17c478bd9Sstevel@tonic-gate /* 27c478bd9Sstevel@tonic-gate * CDDL HEADER START 37c478bd9Sstevel@tonic-gate * 47c478bd9Sstevel@tonic-gate * The contents of this file are subject to the terms of the 58bc68872Selowe * Common Development and Distribution License (the "License"). 68bc68872Selowe * You may not use this file except in compliance with the License. 77c478bd9Sstevel@tonic-gate * 87c478bd9Sstevel@tonic-gate * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 97c478bd9Sstevel@tonic-gate * or http://www.opensolaris.org/os/licensing. 107c478bd9Sstevel@tonic-gate * See the License for the specific language governing permissions 117c478bd9Sstevel@tonic-gate * and limitations under the License. 127c478bd9Sstevel@tonic-gate * 137c478bd9Sstevel@tonic-gate * When distributing Covered Code, include this CDDL HEADER in each 147c478bd9Sstevel@tonic-gate * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 157c478bd9Sstevel@tonic-gate * If applicable, add the following below this CDDL HEADER, with the 167c478bd9Sstevel@tonic-gate * fields enclosed by brackets "[]" replaced with your own identifying 177c478bd9Sstevel@tonic-gate * information: Portions Copyright [yyyy] [name of copyright owner] 187c478bd9Sstevel@tonic-gate * 197c478bd9Sstevel@tonic-gate * CDDL HEADER END 207c478bd9Sstevel@tonic-gate */ 217c478bd9Sstevel@tonic-gate /* 228bc68872Selowe * Copyright 2006 Sun Microsystems, Inc. All rights reserved. 237c478bd9Sstevel@tonic-gate * Use is subject to license terms. 247c478bd9Sstevel@tonic-gate */ 257c478bd9Sstevel@tonic-gate 267c478bd9Sstevel@tonic-gate #pragma ident "%Z%%M% %I% %E% SMI" 277c478bd9Sstevel@tonic-gate 287c478bd9Sstevel@tonic-gate /* 297c478bd9Sstevel@tonic-gate * VM - page locking primitives 307c478bd9Sstevel@tonic-gate */ 317c478bd9Sstevel@tonic-gate #include <sys/param.h> 327c478bd9Sstevel@tonic-gate #include <sys/t_lock.h> 337c478bd9Sstevel@tonic-gate #include <sys/vtrace.h> 347c478bd9Sstevel@tonic-gate #include <sys/debug.h> 357c478bd9Sstevel@tonic-gate #include <sys/cmn_err.h> 367c478bd9Sstevel@tonic-gate #include <sys/vnode.h> 377c478bd9Sstevel@tonic-gate #include <sys/bitmap.h> 387c478bd9Sstevel@tonic-gate #include <sys/lockstat.h> 397c478bd9Sstevel@tonic-gate #include <sys/condvar_impl.h> 407c478bd9Sstevel@tonic-gate #include <vm/page.h> 417c478bd9Sstevel@tonic-gate #include <vm/seg_enum.h> 427c478bd9Sstevel@tonic-gate #include <vm/vm_dep.h> 437c478bd9Sstevel@tonic-gate 447c478bd9Sstevel@tonic-gate /* 457c478bd9Sstevel@tonic-gate * This global mutex is for logical page locking. 467c478bd9Sstevel@tonic-gate * The following fields in the page structure are protected 477c478bd9Sstevel@tonic-gate * by this lock: 487c478bd9Sstevel@tonic-gate * 497c478bd9Sstevel@tonic-gate * p_lckcnt 507c478bd9Sstevel@tonic-gate * p_cowcnt 517c478bd9Sstevel@tonic-gate */ 527c478bd9Sstevel@tonic-gate kmutex_t page_llock; 537c478bd9Sstevel@tonic-gate 547c478bd9Sstevel@tonic-gate /* 557c478bd9Sstevel@tonic-gate * This is a global lock for the logical page free list. The 567c478bd9Sstevel@tonic-gate * logical free list, in this implementation, is maintained as two 577c478bd9Sstevel@tonic-gate * separate physical lists - the cache list and the free list. 587c478bd9Sstevel@tonic-gate */ 597c478bd9Sstevel@tonic-gate kmutex_t page_freelock; 607c478bd9Sstevel@tonic-gate 617c478bd9Sstevel@tonic-gate /* 627c478bd9Sstevel@tonic-gate * The hash table, page_hash[], the p_selock fields, and the 637c478bd9Sstevel@tonic-gate * list of pages associated with vnodes are protected by arrays of mutexes. 647c478bd9Sstevel@tonic-gate * 657c478bd9Sstevel@tonic-gate * Unless the hashes are changed radically, the table sizes must be 667c478bd9Sstevel@tonic-gate * a power of two. Also, we typically need more mutexes for the 677c478bd9Sstevel@tonic-gate * vnodes since these locks are occasionally held for long periods. 687c478bd9Sstevel@tonic-gate * And since there seem to be two special vnodes (kvp and swapvp), 697c478bd9Sstevel@tonic-gate * we make room for private mutexes for them. 707c478bd9Sstevel@tonic-gate * 717c478bd9Sstevel@tonic-gate * The pse_mutex[] array holds the mutexes to protect the p_selock 727c478bd9Sstevel@tonic-gate * fields of all page_t structures. 737c478bd9Sstevel@tonic-gate * 747c478bd9Sstevel@tonic-gate * PAGE_SE_MUTEX(pp) returns the address of the appropriate mutex 757c478bd9Sstevel@tonic-gate * when given a pointer to a page_t. 767c478bd9Sstevel@tonic-gate * 777c478bd9Sstevel@tonic-gate * PSE_TABLE_SIZE must be a power of two. One could argue that we 787c478bd9Sstevel@tonic-gate * should go to the trouble of setting it up at run time and base it 797c478bd9Sstevel@tonic-gate * on memory size rather than the number of compile time CPUs. 807c478bd9Sstevel@tonic-gate * 817c478bd9Sstevel@tonic-gate * XX64 We should be using physmem size to calculate PSE_TABLE_SIZE, 827c478bd9Sstevel@tonic-gate * PSE_SHIFT, PIO_SHIFT. 837c478bd9Sstevel@tonic-gate * 847c478bd9Sstevel@tonic-gate * These might break in 64 bit world. 857c478bd9Sstevel@tonic-gate */ 867c478bd9Sstevel@tonic-gate #define PSE_SHIFT 7 /* log2(PSE_TABLE_SIZE) */ 877c478bd9Sstevel@tonic-gate 887c478bd9Sstevel@tonic-gate #define PSE_TABLE_SIZE 128 /* number of mutexes to have */ 897c478bd9Sstevel@tonic-gate 907c478bd9Sstevel@tonic-gate #define PIO_SHIFT PSE_SHIFT /* next power of 2 bigger than page_t */ 917c478bd9Sstevel@tonic-gate #define PIO_TABLE_SIZE PSE_TABLE_SIZE /* number of io mutexes to have */ 927c478bd9Sstevel@tonic-gate 937c478bd9Sstevel@tonic-gate pad_mutex_t ph_mutex[PH_TABLE_SIZE]; 947c478bd9Sstevel@tonic-gate pad_mutex_t pse_mutex[PSE_TABLE_SIZE]; 957c478bd9Sstevel@tonic-gate kmutex_t pio_mutex[PIO_TABLE_SIZE]; 967c478bd9Sstevel@tonic-gate 977c478bd9Sstevel@tonic-gate #define PAGE_SE_MUTEX(pp) \ 987c478bd9Sstevel@tonic-gate &pse_mutex[((((uintptr_t)(pp) >> PSE_SHIFT) ^ \ 997c478bd9Sstevel@tonic-gate ((uintptr_t)(pp) >> (PSE_SHIFT << 1))) & \ 1007c478bd9Sstevel@tonic-gate (PSE_TABLE_SIZE - 1))].pad_mutex 1017c478bd9Sstevel@tonic-gate 1027c478bd9Sstevel@tonic-gate #define PAGE_IO_MUTEX(pp) \ 1037c478bd9Sstevel@tonic-gate &pio_mutex[(((uintptr_t)pp) >> PIO_SHIFT) & (PIO_TABLE_SIZE - 1)] 1047c478bd9Sstevel@tonic-gate 1057c478bd9Sstevel@tonic-gate #define PSZC_MTX_TABLE_SIZE 128 1067c478bd9Sstevel@tonic-gate #define PSZC_MTX_TABLE_SHIFT 7 1077c478bd9Sstevel@tonic-gate 1087c478bd9Sstevel@tonic-gate static pad_mutex_t pszc_mutex[PSZC_MTX_TABLE_SIZE]; 1097c478bd9Sstevel@tonic-gate 1107c478bd9Sstevel@tonic-gate #define PAGE_SZC_MUTEX(_pp) \ 1117c478bd9Sstevel@tonic-gate &pszc_mutex[((((uintptr_t)(_pp) >> PSZC_MTX_TABLE_SHIFT) ^ \ 1127c478bd9Sstevel@tonic-gate ((uintptr_t)(_pp) >> (PSZC_MTX_TABLE_SHIFT << 1)) ^ \ 1137c478bd9Sstevel@tonic-gate ((uintptr_t)(_pp) >> (3 * PSZC_MTX_TABLE_SHIFT))) & \ 1147c478bd9Sstevel@tonic-gate (PSZC_MTX_TABLE_SIZE - 1))].pad_mutex 1157c478bd9Sstevel@tonic-gate 1167c478bd9Sstevel@tonic-gate /* 1177c478bd9Sstevel@tonic-gate * The vph_mutex[] array holds the mutexes to protect the vnode chains, 1187c478bd9Sstevel@tonic-gate * (i.e., the list of pages anchored by v_pages and connected via p_vpprev 1197c478bd9Sstevel@tonic-gate * and p_vpnext). 