18355f576SJeff Roberson /* 28355f576SJeff Roberson * Copyright (c) 2002, Jeffrey Roberson <jroberson@chesapeake.net> 38355f576SJeff Roberson * All rights reserved. 48355f576SJeff Roberson * 58355f576SJeff Roberson * Redistribution and use in source and binary forms, with or without 68355f576SJeff Roberson * modification, are permitted provided that the following conditions 78355f576SJeff Roberson * are met: 88355f576SJeff Roberson * 1. Redistributions of source code must retain the above copyright 98355f576SJeff Roberson * notice unmodified, this list of conditions, and the following 108355f576SJeff Roberson * disclaimer. 118355f576SJeff Roberson * 2. Redistributions in binary form must reproduce the above copyright 128355f576SJeff Roberson * notice, this list of conditions and the following disclaimer in the 138355f576SJeff Roberson * documentation and/or other materials provided with the distribution. 148355f576SJeff Roberson * 158355f576SJeff Roberson * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 168355f576SJeff Roberson * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 178355f576SJeff Roberson * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 188355f576SJeff Roberson * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 198355f576SJeff Roberson * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 208355f576SJeff Roberson * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 218355f576SJeff Roberson * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 228355f576SJeff Roberson * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 238355f576SJeff Roberson * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 248355f576SJeff Roberson * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 258355f576SJeff Roberson * 268355f576SJeff Roberson * $FreeBSD$ 278355f576SJeff Roberson * 288355f576SJeff Roberson */ 298355f576SJeff Roberson 308355f576SJeff Roberson /* 318355f576SJeff Roberson * uma_core.c Implementation of the Universal Memory allocator 328355f576SJeff Roberson * 338355f576SJeff Roberson * This allocator is intended to replace the multitude of similar object caches 348355f576SJeff Roberson * in the standard FreeBSD kernel. The intent is to be flexible as well as 358355f576SJeff Roberson * effecient. A primary design goal is to return unused memory to the rest of 368355f576SJeff Roberson * the system. This will make the system as a whole more flexible due to the 378355f576SJeff Roberson * ability to move memory to subsystems which most need it instead of leaving 388355f576SJeff Roberson * pools of reserved memory unused. 398355f576SJeff Roberson * 408355f576SJeff Roberson * The basic ideas stem from similar slab/zone based allocators whose algorithms 418355f576SJeff Roberson * are well known. 428355f576SJeff Roberson * 438355f576SJeff Roberson */ 448355f576SJeff Roberson 458355f576SJeff Roberson /* 468355f576SJeff Roberson * TODO: 478355f576SJeff Roberson * - Improve memory usage for large allocations 488355f576SJeff Roberson * - Improve INVARIANTS (0xdeadc0de write out) 498355f576SJeff Roberson * - Investigate cache size adjustments 508355f576SJeff Roberson */ 518355f576SJeff Roberson 528355f576SJeff Roberson /* I should really use ktr.. */ 538355f576SJeff Roberson /* 548355f576SJeff Roberson #define UMA_DEBUG 1 558355f576SJeff Roberson #define UMA_DEBUG_ALLOC 1 568355f576SJeff Roberson #define UMA_DEBUG_ALLOC_1 1 578355f576SJeff Roberson */ 588355f576SJeff Roberson 598355f576SJeff Roberson 608355f576SJeff Roberson #include "opt_param.h" 618355f576SJeff Roberson #include <sys/param.h> 628355f576SJeff Roberson #include <sys/systm.h> 638355f576SJeff Roberson #include <sys/kernel.h> 648355f576SJeff Roberson #include <sys/types.h> 658355f576SJeff Roberson #include <sys/queue.h> 668355f576SJeff Roberson #include <sys/malloc.h> 678355f576SJeff Roberson #include <sys/lock.h> 688355f576SJeff Roberson #include <sys/sysctl.h> 698355f576SJeff Roberson #include <sys/mutex.h> 708355f576SJeff Roberson #include <sys/smp.h> 7186bbae32SJeff Roberson #include <sys/vmmeter.h> 7286bbae32SJeff Roberson 7386bbae32SJeff Roberson #include <machine/types.h> 748355f576SJeff Roberson 758355f576SJeff Roberson #include <vm/vm.h> 768355f576SJeff Roberson #include <vm/vm_object.h> 778355f576SJeff Roberson #include <vm/vm_page.h> 788355f576SJeff Roberson #include <vm/vm_param.h> 798355f576SJeff Roberson #include <vm/vm_map.h> 808355f576SJeff Roberson #include <vm/vm_kern.h> 818355f576SJeff Roberson #include <vm/vm_extern.h> 828355f576SJeff Roberson #include <vm/uma.h> 838355f576SJeff Roberson #include <vm/uma_int.h> 848355f576SJeff Roberson 858355f576SJeff Roberson /* 868355f576SJeff Roberson * This is the zone from which all zones are spawned. The idea is that even 878355f576SJeff Roberson * the zone heads are allocated from the allocator, so we use the bss section 888355f576SJeff Roberson * to bootstrap us. 898355f576SJeff Roberson */ 9086bbae32SJeff Roberson static struct uma_zone masterzone; 9186bbae32SJeff Roberson static uma_zone_t zones = &masterzone; 928355f576SJeff Roberson 938355f576SJeff Roberson /* This is the zone from which all of uma_slab_t's are allocated. */ 948355f576SJeff Roberson static uma_zone_t slabzone; 958355f576SJeff Roberson 968355f576SJeff Roberson /* 978355f576SJeff Roberson * The initial hash tables come out of this zone so they can be allocated 988355f576SJeff Roberson * prior to malloc coming up. 998355f576SJeff Roberson */ 1008355f576SJeff Roberson static uma_zone_t hashzone; 1018355f576SJeff Roberson 1028355f576SJeff Roberson /* 1038355f576SJeff Roberson * Zone that buckets come from. 1048355f576SJeff Roberson */ 1058355f576SJeff Roberson static uma_zone_t bucketzone; 1068355f576SJeff Roberson 10786bbae32SJeff Roberson /* 10886bbae32SJeff Roberson * Are we allowed to allocate buckets? 10986bbae32SJeff Roberson */ 11086bbae32SJeff Roberson static int bucketdisable = 1; 11186bbae32SJeff Roberson 1128355f576SJeff Roberson /* Linked list of all zones in the system */ 1138355f576SJeff Roberson static LIST_HEAD(,uma_zone) uma_zones = LIST_HEAD_INITIALIZER(&uma_zones); 1148355f576SJeff Roberson 1158355f576SJeff Roberson /* This mutex protects the zone list */ 1168355f576SJeff Roberson static struct mtx uma_mtx; 1178355f576SJeff Roberson 1188355f576SJeff Roberson /* Linked list of boot time pages */ 1198355f576SJeff Roberson static LIST_HEAD(,uma_slab) uma_boot_pages = 1208355f576SJeff Roberson LIST_HEAD_INITIALIZER(&uma_boot_pages); 1218355f576SJeff Roberson 1228355f576SJeff Roberson /* Count of free boottime pages */ 1238355f576SJeff Roberson static int uma_boot_free = 0; 1248355f576SJeff Roberson 1258355f576SJeff Roberson /* Is the VM done starting up? */ 1268355f576SJeff Roberson static int booted = 0; 1278355f576SJeff Roberson 1288355f576SJeff Roberson /* This is the handle used to schedule our working set calculator */ 1298355f576SJeff Roberson static struct callout uma_callout; 1308355f576SJeff Roberson 1318355f576SJeff Roberson /* This is mp_maxid + 1, for use while looping over each cpu */ 1328355f576SJeff Roberson static int maxcpu; 1338355f576SJeff Roberson 1348355f576SJeff Roberson /* 1358355f576SJeff Roberson * This structure is passed as the zone ctor arg so that I don't have to create 1368355f576SJeff Roberson * a special allocation function just for zones. 1378355f576SJeff Roberson */ 1388355f576SJeff Roberson struct uma_zctor_args { 1398355f576SJeff Roberson char *name; 1408355f576SJeff Roberson int size; 1418355f576SJeff Roberson uma_ctor ctor; 1428355f576SJeff Roberson uma_dtor dtor; 1438355f576SJeff Roberson uma_init uminit; 1448355f576SJeff Roberson uma_fini fini; 1458355f576SJeff Roberson int align; 1468355f576SJeff Roberson u_int16_t flags; 1478355f576SJeff Roberson }; 1488355f576SJeff Roberson 1498355f576SJeff Roberson /* 1508355f576SJeff Roberson * This is the malloc hash table which is used to find the zone that a 1518355f576SJeff Roberson * malloc allocation came from. It is not currently resizeable. The 1528355f576SJeff Roberson * memory for the actual hash bucket is allocated in kmeminit. 1538355f576SJeff Roberson */ 1548355f576SJeff Roberson struct uma_hash mhash; 1558355f576SJeff Roberson struct uma_hash *mallochash = &mhash; 1568355f576SJeff Roberson 1578355f576SJeff Roberson /* Prototypes.. */ 1588355f576SJeff Roberson 1598355f576SJeff Roberson static void *obj_alloc(uma_zone_t, int, u_int8_t *, int); 1608355f576SJeff Roberson static void *page_alloc(uma_zone_t, int, u_int8_t *, int); 1618355f576SJeff Roberson static void page_free(void *, int, u_int8_t); 1628355f576SJeff Roberson static uma_slab_t slab_zalloc(uma_zone_t, int); 1638355f576SJeff Roberson static void cache_drain(uma_zone_t); 1648355f576SJeff Roberson static void bucket_drain(uma_zone_t, uma_bucket_t); 1658355f576SJeff Roberson static void zone_drain(uma_zone_t); 1668355f576SJeff Roberson static void zone_ctor(void *, int, void *); 1679c2cd7e5SJeff Roberson static void zone_dtor(void *, int, void *); 1688355f576SJeff Roberson static void zero_init(void *, int); 1698355f576SJeff Roberson static void zone_small_init(uma_zone_t zone); 1708355f576SJeff Roberson static void zone_large_init(uma_zone_t zone); 1718355f576SJeff Roberson static void zone_foreach(void (*zfunc)(uma_zone_t)); 1728355f576SJeff Roberson static void zone_timeout(uma_zone_t zone); 1738355f576SJeff Roberson static void hash_expand(struct uma_hash *); 1749c2cd7e5SJeff Roberson static void hash_free(struct uma_hash *hash); 1758355f576SJeff Roberson static void uma_timeout(void *); 1768355f576SJeff Roberson static void uma_startup3(void); 177a553d4b8SJeff Roberson static void *uma_zalloc_internal(uma_zone_t, void *, int, uma_bucket_t); 17886bbae32SJeff Roberson static void uma_zfree_internal(uma_zone_t, void *, void *, int); 17986bbae32SJeff Roberson static void bucket_enable(void); 1808355f576SJeff Roberson void uma_print_zone(uma_zone_t); 1818355f576SJeff Roberson void uma_print_stats(void); 1828355f576SJeff Roberson static int sysctl_vm_zone(SYSCTL_HANDLER_ARGS); 1838355f576SJeff Roberson 1848355f576SJeff Roberson SYSCTL_OID(_vm, OID_AUTO, zone, CTLTYPE_STRING|CTLFLAG_RD, 1858355f576SJeff Roberson NULL, 0, sysctl_vm_zone, "A", "Zone Info"); 1868355f576SJeff Roberson SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 1878355f576SJeff Roberson 18886bbae32SJeff Roberson /* 18986bbae32SJeff Roberson * This routine checks to see whether or not it's safe to enable buckets. 19086bbae32SJeff Roberson */ 19186bbae32SJeff Roberson 19286bbae32SJeff Roberson static void 19386bbae32SJeff Roberson bucket_enable(void) 19486bbae32SJeff Roberson { 19586bbae32SJeff Roberson if (cnt.v_free_count < cnt.v_free_min) 19686bbae32SJeff Roberson bucketdisable = 1; 19786bbae32SJeff Roberson else 19886bbae32SJeff Roberson bucketdisable = 0; 19986bbae32SJeff Roberson } 20086bbae32SJeff Roberson 2018355f576SJeff Roberson 2028355f576SJeff Roberson /* 2038355f576SJeff Roberson * Routine called by timeout which is used to fire off some time interval 2048355f576SJeff Roberson * based calculations. (working set, stats, etc.) 2058355f576SJeff Roberson * 2068355f576SJeff Roberson * Arguments: 2078355f576SJeff Roberson * arg Unused 2088355f576SJeff Roberson * 2098355f576SJeff Roberson * Returns: 2108355f576SJeff Roberson * Nothing 2118355f576SJeff Roberson */ 2128355f576SJeff Roberson static void 2138355f576SJeff Roberson uma_timeout(void *unused) 2148355f576SJeff Roberson { 21586bbae32SJeff Roberson bucket_enable(); 2168355f576SJeff Roberson zone_foreach(zone_timeout); 2178355f576SJeff Roberson 2188355f576SJeff Roberson /* Reschedule this event */ 2198355f576SJeff Roberson callout_reset(&uma_callout, UMA_WORKING_TIME * hz, uma_timeout, NULL); 2208355f576SJeff Roberson } 2218355f576SJeff Roberson 2228355f576SJeff Roberson /* 2238355f576SJeff Roberson * Routine to perform timeout driven calculations. This does the working set 2248355f576SJeff Roberson * as well as hash expanding, and per cpu statistics aggregation. 