Memcached源码分析之slabs.c

1. #include "memcached.h"
2. #include <sys/stat.h>
3. #include <sys/socket.h>
4. #include <sys/signal.h>
5. #include <sys/resource.h>
6. #include <fcntl.h>
7. #include <netinet/in.h>
8. #include <errno.h>
9. #include <stdlib.h>
10. #include <stdio.h>
11. #include <string.h>
12. #include <assert.h>
13. #include <pthread.h>
14.  
15. typedef struct {
16. unsigned int size;      /* sizes of items */     //item或者说chunk的大小
17. unsigned int perslab;   /* how many items per slab */ //每个slab有多少个item，slab又称“页”
18.  
19. /**
20. 当前slabclass的空闲item链表，也是可用item链表，当前slabclass一切可以用的内存空间都在此，
21. 这里是内存分配的入口，分配内存的时候都是在这个链表上挤一个出去。
22.  
23. ps：memcached的新版本才开始把slots作为“所有空闲的item链接”的用途，以前的版本slots链表保存的是“回收的item”的意思，
24. 而旧版本新分配的slab，是用end_page_ptr指针及end_page_free来控制，此版本已不用。
25. */
26. void *slots;           /* list of item ptrs */
27. unsigned int sl_curr;   /* total free items in list */  //当前slabclass还剩多少空闲的item，即上面的slots数
28.  
29. unsigned int slabs;     /* how many slabs were allocated for this class */ //这个slabclass分配了多少个slab了
30.  
31. /**
32. slab_list是这个slabclass下的slabs列表，逻辑上是一个数组，每个元素是一个slab指针。
33. list_size是slab_list的元素个数。
34. 注意这个list_size和上面的slabs的不同：
35. 由于slab_list是一个空间大小固定的数组，是数组！而list_size是这个数组元素的个数，代表slab_list的空间大小。
36. slabs代表已经分配出去的slabs数，list_size则代表可以有多少个slabs数
37. 所以当slabs等于list_size的时候代表这个slab_list已经满了，得增大空间。
38.  
39. */
40. void **slab_list;       /* array of slab pointers */
41. unsigned int list_size; /* size of prev array */
42.  
43. unsigned int killing;  /* index+1 of dying slab, or zero if none */
44. size_t requested; /* The number of requested bytes */
45. } slabclass_t;
46.  
47. static slabclass_t slabclass[MAX_NUMBER_OF_SLAB_CLASSES];
48. static size_t mem_limit = 0; //内存上限
49. static size_t mem_malloced = 0; //已分配的内存
50. static int power_largest;
51.  
52. static void *mem_base = NULL; //预分配的内存空间
53. static void *mem_current = NULL;
54. static size_t mem_avail = 0;
55.  
56. static pthread_mutex_t slabs_lock = PTHREAD_MUTEX_INITIALIZER;
57. static pthread_mutex_t slabs_rebalance_lock = PTHREAD_MUTEX_INITIALIZER;
58.  
59. static int do_slabs_newslab(const unsigned int id);
60. static void *memory_allocate(size_t size);
61. static void do_slabs_free(void *ptr, const size_t size, unsigned int id);
62.  
63. static void slabs_preallocate (const unsigned int maxslabs);
64.  
65. //根据item大小找到合适的slabclass
66. unsigned int slabs_clsid(const size_t size) {
67. int res = POWER_SMALLEST;
68.  
69. if (size == 0)
70. return 0;
71. while (size > slabclass[res].size)
72. if (res++ == power_largest)     /* won't fit in the biggest slab */
73. return 0;
74. return res;
75. }
76.  
77. /**
78. 初始化slabs，这里会对一些内存管理进行初始化
79. */
80. void slabs_init(const size_t limit, const double factor, const bool prealloc) {
81. int i = POWER_SMALLEST - 1;
82. unsigned int size = sizeof(item) + settings.chunk_size;
83.  
84. mem_limit = limit; //这个limit就是启动时候用户设置的-m xx中的xx，最大的内存上限
85.  
86. if (prealloc) {
87. /**
88. 如果用户开启了预分配，则先把上限的内存先分配出来，放到mem_base全局变量中。
89. 所以这个时候服务就拥有了一大坨内存，以后要分配的内存都是从这一坨里面割下来。
90. */
91. mem_base = malloc(mem_limit);
92. if (mem_base != NULL) {
93. mem_current = mem_base;
94. mem_avail = mem_limit;
95. } else {
96. fprintf(stderr, "Warning: Failed to allocate requested memory in"
97. " one large chunk.\nWill allocate in smaller chunks\n");
98. }
99. }
100.  
