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 retrieving revision 1.158.2.5 retrieving revision 1.159 diff -u -p -r1.158.2.5 -r1.159 --- src/sys/kern/subr_pool.c 2010/08/11 22:54:42 1.158.2.5 +++ src/sys/kern/subr_pool.c 2008/04/28 15:36:01 1.159 @@ -1,8 +1,7 @@ -/* $NetBSD: subr_pool.c,v 1.158.2.5 2010/08/11 22:54:42 yamt Exp $ */ +/* $NetBSD: subr_pool.c,v 1.159 2008/04/28 15:36:01 ad Exp $ */ /*- - * Copyright (c) 1997, 1999, 2000, 2002, 2007, 2008, 2010 - * The NetBSD Foundation, Inc. + * Copyright (c) 1997, 1999, 2000, 2002, 2007 The NetBSD Foundation, Inc. * All rights reserved. * * This code is derived from software contributed to The NetBSD Foundation @@ -17,6 +16,13 @@ * 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 @@ -32,7 +38,7 @@ */ #include -__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.158.2.5 2010/08/11 22:54:42 yamt Exp $"); +__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.159 2008/04/28 15:36:01 ad Exp $"); #include "opt_ddb.h" #include "opt_pool.h" @@ -70,7 +76,7 @@ __KERNEL_RCSID(0, "$NetBSD: subr_pool.c, */ /* List of all pools */ -static TAILQ_HEAD(, pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head); +TAILQ_HEAD(,pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head); /* Private pool for page header structures */ #define PHPOOL_MAX 8 @@ -105,9 +111,6 @@ 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) @@ -183,16 +186,15 @@ static struct pool cache_cpu_pool; TAILQ_HEAD(,pool_cache) pool_cache_head = TAILQ_HEAD_INITIALIZER(pool_cache_head); -int pool_cache_disable; /* global disable for caching */ -static const pcg_t pcg_dummy; /* zero sized: always empty, yet always full */ +int pool_cache_disable; + -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 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 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 *); @@ -235,28 +237,16 @@ int pool_logsize = POOL_LOGSIZE; static inline void pr_log(struct pool *pp, void *v, int action, const char *file, long line) { - int n; + int n = pp->pr_curlogentry; 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; @@ -274,7 +264,7 @@ pr_printlog(struct pool *pp, struct pool int i = pp->pr_logsize; int n = pp->pr_curlogentry; - if (pp->pr_log == NULL) + if ((pp->pr_roflags & PR_LOGGING) == 0) return; /* @@ -562,11 +552,6 @@ pool_reclaim_register(struct pool *pp) callback_register(&vm_map_to_kernel(map)->vmk_reclaim_callback, &pp->pr_reclaimerentry, pp, pool_reclaim_callback); splx(s); - -#ifdef DIAGNOSTIC - /* Diagnostic drain attempt. */ - uvm_km_va_drain(map, 0); -#endif } static void @@ -609,11 +594,17 @@ 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); @@ -672,9 +663,7 @@ pool_init(struct pool *pp, size_t size, palloc = &pool_allocator_nointr_fullpage; } #endif /* POOL_SUBPAGE */ - if (!cold) - mutex_enter(&pool_allocator_lock); - if (palloc->pa_refcnt++ == 0) { + if ((palloc->pa_flags & PA_INITIALIZED) == 0) { if (palloc->pa_pagesz == 0) palloc->pa_pagesz = PAGE_SIZE; @@ -687,9 +676,8 @@ 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); @@ -812,7 +800,16 @@ pool_init(struct pool *pp, size_t size, pp->pr_nidle = 0; pp->pr_refcnt = 0; - pp->pr_log = NULL; +#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_entered_file = NULL; pp->pr_entered_line = 0; @@ -861,7 +858,7 @@ pool_init(struct pool *pp, size_t size, } /* Insert into the list of all pools. */ - if (!cold) + if (__predict_true(!cold)) mutex_enter(&pool_head_lock); TAILQ_FOREACH(pp1, &pool_head, pr_poollist) { if (strcmp(pp1->pr_wchan, pp->pr_wchan) > 0) @@ -871,14 +868,14 @@ 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 (!cold) + if (__predict_true(!cold)) mutex_exit(&pool_head_lock); - /* Insert this into the list of pools using this allocator. */ - if (!cold) + /* Insert this into the list of pools using this allocator. */ + if (__predict_true(!cold)) mutex_enter(&palloc->pa_lock); TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list); - if (!cold) + if (__predict_true(!cold)) mutex_exit(&palloc->pa_lock); pool_reclaim_register(pp); @@ -908,11 +905,6 @@ 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); @@ -938,10 +930,8 @@ pool_destroy(struct pool *pp) pr_pagelist_free(pp, &pq); #ifdef POOL_DIAGNOSTIC - if (pp->pr_log != NULL) { + if ((pp->pr_roflags & PR_LOGGING) != 0) free(pp->pr_log, M_TEMP); - pp->pr_log = NULL; - } #endif cv_destroy(&pp->pr_cv); @@ -989,17 +979,19 @@ pool_get(struct pool *pp, int flags) void *v; #ifdef DIAGNOSTIC - if (pp->pr_itemsperpage == 0) - panic("pool_get: pool '%s': pr_itemsperpage is zero, " - "pool not initialized?", pp->pr_wchan); - if ((cpu_intr_p() || cpu_softintr_p()) && pp->pr_ipl == IPL_NONE && - !cold && panicstr == NULL) - panic("pool '%s' is IPL_NONE, but called from " - "interrupt context\n", pp->pr_wchan); -#endif + if (__predict_false(pp->pr_itemsperpage == 0)) + panic("pool_get: pool %p: pr_itemsperpage is zero, " + "pool not initialized?", pp); + if (__predict_false(curlwp == NULL && doing_shutdown == 0 && + (flags & PR_WAITOK) != 0)) + panic("pool_get: %s: must have NOWAIT", pp->pr_wchan); + +#endif /* DIAGNOSTIC */ +#ifdef LOCKDEBUG if (flags & PR_WAITOK) { ASSERT_SLEEPABLE(); } +#endif mutex_enter(&pp->pr_lock); pr_enter(pp, file, line); @@ -1268,7 +1260,10 @@ 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; } /* @@ -1542,8 +1537,6 @@ 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 @@ -1607,8 +1600,6 @@ pool_sethardlimit(struct pool *pp, int n /* * Release all complete pages that have not been used recently. - * - * Might be called from interrupt context. */ int #ifdef POOL_DIAGNOSTIC @@ -1623,10 +1614,6 @@ pool_reclaim(struct pool *pp) bool klock; int rv; - if (cpu_intr_p() || cpu_softintr_p()) { - KASSERT(pp->pr_ipl != IPL_NONE); - } - if (pp->pr_drain_hook != NULL) { /* * The drain hook must be called with the pool unlocked. @@ -1745,13 +1732,12 @@ pool_drain_start(struct pool **ppp, uint } } -bool +void pool_drain_end(struct pool *pp, uint64_t where) { - bool reclaimed; if (pp == NULL) - return false; + return; KASSERT(pp->pr_refcnt > 0); @@ -1760,15 +1746,13 @@ pool_drain_end(struct pool *pp, uint64_t xc_wait(where); /* Drain the cache (if any) and pool.. */ - reclaimed = pool_reclaim(pp); + pool_reclaim(pp); /* Finally, unlock the pool. */ mutex_enter(&pool_head_lock); pp->pr_refcnt--; cv_broadcast(&pool_busy); mutex_exit(&pool_head_lock); - - return reclaimed; } /* @@ -1917,7 +1901,7 @@ pool_print1(struct pool *pp, const char if (pc != NULL) { cpuhit = 0; cpumiss = 0; - for (i = 0; i < __arraycount(pc->pc_cpus); i++) { + for (i = 0; i < MAXCPUS; i++) { if ((cc = pc->pc_cpus[i]) == NULL) continue; cpuhit += cc->cc_hits; @@ -2108,10 +2092,8 @@ 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. */ @@ -2153,7 +2135,9 @@ void pool_cache_destroy(pool_cache_t pc) { struct pool *pp = &pc->pc_pool; - u_int