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_object.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_pageout.h>
90 #include <vm/vm_phys.h>
91 #include <vm/vm_pagequeue.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_extern.h>
95 #include <vm/vm_dumpset.h>
96 #include <vm/uma.h>
97 #include <vm/uma_int.h>
98 #include <vm/uma_dbg.h>
99
100 #include <ddb/ddb.h>
101
102 #ifdef DEBUG_MEMGUARD
103 #include <vm/memguard.h>
104 #endif
105
106 #include <machine/md_var.h>
107
108 #ifdef INVARIANTS
109 #define UMA_ALWAYS_CTORDTOR 1
110 #else
111 #define UMA_ALWAYS_CTORDTOR 0
112 #endif
113
114 /*
115 * This is the zone and keg from which all zones are spawned.
116 */
117 static uma_zone_t kegs;
118 static uma_zone_t zones;
119
120 /*
121 * On INVARIANTS builds, the slab contains a second bitset of the same size,
122 * "dbg_bits", which is laid out immediately after us_free.
123 */
124 #ifdef INVARIANTS
125 #define SLAB_BITSETS 2
126 #else
127 #define SLAB_BITSETS 1
128 #endif
129
130 /*
131 * These are the two zones from which all offpage uma_slab_ts are allocated.
132 *
133 * One zone is for slab headers that can represent a larger number of items,
134 * making the slabs themselves more efficient, and the other zone is for
135 * headers that are smaller and represent fewer items, making the headers more
136 * efficient.
137 */
138 #define SLABZONE_SIZE(setsize) \
139 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
140 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16)
141 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE
142 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE)
143 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE)
144 static uma_zone_t slabzones[2];
145
146 /*
147 * The initial hash tables come out of this zone so they can be allocated
148 * prior to malloc coming up.
149 */
150 static uma_zone_t hashzone;
151
152 /* The boot-time adjusted value for cache line alignment. */
153 static unsigned int uma_cache_align_mask = 64 - 1;
154
155 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
156 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
157
158 /*
159 * Are we allowed to allocate buckets?
160 */
161 static int bucketdisable = 1;
162
163 /* Linked list of all kegs in the system */
164 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
165
166 /* Linked list of all cache-only zones in the system */
167 static LIST_HEAD(,uma_zone) uma_cachezones =
168 LIST_HEAD_INITIALIZER(uma_cachezones);
169
170 /*
171 * Mutex for global lists: uma_kegs, uma_cachezones, and the per-keg list of
172 * zones.
173 */
174 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
175
176 static struct sx uma_reclaim_lock;
177
178 /*
179 * First available virual address for boot time allocations.
180 */
181 static vm_offset_t bootstart;
182 static vm_offset_t bootmem;
183
184 /*
185 * kmem soft limit, initialized by uma_set_limit(). Ensure that early
186 * allocations don't trigger a wakeup of the reclaim thread.
187 */
188 unsigned long uma_kmem_limit = LONG_MAX;
189 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
190 "UMA kernel memory soft limit");
191 unsigned long uma_kmem_total;
192 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
193 "UMA kernel memory usage");
194
195 /* Is the VM done starting up? */
196 static enum {
197 BOOT_COLD,
198 BOOT_KVA,
199 BOOT_PCPU,
200 BOOT_RUNNING,
201 BOOT_SHUTDOWN,
202 } booted = BOOT_COLD;
203
204 /*
205 * This is the handle used to schedule events that need to happen
206 * outside of the allocation fast path.
207 */
208 static struct timeout_task uma_timeout_task;
209 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
210
211 /*
212 * This structure is passed as the zone ctor arg so that I don't have to create
213 * a special allocation function just for zones.
214 */
215 struct uma_zctor_args {
216 const char *name;
217 size_t size;
218 uma_ctor ctor;
219 uma_dtor dtor;
220 uma_init uminit;
221 uma_fini fini;
222 uma_import import;
223 uma_release release;
224 void *arg;
225 uma_keg_t keg;
226 int align;
227 uint32_t flags;
228 };
229
230 struct uma_kctor_args {
231 uma_zone_t zone;
232 size_t size;
233 uma_init uminit;
234 uma_fini fini;
235 int align;
236 uint32_t flags;
237 };
238
239 struct uma_bucket_zone {
240 uma_zone_t ubz_zone;
241 const char *ubz_name;
242 int ubz_entries; /* Number of items it can hold. */
243 int ubz_maxsize; /* Maximum allocation size per-item. */
244 };
245
246 /*
247 * Compute the actual number of bucket entries to pack them in power
248 * of two sizes for more efficient space utilization.
249 */
250 #define BUCKET_SIZE(n) \
251 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
252
253 #define BUCKET_MAX BUCKET_SIZE(256)
254
255 struct uma_bucket_zone bucket_zones[] = {
256 /* Literal bucket sizes. */
257 { NULL, "2 Bucket", 2, 4096 },
258 { NULL, "4 Bucket", 4, 3072 },
259 { NULL, "8 Bucket", 8, 2048 },
260 { NULL, "16 Bucket", 16, 1024 },
261 /* Rounded down power of 2 sizes for efficiency. */
262 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
263 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
264 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
265 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
266 { NULL, NULL, 0}
267 };
268
269 /*
270 * Flags and enumerations to be passed to internal functions.
271 */
272 enum zfreeskip {
273 SKIP_NONE = 0,
274 SKIP_CNT = 0x00000001,
275 SKIP_DTOR = 0x00010000,
276 SKIP_FINI = 0x00020000,
277 };
278
279 /* Prototypes.. */
280
281 void uma_startup1(vm_offset_t);
282 void uma_startup2(void);
283
284 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
285 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
286 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
287 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
288 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
289 static void page_free(void *, vm_size_t, uint8_t);
290 static void pcpu_page_free(void *, vm_size_t, uint8_t);
291 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
292 static void cache_drain(uma_zone_t);
293 static void bucket_drain(uma_zone_t, uma_bucket_t);
294 static void bucket_cache_reclaim(uma_zone_t zone, bool, int);
295 static bool bucket_cache_reclaim_domain(uma_zone_t, bool, bool, int);
296 static int keg_ctor(void *, int, void *, int);
297 static void keg_dtor(void *, int, void *);
298 static void keg_drain(uma_keg_t keg, int domain);
299 static int zone_ctor(void *, int, void *, int);
300 static void zone_dtor(void *, int, void *);
301 static inline void item_dtor(uma_zone_t zone, void *item, int size,
302 void *udata, enum zfreeskip skip);
303 static int zero_init(void *, int, int);
304 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
305 int itemdomain, bool ws);
306 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
307 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
308 static void zone_timeout(uma_zone_t zone, void *);
309 static int hash_alloc(struct uma_hash *, u_int);
310 static int hash_expand(struct uma_hash *, struct uma_hash *);
311 static void hash_free(struct uma_hash *hash);
312 static void uma_timeout(void *, int);
313 static void uma_shutdown(void);
314 static void *zone_alloc_item(uma_zone_t, void *, int, int);
315 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
316 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
317 static void zone_free_limit(uma_zone_t zone, int count);
318 static void bucket_enable(void);
319 static void bucket_init(void);
320 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
321 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
322 static void bucket_zone_drain(int domain);
323 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
324 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
325 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
326 static size_t slab_sizeof(int nitems);
327 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
328 uma_fini fini, int align, uint32_t flags);
329 static int zone_import(void *, void **, int, int, int);
330 static void zone_release(void *, void **, int);
331 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
332 static bool cache_free(uma_zone_t, uma_cache_t, void *, int);
333
334 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
335 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
336 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
337 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
338 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
339 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
340 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
341
342 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
343
344 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
345 "Memory allocation debugging");
346
347 #ifdef INVARIANTS
348 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
349 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
350
351 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
352 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
353 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
354 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
355
356 static u_int dbg_divisor = 1;
357 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
358 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
359 "Debug & thrash every this item in memory allocator");
360
361 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
362 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
363 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
364 &uma_dbg_cnt, "memory items debugged");
365 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
366 &uma_skip_cnt, "memory items skipped, not debugged");
367 #endif
368
369 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
370 "Universal Memory Allocator");
371
372 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
373 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
374
375 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
376 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
377
378 static int zone_warnings = 1;
379 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
380 "Warn when UMA zones becomes full");
381
382 static int multipage_slabs = 1;
383 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
384 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
385 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
386 "UMA may choose larger slab sizes for better efficiency");
387
388 /*
389 * Select the slab zone for an offpage slab with the given maximum item count.
390 */
391 static inline uma_zone_t
slabzone(int ipers)392 slabzone(int ipers)
393 {
394
395 return (slabzones[ipers > SLABZONE0_SETSIZE]);
396 }
397
398 /*
399 * This routine checks to see whether or not it's safe to enable buckets.
400 */
401 static void
bucket_enable(void)402 bucket_enable(void)
403 {
404
405 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
406 bucketdisable = vm_page_count_min();
407 }
408
409 /*
410 * Initialize bucket_zones, the array of zones of buckets of various sizes.
411 *
412 * For each zone, calculate the memory required for each bucket, consisting
413 * of the header and an array of pointers.
414 */
415 static void
bucket_init(void)416 bucket_init(void)
417 {
418 struct uma_bucket_zone *ubz;
419 int size;
420
421 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
422 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
423 size += sizeof(void *) * ubz->ubz_entries;
424 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
425 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
426 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
427 UMA_ZONE_FIRSTTOUCH);
428 }
429 }
430
431 /*
432 * Given a desired number of entries for a bucket, return the zone from which
433 * to allocate the bucket.
434 */
435 static struct uma_bucket_zone *
bucket_zone_lookup(int entries)436 bucket_zone_lookup(int entries)
437 {
438 struct uma_bucket_zone *ubz;
439
440 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
441 if (ubz->ubz_entries >= entries)
442 return (ubz);
443 ubz--;
444 return (ubz);
445 }
446
447 static int
bucket_select(int size)448 bucket_select(int size)
449 {
450 struct uma_bucket_zone *ubz;
451
452 ubz = &bucket_zones[0];
453 if (size > ubz->ubz_maxsize)
454 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
455
456 for (; ubz->ubz_entries != 0; ubz++)
457 if (ubz->ubz_maxsize < size)
458 break;
459 ubz--;
460 return (ubz->ubz_entries);
461 }
462
463 static uma_bucket_t
bucket_alloc(uma_zone_t zone,void * udata,int flags)464 bucket_alloc(uma_zone_t zone, void *udata, int flags)
465 {
466 struct uma_bucket_zone *ubz;
467 uma_bucket_t bucket;
468
469 /*
470 * Don't allocate buckets early in boot.
471 */
472 if (__predict_false(booted < BOOT_KVA))
473 return (NULL);
474
475 /*
476 * To limit bucket recursion we store the original zone flags
477 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
478 * NOVM flag to persist even through deep recursions. We also
479 * store ZFLAG_BUCKET once we have recursed attempting to allocate
480 * a bucket for a bucket zone so we do not allow infinite bucket
481 * recursion. This cookie will even persist to frees of unused
482 * buckets via the allocation path or bucket allocations in the
483 * free path.
484 */
485 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
486 udata = (void *)(uintptr_t)zone->uz_flags;
487 else {
488 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
489 return (NULL);
490 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
491 }
492 if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
493 flags |= M_NOVM;
494 ubz = bucket_zone_lookup(atomic_load_16(&zone->uz_bucket_size));
495 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
496 ubz++;
497 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
498 if (bucket) {
499 #ifdef INVARIANTS
500 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
501 #endif
502 bucket->ub_cnt = 0;
503 bucket->ub_entries = min(ubz->ubz_entries,
504 zone->uz_bucket_size_max);
505 bucket->ub_seq = SMR_SEQ_INVALID;
506 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
507 zone->uz_name, zone, bucket);
508 }
509
510 return (bucket);
511 }
512
513 static void
bucket_free(uma_zone_t zone,uma_bucket_t bucket,void * udata)514 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
515 {
516 struct uma_bucket_zone *ubz;
517
518 if (bucket->ub_cnt != 0)
519 bucket_drain(zone, bucket);
520
521 KASSERT(bucket->ub_cnt == 0,
522 ("bucket_free: Freeing a non free bucket."));
523 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
524 ("bucket_free: Freeing an SMR bucket."));
525 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
526 udata = (void *)(uintptr_t)zone->uz_flags;
527 ubz = bucket_zone_lookup(bucket->ub_entries);
528 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
529 }
530
531 static void
bucket_zone_drain(int domain)532 bucket_zone_drain(int domain)
533 {
534 struct uma_bucket_zone *ubz;
535
536 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
537 uma_zone_reclaim_domain(ubz->ubz_zone, UMA_RECLAIM_DRAIN,
538 domain);
539 }
540
541 #ifdef KASAN
542 _Static_assert(UMA_SMALLEST_UNIT % KASAN_SHADOW_SCALE == 0,
543 "Base UMA allocation size not a multiple of the KASAN scale factor");
544
545 static void
kasan_mark_item_valid(uma_zone_t zone,void * item)546 kasan_mark_item_valid(uma_zone_t zone, void *item)
547 {
548 void *pcpu_item;
549 size_t sz, rsz;
550 int i;
551
552 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
553 return;
554
555 sz = zone->uz_size;
556 rsz = roundup2(sz, KASAN_SHADOW_SCALE);
557 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
558 kasan_mark(item, sz, rsz, KASAN_GENERIC_REDZONE);
559 } else {
560 pcpu_item = zpcpu_base_to_offset(item);
561 for (i = 0; i <= mp_maxid; i++)
562 kasan_mark(zpcpu_get_cpu(pcpu_item, i), sz, rsz,
563 KASAN_GENERIC_REDZONE);
564 }
565 }
566
567 static void
kasan_mark_item_invalid(uma_zone_t zone,void * item)568 kasan_mark_item_invalid(uma_zone_t zone, void *item)
569 {
570 void *pcpu_item;
571 size_t sz;
572 int i;
573
574 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
575 return;
576
577 sz = roundup2(zone->uz_size, KASAN_SHADOW_SCALE);
578 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
579 kasan_mark(item, 0, sz, KASAN_UMA_FREED);
580 } else {
581 pcpu_item = zpcpu_base_to_offset(item);
582 for (i = 0; i <= mp_maxid; i++)
583 kasan_mark(zpcpu_get_cpu(pcpu_item, i), 0, sz,
584 KASAN_UMA_FREED);
585 }
586 }
587
588 static void
kasan_mark_slab_valid(uma_keg_t keg,void * mem)589 kasan_mark_slab_valid(uma_keg_t keg, void *mem)
590 {
591 size_t sz;
592
593 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
594 sz = keg->uk_ppera * PAGE_SIZE;
595 kasan_mark(mem, sz, sz, 0);
596 }
597 }
598
599 static void
kasan_mark_slab_invalid(uma_keg_t keg,void * mem)600 kasan_mark_slab_invalid(uma_keg_t keg, void *mem)
601 {
602 size_t sz;
603
604 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
605 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
606 sz = keg->uk_ppera * PAGE_SIZE;
607 else
608 sz = keg->uk_pgoff;
609 kasan_mark(mem, 0, sz, KASAN_UMA_FREED);
610 }
611 }
612 #else /* !KASAN */
613 static void
kasan_mark_item_valid(uma_zone_t zone __unused,void * item __unused)614 kasan_mark_item_valid(uma_zone_t zone __unused, void *item __unused)
615 {
616 }
617
618 static void
kasan_mark_item_invalid(uma_zone_t zone __unused,void * item __unused)619 kasan_mark_item_invalid(uma_zone_t zone __unused, void *item __unused)
620 {
621 }
622
623 static void
kasan_mark_slab_valid(uma_keg_t keg __unused,void * mem __unused)624 kasan_mark_slab_valid(uma_keg_t keg __unused, void *mem __unused)
625 {
626 }
627
628 static void
kasan_mark_slab_invalid(uma_keg_t keg __unused,void * mem __unused)629 kasan_mark_slab_invalid(uma_keg_t keg __unused, void *mem __unused)
630 {
631 }
632 #endif /* KASAN */
633
634 #ifdef KMSAN
635 static inline void
kmsan_mark_item_uninitialized(uma_zone_t zone,void * item)636 kmsan_mark_item_uninitialized(uma_zone_t zone, void *item)
637 {
638 void *pcpu_item;
639 size_t sz;
640 int i;
641
642 if ((zone->uz_flags &
643 (UMA_ZFLAG_CACHE | UMA_ZONE_SECONDARY | UMA_ZONE_MALLOC)) != 0) {
644 /*
645 * Cache zones should not be instrumented by default, as UMA
646 * does not have enough information to do so correctly.
647 * Consumers can mark items themselves if it makes sense to do
648 * so.
649 *
650 * Items from secondary zones are initialized by the parent
651 * zone and thus cannot safely be marked by UMA.
652 *
653 * malloc zones are handled directly by malloc(9) and friends,
654 * since they can provide more precise origin tracking.
655 */
656 return;
657 }
658 if (zone->uz_keg->uk_init != NULL) {
659 /*
660 * By definition, initialized items cannot be marked. The
661 * best we can do is mark items from these zones after they
662 * are freed to the keg.
663 */
664 return;
665 }
666
667 sz = zone->uz_size;
668 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
669 kmsan_orig(item, sz, KMSAN_TYPE_UMA, KMSAN_RET_ADDR);
670 kmsan_mark(item, sz, KMSAN_STATE_UNINIT);
671 } else {
672 pcpu_item = zpcpu_base_to_offset(item);
673 for (i = 0; i <= mp_maxid; i++) {
674 kmsan_orig(zpcpu_get_cpu(pcpu_item, i), sz,
675 KMSAN_TYPE_UMA, KMSAN_RET_ADDR);
676 kmsan_mark(zpcpu_get_cpu(pcpu_item, i), sz,
677 KMSAN_STATE_INITED);
678 }
679 }
680 }
681 #else /* !KMSAN */
682 static inline void
kmsan_mark_item_uninitialized(uma_zone_t zone __unused,void * item __unused)683 kmsan_mark_item_uninitialized(uma_zone_t zone __unused, void *item __unused)
684 {
685 }
686 #endif /* KMSAN */
687
688 /*
689 * Acquire the domain lock and record contention.
690 */
691 static uma_zone_domain_t
zone_domain_lock(uma_zone_t zone,int domain)692 zone_domain_lock(uma_zone_t zone, int domain)
693 {
694 uma_zone_domain_t zdom;
695 bool lockfail;
696
697 zdom = ZDOM_GET(zone, domain);
698 lockfail = false;
699 if (ZDOM_OWNED(zdom))
700 lockfail = true;
701 ZDOM_LOCK(zdom);
702 /* This is unsynchronized. The counter does not need to be precise. */
703 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
704 zone->uz_bucket_size++;
705 return (zdom);
706 }
707
708 /*
709 * Search for the domain with the least cached items and return it if it
710 * is out of balance with the preferred domain.
711 */
712 static __noinline int
zone_domain_lowest(uma_zone_t zone,int pref)713 zone_domain_lowest(uma_zone_t zone, int pref)
714 {
715 long least, nitems, prefitems;
716 int domain;
717 int i;
718
719 prefitems = least = LONG_MAX;
720 domain = 0;
721 for (i = 0; i < vm_ndomains; i++) {
722 nitems = ZDOM_GET(zone, i)->uzd_nitems;
723 if (nitems < least) {
724 domain = i;
725 least = nitems;
726 }
727 if (domain == pref)
728 prefitems = nitems;
729 }
730 if (prefitems < least * 2)
731 return (pref);
732
733 return (domain);
734 }
735
736 /*
737 * Search for the domain with the most cached items and return it or the
738 * preferred domain if it has enough to proceed.
739 */
740 static __noinline int
zone_domain_highest(uma_zone_t zone,int pref)741 zone_domain_highest(uma_zone_t zone, int pref)
742 {
743 long most, nitems;
744 int domain;
745 int i;
746
747 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
748 return (pref);
749
750 most = 0;
751 domain = 0;
752 for (i = 0; i < vm_ndomains; i++) {
753 nitems = ZDOM_GET(zone, i)->uzd_nitems;
754 if (nitems > most) {
755 domain = i;
756 most = nitems;
757 }
758 }
759
760 return (domain);
761 }
762
763 /*
764 * Set the maximum imax value.
765 */
766 static void
zone_domain_imax_set(uma_zone_domain_t zdom,int nitems)767 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
768 {
769 long old;
770
771 old = zdom->uzd_imax;
772 do {
773 if (old >= nitems)
774 return;
775 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
776
777 /*
778 * We are at new maximum, so do the last WSS update for the old
779 * bimin and prepare to measure next allocation batch.
780 */
781 if (zdom->uzd_wss < old - zdom->uzd_bimin)
782 zdom->uzd_wss = old - zdom->uzd_bimin;
783 zdom->uzd_bimin = nitems;
784 }
785
786 /*
787 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
788 * zone's caches. If a bucket is found the zone is not locked on return.
789 */
790 static uma_bucket_t
zone_fetch_bucket(uma_zone_t zone,uma_zone_domain_t zdom,bool reclaim)791 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
792 {
793 uma_bucket_t bucket;
794 long cnt;
795 int i;
796 bool dtor = false;
797
798 ZDOM_LOCK_ASSERT(zdom);
799
800 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
801 return (NULL);
802
803 /* SMR Buckets can not be re-used until readers expire. */
804 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
805 bucket->ub_seq != SMR_SEQ_INVALID) {
806 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
807 return (NULL);
808 bucket->ub_seq = SMR_SEQ_INVALID;
809 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
810 if (STAILQ_NEXT(bucket, ub_link) != NULL)
811 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
812 }
813 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
814
815 KASSERT(zdom->uzd_nitems >= bucket->ub_cnt,
816 ("%s: item count underflow (%ld, %d)",
817 __func__, zdom->uzd_nitems, bucket->ub_cnt));
818 KASSERT(bucket->ub_cnt > 0,
819 ("%s: empty bucket in bucket cache", __func__));
820 zdom->uzd_nitems -= bucket->ub_cnt;
821
822 if (reclaim) {
823 /*
824 * Shift the bounds of the current WSS interval to avoid
825 * perturbing the estimates.
826 */
827 cnt = lmin(zdom->uzd_bimin, bucket->ub_cnt);
828 atomic_subtract_long(&zdom->uzd_imax, cnt);
829 zdom->uzd_bimin -= cnt;
830 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
831 if (zdom->uzd_limin >= bucket->ub_cnt) {
832 zdom->uzd_limin -= bucket->ub_cnt;
833 } else {
834 zdom->uzd_limin = 0;
835 zdom->uzd_timin = 0;
836 }
837 } else if (zdom->uzd_bimin > zdom->uzd_nitems) {
838 zdom->uzd_bimin = zdom->uzd_nitems;
839 if (zdom->uzd_imin > zdom->uzd_nitems)
840 zdom->uzd_imin = zdom->uzd_nitems;
841 }
842
843 ZDOM_UNLOCK(zdom);
844 if (dtor)
845 for (i = 0; i < bucket->ub_cnt; i++)
846 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
847 NULL, SKIP_NONE);
848
849 return (bucket);
850 }
851
852 /*
853 * Insert a full bucket into the specified cache. The "ws" parameter indicates
854 * whether the bucket's contents should be counted as part of the zone's working
855 * set. The bucket may be freed if it exceeds the bucket limit.
