1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause
3 *
4 * Copyright (c) 2002-2019 Jeffrey Roberson <jeff@FreeBSD.org>
5 * Copyright (c) 2004, 2005 Bosko Milekic <bmilekic@FreeBSD.org>
6 * Copyright (c) 2004-2006 Robert N. M. Watson
7 * All rights reserved.
8 *
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
11 * are met:
12 * 1. Redistributions of source code must retain the above copyright
13 * notice unmodified, this list of conditions, and the following
14 * disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
20 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
21 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
22 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
24 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
28 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 */
30
31 /*
32 * uma_core.c Implementation of the Universal Memory allocator
33 *
34 * This allocator is intended to replace the multitude of similar object caches
35 * in the standard FreeBSD kernel. The intent is to be flexible as well as
36 * efficient. A primary design goal is to return unused memory to the rest of
37 * the system. This will make the system as a whole more flexible due to the
38 * ability to move memory to subsystems which most need it instead of leaving
39 * pools of reserved memory unused.
40 *
41 * The basic ideas stem from similar slab/zone based allocators whose algorithms
42 * are well known.
43 *
44 */
45
46 /*
47 * TODO:
48 * - Improve memory usage for large allocations
49 * - Investigate cache size adjustments
50 */
51
52 #include <sys/cdefs.h>
53 #include "opt_ddb.h"
54 #include "opt_param.h"
55 #include "opt_vm.h"
56
57 #include <sys/param.h>
58 #include <sys/systm.h>
59 #include <sys/asan.h>
60 #include <sys/bitset.h>
61 #include <sys/domainset.h>
62 #include <sys/eventhandler.h>
63 #include <sys/kernel.h>
64 #include <sys/types.h>
65 #include <sys/limits.h>
66 #include <sys/queue.h>
67 #include <sys/malloc.h>
68 #include <sys/ktr.h>
69 #include <sys/lock.h>
70 #include <sys/msan.h>
71 #include <sys/mutex.h>
72 #include <sys/proc.h>
73 #include <sys/random.h>
74 #include <sys/rwlock.h>
75 #include <sys/sbuf.h>
76 #include <sys/sched.h>
77 #include <sys/sleepqueue.h>
78 #include <sys/smp.h>
79 #include <sys/smr.h>
80 #include <sys/sysctl.h>
81 #include <sys/taskqueue.h>
82 #include <sys/vmmeter.h>
83
84 #include <vm/vm.h>
85 #include <vm/vm_param.h>
86 #include <vm/vm_domainset.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_phys.h>
90 #include <vm/vm_pagequeue.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_kern.h>
93 #include <vm/vm_extern.h>
94 #include <vm/vm_dumpset.h>
95 #include <vm/uma.h>
96 #include <vm/uma_int.h>
97 #include <vm/uma_dbg.h>
98
99 #include <ddb/ddb.h>
100
101 #ifdef DEBUG_MEMGUARD
102 #include <vm/memguard.h>
103 #endif
104
105 #include <machine/md_var.h>
106
107 #ifdef INVARIANTS
108 #define UMA_ALWAYS_CTORDTOR 1
109 #else
110 #define UMA_ALWAYS_CTORDTOR 0
111 #endif
112
113 /*
114 * This is the zone and keg from which all zones are spawned.
115 */
116 static uma_zone_t kegs;
117 static uma_zone_t zones;
118
119 /*
120 * On INVARIANTS builds, the slab contains a second bitset of the same size,
121 * "dbg_bits", which is laid out immediately after us_free.
122 */
123 #ifdef INVARIANTS
124 #define SLAB_BITSETS 2
125 #else
126 #define SLAB_BITSETS 1
127 #endif
128
129 /*
130 * These are the two zones from which all offpage uma_slab_ts are allocated.
131 *
132 * One zone is for slab headers that can represent a larger number of items,
133 * making the slabs themselves more efficient, and the other zone is for
134 * headers that are smaller and represent fewer items, making the headers more
135 * efficient.
136 */
137 #define SLABZONE_SIZE(setsize) \
138 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
139 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16)
140 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE
141 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE)
142 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE)
143 static uma_zone_t slabzones[2];
144
145 /*
146 * The initial hash tables come out of this zone so they can be allocated
147 * prior to malloc coming up.
148 */
149 static uma_zone_t hashzone;
150
151 /* The boot-time adjusted value for cache line alignment. */
152 static unsigned int uma_cache_align_mask = 64 - 1;
153
154 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
155 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
156
157 /*
158 * Are we allowed to allocate buckets?
159 */
160 static int bucketdisable = 1;
161
162 /* Linked list of all kegs in the system */
163 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
164
165 /* Linked list of all cache-only zones in the system */
166 static LIST_HEAD(,uma_zone) uma_cachezones =
167 LIST_HEAD_INITIALIZER(uma_cachezones);
168
169 /*
170 * Mutex for global lists: uma_kegs, uma_cachezones, and the per-keg list of
171 * zones.
172 */
173 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
174
175 static struct sx uma_reclaim_lock;
176
177 /*
178 * First available virual address for boot time allocations.
179 */
180 static vm_offset_t bootstart;
181 static vm_offset_t bootmem;
182
183 /*
184 * kmem soft limit, initialized by uma_set_limit(). Ensure that early
185 * allocations don't trigger a wakeup of the reclaim thread.
186 */
187 unsigned long uma_kmem_limit = LONG_MAX;
188 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
189 "UMA kernel memory soft limit");
190 unsigned long uma_kmem_total;
191 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
192 "UMA kernel memory usage");
193
194 /* Is the VM done starting up? */
195 static enum {
196 BOOT_COLD,
197 BOOT_KVA,
198 BOOT_PCPU,
199 BOOT_RUNNING,
200 BOOT_SHUTDOWN,
201 } booted = BOOT_COLD;
202
203 /*
204 * This is the handle used to schedule events that need to happen
205 * outside of the allocation fast path.
206 */
207 static struct timeout_task uma_timeout_task;
208 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
209
210 /*
211 * This structure is passed as the zone ctor arg so that I don't have to create
212 * a special allocation function just for zones.
213 */
214 struct uma_zctor_args {
215 const char *name;
216 size_t size;
217 uma_ctor ctor;
218 uma_dtor dtor;
219 uma_init uminit;
220 uma_fini fini;
221 uma_import import;
222 uma_release release;
223 void *arg;
224 uma_keg_t keg;
225 int align;
226 uint32_t flags;
227 };
228
229 struct uma_kctor_args {
230 uma_zone_t zone;
231 size_t size;
232 uma_init uminit;
233 uma_fini fini;
234 int align;
235 uint32_t flags;
236 };
237
238 struct uma_bucket_zone {
239 uma_zone_t ubz_zone;
240 const char *ubz_name;
241 int ubz_entries; /* Number of items it can hold. */
242 int ubz_maxsize; /* Maximum allocation size per-item. */
243 };
244
245 /*
246 * Compute the actual number of bucket entries to pack them in power
247 * of two sizes for more efficient space utilization.
248 */
249 #define BUCKET_SIZE(n) \
250 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
251
252 #define BUCKET_MAX BUCKET_SIZE(256)
253
254 struct uma_bucket_zone bucket_zones[] = {
255 /* Literal bucket sizes. */
256 { NULL, "2 Bucket", 2, 4096 },
257 { NULL, "4 Bucket", 4, 3072 },
258 { NULL, "8 Bucket", 8, 2048 },
259 { NULL, "16 Bucket", 16, 1024 },
260 /* Rounded down power of 2 sizes for efficiency. */
261 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
262 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
263 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
264 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
265 { NULL, NULL, 0}
266 };
267
268 /*
269 * Flags and enumerations to be passed to internal functions.
270 */
271 enum zfreeskip {
272 SKIP_NONE = 0,
273 SKIP_CNT = 0x00000001,
274 SKIP_DTOR = 0x00010000,
275 SKIP_FINI = 0x00020000,
276 };
277
278 /* Prototypes.. */
279
280 void uma_startup1(vm_offset_t);
281 void uma_startup2(void);
282
283 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
284 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
285 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
286 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
287 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
288 static void page_free(void *, vm_size_t, uint8_t);
289 static void pcpu_page_free(void *, vm_size_t, uint8_t);
290 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
291 static void cache_drain(uma_zone_t);
292 static void bucket_drain(uma_zone_t, uma_bucket_t);
293 static void bucket_cache_reclaim(uma_zone_t zone, bool, int);
294 static bool bucket_cache_reclaim_domain(uma_zone_t, bool, bool, int);
295 static int keg_ctor(void *, int, void *, int);
296 static void keg_dtor(void *, int, void *);
297 static void keg_drain(uma_keg_t keg, int domain);
298 static int zone_ctor(void *, int, void *, int);
299 static void zone_dtor(void *, int, void *);
300 static inline void item_dtor(uma_zone_t zone, void *item, int size,
301 void *udata, enum zfreeskip skip);
302 static int zero_init(void *, int, int);
303 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
304 int itemdomain, bool ws);
305 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
306 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
307 static void zone_timeout(uma_zone_t zone, void *);
308 static int hash_alloc(struct uma_hash *, u_int);
309 static int hash_expand(struct uma_hash *, struct uma_hash *);
310 static void hash_free(struct uma_hash *hash);
311 static void uma_timeout(void *, int);
312 static void uma_shutdown(void);
313 static void *zone_alloc_item(uma_zone_t, void *, int, int);
314 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
315 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
316 static void zone_free_limit(uma_zone_t zone, int count);
317 static void bucket_enable(void);
318 static void bucket_init(void);
319 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
320 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
321 static void bucket_zone_drain(int domain);
322 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
323 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
324 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
325 static size_t slab_sizeof(int nitems);
326 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
327 uma_fini fini, int align, uint32_t flags);
328 static int zone_import(void *, void **, int, int, int);
329 static void zone_release(void *, void **, int);
330 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
331 static bool cache_free(uma_zone_t, uma_cache_t, void *, int);
332
333 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
334 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
335 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
336 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
337 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
338 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
339 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
340
341 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
342
343 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
344 "Memory allocation debugging");
345
346 #ifdef INVARIANTS
347 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
348 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
349
350 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
351 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
352 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
353 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
354
355 static u_int dbg_divisor = 1;
356 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
357 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
358 "Debug & thrash every this item in memory allocator");
359
360 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
361 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
362 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
363 &uma_dbg_cnt, "memory items debugged");
364 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
365 &uma_skip_cnt, "memory items skipped, not debugged");
366 #endif
367
368 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
369 "Universal Memory Allocator");
370
371 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
372 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
373
374 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
375 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
376
377 static int zone_warnings = 1;
378 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
379 "Warn when UMA zones becomes full");
380
381 static int multipage_slabs = 1;
382 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
383 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
384 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
385 "UMA may choose larger slab sizes for better efficiency");
386
387 /*
388 * Select the slab zone for an offpage slab with the given maximum item count.
389 */
390 static inline uma_zone_t
slabzone(int ipers)391 slabzone(int ipers)
392 {
393
394 return (slabzones[ipers > SLABZONE0_SETSIZE]);
395 }
396
397 /*
398 * This routine checks to see whether or not it's safe to enable buckets.
399 */
400 static void
bucket_enable(void)401 bucket_enable(void)
402 {
403
404 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
405 bucketdisable = vm_page_count_min();
406 }
407
408 /*
409 * Initialize bucket_zones, the array of zones of buckets of various sizes.
410 *
411 * For each zone, calculate the memory required for each bucket, consisting
412 * of the header and an array of pointers.
413 */
414 static void
bucket_init(void)415 bucket_init(void)
416 {
417 struct uma_bucket_zone *ubz;
418 int size;
419
420 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
421 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
422 size += sizeof(void *) * ubz->ubz_entries;
423 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
424 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
425 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
426 UMA_ZONE_FIRSTTOUCH);
427 }
428 }
429
430 /*
431 * Given a desired number of entries for a bucket, return the zone from which
432 * to allocate the bucket.
433 */
434 static struct uma_bucket_zone *
bucket_zone_lookup(int entries)435 bucket_zone_lookup(int entries)
436 {
437 struct uma_bucket_zone *ubz;
438
439 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
440 if (ubz->ubz_entries >= entries)
441 return (ubz);
442 ubz--;
443 return (ubz);
444 }
445
446 static int
bucket_select(int size)447 bucket_select(int size)
448 {
449 struct uma_bucket_zone *ubz;
450
451 ubz = &bucket_zones[0];
452 if (size > ubz->ubz_maxsize)
453 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
454
455 for (; ubz->ubz_entries != 0; ubz++)
456 if (ubz->ubz_maxsize < size)
457 break;
458 ubz--;
459 return (ubz->ubz_entries);
460 }
461
462 static uma_bucket_t
bucket_alloc(uma_zone_t zone,void * udata,int flags)463 bucket_alloc(uma_zone_t zone, void *udata, int flags)
464 {
465 struct uma_bucket_zone *ubz;
466 uma_bucket_t bucket;
467
468 /*
469 * Don't allocate buckets early in boot.
470 */
471 if (__predict_false(booted < BOOT_KVA))
472 return (NULL);
473
474 /*
475 * To limit bucket recursion we store the original zone flags
476 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
477 * NOVM flag to persist even through deep recursions. We also
478 * store ZFLAG_BUCKET once we have recursed attempting to allocate
479 * a bucket for a bucket zone so we do not allow infinite bucket
480 * recursion. This cookie will even persist to frees of unused
481 * buckets via the allocation path or bucket allocations in the
482 * free path.
483 */
484 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
485 udata = (void *)(uintptr_t)zone->uz_flags;
486 else {
487 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
488 return (NULL);
489 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
490 }
491 if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
492 flags |= M_NOVM;
493 ubz = bucket_zone_lookup(atomic_load_16(&zone->uz_bucket_size));
494 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
495 ubz++;
496 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
497 if (bucket) {
498 #ifdef INVARIANTS
499 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
500 #endif
501 bucket->ub_cnt = 0;
502 bucket->ub_entries = min(ubz->ubz_entries,
503 zone->uz_bucket_size_max);
504 bucket->ub_seq = SMR_SEQ_INVALID;
505 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
506 zone->uz_name, zone, bucket);
507 }
508
509 return (bucket);
510 }
511
512 static void
bucket_free(uma_zone_t zone,uma_bucket_t bucket,void * udata)513 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
514 {
515 struct uma_bucket_zone *ubz;
516
517 if (bucket->ub_cnt != 0)
518 bucket_drain(zone, bucket);
519
520 KASSERT(bucket->ub_cnt == 0,
521 ("bucket_free: Freeing a non free bucket."));
522 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
523 ("bucket_free: Freeing an SMR bucket."));
524 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
525 udata = (void *)(uintptr_t)zone->uz_flags;
526 ubz = bucket_zone_lookup(bucket->ub_entries);
527 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
528 }
529
530 static void
bucket_zone_drain(int domain)531 bucket_zone_drain(int domain)
532 {
533 struct uma_bucket_zone *ubz;
534
535 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
536 uma_zone_reclaim_domain(ubz->ubz_zone, UMA_RECLAIM_DRAIN,
537 domain);
538 }
539
540 #ifdef KASAN
541 _Static_assert(UMA_SMALLEST_UNIT % KASAN_SHADOW_SCALE == 0,
542 "Base UMA allocation size not a multiple of the KASAN scale factor");
543
544 static void
kasan_mark_item_valid(uma_zone_t zone,void * item)545 kasan_mark_item_valid(uma_zone_t zone, void *item)
546 {
547 void *pcpu_item;
548 size_t sz, rsz;
549 int i;
550
551 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
552 return;
553
554 sz = zone->uz_size;
555 rsz = roundup2(sz, KASAN_SHADOW_SCALE);
556 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
557 kasan_mark(item, sz, rsz, KASAN_GENERIC_REDZONE);
558 } else {
559 pcpu_item = zpcpu_base_to_offset(item);
560 for (i = 0; i <= mp_maxid; i++)
561 kasan_mark(zpcpu_get_cpu(pcpu_item, i), sz, rsz,
562 KASAN_GENERIC_REDZONE);
563 }
564 }
565
566 static void
kasan_mark_item_invalid(uma_zone_t zone,void * item)567 kasan_mark_item_invalid(uma_zone_t zone, void *item)
568 {
569 void *pcpu_item;
570 size_t sz;
571 int i;
572
573 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
574 return;
575
576 sz = roundup2(zone->uz_size, KASAN_SHADOW_SCALE);
577 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
578 kasan_mark(item, 0, sz, KASAN_UMA_FREED);
579 } else {
580 pcpu_item = zpcpu_base_to_offset(item);
581 for (i = 0; i <= mp_maxid; i++)
582 kasan_mark(zpcpu_get_cpu(pcpu_item, i), 0, sz,
583 KASAN_UMA_FREED);
584 }
585 }
586
587 static void
kasan_mark_slab_valid(uma_keg_t keg,void * mem)588 kasan_mark_slab_valid(uma_keg_t keg, void *mem)
589 {
590 size_t sz;
591
592 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
593 sz = keg->uk_ppera * PAGE_SIZE;
594 kasan_mark(mem, sz, sz, 0);
595 }
596 }
597
598 static void
kasan_mark_slab_invalid(uma_keg_t keg,void * mem)599 kasan_mark_slab_invalid(uma_keg_t keg, void *mem)
600 {
601 size_t sz;
602
603 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
604 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
605 sz = keg->uk_ppera * PAGE_SIZE;
606 else
607 sz = keg->uk_pgoff;
608 kasan_mark(mem, 0, sz, KASAN_UMA_FREED);
609 }
610 }
611 #else /* !KASAN */
612 static void
kasan_mark_item_valid(uma_zone_t zone __unused,void * item __unused)613 kasan_mark_item_valid(uma_zone_t zone __unused, void *item __unused)
614 {
615 }
616
617 static void
kasan_mark_item_invalid(uma_zone_t zone __unused,void * item __unused)618 kasan_mark_item_invalid(uma_zone_t zone __unused, void *item __unused)
619 {
620 }
621
622 static void
kasan_mark_slab_valid(uma_keg_t keg __unused,void * mem __unused)623 kasan_mark_slab_valid(uma_keg_t keg __unused, void *mem __unused)
624 {
625 }
626
627 static void
kasan_mark_slab_invalid(uma_keg_t keg __unused,void * mem __unused)628 kasan_mark_slab_invalid(uma_keg_t keg __unused, void *mem __unused)
629 {
630 }
631 #endif /* KASAN */
632
633 #ifdef KMSAN
634 static inline void
kmsan_mark_item_uninitialized(uma_zone_t zone,void * item)635 kmsan_mark_item_uninitialized(uma_zone_t zone, void *item)
636 {
637 void *pcpu_item;
638 size_t sz;
639 int i;
640
641 if ((zone->uz_flags &
642 (UMA_ZFLAG_CACHE | UMA_ZONE_SECONDARY | UMA_ZONE_MALLOC)) != 0) {
643 /*
644 * Cache zones should not be instrumented by default, as UMA
645 * does not have enough information to do so correctly.
646 * Consumers can mark items themselves if it makes sense to do
647 * so.
648 *
649 * Items from secondary zones are initialized by the parent
650 * zone and thus cannot safely be marked by UMA.
651 *
652 * malloc zones are handled directly by malloc(9) and friends,
653 * since they can provide more precise origin tracking.
654 */
655 return;
656 }
657 if (zone->uz_keg->uk_init != NULL) {
658 /*
659 * By definition, initialized items cannot be marked. The
660 * best we can do is mark items from these zones after they
661 * are freed to the keg.
662 */
663 return;
664 }
665
666 sz = zone->uz_size;
667 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
668 kmsan_orig(item, sz, KMSAN_TYPE_UMA, KMSAN_RET_ADDR);
669 kmsan_mark(item, sz, KMSAN_STATE_UNINIT);
670 } else {
671 pcpu_item = zpcpu_base_to_offset(item);
672 for (i = 0; i <= mp_maxid; i++) {
673 kmsan_orig(zpcpu_get_cpu(pcpu_item, i), sz,
674 KMSAN_TYPE_UMA, KMSAN_RET_ADDR);
675 kmsan_mark(zpcpu_get_cpu(pcpu_item, i), sz,
676 KMSAN_STATE_INITED);
677 }
678 }
679 }
680 #else /* !KMSAN */
681 static inline void
kmsan_mark_item_uninitialized(uma_zone_t zone __unused,void * item __unused)682 kmsan_mark_item_uninitialized(uma_zone_t zone __unused, void *item __unused)
683 {
684 }
685 #endif /* KMSAN */
686
687 /*
688 * Acquire the domain lock and record contention.
689 */
690 static uma_zone_domain_t
zone_domain_lock(uma_zone_t zone,int domain)691 zone_domain_lock(uma_zone_t zone, int domain)
692 {
693 uma_zone_domain_t zdom;
694 bool lockfail;
695
696 zdom = ZDOM_GET(zone, domain);
697 lockfail = false;
698 if (ZDOM_OWNED(zdom))
699 lockfail = true;
700 ZDOM_LOCK(zdom);
701 /* This is unsynchronized. The counter does not need to be precise. */
702 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
703 zone->uz_bucket_size++;
704 return (zdom);
705 }
706
707 /*
708 * Search for the domain with the least cached items and return it if it
709 * is out of balance with the preferred domain.
710 */
711 static __noinline int
zone_domain_lowest(uma_zone_t zone,int pref)712 zone_domain_lowest(uma_zone_t zone, int pref)
713 {
714 long least, nitems, prefitems;
715 int domain;
716 int i;
717
718 prefitems = least = LONG_MAX;
719 domain = 0;
720 for (i = 0; i < vm_ndomains; i++) {
721 nitems = ZDOM_GET(zone, i)->uzd_nitems;
722 if (nitems < least) {
723 domain = i;
724 least = nitems;
725 }
726 if (domain == pref)
727 prefitems = nitems;
728 }
729 if (prefitems < least * 2)
730 return (pref);
731
732 return (domain);
733 }
734
735 /*
736 * Search for the domain with the most cached items and return it or the
737 * preferred domain if it has enough to proceed.
738 */
739 static __noinline int
zone_domain_highest(uma_zone_t zone,int pref)740 zone_domain_highest(uma_zone_t zone, int pref)
741 {
742 long most, nitems;
743 int domain;
744 int i;
745
746 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
747 return (pref);
748
749 most = 0;
750 domain = 0;
751 for (i = 0; i < vm_ndomains; i++) {
752 nitems = ZDOM_GET(zone, i)->uzd_nitems;
753 if (nitems > most) {
754 domain = i;
755 most = nitems;
756 }
757 }
758
759 return (domain);
760 }
761
762 /*
763 * Set the maximum imax value.
764 */
765 static void
zone_domain_imax_set(uma_zone_domain_t zdom,int nitems)766 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
767 {
768 long old;
769
770 old = zdom->uzd_imax;
771 do {
772 if (old >= nitems)
773 return;
774 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
775
776 /*
777 * We are at new maximum, so do the last WSS update for the old
778 * bimin and prepare to measure next allocation batch.
779 */
780 if (zdom->uzd_wss < old - zdom->uzd_bimin)
781 zdom->uzd_wss = old - zdom->uzd_bimin;
782 zdom->uzd_bimin = nitems;
783 }
784
785 /*
786 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
787 * zone's caches. If a bucket is found the zone is not locked on return.
