| File: | src/defrag.c |
| Warning: | line 1127, column 17 Value stored to 'quit' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | /* |
| 2 | * Active memory defragmentation |
| 3 | * Try to find key / value allocations that need to be re-allocated in order |
| 4 | * to reduce external fragmentation. |
| 5 | * We do that by scanning the keyspace and for each pointer we have, we can try to |
| 6 | * ask the allocator if moving it to a new address will help reduce fragmentation. |
| 7 | * |
| 8 | * Copyright (c) 2020, Redis Labs, Inc |
| 9 | * All rights reserved. |
| 10 | * |
| 11 | * Redistribution and use in source and binary forms, with or without |
| 12 | * modification, are permitted provided that the following conditions are met: |
| 13 | * |
| 14 | * * Redistributions of source code must retain the above copyright notice, |
| 15 | * this list of conditions and the following disclaimer. |
| 16 | * * Redistributions in binary form must reproduce the above copyright |
| 17 | * notice, this list of conditions and the following disclaimer in the |
| 18 | * documentation and/or other materials provided with the distribution. |
| 19 | * * Neither the name of Redis nor the names of its contributors may be used |
| 20 | * to endorse or promote products derived from this software without |
| 21 | * specific prior written permission. |
| 22 | * |
| 23 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" |
| 24 | * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 25 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
| 26 | * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE |
| 27 | * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR |
| 28 | * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF |
| 29 | * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS |
| 30 | * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN |
| 31 | * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) |
| 32 | * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE |
| 33 | * POSSIBILITY OF SUCH DAMAGE. |
| 34 | */ |
| 35 | |
| 36 | #include "server.h" |
| 37 | #include <time.h> |
| 38 | #include <assert.h> |
| 39 | #include <stddef.h> |
| 40 | |
| 41 | #ifdef HAVE_DEFRAG |
| 42 | |
| 43 | /* this method was added to jemalloc in order to help us understand which |
| 44 | * pointers are worthwhile moving and which aren't */ |
| 45 | int je_get_defrag_hint(void* ptr); |
| 46 | |
| 47 | /* forward declarations*/ |
| 48 | void defragDictBucketCallback(void *privdata, dictEntry **bucketref); |
| 49 | dictEntry* replaceSatelliteDictKeyPtrAndOrDefragDictEntry(dict *d, sds oldkey, sds newkey, uint64_t hash, long *defragged); |
| 50 | |
| 51 | /* Defrag helper for generic allocations. |
| 52 | * |
| 53 | * returns NULL in case the allocation wasn't moved. |
| 54 | * when it returns a non-null value, the old pointer was already released |
| 55 | * and should NOT be accessed. */ |
| 56 | void* activeDefragAlloc(void *ptr) { |
| 57 | size_t size; |
| 58 | void *newptr; |
| 59 | if(!je_get_defrag_hint(ptr)) { |
| 60 | server.stat_active_defrag_misses++; |
| 61 | return NULL((void*)0); |
| 62 | } |
| 63 | /* move this allocation to a new allocation. |
| 64 | * make sure not to use the thread cache. so that we don't get back the same |
| 65 | * pointers we try to free */ |
| 66 | size = zmalloc_size(ptr)je_malloc_usable_size(ptr); |
| 67 | newptr = zmalloc_no_tcache(size); |
| 68 | memcpy(newptr, ptr, size); |
| 69 | zfree_no_tcache(ptr); |
| 70 | return newptr; |
| 71 | } |
| 72 | |
| 73 | /*Defrag helper for sds strings |
| 74 | * |
| 75 | * returns NULL in case the allocation wasn't moved. |
| 76 | * when it returns a non-null value, the old pointer was already released |
| 77 | * and should NOT be accessed. */ |
| 78 | sds activeDefragSds(sds sdsptr) { |
| 79 | void* ptr = sdsAllocPtr(sdsptr); |
| 80 | void* newptr = activeDefragAlloc(ptr); |
| 81 | if (newptr) { |
| 82 | size_t offset = sdsptr - (char*)ptr; |
| 83 | sdsptr = (char*)newptr + offset; |
| 84 | return sdsptr; |
| 85 | } |
| 86 | return NULL((void*)0); |
| 87 | } |
| 88 | |
| 89 | /* Defrag helper for robj and/or string objects |
| 90 | * |
| 91 | * returns NULL in case the allocation wasn't moved. |
| 92 | * when it returns a non-null value, the old pointer was already released |
| 93 | * and should NOT be accessed. */ |
| 94 | robj *activeDefragStringOb(robj* ob, long *defragged) { |
| 95 | robj *ret = NULL((void*)0); |
| 96 | if (ob->refcount!=1) |
| 97 | return NULL((void*)0); |
| 98 | |
| 99 | /* try to defrag robj (only if not an EMBSTR type (handled below). */ |
| 100 | if (ob->type!=OBJ_STRING0 || ob->encoding!=OBJ_ENCODING_EMBSTR8) { |
| 101 | if ((ret = activeDefragAlloc(ob))) { |
| 102 | ob = ret; |
| 103 | (*defragged)++; |
| 104 | } |
| 105 | } |
| 106 | |
| 107 | /* try to defrag string object */ |
| 108 | if (ob->type == OBJ_STRING0) { |
| 109 | if(ob->encoding==OBJ_ENCODING_RAW0) { |
| 110 | sds newsds = activeDefragSds((sds)ob->ptr); |
| 111 | if (newsds) { |
| 112 | ob->ptr = newsds; |
| 113 | (*defragged)++; |
| 114 | } |
| 115 | } else if (ob->encoding==OBJ_ENCODING_EMBSTR8) { |
| 116 | /* The sds is embedded in the object allocation, calculate the |
| 117 | * offset and update the pointer in the new allocation. */ |
| 118 | long ofs = (intptr_t)ob->ptr - (intptr_t)ob; |
| 119 | if ((ret = activeDefragAlloc(ob))) { |
| 120 | ret->ptr = (void*)((intptr_t)ret + ofs); |
| 121 | (*defragged)++; |
| 122 | } |
| 123 | } else if (ob->encoding!=OBJ_ENCODING_INT1) { |
| 124 | serverPanic("Unknown string encoding")_serverPanic("defrag.c",124,"Unknown string encoding"),__builtin_unreachable (); |
| 125 | } |
| 126 | } |
| 127 | return ret; |
| 128 | } |
| 129 | |
| 130 | /* Defrag helper for dictEntries to be used during dict iteration (called on |
| 131 | * each step). Returns a stat of how many pointers were moved. */ |
| 132 | long dictIterDefragEntry(dictIterator *iter) { |
| 133 | /* This function is a little bit dirty since it messes with the internals |
| 134 | * of the dict and it's iterator, but the benefit is that it is very easy |
| 135 | * to use, and require no other changes in the dict. */ |
| 136 | long defragged = 0; |
| 137 | dictht *ht; |
| 138 | /* Handle the next entry (if there is one), and update the pointer in the |
| 139 | * current entry. */ |
| 140 | if (iter->nextEntry) { |
| 141 | dictEntry *newde = activeDefragAlloc(iter->nextEntry); |
| 142 | if (newde) { |
| 143 | defragged++; |
| 144 | iter->nextEntry = newde; |
| 145 | iter->entry->next = newde; |
| 146 | } |
| 147 | } |
| 148 | /* handle the case of the first entry in the hash bucket. */ |
| 149 | ht = &iter->d->ht[iter->table]; |
| 150 | if (ht->table[iter->index] == iter->entry) { |
| 151 | dictEntry *newde = activeDefragAlloc(iter->entry); |
| 152 | if (newde) { |
| 153 | iter->entry = newde; |
| 154 | ht->table[iter->index] = newde; |
| 155 | defragged++; |
| 156 | } |
| 157 | } |