1207c478bd9Sstevel@tonic-gate * 1217c478bd9Sstevel@tonic-gate * The page_vnode_mutex(vp) function returns the address of the appropriate 1227c478bd9Sstevel@tonic-gate * mutex from this array given a pointer to a vnode. It is complicated 1237c478bd9Sstevel@tonic-gate * by the fact that the kernel's vnode and the swapfs vnode are referenced 1247c478bd9Sstevel@tonic-gate * frequently enough to warrent their own mutexes. 1257c478bd9Sstevel@tonic-gate * 1267c478bd9Sstevel@tonic-gate * The VP_HASH_FUNC returns the index into the vph_mutex array given 1277c478bd9Sstevel@tonic-gate * an address of a vnode. 1287c478bd9Sstevel@tonic-gate */ 1297c478bd9Sstevel@tonic-gate 1307c478bd9Sstevel@tonic-gate /* 1317c478bd9Sstevel@tonic-gate * XX64 VPH_TABLE_SIZE and VP_HASH_FUNC might break in 64 bit world. 1327c478bd9Sstevel@tonic-gate * Need to review again. 1337c478bd9Sstevel@tonic-gate */ 1347c478bd9Sstevel@tonic-gate #define VPH_TABLE_SIZE (2 << VP_SHIFT) 1357c478bd9Sstevel@tonic-gate 1367c478bd9Sstevel@tonic-gate #define VP_HASH_FUNC(vp) \ 1377c478bd9Sstevel@tonic-gate ((((uintptr_t)(vp) >> 6) + \ 1387c478bd9Sstevel@tonic-gate ((uintptr_t)(vp) >> 8) + \ 1397c478bd9Sstevel@tonic-gate ((uintptr_t)(vp) >> 10) + \ 1407c478bd9Sstevel@tonic-gate ((uintptr_t)(vp) >> 12)) \ 1417c478bd9Sstevel@tonic-gate & (VPH_TABLE_SIZE - 1)) 1427c478bd9Sstevel@tonic-gate 1437c478bd9Sstevel@tonic-gate extern struct vnode kvp; 1447c478bd9Sstevel@tonic-gate 1457c478bd9Sstevel@tonic-gate kmutex_t vph_mutex[VPH_TABLE_SIZE + 2]; 1467c478bd9Sstevel@tonic-gate 1477c478bd9Sstevel@tonic-gate /* 1487c478bd9Sstevel@tonic-gate * Initialize the locks used by the Virtual Memory Management system. 1497c478bd9Sstevel@tonic-gate */ 1507c478bd9Sstevel@tonic-gate void 1517c478bd9Sstevel@tonic-gate page_lock_init() 1527c478bd9Sstevel@tonic-gate { 1537c478bd9Sstevel@tonic-gate } 1547c478bd9Sstevel@tonic-gate 1557c478bd9Sstevel@tonic-gate /* 1567c478bd9Sstevel@tonic-gate * At present we only use page ownership to aid debugging, so it's 1577c478bd9Sstevel@tonic-gate * OK if the owner field isn't exact. In the 32-bit world two thread ids 1587c478bd9Sstevel@tonic-gate * can map to the same owner because we just 'or' in 0x80000000 and 1597c478bd9Sstevel@tonic-gate * then clear the second highest bit, so that (for example) 0x2faced00 1607c478bd9Sstevel@tonic-gate * and 0xafaced00 both map to 0xafaced00. 1617c478bd9Sstevel@tonic-gate * In the 64-bit world, p_selock may not be large enough to hold a full 1627c478bd9Sstevel@tonic-gate * thread pointer. If we ever need precise ownership (e.g. if we implement 1637c478bd9Sstevel@tonic-gate * priority inheritance for page locks) then p_selock should become a 1647c478bd9Sstevel@tonic-gate * uintptr_t and SE_WRITER should be -((uintptr_t)curthread >> 2). 1657c478bd9Sstevel@tonic-gate */ 1667c478bd9Sstevel@tonic-gate #define SE_WRITER (((selock_t)(ulong_t)curthread | INT_MIN) & ~SE_EWANTED) 1677c478bd9Sstevel@tonic-gate #define SE_READER 1 1687c478bd9Sstevel@tonic-gate 1697c478bd9Sstevel@tonic-gate /* 1707c478bd9Sstevel@tonic-gate * A page that is deleted must be marked as such using the 1717c478bd9Sstevel@tonic-gate * page_lock_delete() function. The page must be exclusively locked. 1727c478bd9Sstevel@tonic-gate * The SE_DELETED marker is put in p_selock when this function is called. 1737c478bd9Sstevel@tonic-gate * SE_DELETED must be distinct from any SE_WRITER value. 1747c478bd9Sstevel@tonic-gate */ 1757c478bd9Sstevel@tonic-gate #define SE_DELETED (1 | INT_MIN) 1767c478bd9Sstevel@tonic-gate 1777c478bd9Sstevel@tonic-gate #ifdef VM_STATS 1787c478bd9Sstevel@tonic-gate uint_t vph_kvp_count; 1797c478bd9Sstevel@tonic-gate uint_t vph_swapfsvp_count; 1807c478bd9Sstevel@tonic-gate uint_t vph_other; 1817c478bd9Sstevel@tonic-gate #endif /* VM_STATS */ 1827c478bd9Sstevel@tonic-gate 1837c478bd9Sstevel@tonic-gate #ifdef VM_STATS 1847c478bd9Sstevel@tonic-gate uint_t page_lock_count; 1857c478bd9Sstevel@tonic-gate uint_t page_lock_miss; 1867c478bd9Sstevel@tonic-gate uint_t page_lock_miss_lock; 1877c478bd9Sstevel@tonic-gate uint_t page_lock_reclaim; 1887c478bd9Sstevel@tonic-gate uint_t page_lock_bad_reclaim; 1897c478bd9Sstevel@tonic-gate uint_t page_lock_same_page; 1907c478bd9Sstevel@tonic-gate uint_t page_lock_upgrade; 191db874c57Selowe uint_t page_lock_retired; 1927c478bd9Sstevel@tonic-gate uint_t page_lock_upgrade_failed; 1937c478bd9Sstevel@tonic-gate uint_t page_lock_deleted; 1947c478bd9Sstevel@tonic-gate 1957c478bd9Sstevel@tonic-gate uint_t page_trylock_locked; 196db874c57Selowe uint_t page_trylock_failed; 1977c478bd9Sstevel@tonic-gate uint_t page_trylock_missed; 1987c478bd9Sstevel@tonic-gate 1997c478bd9Sstevel@tonic-gate uint_t page_try_reclaim_upgrade; 2007c478bd9Sstevel@tonic-gate #endif /* VM_STATS */ 2017c478bd9Sstevel@tonic-gate 2027c478bd9Sstevel@tonic-gate /* 2037c478bd9Sstevel@tonic-gate * Acquire the "shared/exclusive" lock on a page. 2047c478bd9Sstevel@tonic-gate * 2057c478bd9Sstevel@tonic-gate * Returns 1 on success and locks the page appropriately. 2067c478bd9Sstevel@tonic-gate * 0 on failure and does not lock the page. 2077c478bd9Sstevel@tonic-gate * 2087c478bd9Sstevel@tonic-gate * If `lock' is non-NULL, it will be dropped and reacquired in the 2097c478bd9Sstevel@tonic-gate * failure case. This routine can block, and if it does 2107c478bd9Sstevel@tonic-gate * it will always return a failure since the page identity [vp, off] 2117c478bd9Sstevel@tonic-gate * or state may have changed. 