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 void
kasan_mark_item_valid(uma_zone_t zone,void * item)542 kasan_mark_item_valid(uma_zone_t zone, void *item)
543 {
544 void *pcpu_item;
545 size_t sz, rsz;
546 int i;
547
548 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
549 return;
550
551 sz = zone->uz_size;
552 rsz = roundup2(sz, KASAN_SHADOW_SCALE);
553 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
554 kasan_mark(item, sz, rsz, KASAN_GENERIC_REDZONE);
555 } else {
556 pcpu_item = zpcpu_base_to_offset(item);
557 for (i = 0; i <= mp_maxid; i++)
558 kasan_mark(zpcpu_get_cpu(pcpu_item, i), sz, rsz,
559 KASAN_GENERIC_REDZONE);
560 }
561 }
562
563 static void
kasan_mark_item_invalid(uma_zone_t zone,void * item)564 kasan_mark_item_invalid(uma_zone_t zone, void *item)
565 {
566 void *pcpu_item;
567 size_t sz;
568 int i;
569
570 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
571 return;
572
573 sz = roundup2(zone->uz_size, KASAN_SHADOW_SCALE);
574 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
575 kasan_mark(item, 0, sz, KASAN_UMA_FREED);
576 } else {
577 pcpu_item = zpcpu_base_to_offset(item);
578 for (i = 0; i <= mp_maxid; i++)
579 kasan_mark(zpcpu_get_cpu(pcpu_item, i), 0, sz,
580 KASAN_UMA_FREED);
581 }
582 }
583
584 static void
kasan_mark_slab_valid(uma_keg_t keg,void * mem)585 kasan_mark_slab_valid(uma_keg_t keg, void *mem)
586 {
587 size_t sz;
588
589 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
590 sz = keg->uk_ppera * PAGE_SIZE;
591 kasan_mark(mem, sz, sz, 0);
592 }
593 }
594
595 static void
kasan_mark_slab_invalid(uma_keg_t keg,void * mem)596 kasan_mark_slab_invalid(uma_keg_t keg, void *mem)
597 {
598 size_t sz;
599
600 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
601 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
602 sz = keg->uk_ppera * PAGE_SIZE;
603 else
604 sz = keg->uk_pgoff;
605 kasan_mark(mem, 0, sz, KASAN_GENERIC_REDZONE);
606 }
607 }
608 #else /* !KASAN */
609 static void
kasan_mark_item_valid(uma_zone_t zone __unused,void * item __unused)610 kasan_mark_item_valid(uma_zone_t zone __unused, void *item __unused)
611 {
612 }
613
614 static void
kasan_mark_item_invalid(uma_zone_t zone __unused,void * item __unused)615 kasan_mark_item_invalid(uma_zone_t zone __unused, void *item __unused)
616 {
617 }
618
619 static void
kasan_mark_slab_valid(uma_keg_t keg __unused,void * mem __unused)620 kasan_mark_slab_valid(uma_keg_t keg __unused, void *mem __unused)
621 {
622 }
623
624 static void
kasan_mark_slab_invalid(uma_keg_t keg __unused,void * mem __unused)625 kasan_mark_slab_invalid(uma_keg_t keg __unused, void *mem __unused)
626 {
627 }
628 #endif /* KASAN */
629
630 #ifdef KMSAN
631 static inline void
kmsan_mark_item_uninitialized(uma_zone_t zone,void * item)632 kmsan_mark_item_uninitialized(uma_zone_t zone, void *item)
633 {
634 void *pcpu_item;
635 size_t sz;
636 int i;
637
638 if ((zone->uz_flags & (UMA_ZFLAG_CACHE | UMA_ZONE_SECONDARY)) != 0) {
639 /*
640 * Cache zones should not be instrumented by default, as UMA
641 * does not have enough information to do so correctly.
642 * Consumers can mark items themselves if it makes sense to do
643 * so.
644 *
645 * Items from secondary zones are initialized by the parent
646 * zone and thus cannot safely be marked by UMA.
647 */
648 return;
649 }
650 if (zone->uz_keg->uk_init != NULL) {
651 /*
652 * By definition, initialized items cannot be marked. The
653 * best we can do is mark items from these zones after they
654 * are freed to the keg.
655 */
656 return;
657 }
658
659 sz = zone->uz_size;
660 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
661 /* malloc(9) updates the origin map itself. */
662 if ((zone->uz_flags & UMA_ZONE_MALLOC) == 0)
663 kmsan_orig(item, sz, KMSAN_TYPE_UMA, KMSAN_RET_ADDR);
664 kmsan_mark(item, sz, KMSAN_STATE_UNINIT);
665 } else {
666 pcpu_item = zpcpu_base_to_offset(item);
667 for (i = 0; i <= mp_maxid; i++) {
668 kmsan_orig(zpcpu_get_cpu(pcpu_item, i), sz,
669 KMSAN_TYPE_UMA, KMSAN_RET_ADDR);
670 kmsan_mark(zpcpu_get_cpu(pcpu_item, i), sz,
671 KMSAN_STATE_INITED);
672 }
673 }
674 }
675 #else /* !KMSAN */
676 static inline void
kmsan_mark_item_uninitialized(uma_zone_t zone __unused,void * item __unused)677 kmsan_mark_item_uninitialized(uma_zone_t zone __unused, void *item __unused)
678 {
679 }
680 #endif /* KMSAN */
681
682 /*
683 * Acquire the domain lock and record contention.
684 */
685 static uma_zone_domain_t
zone_domain_lock(uma_zone_t zone,int domain)686 zone_domain_lock(uma_zone_t zone, int domain)
687 {
688 uma_zone_domain_t zdom;
689 bool lockfail;
690
691 zdom = ZDOM_GET(zone, domain);
692 lockfail = false;
693 if (ZDOM_OWNED(zdom))
694 lockfail = true;
695 ZDOM_LOCK(zdom);
696 /* This is unsynchronized. The counter does not need to be precise. */
697 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
698 zone->uz_bucket_size++;
699 return (zdom);
700 }
701
702 /*
703 * Search for the domain with the least cached items and return it if it
704 * is out of balance with the preferred domain.
705 */
706 static __noinline int
zone_domain_lowest(uma_zone_t zone,int pref)707 zone_domain_lowest(uma_zone_t zone, int pref)
708 {
709 long least, nitems, prefitems;
710 int domain;
711 int i;
712
713 prefitems = least = LONG_MAX;
714 domain = 0;
715 for (i = 0; i < vm_ndomains; i++) {
716 nitems = ZDOM_GET(zone, i)->uzd_nitems;
717 if (nitems < least) {
718 domain = i;
719 least = nitems;
720 }
721 if (domain == pref)
722 prefitems = nitems;
723 }
724 if (prefitems < least * 2)
725 return (pref);
726
727 return (domain);
728 }
729
730 /*
731 * Search for the domain with the most cached items and return it or the
732 * preferred domain if it has enough to proceed.
733 */
734 static __noinline int
zone_domain_highest(uma_zone_t zone,int pref)735 zone_domain_highest(uma_zone_t zone, int pref)
736 {
737 long most, nitems;
738 int domain;
739 int i;
740
741 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
742 return (pref);
743
744 most = 0;
745 domain = 0;
746 for (i = 0; i < vm_ndomains; i++) {
747 nitems = ZDOM_GET(zone, i)->uzd_nitems;
748 if (nitems > most) {
749 domain = i;
750 most = nitems;
751 }
752 }
753
754 return (domain);
755 }
756
757 /*
758 * Set the maximum imax value.
759 */
760 static void
zone_domain_imax_set(uma_zone_domain_t zdom,int nitems)761 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
762 {
763 long old;
764
765 old = zdom->uzd_imax;
766 do {
767 if (old >= nitems)
768 return;
769 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
770
771 /*
772 * We are at new maximum, so do the last WSS update for the old
773 * bimin and prepare to measure next allocation batch.
774 */
775 if (zdom->uzd_wss < old - zdom->uzd_bimin)
776 zdom->uzd_wss = old - zdom->uzd_bimin;
777 zdom->uzd_bimin = nitems;
778 }
779
780 /*
781 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
782 * zone's caches. If a bucket is found the zone is not locked on return.
783 */
784 static uma_bucket_t
zone_fetch_bucket(uma_zone_t zone,uma_zone_domain_t zdom,bool reclaim)785 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
786 {
787 uma_bucket_t bucket;
788 long cnt;
789 int i;
790 bool dtor = false;
791
792 ZDOM_LOCK_ASSERT(zdom);
793
794 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
795 return (NULL);
796
797 /* SMR Buckets can not be re-used until readers expire. */
798 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
799 bucket->ub_seq != SMR_SEQ_INVALID) {
800 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
801 return (NULL);
802 bucket->ub_seq = SMR_SEQ_INVALID;
803 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
804 if (STAILQ_NEXT(bucket, ub_link) != NULL)
805 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
806 }
807 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
808
809 KASSERT(zdom->uzd_nitems >= bucket->ub_cnt,
810 ("%s: item count underflow (%ld, %d)",
811 __func__, zdom->uzd_nitems, bucket->ub_cnt));
812 KASSERT(bucket->ub_cnt > 0,
813 ("%s: empty bucket in bucket cache", __func__));
814 zdom->uzd_nitems -= bucket->ub_cnt;
815
816 if (reclaim) {
817 /*
818 * Shift the bounds of the current WSS interval to avoid
819 * perturbing the estimates.
820 */
821 cnt = lmin(zdom->uzd_bimin, bucket->ub_cnt);
822 atomic_subtract_long(&zdom->uzd_imax, cnt);
823 zdom->uzd_bimin -= cnt;
824 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
825 if (zdom->uzd_limin >= bucket->ub_cnt) {
826 zdom->uzd_limin -= bucket->ub_cnt;
827 } else {
828 zdom->uzd_limin = 0;
829 zdom->uzd_timin = 0;
830 }
831 } else if (zdom->uzd_bimin > zdom->uzd_nitems) {
832 zdom->uzd_bimin = zdom->uzd_nitems;
833 if (zdom->uzd_imin > zdom->uzd_nitems)
834 zdom->uzd_imin = zdom->uzd_nitems;
835 }
836
837 ZDOM_UNLOCK(zdom);
838 if (dtor)
839 for (i = 0; i < bucket->ub_cnt; i++)
840 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
841 NULL, SKIP_NONE);
842
843 return (bucket);
844 }
845
846 /*
847 * Insert a full bucket into the specified cache. The "ws" parameter indicates
848 * whether the bucket's contents should be counted as part of the zone's working
849 * set. The bucket may be freed if it exceeds the bucket limit.
850 */
851 static void
zone_put_bucket(uma_zone_t zone,int domain,uma_bucket_t bucket,void * udata,const bool ws)852 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
853 const bool ws)
854 {
855 uma_zone_domain_t zdom;
856
857 /* We don't cache empty buckets. This can happen after a reclaim. */
858 if (bucket->ub_cnt == 0)
859 goto out;
860 zdom = zone_domain_lock(zone, domain);
861
862 /*
863 * Conditionally set the maximum number of items.
864 */
865 zdom->uzd_nitems += bucket->ub_cnt;
866 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
867 bool head;
868
869 if (ws) {
870 zone_domain_imax_set(zdom, zdom->uzd_nitems);
871 } else {
872 /*
873 * Shift the bounds of the current WSS interval to
874 * avoid perturbing the estimates.
875 */
876 atomic_add_long(&zdom->uzd_imax, bucket->ub_cnt);
877 zdom->uzd_imin += bucket->ub_cnt;
878 zdom->uzd_bimin += bucket->ub_cnt;
879 zdom->uzd_limin += bucket->ub_cnt;
880 }
881 if (STAILQ_EMPTY(&zdom->uzd_buckets))
882 zdom->uzd_seq = bucket->ub_seq;
883
884 /*
885 * Try to promote reuse of recently used items. For items
886 * protected by SMR, try to defer reuse to minimize polling.
887 * If KASAN is configured, try to defer reuse to improve UAF
888 * detection.
889 */
890 head = bucket->ub_seq == SMR_SEQ_INVALID;
891 #ifdef KASAN
892 head = head && (zone->uz_flags & UMA_ZONE_NOKASAN) != 0;
893 #endif
894 if (head)
895 STAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
896 else
897 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
898 ZDOM_UNLOCK(zdom);
899 return;
900 }
901 zdom->uzd_nitems -= bucket->ub_cnt;
902 ZDOM_UNLOCK(zdom);
903 out:
904 bucket_free(zone, bucket, udata);
905 }
906
907 /* Pops an item out of a per-cpu cache bucket. */
908 static inline void *
cache_bucket_pop(uma_cache_t cache,uma_cache_bucket_t bucket)909 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
910 {
911 void *item;
912
913 CRITICAL_ASSERT(curthread);
914
915 bucket->ucb_cnt--;
916 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
917 #ifdef INVARIANTS
918 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
919 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
920 #endif
921 cache->uc_allocs++;
922
923 return (item);
924 }
925
926 /* Pushes an item into a per-cpu cache bucket. */
927 static inline void
cache_bucket_push(uma_cache_t cache,uma_cache_bucket_t bucket,void * item)928 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
929 {
930
931 CRITICAL_ASSERT(curthread);
932 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
933 ("uma_zfree: Freeing to non free bucket index."));
934
935 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
936 bucket->ucb_cnt++;
937 cache->uc_frees++;
938 }
939
940 /*
941 * Unload a UMA bucket from a per-cpu cache.
942 */
943 static inline uma_bucket_t
cache_bucket_unload(uma_cache_bucket_t bucket)944 cache_bucket_unload(uma_cache_bucket_t bucket)
945 {
946 uma_bucket_t b;
947
948 b = bucket->ucb_bucket;
949 if (b != NULL) {
950 MPASS(b->ub_entries == bucket->ucb_entries);
951 b->ub_cnt = bucket->ucb_cnt;
952 bucket->ucb_bucket = NULL;
953 bucket->ucb_entries = bucket->ucb_cnt = 0;
954 }
955
956 return (b);
957 }
958
959 static inline uma_bucket_t
cache_bucket_unload_alloc(uma_cache_t cache)960 cache_bucket_unload_alloc(uma_cache_t cache)
961 {
962
963 return (cache_bucket_unload(&cache->uc_allocbucket));
964 }
965
966 static inline uma_bucket_t
cache_bucket_unload_free(uma_cache_t cache)967 cache_bucket_unload_free(uma_cache_t cache)
968 {
969
970 return (cache_bucket_unload(&cache->uc_freebucket));
971 }
972
973 static inline uma_bucket_t
cache_bucket_unload_cross(uma_cache_t cache)974 cache_bucket_unload_cross(uma_cache_t cache)
975 {
976
977 return (cache_bucket_unload(&cache->uc_crossbucket));
978 }
979
980 /*
981 * Load a bucket into a per-cpu cache bucket.
982 */
983 static inline void
cache_bucket_load(uma_cache_bucket_t bucket,uma_bucket_t b)984 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
985 {
986
987 CRITICAL_ASSERT(curthread);
988 MPASS(bucket->ucb_bucket == NULL);
989 MPASS(b->ub_seq == SMR_SEQ_INVALID);
990
991 bucket->ucb_bucket = b;
992 bucket->ucb_cnt = b->ub_cnt;
993 bucket->ucb_entries = b->ub_entries;
994 }
995
996 static inline void
cache_bucket_load_alloc(uma_cache_t cache,uma_bucket_t b)997 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
998 {
999
1000 cache_bucket_load(&cache->uc_allocbucket, b);
1001 }
1002
1003 static inline void
cache_bucket_load_free(uma_cache_t cache,uma_bucket_t b)1004 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
1005 {
1006
1007 cache_bucket_load(&cache->uc_freebucket, b);
1008 }
1009
1010 #ifdef NUMA
1011 static inline void
cache_bucket_load_cross(uma_cache_t cache,uma_bucket_t b)1012 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
1013 {
1014
1015 cache_bucket_load(&cache->uc_crossbucket, b);
1016 }
1017 #endif
1018
1019 /*
1020 * Copy and preserve ucb_spare.
1021 */
1022 static inline void
cache_bucket_copy(uma_cache_bucket_t b1,uma_cache_bucket_t b2)1023 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
1024 {
1025
1026 b1->ucb_bucket = b2->ucb_bucket;
1027 b1->ucb_entries = b2->ucb_entries;
1028 b1->ucb_cnt = b2->ucb_cnt;
1029 }
1030
1031 /*
1032 * Swap two cache buckets.
1033 */
1034 static inline void
cache_bucket_swap(uma_cache_bucket_t b1,uma_cache_bucket_t b2)1035 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
1036 {
1037 struct uma_cache_bucket b3;
1038
1039 CRITICAL_ASSERT(curthread);
1040
1041 cache_bucket_copy(&b3, b1);
1042 cache_bucket_copy(b1, b2);
1043 cache_bucket_copy(b2, &b3);
1044 }
1045
1046 /*
1047 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
1048 */
1049 static uma_bucket_t
cache_fetch_bucket(uma_zone_t zone,uma_cache_t cache,int domain)1050 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
1051 {
1052 uma_zone_domain_t zdom;
1053 uma_bucket_t bucket;
1054 smr_seq_t seq;
1055
1056 /*
1057 * Avoid the lock if possible.
1058 */
1059 zdom = ZDOM_GET(zone, domain);
1060 if (zdom->uzd_nitems == 0)
1061 return (NULL);
1062
1063 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
1064 (seq = atomic_load_32(&zdom->uzd_seq)) != SMR_SEQ_INVALID &&
1065 !smr_poll(zone->uz_smr, seq, false))
1066 return (NULL);
1067
1068 /*
1069 * Check the zone's cache of buckets.
1070 */
1071 zdom = zone_domain_lock(zone, domain);
1072 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL)
1073 return (bucket);
1074 ZDOM_UNLOCK(zdom);
1075
1076 return (NULL);
1077 }
1078
1079 static void
zone_log_warning(uma_zone_t zone)1080 zone_log_warning(uma_zone_t zone)
1081 {
1082 static const struct timeval warninterval = { 300, 0 };
1083
1084 if (!zone_warnings || zone->uz_warning == NULL)
1085 return;
1086
1087 if (ratecheck(&zone->uz_ratecheck, &warninterval))
1088 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
1089 }
1090
1091 static inline void
zone_maxaction(uma_zone_t zone)1092 zone_maxaction(uma_zone_t zone)
1093 {
1094
1095 if (zone->uz_maxaction.ta_func != NULL)
1096 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
1097 }
1098
1099 /*
1100 * Routine called by timeout which is used to fire off some time interval
1101 * based calculations. (stats, hash size, etc.)
1102 *
1103 * Arguments:
1104 * arg Unused
1105 *
1106 * Returns:
1107 * Nothing
1108 */
1109 static void
uma_timeout(void * context __unused,int pending __unused)1110 uma_timeout(void *context __unused, int pending __unused)
1111 {
1112 bucket_enable();
1113 zone_foreach(zone_timeout, NULL);
1114
1115 /* Reschedule this event */
1116 taskqueue_enqueue_timeout(taskqueue_thread, &uma_timeout_task,
1117 UMA_TIMEOUT * hz);
1118 }
1119
1120 /*
1121 * Update the working set size estimates for the zone's bucket cache.
1122 * The constants chosen here are somewhat arbitrary.
1123 */
1124 static void
zone_domain_update_wss(uma_zone_domain_t zdom)1125 zone_domain_update_wss(uma_zone_domain_t zdom)
1126 {
1127 long m;
1128
1129 ZDOM_LOCK_ASSERT(zdom);
1130 MPASS(zdom->uzd_imax >= zdom->uzd_nitems);
1131 MPASS(zdom->uzd_nitems >= zdom->uzd_bimin);
1132 MPASS(zdom->uzd_bimin >= zdom->uzd_imin);
1133
1134 /*
1135 * Estimate WSS as modified moving average of biggest allocation
1136 * batches for each period over few minutes (UMA_TIMEOUT of 20s).
1137 */
1138 zdom->uzd_wss = lmax(zdom->uzd_wss * 3 / 4,
1139 zdom->uzd_imax - zdom->uzd_bimin);
1140
1141 /*
1142 * Estimate longtime minimum item count as a combination of recent
1143 * minimum item count, adjusted by WSS for safety, and the modified
1144 * moving average over the last several hours (UMA_TIMEOUT of 20s).
1145 * timin measures time since limin tried to go negative, that means
1146 * we were dangerously close to or got out of cache.
1147 */
1148 m = zdom->uzd_imin - zdom->uzd_wss;
1149 if (m >= 0) {
1150 if (zdom->uzd_limin >= m)
1151 zdom->uzd_limin = m;
1152 else
1153 zdom->uzd_limin = (m + zdom->uzd_limin * 255) / 256;
1154 zdom->uzd_timin++;
1155 } else {
1156 zdom->uzd_limin = 0;
1157 zdom->uzd_timin = 0;
1158 }
1159
1160 /* To reduce period edge effects on WSS keep half of the imax. */
1161 atomic_subtract_long(&zdom->uzd_imax,
1162 (zdom->uzd_imax - zdom->uzd_nitems + 1) / 2);
1163 zdom->uzd_imin = zdom->uzd_bimin = zdom->uzd_nitems;
1164 }
1165
1166 /*
1167 * Routine to perform timeout driven calculations. This expands the
1168 * hashes and does per cpu statistics aggregation.
1169 *
1170 * Returns nothing.
1171 */
1172 static void
zone_timeout(uma_zone_t zone,void * unused)1173 zone_timeout(uma_zone_t zone, void *unused)
1174 {
1175 uma_keg_t keg;
1176 u_int slabs, pages;
1177
1178 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
1179 goto trim;
1180
1181 keg = zone->uz_keg;
1182
1183 /*
1184 * Hash zones are non-numa by definition so the first domain
1185 * is the only one present.
