xref: /illumos-gate/usr/src/uts/common/vm/seg_spt.c (revision 528737823843346cf95a4a701612f82089135554)
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) 1993, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright 2018 Joyent, Inc.
24  * Copyright (c) 2016 by Delphix. All rights reserved.
25  */
26 
27 #include <sys/param.h>
28 #include <sys/user.h>
29 #include <sys/mman.h>
30 #include <sys/kmem.h>
31 #include <sys/sysmacros.h>
32 #include <sys/cmn_err.h>
33 #include <sys/systm.h>
34 #include <sys/tuneable.h>
35 #include <vm/hat.h>
36 #include <vm/seg.h>
37 #include <vm/as.h>
38 #include <vm/anon.h>
39 #include <vm/page.h>
40 #include <sys/buf.h>
41 #include <sys/swap.h>
42 #include <sys/atomic.h>
43 #include <vm/seg_spt.h>
44 #include <sys/debug.h>
45 #include <sys/vtrace.h>
46 #include <sys/shm.h>
47 #include <sys/shm_impl.h>
48 #include <sys/lgrp.h>
49 #include <sys/vmsystm.h>
50 #include <sys/policy.h>
51 #include <sys/project.h>
52 #include <sys/tnf_probe.h>
53 #include <sys/zone.h>
54 
55 #define	SEGSPTADDR	(caddr_t)0x0
56 
57 /*
58  * # pages used for spt
59  */
60 size_t	spt_used;
61 
62 /*
63  * segspt_minfree is the memory left for system after ISM
64  * locked its pages; it is set up to 5% of availrmem in
65  * sptcreate when ISM is created.  ISM should not use more
66  * than ~90% of availrmem; if it does, then the performance
67  * of the system may decrease. Machines with large memories may
68  * be able to use up more memory for ISM so we set the default
69  * segspt_minfree to 5% (which gives ISM max 95% of availrmem.
70  * If somebody wants even more memory for ISM (risking hanging
71  * the system) they can patch the segspt_minfree to smaller number.
72  */
73 pgcnt_t segspt_minfree = 0;
74 
75 static int segspt_create(struct seg **segpp, void *argsp);
76 static int segspt_unmap(struct seg *seg, caddr_t raddr, size_t ssize);
77 static void segspt_free(struct seg *seg);
78 static void segspt_free_pages(struct seg *seg, caddr_t addr, size_t len);
79 static lgrp_mem_policy_info_t *segspt_getpolicy(struct seg *seg, caddr_t addr);
80 
81 /* ARGSUSED */
82 __NORETURN static int
83 segspt_badop_dup(struct seg *seg __unused, struct seg *newseg __unused)
84 {
85 	panic("%s called", __func__);
86 }
87 
88 /* ARGSUSED */
89 __NORETURN static faultcode_t
90 segspt_badop_fault(struct hat *hat, struct seg *seg, caddr_t addr,
91     size_t len, enum fault_type type, enum seg_rw rw)
92 {
93 	panic("%s called", __func__);
94 }
95 
96 /* ARGSUSED */
97 __NORETURN static faultcode_t
98 segspt_badop_faulta(struct seg *seg __unused, caddr_t addr __unused)
99 {
100 	panic("%s called", __func__);
101 }
102 
103 /* ARGSUSED */
104 __NORETURN static int
105 segspt_badop_prot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
106 {
107 	panic("%s called", __func__);
108 }
109 
110 /* ARGSUSED */
111 __NORETURN static int
112 segspt_badop_checkprot(struct seg *seg, caddr_t addr, size_t size, uint_t prot)
113 {
114 	panic("%s called", __func__);
115 }
116 
117 /* ARGSUSED */
118 __NORETURN static int
119 segspt_badop_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
120 {
121 	panic("%s called", __func__);
122 }
123 
124 /* ARGSUSED */
125 __NORETURN static size_t
126 segspt_badop_swapout(struct seg *seg)
127 {
128 	panic("%s called", __func__);
129 }
130 
131 /* ARGSUSED */
132 __NORETURN static int
133 segspt_badop_sync(struct seg *seg, caddr_t addr, size_t len, int attr,
134     uint_t flags)
135 {
136 	panic("%s called", __func__);
137 }
138 
139 /* ARGSUSED */
140 __NORETURN
141 static size_t
142 segspt_badop_incore(struct seg *seg, caddr_t addr, size_t len, char *vec)
143 {
144 	panic("%s called", __func__);
145 }
146 
147 /* ARGSUSED */
148 __NORETURN static int
149 segspt_badop_lockop(struct seg *seg, caddr_t addr, size_t len, int attr,
150     int op, ulong_t *lockmap, size_t pos)
151 {
152 	panic("%s called", __func__);
153 }
154 
155 /* ARGSUSED */
156 __NORETURN static int
157 segspt_badop_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
158 {
159 	panic("%s called", __func__);
160 }
161 
162 /* ARGSUSED */
163 __NORETURN static u_offset_t
164 segspt_badop_getoffset(struct seg *seg, caddr_t addr)
165 {
166 	panic("%s called", __func__);
167 }
168 
169 /* ARGSUSED */
170 __NORETURN static int
171 segspt_badop_gettype(struct seg *seg, caddr_t addr)
172 {
173 	panic("%s called", __func__);
174 }
175 
176 /* ARGSUSED */
177 __NORETURN static int
178 segspt_badop_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
179 {
180 	panic("%s called", __func__);
181 }
182 
183 /* ARGSUSED */
184 __NORETURN static int
185 segspt_badop_advise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
186 {
187 	panic("%s called", __func__);
188 }
189 
190 /* ARGSUSED */
191 __NORETURN static void
192 segspt_badop_dump(struct seg *seg)
193 {
194 	panic("%s called", __func__);
195 }
196 
197 /* ARGSUSED */
198 __NORETURN static int
199 segspt_badop_pagelock(struct seg *seg, caddr_t addr, size_t len,
200     struct page ***ppp, enum lock_type type, enum seg_rw rw)
201 {
202 	panic("%s called", __func__);
203 }
204 
205 /* ARGSUSED */
206 __NORETURN static int
207 segspt_badop_setpgsz(struct seg *seg, caddr_t addr, size_t len, uint_t szc)
208 {
209 	panic("%s called", __func__);
210 }
211 
212 /* ARGSUSED */
213 __NORETURN static int
214 segspt_badop_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
215 {
216 	panic("%s called", __func__);
217 }
218 
219 /* ARGSUSED */
220 __NORETURN static int
221 segspt_badop_capable(struct seg *seg, segcapability_t capability)
222 {
223 	panic("%s called", __func__);
224 }
225 
226 struct seg_ops segspt_ops = {
227 	segspt_badop_dup,		/* dup */
228 	segspt_unmap,
229 	segspt_free,
230 	segspt_badop_fault,		/* fault */
231 	segspt_badop_faulta,		/* faulta */
232 	segspt_badop_prot,		/* setprot */
233 	segspt_badop_checkprot,		/* checkprot */
234 	segspt_badop_kluster,		/* kluster */
235 	segspt_badop_swapout,		/* swapout */
236 	segspt_badop_sync,		/* sync */
237 	segspt_badop_incore,		/* incore */
238 	segspt_badop_lockop,		/* lockop */
239 	segspt_badop_getprot,		/* getprot */
240 	segspt_badop_getoffset,		/* getoffset */
241 	segspt_badop_gettype,		/* gettype */
242 	segspt_badop_getvp,		/* getvp */
243 	segspt_badop_advise,		/* advise */
244 	segspt_badop_dump,		/* dump */
245 	segspt_badop_pagelock,		/* pagelock */
246 	segspt_badop_setpgsz,		/* setpgsz */
247 	segspt_badop_getmemid,		/* getmemid */
248 	segspt_getpolicy,		/* getpolicy */
249 	segspt_badop_capable,		/* capable */
250 	seg_inherit_notsup		/* inherit */
251 };
252 
253 static int segspt_shmdup(struct seg *seg, struct seg *newseg);
254 static int segspt_shmunmap(struct seg *seg, caddr_t raddr, size_t ssize);
255 static void segspt_shmfree(struct seg *seg);
256 static faultcode_t segspt_shmfault(struct hat *hat, struct seg *seg,
257 		caddr_t addr, size_t len, enum fault_type type, enum seg_rw rw);
258 static faultcode_t segspt_shmfaulta(struct seg *seg, caddr_t addr);
259 static int segspt_shmsetprot(struct seg *seg, caddr_t addr, size_t len,
260 		uint_t prot);
261 static int segspt_shmcheckprot(struct seg *seg, caddr_t addr, size_t size,
262 		uint_t prot);
263 static int	segspt_shmkluster(struct seg *seg, caddr_t addr, ssize_t delta);
264 static size_t	segspt_shmswapout(struct seg *seg);
265 static size_t segspt_shmincore(struct seg *seg, caddr_t addr, size_t len,
266 		char *vec);
267 static int segspt_shmsync(struct seg *seg, caddr_t addr, size_t len,
268 		int attr, uint_t flags);
269 static int segspt_shmlockop(struct seg *seg, caddr_t addr, size_t len,
270 		int attr, int op, ulong_t *lockmap, size_t pos);
271 static int segspt_shmgetprot(struct seg *seg, caddr_t addr, size_t len,
272 		uint_t *protv);
273 static u_offset_t segspt_shmgetoffset(struct seg *seg, caddr_t addr);
274 static int segspt_shmgettype(struct seg *seg, caddr_t addr);
275 static int segspt_shmgetvp(struct seg *seg, caddr_t addr, struct vnode **vpp);
276 static int segspt_shmadvise(struct seg *seg, caddr_t addr, size_t len,
277 		uint_t behav);
278 static void segspt_shmdump(struct seg *seg);
279 static int segspt_shmpagelock(struct seg *, caddr_t, size_t,
280 		struct page ***, enum lock_type, enum seg_rw);
281 static int segspt_shmsetpgsz(struct seg *, caddr_t, size_t, uint_t);
282 static int segspt_shmgetmemid(struct seg *, caddr_t, memid_t *);
283 static lgrp_mem_policy_info_t *segspt_shmgetpolicy(struct seg *, caddr_t);
284 static int segspt_shmcapable(struct seg *, segcapability_t);
285 
286 struct seg_ops segspt_shmops = {
287 	segspt_shmdup,
288 	segspt_shmunmap,
289 	segspt_shmfree,
290 	segspt_shmfault,
291 	segspt_shmfaulta,
292 	segspt_shmsetprot,
293 	segspt_shmcheckprot,
294 	segspt_shmkluster,
295 	segspt_shmswapout,
296 	segspt_shmsync,
297 	segspt_shmincore,
298 	segspt_shmlockop,
299 	segspt_shmgetprot,
300 	segspt_shmgetoffset,
301 	segspt_shmgettype,
302 	segspt_shmgetvp,
303 	segspt_shmadvise,	/* advise */
304 	segspt_shmdump,
305 	segspt_shmpagelock,
306 	segspt_shmsetpgsz,
307 	segspt_shmgetmemid,
308 	segspt_shmgetpolicy,
309 	segspt_shmcapable,
310 	seg_inherit_notsup
311 };
312 
313 static void segspt_purge(struct seg *seg);
314 static int segspt_reclaim(void *, caddr_t, size_t, struct page **,
315 		enum seg_rw, int);
316 static int spt_anon_getpages(struct seg *seg, caddr_t addr, size_t len,
317 		page_t **ppa);
318 
319 
320 
321 /*ARGSUSED*/
322 int
323 sptcreate(size_t size, struct seg **sptseg, struct anon_map *amp,
324     uint_t prot, uint_t flags, uint_t share_szc)
325 {
326 	int	err;
327 	struct	as	*newas;
328 	struct	segspt_crargs sptcargs;
329 
330 #ifdef DEBUG
331 	TNF_PROBE_1(sptcreate, "spt", /* CSTYLED */,
332 			tnf_ulong, size, size );
333 #endif
334 	if (segspt_minfree == 0)	/* leave min 5% of availrmem for */
335 		segspt_minfree = availrmem/20;	/* for the system */
336 
337 	if (!hat_supported(HAT_SHARED_PT, (void *)0))
338 		return (EINVAL);
339 
340 	/*
341 	 * get a new as for this shared memory segment
342 	 */
343 	newas = as_alloc();
344 	newas->a_proc = NULL;
345 	sptcargs.amp = amp;
346 	sptcargs.prot = prot;
347 	sptcargs.flags = flags;
348 	sptcargs.szc = share_szc;
349 	/*
350 	 * create a shared page table (spt) segment
351 	 */
352 
353 	if (err = as_map(newas, SEGSPTADDR, size, segspt_create, &sptcargs)) {
354 		as_free(newas);
355 		return (err);
356 	}
357 	*sptseg = sptcargs.seg_spt;
358 	return (0);
359 }
360 
361 void
362 sptdestroy(struct as *as, struct anon_map *amp)
363 {
364 
365 #ifdef DEBUG
366 	TNF_PROBE_0(sptdestroy, "spt", /* CSTYLED */);
367 #endif
368 	(void) as_unmap(as, SEGSPTADDR, amp->size);
369 	as_free(as);
370 }
371 
372 /*
373  * called from seg_free().
374  * free (i.e., unlock, unmap, return to free list)
375  *  all the pages in the given seg.
376  */
377 void
378 segspt_free(struct seg	*seg)
379 {
380 	struct spt_data *sptd = (struct spt_data *)seg->s_data;
381 
382 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
383 
384 	if (sptd != NULL) {
385 		if (sptd->spt_realsize)
386 			segspt_free_pages(seg, seg->s_base, sptd->spt_realsize);
387 
388 		if (sptd->spt_ppa_lckcnt) {
389 			kmem_free(sptd->spt_ppa_lckcnt,
390 			    sizeof (*sptd->spt_ppa_lckcnt)
391 			    * btopr(sptd->spt_amp->size));
392 		}
393 		kmem_free(sptd->spt_vp, sizeof (*sptd->spt_vp));
394 		cv_destroy(&sptd->spt_cv);
395 		mutex_destroy(&sptd->spt_lock);
396 		kmem_free(sptd, sizeof (*sptd));
397 	}
398 }
399 
400 /*ARGSUSED*/
401 static int
402 segspt_shmsync(struct seg *seg, caddr_t addr, size_t len, int attr,
403     uint_t flags)
404 {
405 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
406 
407 	return (0);
408 }
409 
410 /*ARGSUSED*/
411 static size_t
412 segspt_shmincore(struct seg *seg, caddr_t addr, size_t len, char *vec)
413 {
414 	caddr_t	eo_seg;
415 	pgcnt_t	npages;
416 	struct shm_data *shmd = (struct shm_data *)seg->s_data;
417 	struct seg	*sptseg;
418 	struct spt_data *sptd;
419 
420 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
421 #ifdef lint
422 	seg = seg;
423 #endif
424 	sptseg = shmd->shm_sptseg;
425 	sptd = sptseg->s_data;
426 
427 	if ((sptd->spt_flags & SHM_PAGEABLE) == 0) {
428 		eo_seg = addr + len;
429 		while (addr < eo_seg) {
430 			/* page exists, and it's locked. */
431 			*vec++ = SEG_PAGE_INCORE | SEG_PAGE_LOCKED |
432 			    SEG_PAGE_ANON;
433 			addr += PAGESIZE;
434 		}
435 		return (len);
436 	} else {
437 		struct  anon_map *amp = shmd->shm_amp;
438 		struct  anon	*ap;
439 		page_t		*pp;
440 		pgcnt_t		anon_index;
441 		struct vnode	*vp;
442 		u_offset_t	off;
443 		ulong_t		i;
444 		int		ret;
445 		anon_sync_obj_t	cookie;
446 
447 		addr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
448 		anon_index = seg_page(seg, addr);
449 		npages = btopr(len);
450 		if (anon_index + npages > btopr(shmd->shm_amp->size)) {
451 			return (EINVAL);
452 		}
453 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
454 		for (i = 0; i < npages; i++, anon_index++) {
455 			ret = 0;
456 			anon_array_enter(amp, anon_index, &cookie);
457 			ap = anon_get_ptr(amp->ahp, anon_index);
458 			if (ap != NULL) {
459 				swap_xlate(ap, &vp, &off);
460 				anon_array_exit(&cookie);
461 				pp = page_lookup_nowait(vp, off, SE_SHARED);
462 				if (pp != NULL) {
463 					ret |= SEG_PAGE_INCORE | SEG_PAGE_ANON;
464 					page_unlock(pp);
465 				}
466 			} else {
467 				anon_array_exit(&cookie);
468 			}
469 			if (shmd->shm_vpage[anon_index] & DISM_PG_LOCKED) {
470 				ret |= SEG_PAGE_LOCKED;
471 			}
472 			*vec++ = (char)ret;
473 		}
474 		ANON_LOCK_EXIT(&amp->a_rwlock);
475 		return (len);
476 	}
477 }
478 
479 static int
480 segspt_unmap(struct seg *seg, caddr_t raddr, size_t ssize)
481 {
482 	size_t share_size;
483 
484 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
485 
486 	/*
487 	 * seg.s_size may have been rounded up to the largest page size
488 	 * in shmat().
