117static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
118
119/*
120 * Are we allowed to allocate buckets?
121 */
122static int bucketdisable = 1;
123
124/* Linked list of all kegs in the system */
125static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(&uma_kegs);
126
127/* This mutex protects the keg list */
128static struct mtx uma_mtx;
129
130/* Linked list of boot time pages */
131static LIST_HEAD(,uma_slab) uma_boot_pages =
132 LIST_HEAD_INITIALIZER(&uma_boot_pages);
133
134/* This mutex protects the boot time pages list */
135static struct mtx uma_boot_pages_mtx;
136
137/* Is the VM done starting up? */
138static int booted = 0;
139
140/* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */
141static u_int uma_max_ipers;
142static 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 */
148static 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 */
155struct 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
167struct 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
176struct 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
184struct 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 */
199static uint8_t bucket_size[BUCKET_ZONES];
200
201/*
202 * Flags and enumerations to be passed to internal functions.
203 */
204enum 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
211static void *obj_alloc(uma_zone_t, int, u_int8_t *, int);
212static void *page_alloc(uma_zone_t, int, u_int8_t *, int);
213static void *startup_alloc(uma_zone_t, int, u_int8_t *, int);
214static void page_free(void *, int, u_int8_t);
215static uma_slab_t slab_zalloc(uma_zone_t, int);
216static void cache_drain(uma_zone_t);
217static void bucket_drain(uma_zone_t, uma_bucket_t);
218static void bucket_cache_drain(uma_zone_t zone);
219static int keg_ctor(void *, int, void *, int);
220static void keg_dtor(void *, int, void *);
221static int zone_ctor(void *, int, void *, int);
222static void zone_dtor(void *, int, void *);
223static int zero_init(void *, int, int);
224static void zone_small_init(uma_zone_t zone);
225static void zone_large_init(uma_zone_t zone);
226static void zone_foreach(void (*zfunc)(uma_zone_t));
227static void zone_timeout(uma_zone_t zone);
228static int hash_alloc(struct uma_hash *);
229static int hash_expand(struct uma_hash *, struct uma_hash *);
230static void hash_free(struct uma_hash *hash);
231static void uma_timeout(void *);
232static void uma_startup3(void);
233static void *uma_zalloc_internal(uma_zone_t, void *, int);
234static void uma_zfree_internal(uma_zone_t, void *, void *, enum zfreeskip,
235 int);
236static void bucket_enable(void);
237static void bucket_init(void);
238static uma_bucket_t bucket_alloc(int, int);
239static void bucket_free(uma_bucket_t);
240static void bucket_zone_drain(void);
241static int uma_zalloc_bucket(uma_zone_t zone, int flags);
242static uma_slab_t uma_zone_slab(uma_zone_t zone, int flags);
243static void *uma_slab_alloc(uma_zone_t zone, uma_slab_t slab);
244static 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
247void uma_print_zone(uma_zone_t);
248void uma_print_stats(void);
249static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
250static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
251
252#ifdef WITNESS
253static int nosleepwithlocks = 1;
254#else
255static int nosleepwithlocks = 0;
256#endif
257SYSCTL_INT(_debug, OID_AUTO, nosleepwithlocks, CTLFLAG_RW, &nosleepwithlocks,
258 0, "Convert M_WAITOK to M_NOWAIT to avoid lock-held-across-sleep paths");
259SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
260
261SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT,
262 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
263
264SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT,
265 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
266
267/*
268 * This routine checks to see whether or not it's safe to enable buckets.
269 */
270
271static void
272bucket_enable(void)
273{
274 if (cnt.v_free_count < cnt.v_free_min)
275 bucketdisable = 1;
276 else
277 bucketdisable = 0;
278}
279
280/*
281 * Initialize bucket_zones, the array of zones of buckets of various sizes.
282 *
283 * For each zone, calculate the memory required for each bucket, consisting
284 * of the header and an array of pointers. Initialize bucket_size[] to point
285 * the range of appropriate bucket sizes at the zone.
286 */
287static void
288bucket_init(void)
289{
290 struct uma_bucket_zone *ubz;
291 int i;
292 int j;
293
294 for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) {
295 int size;
296
297 ubz = &bucket_zones[j];
298 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
299 size += sizeof(void *) * ubz->ubz_entries;
300 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
301 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
302 for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT))
303 bucket_size[i >> BUCKET_SHIFT] = j;
304 }
305}
306
307/*
308 * Given a desired number of entries for a bucket, return the zone from which
309 * to allocate the bucket.
310 */
311static struct uma_bucket_zone *
312bucket_zone_lookup(int entries)
313{
314 int idx;
315
316 idx = howmany(entries, 1 << BUCKET_SHIFT);
317 return (&bucket_zones[bucket_size[idx]]);
318}
319
320static uma_bucket_t
321bucket_alloc(int entries, int bflags)
322{
323 struct uma_bucket_zone *ubz;
324 uma_bucket_t bucket;
325
326 /*
327 * This is to stop us from allocating per cpu buckets while we're
328 * running out of vm.boot_pages. Otherwise, we would exhaust the
329 * boot pages. This also prevents us from allocating buckets in
330 * low memory situations.
331 */
332 if (bucketdisable)
333 return (NULL);
334
335 ubz = bucket_zone_lookup(entries);
336 bucket = uma_zalloc_internal(ubz->ubz_zone, NULL, bflags);
337 if (bucket) {
338#ifdef INVARIANTS
339 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
340#endif
341 bucket->ub_cnt = 0;
342 bucket->ub_entries = ubz->ubz_entries;
343 }
344
345 return (bucket);
346}
347
348static void
349bucket_free(uma_bucket_t bucket)
350{
351 struct uma_bucket_zone *ubz;
352
353 ubz = bucket_zone_lookup(bucket->ub_entries);
354 uma_zfree_internal(ubz->ubz_zone, bucket, NULL, SKIP_NONE,
355 ZFREE_STATFREE);
356}
357
358static void
359bucket_zone_drain(void)
360{
361 struct uma_bucket_zone *ubz;
362
363 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
364 zone_drain(ubz->ubz_zone);
365}
366
367
368/*
369 * Routine called by timeout which is used to fire off some time interval
370 * based calculations. (stats, hash size, etc.)
371 *
372 * Arguments:
373 * arg Unused
374 *
375 * Returns:
376 * Nothing
377 */
378static void
379uma_timeout(void *unused)
380{
381 bucket_enable();
382 zone_foreach(zone_timeout);
383
384 /* Reschedule this event */
385 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
386}
387
388/*
389 * Routine to perform timeout driven calculations. This expands the
390 * hashes and does per cpu statistics aggregation.
391 *
392 * Arguments:
393 * zone The zone to operate on
394 *
395 * Returns:
396 * Nothing
397 */
398static void
399zone_timeout(uma_zone_t zone)
400{
401 uma_keg_t keg;
402 u_int64_t alloc;
403
404 keg = zone->uz_keg;
405 alloc = 0;
406
407 /*
408 * Expand the zone hash table.
409 *
410 * This is done if the number of slabs is larger than the hash size.
411 * What I'm trying to do here is completely reduce collisions. This
412 * may be a little aggressive. Should I allow for two collisions max?
413 */
414 ZONE_LOCK(zone);
415 if (keg->uk_flags & UMA_ZONE_HASH &&
416 keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) {
417 struct uma_hash newhash;
418 struct uma_hash oldhash;
419 int ret;
420
421 /*
422 * This is so involved because allocating and freeing
423 * while the zone lock is held will lead to deadlock.
424 * I have to do everything in stages and check for
425 * races.
426 */
427 newhash = keg->uk_hash;
428 ZONE_UNLOCK(zone);
429 ret = hash_alloc(&newhash);
430 ZONE_LOCK(zone);
431 if (ret) {
432 if (hash_expand(&keg->uk_hash, &newhash)) {
433 oldhash = keg->uk_hash;
434 keg->uk_hash = newhash;
435 } else
436 oldhash = newhash;
437
438 ZONE_UNLOCK(zone);
439 hash_free(&oldhash);
440 ZONE_LOCK(zone);
441 }
442 }
443 ZONE_UNLOCK(zone);
444}
445
446/*
447 * Allocate and zero fill the next sized hash table from the appropriate
448 * backing store.
449 *
450 * Arguments:
451 * hash A new hash structure with the old hash size in uh_hashsize
452 *
453 * Returns:
454 * 1 on sucess and 0 on failure.
455 */
456static int
457hash_alloc(struct uma_hash *hash)
458{
459 int oldsize;
460 int alloc;
461
462 oldsize = hash->uh_hashsize;
463
464 /* We're just going to go to a power of two greater */
465 if (oldsize) {
466 hash->uh_hashsize = oldsize * 2;
467 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
468 hash->uh_slab_hash = (struct slabhead *)malloc(alloc,
469 M_UMAHASH, M_NOWAIT);
470 } else {
471 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
472 hash->uh_slab_hash = uma_zalloc_internal(hashzone, NULL,
473 M_WAITOK);
474 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
475 }
476 if (hash->uh_slab_hash) {
477 bzero(hash->uh_slab_hash, alloc);
478 hash->uh_hashmask = hash->uh_hashsize - 1;
479 return (1);
480 }
481
482 return (0);
483}
484
485/*
486 * Expands the hash table for HASH zones. This is done from zone_timeout
487 * to reduce collisions. This must not be done in the regular allocation
488 * path, otherwise, we can recurse on the vm while allocating pages.
489 *
490 * Arguments:
491 * oldhash The hash you want to expand
492 * newhash The hash structure for the new table
493 *
494 * Returns:
495 * Nothing
496 *
497 * Discussion:
498 */
499static int
500hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
501{
502 uma_slab_t slab;
503 int hval;
504 int i;
505
506 if (!newhash->uh_slab_hash)
507 return (0);
508
509 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
510 return (0);
511
512 /*
513 * I need to investigate hash algorithms for resizing without a
514 * full rehash.
515 */
516
517 for (i = 0; i < oldhash->uh_hashsize; i++)
518 while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) {
519 slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]);
520 SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink);
521 hval = UMA_HASH(newhash, slab->us_data);
522 SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
523 slab, us_hlink);
524 }
525
526 return (1);
527}
528
529/*
530 * Free the hash bucket to the appropriate backing store.
531 *
532 * Arguments:
533 * slab_hash The hash bucket we're freeing
534 * hashsize The number of entries in that hash bucket
535 *
536 * Returns:
537 * Nothing
538 */
539static void
540hash_free(struct uma_hash *hash)
541{
542 if (hash->uh_slab_hash == NULL)
543 return;
544 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
545 uma_zfree_internal(hashzone,
546 hash->uh_slab_hash, NULL, SKIP_NONE, ZFREE_STATFREE);
547 else
548 free(hash->uh_slab_hash, M_UMAHASH);
549}
550
551/*
552 * Frees all outstanding items in a bucket
553 *
554 * Arguments:
555 * zone The zone to free to, must be unlocked.
556 * bucket The free/alloc bucket with items, cpu queue must be locked.
557 *
558 * Returns:
559 * Nothing
560 */
561
562static void
563bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
564{
565 uma_slab_t slab;
566 int mzone;
567 void *item;
568
569 if (bucket == NULL)
570 return;
571
572 slab = NULL;
573 mzone = 0;
574
575 /* We have to lookup the slab again for malloc.. */
576 if (zone->uz_keg->uk_flags & UMA_ZONE_MALLOC)
577 mzone = 1;
578
579 while (bucket->ub_cnt > 0) {
580 bucket->ub_cnt--;
581 item = bucket->ub_bucket[bucket->ub_cnt];
582#ifdef INVARIANTS
583 bucket->ub_bucket[bucket->ub_cnt] = NULL;
584 KASSERT(item != NULL,
585 ("bucket_drain: botched ptr, item is NULL"));
586#endif
587 /*
588 * This is extremely inefficient. The slab pointer was passed
589 * to uma_zfree_arg, but we lost it because the buckets don't
590 * hold them. This will go away when free() gets a size passed
591 * to it.
