svd->softlockcnt,
4686 				    pages);
4687 			}
4688 			anon_array_enter(amp, aindx, &cookie);
4689 			ppa_szc = (uint_t)-1;
4690 			ierr = anon_map_getpages(amp, aindx, szc, seg, a,
4691 			    prot, &vpprot, ppa, &ppa_szc, vpage, rw, brkcow,
4692 			    segvn_anypgsz, pgflags, svd->cred);
4693 			if (ierr != 0) {
4694 				anon_array_exit(&cookie);
4695 				VM_STAT_ADD(segvnvmstats.fltanpages[4]);
4696 				if (type == F_SOFTLOCK) {
4697 					atomic_add_long(
4698 					    (ulong_t *)&svd->softlockcnt,
4699 					    -pages);
4700 				}
4701 				if (ierr > 0) {
4702 					VM_STAT_ADD(segvnvmstats.fltanpages[6]);
4703 					err = FC_MAKE_ERR(ierr);
4704 					goto error;
4705 				}
4706 				break;
4707 			}
4708 
4709 			ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4710 
4711 			ASSERT(segtype == MAP_SHARED ||
4712 			    ppa[0]->p_szc <= szc);
4713 			ASSERT(segtype == MAP_PRIVATE ||
4714 			    ppa[0]->p_szc >= szc);
4715 
4716 			/*
4717 			 * Handle pages that have been marked for migration
4718 			 */
4719 			if (lgrp_optimizations())
4720 				page_migrate(seg, a, ppa, pages);
4721 
4722 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
4723 
4724 			if (segtype == MAP_SHARED) {
4725 				vpprot |= PROT_WRITE;
4726 			}
4727 
4728 			hat_memload_array(hat, a, pgsz, ppa,
4729 			    prot & vpprot, hat_flag);
4730 
4731 			if (hat_flag & HAT_LOAD_LOCK) {
4732 				VM_STAT_ADD(segvnvmstats.fltanpages[7]);
4733 			} else {
4734 				VM_STAT_ADD(segvnvmstats.fltanpages[8]);
4735 				for (i = 0; i < pages; i++)
4736 					page_unlock(ppa[i]);
4737 			}
4738 			if (vpage != NULL)
4739 				vpage += pages;
4740 
4741 			anon_array_exit(&cookie);
4742 			adjszc_chk = 1;
4743 		}
4744 		if (a == lpgeaddr)
4745 			break;
4746 		ASSERT(a < lpgeaddr);
4747 		/*
4748 		 * ierr == -1 means we failed to allocate a large page.
4749 		 * so do a size down operation.
4750 		 *
4751 		 * ierr == -2 means some other process that privately shares
4752 		 * pages with this process has allocated a larger page and we
4753 		 * need to retry with larger pages. So do a size up
4754 		 * operation. This relies on the fact that large pages are
4755 		 * never partially shared i.e. if we share any constituent
4756 		 * page of a large page with another process we must share the
4757 		 * entire large page. Note this cannot happen for SOFTLOCK
4758 		 * case, unless current address (a) is at the beginning of the
4759 		 * next page size boundary because the other process couldn't
4760 		 * have relocated locked pages.
4761 		 */
4762 		ASSERT(ierr == -1 || ierr == -2);
4763 
4764 		if (segvn_anypgsz) {
4765 			ASSERT(ierr == -2 || szc != 0);
4766 			ASSERT(ierr == -1 || szc < seg->s_szc);
4767 			szc = (ierr == -1) ? szc - 1 : szc + 1;
4768 		} else {
4769 			/*
4770 			 * For non COW faults and segvn_anypgsz == 0
4771 			 * we need to be careful not to loop forever
4772 			 * if existing page is found with szc other
4773 			 * than 0 or seg->s_szc. This could be due
4774 			 * to page relocations on behalf of DR or
4775 			 * more likely large page creation. For this
4776 			 * case simply re-size to existing page's szc
4777 			 * if returned by anon_map_getpages().
4778 			 */
4779 			if (ppa_szc == (uint_t)-1) {
4780 				szc = (ierr == -1) ? 0 : seg->s_szc;
4781 			} else {
4782 				ASSERT(ppa_szc <= seg->s_szc);
4783 				ASSERT(ierr == -2 || ppa_szc < szc);
4784 				ASSERT(ierr == -1 || ppa_szc > szc);
4785 				szc = ppa_szc;
4786 			}
4787 		}
4788 
4789 		pgsz = page_get_pagesize(szc);
4790 		pages = btop(pgsz);
4791 		ASSERT(type != F_SOFTLOCK || ierr == -1 ||
4792 		    (IS_P2ALIGNED(a, pgsz) && IS_P2ALIGNED(lpgeaddr, pgsz)));
4793 		if (type == F_SOFTLOCK) {
4794 			/*
4795 			 * For softlocks we cannot reduce the fault area
4796 			 * (calculated based on the largest page size for this
4797 			 * segment) for size down and a is already next
4798 			 * page size aligned as assertted above for size
4799 			 * ups. Therefore just continue in case of softlock.
4800 			 */
4801 			VM_STAT_ADD(segvnvmstats.fltanpages[9]);
4802 			continue; /* keep lint happy */
4803 		} else if (ierr == -2) {
4804 
4805 			/*
4806 			 * Size up case. Note lpgaddr may only be needed for
4807 			 * softlock case so we don't adjust it here.
4808 			 */
4809 			VM_STAT_ADD(segvnvmstats.fltanpages[10]);
4810 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4811 			ASSERT(a >= lpgaddr);
4812 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4813 			aindx = svd->anon_index + seg_page(seg, a);
4814 			vpage = (svd->vpage != NULL) ?
4815 			    &svd->vpage[seg_page(seg, a)] : NULL;
4816 		} else {
4817 			/*
4818 			 * Size down 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[11]);
4822 			ASSERT(IS_P2ALIGNED(a, pgsz));
4823 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4824 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4825 			ASSERT(a < lpgeaddr);
4826 			if (a < addr) {
4827 				/*
4828 				 * The beginning of the large page region can
4829 				 * be pulled to the right to make a smaller
4830 				 * region. We haven't yet faulted a single
4831 				 * page.
4832 				 */
4833 				VM_STAT_ADD(segvnvmstats.fltanpages[12]);
4834 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4835 				ASSERT(a >= lpgaddr);
4836 				aindx = svd->anon_index + seg_page(seg, a);
4837 				vpage = (svd->vpage != NULL) ?
4838 				    &svd->vpage[seg_page(seg, a)] : NULL;
4839 			}
4840 		}
4841 	}
4842 	VM_STAT_ADD(segvnvmstats.fltanpages[13]);
4843 	ANON_LOCK_EXIT(&amp->a_rwlock);
4844 	kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4845 	return (0);
4846 error:
4847 	VM_STAT_ADD(segvnvmstats.fltanpages[14]);
4848 	ANON_LOCK_EXIT(&amp->a_rwlock);
4849 	kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4850 	if (type == F_SOFTLOCK && a > lpgaddr) {
4851 		VM_STAT_ADD(segvnvmstats.fltanpages[15]);
4852 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4853 	}
4854 	return (err);
4855 }
4856 
4857 int fltadvice = 1;	/* set to free behind pages for sequential access */
4858 
4859 /*
4860  * This routine is called via a machine specific fault handling routine.
4861  * It is also called by software routines wishing to lock or unlock
4862  * a range of addresses.
4863  *
4864  * Here is the basic algorithm:
4865  *	If unlocking
4866  *		Call segvn_softunlock
4867  *		Return
4868  *	endif
4869  *	Checking and set up work
4870  *	If we will need some non-anonymous pages
4871  *		Call VOP_GETPAGE over the range of non-anonymous pages
4872  *	endif
4873  *	Loop over all addresses requested
4874  *		Call segvn_faultpage passing in page list
4875  *		    to load up translations and handle anonymous pages
4876  *	endloop
4877  *	Load up translation to any additional pages in page list not
4878  *	    already handled that fit into this segment
4879  */
4880 static faultcode_t
4881 segvn_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
4882     enum fault_type type, enum seg_rw rw)
4883 {
4884 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4885 	page_t **plp, **ppp, *pp;
4886 	u_offset_t off;
4887 	caddr_t a;
4888 	struct vpage *vpage;
4889 	uint_t vpprot, prot;
4890 	int err;
4891 	page_t *pl[PVN_GETPAGE_NUM + 1];
4892 	size_t plsz, pl_alloc_sz;
4893 	size_t page;
4894 	ulong_t anon_index;
4895 	struct anon_map *amp;
4896 	int dogetpage = 0;
4897 	caddr_t	lpgaddr, lpgeaddr;
4898 	size_t pgsz;
4899 	anon_sync_obj_t cookie;
4900 	int brkcow = BREAK_COW_SHARE(rw, type, svd->type);
4901 
4902 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
4903 	ASSERT(svd->amp == NULL || svd->rcookie == HAT_INVALID_REGION_COOKIE);
4904 
4905 	/*
4906 	 * First handle the easy stuff
4907 	 */
4908 	if (type == F_SOFTUNLOCK) {
4909 		if (rw == S_READ_NOCOW) {
4910 			rw = S_READ;
4911 			ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
4912 		}
4913 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4914 		pgsz = (seg->s_szc == 0) ? PAGESIZE :
4915 		    page_get_pagesize(seg->s_szc);
4916 		VM_STAT_COND_ADD(pgsz > PAGESIZE, segvnvmstats.fltanpages[16]);
4917 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
4918 		segvn_softunlock(seg, lpgaddr, lpgeaddr - lpgaddr, rw);
4919 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4920 		return (0);
4921 	}
4922 
4923 	ASSERT(svd->tr_state == SEGVN_TR_OFF ||
4924 	    !HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
4925 	if (brkcow == 0) {
4926 		if (svd->tr_state == SEGVN_TR_INIT) {
4927 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4928 			if (svd->tr_state == SEGVN_TR_INIT) {
4929 				ASSERT(svd->vp != NULL && svd->amp == NULL);
4930 				ASSERT(svd->flags & MAP_TEXT);
4931 				ASSERT(svd->type == MAP_PRIVATE);
4932 				segvn_textrepl(seg);
4933 				ASSERT(svd->tr_state != SEGVN_TR_INIT);
4934 				ASSERT(svd->tr_state != SEGVN_TR_ON ||
4935 				    svd->amp != NULL);
4936 			}
4937 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4938 		}
4939 	} else if (svd->tr_state != SEGVN_TR_OFF) {
4940 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4941 
4942 		if (rw == S_WRITE && svd->tr_state != SEGVN_TR_OFF) {
4943 			ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
4944 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4945 			return (FC_PROT);
4946 		}
4947 
4948 		if (svd->tr_state == SEGVN_TR_ON) {
4949 			ASSERT(svd->vp != NULL && svd->amp != NULL);
4950 			segvn_textunrepl(seg, 0);
4951 			ASSERT(svd->amp == NULL &&
4952 			    svd->tr_state == SEGVN_TR_OFF);
4953 		} else if (svd->tr_state != SEGVN_TR_OFF) {
4954 			svd->tr_state = SEGVN_TR_OFF;
4955 		}
4956 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
4957 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4958 	}
4959 
4960 top:
4961 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4962 
4963 	/*
4964 	 * If we have the same protections for the entire segment,
4965 	 * insure that the access being attempted is legitimate.
4966 	 */
4967 
4968 	if (svd->pageprot == 0) {
4969 		uint_t protchk;
4970 
4971 		switch (rw) {
4972 		case S_READ:
4973 		case S_READ_NOCOW:
4974 			protchk = PROT_READ;
4975 			break;
4976 		case S_WRITE:
4977 			protchk = PROT_WRITE;
4978 			break;
4979 		case S_EXEC:
4980 			protchk = PROT_EXEC;
4981 			break;
4982 		case S_OTHER:
4983 		default:
4984 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4985 			break;
4986 		}
4987 
4988 		if ((svd->prot & protchk) == 0) {
4989 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4990 			return (FC_PROT);	/* illegal access type */
4991 		}
4992 	}
4993 
4994 	if (brkcow && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
4995 		/* this must be SOFTLOCK S_READ fault */
4996 		ASSERT(svd->amp == NULL);
4997 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
4998 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4999 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5000 		if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5001 			/*
5002 			 * this must be the first ever non S_READ_NOCOW
5003 			 * softlock for this segment.
5004 			 */
5005 			ASSERT(svd->softlockcnt == 0);
5006 			hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5007 			    HAT_REGION_TEXT);
5008 			svd->rcookie = HAT_INVALID_REGION_COOKIE;
5009 		}
5010 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5011 		goto top;
5012 	}
5013 
5014 	/*
5015 	 * We can't allow the long term use of softlocks for vmpss segments,
5016 	 * because in some file truncation cases we should be able to demote
5017 	 * the segment, which requires that there are no softlocks.  The
5018 	 * only case where it's ok to allow a SOFTLOCK fault against a vmpss
5019 	 * segment is S_READ_NOCOW, where the caller holds the address space
5020 	 * locked as writer and calls softunlock before dropping the as lock.
5021 	 * S_READ_NOCOW is used by /proc to read memory from another user.
5022 	 *
5023 	 * Another deadlock between SOFTLOCK and file truncation can happen
5024 	 * because segvn_fault_vnodepages() calls the FS one pagesize at
5025 	 * a time. A second VOP_GETPAGE() call by segvn_fault_vnodepages()
5026 	 * can cause a deadlock because the first set of page_t's remain
5027 	 * locked SE_SHARED.  To avoid this, we demote segments on a first
5028 	 * SOFTLOCK if they have a length greater than the segment's
5029 	 * page size.
5030 	 *
5031 	 * So for now, we only avoid demoting a segment on a SOFTLOCK when
5032 	 * the access type is S_READ_NOCOW and the fault length is less than
5033 	 * or equal to the segment's page size. While this is quite restrictive,
5034 	 * it should be the most common case of SOFTLOCK against a vmpss
5035 	 * segment.
5036 	 *
5037 	 * For S_READ_NOCOW, it's safe not to do a copy on write because the
5038 	 * caller makes sure no COW will be caused by another thread for a
5039 	 * softlocked page.
5040 	 */
5041 	if (type == F_SOFTLOCK && svd->vp != NULL && seg->s_szc != 0) {
5042 		int demote = 0;
5043 
5044 		if (rw != S_READ_NOCOW) {
5045 			demote = 1;
5046 		}
5047 		if (!demote && len > PAGESIZE) {
5048 			pgsz = page_get_pagesize(seg->s_szc);
5049 			CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr,
5050 			    lpgeaddr);
5051 			if (lpgeaddr - lpgaddr > pgsz) {
5052 				demote = 1;
5053 			}
5054 		}
5055 
5056 		ASSERT(demote || AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5057 
5058 		if (demote) {
5059 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5060 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5061 			if (seg->s_szc != 0) {
5062 				segvn_vmpss_clrszc_cnt++;
5063 				ASSERT(svd->softlockcnt == 0);
5064 				err = segvn_clrszc(seg);
5065 				if (err) {
5066 					segvn_vmpss_clrszc_err++;
5067 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5068 					return (FC_MAKE_ERR(err));
5069 				}
5070 			}
5071 			ASSERT(seg->s_szc == 0);
5072 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5073 			goto top;
5074 		}
5075 	}
5076 
5077 	/*
5078 	 * Check to see if we need to allocate an anon_map structure.
5079 	 */
5080 	if (svd->amp == NULL && (svd->vp == NULL || brkcow)) {
5081 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5082 		/*
5083 		 * Drop the "read" lock on the segment and acquire
5084 		 * the "write" version since we have to allocate the
5085 		 * anon_map.
5086 		 */
5087 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5088 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5089 
5090 		if (svd->amp == NULL) {
5091 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
5092 			svd->amp->a_szc = seg->s_szc;
5093 		}
5094 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5095 
5096 		/*
5097 		 * Start all over again since segment protections
5098 		 * may have changed after we dropped the "read" lock.
5099 		 */
5100 		goto top;
5101 	}
5102 
5103 	/*
5104 	 * S_READ_NOCOW vs S_READ distinction was
5105 	 * only needed for the code above. After
5106 	 * that we treat it as S_READ.
5107 	 */
5108 	if (rw == S_READ_NOCOW) {
5109 		ASSERT(type == F_SOFTLOCK);
5110 		ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5111 		rw = S_READ;
5112 	}
5113 
5114 	amp = svd->amp;
5115 
5116 	/*
5117 	 * MADV_SEQUENTIAL work is ignored for large page segments.
5118 	 */
5119 	if (seg->s_szc != 0) {
5120 		pgsz = page_get_pagesize(seg->s_szc);
5121 		ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
5122 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
5123 		if (svd->vp == NULL) {
5124 			err = segvn_fault_anonpages(hat, seg, lpgaddr,
5125 			    lpgeaddr, type, rw, addr, addr + len, brkcow);
5126 		} else {
5127 			err = segvn_fault_vnodepages(hat, seg, lpgaddr,
5128 			    lpgeaddr, type, rw, addr, addr + len, brkcow);
5129 			if (err == IE_RETRY) {
5130 				ASSERT(seg->s_szc == 0);
5131 				ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
5132 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5133 				goto top;
5134 			}
5135 		}
5136 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5137 		return (err);
5138 	}
5139 
5140 	page = seg_page(seg, addr);
5141 	if (amp != NULL) {
5142 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5143 		anon_index = svd->anon_index + page;
5144 
5145 		if (type == F_PROT && rw == S_READ &&
5146 		    svd->tr_state == SEGVN_TR_OFF &&
5147 		    svd->type == MAP_PRIVATE && svd->pageprot == 0) {
5148 			size_t index = anon_index;
5149 			struct anon *ap;
5150 
5151 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5152 			/*
5153 			 * The fast path could apply to S_WRITE also, except
5154 			 * that the protection fault could be caused by lazy
5155 			 * tlb flush when ro->rw. In this case, the pte is
5156 			 * RW already. But RO in the other cpu's tlb causes
5157 			 * the fault. Since hat_chgprot won't do anything if
5158 			 * pte doesn't change, we may end up faulting
5159 			 * indefinitely until the RO tlb entry gets replaced.
5160 			 */
5161 			for (a = addr; a < addr + len; a += PAGESIZE, index++) {
5162 				anon_array_enter(amp, index, &cookie);
5163 				ap = anon_get_ptr(amp->ahp, index);
5164 				anon_array_exit(&cookie);
5165 				if ((ap == NULL) || (ap->an_refcnt != 1)) {
5166 					ANON_LOCK_EXIT(&amp->a_rwlock);
5167 					goto slow;
5168 				}
5169 			}
5170 			hat_chgprot(seg->s_as->a_hat, addr, len, svd->prot);
5171 			ANON_LOCK_EXIT(&amp->a_rwlock);
5172 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5173 			return (0);
5174 		}
5175 	}
5176 slow:
5177 
5178 	if (svd->vpage == NULL)
5179 		vpage = NULL;
5180 	else
5181 		vpage = &svd->vpage[page];
5182 
5183 	off = svd->offset + (uintptr_t)(addr - seg->s_base);
5184 
5185 	/*
5186 	 * If MADV_SEQUENTIAL has been set for the particular page we
5187 	 * are faulting on, free behind all pages in the segment and put
5188 	 * them on the free list.
5189 	 */
5190 
5191 	if ((page != 0) && fltadvice && svd->tr_state != SEGVN_TR_ON) {
5192 		struct vpage *vpp;
5193 		ulong_t fanon_index;
5194 		size_t fpage;
5195 		u_offset_t pgoff, fpgoff;
5196 		struct vnode *fvp;
5197 		struct anon *fap = NULL;
5198 
5199 		if (svd->advice == MADV_SEQUENTIAL ||
5200 		    (svd->pageadvice &&
5201 		    VPP_ADVICE(vpage) == MADV_SEQUENTIAL)) {
5202 			pgoff = off - PAGESIZE;
5203 			fpage = page - 1;
5204 			if (vpage != NULL)
5205 				vpp = &svd->vpage[fpage];
5206 			if (amp != NULL)
5207 				fanon_index = svd->anon_index + fpage;
5208 
5209 			while (pgoff > svd->offset) {
5210 				if (svd->advice != MADV_SEQUENTIAL &&
5211 				    (!svd->pageadvice || (vpage &&
5212 				    VPP_ADVICE(vpp) != MADV_SEQUENTIAL)))
5213 					break;
5214 
5215 				/*
5216 				 * If this is an anon page, we must find the
5217 				 * correct <vp, offset> for it
5218 				 */
5219 				fap = NULL;
5220 				if (amp != NULL) {
5221 					ANON_LOCK_ENTER(&amp->a_rwlock,
5222 					    RW_READER);
5223 					anon_array_enter(amp, fanon_index,
5224 					    &cookie);
5225 					fap = anon_get_ptr(amp->ahp,
5226 					    fanon_index);
5227 					if (fap != NULL) {
5228 						swap_xlate(fap, &fvp, &fpgoff);
5229 					} else {
5230 						fpgoff = pgoff;
5231 						fvp = svd->vp;
5232 					}
5233 					anon_array_exit(&cookie);
5234 					ANON_LOCK_EXIT(&amp->a_rwlock);
5235 				} else {
5236 					fpgoff = pgoff;
5237 					fvp = svd->vp;
5238 				}
5239 				if (fvp == NULL)
5240 					break;	/* XXX */
5241 				/*
5242 				 * Skip pages that are free or have an
5243 				 * "exclusive" lock.
5244 				 */
5245 				pp = page_lookup_nowait(fvp, fpgoff, SE_SHARED);
5246 				if (pp == NULL)
5247 					break;
5248 				/*
5249 				 * We don't need the page_struct_lock to test
5250 				 * as this is only advisory; even if we
5251 				 * acquire it someone might race in and lock
5252 				 * the page after we unlock and before the
5253 				 * PUTPAGE, then VOP_PUTPAGE will do nothing.
5254 				 */
5255 				if (pp->p_lckcnt == 0 && pp->p_cowcnt == 0) {
5256 					/*
5257 					 * Hold the vnode before releasing
5258 					 * the page lock to prevent it from
5259 					 * being freed and re-used by some
5260 					 * other thread.
5261 					 */
5262 					VN_HOLD(fvp);
5263 					page_unlock(pp);
5264 					/*
5265 					 * We should build a page list
5266 					 * to kluster putpages XXX
5267 					 */
5268 					(void) VOP_PUTPAGE(fvp,
5269 					    (offset_t)fpgoff, PAGESIZE,
5270 					    (B_DONTNEED|B_FREE|B_ASYNC),
5271 					    svd->cred, NULL);
5272 					VN_RELE(fvp);
5273 				} else {
5274 					/*
5275 					 * XXX - Should the loop terminate if
5276 					 * the page is `locked'?
5277 					 */
5278 					page_unlock(pp);
5279 				}
5280 				--vpp;
5281 				--fanon_index;
5282 				pgoff -= PAGESIZE;
5283 			}
5284 		}
5285 	}
5286 
5287 	plp = pl;
5288 	*plp = NULL;
5289 	pl_alloc_sz = 0;
5290 
5291 	/*
5292 	 * See if we need to call VOP_GETPAGE for
5293 	 * *any* of the range being faulted on.
5294 	 * We can skip all of this work if there
5295 	 * was no original vnode.
5296 	 */
5297 	if (svd->vp != NULL) {
5298 		u_offset_t vp_off;
5299 		size_t vp_len;
5300 		struct anon *ap;
5301 		vnode_t *vp;
5302 
5303 		vp_off = off;
5304 		vp_len = len;
5305 
5306 		if (amp == NULL)
5307 			dogetpage = 1;
5308 		else {
5309 			/*
5310 			 * Only acquire reader lock to prevent amp->ahp
5311 			 * from being changed.  It's ok to miss pages,
5312 			 * hence we don't do anon_array_enter
5313 			 */
5314 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5315 			ap = anon_get_ptr(amp->ahp, anon_index);
5316 
5317 			if (len <= PAGESIZE)
5318 				/* inline non_anon() */
5319 				dogetpage = (ap == NULL);
5320 			else
5321 				dogetpage = non_anon(amp->ahp, anon_index,
5322 				    &vp_off, &vp_len);
5323 			ANON_LOCK_EXIT(&amp->a_rwlock);
5324 		}
5325 
5326 		if (dogetpage) {
5327 			enum seg_rw arw;
5328 			struct as *as = seg->s_as;
5329 
5330 			if (len > ptob((sizeof (pl) / sizeof (pl[0])) - 1)) {
5331 				/*
5332 				 * Page list won't fit in local array,
5333 				 * allocate one of the needed size.
5334 				 */
5335 				pl_alloc_sz =
5336 				    (btop(len) + 1) * sizeof (page_t *);
5337 				plp = kmem_alloc(pl_alloc_sz, KM_SLEEP);
5338 				plp[0] = NULL;
5339 				plsz = len;
5340 			} else if (rw == S_WRITE && svd->type == MAP_PRIVATE ||
5341 			    svd->tr_state == SEGVN_TR_ON || rw == S_OTHER ||
5342 			    (((size_t)(addr + PAGESIZE) <
5343 			    (size_t)(seg->s_base + seg->s_size)) &&
5344 			    hat_probe(as->a_hat, addr + PAGESIZE))) {
5345 				/*
5346 				 * Ask VOP_GETPAGE to return the exact number
5347 				 * of pages if
5348 				 * (a) this is a COW fault, or
5349 				 * (b) this is a software fault, or
5350 				 * (c) next page is already mapped.
5351 				 */
5352 				plsz = len;
5353 			} else {
5354 				/*
5355 				 * Ask VOP_GETPAGE to return adjacent pages
5356 				 * within the segment.
5357 				 */
5358 				plsz = MIN((size_t)PVN_GETPAGE_SZ, (size_t)
5359 				    ((seg->s_base + seg->s_size) - addr));
5360 				ASSERT((addr + plsz) <=
5361 				    (seg->s_base + seg->s_size));
5362 			}
5363 
5364 			/*
5365 			 * Need to get some non-anonymous pages.
5366 			 * We need to make only one call to GETPAGE to do
5367 			 * this to prevent certain deadlocking conditions
5368 			 * when we are doing locking.  In this case
5369 			 * non_anon() should have picked up the smallest
5370 			 * range which includes all the non-anonymous
5371 			 * pages in the requested range.  We have to
5372 			 * be careful regarding which rw flag to pass in
5373 			 * because on a private mapping, the underlying
5374 			 * object is never allowed to be written.
