FreeBSD/Linux Kernel Cross Reference
sys/vm/uma_core.c
1 /*-
2 * Copyright (c) 2002, 2003, 2004, 2005 Jeffrey Roberson <jeff@FreeBSD.org>
3 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
4 * Copyright (c) 2004-2006 Robert N. M. Watson
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice unmodified, this list of conditions, and the following
12 * disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 */
28
29 /*
30 * uma_core.c Implementation of the Universal Memory allocator
31 *
32 * This allocator is intended to replace the multitude of similar object caches
33 * in the standard FreeBSD kernel. The intent is to be flexible as well as
34 * effecient. A primary design goal is to return unused memory to the rest of
35 * the system. This will make the system as a whole more flexible due to the
36 * ability to move memory to subsystems which most need it instead of leaving
37 * pools of reserved memory unused.
38 *
39 * The basic ideas stem from similar slab/zone based allocators whose algorithms
40 * are well known.
41 *
42 */
43
44 /*
45 * TODO:
46 * - Improve memory usage for large allocations
47 * - Investigate cache size adjustments
48 */
49
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
52
53 /* I should really use ktr.. */
54 /*
55 #define UMA_DEBUG 1
56 #define UMA_DEBUG_ALLOC 1
57 #define UMA_DEBUG_ALLOC_1 1
58 */
59
60 #include "opt_ddb.h"
61 #include "opt_param.h"
62
63 #include <sys/param.h>
64 #include <sys/systm.h>
65 #include <sys/kernel.h>
66 #include <sys/types.h>
67 #include <sys/queue.h>
68 #include <sys/malloc.h>
69 #include <sys/ktr.h>
70 #include <sys/lock.h>
71 #include <sys/sysctl.h>
72 #include <sys/mutex.h>
73 #include <sys/proc.h>
74 #include <sys/sbuf.h>
75 #include <sys/smp.h>
76 #include <sys/vmmeter.h>
77
78 #include <vm/vm.h>
79 #include <vm/vm_object.h>
80 #include <vm/vm_page.h>
81 #include <vm/vm_param.h>
82 #include <vm/vm_map.h>
83 #include <vm/vm_kern.h>
84 #include <vm/vm_extern.h>
85 #include <vm/uma.h>
86 #include <vm/uma_int.h>
87 #include <vm/uma_dbg.h>
88
89 #include <machine/vmparam.h>
90
91 #include <ddb/ddb.h>
92
93 /*
94 * This is the zone and keg from which all zones are spawned. The idea is that
95 * even the zone & keg heads are allocated from the allocator, so we use the
96 * bss section to bootstrap us.
97 */
98 static struct uma_keg masterkeg;
99 static struct uma_zone masterzone_k;
100 static struct uma_zone masterzone_z;
101 static uma_zone_t kegs = &masterzone_k;
102 static uma_zone_t zones = &masterzone_z;
103
104 /* This is the zone from which all of uma_slab_t's are allocated. */
105 static uma_zone_t slabzone;
106 static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */
107
108 /*
109 * The initial hash tables come out of this zone so they can be allocated
110 * prior to malloc coming up.
111 */
112 static uma_zone_t hashzone;
113
114 /* The boot-time adjusted value for cache line alignment. */
115 static int uma_align_cache = 16 - 1;
116
117 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
118
119 /*
120 * Are we allowed to allocate buckets?
121 */
122 static int bucketdisable = 1;
123
124 /* Linked list of all kegs in the system */
125 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(&uma_kegs);
126
127 /* This mutex protects the keg list */
128 static struct mtx uma_mtx;
129
130 /* Linked list of boot time pages */
131 static LIST_HEAD(,uma_slab) uma_boot_pages =
132 LIST_HEAD_INITIALIZER(&uma_boot_pages);
133
134 /* This mutex protects the boot time pages list */
135 static struct mtx uma_boot_pages_mtx;
136
137 /* Is the VM done starting up? */
138 static int booted = 0;
139
140 /* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
141 static u_int uma_max_ipers;
142 static u_int uma_max_ipers_ref;
143
144 /*
145 * This is the handle used to schedule events that need to happen
146 * outside of the allocation fast path.
147 */
148 static struct callout uma_callout;
149 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
150
151 /*
152 * This structure is passed as the zone ctor arg so that I don't have to create
153 * a special allocation function just for zones.
154 */
155 struct uma_zctor_args {
156 char *name;
157 size_t size;
158 uma_ctor ctor;
159 uma_dtor dtor;
160 uma_init uminit;
161 uma_fini fini;
162 uma_keg_t keg;
163 int align;
164 u_int32_t flags;
165 };
166
167 struct uma_kctor_args {
168 uma_zone_t zone;
169 size_t size;
170 uma_init uminit;
171 uma_fini fini;
172 int align;
173 u_int32_t flags;
174 };
175
176 struct uma_bucket_zone {
177 uma_zone_t ubz_zone;
178 char *ubz_name;
179 int ubz_entries;
180 };
181
182 #define BUCKET_MAX 128
183
184 struct uma_bucket_zone bucket_zones[] = {
185 { NULL, "16 Bucket", 16 },
186 { NULL, "32 Bucket", 32 },
187 { NULL, "64 Bucket", 64 },
188 { NULL, "128 Bucket", 128 },
189 { NULL, NULL, 0}
190 };
191
192 #define BUCKET_SHIFT 4
193 #define BUCKET_ZONES ((BUCKET_MAX >> BUCKET_SHIFT) + 1)
194
195 /*
196 * bucket_size[] maps requested bucket sizes to zones that allocate a bucket
197 * of approximately the right size.
198 */
199 static uint8_t bucket_size[BUCKET_ZONES];
200
201 /*
202 * Flags and enumerations to be passed to internal functions.
203 */
204 enum zfreeskip { SKIP_NONE, SKIP_DTOR, SKIP_FINI };
205
206 #define ZFREE_STATFAIL 0x00000001 /* Update zone failure statistic. */
207 #define ZFREE_STATFREE 0x00000002 /* Update zone free statistic. */
208
209 /* Prototypes.. */
210
211 static void *obj_alloc(uma_zone_t, int, u_int8_t *, int);
212 static void *page_alloc(uma_zone_t, int, u_int8_t *, int);
213 static void *startup_alloc(uma_zone_t, int, u_int8_t *, int);
214 static void page_free(void *, int, u_int8_t);
215 static uma_slab_t slab_zalloc(uma_zone_t, int);
216 static void cache_drain(uma_zone_t);
217 static void bucket_drain(uma_zone_t, uma_bucket_t);
218 static void bucket_cache_drain(uma_zone_t zone);
219 static int keg_ctor(void *, int, void *, int);
220 static void keg_dtor(void *, int, void *);
221 static int zone_ctor(void *, int, void *, int);
222 static void zone_dtor(void *, int, void *);
223 static int zero_init(void *, int, int);
224 static void zone_small_init(uma_zone_t zone);
225 static void zone_large_init(uma_zone_t zone);
226 static void zone_foreach(void (*zfunc)(uma_zone_t));
227 static void zone_timeout(uma_zone_t zone);
228 static int hash_alloc(struct uma_hash *);
229 static int hash_expand(struct uma_hash *, struct uma_hash *);
230 static void hash_free(struct uma_hash *hash);
231 static void uma_timeout(void *);
232 static void uma_startup3(void);
233 static void *uma_zalloc_internal(uma_zone_t, void *, int);
234 static void uma_zfree_internal(uma_zone_t, void *, void *, enum zfreeskip,
235 int);
236 static void bucket_enable(void);
237 static void bucket_init(void);
238 static uma_bucket_t bucket_alloc(int, int);
239 static void bucket_free(uma_bucket_t);
240 static void bucket_zone_drain(void);
241 static int uma_zalloc_bucket(uma_zone_t zone, int flags);
242 static uma_slab_t uma_zone_slab(uma_zone_t zone, int flags);
243 static void *uma_slab_alloc(uma_zone_t zone, uma_slab_t slab);
244 static uma_zone_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
245 uma_fini fini, int align, u_int32_t flags);
246
247 void uma_print_zone(uma_zone_t);
248 void uma_print_stats(void);
249 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
250 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
251
252 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
253
254 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
255 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
256
257 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
258 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
259
260 /*
261 * This routine checks to see whether or not it's safe to enable buckets.
262 */
263
264 static void
265 bucket_enable(void)
266 {
267 if (cnt.v_free_count < cnt.v_free_min)
268 bucketdisable = 1;
269 else
270 bucketdisable = 0;
271 }
272
273 /*
274 * Initialize bucket_zones, the array of zones of buckets of various sizes.
275 *
276 * For each zone, calculate the memory required for each bucket, consisting
277 * of the header and an array of pointers. Initialize bucket_size[] to point
278 * the range of appropriate bucket sizes at the zone.
279 */
280 static void
281 bucket_init(void)
282 {
283 struct uma_bucket_zone *ubz;
284 int i;
285 int j;
286
287 for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) {
288 int size;
289
290 ubz = &bucket_zones[j];
291 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
292 size += sizeof(void *) * ubz->ubz_entries;
293 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
294 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
295 for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT))
296 bucket_size[i >> BUCKET_SHIFT] = j;
297 }
298 }
299
300 /*
301 * Given a desired number of entries for a bucket, return the zone from which
302 * to allocate the bucket.
303 */
304 static struct uma_bucket_zone *
305 bucket_zone_lookup(int entries)
306 {
307 int idx;
308
309 idx = howmany(entries, 1 << BUCKET_SHIFT);
310 return (&bucket_zones[bucket_size[idx]]);
311 }
312
313 static uma_bucket_t
314 bucket_alloc(int entries, int bflags)
315 {
316 struct uma_bucket_zone *ubz;
317 uma_bucket_t bucket;
318
319 /*
320 * This is to stop us from allocating per cpu buckets while we're
321 * running out of vm.boot_pages. Otherwise, we would exhaust the
322 * boot pages. This also prevents us from allocating buckets in
323 * low memory situations.
324 */
325 if (bucketdisable)
326 return (NULL);
327
328 ubz = bucket_zone_lookup(entries);
329 bucket = uma_zalloc_internal(ubz->ubz_zone, NULL, bflags);
330 if (bucket) {
331 #ifdef INVARIANTS
332 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
333 #endif
334 bucket->ub_cnt = 0;
335 bucket->ub_entries = ubz->ubz_entries;
336 }
337
338 return (bucket);
339 }
340
341 static void
342 bucket_free(uma_bucket_t bucket)
343 {
344 struct uma_bucket_zone *ubz;
345
346 ubz = bucket_zone_lookup(bucket->ub_entries);
347 uma_zfree_internal(ubz->ubz_zone, bucket, NULL, SKIP_NONE,
348 ZFREE_STATFREE);
349 }
350
351 static void
352 bucket_zone_drain(void)
353 {
354 struct uma_bucket_zone *ubz;
355
356 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
357 zone_drain(ubz->ubz_zone);
358 }
359
360
361 /*
362 * Routine called by timeout which is used to fire off some time interval
363 * based calculations. (stats, hash size, etc.)
364 *
365 * Arguments:
366 * arg Unused
367 *
368 * Returns:
369 * Nothing
370 */
371 static void
372 uma_timeout(void *unused)
373 {
374 bucket_enable();
375 zone_foreach(zone_timeout);
376
377 /* Reschedule this event */
378 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
379 }
380
381 /*
382 * Routine to perform timeout driven calculations. This expands the
383 * hashes and does per cpu statistics aggregation.
384 *
385 * Arguments:
386 * zone The zone to operate on
387 *
388 * Returns:
389 * Nothing
390 */
391 static void
392 zone_timeout(uma_zone_t zone)
393 {
394 uma_keg_t keg;
395 u_int64_t alloc;
396
397 keg = zone->uz_keg;
398 alloc = 0;
399
400 /*
401 * Expand the zone hash table.
402 *
403 * This is done if the number of slabs is larger than the hash size.
404 * What I'm trying to do here is completely reduce collisions. This
405 * may be a little aggressive. Should I allow for two collisions max?
406 */
407 ZONE_LOCK(zone);
408 if (keg->uk_flags & UMA_ZONE_HASH &&
409 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
410 struct uma_hash newhash;
411 struct uma_hash oldhash;
412 int ret;
413
414 /*
415 * This is so involved because allocating and freeing
416 * while the zone lock is held will lead to deadlock.