788 */
789 static uma_bucket_t
zone_fetch_bucket(uma_zone_t zone,uma_zone_domain_t zdom,bool reclaim)790 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
791 {
792 uma_bucket_t bucket;
793 long cnt;
794 int i;
795 bool dtor = false;
796
797 ZDOM_LOCK_ASSERT(zdom);
798
799 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
800 return (NULL);
801
802 /* SMR Buckets can not be re-used until readers expire. */
803 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
804 bucket->ub_seq != SMR_SEQ_INVALID) {
805 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
806 return (NULL);
807 bucket->ub_seq = SMR_SEQ_INVALID;
808 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
809 if (STAILQ_NEXT(bucket, ub_link) != NULL)
810 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
811 }
812 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
813
814 KASSERT(zdom->uzd_nitems >= bucket->ub_cnt,
815 ("%s: item count underflow (%ld, %d)",
816 __func__, zdom->uzd_nitems, bucket->ub_cnt));
817 KASSERT(bucket->ub_cnt > 0,
818 ("%s: empty bucket in bucket cache", __func__));
819 zdom->uzd_nitems -= bucket->ub_cnt;
820
821 if (reclaim) {
822 /*
823 * Shift the bounds of the current WSS interval to avoid
824 * perturbing the estimates.
825 */
826 cnt = lmin(zdom->uzd_bimin, bucket->ub_cnt);
827 atomic_subtract_long(&zdom->uzd_imax, cnt);
828 zdom->uzd_bimin -= cnt;
829 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
830 if (zdom->uzd_limin >= bucket->ub_cnt) {
831 zdom->uzd_limin -= bucket->ub_cnt;
832 } else {
833 zdom->uzd_limin = 0;
834 zdom->uzd_timin = 0;
835 }
836 } else if (zdom->uzd_bimin > zdom->uzd_nitems) {
837 zdom->uzd_bimin = zdom->uzd_nitems;
838 if (zdom->uzd_imin > zdom->uzd_nitems)
839 zdom->uzd_imin = zdom->uzd_nitems;
840 }
841
842 ZDOM_UNLOCK(zdom);
843 if (dtor)
844 for (i = 0; i < bucket->ub_cnt; i++)
845 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
846 NULL, SKIP_NONE);
847
848 return (bucket);
849 }
850
851 /*
852 * Insert a full bucket into the specified cache. The "ws" parameter indicates
853 * whether the bucket's contents should be counted as part of the zone's working
854 * set. The bucket may be freed if it exceeds the bucket limit.
855 */
856 static void
zone_put_bucket(uma_zone_t zone,int domain,uma_bucket_t bucket,void * udata,const bool ws)857 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
858 const bool ws)
859 {
860 uma_zone_domain_t zdom;
861
862 /* We don't cache empty buckets. This can happen after a reclaim. */
863 if (bucket->ub_cnt == 0)
864 goto out;
865 zdom = zone_domain_lock(zone, domain);
866
867 /*
868 * Conditionally set the maximum number of items.
869 */
870 zdom->uzd_nitems += bucket->ub_cnt;
871 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
872 if (ws) {
873 zone_domain_imax_set(zdom, zdom->uzd_nitems);
874 } else {
875 /*
876 * Shift the bounds of the current WSS interval to
877 * avoid perturbing the estimates.
878 */
879 atomic_add_long(&zdom->uzd_imax, bucket->ub_cnt);
880 zdom->uzd_imin += bucket->ub_cnt;
881 zdom->uzd_bimin += bucket->ub_cnt;
882 zdom->uzd_limin += bucket->ub_cnt;
883 }
884 if (STAILQ_EMPTY(&zdom->uzd_buckets))
885 zdom->uzd_seq = bucket->ub_seq;
886
887 /*
888 * Try to promote reuse of recently used items. For items
889 * protected by SMR, try to defer reuse to minimize polling.
890 */
891 if (bucket->ub_seq == SMR_SEQ_INVALID)
892 STAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
893 else
894 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
895 ZDOM_UNLOCK(zdom);
896 return;
897 }
898 zdom->uzd_nitems -= bucket->ub_cnt;
899 ZDOM_UNLOCK(zdom);
900 out:
901 bucket_free(zone, bucket, udata);
902 }
903
904 /* Pops an item out of a per-cpu cache bucket. */
905 static inline void *
cache_bucket_pop(uma_cache_t cache,uma_cache_bucket_t bucket)906 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
907 {
908 void *item;
909
910 CRITICAL_ASSERT(curthread);
911
912 bucket->ucb_cnt--;
913 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
914 #ifdef INVARIANTS
915 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
916 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
917 #endif
918 cache->uc_allocs++;
919
920 return (item);
921 }
922
923 /* Pushes an item into a per-cpu cache bucket. */
924 static inline void
cache_bucket_push(uma_cache_t cache,uma_cache_bucket_t bucket,void * item)925 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
926 {
927
928 CRITICAL_ASSERT(curthread);
929 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
930 ("uma_zfree: Freeing to non free bucket index."));
931
932 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
933 bucket->ucb_cnt++;
934 cache->uc_frees++;
935 }
936
937 /*
938 * Unload a UMA bucket from a per-cpu cache.
939 */
940 static inline uma_bucket_t
cache_bucket_unload(uma_cache_bucket_t bucket)941 cache_bucket_unload(uma_cache_bucket_t bucket)
942 {
943 uma_bucket_t b;
944
945 b = bucket->ucb_bucket;
946 if (b != NULL) {
947 MPASS(b->ub_entries == bucket->ucb_entries);
948 b->ub_cnt = bucket->ucb_cnt;
949 bucket->ucb_bucket = NULL;
950 bucket->ucb_entries = bucket->ucb_cnt = 0;
951 }
952
953 return (b);
954 }
955
956 static inline uma_bucket_t
cache_bucket_unload_alloc(uma_cache_t cache)957 cache_bucket_unload_alloc(uma_cache_t cache)
958 {
959
960 return (cache_bucket_unload(&cache->uc_allocbucket));
961 }
962
963 static inline uma_bucket_t
cache_bucket_unload_free(uma_cache_t cache)964 cache_bucket_unload_free(uma_cache_t cache)
965 {
966
967 return (cache_bucket_unload(&cache->uc_freebucket));
968 }
969
970 static inline uma_bucket_t
cache_bucket_unload_cross(uma_cache_t cache)971 cache_bucket_unload_cross(uma_cache_t cache)
972 {
973
974 return (cache_bucket_unload(&cache->uc_crossbucket));
975 }
976
977 /*
978 * Load a bucket into a per-cpu cache bucket.
979 */
980 static inline void
cache_bucket_load(uma_cache_bucket_t bucket,uma_bucket_t b)981 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
982 {
983
984 CRITICAL_ASSERT(curthread);
985 MPASS(bucket->ucb_bucket == NULL);
986 MPASS(b->ub_seq == SMR_SEQ_INVALID);
987
988 bucket->ucb_bucket = b;
989 bucket->ucb_cnt = b->ub_cnt;
990 bucket->ucb_entries = b->ub_entries;
991 }
992
993 static inline void
cache_bucket_load_alloc(uma_cache_t cache,uma_bucket_t b)994 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
995 {
996
997 cache_bucket_load(&cache->uc_allocbucket, b);
998 }
999
1000 static inline void
cache_bucket_load_free(uma_cache_t cache,uma_bucket_t b)1001 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
1002 {
1003
1004 cache_bucket_load(&cache->uc_freebucket, b);
1005 }
1006
1007 #ifdef NUMA
1008 static inline void
cache_bucket_load_cross(uma_cache_t cache,uma_bucket_t b)1009 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
1010 {
1011
1012 cache_bucket_load(&cache->uc_crossbucket, b);
1013 }
1014 #endif
1015
1016 /*
1017 * Copy and preserve ucb_spare.
1018 */
1019 static inline void
cache_bucket_copy(uma_cache_bucket_t b1,uma_cache_bucket_t b2)1020 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
1021 {
1022
1023 b1->ucb_bucket = b2->ucb_bucket;
1024 b1->ucb_entries = b2->ucb_entries;
1025 b1->ucb_cnt = b2->ucb_cnt;
1026 }
1027
1028 /*
1029 * Swap two cache buckets.
1030 */
1031 static inline void
cache_bucket_swap(uma_cache_bucket_t b1,uma_cache_bucket_t b2)1032 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
1033 {
1034 struct uma_cache_bucket b3;
1035
1036 CRITICAL_ASSERT(curthread);
1037
1038 cache_bucket_copy(&b3, b1);
1039 cache_bucket_copy(b1, b2);
1040 cache_bucket_copy(b2, &b3);
1041 }
1042
1043 /*
1044 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
1045 */
1046 static uma_bucket_t
cache_fetch_bucket(uma_zone_t zone,uma_cache_t cache,int domain)1047 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
1048 {
1049 uma_zone_domain_t zdom;
1050 uma_bucket_t bucket;
1051 smr_seq_t seq;
1052
1053 /*
1054 * Avoid the lock if possible.
1055 */
1056 zdom = ZDOM_GET(zone, domain);
1057 if (zdom->uzd_nitems == 0)
1058 return (NULL);
1059
1060 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
1061 (seq = atomic_load_32(&zdom->uzd_seq)) != SMR_SEQ_INVALID &&
1062 !smr_poll(zone->uz_smr, seq, false))
1063 return (NULL);
1064
1065 /*
1066 * Check the zone's cache of buckets.
1067 */
1068 zdom = zone_domain_lock(zone, domain);
1069 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL)
1070 return (bucket);
1071 ZDOM_UNLOCK(zdom);
1072
1073 return (NULL);
1074 }
1075
1076 static void
zone_log_warning(uma_zone_t zone)1077 zone_log_warning(uma_zone_t zone)
1078 {
1079 static const struct timeval warninterval = { 300, 0 };
1080
1081 if (!zone_warnings || zone->uz_warning == NULL)
1082 return;
1083
1084 if (ratecheck(&zone->uz_ratecheck, &warninterval))
1085 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
1086 }
1087
1088 static inline void
zone_maxaction(uma_zone_t zone)1089 zone_maxaction(uma_zone_t zone)
1090 {
1091
1092 if (zone->uz_maxaction.ta_func != NULL)
1093 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
1094 }
1095
1096 /*
1097 * Routine called by timeout which is used to fire off some time interval
1098 * based calculations. (stats, hash size, etc.)
1099 *
1100 * Arguments:
1101 * arg Unused
1102 *
1103 * Returns:
1104 * Nothing
1105 */
1106 static void
uma_timeout(void * context __unused,int pending __unused)1107 uma_timeout(void *context __unused, int pending __unused)
1108 {
1109 bucket_enable();
1110 zone_foreach(zone_timeout, NULL);
1111
1112 /* Reschedule this event */
1113 taskqueue_enqueue_timeout(taskqueue_thread, &uma_timeout_task,
1114 UMA_TIMEOUT * hz);
1115 }
1116
1117 /*
1118 * Update the working set size estimates for the zone's bucket cache.
1119 * The constants chosen here are somewhat arbitrary.
1120 */
1121 static void
zone_domain_update_wss(uma_zone_domain_t zdom)1122 zone_domain_update_wss(uma_zone_domain_t zdom)
1123 {
1124 long m;
1125
1126 ZDOM_LOCK_ASSERT(zdom);
1127 MPASS(zdom->uzd_imax >= zdom->uzd_nitems);
1128 MPASS(zdom->uzd_nitems >= zdom->uzd_bimin);
1129 MPASS(zdom->uzd_bimin >= zdom->uzd_imin);
1130
1131 /*
1132 * Estimate WSS as modified moving average of biggest allocation
1133 * batches for each period over few minutes (UMA_TIMEOUT of 20s).
1134 */
1135 zdom->uzd_wss = lmax(zdom->uzd_wss * 3 / 4,
1136 zdom->uzd_imax - zdom->uzd_bimin);
1137
1138 /*
1139 * Estimate longtime minimum item count as a combination of recent
1140 * minimum item count, adjusted by WSS for safety, and the modified
1141 * moving average over the last several hours (UMA_TIMEOUT of 20s).
1142 * timin measures time since limin tried to go negative, that means
1143 * we were dangerously close to or got out of cache.
1144 */
1145 m = zdom->uzd_imin - zdom->uzd_wss;
1146 if (m >= 0) {
1147 if (zdom->uzd_limin >= m)
1148 zdom->uzd_limin = m;
1149 else
1150 zdom->uzd_limin = (m + zdom->uzd_limin * 255) / 256;
1151 zdom->uzd_timin++;
1152 } else {
1153 zdom->uzd_limin = 0;
1154 zdom->uzd_timin = 0;
1155 }
1156
1157 /* To reduce period edge effects on WSS keep half of the imax. */
1158 atomic_subtract_long(&zdom->uzd_imax,
1159 (zdom->uzd_imax - zdom->uzd_nitems + 1) / 2);
1160 zdom->uzd_imin = zdom->uzd_bimin = zdom->uzd_nitems;
1161 }
1162
1163 /*
1164 * Routine to perform timeout driven calculations. This expands the
1165 * hashes and does per cpu statistics aggregation.
1166 *
1167 * Returns nothing.
1168 */
1169 static void
zone_timeout(uma_zone_t zone,void * unused)1170 zone_timeout(uma_zone_t zone, void *unused)
1171 {
1172 uma_keg_t keg;
1173 u_int slabs, pages;
1174
1175 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
1176 goto trim;
1177
1178 keg = zone->uz_keg;
1179
1180 /*
1181 * Hash zones are non-numa by definition so the first domain
1182 * is the only one present.
1183 */
1184 KEG_LOCK(keg, 0);
1185 pages = keg->uk_domain[0].ud_pages;
1186
1187 /*
1188 * Expand the keg hash table.
1189 *
1190 * This is done if the number of slabs is larger than the hash size.
1191 * What I'm trying to do here is completely reduce collisions. This
1192 * may be a little aggressive. Should I allow for two collisions max?
1193 */
1194 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1195 struct uma_hash newhash;
1196 struct uma_hash oldhash;
1197 int ret;
1198
1199 /*
1200 * This is so involved because allocating and freeing
1201 * while the keg lock is held will lead to deadlock.
1202 * I have to do everything in stages and check for
1203 * races.
1204 */
1205 KEG_UNLOCK(keg, 0);
1206 ret = hash_alloc(&newhash, 1 << fls(slabs));
1207 KEG_LOCK(keg, 0);
1208 if (ret) {
1209 if (hash_expand(&keg->uk_hash, &newhash)) {
1210 oldhash = keg->uk_hash;
1211 keg->uk_hash = newhash;
1212 } else
1213 oldhash = newhash;
1214
1215 KEG_UNLOCK(keg, 0);
1216 hash_free(&oldhash);
1217 goto trim;
1218 }
1219 }
1220 KEG_UNLOCK(keg, 0);
1221
1222 trim:
1223 /* Trim caches not used for a long time. */
1224 if ((zone->uz_flags & (UMA_ZONE_UNMANAGED | UMA_ZONE_NOTRIM)) == 0) {
1225 for (int i = 0; i < vm_ndomains; i++) {
1226 if (bucket_cache_reclaim_domain(zone, false, false, i) &&
1227 (zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1228 keg_drain(zone->uz_keg, i);
1229 }
1230 }
1231 }
1232
1233 /*
1234 * Allocate and zero fill the next sized hash table from the appropriate
1235 * backing store.
1236 *
1237 * Arguments:
1238 * hash A new hash structure with the old hash size in uh_hashsize
1239 *
1240 * Returns:
1241 * 1 on success and 0 on failure.
1242 */
1243 static int
hash_alloc(struct uma_hash * hash,u_int size)1244 hash_alloc(struct uma_hash *hash, u_int size)
1245 {
1246 size_t alloc;
1247
1248 KASSERT(powerof2(size), ("hash size must be power of 2"));
1249 if (size > UMA_HASH_SIZE_INIT) {
1250 hash->uh_hashsize = size;
1251 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1252 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1253 } else {
1254 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1255 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1256 UMA_ANYDOMAIN, M_WAITOK);
1257 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1258 }
1259 if (hash->uh_slab_hash) {
1260 bzero(hash->uh_slab_hash, alloc);
1261 hash->uh_hashmask = hash->uh_hashsize - 1;
1262 return (1);
1263 }
1264
1265 return (0);
1266 }
1267
1268 /*
1269 * Expands the hash table for HASH zones. This is done from zone_timeout
1270 * to reduce collisions. This must not be done in the regular allocation
1271 * path, otherwise, we can recurse on the vm while allocating pages.
1272 *
1273 * Arguments:
1274 * oldhash The hash you want to expand
1275 * newhash The hash structure for the new table
1276 *
1277 * Returns:
1278 * Nothing
1279 *
1280 * Discussion:
1281 */
1282 static int
hash_expand(struct uma_hash * oldhash,struct uma_hash * newhash)1283 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1284 {
1285 uma_hash_slab_t slab;
1286 u_int hval;
1287 u_int idx;
1288
1289 if (!newhash->uh_slab_hash)
1290 return (0);
1291
1292 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1293 return (0);
1294
1295 /*
1296 * I need to investigate hash algorithms for resizing without a
1297 * full rehash.
1298 */
1299
1300 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1301 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1302 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1303 LIST_REMOVE(slab, uhs_hlink);
1304 hval = UMA_HASH(newhash, slab->uhs_data);
1305 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1306 slab, uhs_hlink);
1307 }
1308
1309 return (1);
1310 }
1311
1312 /*
1313 * Free the hash bucket to the appropriate backing store.
1314 *
1315 * Arguments:
1316 * slab_hash The hash bucket we're freeing
1317 * hashsize The number of entries in that hash bucket
1318 *
1319 * Returns:
1320 * Nothing
1321 */
1322 static void
hash_free(struct uma_hash * hash)1323 hash_free(struct uma_hash *hash)
1324 {
1325 if (hash->uh_slab_hash == NULL)
1326 return;
1327 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1328 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1329 else
1330 free(hash->uh_slab_hash, M_UMAHASH);
1331 }
1332
1333 /*
1334 * Frees all outstanding items in a bucket
1335 *
1336 * Arguments:
1337 * zone The zone to free to, must be unlocked.
1338 * bucket The free/alloc bucket with items.
1339 *
1340 * Returns:
1341 * Nothing
1342 */
1343 static void
bucket_drain(uma_zone_t zone,uma_bucket_t bucket)1344 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1345 {
1346 int i;
1347
1348 if (bucket->ub_cnt == 0)
1349 return;
1350
1351 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1352 bucket->ub_seq != SMR_SEQ_INVALID) {
1353 smr_wait(zone->uz_smr, bucket->ub_seq);
1354 bucket->ub_seq = SMR_SEQ_INVALID;
1355 for (i = 0; i < bucket->ub_cnt; i++)
1356 item_dtor(zone, bucket->ub_bucket[i],
1357 zone->uz_size, NULL, SKIP_NONE);
1358 }
1359 if (zone->uz_fini)
1360 for (i = 0; i < bucket->ub_cnt; i++) {
1361 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
1362 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1363 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
1364 }
1365 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1366 if (zone->uz_max_items > 0)
1367 zone_free_limit(zone, bucket->ub_cnt);
1368 #ifdef INVARIANTS
1369 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1370 #endif
1371 bucket->ub_cnt = 0;
1372 }
1373
1374 /*
1375 * Drains the per cpu caches for a zone.
1376 *
1377 * NOTE: This may only be called while the zone is being torn down, and not
1378 * during normal operation. This is necessary in order that we do not have
1379 * to migrate CPUs to drain the per-CPU caches.
1380 *
1381 * Arguments:
1382 * zone The zone to drain, must be unlocked.
1383 *
1384 * Returns:
1385 * Nothing
1386 */
1387 static void
cache_drain(uma_zone_t zone)1388 cache_drain(uma_zone_t zone)
1389 {
1390 uma_cache_t cache;
1391 uma_bucket_t bucket;
1392 smr_seq_t seq;
1393 int cpu;
1394
1395 /*
1396 * XXX: It is safe to not lock the per-CPU caches, because we're
1397 * tearing down the zone anyway. I.e., there will be no further use
1398 * of the caches at this point.
1399 *
1400 * XXX: It would good to be able to assert that the zone is being
1401 * torn down to prevent improper use of cache_drain().
1402 */
1403 seq = SMR_SEQ_INVALID;
1404 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1405 seq = smr_advance(zone->uz_smr);
1406 CPU_FOREACH(cpu) {
1407 cache = &zone->uz_cpu[cpu];
1408 bucket = cache_bucket_unload_alloc(cache);
1409 if (bucket != NULL)
1410 bucket_free(zone, bucket, NULL);
1411 bucket = cache_bucket_unload_free(cache);
1412 if (bucket != NULL) {
1413 bucket->ub_seq = seq;
1414 bucket_free(zone, bucket, NULL);
1415 }
1416 bucket = cache_bucket_unload_cross(cache);
1417 if (bucket != NULL) {
1418 bucket->ub_seq = seq;
1419 bucket_free(zone, bucket, NULL);
1420 }
1421 }
1422 bucket_cache_reclaim(zone, true, UMA_ANYDOMAIN);
1423 }
1424
1425 static void
cache_shrink(uma_zone_t zone,void * unused)1426 cache_shrink(uma_zone_t zone, void *unused)
1427 {
1428
1429 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1430 return;
1431
1432 ZONE_LOCK(zone);
1433 zone->uz_bucket_size =
1434 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1435 ZONE_UNLOCK(zone);
1436 }
1437
1438 static void
cache_drain_safe_cpu(uma_zone_t zone,void * unused)1439 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1440 {
1441 uma_cache_t cache;
1442 uma_bucket_t b1, b2, b3;
1443 int domain;
1444
1445 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1446 return;
1447
1448 b1 = b2 = b3 = NULL;
1449 critical_enter();
1450 cache = &zone->uz_cpu[curcpu];
1451 domain = PCPU_GET(domain);
1452 b1 = cache_bucket_unload_alloc(cache);
1453
1454 /*
1455 * Don't flush SMR zone buckets. This leaves the zone without a
1456 * bucket and forces every free to synchronize().
1457 */
1458 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1459 b2 = cache_bucket_unload_free(cache);
1460 b3 = cache_bucket_unload_cross(cache);
1461 }
1462 critical_exit();
1463
1464 if (b1 != NULL)
1465 zone_free_bucket(zone, b1, NULL, domain, false);
1466 if (b2 != NULL)
1467 zone_free_bucket(zone, b2, NULL, domain, false);
1468 if (b3 != NULL) {
1469 /* Adjust the domain so it goes to zone_free_cross. */
1470 domain = (domain + 1) % vm_ndomains;
1471 zone_free_bucket(zone, b3, NULL, domain, false);
1472 }
1473 }
1474
1475 /*
1476 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1477 * This is an expensive call because it needs to bind to all CPUs
1478 * one by one and enter a critical section on each of them in order
1479 * to safely access their cache buckets.
1480 * Zone lock must not be held on call this function.
