1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2002-2019 Jeffrey Roberson <[email protected]>
5 * Copyright (c) 2004, 2005 Bosko Milekic <[email protected]>
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 __FBSDID("$FreeBSD$");
54
55 #include "opt_ddb.h"
56 #include "opt_param.h"
57 #include "opt_vm.h"
58
59 #include <sys/param.h>
60 #include <sys/systm.h>
61 #include <sys/asan.h>
62 #include <sys/bitset.h>
63 #include <sys/domainset.h>
64 #include <sys/eventhandler.h>
65 #include <sys/kernel.h>
66 #include <sys/types.h>
67 #include <sys/limits.h>
68 #include <sys/queue.h>
69 #include <sys/malloc.h>
70 #include <sys/ktr.h>
71 #include <sys/lock.h>
72 #include <sys/sysctl.h>
73 #include <sys/mutex.h>
74 #include <sys/proc.h>
75 #include <sys/random.h>
76 #include <sys/rwlock.h>
77 #include <sys/sbuf.h>
78 #include <sys/sched.h>
79 #include <sys/sleepqueue.h>
80 #include <sys/smp.h>
81 #include <sys/smr.h>
82 #include <sys/taskqueue.h>
83 #include <sys/vmmeter.h>
84
85 #include <vm/vm.h>
86 #include <vm/vm_param.h>
87 #include <vm/vm_domainset.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pageout.h>
91 #include <vm/vm_phys.h>
92 #include <vm/vm_pagequeue.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_dumpset.h>
97 #include <vm/uma.h>
98 #include <vm/uma_int.h>
99 #include <vm/uma_dbg.h>
100
101 #include <ddb/ddb.h>
102
103 #ifdef DEBUG_MEMGUARD
104 #include <vm/memguard.h>
105 #endif
106
107 #include <machine/md_var.h>
108
109 #ifdef INVARIANTS
110 #define UMA_ALWAYS_CTORDTOR 1
111 #else
112 #define UMA_ALWAYS_CTORDTOR 0
113 #endif
114
115 /*
116 * This is the zone and keg from which all zones are spawned.
117 */
118 static uma_zone_t kegs;
119 static uma_zone_t zones;
120
121 /*
122 * On INVARIANTS builds, the slab contains a second bitset of the same size,
123 * "dbg_bits", which is laid out immediately after us_free.
124 */
125 #ifdef INVARIANTS
126 #define SLAB_BITSETS 2
127 #else
128 #define SLAB_BITSETS 1
129 #endif
130
131 /*
132 * These are the two zones from which all offpage uma_slab_ts are allocated.
133 *
134 * One zone is for slab headers that can represent a larger number of items,
135 * making the slabs themselves more efficient, and the other zone is for
136 * headers that are smaller and represent fewer items, making the headers more
137 * efficient.
138 */
139 #define SLABZONE_SIZE(setsize) \
140 (sizeof(struct uma_hash_slab) + BITSET_SIZE(setsize) * SLAB_BITSETS)
141 #define SLABZONE0_SETSIZE (PAGE_SIZE / 16)
142 #define SLABZONE1_SETSIZE SLAB_MAX_SETSIZE
143 #define SLABZONE0_SIZE SLABZONE_SIZE(SLABZONE0_SETSIZE)
144 #define SLABZONE1_SIZE SLABZONE_SIZE(SLABZONE1_SETSIZE)
145 static uma_zone_t slabzones[2];
146
147 /*
148 * The initial hash tables come out of this zone so they can be allocated
149 * prior to malloc coming up.
150 */
151 static uma_zone_t hashzone;
152
153 /* The boot-time adjusted value for cache line alignment. */
154 int uma_align_cache = 64 - 1;
155
156 static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets");
157 static MALLOC_DEFINE(M_UMA, "UMA", "UMA Misc");
158
159 /*
160 * Are we allowed to allocate buckets?
161 */
162 static int bucketdisable = 1;
163
164 /* Linked list of all kegs in the system */
165 static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs);
166
167 /* Linked list of all cache-only zones in the system */
168 static LIST_HEAD(,uma_zone) uma_cachezones =
169 LIST_HEAD_INITIALIZER(uma_cachezones);
170
171 /*
172 * Mutex for global lists: uma_kegs, uma_cachezones, and the per-keg list of
173 * zones.
174 */
175 static struct rwlock_padalign __exclusive_cache_line uma_rwlock;
176
177 static struct sx uma_reclaim_lock;
178
179 /*
180 * First available virual address for boot time allocations.
181 */
182 static vm_offset_t bootstart;
183 static vm_offset_t bootmem;
184
185 /*
186 * kmem soft limit, initialized by uma_set_limit(). Ensure that early
187 * allocations don't trigger a wakeup of the reclaim thread.
188 */
189 unsigned long uma_kmem_limit = LONG_MAX;
190 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_limit, CTLFLAG_RD, &uma_kmem_limit, 0,
191 "UMA kernel memory soft limit");
192 unsigned long uma_kmem_total;
193 SYSCTL_ULONG(_vm, OID_AUTO, uma_kmem_total, CTLFLAG_RD, &uma_kmem_total, 0,
194 "UMA kernel memory usage");
195
196 /* Is the VM done starting up? */
197 static enum {
198 BOOT_COLD,
199 BOOT_KVA,
200 BOOT_PCPU,
201 BOOT_RUNNING,
202 BOOT_SHUTDOWN,
203 } booted = BOOT_COLD;
204
205 /*
206 * This is the handle used to schedule events that need to happen
207 * outside of the allocation fast path.
208 */
209 static struct callout uma_callout;
210 #define UMA_TIMEOUT 20 /* Seconds for callout interval. */
211
212 /*
213 * This structure is passed as the zone ctor arg so that I don't have to create
214 * a special allocation function just for zones.
215 */
216 struct uma_zctor_args {
217 const char *name;
218 size_t size;
219 uma_ctor ctor;
220 uma_dtor dtor;
221 uma_init uminit;
222 uma_fini fini;
223 uma_import import;
224 uma_release release;
225 void *arg;
226 uma_keg_t keg;
227 int align;
228 uint32_t flags;
229 };
230
231 struct uma_kctor_args {
232 uma_zone_t zone;
233 size_t size;
234 uma_init uminit;
235 uma_fini fini;
236 int align;
237 uint32_t flags;
238 };
239
240 struct uma_bucket_zone {
241 uma_zone_t ubz_zone;
242 const char *ubz_name;
243 int ubz_entries; /* Number of items it can hold. */
244 int ubz_maxsize; /* Maximum allocation size per-item. */
245 };
246
247 /*
248 * Compute the actual number of bucket entries to pack them in power
249 * of two sizes for more efficient space utilization.
250 */
251 #define BUCKET_SIZE(n) \
252 (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *))
253
254 #define BUCKET_MAX BUCKET_SIZE(256)
255
256 struct uma_bucket_zone bucket_zones[] = {
257 /* Literal bucket sizes. */
258 { NULL, "2 Bucket", 2, 4096 },
259 { NULL, "4 Bucket", 4, 3072 },
260 { NULL, "8 Bucket", 8, 2048 },
261 { NULL, "16 Bucket", 16, 1024 },
262 /* Rounded down power of 2 sizes for efficiency. */
263 { NULL, "32 Bucket", BUCKET_SIZE(32), 512 },
264 { NULL, "64 Bucket", BUCKET_SIZE(64), 256 },
265 { NULL, "128 Bucket", BUCKET_SIZE(128), 128 },
266 { NULL, "256 Bucket", BUCKET_SIZE(256), 64 },
267 { NULL, NULL, 0}
268 };
269
270 /*
271 * Flags and enumerations to be passed to internal functions.
272 */
273 enum zfreeskip {
274 SKIP_NONE = 0,
275 SKIP_CNT = 0x00000001,
276 SKIP_DTOR = 0x00010000,
277 SKIP_FINI = 0x00020000,
278 };
279
280 /* Prototypes.. */
281
282 void uma_startup1(vm_offset_t);
283 void uma_startup2(void);
284
285 static void *noobj_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
286 static void *page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
287 static void *pcpu_page_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
288 static void *startup_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
289 static void *contig_alloc(uma_zone_t, vm_size_t, int, uint8_t *, int);
290 static void page_free(void *, vm_size_t, uint8_t);
291 static void pcpu_page_free(void *, vm_size_t, uint8_t);
292 static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int, int, int);
293 static void cache_drain(uma_zone_t);
294 static void bucket_drain(uma_zone_t, uma_bucket_t);
295 static void bucket_cache_reclaim(uma_zone_t zone, bool, int);
296 static bool bucket_cache_reclaim_domain(uma_zone_t, bool, bool, int);
297 static int keg_ctor(void *, int, void *, int);
298 static void keg_dtor(void *, int, void *);
299 static void keg_drain(uma_keg_t keg, int domain);
300 static int zone_ctor(void *, int, void *, int);
301 static void zone_dtor(void *, int, void *);
302 static inline void item_dtor(uma_zone_t zone, void *item, int size,
303 void *udata, enum zfreeskip skip);
304 static int zero_init(void *, int, int);
305 static void zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
306 int itemdomain, bool ws);
307 static void zone_foreach(void (*zfunc)(uma_zone_t, void *), void *);
308 static void zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *), void *);
309 static void zone_timeout(uma_zone_t zone, void *);
310 static int hash_alloc(struct uma_hash *, u_int);
311 static int hash_expand(struct uma_hash *, struct uma_hash *);
312 static void hash_free(struct uma_hash *hash);
313 static void uma_timeout(void *);
314 static void uma_shutdown(void);
315 static void *zone_alloc_item(uma_zone_t, void *, int, int);
316 static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip);
317 static int zone_alloc_limit(uma_zone_t zone, int count, int flags);
318 static void zone_free_limit(uma_zone_t zone, int count);
319 static void bucket_enable(void);
320 static void bucket_init(void);
321 static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int);
322 static void bucket_free(uma_zone_t zone, uma_bucket_t, void *);
323 static void bucket_zone_drain(int domain);
324 static uma_bucket_t zone_alloc_bucket(uma_zone_t, void *, int, int);
325 static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab);
326 static void slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item);
327 static size_t slab_sizeof(int nitems);
328 static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit,
329 uma_fini fini, int align, uint32_t flags);
330 static int zone_import(void *, void **, int, int, int);
331 static void zone_release(void *, void **, int);
332 static bool cache_alloc(uma_zone_t, uma_cache_t, void *, int);
333 static bool cache_free(uma_zone_t, uma_cache_t, void *, void *, int);
334
335 static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS);
336 static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS);
337 static int sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS);
338 static int sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS);
339 static int sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS);
340 static int sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS);
341 static int sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS);
342
343 static uint64_t uma_zone_get_allocs(uma_zone_t zone);
344
345 static SYSCTL_NODE(_vm, OID_AUTO, debug, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
346 "Memory allocation debugging");
347
348 #ifdef INVARIANTS
349 static uint64_t uma_keg_get_allocs(uma_keg_t zone);
350 static inline struct noslabbits *slab_dbg_bits(uma_slab_t slab, uma_keg_t keg);
351
352 static bool uma_dbg_kskip(uma_keg_t keg, void *mem);
353 static bool uma_dbg_zskip(uma_zone_t zone, void *mem);
354 static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item);
355 static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item);
356
357 static u_int dbg_divisor = 1;
358 SYSCTL_UINT(_vm_debug, OID_AUTO, divisor,
359 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &dbg_divisor, 0,
360 "Debug & thrash every this item in memory allocator");
361
362 static counter_u64_t uma_dbg_cnt = EARLY_COUNTER;
363 static counter_u64_t uma_skip_cnt = EARLY_COUNTER;
364 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, trashed, CTLFLAG_RD,
365 &uma_dbg_cnt, "memory items debugged");
366 SYSCTL_COUNTER_U64(_vm_debug, OID_AUTO, skipped, CTLFLAG_RD,
367 &uma_skip_cnt, "memory items skipped, not debugged");
368 #endif
369
370 SYSCTL_NODE(_vm, OID_AUTO, uma, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
371 "Universal Memory Allocator");
372
373 SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_INT,
374 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones");
375
376 SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLFLAG_MPSAFE|CTLTYPE_STRUCT,
377 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats");
378
379 static int zone_warnings = 1;
380 SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0,
381 "Warn when UMA zones becomes full");
382
383 static int multipage_slabs = 1;
384 TUNABLE_INT("vm.debug.uma_multipage_slabs", &multipage_slabs);
385 SYSCTL_INT(_vm_debug, OID_AUTO, uma_multipage_slabs,
386 CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &multipage_slabs, 0,
387 "UMA may choose larger slab sizes for better efficiency");
388
389 /*
390 * Select the slab zone for an offpage slab with the given maximum item count.
391 */
392 static inline uma_zone_t
slabzone(int ipers)393 slabzone(int ipers)
394 {
395
396 return (slabzones[ipers > SLABZONE0_SETSIZE]);
397 }
398
399 /*
400 * This routine checks to see whether or not it's safe to enable buckets.
401 */
402 static void
bucket_enable(void)403 bucket_enable(void)
404 {
405
406 KASSERT(booted >= BOOT_KVA, ("Bucket enable before init"));
407 bucketdisable = vm_page_count_min();
408 }
409
410 /*
411 * Initialize bucket_zones, the array of zones of buckets of various sizes.
412 *
413 * For each zone, calculate the memory required for each bucket, consisting
414 * of the header and an array of pointers.
415 */
416 static void
bucket_init(void)417 bucket_init(void)
418 {
419 struct uma_bucket_zone *ubz;
420 int size;
421
422 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) {
423 size = roundup(sizeof(struct uma_bucket), sizeof(void *));
424 size += sizeof(void *) * ubz->ubz_entries;
425 ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size,
426 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
427 UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET |
428 UMA_ZONE_FIRSTTOUCH);
429 }
430 }
431
432 /*
433 * Given a desired number of entries for a bucket, return the zone from which
434 * to allocate the bucket.
435 */
436 static struct uma_bucket_zone *
bucket_zone_lookup(int entries)437 bucket_zone_lookup(int entries)
438 {
439 struct uma_bucket_zone *ubz;
440
441 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
442 if (ubz->ubz_entries >= entries)
443 return (ubz);
444 ubz--;
445 return (ubz);
446 }
447
448 static int
bucket_select(int size)449 bucket_select(int size)
450 {
451 struct uma_bucket_zone *ubz;
452
453 ubz = &bucket_zones[0];
454 if (size > ubz->ubz_maxsize)
455 return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1);
456
457 for (; ubz->ubz_entries != 0; ubz++)
458 if (ubz->ubz_maxsize < size)
459 break;
460 ubz--;
461 return (ubz->ubz_entries);
462 }
463
464 static uma_bucket_t
bucket_alloc(uma_zone_t zone,void * udata,int flags)465 bucket_alloc(uma_zone_t zone, void *udata, int flags)
466 {
467 struct uma_bucket_zone *ubz;
468 uma_bucket_t bucket;
469
470 /*
471 * Don't allocate buckets early in boot.
472 */
473 if (__predict_false(booted < BOOT_KVA))
474 return (NULL);
475
476 /*
477 * To limit bucket recursion we store the original zone flags
478 * in a cookie passed via zalloc_arg/zfree_arg. This allows the
479 * NOVM flag to persist even through deep recursions. We also
480 * store ZFLAG_BUCKET once we have recursed attempting to allocate
481 * a bucket for a bucket zone so we do not allow infinite bucket
482 * recursion. This cookie will even persist to frees of unused
483 * buckets via the allocation path or bucket allocations in the
484 * free path.
485 */
486 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
487 udata = (void *)(uintptr_t)zone->uz_flags;
488 else {
489 if ((uintptr_t)udata & UMA_ZFLAG_BUCKET)
490 return (NULL);
491 udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET);
492 }
493 if (((uintptr_t)udata & UMA_ZONE_VM) != 0)
494 flags |= M_NOVM;
495 ubz = bucket_zone_lookup(atomic_load_16(&zone->uz_bucket_size));
496 if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0)
497 ubz++;
498 bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags);
499 if (bucket) {
500 #ifdef INVARIANTS
501 bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries);
502 #endif
503 bucket->ub_cnt = 0;
504 bucket->ub_entries = min(ubz->ubz_entries,
505 zone->uz_bucket_size_max);
506 bucket->ub_seq = SMR_SEQ_INVALID;
507 CTR3(KTR_UMA, "bucket_alloc: zone %s(%p) allocated bucket %p",
508 zone->uz_name, zone, bucket);
509 }
510
511 return (bucket);
512 }
513
514 static void
bucket_free(uma_zone_t zone,uma_bucket_t bucket,void * udata)515 bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata)
516 {
517 struct uma_bucket_zone *ubz;
518
519 if (bucket->ub_cnt != 0)
520 bucket_drain(zone, bucket);
521
522 KASSERT(bucket->ub_cnt == 0,
523 ("bucket_free: Freeing a non free bucket."));
524 KASSERT(bucket->ub_seq == SMR_SEQ_INVALID,
525 ("bucket_free: Freeing an SMR bucket."));
526 if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0)
527 udata = (void *)(uintptr_t)zone->uz_flags;
528 ubz = bucket_zone_lookup(bucket->ub_entries);
529 uma_zfree_arg(ubz->ubz_zone, bucket, udata);
530 }
531
532 static void
bucket_zone_drain(int domain)533 bucket_zone_drain(int domain)
534 {
535 struct uma_bucket_zone *ubz;
536
537 for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++)
538 uma_zone_reclaim_domain(ubz->ubz_zone, UMA_RECLAIM_DRAIN,
539 domain);
540 }
541
542 #ifdef KASAN
543 _Static_assert(UMA_SMALLEST_UNIT % KASAN_SHADOW_SCALE == 0,
544 "Base UMA allocation size not a multiple of the KASAN scale factor");
545
546 static void
kasan_mark_item_valid(uma_zone_t zone,void * item)547 kasan_mark_item_valid(uma_zone_t zone, void *item)
548 {
549 void *pcpu_item;
550 size_t sz, rsz;
551 int i;
552
553 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
554 return;
555
556 sz = zone->uz_size;
557 rsz = roundup2(sz, KASAN_SHADOW_SCALE);
558 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
559 kasan_mark(item, sz, rsz, KASAN_GENERIC_REDZONE);
560 } else {
561 pcpu_item = zpcpu_base_to_offset(item);
562 for (i = 0; i <= mp_maxid; i++)
563 kasan_mark(zpcpu_get_cpu(pcpu_item, i), sz, rsz,
564 KASAN_GENERIC_REDZONE);
565 }
566 }
567
568 static void
kasan_mark_item_invalid(uma_zone_t zone,void * item)569 kasan_mark_item_invalid(uma_zone_t zone, void *item)
570 {
571 void *pcpu_item;
572 size_t sz;
573 int i;
574
575 if ((zone->uz_flags & UMA_ZONE_NOKASAN) != 0)
576 return;
577
578 sz = roundup2(zone->uz_size, KASAN_SHADOW_SCALE);
579 if ((zone->uz_flags & UMA_ZONE_PCPU) == 0) {
580 kasan_mark(item, 0, sz, KASAN_UMA_FREED);
581 } else {
582 pcpu_item = zpcpu_base_to_offset(item);
583 for (i = 0; i <= mp_maxid; i++)
584 kasan_mark(zpcpu_get_cpu(pcpu_item, i), 0, sz,
585 KASAN_UMA_FREED);
586 }
587 }
588
589 static void
kasan_mark_slab_valid(uma_keg_t keg,void * mem)590 kasan_mark_slab_valid(uma_keg_t keg, void *mem)
591 {
592 size_t sz;
593
594 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
595 sz = keg->uk_ppera * PAGE_SIZE;
596 kasan_mark(mem, sz, sz, 0);
597 }
598 }
599
600 static void
kasan_mark_slab_invalid(uma_keg_t keg,void * mem)601 kasan_mark_slab_invalid(uma_keg_t keg, void *mem)
602 {
603 size_t sz;
604
605 if ((keg->uk_flags & UMA_ZONE_NOKASAN) == 0) {
606 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
607 sz = keg->uk_ppera * PAGE_SIZE;
608 else
609 sz = keg->uk_pgoff;
610 kasan_mark(mem, 0, sz, KASAN_UMA_FREED);
611 }
612 }
613 #else /* !KASAN */
614 static void
kasan_mark_item_valid(uma_zone_t zone __unused,void * item __unused)615 kasan_mark_item_valid(uma_zone_t zone __unused, void *item __unused)
616 {
617 }
618
619 static void
kasan_mark_item_invalid(uma_zone_t zone __unused,void * item __unused)620 kasan_mark_item_invalid(uma_zone_t zone __unused, void *item __unused)
621 {
622 }
623
624 static void
kasan_mark_slab_valid(uma_keg_t keg __unused,void * mem __unused)625 kasan_mark_slab_valid(uma_keg_t keg __unused, void *mem __unused)
626 {
627 }
628
629 static void
kasan_mark_slab_invalid(uma_keg_t keg __unused,void * mem __unused)630 kasan_mark_slab_invalid(uma_keg_t keg __unused, void *mem __unused)
631 {
632 }
633 #endif /* KASAN */
634
635 /*
636 * Acquire the domain lock and record contention.
637 */
638 static uma_zone_domain_t
zone_domain_lock(uma_zone_t zone,int domain)639 zone_domain_lock(uma_zone_t zone, int domain)
640 {
641 uma_zone_domain_t zdom;
642 bool lockfail;
643
644 zdom = ZDOM_GET(zone, domain);
645 lockfail = false;
646 if (ZDOM_OWNED(zdom))
647 lockfail = true;
648 ZDOM_LOCK(zdom);
649 /* This is unsynchronized. The counter does not need to be precise. */
650 if (lockfail && zone->uz_bucket_size < zone->uz_bucket_size_max)
651 zone->uz_bucket_size++;
652 return (zdom);
653 }
654
655 /*
656 * Search for the domain with the least cached items and return it if it
657 * is out of balance with the preferred domain.
658 */
659 static __noinline int
zone_domain_lowest(uma_zone_t zone,int pref)660 zone_domain_lowest(uma_zone_t zone, int pref)
661 {
662 long least, nitems, prefitems;
663 int domain;
664 int i;
665
666 prefitems = least = LONG_MAX;
667 domain = 0;
668 for (i = 0; i < vm_ndomains; i++) {
669 nitems = ZDOM_GET(zone, i)->uzd_nitems;
670 if (nitems < least) {
671 domain = i;
672 least = nitems;
673 }
674 if (domain == pref)
675 prefitems = nitems;
676 }
677 if (prefitems < least * 2)
678 return (pref);
679
680 return (domain);
681 }
682
683 /*
684 * Search for the domain with the most cached items and return it or the
685 * preferred domain if it has enough to proceed.
686 */
687 static __noinline int
zone_domain_highest(uma_zone_t zone,int pref)688 zone_domain_highest(uma_zone_t zone, int pref)
689 {
690 long most, nitems;
691 int domain;
692 int i;
693
694 if (ZDOM_GET(zone, pref)->uzd_nitems > BUCKET_MAX)
695 return (pref);
696
697 most = 0;
698 domain = 0;
699 for (i = 0; i < vm_ndomains; i++) {
700 nitems = ZDOM_GET(zone, i)->uzd_nitems;
701 if (nitems > most) {
702 domain = i;
703 most = nitems;
704 }
705 }
706
707 return (domain);
708 }
709
710 /*
711 * Set the maximum imax value.
712 */
713 static void
zone_domain_imax_set(uma_zone_domain_t zdom,int nitems)714 zone_domain_imax_set(uma_zone_domain_t zdom, int nitems)
715 {
716 long old;
717
718 old = zdom->uzd_imax;
719 do {
720 if (old >= nitems)
721 return;
722 } while (atomic_fcmpset_long(&zdom->uzd_imax, &old, nitems) == 0);
723
724 /*
725 * We are at new maximum, so do the last WSS update for the old
726 * bimin and prepare to measure next allocation batch.
727 */
728 if (zdom->uzd_wss < old - zdom->uzd_bimin)
729 zdom->uzd_wss = old - zdom->uzd_bimin;
730 zdom->uzd_bimin = nitems;
731 }
732
733 /*
734 * Attempt to satisfy an allocation by retrieving a full bucket from one of the
735 * zone's caches. If a bucket is found the zone is not locked on return.
736 */
737 static uma_bucket_t
zone_fetch_bucket(uma_zone_t zone,uma_zone_domain_t zdom,bool reclaim)738 zone_fetch_bucket(uma_zone_t zone, uma_zone_domain_t zdom, bool reclaim)
739 {
740 uma_bucket_t bucket;
741 long cnt;
742 int i;
743 bool dtor = false;
744
745 ZDOM_LOCK_ASSERT(zdom);
746
747 if ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) == NULL)
748 return (NULL);
749
750 /* SMR Buckets can not be re-used until readers expire. */
751 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
752 bucket->ub_seq != SMR_SEQ_INVALID) {
753 if (!smr_poll(zone->uz_smr, bucket->ub_seq, false))
754 return (NULL);
755 bucket->ub_seq = SMR_SEQ_INVALID;
756 dtor = (zone->uz_dtor != NULL) || UMA_ALWAYS_CTORDTOR;
757 if (STAILQ_NEXT(bucket, ub_link) != NULL)
758 zdom->uzd_seq = STAILQ_NEXT(bucket, ub_link)->ub_seq;
759 }
760 STAILQ_REMOVE_HEAD(&zdom->uzd_buckets, ub_link);
761
762 KASSERT(zdom->uzd_nitems >= bucket->ub_cnt,
763 ("%s: item count underflow (%ld, %d)",
764 __func__, zdom->uzd_nitems, bucket->ub_cnt));
765 KASSERT(bucket->ub_cnt > 0,
766 ("%s: empty bucket in bucket cache", __func__));
767 zdom->uzd_nitems -= bucket->ub_cnt;
768
769 if (reclaim) {
770 /*
771 * Shift the bounds of the current WSS interval to avoid
772 * perturbing the estimates.
773 */
774 cnt = lmin(zdom->uzd_bimin, bucket->ub_cnt);
775 atomic_subtract_long(&zdom->uzd_imax, cnt);
776 zdom->uzd_bimin -= cnt;
777 zdom->uzd_imin -= lmin(zdom->uzd_imin, bucket->ub_cnt);
778 if (zdom->uzd_limin >= bucket->ub_cnt) {
779 zdom->uzd_limin -= bucket->ub_cnt;
780 } else {
781 zdom->uzd_limin = 0;
782 zdom->uzd_timin = 0;
783 }
784 } else if (zdom->uzd_bimin > zdom->uzd_nitems) {
785 zdom->uzd_bimin = zdom->uzd_nitems;
786 if (zdom->uzd_imin > zdom->uzd_nitems)
787 zdom->uzd_imin = zdom->uzd_nitems;
788 }
789
790 ZDOM_UNLOCK(zdom);
791 if (dtor)
792 for (i = 0; i < bucket->ub_cnt; i++)
793 item_dtor(zone, bucket->ub_bucket[i], zone->uz_size,
794 NULL, SKIP_NONE);
795
796 return (bucket);
797 }
798
799 /*
800 * Insert a full bucket into the specified cache. The "ws" parameter indicates
801 * whether the bucket's contents should be counted as part of the zone's working
802 * set. The bucket may be freed if it exceeds the bucket limit.