417 * I have to do everything in stages and check for
418 * races.
419 */
420 newhash = keg->uk_hash;
421 ZONE_UNLOCK(zone);
422 ret = hash_alloc(&newhash);
423 ZONE_LOCK(zone);
424 if (ret) {
425 if (hash_expand(&keg->uk_hash, &newhash)) {
426 oldhash = keg->uk_hash;
427 keg->uk_hash = newhash;
428 } else
429 oldhash = newhash;
430
431 ZONE_UNLOCK(zone);
432 hash_free(&oldhash);
433 ZONE_LOCK(zone);
434 }
435 }
436 ZONE_UNLOCK(zone);
437 }
438
439 /*
440 * Allocate and zero fill the next sized hash table from the appropriate
441 * backing store.
442 *
443 * Arguments:
444 * hash A new hash structure with the old hash size in uh_hashsize
445 *
446 * Returns:
447 * 1 on sucess and 0 on failure.
448 */
449 static int
450 hash_alloc(struct uma_hash *hash)
451 {
452 int oldsize;
453 int alloc;
454
455 oldsize = hash->uh_hashsize;
456
457 /* We're just going to go to a power of two greater */
458 if (oldsize) {
459 hash->uh_hashsize = oldsize * 2;
460 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
461 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
462 M_UMAHASH, M_NOWAIT);
463 } else {
464 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
465 hash->uh_slab_hash = uma_zalloc_internal(hashzone, NULL,
466 M_WAITOK);
467 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
468 }
469 if (hash->uh_slab_hash) {
470 bzero(hash->uh_slab_hash, alloc);
471 hash->uh_hashmask = hash->uh_hashsize - 1;
472 return (1);
473 }
474
475 return (0);
476 }
477
478 /*
479 * Expands the hash table for HASH zones. This is done from zone_timeout
480 * to reduce collisions. This must not be done in the regular allocation
481 * path, otherwise, we can recurse on the vm while allocating pages.
482 *
483 * Arguments:
484 * oldhash The hash you want to expand
485 * newhash The hash structure for the new table
486 *
487 * Returns:
488 * Nothing
489 *
490 * Discussion:
491 */
492 static int
493 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
494 {
495 uma_slab_t slab;
496 int hval;
497 int i;
498
499 if (!newhash->uh_slab_hash)
500 return (0);
501
502 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
503 return (0);
504
505 /*
506 * I need to investigate hash algorithms for resizing without a
507 * full rehash.
508 */
509
510 for (i = 0; i < oldhash->uh_hashsize; i++)
511 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
512 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
513 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
514 hval = UMA_HASH(newhash, slab->us_data);
515 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
516 slab, us_hlink);
517 }
518
519 return (1);
520 }
521
522 /*
523 * Free the hash bucket to the appropriate backing store.
524 *
525 * Arguments:
526 * slab_hash The hash bucket we're freeing
527 * hashsize The number of entries in that hash bucket
528 *
529 * Returns:
530 * Nothing
531 */
532 static void
533 hash_free(struct uma_hash *hash)
534 {
535 if (hash->uh_slab_hash == NULL)
536 return;
537 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
538 uma_zfree_internal(hashzone,
539 hash->uh_slab_hash, NULL, SKIP_NONE, ZFREE_STATFREE);
540 else
541 free(hash->uh_slab_hash, M_UMAHASH);
542 }
543
544 /*
545 * Frees all outstanding items in a bucket
546 *
547 * Arguments:
548 * zone The zone to free to, must be unlocked.
549 * bucket The free/alloc bucket with items, cpu queue must be locked.
550 *
551 * Returns:
552 * Nothing
553 */
554
555 static void
556 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
557 {
558 uma_slab_t slab;
559 int mzone;
560 void *item;
561
562 if (bucket == NULL)
563 return;
564
565 slab = NULL;
566 mzone = 0;
567
568 /* We have to lookup the slab again for malloc.. */
569 if (zone->uz_keg->uk_flags & UMA_ZONE_MALLOC)
570 mzone = 1;
571
572 while (bucket->ub_cnt > 0) {
573 bucket->ub_cnt--;
574 item = bucket->ub_bucket[bucket->ub_cnt];
575 #ifdef INVARIANTS
576 bucket->ub_bucket[bucket->ub_cnt] = NULL;
577 KASSERT(item != NULL,
578 ("bucket_drain: botched ptr, item is NULL"));
579 #endif
580 /*
581 * This is extremely inefficient. The slab pointer was passed
582 * to uma_zfree_arg, but we lost it because the buckets don't
583 * hold them. This will go away when free() gets a size passed
584 * to it.
585 */
586 if (mzone)
587 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
588 uma_zfree_internal(zone, item, slab, SKIP_DTOR, 0);
589 }
590 }
591
592 /*
593 * Drains the per cpu caches for a zone.
594 *
595 * NOTE: This may only be called while the zone is being turn down, and not
596 * during normal operation. This is necessary in order that we do not have
597 * to migrate CPUs to drain the per-CPU caches.
598 *
599 * Arguments:
600 * zone The zone to drain, must be unlocked.
601 *
602 * Returns:
603 * Nothing
604 */
605 static void
606 cache_drain(uma_zone_t zone)
607 {
608 uma_cache_t cache;
609 int cpu;
610
611 /*
612 * XXX: It is safe to not lock the per-CPU caches, because we're
613 * tearing down the zone anyway. I.e., there will be no further use
614 * of the caches at this point.
615 *
616 * XXX: It would good to be able to assert that the zone is being
617 * torn down to prevent improper use of cache_drain().
618 *
619 * XXX: We lock the zone before passing into bucket_cache_drain() as
620 * it is used elsewhere. Should the tear-down path be made special
621 * there in some form?
622 */
623 for (cpu = 0; cpu <= mp_maxid; cpu++) {
624 if (CPU_ABSENT(cpu))
625 continue;
626 cache = &zone->uz_cpu[cpu];
627 bucket_drain(zone, cache->uc_allocbucket);
628 bucket_drain(zone, cache->uc_freebucket);
629 if (cache->uc_allocbucket != NULL)
630 bucket_free(cache->uc_allocbucket);
631 if (cache->uc_freebucket != NULL)
632 bucket_free(cache->uc_freebucket);
633 cache->uc_allocbucket = cache->uc_freebucket = NULL;
634 }
635 ZONE_LOCK(zone);
636 bucket_cache_drain(zone);
637 ZONE_UNLOCK(zone);
638 }
639
640 /*
641 * Drain the cached buckets from a zone. Expects a locked zone on entry.
642 */
643 static void
644 bucket_cache_drain(uma_zone_t zone)
645 {
646 uma_bucket_t bucket;
647
648 /*
649 * Drain the bucket queues and free the buckets, we just keep two per
650 * cpu (alloc/free).
651 */
652 while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
653 LIST_REMOVE(bucket, ub_link);
654 ZONE_UNLOCK(zone);
655 bucket_drain(zone, bucket);
656 bucket_free(bucket);
657 ZONE_LOCK(zone);
658 }
659
660 /* Now we do the free queue.. */
661 while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
662 LIST_REMOVE(bucket, ub_link);
663 bucket_free(bucket);
664 }
665 }
666
667 /*
668 * Frees pages from a zone back to the system. This is done on demand from
669 * the pageout daemon.
670 *
671 * Arguments:
672 * zone The zone to free pages from
673 * all Should we drain all items?
674 *
675 * Returns:
676 * Nothing.
677 */
678 void
679 zone_drain(uma_zone_t zone)
680 {
681 struct slabhead freeslabs = { 0 };
682 uma_keg_t keg;
683 uma_slab_t slab;
684 uma_slab_t n;
685 u_int8_t flags;
686 u_int8_t *mem;
687 int i;
688
689 keg = zone->uz_keg;
690
691 /*
692 * We don't want to take pages from statically allocated zones at this
693 * time
694 */
695 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
696 return;
697
698 ZONE_LOCK(zone);
699
700 #ifdef UMA_DEBUG
701 printf("%s free items: %u\n", zone->uz_name, keg->uk_free);
702 #endif
703 bucket_cache_drain(zone);
704 if (keg->uk_free == 0)
705 goto finished;
706
707 slab = LIST_FIRST(&keg->uk_free_slab);
708 while (slab) {
709 n = LIST_NEXT(slab, us_link);
710
711 /* We have no where to free these to */
712 if (slab->us_flags & UMA_SLAB_BOOT) {
713 slab = n;
714 continue;
715 }
716
717 LIST_REMOVE(slab, us_link);
718 keg->uk_pages -= keg->uk_ppera;
719 keg->uk_free -= keg->uk_ipers;
720
721 if (keg->uk_flags & UMA_ZONE_HASH)
722 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
723
724 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
725
726 slab = n;
727 }
728 finished:
729 ZONE_UNLOCK(zone);
730
731 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
732 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
733 if (keg->uk_fini)
734 for (i = 0; i < keg->uk_ipers; i++)
735 keg->uk_fini(
736 slab->us_data + (keg->uk_rsize * i),
737 keg->uk_size);
738 flags = slab->us_flags;
739 mem = slab->us_data;
740
741 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
742 (keg->uk_flags & UMA_ZONE_REFCNT)) {
743 vm_object_t obj;
744
745 if (flags & UMA_SLAB_KMEM)
746 obj = kmem_object;
747 else if (flags & UMA_SLAB_KERNEL)
748 obj = kernel_object;
749 else
750 obj = NULL;
751 for (i = 0; i < keg->uk_ppera; i++)
752 vsetobj((vm_offset_t)mem + (i * PAGE_SIZE),
753 obj);
754 }
755 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
756 uma_zfree_internal(keg->uk_slabzone, slab, NULL,
757 SKIP_NONE, ZFREE_STATFREE);
758 #ifdef UMA_DEBUG
759 printf("%s: Returning %d bytes.\n",
760 zone->uz_name, UMA_SLAB_SIZE * keg->uk_ppera);
761 #endif
762 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, flags);
763 }
764 }
765
766 /*
767 * Allocate a new slab for a zone. This does not insert the slab onto a list.
768 *
769 * Arguments:
770 * zone The zone to allocate slabs for
771 * wait Shall we wait?
772 *
773 * Returns:
774 * The slab that was allocated or NULL if there is no memory and the
775 * caller specified M_NOWAIT.
776 */
777 static uma_slab_t
778 slab_zalloc(uma_zone_t zone, int wait)
779 {
780 uma_slabrefcnt_t slabref;
781 uma_slab_t slab;
782 uma_keg_t keg;
783 u_int8_t *mem;
784 u_int8_t flags;
785 int i;
786
787 slab = NULL;
788 keg = zone->uz_keg;
789
790 #ifdef UMA_DEBUG
791 printf("slab_zalloc: Allocating a new slab for %s\n", zone->uz_name);
792 #endif
793 ZONE_UNLOCK(zone);
794
795 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
796 slab = uma_zalloc_internal(keg->uk_slabzone, NULL, wait);
797 if (slab == NULL) {
798 ZONE_LOCK(zone);
799 return NULL;
800 }
801 }
802
803 /*
804 * This reproduces the old vm_zone behavior of zero filling pages the
805 * first time they are added to a zone.
806 *
807 * Malloced items are zeroed in uma_zalloc.