5375 			 */
5376 			if (rw == S_WRITE && svd->type == MAP_PRIVATE) {
5377 				arw = S_READ;
5378 			} else {
5379 				arw = rw;
5380 			}
5381 			vp = svd->vp;
5382 			TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5383 			    "segvn_getpage:seg %p addr %p vp %p",
5384 			    seg, addr, vp);
5385 			err = VOP_GETPAGE(vp, (offset_t)vp_off, vp_len,
5386 			    &vpprot, plp, plsz, seg, addr + (vp_off - off), arw,
5387 			    svd->cred, NULL);
5388 			if (err) {
5389 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5390 				segvn_pagelist_rele(plp);
5391 				if (pl_alloc_sz)
5392 					kmem_free(plp, pl_alloc_sz);
5393 				return (FC_MAKE_ERR(err));
5394 			}
5395 			if (svd->type == MAP_PRIVATE)
5396 				vpprot &= ~PROT_WRITE;
5397 		}
5398 	}
5399 
5400 	/*
5401 	 * N.B. at this time the plp array has all the needed non-anon
5402 	 * pages in addition to (possibly) having some adjacent pages.
5403 	 */
5404 
5405 	/*
5406 	 * Always acquire the anon_array_lock to prevent
5407 	 * 2 threads from allocating separate anon slots for
5408 	 * the same "addr".
5409 	 *
5410 	 * If this is a copy-on-write fault and we don't already
5411 	 * have the anon_array_lock, acquire it to prevent the
5412 	 * fault routine from handling multiple copy-on-write faults
5413 	 * on the same "addr" in the same address space.
5414 	 *
5415 	 * Only one thread should deal with the fault since after
5416 	 * it is handled, the other threads can acquire a translation
5417 	 * to the newly created private page.  This prevents two or
5418 	 * more threads from creating different private pages for the
5419 	 * same fault.
5420 	 *
5421 	 * We grab "serialization" lock here if this is a MAP_PRIVATE segment
5422 	 * to prevent deadlock between this thread and another thread
5423 	 * which has soft-locked this page and wants to acquire serial_lock.
5424 	 * ( bug 4026339 )
5425 	 *
5426 	 * The fix for bug 4026339 becomes unnecessary when using the
5427 	 * locking scheme with per amp rwlock and a global set of hash
5428 	 * lock, anon_array_lock.  If we steal a vnode page when low
5429 	 * on memory and upgrad the page lock through page_rename,
5430 	 * then the page is PAGE_HANDLED, nothing needs to be done
5431 	 * for this page after returning from segvn_faultpage.
5432 	 *
5433 	 * But really, the page lock should be downgraded after
5434 	 * the stolen page is page_rename'd.
5435 	 */
5436 
5437 	if (amp != NULL)
5438 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5439 
5440 	/*
5441 	 * Ok, now loop over the address range and handle faults
5442 	 */
5443 	for (a = addr; a < addr + len; a += PAGESIZE, off += PAGESIZE) {
5444 		err = segvn_faultpage(hat, seg, a, off, vpage, plp, vpprot,
5445 		    type, rw, brkcow);
5446 		if (err) {
5447 			if (amp != NULL)
5448 				ANON_LOCK_EXIT(&amp->a_rwlock);
5449 			if (type == F_SOFTLOCK && a > addr) {
5450 				segvn_softunlock(seg, addr, (a - addr),
5451 				    S_OTHER);
5452 			}
5453 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5454 			segvn_pagelist_rele(plp);
5455 			if (pl_alloc_sz)
5456 				kmem_free(plp, pl_alloc_sz);
5457 			return (err);
5458 		}
5459 		if (vpage) {
5460 			vpage++;
5461 		} else if (svd->vpage) {
5462 			page = seg_page(seg, addr);
5463 			vpage = &svd->vpage[++page];
5464 		}
5465 	}
5466 
5467 	/* Didn't get pages from the underlying fs so we're done */
5468 	if (!dogetpage)
5469 		goto done;
5470 
5471 	/*
5472 	 * Now handle any other pages in the list returned.
5473 	 * If the page can be used, load up the translations now.
5474 	 * Note that the for loop will only be entered if "plp"
5475 	 * is pointing to a non-NULL page pointer which means that
5476 	 * VOP_GETPAGE() was called and vpprot has been initialized.
5477 	 */
5478 	if (svd->pageprot == 0)
5479 		prot = svd->prot & vpprot;
5480 
5481 
5482 	/*
5483 	 * Large Files: diff should be unsigned value because we started
5484 	 * supporting > 2GB segment sizes from 2.5.1 and when a
5485 	 * large file of size > 2GB gets mapped to address space
5486 	 * the diff value can be > 2GB.
5487 	 */
5488 
5489 	for (ppp = plp; (pp = *ppp) != NULL; ppp++) {
5490 		size_t diff;
5491 		struct anon *ap;
5492 		int anon_index;
5493 		anon_sync_obj_t cookie;
5494 		int hat_flag = HAT_LOAD_ADV;
5495 
5496 		if (svd->flags & MAP_TEXT) {
5497 			hat_flag |= HAT_LOAD_TEXT;
5498 		}
5499 
5500 		if (pp == PAGE_HANDLED)
5501 			continue;
5502 
5503 		if (svd->tr_state != SEGVN_TR_ON &&
5504 		    pp->p_offset >=  svd->offset &&
5505 		    pp->p_offset < svd->offset + seg->s_size) {
5506 
5507 			diff = pp->p_offset - svd->offset;
5508 
5509 			/*
5510 			 * Large Files: Following is the assertion
5511 			 * validating the above cast.
5512 			 */
5513 			ASSERT(svd->vp == pp->p_vnode);
5514 
5515 			page = btop(diff);
5516 			if (svd->pageprot)
5517 				prot = VPP_PROT(&svd->vpage[page]) & vpprot;
5518 
5519 			/*
5520 			 * Prevent other threads in the address space from
5521 			 * creating private pages (i.e., allocating anon slots)
5522 			 * while we are in the process of loading translations
5523 			 * to additional pages returned by the underlying
5524 			 * object.
5525 			 */
5526 			if (amp != NULL) {
5527 				anon_index = svd->anon_index + page;
5528 				anon_array_enter(amp, anon_index, &cookie);
5529 				ap = anon_get_ptr(amp->ahp, anon_index);
5530 			}
5531 			if ((amp == NULL) || (ap == NULL)) {
5532 				if (IS_VMODSORT(pp->p_vnode) ||
5533 				    enable_mbit_wa) {
5534 					if (rw == S_WRITE)
5535 						hat_setmod(pp);
5536 					else if (rw != S_OTHER &&
5537 					    !hat_ismod(pp))
5538 						prot &= ~PROT_WRITE;
5539 				}
5540 				/*
5541 				 * Skip mapping read ahead pages marked
5542 				 * for migration, so they will get migrated
5543 				 * properly on fault
5544 				 */
5545 				ASSERT(amp == NULL ||
5546 				    svd->rcookie == HAT_INVALID_REGION_COOKIE);
5547 				if ((prot & PROT_READ) && !PP_ISMIGRATE(pp)) {
5548 					hat_memload_region(hat,
5549 					    seg->s_base + diff,
5550 					    pp, prot, hat_flag,
5551 					    svd->rcookie);
5552 				}
5553 			}
5554 			if (amp != NULL)
5555 				anon_array_exit(&cookie);
5556 		}
5557 		page_unlock(pp);
5558 	}
5559 done:
5560 	if (amp != NULL)
5561 		ANON_LOCK_EXIT(&amp->a_rwlock);
5562 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5563 	if (pl_alloc_sz)
5564 		kmem_free(plp, pl_alloc_sz);
5565 	return (0);
5566 }
5567 
5568 /*
5569  * This routine is used to start I/O on pages asynchronously.  XXX it will
5570  * only create PAGESIZE pages. At fault time they will be relocated into
5571  * larger pages.
5572  */
5573 static faultcode_t
5574 segvn_faulta(struct seg *seg, caddr_t addr)
5575 {
5576 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5577 	int err;
5578 	struct anon_map *amp;
5579 	vnode_t *vp;
5580 
5581 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5582 
5583 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5584 	if ((amp = svd->amp) != NULL) {
5585 		struct anon *ap;
5586 
5587 		/*
5588 		 * Reader lock to prevent amp->ahp from being changed.
5589 		 * This is advisory, it's ok to miss a page, so
5590 		 * we don't do anon_array_enter lock.
5591 		 */
5592 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5593 		if ((ap = anon_get_ptr(amp->ahp,
5594 		    svd->anon_index + seg_page(seg, addr))) != NULL) {
5595 
5596 			err = anon_getpage(&ap, NULL, NULL,
5597 			    0, seg, addr, S_READ, svd->cred);
5598 
5599 			ANON_LOCK_EXIT(&amp->a_rwlock);
5600 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5601 			if (err)
5602 				return (FC_MAKE_ERR(err));
5603 			return (0);
5604 		}
5605 		ANON_LOCK_EXIT(&amp->a_rwlock);
5606 	}
5607 
5608 	if (svd->vp == NULL) {
5609 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5610 		return (0);			/* zfod page - do nothing now */
5611 	}
5612 
5613 	vp = svd->vp;
5614 	TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5615 	    "segvn_getpage:seg %p addr %p vp %p", seg, addr, vp);
5616 	err = VOP_GETPAGE(vp,
5617 	    (offset_t)(svd->offset + (uintptr_t)(addr - seg->s_base)),
5618 	    PAGESIZE, NULL, NULL, 0, seg, addr,
5619 	    S_OTHER, svd->cred, NULL);
5620 
5621 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5622 	if (err)
5623 		return (FC_MAKE_ERR(err));
5624 	return (0);
5625 }
5626 
5627 static int
5628 segvn_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
5629 {
5630 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5631 	struct vpage *cvp, *svp, *evp;
5632 	struct vnode *vp;
5633 	size_t pgsz;
5634 	pgcnt_t pgcnt;
5635 	anon_sync_obj_t cookie;
5636 	int unload_done = 0;
5637 
5638 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5639 
5640 	if ((svd->maxprot & prot) != prot)
5641 		return (EACCES);			/* violated maxprot */
5642 
5643 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5644 
5645 	/* return if prot is the same */
5646 	if (!svd->pageprot && svd->prot == prot) {
5647 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5648 		return (0);
5649 	}
5650 
5651 	/*
5652 	 * Since we change protections we first have to flush the cache.
5653 	 * This makes sure all the pagelock calls have to recheck
5654 	 * protections.
5655 	 */
5656 	if (svd->softlockcnt > 0) {
5657 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5658 
5659 		/*
5660 		 * If this is shared segment non 0 softlockcnt
5661 		 * means locked pages are still in use.
5662 		 */
5663 		if (svd->type == MAP_SHARED) {
5664 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5665 			return (EAGAIN);
5666 		}
5667 
5668 		/*
5669 		 * Since we do have the segvn writers lock nobody can fill
5670 		 * the cache with entries belonging to this seg during
5671 		 * the purge. The flush either succeeds or we still have
5672 		 * pending I/Os.
5673 		 */
5674 		segvn_purge(seg);
5675 		if (svd->softlockcnt > 0) {
5676 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5677 			return (EAGAIN);
5678 		}
5679 	}
5680 
5681 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5682 		ASSERT(svd->amp == NULL);
5683 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5684 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5685 		    HAT_REGION_TEXT);
5686 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
5687 		unload_done = 1;
5688 	} else if (svd->tr_state == SEGVN_TR_INIT) {
5689 		svd->tr_state = SEGVN_TR_OFF;
5690 	} else if (svd->tr_state == SEGVN_TR_ON) {
5691 		ASSERT(svd->amp != NULL);
5692 		segvn_textunrepl(seg, 0);
5693 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
5694 		unload_done = 1;
5695 	}
5696 
5697 	if ((prot & PROT_WRITE) && svd->type == MAP_SHARED &&
5698 	    svd->vp != NULL && (svd->vp->v_flag & VVMEXEC)) {
5699 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
5700 		segvn_inval_trcache(svd->vp);
5701 	}
5702 	if (seg->s_szc != 0) {
5703 		int err;
5704 		pgsz = page_get_pagesize(seg->s_szc);
5705 		pgcnt = pgsz >> PAGESHIFT;
5706 		ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
5707 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
5708 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5709 			ASSERT(seg->s_base != addr || seg->s_size != len);
5710 			/*
5711 			 * If we are holding the as lock as a reader then
5712 			 * we need to return IE_RETRY and let the as
5713 			 * layer drop and re-acquire the lock as a writer.
5714 			 */
5715 			if (AS_READ_HELD(seg->s_as, &seg->s_as->a_lock))
5716 				return (IE_RETRY);
5717 			VM_STAT_ADD(segvnvmstats.demoterange[1]);
5718 			if (svd->type == MAP_PRIVATE || svd->vp != NULL) {
5719 				err = segvn_demote_range(seg, addr, len,
5720 				    SDR_END, 0);
5721 			} else {
5722 				uint_t szcvec = map_pgszcvec(seg->s_base,
5723 				    pgsz, (uintptr_t)seg->s_base,
5724 				    (svd->flags & MAP_TEXT), MAPPGSZC_SHM, 0);
5725 				err = segvn_demote_range(seg, addr, len,
5726 				    SDR_END, szcvec);
5727 			}
5728 			if (err == 0)
5729 				return (IE_RETRY);
5730 			if (err == ENOMEM)
5731 				return (IE_NOMEM);
5732 			return (err);
5733 		}
5734 	}
5735 
5736 
5737 	/*
5738 	 * If it's a private mapping and we're making it writable then we
5739 	 * may have to reserve the additional swap space now. If we are
5740 	 * making writable only a part of the segment then we use its vpage
5741 	 * array to keep a record of the pages for which we have reserved
5742 	 * swap. In this case we set the pageswap field in the segment's
5743 	 * segvn structure to record this.
5744 	 *
5745 	 * If it's a private mapping to a file (i.e., vp != NULL) and we're
5746 	 * removing write permission on the entire segment and we haven't
5747 	 * modified any pages, we can release the swap space.
5748 	 */
5749 	if (svd->type == MAP_PRIVATE) {
5750 		if (prot & PROT_WRITE) {
5751 			if (!(svd->flags & MAP_NORESERVE) &&
5752 			    !(svd->swresv && svd->pageswap == 0)) {
5753 				size_t sz = 0;
5754 
5755 				/*
5756 				 * Start by determining how much swap
5757 				 * space is required.
5758 				 */
5759 				if (addr == seg->s_base &&
5760 				    len == seg->s_size &&
5761 				    svd->pageswap == 0) {
5762 					/* The whole segment */
5763 					sz = seg->s_size;
5764 				} else {
5765 					/*
5766 					 * Make sure that the vpage array
5767 					 * exists, and make a note of the
5768 					 * range of elements corresponding
5769 					 * to len.
5770 					 */
5771 					segvn_vpage(seg);
5772 					svp = &svd->vpage[seg_page(seg, addr)];
5773 					evp = &svd->vpage[seg_page(seg,
5774 					    addr + len)];
5775 
5776 					if (svd->pageswap == 0) {
5777 						/*
5778 						 * This is the first time we've
5779 						 * asked for a part of this
5780 						 * segment, so we need to
5781 						 * reserve everything we've
5782 						 * been asked for.
5783 						 */
5784 						sz = len;
5785 					} else {
5786 						/*
5787 						 * We have to count the number
5788 						 * of pages required.
5789 						 */
5790 						for (cvp = svp;  cvp < evp;
5791 						    cvp++) {
5792 							if (!VPP_ISSWAPRES(cvp))
5793 								sz++;
5794 						}
5795 						sz <<= PAGESHIFT;
5796 					}
5797 				}
5798 
5799 				/* Try to reserve the necessary swap. */
5800 				if (anon_resv_zone(sz,
5801 				    seg->s_as->a_proc->p_zone) == 0) {
5802 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5803 					return (IE_NOMEM);
5804 				}
5805 
5806 				/*
5807 				 * Make a note of how much swap space
5808 				 * we've reserved.
5809 				 */
5810 				if (svd->pageswap == 0 && sz == seg->s_size) {
5811 					svd->swresv = sz;
5812 				} else {
5813 					ASSERT(svd->vpage != NULL);
5814 					svd->swresv += sz;
5815 					svd->pageswap = 1;
5816 					for (cvp = svp; cvp < evp; cvp++) {
5817 						if (!VPP_ISSWAPRES(cvp))
5818 							VPP_SETSWAPRES(cvp);
5819 					}
5820 				}
5821 			}
5822 		} else {
5823 			/*
5824 			 * Swap space is released only if this segment
5825 			 * does not map anonymous memory, since read faults
5826 			 * on such segments still need an anon slot to read
5827 			 * in the data.
5828 			 */
5829 			if (svd->swresv != 0 && svd->vp != NULL &&
5830 			    svd->amp == NULL && addr == seg->s_base &&
5831 			    len == seg->s_size && svd->pageprot == 0) {
5832 				ASSERT(svd->pageswap == 0);
5833 				anon_unresv_zone(svd->swresv,
5834 				    seg->s_as->a_proc->p_zone);
5835 				svd->swresv = 0;
5836 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
5837 				    "anon proc:%p %lu %u", seg, 0, 0);
5838 			}
5839 		}
5840 	}
5841 
5842 	if (addr == seg->s_base && len == seg->s_size && svd->vpage == NULL) {
5843 		if (svd->prot == prot) {
5844 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5845 			return (0);			/* all done */
5846 		}
5847 		svd->prot = (uchar_t)prot;
5848 	} else if (svd->type == MAP_PRIVATE) {
5849 		struct anon *ap = NULL;
5850 		page_t *pp;
5851 		u_offset_t offset, off;
5852 		struct anon_map *amp;
5853 		ulong_t anon_idx = 0;
5854 
5855 		/*
5856 		 * A vpage structure exists or else the change does not
5857 		 * involve the entire segment.  Establish a vpage structure
5858 		 * if none is there.  Then, for each page in the range,
5859 		 * adjust its individual permissions.  Note that write-
5860 		 * enabling a MAP_PRIVATE page can affect the claims for
5861 		 * locked down memory.  Overcommitting memory terminates
5862 		 * the operation.
5863 		 */
5864 		segvn_vpage(seg);
5865 		svd->pageprot = 1;
5866 		if ((amp = svd->amp) != NULL) {
5867 			anon_idx = svd->anon_index + seg_page(seg, addr);
5868 			ASSERT(seg->s_szc == 0 ||
5869 			    IS_P2ALIGNED(anon_idx, pgcnt));
5870 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5871 		}
5872 
5873 		offset = svd->offset + (uintptr_t)(addr - seg->s_base);
5874 		evp = &svd->vpage[seg_page(seg, addr + len)];
5875 
5876 		/*
5877 		 * See Statement at the beginning of segvn_lockop regarding
5878 		 * the way cowcnts and lckcnts are handled.
5879 		 */
5880 		for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5881 
5882 			if (seg->s_szc != 0) {
5883 				if (amp != NULL) {
5884 					anon_array_enter(amp, anon_idx,
5885 					    &cookie);
5886 				}
5887 				if (IS_P2ALIGNED(anon_idx, pgcnt) &&
5888 				    !segvn_claim_pages(seg, svp, offset,
5889 				    anon_idx, prot)) {
5890 					if (amp != NULL) {
5891 						anon_array_exit(&cookie);
5892 					}
5893 					break;
5894 				}
5895 				if (amp != NULL) {
5896 					anon_array_exit(&cookie);
5897 				}
5898 				anon_idx++;
5899 			} else {
5900 				if (amp != NULL) {
5901 					anon_array_enter(amp, anon_idx,
5902 					    &cookie);
5903 					ap = anon_get_ptr(amp->ahp, anon_idx++);
5904 				}
5905 
5906 				if (VPP_ISPPLOCK(svp) &&
5907 				    VPP_PROT(svp) != prot) {
5908 
5909 					if (amp == NULL || ap == NULL) {
5910 						vp = svd->vp;
5911 						off = offset;
5912 					} else
5913 						swap_xlate(ap, &vp, &off);
5914 					if (amp != NULL)
5915 						anon_array_exit(&cookie);
5916 
5917 					if ((pp = page_lookup(vp, off,
5918 					    SE_SHARED)) == NULL) {
5919 						panic("segvn_setprot: no page");
5920 						/*NOTREACHED*/
5921 					}
5922 					ASSERT(seg->s_szc == 0);
5923 					if ((VPP_PROT(svp) ^ prot) &
5924 					    PROT_WRITE) {
5925 						if (prot & PROT_WRITE) {
5926 							if (!page_addclaim(
5927 							    pp)) {
5928 								page_unlock(pp);
5929 								break;
5930 							}
5931 						} else {
5932 							if (!page_subclaim(
5933 							    pp)) {
5934 								page_unlock(pp);
5935 								break;
5936 							}
5937 						}
5938 					}
5939 					page_unlock(pp);
5940 				} else if (amp != NULL)
5941 					anon_array_exit(&cookie);
5942 			}
5943 			VPP_SETPROT(svp, prot);
5944 			offset += PAGESIZE;
5945 		}
5946 		if (amp != NULL)
5947 			ANON_LOCK_EXIT(&amp->a_rwlock);
5948 
5949 		/*
5950 		 * Did we terminate prematurely?  If so, simply unload
5951 		 * the translations to the things we've updated so far.
5952 		 */
5953 		if (svp != evp) {
5954 			if (unload_done) {
5955 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5956 				return (IE_NOMEM);
5957 			}
5958 			len = (svp - &svd->vpage[seg_page(seg, addr)]) *
5959 			    PAGESIZE;
5960 			ASSERT(seg->s_szc == 0 || IS_P2ALIGNED(len, pgsz));
5961 			if (len != 0)
5962 				hat_unload(seg->s_as->a_hat, addr,
5963 				    len, HAT_UNLOAD);
5964 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5965 			return (IE_NOMEM);
5966 		}
5967 	} else {
5968 		segvn_vpage(seg);
5969 		svd->pageprot = 1;
5970 		evp = &svd->vpage[seg_page(seg, addr + len)];
5971 		for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5972 			VPP_SETPROT(svp, prot);
5973 		}
5974 	}
5975 
5976 	if (unload_done) {
5977 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5978 		return (0);
5979 	}
5980 
5981 	if (((prot & PROT_WRITE) != 0 &&
5982 	    (svd->vp != NULL || svd->type == MAP_PRIVATE)) ||
5983 	    (prot & ~PROT_USER) == PROT_NONE) {
5984 		/*
5985 		 * Either private or shared data with write access (in
5986 		 * which case we need to throw out all former translations
5987 		 * so that we get the right translations set up on fault
5988 		 * and we don't allow write access to any copy-on-write pages
5989 		 * that might be around or to prevent write access to pages
5990 		 * representing holes in a file), or we don't have permission
5991 		 * to access the memory at all (in which case we have to
5992 		 * unload any current translations that might exist).
5993 		 */
5994 		hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
5995 	} else {
5996 		/*
5997 		 * A shared mapping or a private mapping in which write
5998 		 * protection is going to be denied - just change all the
5999 		 * protections over the range of addresses in question.
6000 		 * segvn does not support any other attributes other
6001 		 * than prot so we can use hat_chgattr.
6002 		 */
6003 		hat_chgattr(seg->s_as->a_hat, addr, len, prot);
6004 	}
6005 
6006 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6007 
6008 	return (0);
6009 }
6010 
6011 /*
6012  * segvn_setpagesize is called via SEGOP_SETPAGESIZE from as_setpagesize,
6013  * to determine if the seg is capable of mapping the requested szc.
6014  */
6015 static int
6016 segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len, uint_t szc)
6017 {
6018 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6019 	struct segvn_data *nsvd;
6020 	struct anon_map *amp = svd->amp;
6021 	struct seg *nseg;
6022 	caddr_t eaddr = addr + len, a;
6023 	size_t pgsz = page_get_pagesize(szc);
6024 	pgcnt_t pgcnt = page_get_pagecnt(szc);
6025 	int err;
6026 	u_offset_t off = svd->offset + (uintptr_t)(addr - seg->s_base);
6027 
6028 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6029 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6030 
6031 	if (seg->s_szc == szc || segvn_lpg_disable != 0) {
6032 		return (0);
6033 	}
6034 
6035 	/*
6036 	 * addr should always be pgsz aligned but eaddr may be misaligned if
6037 	 * it's at the end of the segment.
6038 	 *
6039 	 * XXX we should assert this condition since as_setpagesize() logic
6040 	 * guarantees it.
6041 	 */
6042 	if (!IS_P2ALIGNED(addr, pgsz) ||
6043 	    (!IS_P2ALIGNED(eaddr, pgsz) &&
6044 	    eaddr != seg->s_base + seg->s_size)) {
6045 
6046 		segvn_setpgsz_align_err++;
6047 		return (EINVAL);
6048 	}
6049 
6050 	if (amp != NULL && svd->type == MAP_SHARED) {
6051 		ulong_t an_idx = svd->anon_index + seg_page(seg, addr);
6052 		if (!IS_P2ALIGNED(an_idx, pgcnt)) {
6053 
6054 			segvn_setpgsz_anon_align_err++;
6055 			return (EINVAL);
6056 		}
6057 	}
6058 
6059 	if ((svd->flags & MAP_NORESERVE) || seg->s_as == &kas ||
6060 	    szc > segvn_maxpgszc) {
6061 		return (EINVAL);
6062 	}
6063 
6064 	/* paranoid check */
6065 	if (svd->vp != NULL &&
6066 	    (IS_SWAPFSVP(svd->vp) || VN_ISKAS(svd->vp))) {
6067 		return (EINVAL);
6068 	}
6069 
6070 	if (seg->s_szc == 0 && svd->vp != NULL &&
6071 	    map_addr_vacalign_check(addr, off)) {
6072 		return (EINVAL);
6073 	}
6074 
6075 	/*
6076 	 * Check that protections are the same within new page
6077 	 * size boundaries.
