xref: /illumos-gate/usr/src/uts/common/vm/page_lock.c (revision 8bc68872f6b178bf5e1d324c663e29fb6ccb1eab)
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 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 /*
29  * VM - page locking primitives
30  */
31 #include <sys/param.h>
32 #include <sys/t_lock.h>
33 #include <sys/vtrace.h>
34 #include <sys/debug.h>
35 #include <sys/cmn_err.h>
36 #include <sys/vnode.h>
37 #include <sys/bitmap.h>
38 #include <sys/lockstat.h>
39 #include <sys/condvar_impl.h>
40 #include <vm/page.h>
41 #include <vm/seg_enum.h>
42 #include <vm/vm_dep.h>
43 
44 /*
45  * This global mutex is for logical page locking.
46  * The following fields in the page structure are protected
47  * by this lock:
48  *
49  *	p_lckcnt
50  *	p_cowcnt
51  */
52 kmutex_t page_llock;
53 
54 /*
55  * This is a global lock for the logical page free list.  The
56  * logical free list, in this implementation, is maintained as two
57  * separate physical lists - the cache list and the free list.
58  */
59 kmutex_t  page_freelock;
60 
61 /*
62  * The hash table, page_hash[], the p_selock fields, and the
63  * list of pages associated with vnodes are protected by arrays of mutexes.
64  *
65  * Unless the hashes are changed radically, the table sizes must be
66  * a power of two.  Also, we typically need more mutexes for the
67  * vnodes since these locks are occasionally held for long periods.
68  * And since there seem to be two special vnodes (kvp and swapvp),
69  * we make room for private mutexes for them.
70  *
71  * The pse_mutex[] array holds the mutexes to protect the p_selock
72  * fields of all page_t structures.
73  *
74  * PAGE_SE_MUTEX(pp) returns the address of the appropriate mutex
75  * when given a pointer to a page_t.
76  *
77  * PSE_TABLE_SIZE must be a power of two.  One could argue that we
78  * should go to the trouble of setting it up at run time and base it
79  * on memory size rather than the number of compile time CPUs.
80  *
81  * XX64	We should be using physmem size to calculate PSE_TABLE_SIZE,
82  *	PSE_SHIFT, PIO_SHIFT.
83  *
84  *	These might break in 64 bit world.
85  */
86 #define	PSE_SHIFT	7		/* log2(PSE_TABLE_SIZE) */
87 
88 #define	PSE_TABLE_SIZE	128		/* number of mutexes to have */
89 
90 #define	PIO_SHIFT	PSE_SHIFT	/* next power of 2 bigger than page_t */
91 #define	PIO_TABLE_SIZE	PSE_TABLE_SIZE	/* number of io mutexes to have */
92 
93 pad_mutex_t	ph_mutex[PH_TABLE_SIZE];
94 pad_mutex_t	pse_mutex[PSE_TABLE_SIZE];
95 kmutex_t	pio_mutex[PIO_TABLE_SIZE];
96 
97 #define	PAGE_SE_MUTEX(pp) \
98 	    &pse_mutex[((((uintptr_t)(pp) >> PSE_SHIFT) ^ \
99 		((uintptr_t)(pp) >> (PSE_SHIFT << 1))) & \
100 		(PSE_TABLE_SIZE - 1))].pad_mutex
101 
102 #define	PAGE_IO_MUTEX(pp) \
103 	    &pio_mutex[(((uintptr_t)pp) >> PIO_SHIFT) & (PIO_TABLE_SIZE - 1)]
104 
105 #define	PSZC_MTX_TABLE_SIZE	128
106 #define	PSZC_MTX_TABLE_SHIFT	7
107 
108 static pad_mutex_t	pszc_mutex[PSZC_MTX_TABLE_SIZE];
109 
110 #define	PAGE_SZC_MUTEX(_pp) \
111 	    &pszc_mutex[((((uintptr_t)(_pp) >> PSZC_MTX_TABLE_SHIFT) ^ \
112 		((uintptr_t)(_pp) >> (PSZC_MTX_TABLE_SHIFT << 1)) ^ \
113 		((uintptr_t)(_pp) >> (3 * PSZC_MTX_TABLE_SHIFT))) & \
114 		(PSZC_MTX_TABLE_SIZE - 1))].pad_mutex
115 
116 /*
117  * The vph_mutex[] array  holds the mutexes to protect the vnode chains,
118  * (i.e., the list of pages anchored by v_pages and connected via p_vpprev
119  * and p_vpnext).
