#include <sys/uio.h>
#include <limits.h>
#include <math.h>
+#include <pthread.h>
#if defined(__sun)
#include "solarisfixes.h"
#include "lzf.h" /* LZF compression library */
#include "pqsort.h" /* Partial qsort for SORT+LIMIT */
+/* #define REDIS_HELGRIND_FRIENDLY */
+#if defined(__GNUC__) && defined(REDIS_HELGRIND_FRIENDLY)
+#warning "Remember to undef REDIS_HELGRIND_FRIENDLY before to commit"
+#endif
+
/* Error codes */
#define REDIS_OK 0
#define REDIS_ERR -1
* Check vmFindContiguousPages() to know more about this magic numbers. */
#define REDIS_VM_MAX_NEAR_PAGES 65536
#define REDIS_VM_MAX_RANDOM_JUMP 4096
+#define REDIS_VM_MAX_THREADS 32
+#define REDIS_THREAD_STACK_SIZE (1024*1024*4)
+/* The following is the number of completed I/O jobs to process when the
+ * handelr is called. 1 is the minimum, and also the default, as it allows
+ * to block as little as possible other accessing clients. While Virtual
+ * Memory I/O operations are performed by threads, this operations must
+ * be processed by the main thread when completed to take effect. */
+#define REDIS_MAX_COMPLETED_JOBS_PROCESSED 1
/* Client flags */
#define REDIS_CLOSE 1 /* This client connection should be closed ASAP */
#define REDIS_MONITOR 8 /* This client is a slave monitor, see MONITOR */
#define REDIS_MULTI 16 /* This client is in a MULTI context */
#define REDIS_BLOCKED 32 /* The client is waiting in a blocking operation */
+#define REDIS_IO_WAIT 64 /* The client is waiting for Virtual Memory I/O */
/* Slave replication state - slave side */
#define REDIS_REPL_NONE 0 /* No active replication */
/* Log levels */
#define REDIS_DEBUG 0
-#define REDIS_NOTICE 1
-#define REDIS_WARNING 2
+#define REDIS_VERBOSE 1
+#define REDIS_NOTICE 2
+#define REDIS_WARNING 3
/* Anti-warning macro... */
#define REDIS_NOTUSED(V) ((void) V)
#define APPENDFSYNC_EVERYSEC 2
/* We can print the stacktrace, so our assert is defined this way: */
-#define redisAssert(_e) ((_e)?(void)0 : (_redisAssert(#_e),exit(1)))
-static void _redisAssert(char *estr);
+#define redisAssert(_e) ((_e)?(void)0 : (_redisAssert(#_e,__FILE__,__LINE__),exit(1)))
+static void _redisAssert(char *estr, char *file, int line);
/*================================= Data types ============================== */
int blockingkeysnum; /* Number of blocking keys */
time_t blockingto; /* Blocking operation timeout. If UNIX current time
* is >= blockingto then the operation timed out. */
+ list *io_keys; /* Keys this client is waiting to be loaded from the
+ * swap file in order to continue. */
} redisClient;
struct saveparam {
off_t vm_near_pages; /* Number of pages allocated sequentially */
unsigned char *vm_bitmap; /* Bitmap of free/used pages */
time_t unixtime; /* Unix time sampled every second. */
+ /* Virtual memory I/O threads stuff */
+ /* An I/O thread process an element taken from the io_jobs queue and
+ * put the result of the operation in the io_done list. While the
+ * job is being processed, it's put on io_processing queue. */
+ list *io_newjobs; /* List of VM I/O jobs yet to be processed */
+ list *io_processing; /* List of VM I/O jobs being processed */
+ list *io_processed; /* List of VM I/O jobs already processed */
+ list *io_clients; /* All the clients waiting for SWAP I/O operations */
+ pthread_mutex_t io_mutex; /* lock to access io_jobs/io_done/io_thread_job */
+ pthread_mutex_t obj_freelist_mutex; /* safe redis objects creation/free */
+ pthread_mutex_t io_swapfile_mutex; /* So we can lseek + write */
+ pthread_attr_t io_threads_attr; /* attributes for threads creation */
+ int io_active_threads; /* Number of running I/O threads */
+ int vm_max_threads; /* Max number of I/O threads running at the same time */
+ /* Our main thread is blocked on the event loop, locking for sockets ready
+ * to be read or written, so when a threaded I/O operation is ready to be
+ * processed by the main thread, the I/O thread will use a unix pipe to
+ * awake the main thread. The followings are the two pipe FDs. */
+ int io_ready_pipe_read;
+ int io_ready_pipe_write;
+ /* Virtual memory stats */
+ unsigned long long vm_stats_used_pages;
