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 1.128.2.12 retrieving revision 1.182.4.3 diff -u -p -r1.128.2.12 -r1.182.4.3 --- src/sys/kern/subr_pool.c 2007/10/29 16:37:44 1.128.2.12 +++ src/sys/kern/subr_pool.c 2011/03/05 20:55:19 1.182.4.3 @@ -1,7 +1,8 @@ -/* $NetBSD: subr_pool.c,v 1.128.2.12 2007/10/29 16:37:44 ad Exp $ */ +/* $NetBSD: subr_pool.c,v 1.182.4.3 2011/03/05 20:55:19 rmind Exp $ */ /*- - * Copyright (c) 1997, 1999, 2000, 2002, 2007 The NetBSD Foundation, Inc. + * Copyright (c) 1997, 1999, 2000, 2002, 2007, 2008, 2010 + * The NetBSD Foundation, Inc. * All rights reserved. * * This code is derived from software contributed to The NetBSD Foundation @@ -16,13 +17,6 @@ * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. - * 3. All advertising materials mentioning features or use of this software - * must display the following acknowledgement: - * This product includes software developed by the NetBSD - * Foundation, Inc. and its contributors. - * 4. Neither the name of The NetBSD Foundation nor the names of its - * contributors may be used to endorse or promote products derived - * from this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED @@ -38,26 +32,32 @@ */ #include -__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.128.2.12 2007/10/29 16:37:44 ad Exp $"); +__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.182.4.3 2011/03/05 20:55:19 rmind Exp $"); +#include "opt_ddb.h" #include "opt_pool.h" #include "opt_poollog.h" #include "opt_lockdebug.h" #include #include +#include #include #include #include #include -#include #include #include #include #include #include +#include +#include -#include +#include +#ifdef DIAGNOSTIC +#include /* uvm_km_va_drain */ +#endif /* * Pool resource management utility. @@ -73,16 +73,13 @@ __KERNEL_RCSID(0, "$NetBSD: subr_pool.c, */ /* List of all pools */ -LIST_HEAD(,pool) pool_head = LIST_HEAD_INITIALIZER(pool_head); - -/* List of all caches. */ -LIST_HEAD(,pool_cache) pool_cache_head = - LIST_HEAD_INITIALIZER(pool_cache_head); +static TAILQ_HEAD(, pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head); /* Private pool for page header structures */ #define PHPOOL_MAX 8 static struct pool phpool[PHPOOL_MAX]; -#define PHPOOL_FREELIST_NELEM(idx) (((idx) == 0) ? 0 : (1 << (idx))) +#define PHPOOL_FREELIST_NELEM(idx) \ + (((idx) == 0) ? 0 : BITMAP_SIZE * (1 << (idx))) #ifdef POOL_SUBPAGE /* Pool of subpages for use by normal pools. */ @@ -111,7 +108,12 @@ static struct pool *drainpp; static kmutex_t pool_head_lock; static kcondvar_t pool_busy; -typedef uint8_t pool_item_freelist_t; +/* This lock protects initialization of a potentially shared pool allocator */ +static kmutex_t pool_allocator_lock; + +typedef uint32_t pool_item_bitmap_t; +#define BITMAP_SIZE (CHAR_BIT * sizeof(pool_item_bitmap_t)) +#define BITMAP_MASK (BITMAP_SIZE - 1) struct pool_item_header { /* Page headers */ @@ -120,7 +122,9 @@ struct pool_item_header { SPLAY_ENTRY(pool_item_header) ph_node; /* Off-page page headers */ void * ph_page; /* this page's address */ - struct timeval ph_time; /* last referenced */ + uint32_t ph_time; /* last referenced */ + uint16_t ph_nmissing; /* # of chunks in use */ + uint16_t ph_off; /* start offset in page */ union { /* !PR_NOTOUCH */ struct { @@ -129,21 +133,12 @@ struct pool_item_header { } phu_normal; /* PR_NOTOUCH */ struct { - uint16_t - phu_off; /* start offset in page */ - pool_item_freelist_t - phu_firstfree; /* first free item */ - /* - * XXX it might be better to use - * a simple bitmap and ffs(3) - */ + pool_item_bitmap_t phu_bitmap[1]; } phu_notouch; } ph_u; - uint16_t ph_nmissing; /* # of chunks in use */ }; #define ph_itemlist ph_u.phu_normal.phu_itemlist -#define ph_off ph_u.phu_notouch.phu_off -#define ph_firstfree ph_u.phu_notouch.phu_firstfree +#define ph_bitmap ph_u.phu_notouch.phu_bitmap struct pool_item { #ifdef DIAGNOSTIC @@ -182,16 +177,25 @@ struct pool_item { * from it. */ -static struct pool pcgpool; +static struct pool pcg_normal_pool; +static struct pool pcg_large_pool; static struct pool cache_pool; static struct pool cache_cpu_pool; -static pool_cache_cpu_t *pool_cache_put_slow(pool_cache_cpu_t *, int *, - void *, paddr_t); -static pool_cache_cpu_t *pool_cache_get_slow(pool_cache_cpu_t *, int *, - void **, paddr_t *, int); +/* List of all caches. */ +TAILQ_HEAD(,pool_cache) pool_cache_head = + TAILQ_HEAD_INITIALIZER(pool_cache_head); + +int pool_cache_disable; /* global disable for caching */ +static const pcg_t pcg_dummy; /* zero sized: always empty, yet always full */ + +static bool pool_cache_put_slow(pool_cache_cpu_t *, int, + void *); +static bool pool_cache_get_slow(pool_cache_cpu_t *, int, + void **, paddr_t *, int); static void pool_cache_cpu_init1(struct cpu_info *, pool_cache_t); static void pool_cache_invalidate_groups(pool_cache_t, pcg_t *); +static void pool_cache_invalidate_cpu(pool_cache_t, u_int); static void pool_cache_xcall(pool_cache_t); static int pool_catchup(struct pool *); @@ -234,16 +238,28 @@ int pool_logsize = POOL_LOGSIZE; static inline void pr_log(struct pool *pp, void *v, int action, const char *file, long line) { - int n = pp->pr_curlogentry; + int n; struct pool_log *pl; if ((pp->pr_roflags & PR_LOGGING) == 0) return; + if (pp->pr_log == NULL) { + if (kmem_map != NULL) + pp->pr_log = malloc( + pool_logsize * sizeof(struct pool_log), + M_TEMP, M_NOWAIT | M_ZERO); + if (pp->pr_log == NULL) + return; + pp->pr_curlogentry = 0; + pp->pr_logsize = pool_logsize; + } + /* * Fill in the current entry. Wrap around and overwrite * the oldest entry if necessary. */ + n = pp->pr_curlogentry; pl = &pp->pr_log[n]; pl->pl_file = file; pl->pl_line = line; @@ -261,7 +277,7 @@ pr_printlog(struct pool *pp, struct pool int i = pp->pr_logsize; int n = pp->pr_curlogentry; - if ((pp->pr_roflags & PR_LOGGING) == 0) + if (pp->pr_log == NULL) return; /* @@ -329,12 +345,12 @@ pr_enter_check(struct pool *pp, void (*p #define pr_enter_check(pp, pr) #endif /* POOL_DIAGNOSTIC */ -static inline int +static inline unsigned int pr_item_notouch_index(const struct pool *pp, const struct pool_item_header *ph, const void *v) { const char *cp = v; - int idx; + unsigned int idx; KASSERT(pp->pr_roflags & PR_NOTOUCH); idx = (cp - (char *)ph->ph_page - ph->ph_off) / pp->pr_size; @@ -342,37 +358,57 @@ pr_item_notouch_index(const struct pool return idx; } -#define PR_FREELIST_ALIGN(p) \ - roundup((uintptr_t)(p), sizeof(pool_item_freelist_t)) -#define PR_FREELIST(ph) ((pool_item_freelist_t *)PR_FREELIST_ALIGN((ph) + 1)) -#define PR_INDEX_USED ((pool_item_freelist_t)-1) -#define PR_INDEX_EOL ((pool_item_freelist_t)-2) - static inline void pr_item_notouch_put(const struct pool *pp, struct pool_item_header *ph, void *obj) { - int idx = pr_item_notouch_index(pp, ph, obj); - pool_item_freelist_t *freelist = PR_FREELIST(ph); + unsigned int idx = pr_item_notouch_index(pp, ph, obj); + pool_item_bitmap_t *bitmap = ph->ph_bitmap + (idx / BITMAP_SIZE); + pool_item_bitmap_t mask = 1 << (idx & BITMAP_MASK); - KASSERT(freelist[idx] == PR_INDEX_USED); - freelist[idx] = ph->ph_firstfree; - ph->ph_firstfree = idx; + KASSERT((*bitmap & mask) == 0); + *bitmap |= mask; } static inline void * pr_item_notouch_get(const struct pool *pp, struct pool_item_header *ph) { - int idx = ph->ph_firstfree; - pool_item_freelist_t *freelist = PR_FREELIST(ph); + pool_item_bitmap_t *bitmap = ph->ph_bitmap; + unsigned int idx; + int i; - KASSERT(freelist[idx] != PR_INDEX_USED); - ph->ph_firstfree = freelist[idx]; - freelist[idx] = PR_INDEX_USED; + for (i = 0; ; i++) { + int bit; + KASSERT((i * BITMAP_SIZE) < pp->pr_itemsperpage); + bit = ffs32(bitmap[i]); + if (bit) { + pool_item_bitmap_t mask; + + bit--; + idx = (i * BITMAP_SIZE) + bit; + mask = 1 << bit; + KASSERT((bitmap[i] & mask) != 0); + bitmap[i] &= ~mask; + break; + } + } + KASSERT(idx < pp->pr_itemsperpage); return (char *)ph->ph_page + ph->ph_off + idx * pp->pr_size; } +static inline void +pr_item_notouch_init(const struct pool *pp, struct pool_item_header *ph) +{ + pool_item_bitmap_t *bitmap = ph->ph_bitmap; + const int n = howmany(pp->pr_itemsperpage, BITMAP_SIZE); + int i; + + for (i = 0; i < n; i++) { + bitmap[i] = (pool_item_bitmap_t)-1; + } +} + static inline int phtree_compare(struct pool_item_header *a, struct pool_item_header *b) { @@ -393,6 +429,24 @@ phtree_compare(struct pool_item_header * SPLAY_PROTOTYPE(phtree, pool_item_header, ph_node, phtree_compare); SPLAY_GENERATE(phtree, pool_item_header, ph_node, phtree_compare); +static inline struct pool_item_header * +pr_find_pagehead_noalign(struct pool *pp, void *v) +{ + struct pool_item_header *ph, tmp; + + tmp.ph_page = (void *)(uintptr_t)v; + ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp); + if (ph == NULL) { + ph = SPLAY_ROOT(&pp->pr_phtree); + if (ph != NULL && phtree_compare(&tmp, ph) >= 0) { + ph = SPLAY_NEXT(phtree, &pp->pr_phtree, ph); + } + KASSERT(ph == NULL || phtree_compare(&tmp, ph) < 0); + } + + return ph; +} + /* * Return the pool page header based on item address. */ @@ -402,15 +456,7 @@ pr_find_pagehead(struct pool *pp, void * struct pool_item_header *ph, tmp; if ((pp->pr_roflags & PR_NOALIGN) != 0) { - tmp.ph_page = (void *)(uintptr_t)v; - ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp); - if (ph == NULL) { - ph = SPLAY_ROOT(&pp->pr_phtree); - if (ph != NULL && phtree_compare(&tmp, ph) >= 0) { - ph = SPLAY_NEXT(phtree, &pp->pr_phtree, ph); - } - KASSERT(ph == NULL || phtree_compare(&tmp, ph) < 0); - } + ph = pr_find_pagehead_noalign(pp, v); } else { void *page = (void *)((uintptr_t)v & pp->pr_alloc->pa_pagemask); @@ -519,6 +565,11 @@ pool_reclaim_register(struct pool *pp) callback_register(&vm_map_to_kernel(map)->vmk_reclaim_callback, &pp->pr_reclaimerentry, pp, pool_reclaim_callback); splx(s); + +#ifdef DIAGNOSTIC + /* Diagnostic drain attempt. */ + uvm_km_va_drain(map, 0); +#endif } static void @@ -561,17 +612,11 @@ void pool_subsystem_init(void) { struct pool_allocator *pa; - __link_set_decl(pools, struct link_pool_init); - struct link_pool_init * const *pi; mutex_init(&pool_head_lock, MUTEX_DEFAULT, IPL_NONE); + mutex_init(&pool_allocator_lock, MUTEX_DEFAULT, IPL_NONE); cv_init(&pool_busy, "poolbusy"); - __link_set_foreach(pi, pools) - pool_init((*pi)->pp, (*pi)->size, (*pi)->align, - (*pi)->align_offset, (*pi)->flags, (*pi)->wchan, - (*pi)->palloc, (*pi)->ipl); - while ((pa = SLIST_FIRST(&pa_deferinitq)) != NULL) { KASSERT(pa->pa_backingmapptr != NULL); KASSERT(*pa->pa_backingmapptr != NULL); @@ -579,10 +624,10 @@ pool_subsystem_init(void) pa_reclaim_register(pa); } - pool_init(&cache_pool, sizeof(struct pool_cache), CACHE_LINE_SIZE, + pool_init(&cache_pool, sizeof(struct pool_cache), coherency_unit, 0, 0, "pcache", &pool_allocator_nointr, IPL_NONE); - pool_init(&cache_cpu_pool, sizeof(pool_cache_cpu_t), CACHE_LINE_SIZE, + pool_init(&cache_cpu_pool, sizeof(pool_cache_cpu_t), coherency_unit, 0, 0, "pcachecpu", &pool_allocator_nointr, IPL_NONE); } @@ -596,21 +641,16 @@ void pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags, const char *wchan, struct pool_allocator *palloc, int ipl) { -#ifdef DEBUG struct pool *pp1; -#endif size_t trysize, phsize; int off, slack; - KASSERT((1UL << (CHAR_BIT * sizeof(pool_item_freelist_t))) - 2 >= - PHPOOL_FREELIST_NELEM(PHPOOL_MAX - 1)); - #ifdef DEBUG /* * Check that the pool hasn't already been