2258355f576SJeff Roberson * 2268355f576SJeff Roberson * Arguments: 2278355f576SJeff Roberson * zone The zone to operate on 2288355f576SJeff Roberson * 2298355f576SJeff Roberson * Returns: 2308355f576SJeff Roberson * Nothing 2318355f576SJeff Roberson */ 2328355f576SJeff Roberson static void 2338355f576SJeff Roberson zone_timeout(uma_zone_t zone) 2348355f576SJeff Roberson { 2358355f576SJeff Roberson uma_cache_t cache; 2368355f576SJeff Roberson u_int64_t alloc; 2378355f576SJeff Roberson int free; 2388355f576SJeff Roberson int cpu; 2398355f576SJeff Roberson 2408355f576SJeff Roberson alloc = 0; 2418355f576SJeff Roberson free = 0; 2428355f576SJeff Roberson 2438355f576SJeff Roberson /* 2448355f576SJeff Roberson * Aggregate per cpu cache statistics back to the zone. 2458355f576SJeff Roberson * 2468355f576SJeff Roberson * I may rewrite this to set a flag in the per cpu cache instead of 2478355f576SJeff Roberson * locking. If the flag is not cleared on the next round I will have 2488355f576SJeff Roberson * to lock and do it here instead so that the statistics don't get too 2498355f576SJeff Roberson * far out of sync. 2508355f576SJeff Roberson */ 2518355f576SJeff Roberson if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) { 2528355f576SJeff Roberson for (cpu = 0; cpu < maxcpu; cpu++) { 2538355f576SJeff Roberson if (CPU_ABSENT(cpu)) 2548355f576SJeff Roberson continue; 2558355f576SJeff Roberson CPU_LOCK(zone, cpu); 2568355f576SJeff Roberson cache = &zone->uz_cpu[cpu]; 2578355f576SJeff Roberson /* Add them up, and reset */ 2588355f576SJeff Roberson alloc += cache->uc_allocs; 2598355f576SJeff Roberson cache->uc_allocs = 0; 2608355f576SJeff Roberson if (cache->uc_allocbucket) 2618355f576SJeff Roberson free += cache->uc_allocbucket->ub_ptr + 1; 2628355f576SJeff Roberson if (cache->uc_freebucket) 2638355f576SJeff Roberson free += cache->uc_freebucket->ub_ptr + 1; 2648355f576SJeff Roberson CPU_UNLOCK(zone, cpu); 2658355f576SJeff Roberson } 2668355f576SJeff Roberson } 2678355f576SJeff Roberson 2688355f576SJeff Roberson /* Now push these stats back into the zone.. */ 2698355f576SJeff Roberson ZONE_LOCK(zone); 2708355f576SJeff Roberson zone->uz_allocs += alloc; 2718355f576SJeff Roberson 2728355f576SJeff Roberson /* 2738355f576SJeff Roberson * cachefree is an instantanious snapshot of what is in the per cpu 2748355f576SJeff Roberson * caches, not an accurate counter 2758355f576SJeff Roberson */ 2768355f576SJeff Roberson zone->uz_cachefree = free; 2778355f576SJeff Roberson 2788355f576SJeff Roberson /* 2798355f576SJeff Roberson * Expand the zone hash table. 2808355f576SJeff Roberson * 2818355f576SJeff Roberson * This is done if the number of slabs is larger than the hash size. 2828355f576SJeff Roberson * What I'm trying to do here is completely reduce collisions. This 2838355f576SJeff Roberson * may be a little aggressive. Should I allow for two collisions max? 2848355f576SJeff Roberson */ 2858355f576SJeff Roberson 2868355f576SJeff Roberson if ((zone->uz_flags & UMA_ZFLAG_OFFPAGE) && 2878355f576SJeff Roberson !(zone->uz_flags & UMA_ZFLAG_MALLOC)) { 2888355f576SJeff Roberson if (zone->uz_pages / zone->uz_ppera 2898355f576SJeff Roberson >= zone->uz_hash.uh_hashsize) 2908355f576SJeff Roberson hash_expand(&zone->uz_hash); 2918355f576SJeff Roberson } 2928355f576SJeff Roberson 2938355f576SJeff Roberson /* 2948355f576SJeff Roberson * Here we compute the working set size as the total number of items 2958355f576SJeff Roberson * left outstanding since the last time interval. This is slightly 2968355f576SJeff Roberson * suboptimal. What we really want is the highest number of outstanding 2978355f576SJeff Roberson * items during the last time quantum. This should be close enough. 2988355f576SJeff Roberson * 2998355f576SJeff Roberson * The working set size is used to throttle the zone_drain function. 3008355f576SJeff Roberson * We don't want to return memory that we may need again immediately. 3018355f576SJeff Roberson */ 3028355f576SJeff Roberson alloc = zone->uz_allocs - zone->uz_oallocs; 3038355f576SJeff Roberson zone->uz_oallocs = zone->uz_allocs; 3048355f576SJeff Roberson zone->uz_wssize = alloc; 3058355f576SJeff Roberson 3068355f576SJeff Roberson ZONE_UNLOCK(zone); 3078355f576SJeff Roberson } 3088355f576SJeff Roberson 3098355f576SJeff Roberson /* 3108355f576SJeff Roberson * Expands the hash table for OFFPAGE zones. This is done from zone_timeout 3118355f576SJeff Roberson * to reduce collisions. This must not be done in the regular allocation path, 3128355f576SJeff Roberson * otherwise, we can recurse on the vm while allocating pages. 3138355f576SJeff Roberson * 3148355f576SJeff Roberson * Arguments: 3158355f576SJeff Roberson * hash The hash you want to expand by a factor of two. 3168355f576SJeff Roberson * 3178355f576SJeff Roberson * Returns: 3188355f576SJeff Roberson * Nothing 3198355f576SJeff Roberson * 3208355f576SJeff Roberson * Discussion: 3218355f576SJeff Roberson */ 3228355f576SJeff Roberson static void 3238355f576SJeff Roberson hash_expand(struct uma_hash *hash) 3248355f576SJeff Roberson { 3258355f576SJeff Roberson struct slabhead *newhash; 3268355f576SJeff Roberson struct slabhead *oldhash; 3278355f576SJeff Roberson uma_slab_t slab; 3289c2cd7e5SJeff Roberson int oldsize; 3299c2cd7e5SJeff Roberson int newsize; 3308355f576SJeff Roberson int alloc; 3318355f576SJeff Roberson int hval; 3328355f576SJeff Roberson int i; 3338355f576SJeff Roberson 3348355f576SJeff Roberson 3358355f576SJeff Roberson /* 3368355f576SJeff Roberson * Remember the old hash size and see if it has to go back to the 3378355f576SJeff Roberson * hash zone, or malloc. The hash zone is used for the initial hash 3388355f576SJeff Roberson */ 3398355f576SJeff Roberson 3409c2cd7e5SJeff Roberson oldsize = hash->uh_hashsize; 3418355f576SJeff Roberson oldhash = hash->uh_slab_hash; 3428355f576SJeff Roberson 3438355f576SJeff Roberson /* We're just going to go to a power of two greater */ 3448355f576SJeff Roberson if (hash->uh_hashsize) { 3459c2cd7e5SJeff Roberson newsize = oldsize * 2; 3469c2cd7e5SJeff Roberson alloc = sizeof(hash->uh_slab_hash[0]) * newsize; 3478355f576SJeff Roberson /* XXX Shouldn't be abusing DEVBUF here */ 3488355f576SJeff Roberson newhash = (struct slabhead *)malloc(alloc, M_DEVBUF, M_NOWAIT); 3498355f576SJeff Roberson if (newhash == NULL) { 3508355f576SJeff Roberson return; 3518355f576SJeff Roberson } 3528355f576SJeff Roberson } else { 3538355f576SJeff Roberson alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 354a553d4b8SJeff Roberson newhash = uma_zalloc_internal(hashzone, NULL, M_WAITOK, NULL); 3559c2cd7e5SJeff Roberson newsize = UMA_HASH_SIZE_INIT; 3568355f576SJeff Roberson } 3578355f576SJeff Roberson 3588355f576SJeff Roberson bzero(newhash, alloc); 3598355f576SJeff Roberson 3609c2cd7e5SJeff Roberson hash->uh_hashmask = newsize - 1; 3618355f576SJeff Roberson 3628355f576SJeff Roberson /* 3638355f576SJeff Roberson * I need to investigate hash algorithms for resizing without a 3648355f576SJeff Roberson * full rehash. 3658355f576SJeff Roberson */ 3668355f576SJeff Roberson 3679c2cd7e5SJeff Roberson for (i = 0; i < oldsize; i++) 3688355f576SJeff Roberson while (!SLIST_EMPTY(&hash->uh_slab_hash[i])) { 3698355f576SJeff Roberson slab = SLIST_FIRST(&hash->uh_slab_hash[i]); 3708355f576SJeff Roberson SLIST_REMOVE_HEAD(&hash->uh_slab_hash[i], us_hlink); 3718355f576SJeff Roberson hval = UMA_HASH(hash, slab->us_data); 3728355f576SJeff Roberson SLIST_INSERT_HEAD(&newhash[hval], slab, us_hlink); 3738355f576SJeff Roberson } 3748355f576SJeff Roberson 3759c2cd7e5SJeff Roberson if (oldhash) 3769c2cd7e5SJeff Roberson hash_free(hash); 3779c2cd7e5SJeff Roberson 3789c2cd7e5SJeff Roberson hash->uh_slab_hash = newhash; 3799c2cd7e5SJeff Roberson hash->uh_hashsize = newsize; 3809c2cd7e5SJeff Roberson 3819c2cd7e5SJeff Roberson return; 3829c2cd7e5SJeff Roberson } 3839c2cd7e5SJeff Roberson 3849c2cd7e5SJeff Roberson static void 3859c2cd7e5SJeff Roberson hash_free(struct uma_hash *hash) 3869c2cd7e5SJeff Roberson { 3879c2cd7e5SJeff Roberson if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 3888355f576SJeff Roberson uma_zfree_internal(hashzone, 3898355f576SJeff Roberson hash->uh_slab_hash, NULL, 0); 3908355f576SJeff Roberson else 3918355f576SJeff Roberson free(hash->uh_slab_hash, M_DEVBUF); 3928355f576SJeff Roberson 3939c2cd7e5SJeff Roberson hash->uh_slab_hash = NULL; 3948355f576SJeff Roberson } 3958355f576SJeff Roberson 3968355f576SJeff Roberson /* 3978355f576SJeff Roberson * Frees all outstanding items in a bucket 3988355f576SJeff Roberson * 3998355f576SJeff Roberson * Arguments: 4008355f576SJeff Roberson * zone The zone to free to, must be unlocked. 4018355f576SJeff Roberson * bucket The free/alloc bucket with items, cpu queue must be locked. 4028355f576SJeff Roberson * 4038355f576SJeff Roberson * Returns: 4048355f576SJeff Roberson * Nothing 4058355f576SJeff Roberson */ 4068355f576SJeff Roberson 4078355f576SJeff Roberson static void 4088355f576SJeff Roberson bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 4098355f576SJeff Roberson { 4108355f576SJeff Roberson uma_slab_t slab; 4118355f576SJeff Roberson int mzone; 4128355f576SJeff Roberson void *item; 4138355f576SJeff Roberson 4148355f576SJeff Roberson if (bucket == NULL) 4158355f576SJeff Roberson return; 4168355f576SJeff Roberson 4178355f576SJeff Roberson slab = NULL; 4188355f576SJeff Roberson mzone = 0; 4198355f576SJeff Roberson 4208355f576SJeff Roberson /* We have to lookup the slab again for malloc.. */ 4218355f576SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_MALLOC) 4228355f576SJeff Roberson mzone = 1; 4238355f576SJeff Roberson 4248355f576SJeff Roberson while (bucket->ub_ptr > -1) { 4258355f576SJeff Roberson item = bucket->ub_bucket[bucket->ub_ptr]; 4268355f576SJeff Roberson #ifdef INVARIANTS 4278355f576SJeff Roberson bucket->ub_bucket[bucket->ub_ptr] = NULL; 4288355f576SJeff Roberson KASSERT(item != NULL, 4298355f576SJeff Roberson ("bucket_drain: botched ptr, item is NULL")); 4308355f576SJeff Roberson #endif 4318355f576SJeff Roberson bucket->ub_ptr--; 4328355f576SJeff Roberson /* 4338355f576SJeff Roberson * This is extremely inefficient. The slab pointer was passed 4348355f576SJeff Roberson * to uma_zfree_arg, but we lost it because the buckets don't 4358355f576SJeff Roberson * hold them. This will go away when free() gets a size passed 4368355f576SJeff Roberson * to it. 4378355f576SJeff Roberson */ 4388355f576SJeff Roberson if (mzone) 4398355f576SJeff Roberson slab = hash_sfind(mallochash, 4408355f576SJeff Roberson (u_int8_t *)((unsigned long)item & 4418355f576SJeff Roberson (~UMA_SLAB_MASK))); 4428355f576SJeff Roberson uma_zfree_internal(zone, item, slab, 1); 4438355f576SJeff Roberson } 4448355f576SJeff Roberson } 4458355f576SJeff Roberson 4468355f576SJeff Roberson /* 4478355f576SJeff Roberson * Drains the per cpu caches for a zone. 4488355f576SJeff Roberson * 4498355f576SJeff Roberson * Arguments: 4508355f576SJeff Roberson * zone The zone to drain, must be unlocked. 4518355f576SJeff Roberson * 4528355f576SJeff Roberson * Returns: 4538355f576SJeff Roberson * Nothing 4548355f576SJeff Roberson * 4558355f576SJeff Roberson * This function returns with the zone locked so that the per cpu queues can 4568355f576SJeff Roberson * not be filled until zone_drain is finished. 4578355f576SJeff Roberson * 4588355f576SJeff Roberson */ 4598355f576SJeff Roberson static void 4608355f576SJeff Roberson cache_drain(uma_zone_t zone) 4618355f576SJeff Roberson { 4628355f576SJeff Roberson uma_bucket_t bucket; 4638355f576SJeff Roberson uma_cache_t cache; 4648355f576SJeff Roberson int cpu; 4658355f576SJeff Roberson 4668355f576SJeff Roberson /* 4678355f576SJeff Roberson * Flush out the per cpu queues. 4688355f576SJeff Roberson * 469157d7b35SAlfred Perlstein * XXX This causes unnecessary thrashing due to immediately having 4708355f576SJeff Roberson * empty per cpu queues. I need to improve this. 4718355f576SJeff Roberson */ 4728355f576SJeff Roberson 4738355f576SJeff Roberson /* 4748355f576SJeff Roberson * We have to lock each cpu cache before locking the zone 4758355f576SJeff Roberson */ 4768355f576SJeff Roberson ZONE_UNLOCK(zone); 4778355f576SJeff Roberson 4788355f576SJeff Roberson for (cpu = 0; cpu < maxcpu; cpu++) { 4798355f576SJeff Roberson if (CPU_ABSENT(cpu)) 4808355f576SJeff Roberson continue; 4818355f576SJeff Roberson CPU_LOCK(zone, cpu); 4828355f576SJeff Roberson cache = &zone->uz_cpu[cpu]; 4838355f576SJeff Roberson bucket_drain(zone, cache->uc_allocbucket); 4848355f576SJeff Roberson bucket_drain(zone, cache->uc_freebucket); 4858355f576SJeff Roberson } 4868355f576SJeff Roberson 4878355f576SJeff Roberson /* 4888355f576SJeff Roberson * Drain the bucket queues and free the buckets, we just keep two per 4898355f576SJeff Roberson * cpu (alloc/free). 