1 | /** |
---|
2 | * @file |
---|
3 | * |
---|
4 | * @ingroup rtems_bsd_rtems |
---|
5 | * |
---|
6 | * @brief TODO. |
---|
7 | */ |
---|
8 | |
---|
9 | /* |
---|
10 | * Copyright (c) 2009, 2010 embedded brains GmbH. All rights reserved. |
---|
11 | * |
---|
12 | * embedded brains GmbH |
---|
13 | * Obere Lagerstr. 30 |
---|
14 | * 82178 Puchheim |
---|
15 | * Germany |
---|
16 | * <rtems@embedded-brains.de> |
---|
17 | * |
---|
18 | * The license and distribution terms for this file may be |
---|
19 | * found in the file LICENSE in this distribution or at |
---|
20 | * http://www.rtems.com/license/LICENSE. |
---|
21 | */ |
---|
22 | |
---|
23 | #include <rtems/freebsd/machine/rtems-bsd-config.h> |
---|
24 | |
---|
25 | #include <rtems/freebsd/sys/param.h> |
---|
26 | #include <rtems/freebsd/sys/types.h> |
---|
27 | #include <rtems/freebsd/sys/systm.h> |
---|
28 | #include <rtems/freebsd/sys/malloc.h> |
---|
29 | #include <rtems/freebsd/sys/kernel.h> |
---|
30 | #include <rtems/freebsd/sys/lock.h> |
---|
31 | #include <rtems/freebsd/sys/mutex.h> |
---|
32 | #include <rtems/freebsd/sys/ktr.h> |
---|
33 | #include <rtems/freebsd/vm/uma.h> |
---|
34 | #include <rtems/freebsd/vm/uma_int.h> |
---|
35 | #include <rtems/freebsd/vm/uma_dbg.h> |
---|
36 | |
---|
37 | /* |
---|
38 | * This is the zone and keg from which all zones are spawned. The idea is that |
---|
39 | * even the zone & keg heads are allocated from the allocator, so we use the |
---|
40 | * bss section to bootstrap us. |
---|
41 | */ |
---|
42 | static struct uma_keg masterkeg; |
---|
43 | static struct uma_zone masterzone_k; |
---|
44 | static struct uma_zone masterzone_z; |
---|
45 | static uma_zone_t kegs = &masterzone_k; |
---|
46 | static uma_zone_t zones = &masterzone_z; |
---|
47 | |
---|
48 | /* This is the zone from which all of uma_slab_t's are allocated. */ |
---|
49 | static uma_zone_t slabzone; |
---|
50 | static uma_zone_t slabrefzone; /* With refcounters (for UMA_ZONE_REFCNT) */ |
---|
51 | |
---|
52 | static u_int mp_maxid = 0; /* simulate 1 CPU. This should really come from RTEMS SMP. AT this time, RTEMS SMP is not functional */ |
---|
53 | #define CPU_ABSENT(x_cpu) 0 /* force all cpus to be present. This should really come from RTEMS SMP. */ |
---|
54 | #define CPU_FOREACH(i) \ |
---|
55 | for ((i) = 0; (i) <= mp_maxid; (i)++) \ |
---|
56 | if (!CPU_ABSENT((i))) |
---|
57 | |
---|
58 | /* |
---|
59 | * The initial hash tables come out of this zone so they can be allocated |
---|
60 | * prior to malloc coming up. |
---|
61 | */ |
---|
62 | static uma_zone_t hashzone; |
---|
63 | |
---|
64 | /* The boot-time adjusted value for cache line alignment. */ |
---|
65 | static int uma_align_cache = 64 - 1; |
---|
66 | |
---|
67 | static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); |
---|
68 | |
---|
69 | /* |
---|
70 | * Are we allowed to allocate buckets? |
---|
71 | */ |
---|
72 | static int bucketdisable = 1; |
---|
73 | |
---|
74 | /* Linked list of all kegs in the system */ |
---|
75 | static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); |
---|
76 | |
---|
77 | /* This mutex protects the keg list */ |
---|
78 | static struct mtx uma_mtx; |
---|
79 | |
---|
80 | /* Linked list of boot time pages */ |
---|
81 | static LIST_HEAD(,uma_slab) uma_boot_pages = |
---|
82 | LIST_HEAD_INITIALIZER(uma_boot_pages); |
---|
83 | |
---|
84 | /* This mutex protects the boot time pages list */ |
---|
85 | static struct mtx uma_boot_pages_mtx; |
---|
86 | |
---|
87 | /* Is the VM done starting up? */ |
---|
88 | static int booted = 0; |
---|
89 | |
---|
90 | /* Maximum number of allowed items-per-slab if the slab header is OFFPAGE */ |
---|
91 | static u_int uma_max_ipers; |
---|
92 | static u_int uma_max_ipers_ref; |
---|
93 | |
---|
94 | /* |
---|
95 | * This is the handle used to schedule events that need to happen |
---|
96 | * outside of the allocation fast path. |
---|
97 | */ |
---|
98 | static struct callout uma_callout; |
---|
99 | #define UMA_TIMEOUT 20 /* Seconds for callout interval. */ |
---|
100 | |
---|
101 | /* |
---|
102 | * This structure is passed as the zone ctor arg so that I don't have to create |
---|
103 | * a special allocation function just for zones. |
---|
104 | */ |
---|
105 | struct uma_zctor_args { |
---|
106 | char *name; |
---|
107 | size_t size; |
---|
108 | uma_ctor ctor; |
---|
109 | uma_dtor dtor; |
---|
110 | uma_init uminit; |
---|
111 | uma_fini fini; |
---|
112 | uma_keg_t keg; |
---|
113 | int align; |
---|
114 | u_int32_t flags; |
---|
115 | }; |
---|
116 | |
---|
117 | struct uma_kctor_args { |
---|
118 | uma_zone_t zone; |
---|
119 | size_t size; |
---|
120 | uma_init uminit; |
---|
121 | uma_fini fini; |
---|
122 | int align; |
---|
123 | u_int32_t flags; |
---|
124 | }; |
---|
125 | |
---|
126 | struct uma_bucket_zone { |
---|
127 | uma_zone_t ubz_zone; |
---|
128 | char *ubz_name; |
---|
129 | int ubz_entries; |
---|
130 | }; |
---|
131 | |
---|
132 | #define BUCKET_MAX 128 |
---|
133 | |
---|
134 | struct uma_bucket_zone bucket_zones[] = { |
---|
135 | { NULL, "16 Bucket", 16 }, |
---|
136 | { NULL, "32 Bucket", 32 }, |
---|
137 | { NULL, "64 Bucket", 64 }, |
---|
138 | { NULL, "128 Bucket", 128 }, |
---|
139 | { NULL, NULL, 0} |
---|
140 | }; |
---|
141 | |
---|
142 | #define BUCKET_SHIFT 4 |
---|
143 | #define BUCKET_ZONES ((BUCKET_MAX >> BUCKET_SHIFT) + 1) |
---|
144 | |
---|
145 | /* |
---|
146 | * bucket_size[] maps requested bucket sizes to zones that allocate a bucket |
---|
147 | * of approximately the right size. |
---|
148 | */ |
---|
149 | static uint8_t bucket_size[BUCKET_ZONES]; |
---|
150 | |
---|
151 | /* |
---|
152 | * Flags and enumerations to be passed to internal functions. |
---|
153 | */ |
---|
154 | enum zfreeskip { SKIP_NONE, SKIP_DTOR, SKIP_FINI }; |
---|
155 | |
---|
156 | #define ZFREE_STATFAIL 0x00000001 /* Update zone failure statistic. */ |
---|
157 | #define ZFREE_STATFREE 0x00000002 /* Update zone free statistic. */ |
---|
158 | |
---|
159 | /* Prototypes.. */ |
---|
160 | |
---|
161 | static void *page_alloc(uma_zone_t, int, u_int8_t *, int); |
---|
162 | static void *startup_alloc(uma_zone_t, int, u_int8_t *, int); |
---|
163 | static void page_free(void *, int, u_int8_t); |
---|
164 | static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int); |
---|
165 | static void cache_drain(uma_zone_t); |
---|
166 | static void bucket_drain(uma_zone_t, uma_bucket_t); |
---|
167 | static void bucket_cache_drain(uma_zone_t zone); |
---|
168 | static int keg_ctor(void *, int, void *, int); |
---|
169 | static void keg_dtor(void *, int, void *); |
---|
170 | static int zone_ctor(void *, int, void *, int); |
---|
171 | static void zone_dtor(void *, int, void *); |
---|
172 | static int zero_init(void *, int, int); |
---|
173 | static void keg_small_init(uma_keg_t keg); |
---|
174 | static void keg_large_init(uma_keg_t keg); |
---|
175 | static void zone_foreach(void (*zfunc)(uma_zone_t)); |
---|
176 | static void zone_timeout(uma_zone_t zone); |
---|
177 | static int hash_alloc(struct uma_hash *); |
---|
178 | static int hash_expand(struct uma_hash *, struct uma_hash *); |
---|
179 | static void hash_free(struct uma_hash *hash); |
---|
180 | static void *zone_alloc_item(uma_zone_t, void *, int); |
---|
181 | static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip, |
---|
182 | int); |
---|
183 | static void bucket_init(void); |
---|
184 | static uma_bucket_t bucket_alloc(int, int); |
---|
185 | static void bucket_free(uma_bucket_t); |
---|
186 | static void bucket_zone_drain(void); |
---|
187 | static int zone_alloc_bucket(uma_zone_t zone, int flags); |
---|
188 | static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags); |
---|
189 | static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags); |
---|
190 | static void *slab_alloc_item(uma_zone_t zone, uma_slab_t slab); |
---|
191 | static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, |
---|
192 | uma_fini fini, int align, u_int32_t flags); |
---|
193 | static inline void zone_relock(uma_zone_t zone, uma_keg_t keg); |
---|
194 | static inline void keg_relock(uma_keg_t keg, uma_zone_t zone); |
---|
195 | |
---|
196 | void uma_print_zone(uma_zone_t); |
---|
197 | void uma_print_stats(void); |
---|
198 | |
---|
199 | /* |
---|
200 | * Initialize bucket_zones, the array of zones of buckets of various sizes. |
---|
201 | * |
---|
202 | * For each zone, calculate the memory required for each bucket, consisting |
---|
203 | * of the header and an array of pointers. Initialize bucket_size[] to point |
---|
204 | * the range of appropriate bucket sizes at the zone. |
---|
205 | */ |
---|
206 | static void |
---|
207 | bucket_init(void) |
---|
208 | { |
---|
209 | struct uma_bucket_zone *ubz; |
---|
210 | int i; |
---|
211 | int j; |
---|
212 | |
---|
213 | for (i = 0, j = 0; bucket_zones[j].ubz_entries != 0; j++) { |
---|
214 | int size; |
---|
215 | |
---|
216 | ubz = &bucket_zones[j]; |
---|
217 | size = roundup(sizeof(struct uma_bucket), sizeof(void *)); |
---|
218 | size += sizeof(void *) * ubz->ubz_entries; |
---|
219 | ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, |
---|
220 | NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, |
---|
221 | UMA_ZFLAG_INTERNAL | UMA_ZFLAG_BUCKET); |
---|
222 | for (; i <= ubz->ubz_entries; i += (1 << BUCKET_SHIFT)) |
---|
223 | bucket_size[i >> BUCKET_SHIFT] = j; |
---|
224 | } |
---|
225 | } |
---|
226 | |
---|
227 | /* |
---|
228 | * Given a desired number of entries for a bucket, return the zone from which |
---|
229 | * to allocate the bucket. |
---|
230 | */ |
---|
231 | static struct uma_bucket_zone * |
---|
232 | bucket_zone_lookup(int entries) |
---|
233 | { |
---|
234 | int idx; |
---|
235 | |
---|
236 | idx = howmany(entries, 1 << BUCKET_SHIFT); |
---|
237 | return (&bucket_zones[bucket_size[idx]]); |
---|
238 | } |
---|
239 | |
---|
240 | static uma_bucket_t |
---|
241 | bucket_alloc(int entries, int bflags) |
---|
242 | { |
---|
243 | struct uma_bucket_zone *ubz; |
---|
244 | uma_bucket_t bucket; |
---|
245 | |
---|
246 | /* |
---|
247 | * This is to stop us from allocating per cpu buckets while we're |
---|
248 | * running out of vm.boot_pages. Otherwise, we would exhaust the |
---|
249 | * boot pages. This also prevents us from allocating buckets in |
---|
250 | * low memory situations. |
---|
251 | */ |
---|
252 | if (bucketdisable) |
---|
253 | return (NULL); |
---|
254 | |
---|
255 | ubz = bucket_zone_lookup(entries); |
---|
256 | bucket = zone_alloc_item(ubz->ubz_zone, NULL, bflags); |
---|
257 | if (bucket) { |
---|
258 | #ifdef INVARIANTS |
---|
259 | bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); |
---|
260 | #endif |
---|
261 | bucket->ub_cnt = 0; |
---|
262 | bucket->ub_entries = ubz->ubz_entries; |
---|
263 | } |
---|
264 | |
---|
265 | return (bucket); |
---|
266 | } |
---|
267 | |
---|
268 | static void |
---|
269 | bucket_free(uma_bucket_t bucket) |
---|
270 | { |
---|
271 | struct uma_bucket_zone *ubz; |
---|
272 | |
---|
273 | ubz = bucket_zone_lookup(bucket->ub_entries); |
---|
274 | zone_free_item(ubz->ubz_zone, bucket, NULL, SKIP_NONE, |
---|
275 | ZFREE_STATFREE); |
---|
276 | } |
---|
277 | |
---|
278 | static void |
---|
279 | bucket_zone_drain(void) |
---|
280 | { |
---|
281 | struct uma_bucket_zone *ubz; |
---|
282 | |
---|
283 | for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) |
---|
284 | zone_drain(ubz->ubz_zone); |
---|
285 | } |
---|
286 | |
---|
287 | static inline uma_keg_t |
---|
288 | zone_first_keg(uma_zone_t zone) |
---|
289 | { |
---|
290 | |
---|
291 | return (LIST_FIRST(&zone->uz_kegs)->kl_keg); |
---|
292 | } |
---|
293 | |
---|
294 | static void |
---|
295 | zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) |
---|
296 | { |
---|
297 | uma_klink_t klink; |
---|
298 | |
---|
299 | LIST_FOREACH(klink, &zone->uz_kegs, kl_link) |
---|
300 | kegfn(klink->kl_keg); |
---|
301 | } |
---|
302 | |
---|
303 | /* |
---|
304 | * Routine to perform timeout driven calculations. This expands the |
---|
305 | * hashes and does per cpu statistics aggregation. |
---|
306 | * |
---|
307 | * Returns nothing. |
---|
308 | */ |
---|
309 | static void |
---|
310 | keg_timeout(uma_keg_t keg) |
---|
311 | { |
---|
312 | |
---|
313 | KEG_LOCK(keg); |
---|
314 | /* |
---|
315 | * Expand the keg hash table. |
---|
316 | * |
---|
317 | * This is done if the number of slabs is larger than the hash size. |
---|
318 | * What I'm trying to do here is completely reduce collisions. This |
---|
319 | * may be a little aggressive. Should I allow for two collisions max? |
---|
320 | */ |
---|
321 | if (keg->uk_flags & UMA_ZONE_HASH && |
---|
322 | keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { |
---|
323 | struct uma_hash newhash; |
---|
324 | struct uma_hash oldhash; |
---|
325 | int ret; |
---|
326 | |
---|
327 | /* |
---|
328 | * This is so involved because allocating and freeing |
---|
329 | * while the keg lock is held will lead to deadlock. |
---|
330 | * I have to do everything in stages and check for |
