18355f576SJeff Roberson /* 2f461cf22SJeff Roberson * Copyright (c) 2002, Jeffrey Roberson <jeff@freebsd.org> 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 278355f576SJeff Roberson /* 288355f576SJeff Roberson * uma_core.c Implementation of the Universal Memory allocator 298355f576SJeff Roberson * 308355f576SJeff Roberson * This allocator is intended to replace the multitude of similar object caches 318355f576SJeff Roberson * in the standard FreeBSD kernel. The intent is to be flexible as well as 328355f576SJeff Roberson * effecient. A primary design goal is to return unused memory to the rest of 338355f576SJeff Roberson * the system. This will make the system as a whole more flexible due to the 348355f576SJeff Roberson * ability to move memory to subsystems which most need it instead of leaving 358355f576SJeff Roberson * pools of reserved memory unused. 368355f576SJeff Roberson * 378355f576SJeff Roberson * The basic ideas stem from similar slab/zone based allocators whose algorithms 388355f576SJeff Roberson * are well known. 398355f576SJeff Roberson * 408355f576SJeff Roberson */ 418355f576SJeff Roberson 428355f576SJeff Roberson /* 438355f576SJeff Roberson * TODO: 448355f576SJeff Roberson * - Improve memory usage for large allocations 458355f576SJeff Roberson * - Investigate cache size adjustments 468355f576SJeff Roberson */ 478355f576SJeff Roberson 48874651b1SDavid E. O'Brien #include <sys/cdefs.h> 49874651b1SDavid E. O'Brien __FBSDID("$FreeBSD$"); 50874651b1SDavid E. O'Brien 518355f576SJeff Roberson /* I should really use ktr.. */ 528355f576SJeff Roberson /* 538355f576SJeff Roberson #define UMA_DEBUG 1 548355f576SJeff Roberson #define UMA_DEBUG_ALLOC 1 558355f576SJeff Roberson #define UMA_DEBUG_ALLOC_1 1 568355f576SJeff Roberson */ 578355f576SJeff Roberson 588355f576SJeff Roberson #include "opt_param.h" 598355f576SJeff Roberson #include <sys/param.h> 608355f576SJeff Roberson #include <sys/systm.h> 618355f576SJeff Roberson #include <sys/kernel.h> 628355f576SJeff Roberson #include <sys/types.h> 638355f576SJeff Roberson #include <sys/queue.h> 648355f576SJeff Roberson #include <sys/malloc.h> 658355f576SJeff Roberson #include <sys/lock.h> 668355f576SJeff Roberson #include <sys/sysctl.h> 678355f576SJeff Roberson #include <sys/mutex.h> 684c1cc01cSJohn Baldwin #include <sys/proc.h> 698355f576SJeff Roberson #include <sys/smp.h> 7086bbae32SJeff Roberson #include <sys/vmmeter.h> 718522511bSHartmut Brandt #include <sys/mbuf.h> 7286bbae32SJeff Roberson 738355f576SJeff Roberson #include <vm/vm.h> 748355f576SJeff Roberson #include <vm/vm_object.h> 758355f576SJeff Roberson #include <vm/vm_page.h> 768355f576SJeff Roberson #include <vm/vm_param.h> 778355f576SJeff Roberson #include <vm/vm_map.h> 788355f576SJeff Roberson #include <vm/vm_kern.h> 798355f576SJeff Roberson #include <vm/vm_extern.h> 808355f576SJeff Roberson #include <vm/uma.h> 818355f576SJeff Roberson #include <vm/uma_int.h> 82639c9550SJeff Roberson #include <vm/uma_dbg.h> 838355f576SJeff Roberson 8448eea375SJeff Roberson #include <machine/vmparam.h> 8548eea375SJeff Roberson 868355f576SJeff Roberson /* 878355f576SJeff Roberson * This is the zone from which all zones are spawned. The idea is that even 888355f576SJeff Roberson * the zone heads are allocated from the allocator, so we use the bss section 898355f576SJeff Roberson * to bootstrap us. 908355f576SJeff Roberson */ 9186bbae32SJeff Roberson static struct uma_zone masterzone; 9286bbae32SJeff Roberson static uma_zone_t zones = &masterzone; 938355f576SJeff Roberson 948355f576SJeff Roberson /* This is the zone from which all of uma_slab_t's are allocated. */ 958355f576SJeff Roberson static uma_zone_t slabzone; 968355f576SJeff Roberson 978355f576SJeff Roberson /* 988355f576SJeff Roberson * The initial hash tables come out of this zone so they can be allocated 998355f576SJeff Roberson * prior to malloc coming up. 1008355f576SJeff Roberson */ 1018355f576SJeff Roberson static uma_zone_t hashzone; 1028355f576SJeff Roberson 1038355f576SJeff Roberson /* 10486bbae32SJeff Roberson * Are we allowed to allocate buckets? 10586bbae32SJeff Roberson */ 10686bbae32SJeff Roberson static int bucketdisable = 1; 10786bbae32SJeff Roberson 1088355f576SJeff Roberson /* Linked list of all zones in the system */ 1098355f576SJeff Roberson static LIST_HEAD(,uma_zone) uma_zones = LIST_HEAD_INITIALIZER(&uma_zones); 1108355f576SJeff Roberson 1118355f576SJeff Roberson /* This mutex protects the zone list */ 1128355f576SJeff Roberson static struct mtx uma_mtx; 1138355f576SJeff Roberson 114d88797c2SBosko Milekic /* These are the pcpu cache locks */ 115d88797c2SBosko Milekic static struct mtx uma_pcpu_mtx[MAXCPU]; 116d88797c2SBosko Milekic 1178355f576SJeff Roberson /* Linked list of boot time pages */ 1188355f576SJeff Roberson static LIST_HEAD(,uma_slab) uma_boot_pages = 1198355f576SJeff Roberson LIST_HEAD_INITIALIZER(&uma_boot_pages); 1208355f576SJeff Roberson 1218355f576SJeff Roberson /* Count of free boottime pages */ 1228355f576SJeff Roberson static int uma_boot_free = 0; 1238355f576SJeff Roberson 1248355f576SJeff Roberson /* Is the VM done starting up? */ 1258355f576SJeff Roberson static int booted = 0; 1268355f576SJeff Roberson 1278355f576SJeff Roberson /* This is the handle used to schedule our working set calculator */ 1288355f576SJeff Roberson static struct callout uma_callout; 1298355f576SJeff Roberson 1308355f576SJeff Roberson /* This is mp_maxid + 1, for use while looping over each cpu */ 1318355f576SJeff Roberson static int maxcpu; 1328355f576SJeff Roberson 1338355f576SJeff Roberson /* 1348355f576SJeff Roberson * This structure is passed as the zone ctor arg so that I don't have to create 1358355f576SJeff Roberson * a special allocation function just for zones. 1368355f576SJeff Roberson */ 1378355f576SJeff Roberson struct uma_zctor_args { 1388355f576SJeff Roberson char *name; 139c3bdc05fSAndrew R. Reiter size_t size; 1408355f576SJeff Roberson uma_ctor ctor; 1418355f576SJeff Roberson uma_dtor dtor; 1428355f576SJeff Roberson uma_init uminit; 1438355f576SJeff Roberson uma_fini fini; 1448355f576SJeff Roberson int align; 1458355f576SJeff Roberson u_int16_t flags; 1468355f576SJeff Roberson }; 1478355f576SJeff Roberson 148cae33c14SJeff Roberson struct uma_bucket_zone { 149cae33c14SJeff Roberson uma_zone_t ubz_zone; 150cae33c14SJeff Roberson char *ubz_name; 151cae33c14SJeff Roberson int ubz_entries; 152cae33c14SJeff Roberson }; 153cae33c14SJeff Roberson 154cae33c14SJeff Roberson #define BUCKET_MAX 128 155cae33c14SJeff Roberson 156cae33c14SJeff Roberson struct uma_bucket_zone bucket_zones[] = { 157cae33c14SJeff Roberson { NULL, "16 Bucket", 16 }, 158cae33c14SJeff Roberson { NULL, "32 Bucket", 32 }, 159cae33c14SJeff Roberson { NULL, "64 Bucket", 64 }, 160cae33c14SJeff Roberson { NULL, "128 Bucket", 128 }, 161cae33c14SJeff Roberson { NULL, NULL, 0} 162cae33c14SJeff Roberson }; 163cae33c14SJeff Roberson 164cae33c14SJeff Roberson #define BUCKET_SHIFT 4 165cae33c14SJeff Roberson #define BUCKET_ZONES ((BUCKET_MAX >> BUCKET_SHIFT) + 1) 166cae33c14SJeff Roberson 167cae33c14SJeff Roberson uint8_t bucket_size[BUCKET_ZONES]; 168cae33c14SJeff Roberson 1698355f576SJeff Roberson /* Prototypes.. */ 1708355f576SJeff Roberson 1718355f576SJeff Roberson static void *obj_alloc(uma_zone_t, int, u_int8_t *, int); 1728355f576SJeff Roberson static void *page_alloc(uma_zone_t, int, u_int8_t *, int); 1738355f576SJeff Roberson static void page_free(void *, int, u_int8_t); 1748355f576SJeff Roberson static uma_slab_t slab_zalloc(uma_zone_t, int); 175d56368d7SBosko Milekic static void cache_drain(uma_zone_t, int); 1768355f576SJeff Roberson static void bucket_drain(uma_zone_t, uma_bucket_t); 177d56368d7SBosko Milekic static void zone_drain_common(uma_zone_t, int); 1788355f576SJeff Roberson static void zone_ctor(void *, int, void *); 1799c2cd7e5SJeff Roberson static void zone_dtor(void *, int, void *); 1808355f576SJeff Roberson static void zero_init(void *, int); 1818355f576SJeff Roberson static void zone_small_init(uma_zone_t zone); 1828355f576SJeff Roberson static void zone_large_init(uma_zone_t zone); 1838355f576SJeff Roberson static void zone_foreach(void (*zfunc)(uma_zone_t)); 1848355f576SJeff Roberson static void zone_timeout(uma_zone_t zone); 1850aef6126SJeff Roberson static int hash_alloc(struct uma_hash *); 1860aef6126SJeff Roberson static int hash_expand(struct uma_hash *, struct uma_hash *); 1870aef6126SJeff Roberson static void hash_free(struct uma_hash *hash); 1888355f576SJeff Roberson static void uma_timeout(void *); 1898355f576SJeff Roberson static void uma_startup3(void); 190bbee39c6SJeff Roberson static void *uma_zalloc_internal(uma_zone_t, void *, int); 19186bbae32SJeff Roberson static void uma_zfree_internal(uma_zone_t, void *, void *, int); 19286bbae32SJeff Roberson static void bucket_enable(void); 193cae33c14SJeff Roberson static void bucket_init(void); 194cae33c14SJeff Roberson static uma_bucket_t bucket_alloc(int, int); 195cae33c14SJeff Roberson static void bucket_free(uma_bucket_t); 196cae33c14SJeff Roberson static void bucket_zone_drain(void); 197bbee39c6SJeff Roberson static int uma_zalloc_bucket(uma_zone_t zone, int flags); 198bbee39c6SJeff Roberson static uma_slab_t uma_zone_slab(uma_zone_t zone, int flags); 199bbee39c6SJeff Roberson static void *uma_slab_alloc(uma_zone_t zone, uma_slab_t slab); 200d56368d7SBosko Milekic static __inline void zone_drain(uma_zone_t); 201bbee39c6SJeff Roberson 2028355f576SJeff Roberson void uma_print_zone(uma_zone_t); 2038355f576SJeff Roberson void uma_print_stats(void); 2048355f576SJeff Roberson static int sysctl_vm_zone(SYSCTL_HANDLER_ARGS); 2058355f576SJeff Roberson 2068355f576SJeff Roberson SYSCTL_OID(_vm, OID_AUTO, zone, CTLTYPE_STRING|CTLFLAG_RD, 2078355f576SJeff Roberson NULL, 0, sysctl_vm_zone, "A", "Zone Info"); 2088355f576SJeff Roberson SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); 2098355f576SJeff Roberson 21086bbae32SJeff Roberson /* 21186bbae32SJeff Roberson * This routine checks to see whether or not it's safe to enable buckets. 21286bbae32SJeff Roberson */ 21386bbae32SJeff Roberson 21486bbae32SJeff Roberson static void 21586bbae32SJeff Roberson bucket_enable(void) 21686bbae32SJeff Roberson { 21786bbae32SJeff Roberson if (cnt.v_free_count < cnt.v_free_min) 21886bbae32SJeff Roberson bucketdisable = 1; 21986bbae32SJeff Roberson else 22086bbae32SJeff Roberson bucketdisable = 0; 22186bbae32SJeff Roberson } 22286bbae32SJeff Roberson 223cae33c14SJeff Roberson static void 224cae33c14SJeff Roberson bucket_init(void) 225cae33c14SJeff Roberson { 226cae33c14SJeff Roberson struct uma_bucket_zone *ubz; 227cae33c14SJeff Roberson int i; 228cae33c14SJeff Roberson int j; 229cae33c14SJeff Roberson 230cae33c14SJeff Roberson for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) { 231cae33c14SJeff Roberson int size; 232cae33c14SJeff Roberson 233cae33c14SJeff Roberson ubz = &bucket_zones[j]; 234cae33c14SJeff Roberson size = roundup(sizeof(struct uma_bucket), sizeof(void *)); 235cae33c14SJeff Roberson size += sizeof(void *) * ubz->ubz_entries; 236cae33c14SJeff Roberson ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, 237cae33c14SJeff Roberson NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_INTERNAL); 238cae33c14SJeff Roberson for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT)) 239cae33c14SJeff Roberson bucket_size[i >> BUCKET_SHIFT] = j; 240cae33c14SJeff Roberson } 241cae33c14SJeff Roberson } 242cae33c14SJeff Roberson 243cae33c14SJeff Roberson static uma_bucket_t 244cae33c14SJeff Roberson bucket_alloc(int entries, int bflags) 245cae33c14SJeff Roberson { 246cae33c14SJeff Roberson struct uma_bucket_zone *ubz; 247cae33c14SJeff Roberson uma_bucket_t bucket; 248cae33c14SJeff Roberson int idx; 249cae33c14SJeff Roberson 250cae33c14SJeff Roberson /* 251cae33c14SJeff Roberson * This is to stop us from allocating per cpu buckets while we're 252cae33c14SJeff Roberson * running out of UMA_BOOT_PAGES. Otherwise, we would exhaust the 253cae33c14SJeff Roberson * boot pages. This also prevents us from allocating buckets in 254cae33c14SJeff Roberson * low memory situations. 255cae33c14SJeff Roberson */ 256cae33c14SJeff Roberson 257cae33c14SJeff Roberson if (bucketdisable) 258cae33c14SJeff Roberson return (NULL); 259cae33c14SJeff Roberson idx = howmany(entries, 1 << BUCKET_SHIFT); 260cae33c14SJeff Roberson ubz = &bucket_zones[bucket_size[idx]]; 261cae33c14SJeff Roberson bucket = uma_zalloc_internal(ubz->ubz_zone, NULL, bflags); 262cae33c14SJeff Roberson if (bucket) { 263cae33c14SJeff Roberson #ifdef INVARIANTS 264cae33c14SJeff Roberson bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); 265cae33c14SJeff Roberson #endif 266cae33c14SJeff Roberson bucket->ub_cnt = 0; 267cae33c14SJeff Roberson bucket->ub_entries = ubz->ubz_entries; 268cae33c14SJeff Roberson } 269cae33c14SJeff Roberson 270cae33c14SJeff Roberson return (bucket); 271cae33c14SJeff Roberson } 272cae33c14SJeff Roberson 273cae33c14SJeff Roberson static void 274cae33c14SJeff Roberson bucket_free(uma_bucket_t bucket) 275cae33c14SJeff Roberson { 276cae33c14SJeff Roberson struct uma_bucket_zone *ubz; 277cae33c14SJeff Roberson int idx; 278cae33c14SJeff Roberson 279cae33c14SJeff Roberson idx = howmany(bucket->ub_entries, 1 << BUCKET_SHIFT); 280cae33c14SJeff Roberson ubz = &bucket_zones[bucket_size[idx]]; 281cae33c14SJeff Roberson uma_zfree_internal(ubz->ubz_zone, bucket, NULL, 0); 282cae33c14SJeff Roberson } 283cae33c14SJeff Roberson 284cae33c14SJeff Roberson static void 285cae33c14SJeff Roberson bucket_zone_drain(void) 286cae33c14SJeff Roberson { 287cae33c14SJeff Roberson struct uma_bucket_zone *ubz; 288cae33c14SJeff Roberson 289cae33c14SJeff Roberson for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) 290cae33c14SJeff Roberson zone_drain(ubz->ubz_zone); 291cae33c14SJeff Roberson } 292cae33c14SJeff Roberson 2938355f576SJeff Roberson 2948355f576SJeff Roberson /* 2958355f576SJeff Roberson * Routine called by timeout which is used to fire off some time interval 2968355f576SJeff Roberson * based calculations. (working set, stats, etc.) 2978355f576SJeff Roberson * 2988355f576SJeff Roberson * Arguments: 2998355f576SJeff Roberson * arg Unused 3008355f576SJeff Roberson * 3018355f576SJeff Roberson * Returns: 3028355f576SJeff Roberson * Nothing 3038355f576SJeff Roberson */ 3048355f576SJeff Roberson static void 3058355f576SJeff Roberson uma_timeout(void *unused) 3068355f576SJeff Roberson { 30786bbae32SJeff Roberson bucket_enable(); 3088355f576SJeff Roberson zone_foreach(zone_timeout); 3098355f576SJeff Roberson 3108355f576SJeff Roberson /* Reschedule this event */ 3118355f576SJeff Roberson callout_reset(&uma_callout, UMA_WORKING_TIME * hz, uma_timeout, NULL); 3128355f576SJeff Roberson } 3138355f576SJeff Roberson 3148355f576SJeff Roberson /* 3158355f576SJeff Roberson * Routine to perform timeout driven calculations. This does the working set 3168355f576SJeff Roberson * as well as hash expanding, and per cpu statistics aggregation. 