4908355f576SJeff Roberson */ 4918355f576SJeff Roberson ZONE_LOCK(zone); 4928355f576SJeff Roberson while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) { 4938355f576SJeff Roberson LIST_REMOVE(bucket, ub_link); 4948355f576SJeff Roberson ZONE_UNLOCK(zone); 4958355f576SJeff Roberson bucket_drain(zone, bucket); 4968355f576SJeff Roberson uma_zfree_internal(bucketzone, bucket, NULL, 0); 4978355f576SJeff Roberson ZONE_LOCK(zone); 4988355f576SJeff Roberson } 4998355f576SJeff Roberson 5008355f576SJeff Roberson /* Now we do the free queue.. */ 5018355f576SJeff Roberson while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) { 5028355f576SJeff Roberson LIST_REMOVE(bucket, ub_link); 5038355f576SJeff Roberson uma_zfree_internal(bucketzone, bucket, NULL, 0); 5048355f576SJeff Roberson } 5058355f576SJeff Roberson 5068355f576SJeff Roberson /* We unlock here, but they will all block until the zone is unlocked */ 5078355f576SJeff Roberson for (cpu = 0; cpu < maxcpu; cpu++) { 5088355f576SJeff Roberson if (CPU_ABSENT(cpu)) 5098355f576SJeff Roberson continue; 5108355f576SJeff Roberson CPU_UNLOCK(zone, cpu); 5118355f576SJeff Roberson } 512f4af24d5SJeff Roberson 513f4af24d5SJeff Roberson zone->uz_cachefree = 0; 5148355f576SJeff Roberson } 5158355f576SJeff Roberson 5168355f576SJeff Roberson /* 5178355f576SJeff Roberson * Frees pages from a zone back to the system. This is done on demand from 5188355f576SJeff Roberson * the pageout daemon. 5198355f576SJeff Roberson * 5208355f576SJeff Roberson * Arguments: 5218355f576SJeff Roberson * zone The zone to free pages from 5229c2cd7e5SJeff Roberson * all Should we drain all items? 5238355f576SJeff Roberson * 5248355f576SJeff Roberson * Returns: 5258355f576SJeff Roberson * Nothing. 5268355f576SJeff Roberson */ 5278355f576SJeff Roberson static void 5288355f576SJeff Roberson zone_drain(uma_zone_t zone) 5298355f576SJeff Roberson { 5308355f576SJeff Roberson uma_slab_t slab; 5318355f576SJeff Roberson uma_slab_t n; 5328355f576SJeff Roberson u_int64_t extra; 5338355f576SJeff Roberson u_int8_t flags; 5348355f576SJeff Roberson u_int8_t *mem; 5358355f576SJeff Roberson int i; 5368355f576SJeff Roberson 5378355f576SJeff Roberson /* 5388355f576SJeff Roberson * We don't want to take pages from staticly allocated zones at this 5398355f576SJeff Roberson * time 5408355f576SJeff Roberson */ 5418355f576SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_NOFREE || zone->uz_freef == NULL) 5428355f576SJeff Roberson return; 5438355f576SJeff Roberson 5448355f576SJeff Roberson ZONE_LOCK(zone); 5458355f576SJeff Roberson 5468355f576SJeff Roberson if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 5478355f576SJeff Roberson cache_drain(zone); 5488355f576SJeff Roberson 5498355f576SJeff Roberson if (zone->uz_free < zone->uz_wssize) 5508355f576SJeff Roberson goto finished; 5518355f576SJeff Roberson #ifdef UMA_DEBUG 5528355f576SJeff Roberson printf("%s working set size: %llu free items: %u\n", 5538355f576SJeff Roberson zone->uz_name, (unsigned long long)zone->uz_wssize, zone->uz_free); 5548355f576SJeff Roberson #endif 5559c2cd7e5SJeff Roberson extra = zone->uz_free - zone->uz_wssize; 5568355f576SJeff Roberson extra /= zone->uz_ipers; 5578355f576SJeff Roberson 5588355f576SJeff Roberson /* extra is now the number of extra slabs that we can free */ 5598355f576SJeff Roberson 5608355f576SJeff Roberson if (extra == 0) 5618355f576SJeff Roberson goto finished; 5628355f576SJeff Roberson 5638355f576SJeff Roberson slab = LIST_FIRST(&zone->uz_free_slab); 5648355f576SJeff Roberson while (slab && extra) { 5658355f576SJeff Roberson n = LIST_NEXT(slab, us_link); 5668355f576SJeff Roberson 5678355f576SJeff Roberson /* We have no where to free these to */ 5688355f576SJeff Roberson if (slab->us_flags & UMA_SLAB_BOOT) { 5698355f576SJeff Roberson slab = n; 5708355f576SJeff Roberson continue; 5718355f576SJeff Roberson } 5728355f576SJeff Roberson 5738355f576SJeff Roberson LIST_REMOVE(slab, us_link); 5748355f576SJeff Roberson zone->uz_pages -= zone->uz_ppera; 5758355f576SJeff Roberson zone->uz_free -= zone->uz_ipers; 5768355f576SJeff Roberson if (zone->uz_fini) 5778355f576SJeff Roberson for (i = 0; i < zone->uz_ipers; i++) 5788355f576SJeff Roberson zone->uz_fini( 5798355f576SJeff Roberson slab->us_data + (zone->uz_rsize * i), 5808355f576SJeff Roberson zone->uz_size); 5818355f576SJeff Roberson flags = slab->us_flags; 5828355f576SJeff Roberson mem = slab->us_data; 5838355f576SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_OFFPAGE) { 5848355f576SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_MALLOC) { 5858355f576SJeff Roberson UMA_HASH_REMOVE(mallochash, 5868355f576SJeff Roberson slab, slab->us_data); 5878355f576SJeff Roberson } else { 5888355f576SJeff Roberson UMA_HASH_REMOVE(&zone->uz_hash, 5898355f576SJeff Roberson slab, slab->us_data); 5908355f576SJeff Roberson } 5918355f576SJeff Roberson uma_zfree_internal(slabzone, slab, NULL, 0); 5928355f576SJeff Roberson } else if (zone->uz_flags & UMA_ZFLAG_MALLOC) 5938355f576SJeff Roberson UMA_HASH_REMOVE(mallochash, slab, slab->us_data); 5948355f576SJeff Roberson #ifdef UMA_DEBUG 5958355f576SJeff Roberson printf("%s: Returning %d bytes.\n", 5968355f576SJeff Roberson zone->uz_name, UMA_SLAB_SIZE * zone->uz_ppera); 5978355f576SJeff Roberson #endif 5988355f576SJeff Roberson zone->uz_freef(mem, UMA_SLAB_SIZE * zone->uz_ppera, flags); 5998355f576SJeff Roberson 6008355f576SJeff Roberson slab = n; 6018355f576SJeff Roberson extra--; 6028355f576SJeff Roberson } 6038355f576SJeff Roberson 6048355f576SJeff Roberson finished: 6058355f576SJeff Roberson ZONE_UNLOCK(zone); 6068355f576SJeff Roberson } 6078355f576SJeff Roberson 6088355f576SJeff Roberson /* 6098355f576SJeff Roberson * Allocate a new slab for a zone. This does not insert the slab onto a list. 6108355f576SJeff Roberson * 6118355f576SJeff Roberson * Arguments: 6128355f576SJeff Roberson * zone The zone to allocate slabs for 6138355f576SJeff Roberson * wait Shall we wait? 6148355f576SJeff Roberson * 6158355f576SJeff Roberson * Returns: 6168355f576SJeff Roberson * The slab that was allocated or NULL if there is no memory and the 6178355f576SJeff Roberson * caller specified M_NOWAIT. 6188355f576SJeff Roberson * 6198355f576SJeff Roberson */ 6208355f576SJeff Roberson static uma_slab_t 6218355f576SJeff Roberson slab_zalloc(uma_zone_t zone, int wait) 6228355f576SJeff Roberson { 6238355f576SJeff Roberson uma_slab_t slab; /* Starting slab */ 6248355f576SJeff Roberson u_int8_t *mem; 6258355f576SJeff Roberson u_int8_t flags; 6268355f576SJeff Roberson int i; 6278355f576SJeff Roberson 628a553d4b8SJeff Roberson slab = NULL; 629a553d4b8SJeff Roberson 6308355f576SJeff Roberson #ifdef UMA_DEBUG 6318355f576SJeff Roberson printf("slab_zalloc: Allocating a new slab for %s\n", zone->uz_name); 6328355f576SJeff Roberson #endif 6338355f576SJeff Roberson ZONE_UNLOCK(zone); 634a553d4b8SJeff Roberson 635a553d4b8SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_OFFPAGE) { 636a553d4b8SJeff Roberson slab = uma_zalloc_internal(slabzone, NULL, wait, NULL); 637a553d4b8SJeff Roberson if (slab == NULL) { 638a553d4b8SJeff Roberson ZONE_LOCK(zone); 639a553d4b8SJeff Roberson return NULL; 640a553d4b8SJeff Roberson } 641a553d4b8SJeff Roberson } 642a553d4b8SJeff Roberson 643a553d4b8SJeff Roberson if (booted || (zone->uz_flags & UMA_ZFLAG_PRIVALLOC)) { 6448355f576SJeff Roberson mtx_lock(&Giant); 645a553d4b8SJeff Roberson mem = zone->uz_allocf(zone, 6468355f576SJeff Roberson zone->uz_ppera * UMA_SLAB_SIZE, &flags, wait); 6478355f576SJeff Roberson mtx_unlock(&Giant); 648a553d4b8SJeff Roberson if (mem == NULL) { 6498355f576SJeff Roberson ZONE_LOCK(zone); 6508355f576SJeff Roberson return (NULL); 651a553d4b8SJeff Roberson } 6528355f576SJeff Roberson } else { 653a553d4b8SJeff Roberson uma_slab_t tmps; 6548355f576SJeff Roberson 6558355f576SJeff Roberson if (zone->uz_ppera > 1) 6568355f576SJeff Roberson panic("UMA: Attemping to allocate multiple pages before vm has started.\n"); 6578355f576SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_MALLOC) 6588355f576SJeff Roberson panic("Mallocing before uma_startup2 has been called.\n"); 6598355f576SJeff Roberson if (uma_boot_free == 0) 6608355f576SJeff Roberson panic("UMA: Ran out of pre init pages, increase UMA_BOOT_PAGES\n"); 661a553d4b8SJeff Roberson tmps = LIST_FIRST(&uma_boot_pages); 662a553d4b8SJeff Roberson LIST_REMOVE(tmps, us_link); 6638355f576SJeff Roberson uma_boot_free--; 664a553d4b8SJeff Roberson mem = tmps->us_data; 6658355f576SJeff Roberson } 6668355f576SJeff Roberson 667a553d4b8SJeff Roberson ZONE_LOCK(zone); 6688355f576SJeff Roberson 6698355f576SJeff Roberson /* Alloc slab structure for offpage, otherwise adjust it's position */ 6708355f576SJeff Roberson if (!(zone->uz_flags & UMA_ZFLAG_OFFPAGE)) { 6718355f576SJeff Roberson slab = (uma_slab_t )(mem + zone->uz_pgoff); 6728355f576SJeff Roberson } else { 6738355f576SJeff Roberson if (!(zone->uz_flags & UMA_ZFLAG_MALLOC)) 6748355f576SJeff Roberson UMA_HASH_INSERT(&zone->uz_hash, slab, mem); 6758355f576SJeff Roberson } 6768355f576SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_MALLOC) { 6778355f576SJeff Roberson #ifdef UMA_DEBUG 6788355f576SJeff Roberson printf("Inserting %p into malloc hash from slab %p\n", 6798355f576SJeff Roberson mem, slab); 6808355f576SJeff Roberson #endif 681a553d4b8SJeff Roberson /* XXX Yikes! No lock on the malloc hash! */ 6828355f576SJeff Roberson UMA_HASH_INSERT(mallochash, slab, mem); 6838355f576SJeff Roberson } 6848355f576SJeff Roberson 6858355f576SJeff Roberson slab->us_zone = zone; 6868355f576SJeff Roberson slab->us_data = mem; 6878355f576SJeff Roberson 6888355f576SJeff Roberson /* 6898355f576SJeff Roberson * This is intended to spread data out across cache lines. 6908355f576SJeff Roberson * 6918355f576SJeff Roberson * This code doesn't seem to work properly on x86, and on alpha 6928355f576SJeff Roberson * it makes absolutely no performance difference. I'm sure it could 6938355f576SJeff Roberson * use some tuning, but sun makes outrageous claims about it's 6948355f576SJeff Roberson * performance. 6958355f576SJeff Roberson */ 6968355f576SJeff Roberson #if 0 6978355f576SJeff Roberson if (zone->uz_cachemax) { 6988355f576SJeff Roberson slab->us_data += zone->uz_cacheoff; 6998355f576SJeff Roberson zone->uz_cacheoff += UMA_CACHE_INC; 7008355f576SJeff Roberson if (zone->uz_cacheoff > zone->uz_cachemax) 7018355f576SJeff Roberson zone->uz_cacheoff = 0; 7028355f576SJeff Roberson } 7038355f576SJeff Roberson #endif 7048355f576SJeff Roberson 7058355f576SJeff Roberson slab->us_freecount = zone->uz_ipers; 7068355f576SJeff Roberson slab->us_firstfree = 0; 7078355f576SJeff Roberson slab->us_flags = flags; 7088355f576SJeff Roberson for (i = 0; i < zone->uz_ipers; i++) 7098355f576SJeff Roberson slab->us_freelist[i] = i+1; 7108355f576SJeff Roberson 7118355f576SJeff Roberson if (zone->uz_init) 7128355f576SJeff Roberson for (i = 0; i < zone->uz_ipers; i++) 7138355f576SJeff Roberson zone->uz_init(slab->us_data + (zone->uz_rsize * i), 7148355f576SJeff Roberson zone->uz_size); 7158355f576SJeff Roberson 7168355f576SJeff Roberson zone->uz_pages += zone->uz_ppera; 7178355f576SJeff Roberson zone->uz_free += zone->uz_ipers; 7188355f576SJeff Roberson 7198355f576SJeff Roberson return (slab); 7208355f576SJeff Roberson } 7218355f576SJeff Roberson 7228355f576SJeff Roberson /* 7238355f576SJeff Roberson * Allocates a number of pages from the system 7248355f576SJeff Roberson * 7258355f576SJeff Roberson * Arguments: 7268355f576SJeff Roberson * zone Unused 7278355f576SJeff Roberson * bytes The number of bytes requested 7288355f576SJeff Roberson * wait Shall we wait? 7298355f576SJeff Roberson * 7308355f576SJeff Roberson * Returns: 7318355f576SJeff Roberson * A pointer to the alloced memory or possibly 7328355f576SJeff Roberson * NULL if M_NOWAIT is set. 7338355f576SJeff Roberson */ 7348355f576SJeff Roberson static void * 7358355f576SJeff Roberson page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait) 7368355f576SJeff Roberson { 7378355f576SJeff Roberson void *p; /* Returned page */ 7388355f576SJeff Roberson 7398355f576SJeff Roberson /* 7408355f576SJeff Roberson * XXX The original zone allocator did this, but I don't think it's 741157d7b35SAlfred Perlstein * necessary in current. 