489 	 * XXX This should be cleanedup. sptdestroy should take a length
490 	 * argument which should be the same as sptcreate. Then
491 	 * this rounding would not be needed (or is done in shm.c)
492 	 * Only the check for full segment will be needed.
493 	 *
494 	 * XXX -- shouldn't raddr == 0 always? These tests don't seem
495 	 * to be useful at all.
496 	 */
497 	share_size = page_get_pagesize(seg->s_szc);
498 	ssize = P2ROUNDUP(ssize, share_size);
499 
500 	if (raddr == seg->s_base && ssize == seg->s_size) {
501 		seg_free(seg);
502 		return (0);
503 	} else
504 		return (EINVAL);
505 }
506 
507 int
508 segspt_create(struct seg **segpp, void *argsp)
509 {
510 	struct seg	*seg = *segpp;
511 	int		err;
512 	caddr_t		addr = seg->s_base;
513 	struct spt_data *sptd;
514 	struct segspt_crargs *sptcargs = (struct segspt_crargs *)argsp;
515 	struct anon_map *amp = sptcargs->amp;
516 	struct kshmid	*sp = amp->a_sp;
517 	struct	cred	*cred = CRED();
518 	ulong_t		i, j, anon_index = 0;
519 	pgcnt_t		npages = btopr(amp->size);
520 	struct vnode	*vp;
521 	page_t		**ppa;
522 	uint_t		hat_flags;
523 	size_t		pgsz;
524 	pgcnt_t		pgcnt;
525 	caddr_t		a;
526 	pgcnt_t		pidx;
527 	size_t		sz;
528 	proc_t		*procp = curproc;
529 	rctl_qty_t	lockedbytes = 0;
530 	kproject_t	*proj;
531 
532 	/*
533 	 * We are holding the a_lock on the underlying dummy as,
534 	 * so we can make calls to the HAT layer.
535 	 */
536 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
537 	ASSERT(sp != NULL);
538 
539 #ifdef DEBUG
540 	TNF_PROBE_2(segspt_create, "spt", /* CSTYLED */,
541 	    tnf_opaque, addr, addr, tnf_ulong, len, seg->s_size);
542 #endif
543 	if ((sptcargs->flags & SHM_PAGEABLE) == 0) {
544 		if (err = anon_swap_adjust(npages))
545 			return (err);
546 	}
547 	err = ENOMEM;
548 
549 	if ((sptd = kmem_zalloc(sizeof (*sptd), KM_NOSLEEP)) == NULL)
550 		goto out1;
551 
552 	ppa = NULL;
553 	if ((sptcargs->flags & SHM_PAGEABLE) == 0) {
554 		if ((ppa = kmem_zalloc(((sizeof (page_t *)) * npages),
555 		    KM_NOSLEEP)) == NULL)
556 			goto out2;
557 	}
558 
559 	mutex_init(&sptd->spt_lock, NULL, MUTEX_DEFAULT, NULL);
560 
561 	if ((vp = kmem_zalloc(sizeof (*vp), KM_NOSLEEP)) == NULL)
562 		goto out3;
563 
564 	seg->s_ops = &segspt_ops;
565 	sptd->spt_vp = vp;
566 	sptd->spt_amp = amp;
567 	sptd->spt_prot = sptcargs->prot;
568 	sptd->spt_flags = sptcargs->flags;
569 	seg->s_data = (caddr_t)sptd;
570 	sptd->spt_ppa = NULL;
571 	sptd->spt_ppa_lckcnt = NULL;
572 	seg->s_szc = sptcargs->szc;
573 	cv_init(&sptd->spt_cv, NULL, CV_DEFAULT, NULL);
574 	sptd->spt_gen = 0;
575 
576 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
577 	if (seg->s_szc > amp->a_szc) {
578 		amp->a_szc = seg->s_szc;
579 	}
580 	ANON_LOCK_EXIT(&amp->a_rwlock);
581 
582 	/*
583 	 * Set policy to affect initial allocation of pages in
584 	 * anon_map_createpages()
585 	 */
586 	(void) lgrp_shm_policy_set(LGRP_MEM_POLICY_DEFAULT, amp, anon_index,
587 	    NULL, 0, ptob(npages));
588 
589 	if (sptcargs->flags & SHM_PAGEABLE) {
590 		size_t  share_sz;
591 		pgcnt_t new_npgs, more_pgs;
592 		struct anon_hdr *nahp;
593 		zone_t *zone;
594 
595 		share_sz = page_get_pagesize(seg->s_szc);
596 		if (!IS_P2ALIGNED(amp->size, share_sz)) {
597 			/*
598 			 * We are rounding up the size of the anon array
599 			 * on 4 M boundary because we always create 4 M
600 			 * of page(s) when locking, faulting pages and we
601 			 * don't have to check for all corner cases e.g.
602 			 * if there is enough space to allocate 4 M
603 			 * page.
604 			 */
605 			new_npgs = btop(P2ROUNDUP(amp->size, share_sz));
606 			more_pgs = new_npgs - npages;
607 
608 			/*
609 			 * The zone will never be NULL, as a fully created
610 			 * shm always has an owning zone.
611 			 */
612 			zone = sp->shm_perm.ipc_zone_ref.zref_zone;
613 			ASSERT(zone != NULL);
614 			if (anon_resv_zone(ptob(more_pgs), zone) == 0) {
615 				err = ENOMEM;
616 				goto out4;
617 			}
618 
619 			nahp = anon_create(new_npgs, ANON_SLEEP);
620 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
621 			(void) anon_copy_ptr(amp->ahp, 0, nahp, 0, npages,
622 			    ANON_SLEEP);
623 			anon_release(amp->ahp, npages);
624 			amp->ahp = nahp;
625 			ASSERT(amp->swresv == ptob(npages));
626 			amp->swresv = amp->size = ptob(new_npgs);
627 			ANON_LOCK_EXIT(&amp->a_rwlock);
628 			npages = new_npgs;
629 		}
630 
631 		sptd->spt_ppa_lckcnt = kmem_zalloc(npages *
632 		    sizeof (*sptd->spt_ppa_lckcnt), KM_SLEEP);
633 		sptd->spt_pcachecnt = 0;
634 		sptd->spt_realsize = ptob(npages);
635 		sptcargs->seg_spt = seg;
636 		return (0);
637 	}
638 
639 	/*
640 	 * get array of pages for each anon slot in amp
641 	 */
642 	if ((err = anon_map_createpages(amp, anon_index, ptob(npages), ppa,
643 	    seg, addr, S_CREATE, cred)) != 0)
644 		goto out4;
645 
646 	mutex_enter(&sp->shm_mlock);
647 
648 	/* May be partially locked, so, count bytes to charge for locking */
649 	for (i = 0; i < npages; i++)
650 		if (ppa[i]->p_lckcnt == 0)
651 			lockedbytes += PAGESIZE;
652 
653 	proj = sp->shm_perm.ipc_proj;
654 
655 	if (lockedbytes > 0) {
656 		mutex_enter(&procp->p_lock);
657 		if (rctl_incr_locked_mem(procp, proj, lockedbytes, 0)) {
658 			mutex_exit(&procp->p_lock);
659 			mutex_exit(&sp->shm_mlock);
660 			for (i = 0; i < npages; i++)
661 				page_unlock(ppa[i]);
662 			err = ENOMEM;
663 			goto out4;
664 		}
665 		mutex_exit(&procp->p_lock);
666 	}
667 
668 	/*
669 	 * addr is initial address corresponding to the first page on ppa list
670 	 */
671 	for (i = 0; i < npages; i++) {
672 		/* attempt to lock all pages */
673 		if (page_pp_lock(ppa[i], 0, 1) == 0) {
674 			/*
675 			 * if unable to lock any page, unlock all
676 			 * of them and return error
677 			 */
678 			for (j = 0; j < i; j++)
679 				page_pp_unlock(ppa[j], 0, 1);
680 			for (i = 0; i < npages; i++)
681 				page_unlock(ppa[i]);
682 			rctl_decr_locked_mem(NULL, proj, lockedbytes, 0);
683 			mutex_exit(&sp->shm_mlock);
684 			err = ENOMEM;
685 			goto out4;
686 		}
687 	}
688 	mutex_exit(&sp->shm_mlock);
689 
690 	/*
691 	 * Some platforms assume that ISM mappings are HAT_LOAD_LOCK
692 	 * for the entire life of the segment. For example platforms
693 	 * that do not support Dynamic Reconfiguration.
694 	 */
695 	hat_flags = HAT_LOAD_SHARE;
696 	if (!hat_supported(HAT_DYNAMIC_ISM_UNMAP, NULL))
697 		hat_flags |= HAT_LOAD_LOCK;
698 
699 	/*
700 	 * Load translations one lare page at a time
701 	 * to make sure we don't create mappings bigger than
702 	 * segment's size code in case underlying pages
703 	 * are shared with segvn's segment that uses bigger
704 	 * size code than we do.
705 	 */
706 	pgsz = page_get_pagesize(seg->s_szc);
707 	pgcnt = page_get_pagecnt(seg->s_szc);
708 	for (a = addr, pidx = 0; pidx < npages; a += pgsz, pidx += pgcnt) {
709 		sz = MIN(pgsz, ptob(npages - pidx));
710 		hat_memload_array(seg->s_as->a_hat, a, sz,
711 		    &ppa[pidx], sptd->spt_prot, hat_flags);
712 	}
713 
714 	/*
715 	 * On platforms that do not support HAT_DYNAMIC_ISM_UNMAP,
716 	 * we will leave the pages locked SE_SHARED for the life
717 	 * of the ISM segment. This will prevent any calls to
718 	 * hat_pageunload() on this ISM segment for those platforms.
719 	 */
720 	if (!(hat_flags & HAT_LOAD_LOCK)) {
721 		/*
722 		 * On platforms that support HAT_DYNAMIC_ISM_UNMAP,
723 		 * we no longer need to hold the SE_SHARED lock on the pages,
724 		 * since L_PAGELOCK and F_SOFTLOCK calls will grab the
725 		 * SE_SHARED lock on the pages as necessary.
726 		 */
727 		for (i = 0; i < npages; i++)
728 			page_unlock(ppa[i]);
729 	}
730 	sptd->spt_pcachecnt = 0;
731 	kmem_free(ppa, ((sizeof (page_t *)) * npages));
732 	sptd->spt_realsize = ptob(npages);
733 	atomic_add_long(&spt_used, npages);
734 	sptcargs->seg_spt = seg;
735 	return (0);
736 
737 out4:
738 	seg->s_data = NULL;
739 	kmem_free(vp, sizeof (*vp));
740 	cv_destroy(&sptd->spt_cv);
741 out3:
742 	mutex_destroy(&sptd->spt_lock);
743 	if ((sptcargs->flags & SHM_PAGEABLE) == 0)
744 		kmem_free(ppa, (sizeof (*ppa) * npages));
745 out2:
746 	kmem_free(sptd, sizeof (*sptd));
747 out1:
748 	if ((sptcargs->flags & SHM_PAGEABLE) == 0)
749 		anon_swap_restore(npages);
750 	return (err);
751 }
752 
753 /*ARGSUSED*/
754 void
755 segspt_free_pages(struct seg *seg, caddr_t addr, size_t len)
756 {
757 	struct page	*pp;
758 	struct spt_data *sptd = (struct spt_data *)seg->s_data;
759 	pgcnt_t		npages;
760 	ulong_t		anon_idx;
761 	struct anon_map *amp;
762 	struct anon	*ap;
763 	struct vnode	*vp;
764 	u_offset_t	off;
765 	uint_t		hat_flags;
766 	int		root = 0;
767 	pgcnt_t		pgs, curnpgs = 0;
768 	page_t		*rootpp;
769 	rctl_qty_t	unlocked_bytes = 0;
770 	kproject_t	*proj;
771 	kshmid_t	*sp;
772 
773 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
774 
775 	len = P2ROUNDUP(len, PAGESIZE);
776 
777 	npages = btop(len);
778 
779 	hat_flags = HAT_UNLOAD_UNLOCK | HAT_UNLOAD_UNMAP;
780 	if ((hat_supported(HAT_DYNAMIC_ISM_UNMAP, (void *)0)) ||
781 	    (sptd->spt_flags & SHM_PAGEABLE)) {
782 		hat_flags = HAT_UNLOAD_UNMAP;
783 	}
784 
785 	hat_unload(seg->s_as->a_hat, addr, len, hat_flags);
786 
787 	amp = sptd->spt_amp;
788 	if (sptd->spt_flags & SHM_PAGEABLE)
789 		npages = btop(amp->size);
790 
791 	ASSERT(amp != NULL);
792 
793 	proj = NULL;
794 	rootpp = NULL;
795 	sp = NULL;
796 	if ((sptd->spt_flags & SHM_PAGEABLE) == 0) {
797 		sp = amp->a_sp;
798 		proj = sp->shm_perm.ipc_proj;
799 		mutex_enter(&sp->shm_mlock);
800 	}
801 	for (anon_idx = 0; anon_idx < npages; anon_idx++) {
802 		if ((sptd->spt_flags & SHM_PAGEABLE) == 0) {
803 			if ((ap = anon_get_ptr(amp->ahp, anon_idx)) == NULL) {
804 				panic("segspt_free_pages: null app");
805 				/*NOTREACHED*/
806 			}
807 		} else {
808 			if ((ap = anon_get_next_ptr(amp->ahp, &anon_idx))
809 			    == NULL)
810 				continue;
811 		}
812 		ASSERT(ANON_ISBUSY(anon_get_slot(amp->ahp, anon_idx)) == 0);
813 		swap_xlate(ap, &vp, &off);
814 
815 		/*
816 		 * If this platform supports HAT_DYNAMIC_ISM_UNMAP,
817 		 * the pages won't be having SE_SHARED lock at this
818 		 * point.
819 		 *
820 		 * On platforms that do not support HAT_DYNAMIC_ISM_UNMAP,
821 		 * the pages are still held SE_SHARED locked from the
822 		 * original segspt_create()
823 		 *
824 		 * Our goal is to get SE_EXCL lock on each page, remove
825 		 * permanent lock on it and invalidate the page.
826 		 */
827 		if ((sptd->spt_flags & SHM_PAGEABLE) == 0) {
828 			if (hat_flags == HAT_UNLOAD_UNMAP)
829 				pp = page_lookup(vp, off, SE_EXCL);
830 			else {
831 				if ((pp = page_find(vp, off)) == NULL) {
832 					panic("segspt_free_pages: "
833 					    "page not locked");
834 					/*NOTREACHED*/
835 				}
836 				if (!page_tryupgrade(pp)) {
837 					page_unlock(pp);
838 					pp = page_lookup(vp, off, SE_EXCL);
839 				}
840 			}
841 			if (pp == NULL) {
842 				panic("segspt_free_pages: "
843 				    "page not in the system");
844 				/*NOTREACHED*/
845 			}
846 			ASSERT(pp->p_lckcnt > 0);
847 			page_pp_unlock(pp, 0, 1);
848 			if (pp->p_lckcnt == 0)
849 				unlocked_bytes += PAGESIZE;
850 		} else {
851 			if ((pp = page_lookup(vp, off, SE_EXCL)) == NULL)
852 				continue;
853 		}
854 		/*
855 		 * It's logical to invalidate the pages here as in most cases
856 		 * these were created by segspt.