808 */
809
810 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
811 wait |= M_ZERO;
812 else
813 wait &= ~M_ZERO;
814
815 mem = keg->uk_allocf(zone, keg->uk_ppera * UMA_SLAB_SIZE,
816 &flags, wait);
817 if (mem == NULL) {
818 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
819 uma_zfree_internal(keg->uk_slabzone, slab, NULL,
820 SKIP_NONE, ZFREE_STATFREE);
821 ZONE_LOCK(zone);
822 return (NULL);
823 }
824
825 /* Point the slab into the allocated memory */
826 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
827 slab = (uma_slab_t )(mem + keg->uk_pgoff);
828
829 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
830 (keg->uk_flags & UMA_ZONE_REFCNT))
831 for (i = 0; i < keg->uk_ppera; i++)
832 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
833
834 slab->us_keg = keg;
835 slab->us_data = mem;
836 slab->us_freecount = keg->uk_ipers;
837 slab->us_firstfree = 0;
838 slab->us_flags = flags;
839
840 if (keg->uk_flags & UMA_ZONE_REFCNT) {
841 slabref = (uma_slabrefcnt_t)slab;
842 for (i = 0; i < keg->uk_ipers; i++) {
843 slabref->us_freelist[i].us_refcnt = 0;
844 slabref->us_freelist[i].us_item = i+1;
845 }
846 } else {
847 for (i = 0; i < keg->uk_ipers; i++)
848 slab->us_freelist[i].us_item = i+1;
849 }
850
851 if (keg->uk_init != NULL) {
852 for (i = 0; i < keg->uk_ipers; i++)
853 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
854 keg->uk_size, wait) != 0)
855 break;
856 if (i != keg->uk_ipers) {
857 if (keg->uk_fini != NULL) {
858 for (i--; i > -1; i--)
859 keg->uk_fini(slab->us_data +
860 (keg->uk_rsize * i),
861 keg->uk_size);
862 }
863 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
864 (keg->uk_flags & UMA_ZONE_REFCNT)) {
865 vm_object_t obj;
866
867 if (flags & UMA_SLAB_KMEM)
868 obj = kmem_object;
869 else if (flags & UMA_SLAB_KERNEL)
870 obj = kernel_object;
871 else
872 obj = NULL;
873 for (i = 0; i < keg->uk_ppera; i++)
874 vsetobj((vm_offset_t)mem +
875 (i * PAGE_SIZE), obj);
876 }
877 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
878 uma_zfree_internal(keg->uk_slabzone, slab,
879 NULL, SKIP_NONE, ZFREE_STATFREE);
880 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera,
881 flags);
882 ZONE_LOCK(zone);
883 return (NULL);
884 }
885 }
886 ZONE_LOCK(zone);
887
888 if (keg->uk_flags & UMA_ZONE_HASH)
889 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
890
891 keg->uk_pages += keg->uk_ppera;
892 keg->uk_free += keg->uk_ipers;
893
894 return (slab);
895 }
896
897 /*
898 * This function is intended to be used early on in place of page_alloc() so
899 * that we may use the boot time page cache to satisfy allocations before
900 * the VM is ready.
901 */
902 static void *
903 startup_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
904 {
905 uma_keg_t keg;
906 uma_slab_t tmps;
907
908 keg = zone->uz_keg;
909
910 /*
911 * Check our small startup cache to see if it has pages remaining.
912 */
913 mtx_lock(&uma_boot_pages_mtx);
914 if ((tmps = LIST_FIRST(&uma_boot_pages)) != NULL) {
915 LIST_REMOVE(tmps, us_link);
916 mtx_unlock(&uma_boot_pages_mtx);
917 *pflag = tmps->us_flags;
918 return (tmps->us_data);
919 }
920 mtx_unlock(&uma_boot_pages_mtx);
921 if (booted == 0)
922 panic("UMA: Increase vm.boot_pages");
923 /*
924 * Now that we've booted reset these users to their real allocator.
925 */
926 #ifdef UMA_MD_SMALL_ALLOC
927 keg->uk_allocf = uma_small_alloc;
928 #else
929 keg->uk_allocf = page_alloc;
930 #endif
931 return keg->uk_allocf(zone, bytes, pflag, wait);
932 }
933
934 /*
935 * Allocates a number of pages from the system
936 *
937 * Arguments:
938 * zone Unused
939 * bytes The number of bytes requested
940 * wait Shall we wait?
941 *
942 * Returns:
943 * A pointer to the alloced memory or possibly
944 * NULL if M_NOWAIT is set.
945 */
946 static void *
947 page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
948 {
949 void *p; /* Returned page */
950
951 *pflag = UMA_SLAB_KMEM;
952 p = (void *) kmem_malloc(kmem_map, bytes, wait);
953
954 return (p);
955 }
956
957 /*
958 * Allocates a number of pages from within an object
959 *
960 * Arguments:
961 * zone Unused
962 * bytes The number of bytes requested
963 * wait Shall we wait?
964 *
965 * Returns:
966 * A pointer to the alloced memory or possibly
967 * NULL if M_NOWAIT is set.
968 */
969 static void *
970 obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
971 {
972 vm_object_t object;
973 vm_offset_t retkva, zkva;
974 vm_page_t p;
975 int pages, startpages;
976
977 object = zone->uz_keg->uk_obj;
978 retkva = 0;
979
980 /*
981 * This looks a little weird since we're getting one page at a time.
982 */
983 VM_OBJECT_LOCK(object);
984 p = TAILQ_LAST(&object->memq, pglist);
985 pages = p != NULL ? p->pindex + 1 : 0;
986 startpages = pages;
987 zkva = zone->uz_keg->uk_kva + pages * PAGE_SIZE;
988 for (; bytes > 0; bytes -= PAGE_SIZE) {
989 p = vm_page_alloc(object, pages,
990 VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED);
991 if (p == NULL) {
992 if (pages != startpages)
993 pmap_qremove(retkva, pages - startpages);
994 while (pages != startpages) {
995 pages--;
996 p = TAILQ_LAST(&object->memq, pglist);
997 vm_page_lock_queues();
998 vm_page_unwire(p, 0);
999 vm_page_free(p);
1000 vm_page_unlock_queues();
1001 }
1002 retkva = 0;
1003 goto done;
1004 }
1005 pmap_qenter(zkva, &p, 1);
1006 if (retkva == 0)
1007 retkva = zkva;
1008 zkva += PAGE_SIZE;
1009 pages += 1;
1010 }
1011 done:
1012 VM_OBJECT_UNLOCK(object);
1013 *flags = UMA_SLAB_PRIV;
1014
1015 return ((void *)retkva);
1016 }
1017
1018 /*
1019 * Frees a number of pages to the system
1020 *
1021 * Arguments:
1022 * mem A pointer to the memory to be freed
1023 * size The size of the memory being freed
1024 * flags The original p->us_flags field
1025 *
1026 * Returns:
1027 * Nothing
1028 */
1029 static void
1030 page_free(void *mem, int size, u_int8_t flags)
1031 {
1032 vm_map_t map;
1033
1034 if (flags & UMA_SLAB_KMEM)
1035 map = kmem_map;
1036 else
1037 panic("UMA: page_free used with invalid flags %d\n", flags);
1038
1039 kmem_free(map, (vm_offset_t)mem, size);
1040 }
1041
1042 /*
1043 * Zero fill initializer
1044 *
1045 * Arguments/Returns follow uma_init specifications
1046 */
1047 static int
1048 zero_init(void *mem, int size, int flags)
1049 {
1050 bzero(mem, size);
1051 return (0);
1052 }
1053
1054 /*
1055 * Finish creating a small uma zone. This calculates ipers, and the zone size.
1056 *
1057 * Arguments
1058 * zone The zone we should initialize
1059 *
1060 * Returns
1061 * Nothing
1062 */
1063 static void
1064 zone_small_init(uma_zone_t zone)
1065 {
1066 uma_keg_t keg;
1067 u_int rsize;
1068 u_int memused;
1069 u_int wastedspace;
1070 u_int shsize;
1071
1072 keg = zone->uz_keg;
1073 KASSERT(keg != NULL, ("Keg is null in zone_small_init"));
1074 rsize = keg->uk_size;
1075
1076 if (rsize < UMA_SMALLEST_UNIT)
1077 rsize = UMA_SMALLEST_UNIT;
1078 if (rsize & keg->uk_align)
1079 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1080
1081 keg->uk_rsize = rsize;
1082 keg->uk_ppera = 1;
1083
1084 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1085 rsize += UMA_FRITMREF_SZ; /* linkage & refcnt */
1086 shsize = sizeof(struct uma_slab_refcnt);
1087 } else {
1088 rsize += UMA_FRITM_SZ; /* Account for linkage */
1089 shsize = sizeof(struct uma_slab);
1090 }
1091
1092 keg->uk_ipers = (UMA_SLAB_SIZE - shsize) / rsize;
1093 KASSERT(keg->uk_ipers != 0, ("zone_small_init: ipers is 0"));
1094 memused = keg->uk_ipers * rsize + shsize;
1095 wastedspace = UMA_SLAB_SIZE - memused;
1096
1097 /*
1098 * We can't do OFFPAGE if we're internal or if we've been
1099 * asked to not go to the VM for buckets. If we do this we
1100 * may end up going to the VM (kmem_map) for slabs which we
1101 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a
1102 * result of UMA_ZONE_VM, which clearly forbids it.
1103 */
1104 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1105 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1106 return;
1107
1108 if ((wastedspace >= UMA_MAX_WASTE) &&
1109 (keg->uk_ipers < (UMA_SLAB_SIZE / keg->uk_rsize))) {
1110 keg->uk_ipers = UMA_SLAB_SIZE / keg->uk_rsize;
1111 KASSERT(keg->uk_ipers <= 255,
1112 ("zone_small_init: keg->uk_ipers too high!"));
1113 #ifdef UMA_DEBUG
1114 printf("UMA decided we need offpage slab headers for "
1115 "zone: %s, calculated wastedspace = %d, "
1116 "maximum wasted space allowed = %d, "
1117 "calculated ipers = %d, "
1118 "new wasted space = %d\n", zone->uz_name, wastedspace,
1119 UMA_MAX_WASTE, keg->uk_ipers,
1120 UMA_SLAB_SIZE - keg->uk_ipers * keg->uk_rsize);
1121 #endif
1122 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1123 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1124 keg->uk_flags |= UMA_ZONE_HASH;
1125 }
1126 }
1127
1128 /*
1129 * Finish creating a large (> UMA_SLAB_SIZE) uma zone. Just give in and do
1130 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1131 * more complicated.
1132 *
1133 * Arguments
1134 * zone The zone we should initialize
1135 *
1136 * Returns
1137 * Nothing
1138 */
1139 static void
1140 zone_large_init(uma_zone_t zone)
1141 {
1142 uma_keg_t keg;
1143 int pages;
1144
1145 keg = zone->uz_keg;
1146
1147 KASSERT(keg != NULL, ("Keg is null in zone_large_init"));
1148 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1149 ("zone_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY zone"));
1150
1151 pages = keg->uk_size / UMA_SLAB_SIZE;
1152
1153 /* Account for remainder */
1154 if ((pages * UMA_SLAB_SIZE) < keg->uk_size)
1155 pages++;
1156
1157 keg->uk_ppera = pages;
1158 keg->uk_ipers = 1;
1159
1160 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1161 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1162 keg->uk_flags |= UMA_ZONE_HASH;
1163
1164 keg->uk_rsize = keg->uk_size;
1165 }
1166
1167 /*
1168 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1169 * the keg onto the global keg list.
1170 *
1171 * Arguments/Returns follow uma_ctor specifications
1172 * udata Actually uma_kctor_args
1173 */
1174 static int
1175 keg_ctor(void *mem, int size, void *udata, int flags)
1176 {
1177 struct uma_kctor_args *arg = udata;
1178 uma_keg_t keg = mem;
1179 uma_zone_t zone;
1180
1181 bzero(keg, size);
1182 keg->uk_size = arg->size;
1183 keg->uk_init = arg->uminit;
1184 keg->uk_fini = arg->fini;
1185 keg->uk_align = arg->align;
1186 keg->uk_free = 0;
1187 keg->uk_pages = 0;
1188 keg->uk_flags = arg->flags;
1189 keg->uk_allocf = page_alloc;
1190 keg->uk_freef = page_free;
1191 keg->uk_recurse = 0;
1192 keg->uk_slabzone = NULL;
1193
1194 /*
1195 * The master zone is passed to us at keg-creation time.
1196 */
1197 zone = arg->zone;
1198 zone->uz_keg = keg;
1199
1200 if (arg->flags & UMA_ZONE_VM)
1201 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1202
1203 if (arg->flags & UMA_ZONE_ZINIT)
1204 keg->uk_init = zero_init;
1205
1206 /*
1207 * The +UMA_FRITM_SZ added to uk_size is to account for the
1208 * linkage that is added to the size in zone_small_init(). If
1209 * we don't account for this here then we may end up in
1210 * zone_small_init() with a calculated 'ipers' of 0.
1211 */
1212 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1213 if ((keg->uk_size+UMA_FRITMREF_SZ) >
1214 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)))
1215 zone_large_init(zone);
1216 else
1217 zone_small_init(zone);
1218 } else {
1219 if ((keg->uk_size+UMA_FRITM_SZ) >
1220 (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1221 zone_large_init(zone);
1222 else
1223 zone_small_init(zone);
1224 }
1225
1226 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1227 if (keg->uk_flags & UMA_ZONE_REFCNT)
1228 keg->uk_slabzone = slabrefzone;
1229 else
1230 keg->uk_slabzone = slabzone;
1231 }
1232
1233 /*
1234 * If we haven't booted yet we need allocations to go through the
1235 * startup cache until the vm is ready.