1481 */
1482 static void
pcpu_cache_drain_safe(uma_zone_t zone)1483 pcpu_cache_drain_safe(uma_zone_t zone)
1484 {
1485 int cpu;
1486
1487 /*
1488 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1489 */
1490 if (zone)
1491 cache_shrink(zone, NULL);
1492 else
1493 zone_foreach(cache_shrink, NULL);
1494
1495 CPU_FOREACH(cpu) {
1496 thread_lock(curthread);
1497 sched_bind(curthread, cpu);
1498 thread_unlock(curthread);
1499
1500 if (zone)
1501 cache_drain_safe_cpu(zone, NULL);
1502 else
1503 zone_foreach(cache_drain_safe_cpu, NULL);
1504 }
1505 thread_lock(curthread);
1506 sched_unbind(curthread);
1507 thread_unlock(curthread);
1508 }
1509
1510 /*
1511 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1512 * requested a drain, otherwise the per-domain caches are trimmed to either
1513 * estimated working set size.
1514 */
1515 static bool
bucket_cache_reclaim_domain(uma_zone_t zone,bool drain,bool trim,int domain)1516 bucket_cache_reclaim_domain(uma_zone_t zone, bool drain, bool trim, int domain)
1517 {
1518 uma_zone_domain_t zdom;
1519 uma_bucket_t bucket;
1520 long target;
1521 bool done = false;
1522
1523 /*
1524 * The cross bucket is partially filled and not part of
1525 * the item count. Reclaim it individually here.
1526 */
1527 zdom = ZDOM_GET(zone, domain);
1528 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1529 ZONE_CROSS_LOCK(zone);
1530 bucket = zdom->uzd_cross;
1531 zdom->uzd_cross = NULL;
1532 ZONE_CROSS_UNLOCK(zone);
1533 if (bucket != NULL)
1534 bucket_free(zone, bucket, NULL);
1535 }
1536
1537 /*
1538 * If we were asked to drain the zone, we are done only once
1539 * this bucket cache is empty. If trim, we reclaim items in
1540 * excess of the zone's estimated working set size. Multiple
1541 * consecutive calls will shrink the WSS and so reclaim more.
1542 * If neither drain nor trim, then voluntarily reclaim 1/4
1543 * (to reduce first spike) of items not used for a long time.
1544 */
1545 ZDOM_LOCK(zdom);
1546 zone_domain_update_wss(zdom);
1547 if (drain)
1548 target = 0;
1549 else if (trim)
1550 target = zdom->uzd_wss;
1551 else if (zdom->uzd_timin > 900 / UMA_TIMEOUT)
1552 target = zdom->uzd_nitems - zdom->uzd_limin / 4;
1553 else {
1554 ZDOM_UNLOCK(zdom);
1555 return (done);
1556 }
1557 while ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) != NULL &&
1558 zdom->uzd_nitems >= target + bucket->ub_cnt) {
1559 bucket = zone_fetch_bucket(zone, zdom, true);
1560 if (bucket == NULL)
1561 break;
1562 bucket_free(zone, bucket, NULL);
1563 done = true;
1564 ZDOM_LOCK(zdom);
1565 }
1566 ZDOM_UNLOCK(zdom);
1567 return (done);
1568 }
1569
1570 static void
bucket_cache_reclaim(uma_zone_t zone,bool drain,int domain)1571 bucket_cache_reclaim(uma_zone_t zone, bool drain, int domain)
1572 {
1573 int i;
1574
1575 /*
1576 * Shrink the zone bucket size to ensure that the per-CPU caches
1577 * don't grow too large.
1578 */
1579 if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1580 zone->uz_bucket_size--;
1581
1582 if (domain != UMA_ANYDOMAIN &&
1583 (zone->uz_flags & UMA_ZONE_ROUNDROBIN) == 0) {
1584 bucket_cache_reclaim_domain(zone, drain, true, domain);
1585 } else {
1586 for (i = 0; i < vm_ndomains; i++)
1587 bucket_cache_reclaim_domain(zone, drain, true, i);
1588 }
1589 }
1590
1591 static void
keg_free_slab(uma_keg_t keg,uma_slab_t slab,int start)1592 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1593 {
1594 uint8_t *mem;
1595 size_t size;
1596 int i;
1597 uint8_t flags;
1598
1599 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1600 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1601
1602 mem = slab_data(slab, keg);
1603 size = PAGE_SIZE * keg->uk_ppera;
1604
1605 kasan_mark_slab_valid(keg, mem);
1606 if (keg->uk_fini != NULL) {
1607 for (i = start - 1; i > -1; i--)
1608 #ifdef INVARIANTS
1609 /*
1610 * trash_fini implies that dtor was trash_dtor. trash_fini
1611 * would check that memory hasn't been modified since free,
1612 * which executed trash_dtor.
1613 * That's why we need to run uma_dbg_kskip() check here,
1614 * albeit we don't make skip check for other init/fini
1615 * invocations.
1616 */
1617 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1618 keg->uk_fini != trash_fini)
1619 #endif
1620 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1621 }
1622 flags = slab->us_flags;
1623 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1624 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1625 NULL, SKIP_NONE);
1626 }
1627 keg->uk_freef(mem, size, flags);
1628 uma_total_dec(size);
1629 }
1630
1631 static void
keg_drain_domain(uma_keg_t keg,int domain)1632 keg_drain_domain(uma_keg_t keg, int domain)
1633 {
1634 struct slabhead freeslabs;
1635 uma_domain_t dom;
1636 uma_slab_t slab, tmp;
1637 uint32_t i, stofree, stokeep, partial;
1638
1639 dom = &keg->uk_domain[domain];
1640 LIST_INIT(&freeslabs);
1641
1642 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1643 keg->uk_name, keg, domain, dom->ud_free_items);
1644
1645 KEG_LOCK(keg, domain);
1646
1647 /*
1648 * Are the free items in partially allocated slabs sufficient to meet
1649 * the reserve? If not, compute the number of fully free slabs that must
1650 * be kept.
1651 */
1652 partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers;
1653 if (partial < keg->uk_reserve) {
1654 stokeep = min(dom->ud_free_slabs,
1655 howmany(keg->uk_reserve - partial, keg->uk_ipers));
1656 } else {
1657 stokeep = 0;
1658 }
1659 stofree = dom->ud_free_slabs - stokeep;
1660
1661 /*
1662 * Partition the free slabs into two sets: those that must be kept in
1663 * order to maintain the reserve, and those that may be released back to
1664 * the system. Since one set may be much larger than the other,
1665 * populate the smaller of the two sets and swap them if necessary.
1666 */
1667 for (i = min(stofree, stokeep); i > 0; i--) {
1668 slab = LIST_FIRST(&dom->ud_free_slab);
1669 LIST_REMOVE(slab, us_link);
1670 LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1671 }
1672 if (stofree > stokeep)
1673 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1674
1675 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1676 LIST_FOREACH(slab, &freeslabs, us_link)
1677 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1678 }
1679 dom->ud_free_items -= stofree * keg->uk_ipers;
1680 dom->ud_free_slabs -= stofree;
1681 dom->ud_pages -= stofree * keg->uk_ppera;
1682 KEG_UNLOCK(keg, domain);
1683
1684 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1685 keg_free_slab(keg, slab, keg->uk_ipers);
1686 }
1687
1688 /*
1689 * Frees pages from a keg back to the system. This is done on demand from
1690 * the pageout daemon.
1691 *
1692 * Returns nothing.
1693 */
1694 static void
keg_drain(uma_keg_t keg,int domain)1695 keg_drain(uma_keg_t keg, int domain)
1696 {
1697 int i;
1698
1699 if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0)
1700 return;
1701 if (domain != UMA_ANYDOMAIN) {
1702 keg_drain_domain(keg, domain);
1703 } else {
1704 for (i = 0; i < vm_ndomains; i++)
1705 keg_drain_domain(keg, i);
1706 }
1707 }
1708
1709 static void
zone_reclaim(uma_zone_t zone,int domain,int waitok,bool drain)1710 zone_reclaim(uma_zone_t zone, int domain, int waitok, bool drain)
1711 {
1712 /*
1713 * Count active reclaim operations in order to interlock with
1714 * zone_dtor(), which removes the zone from global lists before
1715 * attempting to reclaim items itself.
1716 *
1717 * The zone may be destroyed while sleeping, so only zone_dtor() should
1718 * specify M_WAITOK.
1719 */
1720 ZONE_LOCK(zone);
1721 if (waitok == M_WAITOK) {
1722 while (zone->uz_reclaimers > 0)
1723 msleep(zone, ZONE_LOCKPTR(zone), PVM, "zonedrain", 1);
1724 }
1725 zone->uz_reclaimers++;
1726 ZONE_UNLOCK(zone);
1727 bucket_cache_reclaim(zone, drain, domain);
1728
1729 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1730 keg_drain(zone->uz_keg, domain);
1731 ZONE_LOCK(zone);
1732 zone->uz_reclaimers--;
1733 if (zone->uz_reclaimers == 0)
1734 wakeup(zone);
1735 ZONE_UNLOCK(zone);
1736 }
1737
1738 /*
1739 * Allocate a new slab for a keg and inserts it into the partial slab list.
1740 * The keg should be unlocked on entry. If the allocation succeeds it will
1741 * be locked on return.
1742 *
1743 * Arguments:
1744 * flags Wait flags for the item initialization routine
1745 * aflags Wait flags for the slab allocation
1746 *
1747 * Returns:
1748 * The slab that was allocated or NULL if there is no memory and the
1749 * caller specified M_NOWAIT.
1750 */
1751 static uma_slab_t
keg_alloc_slab(uma_keg_t keg,uma_zone_t zone,int domain,int flags,int aflags)1752 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1753 int aflags)
1754 {
1755 uma_domain_t dom;
1756 uma_slab_t slab;
1757 unsigned long size;
1758 uint8_t *mem;
1759 uint8_t sflags;
1760 int i;
1761
1762 TSENTER();
1763
1764 KASSERT(domain >= 0 && domain < vm_ndomains,
1765 ("keg_alloc_slab: domain %d out of range", domain));
1766
1767 slab = NULL;
1768 mem = NULL;
1769 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1770 uma_hash_slab_t hslab;
1771 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1772 domain, aflags);
1773 if (hslab == NULL)
1774 goto fail;
1775 slab = &hslab->uhs_slab;
1776 }
1777
1778 /*
1779 * This reproduces the old vm_zone behavior of zero filling pages the
1780 * first time they are added to a zone.
1781 *
1782 * Malloced items are zeroed in uma_zalloc.
1783 */
1784
1785 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1786 aflags |= M_ZERO;
1787 else
1788 aflags &= ~M_ZERO;
1789
1790 if (keg->uk_flags & UMA_ZONE_NODUMP)
1791 aflags |= M_NODUMP;
1792
1793 if (keg->uk_flags & UMA_ZONE_NOFREE)
1794 aflags |= M_NEVERFREED;
1795
1796 /* zone is passed for legacy reasons. */
1797 size = keg->uk_ppera * PAGE_SIZE;
1798 mem = keg->uk_allocf(zone, size, domain, &sflags, aflags);
1799 if (mem == NULL) {
1800 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1801 zone_free_item(slabzone(keg->uk_ipers),
1802 slab_tohashslab(slab), NULL, SKIP_NONE);
1803 goto fail;
1804 }
1805 uma_total_inc(size);
1806
1807 /* For HASH zones all pages go to the same uma_domain. */
1808 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1809 domain = 0;
1810
1811 kmsan_mark(mem, size,
1812 (aflags & M_ZERO) != 0 ? KMSAN_STATE_INITED : KMSAN_STATE_UNINIT);
1813
1814 /* Point the slab into the allocated memory */
1815 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1816 slab = (uma_slab_t)(mem + keg->uk_pgoff);
1817 else
1818 slab_tohashslab(slab)->uhs_data = mem;
1819
1820 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1821 for (i = 0; i < keg->uk_ppera; i++)
1822 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1823 zone, slab);
1824
1825 slab->us_freecount = keg->uk_ipers;
1826 slab->us_flags = sflags;
1827 slab->us_domain = domain;
1828
1829 BIT_FILL(keg->uk_ipers, &slab->us_free);
1830 #ifdef INVARIANTS
1831 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1832 #endif
1833
1834 if (keg->uk_init != NULL) {
1835 for (i = 0; i < keg->uk_ipers; i++)
1836 if (keg->uk_init(slab_item(slab, keg, i),
1837 keg->uk_size, flags) != 0)
1838 break;
1839 if (i != keg->uk_ipers) {
1840 keg_free_slab(keg, slab, i);
1841 goto fail;
1842 }
1843 }
1844 kasan_mark_slab_invalid(keg, mem);
1845 KEG_LOCK(keg, domain);
1846
1847 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1848 slab, keg->uk_name, keg);
1849
1850 if (keg->uk_flags & UMA_ZFLAG_HASH)
1851 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1852
1853 /*
1854 * If we got a slab here it's safe to mark it partially used
1855 * and return. We assume that the caller is going to remove
1856 * at least one item.
1857 */
1858 dom = &keg->uk_domain[domain];
1859 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1860 dom->ud_pages += keg->uk_ppera;
1861 dom->ud_free_items += keg->uk_ipers;
1862
1863 TSEXIT();
1864 return (slab);
1865
1866 fail:
1867 return (NULL);
1868 }
1869
1870 /*
1871 * This function is intended to be used early on in place of page_alloc(). It
1872 * performs contiguous physical memory allocations and uses a bump allocator for
1873 * KVA, so is usable before the kernel map is initialized.
1874 */
1875 static void *
startup_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)1876 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1877 int wait)
1878 {
1879 vm_paddr_t pa;
1880 vm_page_t m;
1881 int i, pages;
1882
1883 pages = howmany(bytes, PAGE_SIZE);
1884 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1885
1886 *pflag = UMA_SLAB_BOOT;
1887 m = vm_page_alloc_noobj_contig_domain(domain, malloc2vm_flags(wait) |
1888 VM_ALLOC_WIRED, pages, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
1889 VM_MEMATTR_DEFAULT);
1890 if (m == NULL)
1891 return (NULL);
1892
1893 pa = VM_PAGE_TO_PHYS(m);
1894 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1895 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING
1896 if ((wait & M_NODUMP) == 0)
1897 dump_add_page(pa);
1898 #endif
1899 }
1900
1901 /* Allocate KVA and indirectly advance bootmem. */
1902 return ((void *)pmap_map(&bootmem, m->phys_addr,
1903 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE));
1904 }
1905
1906 static void
startup_free(void * mem,vm_size_t bytes)1907 startup_free(void *mem, vm_size_t bytes)
1908 {
1909 vm_offset_t va;
1910 vm_page_t m;
1911
1912 va = (vm_offset_t)mem;
1913 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1914
1915 /*
1916 * startup_alloc() returns direct-mapped slabs on some platforms. Avoid
1917 * unmapping ranges of the direct map.
1918 */
1919 if (va >= bootstart && va + bytes <= bootmem)
1920 pmap_remove(kernel_pmap, va, va + bytes);
1921 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1922 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING
1923 dump_drop_page(VM_PAGE_TO_PHYS(m));
1924 #endif
1925 vm_page_unwire_noq(m);
1926 vm_page_free(m);
1927 }
1928 }
1929
1930 /*
1931 * Allocates a number of pages from the system
1932 *
1933 * Arguments:
1934 * bytes The number of bytes requested
1935 * wait Shall we wait?
1936 *
1937 * Returns:
1938 * A pointer to the alloced memory or possibly
1939 * NULL if M_NOWAIT is set.
1940 */
1941 static void *
page_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)1942 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1943 int wait)
1944 {
1945 void *p; /* Returned page */
1946
1947 *pflag = UMA_SLAB_KERNEL;
1948 p = kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1949
1950 return (p);
1951 }
1952
1953 static void *
pcpu_page_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)1954 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1955 int wait)
1956 {
1957 struct pglist alloctail;
1958 vm_offset_t addr, zkva;
1959 int cpu, flags;
1960 vm_page_t p, p_next;
1961 #ifdef NUMA
1962 struct pcpu *pc;
1963 #endif
1964
1965 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1966
1967 TAILQ_INIT(&alloctail);
1968 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | malloc2vm_flags(wait);
1969 *pflag = UMA_SLAB_KERNEL;
1970 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1971 if (CPU_ABSENT(cpu)) {
1972 p = vm_page_alloc_noobj(flags);
1973 } else {
1974 #ifndef NUMA
1975 p = vm_page_alloc_noobj(flags);
1976 #else
1977 pc = pcpu_find(cpu);
1978 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1979 p = NULL;
1980 else
1981 p = vm_page_alloc_noobj_domain(pc->pc_domain,
1982 flags);
1983 if (__predict_false(p == NULL))
1984 p = vm_page_alloc_noobj(flags);
1985 #endif
1986 }
1987 if (__predict_false(p == NULL))
1988 goto fail;
1989 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1990 }
1991 if ((addr = kva_alloc(bytes)) == 0)
1992 goto fail;
1993 zkva = addr;
1994 TAILQ_FOREACH(p, &alloctail, listq) {
1995 pmap_qenter(zkva, &p, 1);
1996 zkva += PAGE_SIZE;
1997 }
1998 return ((void*)addr);
1999 fail:
2000 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2001 vm_page_unwire_noq(p);
2002 vm_page_free(p);
2003 }
2004 return (NULL);
2005 }
2006
2007 /*
2008 * Allocates a number of pages not belonging to a VM object
2009 *
2010 * Arguments:
2011 * bytes The number of bytes requested
2012 * wait Shall we wait?
2013 *
2014 * Returns:
2015 * A pointer to the alloced memory or possibly
2016 * NULL if M_NOWAIT is set.
2017 */
2018 static void *
noobj_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * flags,int wait)2019 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
2020 int wait)
2021 {
2022 TAILQ_HEAD(, vm_page) alloctail;
2023 u_long npages;
2024 vm_offset_t retkva, zkva;
2025 vm_page_t p, p_next;
2026 uma_keg_t keg;
2027 int req;
2028
2029 TAILQ_INIT(&alloctail);
2030 keg = zone->uz_keg;
2031 req = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
2032 if ((wait & M_WAITOK) != 0)
2033 req |= VM_ALLOC_WAITOK;
2034
2035 npages = howmany(bytes, PAGE_SIZE);
2036 while (npages > 0) {
2037 p = vm_page_alloc_noobj_domain(domain, req);
2038 if (p != NULL) {
2039 /*
2040 * Since the page does not belong to an object, its
2041 * listq is unused.
2042 */
2043 TAILQ_INSERT_TAIL(&alloctail, p, listq);
2044 npages--;
2045 continue;
2046 }
2047 /*
2048 * Page allocation failed, free intermediate pages and
2049 * exit.
2050 */
2051 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2052 vm_page_unwire_noq(p);
2053 vm_page_free(p);
2054 }
2055 return (NULL);
2056 }
2057 *flags = UMA_SLAB_PRIV;
2058 zkva = keg->uk_kva +
2059 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
2060 retkva = zkva;
2061 TAILQ_FOREACH(p, &alloctail, listq) {
2062 pmap_qenter(zkva, &p, 1);
2063 zkva += PAGE_SIZE;
2064 }
2065
2066 return ((void *)retkva);
2067 }
2068
2069 /*
2070 * Allocate physically contiguous pages.
2071 */
2072 static void *
contig_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)2073 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
2074 int wait)
2075 {
2076
2077 *pflag = UMA_SLAB_KERNEL;
2078 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
2079 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
2080 }
2081
2082 #if defined(UMA_USE_DMAP) && !defined(UMA_MD_SMALL_ALLOC)
2083 void *
uma_small_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * flags,int wait)2084 uma_small_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
2085 int wait)
2086 {
2087 vm_page_t m;
2088 vm_paddr_t pa;
2089 void *va;
2090
2091 *flags = UMA_SLAB_PRIV;
2092 m = vm_page_alloc_noobj_domain(domain,
2093 malloc2vm_flags(wait) | VM_ALLOC_WIRED);
2094 if (m == NULL)
2095 return (NULL);
2096 pa = m->phys_addr;
2097 if ((wait & M_NODUMP) == 0)
2098 dump_add_page(pa);
2099 va = (void *)PHYS_TO_DMAP(pa);
2100 return (va);
2101 }
2102 #endif
2103
2104 /*
2105 * Frees a number of pages to the system
2106 *
2107 * Arguments:
2108 * mem A pointer to the memory to be freed
2109 * size The size of the memory being freed
2110 * flags The original p->us_flags field
2111 *
2112 * Returns:
2113 * Nothing
2114 */
2115 static void
page_free(void * mem,vm_size_t size,uint8_t flags)2116 page_free(void *mem, vm_size_t size, uint8_t flags)
2117 {
2118
2119 if ((flags & UMA_SLAB_BOOT) != 0) {
2120 startup_free(mem, size);
2121 return;
2122 }
2123
2124 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
2125 ("UMA: page_free used with invalid flags %x", flags));
2126
2127 kmem_free(mem, size);
2128 }
2129
2130 /*
2131 * Frees pcpu zone allocations
2132 *
2133 * Arguments:
2134 * mem A pointer to the memory to be freed
2135 * size The size of the memory being freed
2136 * flags The original p->us_flags field
2137 *
2138 * Returns:
2139 * Nothing
2140 */
2141 static void
pcpu_page_free(void * mem,vm_size_t size,uint8_t flags)2142 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
2143 {
2144 vm_offset_t sva, curva;
2145 vm_paddr_t paddr;
2146 vm_page_t m;
2147
2148 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
2149
2150 if ((flags & UMA_SLAB_BOOT) != 0) {
2151 startup_free(mem, size);
2152 return;
2153 }
2154
2155 sva = (vm_offset_t)mem;
2156 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
2157 paddr = pmap_kextract(curva);
2158 m = PHYS_TO_VM_PAGE(paddr);
2159 vm_page_unwire_noq(m);
2160 vm_page_free(m);
2161 }
2162 pmap_qremove(sva, size >> PAGE_SHIFT);
2163 kva_free(sva, size);
2164 }
2165
2166 #if defined(UMA_USE_DMAP) && !defined(UMA_MD_SMALL_ALLOC)
2167 void
uma_small_free(void * mem,vm_size_t size,uint8_t flags)2168 uma_small_free(void *mem, vm_size_t size, uint8_t flags)
2169 {
2170 vm_page_t m;
2171 vm_paddr_t pa;
2172
2173 pa = DMAP_TO_PHYS((vm_offset_t)mem);
2174 dump_drop_page(pa);
2175 m = PHYS_TO_VM_PAGE(pa);
2176 vm_page_unwire_noq(m);
2177 vm_page_free(m);
2178 }
2179 #endif
2180
2181 /*
2182 * Zero fill initializer
2183 *
2184 * Arguments/Returns follow uma_init specifications
2185 */
2186 static int
zero_init(void * mem,int size,int flags)2187 zero_init(void *mem, int size, int flags)
2188 {
2189 bzero(mem, size);
2190 return (0);
2191 }
2192
2193 #ifdef INVARIANTS
2194 static struct noslabbits *
slab_dbg_bits(uma_slab_t slab,uma_keg_t keg)2195 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
2196 {
2197
2198 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
2199 }
2200 #endif
2201
2202 /*
2203 * Actual size of embedded struct slab (!OFFPAGE).
2204 */
2205 static size_t
slab_sizeof(int nitems)2206 slab_sizeof(int nitems)
2207 {
2208 size_t s;
2209
2210 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
2211 return (roundup(s, UMA_ALIGN_PTR + 1));
2212 }
2213
2214 #define UMA_FIXPT_SHIFT 31
2215 #define UMA_FRAC_FIXPT(n, d) \
2216 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
2217 #define UMA_FIXPT_PCT(f) \
2218 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
2219 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
2220 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
2221
2222 /*
2223 * Compute the number of items that will fit in a slab. If hdr is true, the
2224 * item count may be limited to provide space in the slab for an inline slab
2225 * header. Otherwise, all slab space will be provided for item storage.