856 */
857 static void
zone_put_bucket(uma_zone_t zone,int domain,uma_bucket_t bucket,void * udata,const bool ws)858 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
859 const bool ws)
860 {
861 uma_zone_domain_t zdom;
862
863 /* We don't cache empty buckets. This can happen after a reclaim. */
864 if (bucket->ub_cnt == 0)
865 goto out;
866 zdom = zone_domain_lock(zone, domain);
867
868 /*
869 * Conditionally set the maximum number of items.
870 */
871 zdom->uzd_nitems += bucket->ub_cnt;
872 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
873 if (ws) {
874 zone_domain_imax_set(zdom, zdom->uzd_nitems);
875 } else {
876 /*
877 * Shift the bounds of the current WSS interval to
878 * avoid perturbing the estimates.
879 */
880 atomic_add_long(&zdom->uzd_imax, bucket->ub_cnt);
881 zdom->uzd_imin += bucket->ub_cnt;
882 zdom->uzd_bimin += bucket->ub_cnt;
883 zdom->uzd_limin += bucket->ub_cnt;
884 }
885 if (STAILQ_EMPTY(&zdom->uzd_buckets))
886 zdom->uzd_seq = bucket->ub_seq;
887
888 /*
889 * Try to promote reuse of recently used items. For items
890 * protected by SMR, try to defer reuse to minimize polling.
891 */
892 if (bucket->ub_seq == SMR_SEQ_INVALID)
893 STAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
894 else
895 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
896 ZDOM_UNLOCK(zdom);
897 return;
898 }
899 zdom->uzd_nitems -= bucket->ub_cnt;
900 ZDOM_UNLOCK(zdom);
901 out:
902 bucket_free(zone, bucket, udata);
903 }
904
905 /* Pops an item out of a per-cpu cache bucket. */
906 static inline void *
cache_bucket_pop(uma_cache_t cache,uma_cache_bucket_t bucket)907 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
908 {
909 void *item;
910
911 CRITICAL_ASSERT(curthread);
912
913 bucket->ucb_cnt--;
914 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
915 #ifdef INVARIANTS
916 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
917 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
918 #endif
919 cache->uc_allocs++;
920
921 return (item);
922 }
923
924 /* Pushes an item into a per-cpu cache bucket. */
925 static inline void
cache_bucket_push(uma_cache_t cache,uma_cache_bucket_t bucket,void * item)926 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
927 {
928
929 CRITICAL_ASSERT(curthread);
930 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
931 ("uma_zfree: Freeing to non free bucket index."));
932
933 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
934 bucket->ucb_cnt++;
935 cache->uc_frees++;
936 }
937
938 /*
939 * Unload a UMA bucket from a per-cpu cache.
940 */
941 static inline uma_bucket_t
cache_bucket_unload(uma_cache_bucket_t bucket)942 cache_bucket_unload(uma_cache_bucket_t bucket)
943 {
944 uma_bucket_t b;
945
946 b = bucket->ucb_bucket;
947 if (b != NULL) {
948 MPASS(b->ub_entries == bucket->ucb_entries);
949 b->ub_cnt = bucket->ucb_cnt;
950 bucket->ucb_bucket = NULL;
951 bucket->ucb_entries = bucket->ucb_cnt = 0;
952 }
953
954 return (b);
955 }
956
957 static inline uma_bucket_t
cache_bucket_unload_alloc(uma_cache_t cache)958 cache_bucket_unload_alloc(uma_cache_t cache)
959 {
960
961 return (cache_bucket_unload(&cache->uc_allocbucket));
962 }
963
964 static inline uma_bucket_t
cache_bucket_unload_free(uma_cache_t cache)965 cache_bucket_unload_free(uma_cache_t cache)
966 {
967
968 return (cache_bucket_unload(&cache->uc_freebucket));
969 }
970
971 static inline uma_bucket_t
cache_bucket_unload_cross(uma_cache_t cache)972 cache_bucket_unload_cross(uma_cache_t cache)
973 {
974
975 return (cache_bucket_unload(&cache->uc_crossbucket));
976 }
977
978 /*
979 * Load a bucket into a per-cpu cache bucket.
980 */
981 static inline void
cache_bucket_load(uma_cache_bucket_t bucket,uma_bucket_t b)982 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
983 {
984
985 CRITICAL_ASSERT(curthread);
986 MPASS(bucket->ucb_bucket == NULL);
987 MPASS(b->ub_seq == SMR_SEQ_INVALID);
988
989 bucket->ucb_bucket = b;
990 bucket->ucb_cnt = b->ub_cnt;
991 bucket->ucb_entries = b->ub_entries;
992 }
993
994 static inline void
cache_bucket_load_alloc(uma_cache_t cache,uma_bucket_t b)995 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
996 {
997
998 cache_bucket_load(&cache->uc_allocbucket, b);
999 }
1000
1001 static inline void
cache_bucket_load_free(uma_cache_t cache,uma_bucket_t b)1002 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
1003 {
1004
1005 cache_bucket_load(&cache->uc_freebucket, b);
1006 }
1007
1008 #ifdef NUMA
1009 static inline void
cache_bucket_load_cross(uma_cache_t cache,uma_bucket_t b)1010 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
1011 {
1012
1013 cache_bucket_load(&cache->uc_crossbucket, b);
1014 }
1015 #endif
1016
1017 /*
1018 * Copy and preserve ucb_spare.
1019 */
1020 static inline void
cache_bucket_copy(uma_cache_bucket_t b1,uma_cache_bucket_t b2)1021 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
1022 {
1023
1024 b1->ucb_bucket = b2->ucb_bucket;
1025 b1->ucb_entries = b2->ucb_entries;
1026 b1->ucb_cnt = b2->ucb_cnt;
1027 }
1028
1029 /*
1030 * Swap two cache buckets.
1031 */
1032 static inline void
cache_bucket_swap(uma_cache_bucket_t b1,uma_cache_bucket_t b2)1033 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
1034 {
1035 struct uma_cache_bucket b3;
1036
1037 CRITICAL_ASSERT(curthread);
1038
1039 cache_bucket_copy(&b3, b1);
1040 cache_bucket_copy(b1, b2);
1041 cache_bucket_copy(b2, &b3);
1042 }
1043
1044 /*
1045 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
1046 */
1047 static uma_bucket_t
cache_fetch_bucket(uma_zone_t zone,uma_cache_t cache,int domain)1048 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
1049 {
1050 uma_zone_domain_t zdom;
1051 uma_bucket_t bucket;
1052 smr_seq_t seq;
1053
1054 /*
1055 * Avoid the lock if possible.
1056 */
1057 zdom = ZDOM_GET(zone, domain);
1058 if (zdom->uzd_nitems == 0)
1059 return (NULL);
1060
1061 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
1062 (seq = atomic_load_32(&zdom->uzd_seq)) != SMR_SEQ_INVALID &&
1063 !smr_poll(zone->uz_smr, seq, false))
1064 return (NULL);
1065
1066 /*
1067 * Check the zone's cache of buckets.
1068 */
1069 zdom = zone_domain_lock(zone, domain);
1070 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL)
1071 return (bucket);
1072 ZDOM_UNLOCK(zdom);
1073
1074 return (NULL);
1075 }
1076
1077 static void
zone_log_warning(uma_zone_t zone)1078 zone_log_warning(uma_zone_t zone)
1079 {
1080 static const struct timeval warninterval = { 300, 0 };
1081
1082 if (!zone_warnings || zone->uz_warning == NULL)
1083 return;
1084
1085 if (ratecheck(&zone->uz_ratecheck, &warninterval))
1086 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
1087 }
1088
1089 static inline void
zone_maxaction(uma_zone_t zone)1090 zone_maxaction(uma_zone_t zone)
1091 {
1092
1093 if (zone->uz_maxaction.ta_func != NULL)
1094 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
1095 }
1096
1097 /*
1098 * Routine called by timeout which is used to fire off some time interval
1099 * based calculations. (stats, hash size, etc.)
1100 *
1101 * Arguments:
1102 * arg Unused
1103 *
1104 * Returns:
1105 * Nothing
1106 */
1107 static void
uma_timeout(void * context __unused,int pending __unused)1108 uma_timeout(void *context __unused, int pending __unused)
1109 {
1110 bucket_enable();
1111 zone_foreach(zone_timeout, NULL);
1112
1113 /* Reschedule this event */
1114 taskqueue_enqueue_timeout(taskqueue_thread, &uma_timeout_task,
1115 UMA_TIMEOUT * hz);
1116 }
1117
1118 /*
1119 * Update the working set size estimates for the zone's bucket cache.
1120 * The constants chosen here are somewhat arbitrary.
1121 */
1122 static void
zone_domain_update_wss(uma_zone_domain_t zdom)1123 zone_domain_update_wss(uma_zone_domain_t zdom)
1124 {
1125 long m;
1126
1127 ZDOM_LOCK_ASSERT(zdom);
1128 MPASS(zdom->uzd_imax >= zdom->uzd_nitems);
1129 MPASS(zdom->uzd_nitems >= zdom->uzd_bimin);
1130 MPASS(zdom->uzd_bimin >= zdom->uzd_imin);
1131
1132 /*
1133 * Estimate WSS as modified moving average of biggest allocation
1134 * batches for each period over few minutes (UMA_TIMEOUT of 20s).
1135 */
1136 zdom->uzd_wss = lmax(zdom->uzd_wss * 3 / 4,
1137 zdom->uzd_imax - zdom->uzd_bimin);
1138
1139 /*
1140 * Estimate longtime minimum item count as a combination of recent
1141 * minimum item count, adjusted by WSS for safety, and the modified
1142 * moving average over the last several hours (UMA_TIMEOUT of 20s).
1143 * timin measures time since limin tried to go negative, that means
1144 * we were dangerously close to or got out of cache.
1145 */
1146 m = zdom->uzd_imin - zdom->uzd_wss;
1147 if (m >= 0) {
1148 if (zdom->uzd_limin >= m)
1149 zdom->uzd_limin = m;
1150 else
1151 zdom->uzd_limin = (m + zdom->uzd_limin * 255) / 256;
1152 zdom->uzd_timin++;
1153 } else {
1154 zdom->uzd_limin = 0;
1155 zdom->uzd_timin = 0;
1156 }
1157
1158 /* To reduce period edge effects on WSS keep half of the imax. */
1159 atomic_subtract_long(&zdom->uzd_imax,
1160 (zdom->uzd_imax - zdom->uzd_nitems + 1) / 2);
1161 zdom->uzd_imin = zdom->uzd_bimin = zdom->uzd_nitems;
1162 }
1163
1164 /*
1165 * Routine to perform timeout driven calculations. This expands the
1166 * hashes and does per cpu statistics aggregation.
1167 *
1168 * Returns nothing.
1169 */
1170 static void
zone_timeout(uma_zone_t zone,void * unused)1171 zone_timeout(uma_zone_t zone, void *unused)
1172 {
1173 uma_keg_t keg;
1174 u_int slabs, pages;
1175
1176 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
1177 goto trim;
1178
1179 keg = zone->uz_keg;
1180
1181 /*
1182 * Hash zones are non-numa by definition so the first domain
1183 * is the only one present.
1184 */
1185 KEG_LOCK(keg, 0);
1186 pages = keg->uk_domain[0].ud_pages;
1187
1188 /*
1189 * Expand the keg hash table.
1190 *
1191 * This is done if the number of slabs is larger than the hash size.
1192 * What I'm trying to do here is completely reduce collisions. This
1193 * may be a little aggressive. Should I allow for two collisions max?
1194 */
1195 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1196 struct uma_hash newhash;
1197 struct uma_hash oldhash;
1198 int ret;
1199
1200 /*
1201 * This is so involved because allocating and freeing
1202 * while the keg lock is held will lead to deadlock.
1203 * I have to do everything in stages and check for
1204 * races.
1205 */
1206 KEG_UNLOCK(keg, 0);
1207 ret = hash_alloc(&newhash, 1 << fls(slabs));
1208 KEG_LOCK(keg, 0);
1209 if (ret) {
1210 if (hash_expand(&keg->uk_hash, &newhash)) {
1211 oldhash = keg->uk_hash;
1212 keg->uk_hash = newhash;
1213 } else
1214 oldhash = newhash;
1215
1216 KEG_UNLOCK(keg, 0);
1217 hash_free(&oldhash);
1218 goto trim;
1219 }
1220 }
1221 KEG_UNLOCK(keg, 0);
1222
1223 trim:
1224 /* Trim caches not used for a long time. */
1225 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0) {
1226 for (int i = 0; i < vm_ndomains; i++) {
1227 if (bucket_cache_reclaim_domain(zone, false, false, i) &&
1228 (zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1229 keg_drain(zone->uz_keg, i);
1230 }
1231 }
1232 }
1233
1234 /*
1235 * Allocate and zero fill the next sized hash table from the appropriate
1236 * backing store.
1237 *
1238 * Arguments:
1239 * hash A new hash structure with the old hash size in uh_hashsize
1240 *
1241 * Returns:
1242 * 1 on success and 0 on failure.
1243 */
1244 static int
hash_alloc(struct uma_hash * hash,u_int size)1245 hash_alloc(struct uma_hash *hash, u_int size)
1246 {
1247 size_t alloc;
1248
1249 KASSERT(powerof2(size), ("hash size must be power of 2"));
1250 if (size > UMA_HASH_SIZE_INIT) {
1251 hash->uh_hashsize = size;
1252 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1253 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1254 } else {
1255 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1256 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1257 UMA_ANYDOMAIN, M_WAITOK);
1258 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1259 }
1260 if (hash->uh_slab_hash) {
1261 bzero(hash->uh_slab_hash, alloc);
1262 hash->uh_hashmask = hash->uh_hashsize - 1;
1263 return (1);
1264 }
1265
1266 return (0);
1267 }
1268
1269 /*
1270 * Expands the hash table for HASH zones. This is done from zone_timeout
1271 * to reduce collisions. This must not be done in the regular allocation
1272 * path, otherwise, we can recurse on the vm while allocating pages.
1273 *
1274 * Arguments:
1275 * oldhash The hash you want to expand
1276 * newhash The hash structure for the new table
1277 *
1278 * Returns:
1279 * Nothing
1280 *
1281 * Discussion:
1282 */
1283 static int
hash_expand(struct uma_hash * oldhash,struct uma_hash * newhash)1284 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1285 {
1286 uma_hash_slab_t slab;
1287 u_int hval;
1288 u_int idx;
1289
1290 if (!newhash->uh_slab_hash)
1291 return (0);
1292
1293 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1294 return (0);
1295
1296 /*
1297 * I need to investigate hash algorithms for resizing without a
1298 * full rehash.
1299 */
1300
1301 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1302 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1303 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1304 LIST_REMOVE(slab, uhs_hlink);
1305 hval = UMA_HASH(newhash, slab->uhs_data);
1306 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1307 slab, uhs_hlink);
1308 }
1309
1310 return (1);
1311 }
1312
1313 /*
1314 * Free the hash bucket to the appropriate backing store.
1315 *
1316 * Arguments:
1317 * slab_hash The hash bucket we're freeing
1318 * hashsize The number of entries in that hash bucket
1319 *
1320 * Returns:
1321 * Nothing
1322 */
1323 static void
hash_free(struct uma_hash * hash)1324 hash_free(struct uma_hash *hash)
1325 {
1326 if (hash->uh_slab_hash == NULL)
1327 return;
1328 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1329 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1330 else
1331 free(hash->uh_slab_hash, M_UMAHASH);
1332 }
1333
1334 /*
1335 * Frees all outstanding items in a bucket
1336 *
1337 * Arguments:
1338 * zone The zone to free to, must be unlocked.
1339 * bucket The free/alloc bucket with items.
1340 *
1341 * Returns:
1342 * Nothing
1343 */
1344 static void
bucket_drain(uma_zone_t zone,uma_bucket_t bucket)1345 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1346 {
1347 int i;
1348
1349 if (bucket->ub_cnt == 0)
1350 return;
1351
1352 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1353 bucket->ub_seq != SMR_SEQ_INVALID) {
1354 smr_wait(zone->uz_smr, bucket->ub_seq);
1355 bucket->ub_seq = SMR_SEQ_INVALID;
1356 for (i = 0; i < bucket->ub_cnt; i++)
1357 item_dtor(zone, bucket->ub_bucket[i],
1358 zone->uz_size, NULL, SKIP_NONE);
1359 }
1360 if (zone->uz_fini)
1361 for (i = 0; i < bucket->ub_cnt; i++) {
1362 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
1363 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1364 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
1365 }
1366 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1367 if (zone->uz_max_items > 0)
1368 zone_free_limit(zone, bucket->ub_cnt);
1369 #ifdef INVARIANTS
1370 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1371 #endif
1372 bucket->ub_cnt = 0;
1373 }
1374
1375 /*
1376 * Drains the per cpu caches for a zone.
1377 *
1378 * NOTE: This may only be called while the zone is being torn down, and not
1379 * during normal operation. This is necessary in order that we do not have
1380 * to migrate CPUs to drain the per-CPU caches.
1381 *
1382 * Arguments:
1383 * zone The zone to drain, must be unlocked.
1384 *
1385 * Returns:
1386 * Nothing
1387 */
1388 static void
cache_drain(uma_zone_t zone)1389 cache_drain(uma_zone_t zone)
1390 {
1391 uma_cache_t cache;
1392 uma_bucket_t bucket;
1393 smr_seq_t seq;
1394 int cpu;
1395
1396 /*
1397 * XXX: It is safe to not lock the per-CPU caches, because we're
1398 * tearing down the zone anyway. I.e., there will be no further use
1399 * of the caches at this point.
1400 *
1401 * XXX: It would good to be able to assert that the zone is being
1402 * torn down to prevent improper use of cache_drain().
1403 */
1404 seq = SMR_SEQ_INVALID;
1405 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1406 seq = smr_advance(zone->uz_smr);
1407 CPU_FOREACH(cpu) {
1408 cache = &zone->uz_cpu[cpu];
1409 bucket = cache_bucket_unload_alloc(cache);
1410 if (bucket != NULL)
1411 bucket_free(zone, bucket, NULL);
1412 bucket = cache_bucket_unload_free(cache);
1413 if (bucket != NULL) {
1414 bucket->ub_seq = seq;
1415 bucket_free(zone, bucket, NULL);
1416 }
1417 bucket = cache_bucket_unload_cross(cache);
1418 if (bucket != NULL) {
1419 bucket->ub_seq = seq;
1420 bucket_free(zone, bucket, NULL);
1421 }
1422 }
1423 bucket_cache_reclaim(zone, true, UMA_ANYDOMAIN);
1424 }
1425
1426 static void
cache_shrink(uma_zone_t zone,void * unused)1427 cache_shrink(uma_zone_t zone, void *unused)
1428 {
1429
1430 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1431 return;
1432
1433 ZONE_LOCK(zone);
1434 zone->uz_bucket_size =
1435 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1436 ZONE_UNLOCK(zone);
1437 }
1438
1439 static void
cache_drain_safe_cpu(uma_zone_t zone,void * unused)1440 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1441 {
1442 uma_cache_t cache;
1443 uma_bucket_t b1, b2, b3;
1444 int domain;
1445
1446 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1447 return;
1448
1449 b1 = b2 = b3 = NULL;
1450 critical_enter();
1451 cache = &zone->uz_cpu[curcpu];
1452 domain = PCPU_GET(domain);
1453 b1 = cache_bucket_unload_alloc(cache);
1454
1455 /*
1456 * Don't flush SMR zone buckets. This leaves the zone without a
1457 * bucket and forces every free to synchronize().
1458 */
1459 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1460 b2 = cache_bucket_unload_free(cache);
1461 b3 = cache_bucket_unload_cross(cache);
1462 }
1463 critical_exit();
1464
1465 if (b1 != NULL)
1466 zone_free_bucket(zone, b1, NULL, domain, false);
1467 if (b2 != NULL)
1468 zone_free_bucket(zone, b2, NULL, domain, false);
1469 if (b3 != NULL) {
1470 /* Adjust the domain so it goes to zone_free_cross. */
1471 domain = (domain + 1) % vm_ndomains;
1472 zone_free_bucket(zone, b3, NULL, domain, false);
1473 }
1474 }
1475
1476 /*
1477 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1478 * This is an expensive call because it needs to bind to all CPUs
1479 * one by one and enter a critical section on each of them in order
1480 * to safely access their cache buckets.
1481 * Zone lock must not be held on call this function.
1482 */
1483 static void
pcpu_cache_drain_safe(uma_zone_t zone)1484 pcpu_cache_drain_safe(uma_zone_t zone)
1485 {
1486 int cpu;
1487
1488 /*
1489 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1490 */
1491 if (zone)
1492 cache_shrink(zone, NULL);
1493 else
1494 zone_foreach(cache_shrink, NULL);
1495
1496 CPU_FOREACH(cpu) {
1497 thread_lock(curthread);
1498 sched_bind(curthread, cpu);
1499 thread_unlock(curthread);
1500
1501 if (zone)
1502 cache_drain_safe_cpu(zone, NULL);
1503 else
1504 zone_foreach(cache_drain_safe_cpu, NULL);
1505 }
1506 thread_lock(curthread);
1507 sched_unbind(curthread);
1508 thread_unlock(curthread);
1509 }
1510
1511 /*
1512 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1513 * requested a drain, otherwise the per-domain caches are trimmed to either
1514 * estimated working set size.
1515 */
1516 static bool
bucket_cache_reclaim_domain(uma_zone_t zone,bool drain,bool trim,int domain)1517 bucket_cache_reclaim_domain(uma_zone_t zone, bool drain, bool trim, int domain)
1518 {
1519 uma_zone_domain_t zdom;
1520 uma_bucket_t bucket;
1521 long target;
1522 bool done = false;
1523
1524 /*
1525 * The cross bucket is partially filled and not part of
1526 * the item count. Reclaim it individually here.