initialised and * added to the list of all pools. */ - LIST_FOREACH(pp1, &pool_head, pr_poollist) { + TAILQ_FOREACH(pp1, &pool_head, pr_poollist) { if (pp == pp1) panic("pool_init: pool %s already initialised", wchan); @@ -635,7 +675,9 @@ pool_init(struct pool *pp, size_t size, palloc = &pool_allocator_nointr_fullpage; } #endif /* POOL_SUBPAGE */ - if ((palloc->pa_flags & PA_INITIALIZED) == 0) { + if (!cold) + mutex_enter(&pool_allocator_lock); + if (palloc->pa_refcnt++ == 0) { if (palloc->pa_pagesz == 0) palloc->pa_pagesz = PAGE_SIZE; @@ -648,8 +690,9 @@ pool_init(struct pool *pp, size_t size, if (palloc->pa_backingmapptr != NULL) { pa_reclaim_register(palloc); } - palloc->pa_flags |= PA_INITIALIZED; } + if (!cold) + mutex_exit(&pool_allocator_lock); if (align == 0) align = ALIGN(1); @@ -772,16 +815,7 @@ pool_init(struct pool *pp, size_t size, pp->pr_nidle = 0; pp->pr_refcnt = 0; -#ifdef POOL_DIAGNOSTIC - if (flags & PR_LOGGING) { - if (kmem_map == NULL || - (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log), - M_TEMP, M_NOWAIT)) == NULL) - pp->pr_roflags &= ~PR_LOGGING; - pp->pr_curlogentry = 0; - pp->pr_logsize = pool_logsize; - } -#endif + pp->pr_log = NULL; pp->pr_entered_file = NULL; pp->pr_entered_line = 0; @@ -807,8 +841,8 @@ pool_init(struct pool *pp, size_t size, "phpool-%d", nelem); sz = sizeof(struct pool_item_header); if (nelem) { - sz = PR_FREELIST_ALIGN(sz) - + nelem * sizeof(pool_item_freelist_t); + sz = offsetof(struct pool_item_header, + ph_bitmap[howmany(nelem, BITMAP_SIZE)]); } pool_init(&phpool[idx], sz, 0, 0, 0, phpool_names[idx], &pool_allocator_meta, IPL_VM); @@ -817,24 +851,38 @@ pool_init(struct pool *pp, size_t size, pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0, PR_RECURSIVE, "psppool", &pool_allocator_meta, IPL_VM); #endif - pool_init(&pcgpool, sizeof(pcg_t), CACHE_LINE_SIZE, 0, 0, - "cachegrp", &pool_allocator_meta, IPL_VM); + + size = sizeof(pcg_t) + + (PCG_NOBJECTS_NORMAL - 1) * sizeof(pcgpair_t); + pool_init(&pcg_normal_pool, size, coherency_unit, 0, 0, + "pcgnormal", &pool_allocator_meta, IPL_VM); + + size = sizeof(pcg_t) + + (PCG_NOBJECTS_LARGE - 1) * sizeof(pcgpair_t); + pool_init(&pcg_large_pool, size, coherency_unit, 0, 0, + "pcglarge", &pool_allocator_meta, IPL_VM); } - if (__predict_true(!cold)) { - /* Insert into the list of all pools. */ + /* Insert into the list of all pools. */ + if (!cold) mutex_enter(&pool_head_lock); - LIST_INSERT_HEAD(&pool_head, pp, pr_poollist); + TAILQ_FOREACH(pp1, &pool_head, pr_poollist) { + if (strcmp(pp1->pr_wchan, pp->pr_wchan) > 0) + break; + } + if (pp1 == NULL) + TAILQ_INSERT_TAIL(&pool_head, pp, pr_poollist); + else + TAILQ_INSERT_BEFORE(pp1, pp, pr_poollist); + if (!cold) mutex_exit(&pool_head_lock); - /* Insert this into the list of pools using this allocator. */ + /* Insert this into the list of pools using this allocator. */ + if (!cold) mutex_enter(&palloc->pa_lock); - TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list); + TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list); + if (!cold) mutex_exit(&palloc->pa_lock); - } else { - LIST_INSERT_HEAD(&pool_head, pp, pr_poollist); - TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list); - } pool_reclaim_register(pp); } @@ -852,7 +900,7 @@ pool_destroy(struct pool *pp) mutex_enter(&pool_head_lock); while (pp->pr_refcnt != 0) cv_wait(&pool_busy, &pool_head_lock); - LIST_REMOVE(pp, pr_poollist); + TAILQ_REMOVE(&pool_head, pp, pr_poollist); if (drainpp == pp) drainpp = NULL; mutex_exit(&pool_head_lock); @@ -863,6 +911,11 @@ pool_destroy(struct pool *pp) TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list); mutex_exit(&pp->pr_alloc->pa_lock); + mutex_enter(&pool_allocator_lock); + if (--pp->pr_alloc->pa_refcnt == 0) + mutex_destroy(&pp->pr_alloc->pa_lock); + mutex_exit(&pool_allocator_lock); + mutex_enter(&pp->pr_lock); KASSERT(pp->pr_cache == NULL); @@ -888,8 +941,10 @@ pool_destroy(struct pool *pp) pr_pagelist_free(pp, &pq); #ifdef POOL_DIAGNOSTIC - if ((pp->pr_roflags & PR_LOGGING) != 0) + if (pp->pr_log != NULL) { free(pp->pr_log, M_TEMP); + pp->pr_log = NULL; + } #endif cv_destroy(&pp->pr_cv); @@ -937,18 +992,17 @@ pool_get(struct pool *pp, int flags) void *v; #ifdef DIAGNOSTIC - if (__predict_false(pp->pr_itemsperpage == 0)) - panic("pool_get: pool %p: pr_itemsperpage is zero, " - "pool not initialized?", pp); - if (__predict_false(curlwp == NULL && doing_shutdown == 0 && - (flags & PR_WAITOK) != 0)) - panic("pool_get: %s: must have NOWAIT", pp->pr_wchan); - -#endif /* DIAGNOSTIC */ -#ifdef LOCKDEBUG - if (flags & PR_WAITOK) - ASSERT_SLEEPABLE(NULL, "pool_get(PR_WAITOK)"); + if (pp->pr_itemsperpage == 0) + panic("pool_get: pool '%s': pr_itemsperpage is zero, " + "pool not initialized?", pp->pr_wchan); + if ((cpu_intr_p() || cpu_softintr_p()) && pp->pr_ipl == IPL_NONE && + !cold && panicstr == NULL) + panic("pool '%s' is IPL_NONE, but called from " + "interrupt context\n", pp->pr_wchan); #endif + if (flags & PR_WAITOK) { + ASSERT_SLEEPABLE(); + } mutex_enter(&pp->pr_lock); pr_enter(pp, file, line); @@ -1217,10 +1271,7 @@ pool_do_put(struct pool *pp, void *v, st if (pp->pr_flags & PR_WANTED) { pp->pr_flags &= ~PR_WANTED; - if (ph->ph_nmissing == 0) - pp->pr_nidle++; cv_broadcast(&pp->pr_cv); - return; } /* @@ -1239,8 +1290,7 @@ pool_do_put(struct pool *pp, void *v, st if (ph->ph_nmissing == 0) { pp->pr_nidle++; if (pp->pr_npages > pp->pr_minpages && - (pp->pr_npages > pp->pr_maxpages || - pa_starved_p(pp->pr_alloc))) { + pp->pr_npages > pp->pr_maxpages) { pr_rmpage(pp, ph, pq); } else { LIST_REMOVE(ph, ph_pagelist); @@ -1251,8 +1301,11 @@ pool_do_put(struct pool *pp, void *v, st * be idle for some period