2226 */
2227 static u_int
slab_ipers_hdr(u_int size,u_int rsize,u_int slabsize,bool hdr)2228 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
2229 {
2230 u_int ipers;
2231 u_int padpi;
2232
2233 /* The padding between items is not needed after the last item. */
2234 padpi = rsize - size;
2235
2236 if (hdr) {
2237 /*
2238 * Start with the maximum item count and remove items until
2239 * the slab header first alongside the allocatable memory.
2240 */
2241 for (ipers = MIN(SLAB_MAX_SETSIZE,
2242 (slabsize + padpi - slab_sizeof(1)) / rsize);
2243 ipers > 0 &&
2244 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2245 ipers--)
2246 continue;
2247 } else {
2248 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2249 }
2250
2251 return (ipers);
2252 }
2253
2254 struct keg_layout_result {
2255 u_int format;
2256 u_int slabsize;
2257 u_int ipers;
2258 u_int eff;
2259 };
2260
2261 static void
keg_layout_one(uma_keg_t keg,u_int rsize,u_int slabsize,u_int fmt,struct keg_layout_result * kl)2262 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2263 struct keg_layout_result *kl)
2264 {
2265 u_int total;
2266
2267 kl->format = fmt;
2268 kl->slabsize = slabsize;
2269
2270 /* Handle INTERNAL as inline with an extra page. */
2271 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2272 kl->format &= ~UMA_ZFLAG_INTERNAL;
2273 kl->slabsize += PAGE_SIZE;
2274 }
2275
2276 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2277 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2278
2279 /* Account for memory used by an offpage slab header. */
2280 total = kl->slabsize;
2281 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2282 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2283
2284 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2285 }
2286
2287 /*
2288 * Determine the format of a uma keg. This determines where the slab header
2289 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2290 *
2291 * Arguments
2292 * keg The zone we should initialize
2293 *
2294 * Returns
2295 * Nothing
2296 */
2297 static void
keg_layout(uma_keg_t keg)2298 keg_layout(uma_keg_t keg)
2299 {
2300 struct keg_layout_result kl = {}, kl_tmp;
2301 u_int fmts[2];
2302 u_int alignsize;
2303 u_int nfmt;
2304 u_int pages;
2305 u_int rsize;
2306 u_int slabsize;
2307 u_int i, j;
2308
2309 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2310 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2311 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2312 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2313 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2314 PRINT_UMA_ZFLAGS));
2315 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2316 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2317 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2318 PRINT_UMA_ZFLAGS));
2319
2320 alignsize = keg->uk_align + 1;
2321 #ifdef KASAN
2322 /*
2323 * ASAN requires that each allocation be aligned to the shadow map
2324 * scale factor.
2325 */
2326 if (alignsize < KASAN_SHADOW_SCALE)
2327 alignsize = KASAN_SHADOW_SCALE;
2328 #endif
2329
2330 /*
2331 * Calculate the size of each allocation (rsize) according to
2332 * alignment. If the requested size is smaller than we have
2333 * allocation bits for we round it up.
2334 */
2335 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2336 rsize = roundup2(rsize, alignsize);
2337
2338 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2339 /*
2340 * We want one item to start on every align boundary in a page.
2341 * To do this we will span pages. We will also extend the item
2342 * by the size of align if it is an even multiple of align.
2343 * Otherwise, it would fall on the same boundary every time.
2344 */
2345 if ((rsize & alignsize) == 0)
2346 rsize += alignsize;
2347 slabsize = rsize * (PAGE_SIZE / alignsize);
2348 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2349 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2350 slabsize = round_page(slabsize);
2351 } else {
2352 /*
2353 * Start with a slab size of as many pages as it takes to
2354 * represent a single item. We will try to fit as many
2355 * additional items into the slab as possible.
2356 */
2357 slabsize = round_page(keg->uk_size);
2358 }
2359
2360 /* Build a list of all of the available formats for this keg. */
2361 nfmt = 0;
2362
2363 /* Evaluate an inline slab layout. */
2364 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2365 fmts[nfmt++] = 0;
2366
2367 /* TODO: vm_page-embedded slab. */
2368
2369 /*
2370 * We can't do OFFPAGE if we're internal or if we've been
2371 * asked to not go to the VM for buckets. If we do this we
2372 * may end up going to the VM for slabs which we do not want
2373 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2374 * In those cases, evaluate a pseudo-format called INTERNAL
2375 * which has an inline slab header and one extra page to
2376 * guarantee that it fits.
2377 *
2378 * Otherwise, see if using an OFFPAGE slab will improve our
2379 * efficiency.
2380 */
2381 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2382 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2383 else
2384 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2385
2386 /*
2387 * Choose a slab size and format which satisfy the minimum efficiency.
2388 * Prefer the smallest slab size that meets the constraints.
2389 *
2390 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2391 * for small items (up to PAGE_SIZE), the iteration increment is one
2392 * page; and for large items, the increment is one item.
2393 */
2394 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2395 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2396 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2397 rsize, i));
2398 for ( ; ; i++) {
2399 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2400 round_page(rsize * (i - 1) + keg->uk_size);
2401
2402 for (j = 0; j < nfmt; j++) {
2403 /* Only if we have no viable format yet. */
2404 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2405 kl.ipers > 0)
2406 continue;
2407
2408 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2409 if (kl_tmp.eff <= kl.eff)
2410 continue;
2411
2412 kl = kl_tmp;
2413
2414 CTR6(KTR_UMA, "keg %s layout: format %#x "
2415 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2416 keg->uk_name, kl.format, kl.ipers, rsize,
2417 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2418
2419 /* Stop when we reach the minimum efficiency. */
2420 if (kl.eff >= UMA_MIN_EFF)
2421 break;
2422 }
2423
2424 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2425 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2426 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2427 break;
2428 }
2429
2430 pages = atop(kl.slabsize);
2431 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2432 pages *= mp_maxid + 1;
2433
2434 keg->uk_rsize = rsize;
2435 keg->uk_ipers = kl.ipers;
2436 keg->uk_ppera = pages;
2437 keg->uk_flags |= kl.format;
2438
2439 /*
2440 * How do we find the slab header if it is offpage or if not all item
2441 * start addresses are in the same page? We could solve the latter
2442 * case with vaddr alignment, but we don't.
2443 */
2444 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2445 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2446 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2447 keg->uk_flags |= UMA_ZFLAG_HASH;
2448 else
2449 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2450 }
2451
2452 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2453 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2454 pages);
2455 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2456 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2457 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2458 keg->uk_ipers, pages));
2459 }
2460
2461 /*
2462 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2463 * the keg onto the global keg list.
2464 *
2465 * Arguments/Returns follow uma_ctor specifications
2466 * udata Actually uma_kctor_args
2467 */
2468 static int
keg_ctor(void * mem,int size,void * udata,int flags)2469 keg_ctor(void *mem, int size, void *udata, int flags)
2470 {
2471 struct uma_kctor_args *arg = udata;
2472 uma_keg_t keg = mem;
2473 uma_zone_t zone;
2474 int i;
2475
2476 bzero(keg, size);
2477 keg->uk_size = arg->size;
2478 keg->uk_init = arg->uminit;
2479 keg->uk_fini = arg->fini;
2480 keg->uk_align = arg->align;
2481 keg->uk_reserve = 0;
2482 keg->uk_flags = arg->flags;
2483
2484 /*
2485 * We use a global round-robin policy by default. Zones with
2486 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2487 * case the iterator is never run.
2488 */
2489 keg->uk_dr.dr_policy = DOMAINSET_RR();
2490 keg->uk_dr.dr_iter = 0;
2491
2492 /*
2493 * The primary zone is passed to us at keg-creation time.
2494 */
2495 zone = arg->zone;
2496 keg->uk_name = zone->uz_name;
2497
2498 if (arg->flags & UMA_ZONE_ZINIT)
2499 keg->uk_init = zero_init;
2500
2501 if (arg->flags & UMA_ZONE_MALLOC)
2502 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2503
2504 #ifndef SMP
2505 keg->uk_flags &= ~UMA_ZONE_PCPU;
2506 #endif
2507
2508 keg_layout(keg);
2509
2510 /*
2511 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2512 * work on. Use round-robin for everything else.
2513 *
2514 * Zones may override the default by specifying either.
2515 */
2516 #ifdef NUMA
2517 if ((keg->uk_flags &
2518 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2519 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2520 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2521 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2522 #endif
2523
2524 /*
2525 * If we haven't booted yet we need allocations to go through the
2526 * startup cache until the vm is ready.
2527 */
2528 #ifdef UMA_USE_DMAP
2529 if (keg->uk_ppera == 1)
2530 keg->uk_allocf = uma_small_alloc;
2531 else
2532 #endif
2533 if (booted < BOOT_KVA)
2534 keg->uk_allocf = startup_alloc;
2535 else if (keg->uk_flags & UMA_ZONE_PCPU)
2536 keg->uk_allocf = pcpu_page_alloc;
2537 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2538 keg->uk_allocf = contig_alloc;
2539 else
2540 keg->uk_allocf = page_alloc;
2541 #ifdef UMA_USE_DMAP
2542 if (keg->uk_ppera == 1)
2543 keg->uk_freef = uma_small_free;
2544 else
2545 #endif
2546 if (keg->uk_flags & UMA_ZONE_PCPU)
2547 keg->uk_freef = pcpu_page_free;
2548 else
2549 keg->uk_freef = page_free;
2550
2551 /*
2552 * Initialize keg's locks.
2553 */
2554 for (i = 0; i < vm_ndomains; i++)
2555 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2556
2557 /*
2558 * If we're putting the slab header in the actual page we need to
2559 * figure out where in each page it goes. See slab_sizeof
2560 * definition.
2561 */
2562 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2563 size_t shsize;
2564
2565 shsize = slab_sizeof(keg->uk_ipers);
2566 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2567 /*
2568 * The only way the following is possible is if with our
2569 * UMA_ALIGN_PTR adjustments we are now bigger than
2570 * UMA_SLAB_SIZE. I haven't checked whether this is
2571 * mathematically possible for all cases, so we make
2572 * sure here anyway.
2573 */
2574 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2575 ("zone %s ipers %d rsize %d size %d slab won't fit",
2576 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2577 }
2578
2579 if (keg->uk_flags & UMA_ZFLAG_HASH)
2580 hash_alloc(&keg->uk_hash, 0);
2581
2582 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2583
2584 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2585
2586 rw_wlock(&uma_rwlock);
2587 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2588 rw_wunlock(&uma_rwlock);
2589 return (0);
2590 }
2591
2592 static void
zone_kva_available(uma_zone_t zone,void * unused)2593 zone_kva_available(uma_zone_t zone, void *unused)
2594 {
2595 uma_keg_t keg;
2596
2597 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2598 return;
2599 KEG_GET(zone, keg);
2600
2601 if (keg->uk_allocf == startup_alloc) {
2602 /* Switch to the real allocator. */
2603 if (keg->uk_flags & UMA_ZONE_PCPU)
2604 keg->uk_allocf = pcpu_page_alloc;
2605 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2606 keg->uk_ppera > 1)
2607 keg->uk_allocf = contig_alloc;
2608 else
2609 keg->uk_allocf = page_alloc;
2610 }
2611 }
2612
2613 static void
zone_alloc_counters(uma_zone_t zone,void * unused)2614 zone_alloc_counters(uma_zone_t zone, void *unused)
2615 {
2616
2617 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2618 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2619 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2620 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2621 }
2622
2623 static void
zone_alloc_sysctl(uma_zone_t zone,void * unused)2624 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2625 {
2626 uma_zone_domain_t zdom;
2627 uma_domain_t dom;
2628 uma_keg_t keg;
2629 struct sysctl_oid *oid, *domainoid;
2630 int domains, i, cnt;
2631 static const char *nokeg = "cache zone";
2632 char *c;
2633
2634 /*
2635 * Make a sysctl safe copy of the zone name by removing
2636 * any special characters and handling dups by appending
2637 * an index.
2638 */
2639 if (zone->uz_namecnt != 0) {
2640 /* Count the number of decimal digits and '_' separator. */
2641 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2642 cnt /= 10;
2643 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2644 M_UMA, M_WAITOK);
2645 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2646 zone->uz_namecnt);
2647 } else
2648 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2649 for (c = zone->uz_ctlname; *c != '\0'; c++)
2650 if (strchr("./\\ -", *c) != NULL)
2651 *c = '_';
2652
2653 /*
2654 * Basic parameters at the root.
2655 */
2656 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2657 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2658 oid = zone->uz_oid;
2659 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2660 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2661 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2662 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2663 zone, 0, sysctl_handle_uma_zone_flags, "A",
2664 "Allocator configuration flags");
2665 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2666 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2667 "Desired per-cpu cache size");
2668 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2669 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2670 "Maximum allowed per-cpu cache size");
2671
2672 /*
2673 * keg if present.
2674 */
2675 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2676 domains = vm_ndomains;
2677 else
2678 domains = 1;
2679 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2680 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2681 keg = zone->uz_keg;
2682 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2683 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2684 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2685 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2686 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2687 "Real object size with alignment");
2688 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2689 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2690 "pages per-slab allocation");
2691 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2692 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2693 "items available per-slab");
2694 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2695 "align", CTLFLAG_RD, &keg->uk_align, 0,
2696 "item alignment mask");
2697 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2698 "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2699 "number of reserved items");
2700 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2701 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2702 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2703 "Slab utilization (100 - internal fragmentation %)");
2704 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2705 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2706 for (i = 0; i < domains; i++) {
2707 dom = &keg->uk_domain[i];
2708 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2709 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2710 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2711 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2712 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2713 "Total pages currently allocated from VM");
2714 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2715 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2716 "Items free in the slab layer");
2717 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2718 "free_slabs", CTLFLAG_RD, &dom->ud_free_slabs, 0,
2719 "Unused slabs");
2720 }
2721 } else
2722 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2723 "name", CTLFLAG_RD, nokeg, "Keg name");
2724
2725 /*
2726 * Information about zone limits.
2727 */
2728 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2729 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2730 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2731 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2732 zone, 0, sysctl_handle_uma_zone_items, "QU",
2733 "Current number of allocated items if limit is set");
2734 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2735 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2736 "Maximum number of allocated and cached items");
2737 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2738 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2739 "Number of threads sleeping at limit");
2740 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2741 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2742 "Total zone limit sleeps");
2743 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2744 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2745 "Maximum number of items in each domain's bucket cache");
2746
2747 /*
2748 * Per-domain zone information.
2749 */
2750 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2751 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2752 for (i = 0; i < domains; i++) {
2753 zdom = ZDOM_GET(zone, i);
2754 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2755 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2756 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2757 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2758 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2759 "number of items in this domain");
2760 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2761 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2762 "maximum item count in this period");
2763 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2764 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2765 "minimum item count in this period");
2766 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2767 "bimin", CTLFLAG_RD, &zdom->uzd_bimin,
2768 "Minimum item count in this batch");
2769 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2770 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2771 "Working set size");
2772 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2773 "limin", CTLFLAG_RD, &zdom->uzd_limin,
2774 "Long time minimum item count");
2775 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2776 "timin", CTLFLAG_RD, &zdom->uzd_timin, 0,
2777 "Time since zero long time minimum item count");
2778 }
2779
2780 /*
2781 * General statistics.
2782 */
2783 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2784 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2785 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2786 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2787 zone, 1, sysctl_handle_uma_zone_cur, "I",
2788 "Current number of allocated items");
2789 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2790 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2791 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2792 "Total allocation calls");
2793 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2794 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2795 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2796 "Total free calls");
2797 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2798 "fails", CTLFLAG_RD, &zone->uz_fails,
2799 "Number of allocation failures");
2800 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2801 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2802 "Free calls from the wrong domain");
2803 }
2804
2805 struct uma_zone_count {
2806 const char *name;
2807 int count;
2808 };
2809
2810 static void
zone_count(uma_zone_t zone,void * arg)2811 zone_count(uma_zone_t zone, void *arg)
2812 {
2813 struct uma_zone_count *cnt;
2814
2815 cnt = arg;
2816 /*
2817 * Some zones are rapidly created with identical names and
2818 * destroyed out of order. This can lead to gaps in the count.
2819 * Use one greater than the maximum observed for this name.
2820 */
2821 if (strcmp(zone->uz_name, cnt->name) == 0)
2822 cnt->count = MAX(cnt->count,
2823 zone->uz_namecnt + 1);
2824 }
2825
2826 static void
zone_update_caches(uma_zone_t zone)2827 zone_update_caches(uma_zone_t zone)
2828 {
2829 int i;
2830
2831 for (i = 0; i <= mp_maxid; i++) {
2832 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2833 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2834 }
2835 }
2836
2837 /*
2838 * Zone header ctor. This initializes all fields, locks, etc.
2839 *
2840 * Arguments/Returns follow uma_ctor specifications
2841 * udata Actually uma_zctor_args
2842 */
2843 static int
zone_ctor(void * mem,int size,void * udata,int flags)2844 zone_ctor(void *mem, int size, void *udata, int flags)
2845 {
2846 struct uma_zone_count cnt;
2847 struct uma_zctor_args *arg = udata;
2848 uma_zone_domain_t zdom;
2849 uma_zone_t zone = mem;
2850 uma_zone_t z;
2851 uma_keg_t keg;
2852 int i;
2853
2854 bzero(zone, size);
2855 zone->uz_name = arg->name;
2856 zone->uz_ctor = arg->ctor;
2857 zone->uz_dtor = arg->dtor;
2858 zone->uz_init = NULL;
2859 zone->uz_fini = NULL;
2860 zone->uz_sleeps = 0;
2861 zone->uz_bucket_size = 0;
2862 zone->uz_bucket_size_min = 0;
2863 zone->uz_bucket_size_max = BUCKET_MAX;
2864 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2865 zone->uz_warning = NULL;
2866 /* The domain structures follow the cpu structures. */
2867 zone->uz_bucket_max = ULONG_MAX;
2868 timevalclear(&zone->uz_ratecheck);
2869
2870 /* Count the number of duplicate names. */
2871 cnt.name = arg->name;
2872 cnt.count = 0;
2873 zone_foreach(zone_count, &cnt);
2874 zone->uz_namecnt = cnt.count;
2875 ZONE_CROSS_LOCK_INIT(zone);
2876
2877 for (i = 0; i < vm_ndomains; i++) {
2878 zdom = ZDOM_GET(zone, i);
2879 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2880 STAILQ_INIT(&zdom->uzd_buckets);
2881 }
2882
2883 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
2884 if (arg->uminit == trash_init && arg->fini == trash_fini)
2885 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2886 #elif defined(KASAN)
2887 if ((arg->flags & (UMA_ZONE_NOFREE | UMA_ZFLAG_CACHE)) != 0)
2888 arg->flags |= UMA_ZONE_NOKASAN;
2889 #endif
2890
2891 /*
2892 * This is a pure cache zone, no kegs.
2893 */
2894 if (arg->import) {
2895 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2896 ("zone_ctor: Import specified for non-cache zone."));
2897 zone->uz_flags = arg->flags;
2898 zone->uz_size = arg->size;
2899 zone->uz_import = arg->import;
2900 zone->uz_release = arg->release;
2901 zone->uz_arg = arg->arg;
2902 #ifdef NUMA
2903 /*
2904 * Cache zones are round-robin unless a policy is
2905 * specified because they may have incompatible
2906 * constraints.
2907 */
2908 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2909 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2910 #endif
2911 rw_wlock(&uma_rwlock);
2912 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2913 rw_wunlock(&uma_rwlock);
2914 goto out;
2915 }
2916
2917 /*
2918 * Use the regular zone/keg/slab allocator.
2919 */
2920 zone->uz_import = zone_import;
2921 zone->uz_release = zone_release;
2922 zone->uz_arg = zone;
2923 keg = arg->keg;
2924
2925 if (arg->flags & UMA_ZONE_SECONDARY) {
2926 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2927 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2928 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2929 zone->uz_init = arg->uminit;
2930 zone->uz_fini = arg->fini;
2931 zone->uz_flags |= UMA_ZONE_SECONDARY;
2932 rw_wlock(&uma_rwlock);
2933 ZONE_LOCK(zone);
2934 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2935 if (LIST_NEXT(z, uz_link) == NULL) {
2936 LIST_INSERT_AFTER(z, zone, uz_link);
2937 break;
2938 }
2939 }
2940 ZONE_UNLOCK(zone);
2941 rw_wunlock(&uma_rwlock);
2942 } else if (keg == NULL) {
2943 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2944 arg->align, arg->flags)) == NULL)
2945 return (ENOMEM);
2946 } else {
2947 struct uma_kctor_args karg;
2948 int error;
2949
2950 /* We should only be here from uma_startup() */
2951 karg.size = arg->size;
2952 karg.uminit = arg->uminit;
2953 karg.fini = arg->fini;
2954 karg.align = arg->align;
2955 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2956 karg.zone = zone;
2957 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2958 flags);
2959 if (error)
2960 return (error);
2961 }
2962
2963 /* Inherit properties from the keg. */
2964 zone->uz_keg = keg;
2965 zone->uz_size = keg->uk_size;
2966 zone->uz_flags |= (keg->uk_flags &
2967 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2968
2969 out:
2970 if (booted >= BOOT_PCPU) {
2971 zone_alloc_counters(zone, NULL);
2972 if (booted >= BOOT_RUNNING)
2973 zone_alloc_sysctl(zone, NULL);
2974 } else {
2975 zone->uz_allocs = EARLY_COUNTER;
2976 zone->uz_frees = EARLY_COUNTER;
2977 zone->uz_fails = EARLY_COUNTER;
2978 }
2979
2980 /* Caller requests a private SMR context. */
2981 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2982 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2983
2984 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2985 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2986 ("Invalid zone flag combination"));
2987 if (arg->flags & UMA_ZFLAG_INTERNAL)
2988 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2989 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2990 zone->uz_bucket_size = BUCKET_MAX;
2991 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2992 zone->uz_bucket_size = 0;
2993 else
2994 zone->uz_bucket_size = bucket_select(zone->uz_size);
2995 zone->uz_bucket_size_min = zone->uz_bucket_size;
2996 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2997 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2998 zone_update_caches(zone);
2999
3000 return (0);
3001 }
3002
3003 /*
3004 * Keg header dtor. This frees all data, destroys locks, frees the hash
3005 * table and removes the keg from the global list.
3006 *
3007 * Arguments/Returns follow uma_dtor specifications
3008 * udata unused
3009 */
3010 static void
keg_dtor(void * arg,int size,void * udata)3011 keg_dtor(void *arg, int size, void *udata)
3012 {
3013 uma_keg_t keg;
3014 uint32_t free, pages;
3015 int i;
3016
3017 keg = (uma_keg_t)arg;
3018 free = pages = 0;
3019 for (i = 0; i < vm_ndomains; i++) {
3020 free += keg->uk_domain[i].ud_free_items;
3021 pages += keg->uk_domain[i].ud_pages;
3022 KEG_LOCK_FINI(keg, i);
3023 }
3024 if (pages != 0)
3025 printf("Freed UMA keg (%s) was not empty (%u items). "
3026 " Lost %u pages of memory.\n",
3027 keg->uk_name ? keg->uk_name : "",
3028 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
3029
3030 hash_free(&keg->uk_hash);
3031 }
3032
3033 /*
3034 * Zone header dtor.
