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 1.50.2.8 retrieving revision 1.51 diff -u -p -r1.50.2.8 -r1.51 --- src/sys/kern/subr_pool.c 2002/06/24 22:10:56 1.50.2.8 +++ src/sys/kern/subr_pool.c 2001/05/04 19:41:26 1.51 @@ -1,4 +1,4 @@ -/* $NetBSD: subr_pool.c,v 1.50.2.8 2002/06/24 22:10:56 nathanw Exp $ */ +/* $NetBSD: subr_pool.c,v 1.51 2001/05/04 19:41:26 thorpej Exp $ */ /*- * Copyright (c) 1997, 1999, 2000 The NetBSD Foundation, Inc. @@ -37,9 +37,6 @@ * POSSIBILITY OF SUCH DAMAGE. */ -#include -__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.50.2.8 2002/06/24 22:10:56 nathanw Exp $"); - #include "opt_pool.h" #include "opt_poollog.h" #include "opt_lockdebug.h" @@ -73,11 +70,6 @@ TAILQ_HEAD(,pool) pool_head = TAILQ_HEAD /* Private pool for page header structures */ static struct pool phpool; -#ifdef POOL_SUBPAGE -/* Pool of subpages for use by normal pools. */ -static struct pool psppool; -#endif - /* # of seconds to retain page after last use */ int pool_inactive_time = 10; @@ -98,7 +90,6 @@ struct pool_item_header { caddr_t ph_page; /* this page's address */ struct timeval ph_time; /* last referenced */ }; -TAILQ_HEAD(pool_pagelist,pool_item_header); struct pool_item { #ifdef DIAGNOSTIC @@ -110,11 +101,7 @@ struct pool_item { }; #define PR_HASH_INDEX(pp,addr) \ - (((u_long)(addr) >> (pp)->pr_alloc->pa_pageshift) & \ - (PR_HASHTABSIZE - 1)) - -#define POOL_NEEDS_CATCHUP(pp) \ - ((pp)->pr_nitems < (pp)->pr_minitems) + (((u_long)(addr) >> (pp)->pr_pageshift) & (PR_HASHTABSIZE - 1)) /* * Pool cache management. @@ -145,20 +132,28 @@ struct pool_item { /* 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, - struct pool_item_header *); - -void *pool_allocator_alloc(struct pool *, int); -void pool_allocator_free(struct pool *, void *); +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); static void pool_print1(struct pool *, const char *, void (*)(const char *, ...)); /* - * Pool log entry. An array of these is allocated in pool_init(). + * Pool log entry. An array of these is allocated in pool_create(). */ struct pool_log { const char *pl_file; @@ -176,7 +171,7 @@ struct pool_log { int pool_logsize = POOL_LOGSIZE; -#ifdef POOL_DIAGNOSTIC +#ifdef DIAGNOSTIC static __inline void pr_log(struct pool *pp, void *v, int action, const char *file, long line) { @@ -273,7 +268,7 @@ pr_enter_check(struct pool *pp, void (*p #define pr_enter(pp, file, line) #define pr_leave(pp) #define pr_enter_check(pp, pr) -#endif /* POOL_DIAGNOSTIC */ +#endif /* DIAGNOSTIC */ /* * Return the pool page header based on page address. @@ -299,10 +294,8 @@ pr_find_pagehead(struct pool *pp, caddr_ * Remove a page from the pool. */ static __inline void -pr_rmpage(struct pool *pp, struct pool_item_header *ph, - struct pool_pagelist *pq) +pr_rmpage(struct pool *pp, struct pool_item_header *ph) { - int s; /* * If the page was idle, decrement the idle page count. @@ -320,30 +313,29 @@ pr_rmpage(struct pool *pp, struct pool_i pp->pr_nitems -= pp->pr_itemsperpage; /* - * Unlink a page from the pool and release it (or queue it for release). + * Unlink a page from the pool and release it. */ TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist); - if (pq) { - TAILQ_INSERT_HEAD(pq, ph, ph_pagelist); - } else { - pool_allocator_free(pp, ph->ph_page); - if ((pp->pr_roflags & PR_PHINPAGE) == 0) { - LIST_REMOVE(ph, ph_hashlist); - s = splhigh(); - pool_put(&phpool, ph); - splx(s); - } - } + (*pp->pr_free)(ph->ph_page, pp->pr_pagesz, pp->pr_mtype); pp->pr_npages--; pp->pr_npagefree++; + if ((pp->pr_roflags & PR_PHINPAGE) == 0) { + int s; + LIST_REMOVE(ph, ph_hashlist); + s = splhigh(); + pool_put(&phpool, ph); + splx(s); + } + if (pp->pr_curpage == ph) { /* * Find a new non-empty page header, if any. * Start search from the page head, to increase the * chance for "high water" pages to be freed. */ - TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist) + for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; + ph = TAILQ_NEXT(ph, ph_pagelist)) if (TAILQ_FIRST(&ph->ph_itemlist) != NULL) break; @@ -352,6 +344,37 @@ 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 @@ -359,7 +382,10 @@ pr_rmpage(struct pool *pp, struct pool_i */ void pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags, - const char *wchan, struct pool_allocator *palloc) + const char *wchan, size_t pagesz, + void *(*alloc)(unsigned