3178355f576SJeff Roberson * 3188355f576SJeff Roberson * Arguments: 3198355f576SJeff Roberson * zone The zone to operate on 3208355f576SJeff Roberson * 3218355f576SJeff Roberson * Returns: 3228355f576SJeff Roberson * Nothing 3238355f576SJeff Roberson */ 3248355f576SJeff Roberson static void 3258355f576SJeff Roberson zone_timeout(uma_zone_t zone) 3268355f576SJeff Roberson { 3278355f576SJeff Roberson uma_cache_t cache; 3288355f576SJeff Roberson u_int64_t alloc; 3298355f576SJeff Roberson int cpu; 3308355f576SJeff Roberson 3318355f576SJeff Roberson alloc = 0; 3328355f576SJeff Roberson 3338355f576SJeff Roberson /* 3348355f576SJeff Roberson * Aggregate per cpu cache statistics back to the zone. 3358355f576SJeff Roberson * 3368355f576SJeff Roberson * I may rewrite this to set a flag in the per cpu cache instead of 3378355f576SJeff Roberson * locking. If the flag is not cleared on the next round I will have 3388355f576SJeff Roberson * to lock and do it here instead so that the statistics don't get too 3398355f576SJeff Roberson * far out of sync. 3408355f576SJeff Roberson */ 3418355f576SJeff Roberson if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) { 3428355f576SJeff Roberson for (cpu = 0; cpu < maxcpu; cpu++) { 3438355f576SJeff Roberson if (CPU_ABSENT(cpu)) 3448355f576SJeff Roberson continue; 345d88797c2SBosko Milekic CPU_LOCK(cpu); 3468355f576SJeff Roberson cache = &zone->uz_cpu[cpu]; 3478355f576SJeff Roberson /* Add them up, and reset */ 3488355f576SJeff Roberson alloc += cache->uc_allocs; 3498355f576SJeff Roberson cache->uc_allocs = 0; 350d88797c2SBosko Milekic CPU_UNLOCK(cpu); 3518355f576SJeff Roberson } 3528355f576SJeff Roberson } 3538355f576SJeff Roberson 3548355f576SJeff Roberson /* Now push these stats back into the zone.. */ 3558355f576SJeff Roberson ZONE_LOCK(zone); 3568355f576SJeff Roberson zone->uz_allocs += alloc; 3578355f576SJeff Roberson 3588355f576SJeff Roberson /* 3598355f576SJeff Roberson * Expand the zone hash table. 3608355f576SJeff Roberson * 3618355f576SJeff Roberson * This is done if the number of slabs is larger than the hash size. 3628355f576SJeff Roberson * What I'm trying to do here is completely reduce collisions. This 3638355f576SJeff Roberson * may be a little aggressive. Should I allow for two collisions max? 3648355f576SJeff Roberson */ 3658355f576SJeff Roberson 36699571dc3SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_HASH && 36799571dc3SJeff Roberson zone->uz_pages / zone->uz_ppera >= zone->uz_hash.uh_hashsize) { 3680aef6126SJeff Roberson struct uma_hash newhash; 3690aef6126SJeff Roberson struct uma_hash oldhash; 3700aef6126SJeff Roberson int ret; 3715300d9ddSJeff Roberson 3720aef6126SJeff Roberson /* 3730aef6126SJeff Roberson * This is so involved because allocating and freeing 3740aef6126SJeff Roberson * while the zone lock is held will lead to deadlock. 3750aef6126SJeff Roberson * I have to do everything in stages and check for 3760aef6126SJeff Roberson * races. 3770aef6126SJeff Roberson */ 3780aef6126SJeff Roberson newhash = zone->uz_hash; 3795300d9ddSJeff Roberson ZONE_UNLOCK(zone); 3800aef6126SJeff Roberson ret = hash_alloc(&newhash); 3815300d9ddSJeff Roberson ZONE_LOCK(zone); 3820aef6126SJeff Roberson if (ret) { 3830aef6126SJeff Roberson if (hash_expand(&zone->uz_hash, &newhash)) { 3840aef6126SJeff Roberson oldhash = zone->uz_hash; 3850aef6126SJeff Roberson zone->uz_hash = newhash; 3860aef6126SJeff Roberson } else 3870aef6126SJeff Roberson oldhash = newhash; 3880aef6126SJeff Roberson 3890aef6126SJeff Roberson ZONE_UNLOCK(zone); 3900aef6126SJeff Roberson hash_free(&oldhash); 3910aef6126SJeff Roberson ZONE_LOCK(zone); 3920aef6126SJeff Roberson } 3935300d9ddSJeff Roberson } 3948355f576SJeff Roberson 3958355f576SJeff Roberson /* 3968355f576SJeff Roberson * Here we compute the working set size as the total number of items 3978355f576SJeff Roberson * left outstanding since the last time interval. This is slightly 3988355f576SJeff Roberson * suboptimal. What we really want is the highest number of outstanding 3998355f576SJeff Roberson * items during the last time quantum. This should be close enough. 4008355f576SJeff Roberson * 4018355f576SJeff Roberson * The working set size is used to throttle the zone_drain function. 4028355f576SJeff Roberson * We don't want to return memory that we may need again immediately. 4038355f576SJeff Roberson */ 4048355f576SJeff Roberson alloc = zone->uz_allocs - zone->uz_oallocs; 4058355f576SJeff Roberson zone->uz_oallocs = zone->uz_allocs; 4068355f576SJeff Roberson zone->uz_wssize = alloc; 4078355f576SJeff Roberson 4088355f576SJeff Roberson ZONE_UNLOCK(zone); 4098355f576SJeff Roberson } 4108355f576SJeff Roberson 4118355f576SJeff Roberson /* 4125300d9ddSJeff Roberson * Allocate and zero fill the next sized hash table from the appropriate 4135300d9ddSJeff Roberson * backing store. 4145300d9ddSJeff Roberson * 4155300d9ddSJeff Roberson * Arguments: 4160aef6126SJeff Roberson * hash A new hash structure with the old hash size in uh_hashsize 4175300d9ddSJeff Roberson * 4185300d9ddSJeff Roberson * Returns: 4190aef6126SJeff Roberson * 1 on sucess and 0 on failure. 4205300d9ddSJeff Roberson */ 42137c84183SPoul-Henning Kamp static int 4220aef6126SJeff Roberson hash_alloc(struct uma_hash *hash) 4235300d9ddSJeff Roberson { 4240aef6126SJeff Roberson int oldsize; 4255300d9ddSJeff Roberson int alloc; 4265300d9ddSJeff Roberson 4270aef6126SJeff Roberson oldsize = hash->uh_hashsize; 4280aef6126SJeff Roberson 4295300d9ddSJeff Roberson /* We're just going to go to a power of two greater */ 4300aef6126SJeff Roberson if (oldsize) { 4310aef6126SJeff Roberson hash->uh_hashsize = oldsize * 2; 4320aef6126SJeff Roberson alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; 4335300d9ddSJeff Roberson /* XXX Shouldn't be abusing DEVBUF here */ 4340aef6126SJeff Roberson hash->uh_slab_hash = (struct slabhead *)malloc(alloc, 4350aef6126SJeff Roberson M_DEVBUF, M_NOWAIT); 4365300d9ddSJeff Roberson } else { 4370aef6126SJeff Roberson alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; 4380aef6126SJeff Roberson hash->uh_slab_hash = uma_zalloc_internal(hashzone, NULL, 439a163d034SWarner Losh M_WAITOK); 4400aef6126SJeff Roberson hash->uh_hashsize = UMA_HASH_SIZE_INIT; 4415300d9ddSJeff Roberson } 4420aef6126SJeff Roberson if (hash->uh_slab_hash) { 4430aef6126SJeff Roberson bzero(hash->uh_slab_hash, alloc); 4440aef6126SJeff Roberson hash->uh_hashmask = hash->uh_hashsize - 1; 4450aef6126SJeff Roberson return (1); 4460aef6126SJeff Roberson } 4475300d9ddSJeff Roberson 4480aef6126SJeff Roberson return (0); 4495300d9ddSJeff Roberson } 4505300d9ddSJeff Roberson 4515300d9ddSJeff Roberson /* 4528355f576SJeff Roberson * Expands the hash table for OFFPAGE zones. This is done from zone_timeout 4538355f576SJeff Roberson * to reduce collisions. This must not be done in the regular allocation path, 4548355f576SJeff Roberson * otherwise, we can recurse on the vm while allocating pages. 4558355f576SJeff Roberson * 4568355f576SJeff Roberson * Arguments: 4570aef6126SJeff Roberson * oldhash The hash you want to expand 4580aef6126SJeff Roberson * newhash The hash structure for the new table 4598355f576SJeff Roberson * 4608355f576SJeff Roberson * Returns: 4618355f576SJeff Roberson * Nothing 4628355f576SJeff Roberson * 4638355f576SJeff Roberson * Discussion: 4648355f576SJeff Roberson */ 4650aef6126SJeff Roberson static int 4660aef6126SJeff Roberson hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) 4678355f576SJeff Roberson { 4688355f576SJeff Roberson uma_slab_t slab; 4698355f576SJeff Roberson int hval; 4708355f576SJeff Roberson int i; 4718355f576SJeff Roberson 4720aef6126SJeff Roberson if (!newhash->uh_slab_hash) 4730aef6126SJeff Roberson return (0); 4748355f576SJeff Roberson 4750aef6126SJeff Roberson if (oldhash->uh_hashsize >= newhash->uh_hashsize) 4760aef6126SJeff Roberson return (0); 4778355f576SJeff Roberson 4788355f576SJeff Roberson /* 4798355f576SJeff Roberson * I need to investigate hash algorithms for resizing without a 4808355f576SJeff Roberson * full rehash. 4818355f576SJeff Roberson */ 4828355f576SJeff Roberson 4830aef6126SJeff Roberson for (i = 0; i < oldhash->uh_hashsize; i++) 4840aef6126SJeff Roberson while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { 4850aef6126SJeff Roberson slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); 4860aef6126SJeff Roberson SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); 4870aef6126SJeff Roberson hval = UMA_HASH(newhash, slab->us_data); 4880aef6126SJeff Roberson SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], 4890aef6126SJeff Roberson slab, us_hlink); 4908355f576SJeff Roberson } 4918355f576SJeff Roberson 4920aef6126SJeff Roberson return (1); 4939c2cd7e5SJeff Roberson } 4949c2cd7e5SJeff Roberson 4955300d9ddSJeff Roberson /* 4965300d9ddSJeff Roberson * Free the hash bucket to the appropriate backing store. 4975300d9ddSJeff Roberson * 4985300d9ddSJeff Roberson * Arguments: 4995300d9ddSJeff Roberson * slab_hash The hash bucket we're freeing 5005300d9ddSJeff Roberson * hashsize The number of entries in that hash bucket 5015300d9ddSJeff Roberson * 5025300d9ddSJeff Roberson * Returns: 5035300d9ddSJeff Roberson * Nothing 5045300d9ddSJeff Roberson */ 5059c2cd7e5SJeff Roberson static void 5060aef6126SJeff Roberson hash_free(struct uma_hash *hash) 5079c2cd7e5SJeff Roberson { 5080aef6126SJeff Roberson if (hash->uh_slab_hash == NULL) 5090aef6126SJeff Roberson return; 5100aef6126SJeff Roberson if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) 5118355f576SJeff Roberson uma_zfree_internal(hashzone, 5120aef6126SJeff Roberson hash->uh_slab_hash, NULL, 0); 5138355f576SJeff Roberson else 5140aef6126SJeff Roberson free(hash->uh_slab_hash, M_DEVBUF); 5158355f576SJeff Roberson } 5168355f576SJeff Roberson 5178355f576SJeff Roberson /* 5188355f576SJeff Roberson * Frees all outstanding items in a bucket 5198355f576SJeff Roberson * 5208355f576SJeff Roberson * Arguments: 5218355f576SJeff Roberson * zone The zone to free to, must be unlocked. 5228355f576SJeff Roberson * bucket The free/alloc bucket with items, cpu queue must be locked. 5238355f576SJeff Roberson * 5248355f576SJeff Roberson * Returns: 5258355f576SJeff Roberson * Nothing 5268355f576SJeff Roberson */ 5278355f576SJeff Roberson 5288355f576SJeff Roberson static void 5298355f576SJeff Roberson bucket_drain(uma_zone_t zone, uma_bucket_t bucket) 5308355f576SJeff Roberson { 5318355f576SJeff Roberson uma_slab_t slab; 5328355f576SJeff Roberson int mzone; 5338355f576SJeff Roberson void *item; 5348355f576SJeff Roberson 5358355f576SJeff Roberson if (bucket == NULL) 5368355f576SJeff Roberson return; 5378355f576SJeff Roberson 5388355f576SJeff Roberson slab = NULL; 5398355f576SJeff Roberson mzone = 0; 5408355f576SJeff Roberson 5418355f576SJeff Roberson /* We have to lookup the slab again for malloc.. */ 5428355f576SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_MALLOC) 5438355f576SJeff Roberson mzone = 1; 5448355f576SJeff Roberson 545cae33c14SJeff Roberson while (bucket->ub_cnt > 0) { 546cae33c14SJeff Roberson bucket->ub_cnt--; 547cae33c14SJeff Roberson item = bucket->ub_bucket[bucket->ub_cnt]; 5488355f576SJeff Roberson #ifdef INVARIANTS 549cae33c14SJeff Roberson bucket->ub_bucket[bucket->ub_cnt] = NULL; 5508355f576SJeff Roberson KASSERT(item != NULL, 5518355f576SJeff Roberson ("bucket_drain: botched ptr, item is NULL")); 5528355f576SJeff Roberson #endif 5538355f576SJeff Roberson /* 5548355f576SJeff Roberson * This is extremely inefficient. The slab pointer was passed 5558355f576SJeff Roberson * to uma_zfree_arg, but we lost it because the buckets don't 5568355f576SJeff Roberson * hold them. This will go away when free() gets a size passed 5578355f576SJeff Roberson * to it. 5588355f576SJeff Roberson */ 55999571dc3SJeff Roberson if (mzone) 56099571dc3SJeff Roberson slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK)); 5618355f576SJeff Roberson uma_zfree_internal(zone, item, slab, 1); 5628355f576SJeff Roberson } 5638355f576SJeff Roberson } 5648355f576SJeff Roberson 5658355f576SJeff Roberson /* 5668355f576SJeff Roberson * Drains the per cpu caches for a zone. 5678355f576SJeff Roberson * 5688355f576SJeff Roberson * Arguments: 5698355f576SJeff Roberson * zone The zone to drain, must be unlocked. 