857 		 */
858 		if (pp->p_szc != 0) {
859 			if (root == 0) {
860 				ASSERT(curnpgs == 0);
861 				root = 1;
862 				rootpp = pp;
863 				pgs = curnpgs = page_get_pagecnt(pp->p_szc);
864 				ASSERT(pgs > 1);
865 				ASSERT(IS_P2ALIGNED(pgs, pgs));
866 				ASSERT(!(page_pptonum(pp) & (pgs - 1)));
867 				curnpgs--;
868 			} else if ((page_pptonum(pp) & (pgs - 1)) == pgs - 1) {
869 				ASSERT(curnpgs == 1);
870 				ASSERT(page_pptonum(pp) ==
871 				    page_pptonum(rootpp) + (pgs - 1));
872 				page_destroy_pages(rootpp);
873 				root = 0;
874 				curnpgs = 0;
875 			} else {
876 				ASSERT(curnpgs > 1);
877 				ASSERT(page_pptonum(pp) ==
878 				    page_pptonum(rootpp) + (pgs - curnpgs));
879 				curnpgs--;
880 			}
881 		} else {
882 			if (root != 0 || curnpgs != 0) {
883 				panic("segspt_free_pages: bad large page");
884 				/*NOTREACHED*/
885 			}
886 			/*
887 			 * Before destroying the pages, we need to take care
888 			 * of the rctl locked memory accounting. For that
889 			 * we need to calculte the unlocked_bytes.
890 			 */
891 			if (pp->p_lckcnt > 0)
892 				unlocked_bytes += PAGESIZE;
893 			/*LINTED: constant in conditional context */
894 			VN_DISPOSE(pp, B_INVAL, 0, kcred);
895 		}
896 	}
897 	if ((sptd->spt_flags & SHM_PAGEABLE) == 0) {
898 		if (unlocked_bytes > 0)
899 			rctl_decr_locked_mem(NULL, proj, unlocked_bytes, 0);
900 		mutex_exit(&sp->shm_mlock);
901 	}
902 	if (root != 0 || curnpgs != 0) {
903 		panic("segspt_free_pages: bad large page");
904 		/*NOTREACHED*/
905 	}
906 
907 	/*
908 	 * mark that pages have been released
909 	 */
910 	sptd->spt_realsize = 0;
911 
912 	if ((sptd->spt_flags & SHM_PAGEABLE) == 0) {
913 		atomic_add_long(&spt_used, -npages);
914 		anon_swap_restore(npages);
915 	}
916 }
917 
918 /*
919  * Get memory allocation policy info for specified address in given segment
920  */
921 static lgrp_mem_policy_info_t *
922 segspt_getpolicy(struct seg *seg, caddr_t addr)
923 {
924 	struct anon_map		*amp;
925 	ulong_t			anon_index;
926 	lgrp_mem_policy_info_t	*policy_info;
927 	struct spt_data		*spt_data;
928 
929 	ASSERT(seg != NULL);
930 
931 	/*
932 	 * Get anon_map from segspt
933 	 *
934 	 * Assume that no lock needs to be held on anon_map, since
935 	 * it should be protected by its reference count which must be
936 	 * nonzero for an existing segment
937 	 * Need to grab readers lock on policy tree though
938 	 */
939 	spt_data = (struct spt_data *)seg->s_data;
940 	if (spt_data == NULL)
941 		return (NULL);
942 	amp = spt_data->spt_amp;
943 	ASSERT(amp->refcnt != 0);
944 
945 	/*
946 	 * Get policy info
947 	 *
948 	 * Assume starting anon index of 0
949 	 */
950 	anon_index = seg_page(seg, addr);
951 	policy_info = lgrp_shm_policy_get(amp, anon_index, NULL, 0);
952 
953 	return (policy_info);
954 }
955 
956 /*
957  * DISM only.
958  * Return locked pages over a given range.
959  *
960  * We will cache all DISM locked pages and save the pplist for the
961  * entire segment in the ppa field of the underlying DISM segment structure.
962  * Later, during a call to segspt_reclaim() we will use this ppa array
963  * to page_unlock() all of the pages and then we will free this ppa list.
964  */
965 /*ARGSUSED*/
966 static int
967 segspt_dismpagelock(struct seg *seg, caddr_t addr, size_t len,
968     struct page ***ppp, enum lock_type type, enum seg_rw rw)
969 {
970 	struct  shm_data *shmd = (struct shm_data *)seg->s_data;
971 	struct  seg	*sptseg = shmd->shm_sptseg;
972 	struct  spt_data *sptd = sptseg->s_data;
973 	pgcnt_t pg_idx, npages, tot_npages, npgs;
974 	struct  page **pplist, **pl, **ppa, *pp;
975 	struct  anon_map *amp;
976 	spgcnt_t	an_idx;
977 	int	ret = ENOTSUP;
978 	uint_t	pl_built = 0;
979 	struct  anon *ap;
980 	struct  vnode *vp;
981 	u_offset_t off;
982 	pgcnt_t claim_availrmem = 0;
983 	uint_t	szc;
984 
985 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
986 	ASSERT(type == L_PAGELOCK || type == L_PAGEUNLOCK);
987 
988 	/*
989 	 * We want to lock/unlock the entire ISM segment. Therefore,
990 	 * we will be using the underlying sptseg and it's base address
991 	 * and length for the caching arguments.
992 	 */
993 	ASSERT(sptseg);
994 	ASSERT(sptd);
995 
996 	pg_idx = seg_page(seg, addr);
997 	npages = btopr(len);
998 
999 	/*
1000 	 * check if the request is larger than number of pages covered
1001 	 * by amp
1002 	 */
1003 	if (pg_idx + npages > btopr(sptd->spt_amp->size)) {
1004 		*ppp = NULL;
1005 		return (ENOTSUP);
1006 	}
1007 
1008 	if (type == L_PAGEUNLOCK) {
1009 		ASSERT(sptd->spt_ppa != NULL);
1010 
1011 		seg_pinactive(seg, NULL, seg->s_base, sptd->spt_amp->size,
1012 		    sptd->spt_ppa, S_WRITE, SEGP_FORCE_WIRED, segspt_reclaim);
1013 
1014 		/*
1015 		 * If someone is blocked while unmapping, we purge
1016 		 * segment page cache and thus reclaim pplist synchronously
1017 		 * without waiting for seg_pasync_thread. This speeds up
1018 		 * unmapping in cases where munmap(2) is called, while
1019 		 * raw async i/o is still in progress or where a thread
1020 		 * exits on data fault in a multithreaded application.
1021 		 */
1022 		if ((sptd->spt_flags & DISM_PPA_CHANGED) ||
1023 		    (AS_ISUNMAPWAIT(seg->s_as) &&
1024 		    shmd->shm_softlockcnt > 0)) {
1025 			segspt_purge(seg);
1026 		}
1027 		return (0);
1028 	}
1029 
1030 	/* The L_PAGELOCK case ... */
1031 
1032 	if (sptd->spt_flags & DISM_PPA_CHANGED) {
1033 		segspt_purge(seg);
1034 		/*
1035 		 * for DISM ppa needs to be rebuild since
1036 		 * number of locked pages could be changed
1037 		 */
1038 		*ppp = NULL;
1039 		return (ENOTSUP);
1040 	}
1041 
1042 	/*
1043 	 * First try to find pages in segment page cache, without
1044 	 * holding the segment lock.
1045 	 */
1046 	pplist = seg_plookup(seg, NULL, seg->s_base, sptd->spt_amp->size,
1047 	    S_WRITE, SEGP_FORCE_WIRED);
1048 	if (pplist != NULL) {
1049 		ASSERT(sptd->spt_ppa != NULL);
1050 		ASSERT(sptd->spt_ppa == pplist);
1051 		ppa = sptd->spt_ppa;
1052 		for (an_idx = pg_idx; an_idx < pg_idx + npages; ) {
1053 			if (ppa[an_idx] == NULL) {
1054 				seg_pinactive(seg, NULL, seg->s_base,
1055 				    sptd->spt_amp->size, ppa,
1056 				    S_WRITE, SEGP_FORCE_WIRED, segspt_reclaim);
1057 				*ppp = NULL;
1058 				return (ENOTSUP);
1059 			}
1060 			if ((szc = ppa[an_idx]->p_szc) != 0) {
1061 				npgs = page_get_pagecnt(szc);
1062 				an_idx = P2ROUNDUP(an_idx + 1, npgs);
1063 			} else {
1064 				an_idx++;
1065 			}
1066 		}
1067 		/*
1068 		 * Since we cache the entire DISM segment, we want to
1069 		 * set ppp to point to the first slot that corresponds
1070 		 * to the requested addr, i.e. pg_idx.
1071 		 */
1072 		*ppp = &(sptd->spt_ppa[pg_idx]);
1073 		return (0);
1074 	}
1075 
1076 	mutex_enter(&sptd->spt_lock);
1077 	/*
1078 	 * try to find pages in segment page cache with mutex
1079 	 */
1080 	pplist = seg_plookup(seg, NULL, seg->s_base, sptd->spt_amp->size,
1081 	    S_WRITE, SEGP_FORCE_WIRED);
1082 	if (pplist != NULL) {
1083 		ASSERT(sptd->spt_ppa != NULL);
1084 		ASSERT(sptd->spt_ppa == pplist);
1085 		ppa = sptd->spt_ppa;
1086 		for (an_idx = pg_idx; an_idx < pg_idx + npages; ) {
1087 			if (ppa[an_idx] == NULL) {
1088 				mutex_exit(&sptd->spt_lock);
1089 				seg_pinactive(seg, NULL, seg->s_base,
1090 				    sptd->spt_amp->size, ppa,
1091 				    S_WRITE, SEGP_FORCE_WIRED, segspt_reclaim);
1092 				*ppp = NULL;
1093 				return (ENOTSUP);
1094 			}
1095 			if ((szc = ppa[an_idx]->p_szc) != 0) {
1096 				npgs = page_get_pagecnt(szc);
1097 				an_idx = P2ROUNDUP(an_idx + 1, npgs);
1098 			} else {
1099 				an_idx++;
1100 			}
1101 		}
1102 		/*
1103 		 * Since we cache the entire DISM segment, we want to
1104 		 * set ppp to point to the first slot that corresponds
1105 		 * to the requested addr, i.e. pg_idx.
1106 		 */
1107 		mutex_exit(&sptd->spt_lock);
1108 		*ppp = &(sptd->spt_ppa[pg_idx]);
1109 		return (0);
1110 	}
1111 	if (seg_pinsert_check(seg, NULL, seg->s_base, sptd->spt_amp->size,
1112 	    SEGP_FORCE_WIRED) == SEGP_FAIL) {
1113 		mutex_exit(&sptd->spt_lock);
1114 		*ppp = NULL;
1115 		return (ENOTSUP);
1116 	}
1117 
1118 	/*
1119 	 * No need to worry about protections because DISM pages are always rw.
1120 	 */
1121 	pl = pplist = NULL;
1122 	amp = sptd->spt_amp;
1123 
1124 	/*
1125 	 * Do we need to build the ppa array?
1126 	 */
1127 	if (sptd->spt_ppa == NULL) {
1128 		pgcnt_t lpg_cnt = 0;
1129 
1130 		pl_built = 1;
1131 		tot_npages = btopr(sptd->spt_amp->size);
1132 
1133 		ASSERT(sptd->spt_pcachecnt == 0);
1134 		pplist = kmem_zalloc(sizeof (page_t *) * tot_npages, KM_SLEEP);
1135 		pl = pplist;
1136 
1137 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1138 		for (an_idx = 0; an_idx < tot_npages; ) {
1139 			ap = anon_get_ptr(amp->ahp, an_idx);
1140 			/*
1141 			 * Cache only mlocked pages. For large pages
1142 			 * if one (constituent) page is mlocked
1143 			 * all pages for that large page
1144 			 * are cached also. This is for quick
1145 			 * lookups of ppa array;
1146 			 */
1147 			if ((ap != NULL) && (lpg_cnt != 0 ||
1148 			    (sptd->spt_ppa_lckcnt[an_idx] != 0))) {
1149 
1150 				swap_xlate(ap, &vp, &off);
1151 				pp = page_lookup(vp, off, SE_SHARED);
1152 				ASSERT(pp != NULL);
1153 				if (lpg_cnt == 0) {
1154 					lpg_cnt++;
1155 					/*
1156 					 * For a small page, we are done --
1157 					 * lpg_count is reset to 0 below.
1158 					 *
1159 					 * For a large page, we are guaranteed
1160 					 * to find the anon structures of all
1161 					 * constituent pages and a non-zero
1162 					 * lpg_cnt ensures that we don't test
1163 					 * for mlock for these. We are done
1164 					 * when lpg_count reaches (npgs + 1).
1165 					 * If we are not the first constituent
1166 					 * page, restart at the first one.
1167 					 */
1168 					npgs = page_get_pagecnt(pp->p_szc);
1169 					if (!IS_P2ALIGNED(an_idx, npgs)) {
1170 						an_idx = P2ALIGN(an_idx, npgs);
1171 						page_unlock(pp);
1172 						continue;
1173 					}
1174 				}
1175 				if (++lpg_cnt > npgs)
1176 					lpg_cnt = 0;
1177 
1178 				/*
1179 				 * availrmem is decremented only
1180 				 * for unlocked pages
1181 				 */
1182 				if (sptd->spt_ppa_lckcnt[an_idx] == 0)
1183 					claim_availrmem++;
1184 				pplist[an_idx] = pp;
1185 			}
1186 			an_idx++;
1187 		}
1188 		ANON_LOCK_EXIT(&amp->a_rwlock);
1189 
1190 		if (claim_availrmem) {
1191 			mutex_enter(&freemem_lock);
1192 			if (availrmem < tune.t_minarmem + claim_availrmem) {
1193 				mutex_exit(&freemem_lock);
1194 				ret = ENOTSUP;
1195 				claim_availrmem = 0;
1196 				goto insert_fail;
1197 			} else {
1198 				availrmem -= claim_availrmem;
1199 			}
1200 			mutex_exit(&freemem_lock);
1201 		}
1202 
1203 		sptd->spt_ppa = pl;
1204 	} else {
1205 		/*
1206 		 * We already have a valid ppa[].
1207 		 */
1208 		pl = sptd->spt_ppa;
1209 	}
1210 
1211 	ASSERT(pl != NULL);
1212 
1213 	ret = seg_pinsert(seg, NULL, seg->s_base, sptd->spt_amp->size,
1214 	    sptd->spt_amp->size, pl, S_WRITE, SEGP_FORCE_WIRED,
1215 	    segspt_reclaim);
1216 	if (ret == SEGP_FAIL) {
1217 		/*
1218 		 * seg_pinsert failed. We return
1219 		 * ENOTSUP, so that the as_pagelock() code will
1220 		 * then try the slower F_SOFTLOCK path.
1221 		 */
1222 		if (pl_built) {
1223 			/*
1224 			 * No one else has referenced the ppa[].
1225 			 * We created it and we need to destroy it.
1226 			 */
1227 			sptd->spt_ppa = NULL;
1228 		}
1229 		ret = ENOTSUP;
1230 		goto insert_fail;
1231 	}
1232 
1233 	/*
1234 	 * In either case, we increment softlockcnt on the 'real' segment.
1235 	 */
1236 	sptd->spt_pcachecnt++;
1237 	atomic_inc_ulong((ulong_t *)(&(shmd->shm_softlockcnt)));
1238 
1239 	ppa = sptd->spt_ppa;
1240 	for (an_idx = pg_idx; an_idx < pg_idx + npages; ) {
1241 		if (ppa[an_idx] == NULL) {
1242 			mutex_exit(&sptd->spt_lock);
1243 			seg_pinactive(seg, NULL, seg->s_base,
1244 			    sptd->spt_amp->size,
1245 			    pl, S_WRITE, SEGP_FORCE_WIRED, segspt_reclaim);
1246 			*ppp = NULL;
1247 			return (ENOTSUP);
1248 		}
1249 		if ((szc = ppa[an_idx]->p_szc) != 0) {
1250 			npgs = page_get_pagecnt(szc);
1251 			an_idx = P2ROUNDUP(an_idx + 1, npgs);
1252 		} else {
1253 			an_idx++;
1254 		}
1255 	}
1256 	/*
1257 	 * We can now drop the sptd->spt_lock since the ppa[]
1258 	 * exists and we have incremented pacachecnt.
1259 	 */
1260 	mutex_exit(&sptd->spt_lock);
1261 
1262 	/*
1263 	 * Since we cache the entire segment, we want to
1264 	 * set ppp to point to the first slot that corresponds
1265 	 * to the requested addr, i.e. pg_idx.
1266 	 */
1267 	*ppp = &(sptd->spt_ppa[pg_idx]);
1268 	return (0);
1269 
1270 insert_fail:
1271 	/*
1272 	 * We will only reach this code if we tried and failed.