1236 */
1237 if (keg->uk_ppera == 1) {
1238 #ifdef UMA_MD_SMALL_ALLOC
1239 keg->uk_allocf = uma_small_alloc;
1240 keg->uk_freef = uma_small_free;
1241 #endif
1242 if (booted == 0)
1243 keg->uk_allocf = startup_alloc;
1244 }
1245
1246 /*
1247 * Initialize keg's lock (shared among zones) through
1248 * Master zone
1249 */
1250 zone->uz_lock = &keg->uk_lock;
1251 if (arg->flags & UMA_ZONE_MTXCLASS)
1252 ZONE_LOCK_INIT(zone, 1);
1253 else
1254 ZONE_LOCK_INIT(zone, 0);
1255
1256 /*
1257 * If we're putting the slab header in the actual page we need to
1258 * figure out where in each page it goes. This calculates a right
1259 * justified offset into the memory on an ALIGN_PTR boundary.
1260 */
1261 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1262 u_int totsize;
1263
1264 /* Size of the slab struct and free list */
1265 if (keg->uk_flags & UMA_ZONE_REFCNT)
1266 totsize = sizeof(struct uma_slab_refcnt) +
1267 keg->uk_ipers * UMA_FRITMREF_SZ;
1268 else
1269 totsize = sizeof(struct uma_slab) +
1270 keg->uk_ipers * UMA_FRITM_SZ;
1271
1272 if (totsize & UMA_ALIGN_PTR)
1273 totsize = (totsize & ~UMA_ALIGN_PTR) +
1274 (UMA_ALIGN_PTR + 1);
1275 keg->uk_pgoff = UMA_SLAB_SIZE - totsize;
1276
1277 if (keg->uk_flags & UMA_ZONE_REFCNT)
1278 totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt)
1279 + keg->uk_ipers * UMA_FRITMREF_SZ;
1280 else
1281 totsize = keg->uk_pgoff + sizeof(struct uma_slab)
1282 + keg->uk_ipers * UMA_FRITM_SZ;
1283
1284 /*
1285 * The only way the following is possible is if with our
1286 * UMA_ALIGN_PTR adjustments we are now bigger than
1287 * UMA_SLAB_SIZE. I haven't checked whether this is
1288 * mathematically possible for all cases, so we make
1289 * sure here anyway.
1290 */
1291 if (totsize > UMA_SLAB_SIZE) {
1292 printf("zone %s ipers %d rsize %d size %d\n",
1293 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1294 keg->uk_size);
1295 panic("UMA slab won't fit.\n");
1296 }
1297 }
1298
1299 if (keg->uk_flags & UMA_ZONE_HASH)
1300 hash_alloc(&keg->uk_hash);
1301
1302 #ifdef UMA_DEBUG
1303 printf("%s(%p) size = %d ipers = %d ppera = %d pgoff = %d\n",
1304 zone->uz_name, zone,
1305 keg->uk_size, keg->uk_ipers,
1306 keg->uk_ppera, keg->uk_pgoff);
1307 #endif
1308
1309 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1310
1311 mtx_lock(&uma_mtx);
1312 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1313 mtx_unlock(&uma_mtx);
1314 return (0);
1315 }
1316
1317 /*
1318 * Zone header ctor. This initializes all fields, locks, etc.
1319 *
1320 * Arguments/Returns follow uma_ctor specifications
1321 * udata Actually uma_zctor_args
1322 */
1323
1324 static int
1325 zone_ctor(void *mem, int size, void *udata, int flags)
1326 {
1327 struct uma_zctor_args *arg = udata;
1328 uma_zone_t zone = mem;
1329 uma_zone_t z;
1330 uma_keg_t keg;
1331
1332 bzero(zone, size);
1333 zone->uz_name = arg->name;
1334 zone->uz_ctor = arg->ctor;
1335 zone->uz_dtor = arg->dtor;
1336 zone->uz_init = NULL;
1337 zone->uz_fini = NULL;
1338 zone->uz_allocs = 0;
1339 zone->uz_frees = 0;
1340 zone->uz_fails = 0;
1341 zone->uz_fills = zone->uz_count = 0;
1342
1343 if (arg->flags & UMA_ZONE_SECONDARY) {
1344 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1345 keg = arg->keg;
1346 zone->uz_keg = keg;
1347 zone->uz_init = arg->uminit;
1348 zone->uz_fini = arg->fini;
1349 zone->uz_lock = &keg->uk_lock;
1350 mtx_lock(&uma_mtx);
1351 ZONE_LOCK(zone);
1352 keg->uk_flags |= UMA_ZONE_SECONDARY;
1353 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1354 if (LIST_NEXT(z, uz_link) == NULL) {
1355 LIST_INSERT_AFTER(z, zone, uz_link);
1356 break;
1357 }
1358 }
1359 ZONE_UNLOCK(zone);
1360 mtx_unlock(&uma_mtx);
1361 } else if (arg->keg == NULL) {
1362 if (uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1363 arg->align, arg->flags) == NULL)
1364 return (ENOMEM);
1365 } else {
1366 struct uma_kctor_args karg;
1367 int error;
1368
1369 /* We should only be here from uma_startup() */
1370 karg.size = arg->size;
1371 karg.uminit = arg->uminit;
1372 karg.fini = arg->fini;
1373 karg.align = arg->align;
1374 karg.flags = arg->flags;
1375 karg.zone = zone;
1376 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1377 flags);
1378 if (error)
1379 return (error);
1380 }
1381 keg = zone->uz_keg;
1382 zone->uz_lock = &keg->uk_lock;
1383
1384 /*
1385 * Some internal zones don't have room allocated for the per cpu
1386 * caches. If we're internal, bail out here.
1387 */
1388 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1389 KASSERT((keg->uk_flags & UMA_ZONE_SECONDARY) == 0,
1390 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1391 return (0);
1392 }
1393
1394 if (keg->uk_flags & UMA_ZONE_MAXBUCKET)
1395 zone->uz_count = BUCKET_MAX;
1396 else if (keg->uk_ipers <= BUCKET_MAX)
1397 zone->uz_count = keg->uk_ipers;
1398 else
1399 zone->uz_count = BUCKET_MAX;
1400 return (0);
1401 }
1402
1403 /*
1404 * Keg header dtor. This frees all data, destroys locks, frees the hash
1405 * table and removes the keg from the global list.
1406 *
1407 * Arguments/Returns follow uma_dtor specifications
1408 * udata unused
1409 */
1410 static void
1411 keg_dtor(void *arg, int size, void *udata)
1412 {
1413 uma_keg_t keg;
1414
1415 keg = (uma_keg_t)arg;
1416 mtx_lock(&keg->uk_lock);
1417 if (keg->uk_free != 0) {
1418 printf("Freed UMA keg was not empty (%d items). "
1419 " Lost %d pages of memory.\n",
1420 keg->uk_free, keg->uk_pages);
1421 }
1422 mtx_unlock(&keg->uk_lock);
1423
1424 if (keg->uk_flags & UMA_ZONE_HASH)
1425 hash_free(&keg->uk_hash);
1426
1427 mtx_destroy(&keg->uk_lock);
1428 }
1429
1430 /*
1431 * Zone header dtor.
1432 *
1433 * Arguments/Returns follow uma_dtor specifications
1434 * udata unused
1435 */
1436 static void
1437 zone_dtor(void *arg, int size, void *udata)
1438 {
1439 uma_zone_t zone;
1440 uma_keg_t keg;
1441
1442 zone = (uma_zone_t)arg;
1443 keg = zone->uz_keg;
1444
1445 if (!(keg->uk_flags & UMA_ZFLAG_INTERNAL))
1446 cache_drain(zone);
1447
1448 mtx_lock(&uma_mtx);
1449 zone_drain(zone);
1450 if (keg->uk_flags & UMA_ZONE_SECONDARY) {
1451 LIST_REMOVE(zone, uz_link);
1452 /*
1453 * XXX there are some races here where
1454 * the zone can be drained but zone lock
1455 * released and then refilled before we
1456 * remove it... we dont care for now
1457 */
1458 ZONE_LOCK(zone);
1459 if (LIST_EMPTY(&keg->uk_zones))
1460 keg->uk_flags &= ~UMA_ZONE_SECONDARY;
1461 ZONE_UNLOCK(zone);
1462 mtx_unlock(&uma_mtx);
1463 } else {
1464 LIST_REMOVE(keg, uk_link);
1465 LIST_REMOVE(zone, uz_link);
1466 mtx_unlock(&uma_mtx);
1467 uma_zfree_internal(kegs, keg, NULL, SKIP_NONE,
1468 ZFREE_STATFREE);
1469 }
1470 zone->uz_keg = NULL;
1471 }
1472
1473 /*
1474 * Traverses every zone in the system and calls a callback
1475 *
1476 * Arguments:
1477 * zfunc A pointer to a function which accepts a zone
1478 * as an argument.
1479 *
1480 * Returns:
1481 * Nothing
1482 */
1483 static void
1484 zone_foreach(void (*zfunc)(uma_zone_t))
1485 {
1486 uma_keg_t keg;
1487 uma_zone_t zone;
1488
1489 mtx_lock(&uma_mtx);
1490 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1491 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1492 zfunc(zone);
1493 }
1494 mtx_unlock(&uma_mtx);
1495 }
1496
1497 /* Public functions */
1498 /* See uma.h */
1499 void
1500 uma_startup(void *bootmem, int boot_pages)
1501 {
1502 struct uma_zctor_args args;
1503 uma_slab_t slab;
1504 u_int slabsize;
1505 u_int objsize, totsize, wsize;
1506 int i;
1507
1508 #ifdef UMA_DEBUG
1509 printf("Creating uma keg headers zone and keg.\n");
1510 #endif
1511 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1512
1513 /*
1514 * Figure out the maximum number of items-per-slab we'll have if
1515 * we're using the OFFPAGE slab header to track free items, given
1516 * all possible object sizes and the maximum desired wastage
1517 * (UMA_MAX_WASTE).
1518 *
1519 * We iterate until we find an object size for
1520 * which the calculated wastage in zone_small_init() will be
1521 * enough to warrant OFFPAGE. Since wastedspace versus objsize
1522 * is an overall increasing see-saw function, we find the smallest
1523 * objsize such that the wastage is always acceptable for objects
1524 * with that objsize or smaller. Since a smaller objsize always
1525 * generates a larger possible uma_max_ipers, we use this computed
1526 * objsize to calculate the largest ipers possible. Since the
1527 * ipers calculated for OFFPAGE slab headers is always larger than
1528 * the ipers initially calculated in zone_small_init(), we use
1529 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to
1530 * obtain the maximum ipers possible for offpage slab headers.
1531 *
1532 * It should be noted that ipers versus objsize is an inversly
1533 * proportional function which drops off rather quickly so as
1534 * long as our UMA_MAX_WASTE is such that the objsize we calculate
1535 * falls into the portion of the inverse relation AFTER the steep
1536 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386).
1537 *
1538 * Note that we have 8-bits (1 byte) to use as a freelist index
1539 * inside the actual slab header itself and this is enough to
1540 * accomodate us. In the worst case, a UMA_SMALLEST_UNIT sized
1541 * object with offpage slab header would have ipers =
1542 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is
1543 * 1 greater than what our byte-integer freelist index can
1544 * accomodate, but we know that this situation never occurs as
1545 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate
1546 * that we need to go to offpage slab headers. Or, if we do,
1547 * then we trap that condition below and panic in the INVARIANTS case.