570d56368d7SBosko Milekic * destroy Whether or not to destroy the pcpu buckets (from zone_dtor) 5718355f576SJeff Roberson * 5728355f576SJeff Roberson * Returns: 5738355f576SJeff Roberson * Nothing 5748355f576SJeff Roberson * 5758355f576SJeff Roberson * This function returns with the zone locked so that the per cpu queues can 5768355f576SJeff Roberson * not be filled until zone_drain is finished. 5778355f576SJeff Roberson * 5788355f576SJeff Roberson */ 5798355f576SJeff Roberson static void 580d56368d7SBosko Milekic cache_drain(uma_zone_t zone, int destroy) 5818355f576SJeff Roberson { 5828355f576SJeff Roberson uma_bucket_t bucket; 5838355f576SJeff Roberson uma_cache_t cache; 5848355f576SJeff Roberson int cpu; 5858355f576SJeff Roberson 5868355f576SJeff Roberson /* 5878355f576SJeff Roberson * Flush out the per cpu queues. 5888355f576SJeff Roberson * 589157d7b35SAlfred Perlstein * XXX This causes unnecessary thrashing due to immediately having 5908355f576SJeff Roberson * empty per cpu queues. I need to improve this. 5918355f576SJeff Roberson */ 5928355f576SJeff Roberson 5938355f576SJeff Roberson /* 5948355f576SJeff Roberson * We have to lock each cpu cache before locking the zone 5958355f576SJeff Roberson */ 5968355f576SJeff Roberson ZONE_UNLOCK(zone); 5978355f576SJeff Roberson 5988355f576SJeff Roberson for (cpu = 0; cpu < maxcpu; cpu++) { 5998355f576SJeff Roberson if (CPU_ABSENT(cpu)) 6008355f576SJeff Roberson continue; 601d88797c2SBosko Milekic CPU_LOCK(cpu); 6028355f576SJeff Roberson cache = &zone->uz_cpu[cpu]; 6038355f576SJeff Roberson bucket_drain(zone, cache->uc_allocbucket); 6048355f576SJeff Roberson bucket_drain(zone, cache->uc_freebucket); 605d56368d7SBosko Milekic if (destroy) { 606174ab450SBosko Milekic if (cache->uc_allocbucket != NULL) 607cae33c14SJeff Roberson bucket_free(cache->uc_allocbucket); 608174ab450SBosko Milekic if (cache->uc_freebucket != NULL) 609cae33c14SJeff Roberson bucket_free(cache->uc_freebucket); 610d56368d7SBosko Milekic cache->uc_allocbucket = cache->uc_freebucket = NULL; 611d56368d7SBosko Milekic } 6128355f576SJeff Roberson } 6138355f576SJeff Roberson 6148355f576SJeff Roberson /* 6158355f576SJeff Roberson * Drain the bucket queues and free the buckets, we just keep two per 6168355f576SJeff Roberson * cpu (alloc/free). 6178355f576SJeff Roberson */ 6188355f576SJeff Roberson ZONE_LOCK(zone); 6198355f576SJeff Roberson while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) { 6208355f576SJeff Roberson LIST_REMOVE(bucket, ub_link); 6218355f576SJeff Roberson ZONE_UNLOCK(zone); 6228355f576SJeff Roberson bucket_drain(zone, bucket); 623cae33c14SJeff Roberson bucket_free(bucket); 6248355f576SJeff Roberson ZONE_LOCK(zone); 6258355f576SJeff Roberson } 6268355f576SJeff Roberson 6278355f576SJeff Roberson /* Now we do the free queue.. */ 6288355f576SJeff Roberson while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) { 6298355f576SJeff Roberson LIST_REMOVE(bucket, ub_link); 630cae33c14SJeff Roberson bucket_free(bucket); 6318355f576SJeff Roberson } 6328355f576SJeff Roberson 6338355f576SJeff Roberson /* We unlock here, but they will all block until the zone is unlocked */ 6348355f576SJeff Roberson for (cpu = 0; cpu < maxcpu; cpu++) { 6358355f576SJeff Roberson if (CPU_ABSENT(cpu)) 6368355f576SJeff Roberson continue; 637d88797c2SBosko Milekic CPU_UNLOCK(cpu); 6388355f576SJeff Roberson } 6398355f576SJeff Roberson } 6408355f576SJeff Roberson 6418355f576SJeff Roberson /* 6428355f576SJeff Roberson * Frees pages from a zone back to the system. This is done on demand from 6438355f576SJeff Roberson * the pageout daemon. 6448355f576SJeff Roberson * 6458355f576SJeff Roberson * Arguments: 6468355f576SJeff Roberson * zone The zone to free pages from 6479c2cd7e5SJeff Roberson * all Should we drain all items? 648d56368d7SBosko Milekic * destroy Whether to destroy the zone and pcpu buckets (from zone_dtor) 6498355f576SJeff Roberson * 6508355f576SJeff Roberson * Returns: 6518355f576SJeff Roberson * Nothing. 6528355f576SJeff Roberson */ 6538355f576SJeff Roberson static void 654d56368d7SBosko Milekic zone_drain_common(uma_zone_t zone, int destroy) 6558355f576SJeff Roberson { 656713deb36SJeff Roberson struct slabhead freeslabs = {}; 6578355f576SJeff Roberson uma_slab_t slab; 6588355f576SJeff Roberson uma_slab_t n; 6598355f576SJeff Roberson u_int64_t extra; 6608355f576SJeff Roberson u_int8_t flags; 6618355f576SJeff Roberson u_int8_t *mem; 6628355f576SJeff Roberson int i; 6638355f576SJeff Roberson 6648355f576SJeff Roberson /* 6658355f576SJeff Roberson * We don't want to take pages from staticly allocated zones at this 6668355f576SJeff Roberson * time 6678355f576SJeff Roberson */ 6688355f576SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_NOFREE || zone->uz_freef == NULL) 6698355f576SJeff Roberson return; 6708355f576SJeff Roberson 6718355f576SJeff Roberson ZONE_LOCK(zone); 6728355f576SJeff Roberson 6738355f576SJeff Roberson if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) 674d56368d7SBosko Milekic cache_drain(zone, destroy); 675d56368d7SBosko Milekic 676d56368d7SBosko Milekic if (destroy) 677d56368d7SBosko Milekic zone->uz_wssize = 0; 6788355f576SJeff Roberson 6798355f576SJeff Roberson if (zone->uz_free < zone->uz_wssize) 6808355f576SJeff Roberson goto finished; 6818355f576SJeff Roberson #ifdef UMA_DEBUG 6828355f576SJeff Roberson printf("%s working set size: %llu free items: %u\n", 6838355f576SJeff Roberson zone->uz_name, (unsigned long long)zone->uz_wssize, zone->uz_free); 6848355f576SJeff Roberson #endif 6859c2cd7e5SJeff Roberson extra = zone->uz_free - zone->uz_wssize; 6868355f576SJeff Roberson extra /= zone->uz_ipers; 6878355f576SJeff Roberson 6888355f576SJeff Roberson /* extra is now the number of extra slabs that we can free */ 6898355f576SJeff Roberson 6908355f576SJeff Roberson if (extra == 0) 6918355f576SJeff Roberson goto finished; 6928355f576SJeff Roberson 6938355f576SJeff Roberson slab = LIST_FIRST(&zone->uz_free_slab); 6948355f576SJeff Roberson while (slab && extra) { 6958355f576SJeff Roberson n = LIST_NEXT(slab, us_link); 6968355f576SJeff Roberson 6978355f576SJeff Roberson /* We have no where to free these to */ 6988355f576SJeff Roberson if (slab->us_flags & UMA_SLAB_BOOT) { 6998355f576SJeff Roberson slab = n; 7008355f576SJeff Roberson continue; 7018355f576SJeff Roberson } 7028355f576SJeff Roberson 7038355f576SJeff Roberson LIST_REMOVE(slab, us_link); 7048355f576SJeff Roberson zone->uz_pages -= zone->uz_ppera; 7058355f576SJeff Roberson zone->uz_free -= zone->uz_ipers; 706713deb36SJeff Roberson 70799571dc3SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_HASH) 708713deb36SJeff Roberson UMA_HASH_REMOVE(&zone->uz_hash, slab, slab->us_data); 709713deb36SJeff Roberson 710713deb36SJeff Roberson SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); 711713deb36SJeff Roberson 712713deb36SJeff Roberson slab = n; 713713deb36SJeff Roberson extra--; 714713deb36SJeff Roberson } 715713deb36SJeff Roberson finished: 716713deb36SJeff Roberson ZONE_UNLOCK(zone); 717713deb36SJeff Roberson 718713deb36SJeff Roberson while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { 719713deb36SJeff Roberson SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); 7208355f576SJeff Roberson if (zone->uz_fini) 7218355f576SJeff Roberson for (i = 0; i < zone->uz_ipers; i++) 7228355f576SJeff Roberson zone->uz_fini( 7238355f576SJeff Roberson slab->us_data + (zone->uz_rsize * i), 7248355f576SJeff Roberson zone->uz_size); 7258355f576SJeff Roberson flags = slab->us_flags; 7268355f576SJeff Roberson mem = slab->us_data; 72799571dc3SJeff Roberson 72899571dc3SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_OFFPAGE) 7298355f576SJeff Roberson uma_zfree_internal(slabzone, slab, NULL, 0); 73048eea375SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_MALLOC) { 73148eea375SJeff Roberson vm_object_t obj; 73248eea375SJeff Roberson 73348eea375SJeff Roberson if (flags & UMA_SLAB_KMEM) 73448eea375SJeff Roberson obj = kmem_object; 73548eea375SJeff Roberson else 73648eea375SJeff Roberson obj = NULL; 73799571dc3SJeff Roberson for (i = 0; i < zone->uz_ppera; i++) 73899571dc3SJeff Roberson vsetobj((vm_offset_t)mem + (i * PAGE_SIZE), 73948eea375SJeff Roberson obj); 74048eea375SJeff Roberson } 7418355f576SJeff Roberson #ifdef UMA_DEBUG 7428355f576SJeff Roberson printf("%s: Returning %d bytes.\n", 7438355f576SJeff Roberson zone->uz_name, UMA_SLAB_SIZE * zone->uz_ppera); 7448355f576SJeff Roberson #endif 7458355f576SJeff Roberson zone->uz_freef(mem, UMA_SLAB_SIZE * zone->uz_ppera, flags); 7468355f576SJeff Roberson } 7478355f576SJeff Roberson 7488355f576SJeff Roberson } 7498355f576SJeff Roberson 750cae33c14SJeff Roberson static __inline void 751d56368d7SBosko Milekic zone_drain(uma_zone_t zone) 752d56368d7SBosko Milekic { 753d56368d7SBosko Milekic zone_drain_common(zone, 0); 754d56368d7SBosko Milekic } 755d56368d7SBosko Milekic 7568355f576SJeff Roberson /* 7578355f576SJeff Roberson * Allocate a new slab for a zone. This does not insert the slab onto a list. 7588355f576SJeff Roberson * 7598355f576SJeff Roberson * Arguments: 7608355f576SJeff Roberson * zone The zone to allocate slabs for 7618355f576SJeff Roberson * wait Shall we wait? 7628355f576SJeff Roberson * 7638355f576SJeff Roberson * Returns: 7648355f576SJeff Roberson * The slab that was allocated or NULL if there is no memory and the 7658355f576SJeff Roberson * caller specified M_NOWAIT. 7668355f576SJeff Roberson * 7678355f576SJeff Roberson */ 7688355f576SJeff Roberson static uma_slab_t 7698355f576SJeff Roberson slab_zalloc(uma_zone_t zone, int wait) 7708355f576SJeff Roberson { 7718355f576SJeff Roberson uma_slab_t slab; /* Starting slab */ 7728355f576SJeff Roberson u_int8_t *mem; 7738355f576SJeff Roberson u_int8_t flags; 7748355f576SJeff Roberson int i; 7758355f576SJeff Roberson 776a553d4b8SJeff Roberson slab = NULL; 777a553d4b8SJeff Roberson 7788355f576SJeff Roberson #ifdef UMA_DEBUG 7798355f576SJeff Roberson printf("slab_zalloc: Allocating a new slab for %s\n", zone->uz_name); 7808355f576SJeff Roberson #endif 7818355f576SJeff Roberson ZONE_UNLOCK(zone); 782a553d4b8SJeff Roberson 783a553d4b8SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_OFFPAGE) { 784bbee39c6SJeff Roberson slab = uma_zalloc_internal(slabzone, NULL, wait); 785a553d4b8SJeff Roberson if (slab == NULL) { 786a553d4b8SJeff Roberson ZONE_LOCK(zone); 787a553d4b8SJeff Roberson return NULL; 788a553d4b8SJeff Roberson } 789a553d4b8SJeff Roberson } 790a553d4b8SJeff Roberson 7913370c5bfSJeff Roberson /* 7923370c5bfSJeff Roberson * This reproduces the old vm_zone behavior of zero filling pages the 7933370c5bfSJeff Roberson * first time they are added to a zone. 7943370c5bfSJeff Roberson * 7953370c5bfSJeff Roberson * Malloced items are zeroed in uma_zalloc. 7963370c5bfSJeff Roberson */ 7973370c5bfSJeff Roberson 7983370c5bfSJeff Roberson if ((zone->uz_flags & UMA_ZFLAG_MALLOC) == 0) 7993370c5bfSJeff Roberson wait |= M_ZERO; 8003370c5bfSJeff Roberson else 8013370c5bfSJeff Roberson wait &= ~M_ZERO; 8023370c5bfSJeff Roberson 803a553d4b8SJeff Roberson if (booted || (zone->uz_flags & UMA_ZFLAG_PRIVALLOC)) { 804234c7726SAlan Cox mem = zone->uz_allocf(zone, zone->uz_ppera * UMA_SLAB_SIZE, 805234c7726SAlan Cox &flags, wait); 806a553d4b8SJeff Roberson if (mem == NULL) { 8078355f576SJeff Roberson ZONE_LOCK(zone); 8088355f576SJeff Roberson return (NULL); 809a553d4b8SJeff Roberson } 8108355f576SJeff Roberson } else { 811a553d4b8SJeff Roberson uma_slab_t tmps; 8128355f576SJeff Roberson 8138355f576SJeff Roberson if (zone->uz_ppera > 1) 8148355f576SJeff Roberson panic("UMA: Attemping to allocate multiple pages before vm has started.\n"); 8158355f576SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_MALLOC) 8168355f576SJeff Roberson panic("Mallocing before uma_startup2 has been called.\n"); 8178355f576SJeff Roberson if (uma_boot_free == 0) 8188355f576SJeff Roberson panic("UMA: Ran out of pre init pages, increase UMA_BOOT_PAGES\n"); 819a553d4b8SJeff Roberson tmps = LIST_FIRST(&uma_boot_pages); 820a553d4b8SJeff Roberson LIST_REMOVE(tmps, us_link); 8218355f576SJeff Roberson uma_boot_free--; 822a553d4b8SJeff Roberson mem = tmps->us_data; 823f3da1873SJeff Roberson flags = tmps->us_flags; 8248355f576SJeff Roberson } 8258355f576SJeff Roberson 8265c0e403bSJeff Roberson /* Point the slab into the allocated memory */ 82799571dc3SJeff Roberson if (!(zone->uz_flags & UMA_ZFLAG_OFFPAGE)) 8288355f576SJeff Roberson slab = (uma_slab_t )(mem + zone->uz_pgoff); 8295c0e403bSJeff Roberson 83099571dc3SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_MALLOC) 83199571dc3SJeff Roberson for (i = 0; i < zone->uz_ppera; i++) 83299571dc3SJeff Roberson vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); 8338355f576SJeff Roberson 8348355f576SJeff Roberson slab->us_zone = zone; 8358355f576SJeff Roberson slab->us_data = mem; 8368355f576SJeff Roberson 8378355f576SJeff Roberson /* 8388355f576SJeff Roberson * This is intended to spread data out across cache lines. 8398355f576SJeff Roberson * 8408355f576SJeff Roberson * This code doesn't seem to work properly on x86, and on alpha 8418355f576SJeff Roberson * it makes absolutely no performance difference. I'm sure it could 8428355f576SJeff Roberson * use some tuning, but sun makes outrageous claims about it's 8438355f576SJeff Roberson * performance. 