803 */
804 static void
zone_put_bucket(uma_zone_t zone,int domain,uma_bucket_t bucket,void * udata,const bool ws)805 zone_put_bucket(uma_zone_t zone, int domain, uma_bucket_t bucket, void *udata,
806 const bool ws)
807 {
808 uma_zone_domain_t zdom;
809
810 /* We don't cache empty buckets. This can happen after a reclaim. */
811 if (bucket->ub_cnt == 0)
812 goto out;
813 zdom = zone_domain_lock(zone, domain);
814
815 /*
816 * Conditionally set the maximum number of items.
817 */
818 zdom->uzd_nitems += bucket->ub_cnt;
819 if (__predict_true(zdom->uzd_nitems < zone->uz_bucket_max)) {
820 if (ws) {
821 zone_domain_imax_set(zdom, zdom->uzd_nitems);
822 } else {
823 /*
824 * Shift the bounds of the current WSS interval to
825 * avoid perturbing the estimates.
826 */
827 atomic_add_long(&zdom->uzd_imax, bucket->ub_cnt);
828 zdom->uzd_imin += bucket->ub_cnt;
829 zdom->uzd_bimin += bucket->ub_cnt;
830 zdom->uzd_limin += bucket->ub_cnt;
831 }
832 if (STAILQ_EMPTY(&zdom->uzd_buckets))
833 zdom->uzd_seq = bucket->ub_seq;
834
835 /*
836 * Try to promote reuse of recently used items. For items
837 * protected by SMR, try to defer reuse to minimize polling.
838 */
839 if (bucket->ub_seq == SMR_SEQ_INVALID)
840 STAILQ_INSERT_HEAD(&zdom->uzd_buckets, bucket, ub_link);
841 else
842 STAILQ_INSERT_TAIL(&zdom->uzd_buckets, bucket, ub_link);
843 ZDOM_UNLOCK(zdom);
844 return;
845 }
846 zdom->uzd_nitems -= bucket->ub_cnt;
847 ZDOM_UNLOCK(zdom);
848 out:
849 bucket_free(zone, bucket, udata);
850 }
851
852 /* Pops an item out of a per-cpu cache bucket. */
853 static inline void *
cache_bucket_pop(uma_cache_t cache,uma_cache_bucket_t bucket)854 cache_bucket_pop(uma_cache_t cache, uma_cache_bucket_t bucket)
855 {
856 void *item;
857
858 CRITICAL_ASSERT(curthread);
859
860 bucket->ucb_cnt--;
861 item = bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt];
862 #ifdef INVARIANTS
863 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = NULL;
864 KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled."));
865 #endif
866 cache->uc_allocs++;
867
868 return (item);
869 }
870
871 /* Pushes an item into a per-cpu cache bucket. */
872 static inline void
cache_bucket_push(uma_cache_t cache,uma_cache_bucket_t bucket,void * item)873 cache_bucket_push(uma_cache_t cache, uma_cache_bucket_t bucket, void *item)
874 {
875
876 CRITICAL_ASSERT(curthread);
877 KASSERT(bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] == NULL,
878 ("uma_zfree: Freeing to non free bucket index."));
879
880 bucket->ucb_bucket->ub_bucket[bucket->ucb_cnt] = item;
881 bucket->ucb_cnt++;
882 cache->uc_frees++;
883 }
884
885 /*
886 * Unload a UMA bucket from a per-cpu cache.
887 */
888 static inline uma_bucket_t
cache_bucket_unload(uma_cache_bucket_t bucket)889 cache_bucket_unload(uma_cache_bucket_t bucket)
890 {
891 uma_bucket_t b;
892
893 b = bucket->ucb_bucket;
894 if (b != NULL) {
895 MPASS(b->ub_entries == bucket->ucb_entries);
896 b->ub_cnt = bucket->ucb_cnt;
897 bucket->ucb_bucket = NULL;
898 bucket->ucb_entries = bucket->ucb_cnt = 0;
899 }
900
901 return (b);
902 }
903
904 static inline uma_bucket_t
cache_bucket_unload_alloc(uma_cache_t cache)905 cache_bucket_unload_alloc(uma_cache_t cache)
906 {
907
908 return (cache_bucket_unload(&cache->uc_allocbucket));
909 }
910
911 static inline uma_bucket_t
cache_bucket_unload_free(uma_cache_t cache)912 cache_bucket_unload_free(uma_cache_t cache)
913 {
914
915 return (cache_bucket_unload(&cache->uc_freebucket));
916 }
917
918 static inline uma_bucket_t
cache_bucket_unload_cross(uma_cache_t cache)919 cache_bucket_unload_cross(uma_cache_t cache)
920 {
921
922 return (cache_bucket_unload(&cache->uc_crossbucket));
923 }
924
925 /*
926 * Load a bucket into a per-cpu cache bucket.
927 */
928 static inline void
cache_bucket_load(uma_cache_bucket_t bucket,uma_bucket_t b)929 cache_bucket_load(uma_cache_bucket_t bucket, uma_bucket_t b)
930 {
931
932 CRITICAL_ASSERT(curthread);
933 MPASS(bucket->ucb_bucket == NULL);
934 MPASS(b->ub_seq == SMR_SEQ_INVALID);
935
936 bucket->ucb_bucket = b;
937 bucket->ucb_cnt = b->ub_cnt;
938 bucket->ucb_entries = b->ub_entries;
939 }
940
941 static inline void
cache_bucket_load_alloc(uma_cache_t cache,uma_bucket_t b)942 cache_bucket_load_alloc(uma_cache_t cache, uma_bucket_t b)
943 {
944
945 cache_bucket_load(&cache->uc_allocbucket, b);
946 }
947
948 static inline void
cache_bucket_load_free(uma_cache_t cache,uma_bucket_t b)949 cache_bucket_load_free(uma_cache_t cache, uma_bucket_t b)
950 {
951
952 cache_bucket_load(&cache->uc_freebucket, b);
953 }
954
955 #ifdef NUMA
956 static inline void
cache_bucket_load_cross(uma_cache_t cache,uma_bucket_t b)957 cache_bucket_load_cross(uma_cache_t cache, uma_bucket_t b)
958 {
959
960 cache_bucket_load(&cache->uc_crossbucket, b);
961 }
962 #endif
963
964 /*
965 * Copy and preserve ucb_spare.
966 */
967 static inline void
cache_bucket_copy(uma_cache_bucket_t b1,uma_cache_bucket_t b2)968 cache_bucket_copy(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
969 {
970
971 b1->ucb_bucket = b2->ucb_bucket;
972 b1->ucb_entries = b2->ucb_entries;
973 b1->ucb_cnt = b2->ucb_cnt;
974 }
975
976 /*
977 * Swap two cache buckets.
978 */
979 static inline void
cache_bucket_swap(uma_cache_bucket_t b1,uma_cache_bucket_t b2)980 cache_bucket_swap(uma_cache_bucket_t b1, uma_cache_bucket_t b2)
981 {
982 struct uma_cache_bucket b3;
983
984 CRITICAL_ASSERT(curthread);
985
986 cache_bucket_copy(&b3, b1);
987 cache_bucket_copy(b1, b2);
988 cache_bucket_copy(b2, &b3);
989 }
990
991 /*
992 * Attempt to fetch a bucket from a zone on behalf of the current cpu cache.
993 */
994 static uma_bucket_t
cache_fetch_bucket(uma_zone_t zone,uma_cache_t cache,int domain)995 cache_fetch_bucket(uma_zone_t zone, uma_cache_t cache, int domain)
996 {
997 uma_zone_domain_t zdom;
998 uma_bucket_t bucket;
999 smr_seq_t seq;
1000
1001 /*
1002 * Avoid the lock if possible.
1003 */
1004 zdom = ZDOM_GET(zone, domain);
1005 if (zdom->uzd_nitems == 0)
1006 return (NULL);
1007
1008 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) != 0 &&
1009 (seq = atomic_load_32(&zdom->uzd_seq)) != SMR_SEQ_INVALID &&
1010 !smr_poll(zone->uz_smr, seq, false))
1011 return (NULL);
1012
1013 /*
1014 * Check the zone's cache of buckets.
1015 */
1016 zdom = zone_domain_lock(zone, domain);
1017 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL)
1018 return (bucket);
1019 ZDOM_UNLOCK(zdom);
1020
1021 return (NULL);
1022 }
1023
1024 static void
zone_log_warning(uma_zone_t zone)1025 zone_log_warning(uma_zone_t zone)
1026 {
1027 static const struct timeval warninterval = { 300, 0 };
1028
1029 if (!zone_warnings || zone->uz_warning == NULL)
1030 return;
1031
1032 if (ratecheck(&zone->uz_ratecheck, &warninterval))
1033 printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning);
1034 }
1035
1036 static inline void
zone_maxaction(uma_zone_t zone)1037 zone_maxaction(uma_zone_t zone)
1038 {
1039
1040 if (zone->uz_maxaction.ta_func != NULL)
1041 taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction);
1042 }
1043
1044 /*
1045 * Routine called by timeout which is used to fire off some time interval
1046 * based calculations. (stats, hash size, etc.)
1047 *
1048 * Arguments:
1049 * arg Unused
1050 *
1051 * Returns:
1052 * Nothing
1053 */
1054 static void
uma_timeout(void * unused)1055 uma_timeout(void *unused)
1056 {
1057 bucket_enable();
1058 zone_foreach(zone_timeout, NULL);
1059
1060 /* Reschedule this event */
1061 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
1062 }
1063
1064 /*
1065 * Update the working set size estimates for the zone's bucket cache.
1066 * The constants chosen here are somewhat arbitrary.
1067 */
1068 static void
zone_domain_update_wss(uma_zone_domain_t zdom)1069 zone_domain_update_wss(uma_zone_domain_t zdom)
1070 {
1071 long m;
1072
1073 ZDOM_LOCK_ASSERT(zdom);
1074 MPASS(zdom->uzd_imax >= zdom->uzd_nitems);
1075 MPASS(zdom->uzd_nitems >= zdom->uzd_bimin);
1076 MPASS(zdom->uzd_bimin >= zdom->uzd_imin);
1077
1078 /*
1079 * Estimate WSS as modified moving average of biggest allocation
1080 * batches for each period over few minutes (UMA_TIMEOUT of 20s).
1081 */
1082 zdom->uzd_wss = lmax(zdom->uzd_wss * 3 / 4,
1083 zdom->uzd_imax - zdom->uzd_bimin);
1084
1085 /*
1086 * Estimate longtime minimum item count as a combination of recent
1087 * minimum item count, adjusted by WSS for safety, and the modified
1088 * moving average over the last several hours (UMA_TIMEOUT of 20s).
1089 * timin measures time since limin tried to go negative, that means
1090 * we were dangerously close to or got out of cache.
1091 */
1092 m = zdom->uzd_imin - zdom->uzd_wss;
1093 if (m >= 0) {
1094 if (zdom->uzd_limin >= m)
1095 zdom->uzd_limin = m;
1096 else
1097 zdom->uzd_limin = (m + zdom->uzd_limin * 255) / 256;
1098 zdom->uzd_timin++;
1099 } else {
1100 zdom->uzd_limin = 0;
1101 zdom->uzd_timin = 0;
1102 }
1103
1104 /* To reduce period edge effects on WSS keep half of the imax. */
1105 atomic_subtract_long(&zdom->uzd_imax,
1106 (zdom->uzd_imax - zdom->uzd_nitems + 1) / 2);
1107 zdom->uzd_imin = zdom->uzd_bimin = zdom->uzd_nitems;
1108 }
1109
1110 /*
1111 * Routine to perform timeout driven calculations. This expands the
1112 * hashes and does per cpu statistics aggregation.
1113 *
1114 * Returns nothing.
1115 */
1116 static void
zone_timeout(uma_zone_t zone,void * unused)1117 zone_timeout(uma_zone_t zone, void *unused)
1118 {
1119 uma_keg_t keg;
1120 u_int slabs, pages;
1121
1122 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
1123 goto trim;
1124
1125 keg = zone->uz_keg;
1126
1127 /*
1128 * Hash zones are non-numa by definition so the first domain
1129 * is the only one present.
1130 */
1131 KEG_LOCK(keg, 0);
1132 pages = keg->uk_domain[0].ud_pages;
1133
1134 /*
1135 * Expand the keg hash table.
1136 *
1137 * This is done if the number of slabs is larger than the hash size.
1138 * What I'm trying to do here is completely reduce collisions. This
1139 * may be a little aggressive. Should I allow for two collisions max?
1140 */
1141 if ((slabs = pages / keg->uk_ppera) > keg->uk_hash.uh_hashsize) {
1142 struct uma_hash newhash;
1143 struct uma_hash oldhash;
1144 int ret;
1145
1146 /*
1147 * This is so involved because allocating and freeing
1148 * while the keg lock is held will lead to deadlock.
1149 * I have to do everything in stages and check for
1150 * races.
1151 */
1152 KEG_UNLOCK(keg, 0);
1153 ret = hash_alloc(&newhash, 1 << fls(slabs));
1154 KEG_LOCK(keg, 0);
1155 if (ret) {
1156 if (hash_expand(&keg->uk_hash, &newhash)) {
1157 oldhash = keg->uk_hash;
1158 keg->uk_hash = newhash;
1159 } else
1160 oldhash = newhash;
1161
1162 KEG_UNLOCK(keg, 0);
1163 hash_free(&oldhash);
1164 goto trim;
1165 }
1166 }
1167 KEG_UNLOCK(keg, 0);
1168
1169 trim:
1170 /* Trim caches not used for a long time. */
1171 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0) {
1172 for (int i = 0; i < vm_ndomains; i++) {
1173 if (bucket_cache_reclaim_domain(zone, false, false, i) &&
1174 (zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1175 keg_drain(zone->uz_keg, i);
1176 }
1177 }
1178 }
1179
1180 /*
1181 * Allocate and zero fill the next sized hash table from the appropriate
1182 * backing store.
1183 *
1184 * Arguments:
1185 * hash A new hash structure with the old hash size in uh_hashsize
1186 *
1187 * Returns:
1188 * 1 on success and 0 on failure.
1189 */
1190 static int
hash_alloc(struct uma_hash * hash,u_int size)1191 hash_alloc(struct uma_hash *hash, u_int size)
1192 {
1193 size_t alloc;
1194
1195 KASSERT(powerof2(size), ("hash size must be power of 2"));
1196 if (size > UMA_HASH_SIZE_INIT) {
1197 hash->uh_hashsize = size;
1198 alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize;
1199 hash->uh_slab_hash = malloc(alloc, M_UMAHASH, M_NOWAIT);
1200 } else {
1201 alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT;
1202 hash->uh_slab_hash = zone_alloc_item(hashzone, NULL,
1203 UMA_ANYDOMAIN, M_WAITOK);
1204 hash->uh_hashsize = UMA_HASH_SIZE_INIT;
1205 }
1206 if (hash->uh_slab_hash) {
1207 bzero(hash->uh_slab_hash, alloc);
1208 hash->uh_hashmask = hash->uh_hashsize - 1;
1209 return (1);
1210 }
1211
1212 return (0);
1213 }
1214
1215 /*
1216 * Expands the hash table for HASH zones. This is done from zone_timeout
1217 * to reduce collisions. This must not be done in the regular allocation
1218 * path, otherwise, we can recurse on the vm while allocating pages.
1219 *
1220 * Arguments:
1221 * oldhash The hash you want to expand
1222 * newhash The hash structure for the new table
1223 *
1224 * Returns:
1225 * Nothing
1226 *
1227 * Discussion:
1228 */
1229 static int
hash_expand(struct uma_hash * oldhash,struct uma_hash * newhash)1230 hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash)
1231 {
1232 uma_hash_slab_t slab;
1233 u_int hval;
1234 u_int idx;
1235
1236 if (!newhash->uh_slab_hash)
1237 return (0);
1238
1239 if (oldhash->uh_hashsize >= newhash->uh_hashsize)
1240 return (0);
1241
1242 /*
1243 * I need to investigate hash algorithms for resizing without a
1244 * full rehash.
1245 */
1246
1247 for (idx = 0; idx < oldhash->uh_hashsize; idx++)
1248 while (!LIST_EMPTY(&oldhash->uh_slab_hash[idx])) {
1249 slab = LIST_FIRST(&oldhash->uh_slab_hash[idx]);
1250 LIST_REMOVE(slab, uhs_hlink);
1251 hval = UMA_HASH(newhash, slab->uhs_data);
1252 LIST_INSERT_HEAD(&newhash->uh_slab_hash[hval],
1253 slab, uhs_hlink);
1254 }
1255
1256 return (1);
1257 }
1258
1259 /*
1260 * Free the hash bucket to the appropriate backing store.
1261 *
1262 * Arguments:
1263 * slab_hash The hash bucket we're freeing
1264 * hashsize The number of entries in that hash bucket
1265 *
1266 * Returns:
1267 * Nothing
1268 */
1269 static void
hash_free(struct uma_hash * hash)1270 hash_free(struct uma_hash *hash)
1271 {
1272 if (hash->uh_slab_hash == NULL)
1273 return;
1274 if (hash->uh_hashsize == UMA_HASH_SIZE_INIT)
1275 zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE);
1276 else
1277 free(hash->uh_slab_hash, M_UMAHASH);
1278 }
1279
1280 /*
1281 * Frees all outstanding items in a bucket
1282 *
1283 * Arguments:
1284 * zone The zone to free to, must be unlocked.
1285 * bucket The free/alloc bucket with items.
1286 *
1287 * Returns:
1288 * Nothing
1289 */
1290 static void
bucket_drain(uma_zone_t zone,uma_bucket_t bucket)1291 bucket_drain(uma_zone_t zone, uma_bucket_t bucket)
1292 {
1293 int i;
1294
1295 if (bucket->ub_cnt == 0)
1296 return;
1297
1298 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 &&
1299 bucket->ub_seq != SMR_SEQ_INVALID) {
1300 smr_wait(zone->uz_smr, bucket->ub_seq);
1301 bucket->ub_seq = SMR_SEQ_INVALID;
1302 for (i = 0; i < bucket->ub_cnt; i++)
1303 item_dtor(zone, bucket->ub_bucket[i],
1304 zone->uz_size, NULL, SKIP_NONE);
1305 }
1306 if (zone->uz_fini)
1307 for (i = 0; i < bucket->ub_cnt; i++) {
1308 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
1309 zone->uz_fini(bucket->ub_bucket[i], zone->uz_size);
1310 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
1311 }
1312 zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt);
1313 if (zone->uz_max_items > 0)
1314 zone_free_limit(zone, bucket->ub_cnt);
1315 #ifdef INVARIANTS
1316 bzero(bucket->ub_bucket, sizeof(void *) * bucket->ub_cnt);
1317 #endif
1318 bucket->ub_cnt = 0;
1319 }
1320
1321 /*
1322 * Drains the per cpu caches for a zone.
1323 *
1324 * NOTE: This may only be called while the zone is being torn down, and not
1325 * during normal operation. This is necessary in order that we do not have
1326 * to migrate CPUs to drain the per-CPU caches.
1327 *
1328 * Arguments:
1329 * zone The zone to drain, must be unlocked.
1330 *
1331 * Returns:
1332 * Nothing
1333 */
1334 static void
cache_drain(uma_zone_t zone)1335 cache_drain(uma_zone_t zone)
1336 {
1337 uma_cache_t cache;
1338 uma_bucket_t bucket;
1339 smr_seq_t seq;
1340 int cpu;
1341
1342 /*
1343 * XXX: It is safe to not lock the per-CPU caches, because we're
1344 * tearing down the zone anyway. I.e., there will be no further use
1345 * of the caches at this point.
1346 *
1347 * XXX: It would good to be able to assert that the zone is being
1348 * torn down to prevent improper use of cache_drain().
1349 */
1350 seq = SMR_SEQ_INVALID;
1351 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
1352 seq = smr_advance(zone->uz_smr);
1353 CPU_FOREACH(cpu) {
1354 cache = &zone->uz_cpu[cpu];
1355 bucket = cache_bucket_unload_alloc(cache);
1356 if (bucket != NULL)
1357 bucket_free(zone, bucket, NULL);
1358 bucket = cache_bucket_unload_free(cache);
1359 if (bucket != NULL) {
1360 bucket->ub_seq = seq;
1361 bucket_free(zone, bucket, NULL);
1362 }
1363 bucket = cache_bucket_unload_cross(cache);
1364 if (bucket != NULL) {
1365 bucket->ub_seq = seq;
1366 bucket_free(zone, bucket, NULL);
1367 }
1368 }
1369 bucket_cache_reclaim(zone, true, UMA_ANYDOMAIN);
1370 }
1371
1372 static void
cache_shrink(uma_zone_t zone,void * unused)1373 cache_shrink(uma_zone_t zone, void *unused)
1374 {
1375
1376 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1377 return;
1378
1379 ZONE_LOCK(zone);
1380 zone->uz_bucket_size =
1381 (zone->uz_bucket_size_min + zone->uz_bucket_size) / 2;
1382 ZONE_UNLOCK(zone);
1383 }
1384
1385 static void
cache_drain_safe_cpu(uma_zone_t zone,void * unused)1386 cache_drain_safe_cpu(uma_zone_t zone, void *unused)
1387 {
1388 uma_cache_t cache;
1389 uma_bucket_t b1, b2, b3;
1390 int domain;
1391
1392 if (zone->uz_flags & UMA_ZFLAG_INTERNAL)
1393 return;
1394
1395 b1 = b2 = b3 = NULL;
1396 critical_enter();
1397 cache = &zone->uz_cpu[curcpu];
1398 domain = PCPU_GET(domain);
1399 b1 = cache_bucket_unload_alloc(cache);
1400
1401 /*
1402 * Don't flush SMR zone buckets. This leaves the zone without a
1403 * bucket and forces every free to synchronize().
1404 */
1405 if ((zone->uz_flags & UMA_ZONE_SMR) == 0) {
1406 b2 = cache_bucket_unload_free(cache);
1407 b3 = cache_bucket_unload_cross(cache);
1408 }
1409 critical_exit();
1410
1411 if (b1 != NULL)
1412 zone_free_bucket(zone, b1, NULL, domain, false);
1413 if (b2 != NULL)
1414 zone_free_bucket(zone, b2, NULL, domain, false);
1415 if (b3 != NULL) {
1416 /* Adjust the domain so it goes to zone_free_cross. */
1417 domain = (domain + 1) % vm_ndomains;
1418 zone_free_bucket(zone, b3, NULL, domain, false);
1419 }
1420 }
1421
1422 /*
1423 * Safely drain per-CPU caches of a zone(s) to alloc bucket.
1424 * This is an expensive call because it needs to bind to all CPUs
1425 * one by one and enter a critical section on each of them in order
1426 * to safely access their cache buckets.
1427 * Zone lock must not be held on call this function.
1428 */
1429 static void
pcpu_cache_drain_safe(uma_zone_t zone)1430 pcpu_cache_drain_safe(uma_zone_t zone)
1431 {
1432 int cpu;
1433
1434 /*
1435 * Polite bucket sizes shrinking was not enough, shrink aggressively.
1436 */
1437 if (zone)
1438 cache_shrink(zone, NULL);
1439 else
1440 zone_foreach(cache_shrink, NULL);
1441
1442 CPU_FOREACH(cpu) {
1443 thread_lock(curthread);
1444 sched_bind(curthread, cpu);
1445 thread_unlock(curthread);
1446
1447 if (zone)
1448 cache_drain_safe_cpu(zone, NULL);
1449 else
1450 zone_foreach(cache_drain_safe_cpu, NULL);
1451 }
1452 thread_lock(curthread);
1453 sched_unbind(curthread);
1454 thread_unlock(curthread);
1455 }
1456
1457 /*
1458 * Reclaim cached buckets from a zone. All buckets are reclaimed if the caller
1459 * requested a drain, otherwise the per-domain caches are trimmed to either
1460 * estimated working set size.
1461 */
1462 static bool
bucket_cache_reclaim_domain(uma_zone_t zone,bool drain,bool trim,int domain)1463 bucket_cache_reclaim_domain(uma_zone_t zone, bool drain, bool trim, int domain)
1464 {
1465 uma_zone_domain_t zdom;
1466 uma_bucket_t bucket;
1467 long target;
1468 bool done = false;
1469
1470 /*
1471 * The cross bucket is partially filled and not part of
1472 * the item count. Reclaim it individually here.
1473 */
1474 zdom = ZDOM_GET(zone, domain);
1475 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 || drain) {
1476 ZONE_CROSS_LOCK(zone);
1477 bucket = zdom->uzd_cross;
1478 zdom->uzd_cross = NULL;
1479 ZONE_CROSS_UNLOCK(zone);
1480 if (bucket != NULL)
1481 bucket_free(zone, bucket, NULL);
1482 }
1483
1484 /*
1485 * If we were asked to drain the zone, we are done only once
1486 * this bucket cache is empty. If trim, we reclaim items in
1487 * excess of the zone's estimated working set size. Multiple
1488 * consecutive calls will shrink the WSS and so reclaim more.
1489 * If neither drain nor trim, then voluntarily reclaim 1/4
1490 * (to reduce first spike) of items not used for a long time.
1491 */
1492 ZDOM_LOCK(zdom);
1493 zone_domain_update_wss(zdom);
1494 if (drain)
1495 target = 0;
1496 else if (trim)
1497 target = zdom->uzd_wss;
1498 else if (zdom->uzd_timin > 900 / UMA_TIMEOUT)
1499 target = zdom->uzd_nitems - zdom->uzd_limin / 4;
1500 else {
1501 ZDOM_UNLOCK(zdom);
1502 return (done);
1503 }
1504 while ((bucket = STAILQ_FIRST(&zdom->uzd_buckets)) != NULL &&
1505 zdom->uzd_nitems >= target + bucket->ub_cnt) {
1506 bucket = zone_fetch_bucket(zone, zdom, true);
1507 if (bucket == NULL)
1508 break;
1509 bucket_free(zone, bucket, NULL);
1510 done = true;
1511 ZDOM_LOCK(zdom);
1512 }
1513 ZDOM_UNLOCK(zdom);
1514 return (done);
1515 }
1516
1517 static void
bucket_cache_reclaim(uma_zone_t zone,bool drain,int domain)1518 bucket_cache_reclaim(uma_zone_t zone, bool drain, int domain)
1519 {
1520 int i;
1521
1522 /*
1523 * Shrink the zone bucket size to ensure that the per-CPU caches
1524 * don't grow too large.