---|
331 | * races. |
---|
332 | */ |
---|
333 | newhash = keg->uk_hash; |
---|
334 | KEG_UNLOCK(keg); |
---|
335 | ret = hash_alloc(&newhash); |
---|
336 | KEG_LOCK(keg); |
---|
337 | if (ret) { |
---|
338 | if (hash_expand(&keg->uk_hash, &newhash)) { |
---|
339 | oldhash = keg->uk_hash; |
---|
340 | keg->uk_hash = newhash; |
---|
341 | } else |
---|
342 | oldhash = newhash; |
---|
343 | |
---|
344 | KEG_UNLOCK(keg); |
---|
345 | hash_free(&oldhash); |
---|
346 | KEG_LOCK(keg); |
---|
347 | } |
---|
348 | } |
---|
349 | KEG_UNLOCK(keg); |
---|
350 | } |
---|
351 | |
---|
352 | static void |
---|
353 | zone_timeout(uma_zone_t zone) |
---|
354 | { |
---|
355 | |
---|
356 | zone_foreach_keg(zone, &keg_timeout); |
---|
357 | } |
---|
358 | |
---|
359 | /* |
---|
360 | * Allocate and zero fill the next sized hash table from the appropriate |
---|
361 | * backing store. |
---|
362 | * |
---|
363 | * Arguments: |
---|
364 | * hash A new hash structure with the old hash size in uh_hashsize |
---|
365 | * |
---|
366 | * Returns: |
---|
367 | * 1 on sucess and 0 on failure. |
---|
368 | */ |
---|
369 | static int |
---|
370 | hash_alloc(struct uma_hash *hash) |
---|
371 | { |
---|
372 | int oldsize; |
---|
373 | int alloc; |
---|
374 | |
---|
375 | oldsize = hash->uh_hashsize; |
---|
376 | |
---|
377 | /* We're just going to go to a power of two greater */ |
---|
378 | if (oldsize) { |
---|
379 | hash->uh_hashsize = oldsize * 2; |
---|
380 | alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; |
---|
381 | hash->uh_slab_hash = (struct slabhead *)malloc(alloc, |
---|
382 | M_UMAHASH, M_NOWAIT); |
---|
383 | } else { |
---|
384 | alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; |
---|
385 | hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, |
---|
386 | M_WAITOK); |
---|
387 | hash->uh_hashsize = UMA_HASH_SIZE_INIT; |
---|
388 | } |
---|
389 | if (hash->uh_slab_hash) { |
---|
390 | bzero(hash->uh_slab_hash, alloc); |
---|
391 | hash->uh_hashmask = hash->uh_hashsize - 1; |
---|
392 | return (1); |
---|
393 | } |
---|
394 | |
---|
395 | return (0); |
---|
396 | } |
---|
397 | |
---|
398 | /* |
---|
399 | * Expands the hash table for HASH zones. This is done from zone_timeout |
---|
400 | * to reduce collisions. This must not be done in the regular allocation |
---|
401 | * path, otherwise, we can recurse on the vm while allocating pages. |
---|
402 | * |
---|
403 | * Arguments: |
---|
404 | * oldhash The hash you want to expand |
---|
405 | * newhash The hash structure for the new table |
---|
406 | * |
---|
407 | * Returns: |
---|
408 | * Nothing |
---|
409 | * |
---|
410 | * Discussion: |
---|
411 | */ |
---|
412 | static int |
---|
413 | hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) |
---|
414 | { |
---|
415 | uma_slab_t slab; |
---|
416 | int hval; |
---|
417 | int i; |
---|
418 | |
---|
419 | if (!newhash->uh_slab_hash) |
---|
420 | return (0); |
---|
421 | |
---|
422 | if (oldhash->uh_hashsize >= newhash->uh_hashsize) |
---|
423 | return (0); |
---|
424 | |
---|
425 | /* |
---|
426 | * I need to investigate hash algorithms for resizing without a |
---|
427 | * full rehash. |
---|
428 | */ |
---|
429 | |
---|
430 | for (i = 0; i < oldhash->uh_hashsize; i++) |
---|
431 | while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { |
---|
432 | slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); |
---|
433 | SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); |
---|
434 | hval = UMA_HASH(newhash, slab->us_data); |
---|
435 | SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], |
---|
436 | slab, us_hlink); |
---|
437 | } |
---|
438 | |
---|
439 | return (1); |
---|
440 | } |
---|
441 | |
---|
442 | /* |
---|
443 | * Free the hash bucket to the appropriate backing store. |
---|
444 | * |
---|
445 | * Arguments: |
---|
446 | * slab_hash The hash bucket we're freeing |
---|
447 | * hashsize The number of entries in that hash bucket |
---|
448 | * |
---|
449 | * Returns: |
---|
450 | * Nothing |
---|
451 | */ |
---|
452 | static void |
---|
453 | hash_free(struct uma_hash *hash) |
---|
454 | { |
---|
455 | if (hash->uh_slab_hash == NULL) |
---|
456 | return; |
---|
457 | if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) |
---|
458 | zone_free_item(hashzone, |
---|
459 | hash->uh_slab_hash, NULL, SKIP_NONE, ZFREE_STATFREE); |
---|
460 | else |
---|
461 | free(hash->uh_slab_hash, M_UMAHASH); |
---|
462 | } |
---|
463 | |
---|
464 | /* |
---|
465 | * Frees all outstanding items in a bucket |
---|
466 | * |
---|
467 | * Arguments: |
---|
468 | * zone The zone to free to, must be unlocked. |
---|
469 | * bucket The free/alloc bucket with items, cpu queue must be locked. |
---|
470 | * |
---|
471 | * Returns: |
---|
472 | * Nothing |
---|
473 | */ |
---|
474 | |
---|
475 | static void |
---|
476 | bucket_drain(uma_zone_t zone, uma_bucket_t bucket) |
---|
477 | { |
---|
478 | void *item; |
---|
479 | |
---|
480 | if (bucket == NULL) |
---|
481 | return; |
---|
482 | |
---|
483 | while (bucket->ub_cnt > 0) { |
---|
484 | bucket->ub_cnt--; |
---|
485 | item = bucket->ub_bucket[bucket->ub_cnt]; |
---|
486 | #ifdef INVARIANTS |
---|
487 | bucket->ub_bucket[bucket->ub_cnt] = NULL; |
---|
488 | KASSERT(item != NULL, |
---|
489 | ("bucket_drain: botched ptr, item is NULL")); |
---|
490 | #endif |
---|
491 | zone_free_item(zone, item, NULL, SKIP_DTOR, 0); |
---|
492 | } |
---|
493 | } |
---|
494 | |
---|
495 | /* |
---|
496 | * Drains the per cpu caches for a zone. |
---|
497 | * |
---|
498 | * NOTE: This may only be called while the zone is being turn down, and not |
---|
499 | * during normal operation. This is necessary in order that we do not have |
---|
500 | * to migrate CPUs to drain the per-CPU caches. |
---|
501 | * |
---|
502 | * Arguments: |
---|
503 | * zone The zone to drain, must be unlocked. |
---|
504 | * |
---|
505 | * Returns: |
---|
506 | * Nothing |
---|
507 | */ |
---|
508 | static void |
---|
509 | cache_drain(uma_zone_t zone) |
---|
510 | { |
---|
511 | uma_cache_t cache; |
---|
512 | int cpu; |
---|
513 | |
---|
514 | /* |
---|
515 | * XXX: It is safe to not lock the per-CPU caches, because we're |
---|
516 | * tearing down the zone anyway. I.e., there will be no further use |
---|
517 | * of the caches at this point. |
---|
518 | * |
---|
519 | * XXX: It would good to be able to assert that the zone is being |
---|
520 | * torn down to prevent improper use of cache_drain(). |
---|
521 | * |
---|
522 | * XXX: We lock the zone before passing into bucket_cache_drain() as |
---|
523 | * it is used elsewhere. Should the tear-down path be made special |
---|
524 | * there in some form? |
---|
525 | */ |
---|
526 | for (cpu = 0; cpu <= mp_maxid; cpu++) { |
---|
527 | if (CPU_ABSENT(cpu)) |
---|
528 | continue; |
---|
529 | cache = &zone->uz_cpu[cpu]; |
---|
530 | bucket_drain(zone, cache->uc_allocbucket); |
---|
531 | bucket_drain(zone, cache->uc_freebucket); |
---|
532 | if (cache->uc_allocbucket != NULL) |
---|
533 | bucket_free(cache->uc_allocbucket); |
---|
534 | if (cache->uc_freebucket != NULL) |
---|
535 | bucket_free(cache->uc_freebucket); |
---|
536 | cache->uc_allocbucket = cache->uc_freebucket = NULL; |
---|
537 | } |
---|
538 | ZONE_LOCK(zone); |
---|
539 | bucket_cache_drain(zone); |
---|
540 | ZONE_UNLOCK(zone); |
---|
541 | } |
---|
542 | |
---|
543 | /* |
---|
544 | * Drain the cached buckets from a zone. Expects a locked zone on entry. |
---|
545 | */ |
---|
546 | static void |
---|
547 | bucket_cache_drain(uma_zone_t zone) |
---|
548 | { |
---|
549 | uma_bucket_t bucket; |
---|
550 | |
---|
551 | /* |
---|
552 | * Drain the bucket queues and free the buckets, we just keep two per |
---|
553 | * cpu (alloc/free). |
---|
554 | */ |
---|
555 | while ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) { |
---|
556 | LIST_REMOVE(bucket, ub_link); |
---|
557 | ZONE_UNLOCK(zone); |
---|
558 | bucket_drain(zone, bucket); |
---|
559 | bucket_free(bucket); |
---|
560 | ZONE_LOCK(zone); |
---|
561 | } |
---|
562 | |
---|
563 | /* Now we do the free queue.. */ |
---|
564 | while ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) { |
---|
565 | LIST_REMOVE(bucket, ub_link); |
---|
566 | bucket_free(bucket); |
---|
567 | } |
---|
568 | } |
---|
569 | |
---|
570 | /* |
---|
571 | * Frees pages from a keg back to the system. This is done on demand from |
---|
572 | * the pageout daemon. |
---|
573 | * |
---|
574 | * Returns nothing. |
---|
575 | */ |
---|
576 | static void |
---|
577 | keg_drain(uma_keg_t keg) |
---|
578 | { |
---|
579 | struct slabhead freeslabs = { 0 }; |
---|
580 | uma_slab_t slab; |
---|
581 | uma_slab_t n; |
---|
582 | u_int8_t flags; |
---|
583 | u_int8_t *mem; |
---|
584 | int i; |
---|
585 | |
---|
586 | /* |
---|
587 | * We don't want to take pages from statically allocated kegs at this |
---|
588 | * time |
---|
589 | */ |
---|
590 | if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) |
---|
591 | return; |
---|
592 | |
---|
593 | #ifdef UMA_DEBUG |
---|
594 | printf("%s free items: %u\n", keg->uk_name, keg->uk_free); |
---|
595 | #endif |
---|
596 | KEG_LOCK(keg); |
---|
597 | if (keg->uk_free == 0) |
---|
598 | goto finished; |
---|
599 | |
---|
600 | slab = LIST_FIRST(&keg->uk_free_slab); |
---|
601 | while (slab) { |
---|
602 | n = LIST_NEXT(slab, us_link); |
---|
603 | |
---|
604 | /* We have no where to free these to */ |
---|
605 | if (slab->us_flags & UMA_SLAB_BOOT) { |
---|
606 | slab = n; |
---|
607 | continue; |
---|
608 | } |
---|
609 | |
---|
610 | LIST_REMOVE(slab, us_link); |
---|
611 | keg->uk_pages -= keg->uk_ppera; |
---|
612 | keg->uk_free -= keg->uk_ipers; |
---|
613 | |
---|
614 | if (keg->uk_flags & UMA_ZONE_HASH) |
---|
615 | UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data); |
---|
616 | |
---|
617 | SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); |
---|
618 | |
---|
619 | slab = n; |
---|
620 | } |
---|
621 | finished: |
---|
622 | KEG_UNLOCK(keg); |
---|
623 | |
---|
624 | while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { |
---|
625 | SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); |
---|
626 | if (keg->uk_fini) |
---|
627 | for (i = 0; i < keg->uk_ipers; i++) |
---|
628 | keg->uk_fini( |
---|
629 | slab->us_data + (keg->uk_rsize * i), |
---|
630 | keg->uk_size); |
---|
631 | flags = slab->us_flags; |
---|
632 | mem = slab->us_data; |
---|
633 | |
---|
634 | if (keg->uk_flags & UMA_ZONE_OFFPAGE) |
---|
635 | zone_free_item(keg->uk_slabzone, slab, NULL, |
---|
636 | SKIP_NONE, ZFREE_STATFREE); |
---|
637 | #ifdef UMA_DEBUG |
---|
638 | printf("%s: Returning %d bytes.\n", |
---|
639 | keg->uk_name, UMA_SLAB_SIZE * keg->uk_ppera); |
---|
640 | #endif |
---|
641 | keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, flags); |
---|
642 | } |
---|
643 | } |
---|
644 | |
---|
645 | static void |
---|
646 | zone_drain_wait(uma_zone_t zone, int waitok) |
---|
647 | { |
---|
648 | |
---|
649 | /* |
---|
650 | * Set draining to interlock with zone_dtor() so we can release our |
---|
651 | * locks as we go. Only dtor() should do a WAITOK call since it |
---|
652 | * is the only call that knows the structure will still be available |
---|
653 | * when it wakes up. |
---|
654 | */ |
---|
655 | ZONE_LOCK(zone); |
---|
656 | while (zone->uz_flags & UMA_ZFLAG_DRAINING) { |
---|
657 | if (waitok == M_NOWAIT) |
---|
658 | goto out; |
---|
659 | mtx_unlock(&uma_mtx); |
---|
660 | msleep(zone, zone->uz_lock, PVM, "zonedrain", 1); |
---|
661 | mtx_lock(&uma_mtx); |
---|
662 | } |
---|
663 | zone->uz_flags |= UMA_ZFLAG_DRAINING; |
---|
664 | bucket_cache_drain(zone); |
---|
665 | ZONE_UNLOCK(zone); |
---|
666 | /* |
---|
667 | * The DRAINING flag protects us from being freed while |
---|
668 | * we're running. Normally the uma_mtx would protect us but we |
---|
669 | * must be able to release and acquire the right lock for each keg. |
---|
670 | */ |
---|
671 | zone_foreach_keg(zone, &keg_drain); |
---|
672 | ZONE_LOCK(zone); |
---|
673 | zone->uz_flags &= ~UMA_ZFLAG_DRAINING; |
---|
674 | wakeup(zone); |
---|
675 | out: |
---|
676 | ZONE_UNLOCK(zone); |
---|
677 | } |
---|
678 | |
---|
679 | void |
---|
680 | zone_drain(uma_zone_t zone) |
---|
681 | { |
---|
682 | |
---|
683 | zone_drain_wait(zone, M_NOWAIT); |
---|
684 | } |
---|
685 | |
---|
686 | /* |
---|
687 | * Allocate a new slab for a keg. This does not insert the slab onto a list. |
---|
688 | * |
---|
689 | * Arguments: |
---|
690 | * wait Shall we wait? |
---|
691 | * |
---|
692 | * Returns: |
---|
693 | * The slab that was allocated or NULL if there is no memory and the |
---|
694 | * caller specified M_NOWAIT. |
---|
695 | */ |
---|
696 | static uma_slab_t |
---|
697 | keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait) |
---|
698 | { |
---|
699 | uma_slabrefcnt_t slabref; |
---|
700 | uma_alloc allocf; |
---|
701 | uma_slab_t slab; |
---|
702 | u_int8_t *mem; |
---|
703 | u_int8_t flags; |
---|
704 | int i; |
---|
705 | |
---|
706 | mtx_assert(&keg->uk_lock, MA_OWNED); |
---|
707 | slab = NULL; |
---|
708 | |