3035 *
3036 * Arguments/Returns follow uma_dtor specifications
3037 * udata unused
3038 */
3039 static void
zone_dtor(void * arg,int size,void * udata)3040 zone_dtor(void *arg, int size, void *udata)
3041 {
3042 uma_zone_t zone;
3043 uma_keg_t keg;
3044 int i;
3045
3046 zone = (uma_zone_t)arg;
3047
3048 sysctl_remove_oid(zone->uz_oid, 1, 1);
3049
3050 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
3051 cache_drain(zone);
3052
3053 rw_wlock(&uma_rwlock);
3054 LIST_REMOVE(zone, uz_link);
3055 rw_wunlock(&uma_rwlock);
3056 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3057 keg = zone->uz_keg;
3058 keg->uk_reserve = 0;
3059 }
3060 zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true);
3061
3062 /*
3063 * We only destroy kegs from non secondary/non cache zones.
3064 */
3065 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3066 keg = zone->uz_keg;
3067 rw_wlock(&uma_rwlock);
3068 LIST_REMOVE(keg, uk_link);
3069 rw_wunlock(&uma_rwlock);
3070 zone_free_item(kegs, keg, NULL, SKIP_NONE);
3071 }
3072 counter_u64_free(zone->uz_allocs);
3073 counter_u64_free(zone->uz_frees);
3074 counter_u64_free(zone->uz_fails);
3075 counter_u64_free(zone->uz_xdomain);
3076 free(zone->uz_ctlname, M_UMA);
3077 for (i = 0; i < vm_ndomains; i++)
3078 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
3079 ZONE_CROSS_LOCK_FINI(zone);
3080 }
3081
3082 static void
zone_foreach_unlocked(void (* zfunc)(uma_zone_t,void * arg),void * arg)3083 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3084 {
3085 uma_keg_t keg;
3086 uma_zone_t zone;
3087
3088 LIST_FOREACH(keg, &uma_kegs, uk_link) {
3089 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
3090 zfunc(zone, arg);
3091 }
3092 LIST_FOREACH(zone, &uma_cachezones, uz_link)
3093 zfunc(zone, arg);
3094 }
3095
3096 /*
3097 * Traverses every zone in the system and calls a callback
3098 *
3099 * Arguments:
3100 * zfunc A pointer to a function which accepts a zone
3101 * as an argument.
3102 *
3103 * Returns:
3104 * Nothing
3105 */
3106 static void
zone_foreach(void (* zfunc)(uma_zone_t,void * arg),void * arg)3107 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3108 {
3109
3110 rw_rlock(&uma_rwlock);
3111 zone_foreach_unlocked(zfunc, arg);
3112 rw_runlock(&uma_rwlock);
3113 }
3114
3115 /*
3116 * Initialize the kernel memory allocator. This is done after pages can be
3117 * allocated but before general KVA is available.
3118 */
3119 void
uma_startup1(vm_offset_t virtual_avail)3120 uma_startup1(vm_offset_t virtual_avail)
3121 {
3122 struct uma_zctor_args args;
3123 size_t ksize, zsize, size;
3124 uma_keg_t primarykeg;
3125 uintptr_t m;
3126 int domain;
3127 uint8_t pflag;
3128
3129 bootstart = bootmem = virtual_avail;
3130
3131 rw_init(&uma_rwlock, "UMA lock");
3132 sx_init(&uma_reclaim_lock, "umareclaim");
3133
3134 ksize = sizeof(struct uma_keg) +
3135 (sizeof(struct uma_domain) * vm_ndomains);
3136 ksize = roundup(ksize, UMA_SUPER_ALIGN);
3137 zsize = sizeof(struct uma_zone) +
3138 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
3139 (sizeof(struct uma_zone_domain) * vm_ndomains);
3140 zsize = roundup(zsize, UMA_SUPER_ALIGN);
3141
3142 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
3143 size = (zsize * 2) + ksize;
3144 for (domain = 0; domain < vm_ndomains; domain++) {
3145 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
3146 M_NOWAIT | M_ZERO);
3147 if (m != 0)
3148 break;
3149 }
3150 zones = (uma_zone_t)m;
3151 m += zsize;
3152 kegs = (uma_zone_t)m;
3153 m += zsize;
3154 primarykeg = (uma_keg_t)m;
3155
3156 /* "manually" create the initial zone */
3157 memset(&args, 0, sizeof(args));
3158 args.name = "UMA Kegs";
3159 args.size = ksize;
3160 args.ctor = keg_ctor;
3161 args.dtor = keg_dtor;
3162 args.uminit = zero_init;
3163 args.fini = NULL;
3164 args.keg = primarykeg;
3165 args.align = UMA_SUPER_ALIGN - 1;
3166 args.flags = UMA_ZFLAG_INTERNAL;
3167 zone_ctor(kegs, zsize, &args, M_WAITOK);
3168
3169 args.name = "UMA Zones";
3170 args.size = zsize;
3171 args.ctor = zone_ctor;
3172 args.dtor = zone_dtor;
3173 args.uminit = zero_init;
3174 args.fini = NULL;
3175 args.keg = NULL;
3176 args.align = UMA_SUPER_ALIGN - 1;
3177 args.flags = UMA_ZFLAG_INTERNAL;
3178 zone_ctor(zones, zsize, &args, M_WAITOK);
3179
3180 /* Now make zones for slab headers */
3181 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
3182 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3183 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
3184 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3185
3186 hashzone = uma_zcreate("UMA Hash",
3187 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
3188 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3189
3190 bucket_init();
3191 smr_init();
3192 }
3193
3194 #ifndef UMA_USE_DMAP
3195 extern void vm_radix_reserve_kva(void);
3196 #endif
3197
3198 /*
3199 * Advertise the availability of normal kva allocations and switch to
3200 * the default back-end allocator. Marks the KVA we consumed on startup
3201 * as used in the map.
3202 */
3203 void
uma_startup2(void)3204 uma_startup2(void)
3205 {
3206
3207 if (bootstart != bootmem) {
3208 vm_map_lock(kernel_map);
3209 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
3210 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
3211 vm_map_unlock(kernel_map);
3212 }
3213
3214 #ifndef UMA_USE_DMAP
3215 /* Set up radix zone to use noobj_alloc. */
3216 vm_radix_reserve_kva();
3217 #endif
3218
3219 booted = BOOT_KVA;
3220 zone_foreach_unlocked(zone_kva_available, NULL);
3221 bucket_enable();
3222 }
3223
3224 /*
3225 * Allocate counters as early as possible so that boot-time allocations are
3226 * accounted more precisely.
3227 */
3228 static void
uma_startup_pcpu(void * arg __unused)3229 uma_startup_pcpu(void *arg __unused)
3230 {
3231
3232 zone_foreach_unlocked(zone_alloc_counters, NULL);
3233 booted = BOOT_PCPU;
3234 }
3235 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
3236
3237 /*
3238 * Finish our initialization steps.
3239 */
3240 static void
uma_startup3(void * arg __unused)3241 uma_startup3(void *arg __unused)
3242 {
3243
3244 #ifdef INVARIANTS
3245 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
3246 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
3247 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
3248 #endif
3249 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
3250 booted = BOOT_RUNNING;
3251
3252 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
3253 EVENTHANDLER_PRI_FIRST);
3254 }
3255 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
3256
3257 static void
uma_startup4(void * arg __unused)3258 uma_startup4(void *arg __unused)
3259 {
3260 TIMEOUT_TASK_INIT(taskqueue_thread, &uma_timeout_task, 0, uma_timeout,
3261 NULL);
3262 taskqueue_enqueue_timeout(taskqueue_thread, &uma_timeout_task,
3263 UMA_TIMEOUT * hz);
3264 }
3265 SYSINIT(uma_startup4, SI_SUB_TASKQ, SI_ORDER_ANY, uma_startup4, NULL);
3266
3267 static void
uma_shutdown(void)3268 uma_shutdown(void)
3269 {
3270
3271 booted = BOOT_SHUTDOWN;
3272 }
3273
3274 static uma_keg_t
uma_kcreate(uma_zone_t zone,size_t size,uma_init uminit,uma_fini fini,int align,uint32_t flags)3275 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
3276 int align, uint32_t flags)
3277 {
3278 struct uma_kctor_args args;
3279
3280 args.size = size;
3281 args.uminit = uminit;
3282 args.fini = fini;
3283 args.align = align;
3284 args.flags = flags;
3285 args.zone = zone;
3286 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3287 }
3288
3289
3290 static void
check_align_mask(unsigned int mask)3291 check_align_mask(unsigned int mask)
3292 {
3293
3294 KASSERT(powerof2(mask + 1),
3295 ("UMA: %s: Not the mask of a power of 2 (%#x)", __func__, mask));
3296 /*
3297 * Make sure the stored align mask doesn't have its highest bit set,
3298 * which would cause implementation-defined behavior when passing it as
3299 * the 'align' argument of uma_zcreate(). Such very large alignments do
3300 * not make sense anyway.
3301 */
3302 KASSERT(mask <= INT_MAX,
3303 ("UMA: %s: Mask too big (%#x)", __func__, mask));
3304 }
3305
3306 /* Public functions */
3307 /* See uma.h */
3308 void
uma_set_cache_align_mask(unsigned int mask)3309 uma_set_cache_align_mask(unsigned int mask)
3310 {
3311
3312 check_align_mask(mask);
3313 uma_cache_align_mask = mask;
3314 }
3315
3316 /* Returns the alignment mask to use to request cache alignment. */
3317 unsigned int
uma_get_cache_align_mask(void)3318 uma_get_cache_align_mask(void)
3319 {
3320 return (uma_cache_align_mask);
3321 }
3322
3323 /* See uma.h */
3324 uma_zone_t
uma_zcreate(const char * name,size_t size,uma_ctor ctor,uma_dtor dtor,uma_init uminit,uma_fini fini,int align,uint32_t flags)3325 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3326 uma_init uminit, uma_fini fini, int align, uint32_t flags)
3327
3328 {
3329 struct uma_zctor_args args;
3330 uma_zone_t res;
3331
3332 check_align_mask(align);
3333
3334 /* This stuff is essential for the zone ctor */
3335 memset(&args, 0, sizeof(args));
3336 args.name = name;
3337 args.size = size;
3338 args.ctor = ctor;
3339 args.dtor = dtor;
3340 args.uminit = uminit;
3341 args.fini = fini;
3342 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
3343 /*
3344 * Inject procedures which check for memory use after free if we are
3345 * allowed to scramble the memory while it is not allocated. This
3346 * requires that: UMA is actually able to access the memory, no init
3347 * or fini procedures, no dependency on the initial value of the
3348 * memory, and no (legitimate) use of the memory after free. Note,
3349 * the ctor and dtor do not need to be empty.
3350 */
3351 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3352 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3353 args.uminit = trash_init;
3354 args.fini = trash_fini;
3355 }
3356 #endif
3357 args.align = align;
3358 args.flags = flags;
3359 args.keg = NULL;
3360
3361 sx_xlock(&uma_reclaim_lock);
3362 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3363 sx_xunlock(&uma_reclaim_lock);
3364
3365 return (res);
3366 }
3367
3368 /* See uma.h */
3369 uma_zone_t
uma_zsecond_create(const char * name,uma_ctor ctor,uma_dtor dtor,uma_init zinit,uma_fini zfini,uma_zone_t primary)3370 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3371 uma_init zinit, uma_fini zfini, uma_zone_t primary)
3372 {
3373 struct uma_zctor_args args;
3374 uma_keg_t keg;
3375 uma_zone_t res;
3376
3377 keg = primary->uz_keg;
3378 memset(&args, 0, sizeof(args));
3379 args.name = name;
3380 args.size = keg->uk_size;
3381 args.ctor = ctor;
3382 args.dtor = dtor;
3383 args.uminit = zinit;
3384 args.fini = zfini;
3385 args.align = keg->uk_align;
3386 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3387 args.keg = keg;
3388
3389 sx_xlock(&uma_reclaim_lock);
3390 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3391 sx_xunlock(&uma_reclaim_lock);
3392
3393 return (res);
3394 }
3395
3396 /* See uma.h */
3397 uma_zone_t
uma_zcache_create(const char * name,int size,uma_ctor ctor,uma_dtor dtor,uma_init zinit,uma_fini zfini,uma_import zimport,uma_release zrelease,void * arg,int flags)3398 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3399 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3400 void *arg, int flags)
3401 {
3402 struct uma_zctor_args args;
3403
3404 memset(&args, 0, sizeof(args));
3405 args.name = name;
3406 args.size = size;
3407 args.ctor = ctor;
3408 args.dtor = dtor;
3409 args.uminit = zinit;
3410 args.fini = zfini;
3411 args.import = zimport;
3412 args.release = zrelease;
3413 args.arg = arg;
3414 args.align = 0;
3415 args.flags = flags | UMA_ZFLAG_CACHE;
3416
3417 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3418 }
3419
3420 /* See uma.h */
3421 void
uma_zdestroy(uma_zone_t zone)3422 uma_zdestroy(uma_zone_t zone)
3423 {
3424
3425 /*
3426 * Large slabs are expensive to reclaim, so don't bother doing
3427 * unnecessary work if we're shutting down.
3428 */
3429 if (booted == BOOT_SHUTDOWN &&
3430 zone->uz_fini == NULL && zone->uz_release == zone_release)
3431 return;
3432 sx_xlock(&uma_reclaim_lock);
3433 zone_free_item(zones, zone, NULL, SKIP_NONE);
3434 sx_xunlock(&uma_reclaim_lock);
3435 }
3436
3437 void
uma_zwait(uma_zone_t zone)3438 uma_zwait(uma_zone_t zone)
3439 {
3440
3441 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3442 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3443 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3444 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3445 else
3446 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3447 }
3448
3449 void *
uma_zalloc_pcpu_arg(uma_zone_t zone,void * udata,int flags)3450 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3451 {
3452 void *item, *pcpu_item;
3453 #ifdef SMP
3454 int i;
3455
3456 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3457 #endif
3458 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3459 if (item == NULL)
3460 return (NULL);
3461 pcpu_item = zpcpu_base_to_offset(item);
3462 if (flags & M_ZERO) {
3463 #ifdef SMP
3464 for (i = 0; i <= mp_maxid; i++)
3465 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3466 #else
3467 bzero(item, zone->uz_size);
3468 #endif
3469 }
3470 return (pcpu_item);
3471 }
3472
3473 /*
3474 * A stub while both regular and pcpu cases are identical.
3475 */
3476 void
uma_zfree_pcpu_arg(uma_zone_t zone,void * pcpu_item,void * udata)3477 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3478 {
3479 void *item;
3480
3481 #ifdef SMP
3482 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3483 #endif
3484
3485 /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */
3486 if (pcpu_item == NULL)
3487 return;
3488
3489 item = zpcpu_offset_to_base(pcpu_item);
3490 uma_zfree_arg(zone, item, udata);
3491 }
3492
3493 static inline void *
item_ctor(uma_zone_t zone,int uz_flags,int size,void * udata,int flags,void * item)3494 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3495 void *item)
3496 {
3497 #ifdef INVARIANTS
3498 bool skipdbg;
3499 #endif
3500
3501 kasan_mark_item_valid(zone, item);
3502 kmsan_mark_item_uninitialized(zone, item);
3503
3504 #ifdef INVARIANTS
3505 skipdbg = uma_dbg_zskip(zone, item);
3506 if (!skipdbg && (uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3507 zone->uz_ctor != trash_ctor)
3508 trash_ctor(item, size, zone, flags);
3509 #endif
3510
3511 /* Check flags before loading ctor pointer. */
3512 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3513 __predict_false(zone->uz_ctor != NULL) &&
3514 zone->uz_ctor(item, size, udata, flags) != 0) {
3515 counter_u64_add(zone->uz_fails, 1);
3516 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3517 return (NULL);
3518 }
3519 #ifdef INVARIANTS
3520 if (!skipdbg)
3521 uma_dbg_alloc(zone, NULL, item);
3522 #endif
3523 if (__predict_false(flags & M_ZERO))
3524 return (memset(item, 0, size));
3525
3526 return (item);
3527 }
3528
3529 static inline void
item_dtor(uma_zone_t zone,void * item,int size,void * udata,enum zfreeskip skip)3530 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3531 enum zfreeskip skip)
3532 {
3533 #ifdef INVARIANTS
3534 bool skipdbg;
3535
3536 skipdbg = uma_dbg_zskip(zone, item);
3537 if (skip == SKIP_NONE && !skipdbg) {
3538 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3539 uma_dbg_free(zone, udata, item);
3540 else
3541 uma_dbg_free(zone, NULL, item);
3542 }
3543 #endif
3544 if (__predict_true(skip < SKIP_DTOR)) {
3545 if (zone->uz_dtor != NULL)
3546 zone->uz_dtor(item, size, udata);
3547 #ifdef INVARIANTS
3548 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3549 zone->uz_dtor != trash_dtor)
3550 trash_dtor(item, size, zone);
3551 #endif
3552 }
3553 kasan_mark_item_invalid(zone, item);
3554 }
3555
3556 #ifdef NUMA
3557 static int
item_domain(void * item)3558 item_domain(void *item)
3559 {
3560 int domain;
3561
3562 domain = vm_phys_domain(vtophys(item));
3563 KASSERT(domain >= 0 && domain < vm_ndomains,
3564 ("%s: unknown domain for item %p", __func__, item));
3565 return (domain);
3566 }
3567 #endif
3568
3569 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3570 #if defined(INVARIANTS) && (defined(DDB) || defined(STACK))
3571 #include <sys/stack.h>
3572 #endif
3573 #define UMA_ZALLOC_DEBUG
3574 static int
uma_zalloc_debug(uma_zone_t zone,void ** itemp,void * udata,int flags)3575 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3576 {
3577 int error;
3578
3579 error = 0;
3580 #ifdef WITNESS
3581 if (flags & M_WAITOK) {
3582 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3583 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3584 }
3585 #endif
3586
3587 #ifdef INVARIANTS
3588 KASSERT((flags & M_EXEC) == 0,
3589 ("uma_zalloc_debug: called with M_EXEC"));
3590 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3591 ("uma_zalloc_debug: called within spinlock or critical section"));
3592 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3593 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3594
3595 _Static_assert(M_NOWAIT != 0 && M_WAITOK != 0,
3596 "M_NOWAIT and M_WAITOK must be non-zero for this assertion:");
3597 #if 0
3598 /*
3599 * Give the #elif clause time to find problems, then remove it
3600 * and enable this. (Remove <sys/stack.h> above, too.)
3601 */
3602 KASSERT((flags & (M_NOWAIT|M_WAITOK)) == M_NOWAIT ||
3603 (flags & (M_NOWAIT|M_WAITOK)) == M_WAITOK,
3604 ("uma_zalloc_debug: must pass one of M_NOWAIT or M_WAITOK"));
3605 #elif defined(DDB) || defined(STACK)
3606 if (__predict_false((flags & (M_NOWAIT|M_WAITOK)) != M_NOWAIT &&
3607 (flags & (M_NOWAIT|M_WAITOK)) != M_WAITOK)) {
3608 static int stack_count;
3609 struct stack st;
3610
3611 if (stack_count < 10) {
3612 ++stack_count;
3613 printf("uma_zalloc* called with bad WAIT flags:\n");
3614 stack_save(&st);
3615 stack_print(&st);
3616 }
3617 }
3618 #endif
3619 #endif
3620
3621 #ifdef DEBUG_MEMGUARD
3622 if ((zone->uz_flags & (UMA_ZONE_SMR | UMA_ZFLAG_CACHE)) == 0 &&
3623 memguard_cmp_zone(zone)) {
3624 void *item;
3625 item = memguard_alloc(zone->uz_size, flags);
3626 if (item != NULL) {
3627 error = EJUSTRETURN;
3628 if (zone->uz_init != NULL &&
3629 zone->uz_init(item, zone->uz_size, flags) != 0) {
3630 *itemp = NULL;
3631 return (error);
3632 }
3633 if (zone->uz_ctor != NULL &&
3634 zone->uz_ctor(item, zone->uz_size, udata,
3635 flags) != 0) {
3636 counter_u64_add(zone->uz_fails, 1);
3637 if (zone->uz_fini != NULL)
3638 zone->uz_fini(item, zone->uz_size);
3639 *itemp = NULL;
3640 return (error);
3641 }
3642 *itemp = item;
3643 return (error);
3644 }
3645 /* This is unfortunate but should not be fatal. */
3646 }
3647 #endif
3648 return (error);
3649 }
3650
3651 static int
uma_zfree_debug(uma_zone_t zone,void * item,void * udata)3652 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3653 {
3654 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3655 ("uma_zfree_debug: called with spinlock or critical section held"));
3656
3657 #ifdef DEBUG_MEMGUARD
3658 if ((zone->uz_flags & (UMA_ZONE_SMR | UMA_ZFLAG_CACHE)) == 0 &&
3659 is_memguard_addr(item)) {
3660 if (zone->uz_dtor != NULL)
3661 zone->uz_dtor(item, zone->uz_size, udata);
3662 if (zone->uz_fini != NULL)
3663 zone->uz_fini(item, zone->uz_size);
3664 memguard_free(item);
3665 return (EJUSTRETURN);
3666 }
3667 #endif
3668 return (0);
3669 }
3670 #endif
3671
3672 static inline void *
cache_alloc_item(uma_zone_t zone,uma_cache_t cache,uma_cache_bucket_t bucket,void * udata,int flags)3673 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3674 void *udata, int flags)
3675 {
3676 void *item;
3677 int size, uz_flags;
3678
3679 item = cache_bucket_pop(cache, bucket);
3680 size = cache_uz_size(cache);
3681 uz_flags = cache_uz_flags(cache);
3682 critical_exit();
3683 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3684 }
3685
3686 static __noinline void *
cache_alloc_retry(uma_zone_t zone,uma_cache_t cache,void * udata,int flags)3687 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3688 {
3689 uma_cache_bucket_t bucket;
3690 int domain;
3691
3692 while (cache_alloc(zone, cache, udata, flags)) {
3693 cache = &zone->uz_cpu[curcpu];
3694 bucket = &cache->uc_allocbucket;
3695 if (__predict_false(bucket->ucb_cnt == 0))
3696 continue;
3697 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3698 }
3699 critical_exit();
3700
3701 /*
3702 * We can not get a bucket so try to return a single item.