1273 	 *
1274 	 * And we can drop the lock on the dummy seg, once we've failed
1275 	 * to set up a new ppa[].
1276 	 */
1277 	mutex_exit(&sptd->spt_lock);
1278 
1279 	if (pl_built) {
1280 		if (claim_availrmem) {
1281 			mutex_enter(&freemem_lock);
1282 			availrmem += claim_availrmem;
1283 			mutex_exit(&freemem_lock);
1284 		}
1285 
1286 		/*
1287 		 * We created pl and we need to destroy it.
1288 		 */
1289 		pplist = pl;
1290 		for (an_idx = 0; an_idx < tot_npages; an_idx++) {
1291 			if (pplist[an_idx] != NULL)
1292 				page_unlock(pplist[an_idx]);
1293 		}
1294 		kmem_free(pl, sizeof (page_t *) * tot_npages);
1295 	}
1296 
1297 	if (shmd->shm_softlockcnt <= 0) {
1298 		if (AS_ISUNMAPWAIT(seg->s_as)) {
1299 			mutex_enter(&seg->s_as->a_contents);
1300 			if (AS_ISUNMAPWAIT(seg->s_as)) {
1301 				AS_CLRUNMAPWAIT(seg->s_as);
1302 				cv_broadcast(&seg->s_as->a_cv);
1303 			}
1304 			mutex_exit(&seg->s_as->a_contents);
1305 		}
1306 	}
1307 	*ppp = NULL;
1308 	return (ret);
1309 }
1310 
1311 
1312 
1313 /*
1314  * return locked pages over a given range.
1315  *
1316  * We will cache the entire ISM segment and save the pplist for the
1317  * entire segment in the ppa field of the underlying ISM segment structure.
1318  * Later, during a call to segspt_reclaim() we will use this ppa array
1319  * to page_unlock() all of the pages and then we will free this ppa list.
1320  */
1321 /*ARGSUSED*/
1322 static int
1323 segspt_shmpagelock(struct seg *seg, caddr_t addr, size_t len,
1324     struct page ***ppp, enum lock_type type, enum seg_rw rw)
1325 {
1326 	struct shm_data *shmd = (struct shm_data *)seg->s_data;
1327 	struct seg	*sptseg = shmd->shm_sptseg;
1328 	struct spt_data *sptd = sptseg->s_data;
1329 	pgcnt_t np, page_index, npages;
1330 	caddr_t a, spt_base;
1331 	struct page **pplist, **pl, *pp;
1332 	struct anon_map *amp;
1333 	ulong_t anon_index;
1334 	int ret = ENOTSUP;
1335 	uint_t	pl_built = 0;
1336 	struct anon *ap;
1337 	struct vnode *vp;
1338 	u_offset_t off;
1339 
1340 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
1341 	ASSERT(type == L_PAGELOCK || type == L_PAGEUNLOCK);
1342 
1343 
1344 	/*
1345 	 * We want to lock/unlock the entire ISM segment. Therefore,
1346 	 * we will be using the underlying sptseg and it's base address
1347 	 * and length for the caching arguments.
1348 	 */
1349 	ASSERT(sptseg);
1350 	ASSERT(sptd);
1351 
1352 	if (sptd->spt_flags & SHM_PAGEABLE) {
1353 		return (segspt_dismpagelock(seg, addr, len, ppp, type, rw));
1354 	}
1355 
1356 	page_index = seg_page(seg, addr);
1357 	npages = btopr(len);
1358 
1359 	/*
1360 	 * check if the request is larger than number of pages covered
1361 	 * by amp
1362 	 */
1363 	if (page_index + npages > btopr(sptd->spt_amp->size)) {
1364 		*ppp = NULL;
1365 		return (ENOTSUP);
1366 	}
1367 
1368 	if (type == L_PAGEUNLOCK) {
1369 
1370 		ASSERT(sptd->spt_ppa != NULL);
1371 
1372 		seg_pinactive(seg, NULL, seg->s_base, sptd->spt_amp->size,
1373 		    sptd->spt_ppa, S_WRITE, SEGP_FORCE_WIRED, segspt_reclaim);
1374 
1375 		/*
1376 		 * If someone is blocked while unmapping, we purge
1377 		 * segment page cache and thus reclaim pplist synchronously
1378 		 * without waiting for seg_pasync_thread. This speeds up
1379 		 * unmapping in cases where munmap(2) is called, while
1380 		 * raw async i/o is still in progress or where a thread
1381 		 * exits on data fault in a multithreaded application.
1382 		 */
1383 		if (AS_ISUNMAPWAIT(seg->s_as) && (shmd->shm_softlockcnt > 0)) {
1384 			segspt_purge(seg);
1385 		}
1386 		return (0);
1387 	}
1388 
1389 	/* The L_PAGELOCK case... */
1390 
1391 	/*
1392 	 * First try to find pages in segment page cache, without
1393 	 * holding the segment lock.
1394 	 */
1395 	pplist = seg_plookup(seg, NULL, seg->s_base, sptd->spt_amp->size,
1396 	    S_WRITE, SEGP_FORCE_WIRED);
1397 	if (pplist != NULL) {
1398 		ASSERT(sptd->spt_ppa == pplist);
1399 		ASSERT(sptd->spt_ppa[page_index]);
1400 		/*
1401 		 * Since we cache the entire ISM segment, we want to
1402 		 * set ppp to point to the first slot that corresponds
1403 		 * to the requested addr, i.e. page_index.
1404 		 */
1405 		*ppp = &(sptd->spt_ppa[page_index]);
1406 		return (0);
1407 	}
1408 
1409 	mutex_enter(&sptd->spt_lock);
1410 
1411 	/*
1412 	 * try to find pages in segment page cache
1413 	 */
1414 	pplist = seg_plookup(seg, NULL, seg->s_base, sptd->spt_amp->size,
1415 	    S_WRITE, SEGP_FORCE_WIRED);
1416 	if (pplist != NULL) {
1417 		ASSERT(sptd->spt_ppa == pplist);
1418 		/*
1419 		 * Since we cache the entire segment, we want to
1420 		 * set ppp to point to the first slot that corresponds
1421 		 * to the requested addr, i.e. page_index.
1422 		 */
1423 		mutex_exit(&sptd->spt_lock);
1424 		*ppp = &(sptd->spt_ppa[page_index]);
1425 		return (0);
1426 	}
1427 
1428 	if (seg_pinsert_check(seg, NULL, seg->s_base, sptd->spt_amp->size,
1429 	    SEGP_FORCE_WIRED) == SEGP_FAIL) {
1430 		mutex_exit(&sptd->spt_lock);
1431 		*ppp = NULL;
1432 		return (ENOTSUP);
1433 	}
1434 
1435 	/*
1436 	 * No need to worry about protections because ISM pages
1437 	 * are always rw.
1438 	 */
1439 	pl = pplist = NULL;
1440 
1441 	/*
1442 	 * Do we need to build the ppa array?
1443 	 */
1444 	if (sptd->spt_ppa == NULL) {
1445 		ASSERT(sptd->spt_ppa == pplist);
1446 
1447 		spt_base = sptseg->s_base;
1448 		pl_built = 1;
1449 
1450 		/*
1451 		 * availrmem is decremented once during anon_swap_adjust()
1452 		 * and is incremented during the anon_unresv(), which is
1453 		 * called from shm_rm_amp() when the segment is destroyed.
1454 		 */
1455 		amp = sptd->spt_amp;
1456 		ASSERT(amp != NULL);
1457 
1458 		/* pcachecnt is protected by sptd->spt_lock */
1459 		ASSERT(sptd->spt_pcachecnt == 0);
1460 		pplist = kmem_zalloc(sizeof (page_t *)
1461 		    * btopr(sptd->spt_amp->size), KM_SLEEP);
1462 		pl = pplist;
1463 
1464 		anon_index = seg_page(sptseg, spt_base);
1465 
1466 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1467 		for (a = spt_base; a < (spt_base + sptd->spt_amp->size);
1468 		    a += PAGESIZE, anon_index++, pplist++) {
1469 			ap = anon_get_ptr(amp->ahp, anon_index);
1470 			ASSERT(ap != NULL);
1471 			swap_xlate(ap, &vp, &off);
1472 			pp = page_lookup(vp, off, SE_SHARED);
1473 			ASSERT(pp != NULL);
1474 			*pplist = pp;
1475 		}
1476 		ANON_LOCK_EXIT(&amp->a_rwlock);
1477 
1478 		if (a < (spt_base + sptd->spt_amp->size)) {
1479 			ret = ENOTSUP;
1480 			goto insert_fail;
1481 		}
1482 		sptd->spt_ppa = pl;
1483 	} else {
1484 		/*
1485 		 * We already have a valid ppa[].
1486 		 */
1487 		pl = sptd->spt_ppa;
1488 	}
1489 
1490 	ASSERT(pl != NULL);
1491 
1492 	ret = seg_pinsert(seg, NULL, seg->s_base, sptd->spt_amp->size,
1493 	    sptd->spt_amp->size, pl, S_WRITE, SEGP_FORCE_WIRED,
1494 	    segspt_reclaim);
1495 	if (ret == SEGP_FAIL) {
1496 		/*
1497 		 * seg_pinsert failed. We return
1498 		 * ENOTSUP, so that the as_pagelock() code will
1499 		 * then try the slower F_SOFTLOCK path.
1500 		 */
1501 		if (pl_built) {
1502 			/*
1503 			 * No one else has referenced the ppa[].
1504 			 * We created it and we need to destroy it.
1505 			 */
1506 			sptd->spt_ppa = NULL;
1507 		}
1508 		ret = ENOTSUP;
1509 		goto insert_fail;
1510 	}
1511 
1512 	/*
1513 	 * In either case, we increment softlockcnt on the 'real' segment.
1514 	 */
1515 	sptd->spt_pcachecnt++;
1516 	atomic_inc_ulong((ulong_t *)(&(shmd->shm_softlockcnt)));
1517 
1518 	/*
1519 	 * We can now drop the sptd->spt_lock since the ppa[]
1520 	 * exists and we have incremented pacachecnt.
1521 	 */
1522 	mutex_exit(&sptd->spt_lock);
1523 
1524 	/*
1525 	 * Since we cache the entire segment, we want to
1526 	 * set ppp to point to the first slot that corresponds
1527 	 * to the requested addr, i.e. page_index.
1528 	 */
1529 	*ppp = &(sptd->spt_ppa[page_index]);
1530 	return (0);
1531 
1532 insert_fail:
1533 	/*
1534 	 * We will only reach this code if we tried and failed.
1535 	 *
1536 	 * And we can drop the lock on the dummy seg, once we've failed
1537 	 * to set up a new ppa[].
1538 	 */
1539 	mutex_exit(&sptd->spt_lock);
1540 
1541 	if (pl_built) {
1542 		/*
1543 		 * We created pl and we need to destroy it.
1544 		 */
1545 		pplist = pl;
1546 		np = (((uintptr_t)(a - spt_base)) >> PAGESHIFT);
1547 		while (np) {
1548 			page_unlock(*pplist);
1549 			np--;
1550 			pplist++;
1551 		}
1552 		kmem_free(pl, sizeof (page_t *) * btopr(sptd->spt_amp->size));
1553 	}
1554 	if (shmd->shm_softlockcnt <= 0) {
1555 		if (AS_ISUNMAPWAIT(seg->s_as)) {
1556 			mutex_enter(&seg->s_as->a_contents);
1557 			if (AS_ISUNMAPWAIT(seg->s_as)) {
1558 				AS_CLRUNMAPWAIT(seg->s_as);
1559 				cv_broadcast(&seg->s_as->a_cv);
1560 			}
1561 			mutex_exit(&seg->s_as->a_contents);
1562 		}
1563 	}
1564 	*ppp = NULL;
1565 	return (ret);
1566 }
1567 
1568 /*
1569  * purge any cached pages in the I/O page cache
1570  */
1571 static void
1572 segspt_purge(struct seg *seg)
1573 {
1574 	seg_ppurge(seg, NULL, SEGP_FORCE_WIRED);
1575 }
1576 
1577 static int
1578 segspt_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
1579     enum seg_rw rw, int async)
1580 {
1581 	struct seg *seg = (struct seg *)ptag;
1582 	struct	shm_data *shmd = (struct shm_data *)seg->s_data;
1583 	struct	seg	*sptseg;
1584 	struct	spt_data *sptd;
1585 	pgcnt_t npages, i, free_availrmem = 0;
1586 	int	done = 0;
1587 
1588 #ifdef lint
1589 	addr = addr;
1590 #endif
1591 	sptseg = shmd->shm_sptseg;
1592 	sptd = sptseg->s_data;
1593 	npages = (len >> PAGESHIFT);
1594 	ASSERT(npages);
1595 	ASSERT(sptd->spt_pcachecnt != 0);
1596 	ASSERT(sptd->spt_ppa == pplist);
1597 	ASSERT(npages == btopr(sptd->spt_amp->size));
1598 	ASSERT(async || AS_LOCK_HELD(seg->s_as));
1599 
1600 	/*
1601 	 * Acquire the lock on the dummy seg and destroy the
1602 	 * ppa array IF this is the last pcachecnt.
1603 	 */
1604 	mutex_enter(&sptd->spt_lock);
1605 	if (--sptd->spt_pcachecnt == 0) {
1606 		for (i = 0; i < npages; i++) {
1607 			if (pplist[i] == NULL) {
1608 				continue;
1609 			}
1610 			if (rw == S_WRITE) {
1611 				hat_setrefmod(pplist[i]);
1612 			} else {
1613 				hat_setref(pplist[i]);
1614 			}
1615 			if ((sptd->spt_flags & SHM_PAGEABLE) &&
1616 			    (sptd->spt_ppa_lckcnt[i] == 0))
1617 				free_availrmem++;
1618 			page_unlock(pplist[i]);
1619 		}
1620 		if ((sptd->spt_flags & SHM_PAGEABLE) && free_availrmem) {
1621 			mutex_enter(&freemem_lock);
1622 			availrmem += free_availrmem;
1623 			mutex_exit(&freemem_lock);
1624 		}
1625 		/*
1626 		 * Since we want to cach/uncache the entire ISM segment,
1627 		 * we will track the pplist in a segspt specific field
1628 		 * ppa, that is initialized at the time we add an entry to
1629 		 * the cache.
1630 		 */
1631 		ASSERT(sptd->spt_pcachecnt == 0);
1632 		kmem_free(pplist, sizeof (page_t *) * npages);
1633 		sptd->spt_ppa = NULL;
1634 		sptd->spt_flags &= ~DISM_PPA_CHANGED;
1635 		sptd->spt_gen++;
1636 		cv_broadcast(&sptd->spt_cv);
1637 		done = 1;
1638 	}
1639 	mutex_exit(&sptd->spt_lock);
1640 
1641 	/*
1642 	 * If we are pcache async thread or called via seg_ppurge_wiredpp() we
1643 	 * may not hold AS lock (in this case async argument is not 0). This
1644 	 * means if softlockcnt drops to 0 after the decrement below address
1645 	 * space may get freed. We can't allow it since after softlock
1646 	 * derement to 0 we still need to access as structure for possible
1647 	 * wakeup of unmap waiters. To prevent the disappearance of as we take
1648 	 * this segment's shm_segfree_syncmtx. segspt_shmfree() also takes
1649 	 * this mutex as a barrier to make sure this routine completes before
1650 	 * segment is freed.
1651 	 *
1652 	 * The second complication we have to deal with in async case is a
1653 	 * possibility of missed wake up of unmap wait thread. When we don't
1654 	 * hold as lock here we may take a_contents lock before unmap wait
1655 	 * thread that was first to see softlockcnt was still not 0. As a
1656 	 * result we'll fail to wake up an unmap wait thread. To avoid this
1657 	 * race we set nounmapwait flag in as structure if we drop softlockcnt
1658 	 * to 0 if async is not 0.  unmapwait thread
1659 	 * will not block if this flag is set.
1660 	 */
1661 	if (async)
1662 		mutex_enter(&shmd->shm_segfree_syncmtx);
1663 
1664 	/*
1665 	 * Now decrement softlockcnt.