1525 */
1526 if (zone->uz_bucket_size > zone->uz_bucket_size_min)
1527 zone->uz_bucket_size--;
1528
1529 if (domain != UMA_ANYDOMAIN &&
1530 (zone->uz_flags & UMA_ZONE_ROUNDROBIN) == 0) {
1531 bucket_cache_reclaim_domain(zone, drain, true, domain);
1532 } else {
1533 for (i = 0; i < vm_ndomains; i++)
1534 bucket_cache_reclaim_domain(zone, drain, true, i);
1535 }
1536 }
1537
1538 static void
keg_free_slab(uma_keg_t keg,uma_slab_t slab,int start)1539 keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start)
1540 {
1541 uint8_t *mem;
1542 size_t size;
1543 int i;
1544 uint8_t flags;
1545
1546 CTR4(KTR_UMA, "keg_free_slab keg %s(%p) slab %p, returning %d bytes",
1547 keg->uk_name, keg, slab, PAGE_SIZE * keg->uk_ppera);
1548
1549 mem = slab_data(slab, keg);
1550 size = PAGE_SIZE * keg->uk_ppera;
1551
1552 kasan_mark_slab_valid(keg, mem);
1553 if (keg->uk_fini != NULL) {
1554 for (i = start - 1; i > -1; i--)
1555 #ifdef INVARIANTS
1556 /*
1557 * trash_fini implies that dtor was trash_dtor. trash_fini
1558 * would check that memory hasn't been modified since free,
1559 * which executed trash_dtor.
1560 * That's why we need to run uma_dbg_kskip() check here,
1561 * albeit we don't make skip check for other init/fini
1562 * invocations.
1563 */
1564 if (!uma_dbg_kskip(keg, slab_item(slab, keg, i)) ||
1565 keg->uk_fini != trash_fini)
1566 #endif
1567 keg->uk_fini(slab_item(slab, keg, i), keg->uk_size);
1568 }
1569 flags = slab->us_flags;
1570 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1571 zone_free_item(slabzone(keg->uk_ipers), slab_tohashslab(slab),
1572 NULL, SKIP_NONE);
1573 }
1574 keg->uk_freef(mem, size, flags);
1575 uma_total_dec(size);
1576 }
1577
1578 static void
keg_drain_domain(uma_keg_t keg,int domain)1579 keg_drain_domain(uma_keg_t keg, int domain)
1580 {
1581 struct slabhead freeslabs;
1582 uma_domain_t dom;
1583 uma_slab_t slab, tmp;
1584 uint32_t i, stofree, stokeep, partial;
1585
1586 dom = &keg->uk_domain[domain];
1587 LIST_INIT(&freeslabs);
1588
1589 CTR4(KTR_UMA, "keg_drain %s(%p) domain %d free items: %u",
1590 keg->uk_name, keg, domain, dom->ud_free_items);
1591
1592 KEG_LOCK(keg, domain);
1593
1594 /*
1595 * Are the free items in partially allocated slabs sufficient to meet
1596 * the reserve? If not, compute the number of fully free slabs that must
1597 * be kept.
1598 */
1599 partial = dom->ud_free_items - dom->ud_free_slabs * keg->uk_ipers;
1600 if (partial < keg->uk_reserve) {
1601 stokeep = min(dom->ud_free_slabs,
1602 howmany(keg->uk_reserve - partial, keg->uk_ipers));
1603 } else {
1604 stokeep = 0;
1605 }
1606 stofree = dom->ud_free_slabs - stokeep;
1607
1608 /*
1609 * Partition the free slabs into two sets: those that must be kept in
1610 * order to maintain the reserve, and those that may be released back to
1611 * the system. Since one set may be much larger than the other,
1612 * populate the smaller of the two sets and swap them if necessary.
1613 */
1614 for (i = min(stofree, stokeep); i > 0; i--) {
1615 slab = LIST_FIRST(&dom->ud_free_slab);
1616 LIST_REMOVE(slab, us_link);
1617 LIST_INSERT_HEAD(&freeslabs, slab, us_link);
1618 }
1619 if (stofree > stokeep)
1620 LIST_SWAP(&freeslabs, &dom->ud_free_slab, uma_slab, us_link);
1621
1622 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0) {
1623 LIST_FOREACH(slab, &freeslabs, us_link)
1624 UMA_HASH_REMOVE(&keg->uk_hash, slab);
1625 }
1626 dom->ud_free_items -= stofree * keg->uk_ipers;
1627 dom->ud_free_slabs -= stofree;
1628 dom->ud_pages -= stofree * keg->uk_ppera;
1629 KEG_UNLOCK(keg, domain);
1630
1631 LIST_FOREACH_SAFE(slab, &freeslabs, us_link, tmp)
1632 keg_free_slab(keg, slab, keg->uk_ipers);
1633 }
1634
1635 /*
1636 * Frees pages from a keg back to the system. This is done on demand from
1637 * the pageout daemon.
1638 *
1639 * Returns nothing.
1640 */
1641 static void
keg_drain(uma_keg_t keg,int domain)1642 keg_drain(uma_keg_t keg, int domain)
1643 {
1644 int i;
1645
1646 if ((keg->uk_flags & UMA_ZONE_NOFREE) != 0)
1647 return;
1648 if (domain != UMA_ANYDOMAIN) {
1649 keg_drain_domain(keg, domain);
1650 } else {
1651 for (i = 0; i < vm_ndomains; i++)
1652 keg_drain_domain(keg, i);
1653 }
1654 }
1655
1656 static void
zone_reclaim(uma_zone_t zone,int domain,int waitok,bool drain)1657 zone_reclaim(uma_zone_t zone, int domain, int waitok, bool drain)
1658 {
1659 /*
1660 * Count active reclaim operations in order to interlock with
1661 * zone_dtor(), which removes the zone from global lists before
1662 * attempting to reclaim items itself.
1663 *
1664 * The zone may be destroyed while sleeping, so only zone_dtor() should
1665 * specify M_WAITOK.
1666 */
1667 ZONE_LOCK(zone);
1668 if (waitok == M_WAITOK) {
1669 while (zone->uz_reclaimers > 0)
1670 msleep(zone, ZONE_LOCKPTR(zone), PVM, "zonedrain", 1);
1671 }
1672 zone->uz_reclaimers++;
1673 ZONE_UNLOCK(zone);
1674 bucket_cache_reclaim(zone, drain, domain);
1675
1676 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0)
1677 keg_drain(zone->uz_keg, domain);
1678 ZONE_LOCK(zone);
1679 zone->uz_reclaimers--;
1680 if (zone->uz_reclaimers == 0)
1681 wakeup(zone);
1682 ZONE_UNLOCK(zone);
1683 }
1684
1685 /*
1686 * Allocate a new slab for a keg and inserts it into the partial slab list.
1687 * The keg should be unlocked on entry. If the allocation succeeds it will
1688 * be locked on return.
1689 *
1690 * Arguments:
1691 * flags Wait flags for the item initialization routine
1692 * aflags Wait flags for the slab allocation
1693 *
1694 * Returns:
1695 * The slab that was allocated or NULL if there is no memory and the
1696 * caller specified M_NOWAIT.
1697 */
1698 static uma_slab_t
keg_alloc_slab(uma_keg_t keg,uma_zone_t zone,int domain,int flags,int aflags)1699 keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int domain, int flags,
1700 int aflags)
1701 {
1702 uma_domain_t dom;
1703 uma_slab_t slab;
1704 unsigned long size;
1705 uint8_t *mem;
1706 uint8_t sflags;
1707 int i;
1708
1709 KASSERT(domain >= 0 && domain < vm_ndomains,
1710 ("keg_alloc_slab: domain %d out of range", domain));
1711
1712 slab = NULL;
1713 mem = NULL;
1714 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE) {
1715 uma_hash_slab_t hslab;
1716 hslab = zone_alloc_item(slabzone(keg->uk_ipers), NULL,
1717 domain, aflags);
1718 if (hslab == NULL)
1719 goto fail;
1720 slab = &hslab->uhs_slab;
1721 }
1722
1723 /*
1724 * This reproduces the old vm_zone behavior of zero filling pages the
1725 * first time they are added to a zone.
1726 *
1727 * Malloced items are zeroed in uma_zalloc.
1728 */
1729
1730 if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0)
1731 aflags |= M_ZERO;
1732 else
1733 aflags &= ~M_ZERO;
1734
1735 if (keg->uk_flags & UMA_ZONE_NODUMP)
1736 aflags |= M_NODUMP;
1737
1738 /* zone is passed for legacy reasons. */
1739 size = keg->uk_ppera * PAGE_SIZE;
1740 mem = keg->uk_allocf(zone, size, domain, &sflags, aflags);
1741 if (mem == NULL) {
1742 if (keg->uk_flags & UMA_ZFLAG_OFFPAGE)
1743 zone_free_item(slabzone(keg->uk_ipers),
1744 slab_tohashslab(slab), NULL, SKIP_NONE);
1745 goto fail;
1746 }
1747 uma_total_inc(size);
1748
1749 /* For HASH zones all pages go to the same uma_domain. */
1750 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
1751 domain = 0;
1752
1753 /* Point the slab into the allocated memory */
1754 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE))
1755 slab = (uma_slab_t)(mem + keg->uk_pgoff);
1756 else
1757 slab_tohashslab(slab)->uhs_data = mem;
1758
1759 if (keg->uk_flags & UMA_ZFLAG_VTOSLAB)
1760 for (i = 0; i < keg->uk_ppera; i++)
1761 vsetzoneslab((vm_offset_t)mem + (i * PAGE_SIZE),
1762 zone, slab);
1763
1764 slab->us_freecount = keg->uk_ipers;
1765 slab->us_flags = sflags;
1766 slab->us_domain = domain;
1767
1768 BIT_FILL(keg->uk_ipers, &slab->us_free);
1769 #ifdef INVARIANTS
1770 BIT_ZERO(keg->uk_ipers, slab_dbg_bits(slab, keg));
1771 #endif
1772
1773 if (keg->uk_init != NULL) {
1774 for (i = 0; i < keg->uk_ipers; i++)
1775 if (keg->uk_init(slab_item(slab, keg, i),
1776 keg->uk_size, flags) != 0)
1777 break;
1778 if (i != keg->uk_ipers) {
1779 keg_free_slab(keg, slab, i);
1780 goto fail;
1781 }
1782 }
1783 kasan_mark_slab_invalid(keg, mem);
1784 KEG_LOCK(keg, domain);
1785
1786 CTR3(KTR_UMA, "keg_alloc_slab: allocated slab %p for %s(%p)",
1787 slab, keg->uk_name, keg);
1788
1789 if (keg->uk_flags & UMA_ZFLAG_HASH)
1790 UMA_HASH_INSERT(&keg->uk_hash, slab, mem);
1791
1792 /*
1793 * If we got a slab here it's safe to mark it partially used
1794 * and return. We assume that the caller is going to remove
1795 * at least one item.
1796 */
1797 dom = &keg->uk_domain[domain];
1798 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
1799 dom->ud_pages += keg->uk_ppera;
1800 dom->ud_free_items += keg->uk_ipers;
1801
1802 return (slab);
1803
1804 fail:
1805 return (NULL);
1806 }
1807
1808 /*
1809 * This function is intended to be used early on in place of page_alloc(). It
1810 * performs contiguous physical memory allocations and uses a bump allocator for
1811 * KVA, so is usable before the kernel map is initialized.
1812 */
1813 static void *
startup_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)1814 startup_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1815 int wait)
1816 {
1817 vm_paddr_t pa;
1818 vm_page_t m;
1819 int i, pages;
1820
1821 pages = howmany(bytes, PAGE_SIZE);
1822 KASSERT(pages > 0, ("%s can't reserve 0 pages", __func__));
1823
1824 *pflag = UMA_SLAB_BOOT;
1825 m = vm_page_alloc_noobj_contig_domain(domain, malloc2vm_flags(wait) |
1826 VM_ALLOC_WIRED, pages, (vm_paddr_t)0, ~(vm_paddr_t)0, 1, 0,
1827 VM_MEMATTR_DEFAULT);
1828 if (m == NULL)
1829 return (NULL);
1830
1831 pa = VM_PAGE_TO_PHYS(m);
1832 for (i = 0; i < pages; i++, pa += PAGE_SIZE) {
1833 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1834 defined(__riscv) || defined(__powerpc64__)
1835 if ((wait & M_NODUMP) == 0)
1836 dump_add_page(pa);
1837 #endif
1838 }
1839
1840 /* Allocate KVA and indirectly advance bootmem. */
1841 return ((void *)pmap_map(&bootmem, m->phys_addr,
1842 m->phys_addr + (pages * PAGE_SIZE), VM_PROT_READ | VM_PROT_WRITE));
1843 }
1844
1845 static void
startup_free(void * mem,vm_size_t bytes)1846 startup_free(void *mem, vm_size_t bytes)
1847 {
1848 vm_offset_t va;
1849 vm_page_t m;
1850
1851 va = (vm_offset_t)mem;
1852 m = PHYS_TO_VM_PAGE(pmap_kextract(va));
1853
1854 /*
1855 * startup_alloc() returns direct-mapped slabs on some platforms. Avoid
1856 * unmapping ranges of the direct map.
1857 */
1858 if (va >= bootstart && va + bytes <= bootmem)
1859 pmap_remove(kernel_pmap, va, va + bytes);
1860 for (; bytes != 0; bytes -= PAGE_SIZE, m++) {
1861 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__) || \
1862 defined(__riscv) || defined(__powerpc64__)
1863 dump_drop_page(VM_PAGE_TO_PHYS(m));
1864 #endif
1865 vm_page_unwire_noq(m);
1866 vm_page_free(m);
1867 }
1868 }
1869
1870 /*
1871 * Allocates a number of pages from the system
1872 *
1873 * Arguments:
1874 * bytes The number of bytes requested
1875 * wait Shall we wait?
1876 *
1877 * Returns:
1878 * A pointer to the alloced memory or possibly
1879 * NULL if M_NOWAIT is set.
1880 */
1881 static void *
page_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)1882 page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1883 int wait)
1884 {
1885 void *p; /* Returned page */
1886
1887 *pflag = UMA_SLAB_KERNEL;
1888 p = (void *)kmem_malloc_domainset(DOMAINSET_FIXED(domain), bytes, wait);
1889
1890 return (p);
1891 }
1892
1893 static void *
pcpu_page_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)1894 pcpu_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
1895 int wait)
1896 {
1897 struct pglist alloctail;
1898 vm_offset_t addr, zkva;
1899 int cpu, flags;
1900 vm_page_t p, p_next;
1901 #ifdef NUMA
1902 struct pcpu *pc;
1903 #endif
1904
1905 MPASS(bytes == (mp_maxid + 1) * PAGE_SIZE);
1906
1907 TAILQ_INIT(&alloctail);
1908 flags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED | malloc2vm_flags(wait);
1909 *pflag = UMA_SLAB_KERNEL;
1910 for (cpu = 0; cpu <= mp_maxid; cpu++) {
1911 if (CPU_ABSENT(cpu)) {
1912 p = vm_page_alloc_noobj(flags);
1913 } else {
1914 #ifndef NUMA
1915 p = vm_page_alloc_noobj(flags);
1916 #else
1917 pc = pcpu_find(cpu);
1918 if (__predict_false(VM_DOMAIN_EMPTY(pc->pc_domain)))
1919 p = NULL;
1920 else
1921 p = vm_page_alloc_noobj_domain(pc->pc_domain,
1922 flags);
1923 if (__predict_false(p == NULL))
1924 p = vm_page_alloc_noobj(flags);
1925 #endif
1926 }
1927 if (__predict_false(p == NULL))
1928 goto fail;
1929 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1930 }
1931 if ((addr = kva_alloc(bytes)) == 0)
1932 goto fail;
1933 zkva = addr;
1934 TAILQ_FOREACH(p, &alloctail, listq) {
1935 pmap_qenter(zkva, &p, 1);
1936 zkva += PAGE_SIZE;
1937 }
1938 return ((void*)addr);
1939 fail:
1940 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1941 vm_page_unwire_noq(p);
1942 vm_page_free(p);
1943 }
1944 return (NULL);
1945 }
1946
1947 /*
1948 * Allocates a number of pages from within an object
1949 *
1950 * Arguments:
1951 * bytes The number of bytes requested
1952 * wait Shall we wait?
1953 *
1954 * Returns:
1955 * A pointer to the alloced memory or possibly
1956 * NULL if M_NOWAIT is set.
1957 */
1958 static void *
noobj_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * flags,int wait)1959 noobj_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags,
1960 int wait)
1961 {
1962 TAILQ_HEAD(, vm_page) alloctail;
1963 u_long npages;
1964 vm_offset_t retkva, zkva;
1965 vm_page_t p, p_next;
1966 uma_keg_t keg;
1967 int req;
1968
1969 TAILQ_INIT(&alloctail);
1970 keg = zone->uz_keg;
1971 req = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
1972 if ((wait & M_WAITOK) != 0)
1973 req |= VM_ALLOC_WAITOK;
1974
1975 npages = howmany(bytes, PAGE_SIZE);
1976 while (npages > 0) {
1977 p = vm_page_alloc_noobj_domain(domain, req);
1978 if (p != NULL) {
1979 /*
1980 * Since the page does not belong to an object, its
1981 * listq is unused.
1982 */
1983 TAILQ_INSERT_TAIL(&alloctail, p, listq);
1984 npages--;
1985 continue;
1986 }
1987 /*
1988 * Page allocation failed, free intermediate pages and
1989 * exit.
1990 */
1991 TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) {
1992 vm_page_unwire_noq(p);
1993 vm_page_free(p);
1994 }
1995 return (NULL);
1996 }
1997 *flags = UMA_SLAB_PRIV;
1998 zkva = keg->uk_kva +
1999 atomic_fetchadd_long(&keg->uk_offset, round_page(bytes));
2000 retkva = zkva;
2001 TAILQ_FOREACH(p, &alloctail, listq) {
2002 pmap_qenter(zkva, &p, 1);
2003 zkva += PAGE_SIZE;
2004 }
2005
2006 return ((void *)retkva);
2007 }
2008
2009 /*
2010 * Allocate physically contiguous pages.
2011 */
2012 static void *
contig_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)2013 contig_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
2014 int wait)
2015 {
2016
2017 *pflag = UMA_SLAB_KERNEL;
2018 return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain),
2019 bytes, wait, 0, ~(vm_paddr_t)0, 1, 0, VM_MEMATTR_DEFAULT));
2020 }
2021
2022 /*
2023 * Frees a number of pages to the system
2024 *
2025 * Arguments:
2026 * mem A pointer to the memory to be freed
2027 * size The size of the memory being freed
2028 * flags The original p->us_flags field
2029 *
2030 * Returns:
2031 * Nothing
2032 */
2033 static void
page_free(void * mem,vm_size_t size,uint8_t flags)2034 page_free(void *mem, vm_size_t size, uint8_t flags)
2035 {
2036
2037 if ((flags & UMA_SLAB_BOOT) != 0) {
2038 startup_free(mem, size);
2039 return;
2040 }
2041
2042 KASSERT((flags & UMA_SLAB_KERNEL) != 0,
2043 ("UMA: page_free used with invalid flags %x", flags));
2044
2045 kmem_free((vm_offset_t)mem, size);
2046 }
2047
2048 /*
2049 * Frees pcpu zone allocations
2050 *
2051 * Arguments:
2052 * mem A pointer to the memory to be freed
2053 * size The size of the memory being freed
2054 * flags The original p->us_flags field
2055 *
2056 * Returns:
2057 * Nothing
2058 */
2059 static void
pcpu_page_free(void * mem,vm_size_t size,uint8_t flags)2060 pcpu_page_free(void *mem, vm_size_t size, uint8_t flags)
2061 {
2062 vm_offset_t sva, curva;
2063 vm_paddr_t paddr;
2064 vm_page_t m;
2065
2066 MPASS(size == (mp_maxid+1)*PAGE_SIZE);
2067
2068 if ((flags & UMA_SLAB_BOOT) != 0) {
2069 startup_free(mem, size);
2070 return;
2071 }
2072
2073 sva = (vm_offset_t)mem;
2074 for (curva = sva; curva < sva + size; curva += PAGE_SIZE) {
2075 paddr = pmap_kextract(curva);
2076 m = PHYS_TO_VM_PAGE(paddr);
2077 vm_page_unwire_noq(m);
2078 vm_page_free(m);
2079 }
2080 pmap_qremove(sva, size >> PAGE_SHIFT);
2081 kva_free(sva, size);
2082 }
2083
2084 /*
2085 * Zero fill initializer
2086 *
2087 * Arguments/Returns follow uma_init specifications
2088 */
2089 static int
zero_init(void * mem,int size,int flags)2090 zero_init(void *mem, int size, int flags)
2091 {
2092 bzero(mem, size);
2093 return (0);
2094 }
2095
2096 #ifdef INVARIANTS
2097 static struct noslabbits *
slab_dbg_bits(uma_slab_t slab,uma_keg_t keg)2098 slab_dbg_bits(uma_slab_t slab, uma_keg_t keg)
2099 {
2100
2101 return ((void *)((char *)&slab->us_free + BITSET_SIZE(keg->uk_ipers)));
2102 }
2103 #endif
2104
2105 /*
2106 * Actual size of embedded struct slab (!OFFPAGE).
2107 */
2108 static size_t
slab_sizeof(int nitems)2109 slab_sizeof(int nitems)
2110 {
2111 size_t s;
2112
2113 s = sizeof(struct uma_slab) + BITSET_SIZE(nitems) * SLAB_BITSETS;
2114 return (roundup(s, UMA_ALIGN_PTR + 1));
2115 }
2116
2117 #define UMA_FIXPT_SHIFT 31
2118 #define UMA_FRAC_FIXPT(n, d) \
2119 ((uint32_t)(((uint64_t)(n) << UMA_FIXPT_SHIFT) / (d)))
2120 #define UMA_FIXPT_PCT(f) \
2121 ((u_int)(((uint64_t)100 * (f)) >> UMA_FIXPT_SHIFT))
2122 #define UMA_PCT_FIXPT(pct) UMA_FRAC_FIXPT((pct), 100)
2123 #define UMA_MIN_EFF UMA_PCT_FIXPT(100 - UMA_MAX_WASTE)
2124
2125 /*
2126 * Compute the number of items that will fit in a slab. If hdr is true, the
2127 * item count may be limited to provide space in the slab for an inline slab
2128 * header. Otherwise, all slab space will be provided for item storage.
2129 */
2130 static u_int
slab_ipers_hdr(u_int size,u_int rsize,u_int slabsize,bool hdr)2131 slab_ipers_hdr(u_int size, u_int rsize, u_int slabsize, bool hdr)
2132 {
2133 u_int ipers;
2134 u_int padpi;
2135
2136 /* The padding between items is not needed after the last item. */
2137 padpi = rsize - size;
2138
2139 if (hdr) {
2140 /*
2141 * Start with the maximum item count and remove items until
2142 * the slab header first alongside the allocatable memory.
2143 */
2144 for (ipers = MIN(SLAB_MAX_SETSIZE,
2145 (slabsize + padpi - slab_sizeof(1)) / rsize);
2146 ipers > 0 &&
2147 ipers * rsize - padpi + slab_sizeof(ipers) > slabsize;
2148 ipers--)
2149 continue;
2150 } else {
2151 ipers = MIN((slabsize + padpi) / rsize, SLAB_MAX_SETSIZE);
2152 }
2153
2154 return (ipers);
2155 }
2156
2157 struct keg_layout_result {
2158 u_int format;
2159 u_int slabsize;
2160 u_int ipers;
2161 u_int eff;
2162 };
2163
2164 static void
keg_layout_one(uma_keg_t keg,u_int rsize,u_int slabsize,u_int fmt,struct keg_layout_result * kl)2165 keg_layout_one(uma_keg_t keg, u_int rsize, u_int slabsize, u_int fmt,
2166 struct keg_layout_result *kl)
2167 {
2168 u_int total;
2169
2170 kl->format = fmt;
2171 kl->slabsize = slabsize;
2172
2173 /* Handle INTERNAL as inline with an extra page. */
2174 if ((fmt & UMA_ZFLAG_INTERNAL) != 0) {
2175 kl->format &= ~UMA_ZFLAG_INTERNAL;
2176 kl->slabsize += PAGE_SIZE;
2177 }
2178
2179 kl->ipers = slab_ipers_hdr(keg->uk_size, rsize, kl->slabsize,
2180 (fmt & UMA_ZFLAG_OFFPAGE) == 0);
2181
2182 /* Account for memory used by an offpage slab header. */
2183 total = kl->slabsize;
2184 if ((fmt & UMA_ZFLAG_OFFPAGE) != 0)
2185 total += slabzone(kl->ipers)->uz_keg->uk_rsize;
2186
2187 kl->eff = UMA_FRAC_FIXPT(kl->ipers * rsize, total);
2188 }
2189
2190 /*
2191 * Determine the format of a uma keg. This determines where the slab header
2192 * will be placed (inline or offpage) and calculates ipers, rsize, and ppera.
2193 *
2194 * Arguments
2195 * keg The zone we should initialize
2196 *
2197 * Returns
2198 * Nothing
2199 */
2200 static void
keg_layout(uma_keg_t keg)2201 keg_layout(uma_keg_t keg)
2202 {
2203 struct keg_layout_result kl = {}, kl_tmp;
2204 u_int fmts[2];
2205 u_int alignsize;
2206 u_int nfmt;
2207 u_int pages;
2208 u_int rsize;
2209 u_int slabsize;
2210 u_int i, j;
2211
2212 KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 ||
2213 (keg->uk_size <= UMA_PCPU_ALLOC_SIZE &&
2214 (keg->uk_flags & UMA_ZONE_CACHESPREAD) == 0),
2215 ("%s: cannot configure for PCPU: keg=%s, size=%u, flags=0x%b",
2216 __func__, keg->uk_name, keg->uk_size, keg->uk_flags,
2217 PRINT_UMA_ZFLAGS));
2218 KASSERT((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) == 0 ||
2219 (keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0,
2220 ("%s: incompatible flags 0x%b", __func__, keg->uk_flags,
2221 PRINT_UMA_ZFLAGS));
2222
2223 alignsize = keg->uk_align + 1;
2224 #ifdef KASAN
2225 /*
2226 * ASAN requires that each allocation be aligned to the shadow map
2227 * scale factor.
2228 */
2229 if (alignsize < KASAN_SHADOW_SCALE)
2230 alignsize = KASAN_SHADOW_SCALE;
2231 #endif
2232
2233 /*
2234 * Calculate the size of each allocation (rsize) according to
2235 * alignment. If the requested size is smaller than we have
2236 * allocation bits for we round it up.
2237 */
2238 rsize = MAX(keg->uk_size, UMA_SMALLEST_UNIT);
2239 rsize = roundup2(rsize, alignsize);
2240
2241 if ((keg->uk_flags & UMA_ZONE_CACHESPREAD) != 0) {
2242 /*
2243 * We want one item to start on every align boundary in a page.
2244 * To do this we will span pages. We will also extend the item
2245 * by the size of align if it is an even multiple of align.
2246 * Otherwise, it would fall on the same boundary every time.
2247 */
2248 if ((rsize & alignsize) == 0)
2249 rsize += alignsize;
2250 slabsize = rsize * (PAGE_SIZE / alignsize);
2251 slabsize = MIN(slabsize, rsize * SLAB_MAX_SETSIZE);
2252 slabsize = MIN(slabsize, UMA_CACHESPREAD_MAX_SIZE);
2253 slabsize = round_page(slabsize);
2254 } else {
2255 /*
2256 * Start with a slab size of as many pages as it takes to
2257 * represent a single item. We will try to fit as many
2258 * additional items into the slab as possible.