2127c478bd9Sstevel@tonic-gate */ 2137c478bd9Sstevel@tonic-gate 2147c478bd9Sstevel@tonic-gate int 2157c478bd9Sstevel@tonic-gate page_lock(page_t *pp, se_t se, kmutex_t *lock, reclaim_t reclaim) 2167c478bd9Sstevel@tonic-gate { 2177c478bd9Sstevel@tonic-gate return (page_lock_es(pp, se, lock, reclaim, 0)); 2187c478bd9Sstevel@tonic-gate } 2197c478bd9Sstevel@tonic-gate 2207c478bd9Sstevel@tonic-gate /* 2217c478bd9Sstevel@tonic-gate * With the addition of reader-writer lock semantics to page_lock_es, 2227c478bd9Sstevel@tonic-gate * callers wanting an exclusive (writer) lock may prevent shared-lock 2237c478bd9Sstevel@tonic-gate * (reader) starvation by setting the es parameter to SE_EXCL_WANTED. 2247c478bd9Sstevel@tonic-gate * In this case, when an exclusive lock cannot be acquired, p_selock's 225db874c57Selowe * SE_EWANTED bit is set. Shared-lock (reader) requests are also denied 226db874c57Selowe * if the page is slated for retirement. 227db874c57Selowe * 228db874c57Selowe * The se and es parameters determine if the lock should be granted 229db874c57Selowe * based on the following decision table: 230db874c57Selowe * 231db874c57Selowe * Lock wanted es flags p_selock/SE_EWANTED Action 232db874c57Selowe * ----------- -------------- ------------------- --------- 233db874c57Selowe * SE_EXCL any [1][2] unlocked/any grant lock, clear SE_EWANTED 234db874c57Selowe * SE_EXCL SE_EWANTED any lock/any deny, set SE_EWANTED 235db874c57Selowe * SE_EXCL none any lock/any deny 2368bc68872Selowe * SE_SHARED n/a [2] shared/0 grant 2378bc68872Selowe * SE_SHARED n/a [2] unlocked/0 grant 238db874c57Selowe * SE_SHARED n/a shared/1 deny 239db874c57Selowe * SE_SHARED n/a unlocked/1 deny 240db874c57Selowe * SE_SHARED n/a excl/any deny 241db874c57Selowe * 242db874c57Selowe * Notes: 243db874c57Selowe * [1] The code grants an exclusive lock to the caller and clears the bit 244db874c57Selowe * SE_EWANTED whenever p_selock is unlocked, regardless of the SE_EWANTED 245db874c57Selowe * bit's value. This was deemed acceptable as we are not concerned about 246db874c57Selowe * exclusive-lock starvation. If this ever becomes an issue, a priority or 247db874c57Selowe * fifo mechanism should also be implemented. Meantime, the thread that 248db874c57Selowe * set SE_EWANTED should be prepared to catch this condition and reset it 249db874c57Selowe * 250db874c57Selowe * [2] Retired pages may not be locked at any time, regardless of the 251db874c57Selowe * dispostion of se, unless the es parameter has SE_RETIRED flag set. 2527c478bd9Sstevel@tonic-gate * 253db874c57Selowe * Notes on values of "es": 254db874c57Selowe * 255db874c57Selowe * es & 1: page_lookup_create will attempt page relocation 256db874c57Selowe * es & SE_EXCL_WANTED: caller wants SE_EWANTED set (eg. delete 257db874c57Selowe * memory thread); this prevents reader-starvation of waiting 258db874c57Selowe * writer thread(s) by giving priority to writers over readers. 259db874c57Selowe * es & SE_RETIRED: caller wants to lock pages even if they are 260db874c57Selowe * retired. Default is to deny the lock if the page is retired. 261db874c57Selowe * 262db874c57Selowe * And yes, we know, the semantics of this function are too complicated. 263db874c57Selowe * It's on the list to be cleaned up. 2647c478bd9Sstevel@tonic-gate */ 2657c478bd9Sstevel@tonic-gate int 2667c478bd9Sstevel@tonic-gate page_lock_es(page_t *pp, se_t se, kmutex_t *lock, reclaim_t reclaim, int es) 2677c478bd9Sstevel@tonic-gate { 2687c478bd9Sstevel@tonic-gate int retval; 2697c478bd9Sstevel@tonic-gate kmutex_t *pse = PAGE_SE_MUTEX(pp); 2707c478bd9Sstevel@tonic-gate int upgraded; 2717c478bd9Sstevel@tonic-gate int reclaim_it; 2727c478bd9Sstevel@tonic-gate 2737c478bd9Sstevel@tonic-gate ASSERT(lock != NULL ? MUTEX_HELD(lock) : 1); 2747c478bd9Sstevel@tonic-gate 2757c478bd9Sstevel@tonic-gate VM_STAT_ADD(page_lock_count); 2767c478bd9Sstevel@tonic-gate 2777c478bd9Sstevel@tonic-gate upgraded = 0; 2787c478bd9Sstevel@tonic-gate reclaim_it = 0; 2797c478bd9Sstevel@tonic-gate 2807c478bd9Sstevel@tonic-gate mutex_enter(pse); 2817c478bd9Sstevel@tonic-gate 2827c478bd9Sstevel@tonic-gate ASSERT(((es & SE_EXCL_WANTED) == 0) || 283db874c57Selowe ((es & SE_EXCL_WANTED) && (se == SE_EXCL))); 284db874c57Selowe 285db874c57Selowe if (PP_RETIRED(pp) && !(es & SE_RETIRED)) { 286db874c57Selowe mutex_exit(pse); 287db874c57Selowe VM_STAT_ADD(page_lock_retired); 288db874c57Selowe return (0); 289db874c57Selowe } 2907c478bd9Sstevel@tonic-gate 2917c478bd9Sstevel@tonic-gate if (se == SE_SHARED && es == 1 && pp->p_selock == 0) { 2927c478bd9Sstevel@tonic-gate se = SE_EXCL; 2937c478bd9Sstevel@tonic-gate } 2947c478bd9Sstevel@tonic-gate 2957c478bd9Sstevel@tonic-gate if ((reclaim == P_RECLAIM) && (PP_ISFREE(pp))) { 2967c478bd9Sstevel@tonic-gate 2977c478bd9Sstevel@tonic-gate reclaim_it = 1; 2987c478bd9Sstevel@tonic-gate if (se == SE_SHARED) { 2997c478bd9Sstevel@tonic-gate /* 3007c478bd9Sstevel@tonic-gate * This is an interesting situation. 3017c478bd9Sstevel@tonic-gate * 3027c478bd9Sstevel@tonic-gate * Remember that p_free can only change if 3037c478bd9Sstevel@tonic-gate * p_selock < 0. 3047c478bd9Sstevel@tonic-gate * p_free does not depend on our holding `pse'. 3057c478bd9Sstevel@tonic-gate * And, since we hold `pse', p_selock can not change. 3067c478bd9Sstevel@tonic-gate * So, if p_free changes on us, the page is already 3077c478bd9Sstevel@tonic-gate * exclusively held, and we would fail to get p_selock 3087c478bd9Sstevel@tonic-gate * regardless. 3097c478bd9Sstevel@tonic-gate * 3107c478bd9Sstevel@tonic-gate * We want to avoid getting the share 3117c478bd9Sstevel@tonic-gate * lock on a free page that needs to be reclaimed. 3127c478bd9Sstevel@tonic-gate * It is possible that some other thread has the share 3137c478bd9Sstevel@tonic-gate * lock and has left the free page on the cache list. 3147c478bd9Sstevel@tonic-gate * pvn_vplist_dirty() does this for brief periods. 3157c478bd9Sstevel@tonic-gate * If the se_share is currently SE_EXCL, we will fail 3167c478bd9Sstevel@tonic-gate * to acquire p_selock anyway. Blocking is the 3177c478bd9Sstevel@tonic-gate * right thing to do. 3187c478bd9Sstevel@tonic-gate * If we need to reclaim this page, we must get 3197c478bd9Sstevel@tonic-gate * exclusive access to it, force the upgrade now. 3207c478bd9Sstevel@tonic-gate * Again, we will fail to acquire p_selock if the 3217c478bd9Sstevel@tonic-gate * page is not free and block. 3227c478bd9Sstevel@tonic-gate */ 3237c478bd9Sstevel@tonic-gate upgraded = 1; 3247c478bd9Sstevel@tonic-gate se = SE_EXCL; 3257c478bd9Sstevel@tonic-gate VM_STAT_ADD(page_lock_upgrade); 3267c478bd9Sstevel@tonic-gate } 3277c478bd9Sstevel@tonic-gate } 3287c478bd9Sstevel@tonic-gate 3297c478bd9Sstevel@tonic-gate if (se == SE_EXCL) { 330db874c57Selowe if (!