1548 */
1549 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - UMA_MAX_WASTE;
1550 totsize = wsize;
1551 objsize = UMA_SMALLEST_UNIT;
1552 while (totsize >= wsize) {
1553 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) /
1554 (objsize + UMA_FRITM_SZ);
1555 totsize *= (UMA_FRITM_SZ + objsize);
1556 objsize++;
1557 }
1558 if (objsize > UMA_SMALLEST_UNIT)
1559 objsize--;
1560 uma_max_ipers = UMA_SLAB_SIZE / objsize;
1561
1562 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - UMA_MAX_WASTE;
1563 totsize = wsize;
1564 objsize = UMA_SMALLEST_UNIT;
1565 while (totsize >= wsize) {
1566 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) /
1567 (objsize + UMA_FRITMREF_SZ);
1568 totsize *= (UMA_FRITMREF_SZ + objsize);
1569 objsize++;
1570 }
1571 if (objsize > UMA_SMALLEST_UNIT)
1572 objsize--;
1573 uma_max_ipers_ref = UMA_SLAB_SIZE / objsize;
1574
1575 KASSERT((uma_max_ipers_ref <= 255) && (uma_max_ipers <= 255),
1576 ("uma_startup: calculated uma_max_ipers values too large!"));
1577
1578 #ifdef UMA_DEBUG
1579 printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers);
1580 printf("Calculated uma_max_ipers_slab (for OFFPAGE) is %d\n",
1581 uma_max_ipers_ref);
1582 #endif
1583
1584 /* "manually" create the initial zone */
1585 args.name = "UMA Kegs";
1586 args.size = sizeof(struct uma_keg);
1587 args.ctor = keg_ctor;
1588 args.dtor = keg_dtor;
1589 args.uminit = zero_init;
1590 args.fini = NULL;
1591 args.keg = &masterkeg;
1592 args.align = 32 - 1;
1593 args.flags = UMA_ZFLAG_INTERNAL;
1594 /* The initial zone has no Per cpu queues so it's smaller */
1595 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1596
1597 #ifdef UMA_DEBUG
1598 printf("Filling boot free list.\n");
1599 #endif
1600 for (i = 0; i < boot_pages; i++) {
1601 slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE));
1602 slab->us_data = (u_int8_t *)slab;
1603 slab->us_flags = UMA_SLAB_BOOT;
1604 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1605 }
1606 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1607
1608 #ifdef UMA_DEBUG
1609 printf("Creating uma zone headers zone and keg.\n");
1610 #endif
1611 args.name = "UMA Zones";
1612 args.size = sizeof(struct uma_zone) +
1613 (sizeof(struct uma_cache) * (mp_maxid + 1));
1614 args.ctor = zone_ctor;
1615 args.dtor = zone_dtor;
1616 args.uminit = zero_init;
1617 args.fini = NULL;
1618 args.keg = NULL;
1619 args.align = 32 - 1;
1620 args.flags = UMA_ZFLAG_INTERNAL;
1621 /* The initial zone has no Per cpu queues so it's smaller */
1622 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1623
1624 #ifdef UMA_DEBUG
1625 printf("Initializing pcpu cache locks.\n");
1626 #endif
1627 #ifdef UMA_DEBUG
1628 printf("Creating slab and hash zones.\n");
1629 #endif
1630
1631 /*
1632 * This is the max number of free list items we'll have with
1633 * offpage slabs.
1634 */
1635 slabsize = uma_max_ipers * UMA_FRITM_SZ;
1636 slabsize += sizeof(struct uma_slab);
1637
1638 /* Now make a zone for slab headers */
1639 slabzone = uma_zcreate("UMA Slabs",
1640 slabsize,
1641 NULL, NULL, NULL, NULL,
1642 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1643
1644 /*
1645 * We also create a zone for the bigger slabs with reference
1646 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1647 */
1648 slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ;
1649 slabsize += sizeof(struct uma_slab_refcnt);
1650 slabrefzone = uma_zcreate("UMA RCntSlabs",
1651 slabsize,
1652 NULL, NULL, NULL, NULL,
1653 UMA_ALIGN_PTR,
1654 UMA_ZFLAG_INTERNAL);
1655
1656 hashzone = uma_zcreate("UMA Hash",
1657 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1658 NULL, NULL, NULL, NULL,
1659 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1660
1661 bucket_init();
1662
1663 #ifdef UMA_MD_SMALL_ALLOC
1664 booted = 1;
1665 #endif
1666
1667 #ifdef UMA_DEBUG
1668 printf("UMA startup complete.\n");
1669 #endif
1670 }
1671
1672 /* see uma.h */
1673 void
1674 uma_startup2(void)
1675 {
1676 booted = 1;
1677 bucket_enable();
1678 #ifdef UMA_DEBUG
1679 printf("UMA startup2 complete.\n");
1680 #endif
1681 }
1682
1683 /*
1684 * Initialize our callout handle
1685 *
1686 */
1687
1688 static void
1689 uma_startup3(void)
1690 {
1691 #ifdef UMA_DEBUG
1692 printf("Starting callout.\n");
1693 #endif
1694 callout_init(&uma_callout, CALLOUT_MPSAFE);
1695 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1696 #ifdef UMA_DEBUG
1697 printf("UMA startup3 complete.\n");
1698 #endif
1699 }
1700
1701 static uma_zone_t
1702 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1703 int align, u_int32_t flags)
1704 {
1705 struct uma_kctor_args args;
1706
1707 args.size = size;
1708 args.uminit = uminit;
1709 args.fini = fini;
1710 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
1711 args.flags = flags;
1712 args.zone = zone;
1713 return (uma_zalloc_internal(kegs, &args, M_WAITOK));
1714 }
1715
1716 /* See uma.h */
1717 void
1718 uma_set_align(int align)
1719 {
1720
1721 if (align != UMA_ALIGN_CACHE)
1722 uma_align_cache = align;
1723 }
1724
1725 /* See uma.h */
1726 uma_zone_t
1727 uma_zcreate(char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1728 uma_init uminit, uma_fini fini, int align, u_int32_t flags)
1729
1730 {
1731 struct uma_zctor_args args;
1732
1733 /* This stuff is essential for the zone ctor */
1734 args.name = name;
1735 args.size = size;
1736 args.ctor = ctor;
1737 args.dtor = dtor;
1738 args.uminit = uminit;
1739 args.fini = fini;
1740 args.align = align;
1741 args.flags = flags;
1742 args.keg = NULL;
1743
1744 return (uma_zalloc_internal(zones, &args, M_WAITOK));
1745 }
1746
1747 /* See uma.h */
1748 uma_zone_t
1749 uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1750 uma_init zinit, uma_fini zfini, uma_zone_t master)
1751 {
1752 struct uma_zctor_args args;
1753
1754 args.name = name;
1755 args.size = master->uz_keg->uk_size;
1756 args.ctor = ctor;
1757 args.dtor = dtor;
1758 args.uminit = zinit;
1759 args.fini = zfini;
1760 args.align = master->uz_keg->uk_align;
1761 args.flags = master->uz_keg->uk_flags | UMA_ZONE_SECONDARY;
1762 args.keg = master->uz_keg;
1763
1764 return (uma_zalloc_internal(zones, &args, M_WAITOK));
1765 }
1766
1767 /* See uma.h */
1768 void
1769 uma_zdestroy(uma_zone_t zone)
1770 {
1771
1772 uma_zfree_internal(zones, zone, NULL, SKIP_NONE, ZFREE_STATFREE);
1773 }
1774
1775 /* See uma.h */
1776 void *
1777 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1778 {
1779 void *item;
1780 uma_cache_t cache;
1781 uma_bucket_t bucket;
1782 int cpu;
1783
1784 /* This is the fast path allocation */
1785 #ifdef UMA_DEBUG_ALLOC_1
1786 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1787 #endif
1788 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1789 zone->uz_name, flags);
1790
1791 if (flags & M_WAITOK) {
1792 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
1793 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
1794 }
1795
1796 /*
1797 * If possible, allocate from the per-CPU cache. There are two
1798 * requirements for safe access to the per-CPU cache: (1) the thread
1799 * accessing the cache must not be preempted or yield during access,
1800 * and (2) the thread must not migrate CPUs without switching which
1801 * cache it accesses. We rely on a critical section to prevent
1802 * preemption and migration. We release the critical section in
1803 * order to acquire the zone mutex if we are unable to allocate from
1804 * the current cache; when we re-acquire the critical section, we
1805 * must detect and handle migration if it has occurred.
1806 */
1807 zalloc_restart:
1808 critical_enter();
1809 cpu = curcpu;
1810 cache = &zone->uz_cpu[cpu];
1811
1812 zalloc_start:
1813 bucket = cache->uc_allocbucket;
1814
1815 if (bucket) {
1816 if (bucket->ub_cnt > 0) {
1817 bucket->ub_cnt--;
1818 item = bucket->ub_bucket[bucket->ub_cnt];
1819 #ifdef INVARIANTS
1820 bucket->ub_bucket[bucket->ub_cnt] = NULL;
1821 #endif
1822 KASSERT(item != NULL,
1823 ("uma_zalloc: Bucket pointer mangled."));
1824 cache->uc_allocs++;
1825 critical_exit();
1826 #ifdef INVARIANTS
1827 ZONE_LOCK(zone);
1828 uma_dbg_alloc(zone, NULL, item);
1829 ZONE_UNLOCK(zone);
1830 #endif
1831 if (zone->uz_ctor != NULL) {
1832 if (zone->uz_ctor(item, zone->uz_keg->uk_size,
1833 udata, flags) != 0) {
1834 uma_zfree_internal(zone, item, udata,
1835 SKIP_DTOR, ZFREE_STATFAIL |
1836 ZFREE_STATFREE);
1837 return (NULL);
1838 }
1839 }
1840 if (flags & M_ZERO)
1841 bzero(item, zone->uz_keg->uk_size);
1842 return (item);
1843 } else if (cache->uc_freebucket) {
1844 /*
1845 * We have run out of items in our allocbucket.
1846 * See if we can switch with our free bucket.
1847 */
1848 if (cache->uc_freebucket->ub_cnt > 0) {
1849 #ifdef UMA_DEBUG_ALLOC
1850 printf("uma_zalloc: Swapping empty with"
1851 " alloc.\n");
1852 #endif
1853 bucket = cache->uc_freebucket;
1854 cache->uc_freebucket = cache->uc_allocbucket;
1855 cache->uc_allocbucket = bucket;
1856
1857 goto zalloc_start;
1858 }
1859 }
1860 }
1861 /*
1862 * Attempt to retrieve the item from the per-CPU cache has failed, so
1863 * we must go back to the zone. This requires the zone lock, so we
1864 * must drop the critical section, then re-acquire it when we go back
1865 * to the cache. Since the critical section is released, we may be
1866 * preempted or migrate. As such, make sure not to maintain any
1867 * thread-local state specific to the cache from prior to releasing
1868 * the critical section.
1869 */
1870 critical_exit();
1871 ZONE_LOCK(zone);
1872 critical_enter();
1873 cpu = curcpu;
1874 cache = &zone->uz_cpu[cpu];
1875 bucket = cache->uc_allocbucket;
1876 if (bucket != NULL) {
1877 if (bucket->ub_cnt > 0) {
1878 ZONE_UNLOCK(zone);
1879 goto zalloc_start;
1880 }
1881 bucket = cache->uc_freebucket;
1882 if (bucket != NULL && bucket->ub_cnt > 0) {
1883 ZONE_UNLOCK(zone);
1884 goto zalloc_start;
1885 }
1886 }
1887
1888 /* Since we have locked the zone we may as well send back our stats */
1889 zone->uz_allocs += cache->uc_allocs;
1890 cache->uc_allocs = 0;
1891 zone->uz_frees += cache->uc_frees;
1892 cache->uc_frees = 0;
1893
1894 /* Our old one is now a free bucket */
1895 if (cache->uc_allocbucket) {
1896 KASSERT(cache->uc_allocbucket->ub_cnt == 0,
1897 ("uma_zalloc_arg: Freeing a non free bucket."));
1898 LIST_INSERT_HEAD(&zone->uz_free_bucket,
1899 cache->uc_allocbucket, ub_link);
1900 cache->uc_allocbucket = NULL;
1901 }
1902
1903 /* Check the free list for a new alloc bucket */
1904 if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
1905 KASSERT(bucket->ub_cnt != 0,
1906 ("uma_zalloc_arg: Returning an empty bucket."));
1907
1908 LIST_REMOVE(bucket, ub_link);
1909 cache->uc_allocbucket = bucket;
1910 ZONE_UNLOCK(zone);
1911 goto zalloc_start;
1912 }
1913 /* We are no longer associated with this CPU. */
1914 critical_exit();
1915
1916 /* Bump up our uz_count so we get here less */
1917 if (zone->uz_count < BUCKET_MAX)
1918 zone->uz_count++;
1919
1920 /*
1921 * Now lets just fill a bucket and put it on the free list. If that
1922 * works we'll restart the allocation from the begining.
