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 1.101.2.10 retrieving revision 1.111.4.1 diff -u -p -r1.101.2.10 -r1.111.4.1 --- src/sys/kern/subr_pool.c 2008/02/11 14:59:58 1.101.2.10 +++ src/sys/kern/subr_pool.c 2006/02/04 14:30:17 1.111.4.1 @@ -1,12 +1,12 @@ -/* $NetBSD: subr_pool.c,v 1.101.2.10 2008/02/11 14:59:58 yamt Exp $ */ +/* $NetBSD: subr_pool.c,v 1.111.4.1 2006/02/04 14:30:17 simonb Exp $ */ /*- - * Copyright (c) 1997, 1999, 2000, 2002, 2007 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 * by Paul Kranenburg; by Jason R. Thorpe of the Numerical Aerospace - * Simulation Facility, NASA Ames Research Center, and by Andrew Doran. + * Simulation Facility, NASA Ames Research Center. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions @@ -38,27 +38,21 @@ */ #include -__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.101.2.10 2008/02/11 14:59:58 yamt Exp $"); +__KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.111.4.1 2006/02/04 14:30:17 simonb Exp $"); -#include "opt_ddb.h" #include "opt_pool.h" #include "opt_poollog.h" #include "opt_lockdebug.h" #include #include -#include #include #include #include #include +#include #include #include -#include -#include -#include -#include -#include #include @@ -76,29 +70,24 @@ __KERNEL_RCSID(0, "$NetBSD: subr_pool.c, */ /* List of all pools */ -TAILQ_HEAD(,pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head); +LIST_HEAD(,pool) pool_head = LIST_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 : BITMAP_SIZE * (1 << (idx))) +#define PHPOOL_FREELIST_NELEM(idx) (((idx) == 0) ? 0 : (1 << (idx))) #ifdef POOL_SUBPAGE /* Pool of subpages for use by normal pools. */ static struct pool psppool; #endif -static SLIST_HEAD(, pool_allocator) pa_deferinitq = - SLIST_HEAD_INITIALIZER(pa_deferinitq); - static void *pool_page_alloc_meta(struct pool *, int); static void pool_page_free_meta(struct pool *, void *); /* allocator for pool metadata */ -struct pool_allocator pool_allocator_meta = { - pool_page_alloc_meta, pool_page_free_meta, - .pa_backingmapptr = &kmem_map, +static struct pool_allocator pool_allocator_meta = { + pool_page_alloc_meta, pool_page_free_meta }; /* # of seconds to retain page after last use */ @@ -107,13 +96,10 @@ int pool_inactive_time = 10; /* Next candidate for drainage (see pool_drain()) */ static struct pool *drainpp; -/* This lock protects both pool_head and drainpp. */ -static kmutex_t pool_head_lock; -static kcondvar_t pool_busy; - -typedef uint32_t pool_item_bitmap_t; -#define BITMAP_SIZE (CHAR_BIT * sizeof(pool_item_bitmap_t)) -#define BITMAP_MASK (BITMAP_SIZE - 1) +/* This spin lock protects both pool_head and drainpp. */ +struct simplelock pool_head_slock = SIMPLELOCK_INITIALIZER; + +typedef uint8_t pool_item_freelist_t; struct pool_item_header { /* Page headers */ @@ -121,10 +107,8 @@ struct pool_item_header { ph_pagelist; /* pool page list */ SPLAY_ENTRY(pool_item_header) ph_node; /* Off-page page headers */ - void * ph_page; /* this page's address */ + caddr_t ph_page; /* this page's address */ struct timeval ph_time; /* last referenced */ - uint16_t ph_nmissing; /* # of chunks in use */ - uint16_t ph_off; /* start offset in page */ union { /* !PR_NOTOUCH */ struct { @@ -133,18 +117,27 @@ struct pool_item_header { } phu_normal; /* PR_NOTOUCH */ struct { - pool_item_bitmap_t phu_bitmap[1]; + 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) + */ } phu_notouch; } ph_u; + uint16_t ph_nmissing; /* # of chunks in use */ }; #define ph_itemlist ph_u.phu_normal.phu_itemlist -#define ph_bitmap ph_u.phu_notouch.phu_bitmap +#define ph_off ph_u.phu_notouch.phu_off +#define ph_firstfree ph_u.phu_notouch.phu_firstfree struct pool_item { #ifdef DIAGNOSTIC u_int pi_magic; #endif -#define PI_MAGIC 0xdeaddeadU +#define PI_MAGIC 0xdeadbeefU /* Other entries use only this list entry */ LIST_ENTRY(pool_item) pi_list; }; @@ -160,51 +153,38 @@ struct pool_item { * needless object construction/destruction; it is deferred until absolutely * necessary. * - * Caches are grouped into cache groups. Each cache group references up - * to PCG_NUMOBJECTS 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. - * - * 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. + * 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. */ -static struct pool pcg_normal_pool; -static struct pool pcg_large_pool; -static struct pool cache_pool; -static struct pool cache_cpu_pool; - -/* List of all caches. */ -TAILQ_HEAD(,pool_cache) pool_cache_head = - TAILQ_HEAD_INITIALIZER(pool_cache_head); - -int pool_cache_disable; - - -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); -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_xcall(pool_cache_t); +/* The cache group pool. */ +static struct pool pcgpool; + +static void pool_cache_reclaim(struct pool_cache *, struct pool_pagelist *, + struct pool_cache_grouplist *); +static void pcg_grouplist_free(struct pool_cache_grouplist *); static int pool_catchup(struct pool *); -static void pool_prime_page(struct pool *, void *, +static void pool_prime_page(struct pool *, caddr_t, struct pool_item_header *); static void pool_update_curpage(struct pool *); -static int pool_grow(struct pool *, int); -static void *pool_allocator_alloc(struct pool *, int); -static void pool_allocator_free(struct pool *, void *); +void *pool_allocator_alloc(struct pool *, int); +void pool_allocator_free(struct pool *, void *); static void pool_print_pagelist(struct pool *, struct pool_pagelist *, void (*)(const char *, ...)); @@ -332,83 +312,57 @@ pr_enter_check(struct pool *pp, void (*p #define pr_enter_check(pp, pr) #endif /* POOL_DIAGNOSTIC */ -static inline unsigned int +static inline int pr_item_notouch_index(const struct pool *pp, const struct pool_item_header *ph, const void *v) { const char *cp = v; - unsigned int idx; + int idx; KASSERT(pp->pr_roflags & PR_NOTOUCH); - idx = (cp - (char *)ph->ph_page - ph->ph_off) / pp->pr_size; + idx = (cp - ph->ph_page - ph->ph_off) / pp->pr_size; KASSERT(idx < pp->pr_itemsperpage); 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) { - 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); + int idx = pr_item_notouch_index(pp, ph, obj); + pool_item_freelist_t *freelist = PR_FREELIST(ph); - KASSERT((*bitmap & mask) == 0); - *bitmap |= mask; + KASSERT(freelist[idx] == PR_INDEX_USED); + freelist[idx] = ph->ph_firstfree; + ph->ph_firstfree = idx; } static inline void * pr_item_notouch_get(const struct pool *pp, struct pool_item_header *ph) { - pool_item_bitmap_t *bitmap = ph->ph_bitmap; - unsigned int idx; - int i; - - 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; -} + int idx = ph->ph_firstfree; + pool_item_freelist_t *freelist = PR_FREELIST(ph); -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; + KASSERT(freelist[idx] != PR_INDEX_USED); + ph->ph_firstfree = freelist[idx]; + freelist[idx] = PR_INDEX_USED; - for (i = 0; i < n; i++) { - bitmap[i] = (pool_item_bitmap_t)-1; - } + return ph->ph_page + ph->ph_off + idx * pp->pr_size; } static inline int phtree_compare(struct pool_item_header *a, struct pool_item_header *b) { - - /* - * we consider pool_item_header with smaller ph_page bigger. - * (this unnatural ordering is for the benefit of pr_find_pagehead.) - */ - if (a->ph_page < b->ph_page) - return (1); - else if (a->ph_page > b->ph_page) return (-1); + else if (a->ph_page > b->ph_page) + return (1); else return (0); } @@ -416,49 +370,19 @@ 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. + * Return the pool page header based on page address. */ static inline struct pool_item_header * -pr_find_pagehead(struct pool *pp, void *v) +pr_find_pagehead(struct pool *pp, caddr_t page) { struct pool_item_header *ph, tmp; - if ((pp->pr_roflags & PR_NOALIGN) != 0) { - ph = pr_find_pagehead_noalign(pp, v); - } else { - void *page = - (void *)((uintptr_t)v & pp->pr_alloc->pa_pagemask); - - if ((pp->pr_roflags & PR_PHINPAGE) != 0) { - ph = (struct pool_item_header *)((char *)page + pp->pr_phoffset); - } else { - tmp.ph_page = page; - ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp); - } - } + if ((pp->pr_roflags & PR_PHINPAGE) != 0) + return ((struct pool_item_header *)(page + pp->pr_phoffset)); - KASSERT(ph == NULL || ((pp->pr_roflags & PR_PHINPAGE) != 0) || - ((char *)ph->ph_page <= (char *)v && - (char *)v < (char *)ph->ph_page + pp->pr_alloc->pa_pagesz)); + tmp.ph_page = page; + ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp); return ph; } @@ -466,12 +390,16 @@ static void pr_pagelist_free(struct pool *pp, struct pool_pagelist *pq) { struct pool_item_header *ph; + int s; while ((ph = LIST_FIRST(pq)) != NULL) { LIST_REMOVE(ph, ph_pagelist); pool_allocator_free(pp, ph->ph_page); - if ((pp->pr_roflags & PR_PHINPAGE) == 0) + if ((pp->pr_roflags & PR_PHINPAGE) == 0) { + s = splvm(); pool_put(pp->pr_phpool, ph); + splx(s); + } } } @@ -483,7 +411,7 @@ pr_rmpage(struct pool *pp, struct pool_i struct pool_pagelist *pq) { - KASSERT(mutex_owned(&pp->pr_lock)); + LOCK_ASSERT(simple_lock_held(&pp->pr_slock)); /* * If the page was idle, decrement the idle page count. @@ -514,109 +442,19 @@ pr_rmpage(struct pool *pp, struct pool_i pool_update_curpage(pp); } -static bool -pa_starved_p(struct pool_allocator *pa) -{ - - if (pa->pa_backingmap != NULL) { - return vm_map_starved_p(pa->pa_backingmap); - } - return false; -} - -static int -pool_reclaim_callback(struct callback_entry *ce, void *obj, void *arg) -{ - struct pool *pp = obj; - struct pool_allocator *pa = pp->pr_alloc; - - KASSERT(&pp->pr_reclaimerentry == ce); - pool_reclaim(pp); - if (!pa_starved_p(pa)) { - return CALLBACK_CHAIN_ABORT; - } - return CALLBACK_CHAIN_CONTINUE; -} - -static void -pool_reclaim_register(struct pool *pp) -{ - struct vm_map *map = pp->pr_alloc->pa_backingmap; - int s; - - if (map == NULL) { - return; - } - - s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */ - callback_register(&vm_map_to_kernel(map)->vmk_reclaim_callback, - &pp->pr_reclaimerentry, pp, pool_reclaim_callback); - splx(s); -} - -static void -pool_reclaim_unregister(struct pool *pp) -{ - struct vm_map *map = pp->pr_alloc->pa_backingmap; - int s; - - if (map == NULL) { - return; - } - - s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */ - callback_unregister(&vm_map_to_kernel(map)->vmk_reclaim_callback, - &pp->pr_reclaimerentry); - splx(s); -} - -static void -pa_reclaim_register(struct pool_allocator *pa) -{ - struct vm_map *map = *pa->pa_backingmapptr; - struct pool *pp; - - KASSERT(pa->pa_backingmap == NULL); - if (map == NULL) { - SLIST_INSERT_HEAD(&pa_deferinitq, pa, pa_q); - return; - } - pa->pa_backingmap = map; - TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) { - pool_reclaim_register(pp); - } -} - /* * Initialize all the pools listed in the "pools" link set. */ void -pool_subsystem_init(void) +link_pool_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); - 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); - SLIST_REMOVE_HEAD(&pa_deferinitq, pa_q); - pa_reclaim_register(pa); - } - - pool_init(&cache_pool, sizeof(struct pool_cache), CACHE_LINE_SIZE, - 0, 0, "pcache", &pool_allocator_nointr, IPL_NONE); - - pool_init(&cache_cpu_pool, sizeof(pool_cache_cpu_t), CACHE_LINE_SIZE, - 0, 0, "pcachecpu", &pool_allocator_nointr, IPL_NONE); + (*pi)->palloc); } /* @@ -627,23 +465,14 @@ pool_subsystem_init(void) */ 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) + const char *wchan, struct pool_allocator *palloc) { - struct pool *pp1; - size_t trysize, phsize; int off, slack; + size_t trysize, phsize; + int s; -#ifdef DEBUG - /* - * Check that the pool hasn't already been initialised and - * added to the list of all pools. - */ - TAILQ_FOREACH(pp1, &pool_head, pr_poollist) { - if (pp == pp1) - panic("pool_init: pool %s already initialised", - wchan); - } -#endif + KASSERT((1UL << (CHAR_BIT * sizeof(pool_item_freelist_t))) - 2 >= + PHPOOL_FREELIST_NELEM(PHPOOL_MAX - 1)); #ifdef POOL_DIAGNOSTIC /* @@ -653,42 +482,55 @@ pool_init(struct pool *pp, size_t size, flags |= PR_LOGGING; #endif +#ifdef POOL_SUBPAGE + /* + * XXX We don't provide a real `nointr' back-end + * yet; all sub-pages come from a kmem back-end. + * maybe some day... + */ + if (palloc == NULL) { + extern struct pool_allocator pool_allocator_kmem_subpage; + palloc = &pool_allocator_kmem_subpage; + } + /* + * We'll assume any user-specified back-end allocator + * will deal with sub-pages, or simply don't care. + */ +#else if (palloc == NULL) palloc = &pool_allocator_kmem; -#ifdef POOL_SUBPAGE - if (size > palloc->pa_pagesz) { - if (palloc == &pool_allocator_kmem) - palloc = &pool_allocator_kmem_fullpage; - else if (palloc == &pool_allocator_nointr) - palloc = &pool_allocator_nointr_fullpage; - } #endif /* POOL_SUBPAGE */ if ((palloc->pa_flags & PA_INITIALIZED) == 0) { - if (palloc->pa_pagesz == 0) + if (palloc->pa_pagesz == 0) { +#ifdef POOL_SUBPAGE + if (palloc == &pool_allocator_kmem) + palloc->pa_pagesz = PAGE_SIZE; + else + palloc->pa_pagesz = POOL_SUBPAGE; +#else palloc->pa_pagesz = PAGE_SIZE; +#endif /* POOL_SUBPAGE */ + } TAILQ_INIT(&palloc->pa_list); - mutex_init(&palloc->pa_lock, MUTEX_DEFAULT, IPL_VM); + simple_lock_init(&palloc->pa_slock); palloc->pa_pagemask = ~(palloc->pa_pagesz - 1); palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1; - - if (palloc->pa_backingmapptr != NULL) { - pa_reclaim_register(palloc); - } palloc->pa_flags |= PA_INITIALIZED; } if (align == 0) align = ALIGN(1); - if ((flags & PR_NOTOUCH) == 0 && size < sizeof(struct pool_item)) + if (size < sizeof(struct pool_item)) size = sizeof(struct pool_item); size = roundup(size, align); #ifdef DIAGNOSTIC if (size > palloc->pa_pagesz) - panic("pool_init: pool item size (%zu) too large", size); + panic("pool_init: pool item size (%lu) too large", + (u_long)size); #endif /* @@ -697,7 +539,7 @@ pool_init(struct pool *pp, size_t size, LIST_INIT(&pp->pr_emptypages); LIST_INIT(&pp->pr_fullpages); LIST_INIT(&pp->pr_partpages); - pp->pr_cache = NULL; + LIST_INIT(&pp->pr_cachelist); pp->pr_curpage = NULL; pp->pr_npages = 0; pp->pr_minitems = 0; @@ -719,7 +561,6 @@ pool_init(struct pool *pp, size_t size, pp->pr_hardlimit_warning_last.tv_usec = 0; pp->pr_drain_hook = NULL; pp->pr_drain_hook_arg = NULL; - pp->pr_freecheck = NULL; /* * Decide whether to put the page header off page to avoid @@ -738,7 +579,7 @@ pool_init(struct pool *pp, size_t size, /* See the comment below about reserved bytes. */ trysize = palloc->pa_pagesz - ((align - ioff) % align); phsize = ALIGN(sizeof(struct pool_item_header)); - if ((pp->pr_roflags & (PR_NOTOUCH | PR_NOALIGN)) == 0 && + if ((pp->pr_roflags & PR_NOTOUCH) == 0 && (pp->pr_size < MIN(palloc->pa_pagesz / 16, phsize << 3) || trysize / pp->pr_size == (trysize - phsize) / pp->pr_size)) { /* Use the end of the page for the page header */ @@ -798,7 +639,6 @@ pool_init(struct pool *pp, size_t size, pp->pr_npagefree = 0; pp->pr_hiwat = 0; pp->pr_nidle = 0; - pp->pr_refcnt = 0; #ifdef POOL_DIAGNOSTIC if (flags & PR_LOGGING) { @@ -814,16 +654,7 @@ pool_init(struct pool *pp, size_t size, pp->pr_entered_file = NULL; pp->pr_entered_line = 0; - /* - * XXXAD hack to prevent IP input processing from blocking. - */ - if (ipl == IPL_SOFTNET) { - mutex_init(&pp->pr_lock, MUTEX_DEFAULT, IPL_VM); - } else { - mutex_init(&pp->pr_lock, MUTEX_DEFAULT, ipl); - } - cv_init(&pp->pr_cv, wchan); - pp->pr_ipl = ipl; + simple_lock_init(&pp->pr_slock); /* * Initialize private page header pool and cache magazine pool if we @@ -842,50 +673,31 @@ pool_init(struct pool *pp, size_t size, "phpool-%d", nelem); sz = sizeof(struct pool_item_header); if (nelem) { - sz = offsetof(struct pool_item_header, - ph_bitmap[howmany(nelem, BITMAP_SIZE)]); + sz = PR_FREELIST_ALIGN(sz) + + nelem * sizeof(pool_item_freelist_t); } pool_init(&phpool[idx], sz, 0, 0, 0, - phpool_names[idx], &pool_allocator_meta, IPL_VM); + phpool_names[idx], &pool_allocator_meta); } #ifdef POOL_SUBPAGE pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0, - PR_RECURSIVE, "psppool", &pool_allocator_meta, IPL_VM); + PR_RECURSIVE, "psppool", &pool_allocator_meta); #endif - - size = sizeof(pcg_t) + - (PCG_NOBJECTS_NORMAL - 1) * sizeof(pcgpair_t); - pool_init(&pcg_normal_pool, size, CACHE_LINE_SIZE, 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, CACHE_LINE_SIZE, 0, 0, - "pcglarge", &pool_allocator_meta, IPL_VM); + pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0, + 0, "pcgpool", &pool_allocator_meta); } /* Insert into the list of all pools. */ - if (__predict_true(!cold)) - mutex_enter(&pool_head_lock); - 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 (__predict_true(!cold)) - mutex_exit(&pool_head_lock); - - /* Insert this into the list of pools using this allocator. */ - if (__predict_true(!cold)) - mutex_enter(&palloc->pa_lock); - TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list); - if (__predict_true(!cold)) - mutex_exit(&palloc->pa_lock); + simple_lock(&pool_head_slock); + LIST_INSERT_HEAD(&pool_head, pp, pr_poollist); + simple_unlock(&pool_head_slock); - pool_reclaim_register(pp); + /* Insert this into the list of pools using this allocator. */ + s = splvm(); + simple_lock(&palloc->pa_slock); + TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list); + simple_unlock(&palloc->pa_slock); + splx(s); } /* @@ -896,25 +708,26 @@ pool_destroy(struct pool *pp) { struct pool_pagelist pq; struct pool_item_header *ph; + int s; /* Remove from global pool list */ - mutex_enter(&pool_head_lock); - while (pp->pr_refcnt != 0) - cv_wait(&pool_busy, &pool_head_lock); - TAILQ_REMOVE(&pool_head, pp, pr_poollist); + simple_lock(&pool_head_slock); + LIST_REMOVE(pp, pr_poollist); if (drainpp == pp) drainpp = NULL; - mutex_exit(&pool_head_lock); + simple_unlock(&pool_head_slock); /* Remove this pool from its allocator's list of pools. */ - pool_reclaim_unregister(pp); - mutex_enter(&pp->pr_alloc->pa_lock); + s = splvm(); + simple_lock(&pp->pr_alloc->pa_slock); TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list); - mutex_exit(&pp->pr_alloc->pa_lock); + simple_unlock(&pp->pr_alloc->pa_slock); + splx(s); - mutex_enter(&pp->pr_lock); + s = splvm(); + simple_lock(&pp->pr_slock); - KASSERT(pp->pr_cache == NULL); + KASSERT(LIST_EMPTY(&pp->pr_cachelist)); #ifdef DIAGNOSTIC if (pp->pr_nout != 0) { @@ -932,7 +745,8 @@ pool_destroy(struct pool *pp) while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL) pr_rmpage(pp, ph, &pq); - mutex_exit(&pp->pr_lock); + simple_unlock(&pp->pr_slock); + splx(s); pr_pagelist_free(pp, &pq); @@ -940,9 +754,6 @@ pool_destroy(struct pool *pp) if ((pp->pr_roflags & PR_LOGGING) != 0) free(pp->pr_log, M_TEMP); #endif - - cv_destroy(&pp->pr_cv); - mutex_destroy(&pp->pr_lock); } void @@ -959,20 +770,26 @@ pool_set_drain_hook(struct pool *pp, voi } static struct pool_item_header * -pool_alloc_item_header(struct pool *pp, void *storage, int flags) +pool_alloc_item_header(struct pool *pp, caddr_t storage, int flags) { struct pool_item_header *ph; + int s; + + LOCK_ASSERT(simple_lock_held(&pp->pr_slock) == 0); if ((pp->pr_roflags & PR_PHINPAGE) != 0) - ph = (struct pool_item_header *) ((char *)storage + pp->pr_phoffset); - else + ph = (struct pool_item_header *) (storage + pp->pr_phoffset); + else { + s = splvm(); ph = pool_get(pp->pr_phpool, flags); + splx(s); + } return (ph); } /* - * Grab an item from the pool. + * Grab an item from the pool; must be called at appropriate spl level */ void * #ifdef POOL_DIAGNOSTIC @@ -996,10 +813,11 @@ pool_get(struct pool *pp, int flags) #endif /* DIAGNOSTIC */ #ifdef LOCKDEBUG if (flags & PR_WAITOK) - ASSERT_SLEEPABLE(NULL, "pool_get(PR_WAITOK)"); + simple_lock_only_held(NULL, "pool_get(PR_WAITOK)"); + SCHED_ASSERT_UNLOCKED(); #endif - mutex_enter(&pp->pr_lock); + simple_lock(&pp->pr_slock); pr_enter(pp, file, line); startover: @@ -1011,7 +829,7 @@ pool_get(struct pool *pp, int flags) #ifdef DIAGNOSTIC if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) { pr_leave(pp); - mutex_exit(&pp->pr_lock); + simple_unlock(&pp->pr_slock); panic("pool_get: %s: crossed hard limit", pp->pr_wchan); } #endif @@ -1023,9 +841,9 @@ pool_get(struct pool *pp, int flags) * and check the hardlimit condition again. */ pr_leave(pp); - mutex_exit(&pp->pr_lock); + simple_unlock(&pp->pr_slock); (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags); - mutex_enter(&pp->pr_lock); + simple_lock(&pp->pr_slock); pr_enter(pp, file, line); if (pp->pr_nout < pp->pr_hardlimit) goto startover; @@ -1038,7 +856,7 @@ pool_get(struct pool *pp, int flags) */ pp->pr_flags |= PR_WANTED; pr_leave(pp); - cv_wait(&pp->pr_cv, &pp->pr_lock); + ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock); pr_enter(pp, file, line); goto startover; } @@ -1054,7 +872,7 @@ pool_get(struct pool *pp, int flags) pp->pr_nfail++; pr_leave(pp); - mutex_exit(&pp->pr_lock); + simple_unlock(&pp->pr_slock); return (NULL); } @@ -1065,11 +883,9 @@ pool_get(struct pool *pp, int flags) * has no items in its bucket. */ if ((ph = pp->pr_curpage) == NULL) { - int error; - #ifdef DIAGNOSTIC if (pp->pr_nitems != 0) { - mutex_exit(&pp->pr_lock); + simple_unlock(&pp->pr_slock); printf("pool_get: %s: curpage NULL, nitems %u\n", pp->pr_wchan, pp->pr_nitems); panic("pool_get: nitems inconsistent"); @@ -1082,9 +898,18 @@ pool_get(struct pool *pp, int flags) * may block. */ pr_leave(pp); - error = pool_grow(pp, flags); - pr_enter(pp, file, line); - if (error != 0) { + simple_unlock(&pp->pr_slock); + v = pool_allocator_alloc(pp, flags); + if (__predict_true(v != NULL)) + ph = pool_alloc_item_header(pp, v, flags); + + if (__predict_false(v == NULL || ph == NULL)) { + if (v != NULL) + pool_allocator_free(pp, v); + + simple_lock(&pp->pr_slock); + pr_enter(pp, file, line); + /* * We were unable to allocate a page or item * header, but we released the lock during @@ -1094,12 +919,33 @@ pool_get(struct pool *pp, int flags) if (pp->pr_curpage != NULL) goto startover; - pp->pr_nfail++; + if ((flags & PR_WAITOK) == 0) { + pp->pr_nfail++; + pr_leave(pp); + simple_unlock(&pp->pr_slock); + return (NULL); + } + + /* + * Wait for items to be returned to this pool. + * + * wake up once a second and try again, + * as the check in pool_cache_put_paddr() is racy. + */ + pp->pr_flags |= PR_WANTED; + /* PA_WANTED is already set on the allocator. */ pr_leave(pp); - mutex_exit(&pp->pr_lock); - return (NULL); + ltsleep(pp, PSWP, pp->pr_wchan, hz, &pp->pr_slock); + pr_enter(pp, file, line); + goto startover; } + /* We have more memory; add it to the pool */ + simple_lock(&pp->pr_slock); + pr_enter(pp, file, line); + pool_prime_page(pp, v, ph); + pp->pr_npagealloc++; + /* Start the allocation process over. */ goto startover; } @@ -1107,7 +953,7 @@ pool_get(struct pool *pp, int flags) #ifdef DIAGNOSTIC if (__predict_false(ph->ph_nmissing == pp->pr_itemsperpage)) { pr_leave(pp); - mutex_exit(&pp->pr_lock); + simple_unlock(&pp->pr_slock); panic("pool_get: %s: page empty", pp->pr_wchan); } #endif @@ -1119,13 +965,13 @@ pool_get(struct pool *pp, int flags) v = pi = LIST_FIRST(&ph->ph_itemlist); if (__predict_false(v == NULL)) { pr_leave(pp); - mutex_exit(&pp->pr_lock); + simple_unlock(&pp->pr_slock); panic("pool_get: %s: page empty", pp->pr_wchan); } #ifdef DIAGNOSTIC if (__predict_false(pp->pr_nitems == 0)) { pr_leave(pp); - mutex_exit(&pp->pr_lock); + simple_unlock(&pp->pr_slock); printf("pool_get: %s: items on itemlist, nitems %u\n", pp->pr_wchan, pp->pr_nitems); panic("pool_get: nitems inconsistent"); @@ -1172,7 +1018,7 @@ pool_get(struct pool *pp, int flags) if (__predict_false((pp->pr_roflags & PR_NOTOUCH) == 0 && !LIST_EMPTY(&ph->ph_itemlist))) { pr_leave(pp); - mutex_exit(&pp->pr_lock); + simple_unlock(&pp->pr_slock); panic("pool_get: %s: nmissing inconsistent", pp->pr_wchan); } @@ -1201,9 +1047,7 @@ pool_get(struct pool *pp, int flags) */ } - mutex_exit(&pp->pr_lock); - KASSERT((((vaddr_t)v + pp->pr_itemoffset) & (pp->pr_align - 1)) == 0); - FREECHECK_OUT(&pp->pr_freecheck, v); + simple_unlock(&pp->pr_slock); return (v); } @@ -1215,10 +1059,12 @@ pool_do_put(struct pool *pp, void *v, st { struct pool_item *pi = v; struct pool_item_header *ph; + caddr_t page; - KASSERT(mutex_owned(&pp->pr_lock)); - FREECHECK_IN(&pp->pr_freecheck, v); - LOCKDEBUG_MEM_CHECK(v, pp->pr_size); + LOCK_ASSERT(simple_lock_held(&pp->pr_slock)); + SCHED_ASSERT_UNLOCKED(); + + page = (caddr_t)((u_long)v & pp->pr_alloc->pa_pagemask); #ifdef DIAGNOSTIC if (__predict_false(pp->pr_nout == 0)) { @@ -1228,11 +1074,18 @@ pool_do_put(struct pool *pp, void *v, st } #endif - if (__predict_false((ph = pr_find_pagehead(pp, v)) == NULL)) { + 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. */ @@ -1268,7 +1121,7 @@ pool_do_put(struct pool *pp, void *v, st pp->pr_flags &= ~PR_WANTED; if (ph->ph_nmissing == 0) pp->pr_nidle++; - cv_broadcast(&pp->pr_cv); + wakeup((caddr_t)pp); return; } @@ -1289,7 +1142,7 @@ pool_do_put(struct pool *pp, void *v, st pp->pr_nidle++; if (pp->pr_npages > pp->pr_minpages && (pp->pr_npages > pp->pr_maxpages || - pa_starved_p(pp->pr_alloc))) { + (pp->pr_alloc->pa_flags & PA_WANT) != 0)) { pr_rmpage(pp, ph, pq); } else { LIST_REMOVE(ph, ph_pagelist); @@ -1320,7 +1173,7 @@ pool_do_put(struct pool *pp, void *v, st } /* - * Return resource to the pool. + * Return resource to the pool; must be called at appropriate spl level */ #ifdef POOL_DIAGNOSTIC void @@ -1330,7 +1183,7 @@ _pool_put(struct pool *pp, void *v, cons LIST_INIT(&pq); - mutex_enter(&pp->pr_lock); + simple_lock(&pp->pr_slock); pr_enter(pp, file, line); pr_log(pp, v, PRLOG_PUT, file, line); @@ -1338,7 +1191,7 @@ _pool_put(struct pool *pp, void *v, cons pool_do_put(pp, v, &pq); pr_leave(pp); - mutex_exit(&pp->pr_lock); + simple_unlock(&pp->pr_slock); pr_pagelist_free(pp, &pq); } @@ -1352,9 +1205,9 @@ pool_put(struct pool *pp, void *v) LIST_INIT(&pq); - mutex_enter(&pp->pr_lock); + simple_lock(&pp->pr_slock); pool_do_put(pp, v, &pq); - mutex_exit(&pp->pr_lock); + simple_unlock(&pp->pr_slock); pr_pagelist_free(pp, &pq); } @@ -1364,64 +1217,43 @@ pool_put(struct pool *pp, void *v) #endif /* - * pool_grow: grow a pool by a page. - * - * => called with pool locked. - * => unlock and relock the pool. - * => return with pool locked. - */ - -static int -pool_grow(struct pool *pp, int flags) -{ - struct pool_item_header *ph = NULL; - char *cp; - - mutex_exit(&pp->pr_lock); - cp = pool_allocator_alloc(pp, flags); - if (__predict_true(cp != NULL)) { - ph = pool_alloc_item_header(pp, cp, flags); - } - if (__predict_false(cp == NULL || ph == NULL)) { - if (cp != NULL) { - pool_allocator_free(pp, cp); - } - mutex_enter(&pp->pr_lock); - return ENOMEM; - } - - mutex_enter(&pp->pr_lock); - pool_prime_page(pp, cp, ph); - pp->pr_npagealloc++; - return 0; -} - -/* * Add N items to the pool. */ int pool_prime(struct pool *pp, int n) { + struct pool_item_header *ph = NULL; + caddr_t cp; int newpages; - int error = 0; - mutex_enter(&pp->pr_lock); + simple_lock(&pp->pr_slock); newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage; while (newpages-- > 0) { - error = pool_grow(pp, PR_NOWAIT); - if (error) { + simple_unlock(&pp->pr_slock); + cp = pool_allocator_alloc(pp, PR_NOWAIT); + if (__predict_true(cp != NULL)) + ph = pool_alloc_item_header(pp, cp, PR_NOWAIT); + + if (__predict_false(cp == NULL || ph == NULL)) { + if (cp != NULL) + pool_allocator_free(pp, cp); + simple_lock(&pp->pr_slock); break; } + + simple_lock(&pp->pr_slock); + pool_prime_page(pp, cp, ph); + pp->pr_npagealloc++; pp->pr_minpages++; } if (pp->pr_minpages >= pp->pr_maxpages) pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */ - mutex_exit(&pp->pr_lock); - return error; + simple_unlock(&pp->pr_slock); + return (0); } /* @@ -1430,19 +1262,18 @@ pool_prime(struct pool *pp, int n) * Note, we must be called with the pool descriptor LOCKED. */ static void -pool_prime_page(struct pool *pp, void *storage, struct pool_item_header *ph) +pool_prime_page(struct pool *pp, caddr_t storage, struct pool_item_header *ph) { struct pool_item *pi; - void *cp = storage; - const unsigned int align = pp->pr_align; - const unsigned int ioff = pp->pr_itemoffset; + caddr_t cp = storage; + unsigned int align = pp->pr_align; + unsigned int ioff = pp->pr_itemoffset; int n; - KASSERT(mutex_owned(&pp->pr_lock)); + LOCK_ASSERT(simple_lock_held(&pp->pr_slock)); #ifdef DIAGNOSTIC - if ((pp->pr_roflags & PR_NOALIGN) == 0 && - ((uintptr_t)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0) + if (((u_long)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0) panic("pool_prime_page: %s: unaligned page", pp->pr_wchan); #endif @@ -1462,8 +1293,7 @@ pool_prime_page(struct pool *pp, void *s /* * Color this page. */ - ph->ph_off = pp->pr_curcolor; - cp = (char *)cp + ph->ph_off; + cp = (caddr_t)(cp + pp->pr_curcolor); if ((pp->pr_curcolor += align) > pp->pr_maxcolor) pp->pr_curcolor = 0; @@ -1471,9 +1301,7 @@ pool_prime_page(struct pool *pp, void *s * Adjust storage to apply aligment to `pr_itemoffset' in each item. */ if (ioff != 0) - cp = (char *)cp + align - ioff; - - KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0); + cp = (caddr_t)(cp + (align - ioff)); /* * Insert remaining chunks on the bucket list. @@ -1482,7 +1310,14 @@ pool_prime_page(struct pool *pp, void *s pp->pr_nitems += n; if (pp->pr_roflags & PR_NOTOUCH) { - pr_item_notouch_init(pp, ph); + pool_item_freelist_t *freelist = PR_FREELIST(ph); + int i; + + ph->ph_off = cp - storage; + ph->ph_firstfree = 0; + for (i = 0; i < n - 1; i++) + freelist[i] = i + 1; + freelist[n - 1] = PR_INDEX_EOL; } else { while (n--) { pi = (struct pool_item *)cp; @@ -1494,9 +1329,7 @@ pool_prime_page(struct pool *pp, void *s #ifdef DIAGNOSTIC pi->pi_magic = PI_MAGIC; #endif - cp = (char *)cp + pp->pr_size; - - KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0); + cp = (caddr_t)(cp + pp->pr_size); } } @@ -1522,15 +1355,34 @@ pool_prime_page(struct pool *pp, void *s static int pool_catchup(struct pool *pp) { + struct pool_item_header *ph = NULL; + caddr_t cp; int error = 0; while (POOL_NEEDS_CATCHUP(pp)) { - error = pool_grow(pp, PR_NOWAIT); - if (error) { + /* + * Call the page back-end allocator for more memory. + * + * XXX: We never wait, so should we bother unlocking + * the pool descriptor? + */ + simple_unlock(&pp->pr_slock); + cp = pool_allocator_alloc(pp, PR_NOWAIT); + if (__predict_true(cp != NULL)) + ph = pool_alloc_item_header(pp, cp, PR_NOWAIT); + if (__predict_false(cp == NULL || ph == NULL)) { + if (cp != NULL) + pool_allocator_free(pp, cp); + error = ENOMEM; + simple_lock(&pp->pr_slock); break; } + simple_lock(&pp->pr_slock); + pool_prime_page(pp, cp, ph); + pp->pr_npagealloc++; } - return error; + + return (error); } static void @@ -1547,7 +1399,7 @@ void pool_setlowat(struct pool *pp, int n) { - mutex_enter(&pp->pr_lock); + simple_lock(&pp->pr_slock); pp->pr_minitems = n; pp->pr_minpages = (n == 0) @@ -1563,27 +1415,27 @@ pool_setlowat(struct pool *pp, int n) */ } - mutex_exit(&pp->pr_lock); + simple_unlock(&pp->pr_slock); } void pool_sethiwat(struct pool *pp, int n) { - mutex_enter(&pp->pr_lock); + simple_lock(&pp->pr_slock); pp->pr_maxpages = (n == 0) ? 0 : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage; - mutex_exit(&pp->pr_lock); + simple_unlock(&pp->pr_slock); } void pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap) { - mutex_enter(&pp->pr_lock); + simple_lock(&pp->pr_slock); pp->pr_hardlimit = n; pp->pr_hardlimit_warning = warnmess; @@ -1599,7 +1451,7 @@ pool_sethardlimit(struct pool *pp, int n ? 0 : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage; - mutex_exit(&pp->pr_lock); + simple_unlock(&pp->pr_slock); } /* @@ -1613,10 +1465,10 @@ pool_reclaim(struct pool *pp) #endif { struct pool_item_header *ph, *phnext; + struct pool_cache *pc; struct pool_pagelist pq; + struct pool_cache_grouplist pcgl; struct timeval curtime, diff; - bool klock; - int rv; if (pp->pr_drain_hook != NULL) { /* @@ -1625,31 +1477,18 @@ pool_reclaim(struct pool *pp) (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT); } - /* - * XXXSMP Because mutexes at IPL_SOFTXXX are still spinlocks, - * and we are called from the pagedaemon without kernel_lock. - * Does not apply to IPL_SOFTBIO. - */ - 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 (klock) { - KERNEL_UNLOCK_ONE(NULL); - } + if (simple_lock_try(&pp->pr_slock) == 0) return (0); - } pr_enter(pp, file, line); LIST_INIT(&pq); + LIST_INIT(&pcgl); + + /* + * Reclaim items from the pool's caches. + */ + LIST_FOREACH(pc, &pp->pr_cachelist, pc_poollist) + pool_cache_reclaim(pc, &pq, &pcgl); getmicrotime(&curtime); @@ -1662,8 +1501,7 @@ pool_reclaim(struct pool *pp) KASSERT(ph->ph_nmissing == 0); timersub(&curtime, &ph->ph_time, &diff); - if (diff.tv_sec < pool_inactive_time - && !pa_starved_p(pp->pr_alloc)) + if (diff.tv_sec < pool_inactive_time) continue; /* @@ -1678,87 +1516,39 @@ pool_reclaim(struct pool *pp) } pr_leave(pp); - mutex_exit(&pp->pr_lock); - - if (LIST_EMPTY(&pq)) - rv = 0; - else { - pr_pagelist_free(pp, &pq); - rv = 1; - } - - if (klock) { - KERNEL_UNLOCK_ONE(NULL); - } + simple_unlock(&pp->pr_slock); + if (LIST_EMPTY(&pq) && LIST_EMPTY(&pcgl)) + return 0; - return (rv); + pr_pagelist_free(pp, &pq); + pcg_grouplist_free(&pcgl); + return (1); } /* - * Drain pools, one at a time. This is a two stage process; - * drain_start kicks off a cross call to drain CPU-level caches - * if the pool has an associated pool_cache. drain_end waits - * for those cross calls to finish, and then drains the cache - * (if any) and pool. + * Drain pools, one at a time. * - * Note, must never be called from interrupt context. + * Note, we must never be called from an interrupt context. */ void -pool_drain_start(struct pool **ppp, uint64_t *wp) +pool_drain(void *arg) { struct pool *pp; - - KASSERT(!TAILQ_EMPTY(&pool_head)); + int s; pp = NULL; - - /* Find next pool to drain, and add a reference. */ - mutex_enter(&pool_head_lock); - do { - if (drainpp == NULL) { - drainpp = TAILQ_FIRST(&pool_head); - } - if (drainpp != NULL) { - pp = drainpp; - drainpp = TAILQ_NEXT(pp, pr_poollist); - } - /* - * Skip completely idle pools. We depend on at least - * one pool in the system being active. - */ - } while (pp == NULL || pp->pr_npages == 0); - pp->pr_refcnt++; - mutex_exit(&pool_head_lock); - - /* If there is a pool_cache, drain CPU level caches. */ - *ppp = pp; - if (pp->pr_cache != NULL) { - *wp = xc_broadcast(0, (xcfunc_t)pool_cache_xcall, - pp->pr_cache, NULL); + s = splvm(); + simple_lock(&pool_head_slock); + if (drainpp == NULL) { + drainpp = LIST_FIRST(&pool_head); + } + if (drainpp) { + pp = drainpp; + drainpp = LIST_NEXT(pp, pr_poollist); } -} - -void -pool_drain_end(struct pool *pp, uint64_t where) -{ - - if (pp == NULL) - return; - - KASSERT(pp->pr_refcnt > 0); - - /* Wait for remote draining to complete. */ - if (pp->pr_cache != NULL) - xc_wait(where); - - /* Drain the cache (if any) and pool.. */ + simple_unlock(&pool_head_slock); pool_reclaim(pp); - - /* Finally, unlock the pool. */ - mutex_enter(&pool_head_lock); - pp->pr_refcnt--; - cv_broadcast(&pool_busy); - mutex_exit(&pool_head_lock); + splx(s); } /* @@ -1767,8 +1557,18 @@ pool_drain_end(struct pool *pp, uint64_t void pool_print(struct pool *pp, const char *modif) { + int s; + s = splvm(); + 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 @@ -1776,7 +1576,13 @@ pool_printall(const char *modif, void (* { struct pool *pp; - TAILQ_FOREACH(pp, &pool_head, pr_poollist) { + if (simple_lock_try(&pool_head_slock) == 0) { + (*pr)("WARNING: pool_head_slock is locked\n"); + } else { + simple_unlock(&pool_head_slock); + } + + LIST_FOREACH(pp, &pool_head, pr_poollist) { pool_printit(pp, modif, pr); } } @@ -1790,6 +1596,20 @@ pool_printit(struct pool *pp, const char 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 + simple_unlock(&pp->pr_slock); + pool_print1(pp, modif, pr); } @@ -1824,10 +1644,8 @@ static void pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...)) { struct pool_item_header *ph; - pool_cache_t pc; - pcg_t *pcg; - pool_cache_cpu_t *cc; - uint64_t cpuhit, cpumiss; + struct pool_cache *pc; + struct pool_cache_group *pcg; int i, print_log = 0, print_pagelist = 0, print_cache = 0; char c; @@ -1840,13 +1658,7 @@ pool_print1(struct pool *pp, const char print_cache = 1; } - if ((pc = pp->pr_cache) != NULL) { - (*pr)("POOL CACHE"); - } else { - (*pr)("POOL"); - } - - (*pr)(" %s: size %u, align %u, ioff %u, roflags 0x%08x\n", + (*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n", pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset, pp->pr_roflags); (*pr)("\talloc %p\n", pp->pr_alloc); @@ -1855,7 +1667,7 @@ pool_print1(struct pool *pp, const char (*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n", pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit); - (*pr)("\tnget %lu, nfail %lu, nput %lu\n", + (*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); @@ -1885,15 +1697,16 @@ pool_print1(struct pool *pp, const char (*pr)("\n"); if ((pp->pr_roflags & PR_LOGGING) == 0) (*pr)("\tno log\n"); - else { + else pr_printlog(pp, NULL, pr); - } skip_log: + if (print_cache == 0) + goto skip_cache; #define PR_GROUPLIST(pcg) \ (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail); \ - for (i = 0; i < pcg->pcg_size; i++) { \ + for (i = 0; i < PCG_NOBJECTS; i++) { \ if (pcg->pcg_objects[i].pcgo_pa != \ POOL_PADDR_INVALID) { \ (*pr)("\t\t\t%p, 0x%llx\n", \ @@ -1906,38 +1719,26 @@ pool_print1(struct pool *pp, const char } \ } - if (pc != NULL) { - cpuhit = 0; - cpumiss = 0; - for (i = 0; i < MAXCPUS; i++) { - if ((cc = pc->pc_cpus[i]) == NULL) - continue; - cpuhit += cc->cc_hits; - cpumiss += cc->cc_misses; - } - (*pr)("\tcpu layer hits %llu misses %llu\n", cpuhit, cpumiss); - (*pr)("\tcache layer hits %llu misses %llu\n", - pc->pc_hits, pc->pc_misses); - (*pr)("\tcache layer entry uncontended %llu contended %llu\n", - pc->pc_hits + pc->pc_misses - pc->pc_contended, - pc->pc_contended); - (*pr)("\tcache layer empty groups %u full groups %u\n", - pc->pc_nempty, pc->pc_nfull); - if (print_cache) { - (*pr)("\tfull cache groups:\n"); - for (pcg = pc->pc_fullgroups; pcg != NULL; - pcg = pcg->pcg_next) { - PR_GROUPLIST(pcg); - } - (*pr)("\tempty cache groups:\n"); - for (pcg = pc->pc_emptygroups; pcg != NULL; - pcg = pcg->pcg_next) { - PR_GROUPLIST(pcg); - } + LIST_FOREACH(pc, &pp->pr_cachelist, pc_poollist) { + (*pr)("\tcache %p\n", pc); + (*pr)("\t hits %lu misses %lu ngroups %lu nitems %lu\n", + pc->pc_hits, pc->pc_misses, pc->pc_ngroups, pc->pc_nitems); + (*pr)("\t full groups:\n"); + LIST_FOREACH(pcg, &pc->pc_fullgroups, pcg_list) { + PR_GROUPLIST(pcg); + } + (*pr)("\t partial groups:\n"); + LIST_FOREACH(pcg, &pc->pc_partgroups, pcg_list) { + PR_GROUPLIST(pcg); + } + (*pr)("\t empty groups:\n"); + LIST_FOREACH(pcg, &pc->pc_emptygroups, pcg_list) { + PR_GROUPLIST(pcg); } } #undef PR_GROUPLIST + skip_cache: pr_enter_check(pp, pr); } @@ -1945,21 +1746,19 @@ static int pool_chk_page(struct pool *pp, const char *label, struct pool_item_header *ph) { struct pool_item *pi; - void *page; + caddr_t page; int n; - if ((pp->pr_roflags & PR_NOALIGN) == 0) { - page = (void *)((uintptr_t)ph & pp->pr_alloc->pa_pagemask); - if (page != ph->ph_page && - (pp->pr_roflags & PR_PHINPAGE) != 0) { - if (label != NULL) - printf("%s: ", label); - printf("pool(%p:%s): page inconsistency: page %p;" - " at page head addr %p (p %p)\n", pp, - pp->pr_wchan, ph->ph_page, - ph, page); - return 1; - } + page = (caddr_t)((u_long)ph & pp->pr_alloc->pa_pagemask); + if (page != ph->ph_page && + (pp->pr_roflags & PR_PHINPAGE) != 0) { + if (label != NULL) + printf("%s: ", label); + printf("pool(%p:%s): page inconsistency: page %p;" + " at page head addr %p (p %p)\n", pp, + pp->pr_wchan, ph->ph_page, + ph, page); + return 1; } if ((pp->pr_roflags & PR_NOTOUCH) != 0) @@ -1974,16 +1773,15 @@ pool_chk_page(struct pool *pp, const cha if (label != NULL) printf("%s: ", label); printf("pool(%s): free list modified: magic=%x;" - " page %p; item ordinal %d; addr %p\n", + " page %p; item ordinal %d;" + " addr %p (p %p)\n", pp->pr_wchan, pi->pi_magic, ph->ph_page, - n, pi); + n, pi, page); panic("pool"); } #endif - if ((pp->pr_roflags & PR_NOALIGN) != 0) { - continue; - } - page = (void *)((uintptr_t)pi & pp->pr_alloc->pa_pagemask); + page = + (caddr_t)((u_long)pi & pp->pr_alloc->pa_pagemask); if (page == ph->ph_page) continue; @@ -2005,7 +1803,7 @@ pool_chk(struct pool *pp, const char *la struct pool_item_header *ph; int r = 0; - mutex_enter(&pp->pr_lock); + simple_lock(&pp->pr_slock); LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) { r = pool_chk_page(pp, label, ph); if (r) { @@ -2026,7 +1824,7 @@ pool_chk(struct pool *pp, const char *la } out: - mutex_exit(&pp->pr_lock); + simple_unlock(&pp->pr_slock); return (r); } @@ -2034,112 +1832,38 @@ out: * 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. */ -pool_cache_t -pool_cache_init(size_t size, u_int align, u_int align_offset, u_int flags, - const char *wchan, struct pool_allocator *palloc, int ipl, - int (*ctor)(void *, void *, int), void (*dtor)(void *, void *), void *arg) +void +pool_cache_init(struct pool_cache *pc, struct pool *pp, + int (*ctor)(void *, void *, int), + void (*dtor)(void *, void *), + void *arg) { - pool_cache_t pc; - - pc = pool_get(&cache_pool, PR_WAITOK); - if (pc == NULL) - return NULL; - pool_cache_bootstrap(pc, size, align, align_offset, flags, wchan, - palloc, ipl, ctor, dtor, arg); + LIST_INIT(&pc->pc_emptygroups); + LIST_INIT(&pc->pc_fullgroups); + LIST_INIT(&pc->pc_partgroups); + simple_lock_init(&pc->pc_slock); - return pc; -} + pc->pc_pool = pp; -/* - * pool_cache_bootstrap: - * - * Kernel-private version of pool_cache_init(). The caller - * provides initial storage. - */ -void -pool_cache_bootstrap(pool_cache_t pc, size_t size, u_int align, - u_int align_offset, u_int flags, const char *wchan, - struct pool_allocator *palloc, int ipl, - int (*ctor)(void *, void *, int), void (*dtor)(void *, void *), - void *arg) -{ - CPU_INFO_ITERATOR cii; - pool_cache_t pc1; - struct cpu_info *ci; - struct pool *pp; - - pp = &pc->pc_pool; - if (palloc == NULL && ipl == IPL_NONE) - palloc = &pool_allocator_nointr; - pool_init(pp, size, align, align_offset, flags, wchan, palloc, ipl); - - /* - * XXXAD hack to prevent IP input processing from blocking. - */ - if (ipl == IPL_SOFTNET) { - mutex_init(&pc->pc_lock, MUTEX_DEFAULT, IPL_VM); - } else { - mutex_init(&pc->pc_lock, MUTEX_DEFAULT, ipl); - } - - if (ctor == NULL) { - ctor = (int (*)(void *, void *, int))nullop; - } - if (dtor == NULL) { - dtor = (void (*)(void *, void *))nullop; - } - - pc->pc_emptygroups = NULL; - pc->pc_fullgroups = NULL; - pc->pc_partgroups = NULL; pc->pc_ctor = ctor; pc->pc_dtor = dtor; pc->pc_arg = arg; - pc->pc_hits = 0; - pc->pc_misses = 0; - pc->pc_nempty = 0; - pc->pc_npart = 0; - 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; - } else { - pc->pc_pcgsize = PCG_NOBJECTS_NORMAL; - } + pc->pc_hits = 0; + pc->pc_misses = 0; - /* Allocate per-CPU caches. */ - memset(pc->pc_cpus, 0, sizeof(pc->pc_cpus)); - pc->pc_ncpu = 0; - 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); - } - } + pc->pc_ngroups = 0; - /* Add