2259 */
2260 slabsize = round_page(keg->uk_size);
2261 }
2262
2263 /* Build a list of all of the available formats for this keg. */
2264 nfmt = 0;
2265
2266 /* Evaluate an inline slab layout. */
2267 if ((keg->uk_flags & (UMA_ZONE_NOTOUCH | UMA_ZONE_PCPU)) == 0)
2268 fmts[nfmt++] = 0;
2269
2270 /* TODO: vm_page-embedded slab. */
2271
2272 /*
2273 * We can't do OFFPAGE if we're internal or if we've been
2274 * asked to not go to the VM for buckets. If we do this we
2275 * may end up going to the VM for slabs which we do not want
2276 * to do if we're UMA_ZONE_VM, which clearly forbids it.
2277 * In those cases, evaluate a pseudo-format called INTERNAL
2278 * which has an inline slab header and one extra page to
2279 * guarantee that it fits.
2280 *
2281 * Otherwise, see if using an OFFPAGE slab will improve our
2282 * efficiency.
2283 */
2284 if ((keg->uk_flags & (UMA_ZFLAG_INTERNAL | UMA_ZONE_VM)) != 0)
2285 fmts[nfmt++] = UMA_ZFLAG_INTERNAL;
2286 else
2287 fmts[nfmt++] = UMA_ZFLAG_OFFPAGE;
2288
2289 /*
2290 * Choose a slab size and format which satisfy the minimum efficiency.
2291 * Prefer the smallest slab size that meets the constraints.
2292 *
2293 * Start with a minimum slab size, to accommodate CACHESPREAD. Then,
2294 * for small items (up to PAGE_SIZE), the iteration increment is one
2295 * page; and for large items, the increment is one item.
2296 */
2297 i = (slabsize + rsize - keg->uk_size) / MAX(PAGE_SIZE, rsize);
2298 KASSERT(i >= 1, ("keg %s(%p) flags=0x%b slabsize=%u, rsize=%u, i=%u",
2299 keg->uk_name, keg, keg->uk_flags, PRINT_UMA_ZFLAGS, slabsize,
2300 rsize, i));
2301 for ( ; ; i++) {
2302 slabsize = (rsize <= PAGE_SIZE) ? ptoa(i) :
2303 round_page(rsize * (i - 1) + keg->uk_size);
2304
2305 for (j = 0; j < nfmt; j++) {
2306 /* Only if we have no viable format yet. */
2307 if ((fmts[j] & UMA_ZFLAG_INTERNAL) != 0 &&
2308 kl.ipers > 0)
2309 continue;
2310
2311 keg_layout_one(keg, rsize, slabsize, fmts[j], &kl_tmp);
2312 if (kl_tmp.eff <= kl.eff)
2313 continue;
2314
2315 kl = kl_tmp;
2316
2317 CTR6(KTR_UMA, "keg %s layout: format %#x "
2318 "(ipers %u * rsize %u) / slabsize %#x = %u%% eff",
2319 keg->uk_name, kl.format, kl.ipers, rsize,
2320 kl.slabsize, UMA_FIXPT_PCT(kl.eff));
2321
2322 /* Stop when we reach the minimum efficiency. */
2323 if (kl.eff >= UMA_MIN_EFF)
2324 break;
2325 }
2326
2327 if (kl.eff >= UMA_MIN_EFF || !multipage_slabs ||
2328 slabsize >= SLAB_MAX_SETSIZE * rsize ||
2329 (keg->uk_flags & (UMA_ZONE_PCPU | UMA_ZONE_CONTIG)) != 0)
2330 break;
2331 }
2332
2333 pages = atop(kl.slabsize);
2334 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
2335 pages *= mp_maxid + 1;
2336
2337 keg->uk_rsize = rsize;
2338 keg->uk_ipers = kl.ipers;
2339 keg->uk_ppera = pages;
2340 keg->uk_flags |= kl.format;
2341
2342 /*
2343 * How do we find the slab header if it is offpage or if not all item
2344 * start addresses are in the same page? We could solve the latter
2345 * case with vaddr alignment, but we don't.
2346 */
2347 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0 ||
2348 (keg->uk_ipers - 1) * rsize >= PAGE_SIZE) {
2349 if ((keg->uk_flags & UMA_ZONE_NOTPAGE) != 0)
2350 keg->uk_flags |= UMA_ZFLAG_HASH;
2351 else
2352 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2353 }
2354
2355 CTR6(KTR_UMA, "%s: keg=%s, flags=%#x, rsize=%u, ipers=%u, ppera=%u",
2356 __func__, keg->uk_name, keg->uk_flags, rsize, keg->uk_ipers,
2357 pages);
2358 KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE,
2359 ("%s: keg=%s, flags=0x%b, rsize=%u, ipers=%u, ppera=%u", __func__,
2360 keg->uk_name, keg->uk_flags, PRINT_UMA_ZFLAGS, rsize,
2361 keg->uk_ipers, pages));
2362 }
2363
2364 /*
2365 * Keg header ctor. This initializes all fields, locks, etc. And inserts
2366 * the keg onto the global keg list.
2367 *
2368 * Arguments/Returns follow uma_ctor specifications
2369 * udata Actually uma_kctor_args
2370 */
2371 static int
keg_ctor(void * mem,int size,void * udata,int flags)2372 keg_ctor(void *mem, int size, void *udata, int flags)
2373 {
2374 struct uma_kctor_args *arg = udata;
2375 uma_keg_t keg = mem;
2376 uma_zone_t zone;
2377 int i;
2378
2379 bzero(keg, size);
2380 keg->uk_size = arg->size;
2381 keg->uk_init = arg->uminit;
2382 keg->uk_fini = arg->fini;
2383 keg->uk_align = arg->align;
2384 keg->uk_reserve = 0;
2385 keg->uk_flags = arg->flags;
2386
2387 /*
2388 * We use a global round-robin policy by default. Zones with
2389 * UMA_ZONE_FIRSTTOUCH set will use first-touch instead, in which
2390 * case the iterator is never run.
2391 */
2392 keg->uk_dr.dr_policy = DOMAINSET_RR();
2393 keg->uk_dr.dr_iter = 0;
2394
2395 /*
2396 * The primary zone is passed to us at keg-creation time.
2397 */
2398 zone = arg->zone;
2399 keg->uk_name = zone->uz_name;
2400
2401 if (arg->flags & UMA_ZONE_ZINIT)
2402 keg->uk_init = zero_init;
2403
2404 if (arg->flags & UMA_ZONE_MALLOC)
2405 keg->uk_flags |= UMA_ZFLAG_VTOSLAB;
2406
2407 #ifndef SMP
2408 keg->uk_flags &= ~UMA_ZONE_PCPU;
2409 #endif
2410
2411 keg_layout(keg);
2412
2413 /*
2414 * Use a first-touch NUMA policy for kegs that pmap_extract() will
2415 * work on. Use round-robin for everything else.
2416 *
2417 * Zones may override the default by specifying either.
2418 */
2419 #ifdef NUMA
2420 if ((keg->uk_flags &
2421 (UMA_ZONE_ROUNDROBIN | UMA_ZFLAG_CACHE | UMA_ZONE_NOTPAGE)) == 0)
2422 keg->uk_flags |= UMA_ZONE_FIRSTTOUCH;
2423 else if ((keg->uk_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2424 keg->uk_flags |= UMA_ZONE_ROUNDROBIN;
2425 #endif
2426
2427 /*
2428 * If we haven't booted yet we need allocations to go through the
2429 * startup cache until the vm is ready.
2430 */
2431 #ifdef UMA_MD_SMALL_ALLOC
2432 if (keg->uk_ppera == 1)
2433 keg->uk_allocf = uma_small_alloc;
2434 else
2435 #endif
2436 if (booted < BOOT_KVA)
2437 keg->uk_allocf = startup_alloc;
2438 else if (keg->uk_flags & UMA_ZONE_PCPU)
2439 keg->uk_allocf = pcpu_page_alloc;
2440 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 && keg->uk_ppera > 1)
2441 keg->uk_allocf = contig_alloc;
2442 else
2443 keg->uk_allocf = page_alloc;
2444 #ifdef UMA_MD_SMALL_ALLOC
2445 if (keg->uk_ppera == 1)
2446 keg->uk_freef = uma_small_free;
2447 else
2448 #endif
2449 if (keg->uk_flags & UMA_ZONE_PCPU)
2450 keg->uk_freef = pcpu_page_free;
2451 else
2452 keg->uk_freef = page_free;
2453
2454 /*
2455 * Initialize keg's locks.
2456 */
2457 for (i = 0; i < vm_ndomains; i++)
2458 KEG_LOCK_INIT(keg, i, (arg->flags & UMA_ZONE_MTXCLASS));
2459
2460 /*
2461 * If we're putting the slab header in the actual page we need to
2462 * figure out where in each page it goes. See slab_sizeof
2463 * definition.
2464 */
2465 if (!(keg->uk_flags & UMA_ZFLAG_OFFPAGE)) {
2466 size_t shsize;
2467
2468 shsize = slab_sizeof(keg->uk_ipers);
2469 keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize;
2470 /*
2471 * The only way the following is possible is if with our
2472 * UMA_ALIGN_PTR adjustments we are now bigger than
2473 * UMA_SLAB_SIZE. I haven't checked whether this is
2474 * mathematically possible for all cases, so we make
2475 * sure here anyway.
2476 */
2477 KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera,
2478 ("zone %s ipers %d rsize %d size %d slab won't fit",
2479 zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size));
2480 }
2481
2482 if (keg->uk_flags & UMA_ZFLAG_HASH)
2483 hash_alloc(&keg->uk_hash, 0);
2484
2485 CTR3(KTR_UMA, "keg_ctor %p zone %s(%p)", keg, zone->uz_name, zone);
2486
2487 LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link);
2488
2489 rw_wlock(&uma_rwlock);
2490 LIST_INSERT_HEAD(&uma_kegs, keg, uk_link);
2491 rw_wunlock(&uma_rwlock);
2492 return (0);
2493 }
2494
2495 static void
zone_kva_available(uma_zone_t zone,void * unused)2496 zone_kva_available(uma_zone_t zone, void *unused)
2497 {
2498 uma_keg_t keg;
2499
2500 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
2501 return;
2502 KEG_GET(zone, keg);
2503
2504 if (keg->uk_allocf == startup_alloc) {
2505 /* Switch to the real allocator. */
2506 if (keg->uk_flags & UMA_ZONE_PCPU)
2507 keg->uk_allocf = pcpu_page_alloc;
2508 else if ((keg->uk_flags & UMA_ZONE_CONTIG) != 0 &&
2509 keg->uk_ppera > 1)
2510 keg->uk_allocf = contig_alloc;
2511 else
2512 keg->uk_allocf = page_alloc;
2513 }
2514 }
2515
2516 static void
zone_alloc_counters(uma_zone_t zone,void * unused)2517 zone_alloc_counters(uma_zone_t zone, void *unused)
2518 {
2519
2520 zone->uz_allocs = counter_u64_alloc(M_WAITOK);
2521 zone->uz_frees = counter_u64_alloc(M_WAITOK);
2522 zone->uz_fails = counter_u64_alloc(M_WAITOK);
2523 zone->uz_xdomain = counter_u64_alloc(M_WAITOK);
2524 }
2525
2526 static void
zone_alloc_sysctl(uma_zone_t zone,void * unused)2527 zone_alloc_sysctl(uma_zone_t zone, void *unused)
2528 {
2529 uma_zone_domain_t zdom;
2530 uma_domain_t dom;
2531 uma_keg_t keg;
2532 struct sysctl_oid *oid, *domainoid;
2533 int domains, i, cnt;
2534 static const char *nokeg = "cache zone";
2535 char *c;
2536
2537 /*
2538 * Make a sysctl safe copy of the zone name by removing
2539 * any special characters and handling dups by appending
2540 * an index.
2541 */
2542 if (zone->uz_namecnt != 0) {
2543 /* Count the number of decimal digits and '_' separator. */
2544 for (i = 1, cnt = zone->uz_namecnt; cnt != 0; i++)
2545 cnt /= 10;
2546 zone->uz_ctlname = malloc(strlen(zone->uz_name) + i + 1,
2547 M_UMA, M_WAITOK);
2548 sprintf(zone->uz_ctlname, "%s_%d", zone->uz_name,
2549 zone->uz_namecnt);
2550 } else
2551 zone->uz_ctlname = strdup(zone->uz_name, M_UMA);
2552 for (c = zone->uz_ctlname; *c != '\0'; c++)
2553 if (strchr("./\\ -", *c) != NULL)
2554 *c = '_';
2555
2556 /*
2557 * Basic parameters at the root.
2558 */
2559 zone->uz_oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_vm_uma),
2560 OID_AUTO, zone->uz_ctlname, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2561 oid = zone->uz_oid;
2562 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2563 "size", CTLFLAG_RD, &zone->uz_size, 0, "Allocation size");
2564 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2565 "flags", CTLFLAG_RD | CTLTYPE_STRING | CTLFLAG_MPSAFE,
2566 zone, 0, sysctl_handle_uma_zone_flags, "A",
2567 "Allocator configuration flags");
2568 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2569 "bucket_size", CTLFLAG_RD, &zone->uz_bucket_size, 0,
2570 "Desired per-cpu cache size");
2571 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2572 "bucket_size_max", CTLFLAG_RD, &zone->uz_bucket_size_max, 0,
2573 "Maximum allowed per-cpu cache size");
2574
2575 /*
2576 * keg if present.
2577 */
2578 if ((zone->uz_flags & UMA_ZFLAG_HASH) == 0)
2579 domains = vm_ndomains;
2580 else
2581 domains = 1;
2582 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2583 "keg", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2584 keg = zone->uz_keg;
2585 if ((zone->uz_flags & UMA_ZFLAG_CACHE) == 0) {
2586 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2587 "name", CTLFLAG_RD, keg->uk_name, "Keg name");
2588 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2589 "rsize", CTLFLAG_RD, &keg->uk_rsize, 0,
2590 "Real object size with alignment");
2591 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2592 "ppera", CTLFLAG_RD, &keg->uk_ppera, 0,
2593 "pages per-slab allocation");
2594 SYSCTL_ADD_U16(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2595 "ipers", CTLFLAG_RD, &keg->uk_ipers, 0,
2596 "items available per-slab");
2597 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2598 "align", CTLFLAG_RD, &keg->uk_align, 0,
2599 "item alignment mask");
2600 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2601 "reserve", CTLFLAG_RD, &keg->uk_reserve, 0,
2602 "number of reserved items");
2603 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2604 "efficiency", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2605 keg, 0, sysctl_handle_uma_slab_efficiency, "I",
2606 "Slab utilization (100 - internal fragmentation %)");
2607 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(oid),
2608 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2609 for (i = 0; i < domains; i++) {
2610 dom = &keg->uk_domain[i];
2611 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2612 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2613 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2614 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2615 "pages", CTLFLAG_RD, &dom->ud_pages, 0,
2616 "Total pages currently allocated from VM");
2617 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2618 "free_items", CTLFLAG_RD, &dom->ud_free_items, 0,
2619 "Items free in the slab layer");
2620 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2621 "free_slabs", CTLFLAG_RD, &dom->ud_free_slabs, 0,
2622 "Unused slabs");
2623 }
2624 } else
2625 SYSCTL_ADD_CONST_STRING(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2626 "name", CTLFLAG_RD, nokeg, "Keg name");
2627
2628 /*
2629 * Information about zone limits.
2630 */
2631 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2632 "limit", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2633 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2634 "items", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2635 zone, 0, sysctl_handle_uma_zone_items, "QU",
2636 "Current number of allocated items if limit is set");
2637 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2638 "max_items", CTLFLAG_RD, &zone->uz_max_items, 0,
2639 "Maximum number of allocated and cached items");
2640 SYSCTL_ADD_U32(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2641 "sleepers", CTLFLAG_RD, &zone->uz_sleepers, 0,
2642 "Number of threads sleeping at limit");
2643 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2644 "sleeps", CTLFLAG_RD, &zone->uz_sleeps, 0,
2645 "Total zone limit sleeps");
2646 SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2647 "bucket_max", CTLFLAG_RD, &zone->uz_bucket_max, 0,
2648 "Maximum number of items in each domain's bucket cache");
2649
2650 /*
2651 * Per-domain zone information.
2652 */
2653 domainoid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid),
2654 OID_AUTO, "domain", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2655 for (i = 0; i < domains; i++) {
2656 zdom = ZDOM_GET(zone, i);
2657 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(domainoid),
2658 OID_AUTO, VM_DOMAIN(i)->vmd_name,
2659 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2660 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2661 "nitems", CTLFLAG_RD, &zdom->uzd_nitems,
2662 "number of items in this domain");
2663 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2664 "imax", CTLFLAG_RD, &zdom->uzd_imax,
2665 "maximum item count in this period");
2666 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2667 "imin", CTLFLAG_RD, &zdom->uzd_imin,
2668 "minimum item count in this period");
2669 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2670 "bimin", CTLFLAG_RD, &zdom->uzd_bimin,
2671 "Minimum item count in this batch");
2672 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2673 "wss", CTLFLAG_RD, &zdom->uzd_wss,
2674 "Working set size");
2675 SYSCTL_ADD_LONG(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2676 "limin", CTLFLAG_RD, &zdom->uzd_limin,
2677 "Long time minimum item count");
2678 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2679 "timin", CTLFLAG_RD, &zdom->uzd_timin, 0,
2680 "Time since zero long time minimum item count");
2681 }
2682
2683 /*
2684 * General statistics.
2685 */
2686 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(zone->uz_oid), OID_AUTO,
2687 "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2688 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2689 "current", CTLFLAG_RD | CTLTYPE_INT | CTLFLAG_MPSAFE,
2690 zone, 1, sysctl_handle_uma_zone_cur, "I",
2691 "Current number of allocated items");
2692 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2693 "allocs", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2694 zone, 0, sysctl_handle_uma_zone_allocs, "QU",
2695 "Total allocation calls");
2696 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2697 "frees", CTLFLAG_RD | CTLTYPE_U64 | CTLFLAG_MPSAFE,
2698 zone, 0, sysctl_handle_uma_zone_frees, "QU",
2699 "Total free calls");
2700 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2701 "fails", CTLFLAG_RD, &zone->uz_fails,
2702 "Number of allocation failures");
2703 SYSCTL_ADD_COUNTER_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO,
2704 "xdomain", CTLFLAG_RD, &zone->uz_xdomain,
2705 "Free calls from the wrong domain");
2706 }
2707
2708 struct uma_zone_count {
2709 const char *name;
2710 int count;
2711 };
2712
2713 static void
zone_count(uma_zone_t zone,void * arg)2714 zone_count(uma_zone_t zone, void *arg)
2715 {
2716 struct uma_zone_count *cnt;
2717
2718 cnt = arg;
2719 /*
2720 * Some zones are rapidly created with identical names and
2721 * destroyed out of order. This can lead to gaps in the count.
2722 * Use one greater than the maximum observed for this name.
2723 */
2724 if (strcmp(zone->uz_name, cnt->name) == 0)
2725 cnt->count = MAX(cnt->count,
2726 zone->uz_namecnt + 1);
2727 }
2728
2729 static void
zone_update_caches(uma_zone_t zone)2730 zone_update_caches(uma_zone_t zone)
2731 {
2732 int i;
2733
2734 for (i = 0; i <= mp_maxid; i++) {
2735 cache_set_uz_size(&zone->uz_cpu[i], zone->uz_size);
2736 cache_set_uz_flags(&zone->uz_cpu[i], zone->uz_flags);
2737 }
2738 }
2739
2740 /*
2741 * Zone header ctor. This initializes all fields, locks, etc.
2742 *
2743 * Arguments/Returns follow uma_ctor specifications
2744 * udata Actually uma_zctor_args
2745 */
2746 static int
zone_ctor(void * mem,int size,void * udata,int flags)2747 zone_ctor(void *mem, int size, void *udata, int flags)
2748 {
2749 struct uma_zone_count cnt;
2750 struct uma_zctor_args *arg = udata;
2751 uma_zone_domain_t zdom;
2752 uma_zone_t zone = mem;
2753 uma_zone_t z;
2754 uma_keg_t keg;
2755 int i;
2756
2757 bzero(zone, size);
2758 zone->uz_name = arg->name;
2759 zone->uz_ctor = arg->ctor;
2760 zone->uz_dtor = arg->dtor;
2761 zone->uz_init = NULL;
2762 zone->uz_fini = NULL;
2763 zone->uz_sleeps = 0;
2764 zone->uz_bucket_size = 0;
2765 zone->uz_bucket_size_min = 0;
2766 zone->uz_bucket_size_max = BUCKET_MAX;
2767 zone->uz_flags = (arg->flags & UMA_ZONE_SMR);
2768 zone->uz_warning = NULL;
2769 /* The domain structures follow the cpu structures. */
2770 zone->uz_bucket_max = ULONG_MAX;
2771 timevalclear(&zone->uz_ratecheck);
2772
2773 /* Count the number of duplicate names. */
2774 cnt.name = arg->name;
2775 cnt.count = 0;
2776 zone_foreach(zone_count, &cnt);
2777 zone->uz_namecnt = cnt.count;
2778 ZONE_CROSS_LOCK_INIT(zone);
2779
2780 for (i = 0; i < vm_ndomains; i++) {
2781 zdom = ZDOM_GET(zone, i);
2782 ZDOM_LOCK_INIT(zone, zdom, (arg->flags & UMA_ZONE_MTXCLASS));
2783 STAILQ_INIT(&zdom->uzd_buckets);
2784 }
2785
2786 #if defined(INVARIANTS) && !defined(KASAN)
2787 if (arg->uminit == trash_init && arg->fini == trash_fini)
2788 zone->uz_flags |= UMA_ZFLAG_TRASH | UMA_ZFLAG_CTORDTOR;
2789 #elif defined(KASAN)
2790 if ((arg->flags & (UMA_ZONE_NOFREE | UMA_ZFLAG_CACHE)) != 0)
2791 arg->flags |= UMA_ZONE_NOKASAN;
2792 #endif
2793
2794 /*
2795 * This is a pure cache zone, no kegs.
2796 */
2797 if (arg->import) {
2798 KASSERT((arg->flags & UMA_ZFLAG_CACHE) != 0,
2799 ("zone_ctor: Import specified for non-cache zone."));
2800 zone->uz_flags = arg->flags;
2801 zone->uz_size = arg->size;
2802 zone->uz_import = arg->import;
2803 zone->uz_release = arg->release;
2804 zone->uz_arg = arg->arg;
2805 #ifdef NUMA
2806 /*
2807 * Cache zones are round-robin unless a policy is
2808 * specified because they may have incompatible
2809 * constraints.
2810 */
2811 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) == 0)
2812 zone->uz_flags |= UMA_ZONE_ROUNDROBIN;
2813 #endif
2814 rw_wlock(&uma_rwlock);
2815 LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link);
2816 rw_wunlock(&uma_rwlock);
2817 goto out;
2818 }
2819
2820 /*
2821 * Use the regular zone/keg/slab allocator.
2822 */
2823 zone->uz_import = zone_import;
2824 zone->uz_release = zone_release;
2825 zone->uz_arg = zone;
2826 keg = arg->keg;
2827
2828 if (arg->flags & UMA_ZONE_SECONDARY) {
2829 KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0,
2830 ("Secondary zone requested UMA_ZFLAG_INTERNAL"));
2831 KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg"));
2832 zone->uz_init = arg->uminit;
2833 zone->uz_fini = arg->fini;
2834 zone->uz_flags |= UMA_ZONE_SECONDARY;
2835 rw_wlock(&uma_rwlock);
2836 ZONE_LOCK(zone);
2837 LIST_FOREACH(z, &keg->uk_zones, uz_link) {
2838 if (LIST_NEXT(z, uz_link) == NULL) {
2839 LIST_INSERT_AFTER(z, zone, uz_link);
2840 break;
2841 }
2842 }
2843 ZONE_UNLOCK(zone);
2844 rw_wunlock(&uma_rwlock);
2845 } else if (keg == NULL) {
2846 if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini,
2847 arg->align, arg->flags)) == NULL)
2848 return (ENOMEM);
2849 } else {
2850 struct uma_kctor_args karg;
2851 int error;
2852
2853 /* We should only be here from uma_startup() */
2854 karg.size = arg->size;
2855 karg.uminit = arg->uminit;
2856 karg.fini = arg->fini;
2857 karg.align = arg->align;
2858 karg.flags = (arg->flags & ~UMA_ZONE_SMR);
2859 karg.zone = zone;
2860 error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg,
2861 flags);
2862 if (error)
2863 return (error);
2864 }
2865
2866 /* Inherit properties from the keg. */
2867 zone->uz_keg = keg;
2868 zone->uz_size = keg->uk_size;
2869 zone->uz_flags |= (keg->uk_flags &
2870 (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT));
2871
2872 out:
2873 if (booted >= BOOT_PCPU) {
2874 zone_alloc_counters(zone, NULL);
2875 if (booted >= BOOT_RUNNING)
2876 zone_alloc_sysctl(zone, NULL);
2877 } else {
2878 zone->uz_allocs = EARLY_COUNTER;
2879 zone->uz_frees = EARLY_COUNTER;
2880 zone->uz_fails = EARLY_COUNTER;
2881 }
2882
2883 /* Caller requests a private SMR context. */
2884 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
2885 zone->uz_smr = smr_create(zone->uz_name, 0, 0);
2886
2887 KASSERT((arg->flags & (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET)) !=
2888 (UMA_ZONE_MAXBUCKET | UMA_ZONE_NOBUCKET),
2889 ("Invalid zone flag combination"));
2890 if (arg->flags & UMA_ZFLAG_INTERNAL)
2891 zone->uz_bucket_size_max = zone->uz_bucket_size = 0;
2892 if ((arg->flags & UMA_ZONE_MAXBUCKET) != 0)
2893 zone->uz_bucket_size = BUCKET_MAX;
2894 else if ((arg->flags & UMA_ZONE_NOBUCKET) != 0)
2895 zone->uz_bucket_size = 0;
2896 else
2897 zone->uz_bucket_size = bucket_select(zone->uz_size);
2898 zone->uz_bucket_size_min = zone->uz_bucket_size;
2899 if (zone->uz_dtor != NULL || zone->uz_ctor != NULL)
2900 zone->uz_flags |= UMA_ZFLAG_CTORDTOR;
2901 zone_update_caches(zone);
2902
2903 return (0);
2904 }
2905
2906 /*
2907 * Keg header dtor. This frees all data, destroys locks, frees the hash
2908 * table and removes the keg from the global list.
2909 *
2910 * Arguments/Returns follow uma_dtor specifications
2911 * udata unused
2912 */
2913 static void
keg_dtor(void * arg,int size,void * udata)2914 keg_dtor(void *arg, int size, void *udata)
2915 {
2916 uma_keg_t keg;
2917 uint32_t free, pages;
2918 int i;
2919
2920 keg = (uma_keg_t)arg;
2921 free = pages = 0;
2922 for (i = 0; i < vm_ndomains; i++) {
2923 free += keg->uk_domain[i].ud_free_items;
2924 pages += keg->uk_domain[i].ud_pages;
2925 KEG_LOCK_FINI(keg, i);
2926 }
2927 if (pages != 0)
2928 printf("Freed UMA keg (%s) was not empty (%u items). "
2929 " Lost %u pages of memory.\n",
2930 keg->uk_name ? keg->uk_name : "",
2931 pages / keg->uk_ppera * keg->uk_ipers - free, pages);
2932
2933 hash_free(&keg->uk_hash);
2934 }
2935
2936 /*
2937 * Zone header dtor.
