+/* =================== Virtual Memory - Threaded I/O ======================= */
+
+static void freeIOJob(iojob *j) {
+ if ((j->type == REDIS_IOJOB_PREPARE_SWAP ||
+ j->type == REDIS_IOJOB_DO_SWAP ||
+ j->type == REDIS_IOJOB_LOAD) && j->val != NULL)
+ decrRefCount(j->val);
+ decrRefCount(j->key);
+ zfree(j);
+}
+
+/* Every time a thread finished a Job, it writes a byte into the write side
+ * of an unix pipe in order to "awake" the main thread, and this function
+ * is called. */
+static void vmThreadedIOCompletedJob(aeEventLoop *el, int fd, void *privdata,
+ int mask)
+{
+ char buf[1];
+ int retval, processed = 0, toprocess = -1, trytoswap = 1;
+ REDIS_NOTUSED(el);
+ REDIS_NOTUSED(mask);
+ REDIS_NOTUSED(privdata);
+
+ /* For every byte we read in the read side of the pipe, there is one
+ * I/O job completed to process. */
+ while((retval = read(fd,buf,1)) == 1) {
+ iojob *j;
+ listNode *ln;
+ robj *key;
+ struct dictEntry *de;
+
+ redisLog(REDIS_DEBUG,"Processing I/O completed job");
+
+ /* Get the processed element (the oldest one) */
+ lockThreadedIO();
+ assert(listLength(server.io_processed) != 0);
+ if (toprocess == -1) {
+ toprocess = (listLength(server.io_processed)*REDIS_MAX_COMPLETED_JOBS_PROCESSED)/100;
+ if (toprocess <= 0) toprocess = 1;
+ }
+ ln = listFirst(server.io_processed);
+ j = ln->value;
+ listDelNode(server.io_processed,ln);
+ unlockThreadedIO();
+ /* If this job is marked as canceled, just ignore it */
+ if (j->canceled) {
+ freeIOJob(j);
+ continue;
+ }
+ /* Post process it in the main thread, as there are things we
+ * can do just here to avoid race conditions and/or invasive locks */
+ redisLog(REDIS_DEBUG,"Job %p type: %d, key at %p (%s) refcount: %d\n", (void*) j, j->type, (void*)j->key, (char*)j->key->ptr, j->key->refcount);
+ de = dictFind(j->db->dict,j->key);
+ assert(de != NULL);
+ key = dictGetEntryKey(de);
+ if (j->type == REDIS_IOJOB_LOAD) {
+ redisDb *db;
+
+ /* Key loaded, bring it at home */
+ 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 (threaded)",
+ (unsigned char*) key->ptr);
+ server.vm_stats_swapped_objects--;
+ server.vm_stats_swapins++;
+ dictGetEntryVal(de) = j->val;
+ incrRefCount(j->val);
+ db = j->db;
+ freeIOJob(j);
+ /* Handle clients waiting for this key to be loaded. */
+ handleClientsBlockedOnSwappedKey(db,key);
+ } else if (j->type == REDIS_IOJOB_PREPARE_SWAP) {
+ /* Now we know the amount of pages required to swap this object.