(es & SE_EXCL_WANTED) && (pp->p_selock & SE_EWANTED)) { 3317c478bd9Sstevel@tonic-gate /* 3327c478bd9Sstevel@tonic-gate * if the caller wants a writer lock (but did not 3337c478bd9Sstevel@tonic-gate * specify exclusive access), and there is a pending 3347c478bd9Sstevel@tonic-gate * writer that wants exclusive access, return failure 3357c478bd9Sstevel@tonic-gate */ 3367c478bd9Sstevel@tonic-gate retval = 0; 3377c478bd9Sstevel@tonic-gate } else if ((pp->p_selock & ~SE_EWANTED) == 0) { 3387c478bd9Sstevel@tonic-gate /* no reader/writer lock held */ 3397c478bd9Sstevel@tonic-gate THREAD_KPRI_REQUEST(); 3407c478bd9Sstevel@tonic-gate /* this clears our setting of the SE_EWANTED bit */ 3417c478bd9Sstevel@tonic-gate pp->p_selock = SE_WRITER; 3427c478bd9Sstevel@tonic-gate retval = 1; 3437c478bd9Sstevel@tonic-gate } else { 3447c478bd9Sstevel@tonic-gate /* page is locked */ 345db874c57Selowe if (es & SE_EXCL_WANTED) { 3467c478bd9Sstevel@tonic-gate /* set the SE_EWANTED bit */ 3477c478bd9Sstevel@tonic-gate pp->p_selock |= SE_EWANTED; 3487c478bd9Sstevel@tonic-gate } 3497c478bd9Sstevel@tonic-gate retval = 0; 3507c478bd9Sstevel@tonic-gate } 3517c478bd9Sstevel@tonic-gate } else { 3527c478bd9Sstevel@tonic-gate retval = 0; 3537c478bd9Sstevel@tonic-gate if (pp->p_selock >= 0) { 354db874c57Selowe if ((pp->p_selock & SE_EWANTED) == 0) { 3558bc68872Selowe pp->p_selock += SE_READER; 3568bc68872Selowe retval = 1; 3577c478bd9Sstevel@tonic-gate } 3587c478bd9Sstevel@tonic-gate } 3597c478bd9Sstevel@tonic-gate } 3607c478bd9Sstevel@tonic-gate 3617c478bd9Sstevel@tonic-gate if (retval == 0) { 3627c478bd9Sstevel@tonic-gate if ((pp->p_selock & ~SE_EWANTED) == SE_DELETED) { 3637c478bd9Sstevel@tonic-gate VM_STAT_ADD(page_lock_deleted); 3647c478bd9Sstevel@tonic-gate mutex_exit(pse); 3657c478bd9Sstevel@tonic-gate return (retval); 3667c478bd9Sstevel@tonic-gate } 3677c478bd9Sstevel@tonic-gate 3687c478bd9Sstevel@tonic-gate #ifdef VM_STATS 3697c478bd9Sstevel@tonic-gate VM_STAT_ADD(page_lock_miss); 3707c478bd9Sstevel@tonic-gate if (upgraded) { 3717c478bd9Sstevel@tonic-gate VM_STAT_ADD(page_lock_upgrade_failed); 3727c478bd9Sstevel@tonic-gate } 3737c478bd9Sstevel@tonic-gate #endif 3747c478bd9Sstevel@tonic-gate if (lock) { 3757c478bd9Sstevel@tonic-gate VM_STAT_ADD(page_lock_miss_lock); 3767c478bd9Sstevel@tonic-gate mutex_exit(lock); 3777c478bd9Sstevel@tonic-gate } 3787c478bd9Sstevel@tonic-gate 3797c478bd9Sstevel@tonic-gate /* 3807c478bd9Sstevel@tonic-gate * Now, wait for the page to be unlocked and 3817c478bd9Sstevel@tonic-gate * release the lock protecting p_cv and p_selock. 3827c478bd9Sstevel@tonic-gate */ 3837c478bd9Sstevel@tonic-gate cv_wait(&pp->p_cv, pse); 3847c478bd9Sstevel@tonic-gate mutex_exit(pse); 3857c478bd9Sstevel@tonic-gate 3867c478bd9Sstevel@tonic-gate /* 3877c478bd9Sstevel@tonic-gate * The page identity may have changed while we were 3887c478bd9Sstevel@tonic-gate * blocked. If we are willing to depend on "pp" 3897c478bd9Sstevel@tonic-gate * still pointing to a valid page structure (i.e., 3907c478bd9Sstevel@tonic-gate * assuming page structures are not dynamically allocated 3917c478bd9Sstevel@tonic-gate * or freed), we could try to lock the page if its 3927c478bd9Sstevel@tonic-gate * identity hasn't changed. 3937c478bd9Sstevel@tonic-gate * 3947c478bd9Sstevel@tonic-gate * This needs to be measured, since we come back from 3957c478bd9Sstevel@tonic-gate * cv_wait holding pse (the expensive part of this 3967c478bd9Sstevel@tonic-gate * operation) we might as well try the cheap part. 3977c478bd9Sstevel@tonic-gate * Though we would also have to confirm that dropping 3987c478bd9Sstevel@tonic-gate * `lock' did not cause any grief to the callers. 3997c478bd9Sstevel@tonic-gate */ 4007c478bd9Sstevel@tonic-gate if (lock) { 4017c478bd9Sstevel@tonic-gate mutex_enter(lock); 4027c478bd9Sstevel@tonic-gate } 4037c478bd9Sstevel@tonic-gate } else { 4047c478bd9Sstevel@tonic-gate /* 4057c478bd9Sstevel@tonic-gate * We have the page lock. 4067c478bd9Sstevel@tonic-gate * If we needed to reclaim the page, and the page 4077c478bd9Sstevel@tonic-gate * needed reclaiming (ie, it was free), then we 4087c478bd9Sstevel@tonic-gate * have the page exclusively locked. We may need 4097c478bd9Sstevel@tonic-gate * to downgrade the page. 4107c478bd9Sstevel@tonic-gate */ 4117c478bd9Sstevel@tonic-gate ASSERT((upgraded) ? 4127c478bd9Sstevel@tonic-gate ((PP_ISFREE(pp)) && PAGE_EXCL(pp)) : 1); 4137c478bd9Sstevel@tonic-gate mutex_exit(pse); 4147c478bd9Sstevel@tonic-gate 4157c478bd9Sstevel@tonic-gate /* 4167c478bd9Sstevel@tonic-gate * We now hold this page's lock, either shared or 4177c478bd9Sstevel@tonic-gate * exclusive. This will prevent its identity from changing. 4187c478bd9Sstevel@tonic-gate * The page, however, may or may not be free. If the caller 4197c478bd9Sstevel@tonic-gate * requested, and it is free, go reclaim it from the 4207c478bd9Sstevel@tonic-gate * free list. If the page can't be reclaimed, return failure 4217c478bd9Sstevel@tonic-gate * so that the caller can start all over again. 4227c478bd9Sstevel@tonic-gate * 4237c478bd9Sstevel@tonic-gate * NOTE:page_reclaim() releases the page lock (p_selock) 4247c478bd9Sstevel@tonic-gate * if it can't be reclaimed. 4257c478bd9Sstevel@tonic-gate */ 4267c478bd9Sstevel@tonic-gate if (reclaim_it) { 4277c478bd9Sstevel@tonic-gate if (!page_reclaim(pp, lock)) { 4287c478bd9Sstevel@tonic-gate VM_STAT_ADD(page_lock_bad_reclaim); 4297c478bd9Sstevel@tonic-gate retval = 0; 4307c478bd9Sstevel@tonic-gate } else { 4317c478bd9Sstevel@tonic-gate VM_STAT_ADD(page_lock_reclaim); 4327c478bd9Sstevel@tonic-gate if (upgraded) { 4337c478bd9Sstevel@tonic-gate page_downgrade(pp); 4347c478bd9Sstevel@tonic-gate } 4357c478bd9Sstevel@tonic-gate } 4367c478bd9Sstevel@tonic-gate } 4377c478bd9Sstevel@tonic-gate } 4387c478bd9Sstevel@tonic-gate return (retval); 4397c478bd9Sstevel@tonic-gate } 4407c478bd9Sstevel@tonic-gate 4417c478bd9Sstevel@tonic-gate /* 4427c478bd9Sstevel@tonic-gate * Clear the SE_EWANTED bit from p_selock. This function allows 4437c478bd9Sstevel@tonic-gate * callers of page_lock_es and page_try_reclaim_lock to clear 4447c478bd9Sstevel@tonic-gate * their setting of this bit if they decide they no longer wish 4457c478bd9Sstevel@tonic-gate * to gain exclusive access to the page. Currently only 4467c478bd9Sstevel@tonic-gate * delete_memory_thread uses this when the delete memory 4477c478bd9Sstevel@tonic-gate * operation is cancelled. 