long, int, int), + void (*release)(void *, unsigned long, int), + int mtype) { int off, slack, i; @@ -371,43 +397,23 @@ pool_init(struct pool *pp, size_t size, flags |= PR_LOGGING; #endif -#ifdef POOL_SUBPAGE - /* - * XXX We don't provide a real `nointr' back-end - * yet; all sub-pages come from a kmem back-end. - * maybe some day... - */ - if (palloc == NULL) { - extern struct pool_allocator pool_allocator_kmem_subpage; - palloc = &pool_allocator_kmem_subpage; - } /* - * We'll assume any user-specified back-end allocator - * will deal with sub-pages, or simply don't care. + * Check arguments and construct default values. */ -#else - if (palloc == NULL) - palloc = &pool_allocator_kmem; -#endif /* POOL_SUBPAGE */ - if ((palloc->pa_flags & PA_INITIALIZED) == 0) { - if (palloc->pa_pagesz == 0) { -#ifdef POOL_SUBPAGE - if (palloc == &pool_allocator_kmem) - palloc->pa_pagesz = PAGE_SIZE; - else - palloc->pa_pagesz = POOL_SUBPAGE; -#else - palloc->pa_pagesz = PAGE_SIZE; -#endif /* POOL_SUBPAGE */ - } - - TAILQ_INIT(&palloc->pa_list); + if (!powerof2(pagesz)) + panic("pool_init: page size invalid (%lx)\n", (u_long)pagesz); - simple_lock_init(&palloc->pa_slock); - palloc->pa_pagemask = ~(palloc->pa_pagesz - 1); - palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1; - palloc->pa_flags |= PA_INITIALIZED; - } + if (alloc == NULL && release == NULL) { + alloc = pool_page_alloc; + release = pool_page_free; + pagesz = PAGE_SIZE; /* Rounds to PAGE_SIZE anyhow. */ + } else if ((alloc != NULL && release != NULL) == 0) { + /* If you specifiy one, must specify both. */ + panic("pool_init: must specify alloc and release together"); + } + + if (pagesz == 0) + pagesz = PAGE_SIZE; if (align == 0) align = ALIGN(1); @@ -416,11 +422,9 @@ pool_init(struct pool *pp, size_t size, size = sizeof(struct pool_item); size = ALIGN(size); -#ifdef DIAGNOSTIC - if (size > palloc->pa_pagesz) + if (size > pagesz) panic("pool_init: pool item size (%lu) too large", (u_long)size); -#endif /* * Initialize the pool structure. @@ -437,7 +441,12 @@ pool_init(struct pool *pp, size_t size, pp->pr_size = size; pp->pr_align = align; pp->pr_wchan = wchan; - pp->pr_alloc = palloc; + pp->pr_mtype = mtype; + pp->pr_alloc = alloc; + pp->pr_free = release; + pp->pr_pagesz = pagesz; + pp->pr_pagemask = ~(pagesz - 1); + pp->pr_pageshift = ffs(pagesz) - 1; pp->pr_nitems = 0; pp->pr_nout = 0; pp->pr_hardlimit = UINT_MAX; @@ -446,8 +455,6 @@ pool_init(struct pool *pp, size_t size, pp->pr_hardlimit_ratecap.tv_usec = 0; pp->pr_hardlimit_warning_last.tv_sec = 0; pp->pr_hardlimit_warning_last.tv_usec = 0; - pp->pr_drain_hook = NULL; - pp->pr_drain_hook_arg = NULL; /* * Decide whether to put the page header off page to avoid @@ -456,15 +463,15 @@ pool_init(struct pool *pp, size_t size, * with its header based on the page address. * We use 1/16 of the page size as the threshold (XXX: tune) */ - if (pp->pr_size < palloc->pa_pagesz/16) { + if (pp->pr_size < pagesz/16) { /* Use the end of the page for the page header */ pp->pr_roflags |= PR_PHINPAGE; - pp->pr_phoffset = off = palloc->pa_pagesz - - ALIGN(sizeof(struct pool_item_header)); + pp->pr_phoffset = off = + pagesz - ALIGN(sizeof(struct pool_item_header)); } else { /* The page header will be taken from our page header pool */ pp->pr_phoffset = 0; - off = palloc->pa_pagesz; + off = pagesz; for (i = 0; i < PR_HASHTABSIZE; i++) { LIST_INIT(&pp->pr_hashtab[i]); } @@ -497,7 +504,6 @@ pool_init(struct pool *pp, size_t size, pp->pr_hiwat = 0; pp->pr_nidle = 0; -#ifdef POOL_DIAGNOSTIC if (flags & PR_LOGGING) { if (kmem_map == NULL || (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log), @@ -506,7 +512,6 @@ pool_init(struct pool *pp, size_t size, pp->pr_curlogentry = 0; pp->pr_logsize = pool_logsize; } -#endif pp->pr_entered_file = NULL; pp->pr_entered_line = 0; @@ -519,28 +524,16 @@ pool_init(struct pool *pp, size_t size, * XXX LOCKING. */ if (phpool.pr_size == 0) { -#ifdef POOL_SUBPAGE - pool_init(&phpool, sizeof(struct pool_item_header), 0, 0, 0, - "phpool", &pool_allocator_kmem); - pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0, - PR_RECURSIVE, "psppool", &pool_allocator_kmem); -#else pool_init(&phpool, sizeof(struct pool_item_header), 0, 0, - 0, "phpool", NULL); -#endif + 0, "phpool", 0, 0, 0, 0); pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0, - 0, "pcgpool", NULL); + 