+ * Let's find some space for it, and queue this task again
+ * rebranded as REDIS_IOJOB_DO_SWAP. */
+ if (!vmCanSwapOut() ||
+ vmFindContiguousPages(&j->page,j->pages) == REDIS_ERR)
+ {
+ /* Ooops... no space or we can't swap as there is
+ * a fork()ed Redis trying to save stuff on disk. */
+ freeIOJob(j);
+ key->storage = REDIS_VM_MEMORY; /* undo operation */
+ } else {
+ /* Note that we need to mark this pages as used now,
+ * if the job will be canceled, we'll mark them as freed
+ * again. */
+ vmMarkPagesUsed(j->page,j->pages);
+ j->type = REDIS_IOJOB_DO_SWAP;
+ lockThreadedIO();
+ queueIOJob(j);
+ unlockThreadedIO();
+ }
+ } else if (j->type == REDIS_IOJOB_DO_SWAP) {
+ robj *val;
+
+ /* Key swapped. We can finally free some memory. */
+ if (key->storage != REDIS_VM_SWAPPING) {
+ printf("key->storage: %d\n",key->storage);
+ printf("key->name: %s\n",(char*)key->ptr);
+ printf("key->refcount: %d\n",key->refcount);
+ printf("val: %p\n",(void*)j->val);
+ printf("val->type: %d\n",j->val->type);
+ printf("val->ptr: %s\n",(char*)j->val->ptr);
+ }
+ redisAssert(key->storage == REDIS_VM_SWAPPING);
+ val = dictGetEntryVal(de);
+ key->vm.page = j->page;
+ key->vm.usedpages = j->pages;
+ key->storage = REDIS_VM_SWAPPED;
+ key->vtype = j->val->type;
+ decrRefCount(val); /* Deallocate the object from memory. */
+ dictGetEntryVal(de) = NULL;
+ redisLog(REDIS_DEBUG,
+ "VM: object %s swapped out at %lld (%lld pages) (threaded)",
+ (unsigned char*) key->ptr,
+ (unsigned long long) j->page, (unsigned long long) j->pages);
+ server.vm_stats_swapped_objects++;
+ server.vm_stats_swapouts++;
+ freeIOJob(j);
+ /* Put a few more swap requests in queue if we are still
+ * out of memory */
+ if (trytoswap && vmCanSwapOut() &&
+ zmalloc_used_memory() > server.vm_max_memory)
+ {
+ int more = 1;
+ while(more) {
+ lockThreadedIO();
+ more = listLength(server.io_newjobs) <
+ (unsigned) server.vm_max_threads;
+ unlockThreadedIO();
+ /* Don't waste CPU time if swappable objects are rare. */
+ if (vmSwapOneObjectThreaded() == REDIS_ERR) {
+ trytoswap = 0;
+ break;
+ }
+ }
+ }
+ }
+ processed++;
+ if (processed == toprocess) return;
+ }
+ if (retval < 0 && errno != EAGAIN) {
+ redisLog(REDIS_WARNING,
+ "WARNING: read(2) error in vmThreadedIOCompletedJob() %s",
+ strerror(errno));
+ }
+}
+
+static void lockThreadedIO(void) {
+ pthread_mutex_lock(&server.io_mutex);
+}
+
+static void unlockThreadedIO(void) {
+ pthread_mutex_unlock(&server.io_mutex);
+}
+
+/* Remove the specified object from the threaded I/O queue if still not
+ * processed, otherwise make sure to flag it as canceled. */
+static void vmCancelThreadedIOJob(robj *o) {
+ list *lists[3] = {
+ server.io_newjobs, /* 0 */
+ server.io_processing, /* 1 */
+ server.io_processed /* 2 */
+ };
+ int i;
+
+ assert(o->storage == REDIS_VM_LOADING || o->storage == REDIS_VM_SWAPPING);
+again:
+ lockThreadedIO();
+ /* Search for a matching key in one of the queues */
+ for (i = 0; i < 3; i++) {
+ listNode *ln;
+ listIter li;
+
+ listRewind(lists[i],&li);
+ while ((ln = listNext(&li)) != NULL) {
+ iojob *job = ln->value;
+
+ if (job->canceled) continue; /* Skip this, already canceled. */
+ if (compareStringObjects(job->key,o) == 0) {
+ redisLog(REDIS_DEBUG,"*** CANCELED %p (%s) (type %d) (LIST ID %d)\n",
+ (void*)job, (char*)o->ptr, job->type, i);
+ /* Mark the pages as free since the swap didn't happened
+ * or happened but is now discarded. */
+ if (i != 1 && job->type == REDIS_IOJOB_DO_SWAP)
+ vmMarkPagesFree(job->page,job->pages);
+ /* Cancel the job. It depends on the list the job is
+ * living in. */
+ switch(i) {
+ case 0: /* io_newjobs */
+ /* If the job was yet not processed the best thing to do
+ * is to remove it from the queue at all */
+ freeIOJob(job);
+ listDelNode(lists[i],ln);
+ break;
+ case 1: /* io_processing */
+ /* Oh Shi- the thread is messing with the Job:
+ *
+ * Probably it's accessing the object if this is a
+ * PREPARE_SWAP or DO_SWAP job.