8448355f576SJeff Roberson */ 8458355f576SJeff Roberson #if 0 8468355f576SJeff Roberson if (zone->uz_cachemax) { 8478355f576SJeff Roberson slab->us_data += zone->uz_cacheoff; 8488355f576SJeff Roberson zone->uz_cacheoff += UMA_CACHE_INC; 8498355f576SJeff Roberson if (zone->uz_cacheoff > zone->uz_cachemax) 8508355f576SJeff Roberson zone->uz_cacheoff = 0; 8518355f576SJeff Roberson } 8528355f576SJeff Roberson #endif 8538355f576SJeff Roberson 8548355f576SJeff Roberson slab->us_freecount = zone->uz_ipers; 8558355f576SJeff Roberson slab->us_firstfree = 0; 8568355f576SJeff Roberson slab->us_flags = flags; 8578355f576SJeff Roberson for (i = 0; i < zone->uz_ipers; i++) 8588355f576SJeff Roberson slab->us_freelist[i] = i+1; 8598355f576SJeff Roberson 8608355f576SJeff Roberson if (zone->uz_init) 8618355f576SJeff Roberson for (i = 0; i < zone->uz_ipers; i++) 8628355f576SJeff Roberson zone->uz_init(slab->us_data + (zone->uz_rsize * i), 8638355f576SJeff Roberson zone->uz_size); 8645c0e403bSJeff Roberson ZONE_LOCK(zone); 8655c0e403bSJeff Roberson 86699571dc3SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_HASH) 8675c0e403bSJeff Roberson UMA_HASH_INSERT(&zone->uz_hash, slab, mem); 8688355f576SJeff Roberson 8698355f576SJeff Roberson zone->uz_pages += zone->uz_ppera; 8708355f576SJeff Roberson zone->uz_free += zone->uz_ipers; 8718355f576SJeff Roberson 8725c0e403bSJeff Roberson 8738355f576SJeff Roberson return (slab); 8748355f576SJeff Roberson } 8758355f576SJeff Roberson 8768355f576SJeff Roberson /* 8778355f576SJeff Roberson * Allocates a number of pages from the system 8788355f576SJeff Roberson * 8798355f576SJeff Roberson * Arguments: 8808355f576SJeff Roberson * zone Unused 8818355f576SJeff Roberson * bytes The number of bytes requested 8828355f576SJeff Roberson * wait Shall we wait? 8838355f576SJeff Roberson * 8848355f576SJeff Roberson * Returns: 8858355f576SJeff Roberson * A pointer to the alloced memory or possibly 8868355f576SJeff Roberson * NULL if M_NOWAIT is set. 8878355f576SJeff Roberson */ 8888355f576SJeff Roberson static void * 8898355f576SJeff Roberson page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait) 8908355f576SJeff Roberson { 8918355f576SJeff Roberson void *p; /* Returned page */ 8928355f576SJeff Roberson 8938355f576SJeff Roberson *pflag = UMA_SLAB_KMEM; 8948355f576SJeff Roberson p = (void *) kmem_malloc(kmem_map, bytes, wait); 8958355f576SJeff Roberson 8968355f576SJeff Roberson return (p); 8978355f576SJeff Roberson } 8988355f576SJeff Roberson 8998355f576SJeff Roberson /* 9008355f576SJeff Roberson * Allocates a number of pages from within an object 9018355f576SJeff Roberson * 9028355f576SJeff Roberson * Arguments: 9038355f576SJeff Roberson * zone Unused 9048355f576SJeff Roberson * bytes The number of bytes requested 9058355f576SJeff Roberson * wait Shall we wait? 9068355f576SJeff Roberson * 9078355f576SJeff Roberson * Returns: 9088355f576SJeff Roberson * A pointer to the alloced memory or possibly 9098355f576SJeff Roberson * NULL if M_NOWAIT is set. 910494273beSJeff Roberson * 9118355f576SJeff Roberson */ 9128355f576SJeff Roberson static void * 9138355f576SJeff Roberson obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait) 9148355f576SJeff Roberson { 915b245ac95SAlan Cox vm_object_t object; 916b245ac95SAlan Cox vm_offset_t retkva, zkva; 9178355f576SJeff Roberson vm_page_t p; 918b245ac95SAlan Cox int pages, startpages; 9198355f576SJeff Roberson 920b245ac95SAlan Cox object = zone->uz_obj; 92155f7c614SArchie Cobbs retkva = 0; 9228355f576SJeff Roberson 9238355f576SJeff Roberson /* 9248355f576SJeff Roberson * This looks a little weird since we're getting one page at a time 9258355f576SJeff Roberson */ 926b245ac95SAlan Cox VM_OBJECT_LOCK(object); 927b245ac95SAlan Cox p = TAILQ_LAST(&object->memq, pglist); 928b245ac95SAlan Cox pages = p != NULL ? p->pindex + 1 : 0; 929b245ac95SAlan Cox startpages = pages; 9308355f576SJeff Roberson zkva = zone->uz_kva + pages * PAGE_SIZE; 931b245ac95SAlan Cox for (; bytes > 0; bytes -= PAGE_SIZE) { 932b245ac95SAlan Cox p = vm_page_alloc(object, pages, 933b245ac95SAlan Cox VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED); 934b245ac95SAlan Cox if (p == NULL) { 935b245ac95SAlan Cox if (pages != startpages) 936b245ac95SAlan Cox pmap_qremove(retkva, pages - startpages); 937b245ac95SAlan Cox while (pages != startpages) { 938b245ac95SAlan Cox pages--; 939b245ac95SAlan Cox p = TAILQ_LAST(&object->memq, pglist); 940b245ac95SAlan Cox vm_page_lock_queues(); 941b245ac95SAlan Cox vm_page_unwire(p, 0); 942b245ac95SAlan Cox vm_page_free(p); 943b245ac95SAlan Cox vm_page_unlock_queues(); 944b245ac95SAlan Cox } 945b245ac95SAlan Cox retkva = 0; 946b245ac95SAlan Cox goto done; 947b245ac95SAlan Cox } 948b245ac95SAlan Cox pmap_qenter(zkva, &p, 1); 94955f7c614SArchie Cobbs if (retkva == 0) 9508355f576SJeff Roberson retkva = zkva; 951b245ac95SAlan Cox zkva += PAGE_SIZE; 9528355f576SJeff Roberson pages += 1; 9538355f576SJeff Roberson } 954b245ac95SAlan Cox done: 955b245ac95SAlan Cox VM_OBJECT_UNLOCK(object); 9568355f576SJeff Roberson 9578355f576SJeff Roberson *flags = UMA_SLAB_PRIV; 9588355f576SJeff Roberson 9598355f576SJeff Roberson return ((void *)retkva); 9608355f576SJeff Roberson } 9618355f576SJeff Roberson 9628355f576SJeff Roberson /* 9638355f576SJeff Roberson * Frees a number of pages to the system 9648355f576SJeff Roberson * 9658355f576SJeff Roberson * Arguments: 9668355f576SJeff Roberson * mem A pointer to the memory to be freed 9678355f576SJeff Roberson * size The size of the memory being freed 9688355f576SJeff Roberson * flags The original p->us_flags field 9698355f576SJeff Roberson * 9708355f576SJeff Roberson * Returns: 9718355f576SJeff Roberson * Nothing 9728355f576SJeff Roberson * 9738355f576SJeff Roberson */ 9748355f576SJeff Roberson static void 9758355f576SJeff Roberson page_free(void *mem, int size, u_int8_t flags) 9768355f576SJeff Roberson { 9778355f576SJeff Roberson vm_map_t map; 9783370c5bfSJeff Roberson 9798355f576SJeff Roberson if (flags & UMA_SLAB_KMEM) 9808355f576SJeff Roberson map = kmem_map; 9818355f576SJeff Roberson else 9828355f576SJeff Roberson panic("UMA: page_free used with invalid flags %d\n", flags); 9838355f576SJeff Roberson 9848355f576SJeff Roberson kmem_free(map, (vm_offset_t)mem, size); 9858355f576SJeff Roberson } 9868355f576SJeff Roberson 9878355f576SJeff Roberson /* 9888355f576SJeff Roberson * Zero fill initializer 9898355f576SJeff Roberson * 9908355f576SJeff Roberson * Arguments/Returns follow uma_init specifications 9918355f576SJeff Roberson * 9928355f576SJeff Roberson */ 9938355f576SJeff Roberson static void 9948355f576SJeff Roberson zero_init(void *mem, int size) 9958355f576SJeff Roberson { 9968355f576SJeff Roberson bzero(mem, size); 9978355f576SJeff Roberson } 9988355f576SJeff Roberson 9998355f576SJeff Roberson /* 10008355f576SJeff Roberson * Finish creating a small uma zone. This calculates ipers, and the zone size. 10018355f576SJeff Roberson * 10028355f576SJeff Roberson * Arguments 10038355f576SJeff Roberson * zone The zone we should initialize 10048355f576SJeff Roberson * 10058355f576SJeff Roberson * Returns 10068355f576SJeff Roberson * Nothing 10078355f576SJeff Roberson */ 10088355f576SJeff Roberson static void 10098355f576SJeff Roberson zone_small_init(uma_zone_t zone) 10108355f576SJeff Roberson { 10118355f576SJeff Roberson int rsize; 10128355f576SJeff Roberson int memused; 10138355f576SJeff Roberson int ipers; 10148355f576SJeff Roberson 10158355f576SJeff Roberson rsize = zone->uz_size; 10168355f576SJeff Roberson 10178355f576SJeff Roberson if (rsize < UMA_SMALLEST_UNIT) 10188355f576SJeff Roberson rsize = UMA_SMALLEST_UNIT; 10198355f576SJeff Roberson 10208355f576SJeff Roberson if (rsize & zone->uz_align) 10218355f576SJeff Roberson rsize = (rsize & ~zone->uz_align) + (zone->uz_align + 1); 10228355f576SJeff Roberson 10238355f576SJeff Roberson zone->uz_rsize = rsize; 10248355f576SJeff Roberson 10258355f576SJeff Roberson rsize += 1; /* Account for the byte of linkage */ 10268355f576SJeff Roberson zone->uz_ipers = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) / rsize; 10278355f576SJeff Roberson zone->uz_ppera = 1; 10288355f576SJeff Roberson 102920e8e865SBosko Milekic KASSERT(zone->uz_ipers != 0, ("zone_small_init: ipers is 0, uh-oh!")); 10308355f576SJeff Roberson memused = zone->uz_ipers * zone->uz_rsize; 10318355f576SJeff Roberson 10328355f576SJeff Roberson /* Can we do any better? */ 10338355f576SJeff Roberson if ((UMA_SLAB_SIZE - memused) >= UMA_MAX_WASTE) { 103420e8e865SBosko Milekic /* 103520e8e865SBosko Milekic * We can't do this if we're internal or if we've been 103620e8e865SBosko Milekic * asked to not go to the VM for buckets. If we do this we 103720e8e865SBosko Milekic * may end up going to the VM (kmem_map) for slabs which we 103820e8e865SBosko Milekic * do not want to do if we're UMA_ZFLAG_CACHEONLY as a 103920e8e865SBosko Milekic * result of UMA_ZONE_VM, which clearly forbids it. 104020e8e865SBosko Milekic */ 104120e8e865SBosko Milekic if ((zone->uz_flags & UMA_ZFLAG_INTERNAL) || 104220e8e865SBosko Milekic (zone->uz_flags & UMA_ZFLAG_CACHEONLY)) 10438355f576SJeff Roberson return; 10448355f576SJeff Roberson ipers = UMA_SLAB_SIZE / zone->uz_rsize; 10458355f576SJeff Roberson if (ipers > zone->uz_ipers) { 10468355f576SJeff Roberson zone->uz_flags |= UMA_ZFLAG_OFFPAGE; 104799571dc3SJeff Roberson if ((zone->uz_flags & UMA_ZFLAG_MALLOC) == 0) 104899571dc3SJeff Roberson zone->uz_flags |= UMA_ZFLAG_HASH; 10498355f576SJeff Roberson zone->uz_ipers = ipers; 10508355f576SJeff Roberson } 10518355f576SJeff Roberson } 10528355f576SJeff Roberson 10538355f576SJeff Roberson } 10548355f576SJeff Roberson 10558355f576SJeff Roberson /* 10568355f576SJeff Roberson * Finish creating a large (> UMA_SLAB_SIZE) uma zone. Just give in and do 10578355f576SJeff Roberson * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be 10588355f576SJeff Roberson * more complicated. 10598355f576SJeff Roberson * 10608355f576SJeff Roberson * Arguments 10618355f576SJeff Roberson * zone The zone we should initialize 10628355f576SJeff Roberson * 10638355f576SJeff Roberson * Returns 10648355f576SJeff Roberson * Nothing 10658355f576SJeff Roberson */ 10668355f576SJeff Roberson static void 10678355f576SJeff Roberson zone_large_init(uma_zone_t zone) 10688355f576SJeff Roberson { 10698355f576SJeff Roberson int pages; 10708355f576SJeff Roberson 107120e8e865SBosko Milekic KASSERT((zone->uz_flags & UMA_ZFLAG_CACHEONLY) == 0, 107220e8e865SBosko Milekic ("zone_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY zone")); 107320e8e865SBosko Milekic 10748355f576SJeff Roberson pages = zone->uz_size / UMA_SLAB_SIZE; 10758355f576SJeff Roberson 10768355f576SJeff Roberson /* Account for remainder */ 10778355f576SJeff Roberson if ((pages * UMA_SLAB_SIZE) < zone->uz_size) 10788355f576SJeff Roberson pages++; 10798355f576SJeff Roberson 10808355f576SJeff Roberson zone->uz_ppera = pages; 10818355f576SJeff Roberson zone->uz_ipers = 1; 10828355f576SJeff Roberson 10838355f576SJeff Roberson zone->uz_flags |= UMA_ZFLAG_OFFPAGE; 108499571dc3SJeff Roberson if ((zone->uz_flags & UMA_ZFLAG_MALLOC) == 0) 108599571dc3SJeff Roberson zone->uz_flags |= UMA_ZFLAG_HASH; 108699571dc3SJeff Roberson 10878355f576SJeff Roberson zone->uz_rsize = zone->uz_size; 10888355f576SJeff Roberson } 10898355f576SJeff Roberson 10908355f576SJeff Roberson /* 10918355f576SJeff Roberson * Zone header ctor. This initializes all fields, locks, etc. And inserts 10928355f576SJeff Roberson * the zone onto the global zone list. 10938355f576SJeff Roberson * 10948355f576SJeff Roberson * Arguments/Returns follow uma_ctor specifications 10958355f576SJeff Roberson * udata Actually uma_zcreat_args 10968355f576SJeff Roberson * 10978355f576SJeff Roberson */ 10988355f576SJeff Roberson 10998355f576SJeff Roberson static void 11008355f576SJeff Roberson zone_ctor(void *mem, int size, void *udata) 11018355f576SJeff Roberson { 11028355f576SJeff Roberson struct uma_zctor_args *arg = udata; 11038355f576SJeff Roberson uma_zone_t zone = mem; 110428bc4419SJeff Roberson int privlc; 11058355f576SJeff Roberson 11068355f576SJeff Roberson bzero(zone, size); 11078355f576SJeff Roberson zone->uz_name = arg->name; 11088355f576SJeff Roberson zone->uz_size = arg->size; 11098355f576SJeff Roberson zone->uz_ctor = arg->ctor; 11108355f576SJeff Roberson zone->uz_dtor = arg->dtor; 11118355f576SJeff Roberson zone->uz_init = arg->uminit; 1112e221e841SJeff Roberson zone->uz_fini = arg->fini; 11138355f576SJeff Roberson zone->uz_align = arg->align; 11148355f576SJeff Roberson zone->uz_free = 0; 11158355f576SJeff Roberson zone->uz_pages = 0; 11168355f576SJeff Roberson zone->uz_flags = 0; 11178355f576SJeff Roberson zone->uz_allocf = page_alloc; 11188355f576SJeff Roberson zone->uz_freef = page_free; 11198355f576SJeff Roberson 11208355f576SJeff Roberson if (arg->flags & UMA_ZONE_ZINIT) 11218355f576SJeff Roberson zone->uz_init = zero_init; 11228355f576SJeff Roberson 11238355f576SJeff Roberson if (arg->flags & UMA_ZONE_INTERNAL) 11248355f576SJeff Roberson zone->uz_flags |= UMA_ZFLAG_INTERNAL; 11258355f576SJeff Roberson 11268355f576SJeff Roberson if (arg->flags & UMA_ZONE_MALLOC) 11278355f576SJeff Roberson zone->uz_flags |= UMA_ZFLAG_MALLOC; 11288355f576SJeff Roberson 11298355f576SJeff Roberson if (arg->flags & UMA_ZONE_NOFREE) 11308355f576SJeff Roberson zone->uz_flags |= UMA_ZFLAG_NOFREE; 11318355f576SJeff Roberson 113218aa2de5SJeff Roberson if (arg->flags & UMA_ZONE_VM) 113320e8e865SBosko Milekic zone->uz_flags |= UMA_ZFLAG_CACHEONLY; 113418aa2de5SJeff Roberson 113520e8e865SBosko Milekic /* 113620e8e865SBosko Milekic * XXX: 113720e8e865SBosko Milekic * The +1 byte added to uz_size is to account for the byte of 113820e8e865SBosko Milekic * linkage that is added to the size in zone_small_init(). If 113920e8e865SBosko Milekic * we don't account for this here then we may end up in 114020e8e865SBosko Milekic * zone_small_init() with a calculated 'ipers' of 0. 