3703 */
3704 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3705 domain = PCPU_GET(domain);
3706 else
3707 domain = UMA_ANYDOMAIN;
3708 return (zone_alloc_item(zone, udata, domain, flags));
3709 }
3710
3711 /* See uma.h */
3712 void *
uma_zalloc_smr(uma_zone_t zone,int flags)3713 uma_zalloc_smr(uma_zone_t zone, int flags)
3714 {
3715 uma_cache_bucket_t bucket;
3716 uma_cache_t cache;
3717
3718 CTR3(KTR_UMA, "uma_zalloc_smr zone %s(%p) flags %d", zone->uz_name,
3719 zone, flags);
3720
3721 #ifdef UMA_ZALLOC_DEBUG
3722 void *item;
3723
3724 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3725 ("uma_zalloc_arg: called with non-SMR zone."));
3726 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3727 return (item);
3728 #endif
3729
3730 critical_enter();
3731 cache = &zone->uz_cpu[curcpu];
3732 bucket = &cache->uc_allocbucket;
3733 if (__predict_false(bucket->ucb_cnt == 0))
3734 return (cache_alloc_retry(zone, cache, NULL, flags));
3735 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3736 }
3737
3738 /* See uma.h */
3739 void *
uma_zalloc_arg(uma_zone_t zone,void * udata,int flags)3740 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3741 {
3742 uma_cache_bucket_t bucket;
3743 uma_cache_t cache;
3744
3745 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3746 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3747
3748 /* This is the fast path allocation */
3749 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3750 zone, flags);
3751
3752 #ifdef UMA_ZALLOC_DEBUG
3753 void *item;
3754
3755 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3756 ("uma_zalloc_arg: called with SMR zone."));
3757 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3758 return (item);
3759 #endif
3760
3761 /*
3762 * If possible, allocate from the per-CPU cache. There are two
3763 * requirements for safe access to the per-CPU cache: (1) the thread
3764 * accessing the cache must not be preempted or yield during access,
3765 * and (2) the thread must not migrate CPUs without switching which
3766 * cache it accesses. We rely on a critical section to prevent
3767 * preemption and migration. We release the critical section in
3768 * order to acquire the zone mutex if we are unable to allocate from
3769 * the current cache; when we re-acquire the critical section, we
3770 * must detect and handle migration if it has occurred.
3771 */
3772 critical_enter();
3773 cache = &zone->uz_cpu[curcpu];
3774 bucket = &cache->uc_allocbucket;
3775 if (__predict_false(bucket->ucb_cnt == 0))
3776 return (cache_alloc_retry(zone, cache, udata, flags));
3777 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3778 }
3779
3780 /*
3781 * Replenish an alloc bucket and possibly restore an old one. Called in
3782 * a critical section. Returns in a critical section.
3783 *
3784 * A false return value indicates an allocation failure.
3785 * A true return value indicates success and the caller should retry.
3786 */
3787 static __noinline bool
cache_alloc(uma_zone_t zone,uma_cache_t cache,void * udata,int flags)3788 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3789 {
3790 uma_bucket_t bucket;
3791 int curdomain, domain;
3792 bool new;
3793
3794 CRITICAL_ASSERT(curthread);
3795
3796 /*
3797 * If we have run out of items in our alloc bucket see
3798 * if we can switch with the free bucket.
3799 *
3800 * SMR Zones can't re-use the free bucket until the sequence has
3801 * expired.
3802 */
3803 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3804 cache->uc_freebucket.ucb_cnt != 0) {
3805 cache_bucket_swap(&cache->uc_freebucket,
3806 &cache->uc_allocbucket);
3807 return (true);
3808 }
3809
3810 /*
3811 * Discard any empty allocation bucket while we hold no locks.
3812 */
3813 bucket = cache_bucket_unload_alloc(cache);
3814 critical_exit();
3815
3816 if (bucket != NULL) {
3817 KASSERT(bucket->ub_cnt == 0,
3818 ("cache_alloc: Entered with non-empty alloc bucket."));
3819 bucket_free(zone, bucket, udata);
3820 }
3821
3822 /*
3823 * Attempt to retrieve the item from the per-CPU cache has failed, so
3824 * we must go back to the zone. This requires the zdom lock, so we
3825 * must drop the critical section, then re-acquire it when we go back
3826 * to the cache. Since the critical section is released, we may be
3827 * preempted or migrate. As such, make sure not to maintain any
3828 * thread-local state specific to the cache from prior to releasing
3829 * the critical section.
3830 */
3831 domain = PCPU_GET(domain);
3832 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3833 VM_DOMAIN_EMPTY(domain))
3834 domain = zone_domain_highest(zone, domain);
3835 bucket = cache_fetch_bucket(zone, cache, domain);
3836 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3837 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3838 new = true;
3839 } else {
3840 new = false;
3841 }
3842
3843 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3844 zone->uz_name, zone, bucket);
3845 if (bucket == NULL) {
3846 critical_enter();
3847 return (false);
3848 }
3849
3850 /*
3851 * See if we lost the race or were migrated. Cache the
3852 * initialized bucket to make this less likely or claim
3853 * the memory directly.
3854 */
3855 critical_enter();
3856 cache = &zone->uz_cpu[curcpu];
3857 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3858 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3859 (curdomain = PCPU_GET(domain)) == domain ||
3860 VM_DOMAIN_EMPTY(curdomain))) {
3861 if (new)
3862 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3863 bucket->ub_cnt);
3864 cache_bucket_load_alloc(cache, bucket);
3865 return (true);
3866 }
3867
3868 /*
3869 * We lost the race, release this bucket and start over.
3870 */
3871 critical_exit();
3872 zone_put_bucket(zone, domain, bucket, udata, !new);
3873 critical_enter();
3874
3875 return (true);
3876 }
3877
3878 void *
uma_zalloc_domain(uma_zone_t zone,void * udata,int domain,int flags)3879 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3880 {
3881 #ifdef NUMA
3882 uma_bucket_t bucket;
3883 uma_zone_domain_t zdom;
3884 void *item;
3885 #endif
3886
3887 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3888 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3889
3890 /* This is the fast path allocation */
3891 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3892 zone->uz_name, zone, domain, flags);
3893
3894 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3895 ("uma_zalloc_domain: called with SMR zone."));
3896 #ifdef NUMA
3897 KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3898 ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3899
3900 if (vm_ndomains == 1)
3901 return (uma_zalloc_arg(zone, udata, flags));
3902
3903 #ifdef UMA_ZALLOC_DEBUG
3904 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3905 return (item);
3906 #endif
3907
3908 /*
3909 * Try to allocate from the bucket cache before falling back to the keg.
3910 * We could try harder and attempt to allocate from per-CPU caches or
3911 * the per-domain cross-domain buckets, but the complexity is probably
3912 * not worth it. It is more important that frees of previous
3913 * cross-domain allocations do not blow up the cache.
3914 */
3915 zdom = zone_domain_lock(zone, domain);
3916 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3917 item = bucket->ub_bucket[bucket->ub_cnt - 1];
3918 #ifdef INVARIANTS
3919 bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3920 #endif
3921 bucket->ub_cnt--;
3922 zone_put_bucket(zone, domain, bucket, udata, true);
3923 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3924 flags, item);
3925 if (item != NULL) {
3926 KASSERT(item_domain(item) == domain,
3927 ("%s: bucket cache item %p from wrong domain",
3928 __func__, item));
3929 counter_u64_add(zone->uz_allocs, 1);
3930 }
3931 return (item);
3932 }
3933 ZDOM_UNLOCK(zdom);
3934 return (zone_alloc_item(zone, udata, domain, flags));
3935 #else
3936 return (uma_zalloc_arg(zone, udata, flags));
3937 #endif
3938 }
3939
3940 /*
3941 * Find a slab with some space. Prefer slabs that are partially used over those
3942 * that are totally full. This helps to reduce fragmentation.
3943 *
3944 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3945 * only 'domain'.
3946 */
3947 static uma_slab_t
keg_first_slab(uma_keg_t keg,int domain,bool rr)3948 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3949 {
3950 uma_domain_t dom;
3951 uma_slab_t slab;
3952 int start;
3953
3954 KASSERT(domain >= 0 && domain < vm_ndomains,
3955 ("keg_first_slab: domain %d out of range", domain));
3956 KEG_LOCK_ASSERT(keg, domain);
3957
3958 slab = NULL;
3959 start = domain;
3960 do {
3961 dom = &keg->uk_domain[domain];
3962 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3963 return (slab);
3964 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3965 LIST_REMOVE(slab, us_link);
3966 dom->ud_free_slabs--;
3967 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3968 return (slab);
3969 }
3970 if (rr)
3971 domain = (domain + 1) % vm_ndomains;
3972 } while (domain != start);
3973
3974 return (NULL);
3975 }
3976
3977 /*
3978 * Fetch an existing slab from a free or partial list. Returns with the
3979 * keg domain lock held if a slab was found or unlocked if not.
3980 */
3981 static uma_slab_t
keg_fetch_free_slab(uma_keg_t keg,int domain,bool rr,int flags)3982 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3983 {
3984 uma_slab_t slab;
3985 uint32_t reserve;
3986
3987 /* HASH has a single free list. */
3988 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3989 domain = 0;
3990
3991 KEG_LOCK(keg, domain);
3992 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3993 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3994 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3995 KEG_UNLOCK(keg, domain);
3996 return (NULL);
3997 }
3998 return (slab);
3999 }
4000
4001 static uma_slab_t
keg_fetch_slab(uma_keg_t keg,uma_zone_t zone,int rdomain,const int flags)4002 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
4003 {
4004 struct vm_domainset_iter di;
4005 uma_slab_t slab;
4006 int aflags, domain;
4007 bool rr;
4008
4009 KASSERT((flags & (M_WAITOK | M_NOVM)) != (M_WAITOK | M_NOVM),
4010 ("%s: invalid flags %#x", __func__, flags));
4011
4012 restart:
4013 /*
4014 * Use the keg's policy if upper layers haven't already specified a
4015 * domain (as happens with first-touch zones).
4016 *
4017 * To avoid races we run the iterator with the keg lock held, but that
4018 * means that we cannot allow the vm_domainset layer to sleep. Thus,
4019 * clear M_WAITOK and handle low memory conditions locally.
4020 */
4021 rr = rdomain == UMA_ANYDOMAIN;
4022 if (rr) {
4023 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
4024 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4025 &aflags);
4026 } else {
4027 aflags = flags;
4028 domain = rdomain;
4029 }
4030
4031 for (;;) {
4032 slab = keg_fetch_free_slab(keg, domain, rr, flags);
4033 if (slab != NULL)
4034 return (slab);
4035
4036 /*
4037 * M_NOVM is used to break the recursion that can otherwise
4038 * occur if low-level memory management routines use UMA.
4039 */
4040 if ((flags & M_NOVM) == 0) {
4041 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
4042 if (slab != NULL)
4043 return (slab);
4044 }
4045
4046 if (!rr) {
4047 if ((flags & M_USE_RESERVE) != 0) {
4048 /*
4049 * Drain reserves from other domains before
4050 * giving up or sleeping. It may be useful to
4051 * support per-domain reserves eventually.
4052 */
4053 rdomain = UMA_ANYDOMAIN;
4054 goto restart;
4055 }
4056 if ((flags & M_WAITOK) == 0)
4057 break;
4058 vm_wait_domain(domain);
4059 } else if (vm_domainset_iter_policy(&di, &domain) != 0) {
4060 if ((flags & M_WAITOK) != 0) {
4061 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
4062 goto restart;
4063 }
4064 break;
4065 }
4066 }
4067
4068 /*
4069 * We might not have been able to get a slab but another cpu
4070 * could have while we were unlocked. Check again before we
4071 * fail.
4072 */
4073 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
4074 return (slab);
4075
4076 return (NULL);
4077 }
4078
4079 static void *
slab_alloc_item(uma_keg_t keg,uma_slab_t slab)4080 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
4081 {
4082 uma_domain_t dom;
4083 void *item;
4084 int freei;
4085
4086 KEG_LOCK_ASSERT(keg, slab->us_domain);
4087
4088 dom = &keg->uk_domain[slab->us_domain];
4089 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
4090 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
4091 item = slab_item(slab, keg, freei);
4092 slab->us_freecount--;
4093 dom->ud_free_items--;
4094
4095 /*
4096 * Move this slab to the full list. It must be on the partial list, so
4097 * we do not need to update the free slab count. In particular,
4098 * keg_fetch_slab() always returns slabs on the partial list.
4099 */
4100 if (slab->us_freecount == 0) {
4101 LIST_REMOVE(slab, us_link);
4102 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
4103 }
4104
4105 return (item);
4106 }
4107
4108 static int
zone_import(void * arg,void ** bucket,int max,int domain,int flags)4109 zone_import(void *arg, void **bucket, int max, int domain, int flags)
4110 {
4111 uma_domain_t dom;
4112 uma_zone_t zone;
4113 uma_slab_t slab;
4114 uma_keg_t keg;
4115 #ifdef NUMA
4116 int stripe;
4117 #endif
4118 int i;
4119
4120 zone = arg;
4121 slab = NULL;
4122 keg = zone->uz_keg;
4123 /* Try to keep the buckets totally full */
4124 for (i = 0; i < max; ) {
4125 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
4126 break;
4127 #ifdef NUMA
4128 stripe = howmany(max, vm_ndomains);
4129 #endif
4130 dom = &keg->uk_domain[slab->us_domain];
4131 do {
4132 bucket[i++] = slab_alloc_item(keg, slab);
4133 if (keg->uk_reserve > 0 &&
4134 dom->ud_free_items <= keg->uk_reserve) {
4135 /*
4136 * Avoid depleting the reserve after a
4137 * successful item allocation, even if
4138 * M_USE_RESERVE is specified.
4139 */
4140 KEG_UNLOCK(keg, slab->us_domain);
4141 goto out;
4142 }
4143 #ifdef NUMA
4144 /*
4145 * If the zone is striped we pick a new slab for every
4146 * N allocations. Eliminating this conditional will
4147 * instead pick a new domain for each bucket rather
4148 * than stripe within each bucket. The current option
4149 * produces more fragmentation and requires more cpu
4150 * time but yields better distribution.
4151 */
4152 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
4153 vm_ndomains > 1 && --stripe == 0)
4154 break;
4155 #endif
4156 } while (slab->us_freecount != 0 && i < max);
4157 KEG_UNLOCK(keg, slab->us_domain);
4158
4159 /* Don't block if we allocated any successfully. */
4160 flags &= ~M_WAITOK;
4161 flags |= M_NOWAIT;
4162 }
4163 out:
4164 return i;
4165 }
4166
4167 static int
zone_alloc_limit_hard(uma_zone_t zone,int count,int flags)4168 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
4169 {
4170 uint64_t old, new, total, max;
4171
4172 /*
4173 * The hard case. We're going to sleep because there were existing
4174 * sleepers or because we ran out of items. This routine enforces
4175 * fairness by keeping fifo order.
4176 *
4177 * First release our ill gotten gains and make some noise.
4178 */
4179 for (;;) {
4180 zone_free_limit(zone, count);
4181 zone_log_warning(zone);
4182 zone_maxaction(zone);
4183 if (flags & M_NOWAIT)
4184 return (0);
4185
4186 /*
4187 * We need to allocate an item or set ourself as a sleeper
4188 * while the sleepq lock is held to avoid wakeup races. This
4189 * is essentially a home rolled semaphore.
4190 */
4191 sleepq_lock(&zone->uz_max_items);
4192 old = zone->uz_items;
4193 do {
4194 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
4195 /* Cache the max since we will evaluate twice. */
4196 max = zone->uz_max_items;
4197 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
4198 UZ_ITEMS_COUNT(old) >= max)
4199 new = old + UZ_ITEMS_SLEEPER;
4200 else
4201 new = old + MIN(count, max - old);
4202 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
4203
4204 /* We may have successfully allocated under the sleepq lock. */
4205 if (UZ_ITEMS_SLEEPERS(new) == 0) {
4206 sleepq_release(&zone->uz_max_items);
4207 return (new - old);
4208 }
4209
4210 /*
4211 * This is in a different cacheline from uz_items so that we
4212 * don't constantly invalidate the fastpath cacheline when we
4213 * adjust item counts. This could be limited to toggling on
4214 * transitions.
4215 */
4216 atomic_add_32(&zone->uz_sleepers, 1);
4217 atomic_add_64(&zone->uz_sleeps, 1);
4218
4219 /*
4220 * We have added ourselves as a sleeper. The sleepq lock
4221 * protects us from wakeup races. Sleep now and then retry.
4222 */
4223 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
4224 sleepq_wait(&zone->uz_max_items, PVM);
4225
4226 /*
4227 * After wakeup, remove ourselves as a sleeper and try
4228 * again. We no longer have the sleepq lock for protection.
4229 *
4230 * Subract ourselves as a sleeper while attempting to add
4231 * our count.
4232 */
4233 atomic_subtract_32(&zone->uz_sleepers, 1);
4234 old = atomic_fetchadd_64(&zone->uz_items,
4235 -(UZ_ITEMS_SLEEPER - count));
4236 /* We're no longer a sleeper. */
4237 old -= UZ_ITEMS_SLEEPER;
4238
4239 /*
4240 * If we're still at the limit, restart. Notably do not
4241 * block on other sleepers. Cache the max value to protect
4242 * against changes via sysctl.
4243 */
4244 total = UZ_ITEMS_COUNT(old);
4245 max = zone->uz_max_items;
4246 if (total >= max)
4247 continue;
4248 /* Truncate if necessary, otherwise wake other sleepers. */
4249 if (total + count > max) {
4250 zone_free_limit(zone, total + count - max);
4251 count = max - total;
4252 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
4253 wakeup_one(&zone->uz_max_items);
4254
4255 return (count);
4256 }
4257 }
4258
4259 /*
4260 * Allocate 'count' items from our max_items limit. Returns the number
4261 * available. If M_NOWAIT is not specified it will sleep until at least
4262 * one item can be allocated.
4263 */
4264 static int
zone_alloc_limit(uma_zone_t zone,int count,int flags)4265 zone_alloc_limit(uma_zone_t zone, int count, int flags)
4266 {
4267 uint64_t old;
4268 uint64_t max;
4269
4270 max = zone->uz_max_items;
4271 MPASS(max > 0);
4272
4273 /*
4274 * We expect normal allocations to succeed with a simple
4275 * fetchadd.
4276 */
4277 old = atomic_fetchadd_64(&zone->uz_items, count);
4278 if (__predict_true(old + count <= max))
4279 return (count);
4280
4281 /*
4282 * If we had some items and no sleepers just return the
4283 * truncated value. We have to release the excess space
4284 * though because that may wake sleepers who weren't woken
4285 * because we were temporarily over the limit.
4286 */
4287 if (old < max) {
4288 zone_free_limit(zone, (old + count) - max);
4289 return (max - old);
4290 }
4291 return (zone_alloc_limit_hard(zone, count, flags));
4292 }
4293
4294 /*
4295 * Free a number of items back to the limit.
4296 */
4297 static void
zone_free_limit(uma_zone_t zone,int count)4298 zone_free_limit(uma_zone_t zone, int count)
4299 {
4300 uint64_t old;
4301
4302 MPASS(count > 0);
4303
4304 /*
4305 * In the common case we either have no sleepers or
4306 * are still over the limit and can just return.
4307 */
4308 old = atomic_fetchadd_64(&zone->uz_items, -count);
4309 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
4310 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
4311 return;
4312
4313 /*
4314 * Moderate the rate of wakeups. Sleepers will continue
4315 * to generate wakeups if necessary.
4316 */
4317 wakeup_one(&zone->uz_max_items);
4318 }
4319
4320 static uma_bucket_t
zone_alloc_bucket(uma_zone_t zone,void * udata,int domain,int flags)4321 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
4322 {
4323 uma_bucket_t bucket;
4324 int error, maxbucket, cnt;
4325
4326 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
4327 zone, domain);
4328
4329 /* Avoid allocs targeting empty domains. */
4330 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4331 domain = UMA_ANYDOMAIN;
4332 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4333 domain = UMA_ANYDOMAIN;
4334
4335 if (zone->uz_max_items > 0)
4336 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
4337 M_NOWAIT);
4338 else
4339 maxbucket = zone->uz_bucket_size;
4340 if (maxbucket == 0)
4341 return (NULL);
4342
4343 /* Don't wait for buckets, preserve caller's NOVM setting. */
4344 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
4345 if (bucket == NULL) {
4346 cnt = 0;
4347 goto out;
4348 }
4349
4350 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
4351 MIN(maxbucket, bucket->ub_entries), domain, flags);
4352
4353 /*
4354 * Initialize the memory if necessary.
4355 */
4356 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4357 int i;
4358
4359 for (i = 0; i < bucket->ub_cnt; i++) {
4360 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
4361 error = zone->uz_init(bucket->ub_bucket[i],
4362 zone->uz_size, flags);
4363 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
4364 if (error != 0)
4365 break;
4366 }
4367
4368 /*
4369 * If we couldn't initialize the whole bucket, put the
4370 * rest back onto the freelist.
4371 */
4372 if (i != bucket->ub_cnt) {
4373 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4374 bucket->ub_cnt - i);
4375 #ifdef INVARIANTS
4376 bzero(&bucket->ub_bucket[i],
4377 sizeof(void *) * (bucket->ub_cnt - i));
4378 #endif
4379 bucket->ub_cnt = i;
4380 }
4381 }
4382
4383 cnt = bucket->ub_cnt;
4384 if (bucket->ub_cnt == 0) {
4385 bucket_free(zone, bucket, udata);
4386 counter_u64_add(zone->uz_fails, 1);
4387 bucket = NULL;
4388 }
4389 out:
4390 if (zone->uz_max_items > 0 && cnt < maxbucket)
4391 zone_free_limit(zone, maxbucket - cnt);
4392
4393 return (bucket);
4394 }
4395
4396 /*
4397 * Allocates a single item from a zone.
4398 *
4399 * Arguments
4400 * zone The zone to alloc for.
4401 * udata The data to be passed to the constructor.
4402 * domain The domain to allocate from or UMA_ANYDOMAIN.
4403 * flags M_WAITOK, M_NOWAIT, M_ZERO.
4404 *
4405 * Returns
4406 * NULL if there is no memory and M_NOWAIT is set
4407 * An item if successful
4408 */
4409
4410 static void *
zone_alloc_item(uma_zone_t zone,void * udata,int domain,int flags)4411 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4412 {
4413 void *item;
4414
4415 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4416 counter_u64_add(zone->uz_fails, 1);
4417 return (NULL);
4418 }
4419
4420 /* Avoid allocs targeting empty domains. */
4421 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4422 domain = UMA_ANYDOMAIN;
4423
4424 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4425 goto fail_cnt;
4426
4427 /*
4428 * We have to call both the zone's init (not the keg's init)
4429 * and the zone's ctor. This is because the item is going from
4430 * a keg slab directly to the user, and the user is expecting it
4431 * to be both zone-init'd as well as zone-ctor'd.