1186 */
1187 KEG_LOCK(keg, 0);
1188 pages = keg->uk_domain[0].ud_pages;
1189
1190 /*
1191 * Expand the keg hash table.
1192 *
1193 * This is done if the number of slabs is larger than the hash size.
1194 * What I'm trying to do here is completely reduce collisions. This
1195 * may be a little aggressive. Should I allow for two collisions max?
1196 */
1197 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1198 struct uma_hash newhash;
1199 struct uma_hash oldhash;
1200 int ret;
1201
1202 /*
1203 * This is so involved because allocating and freeing
1204 * while the keg lock is held will lead to deadlock.
1205 * I have to do everything in stages and check for
1206 * races.
1207 */
1208 KEG_UNLOCK(keg, 0);
1209 ret = hash_alloc(&newhash, 1 << fls(slabs));
1210 KEG_LOCK(keg, 0);
1211 if (ret) {
1212 if (hash_expand(&keg->uk_hash, &newhash)) {
1213 oldhash = keg->uk_hash;
1214 keg->uk_hash = newhash;
1215 } else
1216 oldhash = newhash;
1217
1218 KEG_UNLOCK(keg, 0);
1219 hash_free(&oldhash);
1220 goto trim;
1221 }
1222 }
1223 KEG_UNLOCK(keg, 0);
1224
1225 trim:
1226 /* Trim caches not used for a long time. */
1227 if ((zone->uz_flags & (UMA_ZONE_UNMANAGED | UMA_ZONE_NOTRIM)) == 0) {
1228 for (int i = 0; i < vm_ndomains; i++) {
1229 if (bucket_cache_reclaim_domain(zone, false, false, i) &&
1230 (zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1231 keg_drain(zone->uz_keg, i);
1232 }
1233 }
1234 }
1235
1236 /*
1237 * Allocate and zero fill the next sized hash table from the appropriate
1238 * backing store.
1239 *
1240 * Arguments:
1241 * hash A new hash structure with the old hash size in uh_hashsize
1242 *
1243 * Returns:
1244 * 1 on success and 0 on failure.
1245 */
1246 static int
hash_alloc(struct uma_hash * hash,u_int size)1247 hash_alloc(struct uma_hash *hash, u_int size)
1248 {
1249 size_t alloc;
1250
1251 KASSERT(powerof2(size), ("hash size must be power of 2"));
1252 if (size > UMA_HASH_SIZE_INIT) {
1253 hash->uh_hashsize = size;
1254 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1255 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1256 } else {
1257 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1258 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1259 UMA_ANYDOMAIN, M_WAITOK);
1260 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1261 }
1262 if (hash->uh_slab_hash) {
1263 bzero(hash->uh_slab_hash, alloc);
1264 hash->uh_hashmask = hash->uh_hashsize - 1;
1265 return (1);
1266 }
1267
1268 return (0);
1269 }
1270
1271 /*
1272 * Expands the hash table for HASH zones. This is done from zone_timeout
1273 * to reduce collisions. This must not be done in the regular allocation
1274 * path, otherwise, we can recurse on the vm while allocating pages.
1275 *
1276 * Arguments:
1277 * oldhash The hash you want to expand
1278 * newhash The hash structure for the new table
1279 *
1280 * Returns:
1281 * Nothing
1282 *
1283 * Discussion:
1284 */
1285 static int
hash_expand(struct uma_hash * oldhash,struct uma_hash * newhash)1286 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1287 {
1288 uma_hash_slab_t slab;
1289 u_int hval;
1290 u_int idx;
1291
1292 if (!newhash->uh_slab_hash)
1293 return (0);
1294
1295 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1296 return (0);
1297
1298 /*
1299 * I need to investigate hash algorithms for resizing without a
1300 * full rehash.
1301 */
1302
1303 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1304 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1305 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1306 LIST_REMOVE(slab, uhs_hlink);
1307 hval = UMA_HASH(newhash, slab->uhs_data);
1308 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1309 slab, uhs_hlink);
1310 }
1311
1312 return (1);
1313 }
1314
1315 /*
1316 * Free the hash bucket to the appropriate backing store.
1317 *
1318 * Arguments:
1319 * slab_hash The hash bucket we're freeing
1320 * hashsize The number of entries in that hash bucket
1321 *
1322 * Returns:
1323 * Nothing
1324 */
1325 static void
hash_free(struct uma_hash * hash)1326 hash_free(struct uma_hash *hash)
1327 {
1328 if (hash->uh_slab_hash == NULL)
1329 return;
1330 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1331 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1332 else
1333 free(hash->uh_slab_hash, M_UMAHASH);
1334 }
1335
1336 /*
1337 * Frees all outstanding items in a bucket
1338 *
1339 * Arguments:
1340 * zone The zone to free to, must be unlocked.
1341 * bucket The free/alloc bucket with items.
1342 *
1343 * Returns:
1344 * Nothing
1345 */
1346 static void
bucket_drain(uma_zone_t zone,uma_bucket_t bucket)1347 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1348 {
1349 int i;
1350
1351 if (bucket->ub_cnt == 0)
1352 return;
1353
1354 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1355 bucket->ub_seq != SMR_SEQ_INVALID) {
1356 smr_wait(zone->uz_smr, bucket->ub_seq);
1357 bucket->ub_seq = SMR_SEQ_INVALID;
1358 for (i = 0; i < bucket->ub_cnt; i++)
1359 item_dtor(zone, bucket->ub_bucket[i],
1360 zone->uz_size, NULL, SKIP_NONE);
1361 }
1362 if (zone->uz_fini)
1363 for (i = 0; i < bucket->ub_cnt; i++) {
1364 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
1365 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1366 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
1367 }
1368 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1369 if (zone->uz_max_items > 0)
1370 zone_free_limit(zone, bucket->ub_cnt);
1371 #ifdef INVARIANTS
1372 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1373 #endif
1374 bucket->ub_cnt = 0;
1375 }
1376
1377 /*
1378 * Drains the per cpu caches for a zone.
1379 *
1380 * NOTE: This may only be called while the zone is being torn down, and not
1381 * during normal operation. This is necessary in order that we do not have
1382 * to migrate CPUs to drain the per-CPU caches.
1383 *
1384 * Arguments:
1385 * zone The zone to drain, must be unlocked.
1386 *
1387 * Returns:
1388 * Nothing
1389 */
1390 static void
cache_drain(uma_zone_t zone)1391 cache_drain(uma_zone_t zone)
1392 {
1393 uma_cache_t cache;
1394 uma_bucket_t bucket;
1395 smr_seq_t seq;
1396 int cpu;
1397
1398 /*
1399 * XXX: It is safe to not lock the per-CPU caches, because we're
1400 * tearing down the zone anyway. I.e., there will be no further use
1401 * of the caches at this point.
1402 *
1403 * XXX: It would good to be able to assert that the zone is being
1404 * torn down to prevent improper use of cache_drain().
1405 */
1406 seq = SMR_SEQ_INVALID;
1407 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1408 seq = smr_advance(zone->uz_smr);
1409 CPU_FOREACH(cpu) {
1410 cache = &zone->uz_cpu[cpu];
1411 bucket = cache_bucket_unload_alloc(cache);
1412 if (bucket != NULL)
1413 bucket_free(zone, bucket, NULL);
1414 bucket = cache_bucket_unload_free(cache);
1415 if (bucket != NULL) {
1416 bucket->ub_seq = seq;
1417 bucket_free(zone, bucket, NULL);
1418 }
1419 bucket = cache_bucket_unload_cross(cache);
1420 if (bucket != NULL) {
1421 bucket->ub_seq = seq;
1422 bucket_free(zone, bucket, NULL);
1423 }
1424 }
1425 bucket_cache_reclaim(zone, true, UMA_ANYDOMAIN);
1426 }
1427
1428 static void
cache_shrink(uma_zone_t zone,void * unused)1429 cache_shrink(uma_zone_t zone, void *unused)
1430 {
1431
1432 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1433 return;
1434
1435 ZONE_LOCK(zone);
1436 zone->uz_bucket_size =
1437 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1438 ZONE_UNLOCK(zone);
1439 }
1440
1441 static void
cache_drain_safe_cpu(uma_zone_t zone,void * unused)1442 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1443 {
1444 uma_cache_t cache;
1445 uma_bucket_t b1, b2, b3;
1446 int domain;
1447
1448 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1449 return;
1450
1451 b1 = b2 = b3 = NULL;
1452 critical_enter();
1453 cache = &zone->uz_cpu[curcpu];
1454 domain = PCPU_GET(domain);
1455 b1 = cache_bucket_unload_alloc(cache);
1456
1457 /*
1458 * Don't flush SMR zone buckets. This leaves the zone without a
1459 * bucket and forces every free to synchronize().
1460 */
1461 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1462 b2 = cache_bucket_unload_free(cache);
1463 b3 = cache_bucket_unload_cross(cache);
1464 }
1465 critical_exit();
1466
1467 if (b1 != NULL)
1468 zone_free_bucket(zone, b1, NULL, domain, false);
1469 if (b2 != NULL)
1470 zone_free_bucket(zone, b2, NULL, domain, false);
1471 if (b3 != NULL) {
1472 /* Adjust the domain so it goes to zone_free_cross. */
1473 domain = (domain + 1) % vm_ndomains;
1474 zone_free_bucket(zone, b3, NULL, domain, false);
1475 }
1476 }
1477
1478 /*
1479 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1480 * This is an expensive call because it needs to bind to all CPUs
1481 * one by one and enter a critical section on each of them in order
1482 * to safely access their cache buckets.
1483 * Zone lock must not be held on call this function.
1484 */
1485 static void
pcpu_cache_drain_safe(uma_zone_t zone)1486 pcpu_cache_drain_safe(uma_zone_t zone)
1487 {
1488 int cpu;
1489
1490 /*
1491 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1492 */
1493 if (zone)
1494 cache_shrink(zone, NULL);
1495 else
1496 zone_foreach(cache_shrink, NULL);
1497
1498 CPU_FOREACH(cpu) {
1499 thread_lock(curthread);
1500 sched_bind(curthread, cpu);
1501 thread_unlock(curthread);
1502
1503 if (zone)
1504 cache_drain_safe_cpu(zone, NULL);
1505 else
1506 zone_foreach(cache_drain_safe_cpu, NULL);
1507 }
1508 thread_lock(curthread);
1509 sched_unbind(curthread);
1510 thread_unlock(curthread);
1511 }
1512
1513 /*
1514 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1515 * requested a drain, otherwise the per-domain caches are trimmed to either
1516 * estimated working set size.
1517 */
1518 static bool
bucket_cache_reclaim_domain(uma_zone_t zone,bool drain,bool trim,int domain)1519 bucket_cache_reclaim_domain(uma_zone_t zone, bool drain, bool trim, int domain)
1520 {
1521 uma_zone_domain_t zdom;
1522 uma_bucket_t bucket;
1523 long target;
1524 bool done = false;
1525
1526 /*
1527 * The cross bucket is partially filled and not part of
1528 * the item count. Reclaim it individually here.
1529 */
1530 zdom = ZDOM_GET(zone, domain);
1531 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1532 ZONE_CROSS_LOCK(zone);
1533 bucket = zdom->uzd_cross;
1534 zdom->uzd_cross = NULL;
1535 ZONE_CROSS_UNLOCK(zone);
1536 if (bucket != NULL)
1537 bucket_free(zone, bucket, NULL);
1538 }
1539
1540 /*
1541 * If we were asked to drain the zone, we are done only once
1542 * this bucket cache is empty. If trim, we reclaim items in
1543 * excess of the zone's estimated working set size. Multiple
1544 * consecutive calls will shrink the WSS and so reclaim more.
1545 * If neither drain nor trim, then voluntarily reclaim 1/4
1546 * (to reduce first spike) of items not used for a long time.
1547 */
1548 ZDOM_LOCK(zdom);
1549 zone_domain_update_wss(zdom);
1550 if (drain)
1551 target = 0;
1552 else if (trim)
1553 target = zdom->uzd_wss;
1554 else if (zdom->uzd_timin > 900 / UMA_TIMEOUT)
1555 target = zdom->uzd_nitems - zdom->uzd_limin / 4;
1556 else {
1557 ZDOM_UNLOCK(zdom);
1558 return (done);
1559 }
1560 while ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) != NULL &&
1561 zdom->uzd_nitems >= target + bucket->ub_cnt) {
1562 bucket = zone_fetch_bucket(zone, zdom, true);
1563 if (bucket == NULL)
1564 break;
1565 bucket_free(zone, bucket, NULL);
1566 done = true;
1567 ZDOM_LOCK(zdom);
1568 }
1569 ZDOM_UNLOCK(zdom);
1570 return (done);
1571 }
1572
1573 static void
bucket_cache_reclaim(uma_zone_t zone,bool drain,int domain)1574 bucket_cache_reclaim(uma_zone_t zone, bool drain, int domain)
1575 {
1576 int i;
1577
1578 /*
1579 * Shrink the zone bucket size to ensure that the per-CPU caches
1580 * don't grow too large.
1581 */
1582 if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1583 zone->uz_bucket_size--;
1584
1585 if (domain != UMA_ANYDOMAIN &&
1586 (zone->uz_flags & UMA_ZONE_ROUNDROBIN) == 0) {
1587 bucket_cache_reclaim_domain(zone, drain, true, domain);
1588 } else {
1589 for (i = 0; i < vm_ndomains; i++)
1590 bucket_cache_reclaim_domain(zone, drain, true, i);
1591 }
1592 }
1593
1594 static void
keg_free_slab(uma_keg_t keg,uma_slab_t slab,int start)1595 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1596 {
1597 uint8_t *mem;
1598 size_t size;
1599 int i;
1600 uint8_t flags;
1601
1602 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1603 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1604
1605 mem = slab_data(slab, keg);
1606 size = PAGE_SIZE * keg->uk_ppera;
1607
1608 kasan_mark_slab_valid(keg, mem);
1609 if (keg->uk_fini != NULL) {
1610 for (i = start - 1; i > -1; i--)
1611 #ifdef INVARIANTS
1612 /*
1613 * trash_fini implies that dtor was trash_dtor. trash_fini
1614 * would check that memory hasn't been modified since free,
1615 * which executed trash_dtor.
1616 * That's why we need to run uma_dbg_kskip() check here,
1617 * albeit we don't make skip check for other init/fini
1618 * invocations.
1619 */
1620 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1621 keg->uk_fini != trash_fini)
1622 #endif
1623 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1624 }
1625 flags = slab->us_flags;
1626 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1627 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1628 NULL, SKIP_NONE);
1629 }
1630 keg->uk_freef(mem, size, flags);
1631 uma_total_dec(size);
1632 }
1633
1634 static void
keg_drain_domain(uma_keg_t keg,int domain)1635 keg_drain_domain(uma_keg_t keg, int domain)
1636 {
1637 struct slabhead freeslabs;
1638 uma_domain_t dom;
1639 uma_slab_t slab, tmp;
1640 uint32_t i, stofree, stokeep, partial;
1641
1642 dom = &keg->uk_domain[domain];
1643 LIST_INIT(&freeslabs);
1644
1645 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1646 keg->uk_name, keg, domain, dom->ud_free_items);
1647
1648 KEG_LOCK(keg, domain);
1649
1650 /*
1651 * Are the free items in partially allocated slabs sufficient to meet
1652 * the reserve? If not, compute the number of fully free slabs that must
1653 * be kept.
1654 */
1655 partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers;
1656 if (partial < keg->uk_reserve) {
1657 stokeep = min(dom->ud_free_slabs,
1658 howmany(keg->uk_reserve - partial, keg->uk_ipers));
1659 } else {
1660 stokeep = 0;
1661 }
1662 stofree = dom->ud_free_slabs - stokeep;
1663
1664 /*
1665 * Partition the free slabs into two sets: those that must be kept in
1666 * order to maintain the reserve, and those that may be released back to
1667 * the system. Since one set may be much larger than the other,
1668 * populate the smaller of the two sets and swap them if necessary.
1669 */
1670 for (i = min(stofree, stokeep); i > 0; i--) {
1671 slab = LIST_FIRST(&dom->ud_free_slab);
1672 LIST_REMOVE(slab, us_link);
1673 LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1674 }
1675 if (stofree > stokeep)
1676 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1677
1678 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1679 LIST_FOREACH(slab, &freeslabs, us_link)
1680 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1681 }
1682 dom->ud_free_items -= stofree * keg->uk_ipers;
1683 dom->ud_free_slabs -= stofree;
1684 dom->ud_pages -= stofree * keg->uk_ppera;
1685 KEG_UNLOCK(keg, domain);
1686
1687 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1688 keg_free_slab(keg, slab, keg->uk_ipers);
1689 }
1690
1691 /*
1692 * Frees pages from a keg back to the system. This is done on demand from
1693 * the pageout daemon.
1694 *
1695 * Returns nothing.
1696 */
1697 static void
keg_drain(uma_keg_t keg,int domain)1698 keg_drain(uma_keg_t keg, int domain)
1699 {
1700 int i;
1701
1702 if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0)
1703 return;
1704 if (domain != UMA_ANYDOMAIN) {
1705 keg_drain_domain(keg, domain);
1706 } else {
1707 for (i = 0; i < vm_ndomains; i++)
1708 keg_drain_domain(keg, i);
1709 }
1710 }
1711
1712 static void
zone_reclaim(uma_zone_t zone,int domain,int waitok,bool drain)1713 zone_reclaim(uma_zone_t zone, int domain, int waitok, bool drain)
1714 {
1715 /*
1716 * Count active reclaim operations in order to interlock with
1717 * zone_dtor(), which removes the zone from global lists before
1718 * attempting to reclaim items itself.
1719 *
1720 * The zone may be destroyed while sleeping, so only zone_dtor() should
1721 * specify M_WAITOK.
1722 */
1723 ZONE_LOCK(zone);
1724 if (waitok == M_WAITOK) {
1725 while (zone->uz_reclaimers > 0)
1726 msleep(zone, ZONE_LOCKPTR(zone), PVM, "zonedrain", 1);
1727 }
1728 zone->uz_reclaimers++;
1729 ZONE_UNLOCK(zone);
1730 bucket_cache_reclaim(zone, drain, domain);
1731
1732 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1733 keg_drain(zone->uz_keg, domain);
1734 ZONE_LOCK(zone);
1735 zone->uz_reclaimers--;
1736 if (zone->uz_reclaimers == 0)
1737 wakeup(zone);
1738 ZONE_UNLOCK(zone);
1739 }
1740
1741 /*
1742 * Allocate a new slab for a keg and inserts it into the partial slab list.
1743 * The keg should be unlocked on entry. If the allocation succeeds it will
1744 * be locked on return.
1745 *
1746 * Arguments:
1747 * flags Wait flags for the item initialization routine
1748 * aflags Wait flags for the slab allocation
1749 *
1750 * Returns:
1751 * The slab that was allocated or NULL if there is no memory and the
1752 * caller specified M_NOWAIT.
1753 */
1754 static uma_slab_t
keg_alloc_slab(uma_keg_t keg,uma_zone_t zone,int domain,int flags,int aflags)1755 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1756 int aflags)
1757 {
1758 uma_domain_t dom;
1759 uma_slab_t slab;
1760 unsigned long size;
1761 uint8_t *mem;
1762 uint8_t sflags;
1763 int i;
1764
1765 TSENTER();
1766
1767 KASSERT(domain >= 0 && domain < vm_ndomains,
1768 ("keg_alloc_slab: domain %d out of range", domain));
1769
1770 slab = NULL;
1771 mem = NULL;
1772 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1773 uma_hash_slab_t hslab;
1774 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1775 domain, aflags);
1776 if (hslab == NULL)
1777 goto fail;
1778 slab = &hslab->uhs_slab;
1779 }
1780
1781 /*
1782 * This reproduces the old vm_zone behavior of zero filling pages the
1783 * first time they are added to a zone.
1784 *
1785 * Malloced items are zeroed in uma_zalloc.
1786 */
1787
1788 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1789 aflags |= M_ZERO;
1790 else
1791 aflags &= ~M_ZERO;
1792
1793 if (keg->uk_flags & UMA_ZONE_NODUMP)
1794 aflags |= M_NODUMP;
1795
1796 if (keg->uk_flags & UMA_ZONE_NOFREE)
1797 aflags |= M_NEVERFREED;
1798
1799 /* zone is passed for legacy reasons. */
1800 size = keg->uk_ppera * PAGE_SIZE;
1801 mem = keg->uk_allocf(zone, size, domain, &sflags, aflags);
1802 if (mem == NULL) {
1803 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1804 zone_free_item(slabzone(keg->uk_ipers),
1805 slab_tohashslab(slab), NULL, SKIP_NONE);
1806 goto fail;
1807 }
1808 uma_total_inc(size);
1809
1810 /* For HASH zones all pages go to the same uma_domain. */
1811 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1812 domain = 0;
1813
1814 kmsan_mark(mem, size,
1815 (aflags & M_ZERO) != 0 ? KMSAN_STATE_INITED : KMSAN_STATE_UNINIT);
1816
1817 /* Point the slab into the allocated memory */
1818 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1819 slab = (uma_slab_t)(mem + keg->uk_pgoff);
1820 else
1821 slab_tohashslab(slab)->uhs_data = mem;
1822
1823 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1824 for (i = 0; i < keg->uk_ppera; i++)
1825 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1826 zone, slab);
1827
1828 slab->us_freecount = keg->uk_ipers;
1829 slab->us_flags = sflags;
1830 slab->us_domain = domain;
1831
1832 BIT_FILL(keg->uk_ipers, &slab->us_free);
1833 #ifdef INVARIANTS
1834 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1835 #endif
1836
1837 if (keg->uk_init != NULL) {
1838 for (i = 0; i < keg->uk_ipers; i++)
1839 if (keg->uk_init(slab_item(slab, keg, i),
1840 keg->uk_size, flags) != 0)
1841 break;
1842 if (i != keg->uk_ipers) {
1843 keg_free_slab(keg, slab, i);
1844 goto fail;
1845 }
1846 }
1847 kasan_mark_slab_invalid(keg, mem);
1848 KEG_LOCK(keg, domain);
1849
1850 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1851 slab, keg->uk_name, keg);
1852
1853 if (keg->uk_flags & UMA_ZFLAG_HASH)
1854 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1855
1856 /*
1857 * If we got a slab here it's safe to mark it partially used
1858 * and return. We assume that the caller is going to remove
1859 * at least one item.