| 158 | return defragged; |
| 159 | } |
| 160 | |
| 161 | /* Defrag helper for dict main allocations (dict struct, and hash tables). |
| 162 | * receives a pointer to the dict* and implicitly updates it when the dict |
| 163 | * struct itself was moved. Returns a stat of how many pointers were moved. */ |
| 164 | long dictDefragTables(dict* d) { |
| 165 | dictEntry **newtable; |
| 166 | long defragged = 0; |
| 167 | /* handle the first hash table */ |
| 168 | newtable = activeDefragAlloc(d->ht[0].table); |
| 169 | if (newtable) |
| 170 | defragged++, d->ht[0].table = newtable; |
| 171 | /* handle the second hash table */ |
| 172 | if (d->ht[1].table) { |
| 173 | newtable = activeDefragAlloc(d->ht[1].table); |
| 174 | if (newtable) |
| 175 | defragged++, d->ht[1].table = newtable; |
| 176 | } |
| 177 | return defragged; |
| 178 | } |
| 179 | |
| 180 | /* Internal function used by zslDefrag */ |
| 181 | void zslUpdateNode(zskiplist *zsl, zskiplistNode *oldnode, zskiplistNode *newnode, zskiplistNode **update) { |
| 182 | int i; |
| 183 | for (i = 0; i < zsl->level; i++) { |
| 184 | if (update[i]->level[i].forward == oldnode) |
| 185 | update[i]->level[i].forward = newnode; |
| 186 | } |
| 187 | serverAssert(zsl->header!=oldnode)((zsl->header!=oldnode)?(void)0 : (_serverAssert("zsl->header!=oldnode" ,"defrag.c",187),__builtin_unreachable())); |
| 188 | if (newnode->level[0].forward) { |
| 189 | serverAssert(newnode->level[0].forward->backward==oldnode)((newnode->level[0].forward->backward==oldnode)?(void)0 : (_serverAssert("newnode->level[0].forward->backward==oldnode" ,"defrag.c",189),__builtin_unreachable())); |
| 190 | newnode->level[0].forward->backward = newnode; |
| 191 | } else { |
| 192 | serverAssert(zsl->tail==oldnode)((zsl->tail==oldnode)?(void)0 : (_serverAssert("zsl->tail==oldnode" ,"defrag.c",192),__builtin_unreachable())); |
| 193 | zsl->tail = newnode; |
| 194 | } |
| 195 | } |
| 196 | |
| 197 | /* Defrag helper for sorted set. |
| 198 | * Update the robj pointer, defrag the skiplist struct and return the new score |
| 199 | * reference. We may not access oldele pointer (not even the pointer stored in |
| 200 | * the skiplist), as it was already freed. Newele may be null, in which case we |
| 201 | * only need to defrag the skiplist, but not update the obj pointer. |
| 202 | * When return value is non-NULL, it is the score reference that must be updated |
| 203 | * in the dict record. */ |
| 204 | double *zslDefrag(zskiplist *zsl, double score, sds oldele, sds newele) { |
| 205 | zskiplistNode *update[ZSKIPLIST_MAXLEVEL32], *x, *newx; |
| 206 | int i; |
| 207 | sds ele = newele? newele: oldele; |
| 208 | |
| 209 | /* find the skiplist node referring to the object that was moved, |
| 210 | * and all pointers that need to be updated if we'll end up moving the skiplist node. */ |
| 211 | x = zsl->header; |
| 212 | for (i = zsl->level-1; i >= 0; i--) { |
| 213 | while (x->level[i].forward && |
| 214 | x->level[i].forward->ele != oldele && /* make sure not to access the |
| 215 | ->obj pointer if it matches |
| 216 | oldele */ |
| 217 | (x->level[i].forward->score < score || |
| 218 | (x->level[i].forward->score == score && |
| 219 | sdscmp(x->level[i].forward->ele,ele) < 0))) |
| 220 | x = x->level[i].forward; |
| 221 | update[i] = x; |
| 222 | } |
| 223 | |
| 224 | /* update the robj pointer inside the skip list record. */ |
| 225 | x = x->level[0].forward; |
| 226 | serverAssert(x && score == x->score && x->ele==oldele)((x && score == x->score && x->ele==oldele )?(void)0 : (_serverAssert("x && score == x->score && x->ele==oldele" ,"defrag.c",226),__builtin_unreachable())); |
| 227 | if (newele) |
| 228 | x->ele = newele; |
| 229 | |
| 230 | /* try to defrag the skiplist record itself */ |
| 231 | newx = activeDefragAlloc(x); |
| 232 | if (newx) { |
| 233 | zslUpdateNode(zsl, x, newx, update); |
| 234 | return &newx->score; |
| 235 | } |
| 236 | return NULL((void*)0); |
| 237 | } |
| 238 | |
| 239 | /* Defrag helper for sorted set. |
| 240 | * Defrag a single dict entry key name, and corresponding skiplist struct */ |
| 241 | long activeDefragZsetEntry(zset *zs, dictEntry *de) { |
| 242 | sds newsds; |
| 243 | double* newscore; |
| 244 | long defragged = 0; |
| 245 | sds sdsele = dictGetKey(de)((de)->key); |
| 246 | if ((newsds = activeDefragSds(sdsele))) |
| 247 | defragged++, de->key = newsds; |
| 248 | newscore = zslDefrag(zs->zsl, *(double*)dictGetVal(de)((de)->v.val), sdsele, newsds); |
| 249 | if (newscore) { |
| 250 | dictSetVal(zs->dict, de, newscore)do { if ((zs->dict)->type->valDup) (de)->v.val = ( zs->dict)->type->valDup((zs->dict)->privdata, newscore ); else (de)->v.val = (newscore); } while(0); |
| 251 | defragged++; |
| 252 | } |
| 253 | return defragged; |
| 254 | } |
| 255 | |
| 256 | #define DEFRAG_SDS_DICT_NO_VAL0 0 |
| 257 | #define DEFRAG_SDS_DICT_VAL_IS_SDS1 1 |
| 258 | #define DEFRAG_SDS_DICT_VAL_IS_STROB2 2 |
| 259 | #define DEFRAG_SDS_DICT_VAL_VOID_PTR3 3 |
| 260 | |
| 261 | /* Defrag a dict with sds key and optional value (either ptr, sds or robj string) */ |
| 262 | long activeDefragSdsDict(dict* d, int val_type) { |
| 263 | dictIterator *di; |
| 264 | dictEntry *de; |
| 265 | long defragged = 0; |
| 266 | di = dictGetIterator(d); |
| 267 | while((de = dictNext(di)) != NULL((void*)0)) { |
| 268 | sds sdsele = dictGetKey(de)((de)->key), newsds; |
| 269 | if ((newsds = activeDefragSds(sdsele))) |
| 270 | de->key = newsds, defragged++; |
| 271 | /* defrag the value */ |
| 272 | if (val_type == DEFRAG_SDS_DICT_VAL_IS_SDS1) { |
| 273 | sdsele = dictGetVal(de)((de)->v.val); |
| 274 | if ((newsds = activeDefragSds(sdsele))) |
| 275 | de->v.val = newsds, defragged++; |
| 276 | } else if (val_type == DEFRAG_SDS_DICT_VAL_IS_STROB2) { |
| 277 | robj *newele, *ele = dictGetVal(de)((de)->v.val); |
| 278 | if ((newele = activeDefragStringOb(ele, &defragged))) |
| 279 | de->v.val = newele; |
| 280 | } else if (val_type == DEFRAG_SDS_DICT_VAL_VOID_PTR3) { |
| 281 | void *newptr, *ptr = dictGetVal(de)((de)->v.val); |
| 282 | if ((newptr = activeDefragAlloc(ptr))) |
| 283 | de->v.val = newptr, defragged++; |
| 284 | } |
| 285 | defragged += dictIterDefragEntry(di); |
| 286 | } |
| 287 | dictReleaseIterator(di); |
| 288 | return defragged; |
| 289 | } |
| 290 | |
| 291 | /* Defrag a list of ptr, sds or robj string values */ |
| 292 | long activeDefragList(list *l, int val_type) { |
| 293 | long defragged = 0; |
| 294 | listNode *ln, *newln; |
| 295 | for (ln = l->head; ln; ln = ln->next) { |
| 296 | if ((newln = activeDefragAlloc(ln))) { |
| 297 | if (newln->prev) |
| 298 | newln->prev->next = newln; |
| 299 | else |
| 300 | l->head = newln; |
| 301 | if (newln->next) |
| 302 | newln->next->prev = newln; |
| 303 | else |
| 304 | l->tail = newln; |
| 305 | ln = newln; |
| 306 | defragged++; |
| 307 | } |
| 308 | if (val_type == DEFRAG_SDS_DICT_VAL_IS_SDS1) { |
| 309 | sds newsds, sdsele = ln->value; |
| 310 | if ((newsds = activeDefragSds(sdsele))) |
