xref: /titanic_41/usr/src/uts/common/vm/page_lock.c (revision 5aefb6555731130ca4fd295960123d71f2d21fe8)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 /*
30  * VM - page locking primitives
31  */
32 #include <sys/param.h>
33 #include <sys/t_lock.h>
34 #include <sys/vtrace.h>
35 #include <sys/debug.h>
36 #include <sys/cmn_err.h>
37 #include <sys/vnode.h>
38 #include <sys/bitmap.h>
39 #include <sys/lockstat.h>
40 #include <sys/condvar_impl.h>
41 #include <vm/page.h>
42 #include <vm/seg_enum.h>
43 #include <vm/vm_dep.h>
44 
45 /*
46  * This global mutex is for logical page locking.
47  * The following fields in the page structure are protected
48  * by this lock:
49  *
50  *	p_lckcnt
51  *	p_cowcnt
52  */
53 kmutex_t page_llock;
54 
55 /*
56  * This is a global lock for the logical page free list.  The
57  * logical free list, in this implementation, is maintained as two
58  * separate physical lists - the cache list and the free list.
59  */
60 kmutex_t  page_freelock;
61 
62 /*
63  * The hash table, page_hash[], the p_selock fields, and the
64  * list of pages associated with vnodes are protected by arrays of mutexes.
65  *
66  * Unless the hashes are changed radically, the table sizes must be
67  * a power of two.  Also, we typically need more mutexes for the
68  * vnodes since these locks are occasionally held for long periods.
69  * And since there seem to be two special vnodes (kvp and swapvp),
70  * we make room for private mutexes for them.
71  *
72  * The pse_mutex[] array holds the mutexes to protect the p_selock
73  * fields of all page_t structures.
74  *
75  * PAGE_SE_MUTEX(pp) returns the address of the appropriate mutex
76  * when given a pointer to a page_t.
77  *
78  * PSE_TABLE_SIZE must be a power of two.  One could argue that we
79  * should go to the trouble of setting it up at run time and base it
80  * on memory size rather than the number of compile time CPUs.
81  *
82  * XX64	We should be using physmem size to calculate PSE_TABLE_SIZE,
83  *	PSE_SHIFT, PIO_SHIFT.
84  *
85  *	These might break in 64 bit world.
86  */
87 #define	PSE_SHIFT	7		/* log2(PSE_TABLE_SIZE) */
88 
89 #define	PSE_TABLE_SIZE	128		/* number of mutexes to have */
90 
91 #define	PIO_SHIFT	PSE_SHIFT	/* next power of 2 bigger than page_t */
92 #define	PIO_TABLE_SIZE	PSE_TABLE_SIZE	/* number of io mutexes to have */
93 
94 pad_mutex_t	ph_mutex[PH_TABLE_SIZE];
95 pad_mutex_t	pse_mutex[PSE_TABLE_SIZE];
96 kmutex_t	pio_mutex[PIO_TABLE_SIZE];
97 
98 #define	PAGE_SE_MUTEX(pp) \
99 	    &pse_mutex[((((uintptr_t)(pp) >> PSE_SHIFT) ^ \
100 		((uintptr_t)(pp) >> (PSE_SHIFT << 1))) & \
101 		(PSE_TABLE_SIZE - 1))].pad_mutex
102 
103 #define	PAGE_IO_MUTEX(pp) \
104 	    &pio_mutex[(((uintptr_t)pp) >> PIO_SHIFT) & (PIO_TABLE_SIZE - 1)]
105 
106 #define	PSZC_MTX_TABLE_SIZE	128
107 #define	PSZC_MTX_TABLE_SHIFT	7
108 
109 static pad_mutex_t	pszc_mutex[PSZC_MTX_TABLE_SIZE];
110 
111 #define	PAGE_SZC_MUTEX(_pp) \
112 	    &pszc_mutex[((((uintptr_t)(_pp) >> PSZC_MTX_TABLE_SHIFT) ^ \
113 		((uintptr_t)(_pp) >> (PSZC_MTX_TABLE_SHIFT << 1)) ^ \
114 		((uintptr_t)(_pp) >> (3 * PSZC_MTX_TABLE_SHIFT))) & \
115 		(PSZC_MTX_TABLE_SIZE - 1))].pad_mutex
116 
117 /*
118  * The vph_mutex[] array  holds the mutexes to protect the vnode chains,
119  * (i.e., the list of pages anchored by v_pages and connected via p_vpprev
120  * and p_vpnext).