+ unsigned long long vm_stats_swapped_objects;
+ unsigned long long vm_stats_swapouts;
+ unsigned long long vm_stats_swapins;
+ FILE *devnull;
};
typedef void redisCommandProc(redisClient *c);
static double R_Zero, R_PosInf, R_NegInf, R_Nan;
+/* VM threaded I/O request message */
+#define REDIS_IOJOB_LOAD 0 /* Load from disk to memory */
+#define REDIS_IOJOB_PREPARE_SWAP 1 /* Compute needed pages */
+#define REDIS_IOJOB_DO_SWAP 2 /* Swap from memory to disk */
+typedef struct iojon {
+ int type; /* Request type, REDIS_IOJOB_* */
+ redisDb *db;/* Redis database */
+ robj *key; /* This I/O request is about swapping this key */
+ robj *val; /* the value to swap for REDIS_IOREQ_*_SWAP, otherwise this
+ * field is populated by the I/O thread for REDIS_IOREQ_LOAD. */
+ off_t page; /* Swap page where to read/write the object */
+ off_t pages; /* Swap pages needed to safe object. PREPARE_SWAP return val */
+ int canceled; /* True if this command was canceled by blocking side of VM */
+ pthread_t thread; /* ID of the thread processing this entry */
+} iojob;
+
/*================================ Prototypes =============================== */
static void freeStringObject(robj *o);
static int removeExpire(redisDb *db, robj *key);
static int expireIfNeeded(redisDb *db, robj *key);
static int deleteIfVolatile(redisDb *db, robj *key);
+static int deleteIfSwapped(redisDb *db, robj *key);
static int deleteKey(redisDb *db, robj *key);
static time_t getExpire(redisDb *db, robj *key);
static int setExpire(redisDb *db, robj *key, time_t when);
static void vmInit(void);
static void vmMarkPagesFree(off_t page, off_t count);
static robj *vmLoadObject(robj *key);
-static int vmSwapOneObject(void);
+static robj *vmPreviewObject(robj *key);
+static int vmSwapOneObjectBlocking(void);
+static int vmSwapOneObjectThreaded(void);
+static int vmCanSwapOut(void);
+static int tryFreeOneObjectFromFreelist(void);
+static void acceptHandler(aeEventLoop *el, int fd, void *privdata, int mask);
+static void vmThreadedIOCompletedJob(aeEventLoop *el, int fd, void *privdata, int mask);
+static void vmCancelThreadedIOJob(robj *o);
+static void lockThreadedIO(void);
+static void unlockThreadedIO(void);
+static int vmSwapObjectThreaded(robj *key, robj *val, redisDb *db);
+static void freeIOJob(iojob *j);
+static void queueIOJob(iojob *j);
+static int vmWriteObjectOnSwap(robj *o, off_t page);
+static robj *vmReadObjectFromSwap(off_t page, int type);
+static void waitZeroActiveThreads(void);
static void authCommand(redisClient *c);
static void pingCommand(redisClient *c);
return hash;
}
+/* Sets type and expires */
static dictType setDictType = {
dictEncObjHash, /* hash function */
NULL, /* key dup */
NULL /* val destructor */
};
+/* Sorted sets hash (note: a skiplist is used in addition to the hash table) */
static dictType zsetDictType = {
dictEncObjHash, /* hash function */
NULL, /* key dup */
dictVanillaFree /* val destructor of malloc(sizeof(double)) */
};
+/* Db->dict */
static dictType hashDictType = {
dictObjHash, /* hash function */
NULL, /* key dup */
dictRedisObjectDestructor /* val destructor */
};
+/* Db->expires */
+static dictType keyptrDictType = {
+ dictObjHash, /* hash function */
+ NULL, /* key dup */
+ NULL, /* val dup */
+ dictObjKeyCompare, /* key compare */
+ dictRedisObjectDestructor, /* key destructor */
+ NULL /* val destructor */
+};
+
/* Keylist hash table type has unencoded redis objects as keys and
* lists as values. It's used for blocking operations (BLPOP) */
static dictType keylistDictType = {
redisClient *c;
listNode *ln;
time_t now = time(NULL);
+ listIter li;
- listRewind(server.clients);
- while ((ln = listYield(server.clients)) != NULL) {
+ listRewind(server.clients,&li);
+ while ((ln = listNext(&li)) != NULL) {
c = listNodeValue(ln);
if (server.maxidletime &&
!(c->flags & REDIS_SLAVE) && /* no timeout for slaves */
!(c->flags & REDIS_MASTER) && /* no timeout for masters */
(now - c->lastinteraction > server.maxidletime))
{
- redisLog(REDIS_DEBUG,"Closing idle client");
+ redisLog(REDIS_VERBOSE,"Closing idle client");
freeClient(c);
} else if (c->flags & REDIS_BLOCKED) {