7428355f576SJeff Roberson */ 7438355f576SJeff Roberson 7448355f576SJeff Roberson if (lockstatus(&kernel_map->lock, NULL)) { 7458355f576SJeff Roberson *pflag = UMA_SLAB_KMEM; 7468355f576SJeff Roberson p = (void *) kmem_malloc(kmem_map, bytes, wait); 7478355f576SJeff Roberson } else { 7488355f576SJeff Roberson *pflag = UMA_SLAB_KMAP; 7498355f576SJeff Roberson p = (void *) kmem_alloc(kernel_map, bytes); 7508355f576SJeff Roberson } 7518355f576SJeff Roberson 7528355f576SJeff Roberson return (p); 7538355f576SJeff Roberson } 7548355f576SJeff Roberson 7558355f576SJeff Roberson /* 7568355f576SJeff Roberson * Allocates a number of pages from within an object 7578355f576SJeff Roberson * 7588355f576SJeff Roberson * Arguments: 7598355f576SJeff Roberson * zone Unused 7608355f576SJeff Roberson * bytes The number of bytes requested 7618355f576SJeff Roberson * wait Shall we wait? 7628355f576SJeff Roberson * 7638355f576SJeff Roberson * Returns: 7648355f576SJeff Roberson * A pointer to the alloced memory or possibly 7658355f576SJeff Roberson * NULL if M_NOWAIT is set. 7668355f576SJeff Roberson */ 7678355f576SJeff Roberson static void * 7688355f576SJeff Roberson obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 7698355f576SJeff Roberson { 7708355f576SJeff Roberson vm_offset_t zkva; 7718355f576SJeff Roberson vm_offset_t retkva; 7728355f576SJeff Roberson vm_page_t p; 7738355f576SJeff Roberson int pages; 7748355f576SJeff Roberson 7758355f576SJeff Roberson retkva = NULL; 7768355f576SJeff Roberson pages = zone->uz_pages; 7778355f576SJeff Roberson 7788355f576SJeff Roberson /* 7798355f576SJeff Roberson * This looks a little weird since we're getting one page at a time 7808355f576SJeff Roberson */ 7818355f576SJeff Roberson while (bytes > 0) { 7828355f576SJeff Roberson p = vm_page_alloc(zone->uz_obj, pages, 7838355f576SJeff Roberson VM_ALLOC_INTERRUPT); 7848355f576SJeff Roberson if (p == NULL) 7858355f576SJeff Roberson return (NULL); 7868355f576SJeff Roberson 7878355f576SJeff Roberson zkva = zone->uz_kva + pages * PAGE_SIZE; 7888355f576SJeff Roberson if (retkva == NULL) 7898355f576SJeff Roberson retkva = zkva; 7908355f576SJeff Roberson pmap_qenter(zkva, &p, 1); 7918355f576SJeff Roberson bytes -= PAGE_SIZE; 7928355f576SJeff Roberson pages += 1; 7938355f576SJeff Roberson } 7948355f576SJeff Roberson 7958355f576SJeff Roberson *flags = UMA_SLAB_PRIV; 7968355f576SJeff Roberson 7978355f576SJeff Roberson return ((void *)retkva); 7988355f576SJeff Roberson } 7998355f576SJeff Roberson 8008355f576SJeff Roberson /* 8018355f576SJeff Roberson * Frees a number of pages to the system 8028355f576SJeff Roberson * 8038355f576SJeff Roberson * Arguments: 8048355f576SJeff Roberson * mem A pointer to the memory to be freed 8058355f576SJeff Roberson * size The size of the memory being freed 8068355f576SJeff Roberson * flags The original p->us_flags field 8078355f576SJeff Roberson * 8088355f576SJeff Roberson * Returns: 8098355f576SJeff Roberson * Nothing 8108355f576SJeff Roberson * 8118355f576SJeff Roberson */ 8128355f576SJeff Roberson static void 8138355f576SJeff Roberson page_free(void *mem, int size, u_int8_t flags) 8148355f576SJeff Roberson { 8158355f576SJeff Roberson vm_map_t map; 8168355f576SJeff Roberson if (flags & UMA_SLAB_KMEM) 8178355f576SJeff Roberson map = kmem_map; 8188355f576SJeff Roberson else if (flags & UMA_SLAB_KMAP) 8198355f576SJeff Roberson map = kernel_map; 8208355f576SJeff Roberson else 8218355f576SJeff Roberson panic("UMA: page_free used with invalid flags %d\n", flags); 8228355f576SJeff Roberson 8238355f576SJeff Roberson kmem_free(map, (vm_offset_t)mem, size); 8248355f576SJeff Roberson } 8258355f576SJeff Roberson 8268355f576SJeff Roberson /* 8278355f576SJeff Roberson * Zero fill initializer 8288355f576SJeff Roberson * 8298355f576SJeff Roberson * Arguments/Returns follow uma_init specifications 8308355f576SJeff Roberson * 8318355f576SJeff Roberson */ 8328355f576SJeff Roberson static void 8338355f576SJeff Roberson zero_init(void *mem, int size) 8348355f576SJeff Roberson { 8358355f576SJeff Roberson bzero(mem, size); 8368355f576SJeff Roberson } 8378355f576SJeff Roberson 8388355f576SJeff Roberson /* 8398355f576SJeff Roberson * Finish creating a small uma zone. This calculates ipers, and the zone size. 8408355f576SJeff Roberson * 8418355f576SJeff Roberson * Arguments 8428355f576SJeff Roberson * zone The zone we should initialize 8438355f576SJeff Roberson * 8448355f576SJeff Roberson * Returns 8458355f576SJeff Roberson * Nothing 8468355f576SJeff Roberson */ 8478355f576SJeff Roberson static void 8488355f576SJeff Roberson zone_small_init(uma_zone_t zone) 8498355f576SJeff Roberson { 8508355f576SJeff Roberson int rsize; 8518355f576SJeff Roberson int memused; 8528355f576SJeff Roberson int ipers; 8538355f576SJeff Roberson 8548355f576SJeff Roberson rsize = zone->uz_size; 8558355f576SJeff Roberson 8568355f576SJeff Roberson if (rsize < UMA_SMALLEST_UNIT) 8578355f576SJeff Roberson rsize = UMA_SMALLEST_UNIT; 8588355f576SJeff Roberson 8598355f576SJeff Roberson if (rsize & zone->uz_align) 8608355f576SJeff Roberson rsize = (rsize & ~zone->uz_align) + (zone->uz_align + 1); 8618355f576SJeff Roberson 8628355f576SJeff Roberson zone->uz_rsize = rsize; 8638355f576SJeff Roberson 8648355f576SJeff Roberson rsize += 1; /* Account for the byte of linkage */ 8658355f576SJeff Roberson zone->uz_ipers = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) / rsize; 8668355f576SJeff Roberson zone->uz_ppera = 1; 8678355f576SJeff Roberson 8688355f576SJeff Roberson memused = zone->uz_ipers * zone->uz_rsize; 8698355f576SJeff Roberson 8708355f576SJeff Roberson /* Can we do any better? */ 8718355f576SJeff Roberson if ((UMA_SLAB_SIZE - memused) >= UMA_MAX_WASTE) { 8728355f576SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 8738355f576SJeff Roberson return; 8748355f576SJeff Roberson ipers = UMA_SLAB_SIZE / zone->uz_rsize; 8758355f576SJeff Roberson if (ipers > zone->uz_ipers) { 8768355f576SJeff Roberson zone->uz_flags |= UMA_ZFLAG_OFFPAGE; 8778355f576SJeff Roberson zone->uz_ipers = ipers; 8788355f576SJeff Roberson } 8798355f576SJeff Roberson } 8808355f576SJeff Roberson 8818355f576SJeff Roberson } 8828355f576SJeff Roberson 8838355f576SJeff Roberson /* 8848355f576SJeff Roberson * Finish creating a large (> UMA_SLAB_SIZE) uma zone. Just give in and do 8858355f576SJeff Roberson * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 8868355f576SJeff Roberson * more complicated. 8878355f576SJeff Roberson * 8888355f576SJeff Roberson * Arguments 8898355f576SJeff Roberson * zone The zone we should initialize 8908355f576SJeff Roberson * 8918355f576SJeff Roberson * Returns 8928355f576SJeff Roberson * Nothing 8938355f576SJeff Roberson */ 8948355f576SJeff Roberson static void 8958355f576SJeff Roberson zone_large_init(uma_zone_t zone) 8968355f576SJeff Roberson { 8978355f576SJeff Roberson int pages; 8988355f576SJeff Roberson 8998355f576SJeff Roberson pages = zone->uz_size / UMA_SLAB_SIZE; 9008355f576SJeff Roberson 9018355f576SJeff Roberson /* Account for remainder */ 9028355f576SJeff Roberson if ((pages * UMA_SLAB_SIZE) < zone->uz_size) 9038355f576SJeff Roberson pages++; 9048355f576SJeff Roberson 9058355f576SJeff Roberson zone->uz_ppera = pages; 9068355f576SJeff Roberson zone->uz_ipers = 1; 9078355f576SJeff Roberson 9088355f576SJeff Roberson zone->uz_flags |= UMA_ZFLAG_OFFPAGE; 9098355f576SJeff Roberson zone->uz_rsize = zone->uz_size; 9108355f576SJeff Roberson } 9118355f576SJeff Roberson 9128355f576SJeff Roberson /* 9138355f576SJeff Roberson * Zone header ctor. This initializes all fields, locks, etc. And inserts 9148355f576SJeff Roberson * the zone onto the global zone list. 9158355f576SJeff Roberson * 9168355f576SJeff Roberson * Arguments/Returns follow uma_ctor specifications 9178355f576SJeff Roberson * udata Actually uma_zcreat_args 9188355f576SJeff Roberson * 9198355f576SJeff Roberson */ 9208355f576SJeff Roberson 9218355f576SJeff Roberson static void 9228355f576SJeff Roberson zone_ctor(void *mem, int size, void *udata) 9238355f576SJeff Roberson { 9248355f576SJeff Roberson struct uma_zctor_args *arg = udata; 9258355f576SJeff Roberson uma_zone_t zone = mem; 9268355f576SJeff Roberson int cplen; 9278355f576SJeff Roberson int cpu; 9288355f576SJeff Roberson 9298355f576SJeff Roberson bzero(zone, size); 9308355f576SJeff Roberson zone->uz_name = arg->name; 9318355f576SJeff Roberson zone->uz_size = arg->size; 9328355f576SJeff Roberson zone->uz_ctor = arg->ctor; 9338355f576SJeff Roberson zone->uz_dtor = arg->dtor; 9348355f576SJeff Roberson zone->uz_init = arg->uminit; 9358355f576SJeff Roberson zone->uz_align = arg->align; 9368355f576SJeff Roberson zone->uz_free = 0; 9378355f576SJeff Roberson zone->uz_pages = 0; 9388355f576SJeff Roberson zone->uz_flags = 0; 9398355f576SJeff Roberson zone->uz_allocf = page_alloc; 9408355f576SJeff Roberson zone->uz_freef = page_free; 9418355f576SJeff Roberson 9428355f576SJeff Roberson if (arg->flags & UMA_ZONE_ZINIT) 9438355f576SJeff Roberson zone->uz_init = zero_init; 9448355f576SJeff Roberson 9458355f576SJeff Roberson if (arg->flags & UMA_ZONE_INTERNAL) 9468355f576SJeff Roberson zone->uz_flags |= UMA_ZFLAG_INTERNAL; 9478355f576SJeff Roberson 9488355f576SJeff Roberson if (arg->flags & UMA_ZONE_MALLOC) 9498355f576SJeff Roberson zone->uz_flags |= UMA_ZFLAG_MALLOC; 9508355f576SJeff Roberson 9518355f576SJeff Roberson if (arg->flags & UMA_ZONE_NOFREE) 9528355f576SJeff Roberson zone->uz_flags |= UMA_ZFLAG_NOFREE; 9538355f576SJeff Roberson 9548355f576SJeff Roberson if (zone->uz_size > UMA_SLAB_SIZE) 9558355f576SJeff Roberson zone_large_init(zone); 9568355f576SJeff Roberson else 9578355f576SJeff Roberson zone_small_init(zone); 9588355f576SJeff Roberson 9598355f576SJeff Roberson /* We do this so that the per cpu lock name is unique for each zone */ 9608355f576SJeff Roberson memcpy(zone->uz_lname, "PCPU ", 5); 9618355f576SJeff Roberson cplen = min(strlen(zone->uz_name) + 1, LOCKNAME_LEN - 6); 9628355f576SJeff Roberson memcpy(zone->uz_lname+5, zone->uz_name, cplen); 9638355f576SJeff Roberson zone->uz_lname[LOCKNAME_LEN - 1] = '\0'; 9648355f576SJeff Roberson 9658355f576SJeff Roberson /* 9668355f576SJeff Roberson * If we're putting the slab header in the actual page we need to 9678355f576SJeff Roberson * figure out where in each page it goes. This calculates a right 9688355f576SJeff Roberson * justified offset into the memory on a ALIGN_PTR boundary. 9698355f576SJeff Roberson */ 9708355f576SJeff Roberson if (!(zone->uz_flags & UMA_ZFLAG_OFFPAGE)) { 9718355f576SJeff Roberson int totsize; 9728355f576SJeff Roberson int waste; 9738355f576SJeff Roberson 9748355f576SJeff Roberson /* Size of the slab struct and free list */ 9758355f576SJeff Roberson totsize = sizeof(struct uma_slab) + zone->uz_ipers; 9768355f576SJeff Roberson if (totsize & UMA_ALIGN_PTR) 9778355f576SJeff Roberson totsize = (totsize & ~UMA_ALIGN_PTR) + 9788355f576SJeff Roberson (UMA_ALIGN_PTR + 1); 9798355f576SJeff Roberson zone->uz_pgoff = UMA_SLAB_SIZE - totsize; 9808355f576SJeff Roberson 9818355f576SJeff Roberson waste = zone->uz_pgoff; 9828355f576SJeff Roberson waste -= (zone->uz_ipers * zone->uz_rsize); 9838355f576SJeff Roberson 9848355f576SJeff Roberson /* 9858355f576SJeff Roberson * This calculates how much space we have for cache line size 9868355f576SJeff Roberson * optimizations. It works by offseting each slab slightly. 9878355f576SJeff Roberson * Currently it breaks on x86, and so it is disabled. 