120  *
121  * The page_vnode_mutex(vp) function returns the address of the appropriate
122  * mutex from this array given a pointer to a vnode.  It is complicated
123  * by the fact that the kernel's vnode and the swapfs vnode are referenced
124  * frequently enough to warrent their own mutexes.
125  *
126  * The VP_HASH_FUNC returns the index into the vph_mutex array given
127  * an address of a vnode.
128  */
129 
130 /*
131  * XX64	VPH_TABLE_SIZE and VP_HASH_FUNC might break in 64 bit world.
132  *	Need to review again.
133  */
134 #define	VPH_TABLE_SIZE	(2 << VP_SHIFT)
135 
136 #define	VP_HASH_FUNC(vp) \
137 	((((uintptr_t)(vp) >> 6) + \
138 	    ((uintptr_t)(vp) >> 8) + \
139 	    ((uintptr_t)(vp) >> 10) + \
140 	    ((uintptr_t)(vp) >> 12)) \
141 	    & (VPH_TABLE_SIZE - 1))
142 
143 extern	struct vnode	kvp;
144 
145 kmutex_t	vph_mutex[VPH_TABLE_SIZE + 2];
146 
147 /*
148  * Initialize the locks used by the Virtual Memory Management system.
149  */
150 void
151 page_lock_init()
152 {
153 }
154 
155 /*
156  * At present we only use page ownership to aid debugging, so it's
157  * OK if the owner field isn't exact.  In the 32-bit world two thread ids
158  * can map to the same owner because we just 'or' in 0x80000000 and
159  * then clear the second highest bit, so that (for example) 0x2faced00
160  * and 0xafaced00 both map to 0xafaced00.
161  * In the 64-bit world, p_selock may not be large enough to hold a full
162  * thread pointer.  If we ever need precise ownership (e.g. if we implement
163  * priority inheritance for page locks) then p_selock should become a
164  * uintptr_t and SE_WRITER should be -((uintptr_t)curthread >> 2).
165  */
166 #define	SE_WRITER	(((selock_t)(ulong_t)curthread | INT_MIN) & ~SE_EWANTED)
167 #define	SE_READER	1
168 
169 /*
170  * A page that is deleted must be marked as such using the
171  * page_lock_delete() function. The page must be exclusively locked.
172  * The SE_DELETED marker is put in p_selock when this function is called.
173  * SE_DELETED must be distinct from any SE_WRITER value.
174  */
175 #define	SE_DELETED	(1 | INT_MIN)
176 
177 #ifdef VM_STATS
178 uint_t	vph_kvp_count;
179 uint_t	vph_swapfsvp_count;
180 uint_t	vph_other;
181 #endif /* VM_STATS */
182 
183 #ifdef VM_STATS
184 uint_t	page_lock_count;
185 uint_t	page_lock_miss;
186 uint_t	page_lock_miss_lock;
187 uint_t	page_lock_reclaim;
188 uint_t	page_lock_bad_reclaim;
189 uint_t	page_lock_same_page;
190 uint_t	page_lock_upgrade;
191 uint_t	page_lock_retired;
192 uint_t	page_lock_upgrade_failed;
193 uint_t	page_lock_deleted;
194 
195 uint_t	page_trylock_locked;
196 uint_t	page_trylock_failed;
197 uint_t	page_trylock_missed;
198 
199 uint_t	page_try_reclaim_upgrade;
200 #endif /* VM_STATS */
201 
202 /*
203  * Acquire the "shared/exclusive" lock on a page.
204  *
205  * Returns 1 on success and locks the page appropriately.
206  *	   0 on failure and does not lock the page.
207  *
208  * If `lock' is non-NULL, it will be dropped and reacquired in the
209  * failure case.  This routine can block, and if it does
210  * it will always return a failure since the page identity [vp, off]
211  * or state may have changed.
212  */
213 
214 int
215 page_lock(page_t *pp, se_t se, kmutex_t *lock, reclaim_t reclaim)
216 {
217 	return (page_lock_es(pp, se, lock, reclaim, 0));
218 }
219 
220 /*
221  * With the addition of reader-writer lock semantics to page_lock_es,
222  * callers wanting an exclusive (writer) lock may prevent shared-lock
223  * (reader) starvation by setting the es parameter to SE_EXCL_WANTED.
224  * In this case, when an exclusive lock cannot be acquired, p_selock's
225  * SE_EWANTED bit is set. Shared-lock (reader) requests are also denied
226  * if the page is slated for retirement.