1860 */
1861 dom = &keg->uk_domain[domain];
1862 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1863 dom->ud_pages += keg->uk_ppera;
1864 dom->ud_free_items += keg->uk_ipers;
1865
1866 TSEXIT();
1867 return (slab);
1868
1869 fail:
1870 return (NULL);
1871 }
1872
1873 /*
1874 * This function is intended to be used early on in place of page_alloc(). It
1875 * performs contiguous physical memory allocations and uses a bump allocator for
1876 * KVA, so is usable before the kernel map is initialized.
1877 */
1878 static void *
startup_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)1879 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1880 int wait)
1881 {
1882 vm_paddr_t pa;
1883 vm_page_t m;
1884 int i, pages;
1885
1886 pages = howmany(bytes, PAGE_SIZE);
1887 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1888
1889 *pflag = UMA_SLAB_BOOT;
1890 m = vm_page_alloc_noobj_contig_domain(domain, malloc2vm_flags(wait) |
1891 VM_ALLOC_WIRED, pages, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
1892 VM_MEMATTR_DEFAULT);
1893 if (m == NULL)
1894 return (NULL);
1895
1896 pa = VM_PAGE_TO_PHYS(m);
1897 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1898 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING
1899 if ((wait & M_NODUMP) == 0)
1900 dump_add_page(pa);
1901 #endif
1902 }
1903
1904 /* Allocate KVA and indirectly advance bootmem. */
1905 return (pmap_map(&bootmem, m->phys_addr,
1906 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE));
1907 }
1908
1909 static void
startup_free(void * mem,vm_size_t bytes)1910 startup_free(void *mem, vm_size_t bytes)
1911 {
1912 vm_offset_t va;
1913 vm_page_t m;
1914
1915 va = (vm_offset_t)mem;
1916 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1917
1918 /*
1919 * startup_alloc() returns direct-mapped slabs on some platforms. Avoid
1920 * unmapping ranges of the direct map.
1921 */
1922 if (va >= bootstart && va + bytes <= bootmem)
1923 pmap_remove(kernel_pmap, va, va + bytes);
1924 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1925 #if MINIDUMP_PAGE_TRACKING && MINIDUMP_STARTUP_PAGE_TRACKING
1926 dump_drop_page(VM_PAGE_TO_PHYS(m));
1927 #endif
1928 vm_page_unwire_noq(m);
1929 vm_page_free(m);
1930 }
1931 }
1932
1933 /*
1934 * Allocates a number of pages from the system
1935 *
1936 * Arguments:
1937 * bytes The number of bytes requested
1938 * wait Shall we wait?
1939 *
1940 * Returns:
1941 * A pointer to the alloced memory or possibly
1942 * NULL if M_NOWAIT is set.
1943 */
1944 static void *
page_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)1945 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1946 int wait)
1947 {
1948 void *p; /* Returned page */
1949
1950 *pflag = UMA_SLAB_KERNEL;
1951 p = kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1952
1953 return (p);
1954 }
1955
1956 static void *
pcpu_page_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)1957 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1958 int wait)
1959 {
1960 struct pglist alloctail;
1961 void *addr;
1962 char *zkva;
1963 int cpu, flags;
1964 vm_page_t p, p_next;
1965 #ifdef NUMA
1966 struct pcpu *pc;
1967 #endif
1968
1969 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1970
1971 TAILQ_INIT(&alloctail);
1972 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | malloc2vm_flags(wait);
1973 *pflag = UMA_SLAB_KERNEL;
1974 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1975 if (CPU_ABSENT(cpu)) {
1976 p = vm_page_alloc_noobj(flags);
1977 } else {
1978 #ifndef NUMA
1979 p = vm_page_alloc_noobj(flags);
1980 #else
1981 pc = pcpu_find(cpu);
1982 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1983 p = NULL;
1984 else
1985 p = vm_page_alloc_noobj_domain(pc->pc_domain,
1986 flags);
1987 if (__predict_false(p == NULL))
1988 p = vm_page_alloc_noobj(flags);
1989 #endif
1990 }
1991 if (__predict_false(p == NULL))
1992 goto fail;
1993 TAILQ_INSERT_TAIL(&alloctail, p, plinks.q);
1994 }
1995 if ((addr = kva_alloc(bytes)) == NULL)
1996 goto fail;
1997 zkva = addr;
1998 TAILQ_FOREACH(p, &alloctail, plinks.q) {
1999 pmap_qenter(zkva, &p, 1);
2000 zkva += PAGE_SIZE;
2001 }
2002 return (addr);
2003 fail:
2004 TAILQ_FOREACH_SAFE(p, &alloctail, plinks.q, p_next) {
2005 vm_page_unwire_noq(p);
2006 vm_page_free(p);
2007 }
2008 return (NULL);
2009 }
2010
2011 /*
2012 * Allocates a number of pages not belonging to a VM object
2013 *
2014 * Arguments:
2015 * bytes The number of bytes requested
2016 * wait Shall we wait?
2017 *
2018 * Returns:
2019 * A pointer to the alloced memory or possibly
2020 * NULL if M_NOWAIT is set.
2021 */
2022 static void *
noobj_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * flags,int wait)2023 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
2024 int wait)
2025 {
2026 TAILQ_HEAD(, vm_page) alloctail;
2027 u_long npages;
2028 void *retkva;
2029 char *zkva;
2030 vm_page_t p, p_next;
2031 uma_keg_t keg;
2032 int req;
2033
2034 TAILQ_INIT(&alloctail);
2035 keg = zone->uz_keg;
2036 req = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
2037 if ((wait & M_WAITOK) != 0)
2038 req |= VM_ALLOC_WAITOK;
2039
2040 npages = howmany(bytes, PAGE_SIZE);
2041 while (npages > 0) {
2042 p = vm_page_alloc_noobj_domain(domain, req);
2043 if (p != NULL) {
2044 TAILQ_INSERT_TAIL(&alloctail, p, plinks.q);
2045 npages--;
2046 continue;
2047 }
2048 /*
2049 * Page allocation failed, free intermediate pages and
2050 * exit.
2051 */
2052 TAILQ_FOREACH_SAFE(p, &alloctail, plinks.q, p_next) {
2053 vm_page_unwire_noq(p);
2054 vm_page_free(p);
2055 }
2056 return (NULL);
2057 }
2058 *flags = UMA_SLAB_PRIV;
2059 zkva = (char *)keg->uk_kva +
2060 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
2061 retkva = zkva;
2062 TAILQ_FOREACH(p, &alloctail, plinks.q) {
2063 pmap_qenter(zkva, &p, 1);
2064 zkva += PAGE_SIZE;
2065 }
2066
2067 return (retkva);
2068 }
2069
2070 /*
2071 * Allocate physically contiguous pages.
2072 */
2073 static void *
contig_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)2074 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
2075 int wait)
2076 {
2077
2078 *pflag = UMA_SLAB_KERNEL;
2079 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
2080 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
2081 }
2082
2083 #if defined(UMA_USE_DMAP) && !defined(UMA_MD_SMALL_ALLOC)
2084 void *
uma_small_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * flags,int wait)2085 uma_small_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
2086 int wait)
2087 {
2088 vm_page_t m;
2089
2090 *flags = UMA_SLAB_PRIV;
2091 m = vm_page_alloc_noobj_domain(domain,
2092 malloc2vm_flags(wait) | VM_ALLOC_WIRED);
2093 if (m == NULL)
2094 return (NULL);
2095 if ((wait & M_NODUMP) == 0)
2096 dump_add_page(VM_PAGE_TO_PHYS(m));
2097 return (VM_PAGE_TO_DMAP(m));
2098 }
2099 #endif
2100
2101 /*
2102 * Frees a number of pages to the system
2103 *
2104 * Arguments:
2105 * mem A pointer to the memory to be freed
2106 * size The size of the memory being freed
2107 * flags The original p->us_flags field
2108 *
2109 * Returns:
2110 * Nothing
2111 */
2112 static void
page_free(void * mem,vm_size_t size,uint8_t flags)2113 page_free(void *mem, vm_size_t size, uint8_t flags)
2114 {
2115
2116 if ((flags & UMA_SLAB_BOOT) != 0) {
2117 startup_free(mem, size);
2118 return;
2119 }
2120
2121 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
2122 ("UMA: page_free used with invalid flags %x", flags));
2123
2124 kmem_free(mem, size);
2125 }
2126
2127 /*
2128 * Frees pcpu zone allocations
2129 *
2130 * Arguments:
2131 * mem A pointer to the memory to be freed
2132 * size The size of the memory being freed
2133 * flags The original p->us_flags field
2134 *
2135 * Returns:
2136 * Nothing
2137 */
2138 static void
pcpu_page_free(void * mem,vm_size_t size,uint8_t flags)2139 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
2140 {
2141 vm_offset_t sva, curva;
2142 vm_paddr_t paddr;
2143 vm_page_t m;
2144
2145 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
2146
2147 if ((flags & UMA_SLAB_BOOT) != 0) {
2148 startup_free(mem, size);
2149 return;
2150 }
2151
2152 sva = (vm_offset_t)mem;
2153 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
2154 paddr = pmap_kextract(curva);
2155 m = PHYS_TO_VM_PAGE(paddr);
2156 vm_page_unwire_noq(m);
2157 vm_page_free(m);
2158 }
2159 pmap_qremove(mem, size >> PAGE_SHIFT);
2160 kva_free(mem, size);
2161 }
2162
2163 #if defined(UMA_USE_DMAP) && !defined(UMA_MD_SMALL_ALLOC)
2164 void
uma_small_free(void * mem,vm_size_t size,uint8_t flags)2165 uma_small_free(void *mem, vm_size_t size, uint8_t flags)
2166 {
2167 vm_page_t m;
2168 vm_paddr_t pa;
2169
2170 pa = DMAP_TO_PHYS(mem);
2171 dump_drop_page(pa);
2172 m = PHYS_TO_VM_PAGE(pa);
2173 vm_page_unwire_noq(m);
2174 vm_page_free(m);
2175 }
2176 #endif
2177
2178 /*
2179 * Zero fill initializer
2180 *
2181 * Arguments/Returns follow uma_init specifications
2182 */
2183 static int
zero_init(void * mem,int size,int flags)2184 zero_init(void *mem, int size, int flags)
2185 {
2186 bzero(mem, size);
2187 return (0);
2188 }
2189
2190 #ifdef INVARIANTS
2191 static struct noslabbits *
slab_dbg_bits(uma_slab_t slab,uma_keg_t keg)2192 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
2193 {
2194
2195 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
2196 }
2197 #endif
2198
2199 /*
2200 * Actual size of embedded struct slab (!OFFPAGE).
2201 */
2202 static size_t
slab_sizeof(int nitems)2203 slab_sizeof(int nitems)
2204 {
2205 size_t s;
2206
2207 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
2208 return (roundup(s, UMA_ALIGN_PTR + 1));
2209 }
2210
2211 #define UMA_FIXPT_SHIFT 31
2212 #define UMA_FRAC_FIXPT(n, d) \
2213 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
2214 #define UMA_FIXPT_PCT(f) \
2215 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
2216 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
2217 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
2218
2219 /*
2220 * Compute the number of items that will fit in a slab. If hdr is true, the
2221 * item count may be limited to provide space in the slab for an inline slab
2222 * header. Otherwise, all slab space will be provided for item storage.
2223 */
2224 static u_int
slab_ipers_hdr(u_int size,u_int rsize,u_int slabsize,bool hdr)2225 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
2226 {
2227 u_int ipers;
2228 u_int padpi;
2229
2230 /* The padding between items is not needed after the last item. */
2231 padpi = rsize - size;
2232
2233 if (hdr) {
2234 /*
2235 * Start with the maximum item count and remove items until
2236 * the slab header first alongside the allocatable memory.
2237 */
2238 for (ipers = MIN(SLAB_MAX_SETSIZE,
2239 (slabsize + padpi - slab_sizeof(1)) / rsize);
2240 ipers > 0 &&
2241 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2242 ipers--)
2243 continue;
2244 } else {
2245 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2246 }
2247
2248 return (ipers);
2249 }
2250
2251 struct keg_layout_result {
2252 u_int format;
2253 u_int slabsize;
2254 u_int ipers;
2255 u_int eff;
2256 };
2257
2258 static void
keg_layout_one(uma_keg_t keg,u_int size,u_int rsize,u_int slabsize,u_int fmt,struct keg_layout_result * kl)2259 keg_layout_one(uma_keg_t keg, u_int size, u_int rsize, u_int slabsize,
2260 u_int fmt, struct keg_layout_result *kl)
2261 {
2262 u_int total;
2263
2264 kl->format = fmt;
2265 kl->slabsize = slabsize;
2266
2267 /* Handle INTERNAL as inline with an extra page. */
2268 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2269 kl->format &= ~UMA_ZFLAG_INTERNAL;
2270 kl->slabsize += PAGE_SIZE;
2271 }
2272
2273 kl->ipers = slab_ipers_hdr(size, rsize, kl->slabsize,
2274 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2275
2276 /* Account for memory used by an offpage slab header. */
2277 total = kl->slabsize;
2278 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2279 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2280
2281 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2282 }
2283
2284 /*
2285 * Determine the format of a uma keg. This determines where the slab header
2286 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2287 *
2288 * Arguments
2289 * keg The zone we should initialize
2290 *
2291 * Returns
2292 * Nothing
2293 */
2294 static void
keg_layout(uma_keg_t keg)2295 keg_layout(uma_keg_t keg)
2296 {
2297 struct keg_layout_result kl = {}, kl_tmp;
2298 u_int fmts[2];
2299 u_int alignsize;
2300 u_int nfmt;
2301 u_int pages;
2302 u_int size, rsize;
2303 u_int slabsize;
2304 u_int i, j;
2305
2306 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2307 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2308 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2309 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2310 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2311 PRINT_UMA_ZFLAGS));
2312 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2313 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2314 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2315 PRINT_UMA_ZFLAGS));
2316
2317 alignsize = keg->uk_align + 1;
2318
2319 /*
2320 * Calculate the size of each allocation. uk_size is the originally
2321 * requested item size that the consumer expects to use. "size" is the
2322 * requested size after adjusting for an optional redzone after each
2323 * item (currently used only by KASAN). rsize is the final size between
2324 * item start addresses after adjusting for alignment and minimum
2325 * allocation size requirements.
2326 *
2327 * The padding given by the difference rsize - size may not be present
2328 * for the last item in a slab.
2329 */
2330 size = keg->uk_size;
2331
2332 #ifdef KASAN
2333 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
2334 /*
2335 * kasan_mark() requires that each allocation be aligned to the
2336 * shadow map scale factor.
2337 */
2338 if (alignsize < KASAN_SHADOW_SCALE)
2339 alignsize = KASAN_SHADOW_SCALE;
2340 size += KASAN_SHADOW_SCALE;
2341 }
2342 #endif
2343
2344 rsize = MAX(size, UMA_SMALLEST_UNIT);
2345 rsize = roundup2(rsize, alignsize);
2346
2347 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2348 /*
2349 * We want one item to start on every align boundary in a page.
2350 * To do this we will span pages. We will also extend the item
2351 * by the size of align if it is an even multiple of align.
2352 * Otherwise, it would fall on the same boundary every time.
2353 */
2354 if ((rsize & alignsize) == 0)
2355 rsize += alignsize;
2356 slabsize = rsize * (PAGE_SIZE / alignsize);
2357 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2358 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2359 slabsize = round_page(slabsize);
2360 } else {
2361 /*
2362 * Start with a slab size of as many pages as it takes to
2363 * represent a single item. We will try to fit as many
2364 * additional items into the slab as possible.
2365 */
2366 slabsize = round_page(size);
2367 }
2368
2369 /* Build a list of all of the available formats for this keg. */
2370 nfmt = 0;
2371
2372 /* Evaluate an inline slab layout. */
2373 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2374 fmts[nfmt++] = 0;
2375
2376 /* TODO: vm_page-embedded slab. */
2377
2378 /*
2379 * We can't do OFFPAGE if we're internal or if we've been
2380 * asked to not go to the VM for buckets. If we do this we
2381 * may end up going to the VM for slabs which we do not want
2382 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2383 * In those cases, evaluate a pseudo-format called INTERNAL
2384 * which has an inline slab header and one extra page to
2385 * guarantee that it fits.
2386 *
2387 * Otherwise, see if using an OFFPAGE slab will improve our
2388 * efficiency.
2389 */
2390 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2391 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2392 else
2393 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2394
2395 /*
2396 * Choose a slab size and format which satisfy the minimum efficiency.
2397 * Prefer the smallest slab size that meets the constraints.
2398 *
2399 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2400 * for small items (up to PAGE_SIZE), the iteration increment is one
2401 * page; and for large items, the increment is one item.
2402 */
2403 i = (slabsize + rsize - size) / MAX(PAGE_SIZE, rsize);
2404 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2405 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2406 rsize, i));
2407 for ( ; ; i++) {
2408 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2409 round_page(rsize * (i - 1) + size);
2410
2411 for (j = 0; j < nfmt; j++) {
2412 /* Only if we have no viable format yet. */
2413 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2414 kl.ipers > 0)
2415 continue;
2416
2417 keg_layout_one(keg, size, rsize, slabsize, fmts[j],
2418 &kl_tmp);
2419 if (kl_tmp.eff <= kl.eff)
2420 continue;
2421
2422 kl = kl_tmp;
2423
2424 CTR6(KTR_UMA, "keg %s layout: format %#x "
2425 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2426 keg->uk_name, kl.format, kl.ipers, rsize,
2427 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2428
2429 /* Stop when we reach the minimum efficiency. */
2430 if (kl.eff >= UMA_MIN_EFF)
2431 break;
2432 }
2433
2434 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2435 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2436 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2437 break;
2438 }
2439
2440 pages = atop(kl.slabsize);
2441 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2442 pages *= mp_maxid + 1;
2443
2444 keg->uk_rsize = rsize;
2445 keg->uk_ipers = kl.ipers;
2446 keg->uk_ppera = pages;
2447 keg->uk_flags |= kl.format;
2448
2449 /*
2450 * How do we find the slab header if it is offpage or if not all item
2451 * start addresses are in the same page? We could solve the latter
2452 * case with vaddr alignment, but we don't.
2453 */
2454 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2455 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2456 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2457 keg->uk_flags |= UMA_ZFLAG_HASH;
2458 else
2459 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2460 }
2461
2462 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2463 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2464 pages);
2465 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2466 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2467 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2468 keg->uk_ipers, pages));
2469 }
2470
2471 /*
2472 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2473 * the keg onto the global keg list.
2474 *
2475 * Arguments/Returns follow uma_ctor specifications
2476 * udata Actually uma_kctor_args
2477 */
2478 static int
keg_ctor(void * mem,int size,void * udata,int flags)2479 keg_ctor(void *mem, int size, void *udata, int flags)
2480 {
2481 struct uma_kctor_args *arg = udata;
2482 uma_keg_t keg = mem;
2483 uma_zone_t zone;
2484 int i;
2485
2486 bzero(keg, size);
2487 keg->uk_size = arg->size;
2488 keg->uk_init = arg->uminit;
2489 keg->uk_fini = arg->fini;
2490 keg->uk_align = arg->align;
2491 keg->uk_reserve = 0;
2492 keg->uk_flags = arg->flags;
2493
2494 /*
2495 * We use a global round-robin policy by default. Zones with
2496 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2497 * case the iterator is never run.
2498 */
2499 keg->uk_dr.dr_policy = DOMAINSET_RR();
2500 keg->uk_dr.dr_iter = 0;
2501
2502 /*
2503 * The primary zone is passed to us at keg-creation time.
2504 */
2505 zone = arg->zone;
2506 keg->uk_name = zone->uz_name;
2507
2508 if (arg->flags & UMA_ZONE_ZINIT)
2509 keg->uk_init = zero_init;
2510
2511 if (arg->flags & UMA_ZONE_MALLOC)
2512 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2513
2514 #ifndef SMP
2515 keg->uk_flags &= ~UMA_ZONE_PCPU;
2516 #endif
2517
2518 keg_layout(keg);
2519
2520 /*
2521 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2522 * work on. Use round-robin for everything else.
2523 *
2524 * Zones may override the default by specifying either.
2525 */
2526 #ifdef NUMA
2527 if ((keg->uk_flags &
2528 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2529 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2530 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2531 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2532 #endif
2533
2534 /*
2535 * If we haven't booted yet we need allocations to go through the
2536 * startup cache until the vm is ready.
2537 */
2538 #ifdef UMA_USE_DMAP
2539 if (keg->uk_ppera == 1)
2540 keg->uk_allocf = uma_small_alloc;
2541 else
2542 #endif
2543 if (booted < BOOT_KVA)
2544 keg->uk_allocf = startup_alloc;
2545 else if (keg->uk_flags & UMA_ZONE_PCPU)
2546 keg->uk_allocf = pcpu_page_alloc;
2547 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2548 keg->uk_allocf = contig_alloc;
2549 else
2550 keg->uk_allocf = page_alloc;
2551 #ifdef UMA_USE_DMAP
2552 if (keg->uk_ppera == 1)
2553 keg->uk_freef = uma_small_free;
2554 else
2555 #endif
2556 if (keg->uk_flags & UMA_ZONE_PCPU)
2557 keg->uk_freef = pcpu_page_free;
2558 else
2559 keg->uk_freef = page_free;
2560
2561 /*
2562 * Initialize keg's locks.