1923 */
1924 if (uma_zalloc_bucket(zone, flags)) {
1925 ZONE_UNLOCK(zone);
1926 goto zalloc_restart;
1927 }
1928 ZONE_UNLOCK(zone);
1929 /*
1930 * We may not be able to get a bucket so return an actual item.
1931 */
1932 #ifdef UMA_DEBUG
1933 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
1934 #endif
1935
1936 return (uma_zalloc_internal(zone, udata, flags));
1937 }
1938
1939 static uma_slab_t
1940 uma_zone_slab(uma_zone_t zone, int flags)
1941 {
1942 uma_slab_t slab;
1943 uma_keg_t keg;
1944
1945 keg = zone->uz_keg;
1946
1947 /*
1948 * This is to prevent us from recursively trying to allocate
1949 * buckets. The problem is that if an allocation forces us to
1950 * grab a new bucket we will call page_alloc, which will go off
1951 * and cause the vm to allocate vm_map_entries. If we need new
1952 * buckets there too we will recurse in kmem_alloc and bad
1953 * things happen. So instead we return a NULL bucket, and make
1954 * the code that allocates buckets smart enough to deal with it
1955 *
1956 * XXX: While we want this protection for the bucket zones so that
1957 * recursion from the VM is handled (and the calling code that
1958 * allocates buckets knows how to deal with it), we do not want
1959 * to prevent allocation from the slab header zones (slabzone
1960 * and slabrefzone) if uk_recurse is not zero for them. The
1961 * reason is that it could lead to NULL being returned for
1962 * slab header allocations even in the M_WAITOK case, and the
1963 * caller can't handle that.
1964 */
1965 if (keg->uk_flags & UMA_ZFLAG_INTERNAL && keg->uk_recurse != 0)
1966 if (zone != slabzone && zone != slabrefzone && zone != zones)
1967 return (NULL);
1968
1969 slab = NULL;
1970
1971 for (;;) {
1972 /*
1973 * Find a slab with some space. Prefer slabs that are partially
1974 * used over those that are totally full. This helps to reduce
1975 * fragmentation.
1976 */
1977 if (keg->uk_free != 0) {
1978 if (!LIST_EMPTY(&keg->uk_part_slab)) {
1979 slab = LIST_FIRST(&keg->uk_part_slab);
1980 } else {
1981 slab = LIST_FIRST(&keg->uk_free_slab);
1982 LIST_REMOVE(slab, us_link);
1983 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
1984 us_link);
1985 }
1986 return (slab);
1987 }
1988
1989 /*
1990 * M_NOVM means don't ask at all!
1991 */
1992 if (flags & M_NOVM)
1993 break;
1994
1995 if (keg->uk_maxpages &&
1996 keg->uk_pages >= keg->uk_maxpages) {
1997 keg->uk_flags |= UMA_ZFLAG_FULL;
1998
1999 if (flags & M_NOWAIT)
2000 break;
2001 else
2002 msleep(keg, &keg->uk_lock, PVM,
2003 "zonelimit", 0);
2004 continue;
2005 }
2006 keg->uk_recurse++;
2007 slab = slab_zalloc(zone, flags);
2008 keg->uk_recurse--;
2009
2010 /*
2011 * If we got a slab here it's safe to mark it partially used
2012 * and return. We assume that the caller is going to remove
2013 * at least one item.
2014 */
2015 if (slab) {
2016 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2017 return (slab);
2018 }
2019 /*
2020 * We might not have been able to get a slab but another cpu
2021 * could have while we were unlocked. Check again before we
2022 * fail.
2023 */
2024 if (flags & M_NOWAIT)
2025 flags |= M_NOVM;
2026 }
2027 return (slab);
2028 }
2029
2030 static void *
2031 uma_slab_alloc(uma_zone_t zone, uma_slab_t slab)
2032 {
2033 uma_keg_t keg;
2034 uma_slabrefcnt_t slabref;
2035 void *item;
2036 u_int8_t freei;
2037
2038 keg = zone->uz_keg;
2039
2040 freei = slab->us_firstfree;
2041 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2042 slabref = (uma_slabrefcnt_t)slab;
2043 slab->us_firstfree = slabref->us_freelist[freei].us_item;
2044 } else {
2045 slab->us_firstfree = slab->us_freelist[freei].us_item;
2046 }
2047 item = slab->us_data + (keg->uk_rsize * freei);
2048
2049 slab->us_freecount--;
2050 keg->uk_free--;
2051 #ifdef INVARIANTS
2052 uma_dbg_alloc(zone, slab, item);
2053 #endif
2054 /* Move this slab to the full list */
2055 if (slab->us_freecount == 0) {
2056 LIST_REMOVE(slab, us_link);
2057 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2058 }
2059
2060 return (item);
2061 }
2062
2063 static int
2064 uma_zalloc_bucket(uma_zone_t zone, int flags)
2065 {
2066 uma_bucket_t bucket;
2067 uma_slab_t slab;
2068 int16_t saved;
2069 int max, origflags = flags;
2070
2071 /*
2072 * Try this zone's free list first so we don't allocate extra buckets.
2073 */
2074 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2075 KASSERT(bucket->ub_cnt == 0,
2076 ("uma_zalloc_bucket: Bucket on free list is not empty."));
2077 LIST_REMOVE(bucket, ub_link);
2078 } else {
2079 int bflags;
2080
2081 bflags = (flags & ~M_ZERO);
2082 if (zone->uz_keg->uk_flags & UMA_ZFLAG_CACHEONLY)
2083 bflags |= M_NOVM;
2084
2085 ZONE_UNLOCK(zone);
2086 bucket = bucket_alloc(zone->uz_count, bflags);
2087 ZONE_LOCK(zone);
2088 }
2089
2090 if (bucket == NULL)
2091 return (0);
2092
2093 #ifdef SMP
2094 /*
2095 * This code is here to limit the number of simultaneous bucket fills
2096 * for any given zone to the number of per cpu caches in this zone. This
2097 * is done so that we don't allocate more memory than we really need.
2098 */
2099 if (zone->uz_fills >= mp_ncpus)
2100 goto done;
2101
2102 #endif
2103 zone->uz_fills++;
2104
2105 max = MIN(bucket->ub_entries, zone->uz_count);
2106 /* Try to keep the buckets totally full */
2107 saved = bucket->ub_cnt;
2108 while (bucket->ub_cnt < max &&
2109 (slab = uma_zone_slab(zone, flags)) != NULL) {
2110 while (slab->us_freecount && bucket->ub_cnt < max) {
2111 bucket->ub_bucket[bucket->ub_cnt++] =
2112 uma_slab_alloc(zone, slab);
2113 }
2114
2115 /* Don't block on the next fill */
2116 flags |= M_NOWAIT;
2117 }
2118
2119 /*
2120 * We unlock here because we need to call the zone's init.
2121 * It should be safe to unlock because the slab dealt with
2122 * above is already on the appropriate list within the keg
2123 * and the bucket we filled is not yet on any list, so we
2124 * own it.
2125 */
2126 if (zone->uz_init != NULL) {
2127 int i;
2128
2129 ZONE_UNLOCK(zone);
2130 for (i = saved; i < bucket->ub_cnt; i++)
2131 if (zone->uz_init(bucket->ub_bucket[i],
2132 zone->uz_keg->uk_size, origflags) != 0)
2133 break;
2134 /*
2135 * If we couldn't initialize the whole bucket, put the
2136 * rest back onto the freelist.
2137 */
2138 if (i != bucket->ub_cnt) {
2139 int j;
2140
2141 for (j = i; j < bucket->ub_cnt; j++) {
2142 uma_zfree_internal(zone, bucket->ub_bucket[j],
2143 NULL, SKIP_FINI, 0);
2144 #ifdef INVARIANTS
2145 bucket->ub_bucket[j] = NULL;
2146 #endif
2147 }
2148 bucket->ub_cnt = i;
2149 }
2150 ZONE_LOCK(zone);
2151 }
2152
2153 zone->uz_fills--;
2154 if (bucket->ub_cnt != 0) {
2155 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2156 bucket, ub_link);
2157 return (1);
2158 }
2159 #ifdef SMP
2160 done:
2161 #endif
2162 bucket_free(bucket);
2163
2164 return (0);
2165 }
2166 /*
2167 * Allocates an item for an internal zone
2168 *
2169 * Arguments
2170 * zone The zone to alloc for.
2171 * udata The data to be passed to the constructor.
2172 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2173 *
2174 * Returns
2175 * NULL if there is no memory and M_NOWAIT is set
2176 * An item if successful
2177 */
2178
2179 static void *
2180 uma_zalloc_internal(uma_zone_t zone, void *udata, int flags)
2181 {
2182 uma_keg_t keg;
2183 uma_slab_t slab;
2184 void *item;
2185
2186 item = NULL;
2187 keg = zone->uz_keg;
2188
2189 #ifdef UMA_DEBUG_ALLOC
2190 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2191 #endif
2192 ZONE_LOCK(zone);
2193
2194 slab = uma_zone_slab(zone, flags);
2195 if (slab == NULL) {
2196 zone->uz_fails++;
2197 ZONE_UNLOCK(zone);
2198 return (NULL);
2199 }
2200
2201 item = uma_slab_alloc(zone, slab);
2202
2203 zone->uz_allocs++;
2204
2205 ZONE_UNLOCK(zone);
2206
2207 /*
2208 * We have to call both the zone's init (not the keg's init)
2209 * and the zone's ctor. This is because the item is going from
2210 * a keg slab directly to the user, and the user is expecting it
2211 * to be both zone-init'd as well as zone-ctor'd.
2212 */
2213 if (zone->uz_init != NULL) {
2214 if (zone->uz_init(item, keg->uk_size, flags) != 0) {
2215 uma_zfree_internal(zone, item, udata, SKIP_FINI,
2216 ZFREE_STATFAIL | ZFREE_STATFREE);
2217 return (NULL);
2218 }
2219 }
2220 if (zone->uz_ctor != NULL) {
2221 if (zone->uz_ctor(item, keg->uk_size, udata, flags) != 0) {
2222 uma_zfree_internal(zone, item, udata, SKIP_DTOR,
2223 ZFREE_STATFAIL | ZFREE_STATFREE);
2224 return (NULL);
2225 }
2226 }
2227 if (flags & M_ZERO)
2228 bzero(item, keg->uk_size);
2229
2230 return (item);
2231 }
2232
2233 /* See uma.h */
2234 void
2235 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2236 {
2237 uma_keg_t keg;
2238 uma_cache_t cache;
2239 uma_bucket_t bucket;
2240 int bflags;
2241 int cpu;
2242
2243 keg = zone->uz_keg;
2244
2245 #ifdef UMA_DEBUG_ALLOC_1
2246 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2247 #endif
2248 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2249 zone->uz_name);
2250
2251 if (zone->uz_dtor)
2252 zone->uz_dtor(item, keg->uk_size, udata);
2253 #ifdef INVARIANTS
2254 ZONE_LOCK(zone);
2255 if (keg->uk_flags & UMA_ZONE_MALLOC)
2256 uma_dbg_free(zone, udata, item);
2257 else
2258 uma_dbg_free(zone, NULL, item);
2259 ZONE_UNLOCK(zone);
2260 #endif
2261 /*
2262 * The race here is acceptable. If we miss it we'll just have to wait
2263 * a little longer for the limits to be reset.
2264 */
2265 if (keg->uk_flags & UMA_ZFLAG_FULL)
2266 goto zfree_internal;
2267
2268 /*
2269 * If possible, free to the per-CPU cache. There are two
2270 * requirements for safe access to the per-CPU cache: (1) the thread
2271 * accessing the cache must not be preempted or yield during access,
2272 * and (2) the thread must not migrate CPUs without switching which
2273 * cache it accesses. We rely on a critical section to prevent
2274 * preemption and migration. We release the critical section in
2275 * order to acquire the zone mutex if we are unable to free to the
2276 * current cache; when we re-acquire the critical section, we must
2277 * detect and handle migration if it has occurred.
2278 */
2279 zfree_restart:
2280 critical_enter();
2281 cpu = curcpu;
2282 cache = &zone->uz_cpu[cpu];
2283
2284 zfree_start:
2285 bucket = cache->uc_freebucket;
2286
2287 if (bucket) {
2288 /*
2289 * Do we have room in our bucket? It is OK for this uz count
2290 * check to be slightly out of sync.