| 311 | ln->value = newsds, defragged++; |
| 312 | } else if (val_type == DEFRAG_SDS_DICT_VAL_IS_STROB2) { |
| 313 | robj *newele, *ele = ln->value; |
| 314 | if ((newele = activeDefragStringOb(ele, &defragged))) |
| 315 | ln->value = newele; |
| 316 | } else if (val_type == DEFRAG_SDS_DICT_VAL_VOID_PTR3) { |
| 317 | void *newptr, *ptr = ln->value; |
| 318 | if ((newptr = activeDefragAlloc(ptr))) |
| 319 | ln->value = newptr, defragged++; |
| 320 | } |
| 321 | } |
| 322 | return defragged; |
| 323 | } |
| 324 | |
| 325 | /* Defrag a list of sds values and a dict with the same sds keys */ |
| 326 | long activeDefragSdsListAndDict(list *l, dict *d, int dict_val_type) { |
| 327 | long defragged = 0; |
| 328 | sds newsds, sdsele; |
| 329 | listNode *ln, *newln; |
| 330 | dictIterator *di; |
| 331 | dictEntry *de; |
| 332 | /* Defrag the list and it's sds values */ |
| 333 | for (ln = l->head; ln; ln = ln->next) { |
| 334 | if ((newln = activeDefragAlloc(ln))) { |
| 335 | if (newln->prev) |
| 336 | newln->prev->next = newln; |
| 337 | else |
| 338 | l->head = newln; |
| 339 | if (newln->next) |
| 340 | newln->next->prev = newln; |
| 341 | else |
| 342 | l->tail = newln; |
| 343 | ln = newln; |
| 344 | defragged++; |
| 345 | } |
| 346 | sdsele = ln->value; |
| 347 | if ((newsds = activeDefragSds(sdsele))) { |
| 348 | /* When defragging an sds value, we need to update the dict key */ |
| 349 | uint64_t hash = dictGetHash(d, newsds); |
| 350 | replaceSatelliteDictKeyPtrAndOrDefragDictEntry(d, sdsele, newsds, hash, &defragged); |
| 351 | ln->value = newsds; |
| 352 | defragged++; |
| 353 | } |
| 354 | } |
| 355 | |
| 356 | /* Defrag the dict values (keys were already handled) */ |
| 357 | di = dictGetIterator(d); |
| 358 | while((de = dictNext(di)) != NULL((void*)0)) { |
| 359 | if (dict_val_type == DEFRAG_SDS_DICT_VAL_IS_SDS1) { |
| 360 | sds newsds, sdsele = dictGetVal(de)((de)->v.val); |
| 361 | if ((newsds = activeDefragSds(sdsele))) |
| 362 | de->v.val = newsds, defragged++; |
| 363 | } else if (dict_val_type == DEFRAG_SDS_DICT_VAL_IS_STROB2) { |
| 364 | robj *newele, *ele = dictGetVal(de)((de)->v.val); |
| 365 | if ((newele = activeDefragStringOb(ele, &defragged))) |
| 366 | de->v.val = newele, defragged++; |
| 367 | } else if (dict_val_type == DEFRAG_SDS_DICT_VAL_VOID_PTR3) { |
| 368 | void *newptr, *ptr = dictGetVal(de)((de)->v.val); |
| 369 | if ((newptr = activeDefragAlloc(ptr))) |
| 370 | de->v.val = newptr, defragged++; |
| 371 | } |
| 372 | defragged += dictIterDefragEntry(di); |
| 373 | } |
| 374 | dictReleaseIterator(di); |
| 375 | |
| 376 | return defragged; |
| 377 | } |
| 378 | |
| 379 | /* Utility function that replaces an old key pointer in the dictionary with a |
| 380 | * new pointer. Additionally, we try to defrag the dictEntry in that dict. |
| 381 | * Oldkey mey be a dead pointer and should not be accessed (we get a |
| 382 | * pre-calculated hash value). Newkey may be null if the key pointer wasn't |
| 383 | * moved. Return value is the the dictEntry if found, or NULL if not found. |
| 384 | * NOTE: this is very ugly code, but it let's us avoid the complication of |
| 385 | * doing a scan on another dict. */ |
| 386 | dictEntry* replaceSatelliteDictKeyPtrAndOrDefragDictEntry(dict *d, sds oldkey, sds newkey, uint64_t hash, long *defragged) { |
| 387 | dictEntry **deref = dictFindEntryRefByPtrAndHash(d, oldkey, hash); |
| 388 | if (deref) { |
| 389 | dictEntry *de = *deref; |
| 390 | dictEntry *newde = activeDefragAlloc(de); |
| 391 | if (newde) { |
| 392 | de = *deref = newde; |
| 393 | (*defragged)++; |
| 394 | } |
| 395 | if (newkey) |
| 396 | de->key = newkey; |
| 397 | return de; |
| 398 | } |
| 399 | return NULL((void*)0); |
| 400 | } |
| 401 | |
| 402 | long activeDefragQuickListNode(quicklist *ql, quicklistNode **node_ref) { |
| 403 | quicklistNode *newnode, *node = *node_ref; |
| 404 | long defragged = 0; |
| 405 | unsigned char *newzl; |
| 406 | if ((newnode = activeDefragAlloc(node))) { |
| 407 | if (newnode->prev) |
| 408 | newnode->prev->next = newnode; |
| 409 | else |
| 410 | ql->head = newnode; |
| 411 | if (newnode->next) |
| 412 | newnode->next->prev = newnode; |
| 413 | else |
| 414 | ql->tail = newnode; |
| 415 | *node_ref = node = newnode; |
| 416 | defragged++; |
| 417 | } |
| 418 | if ((newzl = activeDefragAlloc(node->zl))) |
| 419 | defragged++, node->zl = newzl; |
| 420 | return defragged; |
| 421 | } |
| 422 | |
| 423 | long activeDefragQuickListNodes(quicklist *ql) { |
| 424 | quicklistNode *node = ql->head; |
| 425 | long defragged = 0; |
| 426 | while (node) { |
| 427 | defragged += activeDefragQuickListNode(ql, &node); |
| 428 | node = node->next; |
| 429 | } |
| 430 | return defragged; |
| 431 | } |
| 432 | |
| 433 | /* when the value has lots of elements, we want to handle it later and not as |
| 434 | * part of the main dictionary scan. this is needed in order to prevent latency |
| 435 | * spikes when handling large items */ |
| 436 | void defragLater(redisDb *db, dictEntry *kde) { |
| 437 | sds key = sdsdup(dictGetKey(kde)((kde)->key)); |
| 438 | listAddNodeTail(db->defrag_later, key); |
| 439 | } |
| 440 | |
| 441 | /* returns 0 if no more work needs to be been done, and 1 if time is up and more work is needed. */ |
| 442 | long scanLaterList(robj *ob, unsigned long *cursor, long long endtime, long long *defragged) { |
| 443 | quicklist *ql = ob->ptr; |
| 444 | quicklistNode *node; |
| 445 | long iterations = 0; |
| 446 | int bookmark_failed = 0; |
| 447 | if (ob->type != OBJ_LIST1 || ob->encoding != OBJ_ENCODING_QUICKLIST9) |
| 448 | return 0; |
| 449 | |
| 450 | if (*cursor == 0) { |
| 451 | /* if cursor is 0, we start new iteration */ |
| 452 | node = ql->head; |
| 453 | } else { |
| 454 | node = quicklistBookmarkFind(ql, "_AD"); |
| 455 | if (!node) { |
| 456 | /* if the bookmark was deleted, it means we reached the end. */ |
| 457 | *cursor = 0; |
| 458 | return 0; |
| 459 | } |
| 460 | node = node->next; |
| 461 | } |
| 462 | |
| 463 | (*cursor)++; |
| 464 | while (node) { |
| 465 | (*defragged) += activeDefragQuickListNode(ql, &node); |
| 466 | server.stat_active_defrag_scanned++; |
| 467 | if (++iterations > 128 && !bookmark_failed) { |
| 468 | if (ustime() > endtime) { |
| 469 | if (!quicklistBookmarkCreate(&ql, "_AD", node)) { |
| 470 | bookmark_failed = 1; |
| 471 | } else { |
| 472 | ob->ptr = ql; /* bookmark creation may have re-allocated the quicklist */ |
| 473 | return 1; |
| 474 | } |
| 475 | } |
| 476 | iterations = 0; |
| 477 | } |
| 478 | node = node->next; |
| 479 | } |
| 480 | quicklistBookmarkDelete(ql, "_AD"); |
| 481 | *cursor = 0; |
| 482 | return bookmark_failed? 1: 0; |
| 483 | } |
| 484 | |
| 485 | typedef struct { |
| 486 | zset *zs; |
| 487 | long defragged; |
| 488 | } scanLaterZsetData; |
| 489 | |
| 490 | void scanLaterZsetCallback(void *privdata, const dictEntry *_de) { |
| 491 | dictEntry *de = (dictEntry*)_de; |