of time before it can * be reclaimed by the pagedaemon. This minimizes * ping-pong'ing for memory. + * + * note for 64-bit time_t: truncating to 32-bit is not + * a problem for our usage. */ - getmicrotime(&ph->ph_time); + ph->ph_time = time_uptime; } pool_update_curpage(pp); } @@ -1404,7 +1457,7 @@ pool_prime_page(struct pool *pp, void *s LIST_INIT(&ph->ph_itemlist); ph->ph_page = storage; ph->ph_nmissing = 0; - getmicrotime(&ph->ph_time); + ph->ph_time = time_uptime; if ((pp->pr_roflags & PR_PHINPAGE) == 0) SPLAY_INSERT(phtree, &pp->pr_phtree, ph); @@ -1413,7 +1466,8 @@ pool_prime_page(struct pool *pp, void *s /* * Color this page. */ - cp = (char *)cp + pp->pr_curcolor; + ph->ph_off = pp->pr_curcolor; + cp = (char *)cp + ph->ph_off; if ((pp->pr_curcolor += align) > pp->pr_maxcolor) pp->pr_curcolor = 0; @@ -1432,14 +1486,7 @@ pool_prime_page(struct pool *pp, void *s pp->pr_nitems += n; if (pp->pr_roflags & PR_NOTOUCH) { - pool_item_freelist_t *freelist = PR_FREELIST(ph); - int i; - - ph->ph_off = (char *)cp - (char *)storage; - ph->ph_firstfree = 0; - for (i = 0; i < n - 1; i++) - freelist[i] = i + 1; - freelist[n - 1] = PR_INDEX_EOL; + pr_item_notouch_init(pp, ph); } else { while (n--) { pi = (struct pool_item *)cp; @@ -1498,6 +1545,8 @@ pool_update_curpage(struct pool *pp) if (pp->pr_curpage == NULL) { pp->pr_curpage = LIST_FIRST(&pp->pr_emptypages); } + KASSERT((pp->pr_curpage == NULL && pp->pr_nitems == 0) || + (pp->pr_curpage != NULL && pp->pr_nitems > 0)); } void @@ -1561,6 +1610,8 @@ pool_sethardlimit(struct pool *pp, int n /* * Release all complete pages that have not been used recently. + * + * Might be called from interrupt context. */ int #ifdef POOL_DIAGNOSTIC @@ -1571,7 +1622,13 @@ pool_reclaim(struct pool *pp) { struct pool_item_header *ph, *phnext; struct pool_pagelist pq; - struct timeval curtime, diff; + uint32_t curtime; + bool klock; + int rv; + + if (cpu_intr_p() || cpu_softintr_p()) { + KASSERT(pp->pr_ipl != IPL_NONE); + } if (pp->pr_drain_hook != NULL) { /* @@ -1580,17 +1637,32 @@ pool_reclaim(struct pool *pp) (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT); } + /* + * XXXSMP Because we do not want to cause non-MPSAFE code + * to block. + */ + if (pp->pr_ipl == IPL_SOFTNET || pp->pr_ipl == IPL_SOFTCLOCK || + pp->pr_ipl == IPL_SOFTSERIAL) { + KERNEL_LOCK(1, NULL); + klock = true; + } else + klock = false; + /* Reclaim items from the pool's cache (if any). */ if (pp->pr_cache != NULL) pool_cache_invalidate(pp->pr_cache); - if (mutex_tryenter(&pp->pr_lock) == 0) + if (mutex_tryenter(&pp->pr_lock) == 0) { + if (klock) { + KERNEL_UNLOCK_ONE(NULL); + } return (0); + } pr_enter(pp, file, line); LIST_INIT(&pq); - getmicrotime(&curtime); + curtime = time_uptime; for (ph = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) { phnext = LIST_NEXT(ph, ph_pagelist); @@ -1600,8 +1672,7 @@ pool_reclaim(struct pool *pp) break; KASSERT(ph->ph_nmissing == 0); - timersub(&curtime, &ph->ph_time, &diff); - if (diff.tv_sec < pool_inactive_time + if (curtime - ph->ph_time < pool_inactive_time && !pa_starved_p(pp->pr_alloc)) continue; @@ -1618,12 +1689,19 @@ pool_reclaim(struct pool *pp) pr_leave(pp); mutex_exit(&pp->pr_lock); + if (LIST_EMPTY(&pq)) - return 0; + rv = 0; + else { + pr_pagelist_free(pp, &pq); + rv = 1; + } - pr_pagelist_free(pp, &pq); + if (klock) { + KERNEL_UNLOCK_ONE(NULL); + } - return (1); + return (rv); } /* @@ -1640,7 +1718,7 @@ pool_drain_start(struct pool **ppp, uint { struct pool *pp; - KASSERT(!LIST_EMPTY(&pool_head)); + KASSERT(!TAILQ_EMPTY(&pool_head)); pp = NULL; @@ -1648,11 +1726,11 @@ pool_drain_start(struct pool **ppp, uint mutex_enter(&pool_head_lock); do { if (drainpp == NULL) { - drainpp = LIST_FIRST(&pool_head); + drainpp = TAILQ_FIRST(&pool_head); } if (drainpp != NULL) { pp = drainpp; - drainpp = LIST_NEXT(pp, pr_poollist); + drainpp = TAILQ_NEXT(pp, pr_poollist); } /* * Skip completely idle pools. We depend on at least @@ -1670,12 +1748,13 @@ pool_drain_start(struct pool **ppp, uint } } -void +bool pool_drain_end(struct pool *pp, uint64_t where) { + bool reclaimed; if (pp == NULL) - return; + return false; KASSERT(pp->pr_refcnt > 0); @@ -1684,13 +1763,15 @@ pool_drain_end(struct pool *pp, uint64_t xc_wait(where); /* Drain the cache (if any) and pool.. */ - pool_reclaim(pp); + reclaimed = pool_reclaim(pp); /* Finally, unlock the pool. */ mutex_enter(&pool_head_lock); pp->pr_refcnt--; cv_broadcast(&pool_busy); mutex_exit(&pool_head_lock); + + return reclaimed; } /* @@ -1708,7 +1789,7 @@ pool_printall(const char *modif, void (* { struct pool *pp; - LIST_FOREACH(pp, &pool_head, pr_poollist) { + TAILQ_FOREACH(pp, &pool_head, pr_poollist) { pool_printit(pp, modif, pr); } } @@ -1735,10 +1816,8 @@ pool_print_pagelist(struct pool *pp, str #endif LIST_FOREACH(ph, pl, 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); + (*pr)("\t\tpage %p, nmissing %d, time %" PRIu32 "\n", + ph->ph_page, ph->ph_nmissing, ph->ph_time); #ifdef DIAGNOSTIC if (!