9888355f576SJeff Roberson */ 9898355f576SJeff Roberson 9908355f576SJeff Roberson if (zone->uz_align < UMA_CACHE_INC && waste > UMA_CACHE_INC) { 9918355f576SJeff Roberson zone->uz_cachemax = waste - UMA_CACHE_INC; 9928355f576SJeff Roberson zone->uz_cacheoff = 0; 9938355f576SJeff Roberson } 9948355f576SJeff Roberson 9958355f576SJeff Roberson totsize = zone->uz_pgoff + sizeof(struct uma_slab) 9968355f576SJeff Roberson + zone->uz_ipers; 9978355f576SJeff Roberson /* I don't think it's possible, but I'll make sure anyway */ 9988355f576SJeff Roberson if (totsize > UMA_SLAB_SIZE) { 9998355f576SJeff Roberson printf("zone %s ipers %d rsize %d size %d\n", 10008355f576SJeff Roberson zone->uz_name, zone->uz_ipers, zone->uz_rsize, 10018355f576SJeff Roberson zone->uz_size); 10028355f576SJeff Roberson panic("UMA slab won't fit.\n"); 10038355f576SJeff Roberson } 10048355f576SJeff Roberson } else { 10058355f576SJeff Roberson /* hash_expand here to allocate the initial hash table */ 10068355f576SJeff Roberson hash_expand(&zone->uz_hash); 10078355f576SJeff Roberson zone->uz_pgoff = 0; 10088355f576SJeff Roberson } 10098355f576SJeff Roberson 10108355f576SJeff Roberson #ifdef UMA_DEBUG 10118355f576SJeff Roberson printf("%s(%p) size = %d ipers = %d ppera = %d pgoff = %d\n", 10128355f576SJeff Roberson zone->uz_name, zone, 10138355f576SJeff Roberson zone->uz_size, zone->uz_ipers, 10148355f576SJeff Roberson zone->uz_ppera, zone->uz_pgoff); 10158355f576SJeff Roberson #endif 10168355f576SJeff Roberson ZONE_LOCK_INIT(zone); 10178355f576SJeff Roberson 10188355f576SJeff Roberson mtx_lock(&uma_mtx); 10198355f576SJeff Roberson LIST_INSERT_HEAD(&uma_zones, zone, uz_link); 10208355f576SJeff Roberson mtx_unlock(&uma_mtx); 10218355f576SJeff Roberson 10228355f576SJeff Roberson /* 10238355f576SJeff Roberson * Some internal zones don't have room allocated for the per cpu 10248355f576SJeff Roberson * caches. If we're internal, bail out here. 10258355f576SJeff Roberson */ 10268355f576SJeff Roberson 10278355f576SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 10288355f576SJeff Roberson return; 10298355f576SJeff Roberson 10308355f576SJeff Roberson if (zone->uz_ipers < UMA_BUCKET_SIZE) 1031a553d4b8SJeff Roberson zone->uz_count = zone->uz_ipers - 1; 10328355f576SJeff Roberson else 1033a553d4b8SJeff Roberson zone->uz_count = UMA_BUCKET_SIZE - 1; 1034a553d4b8SJeff Roberson 1035a553d4b8SJeff Roberson for (cpu = 0; cpu < maxcpu; cpu++) 10368355f576SJeff Roberson CPU_LOCK_INIT(zone, cpu); 10378355f576SJeff Roberson } 10388355f576SJeff Roberson 10398355f576SJeff Roberson /* 10409c2cd7e5SJeff Roberson * Zone header dtor. This frees all data, destroys locks, frees the hash table 10419c2cd7e5SJeff Roberson * and removes the zone from the global list. 10429c2cd7e5SJeff Roberson * 10439c2cd7e5SJeff Roberson * Arguments/Returns follow uma_dtor specifications 10449c2cd7e5SJeff Roberson * udata unused 10459c2cd7e5SJeff Roberson */ 10469c2cd7e5SJeff Roberson 10479c2cd7e5SJeff Roberson static void 10489c2cd7e5SJeff Roberson zone_dtor(void *arg, int size, void *udata) 10499c2cd7e5SJeff Roberson { 10509c2cd7e5SJeff Roberson uma_zone_t zone; 10519c2cd7e5SJeff Roberson int cpu; 10529c2cd7e5SJeff Roberson 10539c2cd7e5SJeff Roberson zone = (uma_zone_t)arg; 10549c2cd7e5SJeff Roberson 10559c2cd7e5SJeff Roberson mtx_lock(&uma_mtx); 10569c2cd7e5SJeff Roberson LIST_REMOVE(zone, uz_link); 10579c2cd7e5SJeff Roberson mtx_unlock(&uma_mtx); 10589c2cd7e5SJeff Roberson 10599c2cd7e5SJeff Roberson ZONE_LOCK(zone); 10609c2cd7e5SJeff Roberson zone->uz_wssize = 0; 10619c2cd7e5SJeff Roberson ZONE_UNLOCK(zone); 10629c2cd7e5SJeff Roberson 10639c2cd7e5SJeff Roberson zone_drain(zone); 10649c2cd7e5SJeff Roberson ZONE_LOCK(zone); 10659c2cd7e5SJeff Roberson if (zone->uz_free != 0) 10669c2cd7e5SJeff Roberson printf("Zone %s was not empty. Lost %d pages of memory.\n", 10679c2cd7e5SJeff Roberson zone->uz_name, zone->uz_pages); 10689c2cd7e5SJeff Roberson 10699c2cd7e5SJeff Roberson if ((zone->uz_flags & UMA_ZFLAG_INTERNAL) != 0) 10709c2cd7e5SJeff Roberson for (cpu = 0; cpu < maxcpu; cpu++) 10719c2cd7e5SJeff Roberson CPU_LOCK_FINI(zone, cpu); 10729c2cd7e5SJeff Roberson 10739c2cd7e5SJeff Roberson if ((zone->uz_flags & UMA_ZFLAG_OFFPAGE) != 0) 10749c2cd7e5SJeff Roberson hash_free(&zone->uz_hash); 10759c2cd7e5SJeff Roberson 10769c2cd7e5SJeff Roberson ZONE_UNLOCK(zone); 10779c2cd7e5SJeff Roberson ZONE_LOCK_FINI(zone); 10789c2cd7e5SJeff Roberson } 10799c2cd7e5SJeff Roberson /* 10808355f576SJeff Roberson * Traverses every zone in the system and calls a callback 10818355f576SJeff Roberson * 10828355f576SJeff Roberson * Arguments: 10838355f576SJeff Roberson * zfunc A pointer to a function which accepts a zone 10848355f576SJeff Roberson * as an argument. 10858355f576SJeff Roberson * 10868355f576SJeff Roberson * Returns: 10878355f576SJeff Roberson * Nothing 10888355f576SJeff Roberson */ 10898355f576SJeff Roberson static void 10908355f576SJeff Roberson zone_foreach(void (*zfunc)(uma_zone_t)) 10918355f576SJeff Roberson { 10928355f576SJeff Roberson uma_zone_t zone; 10938355f576SJeff Roberson 10948355f576SJeff Roberson mtx_lock(&uma_mtx); 10958355f576SJeff Roberson LIST_FOREACH(zone, &uma_zones, uz_link) { 10968355f576SJeff Roberson zfunc(zone); 10978355f576SJeff Roberson } 10988355f576SJeff Roberson mtx_unlock(&uma_mtx); 10998355f576SJeff Roberson } 11008355f576SJeff Roberson 11018355f576SJeff Roberson /* Public functions */ 11028355f576SJeff Roberson /* See uma.h */ 11038355f576SJeff Roberson void 11048355f576SJeff Roberson uma_startup(void *bootmem) 11058355f576SJeff Roberson { 11068355f576SJeff Roberson struct uma_zctor_args args; 11078355f576SJeff Roberson uma_slab_t slab; 11088355f576SJeff Roberson int slabsize; 11098355f576SJeff Roberson int i; 11108355f576SJeff Roberson 11118355f576SJeff Roberson #ifdef UMA_DEBUG 11128355f576SJeff Roberson printf("Creating uma zone headers zone.\n"); 11138355f576SJeff Roberson #endif 11148355f576SJeff Roberson #ifdef SMP 11158355f576SJeff Roberson maxcpu = mp_maxid + 1; 11168355f576SJeff Roberson #else 11178355f576SJeff Roberson maxcpu = 1; 11188355f576SJeff Roberson #endif 11198355f576SJeff Roberson #ifdef UMA_DEBUG 11208355f576SJeff Roberson printf("Max cpu = %d, mp_maxid = %d\n", maxcpu, mp_maxid); 11218355f576SJeff Roberson Debugger("stop"); 11228355f576SJeff Roberson #endif 11236008862bSJohn Baldwin mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF); 11248355f576SJeff Roberson /* "manually" Create the initial zone */ 11258355f576SJeff Roberson args.name = "UMA Zones"; 11268355f576SJeff Roberson args.size = sizeof(struct uma_zone) + 11278355f576SJeff Roberson (sizeof(struct uma_cache) * (maxcpu - 1)); 11288355f576SJeff Roberson args.ctor = zone_ctor; 11299c2cd7e5SJeff Roberson args.dtor = zone_dtor; 11308355f576SJeff Roberson args.uminit = zero_init; 11318355f576SJeff Roberson args.fini = NULL; 11328355f576SJeff Roberson args.align = 32 - 1; 11338355f576SJeff Roberson args.flags = UMA_ZONE_INTERNAL; 11348355f576SJeff Roberson /* The initial zone has no Per cpu queues so it's smaller */ 11358355f576SJeff Roberson zone_ctor(zones, sizeof(struct uma_zone), &args); 11368355f576SJeff Roberson 11378355f576SJeff Roberson #ifdef UMA_DEBUG 11388355f576SJeff Roberson printf("Filling boot free list.\n"); 11398355f576SJeff Roberson #endif 11408355f576SJeff Roberson for (i = 0; i < UMA_BOOT_PAGES; i++) { 11418355f576SJeff Roberson slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE)); 11428355f576SJeff Roberson slab->us_data = (u_int8_t *)slab; 11438355f576SJeff Roberson slab->us_flags = UMA_SLAB_BOOT; 11448355f576SJeff Roberson LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link); 11458355f576SJeff Roberson uma_boot_free++; 11468355f576SJeff Roberson } 11478355f576SJeff Roberson 11488355f576SJeff Roberson #ifdef UMA_DEBUG 11498355f576SJeff Roberson printf("Creating slab zone.\n"); 11508355f576SJeff Roberson #endif 11518355f576SJeff Roberson 11528355f576SJeff Roberson /* 11538355f576SJeff Roberson * This is the max number of free list items we'll have with 11548355f576SJeff Roberson * offpage slabs. 11558355f576SJeff Roberson */ 11568355f576SJeff Roberson 11578355f576SJeff Roberson slabsize = UMA_SLAB_SIZE - sizeof(struct uma_slab); 11588355f576SJeff Roberson slabsize /= UMA_MAX_WASTE; 11598355f576SJeff Roberson slabsize++; /* In case there it's rounded */ 11608355f576SJeff Roberson slabsize += sizeof(struct uma_slab); 11618355f576SJeff Roberson 11628355f576SJeff Roberson /* Now make a zone for slab headers */ 11638355f576SJeff Roberson slabzone = uma_zcreate("UMA Slabs", 11648355f576SJeff Roberson slabsize, 11658355f576SJeff Roberson NULL, NULL, NULL, NULL, 11668355f576SJeff Roberson UMA_ALIGN_PTR, UMA_ZONE_INTERNAL); 11678355f576SJeff Roberson 11688355f576SJeff Roberson hashzone = uma_zcreate("UMA Hash", 11698355f576SJeff Roberson sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 11708355f576SJeff Roberson NULL, NULL, NULL, NULL, 11718355f576SJeff Roberson UMA_ALIGN_PTR, UMA_ZONE_INTERNAL); 11728355f576SJeff Roberson 11738355f576SJeff Roberson bucketzone = uma_zcreate("UMA Buckets", sizeof(struct uma_bucket), 11748355f576SJeff Roberson NULL, NULL, NULL, NULL, 11758355f576SJeff Roberson UMA_ALIGN_PTR, UMA_ZONE_INTERNAL); 11768355f576SJeff Roberson 11778355f576SJeff Roberson 11788355f576SJeff Roberson #ifdef UMA_DEBUG 11798355f576SJeff Roberson printf("UMA startup complete.\n"); 11808355f576SJeff Roberson #endif 11818355f576SJeff Roberson } 11828355f576SJeff Roberson 11838355f576SJeff Roberson /* see uma.h */ 11848355f576SJeff Roberson void 11858355f576SJeff Roberson uma_startup2(void *hashmem, u_long elems) 11868355f576SJeff Roberson { 11878355f576SJeff Roberson bzero(hashmem, elems * sizeof(void *)); 11888355f576SJeff Roberson mallochash->uh_slab_hash = hashmem; 11898355f576SJeff Roberson mallochash->uh_hashsize = elems; 11908355f576SJeff Roberson mallochash->uh_hashmask = elems - 1; 11918355f576SJeff Roberson booted = 1; 119286bbae32SJeff Roberson bucket_enable(); 11938355f576SJeff Roberson #ifdef UMA_DEBUG 11948355f576SJeff Roberson printf("UMA startup2 complete.\n"); 11958355f576SJeff Roberson #endif 11968355f576SJeff Roberson } 11978355f576SJeff Roberson 11988355f576SJeff Roberson /* 11998355f576SJeff Roberson * Initialize our callout handle 12008355f576SJeff Roberson * 12018355f576SJeff Roberson */ 12028355f576SJeff Roberson 12038355f576SJeff Roberson static void 12048355f576SJeff Roberson uma_startup3(void) 12058355f576SJeff Roberson { 12068355f576SJeff Roberson #ifdef UMA_DEBUG 12078355f576SJeff Roberson printf("Starting callout.\n"); 12088355f576SJeff Roberson #endif 12098355f576SJeff Roberson callout_init(&uma_callout, 0); 12108355f576SJeff Roberson callout_reset(&uma_callout, UMA_WORKING_TIME * hz, uma_timeout, NULL); 12118355f576SJeff Roberson #ifdef UMA_DEBUG 12128355f576SJeff Roberson printf("UMA startup3 complete.