1666 	 */
1667 	ASSERT(shmd->shm_softlockcnt > 0);
1668 	atomic_dec_ulong((ulong_t *)(&(shmd->shm_softlockcnt)));
1669 
1670 	if (shmd->shm_softlockcnt <= 0) {
1671 		if (async || AS_ISUNMAPWAIT(seg->s_as)) {
1672 			mutex_enter(&seg->s_as->a_contents);
1673 			if (async)
1674 				AS_SETNOUNMAPWAIT(seg->s_as);
1675 			if (AS_ISUNMAPWAIT(seg->s_as)) {
1676 				AS_CLRUNMAPWAIT(seg->s_as);
1677 				cv_broadcast(&seg->s_as->a_cv);
1678 			}
1679 			mutex_exit(&seg->s_as->a_contents);
1680 		}
1681 	}
1682 
1683 	if (async)
1684 		mutex_exit(&shmd->shm_segfree_syncmtx);
1685 
1686 	return (done);
1687 }
1688 
1689 /*
1690  * Do a F_SOFTUNLOCK call over the range requested.
1691  * The range must have already been F_SOFTLOCK'ed.
1692  *
1693  * The calls to acquire and release the anon map lock mutex were
1694  * removed in order to avoid a deadly embrace during a DR
1695  * memory delete operation.  (Eg. DR blocks while waiting for a
1696  * exclusive lock on a page that is being used for kaio; the
1697  * thread that will complete the kaio and call segspt_softunlock
1698  * blocks on the anon map lock; another thread holding the anon
1699  * map lock blocks on another page lock via the segspt_shmfault
1700  * -> page_lookup -> page_lookup_create -> page_lock_es code flow.)
1701  *
1702  * The appropriateness of the removal is based upon the following:
1703  * 1. If we are holding a segment's reader lock and the page is held
1704  * shared, then the corresponding element in anonmap which points to
1705  * anon struct cannot change and there is no need to acquire the
1706  * anonymous map lock.
1707  * 2. Threads in segspt_softunlock have a reader lock on the segment
1708  * and already have the shared page lock, so we are guaranteed that
1709  * the anon map slot cannot change and therefore can call anon_get_ptr()
1710  * without grabbing the anonymous map lock.
1711  * 3. Threads that softlock a shared page break copy-on-write, even if
1712  * its a read.  Thus cow faults can be ignored with respect to soft
1713  * unlocking, since the breaking of cow means that the anon slot(s) will
1714  * not be shared.
1715  */
1716 static void
1717 segspt_softunlock(struct seg *seg, caddr_t sptseg_addr,
1718     size_t len, enum seg_rw rw)
1719 {
1720 	struct shm_data *shmd = (struct shm_data *)seg->s_data;
1721 	struct seg	*sptseg;
1722 	struct spt_data *sptd;
1723 	page_t *pp;
1724 	caddr_t adr;
1725 	struct vnode *vp;
1726 	u_offset_t offset;
1727 	ulong_t anon_index;
1728 	struct anon_map *amp;		/* XXX - for locknest */
1729 	struct anon *ap = NULL;
1730 	pgcnt_t npages;
1731 
1732 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
1733 
1734 	sptseg = shmd->shm_sptseg;
1735 	sptd = sptseg->s_data;
1736 
1737 	/*
1738 	 * Some platforms assume that ISM mappings are HAT_LOAD_LOCK
1739 	 * and therefore their pages are SE_SHARED locked
1740 	 * for the entire life of the segment.
1741 	 */
1742 	if ((!hat_supported(HAT_DYNAMIC_ISM_UNMAP, (void *)0)) &&
1743 	    ((sptd->spt_flags & SHM_PAGEABLE) == 0)) {
1744 		goto softlock_decrement;
1745 	}
1746 
1747 	/*
1748 	 * Any thread is free to do a page_find and
1749 	 * page_unlock() on the pages within this seg.
1750 	 *
1751 	 * We are already holding the as->a_lock on the user's
1752 	 * real segment, but we need to hold the a_lock on the
1753 	 * underlying dummy as. This is mostly to satisfy the
1754 	 * underlying HAT layer.
1755 	 */
1756 	AS_LOCK_ENTER(sptseg->s_as, RW_READER);
1757 	hat_unlock(sptseg->s_as->a_hat, sptseg_addr, len);
1758 	AS_LOCK_EXIT(sptseg->s_as);
1759 
1760 	amp = sptd->spt_amp;
1761 	ASSERT(amp != NULL);
1762 	anon_index = seg_page(sptseg, sptseg_addr);
1763 
1764 	for (adr = sptseg_addr; adr < sptseg_addr + len; adr += PAGESIZE) {
1765 		ap = anon_get_ptr(amp->ahp, anon_index++);
1766 		ASSERT(ap != NULL);
1767 		swap_xlate(ap, &vp, &offset);
1768 
1769 		/*
1770 		 * Use page_find() instead of page_lookup() to
1771 		 * find the page since we know that it has a
1772 		 * "shared" lock.
1773 		 */
1774 		pp = page_find(vp, offset);
1775 		ASSERT(ap == anon_get_ptr(amp->ahp, anon_index - 1));
1776 		if (pp == NULL) {
1777 			panic("segspt_softunlock: "
1778 			    "addr %p, ap %p, vp %p, off %llx",
1779 			    (void *)adr, (void *)ap, (void *)vp, offset);
1780 			/*NOTREACHED*/
1781 		}
1782 
1783 		if (rw == S_WRITE) {
1784 			hat_setrefmod(pp);
1785 		} else if (rw != S_OTHER) {
1786 			hat_setref(pp);
1787 		}
1788 		page_unlock(pp);
1789 	}
1790 
1791 softlock_decrement:
1792 	npages = btopr(len);
1793 	ASSERT(shmd->shm_softlockcnt >= npages);
1794 	atomic_add_long((ulong_t *)(&(shmd->shm_softlockcnt)), -npages);
1795 	if (shmd->shm_softlockcnt == 0) {
1796 		/*
1797 		 * All SOFTLOCKS are gone. Wakeup any waiting
1798 		 * unmappers so they can try again to unmap.
1799 		 * Check for waiters first without the mutex
1800 		 * held so we don't always grab the mutex on
1801 		 * softunlocks.
1802 		 */
1803 		if (AS_ISUNMAPWAIT(seg->s_as)) {
1804 			mutex_enter(&seg->s_as->a_contents);
1805 			if (AS_ISUNMAPWAIT(seg->s_as)) {
1806 				AS_CLRUNMAPWAIT(seg->s_as);
1807 				cv_broadcast(&seg->s_as->a_cv);
1808 			}
1809 			mutex_exit(&seg->s_as->a_contents);
1810 		}
1811 	}
1812 }
1813 
1814 int
1815 segspt_shmattach(struct seg **segpp, void *argsp)
1816 {
1817 	struct seg *seg = *segpp;
1818 	struct shm_data *shmd_arg = (struct shm_data *)argsp;
1819 	struct shm_data *shmd;
1820 	struct anon_map *shm_amp = shmd_arg->shm_amp;
1821 	struct spt_data *sptd;
1822 	int error = 0;
1823 
1824 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1825 
1826 	shmd = kmem_zalloc((sizeof (*shmd)), KM_NOSLEEP);
1827 	if (shmd == NULL)
1828 		return (ENOMEM);
1829 
1830 	shmd->shm_sptas = shmd_arg->shm_sptas;
1831 	shmd->shm_amp = shm_amp;
1832 	shmd->shm_sptseg = shmd_arg->shm_sptseg;
1833 
1834 	(void) lgrp_shm_policy_set(LGRP_MEM_POLICY_DEFAULT, shm_amp, 0,
1835 	    NULL, 0, seg->s_size);
1836 
1837 	mutex_init(&shmd->shm_segfree_syncmtx, NULL, MUTEX_DEFAULT, NULL);
1838 
1839 	seg->s_data = (void *)shmd;
1840 	seg->s_ops = &segspt_shmops;
1841 	seg->s_szc = shmd->shm_sptseg->s_szc;
1842 	sptd = shmd->shm_sptseg->s_data;
1843 
1844 	if (sptd->spt_flags & SHM_PAGEABLE) {
1845 		if ((shmd->shm_vpage = kmem_zalloc(btopr(shm_amp->size),
1846 		    KM_NOSLEEP)) == NULL) {
1847 			seg->s_data = (void *)NULL;
1848 			kmem_free(shmd, (sizeof (*shmd)));
1849 			return (ENOMEM);
1850 		}
1851 		shmd->shm_lckpgs = 0;
1852 		if (hat_supported(HAT_DYNAMIC_ISM_UNMAP, (void *)0)) {
1853 			if ((error = hat_share(seg->s_as->a_hat, seg->s_base,
1854 			    shmd_arg->shm_sptas->a_hat, SEGSPTADDR,
1855 			    seg->s_size, seg->s_szc)) != 0) {
1856 				kmem_free(shmd->shm_vpage,
1857 				    btopr(shm_amp->size));
1858 			}
1859 		}
1860 	} else {
1861 		error = hat_share(seg->s_as->a_hat, seg->s_base,
1862 		    shmd_arg->shm_sptas->a_hat, SEGSPTADDR,
1863 		    seg->s_size, seg->s_szc);
1864 	}
1865 	if (error) {
1866 		seg->s_szc = 0;
1867 		seg->s_data = (void *)NULL;
1868 		kmem_free(shmd, (sizeof (*shmd)));
1869 	} else {
1870 		ANON_LOCK_ENTER(&shm_amp->a_rwlock, RW_WRITER);
1871 		shm_amp->refcnt++;
1872 		ANON_LOCK_EXIT(&shm_amp->a_rwlock);
1873 	}
1874 	return (error);
1875 }
1876 
1877 int
1878 segspt_shmunmap(struct seg *seg, caddr_t raddr, size_t ssize)
1879 {
1880 	struct shm_data *shmd = (struct shm_data *)seg->s_data;
1881 	int reclaim = 1;
1882 
1883 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1884 retry:
1885 	if (shmd->shm_softlockcnt > 0) {
1886 		if (reclaim == 1) {
1887 			segspt_purge(seg);
1888 			reclaim = 0;
1889 			goto retry;
1890 		}
1891 		return (EAGAIN);
1892 	}
1893 
1894 	if (ssize != seg->s_size) {
1895 #ifdef DEBUG
1896 		cmn_err(CE_WARN, "Incompatible ssize %lx s_size %lx\n",
1897 		    ssize, seg->s_size);
1898 #endif
1899 		return (EINVAL);
1900 	}
1901 
1902 	(void) segspt_shmlockop(seg, raddr, shmd->shm_amp->size, 0, MC_UNLOCK,
1903 	    NULL, 0);
1904 	hat_unshare(seg->s_as->a_hat, raddr, ssize, seg->s_szc);
1905 
1906 	seg_free(seg);
1907 
1908 	return (0);
1909 }
1910 
1911 void
1912 segspt_shmfree(struct seg *seg)
1913 {
1914 	struct shm_data *shmd = (struct shm_data *)seg->s_data;
1915 	struct anon_map *shm_amp = shmd->shm_amp;
1916 
1917 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1918 
1919 	(void) segspt_shmlockop(seg, seg->s_base, shm_amp->size, 0,
1920 	    MC_UNLOCK, NULL, 0);
1921 
1922 	/*
1923 	 * Need to increment refcnt when attaching
1924 	 * and decrement when detaching because of dup().
1925 	 */
1926 	ANON_LOCK_ENTER(&shm_amp->a_rwlock, RW_WRITER);
1927 	shm_amp->refcnt--;
1928 	ANON_LOCK_EXIT(&shm_amp->a_rwlock);
1929 
1930 	if (shmd->shm_vpage) {	/* only for DISM */
1931 		kmem_free(shmd->shm_vpage, btopr(shm_amp->size));
1932 		shmd->shm_vpage = NULL;
1933 	}
1934 
1935 	/*
1936 	 * Take shm_segfree_syncmtx lock to let segspt_reclaim() finish if it's
1937 	 * still working with this segment without holding as lock.
1938 	 */
1939 	ASSERT(shmd->shm_softlockcnt == 0);
1940 	mutex_enter(&shmd->shm_segfree_syncmtx);
1941 	mutex_destroy(&shmd->shm_segfree_syncmtx);
1942 
1943 	kmem_free(shmd, sizeof (*shmd));
1944 }
1945 
1946 /*ARGSUSED*/
1947 int
1948 segspt_shmsetprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
1949 {
1950 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
1951 
1952 	/*
1953 	 * Shared page table is more than shared mapping.
1954 	 *  Individual process sharing page tables can't change prot
1955 	 *  because there is only one set of page tables.
1956 	 *  This will be allowed after private page table is
1957 	 *  supported.
1958 	 */
1959 /* need to return correct status error? */
1960 	return (0);
1961 }
1962 
1963 
1964 faultcode_t
1965 segspt_dismfault(struct hat *hat, struct seg *seg, caddr_t addr,
1966     size_t len, enum fault_type type, enum seg_rw rw)
1967 {
1968 	struct  shm_data	*shmd = (struct shm_data *)seg->s_data;
1969 	struct  seg		*sptseg = shmd->shm_sptseg;
1970 	struct  as		*curspt = shmd->shm_sptas;
1971 	struct  spt_data	*sptd = sptseg->s_data;
1972 	pgcnt_t npages;
1973 	size_t  size;
1974 	caddr_t segspt_addr, shm_addr;
1975 	page_t  **ppa;
1976 	int	i;
1977 	ulong_t an_idx = 0;
1978 	int	err = 0;
1979 	int	dyn_ism_unmap = hat_supported(HAT_DYNAMIC_ISM_UNMAP, (void *)0);
1980 	size_t	pgsz;
1981 	pgcnt_t	pgcnt;
1982 	caddr_t	a;
1983 	pgcnt_t	pidx;
1984 
1985 #ifdef lint
1986 	hat = hat;
1987 #endif
1988 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
1989 
1990 	/*
1991 	 * Because of the way spt is implemented
1992 	 * the realsize of the segment does not have to be
1993 	 * equal to the segment size itself. The segment size is
1994 	 * often in multiples of a page size larger than PAGESIZE.
1995 	 * The realsize is rounded up to the nearest PAGESIZE
1996 	 * based on what the user requested. This is a bit of
1997 	 * ungliness that is historical but not easily fixed
1998 	 * without re-designing the higher levels of ISM.
1999 	 */
2000 	ASSERT(addr >= seg->s_base);
2001 	if (((addr + len) - seg->s_base) > sptd->spt_realsize)
2002 		return (FC_NOMAP);
2003 	/*
2004 	 * For all of the following cases except F_PROT, we need to
2005 	 * make any necessary adjustments to addr and len
2006 	 * and get all of the necessary page_t's into an array called ppa[].
2007 	 *
2008 	 * The code in shmat() forces base addr and len of ISM segment
2009 	 * to be aligned to largest page size supported. Therefore,
2010 	 * we are able to handle F_SOFTLOCK and F_INVAL calls in "large
2011 	 * pagesize" chunks. We want to make sure that we HAT_LOAD_LOCK
2012 	 * in large pagesize chunks, or else we will screw up the HAT
2013 	 * layer by calling hat_memload_array() with differing page sizes
2014 	 * over a given virtual range.
2015 	 */
2016 	pgsz = page_get_pagesize(sptseg->s_szc);
2017 	pgcnt = page_get_pagecnt(sptseg->s_szc);
2018 	shm_addr = (caddr_t)P2ALIGN((uintptr_t)(addr), pgsz);
2019 	size = P2ROUNDUP((uintptr_t)(((addr + len) - shm_addr)), pgsz);
2020 	npages = btopr(size);
2021 
2022 	/*
2023 	 * Now we need to convert from addr in segshm to addr in segspt.
2024 	 */
2025 	an_idx = seg_page(seg, shm_addr);
2026 	segspt_addr = sptseg->s_base + ptob(an_idx);
2027 
2028 	ASSERT((segspt_addr + ptob(npages)) <=
2029 	    (sptseg->s_base + sptd->spt_realsize));
2030 	ASSERT(segspt_addr < (sptseg->s_base + sptseg->s_size));
2031 
2032 	switch (type) {
2033 
2034 	case F_SOFTLOCK:
2035 
2036 		atomic_add_long((ulong_t *)(&(shmd->shm_softlockcnt)), npages);
2037 		/*
2038 		 * Fall through to the F_INVAL case to load up the hat layer
2039 		 * entries with the HAT_LOAD_LOCK flag.