6078 	 */
6079 	if (svd->pageprot) {
6080 		for (a = addr; a < eaddr; a += pgsz) {
6081 			if ((a + pgsz) > eaddr) {
6082 				if (!sameprot(seg, a, eaddr - a)) {
6083 					return (EINVAL);
6084 				}
6085 			} else {
6086 				if (!sameprot(seg, a, pgsz)) {
6087 					return (EINVAL);
6088 				}
6089 			}
6090 		}
6091 	}
6092 
6093 	/*
6094 	 * Since we are changing page size we first have to flush
6095 	 * the cache. This makes sure all the pagelock calls have
6096 	 * to recheck protections.
6097 	 */
6098 	if (svd->softlockcnt > 0) {
6099 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6100 
6101 		/*
6102 		 * If this is shared segment non 0 softlockcnt
6103 		 * means locked pages are still in use.
6104 		 */
6105 		if (svd->type == MAP_SHARED) {
6106 			return (EAGAIN);
6107 		}
6108 
6109 		/*
6110 		 * Since we do have the segvn writers lock nobody can fill
6111 		 * the cache with entries belonging to this seg during
6112 		 * the purge. The flush either succeeds or we still have
6113 		 * pending I/Os.
6114 		 */
6115 		segvn_purge(seg);
6116 		if (svd->softlockcnt > 0) {
6117 			return (EAGAIN);
6118 		}
6119 	}
6120 
6121 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6122 		ASSERT(svd->amp == NULL);
6123 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6124 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6125 		    HAT_REGION_TEXT);
6126 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6127 	} else if (svd->tr_state == SEGVN_TR_INIT) {
6128 		svd->tr_state = SEGVN_TR_OFF;
6129 	} else if (svd->tr_state == SEGVN_TR_ON) {
6130 		ASSERT(svd->amp != NULL);
6131 		segvn_textunrepl(seg, 1);
6132 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6133 		amp = NULL;
6134 	}
6135 
6136 	/*
6137 	 * Operation for sub range of existing segment.
6138 	 */
6139 	if (addr != seg->s_base || eaddr != (seg->s_base + seg->s_size)) {
6140 		if (szc < seg->s_szc) {
6141 			VM_STAT_ADD(segvnvmstats.demoterange[2]);
6142 			err = segvn_demote_range(seg, addr, len, SDR_RANGE, 0);
6143 			if (err == 0) {
6144 				return (IE_RETRY);
6145 			}
6146 			if (err == ENOMEM) {
6147 				return (IE_NOMEM);
6148 			}
6149 			return (err);
6150 		}
6151 		if (addr != seg->s_base) {
6152 			nseg = segvn_split_seg(seg, addr);
6153 			if (eaddr != (nseg->s_base + nseg->s_size)) {
6154 				/* eaddr is szc aligned */
6155 				(void) segvn_split_seg(nseg, eaddr);
6156 			}
6157 			return (IE_RETRY);
6158 		}
6159 		if (eaddr != (seg->s_base + seg->s_size)) {
6160 			/* eaddr is szc aligned */
6161 			(void) segvn_split_seg(seg, eaddr);
6162 		}
6163 		return (IE_RETRY);
6164 	}
6165 
6166 	/*
6167 	 * Break any low level sharing and reset seg->s_szc to 0.
6168 	 */
6169 	if ((err = segvn_clrszc(seg)) != 0) {
6170 		if (err == ENOMEM) {
6171 			err = IE_NOMEM;
6172 		}
6173 		return (err);
6174 	}
6175 	ASSERT(seg->s_szc == 0);
6176 
6177 	/*
6178 	 * If the end of the current segment is not pgsz aligned
6179 	 * then attempt to concatenate with the next segment.
6180 	 */
6181 	if (!IS_P2ALIGNED(eaddr, pgsz)) {
6182 		nseg = AS_SEGNEXT(seg->s_as, seg);
6183 		if (nseg == NULL || nseg == seg || eaddr != nseg->s_base) {
6184 			return (ENOMEM);
6185 		}
6186 		if (nseg->s_ops != &segvn_ops) {
6187 			return (EINVAL);
6188 		}
6189 		nsvd = (struct segvn_data *)nseg->s_data;
6190 		if (nsvd->softlockcnt > 0) {
6191 			/*
6192 			 * If this is shared segment non 0 softlockcnt
6193 			 * means locked pages are still in use.
6194 			 */
6195 			if (nsvd->type == MAP_SHARED) {
6196 				return (EAGAIN);
6197 			}
6198 			segvn_purge(nseg);
6199 			if (nsvd->softlockcnt > 0) {
6200 				return (EAGAIN);
6201 			}
6202 		}
6203 		err = segvn_clrszc(nseg);
6204 		if (err == ENOMEM) {
6205 			err = IE_NOMEM;
6206 		}
6207 		if (err != 0) {
6208 			return (err);
6209 		}
6210 		ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6211 		err = segvn_concat(seg, nseg, 1);
6212 		if (err == -1) {
6213 			return (EINVAL);
6214 		}
6215 		if (err == -2) {
6216 			return (IE_NOMEM);
6217 		}
6218 		return (IE_RETRY);
6219 	}
6220 
6221 	/*
6222 	 * May need to re-align anon array to
6223 	 * new szc.
6224 	 */
6225 	if (amp != NULL) {
6226 		if (!IS_P2ALIGNED(svd->anon_index, pgcnt)) {
6227 			struct anon_hdr *nahp;
6228 
6229 			ASSERT(svd->type == MAP_PRIVATE);
6230 
6231 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6232 			ASSERT(amp->refcnt == 1);
6233 			nahp = anon_create(btop(amp->size), ANON_NOSLEEP);
6234 			if (nahp == NULL) {
6235 				ANON_LOCK_EXIT(&amp->a_rwlock);
6236 				return (IE_NOMEM);
6237 			}
6238 			if (anon_copy_ptr(amp->ahp, svd->anon_index,
6239 			    nahp, 0, btop(seg->s_size), ANON_NOSLEEP)) {
6240 				anon_release(nahp, btop(amp->size));
6241 				ANON_LOCK_EXIT(&amp->a_rwlock);
6242 				return (IE_NOMEM);
6243 			}
6244 			anon_release(amp->ahp, btop(amp->size));
6245 			amp->ahp = nahp;
6246 			svd->anon_index = 0;
6247 			ANON_LOCK_EXIT(&amp->a_rwlock);
6248 		}
6249 	}
6250 	if (svd->vp != NULL && szc != 0) {
6251 		struct vattr va;
6252 		u_offset_t eoffpage = svd->offset;
6253 		va.va_mask = AT_SIZE;
6254 		eoffpage += seg->s_size;
6255 		eoffpage = btopr(eoffpage);
6256 		if (VOP_GETATTR(svd->vp, &va, 0, svd->cred, NULL) != 0) {
6257 			segvn_setpgsz_getattr_err++;
6258 			return (EINVAL);
6259 		}
6260 		if (btopr(va.va_size) < eoffpage) {
6261 			segvn_setpgsz_eof_err++;
6262 			return (EINVAL);
6263 		}
6264 		if (amp != NULL) {
6265 			/*
6266 			 * anon_fill_cow_holes() may call VOP_GETPAGE().
6267 			 * don't take anon map lock here to avoid holding it
6268 			 * across VOP_GETPAGE() calls that may call back into
6269 			 * segvn for klsutering checks. We don't really need
6270 			 * anon map lock here since it's a private segment and
6271 			 * we hold as level lock as writers.
6272 			 */
6273 			if ((err = anon_fill_cow_holes(seg, seg->s_base,
6274 			    amp->ahp, svd->anon_index, svd->vp, svd->offset,
6275 			    seg->s_size, szc, svd->prot, svd->vpage,
6276 			    svd->cred)) != 0) {
6277 				return (EINVAL);
6278 			}
6279 		}
6280 		segvn_setvnode_mpss(svd->vp);
6281 	}
6282 
6283 	if (amp != NULL) {
6284 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6285 		if (svd->type == MAP_PRIVATE) {
6286 			amp->a_szc = szc;
6287 		} else if (szc > amp->a_szc) {
6288 			amp->a_szc = szc;
6289 		}
6290 		ANON_LOCK_EXIT(&amp->a_rwlock);
6291 	}
6292 
6293 	seg->s_szc = szc;
6294 
6295 	return (0);
6296 }
6297 
6298 static int
6299 segvn_clrszc(struct seg *seg)
6300 {
6301 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6302 	struct anon_map *amp = svd->amp;
6303 	size_t pgsz;
6304 	pgcnt_t pages;
6305 	int err = 0;
6306 	caddr_t a = seg->s_base;
6307 	caddr_t ea = a + seg->s_size;
6308 	ulong_t an_idx = svd->anon_index;
6309 	vnode_t *vp = svd->vp;
6310 	struct vpage *vpage = svd->vpage;
6311 	page_t *anon_pl[1 + 1], *pp;
6312 	struct anon *ap, *oldap;
6313 	uint_t prot = svd->prot, vpprot;
6314 	int pageflag = 0;
6315 
6316 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
6317 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
6318 	ASSERT(svd->softlockcnt == 0);
6319 
6320 	if (vp == NULL && amp == NULL) {
6321 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6322 		seg->s_szc = 0;
6323 		return (0);
6324 	}
6325 
6326 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6327 		ASSERT(svd->amp == NULL);
6328 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6329 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6330 		    HAT_REGION_TEXT);
6331 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6332 	} else if (svd->tr_state == SEGVN_TR_ON) {
6333 		ASSERT(svd->amp != NULL);
6334 		segvn_textunrepl(seg, 1);
6335 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6336 		amp = NULL;
6337 	} else {
6338 		if (svd->tr_state != SEGVN_TR_OFF) {
6339 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
6340 			svd->tr_state = SEGVN_TR_OFF;
6341 		}
6342 
6343 		/*
6344 		 * do HAT_UNLOAD_UNMAP since we are changing the pagesize.
6345 		 * unload argument is 0 when we are freeing the segment
6346 		 * and unload was already done.
6347 		 */
6348 		hat_unload(seg->s_as->a_hat, seg->s_base, seg->s_size,
6349 		    HAT_UNLOAD_UNMAP);
6350 	}
6351 
6352 	if (amp == NULL || svd->type == MAP_SHARED) {
6353 		seg->s_szc = 0;
6354 		return (0);
6355 	}
6356 
6357 	pgsz = page_get_pagesize(seg->s_szc);
6358 	pages = btop(pgsz);
6359 
6360 	/*
6361 	 * XXX anon rwlock is not really needed because this is a
6362 	 * private segment and we are writers.
6363 	 */
6364 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6365 
6366 	for (; a < ea; a += pgsz, an_idx += pages) {
6367 		if ((oldap = anon_get_ptr(amp->ahp, an_idx)) != NULL) {
6368 			ASSERT(vpage != NULL || svd->pageprot == 0);
6369 			if (vpage != NULL) {
6370 				ASSERT(sameprot(seg, a, pgsz));
6371 				prot = VPP_PROT(vpage);
6372 				pageflag = VPP_ISPPLOCK(vpage) ? LOCK_PAGE : 0;
6373 			}
6374 			if (seg->s_szc != 0) {
6375 				ASSERT(vp == NULL || anon_pages(amp->ahp,
6376 				    an_idx, pages) == pages);
6377 				if ((err = anon_map_demotepages(amp, an_idx,
6378 				    seg, a, prot, vpage, svd->cred)) != 0) {
6379 					goto out;
6380 				}
6381 			} else {
6382 				if (oldap->an_refcnt == 1) {
6383 					continue;
6384 				}
6385 				if ((err = anon_getpage(&oldap, &vpprot,
6386 				    anon_pl, PAGESIZE, seg, a, S_READ,
6387 				    svd->cred))) {
6388 					goto out;
6389 				}
6390 				if ((pp = anon_private(&ap, seg, a, prot,
6391 				    anon_pl[0], pageflag, svd->cred)) == NULL) {
6392 					err = ENOMEM;
6393 					goto out;
6394 				}
6395 				anon_decref(oldap);
6396 				(void) anon_set_ptr(amp->ahp, an_idx, ap,
6397 				    ANON_SLEEP);
6398 				page_unlock(pp);
6399 			}
6400 		}
6401 		vpage = (vpage == NULL) ? NULL : vpage + pages;
6402 	}
6403 
6404 	amp->a_szc = 0;
6405 	seg->s_szc = 0;
6406 out:
6407 	ANON_LOCK_EXIT(&amp->a_rwlock);
6408 	return (err);
6409 }
6410 
6411 static int
6412 segvn_claim_pages(
6413 	struct seg *seg,
6414 	struct vpage *svp,
6415 	u_offset_t off,
6416 	ulong_t anon_idx,
6417 	uint_t prot)
6418 {
6419 	pgcnt_t	pgcnt = page_get_pagecnt(seg->s_szc);
6420 	size_t ppasize = (pgcnt + 1) * sizeof (page_t *);
6421 	page_t	**ppa;
6422 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6423 	struct anon_map *amp = svd->amp;
6424 	struct vpage *evp = svp + pgcnt;
6425 	caddr_t addr = ((uintptr_t)(svp - svd->vpage) << PAGESHIFT)
6426 	    + seg->s_base;
6427 	struct anon *ap;
6428 	struct vnode *vp = svd->vp;
6429 	page_t *pp;
6430 	pgcnt_t pg_idx, i;
6431 	int err = 0;
6432 	anoff_t aoff;
6433 	int anon = (amp != NULL) ? 1 : 0;
6434 
6435 	ASSERT(svd->type == MAP_PRIVATE);
6436 	ASSERT(svd->vpage != NULL);
6437 	ASSERT(seg->s_szc != 0);
6438 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
6439 	ASSERT(amp == NULL || IS_P2ALIGNED(anon_idx, pgcnt));
6440 	ASSERT(sameprot(seg, addr, pgcnt << PAGESHIFT));
6441 
6442 	if (VPP_PROT(svp) == prot)
6443 		return (1);
6444 	if (!((VPP_PROT(svp) ^ prot) & PROT_WRITE))
6445 		return (1);
6446 
6447 	ppa = kmem_alloc(ppasize, KM_SLEEP);
6448 	if (anon && vp != NULL) {
6449 		if (anon_get_ptr(amp->ahp, anon_idx) == NULL) {
6450 			anon = 0;
6451 			ASSERT(!anon_pages(amp->ahp, anon_idx, pgcnt));
6452 		}
6453 		ASSERT(!anon ||
6454 		    anon_pages(amp->ahp, anon_idx, pgcnt) == pgcnt);
6455 	}
6456 
6457 	for (*ppa = NULL, pg_idx = 0; svp < evp; svp++, anon_idx++) {
6458 		if (!VPP_ISPPLOCK(svp))
6459 			continue;
6460 		if (anon) {
6461 			ap = anon_get_ptr(amp->ahp, anon_idx);
6462 			if (ap == NULL) {
6463 				panic("segvn_claim_pages: no anon slot");
6464 			}
6465 			swap_xlate(ap, &vp, &aoff);
6466 			off = (u_offset_t)aoff;
6467 		}
6468 		ASSERT(vp != NULL);
6469 		if ((pp = page_lookup(vp,
6470 		    (u_offset_t)off, SE_SHARED)) == NULL) {
6471 			panic("segvn_claim_pages: no page");
6472 		}
6473 		ppa[pg_idx++] = pp;
6474 		off += PAGESIZE;
6475 	}
6476 
6477 	if (ppa[0] == NULL) {
6478 		kmem_free(ppa, ppasize);
6479 		return (1);
6480 	}
6481 
6482 	ASSERT(pg_idx <= pgcnt);
6483 	ppa[pg_idx] = NULL;
6484 
6485 
6486 	/* Find each large page within ppa, and adjust its claim */
6487 
6488 	/* Does ppa cover a single large page? */
6489 	if (ppa[0]->p_szc == seg->s_szc) {
6490 		if (prot & PROT_WRITE)
6491 			err = page_addclaim_pages(ppa);
6492 		else
6493 			err = page_subclaim_pages(ppa);
6494 	} else {
6495 		for (i = 0; ppa[i]; i += pgcnt) {
6496 			ASSERT(IS_P2ALIGNED(page_pptonum(ppa[i]), pgcnt));
6497 			if (prot & PROT_WRITE)
6498 				err = page_addclaim_pages(&ppa[i]);
6499 			else
6500 				err = page_subclaim_pages(&ppa[i]);
6501 			if (err == 0)
6502 				break;
6503 		}
6504 	}
6505 
6506 	for (i = 0; i < pg_idx; i++) {
6507 		ASSERT(ppa[i] != NULL);
6508 		page_unlock(ppa[i]);
6509 	}
6510 
6511 	kmem_free(ppa, ppasize);
6512 	return (err);
6513 }
6514 
6515 /*
6516  * Returns right (upper address) segment if split occurred.
6517  * If the address is equal to the beginning or end of its segment it returns
6518  * the current segment.
6519  */
6520 static struct seg *
6521 segvn_split_seg(struct seg *seg, caddr_t addr)
6522 {
6523 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6524 	struct seg *nseg;
6525 	size_t nsize;
6526 	struct segvn_data *nsvd;
6527 
6528 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6529 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6530 
6531 	ASSERT(addr >= seg->s_base);
6532 	ASSERT(addr <= seg->s_base + seg->s_size);
6533 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6534 
6535 	if (addr == seg->s_base || addr == seg->s_base + seg->s_size)
6536 		return (seg);
6537 
6538 	nsize = seg->s_base + seg->s_size - addr;
6539 	seg->s_size = addr - seg->s_base;
6540 	nseg = seg_alloc(seg->s_as, addr, nsize);
6541 	ASSERT(nseg != NULL);
6542 	nseg->s_ops = seg->s_ops;
6543 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
6544 	nseg->s_data = (void *)nsvd;
6545 	nseg->s_szc = seg->s_szc;
6546 	*nsvd = *svd;
6547 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6548 	nsvd->seg = nseg;
6549 	rw_init(&nsvd->lock, NULL, RW_DEFAULT, NULL);
6550 
6551 	if (nsvd->vp != NULL) {
6552 		VN_HOLD(nsvd->vp);
6553 		nsvd->offset = svd->offset +
6554 		    (uintptr_t)(nseg->s_base - seg->s_base);
6555 		if (nsvd->type == MAP_SHARED)
6556 			lgrp_shm_policy_init(NULL, nsvd->vp);
6557 	} else {
6558 		/*
6559 		 * The offset for an anonymous segment has no signifigance in
6560 		 * terms of an offset into a file. If we were to use the above
6561 		 * calculation instead, the structures read out of
6562 		 * /proc/<pid>/xmap would be more difficult to decipher since
6563 		 * it would be unclear whether two seemingly contiguous
6564 		 * prxmap_t structures represented different segments or a
6565 		 * single segment that had been split up into multiple prxmap_t
6566 		 * structures (e.g. if some part of the segment had not yet
6567 		 * been faulted in).
6568 		 */
6569 		nsvd->offset = 0;
6570 	}
6571 
6572 	ASSERT(svd->softlockcnt == 0);
6573 	ASSERT(svd->softlockcnt_sbase == 0);
6574 	ASSERT(svd->softlockcnt_send == 0);
6575 	crhold(svd->cred);
6576 
6577 	if (svd->vpage != NULL) {
6578 		size_t bytes = vpgtob(seg_pages(seg));
6579 		size_t nbytes = vpgtob(seg_pages(nseg));
6580 		struct vpage *ovpage = svd->vpage;
6581 
6582 		svd->vpage = kmem_alloc(bytes, KM_SLEEP);
6583 		bcopy(ovpage, svd->vpage, bytes);
6584 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
6585 		bcopy(ovpage + seg_pages(seg), nsvd->vpage, nbytes);
6586 		kmem_free(ovpage, bytes + nbytes);
6587 	}
6588 	if (svd->amp != NULL && svd->type == MAP_PRIVATE) {
6589 		struct anon_map *oamp = svd->amp, *namp;
6590 		struct anon_hdr *nahp;
6591 
6592 		ANON_LOCK_ENTER(&oamp->a_rwlock, RW_WRITER);
6593 		ASSERT(oamp->refcnt == 1);
6594 		nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
6595 		(void) anon_copy_ptr(oamp->ahp, svd->anon_index,
6596 		    nahp, 0, btop(seg->s_size), ANON_SLEEP);
6597 
6598 		namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
6599 		namp->a_szc = nseg->s_szc;
6600 		(void) anon_copy_ptr(oamp->ahp,
6601 		    svd->anon_index + btop(seg->s_size),
6602 		    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
6603 		anon_release(oamp->ahp, btop(oamp->size));
6604 		oamp->ahp = nahp;
6605 		oamp->size = seg->s_size;
6606 		svd->anon_index = 0;
6607 		nsvd->amp = namp;
6608 		nsvd->anon_index = 0;
6609 		ANON_LOCK_EXIT(&oamp->a_rwlock);
6610 	} else if (svd->amp != NULL) {
6611 		pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6612 		ASSERT(svd->amp == nsvd->amp);
6613 		ASSERT(seg->s_szc <= svd->amp->a_szc);
6614 		nsvd->anon_index = svd->anon_index + seg_pages(seg);
6615 		ASSERT(IS_P2ALIGNED(nsvd->anon_index, pgcnt));
6616 		ANON_LOCK_ENTER(&svd->amp->a_rwlock, RW_WRITER);
6617 		svd->amp->refcnt++;
6618 		ANON_LOCK_EXIT(&svd->amp->a_rwlock);
6619 	}
6620 
6621 	/*
6622 	 * Split the amount of swap reserved.
6623 	 */
6624 	if (svd->swresv) {
6625 		/*
6626 		 * For MAP_NORESERVE, only allocate swap reserve for pages
6627 		 * being used.  Other segments get enough to cover whole
6628 		 * segment.
6629 		 */
6630 		if (svd->flags & MAP_NORESERVE) {
6631 			size_t	oswresv;
6632 
6633 			ASSERT(svd->amp);
6634 			oswresv = svd->swresv;
6635 			svd->swresv = ptob(anon_pages(svd->amp->ahp,
6636 			    svd->anon_index, btop(seg->s_size)));
6637 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
6638 			    nsvd->anon_index, btop(nseg->s_size)));
6639 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
6640 		} else {
6641 			if (svd->pageswap) {
6642 				svd->swresv = segvn_count_swap_by_vpages(seg);
6643 				ASSERT(nsvd->swresv >= svd->swresv);
6644 				nsvd->swresv -= svd->swresv;
6645 			} else {
6646 				ASSERT(svd->swresv == seg->s_size +
6647 				    nseg->s_size);
6648 				svd->swresv = seg->s_size;
6649 				nsvd->swresv = nseg->s_size;
6650 			}
6651 		}
6652 	}
6653 
6654 	return (nseg);
6655 }
6656 
6657 /*
6658  * called on memory operations (unmap, setprot, setpagesize) for a subset
6659  * of a large page segment to either demote the memory range (SDR_RANGE)
6660  * or the ends (SDR_END) by addr/len.
6661  *
6662  * returns 0 on success. returns errno, including ENOMEM, on failure.
6663  */
6664 static int
6665 segvn_demote_range(
6666 	struct seg *seg,
6667 	caddr_t addr,
6668 	size_t len,
6669 	int flag,
6670 	uint_t szcvec)
6671 {
6672 	caddr_t eaddr = addr + len;
6673 	caddr_t lpgaddr, lpgeaddr;
6674 	struct seg *nseg;
6675 	struct seg *badseg1 = NULL;
6676 	struct seg *badseg2 = NULL;
6677 	size_t pgsz;
6678 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6679 	int err;
6680 	uint_t szc = seg->s_szc;
6681 	uint_t tszcvec;
6682 
6683 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6684 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6685 	ASSERT(szc != 0);
6686 	pgsz = page_get_pagesize(szc);
6687 	ASSERT(seg->s_base != addr || seg->s_size != len);
6688 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6689 	ASSERT(svd->softlockcnt == 0);
6690 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6691 	ASSERT(szcvec == 0 || (flag == SDR_END && svd->type == MAP_SHARED));
6692 
6693 	CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
6694 	ASSERT(flag == SDR_RANGE || eaddr < lpgeaddr || addr > lpgaddr);
6695 	if (flag == SDR_RANGE) {
6696 		/* demote entire range */
6697 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6698 		(void) segvn_split_seg(nseg, lpgeaddr);
6699 		ASSERT(badseg1->s_base == lpgaddr);
6700 		ASSERT(badseg1->s_size == lpgeaddr - lpgaddr);
6701 	} else if (addr != lpgaddr) {
6702 		ASSERT(flag == SDR_END);
6703 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6704 		if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz &&
6705 		    eaddr < lpgaddr + 2 * pgsz) {
6706 			(void) segvn_split_seg(nseg, lpgeaddr);
6707 			ASSERT(badseg1->s_base == lpgaddr);
6708 			ASSERT(badseg1->s_size == 2 * pgsz);
6709 		} else {
6710 			nseg = segvn_split_seg(nseg, lpgaddr + pgsz);
6711 			ASSERT(badseg1->s_base == lpgaddr);
6712 			ASSERT(badseg1->s_size == pgsz);
6713 			if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz) {
6714 				ASSERT(lpgeaddr - lpgaddr > 2 * pgsz);
6715 				nseg = segvn_split_seg(nseg, lpgeaddr - pgsz);
6716 				badseg2 = nseg;
6717 				(void) segvn_split_seg(nseg, lpgeaddr);
6718 				ASSERT(badseg2->s_base == lpgeaddr - pgsz);
6719 				ASSERT(badseg2->s_size == pgsz);
6720 			}
6721 		}
6722 	} else {
6723 		ASSERT(flag == SDR_END);
6724 		ASSERT(eaddr < lpgeaddr);
6725 		badseg1 = nseg = segvn_split_seg(seg, lpgeaddr - pgsz);
6726 		(void) segvn_split_seg(nseg, lpgeaddr);
6727 		ASSERT(badseg1->s_base == lpgeaddr - pgsz);
6728 		ASSERT(badseg1->s_size == pgsz);
6729 	}
6730 
6731 	ASSERT(badseg1 != NULL);
6732 	ASSERT(badseg1->s_szc == szc);
6733 	ASSERT(flag == SDR_RANGE || badseg1->s_size == pgsz ||
6734 	    badseg1->s_size == 2 * pgsz);
6735 	ASSERT(sameprot(badseg1, badseg1->s_base, pgsz));
6736 	ASSERT(badseg1->s_size == pgsz ||
6737 	    sameprot(badseg1, badseg1->s_base + pgsz, pgsz));
6738 	if (err = segvn_clrszc(badseg1)) {
6739 		return (err);
6740 	}
6741 	ASSERT(badseg1->s_szc == 0);
6742 
6743 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6744 		uint_t tszc = highbit(tszcvec) - 1;
6745 		caddr_t ta = MAX(addr, badseg1->s_base);
6746 		caddr_t te;
6747 		size_t tpgsz = page_get_pagesize(tszc);
6748 
6749 		ASSERT(svd->type == MAP_SHARED);
6750 		ASSERT(flag == SDR_END);
6751 		ASSERT(tszc < szc && tszc > 0);
6752 
6753 		if (eaddr > badseg1->s_base + badseg1->s_size) {
6754 			te = badseg1->s_base + badseg1->s_size;
6755 		} else {
6756 			te = eaddr;
6757 		}
6758 
6759 		ASSERT(ta <= te);
6760 		badseg1->s_szc = tszc;
6761 		if (!IS_P2ALIGNED(ta, tpgsz) || !IS_P2ALIGNED(te, tpgsz)) {
6762 			if (badseg2 != NULL) {
6763 				err = segvn_demote_range(badseg1, ta, te - ta,
6764 				    SDR_END, tszcvec);
6765 				if (err != 0) {
6766 					return (err);
6767 				}
6768 			} else {
6769 				return (segvn_demote_range(badseg1, ta,
6770 				    te - ta, SDR_END, tszcvec));
6771 			}
6772 		}
6773 	}
6774 
6775 	if (badseg2 == NULL)
6776 		return (0);
6777 	ASSERT(badseg2->s_szc == szc);
6778 	ASSERT(badseg2->s_size == pgsz);
6779 	ASSERT(sameprot(badseg2, badseg2->s_base, badseg2->s_size));
6780 	if (err = segvn_clrszc(badseg2)) {
6781 		return (err);
6782 	}
6783 	ASSERT(badseg2->s_szc == 0);
6784 
6785 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6786 		uint_t tszc = highbit(tszcvec) - 1;
6787 		size_t tpgsz = page_get_pagesize(tszc);
6788 
6789 		ASSERT(svd->type == MAP_SHARED);
6790 		ASSERT(flag == SDR_END);
6791 		ASSERT(tszc < szc && tszc > 0);
6792 		ASSERT(badseg2->s_base > addr);
6793 		ASSERT(eaddr > badseg2->s_base);
6794 		ASSERT(eaddr < badseg2->s_base + badseg2->s_size);
6795 
6796 		badseg2->s_szc = tszc;
6797 		if (!IS_P2ALIGNED(eaddr, tpgsz)) {
6798 			return (segvn_demote_range(badseg2, badseg2->s_base,
6799 			    eaddr - badseg2->s_base, SDR_END, tszcvec));
6800 		}
6801 	}
6802 
6803 	return (0);
6804 }
6805 
6806 static int
6807 segvn_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
6808 {
6809 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6810 	struct vpage *vp, *evp;
6811 
6812 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6813 
6814 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6815 	/*
6816 	 * If segment protection can be used, simply check against them.