\n"); 12138355f576SJeff Roberson #endif 12148355f576SJeff Roberson } 12158355f576SJeff Roberson 12168355f576SJeff Roberson /* See uma.h */ 12178355f576SJeff Roberson uma_zone_t 12188355f576SJeff Roberson uma_zcreate(char *name, int size, uma_ctor ctor, uma_dtor dtor, uma_init uminit, 12198355f576SJeff Roberson uma_fini fini, int align, u_int16_t flags) 12208355f576SJeff Roberson 12218355f576SJeff Roberson { 12228355f576SJeff Roberson struct uma_zctor_args args; 12238355f576SJeff Roberson 12248355f576SJeff Roberson /* This stuff is essential for the zone ctor */ 12258355f576SJeff Roberson args.name = name; 12268355f576SJeff Roberson args.size = size; 12278355f576SJeff Roberson args.ctor = ctor; 12288355f576SJeff Roberson args.dtor = dtor; 12298355f576SJeff Roberson args.uminit = uminit; 12308355f576SJeff Roberson args.fini = fini; 12318355f576SJeff Roberson args.align = align; 12328355f576SJeff Roberson args.flags = flags; 12338355f576SJeff Roberson 1234a553d4b8SJeff Roberson return (uma_zalloc_internal(zones, &args, M_WAITOK, NULL)); 12358355f576SJeff Roberson } 12368355f576SJeff Roberson 12378355f576SJeff Roberson /* See uma.h */ 12389c2cd7e5SJeff Roberson void 12399c2cd7e5SJeff Roberson uma_zdestroy(uma_zone_t zone) 12409c2cd7e5SJeff Roberson { 12419c2cd7e5SJeff Roberson uma_zfree_internal(zones, zone, NULL, 0); 12429c2cd7e5SJeff Roberson } 12439c2cd7e5SJeff Roberson 12449c2cd7e5SJeff Roberson /* See uma.h */ 12458355f576SJeff Roberson void * 12468355f576SJeff Roberson uma_zalloc_arg(uma_zone_t zone, void *udata, int wait) 12478355f576SJeff Roberson { 12488355f576SJeff Roberson void *item; 12498355f576SJeff Roberson uma_cache_t cache; 12508355f576SJeff Roberson uma_bucket_t bucket; 12518355f576SJeff Roberson int cpu; 12528355f576SJeff Roberson 12538355f576SJeff Roberson /* This is the fast path allocation */ 12548355f576SJeff Roberson #ifdef UMA_DEBUG_ALLOC_1 12558355f576SJeff Roberson printf("Allocating one item from %s(%p)\n", zone->uz_name, zone); 12568355f576SJeff Roberson #endif 1257a553d4b8SJeff Roberson 1258a553d4b8SJeff Roberson zalloc_restart: 12598355f576SJeff Roberson cpu = PCPU_GET(cpuid); 12608355f576SJeff Roberson CPU_LOCK(zone, cpu); 12618355f576SJeff Roberson cache = &zone->uz_cpu[cpu]; 12628355f576SJeff Roberson 12638355f576SJeff Roberson zalloc_start: 12648355f576SJeff Roberson bucket = cache->uc_allocbucket; 12658355f576SJeff Roberson 12668355f576SJeff Roberson if (bucket) { 12678355f576SJeff Roberson if (bucket->ub_ptr > -1) { 12688355f576SJeff Roberson item = bucket->ub_bucket[bucket->ub_ptr]; 12698355f576SJeff Roberson #ifdef INVARIANTS 12708355f576SJeff Roberson bucket->ub_bucket[bucket->ub_ptr] = NULL; 12718355f576SJeff Roberson #endif 12728355f576SJeff Roberson bucket->ub_ptr--; 12738355f576SJeff Roberson KASSERT(item != NULL, 12748355f576SJeff Roberson ("uma_zalloc: Bucket pointer mangled.")); 12758355f576SJeff Roberson cache->uc_allocs++; 12768355f576SJeff Roberson CPU_UNLOCK(zone, cpu); 12778355f576SJeff Roberson if (zone->uz_ctor) 12788355f576SJeff Roberson zone->uz_ctor(item, zone->uz_size, udata); 12798355f576SJeff Roberson return (item); 12808355f576SJeff Roberson } else if (cache->uc_freebucket) { 12818355f576SJeff Roberson /* 12828355f576SJeff Roberson * We have run out of items in our allocbucket. 12838355f576SJeff Roberson * See if we can switch with our free bucket. 12848355f576SJeff Roberson */ 12858355f576SJeff Roberson if (cache->uc_freebucket->ub_ptr > -1) { 12868355f576SJeff Roberson uma_bucket_t swap; 12878355f576SJeff Roberson 12888355f576SJeff Roberson #ifdef UMA_DEBUG_ALLOC 12898355f576SJeff Roberson printf("uma_zalloc: Swapping empty with alloc.\n"); 12908355f576SJeff Roberson #endif 12918355f576SJeff Roberson swap = cache->uc_freebucket; 12928355f576SJeff Roberson cache->uc_freebucket = cache->uc_allocbucket; 12938355f576SJeff Roberson cache->uc_allocbucket = swap; 12948355f576SJeff Roberson 12958355f576SJeff Roberson goto zalloc_start; 12968355f576SJeff Roberson } 12978355f576SJeff Roberson } 12988355f576SJeff Roberson } 1299a553d4b8SJeff Roberson ZONE_LOCK(zone); 1300a553d4b8SJeff Roberson /* Since we have locked the zone we may as well send back our stats */ 1301a553d4b8SJeff Roberson zone->uz_allocs += cache->uc_allocs; 1302a553d4b8SJeff Roberson cache->uc_allocs = 0; 13038355f576SJeff Roberson 1304a553d4b8SJeff Roberson /* Our old one is now a free bucket */ 1305a553d4b8SJeff Roberson if (cache->uc_allocbucket) { 1306a553d4b8SJeff Roberson KASSERT(cache->uc_allocbucket->ub_ptr == -1, 1307a553d4b8SJeff Roberson ("uma_zalloc_arg: Freeing a non free bucket.")); 1308a553d4b8SJeff Roberson LIST_INSERT_HEAD(&zone->uz_free_bucket, 1309a553d4b8SJeff Roberson cache->uc_allocbucket, ub_link); 1310a553d4b8SJeff Roberson cache->uc_allocbucket = NULL; 1311a553d4b8SJeff Roberson } 13128355f576SJeff Roberson 1313a553d4b8SJeff Roberson /* Check the free list for a new alloc bucket */ 1314a553d4b8SJeff Roberson if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) { 1315a553d4b8SJeff Roberson KASSERT(bucket->ub_ptr != -1, 1316a553d4b8SJeff Roberson ("uma_zalloc_arg: Returning an empty bucket.")); 13178355f576SJeff Roberson 1318a553d4b8SJeff Roberson LIST_REMOVE(bucket, ub_link); 1319a553d4b8SJeff Roberson cache->uc_allocbucket = bucket; 1320a553d4b8SJeff Roberson ZONE_UNLOCK(zone); 13218355f576SJeff Roberson goto zalloc_start; 1322a553d4b8SJeff Roberson } 1323a553d4b8SJeff Roberson /* Bump up our uz_count so we get here less */ 1324a553d4b8SJeff Roberson if (zone->uz_count < UMA_BUCKET_SIZE - 1) 1325a553d4b8SJeff Roberson zone->uz_count++; 1326a553d4b8SJeff Roberson 1327a553d4b8SJeff Roberson /* We are no longer associated with this cpu!!! */ 1328a553d4b8SJeff Roberson CPU_UNLOCK(zone, cpu); 13298355f576SJeff Roberson 13308355f576SJeff Roberson /* 1331a553d4b8SJeff Roberson * Now lets just fill a bucket and put it on the free list. If that 1332a553d4b8SJeff Roberson * works we'll restart the allocation from the begining. 1333a553d4b8SJeff Roberson * 1334a553d4b8SJeff Roberson * Try this zone's free list first so we don't allocate extra buckets. 13358355f576SJeff Roberson */ 1336a553d4b8SJeff Roberson 1337a553d4b8SJeff Roberson if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) 1338a553d4b8SJeff Roberson LIST_REMOVE(bucket, ub_link); 1339a553d4b8SJeff Roberson 1340a553d4b8SJeff Roberson /* Now we no longer need the zone lock. */ 1341a553d4b8SJeff Roberson ZONE_UNLOCK(zone); 1342a553d4b8SJeff Roberson 1343a553d4b8SJeff Roberson if (bucket == NULL) 1344a553d4b8SJeff Roberson bucket = uma_zalloc_internal(bucketzone, 1345a553d4b8SJeff Roberson NULL, wait, NULL); 1346a553d4b8SJeff Roberson 1347a553d4b8SJeff Roberson if (bucket != NULL) { 1348a553d4b8SJeff Roberson #ifdef INVARIANTS 1349a553d4b8SJeff Roberson bzero(bucket, bucketzone->uz_size); 1350a553d4b8SJeff Roberson #endif 1351a553d4b8SJeff Roberson bucket->ub_ptr = -1; 1352a553d4b8SJeff Roberson 1353a553d4b8SJeff Roberson if (uma_zalloc_internal(zone, udata, wait, bucket)) 1354a553d4b8SJeff Roberson goto zalloc_restart; 1355a553d4b8SJeff Roberson else 1356a553d4b8SJeff Roberson uma_zfree_internal(bucketzone, bucket, NULL, 0); 1357a553d4b8SJeff Roberson } 1358a553d4b8SJeff Roberson /* 1359a553d4b8SJeff Roberson * We may not get a bucket if we recurse, so 1360a553d4b8SJeff Roberson * return an actual item. 1361a553d4b8SJeff Roberson */ 1362a553d4b8SJeff Roberson #ifdef UMA_DEBUG 1363a553d4b8SJeff Roberson printf("uma_zalloc_arg: Bucketzone returned NULL\n"); 1364a553d4b8SJeff Roberson #endif 1365a553d4b8SJeff Roberson 1366a553d4b8SJeff Roberson return (uma_zalloc_internal(zone, udata, wait, NULL)); 13678355f576SJeff Roberson } 13688355f576SJeff Roberson 13698355f576SJeff Roberson /* 13708355f576SJeff Roberson * Allocates an item for an internal zone OR fills a bucket 13718355f576SJeff Roberson * 13728355f576SJeff Roberson * Arguments 13738355f576SJeff Roberson * zone The zone to alloc for. 13748355f576SJeff Roberson * udata The data to be passed to the constructor. 13758355f576SJeff Roberson * wait M_WAITOK or M_NOWAIT. 1376a553d4b8SJeff Roberson * bucket The bucket to fill or NULL 13778355f576SJeff Roberson * 13788355f576SJeff Roberson * Returns 13798355f576SJeff Roberson * NULL if there is no memory and M_NOWAIT is set 13808355f576SJeff Roberson * An item if called on an interal zone 13818355f576SJeff Roberson * Non NULL if called to fill a bucket and it was successful. 13828355f576SJeff Roberson * 13838355f576SJeff Roberson * Discussion: 13848355f576SJeff Roberson * This was much cleaner before it had to do per cpu caches. It is 13858355f576SJeff Roberson * complicated now because it has to handle the simple internal case, and 1386a553d4b8SJeff Roberson * the more involved bucket filling and allocation. 13878355f576SJeff Roberson */ 13888355f576SJeff Roberson 13898355f576SJeff Roberson static void * 1390a553d4b8SJeff Roberson uma_zalloc_internal(uma_zone_t zone, void *udata, int wait, uma_bucket_t bucket) 13918355f576SJeff Roberson { 13928355f576SJeff Roberson uma_slab_t slab; 13938355f576SJeff Roberson u_int8_t freei; 13948355f576SJeff Roberson void *item; 13958355f576SJeff Roberson 13968355f576SJeff Roberson item = NULL; 13978355f576SJeff Roberson 13988355f576SJeff Roberson /* 1399a553d4b8SJeff Roberson * This is to stop us from allocating per cpu buckets while we're 1400a553d4b8SJeff Roberson * running out of UMA_BOOT_PAGES. Otherwise, we would exhaust the 1401a553d4b8SJeff Roberson * boot pages. 14028355f576SJeff Roberson */ 14038355f576SJeff Roberson 140486bbae32SJeff Roberson if (bucketdisable && zone == bucketzone) 14058355f576SJeff Roberson return (NULL); 14068355f576SJeff Roberson 14078355f576SJeff Roberson #ifdef UMA_DEBUG_ALLOC 14088355f576SJeff Roberson printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone); 14098355f576SJeff Roberson #endif 14108355f576SJeff Roberson ZONE_LOCK(zone); 14118355f576SJeff Roberson 14128355f576SJeff Roberson /* 1413a553d4b8SJeff Roberson * This code is here to limit the number of simultaneous bucket fills 1414a553d4b8SJeff Roberson * for any given zone to the number of per cpu caches in this zone. This 1415a553d4b8SJeff Roberson * is done so that we don't allocate more memory than we really need. 14168355f576SJeff Roberson */ 1417a553d4b8SJeff Roberson 1418a553d4b8SJeff Roberson if (bucket) { 1419a553d4b8SJeff Roberson #ifdef SMP 1420bce97791SJeff Roberson if (zone->uz_fills >= mp_ncpus) { 1421a553d4b8SJeff Roberson #else 1422bce97791SJeff Roberson if (zone->uz_fills > 1) { 14238355f576SJeff Roberson #endif 1424bce97791SJeff Roberson ZONE_UNLOCK(zone); 1425a553d4b8SJeff Roberson return (NULL); 1426bce97791SJeff Roberson } 1427a553d4b8SJeff Roberson 1428a553d4b8SJeff Roberson zone->uz_fills++; 14298355f576SJeff Roberson } 14308355f576SJeff Roberson 14318355f576SJeff Roberson new_slab: 14328355f576SJeff Roberson 14338355f576SJeff Roberson /* Find a slab with some space */ 14348355f576SJeff Roberson if (zone->uz_free) { 14358355f576SJeff Roberson if (!LIST_EMPTY(&zone->uz_part_slab)) { 14368355f576SJeff Roberson slab = LIST_FIRST(&zone->uz_part_slab); 14378355f576SJeff Roberson } else { 14388355f576SJeff Roberson slab = LIST_FIRST(&zone->uz_free_slab); 14398355f576SJeff Roberson LIST_REMOVE(slab, us_link); 14408355f576SJeff Roberson LIST_INSERT_HEAD(&zone->uz_part_slab, slab, us_link); 14418355f576SJeff Roberson } 14428355f576SJeff Roberson } else { 14438355f576SJeff Roberson /* 14448355f576SJeff Roberson * This is to prevent us from recursively trying to allocate 14458355f576SJeff Roberson * buckets. The problem is that if an allocation forces us to 14468355f576SJeff Roberson * grab a new bucket we will call page_alloc, which will go off 14478355f576SJeff Roberson * and cause the vm to allocate vm_map_entries. If we need new 14488355f576SJeff Roberson * buckets there too we will recurse in kmem_alloc and bad 14498355f576SJeff Roberson * things happen. So instead we return a NULL bucket, and make 14508355f576SJeff Roberson * the code that allocates buckets smart enough to deal with it */ 14518355f576SJeff Roberson if (zone == bucketzone && zone->uz_recurse != 0) { 14528355f576SJeff Roberson ZONE_UNLOCK(zone); 14538355f576SJeff Roberson return (NULL); 14548355f576SJeff Roberson } 1455af7f9b97SJeff Roberson while (zone->uz_maxpages && 1456af7f9b97SJeff Roberson zone->uz_pages >= zone->uz_maxpages) { 1457af7f9b97SJeff Roberson zone->uz_flags |= UMA_ZFLAG_FULL; 1458af7f9b97SJeff Roberson 1459af7f9b97SJeff Roberson if (wait & M_WAITOK) 1460af7f9b97SJeff Roberson msleep(zone, &zone->uz_lock, PVM, "zonelimit", 0); 1461af7f9b97SJeff Roberson else 1462af7f9b97SJeff Roberson goto alloc_fail; 1463af7f9b97SJeff Roberson 1464af7f9b97SJeff Roberson goto new_slab; 1465af7f9b97SJeff Roberson } 1466af7f9b97SJeff Roberson 14678355f576SJeff Roberson zone->uz_recurse++; 14688355f576SJeff Roberson slab = slab_zalloc(zone, wait); 14698355f576SJeff Roberson zone->uz_recurse--; 14708355f576SJeff Roberson /* 1471af7f9b97SJeff Roberson * We might not have been able to get a slab but another cpu 1472af7f9b97SJeff Roberson * could have while we were unlocked. If we did get a slab put 1473af7f9b97SJeff Roberson * it on the partially used slab list. If not check the free 1474af7f9b97SJeff Roberson * count and restart or fail accordingly. 