121  *
122  * The page_vnode_mutex(vp) function returns the address of the appropriate
123  * mutex from this array given a pointer to a vnode.  It is complicated
124  * by the fact that the kernel's vnode and the swapfs vnode are referenced
125  * frequently enough to warrent their own mutexes.
126  *
127  * The VP_HASH_FUNC returns the index into the vph_mutex array given
128  * an address of a vnode.
129  */
130 
131 /*
132  * XX64	VPH_TABLE_SIZE and VP_HASH_FUNC might break in 64 bit world.
133  *	Need to review again.
134  */
135 #define	VPH_TABLE_SIZE	(2 << VP_SHIFT)
136 
137 #define	VP_HASH_FUNC(vp) \
138 	((((uintptr_t)(vp) >> 6) + \
139 	    ((uintptr_t)(vp) >> 8) + \
140 	    ((uintptr_t)(vp) >> 10) + \
141 	    ((uintptr_t)(vp) >> 12)) \
142 	    & (VPH_TABLE_SIZE - 1))
143 
144 extern	struct vnode	kvp;
145 
146 kmutex_t	vph_mutex[VPH_TABLE_SIZE + 2];
147 
148 /*
149  * Initialize the locks used by the Virtual Memory Management system.
150  */
151 void
152 page_lock_init()
153 {
154 }
155 
156 /*
157  * At present we only use page ownership to aid debugging, so it's
158  * OK if the owner field isn't exact.  In the 32-bit world two thread ids
159  * can map to the same owner because we just 'or' in 0x80000000 and
160  * then clear the second highest bit, so that (for example) 0x2faced00
161  * and 0xafaced00 both map to 0xafaced00.
162  * In the 64-bit world, p_selock may not be large enough to hold a full
163  * thread pointer.  If we ever need precise ownership (e.g. if we implement
164  * priority inheritance for page locks) then p_selock should become a
165  * uintptr_t and SE_WRITER should be -((uintptr_t)curthread >> 2).
166  */
167 #define	SE_WRITER	(((selock_t)(ulong_t)curthread | INT_MIN) & ~SE_EWANTED)
168 #define	SE_READER	1
169 
170 /*
171  * A page that is deleted must be marked as such using the
172  * page_lock_delete() function. The page must be exclusively locked.
173  * The SE_DELETED marker is put in p_selock when this function is called.
174  * SE_DELETED must be distinct from any SE_WRITER value.
175  */
176 #define	SE_DELETED	(1 | INT_MIN)
177 
178 #ifdef VM_STATS
179 uint_t	vph_kvp_count;
180 uint_t	vph_swapfsvp_count;
181 uint_t	vph_other;
182 #endif /* VM_STATS */
183 
184 #ifdef VM_STATS
185 uint_t	page_lock_count;
186 uint_t	page_lock_miss;
187 uint_t	page_lock_miss_lock;
188 uint_t	page_lock_reclaim;
189 uint_t	page_lock_bad_reclaim;
190 uint_t	page_lock_same_page;
191 uint_t	page_lock_upgrade;
192 uint_t	page_lock_retired;
193 uint_t	page_lock_upgrade_failed;
194 uint_t	page_lock_deleted;
195 
196 uint_t	page_trylock_locked;
197 uint_t	page_trylock_failed;
198 uint_t	page_trylock_missed;
199 
200 uint_t	page_try_reclaim_upgrade;
201 #endif /* VM_STATS */
202 
203 /*
204  * Acquire the "shared/exclusive" lock on a page.
205  *
206  * Returns 1 on success and locks the page appropriately.
207  *	   0 on failure and does not lock the page.
208  *
209  * If `lock' is non-NULL, it will be dropped and reacquired in the
210  * failure case.  This routine can block, and if it does
211  * it will always return a failure since the page identity [vp, off]
212  * or state may have changed.
213  */
214 
215 int
216 page_lock(page_t *pp, se_t se, kmutex_t *lock, reclaim_t reclaim)
217 {
218 	return (page_lock_es(pp, se, lock, reclaim, 0));
219 }
220 
221 /*
222  * With the addition of reader-writer lock semantics to page_lock_es,
223  * callers wanting an exclusive (writer) lock may prevent shared-lock
224  * (reader) starvation by setting the es parameter to SE_EXCL_WANTED.
225  * In this case, when an exclusive lock cannot be acquired, p_selock's
226  * SE_EWANTED bit is set. Shared-lock (reader) requests are also denied
227  * if the page is slated for retirement.