1527 */
1528 zdom = ZDOM_GET(zone, domain);
1529 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1530 ZONE_CROSS_LOCK(zone);
1531 bucket = zdom->uzd_cross;
1532 zdom->uzd_cross = NULL;
1533 ZONE_CROSS_UNLOCK(zone);
1534 if (bucket != NULL)
1535 bucket_free(zone, bucket, NULL);
1536 }
1537
1538 /*
1539 * If we were asked to drain the zone, we are done only once
1540 * this bucket cache is empty. If trim, we reclaim items in
1541 * excess of the zone's estimated working set size. Multiple
1542 * consecutive calls will shrink the WSS and so reclaim more.
1543 * If neither drain nor trim, then voluntarily reclaim 1/4
1544 * (to reduce first spike) of items not used for a long time.
1545 */
1546 ZDOM_LOCK(zdom);
1547 zone_domain_update_wss(zdom);
1548 if (drain)
1549 target = 0;
1550 else if (trim)
1551 target = zdom->uzd_wss;
1552 else if (zdom->uzd_timin > 900 / UMA_TIMEOUT)
1553 target = zdom->uzd_nitems - zdom->uzd_limin / 4;
1554 else {
1555 ZDOM_UNLOCK(zdom);
1556 return (done);
1557 }
1558 while ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) != NULL &&
1559 zdom->uzd_nitems >= target + bucket->ub_cnt) {
1560 bucket = zone_fetch_bucket(zone, zdom, true);
1561 if (bucket == NULL)
1562 break;
1563 bucket_free(zone, bucket, NULL);
1564 done = true;
1565 ZDOM_LOCK(zdom);
1566 }
1567 ZDOM_UNLOCK(zdom);
1568 return (done);
1569 }
1570
1571 static void
bucket_cache_reclaim(uma_zone_t zone,bool drain,int domain)1572 bucket_cache_reclaim(uma_zone_t zone, bool drain, int domain)
1573 {
1574 int i;
1575
1576 /*
1577 * Shrink the zone bucket size to ensure that the per-CPU caches
1578 * don't grow too large.
1579 */
1580 if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1581 zone->uz_bucket_size--;
1582
1583 if (domain != UMA_ANYDOMAIN &&
1584 (zone->uz_flags & UMA_ZONE_ROUNDROBIN) == 0) {
1585 bucket_cache_reclaim_domain(zone, drain, true, domain);
1586 } else {
1587 for (i = 0; i < vm_ndomains; i++)
1588 bucket_cache_reclaim_domain(zone, drain, true, i);
1589 }
1590 }
1591
1592 static void
keg_free_slab(uma_keg_t keg,uma_slab_t slab,int start)1593 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1594 {
1595 uint8_t *mem;
1596 size_t size;
1597 int i;
1598 uint8_t flags;
1599
1600 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1601 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1602
1603 mem = slab_data(slab, keg);
1604 size = PAGE_SIZE * keg->uk_ppera;
1605
1606 kasan_mark_slab_valid(keg, mem);
1607 if (keg->uk_fini != NULL) {
1608 for (i = start - 1; i > -1; i--)
1609 #ifdef INVARIANTS
1610 /*
1611 * trash_fini implies that dtor was trash_dtor. trash_fini
1612 * would check that memory hasn't been modified since free,
1613 * which executed trash_dtor.
1614 * That's why we need to run uma_dbg_kskip() check here,
1615 * albeit we don't make skip check for other init/fini
1616 * invocations.
1617 */
1618 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1619 keg->uk_fini != trash_fini)
1620 #endif
1621 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1622 }
1623 flags = slab->us_flags;
1624 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1625 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1626 NULL, SKIP_NONE);
1627 }
1628 keg->uk_freef(mem, size, flags);
1629 uma_total_dec(size);
1630 }
1631
1632 static void
keg_drain_domain(uma_keg_t keg,int domain)1633 keg_drain_domain(uma_keg_t keg, int domain)
1634 {
1635 struct slabhead freeslabs;
1636 uma_domain_t dom;
1637 uma_slab_t slab, tmp;
1638 uint32_t i, stofree, stokeep, partial;
1639
1640 dom = &keg->uk_domain[domain];
1641 LIST_INIT(&freeslabs);
1642
1643 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1644 keg->uk_name, keg, domain, dom->ud_free_items);
1645
1646 KEG_LOCK(keg, domain);
1647
1648 /*
1649 * Are the free items in partially allocated slabs sufficient to meet
1650 * the reserve? If not, compute the number of fully free slabs that must
1651 * be kept.
1652 */
1653 partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers;
1654 if (partial < keg->uk_reserve) {
1655 stokeep = min(dom->ud_free_slabs,
1656 howmany(keg->uk_reserve - partial, keg->uk_ipers));
1657 } else {
1658 stokeep = 0;
1659 }
1660 stofree = dom->ud_free_slabs - stokeep;
1661
1662 /*
1663 * Partition the free slabs into two sets: those that must be kept in
1664 * order to maintain the reserve, and those that may be released back to
1665 * the system. Since one set may be much larger than the other,
1666 * populate the smaller of the two sets and swap them if necessary.
1667 */
1668 for (i = min(stofree, stokeep); i > 0; i--) {
1669 slab = LIST_FIRST(&dom->ud_free_slab);
1670 LIST_REMOVE(slab, us_link);
1671 LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1672 }
1673 if (stofree > stokeep)
1674 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1675
1676 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1677 LIST_FOREACH(slab, &freeslabs, us_link)
1678 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1679 }
1680 dom->ud_free_items -= stofree * keg->uk_ipers;
1681 dom->ud_free_slabs -= stofree;
1682 dom->ud_pages -= stofree * keg->uk_ppera;
1683 KEG_UNLOCK(keg, domain);
1684
1685 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1686 keg_free_slab(keg, slab, keg->uk_ipers);
1687 }
1688
1689 /*
1690 * Frees pages from a keg back to the system. This is done on demand from
1691 * the pageout daemon.
1692 *
1693 * Returns nothing.
1694 */
1695 static void
keg_drain(uma_keg_t keg,int domain)1696 keg_drain(uma_keg_t keg, int domain)
1697 {
1698 int i;
1699
1700 if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0)
1701 return;
1702 if (domain != UMA_ANYDOMAIN) {
1703 keg_drain_domain(keg, domain);
1704 } else {
1705 for (i = 0; i < vm_ndomains; i++)
1706 keg_drain_domain(keg, i);
1707 }
1708 }
1709
1710 static void
zone_reclaim(uma_zone_t zone,int domain,int waitok,bool drain)1711 zone_reclaim(uma_zone_t zone, int domain, int waitok, bool drain)
1712 {
1713 /*
1714 * Count active reclaim operations in order to interlock with
1715 * zone_dtor(), which removes the zone from global lists before
1716 * attempting to reclaim items itself.
1717 *
1718 * The zone may be destroyed while sleeping, so only zone_dtor() should
1719 * specify M_WAITOK.
1720 */
1721 ZONE_LOCK(zone);
1722 if (waitok == M_WAITOK) {
1723 while (zone->uz_reclaimers > 0)
1724 msleep(zone, ZONE_LOCKPTR(zone), PVM, "zonedrain", 1);
1725 }
1726 zone->uz_reclaimers++;
1727 ZONE_UNLOCK(zone);
1728 bucket_cache_reclaim(zone, drain, domain);
1729
1730 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1731 keg_drain(zone->uz_keg, domain);
1732 ZONE_LOCK(zone);
1733 zone->uz_reclaimers--;
1734 if (zone->uz_reclaimers == 0)
1735 wakeup(zone);
1736 ZONE_UNLOCK(zone);
1737 }
1738
1739 /*
1740 * Allocate a new slab for a keg and inserts it into the partial slab list.
1741 * The keg should be unlocked on entry. If the allocation succeeds it will
1742 * be locked on return.
1743 *
1744 * Arguments:
1745 * flags Wait flags for the item initialization routine
1746 * aflags Wait flags for the slab allocation
1747 *
1748 * Returns:
1749 * The slab that was allocated or NULL if there is no memory and the
1750 * caller specified M_NOWAIT.
1751 */
1752 static uma_slab_t
keg_alloc_slab(uma_keg_t keg,uma_zone_t zone,int domain,int flags,int aflags)1753 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1754 int aflags)
1755 {
1756 uma_domain_t dom;
1757 uma_slab_t slab;
1758 unsigned long size;
1759 uint8_t *mem;
1760 uint8_t sflags;
1761 int i;
1762
1763 TSENTER();
1764
1765 KASSERT(domain >= 0 && domain < vm_ndomains,
1766 ("keg_alloc_slab: domain %d out of range", domain));
1767
1768 slab = NULL;
1769 mem = NULL;
1770 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1771 uma_hash_slab_t hslab;
1772 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1773 domain, aflags);
1774 if (hslab == NULL)
1775 goto fail;
1776 slab = &hslab->uhs_slab;
1777 }
1778
1779 /*
1780 * This reproduces the old vm_zone behavior of zero filling pages the
1781 * first time they are added to a zone.
1782 *
1783 * Malloced items are zeroed in uma_zalloc.
1784 */
1785
1786 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1787 aflags |= M_ZERO;
1788 else
1789 aflags &= ~M_ZERO;
1790
1791 if (keg->uk_flags & UMA_ZONE_NODUMP)
1792 aflags |= M_NODUMP;
1793
1794 /* zone is passed for legacy reasons. */
1795 size = keg->uk_ppera * PAGE_SIZE;
1796 mem = keg->uk_allocf(zone, size, domain, &sflags, aflags);
1797 if (mem == NULL) {
1798 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1799 zone_free_item(slabzone(keg->uk_ipers),
1800 slab_tohashslab(slab), NULL, SKIP_NONE);
1801 goto fail;
1802 }
1803 uma_total_inc(size);
1804
1805 /* For HASH zones all pages go to the same uma_domain. */
1806 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1807 domain = 0;
1808
1809 kmsan_mark(mem, size,
1810 (aflags & M_ZERO) != 0 ? KMSAN_STATE_INITED : KMSAN_STATE_UNINIT);
1811
1812 /* Point the slab into the allocated memory */
1813 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1814 slab = (uma_slab_t)(mem + keg->uk_pgoff);
1815 else
1816 slab_tohashslab(slab)->uhs_data = mem;
1817
1818 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1819 for (i = 0; i < keg->uk_ppera; i++)
1820 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1821 zone, slab);
1822
1823 slab->us_freecount = keg->uk_ipers;
1824 slab->us_flags = sflags;
1825 slab->us_domain = domain;
1826
1827 BIT_FILL(keg->uk_ipers, &slab->us_free);
1828 #ifdef INVARIANTS
1829 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1830 #endif
1831
1832 if (keg->uk_init != NULL) {
1833 for (i = 0; i < keg->uk_ipers; i++)
1834 if (keg->uk_init(slab_item(slab, keg, i),
1835 keg->uk_size, flags) != 0)
1836 break;
1837 if (i != keg->uk_ipers) {
1838 keg_free_slab(keg, slab, i);
1839 goto fail;
1840 }
1841 }
1842 kasan_mark_slab_invalid(keg, mem);
1843 KEG_LOCK(keg, domain);
1844
1845 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1846 slab, keg->uk_name, keg);
1847
1848 if (keg->uk_flags & UMA_ZFLAG_HASH)
1849 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1850
1851 /*
1852 * If we got a slab here it's safe to mark it partially used
1853 * and return. We assume that the caller is going to remove
1854 * at least one item.
1855 */
1856 dom = &keg->uk_domain[domain];
1857 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1858 dom->ud_pages += keg->uk_ppera;
1859 dom->ud_free_items += keg->uk_ipers;
1860
1861 TSEXIT();
1862 return (slab);
1863
1864 fail:
1865 return (NULL);
1866 }
1867
1868 /*
1869 * This function is intended to be used early on in place of page_alloc(). It
1870 * performs contiguous physical memory allocations and uses a bump allocator for
1871 * KVA, so is usable before the kernel map is initialized.
1872 */
1873 static void *
startup_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)1874 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1875 int wait)
1876 {
1877 vm_paddr_t pa;
1878 vm_page_t m;
1879 int i, pages;
1880
1881 pages = howmany(bytes, PAGE_SIZE);
1882 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1883
1884 *pflag = UMA_SLAB_BOOT;
1885 m = vm_page_alloc_noobj_contig_domain(domain, malloc2vm_flags(wait) |
1886 VM_ALLOC_WIRED, pages, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
1887 VM_MEMATTR_DEFAULT);
1888 if (m == NULL)
1889 return (NULL);
1890
1891 pa = VM_PAGE_TO_PHYS(m);
1892 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1893 #if defined(__aarch64__) || defined(__amd64__) || \
1894 defined(__riscv) || defined(__powerpc64__)
1895 if ((wait & M_NODUMP) == 0)
1896 dump_add_page(pa);
1897 #endif
1898 }
1899
1900 /* Allocate KVA and indirectly advance bootmem. */
1901 return ((void *)pmap_map(&bootmem, m->phys_addr,
1902 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE));
1903 }
1904
1905 static void
startup_free(void * mem,vm_size_t bytes)1906 startup_free(void *mem, vm_size_t bytes)
1907 {
1908 vm_offset_t va;
1909 vm_page_t m;
1910
1911 va = (vm_offset_t)mem;
1912 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1913
1914 /*
1915 * startup_alloc() returns direct-mapped slabs on some platforms. Avoid
1916 * unmapping ranges of the direct map.
1917 */
1918 if (va >= bootstart && va + bytes <= bootmem)
1919 pmap_remove(kernel_pmap, va, va + bytes);
1920 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1921 #if defined(__aarch64__) || defined(__amd64__) || \
1922 defined(__riscv) || defined(__powerpc64__)
1923 dump_drop_page(VM_PAGE_TO_PHYS(m));
1924 #endif
1925 vm_page_unwire_noq(m);
1926 vm_page_free(m);
1927 }
1928 }
1929
1930 /*
1931 * Allocates a number of pages from the system
1932 *
1933 * Arguments:
1934 * bytes The number of bytes requested
1935 * wait Shall we wait?
1936 *
1937 * Returns:
1938 * A pointer to the alloced memory or possibly
1939 * NULL if M_NOWAIT is set.
1940 */
1941 static void *
page_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)1942 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1943 int wait)
1944 {
1945 void *p; /* Returned page */
1946
1947 *pflag = UMA_SLAB_KERNEL;
1948 p = kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1949
1950 return (p);
1951 }
1952
1953 static void *
pcpu_page_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)1954 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1955 int wait)
1956 {
1957 struct pglist alloctail;
1958 vm_offset_t addr, zkva;
1959 int cpu, flags;
1960 vm_page_t p, p_next;
1961 #ifdef NUMA
1962 struct pcpu *pc;
1963 #endif
1964
1965 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1966
1967 TAILQ_INIT(&alloctail);
1968 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | malloc2vm_flags(wait);
1969 *pflag = UMA_SLAB_KERNEL;
1970 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1971 if (CPU_ABSENT(cpu)) {
1972 p = vm_page_alloc_noobj(flags);
1973 } else {
1974 #ifndef NUMA
1975 p = vm_page_alloc_noobj(flags);
1976 #else
1977 pc = pcpu_find(cpu);
1978 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1979 p = NULL;
1980 else
1981 p = vm_page_alloc_noobj_domain(pc->pc_domain,
1982 flags);
1983 if (__predict_false(p == NULL))
1984 p = vm_page_alloc_noobj(flags);
1985 #endif
1986 }
1987 if (__predict_false(p == NULL))
1988 goto fail;
1989 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1990 }
1991 if ((addr = kva_alloc(bytes)) == 0)
1992 goto fail;
1993 zkva = addr;
1994 TAILQ_FOREACH(p, &alloctail, listq) {
1995 pmap_qenter(zkva, &p, 1);
1996 zkva += PAGE_SIZE;
1997 }
1998 return ((void*)addr);
1999 fail:
2000 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2001 vm_page_unwire_noq(p);
2002 vm_page_free(p);
2003 }
2004 return (NULL);
2005 }
2006
2007 /*
2008 * Allocates a number of pages not belonging to a VM object
2009 *
2010 * Arguments:
2011 * bytes The number of bytes requested
2012 * wait Shall we wait?
2013 *
2014 * Returns:
2015 * A pointer to the alloced memory or possibly
2016 * NULL if M_NOWAIT is set.
2017 */
2018 static void *
noobj_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * flags,int wait)2019 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
2020 int wait)
2021 {
2022 TAILQ_HEAD(, vm_page) alloctail;
2023 u_long npages;
2024 vm_offset_t retkva, zkva;
2025 vm_page_t p, p_next;
2026 uma_keg_t keg;
2027 int req;
2028
2029 TAILQ_INIT(&alloctail);
2030 keg = zone->uz_keg;
2031 req = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
2032 if ((wait & M_WAITOK) != 0)
2033 req |= VM_ALLOC_WAITOK;
2034
2035 npages = howmany(bytes, PAGE_SIZE);
2036 while (npages > 0) {
2037 p = vm_page_alloc_noobj_domain(domain, req);
2038 if (p != NULL) {
2039 /*
2040 * Since the page does not belong to an object, its
2041 * listq is unused.
2042 */
2043 TAILQ_INSERT_TAIL(&alloctail, p, listq);
2044 npages--;
2045 continue;
2046 }
2047 /*
2048 * Page allocation failed, free intermediate pages and
2049 * exit.
2050 */
2051 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
2052 vm_page_unwire_noq(p);
2053 vm_page_free(p);
2054 }
2055 return (NULL);
2056 }
2057 *flags = UMA_SLAB_PRIV;
2058 zkva = keg->uk_kva +
2059 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
2060 retkva = zkva;
2061 TAILQ_FOREACH(p, &alloctail, listq) {
2062 pmap_qenter(zkva, &p, 1);
2063 zkva += PAGE_SIZE;
2064 }
2065
2066 return ((void *)retkva);
2067 }
2068
2069 /*
2070 * Allocate physically contiguous pages.
2071 */
2072 static void *
contig_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)2073 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
2074 int wait)
2075 {
2076
2077 *pflag = UMA_SLAB_KERNEL;
2078 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
2079 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
2080 }
2081
2082 /*
2083 * Frees a number of pages to the system
2084 *
2085 * Arguments:
2086 * mem A pointer to the memory to be freed
2087 * size The size of the memory being freed
2088 * flags The original p->us_flags field
2089 *
2090 * Returns:
2091 * Nothing
2092 */
2093 static void
page_free(void * mem,vm_size_t size,uint8_t flags)2094 page_free(void *mem, vm_size_t size, uint8_t flags)
2095 {
2096
2097 if ((flags & UMA_SLAB_BOOT) != 0) {
2098 startup_free(mem, size);
2099 return;
2100 }
2101
2102 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
2103 ("UMA: page_free used with invalid flags %x", flags));
2104
2105 kmem_free(mem, size);
2106 }
2107
2108 /*
2109 * Frees pcpu zone allocations
2110 *
2111 * Arguments:
2112 * mem A pointer to the memory to be freed
2113 * size The size of the memory being freed
2114 * flags The original p->us_flags field
2115 *
2116 * Returns:
2117 * Nothing
2118 */
2119 static void
pcpu_page_free(void * mem,vm_size_t size,uint8_t flags)2120 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
2121 {
2122 vm_offset_t sva, curva;
2123 vm_paddr_t paddr;
2124 vm_page_t m;
2125
2126 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
2127
2128 if ((flags & UMA_SLAB_BOOT) != 0) {
2129 startup_free(mem, size);
2130 return;
2131 }
2132
2133 sva = (vm_offset_t)mem;
2134 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
2135 paddr = pmap_kextract(curva);
2136 m = PHYS_TO_VM_PAGE(paddr);
2137 vm_page_unwire_noq(m);
2138 vm_page_free(m);
2139 }
2140 pmap_qremove(sva, size >> PAGE_SHIFT);
2141 kva_free(sva, size);
2142 }
2143
2144 /*
2145 * Zero fill initializer
2146 *
2147 * Arguments/Returns follow uma_init specifications
2148 */
2149 static int
zero_init(void * mem,int size,int flags)2150 zero_init(void *mem, int size, int flags)
2151 {
2152 bzero(mem, size);
2153 return (0);
2154 }
2155
2156 #ifdef INVARIANTS
2157 static struct noslabbits *
slab_dbg_bits(uma_slab_t slab,uma_keg_t keg)2158 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
2159 {
2160
2161 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
2162 }
2163 #endif
2164
2165 /*
2166 * Actual size of embedded struct slab (!OFFPAGE).
2167 */
2168 static size_t
slab_sizeof(int nitems)2169 slab_sizeof(int nitems)
2170 {
2171 size_t s;
2172
2173 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
2174 return (roundup(s, UMA_ALIGN_PTR + 1));
2175 }
2176
2177 #define UMA_FIXPT_SHIFT 31
2178 #define UMA_FRAC_FIXPT(n, d) \
2179 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
2180 #define UMA_FIXPT_PCT(f) \
2181 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
2182 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
2183 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
2184
2185 /*
2186 * Compute the number of items that will fit in a slab. If hdr is true, the
2187 * item count may be limited to provide space in the slab for an inline slab
2188 * header. Otherwise, all slab space will be provided for item storage.
2189 */
2190 static u_int
slab_ipers_hdr(u_int size,u_int rsize,u_int slabsize,bool hdr)2191 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
2192 {
2193 u_int ipers;
2194 u_int padpi;
2195
2196 /* The padding between items is not needed after the last item. */
2197 padpi = rsize - size;
2198
2199 if (hdr) {
2200 /*
2201 * Start with the maximum item count and remove items until
2202 * the slab header first alongside the allocatable memory.
2203 */
2204 for (ipers = MIN(SLAB_MAX_SETSIZE,
2205 (slabsize + padpi - slab_sizeof(1)) / rsize);
2206 ipers > 0 &&
2207 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2208 ipers--)
2209 continue;
2210 } else {
2211 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2212 }
2213
2214 return (ipers);
2215 }
2216
2217 struct keg_layout_result {
2218 u_int format;
2219 u_int slabsize;
2220 u_int ipers;
2221 u_int eff;
2222 };
2223
2224 static void
keg_layout_one(uma_keg_t keg,u_int rsize,u_int slabsize,u_int fmt,struct keg_layout_result * kl)2225 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2226 struct keg_layout_result *kl)
2227 {
2228 u_int total;
2229
2230 kl->format = fmt;
2231 kl->slabsize = slabsize;
2232
2233 /* Handle INTERNAL as inline with an extra page. */
2234 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2235 kl->format &= ~UMA_ZFLAG_INTERNAL;
2236 kl->slabsize += PAGE_SIZE;
2237 }
2238
2239 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2240 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2241
2242 /* Account for memory used by an offpage slab header. */
2243 total = kl->slabsize;
2244 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2245 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2246
2247 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2248 }
2249
2250 /*
2251 * Determine the format of a uma keg. This determines where the slab header
2252 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2253 *
2254 * Arguments
2255 * keg The zone we should initialize
2256 *
2257 * Returns
2258 * Nothing
2259 */
2260 static void
keg_layout(uma_keg_t keg)2261 keg_layout(uma_keg_t keg)
2262 {
2263 struct keg_layout_result kl = {}, kl_tmp;
2264 u_int fmts[2];
2265 u_int alignsize;
2266 u_int nfmt;
2267 u_int pages;
2268 u_int rsize;
2269 u_int slabsize;
2270 u_int i, j;
2271
2272 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2273 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2274 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2275 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2276 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2277 PRINT_UMA_ZFLAGS));
2278 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2279 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2280 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2281 PRINT_UMA_ZFLAGS));
2282
2283 alignsize = keg->uk_align + 1;
2284 #ifdef KASAN
2285 /*
2286 * ASAN requires that each allocation be aligned to the shadow map
2287 * scale factor.