2563 */
2564 for (i = 0; i < vm_ndomains; i++)
2565 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2566
2567 /*
2568 * If we're putting the slab header in the actual page we need to
2569 * figure out where in each page it goes. See slab_sizeof
2570 * definition.
2571 */
2572 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2573 size_t shsize;
2574
2575 shsize = slab_sizeof(keg->uk_ipers);
2576 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2577 /*
2578 * The only way the following is possible is if with our
2579 * UMA_ALIGN_PTR adjustments we are now bigger than
2580 * UMA_SLAB_SIZE. I haven't checked whether this is
2581 * mathematically possible for all cases, so we make
2582 * sure here anyway.
2583 */
2584 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2585 ("zone %s ipers %d rsize %d size %d slab won't fit",
2586 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2587 }
2588
2589 if (keg->uk_flags & UMA_ZFLAG_HASH)
2590 hash_alloc(&keg->uk_hash, 0);
2591
2592 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2593
2594 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2595
2596 rw_wlock(&uma_rwlock);
2597 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2598 rw_wunlock(&uma_rwlock);
2599 return (0);
2600 }
2601
2602 static void
zone_kva_available(uma_zone_t zone,void * unused)2603 zone_kva_available(uma_zone_t zone, void *unused)
2604 {
2605 uma_keg_t keg;
2606
2607 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2608 return;
2609 KEG_GET(zone, keg);
2610
2611 if (keg->uk_allocf == startup_alloc) {
2612 /* Switch to the real allocator. */
2613 if (keg->uk_flags & UMA_ZONE_PCPU)
2614 keg->uk_allocf = pcpu_page_alloc;
2615 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2616 keg->uk_ppera > 1)
2617 keg->uk_allocf = contig_alloc;
2618 else
2619 keg->uk_allocf = page_alloc;
2620 }
2621 }
2622
2623 static void
zone_alloc_counters(uma_zone_t zone,void * unused)2624 zone_alloc_counters(uma_zone_t zone, void *unused)
2625 {
2626
2627 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2628 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2629 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2630 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2631 }
2632
2633 static void
zone_alloc_sysctl(uma_zone_t zone,void * unused)2634 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2635 {
2636 uma_zone_domain_t zdom;
2637 uma_domain_t dom;
2638 uma_keg_t keg;
2639 struct sysctl_oid *oid, *domainoid;
2640 int domains, i, cnt;
2641 static const char *nokeg = "cache zone";
2642 char *c;
2643
2644 /*
2645 * Make a sysctl safe copy of the zone name by removing
2646 * any special characters and handling dups by appending
2647 * an index.
2648 */
2649 if (zone->uz_namecnt != 0) {
2650 /* Count the number of decimal digits and '_' separator. */
2651 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2652 cnt /= 10;
2653 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2654 M_UMA, M_WAITOK);
2655 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2656 zone->uz_namecnt);
2657 } else
2658 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2659 for (c = zone->uz_ctlname; *c != '\0'; c++)
2660 if (strchr("./\\ -", *c) != NULL)
2661 *c = '_';
2662
2663 /*
2664 * Basic parameters at the root.
2665 */
2666 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2667 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2668 oid = zone->uz_oid;
2669 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2670 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2671 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2672 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2673 zone, 0, sysctl_handle_uma_zone_flags, "A",
2674 "Allocator configuration flags");
2675 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2676 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2677 "Desired per-cpu cache size");
2678 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2679 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2680 "Maximum allowed per-cpu cache size");
2681
2682 /*
2683 * keg if present.
2684 */
2685 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2686 domains = vm_ndomains;
2687 else
2688 domains = 1;
2689 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2690 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2691 keg = zone->uz_keg;
2692 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2693 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2694 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2695 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2696 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2697 "Real object size with alignment");
2698 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2699 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2700 "pages per-slab allocation");
2701 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2702 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2703 "items available per-slab");
2704 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2705 "align", CTLFLAG_RD, &keg->uk_align, 0,
2706 "item alignment mask");
2707 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2708 "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2709 "number of reserved items");
2710 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2711 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2712 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2713 "Slab utilization (100 - internal fragmentation %)");
2714 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2715 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2716 for (i = 0; i < domains; i++) {
2717 dom = &keg->uk_domain[i];
2718 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2719 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2720 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2721 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2722 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2723 "Total pages currently allocated from VM");
2724 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2725 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2726 "Items free in the slab layer");
2727 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2728 "free_slabs", CTLFLAG_RD, &dom->ud_free_slabs, 0,
2729 "Unused slabs");
2730 }
2731 } else
2732 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2733 "name", CTLFLAG_RD, nokeg, "Keg name");
2734
2735 /*
2736 * Information about zone limits.
2737 */
2738 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2739 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2740 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2741 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2742 zone, 0, sysctl_handle_uma_zone_items, "QU",
2743 "Current number of allocated items if limit is set");
2744 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2745 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2746 "Maximum number of allocated and cached items");
2747 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2748 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2749 "Number of threads sleeping at limit");
2750 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2751 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2752 "Total zone limit sleeps");
2753 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2754 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2755 "Maximum number of items in each domain's bucket cache");
2756
2757 /*
2758 * Per-domain zone information.
2759 */
2760 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2761 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2762 for (i = 0; i < domains; i++) {
2763 zdom = ZDOM_GET(zone, i);
2764 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2765 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2766 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2767 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2768 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2769 "number of items in this domain");
2770 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2771 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2772 "maximum item count in this period");
2773 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2774 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2775 "minimum item count in this period");
2776 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2777 "bimin", CTLFLAG_RD, &zdom->uzd_bimin,
2778 "Minimum item count in this batch");
2779 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2780 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2781 "Working set size");
2782 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2783 "limin", CTLFLAG_RD, &zdom->uzd_limin,
2784 "Long time minimum item count");
2785 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2786 "timin", CTLFLAG_RD, &zdom->uzd_timin, 0,
2787 "Time since zero long time minimum item count");
2788 }
2789
2790 /*
2791 * General statistics.
2792 */
2793 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2794 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2795 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2796 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2797 zone, 1, sysctl_handle_uma_zone_cur, "I",
2798 "Current number of allocated items");
2799 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2800 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2801 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2802 "Total allocation calls");
2803 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2804 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2805 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2806 "Total free calls");
2807 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2808 "fails", CTLFLAG_RD, &zone->uz_fails,
2809 "Number of allocation failures");
2810 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2811 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2812 "Free calls from the wrong domain");
2813 }
2814
2815 struct uma_zone_count {
2816 const char *name;
2817 int count;
2818 };
2819
2820 static void
zone_count(uma_zone_t zone,void * arg)2821 zone_count(uma_zone_t zone, void *arg)
2822 {
2823 struct uma_zone_count *cnt;
2824
2825 cnt = arg;
2826 /*
2827 * Some zones are rapidly created with identical names and
2828 * destroyed out of order. This can lead to gaps in the count.
2829 * Use one greater than the maximum observed for this name.
2830 */
2831 if (strcmp(zone->uz_name, cnt->name) == 0)
2832 cnt->count = MAX(cnt->count,
2833 zone->uz_namecnt + 1);
2834 }
2835
2836 static void
zone_update_caches(uma_zone_t zone)2837 zone_update_caches(uma_zone_t zone)
2838 {
2839 int i;
2840
2841 for (i = 0; i <= mp_maxid; i++) {
2842 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2843 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2844 }
2845 }
2846
2847 /*
2848 * Zone header ctor. This initializes all fields, locks, etc.
2849 *
2850 * Arguments/Returns follow uma_ctor specifications
2851 * udata Actually uma_zctor_args
2852 */
2853 static int
zone_ctor(void * mem,int size,void * udata,int flags)2854 zone_ctor(void *mem, int size, void *udata, int flags)
2855 {
2856 struct uma_zone_count cnt;
2857 struct uma_zctor_args *arg = udata;
2858 uma_zone_domain_t zdom;
2859 uma_zone_t zone = mem;
2860 uma_zone_t z;
2861 uma_keg_t keg;
2862 int i;
2863
2864 bzero(zone, size);
2865 zone->uz_name = arg->name;
2866 zone->uz_ctor = arg->ctor;
2867 zone->uz_dtor = arg->dtor;
2868 zone->uz_init = NULL;
2869 zone->uz_fini = NULL;
2870 zone->uz_sleeps = 0;
2871 zone->uz_bucket_size = 0;
2872 zone->uz_bucket_size_min = 0;
2873 zone->uz_bucket_size_max = BUCKET_MAX;
2874 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2875 zone->uz_warning = NULL;
2876 /* The domain structures follow the cpu structures. */
2877 zone->uz_bucket_max = ULONG_MAX;
2878 timevalclear(&zone->uz_ratecheck);
2879
2880 /* Count the number of duplicate names. */
2881 cnt.name = arg->name;
2882 cnt.count = 0;
2883 zone_foreach(zone_count, &cnt);
2884 zone->uz_namecnt = cnt.count;
2885 ZONE_CROSS_LOCK_INIT(zone);
2886
2887 for (i = 0; i < vm_ndomains; i++) {
2888 zdom = ZDOM_GET(zone, i);
2889 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2890 STAILQ_INIT(&zdom->uzd_buckets);
2891 }
2892
2893 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
2894 if (arg->uminit == trash_init && arg->fini == trash_fini)
2895 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2896 #elif defined(KASAN)
2897 if ((arg->flags & (UMA_ZONE_NOFREE | UMA_ZFLAG_CACHE)) != 0)
2898 arg->flags |= UMA_ZONE_NOKASAN;
2899 #endif
2900
2901 /*
2902 * This is a pure cache zone, no kegs.
2903 */
2904 if (arg->import) {
2905 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2906 ("zone_ctor: Import specified for non-cache zone."));
2907 zone->uz_flags = arg->flags;
2908 zone->uz_size = arg->size;
2909 zone->uz_import = arg->import;
2910 zone->uz_release = arg->release;
2911 zone->uz_arg = arg->arg;
2912 #ifdef NUMA
2913 /*
2914 * Cache zones are round-robin unless a policy is
2915 * specified because they may have incompatible
2916 * constraints.
2917 */
2918 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2919 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2920 #endif
2921 rw_wlock(&uma_rwlock);
2922 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2923 rw_wunlock(&uma_rwlock);
2924 goto out;
2925 }
2926
2927 /*
2928 * Use the regular zone/keg/slab allocator.
2929 */
2930 zone->uz_import = zone_import;
2931 zone->uz_release = zone_release;
2932 zone->uz_arg = zone;
2933 keg = arg->keg;
2934
2935 if (arg->flags & UMA_ZONE_SECONDARY) {
2936 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2937 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2938 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2939 zone->uz_init = arg->uminit;
2940 zone->uz_fini = arg->fini;
2941 zone->uz_flags |= UMA_ZONE_SECONDARY;
2942 rw_wlock(&uma_rwlock);
2943 ZONE_LOCK(zone);
2944 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2945 if (LIST_NEXT(z, uz_link) == NULL) {
2946 LIST_INSERT_AFTER(z, zone, uz_link);
2947 break;
2948 }
2949 }
2950 ZONE_UNLOCK(zone);
2951 rw_wunlock(&uma_rwlock);
2952 } else if (keg == NULL) {
2953 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2954 arg->align, arg->flags)) == NULL)
2955 return (ENOMEM);
2956 } else {
2957 struct uma_kctor_args karg;
2958 int error;
2959
2960 /* We should only be here from uma_startup() */
2961 karg.size = arg->size;
2962 karg.uminit = arg->uminit;
2963 karg.fini = arg->fini;
2964 karg.align = arg->align;
2965 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2966 karg.zone = zone;
2967 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2968 flags);
2969 if (error)
2970 return (error);
2971 }
2972
2973 /* Inherit properties from the keg. */
2974 zone->uz_keg = keg;
2975 zone->uz_size = keg->uk_size;
2976 zone->uz_flags |= (keg->uk_flags &
2977 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2978
2979 out:
2980 if (booted >= BOOT_PCPU) {
2981 zone_alloc_counters(zone, NULL);
2982 if (booted >= BOOT_RUNNING)
2983 zone_alloc_sysctl(zone, NULL);
2984 } else {
2985 zone->uz_allocs = EARLY_COUNTER;
2986 zone->uz_frees = EARLY_COUNTER;
2987 zone->uz_fails = EARLY_COUNTER;
2988 }
2989
2990 /* Caller requests a private SMR context. */
2991 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2992 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2993
2994 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2995 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2996 ("Invalid zone flag combination"));
2997 if (arg->flags & UMA_ZFLAG_INTERNAL)
2998 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2999 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
3000 zone->uz_bucket_size = BUCKET_MAX;
3001 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
3002 zone->uz_bucket_size = 0;
3003 else
3004 zone->uz_bucket_size = bucket_select(zone->uz_size);
3005 zone->uz_bucket_size_min = zone->uz_bucket_size;
3006 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
3007 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
3008 zone_update_caches(zone);
3009
3010 return (0);
3011 }
3012
3013 /*
3014 * Keg header dtor. This frees all data, destroys locks, frees the hash
3015 * table and removes the keg from the global list.
3016 *
3017 * Arguments/Returns follow uma_dtor specifications
3018 * udata unused
3019 */
3020 static void
keg_dtor(void * arg,int size,void * udata)3021 keg_dtor(void *arg, int size, void *udata)
3022 {
3023 uma_keg_t keg;
3024 uint32_t free, pages;
3025 int i;
3026
3027 keg = (uma_keg_t)arg;
3028 free = pages = 0;
3029 for (i = 0; i < vm_ndomains; i++) {
3030 free += keg->uk_domain[i].ud_free_items;
3031 pages += keg->uk_domain[i].ud_pages;
3032 KEG_LOCK_FINI(keg, i);
3033 }
3034 if (pages != 0)
3035 printf("Freed UMA keg (%s) was not empty (%u items). "
3036 " Lost %u pages of memory.\n",
3037 keg->uk_name ? keg->uk_name : "",
3038 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
3039
3040 hash_free(&keg->uk_hash);
3041 }
3042
3043 /*
3044 * Zone header dtor.
3045 *
3046 * Arguments/Returns follow uma_dtor specifications
3047 * udata unused
3048 */
3049 static void
zone_dtor(void * arg,int size,void * udata)3050 zone_dtor(void *arg, int size, void *udata)
3051 {
3052 uma_zone_t zone;
3053 uma_keg_t keg;
3054 int i;
3055
3056 zone = (uma_zone_t)arg;
3057
3058 sysctl_remove_oid(zone->uz_oid, 1, 1);
3059
3060 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
3061 cache_drain(zone);
3062
3063 rw_wlock(&uma_rwlock);
3064 LIST_REMOVE(zone, uz_link);
3065 rw_wunlock(&uma_rwlock);
3066 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3067 keg = zone->uz_keg;
3068 keg->uk_reserve = 0;
3069 }
3070 zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true);
3071
3072 /*
3073 * We only destroy kegs from non secondary/non cache zones.
3074 */
3075 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3076 keg = zone->uz_keg;
3077 rw_wlock(&uma_rwlock);
3078 LIST_REMOVE(keg, uk_link);
3079 rw_wunlock(&uma_rwlock);
3080 zone_free_item(kegs, keg, NULL, SKIP_NONE);
3081 }
3082 counter_u64_free(zone->uz_allocs);
3083 counter_u64_free(zone->uz_frees);
3084 counter_u64_free(zone->uz_fails);
3085 counter_u64_free(zone->uz_xdomain);
3086 free(zone->uz_ctlname, M_UMA);
3087 for (i = 0; i < vm_ndomains; i++)
3088 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
3089 ZONE_CROSS_LOCK_FINI(zone);
3090 }
3091
3092 static void
zone_foreach_unlocked(void (* zfunc)(uma_zone_t,void * arg),void * arg)3093 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3094 {
3095 uma_keg_t keg;
3096 uma_zone_t zone;
3097
3098 LIST_FOREACH(keg, &uma_kegs, uk_link) {
3099 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
3100 zfunc(zone, arg);
3101 }
3102 LIST_FOREACH(zone, &uma_cachezones, uz_link)
3103 zfunc(zone, arg);
3104 }
3105
3106 /*
3107 * Traverses every zone in the system and calls a callback
3108 *
3109 * Arguments:
3110 * zfunc A pointer to a function which accepts a zone
3111 * as an argument.
3112 *
3113 * Returns:
3114 * Nothing
3115 */
3116 static void
zone_foreach(void (* zfunc)(uma_zone_t,void * arg),void * arg)3117 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3118 {
3119
3120 rw_rlock(&uma_rwlock);
3121 zone_foreach_unlocked(zfunc, arg);
3122 rw_runlock(&uma_rwlock);
3123 }
3124
3125 /*
3126 * Initialize the kernel memory allocator. This is done after pages can be
3127 * allocated but before general KVA is available.
3128 */
3129 void
uma_startup1(vm_offset_t virtual_avail)3130 uma_startup1(vm_offset_t virtual_avail)
3131 {
3132 struct uma_zctor_args args;
3133 size_t ksize, zsize, size;
3134 uma_keg_t primarykeg;
3135 uintptr_t m;
3136 int domain;
3137 uint8_t pflag;
3138
3139 bootstart = bootmem = virtual_avail;
3140
3141 rw_init(&uma_rwlock, "UMA lock");
3142 sx_init(&uma_reclaim_lock, "umareclaim");
3143
3144 ksize = sizeof(struct uma_keg) +
3145 (sizeof(struct uma_domain) * vm_ndomains);
3146 ksize = roundup(ksize, UMA_SUPER_ALIGN);
3147 zsize = sizeof(struct uma_zone) +
3148 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
3149 (sizeof(struct uma_zone_domain) * vm_ndomains);
3150 zsize = roundup(zsize, UMA_SUPER_ALIGN);
3151
3152 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
3153 size = (zsize * 2) + ksize;
3154 for (domain = 0; domain < vm_ndomains; domain++) {
3155 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
3156 M_NOWAIT | M_ZERO);
3157 if (m != 0)
3158 break;
3159 }
3160 zones = (uma_zone_t)m;
3161 m += zsize;
3162 kegs = (uma_zone_t)m;
3163 m += zsize;
3164 primarykeg = (uma_keg_t)m;
3165
3166 /* "manually" create the initial zone */
3167 memset(&args, 0, sizeof(args));
3168 args.name = "UMA Kegs";
3169 args.size = ksize;
3170 args.ctor = keg_ctor;
3171 args.dtor = keg_dtor;
3172 args.uminit = zero_init;
3173 args.fini = NULL;
3174 args.keg = primarykeg;
3175 args.align = UMA_SUPER_ALIGN - 1;
3176 args.flags = UMA_ZFLAG_INTERNAL;
3177 zone_ctor(kegs, zsize, &args, M_WAITOK);
3178
3179 args.name = "UMA Zones";
3180 args.size = zsize;
3181 args.ctor = zone_ctor;
3182 args.dtor = zone_dtor;
3183 args.uminit = zero_init;
3184 args.fini = NULL;
3185 args.keg = NULL;
3186 args.align = UMA_SUPER_ALIGN - 1;
3187 args.flags = UMA_ZFLAG_INTERNAL;
3188 zone_ctor(zones, zsize, &args, M_WAITOK);
3189
3190 /* Now make zones for slab headers */
3191 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
3192 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3193 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
3194 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3195
3196 hashzone = uma_zcreate("UMA Hash",
3197 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
3198 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3199
3200 bucket_init();
3201 smr_init();
3202 }
3203
3204 #ifndef UMA_USE_DMAP
3205 extern void vm_radix_reserve_kva(void);
3206 #endif
3207
3208 /*
3209 * Advertise the availability of normal kva allocations and switch to
3210 * the default back-end allocator. Marks the KVA we consumed on startup
3211 * as used in the map.
3212 */
3213 void
uma_startup2(void)3214 uma_startup2(void)
3215 {
3216
3217 if (bootstart != bootmem) {
3218 vm_map_lock(kernel_map);
3219 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
3220 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
3221 vm_map_unlock(kernel_map);
3222 }
3223
3224 #ifndef UMA_USE_DMAP
3225 /* Set up radix zone to use noobj_alloc. */
3226 vm_radix_reserve_kva();
3227 #endif
3228
3229 booted = BOOT_KVA;
3230 zone_foreach_unlocked(zone_kva_available, NULL);
3231 bucket_enable();
3232 }
3233
3234 /*
3235 * Allocate counters as early as possible so that boot-time allocations are
3236 * accounted more precisely.
3237 */
3238 static void
uma_startup_pcpu(void * arg __unused)3239 uma_startup_pcpu(void *arg __unused)
3240 {
3241
3242 zone_foreach_unlocked(zone_alloc_counters, NULL);
3243 booted = BOOT_PCPU;
3244 }
3245 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
3246
3247 /*
3248 * Finish our initialization steps.