592 */
593 if (mzone)
594 slab = vtoslab((vm_offset_t)item & (~UMA_SLAB_MASK));
595 uma_zfree_internal(zone, item, slab, SKIP_DTOR, 0);
596 }
597}
598
599/*
600 * Drains the per cpu caches for a zone.
601 *
602 * NOTE: This may only be called while the zone is being turn down, and not
603 * during normal operation. This is necessary in order that we do not have
604 * to migrate CPUs to drain the per-CPU caches.
605 *
606 * Arguments:
607 * zone The zone to drain, must be unlocked.
608 *
609 * Returns:
610 * Nothing
611 */
612static void
613cache_drain(uma_zone_t zone)
614{
615 uma_cache_t cache;
616 int cpu;
617
618 /*
619 * XXX: It is safe to not lock the per-CPU caches, because we're
620 * tearing down the zone anyway. I.e., there will be no further use
621 * of the caches at this point.
622 *
623 * XXX: It would good to be able to assert that the zone is being
624 * torn down to prevent improper use of cache_drain().
625 *
626 * XXX: We lock the zone before passing into bucket_cache_drain() as
627 * it is used elsewhere. Should the tear-down path be made special
628 * there in some form?
629 */
630 for (cpu = 0; cpu <= mp_maxid; cpu++) {
631 if (CPU_ABSENT(cpu))
632 continue;
633 cache = &zone->uz_cpu[cpu];
634 bucket_drain(zone, cache->uc_allocbucket);
635 bucket_drain(zone, cache->uc_freebucket);
636 if (cache->uc_allocbucket != NULL)
637 bucket_free(cache->uc_allocbucket);
638 if (cache->uc_freebucket != NULL)
639 bucket_free(cache->uc_freebucket);
640 cache->uc_allocbucket = cache->uc_freebucket = NULL;
641 }
642 ZONE_LOCK(zone);
643 bucket_cache_drain(zone);
644 ZONE_UNLOCK(zone);
645}
646
647/*
648 * Drain the cached buckets from a zone. Expects a locked zone on entry.
649 */
650static void
651bucket_cache_drain(uma_zone_t zone)
652{
653 uma_bucket_t bucket;
654
655 /*
656 * Drain the bucket queues and free the buckets, we just keep two per
657 * cpu (alloc/free).
658 */
659 while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
660 LIST_REMOVE(bucket, ub_link);
661 ZONE_UNLOCK(zone);
662 bucket_drain(zone, bucket);
663 bucket_free(bucket);
664 ZONE_LOCK(zone);
665 }
666
667 /* Now we do the free queue.. */
668 while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
669 LIST_REMOVE(bucket, ub_link);
670 bucket_free(bucket);
671 }
672}
673
674/*
675 * Frees pages from a zone back to the system. This is done on demand from
676 * the pageout daemon.
677 *
678 * Arguments:
679 * zone The zone to free pages from
680 * all Should we drain all items?
681 *
682 * Returns:
683 * Nothing.
684 */
685void
686zone_drain(uma_zone_t zone)
687{
688 struct slabhead freeslabs = { 0 };
689 uma_keg_t keg;
690 uma_slab_t slab;
691 uma_slab_t n;
692 u_int8_t flags;
693 u_int8_t *mem;
694 int i;
695
696 keg = zone->uz_keg;
697
698 /*
699 * We don't want to take pages from statically allocated zones at this
700 * time
701 */
702 if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL)
703 return;
704
705 ZONE_LOCK(zone);
706
707#ifdef UMA_DEBUG
708 printf("%s free items: %u\n", zone->uz_name, keg->uk_free);
709#endif
710 bucket_cache_drain(zone);
711 if (keg->uk_free == 0)
712 goto finished;
713
714 slab = LIST_FIRST(&keg->uk_free_slab);
715 while (slab) {
716 n = LIST_NEXT(slab, us_link);
717
718 /* We have no where to free these to */
719 if (slab->us_flags & UMA_SLAB_BOOT) {
720 slab = n;
721 continue;
722 }
723
724 LIST_REMOVE(slab, us_link);
725 keg->uk_pages -= keg->uk_ppera;
726 keg->uk_free -= keg->uk_ipers;
727
728 if (keg->uk_flags & UMA_ZONE_HASH)
729 UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data);
730
731 SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink);
732
733 slab = n;
734 }
735finished:
736 ZONE_UNLOCK(zone);
737
738 while ((slab = SLIST_FIRST(&freeslabs)) != NULL) {
739 SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink);
740 if (keg->uk_fini)
741 for (i = 0; i < keg->uk_ipers; i++)
742 keg->uk_fini(
743 slab->us_data + (keg->uk_rsize * i),
744 keg->uk_size);
745 flags = slab->us_flags;
746 mem = slab->us_data;
747
748 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
749 (keg->uk_flags & UMA_ZONE_REFCNT)) {
750 vm_object_t obj;
751
752 if (flags & UMA_SLAB_KMEM)
753 obj = kmem_object;
754 else
755 obj = NULL;
756 for (i = 0; i < keg->uk_ppera; i++)
757 vsetobj((vm_offset_t)mem + (i * PAGE_SIZE),
758 obj);
759 }
760 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
761 uma_zfree_internal(keg->uk_slabzone, slab, NULL,
762 SKIP_NONE, ZFREE_STATFREE);
763#ifdef UMA_DEBUG
764 printf("%s: Returning %d bytes.\n",
765 zone->uz_name, UMA_SLAB_SIZE * keg->uk_ppera);
766#endif
767 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, flags);
768 }
769}
770
771/*
772 * Allocate a new slab for a zone. This does not insert the slab onto a list.
773 *
774 * Arguments:
775 * zone The zone to allocate slabs for
776 * wait Shall we wait?
777 *
778 * Returns:
779 * The slab that was allocated or NULL if there is no memory and the
780 * caller specified M_NOWAIT.
781 */
782static uma_slab_t
783slab_zalloc(uma_zone_t zone, int wait)
784{
785 uma_slabrefcnt_t slabref;
786 uma_slab_t slab;
787 uma_keg_t keg;
788 u_int8_t *mem;
789 u_int8_t flags;
790 int i;
791
792 slab = NULL;
793 keg = zone->uz_keg;
794
795#ifdef UMA_DEBUG
796 printf("slab_zalloc: Allocating a new slab for %s\n", zone->uz_name);
797#endif
798 ZONE_UNLOCK(zone);
799
800 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
801 slab = uma_zalloc_internal(keg->uk_slabzone, NULL, wait);
802 if (slab == NULL) {
803 ZONE_LOCK(zone);
804 return NULL;
805 }
806 }
807
808 /*
809 * This reproduces the old vm_zone behavior of zero filling pages the
810 * first time they are added to a zone.
811 *
812 * Malloced items are zeroed in uma_zalloc.
813 */
814
815 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
816 wait |= M_ZERO;
817 else
818 wait &= ~M_ZERO;
819
820 mem = keg->uk_allocf(zone, keg->uk_ppera * UMA_SLAB_SIZE,
821 &flags, wait);
822 if (mem == NULL) {
823 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
824 uma_zfree_internal(keg->uk_slabzone, slab, NULL,
825 SKIP_NONE, ZFREE_STATFREE);
826 ZONE_LOCK(zone);
827 return (NULL);
828 }
829
830 /* Point the slab into the allocated memory */
831 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE))
832 slab = (uma_slab_t )(mem + keg->uk_pgoff);
833
834 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
835 (keg->uk_flags & UMA_ZONE_REFCNT))
836 for (i = 0; i < keg->uk_ppera; i++)
837 vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab);
838
839 slab->us_keg = keg;
840 slab->us_data = mem;
841 slab->us_freecount = keg->uk_ipers;
842 slab->us_firstfree = 0;
843 slab->us_flags = flags;
844
845 if (keg->uk_flags & UMA_ZONE_REFCNT) {
846 slabref = (uma_slabrefcnt_t)slab;
847 for (i = 0; i < keg->uk_ipers; i++) {
848 slabref->us_freelist[i].us_refcnt = 0;
849 slabref->us_freelist[i].us_item = i+1;
850 }
851 } else {
852 for (i = 0; i < keg->uk_ipers; i++)
853 slab->us_freelist[i].us_item = i+1;
854 }
855
856 if (keg->uk_init != NULL) {
857 for (i = 0; i < keg->uk_ipers; i++)
858 if (keg->uk_init(slab->us_data + (keg->uk_rsize * i),
859 keg->uk_size, wait) != 0)
860 break;
861 if (i != keg->uk_ipers) {
862 if (keg->uk_fini != NULL) {
863 for (i--; i > -1; i--)
864 keg->uk_fini(slab->us_data +
865 (keg->uk_rsize * i),
866 keg->uk_size);
867 }
868 if ((keg->uk_flags & UMA_ZONE_MALLOC) ||
869 (keg->uk_flags & UMA_ZONE_REFCNT)) {
870 vm_object_t obj;
871
872 if (flags & UMA_SLAB_KMEM)
873 obj = kmem_object;
874 else
875 obj = NULL;
876 for (i = 0; i < keg->uk_ppera; i++)
877 vsetobj((vm_offset_t)mem +
878 (i * PAGE_SIZE), obj);
879 }
880 if (keg->uk_flags & UMA_ZONE_OFFPAGE)
881 uma_zfree_internal(keg->uk_slabzone, slab,
882 NULL, SKIP_NONE, ZFREE_STATFREE);
883 keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera,
884 flags);
885 ZONE_LOCK(zone);
886 return (NULL);
887 }
888 }
889 ZONE_LOCK(zone);
890
891 if (keg->uk_flags & UMA_ZONE_HASH)
892 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
893
894 keg->uk_pages += keg->uk_ppera;
895 keg->uk_free += keg->uk_ipers;
896
897 return (slab);
898}
899
900/*
901 * This function is intended to be used early on in place of page_alloc() so
902 * that we may use the boot time page cache to satisfy allocations before
903 * the VM is ready.
904 */
905static void *
906startup_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
907{
908 uma_keg_t keg;
909 uma_slab_t tmps;
910
911 keg = zone->uz_keg;
912
913 /*
914 * Check our small startup cache to see if it has pages remaining.
915 */
916 mtx_lock(&uma_boot_pages_mtx);
917 if ((tmps = LIST_FIRST(&uma_boot_pages)) != NULL) {
918 LIST_REMOVE(tmps, us_link);
919 mtx_unlock(&uma_boot_pages_mtx);
920 *pflag = tmps->us_flags;
921 return (tmps->us_data);
922 }
923 mtx_unlock(&uma_boot_pages_mtx);
924 if (booted == 0)
925 panic("UMA: Increase vm.boot_pages");
926 /*
927 * Now that we've booted reset these users to their real allocator.
928 */
929#ifdef UMA_MD_SMALL_ALLOC
930 keg->uk_allocf = uma_small_alloc;
931#else
932 keg->uk_allocf = page_alloc;
933#endif
934 return keg->uk_allocf(zone, bytes, pflag, wait);
935}
936
937/*
938 * Allocates a number of pages from the system
939 *
940 * Arguments:
941 * zone Unused
942 * bytes The number of bytes requested
943 * wait Shall we wait?
944 *
945 * Returns:
946 * A pointer to the alloced memory or possibly
947 * NULL if M_NOWAIT is set.
948 */
949static void *
950page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait)
951{
952 void *p; /* Returned page */
953
954 *pflag = UMA_SLAB_KMEM;
955 p = (void *) kmem_malloc(kmem_map, bytes, wait);
956
957 return (p);
958}
959
960/*
961 * Allocates a number of pages from within an object
962 *
963 * Arguments:
964 * zone Unused
965 * bytes The number of bytes requested
966 * wait Shall we wait?
967 *
968 * Returns:
969 * A pointer to the alloced memory or possibly
970 * NULL if M_NOWAIT is set.
971 */
972static void *
973obj_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
974{
975 vm_object_t object;
976 vm_offset_t retkva, zkva;
977 vm_page_t p;
978 int pages, startpages;
979
980 object = zone->uz_keg->uk_obj;
981 retkva = 0;
982
983 /*
984 * This looks a little weird since we're getting one page at a time.