227  *
228  * The se and es parameters determine if the lock should be granted
229  * based on the following decision table:
230  *
231  * Lock wanted   es flags     p_selock/SE_EWANTED  Action
232  * ----------- -------------- -------------------  ---------
233  * SE_EXCL        any [1][2]   unlocked/any        grant lock, clear SE_EWANTED
234  * SE_EXCL        SE_EWANTED   any lock/any        deny, set SE_EWANTED
235  * SE_EXCL        none         any lock/any        deny
236  * SE_SHARED      n/a [2]        shared/0          grant
237  * SE_SHARED      n/a [2]      unlocked/0          grant
238  * SE_SHARED      n/a            shared/1          deny
239  * SE_SHARED      n/a          unlocked/1          deny
240  * SE_SHARED      n/a              excl/any        deny
241  *
242  * Notes:
243  * [1] The code grants an exclusive lock to the caller and clears the bit
244  *   SE_EWANTED whenever p_selock is unlocked, regardless of the SE_EWANTED
245  *   bit's value.  This was deemed acceptable as we are not concerned about
246  *   exclusive-lock starvation. If this ever becomes an issue, a priority or
247  *   fifo mechanism should also be implemented. Meantime, the thread that
248  *   set SE_EWANTED should be prepared to catch this condition and reset it
249  *
250  * [2] Retired pages may not be locked at any time, regardless of the
251  *   dispostion of se, unless the es parameter has SE_RETIRED flag set.
252  *
253  * Notes on values of "es":
254  *
255  *   es & 1: page_lookup_create will attempt page relocation
256  *   es & SE_EXCL_WANTED: caller wants SE_EWANTED set (eg. delete
257  *       memory thread); this prevents reader-starvation of waiting
258  *       writer thread(s) by giving priority to writers over readers.
259  *   es & SE_RETIRED: caller wants to lock pages even if they are
260  *       retired.  Default is to deny the lock if the page is retired.
261  *
262  * And yes, we know, the semantics of this function are too complicated.
263  * It's on the list to be cleaned up.
264  */
265 int
266 page_lock_es(page_t *pp, se_t se, kmutex_t *lock, reclaim_t reclaim, int es)
267 {
268 	int		retval;
269 	kmutex_t	*pse = PAGE_SE_MUTEX(pp);
270 	int		upgraded;
271 	int		reclaim_it;
272 
273 	ASSERT(lock != NULL ? MUTEX_HELD(lock) : 1);
274 
275 	VM_STAT_ADD(page_lock_count);
276 
277 	upgraded = 0;
278 	reclaim_it = 0;
279 
280 	mutex_enter(pse);
281 
282 	ASSERT(((es & SE_EXCL_WANTED) == 0) ||
283 	    ((es & SE_EXCL_WANTED) && (se == SE_EXCL)));
284 
285 	if (PP_RETIRED(pp) && !(es & SE_RETIRED)) {
286 		mutex_exit(pse);
287 		VM_STAT_ADD(page_lock_retired);
288 		return (0);
289 	}
290 
291 	if (se == SE_SHARED && es == 1 && pp->p_selock == 0) {
292 		se = SE_EXCL;
293 	}
294 
295 	if ((reclaim == P_RECLAIM) && (PP_ISFREE(pp))) {
296 
297 		reclaim_it = 1;
298 		if (se == SE_SHARED) {
299 			/*
300 			 * This is an interesting situation.
301 			 *
302 			 * Remember that p_free can only change if
303 			 * p_selock < 0.
304 			 * p_free does not depend on our holding `pse'.
305 			 * And, since we hold `pse', p_selock can not change.
306 			 * So, if p_free changes on us, the page is already
307 			 * exclusively held, and we would fail to get p_selock
308 			 * regardless.
309 			 *
310 			 * We want to avoid getting the share
311 			 * lock on a free page that needs to be reclaimed.
312 			 * It is possible that some other thread has the share
313 			 * lock and has left the free page on the cache list.
314 			 * pvn_vplist_dirty() does this for brief periods.
315 			 * If the se_share is currently SE_EXCL, we will fail
316 			 * to acquire p_selock anyway.  Blocking is the
317 			 * right thing to do.
318 			 * If we need to reclaim this page, we must get
319 			 * exclusive access to it, force the upgrade now.
320 			 * Again, we will fail to acquire p_selock if the
321 			 * page is not free and block.