4487c478bd9Sstevel@tonic-gate */ 4497c478bd9Sstevel@tonic-gate void 4507c478bd9Sstevel@tonic-gate page_lock_clr_exclwanted(page_t *pp) 4517c478bd9Sstevel@tonic-gate { 4527c478bd9Sstevel@tonic-gate kmutex_t *pse = PAGE_SE_MUTEX(pp); 4537c478bd9Sstevel@tonic-gate 4547c478bd9Sstevel@tonic-gate mutex_enter(pse); 4557c478bd9Sstevel@tonic-gate pp->p_selock &= ~SE_EWANTED; 4567c478bd9Sstevel@tonic-gate if (CV_HAS_WAITERS(&pp->p_cv)) 4577c478bd9Sstevel@tonic-gate cv_broadcast(&pp->p_cv); 4587c478bd9Sstevel@tonic-gate mutex_exit(pse); 4597c478bd9Sstevel@tonic-gate } 4607c478bd9Sstevel@tonic-gate 4617c478bd9Sstevel@tonic-gate /* 4627c478bd9Sstevel@tonic-gate * Read the comments inside of page_lock_es() carefully. 4637c478bd9Sstevel@tonic-gate * 4647c478bd9Sstevel@tonic-gate * SE_EXCL callers specifying es == SE_EXCL_WANTED will cause the 4657c478bd9Sstevel@tonic-gate * SE_EWANTED bit of p_selock to be set when the lock cannot be obtained. 4667c478bd9Sstevel@tonic-gate * This is used by threads subject to reader-starvation (eg. memory delete). 4677c478bd9Sstevel@tonic-gate * 4687c478bd9Sstevel@tonic-gate * When a thread using SE_EXCL_WANTED does not obtain the SE_EXCL lock, 4697c478bd9Sstevel@tonic-gate * it is expected that it will retry at a later time. Threads that will 4707c478bd9Sstevel@tonic-gate * not retry the lock *must* call page_lock_clr_exclwanted to clear the 4717c478bd9Sstevel@tonic-gate * SE_EWANTED bit. (When a thread using SE_EXCL_WANTED obtains the lock, 4727c478bd9Sstevel@tonic-gate * the bit is cleared.) 4737c478bd9Sstevel@tonic-gate */ 4747c478bd9Sstevel@tonic-gate int 4757c478bd9Sstevel@tonic-gate page_try_reclaim_lock(page_t *pp, se_t se, int es) 4767c478bd9Sstevel@tonic-gate { 4777c478bd9Sstevel@tonic-gate kmutex_t *pse = PAGE_SE_MUTEX(pp); 4787c478bd9Sstevel@tonic-gate selock_t old; 4797c478bd9Sstevel@tonic-gate 4807c478bd9Sstevel@tonic-gate mutex_enter(pse); 4817c478bd9Sstevel@tonic-gate 4827c478bd9Sstevel@tonic-gate old = pp->p_selock; 4837c478bd9Sstevel@tonic-gate 4847c478bd9Sstevel@tonic-gate ASSERT(((es & SE_EXCL_WANTED) == 0) || 485db874c57Selowe ((es & SE_EXCL_WANTED) && (se == SE_EXCL))); 486db874c57Selowe 487db874c57Selowe if (PP_RETIRED(pp) && !(es & SE_RETIRED)) { 488db874c57Selowe mutex_exit(pse); 489db874c57Selowe VM_STAT_ADD(page_trylock_failed); 490db874c57Selowe return (0); 491db874c57Selowe } 4927c478bd9Sstevel@tonic-gate 4937c478bd9Sstevel@tonic-gate if (se == SE_SHARED && es == 1 && old == 0) { 4947c478bd9Sstevel@tonic-gate se = SE_EXCL; 4957c478bd9Sstevel@tonic-gate } 4967c478bd9Sstevel@tonic-gate 4977c478bd9Sstevel@tonic-gate if (se == SE_SHARED) { 4987c478bd9Sstevel@tonic-gate if (!PP_ISFREE(pp)) { 4997c478bd9Sstevel@tonic-gate if (old >= 0) { 500db874c57Selowe /* 501db874c57Selowe * Readers are not allowed when excl wanted 502db874c57Selowe */ 503db874c57Selowe if ((old & SE_EWANTED) == 0) { 5048bc68872Selowe pp->p_selock = old + SE_READER; 5058bc68872Selowe mutex_exit(pse); 5068bc68872Selowe return (1); 5077c478bd9Sstevel@tonic-gate } 5087c478bd9Sstevel@tonic-gate } 5097c478bd9Sstevel@tonic-gate mutex_exit(pse); 5107c478bd9Sstevel@tonic-gate return (0); 5117c478bd9Sstevel@tonic-gate } 5127c478bd9Sstevel@tonic-gate /* 5137c478bd9Sstevel@tonic-gate * The page is free, so we really want SE_EXCL (below) 5147c478bd9Sstevel@tonic-gate */ 5157c478bd9Sstevel@tonic-gate VM_STAT_ADD(page_try_reclaim_upgrade); 5167c478bd9Sstevel@tonic-gate } 5177c478bd9Sstevel@tonic-gate 5187c478bd9Sstevel@tonic-gate /* 5197c478bd9Sstevel@tonic-gate * The caller wants a writer lock. We try for it only if 5207c478bd9Sstevel@tonic-gate * SE_EWANTED is not set, or if the caller specified 5217c478bd9Sstevel@tonic-gate * SE_EXCL_WANTED. 5227c478bd9Sstevel@tonic-gate */ 523db874c57Selowe if (!(old & SE_EWANTED) || (es & SE_EXCL_WANTED)) { 5247c478bd9Sstevel@tonic-gate if ((old & ~SE_EWANTED) == 0) { 5257c478bd9Sstevel@tonic-gate /* no reader/writer lock held */ 5267c478bd9Sstevel@tonic-gate THREAD_KPRI_REQUEST(); 5277c478bd9Sstevel@tonic-gate /* this clears out our setting of the SE_EWANTED bit */ 5287c478bd9Sstevel@tonic-gate pp->p_selock = SE_WRITER; 5297c478bd9Sstevel@tonic-gate mutex_exit(pse); 5307c478bd9Sstevel@tonic-gate return (1); 5317c478bd9Sstevel@tonic-gate } 5327c478bd9Sstevel@tonic-gate } 533db874c57Selowe if (es & SE_EXCL_WANTED) { 5347c478bd9Sstevel@tonic-gate /* page is locked, set the SE_EWANTED bit */ 5357c478bd9Sstevel@tonic-gate pp->p_selock |= SE_EWANTED; 5367c478bd9Sstevel@tonic-gate } 5377c478bd9Sstevel@tonic-gate mutex_exit(pse); 5387c478bd9Sstevel@tonic-gate return (0); 5397c478bd9Sstevel@tonic-gate } 5407c478bd9Sstevel@tonic-gate 5417c478bd9Sstevel@tonic-gate /* 5427c478bd9Sstevel@tonic-gate * Acquire a page's "shared/exclusive" lock, but never block. 5437c478bd9Sstevel@tonic-gate * Returns 1 on success, 0 on failure. 5447c478bd9Sstevel@tonic-gate */ 5457c478bd9Sstevel@tonic-gate int 5467c478bd9Sstevel@tonic-gate page_trylock(page_t *pp, se_t se) 5477c478bd9Sstevel@tonic-gate { 5487c478bd9Sstevel@tonic-gate kmutex_t *pse = PAGE_SE_MUTEX(pp); 5497c478bd9Sstevel@tonic-gate 5507c478bd9Sstevel@tonic-gate mutex_enter(pse); 551db874c57Selowe if (pp->p_selock & SE_EWANTED || PP_RETIRED(pp) || 55224e9c58bSelowe (se == SE_SHARED && PP_PR_NOSHARE(pp))) { 553db874c57Selowe /* 554db874c57Selowe * Fail if a thread wants exclusive access and page is 555db874c57Selowe * retired, if the page is slated for retirement, or a 556db874c57Selowe * share lock is requested. 