Please note that diffs are not public domain; they are subject to the copyright notices on the relevant files. =================================================================== RCS file: /ftp/cvs/cvsroot/src/sys/kern/subr_pool.c,v rcsdiff: /ftp/cvs/cvsroot/src/sys/kern/subr_pool.c,v: warning: Unknown phrases like `commitid ...;' are present. retrieving revision 1.137.2.5 retrieving revision 1.182 diff -u -p -r1.137.2.5 -r1.182 --- src/sys/kern/subr_pool.c 2007/12/26 21:39:42 1.137.2.5 +++ src/sys/kern/subr_pool.c 2010/01/20 23:40:42 1.182 @@ -1,7 +1,7 @@ -/* $NetBSD: subr_pool.c,v 1.137.2.5 2007/12/26 21:39:42 ad Exp $ */ +/* $NetBSD: subr_pool.c,v 1.182 2010/01/20 23:40:42 rmind Exp $ */ /*- - * Copyright (c) 1997, 1999, 2000, 2002, 2007 The NetBSD Foundation, Inc. + * Copyright (c) 1997, 1999, 2000, 2002, 2007, 2008 The NetBSD Foundation, Inc. * All rights reserved. * * This code is derived from software contributed to The NetBSD Foundation @@ -16,13 +16,6 @@ * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. - * 3. All advertising materials mentioning features or use of this software - * must display the following acknowledgement: - * This product includes software developed by the NetBSD - * Foundation, Inc. and its contributors. - * 4. Neither the name of The NetBSD Foundation nor the names of its - * contributors may be used to endorse or promote products derived - * from this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED @@ -38,7 +31,7 @@ */ #include -__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.137.2.5 2007/12/26 21:39:42 ad Exp $"); +__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.182 2010/01/20 23:40:42 rmind Exp $"); #include "opt_ddb.h" #include "opt_pool.h" @@ -52,7 +45,6 @@ __KERNEL_RCSID(0, "$NetBSD: subr_pool.c, #include #include #include -#include #include #include #include @@ -77,7 +69,7 @@ __KERNEL_RCSID(0, "$NetBSD: subr_pool.c, */ /* List of all pools */ -TAILQ_HEAD(,pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head); +static TAILQ_HEAD(, pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head); /* Private pool for page header structures */ #define PHPOOL_MAX 8 @@ -112,6 +104,9 @@ static struct pool *drainpp; static kmutex_t pool_head_lock; static kcondvar_t pool_busy; +/* This lock protects initialization of a potentially shared pool allocator */ +static kmutex_t pool_allocator_lock; + typedef uint32_t pool_item_bitmap_t; #define BITMAP_SIZE (CHAR_BIT * sizeof(pool_item_bitmap_t)) #define BITMAP_MASK (BITMAP_SIZE - 1) @@ -123,7 +118,7 @@ struct pool_item_header { SPLAY_ENTRY(pool_item_header) ph_node; /* Off-page page headers */ void * ph_page; /* this page's address */ - struct timeval ph_time; /* last referenced */ + uint32_t ph_time; /* last referenced */ uint16_t ph_nmissing; /* # of chunks in use */ uint16_t ph_off; /* start offset in page */ union { @@ -187,15 +182,16 @@ static struct pool cache_cpu_pool; TAILQ_HEAD(,pool_cache) pool_cache_head = TAILQ_HEAD_INITIALIZER(pool_cache_head); -int pool_cache_disable; - +int pool_cache_disable; /* global disable for caching */ +static const pcg_t pcg_dummy; /* zero sized: always empty, yet always full */ -static pool_cache_cpu_t *pool_cache_put_slow(pool_cache_cpu_t *, int *, - void *, paddr_t); -static pool_cache_cpu_t *pool_cache_get_slow(pool_cache_cpu_t *, int *, - void **, paddr_t *, int); +static bool pool_cache_put_slow(pool_cache_cpu_t *, int, + void *); +static bool pool_cache_get_slow(pool_cache_cpu_t *, int, + void **, paddr_t *, int); static void