---|
709 | #ifdef UMA_DEBUG |
---|
710 | printf("slab_zalloc: Allocating a new slab for %s\n", keg->uk_name); |
---|
711 | #endif |
---|
712 | allocf = keg->uk_allocf; |
---|
713 | KEG_UNLOCK(keg); |
---|
714 | |
---|
715 | if (keg->uk_flags & UMA_ZONE_OFFPAGE) { |
---|
716 | slab = zone_alloc_item(keg->uk_slabzone, NULL, wait); |
---|
717 | if (slab == NULL) { |
---|
718 | KEG_LOCK(keg); |
---|
719 | return NULL; |
---|
720 | } |
---|
721 | } |
---|
722 | |
---|
723 | /* |
---|
724 | * This reproduces the old vm_zone behavior of zero filling pages the |
---|
725 | * first time they are added to a zone. |
---|
726 | * |
---|
727 | * Malloced items are zeroed in uma_zalloc. |
---|
728 | */ |
---|
729 | |
---|
730 | if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) |
---|
731 | wait |= M_ZERO; |
---|
732 | else |
---|
733 | wait &= ~M_ZERO; |
---|
734 | |
---|
735 | /* zone is passed for legacy reasons. */ |
---|
736 | mem = allocf(zone, keg->uk_ppera * UMA_SLAB_SIZE, &flags, wait); |
---|
737 | if (mem == NULL) { |
---|
738 | if (keg->uk_flags & UMA_ZONE_OFFPAGE) |
---|
739 | zone_free_item(keg->uk_slabzone, slab, NULL, |
---|
740 | SKIP_NONE, ZFREE_STATFREE); |
---|
741 | KEG_LOCK(keg); |
---|
742 | return (NULL); |
---|
743 | } |
---|
744 | |
---|
745 | /* Point the slab into the allocated memory */ |
---|
746 | if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) |
---|
747 | slab = (uma_slab_t )(mem + keg->uk_pgoff); |
---|
748 | |
---|
749 | slab->us_keg = keg; |
---|
750 | slab->us_data = mem; |
---|
751 | slab->us_freecount = keg->uk_ipers; |
---|
752 | slab->us_firstfree = 0; |
---|
753 | slab->us_flags = flags; |
---|
754 | |
---|
755 | if (keg->uk_flags & UMA_ZONE_REFCNT) { |
---|
756 | slabref = (uma_slabrefcnt_t)slab; |
---|
757 | for (i = 0; i < keg->uk_ipers; i++) { |
---|
758 | slabref->us_freelist[i].us_refcnt = 0; |
---|
759 | slabref->us_freelist[i].us_item = i+1; |
---|
760 | } |
---|
761 | } else { |
---|
762 | for (i = 0; i < keg->uk_ipers; i++) |
---|
763 | slab->us_freelist[i].us_item = i+1; |
---|
764 | } |
---|
765 | |
---|
766 | if (keg->uk_init != NULL) { |
---|
767 | for (i = 0; i < keg->uk_ipers; i++) |
---|
768 | if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), |
---|
769 | keg->uk_size, wait) != 0) |
---|
770 | break; |
---|
771 | if (i != keg->uk_ipers) { |
---|
772 | if (keg->uk_fini != NULL) { |
---|
773 | for (i--; i > -1; i--) |
---|
774 | keg->uk_fini(slab->us_data + |
---|
775 | (keg->uk_rsize * i), |
---|
776 | keg->uk_size); |
---|
777 | } |
---|
778 | if (keg->uk_flags & UMA_ZONE_OFFPAGE) |
---|
779 | zone_free_item(keg->uk_slabzone, slab, |
---|
780 | NULL, SKIP_NONE, ZFREE_STATFREE); |
---|
781 | keg->uk_freef(mem, UMA_SLAB_SIZE * keg->uk_ppera, |
---|
782 | flags); |
---|
783 | KEG_LOCK(keg); |
---|
784 | return (NULL); |
---|
785 | } |
---|
786 | } |
---|
787 | KEG_LOCK(keg); |
---|
788 | |
---|
789 | if (keg->uk_flags & UMA_ZONE_HASH) |
---|
790 | UMA_HASH_INSERT(&keg->uk_hash, slab, mem); |
---|
791 | |
---|
792 | keg->uk_pages += keg->uk_ppera; |
---|
793 | keg->uk_free += keg->uk_ipers; |
---|
794 | |
---|
795 | return (slab); |
---|
796 | } |
---|
797 | |
---|
798 | /* |
---|
799 | * This function is intended to be used early on in place of page_alloc() so |
---|
800 | * that we may use the boot time page cache to satisfy allocations before |
---|
801 | * the VM is ready. |
---|
802 | */ |
---|
803 | static void * |
---|
804 | startup_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait) |
---|
805 | { |
---|
806 | uma_keg_t keg; |
---|
807 | uma_slab_t tmps; |
---|
808 | int pages, check_pages; |
---|
809 | |
---|
810 | keg = zone_first_keg(zone); |
---|
811 | pages = howmany(bytes, PAGE_SIZE); |
---|
812 | check_pages = pages - 1; |
---|
813 | KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n")); |
---|
814 | |
---|
815 | /* |
---|
816 | * Check our small startup cache to see if it has pages remaining. |
---|
817 | */ |
---|
818 | mtx_lock(&uma_boot_pages_mtx); |
---|
819 | |
---|
820 | /* First check if we have enough room. */ |
---|
821 | tmps = LIST_FIRST(&uma_boot_pages); |
---|
822 | while (tmps != NULL && check_pages-- > 0) |
---|
823 | tmps = LIST_NEXT(tmps, us_link); |
---|
824 | if (tmps != NULL) { |
---|
825 | /* |
---|
826 | * It's ok to lose tmps references. The last one will |
---|
827 | * have tmps->us_data pointing to the start address of |
---|
828 | * "pages" contiguous pages of memory. |
---|
829 | */ |
---|
830 | while (pages-- > 0) { |
---|
831 | tmps = LIST_FIRST(&uma_boot_pages); |
---|
832 | LIST_REMOVE(tmps, us_link); |
---|
833 | } |
---|
834 | mtx_unlock(&uma_boot_pages_mtx); |
---|
835 | *pflag = tmps->us_flags; |
---|
836 | return (tmps->us_data); |
---|
837 | } |
---|
838 | mtx_unlock(&uma_boot_pages_mtx); |
---|
839 | if (booted == 0) |
---|
840 | panic("UMA: Increase vm.boot_pages"); |
---|
841 | /* |
---|
842 | * Now that we've booted reset these users to their real allocator. |
---|
843 | */ |
---|
844 | #ifdef UMA_MD_SMALL_ALLOC |
---|
845 | keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc; |
---|
846 | #else |
---|
847 | keg->uk_allocf = page_alloc; |
---|
848 | #endif |
---|
849 | return keg->uk_allocf(zone, bytes, pflag, wait); |
---|
850 | } |
---|
851 | |
---|
852 | /* |
---|
853 | * Allocates a number of pages from the system |
---|
854 | * |
---|
855 | * Arguments: |
---|
856 | * bytes The number of bytes requested |
---|
857 | * wait Shall we wait? |
---|
858 | * |
---|
859 | * Returns: |
---|
860 | * A pointer to the alloced memory or possibly |
---|
861 | * NULL if M_NOWAIT is set. |
---|
862 | */ |
---|
863 | static void * |
---|
864 | page_alloc(uma_zone_t zone, int bytes, u_int8_t *pflag, int wait) |
---|
865 | { |
---|
866 | void *p; /* Returned page */ |
---|
867 | |
---|
868 | *pflag = UMA_SLAB_KMEM; |
---|
869 | p = (void *) malloc(bytes, M_TEMP, wait); |
---|
870 | |
---|
871 | return (p); |
---|
872 | } |
---|
873 | |
---|
874 | /* |
---|
875 | * Frees a number of pages to the system |
---|
876 | * |
---|
877 | * Arguments: |
---|
878 | * mem A pointer to the memory to be freed |
---|
879 | * size The size of the memory being freed |
---|
880 | * flags The original p->us_flags field |
---|
881 | * |
---|
882 | * Returns: |
---|
883 | * Nothing |
---|
884 | */ |
---|
885 | static void |
---|
886 | page_free(void *mem, int size, u_int8_t flags) |
---|
887 | { |
---|
888 | free( mem, M_TEMP ); |
---|
889 | } |
---|
890 | |
---|
891 | /* |
---|
892 | * Zero fill initializer |
---|
893 | * |
---|
894 | * Arguments/Returns follow uma_init specifications |
---|
895 | */ |
---|
896 | static int |
---|
897 | zero_init(void *mem, int size, int flags) |
---|
898 | { |
---|
899 | bzero(mem, size); |
---|
900 | return (0); |
---|
901 | } |
---|
902 | |
---|
903 | /* |
---|
904 | * Finish creating a small uma keg. This calculates ipers, and the keg size. |
---|
905 | * |
---|
906 | * Arguments |
---|
907 | * keg The zone we should initialize |
---|
908 | * |
---|
909 | * Returns |
---|
910 | * Nothing |
---|
911 | */ |
---|
912 | static void |
---|
913 | keg_small_init(uma_keg_t keg) |
---|
914 | { |
---|
915 | u_int rsize; |
---|
916 | u_int memused; |
---|
917 | u_int wastedspace; |
---|
918 | u_int shsize; |
---|
919 | |
---|
920 | KASSERT(keg != NULL, ("Keg is null in keg_small_init")); |
---|
921 | rsize = keg->uk_size; |
---|
922 | |
---|
923 | if (rsize < UMA_SMALLEST_UNIT) |
---|
924 | rsize = UMA_SMALLEST_UNIT; |
---|
925 | if (rsize & keg->uk_align) |
---|
926 | rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); |
---|
927 | |
---|
928 | keg->uk_rsize = rsize; |
---|
929 | keg->uk_ppera = 1; |
---|
930 | |
---|
931 | if (keg->uk_flags & UMA_ZONE_REFCNT) { |
---|
932 | rsize += UMA_FRITMREF_SZ; /* linkage & refcnt */ |
---|
933 | shsize = sizeof(struct uma_slab_refcnt); |
---|
934 | } else { |
---|
935 | rsize += UMA_FRITM_SZ; /* Account for linkage */ |
---|
936 | shsize = sizeof(struct uma_slab); |
---|
937 | } |
---|
938 | |
---|
939 | keg->uk_ipers = (UMA_SLAB_SIZE - shsize) / rsize; |
---|
940 | KASSERT(keg->uk_ipers != 0, ("keg_small_init: ipers is 0")); |
---|
941 | memused = keg->uk_ipers * rsize + shsize; |
---|
942 | wastedspace = UMA_SLAB_SIZE - memused; |
---|
943 | |
---|
944 | /* |
---|
945 | * We can't do OFFPAGE if we're internal or if we've been |
---|
946 | * asked to not go to the VM for buckets. If we do this we |
---|
947 | * may end up going to the VM (kmem_map) for slabs which we |
---|
948 | * do not want to do if we're UMA_ZFLAG_CACHEONLY as a |
---|
949 | * result of UMA_ZONE_VM, which clearly forbids it. |
---|
950 | */ |
---|
951 | if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || |
---|
952 | (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) |
---|
953 | return; |
---|
954 | |
---|
955 | if ((wastedspace >= UMA_MAX_WASTE) && |
---|
956 | (keg->uk_ipers < (UMA_SLAB_SIZE / keg->uk_rsize))) { |
---|
957 | keg->uk_ipers = UMA_SLAB_SIZE / keg->uk_rsize; |
---|
958 | KASSERT(keg->uk_ipers <= 255, |
---|
959 | ("keg_small_init: keg->uk_ipers too high!")); |
---|
960 | #ifdef UMA_DEBUG |
---|
961 | printf("UMA decided we need offpage slab headers for " |
---|
962 | "keg: %s, calculated wastedspace = %d, " |
---|
963 | "maximum wasted space allowed = %d, " |
---|
964 | "calculated ipers = %d, " |
---|
965 | "new wasted space = %d\n", keg->uk_name, wastedspace, |
---|
966 | UMA_MAX_WASTE, keg->uk_ipers, |
---|
967 | UMA_SLAB_SIZE - keg->uk_ipers * keg->uk_rsize); |
---|
968 | #endif |
---|
969 | keg->uk_flags |= UMA_ZONE_OFFPAGE; |
---|
970 | if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) |
---|
971 | keg->uk_flags |= UMA_ZONE_HASH; |
---|
972 | } |
---|
973 | } |
---|
974 | |
---|
975 | /* |
---|
976 | * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do |
---|
977 | * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be |
---|
978 | * more complicated. |
---|
979 | * |
---|
980 | * Arguments |
---|
981 | * keg The keg we should initialize |
---|
982 | * |
---|
983 | * Returns |
---|
984 | * Nothing |
---|
985 | */ |
---|
986 | static void |
---|
987 | keg_large_init(uma_keg_t keg) |
---|
988 | { |
---|
989 | int pages; |
---|
990 | |
---|
991 | KASSERT(keg != NULL, ("Keg is null in keg_large_init")); |
---|
992 | KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, |
---|
993 | ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); |
---|
994 | |
---|
995 | pages = keg->uk_size / UMA_SLAB_SIZE; |
---|
996 | |
---|
997 | /* Account for remainder */ |
---|
998 | if ((pages * UMA_SLAB_SIZE) < keg->uk_size) |
---|
999 | pages++; |
---|
1000 | |
---|
1001 | keg->uk_ppera = pages; |
---|
1002 | keg->uk_ipers = 1; |
---|
1003 | keg->uk_rsize = keg->uk_size; |
---|
1004 | |
---|
1005 | /* We can't do OFFPAGE if we're internal, bail out here. */ |
---|
1006 | if (keg->uk_flags & UMA_ZFLAG_INTERNAL) |
---|
1007 | return; |
---|
1008 | |
---|
1009 | keg->uk_flags |= UMA_ZONE_OFFPAGE; |
---|
1010 | if ((keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) |
---|
1011 | keg->uk_flags |= UMA_ZONE_HASH; |
---|
1012 | } |
---|
1013 | |
---|
1014 | static void |
---|
1015 | keg_cachespread_init(uma_keg_t keg) |
---|
1016 | { |
---|
1017 | int alignsize; |
---|
1018 | int trailer; |
---|
1019 | int pages; |
---|
1020 | int rsize; |
---|
1021 | |
---|
1022 | alignsize = keg->uk_align + 1; |
---|
1023 | rsize = keg->uk_size; |
---|
1024 | /* |
---|
1025 | * We want one item to start on every align boundary in a page. To |
---|
1026 | * do this we will span pages. We will also extend the item by the |
---|
1027 | * size of align if it is an even multiple of align. Otherwise, it |
---|
1028 | * would fall on the same boundary every time. |
---|
1029 | */ |
---|
1030 | if (rsize & keg->uk_align) |
---|
1031 | rsize = (rsize & ~keg->uk_align) + alignsize; |
---|
1032 | if ((rsize & alignsize) == 0) |
---|
1033 | rsize += alignsize; |
---|
1034 | trailer = rsize - keg->uk_size; |
---|
1035 | pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; |
---|
1036 | pages = MIN(pages, (128 * 1024) / PAGE_SIZE); |
---|
1037 | keg->uk_rsize = rsize; |
---|
1038 | keg->uk_ppera = pages; |
---|
1039 | keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; |
---|
1040 | //keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; |
---|
1041 | KASSERT(keg->uk_ipers <= uma_max_ipers, |
---|
1042 | ("keg_small_init: keg->uk_ipers too high(%d) increase max_ipers", |
---|
1043 | keg->uk_ipers)); |
---|
1044 | } |
---|
1045 | |
---|
1046 | /* |
---|
1047 | * Keg header ctor. This initializes all fields, locks, etc. And inserts |
---|
1048 | * the keg onto the global keg list. |
---|
1049 | * |
---|
1050 | * Arguments/Returns follow uma_ctor specifications |
---|
1051 | * udata Actually uma_kctor_args |
---|
1052 | */ |
---|
1053 | static int |
---|
1054 | keg_ctor(void *mem, int size, void *udata, int flags) |
---|
1055 | { |
---|
1056 | struct uma_kctor_args *arg = udata; |
---|
1057 | uma_keg_t keg = mem; |
---|
1058 | uma_zone_t zone; |
---|
