+/* =============================== Virtual Memory =========================== */
+static void vmInit(void) {
+ off_t totsize;
+
+ server.vm_fp = fopen("/tmp/redisvm","w+b");
+ if (server.vm_fp == NULL) {
+ redisLog(REDIS_WARNING,"Impossible to open the swap file. Exiting.");
+ exit(1);
+ }
+ server.vm_fd = fileno(server.vm_fp);
+ server.vm_next_page = 0;
+ server.vm_near_pages = 0;
+ server.vm_stats_used_pages = 0;
+ server.vm_stats_swapped_objects = 0;
+ server.vm_stats_swapouts = 0;
+ server.vm_stats_swapins = 0;
+ totsize = server.vm_pages*server.vm_page_size;
+ redisLog(REDIS_NOTICE,"Allocating %lld bytes of swap file",totsize);
+ if (ftruncate(server.vm_fd,totsize) == -1) {
+ redisLog(REDIS_WARNING,"Can't ftruncate swap file: %s. Exiting.",
+ strerror(errno));
+ exit(1);
+ } else {
+ redisLog(REDIS_NOTICE,"Swap file allocated with success");
+ }
+ server.vm_bitmap = zmalloc((server.vm_pages+7)/8);
+ redisLog(REDIS_VERBOSE,"Allocated %lld bytes page table for %lld pages",
+ (long long) (server.vm_pages+7)/8, server.vm_pages);
+ memset(server.vm_bitmap,0,(server.vm_pages+7)/8);
+ /* Try to remove the swap file, so the OS will really delete it from the
+ * file system when Redis exists. */
+ unlink("/tmp/redisvm");
+}
+
+/* Mark the page as used */
+static void vmMarkPageUsed(off_t page) {
+ off_t byte = page/8;
+ int bit = page&7;
+ server.vm_bitmap[byte] |= 1<<bit;
+ redisLog(REDIS_DEBUG,"Mark used: %lld (byte:%lld bit:%d)\n",
+ (long long)page, (long long)byte, bit);
+}
+
+/* Mark N contiguous pages as used, with 'page' being the first. */
+static void vmMarkPagesUsed(off_t page, off_t count) {
+ off_t j;
+
+ for (j = 0; j < count; j++)
+ vmMarkPageUsed(page+j);
+ server.vm_stats_used_pages += count;
+}
+
+/* Mark the page as free */
+static void vmMarkPageFree(off_t page) {
+ off_t byte = page/8;
+ int bit = page&7;
+ server.vm_bitmap[byte] &= ~(1<<bit);
+}
+
+/* Mark N contiguous pages as free, with 'page' being the first. */
+static void vmMarkPagesFree(off_t page, off_t count) {
+ off_t j;
+
+ for (j = 0; j < count; j++)
+ vmMarkPageFree(page+j);
+ server.vm_stats_used_pages -= count;
+}
+
+/* Test if the page is free */
+static int vmFreePage(off_t page) {
+ off_t byte = page/8;
+ int bit = page&7;
+ return (server.vm_bitmap[byte] & (1<<bit)) == 0;
+}
+
+/* Find N contiguous free pages storing the first page of the cluster in *first.
+ * Returns REDIS_OK if it was able to find N contiguous pages, otherwise
+ * REDIS_ERR is returned.
+ *
+ * This function uses a simple algorithm: we try to allocate
+ * REDIS_VM_MAX_NEAR_PAGES sequentially, when we reach this limit we start
+ * again from the start of the swap file searching for free spaces.
+ *
+ * If it looks pretty clear that there are no free pages near our offset
+ * we try to find less populated places doing a forward jump of
+ * REDIS_VM_MAX_RANDOM_JUMP, then we start scanning again a few pages
+ * without hurry, and then we jump again and so forth...
+ *
+ * This function can be improved using a free list to avoid to guess
+ * too much, since we could collect data about freed pages.
+ *
+ * note: I implemented this function just after watching an episode of
+ * Battlestar Galactica, where the hybrid was continuing to say "JUMP!"
