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/sysctl.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/taskqueue.h>
81 #include <sys/vmmeter.h>
82
83 #include <vm/vm.h>
84 #include <vm/vm_param.h>
85 #include <vm/vm_domainset.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_phys.h>
90 #include <vm/vm_pagequeue.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_kern.h>
93 #include <vm/vm_extern.h>
94 #include <vm/vm_dumpset.h>
95 #include <vm/uma.h>
96 #include <vm/uma_int.h>
97 #include <vm/uma_dbg.h>
98
99 #include <ddb/ddb.h>
100
101 #ifdef DEBUG_MEMGUARD
102 #include <vm/memguard.h>
103 #endif
104
105 #include <machine/md_var.h>
106
107 #ifdef INVARIANTS
108 #define UMA_ALWAYS_CTORDTOR 1
109 #else
110 #define UMA_ALWAYS_CTORDTOR 0
111 #endif
112
113 /*
114 * This is the zone and keg from which all zones are spawned.
115 */
116 static uma_zone_t kegs;
117 static uma_zone_t zones;
118
119 /*
120 * On INVARIANTS builds, the slab contains a second bitset of the same size,
121 * "dbg_bits", which is laid out immediately after us_free.
122 */
123 #ifdef INVARIANTS
124 #define SLAB_BITSETS 2
125 #else
126 #define SLAB_BITSETS 1
127 #endif
128
129 /*
130 * These are the two zones from which all offpage uma_slab_ts are allocated.
131 *
132 * One zone is for slab headers that can represent a larger number of items,
133 * making the slabs themselves more efficient, and the other zone is for
134 * headers that are smaller and represent fewer items, making the headers more
135 * efficient.
136 */
137 #define SLABZONE_SIZE(setsize) \
138 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
139 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16)
140 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE
141 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE)
142 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE)
143 static uma_zone_t slabzones[2];
144
145 /*
146 * The initial hash tables come out of this zone so they can be allocated
147 * prior to malloc coming up.
148 */
149 static uma_zone_t hashzone;
150
151 /* The boot-time adjusted value for cache line alignment. */
152 static unsigned int uma_cache_align_mask = 64 - 1;
153
154 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
155 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
156
157 /*
158 * Are we allowed to allocate buckets?
159 */
160 static int bucketdisable = 1;
161
162 /* Linked list of all kegs in the system */
163 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
164
165 /* Linked list of all cache-only zones in the system */
166 static LIST_HEAD(,uma_zone) uma_cachezones =
167 LIST_HEAD_INITIALIZER(uma_cachezones);
168
169 /*
170 * Mutex for global lists: uma_kegs, uma_cachezones, and the per-keg list of
171 * zones.
172 */
173 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
174
175 static struct sx uma_reclaim_lock;
176
177 /*
178 * First available virual address for boot time allocations.
179 */
180 static vm_offset_t bootstart;
181 static vm_offset_t bootmem;
182
183 /*
184 * kmem soft limit, initialized by uma_set_limit(). Ensure that early
185 * allocations don't trigger a wakeup of the reclaim thread.
186 */
187 unsigned long uma_kmem_limit = LONG_MAX;
188 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
189 "UMA kernel memory soft limit");
190 unsigned long uma_kmem_total;
191 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
192 "UMA kernel memory usage");
193
194 /* Is the VM done starting up? */
195 static enum {
196 BOOT_COLD,
197 BOOT_KVA,
198 BOOT_PCPU,
199 BOOT_RUNNING,
200 BOOT_SHUTDOWN,
201 } booted = BOOT_COLD;
202
203 /*
204 * This is the handle used to schedule events that need to happen
205 * outside of the allocation fast path.
206 */
207 static struct timeout_task uma_timeout_task;
208 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
209
210 /*
211 * This structure is passed as the zone ctor arg so that I don't have to create
212 * a special allocation function just for zones.
213 */
214 struct uma_zctor_args {
215 const char *name;
216 size_t size;
217 uma_ctor ctor;
218 uma_dtor dtor;
219 uma_init uminit;
220 uma_fini fini;
221 uma_import import;
222 uma_release release;
223 void *arg;
224 uma_keg_t keg;
225 int align;
226 uint32_t flags;
227 };
228
229 struct uma_kctor_args {
230 uma_zone_t zone;
231 size_t size;
232 uma_init uminit;
233 uma_fini fini;
234 int align;
235 uint32_t flags;
236 };
237
238 struct uma_bucket_zone {
239 uma_zone_t ubz_zone;
240 const char *ubz_name;
241 int ubz_entries; /* Number of items it can hold. */
242 int ubz_maxsize; /* Maximum allocation size per-item. */
243 };
244
245 /*
246 * Compute the actual number of bucket entries to pack them in power
247 * of two sizes for more efficient space utilization.
248 */
249 #define BUCKET_SIZE(n) \
250 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
251
252 #define BUCKET_MAX BUCKET_SIZE(256)
253
254 struct uma_bucket_zone bucket_zones[] = {
255 /* Literal bucket sizes. */
256 { NULL, "2 Bucket", 2, 4096 },
257 { NULL, "4 Bucket", 4, 3072 },
258 { NULL, "8 Bucket", 8, 2048 },
259 { NULL, "16 Bucket", 16, 1024 },
260 /* Rounded down power of 2 sizes for efficiency. */
261 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
262 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
263 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
264 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
265 { NULL, NULL, 0}
266 };
267
268 /*
269 * Flags and enumerations to be passed to internal functions.
270 */
271 enum zfreeskip {
272 SKIP_NONE = 0,
273 SKIP_CNT = 0x00000001,
274 SKIP_DTOR = 0x00010000,
275 SKIP_FINI = 0x00020000,
276 };
277
278 /* Prototypes.. */
279
280 void uma_startup1(vm_offset_t);
281 void uma_startup2(void);
282
283 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
284 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
285 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
286 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
287 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
288 static void page_free(void *, vm_size_t, uint8_t);
289 static void pcpu_page_free(void *, vm_size_t, uint8_t);
290 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
291 static void cache_drain(uma_zone_t);
292 static void bucket_drain(uma_zone_t, uma_bucket_t);
293 static void bucket_cache_reclaim(uma_zone_t zone, bool, int);
294 static bool bucket_cache_reclaim_domain(uma_zone_t, bool, bool, int);
295 static int keg_ctor(void *, int, void *, int);
296 static void keg_dtor(void *, int, void *);
297 static void keg_drain(uma_keg_t keg, int domain);
298 static int zone_ctor(void *, int, void *, int);
299 static void zone_dtor(void *, int, void *);
300 static inline void item_dtor(uma_zone_t zone, void *item, int size,
301 void *udata, enum zfreeskip skip);
302 static int zero_init(void *, int, int);
303 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
304 int itemdomain, bool ws);
305 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
306 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
307 static void zone_timeout(uma_zone_t zone, void *);
308 static int hash_alloc(struct uma_hash *, u_int);
309 static int hash_expand(struct uma_hash *, struct uma_hash *);
310 static void hash_free(struct uma_hash *hash);
311 static void uma_timeout(void *, int);
312 static void uma_shutdown(void);
313 static void *zone_alloc_item(uma_zone_t, void *, int, int);
314 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
315 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
316 static void zone_free_limit(uma_zone_t zone, int count);
317 static void bucket_enable(void);
318 static void bucket_init(void);
319 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
320 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
321 static void bucket_zone_drain(int domain);
322 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
323 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
324 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
325 static size_t slab_sizeof(int nitems);
326 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
327 uma_fini fini, int align, uint32_t flags);
328 static int zone_import(void *, void **, int, int, int);
329 static void zone_release(void *, void **, int);
330 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
331 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
332
333 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
334 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
335 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
336 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
337 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
338 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
339 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
340
341 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
342
343 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
344 "Memory allocation debugging");
345
346 #ifdef INVARIANTS
347 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
348 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
349
350 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
351 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
352 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
353 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
354
355 static u_int dbg_divisor = 1;
356 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
357 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
358 "Debug & thrash every this item in memory allocator");
359
360 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
361 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
362 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
363 &uma_dbg_cnt, "memory items debugged");
364 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
365 &uma_skip_cnt, "memory items skipped, not debugged");
366 #endif
367
368 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
369 "Universal Memory Allocator");
370
371 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
372 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
373
374 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
375 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
376
377 static int zone_warnings = 1;
378 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
379 "Warn when UMA zones becomes full");
380
381 static int multipage_slabs = 1;
382 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
383 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
384 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
385 "UMA may choose larger slab sizes for better efficiency");
386
387 /*
388 * Select the slab zone for an offpage slab with the given maximum item count.
389 */
390 static inline uma_zone_t
slabzone(int ipers)391 slabzone(int ipers)
392 {
393
394 return (slabzones[ipers > SLABZONE0_SETSIZE]);
395 }
396
397 /*
398 * This routine checks to see whether or not it's safe to enable buckets.
399 */
400 static void
bucket_enable(void)401 bucket_enable(void)
402 {
403
404 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
405 bucketdisable = vm_page_count_min();
406 }
407
408 /*
409 * Initialize bucket_zones, the array of zones of buckets of various sizes.
410 *
411 * For each zone, calculate the memory required for each bucket, consisting
412 * of the header and an array of pointers.
413 */
414 static void
bucket_init(void)415 bucket_init(void)
416 {
417 struct uma_bucket_zone *ubz;
418 int size;
419
420 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
421 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
422 size += sizeof(void *) * ubz->ubz_entries;
423 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
424 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
425 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
426 UMA_ZONE_FIRSTTOUCH);
427 }
428 }
429
430 /*
431 * Given a desired number of entries for a bucket, return the zone from which
432 * to allocate the bucket.
433 */
434 static struct uma_bucket_zone *
bucket_zone_lookup(int entries)435 bucket_zone_lookup(int entries)
436 {
437 struct uma_bucket_zone *ubz;
438
439 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
440 if (ubz->ubz_entries >= entries)
441 return (ubz);
442 ubz--;
443 return (ubz);
444 }
445
446 static int
bucket_select(int size)447 bucket_select(int size)
448 {
449 struct uma_bucket_zone *ubz;
450
451 ubz = &bucket_zones[0];
452 if (size > ubz->ubz_maxsize)
453 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
454
455 for (; ubz->ubz_entries != 0; ubz++)
456 if (ubz->ubz_maxsize < size)
457 break;
458 ubz--;
459 return (ubz->ubz_entries);
460 }
461
462 static uma_bucket_t
bucket_alloc(uma_zone_t zone,void * udata,int flags)463 bucket_alloc(uma_zone_t zone, void *udata, int flags)
464 {
465 struct uma_bucket_zone *ubz;
466 uma_bucket_t bucket;
467
468 /*
469 * Don't allocate buckets early in boot.
470 */
471 if (__predict_false(booted < BOOT_KVA))
472 return (NULL);
473
474 /*
475 * To limit bucket recursion we store the original zone flags
476 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
477 * NOVM flag to persist even through deep recursions. We also
478 * store ZFLAG_BUCKET once we have recursed attempting to allocate
479 * a bucket for a bucket zone so we do not allow infinite bucket
480 * recursion. This cookie will even persist to frees of unused
481 * buckets via the allocation path or bucket allocations in the
482 * free path.
483 */
484 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
485 udata = (void *)(uintptr_t)zone->uz_flags;
486 else {
487 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
488 return (NULL);
489 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
490 }
491 if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
492 flags |= M_NOVM;
493 ubz = bucket_zone_lookup(atomic_load_16(&zone->uz_bucket_size));
494 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
495 ubz++;
496 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
497 if (bucket) {
498 #ifdef INVARIANTS
499 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
500 #endif
501 bucket->ub_cnt = 0;
502 bucket->ub_entries = min(ubz->ubz_entries,
503 zone->uz_bucket_size_max);
504 bucket->ub_seq = SMR_SEQ_INVALID;
505 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
506 zone->uz_name, zone, bucket);
507 }
508
509 return (bucket);
510 }
511
512 static void
bucket_free(uma_zone_t zone,uma_bucket_t bucket,void * udata)513 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
514 {
515 struct uma_bucket_zone *ubz;
516
517 if (bucket->ub_cnt != 0)
518 bucket_drain(zone, bucket);
519
520 KASSERT(bucket->ub_cnt == 0,
521 ("bucket_free: Freeing a non free bucket."));
522 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
523 ("bucket_free: Freeing an SMR bucket."));
524 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
525 udata = (void *)(uintptr_t)zone->uz_flags;
526 ubz = bucket_zone_lookup(bucket->ub_entries);
527 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
528 }
529
530 static void
bucket_zone_drain(int domain)531 bucket_zone_drain(int domain)
532 {
533 struct uma_bucket_zone *ubz;
534
535 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
536 uma_zone_reclaim_domain(ubz->ubz_zone, UMA_RECLAIM_DRAIN,
537 domain);
538 }
539
540 #ifdef KASAN
541 _Static_assert(UMA_SMALLEST_UNIT % KASAN_SHADOW_SCALE == 0,
542 "Base UMA allocation size not a multiple of the KASAN scale factor");
543
544 static void
kasan_mark_item_valid(uma_zone_t zone,void * item)545 kasan_mark_item_valid(uma_zone_t zone, void *item)
546 {
547 void *pcpu_item;
548 size_t sz, rsz;
549 int i;
550
551 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
552 return;
553
554 sz = zone->uz_size;
555 rsz = roundup2(sz, KASAN_SHADOW_SCALE);
556 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
557 kasan_mark(item, sz, rsz, KASAN_GENERIC_REDZONE);
558 } else {
559 pcpu_item = zpcpu_base_to_offset(item);
560 for (i = 0; i <= mp_maxid; i++)
561 kasan_mark(zpcpu_get_cpu(pcpu_item, i), sz, rsz,
562 KASAN_GENERIC_REDZONE);
563 }
564 }
565
566 static void
kasan_mark_item_invalid(uma_zone_t zone,void * item)567 kasan_mark_item_invalid(uma_zone_t zone, void *item)
568 {
569 void *pcpu_item;
570 size_t sz;
571 int i;
572
573 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
574 return;
575
576 sz = roundup2(zone->uz_size, KASAN_SHADOW_SCALE);
577 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
578 kasan_mark(item, 0, sz, KASAN_UMA_FREED);
579 } else {
580 pcpu_item = zpcpu_base_to_offset(item);
581 for (i = 0; i <= mp_maxid; i++)
582 kasan_mark(zpcpu_get_cpu(pcpu_item, i), 0, sz,
583 KASAN_UMA_FREED);
584 }
585 }
586
587 static void
kasan_mark_slab_valid(uma_keg_t keg,void * mem)588 kasan_mark_slab_valid(uma_keg_t keg, void *mem)
589 {
590 size_t sz;
591
592 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
593 sz = keg->uk_ppera * PAGE_SIZE;
594 kasan_mark(mem, sz, sz, 0);
595 }
596 }
597
598 static void
kasan_mark_slab_invalid(uma_keg_t keg,void * mem)599 kasan_mark_slab_invalid(uma_keg_t keg, void *mem)
600 {
601 size_t sz;
602
603 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
604 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
605 sz = keg->uk_ppera * PAGE_SIZE;
606 else
607 sz = keg->uk_pgoff;
608 kasan_mark(mem, 0, sz, KASAN_UMA_FREED);
609 }
610 }
611 #else /* !KASAN */
612 static void
kasan_mark_item_valid(uma_zone_t zone __unused,void * item __unused)613 kasan_mark_item_valid(uma_zone_t zone __unused, void *item __unused)
614 {
615 }
616
617 static void
kasan_mark_item_invalid(uma_zone_t zone __unused,void * item __unused)618 kasan_mark_item_invalid(uma_zone_t zone __unused, void *item __unused)
619 {
620 }
621
622 static void
kasan_mark_slab_valid(uma_keg_t keg __unused,void * mem __unused)623 kasan_mark_slab_valid(uma_keg_t keg __unused, void *mem __unused)
624 {
625 }
626
627 static void
kasan_mark_slab_invalid(uma_keg_t keg __unused,void * mem __unused)628 kasan_mark_slab_invalid(uma_keg_t keg __unused, void *mem __unused)
629 {
630 }
631 #endif /* KASAN */
632
633 /*
634 * Acquire the domain lock and record contention.
635 */
636 static uma_zone_domain_t
zone_domain_lock(uma_zone_t zone,int domain)637 zone_domain_lock(uma_zone_t zone, int domain)
638 {
639 uma_zone_domain_t zdom;
640 bool lockfail;
641
642 zdom = ZDOM_GET(zone, domain);
643 lockfail = false;
644 if (ZDOM_OWNED(zdom))
645 lockfail = true;
646 ZDOM_LOCK(zdom);
647 /* This is unsynchronized. The counter does not need to be precise. */
648 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
649 zone->uz_bucket_size++;
650 return (zdom);
651 }
652
653 /*
654 * Search for the domain with the least cached items and return it if it
655 * is out of balance with the preferred domain.
656 */
657 static __noinline int
zone_domain_lowest(uma_zone_t zone,int pref)658 zone_domain_lowest(uma_zone_t zone, int pref)
659 {
660 long least, nitems, prefitems;
661 int domain;
662 int i;
663
664 prefitems = least = LONG_MAX;
665 domain = 0;
666 for (i = 0; i < vm_ndomains; i++) {
667 nitems = ZDOM_GET(zone, i)->uzd_nitems;
668 if (nitems < least) {
669 domain = i;
670 least = nitems;
671 }
672 if (domain == pref)
673 prefitems = nitems;
674 }
675 if (prefitems < least * 2)
676 return (pref);
677
678 return (domain);
679 }
680
681 /*
682 * Search for the domain with the most cached items and return it or the
683 * preferred domain if it has enough to proceed.
684 */
685 static __noinline int
zone_domain_highest(uma_zone_t zone,int pref)686 zone_domain_highest(uma_zone_t zone, int pref)
687 {
688 long most, nitems;
689 int domain;
690 int i;
691
692 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
693 return (pref);
694
695 most = 0;
696 domain = 0;
697 for (i = 0; i < vm_ndomains; i++) {
698 nitems = ZDOM_GET(zone, i)->uzd_nitems;
699 if (nitems > most) {
700 domain = i;
701 most = nitems;
702 }
703 }
704
705 return (domain);
706 }
707
708 /*
709 * Set the maximum imax value.
710 */
711 static void
zone_domain_imax_set(uma_zone_domain_t zdom,int nitems)712 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
713 {
714 long old;
715
716 old = zdom->uzd_imax;
717 do {
718 if (old >= nitems)
719 return;
720 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
721
722 /*
723 * We are at new maximum, so do the last WSS update for the old
724 * bimin and prepare to measure next allocation batch.
725 */
726 if (zdom->uzd_wss < old - zdom->uzd_bimin)
727 zdom->uzd_wss = old - zdom->uzd_bimin;
728 zdom->uzd_bimin = nitems;
729 }
730
731 /*
732 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
733 * zone's caches. If a bucket is found the zone is not locked on return.
734 */
735 static uma_bucket_t
zone_fetch_bucket(uma_zone_t zone,uma_zone_domain_t zdom,bool reclaim)736 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
737 {
738 uma_bucket_t bucket;
739 long cnt;
740 int i;
741 bool dtor = false;
742
743 ZDOM_LOCK_ASSERT(zdom);
744
745 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
746 return (NULL);
747
748 /* SMR Buckets can not be re-used until readers expire. */
749 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
750 bucket->ub_seq != SMR_SEQ_INVALID) {
751 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
752 return (NULL);
753 bucket->ub_seq = SMR_SEQ_INVALID;
754 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
755 if (STAILQ_NEXT(bucket, ub_link) != NULL)
756 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
757 }
758 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
759
760 KASSERT(zdom->uzd_nitems >= bucket->ub_cnt,
761 ("%s: item count underflow (%ld, %d)",
762 __func__, zdom->uzd_nitems, bucket->ub_cnt));
763 KASSERT(bucket->ub_cnt > 0,
764 ("%s: empty bucket in bucket cache", __func__));
765 zdom->uzd_nitems -= bucket->ub_cnt;
766
767 if (reclaim) {
768 /*
769 * Shift the bounds of the current WSS interval to avoid
770 * perturbing the estimates.
771 */
772 cnt = lmin(zdom->uzd_bimin, bucket->ub_cnt);
773 atomic_subtract_long(&zdom->uzd_imax, cnt);
774 zdom->uzd_bimin -= cnt;
775 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
776 if (zdom->uzd_limin >= bucket->ub_cnt) {
777 zdom->uzd_limin -= bucket->ub_cnt;
778 } else {
779 zdom->uzd_limin = 0;
780 zdom->uzd_timin = 0;
781 }
782 } else if (zdom->uzd_bimin > zdom->uzd_nitems) {
783 zdom->uzd_bimin = zdom->uzd_nitems;
784 if (zdom->uzd_imin > zdom->uzd_nitems)
785 zdom->uzd_imin = zdom->uzd_nitems;
786 }
787
788 ZDOM_UNLOCK(zdom);
789 if (dtor)
790 for (i = 0; i < bucket->ub_cnt; i++)
791 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
792 NULL, SKIP_NONE);
793
794 return (bucket);
795 }
796
797 /*
798 * Insert a full bucket into the specified cache. The "ws" parameter indicates
799 * whether the bucket's contents should be counted as part of the zone's working
800 * set. The bucket may be freed if it exceeds the bucket limit.
801 */
802 static void
zone_put_bucket(uma_zone_t zone,int domain,uma_bucket_t bucket,void * udata,const bool ws)803 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
804 const bool ws)
805 {
806 uma_zone_domain_t zdom;
807
808 /* We don't cache empty buckets. This can happen after a reclaim. */
809 if (bucket->ub_cnt == 0)
810 goto out;
811 zdom = zone_domain_lock(zone, domain);
812
813 /*
814 * Conditionally set the maximum number of items.
815 */
816 zdom->uzd_nitems += bucket->ub_cnt;
817 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
818 if (ws) {
819 zone_domain_imax_set(zdom, zdom->uzd_nitems);
820 } else {
821 /*
822 * Shift the bounds of the current WSS interval to
823 * avoid perturbing the estimates.