14758355f576SJeff Roberson */ 1476af7f9b97SJeff Roberson if (slab) 1477af7f9b97SJeff Roberson LIST_INSERT_HEAD(&zone->uz_part_slab, slab, us_link); 1478af7f9b97SJeff Roberson else if (zone->uz_free == 0) 1479af7f9b97SJeff Roberson goto alloc_fail; 1480a553d4b8SJeff Roberson else 1481af7f9b97SJeff Roberson goto new_slab; 14828355f576SJeff Roberson } 1483a553d4b8SJeff Roberson /* 1484a553d4b8SJeff Roberson * If this is our first time though put this guy on the list. 1485a553d4b8SJeff Roberson */ 1486a553d4b8SJeff Roberson if (bucket != NULL && bucket->ub_ptr == -1) 1487a553d4b8SJeff Roberson LIST_INSERT_HEAD(&zone->uz_full_bucket, 1488a553d4b8SJeff Roberson bucket, ub_link); 1489a553d4b8SJeff Roberson 14908355f576SJeff Roberson 14918355f576SJeff Roberson while (slab->us_freecount) { 14928355f576SJeff Roberson freei = slab->us_firstfree; 14938355f576SJeff Roberson slab->us_firstfree = slab->us_freelist[freei]; 14948355f576SJeff Roberson #ifdef INVARIANTS 14958355f576SJeff Roberson slab->us_freelist[freei] = 255; 14968355f576SJeff Roberson #endif 14978355f576SJeff Roberson slab->us_freecount--; 14988355f576SJeff Roberson zone->uz_free--; 14998355f576SJeff Roberson item = slab->us_data + (zone->uz_rsize * freei); 15008355f576SJeff Roberson 1501a553d4b8SJeff Roberson if (bucket == NULL) { 15028355f576SJeff Roberson zone->uz_allocs++; 15038355f576SJeff Roberson break; 15048355f576SJeff Roberson } 15058355f576SJeff Roberson bucket->ub_bucket[++bucket->ub_ptr] = item; 15068355f576SJeff Roberson 15078355f576SJeff Roberson /* Don't overfill the bucket! */ 1508a553d4b8SJeff Roberson if (bucket->ub_ptr == zone->uz_count) 15098355f576SJeff Roberson break; 15108355f576SJeff Roberson } 15118355f576SJeff Roberson 15128355f576SJeff Roberson /* Move this slab to the full list */ 15138355f576SJeff Roberson if (slab->us_freecount == 0) { 15148355f576SJeff Roberson LIST_REMOVE(slab, us_link); 15158355f576SJeff Roberson LIST_INSERT_HEAD(&zone->uz_full_slab, slab, us_link); 15168355f576SJeff Roberson } 15178355f576SJeff Roberson 1518a553d4b8SJeff Roberson if (bucket != NULL) { 15198355f576SJeff Roberson /* Try to keep the buckets totally full, but don't block */ 1520a553d4b8SJeff Roberson if (bucket->ub_ptr < zone->uz_count) { 15218355f576SJeff Roberson wait = M_NOWAIT; 15228355f576SJeff Roberson goto new_slab; 1523a553d4b8SJeff Roberson } else 1524a553d4b8SJeff Roberson zone->uz_fills--; 15258355f576SJeff Roberson } 15268355f576SJeff Roberson 15278355f576SJeff Roberson ZONE_UNLOCK(zone); 15288355f576SJeff Roberson 15298355f576SJeff Roberson /* Only construct at this time if we're not filling a bucket */ 1530a553d4b8SJeff Roberson if (bucket == NULL && zone->uz_ctor != NULL) 15318355f576SJeff Roberson zone->uz_ctor(item, zone->uz_size, udata); 15328355f576SJeff Roberson 15338355f576SJeff Roberson return (item); 1534af7f9b97SJeff Roberson 1535af7f9b97SJeff Roberson alloc_fail: 1536af7f9b97SJeff Roberson if (bucket != NULL) 1537af7f9b97SJeff Roberson zone->uz_fills--; 1538af7f9b97SJeff Roberson ZONE_UNLOCK(zone); 1539af7f9b97SJeff Roberson 1540af7f9b97SJeff Roberson if (bucket != NULL && bucket->ub_ptr != -1) 1541af7f9b97SJeff Roberson return (bucket); 1542af7f9b97SJeff Roberson 1543af7f9b97SJeff Roberson return (NULL); 15448355f576SJeff Roberson } 15458355f576SJeff Roberson 15468355f576SJeff Roberson /* See uma.h */ 15478355f576SJeff Roberson void 15488355f576SJeff Roberson uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 15498355f576SJeff Roberson { 15508355f576SJeff Roberson uma_cache_t cache; 15518355f576SJeff Roberson uma_bucket_t bucket; 15528355f576SJeff Roberson int cpu; 15538355f576SJeff Roberson 15548355f576SJeff Roberson /* This is the fast path free */ 15558355f576SJeff Roberson #ifdef UMA_DEBUG_ALLOC_1 15568355f576SJeff Roberson printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone); 15578355f576SJeff Roberson #endif 1558af7f9b97SJeff Roberson /* 1559af7f9b97SJeff Roberson * The race here is acceptable. If we miss it we'll just have to wait 1560af7f9b97SJeff Roberson * a little longer for the limits to be reset. 1561af7f9b97SJeff Roberson */ 1562af7f9b97SJeff Roberson 1563af7f9b97SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_FULL) 1564af7f9b97SJeff Roberson goto zfree_internal; 1565af7f9b97SJeff Roberson 1566a553d4b8SJeff Roberson zfree_restart: 15678355f576SJeff Roberson cpu = PCPU_GET(cpuid); 15688355f576SJeff Roberson CPU_LOCK(zone, cpu); 15698355f576SJeff Roberson cache = &zone->uz_cpu[cpu]; 15708355f576SJeff Roberson 15718355f576SJeff Roberson zfree_start: 15728355f576SJeff Roberson bucket = cache->uc_freebucket; 15738355f576SJeff Roberson 15748355f576SJeff Roberson if (bucket) { 1575a553d4b8SJeff Roberson /* 1576a553d4b8SJeff Roberson * Do we have room in our bucket? It is OK for this uz count 1577a553d4b8SJeff Roberson * check to be slightly out of sync. 1578a553d4b8SJeff Roberson */ 1579a553d4b8SJeff Roberson 1580a553d4b8SJeff Roberson if (bucket->ub_ptr < zone->uz_count) { 15818355f576SJeff Roberson bucket->ub_ptr++; 15828355f576SJeff Roberson KASSERT(bucket->ub_bucket[bucket->ub_ptr] == NULL, 15838355f576SJeff Roberson ("uma_zfree: Freeing to non free bucket index.")); 15848355f576SJeff Roberson bucket->ub_bucket[bucket->ub_ptr] = item; 15858355f576SJeff Roberson if (zone->uz_dtor) 15868355f576SJeff Roberson zone->uz_dtor(item, zone->uz_size, udata); 1587605cbd6aSJeff Roberson CPU_UNLOCK(zone, cpu); 15888355f576SJeff Roberson return; 15898355f576SJeff Roberson } else if (cache->uc_allocbucket) { 15908355f576SJeff Roberson #ifdef UMA_DEBUG_ALLOC 15918355f576SJeff Roberson printf("uma_zfree: Swapping buckets.\n"); 15928355f576SJeff Roberson #endif 15938355f576SJeff Roberson /* 15948355f576SJeff Roberson * We have run out of space in our freebucket. 15958355f576SJeff Roberson * See if we can switch with our alloc bucket. 15968355f576SJeff Roberson */ 15978355f576SJeff Roberson if (cache->uc_allocbucket->ub_ptr < 15988355f576SJeff Roberson cache->uc_freebucket->ub_ptr) { 15998355f576SJeff Roberson uma_bucket_t swap; 16008355f576SJeff Roberson 16018355f576SJeff Roberson swap = cache->uc_freebucket; 16028355f576SJeff Roberson cache->uc_freebucket = cache->uc_allocbucket; 16038355f576SJeff Roberson cache->uc_allocbucket = swap; 16048355f576SJeff Roberson 16058355f576SJeff Roberson goto zfree_start; 16068355f576SJeff Roberson } 16078355f576SJeff Roberson } 16088355f576SJeff Roberson } 16098355f576SJeff Roberson 16108355f576SJeff Roberson /* 1611a553d4b8SJeff Roberson * We can get here for two reasons: 16128355f576SJeff Roberson * 16138355f576SJeff Roberson * 1) The buckets are NULL 1614a553d4b8SJeff Roberson * 2) The alloc and free buckets are both somewhat full. 16158355f576SJeff Roberson * 16168355f576SJeff Roberson */ 16178355f576SJeff Roberson 16188355f576SJeff Roberson ZONE_LOCK(zone); 16198355f576SJeff Roberson 16208355f576SJeff Roberson bucket = cache->uc_freebucket; 16218355f576SJeff Roberson cache->uc_freebucket = NULL; 16228355f576SJeff Roberson 16238355f576SJeff Roberson /* Can we throw this on the zone full list? */ 16248355f576SJeff Roberson if (bucket != NULL) { 16258355f576SJeff Roberson #ifdef UMA_DEBUG_ALLOC 16268355f576SJeff Roberson printf("uma_zfree: Putting old bucket on the free list.\n"); 16278355f576SJeff Roberson #endif 16288355f576SJeff Roberson /* ub_ptr is pointing to the last free item */ 16298355f576SJeff Roberson KASSERT(bucket->ub_ptr != -1, 16308355f576SJeff Roberson ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 16318355f576SJeff Roberson LIST_INSERT_HEAD(&zone->uz_full_bucket, 16328355f576SJeff Roberson bucket, ub_link); 16338355f576SJeff Roberson } 1634a553d4b8SJeff Roberson if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) { 1635a553d4b8SJeff Roberson LIST_REMOVE(bucket, ub_link); 1636a553d4b8SJeff Roberson ZONE_UNLOCK(zone); 1637a553d4b8SJeff Roberson cache->uc_freebucket = bucket; 1638a553d4b8SJeff Roberson goto zfree_start; 1639a553d4b8SJeff Roberson } 1640a553d4b8SJeff Roberson /* We're done with this CPU now */ 1641a553d4b8SJeff Roberson CPU_UNLOCK(zone, cpu); 1642a553d4b8SJeff Roberson 1643a553d4b8SJeff Roberson /* And the zone.. */ 1644a553d4b8SJeff Roberson ZONE_UNLOCK(zone); 1645a553d4b8SJeff Roberson 16468355f576SJeff Roberson #ifdef UMA_DEBUG_ALLOC 16478355f576SJeff Roberson printf("uma_zfree: Allocating new free bucket.\n"); 16488355f576SJeff Roberson #endif 16498355f576SJeff Roberson bucket = uma_zalloc_internal(bucketzone, 1650a553d4b8SJeff Roberson NULL, M_NOWAIT, NULL); 16518355f576SJeff Roberson if (bucket) { 16528355f576SJeff Roberson #ifdef INVARIANTS 16538355f576SJeff Roberson bzero(bucket, bucketzone->uz_size); 16548355f576SJeff Roberson #endif 16558355f576SJeff Roberson bucket->ub_ptr = -1; 1656a553d4b8SJeff Roberson ZONE_LOCK(zone); 1657a553d4b8SJeff Roberson LIST_INSERT_HEAD(&zone->uz_free_bucket, 1658a553d4b8SJeff Roberson bucket, ub_link); 16598355f576SJeff Roberson ZONE_UNLOCK(zone); 1660a553d4b8SJeff Roberson goto zfree_restart; 16618355f576SJeff Roberson } 16628355f576SJeff Roberson 1663a553d4b8SJeff Roberson /* 1664a553d4b8SJeff Roberson * If nothing else caught this, we'll just do an internal free. 1665a553d4b8SJeff Roberson */ 16668355f576SJeff Roberson 1667af7f9b97SJeff Roberson zfree_internal: 1668af7f9b97SJeff Roberson 16698355f576SJeff Roberson uma_zfree_internal(zone, item, udata, 0); 16708355f576SJeff Roberson 16718355f576SJeff Roberson return; 16728355f576SJeff Roberson 16738355f576SJeff Roberson } 16748355f576SJeff Roberson 16758355f576SJeff Roberson /* 16768355f576SJeff Roberson * Frees an item to an INTERNAL zone or allocates a free bucket 16778355f576SJeff Roberson * 16788355f576SJeff Roberson * Arguments: 16798355f576SJeff Roberson * zone The zone to free to 16808355f576SJeff Roberson * item The item we're freeing 16818355f576SJeff Roberson * udata User supplied data for the dtor 16828355f576SJeff Roberson * skip Skip the dtor, it was done in uma_zfree_arg 16838355f576SJeff Roberson */ 16848355f576SJeff Roberson 16858355f576SJeff Roberson static void 16868355f576SJeff Roberson uma_zfree_internal(uma_zone_t zone, void *item, void *udata, int skip) 16878355f576SJeff Roberson { 16888355f576SJeff Roberson uma_slab_t slab; 16898355f576SJeff Roberson u_int8_t *mem; 16908355f576SJeff Roberson u_int8_t freei; 16918355f576SJeff Roberson 16928355f576SJeff Roberson ZONE_LOCK(zone); 16938355f576SJeff Roberson 16948355f576SJeff Roberson if (!(zone->uz_flags & UMA_ZFLAG_MALLOC)) { 16958355f576SJeff Roberson mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK)); 16968355f576SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_OFFPAGE) 16978355f576SJeff Roberson slab = hash_sfind(&zone->uz_hash, mem); 16988355f576SJeff Roberson else { 16998355f576SJeff Roberson mem += zone->uz_pgoff; 17008355f576SJeff Roberson slab = (uma_slab_t)mem; 17018355f576SJeff Roberson } 17028355f576SJeff Roberson } else { 17038355f576SJeff Roberson slab = (uma_slab_t)udata; 17048355f576SJeff Roberson } 17058355f576SJeff Roberson 17068355f576SJeff Roberson /* Do we need to remove from any lists? */ 17078355f576SJeff Roberson if (slab->us_freecount+1 == zone->uz_ipers) { 17088355f576SJeff Roberson LIST_REMOVE(slab, us_link); 17098355f576SJeff Roberson LIST_INSERT_HEAD(&zone->uz_free_slab, slab, us_link); 17108355f576SJeff Roberson } else if (slab->us_freecount == 0) { 17118355f576SJeff Roberson LIST_REMOVE(slab, us_link); 17128355f576SJeff Roberson LIST_INSERT_HEAD(&zone->uz_part_slab, slab, us_link); 17138355f576SJeff Roberson } 17148355f576SJeff Roberson 17158355f576SJeff Roberson /* Slab management stuff */ 17168355f576SJeff Roberson freei = ((unsigned long)item - (unsigned long)slab->us_data) 17178355f576SJeff Roberson / zone->uz_rsize; 17188355f576SJeff Roberson #ifdef INVARIANTS 17198355f576SJeff Roberson if (((freei * zone->uz_rsize) + slab->us_data) != item) 17208355f576SJeff Roberson panic("zone: %s(%p) slab %p freed address %p unaligned.\n", 17218355f576SJeff Roberson zone->uz_name, zone, slab, item); 17228355f576SJeff Roberson if (freei >= zone->uz_ipers) 17238355f576SJeff Roberson panic("zone: %s(%p) slab %p freelist %i out of range 0-%d\n", 17248355f576SJeff Roberson zone->uz_name, zone, slab, freei, zone->uz_ipers-1); 17258355f576SJeff Roberson 17268355f576SJeff Roberson if (slab->us_freelist[freei] != 255) { 17278355f576SJeff Roberson printf("Slab at %p, freei %d = %d.