+ */
+static int vmFindContiguousPages(off_t *first, int n) {
+ off_t base, offset = 0, since_jump = 0, numfree = 0;
+
+ if (server.vm_near_pages == REDIS_VM_MAX_NEAR_PAGES) {
+ server.vm_near_pages = 0;
+ server.vm_next_page = 0;
+ }
+ server.vm_near_pages++; /* Yet another try for pages near to the old ones */
+ base = server.vm_next_page;
+
+ while(offset < server.vm_pages) {
+ off_t this = base+offset;
+
+ redisLog(REDIS_DEBUG, "THIS: %lld (%c)\n", (long long) this, vmFreePage(this) ? 'F' : 'X');
+ /* If we overflow, restart from page zero */
+ if (this >= server.vm_pages) {
+ this -= server.vm_pages;
+ if (this == 0) {
+ /* Just overflowed, what we found on tail is no longer
+ * interesting, as it's no longer contiguous. */
+ numfree = 0;
+ }
+ }
+ if (vmFreePage(this)) {
+ /* This is a free page */
+ numfree++;
+ /* Already got N free pages? Return to the caller, with success */
+ if (numfree == n) {
+ *first = this-(n-1);
+ server.vm_next_page = this+1;
+ return REDIS_OK;
+ }
+ } else {
+ /* The current one is not a free page */
+ numfree = 0;
+ }
+
+ /* Fast-forward if the current page is not free and we already
+ * searched enough near this place. */
+ since_jump++;
+ if (!numfree && since_jump >= REDIS_VM_MAX_RANDOM_JUMP/4) {
+ offset += random() % REDIS_VM_MAX_RANDOM_JUMP;
+ since_jump = 0;
+ /* Note that even if we rewind after the jump, we are don't need
+ * to make sure numfree is set to zero as we only jump *if* it
+ * is set to zero. */
+ } else {
+ /* Otherwise just check the next page */
+ offset++;
+ }
+ }
+ return REDIS_ERR;
+}
+
+/* Swap the 'val' object relative to 'key' into disk. Store all the information
+ * needed to later retrieve the object into the key object.
+ * If we can't find enough contiguous empty pages to swap the object on disk
+ * REDIS_ERR is returned. */
+static int vmSwapObject(robj *key, robj *val) {
+ off_t pages = rdbSavedObjectPages(val);
+ off_t page;
+
+ assert(key->storage == REDIS_VM_MEMORY);
+ assert(key->refcount == 1);
+ if (vmFindContiguousPages(&page,pages) == REDIS_ERR) return REDIS_ERR;
+ if (fseeko(server.vm_fp,page*server.vm_page_size,SEEK_SET) == -1) {
+ redisLog(REDIS_WARNING,
+ "Critical VM problem in vmSwapObject(): can't seek: %s",
+ strerror(errno));
+ return REDIS_ERR;
+ }
+ rdbSaveObject(server.vm_fp,val);
+ key->vm.page = page;
+ key->vm.usedpages = pages;
+ key->storage = REDIS_VM_SWAPPED;
+ key->vtype = val->type;
+ decrRefCount(val); /* Deallocate the object from memory. */
+ vmMarkPagesUsed(page,pages);
+ redisLog(REDIS_DEBUG,"VM: object %s swapped out at %lld (%lld pages)",
+ (unsigned char*) key->ptr,
+ (unsigned long long) page, (unsigned long long) pages);
+ server.vm_stats_swapped_objects++;
+ server.vm_stats_swapouts++;
+ return REDIS_OK;
+}
+
+/* Load the value object relative to the 'key' object from swap to memory.
+ * The newly allocated object is returned.
+ *
+ * If preview is true the unserialized object is returned to the caller but
+ * no changes are made to the key object, nor the pages are marked as freed */
+static robj *vmGenericLoadObject(robj *key, int preview) {
+ robj *val;
+
+ redisAssert(key->storage == REDIS_VM_SWAPPED);
+ if (fseeko(server.vm_fp,key->vm.page*server.vm_page_size,SEEK_SET) == -1) {
+ redisLog(REDIS_WARNING,
+ "Unrecoverable VM problem in vmLoadObject(): can't seek: %s",
+ strerror(errno));
+ exit(1);
+ }
+ val = rdbLoadObject(key->vtype,server.vm_fp);
+ if (val == NULL) {
+ redisLog(REDIS_WARNING, "Unrecoverable VM problem in vmLoadObject(): can't load object from swap file: %s", strerror(errno));
+ exit(1);
+ }
+ if (!preview) {
+ key->storage = REDIS_VM_MEMORY;
+ key->vm.atime = server.unixtime;
+ vmMarkPagesFree(key->vm.page,key->vm.usedpages);
+ redisLog(REDIS_DEBUG, "VM: object %s loaded from disk",
+ (unsigned char*) key->ptr);
+ server.vm_stats_swapped_objects--;
+ } else {
+ redisLog(REDIS_DEBUG, "VM: object %s previewed from disk",
+ (unsigned char*) key->ptr);
+ }
+ server.vm_stats_swapins++;
+ return val;
+}
+
+/* Plain object loading, from swap to memory */
+static robj *vmLoadObject(robj *key) {
+ return vmGenericLoadObject(key,0);
+}
+
+/* Just load the value on disk, without to modify the key.
+ * This is useful when we want to perform some operation on the value
+ * without to really bring it from swap to memory, like while saving the
+ * dataset or rewriting the append only log. */
+static robj *vmPreviewObject(robj *key) {
+ return vmGenericLoadObject(key,1);
+}
+
+/* How a good candidate is this object for swapping?
+ * The better candidate it is, the greater the returned value.
+ *
+ * Currently we try to perform a fast estimation of the object size in
+ * memory, and combine it with aging informations.