228  *
229  * The se and es parameters determine if the lock should be granted
230  * based on the following decision table:
231  *
232  * Lock wanted   es flags     p_selock/SE_EWANTED  Action
233  * ----------- -------------- -------------------  ---------
234  * SE_EXCL        any [1][2]   unlocked/any        grant lock, clear SE_EWANTED
235  * SE_EXCL        SE_EWANTED   any lock/any        deny, set SE_EWANTED
236  * SE_EXCL        none         any lock/any        deny
237  * SE_SHARED      n/a [2][3]     shared/0          grant
238  * SE_SHARED      n/a [2][3]   unlocked/0          grant
239  * SE_SHARED      n/a            shared/1          deny
240  * SE_SHARED      n/a          unlocked/1          deny
241  * SE_SHARED      n/a              excl/any        deny
242  *
243  * Notes:
244  * [1] The code grants an exclusive lock to the caller and clears the bit
245  *   SE_EWANTED whenever p_selock is unlocked, regardless of the SE_EWANTED
246  *   bit's value.  This was deemed acceptable as we are not concerned about
247  *   exclusive-lock starvation. If this ever becomes an issue, a priority or
248  *   fifo mechanism should also be implemented. Meantime, the thread that
249  *   set SE_EWANTED should be prepared to catch this condition and reset it
250  *
251  * [2] Retired pages may not be locked at any time, regardless of the
252  *   dispostion of se, unless the es parameter has SE_RETIRED flag set.
253  *
254  * [3] If the page is slated for retirement by an agent, the lock is denied.
255  *
256  * Notes on values of "es":
257  *
258  *   es & 1: page_lookup_create will attempt page relocation
259  *   es & SE_EXCL_WANTED: caller wants SE_EWANTED set (eg. delete
260  *       memory thread); this prevents reader-starvation of waiting
261  *       writer thread(s) by giving priority to writers over readers.
262  *   es & SE_RETIRED: caller wants to lock pages even if they are
263  *       retired.  Default is to deny the lock if the page is retired.
264  *
265  * And yes, we know, the semantics of this function are too complicated.
266  * It's on the list to be cleaned up.
267  */
268 int
269 page_lock_es(page_t *pp, se_t se, kmutex_t *lock, reclaim_t reclaim, int es)
270 {
271 	int		retval;
272 	kmutex_t	*pse = PAGE_SE_MUTEX(pp);
273 	int		upgraded;
274 	int		reclaim_it;
275 
276 	ASSERT(lock != NULL ? MUTEX_HELD(lock) : 1);
277 
278 	VM_STAT_ADD(page_lock_count);
279 
280 	upgraded = 0;
281 	reclaim_it = 0;
282 
283 	mutex_enter(pse);
284 
285 	ASSERT(((es & SE_EXCL_WANTED) == 0) ||
286 	    ((es & SE_EXCL_WANTED) && (se == SE_EXCL)));
287 
288 	if (PP_RETIRED(pp) && !(es & SE_RETIRED)) {
289 		mutex_exit(pse);
290 		VM_STAT_ADD(page_lock_retired);
291 		return (0);
292 	}
293 
294 	if (se == SE_SHARED && es == 1 && pp->p_selock == 0) {
295 		se = SE_EXCL;
296 	}
297 
298 	if ((reclaim == P_RECLAIM) && (PP_ISFREE(pp))) {
299 
300 		reclaim_it = 1;
301 		if (se == SE_SHARED) {
302 			/*
303 			 * This is an interesting situation.
304 			 *
305 			 * Remember that p_free can only change if
306 			 * p_selock < 0.
307 			 * p_free does not depend on our holding `pse'.
308 			 * And, since we hold `pse', p_selock can not change.
309 			 * So, if p_free changes on us, the page is already
310 			 * exclusively held, and we would fail to get p_selock
311 			 * regardless.
312 			 *
313 			 * We want to avoid getting the share
314 			 * lock on a free page that needs to be reclaimed.
315 			 * It is possible that some other thread has the share
316 			 * lock and has left the free page on the cache list.
317 			 * pvn_vplist_dirty() does this for brief periods.
318 			 * If the se_share is currently SE_EXCL, we will fail
319 			 * to acquire p_selock anyway.  Blocking is the
320 			 * right thing to do.
321 			 * If we need to reclaim this page, we must get
322 			 * exclusive access to it, force the upgrade now.
323 			 * Again, we will fail to acquire p_selock if the
324 			 * page is not free and block.