824 */
825 atomic_add_long(&zdom->uzd_imax, bucket->ub_cnt);
826 zdom->uzd_imin += bucket->ub_cnt;
827 zdom->uzd_bimin += bucket->ub_cnt;
828 zdom->uzd_limin += bucket->ub_cnt;
829 }
830 if (STAILQ_EMPTY(&zdom->uzd_buckets))
831 zdom->uzd_seq = bucket->ub_seq;
832
833 /*
834 * Try to promote reuse of recently used items. For items
835 * protected by SMR, try to defer reuse to minimize polling.
836 */
837 if (bucket->ub_seq == SMR_SEQ_INVALID)
838 STAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
839 else
840 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
841 ZDOM_UNLOCK(zdom);
842 return;
843 }
844 zdom->uzd_nitems -= bucket->ub_cnt;
845 ZDOM_UNLOCK(zdom);
846 out:
847 bucket_free(zone, bucket, udata);
848 }
849
850 /* Pops an item out of a per-cpu cache bucket. */
851 static inline void *
cache_bucket_pop(uma_cache_t cache,uma_cache_bucket_t bucket)852 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
853 {
854 void *item;
855
856 CRITICAL_ASSERT(curthread);
857
858 bucket->ucb_cnt--;
859 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
860 #ifdef INVARIANTS
861 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
862 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
863 #endif
864 cache->uc_allocs++;
865
866 return (item);
867 }
868
869 /* Pushes an item into a per-cpu cache bucket. */
870 static inline void
cache_bucket_push(uma_cache_t cache,uma_cache_bucket_t bucket,void * item)871 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
872 {
873
874 CRITICAL_ASSERT(curthread);
875 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
876 ("uma_zfree: Freeing to non free bucket index."));
877
878 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
879 bucket->ucb_cnt++;
880 cache->uc_frees++;
881 }
882
883 /*
884 * Unload a UMA bucket from a per-cpu cache.
885 */
886 static inline uma_bucket_t
cache_bucket_unload(uma_cache_bucket_t bucket)887 cache_bucket_unload(uma_cache_bucket_t bucket)
888 {
889 uma_bucket_t b;
890
891 b = bucket->ucb_bucket;
892 if (b != NULL) {
893 MPASS(b->ub_entries == bucket->ucb_entries);
894 b->ub_cnt = bucket->ucb_cnt;
895 bucket->ucb_bucket = NULL;
896 bucket->ucb_entries = bucket->ucb_cnt = 0;
897 }
898
899 return (b);
900 }
901
902 static inline uma_bucket_t
cache_bucket_unload_alloc(uma_cache_t cache)903 cache_bucket_unload_alloc(uma_cache_t cache)
904 {
905
906 return (cache_bucket_unload(&cache->uc_allocbucket));
907 }
908
909 static inline uma_bucket_t
cache_bucket_unload_free(uma_cache_t cache)910 cache_bucket_unload_free(uma_cache_t cache)
911 {
912
913 return (cache_bucket_unload(&cache->uc_freebucket));
914 }
915
916 static inline uma_bucket_t
cache_bucket_unload_cross(uma_cache_t cache)917 cache_bucket_unload_cross(uma_cache_t cache)
918 {
919
920 return (cache_bucket_unload(&cache->uc_crossbucket));
921 }
922
923 /*
924 * Load a bucket into a per-cpu cache bucket.
925 */
926 static inline void
cache_bucket_load(uma_cache_bucket_t bucket,uma_bucket_t b)927 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
928 {
929
930 CRITICAL_ASSERT(curthread);
931 MPASS(bucket->ucb_bucket == NULL);
932 MPASS(b->ub_seq == SMR_SEQ_INVALID);
933
934 bucket->ucb_bucket = b;
935 bucket->ucb_cnt = b->ub_cnt;
936 bucket->ucb_entries = b->ub_entries;
937 }
938
939 static inline void
cache_bucket_load_alloc(uma_cache_t cache,uma_bucket_t b)940 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
941 {
942
943 cache_bucket_load(&cache->uc_allocbucket, b);
944 }
945
946 static inline void
cache_bucket_load_free(uma_cache_t cache,uma_bucket_t b)947 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
948 {
949
950 cache_bucket_load(&cache->uc_freebucket, b);
951 }
952
953 #ifdef NUMA
954 static inline void
cache_bucket_load_cross(uma_cache_t cache,uma_bucket_t b)955 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
956 {
957
958 cache_bucket_load(&cache->uc_crossbucket, b);
959 }
960 #endif
961
962 /*
963 * Copy and preserve ucb_spare.
964 */
965 static inline void
cache_bucket_copy(uma_cache_bucket_t b1,uma_cache_bucket_t b2)966 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
967 {
968
969 b1->ucb_bucket = b2->ucb_bucket;
970 b1->ucb_entries = b2->ucb_entries;
971 b1->ucb_cnt = b2->ucb_cnt;
972 }
973
974 /*
975 * Swap two cache buckets.
976 */
977 static inline void
cache_bucket_swap(uma_cache_bucket_t b1,uma_cache_bucket_t b2)978 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
979 {
980 struct uma_cache_bucket b3;
981
982 CRITICAL_ASSERT(curthread);
983
984 cache_bucket_copy(&b3, b1);
985 cache_bucket_copy(b1, b2);
986 cache_bucket_copy(b2, &b3);
987 }
988
989 /*
990 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
991 */
992 static uma_bucket_t
cache_fetch_bucket(uma_zone_t zone,uma_cache_t cache,int domain)993 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
994 {
995 uma_zone_domain_t zdom;
996 uma_bucket_t bucket;
997 smr_seq_t seq;
998
999 /*
1000 * Avoid the lock if possible.
1001 */
1002 zdom = ZDOM_GET(zone, domain);
1003 if (zdom->uzd_nitems == 0)
1004 return (NULL);
1005
1006 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
1007 (seq = atomic_load_32(&zdom->uzd_seq)) != SMR_SEQ_INVALID &&
1008 !smr_poll(zone->uz_smr, seq, false))
1009 return (NULL);
1010
1011 /*
1012 * Check the zone's cache of buckets.
1013 */
1014 zdom = zone_domain_lock(zone, domain);
1015 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL)
1016 return (bucket);
1017 ZDOM_UNLOCK(zdom);
1018
1019 return (NULL);
1020 }
1021
1022 static void
zone_log_warning(uma_zone_t zone)1023 zone_log_warning(uma_zone_t zone)
1024 {
1025 static const struct timeval warninterval = { 300, 0 };
1026
1027 if (!zone_warnings || zone->uz_warning == NULL)
1028 return;
1029
1030 if (ratecheck(&zone->uz_ratecheck, &warninterval))
1031 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
1032 }
1033
1034 static inline void
zone_maxaction(uma_zone_t zone)1035 zone_maxaction(uma_zone_t zone)
1036 {
1037
1038 if (zone->uz_maxaction.ta_func != NULL)
1039 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
1040 }
1041
1042 /*
1043 * Routine called by timeout which is used to fire off some time interval
1044 * based calculations. (stats, hash size, etc.)
1045 *
1046 * Arguments:
1047 * arg Unused
1048 *
1049 * Returns:
1050 * Nothing
1051 */
1052 static void
uma_timeout(void * unused __unused,int pending __unused)1053 uma_timeout(void *unused __unused, int pending __unused)
1054 {
1055 bucket_enable();
1056 zone_foreach(zone_timeout, NULL);
1057
1058 /* Reschedule this event */
1059 taskqueue_enqueue_timeout(taskqueue_thread, &uma_timeout_task,
1060 UMA_TIMEOUT * hz);
1061 }
1062
1063 /*
1064 * Update the working set size estimates for the zone's bucket cache.
1065 * The constants chosen here are somewhat arbitrary.
1066 */
1067 static void
zone_domain_update_wss(uma_zone_domain_t zdom)1068 zone_domain_update_wss(uma_zone_domain_t zdom)
1069 {
1070 long m;
1071
1072 ZDOM_LOCK_ASSERT(zdom);
1073 MPASS(zdom->uzd_imax >= zdom->uzd_nitems);
1074 MPASS(zdom->uzd_nitems >= zdom->uzd_bimin);
1075 MPASS(zdom->uzd_bimin >= zdom->uzd_imin);
1076
1077 /*
1078 * Estimate WSS as modified moving average of biggest allocation
1079 * batches for each period over few minutes (UMA_TIMEOUT of 20s).
1080 */
1081 zdom->uzd_wss = lmax(zdom->uzd_wss * 3 / 4,
1082 zdom->uzd_imax - zdom->uzd_bimin);
1083
1084 /*
1085 * Estimate longtime minimum item count as a combination of recent
1086 * minimum item count, adjusted by WSS for safety, and the modified
1087 * moving average over the last several hours (UMA_TIMEOUT of 20s).
1088 * timin measures time since limin tried to go negative, that means
1089 * we were dangerously close to or got out of cache.
1090 */
1091 m = zdom->uzd_imin - zdom->uzd_wss;
1092 if (m >= 0) {
1093 if (zdom->uzd_limin >= m)
1094 zdom->uzd_limin = m;
1095 else
1096 zdom->uzd_limin = (m + zdom->uzd_limin * 255) / 256;
1097 zdom->uzd_timin++;
1098 } else {
1099 zdom->uzd_limin = 0;
1100 zdom->uzd_timin = 0;
1101 }
1102
1103 /* To reduce period edge effects on WSS keep half of the imax. */
1104 atomic_subtract_long(&zdom->uzd_imax,
1105 (zdom->uzd_imax - zdom->uzd_nitems + 1) / 2);
1106 zdom->uzd_imin = zdom->uzd_bimin = zdom->uzd_nitems;
1107 }
1108
1109 /*
1110 * Routine to perform timeout driven calculations. This expands the
1111 * hashes and does per cpu statistics aggregation.
1112 *
1113 * Returns nothing.
1114 */
1115 static void
zone_timeout(uma_zone_t zone,void * unused)1116 zone_timeout(uma_zone_t zone, void *unused)
1117 {
1118 uma_keg_t keg;
1119 u_int slabs, pages;
1120
1121 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
1122 goto trim;
1123
1124 keg = zone->uz_keg;
1125
1126 /*
1127 * Hash zones are non-numa by definition so the first domain
1128 * is the only one present.
1129 */
1130 KEG_LOCK(keg, 0);
1131 pages = keg->uk_domain[0].ud_pages;
1132
1133 /*
1134 * Expand the keg hash table.
1135 *
1136 * This is done if the number of slabs is larger than the hash size.
1137 * What I'm trying to do here is completely reduce collisions. This
1138 * may be a little aggressive. Should I allow for two collisions max?
1139 */
1140 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1141 struct uma_hash newhash;
1142 struct uma_hash oldhash;
1143 int ret;
1144
1145 /*
1146 * This is so involved because allocating and freeing
1147 * while the keg lock is held will lead to deadlock.
1148 * I have to do everything in stages and check for
1149 * races.
1150 */
1151 KEG_UNLOCK(keg, 0);
1152 ret = hash_alloc(&newhash, 1 << fls(slabs));
1153 KEG_LOCK(keg, 0);
1154 if (ret) {
1155 if (hash_expand(&keg->uk_hash, &newhash)) {
1156 oldhash = keg->uk_hash;
1157 keg->uk_hash = newhash;
1158 } else
1159 oldhash = newhash;
1160
1161 KEG_UNLOCK(keg, 0);
1162 hash_free(&oldhash);
1163 goto trim;
1164 }
1165 }
1166 KEG_UNLOCK(keg, 0);
1167
1168 trim:
1169 /* Trim caches not used for a long time. */
1170 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0) {
1171 for (int i = 0; i < vm_ndomains; i++) {
1172 if (bucket_cache_reclaim_domain(zone, false, false, i) &&
1173 (zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1174 keg_drain(zone->uz_keg, i);
1175 }
1176 }
1177 }
1178
1179 /*
1180 * Allocate and zero fill the next sized hash table from the appropriate
1181 * backing store.
1182 *
1183 * Arguments:
1184 * hash A new hash structure with the old hash size in uh_hashsize
1185 *
1186 * Returns:
1187 * 1 on success and 0 on failure.
1188 */
1189 static int
hash_alloc(struct uma_hash * hash,u_int size)1190 hash_alloc(struct uma_hash *hash, u_int size)
1191 {
1192 size_t alloc;
1193
1194 KASSERT(powerof2(size), ("hash size must be power of 2"));
1195 if (size > UMA_HASH_SIZE_INIT) {
1196 hash->uh_hashsize = size;
1197 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1198 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1199 } else {
1200 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1201 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1202 UMA_ANYDOMAIN, M_WAITOK);
1203 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1204 }
1205 if (hash->uh_slab_hash) {
1206 bzero(hash->uh_slab_hash, alloc);
1207 hash->uh_hashmask = hash->uh_hashsize - 1;
1208 return (1);
1209 }
1210
1211 return (0);
1212 }
1213
1214 /*
1215 * Expands the hash table for HASH zones. This is done from zone_timeout
1216 * to reduce collisions. This must not be done in the regular allocation
1217 * path, otherwise, we can recurse on the vm while allocating pages.
1218 *
1219 * Arguments:
1220 * oldhash The hash you want to expand
1221 * newhash The hash structure for the new table
1222 *
1223 * Returns:
1224 * Nothing
1225 *
1226 * Discussion:
1227 */
1228 static int
hash_expand(struct uma_hash * oldhash,struct uma_hash * newhash)1229 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1230 {
1231 uma_hash_slab_t slab;
1232 u_int hval;
1233 u_int idx;
1234
1235 if (!newhash->uh_slab_hash)
1236 return (0);
1237
1238 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1239 return (0);
1240
1241 /*
1242 * I need to investigate hash algorithms for resizing without a
1243 * full rehash.
1244 */
1245
1246 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1247 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1248 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1249 LIST_REMOVE(slab, uhs_hlink);
1250 hval = UMA_HASH(newhash, slab->uhs_data);
1251 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1252 slab, uhs_hlink);
1253 }
1254
1255 return (1);
1256 }
1257
1258 /*
1259 * Free the hash bucket to the appropriate backing store.
1260 *
1261 * Arguments:
1262 * slab_hash The hash bucket we're freeing
1263 * hashsize The number of entries in that hash bucket
1264 *
1265 * Returns:
1266 * Nothing
1267 */
1268 static void
hash_free(struct uma_hash * hash)1269 hash_free(struct uma_hash *hash)
1270 {
1271 if (hash->uh_slab_hash == NULL)
1272 return;
1273 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1274 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1275 else
1276 free(hash->uh_slab_hash, M_UMAHASH);
1277 }
1278
1279 /*
1280 * Frees all outstanding items in a bucket
1281 *
1282 * Arguments:
1283 * zone The zone to free to, must be unlocked.
1284 * bucket The free/alloc bucket with items.
1285 *
1286 * Returns:
1287 * Nothing
1288 */
1289 static void
bucket_drain(uma_zone_t zone,uma_bucket_t bucket)1290 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1291 {
1292 int i;
1293
1294 if (bucket->ub_cnt == 0)
1295 return;
1296
1297 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1298 bucket->ub_seq != SMR_SEQ_INVALID) {
1299 smr_wait(zone->uz_smr, bucket->ub_seq);
1300 bucket->ub_seq = SMR_SEQ_INVALID;
1301 for (i = 0; i < bucket->ub_cnt; i++)
1302 item_dtor(zone, bucket->ub_bucket[i],
1303 zone->uz_size, NULL, SKIP_NONE);
1304 }
1305 if (zone->uz_fini)
1306 for (i = 0; i < bucket->ub_cnt; i++) {
1307 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
1308 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1309 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
1310 }
1311 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1312 if (zone->uz_max_items > 0)
1313 zone_free_limit(zone, bucket->ub_cnt);
1314 #ifdef INVARIANTS
1315 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1316 #endif
1317 bucket->ub_cnt = 0;
1318 }
1319
1320 /*
1321 * Drains the per cpu caches for a zone.
1322 *
1323 * NOTE: This may only be called while the zone is being torn down, and not
1324 * during normal operation. This is necessary in order that we do not have
1325 * to migrate CPUs to drain the per-CPU caches.
1326 *
1327 * Arguments:
1328 * zone The zone to drain, must be unlocked.
1329 *
1330 * Returns:
1331 * Nothing
1332 */
1333 static void
cache_drain(uma_zone_t zone)1334 cache_drain(uma_zone_t zone)
1335 {
1336 uma_cache_t cache;
1337 uma_bucket_t bucket;
1338 smr_seq_t seq;
1339 int cpu;
1340
1341 /*
1342 * XXX: It is safe to not lock the per-CPU caches, because we're
1343 * tearing down the zone anyway. I.e., there will be no further use
1344 * of the caches at this point.
1345 *
1346 * XXX: It would good to be able to assert that the zone is being
1347 * torn down to prevent improper use of cache_drain().
1348 */
1349 seq = SMR_SEQ_INVALID;
1350 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1351 seq = smr_advance(zone->uz_smr);
1352 CPU_FOREACH(cpu) {
1353 cache = &zone->uz_cpu[cpu];
1354 bucket = cache_bucket_unload_alloc(cache);
1355 if (bucket != NULL)
1356 bucket_free(zone, bucket, NULL);
1357 bucket = cache_bucket_unload_free(cache);
1358 if (bucket != NULL) {
1359 bucket->ub_seq = seq;
1360 bucket_free(zone, bucket, NULL);
1361 }
1362 bucket = cache_bucket_unload_cross(cache);
1363 if (bucket != NULL) {
1364 bucket->ub_seq = seq;
1365 bucket_free(zone, bucket, NULL);
1366 }
1367 }
1368 bucket_cache_reclaim(zone, true, UMA_ANYDOMAIN);
1369 }
1370
1371 static void
cache_shrink(uma_zone_t zone,void * unused)1372 cache_shrink(uma_zone_t zone, void *unused)
1373 {
1374
1375 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1376 return;
1377
1378 ZONE_LOCK(zone);
1379 zone->uz_bucket_size =
1380 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1381 ZONE_UNLOCK(zone);
1382 }
1383
1384 static void
cache_drain_safe_cpu(uma_zone_t zone,void * unused)1385 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1386 {
1387 uma_cache_t cache;
1388 uma_bucket_t b1, b2, b3;
1389 int domain;
1390
1391 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1392 return;
1393
1394 b1 = b2 = b3 = NULL;
1395 critical_enter();
1396 cache = &zone->uz_cpu[curcpu];
1397 domain = PCPU_GET(domain);
1398 b1 = cache_bucket_unload_alloc(cache);
1399
1400 /*
1401 * Don't flush SMR zone buckets. This leaves the zone without a
1402 * bucket and forces every free to synchronize().
1403 */
1404 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1405 b2 = cache_bucket_unload_free(cache);
1406 b3 = cache_bucket_unload_cross(cache);
1407 }
1408 critical_exit();
1409
1410 if (b1 != NULL)
1411 zone_free_bucket(zone, b1, NULL, domain, false);
1412 if (b2 != NULL)
1413 zone_free_bucket(zone, b2, NULL, domain, false);
1414 if (b3 != NULL) {
1415 /* Adjust the domain so it goes to zone_free_cross. */
1416 domain = (domain + 1) % vm_ndomains;
1417 zone_free_bucket(zone, b3, NULL, domain, false);
1418 }
1419 }
1420
1421 /*
1422 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1423 * This is an expensive call because it needs to bind to all CPUs
1424 * one by one and enter a critical section on each of them in order
1425 * to safely access their cache buckets.
1426 * Zone lock must not be held on call this function.
1427 */
1428 static void
pcpu_cache_drain_safe(uma_zone_t zone)1429 pcpu_cache_drain_safe(uma_zone_t zone)
1430 {
1431 int cpu;
1432
1433 /*
1434 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1435 */
1436 if (zone)
1437 cache_shrink(zone, NULL);
1438 else
1439 zone_foreach(cache_shrink, NULL);
1440
1441 CPU_FOREACH(cpu) {
1442 thread_lock(curthread);
1443 sched_bind(curthread, cpu);
1444 thread_unlock(curthread);
1445
1446 if (zone)
1447 cache_drain_safe_cpu(zone, NULL);
1448 else
1449 zone_foreach(cache_drain_safe_cpu, NULL);
1450 }
1451 thread_lock(curthread);
1452 sched_unbind(curthread);
1453 thread_unlock(curthread);
1454 }
1455
1456 /*
1457 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1458 * requested a drain, otherwise the per-domain caches are trimmed to either
1459 * estimated working set size.
1460 */
1461 static bool
bucket_cache_reclaim_domain(uma_zone_t zone,bool drain,bool trim,int domain)1462 bucket_cache_reclaim_domain(uma_zone_t zone, bool drain, bool trim, int domain)
1463 {
1464 uma_zone_domain_t zdom;
1465 uma_bucket_t bucket;
1466 long target;
1467 bool done = false;
1468
1469 /*
1470 * The cross bucket is partially filled and not part of
1471 * the item count. Reclaim it individually here.
1472 */
1473 zdom = ZDOM_GET(zone, domain);
1474 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1475 ZONE_CROSS_LOCK(zone);
1476 bucket = zdom->uzd_cross;
1477 zdom->uzd_cross = NULL;
1478 ZONE_CROSS_UNLOCK(zone);
1479 if (bucket != NULL)
1480 bucket_free(zone, bucket, NULL);
1481 }
1482
1483 /*
1484 * If we were asked to drain the zone, we are done only once
1485 * this bucket cache is empty. If trim, we reclaim items in
1486 * excess of the zone's estimated working set size. Multiple
1487 * consecutive calls will shrink the WSS and so reclaim more.
1488 * If neither drain nor trim, then voluntarily reclaim 1/4
1489 * (to reduce first spike) of items not used for a long time.