(pp->pr_roflags & PR_NOTOUCH)) { LIST_FOREACH(pi, &ph->ph_itemlist, pi_list) { @@ -1825,7 +1904,7 @@ pool_print1(struct pool *pp, const char #define PR_GROUPLIST(pcg) \ (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail); \ - for (i = 0; i < PCG_NOBJECTS; i++) { \ + for (i = 0; i < pcg->pcg_size; i++) { \ if (pcg->pcg_objects[i].pcgo_pa != \ POOL_PADDR_INVALID) { \ (*pr)("\t\t\t%p, 0x%llx\n", \ @@ -1841,7 +1920,7 @@ pool_print1(struct pool *pp, const char if (pc != NULL) { cpuhit = 0; cpumiss = 0; - for (i = 0; i < MAXCPUS; i++) { + for (i = 0; i < __arraycount(pc->pc_cpus); i++) { if ((cc = pc->pc_cpus[i]) == NULL) continue; cpuhit += cc->cc_hits; @@ -1998,6 +2077,7 @@ pool_cache_bootstrap(pool_cache_t pc, si void *arg) { CPU_INFO_ITERATOR cii; + pool_cache_t pc1; struct cpu_info *ci; struct pool *pp; @@ -2005,8 +2085,7 @@ pool_cache_bootstrap(pool_cache_t pc, si if (palloc == NULL && ipl == IPL_NONE) palloc = &pool_allocator_nointr; pool_init(pp, size, align, align_offset, flags, wchan, palloc, ipl); - - mutex_init(&pc->pc_lock, MUTEX_DEFAULT, pp->pr_ipl); + mutex_init(&pc->pc_lock, MUTEX_DEFAULT, ipl); if (ctor == NULL) { ctor = (int (*)(void *, void *, int))nullop; @@ -2028,25 +2107,44 @@ pool_cache_bootstrap(pool_cache_t pc, si pc->pc_nfull = 0; pc->pc_contended = 0; pc->pc_refcnt = 0; + pc->pc_freecheck = NULL; + + if ((flags & PR_LARGECACHE) != 0) { + pc->pc_pcgsize = PCG_NOBJECTS_LARGE; + pc->pc_pcgpool = &pcg_large_pool; + } else { + pc->pc_pcgsize = PCG_NOBJECTS_NORMAL; + pc->pc_pcgpool = &pcg_normal_pool; + } /* Allocate per-CPU caches. */ memset(pc->pc_cpus, 0, sizeof(pc->pc_cpus)); pc->pc_ncpu = 0; - for (CPU_INFO_FOREACH(cii, ci)) { - pool_cache_cpu_init1(ci, pc); + if (ncpu < 2) { + /* XXX For sparc: boot CPU is not attached yet. */ + pool_cache_cpu_init1(curcpu(), pc); + } else { + for (CPU_INFO_FOREACH(cii, ci)) { + pool_cache_cpu_init1(ci, pc); + } } - - if (__predict_true(!cold)) { - mutex_enter(&pp->pr_lock); - pp->pr_cache = pc; - mutex_exit(&pp->pr_lock); + + /* Add to list of all pools. */ + if (__predict_true(!cold)) mutex_enter(&pool_head_lock); - LIST_INSERT_HEAD(&pool_cache_head, pc, pc_cachelist); - mutex_exit(&pool_head_lock); - } else { - pp->pr_cache = pc; - LIST_INSERT_HEAD(&pool_cache_head, pc, pc_cachelist); + TAILQ_FOREACH(pc1, &pool_cache_head, pc_cachelist) { + if (strcmp(pc1->pc_pool.pr_wchan, pc->pc_pool.pr_wchan) > 0) + break; } + if (pc1 == NULL) + TAILQ_INSERT_TAIL(&pool_cache_head, pc, pc_cachelist); + else + TAILQ_INSERT_BEFORE(pc1, pc, pc_cachelist); + if (__predict_true(!cold)) + mutex_exit(&pool_head_lock); + + membar_sync(); + pp->pr_cache = pc; } /* @@ -2058,15 +2156,13 @@ void pool_cache_destroy(pool_cache_t pc) { struct pool *pp = &pc->pc_pool; - pool_cache_cpu_t *cc; - pcg_t *pcg; - int i; + u_int i; /* Remove it from the global list. */ mutex_enter(&pool_head_lock); while (pc->pc_refcnt != 0) cv_wait(&pool_busy, &pool_head_lock); - LIST_REMOVE(pc, pc_cachelist); + TAILQ_REMOVE(&pool_cache_head, pc, pc_cachelist); mutex_exit(&pool_head_lock); /* First, invalidate the entire cache. */ @@ -2078,20 +2174,8 @@ pool_cache_destroy(pool_cache_t pc) mutex_exit(&pp->pr_lock); /* Destroy per-CPU data */ - for (i = 0; i < MAXCPUS; i++) { - if ((cc = pc->pc_cpus[i]) == NULL) - continue; - if ((pcg = cc->cc_current) != NULL) { - pcg->pcg_next = NULL; - pool_cache_invalidate_groups(pc, pcg); - } - if ((pcg = cc->cc_previous) != NULL) { - pcg->pcg_next = NULL; - pool_cache_invalidate_groups(pc, pcg); - } - if (cc != &pc->pc_cpu0) - pool_put(&cache_cpu_pool, cc); - } + for (i = 0; i < __arraycount(pc->pc_cpus); i++) + pool_cache_invalidate_cpu(pc, i); /* Finally, destroy it. */ mutex_destroy(&pc->pc_lock); @@ -2108,11 +2192,14 @@ static void pool_cache_cpu_init1(struct cpu_info *ci, pool_cache_t pc) { pool_cache_cpu_t *cc; + int index; + + index = ci->ci_index; - KASSERT(((uintptr_t)pc->pc_cpus & (CACHE_LINE_SIZE - 1)) == 0); + KASSERT(index < __arraycount(pc->pc_cpus)); - if ((cc = pc->pc_cpus[ci->ci_index]) != NULL) { - KASSERT(cc->cc_cpu = ci); + if ((cc = pc->pc_cpus[index]) != NULL) { + KASSERT(cc->cc_cpuindex == index); return; } @@ -2133,13 +2220,13 @@ pool_cache_cpu_init1(struct cpu_info *ci cc->cc_ipl = pc->pc_pool.pr_ipl; cc->cc_iplcookie = makeiplcookie(cc->cc_ipl); cc->cc_cache = pc; - cc->cc_cpu = ci; + cc->cc_cpuindex = index; cc->cc_hits = 0; cc->cc_misses = 0; - cc->cc_current = NULL; - cc->cc_previous = NULL; + cc->cc_current = __UNCONST(&pcg_dummy); + cc->cc_previous = __UNCONST(&pcg_dummy); - pc->pc_cpus[ci->ci_index] = cc; + pc->pc_cpus[index] = cc; } /* @@ -2153,7 +2240,7 @@ pool_cache_cpu_init(struct cpu_info *ci) pool_cache_t pc; mutex_enter(&pool_head_lock); - LIST_FOREACH(pc, &pool_cache_head, pc_cachelist) { + TAILQ_FOREACH(pc, &pool_cache_head, pc_cachelist) { pc->pc_refcnt++; mutex_exit(&pool_head_lock); @@ -2178,6 +2265,14 @@ pool_cache_reclaim(pool_cache_t pc) return pool_reclaim(&pc->pc_pool); } +static void +pool_cache_destruct_object1(pool_cache_t pc, void *object) +{ + + (*pc->pc_dtor)(pc->pc_arg, object); + pool_put(&pc->pc_pool, object); +} + /* * pool_cache_destruct_object: * @@ -2188,8 +2283,9 @@ void pool_cache_destruct_object(pool_cache_t pc, void *object) { - (*pc->pc_dtor)(pc->pc_arg, object); - pool_put(&pc->pc_pool, object); + FREECHECK_IN(&pc->pc_freecheck, object); + + pool_cache_destruct_object1(pc, object); } /* @@ -2209,10 +2305,15 @@ pool_cache_invalidate_groups(pool_cache_ for (i = 0; i < pcg->pcg_avail; i++) { object = pcg->pcg_objects[i].pcgo_va; - pool_cache_destruct_object(pc, object); + pool_cache_destruct_object1(pc, object); } - pool_put(&pcgpool, pcg); + if (pcg->pcg_size == PCG_NOBJECTS_LARGE) { + pool_put(&pcg_large_pool, pcg); + } else { + KASSERT(pcg->pcg_size == PCG_NOBJECTS_NORMAL); + pool_put(&pcg_normal_pool, pcg); + } } } @@ -2221,12 +2322,35 @@ pool_cache_invalidate_groups(pool_cache_ * * Invalidate a pool cache (destruct and release all of the * cached objects). Does not reclaim objects from the pool. + * + * Note: For pool caches that provide constructed objects, there + * is an assumption that another level of synchronization is occurring + * between the input to the constructor and the cache invalidation. */ void pool_cache_invalidate(pool_cache_t pc) { pcg_t *full, *empty, *part; +#if 0 + uint64_t where; + if (ncpu < 2 || !mp_online) { + /* + * We