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 1.42 retrieving revision 1.52 diff -u -p -r1.42 -r1.52 --- src/sys/kern/subr_pool.c 2000/12/06 18:20:52 1.42 +++ src/sys/kern/subr_pool.c 2001/05/09 23:46:03 1.52 @@ -1,7 +1,7 @@ -/* $NetBSD: subr_pool.c,v 1.42 2000/12/06 18:20:52 thorpej Exp $ */ +/* $NetBSD: subr_pool.c,v 1.52 2001/05/09 23:46:03 thorpej Exp $ */ /*- - * Copyright (c) 1997, 1999 The NetBSD Foundation, Inc. + * Copyright (c) 1997, 1999, 2000 The NetBSD Foundation, Inc. * All rights reserved. * * This code is derived from software contributed to The NetBSD Foundation @@ -100,14 +100,52 @@ struct pool_item { TAILQ_ENTRY(pool_item) pi_list; }; - #define PR_HASH_INDEX(pp,addr) \ (((u_long)(addr) >> (pp)->pr_pageshift) & (PR_HASHTABSIZE - 1)) +/* + * Pool cache management. + * + * Pool caches provide a way for constructed objects to be cached by the + * pool subsystem. This can lead to performance improvements by avoiding + * needless object construction/destruction; it is deferred until absolutely + * necessary. + * + * Caches are grouped into cache groups. Each cache group references + * up to 16 constructed objects. When a cache allocates an object + * from the pool, it calls the object's constructor and places it into + * a cache group. When a cache group frees an object back to the pool, + * it first calls the object's destructor. This allows the object to + * persist in constructed form while freed to the cache. + * + * Multiple caches may exist for each pool. This allows a single + * object type to have multiple constructed forms. The pool references + * each cache, so that when a pool is drained by the pagedaemon, it can + * drain each individual cache as well. Each time a cache is drained, + * the most idle cache group is freed to the pool in its entirety. + * + * Pool caches are layed on top of pools. By layering them, we can avoid + * the complexity of cache management for pools which would not benefit + * from it. + */ +/* The cache group pool. */ +static struct pool pcgpool; + +/* The pool cache group. */ +#define PCG_NOBJECTS 16 +struct pool_cache_group { + TAILQ_ENTRY(pool_cache_group) + pcg_list; /* link in the pool cache's group list */ + u_int pcg_avail; /* # available objects */ + /* pointers to the objects */ + void *pcg_objects[PCG_NOBJECTS]; +}; + +static void pool_cache_reclaim(struct pool_cache *); static int pool_catchup(struct pool *); -static void pool_prime_page(struct pool *, caddr_t); +static int pool_prime_page(struct pool *, caddr_t, int); static void *pool_page_alloc(unsigned long, int, int); static void pool_page_free(void *, unsigned long, int); @@ -115,7 +153,7 @@ static void pool_print1(struct pool *, c void (*)(const char *, ...)); /* - * Pool log entry. An array of these is allocated in pool_create(). + * Pool log entry. An array of these is allocated in pool_init(). */ struct pool_log { const char *pl_file; @@ -306,37 +344,6 @@ pr_rmpage(struct pool *pp, struct pool_i } /* - * Allocate and initialize a pool. - */ -struct pool * -pool_create(size_t size, u_int align, u_int ioff, int nitems, - const char *wchan, size_t pagesz, - void *(*alloc)(unsigned long, int, int), - void (*release)(void *, unsigned long, int), - int mtype) -{ - struct pool *pp; - int flags; - - pp = (struct pool *)malloc(sizeof(*pp), M_POOL, M_NOWAIT); - if (pp == NULL) - return (NULL); - - flags = PR_FREEHEADER; - pool_init(pp, size, align, ioff, flags, wchan, pagesz, - alloc, release, mtype); - - if (nitems != 0) { - if (pool_prime(pp, nitems, NULL) != 0) { - pool_destroy(pp); - return (NULL); - } - } - - return (pp); -} - -/* * Initialize the given pool resource structure. * * We export this routine to allow other kernel parts to declare @@ -384,7 +391,7 @@ pool_init(struct pool *pp, size_t size, size = sizeof(struct pool_item); size = ALIGN(size); - if (size >= pagesz) + if (size > pagesz) panic("pool_init: pool item size (%lu) too large", (u_long)size); @@ -392,6 +399,7 @@ pool_init(struct pool *pp, size_t size, * Initialize the pool structure. */ TAILQ_INIT(&pp->pr_pagelist); + TAILQ_INIT(&pp->pr_cachelist); pp->pr_curpage = NULL; pp->pr_npages = 0; pp->pr_minitems = 0; @@ -447,6 +455,7 @@ pool_init(struct pool *pp, size_t size, */ pp->pr_itemoffset = ioff = ioff % align; pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size; + KASSERT(pp->pr_itemsperpage != 0); /* * Use the slack between the chunks and the page header @@ -479,12 +488,15 @@ pool_init(struct pool *pp, size_t size, simple_lock_init(&pp->pr_slock); /* - * Initialize private page header pool if we haven't done so yet. + * Initialize private page header pool and cache magazine pool if we + * haven't done so yet. * XXX LOCKING. */ if (phpool.pr_size == 0) { pool_init(&phpool, sizeof(struct pool_item_header), 0, 0, - 0, "phpool", 0, 0, 0, 0); + 0, "phpool", 0, 0, 0, 0); + pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0, + 0, "pcgpool", 0, 0, 0, 0); } /* Insert into the list of all pools. */ @@ -500,6 +512,11 @@ void pool_destroy(struct pool *pp) { struct pool_item_header *ph; + struct pool_cache *pc; + + /* Destroy all caches for this pool. */ + while ((pc = TAILQ_FIRST(&pp->pr_cachelist)) != NULL) + pool_cache_destroy(pc); #ifdef DIAGNOSTIC if (pp->pr_nout != 0) { @@ -666,8 +683,18 @@ _pool_get(struct pool *pp, int flags, co } /* We have more memory; add it to the pool */ + if (pool_prime_page(pp, v, flags & PR_WAITOK) != 0) { + /* + * Probably, we don't allowed to wait and + * couldn't allocate a page header. + */ + (*pp->pr_free)(v, pp->pr_pagesz, pp->pr_mtype); + pp->pr_nfail++; + pr_leave(pp); + simple_unlock(&pp->pr_slock); + return (NULL); + } pp->pr_npagealloc++; - pool_prime_page(pp, v); /* Start the allocation process over. */ goto startover; @@ -762,10 +789,10 @@ _pool_get(struct pool *pp, int flags, co } /* - * Return resource to the pool; must be called at appropriate spl level + * Internal version of pool_put(). Pool is already locked/entered. */ -void -_pool_put(struct pool *pp, void *v, const char *file, long line) +static void +pool_do_put(struct pool *pp, void *v, const char *file, long line) { struct pool_item *pi = v; struct pool_item_header *ph; @@ -774,9 +801,6 @@ _pool_put(struct pool *pp, void *v, cons page = (caddr_t)((u_long)v & pp->pr_pagemask); - simple_lock(&pp->pr_slock); - pr_enter(pp, file, line); - #ifdef DIAGNOSTIC if (__predict_false(pp->pr_nout == 0)) { printf("pool %s: putting with none out\n", @@ -829,8 +853,6 @@ _pool_put(struct pool *pp, void *v, cons pp->pr_flags &= ~PR_WANTED; if (ph->ph_nmissing == 0) pp->pr_nidle++; - pr_leave(pp); - simple_unlock(&pp->pr_slock); wakeup((caddr_t)pp); return; } @@ -894,61 +916,22 @@ _pool_put(struct pool *pp, void *v, cons TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist); pp->pr_curpage = ph; } - - pr_leave(pp); - simple_unlock(&pp->pr_slock); - } /* - * Add N items to the pool. + * Return resource to the pool; must be called at appropriate spl level */ -int -pool_prime(struct pool *pp, int n, caddr_t storage) +void +_pool_put(struct pool *pp, void *v, const char *file, long line) { - caddr_t cp; - int newnitems, newpages; - -#ifdef DIAGNOSTIC - if (__predict_false(storage && !