101. //下面是初始化各个slabclass对象
102. memset(slabclass, 0, sizeof(slabclass));
103.  
104. while (++i < POWER_LARGEST && size <= settings.item_size_max / factor) {
105. /* Make sure items are always n-byte aligned */
106. if (size % CHUNK_ALIGN_BYTES)
107. size += CHUNK_ALIGN_BYTES - (size % CHUNK_ALIGN_BYTES);
108.  
109. slabclass[i].size = size;
110. slabclass[i].perslab = settings.item_size_max / slabclass[i].size;
111. size *= factor;
112. if (settings.verbose > 1) {
113. fprintf(stderr, "slab class %3d: chunk size %9u perslab %7u\n",
114. i, slabclass[i].size, slabclass[i].perslab);
115. }
116. }
117.  
118. power_largest = i;
119. slabclass[power_largest].size = settings.item_size_max;
120. slabclass[power_largest].perslab = 1;
121. if (settings.verbose > 1) {
122. fprintf(stderr, "slab class %3d: chunk size %9u perslab %7u\n",
123. i, slabclass[i].size, slabclass[i].perslab);
124. }
125.  
126. {
127. char *t_initial_malloc = getenv("T_MEMD_INITIAL_MALLOC");
128. if (t_initial_malloc) {
129. mem_malloced = (size_t)atol(t_initial_malloc);
130. }
131.  
132. }
133.  
134. if (prealloc) {
135. slabs_preallocate(power_largest);
136. }
137. }
138.  
139. /**
140. 内存预分配，如果用户开启了预分配，则会调用此方法，先从mem_base为分每个slabclass割一个slab大小下来。
141. */
142. static void slabs_preallocate (const unsigned int maxslabs) {
143. int i;
144. unsigned int prealloc = 0;
145.  
146. for (i = POWER_SMALLEST; i <= POWER_LARGEST; i++) {
147. if (++prealloc > maxslabs)
148. return;
149. if (do_slabs_newslab(i) == 0) {
150. fprintf(stderr, "Error while preallocating slab memory!\n"
151. "If using -L or other prealloc options, max memory must be "
152. "at least %d megabytes.\n", power_largest);
153. exit(1);
154. }
155. }
156.  
157. }
158.  
159. static int grow_slab_list (const unsigned int id) {
160. slabclass_t *p = &slabclass[id];
161. /**
162. p->slab_list是一个空间大小固定的数组，是数组！而list_size是这个数组分配的空间。
163. p->slabs代表已经分配出去的slabs数
164. 而p->list_size代表可以用多少个slabs数
165. 所以当slabs等于list_size的时候代表这个slab_list已经满了，得增大空间。
166. */
167. if (p->slabs == p->list_size) {
168. size_t new_size =  (p->list_size != 0) ? p->list_size * 2 : 16;
169. void *new_list = realloc(p->slab_list, new_size * sizeof(void *)); //
170. if (new_list == 0) return 0;
171. p->list_size = new_size;
172. p->slab_list = new_list;
173. }
174. return 1;
175. }
176.  
177. /**
178. 把整个slab打散成一个个（也叫chunk）放到相应的slots链表中
179. */
180. static void split_slab_page_into_freelist(char *ptr, const unsigned int id) {
181. slabclass_t *p = &slabclass[id];
182. int x;
183. for (x = 0; x < p->perslab; x++) {
184. do_slabs_free(ptr, 0, id); //这个函数主要作用是让当前item空间可用，即加到slots链表中。
185. ptr += p->size;
186. }
187. }
188.  
189. /**
190. 为slabclass[id]分配新的slab，仅当当前的slabclass中slots没有空闲的空间才调用
191. 此函数分配新的slab
192. */
193. static int do_slabs_newslab(const unsigned int id) {
194. slabclass_t *p = &slabclass[id];
195. int len = settings.slab_reassign ? settings.item_size_max
196. : p->size * p->perslab; //先判断是否开启了自定义slab大小，如果没有就按默认的，即约1M
197. char *ptr;
198.  
199. /**
200. 下面if的逻辑是：
201. 如果内存超出了限制，分配失败进入if，返回0
202. 否则调用grow_slab_list检查是否要增大slab_list的大小
203. 如果在grow_slab_list返回失败，则不继续分配空间，进入if，返回0
204. 否则分配空间memory_allocate，如果分配失败，同样进入if，返回0；
205. */
206. if ((mem_limit && mem_malloced + len > mem_limit && p->slabs > 0) ||
207. (grow_slab_list(id) == 0) ||
208. ((ptr = memory_allocate((size_t)len)) == 0)) {
209.  