985 */
986 VM_OBJECT_LOCK(object);
987 p = TAILQ_LAST(&object->memq, pglist);
988 pages = p != NULL ? p->pindex + 1 : 0;
989 startpages = pages;
990 zkva = zone->uz_keg->uk_kva + pages * PAGE_SIZE;
991 for (; bytes > 0; bytes -= PAGE_SIZE) {
992 p = vm_page_alloc(object, pages,
993 VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED);
994 if (p == NULL) {
995 if (pages != startpages)
996 pmap_qremove(retkva, pages - startpages);
997 while (pages != startpages) {
998 pages--;
999 p = TAILQ_LAST(&object->memq, pglist);
1000 vm_page_lock_queues();
1001 vm_page_unwire(p, 0);
1002 vm_page_free(p);
1003 vm_page_unlock_queues();
1004 }
1005 retkva = 0;
1006 goto done;
1007 }
1008 pmap_qenter(zkva, &p, 1);
1009 if (retkva == 0)
1010 retkva = zkva;
1011 zkva += PAGE_SIZE;
1012 pages += 1;
1013 }
1014done:
1015 VM_OBJECT_UNLOCK(object);
1016 *flags = UMA_SLAB_PRIV;
1017
1018 return ((void *)retkva);
1019}
1020
1021/*
1022 * Frees a number of pages to the system
1023 *
1024 * Arguments:
1025 * mem A pointer to the memory to be freed
1026 * size The size of the memory being freed
1027 * flags The original p->us_flags field
1028 *
1029 * Returns:
1030 * Nothing
1031 */
1032static void
1033page_free(void *mem, int size, u_int8_t flags)
1034{
1035 vm_map_t map;
1036
1037 if (flags & UMA_SLAB_KMEM)
1038 map = kmem_map;
1039 else
1040 panic("UMA: page_free used with invalid flags %d\n", flags);
1041
1042 kmem_free(map, (vm_offset_t)mem, size);
1043}
1044
1045/*
1046 * Zero fill initializer
1047 *
1048 * Arguments/Returns follow uma_init specifications
1049 */
1050static int
1051zero_init(void *mem, int size, int flags)
1052{
1053 bzero(mem, size);
1054 return (0);
1055}
1056
1057/*
1058 * Finish creating a small uma zone. This calculates ipers, and the zone size.
1059 *
1060 * Arguments
1061 * zone The zone we should initialize
1062 *
1063 * Returns
1064 * Nothing
1065 */
1066static void
1067zone_small_init(uma_zone_t zone)
1068{
1069 uma_keg_t keg;
1070 u_int rsize;
1071 u_int memused;
1072 u_int wastedspace;
1073 u_int shsize;
1074
1075 keg = zone->uz_keg;
1076 KASSERT(keg != NULL, ("Keg is null in zone_small_init"));
1077 rsize = keg->uk_size;
1078
1079 if (rsize < UMA_SMALLEST_UNIT)
1080 rsize = UMA_SMALLEST_UNIT;
1081 if (rsize & keg->uk_align)
1082 rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1);
1083
1084 keg->uk_rsize = rsize;
1085 keg->uk_ppera = 1;
1086
1087 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1088 rsize += UMA_FRITMREF_SZ; /* linkage & refcnt */
1089 shsize = sizeof(struct uma_slab_refcnt);
1090 } else {
1091 rsize += UMA_FRITM_SZ; /* Account for linkage */
1092 shsize = sizeof(struct uma_slab);
1093 }
1094
1095 keg->uk_ipers = (UMA_SLAB_SIZE - shsize) / rsize;
1096 KASSERT(keg->uk_ipers != 0, ("zone_small_init: ipers is 0"));
1097 memused = keg->uk_ipers * rsize + shsize;
1098 wastedspace = UMA_SLAB_SIZE - memused;
1099
1100 /*
1101 * We can't do OFFPAGE if we're internal or if we've been
1102 * asked to not go to the VM for buckets. If we do this we
1103 * may end up going to the VM (kmem_map) for slabs which we
1104 * do not want to do if we're UMA_ZFLAG_CACHEONLY as a
1105 * result of UMA_ZONE_VM, which clearly forbids it.
1106 */
1107 if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) ||
1108 (keg->uk_flags & UMA_ZFLAG_CACHEONLY))
1109 return;
1110
1111 if ((wastedspace >= UMA_MAX_WASTE) &&
1112 (keg->uk_ipers < (UMA_SLAB_SIZE / keg->uk_rsize))) {
1113 keg->uk_ipers = UMA_SLAB_SIZE / keg->uk_rsize;
1114 KASSERT(keg->uk_ipers <= 255,
1115 ("zone_small_init: keg->uk_ipers too high!"));
1116#ifdef UMA_DEBUG
1117 printf("UMA decided we need offpage slab headers for "
1118 "zone: %s, calculated wastedspace = %d, "
1119 "maximum wasted space allowed = %d, "
1120 "calculated ipers = %d, "
1121 "new wasted space = %d\n", zone->uz_name, wastedspace,
1122 UMA_MAX_WASTE, keg->uk_ipers,
1123 UMA_SLAB_SIZE - keg->uk_ipers * keg->uk_rsize);
1124#endif
1125 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1126 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1127 keg->uk_flags |= UMA_ZONE_HASH;
1128 }
1129}
1130
1131/*
1132 * Finish creating a large (> UMA_SLAB_SIZE) uma zone. Just give in and do
1133 * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be
1134 * more complicated.
1135 *
1136 * Arguments
1137 * zone The zone we should initialize
1138 *
1139 * Returns
1140 * Nothing
1141 */
1142static void
1143zone_large_init(uma_zone_t zone)
1144{
1145 uma_keg_t keg;
1146 int pages;
1147
1148 keg = zone->uz_keg;
1149
1150 KASSERT(keg != NULL, ("Keg is null in zone_large_init"));
1151 KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0,
1152 ("zone_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY zone"));
1153
1154 pages = keg->uk_size / UMA_SLAB_SIZE;
1155
1156 /* Account for remainder */
1157 if ((pages * UMA_SLAB_SIZE) < keg->uk_size)
1158 pages++;
1159
1160 keg->uk_ppera = pages;
1161 keg->uk_ipers = 1;
1162
1163 keg->uk_flags |= UMA_ZONE_OFFPAGE;
1164 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1165 keg->uk_flags |= UMA_ZONE_HASH;
1166
1167 keg->uk_rsize = keg->uk_size;
1168}
1169
1170/*
1171 * Keg header ctor. This initializes all fields, locks, etc. And inserts
1172 * the keg onto the global keg list.
1173 *
1174 * Arguments/Returns follow uma_ctor specifications
1175 * udata Actually uma_kctor_args
1176 */
1177static int
1178keg_ctor(void *mem, int size, void *udata, int flags)
1179{
1180 struct uma_kctor_args *arg = udata;
1181 uma_keg_t keg = mem;
1182 uma_zone_t zone;
1183
1184 bzero(keg, size);
1185 keg->uk_size = arg->size;
1186 keg->uk_init = arg->uminit;
1187 keg->uk_fini = arg->fini;
1188 keg->uk_align = arg->align;
1189 keg->uk_free = 0;
1190 keg->uk_pages = 0;
1191 keg->uk_flags = arg->flags;
1192 keg->uk_allocf = page_alloc;
1193 keg->uk_freef = page_free;
1194 keg->uk_recurse = 0;
1195 keg->uk_slabzone = NULL;
1196
1197 /*
1198 * The master zone is passed to us at keg-creation time.
1199 */
1200 zone = arg->zone;
1201 zone->uz_keg = keg;
1202
1203 if (arg->flags & UMA_ZONE_VM)
1204 keg->uk_flags |= UMA_ZFLAG_CACHEONLY;
1205
1206 if (arg->flags & UMA_ZONE_ZINIT)
1207 keg->uk_init = zero_init;
1208
1209 /*
1210 * The +UMA_FRITM_SZ added to uk_size is to account for the
1211 * linkage that is added to the size in zone_small_init(). If
1212 * we don't account for this here then we may end up in
1213 * zone_small_init() with a calculated 'ipers' of 0.
1214 */
1215 if (keg->uk_flags & UMA_ZONE_REFCNT) {
1216 if ((keg->uk_size+UMA_FRITMREF_SZ) >
1217 (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)))
1218 zone_large_init(zone);
1219 else
1220 zone_small_init(zone);
1221 } else {
1222 if ((keg->uk_size+UMA_FRITM_SZ) >
1223 (UMA_SLAB_SIZE - sizeof(struct uma_slab)))
1224 zone_large_init(zone);
1225 else
1226 zone_small_init(zone);
1227 }
1228
1229 if (keg->uk_flags & UMA_ZONE_OFFPAGE) {
1230 if (keg->uk_flags & UMA_ZONE_REFCNT)
1231 keg->uk_slabzone = slabrefzone;
1232 else
1233 keg->uk_slabzone = slabzone;
1234 }
1235
1236 /*
1237 * If we haven't booted yet we need allocations to go through the
1238 * startup cache until the vm is ready.
1239 */
1240 if (keg->uk_ppera == 1) {
1241#ifdef UMA_MD_SMALL_ALLOC
1242 keg->uk_allocf = uma_small_alloc;
1243 keg->uk_freef = uma_small_free;
1244#endif
1245 if (booted == 0)
1246 keg->uk_allocf = startup_alloc;
1247 }
1248
1249 /*
1250 * Initialize keg's lock (shared among zones) through
1251 * Master zone
1252 */
1253 zone->uz_lock = &keg->uk_lock;
1254 if (arg->flags & UMA_ZONE_MTXCLASS)
1255 ZONE_LOCK_INIT(zone, 1);
1256 else
1257 ZONE_LOCK_INIT(zone, 0);
1258
1259 /*
1260 * If we're putting the slab header in the actual page we need to
1261 * figure out where in each page it goes. This calculates a right
1262 * justified offset into the memory on an ALIGN_PTR boundary.
1263 */
1264 if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) {
1265 u_int totsize;
1266
1267 /* Size of the slab struct and free list */
1268 if (keg->uk_flags & UMA_ZONE_REFCNT)
1269 totsize = sizeof(struct uma_slab_refcnt) +
1270 keg->uk_ipers * UMA_FRITMREF_SZ;
1271 else
1272 totsize = sizeof(struct uma_slab) +
1273 keg->uk_ipers * UMA_FRITM_SZ;
1274
1275 if (totsize & UMA_ALIGN_PTR)
1276 totsize = (totsize & ~UMA_ALIGN_PTR) +
1277 (UMA_ALIGN_PTR + 1);
1278 keg->uk_pgoff = UMA_SLAB_SIZE - totsize;
1279
1280 if (keg->uk_flags & UMA_ZONE_REFCNT)
1281 totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt)
1282 + keg->uk_ipers * UMA_FRITMREF_SZ;
1283 else
1284 totsize = keg->uk_pgoff + sizeof(struct uma_slab)
1285 + keg->uk_ipers * UMA_FRITM_SZ;
1286
1287 /*
1288 * The only way the following is possible is if with our
1289 * UMA_ALIGN_PTR adjustments we are now bigger than
1290 * UMA_SLAB_SIZE. I haven't checked whether this is
1291 * mathematically possible for all cases, so we make
1292 * sure here anyway.
1293 */
1294 if (totsize > UMA_SLAB_SIZE) {
1295 printf("zone %s ipers %d rsize %d size %d\n",
1296 zone->uz_name, keg->uk_ipers, keg->uk_rsize,
1297 keg->uk_size);
1298 panic("UMA slab won't fit.\n");
1299 }
1300 }
1301
1302 if (keg->uk_flags & UMA_ZONE_HASH)
1303 hash_alloc(&keg->uk_hash);
1304
1305#ifdef UMA_DEBUG
1306 printf("%s(%p) size = %d ipers = %d ppera = %d pgoff = %d\n",
1307 zone->uz_name, zone,
1308 keg->uk_size, keg->uk_ipers,
1309 keg->uk_ppera, keg->uk_pgoff);
1310#endif
1311
1312 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
1313
1314 mtx_lock(&uma_mtx);
1315 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
1316 mtx_unlock(&uma_mtx);
1317 return (0);
1318}
1319
1320/*
1321 * Zone header ctor. This initializes all fields, locks, etc.