| 492 | scanLaterZsetData *data = privdata; |
| 493 | data->defragged += activeDefragZsetEntry(data->zs, de); |
| 494 | server.stat_active_defrag_scanned++; |
| 495 | } |
| 496 | |
| 497 | long scanLaterZset(robj *ob, unsigned long *cursor) { |
| 498 | if (ob->type != OBJ_ZSET3 || ob->encoding != OBJ_ENCODING_SKIPLIST7) |
| 499 | return 0; |
| 500 | zset *zs = (zset*)ob->ptr; |
| 501 | dict *d = zs->dict; |
| 502 | scanLaterZsetData data = {zs, 0}; |
| 503 | *cursor = dictScan(d, *cursor, scanLaterZsetCallback, defragDictBucketCallback, &data); |
| 504 | return data.defragged; |
| 505 | } |
| 506 | |
| 507 | void scanLaterSetCallback(void *privdata, const dictEntry *_de) { |
| 508 | dictEntry *de = (dictEntry*)_de; |
| 509 | long *defragged = privdata; |
| 510 | sds sdsele = dictGetKey(de)((de)->key), newsds; |
| 511 | if ((newsds = activeDefragSds(sdsele))) |
| 512 | (*defragged)++, de->key = newsds; |
| 513 | server.stat_active_defrag_scanned++; |
| 514 | } |
| 515 | |
| 516 | long scanLaterSet(robj *ob, unsigned long *cursor) { |
| 517 | long defragged = 0; |
| 518 | if (ob->type != OBJ_SET2 || ob->encoding != OBJ_ENCODING_HT2) |
| 519 | return 0; |
| 520 | dict *d = ob->ptr; |
| 521 | *cursor = dictScan(d, *cursor, scanLaterSetCallback, defragDictBucketCallback, &defragged); |
| 522 | return defragged; |
| 523 | } |
| 524 | |
| 525 | void scanLaterHashCallback(void *privdata, const dictEntry *_de) { |
| 526 | dictEntry *de = (dictEntry*)_de; |
| 527 | long *defragged = privdata; |
| 528 | sds sdsele = dictGetKey(de)((de)->key), newsds; |
| 529 | if ((newsds = activeDefragSds(sdsele))) |
| 530 | (*defragged)++, de->key = newsds; |
| 531 | sdsele = dictGetVal(de)((de)->v.val); |
| 532 | if ((newsds = activeDefragSds(sdsele))) |
| 533 | (*defragged)++, de->v.val = newsds; |
| 534 | server.stat_active_defrag_scanned++; |
| 535 | } |
| 536 | |
| 537 | long scanLaterHash(robj *ob, unsigned long *cursor) { |
| 538 | long defragged = 0; |
| 539 | if (ob->type != OBJ_HASH4 || ob->encoding != OBJ_ENCODING_HT2) |
| 540 | return 0; |
| 541 | dict *d = ob->ptr; |
| 542 | *cursor = dictScan(d, *cursor, scanLaterHashCallback, defragDictBucketCallback, &defragged); |
| 543 | return defragged; |
| 544 | } |
| 545 | |
| 546 | long defragQuicklist(redisDb *db, dictEntry *kde) { |
| 547 | robj *ob = dictGetVal(kde)((kde)->v.val); |
| 548 | long defragged = 0; |
| 549 | quicklist *ql = ob->ptr, *newql; |
| 550 | serverAssert(ob->type == OBJ_LIST && ob->encoding == OBJ_ENCODING_QUICKLIST)((ob->type == 1 && ob->encoding == 9)?(void)0 : (_serverAssert("ob->type == OBJ_LIST && ob->encoding == OBJ_ENCODING_QUICKLIST" ,"defrag.c",550),__builtin_unreachable())); |
| 551 | if ((newql = activeDefragAlloc(ql))) |
| 552 | defragged++, ob->ptr = ql = newql; |
| 553 | if (ql->len > server.active_defrag_max_scan_fields) |
| 554 | defragLater(db, kde); |
| 555 | else |
| 556 | defragged += activeDefragQuickListNodes(ql); |
| 557 | return defragged; |
| 558 | } |
| 559 | |
| 560 | long defragZsetSkiplist(redisDb *db, dictEntry *kde) { |
| 561 | robj *ob = dictGetVal(kde)((kde)->v.val); |
| 562 | long defragged = 0; |
| 563 | zset *zs = (zset*)ob->ptr; |
| 564 | zset *newzs; |
| 565 | zskiplist *newzsl; |
| 566 | dict *newdict; |
| 567 | dictEntry *de; |
| 568 | struct zskiplistNode *newheader; |
| 569 | serverAssert(ob->type == OBJ_ZSET && ob->encoding == OBJ_ENCODING_SKIPLIST)((ob->type == 3 && ob->encoding == 7)?(void)0 : (_serverAssert("ob->type == OBJ_ZSET && ob->encoding == OBJ_ENCODING_SKIPLIST" ,"defrag.c",569),__builtin_unreachable())); |
| 570 | if ((newzs = activeDefragAlloc(zs))) |
| 571 | defragged++, ob->ptr = zs = newzs; |
| 572 | if ((newzsl = activeDefragAlloc(zs->zsl))) |
| 573 | defragged++, zs->zsl = newzsl; |
| 574 | if ((newheader = activeDefragAlloc(zs->zsl->header))) |
| 575 | defragged++, zs->zsl->header = newheader; |
| 576 | if (dictSize(zs->dict)((zs->dict)->ht[0].used+(zs->dict)->ht[1].used) > server.active_defrag_max_scan_fields) |
| 577 | defragLater(db, kde); |
| 578 | else { |
| 579 | dictIterator *di = dictGetIterator(zs->dict); |
| 580 | while((de = dictNext(di)) != NULL((void*)0)) { |
| 581 | defragged += activeDefragZsetEntry(zs, de); |
| 582 | } |
| 583 | dictReleaseIterator(di); |
| 584 | } |
| 585 | /* handle the dict struct */ |
| 586 | if ((newdict = activeDefragAlloc(zs->dict))) |
| 587 | defragged++, zs->dict = newdict; |
| 588 | /* defrag the dict tables */ |
| 589 | defragged += dictDefragTables(zs->dict); |
| 590 | return defragged; |
| 591 | } |
| 592 | |
| 593 | long defragHash(redisDb *db, dictEntry *kde) { |
| 594 | long defragged = 0; |
| 595 | robj *ob = dictGetVal(kde)((kde)->v.val); |
| 596 | dict *d, *newd; |
| 597 | serverAssert(ob->type == OBJ_HASH && ob->encoding == OBJ_ENCODING_HT)((ob->type == 4 && ob->encoding == 2)?(void)0 : (_serverAssert("ob->type == OBJ_HASH && ob->encoding == OBJ_ENCODING_HT" ,"defrag.c",597),__builtin_unreachable())); |
| 598 | d = ob->ptr; |
| 599 | if (dictSize(d)((d)->ht[0].used+(d)->ht[1].used) > server.active_defrag_max_scan_fields) |
| 600 | defragLater(db, kde); |
| 601 | else |
| 602 | defragged += activeDefragSdsDict(d, DEFRAG_SDS_DICT_VAL_IS_SDS1); |
| 603 | /* handle the dict struct */ |
| 604 | if ((newd = activeDefragAlloc(ob->ptr))) |
| 605 | defragged++, ob->ptr = newd; |
| 606 | /* defrag the dict tables */ |
| 607 | defragged += dictDefragTables(ob->ptr); |
| 608 | return defragged; |
| 609 | } |
| 610 | |
| 611 | long defragSet(redisDb *db, dictEntry *kde) { |
| 612 | long defragged = 0; |
| 613 | robj *ob = dictGetVal(kde)((kde)->v.val); |
| 614 | dict *d, *newd; |
| 615 | serverAssert(ob->type == OBJ_SET && ob->encoding == OBJ_ENCODING_HT)((ob->type == 2 && ob->encoding == 2)?(void)0 : (_serverAssert("ob->type == OBJ_SET && ob->encoding == OBJ_ENCODING_HT" ,"defrag.c",615),__builtin_unreachable())); |
| 616 | d = ob->ptr; |
| 617 | if (dictSize(d)((d)->ht[0].used+(d)->ht[1].used) > server.active_defrag_max_scan_fields) |
| 618 | defragLater(db, kde); |
| 619 | else |
| 620 | defragged += activeDefragSdsDict(d, DEFRAG_SDS_DICT_NO_VAL0); |
| 621 | /* handle the dict struct */ |
| 622 | if ((newd = activeDefragAlloc(ob->ptr))) |
| 623 | defragged++, ob->ptr = newd; |
| 624 | /* defrag the dict tables */ |
| 625 | defragged += dictDefragTables(ob->ptr); |
| 626 | return defragged; |
| 627 | } |
| 628 | |
| 629 | /* Defrag callback for radix tree iterator, called for each node, |
| 630 | * used in order to defrag the nodes allocations. */ |
| 631 | int defragRaxNode(raxNode **noderef) { |
| 632 | raxNode *newnode = activeDefragAlloc(*noderef); |
| 633 | if (newnode) { |
| 634 | *noderef = newnode; |
| 635 | return 1; |
| 636 | } |
| 637 | return 0; |
| 638 | } |
| 639 | |
| 640 | /* returns 0 if no more work needs to be been done, and 1 if time is up and more work is needed. */ |
| 641 | int scanLaterStreamListpacks(robj *ob, unsigned long *cursor, long long endtime, long long *defragged) { |