322 			 */
323 			upgraded = 1;
324 			se = SE_EXCL;
325 			VM_STAT_ADD(page_lock_upgrade);
326 		}
327 	}
328 
329 	if (se == SE_EXCL) {
330 		if (!(es & SE_EXCL_WANTED) && (pp->p_selock & SE_EWANTED)) {
331 			/*
332 			 * if the caller wants a writer lock (but did not
333 			 * specify exclusive access), and there is a pending
334 			 * writer that wants exclusive access, return failure
335 			 */
336 			retval = 0;
337 		} else if ((pp->p_selock & ~SE_EWANTED) == 0) {
338 			/* no reader/writer lock held */
339 			THREAD_KPRI_REQUEST();
340 			/* this clears our setting of the SE_EWANTED bit */
341 			pp->p_selock = SE_WRITER;
342 			retval = 1;
343 		} else {
344 			/* page is locked */
345 			if (es & SE_EXCL_WANTED) {
346 				/* set the SE_EWANTED bit */
347 				pp->p_selock |= SE_EWANTED;
348 			}
349 			retval = 0;
350 		}
351 	} else {
352 		retval = 0;
353 		if (pp->p_selock >= 0) {
354 			if ((pp->p_selock & SE_EWANTED) == 0) {
355 				pp->p_selock += SE_READER;
356 				retval = 1;
357 			}
358 		}
359 	}
360 
361 	if (retval == 0) {
362 		if ((pp->p_selock & ~SE_EWANTED) == SE_DELETED) {
363 			VM_STAT_ADD(page_lock_deleted);
364 			mutex_exit(pse);
365 			return (retval);
366 		}
367 
368 #ifdef VM_STATS
369 		VM_STAT_ADD(page_lock_miss);
370 		if (upgraded) {
371 			VM_STAT_ADD(page_lock_upgrade_failed);
372 		}
373 #endif
374 		if (lock) {
375 			VM_STAT_ADD(page_lock_miss_lock);
376 			mutex_exit(lock);
377 		}
378 
379 		/*
380 		 * Now, wait for the page to be unlocked and
381 		 * release the lock protecting p_cv and p_selock.
382 		 */
383 		cv_wait(&pp->p_cv, pse);
384 		mutex_exit(pse);
385 
386 		/*
387 		 * The page identity may have changed while we were
388 		 * blocked.  If we are willing to depend on "pp"
389 		 * still pointing to a valid page structure (i.e.,
390 		 * assuming page structures are not dynamically allocated
391 		 * or freed), we could try to lock the page if its
392 		 * identity hasn't changed.
393 		 *
394 		 * This needs to be measured, since we come back from
395 		 * cv_wait holding pse (the expensive part of this
396 		 * operation) we might as well try the cheap part.
397 		 * Though we would also have to confirm that dropping
398 		 * `lock' did not cause any grief to the callers.
399 		 */
400 		if (lock) {
401 			mutex_enter(lock);
402 		}
403 	} else {
404 		/*
405 		 * We have the page lock.
406 		 * If we needed to reclaim the page, and the page
407 		 * needed reclaiming (ie, it was free), then we
408 		 * have the page exclusively locked.  We may need
409 		 * to downgrade the page.
410 		 */
411 		ASSERT((upgraded) ?
412 		    ((PP_ISFREE(pp)) && PAGE_EXCL(pp)) : 1);
413 		mutex_exit(pse);
414 
415 		/*
416 		 * We now hold this page's lock, either shared or
417 		 * exclusive.  This will prevent its identity from changing.
418 		 * The page, however, may or may not be free.  If the caller
419 		 * requested, and it is free, go reclaim it from the
420 		 * free list.  If the page can't be reclaimed, return failure
421 		 * so that the caller can start all over again.
422 		 *
423 		 * NOTE:page_reclaim() releases the page lock (p_selock)
424 		 *	if it can't be reclaimed.
425 		 */
426 		if (reclaim_it) {
427 			if (!page_reclaim(pp, lock)) {
428 				VM_STAT_ADD(page_lock_bad_reclaim);
429 				retval = 0;
430 			} else {
431 				VM_STAT_ADD(page_lock_reclaim);
432 				if (upgraded) {
433 					page_downgrade(pp);
434 				}
435 			}
436 		}
437 	}
438 	return (retval);
439 }
440 
441 /*
442  * Clear the SE_EWANTED bit from p_selock.  This function allows
443  * callers of page_lock_es and page_try_reclaim_lock to clear
444  * their setting of this bit if they decide they no longer wish
445  * to gain exclusive access to the page.  Currently only
446  * delete_memory_thread uses this when the delete memory
447  * operation is cancelled.