1322 *
1323 * Arguments/Returns follow uma_ctor specifications
1324 * udata Actually uma_zctor_args
1325 */
1326
1327static int
1328zone_ctor(void *mem, int size, void *udata, int flags)
1329{
1330 struct uma_zctor_args *arg = udata;
1331 uma_zone_t zone = mem;
1332 uma_zone_t z;
1333 uma_keg_t keg;
1334
1335 bzero(zone, size);
1336 zone->uz_name = arg->name;
1337 zone->uz_ctor = arg->ctor;
1338 zone->uz_dtor = arg->dtor;
1339 zone->uz_init = NULL;
1340 zone->uz_fini = NULL;
1341 zone->uz_allocs = 0;
1342 zone->uz_frees = 0;
1343 zone->uz_fails = 0;
1344 zone->uz_fills = zone->uz_count = 0;
1345
1346 if (arg->flags & UMA_ZONE_SECONDARY) {
1347 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
1348 keg = arg->keg;
1349 zone->uz_keg = keg;
1350 zone->uz_init = arg->uminit;
1351 zone->uz_fini = arg->fini;
1352 zone->uz_lock = &keg->uk_lock;
1353 mtx_lock(&uma_mtx);
1354 ZONE_LOCK(zone);
1355 keg->uk_flags |= UMA_ZONE_SECONDARY;
1356 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
1357 if (LIST_NEXT(z, uz_link) == NULL) {
1358 LIST_INSERT_AFTER(z, zone, uz_link);
1359 break;
1360 }
1361 }
1362 ZONE_UNLOCK(zone);
1363 mtx_unlock(&uma_mtx);
1364 } else if (arg->keg == NULL) {
1365 if (uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
1366 arg->align, arg->flags) == NULL)
1367 return (ENOMEM);
1368 } else {
1369 struct uma_kctor_args karg;
1370 int error;
1371
1372 /* We should only be here from uma_startup() */
1373 karg.size = arg->size;
1374 karg.uminit = arg->uminit;
1375 karg.fini = arg->fini;
1376 karg.align = arg->align;
1377 karg.flags = arg->flags;
1378 karg.zone = zone;
1379 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
1380 flags);
1381 if (error)
1382 return (error);
1383 }
1384 keg = zone->uz_keg;
1385 zone->uz_lock = &keg->uk_lock;
1386
1387 /*
1388 * Some internal zones don't have room allocated for the per cpu
1389 * caches. If we're internal, bail out here.
1390 */
1391 if (keg->uk_flags & UMA_ZFLAG_INTERNAL) {
1392 KASSERT((keg->uk_flags & UMA_ZONE_SECONDARY) == 0,
1393 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
1394 return (0);
1395 }
1396
1397 if (keg->uk_flags & UMA_ZONE_MAXBUCKET)
1398 zone->uz_count = BUCKET_MAX;
1399 else if (keg->uk_ipers <= BUCKET_MAX)
1400 zone->uz_count = keg->uk_ipers;
1401 else
1402 zone->uz_count = BUCKET_MAX;
1403 return (0);
1404}
1405
1406/*
1407 * Keg header dtor. This frees all data, destroys locks, frees the hash
1408 * table and removes the keg from the global list.
1409 *
1410 * Arguments/Returns follow uma_dtor specifications
1411 * udata unused
1412 */
1413static void
1414keg_dtor(void *arg, int size, void *udata)
1415{
1416 uma_keg_t keg;
1417
1418 keg = (uma_keg_t)arg;
1419 mtx_lock(&keg->uk_lock);
1420 if (keg->uk_free != 0) {
1421 printf("Freed UMA keg was not empty (%d items). "
1422 " Lost %d pages of memory.\n",
1423 keg->uk_free, keg->uk_pages);
1424 }
1425 mtx_unlock(&keg->uk_lock);
1426
1427 if (keg->uk_flags & UMA_ZONE_HASH)
1428 hash_free(&keg->uk_hash);
1429
1430 mtx_destroy(&keg->uk_lock);
1431}
1432
1433/*
1434 * Zone header dtor.
1435 *
1436 * Arguments/Returns follow uma_dtor specifications
1437 * udata unused
1438 */
1439static void
1440zone_dtor(void *arg, int size, void *udata)
1441{
1442 uma_zone_t zone;
1443 uma_keg_t keg;
1444
1445 zone = (uma_zone_t)arg;
1446 keg = zone->uz_keg;
1447
1448 if (!(keg->uk_flags & UMA_ZFLAG_INTERNAL))
1449 cache_drain(zone);
1450
1451 mtx_lock(&uma_mtx);
1452 zone_drain(zone);
1453 if (keg->uk_flags & UMA_ZONE_SECONDARY) {
1454 LIST_REMOVE(zone, uz_link);
1455 /*
1456 * XXX there are some races here where
1457 * the zone can be drained but zone lock
1458 * released and then refilled before we
1459 * remove it... we dont care for now
1460 */
1461 ZONE_LOCK(zone);
1462 if (LIST_EMPTY(&keg->uk_zones))
1463 keg->uk_flags &= ~UMA_ZONE_SECONDARY;
1464 ZONE_UNLOCK(zone);
1465 mtx_unlock(&uma_mtx);
1466 } else {
1467 LIST_REMOVE(keg, uk_link);
1468 LIST_REMOVE(zone, uz_link);
1469 mtx_unlock(&uma_mtx);
1470 uma_zfree_internal(kegs, keg, NULL, SKIP_NONE,
1471 ZFREE_STATFREE);
1472 }
1473 zone->uz_keg = NULL;
1474}
1475
1476/*
1477 * Traverses every zone in the system and calls a callback
1478 *
1479 * Arguments:
1480 * zfunc A pointer to a function which accepts a zone
1481 * as an argument.
1482 *
1483 * Returns:
1484 * Nothing
1485 */
1486static void
1487zone_foreach(void (*zfunc)(uma_zone_t))
1488{
1489 uma_keg_t keg;
1490 uma_zone_t zone;
1491
1492 mtx_lock(&uma_mtx);
1493 LIST_FOREACH(keg, &uma_kegs, uk_link) {
1494 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
1495 zfunc(zone);
1496 }
1497 mtx_unlock(&uma_mtx);
1498}
1499
1500/* Public functions */
1501/* See uma.h */
1502void
1503uma_startup(void *bootmem, int boot_pages)
1504{
1505 struct uma_zctor_args args;
1506 uma_slab_t slab;
1507 u_int slabsize;
1508 u_int objsize, totsize, wsize;
1509 int i;
1510
1511#ifdef UMA_DEBUG
1512 printf("Creating uma keg headers zone and keg.\n");
1513#endif
1514 mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF);
1515
1516 /*
1517 * Figure out the maximum number of items-per-slab we'll have if
1518 * we're using the OFFPAGE slab header to track free items, given
1519 * all possible object sizes and the maximum desired wastage
1520 * (UMA_MAX_WASTE).
1521 *
1522 * We iterate until we find an object size for
1523 * which the calculated wastage in zone_small_init() will be
1524 * enough to warrant OFFPAGE. Since wastedspace versus objsize
1525 * is an overall increasing see-saw function, we find the smallest
1526 * objsize such that the wastage is always acceptable for objects
1527 * with that objsize or smaller. Since a smaller objsize always
1528 * generates a larger possible uma_max_ipers, we use this computed
1529 * objsize to calculate the largest ipers possible. Since the
1530 * ipers calculated for OFFPAGE slab headers is always larger than
1531 * the ipers initially calculated in zone_small_init(), we use
1532 * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to
1533 * obtain the maximum ipers possible for offpage slab headers.
1534 *
1535 * It should be noted that ipers versus objsize is an inversly
1536 * proportional function which drops off rather quickly so as
1537 * long as our UMA_MAX_WASTE is such that the objsize we calculate
1538 * falls into the portion of the inverse relation AFTER the steep
1539 * falloff, then uma_max_ipers shouldn't be too high (~10 on i386).
1540 *
1541 * Note that we have 8-bits (1 byte) to use as a freelist index
1542 * inside the actual slab header itself and this is enough to
1543 * accomodate us. In the worst case, a UMA_SMALLEST_UNIT sized
1544 * object with offpage slab header would have ipers =
1545 * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is
1546 * 1 greater than what our byte-integer freelist index can
1547 * accomodate, but we know that this situation never occurs as
1548 * for UMA_SMALLEST_UNIT-sized objects, we will never calculate
1549 * that we need to go to offpage slab headers. Or, if we do,
1550 * then we trap that condition below and panic in the INVARIANTS case.
1551 */
1552 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - UMA_MAX_WASTE;
1553 totsize = wsize;
1554 objsize = UMA_SMALLEST_UNIT;
1555 while (totsize >= wsize) {
1556 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) /
1557 (objsize + UMA_FRITM_SZ);
1558 totsize *= (UMA_FRITM_SZ + objsize);
1559 objsize++;
1560 }
1561 if (objsize > UMA_SMALLEST_UNIT)
1562 objsize--;
1563 uma_max_ipers = UMA_SLAB_SIZE / objsize;
1564
1565 wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - UMA_MAX_WASTE;
1566 totsize = wsize;
1567 objsize = UMA_SMALLEST_UNIT;
1568 while (totsize >= wsize) {
1569 totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) /
1570 (objsize + UMA_FRITMREF_SZ);
1571 totsize *= (UMA_FRITMREF_SZ + objsize);
1572 objsize++;
1573 }
1574 if (objsize > UMA_SMALLEST_UNIT)
1575 objsize--;
1576 uma_max_ipers_ref = UMA_SLAB_SIZE / objsize;
1577
1578 KASSERT((uma_max_ipers_ref <= 255) && (uma_max_ipers <= 255),
1579 ("uma_startup: calculated uma_max_ipers values too large!"));
1580
1581#ifdef UMA_DEBUG
1582 printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers);
1583 printf("Calculated uma_max_ipers_slab (for OFFPAGE) is %d\n",
1584 uma_max_ipers_ref);
1585#endif
1586
1587 /* "manually" create the initial zone */
1588 args.name = "UMA Kegs";
1589 args.size = sizeof(struct uma_keg);
1590 args.ctor = keg_ctor;
1591 args.dtor = keg_dtor;
1592 args.uminit = zero_init;
1593 args.fini = NULL;
1594 args.keg = &masterkeg;
1595 args.align = 32 - 1;
1596 args.flags = UMA_ZFLAG_INTERNAL;
1597 /* The initial zone has no Per cpu queues so it's smaller */
1598 zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK);
1599
1600#ifdef UMA_DEBUG
1601 printf("Filling boot free list.\n");
1602#endif
1603 for (i = 0; i < boot_pages; i++) {
1604 slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE));
1605 slab->us_data = (u_int8_t *)slab;
1606 slab->us_flags = UMA_SLAB_BOOT;
1607 LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link);
1608 }
1609 mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF);
1610
1611#ifdef UMA_DEBUG
1612 printf("Creating uma zone headers zone and keg.\n");
1613#endif
1614 args.name = "UMA Zones";
1615 args.size = sizeof(struct uma_zone) +
1616 (sizeof(struct uma_cache) * (mp_maxid + 1));
1617 args.ctor = zone_ctor;
1618 args.dtor = zone_dtor;
1619 args.uminit = zero_init;
1620 args.fini = NULL;
1621 args.keg = NULL;
1622 args.align = 32 - 1;
1623 args.flags = UMA_ZFLAG_INTERNAL;
1624 /* The initial zone has no Per cpu queues so it's smaller */
1625 zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK);
1626
1627#ifdef UMA_DEBUG
1628 printf("Initializing pcpu cache locks.\n");
1629#endif
1630#ifdef UMA_DEBUG
1631 printf("Creating slab and hash zones.\n");
1632#endif
1633
1634 /*
1635 * This is the max number of free list items we'll have with
1636 * offpage slabs.