1490 */
1491 ZDOM_LOCK(zdom);
1492 zone_domain_update_wss(zdom);
1493 if (drain)
1494 target = 0;
1495 else if (trim)
1496 target = zdom->uzd_wss;
1497 else if (zdom->uzd_timin > 900 / UMA_TIMEOUT)
1498 target = zdom->uzd_nitems - zdom->uzd_limin / 4;
1499 else {
1500 ZDOM_UNLOCK(zdom);
1501 return (done);
1502 }
1503 while ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) != NULL &&
1504 zdom->uzd_nitems >= target + bucket->ub_cnt) {
1505 bucket = zone_fetch_bucket(zone, zdom, true);
1506 if (bucket == NULL)
1507 break;
1508 bucket_free(zone, bucket, NULL);
1509 done = true;
1510 ZDOM_LOCK(zdom);
1511 }
1512 ZDOM_UNLOCK(zdom);
1513 return (done);
1514 }
1515
1516 static void
bucket_cache_reclaim(uma_zone_t zone,bool drain,int domain)1517 bucket_cache_reclaim(uma_zone_t zone, bool drain, int domain)
1518 {
1519 int i;
1520
1521 /*
1522 * Shrink the zone bucket size to ensure that the per-CPU caches
1523 * don't grow too large.
1524 */
1525 if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1526 zone->uz_bucket_size--;
1527
1528 if (domain != UMA_ANYDOMAIN &&
1529 (zone->uz_flags & UMA_ZONE_ROUNDROBIN) == 0) {
1530 bucket_cache_reclaim_domain(zone, drain, true, domain);
1531 } else {
1532 for (i = 0; i < vm_ndomains; i++)
1533 bucket_cache_reclaim_domain(zone, drain, true, i);
1534 }
1535 }
1536
1537 static void
keg_free_slab(uma_keg_t keg,uma_slab_t slab,int start)1538 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1539 {
1540 uint8_t *mem;
1541 size_t size;
1542 int i;
1543 uint8_t flags;
1544
1545 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1546 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1547
1548 mem = slab_data(slab, keg);
1549 size = PAGE_SIZE * keg->uk_ppera;
1550
1551 kasan_mark_slab_valid(keg, mem);
1552 if (keg->uk_fini != NULL) {
1553 for (i = start - 1; i > -1; i--)
1554 #ifdef INVARIANTS
1555 /*
1556 * trash_fini implies that dtor was trash_dtor. trash_fini
1557 * would check that memory hasn't been modified since free,
1558 * which executed trash_dtor.
1559 * That's why we need to run uma_dbg_kskip() check here,
1560 * albeit we don't make skip check for other init/fini
1561 * invocations.
1562 */
1563 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1564 keg->uk_fini != trash_fini)
1565 #endif
1566 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1567 }
1568 flags = slab->us_flags;
1569 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1570 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1571 NULL, SKIP_NONE);
1572 }
1573 keg->uk_freef(mem, size, flags);
1574 uma_total_dec(size);
1575 }
1576
1577 static void
keg_drain_domain(uma_keg_t keg,int domain)1578 keg_drain_domain(uma_keg_t keg, int domain)
1579 {
1580 struct slabhead freeslabs;
1581 uma_domain_t dom;
1582 uma_slab_t slab, tmp;
1583 uint32_t i, stofree, stokeep, partial;
1584
1585 dom = &keg->uk_domain[domain];
1586 LIST_INIT(&freeslabs);
1587
1588 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1589 keg->uk_name, keg, domain, dom->ud_free_items);
1590
1591 KEG_LOCK(keg, domain);
1592
1593 /*
1594 * Are the free items in partially allocated slabs sufficient to meet
1595 * the reserve? If not, compute the number of fully free slabs that must
1596 * be kept.
1597 */
1598 partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers;
1599 if (partial < keg->uk_reserve) {
1600 stokeep = min(dom->ud_free_slabs,
1601 howmany(keg->uk_reserve - partial, keg->uk_ipers));
1602 } else {
1603 stokeep = 0;
1604 }
1605 stofree = dom->ud_free_slabs - stokeep;
1606
1607 /*
1608 * Partition the free slabs into two sets: those that must be kept in
1609 * order to maintain the reserve, and those that may be released back to
1610 * the system. Since one set may be much larger than the other,
1611 * populate the smaller of the two sets and swap them if necessary.
1612 */
1613 for (i = min(stofree, stokeep); i > 0; i--) {
1614 slab = LIST_FIRST(&dom->ud_free_slab);
1615 LIST_REMOVE(slab, us_link);
1616 LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1617 }
1618 if (stofree > stokeep)
1619 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1620
1621 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1622 LIST_FOREACH(slab, &freeslabs, us_link)
1623 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1624 }
1625 dom->ud_free_items -= stofree * keg->uk_ipers;
1626 dom->ud_free_slabs -= stofree;
1627 dom->ud_pages -= stofree * keg->uk_ppera;
1628 KEG_UNLOCK(keg, domain);
1629
1630 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1631 keg_free_slab(keg, slab, keg->uk_ipers);
1632 }
1633
1634 /*
1635 * Frees pages from a keg back to the system. This is done on demand from
1636 * the pageout daemon.
1637 *
1638 * Returns nothing.
1639 */
1640 static void
keg_drain(uma_keg_t keg,int domain)1641 keg_drain(uma_keg_t keg, int domain)
1642 {
1643 int i;
1644
1645 if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0)
1646 return;
1647 if (domain != UMA_ANYDOMAIN) {
1648 keg_drain_domain(keg, domain);
1649 } else {
1650 for (i = 0; i < vm_ndomains; i++)
1651 keg_drain_domain(keg, i);
1652 }
1653 }
1654
1655 static void
zone_reclaim(uma_zone_t zone,int domain,int waitok,bool drain)1656 zone_reclaim(uma_zone_t zone, int domain, int waitok, bool drain)
1657 {
1658 /*
1659 * Count active reclaim operations in order to interlock with
1660 * zone_dtor(), which removes the zone from global lists before
1661 * attempting to reclaim items itself.
1662 *
1663 * The zone may be destroyed while sleeping, so only zone_dtor() should
1664 * specify M_WAITOK.
1665 */
1666 ZONE_LOCK(zone);
1667 if (waitok == M_WAITOK) {
1668 while (zone->uz_reclaimers > 0)
1669 msleep(zone, ZONE_LOCKPTR(zone), PVM, "zonedrain", 1);
1670 }
1671 zone->uz_reclaimers++;
1672 ZONE_UNLOCK(zone);
1673 bucket_cache_reclaim(zone, drain, domain);
1674
1675 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1676 keg_drain(zone->uz_keg, domain);
1677 ZONE_LOCK(zone);
1678 zone->uz_reclaimers--;
1679 if (zone->uz_reclaimers == 0)
1680 wakeup(zone);
1681 ZONE_UNLOCK(zone);
1682 }
1683
1684 /*
1685 * Allocate a new slab for a keg and inserts it into the partial slab list.
1686 * The keg should be unlocked on entry. If the allocation succeeds it will
1687 * be locked on return.
1688 *
1689 * Arguments:
1690 * flags Wait flags for the item initialization routine
1691 * aflags Wait flags for the slab allocation
1692 *
1693 * Returns:
1694 * The slab that was allocated or NULL if there is no memory and the
1695 * caller specified M_NOWAIT.
1696 */
1697 static uma_slab_t
keg_alloc_slab(uma_keg_t keg,uma_zone_t zone,int domain,int flags,int aflags)1698 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1699 int aflags)
1700 {
1701 uma_domain_t dom;
1702 uma_slab_t slab;
1703 unsigned long size;
1704 uint8_t *mem;
1705 uint8_t sflags;
1706 int i;
1707
1708 KASSERT(domain >= 0 && domain < vm_ndomains,
1709 ("keg_alloc_slab: domain %d out of range", domain));
1710
1711 slab = NULL;
1712 mem = NULL;
1713 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1714 uma_hash_slab_t hslab;
1715 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1716 domain, aflags);
1717 if (hslab == NULL)
1718 goto fail;
1719 slab = &hslab->uhs_slab;
1720 }
1721
1722 /*
1723 * This reproduces the old vm_zone behavior of zero filling pages the
1724 * first time they are added to a zone.
1725 *
1726 * Malloced items are zeroed in uma_zalloc.
1727 */
1728
1729 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1730 aflags |= M_ZERO;
1731 else
1732 aflags &= ~M_ZERO;
1733
1734 if (keg->uk_flags & UMA_ZONE_NODUMP)
1735 aflags |= M_NODUMP;
1736
1737 /* zone is passed for legacy reasons. */
1738 size = keg->uk_ppera * PAGE_SIZE;
1739 mem = keg->uk_allocf(zone, size, domain, &sflags, aflags);
1740 if (mem == NULL) {
1741 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1742 zone_free_item(slabzone(keg->uk_ipers),
1743 slab_tohashslab(slab), NULL, SKIP_NONE);
1744 goto fail;
1745 }
1746 uma_total_inc(size);
1747
1748 /* For HASH zones all pages go to the same uma_domain. */
1749 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1750 domain = 0;
1751
1752 /* Point the slab into the allocated memory */
1753 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1754 slab = (uma_slab_t)(mem + keg->uk_pgoff);
1755 else
1756 slab_tohashslab(slab)->uhs_data = mem;
1757
1758 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1759 for (i = 0; i < keg->uk_ppera; i++)
1760 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1761 zone, slab);
1762
1763 slab->us_freecount = keg->uk_ipers;
1764 slab->us_flags = sflags;
1765 slab->us_domain = domain;
1766
1767 BIT_FILL(keg->uk_ipers, &slab->us_free);
1768 #ifdef INVARIANTS
1769 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1770 #endif
1771
1772 if (keg->uk_init != NULL) {
1773 for (i = 0; i < keg->uk_ipers; i++)
1774 if (keg->uk_init(slab_item(slab, keg, i),
1775 keg->uk_size, flags) != 0)
1776 break;
1777 if (i != keg->uk_ipers) {
1778 keg_free_slab(keg, slab, i);
1779 goto fail;
1780 }
1781 }
1782 kasan_mark_slab_invalid(keg, mem);
1783 KEG_LOCK(keg, domain);
1784
1785 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1786 slab, keg->uk_name, keg);
1787
1788 if (keg->uk_flags & UMA_ZFLAG_HASH)
1789 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1790
1791 /*
1792 * If we got a slab here it's safe to mark it partially used
1793 * and return. We assume that the caller is going to remove
1794 * at least one item.
1795 */
1796 dom = &keg->uk_domain[domain];
1797 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1798 dom->ud_pages += keg->uk_ppera;
1799 dom->ud_free_items += keg->uk_ipers;
1800
1801 return (slab);
1802
1803 fail:
1804 return (NULL);
1805 }
1806
1807 /*
1808 * This function is intended to be used early on in place of page_alloc(). It
1809 * performs contiguous physical memory allocations and uses a bump allocator for
1810 * KVA, so is usable before the kernel map is initialized.
1811 */
1812 static void *
startup_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)1813 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1814 int wait)
1815 {
1816 vm_paddr_t pa;
1817 vm_page_t m;
1818 int i, pages;
1819
1820 pages = howmany(bytes, PAGE_SIZE);
1821 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1822
1823 *pflag = UMA_SLAB_BOOT;
1824 m = vm_page_alloc_noobj_contig_domain(domain, malloc2vm_flags(wait) |
1825 VM_ALLOC_WIRED, pages, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
1826 VM_MEMATTR_DEFAULT);
1827 if (m == NULL)
1828 return (NULL);
1829
1830 pa = VM_PAGE_TO_PHYS(m);
1831 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1832 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1833 defined(__riscv) || defined(__powerpc64__)
1834 if ((wait & M_NODUMP) == 0)
1835 dump_add_page(pa);
1836 #endif
1837 }
1838
1839 /* Allocate KVA and indirectly advance bootmem. */
1840 return ((void *)pmap_map(&bootmem, m->phys_addr,
1841 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE));
1842 }
1843
1844 static void
startup_free(void * mem,vm_size_t bytes)1845 startup_free(void *mem, vm_size_t bytes)
1846 {
1847 vm_offset_t va;
1848 vm_page_t m;
1849
1850 va = (vm_offset_t)mem;
1851 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1852
1853 /*
1854 * startup_alloc() returns direct-mapped slabs on some platforms. Avoid
1855 * unmapping ranges of the direct map.
1856 */
1857 if (va >= bootstart && va + bytes <= bootmem)
1858 pmap_remove(kernel_pmap, va, va + bytes);
1859 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1860 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1861 defined(__riscv) || defined(__powerpc64__)
1862 dump_drop_page(VM_PAGE_TO_PHYS(m));
1863 #endif
1864 vm_page_unwire_noq(m);
1865 vm_page_free(m);
1866 }
1867 }
1868
1869 /*
1870 * Allocates a number of pages from the system
1871 *
1872 * Arguments:
1873 * bytes The number of bytes requested
1874 * wait Shall we wait?
1875 *
1876 * Returns:
1877 * A pointer to the alloced memory or possibly
1878 * NULL if M_NOWAIT is set.
1879 */
1880 static void *
page_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)1881 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1882 int wait)
1883 {
1884 void *p; /* Returned page */
1885
1886 *pflag = UMA_SLAB_KERNEL;
1887 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1888
1889 return (p);
1890 }
1891
1892 static void *
pcpu_page_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)1893 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1894 int wait)
1895 {
1896 struct pglist alloctail;
1897 vm_offset_t addr, zkva;
1898 int cpu, flags;
1899 vm_page_t p, p_next;
1900 #ifdef NUMA
1901 struct pcpu *pc;
1902 #endif
1903
1904 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1905
1906 TAILQ_INIT(&alloctail);
1907 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | malloc2vm_flags(wait);
1908 *pflag = UMA_SLAB_KERNEL;
1909 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1910 if (CPU_ABSENT(cpu)) {
1911 p = vm_page_alloc_noobj(flags);
1912 } else {
1913 #ifndef NUMA
1914 p = vm_page_alloc_noobj(flags);
1915 #else
1916 pc = pcpu_find(cpu);
1917 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1918 p = NULL;
1919 else
1920 p = vm_page_alloc_noobj_domain(pc->pc_domain,
1921 flags);
1922 if (__predict_false(p == NULL))
1923 p = vm_page_alloc_noobj(flags);
1924 #endif
1925 }
1926 if (__predict_false(p == NULL))
1927 goto fail;
1928 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1929 }
1930 if ((addr = kva_alloc(bytes)) == 0)
1931 goto fail;
1932 zkva = addr;
1933 TAILQ_FOREACH(p, &alloctail, listq) {
1934 pmap_qenter(zkva, &p, 1);
1935 zkva += PAGE_SIZE;
1936 }
1937 return ((void*)addr);
1938 fail:
1939 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1940 vm_page_unwire_noq(p);
1941 vm_page_free(p);
1942 }
1943 return (NULL);
1944 }
1945
1946 /*
1947 * Allocates a number of pages from within an object
1948 *
1949 * Arguments:
1950 * bytes The number of bytes requested
1951 * wait Shall we wait?
1952 *
1953 * Returns:
1954 * A pointer to the alloced memory or possibly
1955 * NULL if M_NOWAIT is set.
1956 */
1957 static void *
noobj_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * flags,int wait)1958 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1959 int wait)
1960 {
1961 TAILQ_HEAD(, vm_page) alloctail;
1962 u_long npages;
1963 vm_offset_t retkva, zkva;
1964 vm_page_t p, p_next;
1965 uma_keg_t keg;
1966 int req;
1967
1968 TAILQ_INIT(&alloctail);
1969 keg = zone->uz_keg;
1970 req = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
1971 if ((wait & M_WAITOK) != 0)
1972 req |= VM_ALLOC_WAITOK;
1973
1974 npages = howmany(bytes, PAGE_SIZE);
1975 while (npages > 0) {
1976 p = vm_page_alloc_noobj_domain(domain, req);
1977 if (p != NULL) {
1978 /*
1979 * Since the page does not belong to an object, its
1980 * listq is unused.
1981 */
1982 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1983 npages--;
1984 continue;
1985 }
1986 /*
1987 * Page allocation failed, free intermediate pages and
1988 * exit.
1989 */
1990 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1991 vm_page_unwire_noq(p);
1992 vm_page_free(p);
1993 }
1994 return (NULL);
1995 }
1996 *flags = UMA_SLAB_PRIV;
1997 zkva = keg->uk_kva +
1998 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
1999 retkva = zkva;
2000 TAILQ_FOREACH(p, &alloctail, listq) {
2001 pmap_qenter(zkva, &p, 1);
2002 zkva += PAGE_SIZE;
2003 }
2004
2005 return ((void *)retkva);
2006 }
2007
2008 /*
2009 * Allocate physically contiguous pages.
2010 */
2011 static void *
contig_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)2012 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
2013 int wait)
2014 {
2015
2016 *pflag = UMA_SLAB_KERNEL;
2017 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
2018 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
2019 }
2020
2021 /*
2022 * Frees a number of pages to the system
2023 *
2024 * Arguments:
2025 * mem A pointer to the memory to be freed
2026 * size The size of the memory being freed
2027 * flags The original p->us_flags field
2028 *
2029 * Returns:
2030 * Nothing
2031 */
2032 static void
page_free(void * mem,vm_size_t size,uint8_t flags)2033 page_free(void *mem, vm_size_t size, uint8_t flags)
2034 {
2035
2036 if ((flags & UMA_SLAB_BOOT) != 0) {
2037 startup_free(mem, size);
2038 return;
2039 }
2040
2041 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
2042 ("UMA: page_free used with invalid flags %x", flags));
2043
2044 kmem_free((vm_offset_t)mem, size);
2045 }
2046
2047 /*
2048 * Frees pcpu zone allocations
2049 *
2050 * Arguments:
2051 * mem A pointer to the memory to be freed
2052 * size The size of the memory being freed
2053 * flags The original p->us_flags field
2054 *
2055 * Returns:
2056 * Nothing
2057 */
2058 static void
pcpu_page_free(void * mem,vm_size_t size,uint8_t flags)2059 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
2060 {
2061 vm_offset_t sva, curva;
2062 vm_paddr_t paddr;
2063 vm_page_t m;
2064
2065 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
2066
2067 if ((flags & UMA_SLAB_BOOT) != 0) {
2068 startup_free(mem, size);
2069 return;
2070 }
2071
2072 sva = (vm_offset_t)mem;
2073 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
2074 paddr = pmap_kextract(curva);
2075 m = PHYS_TO_VM_PAGE(paddr);
2076 vm_page_unwire_noq(m);
2077 vm_page_free(m);
2078 }
2079 pmap_qremove(sva, size >> PAGE_SHIFT);
2080 kva_free(sva, size);
2081 }
2082
2083 /*
2084 * Zero fill initializer
2085 *
2086 * Arguments/Returns follow uma_init specifications
2087 */
2088 static int
zero_init(void * mem,int size,int flags)2089 zero_init(void *mem, int size, int flags)
2090 {
2091 bzero(mem, size);
2092 return (0);
2093 }
2094
2095 #ifdef INVARIANTS
2096 static struct noslabbits *
slab_dbg_bits(uma_slab_t slab,uma_keg_t keg)2097 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
2098 {
2099
2100 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
2101 }
2102 #endif
2103
2104 /*
2105 * Actual size of embedded struct slab (!OFFPAGE).
2106 */
2107 static size_t
slab_sizeof(int nitems)2108 slab_sizeof(int nitems)
2109 {
2110 size_t s;
2111
2112 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
2113 return (roundup(s, UMA_ALIGN_PTR + 1));
2114 }
2115
2116 #define UMA_FIXPT_SHIFT 31
2117 #define UMA_FRAC_FIXPT(n, d) \
2118 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
2119 #define UMA_FIXPT_PCT(f) \
2120 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
2121 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
2122 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
2123
2124 /*
2125 * Compute the number of items that will fit in a slab. If hdr is true, the
2126 * item count may be limited to provide space in the slab for an inline slab
2127 * header. Otherwise, all slab space will be provided for item storage.
2128 */
2129 static u_int
slab_ipers_hdr(u_int size,u_int rsize,u_int slabsize,bool hdr)2130 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
2131 {
2132 u_int ipers;
2133 u_int padpi;
2134
2135 /* The padding between items is not needed after the last item. */
2136 padpi = rsize - size;
2137
2138 if (hdr) {
2139 /*
2140 * Start with the maximum item count and remove items until
2141 * the slab header first alongside the allocatable memory.
2142 */
2143 for (ipers = MIN(SLAB_MAX_SETSIZE,
2144 (slabsize + padpi - slab_sizeof(1)) / rsize);
2145 ipers > 0 &&
2146 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2147 ipers--)
2148 continue;
2149 } else {
2150 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2151 }
2152
2153 return (ipers);
2154 }
2155
2156 struct keg_layout_result {
2157 u_int format;
2158 u_int slabsize;
2159 u_int ipers;
2160 u_int eff;
2161 };
2162
2163 static void
keg_layout_one(uma_keg_t keg,u_int rsize,u_int slabsize,u_int fmt,struct keg_layout_result * kl)2164 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2165 struct keg_layout_result *kl)
2166 {
2167 u_int total;
2168
2169 kl->format = fmt;
2170 kl->slabsize = slabsize;
2171
2172 /* Handle INTERNAL as inline with an extra page. */
2173 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2174 kl->format &= ~UMA_ZFLAG_INTERNAL;
2175 kl->slabsize += PAGE_SIZE;
2176 }
2177
2178 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2179 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2180
2181 /* Account for memory used by an offpage slab header. */
2182 total = kl->slabsize;
2183 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2184 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2185
2186 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2187 }
2188
2189 /*
2190 * Determine the format of a uma keg. This determines where the slab header
2191 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2192 *
2193 * Arguments
2194 * keg The zone we should initialize
2195 *
2196 * Returns
2197 * Nothing
2198 */
2199 static void
keg_layout(uma_keg_t keg)2200 keg_layout(uma_keg_t keg)
2201 {
2202 struct keg_layout_result kl = {}, kl_tmp;
2203 u_int fmts[2];
2204 u_int alignsize;
2205 u_int nfmt;
2206 u_int pages;
2207 u_int rsize;
2208 u_int slabsize;
2209 u_int i, j;
2210
2211 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2212 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2213 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2214 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2215 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2216 PRINT_UMA_ZFLAGS));
2217 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2218 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2219 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2220 PRINT_UMA_ZFLAGS));
2221
2222 alignsize = keg->uk_align + 1;
2223 #ifdef KASAN
2224 /*
2225 * ASAN requires that each allocation be aligned to the shadow map
2226 * scale factor.