1059 | |
---|
1060 | bzero(keg, size); |
---|
1061 | keg->uk_size = arg->size; |
---|
1062 | keg->uk_init = arg->uminit; |
---|
1063 | keg->uk_fini = arg->fini; |
---|
1064 | keg->uk_align = arg->align; |
---|
1065 | keg->uk_free = 0; |
---|
1066 | keg->uk_pages = 0; |
---|
1067 | keg->uk_flags = arg->flags; |
---|
1068 | keg->uk_allocf = page_alloc; |
---|
1069 | keg->uk_freef = page_free; |
---|
1070 | keg->uk_recurse = 0; |
---|
1071 | keg->uk_slabzone = NULL; |
---|
1072 | |
---|
1073 | /* |
---|
1074 | * The master zone is passed to us at keg-creation time. |
---|
1075 | */ |
---|
1076 | zone = arg->zone; |
---|
1077 | keg->uk_name = zone->uz_name; |
---|
1078 | |
---|
1079 | if (arg->flags & UMA_ZONE_VM) |
---|
1080 | keg->uk_flags |= UMA_ZFLAG_CACHEONLY; |
---|
1081 | |
---|
1082 | if (arg->flags & UMA_ZONE_ZINIT) |
---|
1083 | keg->uk_init = zero_init; |
---|
1084 | |
---|
1085 | /*if (arg->flags & UMA_ZONE_REFCNT || arg->flags & UMA_ZONE_MALLOC) |
---|
1086 | keg->uk_flags |= UMA_ZONE_VTOSLAB;*/ |
---|
1087 | |
---|
1088 | /* |
---|
1089 | * The +UMA_FRITM_SZ added to uk_size is to account for the |
---|
1090 | * linkage that is added to the size in keg_small_init(). If |
---|
1091 | * we don't account for this here then we may end up in |
---|
1092 | * keg_small_init() with a calculated 'ipers' of 0. |
---|
1093 | */ |
---|
1094 | if (keg->uk_flags & UMA_ZONE_REFCNT) { |
---|
1095 | if (keg->uk_flags & UMA_ZONE_CACHESPREAD) |
---|
1096 | keg_cachespread_init(keg); |
---|
1097 | else if ((keg->uk_size+UMA_FRITMREF_SZ) > |
---|
1098 | (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt))) |
---|
1099 | keg_large_init(keg); |
---|
1100 | else |
---|
1101 | keg_small_init(keg); |
---|
1102 | } else { |
---|
1103 | if (keg->uk_flags & UMA_ZONE_CACHESPREAD) |
---|
1104 | keg_cachespread_init(keg); |
---|
1105 | else if ((keg->uk_size+UMA_FRITM_SZ) > |
---|
1106 | (UMA_SLAB_SIZE - sizeof(struct uma_slab))) |
---|
1107 | keg_large_init(keg); |
---|
1108 | else |
---|
1109 | keg_small_init(keg); |
---|
1110 | } |
---|
1111 | |
---|
1112 | if (keg->uk_flags & UMA_ZONE_OFFPAGE) { |
---|
1113 | if (keg->uk_flags & UMA_ZONE_REFCNT) |
---|
1114 | keg->uk_slabzone = slabrefzone; |
---|
1115 | else |
---|
1116 | keg->uk_slabzone = slabzone; |
---|
1117 | } |
---|
1118 | |
---|
1119 | /* |
---|
1120 | * If we haven't booted yet we need allocations to go through the |
---|
1121 | * startup cache until the vm is ready. |
---|
1122 | */ |
---|
1123 | if (keg->uk_ppera == 1) { |
---|
1124 | #ifdef UMA_MD_SMALL_ALLOC |
---|
1125 | keg->uk_allocf = uma_small_alloc; |
---|
1126 | keg->uk_freef = uma_small_free; |
---|
1127 | #endif |
---|
1128 | if (booted == 0) |
---|
1129 | keg->uk_allocf = startup_alloc; |
---|
1130 | } else if (booted == 0 && (keg->uk_flags & UMA_ZFLAG_INTERNAL)) |
---|
1131 | keg->uk_allocf = startup_alloc; |
---|
1132 | |
---|
1133 | /* |
---|
1134 | * Initialize keg's lock (shared among zones). |
---|
1135 | */ |
---|
1136 | if (arg->flags & UMA_ZONE_MTXCLASS) |
---|
1137 | KEG_LOCK_INIT(keg, 1); |
---|
1138 | else |
---|
1139 | KEG_LOCK_INIT(keg, 0); |
---|
1140 | |
---|
1141 | /* |
---|
1142 | * If we're putting the slab header in the actual page we need to |
---|
1143 | * figure out where in each page it goes. This calculates a right |
---|
1144 | * justified offset into the memory on an ALIGN_PTR boundary. |
---|
1145 | */ |
---|
1146 | if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { |
---|
1147 | u_int totsize; |
---|
1148 | |
---|
1149 | /* Size of the slab struct and free list */ |
---|
1150 | if (keg->uk_flags & UMA_ZONE_REFCNT) |
---|
1151 | totsize = sizeof(struct uma_slab_refcnt) + |
---|
1152 | keg->uk_ipers * UMA_FRITMREF_SZ; |
---|
1153 | else |
---|
1154 | totsize = sizeof(struct uma_slab) + |
---|
1155 | keg->uk_ipers * UMA_FRITM_SZ; |
---|
1156 | |
---|
1157 | if (totsize & UMA_ALIGN_PTR) |
---|
1158 | totsize = (totsize & ~UMA_ALIGN_PTR) + |
---|
1159 | (UMA_ALIGN_PTR + 1); |
---|
1160 | keg->uk_pgoff = (UMA_SLAB_SIZE * keg->uk_ppera) - totsize; |
---|
1161 | |
---|
1162 | if (keg->uk_flags & UMA_ZONE_REFCNT) |
---|
1163 | totsize = keg->uk_pgoff + sizeof(struct uma_slab_refcnt) |
---|
1164 | + keg->uk_ipers * UMA_FRITMREF_SZ; |
---|
1165 | else |
---|
1166 | totsize = keg->uk_pgoff + sizeof(struct uma_slab) |
---|
1167 | + keg->uk_ipers * UMA_FRITM_SZ; |
---|
1168 | |
---|
1169 | /* |
---|
1170 | * The only way the following is possible is if with our |
---|
1171 | * UMA_ALIGN_PTR adjustments we are now bigger than |
---|
1172 | * UMA_SLAB_SIZE. I haven't checked whether this is |
---|
1173 | * mathematically possible for all cases, so we make |
---|
1174 | * sure here anyway. |
---|
1175 | */ |
---|
1176 | if (totsize > UMA_SLAB_SIZE * keg->uk_ppera) { |
---|
1177 | printf("zone %s ipers %d rsize %d size %d\n", |
---|
1178 | zone->uz_name, keg->uk_ipers, keg->uk_rsize, |
---|
1179 | keg->uk_size); |
---|
1180 | panic("UMA slab won't fit."); |
---|
1181 | } |
---|
1182 | } |
---|
1183 | |
---|
1184 | if (keg->uk_flags & UMA_ZONE_HASH) |
---|
1185 | hash_alloc(&keg->uk_hash); |
---|
1186 | |
---|
1187 | #ifdef UMA_DEBUG |
---|
1188 | printf("UMA: %s(%p) size %d(%d) flags %d ipers %d ppera %d out %d free %d\n", |
---|
1189 | zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags, |
---|
1190 | keg->uk_ipers, keg->uk_ppera, |
---|
1191 | (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free); |
---|
1192 | #endif |
---|
1193 | |
---|
1194 | LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); |
---|
1195 | |
---|
1196 | mtx_lock(&uma_mtx); |
---|
1197 | LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); |
---|
1198 | mtx_unlock(&uma_mtx); |
---|
1199 | return (0); |
---|
1200 | } |
---|
1201 | |
---|
1202 | /* |
---|
1203 | * Zone header ctor. This initializes all fields, locks, etc. |
---|
1204 | * |
---|
1205 | * Arguments/Returns follow uma_ctor specifications |
---|
1206 | * udata Actually uma_zctor_args |
---|
1207 | */ |
---|
1208 | static int |
---|
1209 | zone_ctor(void *mem, int size, void *udata, int flags) |
---|
1210 | { |
---|
1211 | struct uma_zctor_args *arg = udata; |
---|
1212 | uma_zone_t zone = mem; |
---|
1213 | uma_zone_t z; |
---|
1214 | uma_keg_t keg; |
---|
1215 | |
---|
1216 | bzero(zone, size); |
---|
1217 | zone->uz_name = arg->name; |
---|
1218 | zone->uz_ctor = arg->ctor; |
---|
1219 | zone->uz_dtor = arg->dtor; |
---|
1220 | zone->uz_slab = zone_fetch_slab; |
---|
1221 | zone->uz_init = NULL; |
---|
1222 | zone->uz_fini = NULL; |
---|
1223 | zone->uz_allocs = 0; |
---|
1224 | zone->uz_frees = 0; |
---|
1225 | zone->uz_fails = 0; |
---|
1226 | zone->uz_fills = zone->uz_count = 0; |
---|
1227 | zone->uz_flags = 0; |
---|
1228 | keg = arg->keg; |
---|
1229 | |
---|
1230 | if (arg->flags & UMA_ZONE_SECONDARY) { |
---|
1231 | KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); |
---|
1232 | zone->uz_init = arg->uminit; |
---|
1233 | zone->uz_fini = arg->fini; |
---|
1234 | zone->uz_lock = &keg->uk_lock; |
---|
1235 | zone->uz_flags |= UMA_ZONE_SECONDARY; |
---|
1236 | mtx_lock(&uma_mtx); |
---|
1237 | ZONE_LOCK(zone); |
---|
1238 | LIST_FOREACH(z, &keg->uk_zones, uz_link) { |
---|
1239 | if (LIST_NEXT(z, uz_link) == NULL) { |
---|
1240 | LIST_INSERT_AFTER(z, zone, uz_link); |
---|
1241 | break; |
---|
1242 | } |
---|
1243 | } |
---|
1244 | ZONE_UNLOCK(zone); |
---|
1245 | mtx_unlock(&uma_mtx); |
---|
1246 | } else if (keg == NULL) { |
---|
1247 | if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, |
---|
1248 | arg->align, arg->flags)) == NULL) |
---|
1249 | return (ENOMEM); |
---|
1250 | } else { |
---|
1251 | struct uma_kctor_args karg; |
---|
1252 | int error; |
---|
1253 | |
---|
1254 | /* We should only be here from uma_startup() */ |
---|
1255 | karg.size = arg->size; |
---|
1256 | karg.uminit = arg->uminit; |
---|
1257 | karg.fini = arg->fini; |
---|
1258 | karg.align = arg->align; |
---|
1259 | karg.flags = arg->flags; |
---|
1260 | karg.zone = zone; |
---|
1261 | error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, |
---|
1262 | flags); |
---|
1263 | if (error) |
---|
1264 | return (error); |
---|
1265 | } |
---|
1266 | /* |
---|
1267 | * Link in the first keg. |
---|
1268 | */ |
---|
1269 | zone->uz_klink.kl_keg = keg; |
---|
1270 | LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); |
---|
1271 | zone->uz_lock = &keg->uk_lock; |
---|
1272 | zone->uz_size = keg->uk_size; |
---|
1273 | zone->uz_flags |= (keg->uk_flags & |
---|
1274 | (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); |
---|
1275 | |
---|
1276 | /* |
---|
1277 | * Some internal zones don't have room allocated for the per cpu |
---|
1278 | * caches. If we're internal, bail out here. |
---|
1279 | */ |
---|
1280 | if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { |
---|
1281 | KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, |
---|
1282 | ("Secondary zone requested UMA_ZFLAG_INTERNAL")); |
---|
1283 | return (0); |
---|
1284 | } |
---|
1285 | |
---|
1286 | if (keg->uk_flags & UMA_ZONE_MAXBUCKET) |
---|
1287 | zone->uz_count = BUCKET_MAX; |
---|
1288 | else if (keg->uk_ipers <= BUCKET_MAX) |
---|
1289 | zone->uz_count = keg->uk_ipers; |
---|
1290 | else |
---|
1291 | zone->uz_count = BUCKET_MAX; |
---|
1292 | return (0); |
---|
1293 | } |
---|
1294 | |
---|
1295 | /* |
---|
1296 | * Keg header dtor. This frees all data, destroys locks, frees the hash |
---|
1297 | * table and removes the keg from the global list. |
---|
1298 | * |
---|
1299 | * Arguments/Returns follow uma_dtor specifications |
---|
1300 | * udata unused |
---|
1301 | */ |
---|
1302 | static void |
---|
1303 | keg_dtor(void *arg, int size, void *udata) |
---|
1304 | { |
---|
1305 | uma_keg_t keg; |
---|
1306 | |
---|
1307 | keg = (uma_keg_t)arg; |
---|
1308 | KEG_LOCK(keg); |
---|
1309 | if (keg->uk_free != 0) { |
---|
1310 | printf("Freed UMA keg was not empty (%d items). " |
---|
1311 | " Lost %d pages of memory.\n", |
---|
1312 | keg->uk_free, keg->uk_pages); |
---|
1313 | } |
---|
1314 | KEG_UNLOCK(keg); |
---|
1315 | |
---|
1316 | hash_free(&keg->uk_hash); |
---|
1317 | |
---|
1318 | KEG_LOCK_FINI(keg); |
---|
1319 | } |
---|
1320 | |
---|
1321 | /* |
---|
1322 | * Zone header dtor. |
---|
1323 | * |
---|
1324 | * Arguments/Returns follow uma_dtor specifications |
---|
1325 | * udata unused |
---|
1326 | */ |
---|
1327 | static void |
---|
1328 | zone_dtor(void *arg, int size, void *udata) |
---|
1329 | { |
---|
1330 | uma_klink_t klink; |
---|
1331 | uma_zone_t zone; |
---|
1332 | uma_keg_t keg; |
---|
1333 | |
---|
1334 | zone = (uma_zone_t)arg; |
---|
1335 | keg = zone_first_keg(zone); |
---|
1336 | |
---|
1337 | if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) |
---|
1338 | cache_drain(zone); |
---|
1339 | |
---|
1340 | mtx_lock(&uma_mtx); |
---|
1341 | LIST_REMOVE(zone, uz_link); |
---|
1342 | mtx_unlock(&uma_mtx); |
---|
1343 | /* |
---|
1344 | * XXX there are some races here where |
---|
1345 | * the zone can be drained but zone lock |
---|
1346 | * released and then refilled before we |
---|
1347 | * remove it... we dont care for now |
---|
1348 | */ |
---|
1349 | zone_drain_wait(zone, M_WAITOK); |
---|
1350 | /* |
---|
1351 | * Unlink all of our kegs. |
---|
1352 | */ |
---|
1353 | while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { |
---|
1354 | klink->kl_keg = NULL; |
---|
1355 | LIST_REMOVE(klink, kl_link); |
---|
1356 | if (klink == &zone->uz_klink) |
---|
1357 | continue; |
---|
1358 | free(klink, M_TEMP); |
---|
1359 | } |
---|
1360 | /* |
---|
1361 | * We only destroy kegs from non secondary zones. |
---|
1362 | */ |
---|
1363 | if ((zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { |
---|
1364 | mtx_lock(&uma_mtx); |
---|
1365 | LIST_REMOVE(keg, uk_link); |
---|
1366 | mtx_unlock(&uma_mtx); |
---|
1367 | zone_free_item(kegs, keg, NULL, SKIP_NONE, |
---|
1368 | ZFREE_STATFREE); |
---|
1369 | } |
---|
1370 | } |
---|
1371 | |
---|
1372 | /* |
---|
1373 | * Traverses every zone in the system and calls a callback |
---|
1374 | * |
---|
1375 | * Arguments: |
---|
1376 | * zfunc A pointer to a function which accepts a zone |
---|
1377 | * as an argument. |
---|
1378 | * |
---|
1379 | * Returns: |
---|
1380 | * Nothing |
---|
1381 | */ |
---|
1382 | static void |
---|
1383 | zone_foreach(void (*zfunc)(uma_zone_t)) |
---|
1384 | { |
---|
1385 | uma_keg_t keg; |
---|
1386 | uma_zone_t zone; |
---|
1387 | |
---|
1388 | mtx_lock(&uma_mtx); |
---|
1389 | LIST_FOREACH(keg, &uma_kegs, uk_link) { |
---|
1390 | LIST_FOREACH(zone, &keg->uk_zones, uz_link) |
---|
1391 | zfunc(zone); |
---|
1392 | } |
---|
1393 | mtx_unlock(&uma_mtx); |
---|
1394 | } |
---|
1395 | |
---|
1396 | /* Public functions */ |
---|
1397 | /* See uma.h */ |
---|
1398 | void |
---|
1399 | uma_startup(void *bootmem, int boot_pages) |
---|
1400 | { |
---|
1401 | struct uma_zctor_args args; |
---|
1402 | uma_slab_t slab; |
---|
1403 | u_int slabsize; |
---|
1404 | u_int objsize, totsize, wsize; |
---|
1405 | int i; |
---|
1406 | |
---|
1407 | #ifdef UMA_DEBUG |
