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 1.156.2.2 retrieving revision 1.190.2.1 diff -u -p -r1.156.2.2 -r1.190.2.1 --- src/sys/kern/subr_pool.c 2008/06/04 02:05:39 1.156.2.2 +++ src/sys/kern/subr_pool.c 2012/04/17 00:08:28 1.190.2.1 @@ -1,7 +1,8 @@ -/* $NetBSD: subr_pool.c,v 1.156.2.2 2008/06/04 02:05:39 yamt Exp $ */ +/* $NetBSD: subr_pool.c,v 1.190.2.1 2012/04/17 00:08:28 yamt Exp $ */ /*- - * Copyright (c) 1997, 1999, 2000, 2002, 2007, 2008 The NetBSD Foundation, Inc. + * Copyright (c) 1997, 1999, 2000, 2002, 2007, 2008, 2010 + * The NetBSD Foundation, Inc. * All rights reserved. * * This code is derived from software contributed to The NetBSD Foundation @@ -31,7 +32,7 @@ */ #include -__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.156.2.2 2008/06/04 02:05:39 yamt Exp $"); +__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.190.2.1 2012/04/17 00:08:28 yamt Exp $"); #include "opt_ddb.h" #include "opt_pool.h" @@ -45,6 +46,7 @@ __KERNEL_RCSID(0, "$NetBSD: subr_pool.c, #include #include #include +#include #include #include #include @@ -53,7 +55,7 @@ __KERNEL_RCSID(0, "$NetBSD: subr_pool.c, #include #include -#include +#include /* * Pool resource management utility. @@ -69,7 +71,7 @@ __KERNEL_RCSID(0, "$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 @@ -82,16 +84,14 @@ static struct pool phpool[PHPOOL_MAX]; static struct pool psppool; #endif -static SLIST_HEAD(, pool_allocator) pa_deferinitq = - SLIST_HEAD_INITIALIZER(pa_deferinitq); - static void *pool_page_alloc_meta(struct pool *, int); static void pool_page_free_meta(struct pool *, void *); /* allocator for pool metadata */ struct pool_allocator pool_allocator_meta = { - pool_page_alloc_meta, pool_page_free_meta, - .pa_backingmapptr = &kmem_map, + .pa_alloc = pool_page_alloc_meta, + .pa_free = pool_page_free_meta, + .pa_pagesz = 0 }; /* # of seconds to retain page after last use */ @@ -104,6 +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) @@ -175,19 +178,22 @@ static struct pool pcg_large_pool; static struct pool cache_pool; static struct pool cache_cpu_pool; +pool_cache_t pnbuf_cache; /* pathname buffer cache */ + /* List of all caches. */ 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 *); @@ -230,16 +236,28 @@ 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; @@ -257,7 +275,7 @@ 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; /* @@ -507,109 +525,59 @@ pr_rmpage(struct pool *pp, struct pool_i pool_update_curpage(pp); } -static bool -pa_starved_p(struct pool_allocator *pa) -{ - - if (pa->pa_backingmap != NULL) { - return vm_map_starved_p(pa->pa_backingmap); - } - return false; -} - -static int -pool_reclaim_callback(struct callback_entry *ce, void *obj, void *arg) -{ - struct pool *pp = obj; - struct pool_allocator *pa = pp->pr_alloc; - - KASSERT(&pp->pr_reclaimerentry == ce); - pool_reclaim(pp); - if (!pa_starved_p(pa)) { - return CALLBACK_CHAIN_ABORT; - } - return CALLBACK_CHAIN_CONTINUE; -} - -static void -pool_reclaim_register(struct pool *pp) -{ - struct vm_map *map = pp->pr_alloc->pa_backingmap; - int s; - - if (map == NULL) { - return; - } - - s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */ - callback_register(&vm_map_to_kernel(map)->vmk_reclaim_callback, - &pp->pr_reclaimerentry, pp, pool_reclaim_callback); - splx(s); -} - -static void -pool_reclaim_unregister(struct pool *pp) -{ - struct vm_map *map = pp->pr_alloc->pa_backingmap; - int s; - - if (map == NULL) { - return; - } - - s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */ - callback_unregister(&vm_map_to_kernel(map)->vmk_reclaim_callback, - &pp->pr_reclaimerentry); - splx(s); -} - -static void -pa_reclaim_register(struct pool_allocator *pa) -{ - struct vm_map *map = *pa->pa_backingmapptr; - struct pool *pp; - - KASSERT(pa->pa_backingmap == NULL); - if (map == NULL) { - SLIST_INSERT_HEAD(&pa_deferinitq, pa, pa_q); - return; - } - pa->pa_backingmap = map; - TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) { - pool_reclaim_register(pp); - } -} - /* * Initialize all the pools listed in the "pools" link set. */ void pool_subsystem_init(void) { - struct pool_allocator *pa; - __link_set_decl(pools, struct link_pool_init); - struct link_pool_init * const *pi; + size_t size; + int idx; 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); - SLIST_REMOVE_HEAD(&pa_deferinitq, pa_q); - pa_reclaim_register(pa); + /* + * Initialize private page header pool and cache magazine pool if we + * haven't done so yet. + */ + for (idx = 0; idx < PHPOOL_MAX; idx++) { + static char phpool_names[PHPOOL_MAX][6+1+6+1]; + int nelem; + size_t sz; + + nelem = PHPOOL_FREELIST_NELEM(idx); + snprintf(phpool_names[idx], sizeof(phpool_names[idx]), + "phpool-%d", nelem); + sz = sizeof(struct pool_item_header); + if (nelem) { + sz = offsetof(struct pool_item_header, + ph_bitmap[howmany(nelem, BITMAP_SIZE)]); + } + pool_init(&phpool[idx], sz, 0, 0, 0, + phpool_names[idx], &pool_allocator_meta, IPL_VM); } +#ifdef POOL_SUBPAGE + pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0, + PR_RECURSIVE, "psppool", &pool_allocator_meta, IPL_VM); +#endif + + size = sizeof(pcg_t) + + (PCG_NOBJECTS_NORMAL - 1) * sizeof(pcgpair_t); + 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, coherency_unit, 0, 0, + "pcglarge", &pool_allocator_meta, IPL_VM); pool_init(&cache_pool, sizeof(struct pool_cache), coherency_unit, - 0, 0, "pcache", &pool_allocator_nointr, IPL_NONE); + 0, 0, "pcache", &pool_allocator_meta, IPL_NONE); pool_init(&cache_cpu_pool, sizeof(pool_cache_cpu_t), coherency_unit, - 0, 0, "pcachecpu", &pool_allocator_nointr, IPL_NONE); + 0, 0, "pcachecpu", &pool_allocator_meta, IPL_NONE); } /* @@ -656,7 +624,9 @@ 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; @@ -665,12 +635,9 @@ pool_init(struct pool *pp, size_t size, mutex_init(&palloc->pa_lock, MUTEX_DEFAULT, IPL_VM); palloc->pa_pagemask = ~(palloc->pa_pagesz - 1); palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1; - - 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); @@ -793,16 +760,7 @@ 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; @@ -811,47 +769,8 @@ pool_init(struct pool *pp, size_t size, cv_init(&pp->pr_cv, wchan); pp->pr_ipl = ipl; - /* - * Initialize private page header pool and cache magazine pool if we - * haven't done so yet. - * XXX LOCKING. - */ - if (phpool[0].pr_size == 0) { - int idx; - for (idx = 0; idx < PHPOOL_MAX; idx++) { - static char phpool_names[PHPOOL_MAX][6+1+6+1]; - int nelem; - size_t sz; - - nelem = PHPOOL_FREELIST_NELEM(idx); - snprintf(phpool_names[idx], sizeof(phpool_names[idx]), - "phpool-%d", nelem); - sz = sizeof(struct pool_item_header); - if (nelem) { - sz = offsetof(struct pool_item_header, - ph_bitmap[howmany(nelem, BITMAP_SIZE)]); - } - pool_init(&phpool[idx], sz, 0, 0, 0, - phpool_names[idx], &pool_allocator_meta, IPL_VM); - } -#ifdef POOL_SUBPAGE - pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0, - PR_RECURSIVE, "psppool", &pool_allocator_meta, IPL_VM); -#endif - - size = sizeof(pcg_t) + - (PCG_NOBJECTS_NORMAL - 1) * sizeof(pcgpair_t); - 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, 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) @@ -861,17 +780,15 @@ 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); } /* @@ -893,11 +810,15 @@ pool_destroy(struct pool *pp) mutex_exit(&pool_head_lock); /* Remove this pool from its allocator's list of pools. */ - pool_reclaim_unregister(pp); mutex_enter(&pp->pr_alloc->pa_lock); 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); @@ -923,8 +844,10 @@ 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); @@ -972,19 +895,17 @@ pool_get(struct pool *pp, int flags) void *v; #ifdef DIAGNOSTIC - 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 (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 (flags & PR_WAITOK) { ASSERT_SLEEPABLE(); } -#endif mutex_enter(&pp->pr_lock); pr_enter(pp, file, line); @@ -1253,10 +1174,7 @@ 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; } /* @@ -1530,6 +1448,8 @@ 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 @@ -1593,6 +1513,8 @@ 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 @@ -1607,6 +1529,10 @@ 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. @@ -1649,8 +1575,7 @@ pool_reclaim(struct pool *pp) break; KASSERT(ph->ph_nmissing == 0); - if (curtime - ph->ph_time < pool_inactive_time - && !pa_starved_p(pp->pr_alloc)) + if (curtime - ph->ph_time < pool_inactive_time) continue; /* @@ -1725,12 +1650,13 @@ pool_drain_start(struct pool **ppp, uint } } -void +bool pool_drain_end(struct pool *pp, uint64_t where) { + bool reclaimed; if (pp == NULL) - return; + return false; KASSERT(pp->pr_refcnt > 0); @@ -1739,13 +1665,15 @@ pool_drain_end(struct pool *pp, uint64_t xc_wait(where); /* Drain the cache (if any) and pool.. */ - pool_reclaim(pp); + reclaimed = 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; } /* @@ -1894,7 +1822,7 @@ pool_print1(struct pool *pp, const char if (pc != NULL) { cpuhit = 0; cpumiss = 0; - for (i = 0; i < MAXCPUS; i++) { + for (i = 0; i < __arraycount(pc->pc_cpus); i++) { if ((cc = pc->pc_cpus[i]) == NULL) continue; cpuhit += cc->cc_hits; @@ -2085,8 +2013,10 @@ 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. */ @@ -2127,10 +2057,21 @@ pool_cache_bootstrap(pool_cache_t pc, si void pool_cache_destroy(pool_cache_t pc) { + + pool_cache_bootstrap_destroy(pc); + pool_put(&cache_pool, pc); +} + +/* + * pool_cache_bootstrap_destroy: + * + * Destroy a pool cache. + */ +void +pool_cache_bootstrap_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); @@ -2148,25 +2089,12 @@ 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 < __arraycount(pc->pc_cpus); i++) + pool_cache_invalidate_cpu(pc, i); /* Finally, destroy it. */ mutex_destroy(&pc->pc_lock); pool_destroy(pp); - pool_put(&cache_pool, pc); } /* @@ -2182,7 +2110,7 @@ pool_cache_cpu_init1(struct cpu_info *ci index = ci->ci_index; - KASSERT(index < MAXCPUS); + KASSERT(index < __arraycount(pc->pc_cpus)); if ((cc = pc->pc_cpus[index]) != NULL) { KASSERT(cc->cc_cpuindex == index); @@ -2209,8 +2137,8 @@ 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; } @@ -2308,12 +2236,35 @@ 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; @@ -2331,6 +2282,40 @@ 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) { @@ -2359,39 +2344,8 @@ 