2938 *
2939 * Arguments/Returns follow uma_dtor specifications
2940 * udata unused
2941 */
2942 static void
zone_dtor(void * arg,int size,void * udata)2943 zone_dtor(void *arg, int size, void *udata)
2944 {
2945 uma_zone_t zone;
2946 uma_keg_t keg;
2947 int i;
2948
2949 zone = (uma_zone_t)arg;
2950
2951 sysctl_remove_oid(zone->uz_oid, 1, 1);
2952
2953 if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL))
2954 cache_drain(zone);
2955
2956 rw_wlock(&uma_rwlock);
2957 LIST_REMOVE(zone, uz_link);
2958 rw_wunlock(&uma_rwlock);
2959 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2960 keg = zone->uz_keg;
2961 keg->uk_reserve = 0;
2962 }
2963 zone_reclaim(zone, UMA_ANYDOMAIN, M_WAITOK, true);
2964
2965 /*
2966 * We only destroy kegs from non secondary/non cache zones.
2967 */
2968 if ((zone->uz_flags & (UMA_ZONE_SECONDARY | UMA_ZFLAG_CACHE)) == 0) {
2969 keg = zone->uz_keg;
2970 rw_wlock(&uma_rwlock);
2971 LIST_REMOVE(keg, uk_link);
2972 rw_wunlock(&uma_rwlock);
2973 zone_free_item(kegs, keg, NULL, SKIP_NONE);
2974 }
2975 counter_u64_free(zone->uz_allocs);
2976 counter_u64_free(zone->uz_frees);
2977 counter_u64_free(zone->uz_fails);
2978 counter_u64_free(zone->uz_xdomain);
2979 free(zone->uz_ctlname, M_UMA);
2980 for (i = 0; i < vm_ndomains; i++)
2981 ZDOM_LOCK_FINI(ZDOM_GET(zone, i));
2982 ZONE_CROSS_LOCK_FINI(zone);
2983 }
2984
2985 static void
zone_foreach_unlocked(void (* zfunc)(uma_zone_t,void * arg),void * arg)2986 zone_foreach_unlocked(void (*zfunc)(uma_zone_t, void *arg), void *arg)
2987 {
2988 uma_keg_t keg;
2989 uma_zone_t zone;
2990
2991 LIST_FOREACH(keg, &uma_kegs, uk_link) {
2992 LIST_FOREACH(zone, &keg->uk_zones, uz_link)
2993 zfunc(zone, arg);
2994 }
2995 LIST_FOREACH(zone, &uma_cachezones, uz_link)
2996 zfunc(zone, arg);
2997 }
2998
2999 /*
3000 * Traverses every zone in the system and calls a callback
3001 *
3002 * Arguments:
3003 * zfunc A pointer to a function which accepts a zone
3004 * as an argument.
3005 *
3006 * Returns:
3007 * Nothing
3008 */
3009 static void
zone_foreach(void (* zfunc)(uma_zone_t,void * arg),void * arg)3010 zone_foreach(void (*zfunc)(uma_zone_t, void *arg), void *arg)
3011 {
3012
3013 rw_rlock(&uma_rwlock);
3014 zone_foreach_unlocked(zfunc, arg);
3015 rw_runlock(&uma_rwlock);
3016 }
3017
3018 /*
3019 * Initialize the kernel memory allocator. This is done after pages can be
3020 * allocated but before general KVA is available.
3021 */
3022 void
uma_startup1(vm_offset_t virtual_avail)3023 uma_startup1(vm_offset_t virtual_avail)
3024 {
3025 struct uma_zctor_args args;
3026 size_t ksize, zsize, size;
3027 uma_keg_t primarykeg;
3028 uintptr_t m;
3029 int domain;
3030 uint8_t pflag;
3031
3032 bootstart = bootmem = virtual_avail;
3033
3034 rw_init(&uma_rwlock, "UMA lock");
3035 sx_init(&uma_reclaim_lock, "umareclaim");
3036
3037 ksize = sizeof(struct uma_keg) +
3038 (sizeof(struct uma_domain) * vm_ndomains);
3039 ksize = roundup(ksize, UMA_SUPER_ALIGN);
3040 zsize = sizeof(struct uma_zone) +
3041 (sizeof(struct uma_cache) * (mp_maxid + 1)) +
3042 (sizeof(struct uma_zone_domain) * vm_ndomains);
3043 zsize = roundup(zsize, UMA_SUPER_ALIGN);
3044
3045 /* Allocate the zone of zones, zone of kegs, and zone of zones keg. */
3046 size = (zsize * 2) + ksize;
3047 for (domain = 0; domain < vm_ndomains; domain++) {
3048 m = (uintptr_t)startup_alloc(NULL, size, domain, &pflag,
3049 M_NOWAIT | M_ZERO);
3050 if (m != 0)
3051 break;
3052 }
3053 zones = (uma_zone_t)m;
3054 m += zsize;
3055 kegs = (uma_zone_t)m;
3056 m += zsize;
3057 primarykeg = (uma_keg_t)m;
3058
3059 /* "manually" create the initial zone */
3060 memset(&args, 0, sizeof(args));
3061 args.name = "UMA Kegs";
3062 args.size = ksize;
3063 args.ctor = keg_ctor;
3064 args.dtor = keg_dtor;
3065 args.uminit = zero_init;
3066 args.fini = NULL;
3067 args.keg = primarykeg;
3068 args.align = UMA_SUPER_ALIGN - 1;
3069 args.flags = UMA_ZFLAG_INTERNAL;
3070 zone_ctor(kegs, zsize, &args, M_WAITOK);
3071
3072 args.name = "UMA Zones";
3073 args.size = zsize;
3074 args.ctor = zone_ctor;
3075 args.dtor = zone_dtor;
3076 args.uminit = zero_init;
3077 args.fini = NULL;
3078 args.keg = NULL;
3079 args.align = UMA_SUPER_ALIGN - 1;
3080 args.flags = UMA_ZFLAG_INTERNAL;
3081 zone_ctor(zones, zsize, &args, M_WAITOK);
3082
3083 /* Now make zones for slab headers */
3084 slabzones[0] = uma_zcreate("UMA Slabs 0", SLABZONE0_SIZE,
3085 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3086 slabzones[1] = uma_zcreate("UMA Slabs 1", SLABZONE1_SIZE,
3087 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3088
3089 hashzone = uma_zcreate("UMA Hash",
3090 sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT,
3091 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL);
3092
3093 bucket_init();
3094 smr_init();
3095 }
3096
3097 #ifndef UMA_MD_SMALL_ALLOC
3098 extern void vm_radix_reserve_kva(void);
3099 #endif
3100
3101 /*
3102 * Advertise the availability of normal kva allocations and switch to
3103 * the default back-end allocator. Marks the KVA we consumed on startup
3104 * as used in the map.
3105 */
3106 void
uma_startup2(void)3107 uma_startup2(void)
3108 {
3109
3110 if (bootstart != bootmem) {
3111 vm_map_lock(kernel_map);
3112 (void)vm_map_insert(kernel_map, NULL, 0, bootstart, bootmem,
3113 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT);
3114 vm_map_unlock(kernel_map);
3115 }
3116
3117 #ifndef UMA_MD_SMALL_ALLOC
3118 /* Set up radix zone to use noobj_alloc. */
3119 vm_radix_reserve_kva();
3120 #endif
3121
3122 booted = BOOT_KVA;
3123 zone_foreach_unlocked(zone_kva_available, NULL);
3124 bucket_enable();
3125 }
3126
3127 /*
3128 * Allocate counters as early as possible so that boot-time allocations are
3129 * accounted more precisely.
3130 */
3131 static void
uma_startup_pcpu(void * arg __unused)3132 uma_startup_pcpu(void *arg __unused)
3133 {
3134
3135 zone_foreach_unlocked(zone_alloc_counters, NULL);
3136 booted = BOOT_PCPU;
3137 }
3138 SYSINIT(uma_startup_pcpu, SI_SUB_COUNTER, SI_ORDER_ANY, uma_startup_pcpu, NULL);
3139
3140 /*
3141 * Finish our initialization steps.
3142 */
3143 static void
uma_startup3(void * arg __unused)3144 uma_startup3(void *arg __unused)
3145 {
3146
3147 #ifdef INVARIANTS
3148 TUNABLE_INT_FETCH("vm.debug.divisor", &dbg_divisor);
3149 uma_dbg_cnt = counter_u64_alloc(M_WAITOK);
3150 uma_skip_cnt = counter_u64_alloc(M_WAITOK);
3151 #endif
3152 zone_foreach_unlocked(zone_alloc_sysctl, NULL);
3153 callout_init(&uma_callout, 1);
3154 callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL);
3155 booted = BOOT_RUNNING;
3156
3157 EVENTHANDLER_REGISTER(shutdown_post_sync, uma_shutdown, NULL,
3158 EVENTHANDLER_PRI_FIRST);
3159 }
3160 SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL);
3161
3162 static void
uma_shutdown(void)3163 uma_shutdown(void)
3164 {
3165
3166 booted = BOOT_SHUTDOWN;
3167 }
3168
3169 static uma_keg_t
uma_kcreate(uma_zone_t zone,size_t size,uma_init uminit,uma_fini fini,int align,uint32_t flags)3170 uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini,
3171 int align, uint32_t flags)
3172 {
3173 struct uma_kctor_args args;
3174
3175 args.size = size;
3176 args.uminit = uminit;
3177 args.fini = fini;
3178 args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align;
3179 args.flags = flags;
3180 args.zone = zone;
3181 return (zone_alloc_item(kegs, &args, UMA_ANYDOMAIN, M_WAITOK));
3182 }
3183
3184 /* Public functions */
3185 /* See uma.h */
3186 void
uma_set_align(int align)3187 uma_set_align(int align)
3188 {
3189
3190 if (align != UMA_ALIGN_CACHE)
3191 uma_align_cache = align;
3192 }
3193
3194 /* See uma.h */
3195 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)3196 uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor,
3197 uma_init uminit, uma_fini fini, int align, uint32_t flags)
3198
3199 {
3200 struct uma_zctor_args args;
3201 uma_zone_t res;
3202
3203 KASSERT(powerof2(align + 1), ("invalid zone alignment %d for \"%s\"",
3204 align, name));
3205
3206 /* This stuff is essential for the zone ctor */
3207 memset(&args, 0, sizeof(args));
3208 args.name = name;
3209 args.size = size;
3210 args.ctor = ctor;
3211 args.dtor = dtor;
3212 args.uminit = uminit;
3213 args.fini = fini;
3214 #if defined(INVARIANTS) && !defined(KASAN)
3215 /*
3216 * Inject procedures which check for memory use after free if we are
3217 * allowed to scramble the memory while it is not allocated. This
3218 * requires that: UMA is actually able to access the memory, no init
3219 * or fini procedures, no dependency on the initial value of the
3220 * memory, and no (legitimate) use of the memory after free. Note,
3221 * the ctor and dtor do not need to be empty.
3222 */
3223 if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOTOUCH |
3224 UMA_ZONE_NOFREE))) && uminit == NULL && fini == NULL) {
3225 args.uminit = trash_init;
3226 args.fini = trash_fini;
3227 }
3228 #endif
3229 args.align = align;
3230 args.flags = flags;
3231 args.keg = NULL;
3232
3233 sx_xlock(&uma_reclaim_lock);
3234 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3235 sx_xunlock(&uma_reclaim_lock);
3236
3237 return (res);
3238 }
3239
3240 /* See uma.h */
3241 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)3242 uma_zsecond_create(const char *name, uma_ctor ctor, uma_dtor dtor,
3243 uma_init zinit, uma_fini zfini, uma_zone_t primary)
3244 {
3245 struct uma_zctor_args args;
3246 uma_keg_t keg;
3247 uma_zone_t res;
3248
3249 keg = primary->uz_keg;
3250 memset(&args, 0, sizeof(args));
3251 args.name = name;
3252 args.size = keg->uk_size;
3253 args.ctor = ctor;
3254 args.dtor = dtor;
3255 args.uminit = zinit;
3256 args.fini = zfini;
3257 args.align = keg->uk_align;
3258 args.flags = keg->uk_flags | UMA_ZONE_SECONDARY;
3259 args.keg = keg;
3260
3261 sx_xlock(&uma_reclaim_lock);
3262 res = zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK);
3263 sx_xunlock(&uma_reclaim_lock);
3264
3265 return (res);
3266 }
3267
3268 /* See uma.h */
3269 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)3270 uma_zcache_create(const char *name, int size, uma_ctor ctor, uma_dtor dtor,
3271 uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease,
3272 void *arg, int flags)
3273 {
3274 struct uma_zctor_args args;
3275
3276 memset(&args, 0, sizeof(args));
3277 args.name = name;
3278 args.size = size;
3279 args.ctor = ctor;
3280 args.dtor = dtor;
3281 args.uminit = zinit;
3282 args.fini = zfini;
3283 args.import = zimport;
3284 args.release = zrelease;
3285 args.arg = arg;
3286 args.align = 0;
3287 args.flags = flags | UMA_ZFLAG_CACHE;
3288
3289 return (zone_alloc_item(zones, &args, UMA_ANYDOMAIN, M_WAITOK));
3290 }
3291
3292 /* See uma.h */
3293 void
uma_zdestroy(uma_zone_t zone)3294 uma_zdestroy(uma_zone_t zone)
3295 {
3296
3297 /*
3298 * Large slabs are expensive to reclaim, so don't bother doing
3299 * unnecessary work if we're shutting down.
3300 */
3301 if (booted == BOOT_SHUTDOWN &&
3302 zone->uz_fini == NULL && zone->uz_release == zone_release)
3303 return;
3304 sx_xlock(&uma_reclaim_lock);
3305 zone_free_item(zones, zone, NULL, SKIP_NONE);
3306 sx_xunlock(&uma_reclaim_lock);
3307 }
3308
3309 void
uma_zwait(uma_zone_t zone)3310 uma_zwait(uma_zone_t zone)
3311 {
3312
3313 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
3314 uma_zfree_smr(zone, uma_zalloc_smr(zone, M_WAITOK));
3315 else if ((zone->uz_flags & UMA_ZONE_PCPU) != 0)
3316 uma_zfree_pcpu(zone, uma_zalloc_pcpu(zone, M_WAITOK));
3317 else
3318 uma_zfree(zone, uma_zalloc(zone, M_WAITOK));
3319 }
3320
3321 void *
uma_zalloc_pcpu_arg(uma_zone_t zone,void * udata,int flags)3322 uma_zalloc_pcpu_arg(uma_zone_t zone, void *udata, int flags)
3323 {
3324 void *item, *pcpu_item;
3325 #ifdef SMP
3326 int i;
3327
3328 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3329 #endif
3330 item = uma_zalloc_arg(zone, udata, flags & ~M_ZERO);
3331 if (item == NULL)
3332 return (NULL);
3333 pcpu_item = zpcpu_base_to_offset(item);
3334 if (flags & M_ZERO) {
3335 #ifdef SMP
3336 for (i = 0; i <= mp_maxid; i++)
3337 bzero(zpcpu_get_cpu(pcpu_item, i), zone->uz_size);
3338 #else
3339 bzero(item, zone->uz_size);
3340 #endif
3341 }
3342 return (pcpu_item);
3343 }
3344
3345 /*
3346 * A stub while both regular and pcpu cases are identical.
3347 */
3348 void
uma_zfree_pcpu_arg(uma_zone_t zone,void * pcpu_item,void * udata)3349 uma_zfree_pcpu_arg(uma_zone_t zone, void *pcpu_item, void *udata)
3350 {
3351 void *item;
3352
3353 #ifdef SMP
3354 MPASS(zone->uz_flags & UMA_ZONE_PCPU);
3355 #endif
3356
3357 /* uma_zfree_pcu_*(..., NULL) does nothing, to match free(9). */
3358 if (pcpu_item == NULL)
3359 return;
3360
3361 item = zpcpu_offset_to_base(pcpu_item);
3362 uma_zfree_arg(zone, item, udata);
3363 }
3364
3365 static inline void *
item_ctor(uma_zone_t zone,int uz_flags,int size,void * udata,int flags,void * item)3366 item_ctor(uma_zone_t zone, int uz_flags, int size, void *udata, int flags,
3367 void *item)
3368 {
3369 #ifdef INVARIANTS
3370 bool skipdbg;
3371 #endif
3372
3373 kasan_mark_item_valid(zone, item);
3374
3375 #ifdef INVARIANTS
3376 skipdbg = uma_dbg_zskip(zone, item);
3377 if (!skipdbg && (uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3378 zone->uz_ctor != trash_ctor)
3379 trash_ctor(item, size, udata, flags);
3380 #endif
3381
3382 /* Check flags before loading ctor pointer. */
3383 if (__predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0) &&
3384 __predict_false(zone->uz_ctor != NULL) &&
3385 zone->uz_ctor(item, size, udata, flags) != 0) {
3386 counter_u64_add(zone->uz_fails, 1);
3387 zone_free_item(zone, item, udata, SKIP_DTOR | SKIP_CNT);
3388 return (NULL);
3389 }
3390 #ifdef INVARIANTS
3391 if (!skipdbg)
3392 uma_dbg_alloc(zone, NULL, item);
3393 #endif
3394 if (__predict_false(flags & M_ZERO))
3395 return (memset(item, 0, size));
3396
3397 return (item);
3398 }
3399
3400 static inline void
item_dtor(uma_zone_t zone,void * item,int size,void * udata,enum zfreeskip skip)3401 item_dtor(uma_zone_t zone, void *item, int size, void *udata,
3402 enum zfreeskip skip)
3403 {
3404 #ifdef INVARIANTS
3405 bool skipdbg;
3406
3407 skipdbg = uma_dbg_zskip(zone, item);
3408 if (skip == SKIP_NONE && !skipdbg) {
3409 if ((zone->uz_flags & UMA_ZONE_MALLOC) != 0)
3410 uma_dbg_free(zone, udata, item);
3411 else
3412 uma_dbg_free(zone, NULL, item);
3413 }
3414 #endif
3415 if (__predict_true(skip < SKIP_DTOR)) {
3416 if (zone->uz_dtor != NULL)
3417 zone->uz_dtor(item, size, udata);
3418 #ifdef INVARIANTS
3419 if (!skipdbg && (zone->uz_flags & UMA_ZFLAG_TRASH) != 0 &&
3420 zone->uz_dtor != trash_dtor)
3421 trash_dtor(item, size, udata);
3422 #endif
3423 }
3424 kasan_mark_item_invalid(zone, item);
3425 }
3426
3427 #ifdef NUMA
3428 static int
item_domain(void * item)3429 item_domain(void *item)
3430 {
3431 int domain;
3432
3433 domain = vm_phys_domain(vtophys(item));
3434 KASSERT(domain >= 0 && domain < vm_ndomains,
3435 ("%s: unknown domain for item %p", __func__, item));
3436 return (domain);
3437 }
3438 #endif
3439
3440 #if defined(INVARIANTS) || defined(DEBUG_MEMGUARD) || defined(WITNESS)
3441 #define UMA_ZALLOC_DEBUG
3442 static int
uma_zalloc_debug(uma_zone_t zone,void ** itemp,void * udata,int flags)3443 uma_zalloc_debug(uma_zone_t zone, void **itemp, void *udata, int flags)
3444 {
3445 int error;
3446
3447 error = 0;
3448 #ifdef WITNESS
3449 if (flags & M_WAITOK) {
3450 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3451 "uma_zalloc_debug: zone \"%s\"", zone->uz_name);
3452 }
3453 #endif
3454
3455 #ifdef INVARIANTS
3456 KASSERT((flags & M_EXEC) == 0,
3457 ("uma_zalloc_debug: called with M_EXEC"));
3458 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3459 ("uma_zalloc_debug: called within spinlock or critical section"));
3460 KASSERT((zone->uz_flags & UMA_ZONE_PCPU) == 0 || (flags & M_ZERO) == 0,
3461 ("uma_zalloc_debug: allocating from a pcpu zone with M_ZERO"));
3462 #endif
3463
3464 #ifdef DEBUG_MEMGUARD
3465 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && memguard_cmp_zone(zone)) {
3466 void *item;
3467 item = memguard_alloc(zone->uz_size, flags);
3468 if (item != NULL) {
3469 error = EJUSTRETURN;
3470 if (zone->uz_init != NULL &&
3471 zone->uz_init(item, zone->uz_size, flags) != 0) {
3472 *itemp = NULL;
3473 return (error);
3474 }
3475 if (zone->uz_ctor != NULL &&
3476 zone->uz_ctor(item, zone->uz_size, udata,
3477 flags) != 0) {
3478 counter_u64_add(zone->uz_fails, 1);
3479 if (zone->uz_fini != NULL)
3480 zone->uz_fini(item, zone->uz_size);
3481 *itemp = NULL;
3482 return (error);
3483 }
3484 *itemp = item;
3485 return (error);
3486 }
3487 /* This is unfortunate but should not be fatal. */
3488 }
3489 #endif
3490 return (error);
3491 }
3492
3493 static int
uma_zfree_debug(uma_zone_t zone,void * item,void * udata)3494 uma_zfree_debug(uma_zone_t zone, void *item, void *udata)
3495 {
3496 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3497 ("uma_zfree_debug: called with spinlock or critical section held"));
3498
3499 #ifdef DEBUG_MEMGUARD
3500 if ((zone->uz_flags & UMA_ZONE_SMR) == 0 && is_memguard_addr(item)) {
3501 if (zone->uz_dtor != NULL)
3502 zone->uz_dtor(item, zone->uz_size, udata);
3503 if (zone->uz_fini != NULL)
3504 zone->uz_fini(item, zone->uz_size);
3505 memguard_free(item);
3506 return (EJUSTRETURN);
3507 }
3508 #endif
3509 return (0);
3510 }
3511 #endif
3512
3513 static inline void *
cache_alloc_item(uma_zone_t zone,uma_cache_t cache,uma_cache_bucket_t bucket,void * udata,int flags)3514 cache_alloc_item(uma_zone_t zone, uma_cache_t cache, uma_cache_bucket_t bucket,
3515 void *udata, int flags)
3516 {
3517 void *item;
3518 int size, uz_flags;
3519
3520 item = cache_bucket_pop(cache, bucket);
3521 size = cache_uz_size(cache);
3522 uz_flags = cache_uz_flags(cache);
3523 critical_exit();
3524 return (item_ctor(zone, uz_flags, size, udata, flags, item));
3525 }
3526
3527 static __noinline void *
cache_alloc_retry(uma_zone_t zone,uma_cache_t cache,void * udata,int flags)3528 cache_alloc_retry(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3529 {
3530 uma_cache_bucket_t bucket;
3531 int domain;
3532
3533 while (cache_alloc(zone, cache, udata, flags)) {
3534 cache = &zone->uz_cpu[curcpu];
3535 bucket = &cache->uc_allocbucket;
3536 if (__predict_false(bucket->ucb_cnt == 0))
3537 continue;
3538 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3539 }
3540 critical_exit();
3541
3542 /*
3543 * We can not get a bucket so try to return a single item.
3544 */
3545 if (zone->uz_flags & UMA_ZONE_FIRSTTOUCH)
3546 domain = PCPU_GET(domain);
3547 else
3548 domain = UMA_ANYDOMAIN;
3549 return (zone_alloc_item(zone, udata, domain, flags));
3550 }
3551
3552 /* See uma.h */
3553 void *
uma_zalloc_smr(uma_zone_t zone,int flags)3554 uma_zalloc_smr(uma_zone_t zone, int flags)
3555 {
3556 uma_cache_bucket_t bucket;
3557 uma_cache_t cache;
3558
3559 #ifdef UMA_ZALLOC_DEBUG
3560 void *item;
3561
3562 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
3563 ("uma_zalloc_arg: called with non-SMR zone."));
3564 if (uma_zalloc_debug(zone, &item, NULL, flags) == EJUSTRETURN)
3565 return (item);
3566 #endif
3567
3568 critical_enter();
3569 cache = &zone->uz_cpu[curcpu];
3570 bucket = &cache->uc_allocbucket;
3571 if (__predict_false(bucket->ucb_cnt == 0))
3572 return (cache_alloc_retry(zone, cache, NULL, flags));
3573 return (cache_alloc_item(zone, cache, bucket, NULL, flags));
3574 }
3575
3576 /* See uma.h */
3577 void *
uma_zalloc_arg(uma_zone_t zone,void * udata,int flags)3578 uma_zalloc_arg(uma_zone_t zone, void *udata, int flags)
3579 {
3580 uma_cache_bucket_t bucket;
3581 uma_cache_t cache;
3582
3583 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3584 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3585
3586 /* This is the fast path allocation */
3587 CTR3(KTR_UMA, "uma_zalloc_arg zone %s(%p) flags %d", zone->uz_name,
3588 zone, flags);
3589
3590 #ifdef UMA_ZALLOC_DEBUG
3591 void *item;
3592
3593 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3594 ("uma_zalloc_arg: called with SMR zone."));
3595 if (uma_zalloc_debug(zone, &item, udata, flags) == EJUSTRETURN)
3596 return (item);
3597 #endif
3598
3599 /*
3600 * If possible, allocate from the per-CPU cache. There are two
3601 * requirements for safe access to the per-CPU cache: (1) the thread
3602 * accessing the cache must not be preempted or yield during access,
3603 * and (2) the thread must not migrate CPUs without switching which
3604 * cache it accesses. We rely on a critical section to prevent
3605 * preemption and migration. We release the critical section in
3606 * order to acquire the zone mutex if we are unable to allocate from
3607 * the current cache; when we re-acquire the critical section, we
3608 * must detect and handle migration if it has occurred.
3609 */
3610 critical_enter();
3611 cache = &zone->uz_cpu[curcpu];
3612 bucket = &cache->uc_allocbucket;
3613 if (__predict_false(bucket->ucb_cnt == 0))
3614 return (cache_alloc_retry(zone, cache, udata, flags));
3615 return (cache_alloc_item(zone, cache, bucket, udata, flags));
3616 }
3617
3618 /*
3619 * Replenish an alloc bucket and possibly restore an old one. Called in
3620 * a critical section. Returns in a critical section.
3621 *
3622 * A false return value indicates an allocation failure.
3623 * A true return value indicates success and the caller should retry.
3624 */
3625 static __noinline bool
cache_alloc(uma_zone_t zone,uma_cache_t cache,void * udata,int flags)3626 cache_alloc(uma_zone_t zone, uma_cache_t cache, void *udata, int flags)
3627 {
3628 uma_bucket_t bucket;
3629 int curdomain, domain;
3630 bool new;
3631
3632 CRITICAL_ASSERT(curthread);
3633
3634 /*
3635 * If we have run out of items in our alloc bucket see
3636 * if we can switch with the free bucket.