+ * If it's a LOAD job it may be reading from disk and
+ * if we don't wait for the job to terminate before to
+ * cancel it, maybe in a few microseconds data can be
+ * corrupted in this pages. So the short story is:
+ *
+ * Better to wait for the job to move into the
+ * next queue (processed)... */
+
+ /* We try again and again until the job is completed. */
+ unlockThreadedIO();
+ /* But let's wait some time for the I/O thread
+ * to finish with this job. After all this condition
+ * should be very rare. */
+ usleep(1);
+ goto again;
+ case 2: /* io_processed */
+ /* The job was already processed, that's easy...
+ * just mark it as canceled so that we'll ignore it
+ * when processing completed jobs. */
+ job->canceled = 1;
+ break;
+ }
+ /* Finally we have to adjust the storage type of the object
+ * in order to "UNDO" the operaiton. */
+ if (o->storage == REDIS_VM_LOADING)
+ o->storage = REDIS_VM_SWAPPED;
+ else if (o->storage == REDIS_VM_SWAPPING)
+ o->storage = REDIS_VM_MEMORY;
+ unlockThreadedIO();
+ return;
+ }
+ }
+ }
+ unlockThreadedIO();
+ assert(1 != 1); /* We should never reach this */
+}
+
+static void *IOThreadEntryPoint(void *arg) {
+ iojob *j;
+ listNode *ln;
+ REDIS_NOTUSED(arg);
+
+ pthread_detach(pthread_self());
+ while(1) {
+ /* Get a new job to process */
+ lockThreadedIO();
+ if (listLength(server.io_newjobs) == 0) {
+ /* No new jobs in queue, exit. */
+ redisLog(REDIS_DEBUG,"Thread %ld exiting, nothing to do",
+ (long) pthread_self());
+ server.io_active_threads--;
+ unlockThreadedIO();
+ return NULL;
+ }
+ ln = listFirst(server.io_newjobs);
+ j = ln->value;
+ listDelNode(server.io_newjobs,ln);
+ /* Add the job in the processing queue */
+ j->thread = pthread_self();
+ listAddNodeTail(server.io_processing,j);
+ ln = listLast(server.io_processing); /* We use ln later to remove it */
+ unlockThreadedIO();
+ redisLog(REDIS_DEBUG,"Thread %ld got a new job (type %d): %p about key '%s'",
+ (long) pthread_self(), j->type, (void*)j, (char*)j->key->ptr);
+
+ /* Process the Job */
+ if (j->type == REDIS_IOJOB_LOAD) {
+ j->val = vmReadObjectFromSwap(j->page,j->key->vtype);
+ } else if (j->type == REDIS_IOJOB_PREPARE_SWAP) {
+ FILE *fp = fopen("/dev/null","w+");
+ j->pages = rdbSavedObjectPages(j->val,fp);
+ fclose(fp);
+ } else if (j->type == REDIS_IOJOB_DO_SWAP) {
+ if (vmWriteObjectOnSwap(j->val,j->page) == REDIS_ERR)
+ j->canceled = 1;
+ }
+
+ /* Done: insert the job into the processed queue */
+ redisLog(REDIS_DEBUG,"Thread %ld completed the job: %p (key %s)",
+ (long) pthread_self(), (void*)j, (char*)j->key->ptr);
+ lockThreadedIO();
+ listDelNode(server.io_processing,ln);
+ listAddNodeTail(server.io_processed,j);
+ unlockThreadedIO();
+
+ /* Signal the main thread there is new stuff to process */
+ assert(write(server.io_ready_pipe_write,"x",1) == 1);
+ }
+ return NULL; /* never reached */
+}
+
+static void spawnIOThread(void) {
+ pthread_t thread;
+ sigset_t mask, omask;
+
+ sigemptyset(&mask);
+ sigaddset(&mask,SIGCHLD);
+ sigaddset(&mask,SIGHUP);
+ sigaddset(&mask,SIGPIPE);
+ pthread_sigmask(SIG_SETMASK, &mask, &omask);
+ pthread_create(&thread,&server.io_threads_attr,IOThreadEntryPoint,NULL);
