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 1.21.2.2 retrieving revision 1.44 diff -u -p -r1.21.2.2 -r1.44 --- src/sys/kern/subr_pool.c 1999/04/07 00:34:55 1.21.2.2 +++ src/sys/kern/subr_pool.c 2000/12/07 19:30:31 1.44 @@ -1,7 +1,7 @@ -/* $NetBSD: subr_pool.c,v 1.21.2.2 1999/04/07 00:34:55 thorpej Exp $ */ +/* $NetBSD: subr_pool.c,v 1.44 2000/12/07 19:30:31 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 @@ -37,6 +37,10 @@ * POSSIBILITY OF SUCH DAMAGE. */ +#include "opt_pool.h" +#include "opt_poollog.h" +#include "opt_lockdebug.h" + #include #include #include @@ -47,9 +51,6 @@ #include #include -#include -#include - #include /* @@ -93,31 +94,64 @@ struct pool_item_header { struct pool_item { #ifdef DIAGNOSTIC int pi_magic; -#define PI_MAGIC 0xdeadbeef #endif +#define PI_MAGIC 0xdeadbeef /* Other entries use only this list entry */ TAILQ_ENTRY(pool_item) pi_list; }; - -#define PR_HASH_INDEX(pp,addr) \ +#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 struct pool_item_header - *pr_find_pagehead __P((struct pool *, caddr_t)); -static void pr_rmpage __P((struct pool *, struct pool_item_header *)); -static int pool_catchup __P((struct pool *)); -static void pool_prime_page __P((struct pool *, caddr_t)); -static void *pool_page_alloc __P((unsigned long, int, int)); -static void pool_page_free __P((void *, unsigned long, int)); +static void pool_cache_reclaim(struct pool_cache *); -#if defined(POOL_DIAGNOSTIC) || defined(DEBUG) -static void pool_print1 __P((struct pool *, const char *)); -#endif +static int pool_catchup(struct pool *); +static void pool_prime_page(struct pool *, caddr_t); +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 *, ...)); -#ifdef POOL_DIAGNOSTIC /* * Pool log entry. An array of these is allocated in pool_create(). */ @@ -125,8 +159,8 @@ struct pool_log { const char *pl_file; long pl_line; int pl_action; -#define PRLOG_GET 1 -#define PRLOG_PUT 2 +#define PRLOG_GET 1 +#define PRLOG_PUT 2 void *pl_addr; }; @@ -137,16 +171,9 @@ struct pool_log { int pool_logsize = POOL_LOGSIZE; -static void pr_log __P((struct pool *, void *, int, const char *, long)); -static void pr_printlog __P((struct pool *)); - -static __inline__ void -pr_log(pp, v, action, file, line) - struct pool *pp; - void *v; - int action; - const char *file; - long line; +#ifdef DIAGNOSTIC +static __inline void +pr_log(struct pool *pp, void *v, int action, const char *file, long line) { int n = pp->pr_curlogentry; struct pool_log *pl; @@ -169,8 +196,8 @@ pr_log(pp, v, action, file, line) } static void -pr_printlog(pp) - struct pool *pp; +pr_printlog(struct pool *pp, struct pool_item *pi, + void (*pr)(const char *, ...)) { int i = pp->pr_logsize; int n = pp->pr_curlogentry; @@ -178,38 +205,76 @@ pr_printlog(pp) if ((pp->pr_roflags & PR_LOGGING) == 0) return; - pool_print1(pp, "printlog"); - /* * Print all entries in this pool's log. */ while (i-- > 0) { struct pool_log *pl = &pp->pr_log[n]; if (pl->pl_action != 0) { - printf("log entry %d:\n", i); - printf("\taction = %s, addr = %p\n", - pl->pl_action == PRLOG_GET ? "get" : "put", - pl->pl_addr); - printf("\tfile: %s at line %lu\n", - pl->pl_file, pl->pl_line); + if (pi == NULL || pi == pl->pl_addr) { + (*pr)("\tlog entry %d:\n", i); + (*pr)("\t\taction = %s, addr = %p\n", + pl->pl_action == PRLOG_GET ? "get" : "put", + pl->pl_addr); + (*pr)("\t\tfile: %s at line %lu\n", + pl->pl_file, pl->pl_line); + } } if (++n >= pp->pr_logsize) n = 0; } } -#else -#define pr_log(pp, v, action, file, line) -#define pr_printlog(pp) -#endif +static __inline void +pr_enter(struct pool *pp, const char *file, long line) +{ + + if (__predict_false(pp->pr_entered_file != NULL)) { + printf("pool %s: reentrancy at file %s line %ld\n", + pp->pr_wchan, file, line); + printf(" previous entry at file %s line %ld\n", + pp->pr_entered_file, pp->pr_entered_line); + panic("pr_enter"); + } + + pp->pr_entered_file = file; + pp->pr_entered_line = line; +} + +static __inline void +pr_leave(struct pool *pp) +{ + + if (__predict_false(pp->pr_entered_file == NULL)) { + printf("pool %s not entered?