114120e8e865SBosko Milekic */ 114220e8e865SBosko Milekic if ((zone->uz_size+1) > (UMA_SLAB_SIZE - sizeof(struct uma_slab))) 11438355f576SJeff Roberson zone_large_init(zone); 11448355f576SJeff Roberson else 11458355f576SJeff Roberson zone_small_init(zone); 114648eea375SJeff Roberson #ifdef UMA_MD_SMALL_ALLOC 114748eea375SJeff Roberson if (zone->uz_ppera == 1) { 114848eea375SJeff Roberson zone->uz_allocf = uma_small_alloc; 114948eea375SJeff Roberson zone->uz_freef = uma_small_free; 115048eea375SJeff Roberson } 115148eea375SJeff Roberson #endif /* UMA_MD_SMALL_ALLOC */ 11528355f576SJeff Roberson 115328bc4419SJeff Roberson if (arg->flags & UMA_ZONE_MTXCLASS) 115428bc4419SJeff Roberson privlc = 1; 115528bc4419SJeff Roberson else 115628bc4419SJeff Roberson privlc = 0; 115728bc4419SJeff Roberson 11588355f576SJeff Roberson /* 11598355f576SJeff Roberson * If we're putting the slab header in the actual page we need to 11608355f576SJeff Roberson * figure out where in each page it goes. This calculates a right 11619d5abbddSJens Schweikhardt * justified offset into the memory on an ALIGN_PTR boundary. 11628355f576SJeff Roberson */ 11638355f576SJeff Roberson if (!(zone->uz_flags & UMA_ZFLAG_OFFPAGE)) { 11648355f576SJeff Roberson int totsize; 11658355f576SJeff Roberson int waste; 11668355f576SJeff Roberson 11678355f576SJeff Roberson /* Size of the slab struct and free list */ 11688355f576SJeff Roberson totsize = sizeof(struct uma_slab) + zone->uz_ipers; 11698355f576SJeff Roberson if (totsize & UMA_ALIGN_PTR) 11708355f576SJeff Roberson totsize = (totsize & ~UMA_ALIGN_PTR) + 11718355f576SJeff Roberson (UMA_ALIGN_PTR + 1); 11728355f576SJeff Roberson zone->uz_pgoff = UMA_SLAB_SIZE - totsize; 11738355f576SJeff Roberson 11748355f576SJeff Roberson waste = zone->uz_pgoff; 11758355f576SJeff Roberson waste -= (zone->uz_ipers * zone->uz_rsize); 11768355f576SJeff Roberson 11778355f576SJeff Roberson /* 11788355f576SJeff Roberson * This calculates how much space we have for cache line size 11798355f576SJeff Roberson * optimizations. It works by offseting each slab slightly. 11808355f576SJeff Roberson * Currently it breaks on x86, and so it is disabled. 11818355f576SJeff Roberson */ 11828355f576SJeff Roberson 11838355f576SJeff Roberson if (zone->uz_align < UMA_CACHE_INC && waste > UMA_CACHE_INC) { 11848355f576SJeff Roberson zone->uz_cachemax = waste - UMA_CACHE_INC; 11858355f576SJeff Roberson zone->uz_cacheoff = 0; 11868355f576SJeff Roberson } 11878355f576SJeff Roberson 11888355f576SJeff Roberson totsize = zone->uz_pgoff + sizeof(struct uma_slab) 11898355f576SJeff Roberson + zone->uz_ipers; 11908355f576SJeff Roberson /* I don't think it's possible, but I'll make sure anyway */ 11918355f576SJeff Roberson if (totsize > UMA_SLAB_SIZE) { 11928355f576SJeff Roberson printf("zone %s ipers %d rsize %d size %d\n", 11938355f576SJeff Roberson zone->uz_name, zone->uz_ipers, zone->uz_rsize, 11948355f576SJeff Roberson zone->uz_size); 11958355f576SJeff Roberson panic("UMA slab won't fit.\n"); 11968355f576SJeff Roberson } 11978355f576SJeff Roberson } 11988355f576SJeff Roberson 119999571dc3SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_HASH) 120099571dc3SJeff Roberson hash_alloc(&zone->uz_hash); 120199571dc3SJeff Roberson 12028355f576SJeff Roberson #ifdef UMA_DEBUG 12038355f576SJeff Roberson printf("%s(%p) size = %d ipers = %d ppera = %d pgoff = %d\n", 12048355f576SJeff Roberson zone->uz_name, zone, 12058355f576SJeff Roberson zone->uz_size, zone->uz_ipers, 12068355f576SJeff Roberson zone->uz_ppera, zone->uz_pgoff); 12078355f576SJeff Roberson #endif 120828bc4419SJeff Roberson ZONE_LOCK_INIT(zone, privlc); 12098355f576SJeff Roberson 12108355f576SJeff Roberson mtx_lock(&uma_mtx); 12118355f576SJeff Roberson LIST_INSERT_HEAD(&uma_zones, zone, uz_link); 12128355f576SJeff Roberson mtx_unlock(&uma_mtx); 12138355f576SJeff Roberson 12148355f576SJeff Roberson /* 12158355f576SJeff Roberson * Some internal zones don't have room allocated for the per cpu 12168355f576SJeff Roberson * caches. If we're internal, bail out here. 12178355f576SJeff Roberson */ 12188355f576SJeff Roberson 12198355f576SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_INTERNAL) 12208355f576SJeff Roberson return; 12218355f576SJeff Roberson 1222cae33c14SJeff Roberson if (zone->uz_ipers <= BUCKET_MAX) 1223cae33c14SJeff Roberson zone->uz_count = zone->uz_ipers; 12248355f576SJeff Roberson else 1225cae33c14SJeff Roberson zone->uz_count = BUCKET_MAX; 12268355f576SJeff Roberson } 12278355f576SJeff Roberson 12288355f576SJeff Roberson /* 12299c2cd7e5SJeff Roberson * Zone header dtor. This frees all data, destroys locks, frees the hash table 12309c2cd7e5SJeff Roberson * and removes the zone from the global list. 12319c2cd7e5SJeff Roberson * 12329c2cd7e5SJeff Roberson * Arguments/Returns follow uma_dtor specifications 12339c2cd7e5SJeff Roberson * udata unused 12349c2cd7e5SJeff Roberson */ 12359c2cd7e5SJeff Roberson 12369c2cd7e5SJeff Roberson static void 12379c2cd7e5SJeff Roberson zone_dtor(void *arg, int size, void *udata) 12389c2cd7e5SJeff Roberson { 12399c2cd7e5SJeff Roberson uma_zone_t zone; 12409c2cd7e5SJeff Roberson 12419c2cd7e5SJeff Roberson zone = (uma_zone_t)arg; 124217b9cc49SJeff Roberson mtx_lock(&uma_mtx); 124317b9cc49SJeff Roberson LIST_REMOVE(zone, uz_link); 1244d56368d7SBosko Milekic zone_drain_common(zone, 1); 124517b9cc49SJeff Roberson mtx_unlock(&uma_mtx); 124617b9cc49SJeff Roberson 12479c2cd7e5SJeff Roberson ZONE_LOCK(zone); 12489c2cd7e5SJeff Roberson if (zone->uz_free != 0) 1249886eaaacSPoul-Henning Kamp printf("Zone %s was not empty (%d items). Lost %d pages of memory.\n", 1250886eaaacSPoul-Henning Kamp zone->uz_name, zone->uz_free, zone->uz_pages); 12519c2cd7e5SJeff Roberson 12529c2cd7e5SJeff Roberson ZONE_UNLOCK(zone); 12530aef6126SJeff Roberson if ((zone->uz_flags & UMA_ZFLAG_OFFPAGE) != 0) 12540aef6126SJeff Roberson hash_free(&zone->uz_hash); 12550aef6126SJeff Roberson 12569c2cd7e5SJeff Roberson ZONE_LOCK_FINI(zone); 12579c2cd7e5SJeff Roberson } 12589c2cd7e5SJeff Roberson /* 12598355f576SJeff Roberson * Traverses every zone in the system and calls a callback 12608355f576SJeff Roberson * 12618355f576SJeff Roberson * Arguments: 12628355f576SJeff Roberson * zfunc A pointer to a function which accepts a zone 12638355f576SJeff Roberson * as an argument. 12648355f576SJeff Roberson * 12658355f576SJeff Roberson * Returns: 12668355f576SJeff Roberson * Nothing 12678355f576SJeff Roberson */ 12688355f576SJeff Roberson static void 12698355f576SJeff Roberson zone_foreach(void (*zfunc)(uma_zone_t)) 12708355f576SJeff Roberson { 12718355f576SJeff Roberson uma_zone_t zone; 12728355f576SJeff Roberson 12738355f576SJeff Roberson mtx_lock(&uma_mtx); 12748355f576SJeff Roberson LIST_FOREACH(zone, &uma_zones, uz_link) { 12758355f576SJeff Roberson zfunc(zone); 12768355f576SJeff Roberson } 12778355f576SJeff Roberson mtx_unlock(&uma_mtx); 12788355f576SJeff Roberson } 12798355f576SJeff Roberson 12808355f576SJeff Roberson /* Public functions */ 12818355f576SJeff Roberson /* See uma.h */ 12828355f576SJeff Roberson void 12838355f576SJeff Roberson uma_startup(void *bootmem) 12848355f576SJeff Roberson { 12858355f576SJeff Roberson struct uma_zctor_args args; 12868355f576SJeff Roberson uma_slab_t slab; 12878355f576SJeff Roberson int slabsize; 12888355f576SJeff Roberson int i; 12898355f576SJeff Roberson 12908355f576SJeff Roberson #ifdef UMA_DEBUG 12918355f576SJeff Roberson printf("Creating uma zone headers zone.\n"); 12928355f576SJeff Roberson #endif 12938355f576SJeff Roberson #ifdef SMP 12948355f576SJeff Roberson maxcpu = mp_maxid + 1; 12958355f576SJeff Roberson #else 12968355f576SJeff Roberson maxcpu = 1; 12978355f576SJeff Roberson #endif 12988355f576SJeff Roberson #ifdef UMA_DEBUG 12998355f576SJeff Roberson printf("Max cpu = %d, mp_maxid = %d\n", maxcpu, mp_maxid); 13008355f576SJeff Roberson Debugger("stop"); 13018355f576SJeff Roberson #endif 13026008862bSJohn Baldwin mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF); 13038355f576SJeff Roberson /* "manually" Create the initial zone */ 13048355f576SJeff Roberson args.name = "UMA Zones"; 13058355f576SJeff Roberson args.size = sizeof(struct uma_zone) + 13068355f576SJeff Roberson (sizeof(struct uma_cache) * (maxcpu - 1)); 13078355f576SJeff Roberson args.ctor = zone_ctor; 13089c2cd7e5SJeff Roberson args.dtor = zone_dtor; 13098355f576SJeff Roberson args.uminit = zero_init; 13108355f576SJeff Roberson args.fini = NULL; 13118355f576SJeff Roberson args.align = 32 - 1; 13128355f576SJeff Roberson args.flags = UMA_ZONE_INTERNAL; 13138355f576SJeff Roberson /* The initial zone has no Per cpu queues so it's smaller */ 13148355f576SJeff Roberson zone_ctor(zones, sizeof(struct uma_zone), &args); 13158355f576SJeff Roberson 1316d88797c2SBosko Milekic /* Initialize the pcpu cache lock set once and for all */ 1317d88797c2SBosko Milekic for (i = 0; i < maxcpu; i++) 1318d88797c2SBosko Milekic CPU_LOCK_INIT(i); 1319d88797c2SBosko Milekic 13208355f576SJeff Roberson #ifdef UMA_DEBUG 13218355f576SJeff Roberson printf("Filling boot free list.\n"); 13228355f576SJeff Roberson #endif 13238355f576SJeff Roberson for (i = 0; i < UMA_BOOT_PAGES; i++) { 13248355f576SJeff Roberson slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE)); 13258355f576SJeff Roberson slab->us_data = (u_int8_t *)slab; 13268355f576SJeff Roberson slab->us_flags = UMA_SLAB_BOOT; 13278355f576SJeff Roberson LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link); 13288355f576SJeff Roberson uma_boot_free++; 13298355f576SJeff Roberson } 13308355f576SJeff Roberson 13318355f576SJeff Roberson #ifdef UMA_DEBUG 13328355f576SJeff Roberson printf("Creating slab zone.\n"); 13338355f576SJeff Roberson #endif 13348355f576SJeff Roberson 13358355f576SJeff Roberson /* 13368355f576SJeff Roberson * This is the max number of free list items we'll have with 13378355f576SJeff Roberson * offpage slabs. 13388355f576SJeff Roberson */ 13398355f576SJeff Roberson 13408355f576SJeff Roberson slabsize = UMA_SLAB_SIZE - sizeof(struct uma_slab); 13418355f576SJeff Roberson slabsize /= UMA_MAX_WASTE; 13428355f576SJeff Roberson slabsize++; /* In case there it's rounded */ 13438355f576SJeff Roberson slabsize += sizeof(struct uma_slab); 13448355f576SJeff Roberson 13458355f576SJeff Roberson /* Now make a zone for slab headers */ 13468355f576SJeff Roberson slabzone = uma_zcreate("UMA Slabs", 13478355f576SJeff Roberson slabsize, 13488355f576SJeff Roberson NULL, NULL, NULL, NULL, 13498355f576SJeff Roberson UMA_ALIGN_PTR, UMA_ZONE_INTERNAL); 13508355f576SJeff Roberson 13518355f576SJeff Roberson hashzone = uma_zcreate("UMA Hash", 13528355f576SJeff Roberson sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, 13538355f576SJeff Roberson NULL, NULL, NULL, NULL, 13548355f576SJeff Roberson UMA_ALIGN_PTR, UMA_ZONE_INTERNAL); 13558355f576SJeff Roberson 1356cae33c14SJeff Roberson bucket_init(); 13578355f576SJeff Roberson 135848eea375SJeff Roberson #ifdef UMA_MD_SMALL_ALLOC 135948eea375SJeff Roberson booted = 1; 136048eea375SJeff Roberson #endif 13618355f576SJeff Roberson 13628355f576SJeff Roberson #ifdef UMA_DEBUG 13638355f576SJeff Roberson printf("UMA startup complete.\n"); 13648355f576SJeff Roberson #endif 13658355f576SJeff Roberson } 13668355f576SJeff Roberson 13678355f576SJeff Roberson /* see uma.h */ 13688355f576SJeff Roberson void 136999571dc3SJeff Roberson uma_startup2(void) 13708355f576SJeff Roberson { 13718355f576SJeff Roberson booted = 1; 137286bbae32SJeff Roberson bucket_enable(); 13738355f576SJeff Roberson #ifdef UMA_DEBUG 13748355f576SJeff Roberson printf("UMA startup2 complete.\n"); 13758355f576SJeff Roberson #endif 13768355f576SJeff Roberson } 13778355f576SJeff Roberson 13788355f576SJeff Roberson /* 13798355f576SJeff Roberson * Initialize our callout handle 13808355f576SJeff Roberson * 13818355f576SJeff Roberson */ 13828355f576SJeff Roberson 13838355f576SJeff Roberson static void 13848355f576SJeff Roberson uma_startup3(void) 13858355f576SJeff Roberson { 13868355f576SJeff Roberson #ifdef UMA_DEBUG 13878355f576SJeff Roberson printf("Starting callout.\n"); 13888355f576SJeff Roberson #endif 13898355f576SJeff Roberson callout_init(&uma_callout, 0); 13908355f576SJeff Roberson callout_reset(&uma_callout, UMA_WORKING_TIME * hz, uma_timeout, NULL); 13918355f576SJeff Roberson #ifdef UMA_DEBUG 13928355f576SJeff Roberson printf("UMA startup3 complete.\n"); 13938355f576SJeff Roberson #endif 13948355f576SJeff Roberson } 13958355f576SJeff Roberson 13968355f576SJeff Roberson /* See uma.h */ 13978355f576SJeff Roberson uma_zone_t 1398c3bdc05fSAndrew R. Reiter uma_zcreate(char *name, size_t size, uma_ctor ctor, uma_dtor dtor, 1399c3bdc05fSAndrew R. Reiter uma_init uminit, uma_fini fini, int align, u_int16_t flags) 14008355f576SJeff Roberson 14018355f576SJeff Roberson { 14028355f576SJeff Roberson struct uma_zctor_args args; 14038355f576SJeff Roberson 14048355f576SJeff Roberson /* This stuff is essential for the zone ctor */ 14058355f576SJeff Roberson args.name = name; 14068355f576SJeff Roberson args.size = size; 14078355f576SJeff Roberson args.ctor = ctor; 14088355f576SJeff Roberson args.dtor = dtor; 14098355f576SJeff Roberson args.uminit = uminit; 14108355f576SJeff Roberson args.fini = fini; 14118355f576SJeff Roberson args.align = align; 14128355f576SJeff Roberson args.flags = flags; 14138355f576SJeff Roberson 1414a163d034SWarner Losh return (uma_zalloc_internal(zones, &args, M_WAITOK)); 14158355f576SJeff Roberson } 14168355f576SJeff Roberson 14178355f576SJeff Roberson /* See uma.h */ 14189c2cd7e5SJeff Roberson void 14199c2cd7e5SJeff Roberson uma_zdestroy(uma_zone_t zone) 14209c2cd7e5SJeff Roberson { 14219c2cd7e5SJeff Roberson uma_zfree_internal(zones, zone, NULL, 0); 14229c2cd7e5SJeff Roberson } 14239c2cd7e5SJeff Roberson 14249c2cd7e5SJeff Roberson /* See uma.h */ 14258355f576SJeff Roberson void * 14262cc35ff9SJeff Roberson uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) 14278355f576SJeff Roberson { 14288355f576SJeff Roberson void *item; 14298355f576SJeff Roberson uma_cache_t cache; 14308355f576SJeff Roberson uma_bucket_t bucket; 14318355f576SJeff Roberson int cpu; 14328355f576SJeff Roberson 14338355f576SJeff Roberson /* This is the fast path allocation */ 14348355f576SJeff Roberson #ifdef UMA_DEBUG_ALLOC_1 14358355f576SJeff Roberson printf("Allocating one item from %s(%p)\n", zone->uz_name, zone); 14368355f576SJeff Roberson #endif 1437a553d4b8SJeff Roberson 14388522511bSHartmut Brandt #ifdef INVARIANTS 14398522511bSHartmut Brandt /* 14408522511bSHartmut Brandt * To make sure that WAITOK or NOWAIT is set, but not more than 14418522511bSHartmut Brandt * one, and check against the API botches that are common. 