to list of all pools. */ - if (__predict_true(!cold)) - mutex_enter(&pool_head_lock); - 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); + pc->pc_nitems = 0; - membar_sync(); - pp->pr_cache = pc; + simple_lock(&pp->pr_slock); + LIST_INSERT_HEAD(&pp->pr_cachelist, pc, pc_poollist); + simple_unlock(&pp->pr_slock); } /* @@ -2148,632 +1872,297 @@ pool_cache_bootstrap(pool_cache_t pc, si * Destroy a pool cache. */ void -pool_cache_destroy(pool_cache_t pc) +pool_cache_destroy(struct pool_cache *pc) { - struct pool *pp = &pc->pc_pool; - pool_cache_cpu_t *cc; - pcg_t *pcg; - 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); - TAILQ_REMOVE(&pool_cache_head, pc, pc_cachelist); - mutex_exit(&pool_head_lock); + struct pool *pp = pc->pc_pool; /* First, invalidate the entire cache. */ pool_cache_invalidate(pc); - /* Disassociate it from the pool. */ - mutex_enter(&pp->pr_lock); - pp->pr_cache = NULL; - 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); - } - - /* Finally, destroy it. */ - mutex_destroy(&pc->pc_lock); - pool_destroy(pp); - pool_put(&cache_pool, pc); + /* ...and remove it from the pool's cache list. */ + simple_lock(&pp->pr_slock); + LIST_REMOVE(pc, pc_poollist); + simple_unlock(&pp->pr_slock); } -/* - * pool_cache_cpu_init1: - * - * Called for each pool_cache whenever a new CPU is attached. - */ -static void -pool_cache_cpu_init1(struct cpu_info *ci, pool_cache_t pc) +static inline void * +pcg_get(struct pool_cache_group *pcg, paddr_t *pap) { - pool_cache_cpu_t *cc; - int index; - - index = ci->ci_index; - - KASSERT(index < MAXCPUS); - KASSERT(((uintptr_t)pc->pc_cpus & (CACHE_LINE_SIZE - 1)) == 0); - - if ((cc = pc->pc_cpus[index]) != NULL) { - KASSERT(cc->cc_cpuindex == index); - return; - } - - /* - * The first CPU is 'free'. This needs to be the case for - * bootstrap - we may not be able to allocate yet. - */ - if (pc->pc_ncpu == 0) { - cc = &pc->pc_cpu0; - pc->pc_ncpu = 1; - } else { - mutex_enter(&pc->pc_lock); - pc->pc_ncpu++; - mutex_exit(&pc->pc_lock); - cc = pool_get(&cache_cpu_pool, PR_WAITOK); - } + void *object; + u_int idx; - cc->cc_ipl = pc->pc_pool.pr_ipl; - cc->cc_iplcookie = makeiplcookie(cc->cc_ipl); - cc->cc_cache = pc; - cc->cc_cpuindex = index; - cc->cc_hits = 0; - cc->cc_misses = 0; - cc->cc_current = NULL; - cc->cc_previous = NULL; + KASSERT(pcg->pcg_avail <= PCG_NOBJECTS); + KASSERT(pcg->pcg_avail != 0); + idx = --pcg->pcg_avail; + + KASSERT(pcg->pcg_objects[idx].pcgo_va != NULL); + object = pcg->pcg_objects[idx].pcgo_va; + if (pap != NULL) + *pap = pcg->pcg_objects[idx].pcgo_pa; + pcg->pcg_objects[idx].pcgo_va = NULL; - pc->pc_cpus[index] = cc; + return (object); } -/* - * pool_cache_cpu_init: - * - * Called whenever a new CPU is attached. - */ -void -pool_cache_cpu_init(struct cpu_info *ci) +static inline void +pcg_put(struct pool_cache_group *pcg, void *object, paddr_t pa) { - pool_cache_t pc; + u_int idx; - mutex_enter(&pool_head_lock); - TAILQ_FOREACH(pc, &pool_cache_head, pc_cachelist) { - pc->pc_refcnt++; - mutex_exit(&pool_head_lock); + KASSERT(pcg->pcg_avail < PCG_NOBJECTS); + idx = pcg->pcg_avail++; - pool_cache_cpu_init1(ci, pc); - - mutex_enter(&pool_head_lock); - pc->pc_refcnt--; - cv_broadcast(&pool_busy); - } - mutex_exit(&pool_head_lock); -} - -/* - * pool_cache_reclaim: - * - * Reclaim memory from a pool cache. - */ -bool -pool_cache_reclaim(pool_cache_t pc) -{ - - return pool_reclaim(&pc->pc_pool); + KASSERT(pcg->pcg_objects[idx].pcgo_va == NULL); + pcg->pcg_objects[idx].pcgo_va = object; + pcg->pcg_objects[idx].pcgo_pa = pa; } 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: - * - * Force destruction of an object and its release back into - * the pool. - */ -void -pool_cache_destruct_object(pool_cache_t pc, void *object) +pcg_grouplist_free(struct pool_cache_grouplist *pcgl) { + struct pool_cache_group *pcg; + int s; - FREECHECK_IN(&pc->pc_freecheck, object); - - pool_cache_destruct_object1(pc, object); + s = splvm(); + while ((pcg = LIST_FIRST(pcgl)) != NULL) { + LIST_REMOVE(pcg, pcg_list); + pool_put(&pcgpool, pcg); + } + splx(s); } /* - * pool_cache_invalidate_groups: + * pool_cache_get{,_paddr}: * - * Invalidate a chain of groups and destruct all objects. + * Get an object from a pool cache (optionally returning + * the physical address of the object). */ -static void -pool_cache_invalidate_groups(pool_cache_t pc, pcg_t *pcg) +void * +pool_cache_get_paddr(struct pool_cache *pc, int flags, paddr_t *pap) { + struct pool_cache_group *pcg; void *object; - pcg_t *next; - int i; - for (; pcg != NULL; pcg = next) { - next = pcg->pcg_next; +#ifdef LOCKDEBUG + if (flags & PR_WAITOK) + simple_lock_only_held(NULL, "pool_cache_get(PR_WAITOK)"); +#endif - for (i = 0; i < pcg->pcg_avail; i++) { - object = pcg->pcg_objects[i].pcgo_va; - pool_cache_destruct_object1(pc, object); - } + simple_lock(&pc->pc_slock); - 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); + pcg = LIST_FIRST(&pc->pc_partgroups); + if (pcg == NULL) { + pcg = LIST_FIRST(&pc->pc_fullgroups); + if (pcg != NULL) { + LIST_REMOVE(pcg, pcg_list); + LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list); } } -} - -/* - * pool_cache_invalidate: - * - * Invalidate a pool cache (destruct and release all of the - * cached objects). Does not reclaim objects from the pool. - */ -void -pool_cache_invalidate(pool_cache_t pc) -{ - pcg_t *full, *empty, *part; - - mutex_enter(&pc->pc_lock); - full = pc->pc_fullgroups; - empty = pc->pc_emptygroups; - part = pc->pc_partgroups; - pc->pc_fullgroups = NULL; - pc->pc_emptygroups = NULL; - pc->pc_partgroups = NULL; - pc->pc_nfull = 0; - pc->pc_nempty = 0; - pc->pc_npart = 0; - mutex_exit(&pc->pc_lock); - - pool_cache_invalidate_groups(pc, full); - pool_cache_invalidate_groups(pc, empty); - pool_cache_invalidate_groups(pc, part); -} - -void -pool_cache_set_drain_hook(pool_cache_t pc, void (*fn)(void *, int), void *arg) -{ - - pool_set_drain_hook(&pc->pc_pool, fn, arg); -} - -void -pool_cache_setlowat(pool_cache_t pc, int n) -{ - - pool_setlowat(&pc->pc_pool, n); -} - -void -pool_cache_sethiwat(pool_cache_t pc, int n) -{ - - pool_sethiwat(&pc->pc_pool, n); -} - -void -pool_cache_sethardlimit(pool_cache_t pc, int n, const char *warnmess, int ratecap) -{ + if (pcg == NULL) { - pool_sethardlimit(&pc->pc_pool, n, warnmess, ratecap); -} + /* + * No groups with any available objects. Allocate + * a new object, construct it, and return it to + * the caller. We will allocate a group, if necessary, + * when the object is freed back to the cache. + */ + pc->pc_misses++; + simple_unlock(&pc->pc_slock); + object = pool_get(pc->pc_pool, flags); + if (object != NULL && pc->pc_ctor != NULL) { + if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) { + pool_put(pc->pc_pool, object); + return (NULL); + } + } + if (object != NULL && pap != NULL) { +#ifdef POOL_VTOPHYS + *pap = POOL_VTOPHYS(object); +#else + *pap = POOL_PADDR_INVALID; +#endif + } + return (object); + } -static inline pool_cache_cpu_t * -pool_cache_cpu_enter(pool_cache_t pc, int *s) -{ - pool_cache_cpu_t *cc; + pc->pc_hits++; + pc->pc_nitems--; + object = pcg_get(pcg, pap); - /* - * 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. - */ - crit_enter(); - cc = pc->pc_cpus[curcpu()->ci_index]; - KASSERT(cc->cc_cache == pc); - if (cc->cc_ipl != IPL_NONE) { - *s = splraiseipl(cc->cc_iplcookie); + if (pcg->pcg_avail == 0) { + LIST_REMOVE(pcg, pcg_list); + LIST_INSERT_HEAD(&pc->pc_emptygroups, pcg, pcg_list); } - KASSERT(((uintptr_t)cc & (CACHE_LINE_SIZE - 1)) == 0); + simple_unlock(&pc->pc_slock); - return cc; + return (object); } -static inline void -pool_cache_cpu_exit(pool_cache_cpu_t *cc, int *s) +/* + * pool_cache_put{,_paddr}: + * + * Put an object back to the pool cache (optionally caching the + * physical address of the object). + */ +void +pool_cache_put_paddr(struct pool_cache *pc, void *object, paddr_t pa) { + struct pool_cache_group *pcg; + int s; - /* No longer need exclusive access to the per-CPU data. */ - if (cc->cc_ipl != IPL_NONE) { - splx(*s); + if (__predict_false((pc->pc_pool->pr_flags & PR_WANTED) != 0)) { + goto destruct; } - crit_exit(); -} - -#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, - paddr_t *pap, int flags) -{ - pcg_t *pcg, *cur; - uint64_t ncsw; - pool_cache_t pc; - void *object; - pc = cc->cc_cache; - cc->cc_misses++; - - /* - * Nothing was available locally. Try and grab a group - * from the cache. - */ - if (!mutex_tryenter(&pc->pc_lock)) { - ncsw = curlwp->l_ncsw; - mutex_enter(&pc->pc_lock); - pc->pc_contended++; + simple_lock(&pc->pc_slock); - /* - * If we context switched while locking, then - * our