\n", pp->pr_wchan); + panic("pr_leave"); + } + + pp->pr_entered_file = NULL; + pp->pr_entered_line = 0; +} + +static __inline void +pr_enter_check(struct pool *pp, void (*pr)(const char *, ...)) +{ + + if (pp->pr_entered_file != NULL) + (*pr)("\n\tcurrently entered from file %s line %ld\n", + pp->pr_entered_file, pp->pr_entered_line); +} +#else +#define pr_log(pp, v, action, file, line) +#define pr_printlog(pp, pi, pr) +#define pr_enter(pp, file, line) +#define pr_leave(pp) +#define pr_enter_check(pp, pr) +#endif /* DIAGNOSTIC */ /* * Return the pool page header based on page address. */ -static __inline__ struct pool_item_header * -pr_find_pagehead(pp, page) - struct pool *pp; - caddr_t page; +static __inline struct pool_item_header * +pr_find_pagehead(struct pool *pp, caddr_t page) { struct pool_item_header *ph; @@ -228,10 +293,8 @@ pr_find_pagehead(pp, page) /* * Remove a page from the pool. */ -static __inline__ void -pr_rmpage(pp, ph) - struct pool *pp; - struct pool_item_header *ph; +static __inline void +pr_rmpage(struct pool *pp, struct pool_item_header *ph) { /* @@ -258,8 +321,11 @@ pr_rmpage(pp, ph) 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) { @@ -281,16 +347,11 @@ pr_rmpage(pp, ph) * Allocate and initialize a pool. */ struct pool * -pool_create(size, align, ioff, nitems, wchan, pagesz, alloc, release, mtype) - size_t size; - u_int align; - u_int ioff; - int nitems; - const char *wchan; - size_t pagesz; - void *(*alloc) __P((unsigned long, int, int)); - void (*release) __P((void *, unsigned long, int)); - int mtype; +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; @@ -300,11 +361,6 @@ pool_create(size, align, ioff, nitems, w return (NULL); flags = PR_FREEHEADER; -#ifdef POOL_DIAGNOSTIC - if (pool_logsize != 0) - flags |= PR_LOGGING; -#endif - pool_init(pp, size, align, ioff, flags, wchan, pagesz, alloc, release, mtype); @@ -325,24 +381,26 @@ pool_create(size, align, ioff, nitems, w * static pools that must be initialized before malloc() is available. */ void -pool_init(pp, size, align, ioff, flags, wchan, pagesz, alloc, release, mtype) - struct pool *pp; - size_t size; - u_int align; - u_int ioff; - int flags; - const char *wchan; - size_t pagesz; - void *(*alloc) __P((unsigned long, int, int)); - void (*release) __P((void *, unsigned long, int)); - int mtype; +pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags, + const char *wchan, size_t pagesz, + void *(*alloc)(unsigned long, int, int), + void (*release)(void *, unsigned long, int), + int mtype) { int off, slack, i; +#ifdef POOL_DIAGNOSTIC + /* + * Always log if POOL_DIAGNOSTIC is defined. + */ + if (pool_logsize != 0) + flags |= PR_LOGGING; +#endif + /* * Check arguments and construct default values. */ - if (!powerof2(pagesz) || pagesz > PAGE_SIZE) + if (!powerof2(pagesz)) panic("pool_init: page size invalid (%lx)\n", (u_long)pagesz); if (alloc == NULL && release == NULL) { @@ -363,10 +421,16 @@ pool_init(pp, size, align, ioff, flags, if (size < sizeof(struct pool_item)) size = sizeof(struct pool_item); + size = ALIGN(size); + if (size > pagesz) + panic("pool_init: pool item size (%lu) too