2040 		 */
2041 		/* FALLTHRU */
2042 	case F_INVAL:
2043 
2044 		if ((rw == S_EXEC) && !(sptd->spt_prot & PROT_EXEC))
2045 			return (FC_NOMAP);
2046 
2047 		ppa = kmem_zalloc(npages * sizeof (page_t *), KM_SLEEP);
2048 
2049 		err = spt_anon_getpages(sptseg, segspt_addr, size, ppa);
2050 		if (err != 0) {
2051 			if (type == F_SOFTLOCK) {
2052 				atomic_add_long((ulong_t *)(
2053 				    &(shmd->shm_softlockcnt)), -npages);
2054 			}
2055 			goto dism_err;
2056 		}
2057 		AS_LOCK_ENTER(sptseg->s_as, RW_READER);
2058 		a = segspt_addr;
2059 		pidx = 0;
2060 		if (type == F_SOFTLOCK) {
2061 
2062 			/*
2063 			 * Load up the translation keeping it
2064 			 * locked and don't unlock the page.
2065 			 */
2066 			for (; pidx < npages; a += pgsz, pidx += pgcnt) {
2067 				hat_memload_array(sptseg->s_as->a_hat,
2068 				    a, pgsz, &ppa[pidx], sptd->spt_prot,
2069 				    HAT_LOAD_LOCK | HAT_LOAD_SHARE);
2070 			}
2071 		} else {
2072 			/*
2073 			 * Migrate pages marked for migration
2074 			 */
2075 			if (lgrp_optimizations())
2076 				page_migrate(seg, shm_addr, ppa, npages);
2077 
2078 			for (; pidx < npages; a += pgsz, pidx += pgcnt) {
2079 				hat_memload_array(sptseg->s_as->a_hat,
2080 				    a, pgsz, &ppa[pidx],
2081 				    sptd->spt_prot,
2082 				    HAT_LOAD_SHARE);
2083 			}
2084 
2085 			/*
2086 			 * And now drop the SE_SHARED lock(s).
2087 			 */
2088 			if (dyn_ism_unmap) {
2089 				for (i = 0; i < npages; i++) {
2090 					page_unlock(ppa[i]);
2091 				}
2092 			}
2093 		}
2094 
2095 		if (!dyn_ism_unmap) {
2096 			if (hat_share(seg->s_as->a_hat, shm_addr,
2097 			    curspt->a_hat, segspt_addr, ptob(npages),
2098 			    seg->s_szc) != 0) {
2099 				panic("hat_share err in DISM fault");
2100 				/* NOTREACHED */
2101 			}
2102 			if (type == F_INVAL) {
2103 				for (i = 0; i < npages; i++) {
2104 					page_unlock(ppa[i]);
2105 				}
2106 			}
2107 		}
2108 		AS_LOCK_EXIT(sptseg->s_as);
2109 dism_err:
2110 		kmem_free(ppa, npages * sizeof (page_t *));
2111 		return (err);
2112 
2113 	case F_SOFTUNLOCK:
2114 
2115 		/*
2116 		 * This is a bit ugly, we pass in the real seg pointer,
2117 		 * but the segspt_addr is the virtual address within the
2118 		 * dummy seg.
2119 		 */
2120 		segspt_softunlock(seg, segspt_addr, size, rw);
2121 		return (0);
2122 
2123 	case F_PROT:
2124 
2125 		/*
2126 		 * This takes care of the unusual case where a user
2127 		 * allocates a stack in shared memory and a register
2128 		 * window overflow is written to that stack page before
2129 		 * it is otherwise modified.
2130 		 *
2131 		 * We can get away with this because ISM segments are
2132 		 * always rw. Other than this unusual case, there
2133 		 * should be no instances of protection violations.
2134 		 */
2135 		return (0);
2136 
2137 	default:
2138 #ifdef DEBUG
2139 		panic("segspt_dismfault default type?");
2140 #else
2141 		return (FC_NOMAP);
2142 #endif
2143 	}
2144 }
2145 
2146 
2147 faultcode_t
2148 segspt_shmfault(struct hat *hat, struct seg *seg, caddr_t addr,
2149     size_t len, enum fault_type type, enum seg_rw rw)
2150 {
2151 	struct shm_data		*shmd = (struct shm_data *)seg->s_data;
2152 	struct seg		*sptseg = shmd->shm_sptseg;
2153 	struct as		*curspt = shmd->shm_sptas;
2154 	struct spt_data		*sptd = sptseg->s_data;
2155 	pgcnt_t npages;
2156 	size_t size;
2157 	caddr_t sptseg_addr, shm_addr;
2158 	page_t *pp, **ppa;
2159 	int	i;
2160 	u_offset_t offset;
2161 	ulong_t anon_index = 0;
2162 	struct vnode *vp;
2163 	struct anon_map *amp;		/* XXX - for locknest */
2164 	struct anon *ap = NULL;
2165 	size_t		pgsz;
2166 	pgcnt_t		pgcnt;
2167 	caddr_t		a;
2168 	pgcnt_t		pidx;
2169 	size_t		sz;
2170 
2171 #ifdef lint
2172 	hat = hat;
2173 #endif
2174 
2175 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2176 
2177 	if (sptd->spt_flags & SHM_PAGEABLE) {
2178 		return (segspt_dismfault(hat, seg, addr, len, type, rw));
2179 	}
2180 
2181 	/*
2182 	 * Because of the way spt is implemented
2183 	 * the realsize of the segment does not have to be
2184 	 * equal to the segment size itself. The segment size is
2185 	 * often in multiples of a page size larger than PAGESIZE.
2186 	 * The realsize is rounded up to the nearest PAGESIZE
2187 	 * based on what the user requested. This is a bit of
2188 	 * ungliness that is historical but not easily fixed
2189 	 * without re-designing the higher levels of ISM.
2190 	 */
2191 	ASSERT(addr >= seg->s_base);
2192 	if (((addr + len) - seg->s_base) > sptd->spt_realsize)
2193 		return (FC_NOMAP);
2194 	/*
2195 	 * For all of the following cases except F_PROT, we need to
2196 	 * make any necessary adjustments to addr and len
2197 	 * and get all of the necessary page_t's into an array called ppa[].
2198 	 *
2199 	 * The code in shmat() forces base addr and len of ISM segment
2200 	 * to be aligned to largest page size supported. Therefore,
2201 	 * we are able to handle F_SOFTLOCK and F_INVAL calls in "large
2202 	 * pagesize" chunks. We want to make sure that we HAT_LOAD_LOCK
2203 	 * in large pagesize chunks, or else we will screw up the HAT
2204 	 * layer by calling hat_memload_array() with differing page sizes
2205 	 * over a given virtual range.
2206 	 */
2207 	pgsz = page_get_pagesize(sptseg->s_szc);
2208 	pgcnt = page_get_pagecnt(sptseg->s_szc);
2209 	shm_addr = (caddr_t)P2ALIGN((uintptr_t)(addr), pgsz);
2210 	size = P2ROUNDUP((uintptr_t)(((addr + len) - shm_addr)), pgsz);
2211 	npages = btopr(size);
2212 
2213 	/*
2214 	 * Now we need to convert from addr in segshm to addr in segspt.
2215 	 */
2216 	anon_index = seg_page(seg, shm_addr);
2217 	sptseg_addr = sptseg->s_base + ptob(anon_index);
2218 
2219 	/*
2220 	 * And now we may have to adjust npages downward if we have
2221 	 * exceeded the realsize of the segment or initial anon
2222 	 * allocations.
2223 	 */
2224 	if ((sptseg_addr + ptob(npages)) >
2225 	    (sptseg->s_base + sptd->spt_realsize))
2226 		size = (sptseg->s_base + sptd->spt_realsize) - sptseg_addr;
2227 
2228 	npages = btopr(size);
2229 
2230 	ASSERT(sptseg_addr < (sptseg->s_base + sptseg->s_size));
2231 	ASSERT((sptd->spt_flags & SHM_PAGEABLE) == 0);
2232 
2233 	switch (type) {
2234 
2235 	case F_SOFTLOCK:
2236 
2237 		/*
2238 		 * availrmem is decremented once during anon_swap_adjust()
2239 		 * and is incremented during the anon_unresv(), which is
2240 		 * called from shm_rm_amp() when the segment is destroyed.
2241 		 */
2242 		atomic_add_long((ulong_t *)(&(shmd->shm_softlockcnt)), npages);
2243 		/*
2244 		 * Some platforms assume that ISM pages are SE_SHARED
2245 		 * locked for the entire life of the segment.
2246 		 */
2247 		if (!hat_supported(HAT_DYNAMIC_ISM_UNMAP, (void *)0))
2248 			return (0);
2249 		/*
2250 		 * Fall through to the F_INVAL case to load up the hat layer
2251 		 * entries with the HAT_LOAD_LOCK flag.
2252 		 */
2253 
2254 		/* FALLTHRU */
2255 	case F_INVAL:
2256 
2257 		if ((rw == S_EXEC) && !(sptd->spt_prot & PROT_EXEC))
2258 			return (FC_NOMAP);
2259 
2260 		/*
2261 		 * Some platforms that do NOT support DYNAMIC_ISM_UNMAP
2262 		 * may still rely on this call to hat_share(). That
2263 		 * would imply that those hat's can fault on a
2264 		 * HAT_LOAD_LOCK translation, which would seem
2265 		 * contradictory.
2266 		 */
2267 		if (!hat_supported(HAT_DYNAMIC_ISM_UNMAP, (void *)0)) {
2268 			if (hat_share(seg->s_as->a_hat, seg->s_base,
2269 			    curspt->a_hat, sptseg->s_base,
2270 			    sptseg->s_size, sptseg->s_szc) != 0) {
2271 				panic("hat_share error in ISM fault");
2272 				/*NOTREACHED*/
2273 			}
2274 			return (0);
2275 		}
2276 		ppa = kmem_zalloc(sizeof (page_t *) * npages, KM_SLEEP);
2277 
2278 		/*
2279 		 * I see no need to lock the real seg,
2280 		 * here, because all of our work will be on the underlying
2281 		 * dummy seg.
2282 		 *
2283 		 * sptseg_addr and npages now account for large pages.
2284 		 */
2285 		amp = sptd->spt_amp;
2286 		ASSERT(amp != NULL);
2287 		anon_index = seg_page(sptseg, sptseg_addr);
2288 
2289 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2290 		for (i = 0; i < npages; i++) {
2291 			ap = anon_get_ptr(amp->ahp, anon_index++);
2292 			ASSERT(ap != NULL);
2293 			swap_xlate(ap, &vp, &offset);
2294 			pp = page_lookup(vp, offset, SE_SHARED);
2295 			ASSERT(pp != NULL);
2296 			ppa[i] = pp;
2297 		}
2298 		ANON_LOCK_EXIT(&amp->a_rwlock);
2299 		ASSERT(i == npages);
2300 
2301 		/*
2302 		 * We are already holding the as->a_lock on the user's
2303 		 * real segment, but we need to hold the a_lock on the
2304 		 * underlying dummy as. This is mostly to satisfy the
2305 		 * underlying HAT layer.
2306 		 */
2307 		AS_LOCK_ENTER(sptseg->s_as, RW_READER);
2308 		a = sptseg_addr;
2309 		pidx = 0;
2310 		if (type == F_SOFTLOCK) {
2311 			/*
2312 			 * Load up the translation keeping it
2313 			 * locked and don't unlock the page.
2314 			 */
2315 			for (; pidx < npages; a += pgsz, pidx += pgcnt) {
2316 				sz = MIN(pgsz, ptob(npages - pidx));
2317 				hat_memload_array(sptseg->s_as->a_hat, a,
2318 				    sz, &ppa[pidx], sptd->spt_prot,
2319 				    HAT_LOAD_LOCK | HAT_LOAD_SHARE);
2320 			}
2321 		} else {
2322 			/*
2323 			 * Migrate pages marked for migration.
2324 			 */
2325 			if (lgrp_optimizations())
2326 				page_migrate(seg, shm_addr, ppa, npages);
2327 
2328 			for (; pidx < npages; a += pgsz, pidx += pgcnt) {
2329 				sz = MIN(pgsz, ptob(npages - pidx));
2330 				hat_memload_array(sptseg->s_as->a_hat,
2331 				    a, sz, &ppa[pidx],
2332 				    sptd->spt_prot, HAT_LOAD_SHARE);
2333 			}
2334 
2335 			/*
2336 			 * And now drop the SE_SHARED lock(s).
2337 			 */
2338 			for (i = 0; i < npages; i++)
2339 				page_unlock(ppa[i]);
2340 		}
2341 		AS_LOCK_EXIT(sptseg->s_as);
2342 
2343 		kmem_free(ppa, sizeof (page_t *) * npages);
2344 		return (0);
2345 	case F_SOFTUNLOCK:
2346 
2347 		/*
2348 		 * This is a bit ugly, we pass in the real seg pointer,
2349 		 * but the sptseg_addr is the virtual address within the
2350 		 * dummy seg.
2351 		 */
2352 		segspt_softunlock(seg, sptseg_addr, ptob(npages), rw);
2353 		return (0);
2354 
2355 	case F_PROT:
2356 
2357 		/*
2358 		 * This takes care of the unusual case where a user
2359 		 * allocates a stack in shared memory and a register
2360 		 * window overflow is written to that stack page before
2361 		 * it is otherwise modified.
2362 		 *
2363 		 * We can get away with this because ISM segments are
2364 		 * always rw. Other than this unusual case, there
2365 		 * should be no instances of protection violations.
2366 		 */
2367 		return (0);
2368 
2369 	default:
2370 #ifdef DEBUG
2371 		cmn_err(CE_WARN, "segspt_shmfault default type?");
2372 #endif
2373 		return (FC_NOMAP);
2374 	}
2375 }
2376 
2377 /*ARGSUSED*/
2378 static faultcode_t
2379 segspt_shmfaulta(struct seg *seg, caddr_t addr)
2380 {
2381 	return (0);
2382 }
2383 
2384 /*ARGSUSED*/
2385 static int
2386 segspt_shmkluster(struct seg *seg, caddr_t addr, ssize_t delta)
2387 {
2388 	return (0);
2389 }
2390 
2391 /*ARGSUSED*/
2392 static size_t
2393 segspt_shmswapout(struct seg *seg)
2394 {
2395 	return (0);
2396 }
2397 
2398 /*
2399  * duplicate the shared page tables
2400  */
2401 int
2402 segspt_shmdup(struct seg *seg, struct seg *newseg)
2403 {
2404 	struct shm_data		*shmd = (struct shm_data *)seg->s_data;
2405 	struct anon_map		*amp = shmd->shm_amp;
2406 	struct shm_data		*shmd_new;
2407 	struct seg		*spt_seg = shmd->shm_sptseg;
2408 	struct spt_data		*sptd = spt_seg->s_data;
2409 	int			error = 0;
2410 
2411 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
2412 
2413 	shmd_new = kmem_zalloc((sizeof (*shmd_new)), KM_SLEEP);
2414 	newseg->s_data = (void *)shmd_new;
2415 	shmd_new->shm_sptas = shmd->shm_sptas;
2416 	shmd_new->shm_amp = amp;
2417 	shmd_new->shm_sptseg = shmd->shm_sptseg;
2418 	newseg->s_ops = &segspt_shmops;
2419 	newseg->s_szc = seg->s_szc;
2420 	ASSERT(seg->s_szc == shmd->shm_sptseg->s_szc);
2421 
2422 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2423 	amp->refcnt++;
2424 	ANON_LOCK_EXIT(&amp->a_rwlock);
2425 
2426 	if (sptd->spt_flags & SHM_PAGEABLE) {
2427 		shmd_new->shm_vpage = kmem_zalloc(btopr(amp->size), KM_SLEEP);
2428 		shmd_new->shm_lckpgs = 0;
2429 		if (hat_supported(HAT_DYNAMIC_ISM_UNMAP, (void *)0)) {
2430 			if ((error = hat_share(newseg->s_as->a_hat,
2431 			    newseg->s_base, shmd->shm_sptas->a_hat, SEGSPTADDR,
2432 			    seg->s_size, seg->s_szc)) != 0) {
2433 				kmem_free(shmd_new->shm_vpage,
2434 				    btopr(amp->size));
2435 			}
2436 		}
2437 		return (error);
2438 	} else {
2439 		return (hat_share(newseg->s_as->a_hat, newseg->s_base,
2440 		    shmd->shm_sptas->a_hat, SEGSPTADDR, seg->s_size,
2441 		    seg->s_szc));
2442 
2443 	}
2444 }
2445 
2446 /*ARGSUSED*/
2447 int
2448 segspt_shmcheckprot(struct seg *seg, caddr_t addr, size_t size, uint_t prot)
2449 {
2450 	struct shm_data *shmd = (struct shm_data *)seg->s_data;
2451 	struct spt_data *sptd = (struct spt_data *)shmd->shm_sptseg->s_data;
2452 
2453 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2454 
2455 	/*
2456 	 * ISM segment is always rw.