1637 */
1638 slabsize = uma_max_ipers * UMA_FRITM_SZ;
1639 slabsize += sizeof(struct uma_slab);
1640
1641 /* Now make a zone for slab headers */
1642 slabzone = uma_zcreate("UMA Slabs",
1643 slabsize,
1644 NULL, NULL, NULL, NULL,
1645 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1646
1647 /*
1648 * We also create a zone for the bigger slabs with reference
1649 * counts in them, to accomodate UMA_ZONE_REFCNT zones.
1650 */
1651 slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ;
1652 slabsize += sizeof(struct uma_slab_refcnt);
1653 slabrefzone = uma_zcreate("UMA RCntSlabs",
1654 slabsize,
1655 NULL, NULL, NULL, NULL,
1656 UMA_ALIGN_PTR,
1657 UMA_ZFLAG_INTERNAL);
1658
1659 hashzone = uma_zcreate("UMA Hash",
1660 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
1661 NULL, NULL, NULL, NULL,
1662 UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
1663
1664 bucket_init();
1665
1666#ifdef UMA_MD_SMALL_ALLOC
1667 booted = 1;
1668#endif
1669
1670#ifdef UMA_DEBUG
1671 printf("UMA startup complete.\n");
1672#endif
1673}
1674
1675/* see uma.h */
1676void
1677uma_startup2(void)
1678{
1679 booted = 1;
1680 bucket_enable();
1681#ifdef UMA_DEBUG
1682 printf("UMA startup2 complete.\n");
1683#endif
1684}
1685
1686/*
1687 * Initialize our callout handle
1688 *
1689 */
1690
1691static void
1692uma_startup3(void)
1693{
1694#ifdef UMA_DEBUG
1695 printf("Starting callout.\n");
1696#endif
1697 callout_init(&uma_callout, CALLOUT_MPSAFE);
1698 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1699#ifdef UMA_DEBUG
1700 printf("UMA startup3 complete.\n");
1701#endif
1702}
1703
1704static uma_zone_t
1705uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
1706 int align, u_int32_t flags)
1707{
1708 struct uma_kctor_args args;
1709
1710 args.size = size;
1711 args.uminit = uminit;
1712 args.fini = fini;
|
1729uma_zone_t
1730uma_zcreate(char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
1731 uma_init uminit, uma_fini fini, int align, u_int32_t flags)
1732
1733{
1734 struct uma_zctor_args args;
1735
1736 /* This stuff is essential for the zone ctor */
1737 args.name = name;
1738 args.size = size;
1739 args.ctor = ctor;
1740 args.dtor = dtor;
1741 args.uminit = uminit;
1742 args.fini = fini;
1743 args.align = align;
1744 args.flags = flags;
1745 args.keg = NULL;
1746
1747 return (uma_zalloc_internal(zones, &args, M_WAITOK));
1748}
1749
1750/* See uma.h */
1751uma_zone_t
1752uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor,
1753 uma_init zinit, uma_fini zfini, uma_zone_t master)
1754{
1755 struct uma_zctor_args args;
1756
1757 args.name = name;
1758 args.size = master->uz_keg->uk_size;
1759 args.ctor = ctor;
1760 args.dtor = dtor;
1761 args.uminit = zinit;
1762 args.fini = zfini;
1763 args.align = master->uz_keg->uk_align;
1764 args.flags = master->uz_keg->uk_flags | UMA_ZONE_SECONDARY;
1765 args.keg = master->uz_keg;
1766
1767 return (uma_zalloc_internal(zones, &args, M_WAITOK));
1768}
1769
1770/* See uma.h */
1771void
1772uma_zdestroy(uma_zone_t zone)
1773{
1774
1775 uma_zfree_internal(zones, zone, NULL, SKIP_NONE, ZFREE_STATFREE);
1776}
1777
1778/* See uma.h */
1779void *
1780uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
1781{
1782 void *item;
1783 uma_cache_t cache;
1784 uma_bucket_t bucket;
1785 int cpu;
1786
1787 /* This is the fast path allocation */
1788#ifdef UMA_DEBUG_ALLOC_1
1789 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone);
1790#endif
1791 CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread,
1792 zone->uz_name, flags);
1793
1794 if (flags & M_WAITOK) {
1795 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
1796 "uma_zalloc_arg: zone \"%s\"", zone->uz_name);
1797 }
1798
1799 /*
1800 * If possible, allocate from the per-CPU cache. There are two
1801 * requirements for safe access to the per-CPU cache: (1) the thread
1802 * accessing the cache must not be preempted or yield during access,
1803 * and (2) the thread must not migrate CPUs without switching which
1804 * cache it accesses. We rely on a critical section to prevent
1805 * preemption and migration. We release the critical section in
1806 * order to acquire the zone mutex if we are unable to allocate from
1807 * the current cache; when we re-acquire the critical section, we
1808 * must detect and handle migration if it has occurred.
1809 */
1810zalloc_restart:
1811 critical_enter();
1812 cpu = curcpu;
1813 cache = &zone->uz_cpu[cpu];
1814
1815zalloc_start:
1816 bucket = cache->uc_allocbucket;
1817
1818 if (bucket) {
1819 if (bucket->ub_cnt > 0) {
1820 bucket->ub_cnt--;
1821 item = bucket->ub_bucket[bucket->ub_cnt];
1822#ifdef INVARIANTS
1823 bucket->ub_bucket[bucket->ub_cnt] = NULL;
1824#endif
1825 KASSERT(item != NULL,
1826 ("uma_zalloc: Bucket pointer mangled."));
1827 cache->uc_allocs++;
1828 critical_exit();
1829#ifdef INVARIANTS
1830 ZONE_LOCK(zone);
1831 uma_dbg_alloc(zone, NULL, item);
1832 ZONE_UNLOCK(zone);
1833#endif
1834 if (zone->uz_ctor != NULL) {
1835 if (zone->uz_ctor(item, zone->uz_keg->uk_size,
1836 udata, flags) != 0) {
1837 uma_zfree_internal(zone, item, udata,
1838 SKIP_DTOR, ZFREE_STATFAIL |
1839 ZFREE_STATFREE);
1840 return (NULL);
1841 }
1842 }
1843 if (flags & M_ZERO)
1844 bzero(item, zone->uz_keg->uk_size);
1845 return (item);
1846 } else if (cache->uc_freebucket) {
1847 /*
1848 * We have run out of items in our allocbucket.
1849 * See if we can switch with our free bucket.
1850 */
1851 if (cache->uc_freebucket->ub_cnt > 0) {
1852#ifdef UMA_DEBUG_ALLOC
1853 printf("uma_zalloc: Swapping empty with"
1854 " alloc.\n");
1855#endif
1856 bucket = cache->uc_freebucket;
1857 cache->uc_freebucket = cache->uc_allocbucket;
1858 cache->uc_allocbucket = bucket;
1859
1860 goto zalloc_start;
1861 }
1862 }
1863 }
1864 /*
1865 * Attempt to retrieve the item from the per-CPU cache has failed, so
1866 * we must go back to the zone. This requires the zone lock, so we
1867 * must drop the critical section, then re-acquire it when we go back
1868 * to the cache. Since the critical section is released, we may be
1869 * preempted or migrate. As such, make sure not to maintain any
1870 * thread-local state specific to the cache from prior to releasing
1871 * the critical section.
1872 */
1873 critical_exit();
1874 ZONE_LOCK(zone);
1875 critical_enter();
1876 cpu = curcpu;
1877 cache = &zone->uz_cpu[cpu];
1878 bucket = cache->uc_allocbucket;
1879 if (bucket != NULL) {
1880 if (bucket->ub_cnt > 0) {
1881 ZONE_UNLOCK(zone);
1882 goto zalloc_start;
1883 }
1884 bucket = cache->uc_freebucket;
1885 if (bucket != NULL && bucket->ub_cnt > 0) {
1886 ZONE_UNLOCK(zone);
1887 goto zalloc_start;
1888 }
1889 }
1890
1891 /* Since we have locked the zone we may as well send back our stats */
1892 zone->uz_allocs += cache->uc_allocs;
1893 cache->uc_allocs = 0;
1894 zone->uz_frees += cache->uc_frees;
1895 cache->uc_frees = 0;
1896
1897 /* Our old one is now a free bucket */
1898 if (cache->uc_allocbucket) {
1899 KASSERT(cache->uc_allocbucket->ub_cnt == 0,
1900 ("uma_zalloc_arg: Freeing a non free bucket."));
1901 LIST_INSERT_HEAD(&zone->uz_free_bucket,
1902 cache->uc_allocbucket, ub_link);
1903 cache->uc_allocbucket = NULL;
1904 }
1905
1906 /* Check the free list for a new alloc bucket */
1907 if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) {
1908 KASSERT(bucket->ub_cnt != 0,
1909 ("uma_zalloc_arg: Returning an empty bucket."));
1910
1911 LIST_REMOVE(bucket, ub_link);
1912 cache->uc_allocbucket = bucket;
1913 ZONE_UNLOCK(zone);
1914 goto zalloc_start;
1915 }
1916 /* We are no longer associated with this CPU. */
1917 critical_exit();
1918
1919 /* Bump up our uz_count so we get here less */
1920 if (zone->uz_count < BUCKET_MAX)
1921 zone->uz_count++;
1922
1923 /*
1924 * Now lets just fill a bucket and put it on the free list. If that
1925 * works we'll restart the allocation from the begining.
1926 */
1927 if (uma_zalloc_bucket(zone, flags)) {
1928 ZONE_UNLOCK(zone);
1929 goto zalloc_restart;
1930 }
1931 ZONE_UNLOCK(zone);
1932 /*
1933 * We may not be able to get a bucket so return an actual item.
1934 */
1935#ifdef UMA_DEBUG
1936 printf("uma_zalloc_arg: Bucketzone returned NULL\n");
1937#endif
1938
1939 return (uma_zalloc_internal(zone, udata, flags));
1940}
1941
1942static uma_slab_t
1943uma_zone_slab(uma_zone_t zone, int flags)
1944{
1945 uma_slab_t slab;
1946 uma_keg_t keg;
1947
1948 keg = zone->uz_keg;
1949
1950 /*
1951 * This is to prevent us from recursively trying to allocate
1952 * buckets. The problem is that if an allocation forces us to
1953 * grab a new bucket we will call page_alloc, which will go off
1954 * and cause the vm to allocate vm_map_entries. If we need new
1955 * buckets there too we will recurse in kmem_alloc and bad
1956 * things happen. So instead we return a NULL bucket, and make
1957 * the code that allocates buckets smart enough to deal with it
1958 *
1959 * XXX: While we want this protection for the bucket zones so that
1960 * recursion from the VM is handled (and the calling code that
1961 * allocates buckets knows how to deal with it), we do not want
1962 * to prevent allocation from the slab header zones (slabzone
1963 * and slabrefzone) if uk_recurse is not zero for them. The
1964 * reason is that it could lead to NULL being returned for
1965 * slab header allocations even in the M_WAITOK case, and the
1966 * caller can't handle that.
1967 */
1968 if (keg->uk_flags & UMA_ZFLAG_INTERNAL && keg->uk_recurse != 0)
1969 if ((zone != slabzone) && (zone != slabrefzone))
1970 return (NULL);
1971
1972 slab = NULL;
1973
1974 for (;;) {
1975 /*
1976 * Find a slab with some space. Prefer slabs that are partially
1977 * used over those that are totally full. This helps to reduce
1978 * fragmentation.
1979 */
1980 if (keg->uk_free != 0) {
1981 if (!LIST_EMPTY(&keg->uk_part_slab)) {
1982 slab = LIST_FIRST(&keg->uk_part_slab);
1983 } else {
1984 slab = LIST_FIRST(&keg->uk_free_slab);
1985 LIST_REMOVE(slab, us_link);
1986 LIST_INSERT_HEAD(&keg->uk_part_slab, slab,
1987 us_link);
1988 }
1989 return (slab);
1990 }
1991
1992 /*
1993 * M_NOVM means don't ask at all!