14428522511bSHartmut Brandt * The uma code implies M_WAITOK if M_NOWAIT is not set, so 14438522511bSHartmut Brandt * we default to waiting if none of the flags is set. 14448522511bSHartmut Brandt */ 14458522511bSHartmut Brandt cpu = flags & (M_WAITOK | M_NOWAIT | M_DONTWAIT | M_TRYWAIT); 14468522511bSHartmut Brandt if (cpu != M_NOWAIT && cpu != M_WAITOK) { 14478522511bSHartmut Brandt static struct timeval lasterr; 14488522511bSHartmut Brandt static int curerr, once; 14498522511bSHartmut Brandt if (once == 0 && ppsratecheck(&lasterr, &curerr, 1)) { 14508522511bSHartmut Brandt printf("Bad uma_zalloc flags: %x\n", cpu); 14518522511bSHartmut Brandt backtrace(); 14528522511bSHartmut Brandt once++; 14538522511bSHartmut Brandt } 14548522511bSHartmut Brandt } 14558522511bSHartmut Brandt #endif 14568522511bSHartmut Brandt 14574c1cc01cSJohn Baldwin if (!(flags & M_NOWAIT)) { 14584c1cc01cSJohn Baldwin KASSERT(curthread->td_intr_nesting_level == 0, 1459a163d034SWarner Losh ("malloc(M_WAITOK) in interrupt context")); 146026306795SJohn Baldwin WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 146126306795SJohn Baldwin "malloc() of \"%s\"", zone->uz_name); 14624c1cc01cSJohn Baldwin } 14634c1cc01cSJohn Baldwin 1464a553d4b8SJeff Roberson zalloc_restart: 14658355f576SJeff Roberson cpu = PCPU_GET(cpuid); 1466d88797c2SBosko Milekic CPU_LOCK(cpu); 14678355f576SJeff Roberson cache = &zone->uz_cpu[cpu]; 14688355f576SJeff Roberson 14698355f576SJeff Roberson zalloc_start: 14708355f576SJeff Roberson bucket = cache->uc_allocbucket; 14718355f576SJeff Roberson 14728355f576SJeff Roberson if (bucket) { 1473cae33c14SJeff Roberson if (bucket->ub_cnt > 0) { 1474cae33c14SJeff Roberson bucket->ub_cnt--; 1475cae33c14SJeff Roberson item = bucket->ub_bucket[bucket->ub_cnt]; 14768355f576SJeff Roberson #ifdef INVARIANTS 1477cae33c14SJeff Roberson bucket->ub_bucket[bucket->ub_cnt] = NULL; 14788355f576SJeff Roberson #endif 14798355f576SJeff Roberson KASSERT(item != NULL, 14808355f576SJeff Roberson ("uma_zalloc: Bucket pointer mangled.")); 14818355f576SJeff Roberson cache->uc_allocs++; 1482639c9550SJeff Roberson #ifdef INVARIANTS 148381f71edaSMatt Jacob ZONE_LOCK(zone); 1484639c9550SJeff Roberson uma_dbg_alloc(zone, NULL, item); 148581f71edaSMatt Jacob ZONE_UNLOCK(zone); 1486639c9550SJeff Roberson #endif 1487d88797c2SBosko Milekic CPU_UNLOCK(cpu); 14888355f576SJeff Roberson if (zone->uz_ctor) 14898355f576SJeff Roberson zone->uz_ctor(item, zone->uz_size, udata); 14902cc35ff9SJeff Roberson if (flags & M_ZERO) 14912cc35ff9SJeff Roberson bzero(item, zone->uz_size); 14928355f576SJeff Roberson return (item); 14938355f576SJeff Roberson } else if (cache->uc_freebucket) { 14948355f576SJeff Roberson /* 14958355f576SJeff Roberson * We have run out of items in our allocbucket. 14968355f576SJeff Roberson * See if we can switch with our free bucket. 14978355f576SJeff Roberson */ 1498cae33c14SJeff Roberson if (cache->uc_freebucket->ub_cnt > 0) { 14998355f576SJeff Roberson uma_bucket_t swap; 15008355f576SJeff Roberson 15018355f576SJeff Roberson #ifdef UMA_DEBUG_ALLOC 15028355f576SJeff Roberson printf("uma_zalloc: Swapping empty with alloc.\n"); 15038355f576SJeff Roberson #endif 15048355f576SJeff Roberson swap = cache->uc_freebucket; 15058355f576SJeff Roberson cache->uc_freebucket = cache->uc_allocbucket; 15068355f576SJeff Roberson cache->uc_allocbucket = swap; 15078355f576SJeff Roberson 15088355f576SJeff Roberson goto zalloc_start; 15098355f576SJeff Roberson } 15108355f576SJeff Roberson } 15118355f576SJeff Roberson } 1512a553d4b8SJeff Roberson ZONE_LOCK(zone); 1513a553d4b8SJeff Roberson /* Since we have locked the zone we may as well send back our stats */ 1514a553d4b8SJeff Roberson zone->uz_allocs += cache->uc_allocs; 1515a553d4b8SJeff Roberson cache->uc_allocs = 0; 15168355f576SJeff Roberson 1517a553d4b8SJeff Roberson /* Our old one is now a free bucket */ 1518a553d4b8SJeff Roberson if (cache->uc_allocbucket) { 1519cae33c14SJeff Roberson KASSERT(cache->uc_allocbucket->ub_cnt == 0, 1520a553d4b8SJeff Roberson ("uma_zalloc_arg: Freeing a non free bucket.")); 1521a553d4b8SJeff Roberson LIST_INSERT_HEAD(&zone->uz_free_bucket, 1522a553d4b8SJeff Roberson cache->uc_allocbucket, ub_link); 1523a553d4b8SJeff Roberson cache->uc_allocbucket = NULL; 1524a553d4b8SJeff Roberson } 15258355f576SJeff Roberson 1526a553d4b8SJeff Roberson /* Check the free list for a new alloc bucket */ 1527a553d4b8SJeff Roberson if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) { 1528cae33c14SJeff Roberson KASSERT(bucket->ub_cnt != 0, 1529a553d4b8SJeff Roberson ("uma_zalloc_arg: Returning an empty bucket.")); 15308355f576SJeff Roberson 1531a553d4b8SJeff Roberson LIST_REMOVE(bucket, ub_link); 1532a553d4b8SJeff Roberson cache->uc_allocbucket = bucket; 1533a553d4b8SJeff Roberson ZONE_UNLOCK(zone); 15348355f576SJeff Roberson goto zalloc_start; 1535a553d4b8SJeff Roberson } 1536bbee39c6SJeff Roberson /* We are no longer associated with this cpu!!! */ 1537d88797c2SBosko Milekic CPU_UNLOCK(cpu); 1538bbee39c6SJeff Roberson 1539a553d4b8SJeff Roberson /* Bump up our uz_count so we get here less */ 1540cae33c14SJeff Roberson if (zone->uz_count < BUCKET_MAX) 1541a553d4b8SJeff Roberson zone->uz_count++; 1542a553d4b8SJeff Roberson 15438355f576SJeff Roberson /* 1544a553d4b8SJeff Roberson * Now lets just fill a bucket and put it on the free list. If that 1545a553d4b8SJeff Roberson * works we'll restart the allocation from the begining. 1546bbee39c6SJeff Roberson */ 1547bbee39c6SJeff Roberson 1548bbee39c6SJeff Roberson if (uma_zalloc_bucket(zone, flags)) { 1549bbee39c6SJeff Roberson ZONE_UNLOCK(zone); 1550bbee39c6SJeff Roberson goto zalloc_restart; 1551bbee39c6SJeff Roberson } 1552bbee39c6SJeff Roberson ZONE_UNLOCK(zone); 1553bbee39c6SJeff Roberson /* 1554bbee39c6SJeff Roberson * We may not be able to get a bucket so return an actual item. 1555bbee39c6SJeff Roberson */ 1556bbee39c6SJeff Roberson #ifdef UMA_DEBUG 1557bbee39c6SJeff Roberson printf("uma_zalloc_arg: Bucketzone returned NULL\n"); 1558bbee39c6SJeff Roberson #endif 1559bbee39c6SJeff Roberson 1560bbee39c6SJeff Roberson return (uma_zalloc_internal(zone, udata, flags)); 1561bbee39c6SJeff Roberson } 1562bbee39c6SJeff Roberson 1563bbee39c6SJeff Roberson static uma_slab_t 1564bbee39c6SJeff Roberson uma_zone_slab(uma_zone_t zone, int flags) 1565bbee39c6SJeff Roberson { 1566bbee39c6SJeff Roberson uma_slab_t slab; 1567bbee39c6SJeff Roberson 1568bbee39c6SJeff Roberson /* 1569bbee39c6SJeff Roberson * This is to prevent us from recursively trying to allocate 1570bbee39c6SJeff Roberson * buckets. The problem is that if an allocation forces us to 1571bbee39c6SJeff Roberson * grab a new bucket we will call page_alloc, which will go off 1572bbee39c6SJeff Roberson * and cause the vm to allocate vm_map_entries. If we need new 1573bbee39c6SJeff Roberson * buckets there too we will recurse in kmem_alloc and bad 1574bbee39c6SJeff Roberson * things happen. So instead we return a NULL bucket, and make 1575bbee39c6SJeff Roberson * the code that allocates buckets smart enough to deal with it 1576bbee39c6SJeff Roberson */ 1577cae33c14SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_INTERNAL && zone->uz_recurse != 0) 1578bbee39c6SJeff Roberson return (NULL); 1579bbee39c6SJeff Roberson 1580bbee39c6SJeff Roberson slab = NULL; 1581bbee39c6SJeff Roberson 1582bbee39c6SJeff Roberson for (;;) { 1583bbee39c6SJeff Roberson /* 1584bbee39c6SJeff Roberson * Find a slab with some space. Prefer slabs that are partially 1585bbee39c6SJeff Roberson * used over those that are totally full. This helps to reduce 1586bbee39c6SJeff Roberson * fragmentation. 1587bbee39c6SJeff Roberson */ 1588bbee39c6SJeff Roberson if (zone->uz_free != 0) { 1589bbee39c6SJeff Roberson if (!LIST_EMPTY(&zone->uz_part_slab)) { 1590bbee39c6SJeff Roberson slab = LIST_FIRST(&zone->uz_part_slab); 1591bbee39c6SJeff Roberson } else { 1592bbee39c6SJeff Roberson slab = LIST_FIRST(&zone->uz_free_slab); 1593bbee39c6SJeff Roberson LIST_REMOVE(slab, us_link); 1594bbee39c6SJeff Roberson LIST_INSERT_HEAD(&zone->uz_part_slab, slab, 1595bbee39c6SJeff Roberson us_link); 1596bbee39c6SJeff Roberson } 1597bbee39c6SJeff Roberson return (slab); 1598bbee39c6SJeff Roberson } 1599bbee39c6SJeff Roberson 1600bbee39c6SJeff Roberson /* 1601bbee39c6SJeff Roberson * M_NOVM means don't ask at all! 1602bbee39c6SJeff Roberson */ 1603bbee39c6SJeff Roberson if (flags & M_NOVM) 1604bbee39c6SJeff Roberson break; 1605bbee39c6SJeff Roberson 1606bbee39c6SJeff Roberson if (zone->uz_maxpages && 1607bbee39c6SJeff Roberson zone->uz_pages >= zone->uz_maxpages) { 1608bbee39c6SJeff Roberson zone->uz_flags |= UMA_ZFLAG_FULL; 1609bbee39c6SJeff Roberson 1610ebc85edfSJeff Roberson if (flags & M_NOWAIT) 1611bbee39c6SJeff Roberson break; 1612ebc85edfSJeff Roberson else 1613ebc85edfSJeff Roberson msleep(zone, &zone->uz_lock, PVM, "zonelimit", 0); 1614bbee39c6SJeff Roberson continue; 1615bbee39c6SJeff Roberson } 1616bbee39c6SJeff Roberson zone->uz_recurse++; 1617bbee39c6SJeff Roberson slab = slab_zalloc(zone, flags); 1618bbee39c6SJeff Roberson zone->uz_recurse--; 1619bbee39c6SJeff Roberson /* 1620bbee39c6SJeff Roberson * If we got a slab here it's safe to mark it partially used 1621bbee39c6SJeff Roberson * and return. We assume that the caller is going to remove 1622bbee39c6SJeff Roberson * at least one item. 1623bbee39c6SJeff Roberson */ 1624bbee39c6SJeff Roberson if (slab) { 1625bbee39c6SJeff Roberson LIST_INSERT_HEAD(&zone->uz_part_slab, slab, us_link); 1626bbee39c6SJeff Roberson return (slab); 1627bbee39c6SJeff Roberson } 1628bbee39c6SJeff Roberson /* 1629bbee39c6SJeff Roberson * We might not have been able to get a slab but another cpu 1630bbee39c6SJeff Roberson * could have while we were unlocked. Check again before we 1631bbee39c6SJeff Roberson * fail. 1632bbee39c6SJeff Roberson */ 1633ebc85edfSJeff Roberson if (flags & M_NOWAIT) 1634bbee39c6SJeff Roberson flags |= M_NOVM; 1635bbee39c6SJeff Roberson } 1636bbee39c6SJeff Roberson return (slab); 1637bbee39c6SJeff Roberson } 1638bbee39c6SJeff Roberson 1639d56368d7SBosko Milekic static void * 1640bbee39c6SJeff Roberson uma_slab_alloc(uma_zone_t zone, uma_slab_t slab) 1641bbee39c6SJeff Roberson { 1642bbee39c6SJeff Roberson void *item; 1643bbee39c6SJeff Roberson u_int8_t freei; 1644bbee39c6SJeff Roberson 1645bbee39c6SJeff Roberson freei = slab->us_firstfree; 1646bbee39c6SJeff Roberson slab->us_firstfree = slab->us_freelist[freei]; 1647bbee39c6SJeff Roberson item = slab->us_data + (zone->uz_rsize * freei); 1648bbee39c6SJeff Roberson 1649bbee39c6SJeff Roberson slab->us_freecount--; 1650bbee39c6SJeff Roberson zone->uz_free--; 1651bbee39c6SJeff Roberson #ifdef INVARIANTS 1652bbee39c6SJeff Roberson uma_dbg_alloc(zone, slab, item); 1653bbee39c6SJeff Roberson #endif 1654bbee39c6SJeff Roberson /* Move this slab to the full list */ 1655bbee39c6SJeff Roberson if (slab->us_freecount == 0) { 1656bbee39c6SJeff Roberson LIST_REMOVE(slab, us_link); 1657bbee39c6SJeff Roberson LIST_INSERT_HEAD(&zone->uz_full_slab, slab, us_link); 1658bbee39c6SJeff Roberson } 1659bbee39c6SJeff Roberson 1660bbee39c6SJeff Roberson return (item); 1661bbee39c6SJeff Roberson } 1662bbee39c6SJeff Roberson 1663bbee39c6SJeff Roberson static int 1664bbee39c6SJeff Roberson uma_zalloc_bucket(uma_zone_t zone, int flags) 1665bbee39c6SJeff Roberson { 1666bbee39c6SJeff Roberson uma_bucket_t bucket; 1667bbee39c6SJeff Roberson uma_slab_t slab; 1668bbee39c6SJeff Roberson 1669bbee39c6SJeff Roberson /* 1670a553d4b8SJeff Roberson * Try this zone's free list first so we don't allocate extra buckets. 