0, "pcgpool", 0, 0, 0, 0); } /* Insert into the list of all pools. */ simple_lock(&pool_head_slock); TAILQ_INSERT_TAIL(&pool_head, pp, pr_poollist); simple_unlock(&pool_head_slock); - - /* Insert this into the list of pools using this allocator. */ - simple_lock(&palloc->pa_slock); - TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list); - simple_unlock(&palloc->pa_slock); } /* @@ -552,11 +545,6 @@ pool_destroy(struct pool *pp) struct pool_item_header *ph; struct pool_cache *pc; - /* Locking order: pool_allocator -> pool */ - simple_lock(&pp->pr_alloc->pa_slock); - TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list); - simple_unlock(&pp->pr_alloc->pa_slock); - /* Destroy all caches for this pool. */ while ((pc = TAILQ_FIRST(&pp->pr_cachelist)) != NULL) pool_cache_destroy(pc); @@ -570,80 +558,47 @@ pool_destroy(struct pool *pp) #endif /* Remove all pages */ - while ((ph = TAILQ_FIRST(&pp->pr_pagelist)) != NULL) - pr_rmpage(pp, ph, NULL); + if ((pp->pr_roflags & PR_STATIC) == 0) + while ((ph = pp->pr_pagelist.tqh_first) != NULL) + pr_rmpage(pp, ph); /* Remove from global pool list */ simple_lock(&pool_head_slock); TAILQ_REMOVE(&pool_head, pp, pr_poollist); - if (drainpp == pp) { - drainpp = NULL; - } + /* XXX Only clear this if we were drainpp? */ + drainpp = NULL; simple_unlock(&pool_head_slock); -#ifdef POOL_DIAGNOSTIC if ((pp->pr_roflags & PR_LOGGING) != 0) free(pp->pr_log, M_TEMP); -#endif -} - -void -pool_set_drain_hook(struct pool *pp, void (*fn)(void *, int), void *arg) -{ - /* XXX no locking -- must be used just after pool_init() */ -#ifdef DIAGNOSTIC - if (pp->pr_drain_hook != NULL) - panic("pool_set_drain_hook(%s): already set", pp->pr_wchan); -#endif - pp->pr_drain_hook = fn; - pp->pr_drain_hook_arg = arg; + if (pp->pr_roflags & PR_FREEHEADER) + free(pp, M_POOL); } -static __inline struct pool_item_header * -pool_alloc_item_header(struct pool *pp, caddr_t storage, int flags) -{ - struct pool_item_header *ph; - int s; - - LOCK_ASSERT(simple_lock_held(&pp->pr_slock) == 0); - - if ((pp->pr_roflags & PR_PHINPAGE) != 0) - ph = (struct pool_item_header *) (storage + pp->pr_phoffset); - else { - s = splhigh(); - ph = pool_get(&phpool, flags); - splx(s); - } - - return (ph); -} /* * Grab an item from the pool; must be called at appropriate spl level */ void * -#ifdef POOL_DIAGNOSTIC _pool_get(struct pool *pp, int flags, const char *file, long line) -#else -pool_get(struct pool *pp, int flags) -#endif { + void *v; struct pool_item *pi; struct pool_item_header *ph; - void *v; #ifdef DIAGNOSTIC - if (__predict_false(curlwp == NULL && doing_shutdown == 0 && + if (__predict_false((pp->pr_roflags & PR_STATIC) && + (flags & PR_MALLOCOK))) { + pr_printlog(pp, NULL, printf); + panic("pool_get: static"); + } +#endif + + if (__predict_false(curproc == NULL && doing_shutdown == 0 && (flags & PR_WAITOK) != 0)) panic("pool_get: must have NOWAIT"); -#ifdef LOCKDEBUG - if (flags & PR_WAITOK) - simple_lock_only_held(NULL, "pool_get(PR_WAITOK)"); -#endif -#endif /* DIAGNOSTIC */ - simple_lock(&pp->pr_slock); pr_enter(pp, file, line); @@ -661,21 +616,6 @@ pool_get(struct pool *pp, int flags) } #endif if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) { - if (pp->pr_drain_hook != NULL) { - /* - * Since the drain hook is going to free things - * back to the pool, unlock, call the hook, re-lock, - * and check the hardlimit condition again. - */ - pr_leave(pp); - simple_unlock(&pp->pr_slock); - (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags); - simple_lock(&pp->pr_slock); - pr_enter(pp, file, line); - if (pp->pr_nout < pp->pr_hardlimit) - goto startover; - } - if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) { /* * XXX: A warning isn't logged in this case. Should @@ -696,6 +636,9 @@ pool_get(struct pool *pp, int flags) &pp->pr_hardlimit_ratecap)) log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning); + if (flags & PR_URGENT) + panic("pool_get: urgent"); + pp->pr_nfail++; pr_leave(pp); @@ -710,6 +653,8 @@ pool_get(struct pool *pp, int flags) * has no items in its bucket. */ if ((ph = pp->pr_curpage) == NULL) { + void *v; + #ifdef DIAGNOSTIC if (pp->pr_nitems != 0) { simple_unlock(&pp->pr_slock); @@ -726,25 +671,23 @@ pool_get(struct pool *pp, int flags) */ pr_leave(pp); simple_unlock(&pp->pr_slock); - v = pool_allocator_alloc(pp, flags); - if (__predict_true(v != NULL)) - ph = pool_alloc_item_header(pp, v, flags); + v = (*pp->pr_alloc)(pp->pr_pagesz, flags, pp->pr_mtype); simple_lock(&pp->pr_slock); pr_enter(pp, file, line); - if (__predict_false(v == NULL || ph == NULL)) { - if (v != NULL) - pool_allocator_free(pp, v); - + if (v == NULL) { /* - * We were unable to allocate a page or item - * header, but we released the lock during - * allocation, so perhaps items were freed - * back to the pool. Check for this case. + * We were unable to allocate a page, but + * we released the lock during allocation, + * so perhaps items were freed back to the + * pool. Check for this case. */ if (pp->pr_curpage != NULL) goto startover; + if (flags & PR_URGENT) + panic("pool_get: urgent"); + if ((flags & PR_WAITOK) == 0) { pp->pr_nfail++; pr_leave(pp); @@ -755,11 +698,15 @@ pool_get(struct pool *pp, int flags) /* * Wait for items to be returned to this pool. * + * XXX: we actually want to wait just until + * the page allocator has memory again. Depending + * on this pool's usage, we might get stuck here + * for a long time. + * * XXX: maybe we should wake up once a second and * try again? */ pp->pr_flags |= PR_WANTED; - /* PA_WANTED is already set on the allocator. */ pr_leave(pp); ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock); pr_enter(pp, file, line); @@ -767,7 +714,17 @@ pool_get(struct pool *pp, int flags) } /* We have more memory; add it to the pool */ - pool_prime_page(pp, v, ph); + 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++; /* Start the allocation process over. */ @@ -788,10 +745,7 @@ pool_get(struct pool *pp, int flags) panic("pool_get: nitems inconsistent\n"); } #endif - -#ifdef POOL_DIAGNOSTIC pr_log(pp, v, PRLOG_GET, file, line); -#endif #ifdef DIAGNOSTIC if (__predict_false(pi->pi_magic != PI_MAGIC)) { @@ -838,7 +792,8 @@ pool_get(struct pool *pp, int flags) */ TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist); TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist); - TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist) + for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; + ph = TAILQ_NEXT(ph, ph_pagelist)) if (TAILQ_FIRST(&ph->ph_itemlist) != NULL) break; @@ -851,7 +806,7 @@ pool_get(struct pool *pp, int flags) * If we have a low water mark and we are now below that low * water mark, add more items to the pool. */ - if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) { + if (pp->pr_nitems < pp->pr_minitems && pool_catchup(pp) != 0) { /* * XXX: Should we log a warning? Should we set up a timeout * to try again in a second or so? The latter could break @@ -868,16 +823,14 @@ pool_get(struct pool *pp, int flags) * Internal version of pool_put(). Pool is already locked/entered. */ static void -pool_do_put(struct pool *pp, void *v) +pool_do_put(struct pool *pp, void *v, const char *file, long line) { struct pool_item *pi = v; struct pool_item_header *ph; caddr_t page; int s; - LOCK_ASSERT(simple_lock_held(&pp->pr_slock)); - - page = (caddr_t)((u_long)v & pp->pr_alloc->pa_pagemask); + page = (caddr_t)((u_long)v & pp->pr_pagemask); #ifdef DIAGNOSTIC if (__predict_false(pp->pr_nout == 0)) { @@ -887,6 +840,8 @@ pool_do_put(struct pool *pp, void *v) } #endif + pr_log(pp, v, PRLOG_PUT, file, line); + if (__predict_false((ph = pr_find_pagehead(pp, page)) == NULL)) { pr_printlog(pp, NULL, printf); panic("pool_put: %s: page header missing", pp->pr_wchan); @@ -949,9 +904,8 @@ pool_do_put(struct pool *pp, void *v) */ if (ph->ph_nmissing == 0) { pp->pr_nidle++; - if (pp->pr_npages > pp->pr_maxpages || - (pp->pr_alloc->pa_flags & PA_WANT) != 0) { - pr_rmpage(pp, ph, NULL); + if (pp->pr_npages > pp->pr_maxpages) { + pr_rmpage(pp, ph); } else { TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist); TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist); @@ -974,7 +928,8 @@ pool_do_put(struct pool *pp, void *v) * page with the fewest available items, to minimize * fragmentation? */ - TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist) + for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; + ph = TAILQ_NEXT(ph, ph_pagelist)) if (TAILQ_FIRST(&ph->ph_itemlist) != NULL) break; @@ -997,7 +952,6 @@ pool_do_put(struct pool *pp, void *v) /* * Return resource to the pool; must be called at appropriate spl level */ -#ifdef POOL_DIAGNOSTIC void _pool_put(struct pool *pp, void *v, const char *file, long line) { @@ -1005,63 +959,62 @@ _pool_put(struct pool *pp, void *v, cons simple_lock(&pp->pr_slock); pr_enter(pp, file, line); - pr_log(pp, v, PRLOG_PUT, file, line); - - pool_do_put(pp, v); + pool_do_put(pp, v, file, line); pr_leave(pp); simple_unlock(&pp->pr_slock); } -#undef pool_put -#endif /* POOL_DIAGNOSTIC */ - -void -pool_put(struct pool *pp, void *v) -{ - - simple_lock(&pp->pr_slock); - - pool_do_put(pp, v); - - simple_unlock(&pp->pr_slock); -} - -#ifdef POOL_DIAGNOSTIC -#define pool_put(h, v) _pool_put((h), (v), __FILE__, __LINE__) -#endif /* * Add N items to the pool. */ int -pool_prime(struct pool *pp, int n) +pool_prime(struct pool *pp, int n, caddr_t storage) { - struct pool_item_header *ph; caddr_t cp; - int newpages; + int error, 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); - newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage; + newnitems = pp->pr_minitems + n; + newpages = + roundup(newnitems, pp->pr_itemsperpage) / pp->pr_itemsperpage + - pp->pr_minpages; while (newpages-- > 0) { - simple_unlock(&pp->pr_slock); - cp = pool_allocator_alloc(pp, PR_NOWAIT); - if (__predict_true(cp != NULL)) - ph = pool_alloc_item_header(pp, cp, PR_NOWAIT); - simple_lock(&pp->pr_slock); + 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 (__predict_false(cp == NULL || ph == NULL)) { - if (cp != NULL) - pool_allocator_free(pp, cp); - break; + if (cp == NULL) { + simple_unlock(&pp->pr_slock); + return (ENOMEM); } - pool_prime_page(pp, cp, ph); + if ((error = pool_prime_page(pp, cp, PR_NOWAIT)) != 0) { + if ((pp->pr_roflags & PR_STATIC) == 0) + (*pp->pr_free)(cp, pp->pr_pagesz, + pp->pr_mtype); + simple_unlock(&pp->pr_slock); + return (error); + } pp->pr_npagealloc++; pp->pr_minpages++; } + pp->pr_minitems = newnitems; + if (pp->pr_minpages >= pp->pr_maxpages) pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */ @@ -1074,23 +1027,30 @@ pool_prime(struct pool *pp, int n) * * Note, we must be called with the pool descriptor LOCKED. */ -static void -pool_prime_page(struct pool *pp, caddr_t storage, struct pool_item_header *ph) +static int +pool_prime_page(struct pool *pp, caddr_t storage, int flags) { struct pool_item *pi; + struct pool_item_header *ph; caddr_t cp = storage; unsigned int align = pp->pr_align; unsigned int ioff = pp->pr_itemoffset; - int n; + int s, n; -#ifdef DIAGNOSTIC - if (((u_long)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0) + if (((u_long)cp & (pp->pr_pagesz - 1)) != 0) panic("pool_prime_page: %s: unaligned page", pp->pr_wchan); -#endif - if ((pp->pr_roflags & PR_PHINPAGE) == 0) + if ((pp->pr_roflags & PR_PHINPAGE) != 0) { + ph = (struct pool_item_header *)(cp + pp->pr_phoffset); + } else { + s = splhigh(); + 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); + ph, ph_hashlist); + } /* * Insert page header. @@ -1141,25 +1101,41 @@ pool_prime_page(struct pool *pp, caddr_t if (++pp->pr_npages > pp->pr_hiwat) pp->pr_hiwat = pp->pr_npages; + + return (0); } /* - * Used by pool_get() when nitems drops below the low water mark. This - * is used to catch up nitmes with the low water mark. + * 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. * * Note 1, we never wait for memory here, we let the caller decide what to do. * - * Note 2, we must be called with the pool already locked, and we return + * Note 2, this doesn't work with static pools. + * + * Note 3, we must be called with the pool already locked, and we return * with it locked. */ static int pool_catchup(struct pool *pp) { - struct pool_item_header *ph; caddr_t cp; int error = 0; - while (POOL_NEEDS_CATCHUP(pp)) { + if (pp->pr_roflags & PR_STATIC) { + /* + * We dropped below the low water mark, and this is not a + * good thing. Log a warning. + * + * XXX: rate-limit this? + */ + printf("WARNING: static pool `%s' dropped below low water " + "mark\n", pp->pr_wchan); + return (0); + } + + while (pp->pr_nitems < pp->pr_minitems) { /* * Call the page back-end allocator for more memory. * @@ -1167,17 +1143,16 @@ pool_catchup(struct pool *pp) * the pool descriptor? */ simple_unlock(&pp->pr_slock); - cp = pool_allocator_alloc(pp, PR_NOWAIT); - if (__predict_true(cp != NULL)) - ph = pool_alloc_item_header(pp, cp, PR_NOWAIT); + cp = (*pp->pr_alloc)(pp->pr_pagesz, 0, pp->pr_mtype); simple_lock(&pp->pr_slock); - if (__predict_false(cp == NULL || ph == NULL)) { - if (cp != NULL) - pool_allocator_free(pp, cp); + if (__predict_false(cp == NULL)) { error = ENOMEM; break; } - pool_prime_page(pp, cp, ph); + if ((error = pool_prime_page(pp, cp, PR_NOWAIT)) != 0) { + (*pp->pr_free)(cp, pp->pr_pagesz, pp->pr_mtype); + break; + } pp->pr_npagealloc++; } @@ -1187,6 +1162,7 @@ pool_catchup(struct pool *pp) void pool_setlowat(struct pool *pp, int n) { + int error; simple_lock(&pp->pr_slock); @@ -1196,7 +1172,8 @@ pool_setlowat(struct pool *pp, int n) : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage; /* Make sure we're caught up with the newly-set low water mark. */ - if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) { + if ((pp->pr_nitems < pp->pr_minitems) && + (error = pool_catchup(pp)) != 0) { /* * XXX: Should we log a warning? Should we set up a timeout * to try again in a second or so? The latter could break @@ -1244,38 +1221,67 @@ pool_sethardlimit(struct pool *pp, int n } /* + * Default page allocator. + */ +static void * +pool_page_alloc(unsigned long sz, int flags, int mtype) +{ + boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE; + + return ((void *)uvm_km_alloc_poolpage(waitok)); +} + +static void +pool_page_free(void *v, unsigned long sz, int mtype) +{ + + uvm_km_free_poolpage((vaddr_t)v); +} + +/* + * Alternate pool page allocator for pools that know they will + * never be accessed in interrupt context. + */ +void * +pool_page_alloc_nointr(unsigned long sz, int flags, int mtype) +{ + boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE; + + return ((void *)uvm_km_alloc_poolpage1(kernel_map, uvm.kernel_object, + waitok)); +} + +void +pool_page_free_nointr(void *v, unsigned long sz, int mtype) +{ + + uvm_km_free_poolpage1(kernel_map, (vaddr_t)v); +} + + +/* * Release all complete pages that have not been used recently. */ -int -#ifdef POOL_DIAGNOSTIC +void _pool_reclaim(struct pool *pp, const char *file, long line) -#else -pool_reclaim(struct pool *pp) -#endif { struct pool_item_header *ph, *phnext; struct pool_cache *pc; struct timeval curtime; - struct pool_pagelist pq; int s; - if (pp->pr_drain_hook != NULL) { - /* - * The drain hook must be called with the pool unlocked. - */ - (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT); - } + if (pp->pr_roflags & PR_STATIC) + return; if (simple_lock_try(&pp->pr_slock) == 0) - return (0); + return; pr_enter(pp, file, line); - TAILQ_INIT(&pq); - /* * Reclaim items from the pool's caches. */ - TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist) + for (pc = TAILQ_FIRST(&pp->pr_cachelist); pc != NULL; + pc = TAILQ_NEXT(pc, pc_poollist)) pool_cache_reclaim(pc); s = splclock(); @@ -1303,30 +1309,15 @@ pool_reclaim(struct pool *pp) pp->pr_minitems) break; - pr_rmpage(pp, ph, &pq); + pr_rmpage(pp, ph); } } pr_leave(pp); simple_unlock(&pp->pr_slock); - if (TAILQ_EMPTY(&pq)) - return (0); - - while ((ph = TAILQ_FIRST(&pq)) != NULL) { - TAILQ_REMOVE(&pq, ph, ph_pagelist); - pool_allocator_free(pp, ph->ph_page); - if (pp->pr_roflags & PR_PHINPAGE) { - continue; - } - LIST_REMOVE(ph, ph_hashlist); - s = splhigh(); - pool_put(&phpool, ph); - splx(s); - } - - return (1); } + /* * Drain pools, one at a time. * @@ -1338,21 +1329,23 @@ pool_drain(void *arg) struct pool *pp; int s; - pp = NULL; s = splvm(); simple_lock(&pool_head_slock); - if (drainpp == NULL) { - drainpp = TAILQ_FIRST(&pool_head); - } - if (drainpp) { - pp = drainpp; - drainpp = TAILQ_NEXT(pp, pr_poollist); - } - simple_unlock(&pool_head_slock); + + if (drainpp == NULL && (drainpp = TAILQ_FIRST(&pool_head)) == NULL) + goto out; + + pp = drainpp; + drainpp = TAILQ_NEXT(pp, pr_poollist); + pool_reclaim(pp); + + out: + simple_unlock(&pool_head_slock); splx(s); } + /* * Diagnostic helpers. */ @@ -1428,7 +1421,8 @@ pool_print1(struct pool *pp, const char (*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n", pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset, pp->pr_roflags); - (*pr)("\talloc %p\n", pp->pr_alloc); + (*pr)("\tpagesz %u, mtype %d\n", pp->pr_pagesz, pp->pr_mtype); + (*pr)("\talloc %p, release %p\n", pp->pr_alloc, pp->pr_free); (*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n", pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages); (*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n", @@ -1450,7 +1444,8 @@ pool_print1(struct pool *pp, const char (u_long)ph->ph_time.tv_sec, (u_long)ph->ph_time.tv_usec); #ifdef