---|
1408 | printf("Creating uma keg headers zone and keg.\n"); |
---|
1409 | #endif |
---|
1410 | mtx_init(&uma_mtx, "UMA lock", NULL, MTX_DEF); |
---|
1411 | |
---|
1412 | /* |
---|
1413 | * Figure out the maximum number of items-per-slab we'll have if |
---|
1414 | * we're using the OFFPAGE slab header to track free items, given |
---|
1415 | * all possible object sizes and the maximum desired wastage |
---|
1416 | * (UMA_MAX_WASTE). |
---|
1417 | * |
---|
1418 | * We iterate until we find an object size for |
---|
1419 | * which the calculated wastage in keg_small_init() will be |
---|
1420 | * enough to warrant OFFPAGE. Since wastedspace versus objsize |
---|
1421 | * is an overall increasing see-saw function, we find the smallest |
---|
1422 | * objsize such that the wastage is always acceptable for objects |
---|
1423 | * with that objsize or smaller. Since a smaller objsize always |
---|
1424 | * generates a larger possible uma_max_ipers, we use this computed |
---|
1425 | * objsize to calculate the largest ipers possible. Since the |
---|
1426 | * ipers calculated for OFFPAGE slab headers is always larger than |
---|
1427 | * the ipers initially calculated in keg_small_init(), we use |
---|
1428 | * the former's equation (UMA_SLAB_SIZE / keg->uk_rsize) to |
---|
1429 | * obtain the maximum ipers possible for offpage slab headers. |
---|
1430 | * |
---|
1431 | * It should be noted that ipers versus objsize is an inversly |
---|
1432 | * proportional function which drops off rather quickly so as |
---|
1433 | * long as our UMA_MAX_WASTE is such that the objsize we calculate |
---|
1434 | * falls into the portion of the inverse relation AFTER the steep |
---|
1435 | * falloff, then uma_max_ipers shouldn't be too high (~10 on i386). |
---|
1436 | * |
---|
1437 | * Note that we have 8-bits (1 byte) to use as a freelist index |
---|
1438 | * inside the actual slab header itself and this is enough to |
---|
1439 | * accomodate us. In the worst case, a UMA_SMALLEST_UNIT sized |
---|
1440 | * object with offpage slab header would have ipers = |
---|
1441 | * UMA_SLAB_SIZE / UMA_SMALLEST_UNIT (currently = 256), which is |
---|
1442 | * 1 greater than what our byte-integer freelist index can |
---|
1443 | * accomodate, but we know that this situation never occurs as |
---|
1444 | * for UMA_SMALLEST_UNIT-sized objects, we will never calculate |
---|
1445 | * that we need to go to offpage slab headers. Or, if we do, |
---|
1446 | * then we trap that condition below and panic in the INVARIANTS case. |
---|
1447 | */ |
---|
1448 | wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab) - UMA_MAX_WASTE; |
---|
1449 | totsize = wsize; |
---|
1450 | objsize = UMA_SMALLEST_UNIT; |
---|
1451 | while (totsize >= wsize) { |
---|
1452 | totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab)) / |
---|
1453 | (objsize + UMA_FRITM_SZ); |
---|
1454 | totsize *= (UMA_FRITM_SZ + objsize); |
---|
1455 | objsize++; |
---|
1456 | } |
---|
1457 | if (objsize > UMA_SMALLEST_UNIT) |
---|
1458 | objsize--; |
---|
1459 | uma_max_ipers = MAX(UMA_SLAB_SIZE / objsize, 64); |
---|
1460 | |
---|
1461 | wsize = UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt) - UMA_MAX_WASTE; |
---|
1462 | totsize = wsize; |
---|
1463 | objsize = UMA_SMALLEST_UNIT; |
---|
1464 | while (totsize >= wsize) { |
---|
1465 | totsize = (UMA_SLAB_SIZE - sizeof(struct uma_slab_refcnt)) / |
---|
1466 | (objsize + UMA_FRITMREF_SZ); |
---|
1467 | totsize *= (UMA_FRITMREF_SZ + objsize); |
---|
1468 | objsize++; |
---|
1469 | } |
---|
1470 | if (objsize > UMA_SMALLEST_UNIT) |
---|
1471 | objsize--; |
---|
1472 | uma_max_ipers_ref = MAX(UMA_SLAB_SIZE / objsize, 64); |
---|
1473 | |
---|
1474 | KASSERT((uma_max_ipers_ref <= 255) && (uma_max_ipers <= 255), |
---|
1475 | ("uma_startup: calculated uma_max_ipers values too large!")); |
---|
1476 | |
---|
1477 | #ifdef UMA_DEBUG |
---|
1478 | printf("Calculated uma_max_ipers (for OFFPAGE) is %d\n", uma_max_ipers); |
---|
1479 | printf("Calculated uma_max_ipers_slab (for OFFPAGE) is %d\n", |
---|
1480 | uma_max_ipers_ref); |
---|
1481 | #endif |
---|
1482 | |
---|
1483 | /* "manually" create the initial zone */ |
---|
1484 | args.name = "UMA Kegs"; |
---|
1485 | args.size = sizeof(struct uma_keg); |
---|
1486 | args.ctor = keg_ctor; |
---|
1487 | args.dtor = keg_dtor; |
---|
1488 | args.uminit = zero_init; |
---|
1489 | args.fini = NULL; |
---|
1490 | args.keg = &masterkeg; |
---|
1491 | args.align = 32 - 1; |
---|
1492 | args.flags = UMA_ZFLAG_INTERNAL; |
---|
1493 | /* The initial zone has no Per cpu queues so it's smaller */ |
---|
1494 | zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK); |
---|
1495 | |
---|
1496 | #ifdef UMA_DEBUG |
---|
1497 | printf("Filling boot free list.\n"); |
---|
1498 | #endif |
---|
1499 | for (i = 0; i < boot_pages; i++) { |
---|
1500 | slab = (uma_slab_t)((u_int8_t *)bootmem + (i * UMA_SLAB_SIZE)); |
---|
1501 | slab->us_data = (u_int8_t *)slab; |
---|
1502 | slab->us_flags = UMA_SLAB_BOOT; |
---|
1503 | LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link); |
---|
1504 | } |
---|
1505 | mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF); |
---|
1506 | |
---|
1507 | #ifdef UMA_DEBUG |
---|
1508 | printf("Creating uma zone headers zone and keg.\n"); |
---|
1509 | #endif |
---|
1510 | args.name = "UMA Zones"; |
---|
1511 | args.size = sizeof(struct uma_zone) + |
---|
1512 | (sizeof(struct uma_cache) * (mp_maxid + 1)); |
---|
1513 | args.ctor = zone_ctor; |
---|
1514 | args.dtor = zone_dtor; |
---|
1515 | args.uminit = zero_init; |
---|
1516 | args.fini = NULL; |
---|
1517 | args.keg = NULL; |
---|
1518 | args.align = 32 - 1; |
---|
1519 | args.flags = UMA_ZFLAG_INTERNAL; |
---|
1520 | /* The initial zone has no Per cpu queues so it's smaller */ |
---|
1521 | zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK); |
---|
1522 | |
---|
1523 | #ifdef UMA_DEBUG |
---|
1524 | printf("Initializing pcpu cache locks.\n"); |
---|
1525 | #endif |
---|
1526 | #ifdef UMA_DEBUG |
---|
1527 | printf("Creating slab and hash zones.\n"); |
---|
1528 | #endif |
---|
1529 | |
---|
1530 | /* |
---|
1531 | * This is the max number of free list items we'll have with |
---|
1532 | * offpage slabs. |
---|
1533 | */ |
---|
1534 | slabsize = uma_max_ipers * UMA_FRITM_SZ; |
---|
1535 | slabsize += sizeof(struct uma_slab); |
---|
1536 | |
---|
1537 | /* Now make a zone for slab headers */ |
---|
1538 | slabzone = uma_zcreate("UMA Slabs", |
---|
1539 | slabsize, |
---|
1540 | NULL, NULL, NULL, NULL, |
---|
1541 | UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); |
---|
1542 | |
---|
1543 | /* |
---|
1544 | * We also create a zone for the bigger slabs with reference |
---|
1545 | * counts in them, to accomodate UMA_ZONE_REFCNT zones. |
---|
1546 | */ |
---|
1547 | slabsize = uma_max_ipers_ref * UMA_FRITMREF_SZ; |
---|
1548 | slabsize += sizeof(struct uma_slab_refcnt); |
---|
1549 | slabrefzone = uma_zcreate("UMA RCntSlabs", |
---|
1550 | slabsize, |
---|
1551 | NULL, NULL, NULL, NULL, |
---|
1552 | UMA_ALIGN_PTR, |
---|
1553 | UMA_ZFLAG_INTERNAL); |
---|
1554 | |
---|
1555 | hashzone = uma_zcreate("UMA Hash", |
---|
1556 | sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, |
---|
1557 | NULL, NULL, NULL, NULL, |
---|
1558 | UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); |
---|
1559 | |
---|
1560 | bucket_init(); |
---|
1561 | |
---|
1562 | #if defined(UMA_MD_SMALL_ALLOC) && !defined(UMA_MD_SMALL_ALLOC_NEEDS_VM) |
---|
1563 | booted = 1; |
---|
1564 | #endif |
---|
1565 | |
---|
1566 | #ifdef UMA_DEBUG |
---|
1567 | printf("UMA startup complete.\n"); |
---|
1568 | #endif |
---|
1569 | } |
---|
1570 | |
---|
1571 | static uma_keg_t |
---|
1572 | uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, |
---|
1573 | int align, u_int32_t flags) |
---|
1574 | { |
---|
1575 | struct uma_kctor_args args; |
---|
1576 | |
---|
1577 | args.size = size; |
---|
1578 | args.uminit = uminit; |
---|
1579 | args.fini = fini; |
---|
1580 | args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; |
---|
1581 | args.flags = flags; |
---|
1582 | args.zone = zone; |
---|
1583 | return (zone_alloc_item(kegs, &args, M_WAITOK)); |
---|
1584 | } |
---|
1585 | |
---|
1586 | /* See uma.h */ |
---|
1587 | void |
---|
1588 | uma_set_align(int align) |
---|
1589 | { |
---|
1590 | |
---|
1591 | if (align != UMA_ALIGN_CACHE) |
---|
1592 | uma_align_cache = align; |
---|
1593 | } |
---|
1594 | |
---|
1595 | /* See uma.h */ |
---|
1596 | uma_zone_t |
---|
1597 | uma_zcreate(char *name, size_t size, uma_ctor ctor, uma_dtor dtor, |
---|
1598 | uma_init uminit, uma_fini fini, int align, u_int32_t flags) |
---|
1599 | |
---|
1600 | { |
---|
1601 | struct uma_zctor_args args; |
---|
1602 | |
---|
1603 | /* This stuff is essential for the zone ctor */ |
---|
1604 | args.name = name; |
---|
1605 | args.size = size; |
---|
1606 | args.ctor = ctor; |
---|
1607 | args.dtor = dtor; |
---|
1608 | args.uminit = uminit; |
---|
1609 | args.fini = fini; |
---|
1610 | args.align = align; |
---|
1611 | args.flags = flags; |
---|
1612 | args.keg = NULL; |
---|
1613 | |
---|
1614 | return (zone_alloc_item(zones, &args, M_WAITOK)); |
---|
1615 | } |
---|
1616 | |
---|
1617 | /* See uma.h */ |
---|
1618 | uma_zone_t |
---|
1619 | uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, |
---|
1620 | uma_init zinit, uma_fini zfini, uma_zone_t master) |
---|
1621 | { |
---|
1622 | struct uma_zctor_args args; |
---|
1623 | uma_keg_t keg; |
---|
1624 | |
---|
1625 | keg = zone_first_keg(master); |
---|
1626 | args.name = name; |
---|
1627 | args.size = keg->uk_size; |
---|
1628 | args.ctor = ctor; |
---|
1629 | args.dtor = dtor; |
---|
1630 | args.uminit = zinit; |
---|
1631 | args.fini = zfini; |
---|
1632 | args.align = keg->uk_align; |
---|
1633 | args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; |
---|
1634 | args.keg = keg; |
---|
1635 | |
---|
1636 | /* XXX Attaches only one keg of potentially many. */ |
---|
1637 | return (zone_alloc_item(zones, &args, M_WAITOK)); |
---|
1638 | } |
---|
1639 | |
---|
1640 | static void |
---|
1641 | zone_lock_pair(uma_zone_t a, uma_zone_t b) |
---|
1642 | { |
---|
1643 | if (a < b) { |
---|
1644 | ZONE_LOCK(a); |
---|
1645 | mtx_lock_flags(b->uz_lock, MTX_DUPOK); |
---|
1646 | } else { |
---|
1647 | ZONE_LOCK(b); |
---|
1648 | mtx_lock_flags(a->uz_lock, MTX_DUPOK); |
---|
1649 | } |
---|
1650 | } |
---|
1651 | |
---|
1652 | static void |
---|
1653 | zone_unlock_pair(uma_zone_t a, uma_zone_t b) |
---|
1654 | { |
---|
1655 | |
---|
1656 | ZONE_UNLOCK(a); |
---|
1657 | ZONE_UNLOCK(b); |
---|
1658 | } |
---|
1659 | |
---|
1660 | |
---|
1661 | /* See uma.h */ |
---|
1662 | void |
---|
1663 | uma_zdestroy(uma_zone_t zone) |
---|
1664 | { |
---|
1665 | |
---|
1666 | zone_free_item(zones, zone, NULL, SKIP_NONE, ZFREE_STATFREE); |
---|
1667 | } |
---|
1668 | |
---|
1669 | /* See uma.h */ |
---|
1670 | void * |
---|
1671 | uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) |
---|
1672 | { |
---|
1673 | void *item; |
---|
1674 | uma_cache_t cache; |
---|
1675 | uma_bucket_t bucket; |
---|
1676 | int cpu; |
---|
1677 | |
---|
1678 | /* This is the fast path allocation */ |
---|
1679 | #ifdef UMA_DEBUG_ALLOC_1 |
---|
1680 | printf("Allocating one item from %s(%p)\n", zone->uz_name, zone); |
---|
1681 | #endif |
---|
1682 | CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread, |
---|
1683 | zone->uz_name, flags); |
---|
1684 | |
---|
1685 | if (flags & M_WAITOK) { |
---|
1686 | WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, |
---|
1687 | "uma_zalloc_arg: zone \"%s\"", zone->uz_name); |
---|
1688 | } |
---|
1689 | |
---|
1690 | /* |
---|
1691 | * If possible, allocate from the per-CPU cache. There are two |
---|
1692 | * requirements for safe access to the per-CPU cache: (1) the thread |
---|
1693 | * accessing the cache must not be preempted or yield during access, |
---|
1694 | * and (2) the thread must not migrate CPUs without switching which |
---|
1695 | * cache it accesses. We rely on a critical section to prevent |
---|
1696 | * preemption and migration. We release the critical section in |
---|
1697 | * order to acquire the zone mutex if we are unable to allocate from |
---|
1698 | * the current cache; when we re-acquire the critical section, we |
---|
1699 | * must detect and handle migration if it has occurred. |
---|
1700 | */ |
---|
1701 | zalloc_restart: |
---|
1702 | critical_enter(); |
---|
1703 | cpu = curcpu; |
---|
1704 | cache = &zone->uz_cpu[cpu]; |
---|
1705 | |
---|
1706 | zalloc_start: |
---|
1707 | bucket = cache->uc_allocbucket; |
---|
1708 | |
---|
1709 | if (bucket) { |
---|
1710 | if (bucket->ub_cnt > 0) { |
---|
1711 | bucket->ub_cnt--; |
---|
1712 | item = bucket->ub_bucket[bucket->ub_cnt]; |
---|
1713 | #ifdef INVARIANTS |
---|
1714 | bucket->ub_bucket[bucket->ub_cnt] = NULL; |
---|
1715 | #endif |
---|
1716 | KASSERT(item != NULL, |
---|
1717 | ("uma_zalloc: Bucket pointer mangled.")); |
---|
1718 | cache->uc_allocs++; |
---|
1719 | critical_exit(); |