325 			 */
326 			upgraded = 1;
327 			se = SE_EXCL;
328 			VM_STAT_ADD(page_lock_upgrade);
329 		}
330 	}
331 
332 	if (se == SE_EXCL) {
333 		if (!(es & SE_EXCL_WANTED) && (pp->p_selock & SE_EWANTED)) {
334 			/*
335 			 * if the caller wants a writer lock (but did not
336 			 * specify exclusive access), and there is a pending
337 			 * writer that wants exclusive access, return failure
338 			 */
339 			retval = 0;
340 		} else if ((pp->p_selock & ~SE_EWANTED) == 0) {
341 			/* no reader/writer lock held */
342 			THREAD_KPRI_REQUEST();
343 			/* this clears our setting of the SE_EWANTED bit */
344 			pp->p_selock = SE_WRITER;
345 			retval = 1;
346 		} else {
347 			/* page is locked */
348 			if (es & SE_EXCL_WANTED) {
349 				/* set the SE_EWANTED bit */
350 				pp->p_selock |= SE_EWANTED;
351 			}
352 			retval = 0;
353 		}
354 	} else {
355 		retval = 0;
356 		if (pp->p_selock >= 0) {
357 			/*
358 			 * Readers are not allowed when excl wanted or
359 			 * an FMA retire is pending.
360 			 */
361 			if ((pp->p_selock & SE_EWANTED) == 0) {
362 				if (!PP_PR_NOSHARE(pp)) {
363 					pp->p_selock += SE_READER;
364 					retval = 1;
365 				}
366 			}
367 		}
368 	}
369 
370 	if (retval == 0) {
371 		if ((pp->p_selock & ~SE_EWANTED) == SE_DELETED) {
372 			VM_STAT_ADD(page_lock_deleted);
373 			mutex_exit(pse);
374 			return (retval);
375 		}
376 
377 #ifdef VM_STATS
378 		VM_STAT_ADD(page_lock_miss);
379 		if (upgraded) {
380 			VM_STAT_ADD(page_lock_upgrade_failed);
381 		}
382 #endif
383 		if (lock) {
384 			VM_STAT_ADD(page_lock_miss_lock);
385 			mutex_exit(lock);
386 		}
387 
388 		/*
389 		 * Now, wait for the page to be unlocked and
390 		 * release the lock protecting p_cv and p_selock.
391 		 */
392 		cv_wait(&pp->p_cv, pse);
393 		mutex_exit(pse);
394 
395 		/*
396 		 * The page identity may have changed while we were
397 		 * blocked.  If we are willing to depend on "pp"
398 		 * still pointing to a valid page structure (i.e.,
399 		 * assuming page structures are not dynamically allocated
400 		 * or freed), we could try to lock the page if its
401 		 * identity hasn't changed.
402 		 *
403 		 * This needs to be measured, since we come back from
404 		 * cv_wait holding pse (the expensive part of this
405 		 * operation) we might as well try the cheap part.
406 		 * Though we would also have to confirm that dropping
407 		 * `lock' did not cause any grief to the callers.
408 		 */
409 		if (lock) {
410 			mutex_enter(lock);
411 		}
412 	} else {
413 		/*
414 		 * We have the page lock.
415 		 * If we needed to reclaim the page, and the page
416 		 * needed reclaiming (ie, it was free), then we
417 		 * have the page exclusively locked.  We may need
418 		 * to downgrade the page.
419 		 */
420 		ASSERT((upgraded) ?
421 		    ((PP_ISFREE(pp)) && PAGE_EXCL(pp)) : 1);
422 		mutex_exit(pse);
423 
424 		/*
425 		 * We now hold this page's lock, either shared or
426 		 * exclusive.  This will prevent its identity from changing.
427 		 * The page, however, may or may not be free.  If the caller
428 		 * requested, and it is free, go reclaim it from the
429 		 * free list.  If the page can't be reclaimed, return failure
430 		 * so that the caller can start all over again.
431 		 *
432 		 * NOTE:page_reclaim() releases the page lock (p_selock)
433 		 *	if it can't be reclaimed.
434 		 */
435 		if (reclaim_it) {
436 			if (!page_reclaim(pp, lock)) {
437 				VM_STAT_ADD(page_lock_bad_reclaim);
438 				retval = 0;
439 			} else {
440 				VM_STAT_ADD(page_lock_reclaim);
441 				if (upgraded) {
442 					page_downgrade(pp);
443 				}
444 			}
445 		}
446 	}
447 	return (retval);
448 }
449 
450 /*
451  * Clear the SE_EWANTED bit from p_selock.  This function allows
452  * callers of page_lock_es and page_try_reclaim_lock to clear
453  * their setting of this bit if they decide they no longer wish
454  * to gain exclusive access to the page.  Currently only
455  * delete_memory_thread uses this when the delete memory
456  * operation is cancelled.