2288 */
2289 if (alignsize < KASAN_SHADOW_SCALE)
2290 alignsize = KASAN_SHADOW_SCALE;
2291 #endif
2292
2293 /*
2294 * Calculate the size of each allocation (rsize) according to
2295 * alignment. If the requested size is smaller than we have
2296 * allocation bits for we round it up.
2297 */
2298 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2299 rsize = roundup2(rsize, alignsize);
2300
2301 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2302 /*
2303 * We want one item to start on every align boundary in a page.
2304 * To do this we will span pages. We will also extend the item
2305 * by the size of align if it is an even multiple of align.
2306 * Otherwise, it would fall on the same boundary every time.
2307 */
2308 if ((rsize & alignsize) == 0)
2309 rsize += alignsize;
2310 slabsize = rsize * (PAGE_SIZE / alignsize);
2311 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2312 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2313 slabsize = round_page(slabsize);
2314 } else {
2315 /*
2316 * Start with a slab size of as many pages as it takes to
2317 * represent a single item. We will try to fit as many
2318 * additional items into the slab as possible.
2319 */
2320 slabsize = round_page(keg->uk_size);
2321 }
2322
2323 /* Build a list of all of the available formats for this keg. */
2324 nfmt = 0;
2325
2326 /* Evaluate an inline slab layout. */
2327 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2328 fmts[nfmt++] = 0;
2329
2330 /* TODO: vm_page-embedded slab. */
2331
2332 /*
2333 * We can't do OFFPAGE if we're internal or if we've been
2334 * asked to not go to the VM for buckets. If we do this we
2335 * may end up going to the VM for slabs which we do not want
2336 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2337 * In those cases, evaluate a pseudo-format called INTERNAL
2338 * which has an inline slab header and one extra page to
2339 * guarantee that it fits.
2340 *
2341 * Otherwise, see if using an OFFPAGE slab will improve our
2342 * efficiency.
2343 */
2344 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2345 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2346 else
2347 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2348
2349 /*
2350 * Choose a slab size and format which satisfy the minimum efficiency.
2351 * Prefer the smallest slab size that meets the constraints.
2352 *
2353 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2354 * for small items (up to PAGE_SIZE), the iteration increment is one
2355 * page; and for large items, the increment is one item.
2356 */
2357 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2358 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2359 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2360 rsize, i));
2361 for ( ; ; i++) {
2362 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2363 round_page(rsize * (i - 1) + keg->uk_size);
2364
2365 for (j = 0; j < nfmt; j++) {
2366 /* Only if we have no viable format yet. */
2367 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2368 kl.ipers > 0)
2369 continue;
2370
2371 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2372 if (kl_tmp.eff <= kl.eff)
2373 continue;
2374
2375 kl = kl_tmp;
2376
2377 CTR6(KTR_UMA, "keg %s layout: format %#x "
2378 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2379 keg->uk_name, kl.format, kl.ipers, rsize,
2380 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2381
2382 /* Stop when we reach the minimum efficiency. */
2383 if (kl.eff >= UMA_MIN_EFF)
2384 break;
2385 }
2386
2387 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2388 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2389 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2390 break;
2391 }
2392
2393 pages = atop(kl.slabsize);
2394 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2395 pages *= mp_maxid + 1;
2396
2397 keg->uk_rsize = rsize;
2398 keg->uk_ipers = kl.ipers;
2399 keg->uk_ppera = pages;
2400 keg->uk_flags |= kl.format;
2401
2402 /*
2403 * How do we find the slab header if it is offpage or if not all item
2404 * start addresses are in the same page? We could solve the latter
2405 * case with vaddr alignment, but we don't.
2406 */
2407 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2408 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2409 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2410 keg->uk_flags |= UMA_ZFLAG_HASH;
2411 else
2412 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2413 }
2414
2415 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2416 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2417 pages);
2418 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2419 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2420 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2421 keg->uk_ipers, pages));
2422 }
2423
2424 /*
2425 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2426 * the keg onto the global keg list.
2427 *
2428 * Arguments/Returns follow uma_ctor specifications
2429 * udata Actually uma_kctor_args
2430 */
2431 static int
keg_ctor(void * mem,int size,void * udata,int flags)2432 keg_ctor(void *mem, int size, void *udata, int flags)
2433 {
2434 struct uma_kctor_args *arg = udata;
2435 uma_keg_t keg = mem;
2436 uma_zone_t zone;
2437 int i;
2438
2439 bzero(keg, size);
2440 keg->uk_size = arg->size;
2441 keg->uk_init = arg->uminit;
2442 keg->uk_fini = arg->fini;
2443 keg->uk_align = arg->align;
2444 keg->uk_reserve = 0;
2445 keg->uk_flags = arg->flags;
2446
2447 /*
2448 * We use a global round-robin policy by default. Zones with
2449 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2450 * case the iterator is never run.
2451 */
2452 keg->uk_dr.dr_policy = DOMAINSET_RR();
2453 keg->uk_dr.dr_iter = 0;
2454
2455 /*
2456 * The primary zone is passed to us at keg-creation time.
2457 */
2458 zone = arg->zone;
2459 keg->uk_name = zone->uz_name;
2460
2461 if (arg->flags & UMA_ZONE_ZINIT)
2462 keg->uk_init = zero_init;
2463
2464 if (arg->flags & UMA_ZONE_MALLOC)
2465 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2466
2467 #ifndef SMP
2468 keg->uk_flags &= ~UMA_ZONE_PCPU;
2469 #endif
2470
2471 keg_layout(keg);
2472
2473 /*
2474 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2475 * work on. Use round-robin for everything else.
2476 *
2477 * Zones may override the default by specifying either.
2478 */
2479 #ifdef NUMA
2480 if ((keg->uk_flags &
2481 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2482 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2483 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2484 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2485 #endif
2486
2487 /*
2488 * If we haven't booted yet we need allocations to go through the
2489 * startup cache until the vm is ready.
2490 */
2491 #ifdef UMA_MD_SMALL_ALLOC
2492 if (keg->uk_ppera == 1)
2493 keg->uk_allocf = uma_small_alloc;
2494 else
2495 #endif
2496 if (booted < BOOT_KVA)
2497 keg->uk_allocf = startup_alloc;
2498 else if (keg->uk_flags & UMA_ZONE_PCPU)
2499 keg->uk_allocf = pcpu_page_alloc;
2500 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2501 keg->uk_allocf = contig_alloc;
2502 else
2503 keg->uk_allocf = page_alloc;
2504 #ifdef UMA_MD_SMALL_ALLOC
2505 if (keg->uk_ppera == 1)
2506 keg->uk_freef = uma_small_free;
2507 else
2508 #endif
2509 if (keg->uk_flags & UMA_ZONE_PCPU)
2510 keg->uk_freef = pcpu_page_free;
2511 else
2512 keg->uk_freef = page_free;
2513
2514 /*
2515 * Initialize keg's locks.
2516 */
2517 for (i = 0; i < vm_ndomains; i++)
2518 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2519
2520 /*
2521 * If we're putting the slab header in the actual page we need to
2522 * figure out where in each page it goes. See slab_sizeof
2523 * definition.
2524 */
2525 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2526 size_t shsize;
2527
2528 shsize = slab_sizeof(keg->uk_ipers);
2529 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2530 /*
2531 * The only way the following is possible is if with our
2532 * UMA_ALIGN_PTR adjustments we are now bigger than
2533 * UMA_SLAB_SIZE. I haven't checked whether this is
2534 * mathematically possible for all cases, so we make
2535 * sure here anyway.
2536 */
2537 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2538 ("zone %s ipers %d rsize %d size %d slab won't fit",
2539 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2540 }
2541
2542 if (keg->uk_flags & UMA_ZFLAG_HASH)
2543 hash_alloc(&keg->uk_hash, 0);
2544
2545 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2546
2547 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2548
2549 rw_wlock(&uma_rwlock);
2550 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2551 rw_wunlock(&uma_rwlock);
2552 return (0);
2553 }
2554
2555 static void
zone_kva_available(uma_zone_t zone,void * unused)2556 zone_kva_available(uma_zone_t zone, void *unused)
2557 {
2558 uma_keg_t keg;
2559
2560 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2561 return;
2562 KEG_GET(zone, keg);
2563
2564 if (keg->uk_allocf == startup_alloc) {
2565 /* Switch to the real allocator. */
2566 if (keg->uk_flags & UMA_ZONE_PCPU)
2567 keg->uk_allocf = pcpu_page_alloc;
2568 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2569 keg->uk_ppera > 1)
2570 keg->uk_allocf = contig_alloc;
2571 else
2572 keg->uk_allocf = page_alloc;
2573 }
2574 }
2575
2576 static void
zone_alloc_counters(uma_zone_t zone,void * unused)2577 zone_alloc_counters(uma_zone_t zone, void *unused)
2578 {
2579
2580 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2581 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2582 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2583 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2584 }
2585
2586 static void
zone_alloc_sysctl(uma_zone_t zone,void * unused)2587 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2588 {
2589 uma_zone_domain_t zdom;
2590 uma_domain_t dom;
2591 uma_keg_t keg;
2592 struct sysctl_oid *oid, *domainoid;
2593 int domains, i, cnt;
2594 static const char *nokeg = "cache zone";
2595 char *c;
2596
2597 /*
2598 * Make a sysctl safe copy of the zone name by removing
2599 * any special characters and handling dups by appending
2600 * an index.
2601 */
2602 if (zone->uz_namecnt != 0) {
2603 /* Count the number of decimal digits and '_' separator. */
2604 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2605 cnt /= 10;
2606 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2607 M_UMA, M_WAITOK);
2608 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2609 zone->uz_namecnt);
2610 } else
2611 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2612 for (c = zone->uz_ctlname; *c != '\0'; c++)
2613 if (strchr("./\\ -", *c) != NULL)
2614 *c = '_';
2615
2616 /*
2617 * Basic parameters at the root.
2618 */
2619 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2620 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2621 oid = zone->uz_oid;
2622 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2623 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2624 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2625 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2626 zone, 0, sysctl_handle_uma_zone_flags, "A",
2627 "Allocator configuration flags");
2628 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2629 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2630 "Desired per-cpu cache size");
2631 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2632 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2633 "Maximum allowed per-cpu cache size");
2634
2635 /*
2636 * keg if present.
2637 */
2638 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2639 domains = vm_ndomains;
2640 else
2641 domains = 1;
2642 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2643 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2644 keg = zone->uz_keg;
2645 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2646 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2647 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2648 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2649 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2650 "Real object size with alignment");
2651 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2652 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2653 "pages per-slab allocation");
2654 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2655 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2656 "items available per-slab");
2657 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2658 "align", CTLFLAG_RD, &keg->uk_align, 0,
2659 "item alignment mask");
2660 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2661 "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2662 "number of reserved items");
2663 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2664 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2665 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2666 "Slab utilization (100 - internal fragmentation %)");
2667 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2668 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2669 for (i = 0; i < domains; i++) {
2670 dom = &keg->uk_domain[i];
2671 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2672 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2673 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2674 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2675 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2676 "Total pages currently allocated from VM");
2677 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2678 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2679 "Items free in the slab layer");
2680 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2681 "free_slabs", CTLFLAG_RD, &dom->ud_free_slabs, 0,
2682 "Unused slabs");
2683 }
2684 } else
2685 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2686 "name", CTLFLAG_RD, nokeg, "Keg name");
2687
2688 /*
2689 * Information about zone limits.
2690 */
2691 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2692 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2693 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2694 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2695 zone, 0, sysctl_handle_uma_zone_items, "QU",
2696 "Current number of allocated items if limit is set");
2697 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2698 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2699 "Maximum number of allocated and cached items");
2700 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2701 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2702 "Number of threads sleeping at limit");
2703 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2704 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2705 "Total zone limit sleeps");
2706 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2707 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2708 "Maximum number of items in each domain's bucket cache");
2709
2710 /*
2711 * Per-domain zone information.
2712 */
2713 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2714 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2715 for (i = 0; i < domains; i++) {
2716 zdom = ZDOM_GET(zone, i);
2717 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2718 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2719 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2720 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2721 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2722 "number of items in this domain");
2723 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2724 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2725 "maximum item count in this period");
2726 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2727 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2728 "minimum item count in this period");
2729 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2730 "bimin", CTLFLAG_RD, &zdom->uzd_bimin,
2731 "Minimum item count in this batch");
2732 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2733 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2734 "Working set size");
2735 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2736 "limin", CTLFLAG_RD, &zdom->uzd_limin,
2737 "Long time minimum item count");
2738 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2739 "timin", CTLFLAG_RD, &zdom->uzd_timin, 0,
2740 "Time since zero long time minimum item count");
2741 }
2742
2743 /*
2744 * General statistics.
2745 */
2746 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2747 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2748 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2749 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2750 zone, 1, sysctl_handle_uma_zone_cur, "I",
2751 "Current number of allocated items");
2752 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2753 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2754 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2755 "Total allocation calls");
2756 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2757 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2758 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2759 "Total free calls");
2760 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2761 "fails", CTLFLAG_RD, &zone->uz_fails,
2762 "Number of allocation failures");
2763 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2764 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2765 "Free calls from the wrong domain");
2766 }
2767
2768 struct uma_zone_count {
2769 const char *name;
2770 int count;
2771 };
2772
2773 static void
zone_count(uma_zone_t zone,void * arg)2774 zone_count(uma_zone_t zone, void *arg)
2775 {
2776 struct uma_zone_count *cnt;
2777
2778 cnt = arg;
2779 /*
2780 * Some zones are rapidly created with identical names and
2781 * destroyed out of order. This can lead to gaps in the count.
2782 * Use one greater than the maximum observed for this name.
2783 */
2784 if (strcmp(zone->uz_name, cnt->name) == 0)
2785 cnt->count = MAX(cnt->count,
2786 zone->uz_namecnt + 1);
2787 }
2788
2789 static void
zone_update_caches(uma_zone_t zone)2790 zone_update_caches(uma_zone_t zone)
2791 {
2792 int i;
2793
2794 for (i = 0; i <= mp_maxid; i++) {
2795 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2796 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2797 }
2798 }
2799
2800 /*
2801 * Zone header ctor. This initializes all fields, locks, etc.
2802 *
2803 * Arguments/Returns follow uma_ctor specifications
2804 * udata Actually uma_zctor_args
2805 */
2806 static int
zone_ctor(void * mem,int size,void * udata,int flags)2807 zone_ctor(void *mem, int size, void *udata, int flags)
2808 {
2809 struct uma_zone_count cnt;
2810 struct uma_zctor_args *arg = udata;
2811 uma_zone_domain_t zdom;
2812 uma_zone_t zone = mem;
2813 uma_zone_t z;
2814 uma_keg_t keg;
2815 int i;
2816
2817 bzero(zone, size);
2818 zone->uz_name = arg->name;
2819 zone->uz_ctor = arg->ctor;
2820 zone->uz_dtor = arg->dtor;
2821 zone->uz_init = NULL;
2822 zone->uz_fini = NULL;
2823 zone->uz_sleeps = 0;
2824 zone->uz_bucket_size = 0;
2825 zone->uz_bucket_size_min = 0;
2826 zone->uz_bucket_size_max = BUCKET_MAX;
2827 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2828 zone->uz_warning = NULL;
2829 /* The domain structures follow the cpu structures. */
2830 zone->uz_bucket_max = ULONG_MAX;
2831 timevalclear(&zone->uz_ratecheck);
2832
2833 /* Count the number of duplicate names. */
2834 cnt.name = arg->name;
2835 cnt.count = 0;
2836 zone_foreach(zone_count, &cnt);
2837 zone->uz_namecnt = cnt.count;
2838 ZONE_CROSS_LOCK_INIT(zone);
2839
2840 for (i = 0; i < vm_ndomains; i++) {
2841 zdom = ZDOM_GET(zone, i);
2842 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2843 STAILQ_INIT(&zdom->uzd_buckets);
2844 }
2845
2846 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
2847 if (arg->uminit == trash_init && arg->fini == trash_fini)
2848 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2849 #elif defined(KASAN)
2850 if ((arg->flags & (UMA_ZONE_NOFREE | UMA_ZFLAG_CACHE)) != 0)
2851 arg->flags |= UMA_ZONE_NOKASAN;
2852 #endif
2853
2854 /*
2855 * This is a pure cache zone, no kegs.
2856 */
2857 if (arg->import) {
2858 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2859 ("zone_ctor: Import specified for non-cache zone."));
2860 zone->uz_flags = arg->flags;
2861 zone->uz_size = arg->size;
2862 zone->uz_import = arg->import;
2863 zone->uz_release = arg->release;
2864 zone->uz_arg = arg->arg;
2865 #ifdef NUMA
2866 /*
2867 * Cache zones are round-robin unless a policy is
2868 * specified because they may have incompatible
2869 * constraints.
2870 */
2871 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2872 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2873 #endif
2874 rw_wlock(&uma_rwlock);
2875 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2876 rw_wunlock(&uma_rwlock);
2877 goto out;
2878 }
2879
2880 /*
2881 * Use the regular zone/keg/slab allocator.
2882 */
2883 zone->uz_import = zone_import;
2884 zone->uz_release = zone_release;
2885 zone->uz_arg = zone;
2886 keg = arg->keg;
2887
2888 if (arg->flags & UMA_ZONE_SECONDARY) {
2889 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2890 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2891 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2892 zone->uz_init = arg->uminit;
2893 zone->uz_fini = arg->fini;
2894 zone->uz_flags |= UMA_ZONE_SECONDARY;
2895 rw_wlock(&uma_rwlock);
2896 ZONE_LOCK(zone);
2897 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2898 if (LIST_NEXT(z, uz_link) == NULL) {
2899 LIST_INSERT_AFTER(z, zone, uz_link);
2900 break;
2901 }
2902 }
2903 ZONE_UNLOCK(zone);
2904 rw_wunlock(&uma_rwlock);
2905 } else if (keg == NULL) {
2906 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2907 arg->align, arg->flags)) == NULL)
2908 return (ENOMEM);
2909 } else {
2910 struct uma_kctor_args karg;
2911 int error;
2912
2913 /* We should only be here from uma_startup() */
2914 karg.size = arg->size;
2915 karg.uminit = arg->uminit;
2916 karg.fini = arg->fini;
2917 karg.align = arg->align;
2918 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2919 karg.zone = zone;
2920 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2921 flags);
2922 if (error)
2923 return (error);
2924 }
2925
2926 /* Inherit properties from the keg. */
2927 zone->uz_keg = keg;
2928 zone->uz_size = keg->uk_size;
2929 zone->uz_flags |= (keg->uk_flags &
2930 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2931
2932 out:
2933 if (booted >= BOOT_PCPU) {
2934 zone_alloc_counters(zone, NULL);
2935 if (booted >= BOOT_RUNNING)
2936 zone_alloc_sysctl(zone, NULL);
2937 } else {
2938 zone->uz_allocs = EARLY_COUNTER;
2939 zone->uz_frees = EARLY_COUNTER;
2940 zone->uz_fails = EARLY_COUNTER;
2941 }
2942
2943 /* Caller requests a private SMR context. */
2944 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2945 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2946
2947 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2948 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2949 ("Invalid zone flag combination"));
2950 if (arg->flags & UMA_ZFLAG_INTERNAL)
2951 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2952 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2953 zone->uz_bucket_size = BUCKET_MAX;
2954 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2955 zone->uz_bucket_size = 0;
2956 else
2957 zone->uz_bucket_size = bucket_select(zone->uz_size);
2958 zone->uz_bucket_size_min = zone->uz_bucket_size;
2959 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2960 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2961 zone_update_caches(zone);
2962
2963 return (0);
2964 }
2965
2966 /*
2967 * Keg header dtor. This frees all data, destroys locks, frees the hash
2968 * table and removes the keg from the global list.
2969 *
2970 * Arguments/Returns follow uma_dtor specifications
2971 * udata unused
2972 */
2973 static void
keg_dtor(void * arg,int size,void * udata)2974 keg_dtor(void *arg, int size, void *udata)
2975 {
2976 uma_keg_t keg;
2977 uint32_t free, pages;
2978 int i;
2979
2980 keg = (uma_keg_t)arg;
2981 free = pages = 0;
2982 for (i = 0; i < vm_ndomains; i++) {
2983 free += keg->uk_domain[i].ud_free_items;
2984 pages += keg->uk_domain[i].ud_pages;
2985 KEG_LOCK_FINI(keg, i);
2986 }
2987 if (pages != 0)
2988 printf("Freed UMA keg (%s) was not empty (%u items). "
2989 " Lost %u pages of memory.\n",
2990 keg->uk_name ? keg->uk_name : "",
2991 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2992
2993 hash_free(&keg->uk_hash);
2994 }
2995
2996 /*
2997 * Zone header dtor.
2998 *
2999 * Arguments/Returns follow uma_dtor specifications
3000 * udata unused
3001 */
3002 static void
zone_dtor(void * arg,int size,void * udata)3003 zone_dtor(void *arg, int size, void *udata)
3004 {
3005 uma_zone_t zone;
3006 uma_keg_t keg;
3007 int i;
3008
3009 zone = (uma_zone_t)arg;
3010
3011 sysctl_remove_oid(zone->uz_oid, 1, 1);
3012
3013 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
3014 cache_drain(zone);
3015
3016 rw_wlock(&uma_rwlock);
3017 LIST_REMOVE(zone, uz_link);
3018 rw_wunlock(&uma_rwlock);
3019 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3020 keg = zone->uz_keg;
3021 keg->uk_reserve = 0;
3022 }
3023 zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true);
3024
3025 /*
3026 * We only destroy kegs from non secondary/non cache zones.