pool_cache_cpu_init1(struct cpu_info *, pool_cache_t); static void pool_cache_invalidate_groups(pool_cache_t, pcg_t *); +static void pool_cache_invalidate_cpu(pool_cache_t, u_int); static void pool_cache_xcall(pool_cache_t); static int pool_catchup(struct pool *); @@ -238,16 +234,28 @@ int pool_logsize = POOL_LOGSIZE; static inline void pr_log(struct pool *pp, void *v, int action, const char *file, long line) { - int n = pp->pr_curlogentry; + int n; struct pool_log *pl; if ((pp->pr_roflags & PR_LOGGING) == 0) return; + if (pp->pr_log == NULL) { + if (kmem_map != NULL) + pp->pr_log = malloc( + pool_logsize * sizeof(struct pool_log), + M_TEMP, M_NOWAIT | M_ZERO); + if (pp->pr_log == NULL) + return; + pp->pr_curlogentry = 0; + pp->pr_logsize = pool_logsize; + } + /* * Fill in the current entry. Wrap around and overwrite * the oldest entry if necessary. */ + n = pp->pr_curlogentry; pl = &pp->pr_log[n]; pl->pl_file = file; pl->pl_line = line; @@ -265,7 +273,7 @@ pr_printlog(struct pool *pp, struct pool int i = pp->pr_logsize; int n = pp->pr_curlogentry; - if ((pp->pr_roflags & PR_LOGGING) == 0) + if (pp->pr_log == NULL) return; /* @@ -595,17 +603,11 @@ void pool_subsystem_init(void) { struct pool_allocator *pa; - __link_set_decl(pools, struct link_pool_init); - struct link_pool_init * const *pi; mutex_init(&pool_head_lock, MUTEX_DEFAULT, IPL_NONE); + mutex_init(&pool_allocator_lock, MUTEX_DEFAULT, IPL_NONE); cv_init(&pool_busy, "poolbusy"); - __link_set_foreach(pi, pools) - pool_init((*pi)->pp, (*pi)->size, (*pi)->align, - (*pi)->align_offset, (*pi)->flags, (*pi)->wchan, - (*pi)->palloc, (*pi)->ipl); - while ((pa = SLIST_FIRST(&pa_deferinitq)) != NULL) { KASSERT(pa->pa_backingmapptr != NULL); KASSERT(*pa->pa_backingmapptr != NULL); @@ -613,10 +615,10 @@ pool_subsystem_init(void) pa_reclaim_register(pa); } - pool_init(&cache_pool, sizeof(struct pool_cache), CACHE_LINE_SIZE, + pool_init(&cache_pool, sizeof(struct pool_cache), coherency_unit, 0, 0, "pcache", &pool_allocator_nointr, IPL_NONE); - pool_init(&cache_cpu_pool, sizeof(pool_cache_cpu_t), CACHE_LINE_SIZE, + pool_init(&cache_cpu_pool, sizeof(pool_cache_cpu_t), coherency_unit, 0, 0, "pcachecpu", &pool_allocator_nointr, IPL_NONE); } @@ -664,7 +666,9 @@ pool_init(struct pool *pp, size_t size, palloc = &pool_allocator_nointr_fullpage; } #endif /* POOL_SUBPAGE */ - if ((palloc->pa_flags & PA_INITIALIZED) == 0) { + if (!cold) + mutex_enter(&pool_allocator_lock); + if (palloc->pa_refcnt++ == 0) { if (palloc->pa_pagesz == 0) palloc->pa_pagesz = PAGE_SIZE; @@ -677,8 +681,9 @@ pool_init(struct pool *pp, size_t size, if (palloc->pa_backingmapptr != NULL) { pa_reclaim_register(palloc); } - palloc->pa_flags |= PA_INITIALIZED; } + if (!cold) + mutex_exit(&pool_allocator_lock); if (align == 0) align = ALIGN(1); @@ -801,28 +806,12 @@ pool_init(struct pool *pp, size_t size, pp->pr_nidle = 0; pp->pr_refcnt = 0; -#ifdef POOL_DIAGNOSTIC - if (flags & PR_LOGGING) { - if (kmem_map == NULL || - (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log), - M_TEMP, M_NOWAIT)) == NULL) - pp->pr_roflags &= ~PR_LOGGING; - pp->pr_curlogentry = 0; - pp->pr_logsize = pool_logsize; - } -#endif + pp->pr_log = NULL; pp->pr_entered_file = NULL; pp->pr_entered_line = 0; - /* - * XXXAD hack to prevent IP input processing from blocking. - */ - if (ipl == IPL_SOFTNET) { - mutex_init(&pp->pr_lock, MUTEX_DEFAULT, IPL_VM); - } else { - mutex_init(&pp->pr_lock, MUTEX_DEFAULT, ipl); - } + mutex_init(&pp->pr_lock, MUTEX_DEFAULT, ipl); cv_init(&pp->pr_cv, wchan); pp->pr_ipl = ipl; @@ -856,17 +845,17 @@ pool_init(struct pool *pp, size_t size, size = sizeof(pcg_t) + (PCG_NOBJECTS_NORMAL - 1) * sizeof(pcgpair_t); - pool_init(&pcg_normal_pool, size, CACHE_LINE_SIZE, 0, 0, + pool_init(&pcg_normal_pool, size, coherency_unit, 0, 0, "pcgnormal", &pool_allocator_meta, IPL_VM); size = sizeof(pcg_t) + (PCG_NOBJECTS_LARGE - 1) * sizeof(pcgpair_t); - pool_init(&pcg_large_pool, size, CACHE_LINE_SIZE, 0, 0, + pool_init(&pcg_large_pool, size, coherency_unit, 0, 0, "pcglarge", &pool_allocator_meta, IPL_VM); } /* Insert into the list of all pools. */ - if (__predict_true(!cold)) + if (!cold) mutex_enter(&pool_head_lock); TAILQ_FOREACH(pp1, &pool_head, pr_poollist) { if (strcmp(pp1->pr_wchan, pp->pr_wchan) > 0) @@ -876,14 +865,14 @@ pool_init(struct pool *pp, size_t size, TAILQ_INSERT_TAIL(&pool_head, pp, pr_poollist); else TAILQ_INSERT_BEFORE(pp1, pp, pr_poollist); - if (__predict_true(!cold)) + if (!cold) mutex_exit(&pool_head_lock); - /* Insert this into the list of pools using this allocator. */ - if (__predict_true(!cold)) + /* Insert this into the list of pools using this allocator. */ + if (!cold) mutex_enter(&palloc->pa_lock); TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list); - if (__predict_true(!cold)) + if (!cold) mutex_exit(&palloc->pa_lock); pool_reclaim_register(pp); @@ -913,6 +902,11 @@ pool_destroy(struct pool *pp) TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list); mutex_exit(&pp->pr_alloc->pa_lock); + mutex_enter(&pool_allocator_lock); + if (--pp->pr_alloc->pa_refcnt == 0) + mutex_destroy(&pp->pr_alloc->pa_lock); + mutex_exit(&pool_allocator_lock); + mutex_enter(&pp->pr_lock); KASSERT(pp->pr_cache == NULL); @@ -938,8 +932,10 @@ pool_destroy(struct pool *pp) pr_pagelist_free(pp, &pq); #ifdef POOL_DIAGNOSTIC - if ((pp->pr_roflags & PR_LOGGING) != 0) + if (pp->pr_log != NULL) { free(pp->pr_log, M_TEMP); + pp->pr_log = NULL; + } #endif cv_destroy(&pp->pr_cv); @@ -996,8 +992,9 @@ pool_get(struct pool *pp, int flags) #endif /* DIAGNOSTIC */ #ifdef LOCKDEBUG - if (flags & PR_WAITOK) - ASSERT_SLEEPABLE(NULL, "pool_get(PR_WAITOK)"); + if (flags & PR_WAITOK) { + ASSERT_SLEEPABLE(); + } #endif mutex_enter(&pp->pr_lock); @@ -1267,10 +1264,7 @@ pool_do_put(struct pool *pp, void *v, st if (pp->pr_flags & PR_WANTED) { pp->pr_flags &= ~PR_WANTED; - if (ph->ph_nmissing == 0) - pp->pr_nidle++; cv_broadcast(&pp->pr_cv); - return; } /* @@ -1289,8 +1283,7 @@ pool_do_put(struct pool *pp, void *v, st if (ph->ph_nmissing == 0) { pp->pr_nidle++; if (pp->pr_npages > pp->pr_minpages && - (pp->pr_npages > pp->pr_maxpages || - pa_starved_p(pp->pr_alloc))) { + pp->pr_npages > pp->pr_maxpages) { pr_rmpage(pp, ph, pq); } else { LIST_REMOVE(ph, ph_pagelist); @@ -1301,8 +1294,11 @@ pool_do_put(struct pool *pp, void *v, st * be idle for some period of time before it can * be reclaimed by the pagedaemon. This minimizes * ping-pong'ing for memory. + * + * note for 64-bit time_t: truncating to 32-bit is not + * a problem for our usage. */ - getmicrotime(&ph->ph_time); + ph->ph_time = time_uptime; } pool_update_curpage(pp); } @@ -1454,7 +1450,7 @@ pool_prime_page(struct pool *pp, void *s LIST_INIT(&ph->ph_itemlist); ph->ph_page = storage; ph->ph_nmissing = 0; - getmicrotime(&ph->ph_time); + ph->ph_time = time_uptime; if ((pp->pr_roflags & PR_PHINPAGE) == 0) SPLAY_INSERT(phtree, &pp->pr_phtree, ph); @@ -1542,6 +1538,8 @@ pool_update_curpage(struct pool *pp) if (pp->pr_curpage == NULL) { pp->pr_curpage = LIST_FIRST(&pp->pr_emptypages); } + KASSERT((pp->pr_curpage == NULL && pp->pr_nitems == 