large", + (u_long)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; @@ -374,7 +438,7 @@ pool_init(pp, size, align, ioff, flags, pp->pr_maxpages = UINT_MAX; pp->pr_roflags = flags; pp->pr_flags = 0; - pp->pr_size = ALIGN(size); + pp->pr_size = size; pp->pr_align = align; pp->pr_wchan = wchan; pp->pr_mtype = mtype; @@ -387,9 +451,10 @@ pool_init(pp, size, align, ioff, flags, pp->pr_nout = 0; pp->pr_hardlimit = UINT_MAX; pp->pr_hardlimit_warning = NULL; - pp->pr_hardlimit_ratecap = 0; - memset(&pp->pr_hardlimit_warning_last, 0, - sizeof(pp->pr_hardlimit_warning_last)); + pp->pr_hardlimit_ratecap.tv_sec = 0; + pp->pr_hardlimit_ratecap.tv_usec = 0; + pp->pr_hardlimit_warning_last.tv_sec = 0; + pp->pr_hardlimit_warning_last.tv_usec = 0; /* * Decide whether to put the page header off page to avoid @@ -421,6 +486,7 @@ pool_init(pp, size, align, ioff, flags, */ 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 @@ -438,26 +504,30 @@ pool_init(pp, size, align, ioff, flags, pp->pr_hiwat = 0; pp->pr_nidle = 0; -#ifdef POOL_DIAGNOSTIC - if ((flags & PR_LOGGING) != 0) { - pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log), - M_TEMP, M_NOWAIT); - if (pp->pr_log == NULL) + 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; 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. */ @@ -470,14 +540,18 @@ pool_init(pp, size, align, ioff, flags, * De-commision a pool resource. */ void -pool_destroy(pp) - struct pool *pp; +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) { - pr_printlog(pp); + pr_printlog(pp, NULL, printf); panic("pool_destroy: pool busy: still out: %u\n", pp->pr_nout); } @@ -495,10 +569,8 @@ pool_destroy(pp) drainpp = NULL; simple_unlock(&pool_head_slock); -#ifdef POOL_DIAGNOSTIC if ((pp->pr_roflags & PR_LOGGING) != 0) free(pp->pr_log, M_TEMP); -#endif if (pp->pr_roflags & PR_FREEHEADER) free(pp, M_POOL); @@ -508,35 +580,27 @@ pool_destroy(pp) /* * Grab an item from the pool; must be called at appropriate spl level */ -#ifdef POOL_DIAGNOSTIC void * -_pool_get(pp, flags, file, line) - struct pool *pp; - int flags; - const char *file; - long line; -#else -void * -pool_get(pp, flags) - struct pool *pp; - int flags; -#endif +_pool_get(struct pool *pp, int flags, const char *file, long line) { void *v; struct pool_item *pi; struct pool_item_header *ph; #ifdef DIAGNOSTIC - if ((pp->pr_roflags & PR_STATIC) && (flags & PR_MALLOCOK)) { - pr_printlog(pp); + if (__predict_false((pp->pr_roflags & PR_STATIC) && + (flags & PR_MALLOCOK))) { + pr_printlog(pp, NULL, printf); panic("pool_get: static"); } #endif - if (curproc == NULL && (flags & PR_WAITOK) != 0) + if (__predict_false(curproc == NULL && doing_shutdown == 0 && + (flags & PR_WAITOK) != 0)) panic("pool_get: must have NOWAIT"); simple_lock(&pp->pr_slock); + pr_enter(pp, file, line); startover: /* @@ -545,44 +609,39 @@ pool_get(pp, flags) * the pool. */ #ifdef DIAGNOSTIC - if (pp->pr_nout > pp->pr_hardlimit) { + if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) { + pr_leave(pp); simple_unlock(&pp->pr_slock); panic("pool_get: %s: crossed hard limit", pp->pr_wchan); } #endif - if (pp->pr_nout == pp->pr_hardlimit) { - if (flags & PR_WAITOK) { + if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) { + if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) { /* * XXX: A warning isn't logged in this case. Should * it be? */ pp->pr_flags |= PR_WANTED; - simple_unlock(&pp->pr_slock); - tsleep((caddr_t)pp, PSWP, pp->pr_wchan, 0); - simple_lock(&pp->pr_slock); + pr_leave(pp); + ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock); + pr_enter(pp, file, line); goto startover; } - if (pp->pr_hardlimit_warning != NULL) { - /* - * Log a message that the hard limit has been hit. - */ - struct timeval curtime, logdiff; - int s = splclock(); - curtime = mono_time; - splx(s); - timersub(&curtime, &pp->pr_hardlimit_warning_last, - &logdiff); - if (logdiff.tv_sec >= pp->pr_hardlimit_ratecap) { - pp->pr_hardlimit_warning_last = curtime; - log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning); - } - } + + /* + * Log a message that the hard limit has been hit. + */ + if (pp->pr_hardlimit_warning != NULL && + ratecheck(&pp->pr_hardlimit_warning_last, + &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); simple_unlock(&pp->pr_slock); return (NULL); } @@ -610,9 +669,11 @@ pool_get(pp, flags) * Release the pool lock, as the back-end page allocator * may block. */ + pr_leave(pp); simple_unlock(&pp->pr_slock); v = (*pp->pr_alloc)(pp->pr_pagesz, flags, pp->pr_mtype); simple_lock(&pp->pr_slock); + pr_enter(pp, file, line); if (v == NULL) { /* @@ -629,6 +690,7 @@ pool_get(pp, flags) if ((flags & PR_WAITOK) == 0) { pp->pr_nfail++; + pr_leave(pp); simple_unlock(&pp->pr_slock); return (NULL); } @@ -645,9 +707,9 @@ pool_get(pp, flags) * try again? */ pp->pr_flags |= PR_WANTED; - simple_unlock(&pp->pr_slock); - tsleep((caddr_t)pp, PSWP, pp->pr_wchan, 0); - simple_lock(&pp->pr_slock); + pr_leave(pp); + ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock); + pr_enter(pp, file, line); goto startover; } @@ -659,12 +721,14 @@ pool_get(pp, flags) goto startover; } - if ((v = pi = TAILQ_FIRST(&ph->ph_itemlist)) == NULL) { + if (__predict_false((v = pi = TAILQ_FIRST(&ph->ph_itemlist)) == NULL)) { + pr_leave(pp); simple_unlock(&pp->pr_slock); panic("pool_get: %s: page empty", pp->pr_wchan); } #ifdef DIAGNOSTIC - if (pp->pr_nitems == 0) { + if (__predict_false(pp->pr_nitems == 0)) { + pr_leave(pp); simple_unlock(&pp->pr_slock); printf("pool_get: %s: items on itemlist, nitems %u\n", pp->pr_wchan, pp->pr_nitems); @@ -674,8 +738,8 @@ pool_get(pp, flags) pr_log(pp, v, PRLOG_GET, file, line); #ifdef DIAGNOSTIC - if (pi->pi_magic != PI_MAGIC) { - pr_printlog(pp); + if (__predict_false(pi->pi_magic != PI_MAGIC)) { + pr_printlog(pp, pi, printf); panic("pool_get(%s): free list modified: magic=%x; page %p;" " item addr %p\n", pp->pr_wchan, pi->pi_magic, ph->ph_page, pi); @@ -690,7 +754,7 @@ pool_get(pp, flags) pp->pr_nout++; if (ph->ph_nmissing == 0) { #ifdef DIAGNOSTIC - if (pp->pr_nidle == 0) + if (__predict_false(pp->pr_nidle == 0)) panic("pool_get: nidle inconsistent"); #endif pp->pr_nidle--; @@ -698,7 +762,8 @@ pool_get(pp, flags) ph->ph_nmissing++; if (TAILQ_FIRST(&ph->ph_itemlist) == NULL) { #ifdef DIAGNOSTIC - if (ph->ph_nmissing != pp->pr_itemsperpage) { + if (__predict_false(ph->ph_nmissing != pp->pr_itemsperpage)) { + pr_leave(pp); simple_unlock(&pp->pr_slock); panic("pool_get: %s: nmissing inconsistent", pp->pr_wchan); @@ -739,26 +804,16 @@ pool_get(pp, flags) */ } + pr_leave(pp); simple_unlock(&pp->pr_slock); return (v); } /* - * Return resource to the pool; must be called at appropriate spl level + * Internal version of pool_put(). Pool is already locked/entered. */ -#ifdef POOL_DIAGNOSTIC -void -_pool_put(pp, v, file, line) - struct pool *pp; - void *v; - const char *file; - long line; -#else -void -pool_put(pp, v) - struct pool *pp; - void *v; -#endif +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; @@ -767,21 +822,44 @@ pool_put(pp, v) page = (caddr_t)((u_long)v & pp->pr_pagemask); - simple_lock(&pp->pr_slock); +#ifdef DIAGNOSTIC + if (__predict_false(pp->pr_nout == 0)) { + printf("pool %s: putting with none out\n", + pp->pr_wchan); + panic("pool_put"); + } +#endif