if (c->blockingto != 0 && c->blockingto < now) {
for (j = 0; j < server.dbnum; j++) {
if (htNeedsResize(server.db[j].dict)) {
- redisLog(REDIS_DEBUG,"The hash table %d is too sparse, resize it...",j);
+ redisLog(REDIS_VERBOSE,"The hash table %d is too sparse, resize it...",j);
dictResize(server.db[j].dict);
- redisLog(REDIS_DEBUG,"Hash table %d resized.",j);
+ redisLog(REDIS_VERBOSE,"Hash table %d resized.",j);
}
if (htNeedsResize(server.db[j].expires))
dictResize(server.db[j].expires);
used = dictSize(server.db[j].dict);
vkeys = dictSize(server.db[j].expires);
if (!(loops % 5) && (used || vkeys)) {
- redisLog(REDIS_DEBUG,"DB %d: %lld keys (%lld volatile) in %lld slots HT.",j,used,vkeys,size);
+ redisLog(REDIS_VERBOSE,"DB %d: %lld keys (%lld volatile) in %lld slots HT.",j,used,vkeys,size);
/* dictPrintStats(server.dict); */
}
}
/* Show information about connected clients */
if (!(loops % 5)) {
- redisLog(REDIS_DEBUG,"%d clients connected (%d slaves), %zu bytes in use, %d shared objects",
+ redisLog(REDIS_VERBOSE,"%d clients connected (%d slaves), %zu bytes in use, %d shared objects",
listLength(server.clients)-listLength(server.slaves),
listLength(server.slaves),
server.usedmemory,
}
/* Swap a few keys on disk if we are over the memory limit and VM
- * is enbled. */
- while (server.vm_enabled && zmalloc_used_memory() > server.vm_max_memory) {
- if (vmSwapOneObject() == REDIS_ERR) {
- redisLog(REDIS_WARNING,"WARNING: vm-max-memory limit reached but unable to swap more objects out!");
- break;
+ * is enbled. Try to free objects from the free list first. */
+ if (vmCanSwapOut()) {
+ while (server.vm_enabled && zmalloc_used_memory() >
+ server.vm_max_memory)
+ {
+ int retval;
+
+ if (tryFreeOneObjectFromFreelist() == REDIS_OK) continue;
+ retval = (server.vm_max_threads == 0) ?
+ vmSwapOneObjectBlocking() :
+ vmSwapOneObjectThreaded();
+ if (retval == REDIS_ERR && (loops % 30) == 0 &&
+ zmalloc_used_memory() >
+ (server.vm_max_memory+server.vm_max_memory/10))
+ {
+ redisLog(REDIS_WARNING,"WARNING: vm-max-memory limit exceeded by more than 10%% but unable to swap more objects out!");
+ }
+ /* Note that when using threade I/O we free just one object,
+ * because anyway when the I/O thread in charge to swap this
+ * object out will finish, the handler of completed jobs
+ * will try to swap more objects if we are still out of memory. */
+ if (retval == REDIS_ERR || server.vm_max_threads > 0) break;
}
}
static void initServerConfig() {
server.dbnum = REDIS_DEFAULT_DBNUM;
server.port = REDIS_SERVERPORT;
- server.verbosity = REDIS_DEBUG;
+ server.verbosity = REDIS_VERBOSE;
server.maxidletime = REDIS_MAXIDLETIME;
server.saveparams = NULL;
server.logfile = NULL; /* NULL = log on standard output */
server.vm_page_size = 256; /* 256 bytes per page */
server.vm_pages = 1024*1024*100; /* 104 millions of pages */
server.vm_max_memory = 1024LL*1024*1024*1; /* 1 GB of RAM */
+ server.vm_max_threads = 4;
resetServerSaveParams();
signal(SIGPIPE, SIG_IGN);
setupSigSegvAction();
+ server.devnull = fopen("/dev/null","w");
+ if (server.devnull == NULL) {
+ redisLog(REDIS_WARNING, "Can't open /dev/null: %s", server.neterr);
+ exit(1);
+ }
server.clients = listCreate();
server.slaves = listCreate();
server.monitors = listCreate();
}
for (j = 0; j < server.dbnum; j++) {
server.db[j].dict = dictCreate(&hashDictType,NULL);
- server.db[j].expires = dictCreate(&setDictType,NULL);
+ server.db[j].expires = dictCreate(&keyptrDictType,NULL);
server.db[j].blockingkeys = dictCreate(&keylistDictType,NULL);
server.db[j].id = j;
}
server.stat_starttime = time(NULL);
server.unixtime = time(NULL);
aeCreateTimeEvent(server.el, 1, serverCron, NULL, NULL);
+ if (aeCreateFileEvent(server.el, server.fd, AE_READABLE,
+ acceptHandler, NULL) == AE_ERR) oom("creating file event");
if (server.appendonly) {
server.appendfd = open(server.appendfilename,O_WRONLY|O_APPEND|O_CREAT,0644);
}
} else if (!strcasecmp(argv[0],"loglevel") && argc == 2) {
if (!strcasecmp(argv[1],"debug")) server.verbosity = REDIS_DEBUG;
+ else if (!strcasecmp(argv[1],"verbose")) server.verbosity = REDIS_VERBOSE;