3249 */
3250 static void
uma_startup3(void * arg __unused)3251 uma_startup3(void *arg __unused)
3252 {
3253
3254 #ifdef INVARIANTS
3255 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
3256 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
3257 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
3258 #endif
3259 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
3260 booted = BOOT_RUNNING;
3261
3262 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
3263 EVENTHANDLER_PRI_FIRST);
3264 }
3265 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
3266
3267 static void
uma_startup4(void * arg __unused)3268 uma_startup4(void *arg __unused)
3269 {
3270 TIMEOUT_TASK_INIT(taskqueue_thread, &uma_timeout_task, 0, uma_timeout,
3271 NULL);
3272 taskqueue_enqueue_timeout(taskqueue_thread, &uma_timeout_task,
3273 UMA_TIMEOUT * hz);
3274 }
3275 SYSINIT(uma_startup4, SI_SUB_TASKQ, SI_ORDER_ANY, uma_startup4, NULL);
3276
3277 static void
uma_shutdown(void)3278 uma_shutdown(void)
3279 {
3280
3281 booted = BOOT_SHUTDOWN;
3282 }
3283
3284 static uma_keg_t
uma_kcreate(uma_zone_t zone,size_t size,uma_init uminit,uma_fini fini,int align,uint32_t flags)3285 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
3286 int align, uint32_t flags)
3287 {
3288 struct uma_kctor_args args;
3289
3290 args.size = size;
3291 args.uminit = uminit;
3292 args.fini = fini;
3293 args.align = align;
3294 args.flags = flags;
3295 args.zone = zone;
3296 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3297 }
3298
3299
3300 static void
check_align_mask(unsigned int mask)3301 check_align_mask(unsigned int mask)
3302 {
3303
3304 KASSERT(powerof2(mask + 1),
3305 ("UMA: %s: Not the mask of a power of 2 (%#x)", __func__, mask));
3306 /*
3307 * Make sure the stored align mask doesn't have its highest bit set,
3308 * which would cause implementation-defined behavior when passing it as
3309 * the 'align' argument of uma_zcreate(). Such very large alignments do
3310 * not make sense anyway.
3311 */
3312 KASSERT(mask <= INT_MAX,
3313 ("UMA: %s: Mask too big (%#x)", __func__, mask));
3314 }
3315
3316 /* Public functions */
3317 /* See uma.h */
3318 void
uma_set_cache_align_mask(unsigned int mask)3319 uma_set_cache_align_mask(unsigned int mask)
3320 {
3321
3322 check_align_mask(mask);
3323 uma_cache_align_mask = mask;
3324 }
3325
3326 /* Returns the alignment mask to use to request cache alignment. */
3327 unsigned int
uma_get_cache_align_mask(void)3328 uma_get_cache_align_mask(void)
3329 {
3330 return (uma_cache_align_mask);
3331 }
3332
3333 /* See uma.h */
3334 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)3335 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3336 uma_init uminit, uma_fini fini, int align, uint32_t flags)
3337
3338 {
3339 struct uma_zctor_args args;
3340 uma_zone_t res;
3341
3342 check_align_mask(align);
3343
3344 /* This stuff is essential for the zone ctor */
3345 memset(&args, 0, sizeof(args));
3346 args.name = name;
3347 args.size = size;
3348 args.ctor = ctor;
3349 args.dtor = dtor;
3350 args.uminit = uminit;
3351 args.fini = fini;
3352 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
3353 /*
3354 * Inject procedures which check for memory use after free if we are
3355 * allowed to scramble the memory while it is not allocated. This
3356 * requires that: UMA is actually able to access the memory, no init
3357 * or fini procedures, no dependency on the initial value of the
3358 * memory, and no (legitimate) use of the memory after free. Note,
3359 * the ctor and dtor do not need to be empty.
3360 */
3361 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3362 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3363 args.uminit = trash_init;
3364 args.fini = trash_fini;
3365 }
3366 #endif
3367 args.align = align;
3368 args.flags = flags;
3369 args.keg = NULL;
3370
3371 sx_xlock(&uma_reclaim_lock);
3372 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3373 sx_xunlock(&uma_reclaim_lock);
3374
3375 return (res);
3376 }
3377
3378 /* See uma.h */
3379 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)3380 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3381 uma_init zinit, uma_fini zfini, uma_zone_t primary)
3382 {
3383 struct uma_zctor_args args;
3384 uma_keg_t keg;
3385 uma_zone_t res;
3386
3387 keg = primary->uz_keg;
3388 memset(&args, 0, sizeof(args));
3389 args.name = name;
3390 args.size = keg->uk_size;
3391 args.ctor = ctor;
3392 args.dtor = dtor;
3393 args.uminit = zinit;
3394 args.fini = zfini;
3395 args.align = keg->uk_align;
3396 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3397 args.keg = keg;
3398
3399 sx_xlock(&uma_reclaim_lock);
3400 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3401 sx_xunlock(&uma_reclaim_lock);
3402
3403 return (res);
3404 }
3405
3406 /* See uma.h */
3407 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)3408 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3409 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3410 void *arg, int flags)
3411 {
3412 struct uma_zctor_args args;
3413
3414 memset(&args, 0, sizeof(args));
3415 args.name = name;
3416 args.size = size;
3417 args.ctor = ctor;
3418 args.dtor = dtor;
3419 args.uminit = zinit;
3420 args.fini = zfini;
3421 args.import = zimport;
3422 args.release = zrelease;
3423 args.arg = arg;
3424 args.align = 0;
3425 args.flags = flags | UMA_ZFLAG_CACHE;
3426
3427 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3428 }
3429
3430 /* See uma.h */
3431 void
uma_zdestroy(uma_zone_t zone)3432 uma_zdestroy(uma_zone_t zone)
3433 {
3434
3435 /*
3436 * Large slabs are expensive to reclaim, so don't bother doing
3437 * unnecessary work if we're shutting down.
3438 */
3439 if (booted == BOOT_SHUTDOWN &&
3440 zone->uz_fini == NULL && zone->uz_release == zone_release)
3441 return;
3442 sx_xlock(&uma_reclaim_lock);
3443 zone_free_item(zones, zone, NULL, SKIP_NONE);
3444 sx_xunlock(&uma_reclaim_lock);
3445 }
3446
3447 void
uma_zwait(uma_zone_t zone)3448 uma_zwait(uma_zone_t zone)
3449 {
3450
3451 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3452 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3453 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3454 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3455 else
3456 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3457 }
3458
3459 void *
uma_zalloc_pcpu_arg(uma_zone_t zone,void * udata,int flags)3460 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3461 {
3462 void *item, *pcpu_item;
3463 #ifdef SMP
3464 int i;
3465
3466 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3467 #endif
3468 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3469 if (item == NULL)
3470 return (NULL);
3471 pcpu_item = zpcpu_base_to_offset(item);
3472 if (flags & M_ZERO) {
3473 #ifdef SMP
3474 for (i = 0; i <= mp_maxid; i++)
3475 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3476 #else
3477 bzero(item, zone->uz_size);
3478 #endif
3479 }
3480 return (pcpu_item);
3481 }
3482
3483 /*
3484 * A stub while both regular and pcpu cases are identical.
3485 */
3486 void
uma_zfree_pcpu_arg(uma_zone_t zone,void * pcpu_item,void * udata)3487 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3488 {
3489 void *item;
3490
3491 #ifdef SMP
3492 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3493 #endif
3494
3495 /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */
3496 if (pcpu_item == NULL)
3497 return;
3498
3499 item = zpcpu_offset_to_base(pcpu_item);
3500 uma_zfree_arg(zone, item, udata);
3501 }
3502
3503 static inline void *
item_ctor(uma_zone_t zone,int uz_flags,int size,void * udata,int flags,void * item)3504 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3505 void *item)
3506 {
3507 #ifdef INVARIANTS
3508 bool skipdbg;
3509 #endif
3510
3511 KASSERT(zone->uz_keg == NULL ||
3512 ((uintptr_t)item & zone->uz_keg->uk_align) == 0,
3513 ("item_ctor: underaligned item %p from %s", item, zone->uz_name));
3514
3515 kasan_mark_item_valid(zone, item);
3516 kmsan_mark_item_uninitialized(zone, item);
3517
3518 #ifdef INVARIANTS
3519 skipdbg = uma_dbg_zskip(zone, item);
3520 if (!skipdbg && (uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3521 zone->uz_ctor != trash_ctor)
3522 trash_ctor(item, size, zone, flags);
3523 #endif
3524
3525 /* Check flags before loading ctor pointer. */
3526 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3527 __predict_false(zone->uz_ctor != NULL) &&
3528 zone->uz_ctor(item, size, udata, flags) != 0) {
3529 counter_u64_add(zone->uz_fails, 1);
3530 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3531 return (NULL);
3532 }
3533 #ifdef INVARIANTS
3534 if (!skipdbg)
3535 uma_dbg_alloc(zone, NULL, item);
3536 #endif
3537 if (__predict_false(flags & M_ZERO))
3538 return (memset(item, 0, size));
3539
3540 return (item);
3541 }
3542
3543 static inline void
item_dtor(uma_zone_t zone,void * item,int size,void * udata,enum zfreeskip skip)3544 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3545 enum zfreeskip skip)
3546 {
3547 #ifdef INVARIANTS
3548 bool skipdbg;
3549
3550 skipdbg = uma_dbg_zskip(zone, item);
3551 if (skip == SKIP_NONE && !skipdbg) {
3552 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3553 uma_dbg_free(zone, udata, item);
3554 else
3555 uma_dbg_free(zone, NULL, item);
3556 }
3557 #endif
3558 if (__predict_true(skip < SKIP_DTOR)) {
3559 if (zone->uz_dtor != NULL)
3560 zone->uz_dtor(item, size, udata);
3561 #ifdef INVARIANTS
3562 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3563 zone->uz_dtor != trash_dtor)
3564 trash_dtor(item, size, zone);
3565 #endif
3566 }
3567 kasan_mark_item_invalid(zone, item);
3568 }
3569
3570 #ifdef NUMA
3571 static int
item_domain(void * item)3572 item_domain(void *item)
3573 {
3574 int domain;
3575
3576 domain = vm_phys_domain(vtophys(item));
3577 KASSERT(domain >= 0 && domain < vm_ndomains,
3578 ("%s: unknown domain for item %p", __func__, item));
3579 return (domain);
3580 }
3581 #endif
3582
3583 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3584 #if defined(INVARIANTS) && (defined(DDB) || defined(STACK))
3585 #include <sys/stack.h>
3586 #endif
3587 #define UMA_ZALLOC_DEBUG
3588 static int
uma_zalloc_debug(uma_zone_t zone,void ** itemp,void * udata,int flags)3589 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3590 {
3591 int error;
3592
3593 error = 0;
3594 #ifdef WITNESS
3595 if (flags & M_WAITOK) {
3596 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3597 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3598 }
3599 #endif
3600
3601 #ifdef INVARIANTS
3602 KASSERT((flags & M_EXEC) == 0,
3603 ("uma_zalloc_debug: called with M_EXEC"));
3604 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3605 ("uma_zalloc_debug: called within spinlock or critical section"));
3606 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3607 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3608
3609 _Static_assert(M_NOWAIT != 0 && M_WAITOK != 0,
3610 "M_NOWAIT and M_WAITOK must be non-zero for this assertion:");
3611 #if 0
3612 /*
3613 * Give the #elif clause time to find problems, then remove it
3614 * and enable this. (Remove <sys/stack.h> above, too.)
3615 */
3616 KASSERT((flags & (M_NOWAIT|M_WAITOK)) == M_NOWAIT ||
3617 (flags & (M_NOWAIT|M_WAITOK)) == M_WAITOK,
3618 ("uma_zalloc_debug: must pass one of M_NOWAIT or M_WAITOK"));
3619 #elif defined(DDB) || defined(STACK)
3620 if (__predict_false((flags & (M_NOWAIT|M_WAITOK)) != M_NOWAIT &&
3621 (flags & (M_NOWAIT|M_WAITOK)) != M_WAITOK)) {
3622 static int stack_count;
3623 struct stack st;
3624
3625 if (stack_count < 10) {
3626 ++stack_count;
3627 printf("uma_zalloc* called with bad WAIT flags:\n");
3628 stack_save(&st);
3629 stack_print(&st);
3630 }
3631 }
3632 #endif
3633 #endif
3634
3635 #ifdef DEBUG_MEMGUARD
3636 if ((zone->uz_flags & (UMA_ZONE_SMR | UMA_ZFLAG_CACHE)) == 0 &&
3637 memguard_cmp_zone(zone)) {
3638 void *item;
3639 item = memguard_alloc(zone->uz_size, flags);
3640 if (item != NULL) {
3641 error = EJUSTRETURN;
3642 if (zone->uz_init != NULL &&
3643 zone->uz_init(item, zone->uz_size, flags) != 0) {
3644 *itemp = NULL;
3645 return (error);
3646 }
3647 if (zone->uz_ctor != NULL &&
3648 zone->uz_ctor(item, zone->uz_size, udata,
3649 flags) != 0) {
3650 counter_u64_add(zone->uz_fails, 1);
3651 if (zone->uz_fini != NULL)
3652 zone->uz_fini(item, zone->uz_size);
3653 *itemp = NULL;
3654 return (error);
3655 }
3656 *itemp = item;
3657 return (error);
3658 }
3659 /* This is unfortunate but should not be fatal. */
3660 }
3661 #endif
3662 return (error);
3663 }
3664
3665 static int
uma_zfree_debug(uma_zone_t zone,void * item,void * udata)3666 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3667 {
3668 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3669 ("uma_zfree_debug: called with spinlock or critical section held"));
3670
3671 #ifdef DEBUG_MEMGUARD
3672 if ((zone->uz_flags & (UMA_ZONE_SMR | UMA_ZFLAG_CACHE)) == 0 &&
3673 is_memguard_addr(item)) {
3674 if (zone->uz_dtor != NULL)
3675 zone->uz_dtor(item, zone->uz_size, udata);
3676 if (zone->uz_fini != NULL)
3677 zone->uz_fini(item, zone->uz_size);
3678 memguard_free(item);
3679 return (EJUSTRETURN);
3680 }
3681 #endif
3682 return (0);
3683 }
3684 #endif
3685
3686 static inline void *
cache_alloc_item(uma_zone_t zone,uma_cache_t cache,uma_cache_bucket_t bucket,void * udata,int flags)3687 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3688 void *udata, int flags)
3689 {
3690 void *item;
3691 int size, uz_flags;
3692
3693 item = cache_bucket_pop(cache, bucket);
3694 size = cache_uz_size(cache);
3695 uz_flags = cache_uz_flags(cache);
3696 critical_exit();
3697 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3698 }
3699
3700 static __noinline void *
cache_alloc_retry(uma_zone_t zone,uma_cache_t cache,void * udata,int flags)3701 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3702 {
3703 uma_cache_bucket_t bucket;
3704 int domain;
3705
3706 while (cache_alloc(zone, cache, udata, flags)) {
3707 cache = &zone->uz_cpu[curcpu];
3708 bucket = &cache->uc_allocbucket;
3709 if (__predict_false(bucket->ucb_cnt == 0))
3710 continue;
3711 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3712 }
3713 critical_exit();
3714
3715 /*
3716 * We can not get a bucket so try to return a single item.
3717 */
3718 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3719 domain = PCPU_GET(domain);
3720 else
3721 domain = UMA_ANYDOMAIN;
3722 return (zone_alloc_item(zone, udata, domain, flags));
3723 }
3724
3725 /* See uma.h */
3726 void *
uma_zalloc_smr(uma_zone_t zone,int flags)3727 uma_zalloc_smr(uma_zone_t zone, int flags)
3728 {
3729 uma_cache_bucket_t bucket;
3730 uma_cache_t cache;
3731
3732 CTR3(KTR_UMA, "uma_zalloc_smr zone %s(%p) flags %d", zone->uz_name,
3733 zone, flags);
3734
3735 #ifdef UMA_ZALLOC_DEBUG
3736 void *item;
3737
3738 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3739 ("uma_zalloc_arg: called with non-SMR zone."));
3740 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3741 return (item);
3742 #endif
3743
3744 critical_enter();
3745 cache = &zone->uz_cpu[curcpu];
3746 bucket = &cache->uc_allocbucket;
3747 if (__predict_false(bucket->ucb_cnt == 0))
3748 return (cache_alloc_retry(zone, cache, NULL, flags));
3749 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3750 }
3751
3752 /* See uma.h */
3753 void *
uma_zalloc_arg(uma_zone_t zone,void * udata,int flags)3754 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3755 {
3756 uma_cache_bucket_t bucket;
3757 uma_cache_t cache;
3758
3759 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3760 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3761
3762 /* This is the fast path allocation */
3763 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3764 zone, flags);
3765
3766 #ifdef UMA_ZALLOC_DEBUG
3767 void *item;
3768
3769 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3770 ("uma_zalloc_arg: called with SMR zone."));
3771 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3772 return (item);
3773 #endif
3774
3775 /*
3776 * If possible, allocate from the per-CPU cache. There are two
3777 * requirements for safe access to the per-CPU cache: (1) the thread
3778 * accessing the cache must not be preempted or yield during access,
3779 * and (2) the thread must not migrate CPUs without switching which
3780 * cache it accesses. We rely on a critical section to prevent
3781 * preemption and migration. We release the critical section in
3782 * order to acquire the zone mutex if we are unable to allocate from
3783 * the current cache; when we re-acquire the critical section, we
3784 * must detect and handle migration if it has occurred.
3785 */
3786 critical_enter();
3787 cache = &zone->uz_cpu[curcpu];
3788 bucket = &cache->uc_allocbucket;
3789 if (__predict_false(bucket->ucb_cnt == 0))
3790 return (cache_alloc_retry(zone, cache, udata, flags));
3791 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3792 }
3793
3794 /*
3795 * Replenish an alloc bucket and possibly restore an old one. Called in
3796 * a critical section. Returns in a critical section.
3797 *
3798 * A false return value indicates an allocation failure.
3799 * A true return value indicates success and the caller should retry.
3800 */
3801 static __noinline bool
cache_alloc(uma_zone_t zone,uma_cache_t cache,void * udata,int flags)3802 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3803 {
3804 uma_bucket_t bucket;
3805 uint32_t zflags;
3806 int curdomain, domain;
3807 bool new;
3808
3809 CRITICAL_ASSERT(curthread);
3810
3811 /*
3812 * If we have run out of items in our alloc bucket see
3813 * if we can switch with the free bucket.
3814 *
3815 * SMR zones can't re-use the free bucket until the sequence has
3816 * expired. When KASAN is enabled, we want to avoid re-using free
3817 * items in order to improve reliability of use-after-free detection.
3818 */
3819 zflags = cache_uz_flags(cache);
3820 if ((zflags & UMA_ZONE_SMR) == 0 &&
3821 #ifdef KASAN
3822 (zflags & UMA_ZONE_NOKASAN) != 0 &&
3823 #endif
3824 cache->uc_freebucket.ucb_cnt != 0) {
3825 cache_bucket_swap(&cache->uc_freebucket,
3826 &cache->uc_allocbucket);
3827 return (true);
3828 }
3829
3830 /*
3831 * Discard any empty allocation bucket while we hold no locks.
3832 */
3833 bucket = cache_bucket_unload_alloc(cache);
3834 critical_exit();
3835
3836 if (bucket != NULL) {
3837 KASSERT(bucket->ub_cnt == 0,
3838 ("cache_alloc: Entered with non-empty alloc bucket."));
3839 bucket_free(zone, bucket, udata);
3840 }
3841
3842 /*
3843 * Attempt to retrieve the item from the per-CPU cache has failed, so
3844 * we must go back to the zone. This requires the zdom lock, so we
3845 * must drop the critical section, then re-acquire it when we go back
3846 * to the cache. Since the critical section is released, we may be
3847 * preempted or migrate. As such, make sure not to maintain any
3848 * thread-local state specific to the cache from prior to releasing
3849 * the critical section.
3850 */
3851 domain = PCPU_GET(domain);
3852 if ((zflags & UMA_ZONE_ROUNDROBIN) != 0 || VM_DOMAIN_EMPTY(domain))
3853 domain = zone_domain_highest(zone, domain);
3854 bucket = cache_fetch_bucket(zone, cache, domain);
3855 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3856 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3857 new = true;
3858 } else {
3859 new = false;
3860 }
3861
3862 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3863 zone->uz_name, zone, bucket);
3864 if (bucket == NULL) {
3865 critical_enter();
3866 return (false);
3867 }
3868
3869 /*
3870 * See if we lost the race or were migrated. Cache the
3871 * initialized bucket to make this less likely or claim
3872 * the memory directly.
3873 */
3874 critical_enter();
3875 cache = &zone->uz_cpu[curcpu];
3876 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3877 ((zflags & UMA_ZONE_FIRSTTOUCH) == 0 ||
3878 (curdomain = PCPU_GET(domain)) == domain ||
3879 VM_DOMAIN_EMPTY(curdomain))) {
3880 if (new)
3881 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3882 bucket->ub_cnt);
3883 cache_bucket_load_alloc(cache, bucket);
3884 return (true);
3885 }
3886
3887 /*
3888 * We lost the race, release this bucket and start over.
3889 */
3890 critical_exit();
3891 zone_put_bucket(zone, domain, bucket, udata, !new);
3892 critical_enter();
3893
3894 return (true);
3895 }
3896
3897 void *
uma_zalloc_domain(uma_zone_t zone,void * udata,int domain,int flags)3898 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3899 {
3900 #ifdef NUMA
3901 uma_bucket_t bucket;
3902 uma_zone_domain_t zdom;
3903 void *item;
3904 #endif
3905
3906 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3907 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3908
3909 /* This is the fast path allocation */
3910 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3911 zone->uz_name, zone, domain, flags);
3912
3913 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3914 ("uma_zalloc_domain: called with SMR zone."));
3915 #ifdef NUMA
3916 KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3917 ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3918
3919 if (vm_ndomains == 1)
3920 return (uma_zalloc_arg(zone, udata, flags));
3921
3922 #ifdef UMA_ZALLOC_DEBUG
3923 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3924 return (item);
3925 #endif
3926
3927 /*
3928 * Try to allocate from the bucket cache before falling back to the keg.
3929 * We could try harder and attempt to allocate from per-CPU caches or
3930 * the per-domain cross-domain buckets, but the complexity is probably
3931 * not worth it. It is more important that frees of previous
3932 * cross-domain allocations do not blow up the cache.