+ pthread_sigmask(SIG_SETMASK, &omask, NULL);
+ server.io_active_threads++;
+}
+
+/* We need to wait for the last thread to exit before we are able to
+ * fork() in order to BGSAVE or BGREWRITEAOF. */
+static void waitEmptyIOJobsQueue(void) {
+ while(1) {
+ int io_processed_len;
+
+ lockThreadedIO();
+ if (listLength(server.io_newjobs) == 0 &&
+ listLength(server.io_processing) == 0 &&
+ server.io_active_threads == 0)
+ {
+ unlockThreadedIO();
+ return;
+ }
+ /* While waiting for empty jobs queue condition we post-process some
+ * finshed job, as I/O threads may be hanging trying to write against
+ * the io_ready_pipe_write FD but there are so much pending jobs that
+ * it's blocking. */
+ io_processed_len = listLength(server.io_processed);
+ unlockThreadedIO();
+ if (io_processed_len) {
+ vmThreadedIOCompletedJob(NULL,server.io_ready_pipe_read,NULL,0);
+ usleep(1000); /* 1 millisecond */
+ } else {
+ usleep(10000); /* 10 milliseconds */
+ }
+ }
+}
+
+static void vmReopenSwapFile(void) {
+ /* Note: we don't close the old one as we are in the child process
+ * and don't want to mess at all with the original file object. */
+ server.vm_fp = fopen(server.vm_swap_file,"r+b");
+ if (server.vm_fp == NULL) {
+ redisLog(REDIS_WARNING,"Can't re-open the VM swap file: %s. Exiting.",
+ server.vm_swap_file);
+ _exit(1);
+ }
+ server.vm_fd = fileno(server.vm_fp);
+}
+
+/* This function must be called while with threaded IO locked */
+static void queueIOJob(iojob *j) {
+ redisLog(REDIS_DEBUG,"Queued IO Job %p type %d about key '%s'\n",
+ (void*)j, j->type, (char*)j->key->ptr);
+ listAddNodeTail(server.io_newjobs,j);
+ if (server.io_active_threads < server.vm_max_threads)
+ spawnIOThread();
+}
+
+static int vmSwapObjectThreaded(robj *key, robj *val, redisDb *db) {
+ iojob *j;
+
+ assert(key->storage == REDIS_VM_MEMORY);
+ assert(key->refcount == 1);
+
+ j = zmalloc(sizeof(*j));
+ j->type = REDIS_IOJOB_PREPARE_SWAP;
+ j->db = db;
+ j->key = dupStringObject(key);
+ j->val = val;
+ incrRefCount(val);
+ j->canceled = 0;
+ j->thread = (pthread_t) -1;
+ key->storage = REDIS_VM_SWAPPING;
+
+ lockThreadedIO();
+ queueIOJob(j);
+ unlockThreadedIO();
+ return REDIS_OK;
+}
+
+/* ============ Virtual Memory - Blocking clients on missing keys =========== */
+
+/* This function makes the clinet 'c' waiting for the key 'key' to be loaded.
+ * If there is not already a job loading the key, it is craeted.
+ * The key is added to the io_keys list in the client structure, and also
+ * in the hash table mapping swapped keys to waiting clients, that is,
+ * server.io_waited_keys. */
+static int waitForSwappedKey(redisClient *c, robj *key) {
+ struct dictEntry *de;
+ robj *o;
+ list *l;
+
+ /* If the key does not exist or is already in RAM we don't need to
+ * block the client at all. */
+ de = dictFind(c->db->dict,key);
+ if (de == NULL) return 0;
+ o = dictGetEntryKey(de);
+ if (o->storage == REDIS_VM_MEMORY) {
+ return 0;
+ } else if (o->storage == REDIS_VM_SWAPPING) {
+ /* We were swapping the key, undo it! */
+ vmCancelThreadedIOJob(o);
+ return 0;
+ }
+
+ /* OK: the key is either swapped, or being loaded just now. */
+
+ /* Add the key to the list of keys this client is waiting for.