2227 */
2228 if (alignsize < KASAN_SHADOW_SCALE)
2229 alignsize = KASAN_SHADOW_SCALE;
2230 #endif
2231
2232 /*
2233 * Calculate the size of each allocation (rsize) according to
2234 * alignment. If the requested size is smaller than we have
2235 * allocation bits for we round it up.
2236 */
2237 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2238 rsize = roundup2(rsize, alignsize);
2239
2240 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2241 /*
2242 * We want one item to start on every align boundary in a page.
2243 * To do this we will span pages. We will also extend the item
2244 * by the size of align if it is an even multiple of align.
2245 * Otherwise, it would fall on the same boundary every time.
2246 */
2247 if ((rsize & alignsize) == 0)
2248 rsize += alignsize;
2249 slabsize = rsize * (PAGE_SIZE / alignsize);
2250 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2251 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2252 slabsize = round_page(slabsize);
2253 } else {
2254 /*
2255 * Start with a slab size of as many pages as it takes to
2256 * represent a single item. We will try to fit as many
2257 * additional items into the slab as possible.
2258 */
2259 slabsize = round_page(keg->uk_size);
2260 }
2261
2262 /* Build a list of all of the available formats for this keg. */
2263 nfmt = 0;
2264
2265 /* Evaluate an inline slab layout. */
2266 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2267 fmts[nfmt++] = 0;
2268
2269 /* TODO: vm_page-embedded slab. */
2270
2271 /*
2272 * We can't do OFFPAGE if we're internal or if we've been
2273 * asked to not go to the VM for buckets. If we do this we
2274 * may end up going to the VM for slabs which we do not want
2275 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2276 * In those cases, evaluate a pseudo-format called INTERNAL
2277 * which has an inline slab header and one extra page to
2278 * guarantee that it fits.
2279 *
2280 * Otherwise, see if using an OFFPAGE slab will improve our
2281 * efficiency.
2282 */
2283 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2284 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2285 else
2286 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2287
2288 /*
2289 * Choose a slab size and format which satisfy the minimum efficiency.
2290 * Prefer the smallest slab size that meets the constraints.
2291 *
2292 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2293 * for small items (up to PAGE_SIZE), the iteration increment is one
2294 * page; and for large items, the increment is one item.
2295 */
2296 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2297 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2298 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2299 rsize, i));
2300 for ( ; ; i++) {
2301 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2302 round_page(rsize * (i - 1) + keg->uk_size);
2303
2304 for (j = 0; j < nfmt; j++) {
2305 /* Only if we have no viable format yet. */
2306 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2307 kl.ipers > 0)
2308 continue;
2309
2310 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2311 if (kl_tmp.eff <= kl.eff)
2312 continue;
2313
2314 kl = kl_tmp;
2315
2316 CTR6(KTR_UMA, "keg %s layout: format %#x "
2317 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2318 keg->uk_name, kl.format, kl.ipers, rsize,
2319 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2320
2321 /* Stop when we reach the minimum efficiency. */
2322 if (kl.eff >= UMA_MIN_EFF)
2323 break;
2324 }
2325
2326 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2327 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2328 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2329 break;
2330 }
2331
2332 pages = atop(kl.slabsize);
2333 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2334 pages *= mp_maxid + 1;
2335
2336 keg->uk_rsize = rsize;
2337 keg->uk_ipers = kl.ipers;
2338 keg->uk_ppera = pages;
2339 keg->uk_flags |= kl.format;
2340
2341 /*
2342 * How do we find the slab header if it is offpage or if not all item
2343 * start addresses are in the same page? We could solve the latter
2344 * case with vaddr alignment, but we don't.
2345 */
2346 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2347 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2348 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2349 keg->uk_flags |= UMA_ZFLAG_HASH;
2350 else
2351 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2352 }
2353
2354 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2355 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2356 pages);
2357 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2358 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2359 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2360 keg->uk_ipers, pages));
2361 }
2362
2363 /*
2364 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2365 * the keg onto the global keg list.
2366 *
2367 * Arguments/Returns follow uma_ctor specifications
2368 * udata Actually uma_kctor_args
2369 */
2370 static int
keg_ctor(void * mem,int size,void * udata,int flags)2371 keg_ctor(void *mem, int size, void *udata, int flags)
2372 {
2373 struct uma_kctor_args *arg = udata;
2374 uma_keg_t keg = mem;
2375 uma_zone_t zone;
2376 int i;
2377
2378 bzero(keg, size);
2379 keg->uk_size = arg->size;
2380 keg->uk_init = arg->uminit;
2381 keg->uk_fini = arg->fini;
2382 keg->uk_align = arg->align;
2383 keg->uk_reserve = 0;
2384 keg->uk_flags = arg->flags;
2385
2386 /*
2387 * We use a global round-robin policy by default. Zones with
2388 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2389 * case the iterator is never run.
2390 */
2391 keg->uk_dr.dr_policy = DOMAINSET_RR();
2392 keg->uk_dr.dr_iter = 0;
2393
2394 /*
2395 * The primary zone is passed to us at keg-creation time.
2396 */
2397 zone = arg->zone;
2398 keg->uk_name = zone->uz_name;
2399
2400 if (arg->flags & UMA_ZONE_ZINIT)
2401 keg->uk_init = zero_init;
2402
2403 if (arg->flags & UMA_ZONE_MALLOC)
2404 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2405
2406 #ifndef SMP
2407 keg->uk_flags &= ~UMA_ZONE_PCPU;
2408 #endif
2409
2410 keg_layout(keg);
2411
2412 /*
2413 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2414 * work on. Use round-robin for everything else.
2415 *
2416 * Zones may override the default by specifying either.
2417 */
2418 #ifdef NUMA
2419 if ((keg->uk_flags &
2420 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2421 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2422 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2423 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2424 #endif
2425
2426 /*
2427 * If we haven't booted yet we need allocations to go through the
2428 * startup cache until the vm is ready.
2429 */
2430 #ifdef UMA_MD_SMALL_ALLOC
2431 if (keg->uk_ppera == 1)
2432 keg->uk_allocf = uma_small_alloc;
2433 else
2434 #endif
2435 if (booted < BOOT_KVA)
2436 keg->uk_allocf = startup_alloc;
2437 else if (keg->uk_flags & UMA_ZONE_PCPU)
2438 keg->uk_allocf = pcpu_page_alloc;
2439 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2440 keg->uk_allocf = contig_alloc;
2441 else
2442 keg->uk_allocf = page_alloc;
2443 #ifdef UMA_MD_SMALL_ALLOC
2444 if (keg->uk_ppera == 1)
2445 keg->uk_freef = uma_small_free;
2446 else
2447 #endif
2448 if (keg->uk_flags & UMA_ZONE_PCPU)
2449 keg->uk_freef = pcpu_page_free;
2450 else
2451 keg->uk_freef = page_free;
2452
2453 /*
2454 * Initialize keg's locks.
2455 */
2456 for (i = 0; i < vm_ndomains; i++)
2457 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2458
2459 /*
2460 * If we're putting the slab header in the actual page we need to
2461 * figure out where in each page it goes. See slab_sizeof
2462 * definition.
2463 */
2464 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2465 size_t shsize;
2466
2467 shsize = slab_sizeof(keg->uk_ipers);
2468 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2469 /*
2470 * The only way the following is possible is if with our
2471 * UMA_ALIGN_PTR adjustments we are now bigger than
2472 * UMA_SLAB_SIZE. I haven't checked whether this is
2473 * mathematically possible for all cases, so we make
2474 * sure here anyway.
2475 */
2476 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2477 ("zone %s ipers %d rsize %d size %d slab won't fit",
2478 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2479 }
2480
2481 if (keg->uk_flags & UMA_ZFLAG_HASH)
2482 hash_alloc(&keg->uk_hash, 0);
2483
2484 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2485
2486 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2487
2488 rw_wlock(&uma_rwlock);
2489 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2490 rw_wunlock(&uma_rwlock);
2491 return (0);
2492 }
2493
2494 static void
zone_kva_available(uma_zone_t zone,void * unused)2495 zone_kva_available(uma_zone_t zone, void *unused)
2496 {
2497 uma_keg_t keg;
2498
2499 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2500 return;
2501 KEG_GET(zone, keg);
2502
2503 if (keg->uk_allocf == startup_alloc) {
2504 /* Switch to the real allocator. */
2505 if (keg->uk_flags & UMA_ZONE_PCPU)
2506 keg->uk_allocf = pcpu_page_alloc;
2507 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2508 keg->uk_ppera > 1)
2509 keg->uk_allocf = contig_alloc;
2510 else
2511 keg->uk_allocf = page_alloc;
2512 }
2513 }
2514
2515 static void
zone_alloc_counters(uma_zone_t zone,void * unused)2516 zone_alloc_counters(uma_zone_t zone, void *unused)
2517 {
2518
2519 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2520 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2521 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2522 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2523 }
2524
2525 static void
zone_alloc_sysctl(uma_zone_t zone,void * unused)2526 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2527 {
2528 uma_zone_domain_t zdom;
2529 uma_domain_t dom;
2530 uma_keg_t keg;
2531 struct sysctl_oid *oid, *domainoid;
2532 int domains, i, cnt;
2533 static const char *nokeg = "cache zone";
2534 char *c;
2535
2536 /*
2537 * Make a sysctl safe copy of the zone name by removing
2538 * any special characters and handling dups by appending
2539 * an index.
2540 */
2541 if (zone->uz_namecnt != 0) {
2542 /* Count the number of decimal digits and '_' separator. */
2543 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2544 cnt /= 10;
2545 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2546 M_UMA, M_WAITOK);
2547 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2548 zone->uz_namecnt);
2549 } else
2550 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2551 for (c = zone->uz_ctlname; *c != '\0'; c++)
2552 if (strchr("./\\ -", *c) != NULL)
2553 *c = '_';
2554
2555 /*
2556 * Basic parameters at the root.
2557 */
2558 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2559 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2560 oid = zone->uz_oid;
2561 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2562 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2563 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2564 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2565 zone, 0, sysctl_handle_uma_zone_flags, "A",
2566 "Allocator configuration flags");
2567 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2568 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2569 "Desired per-cpu cache size");
2570 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2571 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2572 "Maximum allowed per-cpu cache size");
2573
2574 /*
2575 * keg if present.
2576 */
2577 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2578 domains = vm_ndomains;
2579 else
2580 domains = 1;
2581 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2582 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2583 keg = zone->uz_keg;
2584 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2585 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2586 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2587 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2588 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2589 "Real object size with alignment");
2590 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2591 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2592 "pages per-slab allocation");
2593 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2594 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2595 "items available per-slab");
2596 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2597 "align", CTLFLAG_RD, &keg->uk_align, 0,
2598 "item alignment mask");
2599 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2600 "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2601 "number of reserved items");
2602 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2603 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2604 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2605 "Slab utilization (100 - internal fragmentation %)");
2606 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2607 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2608 for (i = 0; i < domains; i++) {
2609 dom = &keg->uk_domain[i];
2610 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2611 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2612 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2613 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2614 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2615 "Total pages currently allocated from VM");
2616 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2617 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2618 "Items free in the slab layer");
2619 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2620 "free_slabs", CTLFLAG_RD, &dom->ud_free_slabs, 0,
2621 "Unused slabs");
2622 }
2623 } else
2624 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2625 "name", CTLFLAG_RD, nokeg, "Keg name");
2626
2627 /*
2628 * Information about zone limits.
2629 */
2630 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2631 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2632 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2633 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2634 zone, 0, sysctl_handle_uma_zone_items, "QU",
2635 "Current number of allocated items if limit is set");
2636 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2637 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2638 "Maximum number of allocated and cached items");
2639 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2640 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2641 "Number of threads sleeping at limit");
2642 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2643 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2644 "Total zone limit sleeps");
2645 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2646 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2647 "Maximum number of items in each domain's bucket cache");
2648
2649 /*
2650 * Per-domain zone information.
2651 */
2652 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2653 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2654 for (i = 0; i < domains; i++) {
2655 zdom = ZDOM_GET(zone, i);
2656 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2657 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2658 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2659 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2660 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2661 "number of items in this domain");
2662 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2663 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2664 "maximum item count in this period");
2665 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2666 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2667 "minimum item count in this period");
2668 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2669 "bimin", CTLFLAG_RD, &zdom->uzd_bimin,
2670 "Minimum item count in this batch");
2671 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2672 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2673 "Working set size");
2674 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2675 "limin", CTLFLAG_RD, &zdom->uzd_limin,
2676 "Long time minimum item count");
2677 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2678 "timin", CTLFLAG_RD, &zdom->uzd_timin, 0,
2679 "Time since zero long time minimum item count");
2680 }
2681
2682 /*
2683 * General statistics.
2684 */
2685 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2686 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2687 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2688 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2689 zone, 1, sysctl_handle_uma_zone_cur, "I",
2690 "Current number of allocated items");
2691 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2692 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2693 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2694 "Total allocation calls");
2695 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2696 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2697 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2698 "Total free calls");
2699 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2700 "fails", CTLFLAG_RD, &zone->uz_fails,
2701 "Number of allocation failures");
2702 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2703 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2704 "Free calls from the wrong domain");
2705 }
2706
2707 struct uma_zone_count {
2708 const char *name;
2709 int count;
2710 };
2711
2712 static void
zone_count(uma_zone_t zone,void * arg)2713 zone_count(uma_zone_t zone, void *arg)
2714 {
2715 struct uma_zone_count *cnt;
2716
2717 cnt = arg;
2718 /*
2719 * Some zones are rapidly created with identical names and
2720 * destroyed out of order. This can lead to gaps in the count.
2721 * Use one greater than the maximum observed for this name.
2722 */
2723 if (strcmp(zone->uz_name, cnt->name) == 0)
2724 cnt->count = MAX(cnt->count,
2725 zone->uz_namecnt + 1);
2726 }
2727
2728 static void
zone_update_caches(uma_zone_t zone)2729 zone_update_caches(uma_zone_t zone)
2730 {
2731 int i;
2732
2733 for (i = 0; i <= mp_maxid; i++) {
2734 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2735 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2736 }
2737 }
2738
2739 /*
2740 * Zone header ctor. This initializes all fields, locks, etc.
2741 *
2742 * Arguments/Returns follow uma_ctor specifications
2743 * udata Actually uma_zctor_args
2744 */
2745 static int
zone_ctor(void * mem,int size,void * udata,int flags)2746 zone_ctor(void *mem, int size, void *udata, int flags)
2747 {
2748 struct uma_zone_count cnt;
2749 struct uma_zctor_args *arg = udata;
2750 uma_zone_domain_t zdom;
2751 uma_zone_t zone = mem;
2752 uma_zone_t z;
2753 uma_keg_t keg;
2754 int i;
2755
2756 bzero(zone, size);
2757 zone->uz_name = arg->name;
2758 zone->uz_ctor = arg->ctor;
2759 zone->uz_dtor = arg->dtor;
2760 zone->uz_init = NULL;
2761 zone->uz_fini = NULL;
2762 zone->uz_sleeps = 0;
2763 zone->uz_bucket_size = 0;
2764 zone->uz_bucket_size_min = 0;
2765 zone->uz_bucket_size_max = BUCKET_MAX;
2766 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2767 zone->uz_warning = NULL;
2768 /* The domain structures follow the cpu structures. */
2769 zone->uz_bucket_max = ULONG_MAX;
2770 timevalclear(&zone->uz_ratecheck);
2771
2772 /* Count the number of duplicate names. */
2773 cnt.name = arg->name;
2774 cnt.count = 0;
2775 zone_foreach(zone_count, &cnt);
2776 zone->uz_namecnt = cnt.count;
2777 ZONE_CROSS_LOCK_INIT(zone);
2778
2779 for (i = 0; i < vm_ndomains; i++) {
2780 zdom = ZDOM_GET(zone, i);
2781 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2782 STAILQ_INIT(&zdom->uzd_buckets);
2783 }
2784
2785 #if defined(INVARIANTS) && !defined(KASAN)
2786 if (arg->uminit == trash_init && arg->fini == trash_fini)
2787 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2788 #elif defined(KASAN)
2789 if ((arg->flags & (UMA_ZONE_NOFREE | UMA_ZFLAG_CACHE)) != 0)
2790 arg->flags |= UMA_ZONE_NOKASAN;
2791 #endif
2792
2793 /*
2794 * This is a pure cache zone, no kegs.
2795 */
2796 if (arg->import) {
2797 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2798 ("zone_ctor: Import specified for non-cache zone."));
2799 zone->uz_flags = arg->flags;
2800 zone->uz_size = arg->size;
2801 zone->uz_import = arg->import;
2802 zone->uz_release = arg->release;
2803 zone->uz_arg = arg->arg;
2804 #ifdef NUMA
2805 /*
2806 * Cache zones are round-robin unless a policy is
2807 * specified because they may have incompatible
2808 * constraints.
2809 */
2810 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2811 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2812 #endif
2813 rw_wlock(&uma_rwlock);
2814 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2815 rw_wunlock(&uma_rwlock);
2816 goto out;
2817 }
2818
2819 /*
2820 * Use the regular zone/keg/slab allocator.
2821 */
2822 zone->uz_import = zone_import;
2823 zone->uz_release = zone_release;
2824 zone->uz_arg = zone;
2825 keg = arg->keg;
2826
2827 if (arg->flags & UMA_ZONE_SECONDARY) {
2828 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2829 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2830 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2831 zone->uz_init = arg->uminit;
2832 zone->uz_fini = arg->fini;
2833 zone->uz_flags |= UMA_ZONE_SECONDARY;
2834 rw_wlock(&uma_rwlock);
2835 ZONE_LOCK(zone);
2836 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2837 if (LIST_NEXT(z, uz_link) == NULL) {
2838 LIST_INSERT_AFTER(z, zone, uz_link);
2839 break;
2840 }
2841 }
2842 ZONE_UNLOCK(zone);
2843 rw_wunlock(&uma_rwlock);
2844 } else if (keg == NULL) {
2845 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2846 arg->align, arg->flags)) == NULL)
2847 return (ENOMEM);
2848 } else {
2849 struct uma_kctor_args karg;
2850 int error;
2851
2852 /* We should only be here from uma_startup() */
2853 karg.size = arg->size;
2854 karg.uminit = arg->uminit;
2855 karg.fini = arg->fini;
2856 karg.align = arg->align;
2857 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2858 karg.zone = zone;
2859 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2860 flags);
2861 if (error)
2862 return (error);
2863 }
2864
2865 /* Inherit properties from the keg. */
2866 zone->uz_keg = keg;
2867 zone->uz_size = keg->uk_size;
2868 zone->uz_flags |= (keg->uk_flags &
2869 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2870
2871 out:
2872 if (booted >= BOOT_PCPU) {
2873 zone_alloc_counters(zone, NULL);
2874 if (booted >= BOOT_RUNNING)
2875 zone_alloc_sysctl(zone, NULL);
2876 } else {
2877 zone->uz_allocs = EARLY_COUNTER;
2878 zone->uz_frees = EARLY_COUNTER;
2879 zone->uz_fails = EARLY_COUNTER;
2880 }
2881
2882 /* Caller requests a private SMR context. */
2883 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2884 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2885
2886 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2887 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2888 ("Invalid zone flag combination"));
2889 if (arg->flags & UMA_ZFLAG_INTERNAL)
2890 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2891 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2892 zone->uz_bucket_size = BUCKET_MAX;
2893 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2894 zone->uz_bucket_size = 0;
2895 else
2896 zone->uz_bucket_size = bucket_select(zone->uz_size);
2897 zone->uz_bucket_size_min = zone->uz_bucket_size;
2898 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2899 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2900 zone_update_caches(zone);
2901
2902 return (0);
2903 }
2904
2905 /*
2906 * Keg header dtor. This frees all data, destroys locks, frees the hash
2907 * table and removes the keg from the global list.
2908 *
2909 * Arguments/Returns follow uma_dtor specifications
2910 * udata unused
2911 */
2912 static void
keg_dtor(void * arg,int size,void * udata)2913 keg_dtor(void *arg, int size, void *udata)
2914 {
2915 uma_keg_t keg;
2916 uint32_t free, pages;
2917 int i;
2918
2919 keg = (uma_keg_t)arg;
2920 free = pages = 0;
2921 for (i = 0; i < vm_ndomains; i++) {
2922 free += keg->uk_domain[i].ud_free_items;
2923 pages += keg->uk_domain[i].ud_pages;
2924 KEG_LOCK_FINI(keg, i);
2925 }
2926 if (pages != 0)
2927 printf("Freed UMA keg (%s) was not empty (%u items). "
2928 " Lost %u pages of memory.\n",
2929 keg->uk_name ? keg->uk_name : "",
2930 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2931
2932 hash_free(&keg->uk_hash);
2933 }
2934
2935 /*
2936 * Zone header dtor.
2937 *
2938 * Arguments/Returns follow uma_dtor specifications
2939 * udata unused
2940 */
2941 static void
zone_dtor(void * arg,int size,void * udata)2942 zone_dtor(void *arg, int size, void *udata)
2943 {
2944 uma_zone_t zone;
2945 uma_keg_t keg;
2946 int i;
2947
2948 zone = (uma_zone_t)arg;
2949
2950 sysctl_remove_oid(zone->uz_oid, 1, 1);
2951
2952 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2953 cache_drain(zone);
2954
2955 rw_wlock(&uma_rwlock);
2956 LIST_REMOVE(zone, uz_link);
2957 rw_wunlock(&uma_rwlock);
2958 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2959 keg = zone->uz_keg;
2960 keg->uk_reserve = 0;
2961 }
2962 zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true);
2963
2964 /*
2965 * We only destroy kegs from non secondary/non cache zones.