else if (!strcasecmp(argv[1],"notice")) server.verbosity = REDIS_NOTICE;
else if (!strcasecmp(argv[1],"warning")) server.verbosity = REDIS_WARNING;
else {
server.vm_page_size = strtoll(argv[1], NULL, 10);
} else if (!strcasecmp(argv[0],"vm-pages") && argc == 2) {
server.vm_pages = strtoll(argv[1], NULL, 10);
+ } else if (!strcasecmp(argv[0],"vm-max-threads") && argc == 2) {
+ server.vm_max_threads = strtoll(argv[1], NULL, 10);
} else {
err = "Bad directive or wrong number of arguments"; goto loaderr;
}
listRelease(c->reply);
freeClientArgv(c);
close(c->fd);
+ /* Remove from the list of clients */
ln = listSearchKey(server.clients,c);
redisAssert(ln != NULL);
listDelNode(server.clients,ln);
+ /* Remove from the list of clients waiting for VM operations */
+ if (server.vm_enabled && listLength(c->io_keys)) {
+ ln = listSearchKey(server.io_clients,c);
+ if (ln) listDelNode(server.io_clients,ln);
+ listRelease(c->io_keys);
+ }
+ listRelease(c->io_keys);
+ /* Other cleanup */
if (c->flags & REDIS_SLAVE) {
if (c->replstate == REDIS_REPL_SEND_BULK && c->repldbfd != -1)
close(c->repldbfd);
int copylen = 0;
char buf[GLUEREPLY_UP_TO];
listNode *ln;
+ listIter li;
robj *o;
- listRewind(c->reply);
- while((ln = listYield(c->reply))) {
+ listRewind(c->reply,&li);
+ while((ln = listNext(&li))) {
int objlen;
o = ln->value;
if (errno == EAGAIN) {
nwritten = 0;
} else {
- redisLog(REDIS_DEBUG,
+ redisLog(REDIS_VERBOSE,
"Error writing to client: %s", strerror(errno));
freeClient(c);
return;
/* write all collected blocks at once */
if((nwritten = writev(fd, iov, ion)) < 0) {
if (errno != EAGAIN) {
- redisLog(REDIS_DEBUG,
+ redisLog(REDIS_VERBOSE,
"Error writing to client: %s", strerror(errno));
freeClient(c);
return;
static void replicationFeedSlaves(list *slaves, struct redisCommand *cmd, int dictid, robj **argv, int argc) {
listNode *ln;
+ listIter li;
int outc = 0, j;
robj **outv;
/* (args*2)+1 is enough room for args, spaces, newlines */
* be sure to free objects if there is no slave in a replication state
* able to be feed with commands */
for (j = 0; j < outc; j++) incrRefCount(outv[j]);
- listRewind(slaves);
- while((ln = listYield(slaves))) {
+ listRewind(slaves,&li);
+ while((ln = listNext(&li))) {
redisClient *slave = ln->value;
/* Don't feed slaves that are still waiting for BGSAVE to start */
* would not be called at all, but after the execution of the first commands
* in the input buffer the client may be blocked, and the "goto again"
* will try to reiterate. The following line will make it return asap. */
- if (c->flags & REDIS_BLOCKED) return;
+ if (c->flags & REDIS_BLOCKED || c->flags & REDIS_IO_WAIT) return;
if (c->bulklen == -1) {
/* Read the first line of the query */
char *p = strchr(c->querybuf,'\n');
}
return;
} else if (sdslen(c->querybuf) >= REDIS_REQUEST_MAX_SIZE) {
- redisLog(REDIS_DEBUG, "Client protocol error");
+ redisLog(REDIS_VERBOSE, "Client protocol error");
freeClient(c);
return;
}
if (errno == EAGAIN) {
nread = 0;
} else {
- redisLog(REDIS_DEBUG, "Reading from client: %s",strerror(errno));
+ redisLog(REDIS_VERBOSE, "Reading from client: %s",strerror(errno));
freeClient(c);
return;
}
} else if (nread == 0) {
- redisLog(REDIS_DEBUG, "Client closed connection");
+ redisLog(REDIS_VERBOSE, "Client closed connection");
freeClient(c);
return;
}
c->authenticated = 0;
c->replstate = REDIS_REPL_NONE;
c->reply = listCreate();
- c->blockingkeys = NULL;
- c->blockingkeysnum = 0;
listSetFreeMethod(c->reply,decrRefCount);
listSetDupMethod(c->reply,dupClientReplyValue);
+ c->blockingkeys = NULL;
+ c->blockingkeysnum = 0;
+ c->io_keys = listCreate();
+ listSetFreeMethod(c->io_keys,decrRefCount);
if (aeCreateFileEvent(server.el, c->fd, AE_READABLE,
readQueryFromClient, c) == AE_ERR) {
freeClient(c);
c->replstate == REDIS_REPL_ONLINE) &&
aeCreateFileEvent(server.el, c->fd, AE_WRITABLE,
sendReplyToClient, c) == AE_ERR) return;
+
+ if (server.vm_enabled && obj->storage != REDIS_VM_MEMORY) {
+ obj = dupStringObject(obj);
+ obj->refcount = 0; /* getDecodedObject() will increment the refcount */
+ }
listAddNodeTail(c->reply,getDecodedObject(obj));
}