1994 */
1995 if (flags & M_NOVM)
1996 break;
1997
1998 if (keg->uk_maxpages &&
1999 keg->uk_pages >= keg->uk_maxpages) {
2000 keg->uk_flags |= UMA_ZFLAG_FULL;
2001
2002 if (flags & M_NOWAIT)
2003 break;
2004 else
2005 msleep(keg, &keg->uk_lock, PVM,
2006 "zonelimit", 0);
2007 continue;
2008 }
2009 keg->uk_recurse++;
2010 slab = slab_zalloc(zone, flags);
2011 keg->uk_recurse--;
2012
2013 /*
2014 * If we got a slab here it's safe to mark it partially used
2015 * and return. We assume that the caller is going to remove
2016 * at least one item.
2017 */
2018 if (slab) {
2019 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2020 return (slab);
2021 }
2022 /*
2023 * We might not have been able to get a slab but another cpu
2024 * could have while we were unlocked. Check again before we
2025 * fail.
2026 */
2027 if (flags & M_NOWAIT)
2028 flags |= M_NOVM;
2029 }
2030 return (slab);
2031}
2032
2033static void *
2034uma_slab_alloc(uma_zone_t zone, uma_slab_t slab)
2035{
2036 uma_keg_t keg;
2037 uma_slabrefcnt_t slabref;
2038 void *item;
2039 u_int8_t freei;
2040
2041 keg = zone->uz_keg;
2042
2043 freei = slab->us_firstfree;
2044 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2045 slabref = (uma_slabrefcnt_t)slab;
2046 slab->us_firstfree = slabref->us_freelist[freei].us_item;
2047 } else {
2048 slab->us_firstfree = slab->us_freelist[freei].us_item;
2049 }
2050 item = slab->us_data + (keg->uk_rsize * freei);
2051
2052 slab->us_freecount--;
2053 keg->uk_free--;
2054#ifdef INVARIANTS
2055 uma_dbg_alloc(zone, slab, item);
2056#endif
2057 /* Move this slab to the full list */
2058 if (slab->us_freecount == 0) {
2059 LIST_REMOVE(slab, us_link);
2060 LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link);
2061 }
2062
2063 return (item);
2064}
2065
2066static int
2067uma_zalloc_bucket(uma_zone_t zone, int flags)
2068{
2069 uma_bucket_t bucket;
2070 uma_slab_t slab;
2071 int16_t saved;
2072 int max, origflags = flags;
2073
2074 /*
2075 * Try this zone's free list first so we don't allocate extra buckets.
2076 */
2077 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2078 KASSERT(bucket->ub_cnt == 0,
2079 ("uma_zalloc_bucket: Bucket on free list is not empty."));
2080 LIST_REMOVE(bucket, ub_link);
2081 } else {
2082 int bflags;
2083
2084 bflags = (flags & ~M_ZERO);
2085 if (zone->uz_keg->uk_flags & UMA_ZFLAG_CACHEONLY)
2086 bflags |= M_NOVM;
2087
2088 ZONE_UNLOCK(zone);
2089 bucket = bucket_alloc(zone->uz_count, bflags);
2090 ZONE_LOCK(zone);
2091 }
2092
2093 if (bucket == NULL)
2094 return (0);
2095
2096#ifdef SMP
2097 /*
2098 * This code is here to limit the number of simultaneous bucket fills
2099 * for any given zone to the number of per cpu caches in this zone. This
2100 * is done so that we don't allocate more memory than we really need.
2101 */
2102 if (zone->uz_fills >= mp_ncpus)
2103 goto done;
2104
2105#endif
2106 zone->uz_fills++;
2107
2108 max = MIN(bucket->ub_entries, zone->uz_count);
2109 /* Try to keep the buckets totally full */
2110 saved = bucket->ub_cnt;
2111 while (bucket->ub_cnt < max &&
2112 (slab = uma_zone_slab(zone, flags)) != NULL) {
2113 while (slab->us_freecount && bucket->ub_cnt < max) {
2114 bucket->ub_bucket[bucket->ub_cnt++] =
2115 uma_slab_alloc(zone, slab);
2116 }
2117
2118 /* Don't block on the next fill */
2119 flags |= M_NOWAIT;
2120 }
2121
2122 /*
2123 * We unlock here because we need to call the zone's init.
2124 * It should be safe to unlock because the slab dealt with
2125 * above is already on the appropriate list within the keg
2126 * and the bucket we filled is not yet on any list, so we
2127 * own it.
2128 */
2129 if (zone->uz_init != NULL) {
2130 int i;
2131
2132 ZONE_UNLOCK(zone);
2133 for (i = saved; i < bucket->ub_cnt; i++)
2134 if (zone->uz_init(bucket->ub_bucket[i],
2135 zone->uz_keg->uk_size, origflags) != 0)
2136 break;
2137 /*
2138 * If we couldn't initialize the whole bucket, put the
2139 * rest back onto the freelist.
2140 */
2141 if (i != bucket->ub_cnt) {
2142 int j;
2143
2144 for (j = i; j < bucket->ub_cnt; j++) {
2145 uma_zfree_internal(zone, bucket->ub_bucket[j],
2146 NULL, SKIP_FINI, 0);
2147#ifdef INVARIANTS
2148 bucket->ub_bucket[j] = NULL;
2149#endif
2150 }
2151 bucket->ub_cnt = i;
2152 }
2153 ZONE_LOCK(zone);
2154 }
2155
2156 zone->uz_fills--;
2157 if (bucket->ub_cnt != 0) {
2158 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2159 bucket, ub_link);
2160 return (1);
2161 }
2162#ifdef SMP
2163done:
2164#endif
2165 bucket_free(bucket);
2166
2167 return (0);
2168}
2169/*
2170 * Allocates an item for an internal zone
2171 *
2172 * Arguments
2173 * zone The zone to alloc for.
2174 * udata The data to be passed to the constructor.
2175 * flags M_WAITOK, M_NOWAIT, M_ZERO.
2176 *
2177 * Returns
2178 * NULL if there is no memory and M_NOWAIT is set
2179 * An item if successful
2180 */
2181
2182static void *
2183uma_zalloc_internal(uma_zone_t zone, void *udata, int flags)
2184{
2185 uma_keg_t keg;
2186 uma_slab_t slab;
2187 void *item;
2188
2189 item = NULL;
2190 keg = zone->uz_keg;
2191
2192#ifdef UMA_DEBUG_ALLOC
2193 printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone);
2194#endif
2195 ZONE_LOCK(zone);
2196
2197 slab = uma_zone_slab(zone, flags);
2198 if (slab == NULL) {
2199 zone->uz_fails++;
2200 ZONE_UNLOCK(zone);
2201 return (NULL);
2202 }
2203
2204 item = uma_slab_alloc(zone, slab);
2205
2206 zone->uz_allocs++;
2207
2208 ZONE_UNLOCK(zone);
2209
2210 /*
2211 * We have to call both the zone's init (not the keg's init)
2212 * and the zone's ctor. This is because the item is going from
2213 * a keg slab directly to the user, and the user is expecting it
2214 * to be both zone-init'd as well as zone-ctor'd.
2215 */
2216 if (zone->uz_init != NULL) {
2217 if (zone->uz_init(item, keg->uk_size, flags) != 0) {
2218 uma_zfree_internal(zone, item, udata, SKIP_FINI,
2219 ZFREE_STATFAIL | ZFREE_STATFREE);
2220 return (NULL);
2221 }
2222 }
2223 if (zone->uz_ctor != NULL) {
2224 if (zone->uz_ctor(item, keg->uk_size, udata, flags) != 0) {
2225 uma_zfree_internal(zone, item, udata, SKIP_DTOR,
2226 ZFREE_STATFAIL | ZFREE_STATFREE);
2227 return (NULL);
2228 }
2229 }
2230 if (flags & M_ZERO)
2231 bzero(item, keg->uk_size);
2232
2233 return (item);
2234}
2235
2236/* See uma.h */
2237void
2238uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
2239{
2240 uma_keg_t keg;
2241 uma_cache_t cache;
2242 uma_bucket_t bucket;
2243 int bflags;
2244 int cpu;
2245
2246 keg = zone->uz_keg;
2247
2248#ifdef UMA_DEBUG_ALLOC_1
2249 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone);
2250#endif
2251 CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread,
2252 zone->uz_name);
2253
2254 if (zone->uz_dtor)
2255 zone->uz_dtor(item, keg->uk_size, udata);
2256#ifdef INVARIANTS
2257 ZONE_LOCK(zone);
2258 if (keg->uk_flags & UMA_ZONE_MALLOC)
2259 uma_dbg_free(zone, udata, item);
2260 else
2261 uma_dbg_free(zone, NULL, item);
2262 ZONE_UNLOCK(zone);
2263#endif
2264 /*
2265 * The race here is acceptable. If we miss it we'll just have to wait
2266 * a little longer for the limits to be reset.
2267 */
2268 if (keg->uk_flags & UMA_ZFLAG_FULL)
2269 goto zfree_internal;
2270
2271 /*
2272 * If possible, free to the per-CPU cache. There are two
2273 * requirements for safe access to the per-CPU cache: (1) the thread
2274 * accessing the cache must not be preempted or yield during access,
2275 * and (2) the thread must not migrate CPUs without switching which
2276 * cache it accesses. We rely on a critical section to prevent
2277 * preemption and migration. We release the critical section in
2278 * order to acquire the zone mutex if we are unable to free to the
2279 * current cache; when we re-acquire the critical section, we must
2280 * detect and handle migration if it has occurred.
2281 */
2282zfree_restart:
2283 critical_enter();
2284 cpu = curcpu;
2285 cache = &zone->uz_cpu[cpu];
2286
2287zfree_start:
2288 bucket = cache->uc_freebucket;
2289
2290 if (bucket) {
2291 /*
2292 * Do we have room in our bucket? It is OK for this uz count
2293 * check to be slightly out of sync.
2294 */
2295
2296 if (bucket->ub_cnt < bucket->ub_entries) {
2297 KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL,
2298 ("uma_zfree: Freeing to non free bucket index."));
2299 bucket->ub_bucket[bucket->ub_cnt] = item;
2300 bucket->ub_cnt++;
2301 cache->uc_frees++;
2302 critical_exit();
2303 return;
2304 } else if (cache->uc_allocbucket) {
2305#ifdef UMA_DEBUG_ALLOC
2306 printf("uma_zfree: Swapping buckets.\n");
2307#endif
2308 /*
2309 * We have run out of space in our freebucket.
2310 * See if we can switch with our alloc bucket.
2311 */
2312 if (cache->uc_allocbucket->ub_cnt <
2313 cache->uc_freebucket->ub_cnt) {
2314 bucket = cache->uc_freebucket;
2315 cache->uc_freebucket = cache->uc_allocbucket;
2316 cache->uc_allocbucket = bucket;
2317 goto zfree_start;
2318 }
2319 }
2320 }
2321 /*
2322 * We can get here for two reasons:
2323 *
2324 * 1) The buckets are NULL
2325 * 2) The alloc and free buckets are both somewhat full.
2326 *
2327 * We must go back the zone, which requires acquiring the zone lock,
2328 * which in turn means we must release and re-acquire the critical
2329 * section. Since the critical section is released, we may be
2330 * preempted or migrate. As such, make sure not to maintain any
2331 * thread-local state specific to the cache from prior to releasing
2332 * the critical section.