2966 */
2967 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2968 keg = zone->uz_keg;
2969 rw_wlock(&uma_rwlock);
2970 LIST_REMOVE(keg, uk_link);
2971 rw_wunlock(&uma_rwlock);
2972 zone_free_item(kegs, keg, NULL, SKIP_NONE);
2973 }
2974 counter_u64_free(zone->uz_allocs);
2975 counter_u64_free(zone->uz_frees);
2976 counter_u64_free(zone->uz_fails);
2977 counter_u64_free(zone->uz_xdomain);
2978 free(zone->uz_ctlname, M_UMA);
2979 for (i = 0; i < vm_ndomains; i++)
2980 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2981 ZONE_CROSS_LOCK_FINI(zone);
2982 }
2983
2984 static void
zone_foreach_unlocked(void (* zfunc)(uma_zone_t,void * arg),void * arg)2985 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2986 {
2987 uma_keg_t keg;
2988 uma_zone_t zone;
2989
2990 LIST_FOREACH(keg, &uma_kegs, uk_link) {
2991 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2992 zfunc(zone, arg);
2993 }
2994 LIST_FOREACH(zone, &uma_cachezones, uz_link)
2995 zfunc(zone, arg);
2996 }
2997
2998 /*
2999 * Traverses every zone in the system and calls a callback
3000 *
3001 * Arguments:
3002 * zfunc A pointer to a function which accepts a zone
3003 * as an argument.
3004 *
3005 * Returns:
3006 * Nothing
3007 */
3008 static void
zone_foreach(void (* zfunc)(uma_zone_t,void * arg),void * arg)3009 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3010 {
3011
3012 rw_rlock(&uma_rwlock);
3013 zone_foreach_unlocked(zfunc, arg);
3014 rw_runlock(&uma_rwlock);
3015 }
3016
3017 /*
3018 * Initialize the kernel memory allocator. This is done after pages can be
3019 * allocated but before general KVA is available.
3020 */
3021 void
uma_startup1(vm_offset_t virtual_avail)3022 uma_startup1(vm_offset_t virtual_avail)
3023 {
3024 struct uma_zctor_args args;
3025 size_t ksize, zsize, size;
3026 uma_keg_t primarykeg;
3027 uintptr_t m;
3028 int domain;
3029 uint8_t pflag;
3030
3031 bootstart = bootmem = virtual_avail;
3032
3033 rw_init(&uma_rwlock, "UMA lock");
3034 sx_init(&uma_reclaim_lock, "umareclaim");
3035
3036 ksize = sizeof(struct uma_keg) +
3037 (sizeof(struct uma_domain) * vm_ndomains);
3038 ksize = roundup(ksize, UMA_SUPER_ALIGN);
3039 zsize = sizeof(struct uma_zone) +
3040 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
3041 (sizeof(struct uma_zone_domain) * vm_ndomains);
3042 zsize = roundup(zsize, UMA_SUPER_ALIGN);
3043
3044 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
3045 size = (zsize * 2) + ksize;
3046 for (domain = 0; domain < vm_ndomains; domain++) {
3047 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
3048 M_NOWAIT | M_ZERO);
3049 if (m != 0)
3050 break;
3051 }
3052 zones = (uma_zone_t)m;
3053 m += zsize;
3054 kegs = (uma_zone_t)m;
3055 m += zsize;
3056 primarykeg = (uma_keg_t)m;
3057
3058 /* "manually" create the initial zone */
3059 memset(&args, 0, sizeof(args));
3060 args.name = "UMA Kegs";
3061 args.size = ksize;
3062 args.ctor = keg_ctor;
3063 args.dtor = keg_dtor;
3064 args.uminit = zero_init;
3065 args.fini = NULL;
3066 args.keg = primarykeg;
3067 args.align = UMA_SUPER_ALIGN - 1;
3068 args.flags = UMA_ZFLAG_INTERNAL;
3069 zone_ctor(kegs, zsize, &args, M_WAITOK);
3070
3071 args.name = "UMA Zones";
3072 args.size = zsize;
3073 args.ctor = zone_ctor;
3074 args.dtor = zone_dtor;
3075 args.uminit = zero_init;
3076 args.fini = NULL;
3077 args.keg = NULL;
3078 args.align = UMA_SUPER_ALIGN - 1;
3079 args.flags = UMA_ZFLAG_INTERNAL;
3080 zone_ctor(zones, zsize, &args, M_WAITOK);
3081
3082 /* Now make zones for slab headers */
3083 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
3084 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3085 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
3086 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3087
3088 hashzone = uma_zcreate("UMA Hash",
3089 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
3090 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3091
3092 bucket_init();
3093 smr_init();
3094 }
3095
3096 #ifndef UMA_MD_SMALL_ALLOC
3097 extern void vm_radix_reserve_kva(void);
3098 #endif
3099
3100 /*
3101 * Advertise the availability of normal kva allocations and switch to
3102 * the default back-end allocator. Marks the KVA we consumed on startup
3103 * as used in the map.
3104 */
3105 void
uma_startup2(void)3106 uma_startup2(void)
3107 {
3108
3109 if (bootstart != bootmem) {
3110 vm_map_lock(kernel_map);
3111 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
3112 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
3113 vm_map_unlock(kernel_map);
3114 }
3115
3116 #ifndef UMA_MD_SMALL_ALLOC
3117 /* Set up radix zone to use noobj_alloc. */
3118 vm_radix_reserve_kva();
3119 #endif
3120
3121 booted = BOOT_KVA;
3122 zone_foreach_unlocked(zone_kva_available, NULL);
3123 bucket_enable();
3124 }
3125
3126 /*
3127 * Allocate counters as early as possible so that boot-time allocations are
3128 * accounted more precisely.
3129 */
3130 static void
uma_startup_pcpu(void * arg __unused)3131 uma_startup_pcpu(void *arg __unused)
3132 {
3133
3134 zone_foreach_unlocked(zone_alloc_counters, NULL);
3135 booted = BOOT_PCPU;
3136 }
3137 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
3138
3139 /*
3140 * Finish our initialization steps.
3141 */
3142 static void
uma_startup3(void * arg __unused)3143 uma_startup3(void *arg __unused)
3144 {
3145
3146 #ifdef INVARIANTS
3147 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
3148 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
3149 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
3150 #endif
3151 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
3152 booted = BOOT_RUNNING;
3153
3154 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
3155 EVENTHANDLER_PRI_FIRST);
3156 }
3157 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
3158
3159 static void
uma_startup4(void * arg __unused)3160 uma_startup4(void *arg __unused)
3161 {
3162 TIMEOUT_TASK_INIT(taskqueue_thread, &uma_timeout_task, 0, uma_timeout,
3163 NULL);
3164 taskqueue_enqueue_timeout(taskqueue_thread, &uma_timeout_task,
3165 UMA_TIMEOUT * hz);
3166 }
3167 SYSINIT(uma_startup4, SI_SUB_TASKQ, SI_ORDER_ANY, uma_startup4, NULL);
3168
3169 static void
uma_shutdown(void)3170 uma_shutdown(void)
3171 {
3172
3173 booted = BOOT_SHUTDOWN;
3174 }
3175
3176 static uma_keg_t
uma_kcreate(uma_zone_t zone,size_t size,uma_init uminit,uma_fini fini,int align,uint32_t flags)3177 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
3178 int align, uint32_t flags)
3179 {
3180 struct uma_kctor_args args;
3181
3182 args.size = size;
3183 args.uminit = uminit;
3184 args.fini = fini;
3185 args.align = align;
3186 args.flags = flags;
3187 args.zone = zone;
3188 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3189 }
3190
3191
3192 static void
check_align_mask(unsigned int mask)3193 check_align_mask(unsigned int mask)
3194 {
3195
3196 KASSERT(powerof2(mask + 1),
3197 ("UMA: %s: Not the mask of a power of 2 (%#x)", __func__, mask));
3198 /*
3199 * Make sure the stored align mask doesn't have its highest bit set,
3200 * which would cause implementation-defined behavior when passing it as
3201 * the 'align' argument of uma_zcreate(). Such very large alignments do
3202 * not make sense anyway.
3203 */
3204 KASSERT(mask <= INT_MAX,
3205 ("UMA: %s: Mask too big (%#x)", __func__, mask));
3206 }
3207
3208 /* Public functions */
3209 /* See uma.h */
3210 void
uma_set_cache_align_mask(unsigned int mask)3211 uma_set_cache_align_mask(unsigned int mask)
3212 {
3213
3214 check_align_mask(mask);
3215 uma_cache_align_mask = mask;
3216 }
3217
3218 /* Returns the alignment mask to use to request cache alignment. */
3219 unsigned int
uma_get_cache_align_mask(void)3220 uma_get_cache_align_mask(void)
3221 {
3222 return (uma_cache_align_mask);
3223 }
3224
3225 /* See uma.h */
3226 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)3227 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3228 uma_init uminit, uma_fini fini, int align, uint32_t flags)
3229
3230 {
3231 struct uma_zctor_args args;
3232 uma_zone_t res;
3233
3234 check_align_mask(align);
3235
3236 /* This stuff is essential for the zone ctor */
3237 memset(&args, 0, sizeof(args));
3238 args.name = name;
3239 args.size = size;
3240 args.ctor = ctor;
3241 args.dtor = dtor;
3242 args.uminit = uminit;
3243 args.fini = fini;
3244 #if defined(INVARIANTS) && !defined(KASAN)
3245 /*
3246 * Inject procedures which check for memory use after free if we are
3247 * allowed to scramble the memory while it is not allocated. This
3248 * requires that: UMA is actually able to access the memory, no init
3249 * or fini procedures, no dependency on the initial value of the
3250 * memory, and no (legitimate) use of the memory after free. Note,
3251 * the ctor and dtor do not need to be empty.
3252 */
3253 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3254 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3255 args.uminit = trash_init;
3256 args.fini = trash_fini;
3257 }
3258 #endif
3259 args.align = align;
3260 args.flags = flags;
3261 args.keg = NULL;
3262
3263 sx_xlock(&uma_reclaim_lock);
3264 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3265 sx_xunlock(&uma_reclaim_lock);
3266
3267 return (res);
3268 }
3269
3270 /* See uma.h */
3271 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)3272 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3273 uma_init zinit, uma_fini zfini, uma_zone_t primary)
3274 {
3275 struct uma_zctor_args args;
3276 uma_keg_t keg;
3277 uma_zone_t res;
3278
3279 keg = primary->uz_keg;
3280 memset(&args, 0, sizeof(args));
3281 args.name = name;
3282 args.size = keg->uk_size;
3283 args.ctor = ctor;
3284 args.dtor = dtor;
3285 args.uminit = zinit;
3286 args.fini = zfini;
3287 args.align = keg->uk_align;
3288 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3289 args.keg = keg;
3290
3291 sx_xlock(&uma_reclaim_lock);
3292 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3293 sx_xunlock(&uma_reclaim_lock);
3294
3295 return (res);
3296 }
3297
3298 /* See uma.h */
3299 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)3300 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3301 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3302 void *arg, int flags)
3303 {
3304 struct uma_zctor_args args;
3305
3306 memset(&args, 0, sizeof(args));
3307 args.name = name;
3308 args.size = size;
3309 args.ctor = ctor;
3310 args.dtor = dtor;
3311 args.uminit = zinit;
3312 args.fini = zfini;
3313 args.import = zimport;
3314 args.release = zrelease;
3315 args.arg = arg;
3316 args.align = 0;
3317 args.flags = flags | UMA_ZFLAG_CACHE;
3318
3319 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3320 }
3321
3322 /* See uma.h */
3323 void
uma_zdestroy(uma_zone_t zone)3324 uma_zdestroy(uma_zone_t zone)
3325 {
3326
3327 /*
3328 * Large slabs are expensive to reclaim, so don't bother doing
3329 * unnecessary work if we're shutting down.
3330 */
3331 if (booted == BOOT_SHUTDOWN &&
3332 zone->uz_fini == NULL && zone->uz_release == zone_release)
3333 return;
3334 sx_xlock(&uma_reclaim_lock);
3335 zone_free_item(zones, zone, NULL, SKIP_NONE);
3336 sx_xunlock(&uma_reclaim_lock);
3337 }
3338
3339 void
uma_zwait(uma_zone_t zone)3340 uma_zwait(uma_zone_t zone)
3341 {
3342
3343 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3344 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3345 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3346 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3347 else
3348 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3349 }
3350
3351 void *
uma_zalloc_pcpu_arg(uma_zone_t zone,void * udata,int flags)3352 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3353 {
3354 void *item, *pcpu_item;
3355 #ifdef SMP
3356 int i;
3357
3358 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3359 #endif
3360 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3361 if (item == NULL)
3362 return (NULL);
3363 pcpu_item = zpcpu_base_to_offset(item);
3364 if (flags & M_ZERO) {
3365 #ifdef SMP
3366 for (i = 0; i <= mp_maxid; i++)
3367 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3368 #else
3369 bzero(item, zone->uz_size);
3370 #endif
3371 }
3372 return (pcpu_item);
3373 }
3374
3375 /*
3376 * A stub while both regular and pcpu cases are identical.
3377 */
3378 void
uma_zfree_pcpu_arg(uma_zone_t zone,void * pcpu_item,void * udata)3379 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3380 {
3381 void *item;
3382
3383 #ifdef SMP
3384 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3385 #endif
3386
3387 /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */
3388 if (pcpu_item == NULL)
3389 return;
3390
3391 item = zpcpu_offset_to_base(pcpu_item);
3392 uma_zfree_arg(zone, item, udata);
3393 }
3394
3395 static inline void *
item_ctor(uma_zone_t zone,int uz_flags,int size,void * udata,int flags,void * item)3396 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3397 void *item)
3398 {
3399 #ifdef INVARIANTS
3400 bool skipdbg;
3401 #endif
3402
3403 kasan_mark_item_valid(zone, item);
3404
3405 #ifdef INVARIANTS
3406 skipdbg = uma_dbg_zskip(zone, item);
3407 if (!skipdbg && (uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3408 zone->uz_ctor != trash_ctor)
3409 trash_ctor(item, size, udata, flags);
3410 #endif
3411
3412 /* Check flags before loading ctor pointer. */
3413 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3414 __predict_false(zone->uz_ctor != NULL) &&
3415 zone->uz_ctor(item, size, udata, flags) != 0) {
3416 counter_u64_add(zone->uz_fails, 1);
3417 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3418 return (NULL);
3419 }
3420 #ifdef INVARIANTS
3421 if (!skipdbg)
3422 uma_dbg_alloc(zone, NULL, item);
3423 #endif
3424 if (__predict_false(flags & M_ZERO))
3425 return (memset(item, 0, size));
3426
3427 return (item);
3428 }
3429
3430 static inline void
item_dtor(uma_zone_t zone,void * item,int size,void * udata,enum zfreeskip skip)3431 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3432 enum zfreeskip skip)
3433 {
3434 #ifdef INVARIANTS
3435 bool skipdbg;
3436
3437 skipdbg = uma_dbg_zskip(zone, item);
3438 if (skip == SKIP_NONE && !skipdbg) {
3439 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3440 uma_dbg_free(zone, udata, item);
3441 else
3442 uma_dbg_free(zone, NULL, item);
3443 }
3444 #endif
3445 if (__predict_true(skip < SKIP_DTOR)) {
3446 if (zone->uz_dtor != NULL)
3447 zone->uz_dtor(item, size, udata);
3448 #ifdef INVARIANTS
3449 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3450 zone->uz_dtor != trash_dtor)
3451 trash_dtor(item, size, udata);
3452 #endif
3453 }
3454 kasan_mark_item_invalid(zone, item);
3455 }
3456
3457 #ifdef NUMA
3458 static int
item_domain(void * item)3459 item_domain(void *item)
3460 {
3461 int domain;
3462
3463 domain = vm_phys_domain(vtophys(item));
3464 KASSERT(domain >= 0 && domain < vm_ndomains,
3465 ("%s: unknown domain for item %p", __func__, item));
3466 return (domain);
3467 }
3468 #endif
3469
3470 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3471 #define UMA_ZALLOC_DEBUG
3472 static int
uma_zalloc_debug(uma_zone_t zone,void ** itemp,void * udata,int flags)3473 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3474 {
3475 int error;
3476
3477 error = 0;
3478 #ifdef WITNESS
3479 if (flags & M_WAITOK) {
3480 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3481 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3482 }
3483 #endif
3484
3485 #ifdef INVARIANTS
3486 KASSERT((flags & M_EXEC) == 0,
3487 ("uma_zalloc_debug: called with M_EXEC"));
3488 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3489 ("uma_zalloc_debug: called within spinlock or critical section"));
3490 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3491 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3492 #endif
3493
3494 #ifdef DEBUG_MEMGUARD
3495 if ((zone->uz_flags & (UMA_ZONE_SMR | UMA_ZFLAG_CACHE)) == 0 &&
3496 memguard_cmp_zone(zone)) {
3497 void *item;
3498 item = memguard_alloc(zone->uz_size, flags);
3499 if (item != NULL) {
3500 error = EJUSTRETURN;
3501 if (zone->uz_init != NULL &&
3502 zone->uz_init(item, zone->uz_size, flags) != 0) {
3503 *itemp = NULL;
3504 return (error);
3505 }
3506 if (zone->uz_ctor != NULL &&
3507 zone->uz_ctor(item, zone->uz_size, udata,
3508 flags) != 0) {
3509 counter_u64_add(zone->uz_fails, 1);
3510 if (zone->uz_fini != NULL)
3511 zone->uz_fini(item, zone->uz_size);
3512 *itemp = NULL;
3513 return (error);
3514 }
3515 *itemp = item;
3516 return (error);
3517 }
3518 /* This is unfortunate but should not be fatal. */
3519 }
3520 #endif
3521 return (error);
3522 }
3523
3524 static int
uma_zfree_debug(uma_zone_t zone,void * item,void * udata)3525 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3526 {
3527 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3528 ("uma_zfree_debug: called with spinlock or critical section held"));
3529
3530 #ifdef DEBUG_MEMGUARD
3531 if ((zone->uz_flags & (UMA_ZONE_SMR | UMA_ZFLAG_CACHE)) == 0 &&
3532 is_memguard_addr(item)) {
3533 if (zone->uz_dtor != NULL)
3534 zone->uz_dtor(item, zone->uz_size, udata);
3535 if (zone->uz_fini != NULL)
3536 zone->uz_fini(item, zone->uz_size);
3537 memguard_free(item);
3538 return (EJUSTRETURN);
3539 }
3540 #endif
3541 return (0);
3542 }
3543 #endif
3544
3545 static inline void *
cache_alloc_item(uma_zone_t zone,uma_cache_t cache,uma_cache_bucket_t bucket,void * udata,int flags)3546 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3547 void *udata, int flags)
3548 {
3549 void *item;
3550 int size, uz_flags;
3551
3552 item = cache_bucket_pop(cache, bucket);
3553 size = cache_uz_size(cache);
3554 uz_flags = cache_uz_flags(cache);
3555 critical_exit();
3556 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3557 }
3558
3559 static __noinline void *
cache_alloc_retry(uma_zone_t zone,uma_cache_t cache,void * udata,int flags)3560 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3561 {
3562 uma_cache_bucket_t bucket;
3563 int domain;
3564
3565 while (cache_alloc(zone, cache, udata, flags)) {
3566 cache = &zone->uz_cpu[curcpu];
3567 bucket = &cache->uc_allocbucket;
3568 if (__predict_false(bucket->ucb_cnt == 0))
3569 continue;
3570 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3571 }
3572 critical_exit();
3573
3574 /*
3575 * We can not get a bucket so try to return a single item.
3576 */
3577 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3578 domain = PCPU_GET(domain);
3579 else
3580 domain = UMA_ANYDOMAIN;
3581 return (zone_alloc_item(zone, udata, domain, flags));
3582 }
3583
3584 /* See uma.h */
3585 void *
uma_zalloc_smr(uma_zone_t zone,int flags)3586 uma_zalloc_smr(uma_zone_t zone, int flags)
3587 {
3588 uma_cache_bucket_t bucket;
3589 uma_cache_t cache;
3590
3591 #ifdef UMA_ZALLOC_DEBUG
3592 void *item;
3593
3594 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3595 ("uma_zalloc_arg: called with non-SMR zone."));
3596 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3597 return (item);
3598 #endif
3599
3600 critical_enter();
3601 cache = &zone->uz_cpu[curcpu];
3602 bucket = &cache->uc_allocbucket;
3603 if (__predict_false(bucket->ucb_cnt == 0))
3604 return (cache_alloc_retry(zone, cache, NULL, flags));
3605 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3606 }
3607
3608 /* See uma.h */
3609 void *
uma_zalloc_arg(uma_zone_t zone,void * udata,int flags)3610 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3611 {
3612 uma_cache_bucket_t bucket;
3613 uma_cache_t cache;
3614
3615 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3616 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3617
3618 /* This is the fast path allocation */
3619 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3620 zone, flags);
3621
3622 #ifdef UMA_ZALLOC_DEBUG
3623 void *item;
3624
3625 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3626 ("uma_zalloc_arg: called with SMR zone."));
3627 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3628 return (item);
3629 #endif
3630
3631 /*
3632 * If possible, allocate from the per-CPU cache. There are two
3633 * requirements for safe access to the per-CPU cache: (1) the thread
3634 * accessing the cache must not be preempted or yield during access,
3635 * and (2) the thread must not migrate CPUs without switching which
3636 * cache it accesses. We rely on a critical section to prevent
3637 * preemption and migration. We release the critical section in
3638 * order to acquire the zone mutex if we are unable to allocate from
3639 * the current cache; when we re-acquire the critical section, we
3640 * must detect and handle migration if it has occurred.
3641 */
3642 critical_enter();
3643 cache = &zone->uz_cpu[curcpu];
3644 bucket = &cache->uc_allocbucket;
3645 if (__predict_false(bucket->ucb_cnt == 0))
3646 return (cache_alloc_retry(zone, cache, udata, flags));
3647 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3648 }
3649
3650 /*
3651 * Replenish an alloc bucket and possibly restore an old one. Called in
3652 * a critical section. Returns in a critical section.
3653 *
3654 * A false return value indicates an allocation failure.
3655 * A true return value indicates success and the caller should retry.
3656 */
3657 static __noinline bool
cache_alloc(uma_zone_t zone,uma_cache_t cache,void * udata,int flags)3658 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3659 {
3660 uma_bucket_t bucket;
3661 int curdomain, domain;
3662 bool new;
3663
3664 CRITICAL_ASSERT(curthread);
3665
3666 /*
3667 * If we have run out of items in our alloc bucket see
3668 * if we can switch with the free bucket.
3669 *
3670 * SMR Zones can't re-use the free bucket until the sequence has
3671 * expired.