457  */
458 void
459 page_lock_clr_exclwanted(page_t *pp)
460 {
461 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
462 
463 	mutex_enter(pse);
464 	pp->p_selock &= ~SE_EWANTED;
465 	if (CV_HAS_WAITERS(&pp->p_cv))
466 		cv_broadcast(&pp->p_cv);
467 	mutex_exit(pse);
468 }
469 
470 /*
471  * Read the comments inside of page_lock_es() carefully.
472  *
473  * SE_EXCL callers specifying es == SE_EXCL_WANTED will cause the
474  * SE_EWANTED bit of p_selock to be set when the lock cannot be obtained.
475  * This is used by threads subject to reader-starvation (eg. memory delete).
476  *
477  * When a thread using SE_EXCL_WANTED does not obtain the SE_EXCL lock,
478  * it is expected that it will retry at a later time.  Threads that will
479  * not retry the lock *must* call page_lock_clr_exclwanted to clear the
480  * SE_EWANTED bit.  (When a thread using SE_EXCL_WANTED obtains the lock,
481  * the bit is cleared.)
482  */
483 int
484 page_try_reclaim_lock(page_t *pp, se_t se, int es)
485 {
486 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
487 	selock_t old;
488 
489 	mutex_enter(pse);
490 
491 	old = pp->p_selock;
492 
493 	ASSERT(((es & SE_EXCL_WANTED) == 0) ||
494 	    ((es & SE_EXCL_WANTED) && (se == SE_EXCL)));
495 
496 	if (PP_RETIRED(pp) && !(es & SE_RETIRED)) {
497 		mutex_exit(pse);
498 		VM_STAT_ADD(page_trylock_failed);
499 		return (0);
500 	}
501 
502 	if (se == SE_SHARED && es == 1 && old == 0) {
503 		se = SE_EXCL;
504 	}
505 
506 	if (se == SE_SHARED) {
507 		if (!PP_ISFREE(pp)) {
508 			if (old >= 0) {
509 				/*
510 				 * Readers are not allowed when excl wanted
511 				 * or a retire is pending.
512 				 */
513 				if ((old & SE_EWANTED) == 0) {
514 					if (!PP_PR_NOSHARE(pp)) {
515 						pp->p_selock = old + SE_READER;
516 						mutex_exit(pse);
517 						return (1);
518 					}
519 				}
520 			}
521 			mutex_exit(pse);
522 			return (0);
523 		}
524 		/*
525 		 * The page is free, so we really want SE_EXCL (below)
526 		 */
527 		VM_STAT_ADD(page_try_reclaim_upgrade);
528 	}
529 
530 	/*
531 	 * The caller wants a writer lock.  We try for it only if
532 	 * SE_EWANTED is not set, or if the caller specified
533 	 * SE_EXCL_WANTED.
534 	 */
535 	if (!(old & SE_EWANTED) || (es & SE_EXCL_WANTED)) {
536 		if ((old & ~SE_EWANTED) == 0) {
537 			/* no reader/writer lock held */
538 			THREAD_KPRI_REQUEST();
539 			/* this clears out our setting of the SE_EWANTED bit */
540 			pp->p_selock = SE_WRITER;
541 			mutex_exit(pse);
542 			return (1);
543 		}
544 	}
545 	if (es & SE_EXCL_WANTED) {
546 		/* page is locked, set the SE_EWANTED bit */
547 		pp->p_selock |= SE_EWANTED;
548 	}
549 	mutex_exit(pse);
550 	return (0);
551 }
552 
553 /*
554  * Acquire a page's "shared/exclusive" lock, but never block.
555  * Returns 1 on success, 0 on failure.
556  */
557 int
558 page_trylock(page_t *pp, se_t se)
559 {
560 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
561 
562 	mutex_enter(pse);
563 	if (pp->p_selock & SE_EWANTED || PP_RETIRED(pp) ||
564 	    (se == SE_SHARED && PP_PR_NOSHARE(pp))) {
565 		/*
566 		 * Fail if a thread wants exclusive access and page is
567 		 * retired, if the page is slated for retirement, or a
568 		 * share lock is requested.
