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 1.55 retrieving revision 1.65 diff -u -p -r1.55 -r1.65 --- src/sys/kern/subr_pool.c 2001/05/10 04:51:41 1.55 +++ src/sys/kern/subr_pool.c 2001/11/20 06:57:04 1.65 @@ -1,4 +1,4 @@ -/* $NetBSD: subr_pool.c,v 1.55 2001/05/10 04:51:41 thorpej Exp $ */ +/* $NetBSD: subr_pool.c,v 1.65 2001/11/20 06:57:04 enami Exp $ */ /*- * Copyright (c) 1997, 1999, 2000 The NetBSD Foundation, Inc. @@ -37,6 +37,9 @@ * POSSIBILITY OF SUCH DAMAGE. */ +#include +__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.65 2001/11/20 06:57:04 enami Exp $"); + #include "opt_pool.h" #include "opt_poollog.h" #include "opt_lockdebug.h" @@ -70,6 +73,11 @@ 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; @@ -90,6 +98,7 @@ 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 @@ -152,6 +161,10 @@ static void pool_prime_page(struct pool struct pool_item_header *); static void *pool_page_alloc(unsigned long, int, int); static void pool_page_free(void *, unsigned long, int); +#ifdef POOL_SUBPAGE +static void *pool_subpage_alloc(unsigned long, int, int); +static void pool_subpage_free(void *, unsigned long, int); +#endif static void pool_print1(struct pool *, const char *, void (*)(const char *, ...)); @@ -175,7 +188,7 @@ struct pool_log { int pool_logsize = POOL_LOGSIZE; -#ifdef DIAGNOSTIC +#ifdef POOL_DIAGNOSTIC static __inline void pr_log(struct pool *pp, void *v, int action, const char *file, long line) { @@ -272,7 +285,7 @@ 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 /* DIAGNOSTIC */ +#endif /* POOL_DIAGNOSTIC */ /* * Return the pool page header based on page address. @@ -298,8 +311,10 @@ 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) +pr_rmpage(struct pool *pp, struct pool_item_header *ph, + struct pool_pagelist *pq) { + int s; /* * If the page was idle, decrement the idle page count. @@ -317,29 +332,30 @@ pr_rmpage(struct pool *pp, struct pool_i pp->pr_nitems -= pp->pr_itemsperpage; /* - * Unlink a page from the pool and release it. + * Unlink a page from the pool and release it (or queue it for release). */ TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist); - (*pp->pr_free)(ph->ph_page, pp->pr_pagesz, pp->pr_mtype); + if (pq) { + TAILQ_INSERT_HEAD(pq, ph, ph_pagelist); + } else { + (*pp->pr_free)(ph->ph_page, pp->pr_pagesz, pp->pr_mtype); + if ((pp->pr_roflags & PR_PHINPAGE) == 0) { + LIST_REMOVE(ph, ph_hashlist); + s = splhigh(); + pool_put(&phpool, ph); + splx(s); + } + } 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. */ - for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; - ph = TAILQ_NEXT(ph, ph_pagelist)) + TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist) if (TAILQ_FIRST(&ph->ph_itemlist) != NULL) break; @@ -377,13 +393,24 @@ pool_init(struct pool *pp, size_t size, panic("pool_init: page size invalid (%lx)\n", (u_long)pagesz); if (alloc == NULL && release == NULL) { +#ifdef POOL_SUBPAGE + alloc = pool_subpage_alloc; + release = pool_subpage_free; + pagesz = POOL_SUBPAGE; +#else alloc = pool_page_alloc; release = pool_page_free; pagesz = PAGE_SIZE; /* Rounds to PAGE_SIZE anyhow. */ +#endif } else if ((alloc != NULL && release != NULL) == 0) { /* If you specifiy one, must specify both. */ panic("pool_init: must specify alloc and release together"); } +#ifdef POOL_SUBPAGE + else if (alloc == pool_page_alloc_nointr && + release == pool_page_free_nointr) + pagesz = POOL_SUBPAGE; +#endif if (pagesz == 0) pagesz = PAGE_SIZE; @@ -477,6 +504,7 @@ 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), @@ -485,6 +513,7 @@ 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; @@ -497,8 +526,16 @@ 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", PAGE_SIZE, pool_page_alloc, pool_page_free, 0); + pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0, + PR_RECURSIVE, "psppool", PAGE_SIZE, + pool_page_alloc, pool_page_free, 