DIAGNOSTIC - TAILQ_FOREACH(pi, &ph->ph_itemlist, pi_list) { + for (pi = TAILQ_FIRST(&ph->ph_itemlist); pi != NULL; + pi = TAILQ_NEXT(pi, pi_list)) { if (pi->pi_magic != PI_MAGIC) { (*pr)("\t\t\titem %p, magic 0x%x\n", pi, pi->pi_magic); @@ -1479,12 +1474,14 @@ pool_print1(struct pool *pp, const char if (print_cache == 0) goto skip_cache; - TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist) { + 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); - TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) { + 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]); @@ -1504,12 +1501,14 @@ pool_chk(struct pool *pp, const char *la simple_lock(&pp->pr_slock); - TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist) { + for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; + ph = TAILQ_NEXT(ph, ph_pagelist)) { + struct pool_item *pi; int n; caddr_t page; - page = (caddr_t)((u_long)ph & pp->pr_alloc->pa_pagemask); + page = (caddr_t)((u_long)ph & pp->pr_pagemask); if (page != ph->ph_page && (pp->pr_roflags & PR_PHINPAGE) != 0) { if (label != NULL) @@ -1538,8 +1537,7 @@ pool_chk(struct pool *pp, const char *la panic("pool"); } #endif - page = - (caddr_t)((u_long)pi & pp->pr_alloc->pa_pagemask); + page = (caddr_t)((u_long)pi & pp->pr_pagemask); if (page == ph->ph_page) continue; @@ -1655,15 +1653,11 @@ pool_cache_get(struct pool_cache *pc, in struct pool_cache_group *pcg; void *object; -#ifdef LOCKDEBUG - if (flags & PR_WAITOK) - simple_lock_only_held(NULL, "pool_cache_get(PR_WAITOK)"); -#endif - simple_lock(&pc->pc_slock); if ((pcg = pc->pc_allocfrom) == NULL) { - TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) { + 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; @@ -1710,12 +1704,12 @@ void pool_cache_put(struct pool_cache *pc, void *object) { struct pool_cache_group *pcg; - int s; simple_lock(&pc->pc_slock); if ((pcg = pc->pc_freeto) == NULL) { - TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) { + 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; @@ -1727,9 +1721,7 @@ pool_cache_put(struct pool_cache *pc, vo * allocate one. */ simple_unlock(&pc->pc_slock); - s = splvm(); pcg = pool_get(&pcgpool, PR_NOWAIT); - splx(s); if (pcg != NULL) { memset(pcg, 0, sizeof(*pcg)); simple_lock(&pc->pc_slock); @@ -1781,11 +1773,10 @@ pool_cache_destruct_object(struct pool_c */ static void pool_cache_do_invalidate(struct pool_cache *pc, int free_groups, - void (*putit)(struct pool *, void *)) + void (*putit)(struct pool *, void *, const char *, long)) { struct pool_cache_group *pcg, *npcg; void *object; - int s; for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL; pcg = npcg) { @@ -1797,16 +1788,14 @@ pool_cache_do_invalidate(struct pool_cac pc->pc_allocfrom = NULL; if (pc->pc_dtor != NULL) (*pc->pc_dtor)(pc->pc_arg, object); - (*putit)(pc->pc_pool, 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; - s = splvm(); pool_put(&pcgpool, pcg); - splx(s); } } } @@ -1822,7 +1811,7 @@ pool_cache_invalidate(struct pool_cache { simple_lock(&pc->pc_slock); - pool_cache_do_invalidate(pc, 0, pool_put); + pool_cache_do_invalidate(pc, 0, _pool_put); simple_unlock(&pc->pc_slock); } @@ -1839,218 +1828,3 @@ pool_cache_reclaim(struct pool_cache *pc pool_cache_do_invalidate(pc, 1, pool_do_put); simple_unlock(&pc->pc_slock); } - -/* - * Pool backend allocators. - * - * Each pool has a backend allocator that handles allocation, deallocation, - * and any additional draining that might be needed. - * - * We provide two standard allocators: - * - * pool_allocator_kmem - the default when no allocator is specified - * - * pool_allocator_nointr - used for pools that will not be accessed - * in interrupt context. - */ -void *pool_page_alloc(struct pool *, int); -void pool_page_free(struct pool *, void *); - -struct pool_allocator pool_allocator_kmem = { - pool_page_alloc, pool_page_free, 0, -}; - -void *pool_page_alloc_nointr(struct pool *, int); -void pool_page_free_nointr(struct pool *, void *); - -struct pool_allocator pool_allocator_nointr = { - pool_page_alloc_nointr, pool_page_free_nointr, 0, -}; - -#ifdef POOL_SUBPAGE -void *pool_subpage_alloc(struct pool *, int); -void pool_subpage_free(struct pool *, void *); - -struct pool_allocator pool_allocator_kmem_subpage = { - pool_subpage_alloc, pool_subpage_free, 0, -}; -#endif /* POOL_SUBPAGE */ - -/* - * We