cfd = anetAccept(server.neterr, fd, cip, &cport);
if (cfd == AE_ERR) {
- redisLog(REDIS_DEBUG,"Accepting client connection: %s", server.neterr);
+ redisLog(REDIS_VERBOSE,"Accepting client connection: %s", server.neterr);
return;
}
- redisLog(REDIS_DEBUG,"Accepted %s:%d", cip, cport);
+ redisLog(REDIS_VERBOSE,"Accepted %s:%d", cip, cport);
if ((c = createClient(cfd)) == NULL) {
redisLog(REDIS_WARNING,"Error allocating resoures for the client");
close(cfd); /* May be already closed, just ingore errors */
static robj *createObject(int type, void *ptr) {
robj *o;
+ if (server.vm_enabled) pthread_mutex_lock(&server.obj_freelist_mutex);
if (listLength(server.objfreelist)) {
listNode *head = listFirst(server.objfreelist);
o = listNodeValue(head);
listDelNode(server.objfreelist,head);
+ if (server.vm_enabled) pthread_mutex_unlock(&server.obj_freelist_mutex);
} else {
if (server.vm_enabled) {
+ pthread_mutex_unlock(&server.obj_freelist_mutex);
o = zmalloc(sizeof(*o));
} else {
o = zmalloc(sizeof(*o)-sizeof(struct redisObjectVM));
o->ptr = ptr;
o->refcount = 1;
if (server.vm_enabled) {
+ /* Note that this code may run in the context of an I/O thread
+ * and accessing to server.unixtime in theory is an error
+ * (no locks). But in practice this is safe, and even if we read
+ * garbage Redis will not fail, as it's just a statistical info */
o->vm.atime = server.unixtime;
o->storage = REDIS_VM_MEMORY;
}
}
static robj *dupStringObject(robj *o) {
+ assert(o->encoding == REDIS_ENCODING_RAW);
return createStringObject(o->ptr,sdslen(o->ptr));
}
}
static void incrRefCount(robj *o) {
- assert(!server.vm_enabled || o->storage == REDIS_VM_MEMORY);
+ redisAssert(!server.vm_enabled || o->storage == REDIS_VM_MEMORY);
o->refcount++;
}
static void decrRefCount(void *obj) {
robj *o = obj;
- /* REDIS_VM_SWAPPED */
- if (server.vm_enabled && o->storage == REDIS_VM_SWAPPED) {
- assert(o->refcount == 1);
- assert(o->type == REDIS_STRING);
+ /* Object is swapped out, or in the process of being loaded. */
+ if (server.vm_enabled &&
+ (o->storage == REDIS_VM_SWAPPED || o->storage == REDIS_VM_LOADING))
+ {
+ if (o->storage == REDIS_VM_SWAPPED || o->storage == REDIS_VM_LOADING) {
+ redisAssert(o->refcount == 1);
+ }
+ if (o->storage == REDIS_VM_LOADING) vmCancelThreadedIOJob(obj);
+ redisAssert(o->type == REDIS_STRING);
freeStringObject(o);
vmMarkPagesFree(o->vm.page,o->vm.usedpages);
+ pthread_mutex_lock(&server.obj_freelist_mutex);
if (listLength(server.objfreelist) > REDIS_OBJFREELIST_MAX ||
!listAddNodeHead(server.objfreelist,o))
zfree(o);
+ pthread_mutex_unlock(&server.obj_freelist_mutex);
+ server.vm_stats_swapped_objects--;
return;
}
- /* REDIS_VM_MEMORY */
+ /* Object is in memory, or in the process of being swapped out. */
if (--(o->refcount) == 0) {
+ if (server.vm_enabled && o->storage == REDIS_VM_SWAPPING)
+ vmCancelThreadedIOJob(obj);
switch(o->type) {
case REDIS_STRING: freeStringObject(o); break;
case REDIS_LIST: freeListObject(o); break;
case REDIS_HASH: freeHashObject(o); break;
default: redisAssert(0 != 0); break;
}
+ if (server.vm_enabled) pthread_mutex_lock(&server.obj_freelist_mutex);
if (listLength(server.objfreelist) > REDIS_OBJFREELIST_MAX ||
!listAddNodeHead(server.objfreelist,o))
zfree(o);
+ if (server.vm_enabled) pthread_mutex_unlock(&server.obj_freelist_mutex);
}
}
robj *val = dictGetEntryVal(de);
if (server.vm_enabled) {
- if (key->storage == REDIS_VM_MEMORY) {
+ if (key->storage == REDIS_VM_MEMORY ||
+ key->storage == REDIS_VM_SWAPPING)
+ {
+ /* If we were swapping the object out, stop it, this key
+ * was requested. */
+ if (key->storage == REDIS_VM_SWAPPING)
+ vmCancelThreadedIOJob(key);
/* Update the access time of the key for the aging algorithm. */
key->vm.atime = server.unixtime;
} else {
/* Our value was swapped on disk. Bring it at home. */
- assert(val == NULL);
+ redisAssert(val == NULL);
val = vmLoadObject(key);
dictGetEntryVal(de) = val;
}
static int rdbSaveStringObject(FILE *fp, robj *obj) {
int retval;
- obj = getDecodedObject(obj);
- retval = rdbSaveStringObjectRaw(fp,obj);
- decrRefCount(obj);
+ /* Avoid incr/decr ref count business when possible.