pr_log(pp, v, PRLOG_PUT, file, line); - if ((ph = pr_find_pagehead(pp, page)) == NULL) { - pr_printlog(pp); + 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); } +#ifdef LOCKDEBUG + /* + * Check if we're freeing a locked simple lock. + */ + simple_lock_freecheck((caddr_t)pi, ((caddr_t)pi) + pp->pr_size); +#endif + /* * Return to item list. */ #ifdef DIAGNOSTIC pi->pi_magic = PI_MAGIC; #endif +#ifdef DEBUG + { + int i, *ip = v; + + for (i = 0; i < pp->pr_size / sizeof(int); i++) { + *ip++ = PI_MAGIC; + } + } +#endif + TAILQ_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list); ph->ph_nmissing--; pp->pr_nput++; @@ -796,7 +874,6 @@ pool_put(pp, v) pp->pr_flags &= ~PR_WANTED; if (ph->ph_nmissing == 0) pp->pr_nidle++; - simple_unlock(&pp->pr_slock); wakeup((caddr_t)pp); return; } @@ -860,25 +937,35 @@ pool_put(pp, v) TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist); pp->pr_curpage = ph; } +} - simple_unlock(&pp->pr_slock); +/* + * Return resource to the pool; must be called at appropriate spl level + */ +void +_pool_put(struct pool *pp, void *v, const char *file, long line) +{ + simple_lock(&pp->pr_slock); + pr_enter(pp, file, line); + + pool_do_put(pp, v, file, line); + + pr_leave(pp); + simple_unlock(&pp->pr_slock); } /* * Add N items to the pool. */ int -pool_prime(pp, n, storage) - struct pool *pp; - int n; - caddr_t storage; +pool_prime(struct pool *pp, int n, caddr_t storage) { caddr_t cp; int newnitems, newpages; #ifdef DIAGNOSTIC - if (storage && !(pp->pr_roflags & PR_STATIC)) + if (__predict_false(storage && !(pp->pr_roflags & PR_STATIC))) panic("pool_prime: static"); /* !storage && static caught below */ #endif @@ -905,6 +992,7 @@ pool_prime(pp, n, storage) return (ENOMEM); } + pp->pr_npagealloc++; pool_prime_page(pp, cp); pp->pr_minpages++; } @@ -924,21 +1012,24 @@ pool_prime(pp, n, storage) * Note, we must be called with the pool descriptor LOCKED. */ static void -pool_prime_page(pp, storage) - struct pool *pp; - caddr_t storage; +pool_prime_page(struct pool *pp, caddr_t storage) { 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; + + if (((u_long)cp & (pp->pr_pagesz - 1)) != 0) + panic("pool_prime_page: %s: unaligned page", pp->pr_wchan); if ((pp->pr_roflags & PR_PHINPAGE) != 0) { ph = (struct pool_item_header *)(cp + pp->pr_phoffset); } else { + s = splhigh(); ph = pool_get(&phpool, PR_URGENT); + splx(s); LIST_INSERT_HEAD(&pp->pr_hashtab[PR_HASH_INDEX(pp, cp)], ph, ph_hashlist); } @@ -1007,8 +1098,7 @@ pool_prime_page(pp, storage) * with it locked. */ static int -pool_catchup(pp) - struct pool *pp; +pool_catchup(struct pool *pp) { caddr_t cp; int error = 0; @@ -1035,10 +1125,11 @@ pool_catchup(pp) simple_unlock(&pp->pr_slock); cp = (*pp->pr_alloc)(pp->pr_pagesz, 0, pp->pr_mtype); simple_lock(&pp->pr_slock); - if (cp == NULL) { + if (__predict_false(cp == NULL)) { error = ENOMEM; break; } + pp->pr_npagealloc++; pool_prime_page(pp, cp); } @@ -1046,9 +1137,7 @@ pool_catchup(pp) } void -pool_setlowat(pp, n) - pool_handle_t pp; - int n; +pool_setlowat(struct pool *pp, int n) { int error; @@ -1060,7 +1149,8 @@ pool_setlowat(pp, n) : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage; /* Make sure we're caught up with the newly-set low water mark. */ - if ((error = 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 @@ -1072,9 +1162,7 @@ pool_setlowat(pp, n) } void -pool_sethiwat(pp, n) - pool_handle_t pp; - int n; +pool_sethiwat(struct pool *pp, int n) { simple_lock(&pp->pr_slock); @@ -1087,20 +1175,16 @@ pool_sethiwat(pp, n) } void -pool_sethardlimit(pp, n, warnmess, ratecap) - pool_handle_t