have at least three different resources for the same allocation and - * each resource can be depleted. First, we have the ready elements in the - * pool. Then we have the resource (typically a vm_map) for this allocator. - * Finally, we have physical memory. Waiting for any of these can be - * unnecessary when any other is freed, but the kernel doesn't support - * sleeping on multiple wait channels, so we have to employ another strategy. - * - * The caller sleeps on the pool (so that it can be awakened when an item - * is returned to the pool), but we set PA_WANT on the allocator. When a - * page is returned to the allocator and PA_WANT is set, pool_allocator_free - * will wake up all sleeping pools belonging to this allocator. - * - * XXX Thundering herd. - */ -void * -pool_allocator_alloc(struct pool *org, int flags) -{ - struct pool_allocator *pa = org->pr_alloc; - struct pool *pp, *start; - int s, freed; - void *res; - - do { - if ((res = (*pa->pa_alloc)(org, flags)) != NULL) - return (res); - if ((flags & PR_WAITOK) == 0) { - /* - * We only run the drain hookhere if PR_NOWAIT. - * In other cases, the hook will be run in - * pool_reclaim(). - */ - if (org->pr_drain_hook != NULL) { - (*org->pr_drain_hook)(org->pr_drain_hook_arg, - flags); - if ((res = (*pa->pa_alloc)(org, flags)) != NULL) - return (res); - } - break; - } - - /* - * Drain all pools, except "org", that use this - * allocator. We do this to reclaim VA space. - * pa_alloc is responsible for waiting for - * physical memory. - * - * XXX We risk looping forever if start if someone - * calls pool_destroy on "start". But there is no - * other way to have potentially sleeping pool_reclaim, - * non-sleeping locks on pool_allocator, and some - * stirring of drained pools in the allocator. - * - * XXX Maybe we should use pool_head_slock for locking - * the allocators? - */ - freed = 0; - - s = splvm(); - simple_lock(&pa->pa_slock); - pp = start = TAILQ_FIRST(&pa->pa_list); - do { - TAILQ_REMOVE(&pa->pa_list, pp, pr_alloc_list); - TAILQ_INSERT_TAIL(&pa->pa_list, pp, pr_alloc_list); - if (pp == org) - continue; - simple_unlock(&pa->pa_slock); - freed = pool_reclaim(pp); - simple_lock(&pa->pa_slock); - } while ((pp = TAILQ_FIRST(&pa->pa_list)) != start && - freed == 0); - - if (freed == 0) { - /* - * We set PA_WANT here, the caller will most likely - * sleep waiting for pages (if not, this won't hurt - * that much), and there is no way to set this in - * the caller without violating locking order. - */ - pa->pa_flags |= PA_WANT; - } - simple_unlock(&pa->pa_slock); - splx(s); - } while (freed); - return (NULL); -} - -void -pool_allocator_free(struct pool *pp, void *v) -{ - struct pool_allocator *pa = pp->pr_alloc; - int s; - - (*pa->pa_free)(pp, v); - - s = splvm(); - simple_lock(&pa->pa_slock); - if ((pa->pa_flags & PA_WANT) == 0) { - simple_unlock(&pa->pa_slock); - splx(s); - return; - } - - TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) { - simple_lock(&pp->pr_slock); - if ((pp->pr_flags & PR_WANTED) != 0) { - pp->pr_flags &= ~PR_WANTED; - wakeup(pp); - } - simple_unlock(&pp->pr_slock); - } - pa->pa_flags &= ~PA_WANT; - simple_unlock(&pa->pa_slock); - splx(s); -} - -void * -pool_page_alloc(struct pool *pp, int flags) -{ - boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE; - - return ((void *) uvm_km_alloc_poolpage(waitok)); -} - -void -pool_page_free(struct pool *pp, void *v) -{ - - uvm_km_free_poolpage((vaddr_t) v); -} - -#ifdef POOL_SUBPAGE -/* Sub-page allocator, for machines with large hardware pages. */ -void * -pool_subpage_alloc(struct pool *pp, int flags) -{ - - return (pool_get(&psppool, flags)); -} - -void -pool_subpage_free(struct pool *pp, void *v) -{ - - pool_put(&psppool, v); -} - -/* We don't provide a real nointr allocator. Maybe later. */ -void * -pool_page_alloc_nointr(struct pool *pp, int flags) -{ - - return (pool_subpage_alloc(pp, flags)); -} - -void -pool_page_free_nointr(struct pool *pp, void *v) -{ - - pool_subpage_free(pp, v); -} -#else -void * -pool_page_alloc_nointr(struct pool *pp, int flags) -{ - boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE; - - return ((void *) uvm_km_alloc_poolpage1(kernel_map, - uvm.kernel_object, waitok)); -} - -void -pool_page_free_nointr(struct pool *pp, void *v) -{ - - uvm_km_free_poolpage1(kernel_map, (vaddr_t) v); -} -#endif /* POOL_SUBPAGE */