6817 	 */
6818 	if (svd->pageprot == 0) {
6819 		int err;
6820 
6821 		err = ((svd->prot & prot) != prot) ? EACCES : 0;
6822 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6823 		return (err);
6824 	}
6825 
6826 	/*
6827 	 * Have to check down to the vpage level.
6828 	 */
6829 	evp = &svd->vpage[seg_page(seg, addr + len)];
6830 	for (vp = &svd->vpage[seg_page(seg, addr)]; vp < evp; vp++) {
6831 		if ((VPP_PROT(vp) & prot) != prot) {
6832 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6833 			return (EACCES);
6834 		}
6835 	}
6836 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6837 	return (0);
6838 }
6839 
6840 static int
6841 segvn_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
6842 {
6843 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6844 	size_t pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
6845 
6846 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6847 
6848 	if (pgno != 0) {
6849 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6850 		if (svd->pageprot == 0) {
6851 			do {
6852 				protv[--pgno] = svd->prot;
6853 			} while (pgno != 0);
6854 		} else {
6855 			size_t pgoff = seg_page(seg, addr);
6856 
6857 			do {
6858 				pgno--;
6859 				protv[pgno] = VPP_PROT(&svd->vpage[pgno+pgoff]);
6860 			} while (pgno != 0);
6861 		}
6862 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6863 	}
6864 	return (0);
6865 }
6866 
6867 static u_offset_t
6868 segvn_getoffset(struct seg *seg, caddr_t addr)
6869 {
6870 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6871 
6872 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6873 
6874 	return (svd->offset + (uintptr_t)(addr - seg->s_base));
6875 }
6876 
6877 /*ARGSUSED*/
6878 static int
6879 segvn_gettype(struct seg *seg, caddr_t addr)
6880 {
6881 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6882 
6883 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6884 
6885 	return (svd->type | (svd->flags & (MAP_NORESERVE | MAP_TEXT |
6886 	    MAP_INITDATA)));
6887 }
6888 
6889 /*ARGSUSED*/
6890 static int
6891 segvn_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
6892 {
6893 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6894 
6895 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6896 
6897 	*vpp = svd->vp;
6898 	return (0);
6899 }
6900 
6901 /*
6902  * Check to see if it makes sense to do kluster/read ahead to
6903  * addr + delta relative to the mapping at addr.  We assume here
6904  * that delta is a signed PAGESIZE'd multiple (which can be negative).
6905  *
6906  * For segvn, we currently "approve" of the action if we are
6907  * still in the segment and it maps from the same vp/off,
6908  * or if the advice stored in segvn_data or vpages allows it.
6909  * Currently, klustering is not allowed only if MADV_RANDOM is set.
6910  */
6911 static int
6912 segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
6913 {
6914 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6915 	struct anon *oap, *ap;
6916 	ssize_t pd;
6917 	size_t page;
6918 	struct vnode *vp1, *vp2;
6919 	u_offset_t off1, off2;
6920 	struct anon_map *amp;
6921 
6922 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6923 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
6924 	    SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
6925 
6926 	if (addr + delta < seg->s_base ||
6927 	    addr + delta >= (seg->s_base + seg->s_size))
6928 		return (-1);		/* exceeded segment bounds */
6929 
6930 	pd = delta / (ssize_t)PAGESIZE;	/* divide to preserve sign bit */
6931 	page = seg_page(seg, addr);
6932 
6933 	/*
6934 	 * Check to see if either of the pages addr or addr + delta
6935 	 * have advice set that prevents klustering (if MADV_RANDOM advice
6936 	 * is set for entire segment, or MADV_SEQUENTIAL is set and delta
6937 	 * is negative).
6938 	 */
6939 	if (svd->advice == MADV_RANDOM ||
6940 	    svd->advice == MADV_SEQUENTIAL && delta < 0)
6941 		return (-1);
6942 	else if (svd->pageadvice && svd->vpage) {
6943 		struct vpage *bvpp, *evpp;
6944 
6945 		bvpp = &svd->vpage[page];
6946 		evpp = &svd->vpage[page + pd];
6947 		if (VPP_ADVICE(bvpp) == MADV_RANDOM ||
6948 		    VPP_ADVICE(evpp) == MADV_SEQUENTIAL && delta < 0)
6949 			return (-1);
6950 		if (VPP_ADVICE(bvpp) != VPP_ADVICE(evpp) &&
6951 		    VPP_ADVICE(evpp) == MADV_RANDOM)
6952 			return (-1);
6953 	}
6954 
6955 	if (svd->type == MAP_SHARED)
6956 		return (0);		/* shared mapping - all ok */
6957 
6958 	if ((amp = svd->amp) == NULL)
6959 		return (0);		/* off original vnode */
6960 
6961 	page += svd->anon_index;
6962 
6963 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
6964 
6965 	oap = anon_get_ptr(amp->ahp, page);
6966 	ap = anon_get_ptr(amp->ahp, page + pd);
6967 
6968 	ANON_LOCK_EXIT(&amp->a_rwlock);
6969 
6970 	if ((oap == NULL && ap != NULL) || (oap != NULL && ap == NULL)) {
6971 		return (-1);		/* one with and one without an anon */
6972 	}
6973 
6974 	if (oap == NULL) {		/* implies that ap == NULL */
6975 		return (0);		/* off original vnode */
6976 	}
6977 
6978 	/*
6979 	 * Now we know we have two anon pointers - check to
6980 	 * see if they happen to be properly allocated.
6981 	 */
6982 
6983 	/*
6984 	 * XXX We cheat here and don't lock the anon slots. We can't because
6985 	 * we may have been called from the anon layer which might already
6986 	 * have locked them. We are holding a refcnt on the slots so they
6987 	 * can't disappear. The worst that will happen is we'll get the wrong
6988 	 * names (vp, off) for the slots and make a poor klustering decision.
6989 	 */
6990 	swap_xlate(ap, &vp1, &off1);
6991 	swap_xlate(oap, &vp2, &off2);
6992 
6993 
6994 	if (!VOP_CMP(vp1, vp2, NULL) || off1 - off2 != delta)
6995 		return (-1);
6996 	return (0);
6997 }
6998 
6999 /*
7000  * Swap the pages of seg out to secondary storage, returning the
7001  * number of bytes of storage freed.
7002  *
7003  * The basic idea is first to unload all translations and then to call
7004  * VOP_PUTPAGE() for all newly-unmapped pages, to push them out to the
7005  * swap device.  Pages to which other segments have mappings will remain
7006  * mapped and won't be swapped.  Our caller (as_swapout) has already
7007  * performed the unloading step.
7008  *
7009  * The value returned is intended to correlate well with the process's
7010  * memory requirements.  However, there are some caveats:
7011  * 1)	When given a shared segment as argument, this routine will
7012  *	only succeed in swapping out pages for the last sharer of the
7013  *	segment.  (Previous callers will only have decremented mapping
7014  *	reference counts.)
7015  * 2)	We assume that the hat layer maintains a large enough translation
7016  *	cache to capture process reference patterns.
7017  */
7018 static size_t
7019 segvn_swapout(struct seg *seg)
7020 {
7021 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7022 	struct anon_map *amp;
7023 	pgcnt_t pgcnt = 0;
7024 	pgcnt_t npages;
7025 	pgcnt_t page;
7026 	ulong_t anon_index;
7027 
7028 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7029 
7030 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7031 	/*
7032 	 * Find pages unmapped by our caller and force them
7033 	 * out to the virtual swap device.
7034 	 */
7035 	if ((amp = svd->amp) != NULL)
7036 		anon_index = svd->anon_index;
7037 	npages = seg->s_size >> PAGESHIFT;
7038 	for (page = 0; page < npages; page++) {
7039 		page_t *pp;
7040 		struct anon *ap;
7041 		struct vnode *vp;
7042 		u_offset_t off;
7043 		anon_sync_obj_t cookie;
7044 
7045 		/*
7046 		 * Obtain <vp, off> pair for the page, then look it up.
7047 		 *
7048 		 * Note that this code is willing to consider regular
7049 		 * pages as well as anon pages.  Is this appropriate here?
7050 		 */
7051 		ap = NULL;
7052 		if (amp != NULL) {
7053 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7054 			if (anon_array_try_enter(amp, anon_index + page,
7055 			    &cookie)) {
7056 				ANON_LOCK_EXIT(&amp->a_rwlock);
7057 				continue;
7058 			}
7059 			ap = anon_get_ptr(amp->ahp, anon_index + page);
7060 			if (ap != NULL) {
7061 				swap_xlate(ap, &vp, &off);
7062 			} else {
7063 				vp = svd->vp;
7064 				off = svd->offset + ptob(page);
7065 			}
7066 			anon_array_exit(&cookie);
7067 			ANON_LOCK_EXIT(&amp->a_rwlock);
7068 		} else {
7069 			vp = svd->vp;
7070 			off = svd->offset + ptob(page);
7071 		}
7072 		if (vp == NULL) {		/* untouched zfod page */
7073 			ASSERT(ap == NULL);
7074 			continue;
7075 		}
7076 
7077 		pp = page_lookup_nowait(vp, off, SE_SHARED);
7078 		if (pp == NULL)
7079 			continue;
7080 
7081 
7082 		/*
7083 		 * Examine the page to see whether it can be tossed out,
7084 		 * keeping track of how many we've found.
7085 		 */
7086 		if (!page_tryupgrade(pp)) {
7087 			/*
7088 			 * If the page has an i/o lock and no mappings,
7089 			 * it's very likely that the page is being
7090 			 * written out as a result of klustering.
7091 			 * Assume this is so and take credit for it here.
7092 			 */
7093 			if (!page_io_trylock(pp)) {
7094 				if (!hat_page_is_mapped(pp))
7095 					pgcnt++;
7096 			} else {
7097 				page_io_unlock(pp);
7098 			}
7099 			page_unlock(pp);
7100 			continue;
7101 		}
7102 		ASSERT(!page_iolock_assert(pp));
7103 
7104 
7105 		/*
7106 		 * Skip if page is locked or has mappings.
7107 		 * We don't need the page_struct_lock to look at lckcnt
7108 		 * and cowcnt because the page is exclusive locked.
7109 		 */
7110 		if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0 ||
7111 		    hat_page_is_mapped(pp)) {
7112 			page_unlock(pp);
7113 			continue;
7114 		}
7115 
7116 		/*
7117 		 * dispose skips large pages so try to demote first.
7118 		 */
7119 		if (pp->p_szc != 0 && !page_try_demote_pages(pp)) {
7120 			page_unlock(pp);
7121 			/*
7122 			 * XXX should skip the remaining page_t's of this
7123 			 * large page.
7124 			 */
7125 			continue;
7126 		}
7127 
7128 		ASSERT(pp->p_szc == 0);
7129 
7130 		/*
7131 		 * No longer mapped -- we can toss it out.  How
7132 		 * we do so depends on whether or not it's dirty.
7133 		 */
7134 		if (hat_ismod(pp) && pp->p_vnode) {
7135 			/*
7136 			 * We must clean the page before it can be
7137 			 * freed.  Setting B_FREE will cause pvn_done
7138 			 * to free the page when the i/o completes.
7139 			 * XXX:	This also causes it to be accounted
7140 			 *	as a pageout instead of a swap: need
7141 			 *	B_SWAPOUT bit to use instead of B_FREE.
7142 			 *
7143 			 * Hold the vnode before releasing the page lock
7144 			 * to prevent it from being freed and re-used by
7145 			 * some other thread.
7146 			 */
7147 			VN_HOLD(vp);
7148 			page_unlock(pp);
7149 
7150 			/*
7151 			 * Queue all i/o requests for the pageout thread
7152 			 * to avoid saturating the pageout devices.
7153 			 */
7154 			if (!queue_io_request(vp, off))
7155 				VN_RELE(vp);
7156 		} else {
7157 			/*
7158 			 * The page was clean, free it.
7159 			 *
7160 			 * XXX:	Can we ever encounter modified pages
7161 			 *	with no associated vnode here?
7162 			 */
7163 			ASSERT(pp->p_vnode != NULL);
7164 			/*LINTED: constant in conditional context*/
7165 			VN_DISPOSE(pp, B_FREE, 0, kcred);
7166 		}
7167 
7168 		/*
7169 		 * Credit now even if i/o is in progress.
7170 		 */
7171 		pgcnt++;
7172 	}
7173 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7174 
7175 	/*
7176 	 * Wakeup pageout to initiate i/o on all queued requests.
7177 	 */
7178 	cv_signal_pageout();
7179 	return (ptob(pgcnt));
7180 }
7181 
7182 /*
7183  * Synchronize primary storage cache with real object in virtual memory.
7184  *
7185  * XXX - Anonymous pages should not be sync'ed out at all.
7186  */
7187 static int
7188 segvn_sync(struct seg *seg, caddr_t addr, size_t len, int attr, uint_t flags)
7189 {
7190 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7191 	struct vpage *vpp;
7192 	page_t *pp;
7193 	u_offset_t offset;
7194 	struct vnode *vp;
7195 	u_offset_t off;
7196 	caddr_t eaddr;
7197 	int bflags;
7198 	int err = 0;
7199 	int segtype;
7200 	int pageprot;
7201 	int prot;
7202 	ulong_t anon_index;
7203 	struct anon_map *amp;
7204 	struct anon *ap;
7205 	anon_sync_obj_t cookie;
7206 
7207 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7208 
7209 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7210 
7211 	if (svd->softlockcnt > 0) {
7212 		/*
7213 		 * If this is shared segment non 0 softlockcnt
7214 		 * means locked pages are still in use.
7215 		 */
7216 		if (svd->type == MAP_SHARED) {
7217 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7218 			return (EAGAIN);
7219 		}
7220 
7221 		/*
7222 		 * flush all pages from seg cache
7223 		 * otherwise we may deadlock in swap_putpage
7224 		 * for B_INVAL page (4175402).
7225 		 *
7226 		 * Even if we grab segvn WRITER's lock
7227 		 * here, there might be another thread which could've
7228 		 * successfully performed lookup/insert just before
7229 		 * we acquired the lock here.  So, grabbing either
7230 		 * lock here is of not much use.  Until we devise
7231 		 * a strategy at upper layers to solve the
7232 		 * synchronization issues completely, we expect
7233 		 * applications to handle this appropriately.
7234 		 */
7235 		segvn_purge(seg);
7236 		if (svd->softlockcnt > 0) {
7237 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7238 			return (EAGAIN);
7239 		}
7240 	} else if (svd->type == MAP_SHARED && svd->amp != NULL &&
7241 	    svd->amp->a_softlockcnt > 0) {
7242 		/*
7243 		 * Try to purge this amp's entries from pcache. It will
7244 		 * succeed only if other segments that share the amp have no
7245 		 * outstanding softlock's.
7246 		 */
7247 		segvn_purge(seg);
7248 		if (svd->amp->a_softlockcnt > 0 || svd->softlockcnt > 0) {
7249 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7250 			return (EAGAIN);
7251 		}
7252 	}
7253 
7254 	vpp = svd->vpage;
7255 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7256 	bflags = ((flags & MS_ASYNC) ? B_ASYNC : 0) |
7257 	    ((flags & MS_INVALIDATE) ? B_INVAL : 0);
7258 
7259 	if (attr) {
7260 		pageprot = attr & ~(SHARED|PRIVATE);
7261 		segtype = (attr & SHARED) ? MAP_SHARED : MAP_PRIVATE;
7262 
7263 		/*
7264 		 * We are done if the segment types don't match
7265 		 * or if we have segment level protections and
7266 		 * they don't match.
7267 		 */
7268 		if (svd->type != segtype) {
7269 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7270 			return (0);
7271 		}
7272 		if (vpp == NULL) {
7273 			if (svd->prot != pageprot) {
7274 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7275 				return (0);
7276 			}
7277 			prot = svd->prot;
7278 		} else
7279 			vpp = &svd->vpage[seg_page(seg, addr)];
7280 
7281 	} else if (svd->vp && svd->amp == NULL &&
7282 	    (flags & MS_INVALIDATE) == 0) {
7283 
7284 		/*
7285 		 * No attributes, no anonymous pages and MS_INVALIDATE flag
7286 		 * is not on, just use one big request.
7287 		 */
7288 		err = VOP_PUTPAGE(svd->vp, (offset_t)offset, len,
7289 		    bflags, svd->cred, NULL);
7290 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7291 		return (err);
7292 	}
7293 
7294 	if ((amp = svd->amp) != NULL)
7295 		anon_index = svd->anon_index + seg_page(seg, addr);
7296 
7297 	for (eaddr = addr + len; addr < eaddr; addr += PAGESIZE) {
7298 		ap = NULL;
7299 		if (amp != NULL) {
7300 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7301 			anon_array_enter(amp, anon_index, &cookie);
7302 			ap = anon_get_ptr(amp->ahp, anon_index++);
7303 			if (ap != NULL) {
7304 				swap_xlate(ap, &vp, &off);
7305 			} else {
7306 				vp = svd->vp;
7307 				off = offset;
7308 			}
7309 			anon_array_exit(&cookie);
7310 			ANON_LOCK_EXIT(&amp->a_rwlock);
7311 		} else {
7312 			vp = svd->vp;
7313 			off = offset;
7314 		}
7315 		offset += PAGESIZE;
7316 
7317 		if (vp == NULL)		/* untouched zfod page */
7318 			continue;
7319 
7320 		if (attr) {
7321 			if (vpp) {
7322 				prot = VPP_PROT(vpp);
7323 				vpp++;
7324 			}
7325 			if (prot != pageprot) {
7326 				continue;
7327 			}
7328 		}
7329 
7330 		/*
7331 		 * See if any of these pages are locked --  if so, then we
7332 		 * will have to truncate an invalidate request at the first
7333 		 * locked one. We don't need the page_struct_lock to test
7334 		 * as this is only advisory; even if we acquire it someone
7335 		 * might race in and lock the page after we unlock and before
7336 		 * we do the PUTPAGE, then PUTPAGE simply does nothing.
7337 		 */
7338 		if (flags & MS_INVALIDATE) {
7339 			if ((pp = page_lookup(vp, off, SE_SHARED)) != NULL) {
7340 				if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
7341 					page_unlock(pp);
7342 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7343 					return (EBUSY);
7344 				}
7345 				if (ap != NULL && pp->p_szc != 0 &&
7346 				    page_tryupgrade(pp)) {
7347 					if (pp->p_lckcnt == 0 &&
7348 					    pp->p_cowcnt == 0) {
7349 						/*
7350 						 * swapfs VN_DISPOSE() won't
7351 						 * invalidate large pages.
7352 						 * Attempt to demote.
7353 						 * XXX can't help it if it
7354 						 * fails. But for swapfs
7355 						 * pages it is no big deal.
7356 						 */
7357 						(void) page_try_demote_pages(
7358 						    pp);
7359 					}
7360 				}
7361 				page_unlock(pp);
7362 			}
7363 		} else if (svd->type == MAP_SHARED && amp != NULL) {
7364 			/*
7365 			 * Avoid writing out to disk ISM's large pages
7366 			 * because segspt_free_pages() relies on NULL an_pvp
7367 			 * of anon slots of such pages.
7368 			 */
7369 
7370 			ASSERT(svd->vp == NULL);
7371 			/*
7372 			 * swapfs uses page_lookup_nowait if not freeing or
7373 			 * invalidating and skips a page if
7374 			 * page_lookup_nowait returns NULL.
7375 			 */
7376 			pp = page_lookup_nowait(vp, off, SE_SHARED);
7377 			if (pp == NULL) {
7378 				continue;
7379 			}
7380 			if (pp->p_szc != 0) {
7381 				page_unlock(pp);
7382 				continue;
7383 			}
7384 
7385 			/*
7386 			 * Note ISM pages are created large so (vp, off)'s
7387 			 * page cannot suddenly become large after we unlock
7388 			 * pp.
7389 			 */
7390 			page_unlock(pp);
7391 		}
7392 		/*
7393 		 * XXX - Should ultimately try to kluster
7394 		 * calls to VOP_PUTPAGE() for performance.
7395 		 */
7396 		VN_HOLD(vp);
7397 		err = VOP_PUTPAGE(vp, (offset_t)off, PAGESIZE,
7398 		    (bflags | (IS_SWAPFSVP(vp) ? B_PAGE_NOWAIT : 0)),
7399 		    svd->cred, NULL);
7400 
7401 		VN_RELE(vp);
7402 		if (err)
7403 			break;
7404 	}
7405 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7406 	return (err);
7407 }
7408 
7409 /*
7410  * Determine if we have data corresponding to pages in the
7411  * primary storage virtual memory cache (i.e., "in core").
7412  */
7413 static size_t
7414 segvn_incore(struct seg *seg, caddr_t addr, size_t len, char *vec)
7415 {
7416 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7417 	struct vnode *vp, *avp;
7418 	u_offset_t offset, aoffset;
7419 	size_t p, ep;
7420 	int ret;
7421 	struct vpage *vpp;
7422 	page_t *pp;
7423 	uint_t start;
7424 	struct anon_map *amp;		/* XXX - for locknest */
7425 	struct anon *ap;
7426 	uint_t attr;
7427 	anon_sync_obj_t cookie;
7428 
7429 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7430 
7431 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7432 	if (svd->amp == NULL && svd->vp == NULL) {
7433 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7434 		bzero(vec, btopr(len));
7435 		return (len);	/* no anonymous pages created yet */
7436 	}
7437 
7438 	p = seg_page(seg, addr);
7439 	ep = seg_page(seg, addr + len);
7440 	start = svd->vp ? SEG_PAGE_VNODEBACKED : 0;
7441 
7442 	amp = svd->amp;
7443 	for (; p < ep; p++, addr += PAGESIZE) {
7444 		vpp = (svd->vpage) ? &svd->vpage[p]: NULL;
7445 		ret = start;
7446 		ap = NULL;
7447 		avp = NULL;
7448 		/* Grab the vnode/offset for the anon slot */
7449 		if (amp != NULL) {
7450 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7451 			anon_array_enter(amp, svd->anon_index + p, &cookie);
7452 			ap = anon_get_ptr(amp->ahp, svd->anon_index + p);
7453 			if (ap != NULL) {
7454 				swap_xlate(ap, &avp, &aoffset);
7455 			}
7456 			anon_array_exit(&cookie);
7457 			ANON_LOCK_EXIT(&amp->a_rwlock);
7458 		}
7459 		if ((avp != NULL) && page_exists(avp, aoffset)) {
7460 			/* A page exists for the anon slot */
7461 			ret |= SEG_PAGE_INCORE;
7462 
7463 			/*
7464 			 * If page is mapped and writable
7465 			 */
7466 			attr = (uint_t)0;
7467 			if ((hat_getattr(seg->s_as->a_hat, addr,
7468 			    &attr) != -1) && (attr & PROT_WRITE)) {
7469 				ret |= SEG_PAGE_ANON;
7470 			}
7471 			/*
7472 			 * Don't get page_struct lock for lckcnt and cowcnt,
7473 			 * since this is purely advisory.