16718355f576SJeff Roberson */ 1672a553d4b8SJeff Roberson 1673bbee39c6SJeff Roberson if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) { 1674cae33c14SJeff Roberson KASSERT(bucket->ub_cnt == 0, 1675bbee39c6SJeff Roberson ("uma_zalloc_bucket: Bucket on free list is not empty.")); 1676a553d4b8SJeff Roberson LIST_REMOVE(bucket, ub_link); 1677bbee39c6SJeff Roberson } else { 167818aa2de5SJeff Roberson int bflags; 167918aa2de5SJeff Roberson 1680cae33c14SJeff Roberson bflags = (flags & ~M_ZERO); 168120e8e865SBosko Milekic if (zone->uz_flags & UMA_ZFLAG_CACHEONLY) 168218aa2de5SJeff Roberson bflags |= M_NOVM; 168318aa2de5SJeff Roberson 1684bbee39c6SJeff Roberson ZONE_UNLOCK(zone); 1685cae33c14SJeff Roberson bucket = bucket_alloc(zone->uz_count, bflags); 1686bbee39c6SJeff Roberson ZONE_LOCK(zone); 1687bbee39c6SJeff Roberson } 1688bbee39c6SJeff Roberson 1689bbee39c6SJeff Roberson if (bucket == NULL) 1690bbee39c6SJeff Roberson return (0); 1691bbee39c6SJeff Roberson 1692bbee39c6SJeff Roberson #ifdef SMP 1693a553d4b8SJeff Roberson /* 1694bbee39c6SJeff Roberson * This code is here to limit the number of simultaneous bucket fills 1695bbee39c6SJeff Roberson * for any given zone to the number of per cpu caches in this zone. This 1696bbee39c6SJeff Roberson * is done so that we don't allocate more memory than we really need. 1697a553d4b8SJeff Roberson */ 1698bbee39c6SJeff Roberson if (zone->uz_fills >= mp_ncpus) 1699bbee39c6SJeff Roberson goto done; 1700a553d4b8SJeff Roberson 1701bbee39c6SJeff Roberson #endif 1702bbee39c6SJeff Roberson zone->uz_fills++; 1703bbee39c6SJeff Roberson 1704bbee39c6SJeff Roberson /* Try to keep the buckets totally full */ 1705cae33c14SJeff Roberson while (bucket->ub_cnt < bucket->ub_entries && 1706d11e0ba5SJeff Roberson (slab = uma_zone_slab(zone, flags)) != NULL) { 1707bbee39c6SJeff Roberson while (slab->us_freecount && 1708cae33c14SJeff Roberson bucket->ub_cnt < bucket->ub_entries) { 1709cae33c14SJeff Roberson bucket->ub_bucket[bucket->ub_cnt++] = 1710bbee39c6SJeff Roberson uma_slab_alloc(zone, slab); 1711bbee39c6SJeff Roberson } 1712bbee39c6SJeff Roberson /* Don't block on the next fill */ 1713bbee39c6SJeff Roberson flags |= M_NOWAIT; 17148355f576SJeff Roberson } 17158355f576SJeff Roberson 1716bbee39c6SJeff Roberson zone->uz_fills--; 1717bbee39c6SJeff Roberson 1718cae33c14SJeff Roberson if (bucket->ub_cnt != 0) { 1719bbee39c6SJeff Roberson LIST_INSERT_HEAD(&zone->uz_full_bucket, 1720bbee39c6SJeff Roberson bucket, ub_link); 1721bbee39c6SJeff Roberson return (1); 1722bbee39c6SJeff Roberson } 1723bbee39c6SJeff Roberson #ifdef SMP 1724bbee39c6SJeff Roberson done: 1725bbee39c6SJeff Roberson #endif 1726cae33c14SJeff Roberson bucket_free(bucket); 1727bbee39c6SJeff Roberson 1728bbee39c6SJeff Roberson return (0); 1729bbee39c6SJeff Roberson } 17308355f576SJeff Roberson /* 1731bbee39c6SJeff Roberson * Allocates an item for an internal zone 17328355f576SJeff Roberson * 17338355f576SJeff Roberson * Arguments 17348355f576SJeff Roberson * zone The zone to alloc for. 17358355f576SJeff Roberson * udata The data to be passed to the constructor. 1736a163d034SWarner Losh * flags M_WAITOK, M_NOWAIT, M_ZERO. 17378355f576SJeff Roberson * 17388355f576SJeff Roberson * Returns 17398355f576SJeff Roberson * NULL if there is no memory and M_NOWAIT is set 1740bbee39c6SJeff Roberson * An item if successful 17418355f576SJeff Roberson */ 17428355f576SJeff Roberson 17438355f576SJeff Roberson static void * 1744bbee39c6SJeff Roberson uma_zalloc_internal(uma_zone_t zone, void *udata, int flags) 17458355f576SJeff Roberson { 17468355f576SJeff Roberson uma_slab_t slab; 17478355f576SJeff Roberson void *item; 17488355f576SJeff Roberson 17498355f576SJeff Roberson item = NULL; 17508355f576SJeff Roberson 17518355f576SJeff Roberson #ifdef UMA_DEBUG_ALLOC 17528355f576SJeff Roberson printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone); 17538355f576SJeff Roberson #endif 17548355f576SJeff Roberson ZONE_LOCK(zone); 17558355f576SJeff Roberson 1756bbee39c6SJeff Roberson slab = uma_zone_slab(zone, flags); 1757bbee39c6SJeff Roberson if (slab == NULL) { 1758bce97791SJeff Roberson ZONE_UNLOCK(zone); 1759a553d4b8SJeff Roberson return (NULL); 1760bce97791SJeff Roberson } 1761a553d4b8SJeff Roberson 1762bbee39c6SJeff Roberson item = uma_slab_alloc(zone, slab); 17638355f576SJeff Roberson 17648355f576SJeff Roberson ZONE_UNLOCK(zone); 17658355f576SJeff Roberson 17663370c5bfSJeff Roberson if (zone->uz_ctor != NULL) 17678355f576SJeff Roberson zone->uz_ctor(item, zone->uz_size, udata); 17682cc35ff9SJeff Roberson if (flags & M_ZERO) 17692cc35ff9SJeff Roberson bzero(item, zone->uz_size); 17708355f576SJeff Roberson 17718355f576SJeff Roberson return (item); 17728355f576SJeff Roberson } 17738355f576SJeff Roberson 17748355f576SJeff Roberson /* See uma.h */ 17758355f576SJeff Roberson void 17768355f576SJeff Roberson uma_zfree_arg(uma_zone_t zone, void *item, void *udata) 17778355f576SJeff Roberson { 17788355f576SJeff Roberson uma_cache_t cache; 17798355f576SJeff Roberson uma_bucket_t bucket; 17804741dcbfSJeff Roberson int bflags; 17818355f576SJeff Roberson int cpu; 17825c133dfaSBosko Milekic int skip; 17838355f576SJeff Roberson 17848355f576SJeff Roberson /* This is the fast path free */ 17855c133dfaSBosko Milekic skip = 0; 17868355f576SJeff Roberson #ifdef UMA_DEBUG_ALLOC_1 17878355f576SJeff Roberson printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone); 17888355f576SJeff Roberson #endif 1789af7f9b97SJeff Roberson /* 1790af7f9b97SJeff Roberson * The race here is acceptable. If we miss it we'll just have to wait 1791af7f9b97SJeff Roberson * a little longer for the limits to be reset. 1792af7f9b97SJeff Roberson */ 1793af7f9b97SJeff Roberson 1794af7f9b97SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_FULL) 1795af7f9b97SJeff Roberson goto zfree_internal; 1796af7f9b97SJeff Roberson 17975c133dfaSBosko Milekic if (zone->uz_dtor) { 1798bba739abSJeff Roberson zone->uz_dtor(item, zone->uz_size, udata); 17995c133dfaSBosko Milekic skip = 1; 18005c133dfaSBosko Milekic } 1801bba739abSJeff Roberson 1802a553d4b8SJeff Roberson zfree_restart: 18038355f576SJeff Roberson cpu = PCPU_GET(cpuid); 1804d88797c2SBosko Milekic CPU_LOCK(cpu); 18058355f576SJeff Roberson cache = &zone->uz_cpu[cpu]; 18068355f576SJeff Roberson 18078355f576SJeff Roberson zfree_start: 18088355f576SJeff Roberson bucket = cache->uc_freebucket; 18098355f576SJeff Roberson 18108355f576SJeff Roberson if (bucket) { 1811a553d4b8SJeff Roberson /* 1812a553d4b8SJeff Roberson * Do we have room in our bucket? It is OK for this uz count 1813a553d4b8SJeff Roberson * check to be slightly out of sync. 1814a553d4b8SJeff Roberson */ 1815a553d4b8SJeff Roberson 1816cae33c14SJeff Roberson if (bucket->ub_cnt < bucket->ub_entries) { 1817cae33c14SJeff Roberson KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, 18188355f576SJeff Roberson ("uma_zfree: Freeing to non free bucket index.")); 1819cae33c14SJeff Roberson bucket->ub_bucket[bucket->ub_cnt] = item; 1820cae33c14SJeff Roberson bucket->ub_cnt++; 1821b9ba8931SJeff Roberson #ifdef INVARIANTS 182281f71edaSMatt Jacob ZONE_LOCK(zone); 1823b9ba8931SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_MALLOC) 1824b9ba8931SJeff Roberson uma_dbg_free(zone, udata, item); 1825b9ba8931SJeff Roberson else 1826b9ba8931SJeff Roberson uma_dbg_free(zone, NULL, item); 182781f71edaSMatt Jacob ZONE_UNLOCK(zone); 1828b9ba8931SJeff Roberson #endif 1829d88797c2SBosko Milekic CPU_UNLOCK(cpu); 18308355f576SJeff Roberson return; 18318355f576SJeff Roberson } else if (cache->uc_allocbucket) { 18328355f576SJeff Roberson #ifdef UMA_DEBUG_ALLOC 18338355f576SJeff Roberson printf("uma_zfree: Swapping buckets.\n"); 18348355f576SJeff Roberson #endif 18358355f576SJeff Roberson /* 18368355f576SJeff Roberson * We have run out of space in our freebucket. 18378355f576SJeff Roberson * See if we can switch with our alloc bucket. 18388355f576SJeff Roberson */ 1839cae33c14SJeff Roberson if (cache->uc_allocbucket->ub_cnt < 1840cae33c14SJeff Roberson cache->uc_freebucket->ub_cnt) { 18418355f576SJeff Roberson uma_bucket_t swap; 18428355f576SJeff Roberson 18438355f576SJeff Roberson swap = cache->uc_freebucket; 18448355f576SJeff Roberson cache->uc_freebucket = cache->uc_allocbucket; 18458355f576SJeff Roberson cache->uc_allocbucket = swap; 18468355f576SJeff Roberson 18478355f576SJeff Roberson goto zfree_start; 18488355f576SJeff Roberson } 18498355f576SJeff Roberson } 18508355f576SJeff Roberson } 18518355f576SJeff Roberson 18528355f576SJeff Roberson /* 1853a553d4b8SJeff Roberson * We can get here for two reasons: 18548355f576SJeff Roberson * 18558355f576SJeff Roberson * 1) The buckets are NULL 1856a553d4b8SJeff Roberson * 2) The alloc and free buckets are both somewhat full. 18578355f576SJeff Roberson * 18588355f576SJeff Roberson */ 18598355f576SJeff Roberson 18608355f576SJeff Roberson ZONE_LOCK(zone); 18618355f576SJeff Roberson 18628355f576SJeff Roberson bucket = cache->uc_freebucket; 18638355f576SJeff Roberson cache->uc_freebucket = NULL; 18648355f576SJeff Roberson 18658355f576SJeff Roberson /* Can we throw this on the zone full list? */ 18668355f576SJeff Roberson if (bucket != NULL) { 18678355f576SJeff Roberson #ifdef UMA_DEBUG_ALLOC 18688355f576SJeff Roberson printf("uma_zfree: Putting old bucket on the free list.\n"); 18698355f576SJeff Roberson #endif 1870cae33c14SJeff Roberson /* ub_cnt is pointing to the last free item */ 1871cae33c14SJeff Roberson KASSERT(bucket->ub_cnt != 0, 18728355f576SJeff Roberson ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); 18738355f576SJeff Roberson LIST_INSERT_HEAD(&zone->uz_full_bucket, 18748355f576SJeff Roberson bucket, ub_link); 18758355f576SJeff Roberson } 1876a553d4b8SJeff Roberson if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) { 1877a553d4b8SJeff Roberson LIST_REMOVE(bucket, ub_link); 1878a553d4b8SJeff Roberson ZONE_UNLOCK(zone); 1879a553d4b8SJeff Roberson cache->uc_freebucket = bucket; 1880a553d4b8SJeff Roberson goto zfree_start; 1881a553d4b8SJeff Roberson } 1882a553d4b8SJeff Roberson /* We're done with this CPU now */ 1883d88797c2SBosko Milekic CPU_UNLOCK(cpu); 1884a553d4b8SJeff Roberson 1885a553d4b8SJeff Roberson /* And the zone.. */ 1886a553d4b8SJeff Roberson ZONE_UNLOCK(zone); 1887a553d4b8SJeff Roberson 18888355f576SJeff Roberson #ifdef UMA_DEBUG_ALLOC 18898355f576SJeff Roberson printf("uma_zfree: Allocating new free bucket.\n"); 18908355f576SJeff Roberson #endif 18914741dcbfSJeff Roberson bflags = M_NOWAIT; 18924741dcbfSJeff Roberson 189320e8e865SBosko Milekic if (zone->uz_flags & UMA_ZFLAG_CACHEONLY) 18944741dcbfSJeff Roberson bflags |= M_NOVM; 1895cae33c14SJeff Roberson bucket = bucket_alloc(zone->uz_count, bflags); 18964741dcbfSJeff Roberson if (bucket) { 1897a553d4b8SJeff Roberson ZONE_LOCK(zone); 1898a553d4b8SJeff Roberson LIST_INSERT_HEAD(&zone->uz_free_bucket, 1899a553d4b8SJeff Roberson bucket, ub_link); 19008355f576SJeff Roberson ZONE_UNLOCK(zone); 1901a553d4b8SJeff Roberson goto zfree_restart; 19028355f576SJeff Roberson } 19038355f576SJeff Roberson 1904a553d4b8SJeff Roberson /* 1905a553d4b8SJeff Roberson * If nothing else caught this, we'll just do an internal free. 1906a553d4b8SJeff Roberson */ 19078355f576SJeff Roberson 1908af7f9b97SJeff Roberson zfree_internal: 1909af7f9b97SJeff Roberson 191048bf8725SBosko Milekic #ifdef INVARIANTS 191148bf8725SBosko Milekic /* 191248bf8725SBosko Milekic * If we need to skip the dtor and the uma_dbg_free in uma_zfree_internal 191348bf8725SBosko Milekic * because we've already called the dtor above, but we ended up here, then 191448bf8725SBosko Milekic * we need to make sure that we take care of the uma_dbg_free immediately. 191548bf8725SBosko Milekic */ 191648bf8725SBosko Milekic if (skip) { 191748bf8725SBosko Milekic ZONE_LOCK(zone); 191848bf8725SBosko Milekic if (zone->uz_flags & UMA_ZFLAG_MALLOC) 191948bf8725SBosko Milekic uma_dbg_free(zone, udata, item); 192048bf8725SBosko Milekic else 192148bf8725SBosko Milekic uma_dbg_free(zone, NULL, item); 192248bf8725SBosko Milekic ZONE_UNLOCK(zone); 192348bf8725SBosko Milekic } 192448bf8725SBosko Milekic #endif 19255c133dfaSBosko Milekic uma_zfree_internal(zone, item, udata, skip); 19268355f576SJeff Roberson 19278355f576SJeff Roberson return; 19288355f576SJeff Roberson 19298355f576SJeff Roberson } 19308355f576SJeff Roberson 19318355f576SJeff Roberson /* 19328355f576SJeff Roberson * Frees an item to an INTERNAL zone or allocates a free bucket 19338355f576SJeff Roberson * 19348355f576SJeff Roberson * Arguments: 19358355f576SJeff Roberson * zone The zone to free to 19368355f576SJeff Roberson * item The item we're freeing 19378355f576SJeff Roberson * udata User supplied data for the dtor 19388355f576SJeff Roberson * skip Skip the dtor, it was done in uma_zfree_arg 19398355f576SJeff Roberson */ 19408355f576SJeff Roberson 19418355f576SJeff Roberson static void 19428355f576SJeff Roberson uma_zfree_internal(uma_zone_t zone, void *item, void *udata, int skip) 19438355f576SJeff Roberson { 19448355f576SJeff Roberson uma_slab_t slab; 19458355f576SJeff Roberson u_int8_t *mem; 19468355f576SJeff Roberson u_int8_t freei; 19478355f576SJeff Roberson 1948bba739abSJeff Roberson if (!skip && zone->uz_dtor) 1949bba739abSJeff Roberson zone->uz_dtor(item, zone->uz_size, udata); 1950bba739abSJeff Roberson 19518355f576SJeff Roberson ZONE_LOCK(zone); 19528355f576SJeff Roberson 19538355f576SJeff Roberson if (!