557db874c57Selowe */ 5587c478bd9Sstevel@tonic-gate mutex_exit(pse); 559db874c57Selowe VM_STAT_ADD(page_trylock_failed); 5607c478bd9Sstevel@tonic-gate return (0); 5617c478bd9Sstevel@tonic-gate } 5627c478bd9Sstevel@tonic-gate 5637c478bd9Sstevel@tonic-gate if (se == SE_EXCL) { 5647c478bd9Sstevel@tonic-gate if (pp->p_selock == 0) { 5657c478bd9Sstevel@tonic-gate THREAD_KPRI_REQUEST(); 5667c478bd9Sstevel@tonic-gate pp->p_selock = SE_WRITER; 5677c478bd9Sstevel@tonic-gate mutex_exit(pse); 5687c478bd9Sstevel@tonic-gate return (1); 5697c478bd9Sstevel@tonic-gate } 5707c478bd9Sstevel@tonic-gate } else { 5717c478bd9Sstevel@tonic-gate if (pp->p_selock >= 0) { 5727c478bd9Sstevel@tonic-gate pp->p_selock += SE_READER; 5737c478bd9Sstevel@tonic-gate mutex_exit(pse); 5747c478bd9Sstevel@tonic-gate return (1); 5757c478bd9Sstevel@tonic-gate } 5767c478bd9Sstevel@tonic-gate } 5777c478bd9Sstevel@tonic-gate mutex_exit(pse); 5787c478bd9Sstevel@tonic-gate return (0); 5797c478bd9Sstevel@tonic-gate } 5807c478bd9Sstevel@tonic-gate 581db874c57Selowe /* 582db874c57Selowe * Variant of page_unlock() specifically for the page freelist 583db874c57Selowe * code. The mere existence of this code is a vile hack that 584db874c57Selowe * has resulted due to the backwards locking order of the page 585db874c57Selowe * freelist manager; please don't call it. 586db874c57Selowe */ 587db874c57Selowe void 588*8b464eb8Smec page_unlock_nocapture(page_t *pp) 589db874c57Selowe { 590db874c57Selowe kmutex_t *pse = PAGE_SE_MUTEX(pp); 591db874c57Selowe selock_t old; 592db874c57Selowe 593db874c57Selowe mutex_enter(pse); 594db874c57Selowe 595db874c57Selowe old = pp->p_selock; 596db874c57Selowe if ((old & ~SE_EWANTED) == SE_READER) { 597db874c57Selowe pp->p_selock = old & ~SE_READER; 598db874c57Selowe if (CV_HAS_WAITERS(&pp->p_cv)) 599db874c57Selowe cv_broadcast(&pp->p_cv); 600db874c57Selowe } else if ((old & ~SE_EWANTED) == SE_DELETED) { 601*8b464eb8Smec panic("page_unlock_nocapture: page %p is deleted", pp); 602db874c57Selowe } else if (old < 0) { 603db874c57Selowe THREAD_KPRI_RELEASE(); 604db874c57Selowe pp->p_selock &= SE_EWANTED; 605db874c57Selowe if (CV_HAS_WAITERS(&pp->p_cv)) 606db874c57Selowe cv_broadcast(&pp->p_cv); 607db874c57Selowe } else if ((old & ~SE_EWANTED) > SE_READER) { 608db874c57Selowe pp->p_selock = old - SE_READER; 609db874c57Selowe } else { 610*8b464eb8Smec panic("page_unlock_nocapture: page %p is not locked", pp); 611db874c57Selowe } 612db874c57Selowe 613db874c57Selowe mutex_exit(pse); 614db874c57Selowe } 615db874c57Selowe 6167c478bd9Sstevel@tonic-gate /* 6177c478bd9Sstevel@tonic-gate * Release the page's "shared/exclusive" lock and wake up anyone 6187c478bd9Sstevel@tonic-gate * who might be waiting for it. 6197c478bd9Sstevel@tonic-gate */ 6207c478bd9Sstevel@tonic-gate void 6217c478bd9Sstevel@tonic-gate page_unlock(page_t *pp) 6227c478bd9Sstevel@tonic-gate { 6237c478bd9Sstevel@tonic-gate kmutex_t *pse = PAGE_SE_MUTEX(pp); 6247c478bd9Sstevel@tonic-gate selock_t old; 6257c478bd9Sstevel@tonic-gate 6267c478bd9Sstevel@tonic-gate mutex_enter(pse); 627db874c57Selowe 6287c478bd9Sstevel@tonic-gate old = pp->p_selock; 6297c478bd9Sstevel@tonic-gate if ((old & ~SE_EWANTED) == SE_READER) { 6307c478bd9Sstevel@tonic-gate pp->p_selock = old & ~SE_READER; 6317c478bd9Sstevel@tonic-gate if (CV_HAS_WAITERS(&pp->p_cv)) 6327c478bd9Sstevel@tonic-gate cv_broadcast(&pp->p_cv); 6337c478bd9Sstevel@tonic-gate } else if ((old & ~SE_EWANTED) == SE_DELETED) { 6347c478bd9Sstevel@tonic-gate panic("page_unlock: page %p is deleted", pp); 6357c478bd9Sstevel@tonic-gate } else if (old < 0) { 6367c478bd9Sstevel@tonic-gate THREAD_KPRI_RELEASE(); 6377c478bd9Sstevel@tonic-gate pp->p_selock &= SE_EWANTED; 6387c478bd9Sstevel@tonic-gate if (CV_HAS_WAITERS(&pp->p_cv)) 6397c478bd9Sstevel@tonic-gate cv_broadcast(&pp->p_cv); 6407c478bd9Sstevel@tonic-gate } else if ((old & ~SE_EWANTED) > SE_READER) { 6417c478bd9Sstevel@tonic-gate pp->p_selock = old - SE_READER; 6427c478bd9Sstevel@tonic-gate } else { 6437c478bd9Sstevel@tonic-gate panic("page_unlock: page %p is not locked", pp); 6447c478bd9Sstevel@tonic-gate } 645db874c57Selowe 646*8b464eb8Smec if (pp->p_selock == 0) { 647db874c57Selowe /* 648*8b464eb8Smec * If the T_CAPTURING bit is set, that means that we should 649*8b464eb8Smec * not try and capture the page again as we could recurse 650*8b464eb8Smec * which could lead to a stack overflow panic or spending a 651*8b464eb8Smec * relatively long time in the kernel making no progress. 652db874c57Selowe */ 653*8b464eb8Smec if ((pp->p_toxic & PR_CAPTURE) && 654*8b464eb8Smec !(curthread->t_flag & T_CAPTURING) && 655*8b464eb8Smec !PP_RETIRED(pp)) { 656db874c57Selowe THREAD_KPRI_REQUEST(); 657db874c57Selowe pp->p_selock = SE_WRITER; 658db874c57Selowe mutex_exit(pse); 659*8b464eb8Smec page_unlock_capture(pp); 660db874c57Selowe } else { 661db874c57Selowe mutex_exit(pse); 662db874c57Selowe } 663db874c57Selowe } else { 664db874c57Selowe mutex_exit(pse); 665db874c57Selowe } 6667c478bd9Sstevel@tonic-gate } 6677c478bd9Sstevel@tonic-gate 6687c478bd9Sstevel@tonic-gate /* 6697c478bd9Sstevel@tonic-gate * Try to upgrade the lock on the page from a "shared" to an 6707c478bd9Sstevel@tonic-gate * "exclusive" lock. Since this upgrade operation is done while 6717c478bd9Sstevel@tonic-gate * holding the mutex protecting this page, no one else can acquire this page's 6727c478bd9Sstevel@tonic-gate * lock and change the page. Thus, it is safe to drop the "shared" 6737c478bd9Sstevel@tonic-gate * lock and attempt to acquire the "exclusive" lock. 6747c478bd9Sstevel@tonic-gate * 6757c478bd9Sstevel@tonic-gate * Returns 1 on success, 0 on failure. 6767c478bd9Sstevel@tonic-gate */ 6777c478bd9Sstevel@tonic-gate int 6787c478bd9Sstevel@tonic-gate page_tryupgrade(page_t *pp) 6797c478bd9Sstevel@tonic-gate { 6807c478bd9Sstevel@tonic-gate kmutex_t *pse = PAGE_SE_MUTEX(pp); 6817c478bd9Sstevel@tonic-gate 6827c478bd9Sstevel@tonic-gate mutex_enter(pse); 6837c478bd9Sstevel@tonic-gate if (!(pp->p_selock & SE_EWANTED)) { 6847c478bd9Sstevel@tonic-gate /* no threads want exclusive access, try upgrade */ 6857c478bd9Sstevel@tonic-gate if (pp->p_selock == SE_READER) { 6867c478bd9Sstevel@tonic-gate THREAD_KPRI_REQUEST(); 6877c478bd9Sstevel@tonic-gate /* convert to exclusive lock */ 6887c478bd9Sstevel@tonic-gate pp->p_selock = SE_WRITER; 6897c478bd9Sstevel@tonic-gate mutex_exit(pse); 6907c478bd9Sstevel@tonic-gate return (1); 6917c478bd9Sstevel@tonic-gate } 6927c478bd9Sstevel@tonic-gate } 6937c478bd9Sstevel@tonic-gate mutex_exit(pse); 6947c478bd9Sstevel@tonic-gate return (0); 6957c478bd9Sstevel@tonic-gate } 6967c478bd9Sstevel@tonic-gate 6977c478bd9Sstevel@tonic-gate /* 6987c478bd9Sstevel@tonic-gate * Downgrade the "exclusive" lock on the page to a "shared" lock 6997c478bd9Sstevel@tonic-gate * while holding the mutex protecting this page's p_selock field. 