+ *
+ * Basically swappability = idle-time * log(estimated size)
+ *
+ * Bigger objects are preferred over smaller objects, but not
+ * proportionally, this is why we use the logarithm. This algorithm is
+ * just a first try and will probably be tuned later. */
+static double computeObjectSwappability(robj *o) {
+ time_t age = server.unixtime - o->vm.atime;
+ long asize = 0;
+ list *l;
+ dict *d;
+ struct dictEntry *de;
+ int z;
+
+ if (age <= 0) return 0;
+ switch(o->type) {
+ case REDIS_STRING:
+ if (o->encoding != REDIS_ENCODING_RAW) {
+ asize = sizeof(*o);
+ } else {
+ asize = sdslen(o->ptr)+sizeof(*o)+sizeof(long)*2;
+ }
+ break;
+ case REDIS_LIST:
+ l = o->ptr;
+ listNode *ln = listFirst(l);
+
+ asize = sizeof(list);
+ if (ln) {
+ robj *ele = ln->value;
+ long elesize;
+
+ elesize = (ele->encoding == REDIS_ENCODING_RAW) ?
+ (sizeof(*o)+sdslen(ele->ptr)) :
+ sizeof(*o);
+ asize += (sizeof(listNode)+elesize)*listLength(l);
+ }
+ break;
+ case REDIS_SET:
+ case REDIS_ZSET:
+ z = (o->type == REDIS_ZSET);
+ d = z ? ((zset*)o->ptr)->dict : o->ptr;
+
+ asize = sizeof(dict)+(sizeof(struct dictEntry*)*dictSlots(d));
+ if (z) asize += sizeof(zset)-sizeof(dict);
+ if (dictSize(d)) {
+ long elesize;
+ robj *ele;
+
+ de = dictGetRandomKey(d);
+ ele = dictGetEntryKey(de);
+ elesize = (ele->encoding == REDIS_ENCODING_RAW) ?
+ (sizeof(*o)+sdslen(ele->ptr)) :
+ sizeof(*o);
+ asize += (sizeof(struct dictEntry)+elesize)*dictSize(d);
+ if (z) asize += sizeof(zskiplistNode)*dictSize(d);
+ }
+ break;
+ }
+ return (double)asize*log(1+asize);
+}
+
+/* Try to swap an object that's a good candidate for swapping.
+ * Returns REDIS_OK if the object was swapped, REDIS_ERR if it's not possible
+ * to swap any object at all. */
+static int vmSwapOneObject(void) {
+ int j, i;
+ struct dictEntry *best = NULL;
+ double best_swappability = 0;
+ robj *key, *val;
+
+ for (j = 0; j < server.dbnum; j++) {
+ redisDb *db = server.db+j;
+ int maxtries = 1000;
+
+ if (dictSize(db->dict) == 0) continue;
+ for (i = 0; i < 5; i++) {
+ dictEntry *de;
+ double swappability;
+
+ if (maxtries) maxtries--;
+ de = dictGetRandomKey(db->dict);
+ key = dictGetEntryKey(de);
+ val = dictGetEntryVal(de);
+ if (key->storage != REDIS_VM_MEMORY) {
+ if (maxtries) i--; /* don't count this try */
+ continue;
+ }
+ swappability = computeObjectSwappability(val);
+ if (!best || swappability > best_swappability) {
+ best = de;
+ best_swappability = swappability;
+ }
+ }
+ }
+ if (best == NULL) {
+ redisLog(REDIS_DEBUG,"No swappable key found!");
+ return REDIS_ERR;
+ }
+ key = dictGetEntryKey(best);
+ val = dictGetEntryVal(best);
+
+ redisLog(REDIS_DEBUG,"Key with best swappability: %s, %f",
+ key->ptr, best_swappability);
+
+ /* Unshare the key if needed */
+ if (key->refcount > 1) {
+ robj *newkey = dupStringObject(key);
+ decrRefCount(key);
+ key = dictGetEntryKey(best) = newkey;
+ }
+ /* Swap it */
+ if (vmSwapObject(key,val) == REDIS_OK) {
+ dictGetEntryVal(best) = NULL;
+ return REDIS_OK;
+ } else {
+ return REDIS_ERR;
+ }
+}
+
+/* Return true if it's safe to swap out objects in a given moment.
+ * Basically we don't want to swap objects out while there is a BGSAVE
+ * or a BGAEOREWRITE running in backgroud. */
+static int vmCanSwapOut(void) {
+ return (server.bgsavechildpid == -1 && server.bgrewritechildpid == -1);
+}
+
+/* Delete a key if swapped. Returns 1 if the key was found, was swapped
+ * and was deleted. Otherwise 0 is returned. */
+static int deleteIfSwapped(redisDb *db, robj *key) {
+ dictEntry *de;
+ robj *foundkey;
+
+ if ((de = dictFind(db->dict,key)) == NULL) return 0;
+ foundkey = dictGetEntryKey(de);
+ if (foundkey->storage == REDIS_VM_MEMORY) return 0;
+ deleteKey(db,key);
+ return 1;
+}
+
projects / redis.git / blobdiff