| 642 | static unsigned char last[sizeof(streamID)]; |
| 643 | raxIterator ri; |
| 644 | long iterations = 0; |
| 645 | if (ob->type != OBJ_STREAM6 || ob->encoding != OBJ_ENCODING_STREAM10) { |
| 646 | *cursor = 0; |
| 647 | return 0; |
| 648 | } |
| 649 | |
| 650 | stream *s = ob->ptr; |
| 651 | raxStart(&ri,s->rax); |
| 652 | if (*cursor == 0) { |
| 653 | /* if cursor is 0, we start new iteration */ |
| 654 | defragRaxNode(&s->rax->head); |
| 655 | /* assign the iterator node callback before the seek, so that the |
| 656 | * initial nodes that are processed till the first item are covered */ |
| 657 | ri.node_cb = defragRaxNode; |
| 658 | raxSeek(&ri,"^",NULL((void*)0),0); |
| 659 | } else { |
| 660 | /* if cursor is non-zero, we seek to the static 'last' */ |
| 661 | if (!raxSeek(&ri,">", last, sizeof(last))) { |
| 662 | *cursor = 0; |
| 663 | raxStop(&ri); |
| 664 | return 0; |
| 665 | } |
| 666 | /* assign the iterator node callback after the seek, so that the |
| 667 | * initial nodes that are processed till now aren't covered */ |
| 668 | ri.node_cb = defragRaxNode; |
| 669 | } |
| 670 | |
| 671 | (*cursor)++; |
| 672 | while (raxNext(&ri)) { |
| 673 | void *newdata = activeDefragAlloc(ri.data); |
| 674 | if (newdata) |
| 675 | raxSetData(ri.node, ri.data=newdata), (*defragged)++; |
| 676 | server.stat_active_defrag_scanned++; |
| 677 | if (++iterations > 128) { |
| 678 | if (ustime() > endtime) { |
| 679 | serverAssert(ri.key_len==sizeof(last))((ri.key_len==sizeof(last))?(void)0 : (_serverAssert("ri.key_len==sizeof(last)" ,"defrag.c",679),__builtin_unreachable())); |
| 680 | memcpy(last,ri.key,ri.key_len); |
| 681 | raxStop(&ri); |
| 682 | return 1; |
| 683 | } |
| 684 | iterations = 0; |
| 685 | } |
| 686 | } |
| 687 | raxStop(&ri); |
| 688 | *cursor = 0; |
| 689 | return 0; |
| 690 | } |
| 691 | |
| 692 | /* optional callback used defrag each rax element (not including the element pointer itself) */ |
| 693 | typedef void *(raxDefragFunction)(raxIterator *ri, void *privdata, long *defragged); |
| 694 | |
| 695 | /* defrag radix tree including: |
| 696 | * 1) rax struct |
| 697 | * 2) rax nodes |
| 698 | * 3) rax entry data (only if defrag_data is specified) |
| 699 | * 4) call a callback per element, and allow the callback to return a new pointer for the element */ |
| 700 | long defragRadixTree(rax **raxref, int defrag_data, raxDefragFunction *element_cb, void *element_cb_data) { |
| 701 | long defragged = 0; |
| 702 | raxIterator ri; |
| 703 | rax* rax; |
| 704 | if ((rax = activeDefragAlloc(*raxref))) |
| 705 | defragged++, *raxref = rax; |
| 706 | rax = *raxref; |
| 707 | raxStart(&ri,rax); |
| 708 | ri.node_cb = defragRaxNode; |
| 709 | defragRaxNode(&rax->head); |
| 710 | raxSeek(&ri,"^",NULL((void*)0),0); |
| 711 | while (raxNext(&ri)) { |
| 712 | void *newdata = NULL((void*)0); |
| 713 | if (element_cb) |
| 714 | newdata = element_cb(&ri, element_cb_data, &defragged); |
| 715 | if (defrag_data && !newdata) |
| 716 | newdata = activeDefragAlloc(ri.data); |
| 717 | if (newdata) |
| 718 | raxSetData(ri.node, ri.data=newdata), defragged++; |
| 719 | } |
| 720 | raxStop(&ri); |
| 721 | return defragged; |
| 722 | } |
| 723 | |
| 724 | typedef struct { |
| 725 | streamCG *cg; |
| 726 | streamConsumer *c; |
| 727 | } PendingEntryContext; |
| 728 | |
| 729 | void* defragStreamConsumerPendingEntry(raxIterator *ri, void *privdata, long *defragged) { |
| 730 | UNUSED(defragged)((void) defragged); |
| 731 | PendingEntryContext *ctx = privdata; |
| 732 | streamNACK *nack = ri->data, *newnack; |
| 733 | nack->consumer = ctx->c; /* update nack pointer to consumer */ |
| 734 | newnack = activeDefragAlloc(nack); |
| 735 | if (newnack) { |
| 736 | /* update consumer group pointer to the nack */ |
| 737 | void *prev; |
| 738 | raxInsert(ctx->cg->pel, ri->key, ri->key_len, newnack, &prev); |
| 739 | serverAssert(prev==nack)((prev==nack)?(void)0 : (_serverAssert("prev==nack","defrag.c" ,739),__builtin_unreachable())); |
| 740 | /* note: we don't increment 'defragged' that's done by the caller */ |
| 741 | } |
| 742 | return newnack; |
| 743 | } |
| 744 | |
| 745 | void* defragStreamConsumer(raxIterator *ri, void *privdata, long *defragged) { |
| 746 | streamConsumer *c = ri->data; |
| 747 | streamCG *cg = privdata; |
| 748 | void *newc = activeDefragAlloc(c); |
| 749 | if (newc) { |
| 750 | /* note: we don't increment 'defragged' that's done by the caller */ |
| 751 | c = newc; |
| 752 | } |
| 753 | sds newsds = activeDefragSds(c->name); |
| 754 | if (newsds) |
| 755 | (*defragged)++, c->name = newsds; |
| 756 | if (c->pel) { |
| 757 | PendingEntryContext pel_ctx = {cg, c}; |
| 758 | *defragged += defragRadixTree(&c->pel, 0, defragStreamConsumerPendingEntry, &pel_ctx); |
| 759 | } |
| 760 | return newc; /* returns NULL if c was not defragged */ |
| 761 | } |
| 762 | |
| 763 | void* defragStreamConsumerGroup(raxIterator *ri, void *privdata, long *defragged) { |
| 764 | streamCG *cg = ri->data; |
| 765 | UNUSED(privdata)((void) privdata); |
| 766 | if (cg->consumers) |
| 767 | *defragged += defragRadixTree(&cg->consumers, 0, defragStreamConsumer, cg); |
| 768 | if (cg->pel) |
| 769 | *defragged += defragRadixTree(&cg->pel, 0, NULL((void*)0), NULL((void*)0)); |
| 770 | return NULL((void*)0); |
| 771 | } |
| 772 | |
| 773 | long defragStream(redisDb *db, dictEntry *kde) { |
| 774 | long defragged = 0; |
| 775 | robj *ob = dictGetVal(kde)((kde)->v.val); |
| 776 | serverAssert(ob->type == OBJ_STREAM && ob->encoding == OBJ_ENCODING_STREAM)((ob->type == 6 && ob->encoding == 10)?(void)0 : (_serverAssert("ob->type == OBJ_STREAM && ob->encoding == OBJ_ENCODING_STREAM" ,"defrag.c",776),__builtin_unreachable())); |
| 777 | stream *s = ob->ptr, *news; |
| 778 | |
| 779 | /* handle the main struct */ |
| 780 | if ((news = activeDefragAlloc(s))) |
| 781 | defragged++, ob->ptr = s = news; |
| 782 | |
| 783 | if (raxSize(s->rax) > server.active_defrag_max_scan_fields) { |
| 784 | rax *newrax = activeDefragAlloc(s->rax); |
| 785 | if (newrax) |
| 786 | defragged++, s->rax = newrax; |
| 787 | defragLater(db, kde); |
| 788 | } else |
| 789 | defragged += defragRadixTree(&s->rax, 1, NULL((void*)0), NULL((void*)0)); |
| 790 | |
| 791 | if (s->cgroups) |
| 792 | defragged += defragRadixTree(&s->cgroups, 1, defragStreamConsumerGroup, NULL((void*)0)); |
| 793 | return defragged; |
| 794 | } |
| 795 | |
| 796 | /* Defrag a module key. This is either done immediately or scheduled |
| 797 | * for later. Returns then number of pointers defragged. |
| 798 | */ |
| 799 | long defragModule(redisDb *db, dictEntry *kde) { |
| 800 | robj *obj = dictGetVal(kde)((kde)->v.val); |
| 801 | serverAssert(obj->type == OBJ_MODULE)((obj->type == 5)?(void)0 : (_serverAssert("obj->type == OBJ_MODULE" ,"defrag.c",801),__builtin_unreachable())); |
| 802 | long defragged = 0; |
| 803 | |
| 804 | if (!moduleDefragValue(dictGetKey(kde)((kde)->key), obj, &defragged)) |
| 805 | defragLater(db, kde); |