+ * This plays well with copy-on-write given that we are probably
+ * in a child process (BGSAVE). Also this makes sure key objects
+ * of swapped objects are not incRefCount-ed (an assert does not allow
+ * this in order to avoid bugs) */
+ if (obj->encoding != REDIS_ENCODING_RAW) {
+ obj = getDecodedObject(obj);
+ retval = rdbSaveStringObjectRaw(fp,obj);
+ decrRefCount(obj);
+ } else {
+ retval = rdbSaveStringObjectRaw(fp,obj);
+ }
return retval;
}
} else if (o->type == REDIS_LIST) {
/* Save a list value */
list *list = o->ptr;
+ listIter li;
listNode *ln;
- listRewind(list);
if (rdbSaveLen(fp,listLength(list)) == -1) return -1;
- while((ln = listYield(list))) {
+ listRewind(list,&li);
+ while((ln = listNext(&li))) {
robj *eleobj = listNodeValue(ln);
if (rdbSaveStringObject(fp,eleobj) == -1) return -1;
* the rdbSaveObject() function. Currently we use a trick to get
* this length with very little changes to the code. In the future
* we could switch to a faster solution. */
-static off_t rdbSavedObjectLen(robj *o) {
- static FILE *fp = NULL;
-
- if (fp == NULL) fp = fopen("/dev/null","w");
- assert(fp != NULL);
-
+static off_t rdbSavedObjectLen(robj *o, FILE *fp) {
+ if (fp == NULL) fp = server.devnull;
rewind(fp);
assert(rdbSaveObject(fp,o) != 1);
return ftello(fp);
}
/* Return the number of pages required to save this object in the swap file */
-static off_t rdbSavedObjectPages(robj *o) {
- off_t bytes = rdbSavedObjectLen(o);
+static off_t rdbSavedObjectPages(robj *o, FILE *fp) {
+ off_t bytes = rdbSavedObjectLen(o,fp);
return (bytes+(server.vm_page_size-1))/server.vm_page_size;
}
if (rdbSaveType(fp,REDIS_EXPIRETIME) == -1) goto werr;
if (rdbSaveTime(fp,expiretime) == -1) goto werr;
}
- /* Save the key and associated value */
- if (rdbSaveType(fp,o->type) == -1) goto werr;
- if (rdbSaveStringObject(fp,key) == -1) goto werr;
- /* Save the actual value */
- if (rdbSaveObject(fp,o) == -1) goto werr;
+ /* Save the key and associated value. This requires special
+ * handling if the value is swapped out. */
+ if (!server.vm_enabled || key->storage == REDIS_VM_MEMORY ||
+ key->storage == REDIS_VM_SWAPPING) {
+ /* Save type, key, value */
+ if (rdbSaveType(fp,o->type) == -1) goto werr;
+ if (rdbSaveStringObject(fp,key) == -1) goto werr;
+ if (rdbSaveObject(fp,o) == -1) goto werr;
+ } else {
+ /* REDIS_VM_SWAPPED or REDIS_VM_LOADING */
+ robj *po;
+ /* Get a preview of the object in memory */
+ po = vmPreviewObject(key);
+ /* Save type, key, value */
+ if (rdbSaveType(fp,key->vtype) == -1) goto werr;
+ if (rdbSaveStringObject(fp,key) == -1) goto werr;
+ if (rdbSaveObject(fp,po) == -1) goto werr;
+ /* Remove the loaded object from memory */
+ decrRefCount(po);
+ }
}
dictReleaseIterator(di);
}
pid_t childpid;
if (server.bgsavechildpid != -1) return REDIS_ERR;
+ if (server.vm_enabled) waitZeroActiveThreads();
if ((childpid = fork()) == 0) {
/* Child */
close(server.fd);
redisDb *db = server.db+0;
char buf[1024];
time_t expiretime = -1, now = time(NULL);
+ long long loadedkeys = 0;
fp = fopen(filename,"r");
if (!fp) return REDIS_ERR;
expiretime = -1;
}
keyobj = o = NULL;
+ /* Handle swapping while loading big datasets when VM is on */
+ loadedkeys++;
+ if (server.vm_enabled && (loadedkeys % 5000) == 0) {
+ while (zmalloc_used_memory() > server.vm_max_memory) {
+ if (vmSwapOneObjectBlocking() == REDIS_ERR) break;
+ }
+ }
}
fclose(fp);
return REDIS_OK;
retval = dictAdd(c->db->dict,c->argv[1],c->argv[2]);
if (retval == DICT_ERR) {
if (!nx) {
+ /* If the key is about a swapped value, we want a new key object
+ * to overwrite the old. So we delete the old key in the database.