3637 *
3638 * SMR Zones can't re-use the free bucket until the sequence has
3639 * expired.
3640 */
3641 if ((cache_uz_flags(cache) & UMA_ZONE_SMR) == 0 &&
3642 cache->uc_freebucket.ucb_cnt != 0) {
3643 cache_bucket_swap(&cache->uc_freebucket,
3644 &cache->uc_allocbucket);
3645 return (true);
3646 }
3647
3648 /*
3649 * Discard any empty allocation bucket while we hold no locks.
3650 */
3651 bucket = cache_bucket_unload_alloc(cache);
3652 critical_exit();
3653
3654 if (bucket != NULL) {
3655 KASSERT(bucket->ub_cnt == 0,
3656 ("cache_alloc: Entered with non-empty alloc bucket."));
3657 bucket_free(zone, bucket, udata);
3658 }
3659
3660 /*
3661 * Attempt to retrieve the item from the per-CPU cache has failed, so
3662 * we must go back to the zone. This requires the zdom lock, so we
3663 * must drop the critical section, then re-acquire it when we go back
3664 * to the cache. Since the critical section is released, we may be
3665 * preempted or migrate. As such, make sure not to maintain any
3666 * thread-local state specific to the cache from prior to releasing
3667 * the critical section.
3668 */
3669 domain = PCPU_GET(domain);
3670 if ((cache_uz_flags(cache) & UMA_ZONE_ROUNDROBIN) != 0 ||
3671 VM_DOMAIN_EMPTY(domain))
3672 domain = zone_domain_highest(zone, domain);
3673 bucket = cache_fetch_bucket(zone, cache, domain);
3674 if (bucket == NULL && zone->uz_bucket_size != 0 && !bucketdisable) {
3675 bucket = zone_alloc_bucket(zone, udata, domain, flags);
3676 new = true;
3677 } else {
3678 new = false;
3679 }
3680
3681 CTR3(KTR_UMA, "uma_zalloc: zone %s(%p) bucket zone returned %p",
3682 zone->uz_name, zone, bucket);
3683 if (bucket == NULL) {
3684 critical_enter();
3685 return (false);
3686 }
3687
3688 /*
3689 * See if we lost the race or were migrated. Cache the
3690 * initialized bucket to make this less likely or claim
3691 * the memory directly.
3692 */
3693 critical_enter();
3694 cache = &zone->uz_cpu[curcpu];
3695 if (cache->uc_allocbucket.ucb_bucket == NULL &&
3696 ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) == 0 ||
3697 (curdomain = PCPU_GET(domain)) == domain ||
3698 VM_DOMAIN_EMPTY(curdomain))) {
3699 if (new)
3700 atomic_add_long(&ZDOM_GET(zone, domain)->uzd_imax,
3701 bucket->ub_cnt);
3702 cache_bucket_load_alloc(cache, bucket);
3703 return (true);
3704 }
3705
3706 /*
3707 * We lost the race, release this bucket and start over.
3708 */
3709 critical_exit();
3710 zone_put_bucket(zone, domain, bucket, udata, !new);
3711 critical_enter();
3712
3713 return (true);
3714 }
3715
3716 void *
uma_zalloc_domain(uma_zone_t zone,void * udata,int domain,int flags)3717 uma_zalloc_domain(uma_zone_t zone, void *udata, int domain, int flags)
3718 {
3719 #ifdef NUMA
3720 uma_bucket_t bucket;
3721 uma_zone_domain_t zdom;
3722 void *item;
3723 #endif
3724
3725 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
3726 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
3727
3728 /* This is the fast path allocation */
3729 CTR4(KTR_UMA, "uma_zalloc_domain zone %s(%p) domain %d flags %d",
3730 zone->uz_name, zone, domain, flags);
3731
3732 if (flags & M_WAITOK) {
3733 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL,
3734 "uma_zalloc_domain: zone \"%s\"", zone->uz_name);
3735 }
3736 KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(),
3737 ("uma_zalloc_domain: called with spinlock or critical section held"));
3738 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
3739 ("uma_zalloc_domain: called with SMR zone."));
3740 #ifdef NUMA
3741 KASSERT((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0,
3742 ("uma_zalloc_domain: called with non-FIRSTTOUCH zone."));
3743
3744 if (vm_ndomains == 1)
3745 return (uma_zalloc_arg(zone, udata, flags));
3746
3747 /*
3748 * Try to allocate from the bucket cache before falling back to the keg.
3749 * We could try harder and attempt to allocate from per-CPU caches or
3750 * the per-domain cross-domain buckets, but the complexity is probably
3751 * not worth it. It is more important that frees of previous
3752 * cross-domain allocations do not blow up the cache.
3753 */
3754 zdom = zone_domain_lock(zone, domain);
3755 if ((bucket = zone_fetch_bucket(zone, zdom, false)) != NULL) {
3756 item = bucket->ub_bucket[bucket->ub_cnt - 1];
3757 #ifdef INVARIANTS
3758 bucket->ub_bucket[bucket->ub_cnt - 1] = NULL;
3759 #endif
3760 bucket->ub_cnt--;
3761 zone_put_bucket(zone, domain, bucket, udata, true);
3762 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata,
3763 flags, item);
3764 if (item != NULL) {
3765 KASSERT(item_domain(item) == domain,
3766 ("%s: bucket cache item %p from wrong domain",
3767 __func__, item));
3768 counter_u64_add(zone->uz_allocs, 1);
3769 }
3770 return (item);
3771 }
3772 ZDOM_UNLOCK(zdom);
3773 return (zone_alloc_item(zone, udata, domain, flags));
3774 #else
3775 return (uma_zalloc_arg(zone, udata, flags));
3776 #endif
3777 }
3778
3779 /*
3780 * Find a slab with some space. Prefer slabs that are partially used over those
3781 * that are totally full. This helps to reduce fragmentation.
3782 *
3783 * If 'rr' is 1, search all domains starting from 'domain'. Otherwise check
3784 * only 'domain'.
3785 */
3786 static uma_slab_t
keg_first_slab(uma_keg_t keg,int domain,bool rr)3787 keg_first_slab(uma_keg_t keg, int domain, bool rr)
3788 {
3789 uma_domain_t dom;
3790 uma_slab_t slab;
3791 int start;
3792
3793 KASSERT(domain >= 0 && domain < vm_ndomains,
3794 ("keg_first_slab: domain %d out of range", domain));
3795 KEG_LOCK_ASSERT(keg, domain);
3796
3797 slab = NULL;
3798 start = domain;
3799 do {
3800 dom = &keg->uk_domain[domain];
3801 if ((slab = LIST_FIRST(&dom->ud_part_slab)) != NULL)
3802 return (slab);
3803 if ((slab = LIST_FIRST(&dom->ud_free_slab)) != NULL) {
3804 LIST_REMOVE(slab, us_link);
3805 dom->ud_free_slabs--;
3806 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
3807 return (slab);
3808 }
3809 if (rr)
3810 domain = (domain + 1) % vm_ndomains;
3811 } while (domain != start);
3812
3813 return (NULL);
3814 }
3815
3816 /*
3817 * Fetch an existing slab from a free or partial list. Returns with the
3818 * keg domain lock held if a slab was found or unlocked if not.
3819 */
3820 static uma_slab_t
keg_fetch_free_slab(uma_keg_t keg,int domain,bool rr,int flags)3821 keg_fetch_free_slab(uma_keg_t keg, int domain, bool rr, int flags)
3822 {
3823 uma_slab_t slab;
3824 uint32_t reserve;
3825
3826 /* HASH has a single free list. */
3827 if ((keg->uk_flags & UMA_ZFLAG_HASH) != 0)
3828 domain = 0;
3829
3830 KEG_LOCK(keg, domain);
3831 reserve = (flags & M_USE_RESERVE) != 0 ? 0 : keg->uk_reserve;
3832 if (keg->uk_domain[domain].ud_free_items <= reserve ||
3833 (slab = keg_first_slab(keg, domain, rr)) == NULL) {
3834 KEG_UNLOCK(keg, domain);
3835 return (NULL);
3836 }
3837 return (slab);
3838 }
3839
3840 static uma_slab_t
keg_fetch_slab(uma_keg_t keg,uma_zone_t zone,int rdomain,const int flags)3841 keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int rdomain, const int flags)
3842 {
3843 struct vm_domainset_iter di;
3844 uma_slab_t slab;
3845 int aflags, domain;
3846 bool rr;
3847
3848 KASSERT((flags & (M_WAITOK | M_NOVM)) != (M_WAITOK | M_NOVM),
3849 ("%s: invalid flags %#x", __func__, flags));
3850
3851 restart:
3852 /*
3853 * Use the keg's policy if upper layers haven't already specified a
3854 * domain (as happens with first-touch zones).
3855 *
3856 * To avoid races we run the iterator with the keg lock held, but that
3857 * means that we cannot allow the vm_domainset layer to sleep. Thus,
3858 * clear M_WAITOK and handle low memory conditions locally.
3859 */
3860 rr = rdomain == UMA_ANYDOMAIN;
3861 if (rr) {
3862 aflags = (flags & ~M_WAITOK) | M_NOWAIT;
3863 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
3864 &aflags);
3865 } else {
3866 aflags = flags;
3867 domain = rdomain;
3868 }
3869
3870 for (;;) {
3871 slab = keg_fetch_free_slab(keg, domain, rr, flags);
3872 if (slab != NULL)
3873 return (slab);
3874
3875 /*
3876 * M_NOVM is used to break the recursion that can otherwise
3877 * occur if low-level memory management routines use UMA.
3878 */
3879 if ((flags & M_NOVM) == 0) {
3880 slab = keg_alloc_slab(keg, zone, domain, flags, aflags);
3881 if (slab != NULL)
3882 return (slab);
3883 }
3884
3885 if (!rr) {
3886 if ((flags & M_USE_RESERVE) != 0) {
3887 /*
3888 * Drain reserves from other domains before
3889 * giving up or sleeping. It may be useful to
3890 * support per-domain reserves eventually.
3891 */
3892 rdomain = UMA_ANYDOMAIN;
3893 goto restart;
3894 }
3895 if ((flags & M_WAITOK) == 0)
3896 break;
3897 vm_wait_domain(domain);
3898 } else if (vm_domainset_iter_policy(&di, &domain) != 0) {
3899 if ((flags & M_WAITOK) != 0) {
3900 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
3901 goto restart;
3902 }
3903 break;
3904 }
3905 }
3906
3907 /*
3908 * We might not have been able to get a slab but another cpu
3909 * could have while we were unlocked. Check again before we
3910 * fail.
3911 */
3912 if ((slab = keg_fetch_free_slab(keg, domain, rr, flags)) != NULL)
3913 return (slab);
3914
3915 return (NULL);
3916 }
3917
3918 static void *
slab_alloc_item(uma_keg_t keg,uma_slab_t slab)3919 slab_alloc_item(uma_keg_t keg, uma_slab_t slab)
3920 {
3921 uma_domain_t dom;
3922 void *item;
3923 int freei;
3924
3925 KEG_LOCK_ASSERT(keg, slab->us_domain);
3926
3927 dom = &keg->uk_domain[slab->us_domain];
3928 freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1;
3929 BIT_CLR(keg->uk_ipers, freei, &slab->us_free);
3930 item = slab_item(slab, keg, freei);
3931 slab->us_freecount--;
3932 dom->ud_free_items--;
3933
3934 /*
3935 * Move this slab to the full list. It must be on the partial list, so
3936 * we do not need to update the free slab count. In particular,
3937 * keg_fetch_slab() always returns slabs on the partial list.
3938 */
3939 if (slab->us_freecount == 0) {
3940 LIST_REMOVE(slab, us_link);
3941 LIST_INSERT_HEAD(&dom->ud_full_slab, slab, us_link);
3942 }
3943
3944 return (item);
3945 }
3946
3947 static int
zone_import(void * arg,void ** bucket,int max,int domain,int flags)3948 zone_import(void *arg, void **bucket, int max, int domain, int flags)
3949 {
3950 uma_domain_t dom;
3951 uma_zone_t zone;
3952 uma_slab_t slab;
3953 uma_keg_t keg;
3954 #ifdef NUMA
3955 int stripe;
3956 #endif
3957 int i;
3958
3959 zone = arg;
3960 slab = NULL;
3961 keg = zone->uz_keg;
3962 /* Try to keep the buckets totally full */
3963 for (i = 0; i < max; ) {
3964 if ((slab = keg_fetch_slab(keg, zone, domain, flags)) == NULL)
3965 break;
3966 #ifdef NUMA
3967 stripe = howmany(max, vm_ndomains);
3968 #endif
3969 dom = &keg->uk_domain[slab->us_domain];
3970 do {
3971 bucket[i++] = slab_alloc_item(keg, slab);
3972 if (keg->uk_reserve > 0 &&
3973 dom->ud_free_items <= keg->uk_reserve) {
3974 /*
3975 * Avoid depleting the reserve after a
3976 * successful item allocation, even if
3977 * M_USE_RESERVE is specified.
3978 */
3979 KEG_UNLOCK(keg, slab->us_domain);
3980 goto out;
3981 }
3982 #ifdef NUMA
3983 /*
3984 * If the zone is striped we pick a new slab for every
3985 * N allocations. Eliminating this conditional will
3986 * instead pick a new domain for each bucket rather
3987 * than stripe within each bucket. The current option
3988 * produces more fragmentation and requires more cpu
3989 * time but yields better distribution.
3990 */
3991 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0 &&
3992 vm_ndomains > 1 && --stripe == 0)
3993 break;
3994 #endif
3995 } while (slab->us_freecount != 0 && i < max);
3996 KEG_UNLOCK(keg, slab->us_domain);
3997
3998 /* Don't block if we allocated any successfully. */
3999 flags &= ~M_WAITOK;
4000 flags |= M_NOWAIT;
4001 }
4002 out:
4003 return i;
4004 }
4005
4006 static int
zone_alloc_limit_hard(uma_zone_t zone,int count,int flags)4007 zone_alloc_limit_hard(uma_zone_t zone, int count, int flags)
4008 {
4009 uint64_t old, new, total, max;
4010
4011 /*
4012 * The hard case. We're going to sleep because there were existing
4013 * sleepers or because we ran out of items. This routine enforces
4014 * fairness by keeping fifo order.
4015 *
4016 * First release our ill gotten gains and make some noise.
4017 */
4018 for (;;) {
4019 zone_free_limit(zone, count);
4020 zone_log_warning(zone);
4021 zone_maxaction(zone);
4022 if (flags & M_NOWAIT)
4023 return (0);
4024
4025 /*
4026 * We need to allocate an item or set ourself as a sleeper
4027 * while the sleepq lock is held to avoid wakeup races. This
4028 * is essentially a home rolled semaphore.
4029 */
4030 sleepq_lock(&zone->uz_max_items);
4031 old = zone->uz_items;
4032 do {
4033 MPASS(UZ_ITEMS_SLEEPERS(old) < UZ_ITEMS_SLEEPERS_MAX);
4034 /* Cache the max since we will evaluate twice. */
4035 max = zone->uz_max_items;
4036 if (UZ_ITEMS_SLEEPERS(old) != 0 ||
4037 UZ_ITEMS_COUNT(old) >= max)
4038 new = old + UZ_ITEMS_SLEEPER;
4039 else
4040 new = old + MIN(count, max - old);
4041 } while (atomic_fcmpset_64(&zone->uz_items, &old, new) == 0);
4042
4043 /* We may have successfully allocated under the sleepq lock. */
4044 if (UZ_ITEMS_SLEEPERS(new) == 0) {
4045 sleepq_release(&zone->uz_max_items);
4046 return (new - old);
4047 }
4048
4049 /*
4050 * This is in a different cacheline from uz_items so that we
4051 * don't constantly invalidate the fastpath cacheline when we
4052 * adjust item counts. This could be limited to toggling on
4053 * transitions.
4054 */
4055 atomic_add_32(&zone->uz_sleepers, 1);
4056 atomic_add_64(&zone->uz_sleeps, 1);
4057
4058 /*
4059 * We have added ourselves as a sleeper. The sleepq lock
4060 * protects us from wakeup races. Sleep now and then retry.
4061 */
4062 sleepq_add(&zone->uz_max_items, NULL, "zonelimit", 0, 0);
4063 sleepq_wait(&zone->uz_max_items, PVM);
4064
4065 /*
4066 * After wakeup, remove ourselves as a sleeper and try
4067 * again. We no longer have the sleepq lock for protection.
4068 *
4069 * Subract ourselves as a sleeper while attempting to add
4070 * our count.
4071 */
4072 atomic_subtract_32(&zone->uz_sleepers, 1);
4073 old = atomic_fetchadd_64(&zone->uz_items,
4074 -(UZ_ITEMS_SLEEPER - count));
4075 /* We're no longer a sleeper. */
4076 old -= UZ_ITEMS_SLEEPER;
4077
4078 /*
4079 * If we're still at the limit, restart. Notably do not
4080 * block on other sleepers. Cache the max value to protect
4081 * against changes via sysctl.
4082 */
4083 total = UZ_ITEMS_COUNT(old);
4084 max = zone->uz_max_items;
4085 if (total >= max)
4086 continue;
4087 /* Truncate if necessary, otherwise wake other sleepers. */
4088 if (total + count > max) {
4089 zone_free_limit(zone, total + count - max);
4090 count = max - total;
4091 } else if (total + count < max && UZ_ITEMS_SLEEPERS(old) != 0)
4092 wakeup_one(&zone->uz_max_items);
4093
4094 return (count);
4095 }
4096 }
4097
4098 /*
4099 * Allocate 'count' items from our max_items limit. Returns the number
4100 * available. If M_NOWAIT is not specified it will sleep until at least
4101 * one item can be allocated.
4102 */
4103 static int
zone_alloc_limit(uma_zone_t zone,int count,int flags)4104 zone_alloc_limit(uma_zone_t zone, int count, int flags)
4105 {
4106 uint64_t old;
4107 uint64_t max;
4108
4109 max = zone->uz_max_items;
4110 MPASS(max > 0);
4111
4112 /*
4113 * We expect normal allocations to succeed with a simple
4114 * fetchadd.
4115 */
4116 old = atomic_fetchadd_64(&zone->uz_items, count);
4117 if (__predict_true(old + count <= max))
4118 return (count);
4119
4120 /*
4121 * If we had some items and no sleepers just return the
4122 * truncated value. We have to release the excess space
4123 * though because that may wake sleepers who weren't woken
4124 * because we were temporarily over the limit.
4125 */
4126 if (old < max) {
4127 zone_free_limit(zone, (old + count) - max);
4128 return (max - old);
4129 }
4130 return (zone_alloc_limit_hard(zone, count, flags));
4131 }
4132
4133 /*
4134 * Free a number of items back to the limit.
4135 */
4136 static void
zone_free_limit(uma_zone_t zone,int count)4137 zone_free_limit(uma_zone_t zone, int count)
4138 {
4139 uint64_t old;
4140
4141 MPASS(count > 0);
4142
4143 /*
4144 * In the common case we either have no sleepers or
4145 * are still over the limit and can just return.
4146 */
4147 old = atomic_fetchadd_64(&zone->uz_items, -count);
4148 if (__predict_true(UZ_ITEMS_SLEEPERS(old) == 0 ||
4149 UZ_ITEMS_COUNT(old) - count >= zone->uz_max_items))
4150 return;
4151
4152 /*
4153 * Moderate the rate of wakeups. Sleepers will continue
4154 * to generate wakeups if necessary.
4155 */
4156 wakeup_one(&zone->uz_max_items);
4157 }
4158
4159 static uma_bucket_t
zone_alloc_bucket(uma_zone_t zone,void * udata,int domain,int flags)4160 zone_alloc_bucket(uma_zone_t zone, void *udata, int domain, int flags)
4161 {
4162 uma_bucket_t bucket;
4163 int error, maxbucket, cnt;
4164
4165 CTR3(KTR_UMA, "zone_alloc_bucket zone %s(%p) domain %d", zone->uz_name,
4166 zone, domain);
4167
4168 /* Avoid allocs targeting empty domains. */
4169 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4170 domain = UMA_ANYDOMAIN;
4171 else if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4172 domain = UMA_ANYDOMAIN;
4173
4174 if (zone->uz_max_items > 0)
4175 maxbucket = zone_alloc_limit(zone, zone->uz_bucket_size,
4176 M_NOWAIT);
4177 else
4178 maxbucket = zone->uz_bucket_size;
4179 if (maxbucket == 0)
4180 return (false);
4181
4182 /* Don't wait for buckets, preserve caller's NOVM setting. */
4183 bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM));
4184 if (bucket == NULL) {
4185 cnt = 0;
4186 goto out;
4187 }
4188
4189 bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket,
4190 MIN(maxbucket, bucket->ub_entries), domain, flags);
4191
4192 /*
4193 * Initialize the memory if necessary.
4194 */
4195 if (bucket->ub_cnt != 0 && zone->uz_init != NULL) {
4196 int i;
4197
4198 for (i = 0; i < bucket->ub_cnt; i++) {
4199 kasan_mark_item_valid(zone, bucket->ub_bucket[i]);
4200 error = zone->uz_init(bucket->ub_bucket[i],
4201 zone->uz_size, flags);
4202 kasan_mark_item_invalid(zone, bucket->ub_bucket[i]);
4203 if (error != 0)
4204 break;
4205 }
4206
4207 /*
4208 * If we couldn't initialize the whole bucket, put the
4209 * rest back onto the freelist.
4210 */
4211 if (i != bucket->ub_cnt) {
4212 zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i],
4213 bucket->ub_cnt - i);
4214 #ifdef INVARIANTS
4215 bzero(&bucket->ub_bucket[i],
4216 sizeof(void *) * (bucket->ub_cnt - i));
4217 #endif
4218 bucket->ub_cnt = i;
4219 }
4220 }
4221
4222 cnt = bucket->ub_cnt;
4223 if (bucket->ub_cnt == 0) {
4224 bucket_free(zone, bucket, udata);
4225 counter_u64_add(zone->uz_fails, 1);
4226 bucket = NULL;
4227 }
4228 out:
4229 if (zone->uz_max_items > 0 && cnt < maxbucket)
4230 zone_free_limit(zone, maxbucket - cnt);
4231
4232 return (bucket);
4233 }
4234
4235 /*
4236 * Allocates a single item from a zone.
4237 *
4238 * Arguments
4239 * zone The zone to alloc for.
4240 * udata The data to be passed to the constructor.
4241 * domain The domain to allocate from or UMA_ANYDOMAIN.
4242 * flags M_WAITOK, M_NOWAIT, M_ZERO.
4243 *
4244 * Returns
4245 * NULL if there is no memory and M_NOWAIT is set
4246 * An item if successful
4247 */
4248
4249 static void *
zone_alloc_item(uma_zone_t zone,void * udata,int domain,int flags)4250 zone_alloc_item(uma_zone_t zone, void *udata, int domain, int flags)
4251 {
4252 void *item;
4253
4254 if (zone->uz_max_items > 0 && zone_alloc_limit(zone, 1, flags) == 0) {
4255 counter_u64_add(zone->uz_fails, 1);
4256 return (NULL);
4257 }
4258
4259 /* Avoid allocs targeting empty domains. */
4260 if (domain != UMA_ANYDOMAIN && VM_DOMAIN_EMPTY(domain))
4261 domain = UMA_ANYDOMAIN;
4262
4263 if (zone->uz_import(zone->uz_arg, &item, 1, domain, flags) != 1)
4264 goto fail_cnt;
4265
4266 /*
4267 * We have to call both the zone's init (not the keg's init)
4268 * and the zone's ctor. This is because the item is going from
4269 * a keg slab directly to the user, and the user is expecting it
4270 * to be both zone-init'd as well as zone-ctor'd.
4271 */
4272 if (zone->uz_init != NULL) {
4273 int error;
4274
4275 kasan_mark_item_valid(zone, item);
4276 error = zone->uz_init(item, zone->uz_size, flags);
4277 kasan_mark_item_invalid(zone, item);
4278 if (error != 0) {
4279 zone_free_item(zone, item, udata, SKIP_FINI | SKIP_CNT);
4280 goto fail_cnt;
4281 }
4282 }
4283 item = item_ctor(zone, zone->uz_flags, zone->uz_size, udata, flags,
4284 item);
4285 if (item == NULL)
4286 goto fail;
4287
4288 counter_u64_add(zone->uz_allocs, 1);
4289 CTR3(KTR_UMA, "zone_alloc_item item %p from %s(%p)", item,
4290 zone->uz_name, zone);
4291
4292 return (item);
4293
4294 fail_cnt:
4295 counter_u64_add(zone->uz_fails, 1);
4296 fail:
4297 if (zone->uz_max_items > 0)
4298 zone_free_limit(zone, 1);
4299 CTR2(KTR_UMA, "zone_alloc_item failed from %s(%p)",
4300 zone->uz_name, zone);
4301
4302 return (NULL);
4303 }
4304
4305 /* See uma.h */
4306 void
uma_zfree_smr(uma_zone_t zone,void * item)4307 uma_zfree_smr(uma_zone_t zone, void *item)
4308 {
4309 uma_cache_t cache;
4310 uma_cache_bucket_t bucket;
4311 int itemdomain, uz_flags;
4312
4313 #ifdef UMA_ZALLOC_DEBUG
4314 KASSERT((zone->uz_flags & UMA_ZONE_SMR) != 0,
4315 ("uma_zfree_smr: called with non-SMR zone."));
4316 KASSERT(item != NULL, ("uma_zfree_smr: Called with NULL pointer."));
4317 SMR_ASSERT_NOT_ENTERED(zone->uz_smr);
4318 if (uma_zfree_debug(zone, item, NULL) == EJUSTRETURN)
4319 return;
4320 #endif
4321 cache = &zone->uz_cpu[curcpu];
4322 uz_flags = cache_uz_flags(cache);
4323 itemdomain = 0;
4324 #ifdef NUMA
4325 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4326 itemdomain = item_domain(item);
4327 #endif
4328 critical_enter();
4329 do {
4330 cache = &zone->uz_cpu[curcpu];
4331 /* SMR Zones must free to the free bucket. */
4332 bucket = &cache->uc_freebucket;
4333 #ifdef NUMA
4334 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4335 PCPU_GET(domain) != itemdomain) {
4336 bucket = &cache->uc_crossbucket;
4337 }
4338 #endif
4339 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4340 cache_bucket_push(cache, bucket, item);
4341 critical_exit();
4342 return;
4343 }
4344 } while (cache_free(zone, cache, NULL, item, itemdomain));
4345 critical_exit();
4346
4347 /*
4348 * If nothing else caught this, we'll just do an internal free.