0); +#else pool_init(&phpool, sizeof(struct pool_item_header), 0, 0, 0, "phpool", 0, 0, 0, 0); +#endif pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0, 0, "pcgpool", 0, 0, 0, 0); } @@ -532,18 +569,21 @@ pool_destroy(struct pool *pp) /* Remove all pages */ if ((pp->pr_roflags & PR_STATIC) == 0) - while ((ph = pp->pr_pagelist.tqh_first) != NULL) - pr_rmpage(pp, ph); + while ((ph = TAILQ_FIRST(&pp->pr_pagelist)) != NULL) + pr_rmpage(pp, ph, NULL); /* Remove from global pool list */ simple_lock(&pool_head_slock); TAILQ_REMOVE(&pool_head, pp, pr_poollist); - /* XXX Only clear this if we were drainpp? */ - drainpp = NULL; + if (drainpp == 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); @@ -572,7 +612,11 @@ pool_alloc_item_header(struct pool *pp, * 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 { struct pool_item *pi; struct pool_item_header *ph; @@ -584,12 +628,17 @@ _pool_get(struct pool *pp, int flags, co 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); @@ -729,7 +778,10 @@ _pool_get(struct pool *pp, int flags, co 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)) { @@ -776,8 +828,7 @@ _pool_get(struct pool *pp, int flags, co */ TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist); TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist); - for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; - ph = TAILQ_NEXT(ph, ph_pagelist)) + TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist) if (TAILQ_FIRST(&ph->ph_itemlist) != NULL) break; @@ -807,13 +858,15 @@ _pool_get(struct pool *pp, int flags, co * Internal version of pool_put(). Pool is already locked/entered. */ static void -pool_do_put(struct pool *pp, void *v, const char *file, long line) +pool_do_put(struct pool *pp, void *v) { 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_pagemask); #ifdef DIAGNOSTIC @@ -824,8 +877,6 @@ pool_do_put(struct pool *pp, void *v, co } #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); @@ -889,7 +940,7 @@ pool_do_put(struct pool *pp, void *v, co if (ph->ph_nmissing == 0) { pp->pr_nidle++; if (pp->pr_npages > pp->pr_maxpages) { - pr_rmpage(pp, ph); + pr_rmpage(pp, ph, NULL); } else { TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist); TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist); @@ -912,8 +963,7 @@ pool_do_put(struct pool *pp, void *v, co * page with the fewest available items, to minimize * fragmentation? */ - for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; - ph = TAILQ_NEXT(ph, ph_pagelist)) + TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist) if (TAILQ_FIRST(&ph->ph_itemlist) != NULL) break; @@ -936,6 +986,7 @@ pool_do_put(struct pool *pp, void *v, co /* * 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) { @@ -943,11 +994,30 @@ _pool_put(struct pool *pp, void *v, cons simple_lock(&pp->pr_slock); pr_enter(pp, file, line); - pool_do_put(pp, v, file, line); + pr_log(pp, v, PRLOG_PUT, file, line); + + pool_do_put(pp, v); 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. @@ -1194,6 +1264,41 @@ pool_page_free(void *v, unsigned long sz uvm_km_free_poolpage((vaddr_t)v); } +#ifdef POOL_SUBPAGE +/* + * Sub-page allocator, for machines with large hardware pages. + */ +static void * +pool_subpage_alloc(unsigned long sz, int flags, int mtype) +{ + + return pool_get(&psppool, flags); +} + +static void +pool_subpage_free(void *v, unsigned long sz, int mtype) +{ + + pool_put(&psppool, v); +} +#endif + +#ifdef POOL_SUBPAGE +/* We don't provide a real nointr allocator. Maybe later. */ +void * +pool_page_alloc_nointr(unsigned long sz, int flags, int mtype) +{ + + return pool_subpage_alloc(sz, flags, mtype); +} + +void +pool_page_free_nointr(void *v, unsigned long sz, int mtype) +{ + + pool_subpage_free(v, sz, mtype); +} +#else /* * Alternate pool page allocator for pools that know