+ * This will also make sure that swap pages about the old object
+ * will be marked as free. */
+ if (deleteIfSwapped(c->db,c->argv[1]))
+ incrRefCount(c->argv[1]);
dictReplace(c->db->dict,c->argv[1],c->argv[2]);
incrRefCount(c->argv[2]);
} else {
if (sortval->type == REDIS_LIST) {
list *list = sortval->ptr;
listNode *ln;
+ listIter li;
- listRewind(list);
- while((ln = listYield(list))) {
+ listRewind(list,&li);
+ while((ln = listNext(&li))) {
robj *ele = ln->value;
vector[j].obj = ele;
vector[j].u.score = 0;
addReplySds(c,sdscatprintf(sdsempty(),"*%d\r\n",outputlen));
for (j = start; j <= end; j++) {
listNode *ln;
+ listIter li;
+
if (!getop) {
addReplyBulkLen(c,vector[j].obj);
addReply(c,vector[j].obj);
addReply(c,shared.crlf);
}
- listRewind(operations);
- while((ln = listYield(operations))) {
+ listRewind(operations,&li);
+ while((ln = listNext(&li))) {
redisSortOperation *sop = ln->value;
robj *val = lookupKeyByPattern(c->db,sop->pattern,
vector[j].obj);
/* STORE option specified, set the sorting result as a List object */
for (j = start; j <= end; j++) {
listNode *ln;
+ listIter li;
+
if (!getop) {
listAddNodeTail(listPtr,vector[j].obj);
incrRefCount(vector[j].obj);
}
- listRewind(operations);
- while((ln = listYield(operations))) {
+ listRewind(operations,&li);
+ while((ln = listNext(&li))) {
redisSortOperation *sop = ln->value;
robj *val = lookupKeyByPattern(c->db,sop->pattern,
vector[j].obj);
zfree(vector);
}
+/* Convert an amount of bytes into a human readable string in the form
+ * of 100B, 2G, 100M, 4K, and so forth. */
+static void bytesToHuman(char *s, unsigned long long n) {
+ double d;
+
+ if (n < 1024) {
+ /* Bytes */
+ sprintf(s,"%lluB",n);
+ return;
+ } else if (n < (1024*1024)) {
+ d = (double)n/(1024);
+ sprintf(s,"%.2fK",d);
+ } else if (n < (1024LL*1024*1024)) {
+ d = (double)n/(1024*1024);
+ sprintf(s,"%.2fM",d);
+ } else if (n < (1024LL*1024*1024*1024)) {
+ d = (double)n/(1024LL*1024*1024);
+ sprintf(s,"%.2fM",d);
+ }
+}
+
/* Create the string returned by the INFO command. This is decoupled
* by the INFO command itself as we need to report the same information
* on memory corruption problems. */
sds info;
time_t uptime = time(NULL)-server.stat_starttime;
int j;
-
+ char hmem[64];
+
+ bytesToHuman(hmem,server.usedmemory);
info = sdscatprintf(sdsempty(),
"redis_version:%s\r\n"
"arch_bits:%s\r\n"
"multiplexing_api:%s\r\n"
+ "process_id:%ld\r\n"
"uptime_in_seconds:%ld\r\n"
"uptime_in_days:%ld\r\n"
"connected_clients:%d\r\n"
"connected_slaves:%d\r\n"
"blocked_clients:%d\r\n"
"used_memory:%zu\r\n"
+ "used_memory_human:%s\r\n"
"changes_since_last_save:%lld\r\n"
"bgsave_in_progress:%d\r\n"
"last_save_time:%ld\r\n"
"bgrewriteaof_in_progress:%d\r\n"
"total_connections_received:%lld\r\n"
"total_commands_processed:%lld\r\n"
+ "vm_enabled:%d\r\n"
"role:%s\r\n"
,REDIS_VERSION,
(sizeof(long) == 8) ? "64" : "32",
aeGetApiName(),
+ (long) getpid(),
uptime,
uptime/(3600*24),
listLength(server.clients)-listLength(server.slaves),
listLength(server.slaves),
server.blockedclients,
server.usedmemory,
+ hmem,
server.dirty,
server.bgsavechildpid != -1,
server.lastsave,
server.bgrewritechildpid != -1,
server.stat_numconnections,
server.stat_numcommands,
+ server.vm_enabled != 0,
server.masterhost == NULL ? "master" : "slave"
);
if (server.masterhost) {
server.master ? ((int)(time(NULL)-server.master->lastinteraction)) : -1
);
}
+ if (server.vm_enabled) {
+ lockThreadedIO();
+ info = sdscatprintf(info,
+ "vm_conf_max_memory:%llu\r\n"
+ "vm_conf_page_size:%llu\r\n"
+ "vm_conf_pages:%llu\r\n"
+ "vm_stats_used_pages:%llu\r\n"
+ "vm_stats_swapped_objects:%llu\r\n"
+ "vm_stats_swappin_count:%llu\r\n"
+ "vm_stats_swappout_count:%llu\r\n"
+ "vm_stats_io_newjobs_len:%lu\r\n"
+ "vm_stats_io_processing_len:%lu\r\n"