210. MEMCACHED_SLABS_SLABCLASS_ALLOCATE_FAILED(id);
211. return 0;
212. }
213.  
214. memset(ptr, 0, (size_t)len); //清干净内存空间
215. split_slab_page_into_freelist(ptr, id); //把新申请的slab放到slots中去
216.  
217. p->slab_list[p->slabs++] = ptr; //把新的slab加到slab_list数组中
218. mem_malloced += len; //记下已分配的空间大小
219. MEMCACHED_SLABS_SLABCLASS_ALLOCATE(id);
220.  
221. return 1;
222. }
223.  
224. /**
225. 根据item大小和slabsclass分配空间
226. */
227. static void *do_slabs_alloc(const size_t size, unsigned int id) {
228. slabclass_t *p;
229. void *ret = NULL;
230. item *it = NULL;
231.  
232. if (id < POWER_SMALLEST || id > power_largest) { //默认最大是200，最小是1
233. MEMCACHED_SLABS_ALLOCATE_FAILED(size, 0);
234. return NULL;
235. }
236.  
237. p = &slabclass[id]; //slabclass是一个全局变量，是各个slabclass对象数组，在这取得当前id对应的slabclass
238. assert(p->sl_curr == 0 || ((item *)p->slots)->slabs_clsid == 0);
239.  
240. /* fail unless we have space at the end of a recently allocated page,
241. we have something on our freelist, or we could allocate a new page */
242. /**
243. 下面这个if的逻辑相当于：
244. 如果p->sl_curr==0，即slots链表中没有空闲的空间，则do_slabs_newslab分配新slab
245. 如果p->sl_curr==0，且do_slabs_newslab分配新slab失败，则进入if，ret = NULL，否则进入下面的elseif
246. */
247. if (! (p->sl_curr != 0 || do_slabs_newslab(id) != 0)) {
248. /* We don't have more memory available */
249. ret = NULL;
250. } else if (p->sl_curr != 0) { //如果进入此分支是因为slots链表中还有空闲的空间
251. /* return off our freelist */
252. //把空闲的item分配出去
253. it = (item *)p->slots;
254. p->slots = it->next;
255. if (it->next) it->next->prev = 0;
256. p->sl_curr--;
257. ret = (void *)it;
258. }
259.  
260. if (ret) {
261. p->requested += size; //分配成功，记下已分配的字节数
262. MEMCACHED_SLABS_ALLOCATE(size, id, p->size, ret);
263. } else {
264. MEMCACHED_SLABS_ALLOCATE_FAILED(size, id);
265. }
266.  
267. return ret;
268. }
269.  
270. /**
271. 这个函数的命名虽然叫do_slabs_free，听上去好像是释放空间，其实质是把空间变成可用。
272. 怎样的空间才算可用？就是加到当前slabclass的slots链表中而已。
273. 所以新申请的slab也会调用这个函数，让整个slab变为可用。
274.  
275. ps: 以前的memcached版本slots链表保存的是回收的item空间，而
276. 现在保存的是所有可用的item空间。
277. */
278. static void do_slabs_free(void *ptr, const size_t size, unsigned int id) {
279. slabclass_t *p;
280. item *it;
281.  
282. assert(((item *)ptr)->slabs_clsid == 0);
283. assert(id >= POWER_SMALLEST && id <= power_largest);
284. if (id < POWER_SMALLEST || id > power_largest)
285. return;
286.  
287. MEMCACHED_SLABS_FREE(size, id, ptr);
288. p = &slabclass[id];
289.  
290. it = (item *)ptr;
291. it->it_flags |= ITEM_SLABBED; //把item标记为slabbed状态
292. it->prev = 0;
293. it->next = p->slots;  //插入到slots链表中
294. if (it->next) it->next->prev = it;
295. p->slots = it;
296.  
297. p->sl_curr++; //空闲item数加1
298. p->requested -= size;
299. return;
300. }
301.  
302. static int nz_strcmp(int nzlength, const char *nz, const char *z) {
303. int zlength=strlen(z);
304. return (zlength == nzlength) && (strncmp(nz, z, zlength) == 0) ? 0 : -1;
305. }
306.  
307. bool get_stats(const char *stat_type, int nkey, ADD_STAT add_stats, void *c) {
308. bool ret = true;
309.  