4349 */
4350 zone_free_item(zone, item, NULL, SKIP_NONE);
4351 }
4352
4353 /* See uma.h */
4354 void
uma_zfree_arg(uma_zone_t zone,void * item,void * udata)4355 uma_zfree_arg(uma_zone_t zone, void *item, void *udata)
4356 {
4357 uma_cache_t cache;
4358 uma_cache_bucket_t bucket;
4359 int itemdomain, uz_flags;
4360
4361 /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */
4362 random_harvest_fast_uma(&zone, sizeof(zone), RANDOM_UMA);
4363
4364 CTR2(KTR_UMA, "uma_zfree_arg zone %s(%p)", zone->uz_name, zone);
4365
4366 #ifdef UMA_ZALLOC_DEBUG
4367 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
4368 ("uma_zfree_arg: called with SMR zone."));
4369 if (uma_zfree_debug(zone, item, udata) == EJUSTRETURN)
4370 return;
4371 #endif
4372 /* uma_zfree(..., NULL) does nothing, to match free(9). */
4373 if (item == NULL)
4374 return;
4375
4376 /*
4377 * We are accessing the per-cpu cache without a critical section to
4378 * fetch size and flags. This is acceptable, if we are preempted we
4379 * will simply read another cpu's line.
4380 */
4381 cache = &zone->uz_cpu[curcpu];
4382 uz_flags = cache_uz_flags(cache);
4383 if (UMA_ALWAYS_CTORDTOR ||
4384 __predict_false((uz_flags & UMA_ZFLAG_CTORDTOR) != 0))
4385 item_dtor(zone, item, cache_uz_size(cache), udata, SKIP_NONE);
4386
4387 /*
4388 * The race here is acceptable. If we miss it we'll just have to wait
4389 * a little longer for the limits to be reset.
4390 */
4391 if (__predict_false(uz_flags & UMA_ZFLAG_LIMIT)) {
4392 if (atomic_load_32(&zone->uz_sleepers) > 0)
4393 goto zfree_item;
4394 }
4395
4396 /*
4397 * If possible, free to the per-CPU cache. There are two
4398 * requirements for safe access to the per-CPU cache: (1) the thread
4399 * accessing the cache must not be preempted or yield during access,
4400 * and (2) the thread must not migrate CPUs without switching which
4401 * cache it accesses. We rely on a critical section to prevent
4402 * preemption and migration. We release the critical section in
4403 * order to acquire the zone mutex if we are unable to free to the
4404 * current cache; when we re-acquire the critical section, we must
4405 * detect and handle migration if it has occurred.
4406 */
4407 itemdomain = 0;
4408 #ifdef NUMA
4409 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0)
4410 itemdomain = item_domain(item);
4411 #endif
4412 critical_enter();
4413 do {
4414 cache = &zone->uz_cpu[curcpu];
4415 /*
4416 * Try to free into the allocbucket first to give LIFO
4417 * ordering for cache-hot datastructures. Spill over
4418 * into the freebucket if necessary. Alloc will swap
4419 * them if one runs dry.
4420 */
4421 bucket = &cache->uc_allocbucket;
4422 #ifdef NUMA
4423 if ((uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4424 PCPU_GET(domain) != itemdomain) {
4425 bucket = &cache->uc_crossbucket;
4426 } else
4427 #endif
4428 if (bucket->ucb_cnt == bucket->ucb_entries &&
4429 cache->uc_freebucket.ucb_cnt <
4430 cache->uc_freebucket.ucb_entries)
4431 cache_bucket_swap(&cache->uc_freebucket,
4432 &cache->uc_allocbucket);
4433 if (__predict_true(bucket->ucb_cnt < bucket->ucb_entries)) {
4434 cache_bucket_push(cache, bucket, item);
4435 critical_exit();
4436 return;
4437 }
4438 } while (cache_free(zone, cache, udata, item, itemdomain));
4439 critical_exit();
4440
4441 /*
4442 * If nothing else caught this, we'll just do an internal free.
4443 */
4444 zfree_item:
4445 zone_free_item(zone, item, udata, SKIP_DTOR);
4446 }
4447
4448 #ifdef NUMA
4449 /*
4450 * sort crossdomain free buckets to domain correct buckets and cache
4451 * them.
4452 */
4453 static void
zone_free_cross(uma_zone_t zone,uma_bucket_t bucket,void * udata)4454 zone_free_cross(uma_zone_t zone, uma_bucket_t bucket, void *udata)
4455 {
4456 struct uma_bucketlist emptybuckets, fullbuckets;
4457 uma_zone_domain_t zdom;
4458 uma_bucket_t b;
4459 smr_seq_t seq;
4460 void *item;
4461 int domain;
4462
4463 CTR3(KTR_UMA,
4464 "uma_zfree: zone %s(%p) draining cross bucket %p",
4465 zone->uz_name, zone, bucket);
4466
4467 /*
4468 * It is possible for buckets to arrive here out of order so we fetch
4469 * the current smr seq rather than accepting the bucket's.
4470 */
4471 seq = SMR_SEQ_INVALID;
4472 if ((zone->uz_flags & UMA_ZONE_SMR) != 0)
4473 seq = smr_advance(zone->uz_smr);
4474
4475 /*
4476 * To avoid having ndomain * ndomain buckets for sorting we have a
4477 * lock on the current crossfree bucket. A full matrix with
4478 * per-domain locking could be used if necessary.
4479 */
4480 STAILQ_INIT(&emptybuckets);
4481 STAILQ_INIT(&fullbuckets);
4482 ZONE_CROSS_LOCK(zone);
4483 for (; bucket->ub_cnt > 0; bucket->ub_cnt--) {
4484 item = bucket->ub_bucket[bucket->ub_cnt - 1];
4485 domain = item_domain(item);
4486 zdom = ZDOM_GET(zone, domain);
4487 if (zdom->uzd_cross == NULL) {
4488 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4489 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4490 zdom->uzd_cross = b;
4491 } else {
4492 /*
4493 * Avoid allocating a bucket with the cross lock
4494 * held, since allocation can trigger a
4495 * cross-domain free and bucket zones may
4496 * allocate from each other.
4497 */
4498 ZONE_CROSS_UNLOCK(zone);
4499 b = bucket_alloc(zone, udata, M_NOWAIT);
4500 if (b == NULL)
4501 goto out;
4502 ZONE_CROSS_LOCK(zone);
4503 if (zdom->uzd_cross != NULL) {
4504 STAILQ_INSERT_HEAD(&emptybuckets, b,
4505 ub_link);
4506 } else {
4507 zdom->uzd_cross = b;
4508 }
4509 }
4510 }
4511 b = zdom->uzd_cross;
4512 b->ub_bucket[b->ub_cnt++] = item;
4513 b->ub_seq = seq;
4514 if (b->ub_cnt == b->ub_entries) {
4515 STAILQ_INSERT_HEAD(&fullbuckets, b, ub_link);
4516 if ((b = STAILQ_FIRST(&emptybuckets)) != NULL)
4517 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4518 zdom->uzd_cross = b;
4519 }
4520 }
4521 ZONE_CROSS_UNLOCK(zone);
4522 out:
4523 if (bucket->ub_cnt == 0)
4524 bucket->ub_seq = SMR_SEQ_INVALID;
4525 bucket_free(zone, bucket, udata);
4526
4527 while ((b = STAILQ_FIRST(&emptybuckets)) != NULL) {
4528 STAILQ_REMOVE_HEAD(&emptybuckets, ub_link);
4529 bucket_free(zone, b, udata);
4530 }
4531 while ((b = STAILQ_FIRST(&fullbuckets)) != NULL) {
4532 STAILQ_REMOVE_HEAD(&fullbuckets, ub_link);
4533 domain = item_domain(b->ub_bucket[0]);
4534 zone_put_bucket(zone, domain, b, udata, true);
4535 }
4536 }
4537 #endif
4538
4539 static void
zone_free_bucket(uma_zone_t zone,uma_bucket_t bucket,void * udata,int itemdomain,bool ws)4540 zone_free_bucket(uma_zone_t zone, uma_bucket_t bucket, void *udata,
4541 int itemdomain, bool ws)
4542 {
4543
4544 #ifdef NUMA
4545 /*
4546 * Buckets coming from the wrong domain will be entirely for the
4547 * only other domain on two domain systems. In this case we can
4548 * simply cache them. Otherwise we need to sort them back to
4549 * correct domains.
4550 */
4551 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 &&
4552 vm_ndomains > 2 && PCPU_GET(domain) != itemdomain) {
4553 zone_free_cross(zone, bucket, udata);
4554 return;
4555 }
4556 #endif
4557
4558 /*
4559 * Attempt to save the bucket in the zone's domain bucket cache.
4560 */
4561 CTR3(KTR_UMA,
4562 "uma_zfree: zone %s(%p) putting bucket %p on free list",
4563 zone->uz_name, zone, bucket);
4564 /* ub_cnt is pointing to the last free item */
4565 if ((zone->uz_flags & UMA_ZONE_ROUNDROBIN) != 0)
4566 itemdomain = zone_domain_lowest(zone, itemdomain);
4567 zone_put_bucket(zone, itemdomain, bucket, udata, ws);
4568 }
4569
4570 /*
4571 * Populate a free or cross bucket for the current cpu cache. Free any
4572 * existing full bucket either to the zone cache or back to the slab layer.
4573 *
4574 * Enters and returns in a critical section. false return indicates that
4575 * we can not satisfy this free in the cache layer. true indicates that
4576 * the caller should retry.
4577 */
4578 static __noinline bool
cache_free(uma_zone_t zone,uma_cache_t cache,void * udata,void * item,int itemdomain)4579 cache_free(uma_zone_t zone, uma_cache_t cache, void *udata, void *item,
4580 int itemdomain)
4581 {
4582 uma_cache_bucket_t cbucket;
4583 uma_bucket_t newbucket, bucket;
4584
4585 CRITICAL_ASSERT(curthread);
4586
4587 if (zone->uz_bucket_size == 0)
4588 return false;
4589
4590 cache = &zone->uz_cpu[curcpu];
4591 newbucket = NULL;
4592
4593 /*
4594 * FIRSTTOUCH domains need to free to the correct zdom. When
4595 * enabled this is the zdom of the item. The bucket is the
4596 * cross bucket if the current domain and itemdomain do not match.
4597 */
4598 cbucket = &cache->uc_freebucket;
4599 #ifdef NUMA
4600 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4601 if (PCPU_GET(domain) != itemdomain) {
4602 cbucket = &cache->uc_crossbucket;
4603 if (cbucket->ucb_cnt != 0)
4604 counter_u64_add(zone->uz_xdomain,
4605 cbucket->ucb_cnt);
4606 }
4607 }
4608 #endif
4609 bucket = cache_bucket_unload(cbucket);
4610 KASSERT(bucket == NULL || bucket->ub_cnt == bucket->ub_entries,
4611 ("cache_free: Entered with non-full free bucket."));
4612
4613 /* We are no longer associated with this CPU. */
4614 critical_exit();
4615
4616 /*
4617 * Don't let SMR zones operate without a free bucket. Force
4618 * a synchronize and re-use this one. We will only degrade
4619 * to a synchronize every bucket_size items rather than every
4620 * item if we fail to allocate a bucket.
4621 */
4622 if ((zone->uz_flags & UMA_ZONE_SMR) != 0) {
4623 if (bucket != NULL)
4624 bucket->ub_seq = smr_advance(zone->uz_smr);
4625 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4626 if (newbucket == NULL && bucket != NULL) {
4627 bucket_drain(zone, bucket);
4628 newbucket = bucket;
4629 bucket = NULL;
4630 }
4631 } else if (!bucketdisable)
4632 newbucket = bucket_alloc(zone, udata, M_NOWAIT);
4633
4634 if (bucket != NULL)
4635 zone_free_bucket(zone, bucket, udata, itemdomain, true);
4636
4637 critical_enter();
4638 if ((bucket = newbucket) == NULL)
4639 return (false);
4640 cache = &zone->uz_cpu[curcpu];
4641 #ifdef NUMA
4642 /*
4643 * Check to see if we should be populating the cross bucket. If it
4644 * is already populated we will fall through and attempt to populate
4645 * the free bucket.
4646 */
4647 if ((cache_uz_flags(cache) & UMA_ZONE_FIRSTTOUCH) != 0) {
4648 if (PCPU_GET(domain) != itemdomain &&
4649 cache->uc_crossbucket.ucb_bucket == NULL) {
4650 cache_bucket_load_cross(cache, bucket);
4651 return (true);
4652 }
4653 }
4654 #endif
4655 /*
4656 * We may have lost the race to fill the bucket or switched CPUs.
4657 */
4658 if (cache->uc_freebucket.ucb_bucket != NULL) {
4659 critical_exit();
4660 bucket_free(zone, bucket, udata);
4661 critical_enter();
4662 } else
4663 cache_bucket_load_free(cache, bucket);
4664
4665 return (true);
4666 }
4667
4668 static void
slab_free_item(uma_zone_t zone,uma_slab_t slab,void * item)4669 slab_free_item(uma_zone_t zone, uma_slab_t slab, void *item)
4670 {
4671 uma_keg_t keg;
4672 uma_domain_t dom;
4673 int freei;
4674
4675 keg = zone->uz_keg;
4676 KEG_LOCK_ASSERT(keg, slab->us_domain);
4677
4678 /* Do we need to remove from any lists? */
4679 dom = &keg->uk_domain[slab->us_domain];
4680 if (slab->us_freecount + 1 == keg->uk_ipers) {
4681 LIST_REMOVE(slab, us_link);
4682 LIST_INSERT_HEAD(&dom->ud_free_slab, slab, us_link);
4683 dom->ud_free_slabs++;
4684 } else if (slab->us_freecount == 0) {
4685 LIST_REMOVE(slab, us_link);
4686 LIST_INSERT_HEAD(&dom->ud_part_slab, slab, us_link);
4687 }
4688
4689 /* Slab management. */
4690 freei = slab_item_index(slab, keg, item);
4691 BIT_SET(keg->uk_ipers, freei, &slab->us_free);
4692 slab->us_freecount++;
4693
4694 /* Keg statistics. */
4695 dom->ud_free_items++;
4696 }
4697
4698 static void
zone_release(void * arg,void ** bucket,int cnt)4699 zone_release(void *arg, void **bucket, int cnt)
4700 {
4701 struct mtx *lock;
4702 uma_zone_t zone;
4703 uma_slab_t slab;
4704 uma_keg_t keg;
4705 uint8_t *mem;
4706 void *item;
4707 int i;
4708
4709 zone = arg;
4710 keg = zone->uz_keg;
4711 lock = NULL;
4712 if (__predict_false((zone->uz_flags & UMA_ZFLAG_HASH) != 0))
4713 lock = KEG_LOCK(keg, 0);
4714 for (i = 0; i < cnt; i++) {
4715 item = bucket[i];
4716 if (__predict_true((zone->uz_flags & UMA_ZFLAG_VTOSLAB) != 0)) {
4717 slab = vtoslab((vm_offset_t)item);
4718 } else {
4719 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
4720 if ((zone->uz_flags & UMA_ZFLAG_HASH) != 0)
4721 slab = hash_sfind(&keg->uk_hash, mem);
4722 else
4723 slab = (uma_slab_t)(mem + keg->uk_pgoff);
4724 }
4725 if (lock != KEG_LOCKPTR(keg, slab->us_domain)) {
4726 if (lock != NULL)
4727 mtx_unlock(lock);
4728 lock = KEG_LOCK(keg, slab->us_domain);
4729 }
4730 slab_free_item(zone, slab, item);
4731 }
4732 if (lock != NULL)
4733 mtx_unlock(lock);
4734 }
4735
4736 /*
4737 * Frees a single item to any zone.
4738 *
4739 * Arguments:
4740 * zone The zone to free to
4741 * item The item we're freeing
4742 * udata User supplied data for the dtor
4743 * skip Skip dtors and finis
4744 */
4745 static __noinline void
zone_free_item(uma_zone_t zone,void * item,void * udata,enum zfreeskip skip)4746 zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip)
4747 {
4748
4749 /*
4750 * If a free is sent directly to an SMR zone we have to
4751 * synchronize immediately because the item can instantly
4752 * be reallocated. This should only happen in degenerate
4753 * cases when no memory is available for per-cpu caches.
4754 */
4755 if ((zone->uz_flags & UMA_ZONE_SMR) != 0 && skip == SKIP_NONE)
4756 smr_synchronize(zone->uz_smr);
4757
4758 item_dtor(zone, item, zone->uz_size, udata, skip);
4759
4760 if (skip < SKIP_FINI && zone->uz_fini) {
4761 kasan_mark_item_valid(zone, item);
4762 zone->uz_fini(item, zone->uz_size);
4763 kasan_mark_item_invalid(zone, item);
4764 }
4765
4766 zone->uz_release(zone->uz_arg, &item, 1);
4767
4768 if (skip & SKIP_CNT)
4769 return;
4770
4771 counter_u64_add(zone->uz_frees, 1);
4772
4773 if (zone->uz_max_items > 0)
4774 zone_free_limit(zone, 1);
4775 }
4776
4777 /* See uma.h */
4778 int
uma_zone_set_max(uma_zone_t zone,int nitems)4779 uma_zone_set_max(uma_zone_t zone, int nitems)
4780 {
4781
4782 /*
4783 * If the limit is small, we may need to constrain the maximum per-CPU
4784 * cache size, or disable caching entirely.
4785 */
4786 uma_zone_set_maxcache(zone, nitems);
4787
4788 /*
4789 * XXX This can misbehave if the zone has any allocations with
4790 * no limit and a limit is imposed. There is currently no
4791 * way to clear a limit.
4792 */
4793 ZONE_LOCK(zone);
4794 zone->uz_max_items = nitems;
4795 zone->uz_flags |= UMA_ZFLAG_LIMIT;
4796 zone_update_caches(zone);
4797 /* We may need to wake waiters. */
4798 wakeup(&zone->uz_max_items);
4799 ZONE_UNLOCK(zone);
4800
4801 return (nitems);
4802 }
4803
4804 /* See uma.h */
4805 void
uma_zone_set_maxcache(uma_zone_t zone,int nitems)4806 uma_zone_set_maxcache(uma_zone_t zone, int nitems)
4807 {
4808 int bpcpu, bpdom, bsize, nb;
4809
4810 ZONE_LOCK(zone);
4811
4812 /*
4813 * Compute a lower bound on the number of items that may be cached in
4814 * the zone. Each CPU gets at least two buckets, and for cross-domain
4815 * frees we use an additional bucket per CPU and per domain. Select the
4816 * largest bucket size that does not exceed half of the requested limit,
4817 * with the left over space given to the full bucket cache.
4818 */
4819 bpdom = 0;
4820 bpcpu = 2;
4821 #ifdef NUMA
4822 if ((zone->uz_flags & UMA_ZONE_FIRSTTOUCH) != 0 && vm_ndomains > 1) {
4823 bpcpu++;
4824 bpdom++;
4825 }
4826 #endif
4827 nb = bpcpu * mp_ncpus + bpdom * vm_ndomains;
4828 bsize = nitems / nb / 2;
4829 if (bsize > BUCKET_MAX)
4830 bsize = BUCKET_MAX;
4831 else if (bsize == 0 && nitems / nb > 0)
4832 bsize = 1;
4833 zone->uz_bucket_size_max = zone->uz_bucket_size = bsize;
4834 if (zone->uz_bucket_size_min > zone->uz_bucket_size_max)
4835 zone->uz_bucket_size_min = zone->uz_bucket_size_max;
4836 zone->uz_bucket_max = nitems - nb * bsize;
4837 ZONE_UNLOCK(zone);
4838 }
4839
4840 /* See uma.h */
4841 int
uma_zone_get_max(uma_zone_t zone)4842 uma_zone_get_max(uma_zone_t zone)
4843 {
4844 int nitems;
4845
4846 nitems = atomic_load_64(&zone->uz_max_items);
4847
4848 return (nitems);
4849 }
4850
4851 /* See uma.h */
4852 void
uma_zone_set_warning(uma_zone_t zone,const char * warning)4853 uma_zone_set_warning(uma_zone_t zone, const char *warning)
4854 {
4855
4856 ZONE_ASSERT_COLD(zone);
4857 zone->uz_warning = warning;
4858 }
4859
4860 /* See uma.h */
4861 void
uma_zone_set_maxaction(uma_zone_t zone,uma_maxaction_t maxaction)4862 uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction)
4863 {
4864
4865 ZONE_ASSERT_COLD(zone);
4866 TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone);
4867 }
4868
4869 /* See uma.h */
4870 int
uma_zone_get_cur(uma_zone_t zone)4871 uma_zone_get_cur(uma_zone_t zone)
4872 {
4873 int64_t nitems;
4874 u_int i;
4875
4876 nitems = 0;
4877 if (zone->uz_allocs != EARLY_COUNTER && zone->uz_frees != EARLY_COUNTER)
4878 nitems = counter_u64_fetch(zone->uz_allocs) -
4879 counter_u64_fetch(zone->uz_frees);
4880 CPU_FOREACH(i)
4881 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs) -
4882 atomic_load_64(&zone->uz_cpu[i].uc_frees);
4883
4884 return (nitems < 0 ? 0 : nitems);
4885 }
4886
4887 static uint64_t
uma_zone_get_allocs(uma_zone_t zone)4888 uma_zone_get_allocs(uma_zone_t zone)
4889 {
4890 uint64_t nitems;
4891 u_int i;
4892
4893 nitems = 0;
4894 if (zone->uz_allocs != EARLY_COUNTER)
4895 nitems = counter_u64_fetch(zone->uz_allocs);
4896 CPU_FOREACH(i)
4897 nitems += atomic_load_64(&zone->uz_cpu[i].uc_allocs);
4898
4899 return (nitems);
4900 }
4901
4902 static uint64_t
uma_zone_get_frees(uma_zone_t zone)4903 uma_zone_get_frees(uma_zone_t zone)
4904 {
4905 uint64_t nitems;
4906 u_int i;
4907
4908 nitems = 0;
4909 if (zone->uz_frees != EARLY_COUNTER)
4910 nitems = counter_u64_fetch(zone->uz_frees);
4911 CPU_FOREACH(i)
4912 nitems += atomic_load_64(&zone->uz_cpu[i].uc_frees);
4913
4914 return (nitems);
4915 }
4916
4917 #ifdef INVARIANTS
4918 /* Used only for KEG_ASSERT_COLD(). */
4919 static uint64_t
uma_keg_get_allocs(uma_keg_t keg)4920 uma_keg_get_allocs(uma_keg_t keg)
4921 {
4922 uma_zone_t z;
4923 uint64_t nitems;
4924
4925 nitems = 0;
4926 LIST_FOREACH(z, &keg->uk_zones, uz_link)
4927 nitems += uma_zone_get_allocs(z);
4928
4929 return (nitems);
4930 }
4931 #endif
4932
4933 /* See uma.h */
4934 void
uma_zone_set_init(uma_zone_t zone,uma_init uminit)4935 uma_zone_set_init(uma_zone_t zone, uma_init uminit)
4936 {
4937 uma_keg_t keg;
4938
4939 KEG_GET(zone, keg);
4940 KEG_ASSERT_COLD(keg);
4941 keg->uk_init = uminit;
4942 }
4943
4944 /* See uma.h */
4945 void
uma_zone_set_fini(uma_zone_t zone,uma_fini fini)4946 uma_zone_set_fini(uma_zone_t zone, uma_fini fini)
4947 {
4948 uma_keg_t keg;
4949
4950 KEG_GET(zone, keg);
4951 KEG_ASSERT_COLD(keg);
4952 keg->uk_fini = fini;
4953 }
4954
4955 /* See uma.h */
4956 void
uma_zone_set_zinit(uma_zone_t zone,uma_init zinit)4957 uma_zone_set_zinit(uma_zone_t zone, uma_init zinit)
4958 {
4959
4960 ZONE_ASSERT_COLD(zone);
4961 zone->uz_init = zinit;
4962 }
4963
4964 /* See uma.h */
4965 void
uma_zone_set_zfini(uma_zone_t zone,uma_fini zfini)4966 uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini)
4967 {
4968
4969 ZONE_ASSERT_COLD(zone);
4970 zone->uz_fini = zfini;
4971 }
4972
4973 /* See uma.h */
4974 void
uma_zone_set_freef(uma_zone_t zone,uma_free freef)4975 uma_zone_set_freef(uma_zone_t zone, uma_free freef)
4976 {
4977 uma_keg_t keg;
4978
4979 KEG_GET(zone, keg);
4980 KEG_ASSERT_COLD(keg);
4981 keg->uk_freef = freef;
4982 }
4983
4984 /* See uma.h */
4985 void
uma_zone_set_allocf(uma_zone_t zone,uma_alloc allocf)4986 uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf)
4987 {
4988 uma_keg_t keg;
4989
4990 KEG_GET(zone, keg);
4991 KEG_ASSERT_COLD(keg);
4992 keg->uk_allocf = allocf;
4993 }
4994
4995 /* See uma.h */
4996 void
uma_zone_set_smr(uma_zone_t zone,smr_t smr)4997 uma_zone_set_smr(uma_zone_t zone, smr_t smr)
4998 {
4999
5000 ZONE_ASSERT_COLD(zone);
5001
5002 KASSERT(smr != NULL, ("Got NULL smr"));
5003 KASSERT((zone->uz_flags & UMA_ZONE_SMR) == 0,
5004 ("zone %p (%s) already uses SMR", zone, zone->uz_name));
5005 zone->uz_flags |= UMA_ZONE_SMR;
5006 zone->uz_smr = smr;
5007 zone_update_caches(zone);
5008 }
5009
5010 smr_t
uma_zone_get_smr(uma_zone_t zone)5011 uma_zone_get_smr(uma_zone_t zone)
5012 {
5013
5014 return (zone->uz_smr);
5015 }
5016
5017 /* See uma.h */
5018 void
uma_zone_reserve(uma_zone_t zone,int items)5019 uma_zone_reserve(uma_zone_t zone, int items)
5020 {
5021 uma_keg_t keg;
5022
5023 KEG_GET(zone, keg);
5024 KEG_ASSERT_COLD(keg);
5025 keg->uk_reserve = items;
5026 }
5027
5028 /* See uma.h */
5029 int
uma_zone_reserve_kva(uma_zone_t zone,int count)5030 uma_zone_reserve_kva(uma_zone_t zone, int count)
5031 {
5032 uma_keg_t keg;
5033 vm_offset_t kva;
5034 u_int pages;
5035
5036 KEG_GET(zone, keg);
5037 KEG_ASSERT_COLD(keg);
5038 ZONE_ASSERT_COLD(zone);
5039
5040 pages = howmany(count, keg->uk_ipers) * keg->uk_ppera;
5041
5042 #ifdef UMA_MD_SMALL_ALLOC
5043 if (keg->uk_ppera > 1) {
5044 #else
5045 if (1) {
5046 #endif
5047 kva = kva_alloc((vm_size_t)pages * PAGE_SIZE);
5048 if (kva == 0)
5049 return (0);
5050 } else
5051 kva = 0;
5052
5053 MPASS(keg->uk_kva == 0);
5054 keg->uk_kva = kva;
5055 keg->uk_offset = 0;
5056 zone->uz_max_items = pages * keg->uk_ipers;
5057 #ifdef UMA_MD_SMALL_ALLOC
5058 keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc;
5059 #else
5060 keg->uk_allocf = noobj_alloc;
5061 #endif
5062 keg->uk_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5063 zone->uz_flags |= UMA_ZFLAG_LIMIT | UMA_ZONE_NOFREE;
5064 zone_update_caches(zone);
5065
5066 return (1);
5067 }
5068
5069 /* See uma.h */
5070 void
5071 uma_prealloc(uma_zone_t zone, int items)
5072 {
5073 struct vm_domainset_iter di;
5074 uma_domain_t dom;
5075 uma_slab_t slab;
5076 uma_keg_t keg;
5077 int aflags, domain, slabs;
5078
5079 KEG_GET(zone, keg);
5080 slabs = howmany(items, keg->uk_ipers);
5081 while (slabs-- > 0) {
5082 aflags = M_NOWAIT;
5083 vm_domainset_iter_policy_ref_init(&di, &keg->uk_dr, &domain,
5084 &aflags);
5085 for (;;) {
5086 slab = keg_alloc_slab(keg, zone, domain, M_WAITOK,
5087 aflags);
5088 if (slab != NULL) {
5089 dom = &keg->uk_domain[slab->us_domain];
5090 /*
5091 * keg_alloc_slab() always returns a slab on the
5092 * partial list.