7474 			 */
7475 			if ((pp = page_lookup_nowait(avp, aoffset,
7476 			    SE_SHARED)) != NULL) {
7477 				if (pp->p_lckcnt)
7478 					ret |= SEG_PAGE_SOFTLOCK;
7479 				if (pp->p_cowcnt)
7480 					ret |= SEG_PAGE_HASCOW;
7481 				page_unlock(pp);
7482 			}
7483 		}
7484 
7485 		/* Gather vnode statistics */
7486 		vp = svd->vp;
7487 		offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7488 
7489 		if (vp != NULL) {
7490 			/*
7491 			 * Try to obtain a "shared" lock on the page
7492 			 * without blocking.  If this fails, determine
7493 			 * if the page is in memory.
7494 			 */
7495 			pp = page_lookup_nowait(vp, offset, SE_SHARED);
7496 			if ((pp == NULL) && (page_exists(vp, offset))) {
7497 				/* Page is incore, and is named */
7498 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7499 			}
7500 			/*
7501 			 * Don't get page_struct lock for lckcnt and cowcnt,
7502 			 * since this is purely advisory.
7503 			 */
7504 			if (pp != NULL) {
7505 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7506 				if (pp->p_lckcnt)
7507 					ret |= SEG_PAGE_SOFTLOCK;
7508 				if (pp->p_cowcnt)
7509 					ret |= SEG_PAGE_HASCOW;
7510 				page_unlock(pp);
7511 			}
7512 		}
7513 
7514 		/* Gather virtual page information */
7515 		if (vpp) {
7516 			if (VPP_ISPPLOCK(vpp))
7517 				ret |= SEG_PAGE_LOCKED;
7518 			vpp++;
7519 		}
7520 
7521 		*vec++ = (char)ret;
7522 	}
7523 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7524 	return (len);
7525 }
7526 
7527 /*
7528  * Statement for p_cowcnts/p_lckcnts.
7529  *
7530  * p_cowcnt is updated while mlock/munlocking MAP_PRIVATE and PROT_WRITE region
7531  * irrespective of the following factors or anything else:
7532  *
7533  *	(1) anon slots are populated or not
7534  *	(2) cow is broken or not
7535  *	(3) refcnt on ap is 1 or greater than 1
7536  *
7537  * If it's not MAP_PRIVATE and PROT_WRITE, p_lckcnt is updated during mlock
7538  * and munlock.
7539  *
7540  *
7541  * Handling p_cowcnts/p_lckcnts during copy-on-write fault:
7542  *
7543  *	if vpage has PROT_WRITE
7544  *		transfer cowcnt on the oldpage -> cowcnt on the newpage
7545  *	else
7546  *		transfer lckcnt on the oldpage -> lckcnt on the newpage
7547  *
7548  *	During copy-on-write, decrement p_cowcnt on the oldpage and increment
7549  *	p_cowcnt on the newpage *if* the corresponding vpage has PROT_WRITE.
7550  *
7551  *	We may also break COW if softlocking on read access in the physio case.
7552  *	In this case, vpage may not have PROT_WRITE. So, we need to decrement
7553  *	p_lckcnt on the oldpage and increment p_lckcnt on the newpage *if* the
7554  *	vpage doesn't have PROT_WRITE.
7555  *
7556  *
7557  * Handling p_cowcnts/p_lckcnts during mprotect on mlocked region:
7558  *
7559  * 	If a MAP_PRIVATE region loses PROT_WRITE, we decrement p_cowcnt and
7560  *	increment p_lckcnt by calling page_subclaim() which takes care of
7561  * 	availrmem accounting and p_lckcnt overflow.
7562  *
7563  *	If a MAP_PRIVATE region gains PROT_WRITE, we decrement p_lckcnt and
7564  *	increment p_cowcnt by calling page_addclaim() which takes care of
7565  *	availrmem availability and p_cowcnt overflow.
7566  */
7567 
7568 /*
7569  * Lock down (or unlock) pages mapped by this segment.
7570  *
7571  * XXX only creates PAGESIZE pages if anon slots are not initialized.
7572  * At fault time they will be relocated into larger pages.
7573  */
7574 static int
7575 segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
7576     int attr, int op, ulong_t *lockmap, size_t pos)
7577 {
7578 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7579 	struct vpage *vpp;
7580 	struct vpage *evp;
7581 	page_t *pp;
7582 	u_offset_t offset;
7583 	u_offset_t off;
7584 	int segtype;
7585 	int pageprot;
7586 	int claim;
7587 	struct vnode *vp;
7588 	ulong_t anon_index;
7589 	struct anon_map *amp;
7590 	struct anon *ap;
7591 	struct vattr va;
7592 	anon_sync_obj_t cookie;
7593 	struct kshmid *sp = NULL;
7594 	struct proc	*p = curproc;
7595 	kproject_t	*proj = NULL;
7596 	int chargeproc = 1;
7597 	size_t locked_bytes = 0;
7598 	size_t unlocked_bytes = 0;
7599 	int err = 0;
7600 
7601 	/*
7602 	 * Hold write lock on address space because may split or concatenate
7603 	 * segments
7604 	 */
7605 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7606 
7607 	/*
7608 	 * If this is a shm, use shm's project and zone, else use
7609 	 * project and zone of calling process
7610 	 */
7611 
7612 	/* Determine if this segment backs a sysV shm */
7613 	if (svd->amp != NULL && svd->amp->a_sp != NULL) {
7614 		ASSERT(svd->type == MAP_SHARED);
7615 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
7616 		sp = svd->amp->a_sp;
7617 		proj = sp->shm_perm.ipc_proj;
7618 		chargeproc = 0;
7619 	}
7620 
7621 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7622 	if (attr) {
7623 		pageprot = attr & ~(SHARED|PRIVATE);
7624 		segtype = attr & SHARED ? MAP_SHARED : MAP_PRIVATE;
7625 
7626 		/*
7627 		 * We are done if the segment types don't match
7628 		 * or if we have segment level protections and
7629 		 * they don't match.
7630 		 */
7631 		if (svd->type != segtype) {
7632 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7633 			return (0);
7634 		}
7635 		if (svd->pageprot == 0 && svd->prot != pageprot) {
7636 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7637 			return (0);
7638 		}
7639 	}
7640 
7641 	if (op == MC_LOCK) {
7642 		if (svd->tr_state == SEGVN_TR_INIT) {
7643 			svd->tr_state = SEGVN_TR_OFF;
7644 		} else if (svd->tr_state == SEGVN_TR_ON) {
7645 			ASSERT(svd->amp != NULL);
7646 			segvn_textunrepl(seg, 0);
7647 			ASSERT(svd->amp == NULL &&
7648 			    svd->tr_state == SEGVN_TR_OFF);
7649 		}
7650 	}
7651 
7652 	/*
7653 	 * If we're locking, then we must create a vpage structure if
7654 	 * none exists.  If we're unlocking, then check to see if there
7655 	 * is a vpage --  if not, then we could not have locked anything.
7656 	 */
7657 
7658 	if ((vpp = svd->vpage) == NULL) {
7659 		if (op == MC_LOCK)
7660 			segvn_vpage(seg);
7661 		else {
7662 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7663 			return (0);
7664 		}
7665 	}
7666 
7667 	/*
7668 	 * The anonymous data vector (i.e., previously
7669 	 * unreferenced mapping to swap space) can be allocated
7670 	 * by lazily testing for its existence.
7671 	 */
7672 	if (op == MC_LOCK && svd->amp == NULL && svd->vp == NULL) {
7673 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
7674 		svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
7675 		svd->amp->a_szc = seg->s_szc;
7676 	}
7677 
7678 	if ((amp = svd->amp) != NULL) {
7679 		anon_index = svd->anon_index + seg_page(seg, addr);
7680 	}
7681 
7682 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7683 	evp = &svd->vpage[seg_page(seg, addr + len)];
7684 
7685 	if (sp != NULL)
7686 		mutex_enter(&sp->shm_mlock);
7687 
7688 	/* determine number of unlocked bytes in range for lock operation */
7689 	if (op == MC_LOCK) {
7690 
7691 		if (sp == NULL) {
7692 			for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7693 			    vpp++) {
7694 				if (!VPP_ISPPLOCK(vpp))
7695 					unlocked_bytes += PAGESIZE;
7696 			}
7697 		} else {
7698 			ulong_t		i_idx, i_edx;
7699 			anon_sync_obj_t	i_cookie;
7700 			struct anon	*i_ap;
7701 			struct vnode	*i_vp;
7702 			u_offset_t	i_off;
7703 
7704 			/* Only count sysV pages once for locked memory */
7705 			i_edx = svd->anon_index + seg_page(seg, addr + len);
7706 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7707 			for (i_idx = anon_index; i_idx < i_edx; i_idx++) {
7708 				anon_array_enter(amp, i_idx, &i_cookie);
7709 				i_ap = anon_get_ptr(amp->ahp, i_idx);
7710 				if (i_ap == NULL) {
7711 					unlocked_bytes += PAGESIZE;
7712 					anon_array_exit(&i_cookie);
7713 					continue;
7714 				}
7715 				swap_xlate(i_ap, &i_vp, &i_off);
7716 				anon_array_exit(&i_cookie);
7717 				pp = page_lookup(i_vp, i_off, SE_SHARED);
7718 				if (pp == NULL) {
7719 					unlocked_bytes += PAGESIZE;
7720 					continue;
7721 				} else if (pp->p_lckcnt == 0)
7722 					unlocked_bytes += PAGESIZE;
7723 				page_unlock(pp);
7724 			}
7725 			ANON_LOCK_EXIT(&amp->a_rwlock);
7726 		}
7727 
7728 		mutex_enter(&p->p_lock);
7729 		err = rctl_incr_locked_mem(p, proj, unlocked_bytes,
7730 		    chargeproc);
7731 		mutex_exit(&p->p_lock);
7732 
7733 		if (err) {
7734 			if (sp != NULL)
7735 				mutex_exit(&sp->shm_mlock);
7736 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7737 			return (err);
7738 		}
7739 	}
7740 	/*
7741 	 * Loop over all pages in the range.  Process if we're locking and
7742 	 * page has not already been locked in this mapping; or if we're
7743 	 * unlocking and the page has been locked.
7744 	 */
7745 	for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7746 	    vpp++, pos++, addr += PAGESIZE, offset += PAGESIZE, anon_index++) {
7747 		if ((attr == 0 || VPP_PROT(vpp) == pageprot) &&
7748 		    ((op == MC_LOCK && !VPP_ISPPLOCK(vpp)) ||
7749 		    (op == MC_UNLOCK && VPP_ISPPLOCK(vpp)))) {
7750 
7751 			if (amp != NULL)
7752 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7753 			/*
7754 			 * If this isn't a MAP_NORESERVE segment and
7755 			 * we're locking, allocate anon slots if they
7756 			 * don't exist.  The page is brought in later on.
7757 			 */
7758 			if (op == MC_LOCK && svd->vp == NULL &&
7759 			    ((svd->flags & MAP_NORESERVE) == 0) &&
7760 			    amp != NULL &&
7761 			    ((ap = anon_get_ptr(amp->ahp, anon_index))
7762 			    == NULL)) {
7763 				anon_array_enter(amp, anon_index, &cookie);
7764 
7765 				if ((ap = anon_get_ptr(amp->ahp,
7766 				    anon_index)) == NULL) {
7767 					pp = anon_zero(seg, addr, &ap,
7768 					    svd->cred);
7769 					if (pp == NULL) {
7770 						anon_array_exit(&cookie);
7771 						ANON_LOCK_EXIT(&amp->a_rwlock);
7772 						err = ENOMEM;
7773 						goto out;
7774 					}
7775 					ASSERT(anon_get_ptr(amp->ahp,
7776 					    anon_index) == NULL);
7777 					(void) anon_set_ptr(amp->ahp,
7778 					    anon_index, ap, ANON_SLEEP);
7779 					page_unlock(pp);
7780 				}
7781 				anon_array_exit(&cookie);
7782 			}
7783 
7784 			/*
7785 			 * Get name for page, accounting for
7786 			 * existence of private copy.
7787 			 */
7788 			ap = NULL;
7789 			if (amp != NULL) {
7790 				anon_array_enter(amp, anon_index, &cookie);
7791 				ap = anon_get_ptr(amp->ahp, anon_index);
7792 				if (ap != NULL) {
7793 					swap_xlate(ap, &vp, &off);
7794 				} else {
7795 					if (svd->vp == NULL &&
7796 					    (svd->flags & MAP_NORESERVE)) {
7797 						anon_array_exit(&cookie);
7798 						ANON_LOCK_EXIT(&amp->a_rwlock);
7799 						continue;
7800 					}
7801 					vp = svd->vp;
7802 					off = offset;
7803 				}
7804 				if (op != MC_LOCK || ap == NULL) {
7805 					anon_array_exit(&cookie);
7806 					ANON_LOCK_EXIT(&amp->a_rwlock);
7807 				}
7808 			} else {
7809 				vp = svd->vp;
7810 				off = offset;
7811 			}
7812 
7813 			/*
7814 			 * Get page frame.  It's ok if the page is
7815 			 * not available when we're unlocking, as this
7816 			 * may simply mean that a page we locked got
7817 			 * truncated out of existence after we locked it.
7818 			 *
7819 			 * Invoke VOP_GETPAGE() to obtain the page struct
7820 			 * since we may need to read it from disk if its
7821 			 * been paged out.
7822 			 */
7823 			if (op != MC_LOCK)
7824 				pp = page_lookup(vp, off, SE_SHARED);
7825 			else {
7826 				page_t *pl[1 + 1];
7827 				int error;
7828 
7829 				ASSERT(vp != NULL);
7830 
7831 				error = VOP_GETPAGE(vp, (offset_t)off, PAGESIZE,
7832 				    (uint_t *)NULL, pl, PAGESIZE, seg, addr,
7833 				    S_OTHER, svd->cred, NULL);
7834 
7835 				if (error && ap != NULL) {
7836 					anon_array_exit(&cookie);
7837 					ANON_LOCK_EXIT(&amp->a_rwlock);
7838 				}
7839 
7840 				/*
7841 				 * If the error is EDEADLK then we must bounce
7842 				 * up and drop all vm subsystem locks and then
7843 				 * retry the operation later
7844 				 * This behavior is a temporary measure because
7845 				 * ufs/sds logging is badly designed and will
7846 				 * deadlock if we don't allow this bounce to
7847 				 * happen.  The real solution is to re-design
7848 				 * the logging code to work properly.  See bug
7849 				 * 4125102 for details of the problem.
7850 				 */
7851 				if (error == EDEADLK) {
7852 					err = error;
7853 					goto out;
7854 				}
7855 				/*
7856 				 * Quit if we fail to fault in the page.  Treat
7857 				 * the failure as an error, unless the addr
7858 				 * is mapped beyond the end of a file.
7859 				 */
7860 				if (error && svd->vp) {
7861 					va.va_mask = AT_SIZE;
7862 					if (VOP_GETATTR(svd->vp, &va, 0,
7863 					    svd->cred, NULL) != 0) {
7864 						err = EIO;
7865 						goto out;
7866 					}
7867 					if (btopr(va.va_size) >=
7868 					    btopr(off + 1)) {
7869 						err = EIO;
7870 						goto out;
7871 					}
7872 					goto out;
7873 
7874 				} else if (error) {
7875 					err = EIO;
7876 					goto out;
7877 				}
7878 				pp = pl[0];
7879 				ASSERT(pp != NULL);
7880 			}
7881 
7882 			/*
7883 			 * See Statement at the beginning of this routine.
7884 			 *
7885 			 * claim is always set if MAP_PRIVATE and PROT_WRITE
7886 			 * irrespective of following factors:
7887 			 *
7888 			 * (1) anon slots are populated or not
7889 			 * (2) cow is broken or not
7890 			 * (3) refcnt on ap is 1 or greater than 1
7891 			 *
7892 			 * See 4140683 for details
7893 			 */
7894 			claim = ((VPP_PROT(vpp) & PROT_WRITE) &&
7895 			    (svd->type == MAP_PRIVATE));
7896 
7897 			/*
7898 			 * Perform page-level operation appropriate to
7899 			 * operation.  If locking, undo the SOFTLOCK
7900 			 * performed to bring the page into memory
7901 			 * after setting the lock.  If unlocking,
7902 			 * and no page was found, account for the claim
7903 			 * separately.
7904 			 */
7905 			if (op == MC_LOCK) {
7906 				int ret = 1;	/* Assume success */
7907 
7908 				ASSERT(!VPP_ISPPLOCK(vpp));
7909 
7910 				ret = page_pp_lock(pp, claim, 0);
7911 				if (ap != NULL) {
7912 					if (ap->an_pvp != NULL) {
7913 						anon_swap_free(ap, pp);
7914 					}
7915 					anon_array_exit(&cookie);
7916 					ANON_LOCK_EXIT(&amp->a_rwlock);
7917 				}
7918 				if (ret == 0) {
7919 					/* locking page failed */
7920 					page_unlock(pp);
7921 					err = EAGAIN;
7922 					goto out;
7923 				}
7924 				VPP_SETPPLOCK(vpp);
7925 				if (sp != NULL) {
7926 					if (pp->p_lckcnt == 1)
7927 						locked_bytes += PAGESIZE;
7928 				} else
7929 					locked_bytes += PAGESIZE;
7930 
7931 				if (lockmap != (ulong_t *)NULL)
7932 					BT_SET(lockmap, pos);
7933 
7934 				page_unlock(pp);
7935 			} else {
7936 				ASSERT(VPP_ISPPLOCK(vpp));
7937 				if (pp != NULL) {
7938 					/* sysV pages should be locked */
7939 					ASSERT(sp == NULL || pp->p_lckcnt > 0);
7940 					page_pp_unlock(pp, claim, 0);
7941 					if (sp != NULL) {
7942 						if (pp->p_lckcnt == 0)
7943 							unlocked_bytes
7944 							    += PAGESIZE;
7945 					} else
7946 						unlocked_bytes += PAGESIZE;
7947 					page_unlock(pp);
7948 				} else {
7949 					ASSERT(sp == NULL);
7950 					unlocked_bytes += PAGESIZE;
7951 				}
7952 				VPP_CLRPPLOCK(vpp);
7953 			}
7954 		}
7955 	}
7956 out:
7957 	if (op == MC_LOCK) {
7958 		/* Credit back bytes that did not get locked */
7959 		if ((unlocked_bytes - locked_bytes) > 0) {
7960 			if (proj == NULL)
7961 				mutex_enter(&p->p_lock);
7962 			rctl_decr_locked_mem(p, proj,
7963 			    (unlocked_bytes - locked_bytes), chargeproc);
7964 			if (proj == NULL)
7965 				mutex_exit(&p->p_lock);
7966 		}
7967 
7968 	} else {
7969 		/* Account bytes that were unlocked */
7970 		if (unlocked_bytes > 0) {
7971 			if (proj == NULL)
7972 				mutex_enter(&p->p_lock);
7973 			rctl_decr_locked_mem(p, proj, unlocked_bytes,
7974 			    chargeproc);
7975 			if (proj == NULL)
7976 				mutex_exit(&p->p_lock);
7977 		}
7978 	}
7979 	if (sp != NULL)
7980 		mutex_exit(&sp->shm_mlock);
7981 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7982 
7983 	return (err);
7984 }
7985 
7986 /*
7987  * Set advice from user for specified pages
7988  * There are 5 types of advice:
7989  *	MADV_NORMAL	- Normal (default) behavior (whatever that is)
7990  *	MADV_RANDOM	- Random page references
7991  *				do not allow readahead or 'klustering'
7992  *	MADV_SEQUENTIAL	- Sequential page references
7993  *				Pages previous to the one currently being
7994  *				accessed (determined by fault) are 'not needed'
7995  *				and are freed immediately
7996  *	MADV_WILLNEED	- Pages are likely to be used (fault ahead in mctl)
7997  *	MADV_DONTNEED	- Pages are not needed (synced out in mctl)
7998  *	MADV_FREE	- Contents can be discarded
7999  *	MADV_ACCESS_DEFAULT- Default access
8000  *	MADV_ACCESS_LWP	- Next LWP will access heavily
8001  *	MADV_ACCESS_MANY- Many LWPs or processes will access heavily
8002  */
8003 static int
8004 segvn_advise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
8005 {
8006 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8007 	size_t page;
8008 	int err = 0;
8009 	int already_set;
8010 	struct anon_map *amp;
8011 	ulong_t anon_index;
8012 	struct seg *next;
8013 	lgrp_mem_policy_t policy;
8014 	struct seg *prev;
8015 	struct vnode *vp;
8016 
8017 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
8018 
8019 	/*
8020 	 * In case of MADV_FREE, we won't be modifying any segment private
8021 	 * data structures; so, we only need to grab READER's lock
8022 	 */
8023 	if (behav != MADV_FREE) {
8024 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
8025 		if (svd->tr_state != SEGVN_TR_OFF) {
8026 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8027 			return (0);
8028 		}
8029 	} else {
8030 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8031 	}
8032 
8033 	/*
8034 	 * Large pages are assumed to be only turned on when accesses to the
8035 	 * segment's address range have spatial and temporal locality. That
8036 	 * justifies ignoring MADV_SEQUENTIAL for large page segments.
8037 	 * Also, ignore advice affecting lgroup memory allocation
8038 	 * if don't need to do lgroup optimizations on this system
8039 	 */
8040 
8041 	if ((behav == MADV_SEQUENTIAL &&
8042 	    (seg->s_szc != 0 || HAT_IS_REGION_COOKIE_VALID(svd->rcookie))) ||
8043 	    (!lgrp_optimizations() && (behav == MADV_ACCESS_DEFAULT ||
8044 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY))) {
8045 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8046 		return (0);
8047 	}
8048 
8049 	if (behav == MADV_SEQUENTIAL || behav == MADV_ACCESS_DEFAULT ||
8050 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY) {
8051 		/*
8052 		 * Since we are going to unload hat mappings
8053 		 * we first have to flush the cache. Otherwise
8054 		 * this might lead to system panic if another
8055 		 * thread is doing physio on the range whose
8056 		 * mappings are unloaded by madvise(3C).
8057 		 */
8058 		if (svd->softlockcnt > 0) {
8059 			/*
8060 			 * If this is shared segment non 0 softlockcnt
8061 			 * means locked pages are still in use.
8062 			 */
8063 			if (svd->type == MAP_SHARED) {
8064 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8065 				return (EAGAIN);
8066 			}
8067 			/*
8068 			 * Since we do have the segvn writers lock
8069 			 * nobody can fill the cache with entries
8070 			 * belonging to this seg during the purge.
8071 			 * The flush either succeeds or we still
8072 			 * have pending I/Os. In the later case,
8073 			 * madvise(3C) fails.
8074 			 */
8075 			segvn_purge(seg);
8076 			if (svd->softlockcnt > 0) {
8077 				/*
8078 				 * Since madvise(3C) is advisory and
8079 				 * it's not part of UNIX98, madvise(3C)
8080 				 * failure here doesn't cause any hardship.
8081 				 * Note that we don't block in "as" layer.
8082 				 */
8083 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8084 				return (EAGAIN);
8085 			}
8086 		} else if (svd->type == MAP_SHARED && svd->amp != NULL &&
8087 		    svd->amp->a_softlockcnt > 0) {
8088 			/*
8089 			 * Try to purge this amp's entries from pcache. It
8090 			 * will succeed only if other segments that share the
8091 			 * amp have no outstanding softlock's.
8092 			 */
8093 			segvn_purge(seg);
8094 		}
8095 	}
8096 
8097 	amp = svd->amp;
8098 	vp = svd->vp;
8099 	if (behav == MADV_FREE) {
8100 		/*
8101 		 * MADV_FREE is not supported for segments with
8102 		 * underlying object; if anonmap is NULL, anon slots
8103 		 * are not yet populated and there is nothing for
8104 		 * us to do. As MADV_FREE is advisory, we don't
8105 		 * return error in either case.
8106 		 */
8107 		if (vp != NULL || amp == NULL) {
8108 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8109 			return (0);
8110 		}
8111 
8112 		segvn_purge(seg);
8113 
8114 		page = seg_page(seg, addr);
8115 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8116 		anon_disclaim(amp, svd->anon_index + page, len);
8117 		ANON_LOCK_EXIT(&amp->a_rwlock);
8118 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8119 		return (0);
8120 	}
8121 
8122 	/*
8123 	 * If advice is to be applied to entire segment,
8124 	 * use advice field in seg_data structure
8125 	 * otherwise use appropriate vpage entry.
8126 	 */
8127 	if ((addr == seg->s_base) && (len == seg->s_size)) {
8128 		switch (behav) {
8129 		case MADV_ACCESS_LWP:
8130 		case MADV_ACCESS_MANY:
8131 		case MADV_ACCESS_DEFAULT:
8132 			/*
8133 			 * Set memory allocation policy for this segment
8134 			 */
8135 			policy = lgrp_madv_to_policy(behav, len, svd->type);
8136 			if (svd->type == MAP_SHARED)
8137 				already_set = lgrp_shm_policy_set(policy, amp,
8138 				    svd->anon_index, vp, svd->offset, len);
8139 			else {
8140 				/*
8141 				 * For private memory, need writers lock on
8142 				 * address space because the segment may be
8143 				 * split or concatenated when changing policy
8144 				 */
8145 				if (AS_READ_HELD(seg->s_as,
8146 				    &seg->s_as->a_lock)) {
8147 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8148 					return (IE_RETRY);
8149 				}
8150 
8151 				already_set = lgrp_privm_policy_set(policy,
8152 				    &svd->policy_info, len);
8153 			}
8154 
8155 			/*
8156 			 * If policy set already and it shouldn't be reapplied,
8157 			 * don't do anything.
8158 			 */
8159 			if (already_set &&
8160 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8161 				break;
8162 
8163 			/*
8164 			 * Mark any existing pages in given range for
8165 			 * migration
8166 			 */
8167 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8168 			    vp, svd->offset, 1);
8169 
8170 			/*
8171 			 * If same policy set already or this is a shared
8172 			 * memory segment, don't need to try to concatenate
8173 			 * segment with adjacent ones.
8174 			 */
8175 			if (already_set || svd->type == MAP_SHARED)
8176 				break;
8177 
8178 			/*
8179 			 * Try to concatenate this segment with previous
8180 			 * one and next one, since we changed policy for
8181 			 * this one and it may be compatible with adjacent
8182 			 * ones now.