they will * never be accessed in interrupt context. @@ -1213,17 +1318,23 @@ pool_page_free_nointr(void *v, unsigned uvm_km_free_poolpage1(kernel_map, (vaddr_t)v); } +#endif /* * Release all complete pages that have not been used recently. */ void +#ifdef POOL_DIAGNOSTIC _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_roflags & PR_STATIC) @@ -1232,12 +1343,12 @@ _pool_reclaim(struct pool *pp, const cha if (simple_lock_try(&pp->pr_slock) == 0) return; pr_enter(pp, file, line); + TAILQ_INIT(&pq); /* * Reclaim items from the pool's caches. */ - for (pc = TAILQ_FIRST(&pp->pr_cachelist); pc != NULL; - pc = TAILQ_NEXT(pc, pc_poollist)) + TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist) pool_cache_reclaim(pc); s = splclock(); @@ -1265,12 +1376,26 @@ _pool_reclaim(struct pool *pp, const cha pp->pr_minitems) break; - pr_rmpage(pp, ph); + pr_rmpage(pp, ph, &pq); } } pr_leave(pp); simple_unlock(&pp->pr_slock); + if (TAILQ_EMPTY(&pq)) { + return; + } + while ((ph = TAILQ_FIRST(&pq)) != NULL) { + TAILQ_REMOVE(&pq, ph, ph_pagelist); + (*pp->pr_free)(ph->ph_page, pp->pr_pagesz, pp->pr_mtype); + if (pp->pr_roflags & PR_PHINPAGE) { + continue; + } + LIST_REMOVE(ph, ph_hashlist); + s = splhigh(); + pool_put(&phpool, ph); + splx(s); + } } @@ -1285,19 +1410,18 @@ 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)) == NULL) - goto out; - - pp = drainpp; - drainpp = TAILQ_NEXT(pp, pr_poollist); - - pool_reclaim(pp); - - out: + if (drainpp == NULL) { + drainpp = TAILQ_FIRST(&pool_head); + } + if (drainpp) { + pp = drainpp; + drainpp = TAILQ_NEXT(pp, pr_poollist); + } simple_unlock(&pool_head_slock); + pool_reclaim(pp); splx(s); } @@ -1400,8 +1524,7 @@ pool_print1(struct pool *pp, const char (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)) { + TAILQ_FOREACH(pi, &ph->ph_itemlist, pi_list) { if (pi->pi_magic != PI_MAGIC) { (*pr)("\t\t\titem %p, magic 0x%x\n", pi, pi->pi_magic); @@ -1430,14 +1553,12 @@ pool_print1(struct pool *pp, const char if (print_cache == 0) goto skip_cache; - for (pc = TAILQ_FIRST(&pp->pr_cachelist); pc != NULL; - pc = TAILQ_NEXT(pc, pc_poollist)) { + TAILQ_FOREACH(pc, &pp->pr_cachelist, 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)) { + TAILQ_FOREACH(pcg, &pc->pc_grouplist, 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]); @@ -1457,9 +1578,7 @@ pool_chk(struct pool *pp, const char *la simple_lock(&pp->pr_slock); - for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; - ph = TAILQ_NEXT(ph, ph_pagelist)) { - + TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist) { struct pool_item *pi; int n; caddr_t page; @@ -1609,11 +1728,15 @@ 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) { - for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL; - pcg = TAILQ_NEXT(pcg, pcg_list)) { + TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) { if (pcg->pcg_avail != 0) { pc->pc_allocfrom = pcg; goto have_group; @@ -1660,12 +1783,12 @@ 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) { - for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL; - pcg = TAILQ_NEXT(pcg, pcg_list)) { + TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) { if (pcg->pcg_avail != PCG_NOBJECTS) { pc->pc_freeto = pcg; goto have_group; @@ -1677,7 +1800,9 @@ 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); @@ -1729,10 +1854,11 @@ 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 *, const char *, long)) + void (*putit)(struct pool *, void *)) { struct pool_cache_group *pcg, *npcg; void *object; + int s; for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL; pcg = npcg) { @@ -1744,14 +1870,16 @@ 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, __FILE__, __LINE__); + (*putit)(pc->pc_pool, object); } 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); } } } @@ -1767,7 +1895,7 @@ 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); }