3027 */
3028 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
3029 keg = zone->uz_keg;
3030 rw_wlock(&uma_rwlock);
3031 LIST_REMOVE(keg, uk_link);
3032 rw_wunlock(&uma_rwlock);
3033 zone_free_item(kegs, keg, NULL, SKIP_NONE);
3034 }
3035 counter_u64_free(zone->uz_allocs);
3036 counter_u64_free(zone->uz_frees);
3037 counter_u64_free(zone->uz_fails);
3038 counter_u64_free(zone->uz_xdomain);
3039 free(zone->uz_ctlname, M_UMA);
3040 for (i = 0; i < vm_ndomains; i++)
3041 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
3042 ZONE_CROSS_LOCK_FINI(zone);
3043 }
3044
3045 static void
zone_foreach_unlocked(void (* zfunc)(uma_zone_t,void * arg),void * arg)3046 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3047 {
3048 uma_keg_t keg;
3049 uma_zone_t zone;
3050
3051 LIST_FOREACH(keg, &uma_kegs, uk_link) {
3052 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
3053 zfunc(zone, arg);
3054 }
3055 LIST_FOREACH(zone, &uma_cachezones, uz_link)
3056 zfunc(zone, arg);
3057 }
3058
3059 /*
3060 * Traverses every zone in the system and calls a callback
3061 *
3062 * Arguments:
3063 * zfunc A pointer to a function which accepts a zone
3064 * as an argument.
3065 *
3066 * Returns:
3067 * Nothing
3068 */
3069 static void
zone_foreach(void (* zfunc)(uma_zone_t,void * arg),void * arg)3070 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3071 {
3072
3073 rw_rlock(&uma_rwlock);
3074 zone_foreach_unlocked(zfunc, arg);
3075 rw_runlock(&uma_rwlock);
3076 }
3077
3078 /*
3079 * Initialize the kernel memory allocator. This is done after pages can be
3080 * allocated but before general KVA is available.
3081 */
3082 void
uma_startup1(vm_offset_t virtual_avail)3083 uma_startup1(vm_offset_t virtual_avail)
3084 {
3085 struct uma_zctor_args args;
3086 size_t ksize, zsize, size;
3087 uma_keg_t primarykeg;
3088 uintptr_t m;
3089 int domain;
3090 uint8_t pflag;
3091
3092 bootstart = bootmem = virtual_avail;
3093
3094 rw_init(&uma_rwlock, "UMA lock");
3095 sx_init(&uma_reclaim_lock, "umareclaim");
3096
3097 ksize = sizeof(struct uma_keg) +
3098 (sizeof(struct uma_domain) * vm_ndomains);
3099 ksize = roundup(ksize, UMA_SUPER_ALIGN);
3100 zsize = sizeof(struct uma_zone) +
3101 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
3102 (sizeof(struct uma_zone_domain) * vm_ndomains);
3103 zsize = roundup(zsize, UMA_SUPER_ALIGN);
3104
3105 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
3106 size = (zsize * 2) + ksize;
3107 for (domain = 0; domain < vm_ndomains; domain++) {
3108 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
3109 M_NOWAIT | M_ZERO);
3110 if (m != 0)
3111 break;
3112 }
3113 zones = (uma_zone_t)m;
3114 m += zsize;
3115 kegs = (uma_zone_t)m;
3116 m += zsize;
3117 primarykeg = (uma_keg_t)m;
3118
3119 /* "manually" create the initial zone */
3120 memset(&args, 0, sizeof(args));
3121 args.name = "UMA Kegs";
3122 args.size = ksize;
3123 args.ctor = keg_ctor;
3124 args.dtor = keg_dtor;
3125 args.uminit = zero_init;
3126 args.fini = NULL;
3127 args.keg = primarykeg;
3128 args.align = UMA_SUPER_ALIGN - 1;
3129 args.flags = UMA_ZFLAG_INTERNAL;
3130 zone_ctor(kegs, zsize, &args, M_WAITOK);
3131
3132 args.name = "UMA Zones";
3133 args.size = zsize;
3134 args.ctor = zone_ctor;
3135 args.dtor = zone_dtor;
3136 args.uminit = zero_init;
3137 args.fini = NULL;
3138 args.keg = NULL;
3139 args.align = UMA_SUPER_ALIGN - 1;
3140 args.flags = UMA_ZFLAG_INTERNAL;
3141 zone_ctor(zones, zsize, &args, M_WAITOK);
3142
3143 /* Now make zones for slab headers */
3144 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
3145 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3146 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
3147 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3148
3149 hashzone = uma_zcreate("UMA Hash",
3150 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
3151 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3152
3153 bucket_init();
3154 smr_init();
3155 }
3156
3157 #ifndef UMA_MD_SMALL_ALLOC
3158 extern void vm_radix_reserve_kva(void);
3159 #endif
3160
3161 /*
3162 * Advertise the availability of normal kva allocations and switch to
3163 * the default back-end allocator. Marks the KVA we consumed on startup
3164 * as used in the map.
3165 */
3166 void
uma_startup2(void)3167 uma_startup2(void)
3168 {
3169
3170 if (bootstart != bootmem) {
3171 vm_map_lock(kernel_map);
3172 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
3173 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
3174 vm_map_unlock(kernel_map);
3175 }
3176
3177 #ifndef UMA_MD_SMALL_ALLOC
3178 /* Set up radix zone to use noobj_alloc. */
3179 vm_radix_reserve_kva();
3180 #endif
3181
3182 booted = BOOT_KVA;
3183 zone_foreach_unlocked(zone_kva_available, NULL);
3184 bucket_enable();
3185 }
3186
3187 /*
3188 * Allocate counters as early as possible so that boot-time allocations are
3189 * accounted more precisely.
3190 */
3191 static void
uma_startup_pcpu(void * arg __unused)3192 uma_startup_pcpu(void *arg __unused)
3193 {
3194
3195 zone_foreach_unlocked(zone_alloc_counters, NULL);
3196 booted = BOOT_PCPU;
3197 }
3198 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
3199
3200 /*
3201 * Finish our initialization steps.
3202 */
3203 static void
uma_startup3(void * arg __unused)3204 uma_startup3(void *arg __unused)
3205 {
3206
3207 #ifdef INVARIANTS
3208 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
3209 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
3210 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
3211 #endif
3212 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
3213 booted = BOOT_RUNNING;
3214
3215 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
3216 EVENTHANDLER_PRI_FIRST);
3217 }
3218 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
3219
3220 static void
uma_startup4(void * arg __unused)3221 uma_startup4(void *arg __unused)
3222 {
3223 TIMEOUT_TASK_INIT(taskqueue_thread, &uma_timeout_task, 0, uma_timeout,
3224 NULL);
3225 taskqueue_enqueue_timeout(taskqueue_thread, &uma_timeout_task,
3226 UMA_TIMEOUT * hz);
3227 }
3228 SYSINIT(uma_startup4, SI_SUB_TASKQ, SI_ORDER_ANY, uma_startup4, NULL);
3229
3230 static void
uma_shutdown(void)3231 uma_shutdown(void)
3232 {
3233
3234 booted = BOOT_SHUTDOWN;
3235 }
3236
3237 static uma_keg_t
uma_kcreate(uma_zone_t zone,size_t size,uma_init uminit,uma_fini fini,int align,uint32_t flags)3238 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
3239 int align, uint32_t flags)
3240 {
3241 struct uma_kctor_args args;
3242
3243 args.size = size;
3244 args.uminit = uminit;
3245 args.fini = fini;
3246 args.align = align;
3247 args.flags = flags;
3248 args.zone = zone;
3249 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3250 }
3251
3252
3253 static void
check_align_mask(unsigned int mask)3254 check_align_mask(unsigned int mask)
3255 {
3256
3257 KASSERT(powerof2(mask + 1),
3258 ("UMA: %s: Not the mask of a power of 2 (%#x)", __func__, mask));
3259 /*
3260 * Make sure the stored align mask doesn't have its highest bit set,
3261 * which would cause implementation-defined behavior when passing it as
3262 * the 'align' argument of uma_zcreate(). Such very large alignments do
3263 * not make sense anyway.
3264 */
3265 KASSERT(mask <= INT_MAX,
3266 ("UMA: %s: Mask too big (%#x)", __func__, mask));
3267 }
3268
3269 /* Public functions */
3270 /* See uma.h */
3271 void
uma_set_cache_align_mask(unsigned int mask)3272 uma_set_cache_align_mask(unsigned int mask)
3273 {
3274
3275 check_align_mask(mask);
3276 uma_cache_align_mask = mask;
3277 }
3278
3279 /* Returns the alignment mask to use to request cache alignment. */
3280 unsigned int
uma_get_cache_align_mask(void)3281 uma_get_cache_align_mask(void)
3282 {
3283 return (uma_cache_align_mask);
3284 }
3285
3286 /* See uma.h */
3287 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)3288 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3289 uma_init uminit, uma_fini fini, int align, uint32_t flags)
3290
3291 {
3292 struct uma_zctor_args args;
3293 uma_zone_t res;
3294
3295 check_align_mask(align);
3296
3297 /* This stuff is essential for the zone ctor */
3298 memset(&args, 0, sizeof(args));
3299 args.name = name;
3300 args.size = size;
3301 args.ctor = ctor;
3302 args.dtor = dtor;
3303 args.uminit = uminit;
3304 args.fini = fini;
3305 #if defined(INVARIANTS) && !defined(KASAN) && !defined(KMSAN)
3306 /*
3307 * Inject procedures which check for memory use after free if we are
3308 * allowed to scramble the memory while it is not allocated. This
3309 * requires that: UMA is actually able to access the memory, no init
3310 * or fini procedures, no dependency on the initial value of the
3311 * memory, and no (legitimate) use of the memory after free. Note,
3312 * the ctor and dtor do not need to be empty.
3313 */
3314 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3315 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3316 args.uminit = trash_init;
3317 args.fini = trash_fini;
3318 }
3319 #endif
3320 args.align = align;
3321 args.flags = flags;
3322 args.keg = NULL;
3323
3324 sx_xlock(&uma_reclaim_lock);
3325 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3326 sx_xunlock(&uma_reclaim_lock);
3327
3328 return (res);
3329 }
3330
3331 /* See uma.h */
3332 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)3333 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3334 uma_init zinit, uma_fini zfini, uma_zone_t primary)
3335 {
3336 struct uma_zctor_args args;
3337 uma_keg_t keg;
3338 uma_zone_t res;
3339
3340 keg = primary->uz_keg;
3341 memset(&args, 0, sizeof(args));
3342 args.name = name;
3343 args.size = keg->uk_size;
3344 args.ctor = ctor;
3345 args.dtor = dtor;
3346 args.uminit = zinit;
3347 args.fini = zfini;
3348 args.align = keg->uk_align;
3349 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3350 args.keg = keg;
3351
3352 sx_xlock(&uma_reclaim_lock);
3353 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3354 sx_xunlock(&uma_reclaim_lock);
3355
3356 return (res);
3357 }
3358
3359 /* See uma.h */
3360 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)3361 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3362 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3363 void *arg, int flags)
3364 {
3365 struct uma_zctor_args args;
3366
3367 memset(&args, 0, sizeof(args));
3368 args.name = name;
3369 args.size = size;
3370 args.ctor = ctor;
3371 args.dtor = dtor;
3372 args.uminit = zinit;
3373 args.fini = zfini;
3374 args.import = zimport;
3375 args.release = zrelease;
3376 args.arg = arg;
3377 args.align = 0;
3378 args.flags = flags | UMA_ZFLAG_CACHE;
3379
3380 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3381 }
3382
3383 /* See uma.h */
3384 void
uma_zdestroy(uma_zone_t zone)3385 uma_zdestroy(uma_zone_t zone)
3386 {
3387
3388 /*
3389 * Large slabs are expensive to reclaim, so don't bother doing
3390 * unnecessary work if we're shutting down.
3391 */
3392 if (booted == BOOT_SHUTDOWN &&
3393 zone->uz_fini == NULL && zone->uz_release == zone_release)
3394 return;
3395 sx_xlock(&uma_reclaim_lock);
3396 zone_free_item(zones, zone, NULL, SKIP_NONE);
3397 sx_xunlock(&uma_reclaim_lock);
3398 }
3399
3400 void
uma_zwait(uma_zone_t zone)3401 uma_zwait(uma_zone_t zone)
3402 {
3403
3404 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3405 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3406 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3407 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3408 else
3409 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3410 }
3411
3412 void *
uma_zalloc_pcpu_arg(uma_zone_t zone,void * udata,int flags)3413 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3414 {
3415 void *item, *pcpu_item;
3416 #ifdef SMP
3417 int i;
3418
3419 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3420 #endif
3421 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3422 if (item == NULL)
3423 return (NULL);
3424 pcpu_item = zpcpu_base_to_offset(item);
3425 if (flags & M_ZERO) {
3426 #ifdef SMP
3427 for (i = 0; i <= mp_maxid; i++)
3428 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3429 #else
3430 bzero(item, zone->uz_size);
3431 #endif
3432 }
3433 return (pcpu_item);
3434 }
3435
3436 /*
3437 * A stub while both regular and pcpu cases are identical.
3438 */
3439 void
uma_zfree_pcpu_arg(uma_zone_t zone,void * pcpu_item,void * udata)3440 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3441 {
3442 void *item;
3443
3444 #ifdef SMP
3445 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3446 #endif
3447
3448 /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */
3449 if (pcpu_item == NULL)
3450 return;
3451
3452 item = zpcpu_offset_to_base(pcpu_item);
3453 uma_zfree_arg(zone, item, udata);
3454 }
3455
3456 static inline void *
item_ctor(uma_zone_t zone,int uz_flags,int size,void * udata,int flags,void * item)3457 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3458 void *item)
3459 {
3460 #ifdef INVARIANTS
3461 bool skipdbg;
3462 #endif
3463
3464 kasan_mark_item_valid(zone, item);
3465 kmsan_mark_item_uninitialized(zone, item);
3466
3467 #ifdef INVARIANTS
3468 skipdbg = uma_dbg_zskip(zone, item);
3469 if (!skipdbg && (uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3470 zone->uz_ctor != trash_ctor)
3471 trash_ctor(item, size, zone, flags);
3472 #endif
3473
3474 /* Check flags before loading ctor pointer. */
3475 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3476 __predict_false(zone->uz_ctor != NULL) &&
3477 zone->uz_ctor(item, size, udata, flags) != 0) {
3478 counter_u64_add(zone->uz_fails, 1);
3479 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3480 return (NULL);
3481 }
3482 #ifdef INVARIANTS
3483 if (!skipdbg)
3484 uma_dbg_alloc(zone, NULL, item);
3485 #endif
3486 if (__predict_false(flags & M_ZERO))
3487 return (memset(item, 0, size));
3488
3489 return (item);
3490 }
3491
3492 static inline void
item_dtor(uma_zone_t zone,void * item,int size,void * udata,enum zfreeskip skip)3493 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3494 enum zfreeskip skip)
3495 {
3496 #ifdef INVARIANTS
3497 bool skipdbg;
3498
3499 skipdbg = uma_dbg_zskip(zone, item);
3500 if (skip == SKIP_NONE && !skipdbg) {
3501 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3502 uma_dbg_free(zone, udata, item);
3503 else
3504 uma_dbg_free(zone, NULL, item);
3505 }
3506 #endif
3507 if (__predict_true(skip < SKIP_DTOR)) {
3508 if (zone->uz_dtor != NULL)
3509 zone->uz_dtor(item, size, udata);
3510 #ifdef INVARIANTS
3511 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3512 zone->uz_dtor != trash_dtor)
3513 trash_dtor(item, size, zone);
3514 #endif
3515 }
3516 kasan_mark_item_invalid(zone, item);
3517 }
3518
3519 #ifdef NUMA
3520 static int
item_domain(void * item)3521 item_domain(void *item)
3522 {
3523 int domain;
3524
3525 domain = vm_phys_domain(vtophys(item));
3526 KASSERT(domain >= 0 && domain < vm_ndomains,
3527 ("%s: unknown domain for item %p", __func__, item));
3528 return (domain);
3529 }
3530 #endif
3531
3532 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3533 #if defined(INVARIANTS) && (defined(DDB) || defined(STACK))
3534 #include <sys/stack.h>
3535 #endif
3536 #define UMA_ZALLOC_DEBUG
3537 static int
uma_zalloc_debug(uma_zone_t zone,void ** itemp,void * udata,int flags)3538 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3539 {
3540 int error;
3541
3542 error = 0;
3543 #ifdef WITNESS
3544 if (flags & M_WAITOK) {
3545 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3546 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3547 }
3548 #endif
3549
3550 #ifdef INVARIANTS
3551 KASSERT((flags & M_EXEC) == 0,
3552 ("uma_zalloc_debug: called with M_EXEC"));
3553 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3554 ("uma_zalloc_debug: called within spinlock or critical section"));
3555 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3556 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3557
3558 _Static_assert(M_NOWAIT != 0 && M_WAITOK != 0,
3559 "M_NOWAIT and M_WAITOK must be non-zero for this assertion:");
3560 #if 0
3561 /*
3562 * Give the #elif clause time to find problems, then remove it
3563 * and enable this. (Remove <sys/stack.h> above, too.)
3564 */
3565 KASSERT((flags & (M_NOWAIT|M_WAITOK)) == M_NOWAIT ||
3566 (flags & (M_NOWAIT|M_WAITOK)) == M_WAITOK,
3567 ("uma_zalloc_debug: must pass one of M_NOWAIT or M_WAITOK"));
3568 #elif defined(DDB) || defined(STACK)
3569 if (__predict_false((flags & (M_NOWAIT|M_WAITOK)) != M_NOWAIT &&
3570 (flags & (M_NOWAIT|M_WAITOK)) != M_WAITOK)) {
3571 static int stack_count;
3572 struct stack st;
3573
3574 if (stack_count < 10) {
3575 ++stack_count;
3576 printf("uma_zalloc* called with bad WAIT flags:\n");
3577 stack_save(&st);
3578 stack_print(&st);
3579 }
3580 }
3581 #endif
3582 #endif
3583
3584 #ifdef DEBUG_MEMGUARD
3585 if ((zone->uz_flags & (UMA_ZONE_SMR | UMA_ZFLAG_CACHE)) == 0 &&
3586 memguard_cmp_zone(zone)) {
3587 void *item;
3588 item = memguard_alloc(zone->uz_size, flags);
3589 if (item != NULL) {
3590 error = EJUSTRETURN;
3591 if (zone->uz_init != NULL &&
3592 zone->uz_init(item, zone->uz_size, flags) != 0) {
3593 *itemp = NULL;
3594 return (error);
3595 }
3596 if (zone->uz_ctor != NULL &&
3597 zone->uz_ctor(item, zone->uz_size, udata,
3598 flags) != 0) {
3599 counter_u64_add(zone->uz_fails, 1);
3600 if (zone->uz_fini != NULL)
3601 zone->uz_fini(item, zone->uz_size);
3602 *itemp = NULL;
3603 return (error);
3604 }
3605 *itemp = item;
3606 return (error);
3607 }
3608 /* This is unfortunate but should not be fatal. */
3609 }
3610 #endif
3611 return (error);
3612 }
3613
3614 static int
uma_zfree_debug(uma_zone_t zone,void * item,void * udata)3615 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3616 {
3617 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3618 ("uma_zfree_debug: called with spinlock or critical section held"));
3619
3620 #ifdef DEBUG_MEMGUARD
3621 if ((zone->uz_flags & (UMA_ZONE_SMR | UMA_ZFLAG_CACHE)) == 0 &&
3622 is_memguard_addr(item)) {
3623 if (zone->uz_dtor != NULL)
3624 zone->uz_dtor(item, zone->uz_size, udata);
3625 if (zone->uz_fini != NULL)
3626 zone->uz_fini(item, zone->uz_size);
3627 memguard_free(item);
3628 return (EJUSTRETURN);
3629 }
3630 #endif
3631 return (0);
3632 }
3633 #endif
3634
3635 static inline void *
cache_alloc_item(uma_zone_t zone,uma_cache_t cache,uma_cache_bucket_t bucket,void * udata,int flags)3636 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3637 void *udata, int flags)
3638 {
3639 void *item;
3640 int size, uz_flags;
3641
3642 item = cache_bucket_pop(cache, bucket);
3643 size = cache_uz_size(cache);
3644 uz_flags = cache_uz_flags(cache);
3645 critical_exit();
3646 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3647 }
3648
3649 static __noinline void *
cache_alloc_retry(uma_zone_t zone,uma_cache_t cache,void * udata,int flags)3650 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3651 {
3652 uma_cache_bucket_t bucket;
3653 int domain;
3654
3655 while (cache_alloc(zone, cache, udata, flags)) {
3656 cache = &zone->uz_cpu[curcpu];
3657 bucket = &cache->uc_allocbucket;
3658 if (__predict_false(bucket->ucb_cnt == 0))
3659 continue;
3660 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3661 }
3662 critical_exit();
3663
3664 /*
3665 * We can not get a bucket so try to return a single item.
3666 */
3667 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3668 domain = PCPU_GET(domain);
3669 else
3670 domain = UMA_ANYDOMAIN;
3671 return (zone_alloc_item(zone, udata, domain, flags));
3672 }
3673
3674 /* See uma.h */
3675 void *
uma_zalloc_smr(uma_zone_t zone,int flags)3676 uma_zalloc_smr(uma_zone_t zone, int flags)
3677 {
3678 uma_cache_bucket_t bucket;
3679 uma_cache_t cache;
3680
3681 CTR3(KTR_UMA, "uma_zalloc_smr zone %s(%p) flags %d", zone->uz_name,
3682 zone, flags);
3683
3684 #ifdef UMA_ZALLOC_DEBUG
3685 void *item;
3686
3687 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3688 ("uma_zalloc_arg: called with non-SMR zone."));
3689 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3690 return (item);
3691 #endif
3692
3693 critical_enter();
3694 cache = &zone->uz_cpu[curcpu];
3695 bucket = &cache->uc_allocbucket;
3696 if (__predict_false(bucket->ucb_cnt == 0))
3697 return (cache_alloc_retry(zone, cache, NULL, flags));
3698 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3699 }
3700
3701 /* See uma.h */
3702 void *
uma_zalloc_arg(uma_zone_t zone,void * udata,int flags)3703 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3704 {
3705 uma_cache_bucket_t bucket;
3706 uma_cache_t cache;
3707
3708 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3709 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3710
3711 /* This is the fast path allocation */
3712 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3713 zone, flags);
3714
3715 #ifdef UMA_ZALLOC_DEBUG
3716 void *item;
3717
3718 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3719 ("uma_zalloc_arg: called with SMR zone."));
3720 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3721 return (item);
3722 #endif
3723
3724 /*
3725 * If possible, allocate from the per-CPU cache. There are two
3726 * requirements for safe access to the per-CPU cache: (1) the thread
3727 * accessing the cache must not be preempted or yield during access,
3728 * and (2) the thread must not migrate CPUs without switching which
3729 * cache it accesses. We rely on a critical section to prevent
3730 * preemption and migration. We release the critical section in
3731 * order to acquire the zone mutex if we are unable to allocate from
3732 * the current cache; when we re-acquire the critical section, we
3733 * must detect and handle migration if it has occurred.