---|
1720 | #ifdef INVARIANTS |
---|
1721 | ZONE_LOCK(zone); |
---|
1722 | uma_dbg_alloc(zone, NULL, item); |
---|
1723 | ZONE_UNLOCK(zone); |
---|
1724 | #endif |
---|
1725 | if (zone->uz_ctor != NULL) { |
---|
1726 | if (zone->uz_ctor(item, zone->uz_size, |
---|
1727 | udata, flags) != 0) { |
---|
1728 | zone_free_item(zone, item, udata, |
---|
1729 | SKIP_DTOR, ZFREE_STATFAIL | |
---|
1730 | ZFREE_STATFREE); |
---|
1731 | return (NULL); |
---|
1732 | } |
---|
1733 | } |
---|
1734 | if (flags & M_ZERO) |
---|
1735 | bzero(item, zone->uz_size); |
---|
1736 | return (item); |
---|
1737 | } else if (cache->uc_freebucket) { |
---|
1738 | /* |
---|
1739 | * We have run out of items in our allocbucket. |
---|
1740 | * See if we can switch with our free bucket. |
---|
1741 | */ |
---|
1742 | if (cache->uc_freebucket->ub_cnt > 0) { |
---|
1743 | #ifdef UMA_DEBUG_ALLOC |
---|
1744 | printf("uma_zalloc: Swapping empty with" |
---|
1745 | " alloc.\n"); |
---|
1746 | #endif |
---|
1747 | bucket = cache->uc_freebucket; |
---|
1748 | cache->uc_freebucket = cache->uc_allocbucket; |
---|
1749 | cache->uc_allocbucket = bucket; |
---|
1750 | |
---|
1751 | goto zalloc_start; |
---|
1752 | } |
---|
1753 | } |
---|
1754 | } |
---|
1755 | /* |
---|
1756 | * Attempt to retrieve the item from the per-CPU cache has failed, so |
---|
1757 | * we must go back to the zone. This requires the zone lock, so we |
---|
1758 | * must drop the critical section, then re-acquire it when we go back |
---|
1759 | * to the cache. Since the critical section is released, we may be |
---|
1760 | * preempted or migrate. As such, make sure not to maintain any |
---|
1761 | * thread-local state specific to the cache from prior to releasing |
---|
1762 | * the critical section. |
---|
1763 | */ |
---|
1764 | critical_exit(); |
---|
1765 | ZONE_LOCK(zone); |
---|
1766 | critical_enter(); |
---|
1767 | cpu = curcpu; |
---|
1768 | cache = &zone->uz_cpu[cpu]; |
---|
1769 | bucket = cache->uc_allocbucket; |
---|
1770 | if (bucket != NULL) { |
---|
1771 | if (bucket->ub_cnt > 0) { |
---|
1772 | ZONE_UNLOCK(zone); |
---|
1773 | goto zalloc_start; |
---|
1774 | } |
---|
1775 | bucket = cache->uc_freebucket; |
---|
1776 | if (bucket != NULL && bucket->ub_cnt > 0) { |
---|
1777 | ZONE_UNLOCK(zone); |
---|
1778 | goto zalloc_start; |
---|
1779 | } |
---|
1780 | } |
---|
1781 | |
---|
1782 | /* Since we have locked the zone we may as well send back our stats */ |
---|
1783 | zone->uz_allocs += cache->uc_allocs; |
---|
1784 | cache->uc_allocs = 0; |
---|
1785 | zone->uz_frees += cache->uc_frees; |
---|
1786 | cache->uc_frees = 0; |
---|
1787 | |
---|
1788 | /* Our old one is now a free bucket */ |
---|
1789 | if (cache->uc_allocbucket) { |
---|
1790 | KASSERT(cache->uc_allocbucket->ub_cnt == 0, |
---|
1791 | ("uma_zalloc_arg: Freeing a non free bucket.")); |
---|
1792 | LIST_INSERT_HEAD(&zone->uz_free_bucket, |
---|
1793 | cache->uc_allocbucket, ub_link); |
---|
1794 | cache->uc_allocbucket = NULL; |
---|
1795 | } |
---|
1796 | |
---|
1797 | /* Check the free list for a new alloc bucket */ |
---|
1798 | if ((bucket = LIST_FIRST(&zone->uz_full_bucket)) != NULL) { |
---|
1799 | KASSERT(bucket->ub_cnt != 0, |
---|
1800 | ("uma_zalloc_arg: Returning an empty bucket.")); |
---|
1801 | |
---|
1802 | LIST_REMOVE(bucket, ub_link); |
---|
1803 | cache->uc_allocbucket = bucket; |
---|
1804 | ZONE_UNLOCK(zone); |
---|
1805 | goto zalloc_start; |
---|
1806 | } |
---|
1807 | /* We are no longer associated with this CPU. */ |
---|
1808 | critical_exit(); |
---|
1809 | |
---|
1810 | /* Bump up our uz_count so we get here less */ |
---|
1811 | if (zone->uz_count < BUCKET_MAX) |
---|
1812 | zone->uz_count++; |
---|
1813 | |
---|
1814 | /* |
---|
1815 | * Now lets just fill a bucket and put it on the free list. If that |
---|
1816 | * works we'll restart the allocation from the begining. |
---|
1817 | */ |
---|
1818 | if (zone_alloc_bucket(zone, flags)) { |
---|
1819 | ZONE_UNLOCK(zone); |
---|
1820 | goto zalloc_restart; |
---|
1821 | } |
---|
1822 | ZONE_UNLOCK(zone); |
---|
1823 | /* |
---|
1824 | * We may not be able to get a bucket so return an actual item. |
---|
1825 | */ |
---|
1826 | #ifdef UMA_DEBUG |
---|
1827 | printf("uma_zalloc_arg: Bucketzone returned NULL\n"); |
---|
1828 | #endif |
---|
1829 | |
---|
1830 | item = zone_alloc_item(zone, udata, flags); |
---|
1831 | return (item); |
---|
1832 | } |
---|
1833 | |
---|
1834 | static uma_slab_t |
---|
1835 | keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags) |
---|
1836 | { |
---|
1837 | uma_slab_t slab; |
---|
1838 | |
---|
1839 | mtx_assert(&keg->uk_lock, MA_OWNED); |
---|
1840 | slab = NULL; |
---|
1841 | |
---|
1842 | for (;;) { |
---|
1843 | /* |
---|
1844 | * Find a slab with some space. Prefer slabs that are partially |
---|
1845 | * used over those that are totally full. This helps to reduce |
---|
1846 | * fragmentation. |
---|
1847 | */ |
---|
1848 | if (keg->uk_free != 0) { |
---|
1849 | if (!LIST_EMPTY(&keg->uk_part_slab)) { |
---|
1850 | slab = LIST_FIRST(&keg->uk_part_slab); |
---|
1851 | } else { |
---|
1852 | slab = LIST_FIRST(&keg->uk_free_slab); |
---|
1853 | LIST_REMOVE(slab, us_link); |
---|
1854 | LIST_INSERT_HEAD(&keg->uk_part_slab, slab, |
---|
1855 | us_link); |
---|
1856 | } |
---|
1857 | MPASS(slab->us_keg == keg); |
---|
1858 | return (slab); |
---|
1859 | } |
---|
1860 | |
---|
1861 | /* |
---|
1862 | * M_NOVM means don't ask at all! |
---|
1863 | */ |
---|
1864 | if (flags & M_NOVM) |
---|
1865 | break; |
---|
1866 | |
---|
1867 | if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { |
---|
1868 | keg->uk_flags |= UMA_ZFLAG_FULL; |
---|
1869 | /* |
---|
1870 | * If this is not a multi-zone, set the FULL bit. |
---|
1871 | * Otherwise slab_multi() takes care of it. |
---|
1872 | */ |
---|
1873 | if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) |
---|
1874 | zone->uz_flags |= UMA_ZFLAG_FULL; |
---|
1875 | if (flags & M_NOWAIT) |
---|
1876 | break; |
---|
1877 | msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); |
---|
1878 | continue; |
---|
1879 | } |
---|
1880 | keg->uk_recurse++; |
---|
1881 | slab = keg_alloc_slab(keg, zone, flags); |
---|
1882 | keg->uk_recurse--; |
---|
1883 | /* |
---|
1884 | * If we got a slab here it's safe to mark it partially used |
---|
1885 | * and return. We assume that the caller is going to remove |
---|
1886 | * at least one item. |
---|
1887 | */ |
---|
1888 | if (slab) { |
---|
1889 | MPASS(slab->us_keg == keg); |
---|
1890 | LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); |
---|
1891 | return (slab); |
---|
1892 | } |
---|
1893 | /* |
---|
1894 | * We might not have been able to get a slab but another cpu |
---|
1895 | * could have while we were unlocked. Check again before we |
---|
1896 | * fail. |
---|
1897 | */ |
---|
1898 | flags |= M_NOVM; |
---|
1899 | } |
---|
1900 | return (slab); |
---|
1901 | } |
---|
1902 | |
---|
1903 | static inline void |
---|
1904 | zone_relock(uma_zone_t zone, uma_keg_t keg) |
---|
1905 | { |
---|
1906 | if (zone->uz_lock != &keg->uk_lock) { |
---|
1907 | KEG_UNLOCK(keg); |
---|
1908 | ZONE_LOCK(zone); |
---|
1909 | } |
---|
1910 | } |
---|
1911 | |
---|
1912 | static inline void |
---|
1913 | keg_relock(uma_keg_t keg, uma_zone_t zone) |
---|
1914 | { |
---|
1915 | if (zone->uz_lock != &keg->uk_lock) { |
---|
1916 | ZONE_UNLOCK(zone); |
---|
1917 | KEG_LOCK(keg); |
---|
1918 | } |
---|
1919 | } |
---|
1920 | |
---|
1921 | static uma_slab_t |
---|
1922 | zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags) |
---|
1923 | { |
---|
1924 | uma_slab_t slab; |
---|
1925 | |
---|
1926 | if (keg == NULL) |
---|
1927 | keg = zone_first_keg(zone); |
---|
1928 | /* |
---|
1929 | * This is to prevent us from recursively trying to allocate |
---|
1930 | * buckets. The problem is that if an allocation forces us to |
---|
1931 | * grab a new bucket we will call page_alloc, which will go off |
---|
1932 | * and cause the vm to allocate vm_map_entries. If we need new |
---|
1933 | * buckets there too we will recurse in kmem_alloc and bad |
---|
1934 | * things happen. So instead we return a NULL bucket, and make |
---|
1935 | * the code that allocates buckets smart enough to deal with it |
---|
1936 | */ |
---|
1937 | if (keg->uk_flags & UMA_ZFLAG_BUCKET && keg->uk_recurse != 0) |
---|
1938 | return (NULL); |
---|
1939 | |
---|
1940 | for (;;) { |
---|
1941 | slab = keg_fetch_slab(keg, zone, flags); |
---|
1942 | if (slab) |
---|
1943 | return (slab); |
---|
1944 | if (flags & (M_NOWAIT | M_NOVM)) |
---|
1945 | break; |
---|
1946 | } |
---|
1947 | return (NULL); |
---|
1948 | } |
---|
1949 | |
---|
1950 | /* |
---|
1951 | * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns |
---|
1952 | * with the keg locked. Caller must call zone_relock() afterwards if the |
---|
1953 | * zone lock is required. On NULL the zone lock is held. |
---|
1954 | * |
---|
1955 | * The last pointer is used to seed the search. It is not required. |
---|
1956 | */ |
---|
1957 | static uma_slab_t |
---|
1958 | zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags) |
---|
1959 | { |
---|
1960 | uma_klink_t klink; |
---|
1961 | uma_slab_t slab; |
---|
1962 | uma_keg_t keg; |
---|
1963 | int flags; |
---|
1964 | int empty; |
---|
1965 | int full; |
---|
1966 | |
---|
1967 | /* |
---|
1968 | * Don't wait on the first pass. This will skip limit tests |
---|
1969 | * as well. We don't want to block if we can find a provider |
---|
1970 | * without blocking. |
---|
1971 | */ |
---|
1972 | flags = (rflags & ~M_WAITOK) | M_NOWAIT; |
---|
1973 | /* |
---|
1974 | * Use the last slab allocated as a hint for where to start |
---|
1975 | * the search. |
---|
1976 | */ |
---|
1977 | if (last) { |
---|
1978 | slab = keg_fetch_slab(last, zone, flags); |
---|
1979 | if (slab) |
---|
1980 | return (slab); |
---|
1981 | zone_relock(zone, last); |
---|
1982 | last = NULL; |
---|
1983 | } |
---|
1984 | /* |
---|
1985 | * Loop until we have a slab incase of transient failures |
---|
1986 | * while M_WAITOK is specified. I'm not sure this is 100% |
---|
1987 | * required but we've done it for so long now. |
---|
1988 | */ |
---|
1989 | for (;;) { |
---|
1990 | empty = 0; |
---|
1991 | full = 0; |
---|
1992 | /* |
---|
1993 | * Search the available kegs for slabs. Be careful to hold the |
---|
1994 | * correct lock while calling into the keg layer. |
---|
1995 | */ |
---|
1996 | LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { |
---|
1997 | keg = klink->kl_keg; |
---|
1998 | keg_relock(keg, zone); |
---|
1999 | if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { |
---|
2000 | slab = keg_fetch_slab(keg, zone, flags); |
---|
2001 | if (slab) |
---|
2002 | return (slab); |
---|
2003 | } |
---|
2004 | if (keg->uk_flags & UMA_ZFLAG_FULL) |
---|
2005 | full++; |
---|
2006 | else |
---|
2007 | empty++; |
---|
2008 | zone_relock(zone, keg); |
---|
2009 | } |
---|
2010 | if (rflags & (M_NOWAIT | M_NOVM)) |
---|
2011 | break; |
---|
2012 | flags = rflags; |
---|
2013 | /* |
---|
2014 | * All kegs are full. XXX We can't atomically check all kegs |
---|
2015 | * and sleep so just sleep for a short period and retry. |
---|
2016 | */ |
---|
2017 | if (full && !empty) { |
---|
2018 | zone->uz_flags |= UMA_ZFLAG_FULL; |
---|
2019 | msleep(zone, zone->uz_lock, PVM, "zonelimit", hz/100); |
---|
2020 | zone->uz_flags &= ~UMA_ZFLAG_FULL; |
---|
2021 | continue; |
---|
2022 | } |
---|
2023 | } |
---|
2024 | return (NULL); |
---|
2025 | } |
---|
2026 | |
---|
2027 | static void * |
---|
2028 | slab_alloc_item(uma_zone_t zone, uma_slab_t slab) |
---|
2029 | { |
---|
2030 | uma_keg_t keg; |
---|
2031 | uma_slabrefcnt_t slabref; |
---|
2032 | void *item; |
---|
2033 | u_int8_t freei; |
---|
2034 | |
---|
2035 | keg = slab->us_keg; |
---|
2036 | mtx_assert(&keg->uk_lock, MA_OWNED); |
---|
2037 | |
---|
2038 | freei = slab->us_firstfree; |
---|
2039 | if (keg->uk_flags & UMA_ZONE_REFCNT) { |
---|
2040 | slabref = (uma_slabrefcnt_t)slab; |
---|
2041 | slab->us_firstfree = slabref->us_freelist[freei].us_item; |
---|
2042 | } else { |
---|
2043 | slab->us_firstfree = slab->us_freelist[freei].us_item; |
---|
2044 | } |
---|
2045 | item = slab->us_data + (keg->uk_rsize * freei); |
---|
2046 | |
---|
2047 | slab->us_freecount--; |
---|
2048 | keg->uk_free--; |
---|
2049 | #ifdef INVARIANTS |
---|
2050 | uma_dbg_alloc(zone, slab, item); |
---|
2051 | #endif |
---|
2052 | /* Move this slab to the full list */ |
---|
2053 | if (slab->us_freecount == 0) { |
---|
2054 | LIST_REMOVE(slab, us_link); |
---|
2055 | LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link); |
---|
2056 | } |
---|
2057 | |
---|
2058 | return (item); |
---|
2059 | } |
---|
2060 | |
---|
2061 | static int |
---|
2062 | zone_alloc_bucket(uma_zone_t zone, int flags) |
---|
2063 | { |
---|
2064 | uma_bucket_t bucket; |
---|
2065 | uma_slab_t slab; |
---|
2066 | uma_keg_t keg; |
---|
2067 | int16_t saved; |
---|
2068 | int max, origflags = flags; |
---|
2069 | |
---|
2070 | /* |
---|