2457 	 */
2458 	return (((sptd->spt_prot & prot) != prot) ? EACCES : 0);
2459 }
2460 
2461 /*
2462  * Return an array of locked large pages, for empty slots allocate
2463  * private zero-filled anon pages.
2464  */
2465 static int
2466 spt_anon_getpages(
2467 	struct seg *sptseg,
2468 	caddr_t sptaddr,
2469 	size_t len,
2470 	page_t *ppa[])
2471 {
2472 	struct  spt_data *sptd = sptseg->s_data;
2473 	struct  anon_map *amp = sptd->spt_amp;
2474 	enum	seg_rw rw = sptd->spt_prot;
2475 	uint_t	szc = sptseg->s_szc;
2476 	size_t	pg_sz, share_sz = page_get_pagesize(szc);
2477 	pgcnt_t	lp_npgs;
2478 	caddr_t	lp_addr, e_sptaddr;
2479 	uint_t	vpprot, ppa_szc = 0;
2480 	struct  vpage *vpage = NULL;
2481 	ulong_t	j, ppa_idx;
2482 	int	err, ierr = 0;
2483 	pgcnt_t	an_idx;
2484 	anon_sync_obj_t cookie;
2485 	int anon_locked = 0;
2486 	pgcnt_t amp_pgs;
2487 
2488 
2489 	ASSERT(IS_P2ALIGNED(sptaddr, share_sz) && IS_P2ALIGNED(len, share_sz));
2490 	ASSERT(len != 0);
2491 
2492 	pg_sz = share_sz;
2493 	lp_npgs = btop(pg_sz);
2494 	lp_addr = sptaddr;
2495 	e_sptaddr = sptaddr + len;
2496 	an_idx = seg_page(sptseg, sptaddr);
2497 	ppa_idx = 0;
2498 
2499 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2500 
2501 	amp_pgs = page_get_pagecnt(amp->a_szc);
2502 
2503 	/*CONSTCOND*/
2504 	while (1) {
2505 		for (; lp_addr < e_sptaddr;
2506 		    an_idx += lp_npgs, lp_addr += pg_sz, ppa_idx += lp_npgs) {
2507 
2508 			/*
2509 			 * If we're currently locked, and we get to a new
2510 			 * page, unlock our current anon chunk.
2511 			 */
2512 			if (anon_locked && P2PHASE(an_idx, amp_pgs) == 0) {
2513 				anon_array_exit(&cookie);
2514 				anon_locked = 0;
2515 			}
2516 			if (!anon_locked) {
2517 				anon_array_enter(amp, an_idx, &cookie);
2518 				anon_locked = 1;
2519 			}
2520 			ppa_szc = (uint_t)-1;
2521 			ierr = anon_map_getpages(amp, an_idx, szc, sptseg,
2522 			    lp_addr, sptd->spt_prot, &vpprot, &ppa[ppa_idx],
2523 			    &ppa_szc, vpage, rw, 0, segvn_anypgsz, 0, kcred);
2524 
2525 			if (ierr != 0) {
2526 				if (ierr > 0) {
2527 					err = FC_MAKE_ERR(ierr);
2528 					goto lpgs_err;
2529 				}
2530 				break;
2531 			}
2532 		}
2533 		if (lp_addr == e_sptaddr) {
2534 			break;
2535 		}
2536 		ASSERT(lp_addr < e_sptaddr);
2537 
2538 		/*
2539 		 * ierr == -1 means we failed to allocate a large page.
2540 		 * so do a size down operation.
2541 		 *
2542 		 * ierr == -2 means some other process that privately shares
2543 		 * pages with this process has allocated a larger page and we
2544 		 * need to retry with larger pages. So do a size up
2545 		 * operation. This relies on the fact that large pages are
2546 		 * never partially shared i.e. if we share any constituent
2547 		 * page of a large page with another process we must share the
2548 		 * entire large page. Note this cannot happen for SOFTLOCK
2549 		 * case, unless current address (lpaddr) is at the beginning
2550 		 * of the next page size boundary because the other process
2551 		 * couldn't have relocated locked pages.
2552 		 */
2553 		ASSERT(ierr == -1 || ierr == -2);
2554 		if (segvn_anypgsz) {
2555 			ASSERT(ierr == -2 || szc != 0);
2556 			ASSERT(ierr == -1 || szc < sptseg->s_szc);
2557 			szc = (ierr == -1) ? szc - 1 : szc + 1;
2558 		} else {
2559 			/*
2560 			 * For faults and segvn_anypgsz == 0
2561 			 * we need to be careful not to loop forever
2562 			 * if existing page is found with szc other
2563 			 * than 0 or seg->s_szc. This could be due
2564 			 * to page relocations on behalf of DR or
2565 			 * more likely large page creation. For this
2566 			 * case simply re-size to existing page's szc
2567 			 * if returned by anon_map_getpages().
2568 			 */
2569 			if (ppa_szc == (uint_t)-1) {
2570 				szc = (ierr == -1) ? 0 : sptseg->s_szc;
2571 			} else {
2572 				ASSERT(ppa_szc <= sptseg->s_szc);
2573 				ASSERT(ierr == -2 || ppa_szc < szc);
2574 				ASSERT(ierr == -1 || ppa_szc > szc);
2575 				szc = ppa_szc;
2576 			}
2577 		}
2578 		pg_sz = page_get_pagesize(szc);
2579 		lp_npgs = btop(pg_sz);
2580 		ASSERT(IS_P2ALIGNED(lp_addr, pg_sz));
2581 	}
2582 	if (anon_locked) {
2583 		anon_array_exit(&cookie);
2584 	}
2585 	ANON_LOCK_EXIT(&amp->a_rwlock);
2586 	return (0);
2587 
2588 lpgs_err:
2589 	if (anon_locked) {
2590 		anon_array_exit(&cookie);
2591 	}
2592 	ANON_LOCK_EXIT(&amp->a_rwlock);
2593 	for (j = 0; j < ppa_idx; j++)
2594 		page_unlock(ppa[j]);
2595 	return (err);
2596 }
2597 
2598 /*
2599  * count the number of bytes in a set of spt pages that are currently not
2600  * locked
2601  */
2602 static rctl_qty_t
2603 spt_unlockedbytes(pgcnt_t npages, page_t **ppa)
2604 {
2605 	ulong_t	i;
2606 	rctl_qty_t unlocked = 0;
2607 
2608 	for (i = 0; i < npages; i++) {
2609 		if (ppa[i]->p_lckcnt == 0)
2610 			unlocked += PAGESIZE;
2611 	}
2612 	return (unlocked);
2613 }
2614 
2615 extern	u_longlong_t randtick(void);
2616 /* number of locks to reserve/skip by spt_lockpages() and spt_unlockpages() */
2617 #define	NLCK	(NCPU_P2)
2618 /* Random number with a range [0, n-1], n must be power of two */
2619 #define	RAND_P2(n)	\
2620 	((((long)curthread >> PTR24_LSB) ^ (long)randtick()) & ((n) - 1))
2621 
2622 int
2623 spt_lockpages(struct seg *seg, pgcnt_t anon_index, pgcnt_t npages,
2624     page_t **ppa, ulong_t *lockmap, size_t pos,
2625     rctl_qty_t *locked)
2626 {
2627 	struct	shm_data *shmd = seg->s_data;
2628 	struct	spt_data *sptd = shmd->shm_sptseg->s_data;
2629 	ulong_t	i;
2630 	int	kernel;
2631 	pgcnt_t	nlck = 0;
2632 	int	rv = 0;
2633 	int	use_reserved = 1;
2634 
2635 	/* return the number of bytes actually locked */
2636 	*locked = 0;
2637 
2638 	/*
2639 	 * To avoid contention on freemem_lock, availrmem and pages_locked
2640 	 * global counters are updated only every nlck locked pages instead of
2641 	 * every time.  Reserve nlck locks up front and deduct from this
2642 	 * reservation for each page that requires a lock.  When the reservation
2643 	 * is consumed, reserve again.  nlck is randomized, so the competing
2644 	 * threads do not fall into a cyclic lock contention pattern. When
2645 	 * memory is low, the lock ahead is disabled, and instead page_pp_lock()
2646 	 * is used to lock pages.
2647 	 */
2648 	for (i = 0; i < npages; anon_index++, pos++, i++) {
2649 		if (nlck == 0 && use_reserved == 1) {
2650 			nlck = NLCK + RAND_P2(NLCK);
2651 			/* if fewer loops left, decrease nlck */
2652 			nlck = MIN(nlck, npages - i);
2653 			/*
2654 			 * Reserve nlck locks up front and deduct from this
2655 			 * reservation for each page that requires a lock.  When
2656 			 * the reservation is consumed, reserve again.
2657 			 */
2658 			mutex_enter(&freemem_lock);
2659 			if ((availrmem - nlck) < pages_pp_maximum) {
2660 				/* Do not do advance memory reserves */
2661 				use_reserved = 0;
2662 			} else {
2663 				availrmem	-= nlck;
2664 				pages_locked	+= nlck;
2665 			}
2666 			mutex_exit(&freemem_lock);
2667 		}
2668 		if (!(shmd->shm_vpage[anon_index] & DISM_PG_LOCKED)) {
2669 			if (sptd->spt_ppa_lckcnt[anon_index] <
2670 			    (ushort_t)DISM_LOCK_MAX) {
2671 				if (++sptd->spt_ppa_lckcnt[anon_index] ==
2672 				    (ushort_t)DISM_LOCK_MAX) {
2673 					cmn_err(CE_WARN,
2674 					    "DISM page lock limit "
2675 					    "reached on DISM offset 0x%lx\n",
2676 					    anon_index << PAGESHIFT);
2677 				}
2678 				kernel = (sptd->spt_ppa &&
2679 				    sptd->spt_ppa[anon_index]);
2680 				if (!page_pp_lock(ppa[i], 0, kernel ||
2681 				    use_reserved)) {
2682 					sptd->spt_ppa_lckcnt[anon_index]--;
2683 					rv = EAGAIN;
2684 					break;
2685 				}
2686 				/* if this is a newly locked page, count it */
2687 				if (ppa[i]->p_lckcnt == 1) {
2688 					if (kernel == 0 && use_reserved == 1)
2689 						nlck--;
2690 					*locked += PAGESIZE;
2691 				}
2692 				shmd->shm_lckpgs++;
2693 				shmd->shm_vpage[anon_index] |= DISM_PG_LOCKED;
2694 				if (lockmap != NULL)
2695 					BT_SET(lockmap, pos);
2696 			}
2697 		}
2698 	}
2699 	/* Return unused lock reservation */
2700 	if (nlck != 0 && use_reserved == 1) {
2701 		mutex_enter(&freemem_lock);
2702 		availrmem	+= nlck;
2703 		pages_locked	-= nlck;
2704 		mutex_exit(&freemem_lock);
2705 	}
2706 
2707 	return (rv);
2708 }
2709 
2710 int
2711 spt_unlockpages(struct seg *seg, pgcnt_t anon_index, pgcnt_t npages,
2712     rctl_qty_t *unlocked)
2713 {
2714 	struct shm_data	*shmd = seg->s_data;
2715 	struct spt_data	*sptd = shmd->shm_sptseg->s_data;
2716 	struct anon_map	*amp = sptd->spt_amp;
2717 	struct anon	*ap;
2718 	struct vnode	*vp;
2719 	u_offset_t	off;
2720 	struct page	*pp;
2721 	int		kernel;
2722 	anon_sync_obj_t	cookie;
2723 	ulong_t		i;
2724 	pgcnt_t		nlck = 0;
2725 	pgcnt_t		nlck_limit = NLCK;
2726 
2727 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2728 	for (i = 0; i < npages; i++, anon_index++) {
2729 		if (shmd->shm_vpage[anon_index] & DISM_PG_LOCKED) {
2730 			anon_array_enter(amp, anon_index, &cookie);
2731 			ap = anon_get_ptr(amp->ahp, anon_index);
2732 			ASSERT(ap);
2733 
2734 			swap_xlate(ap, &vp, &off);
2735 			anon_array_exit(&cookie);
2736 			pp = page_lookup(vp, off, SE_SHARED);
2737 			ASSERT(pp);
2738 			/*
2739 			 * availrmem is decremented only for pages which are not
2740 			 * in seg pcache, for pages in seg pcache availrmem was
2741 			 * decremented in _dismpagelock()
2742 			 */
2743 			kernel = (sptd->spt_ppa && sptd->spt_ppa[anon_index]);
2744 			ASSERT(pp->p_lckcnt > 0);
2745 
2746 			/*
2747 			 * lock page but do not change availrmem, we do it
2748 			 * ourselves every nlck loops.
2749 			 */
2750 			page_pp_unlock(pp, 0, 1);
2751 			if (pp->p_lckcnt == 0) {
2752 				if (kernel == 0)
2753 					nlck++;
2754 				*unlocked += PAGESIZE;
2755 			}
2756 			page_unlock(pp);
2757 			shmd->shm_vpage[anon_index] &= ~DISM_PG_LOCKED;
2758 			sptd->spt_ppa_lckcnt[anon_index]--;
2759 			shmd->shm_lckpgs--;
2760 		}
2761 
2762 		/*
2763 		 * To reduce freemem_lock contention, do not update availrmem
2764 		 * until at least NLCK pages have been unlocked.
2765 		 * 1. No need to update if nlck is zero
2766 		 * 2. Always update if the last iteration
2767 		 */
2768 		if (nlck > 0 && (nlck == nlck_limit || i == npages - 1)) {
2769 			mutex_enter(&freemem_lock);
2770 			availrmem	+= nlck;
2771 			pages_locked	-= nlck;
2772 			mutex_exit(&freemem_lock);
2773 			nlck = 0;
2774 			nlck_limit = NLCK + RAND_P2(NLCK);
2775 		}
2776 	}
2777 	ANON_LOCK_EXIT(&amp->a_rwlock);
2778 
2779 	return (0);
2780 }
2781 
2782 /*ARGSUSED*/
2783 static int
2784 segspt_shmlockop(struct seg *seg, caddr_t addr, size_t len,
2785     int attr, int op, ulong_t *lockmap, size_t pos)
2786 {
2787 	struct shm_data *shmd = seg->s_data;
2788 	struct seg	*sptseg = shmd->shm_sptseg;
2789 	struct spt_data *sptd = sptseg->s_data;
2790 	struct kshmid	*sp = sptd->spt_amp->a_sp;
2791 	pgcnt_t		npages, a_npages;
2792 	page_t		**ppa;
2793 	pgcnt_t		an_idx, a_an_idx, ppa_idx;
2794 	caddr_t		spt_addr, a_addr;	/* spt and aligned address */
2795 	size_t		a_len;			/* aligned len */
2796 	size_t		share_sz;
2797 	ulong_t		i;
2798 	int		sts = 0;
2799 	rctl_qty_t	unlocked = 0;
2800 	rctl_qty_t	locked = 0;
2801 	struct proc	*p = curproc;
2802 	kproject_t	*proj;
2803 
2804 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2805 	ASSERT(sp != NULL);
2806 
2807 	if ((sptd->spt_flags & SHM_PAGEABLE) == 0) {
2808 		return (0);
2809 	}
2810 
2811 	addr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
2812 	an_idx = seg_page(seg, addr);
2813 	npages = btopr(len);
2814 
2815 	if (an_idx + npages > btopr(shmd->shm_amp->size)) {
2816 		return (ENOMEM);
2817 	}
2818 
2819 	/*
2820 	 * A shm's project never changes, so no lock needed.
2821 	 * The shm has a hold on the project, so it will not go away.
2822 	 * Since we have a mapping to shm within this zone, we know
2823 	 * that the zone will not go away.
2824 	 */
2825 	proj = sp->shm_perm.ipc_proj;
2826 
2827 	if (op == MC_LOCK) {
2828 
2829 		/*
2830 		 * Need to align addr and size request if they are not
2831 		 * aligned so we can always allocate large page(s) however
2832 		 * we only lock what was requested in initial request.