4432 */
4433 if (zone->uz_init != NULL) {
4434 int error;
4435
4436 kasan_mark_item_valid(zone, item);
4437 error = zone->uz_init(item, zone->uz_size, flags);
4438 kasan_mark_item_invalid(zone, item);
4439 if (error != 0) {
4440 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4441 goto fail_cnt;
4442 }
4443 }
4444 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4445 item);
4446 if (item == NULL)
4447 goto fail;
4448
4449 counter_u64_add(zone->uz_allocs, 1);
4450 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4451 zone->uz_name, zone);
4452
4453 return (item);
4454
4455 fail_cnt:
4456 counter_u64_add(zone->uz_fails, 1);
4457 fail:
4458 if (zone->uz_max_items > 0)
4459 zone_free_limit(zone, 1);
4460 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4461 zone->uz_name, zone);
4462
4463 return (NULL);
4464 }
4465
4466 /* See uma.h */
4467 void
uma_zfree_smr(uma_zone_t zone,void * item)4468 uma_zfree_smr(uma_zone_t zone, void *item)
4469 {
4470 uma_cache_t cache;
4471 uma_cache_bucket_t bucket;
4472 int itemdomain;
4473 #ifdef NUMA
4474 int uz_flags;
4475 #endif
4476
4477 CTR3(KTR_UMA, "uma_zfree_smr zone %s(%p) item %p",
4478 zone->uz_name, zone, item);
4479
4480 #ifdef UMA_ZALLOC_DEBUG
4481 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4482 ("uma_zfree_smr: called with non-SMR zone."));
4483 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4484 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4485 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4486 return;
4487 #endif
4488 cache = &zone->uz_cpu[curcpu];
4489 itemdomain = 0;
4490 #ifdef NUMA
4491 uz_flags = cache_uz_flags(cache);
4492 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4493 itemdomain = item_domain(item);
4494 #endif
4495 critical_enter();
4496 do {
4497 cache = &zone->uz_cpu[curcpu];
4498 /* SMR Zones must free to the free bucket. */
4499 bucket = &cache->uc_freebucket;
4500 #ifdef NUMA
4501 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4502 PCPU_GET(domain) != itemdomain) {
4503 bucket = &cache->uc_crossbucket;
4504 }
4505 #endif
4506 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4507 cache_bucket_push(cache, bucket, item);
4508 critical_exit();
4509 return;
4510 }
4511 } while (cache_free(zone, cache, NULL, itemdomain));
4512 critical_exit();
4513
4514 /*
4515 * If nothing else caught this, we'll just do an internal free.
4516 */
4517 zone_free_item(zone, item, NULL, SKIP_NONE);
4518 }
4519
4520 /* See uma.h */
4521 void
uma_zfree_arg(uma_zone_t zone,void * item,void * udata)4522 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4523 {
4524 uma_cache_t cache;
4525 uma_cache_bucket_t bucket;
4526 int itemdomain, uz_flags;
4527
4528 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4529 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4530
4531 CTR3(KTR_UMA, "uma_zfree_arg zone %s(%p) item %p",
4532 zone->uz_name, zone, item);
4533
4534 #ifdef UMA_ZALLOC_DEBUG
4535 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4536 ("uma_zfree_arg: called with SMR zone."));
4537 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4538 return;
4539 #endif
4540 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4541 if (item == NULL)
4542 return;
4543
4544 /*
4545 * We are accessing the per-cpu cache without a critical section to
4546 * fetch size and flags. This is acceptable, if we are preempted we
4547 * will simply read another cpu's line.
4548 */
4549 cache = &zone->uz_cpu[curcpu];
4550 uz_flags = cache_uz_flags(cache);
4551 if (UMA_ALWAYS_CTORDTOR ||
4552 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4553 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4554
4555 /*
4556 * The race here is acceptable. If we miss it we'll just have to wait
4557 * a little longer for the limits to be reset.
4558 */
4559 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4560 if (atomic_load_32(&zone->uz_sleepers) > 0)
4561 goto zfree_item;
4562 }
4563
4564 /*
4565 * If possible, free to the per-CPU cache. There are two
4566 * requirements for safe access to the per-CPU cache: (1) the thread
4567 * accessing the cache must not be preempted or yield during access,
4568 * and (2) the thread must not migrate CPUs without switching which
4569 * cache it accesses. We rely on a critical section to prevent
4570 * preemption and migration. We release the critical section in
4571 * order to acquire the zone mutex if we are unable to free to the
4572 * current cache; when we re-acquire the critical section, we must
4573 * detect and handle migration if it has occurred.
4574 */
4575 itemdomain = 0;
4576 #ifdef NUMA
4577 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4578 itemdomain = item_domain(item);
4579 #endif
4580 critical_enter();
4581 do {
4582 cache = &zone->uz_cpu[curcpu];
4583 /*
4584 * Try to free into the allocbucket first to give LIFO
4585 * ordering for cache-hot datastructures. Spill over
4586 * into the freebucket if necessary. Alloc will swap
4587 * them if one runs dry.
4588 */
4589 bucket = &cache->uc_allocbucket;
4590 #ifdef NUMA
4591 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4592 PCPU_GET(domain) != itemdomain) {
4593 bucket = &cache->uc_crossbucket;
4594 } else
4595 #endif
4596 if (bucket->ucb_cnt == bucket->ucb_entries &&
4597 cache->uc_freebucket.ucb_cnt <
4598 cache->uc_freebucket.ucb_entries)
4599 cache_bucket_swap(&cache->uc_freebucket,
4600 &cache->uc_allocbucket);
4601 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4602 cache_bucket_push(cache, bucket, item);
4603 critical_exit();
4604 return;
4605 }
4606 } while (cache_free(zone, cache, udata, itemdomain));
4607 critical_exit();
4608
4609 /*
4610 * If nothing else caught this, we'll just do an internal free.
4611 */
4612 zfree_item:
4613 zone_free_item(zone, item, udata, SKIP_DTOR);
4614 }
4615
4616 #ifdef NUMA
4617 /*
4618 * sort crossdomain free buckets to domain correct buckets and cache
4619 * them.
4620 */
4621 static void
zone_free_cross(uma_zone_t zone,uma_bucket_t bucket,void * udata)4622 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4623 {
4624 struct uma_bucketlist emptybuckets, fullbuckets;
4625 uma_zone_domain_t zdom;
4626 uma_bucket_t b;
4627 smr_seq_t seq;
4628 void *item;
4629 int domain;
4630
4631 CTR3(KTR_UMA,
4632 "uma_zfree: zone %s(%p) draining cross bucket %p",
4633 zone->uz_name, zone, bucket);
4634
4635 /*
4636 * It is possible for buckets to arrive here out of order so we fetch
4637 * the current smr seq rather than accepting the bucket's.
4638 */
4639 seq = SMR_SEQ_INVALID;
4640 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4641 seq = smr_advance(zone->uz_smr);
4642
4643 /*
4644 * To avoid having ndomain * ndomain buckets for sorting we have a
4645 * lock on the current crossfree bucket. A full matrix with
4646 * per-domain locking could be used if necessary.
4647 */
4648 STAILQ_INIT(&emptybuckets);
4649 STAILQ_INIT(&fullbuckets);
4650 ZONE_CROSS_LOCK(zone);
4651 for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4652 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4653 domain = item_domain(item);
4654 zdom = ZDOM_GET(zone, domain);
4655 if (zdom->uzd_cross == NULL) {
4656 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4657 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4658 zdom->uzd_cross = b;
4659 } else {
4660 /*
4661 * Avoid allocating a bucket with the cross lock
4662 * held, since allocation can trigger a
4663 * cross-domain free and bucket zones may
4664 * allocate from each other.
4665 */
4666 ZONE_CROSS_UNLOCK(zone);
4667 b = bucket_alloc(zone, udata, M_NOWAIT);
4668 if (b == NULL)
4669 goto out;
4670 ZONE_CROSS_LOCK(zone);
4671 if (zdom->uzd_cross != NULL) {
4672 STAILQ_INSERT_HEAD(&emptybuckets, b,
4673 ub_link);
4674 } else {
4675 zdom->uzd_cross = b;
4676 }
4677 }
4678 }
4679 b = zdom->uzd_cross;
4680 b->ub_bucket[b->ub_cnt++] = item;
4681 b->ub_seq = seq;
4682 if (b->ub_cnt == b->ub_entries) {
4683 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4684 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4685 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4686 zdom->uzd_cross = b;
4687 }
4688 }
4689 ZONE_CROSS_UNLOCK(zone);
4690 out:
4691 if (bucket->ub_cnt == 0)
4692 bucket->ub_seq = SMR_SEQ_INVALID;
4693 bucket_free(zone, bucket, udata);
4694
4695 while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4696 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4697 bucket_free(zone, b, udata);
4698 }
4699 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4700 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4701 domain = item_domain(b->ub_bucket[0]);
4702 zone_put_bucket(zone, domain, b, udata, true);
4703 }
4704 }
4705 #endif
4706
4707 static void
zone_free_bucket(uma_zone_t zone,uma_bucket_t bucket,void * udata,int itemdomain,bool ws)4708 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4709 int itemdomain, bool ws)
4710 {
4711
4712 #ifdef NUMA
4713 /*
4714 * Buckets coming from the wrong domain will be entirely for the
4715 * only other domain on two domain systems. In this case we can
4716 * simply cache them. Otherwise we need to sort them back to
4717 * correct domains.
4718 */
4719 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4720 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4721 zone_free_cross(zone, bucket, udata);
4722 return;
4723 }
4724 #endif
4725
4726 /*
4727 * Attempt to save the bucket in the zone's domain bucket cache.
4728 */
4729 CTR3(KTR_UMA,
4730 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4731 zone->uz_name, zone, bucket);
4732 /* ub_cnt is pointing to the last free item */
4733 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4734 itemdomain = zone_domain_lowest(zone, itemdomain);
4735 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4736 }
4737
4738 /*
4739 * Populate a free or cross bucket for the current cpu cache. Free any
4740 * existing full bucket either to the zone cache or back to the slab layer.
4741 *
4742 * Enters and returns in a critical section. false return indicates that
4743 * we can not satisfy this free in the cache layer. true indicates that
4744 * the caller should retry.
4745 */
4746 static __noinline bool
cache_free(uma_zone_t zone,uma_cache_t cache,void * udata,int itemdomain)4747 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, int itemdomain)
4748 {
4749 uma_cache_bucket_t cbucket;
4750 uma_bucket_t newbucket, bucket;
4751
4752 CRITICAL_ASSERT(curthread);
4753
4754 if (zone->uz_bucket_size == 0)
4755 return false;
4756
4757 cache = &zone->uz_cpu[curcpu];
4758 newbucket = NULL;
4759
4760 /*
4761 * FIRSTTOUCH domains need to free to the correct zdom. When
4762 * enabled this is the zdom of the item. The bucket is the
4763 * cross bucket if the current domain and itemdomain do not match.
4764 */
4765 cbucket = &cache->uc_freebucket;
4766 #ifdef NUMA
4767 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4768 if (PCPU_GET(domain) != itemdomain) {
4769 cbucket = &cache->uc_crossbucket;
4770 if (cbucket->ucb_cnt != 0)
4771 counter_u64_add(zone->uz_xdomain,
4772 cbucket->ucb_cnt);
4773 }
4774 }
4775 #endif
4776 bucket = cache_bucket_unload(cbucket);
4777 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4778 ("cache_free: Entered with non-full free bucket."));
4779
4780 /* We are no longer associated with this CPU. */
4781 critical_exit();
4782
4783 /*
4784 * Don't let SMR zones operate without a free bucket. Force
4785 * a synchronize and re-use this one. We will only degrade
4786 * to a synchronize every bucket_size items rather than every
4787 * item if we fail to allocate a bucket.
4788 */
4789 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4790 if (bucket != NULL)
4791 bucket->ub_seq = smr_advance(zone->uz_smr);
4792 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4793 if (newbucket == NULL && bucket != NULL) {
4794 bucket_drain(zone, bucket);
4795 newbucket = bucket;
4796 bucket = NULL;
4797 }
4798 } else if (!bucketdisable)
4799 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4800
4801 if (bucket != NULL)
4802 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4803
4804 critical_enter();
4805 if ((bucket = newbucket) == NULL)
4806 return (false);
4807 cache = &zone->uz_cpu[curcpu];
4808 #ifdef NUMA
4809 /*
4810 * Check to see if we should be populating the cross bucket. If it
4811 * is already populated we will fall through and attempt to populate
4812 * the free bucket.
4813 */
4814 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4815 if (PCPU_GET(domain) != itemdomain &&
4816 cache->uc_crossbucket.ucb_bucket == NULL) {
4817 cache_bucket_load_cross(cache, bucket);
4818 return (true);
4819 }
4820 }
4821 #endif
4822 /*
4823 * We may have lost the race to fill the bucket or switched CPUs.
4824 */
4825 if (cache->uc_freebucket.ucb_bucket != NULL) {
4826 critical_exit();
4827 bucket_free(zone, bucket, udata);
4828 critical_enter();
4829 } else
4830 cache_bucket_load_free(cache, bucket);
4831
4832 return (true);
4833 }
4834
4835 static void
slab_free_item(uma_zone_t zone,uma_slab_t slab,void * item)4836 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4837 {
4838 uma_keg_t keg;
4839 uma_domain_t dom;
4840 int freei;
4841
4842 keg = zone->uz_keg;
4843 KEG_LOCK_ASSERT(keg, slab->us_domain);
4844
4845 /* Do we need to remove from any lists? */
4846 dom = &keg->uk_domain[slab->us_domain];
4847 if (slab->us_freecount + 1 == keg->uk_ipers) {
4848 LIST_REMOVE(slab, us_link);
4849 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4850 dom->ud_free_slabs++;
4851 } else if (slab->us_freecount == 0) {
4852 LIST_REMOVE(slab, us_link);
4853 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4854 }
4855
4856 /* Slab management. */
4857 freei = slab_item_index(slab, keg, item);
4858 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4859 slab->us_freecount++;
4860
4861 /* Keg statistics. */
4862 dom->ud_free_items++;
4863 }
4864
4865 static void
zone_release(void * arg,void ** bucket,int cnt)4866 zone_release(void *arg, void **bucket, int cnt)
4867 {
4868 struct mtx *lock;
4869 uma_zone_t zone;
4870 uma_slab_t slab;
4871 uma_keg_t keg;
4872 uint8_t *mem;
4873 void *item;
4874 int i;
4875
4876 zone = arg;
4877 keg = zone->uz_keg;
4878 lock = NULL;
4879 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4880 lock = KEG_LOCK(keg, 0);
4881 for (i = 0; i < cnt; i++) {
4882 item = bucket[i];
4883 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4884 slab = vtoslab((vm_offset_t)item);
4885 } else {
4886 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4887 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4888 slab = hash_sfind(&keg->uk_hash, mem);
4889 else
4890 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4891 }
4892 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4893 if (lock != NULL)
4894 mtx_unlock(lock);
4895 lock = KEG_LOCK(keg, slab->us_domain);
4896 }
4897 slab_free_item(zone, slab, item);
4898 }
4899 if (lock != NULL)
4900 mtx_unlock(lock);
4901 }
4902
4903 /*
4904 * Frees a single item to any zone.
4905 *
4906 * Arguments:
4907 * zone The zone to free to
4908 * item The item we're freeing
4909 * udata User supplied data for the dtor
4910 * skip Skip dtors and finis
4911 */
4912 static __noinline void
zone_free_item(uma_zone_t zone,void * item,void * udata,enum zfreeskip skip)4913 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4914 {
4915
4916 /*
4917 * If a free is sent directly to an SMR zone we have to
4918 * synchronize immediately because the item can instantly
4919 * be reallocated. This should only happen in degenerate
4920 * cases when no memory is available for per-cpu caches.
4921 */
4922 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4923 smr_synchronize(zone->uz_smr);
4924
4925 item_dtor(zone, item, zone->uz_size, udata, skip);
4926
4927 if (skip < SKIP_FINI && zone->uz_fini) {
4928 kasan_mark_item_valid(zone, item);
4929 zone->uz_fini(item, zone->uz_size);
4930 kasan_mark_item_invalid(zone, item);
4931 }
4932
4933 zone->uz_release(zone->uz_arg, &item, 1);
4934
4935 if (skip & SKIP_CNT)
4936 return;
4937
4938 counter_u64_add(zone->uz_frees, 1);
4939
4940 if (zone->uz_max_items > 0)
4941 zone_free_limit(zone, 1);
4942 }
4943
4944 /* See uma.h */
4945 int
uma_zone_set_max(uma_zone_t zone,int nitems)4946 uma_zone_set_max(uma_zone_t zone, int nitems)
4947 {
4948
4949 /*
4950 * If the limit is small, we may need to constrain the maximum per-CPU
4951 * cache size, or disable caching entirely.
4952 */
4953 uma_zone_set_maxcache(zone, nitems);
4954
4955 /*
4956 * XXX This can misbehave if the zone has any allocations with
4957 * no limit and a limit is imposed. There is currently no
4958 * way to clear a limit.
4959 */
4960 ZONE_LOCK(zone);
4961 if (zone->uz_max_items == 0)
4962 ZONE_ASSERT_COLD(zone);
4963 zone->uz_max_items = nitems;
4964 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4965 zone_update_caches(zone);
4966 /* We may need to wake waiters. */
4967 wakeup(&zone->uz_max_items);
4968 ZONE_UNLOCK(zone);
4969
4970 return (nitems);
4971 }
4972
4973 /* See uma.h */
4974 void
uma_zone_set_maxcache(uma_zone_t zone,int nitems)4975 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4976 {
4977 int bpcpu, bpdom, bsize, nb;
4978
4979 ZONE_LOCK(zone);
4980
4981 /*
4982 * Compute a lower bound on the number of items that may be cached in
4983 * the zone. Each CPU gets at least two buckets, and for cross-domain
4984 * frees we use an additional bucket per CPU and per domain. Select the
4985 * largest bucket size that does not exceed half of the requested limit,
4986 * with the left over space given to the full bucket cache.
4987 */
4988 bpdom = 0;
4989 bpcpu = 2;
4990 #ifdef NUMA
4991 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
4992 bpcpu++;
4993 bpdom++;
4994 }
4995 #endif
4996 nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
4997 bsize = nitems / nb / 2;
4998 if (bsize > BUCKET_MAX)
4999 bsize = BUCKET_MAX;
5000 else if (bsize == 0 && nitems / nb > 0)
5001 bsize = 1;
5002 zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
5003 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
5004 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
5005 zone->uz_bucket_max = nitems - nb * bsize;
5006 ZONE_UNLOCK(zone);
5007 }
5008
5009 /* See uma.h */
5010 int
uma_zone_get_max(uma_zone_t zone)5011 uma_zone_get_max(uma_zone_t zone)
5012 {
5013 int nitems;
5014
5015 nitems = atomic_load_64(&zone->uz_max_items);
5016
5017 return (nitems);
5018 }
5019
5020 /* See uma.h */
5021 void
uma_zone_set_warning(uma_zone_t zone,const char * warning)5022 uma_zone_set_warning(uma_zone_t zone, const char *warning)
5023 {
5024
5025 ZONE_ASSERT_COLD(zone);
5026 zone->uz_warning = warning;
5027 }
5028
5029 /* See uma.h */
5030 void
uma_zone_set_maxaction(uma_zone_t zone,uma_maxaction_t maxaction)5031 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
5032 {
5033
5034 ZONE_ASSERT_COLD(zone);
5035 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
5036 }
5037
5038 /* See uma.h */
5039 int
uma_zone_get_cur(uma_zone_t zone)5040 uma_zone_get_cur(uma_zone_t zone)
5041 {
5042 int64_t nitems;
5043 u_int i;
5044
5045 nitems = 0;
5046 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
5047 nitems = counter_u64_fetch(zone->uz_allocs) -
5048 counter_u64_fetch(zone->uz_frees);
5049 CPU_FOREACH(i)
5050 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
5051 atomic_load_64(&zone->uz_cpu[i].uc_frees);
5052
5053 return (nitems < 0 ? 0 : nitems);
5054 }
5055
5056 static uint64_t
uma_zone_get_allocs(uma_zone_t zone)5057 uma_zone_get_allocs(uma_zone_t zone)
5058 {
5059 uint64_t nitems;
5060 u_int i;
5061
5062 nitems = 0;
5063 if (zone->uz_allocs != EARLY_COUNTER)
5064 nitems = counter_u64_fetch(zone->uz_allocs);
5065 CPU_FOREACH(i)
5066 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
5067
5068 return (nitems);
5069 }
5070
5071 static uint64_t
uma_zone_get_frees(uma_zone_t zone)5072 uma_zone_get_frees(uma_zone_t zone)
5073 {
5074 uint64_t nitems;
5075 u_int i;
5076
5077 nitems = 0;
5078 if (zone->uz_frees != EARLY_COUNTER)
5079 nitems = counter_u64_fetch(zone->uz_frees);
5080 CPU_FOREACH(i)
5081 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
5082
5083 return (nitems);
5084 }
5085
5086 #ifdef INVARIANTS
5087 /* Used only for KEG_ASSERT_COLD(). */
5088 static uint64_t
uma_keg_get_allocs(uma_keg_t keg)5089 uma_keg_get_allocs(uma_keg_t keg)
5090 {
5091 uma_zone_t z;
5092 uint64_t nitems;
5093
5094 nitems = 0;
5095 LIST_FOREACH(z, &keg->uk_zones, uz_link)
5096 nitems += uma_zone_get_allocs(z);
5097
5098 return (nitems);
5099 }
5100 #endif
5101
5102 /* See uma.h */
5103 void
uma_zone_set_init(uma_zone_t zone,uma_init uminit)5104 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
5105 {
5106 uma_keg_t keg;
5107
5108 KEG_GET(zone, keg);
5109 KEG_ASSERT_COLD(keg);
5110 keg->uk_init = uminit;
5111 }
5112
5113 /* See uma.h */
5114 void
uma_zone_set_fini(uma_zone_t zone,uma_fini fini)5115 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
5116 {
5117 uma_keg_t keg;
5118
5119 KEG_GET(zone, keg);
5120 KEG_ASSERT_COLD(keg);
5121 keg->uk_fini = fini;
5122 }
5123
5124 /* See uma.h */
5125 void
uma_zone_set_zinit(uma_zone_t zone,uma_init zinit)5126 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
5127 {
5128
5129 ZONE_ASSERT_COLD(zone);
5130 zone->uz_init = zinit;
5131 }
5132
5133 /* See uma.h */
5134 void
uma_zone_set_zfini(uma_zone_t zone,uma_fini zfini)5135 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
5136 {
5137
5138 ZONE_ASSERT_COLD(zone);
5139 zone->uz_fini = zfini;
5140 }
5141
5142 /* See uma.h */
5143 void
uma_zone_set_freef(uma_zone_t zone,uma_free freef)5144 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
5145 {
5146 uma_keg_t keg;
5147
5148 KEG_GET(zone, keg);
5149 KEG_ASSERT_COLD(keg);
5150 keg->uk_freef = freef;
5151 }
5152
5153 /* See uma.h */
5154 void
uma_zone_set_allocf(uma_zone_t zone,uma_alloc allocf)5155 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
5156 {
5157 uma_keg_t keg;
5158
5159 KEG_GET(zone, keg);
5160 KEG_ASSERT_COLD(keg);
5161 keg->uk_allocf = allocf;
5162 }
5163
5164 /* See uma.h */
5165 void
uma_zone_set_smr(uma_zone_t zone,smr_t smr)5166 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
5167 {
5168
5169 ZONE_ASSERT_COLD(zone);
5170
5171 KASSERT(smr != NULL, ("Got NULL smr"));
5172 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
5173 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
5174 zone->uz_flags |= UMA_ZONE_SMR;
5175 zone->uz_smr = smr;
5176 zone_update_caches(zone);
5177 }
5178
5179 smr_t
uma_zone_get_smr(uma_zone_t zone)5180 uma_zone_get_smr(uma_zone_t zone)
5181 {
5182
5183 return (zone->uz_smr);
5184 }
5185
5186 /* See uma.h */
5187 void
uma_zone_reserve(uma_zone_t zone,int items)5188 uma_zone_reserve(uma_zone_t zone, int items)
5189 {
5190 uma_keg_t keg;
5191
5192 KEG_GET(zone, keg);
5193 KEG_ASSERT_COLD(keg);
5194 keg->uk_reserve = items;
5195 }
5196
5197 /* See uma.h */
5198 int
uma_zone_reserve_kva(uma_zone_t zone,int count)5199 uma_zone_reserve_kva(uma_zone_t zone, int count)
5200 {
5201 uma_keg_t keg;
5202 vm_offset_t kva;
5203 u_int pages;
5204
5205 KEG_GET(zone, keg);
5206 KEG_ASSERT_COLD(keg);
5207 ZONE_ASSERT_COLD(zone);
5208
5209 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
5210
5211 #ifdef UMA_USE_DMAP
5212 if (keg->uk_ppera > 1) {
5213 #else
5214 if (1) {
5215 #endif
5216 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
5217 if (kva == 0)
5218 return (0);
5219 } else
5220 kva = 0;
5221
5222 MPASS(keg->uk_kva == 0);
5223 keg->uk_kva = kva;
5224 keg->uk_offset = 0;
5225 zone->uz_max_items = pages * keg->uk_ipers;
5226 #ifdef UMA_USE_DMAP
5227 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
5228 #else
5229 keg->uk_allocf = noobj_alloc;
5230 #endif
5231 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5232 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5233 zone_update_caches(zone);
5234
5235 return (1);
5236 }
5237
5238 /* See uma.h */
5239 void
5240 uma_prealloc(uma_zone_t zone, int items)
5241 {
5242 struct vm_domainset_iter di;
5243 uma_domain_t dom;
5244 uma_slab_t slab;
5245 uma_keg_t keg;
5246 int aflags, domain, slabs;
5247
5248 KEG_GET(zone, keg);
5249 slabs = howmany(items, keg->uk_ipers);
5250 while (slabs-- > 0) {
5251 aflags = M_NOWAIT;
5252 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
5253 &aflags);
5254 for (;;) {
5255 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
5256 aflags);
5257 if (slab != NULL) {
5258 dom = &keg->uk_domain[slab->us_domain];
5259 /*
5260 * keg_alloc_slab() always returns a slab on the
5261 * partial list.