(pp->pr_roflags & PR_STATIC))) - panic("pool_prime: static"); - /* !storage && static caught below */ -#endif simple_lock(&pp->pr_slock); + pr_enter(pp, file, line); - newnitems = pp->pr_minitems + n; - newpages = - roundup(newnitems, pp->pr_itemsperpage) / pp->pr_itemsperpage - - pp->pr_minpages; - - while (newpages-- > 0) { - if (pp->pr_roflags & PR_STATIC) { - cp = storage; - storage += pp->pr_pagesz; - } else { - simple_unlock(&pp->pr_slock); - cp = (*pp->pr_alloc)(pp->pr_pagesz, 0, pp->pr_mtype); - simple_lock(&pp->pr_slock); - } - - if (cp == NULL) { - simple_unlock(&pp->pr_slock); - return (ENOMEM); - } - - pp->pr_npagealloc++; - pool_prime_page(pp, cp); - pp->pr_minpages++; - } - - pp->pr_minitems = newnitems; - - if (pp->pr_minpages >= pp->pr_maxpages) - pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */ + pool_do_put(pp, v, file, line); + pr_leave(pp); simple_unlock(&pp->pr_slock); - return (0); } /* @@ -956,8 +939,8 @@ pool_prime(struct pool *pp, int n, caddr * * Note, we must be called with the pool descriptor LOCKED. */ -static void -pool_prime_page(struct pool *pp, caddr_t storage) +static int +pool_prime_page(struct pool *pp, caddr_t storage, int flags) { struct pool_item *pi; struct pool_item_header *ph; @@ -973,8 +956,10 @@ pool_prime_page(struct pool *pp, caddr_t ph = (struct pool_item_header *)(cp + pp->pr_phoffset); } else { s = splhigh(); - ph = pool_get(&phpool, PR_URGENT); + ph = pool_get(&phpool, flags); splx(s); + if (ph == NULL) + return (ENOMEM); LIST_INSERT_HEAD(&pp->pr_hashtab[PR_HASH_INDEX(pp, cp)], ph, ph_hashlist); } @@ -1028,12 +1013,13 @@ pool_prime_page(struct pool *pp, caddr_t if (++pp->pr_npages > pp->pr_hiwat) pp->pr_hiwat = pp->pr_npages; + + return (0); } /* - * Like pool_prime(), except this is used by pool_get() when nitems - * drops below the low water mark. This is used to catch up nitmes - * with the low water mark. + * Used by pool_get() when nitems drops below the low water mark. This + * is used to catch up nitmes with the low water mark. * * Note 1, we never wait for memory here, we let the caller decide what to do. * @@ -1074,8 +1060,11 @@ pool_catchup(struct pool *pp) error = ENOMEM; break; } + if ((error = pool_prime_page(pp, cp, PR_NOWAIT)) != 0) { + (*pp->pr_free)(cp, pp->pr_pagesz, pp->pr_mtype); + break; + } pp->pr_npagealloc++; - pool_prime_page(pp, cp); } return (error); @@ -1188,6 +1177,7 @@ void _pool_reclaim(struct pool *pp, const char *file, long line) { struct pool_item_header *ph, *phnext; + struct pool_cache *pc; struct timeval curtime; int s; @@ -1198,6 +1188,13 @@ _pool_reclaim(struct pool *pp, const cha return; pr_enter(pp, file, line); + /* + * Reclaim items from the pool's caches. + */ + for (pc = TAILQ_FIRST(&pp->pr_cachelist); pc != NULL; + pc = TAILQ_NEXT(pc, pc_poollist)) + pool_cache_reclaim(pc); + s = splclock(); curtime = mono_time; splx(s); @@ -1243,7 +1240,7 @@ pool_drain(void *arg) struct pool *pp; int s; - s = splimp(); + s = splvm(); simple_lock(&pool_head_slock); if (drainpp == NULL && (drainpp = TAILQ_FIRST(&pool_head)) == NULL) @@ -1268,7 +1265,7 @@ pool_print(struct pool *pp, const char * { int s; - s = splimp(); + s = splvm(); if (simple_lock_try(&pp->pr_slock) == 0) { printf("pool %s is locked; try again later\n", pp->pr_wchan); @@ -1314,10 +1311,12 @@ static void pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...)) { struct pool_item_header *ph; + struct pool_cache *pc; + struct pool_cache_group *pcg; #ifdef DIAGNOSTIC struct pool_item *pi; #endif - int print_log = 0, print_pagelist = 0; + int i, print_log = 0, print_pagelist = 0, print_cache = 0; char c; while ((c = *modif++) != '\0') { @@ -1325,6 +1324,8 @@ pool_print1(struct pool *pp, const char print_log = 1; if (c == 'p') print_pagelist = 1; + if (c == 'c') + print_cache = 1; modif++; } @@ -1381,6 +1382,25 @@ pool_print1(struct pool *pp, const char skip_log: + if (print_cache == 0) + goto skip_cache; + + for (pc = TAILQ_FIRST(&pp->pr_cachelist); pc != NULL; + pc = TAILQ_NEXT(pc, pc_poollist)) { + (*pr)("\tcache %p: allocfrom %p freeto %p\n", pc, + pc->pc_allocfrom, pc->pc_freeto); + (*pr)("\t hits %lu misses %lu ngroups %lu nitems %lu\n", + pc->pc_hits, pc->pc_misses, pc->pc_ngroups, pc->pc_nitems); + for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL; + pcg = TAILQ_NEXT(pcg, pcg_list)) { + (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail); + for (i = 0; i < PCG_NOBJECTS; i++) + (*pr)("\t\t\t%p\n", pcg->pcg_objects[i]); + } + } + + skip_cache: + pr_enter_check(pp, pr); } @@ -1446,3 +1466,276 @@ out: simple_unlock(&pp->pr_slock); return (r); } + +/* + * pool_cache_init: + * + * Initialize a pool cache. + * + * NOTE: If the pool must be protected from interrupts, we expect + * to be called at the appropriate interrupt priority level. + */ +void +pool_cache_init(struct pool_cache *pc, struct pool *pp, + int (*ctor)(void *, void *, int), + void (*dtor)(void *, void *), + void *arg) +{ + + TAILQ_INIT(&pc->pc_grouplist); + simple_lock_init(&pc->pc_slock); + + pc->pc_allocfrom = NULL; + pc->pc_freeto = NULL; + pc->pc_pool = pp; + + pc->pc_ctor = ctor; + pc->pc_dtor = dtor; + pc->pc_arg = arg; + + pc->pc_hits = 0; + pc->pc_misses = 0; + + pc->pc_ngroups = 0; + + pc->pc_nitems = 0; + + simple_lock(&pp->pr_slock); + TAILQ_INSERT_TAIL(&pp->pr_cachelist, pc, pc_poollist); + simple_unlock(&pp->pr_slock); +} + +/* + * pool_cache_destroy: + * + * Destroy a pool cache. + */ +void +pool_cache_destroy(struct pool_cache *pc) +{ + struct pool *pp = pc->pc_pool; + + /* First, invalidate the entire cache. */ + pool_cache_invalidate(pc); + + /* ...and remove it from the pool's cache list. */ + simple_lock(&pp->pr_slock); + TAILQ_REMOVE(&pp->pr_cachelist, pc, pc_poollist); + simple_unlock(&pp->pr_slock); +} + +static __inline void * +pcg_get(struct pool_cache_group *pcg) +{ + void *object; + u_int idx; + + KASSERT(pcg->pcg_avail <= PCG_NOBJECTS); + KASSERT(pcg->pcg_avail != 0); + idx = --pcg->pcg_avail; + + KASSERT(pcg->pcg_objects[idx] != NULL); + object = pcg->pcg_objects[idx]; + pcg->pcg_objects[idx] = NULL; + + return (object); +} + +static __inline void +pcg_put(struct pool_cache_group *pcg, void *object) +{ + u_int idx; + + KASSERT(pcg->pcg_avail < PCG_NOBJECTS); + idx = pcg->pcg_avail++; + + KASSERT(pcg->pcg_objects[idx] == NULL); + pcg->pcg_objects[idx] = object; +} + +/* + * pool_cache_get: + * + * Get an object from a pool cache. + */ +void * +pool_cache_get(struct pool_cache *pc, int flags) +{ + struct pool_cache_group *pcg; + void *object; + + simple_lock(&pc->pc_slock); + + if ((pcg = pc->pc_allocfrom) == NULL) { + for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL; + pcg = TAILQ_NEXT(pcg, pcg_list)) { + if (pcg->pcg_avail != 0) { + pc->pc_allocfrom = pcg; + goto have_group; + } + } + + /* + * No groups with any available objects. Allocate + * a new