569 		 */
570 		mutex_exit(pse);
571 		VM_STAT_ADD(page_trylock_failed);
572 		return (0);
573 	}
574 
575 	if (se == SE_EXCL) {
576 		if (pp->p_selock == 0) {
577 			THREAD_KPRI_REQUEST();
578 			pp->p_selock = SE_WRITER;
579 			mutex_exit(pse);
580 			return (1);
581 		}
582 	} else {
583 		if (pp->p_selock >= 0) {
584 			pp->p_selock += SE_READER;
585 			mutex_exit(pse);
586 			return (1);
587 		}
588 	}
589 	mutex_exit(pse);
590 	return (0);
591 }
592 
593 /*
594  * Variant of page_unlock() specifically for the page freelist
595  * code. The mere existence of this code is a vile hack that
596  * has resulted due to the backwards locking order of the page
597  * freelist manager; please don't call it.
598  */
599 void
600 page_unlock_noretire(page_t *pp)
601 {
602 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
603 	selock_t old;
604 
605 	mutex_enter(pse);
606 
607 	old = pp->p_selock;
608 	if ((old & ~SE_EWANTED) == SE_READER) {
609 		pp->p_selock = old & ~SE_READER;
610 		if (CV_HAS_WAITERS(&pp->p_cv))
611 			cv_broadcast(&pp->p_cv);
612 	} else if ((old & ~SE_EWANTED) == SE_DELETED) {
613 		panic("page_unlock_noretire: page %p is deleted", pp);
614 	} else if (old < 0) {
615 		THREAD_KPRI_RELEASE();
616 		pp->p_selock &= SE_EWANTED;
617 		if (CV_HAS_WAITERS(&pp->p_cv))
618 			cv_broadcast(&pp->p_cv);
619 	} else if ((old & ~SE_EWANTED) > SE_READER) {
620 		pp->p_selock = old - SE_READER;
621 	} else {
622 		panic("page_unlock_noretire: page %p is not locked", pp);
623 	}
624 
625 	mutex_exit(pse);
626 }
627 
628 /*
629  * Release the page's "shared/exclusive" lock and wake up anyone
630  * who might be waiting for it.
631  */
632 void
633 page_unlock(page_t *pp)
634 {
635 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
636 	selock_t old;
637 
638 	mutex_enter(pse);
639 
640 	old = pp->p_selock;
641 	if ((old & ~SE_EWANTED) == SE_READER) {
642 		pp->p_selock = old & ~SE_READER;
643 		if (CV_HAS_WAITERS(&pp->p_cv))
644 			cv_broadcast(&pp->p_cv);
645 	} else if ((old & ~SE_EWANTED) == SE_DELETED) {
646 		panic("page_unlock: page %p is deleted", pp);
647 	} else if (old < 0) {
648 		THREAD_KPRI_RELEASE();
649 		pp->p_selock &= SE_EWANTED;
650 		if (CV_HAS_WAITERS(&pp->p_cv))
651 			cv_broadcast(&pp->p_cv);
652 	} else if ((old & ~SE_EWANTED) > SE_READER) {
653 		pp->p_selock = old - SE_READER;
654 	} else {
655 		panic("page_unlock: page %p is not locked", pp);
656 	}
657 
658 	if (pp->p_selock == 0 && PP_PR_REQ(pp)) {
659 		/*
660 		 * Try to retire the page. If it retires, great.
661 		 * If not, oh well, we'll get it in the next unlock
662 		 * request, and repeat the cycle.  Regardless,
663 		 * page_tryretire() will drop the page lock.
664 		 */
665 		if ((pp->p_toxic & PR_BUSY) == 0) {
666 			THREAD_KPRI_REQUEST();
667 			pp->p_selock = SE_WRITER;
668 			page_settoxic(pp, PR_BUSY);
669 			mutex_exit(pse);
670 			page_tryretire(pp);
671 		} else {
672 			pp->p_selock = SE_WRITER;
673 			page_clrtoxic(pp, PR_BUSY);
674 			pp->p_selock = 0;
675 			mutex_exit(pse);
676 		}
677 	} else {
678 		mutex_exit(pse);
679 	}
680 }
681 
682 /*
683  * Try to upgrade the lock on the page from a "shared" to an
684  * "exclusive" lock.  Since this upgrade operation is done while
685  * holding the mutex protecting this page, no one else can acquire this page's
686  * lock and change the page. Thus, it is safe to drop the "shared"
687  * lock and attempt to acquire the "exclusive" lock.
688  *
689  * Returns 1 on success, 0 on failure.