pp; - int n; - const char *warnmess; - int ratecap; +pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap) { simple_lock(&pp->pr_slock); pp->pr_hardlimit = n; pp->pr_hardlimit_warning = warnmess; - pp->pr_hardlimit_ratecap = ratecap; - memset(&pp->pr_hardlimit_warning_last, 0, - sizeof(pp->pr_hardlimit_warning_last)); + pp->pr_hardlimit_ratecap.tv_sec = ratecap; + pp->pr_hardlimit_warning_last.tv_sec = 0; + pp->pr_hardlimit_warning_last.tv_usec = 0; /* * In-line version of pool_sethiwat(), because we don't want to @@ -1117,10 +1201,7 @@ pool_sethardlimit(pp, n, warnmess, ratec * Default page allocator. */ static void * -pool_page_alloc(sz, flags, mtype) - unsigned long sz; - int flags; - int mtype; +pool_page_alloc(unsigned long sz, int flags, int mtype) { boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE; @@ -1128,10 +1209,7 @@ pool_page_alloc(sz, flags, mtype) } static void -pool_page_free(v, sz, mtype) - void *v; - unsigned long sz; - int mtype; +pool_page_free(void *v, unsigned long sz, int mtype) { uvm_km_free_poolpage((vaddr_t)v); @@ -1142,10 +1220,7 @@ pool_page_free(v, sz, mtype) * never be accessed in interrupt context. */ void * -pool_page_alloc_nointr(sz, flags, mtype) - unsigned long sz; - int flags; - int mtype; +pool_page_alloc_nointr(unsigned long sz, int flags, int mtype) { boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE; @@ -1154,10 +1229,7 @@ pool_page_alloc_nointr(sz, flags, mtype) } void -pool_page_free_nointr(v, sz, mtype) - void *v; - unsigned long sz; - int mtype; +pool_page_free_nointr(void *v, unsigned long sz, int mtype) { uvm_km_free_poolpage1(kernel_map, (vaddr_t)v); @@ -1168,10 +1240,10 @@ pool_page_free_nointr(v, sz, mtype) * Release all complete pages that have not been used recently. */ void -pool_reclaim(pp) - pool_handle_t pp; +_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; @@ -1180,6 +1252,14 @@ pool_reclaim(pp) if (simple_lock_try(&pp->pr_slock) == 0) 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; @@ -1210,6 +1290,7 @@ pool_reclaim(pp) } } + pr_leave(pp); simple_unlock(&pp->pr_slock); } @@ -1220,8 +1301,7 @@ pool_reclaim(pp) * Note, we must never be called from an interrupt context. */ void -pool_drain(arg) - void *arg; +pool_drain(void *arg) { struct pool *pp; int s; @@ -1243,52 +1323,153 @@ pool_drain(arg) } -#if defined(POOL_DIAGNOSTIC) || defined(DEBUG) /* * Diagnostic helpers. */ void -pool_print(pp, label) - struct pool *pp; - const char *label; +pool_print(struct pool *pp, const char *modif) { int s; s = splimp(); - simple_lock(&pp->pr_slock); - pool_print1(pp, label); + if (simple_lock_try(&pp->pr_slock) == 0) { + printf("pool %s is locked; try again later\n", + pp->pr_wchan); + splx(s); + return; + } + pool_print1(pp, modif, printf); simple_unlock(&pp->pr_slock); splx(s); } +void +pool_printit(struct pool *pp, const char *modif, void (*pr)(const char *, ...)) +{ + int didlock = 0; + + if (pp == NULL) { + (*pr)("Must specify a pool to print.\n"); + return; + } + + /* + * Called from DDB; interrupts should be blocked, and all + * other processors should be paused. We can skip locking + * the pool in this case. + * + * We do a simple_lock_try() just to print the lock + * status, however. + */ + + if (simple_lock_try(&pp->pr_slock) == 0) + (*pr)("WARNING: pool %s is locked\n", pp->pr_wchan); + else + didlock = 1; + + pool_print1(pp, modif, pr); + + if (didlock) + simple_unlock(&pp->pr_slock); +} + static void -pool_print1(pp, label) - struct pool *pp; - const char *label; +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 i, print_log = 0, print_pagelist = 0, print_cache = 