i; + pool_cache_cpu_t *cc; + pcg_t *pcg; + int i; /* Remove it from the global list. */ mutex_enter(&pool_head_lock); @@ -2171,8 +2155,20 @@ pool_cache_destroy(pool_cache_t pc) mutex_exit(&pp->pr_lock); /* Destroy per-CPU data */ - for (i = 0; i < __arraycount(pc->pc_cpus); i++) - pool_cache_invalidate_cpu(pc, i); + 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); + } /* Finally, destroy it. */ mutex_destroy(&pc->pc_lock); @@ -2193,7 +2189,7 @@ pool_cache_cpu_init1(struct cpu_info *ci index = ci->ci_index; - KASSERT(index < __arraycount(pc->pc_cpus)); + KASSERT(index < MAXCPUS); if ((cc = pc->pc_cpus[index]) != NULL) { KASSERT(cc->cc_cpuindex == index); @@ -2220,8 +2216,8 @@ pool_cache_cpu_init1(struct cpu_info *ci cc->cc_cpuindex = index; cc->cc_hits = 0; cc->cc_misses = 0; - cc->cc_current = __UNCONST(&pcg_dummy); - cc->cc_previous = __UNCONST(&pcg_dummy); + cc->cc_current = NULL; + cc->cc_previous = NULL; pc->pc_cpus[index] = cc; } @@ -2319,35 +2315,12 @@ 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; @@ -2365,40 +2338,6 @@ 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) { @@ -2427,8 +2366,42 @@ pool_cache_sethardlimit(pool_cache_t pc, pool_sethardlimit(&pc->pc_pool, n, warnmess, ratecap); } -static bool __noinline -pool_cache_get_slow(pool_cache_cpu_t *cc, int s, void **objectp, +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. + */ + KPREEMPT_DISABLE(curlwp); + cc = pc->pc_cpus[curcpu()->ci_index]; + KASSERT(cc->cc_cache == pc); + if (cc->cc_ipl != IPL_NONE) { + *s = splraiseipl(cc->cc_iplcookie); + } + + 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); + } + KPREEMPT_ENABLE(curlwp); +} + +#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, paddr_t *pap, int flags) { pcg_t *pcg, *cur; @@ -2436,9 +2409,6 @@ 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++; @@ -2446,7 +2416,7 @@ pool_cache_get_slow(pool_cache_cpu_t *cc * Nothing was available locally. Try and grab a group * from the cache. */ - if (__predict_false(!mutex_tryenter(&pc->pc_lock))) { + if (!mutex_tryenter(&pc->pc_lock)) { ncsw = curlwp->l_ncsw; mutex_enter(&pc->pc_lock); pc->pc_contended++; @@ -2458,17 +2428,18 @@ pool_cache_get_slow(pool_cache_cpu_t *cc */ if (curlwp->l_ncsw != ncsw) { mutex_exit(&pc->pc_lock); - return true; + pool_cache_cpu_exit(cc, s); + return pool_cache_cpu_enter(pc, s); } } - if (__predict_true((pcg = pc->pc_fullgroups) != NULL)) { + if ((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 (__predict_true((cur = cc->cc_current) != &pcg_dummy)) { + if ((cur = cc->cc_current) != NULL) { KASSERT(cur->pcg_avail == 0); cur->pcg_next = pc->pc_emptygroups; pc->pc_emptygroups = cur; @@ -2480,7 +2451,7 @@ pool_cache_get_slow(pool_cache_cpu_t *cc pc->pc_hits++; pc->pc_nfull--; mutex_exit(&pc->pc_lock); - return true; + return cc; } /* @@ -2490,17 +2461,17 @@ pool_cache_get_slow(pool_cache_cpu_t *cc */ pc->pc_misses++; mutex_exit(&pc->pc_lock); - splx(s); + pool_cache_cpu_exit(cc, s); object = pool_get(&pc->pc_pool, flags); *objectp = object; - if (__predict_false(object == NULL)) - return false; + if (object == NULL) + return NULL; - if (__predict_false((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0)) { + if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) { pool_put(&pc->pc_pool, object); *objectp = NULL; - return false; + return NULL; } KASSERT((((vaddr_t)object + pc->pc_pool.pr_itemoffset) & @@ -2515,7 +2486,7 @@ pool_cache_get_slow(pool_cache_cpu_t *cc } FREECHECK_OUT(&pc->pc_freecheck, object); - return false; + return NULL; } /* @@ -2532,33 +2503,27 @@ pool_cache_get_paddr(pool_cache_t pc, in void *object; int s; - KASSERTMSG((!cpu_intr_p() && !cpu_softintr_p()) || - (pc->pc_pool.pr_ipl != IPL_NONE || cold || panicstr != NULL), - ("pool '%s' is IPL_NONE, but called from interrupt context\n", - pc->pc_pool.pr_wchan)); - +#ifdef LOCKDEBUG if (flags & PR_WAITOK) { ASSERT_SLEEPABLE(); } +#endif - /* Lock out interrupts and disable preemption. */ - s = splvm(); - while (/* CONSTCOND */ true) { + cc = pool_cache_cpu_enter(pc, &s); + do { /* 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 (__predict_true(pcg->pcg_avail > 0)) { + if (pcg != NULL && pcg->pcg_avail > 0) { object = pcg->pcg_objects[--pcg->pcg_avail].pcgo_va; - if (__predict_false(pap != NULL)) + if (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); +#endif /* defined(DIAGNOSTIC) */ + KASSERT(pcg->pcg_avail <= pcg->pcg_size); KASSERT(object != NULL); -#endif cc->cc_hits++; - splx(s); + pool_cache_cpu_exit(cc, &s); FREECHECK_OUT(&pc->pc_freecheck, object); return object; } @@ -2568,7 +2533,7 @@ pool_cache_get_paddr(pool_cache_t pc, in * it with the current group and allocate from there. */ pcg = cc->cc_previous; - if (__predict_true(pcg->pcg_avail > 0)) { + if (pcg != NULL && pcg->pcg_avail > 0) { cc->cc_previous = cc->cc_current; cc->cc_current = pcg; continue; @@ -2577,83 +2542,63 @@ 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 false. + * no more objects are available, it will return NULL. * Otherwise, we need to retry. */ - if (!pool_cache_get_slow(cc, s, &object, pap, flags)) - break; - } + cc = pool_cache_get_slow(cc, &s, &object, pap, flags); + } while (cc != NULL); return object; } -static bool __noinline -pool_cache_put_slow(pool_cache_cpu_t *cc, int s, void *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) { pcg_t *pcg, *cur; uint64_t ncsw; pool_cache_t pc; - - KASSERT(cc->cc_current->pcg_avail == cc->cc_current->pcg_size); - KASSERT(cc->cc_previous->pcg_avail == cc->cc_previous->pcg_size); + u_int nobj; pc = cc->cc_cache; - pcg = NULL; cc->cc_misses++; /* - * If there are no empty groups in the cache then allocate one - * while still unlocked. + * No free slots locally. Try to grab an empty, unused + * group from the cache. */ - 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))) { + if (!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 (__predict_false(curlwp->l_ncsw != ncsw)) { + if (curlwp->l_ncsw != ncsw) { mutex_exit(&pc->pc_lock); - if (pcg != NULL) { - pool_put(pc->pc_pcgpool, pcg); - } - return true; + pool_cache_cpu_exit(cc, s); + return pool_cache_cpu_enter(pc, s); } } - /* 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_nempty--; - } - - /* - * If there's a empty group, release our full group back - * to the cache. Install the empty group to the local CPU - * and return. - */ - if (pcg != NULL) { + if ((pcg = pc->pc_emptygroups) != NULL) { + /* + * If there's a empty group, release our full + * group back to the cache. Install the empty + * group and return. + */ KASSERT(pcg->pcg_avail == 0); - if (__predict_false(cc->cc_previous == &pcg_dummy)) { + pc->pc_emptygroups = pcg->pcg_next; + if (cc->cc_previous == NULL) { cc->cc_previous = pcg; } else { - cur = cc->cc_current; - if (__predict_true(cur != &pcg_dummy)) { - KASSERT(cur->pcg_avail == cur->pcg_size); + 