310. if (add_stats != NULL) {
311. if (!stat_type) {
312. /* prepare general statistics for the engine */
313. STATS_LOCK();
314. APPEND_STAT("bytes", "%llu", (unsigned long long)stats.curr_bytes);
315. APPEND_STAT("curr_items", "%u", stats.curr_items);
316. APPEND_STAT("total_items", "%u", stats.total_items);
317. STATS_UNLOCK();
318. item_stats_totals(add_stats, c);
319. } else if (nz_strcmp(nkey, stat_type, "items") == 0) {
320. item_stats(add_stats, c);
321. } else if (nz_strcmp(nkey, stat_type, "slabs") == 0) {
322. slabs_stats(add_stats, c);
323. } else if (nz_strcmp(nkey, stat_type, "sizes") == 0) {
324. item_stats_sizes(add_stats, c);
325. } else {
326. ret = false;
327. }
328. } else {
329. ret = false;
330. }
331.  
332. return ret;
333. }
334.  
335. static void do_slabs_stats(ADD_STAT add_stats, void *c) {
336. int i, total;
337. /* Get the per-thread stats which contain some interesting aggregates */
338. struct thread_stats thread_stats;
339. threadlocal_stats_aggregate(&thread_stats);
340.  
341. total = 0;
342. for(i = POWER_SMALLEST; i <= power_largest; i++) {
343. slabclass_t *p = &slabclass[i];
344. if (p->slabs != 0) {
345. uint32_t perslab, slabs;
346. slabs = p->slabs;
347. perslab = p->perslab;
348.  
349. char key_str[STAT_KEY_LEN];
350. char val_str[STAT_VAL_LEN];
351. int klen = 0, vlen = 0;
352.  
353. APPEND_NUM_STAT(i, "chunk_size", "%u", p->size);
354. APPEND_NUM_STAT(i, "chunks_per_page", "%u", perslab);
355. APPEND_NUM_STAT(i, "total_pages", "%u", slabs);
356. APPEND_NUM_STAT(i, "total_chunks", "%u", slabs * perslab);
357. APPEND_NUM_STAT(i, "used_chunks", "%u",
358. slabs*perslab - p->sl_curr);
359. APPEND_NUM_STAT(i, "free_chunks", "%u", p->sl_curr);
360. /* Stat is dead, but displaying zero instead of removing it. */
361. APPEND_NUM_STAT(i, "free_chunks_end", "%u", 0);
362. APPEND_NUM_STAT(i, "mem_requested", "%llu",
363. (unsigned long long)p->requested);
364. APPEND_NUM_STAT(i, "get_hits", "%llu",
365. (unsigned long long)thread_stats.slab_stats[i].get_hits);
366. APPEND_NUM_STAT(i, "cmd_set", "%llu",
367. (unsigned long long)thread_stats.slab_stats[i].set_cmds);
368. APPEND_NUM_STAT(i, "delete_hits", "%llu",
369. (unsigned long long)thread_stats.slab_stats[i].delete_hits);
370. APPEND_NUM_STAT(i, "incr_hits", "%llu",
371. (unsigned long long)thread_stats.slab_stats[i].incr_hits);
372. APPEND_NUM_STAT(i, "decr_hits", "%llu",
373. (unsigned long long)thread_stats.slab_stats[i].decr_hits);
374. APPEND_NUM_STAT(i, "cas_hits", "%llu",
375. (unsigned long long)thread_stats.slab_stats[i].cas_hits);
376. APPEND_NUM_STAT(i, "cas_badval", "%llu",
377. (unsigned long long)thread_stats.slab_stats[i].cas_badval);
378. APPEND_NUM_STAT(i, "touch_hits", "%llu",
379. (unsigned long long)thread_stats.slab_stats[i].touch_hits);
380. total++;
381. }
382. }
383.  
384. APPEND_STAT("active_slabs", "%d", total);
385. APPEND_STAT("total_malloced", "%llu", (unsigned long long)mem_malloced);
386. add_stats(NULL, 0, NULL, 0, c);
387. }
388.  
389. /**
390. 分配内存空间
391. */
392. static void *memory_allocate(size_t size) {
393. void *ret;
394.  
395. /**
396. 有两种分配策略
397. 1）如果是开启了内存预分配策略，则只需要从预分配好的内存块那里割一块出来。即进入下面的else分支
398. 2）如果没有开启预分配，则malloc分配内存
399.  
400. 关于预分配详见 slabs_init
401. */
402. if (mem_base == NULL) {
403. /* We are not using a preallocated large memory chunk */
404. ret = malloc(size);
405. } else {
406. ret = mem_current;
407.  
408. if (size > mem_avail) {
409. return NULL;
410. }
411.  