3734 */
3735 critical_enter();
3736 cache = &zone->uz_cpu[curcpu];
3737 bucket = &cache->uc_allocbucket;
3738 if (__predict_false(bucket->ucb_cnt == 0))
3739 return (cache_alloc_retry(zone, cache, udata, flags));
3740 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3741 }
3742
3743 /*
3744 * Replenish an alloc bucket and possibly restore an old one. Called in
3745 * a critical section. Returns in a critical section.
3746 *
3747 * A false return value indicates an allocation failure.
3748 * A true return value indicates success and the caller should retry.
3749 */
3750 static __noinline bool
cache_alloc(uma_zone_t zone,uma_cache_t cache,void * udata,int flags)3751 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3752 {
3753 uma_bucket_t bucket;
3754 int curdomain, domain;
3755 bool new;
3756
3757 CRITICAL_ASSERT(curthread);
3758
3759 /*
3760 * If we have run out of items in our alloc bucket see
3761 * if we can switch with the free bucket.
3762 *
3763 * SMR Zones can't re-use the free bucket until the sequence has
3764 * expired.
3765 */
3766 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3767 cache->uc_freebucket.ucb_cnt != 0) {
3768 cache_bucket_swap(&cache->uc_freebucket,
3769 &cache->uc_allocbucket);
3770 return (true);
3771 }
3772
3773 /*
3774 * Discard any empty allocation bucket while we hold no locks.
3775 */
3776 bucket = cache_bucket_unload_alloc(cache);
3777 critical_exit();
3778
3779 if (bucket != NULL) {
3780 KASSERT(bucket->ub_cnt == 0,
3781 ("cache_alloc: Entered with non-empty alloc bucket."));
3782 bucket_free(zone, bucket, udata);
3783 }
3784
3785 /*
3786 * Attempt to retrieve the item from the per-CPU cache has failed, so
3787 * we must go back to the zone. This requires the zdom lock, so we
3788 * must drop the critical section, then re-acquire it when we go back
3789 * to the cache. Since the critical section is released, we may be
3790 * preempted or migrate. As such, make sure not to maintain any
3791 * thread-local state specific to the cache from prior to releasing
3792 * the critical section.
3793 */
3794 domain = PCPU_GET(domain);
3795 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3796 VM_DOMAIN_EMPTY(domain))
3797 domain = zone_domain_highest(zone, domain);
3798 bucket = cache_fetch_bucket(zone, cache, domain);
3799 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3800 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3801 new = true;
3802 } else {
3803 new = false;
3804 }
3805
3806 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3807 zone->uz_name, zone, bucket);
3808 if (bucket == NULL) {
3809 critical_enter();
3810 return (false);
3811 }
3812
3813 /*
3814 * See if we lost the race or were migrated. Cache the
3815 * initialized bucket to make this less likely or claim
3816 * the memory directly.
3817 */
3818 critical_enter();
3819 cache = &zone->uz_cpu[curcpu];
3820 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3821 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3822 (curdomain = PCPU_GET(domain)) == domain ||
3823 VM_DOMAIN_EMPTY(curdomain))) {
3824 if (new)
3825 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3826 bucket->ub_cnt);
3827 cache_bucket_load_alloc(cache, bucket);
3828 return (true);
3829 }
3830
3831 /*
3832 * We lost the race, release this bucket and start over.
3833 */
3834 critical_exit();
3835 zone_put_bucket(zone, domain, bucket, udata, !new);
3836 critical_enter();
3837
3838 return (true);
3839 }
3840
3841 void *
uma_zalloc_domain(uma_zone_t zone,void * udata,int domain,int flags)3842 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3843 {
3844 #ifdef NUMA
3845 uma_bucket_t bucket;
3846 uma_zone_domain_t zdom;
3847 void *item;
3848 #endif
3849
3850 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3851 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3852
3853 /* This is the fast path allocation */
3854 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3855 zone->uz_name, zone, domain, flags);
3856
3857 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3858 ("uma_zalloc_domain: called with SMR zone."));
3859 #ifdef NUMA
3860 KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3861 ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3862
3863 if (vm_ndomains == 1)
3864 return (uma_zalloc_arg(zone, udata, flags));
3865
3866 #ifdef UMA_ZALLOC_DEBUG
3867 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3868 return (item);
3869 #endif
3870
3871 /*
3872 * Try to allocate from the bucket cache before falling back to the keg.
3873 * We could try harder and attempt to allocate from per-CPU caches or
3874 * the per-domain cross-domain buckets, but the complexity is probably
3875 * not worth it. It is more important that frees of previous
3876 * cross-domain allocations do not blow up the cache.
3877 */
3878 zdom = zone_domain_lock(zone, domain);
3879 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3880 item = bucket->ub_bucket[bucket->ub_cnt - 1];
3881 #ifdef INVARIANTS
3882 bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3883 #endif
3884 bucket->ub_cnt--;
3885 zone_put_bucket(zone, domain, bucket, udata, true);
3886 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3887 flags, item);
3888 if (item != NULL) {
3889 KASSERT(item_domain(item) == domain,
3890 ("%s: bucket cache item %p from wrong domain",
3891 __func__, item));
3892 counter_u64_add(zone->uz_allocs, 1);
3893 }
3894 return (item);
3895 }
3896 ZDOM_UNLOCK(zdom);
3897 return (zone_alloc_item(zone, udata, domain, flags));
3898 #else
3899 return (uma_zalloc_arg(zone, udata, flags));
3900 #endif
3901 }
3902
3903 /*
3904 * Find a slab with some space. Prefer slabs that are partially used over those
3905 * that are totally full. This helps to reduce fragmentation.
3906 *
3907 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3908 * only 'domain'.
3909 */
3910 static uma_slab_t
keg_first_slab(uma_keg_t keg,int domain,bool rr)3911 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3912 {
3913 uma_domain_t dom;
3914 uma_slab_t slab;
3915 int start;
3916
3917 KASSERT(domain >= 0 && domain < vm_ndomains,
3918 ("keg_first_slab: domain %d out of range", domain));
3919 KEG_LOCK_ASSERT(keg, domain);
3920
3921 slab = NULL;
3922 start = domain;
3923 do {
3924 dom = &keg->uk_domain[domain];
3925 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3926 return (slab);
3927 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3928 LIST_REMOVE(slab, us_link);
3929 dom->ud_free_slabs--;
3930 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3931 return (slab);
3932 }
3933 if (rr)
3934 domain = (domain + 1) % vm_ndomains;
3935 } while (domain != start);
3936
3937 return (NULL);
3938 }
3939
3940 /*
3941 * Fetch an existing slab from a free or partial list. Returns with the
3942 * keg domain lock held if a slab was found or unlocked if not.
3943 */
3944 static uma_slab_t
keg_fetch_free_slab(uma_keg_t keg,int domain,bool rr,int flags)3945 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3946 {
3947 uma_slab_t slab;
3948 uint32_t reserve;
3949
3950 /* HASH has a single free list. */
3951 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3952 domain = 0;
3953
3954 KEG_LOCK(keg, domain);
3955 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3956 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3957 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3958 KEG_UNLOCK(keg, domain);
3959 return (NULL);
3960 }
3961 return (slab);
3962 }
3963
3964 static uma_slab_t
keg_fetch_slab(uma_keg_t keg,uma_zone_t zone,int rdomain,const int flags)3965 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3966 {
3967 struct vm_domainset_iter di;
3968 uma_slab_t slab;
3969 int aflags, domain;
3970 bool rr;
3971
3972 KASSERT((flags & (M_WAITOK | M_NOVM)) != (M_WAITOK | M_NOVM),
3973 ("%s: invalid flags %#x", __func__, flags));
3974
3975 restart:
3976 /*
3977 * Use the keg's policy if upper layers haven't already specified a
3978 * domain (as happens with first-touch zones).
3979 *
3980 * To avoid races we run the iterator with the keg lock held, but that
3981 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3982 * clear M_WAITOK and handle low memory conditions locally.
3983 */
3984 rr = rdomain == UMA_ANYDOMAIN;
3985 if (rr) {
3986 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3987 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3988 &aflags);
3989 } else {
3990 aflags = flags;
3991 domain = rdomain;
3992 }
3993
3994 for (;;) {
3995 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3996 if (slab != NULL)
3997 return (slab);
3998
3999 /*
4000 * M_NOVM is used to break the recursion that can otherwise
4001 * occur if low-level memory management routines use UMA.
4002 */
4003 if ((flags & M_NOVM) == 0) {
4004 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
4005 if (slab != NULL)
4006 return (slab);
4007 }
4008
4009 if (!rr) {
4010 if ((flags & M_USE_RESERVE) != 0) {
4011 /*
4012 * Drain reserves from other domains before
4013 * giving up or sleeping. It may be useful to
4014 * support per-domain reserves eventually.
4015 */
4016 rdomain = UMA_ANYDOMAIN;
4017 goto restart;
4018 }
4019 if ((flags & M_WAITOK) == 0)
4020 break;
4021 vm_wait_domain(domain);
4022 } else if (vm_domainset_iter_policy(&di, &domain) != 0) {
4023 if ((flags & M_WAITOK) != 0) {
4024 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
4025 goto restart;
4026 }
4027 break;
4028 }
4029 }
4030
4031 /*
4032 * We might not have been able to get a slab but another cpu
4033 * could have while we were unlocked. Check again before we
4034 * fail.
4035 */
4036 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
4037 return (slab);
4038
4039 return (NULL);
4040 }
4041
4042 static void *
slab_alloc_item(uma_keg_t keg,uma_slab_t slab)4043 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
4044 {
4045 uma_domain_t dom;
4046 void *item;
4047 int freei;
4048
4049 KEG_LOCK_ASSERT(keg, slab->us_domain);
4050
4051 dom = &keg->uk_domain[slab->us_domain];
4052 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
4053 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
4054 item = slab_item(slab, keg, freei);
4055 slab->us_freecount--;
4056 dom->ud_free_items--;
4057
4058 /*
4059 * Move this slab to the full list. It must be on the partial list, so
4060 * we do not need to update the free slab count. In particular,
4061 * keg_fetch_slab() always returns slabs on the partial list.
4062 */
4063 if (slab->us_freecount == 0) {
4064 LIST_REMOVE(slab, us_link);
4065 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
4066 }
4067
4068 return (item);
4069 }
4070
4071 static int
zone_import(void * arg,void ** bucket,int max,int domain,int flags)4072 zone_import(void *arg, void **bucket, int max, int domain, int flags)
4073 {
4074 uma_domain_t dom;
4075 uma_zone_t zone;
4076 uma_slab_t slab;
4077 uma_keg_t keg;
4078 #ifdef NUMA
4079 int stripe;
4080 #endif
4081 int i;
4082
4083 zone = arg;
4084 slab = NULL;
4085 keg = zone->uz_keg;
4086 /* Try to keep the buckets totally full */
4087 for (i = 0; i < max; ) {
4088 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
4089 break;
4090 #ifdef NUMA
4091 stripe = howmany(max, vm_ndomains);
4092 #endif
4093 dom = &keg->uk_domain[slab->us_domain];
4094 do {
4095 bucket[i++] = slab_alloc_item(keg, slab);
4096 if (keg->uk_reserve > 0 &&
4097 dom->ud_free_items <= keg->uk_reserve) {
4098 /*
4099 * Avoid depleting the reserve after a
4100 * successful item allocation, even if
4101 * M_USE_RESERVE is specified.
4102 */
4103 KEG_UNLOCK(keg, slab->us_domain);
4104 goto out;
4105 }
4106 #ifdef NUMA
4107 /*
4108 * If the zone is striped we pick a new slab for every
4109 * N allocations. Eliminating this conditional will
4110 * instead pick a new domain for each bucket rather
4111 * than stripe within each bucket. The current option
4112 * produces more fragmentation and requires more cpu
4113 * time but yields better distribution.
4114 */
4115 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
4116 vm_ndomains > 1 && --stripe == 0)
4117 break;
4118 #endif
4119 } while (slab->us_freecount != 0 && i < max);
4120 KEG_UNLOCK(keg, slab->us_domain);
4121
4122 /* Don't block if we allocated any successfully. */
4123 flags &= ~M_WAITOK;
4124 flags |= M_NOWAIT;
4125 }
4126 out:
4127 return i;
4128 }
4129
4130 static int
zone_alloc_limit_hard(uma_zone_t zone,int count,int flags)4131 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
4132 {
4133 uint64_t old, new, total, max;
4134
4135 /*
4136 * The hard case. We're going to sleep because there were existing
4137 * sleepers or because we ran out of items. This routine enforces
4138 * fairness by keeping fifo order.
4139 *
4140 * First release our ill gotten gains and make some noise.
4141 */
4142 for (;;) {
4143 zone_free_limit(zone, count);
4144 zone_log_warning(zone);
4145 zone_maxaction(zone);
4146 if (flags & M_NOWAIT)
4147 return (0);
4148
4149 /*
4150 * We need to allocate an item or set ourself as a sleeper
4151 * while the sleepq lock is held to avoid wakeup races. This
4152 * is essentially a home rolled semaphore.
4153 */
4154 sleepq_lock(&zone->uz_max_items);
4155 old = zone->uz_items;
4156 do {
4157 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
4158 /* Cache the max since we will evaluate twice. */
4159 max = zone->uz_max_items;
4160 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
4161 UZ_ITEMS_COUNT(old) >= max)
4162 new = old + UZ_ITEMS_SLEEPER;
4163 else
4164 new = old + MIN(count, max - old);
4165 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
4166
4167 /* We may have successfully allocated under the sleepq lock. */
4168 if (UZ_ITEMS_SLEEPERS(new) == 0) {
4169 sleepq_release(&zone->uz_max_items);
4170 return (new - old);
4171 }
4172
4173 /*
4174 * This is in a different cacheline from uz_items so that we
4175 * don't constantly invalidate the fastpath cacheline when we
4176 * adjust item counts. This could be limited to toggling on
4177 * transitions.
4178 */
4179 atomic_add_32(&zone->uz_sleepers, 1);
4180 atomic_add_64(&zone->uz_sleeps, 1);
4181
4182 /*
4183 * We have added ourselves as a sleeper. The sleepq lock
4184 * protects us from wakeup races. Sleep now and then retry.
4185 */
4186 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
4187 sleepq_wait(&zone->uz_max_items, PVM);
4188
4189 /*
4190 * After wakeup, remove ourselves as a sleeper and try
4191 * again. We no longer have the sleepq lock for protection.
4192 *
4193 * Subract ourselves as a sleeper while attempting to add
4194 * our count.
4195 */
4196 atomic_subtract_32(&zone->uz_sleepers, 1);
4197 old = atomic_fetchadd_64(&zone->uz_items,
4198 -(UZ_ITEMS_SLEEPER - count));
4199 /* We're no longer a sleeper. */
4200 old -= UZ_ITEMS_SLEEPER;
4201
4202 /*
4203 * If we're still at the limit, restart. Notably do not
4204 * block on other sleepers. Cache the max value to protect
4205 * against changes via sysctl.
4206 */
4207 total = UZ_ITEMS_COUNT(old);
4208 max = zone->uz_max_items;
4209 if (total >= max)
4210 continue;
4211 /* Truncate if necessary, otherwise wake other sleepers. */
4212 if (total + count > max) {
4213 zone_free_limit(zone, total + count - max);
4214 count = max - total;
4215 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
4216 wakeup_one(&zone->uz_max_items);
4217
4218 return (count);
4219 }
4220 }
4221
4222 /*
4223 * Allocate 'count' items from our max_items limit. Returns the number
4224 * available. If M_NOWAIT is not specified it will sleep until at least
4225 * one item can be allocated.
4226 */
4227 static int
zone_alloc_limit(uma_zone_t zone,int count,int flags)4228 zone_alloc_limit(uma_zone_t zone, int count, int flags)
4229 {
4230 uint64_t old;
4231 uint64_t max;
4232
4233 max = zone->uz_max_items;
4234 MPASS(max > 0);
4235
4236 /*
4237 * We expect normal allocations to succeed with a simple
4238 * fetchadd.
4239 */
4240 old = atomic_fetchadd_64(&zone->uz_items, count);
4241 if (__predict_true(old + count <= max))
4242 return (count);
4243
4244 /*
4245 * If we had some items and no sleepers just return the
4246 * truncated value. We have to release the excess space
4247 * though because that may wake sleepers who weren't woken
4248 * because we were temporarily over the limit.
4249 */
4250 if (old < max) {
4251 zone_free_limit(zone, (old + count) - max);
4252 return (max - old);
4253 }
4254 return (zone_alloc_limit_hard(zone, count, flags));
4255 }
4256
4257 /*
4258 * Free a number of items back to the limit.
4259 */
4260 static void
zone_free_limit(uma_zone_t zone,int count)4261 zone_free_limit(uma_zone_t zone, int count)
4262 {
4263 uint64_t old;
4264
4265 MPASS(count > 0);
4266
4267 /*
4268 * In the common case we either have no sleepers or
4269 * are still over the limit and can just return.
4270 */
4271 old = atomic_fetchadd_64(&zone->uz_items, -count);
4272 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
4273 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
4274 return;
4275
4276 /*
4277 * Moderate the rate of wakeups. Sleepers will continue
4278 * to generate wakeups if necessary.
4279 */
4280 wakeup_one(&zone->uz_max_items);
4281 }
4282
4283 static uma_bucket_t
zone_alloc_bucket(uma_zone_t zone,void * udata,int domain,int flags)4284 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
4285 {
4286 uma_bucket_t bucket;
4287 int error, maxbucket, cnt;
4288
4289 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
4290 zone, domain);
4291
4292 /* Avoid allocs targeting empty domains. */
4293 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4294 domain = UMA_ANYDOMAIN;
4295 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4296 domain = UMA_ANYDOMAIN;
4297
4298 if (zone->uz_max_items > 0)
4299 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
4300 M_NOWAIT);
4301 else
4302 maxbucket = zone->uz_bucket_size;
4303 if (maxbucket == 0)
4304 return (NULL);
4305
4306 /* Don't wait for buckets, preserve caller's NOVM setting. */
4307 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
4308 if (bucket == NULL) {
4309 cnt = 0;
4310 goto out;
4311 }
4312
4313 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
4314 MIN(maxbucket, bucket->ub_entries), domain, flags);
4315
4316 /*
4317 * Initialize the memory if necessary.
4318 */
4319 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4320 int i;
4321
4322 for (i = 0; i < bucket->ub_cnt; i++) {
4323 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
4324 error = zone->uz_init(bucket->ub_bucket[i],
4325 zone->uz_size, flags);
4326 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
4327 if (error != 0)
4328 break;
4329 }
4330
4331 /*
4332 * If we couldn't initialize the whole bucket, put the
4333 * rest back onto the freelist.
4334 */
4335 if (i != bucket->ub_cnt) {
4336 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4337 bucket->ub_cnt - i);
4338 #ifdef INVARIANTS
4339 bzero(&bucket->ub_bucket[i],
4340 sizeof(void *) * (bucket->ub_cnt - i));
4341 #endif
4342 bucket->ub_cnt = i;
4343 }
4344 }
4345
4346 cnt = bucket->ub_cnt;
4347 if (bucket->ub_cnt == 0) {
4348 bucket_free(zone, bucket, udata);
4349 counter_u64_add(zone->uz_fails, 1);
4350 bucket = NULL;
4351 }
4352 out:
4353 if (zone->uz_max_items > 0 && cnt < maxbucket)
4354 zone_free_limit(zone, maxbucket - cnt);
4355
4356 return (bucket);
4357 }
4358
4359 /*
4360 * Allocates a single item from a zone.
4361 *
4362 * Arguments
4363 * zone The zone to alloc for.
4364 * udata The data to be passed to the constructor.
4365 * domain The domain to allocate from or UMA_ANYDOMAIN.
4366 * flags M_WAITOK, M_NOWAIT, M_ZERO.
4367 *
4368 * Returns
4369 * NULL if there is no memory and M_NOWAIT is set
4370 * An item if successful
4371 */
4372
4373 static void *
zone_alloc_item(uma_zone_t zone,void * udata,int domain,int flags)4374 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4375 {
4376 void *item;
4377
4378 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4379 counter_u64_add(zone->uz_fails, 1);
4380 return (NULL);
4381 }
4382
4383 /* Avoid allocs targeting empty domains. */
4384 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4385 domain = UMA_ANYDOMAIN;
4386
4387 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4388 goto fail_cnt;
4389
4390 /*
4391 * We have to call both the zone's init (not the keg's init)
4392 * and the zone's ctor. This is because the item is going from
4393 * a keg slab directly to the user, and the user is expecting it
4394 * to be both zone-init'd as well as zone-ctor'd.
4395 */
4396 if (zone->uz_init != NULL) {
4397 int error;
4398
4399 kasan_mark_item_valid(zone, item);
4400 error = zone->uz_init(item, zone->uz_size, flags);
4401 kasan_mark_item_invalid(zone, item);
4402 if (error != 0) {
4403 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4404 goto fail_cnt;
4405 }
4406 }
4407 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4408 item);
4409 if (item == NULL)
4410 goto fail;
4411
4412 counter_u64_add(zone->uz_allocs, 1);
4413 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4414 zone->uz_name, zone);
4415
4416 return (item);
4417
4418 fail_cnt:
4419 counter_u64_add(zone->uz_fails, 1);
4420 fail:
4421 if (zone->uz_max_items > 0)
4422 zone_free_limit(zone, 1);
4423 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4424 zone->uz_name, zone);
4425
4426 return (NULL);
4427 }
4428
4429 /* See uma.h */
4430 void
uma_zfree_smr(uma_zone_t zone,void * item)4431 uma_zfree_smr(uma_zone_t zone, void *item)
4432 {
4433 uma_cache_t cache;
4434 uma_cache_bucket_t bucket;
4435 int itemdomain;
4436 #ifdef NUMA
4437 int uz_flags;
4438 #endif
4439
4440 CTR3(KTR_UMA, "uma_zfree_smr zone %s(%p) item %p",
4441 zone->uz_name, zone, item);
4442
4443 #ifdef UMA_ZALLOC_DEBUG
4444 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4445 ("uma_zfree_smr: called with non-SMR zone."));
4446 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4447 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4448 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4449 return;
4450 #endif
4451 cache = &zone->uz_cpu[curcpu];
4452 itemdomain = 0;
4453 #ifdef NUMA
4454 uz_flags = cache_uz_flags(cache);
4455 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4456 itemdomain = item_domain(item);
4457 #endif
4458 critical_enter();
4459 do {
4460 cache = &zone->uz_cpu[curcpu];
4461 /* SMR Zones must free to the free bucket. */
4462 bucket = &cache->uc_freebucket;
4463 #ifdef NUMA
4464 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4465 PCPU_GET(domain) != itemdomain) {
4466 bucket = &cache->uc_crossbucket;
4467 }
4468 #endif
4469 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4470 cache_bucket_push(cache, bucket, item);
4471 critical_exit();
4472 return;
4473 }
4474 } while (cache_free(zone, cache, NULL, itemdomain));
4475 critical_exit();
4476
4477 /*
4478 * If nothing else caught this, we'll just do an internal free.