3672 */
3673 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3674 cache->uc_freebucket.ucb_cnt != 0) {
3675 cache_bucket_swap(&cache->uc_freebucket,
3676 &cache->uc_allocbucket);
3677 return (true);
3678 }
3679
3680 /*
3681 * Discard any empty allocation bucket while we hold no locks.
3682 */
3683 bucket = cache_bucket_unload_alloc(cache);
3684 critical_exit();
3685
3686 if (bucket != NULL) {
3687 KASSERT(bucket->ub_cnt == 0,
3688 ("cache_alloc: Entered with non-empty alloc bucket."));
3689 bucket_free(zone, bucket, udata);
3690 }
3691
3692 /*
3693 * Attempt to retrieve the item from the per-CPU cache has failed, so
3694 * we must go back to the zone. This requires the zdom lock, so we
3695 * must drop the critical section, then re-acquire it when we go back
3696 * to the cache. Since the critical section is released, we may be
3697 * preempted or migrate. As such, make sure not to maintain any
3698 * thread-local state specific to the cache from prior to releasing
3699 * the critical section.
3700 */
3701 domain = PCPU_GET(domain);
3702 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3703 VM_DOMAIN_EMPTY(domain))
3704 domain = zone_domain_highest(zone, domain);
3705 bucket = cache_fetch_bucket(zone, cache, domain);
3706 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3707 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3708 new = true;
3709 } else {
3710 new = false;
3711 }
3712
3713 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3714 zone->uz_name, zone, bucket);
3715 if (bucket == NULL) {
3716 critical_enter();
3717 return (false);
3718 }
3719
3720 /*
3721 * See if we lost the race or were migrated. Cache the
3722 * initialized bucket to make this less likely or claim
3723 * the memory directly.
3724 */
3725 critical_enter();
3726 cache = &zone->uz_cpu[curcpu];
3727 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3728 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3729 (curdomain = PCPU_GET(domain)) == domain ||
3730 VM_DOMAIN_EMPTY(curdomain))) {
3731 if (new)
3732 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3733 bucket->ub_cnt);
3734 cache_bucket_load_alloc(cache, bucket);
3735 return (true);
3736 }
3737
3738 /*
3739 * We lost the race, release this bucket and start over.
3740 */
3741 critical_exit();
3742 zone_put_bucket(zone, domain, bucket, udata, !new);
3743 critical_enter();
3744
3745 return (true);
3746 }
3747
3748 void *
uma_zalloc_domain(uma_zone_t zone,void * udata,int domain,int flags)3749 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3750 {
3751 #ifdef NUMA
3752 uma_bucket_t bucket;
3753 uma_zone_domain_t zdom;
3754 void *item;
3755 #endif
3756
3757 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3758 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3759
3760 /* This is the fast path allocation */
3761 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3762 zone->uz_name, zone, domain, flags);
3763
3764 if (flags & M_WAITOK) {
3765 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3766 "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3767 }
3768 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3769 ("uma_zalloc_domain: called with spinlock or critical section held"));
3770 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3771 ("uma_zalloc_domain: called with SMR zone."));
3772 #ifdef NUMA
3773 KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3774 ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3775
3776 if (vm_ndomains == 1)
3777 return (uma_zalloc_arg(zone, udata, flags));
3778
3779 /*
3780 * Try to allocate from the bucket cache before falling back to the keg.
3781 * We could try harder and attempt to allocate from per-CPU caches or
3782 * the per-domain cross-domain buckets, but the complexity is probably
3783 * not worth it. It is more important that frees of previous
3784 * cross-domain allocations do not blow up the cache.
3785 */
3786 zdom = zone_domain_lock(zone, domain);
3787 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3788 item = bucket->ub_bucket[bucket->ub_cnt - 1];
3789 #ifdef INVARIANTS
3790 bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3791 #endif
3792 bucket->ub_cnt--;
3793 zone_put_bucket(zone, domain, bucket, udata, true);
3794 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3795 flags, item);
3796 if (item != NULL) {
3797 KASSERT(item_domain(item) == domain,
3798 ("%s: bucket cache item %p from wrong domain",
3799 __func__, item));
3800 counter_u64_add(zone->uz_allocs, 1);
3801 }
3802 return (item);
3803 }
3804 ZDOM_UNLOCK(zdom);
3805 return (zone_alloc_item(zone, udata, domain, flags));
3806 #else
3807 return (uma_zalloc_arg(zone, udata, flags));
3808 #endif
3809 }
3810
3811 /*
3812 * Find a slab with some space. Prefer slabs that are partially used over those
3813 * that are totally full. This helps to reduce fragmentation.
3814 *
3815 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3816 * only 'domain'.
3817 */
3818 static uma_slab_t
keg_first_slab(uma_keg_t keg,int domain,bool rr)3819 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3820 {
3821 uma_domain_t dom;
3822 uma_slab_t slab;
3823 int start;
3824
3825 KASSERT(domain >= 0 && domain < vm_ndomains,
3826 ("keg_first_slab: domain %d out of range", domain));
3827 KEG_LOCK_ASSERT(keg, domain);
3828
3829 slab = NULL;
3830 start = domain;
3831 do {
3832 dom = &keg->uk_domain[domain];
3833 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3834 return (slab);
3835 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3836 LIST_REMOVE(slab, us_link);
3837 dom->ud_free_slabs--;
3838 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3839 return (slab);
3840 }
3841 if (rr)
3842 domain = (domain + 1) % vm_ndomains;
3843 } while (domain != start);
3844
3845 return (NULL);
3846 }
3847
3848 /*
3849 * Fetch an existing slab from a free or partial list. Returns with the
3850 * keg domain lock held if a slab was found or unlocked if not.
3851 */
3852 static uma_slab_t
keg_fetch_free_slab(uma_keg_t keg,int domain,bool rr,int flags)3853 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3854 {
3855 uma_slab_t slab;
3856 uint32_t reserve;
3857
3858 /* HASH has a single free list. */
3859 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3860 domain = 0;
3861
3862 KEG_LOCK(keg, domain);
3863 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3864 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3865 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3866 KEG_UNLOCK(keg, domain);
3867 return (NULL);
3868 }
3869 return (slab);
3870 }
3871
3872 static uma_slab_t
keg_fetch_slab(uma_keg_t keg,uma_zone_t zone,int rdomain,const int flags)3873 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3874 {
3875 struct vm_domainset_iter di;
3876 uma_slab_t slab;
3877 int aflags, domain;
3878 bool rr;
3879
3880 KASSERT((flags & (M_WAITOK | M_NOVM)) != (M_WAITOK | M_NOVM),
3881 ("%s: invalid flags %#x", __func__, flags));
3882
3883 restart:
3884 /*
3885 * Use the keg's policy if upper layers haven't already specified a
3886 * domain (as happens with first-touch zones).
3887 *
3888 * To avoid races we run the iterator with the keg lock held, but that
3889 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3890 * clear M_WAITOK and handle low memory conditions locally.
3891 */
3892 rr = rdomain == UMA_ANYDOMAIN;
3893 if (rr) {
3894 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3895 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3896 &aflags);
3897 } else {
3898 aflags = flags;
3899 domain = rdomain;
3900 }
3901
3902 for (;;) {
3903 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3904 if (slab != NULL)
3905 return (slab);
3906
3907 /*
3908 * M_NOVM is used to break the recursion that can otherwise
3909 * occur if low-level memory management routines use UMA.
3910 */
3911 if ((flags & M_NOVM) == 0) {
3912 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3913 if (slab != NULL)
3914 return (slab);
3915 }
3916
3917 if (!rr) {
3918 if ((flags & M_USE_RESERVE) != 0) {
3919 /*
3920 * Drain reserves from other domains before
3921 * giving up or sleeping. It may be useful to
3922 * support per-domain reserves eventually.
3923 */
3924 rdomain = UMA_ANYDOMAIN;
3925 goto restart;
3926 }
3927 if ((flags & M_WAITOK) == 0)
3928 break;
3929 vm_wait_domain(domain);
3930 } else if (vm_domainset_iter_policy(&di, &domain) != 0) {
3931 if ((flags & M_WAITOK) != 0) {
3932 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
3933 goto restart;
3934 }
3935 break;
3936 }
3937 }
3938
3939 /*
3940 * We might not have been able to get a slab but another cpu
3941 * could have while we were unlocked. Check again before we
3942 * fail.
3943 */
3944 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3945 return (slab);
3946
3947 return (NULL);
3948 }
3949
3950 static void *
slab_alloc_item(uma_keg_t keg,uma_slab_t slab)3951 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3952 {
3953 uma_domain_t dom;
3954 void *item;
3955 int freei;
3956
3957 KEG_LOCK_ASSERT(keg, slab->us_domain);
3958
3959 dom = &keg->uk_domain[slab->us_domain];
3960 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3961 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3962 item = slab_item(slab, keg, freei);
3963 slab->us_freecount--;
3964 dom->ud_free_items--;
3965
3966 /*
3967 * Move this slab to the full list. It must be on the partial list, so
3968 * we do not need to update the free slab count. In particular,
3969 * keg_fetch_slab() always returns slabs on the partial list.
3970 */
3971 if (slab->us_freecount == 0) {
3972 LIST_REMOVE(slab, us_link);
3973 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3974 }
3975
3976 return (item);
3977 }
3978
3979 static int
zone_import(void * arg,void ** bucket,int max,int domain,int flags)3980 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3981 {
3982 uma_domain_t dom;
3983 uma_zone_t zone;
3984 uma_slab_t slab;
3985 uma_keg_t keg;
3986 #ifdef NUMA
3987 int stripe;
3988 #endif
3989 int i;
3990
3991 zone = arg;
3992 slab = NULL;
3993 keg = zone->uz_keg;
3994 /* Try to keep the buckets totally full */
3995 for (i = 0; i < max; ) {
3996 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3997 break;
3998 #ifdef NUMA
3999 stripe = howmany(max, vm_ndomains);
4000 #endif
4001 dom = &keg->uk_domain[slab->us_domain];
4002 do {
4003 bucket[i++] = slab_alloc_item(keg, slab);
4004 if (keg->uk_reserve > 0 &&
4005 dom->ud_free_items <= keg->uk_reserve) {
4006 /*
4007 * Avoid depleting the reserve after a
4008 * successful item allocation, even if
4009 * M_USE_RESERVE is specified.
4010 */
4011 KEG_UNLOCK(keg, slab->us_domain);
4012 goto out;
4013 }
4014 #ifdef NUMA
4015 /*
4016 * If the zone is striped we pick a new slab for every
4017 * N allocations. Eliminating this conditional will
4018 * instead pick a new domain for each bucket rather
4019 * than stripe within each bucket. The current option
4020 * produces more fragmentation and requires more cpu
4021 * time but yields better distribution.
4022 */
4023 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
4024 vm_ndomains > 1 && --stripe == 0)
4025 break;
4026 #endif
4027 } while (slab->us_freecount != 0 && i < max);
4028 KEG_UNLOCK(keg, slab->us_domain);
4029
4030 /* Don't block if we allocated any successfully. */
4031 flags &= ~M_WAITOK;
4032 flags |= M_NOWAIT;
4033 }
4034 out:
4035 return i;
4036 }
4037
4038 static int
zone_alloc_limit_hard(uma_zone_t zone,int count,int flags)4039 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
4040 {
4041 uint64_t old, new, total, max;
4042
4043 /*
4044 * The hard case. We're going to sleep because there were existing
4045 * sleepers or because we ran out of items. This routine enforces
4046 * fairness by keeping fifo order.
4047 *
4048 * First release our ill gotten gains and make some noise.
4049 */
4050 for (;;) {
4051 zone_free_limit(zone, count);
4052 zone_log_warning(zone);
4053 zone_maxaction(zone);
4054 if (flags & M_NOWAIT)
4055 return (0);
4056
4057 /*
4058 * We need to allocate an item or set ourself as a sleeper
4059 * while the sleepq lock is held to avoid wakeup races. This
4060 * is essentially a home rolled semaphore.
4061 */
4062 sleepq_lock(&zone->uz_max_items);
4063 old = zone->uz_items;
4064 do {
4065 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
4066 /* Cache the max since we will evaluate twice. */
4067 max = zone->uz_max_items;
4068 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
4069 UZ_ITEMS_COUNT(old) >= max)
4070 new = old + UZ_ITEMS_SLEEPER;
4071 else
4072 new = old + MIN(count, max - old);
4073 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
4074
4075 /* We may have successfully allocated under the sleepq lock. */
4076 if (UZ_ITEMS_SLEEPERS(new) == 0) {
4077 sleepq_release(&zone->uz_max_items);
4078 return (new - old);
4079 }
4080
4081 /*
4082 * This is in a different cacheline from uz_items so that we
4083 * don't constantly invalidate the fastpath cacheline when we
4084 * adjust item counts. This could be limited to toggling on
4085 * transitions.
4086 */
4087 atomic_add_32(&zone->uz_sleepers, 1);
4088 atomic_add_64(&zone->uz_sleeps, 1);
4089
4090 /*
4091 * We have added ourselves as a sleeper. The sleepq lock
4092 * protects us from wakeup races. Sleep now and then retry.
4093 */
4094 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
4095 sleepq_wait(&zone->uz_max_items, PVM);
4096
4097 /*
4098 * After wakeup, remove ourselves as a sleeper and try
4099 * again. We no longer have the sleepq lock for protection.
4100 *
4101 * Subract ourselves as a sleeper while attempting to add
4102 * our count.
4103 */
4104 atomic_subtract_32(&zone->uz_sleepers, 1);
4105 old = atomic_fetchadd_64(&zone->uz_items,
4106 -(UZ_ITEMS_SLEEPER - count));
4107 /* We're no longer a sleeper. */
4108 old -= UZ_ITEMS_SLEEPER;
4109
4110 /*
4111 * If we're still at the limit, restart. Notably do not
4112 * block on other sleepers. Cache the max value to protect
4113 * against changes via sysctl.
4114 */
4115 total = UZ_ITEMS_COUNT(old);
4116 max = zone->uz_max_items;
4117 if (total >= max)
4118 continue;
4119 /* Truncate if necessary, otherwise wake other sleepers. */
4120 if (total + count > max) {
4121 zone_free_limit(zone, total + count - max);
4122 count = max - total;
4123 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
4124 wakeup_one(&zone->uz_max_items);
4125
4126 return (count);
4127 }
4128 }
4129
4130 /*
4131 * Allocate 'count' items from our max_items limit. Returns the number
4132 * available. If M_NOWAIT is not specified it will sleep until at least
4133 * one item can be allocated.
4134 */
4135 static int
zone_alloc_limit(uma_zone_t zone,int count,int flags)4136 zone_alloc_limit(uma_zone_t zone, int count, int flags)
4137 {
4138 uint64_t old;
4139 uint64_t max;
4140
4141 max = zone->uz_max_items;
4142 MPASS(max > 0);
4143
4144 /*
4145 * We expect normal allocations to succeed with a simple
4146 * fetchadd.
4147 */
4148 old = atomic_fetchadd_64(&zone->uz_items, count);
4149 if (__predict_true(old + count <= max))
4150 return (count);
4151
4152 /*
4153 * If we had some items and no sleepers just return the
4154 * truncated value. We have to release the excess space
4155 * though because that may wake sleepers who weren't woken
4156 * because we were temporarily over the limit.
4157 */
4158 if (old < max) {
4159 zone_free_limit(zone, (old + count) - max);
4160 return (max - old);
4161 }
4162 return (zone_alloc_limit_hard(zone, count, flags));
4163 }
4164
4165 /*
4166 * Free a number of items back to the limit.
4167 */
4168 static void
zone_free_limit(uma_zone_t zone,int count)4169 zone_free_limit(uma_zone_t zone, int count)
4170 {
4171 uint64_t old;
4172
4173 MPASS(count > 0);
4174
4175 /*
4176 * In the common case we either have no sleepers or
4177 * are still over the limit and can just return.
4178 */
4179 old = atomic_fetchadd_64(&zone->uz_items, -count);
4180 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
4181 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
4182 return;
4183
4184 /*
4185 * Moderate the rate of wakeups. Sleepers will continue
4186 * to generate wakeups if necessary.
4187 */
4188 wakeup_one(&zone->uz_max_items);
4189 }
4190
4191 static uma_bucket_t
zone_alloc_bucket(uma_zone_t zone,void * udata,int domain,int flags)4192 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
4193 {
4194 uma_bucket_t bucket;
4195 int error, maxbucket, cnt;
4196
4197 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
4198 zone, domain);
4199
4200 /* Avoid allocs targeting empty domains. */
4201 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4202 domain = UMA_ANYDOMAIN;
4203 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4204 domain = UMA_ANYDOMAIN;
4205
4206 if (zone->uz_max_items > 0)
4207 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
4208 M_NOWAIT);
4209 else
4210 maxbucket = zone->uz_bucket_size;
4211 if (maxbucket == 0)
4212 return (NULL);
4213
4214 /* Don't wait for buckets, preserve caller's NOVM setting. */
4215 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
4216 if (bucket == NULL) {
4217 cnt = 0;
4218 goto out;
4219 }
4220
4221 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
4222 MIN(maxbucket, bucket->ub_entries), domain, flags);
4223
4224 /*
4225 * Initialize the memory if necessary.
4226 */
4227 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4228 int i;
4229
4230 for (i = 0; i < bucket->ub_cnt; i++) {
4231 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
4232 error = zone->uz_init(bucket->ub_bucket[i],
4233 zone->uz_size, flags);
4234 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
4235 if (error != 0)
4236 break;
4237 }
4238
4239 /*
4240 * If we couldn't initialize the whole bucket, put the
4241 * rest back onto the freelist.
4242 */
4243 if (i != bucket->ub_cnt) {
4244 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4245 bucket->ub_cnt - i);
4246 #ifdef INVARIANTS
4247 bzero(&bucket->ub_bucket[i],
4248 sizeof(void *) * (bucket->ub_cnt - i));
4249 #endif
4250 bucket->ub_cnt = i;
4251 }
4252 }
4253
4254 cnt = bucket->ub_cnt;
4255 if (bucket->ub_cnt == 0) {
4256 bucket_free(zone, bucket, udata);
4257 counter_u64_add(zone->uz_fails, 1);
4258 bucket = NULL;
4259 }
4260 out:
4261 if (zone->uz_max_items > 0 && cnt < maxbucket)
4262 zone_free_limit(zone, maxbucket - cnt);
4263
4264 return (bucket);
4265 }
4266
4267 /*
4268 * Allocates a single item from a zone.
4269 *
4270 * Arguments
4271 * zone The zone to alloc for.
4272 * udata The data to be passed to the constructor.
4273 * domain The domain to allocate from or UMA_ANYDOMAIN.
4274 * flags M_WAITOK, M_NOWAIT, M_ZERO.
4275 *
4276 * Returns
4277 * NULL if there is no memory and M_NOWAIT is set
4278 * An item if successful
4279 */
4280
4281 static void *
zone_alloc_item(uma_zone_t zone,void * udata,int domain,int flags)4282 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4283 {
4284 void *item;
4285
4286 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4287 counter_u64_add(zone->uz_fails, 1);
4288 return (NULL);
4289 }
4290
4291 /* Avoid allocs targeting empty domains. */
4292 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4293 domain = UMA_ANYDOMAIN;
4294
4295 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4296 goto fail_cnt;
4297
4298 /*
4299 * We have to call both the zone's init (not the keg's init)
4300 * and the zone's ctor. This is because the item is going from
4301 * a keg slab directly to the user, and the user is expecting it
4302 * to be both zone-init'd as well as zone-ctor'd.
4303 */
4304 if (zone->uz_init != NULL) {
4305 int error;
4306
4307 kasan_mark_item_valid(zone, item);
4308 error = zone->uz_init(item, zone->uz_size, flags);
4309 kasan_mark_item_invalid(zone, item);
4310 if (error != 0) {
4311 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4312 goto fail_cnt;
4313 }
4314 }
4315 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4316 item);
4317 if (item == NULL)
4318 goto fail;
4319
4320 counter_u64_add(zone->uz_allocs, 1);
4321 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4322 zone->uz_name, zone);
4323
4324 return (item);
4325
4326 fail_cnt:
4327 counter_u64_add(zone->uz_fails, 1);
4328 fail:
4329 if (zone->uz_max_items > 0)
4330 zone_free_limit(zone, 1);
4331 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4332 zone->uz_name, zone);
4333
4334 return (NULL);
4335 }
4336
4337 /* See uma.h */
4338 void
uma_zfree_smr(uma_zone_t zone,void * item)4339 uma_zfree_smr(uma_zone_t zone, void *item)
4340 {
4341 uma_cache_t cache;
4342 uma_cache_bucket_t bucket;
4343 int itemdomain, uz_flags;
4344
4345 #ifdef UMA_ZALLOC_DEBUG
4346 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4347 ("uma_zfree_smr: called with non-SMR zone."));
4348 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4349 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4350 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4351 return;
4352 #endif
4353 cache = &zone->uz_cpu[curcpu];
4354 uz_flags = cache_uz_flags(cache);
4355 itemdomain = 0;
4356 #ifdef NUMA
4357 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4358 itemdomain = item_domain(item);
4359 #endif
4360 critical_enter();
4361 do {
4362 cache = &zone->uz_cpu[curcpu];
4363 /* SMR Zones must free to the free bucket. */
4364 bucket = &cache->uc_freebucket;
4365 #ifdef NUMA
4366 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4367 PCPU_GET(domain) != itemdomain) {
4368 bucket = &cache->uc_crossbucket;
4369 }
4370 #endif
4371 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4372 cache_bucket_push(cache, bucket, item);
4373 critical_exit();
4374 return;
4375 }
4376 } while (cache_free(zone, cache, NULL, item, itemdomain));
4377 critical_exit();
4378
4379 /*
4380 * If nothing else caught this, we'll just do an internal free.
4381 */
4382 zone_free_item(zone, item, NULL, SKIP_NONE);
4383 }
4384
4385 /* See uma.h */
4386 void
uma_zfree_arg(uma_zone_t zone,void * item,void * udata)4387 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4388 {
4389 uma_cache_t cache;
4390 uma_cache_bucket_t bucket;
4391 int itemdomain, uz_flags;
4392
4393 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4394 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4395
4396 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4397
4398 #ifdef UMA_ZALLOC_DEBUG
4399 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4400 ("uma_zfree_arg: called with SMR zone."));
4401 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4402 return;
4403 #endif
4404 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4405 if (item == NULL)
4406 return;
4407
4408 /*
4409 * We are accessing the per-cpu cache without a critical section to
4410 * fetch size and flags. This is acceptable, if we are preempted we
4411 * will simply read another cpu's line.