| 806 | |
| 807 | return defragged; |
| 808 | } |
| 809 | |
| 810 | /* for each key we scan in the main dict, this function will attempt to defrag |
| 811 | * all the various pointers it has. Returns a stat of how many pointers were |
| 812 | * moved. */ |
| 813 | long defragKey(redisDb *db, dictEntry *de) { |
| 814 | sds keysds = dictGetKey(de)((de)->key); |
| 815 | robj *newob, *ob; |
| 816 | unsigned char *newzl; |
| 817 | long defragged = 0; |
| 818 | sds newsds; |
| 819 | |
| 820 | /* Try to defrag the key name. */ |
| 821 | newsds = activeDefragSds(keysds); |
| 822 | if (newsds) |
| 823 | defragged++, de->key = newsds; |
| 824 | if (dictSize(db->expires)((db->expires)->ht[0].used+(db->expires)->ht[1].used )) { |
| 825 | /* Dirty code: |
| 826 | * I can't search in db->expires for that key after i already released |
| 827 | * the pointer it holds it won't be able to do the string compare */ |
| 828 | uint64_t hash = dictGetHash(db->dict, de->key); |
| 829 | replaceSatelliteDictKeyPtrAndOrDefragDictEntry(db->expires, keysds, newsds, hash, &defragged); |
| 830 | } |
| 831 | |
| 832 | /* Try to defrag robj and / or string value. */ |
| 833 | ob = dictGetVal(de)((de)->v.val); |
| 834 | if ((newob = activeDefragStringOb(ob, &defragged))) { |
| 835 | de->v.val = newob; |
| 836 | ob = newob; |
| 837 | } |
| 838 | |
| 839 | if (ob->type == OBJ_STRING0) { |
| 840 | /* Already handled in activeDefragStringOb. */ |
| 841 | } else if (ob->type == OBJ_LIST1) { |
| 842 | if (ob->encoding == OBJ_ENCODING_QUICKLIST9) { |
| 843 | defragged += defragQuicklist(db, de); |
| 844 | } else if (ob->encoding == OBJ_ENCODING_ZIPLIST5) { |
| 845 | if ((newzl = activeDefragAlloc(ob->ptr))) |
| 846 | defragged++, ob->ptr = newzl; |
| 847 | } else { |
| 848 | serverPanic("Unknown list encoding")_serverPanic("defrag.c",848,"Unknown list encoding"),__builtin_unreachable (); |
| 849 | } |
| 850 | } else if (ob->type == OBJ_SET2) { |
| 851 | if (ob->encoding == OBJ_ENCODING_HT2) { |
| 852 | defragged += defragSet(db, de); |
| 853 | } else if (ob->encoding == OBJ_ENCODING_INTSET6) { |
| 854 | intset *newis, *is = ob->ptr; |
| 855 | if ((newis = activeDefragAlloc(is))) |
| 856 | defragged++, ob->ptr = newis; |
| 857 | } else { |
| 858 | serverPanic("Unknown set encoding")_serverPanic("defrag.c",858,"Unknown set encoding"),__builtin_unreachable (); |
| 859 | } |
| 860 | } else if (ob->type == OBJ_ZSET3) { |
| 861 | if (ob->encoding == OBJ_ENCODING_ZIPLIST5) { |
| 862 | if ((newzl = activeDefragAlloc(ob->ptr))) |
| 863 | defragged++, ob->ptr = newzl; |
| 864 | } else if (ob->encoding == OBJ_ENCODING_SKIPLIST7) { |
| 865 | defragged += defragZsetSkiplist(db, de); |
| 866 | } else { |
| 867 | serverPanic("Unknown sorted set encoding")_serverPanic("defrag.c",867,"Unknown sorted set encoding"),__builtin_unreachable (); |
| 868 | } |
| 869 | } else if (ob->type == OBJ_HASH4) { |
| 870 | if (ob->encoding == OBJ_ENCODING_ZIPLIST5) { |
| 871 | if ((newzl = activeDefragAlloc(ob->ptr))) |
| 872 | defragged++, ob->ptr = newzl; |
| 873 | } else if (ob->encoding == OBJ_ENCODING_HT2) { |
| 874 | defragged += defragHash(db, de); |
| 875 | } else { |
| 876 | serverPanic("Unknown hash encoding")_serverPanic("defrag.c",876,"Unknown hash encoding"),__builtin_unreachable (); |
| 877 | } |
| 878 | } else if (ob->type == OBJ_STREAM6) { |
| 879 | defragged += defragStream(db, de); |
| 880 | } else if (ob->type == OBJ_MODULE5) { |
| 881 | defragged += defragModule(db, de); |
| 882 | } else { |
| 883 | serverPanic("Unknown object type")_serverPanic("defrag.c",883,"Unknown object type"),__builtin_unreachable (); |
| 884 | } |
| 885 | return defragged; |
| 886 | } |
| 887 | |
| 888 | /* Defrag scan callback for the main db dictionary. */ |
| 889 | void defragScanCallback(void *privdata, const dictEntry *de) { |
| 890 | long defragged = defragKey((redisDb*)privdata, (dictEntry*)de); |
| 891 | server.stat_active_defrag_hits += defragged; |
| 892 | if(defragged) |
| 893 | server.stat_active_defrag_key_hits++; |
| 894 | else |
| 895 | server.stat_active_defrag_key_misses++; |
| 896 | server.stat_active_defrag_scanned++; |
| 897 | } |
| 898 | |
| 899 | /* Defrag scan callback for each hash table bucket, |
| 900 | * used in order to defrag the dictEntry allocations. */ |
| 901 | void defragDictBucketCallback(void *privdata, dictEntry **bucketref) { |
| 902 | UNUSED(privdata)((void) privdata); /* NOTE: this function is also used by both activeDefragCycle and scanLaterHash, etc. don't use privdata */ |
| 903 | while(*bucketref) { |
| 904 | dictEntry *de = *bucketref, *newde; |
| 905 | if ((newde = activeDefragAlloc(de))) { |
| 906 | *bucketref = newde; |
| 907 | } |
| 908 | bucketref = &(*bucketref)->next; |
| 909 | } |
| 910 | } |
| 911 | |
| 912 | /* Utility function to get the fragmentation ratio from jemalloc. |
| 913 | * It is critical to do that by comparing only heap maps that belong to |
| 914 | * jemalloc, and skip ones the jemalloc keeps as spare. Since we use this |
| 915 | * fragmentation ratio in order to decide if a defrag action should be taken |
| 916 | * or not, a false detection can cause the defragmenter to waste a lot of CPU |
| 917 | * without the possibility of getting any results. */ |
| 918 | float getAllocatorFragmentation(size_t *out_frag_bytes) { |
| 919 | size_t resident, active, allocated; |
| 920 | zmalloc_get_allocator_info(&allocated, &active, &resident); |
| 921 | float frag_pct = ((float)active / allocated)*100 - 100; |
| 922 | size_t frag_bytes = active - allocated; |
| 923 | float rss_pct = ((float)resident / allocated)*100 - 100; |
| 924 | size_t rss_bytes = resident - allocated; |
| 925 | if(out_frag_bytes) |
| 926 | *out_frag_bytes = frag_bytes; |
| 927 | serverLog(LL_DEBUG0, |
| 928 | "allocated=%zu, active=%zu, resident=%zu, frag=%.0f%% (%.0f%% rss), frag_bytes=%zu (%zu rss)", |
| 929 | allocated, active, resident, frag_pct, rss_pct, frag_bytes, rss_bytes); |
| 930 | return frag_pct; |
| 931 | } |
| 932 | |
| 933 | /* We may need to defrag other globals, one small allocation can hold a full allocator run. |
| 934 | * so although small, it is still important to defrag these */ |
| 935 | long defragOtherGlobals() { |
| 936 | long defragged = 0; |
| 937 | |
| 938 | /* there are many more pointers to defrag (e.g. client argv, output / aof buffers, etc. |
| 939 | * but we assume most of these are short lived, we only need to defrag allocations |
| 940 | * that remain static for a long time */ |
| 941 | defragged += activeDefragSdsDict(server.lua_scripts, DEFRAG_SDS_DICT_VAL_IS_STROB2); |
| 942 | defragged += activeDefragSdsListAndDict(server.repl_scriptcache_fifo, server.repl_scriptcache_dict, DEFRAG_SDS_DICT_NO_VAL0); |
| 943 | defragged += moduleDefragGlobals(); |
| 944 | return defragged; |
| 945 | } |
| 946 | |
| 947 | /* returns 0 more work may or may not be needed (see non-zero cursor), |
| 948 | * and 1 if time is up and more work is needed. */ |