2833 		 */
2834 		share_sz = page_get_pagesize(sptseg->s_szc);
2835 		a_addr = (caddr_t)P2ALIGN((uintptr_t)(addr), share_sz);
2836 		a_len = P2ROUNDUP((uintptr_t)(((addr + len) - a_addr)),
2837 		    share_sz);
2838 		a_npages = btop(a_len);
2839 		a_an_idx = seg_page(seg, a_addr);
2840 		spt_addr = sptseg->s_base + ptob(a_an_idx);
2841 		ppa_idx = an_idx - a_an_idx;
2842 
2843 		if ((ppa = kmem_zalloc(((sizeof (page_t *)) * a_npages),
2844 		    KM_NOSLEEP)) == NULL) {
2845 			return (ENOMEM);
2846 		}
2847 
2848 		/*
2849 		 * Don't cache any new pages for IO and
2850 		 * flush any cached pages.
2851 		 */
2852 		mutex_enter(&sptd->spt_lock);
2853 		if (sptd->spt_ppa != NULL)
2854 			sptd->spt_flags |= DISM_PPA_CHANGED;
2855 
2856 		sts = spt_anon_getpages(sptseg, spt_addr, a_len, ppa);
2857 		if (sts != 0) {
2858 			mutex_exit(&sptd->spt_lock);
2859 			kmem_free(ppa, ((sizeof (page_t *)) * a_npages));
2860 			return (sts);
2861 		}
2862 
2863 		mutex_enter(&sp->shm_mlock);
2864 		/* enforce locked memory rctl */
2865 		unlocked = spt_unlockedbytes(npages, &ppa[ppa_idx]);
2866 
2867 		mutex_enter(&p->p_lock);
2868 		if (rctl_incr_locked_mem(p, proj, unlocked, 0)) {
2869 			mutex_exit(&p->p_lock);
2870 			sts = EAGAIN;
2871 		} else {
2872 			mutex_exit(&p->p_lock);
2873 			sts = spt_lockpages(seg, an_idx, npages,
2874 			    &ppa[ppa_idx], lockmap, pos, &locked);
2875 
2876 			/*
2877 			 * correct locked count if not all pages could be
2878 			 * locked
2879 			 */
2880 			if ((unlocked - locked) > 0) {
2881 				rctl_decr_locked_mem(NULL, proj,
2882 				    (unlocked - locked), 0);
2883 			}
2884 		}
2885 		/*
2886 		 * unlock pages
2887 		 */
2888 		for (i = 0; i < a_npages; i++)
2889 			page_unlock(ppa[i]);
2890 		if (sptd->spt_ppa != NULL)
2891 			sptd->spt_flags |= DISM_PPA_CHANGED;
2892 		mutex_exit(&sp->shm_mlock);
2893 		mutex_exit(&sptd->spt_lock);
2894 
2895 		kmem_free(ppa, ((sizeof (page_t *)) * a_npages));
2896 
2897 	} else if (op == MC_UNLOCK) { /* unlock */
2898 		page_t		**ppa;
2899 
2900 		mutex_enter(&sptd->spt_lock);
2901 		if (shmd->shm_lckpgs == 0) {
2902 			mutex_exit(&sptd->spt_lock);
2903 			return (0);
2904 		}
2905 		/*
2906 		 * Don't cache new IO pages.
2907 		 */
2908 		if (sptd->spt_ppa != NULL)
2909 			sptd->spt_flags |= DISM_PPA_CHANGED;
2910 
2911 		mutex_enter(&sp->shm_mlock);
2912 		sts = spt_unlockpages(seg, an_idx, npages, &unlocked);
2913 		if ((ppa = sptd->spt_ppa) != NULL)
2914 			sptd->spt_flags |= DISM_PPA_CHANGED;
2915 		mutex_exit(&sptd->spt_lock);
2916 
2917 		rctl_decr_locked_mem(NULL, proj, unlocked, 0);
2918 		mutex_exit(&sp->shm_mlock);
2919 
2920 		if (ppa != NULL)
2921 			seg_ppurge_wiredpp(ppa);
2922 	}
2923 	return (sts);
2924 }
2925 
2926 /*ARGSUSED*/
2927 int
2928 segspt_shmgetprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
2929 {
2930 	struct shm_data *shmd = (struct shm_data *)seg->s_data;
2931 	struct spt_data *sptd = (struct spt_data *)shmd->shm_sptseg->s_data;
2932 	spgcnt_t pgno = seg_page(seg, addr+len) - seg_page(seg, addr) + 1;
2933 
2934 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2935 
2936 	/*
2937 	 * ISM segment is always rw.
2938 	 */
2939 	while (--pgno >= 0)
2940 		*protv++ = sptd->spt_prot;
2941 	return (0);
2942 }
2943 
2944 /*ARGSUSED*/
2945 u_offset_t
2946 segspt_shmgetoffset(struct seg *seg, caddr_t addr)
2947 {
2948 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2949 
2950 	/* Offset does not matter in ISM memory */
2951 
2952 	return ((u_offset_t)0);
2953 }
2954 
2955 /* ARGSUSED */
2956 int
2957 segspt_shmgettype(struct seg *seg, caddr_t addr)
2958 {
2959 	struct shm_data *shmd = (struct shm_data *)seg->s_data;
2960 	struct spt_data *sptd = (struct spt_data *)shmd->shm_sptseg->s_data;
2961 
2962 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2963 
2964 	/*
2965 	 * The shared memory mapping is always MAP_SHARED, SWAP is only
2966 	 * reserved for DISM
2967 	 */
2968 	return (MAP_SHARED |
2969 	    ((sptd->spt_flags & SHM_PAGEABLE) ? 0 : MAP_NORESERVE));
2970 }
2971 
2972 /*ARGSUSED*/
2973 int
2974 segspt_shmgetvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
2975 {
2976 	struct shm_data *shmd = (struct shm_data *)seg->s_data;
2977 	struct spt_data *sptd = (struct spt_data *)shmd->shm_sptseg->s_data;
2978 
2979 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2980 
2981 	*vpp = sptd->spt_vp;
2982 	return (0);
2983 }
2984 
2985 /*
2986  * We need to wait for pending IO to complete to a DISM segment in order for
2987  * pages to get kicked out of the seg_pcache.  120 seconds should be more
2988  * than enough time to wait.
2989  */
2990 static clock_t spt_pcache_wait = 120;
2991 
2992 /*ARGSUSED*/
2993 static int
2994 segspt_shmadvise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
2995 {
2996 	struct shm_data	*shmd = (struct shm_data *)seg->s_data;
2997 	struct spt_data	*sptd = (struct spt_data *)shmd->shm_sptseg->s_data;
2998 	struct anon_map	*amp;
2999 	pgcnt_t pg_idx;
3000 	ushort_t gen;
3001 	clock_t	end_lbolt;
3002 	int writer;
3003 	page_t **ppa;
3004 
3005 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
3006 
3007 	if (behav == MADV_FREE || behav == MADV_PURGE) {
3008 		if ((sptd->spt_flags & SHM_PAGEABLE) == 0)
3009 			return (0);
3010 
3011 		amp = sptd->spt_amp;
3012 		pg_idx = seg_page(seg, addr);
3013 
3014 		mutex_enter(&sptd->spt_lock);
3015 		if ((ppa = sptd->spt_ppa) == NULL) {
3016 			mutex_exit(&sptd->spt_lock);
3017 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
3018 			(void) anon_disclaim(amp, pg_idx, len, behav, NULL);
3019 			ANON_LOCK_EXIT(&amp->a_rwlock);
3020 			return (0);
3021 		}
3022 
3023 		sptd->spt_flags |= DISM_PPA_CHANGED;
3024 		gen = sptd->spt_gen;
3025 
3026 		mutex_exit(&sptd->spt_lock);
3027 
3028 		/*
3029 		 * Purge all DISM cached pages
3030 		 */
3031 		seg_ppurge_wiredpp(ppa);
3032 
3033 		/*
3034 		 * Drop the AS_LOCK so that other threads can grab it
3035 		 * in the as_pageunlock path and hopefully get the segment
3036 		 * kicked out of the seg_pcache.  We bump the shm_softlockcnt
3037 		 * to keep this segment resident.
3038 		 */
3039 		writer = AS_WRITE_HELD(seg->s_as);
3040 		atomic_inc_ulong((ulong_t *)(&(shmd->shm_softlockcnt)));
3041 		AS_LOCK_EXIT(seg->s_as);
3042 
3043 		mutex_enter(&sptd->spt_lock);
3044 
3045 		end_lbolt = ddi_get_lbolt() + (hz * spt_pcache_wait);
3046 
3047 		/*
3048 		 * Try to wait for pages to get kicked out of the seg_pcache.
3049 		 */
3050 		while (sptd->spt_gen == gen &&
3051 		    (sptd->spt_flags & DISM_PPA_CHANGED) &&
3052 		    ddi_get_lbolt() < end_lbolt) {
3053 			if (!cv_timedwait_sig(&sptd->spt_cv,
3054 			    &sptd->spt_lock, end_lbolt)) {
3055 				break;
3056 			}
3057 		}
3058 
3059 		mutex_exit(&sptd->spt_lock);
3060 
3061 		/* Regrab the AS_LOCK and release our hold on the segment */
3062 		AS_LOCK_ENTER(seg->s_as, writer ? RW_WRITER : RW_READER);
3063 		atomic_dec_ulong((ulong_t *)(&(shmd->shm_softlockcnt)));
3064 		if (shmd->shm_softlockcnt <= 0) {
3065 			if (AS_ISUNMAPWAIT(seg->s_as)) {
3066 				mutex_enter(&seg->s_as->a_contents);
3067 				if (AS_ISUNMAPWAIT(seg->s_as)) {
3068 					AS_CLRUNMAPWAIT(seg->s_as);
3069 					cv_broadcast(&seg->s_as->a_cv);
3070 				}
3071 				mutex_exit(&seg->s_as->a_contents);
3072 			}
3073 		}
3074 
3075 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
3076 		(void) anon_disclaim(amp, pg_idx, len, behav, NULL);
3077 		ANON_LOCK_EXIT(&amp->a_rwlock);
3078 	} else if (lgrp_optimizations() && (behav == MADV_ACCESS_LWP ||
3079 	    behav == MADV_ACCESS_MANY || behav == MADV_ACCESS_DEFAULT)) {
3080 		int			already_set;
3081 		ulong_t			anon_index;
3082 		lgrp_mem_policy_t	policy;
3083 		caddr_t			shm_addr;
3084 		size_t			share_size;
3085 		size_t			size;
3086 		struct seg		*sptseg = shmd->shm_sptseg;
3087 		caddr_t			sptseg_addr;
3088 
3089 		/*
3090 		 * Align address and length to page size of underlying segment
3091 		 */
3092 		share_size = page_get_pagesize(shmd->shm_sptseg->s_szc);
3093 		shm_addr = (caddr_t)P2ALIGN((uintptr_t)(addr), share_size);
3094 		size = P2ROUNDUP((uintptr_t)(((addr + len) - shm_addr)),
3095 		    share_size);
3096 
3097 		amp = shmd->shm_amp;
3098 		anon_index = seg_page(seg, shm_addr);
3099 
3100 		/*
3101 		 * And now we may have to adjust size downward if we have
3102 		 * exceeded the realsize of the segment or initial anon
3103 		 * allocations.
3104 		 */
3105 		sptseg_addr = sptseg->s_base + ptob(anon_index);
3106 		if ((sptseg_addr + size) >
3107 		    (sptseg->s_base + sptd->spt_realsize))
3108 			size = (sptseg->s_base + sptd->spt_realsize) -
3109 			    sptseg_addr;
3110 
3111 		/*
3112 		 * Set memory allocation policy for this segment
3113 		 */
3114 		policy = lgrp_madv_to_policy(behav, len, MAP_SHARED);
3115 		already_set = lgrp_shm_policy_set(policy, amp, anon_index,
3116 		    NULL, 0, len);
3117 
3118 		/*
3119 		 * If random memory allocation policy set already,
3120 		 * don't bother reapplying it.
3121 		 */
3122 		if (already_set && !LGRP_MEM_POLICY_REAPPLICABLE(policy))
3123 			return (0);
3124 
3125 		/*
3126 		 * Mark any existing pages in the given range for
3127 		 * migration, flushing the I/O page cache, and using
3128 		 * underlying segment to calculate anon index and get
3129 		 * anonmap and vnode pointer from
3130 		 */
3131 		if (shmd->shm_softlockcnt > 0)
3132 			segspt_purge(seg);
3133 
3134 		page_mark_migrate(seg, shm_addr, size, amp, 0, NULL, 0, 0);
3135 	}
3136 
3137 	return (0);
3138 }
3139 
3140 /*ARGSUSED*/
3141 void
3142 segspt_shmdump(struct seg *seg)
3143 {
3144 	/* no-op for ISM segment */
3145 }
3146 
3147 /*ARGSUSED*/
3148 static int
3149 segspt_shmsetpgsz(struct seg *seg, caddr_t addr, size_t len, uint_t szc)
3150 {
3151 	return (ENOTSUP);
3152 }
3153 
3154 /*
3155  * get a memory ID for an addr in a given segment
3156  */
3157 static int
3158 segspt_shmgetmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
3159 {
3160 	struct shm_data *shmd = (struct shm_data *)seg->s_data;
3161 	struct anon	*ap;
3162 	size_t		anon_index;
3163 	struct anon_map	*amp = shmd->shm_amp;
3164 	struct spt_data	*sptd = shmd->shm_sptseg->s_data;
3165 	struct seg	*sptseg = shmd->shm_sptseg;
3166 	anon_sync_obj_t	cookie;
3167 
3168 	anon_index = seg_page(seg, addr);
3169 
3170 	if (addr > (seg->s_base + sptd->spt_realsize)) {
3171 		return (EFAULT);
3172 	}
3173 
3174 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
3175 	anon_array_enter(amp, anon_index, &cookie);
3176 	ap = anon_get_ptr(amp->ahp, anon_index);
3177 	if (ap == NULL) {
3178 		struct page *pp;
3179 		caddr_t spt_addr = sptseg->s_base + ptob(anon_index);
3180 
3181 		pp = anon_zero(sptseg, spt_addr, &ap, kcred);
3182 		if (pp == NULL) {
3183 			anon_array_exit(&cookie);
3184 			ANON_LOCK_EXIT(&amp->a_rwlock);
3185 			return (ENOMEM);
3186 		}
3187 		(void) anon_set_ptr(amp->ahp, anon_index, ap, ANON_SLEEP);
3188 		page_unlock(pp);
3189 	}
3190 	anon_array_exit(&cookie);
3191 	ANON_LOCK_EXIT(&amp->a_rwlock);
3192 	memidp->val[0] = (uintptr_t)ap;
3193 	memidp->val[1] = (uintptr_t)addr & PAGEOFFSET;
3194 	return (0);
3195 }
3196 
3197 /*
3198  * Get memory allocation policy info for specified address in given segment
3199  */
3200 static lgrp_mem_policy_info_t *
3201 segspt_shmgetpolicy(struct seg *seg, caddr_t addr)
3202 {
3203 	struct anon_map		*amp;
3204 	ulong_t			anon_index;
3205 	lgrp_mem_policy_info_t	*policy_info;
3206 	struct shm_data		*shm_data;
3207 
3208 	ASSERT(seg != NULL);
3209 
3210 	/*
3211 	 * Get anon_map from segshm
3212 	 *
3213 	 * Assume that no lock needs to be held on anon_map, since
3214 	 * it should be protected by its reference count which must be
3215 	 * nonzero for an existing segment
3216 	 * Need to grab readers lock on policy tree though
3217 	 */
3218 	shm_data = (struct shm_data *)seg->s_data;
3219 	if (shm_data == NULL)
3220 		return (NULL);
3221 	amp = shm_data->shm_amp;
3222 	ASSERT(amp->refcnt != 0);
3223 
3224 	/*
3225 	 * Get policy info
3226 	 *
3227 	 * Assume starting anon index of 0
3228 	 */
3229 	anon_index = seg_page(seg, addr);
3230 	policy_info = lgrp_shm_policy_get(amp, anon_index, NULL, 0);
3231 
3232 	return (policy_info);
3233 }
3234 
3235 /*ARGSUSED*/
3236 static int
3237 segspt_shmcapable(struct seg *seg, segcapability_t capability)
3238 {
3239 	return (0);
3240 }
3241