448  */
449 void
450 page_lock_clr_exclwanted(page_t *pp)
451 {
452 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
453 
454 	mutex_enter(pse);
455 	pp->p_selock &= ~SE_EWANTED;
456 	if (CV_HAS_WAITERS(&pp->p_cv))
457 		cv_broadcast(&pp->p_cv);
458 	mutex_exit(pse);
459 }
460 
461 /*
462  * Read the comments inside of page_lock_es() carefully.
463  *
464  * SE_EXCL callers specifying es == SE_EXCL_WANTED will cause the
465  * SE_EWANTED bit of p_selock to be set when the lock cannot be obtained.
466  * This is used by threads subject to reader-starvation (eg. memory delete).
467  *
468  * When a thread using SE_EXCL_WANTED does not obtain the SE_EXCL lock,
469  * it is expected that it will retry at a later time.  Threads that will
470  * not retry the lock *must* call page_lock_clr_exclwanted to clear the
471  * SE_EWANTED bit.  (When a thread using SE_EXCL_WANTED obtains the lock,
472  * the bit is cleared.)
473  */
474 int
475 page_try_reclaim_lock(page_t *pp, se_t se, int es)
476 {
477 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
478 	selock_t old;
479 
480 	mutex_enter(pse);
481 
482 	old = pp->p_selock;
483 
484 	ASSERT(((es & SE_EXCL_WANTED) == 0) ||
485 	    ((es & SE_EXCL_WANTED) && (se == SE_EXCL)));
486 
487 	if (PP_RETIRED(pp) && !(es & SE_RETIRED)) {
488 		mutex_exit(pse);
489 		VM_STAT_ADD(page_trylock_failed);
490 		return (0);
491 	}
492 
493 	if (se == SE_SHARED && es == 1 && old == 0) {
494 		se = SE_EXCL;
495 	}
496 
497 	if (se == SE_SHARED) {
498 		if (!PP_ISFREE(pp)) {
499 			if (old >= 0) {
500 				/*
501 				 * Readers are not allowed when excl wanted
502 				 */
503 				if ((old & SE_EWANTED) == 0) {
504 					pp->p_selock = old + SE_READER;
505 					mutex_exit(pse);
506 					return (1);
507 				}
508 			}
509 			mutex_exit(pse);
510 			return (0);
511 		}
512 		/*
513 		 * The page is free, so we really want SE_EXCL (below)
514 		 */
515 		VM_STAT_ADD(page_try_reclaim_upgrade);
516 	}
517 
518 	/*
519 	 * The caller wants a writer lock.  We try for it only if
520 	 * SE_EWANTED is not set, or if the caller specified
521 	 * SE_EXCL_WANTED.
522 	 */
523 	if (!(old & SE_EWANTED) || (es & SE_EXCL_WANTED)) {
524 		if ((old & ~SE_EWANTED) == 0) {
525 			/* no reader/writer lock held */
526 			THREAD_KPRI_REQUEST();
527 			/* this clears out our setting of the SE_EWANTED bit */
528 			pp->p_selock = SE_WRITER;
529 			mutex_exit(pse);
530 			return (1);
531 		}
532 	}
533 	if (es & SE_EXCL_WANTED) {
534 		/* page is locked, set the SE_EWANTED bit */
535 		pp->p_selock |= SE_EWANTED;
536 	}
537 	mutex_exit(pse);
538 	return (0);
539 }
540 
541 /*
542  * Acquire a page's "shared/exclusive" lock, but never block.
543  * Returns 1 on success, 0 on failure.
544  */
545 int
546 page_trylock(page_t *pp, se_t se)
547 {
548 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
549 
550 	mutex_enter(pse);
551 	if (pp->p_selock & SE_EWANTED || PP_RETIRED(pp) ||
552 	    (se == SE_SHARED && PP_PR_NOSHARE(pp))) {
553 		/*
554 		 * Fail if a thread wants exclusive access and page is
555 		 * retired, if the page is slated for retirement, or a
556 		 * share lock is requested.