0) || + (pp->pr_curpage != NULL && pp->pr_nitems > 0)); } void @@ -1615,7 +1613,7 @@ pool_reclaim(struct pool *pp) { struct pool_item_header *ph, *phnext; struct pool_pagelist pq; - struct timeval curtime, diff; + uint32_t curtime; bool klock; int rv; @@ -1627,9 +1625,8 @@ pool_reclaim(struct pool *pp) } /* - * XXXSMP Because mutexes at IPL_SOFTXXX are still spinlocks, - * and we are called from the pagedaemon without kernel_lock. - * Does not apply to IPL_SOFTBIO. + * XXXSMP Because we do not want to cause non-MPSAFE code + * to block. */ if (pp->pr_ipl == IPL_SOFTNET || pp->pr_ipl == IPL_SOFTCLOCK || pp->pr_ipl == IPL_SOFTSERIAL) { @@ -1652,7 +1649,7 @@ pool_reclaim(struct pool *pp) LIST_INIT(&pq); - getmicrotime(&curtime); + curtime = time_uptime; for (ph = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) { phnext = LIST_NEXT(ph, ph_pagelist); @@ -1662,8 +1659,7 @@ pool_reclaim(struct pool *pp) break; KASSERT(ph->ph_nmissing == 0); - timersub(&curtime, &ph->ph_time, &diff); - if (diff.tv_sec < pool_inactive_time + if (curtime - ph->ph_time < pool_inactive_time && !pa_starved_p(pp->pr_alloc)) continue; @@ -1804,10 +1800,8 @@ pool_print_pagelist(struct pool *pp, str #endif LIST_FOREACH(ph, pl, ph_pagelist) { - (*pr)("\t\tpage %p, nmissing %d, time %lu,%lu\n", - ph->ph_page, ph->ph_nmissing, - (u_long)ph->ph_time.tv_sec, - (u_long)ph->ph_time.tv_usec); + (*pr)("\t\tpage %p, nmissing %d, time %" PRIu32 "\n", + ph->ph_page, ph->ph_nmissing, ph->ph_time); #ifdef DIAGNOSTIC if (!(pp->pr_roflags & PR_NOTOUCH)) { LIST_FOREACH(pi, &ph->ph_itemlist, pi_list) { @@ -2075,15 +2069,7 @@ pool_cache_bootstrap(pool_cache_t pc, si if (palloc == NULL && ipl == IPL_NONE) palloc = &pool_allocator_nointr; pool_init(pp, size, align, align_offset, flags, wchan, palloc, ipl); - - /* - * XXXAD hack to prevent IP input processing from blocking. - */ - if (ipl == IPL_SOFTNET) { - mutex_init(&pc->pc_lock, MUTEX_DEFAULT, IPL_VM); - } else { - mutex_init(&pc->pc_lock, MUTEX_DEFAULT, ipl); - } + mutex_init(&pc->pc_lock, MUTEX_DEFAULT, ipl); if (ctor == NULL) { ctor = (int (*)(void *, void *, int))nullop; @@ -2109,8 +2095,10 @@ pool_cache_bootstrap(pool_cache_t pc, si if ((flags & PR_LARGECACHE) != 0) { pc->pc_pcgsize = PCG_NOBJECTS_LARGE; + pc->pc_pcgpool = &pcg_large_pool; } else { pc->pc_pcgsize = PCG_NOBJECTS_NORMAL; + pc->pc_pcgpool = &pcg_normal_pool; } /* Allocate per-CPU caches. */ @@ -2152,9 +2140,7 @@ void pool_cache_destroy(pool_cache_t pc) { struct pool *pp = &pc->pc_pool; - pool_cache_cpu_t *cc; - pcg_t *pcg; - int i; + u_int i; /* Remove it from the global list. */ mutex_enter(&pool_head_lock); @@ -2172,20 +2158,8 @@ pool_cache_destroy(pool_cache_t pc) mutex_exit(&pp->pr_lock); /* Destroy per-CPU data */ - for (i = 0; i < MAXCPUS; i++) { - if ((cc = pc->pc_cpus[i]) == NULL) - continue; - if ((pcg = cc->cc_current) != NULL) { - pcg->pcg_next = NULL; - pool_cache_invalidate_groups(pc, pcg); - } - if ((pcg = cc->cc_previous) != NULL) { - pcg->pcg_next = NULL; - pool_cache_invalidate_groups(pc, pcg); - } - if (cc != &pc->pc_cpu0) - pool_put(&cache_cpu_pool, cc); - } + for (i = 0; i < MAXCPUS; i++) + pool_cache_invalidate_cpu(pc, i); /* Finally, destroy it. */ mutex_destroy(&pc->pc_lock); @@ -2207,7 +2181,6 @@ pool_cache_cpu_init1(struct cpu_info *ci index = ci->ci_index; KASSERT(index < MAXCPUS); - KASSERT(((uintptr_t)pc->pc_cpus & (CACHE_LINE_SIZE - 1)) == 0); if ((cc = pc->pc_cpus[index]) != NULL) { KASSERT(cc->cc_cpuindex == index); @@ -2234,8 +2207,8 @@ pool_cache_cpu_init1(struct cpu_info *ci cc->cc_cpuindex = index; cc->cc_hits = 0; cc->cc_misses = 0; - cc->cc_current = NULL; - cc->cc_previous = NULL; + cc->cc_current = __UNCONST(&pcg_dummy); + cc->cc_previous = __UNCONST(&pcg_dummy); pc->pc_cpus[index] = cc; } @@ -2333,12 +2306,35 @@ pool_cache_invalidate_groups(pool_cache_ * * Invalidate a pool cache (destruct and release all of the * cached objects). Does not reclaim objects from the pool. + * + * Note: For pool caches that provide constructed objects, there + * is an assumption that another level of synchronization is occurring + * between the input to the constructor and the cache invalidation. */ void pool_cache_invalidate(pool_cache_t pc) { pcg_t *full, *empty, *part; +#if 0 + uint64_t where; + if (ncpu < 2 || !mp_online) { + /* + * We might be called early enough in the boot process + * for the CPU data structures to not be fully initialized. + * In this case, simply gather the local CPU's cache now + * since it will be the only one running. + */ + pool_cache_xcall(pc); + } else { + /* + * Gather all of the CPU-specific caches into the + * global cache. + */ + where = xc_broadcast(0, (xcfunc_t)pool_cache_xcall, pc, NULL); + xc_wait(where); + } +#endif mutex_enter(&pc->pc_lock); full = pc->pc_fullgroups; empty = pc->pc_emptygroups; @@ -2356,6 +2352,40 @@ pool_cache_invalidate(pool_cache_t pc) pool_cache_invalidate_groups(pc, part); } +/* + * pool_cache_invalidate_cpu: + * + * Invalidate all CPU-bound cached objects in pool cache, the CPU being + * identified by its associated index. + * It is caller's responsibility to ensure that no operation is + * taking place on this pool cache while doing this invalidation. + * WARNING: as no inter-CPU locking is enforced, trying to invalidate + * pool cached objects from a CPU different from the one currently running + * may result in an undefined behaviour. + */ +static void +pool_cache_invalidate_cpu(pool_cache_t pc, u_int index) +{ + + pool_cache_cpu_t *cc; + pcg_t *pcg; + + if ((cc = pc->pc_cpus[index]) == NULL) + return; + + if ((pcg = cc->cc_current) != &pcg_dummy) { + pcg->pcg_next = NULL; + pool_cache_invalidate_groups(pc, pcg); + } + if ((pcg = cc->cc_previous) != &pcg_dummy) { + pcg->pcg_next = NULL; + pool_cache_invalidate_groups(pc, pcg); + } + if (cc != &pc->pc_cpu0) + pool_put(&cache_cpu_pool, cc); + +} + void pool_cache_set_drain_hook(pool_cache_t pc, void (*fn)(void *, int), void *arg) { @@ -2384,43 +2414,8 @@ pool_cache_sethardlimit(pool_cache_t pc, pool_sethardlimit(&pc->pc_pool, n, warnmess, ratecap); } -static inline pool_cache_cpu_t * -pool_cache_cpu_enter(pool_cache_t pc, int *s) -{ - pool_cache_cpu_t *cc; - - /* - * Prevent other users of the cache from accessing our - * CPU-local data. To avoid touching shared state, we - * pull the neccessary information from CPU local data. - */ - crit_enter(); - cc = pc->pc_cpus[curcpu()->ci_index]; - KASSERT(cc->cc_cache == pc); - if (cc->cc_ipl != IPL_NONE) { - *s = splraiseipl(cc->cc_iplcookie); - } - KASSERT(((uintptr_t)cc & (CACHE_LINE_SIZE - 1)) == 0); - - return cc; -} - -static inline void -pool_cache_cpu_exit(pool_cache_cpu_t *cc, int *s) -{ - - /* No longer need exclusive access to the per-CPU data. */ - if (cc->cc_ipl != IPL_NONE) { - splx(*s); - } - crit_exit(); -} - -#if __GNUC_PREREQ__(3, 0) -__attribute ((noinline)) -#endif -pool_cache_cpu_t * -pool_cache_get_slow(pool_cache_cpu_t *cc, int *s, void **objectp, +static bool __noinline +pool_cache_get_slow(pool_cache_cpu_t *cc, int s, void **objectp, paddr_t *pap, int flags) { pcg_t *pcg, *cur; @@ -2428,6 +2423,9 @@ pool_cache_get_slow(pool_cache_cpu_t *cc pool_cache_t pc; void *object; + KASSERT(cc->cc_current->pcg_avail == 0); + KASSERT(cc->cc_previous->pcg_avail == 0); + pc = cc->cc_cache; cc->cc_misses++; @@ -2435,7 +2433,7 @@ pool_cache_get_slow(pool_cache_cpu_t *cc * Nothing was available locally. Try and grab a group * from the cache. */ - if (!mutex_tryenter(&pc->pc_lock)) { + if (__predict_false(!mutex_tryenter(&pc->pc_lock))) { ncsw = curlwp->l_ncsw; mutex_enter(&pc->pc_lock); pc->pc_contended++; @@ -2447,18 +2445,17 @@ pool_cache_get_slow(pool_cache_cpu_t *cc */ if (curlwp->l_ncsw != ncsw) { mutex_exit(&pc->pc_lock); - pool_cache_cpu_exit(cc, s); - return pool_cache_cpu_enter(pc, s); + return true; } } - if ((pcg = pc->pc_fullgroups) != NULL) { + if (__predict_true((pcg = pc->pc_fullgroups) != NULL)) { /* * If there's a full group, release our empty * group back to the cache. Install the full * group as cc_current and return. */ - if ((cur = cc->cc_current) != NULL) { + if (__predict_true((cur = cc->cc_current) != &pcg_dummy)) { KASSERT(cur->pcg_avail == 0); cur->pcg_next = pc->pc_emptygroups; pc->pc_emptygroups = cur; @@ -2470,7 +2467,7 @@ pool_cache_get_slow(pool_cache_cpu_t *cc pc->pc_hits++; pc->pc_nfull--; mutex_exit(&pc->pc_lock); - return cc; + return true; } /* @@ -2480,17 +2477,17 @@ pool_cache_get_slow(pool_cache_cpu_t *cc */ pc->pc_misses++; mutex_exit(&pc->pc_lock); - pool_cache_cpu_exit(cc, s); + splx(s); object = pool_get(&pc->pc_pool, flags); *objectp = object; - if (object == NULL) - return NULL; + if (__predict_false(object == NULL)) + return false; - if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) { + if (__predict_false((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0)) { pool_put(&pc->pc_pool, object); *objectp = NULL; - return NULL; + return false; } KASSERT((((vaddr_t)object + pc->pc_pool.pr_itemoffset) & @@ -2505,7 +2502,7 @@ pool_cache_get_slow(pool_cache_cpu_t *cc } FREECHECK_OUT(&pc->pc_freecheck, object); - return NULL; + return false; } /* @@ -2523,23 +2520,29 @@ pool_cache_get_paddr(pool_cache_t pc, in int s; #ifdef LOCKDEBUG - if (flags & PR_WAITOK) - ASSERT_SLEEPABLE(NULL, "pool_cache_get(PR_WAITOK)"); + if (flags & PR_WAITOK) { + ASSERT_SLEEPABLE(); + } #endif - cc = pool_cache_cpu_enter(pc, &s); - do { + /* Lock out interrupts and disable preemption. */ + s = splvm(); + while (/* CONSTCOND */ true) { /* Try and allocate an object from the current group. */ + cc = pc->pc_cpus[curcpu()->ci_index]; + KASSERT(cc->cc_cache == pc); pcg = cc->cc_current; - if (pcg != NULL && pcg->pcg_avail > 0) { + if (__predict_true(pcg->pcg_avail > 0)) { object = pcg->pcg_objects[--pcg->pcg_avail].pcgo_va; - if (pap != NULL) + if (__predict_false(pap != NULL)) *pap = pcg->pcg_objects[pcg->pcg_avail].pcgo_pa; +#if defined(DIAGNOSTIC) pcg->pcg_objects[pcg->pcg_avail].pcgo_va = NULL; - KASSERT(pcg->pcg_avail <= pcg->pcg_size); + KASSERT(pcg->pcg_avail < pcg->pcg_size); KASSERT(object != NULL); +#endif cc->cc_hits++; - pool_cache_cpu_exit(cc, &s); + splx(s); FREECHECK_OUT(&pc->pc_freecheck, object); return object; } @@ -2549,7 +2552,7 @@ pool_cache_get_paddr(pool_cache_t pc, in * it with the current group and allocate from there. */ pcg = cc->cc_previous; - if (pcg != NULL && pcg->pcg_avail > 0) { + if (__predict_true(pcg->pcg_avail > 0)) { cc->cc_previous = cc->cc_current; cc->cc_current = pcg; continue; @@ -2558,107 +2561,105 @@ pool_cache_get_paddr(pool_cache_t pc, in /* * Can't allocate from either group: try the slow path. * If get_slow() allocated an object for