view of the per-CPU data is invalid: - * retry. - */ - if (curlwp->l_ncsw != ncsw) { - mutex_exit(&pc->pc_lock); - pool_cache_cpu_exit(cc, s); - return pool_cache_cpu_enter(pc, s); + pcg = LIST_FIRST(&pc->pc_partgroups); + if (pcg == NULL) { + pcg = LIST_FIRST(&pc->pc_emptygroups); + if (pcg != NULL) { + LIST_REMOVE(pcg, pcg_list); + LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list); } } + if (pcg == NULL) { - if ((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. + * No empty groups to free the object to. Attempt to + * allocate one. */ - if ((cur = cc->cc_current) != NULL) { - KASSERT(cur->pcg_avail == 0); - cur->pcg_next = pc->pc_emptygroups; - pc->pc_emptygroups = cur; - pc->pc_nempty++; - } - 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; - } + simple_unlock(&pc->pc_slock); + s = splvm(); + pcg = pool_get(&pcgpool, PR_NOWAIT); + splx(s); + if (pcg == NULL) { +destruct: - /* - * Nothing available locally or in cache. Take the slow - * path: fetch a new object from the pool and construct - * it. - */ - pc->pc_misses++; - mutex_exit(&pc->pc_lock); - pool_cache_cpu_exit(cc, s); - - object = pool_get(&pc->pc_pool, flags); - *objectp = object; - if (object == NULL) - return NULL; - - if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) { - pool_put(&pc->pc_pool, object); - *objectp = NULL; - return NULL; + /* + * Unable to allocate a cache group; destruct the object + * and free it back to the pool. + */ + pool_cache_destruct_object(pc, object); + return; + } + memset(pcg, 0, sizeof(*pcg)); + simple_lock(&pc->pc_slock); + pc->pc_ngroups++; + LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list); } - KASSERT((((vaddr_t)object + pc->pc_pool.pr_itemoffset) & - (pc->pc_pool.pr_align - 1)) == 0); + pc->pc_nitems++; + pcg_put(pcg, object, pa); - if (pap != NULL) { -#ifdef POOL_VTOPHYS - *pap = POOL_VTOPHYS(object); -#else - *pap = POOL_PADDR_INVALID; -#endif + if (pcg->pcg_avail == PCG_NOBJECTS) { + LIST_REMOVE(pcg, pcg_list); + LIST_INSERT_HEAD(&pc->pc_fullgroups, pcg, pcg_list); } - - FREECHECK_OUT(&pc->pc_freecheck, object); - return NULL; + simple_unlock(&pc->pc_slock); } /* - * pool_cache_get{,_paddr}: + * pool_cache_destruct_object: * - * Get an object from a pool cache (optionally returning - * the physical address of the object). + * Force destruction of an object and its release back into + * the pool. */ -void * -pool_cache_get_paddr(pool_cache_t pc, int flags, paddr_t *pap) +void +pool_cache_destruct_object(struct pool_cache *pc, void *object) { - pool_cache_cpu_t *cc; - pcg_t *pcg; - void *object; - int s; - -#ifdef LOCKDEBUG - if (flags & PR_WAITOK) - ASSERT_SLEEPABLE(NULL, "pool_cache_get(PR_WAITOK)"); -#endif - - cc = pool_cache_cpu_enter(pc, &s); - do { - /* Try and allocate an object from the current group. */ - pcg = cc->cc_current; - if (pcg != NULL && pcg->pcg_avail > 0) { - object = pcg->pcg_objects[--pcg->pcg_avail].pcgo_va; - if (pap != NULL) - *pap = pcg->pcg_objects[pcg->pcg_avail].pcgo_pa; -#if defined(DIAGNOSTIC) - pcg->pcg_objects[pcg->pcg_avail].pcgo_va = NULL; -#endif /* defined(DIAGNOSTIC) */ - KASSERT(pcg->pcg_avail <= pcg->pcg_size); - KASSERT(object != NULL); - cc->cc_hits++; - pool_cache_cpu_exit(cc, &s); - FREECHECK_OUT(&pc->pc_freecheck, object); - return object; - } - - /* - * That failed. If the previous group isn't empty, swap - * it with the current group and allocate from there. - */ - pcg = cc->cc_previous; - if (pcg != NULL && pcg->pcg_avail > 0) { - cc->cc_previous = cc->cc_current; - cc->cc_current = pcg; - continue; - } - - /* - * 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. - * Otherwise, we need to retry. - */ - cc = pool_cache_get_slow(cc, &s, &object, pap, flags); - } while (cc != NULL); - return object; + if (pc->pc_dtor != NULL) + (*pc->pc_dtor)(pc->pc_arg, object); + pool_put(pc->pc_pool, 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 void +pool_do_cache_invalidate_grouplist(struct pool_cache_grouplist *pcgsl, + struct pool_cache *pc, struct pool_pagelist *pq, + struct pool_cache_grouplist *pcgdl) { - pcg_t *pcg, *cur; - uint64_t ncsw; - pool_cache_t pc; - u_int nobj; - - pc = cc->cc_cache; - cc->cc_misses++; - - /* - * No free slots locally. Try to grab an empty, unused - * group from the cache. - */ - if (!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 (curlwp->l_ncsw != ncsw) { - mutex_exit(&pc->pc_lock); - pool_cache_cpu_exit(cc, s); - return pool_cache_cpu_enter(pc, s); - } - } + struct pool_cache_group *pcg, *npcg; + void *object; - if ((pcg = pc->pc_emptygroups) != NULL) { - /* - * If there's a empty group, release our full - * group back to the cache. Install the empty - * group and return. - */ - KASSERT(pcg->pcg_avail == 0); - pc->pc_emptygroups = pcg->pcg_next; - if (cc->cc_previous == NULL) { - cc->cc_previous = pcg; - } else { - if ((cur = cc->cc_current) != NULL) { - KASSERT(cur->pcg_avail == pcg->pcg_size); - cur->pcg_next = pc->pc_fullgroups; - pc->pc_fullgroups = cur; - pc->pc_nfull++; - } - cc->cc_current = pcg; - } - pc->pc_hits++; - pc->pc_nempty--; - mutex_exit(&pc->pc_lock); - return cc; + for (pcg = LIST_FIRST(pcgsl); pcg != NULL; pcg = npcg) { + npcg = LIST_NEXT(pcg, pcg_list); + while (pcg->pcg_avail != 0) { + pc->pc_nitems--; + object = pcg_get(pcg, NULL); + if (pc->pc_dtor != NULL) + (*pc->pc_dtor)(pc->pc_arg, object); + pool_do_put(pc->pc_pool, object, pq); + } + pc->pc_ngroups--; + LIST_REMOVE(pcg, pcg_list); + LIST_INSERT_HEAD(pcgdl, pcg, pcg_list); } +} - /* - * Nothing available locally or in cache. Take the - * slow path and try to allocate a new group that we - * can release to. - */ - pc->pc_misses++; - mutex_exit(&pc->pc_lock); - pool_cache_cpu_exit(cc, s); +static void +pool_do_cache_invalidate(struct pool_cache *pc, struct pool_pagelist *pq, + struct pool_cache_grouplist *pcgl) +{ - /* - * If we can't allocate a new group, just throw the - * object away. - */ - nobj = pc->pc_pcgsize; - if (pool_cache_disable) { - pcg = NULL; - } else if (nobj == PCG_NOBJECTS_LARGE) { - pcg = pool_get(&pcg_large_pool, PR_NOWAIT); - } else { - pcg = pool_get(&pcg_normal_pool, PR_NOWAIT); - } - if (pcg == NULL) { - pool_cache_destruct_object(pc, object); - return NULL; - } - pcg->pcg_avail = 0; - pcg->pcg_size = nobj; + LOCK_ASSERT(simple_lock_held(&pc->pc_slock)); + LOCK_ASSERT(simple_lock_held(&pc->pc_pool->pr_slock)); - /* - * Add the empty group to the cache and try again. - */ - mutex_enter(&pc->pc_lock); - pcg->pcg_next = pc->pc_emptygroups; - pc->pc_emptygroups = pcg; - pc->pc_nempty++; - mutex_exit(&pc->pc_lock); + pool_do_cache_invalidate_grouplist(&pc->pc_fullgroups, pc, pq, pcgl); + pool_do_cache_invalidate_grouplist(&pc->pc_partgroups, pc, pq, pcgl); - return pool_cache_cpu_enter(pc, s); -} + KASSERT(LIST_EMPTY(&pc->pc_partgroups)); + KASSERT(LIST_EMPTY(&pc->pc_fullgroups)); + KASSERT(pc->pc_nitems == 0); +} /* - * pool_cache_put{,_paddr}: + * pool_cache_invalidate: * - * Put an object back to the pool cache (optionally caching the - * physical address of the object). + * Invalidate a pool cache (destruct and release all of the + * cached objects). */ void -pool_cache_put_paddr(pool_cache_t pc, void *object, paddr_t pa) +pool_cache_invalidate(struct pool_cache *pc) { - pool_cache_cpu_t *cc; - pcg_t *pcg; - int s; + struct pool_pagelist pq; + struct pool_cache_grouplist pcgl; - FREECHECK_IN(&pc->pc_freecheck, object); + LIST_INIT(&pq); + LIST_INIT(&pcgl); - cc = pool_cache_cpu_enter(pc, &s); - do { - /* If the current group isn't full, release it there. */ - pcg = cc->cc_current; - if (pcg != NULL && 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); - return; - } + simple_lock(&pc->pc_slock); + simple_lock(&pc->pc_pool->pr_slock); - /* - * That failed. If the previous group is empty, swap - * it with the current group and try again. - */ - pcg = cc->cc_previous; - if (pcg != NULL && pcg->pcg_avail == 0) { - cc->cc_previous = cc->cc_current; - cc->cc_current = pcg; - continue; - } + pool_do_cache_invalidate(pc, &pq, &pcgl); - /* - * 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. - */ - cc = pool_cache_put_slow(cc, &s, object, pa); - } while (cc != NULL); + simple_unlock(&pc->pc_pool->pr_slock); + simple_unlock(&pc->pc_slock); + + pr_pagelist_free(pc->pc_pool, &pq); + pcg_grouplist_free(&pcgl); } /* - * pool_cache_xcall: + * pool_cache_reclaim: * - * Transfer objects from the per-CPU cache to the global cache. - * Run within a cross-call thread. + * Reclaim a pool cache for pool_reclaim(). */ static void -pool_cache_xcall(pool_cache_t pc) +pool_cache_reclaim(struct pool_cache *pc, struct pool_pagelist *pq, + struct