| 949 | int defragLaterItem(dictEntry *de, unsigned long *cursor, long long endtime) { |
| 950 | if (de) { |
| 951 | robj *ob = dictGetVal(de)((de)->v.val); |
| 952 | if (ob->type == OBJ_LIST1) { |
| 953 | return scanLaterList(ob, cursor, endtime, &server.stat_active_defrag_hits); |
| 954 | } else if (ob->type == OBJ_SET2) { |
| 955 | server.stat_active_defrag_hits += scanLaterSet(ob, cursor); |
| 956 | } else if (ob->type == OBJ_ZSET3) { |
| 957 | server.stat_active_defrag_hits += scanLaterZset(ob, cursor); |
| 958 | } else if (ob->type == OBJ_HASH4) { |
| 959 | server.stat_active_defrag_hits += scanLaterHash(ob, cursor); |
| 960 | } else if (ob->type == OBJ_STREAM6) { |
| 961 | return scanLaterStreamListpacks(ob, cursor, endtime, &server.stat_active_defrag_hits); |
| 962 | } else if (ob->type == OBJ_MODULE5) { |
| 963 | return moduleLateDefrag(dictGetKey(de)((de)->key), ob, cursor, endtime, &server.stat_active_defrag_hits); |
| 964 | } else { |
| 965 | *cursor = 0; /* object type may have changed since we schedule it for later */ |
| 966 | } |
| 967 | } else { |
| 968 | *cursor = 0; /* object may have been deleted already */ |
| 969 | } |
| 970 | return 0; |
| 971 | } |
| 972 | |
| 973 | /* static variables serving defragLaterStep to continue scanning a key from were we stopped last time. */ |
| 974 | static sds defrag_later_current_key = NULL((void*)0); |
| 975 | static unsigned long defrag_later_cursor = 0; |
| 976 | |
| 977 | /* returns 0 if no more work needs to be been done, and 1 if time is up and more work is needed. */ |
| 978 | int defragLaterStep(redisDb *db, long long endtime) { |
| 979 | unsigned int iterations = 0; |
| 980 | unsigned long long prev_defragged = server.stat_active_defrag_hits; |
| 981 | unsigned long long prev_scanned = server.stat_active_defrag_scanned; |
| 982 | long long key_defragged; |
| 983 | |
| 984 | do { |
| 985 | /* if we're not continuing a scan from the last call or loop, start a new one */ |
| 986 | if (!defrag_later_cursor) { |
| 987 | listNode *head = listFirst(db->defrag_later)((db->defrag_later)->head); |
| 988 | |
| 989 | /* Move on to next key */ |
| 990 | if (defrag_later_current_key) { |
| 991 | serverAssert(defrag_later_current_key == head->value)((defrag_later_current_key == head->value)?(void)0 : (_serverAssert ("defrag_later_current_key == head->value","defrag.c",991) ,__builtin_unreachable())); |
| 992 | listDelNode(db->defrag_later, head); |
| 993 | defrag_later_cursor = 0; |
| 994 | defrag_later_current_key = NULL((void*)0); |
| 995 | } |
| 996 | |
| 997 | /* stop if we reached the last one. */ |
| 998 | head = listFirst(db->defrag_later)((db->defrag_later)->head); |
| 999 | if (!head) |
| 1000 | return 0; |
| 1001 | |
| 1002 | /* start a new key */ |
| 1003 | defrag_later_current_key = head->value; |
| 1004 | defrag_later_cursor = 0; |
| 1005 | } |
| 1006 | |
| 1007 | /* each time we enter this function we need to fetch the key from the dict again (if it still exists) */ |
| 1008 | dictEntry *de = dictFind(db->dict, defrag_later_current_key); |
| 1009 | key_defragged = server.stat_active_defrag_hits; |
| 1010 | do { |
| 1011 | int quit = 0; |
| 1012 | if (defragLaterItem(de, &defrag_later_cursor, endtime)) |
| 1013 | quit = 1; /* time is up, we didn't finish all the work */ |
| 1014 | |
| 1015 | /* Once in 16 scan iterations, 512 pointer reallocations, or 64 fields |
| 1016 | * (if we have a lot of pointers in one hash bucket, or rehashing), |
| 1017 | * check if we reached the time limit. */ |
| 1018 | if (quit || (++iterations > 16 || |
| 1019 | server.stat_active_defrag_hits - prev_defragged > 512 || |
| 1020 | server.stat_active_defrag_scanned - prev_scanned > 64)) { |
| 1021 | if (quit || ustime() > endtime) { |
| 1022 | if(key_defragged != server.stat_active_defrag_hits) |
| 1023 | server.stat_active_defrag_key_hits++; |
| 1024 | else |
| 1025 | server.stat_active_defrag_key_misses++; |
| 1026 | return 1; |
| 1027 | } |
| 1028 | iterations = 0; |
| 1029 | prev_defragged = server.stat_active_defrag_hits; |
| 1030 | prev_scanned = server.stat_active_defrag_scanned; |
| 1031 | } |
| 1032 | } while(defrag_later_cursor); |
| 1033 | if(key_defragged != server.stat_active_defrag_hits) |
| 1034 | server.stat_active_defrag_key_hits++; |
| 1035 | else |
| 1036 | server.stat_active_defrag_key_misses++; |
| 1037 | } while(1); |
| 1038 | } |
| 1039 | |
| 1040 | #define INTERPOLATE(x, x1, x2, y1, y2)( (y1) + ((x)-(x1)) * ((y2)-(y1)) / ((x2)-(x1)) ) ( (y1) + ((x)-(x1)) * ((y2)-(y1)) / ((x2)-(x1)) ) |
| 1041 | #define LIMIT(y, min, max)((y)<(min)? min: ((y)>(max)? max: (y))) ((y)<(min)? min: ((y)>(max)? max: (y))) |
| 1042 | |
| 1043 | /* decide if defrag is needed, and at what CPU effort to invest in it */ |
| 1044 | void computeDefragCycles() { |
| 1045 | size_t frag_bytes; |
| 1046 | float frag_pct = getAllocatorFragmentation(&frag_bytes); |
| 1047 | /* If we're not already running, and below the threshold, exit. */ |
| 1048 | if (!server.active_defrag_running) { |
| 1049 | if(frag_pct < server.active_defrag_threshold_lower || frag_bytes < server.active_defrag_ignore_bytes) |
| 1050 | return; |
| 1051 | } |
| 1052 | |
| 1053 | /* Calculate the adaptive aggressiveness of the defrag */ |
| 1054 | int cpu_pct = INTERPOLATE(frag_pct,( (server.active_defrag_cycle_min) + ((frag_pct)-(server.active_defrag_threshold_lower )) * ((server.active_defrag_cycle_max)-(server.active_defrag_cycle_min )) / ((server.active_defrag_threshold_upper)-(server.active_defrag_threshold_lower )) ) |
| 1055 | server.active_defrag_threshold_lower,( (server.active_defrag_cycle_min) + ((frag_pct)-(server.active_defrag_threshold_lower )) * ((server.active_defrag_cycle_max)-(server.active_defrag_cycle_min )) / ((server.active_defrag_threshold_upper)-(server.active_defrag_threshold_lower )) ) |
| 1056 | server.active_defrag_threshold_upper,( (server.active_defrag_cycle_min) + ((frag_pct)-(server.active_defrag_threshold_lower )) * ((server.active_defrag_cycle_max)-(server.active_defrag_cycle_min )) / ((server.active_defrag_threshold_upper)-(server.active_defrag_threshold_lower )) ) |
| 1057 | server.active_defrag_cycle_min,( (server.active_defrag_cycle_min) + ((frag_pct)-(server.active_defrag_threshold_lower )) * ((server.active_defrag_cycle_max)-(server.active_defrag_cycle_min )) / ((server.active_defrag_threshold_upper)-(server.active_defrag_threshold_lower )) ) |
| 1058 | server.active_defrag_cycle_max)( (server.active_defrag_cycle_min) + ((frag_pct)-(server.active_defrag_threshold_lower )) * ((server.active_defrag_cycle_max)-(server.active_defrag_cycle_min )) / ((server.active_defrag_threshold_upper)-(server.active_defrag_threshold_lower )) ); |
| 1059 | cpu_pct = LIMIT(cpu_pct,((cpu_pct)<(server.active_defrag_cycle_min)? server.active_defrag_cycle_min : ((cpu_pct)>(server.active_defrag_cycle_max)? server.active_defrag_cycle_max : (cpu_pct))) |