2071 | * Try this zone's free list first so we don't allocate extra buckets. |
---|
2072 | */ |
---|
2073 | if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) { |
---|
2074 | KASSERT(bucket->ub_cnt == 0, |
---|
2075 | ("zone_alloc_bucket: Bucket on free list is not empty.")); |
---|
2076 | LIST_REMOVE(bucket, ub_link); |
---|
2077 | } else { |
---|
2078 | int bflags; |
---|
2079 | |
---|
2080 | bflags = (flags & ~M_ZERO); |
---|
2081 | if (zone->uz_flags & UMA_ZFLAG_CACHEONLY) |
---|
2082 | bflags |= M_NOVM; |
---|
2083 | |
---|
2084 | ZONE_UNLOCK(zone); |
---|
2085 | bucket = bucket_alloc(zone->uz_count, bflags); |
---|
2086 | ZONE_LOCK(zone); |
---|
2087 | } |
---|
2088 | |
---|
2089 | if (bucket == NULL) { |
---|
2090 | return (0); |
---|
2091 | } |
---|
2092 | |
---|
2093 | #ifdef SMP |
---|
2094 | /* |
---|
2095 | * This code is here to limit the number of simultaneous bucket fills |
---|
2096 | * for any given zone to the number of per cpu caches in this zone. This |
---|
2097 | * is done so that we don't allocate more memory than we really need. |
---|
2098 | */ |
---|
2099 | if (zone->uz_fills >= mp_ncpus) |
---|
2100 | goto done; |
---|
2101 | |
---|
2102 | #endif |
---|
2103 | zone->uz_fills++; |
---|
2104 | |
---|
2105 | max = MIN(bucket->ub_entries, zone->uz_count); |
---|
2106 | /* Try to keep the buckets totally full */ |
---|
2107 | saved = bucket->ub_cnt; |
---|
2108 | slab = NULL; |
---|
2109 | keg = NULL; |
---|
2110 | while (bucket->ub_cnt < max && |
---|
2111 | (slab = zone->uz_slab(zone, keg, flags)) != NULL) { |
---|
2112 | keg = slab->us_keg; |
---|
2113 | while (slab->us_freecount && bucket->ub_cnt < max) { |
---|
2114 | bucket->ub_bucket[bucket->ub_cnt++] = |
---|
2115 | slab_alloc_item(zone, slab); |
---|
2116 | } |
---|
2117 | |
---|
2118 | /* Don't block on the next fill */ |
---|
2119 | flags |= M_NOWAIT; |
---|
2120 | } |
---|
2121 | if (slab) |
---|
2122 | zone_relock(zone, keg); |
---|
2123 | |
---|
2124 | /* |
---|
2125 | * We unlock here because we need to call the zone's init. |
---|
2126 | * It should be safe to unlock because the slab dealt with |
---|
2127 | * above is already on the appropriate list within the keg |
---|
2128 | * and the bucket we filled is not yet on any list, so we |
---|
2129 | * own it. |
---|
2130 | */ |
---|
2131 | if (zone->uz_init != NULL) { |
---|
2132 | int i; |
---|
2133 | |
---|
2134 | ZONE_UNLOCK(zone); |
---|
2135 | for (i = saved; i < bucket->ub_cnt; i++) |
---|
2136 | if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, |
---|
2137 | origflags) != 0) |
---|
2138 | break; |
---|
2139 | /* |
---|
2140 | * If we couldn't initialize the whole bucket, put the |
---|
2141 | * rest back onto the freelist. |
---|
2142 | */ |
---|
2143 | if (i != bucket->ub_cnt) { |
---|
2144 | int j; |
---|
2145 | |
---|
2146 | for (j = i; j < bucket->ub_cnt; j++) { |
---|
2147 | zone_free_item(zone, bucket->ub_bucket[j], |
---|
2148 | NULL, SKIP_FINI, 0); |
---|
2149 | #ifdef INVARIANTS |
---|
2150 | bucket->ub_bucket[j] = NULL; |
---|
2151 | #endif |
---|
2152 | } |
---|
2153 | bucket->ub_cnt = i; |
---|
2154 | } |
---|
2155 | ZONE_LOCK(zone); |
---|
2156 | } |
---|
2157 | |
---|
2158 | zone->uz_fills--; |
---|
2159 | if (bucket->ub_cnt != 0) { |
---|
2160 | LIST_INSERT_HEAD(&zone->uz_full_bucket, |
---|
2161 | bucket, ub_link); |
---|
2162 | return (1); |
---|
2163 | } |
---|
2164 | #ifdef SMP |
---|
2165 | done: |
---|
2166 | #endif |
---|
2167 | bucket_free(bucket); |
---|
2168 | |
---|
2169 | return (0); |
---|
2170 | } |
---|
2171 | /* |
---|
2172 | * Allocates an item for an internal zone |
---|
2173 | * |
---|
2174 | * Arguments |
---|
2175 | * zone The zone to alloc for. |
---|
2176 | * udata The data to be passed to the constructor. |
---|
2177 | * flags M_WAITOK, M_NOWAIT, M_ZERO. |
---|
2178 | * |
---|
2179 | * Returns |
---|
2180 | * NULL if there is no memory and M_NOWAIT is set |
---|
2181 | * An item if successful |
---|
2182 | */ |
---|
2183 | |
---|
2184 | static void * |
---|
2185 | zone_alloc_item(uma_zone_t zone, void *udata, int flags) |
---|
2186 | { |
---|
2187 | uma_slab_t slab; |
---|
2188 | void *item; |
---|
2189 | |
---|
2190 | item = NULL; |
---|
2191 | |
---|
2192 | #ifdef UMA_DEBUG_ALLOC |
---|
2193 | printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone); |
---|
2194 | #endif |
---|
2195 | ZONE_LOCK(zone); |
---|
2196 | |
---|
2197 | slab = zone->uz_slab(zone, NULL, flags); |
---|
2198 | if (slab == NULL) { |
---|
2199 | zone->uz_fails++; |
---|
2200 | ZONE_UNLOCK(zone); |
---|
2201 | return (NULL); |
---|
2202 | } |
---|
2203 | |
---|
2204 | item = slab_alloc_item(zone, slab); |
---|
2205 | |
---|
2206 | zone_relock(zone, slab->us_keg); |
---|
2207 | zone->uz_allocs++; |
---|
2208 | ZONE_UNLOCK(zone); |
---|
2209 | |
---|
2210 | /* |
---|
2211 | * We have to call both the zone's init (not the keg's init) |
---|
2212 | * and the zone's ctor. This is because the item is going from |
---|
2213 | * a keg slab directly to the user, and the user is expecting it |
---|
2214 | * to be both zone-init'd as well as zone-ctor'd. |
---|
2215 | */ |
---|
2216 | if (zone->uz_init != NULL) { |
---|
2217 | if (zone->uz_init(item, zone->uz_size, flags) != 0) { |
---|
2218 | zone_free_item(zone, item, udata, SKIP_FINI, |
---|
2219 | ZFREE_STATFAIL | ZFREE_STATFREE); |
---|
2220 | return (NULL); |
---|
2221 | } |
---|
2222 | } |
---|
2223 | if (zone->uz_ctor != NULL) { |
---|
2224 | if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { |
---|
2225 | zone_free_item(zone, item, udata, SKIP_DTOR, |
---|
2226 | ZFREE_STATFAIL | ZFREE_STATFREE); |
---|
2227 | return (NULL); |
---|
2228 | } |
---|
2229 | } |
---|
2230 | if (flags & M_ZERO) |
---|
2231 | bzero(item, zone->uz_size); |
---|
2232 | |
---|
2233 | return (item); |
---|
2234 | } |
---|
2235 | |
---|
2236 | /* See uma.h */ |
---|
2237 | void |
---|
2238 | uma_zfree_arg(uma_zone_t zone, void *item, void *udata) |
---|
2239 | { |
---|
2240 | uma_cache_t cache; |
---|
2241 | uma_bucket_t bucket; |
---|
2242 | int bflags; |
---|
2243 | int cpu; |
---|
2244 | |
---|
2245 | #ifdef UMA_DEBUG_ALLOC_1 |
---|
2246 | printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone); |
---|
2247 | #endif |
---|
2248 | CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, |
---|
2249 | zone->uz_name); |
---|
2250 | |
---|
2251 | /* uma_zfree(..., NULL) does nothing, to match free(9). */ |
---|
2252 | if (item == NULL) |
---|
2253 | return; |
---|
2254 | |
---|
2255 | if (zone->uz_dtor) |
---|
2256 | zone->uz_dtor(item, zone->uz_size, udata); |
---|
2257 | |
---|
2258 | #ifdef INVARIANTS |
---|
2259 | ZONE_LOCK(zone); |
---|
2260 | if (zone->uz_flags & UMA_ZONE_MALLOC) |
---|
2261 | uma_dbg_free(zone, udata, item); |
---|
2262 | else |
---|
2263 | uma_dbg_free(zone, NULL, item); |
---|
2264 | ZONE_UNLOCK(zone); |
---|
2265 | #endif |
---|
2266 | /* |
---|
2267 | * The race here is acceptable. If we miss it we'll just have to wait |
---|
2268 | * a little longer for the limits to be reset. |
---|
2269 | */ |
---|
2270 | if (zone->uz_flags & UMA_ZFLAG_FULL) |
---|
2271 | goto zfree_internal; |
---|
2272 | |
---|
2273 | /* |
---|
2274 | * If possible, free to the per-CPU cache. There are two |
---|
2275 | * requirements for safe access to the per-CPU cache: (1) the thread |
---|
2276 | * accessing the cache must not be preempted or yield during access, |
---|
2277 | * and (2) the thread must not migrate CPUs without switching which |
---|
2278 | * cache it accesses. We rely on a critical section to prevent |
---|
2279 | * preemption and migration. We release the critical section in |
---|
2280 | * order to acquire the zone mutex if we are unable to free to the |
---|
2281 | * current cache; when we re-acquire the critical section, we must |
---|
2282 | * detect and handle migration if it has occurred. |
---|
2283 | */ |
---|
2284 | zfree_restart: |
---|
2285 | critical_enter(); |
---|
2286 | cpu = curcpu; |
---|
2287 | cache = &zone->uz_cpu[cpu]; |
---|
2288 | |
---|
2289 | zfree_start: |
---|
2290 | bucket = cache->uc_freebucket; |
---|
2291 | |
---|
2292 | if (bucket) { |
---|
2293 | /* |
---|
2294 | * Do we have room in our bucket? It is OK for this uz count |
---|
2295 | * check to be slightly out of sync. |
---|
2296 | */ |
---|
2297 | |
---|
2298 | if (bucket->ub_cnt < bucket->ub_entries) { |
---|
2299 | KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, |
---|
2300 | ("uma_zfree: Freeing to non free bucket index.")); |
---|
2301 | bucket->ub_bucket[bucket->ub_cnt] = item; |
---|
2302 | bucket->ub_cnt++; |
---|
2303 | cache->uc_frees++; |
---|
2304 | critical_exit(); |
---|
2305 | return; |
---|
2306 | } else if (cache->uc_allocbucket) { |
---|
2307 | #ifdef UMA_DEBUG_ALLOC |
---|
2308 | printf("uma_zfree: Swapping buckets.\n"); |
---|
2309 | #endif |
---|
2310 | /* |
---|
2311 | * We have run out of space in our freebucket. |
---|
2312 | * See if we can switch with our alloc bucket. |
---|
2313 | */ |
---|
2314 | if (cache->uc_allocbucket->ub_cnt < |
---|
2315 | cache->uc_freebucket->ub_cnt) { |
---|
2316 | bucket = cache->uc_freebucket; |
---|
2317 | cache->uc_freebucket = cache->uc_allocbucket; |
---|
2318 | cache->uc_allocbucket = bucket; |
---|
2319 | goto zfree_start; |
---|
2320 | } |
---|
2321 | } |
---|
2322 | } |
---|
2323 | /* |
---|
2324 | * We can get here for two reasons: |
---|
2325 | * |
---|
2326 | * 1) The buckets are NULL |
---|
2327 | * 2) The alloc and free buckets are both somewhat full. |
---|
2328 | * |
---|
2329 | * We must go back the zone, which requires acquiring the zone lock, |
---|
2330 | * which in turn means we must release and re-acquire the critical |
---|
2331 | * section. Since the critical section is released, we may be |
---|
2332 | * preempted or migrate. As such, make sure not to maintain any |
---|
2333 | * thread-local state specific to the cache from prior to releasing |
---|
2334 | * the critical section. |
---|
2335 | */ |
---|
2336 | critical_exit(); |
---|
2337 | ZONE_LOCK(zone); |
---|
2338 | critical_enter(); |
---|
2339 | cpu = curcpu; |
---|
2340 | cache = &zone->uz_cpu[cpu]; |
---|
2341 | if (cache->uc_freebucket != NULL) { |
---|
2342 | if (cache->uc_freebucket->ub_cnt < |
---|
2343 | cache->uc_freebucket->ub_entries) { |
---|
2344 | ZONE_UNLOCK(zone); |
---|
2345 | goto zfree_start; |
---|
2346 | } |
---|
2347 | if (cache->uc_allocbucket != NULL && |
---|
2348 | (cache->uc_allocbucket->ub_cnt < |
---|
2349 | cache->uc_freebucket->ub_cnt)) { |
---|
2350 | ZONE_UNLOCK(zone); |
---|
2351 | goto zfree_start; |
---|
2352 | } |
---|
2353 | } |
---|
2354 | |
---|
2355 | /* Since we have locked the zone we may as well send back our stats */ |
---|
2356 | zone->uz_allocs += cache->uc_allocs; |
---|
2357 | cache->uc_allocs = 0; |
---|
2358 | zone->uz_frees += cache->uc_frees; |
---|
2359 | cache->uc_frees = 0; |
---|
2360 | |
---|
2361 | bucket = cache->uc_freebucket; |
---|
2362 | cache->uc_freebucket = NULL; |
---|
2363 | |
---|
2364 | /* Can we throw this on the zone full list? */ |
---|
2365 | if (bucket != NULL) { |
---|
2366 | #ifdef UMA_DEBUG_ALLOC |
---|
2367 | printf("uma_zfree: Putting old bucket on the free list.\n"); |
---|
2368 | #endif |
---|
2369 | /* ub_cnt is pointing to the last free item */ |
---|
2370 | KASSERT(bucket->ub_cnt != 0, |
---|
2371 | ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); |
---|
2372 | LIST_INSERT_HEAD(&zone->uz_full_bucket, |
---|
2373 | bucket, ub_link); |
---|
2374 | } |
---|
2375 | if ((bucket = LIST_FIRST(&zone->uz_free_bucket)) != NULL) { |
---|
2376 | LIST_REMOVE(bucket, ub_link); |
---|
2377 | ZONE_UNLOCK(zone); |
---|
2378 | cache->uc_freebucket = bucket; |
---|
2379 | goto zfree_start; |
---|
2380 | } |
---|
2381 | /* We are no longer associated with this CPU. */ |
---|
2382 | critical_exit(); |
---|
2383 | |
---|
2384 | /* And the zone.. */ |
---|
2385 | ZONE_UNLOCK(zone); |
---|
2386 | |
---|
2387 | #ifdef UMA_DEBUG_ALLOC |
---|
2388 | printf("uma_zfree: Allocating new free bucket.\n"); |
---|
2389 | #endif |
---|
2390 | bflags = M_NOWAIT; |
---|
2391 | |
---|
2392 | if (zone->uz_flags & UMA_ZFLAG_CACHEONLY) |
---|
2393 | bflags |= M_NOVM; |
---|
2394 | bucket = bucket_alloc(zone->uz_count, bflags); |
---|
2395 | if (bucket) { |
---|
2396 | ZONE_LOCK(zone); |
---|
2397 | LIST_INSERT_HEAD(&zone->uz_free_bucket, |
---|
2398 | bucket, ub_link); |
---|
2399 | ZONE_UNLOCK(zone); |
---|
2400 | goto zfree_restart; |
---|
2401 | } |
---|
2402 | |
---|
2403 | /* |
---|
2404 | * If nothing else caught this, we'll just do an internal free. |
---|
2405 | */ |
---|
2406 | zfree_internal: |
---|
2407 | zone_free_item(zone, item, udata, SKIP_DTOR, ZFREE_STATFREE); |
---|
2408 | |
---|
2409 | return; |
---|
2410 | } |
---|
2411 | |
---|
2412 | /* |
---|
2413 | * Frees an item to an INTERNAL zone or allocates a free bucket |
---|
2414 | * |
---|
2415 | * Arguments: |
---|
2416 | * zone The zone to free to |
---|
2417 | * item The item we're freeing |
---|
2418 | * udata User supplied data for the dtor |