might be called early enough in the boot process + * for the CPU data structures to not be fully initialized. + * In this case, simply gather the local CPU's cache now + * since it will be the only one running. + */ + pool_cache_xcall(pc); + } else { + /* + * Gather all of the CPU-specific caches into the + * global cache. + */ + where = xc_broadcast(0, (xcfunc_t)pool_cache_xcall, pc, NULL); + xc_wait(where); + } +#endif mutex_enter(&pc->pc_lock); full = pc->pc_fullgroups; empty = pc->pc_emptygroups; @@ -2244,6 +2368,40 @@ pool_cache_invalidate(pool_cache_t pc) pool_cache_invalidate_groups(pc, part); } +/* + * pool_cache_invalidate_cpu: + * + * Invalidate all CPU-bound cached objects in pool cache, the CPU being + * identified by its associated index. + * It is caller's responsibility to ensure that no operation is + * taking place on this pool cache while doing this invalidation. + * WARNING: as no inter-CPU locking is enforced, trying to invalidate + * pool cached objects from a CPU different from the one currently running + * may result in an undefined behaviour. + */ +static void +pool_cache_invalidate_cpu(pool_cache_t pc, u_int index) +{ + + pool_cache_cpu_t *cc; + pcg_t *pcg; + + if ((cc = pc->pc_cpus[index]) == NULL) + return; + + if ((pcg = cc->cc_current) != &pcg_dummy) { + pcg->pcg_next = NULL; + pool_cache_invalidate_groups(pc, pcg); + } + if ((pcg = cc->cc_previous) != &pcg_dummy) { + pcg->pcg_next = NULL; + pool_cache_invalidate_groups(pc, pcg); + } + if (cc != &pc->pc_cpu0) + pool_put(&cache_cpu_pool, cc); + +} + void pool_cache_set_drain_hook(pool_cache_t pc, void (*fn)(void *, int), void *arg) { @@ -2272,58 +2430,8 @@ pool_cache_sethardlimit(pool_cache_t pc, pool_sethardlimit(&pc->pc_pool, n, warnmess, ratecap); } -static inline pool_cache_cpu_t * -pool_cache_cpu_enter(pool_cache_t pc, int *s) -{ - pool_cache_cpu_t *cc; - struct cpu_info *ci; - - /* - * Prevent other users of the cache from accessing our - * CPU-local data. To avoid touching shared state, we - * pull the neccessary information from CPU local data. - */ - ci = curcpu(); - KASSERT(ci->ci_data.cpu_index < MAXCPUS); - cc = pc->pc_cpus[ci->ci_data.cpu_index]; - KASSERT(cc->cc_cache == pc); - if (cc->cc_ipl == IPL_NONE) { - crit_enter(); - } else { - *s = splraiseipl(cc->cc_iplcookie); - } - - /* Moved to another CPU before disabling preemption? */ - if (__predict_false(ci != curcpu())) { - ci = curcpu(); - cc = pc->pc_cpus[ci->ci_data.cpu_index]; - } - -#ifdef DIAGNOSTIC - KASSERT(cc->cc_cpu == ci); - KASSERT(((uintptr_t)cc & (CACHE_LINE_SIZE - 1)) == 0); -#endif - - return cc; -} - -static inline void -pool_cache_cpu_exit(pool_cache_cpu_t *cc, int *s) -{ - - /* No longer need exclusive access to the per-CPU data. */ - if (cc->cc_ipl == IPL_NONE) { - crit_exit(); - } else { - splx(*s); - } -} - -#if __GNUC_PREREQ__(3, 0) -__attribute ((noinline)) -#endif -pool_cache_cpu_t * -pool_cache_get_slow(pool_cache_cpu_t *cc, int *s, void **objectp, +static bool __noinline +pool_cache_get_slow(pool_cache_cpu_t *cc, int s, void **objectp, paddr_t *pap, int flags) { pcg_t *pcg, *cur; @@ -2331,6 +2439,9 @@ pool_cache_get_slow(pool_cache_cpu_t *cc pool_cache_t pc; void *object; + KASSERT(cc->cc_current->pcg_avail == 0); + KASSERT(cc->cc_previous->pcg_avail == 0); + pc = cc->cc_cache; cc->cc_misses++; @@ -2338,7 +2449,7 @@ pool_cache_get_slow(pool_cache_cpu_t *cc * Nothing was available locally. Try and grab a group * from the cache. */ - if (!mutex_tryenter(&pc->pc_lock)) { + if (__predict_false(!mutex_tryenter(&pc->pc_lock))) { ncsw = curlwp->l_ncsw; mutex_enter(&pc->pc_lock); pc->pc_contended++; @@ -2350,30 +2461,29 @@ pool_cache_get_slow(pool_cache_cpu_t *cc */ if (curlwp->l_ncsw != ncsw) { mutex_exit(&pc->pc_lock); - pool_cache_cpu_exit(cc, s); - return pool_cache_cpu_enter(pc, s); + return true; } } - if ((pcg = pc->pc_fullgroups) != NULL) { + if (__predict_true((pcg = pc->pc_fullgroups) != NULL)) { /* * If there's a full group, release our empty * group back to the cache. Install the full * group as cc_current and return. */ - if ((cur = cc->cc_current) != NULL) { + if (__predict_true((cur = cc->cc_current) != &pcg_dummy)) { KASSERT(cur->pcg_avail == 0); cur->pcg_next = pc->pc_emptygroups; pc->pc_emptygroups = cur; pc->pc_nempty++; } - KASSERT(pcg->pcg_avail == PCG_NOBJECTS); + KASSERT(pcg->pcg_avail == pcg->pcg_size); cc->cc_current = pcg; pc->pc_fullgroups = pcg->pcg_next; pc->pc_hits++; pc->pc_nfull--; mutex_exit(&pc->pc_lock); - return cc; + return true; } /* @@ -2383,17 +2493,17 @@ pool_cache_get_slow(pool_cache_cpu_t *cc */ pc->pc_misses++; mutex_exit(&pc->pc_lock); - pool_cache_cpu_exit(cc, s); + splx(s); object = pool_get(&pc->pc_pool, flags); *objectp = object; - if (object == NULL) - return NULL; + if (__predict_false(object == NULL)) + return false; - if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) { + if (__predict_false((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0)) { pool_put(&pc->pc_pool, object); *objectp = NULL; - return NULL; + return false; } KASSERT((((vaddr_t)object + pc->pc_pool.pr_itemoffset) & @@ -2408,7 +2518,7 @@ pool_cache_get_slow(pool_cache_cpu_t *cc } FREECHECK_OUT(&pc->pc_freecheck, object); - return NULL; + return false; } /* @@ -2425,24 +2535,33 @@ pool_cache_get_paddr(pool_cache_t pc, in void *object; int s; -#ifdef LOCKDEBUG - if (flags & PR_WAITOK) - ASSERT_SLEEPABLE(NULL, "pool_cache_get(PR_WAITOK)"); -#endif + KASSERTMSG((!cpu_intr_p() && !cpu_softintr_p()) || + (pc->pc_pool.pr_ipl != IPL_NONE || cold || panicstr != NULL), + ("pool '%s' is IPL_NONE, but called from interrupt context\n", + pc->pc_pool.pr_wchan)); + + if (flags & PR_WAITOK) { + ASSERT_SLEEPABLE(); + } - cc = pool_cache_cpu_enter(pc, &s); - do { + /* Lock out interrupts and disable preemption. */ + s = splvm(); + while (/* CONSTCOND */ true) { /* Try and allocate an object from the current group. */ + cc = pc->pc_cpus[curcpu()->ci_index]; + KASSERT(cc->cc_cache == pc); pcg = cc->cc_current; - if (pcg != NULL && pcg->pcg_avail > 0) { + if (__predict_true(pcg->pcg_avail > 0)) { object = pcg->pcg_objects[--pcg->pcg_avail].pcgo_va; - if (pap != NULL) + if (__predict_false(pap != NULL)) *pap = pcg->pcg_objects[pcg->pcg_avail].pcgo_pa; +#if defined(DIAGNOSTIC) pcg->pcg_objects[pcg->pcg_avail].pcgo_va = NULL; - KASSERT(pcg->pcg_avail <= PCG_NOBJECTS); + KASSERT(pcg->pcg_avail < pcg->pcg_size); KASSERT(object != NULL); +#endif cc->cc_hits++; - pool_cache_cpu_exit(cc, &s); + splx(s); FREECHECK_OUT(&pc->pc_freecheck, object); return object; } @@ -2452,7 +2571,7 @@ pool_cache_get_paddr(pool_cache_t pc, in * it with the current group and allocate from there. */ pcg = cc->cc_previous; - if (pcg != NULL && pcg->pcg_avail > 0) { + if (__predict_true(pcg->pcg_avail > 0)) { cc->cc_previous = cc->cc_current; cc->cc_current = pcg; continue; @@ -2461,104 +2580,105 @@ pool_cache_get_paddr(pool_cache_t pc, in /* * Can't allocate from either group: try the slow path. * If get_slow() allocated an object for us, or if - * no more objects are available, it will return NULL. + * no more objects are available, it will return false. * Otherwise, we need to retry. */ - cc = pool_cache_get_slow(cc, &s, &object, pap, flags); - } while (cc != NULL); + if (!pool_cache_get_slow(cc, s, &object, pap, flags)) + break; + } return object; } -#if __GNUC_PREREQ__(3, 0) -__attribute ((noinline)) -#endif -pool_cache_cpu_t * -pool_cache_put_slow(pool_cache_cpu_t *cc, int *s, void *object, paddr_t pa) +static bool __noinline +pool_cache_put_slow(pool_cache_cpu_t *cc, int s, void *object) { pcg_t *pcg, *cur; uint64_t ncsw; pool_cache_t pc; + KASSERT(cc->cc_current->pcg_avail == cc->cc_current->pcg_size); + KASSERT(cc->cc_previous->pcg_avail == cc->cc_previous->pcg_size); + pc = cc->cc_cache; + pcg = NULL; cc->cc_misses++; /* - * No free slots locally. Try to grab an empty, unused - * group from the cache. + * If there are no empty groups in the cache then allocate one + * while still unlocked. */ - if (!mutex_tryenter(&pc->pc_lock)) { + if (__predict_false(pc->pc_emptygroups == NULL)) { + if (__predict_true(!pool_cache_disable)) { + pcg = pool_get(pc->pc_pcgpool, PR_NOWAIT); + } + if (__predict_true(pcg != NULL)) { + pcg->pcg_avail = 0; + pcg->pcg_size = pc->pc_pcgsize; + } + } + + /* Lock the cache. */ + if (__predict_false(!mutex_tryenter(&pc->pc_lock))) { ncsw = curlwp->l_ncsw; mutex_enter(&pc->pc_lock); pc->pc_contended++; /* - * If we context switched while locking, then - * our view of the per-CPU data is invalid: - * retry. + * If we context switched while locking, then our view of + * the per-CPU data is invalid: retry. */ - if (curlwp->l_ncsw != ncsw) { + if (__predict_false(curlwp->l_ncsw != ncsw)) { mutex_exit(&pc->pc_lock); - pool_cache_cpu_exit(cc, s); - return pool_cache_cpu_enter(pc, s); + if (pcg != NULL) { + pool_put(pc->pc_pcgpool, pcg); + } + return true; } } - if ((pcg = pc->pc_emptygroups) != NULL) { - /* - * If there's a empty group, release our full - * group back to the cache. Install the empty - * group as cc_current and return. - */ - if ((cur = cc->cc_current) != NULL) { - KASSERT(cur->pcg_avail == PCG_NOBJECTS); - cur->pcg_next = pc->pc_fullgroups; - pc->pc_fullgroups = cur; - pc->pc_nfull++; - } - KASSERT(pcg->pcg_avail == 0); - cc->cc_current = pcg; + /* If there are no empty groups in the cache then allocate one. */ + if (pcg == NULL && pc->pc_emptygroups != NULL) { + pcg = pc->pc_emptygroups; pc->pc_emptygroups = pcg->pcg_next; - pc->pc_hits++; pc->pc_nempty--; - mutex_exit(&pc->pc_lock); - return cc; } /* - * Nothing available locally or in cache. Take the - * slow path and try to allocate a new group that we - * can release to. + * If there's a empty group, release our full group back + * to the cache. Install the empty group to the local CPU + * and return. */ - pc->pc_misses++; - mutex_exit(&pc->pc_lock); - pool_cache_cpu_exit(cc, s); - - /* - * If we can't allocate a new group, just throw the - * object away. - */ - pcg = pool_get(&pcgpool, PR_NOWAIT); - if (pcg == NULL) { - pool_cache_destruct_object(pc, object); - return NULL; + if (pcg != NULL) { + KASSERT(pcg->pcg_avail == 0); + if (__predict_false(cc->cc_previous == &pcg_dummy)) { + cc->cc_previous = pcg; + } else { + cur = cc->cc_current; + if (__predict_true(cur != &pcg_dummy)) { + KASSERT(cur->pcg_avail == cur->pcg_size); + cur->pcg_next = pc->pc_fullgroups; + pc->pc_fullgroups = cur; + pc->pc_nfull++; + } + cc->cc_current = pcg; + } + pc->pc_hits++; + mutex_exit(&pc->pc_lock); + return true; } -#ifdef DIAGNOSTIC - memset(pcg, 0, sizeof(*pcg)); -#else - pcg->pcg_avail = 0; -#endif /* - * Add the empty group to the cache and try again. + * Nothing available locally or in cache, and we didn't + * allocate an empty group. Take the slow path and destroy + * the object here and now. */ - mutex_enter(&pc->pc_lock); - pcg->pcg_next = pc->pc_emptygroups; - pc->pc_emptygroups = pcg; - pc->pc_nempty++; + pc->pc_misses++; mutex_exit(&pc->pc_lock); + splx(s); + pool_cache_destruct_object(pc, object); - return pool_cache_cpu_enter(pc, s); + return false; } /* @@ -2574,29 +2694,31 @@ pool_cache_put_paddr(pool_cache_t pc, vo pcg_t *pcg; int s; + KASSERT(object != NULL); FREECHECK_IN(&pc->pc_freecheck, object); - cc = pool_cache_cpu_enter(pc, &s); - do { + /* Lock out interrupts and disable preemption. */ + s = splvm(); + while (/* CONSTCOND */ true) { /* If the current group isn't full, release it there. */ + cc = pc->pc_cpus[curcpu()->ci_index]; + KASSERT(cc->cc_cache == pc); pcg = cc->cc_current; - if (pcg != NULL && pcg->pcg_avail < PCG_NOBJECTS) { - KASSERT(pcg->pcg_objects[pcg->pcg_avail].pcgo_va - == NULL); + if (__predict_true(pcg->pcg_avail < pcg->pcg_size)) { pcg->pcg_objects[pcg->pcg_avail].pcgo_va = object; pcg->pcg_objects[pcg->pcg_avail].pcgo_pa = pa; pcg->pcg_avail++; cc->cc_hits++; - pool_cache_cpu_exit(cc, &s); + splx(s); return; } /* - * That failed. If the previous group is empty, swap + * That failed. If the previous group isn't full, swap * it with the current group and try again. */ pcg = cc->cc_previous; - if (pcg != NULL && pcg->pcg_avail == 0) { + if (__predict_true(pcg->pcg_avail < pcg->pcg_size)) { cc->cc_previous = cc->cc_current; cc->cc_current = pcg; continue; @@ -2605,10 +2727,11 @@ pool_cache_put_paddr(pool_cache_t pc, vo /* * Can't free to either group: try the slow path. * If put_slow() releases the object for us, it - * will return NULL. Otherwise we need to retry. + * will return false. Otherwise we need to retry. */ - cc = pool_cache_put_slow(cc, &s, object, pa); - } while (cc != NULL); + if (!pool_cache_put_slow(cc, s, object)) + break; + } } /* @@ -2622,18 +2745,17 @@ pool_cache_xcall(pool_cache_t pc) { pool_cache_cpu_t *cc; pcg_t *prev, *cur, **list; - int s = 0; /* XXXgcc */ + int s; - cc = pool_cache_cpu_enter(pc, &s); + s = splvm(); + mutex_enter(&pc->pc_lock); + cc = pc->pc_cpus[curcpu()->ci_index]; cur = cc->cc_current; - cc->cc_current = NULL; + cc->cc_current = __UNCONST(&pcg_dummy); prev = cc->cc_previous; - cc->cc_previous = NULL; - pool_cache_cpu_exit(cc, &s); - - mutex_enter(&pc->pc_lock); - if (cur != NULL) { - if (cur->pcg_avail == PCG_NOBJECTS) { + cc->cc_previous = __UNCONST(&pcg_dummy); + if (cur != &pcg_dummy) { + if (cur->pcg_avail == cur->pcg_size) { list = &pc->pc_fullgroups; pc->pc_nfull++; } else if (cur->pcg_avail == 0) { @@ -2646,8 +2768,8 @@ pool_cache_xcall(pool_cache_t pc) cur->pcg_next = *list; *list = cur; } - if (prev != NULL) { - if (prev->pcg_avail == PCG_NOBJECTS) { + if (prev != &pcg_dummy) { + if (prev->pcg_avail == prev->pcg_size) { list = &pc->pc_fullgroups; pc->pc_nfull++; } else if (prev->pcg_avail == 0) { @@ -2661,6 +2783,7 @@ pool_cache_xcall(pool_cache_t pc) *list = prev; } mutex_exit(&pc->pc_lock); + splx(s); } /* @@ -2826,3 +2949,143 @@ pool_page_free_nointr(struct pool *pp, v uvm_km_free_poolpage_cache(kernel_map, (vaddr_t) v); } + +#if defined(DDB) +static bool +pool_in_page(struct pool *pp, struct pool_item_header *ph, uintptr_t addr) +{ + + return (uintptr_t)ph->ph_page <= addr && + addr < (uintptr_t)ph->ph_page + pp->pr_alloc->pa_pagesz; +} + +static bool +pool_in_item(struct pool *pp, void *item, uintptr_t addr) +{ + + return (uintptr_t)item <= addr && addr < (uintptr_t)item + pp->pr_size; +} + +static bool +pool_in_cg(struct pool *pp, struct pool_cache_group *pcg, uintptr_t addr) +{ + int i; + + if (pcg == NULL) { + return false; + } + for (i = 0; i < pcg->pcg_avail; i++) { + if (pool_in_item(pp, pcg->pcg_objects[i].pcgo_va, addr)) { + return true; + } + } + return false; +} + +static bool +pool_allocated(struct pool *pp, struct pool_item_header *ph, uintptr_t addr) +{ + + if ((pp->pr_roflags & PR_NOTOUCH) != 0) { + unsigned int idx = pr_item_notouch_index(pp, ph, (void *)addr); + pool_item_bitmap_t *bitmap = + ph->ph_bitmap + (idx / BITMAP_SIZE); + pool_item_bitmap_t mask = 1 << (idx & BITMAP_MASK); + + return (*bitmap & mask) == 0; + } else { + struct pool_item *pi; + + LIST_FOREACH(pi, &ph->ph_itemlist, pi_list) { + if (pool_in_item(pp, pi, addr)) { + return false; + } + } + return true; + } +} + +void +pool_whatis(uintptr_t addr, void (*pr)(const char *, ...)) +{ + struct pool *pp; + + TAILQ_FOREACH(pp, &pool_head, pr_poollist) { + struct pool_item_header *ph; + uintptr_t item; + bool allocated = true; + bool incache = false; + bool incpucache = false; + char cpucachestr[32]; + + if ((pp->pr_roflags & PR_PHINPAGE) != 0) { + LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) { + if (pool_in_page(pp, ph, addr)) { + goto found; + } + } + LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) { + if (pool_in_page(pp, ph, addr)) { + allocated = + pool_allocated(pp, ph, addr); + goto found; + } + } + LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) { + if (pool_in_page(pp, ph, addr)) { + allocated = false; + goto found; + } + } + continue; + } else { + ph = pr_find_pagehead_noalign(pp, (void *)addr); + if (ph == NULL || !pool_in_page(pp, ph, addr)) { + continue; + } + allocated = pool_allocated(pp, ph, addr); + } +found: + if (allocated && pp->pr_cache) { + pool_cache_t pc = pp->pr_cache; + struct pool_cache_group *pcg; + int i; + + for (pcg = pc->pc_fullgroups; pcg != NULL; + pcg = pcg->pcg_next) { + if (pool_in_cg(pp, pcg, addr)) { + incache = true; + goto print; + } + } + for (i = 0; i < __arraycount(pc->pc_cpus); i++) { + pool_cache_cpu_t *cc; + + if ((cc = pc->pc_cpus[i]) == NULL) { + continue; + } + if (pool_in_cg(pp, cc->cc_current, addr) || + pool_in_cg(pp, cc->cc_previous, addr)) { + struct cpu_info *ci = + cpu_lookup(i); + + incpucache = true; + snprintf(cpucachestr, + sizeof(cpucachestr), + "cached by CPU %u", + ci->ci_index); + goto print; + } + } + } +print: + item = (uintptr_t)ph->ph_page + ph->ph_off; + item = item + rounddown(addr - item, pp->pr_size); + (*pr)("%p is %p+%zu in POOL '%s' (%s)\n", + (void *)addr, item, (size_t)(addr - item), + pp->pr_wchan, + incpucache ? cpucachestr : + incache ? "cached" : allocated ? "allocated" : "free"); + } +} +#endif /* defined(DDB) */