690  */
691 int
692 page_tryupgrade(page_t *pp)
693 {
694 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
695 
696 	mutex_enter(pse);
697 	if (!(pp->p_selock & SE_EWANTED)) {
698 		/* no threads want exclusive access, try upgrade */
699 		if (pp->p_selock == SE_READER) {
700 			THREAD_KPRI_REQUEST();
701 			/* convert to exclusive lock */
702 			pp->p_selock = SE_WRITER;
703 			mutex_exit(pse);
704 			return (1);
705 		}
706 	}
707 	mutex_exit(pse);
708 	return (0);
709 }
710 
711 /*
712  * Downgrade the "exclusive" lock on the page to a "shared" lock
713  * while holding the mutex protecting this page's p_selock field.
714  */
715 void
716 page_downgrade(page_t *pp)
717 {
718 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
719 	int excl_waiting;
720 
721 	ASSERT((pp->p_selock & ~SE_EWANTED) != SE_DELETED);
722 	ASSERT(PAGE_EXCL(pp));
723 
724 	mutex_enter(pse);
725 	excl_waiting =  pp->p_selock & SE_EWANTED;
726 	THREAD_KPRI_RELEASE();
727 	pp->p_selock = SE_READER | excl_waiting;
728 	if (CV_HAS_WAITERS(&pp->p_cv))
729 		cv_broadcast(&pp->p_cv);
730 	mutex_exit(pse);
731 }
732 
733 void
734 page_lock_delete(page_t *pp)
735 {
736 	kmutex_t *pse = PAGE_SE_MUTEX(pp);
737 
738 	ASSERT(PAGE_EXCL(pp));
739 	ASSERT(pp->p_vnode == NULL);
740 	ASSERT(pp->p_offset == (u_offset_t)-1);
741 	ASSERT(!PP_ISFREE(pp));
742 
743 	mutex_enter(pse);
744 	THREAD_KPRI_RELEASE();
745 	pp->p_selock = SE_DELETED;
746 	if (CV_HAS_WAITERS(&pp->p_cv))
747 		cv_broadcast(&pp->p_cv);
748 	mutex_exit(pse);
749 }
750 
751 /*
752  * Implement the io lock for pages
753  */
754 void
755 page_iolock_init(page_t *pp)
756 {
757 	pp->p_iolock_state = 0;
758 	cv_init(&pp->p_io_cv, NULL, CV_DEFAULT, NULL);
759 }
760 
761 /*
762  * Acquire the i/o lock on a page.
763  */
764 void
765 page_io_lock(page_t *pp)
766 {
767 	kmutex_t *pio;
768 
769 	pio = PAGE_IO_MUTEX(pp);
770 	mutex_enter(pio);
771 	while (pp->p_iolock_state & PAGE_IO_INUSE) {
772 		cv_wait(&(pp->p_io_cv), pio);
773 	}
774 	pp->p_iolock_state |= PAGE_IO_INUSE;
775 	mutex_exit(pio);
776 }
777 
778 /*
779  * Release the i/o lock on a page.
780  */
781 void
782 page_io_unlock(page_t *pp)
783 {
784 	kmutex_t *pio;
785 
786 	pio = PAGE_IO_MUTEX(pp);
787 	mutex_enter(pio);
788 	cv_signal(&pp->p_io_cv);
789 	pp->p_iolock_state &= ~PAGE_IO_INUSE;
790 	mutex_exit(pio);
791 }
792 
793 /*
794  * Try to acquire the i/o lock on a page without blocking.
795  * Returns 1 on success, 0 on failure.
796  */
797 int
798 page_io_trylock(page_t *pp)
799 {
800 	kmutex_t *pio;
801 
802 	if (pp->p_iolock_state & PAGE_IO_INUSE)
803 		return (0);
804 
805 	pio = PAGE_IO_MUTEX(pp);
806 	mutex_enter(pio);
807 
808 	if (pp->p_iolock_state & PAGE_IO_INUSE) {
809 		mutex_exit(pio);
810 		return (0);
811 	}
812 	pp->p_iolock_state |= PAGE_IO_INUSE;
813 	mutex_exit(pio);
814 
815 	return (1);
816 }
817 
818 /*
819  * Assert that the i/o lock on a page is held.
820  * Returns 1 on success, 0 on failure.
821  */
822 int
823 page_iolock_assert(page_t *pp)
824 {
825 	return (pp->p_iolock_state & PAGE_IO_INUSE);
826 }
827 
828 /*
829  * Wrapper exported to kernel routines that are built
830  * platform-independent (the macro is platform-dependent;
831  * the size of vph_mutex[] is based on NCPU).
832  *
833  * Note that you can do stress testing on this by setting the
834  * variable page_vnode_mutex_stress to something other than
835  * zero in a DEBUG kernel in a debugger after loading the kernel.
836  * Setting it after the kernel is running may not work correctly.