5093 */
5094 LIST_REMOVE(slab, us_link);
5095 LIST_INSERT_HEAD(&dom->ud_free_slab, slab,
5096 us_link);
5097 dom->ud_free_slabs++;
5098 KEG_UNLOCK(keg, slab->us_domain);
5099 break;
5100 }
5101 if (vm_domainset_iter_policy(&di, &domain) != 0)
5102 vm_wait_doms(&keg->uk_dr.dr_policy->ds_mask, 0);
5103 }
5104 }
5105 }
5106
5107 /*
5108 * Returns a snapshot of memory consumption in bytes.
5109 */
5110 size_t
5111 uma_zone_memory(uma_zone_t zone)
5112 {
5113 size_t sz;
5114 int i;
5115
5116 sz = 0;
5117 if (zone->uz_flags & UMA_ZFLAG_CACHE) {
5118 for (i = 0; i < vm_ndomains; i++)
5119 sz += ZDOM_GET(zone, i)->uzd_nitems;
5120 return (sz * zone->uz_size);
5121 }
5122 for (i = 0; i < vm_ndomains; i++)
5123 sz += zone->uz_keg->uk_domain[i].ud_pages;
5124
5125 return (sz * PAGE_SIZE);
5126 }
5127
5128 struct uma_reclaim_args {
5129 int domain;
5130 int req;
5131 };
5132
5133 static void
5134 uma_reclaim_domain_cb(uma_zone_t zone, void *arg)
5135 {
5136 struct uma_reclaim_args *args;
5137
5138 args = arg;
5139 if ((zone->uz_flags & UMA_ZONE_UNMANAGED) == 0)
5140 uma_zone_reclaim_domain(zone, args->req, args->domain);
5141 }
5142
5143 /* See uma.h */
5144 void
5145 uma_reclaim(int req)
5146 {
5147 uma_reclaim_domain(req, UMA_ANYDOMAIN);
5148 }
5149
5150 void
5151 uma_reclaim_domain(int req, int domain)
5152 {
5153 struct uma_reclaim_args args;
5154
5155 bucket_enable();
5156
5157 args.domain = domain;
5158 args.req = req;
5159
5160 sx_slock(&uma_reclaim_lock);
5161 switch (req) {
5162 case UMA_RECLAIM_TRIM:
5163 case UMA_RECLAIM_DRAIN:
5164 zone_foreach(uma_reclaim_domain_cb, &args);
5165 break;
5166 case UMA_RECLAIM_DRAIN_CPU:
5167 zone_foreach(uma_reclaim_domain_cb, &args);
5168 pcpu_cache_drain_safe(NULL);
5169 zone_foreach(uma_reclaim_domain_cb, &args);
5170 break;
5171 default:
5172 panic("unhandled reclamation request %d", req);
5173 }
5174
5175 /*
5176 * Some slabs may have been freed but this zone will be visited early
5177 * we visit again so that we can free pages that are empty once other
5178 * zones are drained. We have to do the same for buckets.
5179 */
5180 uma_zone_reclaim_domain(slabzones[0], UMA_RECLAIM_DRAIN, domain);
5181 uma_zone_reclaim_domain(slabzones[1], UMA_RECLAIM_DRAIN, domain);
5182 bucket_zone_drain(domain);
5183 sx_sunlock(&uma_reclaim_lock);
5184 }
5185
5186 static volatile int uma_reclaim_needed;
5187
5188 void
5189 uma_reclaim_wakeup(void)
5190 {
5191
5192 if (atomic_fetchadd_int(&uma_reclaim_needed, 1) == 0)
5193 wakeup(uma_reclaim);
5194 }
5195
5196 void
5197 uma_reclaim_worker(void *arg __unused)
5198 {
5199
5200 for (;;) {
5201 sx_xlock(&uma_reclaim_lock);
5202 while (atomic_load_int(&uma_reclaim_needed) == 0)
5203 sx_sleep(uma_reclaim, &uma_reclaim_lock, PVM, "umarcl",
5204 hz);
5205 sx_xunlock(&uma_reclaim_lock);
5206 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_KMEM);
5207 uma_reclaim(UMA_RECLAIM_DRAIN_CPU);
5208 atomic_store_int(&uma_reclaim_needed, 0);
5209 /* Don't fire more than once per-second. */
5210 pause("umarclslp", hz);
5211 }
5212 }
5213
5214 /* See uma.h */
5215 void
5216 uma_zone_reclaim(uma_zone_t zone, int req)
5217 {
5218 uma_zone_reclaim_domain(zone, req, UMA_ANYDOMAIN);
5219 }
5220
5221 void
5222 uma_zone_reclaim_domain(uma_zone_t zone, int req, int domain)
5223 {
5224 switch (req) {
5225 case UMA_RECLAIM_TRIM:
5226 zone_reclaim(zone, domain, M_NOWAIT, false);
5227 break;
5228 case UMA_RECLAIM_DRAIN:
5229 zone_reclaim(zone, domain, M_NOWAIT, true);
5230 break;
5231 case UMA_RECLAIM_DRAIN_CPU:
5232 pcpu_cache_drain_safe(zone);
5233 zone_reclaim(zone, domain, M_NOWAIT, true);
5234 break;
5235 default:
5236 panic("unhandled reclamation request %d", req);
5237 }
5238 }
5239
5240 /* See uma.h */
5241 int
5242 uma_zone_exhausted(uma_zone_t zone)
5243 {
5244
5245 return (atomic_load_32(&zone->uz_sleepers) > 0);
5246 }
5247
5248 unsigned long
5249 uma_limit(void)
5250 {
5251
5252 return (uma_kmem_limit);
5253 }
5254
5255 void
5256 uma_set_limit(unsigned long limit)
5257 {
5258
5259 uma_kmem_limit = limit;
5260 }
5261
5262 unsigned long
5263 uma_size(void)
5264 {
5265
5266 return (atomic_load_long(&uma_kmem_total));
5267 }
5268
5269 long
5270 uma_avail(void)
5271 {
5272
5273 return (uma_kmem_limit - uma_size());
5274 }
5275
5276 #ifdef DDB
5277 /*
5278 * Generate statistics across both the zone and its per-cpu cache's. Return
5279 * desired statistics if the pointer is non-NULL for that statistic.
5280 *
5281 * Note: does not update the zone statistics, as it can't safely clear the
5282 * per-CPU cache statistic.
5283 *
5284 */
5285 static void
5286 uma_zone_sumstat(uma_zone_t z, long *cachefreep, uint64_t *allocsp,
5287 uint64_t *freesp, uint64_t *sleepsp, uint64_t *xdomainp)
5288 {
5289 uma_cache_t cache;
5290 uint64_t allocs, frees, sleeps, xdomain;
5291 int cachefree, cpu;
5292
5293 allocs = frees = sleeps = xdomain = 0;
5294 cachefree = 0;
5295 CPU_FOREACH(cpu) {
5296 cache = &z->uz_cpu[cpu];
5297 cachefree += cache->uc_allocbucket.ucb_cnt;
5298 cachefree += cache->uc_freebucket.ucb_cnt;
5299 xdomain += cache->uc_crossbucket.ucb_cnt;
5300 cachefree += cache->uc_crossbucket.ucb_cnt;
5301 allocs += cache->uc_allocs;
5302 frees += cache->uc_frees;
5303 }
5304 allocs += counter_u64_fetch(z->uz_allocs);
5305 frees += counter_u64_fetch(z->uz_frees);
5306 xdomain += counter_u64_fetch(z->uz_xdomain);
5307 sleeps += z->uz_sleeps;
5308 if (cachefreep != NULL)
5309 *cachefreep = cachefree;
5310 if (allocsp != NULL)
5311 *allocsp = allocs;
5312 if (freesp != NULL)
5313 *freesp = frees;
5314 if (sleepsp != NULL)
5315 *sleepsp = sleeps;
5316 if (xdomainp != NULL)
5317 *xdomainp = xdomain;
5318 }
5319 #endif /* DDB */
5320
5321 static int
5322 sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS)
5323 {
5324 uma_keg_t kz;
5325 uma_zone_t z;
5326 int count;
5327
5328 count = 0;
5329 rw_rlock(&uma_rwlock);
5330 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5331 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5332 count++;
5333 }
5334 LIST_FOREACH(z, &uma_cachezones, uz_link)
5335 count++;
5336
5337 rw_runlock(&uma_rwlock);
5338 return (sysctl_handle_int(oidp, &count, 0, req));
5339 }
5340
5341 static void
5342 uma_vm_zone_stats(struct uma_type_header *uth, uma_zone_t z, struct sbuf *sbuf,
5343 struct uma_percpu_stat *ups, bool internal)
5344 {
5345 uma_zone_domain_t zdom;
5346 uma_cache_t cache;
5347 int i;
5348
5349 for (i = 0; i < vm_ndomains; i++) {
5350 zdom = ZDOM_GET(z, i);
5351 uth->uth_zone_free += zdom->uzd_nitems;
5352 }
5353 uth->uth_allocs = counter_u64_fetch(z->uz_allocs);
5354 uth->uth_frees = counter_u64_fetch(z->uz_frees);
5355 uth->uth_fails = counter_u64_fetch(z->uz_fails);
5356 uth->uth_xdomain = counter_u64_fetch(z->uz_xdomain);
5357 uth->uth_sleeps = z->uz_sleeps;
5358
5359 for (i = 0; i < mp_maxid + 1; i++) {
5360 bzero(&ups[i], sizeof(*ups));
5361 if (internal || CPU_ABSENT(i))
5362 continue;
5363 cache = &z->uz_cpu[i];
5364 ups[i].ups_cache_free += cache->uc_allocbucket.ucb_cnt;
5365 ups[i].ups_cache_free += cache->uc_freebucket.ucb_cnt;
5366 ups[i].ups_cache_free += cache->uc_crossbucket.ucb_cnt;
5367 ups[i].ups_allocs = cache->uc_allocs;
5368 ups[i].ups_frees = cache->uc_frees;
5369 }
5370 }
5371
5372 static int
5373 sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS)
5374 {
5375 struct uma_stream_header ush;
5376 struct uma_type_header uth;
5377 struct uma_percpu_stat *ups;
5378 struct sbuf sbuf;
5379 uma_keg_t kz;
5380 uma_zone_t z;
5381 uint64_t items;
5382 uint32_t kfree, pages;
5383 int count, error, i;
5384
5385 error = sysctl_wire_old_buffer(req, 0);
5386 if (error != 0)
5387 return (error);
5388 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
5389 sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
5390 ups = malloc((mp_maxid + 1) * sizeof(*ups), M_TEMP, M_WAITOK);
5391
5392 count = 0;
5393 rw_rlock(&uma_rwlock);
5394 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5395 LIST_FOREACH(z, &kz->uk_zones, uz_link)
5396 count++;
5397 }
5398
5399 LIST_FOREACH(z, &uma_cachezones, uz_link)
5400 count++;
5401
5402 /*
5403 * Insert stream header.
5404 */
5405 bzero(&ush, sizeof(ush));
5406 ush.ush_version = UMA_STREAM_VERSION;
5407 ush.ush_maxcpus = (mp_maxid + 1);
5408 ush.ush_count = count;
5409 (void)sbuf_bcat(&sbuf, &ush, sizeof(ush));
5410
5411 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5412 kfree = pages = 0;
5413 for (i = 0; i < vm_ndomains; i++) {
5414 kfree += kz->uk_domain[i].ud_free_items;
5415 pages += kz->uk_domain[i].ud_pages;
5416 }
5417 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5418 bzero(&uth, sizeof(uth));
5419 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5420 uth.uth_align = kz->uk_align;
5421 uth.uth_size = kz->uk_size;
5422 uth.uth_rsize = kz->uk_rsize;
5423 if (z->uz_max_items > 0) {
5424 items = UZ_ITEMS_COUNT(z->uz_items);
5425 uth.uth_pages = (items / kz->uk_ipers) *
5426 kz->uk_ppera;
5427 } else
5428 uth.uth_pages = pages;
5429 uth.uth_maxpages = (z->uz_max_items / kz->uk_ipers) *
5430 kz->uk_ppera;
5431 uth.uth_limit = z->uz_max_items;
5432 uth.uth_keg_free = kfree;
5433
5434 /*
5435 * A zone is secondary is it is not the first entry
5436 * on the keg's zone list.
5437 */
5438 if ((z->uz_flags & UMA_ZONE_SECONDARY) &&
5439 (LIST_FIRST(&kz->uk_zones) != z))
5440 uth.uth_zone_flags = UTH_ZONE_SECONDARY;
5441 uma_vm_zone_stats(&uth, z, &sbuf, ups,
5442 kz->uk_flags & UMA_ZFLAG_INTERNAL);
5443 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5444 for (i = 0; i < mp_maxid + 1; i++)
5445 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5446 }
5447 }
5448 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5449 bzero(&uth, sizeof(uth));
5450 strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME);
5451 uth.uth_size = z->uz_size;
5452 uma_vm_zone_stats(&uth, z, &sbuf, ups, false);
5453 (void)sbuf_bcat(&sbuf, &uth, sizeof(uth));
5454 for (i = 0; i < mp_maxid + 1; i++)
5455 (void)sbuf_bcat(&sbuf, &ups[i], sizeof(ups[i]));
5456 }
5457
5458 rw_runlock(&uma_rwlock);
5459 error = sbuf_finish(&sbuf);
5460 sbuf_delete(&sbuf);
5461 free(ups, M_TEMP);
5462 return (error);
5463 }
5464
5465 int
5466 sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS)
5467 {
5468 uma_zone_t zone = *(uma_zone_t *)arg1;
5469 int error, max;
5470
5471 max = uma_zone_get_max(zone);
5472 error = sysctl_handle_int(oidp, &max, 0, req);
5473 if (error || !req->newptr)
5474 return (error);
5475
5476 uma_zone_set_max(zone, max);
5477
5478 return (0);
5479 }
5480
5481 int
5482 sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS)
5483 {
5484 uma_zone_t zone;
5485 int cur;
5486
5487 /*
5488 * Some callers want to add sysctls for global zones that
5489 * may not yet exist so they pass a pointer to a pointer.
5490 */
5491 if (arg2 == 0)
5492 zone = *(uma_zone_t *)arg1;
5493 else
5494 zone = arg1;
5495 cur = uma_zone_get_cur(zone);
5496 return (sysctl_handle_int(oidp, &cur, 0, req));
5497 }
5498
5499 static int
5500 sysctl_handle_uma_zone_allocs(SYSCTL_HANDLER_ARGS)
5501 {
5502 uma_zone_t zone = arg1;
5503 uint64_t cur;
5504
5505 cur = uma_zone_get_allocs(zone);
5506 return (sysctl_handle_64(oidp, &cur, 0, req));
5507 }
5508
5509 static int
5510 sysctl_handle_uma_zone_frees(SYSCTL_HANDLER_ARGS)
5511 {
5512 uma_zone_t zone = arg1;
5513 uint64_t cur;
5514
5515 cur = uma_zone_get_frees(zone);
5516 return (sysctl_handle_64(oidp, &cur, 0, req));
5517 }
5518
5519 static int
5520 sysctl_handle_uma_zone_flags(SYSCTL_HANDLER_ARGS)
5521 {
5522 struct sbuf sbuf;
5523 uma_zone_t zone = arg1;
5524 int error;
5525
5526 sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
5527 if (zone->uz_flags != 0)
5528 sbuf_printf(&sbuf, "0x%b", zone->uz_flags, PRINT_UMA_ZFLAGS);
5529 else
5530 sbuf_printf(&sbuf, "0");
5531 error = sbuf_finish(&sbuf);
5532 sbuf_delete(&sbuf);
5533
5534 return (error);
5535 }
5536
5537 static int
5538 sysctl_handle_uma_slab_efficiency(SYSCTL_HANDLER_ARGS)
5539 {
5540 uma_keg_t keg = arg1;
5541 int avail, effpct, total;
5542
5543 total = keg->uk_ppera * PAGE_SIZE;
5544 if ((keg->uk_flags & UMA_ZFLAG_OFFPAGE) != 0)
5545 total += slabzone(keg->uk_ipers)->uz_keg->uk_rsize;
5546 /*
5547 * We consider the client's requested size and alignment here, not the
5548 * real size determination uk_rsize, because we also adjust the real
5549 * size for internal implementation reasons (max bitset size).
5550 */
5551 avail = keg->uk_ipers * roundup2(keg->uk_size, keg->uk_align + 1);
5552 if ((keg->uk_flags & UMA_ZONE_PCPU) != 0)
5553 avail *= mp_maxid + 1;
5554 effpct = 100 * avail / total;
5555 return (sysctl_handle_int(oidp, &effpct, 0, req));
5556 }
5557
5558 static int
5559 sysctl_handle_uma_zone_items(SYSCTL_HANDLER_ARGS)
5560 {
5561 uma_zone_t zone = arg1;
5562 uint64_t cur;
5563
5564 cur = UZ_ITEMS_COUNT(atomic_load_64(&zone->uz_items));
5565 return (sysctl_handle_64(oidp, &cur, 0, req));
5566 }
5567
5568 #ifdef INVARIANTS
5569 static uma_slab_t
5570 uma_dbg_getslab(uma_zone_t zone, void *item)
5571 {
5572 uma_slab_t slab;
5573 uma_keg_t keg;
5574 uint8_t *mem;
5575
5576 /*
5577 * It is safe to return the slab here even though the
5578 * zone is unlocked because the item's allocation state
5579 * essentially holds a reference.
5580 */
5581 mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK));
5582 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5583 return (NULL);
5584 if (zone->uz_flags & UMA_ZFLAG_VTOSLAB)
5585 return (vtoslab((vm_offset_t)mem));
5586 keg = zone->uz_keg;
5587 if ((keg->uk_flags & UMA_ZFLAG_HASH) == 0)
5588 return ((uma_slab_t)(mem + keg->uk_pgoff));
5589 KEG_LOCK(keg, 0);
5590 slab = hash_sfind(&keg->uk_hash, mem);
5591 KEG_UNLOCK(keg, 0);
5592
5593 return (slab);
5594 }
5595
5596 static bool
5597 uma_dbg_zskip(uma_zone_t zone, void *mem)
5598 {
5599
5600 if ((zone->uz_flags & UMA_ZFLAG_CACHE) != 0)
5601 return (true);
5602
5603 return (uma_dbg_kskip(zone->uz_keg, mem));
5604 }
5605
5606 static bool
5607 uma_dbg_kskip(uma_keg_t keg, void *mem)
5608 {
5609 uintptr_t idx;
5610
5611 if (dbg_divisor == 0)
5612 return (true);
5613
5614 if (dbg_divisor == 1)
5615 return (false);
5616
5617 idx = (uintptr_t)mem >> PAGE_SHIFT;
5618 if (keg->uk_ipers > 1) {
5619 idx *= keg->uk_ipers;
5620 idx += ((uintptr_t)mem & PAGE_MASK) / keg->uk_rsize;
5621 }
5622
5623 if ((idx / dbg_divisor) * dbg_divisor != idx) {
5624 counter_u64_add(uma_skip_cnt, 1);
5625 return (true);
5626 }
5627 counter_u64_add(uma_dbg_cnt, 1);
5628
5629 return (false);
5630 }
5631
5632 /*
5633 * Set up the slab's freei data such that uma_dbg_free can function.
5634 *
5635 */
5636 static void
5637 uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item)
5638 {
5639 uma_keg_t keg;
5640 int freei;
5641
5642 if (slab == NULL) {
5643 slab = uma_dbg_getslab(zone, item);
5644 if (slab == NULL)
5645 panic("uma: item %p did not belong to zone %s",
5646 item, zone->uz_name);
5647 }
5648 keg = zone->uz_keg;
5649 freei = slab_item_index(slab, keg, item);
5650
5651 if (BIT_TEST_SET_ATOMIC(keg->uk_ipers, freei,
5652 slab_dbg_bits(slab, keg)))
5653 panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)",
5654 item, zone, zone->uz_name, slab, freei);
5655 }
5656
5657 /*
5658 * Verifies freed addresses. Checks for alignment, valid slab membership
5659 * and duplicate frees.
5660 *
5661 */
5662 static void
5663 uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item)
5664 {
5665 uma_keg_t keg;
5666 int freei;
5667
5668 if (slab == NULL) {
5669 slab = uma_dbg_getslab(zone, item);
5670 if (slab == NULL)
5671 panic("uma: Freed item %p did not belong to zone %s",
5672 item, zone->uz_name);
5673 }
5674 keg = zone->uz_keg;
5675 freei = slab_item_index(slab, keg, item);
5676
5677 if (freei >= keg->uk_ipers)
5678 panic("Invalid free of %p from zone %p(%s) slab %p(%d)",
5679 item, zone, zone->uz_name, slab, freei);
5680
5681 if (slab_item(slab, keg, freei) != item)
5682 panic("Unaligned free of %p from zone %p(%s) slab %p(%d)",
5683 item, zone, zone->uz_name, slab, freei);
5684
5685 if (!BIT_TEST_CLR_ATOMIC(keg->uk_ipers, freei,
5686 slab_dbg_bits(slab, keg)))
5687 panic("Duplicate free of %p from zone %p(%s) slab %p(%d)",
5688 item, zone, zone->uz_name, slab, freei);
5689 }
5690 #endif /* INVARIANTS */
5691
5692 #ifdef DDB
5693 static int64_t
5694 get_uma_stats(uma_keg_t kz, uma_zone_t z, uint64_t *allocs, uint64_t *used,
5695 uint64_t *sleeps, long *cachefree, uint64_t *xdomain)
5696 {
5697 uint64_t frees;
5698 int i;
5699
5700 if (kz->uk_flags & UMA_ZFLAG_INTERNAL) {
5701 *allocs = counter_u64_fetch(z->uz_allocs);
5702 frees = counter_u64_fetch(z->uz_frees);
5703 *sleeps = z->uz_sleeps;
5704 *cachefree = 0;
5705 *xdomain = 0;
5706 } else
5707 uma_zone_sumstat(z, cachefree, allocs, &frees, sleeps,
5708 xdomain);
5709 for (i = 0; i < vm_ndomains; i++) {
5710 *cachefree += ZDOM_GET(z, i)->uzd_nitems;
5711 if (!((z->uz_flags & UMA_ZONE_SECONDARY) &&
5712 (LIST_FIRST(&kz->uk_zones) != z)))
5713 *cachefree += kz->uk_domain[i].ud_free_items;
5714 }
5715 *used = *allocs - frees;
5716 return (((int64_t)*used + *cachefree) * kz->uk_size);
5717 }
5718
5719 DB_SHOW_COMMAND(uma, db_show_uma)
5720 {
5721 const char *fmt_hdr, *fmt_entry;
5722 uma_keg_t kz;
5723 uma_zone_t z;
5724 uint64_t allocs, used, sleeps, xdomain;
5725 long cachefree;
5726 /* variables for sorting */
5727 uma_keg_t cur_keg;
5728 uma_zone_t cur_zone, last_zone;
5729 int64_t cur_size, last_size, size;
5730 int ties;
5731
5732 /* /i option produces machine-parseable CSV output */
5733 if (modif[0] == 'i') {
5734 fmt_hdr = "%s,%s,%s,%s,%s,%s,%s,%s,%s\n";
5735 fmt_entry = "\"%s\",%ju,%jd,%ld,%ju,%ju,%u,%jd,%ju\n";
5736 } else {
5737 fmt_hdr = "%18s %6s %7s %7s %11s %7s %7s %10s %8s\n";
5738 fmt_entry = "%18s %6ju %7jd %7ld %11ju %7ju %7u %10jd %8ju\n";
5739 }
5740
5741 db_printf(fmt_hdr, "Zone", "Size", "Used", "Free", "Requests",
5742 "Sleeps", "Bucket", "Total Mem", "XFree");
5743
5744 /* Sort the zones with largest size first. */
5745 last_zone = NULL;
5746 last_size = INT64_MAX;
5747 for (;;) {
5748 cur_zone = NULL;
5749 cur_size = -1;
5750 ties = 0;
5751 LIST_FOREACH(kz, &uma_kegs, uk_link) {
5752 LIST_FOREACH(z, &kz->uk_zones, uz_link) {
5753 /*
5754 * In the case of size ties, print out zones
5755 * in the order they are encountered. That is,
5756 * when we encounter the most recently output
5757 * zone, we have already printed all preceding
5758 * ties, and we must print all following ties.
5759 */
5760 if (z == last_zone) {
5761 ties = 1;
5762 continue;
5763 }
5764 size = get_uma_stats(kz, z, &allocs, &used,
5765 &sleeps, &cachefree, &xdomain);
5766 if (size > cur_size && size < last_size + ties)
5767 {
5768 cur_size = size;
5769 cur_zone = z;
5770 cur_keg = kz;
5771 }
5772 }
5773 }
5774 if (cur_zone == NULL)
5775 break;
5776
5777 size = get_uma_stats(cur_keg, cur_zone, &allocs, &used,
5778 &sleeps, &cachefree, &xdomain);
5779 db_printf(fmt_entry, cur_zone->uz_name,
5780 (uintmax_t)cur_keg->uk_size, (intmax_t)used, cachefree,
5781 (uintmax_t)allocs, (uintmax_t)sleeps,
5782 (unsigned)cur_zone->uz_bucket_size, (intmax_t)size,
5783 xdomain);
5784
5785 if (db_pager_quit)
5786 return;
5787 last_zone = cur_zone;
5788 last_size = cur_size;
5789 }
5790 }
5791
5792 DB_SHOW_COMMAND(umacache, db_show_umacache)
5793 {
5794 uma_zone_t z;
5795 uint64_t allocs, frees;
5796 long cachefree;
5797 int i;
5798
5799 db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free",
5800 "Requests", "Bucket");
5801 LIST_FOREACH(z, &uma_cachezones, uz_link) {
5802 uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL, NULL);
5803 for (i = 0; i < vm_ndomains; i++)
5804 cachefree += ZDOM_GET(z, i)->uzd_nitems;
5805 db_printf("%18s %8ju %8jd %8ld %12ju %8u\n",
5806 z->uz_name, (uintmax_t)z->uz_size,
5807 (intmax_t)(allocs - frees), cachefree,
5808 (uintmax_t)allocs, z->uz_bucket_size);
5809 if (db_pager_quit)
5810 return;
5811 }
5812 }
5813 #endif /* DDB */
5814