8183 			 */
8184 			prev = AS_SEGPREV(seg->s_as, seg);
8185 			next = AS_SEGNEXT(seg->s_as, seg);
8186 
8187 			if (next && next->s_ops == &segvn_ops &&
8188 			    addr + len == next->s_base)
8189 				(void) segvn_concat(seg, next, 1);
8190 
8191 			if (prev && prev->s_ops == &segvn_ops &&
8192 			    addr == prev->s_base + prev->s_size) {
8193 				/*
8194 				 * Drop lock for private data of current
8195 				 * segment before concatenating (deleting) it
8196 				 * and return IE_REATTACH to tell as_ctl() that
8197 				 * current segment has changed
8198 				 */
8199 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8200 				if (!segvn_concat(prev, seg, 1))
8201 					err = IE_REATTACH;
8202 
8203 				return (err);
8204 			}
8205 			break;
8206 
8207 		case MADV_SEQUENTIAL:
8208 			/*
8209 			 * unloading mapping guarantees
8210 			 * detection in segvn_fault
8211 			 */
8212 			ASSERT(seg->s_szc == 0);
8213 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8214 			hat_unload(seg->s_as->a_hat, addr, len,
8215 			    HAT_UNLOAD);
8216 			/* FALLTHROUGH */
8217 		case MADV_NORMAL:
8218 		case MADV_RANDOM:
8219 			svd->advice = (uchar_t)behav;
8220 			svd->pageadvice = 0;
8221 			break;
8222 		case MADV_WILLNEED:	/* handled in memcntl */
8223 		case MADV_DONTNEED:	/* handled in memcntl */
8224 		case MADV_FREE:		/* handled above */
8225 			break;
8226 		default:
8227 			err = EINVAL;
8228 		}
8229 	} else {
8230 		caddr_t			eaddr;
8231 		struct seg		*new_seg;
8232 		struct segvn_data	*new_svd;
8233 		u_offset_t		off;
8234 		caddr_t			oldeaddr;
8235 
8236 		page = seg_page(seg, addr);
8237 
8238 		segvn_vpage(seg);
8239 
8240 		switch (behav) {
8241 			struct vpage *bvpp, *evpp;
8242 
8243 		case MADV_ACCESS_LWP:
8244 		case MADV_ACCESS_MANY:
8245 		case MADV_ACCESS_DEFAULT:
8246 			/*
8247 			 * Set memory allocation policy for portion of this
8248 			 * segment
8249 			 */
8250 
8251 			/*
8252 			 * Align address and length of advice to page
8253 			 * boundaries for large pages
8254 			 */
8255 			if (seg->s_szc != 0) {
8256 				size_t	pgsz;
8257 
8258 				pgsz = page_get_pagesize(seg->s_szc);
8259 				addr = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
8260 				len = P2ROUNDUP(len, pgsz);
8261 			}
8262 
8263 			/*
8264 			 * Check to see whether policy is set already
8265 			 */
8266 			policy = lgrp_madv_to_policy(behav, len, svd->type);
8267 
8268 			anon_index = svd->anon_index + page;
8269 			off = svd->offset + (uintptr_t)(addr - seg->s_base);
8270 
8271 			if (svd->type == MAP_SHARED)
8272 				already_set = lgrp_shm_policy_set(policy, amp,
8273 				    anon_index, vp, off, len);
8274 			else
8275 				already_set =
8276 				    (policy == svd->policy_info.mem_policy);
8277 
8278 			/*
8279 			 * If policy set already and it shouldn't be reapplied,
8280 			 * don't do anything.
8281 			 */
8282 			if (already_set &&
8283 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8284 				break;
8285 
8286 			/*
8287 			 * For private memory, need writers lock on
8288 			 * address space because the segment may be
8289 			 * split or concatenated when changing policy
8290 			 */
8291 			if (svd->type == MAP_PRIVATE &&
8292 			    AS_READ_HELD(seg->s_as, &seg->s_as->a_lock)) {
8293 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8294 				return (IE_RETRY);
8295 			}
8296 
8297 			/*
8298 			 * Mark any existing pages in given range for
8299 			 * migration
8300 			 */
8301 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8302 			    vp, svd->offset, 1);
8303 
8304 			/*
8305 			 * Don't need to try to split or concatenate
8306 			 * segments, since policy is same or this is a shared
8307 			 * memory segment
8308 			 */
8309 			if (already_set || svd->type == MAP_SHARED)
8310 				break;
8311 
8312 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
8313 				ASSERT(svd->amp == NULL);
8314 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
8315 				ASSERT(svd->softlockcnt == 0);
8316 				hat_leave_region(seg->s_as->a_hat, svd->rcookie,
8317 				    HAT_REGION_TEXT);
8318 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
8319 			}
8320 
8321 			/*
8322 			 * Split off new segment if advice only applies to a
8323 			 * portion of existing segment starting in middle
8324 			 */
8325 			new_seg = NULL;
8326 			eaddr = addr + len;
8327 			oldeaddr = seg->s_base + seg->s_size;
8328 			if (addr > seg->s_base) {
8329 				/*
8330 				 * Must flush I/O page cache
8331 				 * before splitting segment
8332 				 */
8333 				if (svd->softlockcnt > 0)
8334 					segvn_purge(seg);
8335 
8336 				/*
8337 				 * Split segment and return IE_REATTACH to tell
8338 				 * as_ctl() that current segment changed
8339 				 */
8340 				new_seg = segvn_split_seg(seg, addr);
8341 				new_svd = (struct segvn_data *)new_seg->s_data;
8342 				err = IE_REATTACH;
8343 
8344 				/*
8345 				 * If new segment ends where old one
8346 				 * did, try to concatenate the new
8347 				 * segment with next one.
8348 				 */
8349 				if (eaddr == oldeaddr) {
8350 					/*
8351 					 * Set policy for new segment
8352 					 */
8353 					(void) lgrp_privm_policy_set(policy,
8354 					    &new_svd->policy_info,
8355 					    new_seg->s_size);
8356 
8357 					next = AS_SEGNEXT(new_seg->s_as,
8358 					    new_seg);
8359 
8360 					if (next &&
8361 					    next->s_ops == &segvn_ops &&
8362 					    eaddr == next->s_base)
8363 						(void) segvn_concat(new_seg,
8364 						    next, 1);
8365 				}
8366 			}
8367 
8368 			/*
8369 			 * Split off end of existing segment if advice only
8370 			 * applies to a portion of segment ending before
8371 			 * end of the existing segment
8372 			 */
8373 			if (eaddr < oldeaddr) {
8374 				/*
8375 				 * Must flush I/O page cache
8376 				 * before splitting segment
8377 				 */
8378 				if (svd->softlockcnt > 0)
8379 					segvn_purge(seg);
8380 
8381 				/*
8382 				 * If beginning of old segment was already
8383 				 * split off, use new segment to split end off
8384 				 * from.
8385 				 */
8386 				if (new_seg != NULL && new_seg != seg) {
8387 					/*
8388 					 * Split segment
8389 					 */
8390 					(void) segvn_split_seg(new_seg, eaddr);
8391 
8392 					/*
8393 					 * Set policy for new segment
8394 					 */
8395 					(void) lgrp_privm_policy_set(policy,
8396 					    &new_svd->policy_info,
8397 					    new_seg->s_size);
8398 				} else {
8399 					/*
8400 					 * Split segment and return IE_REATTACH
8401 					 * to tell as_ctl() that current
8402 					 * segment changed
8403 					 */
8404 					(void) segvn_split_seg(seg, eaddr);
8405 					err = IE_REATTACH;
8406 
8407 					(void) lgrp_privm_policy_set(policy,
8408 					    &svd->policy_info, seg->s_size);
8409 
8410 					/*
8411 					 * If new segment starts where old one
8412 					 * did, try to concatenate it with
8413 					 * previous segment.
8414 					 */
8415 					if (addr == seg->s_base) {
8416 						prev = AS_SEGPREV(seg->s_as,
8417 						    seg);
8418 
8419 						/*
8420 						 * Drop lock for private data
8421 						 * of current segment before
8422 						 * concatenating (deleting) it
8423 						 */
8424 						if (prev &&
8425 						    prev->s_ops ==
8426 						    &segvn_ops &&
8427 						    addr == prev->s_base +
8428 						    prev->s_size) {
8429 							SEGVN_LOCK_EXIT(
8430 							    seg->s_as,
8431 							    &svd->lock);
8432 							(void) segvn_concat(
8433 							    prev, seg, 1);
8434 							return (err);
8435 						}
8436 					}
8437 				}
8438 			}
8439 			break;
8440 		case MADV_SEQUENTIAL:
8441 			ASSERT(seg->s_szc == 0);
8442 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8443 			hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
8444 			/* FALLTHROUGH */
8445 		case MADV_NORMAL:
8446 		case MADV_RANDOM:
8447 			bvpp = &svd->vpage[page];
8448 			evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8449 			for (; bvpp < evpp; bvpp++)
8450 				VPP_SETADVICE(bvpp, behav);
8451 			svd->advice = MADV_NORMAL;
8452 			break;
8453 		case MADV_WILLNEED:	/* handled in memcntl */
8454 		case MADV_DONTNEED:	/* handled in memcntl */
8455 		case MADV_FREE:		/* handled above */
8456 			break;
8457 		default:
8458 			err = EINVAL;
8459 		}
8460 	}
8461 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8462 	return (err);
8463 }
8464 
8465 /*
8466  * Create a vpage structure for this seg.
8467  */
8468 static void
8469 segvn_vpage(struct seg *seg)
8470 {
8471 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8472 	struct vpage *vp, *evp;
8473 
8474 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
8475 
8476 	/*
8477 	 * If no vpage structure exists, allocate one.  Copy the protections
8478 	 * and the advice from the segment itself to the individual pages.
8479 	 */
8480 	if (svd->vpage == NULL) {
8481 		svd->pageadvice = 1;
8482 		svd->vpage = kmem_zalloc(seg_pages(seg) * sizeof (struct vpage),
8483 		    KM_SLEEP);
8484 		evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
8485 		for (vp = svd->vpage; vp < evp; vp++) {
8486 			VPP_SETPROT(vp, svd->prot);
8487 			VPP_SETADVICE(vp, svd->advice);
8488 		}
8489 	}
8490 }
8491 
8492 /*
8493  * Dump the pages belonging to this segvn segment.
8494  */
8495 static void
8496 segvn_dump(struct seg *seg)
8497 {
8498 	struct segvn_data *svd;
8499 	page_t *pp;
8500 	struct anon_map *amp;
8501 	ulong_t	anon_index;
8502 	struct vnode *vp;
8503 	u_offset_t off, offset;
8504 	pfn_t pfn;
8505 	pgcnt_t page, npages;
8506 	caddr_t addr;
8507 
8508 	npages = seg_pages(seg);
8509 	svd = (struct segvn_data *)seg->s_data;
8510 	vp = svd->vp;
8511 	off = offset = svd->offset;
8512 	addr = seg->s_base;
8513 
8514 	if ((amp = svd->amp) != NULL) {
8515 		anon_index = svd->anon_index;
8516 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8517 	}
8518 
8519 	for (page = 0; page < npages; page++, offset += PAGESIZE) {
8520 		struct anon *ap;
8521 		int we_own_it = 0;
8522 
8523 		if (amp && (ap = anon_get_ptr(svd->amp->ahp, anon_index++))) {
8524 			swap_xlate_nopanic(ap, &vp, &off);
8525 		} else {
8526 			vp = svd->vp;
8527 			off = offset;
8528 		}
8529 
8530 		/*
8531 		 * If pp == NULL, the page either does not exist
8532 		 * or is exclusively locked.  So determine if it
8533 		 * exists before searching for it.
8534 		 */
8535 
8536 		if ((pp = page_lookup_nowait(vp, off, SE_SHARED)))
8537 			we_own_it = 1;
8538 		else
8539 			pp = page_exists(vp, off);
8540 
8541 		if (pp) {
8542 			pfn = page_pptonum(pp);
8543 			dump_addpage(seg->s_as, addr, pfn);
8544 			if (we_own_it)
8545 				page_unlock(pp);
8546 		}
8547 		addr += PAGESIZE;
8548 		dump_timeleft = dump_timeout;
8549 	}
8550 
8551 	if (amp != NULL)
8552 		ANON_LOCK_EXIT(&amp->a_rwlock);
8553 }
8554 
8555 #ifdef DEBUG
8556 static uint32_t segvn_pglock_mtbf = 0;
8557 #endif
8558 
8559 #define	PCACHE_SHWLIST		((page_t *)-2)
8560 #define	NOPCACHE_SHWLIST	((page_t *)-1)
8561 
8562 /*
8563  * Lock/Unlock anon pages over a given range. Return shadow list. This routine
8564  * uses global segment pcache to cache shadow lists (i.e. pp arrays) of pages
8565  * to avoid the overhead of per page locking, unlocking for subsequent IOs to
8566  * the same parts of the segment. Currently shadow list creation is only
8567  * supported for pure anon segments. MAP_PRIVATE segment pcache entries are
8568  * tagged with segment pointer, starting virtual address and length. This
8569  * approach for MAP_SHARED segments may add many pcache entries for the same
8570  * set of pages and lead to long hash chains that decrease pcache lookup
8571  * performance. To avoid this issue for shared segments shared anon map and
8572  * starting anon index are used for pcache entry tagging. This allows all
8573  * segments to share pcache entries for the same anon range and reduces pcache
8574  * chain's length as well as memory overhead from duplicate shadow lists and
8575  * pcache entries.
8576  *
8577  * softlockcnt field in segvn_data structure counts the number of F_SOFTLOCK'd
8578  * pages via segvn_fault() and pagelock'd pages via this routine. But pagelock
8579  * part of softlockcnt accounting is done differently for private and shared
8580  * segments. In private segment case softlock is only incremented when a new
8581  * shadow list is created but not when an existing one is found via
8582  * seg_plookup(). pcache entries have reference count incremented/decremented
8583  * by each seg_plookup()/seg_pinactive() operation. Only entries that have 0
8584  * reference count can be purged (and purging is needed before segment can be
8585  * freed). When a private segment pcache entry is purged segvn_reclaim() will
8586  * decrement softlockcnt. Since in private segment case each of its pcache
8587  * entries only belongs to this segment we can expect that when
8588  * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8589  * segment purge will succeed and softlockcnt will drop to 0. In shared
8590  * segment case reference count in pcache entry counts active locks from many
8591  * different segments so we can't expect segment purging to succeed even when
8592  * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8593  * segment. To be able to determine when there're no pending pagelocks in
8594  * shared segment case we don't rely on purging to make softlockcnt drop to 0
8595  * but instead softlockcnt is incremented and decremented for every
8596  * segvn_pagelock(L_PAGELOCK/L_PAGEUNLOCK) call regardless if a new shadow
8597  * list was created or an existing one was found. When softlockcnt drops to 0
8598  * this segment no longer has any claims for pcached shadow lists and the
8599  * segment can be freed even if there're still active pcache entries
8600  * shared by this segment anon map. Shared segment pcache entries belong to
8601  * anon map and are typically removed when anon map is freed after all
8602  * processes destroy the segments that use this anon map.
8603  */
8604 static int
8605 segvn_pagelock(struct seg *seg, caddr_t addr, size_t len, struct page ***ppp,
8606     enum lock_type type, enum seg_rw rw)
8607 {
8608 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8609 	size_t np;
8610 	pgcnt_t adjustpages;
8611 	pgcnt_t npages;
8612 	ulong_t anon_index;
8613 	uint_t protchk = (rw == S_READ) ? PROT_READ : PROT_WRITE;
8614 	uint_t error;
8615 	struct anon_map *amp;
8616 	pgcnt_t anpgcnt;
8617 	struct page **pplist, **pl, *pp;
8618 	caddr_t a;
8619 	size_t page;
8620 	caddr_t lpgaddr, lpgeaddr;
8621 	anon_sync_obj_t cookie;
8622 	int anlock;
8623 	struct anon_map *pamp;
8624 	caddr_t paddr;
8625 	seg_preclaim_cbfunc_t preclaim_callback;
8626 	size_t pgsz;
8627 	int use_pcache;
8628 	size_t wlen;
8629 	uint_t pflags = 0;
8630 	int sftlck_sbase = 0;
8631 	int sftlck_send = 0;
8632 
8633 #ifdef DEBUG
8634 	if (type == L_PAGELOCK && segvn_pglock_mtbf) {
8635 		hrtime_t ts = gethrtime();
8636 		if ((ts % segvn_pglock_mtbf) == 0) {
8637 			return (ENOTSUP);
8638 		}
8639 		if ((ts % segvn_pglock_mtbf) == 1) {
8640 			return (EFAULT);
8641 		}
8642 	}
8643 #endif
8644 
8645 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_START,
8646 	    "segvn_pagelock: start seg %p addr %p", seg, addr);
8647 
8648 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
8649 	ASSERT(type == L_PAGELOCK || type == L_PAGEUNLOCK);
8650 
8651 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8652 
8653 	/*
8654 	 * for now we only support pagelock to anon memory. We would have to
8655 	 * check protections for vnode objects and call into the vnode driver.
8656 	 * That's too much for a fast path. Let the fault entry point handle
8657 	 * it.
8658 	 */
8659 	if (svd->vp != NULL) {
8660 		if (type == L_PAGELOCK) {
8661 			error = ENOTSUP;
8662 			goto out;
8663 		}
8664 		panic("segvn_pagelock(L_PAGEUNLOCK): vp != NULL");
8665 	}
8666 	if ((amp = svd->amp) == NULL) {
8667 		if (type == L_PAGELOCK) {
8668 			error = EFAULT;
8669 			goto out;
8670 		}
8671 		panic("segvn_pagelock(L_PAGEUNLOCK): amp == NULL");
8672 	}
8673 	if (rw != S_READ && rw != S_WRITE) {
8674 		if (type == L_PAGELOCK) {
8675 			error = ENOTSUP;
8676 			goto out;
8677 		}
8678 		panic("segvn_pagelock(L_PAGEUNLOCK): bad rw");
8679 	}
8680 
8681 	if (seg->s_szc != 0) {
8682 		/*
8683 		 * We are adjusting the pagelock region to the large page size
8684 		 * boundary because the unlocked part of a large page cannot
8685 		 * be freed anyway unless all constituent pages of a large
8686 		 * page are locked. Bigger regions reduce pcache chain length
8687 		 * and improve lookup performance. The tradeoff is that the
8688 		 * very first segvn_pagelock() call for a given page is more
8689 		 * expensive if only 1 page_t is needed for IO. This is only
8690 		 * an issue if pcache entry doesn't get reused by several
8691 		 * subsequent calls. We optimize here for the case when pcache
8692 		 * is heavily used by repeated IOs to the same address range.
8693 		 *
8694 		 * Note segment's page size cannot change while we are holding
8695 		 * as lock.  And then it cannot change while softlockcnt is
8696 		 * not 0. This will allow us to correctly recalculate large
8697 		 * page size region for the matching pageunlock/reclaim call
8698 		 * since as_pageunlock() caller must always match
8699 		 * as_pagelock() call's addr and len.
8700 		 *
8701 		 * For pageunlock *ppp points to the pointer of page_t that
8702 		 * corresponds to the real unadjusted start address. Similar
8703 		 * for pagelock *ppp must point to the pointer of page_t that
8704 		 * corresponds to the real unadjusted start address.
8705 		 */
8706 		pgsz = page_get_pagesize(seg->s_szc);
8707 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
8708 		adjustpages = btop((uintptr_t)(addr - lpgaddr));
8709 	} else if (len < segvn_pglock_comb_thrshld) {
8710 		lpgaddr = addr;
8711 		lpgeaddr = addr + len;
8712 		adjustpages = 0;
8713 		pgsz = PAGESIZE;
8714 	} else {
8715 		/*
8716 		 * Align the address range of large enough requests to allow
8717 		 * combining of different shadow lists into 1 to reduce memory
8718 		 * overhead from potentially overlapping large shadow lists
8719 		 * (worst case is we have a 1MB IO into buffers with start
8720 		 * addresses separated by 4K).  Alignment is only possible if
8721 		 * padded chunks have sufficient access permissions. Note
8722 		 * permissions won't change between L_PAGELOCK and
8723 		 * L_PAGEUNLOCK calls since non 0 softlockcnt will force
8724 		 * segvn_setprot() to wait until softlockcnt drops to 0. This
8725 		 * allows us to determine in L_PAGEUNLOCK the same range we
8726 		 * computed in L_PAGELOCK.
8727 		 *
8728 		 * If alignment is limited by segment ends set
8729 		 * sftlck_sbase/sftlck_send flags. In L_PAGELOCK case when
8730 		 * these flags are set bump softlockcnt_sbase/softlockcnt_send
8731 		 * per segment counters. In L_PAGEUNLOCK case decrease
8732 		 * softlockcnt_sbase/softlockcnt_send counters if
8733 		 * sftlck_sbase/sftlck_send flags are set.  When
8734 		 * softlockcnt_sbase/softlockcnt_send are non 0
8735 		 * segvn_concat()/segvn_extend_prev()/segvn_extend_next()
8736 		 * won't merge the segments. This restriction combined with
8737 		 * restriction on segment unmapping and splitting for segments
8738 		 * that have non 0 softlockcnt allows L_PAGEUNLOCK to
8739 		 * correctly determine the same range that was previously
8740 		 * locked by matching L_PAGELOCK.
8741 		 */
8742 		pflags = SEGP_PSHIFT | (segvn_pglock_comb_bshift << 16);
8743 		pgsz = PAGESIZE;
8744 		if (svd->type == MAP_PRIVATE) {
8745 			lpgaddr = (caddr_t)P2ALIGN((uintptr_t)addr,
8746 			    segvn_pglock_comb_balign);
8747 			if (lpgaddr < seg->s_base) {
8748 				lpgaddr = seg->s_base;
8749 				sftlck_sbase = 1;
8750 			}
8751 		} else {
8752 			ulong_t aix = svd->anon_index + seg_page(seg, addr);
8753 			ulong_t aaix = P2ALIGN(aix, segvn_pglock_comb_palign);
8754 			if (aaix < svd->anon_index) {
8755 				lpgaddr = seg->s_base;
8756 				sftlck_sbase = 1;
8757 			} else {
8758 				lpgaddr = addr - ptob(aix - aaix);
8759 				ASSERT(lpgaddr >= seg->s_base);
8760 			}
8761 		}
8762 		if (svd->pageprot && lpgaddr != addr) {
8763 			struct vpage *vp = &svd->vpage[seg_page(seg, lpgaddr)];
8764 			struct vpage *evp = &svd->vpage[seg_page(seg, addr)];
8765 			while (vp < evp) {
8766 				if ((VPP_PROT(vp) & protchk) == 0) {
8767 					break;
8768 				}
8769 				vp++;
8770 			}
8771 			if (vp < evp) {
8772 				lpgaddr = addr;
8773 				pflags = 0;
8774 			}
8775 		}
8776 		lpgeaddr = addr + len;
8777 		if (pflags) {
8778 			if (svd->type == MAP_PRIVATE) {
8779 				lpgeaddr = (caddr_t)P2ROUNDUP(
8780 				    (uintptr_t)lpgeaddr,
8781 				    segvn_pglock_comb_balign);
8782 			} else {
8783 				ulong_t aix = svd->anon_index +
8784 				    seg_page(seg, lpgeaddr);
8785 				ulong_t aaix = P2ROUNDUP(aix,
8786 				    segvn_pglock_comb_palign);
8787 				if (aaix < aix) {
8788 					lpgeaddr = 0;
8789 				} else {
8790 					lpgeaddr += ptob(aaix - aix);
8791 				}
8792 			}
8793 			if (lpgeaddr == 0 ||
8794 			    lpgeaddr > seg->s_base + seg->s_size) {
8795 				lpgeaddr = seg->s_base + seg->s_size;
8796 				sftlck_send = 1;
8797 			}
8798 		}
8799 		if (svd->pageprot && lpgeaddr != addr + len) {
8800 			struct vpage *vp;
8801 			struct vpage *evp;
8802 
8803 			vp = &svd->vpage[seg_page(seg, addr + len)];
8804 			evp = &svd->vpage[seg_page(seg, lpgeaddr)];
8805 
8806 			while (vp < evp) {
8807 				if ((VPP_PROT(vp) & protchk) == 0) {
8808 					break;
8809 				}
8810 				vp++;
8811 			}
8812 			if (vp < evp) {
8813 				lpgeaddr = addr + len;
8814 			}
8815 		}
8816 		adjustpages = btop((uintptr_t)(addr - lpgaddr));
8817 	}
8818 
8819 	/*
8820 	 * For MAP_SHARED segments we create pcache entries tagged by amp and
8821 	 * anon index so that we can share pcache entries with other segments
8822 	 * that map this amp.  For private segments pcache entries are tagged
8823 	 * with segment and virtual address.
8824 	 */
8825 	if (svd->type == MAP_SHARED) {
8826 		pamp = amp;
8827 		paddr = (caddr_t)((lpgaddr - seg->s_base) +
8828 		    ptob(svd->anon_index));
8829 		preclaim_callback = shamp_reclaim;
8830 	} else {
8831 		pamp = NULL;
8832 		paddr = lpgaddr;
8833 		preclaim_callback = segvn_reclaim;
8834 	}
8835 
8836 	if (type == L_PAGEUNLOCK) {
8837 		VM_STAT_ADD(segvnvmstats.pagelock[0]);
8838 
8839 		/*
8840 		 * update hat ref bits for /proc. We need to make sure
8841 		 * that threads tracing the ref and mod bits of the
8842 		 * address space get the right data.
8843 		 * Note: page ref and mod bits are updated at reclaim time
8844 		 */
8845 		if (seg->s_as->a_vbits) {
8846 			for (a = addr; a < addr + len; a += PAGESIZE) {
8847 				if (rw == S_WRITE) {
8848 					hat_setstat(seg->s_as, a,
8849 					    PAGESIZE, P_REF | P_MOD);
8850 				} else {
8851 					hat_setstat(seg->s_as, a,
8852 					    PAGESIZE, P_REF);
8853 				}
8854 			}
8855 		}
8856 
8857 		/*
8858 		 * Check the shadow list entry after the last page used in
8859 		 * this IO request. If it's NOPCACHE_SHWLIST the shadow list
8860 		 * was not inserted into pcache and is not large page
8861 		 * adjusted.  In this case call reclaim callback directly and
8862 		 * don't adjust the shadow list start and size for large
8863 		 * pages.