object, construct it, and return it to + * the caller. We will allocate a group, if necessary, + * when the object is freed back to the cache. + */ + pc->pc_misses++; + simple_unlock(&pc->pc_slock); + object = pool_get(pc->pc_pool, flags); + if (object != NULL && pc->pc_ctor != NULL) { + if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) { + pool_put(pc->pc_pool, object); + return (NULL); + } + } + return (object); + } + + have_group: + pc->pc_hits++; + pc->pc_nitems--; + object = pcg_get(pcg); + + if (pcg->pcg_avail == 0) + pc->pc_allocfrom = NULL; + + simple_unlock(&pc->pc_slock); + + return (object); +} + +/* + * pool_cache_put: + * + * Put an object back to the pool cache. + */ +void +pool_cache_put(struct pool_cache *pc, void *object) +{ + struct pool_cache_group *pcg; + + simple_lock(&pc->pc_slock); + + if ((pcg = pc->pc_freeto) == NULL) { + for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL; + pcg = TAILQ_NEXT(pcg, pcg_list)) { + if (pcg->pcg_avail != PCG_NOBJECTS) { + pc->pc_freeto = pcg; + goto have_group; + } + } + + /* + * No empty groups to free the object to. Attempt to + * allocate one. + */ + simple_unlock(&pc->pc_slock); + pcg = pool_get(&pcgpool, PR_NOWAIT); + if (pcg != NULL) { + memset(pcg, 0, sizeof(*pcg)); + simple_lock(&pc->pc_slock); + pc->pc_ngroups++; + TAILQ_INSERT_TAIL(&pc->pc_grouplist, pcg, pcg_list); + if (pc->pc_freeto == NULL) + pc->pc_freeto = pcg; + goto have_group; + } + + /* + * Unable to allocate a cache group; destruct the object + * and free it back to the pool. + */ + pool_cache_destruct_object(pc, object); + return; + } + + have_group: + pc->pc_nitems++; + pcg_put(pcg, object); + + if (pcg->pcg_avail == PCG_NOBJECTS) + pc->pc_freeto = NULL; + + simple_unlock(&pc->pc_slock); +} + +/* + * pool_cache_destruct_object: + * + * Force destruction of an object and its release back into + * the pool. + */ +void +pool_cache_destruct_object(struct pool_cache *pc, void *object) +{ + + if (pc->pc_dtor != NULL) + (*pc->pc_dtor)(pc->pc_arg, object); + pool_put(pc->pc_pool, object); +} + +/* + * pool_cache_do_invalidate: + * + * This internal function implements pool_cache_invalidate() and + * pool_cache_reclaim(). + */ +static void +pool_cache_do_invalidate(struct pool_cache *pc, int free_groups, + void (*putit)(struct pool *, void *, const char *, long)) +{ + struct pool_cache_group *pcg, *npcg; + void *object; + + for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL; + pcg = npcg) { + npcg = TAILQ_NEXT(pcg, pcg_list); + while (pcg->pcg_avail != 0) { + pc->pc_nitems--; + object = pcg_get(pcg); + if (pcg->pcg_avail == 0 && pc->pc_allocfrom == pcg) + pc->pc_allocfrom = NULL; + if (pc->pc_dtor != NULL) + (*pc->pc_dtor)(pc->pc_arg, object); + (*putit)(pc->pc_pool, object, __FILE__, __LINE__); + } + if (free_groups) { + pc->pc_ngroups--; + TAILQ_REMOVE(&pc->pc_grouplist, pcg, pcg_list); + if (pc->pc_freeto == pcg) + pc->pc_freeto = NULL; + pool_put(&pcgpool, pcg); + } + } +} + +/* + * pool_cache_invalidate: + * + * Invalidate a pool cache (destruct and release all of the + * cached objects). + */ +void +pool_cache_invalidate(struct pool_cache *pc) +{ + + simple_lock(&pc->pc_slock); + pool_cache_do_invalidate(pc, 0, _pool_put); + simple_unlock(&pc->pc_slock); +} + +/* + * pool_cache_reclaim: + * + * Reclaim a pool cache for pool_reclaim(). + */ +static void +pool_cache_reclaim(struct pool_cache *pc) +{ + + simple_lock(&pc->pc_slock); + pool_cache_do_invalidate(pc, 1, pool_do_put); + simple_unlock(&pc->pc_slock); +}