+ * This maps clients to keys they are waiting for. */
+ listAddNodeTail(c->io_keys,key);
+ incrRefCount(key);
+
+ /* Add the client to the swapped keys => clients waiting map. */
+ de = dictFind(c->db->io_keys,key);
+ if (de == NULL) {
+ int retval;
+
+ /* For every key we take a list of clients blocked for it */
+ l = listCreate();
+ retval = dictAdd(c->db->io_keys,key,l);
+ incrRefCount(key);
+ assert(retval == DICT_OK);
+ } else {
+ l = dictGetEntryVal(de);
+ }
+ listAddNodeTail(l,c);
+
+ /* Are we already loading the key from disk? If not create a job */
+ if (o->storage == REDIS_VM_SWAPPED) {
+ iojob *j;
+
+ o->storage = REDIS_VM_LOADING;
+ j = zmalloc(sizeof(*j));
+ j->type = REDIS_IOJOB_LOAD;
+ j->db = c->db;
+ j->key = dupStringObject(key);
+ j->key->vtype = o->vtype;
+ j->page = o->vm.page;
+ j->val = NULL;
+ j->canceled = 0;
+ j->thread = (pthread_t) -1;
+ lockThreadedIO();
+ queueIOJob(j);
+ unlockThreadedIO();
+ }
+ return 1;
+}
+
+/* Is this client attempting to run a command against swapped keys?
+ * If so, block it ASAP, load the keys in background, then resume it.
+ *
+ * The important idea about this function is that it can fail! If keys will
+ * still be swapped when the client is resumed, this key lookups will
+ * just block loading keys from disk. In practical terms this should only
+ * happen with SORT BY command or if there is a bug in this function.
+ *
+ * Return 1 if the client is marked as blocked, 0 if the client can
+ * continue as the keys it is going to access appear to be in memory. */
+static int blockClientOnSwappedKeys(struct redisCommand *cmd, redisClient *c) {
+ int j, last;
+
+ if (cmd->vm_firstkey == 0) return 0;
+ last = cmd->vm_lastkey;
+ if (last < 0) last = c->argc+last;
+ for (j = cmd->vm_firstkey; j <= last; j += cmd->vm_keystep)
+ waitForSwappedKey(c,c->argv[j]);
+ /* If the client was blocked for at least one key, mark it as blocked. */
+ if (listLength(c->io_keys)) {
+ c->flags |= REDIS_IO_WAIT;
+ aeDeleteFileEvent(server.el,c->fd,AE_READABLE);
+ server.vm_blocked_clients++;
+ return 1;
+ } else {
+ return 0;
+ }
+}
+
+/* Remove the 'key' from the list of blocked keys for a given client.
+ *
+ * The function returns 1 when there are no longer blocking keys after
+ * the current one was removed (and the client can be unblocked). */
+static int dontWaitForSwappedKey(redisClient *c, robj *key) {
+ list *l;
+ listNode *ln;
+ listIter li;
+ struct dictEntry *de;
+
+ /* Remove the key from the list of keys this client is waiting for. */
+ listRewind(c->io_keys,&li);
+ while ((ln = listNext(&li)) != NULL) {
+ if (compareStringObjects(ln->value,key) == 0) {
+ listDelNode(c->io_keys,ln);
+ break;
+ }
+ }
+ assert(ln != NULL);
+
+ /* Remove the client form the key => waiting clients map. */
+ de = dictFind(c->db->io_keys,key);
+ assert(de != NULL);
+ l = dictGetEntryVal(de);
+ ln = listSearchKey(l,c);
+ assert(ln != NULL);
+ listDelNode(l,ln);
+ if (listLength(l) == 0)
+ dictDelete(c->db->io_keys,key);
+
+ return listLength(c->io_keys) == 0;
+}
+
+static void handleClientsBlockedOnSwappedKey(redisDb *db, robj *key) {
+ struct dictEntry *de;
+ list *l;
+ listNode *ln;
+ int len;
+
+ de = dictFind(db->io_keys,key);
+ if (!de) return;
+
+ l = dictGetEntryVal(de);
+ len = listLength(l);
+ /* Note: we can't use something like while(listLength(l)) as the list
+ * can be freed by the calling function when we remove the last element. */
+ while (len--) {
+ ln = listFirst(l);
+ redisClient *c = ln->value;
+
+ if (dontWaitForSwappedKey(c,key)) {
+ /* Put the client in the list of clients ready to go as we
+ * loaded all the keys about it. */
+ listAddNodeTail(server.io_ready_clients,c);
+ }
+ }
+}
+
projects / redis.git / blobdiff