412. /* mem_current pointer _must_ be aligned!!! */
413. if (size % CHUNK_ALIGN_BYTES) {
414. size += CHUNK_ALIGN_BYTES - (size % CHUNK_ALIGN_BYTES);
415. }
416.  
417. mem_current = ((char*)mem_current) + size;
418. if (size < mem_avail) {
419. mem_avail -= size;
420. } else {
421. mem_avail = 0;
422. }
423. }
424.  
425. return ret;
426. }
427.  
428. void *slabs_alloc(size_t size, unsigned int id) {
429. void *ret;
430.  
431. pthread_mutex_lock(&slabs_lock);
432. ret = do_slabs_alloc(size, id);
433. pthread_mutex_unlock(&slabs_lock);
434. return ret;
435. }
436.  
437. void slabs_free(void *ptr, size_t size, unsigned int id) {
438. pthread_mutex_lock(&slabs_lock);
439. do_slabs_free(ptr, size, id);
440. pthread_mutex_unlock(&slabs_lock);
441. }
442.  
443. void slabs_stats(ADD_STAT add_stats, void *c) {
444. pthread_mutex_lock(&slabs_lock);
445. do_slabs_stats(add_stats, c);
446. pthread_mutex_unlock(&slabs_lock);
447. }
448.  
449. void slabs_adjust_mem_requested(unsigned int id, size_t old, size_t ntotal)
450. {
451. pthread_mutex_lock(&slabs_lock);
452. slabclass_t *p;
453. if (id < POWER_SMALLEST || id > power_largest) {
454. fprintf(stderr, "Internal error! Invalid slab class\n");
455. abort();
456. }
457.  
458. p = &slabclass[id];
459. p->requested = p->requested - old + ntotal;
460. pthread_mutex_unlock(&slabs_lock);
461. }
462.  
463. static pthread_cond_t maintenance_cond = PTHREAD_COND_INITIALIZER;
464. static pthread_cond_t slab_rebalance_cond = PTHREAD_COND_INITIALIZER;
465. static volatile int do_run_slab_thread = 1;
466. static volatile int do_run_slab_rebalance_thread = 1;
467.  
468. #define DEFAULT_SLAB_BULK_CHECK 1
469. int slab_bulk_check = DEFAULT_SLAB_BULK_CHECK;
470.  
471. static int slab_rebalance_start(void) {
472. slabclass_t *s_cls;
473. int no_go = 0;
474.  
475. pthread_mutex_lock(&cache_lock);
476. pthread_mutex_lock(&slabs_lock);
477.  
478. if (slab_rebal.s_clsid < POWER_SMALLEST ||
479. slab_rebal.s_clsid > power_largest  ||
480. slab_rebal.d_clsid < POWER_SMALLEST ||
481. slab_rebal.d_clsid > power_largest  ||
482. slab_rebal.s_clsid == slab_rebal.d_clsid)
483. no_go = -2;
484.  
485. s_cls = &slabclass[slab_rebal.s_clsid];
486.  
487. if (!grow_slab_list(slab_rebal.d_clsid)) {
488. no_go = -1;
489. }
490.  
491. if (s_cls->slabs < 2)
492. no_go = -3;
493.  
494. if (no_go != 0) {
495. pthread_mutex_unlock(&slabs_lock);
496. pthread_mutex_unlock(&cache_lock);
497. return no_go; /* Should use a wrapper function... */
498. }
499.  
500. s_cls->killing = 1;
501.  
502. slab_rebal.slab_start = s_cls->slab_list[s_cls->killing - 1];
503. slab_rebal.slab_end   = (char *)slab_rebal.slab_start +
504. (s_cls->size * s_cls->perslab);
505. slab_rebal.slab_pos   = slab_rebal.slab_start;
506. slab_rebal.done       = 0;
507.  
508. /* Also tells do_item_get to search for items in this slab */
509. slab_rebalance_signal = 2;
510.  
511. if (settings.verbose > 1) {
512. fprintf(stderr, "Started a slab rebalance\n");
513. }
514.  
515. pthread_mutex_unlock(&slabs_lock);
516. pthread_mutex_unlock(&cache_lock);
517.  
518. STATS_LOCK();
519. stats.slab_reassign_running = true;
520. STATS_UNLOCK();
521.  
522. return 0;
523. }
524.  
525. enum move_status {
526. MOVE_PASS=0, MOVE_DONE, MOVE_BUSY, MOVE_LOCKED
527. };
528.  
529. static int slab_rebalance_move(void) {
530. slabclass_t *s_cls;
531. int x;
532. int was_busy = 0;
533. int refcount = 0;
534. enum move_status status = MOVE_PASS;
535.  