3933 */
3934 zdom = zone_domain_lock(zone, domain);
3935 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3936 item = bucket->ub_bucket[bucket->ub_cnt - 1];
3937 #ifdef INVARIANTS
3938 bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3939 #endif
3940 bucket->ub_cnt--;
3941 zone_put_bucket(zone, domain, bucket, udata, true);
3942 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3943 flags, item);
3944 if (item != NULL) {
3945 KASSERT(item_domain(item) == domain,
3946 ("%s: bucket cache item %p from wrong domain",
3947 __func__, item));
3948 counter_u64_add(zone->uz_allocs, 1);
3949 }
3950 return (item);
3951 }
3952 ZDOM_UNLOCK(zdom);
3953 return (zone_alloc_item(zone, udata, domain, flags));
3954 #else
3955 return (uma_zalloc_arg(zone, udata, flags));
3956 #endif
3957 }
3958
3959 /*
3960 * Find a slab with some space. Prefer slabs that are partially used over those
3961 * that are totally full. This helps to reduce fragmentation.
3962 *
3963 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3964 * only 'domain'.
3965 */
3966 static uma_slab_t
keg_first_slab(uma_keg_t keg,int domain,bool rr)3967 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3968 {
3969 uma_domain_t dom;
3970 uma_slab_t slab;
3971 int start;
3972
3973 KASSERT(domain >= 0 && domain < vm_ndomains,
3974 ("keg_first_slab: domain %d out of range", domain));
3975 KEG_LOCK_ASSERT(keg, domain);
3976
3977 slab = NULL;
3978 start = domain;
3979 do {
3980 dom = &keg->uk_domain[domain];
3981 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3982 return (slab);
3983 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3984 LIST_REMOVE(slab, us_link);
3985 dom->ud_free_slabs--;
3986 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3987 return (slab);
3988 }
3989 if (rr)
3990 domain = (domain + 1) % vm_ndomains;
3991 } while (domain != start);
3992
3993 return (NULL);
3994 }
3995
3996 /*
3997 * Fetch an existing slab from a free or partial list. Returns with the
3998 * keg domain lock held if a slab was found or unlocked if not.
3999 */
4000 static uma_slab_t
keg_fetch_free_slab(uma_keg_t keg,int domain,bool rr,int flags)4001 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
4002 {
4003 uma_slab_t slab;
4004 uint32_t reserve;
4005
4006 /* HASH has a single free list. */
4007 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
4008 domain = 0;
4009
4010 KEG_LOCK(keg, domain);
4011 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
4012 if (keg->uk_domain[domain].ud_free_items <= reserve ||
4013 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
4014 KEG_UNLOCK(keg, domain);
4015 return (NULL);
4016 }
4017 return (slab);
4018 }
4019
4020 static uma_slab_t
keg_fetch_slab(uma_keg_t keg,uma_zone_t zone,int rdomain,const int flags)4021 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
4022 {
4023 struct vm_domainset_iter di;
4024 uma_slab_t slab;
4025 int aflags, domain;
4026 bool rr;
4027
4028 KASSERT((flags & (M_WAITOK | M_NOVM)) != (M_WAITOK | M_NOVM),
4029 ("%s: invalid flags %#x", __func__, flags));
4030
4031 restart:
4032 /*
4033 * Use the keg's policy if upper layers haven't already specified a
4034 * domain (as happens with first-touch zones).
4035 */
4036 rr = rdomain == UMA_ANYDOMAIN;
4037 aflags = flags;
4038 if (rr) {
4039 if (vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
4040 &aflags) != 0)
4041 return (NULL);
4042 } else
4043 domain = rdomain;
4044
4045 for (;;) {
4046 slab = keg_fetch_free_slab(keg, domain, rr, flags);
4047 if (slab != NULL)
4048 return (slab);
4049
4050 /*
4051 * M_NOVM is used to break the recursion that can otherwise
4052 * occur if low-level memory management routines use UMA.
4053 */
4054 if ((flags & M_NOVM) == 0) {
4055 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
4056 if (slab != NULL)
4057 return (slab);
4058 }
4059
4060 if (!rr) {
4061 if ((flags & M_USE_RESERVE) != 0) {
4062 /*
4063 * Drain reserves from other domains before
4064 * giving up or sleeping. It may be useful to
4065 * support per-domain reserves eventually.
4066 */
4067 rdomain = UMA_ANYDOMAIN;
4068 goto restart;
4069 }
4070 if ((flags & M_WAITOK) == 0)
4071 break;
4072 vm_wait_domain(domain);
4073 } else if (vm_domainset_iter_policy(&di, &domain) != 0)
4074 break;
4075 }
4076
4077 /*
4078 * We might not have been able to get a slab but another cpu
4079 * could have while we were unlocked. Check again before we
4080 * fail.
4081 */
4082 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
4083 return (slab);
4084
4085 return (NULL);
4086 }
4087
4088 static void *
slab_alloc_item(uma_keg_t keg,uma_slab_t slab)4089 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
4090 {
4091 uma_domain_t dom;
4092 void *item;
4093 int freei;
4094
4095 KEG_LOCK_ASSERT(keg, slab->us_domain);
4096
4097 dom = &keg->uk_domain[slab->us_domain];
4098 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
4099 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
4100 item = slab_item(slab, keg, freei);
4101 slab->us_freecount--;
4102 dom->ud_free_items--;
4103
4104 /*
4105 * Move this slab to the full list. It must be on the partial list, so
4106 * we do not need to update the free slab count. In particular,
4107 * keg_fetch_slab() always returns slabs on the partial list.
4108 */
4109 if (slab->us_freecount == 0) {
4110 LIST_REMOVE(slab, us_link);
4111 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
4112 }
4113
4114 return (item);
4115 }
4116
4117 static int
zone_import(void * arg,void ** bucket,int max,int domain,int flags)4118 zone_import(void *arg, void **bucket, int max, int domain, int flags)
4119 {
4120 uma_domain_t dom;
4121 uma_zone_t zone;
4122 uma_slab_t slab;
4123 uma_keg_t keg;
4124 #ifdef NUMA
4125 int stripe;
4126 #endif
4127 int i;
4128
4129 zone = arg;
4130 slab = NULL;
4131 keg = zone->uz_keg;
4132 /* Try to keep the buckets totally full */
4133 for (i = 0; i < max; ) {
4134 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
4135 break;
4136 #ifdef NUMA
4137 stripe = howmany(max, vm_ndomains);
4138 #endif
4139 dom = &keg->uk_domain[slab->us_domain];
4140 do {
4141 bucket[i++] = slab_alloc_item(keg, slab);
4142 if (keg->uk_reserve > 0 &&
4143 dom->ud_free_items <= keg->uk_reserve) {
4144 /*
4145 * Avoid depleting the reserve after a
4146 * successful item allocation, even if
4147 * M_USE_RESERVE is specified.
4148 */
4149 KEG_UNLOCK(keg, slab->us_domain);
4150 goto out;
4151 }
4152 #ifdef NUMA
4153 /*
4154 * If the zone is striped we pick a new slab for every
4155 * N allocations. Eliminating this conditional will
4156 * instead pick a new domain for each bucket rather
4157 * than stripe within each bucket. The current option
4158 * produces more fragmentation and requires more cpu
4159 * time but yields better distribution.
4160 */
4161 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
4162 vm_ndomains > 1 && --stripe == 0)
4163 break;
4164 #endif
4165 } while (slab->us_freecount != 0 && i < max);
4166 KEG_UNLOCK(keg, slab->us_domain);
4167
4168 /* Don't block if we allocated any successfully. */
4169 flags &= ~M_WAITOK;
4170 flags |= M_NOWAIT;
4171 }
4172 out:
4173 return i;
4174 }
4175
4176 static int
zone_alloc_limit_hard(uma_zone_t zone,int count,int flags)4177 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
4178 {
4179 uint64_t old, new, total, max;
4180
4181 /*
4182 * The hard case. We're going to sleep because there were existing
4183 * sleepers or because we ran out of items. This routine enforces
4184 * fairness by keeping fifo order.
4185 *
4186 * First release our ill gotten gains and make some noise.
4187 */
4188 for (;;) {
4189 zone_free_limit(zone, count);
4190 zone_log_warning(zone);
4191 zone_maxaction(zone);
4192 if (flags & M_NOWAIT)
4193 return (0);
4194
4195 /*
4196 * We need to allocate an item or set ourself as a sleeper
4197 * while the sleepq lock is held to avoid wakeup races. This
4198 * is essentially a home rolled semaphore.
4199 */
4200 sleepq_lock(&zone->uz_max_items);
4201 old = zone->uz_items;
4202 do {
4203 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
4204 /* Cache the max since we will evaluate twice. */
4205 max = zone->uz_max_items;
4206 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
4207 UZ_ITEMS_COUNT(old) >= max)
4208 new = old + UZ_ITEMS_SLEEPER;
4209 else
4210 new = old + MIN(count, max - old);
4211 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
4212
4213 /* We may have successfully allocated under the sleepq lock. */
4214 if (UZ_ITEMS_SLEEPERS(new) == 0) {
4215 sleepq_release(&zone->uz_max_items);
4216 return (new - old);
4217 }
4218
4219 /*
4220 * This is in a different cacheline from uz_items so that we
4221 * don't constantly invalidate the fastpath cacheline when we
4222 * adjust item counts. This could be limited to toggling on
4223 * transitions.
4224 */
4225 atomic_add_32(&zone->uz_sleepers, 1);
4226 atomic_add_64(&zone->uz_sleeps, 1);
4227
4228 /*
4229 * We have added ourselves as a sleeper. The sleepq lock
4230 * protects us from wakeup races. Sleep now and then retry.
4231 */
4232 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
4233 sleepq_wait(&zone->uz_max_items, PVM);
4234
4235 /*
4236 * After wakeup, remove ourselves as a sleeper and try
4237 * again. We no longer have the sleepq lock for protection.
4238 *
4239 * Subract ourselves as a sleeper while attempting to add
4240 * our count.
4241 */
4242 atomic_subtract_32(&zone->uz_sleepers, 1);
4243 old = atomic_fetchadd_64(&zone->uz_items,
4244 -(UZ_ITEMS_SLEEPER - count));
4245 /* We're no longer a sleeper. */
4246 old -= UZ_ITEMS_SLEEPER;
4247
4248 /*
4249 * If we're still at the limit, restart. Notably do not
4250 * block on other sleepers. Cache the max value to protect
4251 * against changes via sysctl.
4252 */
4253 total = UZ_ITEMS_COUNT(old);
4254 max = zone->uz_max_items;
4255 if (total >= max)
4256 continue;
4257 /* Truncate if necessary, otherwise wake other sleepers. */
4258 if (total + count > max) {
4259 zone_free_limit(zone, total + count - max);
4260 count = max - total;
4261 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
4262 wakeup_one(&zone->uz_max_items);
4263
4264 return (count);
4265 }
4266 }
4267
4268 /*
4269 * Allocate 'count' items from our max_items limit. Returns the number
4270 * available. If M_NOWAIT is not specified it will sleep until at least
4271 * one item can be allocated.
4272 */
4273 static int
zone_alloc_limit(uma_zone_t zone,int count,int flags)4274 zone_alloc_limit(uma_zone_t zone, int count, int flags)
4275 {
4276 uint64_t old;
4277 uint64_t max;
4278
4279 max = zone->uz_max_items;
4280 MPASS(max > 0);
4281
4282 /*
4283 * We expect normal allocations to succeed with a simple
4284 * fetchadd.
4285 */
4286 old = atomic_fetchadd_64(&zone->uz_items, count);
4287 if (__predict_true(old + count <= max))
4288 return (count);
4289
4290 /*
4291 * If we had some items and no sleepers just return the
4292 * truncated value. We have to release the excess space
4293 * though because that may wake sleepers who weren't woken
4294 * because we were temporarily over the limit.
4295 */
4296 if (old < max) {
4297 zone_free_limit(zone, (old + count) - max);
4298 return (max - old);
4299 }
4300 return (zone_alloc_limit_hard(zone, count, flags));
4301 }
4302
4303 /*
4304 * Free a number of items back to the limit.
4305 */
4306 static void
zone_free_limit(uma_zone_t zone,int count)4307 zone_free_limit(uma_zone_t zone, int count)
4308 {
4309 uint64_t old;
4310
4311 MPASS(count > 0);
4312
4313 /*
4314 * In the common case we either have no sleepers or
4315 * are still over the limit and can just return.
4316 */
4317 old = atomic_fetchadd_64(&zone->uz_items, -count);
4318 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
4319 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
4320 return;
4321
4322 /*
4323 * Moderate the rate of wakeups. Sleepers will continue
4324 * to generate wakeups if necessary.
4325 */
4326 wakeup_one(&zone->uz_max_items);
4327 }
4328
4329 static uma_bucket_t
zone_alloc_bucket(uma_zone_t zone,void * udata,int domain,int flags)4330 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
4331 {
4332 uma_bucket_t bucket;
4333 int error, maxbucket, cnt;
4334
4335 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
4336 zone, domain);
4337
4338 /* Avoid allocs targeting empty domains. */
4339 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4340 domain = UMA_ANYDOMAIN;
4341 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4342 domain = UMA_ANYDOMAIN;
4343
4344 if (zone->uz_max_items > 0)
4345 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
4346 M_NOWAIT);
4347 else
4348 maxbucket = zone->uz_bucket_size;
4349 if (maxbucket == 0)
4350 return (NULL);
4351
4352 /* Don't wait for buckets, preserve caller's NOVM setting. */
4353 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
4354 if (bucket == NULL) {
4355 cnt = 0;
4356 goto out;
4357 }
4358
4359 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
4360 MIN(maxbucket, bucket->ub_entries), domain, flags);
4361
4362 /*
4363 * Initialize the memory if necessary.
4364 */
4365 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4366 int i;
4367
4368 for (i = 0; i < bucket->ub_cnt; i++) {
4369 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
4370 error = zone->uz_init(bucket->ub_bucket[i],
4371 zone->uz_size, flags);
4372 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
4373 if (error != 0)
4374 break;
4375 }
4376
4377 /*
4378 * If we couldn't initialize the whole bucket, put the
4379 * rest back onto the freelist.
4380 */
4381 if (i != bucket->ub_cnt) {
4382 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4383 bucket->ub_cnt - i);
4384 #ifdef INVARIANTS
4385 bzero(&bucket->ub_bucket[i],
4386 sizeof(void *) * (bucket->ub_cnt - i));
4387 #endif
4388 bucket->ub_cnt = i;
4389 }
4390 }
4391
4392 cnt = bucket->ub_cnt;
4393 if (bucket->ub_cnt == 0) {
4394 bucket_free(zone, bucket, udata);
4395 counter_u64_add(zone->uz_fails, 1);
4396 bucket = NULL;
4397 }
4398 out:
4399 if (zone->uz_max_items > 0 && cnt < maxbucket)
4400 zone_free_limit(zone, maxbucket - cnt);
4401
4402 return (bucket);
4403 }
4404
4405 /*
4406 * Allocates a single item from a zone.
4407 *
4408 * Arguments
4409 * zone The zone to alloc for.
4410 * udata The data to be passed to the constructor.
4411 * domain The domain to allocate from or UMA_ANYDOMAIN.
4412 * flags M_WAITOK, M_NOWAIT, M_ZERO.
4413 *
4414 * Returns
4415 * NULL if there is no memory and M_NOWAIT is set
4416 * An item if successful
4417 */
4418
4419 static void *
zone_alloc_item(uma_zone_t zone,void * udata,int domain,int flags)4420 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4421 {
4422 void *item;
4423
4424 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4425 counter_u64_add(zone->uz_fails, 1);
4426 return (NULL);
4427 }
4428
4429 /* Avoid allocs targeting empty domains. */
4430 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4431 domain = UMA_ANYDOMAIN;
4432
4433 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4434 goto fail_cnt;
4435
4436 /*
4437 * We have to call both the zone's init (not the keg's init)
4438 * and the zone's ctor. This is because the item is going from
4439 * a keg slab directly to the user, and the user is expecting it
4440 * to be both zone-init'd as well as zone-ctor'd.
4441 */
4442 if (zone->uz_init != NULL) {
4443 int error;
4444
4445 kasan_mark_item_valid(zone, item);
4446 error = zone->uz_init(item, zone->uz_size, flags);
4447 kasan_mark_item_invalid(zone, item);
4448 if (error != 0) {
4449 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4450 goto fail_cnt;
4451 }
4452 }
4453 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4454 item);
4455 if (item == NULL)
4456 goto fail;
4457
4458 counter_u64_add(zone->uz_allocs, 1);
4459 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4460 zone->uz_name, zone);
4461
4462 return (item);
4463
4464 fail_cnt:
4465 counter_u64_add(zone->uz_fails, 1);
4466 fail:
4467 if (zone->uz_max_items > 0)
4468 zone_free_limit(zone, 1);
4469 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4470 zone->uz_name, zone);
4471
4472 return (NULL);
4473 }
4474
4475 /*
4476 * Try to free an item to the per-CPU cache, promoting its quick reuse.
4477 */
4478 static __always_inline bool
cache_free_reuse(uma_zone_t zone,int uz_flags,void * item,void * udata)4479 cache_free_reuse(uma_zone_t zone, int uz_flags, void *item, void *udata)
4480 {
4481 uma_cache_t cache;
4482 int itemdomain;
4483
4484 /*
4485 * If possible, free to the per-CPU cache. There are two
4486 * requirements for safe access to the per-CPU cache: (1) the thread
4487 * accessing the cache must not be preempted or yield during access,
4488 * and (2) the thread must not migrate CPUs without switching which
4489 * cache it accesses. We rely on a critical section to prevent
4490 * preemption and migration. We release the critical section in
4491 * order to acquire the zone mutex if we are unable to free to the
4492 * current cache; when we re-acquire the critical section, we must
4493 * detect and handle migration if it has occurred.
4494 */
4495 itemdomain = 0;
4496 #ifdef NUMA
4497 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4498 itemdomain = item_domain(item);
4499 #endif
4500
4501 critical_enter();
4502 do {
4503 uma_cache_bucket_t bucket;
4504
4505 cache = &zone->uz_cpu[curcpu];
4506 /*
4507 * Try to free into the allocbucket first to give LIFO
4508 * ordering for cache-hot datastructures. Spill over
4509 * into the freebucket if necessary. Alloc will swap
4510 * them if one runs dry.
4511 */
4512 bucket = &cache->uc_allocbucket;
4513 #ifdef NUMA
4514 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4515 PCPU_GET(domain) != itemdomain) {
4516 bucket = &cache->uc_crossbucket;
4517 } else
4518 #endif
4519 if (bucket->ucb_cnt == bucket->ucb_entries &&
4520 cache->uc_freebucket.ucb_cnt <
4521 cache->uc_freebucket.ucb_entries)
4522 cache_bucket_swap(&cache->uc_freebucket,
4523 &cache->uc_allocbucket);
4524 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4525 cache_bucket_push(cache, bucket, item);
4526 critical_exit();
4527 return (true);
4528 }
4529 } while (cache_free(zone, cache, udata, itemdomain));
4530 critical_exit();
4531
4532 return (false);
4533 }
4534
4535 /*
4536 * Try to free an object to the per-CPU cache, deferring its reuse. This is
4537 * used by the SMR-protected allocator, which cannot reuse the item until
4538 * smr_poll() guarantees that no threads are still accessing it, and by
4539 * sanitizers, which wish to defer reuse to make UAF detection more effective.
4540 */
4541 static __always_inline bool
cache_free_defer(uma_zone_t zone,void * item,void * udata)4542 cache_free_defer(uma_zone_t zone, void *item, void *udata)
4543 {
4544 uma_cache_t cache;
4545 int itemdomain;
4546 #ifdef NUMA
4547 int uz_flags;
4548 #endif
4549
4550 itemdomain = 0;
4551 #ifdef NUMA
4552 uz_flags = cache_uz_flags(&zone->uz_cpu[curcpu]);
4553 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4554 itemdomain = item_domain(item);
4555 #endif
4556 critical_enter();
4557 do {
4558 uma_cache_bucket_t bucket;
4559
4560 cache = &zone->uz_cpu[curcpu];
4561 bucket = &cache->uc_freebucket;
4562 #ifdef NUMA
4563 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4564 PCPU_GET(domain) != itemdomain) {
4565 bucket = &cache->uc_crossbucket;
4566 }
4567 #endif
4568 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4569 cache_bucket_push(cache, bucket, item);
4570 critical_exit();
4571 return (true);
4572 }
4573 } while (cache_free(zone, cache, udata, itemdomain));
4574 critical_exit();
4575
4576 return (false);
4577 }
4578
4579 /* See uma.h */
4580 void
uma_zfree_smr(uma_zone_t zone,void * item)4581 uma_zfree_smr(uma_zone_t zone, void *item)
4582 {
4583 CTR3(KTR_UMA, "uma_zfree_smr zone %s(%p) item %p",
4584 zone->uz_name, zone, item);
4585
4586 #ifdef UMA_ZALLOC_DEBUG
4587 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4588 ("uma_zfree_smr: called with non-SMR zone."));
4589 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4590 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4591 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4592 return;
4593 #endif
4594
4595 if (cache_free_defer(zone, item, NULL))
4596 return;
4597
4598 /*
4599 * If nothing else caught this, we'll just do an internal free.
4600 */
4601 zone_free_item(zone, item, NULL, SKIP_NONE);
4602 }
4603
4604 /* See uma.h */
4605 void
uma_zfree_arg(uma_zone_t zone,void * item,void * udata)4606 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4607 {
4608 uma_cache_t cache;
4609 int uz_flags;
4610
4611 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4612 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4613
4614 CTR3(KTR_UMA, "uma_zfree_arg zone %s(%p) item %p",
4615 zone->uz_name, zone, item);
4616
4617 #ifdef UMA_ZALLOC_DEBUG
4618 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4619 ("uma_zfree_arg: called with SMR zone."));
4620 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4621 return;
4622 #endif
4623 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4624 if (item == NULL)
4625 return;
4626
4627 /*
4628 * We are accessing the per-cpu cache without a critical section to
4629 * fetch size and flags. This is acceptable, if we are preempted we
4630 * will simply read another cpu's line.