(zone->uz_flags & UMA_ZFLAG_MALLOC)) { 19548355f576SJeff Roberson mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK)); 195599571dc3SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_HASH) 19568355f576SJeff Roberson slab = hash_sfind(&zone->uz_hash, mem); 19578355f576SJeff Roberson else { 19588355f576SJeff Roberson mem += zone->uz_pgoff; 19598355f576SJeff Roberson slab = (uma_slab_t)mem; 19608355f576SJeff Roberson } 19618355f576SJeff Roberson } else { 19628355f576SJeff Roberson slab = (uma_slab_t)udata; 19638355f576SJeff Roberson } 19648355f576SJeff Roberson 19658355f576SJeff Roberson /* Do we need to remove from any lists? */ 19668355f576SJeff Roberson if (slab->us_freecount+1 == zone->uz_ipers) { 19678355f576SJeff Roberson LIST_REMOVE(slab, us_link); 19688355f576SJeff Roberson LIST_INSERT_HEAD(&zone->uz_free_slab, slab, us_link); 19698355f576SJeff Roberson } else if (slab->us_freecount == 0) { 19708355f576SJeff Roberson LIST_REMOVE(slab, us_link); 19718355f576SJeff Roberson LIST_INSERT_HEAD(&zone->uz_part_slab, slab, us_link); 19728355f576SJeff Roberson } 19738355f576SJeff Roberson 19748355f576SJeff Roberson /* Slab management stuff */ 19758355f576SJeff Roberson freei = ((unsigned long)item - (unsigned long)slab->us_data) 19768355f576SJeff Roberson / zone->uz_rsize; 19778355f576SJeff Roberson 1978639c9550SJeff Roberson #ifdef INVARIANTS 1979639c9550SJeff Roberson if (!skip) 1980639c9550SJeff Roberson uma_dbg_free(zone, slab, item); 19818355f576SJeff Roberson #endif 1982639c9550SJeff Roberson 19838355f576SJeff Roberson slab->us_freelist[freei] = slab->us_firstfree; 19848355f576SJeff Roberson slab->us_firstfree = freei; 19858355f576SJeff Roberson slab->us_freecount++; 19868355f576SJeff Roberson 19878355f576SJeff Roberson /* Zone statistics */ 19888355f576SJeff Roberson zone->uz_free++; 19898355f576SJeff Roberson 1990af7f9b97SJeff Roberson if (zone->uz_flags & UMA_ZFLAG_FULL) { 1991af7f9b97SJeff Roberson if (zone->uz_pages < zone->uz_maxpages) 1992af7f9b97SJeff Roberson zone->uz_flags &= ~UMA_ZFLAG_FULL; 1993af7f9b97SJeff Roberson 1994af7f9b97SJeff Roberson /* We can handle one more allocation */ 199574c924b5SJeff Roberson wakeup_one(zone); 1996af7f9b97SJeff Roberson } 1997af7f9b97SJeff Roberson 1998605cbd6aSJeff Roberson ZONE_UNLOCK(zone); 19998355f576SJeff Roberson } 20008355f576SJeff Roberson 20018355f576SJeff Roberson /* See uma.h */ 20028355f576SJeff Roberson void 2003736ee590SJeff Roberson uma_zone_set_max(uma_zone_t zone, int nitems) 2004736ee590SJeff Roberson { 2005736ee590SJeff Roberson ZONE_LOCK(zone); 2006736ee590SJeff Roberson if (zone->uz_ppera > 1) 2007af7f9b97SJeff Roberson zone->uz_maxpages = nitems * zone->uz_ppera; 2008736ee590SJeff Roberson else 2009736ee590SJeff Roberson zone->uz_maxpages = nitems / zone->uz_ipers; 201028bc4419SJeff Roberson 2011d4d6aee5SAndrew R. Reiter if (zone->uz_maxpages * zone->uz_ipers < nitems) 2012d4d6aee5SAndrew R. Reiter zone->uz_maxpages++; 201328bc4419SJeff Roberson 2014736ee590SJeff Roberson ZONE_UNLOCK(zone); 2015736ee590SJeff Roberson } 2016736ee590SJeff Roberson 2017736ee590SJeff Roberson /* See uma.h */ 2018736ee590SJeff Roberson void 20198355f576SJeff Roberson uma_zone_set_freef(uma_zone_t zone, uma_free freef) 20208355f576SJeff Roberson { 20218355f576SJeff Roberson ZONE_LOCK(zone); 20228355f576SJeff Roberson 20238355f576SJeff Roberson zone->uz_freef = freef; 20248355f576SJeff Roberson 20258355f576SJeff Roberson ZONE_UNLOCK(zone); 20268355f576SJeff Roberson } 20278355f576SJeff Roberson 20288355f576SJeff Roberson /* See uma.h */ 20298355f576SJeff Roberson void 20308355f576SJeff Roberson uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) 20318355f576SJeff Roberson { 20328355f576SJeff Roberson ZONE_LOCK(zone); 20338355f576SJeff Roberson 20348355f576SJeff Roberson zone->uz_flags |= UMA_ZFLAG_PRIVALLOC; 20358355f576SJeff Roberson zone->uz_allocf = allocf; 20368355f576SJeff Roberson 20378355f576SJeff Roberson ZONE_UNLOCK(zone); 20388355f576SJeff Roberson } 20398355f576SJeff Roberson 20408355f576SJeff Roberson /* See uma.h */ 20418355f576SJeff Roberson int 20428355f576SJeff Roberson uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count) 20438355f576SJeff Roberson { 20448355f576SJeff Roberson int pages; 20458355f576SJeff Roberson vm_offset_t kva; 20468355f576SJeff Roberson 20478355f576SJeff Roberson mtx_lock(&Giant); 20488355f576SJeff Roberson 20498355f576SJeff Roberson pages = count / zone->uz_ipers; 20508355f576SJeff Roberson 20518355f576SJeff Roberson if (pages * zone->uz_ipers < count) 20528355f576SJeff Roberson pages++; 2053a553d4b8SJeff Roberson 20548355f576SJeff Roberson kva = kmem_alloc_pageable(kernel_map, pages * UMA_SLAB_SIZE); 20558355f576SJeff Roberson 2056a553d4b8SJeff Roberson if (kva == 0) { 2057a553d4b8SJeff Roberson mtx_unlock(&Giant); 20588355f576SJeff Roberson return (0); 20598355f576SJeff Roberson } 20608355f576SJeff Roberson 20618355f576SJeff Roberson 2062a553d4b8SJeff Roberson if (obj == NULL) 2063a553d4b8SJeff Roberson obj = vm_object_allocate(OBJT_DEFAULT, 2064c7173f58SJeff Roberson pages); 206582774d80SAlan Cox else { 206682774d80SAlan Cox VM_OBJECT_LOCK_INIT(obj); 20678355f576SJeff Roberson _vm_object_allocate(OBJT_DEFAULT, 2068c7173f58SJeff Roberson pages, obj); 206982774d80SAlan Cox } 2070a553d4b8SJeff Roberson ZONE_LOCK(zone); 2071a553d4b8SJeff Roberson zone->uz_kva = kva; 2072a553d4b8SJeff Roberson zone->uz_obj = obj; 2073a553d4b8SJeff Roberson zone->uz_maxpages = pages; 20748355f576SJeff Roberson 20758355f576SJeff Roberson zone->uz_allocf = obj_alloc; 20768355f576SJeff Roberson zone->uz_flags |= UMA_ZFLAG_NOFREE | UMA_ZFLAG_PRIVALLOC; 20778355f576SJeff Roberson 20788355f576SJeff Roberson ZONE_UNLOCK(zone); 2079a553d4b8SJeff Roberson mtx_unlock(&Giant); 20808355f576SJeff Roberson 20818355f576SJeff Roberson return (1); 20828355f576SJeff Roberson } 20838355f576SJeff Roberson 20848355f576SJeff Roberson /* See uma.h */ 20858355f576SJeff Roberson void 20868355f576SJeff Roberson uma_prealloc(uma_zone_t zone, int items) 20878355f576SJeff Roberson { 20888355f576SJeff Roberson int slabs; 20898355f576SJeff Roberson uma_slab_t slab; 20908355f576SJeff Roberson 20918355f576SJeff Roberson ZONE_LOCK(zone); 20928355f576SJeff Roberson slabs = items / zone->uz_ipers; 20938355f576SJeff Roberson if (slabs * zone->uz_ipers < items) 20948355f576SJeff Roberson slabs++; 20958355f576SJeff Roberson 20968355f576SJeff Roberson while (slabs > 0) { 2097a163d034SWarner Losh slab = slab_zalloc(zone, M_WAITOK); 20988355f576SJeff Roberson LIST_INSERT_HEAD(&zone->uz_free_slab, slab, us_link); 20998355f576SJeff Roberson slabs--; 21008355f576SJeff Roberson } 21018355f576SJeff Roberson ZONE_UNLOCK(zone); 21028355f576SJeff Roberson } 21038355f576SJeff Roberson 21048355f576SJeff Roberson /* See uma.h */ 21058355f576SJeff Roberson void 21068355f576SJeff Roberson uma_reclaim(void) 21078355f576SJeff Roberson { 21088355f576SJeff Roberson /* 21098355f576SJeff Roberson * You might think that the delay below would improve performance since 21108355f576SJeff Roberson * the allocator will give away memory that it may ask for immediately. 21118355f576SJeff Roberson * Really, it makes things worse, since cpu cycles are so much cheaper 21128355f576SJeff Roberson * than disk activity. 21138355f576SJeff Roberson */ 21148355f576SJeff Roberson #if 0 21158355f576SJeff Roberson static struct timeval tv = {0}; 21168355f576SJeff Roberson struct timeval now; 21178355f576SJeff Roberson getmicrouptime(&now); 21188355f576SJeff Roberson if (now.tv_sec > tv.tv_sec + 30) 21198355f576SJeff Roberson tv = now; 21208355f576SJeff Roberson else 21218355f576SJeff Roberson return; 21228355f576SJeff Roberson #endif 21238355f576SJeff Roberson #ifdef UMA_DEBUG 21248355f576SJeff Roberson printf("UMA: vm asked us to release pages!\n"); 21258355f576SJeff Roberson #endif 212686bbae32SJeff Roberson bucket_enable(); 21278355f576SJeff Roberson zone_foreach(zone_drain); 21288355f576SJeff Roberson 21298355f576SJeff Roberson /* 21308355f576SJeff Roberson * Some slabs may have been freed but this zone will be visited early 21318355f576SJeff Roberson * we visit again so that we can free pages that are empty once other 21328355f576SJeff Roberson * zones are drained. We have to do the same for buckets. 21338355f576SJeff Roberson */ 2134cae33c14SJeff Roberson zone_drain_common(slabzone, 0); 2135cae33c14SJeff Roberson bucket_zone_drain(); 21368355f576SJeff Roberson } 21378355f576SJeff Roberson 21388355f576SJeff Roberson void * 21398355f576SJeff Roberson uma_large_malloc(int size, int wait) 21408355f576SJeff Roberson { 21418355f576SJeff Roberson void *mem; 21428355f576SJeff Roberson uma_slab_t slab; 21438355f576SJeff Roberson u_int8_t flags; 21448355f576SJeff Roberson 2145bbee39c6SJeff Roberson slab = uma_zalloc_internal(slabzone, NULL, wait); 21468355f576SJeff Roberson if (slab == NULL) 21478355f576SJeff Roberson return (NULL); 21488355f576SJeff Roberson 21498355f576SJeff Roberson mem = page_alloc(NULL, size, &flags, wait); 21508355f576SJeff Roberson if (mem) { 215199571dc3SJeff Roberson vsetslab((vm_offset_t)mem, slab); 21528355f576SJeff Roberson slab->us_data = mem; 21538355f576SJeff Roberson slab->us_flags = flags | UMA_SLAB_MALLOC; 21548355f576SJeff Roberson slab->us_size = size; 21558355f576SJeff Roberson } else { 21568355f576SJeff Roberson uma_zfree_internal(slabzone, slab, NULL, 0); 21578355f576SJeff Roberson } 21588355f576SJeff Roberson 21598355f576SJeff Roberson 21608355f576SJeff Roberson return (mem); 21618355f576SJeff Roberson } 21628355f576SJeff Roberson 21638355f576SJeff Roberson void 21648355f576SJeff Roberson uma_large_free(uma_slab_t slab) 21658355f576SJeff Roberson { 216699571dc3SJeff Roberson vsetobj((vm_offset_t)slab->us_data, kmem_object); 2167125ee0d1STor Egge /* 21685103186cSAlan Cox * XXX: We get a lock order reversal if we don't have Giant: 2169125ee0d1STor Egge * vm_map_remove (locks system map) -> vm_map_delete -> 2170125ee0d1STor Egge * vm_map_entry_unwire -> vm_fault_unwire -> mtx_lock(&Giant) 2171125ee0d1STor Egge */ 2172125ee0d1STor Egge if (!mtx_owned(&Giant)) { 2173125ee0d1STor Egge mtx_lock(&Giant); 2174125ee0d1STor Egge page_free(slab->us_data, slab->us_size, slab->us_flags); 2175125ee0d1STor Egge mtx_unlock(&Giant); 2176125ee0d1STor Egge } else 21778355f576SJeff Roberson page_free(slab->us_data, slab->us_size, slab->us_flags); 21788355f576SJeff Roberson uma_zfree_internal(slabzone, slab, NULL, 0); 21798355f576SJeff Roberson } 21808355f576SJeff Roberson 21818355f576SJeff Roberson void 21828355f576SJeff Roberson uma_print_stats(void) 21838355f576SJeff Roberson { 21848355f576SJeff Roberson zone_foreach(uma_print_zone); 21858355f576SJeff Roberson } 21868355f576SJeff Roberson 21878355f576SJeff Roberson void 21888355f576SJeff Roberson uma_print_zone(uma_zone_t zone) 21898355f576SJeff Roberson { 21908355f576SJeff Roberson printf("%s(%p) size %d(%d) flags %d ipers %d ppera %d out %d free %d\n", 21918355f576SJeff Roberson zone->uz_name, zone, zone->uz_size, zone->uz_rsize, zone->uz_flags, 21928355f576SJeff Roberson zone->uz_ipers, zone->uz_ppera, 21938355f576SJeff Roberson (zone->uz_ipers * zone->uz_pages) - zone->uz_free, zone->uz_free); 21948355f576SJeff Roberson } 21958355f576SJeff Roberson 21968355f576SJeff Roberson /* 21978355f576SJeff Roberson * Sysctl handler for vm.zone 21988355f576SJeff Roberson * 21998355f576SJeff Roberson * stolen from vm_zone.c 22008355f576SJeff Roberson */ 22018355f576SJeff Roberson static int 22028355f576SJeff Roberson sysctl_vm_zone(SYSCTL_HANDLER_ARGS) 22038355f576SJeff Roberson { 22048355f576SJeff Roberson int error, len, cnt; 22058355f576SJeff Roberson const int linesize = 128; /* conservative */ 22068355f576SJeff Roberson int totalfree; 22078355f576SJeff Roberson char *tmpbuf, *offset; 22088355f576SJeff Roberson uma_zone_t z; 22098355f576SJeff Roberson char *p; 2210f828e5beSJeff Roberson int cpu; 2211f828e5beSJeff Roberson int cachefree; 2212f828e5beSJeff Roberson uma_bucket_t bucket; 2213f828e5beSJeff Roberson uma_cache_t cache; 22148355f576SJeff Roberson 22158355f576SJeff Roberson cnt = 0; 22160da47b2fSJeff Roberson mtx_lock(&uma_mtx); 22178355f576SJeff Roberson LIST_FOREACH(z, &uma_zones, uz_link) 22188355f576SJeff Roberson cnt++; 22190da47b2fSJeff Roberson mtx_unlock(&uma_mtx); 22208355f576SJeff Roberson MALLOC(tmpbuf, char *, (cnt == 0 ? 1 : cnt) * linesize, 2221a163d034SWarner Losh M_TEMP, M_WAITOK); 22228355f576SJeff Roberson len = snprintf(tmpbuf, linesize, 22238355f576SJeff Roberson "\nITEM SIZE LIMIT USED FREE REQUESTS\n\n"); 22248355f576SJeff Roberson if (cnt == 0) 22258355f576SJeff Roberson tmpbuf[len - 1] = '\0'; 22268355f576SJeff Roberson error = SYSCTL_OUT(req, tmpbuf, cnt == 0 ? len-1 : len); 22278355f576SJeff Roberson if (error || cnt == 0) 22288355f576SJeff Roberson goto out; 22298355f576SJeff Roberson offset = tmpbuf; 2230f4af24d5SJeff Roberson mtx_lock(&uma_mtx); 22318355f576SJeff Roberson LIST_FOREACH(z, &uma_zones, uz_link) { 22328355f576SJeff Roberson if (cnt == 0) /* list may have changed size */ 22338355f576SJeff Roberson break; 22341c35e213SBosko Milekic if (!(z->uz_flags & UMA_ZFLAG_INTERNAL)) { 2235f828e5beSJeff Roberson for (cpu = 0; cpu < maxcpu; cpu++) { 2236f828e5beSJeff Roberson if (CPU_ABSENT(cpu)) 2237f828e5beSJeff Roberson continue; 2238f828e5beSJeff Roberson CPU_LOCK(cpu); 2239f828e5beSJeff Roberson } 22401c35e213SBosko Milekic } 22418355f576SJeff Roberson ZONE_LOCK(z); 2242f828e5beSJeff Roberson cachefree = 0; 22431c35e213SBosko Milekic if (!(z->uz_flags & UMA_ZFLAG_INTERNAL)) { 2244f828e5beSJeff Roberson for (cpu = 0; cpu < maxcpu; cpu++) { 2245f828e5beSJeff Roberson if (CPU_ABSENT(cpu)) 2246f828e5beSJeff Roberson continue; 2247f828e5beSJeff Roberson cache = &z->uz_cpu[cpu]; 2248f828e5beSJeff Roberson if (cache->uc_allocbucket != NULL) 2249cae33c14SJeff Roberson cachefree += cache->uc_allocbucket->ub_cnt; 2250f828e5beSJeff Roberson if (cache->uc_freebucket != NULL) 2251cae33c14SJeff Roberson cachefree += cache->uc_freebucket->ub_cnt; 2252f828e5beSJeff Roberson CPU_UNLOCK(cpu); 2253f828e5beSJeff Roberson } 22541c35e213SBosko Milekic } 2255f828e5beSJeff Roberson LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link) { 2256cae33c14SJeff Roberson cachefree += bucket->ub_cnt; 2257f828e5beSJeff Roberson } 2258f828e5beSJeff Roberson totalfree = z->uz_free + cachefree; 22598355f576SJeff Roberson len = snprintf(offset, linesize, 22608355f576SJeff Roberson "%-12.12s %6.6u, %8.8u, %6.6u, %6.6u, %8.8llu\n", 22618355f576SJeff Roberson z->uz_name, z->uz_size, 22628355f576SJeff Roberson z->uz_maxpages * z->uz_ipers, 22638355f576SJeff Roberson (z->uz_ipers * (z->uz_pages / z->uz_ppera)) - totalfree, 22648355f576SJeff Roberson totalfree, 22658355f576SJeff Roberson (unsigned long long)z->uz_allocs); 22668355f576SJeff Roberson ZONE_UNLOCK(z); 22678355f576SJeff Roberson for (p = offset + 12; p > offset && *p == ' '; --p) 22688355f576SJeff Roberson /* nothing */ ; 22698355f576SJeff Roberson p[1] = ':'; 22708355f576SJeff Roberson cnt--; 22718355f576SJeff Roberson offset += len; 22728355f576SJeff Roberson } 2273f4af24d5SJeff Roberson mtx_unlock(&uma_mtx); 22748355f576SJeff Roberson *offset++ = '\0'; 22758355f576SJeff Roberson error = SYSCTL_OUT(req, tmpbuf, offset - tmpbuf); 22768355f576SJeff Roberson out: 22778355f576SJeff Roberson FREE(tmpbuf, M_TEMP); 22788355f576SJeff Roberson return (error); 22798355f576SJeff Roberson } 2280