536. pthread_mutex_lock(&cache_lock);
537. pthread_mutex_lock(&slabs_lock);
538.  
539. s_cls = &slabclass[slab_rebal.s_clsid];
540.  
541. for (x = 0; x < slab_bulk_check; x++) {
542. item *it = slab_rebal.slab_pos;
543. status = MOVE_PASS;
544. if (it->slabs_clsid != 255) {
545. void *hold_lock = NULL;
546. uint32_t hv = hash(ITEM_key(it), it->nkey);
547. if ((hold_lock = item_trylock(hv)) == NULL) {
548. status = MOVE_LOCKED;
549. } else {
550. refcount = refcount_incr(&it->refcount);
551. if (refcount == 1) { /* item is unlinked, unused */
552. if (it->it_flags & ITEM_SLABBED) {
553. /* remove from slab freelist */
554. if (s_cls->slots == it) {
555. s_cls->slots = it->next;
556. }
557. if (it->next) it->next->prev = it->prev;
558. if (it->prev) it->prev->next = it->next;
559. s_cls->sl_curr--;
560. status = MOVE_DONE;
561. } else {
562. status = MOVE_BUSY;
563. }
564. } else if (refcount == 2) { /* item is linked but not busy */
565. if ((it->it_flags & ITEM_LINKED) != 0) {
566. do_item_unlink_nolock(it, hv);
567. status = MOVE_DONE;
568. } else {
569. /* refcount == 1 + !ITEM_LINKED means the item is being
570. * uploaded to, or was just unlinked but hasn't been freed
571. * yet. Let it bleed off on its own and try again later */
572. status = MOVE_BUSY;
573. }
574. } else {
575. if (settings.verbose > 2) {
576. fprintf(stderr, "Slab reassign hit a busy item: refcount: %d (%d -> %d)\n",
577. it->refcount, slab_rebal.s_clsid, slab_rebal.d_clsid);
578. }
579. status = MOVE_BUSY;
580. }
581. item_trylock_unlock(hold_lock);
582. }
583. }
584.  
585. switch (status) {
586. case MOVE_DONE:
587. it->refcount = 0;
588. it->it_flags = 0;
589. it->slabs_clsid = 255;
590. break;
591. case MOVE_BUSY:
592. refcount_decr(&it->refcount);
593. case MOVE_LOCKED:
594. slab_rebal.busy_items++;
595. was_busy++;
596. break;
597. case MOVE_PASS:
598. break;
599. }
600.  
601. slab_rebal.slab_pos = (char *)slab_rebal.slab_pos + s_cls->size;
602. if (slab_rebal.slab_pos >= slab_rebal.slab_end)
603. break;
604. }
605.  
606. if (slab_rebal.slab_pos >= slab_rebal.slab_end) {
607. /* Some items were busy, start again from the top */
608. if (slab_rebal.busy_items) {
609. slab_rebal.slab_pos = slab_rebal.slab_start;
610. slab_rebal.busy_items = 0;
611. } else {
612. slab_rebal.done++;
613. }
614. }
615.  
616. pthread_mutex_unlock(&slabs_lock);
617. pthread_mutex_unlock(&cache_lock);
618.  
619. return was_busy;
620. }
621.  
622. static void slab_rebalance_finish(void) {
623. slabclass_t *s_cls;
624. slabclass_t *d_cls;
625.  
626. pthread_mutex_lock(&cache_lock);
627. pthread_mutex_lock(&slabs_lock);
628.  
629. s_cls = &slabclass[slab_rebal.s_clsid];
630. d_cls   = &slabclass[slab_rebal.d_clsid];
631.  
632. /* At this point the stolen slab is completely clear */
633. s_cls->slab_list[s_cls->killing - 1] =
634. s_cls->slab_list[s_cls->slabs - 1];
635. s_cls->slabs--;
636. s_cls->killing = 0;
637.  
638. memset(slab_rebal.slab_start, 0, (size_t)settings.item_size_max);
639.  
640. d_cls->slab_list[d_cls->slabs++] = slab_rebal.slab_start;
641. split_slab_page_into_freelist(slab_rebal.slab_start,
642. slab_rebal.d_clsid);
643.  
644. slab_rebal.done       = 0;
645. slab_rebal.s_clsid    = 0;
646. slab_rebal.d_clsid    = 0;
647. slab_rebal.slab_start = NULL;
648. slab_rebal.slab_end   = NULL;
649. slab_rebal.slab_pos   = NULL;
650.  