4412 */
4413 cache = &zone->uz_cpu[curcpu];
4414 uz_flags = cache_uz_flags(cache);
4415 if (UMA_ALWAYS_CTORDTOR ||
4416 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4417 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4418
4419 /*
4420 * The race here is acceptable. If we miss it we'll just have to wait
4421 * a little longer for the limits to be reset.
4422 */
4423 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4424 if (atomic_load_32(&zone->uz_sleepers) > 0)
4425 goto zfree_item;
4426 }
4427
4428 /*
4429 * If possible, free to the per-CPU cache. There are two
4430 * requirements for safe access to the per-CPU cache: (1) the thread
4431 * accessing the cache must not be preempted or yield during access,
4432 * and (2) the thread must not migrate CPUs without switching which
4433 * cache it accesses. We rely on a critical section to prevent
4434 * preemption and migration. We release the critical section in
4435 * order to acquire the zone mutex if we are unable to free to the
4436 * current cache; when we re-acquire the critical section, we must
4437 * detect and handle migration if it has occurred.
4438 */
4439 itemdomain = 0;
4440 #ifdef NUMA
4441 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4442 itemdomain = item_domain(item);
4443 #endif
4444 critical_enter();
4445 do {
4446 cache = &zone->uz_cpu[curcpu];
4447 /*
4448 * Try to free into the allocbucket first to give LIFO
4449 * ordering for cache-hot datastructures. Spill over
4450 * into the freebucket if necessary. Alloc will swap
4451 * them if one runs dry.
4452 */
4453 bucket = &cache->uc_allocbucket;
4454 #ifdef NUMA
4455 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4456 PCPU_GET(domain) != itemdomain) {
4457 bucket = &cache->uc_crossbucket;
4458 } else
4459 #endif
4460 if (bucket->ucb_cnt == bucket->ucb_entries &&
4461 cache->uc_freebucket.ucb_cnt <
4462 cache->uc_freebucket.ucb_entries)
4463 cache_bucket_swap(&cache->uc_freebucket,
4464 &cache->uc_allocbucket);
4465 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4466 cache_bucket_push(cache, bucket, item);
4467 critical_exit();
4468 return;
4469 }
4470 } while (cache_free(zone, cache, udata, item, itemdomain));
4471 critical_exit();
4472
4473 /*
4474 * If nothing else caught this, we'll just do an internal free.
4475 */
4476 zfree_item:
4477 zone_free_item(zone, item, udata, SKIP_DTOR);
4478 }
4479
4480 #ifdef NUMA
4481 /*
4482 * sort crossdomain free buckets to domain correct buckets and cache
4483 * them.
4484 */
4485 static void
zone_free_cross(uma_zone_t zone,uma_bucket_t bucket,void * udata)4486 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4487 {
4488 struct uma_bucketlist emptybuckets, fullbuckets;
4489 uma_zone_domain_t zdom;
4490 uma_bucket_t b;
4491 smr_seq_t seq;
4492 void *item;
4493 int domain;
4494
4495 CTR3(KTR_UMA,
4496 "uma_zfree: zone %s(%p) draining cross bucket %p",
4497 zone->uz_name, zone, bucket);
4498
4499 /*
4500 * It is possible for buckets to arrive here out of order so we fetch
4501 * the current smr seq rather than accepting the bucket's.
4502 */
4503 seq = SMR_SEQ_INVALID;
4504 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4505 seq = smr_advance(zone->uz_smr);
4506
4507 /*
4508 * To avoid having ndomain * ndomain buckets for sorting we have a
4509 * lock on the current crossfree bucket. A full matrix with
4510 * per-domain locking could be used if necessary.
4511 */
4512 STAILQ_INIT(&emptybuckets);
4513 STAILQ_INIT(&fullbuckets);
4514 ZONE_CROSS_LOCK(zone);
4515 for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4516 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4517 domain = item_domain(item);
4518 zdom = ZDOM_GET(zone, domain);
4519 if (zdom->uzd_cross == NULL) {
4520 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4521 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4522 zdom->uzd_cross = b;
4523 } else {
4524 /*
4525 * Avoid allocating a bucket with the cross lock
4526 * held, since allocation can trigger a
4527 * cross-domain free and bucket zones may
4528 * allocate from each other.
4529 */
4530 ZONE_CROSS_UNLOCK(zone);
4531 b = bucket_alloc(zone, udata, M_NOWAIT);
4532 if (b == NULL)
4533 goto out;
4534 ZONE_CROSS_LOCK(zone);
4535 if (zdom->uzd_cross != NULL) {
4536 STAILQ_INSERT_HEAD(&emptybuckets, b,
4537 ub_link);
4538 } else {
4539 zdom->uzd_cross = b;
4540 }
4541 }
4542 }
4543 b = zdom->uzd_cross;
4544 b->ub_bucket[b->ub_cnt++] = item;
4545 b->ub_seq = seq;
4546 if (b->ub_cnt == b->ub_entries) {
4547 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4548 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4549 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4550 zdom->uzd_cross = b;
4551 }
4552 }
4553 ZONE_CROSS_UNLOCK(zone);
4554 out:
4555 if (bucket->ub_cnt == 0)
4556 bucket->ub_seq = SMR_SEQ_INVALID;
4557 bucket_free(zone, bucket, udata);
4558
4559 while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4560 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4561 bucket_free(zone, b, udata);
4562 }
4563 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4564 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4565 domain = item_domain(b->ub_bucket[0]);
4566 zone_put_bucket(zone, domain, b, udata, true);
4567 }
4568 }
4569 #endif
4570
4571 static void
zone_free_bucket(uma_zone_t zone,uma_bucket_t bucket,void * udata,int itemdomain,bool ws)4572 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4573 int itemdomain, bool ws)
4574 {
4575
4576 #ifdef NUMA
4577 /*
4578 * Buckets coming from the wrong domain will be entirely for the
4579 * only other domain on two domain systems. In this case we can
4580 * simply cache them. Otherwise we need to sort them back to
4581 * correct domains.
4582 */
4583 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4584 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4585 zone_free_cross(zone, bucket, udata);
4586 return;
4587 }
4588 #endif
4589
4590 /*
4591 * Attempt to save the bucket in the zone's domain bucket cache.
4592 */
4593 CTR3(KTR_UMA,
4594 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4595 zone->uz_name, zone, bucket);
4596 /* ub_cnt is pointing to the last free item */
4597 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4598 itemdomain = zone_domain_lowest(zone, itemdomain);
4599 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4600 }
4601
4602 /*
4603 * Populate a free or cross bucket for the current cpu cache. Free any
4604 * existing full bucket either to the zone cache or back to the slab layer.
4605 *
4606 * Enters and returns in a critical section. false return indicates that
4607 * we can not satisfy this free in the cache layer. true indicates that
4608 * the caller should retry.
4609 */
4610 static __noinline bool
cache_free(uma_zone_t zone,uma_cache_t cache,void * udata,void * item,int itemdomain)4611 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4612 int itemdomain)
4613 {
4614 uma_cache_bucket_t cbucket;
4615 uma_bucket_t newbucket, bucket;
4616
4617 CRITICAL_ASSERT(curthread);
4618
4619 if (zone->uz_bucket_size == 0)
4620 return false;
4621
4622 cache = &zone->uz_cpu[curcpu];
4623 newbucket = NULL;
4624
4625 /*
4626 * FIRSTTOUCH domains need to free to the correct zdom. When
4627 * enabled this is the zdom of the item. The bucket is the
4628 * cross bucket if the current domain and itemdomain do not match.
4629 */
4630 cbucket = &cache->uc_freebucket;
4631 #ifdef NUMA
4632 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4633 if (PCPU_GET(domain) != itemdomain) {
4634 cbucket = &cache->uc_crossbucket;
4635 if (cbucket->ucb_cnt != 0)
4636 counter_u64_add(zone->uz_xdomain,
4637 cbucket->ucb_cnt);
4638 }
4639 }
4640 #endif
4641 bucket = cache_bucket_unload(cbucket);
4642 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4643 ("cache_free: Entered with non-full free bucket."));
4644
4645 /* We are no longer associated with this CPU. */
4646 critical_exit();
4647
4648 /*
4649 * Don't let SMR zones operate without a free bucket. Force
4650 * a synchronize and re-use this one. We will only degrade
4651 * to a synchronize every bucket_size items rather than every
4652 * item if we fail to allocate a bucket.
4653 */
4654 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4655 if (bucket != NULL)
4656 bucket->ub_seq = smr_advance(zone->uz_smr);
4657 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4658 if (newbucket == NULL && bucket != NULL) {
4659 bucket_drain(zone, bucket);
4660 newbucket = bucket;
4661 bucket = NULL;
4662 }
4663 } else if (!bucketdisable)
4664 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4665
4666 if (bucket != NULL)
4667 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4668
4669 critical_enter();
4670 if ((bucket = newbucket) == NULL)
4671 return (false);
4672 cache = &zone->uz_cpu[curcpu];
4673 #ifdef NUMA
4674 /*
4675 * Check to see if we should be populating the cross bucket. If it
4676 * is already populated we will fall through and attempt to populate
4677 * the free bucket.
4678 */
4679 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4680 if (PCPU_GET(domain) != itemdomain &&
4681 cache->uc_crossbucket.ucb_bucket == NULL) {
4682 cache_bucket_load_cross(cache, bucket);
4683 return (true);
4684 }
4685 }
4686 #endif
4687 /*
4688 * We may have lost the race to fill the bucket or switched CPUs.
4689 */
4690 if (cache->uc_freebucket.ucb_bucket != NULL) {
4691 critical_exit();
4692 bucket_free(zone, bucket, udata);
4693 critical_enter();
4694 } else
4695 cache_bucket_load_free(cache, bucket);
4696
4697 return (true);
4698 }
4699
4700 static void
slab_free_item(uma_zone_t zone,uma_slab_t slab,void * item)4701 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4702 {
4703 uma_keg_t keg;
4704 uma_domain_t dom;
4705 int freei;
4706
4707 keg = zone->uz_keg;
4708 KEG_LOCK_ASSERT(keg, slab->us_domain);
4709
4710 /* Do we need to remove from any lists? */
4711 dom = &keg->uk_domain[slab->us_domain];
4712 if (slab->us_freecount + 1 == keg->uk_ipers) {
4713 LIST_REMOVE(slab, us_link);
4714 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4715 dom->ud_free_slabs++;
4716 } else if (slab->us_freecount == 0) {
4717 LIST_REMOVE(slab, us_link);
4718 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4719 }
4720
4721 /* Slab management. */
4722 freei = slab_item_index(slab, keg, item);
4723 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4724 slab->us_freecount++;
4725
4726 /* Keg statistics. */
4727 dom->ud_free_items++;
4728 }
4729
4730 static void
zone_release(void * arg,void ** bucket,int cnt)4731 zone_release(void *arg, void **bucket, int cnt)
4732 {
4733 struct mtx *lock;
4734 uma_zone_t zone;
4735 uma_slab_t slab;
4736 uma_keg_t keg;
4737 uint8_t *mem;
4738 void *item;
4739 int i;
4740
4741 zone = arg;
4742 keg = zone->uz_keg;
4743 lock = NULL;
4744 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4745 lock = KEG_LOCK(keg, 0);
4746 for (i = 0; i < cnt; i++) {
4747 item = bucket[i];
4748 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4749 slab = vtoslab((vm_offset_t)item);
4750 } else {
4751 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4752 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4753 slab = hash_sfind(&keg->uk_hash, mem);
4754 else
4755 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4756 }
4757 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4758 if (lock != NULL)
4759 mtx_unlock(lock);
4760 lock = KEG_LOCK(keg, slab->us_domain);
4761 }
4762 slab_free_item(zone, slab, item);
4763 }
4764 if (lock != NULL)
4765 mtx_unlock(lock);
4766 }
4767
4768 /*
4769 * Frees a single item to any zone.
4770 *
4771 * Arguments:
4772 * zone The zone to free to
4773 * item The item we're freeing
4774 * udata User supplied data for the dtor
4775 * skip Skip dtors and finis
4776 */
4777 static __noinline void
zone_free_item(uma_zone_t zone,void * item,void * udata,enum zfreeskip skip)4778 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4779 {
4780
4781 /*
4782 * If a free is sent directly to an SMR zone we have to
4783 * synchronize immediately because the item can instantly
4784 * be reallocated. This should only happen in degenerate
4785 * cases when no memory is available for per-cpu caches.
4786 */
4787 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4788 smr_synchronize(zone->uz_smr);
4789
4790 item_dtor(zone, item, zone->uz_size, udata, skip);
4791
4792 if (skip < SKIP_FINI && zone->uz_fini) {
4793 kasan_mark_item_valid(zone, item);
4794 zone->uz_fini(item, zone->uz_size);
4795 kasan_mark_item_invalid(zone, item);
4796 }
4797
4798 zone->uz_release(zone->uz_arg, &item, 1);
4799
4800 if (skip & SKIP_CNT)
4801 return;
4802
4803 counter_u64_add(zone->uz_frees, 1);
4804
4805 if (zone->uz_max_items > 0)
4806 zone_free_limit(zone, 1);
4807 }
4808
4809 /* See uma.h */
4810 int
uma_zone_set_max(uma_zone_t zone,int nitems)4811 uma_zone_set_max(uma_zone_t zone, int nitems)
4812 {
4813
4814 /*
4815 * If the limit is small, we may need to constrain the maximum per-CPU
4816 * cache size, or disable caching entirely.
4817 */
4818 uma_zone_set_maxcache(zone, nitems);
4819
4820 /*
4821 * XXX This can misbehave if the zone has any allocations with
4822 * no limit and a limit is imposed. There is currently no
4823 * way to clear a limit.
4824 */
4825 ZONE_LOCK(zone);
4826 if (zone->uz_max_items == 0)
4827 ZONE_ASSERT_COLD(zone);
4828 zone->uz_max_items = nitems;
4829 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4830 zone_update_caches(zone);
4831 /* We may need to wake waiters. */
4832 wakeup(&zone->uz_max_items);
4833 ZONE_UNLOCK(zone);
4834
4835 return (nitems);
4836 }
4837
4838 /* See uma.h */
4839 void
uma_zone_set_maxcache(uma_zone_t zone,int nitems)4840 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4841 {
4842 int bpcpu, bpdom, bsize, nb;
4843
4844 ZONE_LOCK(zone);
4845
4846 /*
4847 * Compute a lower bound on the number of items that may be cached in
4848 * the zone. Each CPU gets at least two buckets, and for cross-domain
4849 * frees we use an additional bucket per CPU and per domain. Select the
4850 * largest bucket size that does not exceed half of the requested limit,
4851 * with the left over space given to the full bucket cache.
4852 */
4853 bpdom = 0;
4854 bpcpu = 2;
4855 #ifdef NUMA
4856 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
4857 bpcpu++;
4858 bpdom++;
4859 }
4860 #endif
4861 nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
4862 bsize = nitems / nb / 2;
4863 if (bsize > BUCKET_MAX)
4864 bsize = BUCKET_MAX;
4865 else if (bsize == 0 && nitems / nb > 0)
4866 bsize = 1;
4867 zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
4868 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4869 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4870 zone->uz_bucket_max = nitems - nb * bsize;
4871 ZONE_UNLOCK(zone);
4872 }
4873
4874 /* See uma.h */
4875 int
uma_zone_get_max(uma_zone_t zone)4876 uma_zone_get_max(uma_zone_t zone)
4877 {
4878 int nitems;
4879
4880 nitems = atomic_load_64(&zone->uz_max_items);
4881
4882 return (nitems);
4883 }
4884
4885 /* See uma.h */
4886 void
uma_zone_set_warning(uma_zone_t zone,const char * warning)4887 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4888 {
4889
4890 ZONE_ASSERT_COLD(zone);
4891 zone->uz_warning = warning;
4892 }
4893
4894 /* See uma.h */
4895 void
uma_zone_set_maxaction(uma_zone_t zone,uma_maxaction_t maxaction)4896 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4897 {
4898
4899 ZONE_ASSERT_COLD(zone);
4900 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4901 }
4902
4903 /* See uma.h */
4904 int
uma_zone_get_cur(uma_zone_t zone)4905 uma_zone_get_cur(uma_zone_t zone)
4906 {
4907 int64_t nitems;
4908 u_int i;
4909
4910 nitems = 0;
4911 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4912 nitems = counter_u64_fetch(zone->uz_allocs) -
4913 counter_u64_fetch(zone->uz_frees);
4914 CPU_FOREACH(i)
4915 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4916 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4917
4918 return (nitems < 0 ? 0 : nitems);
4919 }
4920
4921 static uint64_t
uma_zone_get_allocs(uma_zone_t zone)4922 uma_zone_get_allocs(uma_zone_t zone)
4923 {
4924 uint64_t nitems;
4925 u_int i;
4926
4927 nitems = 0;
4928 if (zone->uz_allocs != EARLY_COUNTER)
4929 nitems = counter_u64_fetch(zone->uz_allocs);
4930 CPU_FOREACH(i)
4931 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4932
4933 return (nitems);
4934 }
4935
4936 static uint64_t
uma_zone_get_frees(uma_zone_t zone)4937 uma_zone_get_frees(uma_zone_t zone)
4938 {
4939 uint64_t nitems;
4940 u_int i;
4941
4942 nitems = 0;
4943 if (zone->uz_frees != EARLY_COUNTER)
4944 nitems = counter_u64_fetch(zone->uz_frees);
4945 CPU_FOREACH(i)
4946 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4947
4948 return (nitems);
4949 }
4950
4951 #ifdef INVARIANTS
4952 /* Used only for KEG_ASSERT_COLD(). */
4953 static uint64_t
uma_keg_get_allocs(uma_keg_t keg)4954 uma_keg_get_allocs(uma_keg_t keg)
4955 {
4956 uma_zone_t z;
4957 uint64_t nitems;
4958
4959 nitems = 0;
4960 LIST_FOREACH(z, &keg->uk_zones, uz_link)
4961 nitems += uma_zone_get_allocs(z);
4962
4963 return (nitems);
4964 }
4965 #endif
4966
4967 /* See uma.h */
4968 void
uma_zone_set_init(uma_zone_t zone,uma_init uminit)4969 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4970 {
4971 uma_keg_t keg;
4972
4973 KEG_GET(zone, keg);
4974 KEG_ASSERT_COLD(keg);
4975 keg->uk_init = uminit;
4976 }
4977
4978 /* See uma.h */
4979 void
uma_zone_set_fini(uma_zone_t zone,uma_fini fini)4980 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4981 {
4982 uma_keg_t keg;
4983
4984 KEG_GET(zone, keg);
4985 KEG_ASSERT_COLD(keg);
4986 keg->uk_fini = fini;
4987 }
4988
4989 /* See uma.h */
4990 void
uma_zone_set_zinit(uma_zone_t zone,uma_init zinit)4991 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4992 {
4993
4994 ZONE_ASSERT_COLD(zone);
4995 zone->uz_init = zinit;
4996 }
4997
4998 /* See uma.h */
4999 void
uma_zone_set_zfini(uma_zone_t zone,uma_fini zfini)5000 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
5001 {
5002
5003 ZONE_ASSERT_COLD(zone);
5004 zone->uz_fini = zfini;
5005 }
5006
5007 /* See uma.h */
5008 void
uma_zone_set_freef(uma_zone_t zone,uma_free freef)5009 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
5010 {
5011 uma_keg_t keg;
5012
5013 KEG_GET(zone, keg);
5014 KEG_ASSERT_COLD(keg);
5015 keg->uk_freef = freef;
5016 }
5017
5018 /* See uma.h */
5019 void
uma_zone_set_allocf(uma_zone_t zone,uma_alloc allocf)5020 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
5021 {
5022 uma_keg_t keg;
5023
5024 KEG_GET(zone, keg);
5025 KEG_ASSERT_COLD(keg);
5026 keg->uk_allocf = allocf;
5027 }
5028
5029 /* See uma.h */
5030 void
uma_zone_set_smr(uma_zone_t zone,smr_t smr)5031 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
5032 {
5033
5034 ZONE_ASSERT_COLD(zone);
5035
5036 KASSERT(smr != NULL, ("Got NULL smr"));
5037 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
5038 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
5039 zone->uz_flags |= UMA_ZONE_SMR;
5040 zone->uz_smr = smr;
5041 zone_update_caches(zone);
5042 }
5043
5044 smr_t
uma_zone_get_smr(uma_zone_t zone)5045 uma_zone_get_smr(uma_zone_t zone)
5046 {
5047
5048 return (zone->uz_smr);
5049 }
5050
5051 /* See uma.h */
5052 void
uma_zone_reserve(uma_zone_t zone,int items)5053 uma_zone_reserve(uma_zone_t zone, int items)
5054 {
5055 uma_keg_t keg;
5056
5057 KEG_GET(zone, keg);
5058 KEG_ASSERT_COLD(keg);
5059 keg->uk_reserve = items;
5060 }
5061
5062 /* See uma.h */
5063 int
uma_zone_reserve_kva(uma_zone_t zone,int count)5064 uma_zone_reserve_kva(uma_zone_t zone, int count)
5065 {
5066 uma_keg_t keg;
5067 vm_offset_t kva;
5068 u_int pages;
5069
5070 KEG_GET(zone, keg);
5071 KEG_ASSERT_COLD(keg);
5072 ZONE_ASSERT_COLD(zone);
5073
5074 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
5075
5076 #ifdef UMA_MD_SMALL_ALLOC
5077 if (keg->uk_ppera > 1) {
5078 #else
5079 if (1) {
5080 #endif
5081 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
5082 if (kva == 0)
5083 return (0);
5084 } else
5085 kva = 0;
5086
5087 MPASS(keg->uk_kva == 0);
5088 keg->uk_kva = kva;
5089 keg->uk_offset = 0;
5090 zone->uz_max_items = pages * keg->uk_ipers;
5091 #ifdef UMA_MD_SMALL_ALLOC
5092 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
5093 #else
5094 keg->uk_allocf = noobj_alloc;
5095 #endif
5096 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5097 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5098 zone_update_caches(zone);
5099
5100 return (1);
5101 }
5102
5103 /* See uma.h */
5104 void
5105 uma_prealloc(uma_zone_t zone, int items)
5106 {
5107 struct vm_domainset_iter di;
5108 uma_domain_t dom;
5109 uma_slab_t slab;
5110 uma_keg_t keg;
5111 int aflags, domain, slabs;
5112
5113 KEG_GET(zone, keg);
5114 slabs = howmany(items, keg->uk_ipers);
5115 while (slabs-- > 0) {
5116 aflags = M_NOWAIT;
5117 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
5118 &aflags);
5119 for (;;) {
5120 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
5121 aflags);
5122 if (slab != NULL) {
5123 dom = &keg->uk_domain[slab->us_domain];
5124 /*
5125 * keg_alloc_slab() always returns a slab on the
5126 * partial list.