837  */
838 #ifdef DEBUG
839 static int page_vnode_mutex_stress = 0;
840 #endif
841 
842 kmutex_t *
843 page_vnode_mutex(vnode_t *vp)
844 {
845 	if (vp == &kvp)
846 		return (&vph_mutex[VPH_TABLE_SIZE + 0]);
847 #ifdef DEBUG
848 	if (page_vnode_mutex_stress != 0)
849 		return (&vph_mutex[0]);
850 #endif
851 
852 	return (&vph_mutex[VP_HASH_FUNC(vp)]);
853 }
854 
855 kmutex_t *
856 page_se_mutex(page_t *pp)
857 {
858 	return (PAGE_SE_MUTEX(pp));
859 }
860 
861 #ifdef VM_STATS
862 uint_t pszclck_stat[4];
863 #endif
864 /*
865  * Find, take and return a mutex held by hat_page_demote().
866  * Called by page_demote_vp_pages() before hat_page_demote() call and by
867  * routines that want to block hat_page_demote() but can't do it
868  * via locking all constituent pages.
869  *
870  * Return NULL if p_szc is 0.
871  *
872  * It should only be used for pages that can be demoted by hat_page_demote()
873  * i.e. non swapfs file system pages.  The logic here is lifted from
874  * sfmmu_mlspl_enter() except there's no need to worry about p_szc increase
875  * since the page is locked and not free.
876  *
877  * Hash of the root page is used to find the lock.
878  * To find the root in the presense of hat_page_demote() chageing the location
879  * of the root this routine relies on the fact that hat_page_demote() changes
880  * root last.
881  *
882  * If NULL is returned pp's p_szc is guaranteed to be 0. If non NULL is
883  * returned pp's p_szc may be any value.
884  */
885 kmutex_t *
886 page_szc_lock(page_t *pp)
887 {
888 	kmutex_t	*mtx;
889 	page_t		*rootpp;
890 	uint_t		szc;
891 	uint_t		rszc;
892 	uint_t		pszc = pp->p_szc;
893 
894 	ASSERT(pp != NULL);
895 	ASSERT(PAGE_LOCKED(pp));
896 	ASSERT(!PP_ISFREE(pp));
897 	ASSERT(pp->p_vnode != NULL);
898 	ASSERT(!IS_SWAPFSVP(pp->p_vnode));
899 	ASSERT(pp->p_vnode != &kvp);
900 
901 again:
902 	if (pszc == 0) {
903 		VM_STAT_ADD(pszclck_stat[0]);
904 		return (NULL);
905 	}
906 
907 	/* The lock lives in the root page */
908 
909 	rootpp = PP_GROUPLEADER(pp, pszc);
910 	mtx = PAGE_SZC_MUTEX(rootpp);
911 	mutex_enter(mtx);
912 
913 	/*
914 	 * since p_szc can only decrease if pp == rootpp
915 	 * rootpp will be always the same i.e we have the right root
916 	 * regardless of rootpp->p_szc.
917 	 * If location of pp's root didn't change after we took
918 	 * the lock we have the right root. return mutex hashed off it.
919 	 */
920 	if (pp == rootpp || (rszc = rootpp->p_szc) == pszc) {
921 		VM_STAT_ADD(pszclck_stat[1]);
922 		return (mtx);
923 	}
924 
925 	/*
926 	 * root location changed because page got demoted.
927 	 * locate the new root.
928 	 */
929 	if (rszc < pszc) {
930 		szc = pp->p_szc;
931 		ASSERT(szc < pszc);
932 		mutex_exit(mtx);
933 		pszc = szc;
934 		VM_STAT_ADD(pszclck_stat[2]);
935 		goto again;
936 	}
937 
938 	VM_STAT_ADD(pszclck_stat[3]);
939 	/*
940 	 * current hat_page_demote not done yet.
941 	 * wait for it to finish.
942 	 */
943 	mutex_exit(mtx);
944 	rootpp = PP_GROUPLEADER(rootpp, rszc);
945 	mtx = PAGE_SZC_MUTEX(rootpp);
946 	mutex_enter(mtx);
947 	mutex_exit(mtx);
948 	ASSERT(rootpp->p_szc < rszc);
949 	goto again;
950 }
951 
952 int
953 page_szc_lock_assert(page_t *pp)
954 {
955 	page_t *rootpp = PP_PAGEROOT(pp);
956 	kmutex_t *mtx = PAGE_SZC_MUTEX(rootpp);
957 
958 	return (MUTEX_HELD(mtx));
959 }
960