| 1060 | server.active_defrag_cycle_min,((cpu_pct)<(server.active_defrag_cycle_min)? server.active_defrag_cycle_min : ((cpu_pct)>(server.active_defrag_cycle_max)? server.active_defrag_cycle_max : (cpu_pct))) |
| 1061 | server.active_defrag_cycle_max)((cpu_pct)<(server.active_defrag_cycle_min)? server.active_defrag_cycle_min : ((cpu_pct)>(server.active_defrag_cycle_max)? server.active_defrag_cycle_max : (cpu_pct))); |
| 1062 | /* We allow increasing the aggressiveness during a scan, but don't |
| 1063 | * reduce it. */ |
| 1064 | if (!server.active_defrag_running || |
| 1065 | cpu_pct > server.active_defrag_running) |
| 1066 | { |
| 1067 | server.active_defrag_running = cpu_pct; |
| 1068 | serverLog(LL_VERBOSE1, |
| 1069 | "Starting active defrag, frag=%.0f%%, frag_bytes=%zu, cpu=%d%%", |
| 1070 | frag_pct, frag_bytes, cpu_pct); |
| 1071 | } |
| 1072 | } |
| 1073 | |
| 1074 | /* Perform incremental defragmentation work from the serverCron. |
| 1075 | * This works in a similar way to activeExpireCycle, in the sense that |
| 1076 | * we do incremental work across calls. */ |
| 1077 | void activeDefragCycle(void) { |
| 1078 | static int current_db = -1; |
| 1079 | static unsigned long cursor = 0; |
| 1080 | static redisDb *db = NULL((void*)0); |
| 1081 | static long long start_scan, start_stat; |
| 1082 | unsigned int iterations = 0; |
| 1083 | unsigned long long prev_defragged = server.stat_active_defrag_hits; |
| 1084 | unsigned long long prev_scanned = server.stat_active_defrag_scanned; |
| 1085 | long long start, timelimit, endtime; |
| 1086 | mstime_t latency; |
| 1087 | int quit = 0; |
| 1088 | |
| 1089 | if (!server.active_defrag_enabled) { |
| 1090 | if (server.active_defrag_running) { |
| 1091 | /* if active defrag was disabled mid-run, start from fresh next time. */ |
| 1092 | server.active_defrag_running = 0; |
| 1093 | if (db) |
| 1094 | listEmpty(db->defrag_later); |
| 1095 | defrag_later_current_key = NULL((void*)0); |
| 1096 | defrag_later_cursor = 0; |
| 1097 | current_db = -1; |
| 1098 | cursor = 0; |
| 1099 | db = NULL((void*)0); |
| 1100 | } |
| 1101 | return; |
| 1102 | } |
| 1103 | |
| 1104 | if (hasActiveChildProcess()) |
| 1105 | return; /* Defragging memory while there's a fork will just do damage. */ |
| 1106 | |
| 1107 | /* Once a second, check if the fragmentation justfies starting a scan |
| 1108 | * or making it more aggressive. */ |
| 1109 | run_with_period(1000)if ((1000 <= 1000/server.hz) || !(server.cronloops%((1000) /(1000/server.hz)))) { |
| 1110 | computeDefragCycles(); |
| 1111 | } |
| 1112 | if (!server.active_defrag_running) |
| 1113 | return; |
| 1114 | |
| 1115 | /* See activeExpireCycle for how timelimit is handled. */ |
| 1116 | start = ustime(); |
| 1117 | timelimit = 1000000*server.active_defrag_running/server.hz/100; |
| 1118 | if (timelimit <= 0) timelimit = 1; |
| 1119 | endtime = start + timelimit; |
| 1120 | latencyStartMonitor(latency)if (server.latency_monitor_threshold) { latency = mstime(); } else { latency = 0; }; |
| 1121 | |
| 1122 | do { |
| 1123 | /* if we're not continuing a scan from the last call or loop, start a new one */ |
| 1124 | if (!cursor) { |
| 1125 | /* finish any leftovers from previous db before moving to the next one */ |
| 1126 | if (db && defragLaterStep(db, endtime)) { |
| 1127 | quit = 1; /* time is up, we didn't finish all the work */ |
Value stored to 'quit' is never read | |
| 1128 | break; /* this will exit the function and we'll continue on the next cycle */ |
| 1129 | } |
| 1130 | |
| 1131 | /* Move on to next database, and stop if we reached the last one. */ |
| 1132 | if (++current_db >= server.dbnum) { |
| 1133 | /* defrag other items not part of the db / keys */ |
| 1134 | defragOtherGlobals(); |
| 1135 | |
| 1136 | long long now = ustime(); |
| 1137 | size_t frag_bytes; |
| 1138 | float frag_pct = getAllocatorFragmentation(&frag_bytes); |
| 1139 | serverLog(LL_VERBOSE1, |
| 1140 | "Active defrag done in %dms, reallocated=%d, frag=%.0f%%, frag_bytes=%zu", |
| 1141 | (int)((now - start_scan)/1000), (int)(server.stat_active_defrag_hits - start_stat), frag_pct, frag_bytes); |
| 1142 | |
| 1143 | start_scan = now; |
| 1144 | current_db = -1; |
| 1145 | cursor = 0; |
| 1146 | db = NULL((void*)0); |
| 1147 | server.active_defrag_running = 0; |
| 1148 | |
| 1149 | computeDefragCycles(); /* if another scan is needed, start it right away */ |
| 1150 | if (server.active_defrag_running != 0 && ustime() < endtime) |
| 1151 | continue; |
| 1152 | break; |
| 1153 | } |
| 1154 | else if (current_db==0) { |
| 1155 | /* Start a scan from the first database. */ |
| 1156 | start_scan = ustime(); |
| 1157 | start_stat = server.stat_active_defrag_hits; |
| 1158 | } |
| 1159 | |
| 1160 | db = &server.db[current_db]; |
| 1161 | cursor = 0; |
| 1162 | } |
| 1163 | |
| 1164 | do { |
| 1165 | /* before scanning the next bucket, see if we have big keys left from the previous bucket to scan */ |
| 1166 | if (defragLaterStep(db, endtime)) { |
| 1167 | quit = 1; /* time is up, we didn't finish all the work */ |
| 1168 | break; /* this will exit the function and we'll continue on the next cycle */ |
| 1169 | } |
| 1170 | |
| 1171 | cursor = dictScan(db->dict, cursor, defragScanCallback, defragDictBucketCallback, db); |
| 1172 | |
| 1173 | /* Once in 16 scan iterations, 512 pointer reallocations. or 64 keys |
| 1174 | * (if we have a lot of pointers in one hash bucket or rehasing), |
| 1175 | * check if we reached the time limit. |
| 1176 | * But regardless, don't start a new db in this loop, this is because after |
| 1177 | * the last db we call defragOtherGlobals, which must be done in one cycle */ |
| 1178 | if (!cursor || (++iterations > 16 || |
| 1179 | server.stat_active_defrag_hits - prev_defragged > 512 || |
| 1180 | server.stat_active_defrag_scanned - prev_scanned > 64)) { |
| 1181 | if (!cursor || ustime() > endtime) { |
| 1182 | quit = 1; |
| 1183 | break; |
| 1184 | } |
| 1185 | iterations = 0; |
| 1186 | prev_defragged = server.stat_active_defrag_hits; |
| 1187 | prev_scanned = server.stat_active_defrag_scanned; |
| 1188 | } |
| 1189 | } while(cursor && !quit); |
| 1190 | } while(!quit); |
| 1191 | |
| 1192 | latencyEndMonitor(latency)if (server.latency_monitor_threshold) { latency = mstime() - latency ; }; |
| 1193 | latencyAddSampleIfNeeded("active-defrag-cycle",latency)if (server.latency_monitor_threshold && (latency) >= server.latency_monitor_threshold) latencyAddSample(("active-defrag-cycle" ),(latency));; |
| 1194 | } |
| 1195 | |
| 1196 | #else /* HAVE_DEFRAG */ |
| 1197 | |
| 1198 | void activeDefragCycle(void) { |
| 1199 | /* Not implemented yet. */ |
| 1200 | } |
| 1201 | |
| 1202 | void *activeDefragAlloc(void *ptr) { |
| 1203 | UNUSED(ptr)((void) ptr); |
| 1204 | return NULL((void*)0); |
| 1205 | } |
| 1206 | |
| 1207 | robj *activeDefragStringOb(robj *ob, long *defragged) { |
| 1208 | UNUSED(ob)((void) ob); |
| 1209 | UNUSED(defragged)((void) defragged); |
| 1210 | return NULL((void*)0); |
| 1211 | } |
| 1212 | |
| 1213 | #endif |