5127 */
5128 LIST_REMOVE(slab, us_link);
5129 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
5130 us_link);
5131 dom->ud_free_slabs++;
5132 KEG_UNLOCK(keg, slab->us_domain);
5133 break;
5134 }
5135 if (vm_domainset_iter_policy(&di, &domain) != 0)
5136 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
5137 }
5138 }
5139 }
5140
5141 /*
5142 * Returns a snapshot of memory consumption in bytes.
5143 */
5144 size_t
5145 uma_zone_memory(uma_zone_t zone)
5146 {
5147 size_t sz;
5148 int i;
5149
5150 sz = 0;
5151 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
5152 for (i = 0; i < vm_ndomains; i++)
5153 sz += ZDOM_GET(zone, i)->uzd_nitems;
5154 return (sz * zone->uz_size);
5155 }
5156 for (i = 0; i < vm_ndomains; i++)
5157 sz += zone->uz_keg->uk_domain[i].ud_pages;
5158
5159 return (sz * PAGE_SIZE);
5160 }
5161
5162 struct uma_reclaim_args {
5163 int domain;
5164 int req;
5165 };
5166
5167 static void
5168 uma_reclaim_domain_cb(uma_zone_t zone, void *arg)
5169 {
5170 struct uma_reclaim_args *args;
5171
5172 args = arg;
5173 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0)
5174 uma_zone_reclaim_domain(zone, args->req, args->domain);
5175 }
5176
5177 /* See uma.h */
5178 void
5179 uma_reclaim(int req)
5180 {
5181 uma_reclaim_domain(req, UMA_ANYDOMAIN);
5182 }
5183
5184 void
5185 uma_reclaim_domain(int req, int domain)
5186 {
5187 struct uma_reclaim_args args;
5188
5189 bucket_enable();
5190
5191 args.domain = domain;
5192 args.req = req;
5193
5194 sx_slock(&uma_reclaim_lock);
5195 switch (req) {
5196 case UMA_RECLAIM_TRIM:
5197 case UMA_RECLAIM_DRAIN:
5198 zone_foreach(uma_reclaim_domain_cb, &args);
5199 break;
5200 case UMA_RECLAIM_DRAIN_CPU:
5201 /*
5202 * Reclaim globally visible free items from all zones, then drain
5203 * per-CPU buckets, then reclaim items freed while draining.
5204 * This approach minimizes expensive context switching needed to
5205 * drain each zone's per-CPU buckets.
5206 */
5207 args.req = UMA_RECLAIM_DRAIN;
5208 zone_foreach(uma_reclaim_domain_cb, &args);
5209 pcpu_cache_drain_safe(NULL);
5210 zone_foreach(uma_reclaim_domain_cb, &args);
5211 break;
5212 default:
5213 panic("unhandled reclamation request %d", req);
5214 }
5215
5216 /*
5217 * Some slabs may have been freed but this zone will be visited early
5218 * we visit again so that we can free pages that are empty once other
5219 * zones are drained. We have to do the same for buckets.
5220 */
5221 uma_zone_reclaim_domain(slabzones[0], UMA_RECLAIM_DRAIN, domain);
5222 uma_zone_reclaim_domain(slabzones[1], UMA_RECLAIM_DRAIN, domain);
5223 bucket_zone_drain(domain);
5224 sx_sunlock(&uma_reclaim_lock);
5225 }
5226
5227 static volatile int uma_reclaim_needed;
5228
5229 void
5230 uma_reclaim_wakeup(void)
5231 {
5232
5233 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
5234 wakeup(uma_reclaim);
5235 }
5236
5237 void
5238 uma_reclaim_worker(void *arg __unused)
5239 {
5240
5241 for (;;) {
5242 sx_xlock(&uma_reclaim_lock);
5243 while (atomic_load_int(&uma_reclaim_needed) == 0)
5244 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
5245 hz);
5246 sx_xunlock(&uma_reclaim_lock);
5247 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
5248 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
5249 atomic_store_int(&uma_reclaim_needed, 0);
5250 /* Don't fire more than once per-second. */
5251 pause("umarclslp", hz);
5252 }
5253 }
5254
5255 /* See uma.h */
5256 void
5257 uma_zone_reclaim(uma_zone_t zone, int req)
5258 {
5259 uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN);
5260 }
5261
5262 void
5263 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain)
5264 {
5265 switch (req) {
5266 case UMA_RECLAIM_TRIM:
5267 zone_reclaim(zone, domain, M_NOWAIT, false);
5268 break;
5269 case UMA_RECLAIM_DRAIN:
5270 zone_reclaim(zone, domain, M_NOWAIT, true);
5271 break;
5272 case UMA_RECLAIM_DRAIN_CPU:
5273 pcpu_cache_drain_safe(zone);
5274 zone_reclaim(zone, domain, M_NOWAIT, true);
5275 break;
5276 default:
5277 panic("unhandled reclamation request %d", req);
5278 }
5279 }
5280
5281 /* See uma.h */
5282 int
5283 uma_zone_exhausted(uma_zone_t zone)
5284 {
5285
5286 return (atomic_load_32(&zone->uz_sleepers) > 0);
5287 }
5288
5289 unsigned long
5290 uma_limit(void)
5291 {
5292
5293 return (uma_kmem_limit);
5294 }
5295
5296 void
5297 uma_set_limit(unsigned long limit)
5298 {
5299
5300 uma_kmem_limit = limit;
5301 }
5302
5303 unsigned long
5304 uma_size(void)
5305 {
5306
5307 return (atomic_load_long(&uma_kmem_total));
5308 }
5309
5310 long
5311 uma_avail(void)
5312 {
5313
5314 return (uma_kmem_limit - uma_size());
5315 }
5316
5317 #ifdef DDB
5318 /*
5319 * Generate statistics across both the zone and its per-cpu cache's. Return
5320 * desired statistics if the pointer is non-NULL for that statistic.
5321 *
5322 * Note: does not update the zone statistics, as it can't safely clear the
5323 * per-CPU cache statistic.
5324 *
5325 */
5326 static void
5327 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5328 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5329 {
5330 uma_cache_t cache;
5331 uint64_t allocs, frees, sleeps, xdomain;
5332 int cachefree, cpu;
5333
5334 allocs = frees = sleeps = xdomain = 0;
5335 cachefree = 0;
5336 CPU_FOREACH(cpu) {
5337 cache = &z->uz_cpu[cpu];
5338 cachefree += cache->uc_allocbucket.ucb_cnt;
5339 cachefree += cache->uc_freebucket.ucb_cnt;
5340 xdomain += cache->uc_crossbucket.ucb_cnt;
5341 cachefree += cache->uc_crossbucket.ucb_cnt;
5342 allocs += cache->uc_allocs;
5343 frees += cache->uc_frees;
5344 }
5345 allocs += counter_u64_fetch(z->uz_allocs);
5346 frees += counter_u64_fetch(z->uz_frees);
5347 xdomain += counter_u64_fetch(z->uz_xdomain);
5348 sleeps += z->uz_sleeps;
5349 if (cachefreep != NULL)
5350 *cachefreep = cachefree;
5351 if (allocsp != NULL)
5352 *allocsp = allocs;
5353 if (freesp != NULL)
5354 *freesp = frees;
5355 if (sleepsp != NULL)
5356 *sleepsp = sleeps;
5357 if (xdomainp != NULL)
5358 *xdomainp = xdomain;
5359 }
5360 #endif /* DDB */
5361
5362 static int
5363 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5364 {
5365 uma_keg_t kz;
5366 uma_zone_t z;
5367 int count;
5368
5369 count = 0;
5370 rw_rlock(&uma_rwlock);
5371 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5372 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5373 count++;
5374 }
5375 LIST_FOREACH(z, &uma_cachezones, uz_link)
5376 count++;
5377
5378 rw_runlock(&uma_rwlock);
5379 return (sysctl_handle_int(oidp, &count, 0, req));
5380 }
5381
5382 static void
5383 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5384 struct uma_percpu_stat *ups, bool internal)
5385 {
5386 uma_zone_domain_t zdom;
5387 uma_cache_t cache;
5388 int i;
5389
5390 for (i = 0; i < vm_ndomains; i++) {
5391 zdom = ZDOM_GET(z, i);
5392 uth->uth_zone_free += zdom->uzd_nitems;
5393 }
5394 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5395 uth->uth_frees = counter_u64_fetch(z->uz_frees);
5396 uth->uth_fails = counter_u64_fetch(z->uz_fails);
5397 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5398 uth->uth_sleeps = z->uz_sleeps;
5399
5400 for (i = 0; i < mp_maxid + 1; i++) {
5401 bzero(&ups[i], sizeof(*ups));
5402 if (internal || CPU_ABSENT(i))
5403 continue;
5404 cache = &z->uz_cpu[i];
5405 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5406 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5407 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5408 ups[i].ups_allocs = cache->uc_allocs;
5409 ups[i].ups_frees = cache->uc_frees;
5410 }
5411 }
5412
5413 static int
5414 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5415 {
5416 struct uma_stream_header ush;
5417 struct uma_type_header uth;
5418 struct uma_percpu_stat *ups;
5419 struct sbuf sbuf;
5420 uma_keg_t kz;
5421 uma_zone_t z;
5422 uint64_t items;
5423 uint32_t kfree, pages;
5424 int count, error, i;
5425
5426 error = sysctl_wire_old_buffer(req, 0);
5427 if (error != 0)
5428 return (error);
5429 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5430 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5431 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5432
5433 count = 0;
5434 rw_rlock(&uma_rwlock);
5435 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5436 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5437 count++;
5438 }
5439
5440 LIST_FOREACH(z, &uma_cachezones, uz_link)
5441 count++;
5442
5443 /*
5444 * Insert stream header.
5445 */
5446 bzero(&ush, sizeof(ush));
5447 ush.ush_version = UMA_STREAM_VERSION;
5448 ush.ush_maxcpus = (mp_maxid + 1);
5449 ush.ush_count = count;
5450 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5451
5452 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5453 kfree = pages = 0;
5454 for (i = 0; i < vm_ndomains; i++) {
5455 kfree += kz->uk_domain[i].ud_free_items;
5456 pages += kz->uk_domain[i].ud_pages;
5457 }
5458 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5459 bzero(&uth, sizeof(uth));
5460 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5461 uth.uth_align = kz->uk_align;
5462 uth.uth_size = kz->uk_size;
5463 uth.uth_rsize = kz->uk_rsize;
5464 if (z->uz_max_items > 0) {
5465 items = UZ_ITEMS_COUNT(z->uz_items);
5466 uth.uth_pages = (items / kz->uk_ipers) *
5467 kz->uk_ppera;
5468 } else
5469 uth.uth_pages = pages;
5470 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5471 kz->uk_ppera;
5472 uth.uth_limit = z->uz_max_items;
5473 uth.uth_keg_free = kfree;
5474
5475 /*
5476 * A zone is secondary is it is not the first entry
5477 * on the keg's zone list.
5478 */
5479 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5480 (LIST_FIRST(&kz->uk_zones) != z))
5481 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5482 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5483 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5484 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5485 for (i = 0; i < mp_maxid + 1; i++)
5486 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5487 }
5488 }
5489 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5490 bzero(&uth, sizeof(uth));
5491 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5492 uth.uth_size = z->uz_size;
5493 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5494 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5495 for (i = 0; i < mp_maxid + 1; i++)
5496 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5497 }
5498
5499 rw_runlock(&uma_rwlock);
5500 error = sbuf_finish(&sbuf);
5501 sbuf_delete(&sbuf);
5502 free(ups, M_TEMP);
5503 return (error);
5504 }
5505
5506 int
5507 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5508 {
5509 uma_zone_t zone = *(uma_zone_t *)arg1;
5510 int error, max;
5511
5512 max = uma_zone_get_max(zone);
5513 error = sysctl_handle_int(oidp, &max, 0, req);
5514 if (error || !req->newptr)
5515 return (error);
5516
5517 uma_zone_set_max(zone, max);
5518
5519 return (0);
5520 }
5521
5522 int
5523 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5524 {
5525 uma_zone_t zone;
5526 int cur;
5527
5528 /*
5529 * Some callers want to add sysctls for global zones that
5530 * may not yet exist so they pass a pointer to a pointer.
5531 */
5532 if (arg2 == 0)
5533 zone = *(uma_zone_t *)arg1;
5534 else
5535 zone = arg1;
5536 cur = uma_zone_get_cur(zone);
5537 return (sysctl_handle_int(oidp, &cur, 0, req));
5538 }
5539
5540 static int
5541 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5542 {
5543 uma_zone_t zone = arg1;
5544 uint64_t cur;
5545
5546 cur = uma_zone_get_allocs(zone);
5547 return (sysctl_handle_64(oidp, &cur, 0, req));
5548 }
5549
5550 static int
5551 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5552 {
5553 uma_zone_t zone = arg1;
5554 uint64_t cur;
5555
5556 cur = uma_zone_get_frees(zone);
5557 return (sysctl_handle_64(oidp, &cur, 0, req));
5558 }
5559
5560 static int
5561 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5562 {
5563 struct sbuf sbuf;
5564 uma_zone_t zone = arg1;
5565 int error;
5566
5567 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5568 if (zone->uz_flags != 0)
5569 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5570 else
5571 sbuf_printf(&sbuf, "0");
5572 error = sbuf_finish(&sbuf);
5573 sbuf_delete(&sbuf);
5574
5575 return (error);
5576 }
5577
5578 static int
5579 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5580 {
5581 uma_keg_t keg = arg1;
5582 int avail, effpct, total;
5583
5584 total = keg->uk_ppera * PAGE_SIZE;
5585 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5586 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5587 /*
5588 * We consider the client's requested size and alignment here, not the
5589 * real size determination uk_rsize, because we also adjust the real
5590 * size for internal implementation reasons (max bitset size).
5591 */
5592 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5593 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5594 avail *= mp_maxid + 1;
5595 effpct = 100 * avail / total;
5596 return (sysctl_handle_int(oidp, &effpct, 0, req));
5597 }
5598
5599 static int
5600 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5601 {
5602 uma_zone_t zone = arg1;
5603 uint64_t cur;
5604
5605 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5606 return (sysctl_handle_64(oidp, &cur, 0, req));
5607 }
5608
5609 #ifdef INVARIANTS
5610 static uma_slab_t
5611 uma_dbg_getslab(uma_zone_t zone, void *item)
5612 {
5613 uma_slab_t slab;
5614 uma_keg_t keg;
5615 uint8_t *mem;
5616
5617 /*
5618 * It is safe to return the slab here even though the
5619 * zone is unlocked because the item's allocation state
5620 * essentially holds a reference.
5621 */
5622 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5623 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5624 return (NULL);
5625 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5626 return (vtoslab((vm_offset_t)mem));
5627 keg = zone->uz_keg;
5628 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5629 return ((uma_slab_t)(mem + keg->uk_pgoff));
5630 KEG_LOCK(keg, 0);
5631 slab = hash_sfind(&keg->uk_hash, mem);
5632 KEG_UNLOCK(keg, 0);
5633
5634 return (slab);
5635 }
5636
5637 static bool
5638 uma_dbg_zskip(uma_zone_t zone, void *mem)
5639 {
5640
5641 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5642 return (true);
5643
5644 return (uma_dbg_kskip(zone->uz_keg, mem));
5645 }
5646
5647 static bool
5648 uma_dbg_kskip(uma_keg_t keg, void *mem)
5649 {
5650 uintptr_t idx;
5651
5652 if (dbg_divisor == 0)
5653 return (true);
5654
5655 if (dbg_divisor == 1)
5656 return (false);
5657
5658 idx = (uintptr_t)mem >> PAGE_SHIFT;
5659 if (keg->uk_ipers > 1) {
5660 idx *= keg->uk_ipers;
5661 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5662 }
5663
5664 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5665 counter_u64_add(uma_skip_cnt, 1);
5666 return (true);
5667 }
5668 counter_u64_add(uma_dbg_cnt, 1);
5669
5670 return (false);
5671 }
5672
5673 /*
5674 * Set up the slab's freei data such that uma_dbg_free can function.
5675 *
5676 */
5677 static void
5678 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5679 {
5680 uma_keg_t keg;
5681 int freei;
5682
5683 if (slab == NULL) {
5684 slab = uma_dbg_getslab(zone, item);
5685 if (slab == NULL)
5686 panic("uma: item %p did not belong to zone %s",
5687 item, zone->uz_name);
5688 }
5689 keg = zone->uz_keg;
5690 freei = slab_item_index(slab, keg, item);
5691
5692 if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5693 slab_dbg_bits(slab, keg)))
5694 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5695 item, zone, zone->uz_name, slab, freei);
5696 }
5697
5698 /*
5699 * Verifies freed addresses. Checks for alignment, valid slab membership
5700 * and duplicate frees.
5701 *
5702 */
5703 static void
5704 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5705 {
5706 uma_keg_t keg;
5707 int freei;
5708
5709 if (slab == NULL) {
5710 slab = uma_dbg_getslab(zone, item);
5711 if (slab == NULL)
5712 panic("uma: Freed item %p did not belong to zone %s",
5713 item, zone->uz_name);
5714 }
5715 keg = zone->uz_keg;
5716 freei = slab_item_index(slab, keg, item);
5717
5718 if (freei >= keg->uk_ipers)
5719 panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5720 item, zone, zone->uz_name, slab, freei);
5721
5722 if (slab_item(slab, keg, freei) != item)
5723 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5724 item, zone, zone->uz_name, slab, freei);
5725
5726 if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5727 slab_dbg_bits(slab, keg)))
5728 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5729 item, zone, zone->uz_name, slab, freei);
5730 }
5731 #endif /* INVARIANTS */
5732
5733 #ifdef DDB
5734 static int64_t
5735 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5736 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5737 {
5738 uint64_t frees;
5739 int i;
5740
5741 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5742 *allocs = counter_u64_fetch(z->uz_allocs);
5743 frees = counter_u64_fetch(z->uz_frees);
5744 *sleeps = z->uz_sleeps;
5745 *cachefree = 0;
5746 *xdomain = 0;
5747 } else
5748 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5749 xdomain);
5750 for (i = 0; i < vm_ndomains; i++) {
5751 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5752 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5753 (LIST_FIRST(&kz->uk_zones) != z)))
5754 *cachefree += kz->uk_domain[i].ud_free_items;
5755 }
5756 *used = *allocs - frees;
5757 return (((int64_t)*used + *cachefree) * kz->uk_size);
5758 }
5759
5760 DB_SHOW_COMMAND(uma, db_show_uma)
5761 {
5762 const char *fmt_hdr, *fmt_entry;
5763 uma_keg_t kz;
5764 uma_zone_t z;
5765 uint64_t allocs, used, sleeps, xdomain;
5766 long cachefree;
5767 /* variables for sorting */
5768 uma_keg_t cur_keg;
5769 uma_zone_t cur_zone, last_zone;
5770 int64_t cur_size, last_size, size;
5771 int ties;
5772
5773 /* /i option produces machine-parseable CSV output */
5774 if (modif[0] == 'i') {
5775 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5776 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5777 } else {
5778 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5779 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5780 }
5781
5782 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5783 "Sleeps", "Bucket", "Total Mem", "XFree");
5784
5785 /* Sort the zones with largest size first. */
5786 last_zone = NULL;
5787 last_size = INT64_MAX;
5788 for (;;) {
5789 cur_zone = NULL;
5790 cur_size = -1;
5791 ties = 0;
5792 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5793 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5794 /*
5795 * In the case of size ties, print out zones
5796 * in the order they are encountered. That is,
5797 * when we encounter the most recently output
5798 * zone, we have already printed all preceding
5799 * ties, and we must print all following ties.
5800 */
5801 if (z == last_zone) {
5802 ties = 1;
5803 continue;
5804 }
5805 size = get_uma_stats(kz, z, &allocs, &used,
5806 &sleeps, &cachefree, &xdomain);
5807 if (size > cur_size && size < last_size + ties)
5808 {
5809 cur_size = size;
5810 cur_zone = z;
5811 cur_keg = kz;
5812 }
5813 }
5814 }
5815 if (cur_zone == NULL)
5816 break;
5817
5818 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5819 &sleeps, &cachefree, &xdomain);
5820 db_printf(fmt_entry, cur_zone->uz_name,
5821 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5822 (uintmax_t)allocs, (uintmax_t)sleeps,
5823 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5824 xdomain);
5825
5826 if (db_pager_quit)
5827 return;
5828 last_zone = cur_zone;
5829 last_size = cur_size;
5830 }
5831 }
5832
5833 DB_SHOW_COMMAND(umacache, db_show_umacache)
5834 {
5835 uma_zone_t z;
5836 uint64_t allocs, frees;
5837 long cachefree;
5838 int i;
5839
5840 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5841 "Requests", "Bucket");
5842 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5843 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5844 for (i = 0; i < vm_ndomains; i++)
5845 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5846 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5847 z->uz_name, (uintmax_t)z->uz_size,
5848 (intmax_t)(allocs - frees), cachefree,
5849 (uintmax_t)allocs, z->uz_bucket_size);
5850 if (db_pager_quit)
5851 return;
5852 }
5853 }
5854 #endif /* DDB */
5855