\n", 17288355f576SJeff Roberson slab, freei, slab->us_freelist[freei]); 17298355f576SJeff Roberson panic("Duplicate free of item %p from zone %p(%s)\n", 17308355f576SJeff Roberson item, zone, zone->uz_name); 17318355f576SJeff Roberson } 17328355f576SJeff Roberson #endif 17338355f576SJeff Roberson slab->us_freelist[freei] = slab->us_firstfree; 17348355f576SJeff Roberson slab->us_firstfree = freei; 17358355f576SJeff Roberson slab->us_freecount++; 17368355f576SJeff Roberson 17378355f576SJeff Roberson /* Zone statistics */ 17388355f576SJeff Roberson zone->uz_free++; 17398355f576SJeff Roberson 17408355f576SJeff Roberson if (!skip && zone->uz_dtor) 17418355f576SJeff Roberson zone->uz_dtor(item, zone->uz_size, udata); 1742605cbd6aSJeff Roberson 1743af7f9b97SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_FULL) { 1744af7f9b97SJeff Roberson if (zone->uz_pages < zone->uz_maxpages) 1745af7f9b97SJeff Roberson zone->uz_flags &= ~UMA_ZFLAG_FULL; 1746af7f9b97SJeff Roberson 1747af7f9b97SJeff Roberson /* We can handle one more allocation */ 1748af7f9b97SJeff Roberson wakeup_one(&zone); 1749af7f9b97SJeff Roberson } 1750af7f9b97SJeff Roberson 1751605cbd6aSJeff Roberson ZONE_UNLOCK(zone); 17528355f576SJeff Roberson } 17538355f576SJeff Roberson 17548355f576SJeff Roberson /* See uma.h */ 17558355f576SJeff Roberson void 1756736ee590SJeff Roberson uma_zone_set_max(uma_zone_t zone, int nitems) 1757736ee590SJeff Roberson { 1758736ee590SJeff Roberson ZONE_LOCK(zone); 1759736ee590SJeff Roberson if (zone->uz_ppera > 1) 1760af7f9b97SJeff Roberson zone->uz_maxpages = nitems * zone->uz_ppera; 1761736ee590SJeff Roberson else 1762736ee590SJeff Roberson zone->uz_maxpages = nitems / zone->uz_ipers; 1763736ee590SJeff Roberson ZONE_UNLOCK(zone); 1764736ee590SJeff Roberson } 1765736ee590SJeff Roberson 1766736ee590SJeff Roberson /* See uma.h */ 1767736ee590SJeff Roberson void 17688355f576SJeff Roberson uma_zone_set_freef(uma_zone_t zone, uma_free freef) 17698355f576SJeff Roberson { 17708355f576SJeff Roberson ZONE_LOCK(zone); 17718355f576SJeff Roberson 17728355f576SJeff Roberson zone->uz_freef = freef; 17738355f576SJeff Roberson 17748355f576SJeff Roberson ZONE_UNLOCK(zone); 17758355f576SJeff Roberson } 17768355f576SJeff Roberson 17778355f576SJeff Roberson /* See uma.h */ 17788355f576SJeff Roberson void 17798355f576SJeff Roberson uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 17808355f576SJeff Roberson { 17818355f576SJeff Roberson ZONE_LOCK(zone); 17828355f576SJeff Roberson 17838355f576SJeff Roberson zone->uz_flags |= UMA_ZFLAG_PRIVALLOC; 17848355f576SJeff Roberson zone->uz_allocf = allocf; 17858355f576SJeff Roberson 17868355f576SJeff Roberson ZONE_UNLOCK(zone); 17878355f576SJeff Roberson } 17888355f576SJeff Roberson 17898355f576SJeff Roberson /* See uma.h */ 17908355f576SJeff Roberson int 17918355f576SJeff Roberson uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count) 17928355f576SJeff Roberson { 17938355f576SJeff Roberson int pages; 17948355f576SJeff Roberson vm_offset_t kva; 17958355f576SJeff Roberson 17968355f576SJeff Roberson mtx_lock(&Giant); 17978355f576SJeff Roberson 17988355f576SJeff Roberson pages = count / zone->uz_ipers; 17998355f576SJeff Roberson 18008355f576SJeff Roberson if (pages * zone->uz_ipers < count) 18018355f576SJeff Roberson pages++; 1802a553d4b8SJeff Roberson 18038355f576SJeff Roberson kva = kmem_alloc_pageable(kernel_map, pages * UMA_SLAB_SIZE); 18048355f576SJeff Roberson 1805a553d4b8SJeff Roberson if (kva == 0) { 1806a553d4b8SJeff Roberson mtx_unlock(&Giant); 18078355f576SJeff Roberson return (0); 18088355f576SJeff Roberson } 18098355f576SJeff Roberson 18108355f576SJeff Roberson 1811a553d4b8SJeff Roberson if (obj == NULL) 1812a553d4b8SJeff Roberson obj = vm_object_allocate(OBJT_DEFAULT, 18138355f576SJeff Roberson zone->uz_maxpages); 18148355f576SJeff Roberson else 18158355f576SJeff Roberson _vm_object_allocate(OBJT_DEFAULT, 1816a553d4b8SJeff Roberson zone->uz_maxpages, obj); 1817a553d4b8SJeff Roberson 1818a553d4b8SJeff Roberson ZONE_LOCK(zone); 1819a553d4b8SJeff Roberson zone->uz_kva = kva; 1820a553d4b8SJeff Roberson zone->uz_obj = obj; 1821a553d4b8SJeff Roberson zone->uz_maxpages = pages; 18228355f576SJeff Roberson 18238355f576SJeff Roberson zone->uz_allocf = obj_alloc; 18248355f576SJeff Roberson zone->uz_flags |= UMA_ZFLAG_NOFREE | UMA_ZFLAG_PRIVALLOC; 18258355f576SJeff Roberson 18268355f576SJeff Roberson ZONE_UNLOCK(zone); 1827a553d4b8SJeff Roberson mtx_unlock(&Giant); 18288355f576SJeff Roberson 18298355f576SJeff Roberson return (1); 18308355f576SJeff Roberson } 18318355f576SJeff Roberson 18328355f576SJeff Roberson /* See uma.h */ 18338355f576SJeff Roberson void 18348355f576SJeff Roberson uma_prealloc(uma_zone_t zone, int items) 18358355f576SJeff Roberson { 18368355f576SJeff Roberson int slabs; 18378355f576SJeff Roberson uma_slab_t slab; 18388355f576SJeff Roberson 18398355f576SJeff Roberson ZONE_LOCK(zone); 18408355f576SJeff Roberson slabs = items / zone->uz_ipers; 18418355f576SJeff Roberson if (slabs * zone->uz_ipers < items) 18428355f576SJeff Roberson slabs++; 18438355f576SJeff Roberson 18448355f576SJeff Roberson while (slabs > 0) { 18458355f576SJeff Roberson slab = slab_zalloc(zone, M_WAITOK); 18468355f576SJeff Roberson LIST_INSERT_HEAD(&zone->uz_free_slab, slab, us_link); 18478355f576SJeff Roberson slabs--; 18488355f576SJeff Roberson } 18498355f576SJeff Roberson ZONE_UNLOCK(zone); 18508355f576SJeff Roberson } 18518355f576SJeff Roberson 18528355f576SJeff Roberson /* See uma.h */ 18538355f576SJeff Roberson void 18548355f576SJeff Roberson uma_reclaim(void) 18558355f576SJeff Roberson { 18568355f576SJeff Roberson /* 18578355f576SJeff Roberson * You might think that the delay below would improve performance since 18588355f576SJeff Roberson * the allocator will give away memory that it may ask for immediately. 18598355f576SJeff Roberson * Really, it makes things worse, since cpu cycles are so much cheaper 18608355f576SJeff Roberson * than disk activity. 18618355f576SJeff Roberson */ 18628355f576SJeff Roberson #if 0 18638355f576SJeff Roberson static struct timeval tv = {0}; 18648355f576SJeff Roberson struct timeval now; 18658355f576SJeff Roberson getmicrouptime(&now); 18668355f576SJeff Roberson if (now.tv_sec > tv.tv_sec + 30) 18678355f576SJeff Roberson tv = now; 18688355f576SJeff Roberson else 18698355f576SJeff Roberson return; 18708355f576SJeff Roberson #endif 18718355f576SJeff Roberson #ifdef UMA_DEBUG 18728355f576SJeff Roberson printf("UMA: vm asked us to release pages!\n"); 18738355f576SJeff Roberson #endif 187486bbae32SJeff Roberson bucket_enable(); 18758355f576SJeff Roberson zone_foreach(zone_drain); 18768355f576SJeff Roberson 18778355f576SJeff Roberson /* 18788355f576SJeff Roberson * Some slabs may have been freed but this zone will be visited early 18798355f576SJeff Roberson * we visit again so that we can free pages that are empty once other 18808355f576SJeff Roberson * zones are drained. We have to do the same for buckets. 18818355f576SJeff Roberson */ 18828355f576SJeff Roberson zone_drain(slabzone); 18838355f576SJeff Roberson zone_drain(bucketzone); 18848355f576SJeff Roberson } 18858355f576SJeff Roberson 18868355f576SJeff Roberson void * 18878355f576SJeff Roberson uma_large_malloc(int size, int wait) 18888355f576SJeff Roberson { 18898355f576SJeff Roberson void *mem; 18908355f576SJeff Roberson uma_slab_t slab; 18918355f576SJeff Roberson u_int8_t flags; 18928355f576SJeff Roberson 1893a553d4b8SJeff Roberson slab = uma_zalloc_internal(slabzone, NULL, wait, NULL); 18948355f576SJeff Roberson if (slab == NULL) 18958355f576SJeff Roberson return (NULL); 18968355f576SJeff Roberson 18978355f576SJeff Roberson mem = page_alloc(NULL, size, &flags, wait); 18988355f576SJeff Roberson if (mem) { 18998355f576SJeff Roberson slab->us_data = mem; 19008355f576SJeff Roberson slab->us_flags = flags | UMA_SLAB_MALLOC; 19018355f576SJeff Roberson slab->us_size = size; 19028355f576SJeff Roberson UMA_HASH_INSERT(mallochash, slab, mem); 19038355f576SJeff Roberson } else { 19048355f576SJeff Roberson uma_zfree_internal(slabzone, slab, NULL, 0); 19058355f576SJeff Roberson } 19068355f576SJeff Roberson 19078355f576SJeff Roberson 19088355f576SJeff Roberson return (mem); 19098355f576SJeff Roberson } 19108355f576SJeff Roberson 19118355f576SJeff Roberson void 19128355f576SJeff Roberson uma_large_free(uma_slab_t slab) 19138355f576SJeff Roberson { 19148355f576SJeff Roberson UMA_HASH_REMOVE(mallochash, slab, slab->us_data); 19158355f576SJeff Roberson page_free(slab->us_data, slab->us_size, slab->us_flags); 19168355f576SJeff Roberson uma_zfree_internal(slabzone, slab, NULL, 0); 19178355f576SJeff Roberson } 19188355f576SJeff Roberson 19198355f576SJeff Roberson void 19208355f576SJeff Roberson uma_print_stats(void) 19218355f576SJeff Roberson { 19228355f576SJeff Roberson zone_foreach(uma_print_zone); 19238355f576SJeff Roberson } 19248355f576SJeff Roberson 19258355f576SJeff Roberson void 19268355f576SJeff Roberson uma_print_zone(uma_zone_t zone) 19278355f576SJeff Roberson { 19288355f576SJeff Roberson printf("%s(%p) size %d(%d) flags %d ipers %d ppera %d out %d free %d\n", 19298355f576SJeff Roberson zone->uz_name, zone, zone->uz_size, zone->uz_rsize, zone->uz_flags, 19308355f576SJeff Roberson zone->uz_ipers, zone->uz_ppera, 19318355f576SJeff Roberson (zone->uz_ipers * zone->uz_pages) - zone->uz_free, zone->uz_free); 19328355f576SJeff Roberson } 19338355f576SJeff Roberson 19348355f576SJeff Roberson /* 19358355f576SJeff Roberson * Sysctl handler for vm.zone 19368355f576SJeff Roberson * 19378355f576SJeff Roberson * stolen from vm_zone.c 19388355f576SJeff Roberson */ 19398355f576SJeff Roberson static int 19408355f576SJeff Roberson sysctl_vm_zone(SYSCTL_HANDLER_ARGS) 19418355f576SJeff Roberson { 19428355f576SJeff Roberson int error, len, cnt; 19438355f576SJeff Roberson const int linesize = 128; /* conservative */ 19448355f576SJeff Roberson int totalfree; 19458355f576SJeff Roberson char *tmpbuf, *offset; 19468355f576SJeff Roberson uma_zone_t z; 19478355f576SJeff Roberson char *p; 19488355f576SJeff Roberson 19498355f576SJeff Roberson cnt = 0; 19508355f576SJeff Roberson LIST_FOREACH(z, &uma_zones, uz_link) 19518355f576SJeff Roberson cnt++; 19528355f576SJeff Roberson MALLOC(tmpbuf, char *, (cnt == 0 ? 1 : cnt) * linesize, 19538355f576SJeff Roberson M_TEMP, M_WAITOK); 19548355f576SJeff Roberson len = snprintf(tmpbuf, linesize, 19558355f576SJeff Roberson "\nITEM SIZE LIMIT USED FREE REQUESTS\n\n"); 19568355f576SJeff Roberson if (cnt == 0) 19578355f576SJeff Roberson tmpbuf[len - 1] = '\0'; 19588355f576SJeff Roberson error = SYSCTL_OUT(req, tmpbuf, cnt == 0 ? len-1 : len); 19598355f576SJeff Roberson if (error || cnt == 0) 19608355f576SJeff Roberson goto out; 19618355f576SJeff Roberson offset = tmpbuf; 1962f4af24d5SJeff Roberson mtx_lock(&uma_mtx); 19638355f576SJeff Roberson LIST_FOREACH(z, &uma_zones, uz_link) { 19648355f576SJeff Roberson if (cnt == 0) /* list may have changed size */ 19658355f576SJeff Roberson break; 19668355f576SJeff Roberson ZONE_LOCK(z); 19678355f576SJeff Roberson totalfree = z->uz_free + z->uz_cachefree; 19688355f576SJeff Roberson len = snprintf(offset, linesize, 19698355f576SJeff Roberson "%-12.12s %6.6u, %8.8u, %6.6u, %6.6u, %8.8llu\n", 19708355f576SJeff Roberson z->uz_name, z->uz_size, 19718355f576SJeff Roberson z->uz_maxpages * z->uz_ipers, 19728355f576SJeff Roberson (z->uz_ipers * (z->uz_pages / z->uz_ppera)) - totalfree, 19738355f576SJeff Roberson totalfree, 19748355f576SJeff Roberson (unsigned long long)z->uz_allocs); 19758355f576SJeff Roberson ZONE_UNLOCK(z); 19768355f576SJeff Roberson for (p = offset + 12; p > offset && *p == ' '; --p) 19778355f576SJeff Roberson /* nothing */ ; 19788355f576SJeff Roberson p[1] = ':'; 19798355f576SJeff Roberson cnt--; 19808355f576SJeff Roberson offset += len; 19818355f576SJeff Roberson } 1982f4af24d5SJeff Roberson mtx_unlock(&uma_mtx); 19838355f576SJeff Roberson *offset++ = '\0'; 19848355f576SJeff Roberson error = SYSCTL_OUT(req, tmpbuf, offset - tmpbuf); 19858355f576SJeff Roberson out: 19868355f576SJeff Roberson FREE(tmpbuf, M_TEMP); 19878355f576SJeff Roberson return (error); 19888355f576SJeff Roberson } 1989