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. - */ - 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); -} - -pool_cache_cpu_t * __noinline -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; @@ -2399,6 +2353,9 @@ 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++; @@ -2406,7 +2363,7 @@ 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++; @@ -2418,18 +2375,17 @@ 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; @@ -2441,7 +2397,7 @@ pool_cache_get_slow(pool_cache_cpu_t *cc pc->pc_hits++; pc->pc_nfull--; mutex_exit(&pc->pc_lock); - return cc; + return true; } /* @@ -2451,17 +2407,17 @@ 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) & @@ -2476,7 +2432,7 @@ pool_cache_get_slow(pool_cache_cpu_t *cc } FREECHECK_OUT(&pc->pc_freecheck, object); - return NULL; + return false; } /* @@ -2493,27 +2449,33 @@ pool_cache_get_paddr(pool_cache_t pc, in void *object; int s; -#ifdef LOCKDEBUG + 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); + 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; -#endif /* defined(DIAGNOSTIC) */ - 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; } @@ -2523,7 +2485,7 @@ 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; @@ -2532,60 +2494,83 @@ 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; } -pool_cache_cpu_t * __noinline -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 and return. - */ - KASSERT(pcg->pcg_avail == 0); + /* 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; - if (cc->cc_previous == NULL) { + 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) { + KASSERT(pcg->pcg_avail == 0); + if (__predict_false(cc->cc_previous == &pcg_dummy)) { cc->cc_previous = pcg; } else { - if ((cur = cc->cc_current) != NULL) { - KASSERT(cur->pcg_avail == pcg->pcg_size); + 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++; @@ -2593,49 +2578,21 @@ 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 cc; + return true; } /* - * Nothing available locally or in cache. Take the - * slow path and try to allocate a new group that we - * can release to. + * 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. */ 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. - */ - 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); + splx(s); + pool_cache_destruct_object(pc, object); - return pool_cache_cpu_enter(pc, s); + return false; } /* @@ -2651,27 +2608,31 @@ 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; @@ -2680,10 +2641,11 @@ 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; + } } /* @@ -2697,24 +2659,16 @@ 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++; @@ -2728,7 +2682,7 @@ 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++; @@ -2764,28 +2718,29 @@ void pool_page_free(struct pool *, void #ifdef POOL_SUBPAGE struct pool_allocator pool_allocator_kmem_fullpage = { - pool_page_alloc, pool_page_free, 0, - .pa_backingmapptr = &kmem_map, + .pa_alloc = pool_page_alloc, + .pa_free = pool_page_free, + .pa_pagesz = 0 }; #else struct pool_allocator pool_allocator_kmem = { - pool_page_alloc, pool_page_free, 0, - .pa_backingmapptr = &kmem_map, + .pa_alloc = pool_page_alloc, + .pa_free = pool_page_free, + .pa_pagesz = 0 }; #endif -void *pool_page_alloc_nointr(struct pool *, int); -void pool_page_free_nointr(struct pool *, void *); - #ifdef POOL_SUBPAGE struct pool_allocator pool_allocator_nointr_fullpage = { - pool_page_alloc_nointr, pool_page_free_nointr, 0, - .pa_backingmapptr = &kernel_map, + .pa_alloc = pool_page_alloc, + .pa_free = pool_page_free, + .pa_pagesz = 0 }; #else struct pool_allocator pool_allocator_nointr = { - pool_page_alloc_nointr, pool_page_free_nointr, 0, - .pa_backingmapptr = &kernel_map, + .pa_alloc = pool_page_alloc, + .pa_free = pool_page_free, + .pa_pagesz = 0 }; #endif @@ -2794,16 +2749,15 @@ void *pool_subpage_alloc(struct pool *, void pool_subpage_free(struct pool *, void *); struct pool_allocator pool_allocator_kmem = { - pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE, - .pa_backingmapptr = &kmem_map, + .pa_alloc = pool_subpage_alloc, + .pa_free = pool_subpage_free, + .pa_pagesz = POOL_SUBPAGE }; -void *pool_subpage_alloc_nointr(struct pool *, int); -void pool_subpage_free_nointr(struct pool *, void *); - struct pool_allocator pool_allocator_nointr = { - pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE, - .pa_backingmapptr = &kmem_map, + .pa_alloc = pool_subpage_alloc, + .pa_free = pool_subpage_free, + .pa_pagesz = POOL_SUBPAGE }; #endif /* POOL_SUBPAGE */ @@ -2839,31 +2793,41 @@ pool_allocator_free(struct pool *pp, voi void * pool_page_alloc(struct pool *pp, int flags) { - bool waitok = (flags & PR_WAITOK) ? true : false; + const vm_flag_t vflags = (flags & PR_WAITOK) ? VM_SLEEP: VM_NOSLEEP; + vmem_addr_t va; + int ret; - return ((void *) uvm_km_alloc_poolpage_cache(kmem_map, waitok)); + ret = uvm_km_kmem_alloc(kmem_va_arena, pp->pr_alloc->pa_pagesz, + vflags | VM_INSTANTFIT, &va); + + return ret ? NULL : (void *)va; } void pool_page_free(struct pool *pp, void *v) { - uvm_km_free_poolpage_cache(kmem_map, (vaddr_t) v); + uvm_km_kmem_free(kmem_va_arena, (vaddr_t)v, pp->pr_alloc->pa_pagesz); } static void * pool_page_alloc_meta(struct pool *pp, int flags) { - bool waitok = (flags & PR_WAITOK) ? true : false; + const vm_flag_t vflags = (flags & PR_WAITOK) ? VM_SLEEP: VM_NOSLEEP; + vmem_addr_t va; + int ret; + + ret = vmem_alloc(kmem_meta_arena, pp->pr_alloc->pa_pagesz, + vflags | VM_INSTANTFIT, &va); - return ((void *) uvm_km_alloc_poolpage(kmem_map, waitok)); + return ret ? NULL : (void *)va; } static void pool_page_free_meta(struct pool *pp, void *v) { - uvm_km_free_poolpage(kmem_map, (vaddr_t) v); + vmem_free(kmem_meta_arena, (vmem_addr_t)v, pp->pr_alloc->pa_pagesz); } #ifdef POOL_SUBPAGE @@ -2880,35 +2844,7 @@ pool_subpage_free(struct pool *pp, void pool_put(&psppool, v); } -/* We don't provide a real nointr allocator. Maybe later. */ -void * -pool_subpage_alloc_nointr(struct pool *pp, int flags) -{ - - return (pool_subpage_alloc(pp, flags)); -} - -void -pool_subpage_free_nointr(struct pool *pp, void *v) -{ - - pool_subpage_free(pp, v); -} #endif /* POOL_SUBPAGE */ -void * -pool_page_alloc_nointr(struct pool *pp, int flags) -{ - bool waitok = (flags & PR_WAITOK) ? true : false; - - return ((void *) uvm_km_alloc_poolpage_cache(kernel_map, waitok)); -} - -void -pool_page_free_nointr(struct pool *pp, void *v) -{ - - uvm_km_free_poolpage_cache(kernel_map, (vaddr_t) v); -} #if defined(DDB) static bool @@ -3018,7 +2954,7 @@ found: goto print; } } - for (i = 0; i < MAXCPUS; i++) { + for (i = 0; i < __arraycount(pc->pc_cpus); i++) { pool_cache_cpu_t *cc; if ((cc = pc->pc_cpus[i]) == NULL) { @@ -3027,7 +2963,7 @@ found: if (pool_in_cg(pp, cc->cc_current, addr) || pool_in_cg(pp, cc->cc_previous, addr)) { struct cpu_info *ci = - cpu_lookup_byindex(i); + cpu_lookup(i); incpucache = true; snprintf(cpucachestr,