0; + char c; - if (label != NULL) - printf("%s: ", label); + while ((c = *modif++) != '\0') { + if (c == 'l') + print_log = 1; + if (c == 'p') + print_pagelist = 1; + if (c == 'c') + print_cache = 1; + modif++; + } + + (*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)("\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", + pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit); + + (*pr)("\n\tnget %lu, nfail %lu, nput %lu\n", + pp->pr_nget, pp->pr_nfail, pp->pr_nput); + (*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n", + pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle); + + if (print_pagelist == 0) + goto skip_pagelist; + + if ((ph = TAILQ_FIRST(&pp->pr_pagelist)) != NULL) + (*pr)("\n\tpage list:\n"); + for (; ph != NULL; ph = TAILQ_NEXT(ph, ph_pagelist)) { + (*pr)("\t\tpage %p, nmissing %d, time %lu,%lu\n", + ph->ph_page, ph->ph_nmissing, + (u_long)ph->ph_time.tv_sec, + (u_long)ph->ph_time.tv_usec); +#ifdef DIAGNOSTIC + 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); + } + } +#endif + } + if (pp->pr_curpage == NULL) + (*pr)("\tno current page\n"); + else + (*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page); + + skip_pagelist: - printf("pool %s: nalloc %lu nfree %lu npagealloc %lu npagefree %lu\n" - " npages %u minitems %u itemsperpage %u itemoffset %u\n" - " nidle %lu\n", - pp->pr_wchan, - pp->pr_nget, - pp->pr_nput, - pp->pr_npagealloc, - pp->pr_npagefree, - pp->pr_npages, - pp->pr_minitems, - pp->pr_itemsperpage, - pp->pr_itemoffset, - pp->pr_nidle); + if (print_log == 0) + goto skip_log; + + (*pr)("\n"); + if ((pp->pr_roflags & PR_LOGGING) == 0) + (*pr)("\tno log\n"); + else + pr_printlog(pp, NULL, pr); + + 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); + 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); } int -pool_chk(pp, label) - struct pool *pp; - char *label; +pool_chk(struct pool *pp, const char *label) { struct pool_item_header *ph; int r = 0; @@ -1349,4 +1530,250 @@ out: simple_unlock(&pp->pr_slock); return (r); } -#endif /* POOL_DIAGNOSTIC || DEBUG */ + +/* + * 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; + + 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); + 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. + */ + 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: + 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. We don't unlock the cache here, since + * we never block. + */ + pcg = pool_get(&pcgpool, PR_NOWAIT); + if (pcg != NULL) { + memset(pcg, 0, sizeof(*pcg)); + TAILQ_INSERT_TAIL(&pc->pc_grouplist, pcg, pcg_list); + pc->pc_freeto = pcg; + goto have_group; + } + + simple_unlock(&pc->pc_slock); + + /* + * Unable to allocate a cache group; destruct the object + * and free it back to the pool. + */ + if (pc->pc_dtor != NULL) + (*pc->pc_dtor)(pc->pc_arg, object); + pool_put(pc->pc_pool, object); + return; + } + + have_group: + pcg_put(pcg, object); + + if (pcg->pcg_avail == PCG_NOBJECTS) + pc->pc_freeto = NULL; + + simple_unlock(&pc->pc_slock); +} + +/* + * 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) { + object = pcg_get(pcg); + if (pc->pc_dtor != NULL) + (*pc->pc_dtor)(pc->pc_arg, object); + (*putit)(pc->pc_pool, object, __FILE__, __LINE__); + } + if (free_groups) { + TAILQ_REMOVE(&pc->pc_grouplist, pcg, pcg_list); + 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) +{ + + /* + * We're locking in the opposite order (pool already + * locked in pool_reclaim()), so use a try-lock instead. + */ + + if (simple_lock_try(&pc->pc_slock) == 0) + return; + pool_cache_do_invalidate(pc, 1, pool_do_put); + simple_unlock(&pc->pc_slock); +}