2291 */
2292
2293 if (bucket->ub_cnt < bucket->ub_entries) {
2294 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2295 ("uma_zfree: Freeing to non free bucket index."));
2296 bucket->ub_bucket[bucket->ub_cnt] = item;
2297 bucket->ub_cnt++;
2298 cache->uc_frees++;
2299 critical_exit();
2300 return;
2301 } else if (cache->uc_allocbucket) {
2302 #ifdef UMA_DEBUG_ALLOC
2303 printf("uma_zfree: Swapping buckets.\n");
2304 #endif
2305 /*
2306 * We have run out of space in our freebucket.
2307 * See if we can switch with our alloc bucket.
2308 */
2309 if (cache->uc_allocbucket->ub_cnt <
2310 cache->uc_freebucket->ub_cnt) {
2311 bucket = cache->uc_freebucket;
2312 cache->uc_freebucket = cache->uc_allocbucket;
2313 cache->uc_allocbucket = bucket;
2314 goto zfree_start;
2315 }
2316 }
2317 }
2318 /*
2319 * We can get here for two reasons:
2320 *
2321 * 1) The buckets are NULL
2322 * 2) The alloc and free buckets are both somewhat full.
2323 *
2324 * We must go back the zone, which requires acquiring the zone lock,
2325 * which in turn means we must release and re-acquire the critical
2326 * section. Since the critical section is released, we may be
2327 * preempted or migrate. As such, make sure not to maintain any
2328 * thread-local state specific to the cache from prior to releasing
2329 * the critical section.
2330 */
2331 critical_exit();
2332 ZONE_LOCK(zone);
2333 critical_enter();
2334 cpu = curcpu;
2335 cache = &zone->uz_cpu[cpu];
2336 if (cache->uc_freebucket != NULL) {
2337 if (cache->uc_freebucket->ub_cnt <
2338 cache->uc_freebucket->ub_entries) {
2339 ZONE_UNLOCK(zone);
2340 goto zfree_start;
2341 }
2342 if (cache->uc_allocbucket != NULL &&
2343 (cache->uc_allocbucket->ub_cnt <
2344 cache->uc_freebucket->ub_cnt)) {
2345 ZONE_UNLOCK(zone);
2346 goto zfree_start;
2347 }
2348 }
2349
2350 /* Since we have locked the zone we may as well send back our stats */
2351 zone->uz_allocs += cache->uc_allocs;
2352 cache->uc_allocs = 0;
2353 zone->uz_frees += cache->uc_frees;
2354 cache->uc_frees = 0;
2355
2356 bucket = cache->uc_freebucket;
2357 cache->uc_freebucket = NULL;
2358
2359 /* Can we throw this on the zone full list? */
2360 if (bucket != NULL) {
2361 #ifdef UMA_DEBUG_ALLOC
2362 printf("uma_zfree: Putting old bucket on the free list.\n");
2363 #endif
2364 /* ub_cnt is pointing to the last free item */
2365 KASSERT(bucket->ub_cnt != 0,
2366 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2367 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2368 bucket, ub_link);
2369 }
2370 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2371 LIST_REMOVE(bucket, ub_link);
2372 ZONE_UNLOCK(zone);
2373 cache->uc_freebucket = bucket;
2374 goto zfree_start;
2375 }
2376 /* We are no longer associated with this CPU. */
2377 critical_exit();
2378
2379 /* And the zone.. */
2380 ZONE_UNLOCK(zone);
2381
2382 #ifdef UMA_DEBUG_ALLOC
2383 printf("uma_zfree: Allocating new free bucket.\n");
2384 #endif
2385 bflags = M_NOWAIT;
2386
2387 if (keg->uk_flags & UMA_ZFLAG_CACHEONLY)
2388 bflags |= M_NOVM;
2389 bucket = bucket_alloc(zone->uz_count, bflags);
2390 if (bucket) {
2391 ZONE_LOCK(zone);
2392 LIST_INSERT_HEAD(&zone->uz_free_bucket,
2393 bucket, ub_link);
2394 ZONE_UNLOCK(zone);
2395 goto zfree_restart;
2396 }
2397
2398 /*
2399 * If nothing else caught this, we'll just do an internal free.
2400 */
2401 zfree_internal:
2402 uma_zfree_internal(zone, item, udata, SKIP_DTOR, ZFREE_STATFREE);
2403
2404 return;
2405 }
2406
2407 /*
2408 * Frees an item to an INTERNAL zone or allocates a free bucket
2409 *
2410 * Arguments:
2411 * zone The zone to free to
2412 * item The item we're freeing
2413 * udata User supplied data for the dtor
2414 * skip Skip dtors and finis
2415 */
2416 static void
2417 uma_zfree_internal(uma_zone_t zone, void *item, void *udata,
2418 enum zfreeskip skip, int flags)
2419 {
2420 uma_slab_t slab;
2421 uma_slabrefcnt_t slabref;
2422 uma_keg_t keg;
2423 u_int8_t *mem;
2424 u_int8_t freei;
2425
2426 keg = zone->uz_keg;
2427
2428 if (skip < SKIP_DTOR && zone->uz_dtor)
2429 zone->uz_dtor(item, keg->uk_size, udata);
2430 if (skip < SKIP_FINI && zone->uz_fini)
2431 zone->uz_fini(item, keg->uk_size);
2432
2433 ZONE_LOCK(zone);
2434
2435 if (flags & ZFREE_STATFAIL)
2436 zone->uz_fails++;
2437 if (flags & ZFREE_STATFREE)
2438 zone->uz_frees++;
2439
2440 if (!(keg->uk_flags & UMA_ZONE_MALLOC)) {
2441 mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK));
2442 if (keg->uk_flags & UMA_ZONE_HASH)
2443 slab = hash_sfind(&keg->uk_hash, mem);
2444 else {
2445 mem += keg->uk_pgoff;
2446 slab = (uma_slab_t)mem;
2447 }
2448 } else {
2449 slab = (uma_slab_t)udata;
2450 }
2451
2452 /* Do we need to remove from any lists? */
2453 if (slab->us_freecount+1 == keg->uk_ipers) {
2454 LIST_REMOVE(slab, us_link);
2455 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2456 } else if (slab->us_freecount == 0) {
2457 LIST_REMOVE(slab, us_link);
2458 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2459 }
2460
2461 /* Slab management stuff */
2462 freei = ((unsigned long)item - (unsigned long)slab->us_data)
2463 / keg->uk_rsize;
2464
2465 #ifdef INVARIANTS
2466 if (!skip)
2467 uma_dbg_free(zone, slab, item);
2468 #endif
2469
2470 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2471 slabref = (uma_slabrefcnt_t)slab;
2472 slabref->us_freelist[freei].us_item = slab->us_firstfree;
2473 } else {
2474 slab->us_freelist[freei].us_item = slab->us_firstfree;
2475 }
2476 slab->us_firstfree = freei;
2477 slab->us_freecount++;
2478
2479 /* Zone statistics */
2480 keg->uk_free++;
2481
2482 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2483 if (keg->uk_pages < keg->uk_maxpages)
2484 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2485
2486 /*
2487 * We can handle one more allocation. Since we're clearing ZFLAG_FULL,
2488 * wake up all procs blocked on pages. This should be uncommon, so
2489 * keeping this simple for now (rather than adding count of blocked
2490 * threads etc).
2491 */
2492 wakeup(keg);
2493 }
2494
2495 ZONE_UNLOCK(zone);
2496 }
2497
2498 /* See uma.h */
2499 void
2500 uma_zone_set_max(uma_zone_t zone, int nitems)
2501 {
2502 uma_keg_t keg;
2503
2504 keg = zone->uz_keg;
2505 ZONE_LOCK(zone);
2506 if (keg->uk_ppera > 1)
2507 keg->uk_maxpages = nitems * keg->uk_ppera;
2508 else
2509 keg->uk_maxpages = nitems / keg->uk_ipers;
2510
2511 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2512 keg->uk_maxpages++;
2513
2514 ZONE_UNLOCK(zone);
2515 }
2516
2517 /* See uma.h */
2518 void
2519 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2520 {
2521 ZONE_LOCK(zone);
2522 KASSERT(zone->uz_keg->uk_pages == 0,
2523 ("uma_zone_set_init on non-empty keg"));
2524 zone->uz_keg->uk_init = uminit;
2525 ZONE_UNLOCK(zone);
2526 }
2527
2528 /* See uma.h */
2529 void
2530 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2531 {
2532 ZONE_LOCK(zone);
2533 KASSERT(zone->uz_keg->uk_pages == 0,
2534 ("uma_zone_set_fini on non-empty keg"));
2535 zone->uz_keg->uk_fini = fini;
2536 ZONE_UNLOCK(zone);
2537 }
2538
2539 /* See uma.h */
2540 void
2541 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2542 {
2543 ZONE_LOCK(zone);
2544 KASSERT(zone->uz_keg->uk_pages == 0,
2545 ("uma_zone_set_zinit on non-empty keg"));
2546 zone->uz_init = zinit;
2547 ZONE_UNLOCK(zone);
2548 }
2549
2550 /* See uma.h */
2551 void
2552 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2553 {
2554 ZONE_LOCK(zone);
2555 KASSERT(zone->uz_keg->uk_pages == 0,
2556 ("uma_zone_set_zfini on non-empty keg"));
2557 zone->uz_fini = zfini;
2558 ZONE_UNLOCK(zone);
2559 }
2560
2561 /* See uma.h */
2562 /* XXX uk_freef is not actually used with the zone locked */
2563 void
2564 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2565 {
2566 ZONE_LOCK(zone);
2567 zone->uz_keg->uk_freef = freef;
2568 ZONE_UNLOCK(zone);
2569 }
2570
2571 /* See uma.h */
2572 /* XXX uk_allocf is not actually used with the zone locked */
2573 void
2574 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2575 {
2576 ZONE_LOCK(zone);
2577 zone->uz_keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
2578 zone->uz_keg->uk_allocf = allocf;
2579 ZONE_UNLOCK(zone);
2580 }
2581
2582 /* See uma.h */
2583 int
2584 uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count)
2585 {
2586 uma_keg_t keg;
2587 vm_offset_t kva;
2588 int pages;
2589
2590 keg = zone->uz_keg;
2591 pages = count / keg->uk_ipers;
2592
2593 if (pages * keg->uk_ipers < count)
2594 pages++;
2595
2596 kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2597
2598 if (kva == 0)
2599 return (0);
2600 if (obj == NULL) {
2601 obj = vm_object_allocate(OBJT_DEFAULT,
2602 pages);
2603 } else {
2604 VM_OBJECT_LOCK_INIT(obj, "uma object");
2605 _vm_object_allocate(OBJT_DEFAULT,
2606 pages, obj);
2607 }
2608 ZONE_LOCK(zone);
2609 keg->uk_kva = kva;
2610 keg->uk_obj = obj;
2611 keg->uk_maxpages = pages;
2612 keg->uk_allocf = obj_alloc;
2613 keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
2614 ZONE_UNLOCK(zone);
2615 return (1);
2616 }
2617
2618 /* See uma.h */
2619 void
2620 uma_prealloc(uma_zone_t zone, int items)
2621 {
2622 int slabs;
2623 uma_slab_t slab;
2624 uma_keg_t keg;
2625
2626 keg = zone->uz_keg;
2627 ZONE_LOCK(zone);
2628 slabs = items / keg->uk_ipers;
2629 if (slabs * keg->uk_ipers < items)
2630 slabs++;
2631 while (slabs > 0) {
2632 slab = slab_zalloc(zone, M_WAITOK);
2633 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2634 slabs--;
2635 }
2636 ZONE_UNLOCK(zone);
2637 }
2638
2639 /* See uma.h */
2640 u_int32_t *
2641 uma_find_refcnt(uma_zone_t zone, void *item)
2642 {
2643 uma_slabrefcnt_t slabref;
2644 uma_keg_t keg;
2645 u_int32_t *refcnt;
2646 int idx;
2647
2648 keg = zone->uz_keg;
2649 slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item &
2650 (~UMA_SLAB_MASK));
2651 KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT,
2652 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
2653 idx = ((unsigned long)item - (unsigned long)slabref->us_data)
2654 / keg->uk_rsize;
2655 refcnt = &slabref->us_freelist[idx].us_refcnt;
2656 return refcnt;
2657 }
2658
2659 /* See uma.h */
2660 void
2661 uma_reclaim(void)
2662 {
2663 #ifdef UMA_DEBUG
2664 printf("UMA: vm asked us to release pages!\n");
2665 #endif
2666 bucket_enable();
2667 zone_foreach(zone_drain);
2668 /*
2669 * Some slabs may have been freed but this zone will be visited early
2670 * we visit again so that we can free pages that are empty once other
2671 * zones are drained. We have to do the same for buckets.