4631 */
4632 cache = &zone->uz_cpu[curcpu];
4633 uz_flags = cache_uz_flags(cache);
4634 if (UMA_ALWAYS_CTORDTOR ||
4635 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4636 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4637
4638 /*
4639 * The race here is acceptable. If we miss it we'll just have to wait
4640 * a little longer for the limits to be reset.
4641 */
4642 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT) &&
4643 atomic_load_32(&zone->uz_sleepers) > 0) {
4644 /* We will free directly to the zone. */
4645 }
4646 #ifdef KASAN
4647 else if ((uz_flags & UMA_ZONE_NOKASAN) == 0) {
4648 if (cache_free_defer(zone, item, udata))
4649 return;
4650 }
4651 #endif
4652 else if (cache_free_reuse(zone, uz_flags, item, udata)) {
4653 return;
4654 }
4655
4656 /*
4657 * If nothing else caught this, we'll just do an internal free.
4658 */
4659 zone_free_item(zone, item, udata, SKIP_DTOR);
4660 }
4661
4662 #ifdef NUMA
4663 /*
4664 * sort crossdomain free buckets to domain correct buckets and cache
4665 * them.
4666 */
4667 static void
zone_free_cross(uma_zone_t zone,uma_bucket_t bucket,void * udata)4668 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4669 {
4670 struct uma_bucketlist emptybuckets, fullbuckets;
4671 uma_zone_domain_t zdom;
4672 uma_bucket_t b;
4673 smr_seq_t seq;
4674 void *item;
4675 int domain;
4676
4677 CTR3(KTR_UMA,
4678 "uma_zfree: zone %s(%p) draining cross bucket %p",
4679 zone->uz_name, zone, bucket);
4680
4681 /*
4682 * It is possible for buckets to arrive here out of order so we fetch
4683 * the current smr seq rather than accepting the bucket's.
4684 */
4685 seq = SMR_SEQ_INVALID;
4686 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4687 seq = smr_advance(zone->uz_smr);
4688
4689 /*
4690 * To avoid having ndomain * ndomain buckets for sorting we have a
4691 * lock on the current crossfree bucket. A full matrix with
4692 * per-domain locking could be used if necessary.
4693 */
4694 STAILQ_INIT(&emptybuckets);
4695 STAILQ_INIT(&fullbuckets);
4696 ZONE_CROSS_LOCK(zone);
4697 for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4698 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4699 domain = item_domain(item);
4700 zdom = ZDOM_GET(zone, domain);
4701 if (zdom->uzd_cross == NULL) {
4702 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4703 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4704 zdom->uzd_cross = b;
4705 } else {
4706 /*
4707 * Avoid allocating a bucket with the cross lock
4708 * held, since allocation can trigger a
4709 * cross-domain free and bucket zones may
4710 * allocate from each other.
4711 */
4712 ZONE_CROSS_UNLOCK(zone);
4713 b = bucket_alloc(zone, udata, M_NOWAIT);
4714 if (b == NULL)
4715 goto out;
4716 ZONE_CROSS_LOCK(zone);
4717 if (zdom->uzd_cross != NULL) {
4718 STAILQ_INSERT_HEAD(&emptybuckets, b,
4719 ub_link);
4720 } else {
4721 zdom->uzd_cross = b;
4722 }
4723 }
4724 }
4725 b = zdom->uzd_cross;
4726 b->ub_bucket[b->ub_cnt++] = item;
4727 b->ub_seq = seq;
4728 if (b->ub_cnt == b->ub_entries) {
4729 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4730 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4731 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4732 zdom->uzd_cross = b;
4733 }
4734 }
4735 ZONE_CROSS_UNLOCK(zone);
4736 out:
4737 if (bucket->ub_cnt == 0)
4738 bucket->ub_seq = SMR_SEQ_INVALID;
4739 bucket_free(zone, bucket, udata);
4740
4741 while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4742 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4743 bucket_free(zone, b, udata);
4744 }
4745 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4746 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4747 domain = item_domain(b->ub_bucket[0]);
4748 zone_put_bucket(zone, domain, b, udata, true);
4749 }
4750 }
4751 #endif
4752
4753 static void
zone_free_bucket(uma_zone_t zone,uma_bucket_t bucket,void * udata,int itemdomain,bool ws)4754 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4755 int itemdomain, bool ws)
4756 {
4757
4758 #ifdef NUMA
4759 /*
4760 * Buckets coming from the wrong domain will be entirely for the
4761 * only other domain on two domain systems. In this case we can
4762 * simply cache them. Otherwise we need to sort them back to
4763 * correct domains.
4764 */
4765 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4766 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4767 zone_free_cross(zone, bucket, udata);
4768 return;
4769 }
4770 #endif
4771
4772 /*
4773 * Attempt to save the bucket in the zone's domain bucket cache.
4774 */
4775 CTR3(KTR_UMA,
4776 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4777 zone->uz_name, zone, bucket);
4778 /* ub_cnt is pointing to the last free item */
4779 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4780 itemdomain = zone_domain_lowest(zone, itemdomain);
4781 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4782 }
4783
4784 /*
4785 * Populate a free or cross bucket for the current cpu cache. Free any
4786 * existing full bucket either to the zone cache or back to the slab layer.
4787 *
4788 * Enters and returns in a critical section. false return indicates that
4789 * we can not satisfy this free in the cache layer. true indicates that
4790 * the caller should retry.
4791 */
4792 static __noinline bool
cache_free(uma_zone_t zone,uma_cache_t cache,void * udata,int itemdomain)4793 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, int itemdomain)
4794 {
4795 uma_cache_bucket_t cbucket;
4796 uma_bucket_t newbucket, bucket;
4797
4798 CRITICAL_ASSERT(curthread);
4799
4800 if (zone->uz_bucket_size == 0)
4801 return false;
4802
4803 cache = &zone->uz_cpu[curcpu];
4804 newbucket = NULL;
4805
4806 /*
4807 * FIRSTTOUCH domains need to free to the correct zdom. When
4808 * enabled this is the zdom of the item. The bucket is the
4809 * cross bucket if the current domain and itemdomain do not match.
4810 */
4811 cbucket = &cache->uc_freebucket;
4812 #ifdef NUMA
4813 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4814 if (PCPU_GET(domain) != itemdomain) {
4815 cbucket = &cache->uc_crossbucket;
4816 if (cbucket->ucb_cnt != 0)
4817 counter_u64_add(zone->uz_xdomain,
4818 cbucket->ucb_cnt);
4819 }
4820 }
4821 #endif
4822 bucket = cache_bucket_unload(cbucket);
4823 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4824 ("cache_free: Entered with non-full free bucket."));
4825
4826 /* We are no longer associated with this CPU. */
4827 critical_exit();
4828
4829 /*
4830 * Don't let SMR zones operate without a free bucket. Force
4831 * a synchronize and re-use this one. We will only degrade
4832 * to a synchronize every bucket_size items rather than every
4833 * item if we fail to allocate a bucket.
4834 */
4835 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4836 if (bucket != NULL)
4837 bucket->ub_seq = smr_advance(zone->uz_smr);
4838 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4839 if (newbucket == NULL && bucket != NULL) {
4840 bucket_drain(zone, bucket);
4841 newbucket = bucket;
4842 bucket = NULL;
4843 }
4844 } else if (!bucketdisable)
4845 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4846
4847 if (bucket != NULL)
4848 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4849
4850 critical_enter();
4851 if ((bucket = newbucket) == NULL)
4852 return (false);
4853 cache = &zone->uz_cpu[curcpu];
4854 #ifdef NUMA
4855 /*
4856 * Check to see if we should be populating the cross bucket. If it
4857 * is already populated we will fall through and attempt to populate
4858 * the free bucket.
4859 */
4860 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4861 if (PCPU_GET(domain) != itemdomain &&
4862 cache->uc_crossbucket.ucb_bucket == NULL) {
4863 cache_bucket_load_cross(cache, bucket);
4864 return (true);
4865 }
4866 }
4867 #endif
4868 /*
4869 * We may have lost the race to fill the bucket or switched CPUs.
4870 */
4871 if (cache->uc_freebucket.ucb_bucket != NULL) {
4872 critical_exit();
4873 bucket_free(zone, bucket, udata);
4874 critical_enter();
4875 } else
4876 cache_bucket_load_free(cache, bucket);
4877
4878 return (true);
4879 }
4880
4881 static void
slab_free_item(uma_zone_t zone,uma_slab_t slab,void * item)4882 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4883 {
4884 uma_keg_t keg;
4885 uma_domain_t dom;
4886 int freei;
4887
4888 keg = zone->uz_keg;
4889 KEG_LOCK_ASSERT(keg, slab->us_domain);
4890
4891 /* Do we need to remove from any lists? */
4892 dom = &keg->uk_domain[slab->us_domain];
4893 if (slab->us_freecount + 1 == keg->uk_ipers) {
4894 LIST_REMOVE(slab, us_link);
4895 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4896 dom->ud_free_slabs++;
4897 } else if (slab->us_freecount == 0) {
4898 LIST_REMOVE(slab, us_link);
4899 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4900 }
4901
4902 /* Slab management. */
4903 freei = slab_item_index(slab, keg, item);
4904 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4905 slab->us_freecount++;
4906
4907 /* Keg statistics. */
4908 dom->ud_free_items++;
4909 }
4910
4911 static void
zone_release(void * arg,void ** bucket,int cnt)4912 zone_release(void *arg, void **bucket, int cnt)
4913 {
4914 struct mtx *lock;
4915 uma_zone_t zone;
4916 uma_slab_t slab;
4917 uma_keg_t keg;
4918 uint8_t *mem;
4919 void *item;
4920 int i;
4921
4922 zone = arg;
4923 keg = zone->uz_keg;
4924 lock = NULL;
4925 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4926 lock = KEG_LOCK(keg, 0);
4927 for (i = 0; i < cnt; i++) {
4928 item = bucket[i];
4929 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4930 slab = vtoslab((vm_offset_t)item);
4931 } else {
4932 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4933 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4934 slab = hash_sfind(&keg->uk_hash, mem);
4935 else
4936 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4937 }
4938 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4939 if (lock != NULL)
4940 mtx_unlock(lock);
4941 lock = KEG_LOCK(keg, slab->us_domain);
4942 }
4943 slab_free_item(zone, slab, item);
4944 }
4945 if (lock != NULL)
4946 mtx_unlock(lock);
4947 }
4948
4949 /*
4950 * Frees a single item to any zone.
4951 *
4952 * Arguments:
4953 * zone The zone to free to
4954 * item The item we're freeing
4955 * udata User supplied data for the dtor
4956 * skip Skip dtors and finis
4957 */
4958 static __noinline void
zone_free_item(uma_zone_t zone,void * item,void * udata,enum zfreeskip skip)4959 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4960 {
4961
4962 /*
4963 * If a free is sent directly to an SMR zone we have to
4964 * synchronize immediately because the item can instantly
4965 * be reallocated. This should only happen in degenerate
4966 * cases when no memory is available for per-cpu caches.
4967 */
4968 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4969 smr_synchronize(zone->uz_smr);
4970
4971 item_dtor(zone, item, zone->uz_size, udata, skip);
4972
4973 if (skip < SKIP_FINI && zone->uz_fini) {
4974 kasan_mark_item_valid(zone, item);
4975 zone->uz_fini(item, zone->uz_size);
4976 kasan_mark_item_invalid(zone, item);
4977 }
4978
4979 zone->uz_release(zone->uz_arg, &item, 1);
4980
4981 if (skip & SKIP_CNT)
4982 return;
4983
4984 counter_u64_add(zone->uz_frees, 1);
4985
4986 if (zone->uz_max_items > 0)
4987 zone_free_limit(zone, 1);
4988 }
4989
4990 /* See uma.h */
4991 int
uma_zone_set_max(uma_zone_t zone,int nitems)4992 uma_zone_set_max(uma_zone_t zone, int nitems)
4993 {
4994
4995 /*
4996 * If the limit is small, we may need to constrain the maximum per-CPU
4997 * cache size, or disable caching entirely.
4998 */
4999 uma_zone_set_maxcache(zone, nitems);
5000
5001 /*
5002 * XXX This can misbehave if the zone has any allocations with
5003 * no limit and a limit is imposed. There is currently no
5004 * way to clear a limit.
5005 */
5006 ZONE_LOCK(zone);
5007 if (zone->uz_max_items == 0)
5008 ZONE_ASSERT_COLD(zone);
5009 zone->uz_max_items = nitems;
5010 zone->uz_flags |= UMA_ZFLAG_LIMIT;
5011 zone_update_caches(zone);
5012 /* We may need to wake waiters. */
5013 wakeup(&zone->uz_max_items);
5014 ZONE_UNLOCK(zone);
5015
5016 return (nitems);
5017 }
5018
5019 /* See uma.h */
5020 void
uma_zone_set_maxcache(uma_zone_t zone,int nitems)5021 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
5022 {
5023 int bpcpu, bpdom, bsize, nb;
5024
5025 ZONE_LOCK(zone);
5026
5027 /*
5028 * Compute a lower bound on the number of items that may be cached in
5029 * the zone. Each CPU gets at least two buckets, and for cross-domain
5030 * frees we use an additional bucket per CPU and per domain. Select the
5031 * largest bucket size that does not exceed half of the requested limit,
5032 * with the left over space given to the full bucket cache.
5033 */
5034 bpdom = 0;
5035 bpcpu = 2;
5036 #ifdef NUMA
5037 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
5038 bpcpu++;
5039 bpdom++;
5040 }
5041 #endif
5042 nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
5043 bsize = nitems / nb / 2;
5044 if (bsize > BUCKET_MAX)
5045 bsize = BUCKET_MAX;
5046 else if (bsize == 0 && nitems / nb > 0)
5047 bsize = 1;
5048 zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
5049 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
5050 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
5051 zone->uz_bucket_max = nitems - nb * bsize;
5052 ZONE_UNLOCK(zone);
5053 }
5054
5055 /* See uma.h */
5056 int
uma_zone_get_max(uma_zone_t zone)5057 uma_zone_get_max(uma_zone_t zone)
5058 {
5059 int nitems;
5060
5061 nitems = atomic_load_64(&zone->uz_max_items);
5062
5063 return (nitems);
5064 }
5065
5066 /* See uma.h */
5067 void
uma_zone_set_warning(uma_zone_t zone,const char * warning)5068 uma_zone_set_warning(uma_zone_t zone, const char *warning)
5069 {
5070
5071 ZONE_ASSERT_COLD(zone);
5072 zone->uz_warning = warning;
5073 }
5074
5075 /* See uma.h */
5076 void
uma_zone_set_maxaction(uma_zone_t zone,uma_maxaction_t maxaction)5077 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
5078 {
5079
5080 ZONE_ASSERT_COLD(zone);
5081 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
5082 }
5083
5084 /* See uma.h */
5085 int
uma_zone_get_cur(uma_zone_t zone)5086 uma_zone_get_cur(uma_zone_t zone)
5087 {
5088 int64_t nitems;
5089 u_int i;
5090
5091 nitems = 0;
5092 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
5093 nitems = counter_u64_fetch(zone->uz_allocs) -
5094 counter_u64_fetch(zone->uz_frees);
5095 CPU_FOREACH(i)
5096 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
5097 atomic_load_64(&zone->uz_cpu[i].uc_frees);
5098
5099 return (nitems < 0 ? 0 : nitems);
5100 }
5101
5102 static uint64_t
uma_zone_get_allocs(uma_zone_t zone)5103 uma_zone_get_allocs(uma_zone_t zone)
5104 {
5105 uint64_t nitems;
5106 u_int i;
5107
5108 nitems = 0;
5109 if (zone->uz_allocs != EARLY_COUNTER)
5110 nitems = counter_u64_fetch(zone->uz_allocs);
5111 CPU_FOREACH(i)
5112 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
5113
5114 return (nitems);
5115 }
5116
5117 static uint64_t
uma_zone_get_frees(uma_zone_t zone)5118 uma_zone_get_frees(uma_zone_t zone)
5119 {
5120 uint64_t nitems;
5121 u_int i;
5122
5123 nitems = 0;
5124 if (zone->uz_frees != EARLY_COUNTER)
5125 nitems = counter_u64_fetch(zone->uz_frees);
5126 CPU_FOREACH(i)
5127 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
5128
5129 return (nitems);
5130 }
5131
5132 #ifdef INVARIANTS
5133 /* Used only for KEG_ASSERT_COLD(). */
5134 static uint64_t
uma_keg_get_allocs(uma_keg_t keg)5135 uma_keg_get_allocs(uma_keg_t keg)
5136 {
5137 uma_zone_t z;
5138 uint64_t nitems;
5139
5140 nitems = 0;
5141 LIST_FOREACH(z, &keg->uk_zones, uz_link)
5142 nitems += uma_zone_get_allocs(z);
5143
5144 return (nitems);
5145 }
5146 #endif
5147
5148 /* See uma.h */
5149 void
uma_zone_set_init(uma_zone_t zone,uma_init uminit)5150 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
5151 {
5152 uma_keg_t keg;
5153
5154 KEG_GET(zone, keg);
5155 KEG_ASSERT_COLD(keg);
5156 keg->uk_init = uminit;
5157 }
5158
5159 /* See uma.h */
5160 void
uma_zone_set_fini(uma_zone_t zone,uma_fini fini)5161 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
5162 {
5163 uma_keg_t keg;
5164
5165 KEG_GET(zone, keg);
5166 KEG_ASSERT_COLD(keg);
5167 keg->uk_fini = fini;
5168 }
5169
5170 /* See uma.h */
5171 void
uma_zone_set_zinit(uma_zone_t zone,uma_init zinit)5172 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
5173 {
5174
5175 ZONE_ASSERT_COLD(zone);
5176 zone->uz_init = zinit;
5177 }
5178
5179 /* See uma.h */
5180 void
uma_zone_set_zfini(uma_zone_t zone,uma_fini zfini)5181 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
5182 {
5183
5184 ZONE_ASSERT_COLD(zone);
5185 zone->uz_fini = zfini;
5186 }
5187
5188 /* See uma.h */
5189 void
uma_zone_set_freef(uma_zone_t zone,uma_free freef)5190 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
5191 {
5192 uma_keg_t keg;
5193
5194 KEG_GET(zone, keg);
5195 KEG_ASSERT_COLD(keg);
5196 keg->uk_freef = freef;
5197 }
5198
5199 /* See uma.h */
5200 void
uma_zone_set_allocf(uma_zone_t zone,uma_alloc allocf)5201 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
5202 {
5203 uma_keg_t keg;
5204
5205 KEG_GET(zone, keg);
5206 KEG_ASSERT_COLD(keg);
5207 keg->uk_allocf = allocf;
5208 }
5209
5210 /* See uma.h */
5211 void
uma_zone_set_smr(uma_zone_t zone,smr_t smr)5212 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
5213 {
5214
5215 ZONE_ASSERT_COLD(zone);
5216
5217 KASSERT(smr != NULL, ("Got NULL smr"));
5218 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
5219 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
5220 zone->uz_flags |= UMA_ZONE_SMR;
5221 zone->uz_smr = smr;
5222 zone_update_caches(zone);
5223 }
5224
5225 smr_t
uma_zone_get_smr(uma_zone_t zone)5226 uma_zone_get_smr(uma_zone_t zone)
5227 {
5228
5229 return (zone->uz_smr);
5230 }
5231
5232 /* See uma.h */
5233 void
uma_zone_reserve(uma_zone_t zone,int items)5234 uma_zone_reserve(uma_zone_t zone, int items)
5235 {
5236 uma_keg_t keg;
5237
5238 KEG_GET(zone, keg);
5239 KEG_ASSERT_COLD(keg);
5240 keg->uk_reserve = items;
5241 }
5242
5243 /* See uma.h */
5244 int
uma_zone_reserve_kva(uma_zone_t zone,int count)5245 uma_zone_reserve_kva(uma_zone_t zone, int count)
5246 {
5247 uma_keg_t keg;
5248 void *kva;
5249 u_int pages;
5250
5251 KEG_GET(zone, keg);
5252 KEG_ASSERT_COLD(keg);
5253 ZONE_ASSERT_COLD(zone);
5254
5255 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
5256
5257 #ifdef UMA_USE_DMAP
5258 if (keg->uk_ppera > 1) {
5259 #else
5260 if (1) {
5261 #endif
5262 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
5263 if (kva == NULL)
5264 return (0);
5265 } else
5266 kva = NULL;
5267
5268 MPASS(keg->uk_kva == NULL);
5269 keg->uk_kva = kva;
5270 keg->uk_offset = 0;
5271 zone->uz_max_items = pages * keg->uk_ipers;
5272 #ifdef UMA_USE_DMAP
5273 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
5274 #else
5275 keg->uk_allocf = noobj_alloc;
5276 #endif
5277 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5278 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5279 zone_update_caches(zone);
5280
5281 return (1);
5282 }
5283
5284 /* See uma.h */
5285 void
5286 uma_prealloc(uma_zone_t zone, int items)
5287 {
5288 struct vm_domainset_iter di;
5289 uma_domain_t dom;
5290 uma_slab_t slab;
5291 uma_keg_t keg;
5292 int aflags, domain, slabs;
5293
5294 KEG_GET(zone, keg);
5295 slabs = howmany(items, keg->uk_ipers);
5296 while (slabs-- > 0) {
5297 aflags = M_WAITOK;
5298 if (vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
5299 &aflags) != 0)
5300 panic("%s: Domainset is empty", __func__);
5301 for (;;) {
5302 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
5303 aflags);
5304 if (slab != NULL) {
5305 dom = &keg->uk_domain[slab->us_domain];
5306 /*
5307 * keg_alloc_slab() always returns a slab on the
5308 * partial list.