---|
2419 | * skip Skip dtors and finis |
---|
2420 | */ |
---|
2421 | static void |
---|
2422 | zone_free_item(uma_zone_t zone, void *item, void *udata, |
---|
2423 | enum zfreeskip skip, int flags) |
---|
2424 | { |
---|
2425 | uma_slab_t slab; |
---|
2426 | uma_slabrefcnt_t slabref; |
---|
2427 | uma_keg_t keg; |
---|
2428 | u_int8_t *mem; |
---|
2429 | u_int8_t freei; |
---|
2430 | int clearfull; |
---|
2431 | |
---|
2432 | if (skip < SKIP_DTOR && zone->uz_dtor) |
---|
2433 | zone->uz_dtor(item, zone->uz_size, udata); |
---|
2434 | |
---|
2435 | if (skip < SKIP_FINI && zone->uz_fini) |
---|
2436 | zone->uz_fini(item, zone->uz_size); |
---|
2437 | |
---|
2438 | ZONE_LOCK(zone); |
---|
2439 | |
---|
2440 | if (flags & ZFREE_STATFAIL) |
---|
2441 | zone->uz_fails++; |
---|
2442 | if (flags & ZFREE_STATFREE) |
---|
2443 | zone->uz_frees++; |
---|
2444 | |
---|
2445 | if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { |
---|
2446 | mem = (u_int8_t *)((unsigned long)item & (~UMA_SLAB_MASK)); |
---|
2447 | keg = zone_first_keg(zone); /* Must only be one. */ |
---|
2448 | if (zone->uz_flags & UMA_ZONE_HASH) { |
---|
2449 | slab = hash_sfind(&keg->uk_hash, mem); |
---|
2450 | } else { |
---|
2451 | mem += keg->uk_pgoff; |
---|
2452 | slab = (uma_slab_t)mem; |
---|
2453 | } |
---|
2454 | } else { |
---|
2455 | panic("uma virtual memory not supported!" ); |
---|
2456 | } |
---|
2457 | MPASS(keg == slab->us_keg); |
---|
2458 | |
---|
2459 | /* Do we need to remove from any lists? */ |
---|
2460 | if (slab->us_freecount+1 == keg->uk_ipers) { |
---|
2461 | LIST_REMOVE(slab, us_link); |
---|
2462 | LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); |
---|
2463 | } else if (slab->us_freecount == 0) { |
---|
2464 | LIST_REMOVE(slab, us_link); |
---|
2465 | LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); |
---|
2466 | } |
---|
2467 | |
---|
2468 | /* Slab management stuff */ |
---|
2469 | freei = ((unsigned long)item - (unsigned long)slab->us_data) |
---|
2470 | / keg->uk_rsize; |
---|
2471 | |
---|
2472 | #ifdef INVARIANTS |
---|
2473 | if (!skip) |
---|
2474 | uma_dbg_free(zone, slab, item); |
---|
2475 | #endif |
---|
2476 | |
---|
2477 | if (keg->uk_flags & UMA_ZONE_REFCNT) { |
---|
2478 | slabref = (uma_slabrefcnt_t)slab; |
---|
2479 | slabref->us_freelist[freei].us_item = slab->us_firstfree; |
---|
2480 | } else { |
---|
2481 | slab->us_freelist[freei].us_item = slab->us_firstfree; |
---|
2482 | } |
---|
2483 | slab->us_firstfree = freei; |
---|
2484 | slab->us_freecount++; |
---|
2485 | |
---|
2486 | /* Zone statistics */ |
---|
2487 | keg->uk_free++; |
---|
2488 | |
---|
2489 | clearfull = 0; |
---|
2490 | if (keg->uk_flags & UMA_ZFLAG_FULL) { |
---|
2491 | if (keg->uk_pages < keg->uk_maxpages) { |
---|
2492 | keg->uk_flags &= ~UMA_ZFLAG_FULL; |
---|
2493 | clearfull = 1; |
---|
2494 | } |
---|
2495 | |
---|
2496 | /* |
---|
2497 | * We can handle one more allocation. Since we're clearing ZFLAG_FULL, |
---|
2498 | * wake up all procs blocked on pages. This should be uncommon, so |
---|
2499 | * keeping this simple for now (rather than adding count of blocked |
---|
2500 | * threads etc). |
---|
2501 | */ |
---|
2502 | wakeup(keg); |
---|
2503 | } |
---|
2504 | if (clearfull) { |
---|
2505 | zone_relock(zone, keg); |
---|
2506 | zone->uz_flags &= ~UMA_ZFLAG_FULL; |
---|
2507 | wakeup(zone); |
---|
2508 | ZONE_UNLOCK(zone); |
---|
2509 | } else |
---|
2510 | KEG_UNLOCK(keg); |
---|
2511 | } |
---|
2512 | |
---|
2513 | /* See uma.h */ |
---|
2514 | void |
---|
2515 | uma_zone_set_max(uma_zone_t zone, int nitems) |
---|
2516 | { |
---|
2517 | uma_keg_t keg; |
---|
2518 | |
---|
2519 | ZONE_LOCK(zone); |
---|
2520 | keg = zone_first_keg(zone); |
---|
2521 | keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; |
---|
2522 | if (keg->uk_maxpages * keg->uk_ipers < nitems) |
---|
2523 | keg->uk_maxpages += keg->uk_ppera; |
---|
2524 | |
---|
2525 | ZONE_UNLOCK(zone); |
---|
2526 | } |
---|
2527 | |
---|
2528 | /* See uma.h */ |
---|
2529 | int |
---|
2530 | uma_zone_get_max(uma_zone_t zone) |
---|
2531 | { |
---|
2532 | int nitems; |
---|
2533 | uma_keg_t keg; |
---|
2534 | |
---|
2535 | ZONE_LOCK(zone); |
---|
2536 | keg = zone_first_keg(zone); |
---|
2537 | nitems = keg->uk_maxpages * keg->uk_ipers; |
---|
2538 | ZONE_UNLOCK(zone); |
---|
2539 | |
---|
2540 | return (nitems); |
---|
2541 | } |
---|
2542 | |
---|
2543 | /* See uma.h */ |
---|
2544 | int |
---|
2545 | uma_zone_get_cur(uma_zone_t zone) |
---|
2546 | { |
---|
2547 | int64_t nitems; |
---|
2548 | u_int i; |
---|
2549 | |
---|
2550 | ZONE_LOCK(zone); |
---|
2551 | nitems = zone->uz_allocs - zone->uz_frees; |
---|
2552 | CPU_FOREACH(i) { |
---|
2553 | /* |
---|
2554 | * See the comment in sysctl_vm_zone_stats() regarding the |
---|
2555 | * safety of accessing the per-cpu caches. With the zone lock |
---|
2556 | * held, it is safe, but can potentially result in stale data. |
---|
2557 | */ |
---|
2558 | nitems += zone->uz_cpu[i].uc_allocs - |
---|
2559 | zone->uz_cpu[i].uc_frees; |
---|
2560 | } |
---|
2561 | ZONE_UNLOCK(zone); |
---|
2562 | |
---|
2563 | return (nitems < 0 ? 0 : nitems); |
---|
2564 | } |
---|
2565 | |
---|
2566 | /* See uma.h */ |
---|
2567 | void |
---|
2568 | uma_zone_set_init(uma_zone_t zone, uma_init uminit) |
---|
2569 | { |
---|
2570 | uma_keg_t keg; |
---|
2571 | |
---|
2572 | ZONE_LOCK(zone); |
---|
2573 | keg = zone_first_keg(zone); |
---|
2574 | KASSERT(keg->uk_pages == 0, |
---|
2575 | ("uma_zone_set_init on non-empty keg")); |
---|
2576 | keg->uk_init = uminit; |
---|
2577 | ZONE_UNLOCK(zone); |
---|
2578 | } |
---|
2579 | |
---|
2580 | /* See uma.h */ |
---|
2581 | void |
---|
2582 | uma_zone_set_fini(uma_zone_t zone, uma_fini fini) |
---|
2583 | { |
---|
2584 | uma_keg_t keg; |
---|
2585 | |
---|
2586 | ZONE_LOCK(zone); |
---|
2587 | keg = zone_first_keg(zone); |
---|
2588 | KASSERT(keg->uk_pages == 0, |
---|
2589 | ("uma_zone_set_fini on non-empty keg")); |
---|
2590 | keg->uk_fini = fini; |
---|
2591 | ZONE_UNLOCK(zone); |
---|
2592 | } |
---|
2593 | |
---|
2594 | /* See uma.h */ |
---|
2595 | void |
---|
2596 | uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) |
---|
2597 | { |
---|
2598 | ZONE_LOCK(zone); |
---|
2599 | KASSERT(zone_first_keg(zone)->uk_pages == 0, |
---|
2600 | ("uma_zone_set_zinit on non-empty keg")); |
---|
2601 | zone->uz_init = zinit; |
---|
2602 | ZONE_UNLOCK(zone); |
---|
2603 | } |
---|
2604 | |
---|
2605 | /* See uma.h */ |
---|
2606 | void |
---|
2607 | uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) |
---|
2608 | { |
---|
2609 | ZONE_LOCK(zone); |
---|
2610 | KASSERT(zone_first_keg(zone)->uk_pages == 0, |
---|
2611 | ("uma_zone_set_zfini on non-empty keg")); |
---|
2612 | zone->uz_fini = zfini; |
---|
2613 | ZONE_UNLOCK(zone); |
---|
2614 | } |
---|
2615 | |
---|
2616 | /* See uma.h */ |
---|
2617 | /* XXX uk_freef is not actually used with the zone locked */ |
---|
2618 | void |
---|
2619 | uma_zone_set_freef(uma_zone_t zone, uma_free freef) |
---|
2620 | { |
---|
2621 | |
---|
2622 | ZONE_LOCK(zone); |
---|
2623 | zone_first_keg(zone)->uk_freef = freef; |
---|
2624 | ZONE_UNLOCK(zone); |
---|
2625 | } |
---|
2626 | |
---|
2627 | /* See uma.h */ |
---|
2628 | /* XXX uk_allocf is not actually used with the zone locked */ |
---|
2629 | void |
---|
2630 | uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) |
---|
2631 | { |
---|
2632 | uma_keg_t keg; |
---|
2633 | |
---|
2634 | ZONE_LOCK(zone); |
---|
2635 | keg = zone_first_keg(zone); |
---|
2636 | keg->uk_flags |= UMA_ZFLAG_PRIVALLOC; |
---|
2637 | keg->uk_allocf = allocf; |
---|
2638 | ZONE_UNLOCK(zone); |
---|
2639 | } |
---|
2640 | |
---|
2641 | /* See uma.h */ |
---|
2642 | void |
---|
2643 | uma_prealloc(uma_zone_t zone, int items) |
---|
2644 | { |
---|
2645 | int slabs; |
---|
2646 | uma_slab_t slab; |
---|
2647 | uma_keg_t keg; |
---|
2648 | |
---|
2649 | keg = zone_first_keg(zone); |
---|
2650 | ZONE_LOCK(zone); |
---|
2651 | slabs = items / keg->uk_ipers; |
---|
2652 | if (slabs * keg->uk_ipers < items) |
---|
2653 | slabs++; |
---|
2654 | while (slabs > 0) { |
---|
2655 | slab = keg_alloc_slab(keg, zone, M_WAITOK); |
---|
2656 | if (slab == NULL) |
---|
2657 | break; |
---|
2658 | MPASS(slab->us_keg == keg); |
---|
2659 | LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); |
---|
2660 | slabs--; |
---|
2661 | } |
---|
2662 | ZONE_UNLOCK(zone); |
---|
2663 | } |
---|
2664 | |
---|
2665 | /* See uma.h */ |
---|
2666 | void |
---|
2667 | uma_reclaim(void) |
---|
2668 | { |
---|
2669 | #ifdef UMA_DEBUG |
---|
2670 | printf("UMA: vm asked us to release pages!\n"); |
---|
2671 | #endif |
---|
2672 | zone_foreach(zone_drain); |
---|
2673 | /* |
---|
2674 | * Some slabs may have been freed but this zone will be visited early |
---|
2675 | * we visit again so that we can free pages that are empty once other |
---|
2676 | * zones are drained. We have to do the same for buckets. |
---|
2677 | */ |
---|
2678 | zone_drain(slabzone); |
---|
2679 | zone_drain(slabrefzone); |
---|
2680 | bucket_zone_drain(); |
---|
2681 | } |
---|
2682 | |
---|
2683 | /* See uma.h */ |
---|
2684 | int |
---|
2685 | uma_zone_exhausted(uma_zone_t zone) |
---|
2686 | { |
---|
2687 | int full; |
---|
2688 | |
---|
2689 | ZONE_LOCK(zone); |
---|
2690 | full = (zone->uz_flags & UMA_ZFLAG_FULL); |
---|
2691 | ZONE_UNLOCK(zone); |
---|
2692 | return (full); |
---|
2693 | } |
---|
2694 | |
---|
2695 | int |
---|
2696 | uma_zone_exhausted_nolock(uma_zone_t zone) |
---|
2697 | { |
---|
2698 | return (zone->uz_flags & UMA_ZFLAG_FULL); |
---|
2699 | } |
---|
2700 | |
---|
2701 | void * |
---|
2702 | uma_large_malloc(int size, int wait) |
---|
2703 | { |
---|
2704 | void *mem; |
---|
2705 | uma_slab_t slab; |
---|
2706 | u_int8_t flags; |
---|
2707 | |
---|
2708 | slab = zone_alloc_item(slabzone, NULL, wait); |
---|
2709 | if (slab == NULL) |
---|
2710 | return (NULL); |
---|
2711 | mem = page_alloc(NULL, size, &flags, wait); |
---|
2712 | if (mem) { |
---|
2713 | slab->us_data = mem; |
---|
2714 | slab->us_flags = flags | UMA_SLAB_MALLOC; |
---|
2715 | slab->us_size = size; |
---|
2716 | } else { |
---|
2717 | zone_free_item(slabzone, slab, NULL, SKIP_NONE, |
---|
2718 | ZFREE_STATFAIL | ZFREE_STATFREE); |
---|
2719 | } |
---|
2720 | |
---|
2721 | return (mem); |
---|
2722 | } |
---|
2723 | |
---|
2724 | void |
---|
2725 | uma_large_free(uma_slab_t slab) |
---|
2726 | { |
---|
2727 | page_free(slab->us_data, slab->us_size, slab->us_flags); |
---|
2728 | zone_free_item(slabzone, slab, NULL, SKIP_NONE, ZFREE_STATFREE); |
---|
2729 | } |
---|
2730 | |
---|
2731 | void |
---|
2732 | uma_print_stats(void) |
---|
2733 | { |
---|
2734 | zone_foreach(uma_print_zone); |
---|
2735 | } |
---|
2736 | |
---|
2737 | static void |
---|
2738 | slab_print(uma_slab_t slab) |
---|
2739 | { |
---|
2740 | printf("slab: keg %p, data %p, freecount %d, firstfree %d\n", |
---|
2741 | slab->us_keg, slab->us_data, slab->us_freecount, |
---|
2742 | slab->us_firstfree); |
---|
2743 | } |
---|
2744 | |
---|
2745 | static void |
---|
2746 | cache_print(uma_cache_t cache) |
---|
2747 | { |
---|
2748 | printf("alloc: %p(%d), free: %p(%d)\n", |
---|
2749 | cache->uc_allocbucket, |
---|
2750 | cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, |
---|
2751 | cache->uc_freebucket, |
---|
2752 | cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); |
---|
2753 | } |
---|
2754 | |
---|
2755 | static void |
---|
2756 | uma_print_keg(uma_keg_t keg) |
---|
2757 | { |
---|
2758 | uma_slab_t slab; |
---|
2759 | |
---|
2760 | printf("keg: %s(%p) size %d(%d) flags %d ipers %d ppera %d " |
---|
2761 | "out %d free %d limit %d\n", |
---|
2762 | keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, |
---|
2763 | keg->uk_ipers, keg->uk_ppera, |
---|
2764 | (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free, |
---|
2765 | (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); |
---|
2766 | printf("Part slabs:\n"); |
---|
2767 | LIST_FOREACH(slab, &keg->uk_part_slab, us_link) |
---|
2768 | slab_print(slab); |
---|
2769 | printf("Free slabs:\n"); |
---|
2770 | LIST_FOREACH(slab, &keg->uk_free_slab, us_link) |
---|
2771 | slab_print(slab); |
---|
2772 | printf("Full slabs:\n"); |
---|
2773 | LIST_FOREACH(slab, &keg->uk_full_slab, us_link) |
---|
2774 | slab_print(slab); |
---|
2775 | } |
---|
2776 | |
---|
2777 | void |
---|
2778 | uma_print_zone(uma_zone_t zone) |
---|
2779 | { |
---|
2780 | uma_cache_t cache; |
---|
2781 | uma_klink_t kl; |
---|
2782 | int i; |
---|
2783 | |
---|
2784 | printf("zone: %s(%p) size %d flags %d\n", |
---|
2785 | zone->uz_name, zone, zone->uz_size, zone->uz_flags); |
---|
2786 | LIST_FOREACH(kl, &zone->uz_kegs, kl_link) |
---|
2787 | uma_print_keg(kl->kl_keg); |
---|
2788 | for (i = 0; i <= mp_maxid; i++) { |
---|
2789 | if (CPU_ABSENT(i)) |
---|
2790 | continue; |
---|
2791 | cache = &zone->uz_cpu[i]; |
---|
2792 | printf("CPU %d Cache:\n", i); |
---|
2793 | cache_print(cache); |
---|
2794 | } |
---|
2795 | } |
---|
2796 | |
---|