4479 */
4480 zone_free_item(zone, item, NULL, SKIP_NONE);
4481 }
4482
4483 /* See uma.h */
4484 void
uma_zfree_arg(uma_zone_t zone,void * item,void * udata)4485 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4486 {
4487 uma_cache_t cache;
4488 uma_cache_bucket_t bucket;
4489 int itemdomain, uz_flags;
4490
4491 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4492 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4493
4494 CTR3(KTR_UMA, "uma_zfree_arg zone %s(%p) item %p",
4495 zone->uz_name, zone, item);
4496
4497 #ifdef UMA_ZALLOC_DEBUG
4498 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4499 ("uma_zfree_arg: called with SMR zone."));
4500 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4501 return;
4502 #endif
4503 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4504 if (item == NULL)
4505 return;
4506
4507 /*
4508 * We are accessing the per-cpu cache without a critical section to
4509 * fetch size and flags. This is acceptable, if we are preempted we
4510 * will simply read another cpu's line.
4511 */
4512 cache = &zone->uz_cpu[curcpu];
4513 uz_flags = cache_uz_flags(cache);
4514 if (UMA_ALWAYS_CTORDTOR ||
4515 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4516 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4517
4518 /*
4519 * The race here is acceptable. If we miss it we'll just have to wait
4520 * a little longer for the limits to be reset.
4521 */
4522 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4523 if (atomic_load_32(&zone->uz_sleepers) > 0)
4524 goto zfree_item;
4525 }
4526
4527 /*
4528 * If possible, free to the per-CPU cache. There are two
4529 * requirements for safe access to the per-CPU cache: (1) the thread
4530 * accessing the cache must not be preempted or yield during access,
4531 * and (2) the thread must not migrate CPUs without switching which
4532 * cache it accesses. We rely on a critical section to prevent
4533 * preemption and migration. We release the critical section in
4534 * order to acquire the zone mutex if we are unable to free to the
4535 * current cache; when we re-acquire the critical section, we must
4536 * detect and handle migration if it has occurred.
4537 */
4538 itemdomain = 0;
4539 #ifdef NUMA
4540 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4541 itemdomain = item_domain(item);
4542 #endif
4543 critical_enter();
4544 do {
4545 cache = &zone->uz_cpu[curcpu];
4546 /*
4547 * Try to free into the allocbucket first to give LIFO
4548 * ordering for cache-hot datastructures. Spill over
4549 * into the freebucket if necessary. Alloc will swap
4550 * them if one runs dry.
4551 */
4552 bucket = &cache->uc_allocbucket;
4553 #ifdef NUMA
4554 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4555 PCPU_GET(domain) != itemdomain) {
4556 bucket = &cache->uc_crossbucket;
4557 } else
4558 #endif
4559 if (bucket->ucb_cnt == bucket->ucb_entries &&
4560 cache->uc_freebucket.ucb_cnt <
4561 cache->uc_freebucket.ucb_entries)
4562 cache_bucket_swap(&cache->uc_freebucket,
4563 &cache->uc_allocbucket);
4564 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4565 cache_bucket_push(cache, bucket, item);
4566 critical_exit();
4567 return;
4568 }
4569 } while (cache_free(zone, cache, udata, itemdomain));
4570 critical_exit();
4571
4572 /*
4573 * If nothing else caught this, we'll just do an internal free.
4574 */
4575 zfree_item:
4576 zone_free_item(zone, item, udata, SKIP_DTOR);
4577 }
4578
4579 #ifdef NUMA
4580 /*
4581 * sort crossdomain free buckets to domain correct buckets and cache
4582 * them.
4583 */
4584 static void
zone_free_cross(uma_zone_t zone,uma_bucket_t bucket,void * udata)4585 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4586 {
4587 struct uma_bucketlist emptybuckets, fullbuckets;
4588 uma_zone_domain_t zdom;
4589 uma_bucket_t b;
4590 smr_seq_t seq;
4591 void *item;
4592 int domain;
4593
4594 CTR3(KTR_UMA,
4595 "uma_zfree: zone %s(%p) draining cross bucket %p",
4596 zone->uz_name, zone, bucket);
4597
4598 /*
4599 * It is possible for buckets to arrive here out of order so we fetch
4600 * the current smr seq rather than accepting the bucket's.
4601 */
4602 seq = SMR_SEQ_INVALID;
4603 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4604 seq = smr_advance(zone->uz_smr);
4605
4606 /*
4607 * To avoid having ndomain * ndomain buckets for sorting we have a
4608 * lock on the current crossfree bucket. A full matrix with
4609 * per-domain locking could be used if necessary.
4610 */
4611 STAILQ_INIT(&emptybuckets);
4612 STAILQ_INIT(&fullbuckets);
4613 ZONE_CROSS_LOCK(zone);
4614 for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4615 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4616 domain = item_domain(item);
4617 zdom = ZDOM_GET(zone, domain);
4618 if (zdom->uzd_cross == NULL) {
4619 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4620 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4621 zdom->uzd_cross = b;
4622 } else {
4623 /*
4624 * Avoid allocating a bucket with the cross lock
4625 * held, since allocation can trigger a
4626 * cross-domain free and bucket zones may
4627 * allocate from each other.
4628 */
4629 ZONE_CROSS_UNLOCK(zone);
4630 b = bucket_alloc(zone, udata, M_NOWAIT);
4631 if (b == NULL)
4632 goto out;
4633 ZONE_CROSS_LOCK(zone);
4634 if (zdom->uzd_cross != NULL) {
4635 STAILQ_INSERT_HEAD(&emptybuckets, b,
4636 ub_link);
4637 } else {
4638 zdom->uzd_cross = b;
4639 }
4640 }
4641 }
4642 b = zdom->uzd_cross;
4643 b->ub_bucket[b->ub_cnt++] = item;
4644 b->ub_seq = seq;
4645 if (b->ub_cnt == b->ub_entries) {
4646 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4647 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4648 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4649 zdom->uzd_cross = b;
4650 }
4651 }
4652 ZONE_CROSS_UNLOCK(zone);
4653 out:
4654 if (bucket->ub_cnt == 0)
4655 bucket->ub_seq = SMR_SEQ_INVALID;
4656 bucket_free(zone, bucket, udata);
4657
4658 while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4659 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4660 bucket_free(zone, b, udata);
4661 }
4662 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4663 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4664 domain = item_domain(b->ub_bucket[0]);
4665 zone_put_bucket(zone, domain, b, udata, true);
4666 }
4667 }
4668 #endif
4669
4670 static void
zone_free_bucket(uma_zone_t zone,uma_bucket_t bucket,void * udata,int itemdomain,bool ws)4671 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4672 int itemdomain, bool ws)
4673 {
4674
4675 #ifdef NUMA
4676 /*
4677 * Buckets coming from the wrong domain will be entirely for the
4678 * only other domain on two domain systems. In this case we can
4679 * simply cache them. Otherwise we need to sort them back to
4680 * correct domains.
4681 */
4682 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4683 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4684 zone_free_cross(zone, bucket, udata);
4685 return;
4686 }
4687 #endif
4688
4689 /*
4690 * Attempt to save the bucket in the zone's domain bucket cache.
4691 */
4692 CTR3(KTR_UMA,
4693 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4694 zone->uz_name, zone, bucket);
4695 /* ub_cnt is pointing to the last free item */
4696 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4697 itemdomain = zone_domain_lowest(zone, itemdomain);
4698 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4699 }
4700
4701 /*
4702 * Populate a free or cross bucket for the current cpu cache. Free any
4703 * existing full bucket either to the zone cache or back to the slab layer.
4704 *
4705 * Enters and returns in a critical section. false return indicates that
4706 * we can not satisfy this free in the cache layer. true indicates that
4707 * the caller should retry.
4708 */
4709 static __noinline bool
cache_free(uma_zone_t zone,uma_cache_t cache,void * udata,int itemdomain)4710 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, int itemdomain)
4711 {
4712 uma_cache_bucket_t cbucket;
4713 uma_bucket_t newbucket, bucket;
4714
4715 CRITICAL_ASSERT(curthread);
4716
4717 if (zone->uz_bucket_size == 0)
4718 return false;
4719
4720 cache = &zone->uz_cpu[curcpu];
4721 newbucket = NULL;
4722
4723 /*
4724 * FIRSTTOUCH domains need to free to the correct zdom. When
4725 * enabled this is the zdom of the item. The bucket is the
4726 * cross bucket if the current domain and itemdomain do not match.
4727 */
4728 cbucket = &cache->uc_freebucket;
4729 #ifdef NUMA
4730 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4731 if (PCPU_GET(domain) != itemdomain) {
4732 cbucket = &cache->uc_crossbucket;
4733 if (cbucket->ucb_cnt != 0)
4734 counter_u64_add(zone->uz_xdomain,
4735 cbucket->ucb_cnt);
4736 }
4737 }
4738 #endif
4739 bucket = cache_bucket_unload(cbucket);
4740 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4741 ("cache_free: Entered with non-full free bucket."));
4742
4743 /* We are no longer associated with this CPU. */
4744 critical_exit();
4745
4746 /*
4747 * Don't let SMR zones operate without a free bucket. Force
4748 * a synchronize and re-use this one. We will only degrade
4749 * to a synchronize every bucket_size items rather than every
4750 * item if we fail to allocate a bucket.
4751 */
4752 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4753 if (bucket != NULL)
4754 bucket->ub_seq = smr_advance(zone->uz_smr);
4755 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4756 if (newbucket == NULL && bucket != NULL) {
4757 bucket_drain(zone, bucket);
4758 newbucket = bucket;
4759 bucket = NULL;
4760 }
4761 } else if (!bucketdisable)
4762 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4763
4764 if (bucket != NULL)
4765 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4766
4767 critical_enter();
4768 if ((bucket = newbucket) == NULL)
4769 return (false);
4770 cache = &zone->uz_cpu[curcpu];
4771 #ifdef NUMA
4772 /*
4773 * Check to see if we should be populating the cross bucket. If it
4774 * is already populated we will fall through and attempt to populate
4775 * the free bucket.
4776 */
4777 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4778 if (PCPU_GET(domain) != itemdomain &&
4779 cache->uc_crossbucket.ucb_bucket == NULL) {
4780 cache_bucket_load_cross(cache, bucket);
4781 return (true);
4782 }
4783 }
4784 #endif
4785 /*
4786 * We may have lost the race to fill the bucket or switched CPUs.
4787 */
4788 if (cache->uc_freebucket.ucb_bucket != NULL) {
4789 critical_exit();
4790 bucket_free(zone, bucket, udata);
4791 critical_enter();
4792 } else
4793 cache_bucket_load_free(cache, bucket);
4794
4795 return (true);
4796 }
4797
4798 static void
slab_free_item(uma_zone_t zone,uma_slab_t slab,void * item)4799 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4800 {
4801 uma_keg_t keg;
4802 uma_domain_t dom;
4803 int freei;
4804
4805 keg = zone->uz_keg;
4806 KEG_LOCK_ASSERT(keg, slab->us_domain);
4807
4808 /* Do we need to remove from any lists? */
4809 dom = &keg->uk_domain[slab->us_domain];
4810 if (slab->us_freecount + 1 == keg->uk_ipers) {
4811 LIST_REMOVE(slab, us_link);
4812 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4813 dom->ud_free_slabs++;
4814 } else if (slab->us_freecount == 0) {
4815 LIST_REMOVE(slab, us_link);
4816 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4817 }
4818
4819 /* Slab management. */
4820 freei = slab_item_index(slab, keg, item);
4821 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4822 slab->us_freecount++;
4823
4824 /* Keg statistics. */
4825 dom->ud_free_items++;
4826 }
4827
4828 static void
zone_release(void * arg,void ** bucket,int cnt)4829 zone_release(void *arg, void **bucket, int cnt)
4830 {
4831 struct mtx *lock;
4832 uma_zone_t zone;
4833 uma_slab_t slab;
4834 uma_keg_t keg;
4835 uint8_t *mem;
4836 void *item;
4837 int i;
4838
4839 zone = arg;
4840 keg = zone->uz_keg;
4841 lock = NULL;
4842 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4843 lock = KEG_LOCK(keg, 0);
4844 for (i = 0; i < cnt; i++) {
4845 item = bucket[i];
4846 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4847 slab = vtoslab((vm_offset_t)item);
4848 } else {
4849 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4850 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4851 slab = hash_sfind(&keg->uk_hash, mem);
4852 else
4853 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4854 }
4855 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4856 if (lock != NULL)
4857 mtx_unlock(lock);
4858 lock = KEG_LOCK(keg, slab->us_domain);
4859 }
4860 slab_free_item(zone, slab, item);
4861 }
4862 if (lock != NULL)
4863 mtx_unlock(lock);
4864 }
4865
4866 /*
4867 * Frees a single item to any zone.
4868 *
4869 * Arguments:
4870 * zone The zone to free to
4871 * item The item we're freeing
4872 * udata User supplied data for the dtor
4873 * skip Skip dtors and finis
4874 */
4875 static __noinline void
zone_free_item(uma_zone_t zone,void * item,void * udata,enum zfreeskip skip)4876 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4877 {
4878
4879 /*
4880 * If a free is sent directly to an SMR zone we have to
4881 * synchronize immediately because the item can instantly
4882 * be reallocated. This should only happen in degenerate
4883 * cases when no memory is available for per-cpu caches.
4884 */
4885 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4886 smr_synchronize(zone->uz_smr);
4887
4888 item_dtor(zone, item, zone->uz_size, udata, skip);
4889
4890 if (skip < SKIP_FINI && zone->uz_fini) {
4891 kasan_mark_item_valid(zone, item);
4892 zone->uz_fini(item, zone->uz_size);
4893 kasan_mark_item_invalid(zone, item);
4894 }
4895
4896 zone->uz_release(zone->uz_arg, &item, 1);
4897
4898 if (skip & SKIP_CNT)
4899 return;
4900
4901 counter_u64_add(zone->uz_frees, 1);
4902
4903 if (zone->uz_max_items > 0)
4904 zone_free_limit(zone, 1);
4905 }
4906
4907 /* See uma.h */
4908 int
uma_zone_set_max(uma_zone_t zone,int nitems)4909 uma_zone_set_max(uma_zone_t zone, int nitems)
4910 {
4911
4912 /*
4913 * If the limit is small, we may need to constrain the maximum per-CPU
4914 * cache size, or disable caching entirely.
4915 */
4916 uma_zone_set_maxcache(zone, nitems);
4917
4918 /*
4919 * XXX This can misbehave if the zone has any allocations with
4920 * no limit and a limit is imposed. There is currently no
4921 * way to clear a limit.
4922 */
4923 ZONE_LOCK(zone);
4924 if (zone->uz_max_items == 0)
4925 ZONE_ASSERT_COLD(zone);
4926 zone->uz_max_items = nitems;
4927 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4928 zone_update_caches(zone);
4929 /* We may need to wake waiters. */
4930 wakeup(&zone->uz_max_items);
4931 ZONE_UNLOCK(zone);
4932
4933 return (nitems);
4934 }
4935
4936 /* See uma.h */
4937 void
uma_zone_set_maxcache(uma_zone_t zone,int nitems)4938 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4939 {
4940 int bpcpu, bpdom, bsize, nb;
4941
4942 ZONE_LOCK(zone);
4943
4944 /*
4945 * Compute a lower bound on the number of items that may be cached in
4946 * the zone. Each CPU gets at least two buckets, and for cross-domain
4947 * frees we use an additional bucket per CPU and per domain. Select the
4948 * largest bucket size that does not exceed half of the requested limit,
4949 * with the left over space given to the full bucket cache.
4950 */
4951 bpdom = 0;
4952 bpcpu = 2;
4953 #ifdef NUMA
4954 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
4955 bpcpu++;
4956 bpdom++;
4957 }
4958 #endif
4959 nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
4960 bsize = nitems / nb / 2;
4961 if (bsize > BUCKET_MAX)
4962 bsize = BUCKET_MAX;
4963 else if (bsize == 0 && nitems / nb > 0)
4964 bsize = 1;
4965 zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
4966 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4967 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4968 zone->uz_bucket_max = nitems - nb * bsize;
4969 ZONE_UNLOCK(zone);
4970 }
4971
4972 /* See uma.h */
4973 int
uma_zone_get_max(uma_zone_t zone)4974 uma_zone_get_max(uma_zone_t zone)
4975 {
4976 int nitems;
4977
4978 nitems = atomic_load_64(&zone->uz_max_items);
4979
4980 return (nitems);
4981 }
4982
4983 /* See uma.h */
4984 void
uma_zone_set_warning(uma_zone_t zone,const char * warning)4985 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4986 {
4987
4988 ZONE_ASSERT_COLD(zone);
4989 zone->uz_warning = warning;
4990 }
4991
4992 /* See uma.h */
4993 void
uma_zone_set_maxaction(uma_zone_t zone,uma_maxaction_t maxaction)4994 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4995 {
4996
4997 ZONE_ASSERT_COLD(zone);
4998 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4999 }
5000
5001 /* See uma.h */
5002 int
uma_zone_get_cur(uma_zone_t zone)5003 uma_zone_get_cur(uma_zone_t zone)
5004 {
5005 int64_t nitems;
5006 u_int i;
5007
5008 nitems = 0;
5009 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
5010 nitems = counter_u64_fetch(zone->uz_allocs) -
5011 counter_u64_fetch(zone->uz_frees);
5012 CPU_FOREACH(i)
5013 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
5014 atomic_load_64(&zone->uz_cpu[i].uc_frees);
5015
5016 return (nitems < 0 ? 0 : nitems);
5017 }
5018
5019 static uint64_t
uma_zone_get_allocs(uma_zone_t zone)5020 uma_zone_get_allocs(uma_zone_t zone)
5021 {
5022 uint64_t nitems;
5023 u_int i;
5024
5025 nitems = 0;
5026 if (zone->uz_allocs != EARLY_COUNTER)
5027 nitems = counter_u64_fetch(zone->uz_allocs);
5028 CPU_FOREACH(i)
5029 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
5030
5031 return (nitems);
5032 }
5033
5034 static uint64_t
uma_zone_get_frees(uma_zone_t zone)5035 uma_zone_get_frees(uma_zone_t zone)
5036 {
5037 uint64_t nitems;
5038 u_int i;
5039
5040 nitems = 0;
5041 if (zone->uz_frees != EARLY_COUNTER)
5042 nitems = counter_u64_fetch(zone->uz_frees);
5043 CPU_FOREACH(i)
5044 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
5045
5046 return (nitems);
5047 }
5048
5049 #ifdef INVARIANTS
5050 /* Used only for KEG_ASSERT_COLD(). */
5051 static uint64_t
uma_keg_get_allocs(uma_keg_t keg)5052 uma_keg_get_allocs(uma_keg_t keg)
5053 {
5054 uma_zone_t z;
5055 uint64_t nitems;
5056
5057 nitems = 0;
5058 LIST_FOREACH(z, &keg->uk_zones, uz_link)
5059 nitems += uma_zone_get_allocs(z);
5060
5061 return (nitems);
5062 }
5063 #endif
5064
5065 /* See uma.h */
5066 void
uma_zone_set_init(uma_zone_t zone,uma_init uminit)5067 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
5068 {
5069 uma_keg_t keg;
5070
5071 KEG_GET(zone, keg);
5072 KEG_ASSERT_COLD(keg);
5073 keg->uk_init = uminit;
5074 }
5075
5076 /* See uma.h */
5077 void
uma_zone_set_fini(uma_zone_t zone,uma_fini fini)5078 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
5079 {
5080 uma_keg_t keg;
5081
5082 KEG_GET(zone, keg);
5083 KEG_ASSERT_COLD(keg);
5084 keg->uk_fini = fini;
5085 }
5086
5087 /* See uma.h */
5088 void
uma_zone_set_zinit(uma_zone_t zone,uma_init zinit)5089 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
5090 {
5091
5092 ZONE_ASSERT_COLD(zone);
5093 zone->uz_init = zinit;
5094 }
5095
5096 /* See uma.h */
5097 void
uma_zone_set_zfini(uma_zone_t zone,uma_fini zfini)5098 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
5099 {
5100
5101 ZONE_ASSERT_COLD(zone);
5102 zone->uz_fini = zfini;
5103 }
5104
5105 /* See uma.h */
5106 void
uma_zone_set_freef(uma_zone_t zone,uma_free freef)5107 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
5108 {
5109 uma_keg_t keg;
5110
5111 KEG_GET(zone, keg);
5112 KEG_ASSERT_COLD(keg);
5113 keg->uk_freef = freef;
5114 }
5115
5116 /* See uma.h */
5117 void
uma_zone_set_allocf(uma_zone_t zone,uma_alloc allocf)5118 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
5119 {
5120 uma_keg_t keg;
5121
5122 KEG_GET(zone, keg);
5123 KEG_ASSERT_COLD(keg);
5124 keg->uk_allocf = allocf;
5125 }
5126
5127 /* See uma.h */
5128 void
uma_zone_set_smr(uma_zone_t zone,smr_t smr)5129 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
5130 {
5131
5132 ZONE_ASSERT_COLD(zone);
5133
5134 KASSERT(smr != NULL, ("Got NULL smr"));
5135 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
5136 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
5137 zone->uz_flags |= UMA_ZONE_SMR;
5138 zone->uz_smr = smr;
5139 zone_update_caches(zone);
5140 }
5141
5142 smr_t
uma_zone_get_smr(uma_zone_t zone)5143 uma_zone_get_smr(uma_zone_t zone)
5144 {
5145
5146 return (zone->uz_smr);
5147 }
5148
5149 /* See uma.h */
5150 void
uma_zone_reserve(uma_zone_t zone,int items)5151 uma_zone_reserve(uma_zone_t zone, int items)
5152 {
5153 uma_keg_t keg;
5154
5155 KEG_GET(zone, keg);
5156 KEG_ASSERT_COLD(keg);
5157 keg->uk_reserve = items;
5158 }
5159
5160 /* See uma.h */
5161 int
uma_zone_reserve_kva(uma_zone_t zone,int count)5162 uma_zone_reserve_kva(uma_zone_t zone, int count)
5163 {
5164 uma_keg_t keg;
5165 vm_offset_t kva;
5166 u_int pages;
5167
5168 KEG_GET(zone, keg);
5169 KEG_ASSERT_COLD(keg);
5170 ZONE_ASSERT_COLD(zone);
5171
5172 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
5173
5174 #ifdef UMA_MD_SMALL_ALLOC
5175 if (keg->uk_ppera > 1) {
5176 #else
5177 if (1) {
5178 #endif
5179 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
5180 if (kva == 0)
5181 return (0);
5182 } else
5183 kva = 0;
5184
5185 MPASS(keg->uk_kva == 0);
5186 keg->uk_kva = kva;
5187 keg->uk_offset = 0;
5188 zone->uz_max_items = pages * keg->uk_ipers;
5189 #ifdef UMA_MD_SMALL_ALLOC
5190 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
5191 #else
5192 keg->uk_allocf = noobj_alloc;
5193 #endif
5194 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5195 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5196 zone_update_caches(zone);
5197
5198 return (1);
5199 }
5200
5201 /* See uma.h */
5202 void
5203 uma_prealloc(uma_zone_t zone, int items)
5204 {
5205 struct vm_domainset_iter di;
5206 uma_domain_t dom;
5207 uma_slab_t slab;
5208 uma_keg_t keg;
5209 int aflags, domain, slabs;
5210
5211 KEG_GET(zone, keg);
5212 slabs = howmany(items, keg->uk_ipers);
5213 while (slabs-- > 0) {
5214 aflags = M_NOWAIT;
5215 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
5216 &aflags);
5217 for (;;) {
5218 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
5219 aflags);
5220 if (slab != NULL) {
5221 dom = &keg->uk_domain[slab->us_domain];
5222 /*
5223 * keg_alloc_slab() always returns a slab on the
5224 * partial list.