5309 */
5310 LIST_REMOVE(slab, us_link);
5311 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
5312 us_link);
5313 dom->ud_free_slabs++;
5314 KEG_UNLOCK(keg, slab->us_domain);
5315 break;
5316 }
5317 if (vm_domainset_iter_policy(&di, &domain) != 0)
5318 panic("%s: Cannot allocate from any domain",
5319 __func__);
5320 }
5321 }
5322 }
5323
5324 /*
5325 * Returns a snapshot of memory consumption in bytes.
5326 */
5327 size_t
5328 uma_zone_memory(uma_zone_t zone)
5329 {
5330 size_t sz;
5331 int i;
5332
5333 sz = 0;
5334 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
5335 for (i = 0; i < vm_ndomains; i++)
5336 sz += ZDOM_GET(zone, i)->uzd_nitems;
5337 return (sz * zone->uz_size);
5338 }
5339 for (i = 0; i < vm_ndomains; i++)
5340 sz += zone->uz_keg->uk_domain[i].ud_pages;
5341
5342 return (sz * PAGE_SIZE);
5343 }
5344
5345 struct uma_reclaim_args {
5346 int domain;
5347 int req;
5348 };
5349
5350 static void
5351 uma_reclaim_domain_cb(uma_zone_t zone, void *arg)
5352 {
5353 struct uma_reclaim_args *args;
5354
5355 args = arg;
5356 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) != 0)
5357 return;
5358 if ((args->req == UMA_RECLAIM_TRIM) &&
5359 (zone->uz_flags & UMA_ZONE_NOTRIM) !=0)
5360 return;
5361
5362 uma_zone_reclaim_domain(zone, args->req, args->domain);
5363 }
5364
5365 /* See uma.h */
5366 void
5367 uma_reclaim(int req)
5368 {
5369 uma_reclaim_domain(req, UMA_ANYDOMAIN);
5370 }
5371
5372 void
5373 uma_reclaim_domain(int req, int domain)
5374 {
5375 struct uma_reclaim_args args;
5376
5377 bucket_enable();
5378
5379 args.domain = domain;
5380 args.req = req;
5381
5382 sx_slock(&uma_reclaim_lock);
5383 switch (req) {
5384 case UMA_RECLAIM_TRIM:
5385 case UMA_RECLAIM_DRAIN:
5386 zone_foreach(uma_reclaim_domain_cb, &args);
5387 break;
5388 case UMA_RECLAIM_DRAIN_CPU:
5389 /*
5390 * Reclaim globally visible free items from all zones, then drain
5391 * per-CPU buckets, then reclaim items freed while draining.
5392 * This approach minimizes expensive context switching needed to
5393 * drain each zone's per-CPU buckets.
5394 */
5395 args.req = UMA_RECLAIM_DRAIN;
5396 zone_foreach(uma_reclaim_domain_cb, &args);
5397 pcpu_cache_drain_safe(NULL);
5398 zone_foreach(uma_reclaim_domain_cb, &args);
5399 break;
5400 default:
5401 panic("unhandled reclamation request %d", req);
5402 }
5403
5404 /*
5405 * Some slabs may have been freed but this zone will be visited early
5406 * we visit again so that we can free pages that are empty once other
5407 * zones are drained. We have to do the same for buckets.
5408 */
5409 uma_zone_reclaim_domain(slabzones[0], UMA_RECLAIM_DRAIN, domain);
5410 uma_zone_reclaim_domain(slabzones[1], UMA_RECLAIM_DRAIN, domain);
5411 bucket_zone_drain(domain);
5412 sx_sunlock(&uma_reclaim_lock);
5413 }
5414
5415 static volatile int uma_reclaim_needed;
5416
5417 void
5418 uma_reclaim_wakeup(void)
5419 {
5420
5421 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
5422 wakeup(uma_reclaim);
5423 }
5424
5425 void
5426 uma_reclaim_worker(void *arg __unused)
5427 {
5428
5429 for (;;) {
5430 sx_xlock(&uma_reclaim_lock);
5431 while (atomic_load_int(&uma_reclaim_needed) == 0)
5432 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
5433 hz);
5434 sx_xunlock(&uma_reclaim_lock);
5435 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
5436 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
5437 atomic_store_int(&uma_reclaim_needed, 0);
5438 /* Don't fire more than once per-second. */
5439 pause("umarclslp", hz);
5440 }
5441 }
5442
5443 /* See uma.h */
5444 void
5445 uma_zone_reclaim(uma_zone_t zone, int req)
5446 {
5447 uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN);
5448 }
5449
5450 void
5451 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain)
5452 {
5453 switch (req) {
5454 case UMA_RECLAIM_TRIM:
5455 zone_reclaim(zone, domain, M_NOWAIT, false);
5456 break;
5457 case UMA_RECLAIM_DRAIN:
5458 zone_reclaim(zone, domain, M_NOWAIT, true);
5459 break;
5460 case UMA_RECLAIM_DRAIN_CPU:
5461 pcpu_cache_drain_safe(zone);
5462 zone_reclaim(zone, domain, M_NOWAIT, true);
5463 break;
5464 default:
5465 panic("unhandled reclamation request %d", req);
5466 }
5467 }
5468
5469 /* See uma.h */
5470 int
5471 uma_zone_exhausted(uma_zone_t zone)
5472 {
5473
5474 return (atomic_load_32(&zone->uz_sleepers) > 0);
5475 }
5476
5477 unsigned long
5478 uma_limit(void)
5479 {
5480
5481 return (uma_kmem_limit);
5482 }
5483
5484 void
5485 uma_set_limit(unsigned long limit)
5486 {
5487
5488 uma_kmem_limit = limit;
5489 }
5490
5491 unsigned long
5492 uma_size(void)
5493 {
5494
5495 return (atomic_load_long(&uma_kmem_total));
5496 }
5497
5498 long
5499 uma_avail(void)
5500 {
5501
5502 return (uma_kmem_limit - uma_size());
5503 }
5504
5505 #ifdef DDB
5506 /*
5507 * Generate statistics across both the zone and its per-cpu cache's. Return
5508 * desired statistics if the pointer is non-NULL for that statistic.
5509 *
5510 * Note: does not update the zone statistics, as it can't safely clear the
5511 * per-CPU cache statistic.
5512 *
5513 */
5514 static void
5515 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5516 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5517 {
5518 uma_cache_t cache;
5519 uint64_t allocs, frees, sleeps, xdomain;
5520 int cachefree, cpu;
5521
5522 allocs = frees = sleeps = xdomain = 0;
5523 cachefree = 0;
5524 CPU_FOREACH(cpu) {
5525 cache = &z->uz_cpu[cpu];
5526 cachefree += cache->uc_allocbucket.ucb_cnt;
5527 cachefree += cache->uc_freebucket.ucb_cnt;
5528 xdomain += cache->uc_crossbucket.ucb_cnt;
5529 cachefree += cache->uc_crossbucket.ucb_cnt;
5530 allocs += cache->uc_allocs;
5531 frees += cache->uc_frees;
5532 }
5533 allocs += counter_u64_fetch(z->uz_allocs);
5534 frees += counter_u64_fetch(z->uz_frees);
5535 xdomain += counter_u64_fetch(z->uz_xdomain);
5536 sleeps += z->uz_sleeps;
5537 if (cachefreep != NULL)
5538 *cachefreep = cachefree;
5539 if (allocsp != NULL)
5540 *allocsp = allocs;
5541 if (freesp != NULL)
5542 *freesp = frees;
5543 if (sleepsp != NULL)
5544 *sleepsp = sleeps;
5545 if (xdomainp != NULL)
5546 *xdomainp = xdomain;
5547 }
5548 #endif /* DDB */
5549
5550 static int
5551 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5552 {
5553 uma_keg_t kz;
5554 uma_zone_t z;
5555 int count;
5556
5557 count = 0;
5558 rw_rlock(&uma_rwlock);
5559 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5560 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5561 count++;
5562 }
5563 LIST_FOREACH(z, &uma_cachezones, uz_link)
5564 count++;
5565
5566 rw_runlock(&uma_rwlock);
5567 return (sysctl_handle_int(oidp, &count, 0, req));
5568 }
5569
5570 static void
5571 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5572 struct uma_percpu_stat *ups, bool internal)
5573 {
5574 uma_zone_domain_t zdom;
5575 uma_cache_t cache;
5576 int i;
5577
5578 for (i = 0; i < vm_ndomains; i++) {
5579 zdom = ZDOM_GET(z, i);
5580 uth->uth_zone_free += zdom->uzd_nitems;
5581 }
5582 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5583 uth->uth_frees = counter_u64_fetch(z->uz_frees);
5584 uth->uth_fails = counter_u64_fetch(z->uz_fails);
5585 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5586 uth->uth_sleeps = z->uz_sleeps;
5587
5588 for (i = 0; i < mp_maxid + 1; i++) {
5589 bzero(&ups[i], sizeof(*ups));
5590 if (internal || CPU_ABSENT(i))
5591 continue;
5592 cache = &z->uz_cpu[i];
5593 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5594 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5595 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5596 ups[i].ups_allocs = cache->uc_allocs;
5597 ups[i].ups_frees = cache->uc_frees;
5598 }
5599 }
5600
5601 static int
5602 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5603 {
5604 struct uma_stream_header ush;
5605 struct uma_type_header uth;
5606 struct uma_percpu_stat *ups;
5607 struct sbuf sbuf;
5608 uma_keg_t kz;
5609 uma_zone_t z;
5610 uint64_t items;
5611 uint32_t kfree, pages;
5612 int count, error, i;
5613
5614 error = sysctl_wire_old_buffer(req, 0);
5615 if (error != 0)
5616 return (error);
5617 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5618 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5619 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5620
5621 count = 0;
5622 rw_rlock(&uma_rwlock);
5623 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5624 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5625 count++;
5626 }
5627
5628 LIST_FOREACH(z, &uma_cachezones, uz_link)
5629 count++;
5630
5631 /*
5632 * Insert stream header.
5633 */
5634 bzero(&ush, sizeof(ush));
5635 ush.ush_version = UMA_STREAM_VERSION;
5636 ush.ush_maxcpus = (mp_maxid + 1);
5637 ush.ush_count = count;
5638 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5639
5640 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5641 kfree = pages = 0;
5642 for (i = 0; i < vm_ndomains; i++) {
5643 kfree += kz->uk_domain[i].ud_free_items;
5644 pages += kz->uk_domain[i].ud_pages;
5645 }
5646 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5647 bzero(&uth, sizeof(uth));
5648 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5649 uth.uth_align = kz->uk_align;
5650 uth.uth_size = kz->uk_size;
5651 uth.uth_rsize = kz->uk_rsize;
5652 if (z->uz_max_items > 0) {
5653 items = UZ_ITEMS_COUNT(z->uz_items);
5654 uth.uth_pages = (items / kz->uk_ipers) *
5655 kz->uk_ppera;
5656 } else
5657 uth.uth_pages = pages;
5658 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5659 kz->uk_ppera;
5660 uth.uth_limit = z->uz_max_items;
5661 uth.uth_keg_free = kfree;
5662
5663 /*
5664 * A zone is secondary is it is not the first entry
5665 * on the keg's zone list.
5666 */
5667 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5668 (LIST_FIRST(&kz->uk_zones) != z))
5669 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5670 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5671 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5672 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5673 for (i = 0; i < mp_maxid + 1; i++)
5674 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5675 }
5676 }
5677 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5678 bzero(&uth, sizeof(uth));
5679 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5680 uth.uth_size = z->uz_size;
5681 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5682 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5683 for (i = 0; i < mp_maxid + 1; i++)
5684 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5685 }
5686
5687 rw_runlock(&uma_rwlock);
5688 error = sbuf_finish(&sbuf);
5689 sbuf_delete(&sbuf);
5690 free(ups, M_TEMP);
5691 return (error);
5692 }
5693
5694 int
5695 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5696 {
5697 uma_zone_t zone = *(uma_zone_t *)arg1;
5698 int error, max;
5699
5700 max = uma_zone_get_max(zone);
5701 error = sysctl_handle_int(oidp, &max, 0, req);
5702 if (error || !req->newptr)
5703 return (error);
5704
5705 uma_zone_set_max(zone, max);
5706
5707 return (0);
5708 }
5709
5710 int
5711 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5712 {
5713 uma_zone_t zone;
5714 int cur;
5715
5716 /*
5717 * Some callers want to add sysctls for global zones that
5718 * may not yet exist so they pass a pointer to a pointer.
5719 */
5720 if (arg2 == 0)
5721 zone = *(uma_zone_t *)arg1;
5722 else
5723 zone = arg1;
5724 cur = uma_zone_get_cur(zone);
5725 return (sysctl_handle_int(oidp, &cur, 0, req));
5726 }
5727
5728 static int
5729 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5730 {
5731 uma_zone_t zone = arg1;
5732 uint64_t cur;
5733
5734 cur = uma_zone_get_allocs(zone);
5735 return (sysctl_handle_64(oidp, &cur, 0, req));
5736 }
5737
5738 static int
5739 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5740 {
5741 uma_zone_t zone = arg1;
5742 uint64_t cur;
5743
5744 cur = uma_zone_get_frees(zone);
5745 return (sysctl_handle_64(oidp, &cur, 0, req));
5746 }
5747
5748 static int
5749 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5750 {
5751 struct sbuf sbuf;
5752 uma_zone_t zone = arg1;
5753 int error;
5754
5755 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5756 if (zone->uz_flags != 0)
5757 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5758 else
5759 sbuf_printf(&sbuf, "0");
5760 error = sbuf_finish(&sbuf);
5761 sbuf_delete(&sbuf);
5762
5763 return (error);
5764 }
5765
5766 static int
5767 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5768 {
5769 uma_keg_t keg = arg1;
5770 int avail, effpct, total;
5771
5772 total = keg->uk_ppera * PAGE_SIZE;
5773 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5774 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5775 /*
5776 * We consider the client's requested size and alignment here, not the
5777 * real size determination uk_rsize, because we also adjust the real
5778 * size for internal implementation reasons (max bitset size).
5779 */
5780 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5781 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5782 avail *= mp_maxid + 1;
5783 effpct = 100 * avail / total;
5784 return (sysctl_handle_int(oidp, &effpct, 0, req));
5785 }
5786
5787 static int
5788 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5789 {
5790 uma_zone_t zone = arg1;
5791 uint64_t cur;
5792
5793 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5794 return (sysctl_handle_64(oidp, &cur, 0, req));
5795 }
5796
5797 #ifdef INVARIANTS
5798 static uma_slab_t
5799 uma_dbg_getslab(uma_zone_t zone, void *item)
5800 {
5801 uma_slab_t slab;
5802 uma_keg_t keg;
5803 uint8_t *mem;
5804
5805 /*
5806 * It is safe to return the slab here even though the
5807 * zone is unlocked because the item's allocation state
5808 * essentially holds a reference.
5809 */
5810 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5811 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5812 return (NULL);
5813 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5814 return (vtoslab((vm_offset_t)mem));
5815 keg = zone->uz_keg;
5816 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5817 return ((uma_slab_t)(mem + keg->uk_pgoff));
5818 KEG_LOCK(keg, 0);
5819 slab = hash_sfind(&keg->uk_hash, mem);
5820 KEG_UNLOCK(keg, 0);
5821
5822 return (slab);
5823 }
5824
5825 static bool
5826 uma_dbg_zskip(uma_zone_t zone, void *mem)
5827 {
5828
5829 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5830 return (true);
5831
5832 return (uma_dbg_kskip(zone->uz_keg, mem));
5833 }
5834
5835 static bool
5836 uma_dbg_kskip(uma_keg_t keg, void *mem)
5837 {
5838 uintptr_t idx;
5839
5840 if (dbg_divisor == 0)
5841 return (true);
5842
5843 if (dbg_divisor == 1)
5844 return (false);
5845
5846 idx = (uintptr_t)mem >> PAGE_SHIFT;
5847 if (keg->uk_ipers > 1) {
5848 idx *= keg->uk_ipers;
5849 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5850 }
5851
5852 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5853 counter_u64_add(uma_skip_cnt, 1);
5854 return (true);
5855 }
5856 counter_u64_add(uma_dbg_cnt, 1);
5857
5858 return (false);
5859 }
5860
5861 /*
5862 * Set up the slab's freei data such that uma_dbg_free can function.
5863 *
5864 */
5865 static void
5866 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5867 {
5868 uma_keg_t keg;
5869 int freei;
5870
5871 if (slab == NULL) {
5872 slab = uma_dbg_getslab(zone, item);
5873 if (slab == NULL)
5874 panic("uma: item %p did not belong to zone %s",
5875 item, zone->uz_name);
5876 }
5877 keg = zone->uz_keg;
5878 freei = slab_item_index(slab, keg, item);
5879
5880 if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5881 slab_dbg_bits(slab, keg)))
5882 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5883 item, zone, zone->uz_name, slab, freei);
5884 }
5885
5886 /*
5887 * Verifies freed addresses. Checks for alignment, valid slab membership
5888 * and duplicate frees.
5889 *
5890 */
5891 static void
5892 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5893 {
5894 uma_keg_t keg;
5895 int freei;
5896
5897 if (slab == NULL) {
5898 slab = uma_dbg_getslab(zone, item);
5899 if (slab == NULL)
5900 panic("uma: Freed item %p did not belong to zone %s",
5901 item, zone->uz_name);
5902 }
5903 keg = zone->uz_keg;
5904 freei = slab_item_index(slab, keg, item);
5905
5906 if (freei >= keg->uk_ipers)
5907 panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5908 item, zone, zone->uz_name, slab, freei);
5909
5910 if (slab_item(slab, keg, freei) != item)
5911 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5912 item, zone, zone->uz_name, slab, freei);
5913
5914 if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5915 slab_dbg_bits(slab, keg)))
5916 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5917 item, zone, zone->uz_name, slab, freei);
5918 }
5919 #endif /* INVARIANTS */
5920
5921 #ifdef DDB
5922 static int64_t
5923 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5924 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5925 {
5926 uint64_t frees;
5927 int i;
5928
5929 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5930 *allocs = counter_u64_fetch(z->uz_allocs);
5931 frees = counter_u64_fetch(z->uz_frees);
5932 *sleeps = z->uz_sleeps;
5933 *cachefree = 0;
5934 *xdomain = 0;
5935 } else
5936 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5937 xdomain);
5938 for (i = 0; i < vm_ndomains; i++) {
5939 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5940 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5941 (LIST_FIRST(&kz->uk_zones) != z)))
5942 *cachefree += kz->uk_domain[i].ud_free_items;
5943 }
5944 *used = *allocs - frees;
5945 return (((int64_t)*used + *cachefree) * kz->uk_size);
5946 }
5947
5948 DB_SHOW_COMMAND_FLAGS(uma, db_show_uma, DB_CMD_MEMSAFE)
5949 {
5950 const char *fmt_hdr, *fmt_entry;
5951 uma_keg_t kz;
5952 uma_zone_t z;
5953 uint64_t allocs, used, sleeps, xdomain;
5954 long cachefree;
5955 /* variables for sorting */
5956 uma_keg_t cur_keg;
5957 uma_zone_t cur_zone, last_zone;
5958 int64_t cur_size, last_size, size;
5959 int ties;
5960
5961 /* /i option produces machine-parseable CSV output */
5962 if (modif[0] == 'i') {
5963 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5964 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5965 } else {
5966 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5967 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5968 }
5969
5970 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5971 "Sleeps", "Bucket", "Total Mem", "XFree");
5972
5973 /* Sort the zones with largest size first. */
5974 last_zone = NULL;
5975 last_size = INT64_MAX;
5976 for (;;) {
5977 cur_zone = NULL;
5978 cur_size = -1;
5979 ties = 0;
5980 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5981 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5982 /*
5983 * In the case of size ties, print out zones
5984 * in the order they are encountered. That is,
5985 * when we encounter the most recently output
5986 * zone, we have already printed all preceding
5987 * ties, and we must print all following ties.
5988 */
5989 if (z == last_zone) {
5990 ties = 1;
5991 continue;
5992 }
5993 size = get_uma_stats(kz, z, &allocs, &used,
5994 &sleeps, &cachefree, &xdomain);
5995 if (size > cur_size && size < last_size + ties)
5996 {
5997 cur_size = size;
5998 cur_zone = z;
5999 cur_keg = kz;
6000 }
6001 }
6002 }
6003 if (cur_zone == NULL)
6004 break;
6005
6006 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
6007 &sleeps, &cachefree, &xdomain);
6008 db_printf(fmt_entry, cur_zone->uz_name,
6009 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
6010 (uintmax_t)allocs, (uintmax_t)sleeps,
6011 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
6012 xdomain);
6013
6014 if (db_pager_quit)
6015 return;
6016 last_zone = cur_zone;
6017 last_size = cur_size;
6018 }
6019 }
6020
6021 DB_SHOW_COMMAND_FLAGS(umacache, db_show_umacache, DB_CMD_MEMSAFE)
6022 {
6023 uma_zone_t z;
6024 uint64_t allocs, frees;
6025 long cachefree;
6026 int i;
6027
6028 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
6029 "Requests", "Bucket");
6030 LIST_FOREACH(z, &uma_cachezones, uz_link) {
6031 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
6032 for (i = 0; i < vm_ndomains; i++)
6033 cachefree += ZDOM_GET(z, i)->uzd_nitems;
6034 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
6035 z->uz_name, (uintmax_t)z->uz_size,
6036 (intmax_t)(allocs - frees), cachefree,
6037 (uintmax_t)allocs, z->uz_bucket_size);
6038 if (db_pager_quit)
6039 return;
6040 }
6041 }
6042 #endif /* DDB */
6043