us, or if - * no more objects are available, it will return NULL. + * no more objects are available, it will return false. * Otherwise, we need to retry. */ - cc = pool_cache_get_slow(cc, &s, &object, pap, flags); - } while (cc != NULL); + if (!pool_cache_get_slow(cc, s, &object, pap, flags)) + break; + } return object; } -#if __GNUC_PREREQ__(3, 0) -__attribute ((noinline)) -#endif -pool_cache_cpu_t * -pool_cache_put_slow(pool_cache_cpu_t *cc, int *s, void *object, paddr_t pa) +static bool __noinline +pool_cache_put_slow(pool_cache_cpu_t *cc, int s, void *object) { pcg_t *pcg, *cur; uint64_t ncsw; pool_cache_t pc; - u_int nobj; + + KASSERT(cc->cc_current->pcg_avail == cc->cc_current->pcg_size); + KASSERT(cc->cc_previous->pcg_avail == cc->cc_previous->pcg_size); pc = cc->cc_cache; + pcg = NULL; cc->cc_misses++; /* - * No free slots locally. Try to grab an empty, unused - * group from the cache. + * If there are no empty groups in the cache then allocate one + * while still unlocked. */ - if (!mutex_tryenter(&pc->pc_lock)) { + if (__predict_false(pc->pc_emptygroups == NULL)) { + if (__predict_true(!pool_cache_disable)) { + pcg = pool_get(pc->pc_pcgpool, PR_NOWAIT); + } + if (__predict_true(pcg != NULL)) { + pcg->pcg_avail = 0; + pcg->pcg_size = pc->pc_pcgsize; + } + } + + /* Lock the cache. */ + if (__predict_false(!mutex_tryenter(&pc->pc_lock))) { ncsw = curlwp->l_ncsw; mutex_enter(&pc->pc_lock); pc->pc_contended++; /* - * If we context switched while locking, then - * our view of the per-CPU data is invalid: - * retry. + * If we context switched while locking, then our view of + * the per-CPU data is invalid: retry. */ - if (curlwp->l_ncsw != ncsw) { + if (__predict_false(curlwp->l_ncsw != ncsw)) { mutex_exit(&pc->pc_lock); - pool_cache_cpu_exit(cc, s); - return pool_cache_cpu_enter(pc, s); + if (pcg != NULL) { + pool_put(pc->pc_pcgpool, pcg); + } + return true; } } - if ((pcg = pc->pc_emptygroups) != NULL) { - /* - * If there's a empty group, release our full - * group back to the cache. Install the empty - * group as cc_current and return. - */ - if ((cur = cc->cc_current) != NULL) { - KASSERT(cur->pcg_avail == pcg->pcg_size); - cur->pcg_next = pc->pc_fullgroups; - pc->pc_fullgroups = cur; - pc->pc_nfull++; - } - KASSERT(pcg->pcg_avail == 0); - cc->cc_current = pcg; + /* If there are no empty groups in the cache then allocate one. */ + if (pcg == NULL && pc->pc_emptygroups != NULL) { + pcg = pc->pc_emptygroups; pc->pc_emptygroups = pcg->pcg_next; - pc->pc_hits++; pc->pc_nempty--; - mutex_exit(&pc->pc_lock); - return cc; } /* - * Nothing available locally or in cache. Take the - * slow path and try to allocate a new group that we - * can release to. - */ - pc->pc_misses++; - mutex_exit(&pc->pc_lock); - pool_cache_cpu_exit(cc, s); - - /* - * If we can't allocate a new group, just throw the - * object away. + * If there's a empty group, release our full group back + * to the cache. Install the empty group to the local CPU + * and return. */ - nobj = pc->pc_pcgsize; - if (nobj == PCG_NOBJECTS_LARGE) { - pcg = pool_get(&pcg_large_pool, PR_NOWAIT); - } else { - pcg = pool_get(&pcg_normal_pool, PR_NOWAIT); - } - if (pcg == NULL) { - pool_cache_destruct_object(pc, object); - return NULL; + if (pcg != NULL) { + KASSERT(pcg->pcg_avail == 0); + if (__predict_false(cc->cc_previous == &pcg_dummy)) { + cc->cc_previous = pcg; + } else { + cur = cc->cc_current; + if (__predict_true(cur != &pcg_dummy)) { + KASSERT(cur->pcg_avail == cur->pcg_size); + cur->pcg_next = pc->pc_fullgroups; + pc->pc_fullgroups = cur; + pc->pc_nfull++; + } + cc->cc_current = pcg; + } + pc->pc_hits++; + mutex_exit(&pc->pc_lock); + return true; } - pcg->pcg_avail = 0; - pcg->pcg_size = nobj; /* - * Add the empty group to the cache and try again. + * Nothing available locally or in cache, and we didn't + * allocate an empty group. Take the slow path and destroy + * the object here and now. */ - mutex_enter(&pc->pc_lock); - pcg->pcg_next = pc->pc_emptygroups; - pc->pc_emptygroups = pcg; - pc->pc_nempty++; + pc->pc_misses++; mutex_exit(&pc->pc_lock); + splx(s); + pool_cache_destruct_object(pc, object); - return pool_cache_cpu_enter(pc, s); + return false; } /* @@ -2674,27 +2675,31 @@ pool_cache_put_paddr(pool_cache_t pc, vo pcg_t *pcg; int s; + KASSERT(object != NULL); FREECHECK_IN(&pc->pc_freecheck, object); - cc = pool_cache_cpu_enter(pc, &s); - do { + /* Lock out interrupts and disable preemption. */ + s = splvm(); + while (/* CONSTCOND */ true) { /* If the current group isn't full, release it there. */ + cc = pc->pc_cpus[curcpu()->ci_index]; + KASSERT(cc->cc_cache == pc); pcg = cc->cc_current; - if (pcg != NULL && pcg->pcg_avail < pcg->pcg_size) { + if (__predict_true(pcg->pcg_avail < pcg->pcg_size)) { pcg->pcg_objects[pcg->pcg_avail].pcgo_va = object; pcg->pcg_objects[pcg->pcg_avail].pcgo_pa = pa; pcg->pcg_avail++; cc->cc_hits++; - pool_cache_cpu_exit(cc, &s); + splx(s); return; } /* - * That failed. If the previous group is empty, swap + * That failed. If the previous group isn't full, swap * it with the current group and try again. */ pcg = cc->cc_previous; - if (pcg != NULL && pcg->pcg_avail == 0) { + if (__predict_true(pcg->pcg_avail < pcg->pcg_size)) { cc->cc_previous = cc->cc_current; cc->cc_current = pcg; continue; @@ -2703,10 +2708,11 @@ pool_cache_put_paddr(pool_cache_t pc, vo /* * Can't free to either group: try the slow path. * If put_slow() releases the object for us, it - * will return NULL. Otherwise we need to retry. + * will return false. Otherwise we need to retry. */ - cc = pool_cache_put_slow(cc, &s, object, pa); - } while (cc != NULL); + if (!pool_cache_put_slow(cc, s, object)) + break; + } } /* @@ -2720,24 +2726,16 @@ pool_cache_xcall(pool_cache_t pc) { pool_cache_cpu_t *cc; pcg_t *prev, *cur, **list; - int s = 0; /* XXXgcc */ - - cc = pool_cache_cpu_enter(pc, &s); - cur = cc->cc_current; - cc->cc_current = NULL; - prev = cc->cc_previous; - cc->cc_previous = NULL; - pool_cache_cpu_exit(cc, &s); + int s; - /* - * XXXSMP Go to splvm to prevent kernel_lock from being taken, - * because locks at IPL_SOFTXXX are still spinlocks. Does not - * apply to IPL_SOFTBIO. Cross-call threads do not take the - * kernel_lock. - */ s = splvm(); mutex_enter(&pc->pc_lock); - if (cur != NULL) { + cc = pc->pc_cpus[curcpu()->ci_index]; + cur = cc->cc_current; + cc->cc_current = __UNCONST(&pcg_dummy); + prev = cc->cc_previous; + cc->cc_previous = __UNCONST(&pcg_dummy); + if (cur != &pcg_dummy) { if (cur->pcg_avail == cur->pcg_size) { list = &pc->pc_fullgroups; pc->pc_nfull++; @@ -2751,7 +2749,7 @@ pool_cache_xcall(pool_cache_t pc) cur->pcg_next = *list; *list = cur; } - if (prev != NULL) { + if (prev != &pcg_dummy) { if (prev->pcg_avail == prev->pcg_size) { list = &pc->pc_fullgroups; pc->pc_nfull++; @@ -3050,13 +3048,13 @@ found: if (pool_in_cg(pp, cc->cc_current, addr) || pool_in_cg(pp, cc->cc_previous, addr)) { struct cpu_info *ci = - cpu_lookup_byindex(i); + cpu_lookup(i); incpucache = true; snprintf(cpucachestr, sizeof(cpucachestr), "cached by CPU %u", - (u_int)ci->ci_cpuid); + ci->ci_index); goto print; } }