8864 		 */
8865 		npages = btop(len);
8866 		if ((*ppp)[npages] == NOPCACHE_SHWLIST) {
8867 			void *ptag;
8868 			if (pamp != NULL) {
8869 				ASSERT(svd->type == MAP_SHARED);
8870 				ptag = (void *)pamp;
8871 				paddr = (caddr_t)((addr - seg->s_base) +
8872 				    ptob(svd->anon_index));
8873 			} else {
8874 				ptag = (void *)seg;
8875 				paddr = addr;
8876 			}
8877 			(*preclaim_callback)(ptag, paddr, len, *ppp, rw, 0);
8878 		} else {
8879 			ASSERT((*ppp)[npages] == PCACHE_SHWLIST ||
8880 			    IS_SWAPFSVP((*ppp)[npages]->p_vnode));
8881 			len = lpgeaddr - lpgaddr;
8882 			npages = btop(len);
8883 			seg_pinactive(seg, pamp, paddr, len,
8884 			    *ppp - adjustpages, rw, pflags, preclaim_callback);
8885 		}
8886 
8887 		if (pamp != NULL) {
8888 			ASSERT(svd->type == MAP_SHARED);
8889 			ASSERT(svd->softlockcnt >= npages);
8890 			atomic_add_long((ulong_t *)&svd->softlockcnt, -npages);
8891 		}
8892 
8893 		if (sftlck_sbase) {
8894 			ASSERT(svd->softlockcnt_sbase > 0);
8895 			atomic_dec_ulong((ulong_t *)&svd->softlockcnt_sbase);
8896 		}
8897 		if (sftlck_send) {
8898 			ASSERT(svd->softlockcnt_send > 0);
8899 			atomic_dec_ulong((ulong_t *)&svd->softlockcnt_send);
8900 		}
8901 
8902 		/*
8903 		 * If someone is blocked while unmapping, we purge
8904 		 * segment page cache and thus reclaim pplist synchronously
8905 		 * without waiting for seg_pasync_thread. This speeds up
8906 		 * unmapping in cases where munmap(2) is called, while
8907 		 * raw async i/o is still in progress or where a thread
8908 		 * exits on data fault in a multithreaded application.
8909 		 */
8910 		if (AS_ISUNMAPWAIT(seg->s_as)) {
8911 			if (svd->softlockcnt == 0) {
8912 				mutex_enter(&seg->s_as->a_contents);
8913 				if (AS_ISUNMAPWAIT(seg->s_as)) {
8914 					AS_CLRUNMAPWAIT(seg->s_as);
8915 					cv_broadcast(&seg->s_as->a_cv);
8916 				}
8917 				mutex_exit(&seg->s_as->a_contents);
8918 			} else if (pamp == NULL) {
8919 				/*
8920 				 * softlockcnt is not 0 and this is a
8921 				 * MAP_PRIVATE segment. Try to purge its
8922 				 * pcache entries to reduce softlockcnt.
8923 				 * If it drops to 0 segvn_reclaim()
8924 				 * will wake up a thread waiting on
8925 				 * unmapwait flag.
8926 				 *
8927 				 * We don't purge MAP_SHARED segments with non
8928 				 * 0 softlockcnt since IO is still in progress
8929 				 * for such segments.
8930 				 */
8931 				ASSERT(svd->type == MAP_PRIVATE);
8932 				segvn_purge(seg);
8933 			}
8934 		}
8935 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8936 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_UNLOCK_END,
8937 		    "segvn_pagelock: unlock seg %p addr %p", seg, addr);
8938 		return (0);
8939 	}
8940 
8941 	/* The L_PAGELOCK case ... */
8942 
8943 	VM_STAT_ADD(segvnvmstats.pagelock[1]);
8944 
8945 	/*
8946 	 * For MAP_SHARED segments we have to check protections before
8947 	 * seg_plookup() since pcache entries may be shared by many segments
8948 	 * with potentially different page protections.
8949 	 */
8950 	if (pamp != NULL) {
8951 		ASSERT(svd->type == MAP_SHARED);
8952 		if (svd->pageprot == 0) {
8953 			if ((svd->prot & protchk) == 0) {
8954 				error = EACCES;
8955 				goto out;
8956 			}
8957 		} else {
8958 			/*
8959 			 * check page protections
8960 			 */
8961 			caddr_t ea;
8962 
8963 			if (seg->s_szc) {
8964 				a = lpgaddr;
8965 				ea = lpgeaddr;
8966 			} else {
8967 				a = addr;
8968 				ea = addr + len;
8969 			}
8970 			for (; a < ea; a += pgsz) {
8971 				struct vpage *vp;
8972 
8973 				ASSERT(seg->s_szc == 0 ||
8974 				    sameprot(seg, a, pgsz));
8975 				vp = &svd->vpage[seg_page(seg, a)];
8976 				if ((VPP_PROT(vp) & protchk) == 0) {
8977 					error = EACCES;
8978 					goto out;
8979 				}
8980 			}
8981 		}
8982 	}
8983 
8984 	/*
8985 	 * try to find pages in segment page cache
8986 	 */
8987 	pplist = seg_plookup(seg, pamp, paddr, lpgeaddr - lpgaddr, rw, pflags);
8988 	if (pplist != NULL) {
8989 		if (pamp != NULL) {
8990 			npages = btop((uintptr_t)(lpgeaddr - lpgaddr));
8991 			ASSERT(svd->type == MAP_SHARED);
8992 			atomic_add_long((ulong_t *)&svd->softlockcnt,
8993 			    npages);
8994 		}
8995 		if (sftlck_sbase) {
8996 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_sbase);
8997 		}
8998 		if (sftlck_send) {
8999 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_send);
9000 		}
9001 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9002 		*ppp = pplist + adjustpages;
9003 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_HIT_END,
9004 		    "segvn_pagelock: cache hit seg %p addr %p", seg, addr);
9005 		return (0);
9006 	}
9007 
9008 	/*
9009 	 * For MAP_SHARED segments we already verified above that segment
9010 	 * protections allow this pagelock operation.
9011 	 */
9012 	if (pamp == NULL) {
9013 		ASSERT(svd->type == MAP_PRIVATE);
9014 		if (svd->pageprot == 0) {
9015 			if ((svd->prot & protchk) == 0) {
9016 				error = EACCES;
9017 				goto out;
9018 			}
9019 			if (svd->prot & PROT_WRITE) {
9020 				wlen = lpgeaddr - lpgaddr;
9021 			} else {
9022 				wlen = 0;
9023 				ASSERT(rw == S_READ);
9024 			}
9025 		} else {
9026 			int wcont = 1;
9027 			/*
9028 			 * check page protections
9029 			 */
9030 			for (a = lpgaddr, wlen = 0; a < lpgeaddr; a += pgsz) {
9031 				struct vpage *vp;
9032 
9033 				ASSERT(seg->s_szc == 0 ||
9034 				    sameprot(seg, a, pgsz));
9035 				vp = &svd->vpage[seg_page(seg, a)];
9036 				if ((VPP_PROT(vp) & protchk) == 0) {
9037 					error = EACCES;
9038 					goto out;
9039 				}
9040 				if (wcont && (VPP_PROT(vp) & PROT_WRITE)) {
9041 					wlen += pgsz;
9042 				} else {
9043 					wcont = 0;
9044 					ASSERT(rw == S_READ);
9045 				}
9046 			}
9047 		}
9048 		ASSERT(rw == S_READ || wlen == lpgeaddr - lpgaddr);
9049 		ASSERT(rw == S_WRITE || wlen <= lpgeaddr - lpgaddr);
9050 	}
9051 
9052 	/*
9053 	 * Only build large page adjusted shadow list if we expect to insert
9054 	 * it into pcache. For large enough pages it's a big overhead to
9055 	 * create a shadow list of the entire large page. But this overhead
9056 	 * should be amortized over repeated pcache hits on subsequent reuse
9057 	 * of this shadow list (IO into any range within this shadow list will
9058 	 * find it in pcache since we large page align the request for pcache
9059 	 * lookups). pcache performance is improved with bigger shadow lists
9060 	 * as it reduces the time to pcache the entire big segment and reduces
9061 	 * pcache chain length.
9062 	 */
9063 	if (seg_pinsert_check(seg, pamp, paddr,
9064 	    lpgeaddr - lpgaddr, pflags) == SEGP_SUCCESS) {
9065 		addr = lpgaddr;
9066 		len = lpgeaddr - lpgaddr;
9067 		use_pcache = 1;
9068 	} else {
9069 		use_pcache = 0;
9070 		/*
9071 		 * Since this entry will not be inserted into the pcache, we
9072 		 * will not do any adjustments to the starting address or
9073 		 * size of the memory to be locked.
9074 		 */
9075 		adjustpages = 0;
9076 	}
9077 	npages = btop(len);
9078 
9079 	pplist = kmem_alloc(sizeof (page_t *) * (npages + 1), KM_SLEEP);
9080 	pl = pplist;
9081 	*ppp = pplist + adjustpages;
9082 	/*
9083 	 * If use_pcache is 0 this shadow list is not large page adjusted.
9084 	 * Record this info in the last entry of shadow array so that
9085 	 * L_PAGEUNLOCK can determine if it should large page adjust the
9086 	 * address range to find the real range that was locked.
9087 	 */
9088 	pl[npages] = use_pcache ? PCACHE_SHWLIST : NOPCACHE_SHWLIST;
9089 
9090 	page = seg_page(seg, addr);
9091 	anon_index = svd->anon_index + page;
9092 
9093 	anlock = 0;
9094 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
9095 	ASSERT(amp->a_szc >= seg->s_szc);
9096 	anpgcnt = page_get_pagecnt(amp->a_szc);
9097 	for (a = addr; a < addr + len; a += PAGESIZE, anon_index++) {
9098 		struct anon *ap;
9099 		struct vnode *vp;
9100 		u_offset_t off;
9101 
9102 		/*
9103 		 * Lock and unlock anon array only once per large page.
9104 		 * anon_array_enter() locks the root anon slot according to
9105 		 * a_szc which can't change while anon map is locked.  We lock
9106 		 * anon the first time through this loop and each time we
9107 		 * reach anon index that corresponds to a root of a large
9108 		 * page.
9109 		 */
9110 		if (a == addr || P2PHASE(anon_index, anpgcnt) == 0) {
9111 			ASSERT(anlock == 0);
9112 			anon_array_enter(amp, anon_index, &cookie);
9113 			anlock = 1;
9114 		}
9115 		ap = anon_get_ptr(amp->ahp, anon_index);
9116 
9117 		/*
9118 		 * We must never use seg_pcache for COW pages
9119 		 * because we might end up with original page still
9120 		 * lying in seg_pcache even after private page is
9121 		 * created. This leads to data corruption as
9122 		 * aio_write refers to the page still in cache
9123 		 * while all other accesses refer to the private
9124 		 * page.
9125 		 */
9126 		if (ap == NULL || ap->an_refcnt != 1) {
9127 			struct vpage *vpage;
9128 
9129 			if (seg->s_szc) {
9130 				error = EFAULT;
9131 				break;
9132 			}
9133 			if (svd->vpage != NULL) {
9134 				vpage = &svd->vpage[seg_page(seg, a)];
9135 			} else {
9136 				vpage = NULL;
9137 			}
9138 			ASSERT(anlock);
9139 			anon_array_exit(&cookie);
9140 			anlock = 0;
9141 			pp = NULL;
9142 			error = segvn_faultpage(seg->s_as->a_hat, seg, a, 0,
9143 			    vpage, &pp, 0, F_INVAL, rw, 1);
9144 			if (error) {
9145 				error = fc_decode(error);
9146 				break;
9147 			}
9148 			anon_array_enter(amp, anon_index, &cookie);
9149 			anlock = 1;
9150 			ap = anon_get_ptr(amp->ahp, anon_index);
9151 			if (ap == NULL || ap->an_refcnt != 1) {
9152 				error = EFAULT;
9153 				break;
9154 			}
9155 		}
9156 		swap_xlate(ap, &vp, &off);
9157 		pp = page_lookup_nowait(vp, off, SE_SHARED);
9158 		if (pp == NULL) {
9159 			error = EFAULT;
9160 			break;
9161 		}
9162 		if (ap->an_pvp != NULL) {
9163 			anon_swap_free(ap, pp);
9164 		}
9165 		/*
9166 		 * Unlock anon if this is the last slot in a large page.
9167 		 */
9168 		if (P2PHASE(anon_index, anpgcnt) == anpgcnt - 1) {
9169 			ASSERT(anlock);
9170 			anon_array_exit(&cookie);
9171 			anlock = 0;
9172 		}
9173 		*pplist++ = pp;
9174 	}
9175 	if (anlock) {		/* Ensure the lock is dropped */
9176 		anon_array_exit(&cookie);
9177 	}
9178 	ANON_LOCK_EXIT(&amp->a_rwlock);
9179 
9180 	if (a >= addr + len) {
9181 		atomic_add_long((ulong_t *)&svd->softlockcnt, npages);
9182 		if (pamp != NULL) {
9183 			ASSERT(svd->type == MAP_SHARED);
9184 			atomic_add_long((ulong_t *)&pamp->a_softlockcnt,
9185 			    npages);
9186 			wlen = len;
9187 		}
9188 		if (sftlck_sbase) {
9189 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_sbase);
9190 		}
9191 		if (sftlck_send) {
9192 			atomic_inc_ulong((ulong_t *)&svd->softlockcnt_send);
9193 		}
9194 		if (use_pcache) {
9195 			(void) seg_pinsert(seg, pamp, paddr, len, wlen, pl,
9196 			    rw, pflags, preclaim_callback);
9197 		}
9198 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9199 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_FILL_END,
9200 		    "segvn_pagelock: cache fill seg %p addr %p", seg, addr);
9201 		return (0);
9202 	}
9203 
9204 	pplist = pl;
9205 	np = ((uintptr_t)(a - addr)) >> PAGESHIFT;
9206 	while (np > (uint_t)0) {
9207 		ASSERT(PAGE_LOCKED(*pplist));
9208 		page_unlock(*pplist);
9209 		np--;
9210 		pplist++;
9211 	}
9212 	kmem_free(pl, sizeof (page_t *) * (npages + 1));
9213 out:
9214 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9215 	*ppp = NULL;
9216 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
9217 	    "segvn_pagelock: cache miss seg %p addr %p", seg, addr);
9218 	return (error);
9219 }
9220 
9221 /*
9222  * purge any cached pages in the I/O page cache
9223  */
9224 static void
9225 segvn_purge(struct seg *seg)
9226 {
9227 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9228 
9229 	/*
9230 	 * pcache is only used by pure anon segments.
9231 	 */
9232 	if (svd->amp == NULL || svd->vp != NULL) {
9233 		return;
9234 	}
9235 
9236 	/*
9237 	 * For MAP_SHARED segments non 0 segment's softlockcnt means
9238 	 * active IO is still in progress via this segment. So we only
9239 	 * purge MAP_SHARED segments when their softlockcnt is 0.
9240 	 */
9241 	if (svd->type == MAP_PRIVATE) {
9242 		if (svd->softlockcnt) {
9243 			seg_ppurge(seg, NULL, 0);
9244 		}
9245 	} else if (svd->softlockcnt == 0 && svd->amp->a_softlockcnt != 0) {
9246 		seg_ppurge(seg, svd->amp, 0);
9247 	}
9248 }
9249 
9250 /*
9251  * If async argument is not 0 we are called from pcache async thread and don't
9252  * hold AS lock.
9253  */
9254 
9255 /*ARGSUSED*/
9256 static int
9257 segvn_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9258 	enum seg_rw rw, int async)
9259 {
9260 	struct seg *seg = (struct seg *)ptag;
9261 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9262 	pgcnt_t np, npages;
9263 	struct page **pl;
9264 
9265 	npages = np = btop(len);
9266 	ASSERT(npages);
9267 
9268 	ASSERT(svd->vp == NULL && svd->amp != NULL);
9269 	ASSERT(svd->softlockcnt >= npages);
9270 	ASSERT(async || AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
9271 
9272 	pl = pplist;
9273 
9274 	ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9275 	ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9276 
9277 	while (np > (uint_t)0) {
9278 		if (rw == S_WRITE) {
9279 			hat_setrefmod(*pplist);
9280 		} else {
9281 			hat_setref(*pplist);
9282 		}
9283 		page_unlock(*pplist);
9284 		np--;
9285 		pplist++;
9286 	}
9287 
9288 	kmem_free(pl, sizeof (page_t *) * (npages + 1));
9289 
9290 	/*
9291 	 * If we are pcache async thread we don't hold AS lock. This means if
9292 	 * softlockcnt drops to 0 after the decrement below address space may
9293 	 * get freed. We can't allow it since after softlock derement to 0 we
9294 	 * still need to access as structure for possible wakeup of unmap
9295 	 * waiters. To prevent the disappearance of as we take this segment
9296 	 * segfree_syncmtx. segvn_free() also takes this mutex as a barrier to
9297 	 * make sure this routine completes before segment is freed.
9298 	 *
9299 	 * The second complication we have to deal with in async case is a
9300 	 * possibility of missed wake up of unmap wait thread. When we don't
9301 	 * hold as lock here we may take a_contents lock before unmap wait
9302 	 * thread that was first to see softlockcnt was still not 0. As a
9303 	 * result we'll fail to wake up an unmap wait thread. To avoid this
9304 	 * race we set nounmapwait flag in as structure if we drop softlockcnt
9305 	 * to 0 when we were called by pcache async thread.  unmapwait thread
9306 	 * will not block if this flag is set.
9307 	 */
9308 	if (async) {
9309 		mutex_enter(&svd->segfree_syncmtx);
9310 	}
9311 
9312 	if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -npages)) {
9313 		if (async || AS_ISUNMAPWAIT(seg->s_as)) {
9314 			mutex_enter(&seg->s_as->a_contents);
9315 			if (async) {
9316 				AS_SETNOUNMAPWAIT(seg->s_as);
9317 			}
9318 			if (AS_ISUNMAPWAIT(seg->s_as)) {
9319 				AS_CLRUNMAPWAIT(seg->s_as);
9320 				cv_broadcast(&seg->s_as->a_cv);
9321 			}
9322 			mutex_exit(&seg->s_as->a_contents);
9323 		}
9324 	}
9325 
9326 	if (async) {
9327 		mutex_exit(&svd->segfree_syncmtx);
9328 	}
9329 	return (0);
9330 }
9331 
9332 /*ARGSUSED*/
9333 static int
9334 shamp_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9335 	enum seg_rw rw, int async)
9336 {
9337 	amp_t *amp = (amp_t *)ptag;
9338 	pgcnt_t np, npages;
9339 	struct page **pl;
9340 
9341 	npages = np = btop(len);
9342 	ASSERT(npages);
9343 	ASSERT(amp->a_softlockcnt >= npages);
9344 
9345 	pl = pplist;
9346 
9347 	ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9348 	ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9349 
9350 	while (np > (uint_t)0) {
9351 		if (rw == S_WRITE) {
9352 			hat_setrefmod(*pplist);
9353 		} else {
9354 			hat_setref(*pplist);
9355 		}
9356 		page_unlock(*pplist);
9357 		np--;
9358 		pplist++;
9359 	}
9360 
9361 	kmem_free(pl, sizeof (page_t *) * (npages + 1));
9362 
9363 	/*
9364 	 * If somebody sleeps in anonmap_purge() wake them up if a_softlockcnt
9365 	 * drops to 0. anon map can't be freed until a_softlockcnt drops to 0
9366 	 * and anonmap_purge() acquires a_purgemtx.
9367 	 */
9368 	mutex_enter(&amp->a_purgemtx);
9369 	if (!atomic_add_long_nv((ulong_t *)&amp->a_softlockcnt, -npages) &&
9370 	    amp->a_purgewait) {
9371 		amp->a_purgewait = 0;
9372 		cv_broadcast(&amp->a_purgecv);
9373 	}
9374 	mutex_exit(&amp->a_purgemtx);
9375 	return (0);
9376 }
9377 
9378 /*
9379  * get a memory ID for an addr in a given segment
9380  *
9381  * XXX only creates PAGESIZE pages if anon slots are not initialized.
9382  * At fault time they will be relocated into larger pages.
9383  */
9384 static int
9385 segvn_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
9386 {
9387 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9388 	struct anon 	*ap = NULL;
9389 	ulong_t		anon_index;
9390 	struct anon_map	*amp;
9391 	anon_sync_obj_t cookie;
9392 
9393 	if (svd->type == MAP_PRIVATE) {
9394 		memidp->val[0] = (uintptr_t)seg->s_as;
9395 		memidp->val[1] = (uintptr_t)addr;
9396 		return (0);
9397 	}
9398 
9399 	if (svd->type == MAP_SHARED) {
9400 		if (svd->vp) {
9401 			memidp->val[0] = (uintptr_t)svd->vp;
9402 			memidp->val[1] = (u_longlong_t)svd->offset +
9403 			    (uintptr_t)(addr - seg->s_base);
9404 			return (0);
9405 		} else {
9406 
9407 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
9408 			if ((amp = svd->amp) != NULL) {
9409 				anon_index = svd->anon_index +
9410 				    seg_page(seg, addr);
9411 			}
9412 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9413 
9414 			ASSERT(amp != NULL);
9415 
9416 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
9417 			anon_array_enter(amp, anon_index, &cookie);
9418 			ap = anon_get_ptr(amp->ahp, anon_index);
9419 			if (ap == NULL) {
9420 				page_t		*pp;
9421 
9422 				pp = anon_zero(seg, addr, &ap, svd->cred);
9423 				if (pp == NULL) {
9424 					anon_array_exit(&cookie);
9425 					ANON_LOCK_EXIT(&amp->a_rwlock);
9426 					return (ENOMEM);
9427 				}
9428 				ASSERT(anon_get_ptr(amp->ahp, anon_index)
9429 				    == NULL);
9430 				(void) anon_set_ptr(amp->ahp, anon_index,
9431 				    ap, ANON_SLEEP);
9432 				page_unlock(pp);
9433 			}
9434 
9435 			anon_array_exit(&cookie);
9436 			ANON_LOCK_EXIT(&amp->a_rwlock);
9437 
9438 			memidp->val[0] = (uintptr_t)ap;
9439 			memidp->val[1] = (uintptr_t)addr & PAGEOFFSET;
9440 			return (0);
9441 		}
9442 	}
9443 	return (EINVAL);
9444 }
9445 
9446 static int
9447 sameprot(struct seg *seg, caddr_t a, size_t len)
9448 {
9449 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9450 	struct vpage *vpage;
9451 	spgcnt_t pages = btop(len);
9452 	uint_t prot;
9453 
9454 	if (svd->pageprot == 0)
9455 		return (1);
9456 
9457 	ASSERT(svd->vpage != NULL);
9458 
9459 	vpage = &svd->vpage[seg_page(seg, a)];
9460 	prot = VPP_PROT(vpage);
9461 	vpage++;
9462 	pages--;
9463 	while (pages-- > 0) {
9464 		if (prot != VPP_PROT(vpage))
9465 			return (0);
9466 		vpage++;
9467 	}
9468 	return (1);
9469 }
9470 
9471 /*
9472  * Get memory allocation policy info for specified address in given segment
9473  */
9474 static lgrp_mem_policy_info_t *
9475 segvn_getpolicy(struct seg *seg, caddr_t addr)
9476 {
9477 	struct anon_map		*amp;
9478 	ulong_t			anon_index;
9479 	lgrp_mem_policy_info_t	*policy_info;
9480 	struct segvn_data	*svn_data;
9481 	u_offset_t		vn_off;
9482 	vnode_t			*vp;
9483 
9484 	ASSERT(seg != NULL);
9485 
9486 	svn_data = (struct segvn_data *)seg->s_data;
9487 	if (svn_data == NULL)
9488 		return (NULL);
9489 
9490 	/*
9491 	 * Get policy info for private or shared memory
9492 	 */
9493 	if (svn_data->type != MAP_SHARED) {
9494 		if (svn_data->tr_state != SEGVN_TR_ON) {
9495 			policy_info = &svn_data->policy_info;
9496 		} else {
9497 			policy_info = &svn_data->tr_policy_info;
9498 			ASSERT(policy_info->mem_policy ==
9499 			    LGRP_MEM_POLICY_NEXT_SEG);
9500 		}
9501 	} else {
9502 		amp = svn_data->amp;
9503 		anon_index = svn_data->anon_index + seg_page(seg, addr);
9504 		vp = svn_data->vp;
9505 		vn_off = svn_data->offset + (uintptr_t)(addr - seg->s_base);
9506 		policy_info = lgrp_shm_policy_get(amp, anon_index, vp, vn_off);
9507 	}
9508 
9509 	return (policy_info);
9510 }
9511 
9512 /*ARGSUSED*/
9513 static int
9514 segvn_capable(struct seg *seg, segcapability_t capability)
9515 {
9516 	return (0);
9517 }
9518 
9519 /*
9520  * Bind text vnode segment to an amp. If we bind successfully mappings will be
9521  * established to per vnode mapping per lgroup amp pages instead of to vnode
9522  * pages. There's one amp per vnode text mapping per lgroup. Many processes
9523  * may share the same text replication amp. If a suitable amp doesn't already
9524  * exist in svntr hash table create a new one.  We may fail to bind to amp if
9525  * segment is not eligible for text replication.  Code below first checks for
9526  * these conditions. If binding is successful segment tr_state is set to on
9527  * and svd->amp points to the amp to use. Otherwise tr_state is set to off and
9528  * svd->amp remains as NULL.
9529  */
9530 static void
9531 segvn_textrepl(struct seg *seg)
9532 {
9533 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
9534 	vnode_t			*vp = svd->vp;
9535 	u_offset_t		off = svd->offset;
9536 	size_t			size = seg->s_size;
9537 	u_offset_t		eoff = off + size;
9538 	uint_t			szc = seg->s_szc;
9539 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9540 	svntr_t			*svntrp;
9541 	struct vattr		va;
9542 	proc_t			*p = seg->s_as->a_proc;
9543 	lgrp_id_t		lgrp_id;
9544 	lgrp_id_t		olid;
9545 	int			first;
9546 	struct anon_map		*amp;
9547 
9548 	ASSERT(AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
9549 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9550 	ASSERT(p != NULL);
9551 	ASSERT(svd->tr_state == SEGVN_TR_INIT);
9552 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9553 	ASSERT(svd->flags & MAP_TEXT);
9554 	ASSERT(svd->type == MAP_PRIVATE);
9555 	ASSERT(vp != NULL && svd->amp == NULL);
9556 	ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
9557 	ASSERT(!(svd->flags & MAP_NORESERVE) && svd->swresv == 0);
9558 	ASSERT(seg->s_as != &kas);
9559 	ASSERT(off < eoff);
9560 	ASSERT(svntr_hashtab != NULL);
9561 
9562 	/*
9563 	 * If numa optimizations are no longer desired bail out.