pool_cache_grouplist *pcgl) { - pool_cache_cpu_t *cc; - pcg_t *prev, *cur, **list; - int s = 0; /* XXXgcc */ - - cc = pool_cache_cpu_enter(pc, &s); - cur = cc->cc_current; - cc->cc_current = NULL; - prev = cc->cc_previous; - cc->cc_previous = NULL; - pool_cache_cpu_exit(cc, &s); - - /* - * XXXSMP Go to splvm to prevent kernel_lock from being taken, - * because locks at IPL_SOFTXXX are still spinlocks. Does not - * apply to IPL_SOFTBIO. Cross-call threads do not take the - * kernel_lock. + + /* + * We're locking in the wrong order (normally pool_cache -> pool, + * but the pool is already locked when we get here), so we have + * to use trylock. If we can't lock the pool_cache, it's not really + * a big deal here. */ - s = splvm(); - mutex_enter(&pc->pc_lock); - if (cur != NULL) { - if (cur->pcg_avail == cur->pcg_size) { - list = &pc->pc_fullgroups; - pc->pc_nfull++; - } else if (cur->pcg_avail == 0) { - list = &pc->pc_emptygroups; - pc->pc_nempty++; - } else { - list = &pc->pc_partgroups; - pc->pc_npart++; - } - cur->pcg_next = *list; - *list = cur; - } - if (prev != NULL) { - if (prev->pcg_avail == prev->pcg_size) { - list = &pc->pc_fullgroups; - pc->pc_nfull++; - } else if (prev->pcg_avail == 0) { - list = &pc->pc_emptygroups; - pc->pc_nempty++; - } else { - list = &pc->pc_partgroups; - pc->pc_npart++; - } - prev->pcg_next = *list; - *list = prev; - } - mutex_exit(&pc->pc_lock); - splx(s); + if (simple_lock_try(&pc->pc_slock) == 0) + return; + + pool_do_cache_invalidate(pc, pq, pcgl); + + simple_unlock(&pc->pc_slock); } /* @@ -2792,84 +2181,148 @@ pool_cache_xcall(pool_cache_t pc) void *pool_page_alloc(struct pool *, int); void pool_page_free(struct pool *, void *); -#ifdef POOL_SUBPAGE -struct pool_allocator pool_allocator_kmem_fullpage = { - pool_page_alloc, pool_page_free, 0, - .pa_backingmapptr = &kmem_map, -}; -#else struct pool_allocator pool_allocator_kmem = { pool_page_alloc, pool_page_free, 0, - .pa_backingmapptr = &kmem_map, }; -#endif void *pool_page_alloc_nointr(struct pool *, int); void pool_page_free_nointr(struct pool *, void *); -#ifdef POOL_SUBPAGE -struct pool_allocator pool_allocator_nointr_fullpage = { - pool_page_alloc_nointr, pool_page_free_nointr, 0, - .pa_backingmapptr = &kernel_map, -}; -#else struct pool_allocator pool_allocator_nointr = { pool_page_alloc_nointr, pool_page_free_nointr, 0, - .pa_backingmapptr = &kernel_map, }; -#endif #ifdef POOL_SUBPAGE void *pool_subpage_alloc(struct pool *, int); void pool_subpage_free(struct pool *, void *); -struct pool_allocator pool_allocator_kmem = { - pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE, - .pa_backingmapptr = &kmem_map, -}; - -void *pool_subpage_alloc_nointr(struct pool *, int); -void pool_subpage_free_nointr(struct pool *, void *); - -struct pool_allocator pool_allocator_nointr = { - pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE, - .pa_backingmapptr = &kmem_map, +struct pool_allocator pool_allocator_kmem_subpage = { + pool_subpage_alloc, pool_subpage_free, 0, }; #endif /* POOL_SUBPAGE */ -static void * -pool_allocator_alloc(struct pool *pp, int flags) +/* + * We have at least three different resources for the same allocation and + * each resource can be depleted. First, we have the ready elements in the + * pool. Then we have the resource (typically a vm_map) for this allocator. + * Finally, we have physical memory. Waiting for any of these can be + * unnecessary when any other is freed, but the kernel doesn't support + * sleeping on multiple wait channels, so we have to employ another strategy. + * + * The caller sleeps on the pool (so that it can be awakened when an item + * is returned to the pool), but we set PA_WANT on the allocator. When a + * page is returned to the allocator and PA_WANT is set, pool_allocator_free + * will wake up all sleeping pools belonging to this allocator. + * + * XXX Thundering herd. + */ +void * +pool_allocator_alloc(struct pool *org, int flags) { - struct pool_allocator *pa = pp->pr_alloc; + struct pool_allocator *pa = org->pr_alloc; + struct pool *pp, *start; + int s, freed; void *res; - res = (*pa->pa_alloc)(pp, flags); - if (res == NULL && (flags & PR_WAITOK) == 0) { + LOCK_ASSERT(!simple_lock_held(&org->pr_slock)); + + do { + if ((res = (*pa->pa_alloc)(org, flags)) != NULL) + return (res); + if ((flags & PR_WAITOK) == 0) { + /* + * We only run the drain hookhere if PR_NOWAIT. + * In other cases, the hook will be run in + * pool_reclaim(). + */ + if (org->pr_drain_hook != NULL) { + (*org->pr_drain_hook)(org->pr_drain_hook_arg, + flags); + if ((res = (*pa->pa_alloc)(org, flags)) != NULL) + return (res); + } + break; + } + /* - * We only run the drain hook here if PR_NOWAIT. - * In other cases, the hook will be run in - * pool_reclaim(). + * Drain all pools, that use this allocator. + * We do this to reclaim VA space. + * pa_alloc is responsible for waiting for + * physical memory. + * + * XXX We risk looping forever if start if someone + * calls pool_destroy on "start". But there is no + * other way to have potentially sleeping pool_reclaim, + * non-sleeping locks on pool_allocator, and some + * stirring of drained pools in the allocator. + * + * XXX Maybe we should use pool_head_slock for locking + * the allocators? */ - if (pp->pr_drain_hook != NULL) { - (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags); - res = (*pa->pa_alloc)(pp, flags); + freed = 0; + + s = splvm(); + simple_lock(&pa->pa_slock); + pp = start = TAILQ_FIRST(&pa->pa_list); + do { + TAILQ_REMOVE(&pa->pa_list, pp, pr_alloc_list); + TAILQ_INSERT_TAIL(&pa->pa_list, pp, pr_alloc_list); + simple_unlock(&pa->pa_slock); + freed = pool_reclaim(pp); + simple_lock(&pa->pa_slock); + } while ((pp = TAILQ_FIRST(&pa->pa_list)) != start && + freed == 0); + + if (freed == 0) { + /* + * We set PA_WANT here, the caller will most likely + * sleep waiting for pages (if not, this won't hurt + * that much), and there is no way to set this in + * the caller without violating locking order. + */ + pa->pa_flags |= PA_WANT; } - } - return res; + simple_unlock(&pa->pa_slock); + splx(s); + } while (freed); + return (NULL); } -static void +void pool_allocator_free(struct pool *pp, void *v) { struct pool_allocator *pa = pp->pr_alloc; + int s; + + LOCK_ASSERT(!simple_lock_held(&pp->pr_slock)); (*pa->pa_free)(pp, v); + + s = splvm(); + simple_lock(&pa->pa_slock); + if ((pa->pa_flags & PA_WANT) == 0) { + simple_unlock(&pa->pa_slock); + splx(s); + return; + } + + TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) { + simple_lock(&pp->pr_slock); + if ((pp->pr_flags & PR_WANTED) != 0) { + pp->pr_flags &= ~PR_WANTED; + wakeup(pp); + } + simple_unlock(&pp->pr_slock); + } + pa->pa_flags &= ~PA_WANT; + simple_unlock(&pa->pa_slock); + splx(s); } void * pool_page_alloc(struct pool *pp, int flags) { - bool waitok = (flags & PR_WAITOK) ? true : false; + boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE; return ((void *) uvm_km_alloc_poolpage_cache(kmem_map, waitok)); } @@ -2884,7 +2337,7 @@ pool_page_free(struct pool *pp, void *v) static void * pool_page_alloc_meta(struct pool *pp, int flags) { - bool waitok = (flags & PR_WAITOK) ? true : false; + boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE; return ((void *) uvm_km_alloc_poolpage(kmem_map, waitok)); } @@ -2901,34 +2354,42 @@ pool_page_free_meta(struct pool *pp, voi void * pool_subpage_alloc(struct pool *pp, int flags) { - return pool_get(&psppool, flags); + void *v; + int s; + s = splvm(); + v = pool_get(&psppool, flags); + splx(s); + return v; } void pool_subpage_free(struct pool *pp, void *v) { + int s; + s = splvm(); pool_put(&psppool, v); + splx(s); } /* We don't provide a real nointr allocator. Maybe later. */ void * -pool_subpage_alloc_nointr(struct pool *pp, int flags) +pool_page_alloc_nointr(struct pool *pp, int flags) { return (pool_subpage_alloc(pp, flags)); } void -pool_subpage_free_nointr(struct pool *pp, void *v) +pool_page_free_nointr(struct pool *pp, void *v) { pool_subpage_free(pp, v); } -#endif /* POOL_SUBPAGE */ +#else void * pool_page_alloc_nointr(struct pool *pp, int flags) { - bool waitok = (flags & PR_WAITOK) ? true : false; + boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE; return ((void *) uvm_km_alloc_poolpage_cache(kernel_map, waitok)); } @@ -2939,143 +2400,4 @@ 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 < MAXCPUS; 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_byindex(i); - - incpucache = true; - snprintf(cpucachestr, - sizeof(cpucachestr), - "cached by CPU %u", - (u_int)ci->ci_cpuid); - 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) */ +#endif /* POOL_SUBPAGE */