7007c478bd9Sstevel@tonic-gate */ 7017c478bd9Sstevel@tonic-gate void 7027c478bd9Sstevel@tonic-gate page_downgrade(page_t *pp) 7037c478bd9Sstevel@tonic-gate { 7047c478bd9Sstevel@tonic-gate kmutex_t *pse = PAGE_SE_MUTEX(pp); 7057c478bd9Sstevel@tonic-gate int excl_waiting; 7067c478bd9Sstevel@tonic-gate 7077c478bd9Sstevel@tonic-gate ASSERT((pp->p_selock & ~SE_EWANTED) != SE_DELETED); 7087c478bd9Sstevel@tonic-gate ASSERT(PAGE_EXCL(pp)); 7097c478bd9Sstevel@tonic-gate 7107c478bd9Sstevel@tonic-gate mutex_enter(pse); 7117c478bd9Sstevel@tonic-gate excl_waiting = pp->p_selock & SE_EWANTED; 7127c478bd9Sstevel@tonic-gate THREAD_KPRI_RELEASE(); 7137c478bd9Sstevel@tonic-gate pp->p_selock = SE_READER | excl_waiting; 7147c478bd9Sstevel@tonic-gate if (CV_HAS_WAITERS(&pp->p_cv)) 7157c478bd9Sstevel@tonic-gate cv_broadcast(&pp->p_cv); 7167c478bd9Sstevel@tonic-gate mutex_exit(pse); 7177c478bd9Sstevel@tonic-gate } 7187c478bd9Sstevel@tonic-gate 7197c478bd9Sstevel@tonic-gate void 7207c478bd9Sstevel@tonic-gate page_lock_delete(page_t *pp) 7217c478bd9Sstevel@tonic-gate { 7227c478bd9Sstevel@tonic-gate kmutex_t *pse = PAGE_SE_MUTEX(pp); 7237c478bd9Sstevel@tonic-gate 7247c478bd9Sstevel@tonic-gate ASSERT(PAGE_EXCL(pp)); 7257c478bd9Sstevel@tonic-gate ASSERT(pp->p_vnode == NULL); 7267c478bd9Sstevel@tonic-gate ASSERT(pp->p_offset == (u_offset_t)-1); 7277c478bd9Sstevel@tonic-gate ASSERT(!PP_ISFREE(pp)); 7287c478bd9Sstevel@tonic-gate 7297c478bd9Sstevel@tonic-gate mutex_enter(pse); 7307c478bd9Sstevel@tonic-gate THREAD_KPRI_RELEASE(); 7317c478bd9Sstevel@tonic-gate pp->p_selock = SE_DELETED; 7327c478bd9Sstevel@tonic-gate if (CV_HAS_WAITERS(&pp->p_cv)) 7337c478bd9Sstevel@tonic-gate cv_broadcast(&pp->p_cv); 7347c478bd9Sstevel@tonic-gate mutex_exit(pse); 7357c478bd9Sstevel@tonic-gate } 7367c478bd9Sstevel@tonic-gate 737*8b464eb8Smec int 738*8b464eb8Smec page_deleted(page_t *pp) 739*8b464eb8Smec { 740*8b464eb8Smec return (pp->p_selock == SE_DELETED); 741*8b464eb8Smec } 742*8b464eb8Smec 7437c478bd9Sstevel@tonic-gate /* 7447c478bd9Sstevel@tonic-gate * Implement the io lock for pages 7457c478bd9Sstevel@tonic-gate */ 7467c478bd9Sstevel@tonic-gate void 7477c478bd9Sstevel@tonic-gate page_iolock_init(page_t *pp) 7487c478bd9Sstevel@tonic-gate { 7497c478bd9Sstevel@tonic-gate pp->p_iolock_state = 0; 7507c478bd9Sstevel@tonic-gate cv_init(&pp->p_io_cv, NULL, CV_DEFAULT, NULL); 7517c478bd9Sstevel@tonic-gate } 7527c478bd9Sstevel@tonic-gate 7537c478bd9Sstevel@tonic-gate /* 7547c478bd9Sstevel@tonic-gate * Acquire the i/o lock on a page. 7557c478bd9Sstevel@tonic-gate */ 7567c478bd9Sstevel@tonic-gate void 7577c478bd9Sstevel@tonic-gate page_io_lock(page_t *pp) 7587c478bd9Sstevel@tonic-gate { 7597c478bd9Sstevel@tonic-gate kmutex_t *pio; 7607c478bd9Sstevel@tonic-gate 7617c478bd9Sstevel@tonic-gate pio = PAGE_IO_MUTEX(pp); 7627c478bd9Sstevel@tonic-gate mutex_enter(pio); 7637c478bd9Sstevel@tonic-gate while (pp->p_iolock_state & PAGE_IO_INUSE) { 7647c478bd9Sstevel@tonic-gate cv_wait(&(pp->p_io_cv), pio); 7657c478bd9Sstevel@tonic-gate } 7667c478bd9Sstevel@tonic-gate pp->p_iolock_state |= PAGE_IO_INUSE; 7677c478bd9Sstevel@tonic-gate mutex_exit(pio); 7687c478bd9Sstevel@tonic-gate } 7697c478bd9Sstevel@tonic-gate 7707c478bd9Sstevel@tonic-gate /* 7717c478bd9Sstevel@tonic-gate * Release the i/o lock on a page. 7727c478bd9Sstevel@tonic-gate */ 7737c478bd9Sstevel@tonic-gate void 7747c478bd9Sstevel@tonic-gate page_io_unlock(page_t *pp) 7757c478bd9Sstevel@tonic-gate { 7767c478bd9Sstevel@tonic-gate kmutex_t *pio; 7777c478bd9Sstevel@tonic-gate 7787c478bd9Sstevel@tonic-gate pio = PAGE_IO_MUTEX(pp); 7797c478bd9Sstevel@tonic-gate mutex_enter(pio); 780a71e32b6Sstans cv_broadcast(&pp->p_io_cv); 7817c478bd9Sstevel@tonic-gate pp->p_iolock_state &= ~PAGE_IO_INUSE; 7827c478bd9Sstevel@tonic-gate mutex_exit(pio); 7837c478bd9Sstevel@tonic-gate } 7847c478bd9Sstevel@tonic-gate 7857c478bd9Sstevel@tonic-gate /* 7867c478bd9Sstevel@tonic-gate * Try to acquire the i/o lock on a page without blocking. 7877c478bd9Sstevel@tonic-gate * Returns 1 on success, 0 on failure. 7887c478bd9Sstevel@tonic-gate */ 7897c478bd9Sstevel@tonic-gate int 7907c478bd9Sstevel@tonic-gate page_io_trylock(page_t *pp) 7917c478bd9Sstevel@tonic-gate { 7927c478bd9Sstevel@tonic-gate kmutex_t *pio; 7937c478bd9Sstevel@tonic-gate 7947c478bd9Sstevel@tonic-gate if (pp->p_iolock_state & PAGE_IO_INUSE) 7957c478bd9Sstevel@tonic-gate return (0); 7967c478bd9Sstevel@tonic-gate 7977c478bd9Sstevel@tonic-gate pio = PAGE_IO_MUTEX(pp); 7987c478bd9Sstevel@tonic-gate mutex_enter(pio); 7997c478bd9Sstevel@tonic-gate 8007c478bd9Sstevel@tonic-gate if (pp->p_iolock_state & PAGE_IO_INUSE) { 8017c478bd9Sstevel@tonic-gate mutex_exit(pio); 8027c478bd9Sstevel@tonic-gate return (0); 8037c478bd9Sstevel@tonic-gate } 8047c478bd9Sstevel@tonic-gate pp->p_iolock_state |= PAGE_IO_INUSE; 8057c478bd9Sstevel@tonic-gate mutex_exit(pio); 8067c478bd9Sstevel@tonic-gate 8077c478bd9Sstevel@tonic-gate return (1); 8087c478bd9Sstevel@tonic-gate } 8097c478bd9Sstevel@tonic-gate 810a71e32b6Sstans /* 811a71e32b6Sstans * Wait until the i/o lock is not held. 812a71e32b6Sstans */ 813a71e32b6Sstans void 814a71e32b6Sstans page_io_wait(page_t *pp) 815a71e32b6Sstans { 816a71e32b6Sstans kmutex_t *pio; 817a71e32b6Sstans 818a71e32b6Sstans pio = PAGE_IO_MUTEX(pp); 819a71e32b6Sstans mutex_enter(pio); 820a71e32b6Sstans while (pp->p_iolock_state & PAGE_IO_INUSE) { 821a71e32b6Sstans cv_wait(&(pp->p_io_cv), pio); 822a71e32b6Sstans } 823a71e32b6Sstans mutex_exit(pio); 824a71e32b6Sstans } 825a71e32b6Sstans 826a71e32b6Sstans /* 827a71e32b6Sstans * Returns 1 on success, 0 on failure. 828a71e32b6Sstans */ 829a71e32b6Sstans int 830a71e32b6Sstans page_io_locked(page_t *pp) 831a71e32b6Sstans { 832a71e32b6Sstans return (pp->p_iolock_state & PAGE_IO_INUSE); 833a71e32b6Sstans } 834a71e32b6Sstans 8357c478bd9Sstevel@tonic-gate /* 8367c478bd9Sstevel@tonic-gate * Assert that the i/o lock on a page is held. 8377c478bd9Sstevel@tonic-gate * Returns 1 on success, 0 on failure. 8387c478bd9Sstevel@tonic-gate */ 8397c478bd9Sstevel@tonic-gate int 8407c478bd9Sstevel@tonic-gate page_iolock_assert(page_t *pp) 8417c478bd9Sstevel@tonic-gate { 842a71e32b6Sstans return (page_io_locked(pp)); 8437c478bd9Sstevel@tonic-gate } 8447c478bd9Sstevel@tonic-gate 8457c478bd9Sstevel@tonic-gate /* 8467c478bd9Sstevel@tonic-gate * Wrapper exported to kernel routines that are built 8477c478bd9Sstevel@tonic-gate * platform-independent (the macro is platform-dependent; 8487c478bd9Sstevel@tonic-gate * the size of vph_mutex[] is based on NCPU). 8497c478bd9Sstevel@tonic-gate * 8507c478bd9Sstevel@tonic-gate * Note that you can do stress testing on this by setting the 8517c478bd9Sstevel@tonic-gate * variable page_vnode_mutex_stress to something other than 8527c478bd9Sstevel@tonic-gate * zero in a DEBUG kernel in a debugger after loading the kernel. 8537c478bd9Sstevel@tonic-gate * Setting it after the kernel is running may not work correctly. 