9564 	 */
9565 	if (!lgrp_optimizations()) {
9566 		svd->tr_state = SEGVN_TR_OFF;
9567 		return;
9568 	}
9569 
9570 	/*
9571 	 * Avoid creating anon maps with size bigger than the file size.
9572 	 * If VOP_GETATTR() call fails bail out.
9573 	 */
9574 	va.va_mask = AT_SIZE | AT_MTIME | AT_CTIME;
9575 	if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL) != 0) {
9576 		svd->tr_state = SEGVN_TR_OFF;
9577 		SEGVN_TR_ADDSTAT(gaerr);
9578 		return;
9579 	}
9580 	if (btopr(va.va_size) < btopr(eoff)) {
9581 		svd->tr_state = SEGVN_TR_OFF;
9582 		SEGVN_TR_ADDSTAT(overmap);
9583 		return;
9584 	}
9585 
9586 	/*
9587 	 * VVMEXEC may not be set yet if exec() prefaults text segment. Set
9588 	 * this flag now before vn_is_mapped(V_WRITE) so that MAP_SHARED
9589 	 * mapping that checks if trcache for this vnode needs to be
9590 	 * invalidated can't miss us.
9591 	 */
9592 	if (!(vp->v_flag & VVMEXEC)) {
9593 		mutex_enter(&vp->v_lock);
9594 		vp->v_flag |= VVMEXEC;
9595 		mutex_exit(&vp->v_lock);
9596 	}
9597 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9598 	/*
9599 	 * Bail out if potentially MAP_SHARED writable mappings exist to this
9600 	 * vnode.  We don't want to use old file contents from existing
9601 	 * replicas if this mapping was established after the original file
9602 	 * was changed.
9603 	 */
9604 	if (vn_is_mapped(vp, V_WRITE)) {
9605 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9606 		svd->tr_state = SEGVN_TR_OFF;
9607 		SEGVN_TR_ADDSTAT(wrcnt);
9608 		return;
9609 	}
9610 	svntrp = svntr_hashtab[hash].tr_head;
9611 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9612 		ASSERT(svntrp->tr_refcnt != 0);
9613 		if (svntrp->tr_vp != vp) {
9614 			continue;
9615 		}
9616 
9617 		/*
9618 		 * Bail out if the file or its attributes were changed after
9619 		 * this replication entry was created since we need to use the
9620 		 * latest file contents. Note that mtime test alone is not
9621 		 * sufficient because a user can explicitly change mtime via
9622 		 * utimes(2) interfaces back to the old value after modifiying
9623 		 * the file contents. To detect this case we also have to test
9624 		 * ctime which among other things records the time of the last
9625 		 * mtime change by utimes(2). ctime is not changed when the file
9626 		 * is only read or executed so we expect that typically existing
9627 		 * replication amp's can be used most of the time.
9628 		 */
9629 		if (!svntrp->tr_valid ||
9630 		    svntrp->tr_mtime.tv_sec != va.va_mtime.tv_sec ||
9631 		    svntrp->tr_mtime.tv_nsec != va.va_mtime.tv_nsec ||
9632 		    svntrp->tr_ctime.tv_sec != va.va_ctime.tv_sec ||
9633 		    svntrp->tr_ctime.tv_nsec != va.va_ctime.tv_nsec) {
9634 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9635 			svd->tr_state = SEGVN_TR_OFF;
9636 			SEGVN_TR_ADDSTAT(stale);
9637 			return;
9638 		}
9639 		/*
9640 		 * if off, eoff and szc match current segment we found the
9641 		 * existing entry we can use.
9642 		 */
9643 		if (svntrp->tr_off == off && svntrp->tr_eoff == eoff &&
9644 		    svntrp->tr_szc == szc) {
9645 			break;
9646 		}
9647 		/*
9648 		 * Don't create different but overlapping in file offsets
9649 		 * entries to avoid replication of the same file pages more
9650 		 * than once per lgroup.
9651 		 */
9652 		if ((off >= svntrp->tr_off && off < svntrp->tr_eoff) ||
9653 		    (eoff > svntrp->tr_off && eoff <= svntrp->tr_eoff)) {
9654 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9655 			svd->tr_state = SEGVN_TR_OFF;
9656 			SEGVN_TR_ADDSTAT(overlap);
9657 			return;
9658 		}
9659 	}
9660 	/*
9661 	 * If we didn't find existing entry create a new one.
9662 	 */
9663 	if (svntrp == NULL) {
9664 		svntrp = kmem_cache_alloc(svntr_cache, KM_NOSLEEP);
9665 		if (svntrp == NULL) {
9666 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9667 			svd->tr_state = SEGVN_TR_OFF;
9668 			SEGVN_TR_ADDSTAT(nokmem);
9669 			return;
9670 		}
9671 #ifdef DEBUG
9672 		{
9673 			lgrp_id_t i;
9674 			for (i = 0; i < NLGRPS_MAX; i++) {
9675 				ASSERT(svntrp->tr_amp[i] == NULL);
9676 			}
9677 		}
9678 #endif /* DEBUG */
9679 		svntrp->tr_vp = vp;
9680 		svntrp->tr_off = off;
9681 		svntrp->tr_eoff = eoff;
9682 		svntrp->tr_szc = szc;
9683 		svntrp->tr_valid = 1;
9684 		svntrp->tr_mtime = va.va_mtime;
9685 		svntrp->tr_ctime = va.va_ctime;
9686 		svntrp->tr_refcnt = 0;
9687 		svntrp->tr_next = svntr_hashtab[hash].tr_head;
9688 		svntr_hashtab[hash].tr_head = svntrp;
9689 	}
9690 	first = 1;
9691 again:
9692 	/*
9693 	 * We want to pick a replica with pages on main thread's (t_tid = 1,
9694 	 * aka T1) lgrp. Currently text replication is only optimized for
9695 	 * workloads that either have all threads of a process on the same
9696 	 * lgrp or execute their large text primarily on main thread.
9697 	 */
9698 	lgrp_id = p->p_t1_lgrpid;
9699 	if (lgrp_id == LGRP_NONE) {
9700 		/*
9701 		 * In case exec() prefaults text on non main thread use
9702 		 * current thread lgrpid.  It will become main thread anyway
9703 		 * soon.
9704 		 */
9705 		lgrp_id = lgrp_home_id(curthread);
9706 	}
9707 	/*
9708 	 * Set p_tr_lgrpid to lgrpid if it hasn't been set yet.  Otherwise
9709 	 * just set it to NLGRPS_MAX if it's different from current process T1
9710 	 * home lgrp.  p_tr_lgrpid is used to detect if process uses text
9711 	 * replication and T1 new home is different from lgrp used for text
9712 	 * replication. When this happens asyncronous segvn thread rechecks if
9713 	 * segments should change lgrps used for text replication.  If we fail
9714 	 * to set p_tr_lgrpid with atomic_cas_32 then set it to NLGRPS_MAX
9715 	 * without cas if it's not already NLGRPS_MAX and not equal lgrp_id
9716 	 * we want to use.  We don't need to use cas in this case because
9717 	 * another thread that races in between our non atomic check and set
9718 	 * may only change p_tr_lgrpid to NLGRPS_MAX at this point.
9719 	 */
9720 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9721 	olid = p->p_tr_lgrpid;
9722 	if (lgrp_id != olid && olid != NLGRPS_MAX) {
9723 		lgrp_id_t nlid = (olid == LGRP_NONE) ? lgrp_id : NLGRPS_MAX;
9724 		if (atomic_cas_32((uint32_t *)&p->p_tr_lgrpid, olid, nlid) !=
9725 		    olid) {
9726 			olid = p->p_tr_lgrpid;
9727 			ASSERT(olid != LGRP_NONE);
9728 			if (olid != lgrp_id && olid != NLGRPS_MAX) {
9729 				p->p_tr_lgrpid = NLGRPS_MAX;
9730 			}
9731 		}
9732 		ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9733 		membar_producer();
9734 		/*
9735 		 * lgrp_move_thread() won't schedule async recheck after
9736 		 * p->p_t1_lgrpid update unless p->p_tr_lgrpid is not
9737 		 * LGRP_NONE. Recheck p_t1_lgrpid once now that p->p_tr_lgrpid
9738 		 * is not LGRP_NONE.
9739 		 */
9740 		if (first && p->p_t1_lgrpid != LGRP_NONE &&
9741 		    p->p_t1_lgrpid != lgrp_id) {
9742 			first = 0;
9743 			goto again;
9744 		}
9745 	}
9746 	/*
9747 	 * If no amp was created yet for lgrp_id create a new one as long as
9748 	 * we have enough memory to afford it.
9749 	 */
9750 	if ((amp = svntrp->tr_amp[lgrp_id]) == NULL) {
9751 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9752 		if (trmem > segvn_textrepl_max_bytes) {
9753 			SEGVN_TR_ADDSTAT(normem);
9754 			goto fail;
9755 		}
9756 		if (anon_try_resv_zone(size, NULL) == 0) {
9757 			SEGVN_TR_ADDSTAT(noanon);
9758 			goto fail;
9759 		}
9760 		amp = anonmap_alloc(size, size, ANON_NOSLEEP);
9761 		if (amp == NULL) {
9762 			anon_unresv_zone(size, NULL);
9763 			SEGVN_TR_ADDSTAT(nokmem);
9764 			goto fail;
9765 		}
9766 		ASSERT(amp->refcnt == 1);
9767 		amp->a_szc = szc;
9768 		svntrp->tr_amp[lgrp_id] = amp;
9769 		SEGVN_TR_ADDSTAT(newamp);
9770 	}
9771 	svntrp->tr_refcnt++;
9772 	ASSERT(svd->svn_trnext == NULL);
9773 	ASSERT(svd->svn_trprev == NULL);
9774 	svd->svn_trnext = svntrp->tr_svnhead;
9775 	svd->svn_trprev = NULL;
9776 	if (svntrp->tr_svnhead != NULL) {
9777 		svntrp->tr_svnhead->svn_trprev = svd;
9778 	}
9779 	svntrp->tr_svnhead = svd;
9780 	ASSERT(amp->a_szc == szc && amp->size == size && amp->swresv == size);
9781 	ASSERT(amp->refcnt >= 1);
9782 	svd->amp = amp;
9783 	svd->anon_index = 0;
9784 	svd->tr_policy_info.mem_policy = LGRP_MEM_POLICY_NEXT_SEG;
9785 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
9786 	svd->tr_state = SEGVN_TR_ON;
9787 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9788 	SEGVN_TR_ADDSTAT(repl);
9789 	return;
9790 fail:
9791 	ASSERT(segvn_textrepl_bytes >= size);
9792 	atomic_add_long(&segvn_textrepl_bytes, -size);
9793 	ASSERT(svntrp != NULL);
9794 	ASSERT(svntrp->tr_amp[lgrp_id] == NULL);
9795 	if (svntrp->tr_refcnt == 0) {
9796 		ASSERT(svntrp == svntr_hashtab[hash].tr_head);
9797 		svntr_hashtab[hash].tr_head = svntrp->tr_next;
9798 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9799 		kmem_cache_free(svntr_cache, svntrp);
9800 	} else {
9801 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9802 	}
9803 	svd->tr_state = SEGVN_TR_OFF;
9804 }
9805 
9806 /*
9807  * Convert seg back to regular vnode mapping seg by unbinding it from its text
9808  * replication amp.  This routine is most typically called when segment is
9809  * unmapped but can also be called when segment no longer qualifies for text
9810  * replication (e.g. due to protection changes). If unload_unmap is set use
9811  * HAT_UNLOAD_UNMAP flag in hat_unload_callback().  If we are the last user of
9812  * svntr free all its anon maps and remove it from the hash table.
9813  */
9814 static void
9815 segvn_textunrepl(struct seg *seg, int unload_unmap)
9816 {
9817 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
9818 	vnode_t			*vp = svd->vp;
9819 	u_offset_t		off = svd->offset;
9820 	size_t			size = seg->s_size;
9821 	u_offset_t		eoff = off + size;
9822 	uint_t			szc = seg->s_szc;
9823 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9824 	svntr_t			*svntrp;
9825 	svntr_t			**prv_svntrp;
9826 	lgrp_id_t		lgrp_id = svd->tr_policy_info.mem_lgrpid;
9827 	lgrp_id_t		i;
9828 
9829 	ASSERT(AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
9830 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
9831 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9832 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9833 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9834 	ASSERT(svd->amp != NULL);
9835 	ASSERT(svd->amp->refcnt >= 1);
9836 	ASSERT(svd->anon_index == 0);
9837 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9838 	ASSERT(svntr_hashtab != NULL);
9839 
9840 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9841 	prv_svntrp = &svntr_hashtab[hash].tr_head;
9842 	for (; (svntrp = *prv_svntrp) != NULL; prv_svntrp = &svntrp->tr_next) {
9843 		ASSERT(svntrp->tr_refcnt != 0);
9844 		if (svntrp->tr_vp == vp && svntrp->tr_off == off &&
9845 		    svntrp->tr_eoff == eoff && svntrp->tr_szc == szc) {
9846 			break;
9847 		}
9848 	}
9849 	if (svntrp == NULL) {
9850 		panic("segvn_textunrepl: svntr record not found");
9851 	}
9852 	if (svntrp->tr_amp[lgrp_id] != svd->amp) {
9853 		panic("segvn_textunrepl: amp mismatch");
9854 	}
9855 	svd->tr_state = SEGVN_TR_OFF;
9856 	svd->amp = NULL;
9857 	if (svd->svn_trprev == NULL) {
9858 		ASSERT(svntrp->tr_svnhead == svd);
9859 		svntrp->tr_svnhead = svd->svn_trnext;
9860 		if (svntrp->tr_svnhead != NULL) {
9861 			svntrp->tr_svnhead->svn_trprev = NULL;
9862 		}
9863 		svd->svn_trnext = NULL;
9864 	} else {
9865 		svd->svn_trprev->svn_trnext = svd->svn_trnext;
9866 		if (svd->svn_trnext != NULL) {
9867 			svd->svn_trnext->svn_trprev = svd->svn_trprev;
9868 			svd->svn_trnext = NULL;
9869 		}
9870 		svd->svn_trprev = NULL;
9871 	}
9872 	if (--svntrp->tr_refcnt) {
9873 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9874 		goto done;
9875 	}
9876 	*prv_svntrp = svntrp->tr_next;
9877 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9878 	for (i = 0; i < NLGRPS_MAX; i++) {
9879 		struct anon_map *amp = svntrp->tr_amp[i];
9880 		if (amp == NULL) {
9881 			continue;
9882 		}
9883 		ASSERT(amp->refcnt == 1);
9884 		ASSERT(amp->swresv == size);
9885 		ASSERT(amp->size == size);
9886 		ASSERT(amp->a_szc == szc);
9887 		if (amp->a_szc != 0) {
9888 			anon_free_pages(amp->ahp, 0, size, szc);
9889 		} else {
9890 			anon_free(amp->ahp, 0, size);
9891 		}
9892 		svntrp->tr_amp[i] = NULL;
9893 		ASSERT(segvn_textrepl_bytes >= size);
9894 		atomic_add_long(&segvn_textrepl_bytes, -size);
9895 		anon_unresv_zone(amp->swresv, NULL);
9896 		amp->refcnt = 0;
9897 		anonmap_free(amp);
9898 	}
9899 	kmem_cache_free(svntr_cache, svntrp);
9900 done:
9901 	hat_unload_callback(seg->s_as->a_hat, seg->s_base, size,
9902 	    unload_unmap ? HAT_UNLOAD_UNMAP : 0, NULL);
9903 }
9904 
9905 /*
9906  * This is called when a MAP_SHARED writable mapping is created to a vnode
9907  * that is currently used for execution (VVMEXEC flag is set). In this case we
9908  * need to prevent further use of existing replicas.
9909  */
9910 static void
9911 segvn_inval_trcache(vnode_t *vp)
9912 {
9913 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9914 	svntr_t			*svntrp;
9915 
9916 	ASSERT(vp->v_flag & VVMEXEC);
9917 
9918 	if (svntr_hashtab == NULL) {
9919 		return;
9920 	}
9921 
9922 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9923 	svntrp = svntr_hashtab[hash].tr_head;
9924 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9925 		ASSERT(svntrp->tr_refcnt != 0);
9926 		if (svntrp->tr_vp == vp && svntrp->tr_valid) {
9927 			svntrp->tr_valid = 0;
9928 		}
9929 	}
9930 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9931 }
9932 
9933 static void
9934 segvn_trasync_thread(void)
9935 {
9936 	callb_cpr_t cpr_info;
9937 	kmutex_t cpr_lock;	/* just for CPR stuff */
9938 
9939 	mutex_init(&cpr_lock, NULL, MUTEX_DEFAULT, NULL);
9940 
9941 	CALLB_CPR_INIT(&cpr_info, &cpr_lock,
9942 	    callb_generic_cpr, "segvn_async");
9943 
9944 	if (segvn_update_textrepl_interval == 0) {
9945 		segvn_update_textrepl_interval = segvn_update_tr_time * hz;
9946 	} else {
9947 		segvn_update_textrepl_interval *= hz;
9948 	}
9949 	(void) timeout(segvn_trupdate_wakeup, NULL,
9950 	    segvn_update_textrepl_interval);
9951 
9952 	for (;;) {
9953 		mutex_enter(&cpr_lock);
9954 		CALLB_CPR_SAFE_BEGIN(&cpr_info);
9955 		mutex_exit(&cpr_lock);
9956 		sema_p(&segvn_trasync_sem);
9957 		mutex_enter(&cpr_lock);
9958 		CALLB_CPR_SAFE_END(&cpr_info, &cpr_lock);
9959 		mutex_exit(&cpr_lock);
9960 		segvn_trupdate();
9961 	}
9962 }
9963 
9964 static uint64_t segvn_lgrp_trthr_migrs_snpsht = 0;
9965 
9966 static void
9967 segvn_trupdate_wakeup(void *dummy)
9968 {
9969 	uint64_t cur_lgrp_trthr_migrs = lgrp_get_trthr_migrations();
9970 
9971 	if (cur_lgrp_trthr_migrs != segvn_lgrp_trthr_migrs_snpsht) {
9972 		segvn_lgrp_trthr_migrs_snpsht = cur_lgrp_trthr_migrs;
9973 		sema_v(&segvn_trasync_sem);
9974 	}
9975 
9976 	if (!segvn_disable_textrepl_update &&
9977 	    segvn_update_textrepl_interval != 0) {
9978 		(void) timeout(segvn_trupdate_wakeup, dummy,
9979 		    segvn_update_textrepl_interval);
9980 	}
9981 }
9982 
9983 static void
9984 segvn_trupdate(void)
9985 {
9986 	ulong_t		hash;
9987 	svntr_t		*svntrp;
9988 	segvn_data_t	*svd;
9989 
9990 	ASSERT(svntr_hashtab != NULL);
9991 
9992 	for (hash = 0; hash < svntr_hashtab_sz; hash++) {
9993 		mutex_enter(&svntr_hashtab[hash].tr_lock);
9994 		svntrp = svntr_hashtab[hash].tr_head;
9995 		for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9996 			ASSERT(svntrp->tr_refcnt != 0);
9997 			svd = svntrp->tr_svnhead;
9998 			for (; svd != NULL; svd = svd->svn_trnext) {
9999 				segvn_trupdate_seg(svd->seg, svd, svntrp,
10000 				    hash);
10001 			}
10002 		}
10003 		mutex_exit(&svntr_hashtab[hash].tr_lock);
10004 	}
10005 }
10006 
10007 static void
10008 segvn_trupdate_seg(struct seg *seg,
10009 	segvn_data_t *svd,
10010 	svntr_t *svntrp,
10011 	ulong_t hash)
10012 {
10013 	proc_t			*p;
10014 	lgrp_id_t		lgrp_id;
10015 	struct as		*as;
10016 	size_t			size;
10017 	struct anon_map		*amp;
10018 
10019 	ASSERT(svd->vp != NULL);
10020 	ASSERT(svd->vp == svntrp->tr_vp);
10021 	ASSERT(svd->offset == svntrp->tr_off);
10022 	ASSERT(svd->offset + seg->s_size == svntrp->tr_eoff);
10023 	ASSERT(seg != NULL);
10024 	ASSERT(svd->seg == seg);
10025 	ASSERT(seg->s_data == (void *)svd);
10026 	ASSERT(seg->s_szc == svntrp->tr_szc);
10027 	ASSERT(svd->tr_state == SEGVN_TR_ON);
10028 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
10029 	ASSERT(svd->amp != NULL);
10030 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
10031 	ASSERT(svd->tr_policy_info.mem_lgrpid != LGRP_NONE);
10032 	ASSERT(svd->tr_policy_info.mem_lgrpid < NLGRPS_MAX);
10033 	ASSERT(svntrp->tr_amp[svd->tr_policy_info.mem_lgrpid] == svd->amp);
10034 	ASSERT(svntrp->tr_refcnt != 0);
10035 	ASSERT(mutex_owned(&svntr_hashtab[hash].tr_lock));
10036 
10037 	as = seg->s_as;
10038 	ASSERT(as != NULL && as != &kas);
10039 	p = as->a_proc;
10040 	ASSERT(p != NULL);
10041 	ASSERT(p->p_tr_lgrpid != LGRP_NONE);
10042 	lgrp_id = p->p_t1_lgrpid;
10043 	if (lgrp_id == LGRP_NONE) {
10044 		return;
10045 	}
10046 	ASSERT(lgrp_id < NLGRPS_MAX);
10047 	if (svd->tr_policy_info.mem_lgrpid == lgrp_id) {
10048 		return;
10049 	}
10050 
10051 	/*
10052 	 * Use tryenter locking since we are locking as/seg and svntr hash
10053 	 * lock in reverse from syncrounous thread order.
10054 	 */
10055 	if (!AS_LOCK_TRYENTER(as, &as->a_lock, RW_READER)) {
10056 		SEGVN_TR_ADDSTAT(nolock);
10057 		if (segvn_lgrp_trthr_migrs_snpsht) {
10058 			segvn_lgrp_trthr_migrs_snpsht = 0;
10059 		}
10060 		return;
10061 	}
10062 	if (!SEGVN_LOCK_TRYENTER(seg->s_as, &svd->lock, RW_WRITER)) {
10063 		AS_LOCK_EXIT(as, &as->a_lock);
10064 		SEGVN_TR_ADDSTAT(nolock);
10065 		if (segvn_lgrp_trthr_migrs_snpsht) {
10066 			segvn_lgrp_trthr_migrs_snpsht = 0;
10067 		}
10068 		return;
10069 	}
10070 	size = seg->s_size;
10071 	if (svntrp->tr_amp[lgrp_id] == NULL) {
10072 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
10073 		if (trmem > segvn_textrepl_max_bytes) {
10074 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10075 			AS_LOCK_EXIT(as, &as->a_lock);
10076 			atomic_add_long(&segvn_textrepl_bytes, -size);
10077 			SEGVN_TR_ADDSTAT(normem);
10078 			return;
10079 		}
10080 		if (anon_try_resv_zone(size, NULL) == 0) {
10081 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10082 			AS_LOCK_EXIT(as, &as->a_lock);
10083 			atomic_add_long(&segvn_textrepl_bytes, -size);
10084 			SEGVN_TR_ADDSTAT(noanon);
10085 			return;
10086 		}
10087 		amp = anonmap_alloc(size, size, KM_NOSLEEP);
10088 		if (amp == NULL) {
10089 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10090 			AS_LOCK_EXIT(as, &as->a_lock);
10091 			atomic_add_long(&segvn_textrepl_bytes, -size);
10092 			anon_unresv_zone(size, NULL);
10093 			SEGVN_TR_ADDSTAT(nokmem);
10094 			return;
10095 		}
10096 		ASSERT(amp->refcnt == 1);
10097 		amp->a_szc = seg->s_szc;
10098 		svntrp->tr_amp[lgrp_id] = amp;
10099 	}
10100 	/*
10101 	 * We don't need to drop the bucket lock but here we give other
10102 	 * threads a chance.  svntr and svd can't be unlinked as long as
10103 	 * segment lock is held as a writer and AS held as well.  After we
10104 	 * retake bucket lock we'll continue from where we left. We'll be able
10105 	 * to reach the end of either list since new entries are always added
10106 	 * to the beginning of the lists.
10107 	 */
10108 	mutex_exit(&svntr_hashtab[hash].tr_lock);
10109 	hat_unload_callback(as->a_hat, seg->s_base, size, 0, NULL);
10110 	mutex_enter(&svntr_hashtab[hash].tr_lock);
10111 
10112 	ASSERT(svd->tr_state == SEGVN_TR_ON);
10113 	ASSERT(svd->amp != NULL);
10114 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
10115 	ASSERT(svd->tr_policy_info.mem_lgrpid != lgrp_id);
10116 	ASSERT(svd->amp != svntrp->tr_amp[lgrp_id]);
10117 
10118 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
10119 	svd->amp = svntrp->tr_amp[lgrp_id];
10120 	p->p_tr_lgrpid = NLGRPS_MAX;
10121 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10122 	AS_LOCK_EXIT(as, &as->a_lock);
10123 
10124 	ASSERT(svntrp->tr_refcnt != 0);
10125 	ASSERT(svd->vp == svntrp->tr_vp);
10126 	ASSERT(svd->tr_policy_info.mem_lgrpid == lgrp_id);
10127 	ASSERT(svd->amp != NULL && svd->amp == svntrp->tr_amp[lgrp_id]);
10128 	ASSERT(svd->seg == seg);
10129 	ASSERT(svd->tr_state == SEGVN_TR_ON);
10130 
10131 	SEGVN_TR_ADDSTAT(asyncrepl);
10132 }
10133