2672 */
2673 zone_drain(slabzone);
2674 zone_drain(slabrefzone);
2675 bucket_zone_drain();
2676 }
2677
2678 /* See uma.h */
2679 int
2680 uma_zone_exhausted(uma_zone_t zone)
2681 {
2682 int full;
2683
2684 ZONE_LOCK(zone);
2685 full = (zone->uz_keg->uk_flags & UMA_ZFLAG_FULL);
2686 ZONE_UNLOCK(zone);
2687 return (full);
2688 }
2689
2690 int
2691 uma_zone_exhausted_nolock(uma_zone_t zone)
2692 {
2693 return (zone->uz_keg->uk_flags & UMA_ZFLAG_FULL);
2694 }
2695
2696 void *
2697 uma_large_malloc(int size, int wait)
2698 {
2699 void *mem;
2700 uma_slab_t slab;
2701 u_int8_t flags;
2702
2703 slab = uma_zalloc_internal(slabzone, NULL, wait);
2704 if (slab == NULL)
2705 return (NULL);
2706 mem = page_alloc(NULL, size, &flags, wait);
2707 if (mem) {
2708 vsetslab((vm_offset_t)mem, slab);
2709 slab->us_data = mem;
2710 slab->us_flags = flags | UMA_SLAB_MALLOC;
2711 slab->us_size = size;
2712 } else {
2713 uma_zfree_internal(slabzone, slab, NULL, SKIP_NONE,
2714 ZFREE_STATFAIL | ZFREE_STATFREE);
2715 }
2716
2717 return (mem);
2718 }
2719
2720 void
2721 uma_large_free(uma_slab_t slab)
2722 {
2723 vsetobj((vm_offset_t)slab->us_data, kmem_object);
2724 page_free(slab->us_data, slab->us_size, slab->us_flags);
2725 uma_zfree_internal(slabzone, slab, NULL, SKIP_NONE, ZFREE_STATFREE);
2726 }
2727
2728 void
2729 uma_print_stats(void)
2730 {
2731 zone_foreach(uma_print_zone);
2732 }
2733
2734 static void
2735 slab_print(uma_slab_t slab)
2736 {
2737 printf("slab: keg %p, data %p, freecount %d, firstfree %d\n",
2738 slab->us_keg, slab->us_data, slab->us_freecount,
2739 slab->us_firstfree);
2740 }
2741
2742 static void
2743 cache_print(uma_cache_t cache)
2744 {
2745 printf("alloc: %p(%d), free: %p(%d)\n",
2746 cache->uc_allocbucket,
2747 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
2748 cache->uc_freebucket,
2749 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
2750 }
2751
2752 void
2753 uma_print_zone(uma_zone_t zone)
2754 {
2755 uma_cache_t cache;
2756 uma_keg_t keg;
2757 uma_slab_t slab;
2758 int i;
2759
2760 keg = zone->uz_keg;
2761 printf("%s(%p) size %d(%d) flags %d ipers %d ppera %d out %d free %d\n",
2762 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
2763 keg->uk_ipers, keg->uk_ppera,
2764 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
2765 printf("Part slabs:\n");
2766 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
2767 slab_print(slab);
2768 printf("Free slabs:\n");
2769 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
2770 slab_print(slab);
2771 printf("Full slabs:\n");
2772 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
2773 slab_print(slab);
2774 for (i = 0; i <= mp_maxid; i++) {
2775 if (CPU_ABSENT(i))
2776 continue;
2777 cache = &zone->uz_cpu[i];
2778 printf("CPU %d Cache:\n", i);
2779 cache_print(cache);
2780 }
2781 }
2782
2783 #ifdef DDB
2784 /*
2785 * Generate statistics across both the zone and its per-cpu cache's. Return
2786 * desired statistics if the pointer is non-NULL for that statistic.
2787 *
2788 * Note: does not update the zone statistics, as it can't safely clear the
2789 * per-CPU cache statistic.
2790 *
2791 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
2792 * safe from off-CPU; we should modify the caches to track this information
2793 * directly so that we don't have to.
2794 */
2795 static void
2796 uma_zone_sumstat(uma_zone_t z, int *cachefreep, u_int64_t *allocsp,
2797 u_int64_t *freesp)
2798 {
2799 uma_cache_t cache;
2800 u_int64_t allocs, frees;
2801 int cachefree, cpu;
2802
2803 allocs = frees = 0;
2804 cachefree = 0;
2805 for (cpu = 0; cpu <= mp_maxid; cpu++) {
2806 if (CPU_ABSENT(cpu))
2807 continue;
2808 cache = &z->uz_cpu[cpu];
2809 if (cache->uc_allocbucket != NULL)
2810 cachefree += cache->uc_allocbucket->ub_cnt;
2811 if (cache->uc_freebucket != NULL)
2812 cachefree += cache->uc_freebucket->ub_cnt;
2813 allocs += cache->uc_allocs;
2814 frees += cache->uc_frees;
2815 }
2816 allocs += z->uz_allocs;
2817 frees += z->uz_frees;
2818 if (cachefreep != NULL)
2819 *cachefreep = cachefree;
2820 if (allocsp != NULL)
2821 *allocsp = allocs;
2822 if (freesp != NULL)
2823 *freesp = frees;
2824 }
2825 #endif /* DDB */
2826
2827 static int
2828 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
2829 {
2830 uma_keg_t kz;
2831 uma_zone_t z;
2832 int count;
2833
2834 count = 0;
2835 mtx_lock(&uma_mtx);
2836 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2837 LIST_FOREACH(z, &kz->uk_zones, uz_link)
2838 count++;
2839 }
2840 mtx_unlock(&uma_mtx);
2841 return (sysctl_handle_int(oidp, &count, 0, req));
2842 }
2843
2844 static int
2845 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
2846 {
2847 struct uma_stream_header ush;
2848 struct uma_type_header uth;
2849 struct uma_percpu_stat ups;
2850 uma_bucket_t bucket;
2851 struct sbuf sbuf;
2852 uma_cache_t cache;
2853 uma_keg_t kz;
2854 uma_zone_t z;
2855 char *buffer;
2856 int buflen, count, error, i;
2857
2858 mtx_lock(&uma_mtx);
2859 restart:
2860 mtx_assert(&uma_mtx, MA_OWNED);
2861 count = 0;
2862 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2863 LIST_FOREACH(z, &kz->uk_zones, uz_link)
2864 count++;
2865 }
2866 mtx_unlock(&uma_mtx);
2867
2868 buflen = sizeof(ush) + count * (sizeof(uth) + sizeof(ups) *
2869 (mp_maxid + 1)) + 1;
2870 buffer = malloc(buflen, M_TEMP, M_WAITOK | M_ZERO);
2871
2872 mtx_lock(&uma_mtx);
2873 i = 0;
2874 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2875 LIST_FOREACH(z, &kz->uk_zones, uz_link)
2876 i++;
2877 }
2878 if (i > count) {
2879 free(buffer, M_TEMP);
2880 goto restart;
2881 }
2882 count = i;
2883
2884 sbuf_new(&sbuf, buffer, buflen, SBUF_FIXEDLEN);
2885
2886 /*
2887 * Insert stream header.
2888 */
2889 bzero(&ush, sizeof(ush));
2890 ush.ush_version = UMA_STREAM_VERSION;
2891 ush.ush_maxcpus = (mp_maxid + 1);
2892 ush.ush_count = count;
2893 if (sbuf_bcat(&sbuf, &ush, sizeof(ush)) < 0) {
2894 mtx_unlock(&uma_mtx);
2895 error = ENOMEM;
2896 goto out;
2897 }
2898
2899 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2900 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
2901 bzero(&uth, sizeof(uth));
2902 ZONE_LOCK(z);
2903 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
2904 uth.uth_align = kz->uk_align;
2905 uth.uth_pages = kz->uk_pages;
2906 uth.uth_keg_free = kz->uk_free;
2907 uth.uth_size = kz->uk_size;
2908 uth.uth_rsize = kz->uk_rsize;
2909 uth.uth_maxpages = kz->uk_maxpages;
2910 if (kz->uk_ppera > 1)
2911 uth.uth_limit = kz->uk_maxpages /
2912 kz->uk_ppera;
2913 else
2914 uth.uth_limit = kz->uk_maxpages *
2915 kz->uk_ipers;
2916
2917 /*
2918 * A zone is secondary is it is not the first entry
2919 * on the keg's zone list.
2920 */
2921 if ((kz->uk_flags & UMA_ZONE_SECONDARY) &&
2922 (LIST_FIRST(&kz->uk_zones) != z))
2923 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
2924
2925 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
2926 uth.uth_zone_free += bucket->ub_cnt;
2927 uth.uth_allocs = z->uz_allocs;
2928 uth.uth_frees = z->uz_frees;
2929 uth.uth_fails = z->uz_fails;
2930 if (sbuf_bcat(&sbuf, &uth, sizeof(uth)) < 0) {
2931 ZONE_UNLOCK(z);
2932 mtx_unlock(&uma_mtx);
2933 error = ENOMEM;
2934 goto out;
2935 }
2936 /*
2937 * While it is not normally safe to access the cache
2938 * bucket pointers while not on the CPU that owns the
2939 * cache, we only allow the pointers to be exchanged
2940 * without the zone lock held, not invalidated, so
2941 * accept the possible race associated with bucket
2942 * exchange during monitoring.
2943 */
2944 for (i = 0; i < (mp_maxid + 1); i++) {
2945 bzero(&ups, sizeof(ups));
2946 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
2947 goto skip;
2948 if (CPU_ABSENT(i))
2949 goto skip;
2950 cache = &z->uz_cpu[i];
2951 if (cache->uc_allocbucket != NULL)
2952 ups.ups_cache_free +=
2953 cache->uc_allocbucket->ub_cnt;
2954 if (cache->uc_freebucket != NULL)
2955 ups.ups_cache_free +=
2956 cache->uc_freebucket->ub_cnt;
2957 ups.ups_allocs = cache->uc_allocs;
2958 ups.ups_frees = cache->uc_frees;
2959 skip:
2960 if (sbuf_bcat(&sbuf, &ups, sizeof(ups)) < 0) {
2961 ZONE_UNLOCK(z);
2962 mtx_unlock(&uma_mtx);
2963 error = ENOMEM;
2964 goto out;
2965 }
2966 }
2967 ZONE_UNLOCK(z);
2968 }
2969 }
2970 mtx_unlock(&uma_mtx);
2971 sbuf_finish(&sbuf);
2972 error = SYSCTL_OUT(req, sbuf_data(&sbuf), sbuf_len(&sbuf));
2973 out:
2974 free(buffer, M_TEMP);
2975 return (error);
2976 }
2977
2978 #ifdef DDB
2979 DB_SHOW_COMMAND(uma, db_show_uma)
2980 {
2981 u_int64_t allocs, frees;
2982 uma_bucket_t bucket;
2983 uma_keg_t kz;
2984 uma_zone_t z;
2985 int cachefree;
2986
2987 db_printf("%18s %8s %8s %8s %12s\n", "Zone", "Size", "Used", "Free",
2988 "Requests");
2989 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2990 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
2991 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
2992 allocs = z->uz_allocs;
2993 frees = z->uz_frees;
2994 cachefree = 0;
2995 } else
2996 uma_zone_sumstat(z, &cachefree, &allocs,
2997 &frees);
2998 if (!((kz->uk_flags & UMA_ZONE_SECONDARY) &&
2999 (LIST_FIRST(&kz->uk_zones) != z)))
3000 cachefree += kz->uk_free;
3001 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3002 cachefree += bucket->ub_cnt;
3003 db_printf("%18s %8ju %8jd %8d %12ju\n", z->uz_name,
3004 (uintmax_t)kz->uk_size,
3005 (intmax_t)(allocs - frees), cachefree,
3006 (uintmax_t)allocs);
3007 }
3008 }
3009 }
3010 #endif
Cache object: 4f0eb9f2fc635e7c8b913d5dc332cdea
|