8547c478bd9Sstevel@tonic-gate */ 8557c478bd9Sstevel@tonic-gate #ifdef DEBUG 8567c478bd9Sstevel@tonic-gate static int page_vnode_mutex_stress = 0; 8577c478bd9Sstevel@tonic-gate #endif 8587c478bd9Sstevel@tonic-gate 8597c478bd9Sstevel@tonic-gate kmutex_t * 8607c478bd9Sstevel@tonic-gate page_vnode_mutex(vnode_t *vp) 8617c478bd9Sstevel@tonic-gate { 8627c478bd9Sstevel@tonic-gate if (vp == &kvp) 8637c478bd9Sstevel@tonic-gate return (&vph_mutex[VPH_TABLE_SIZE + 0]); 8647c478bd9Sstevel@tonic-gate #ifdef DEBUG 8657c478bd9Sstevel@tonic-gate if (page_vnode_mutex_stress != 0) 8667c478bd9Sstevel@tonic-gate return (&vph_mutex[0]); 8677c478bd9Sstevel@tonic-gate #endif 8687c478bd9Sstevel@tonic-gate 8697c478bd9Sstevel@tonic-gate return (&vph_mutex[VP_HASH_FUNC(vp)]); 8707c478bd9Sstevel@tonic-gate } 8717c478bd9Sstevel@tonic-gate 8727c478bd9Sstevel@tonic-gate kmutex_t * 8737c478bd9Sstevel@tonic-gate page_se_mutex(page_t *pp) 8747c478bd9Sstevel@tonic-gate { 8757c478bd9Sstevel@tonic-gate return (PAGE_SE_MUTEX(pp)); 8767c478bd9Sstevel@tonic-gate } 8777c478bd9Sstevel@tonic-gate 8787c478bd9Sstevel@tonic-gate #ifdef VM_STATS 8797c478bd9Sstevel@tonic-gate uint_t pszclck_stat[4]; 8807c478bd9Sstevel@tonic-gate #endif 8817c478bd9Sstevel@tonic-gate /* 8827c478bd9Sstevel@tonic-gate * Find, take and return a mutex held by hat_page_demote(). 8837c478bd9Sstevel@tonic-gate * Called by page_demote_vp_pages() before hat_page_demote() call and by 8847c478bd9Sstevel@tonic-gate * routines that want to block hat_page_demote() but can't do it 8857c478bd9Sstevel@tonic-gate * via locking all constituent pages. 8867c478bd9Sstevel@tonic-gate * 8877c478bd9Sstevel@tonic-gate * Return NULL if p_szc is 0. 8887c478bd9Sstevel@tonic-gate * 8897c478bd9Sstevel@tonic-gate * It should only be used for pages that can be demoted by hat_page_demote() 8907c478bd9Sstevel@tonic-gate * i.e. non swapfs file system pages. The logic here is lifted from 8917c478bd9Sstevel@tonic-gate * sfmmu_mlspl_enter() except there's no need to worry about p_szc increase 8927c478bd9Sstevel@tonic-gate * since the page is locked and not free. 8937c478bd9Sstevel@tonic-gate * 8947c478bd9Sstevel@tonic-gate * Hash of the root page is used to find the lock. 8957c478bd9Sstevel@tonic-gate * To find the root in the presense of hat_page_demote() chageing the location 8967c478bd9Sstevel@tonic-gate * of the root this routine relies on the fact that hat_page_demote() changes 8977c478bd9Sstevel@tonic-gate * root last. 8987c478bd9Sstevel@tonic-gate * 8997c478bd9Sstevel@tonic-gate * If NULL is returned pp's p_szc is guaranteed to be 0. If non NULL is 9007c478bd9Sstevel@tonic-gate * returned pp's p_szc may be any value. 9017c478bd9Sstevel@tonic-gate */ 9027c478bd9Sstevel@tonic-gate kmutex_t * 9037c478bd9Sstevel@tonic-gate page_szc_lock(page_t *pp) 9047c478bd9Sstevel@tonic-gate { 9057c478bd9Sstevel@tonic-gate kmutex_t *mtx; 9067c478bd9Sstevel@tonic-gate page_t *rootpp; 9077c478bd9Sstevel@tonic-gate uint_t szc; 9087c478bd9Sstevel@tonic-gate uint_t rszc; 9097c478bd9Sstevel@tonic-gate uint_t pszc = pp->p_szc; 9107c478bd9Sstevel@tonic-gate 9117c478bd9Sstevel@tonic-gate ASSERT(pp != NULL); 9127c478bd9Sstevel@tonic-gate ASSERT(PAGE_LOCKED(pp)); 9137c478bd9Sstevel@tonic-gate ASSERT(!PP_ISFREE(pp)); 9147c478bd9Sstevel@tonic-gate ASSERT(pp->p_vnode != NULL); 9157c478bd9Sstevel@tonic-gate ASSERT(!IS_SWAPFSVP(pp->p_vnode)); 9167c478bd9Sstevel@tonic-gate ASSERT(pp->p_vnode != &kvp); 9177c478bd9Sstevel@tonic-gate 9187c478bd9Sstevel@tonic-gate again: 9197c478bd9Sstevel@tonic-gate if (pszc == 0) { 9207c478bd9Sstevel@tonic-gate VM_STAT_ADD(pszclck_stat[0]); 9217c478bd9Sstevel@tonic-gate return (NULL); 9227c478bd9Sstevel@tonic-gate } 9237c478bd9Sstevel@tonic-gate 9247c478bd9Sstevel@tonic-gate /* The lock lives in the root page */ 9257c478bd9Sstevel@tonic-gate 9267c478bd9Sstevel@tonic-gate rootpp = PP_GROUPLEADER(pp, pszc); 9277c478bd9Sstevel@tonic-gate mtx = PAGE_SZC_MUTEX(rootpp); 9287c478bd9Sstevel@tonic-gate mutex_enter(mtx); 9297c478bd9Sstevel@tonic-gate 9307c478bd9Sstevel@tonic-gate /* 9317c478bd9Sstevel@tonic-gate * since p_szc can only decrease if pp == rootpp 9327c478bd9Sstevel@tonic-gate * rootpp will be always the same i.e we have the right root 9337c478bd9Sstevel@tonic-gate * regardless of rootpp->p_szc. 9347c478bd9Sstevel@tonic-gate * If location of pp's root didn't change after we took 9357c478bd9Sstevel@tonic-gate * the lock we have the right root. return mutex hashed off it. 9367c478bd9Sstevel@tonic-gate */ 9377c478bd9Sstevel@tonic-gate if (pp == rootpp || (rszc = rootpp->p_szc) == pszc) { 9387c478bd9Sstevel@tonic-gate VM_STAT_ADD(pszclck_stat[1]); 9397c478bd9Sstevel@tonic-gate return (mtx); 9407c478bd9Sstevel@tonic-gate } 9417c478bd9Sstevel@tonic-gate 9427c478bd9Sstevel@tonic-gate /* 9437c478bd9Sstevel@tonic-gate * root location changed because page got demoted. 9447c478bd9Sstevel@tonic-gate * locate the new root. 9457c478bd9Sstevel@tonic-gate */ 9467c478bd9Sstevel@tonic-gate if (rszc < pszc) { 9477c478bd9Sstevel@tonic-gate szc = pp->p_szc; 9487c478bd9Sstevel@tonic-gate ASSERT(szc < pszc); 9497c478bd9Sstevel@tonic-gate mutex_exit(mtx); 9507c478bd9Sstevel@tonic-gate pszc = szc; 9517c478bd9Sstevel@tonic-gate VM_STAT_ADD(pszclck_stat[2]); 9527c478bd9Sstevel@tonic-gate goto again; 9537c478bd9Sstevel@tonic-gate } 9547c478bd9Sstevel@tonic-gate 9557c478bd9Sstevel@tonic-gate VM_STAT_ADD(pszclck_stat[3]); 9567c478bd9Sstevel@tonic-gate /* 9577c478bd9Sstevel@tonic-gate * current hat_page_demote not done yet. 9587c478bd9Sstevel@tonic-gate * wait for it to finish. 9597c478bd9Sstevel@tonic-gate */ 9607c478bd9Sstevel@tonic-gate mutex_exit(mtx); 9617c478bd9Sstevel@tonic-gate rootpp = PP_GROUPLEADER(rootpp, rszc); 9627c478bd9Sstevel@tonic-gate mtx = PAGE_SZC_MUTEX(rootpp); 9637c478bd9Sstevel@tonic-gate mutex_enter(mtx); 9647c478bd9Sstevel@tonic-gate mutex_exit(mtx); 9657c478bd9Sstevel@tonic-gate ASSERT(rootpp->p_szc < rszc); 9667c478bd9Sstevel@tonic-gate goto again; 9677c478bd9Sstevel@tonic-gate } 9687c478bd9Sstevel@tonic-gate 9697c478bd9Sstevel@tonic-gate int 9707c478bd9Sstevel@tonic-gate page_szc_lock_assert(page_t *pp) 9717c478bd9Sstevel@tonic-gate { 9727c478bd9Sstevel@tonic-gate page_t *rootpp = PP_PAGEROOT(pp); 9737c478bd9Sstevel@tonic-gate kmutex_t *mtx = PAGE_SZC_MUTEX(rootpp); 9747c478bd9Sstevel@tonic-gate 9757c478bd9Sstevel@tonic-gate return (MUTEX_HELD(mtx)); 9767c478bd9Sstevel@tonic-gate } 977