5225 */
5226 LIST_REMOVE(slab, us_link);
5227 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
5228 us_link);
5229 dom->ud_free_slabs++;
5230 KEG_UNLOCK(keg, slab->us_domain);
5231 break;
5232 }
5233 if (vm_domainset_iter_policy(&di, &domain) != 0)
5234 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
5235 }
5236 }
5237 }
5238
5239 /*
5240 * Returns a snapshot of memory consumption in bytes.
5241 */
5242 size_t
5243 uma_zone_memory(uma_zone_t zone)
5244 {
5245 size_t sz;
5246 int i;
5247
5248 sz = 0;
5249 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
5250 for (i = 0; i < vm_ndomains; i++)
5251 sz += ZDOM_GET(zone, i)->uzd_nitems;
5252 return (sz * zone->uz_size);
5253 }
5254 for (i = 0; i < vm_ndomains; i++)
5255 sz += zone->uz_keg->uk_domain[i].ud_pages;
5256
5257 return (sz * PAGE_SIZE);
5258 }
5259
5260 struct uma_reclaim_args {
5261 int domain;
5262 int req;
5263 };
5264
5265 static void
5266 uma_reclaim_domain_cb(uma_zone_t zone, void *arg)
5267 {
5268 struct uma_reclaim_args *args;
5269
5270 args = arg;
5271 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0)
5272 uma_zone_reclaim_domain(zone, args->req, args->domain);
5273 }
5274
5275 /* See uma.h */
5276 void
5277 uma_reclaim(int req)
5278 {
5279 uma_reclaim_domain(req, UMA_ANYDOMAIN);
5280 }
5281
5282 void
5283 uma_reclaim_domain(int req, int domain)
5284 {
5285 struct uma_reclaim_args args;
5286
5287 bucket_enable();
5288
5289 args.domain = domain;
5290 args.req = req;
5291
5292 sx_slock(&uma_reclaim_lock);
5293 switch (req) {
5294 case UMA_RECLAIM_TRIM:
5295 case UMA_RECLAIM_DRAIN:
5296 zone_foreach(uma_reclaim_domain_cb, &args);
5297 break;
5298 case UMA_RECLAIM_DRAIN_CPU:
5299 zone_foreach(uma_reclaim_domain_cb, &args);
5300 pcpu_cache_drain_safe(NULL);
5301 zone_foreach(uma_reclaim_domain_cb, &args);
5302 break;
5303 default:
5304 panic("unhandled reclamation request %d", req);
5305 }
5306
5307 /*
5308 * Some slabs may have been freed but this zone will be visited early
5309 * we visit again so that we can free pages that are empty once other
5310 * zones are drained. We have to do the same for buckets.
5311 */
5312 uma_zone_reclaim_domain(slabzones[0], UMA_RECLAIM_DRAIN, domain);
5313 uma_zone_reclaim_domain(slabzones[1], UMA_RECLAIM_DRAIN, domain);
5314 bucket_zone_drain(domain);
5315 sx_sunlock(&uma_reclaim_lock);
5316 }
5317
5318 static volatile int uma_reclaim_needed;
5319
5320 void
5321 uma_reclaim_wakeup(void)
5322 {
5323
5324 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
5325 wakeup(uma_reclaim);
5326 }
5327
5328 void
5329 uma_reclaim_worker(void *arg __unused)
5330 {
5331
5332 for (;;) {
5333 sx_xlock(&uma_reclaim_lock);
5334 while (atomic_load_int(&uma_reclaim_needed) == 0)
5335 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
5336 hz);
5337 sx_xunlock(&uma_reclaim_lock);
5338 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
5339 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
5340 atomic_store_int(&uma_reclaim_needed, 0);
5341 /* Don't fire more than once per-second. */
5342 pause("umarclslp", hz);
5343 }
5344 }
5345
5346 /* See uma.h */
5347 void
5348 uma_zone_reclaim(uma_zone_t zone, int req)
5349 {
5350 uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN);
5351 }
5352
5353 void
5354 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain)
5355 {
5356 switch (req) {
5357 case UMA_RECLAIM_TRIM:
5358 zone_reclaim(zone, domain, M_NOWAIT, false);
5359 break;
5360 case UMA_RECLAIM_DRAIN:
5361 zone_reclaim(zone, domain, M_NOWAIT, true);
5362 break;
5363 case UMA_RECLAIM_DRAIN_CPU:
5364 pcpu_cache_drain_safe(zone);
5365 zone_reclaim(zone, domain, M_NOWAIT, true);
5366 break;
5367 default:
5368 panic("unhandled reclamation request %d", req);
5369 }
5370 }
5371
5372 /* See uma.h */
5373 int
5374 uma_zone_exhausted(uma_zone_t zone)
5375 {
5376
5377 return (atomic_load_32(&zone->uz_sleepers) > 0);
5378 }
5379
5380 unsigned long
5381 uma_limit(void)
5382 {
5383
5384 return (uma_kmem_limit);
5385 }
5386
5387 void
5388 uma_set_limit(unsigned long limit)
5389 {
5390
5391 uma_kmem_limit = limit;
5392 }
5393
5394 unsigned long
5395 uma_size(void)
5396 {
5397
5398 return (atomic_load_long(&uma_kmem_total));
5399 }
5400
5401 long
5402 uma_avail(void)
5403 {
5404
5405 return (uma_kmem_limit - uma_size());
5406 }
5407
5408 #ifdef DDB
5409 /*
5410 * Generate statistics across both the zone and its per-cpu cache's. Return
5411 * desired statistics if the pointer is non-NULL for that statistic.
5412 *
5413 * Note: does not update the zone statistics, as it can't safely clear the
5414 * per-CPU cache statistic.
5415 *
5416 */
5417 static void
5418 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5419 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5420 {
5421 uma_cache_t cache;
5422 uint64_t allocs, frees, sleeps, xdomain;
5423 int cachefree, cpu;
5424
5425 allocs = frees = sleeps = xdomain = 0;
5426 cachefree = 0;
5427 CPU_FOREACH(cpu) {
5428 cache = &z->uz_cpu[cpu];
5429 cachefree += cache->uc_allocbucket.ucb_cnt;
5430 cachefree += cache->uc_freebucket.ucb_cnt;
5431 xdomain += cache->uc_crossbucket.ucb_cnt;
5432 cachefree += cache->uc_crossbucket.ucb_cnt;
5433 allocs += cache->uc_allocs;
5434 frees += cache->uc_frees;
5435 }
5436 allocs += counter_u64_fetch(z->uz_allocs);
5437 frees += counter_u64_fetch(z->uz_frees);
5438 xdomain += counter_u64_fetch(z->uz_xdomain);
5439 sleeps += z->uz_sleeps;
5440 if (cachefreep != NULL)
5441 *cachefreep = cachefree;
5442 if (allocsp != NULL)
5443 *allocsp = allocs;
5444 if (freesp != NULL)
5445 *freesp = frees;
5446 if (sleepsp != NULL)
5447 *sleepsp = sleeps;
5448 if (xdomainp != NULL)
5449 *xdomainp = xdomain;
5450 }
5451 #endif /* DDB */
5452
5453 static int
5454 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5455 {
5456 uma_keg_t kz;
5457 uma_zone_t z;
5458 int count;
5459
5460 count = 0;
5461 rw_rlock(&uma_rwlock);
5462 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5463 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5464 count++;
5465 }
5466 LIST_FOREACH(z, &uma_cachezones, uz_link)
5467 count++;
5468
5469 rw_runlock(&uma_rwlock);
5470 return (sysctl_handle_int(oidp, &count, 0, req));
5471 }
5472
5473 static void
5474 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5475 struct uma_percpu_stat *ups, bool internal)
5476 {
5477 uma_zone_domain_t zdom;
5478 uma_cache_t cache;
5479 int i;
5480
5481 for (i = 0; i < vm_ndomains; i++) {
5482 zdom = ZDOM_GET(z, i);
5483 uth->uth_zone_free += zdom->uzd_nitems;
5484 }
5485 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5486 uth->uth_frees = counter_u64_fetch(z->uz_frees);
5487 uth->uth_fails = counter_u64_fetch(z->uz_fails);
5488 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5489 uth->uth_sleeps = z->uz_sleeps;
5490
5491 for (i = 0; i < mp_maxid + 1; i++) {
5492 bzero(&ups[i], sizeof(*ups));
5493 if (internal || CPU_ABSENT(i))
5494 continue;
5495 cache = &z->uz_cpu[i];
5496 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5497 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5498 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5499 ups[i].ups_allocs = cache->uc_allocs;
5500 ups[i].ups_frees = cache->uc_frees;
5501 }
5502 }
5503
5504 static int
5505 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5506 {
5507 struct uma_stream_header ush;
5508 struct uma_type_header uth;
5509 struct uma_percpu_stat *ups;
5510 struct sbuf sbuf;
5511 uma_keg_t kz;
5512 uma_zone_t z;
5513 uint64_t items;
5514 uint32_t kfree, pages;
5515 int count, error, i;
5516
5517 error = sysctl_wire_old_buffer(req, 0);
5518 if (error != 0)
5519 return (error);
5520 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5521 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5522 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5523
5524 count = 0;
5525 rw_rlock(&uma_rwlock);
5526 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5527 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5528 count++;
5529 }
5530
5531 LIST_FOREACH(z, &uma_cachezones, uz_link)
5532 count++;
5533
5534 /*
5535 * Insert stream header.
5536 */
5537 bzero(&ush, sizeof(ush));
5538 ush.ush_version = UMA_STREAM_VERSION;
5539 ush.ush_maxcpus = (mp_maxid + 1);
5540 ush.ush_count = count;
5541 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5542
5543 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5544 kfree = pages = 0;
5545 for (i = 0; i < vm_ndomains; i++) {
5546 kfree += kz->uk_domain[i].ud_free_items;
5547 pages += kz->uk_domain[i].ud_pages;
5548 }
5549 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5550 bzero(&uth, sizeof(uth));
5551 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5552 uth.uth_align = kz->uk_align;
5553 uth.uth_size = kz->uk_size;
5554 uth.uth_rsize = kz->uk_rsize;
5555 if (z->uz_max_items > 0) {
5556 items = UZ_ITEMS_COUNT(z->uz_items);
5557 uth.uth_pages = (items / kz->uk_ipers) *
5558 kz->uk_ppera;
5559 } else
5560 uth.uth_pages = pages;
5561 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5562 kz->uk_ppera;
5563 uth.uth_limit = z->uz_max_items;
5564 uth.uth_keg_free = kfree;
5565
5566 /*
5567 * A zone is secondary is it is not the first entry
5568 * on the keg's zone list.
5569 */
5570 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5571 (LIST_FIRST(&kz->uk_zones) != z))
5572 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5573 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5574 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5575 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5576 for (i = 0; i < mp_maxid + 1; i++)
5577 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5578 }
5579 }
5580 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5581 bzero(&uth, sizeof(uth));
5582 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5583 uth.uth_size = z->uz_size;
5584 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5585 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5586 for (i = 0; i < mp_maxid + 1; i++)
5587 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5588 }
5589
5590 rw_runlock(&uma_rwlock);
5591 error = sbuf_finish(&sbuf);
5592 sbuf_delete(&sbuf);
5593 free(ups, M_TEMP);
5594 return (error);
5595 }
5596
5597 int
5598 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5599 {
5600 uma_zone_t zone = *(uma_zone_t *)arg1;
5601 int error, max;
5602
5603 max = uma_zone_get_max(zone);
5604 error = sysctl_handle_int(oidp, &max, 0, req);
5605 if (error || !req->newptr)
5606 return (error);
5607
5608 uma_zone_set_max(zone, max);
5609
5610 return (0);
5611 }
5612
5613 int
5614 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5615 {
5616 uma_zone_t zone;
5617 int cur;
5618
5619 /*
5620 * Some callers want to add sysctls for global zones that
5621 * may not yet exist so they pass a pointer to a pointer.
5622 */
5623 if (arg2 == 0)
5624 zone = *(uma_zone_t *)arg1;
5625 else
5626 zone = arg1;
5627 cur = uma_zone_get_cur(zone);
5628 return (sysctl_handle_int(oidp, &cur, 0, req));
5629 }
5630
5631 static int
5632 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5633 {
5634 uma_zone_t zone = arg1;
5635 uint64_t cur;
5636
5637 cur = uma_zone_get_allocs(zone);
5638 return (sysctl_handle_64(oidp, &cur, 0, req));
5639 }
5640
5641 static int
5642 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5643 {
5644 uma_zone_t zone = arg1;
5645 uint64_t cur;
5646
5647 cur = uma_zone_get_frees(zone);
5648 return (sysctl_handle_64(oidp, &cur, 0, req));
5649 }
5650
5651 static int
5652 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5653 {
5654 struct sbuf sbuf;
5655 uma_zone_t zone = arg1;
5656 int error;
5657
5658 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5659 if (zone->uz_flags != 0)
5660 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5661 else
5662 sbuf_printf(&sbuf, "0");
5663 error = sbuf_finish(&sbuf);
5664 sbuf_delete(&sbuf);
5665
5666 return (error);
5667 }
5668
5669 static int
5670 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5671 {
5672 uma_keg_t keg = arg1;
5673 int avail, effpct, total;
5674
5675 total = keg->uk_ppera * PAGE_SIZE;
5676 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5677 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5678 /*
5679 * We consider the client's requested size and alignment here, not the
5680 * real size determination uk_rsize, because we also adjust the real
5681 * size for internal implementation reasons (max bitset size).
5682 */
5683 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5684 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5685 avail *= mp_maxid + 1;
5686 effpct = 100 * avail / total;
5687 return (sysctl_handle_int(oidp, &effpct, 0, req));
5688 }
5689
5690 static int
5691 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5692 {
5693 uma_zone_t zone = arg1;
5694 uint64_t cur;
5695
5696 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5697 return (sysctl_handle_64(oidp, &cur, 0, req));
5698 }
5699
5700 #ifdef INVARIANTS
5701 static uma_slab_t
5702 uma_dbg_getslab(uma_zone_t zone, void *item)
5703 {
5704 uma_slab_t slab;
5705 uma_keg_t keg;
5706 uint8_t *mem;
5707
5708 /*
5709 * It is safe to return the slab here even though the
5710 * zone is unlocked because the item's allocation state
5711 * essentially holds a reference.
5712 */
5713 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5714 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5715 return (NULL);
5716 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5717 return (vtoslab((vm_offset_t)mem));
5718 keg = zone->uz_keg;
5719 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5720 return ((uma_slab_t)(mem + keg->uk_pgoff));
5721 KEG_LOCK(keg, 0);
5722 slab = hash_sfind(&keg->uk_hash, mem);
5723 KEG_UNLOCK(keg, 0);
5724
5725 return (slab);
5726 }
5727
5728 static bool
5729 uma_dbg_zskip(uma_zone_t zone, void *mem)
5730 {
5731
5732 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5733 return (true);
5734
5735 return (uma_dbg_kskip(zone->uz_keg, mem));
5736 }
5737
5738 static bool
5739 uma_dbg_kskip(uma_keg_t keg, void *mem)
5740 {
5741 uintptr_t idx;
5742
5743 if (dbg_divisor == 0)
5744 return (true);
5745
5746 if (dbg_divisor == 1)
5747 return (false);
5748
5749 idx = (uintptr_t)mem >> PAGE_SHIFT;
5750 if (keg->uk_ipers > 1) {
5751 idx *= keg->uk_ipers;
5752 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5753 }
5754
5755 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5756 counter_u64_add(uma_skip_cnt, 1);
5757 return (true);
5758 }
5759 counter_u64_add(uma_dbg_cnt, 1);
5760
5761 return (false);
5762 }
5763
5764 /*
5765 * Set up the slab's freei data such that uma_dbg_free can function.
5766 *
5767 */
5768 static void
5769 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5770 {
5771 uma_keg_t keg;
5772 int freei;
5773
5774 if (slab == NULL) {
5775 slab = uma_dbg_getslab(zone, item);
5776 if (slab == NULL)
5777 panic("uma: item %p did not belong to zone %s",
5778 item, zone->uz_name);
5779 }
5780 keg = zone->uz_keg;
5781 freei = slab_item_index(slab, keg, item);
5782
5783 if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5784 slab_dbg_bits(slab, keg)))
5785 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5786 item, zone, zone->uz_name, slab, freei);
5787 }
5788
5789 /*
5790 * Verifies freed addresses. Checks for alignment, valid slab membership
5791 * and duplicate frees.
5792 *
5793 */
5794 static void
5795 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5796 {
5797 uma_keg_t keg;
5798 int freei;
5799
5800 if (slab == NULL) {
5801 slab = uma_dbg_getslab(zone, item);
5802 if (slab == NULL)
5803 panic("uma: Freed item %p did not belong to zone %s",
5804 item, zone->uz_name);
5805 }
5806 keg = zone->uz_keg;
5807 freei = slab_item_index(slab, keg, item);
5808
5809 if (freei >= keg->uk_ipers)
5810 panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5811 item, zone, zone->uz_name, slab, freei);
5812
5813 if (slab_item(slab, keg, freei) != item)
5814 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5815 item, zone, zone->uz_name, slab, freei);
5816
5817 if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5818 slab_dbg_bits(slab, keg)))
5819 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5820 item, zone, zone->uz_name, slab, freei);
5821 }
5822 #endif /* INVARIANTS */
5823
5824 #ifdef DDB
5825 static int64_t
5826 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5827 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5828 {
5829 uint64_t frees;
5830 int i;
5831
5832 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5833 *allocs = counter_u64_fetch(z->uz_allocs);
5834 frees = counter_u64_fetch(z->uz_frees);
5835 *sleeps = z->uz_sleeps;
5836 *cachefree = 0;
5837 *xdomain = 0;
5838 } else
5839 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5840 xdomain);
5841 for (i = 0; i < vm_ndomains; i++) {
5842 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5843 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5844 (LIST_FIRST(&kz->uk_zones) != z)))
5845 *cachefree += kz->uk_domain[i].ud_free_items;
5846 }
5847 *used = *allocs - frees;
5848 return (((int64_t)*used + *cachefree) * kz->uk_size);
5849 }
5850
5851 DB_SHOW_COMMAND_FLAGS(uma, db_show_uma, DB_CMD_MEMSAFE)
5852 {
5853 const char *fmt_hdr, *fmt_entry;
5854 uma_keg_t kz;
5855 uma_zone_t z;
5856 uint64_t allocs, used, sleeps, xdomain;
5857 long cachefree;
5858 /* variables for sorting */
5859 uma_keg_t cur_keg;
5860 uma_zone_t cur_zone, last_zone;
5861 int64_t cur_size, last_size, size;
5862 int ties;
5863
5864 /* /i option produces machine-parseable CSV output */
5865 if (modif[0] == 'i') {
5866 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5867 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5868 } else {
5869 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5870 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5871 }
5872
5873 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5874 "Sleeps", "Bucket", "Total Mem", "XFree");
5875
5876 /* Sort the zones with largest size first. */
5877 last_zone = NULL;
5878 last_size = INT64_MAX;
5879 for (;;) {
5880 cur_zone = NULL;
5881 cur_size = -1;
5882 ties = 0;
5883 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5884 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5885 /*
5886 * In the case of size ties, print out zones
5887 * in the order they are encountered. That is,
5888 * when we encounter the most recently output
5889 * zone, we have already printed all preceding
5890 * ties, and we must print all following ties.
5891 */
5892 if (z == last_zone) {
5893 ties = 1;
5894 continue;
5895 }
5896 size = get_uma_stats(kz, z, &allocs, &used,
5897 &sleeps, &cachefree, &xdomain);
5898 if (size > cur_size && size < last_size + ties)
5899 {
5900 cur_size = size;
5901 cur_zone = z;
5902 cur_keg = kz;
5903 }
5904 }
5905 }
5906 if (cur_zone == NULL)
5907 break;
5908
5909 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5910 &sleeps, &cachefree, &xdomain);
5911 db_printf(fmt_entry, cur_zone->uz_name,
5912 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5913 (uintmax_t)allocs, (uintmax_t)sleeps,
5914 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5915 xdomain);
5916
5917 if (db_pager_quit)
5918 return;
5919 last_zone = cur_zone;
5920 last_size = cur_size;
5921 }
5922 }
5923
5924 DB_SHOW_COMMAND_FLAGS(umacache, db_show_umacache, DB_CMD_MEMSAFE)
5925 {
5926 uma_zone_t z;
5927 uint64_t allocs, frees;
5928 long cachefree;
5929 int i;
5930
5931 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5932 "Requests", "Bucket");
5933 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5934 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5935 for (i = 0; i < vm_ndomains; i++)
5936 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5937 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5938 z->uz_name, (uintmax_t)z->uz_size,
5939 (intmax_t)(allocs - frees), cachefree,
5940 (uintmax_t)allocs, z->uz_bucket_size);
5941 if (db_pager_quit)
5942 return;
5943 }
5944 }
5945 #endif /* DDB */
5946