557 		 */
558 		mutex_exit(pse);
559 		VM_STAT_ADD(page_trylock_failed);
560 		return (0);
561 	}
562 
563 	if (se == SE_EXCL) {
564 		if (pp->p_selock == 0) {
565 			THREAD_KPRI_REQUEST();
566 			pp->p_selock = SE_WRITER;
567 			mutex_exit(pse);
568 			return (1);
569 		}
570 	} else {
571 		if (pp->p_selock >= 0) {
572 			pp->p_selock += SE_READER;
573 			mutex_exit(pse);
574 			return (1);
575 		}
576 	}
577 	mutex_exit(pse);
578 	return (0);
579 }
580 
581 /*
582  * Variant of page_unlock() specifically for the page freelist
583  * code. The mere existence of this code is a vile hack that
584  * has resulted due to the backwards locking order of the page
585  * freelist manager; please don't call it.
586  */
587 void
588 page_unlock_noretire(page_t *pp)
589 {
590 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
591 	selock_t old;
592 
593 	mutex_enter(pse);
594 
595 	old = pp->p_selock;
596 	if ((old & ~SE_EWANTED) == SE_READER) {
597 		pp->p_selock = old & ~SE_READER;
598 		if (CV_HAS_WAITERS(&pp->p_cv))
599 			cv_broadcast(&pp->p_cv);
600 	} else if ((old & ~SE_EWANTED) == SE_DELETED) {
601 		panic("page_unlock_noretire: page %p is deleted", pp);
602 	} else if (old < 0) {
603 		THREAD_KPRI_RELEASE();
604 		pp->p_selock &= SE_EWANTED;
605 		if (CV_HAS_WAITERS(&pp->p_cv))
606 			cv_broadcast(&pp->p_cv);
607 	} else if ((old & ~SE_EWANTED) > SE_READER) {
608 		pp->p_selock = old - SE_READER;
609 	} else {
610 		panic("page_unlock_noretire: page %p is not locked", pp);
611 	}
612 
613 	mutex_exit(pse);
614 }
615 
616 /*
617  * Release the page's "shared/exclusive" lock and wake up anyone
618  * who might be waiting for it.
619  */
620 void
621 page_unlock(page_t *pp)
622 {
623 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
624 	selock_t old;
625 
626 	mutex_enter(pse);
627 
628 	old = pp->p_selock;
629 	if ((old & ~SE_EWANTED) == SE_READER) {
630 		pp->p_selock = old & ~SE_READER;
631 		if (CV_HAS_WAITERS(&pp->p_cv))
632 			cv_broadcast(&pp->p_cv);
633 	} else if ((old & ~SE_EWANTED) == SE_DELETED) {
634 		panic("page_unlock: page %p is deleted", pp);
635 	} else if (old < 0) {
636 		THREAD_KPRI_RELEASE();
637 		pp->p_selock &= SE_EWANTED;
638 		if (CV_HAS_WAITERS(&pp->p_cv))
639 			cv_broadcast(&pp->p_cv);
640 	} else if ((old & ~SE_EWANTED) > SE_READER) {
641 		pp->p_selock = old - SE_READER;
642 	} else {
643 		panic("page_unlock: page %p is not locked", pp);
644 	}
645 
646 	if (pp->p_selock == 0 && PP_PR_REQ(pp)) {
647 		/*
648 		 * Try to retire the page. If it retires, great.
649 		 * If not, oh well, we'll get it in the next unlock
650 		 * request, and repeat the cycle.  Regardless,
651 		 * page_tryretire() will drop the page lock.
652 		 */
653 		if ((pp->p_toxic & PR_BUSY) == 0) {
654 			THREAD_KPRI_REQUEST();
655 			pp->p_selock = SE_WRITER;
656 			page_settoxic(pp, PR_BUSY);
657 			mutex_exit(pse);
658 			page_tryretire(pp);
659 		} else {
660 			pp->p_selock = SE_WRITER;
661 			page_clrtoxic(pp, PR_BUSY);
662 			pp->p_selock = 0;
663 			mutex_exit(pse);
664 		}
665 	} else {
666 		mutex_exit(pse);
667 	}
668 }
669 
670 /*
671  * Try to upgrade the lock on the page from a "shared" to an
672  * "exclusive" lock.  Since this upgrade operation is done while
673  * holding the mutex protecting this page, no one else can acquire this page's
674  * lock and change the page. Thus, it is safe to drop the "shared"
675  * lock and attempt to acquire the "exclusive" lock.
676  *
677  * Returns 1 on success, 0 on failure.