651. slab_rebalance_signal = 0;
652.  
653. pthread_mutex_unlock(&slabs_lock);
654. pthread_mutex_unlock(&cache_lock);
655.  
656. STATS_LOCK();
657. stats.slab_reassign_running = false;
658. stats.slabs_moved++;
659. STATS_UNLOCK();
660.  
661. if (settings.verbose > 1) {
662. fprintf(stderr, "finished a slab move\n");
663. }
664. }
665.  
666. /*
667. slab自动重分配时，执行此函数做出重分配方案决定
668. */
669. static int slab_automove_decision(int *src, int *dst) {
670. static uint64_t evicted_old[POWER_LARGEST];
671. static unsigned int slab_zeroes[POWER_LARGEST];
672. static unsigned int slab_winner = 0;
673. static unsigned int slab_wins   = 0;
674. uint64_t evicted_new[POWER_LARGEST];
675. uint64_t evicted_diff = 0;
676. uint64_t evicted_max  = 0;
677. unsigned int highest_slab = 0;
678. unsigned int total_pages[POWER_LARGEST];
679. int i;
680. int source = 0;
681. int dest = 0;
682. static rel_time_t next_run;
683.  
684. /* Run less frequently than the slabmove tester. */
685. if (current_time >= next_run) {
686. next_run = current_time + 10;
687. } else {
688. return 0;
689. }
690.  
691. item_stats_evictions(evicted_new);
692. pthread_mutex_lock(&cache_lock);
693. for (i = POWER_SMALLEST; i < power_largest; i++) {
694. total_pages[i] = slabclass[i].slabs;
695. }
696. pthread_mutex_unlock(&cache_lock);
697.  
698. /* Find a candidate source; something with zero evicts 3+ times */
699. for (i = POWER_SMALLEST; i < power_largest; i++) {
700. evicted_diff = evicted_new[i] - evicted_old[i];
701. if (evicted_diff == 0 && total_pages[i] > 2) {
702. slab_zeroes[i]++;
703. if (source == 0 && slab_zeroes[i] >= 3)
704. source = i;
705. } else {
706. slab_zeroes[i] = 0;
707. if (evicted_diff > evicted_max) {
708. evicted_max = evicted_diff;
709. highest_slab = i;
710. }
711. }
712. evicted_old[i] = evicted_new[i];
713. }
714.  
715. /* Pick a valid destination */
716. if (slab_winner != 0 && slab_winner == highest_slab) {
717. slab_wins++;
718. if (slab_wins >= 3)
719. dest = slab_winner;
720. } else {
721. slab_wins = 1;
722. slab_winner = highest_slab;
723. }
724.  
725. if (source && dest) {
726. *src = source;
727. *dst = dest;
728. return 1;
729. }
730. return 0;
731. }
732.  
733. /* Slab rebalancer thread.
734. * Does not use spinlocks since it is not timing sensitive. Burn less CPU and
735. * go to sleep if locks are contended
736. 运行slab维护线程，slab维护线程的执行入口
737. */
738. static void *slab_maintenance_thread(void *arg) {
739. int src, dest;
740.  
741. while (do_run_slab_thread) {
742. if (settings.slab_automove == 1) {
743. if (slab_automove_decision(&src, &dest) == 1) {
744. /* Blind to the return codes. It will retry on its own */
745. slabs_reassign(src, dest); //移动slab，重分配
746. }
747. sleep(1);
748. } else {
749. /* Don't wake as often if we're not enabled.
750. * This is lazier than setting up a condition right now. */
751. sleep(5);
752. }
753. }
754. return NULL;
755. }
756.  
757. /* Slab mover thread.
758. * Sits waiting for a condition to jump off and shovel some memory about
759. */
760. static void *slab_rebalance_thread(void *arg) {
761. int was_busy = 0;
762. /* So we first pass into cond_wait with the mutex held */
763. mutex_lock(&slabs_rebalance_lock);
764.  
765. while (do_run_slab_rebalance_thread) {
766. if (slab_rebalance_signal == 1) {
767. if (slab_rebalance_start() < 0) {
768. /* Handle errors with more specifity as required. */
769. slab_rebalance_signal = 0;
770. }
771.  
772. was_busy = 0;
773. } else if (slab_rebalance_signal && slab_rebal.slab_start != NULL) {
774. was_busy = slab_rebalance_move();
775. }
776.  
777. if (slab_rebal.done) {
778. slab_rebalance_finish();
779. } else if (was_busy) {
780. /* Stuck waiting for some items to unlock, so slow down a bit
781. * to give them a chance to free up */
782. usleep(50);
783. }
784.  