2333 */
2334 critical_exit();
2335 ZONE_LOCK(zone);
2336 critical_enter();
2337 cpu = curcpu;
2338 cache = &zone->uz_cpu[cpu];
2339 if (cache->uc_freebucket != NULL) {
2340 if (cache->uc_freebucket->ub_cnt <
2341 cache->uc_freebucket->ub_entries) {
2342 ZONE_UNLOCK(zone);
2343 goto zfree_start;
2344 }
2345 if (cache->uc_allocbucket != NULL &&
2346 (cache->uc_allocbucket->ub_cnt <
2347 cache->uc_freebucket->ub_cnt)) {
2348 ZONE_UNLOCK(zone);
2349 goto zfree_start;
2350 }
2351 }
2352
2353 /* Since we have locked the zone we may as well send back our stats */
2354 zone->uz_allocs += cache->uc_allocs;
2355 cache->uc_allocs = 0;
2356 zone->uz_frees += cache->uc_frees;
2357 cache->uc_frees = 0;
2358
2359 bucket = cache->uc_freebucket;
2360 cache->uc_freebucket = NULL;
2361
2362 /* Can we throw this on the zone full list? */
2363 if (bucket != NULL) {
2364#ifdef UMA_DEBUG_ALLOC
2365 printf("uma_zfree: Putting old bucket on the free list.\n");
2366#endif
2367 /* ub_cnt is pointing to the last free item */
2368 KASSERT(bucket->ub_cnt != 0,
2369 ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n"));
2370 LIST_INSERT_HEAD(&zone->uz_full_bucket,
2371 bucket, ub_link);
2372 }
2373 if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) {
2374 LIST_REMOVE(bucket, ub_link);
2375 ZONE_UNLOCK(zone);
2376 cache->uc_freebucket = bucket;
2377 goto zfree_start;
2378 }
2379 /* We are no longer associated with this CPU. */
2380 critical_exit();
2381
2382 /* And the zone.. */
2383 ZONE_UNLOCK(zone);
2384
2385#ifdef UMA_DEBUG_ALLOC
2386 printf("uma_zfree: Allocating new free bucket.\n");
2387#endif
2388 bflags = M_NOWAIT;
2389
2390 if (keg->uk_flags & UMA_ZFLAG_CACHEONLY)
2391 bflags |= M_NOVM;
2392 bucket = bucket_alloc(zone->uz_count, bflags);
2393 if (bucket) {
2394 ZONE_LOCK(zone);
2395 LIST_INSERT_HEAD(&zone->uz_free_bucket,
2396 bucket, ub_link);
2397 ZONE_UNLOCK(zone);
2398 goto zfree_restart;
2399 }
2400
2401 /*
2402 * If nothing else caught this, we'll just do an internal free.
2403 */
2404zfree_internal:
2405 uma_zfree_internal(zone, item, udata, SKIP_DTOR, ZFREE_STATFREE);
2406
2407 return;
2408}
2409
2410/*
2411 * Frees an item to an INTERNAL zone or allocates a free bucket
2412 *
2413 * Arguments:
2414 * zone The zone to free to
2415 * item The item we're freeing
2416 * udata User supplied data for the dtor
2417 * skip Skip dtors and finis
2418 */
2419static void
2420uma_zfree_internal(uma_zone_t zone, void *item, void *udata,
2421 enum zfreeskip skip, int flags)
2422{
2423 uma_slab_t slab;
2424 uma_slabrefcnt_t slabref;
2425 uma_keg_t keg;
2426 u_int8_t *mem;
2427 u_int8_t freei;
2428
2429 keg = zone->uz_keg;
2430
2431 if (skip < SKIP_DTOR && zone->uz_dtor)
2432 zone->uz_dtor(item, keg->uk_size, udata);
2433 if (skip < SKIP_FINI && zone->uz_fini)
2434 zone->uz_fini(item, keg->uk_size);
2435
2436 ZONE_LOCK(zone);
2437
2438 if (flags & ZFREE_STATFAIL)
2439 zone->uz_fails++;
2440 if (flags & ZFREE_STATFREE)
2441 zone->uz_frees++;
2442
2443 if (!(keg->uk_flags & UMA_ZONE_MALLOC)) {
2444 mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK));
2445 if (keg->uk_flags & UMA_ZONE_HASH)
2446 slab = hash_sfind(&keg->uk_hash, mem);
2447 else {
2448 mem += keg->uk_pgoff;
2449 slab = (uma_slab_t)mem;
2450 }
2451 } else {
2452 slab = (uma_slab_t)udata;
2453 }
2454
2455 /* Do we need to remove from any lists? */
2456 if (slab->us_freecount+1 == keg->uk_ipers) {
2457 LIST_REMOVE(slab, us_link);
2458 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2459 } else if (slab->us_freecount == 0) {
2460 LIST_REMOVE(slab, us_link);
2461 LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link);
2462 }
2463
2464 /* Slab management stuff */
2465 freei = ((unsigned long)item - (unsigned long)slab->us_data)
2466 / keg->uk_rsize;
2467
2468#ifdef INVARIANTS
2469 if (!skip)
2470 uma_dbg_free(zone, slab, item);
2471#endif
2472
2473 if (keg->uk_flags & UMA_ZONE_REFCNT) {
2474 slabref = (uma_slabrefcnt_t)slab;
2475 slabref->us_freelist[freei].us_item = slab->us_firstfree;
2476 } else {
2477 slab->us_freelist[freei].us_item = slab->us_firstfree;
2478 }
2479 slab->us_firstfree = freei;
2480 slab->us_freecount++;
2481
2482 /* Zone statistics */
2483 keg->uk_free++;
2484
2485 if (keg->uk_flags & UMA_ZFLAG_FULL) {
2486 if (keg->uk_pages < keg->uk_maxpages)
2487 keg->uk_flags &= ~UMA_ZFLAG_FULL;
2488
2489 /*
2490 * We can handle one more allocation. Since we're clearing ZFLAG_FULL,
2491 * wake up all procs blocked on pages. This should be uncommon, so
2492 * keeping this simple for now (rather than adding count of blocked
2493 * threads etc).
2494 */
2495 wakeup(keg);
2496 }
2497
2498 ZONE_UNLOCK(zone);
2499}
2500
2501/* See uma.h */
2502void
2503uma_zone_set_max(uma_zone_t zone, int nitems)
2504{
2505 uma_keg_t keg;
2506
2507 keg = zone->uz_keg;
2508 ZONE_LOCK(zone);
2509 if (keg->uk_ppera > 1)
2510 keg->uk_maxpages = nitems * keg->uk_ppera;
2511 else
2512 keg->uk_maxpages = nitems / keg->uk_ipers;
2513
2514 if (keg->uk_maxpages * keg->uk_ipers < nitems)
2515 keg->uk_maxpages++;
2516
2517 ZONE_UNLOCK(zone);
2518}
2519
2520/* See uma.h */
2521void
2522uma_zone_set_init(uma_zone_t zone, uma_init uminit)
2523{
2524 ZONE_LOCK(zone);
2525 KASSERT(zone->uz_keg->uk_pages == 0,
2526 ("uma_zone_set_init on non-empty keg"));
2527 zone->uz_keg->uk_init = uminit;
2528 ZONE_UNLOCK(zone);
2529}
2530
2531/* See uma.h */
2532void
2533uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
2534{
2535 ZONE_LOCK(zone);
2536 KASSERT(zone->uz_keg->uk_pages == 0,
2537 ("uma_zone_set_fini on non-empty keg"));
2538 zone->uz_keg->uk_fini = fini;
2539 ZONE_UNLOCK(zone);
2540}
2541
2542/* See uma.h */
2543void
2544uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
2545{
2546 ZONE_LOCK(zone);
2547 KASSERT(zone->uz_keg->uk_pages == 0,
2548 ("uma_zone_set_zinit on non-empty keg"));
2549 zone->uz_init = zinit;
2550 ZONE_UNLOCK(zone);
2551}
2552
2553/* See uma.h */
2554void
2555uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
2556{
2557 ZONE_LOCK(zone);
2558 KASSERT(zone->uz_keg->uk_pages == 0,
2559 ("uma_zone_set_zfini on non-empty keg"));
2560 zone->uz_fini = zfini;
2561 ZONE_UNLOCK(zone);
2562}
2563
2564/* See uma.h */
2565/* XXX uk_freef is not actually used with the zone locked */
2566void
2567uma_zone_set_freef(uma_zone_t zone, uma_free freef)
2568{
2569 ZONE_LOCK(zone);
2570 zone->uz_keg->uk_freef = freef;
2571 ZONE_UNLOCK(zone);
2572}
2573
2574/* See uma.h */
2575/* XXX uk_allocf is not actually used with the zone locked */
2576void
2577uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
2578{
2579 ZONE_LOCK(zone);
2580 zone->uz_keg->uk_flags |= UMA_ZFLAG_PRIVALLOC;
2581 zone->uz_keg->uk_allocf = allocf;
2582 ZONE_UNLOCK(zone);
2583}
2584
2585/* See uma.h */
2586int
2587uma_zone_set_obj(uma_zone_t zone, struct vm_object *obj, int count)
2588{
2589 uma_keg_t keg;
2590 vm_offset_t kva;
2591 int pages;
2592
2593 keg = zone->uz_keg;
2594 pages = count / keg->uk_ipers;
2595
2596 if (pages * keg->uk_ipers < count)
2597 pages++;
2598
2599 kva = kmem_alloc_nofault(kernel_map, pages * UMA_SLAB_SIZE);
2600
2601 if (kva == 0)
2602 return (0);
2603 if (obj == NULL) {
2604 obj = vm_object_allocate(OBJT_DEFAULT,
2605 pages);
2606 } else {
2607 VM_OBJECT_LOCK_INIT(obj, "uma object");
2608 _vm_object_allocate(OBJT_DEFAULT,
2609 pages, obj);
2610 }
2611 ZONE_LOCK(zone);
2612 keg->uk_kva = kva;
2613 keg->uk_obj = obj;
2614 keg->uk_maxpages = pages;
2615 keg->uk_allocf = obj_alloc;
2616 keg->uk_flags |= UMA_ZONE_NOFREE | UMA_ZFLAG_PRIVALLOC;
2617 ZONE_UNLOCK(zone);
2618 return (1);
2619}
2620
2621/* See uma.h */
2622void
2623uma_prealloc(uma_zone_t zone, int items)
2624{
2625 int slabs;
2626 uma_slab_t slab;
2627 uma_keg_t keg;
2628
2629 keg = zone->uz_keg;
2630 ZONE_LOCK(zone);
2631 slabs = items / keg->uk_ipers;
2632 if (slabs * keg->uk_ipers < items)
2633 slabs++;
2634 while (slabs > 0) {
2635 slab = slab_zalloc(zone, M_WAITOK);
2636 LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link);
2637 slabs--;
2638 }
2639 ZONE_UNLOCK(zone);
2640}
2641
2642/* See uma.h */
2643u_int32_t *
2644uma_find_refcnt(uma_zone_t zone, void *item)
2645{
2646 uma_slabrefcnt_t slabref;
2647 uma_keg_t keg;
2648 u_int32_t *refcnt;
2649 int idx;
2650
2651 keg = zone->uz_keg;
2652 slabref = (uma_slabrefcnt_t)vtoslab((vm_offset_t)item &
2653 (~UMA_SLAB_MASK));
2654 KASSERT(slabref != NULL && slabref->us_keg->uk_flags & UMA_ZONE_REFCNT,
2655 ("uma_find_refcnt(): zone possibly not UMA_ZONE_REFCNT"));
2656 idx = ((unsigned long)item - (unsigned long)slabref->us_data)
2657 / keg->uk_rsize;
2658 refcnt = &slabref->us_freelist[idx].us_refcnt;
2659 return refcnt;
2660}
2661
2662/* See uma.h */
2663void
2664uma_reclaim(void)
2665{
2666#ifdef UMA_DEBUG
2667 printf("UMA: vm asked us to release pages!\n");
2668#endif
2669 bucket_enable();
2670 zone_foreach(zone_drain);
2671 /*
2672 * Some slabs may have been freed but this zone will be visited early
2673 * we visit again so that we can free pages that are empty once other
2674 * zones are drained. We have to do the same for buckets.