678  */
679 int
680 page_tryupgrade(page_t *pp)
681 {
682 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
683 
684 	mutex_enter(pse);
685 	if (!(pp->p_selock & SE_EWANTED)) {
686 		/* no threads want exclusive access, try upgrade */
687 		if (pp->p_selock == SE_READER) {
688 			THREAD_KPRI_REQUEST();
689 			/* convert to exclusive lock */
690 			pp->p_selock = SE_WRITER;
691 			mutex_exit(pse);
692 			return (1);
693 		}
694 	}
695 	mutex_exit(pse);
696 	return (0);
697 }
698 
699 /*
700  * Downgrade the "exclusive" lock on the page to a "shared" lock
701  * while holding the mutex protecting this page's p_selock field.
702  */
703 void
704 page_downgrade(page_t *pp)
705 {
706 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
707 	int excl_waiting;
708 
709 	ASSERT((pp->p_selock & ~SE_EWANTED) != SE_DELETED);
710 	ASSERT(PAGE_EXCL(pp));
711 
712 	mutex_enter(pse);
713 	excl_waiting =  pp->p_selock & SE_EWANTED;
714 	THREAD_KPRI_RELEASE();
715 	pp->p_selock = SE_READER | excl_waiting;
716 	if (CV_HAS_WAITERS(&pp->p_cv))
717 		cv_broadcast(&pp->p_cv);
718 	mutex_exit(pse);
719 }
720 
721 void
722 page_lock_delete(page_t *pp)
723 {
724 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
725 
726 	ASSERT(PAGE_EXCL(pp));
727 	ASSERT(pp->p_vnode == NULL);
728 	ASSERT(pp->p_offset == (u_offset_t)-1);
729 	ASSERT(!PP_ISFREE(pp));
730 
731 	mutex_enter(pse);
732 	THREAD_KPRI_RELEASE();
733 	pp->p_selock = SE_DELETED;
734 	if (CV_HAS_WAITERS(&pp->p_cv))
735 		cv_broadcast(&pp->p_cv);
736 	mutex_exit(pse);
737 }
738 
739 /*
740  * Implement the io lock for pages
741  */
742 void
743 page_iolock_init(page_t *pp)
744 {
745 	pp->p_iolock_state = 0;
746 	cv_init(&pp->p_io_cv, NULL, CV_DEFAULT, NULL);
747 }
748 
749 /*
750  * Acquire the i/o lock on a page.
751  */
752 void
753 page_io_lock(page_t *pp)
754 {
755 	kmutex_t *pio;
756 
757 	pio = PAGE_IO_MUTEX(pp);
758 	mutex_enter(pio);
759 	while (pp->p_iolock_state & PAGE_IO_INUSE) {
760 		cv_wait(&(pp->p_io_cv), pio);
761 	}
762 	pp->p_iolock_state |= PAGE_IO_INUSE;
763 	mutex_exit(pio);
764 }
765 
766 /*
767  * Release the i/o lock on a page.
768  */
769 void
770 page_io_unlock(page_t *pp)
771 {
772 	kmutex_t *pio;
773 
774 	pio = PAGE_IO_MUTEX(pp);
775 	mutex_enter(pio);
776 	cv_signal(&pp->p_io_cv);
777 	pp->p_iolock_state &= ~PAGE_IO_INUSE;
778 	mutex_exit(pio);
779 }
780 
781 /*
782  * Try to acquire the i/o lock on a page without blocking.
783  * Returns 1 on success, 0 on failure.
784  */
785 int
786 page_io_trylock(page_t *pp)
787 {
788 	kmutex_t *pio;
789 
790 	if (pp->p_iolock_state & PAGE_IO_INUSE)
791 		return (0);
792 
793 	pio = PAGE_IO_MUTEX(pp);
794 	mutex_enter(pio);
795 
796 	if (pp->p_iolock_state & PAGE_IO_INUSE) {
797 		mutex_exit(pio);
798 		return (0);
799 	}
800 	pp->p_iolock_state |= PAGE_IO_INUSE;
801 	mutex_exit(pio);
802 
803 	return (1);
804 }
805 
806 /*
807  * Assert that the i/o lock on a page is held.
808  * Returns 1 on success, 0 on failure.
809  */
810 int
811 page_iolock_assert(page_t *pp)
812 {
813 	return (pp->p_iolock_state & PAGE_IO_INUSE);
814 }
815 
816 /*
817  * Wrapper exported to kernel routines that are built
818  * platform-independent (the macro is platform-dependent;
819  * the size of vph_mutex[] is based on NCPU).
820  *
821  * Note that you can do stress testing on this by setting the
822  * variable page_vnode_mutex_stress to something other than
823  * zero in a DEBUG kernel in a debugger after loading the kernel.
824  * Setting it after the kernel is running may not work correctly.