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++; @@ -2661,21 +2606,49 @@ pool_cache_put_slow(pool_cache_cpu_t *cc cc->cc_current = pcg; } pc->pc_hits++; + pc->pc_nempty--; mutex_exit(&pc->pc_lock); - return true; + return cc; } /* - * 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. + * 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); - splx(s); - pool_cache_destruct_object(pc, object); + pool_cache_cpu_exit(cc, s); - return false; + /* + * If we can't allocate a new group, just throw the + * object away. + */ + nobj = pc->pc_pcgsize; + if (pool_cache_disable) { + pcg = NULL; + } else 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; + } + pcg->pcg_avail = 0; + pcg->pcg_size = nobj; + + /* + * Add the empty group to the cache and try again. + */ + mutex_enter(&pc->pc_lock); + pcg->pcg_next = pc->pc_emptygroups; + pc->pc_emptygroups = pcg; + pc->pc_nempty++; + mutex_exit(&pc->pc_lock); + + return pool_cache_cpu_enter(pc, s); } /* @@ -2691,31 +2664,27 @@ pool_cache_put_paddr(pool_cache_t pc, vo pcg_t *pcg; int s; - KASSERT(object != NULL); FREECHECK_IN(&pc->pc_freecheck, object); - /* Lock out interrupts and disable preemption. */ - s = splvm(); - while (/* CONSTCOND */ true) { + cc = pool_cache_cpu_enter(pc, &s); + do { /* 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 (__predict_true(pcg->pcg_avail < pcg->pcg_size)) { + if (pcg != NULL && 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++; - splx(s); + pool_cache_cpu_exit(cc, &s); return; } /* - * That failed. If the previous group isn't full, swap + * That failed. If the previous group is empty, swap * it with the current group and try again. */ pcg = cc->cc_previous; - if (__predict_true(pcg->pcg_avail < pcg->pcg_size)) { + if (pcg != NULL && pcg->pcg_avail == 0) { cc->cc_previous = cc->cc_current; cc->cc_current = pcg; continue; @@ -2724,11 +2693,10 @@ 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 false. Otherwise we need to retry. + * will return NULL. Otherwise we need to retry. */ - if (!pool_cache_put_slow(cc, s, object)) - break; - } + cc = pool_cache_put_slow(cc, &s, object, pa); + } while (cc != NULL); } /* @@ -2742,16 +2710,24 @@ pool_cache_xcall(pool_cache_t pc) { pool_cache_cpu_t *cc; pcg_t *prev, *cur, **list; - int s; + int s = 0; /* XXXgcc */ - s = splvm(); - mutex_enter(&pc->pc_lock); - cc = pc->pc_cpus[curcpu()->ci_index]; + cc = pool_cache_cpu_enter(pc, &s); cur = cc->cc_current; - cc->cc_current = __UNCONST(&pcg_dummy); + cc->cc_current = NULL; prev = cc->cc_previous; - cc->cc_previous = __UNCONST(&pcg_dummy); - if (cur != &pcg_dummy) { + cc->cc_previous = NULL; + pool_cache_cpu_exit(cc, &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) { if (cur->pcg_avail == cur->pcg_size) { list = &pc->pc_fullgroups; pc->pc_nfull++; @@ -2765,7 +2741,7 @@ pool_cache_xcall(pool_cache_t pc) cur->pcg_next = *list; *list = cur; } - if (prev != &pcg_dummy) { + if (prev != NULL) { if (prev->pcg_avail == prev->pcg_size) { list = &pc->pc_fullgroups; pc->pc_nfull++; @@ -3055,7 +3031,7 @@ found: goto print; } } - for (i = 0; i < __arraycount(pc->pc_cpus); i++) { + for (i = 0; i < MAXCPUS; i++) { pool_cache_cpu_t *cc; if ((cc = pc->pc_cpus[i]) == NULL) { @@ -3064,7 +3040,7 @@ found: if (pool_in_cg(pp, cc->cc_current, addr) || pool_in_cg(pp, cc->cc_previous, addr)) { struct cpu_info *ci = - cpu_lookup(i); + cpu_lookup_byindex(i); incpucache = true; snprintf(cpucachestr,