2675 */
2676 zone_drain(slabzone);
2677 zone_drain(slabrefzone);
2678 bucket_zone_drain();
2679}
2680
2681/* See uma.h */
2682int
2683uma_zone_exhausted(uma_zone_t zone)
2684{
2685 int full;
2686
2687 ZONE_LOCK(zone);
2688 full = (zone->uz_keg->uk_flags & UMA_ZFLAG_FULL);
2689 ZONE_UNLOCK(zone);
2690 return (full);
2691}
2692
2693int
2694uma_zone_exhausted_nolock(uma_zone_t zone)
2695{
2696 return (zone->uz_keg->uk_flags & UMA_ZFLAG_FULL);
2697}
2698
2699void *
2700uma_large_malloc(int size, int wait)
2701{
2702 void *mem;
2703 uma_slab_t slab;
2704 u_int8_t flags;
2705
2706 slab = uma_zalloc_internal(slabzone, NULL, wait);
2707 if (slab == NULL)
2708 return (NULL);
2709 mem = page_alloc(NULL, size, &flags, wait);
2710 if (mem) {
2711 vsetslab((vm_offset_t)mem, slab);
2712 slab->us_data = mem;
2713 slab->us_flags = flags | UMA_SLAB_MALLOC;
2714 slab->us_size = size;
2715 } else {
2716 uma_zfree_internal(slabzone, slab, NULL, SKIP_NONE,
2717 ZFREE_STATFAIL | ZFREE_STATFREE);
2718 }
2719
2720 return (mem);
2721}
2722
2723void
2724uma_large_free(uma_slab_t slab)
2725{
2726 vsetobj((vm_offset_t)slab->us_data, kmem_object);
2727 page_free(slab->us_data, slab->us_size, slab->us_flags);
2728 uma_zfree_internal(slabzone, slab, NULL, SKIP_NONE, ZFREE_STATFREE);
2729}
2730
2731void
2732uma_print_stats(void)
2733{
2734 zone_foreach(uma_print_zone);
2735}
2736
2737static void
2738slab_print(uma_slab_t slab)
2739{
2740 printf("slab: keg %p, data %p, freecount %d, firstfree %d\n",
2741 slab->us_keg, slab->us_data, slab->us_freecount,
2742 slab->us_firstfree);
2743}
2744
2745static void
2746cache_print(uma_cache_t cache)
2747{
2748 printf("alloc: %p(%d), free: %p(%d)\n",
2749 cache->uc_allocbucket,
2750 cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0,
2751 cache->uc_freebucket,
2752 cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0);
2753}
2754
2755void
2756uma_print_zone(uma_zone_t zone)
2757{
2758 uma_cache_t cache;
2759 uma_keg_t keg;
2760 uma_slab_t slab;
2761 int i;
2762
2763 keg = zone->uz_keg;
2764 printf("%s(%p) size %d(%d) flags %d ipers %d ppera %d out %d free %d\n",
2765 zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags,
2766 keg->uk_ipers, keg->uk_ppera,
2767 (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free);
2768 printf("Part slabs:\n");
2769 LIST_FOREACH(slab, &keg->uk_part_slab, us_link)
2770 slab_print(slab);
2771 printf("Free slabs:\n");
2772 LIST_FOREACH(slab, &keg->uk_free_slab, us_link)
2773 slab_print(slab);
2774 printf("Full slabs:\n");
2775 LIST_FOREACH(slab, &keg->uk_full_slab, us_link)
2776 slab_print(slab);
2777 for (i = 0; i <= mp_maxid; i++) {
2778 if (CPU_ABSENT(i))
2779 continue;
2780 cache = &zone->uz_cpu[i];
2781 printf("CPU %d Cache:\n", i);
2782 cache_print(cache);
2783 }
2784}
2785
2786#ifdef DDB
2787/*
2788 * Generate statistics across both the zone and its per-cpu cache's. Return
2789 * desired statistics if the pointer is non-NULL for that statistic.
2790 *
2791 * Note: does not update the zone statistics, as it can't safely clear the
2792 * per-CPU cache statistic.
2793 *
2794 * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't
2795 * safe from off-CPU; we should modify the caches to track this information
2796 * directly so that we don't have to.
2797 */
2798static void
2799uma_zone_sumstat(uma_zone_t z, int *cachefreep, u_int64_t *allocsp,
2800 u_int64_t *freesp)
2801{
2802 uma_cache_t cache;
2803 u_int64_t allocs, frees;
2804 int cachefree, cpu;
2805
2806 allocs = frees = 0;
2807 cachefree = 0;
2808 for (cpu = 0; cpu <= mp_maxid; cpu++) {
2809 if (CPU_ABSENT(cpu))
2810 continue;
2811 cache = &z->uz_cpu[cpu];
2812 if (cache->uc_allocbucket != NULL)
2813 cachefree += cache->uc_allocbucket->ub_cnt;
2814 if (cache->uc_freebucket != NULL)
2815 cachefree += cache->uc_freebucket->ub_cnt;
2816 allocs += cache->uc_allocs;
2817 frees += cache->uc_frees;
2818 }
2819 allocs += z->uz_allocs;
2820 frees += z->uz_frees;
2821 if (cachefreep != NULL)
2822 *cachefreep = cachefree;
2823 if (allocsp != NULL)
2824 *allocsp = allocs;
2825 if (freesp != NULL)
2826 *freesp = frees;
2827}
2828#endif /* DDB */
2829
2830static int
2831sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
2832{
2833 uma_keg_t kz;
2834 uma_zone_t z;
2835 int count;
2836
2837 count = 0;
2838 mtx_lock(&uma_mtx);
2839 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2840 LIST_FOREACH(z, &kz->uk_zones, uz_link)
2841 count++;
2842 }
2843 mtx_unlock(&uma_mtx);
2844 return (sysctl_handle_int(oidp, &count, 0, req));
2845}
2846
2847static int
2848sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
2849{
2850 struct uma_stream_header ush;
2851 struct uma_type_header uth;
2852 struct uma_percpu_stat ups;
2853 uma_bucket_t bucket;
2854 struct sbuf sbuf;
2855 uma_cache_t cache;
2856 uma_keg_t kz;
2857 uma_zone_t z;
2858 char *buffer;
2859 int buflen, count, error, i;
2860
2861 mtx_lock(&uma_mtx);
2862restart:
2863 mtx_assert(&uma_mtx, MA_OWNED);
2864 count = 0;
2865 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2866 LIST_FOREACH(z, &kz->uk_zones, uz_link)
2867 count++;
2868 }
2869 mtx_unlock(&uma_mtx);
2870
2871 buflen = sizeof(ush) + count * (sizeof(uth) + sizeof(ups) *
2872 (mp_maxid + 1)) + 1;
2873 buffer = malloc(buflen, M_TEMP, M_WAITOK | M_ZERO);
2874
2875 mtx_lock(&uma_mtx);
2876 i = 0;
2877 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2878 LIST_FOREACH(z, &kz->uk_zones, uz_link)
2879 i++;
2880 }
2881 if (i > count) {
2882 free(buffer, M_TEMP);
2883 goto restart;
2884 }
2885 count = i;
2886
2887 sbuf_new(&sbuf, buffer, buflen, SBUF_FIXEDLEN);
2888
2889 /*
2890 * Insert stream header.
2891 */
2892 bzero(&ush, sizeof(ush));
2893 ush.ush_version = UMA_STREAM_VERSION;
2894 ush.ush_maxcpus = (mp_maxid + 1);
2895 ush.ush_count = count;
2896 if (sbuf_bcat(&sbuf, &ush, sizeof(ush)) < 0) {
2897 mtx_unlock(&uma_mtx);
2898 error = ENOMEM;
2899 goto out;
2900 }
2901
2902 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2903 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
2904 bzero(&uth, sizeof(uth));
2905 ZONE_LOCK(z);
2906 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
2907 uth.uth_align = kz->uk_align;
2908 uth.uth_pages = kz->uk_pages;
2909 uth.uth_keg_free = kz->uk_free;
2910 uth.uth_size = kz->uk_size;
2911 uth.uth_rsize = kz->uk_rsize;
2912 uth.uth_maxpages = kz->uk_maxpages;
2913 if (kz->uk_ppera > 1)
2914 uth.uth_limit = kz->uk_maxpages /
2915 kz->uk_ppera;
2916 else
2917 uth.uth_limit = kz->uk_maxpages *
2918 kz->uk_ipers;
2919
2920 /*
2921 * A zone is secondary is it is not the first entry
2922 * on the keg's zone list.
2923 */
2924 if ((kz->uk_flags & UMA_ZONE_SECONDARY) &&
2925 (LIST_FIRST(&kz->uk_zones) != z))
2926 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
2927
2928 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
2929 uth.uth_zone_free += bucket->ub_cnt;
2930 uth.uth_allocs = z->uz_allocs;
2931 uth.uth_frees = z->uz_frees;
2932 uth.uth_fails = z->uz_fails;
2933 if (sbuf_bcat(&sbuf, &uth, sizeof(uth)) < 0) {
2934 ZONE_UNLOCK(z);
2935 mtx_unlock(&uma_mtx);
2936 error = ENOMEM;
2937 goto out;
2938 }
2939 /*
2940 * While it is not normally safe to access the cache
2941 * bucket pointers while not on the CPU that owns the
2942 * cache, we only allow the pointers to be exchanged
2943 * without the zone lock held, not invalidated, so
2944 * accept the possible race associated with bucket
2945 * exchange during monitoring.
2946 */
2947 for (i = 0; i < (mp_maxid + 1); i++) {
2948 bzero(&ups, sizeof(ups));
2949 if (kz->uk_flags & UMA_ZFLAG_INTERNAL)
2950 goto skip;
2951 if (CPU_ABSENT(i))
2952 goto skip;
2953 cache = &z->uz_cpu[i];
2954 if (cache->uc_allocbucket != NULL)
2955 ups.ups_cache_free +=
2956 cache->uc_allocbucket->ub_cnt;
2957 if (cache->uc_freebucket != NULL)
2958 ups.ups_cache_free +=
2959 cache->uc_freebucket->ub_cnt;
2960 ups.ups_allocs = cache->uc_allocs;
2961 ups.ups_frees = cache->uc_frees;
2962skip:
2963 if (sbuf_bcat(&sbuf, &ups, sizeof(ups)) < 0) {
2964 ZONE_UNLOCK(z);
2965 mtx_unlock(&uma_mtx);
2966 error = ENOMEM;
2967 goto out;
2968 }
2969 }
2970 ZONE_UNLOCK(z);
2971 }
2972 }
2973 mtx_unlock(&uma_mtx);
2974 sbuf_finish(&sbuf);
2975 error = SYSCTL_OUT(req, sbuf_data(&sbuf), sbuf_len(&sbuf));
2976out:
2977 free(buffer, M_TEMP);
2978 return (error);
2979}
2980
2981#ifdef DDB
2982DB_SHOW_COMMAND(uma, db_show_uma)
2983{
2984 u_int64_t allocs, frees;
2985 uma_bucket_t bucket;
2986 uma_keg_t kz;
2987 uma_zone_t z;
2988 int cachefree;
2989
2990 db_printf("%18s %8s %8s %8s %12s\n", "Zone", "Size", "Used", "Free",
2991 "Requests");
2992 LIST_FOREACH(kz, &uma_kegs, uk_link) {
2993 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
2994 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
2995 allocs = z->uz_allocs;
2996 frees = z->uz_frees;
2997 cachefree = 0;
2998 } else
2999 uma_zone_sumstat(z, &cachefree, &allocs,
3000 &frees);
3001 if (!((kz->uk_flags & UMA_ZONE_SECONDARY) &&
3002 (LIST_FIRST(&kz->uk_zones) != z)))
3003 cachefree += kz->uk_free;
3004 LIST_FOREACH(bucket, &z->uz_full_bucket, ub_link)
3005 cachefree += bucket->ub_cnt;
3006 db_printf("%18s %8ju %8jd %8d %12ju\n", z->uz_name,
3007 (uintmax_t)kz->uk_size,
3008 (intmax_t)(allocs - frees), cachefree,
3009 (uintmax_t)allocs);
3010 }
3011 }
3012}
3013#endif
|