825  */
826 #ifdef DEBUG
827 static int page_vnode_mutex_stress = 0;
828 #endif
829 
830 kmutex_t *
831 page_vnode_mutex(vnode_t *vp)
832 {
833 	if (vp == &kvp)
834 		return (&vph_mutex[VPH_TABLE_SIZE + 0]);
835 #ifdef DEBUG
836 	if (page_vnode_mutex_stress != 0)
837 		return (&vph_mutex[0]);
838 #endif
839 
840 	return (&vph_mutex[VP_HASH_FUNC(vp)]);
841 }
842 
843 kmutex_t *
844 page_se_mutex(page_t *pp)
845 {
846 	return (PAGE_SE_MUTEX(pp));
847 }
848 
849 #ifdef VM_STATS
850 uint_t pszclck_stat[4];
851 #endif
852 /*
853  * Find, take and return a mutex held by hat_page_demote().
854  * Called by page_demote_vp_pages() before hat_page_demote() call and by
855  * routines that want to block hat_page_demote() but can't do it
856  * via locking all constituent pages.
857  *
858  * Return NULL if p_szc is 0.
859  *
860  * It should only be used for pages that can be demoted by hat_page_demote()
861  * i.e. non swapfs file system pages.  The logic here is lifted from
862  * sfmmu_mlspl_enter() except there's no need to worry about p_szc increase
863  * since the page is locked and not free.
864  *
865  * Hash of the root page is used to find the lock.
866  * To find the root in the presense of hat_page_demote() chageing the location
867  * of the root this routine relies on the fact that hat_page_demote() changes
868  * root last.
869  *
870  * If NULL is returned pp's p_szc is guaranteed to be 0. If non NULL is
871  * returned pp's p_szc may be any value.
872  */
873 kmutex_t *
874 page_szc_lock(page_t *pp)
875 {
876 	kmutex_t	*mtx;
877 	page_t		*rootpp;
878 	uint_t		szc;
879 	uint_t		rszc;
880 	uint_t		pszc = pp->p_szc;
881 
882 	ASSERT(pp != NULL);
883 	ASSERT(PAGE_LOCKED(pp));
884 	ASSERT(!PP_ISFREE(pp));
885 	ASSERT(pp->p_vnode != NULL);
886 	ASSERT(!IS_SWAPFSVP(pp->p_vnode));
887 	ASSERT(pp->p_vnode != &kvp);
888 
889 again:
890 	if (pszc == 0) {
891 		VM_STAT_ADD(pszclck_stat[0]);
892 		return (NULL);
893 	}
894 
895 	/* The lock lives in the root page */
896 
897 	rootpp = PP_GROUPLEADER(pp, pszc);
898 	mtx = PAGE_SZC_MUTEX(rootpp);
899 	mutex_enter(mtx);
900 
901 	/*
902 	 * since p_szc can only decrease if pp == rootpp
903 	 * rootpp will be always the same i.e we have the right root
904 	 * regardless of rootpp->p_szc.
905 	 * If location of pp's root didn't change after we took
906 	 * the lock we have the right root. return mutex hashed off it.
907 	 */
908 	if (pp == rootpp || (rszc = rootpp->p_szc) == pszc) {
909 		VM_STAT_ADD(pszclck_stat[1]);
910 		return (mtx);
911 	}
912 
913 	/*
914 	 * root location changed because page got demoted.
915 	 * locate the new root.
916 	 */
917 	if (rszc < pszc) {
918 		szc = pp->p_szc;
919 		ASSERT(szc < pszc);
920 		mutex_exit(mtx);
921 		pszc = szc;
922 		VM_STAT_ADD(pszclck_stat[2]);
923 		goto again;
924 	}
925 
926 	VM_STAT_ADD(pszclck_stat[3]);
927 	/*
928 	 * current hat_page_demote not done yet.
929 	 * wait for it to finish.
930 	 */
931 	mutex_exit(mtx);
932 	rootpp = PP_GROUPLEADER(rootpp, rszc);
933 	mtx = PAGE_SZC_MUTEX(rootpp);
934 	mutex_enter(mtx);
935 	mutex_exit(mtx);
936 	ASSERT(rootpp->p_szc < rszc);
937 	goto again;
938 }
939 
940 int
941 page_szc_lock_assert(page_t *pp)
942 {
943 	page_t *rootpp = PP_PAGEROOT(pp);
944 	kmutex_t *mtx = PAGE_SZC_MUTEX(rootpp);
945 
946 	return (MUTEX_HELD(mtx));
947 }
948