785. if (slab_rebalance_signal == 0) {
786. /* always hold this lock while we're running */
787. pthread_cond_wait(&slab_rebalance_cond, &slabs_rebalance_lock);
788. }
789. }
790. return NULL;
791. }
792.  
793. static int slabs_reassign_pick_any(int dst) {
794. static int cur = POWER_SMALLEST - 1;
795. int tries = power_largest - POWER_SMALLEST + 1;
796. for (; tries > 0; tries--) {
797. cur++;
798. if (cur > power_largest)
799. cur = POWER_SMALLEST;
800. if (cur == dst)
801. continue;
802. if (slabclass[cur].slabs > 1) {
803. return cur;
804. }
805. }
806. return -1;
807. }
808.  
809. static enum reassign_result_type do_slabs_reassign(int src, int dst) {
810. if (slab_rebalance_signal != 0)
811. return REASSIGN_RUNNING;
812.  
813. if (src == dst)
814. return REASSIGN_SRC_DST_SAME;
815.  
816. /* Special indicator to choose ourselves. */
817. if (src == -1) {
818. src = slabs_reassign_pick_any(dst);
819. /* TODO: If we end up back at -1, return a new error type */
820. }
821.  
822. if (src < POWER_SMALLEST || src > power_largest ||
823. dst < POWER_SMALLEST || dst > power_largest)
824. return REASSIGN_BADCLASS;
825.  
826. if (slabclass[src].slabs < 2)
827. return REASSIGN_NOSPARE;
828.  
829. slab_rebal.s_clsid = src;
830. slab_rebal.d_clsid = dst;
831.  
832. slab_rebalance_signal = 1;
833. pthread_cond_signal(&slab_rebalance_cond);
834.  
835. return REASSIGN_OK;
836. }
837.  
838. enum reassign_result_type slabs_reassign(int src, int dst) {
839. enum reassign_result_type ret;
840. if (pthread_mutex_trylock(&slabs_rebalance_lock) != 0) {
841. return REASSIGN_RUNNING;
842. }
843. ret = do_slabs_reassign(src, dst);
844. pthread_mutex_unlock(&slabs_rebalance_lock);
845. return ret;
846. }
847.  
848. /* If we hold this lock, rebalancer can't wake up or move */
849. void slabs_rebalancer_pause(void) {
850. pthread_mutex_lock(&slabs_rebalance_lock);
851. }
852.  
853. void slabs_rebalancer_resume(void) {
854. pthread_mutex_unlock(&slabs_rebalance_lock);
855. }
856.  
857. static pthread_t maintenance_tid;
858. static pthread_t rebalance_tid;
859.  
860. /**
861. 启动slab维护线程
862. */
863. int start_slab_maintenance_thread(void) {
864. int ret;
865. slab_rebalance_signal = 0;
866. slab_rebal.slab_start = NULL;
867. char *env = getenv("MEMCACHED_SLAB_BULK_CHECK");
868. if (env != NULL) {
869. slab_bulk_check = atoi(env);
870. if (slab_bulk_check == 0) {
871. slab_bulk_check = DEFAULT_SLAB_BULK_CHECK;
872. }
873. }
874.  
875. if (pthread_cond_init(&slab_rebalance_cond, NULL) != 0) {
876. fprintf(stderr, "Can't intiialize rebalance condition\n");
877. return -1;
878. }
879. pthread_mutex_init(&slabs_rebalance_lock, NULL);
880.  
881. if ((ret = pthread_create(&maintenance_tid, NULL,
882. slab_maintenance_thread, NULL)) != 0) {
883. fprintf(stderr, "Can't create slab maint thread: %s\n", strerror(ret));
884. return -1;
885. }
886. if ((ret = pthread_create(&rebalance_tid, NULL,
887. slab_rebalance_thread, NULL)) != 0) {
888. fprintf(stderr, "Can't create rebal thread: %s\n", strerror(ret));
889. return -1;
890. }
891. return 0;
892. }
893.  
894. /**
895. 停止slab维护线程，逻辑和停止哈希表维护线程一样。
896. */
897. void stop_slab_maintenance_thread(void) {
898. mutex_lock(&cache_lock);
899. do_run_slab_thread = 0;
900. do_run_slab_rebalance_thread = 0;
901. pthread_cond_signal(&maintenance_cond);
902. pthread_mutex_unlock(&cache_lock);
903.  
904. /* Wait for the maintenance thread to stop */
905. pthread_join(maintenance_tid, NULL);
906. pthread_join(rebalance_tid, NULL);
907. }