5262 */
5263 LIST_REMOVE(slab, us_link);
5264 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
5265 us_link);
5266 dom->ud_free_slabs++;
5267 KEG_UNLOCK(keg, slab->us_domain);
5268 break;
5269 }
5270 if (vm_domainset_iter_policy(&di, &domain) != 0)
5271 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
5272 }
5273 }
5274 }
5275
5276 /*
5277 * Returns a snapshot of memory consumption in bytes.
5278 */
5279 size_t
5280 uma_zone_memory(uma_zone_t zone)
5281 {
5282 size_t sz;
5283 int i;
5284
5285 sz = 0;
5286 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
5287 for (i = 0; i < vm_ndomains; i++)
5288 sz += ZDOM_GET(zone, i)->uzd_nitems;
5289 return (sz * zone->uz_size);
5290 }
5291 for (i = 0; i < vm_ndomains; i++)
5292 sz += zone->uz_keg->uk_domain[i].ud_pages;
5293
5294 return (sz * PAGE_SIZE);
5295 }
5296
5297 struct uma_reclaim_args {
5298 int domain;
5299 int req;
5300 };
5301
5302 static void
5303 uma_reclaim_domain_cb(uma_zone_t zone, void *arg)
5304 {
5305 struct uma_reclaim_args *args;
5306
5307 args = arg;
5308 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) != 0)
5309 return;
5310 if ((args->req == UMA_RECLAIM_TRIM) &&
5311 (zone->uz_flags & UMA_ZONE_NOTRIM) !=0)
5312 return;
5313
5314 uma_zone_reclaim_domain(zone, args->req, args->domain);
5315 }
5316
5317 /* See uma.h */
5318 void
5319 uma_reclaim(int req)
5320 {
5321 uma_reclaim_domain(req, UMA_ANYDOMAIN);
5322 }
5323
5324 void
5325 uma_reclaim_domain(int req, int domain)
5326 {
5327 struct uma_reclaim_args args;
5328
5329 bucket_enable();
5330
5331 args.domain = domain;
5332 args.req = req;
5333
5334 sx_slock(&uma_reclaim_lock);
5335 switch (req) {
5336 case UMA_RECLAIM_TRIM:
5337 case UMA_RECLAIM_DRAIN:
5338 zone_foreach(uma_reclaim_domain_cb, &args);
5339 break;
5340 case UMA_RECLAIM_DRAIN_CPU:
5341 /*
5342 * Reclaim globally visible free items from all zones, then drain
5343 * per-CPU buckets, then reclaim items freed while draining.
5344 * This approach minimizes expensive context switching needed to
5345 * drain each zone's per-CPU buckets.
5346 */
5347 args.req = UMA_RECLAIM_DRAIN;
5348 zone_foreach(uma_reclaim_domain_cb, &args);
5349 pcpu_cache_drain_safe(NULL);
5350 zone_foreach(uma_reclaim_domain_cb, &args);
5351 break;
5352 default:
5353 panic("unhandled reclamation request %d", req);
5354 }
5355
5356 /*
5357 * Some slabs may have been freed but this zone will be visited early
5358 * we visit again so that we can free pages that are empty once other
5359 * zones are drained. We have to do the same for buckets.
5360 */
5361 uma_zone_reclaim_domain(slabzones[0], UMA_RECLAIM_DRAIN, domain);
5362 uma_zone_reclaim_domain(slabzones[1], UMA_RECLAIM_DRAIN, domain);
5363 bucket_zone_drain(domain);
5364 sx_sunlock(&uma_reclaim_lock);
5365 }
5366
5367 static volatile int uma_reclaim_needed;
5368
5369 void
5370 uma_reclaim_wakeup(void)
5371 {
5372
5373 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
5374 wakeup(uma_reclaim);
5375 }
5376
5377 void
5378 uma_reclaim_worker(void *arg __unused)
5379 {
5380
5381 for (;;) {
5382 sx_xlock(&uma_reclaim_lock);
5383 while (atomic_load_int(&uma_reclaim_needed) == 0)
5384 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
5385 hz);
5386 sx_xunlock(&uma_reclaim_lock);
5387 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
5388 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
5389 atomic_store_int(&uma_reclaim_needed, 0);
5390 /* Don't fire more than once per-second. */
5391 pause("umarclslp", hz);
5392 }
5393 }
5394
5395 /* See uma.h */
5396 void
5397 uma_zone_reclaim(uma_zone_t zone, int req)
5398 {
5399 uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN);
5400 }
5401
5402 void
5403 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain)
5404 {
5405 switch (req) {
5406 case UMA_RECLAIM_TRIM:
5407 zone_reclaim(zone, domain, M_NOWAIT, false);
5408 break;
5409 case UMA_RECLAIM_DRAIN:
5410 zone_reclaim(zone, domain, M_NOWAIT, true);
5411 break;
5412 case UMA_RECLAIM_DRAIN_CPU:
5413 pcpu_cache_drain_safe(zone);
5414 zone_reclaim(zone, domain, M_NOWAIT, true);
5415 break;
5416 default:
5417 panic("unhandled reclamation request %d", req);
5418 }
5419 }
5420
5421 /* See uma.h */
5422 int
5423 uma_zone_exhausted(uma_zone_t zone)
5424 {
5425
5426 return (atomic_load_32(&zone->uz_sleepers) > 0);
5427 }
5428
5429 unsigned long
5430 uma_limit(void)
5431 {
5432
5433 return (uma_kmem_limit);
5434 }
5435
5436 void
5437 uma_set_limit(unsigned long limit)
5438 {
5439
5440 uma_kmem_limit = limit;
5441 }
5442
5443 unsigned long
5444 uma_size(void)
5445 {
5446
5447 return (atomic_load_long(&uma_kmem_total));
5448 }
5449
5450 long
5451 uma_avail(void)
5452 {
5453
5454 return (uma_kmem_limit - uma_size());
5455 }
5456
5457 #ifdef DDB
5458 /*
5459 * Generate statistics across both the zone and its per-cpu cache's. Return
5460 * desired statistics if the pointer is non-NULL for that statistic.
5461 *
5462 * Note: does not update the zone statistics, as it can't safely clear the
5463 * per-CPU cache statistic.
5464 *
5465 */
5466 static void
5467 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5468 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5469 {
5470 uma_cache_t cache;
5471 uint64_t allocs, frees, sleeps, xdomain;
5472 int cachefree, cpu;
5473
5474 allocs = frees = sleeps = xdomain = 0;
5475 cachefree = 0;
5476 CPU_FOREACH(cpu) {
5477 cache = &z->uz_cpu[cpu];
5478 cachefree += cache->uc_allocbucket.ucb_cnt;
5479 cachefree += cache->uc_freebucket.ucb_cnt;
5480 xdomain += cache->uc_crossbucket.ucb_cnt;
5481 cachefree += cache->uc_crossbucket.ucb_cnt;
5482 allocs += cache->uc_allocs;
5483 frees += cache->uc_frees;
5484 }
5485 allocs += counter_u64_fetch(z->uz_allocs);
5486 frees += counter_u64_fetch(z->uz_frees);
5487 xdomain += counter_u64_fetch(z->uz_xdomain);
5488 sleeps += z->uz_sleeps;
5489 if (cachefreep != NULL)
5490 *cachefreep = cachefree;
5491 if (allocsp != NULL)
5492 *allocsp = allocs;
5493 if (freesp != NULL)
5494 *freesp = frees;
5495 if (sleepsp != NULL)
5496 *sleepsp = sleeps;
5497 if (xdomainp != NULL)
5498 *xdomainp = xdomain;
5499 }
5500 #endif /* DDB */
5501
5502 static int
5503 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5504 {
5505 uma_keg_t kz;
5506 uma_zone_t z;
5507 int count;
5508
5509 count = 0;
5510 rw_rlock(&uma_rwlock);
5511 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5512 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5513 count++;
5514 }
5515 LIST_FOREACH(z, &uma_cachezones, uz_link)
5516 count++;
5517
5518 rw_runlock(&uma_rwlock);
5519 return (sysctl_handle_int(oidp, &count, 0, req));
5520 }
5521
5522 static void
5523 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5524 struct uma_percpu_stat *ups, bool internal)
5525 {
5526 uma_zone_domain_t zdom;
5527 uma_cache_t cache;
5528 int i;
5529
5530 for (i = 0; i < vm_ndomains; i++) {
5531 zdom = ZDOM_GET(z, i);
5532 uth->uth_zone_free += zdom->uzd_nitems;
5533 }
5534 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5535 uth->uth_frees = counter_u64_fetch(z->uz_frees);
5536 uth->uth_fails = counter_u64_fetch(z->uz_fails);
5537 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5538 uth->uth_sleeps = z->uz_sleeps;
5539
5540 for (i = 0; i < mp_maxid + 1; i++) {
5541 bzero(&ups[i], sizeof(*ups));
5542 if (internal || CPU_ABSENT(i))
5543 continue;
5544 cache = &z->uz_cpu[i];
5545 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5546 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5547 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5548 ups[i].ups_allocs = cache->uc_allocs;
5549 ups[i].ups_frees = cache->uc_frees;
5550 }
5551 }
5552
5553 static int
5554 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5555 {
5556 struct uma_stream_header ush;
5557 struct uma_type_header uth;
5558 struct uma_percpu_stat *ups;
5559 struct sbuf sbuf;
5560 uma_keg_t kz;
5561 uma_zone_t z;
5562 uint64_t items;
5563 uint32_t kfree, pages;
5564 int count, error, i;
5565
5566 error = sysctl_wire_old_buffer(req, 0);
5567 if (error != 0)
5568 return (error);
5569 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5570 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5571 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5572
5573 count = 0;
5574 rw_rlock(&uma_rwlock);
5575 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5576 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5577 count++;
5578 }
5579
5580 LIST_FOREACH(z, &uma_cachezones, uz_link)
5581 count++;
5582
5583 /*
5584 * Insert stream header.
5585 */
5586 bzero(&ush, sizeof(ush));
5587 ush.ush_version = UMA_STREAM_VERSION;
5588 ush.ush_maxcpus = (mp_maxid + 1);
5589 ush.ush_count = count;
5590 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5591
5592 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5593 kfree = pages = 0;
5594 for (i = 0; i < vm_ndomains; i++) {
5595 kfree += kz->uk_domain[i].ud_free_items;
5596 pages += kz->uk_domain[i].ud_pages;
5597 }
5598 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5599 bzero(&uth, sizeof(uth));
5600 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5601 uth.uth_align = kz->uk_align;
5602 uth.uth_size = kz->uk_size;
5603 uth.uth_rsize = kz->uk_rsize;
5604 if (z->uz_max_items > 0) {
5605 items = UZ_ITEMS_COUNT(z->uz_items);
5606 uth.uth_pages = (items / kz->uk_ipers) *
5607 kz->uk_ppera;
5608 } else
5609 uth.uth_pages = pages;
5610 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5611 kz->uk_ppera;
5612 uth.uth_limit = z->uz_max_items;
5613 uth.uth_keg_free = kfree;
5614
5615 /*
5616 * A zone is secondary is it is not the first entry
5617 * on the keg's zone list.
5618 */
5619 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5620 (LIST_FIRST(&kz->uk_zones) != z))
5621 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5622 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5623 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5624 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5625 for (i = 0; i < mp_maxid + 1; i++)
5626 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5627 }
5628 }
5629 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5630 bzero(&uth, sizeof(uth));
5631 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5632 uth.uth_size = z->uz_size;
5633 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5634 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5635 for (i = 0; i < mp_maxid + 1; i++)
5636 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5637 }
5638
5639 rw_runlock(&uma_rwlock);
5640 error = sbuf_finish(&sbuf);
5641 sbuf_delete(&sbuf);
5642 free(ups, M_TEMP);
5643 return (error);
5644 }
5645
5646 int
5647 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5648 {
5649 uma_zone_t zone = *(uma_zone_t *)arg1;
5650 int error, max;
5651
5652 max = uma_zone_get_max(zone);
5653 error = sysctl_handle_int(oidp, &max, 0, req);
5654 if (error || !req->newptr)
5655 return (error);
5656
5657 uma_zone_set_max(zone, max);
5658
5659 return (0);
5660 }
5661
5662 int
5663 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5664 {
5665 uma_zone_t zone;
5666 int cur;
5667
5668 /*
5669 * Some callers want to add sysctls for global zones that
5670 * may not yet exist so they pass a pointer to a pointer.
5671 */
5672 if (arg2 == 0)
5673 zone = *(uma_zone_t *)arg1;
5674 else
5675 zone = arg1;
5676 cur = uma_zone_get_cur(zone);
5677 return (sysctl_handle_int(oidp, &cur, 0, req));
5678 }
5679
5680 static int
5681 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5682 {
5683 uma_zone_t zone = arg1;
5684 uint64_t cur;
5685
5686 cur = uma_zone_get_allocs(zone);
5687 return (sysctl_handle_64(oidp, &cur, 0, req));
5688 }
5689
5690 static int
5691 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5692 {
5693 uma_zone_t zone = arg1;
5694 uint64_t cur;
5695
5696 cur = uma_zone_get_frees(zone);
5697 return (sysctl_handle_64(oidp, &cur, 0, req));
5698 }
5699
5700 static int
5701 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5702 {
5703 struct sbuf sbuf;
5704 uma_zone_t zone = arg1;
5705 int error;
5706
5707 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5708 if (zone->uz_flags != 0)
5709 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5710 else
5711 sbuf_printf(&sbuf, "0");
5712 error = sbuf_finish(&sbuf);
5713 sbuf_delete(&sbuf);
5714
5715 return (error);
5716 }
5717
5718 static int
5719 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5720 {
5721 uma_keg_t keg = arg1;
5722 int avail, effpct, total;
5723
5724 total = keg->uk_ppera * PAGE_SIZE;
5725 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5726 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5727 /*
5728 * We consider the client's requested size and alignment here, not the
5729 * real size determination uk_rsize, because we also adjust the real
5730 * size for internal implementation reasons (max bitset size).
5731 */
5732 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5733 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5734 avail *= mp_maxid + 1;
5735 effpct = 100 * avail / total;
5736 return (sysctl_handle_int(oidp, &effpct, 0, req));
5737 }
5738
5739 static int
5740 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5741 {
5742 uma_zone_t zone = arg1;
5743 uint64_t cur;
5744
5745 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5746 return (sysctl_handle_64(oidp, &cur, 0, req));
5747 }
5748
5749 #ifdef INVARIANTS
5750 static uma_slab_t
5751 uma_dbg_getslab(uma_zone_t zone, void *item)
5752 {
5753 uma_slab_t slab;
5754 uma_keg_t keg;
5755 uint8_t *mem;
5756
5757 /*
5758 * It is safe to return the slab here even though the
5759 * zone is unlocked because the item's allocation state
5760 * essentially holds a reference.
5761 */
5762 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5763 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5764 return (NULL);
5765 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5766 return (vtoslab((vm_offset_t)mem));
5767 keg = zone->uz_keg;
5768 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5769 return ((uma_slab_t)(mem + keg->uk_pgoff));
5770 KEG_LOCK(keg, 0);
5771 slab = hash_sfind(&keg->uk_hash, mem);
5772 KEG_UNLOCK(keg, 0);
5773
5774 return (slab);
5775 }
5776
5777 static bool
5778 uma_dbg_zskip(uma_zone_t zone, void *mem)
5779 {
5780
5781 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5782 return (true);
5783
5784 return (uma_dbg_kskip(zone->uz_keg, mem));
5785 }
5786
5787 static bool
5788 uma_dbg_kskip(uma_keg_t keg, void *mem)
5789 {
5790 uintptr_t idx;
5791
5792 if (dbg_divisor == 0)
5793 return (true);
5794
5795 if (dbg_divisor == 1)
5796 return (false);
5797
5798 idx = (uintptr_t)mem >> PAGE_SHIFT;
5799 if (keg->uk_ipers > 1) {
5800 idx *= keg->uk_ipers;
5801 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5802 }
5803
5804 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5805 counter_u64_add(uma_skip_cnt, 1);
5806 return (true);
5807 }
5808 counter_u64_add(uma_dbg_cnt, 1);
5809
5810 return (false);
5811 }
5812
5813 /*
5814 * Set up the slab's freei data such that uma_dbg_free can function.
5815 *
5816 */
5817 static void
5818 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5819 {
5820 uma_keg_t keg;
5821 int freei;
5822
5823 if (slab == NULL) {
5824 slab = uma_dbg_getslab(zone, item);
5825 if (slab == NULL)
5826 panic("uma: item %p did not belong to zone %s",
5827 item, zone->uz_name);
5828 }
5829 keg = zone->uz_keg;
5830 freei = slab_item_index(slab, keg, item);
5831
5832 if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5833 slab_dbg_bits(slab, keg)))
5834 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5835 item, zone, zone->uz_name, slab, freei);
5836 }
5837
5838 /*
5839 * Verifies freed addresses. Checks for alignment, valid slab membership
5840 * and duplicate frees.
5841 *
5842 */
5843 static void
5844 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5845 {
5846 uma_keg_t keg;
5847 int freei;
5848
5849 if (slab == NULL) {
5850 slab = uma_dbg_getslab(zone, item);
5851 if (slab == NULL)
5852 panic("uma: Freed item %p did not belong to zone %s",
5853 item, zone->uz_name);
5854 }
5855 keg = zone->uz_keg;
5856 freei = slab_item_index(slab, keg, item);
5857
5858 if (freei >= keg->uk_ipers)
5859 panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5860 item, zone, zone->uz_name, slab, freei);
5861
5862 if (slab_item(slab, keg, freei) != item)
5863 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5864 item, zone, zone->uz_name, slab, freei);
5865
5866 if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5867 slab_dbg_bits(slab, keg)))
5868 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5869 item, zone, zone->uz_name, slab, freei);
5870 }
5871 #endif /* INVARIANTS */
5872
5873 #ifdef DDB
5874 static int64_t
5875 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5876 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5877 {
5878 uint64_t frees;
5879 int i;
5880
5881 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5882 *allocs = counter_u64_fetch(z->uz_allocs);
5883 frees = counter_u64_fetch(z->uz_frees);
5884 *sleeps = z->uz_sleeps;
5885 *cachefree = 0;
5886 *xdomain = 0;
5887 } else
5888 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5889 xdomain);
5890 for (i = 0; i < vm_ndomains; i++) {
5891 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5892 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5893 (LIST_FIRST(&kz->uk_zones) != z)))
5894 *cachefree += kz->uk_domain[i].ud_free_items;
5895 }
5896 *used = *allocs - frees;
5897 return (((int64_t)*used + *cachefree) * kz->uk_size);
5898 }
5899
5900 DB_SHOW_COMMAND_FLAGS(uma, db_show_uma, DB_CMD_MEMSAFE)
5901 {
5902 const char *fmt_hdr, *fmt_entry;
5903 uma_keg_t kz;
5904 uma_zone_t z;
5905 uint64_t allocs, used, sleeps, xdomain;
5906 long cachefree;
5907 /* variables for sorting */
5908 uma_keg_t cur_keg;
5909 uma_zone_t cur_zone, last_zone;
5910 int64_t cur_size, last_size, size;
5911 int ties;
5912
5913 /* /i option produces machine-parseable CSV output */
5914 if (modif[0] == 'i') {
5915 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5916 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5917 } else {
5918 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5919 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5920 }
5921
5922 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5923 "Sleeps", "Bucket", "Total Mem", "XFree");
5924
5925 /* Sort the zones with largest size first. */
5926 last_zone = NULL;
5927 last_size = INT64_MAX;
5928 for (;;) {
5929 cur_zone = NULL;
5930 cur_size = -1;
5931 ties = 0;
5932 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5933 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5934 /*
5935 * In the case of size ties, print out zones
5936 * in the order they are encountered. That is,
5937 * when we encounter the most recently output
5938 * zone, we have already printed all preceding
5939 * ties, and we must print all following ties.
5940 */
5941 if (z == last_zone) {
5942 ties = 1;
5943 continue;
5944 }
5945 size = get_uma_stats(kz, z, &allocs, &used,
5946 &sleeps, &cachefree, &xdomain);
5947 if (size > cur_size && size < last_size + ties)
5948 {
5949 cur_size = size;
5950 cur_zone = z;
5951 cur_keg = kz;
5952 }
5953 }
5954 }
5955 if (cur_zone == NULL)
5956 break;
5957
5958 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5959 &sleeps, &cachefree, &xdomain);
5960 db_printf(fmt_entry, cur_zone->uz_name,
5961 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5962 (uintmax_t)allocs, (uintmax_t)sleeps,
5963 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5964 xdomain);
5965
5966 if (db_pager_quit)
5967 return;
5968 last_zone = cur_zone;
5969 last_size = cur_size;
5970 }
5971 }
5972
5973 DB_SHOW_COMMAND_FLAGS(umacache, db_show_umacache, DB_CMD_MEMSAFE)
5974 {
5975 uma_zone_t z;
5976 uint64_t allocs, frees;
5977 long cachefree;
5978 int i;
5979
5980 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5981 "Requests", "Bucket");
5982 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5983 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5984 for (i = 0; i < vm_ndomains; i++)
5985 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5986 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5987 z->uz_name, (uintmax_t)z->uz_size,
5988 (intmax_t)(allocs - frees), cachefree,
5989 (uintmax_t)allocs, z->uz_bucket_size);
5990 if (db_pager_quit)
5991 return;
5992 }
5993 }
5994 #endif /* DDB */
5995