+ "vm_stats_io_processed_len:%lu\r\n"
+ "vm_stats_io_waiting_clients:%lu\r\n"
+ "vm_stats_io_active_threads:%lu\r\n"
+ ,(unsigned long long) server.vm_max_memory,
+ (unsigned long long) server.vm_page_size,
+ (unsigned long long) server.vm_pages,
+ (unsigned long long) server.vm_stats_used_pages,
+ (unsigned long long) server.vm_stats_swapped_objects,
+ (unsigned long long) server.vm_stats_swapins,
+ (unsigned long long) server.vm_stats_swapouts,
+ (unsigned long) listLength(server.io_newjobs),
+ (unsigned long) listLength(server.io_processing),
+ (unsigned long) listLength(server.io_processed),
+ (unsigned long) listLength(server.io_clients),
+ (unsigned long) server.io_active_threads
+ );
+ unlockThreadedIO();
+ }
for (j = 0; j < server.dbnum; j++) {
long long keys, vkeys;
* registering differences since the server forked to save */
redisClient *slave;
listNode *ln;
+ listIter li;
- listRewind(server.slaves);
- while((ln = listYield(server.slaves))) {
+ listRewind(server.slaves,&li);
+ while((ln = listNext(&li))) {
slave = ln->value;
if (slave->replstate == REDIS_REPL_WAIT_BGSAVE_END) break;
}
return;
}
if ((nwritten = write(fd,buf,buflen)) == -1) {
- redisLog(REDIS_DEBUG,"Write error sending DB to slave: %s",
+ redisLog(REDIS_VERBOSE,"Write error sending DB to slave: %s",
strerror(errno));
freeClient(slave);
return;
static void updateSlavesWaitingBgsave(int bgsaveerr) {
listNode *ln;
int startbgsave = 0;
+ listIter li;
- listRewind(server.slaves);
- while((ln = listYield(server.slaves))) {
+ listRewind(server.slaves,&li);
+ while((ln = listNext(&li))) {
redisClient *slave = ln->value;
if (slave->replstate == REDIS_REPL_WAIT_BGSAVE_START) {
}
if (startbgsave) {
if (rdbSaveBackground(server.dbfilename) != REDIS_OK) {
- listRewind(server.slaves);
+ listIter li;
+
+ listRewind(server.slaves,&li);
redisLog(REDIS_WARNING,"SYNC failed. BGSAVE failed");
- while((ln = listYield(server.slaves))) {
+ while((ln = listNext(&li))) {
redisClient *slave = ln->value;
if (slave->replstate == REDIS_REPL_WAIT_BGSAVE_START)
/* ============================ Maxmemory directive ======================== */
+/* Try to free one object form the pre-allocated objects free list.
+ * This is useful under low mem conditions as by default we take 1 million
+ * free objects allocated. On success REDIS_OK is returned, otherwise
+ * REDIS_ERR. */
+static int tryFreeOneObjectFromFreelist(void) {
+ robj *o;
+
+ if (server.vm_enabled) pthread_mutex_lock(&server.obj_freelist_mutex);
+ if (listLength(server.objfreelist)) {
+ listNode *head = listFirst(server.objfreelist);
+ o = listNodeValue(head);
+ listDelNode(server.objfreelist,head);
+ if (server.vm_enabled) pthread_mutex_unlock(&server.obj_freelist_mutex);
+ zfree(o);
+ return REDIS_OK;
+ } else {
+ if (server.vm_enabled) pthread_mutex_unlock(&server.obj_freelist_mutex);
+ return REDIS_ERR;
+ }
+}
+
/* This function gets called when 'maxmemory' is set on the config file to limit
* the max memory used by the server, and we are out of memory.
* This function will try to, in order:
*/
static void freeMemoryIfNeeded(void) {
while (server.maxmemory && zmalloc_used_memory() > server.maxmemory) {
- if (listLength(server.objfreelist)) {
- robj *o;
-
- listNode *head = listFirst(server.objfreelist);
- o = listNodeValue(head);
- listDelNode(server.objfreelist,head);
- zfree(o);
- } else {
- int j, k, freed = 0;
-
- for (j = 0; j < server.dbnum; j++) {
- int minttl = -1;
- robj *minkey = NULL;
- struct dictEntry *de;
-
- if (dictSize(server.db[j].expires)) {
- freed = 1;
- /* From a sample of three keys drop the one nearest to
- * the natural expire */
- for (k = 0; k < 3; k++) {
- time_t t;
-
- de = dictGetRandomKey(server.db[j].expires);
- t = (time_t) dictGetEntryVal(de);
- if (minttl == -1 || t < minttl) {
- minkey = dictGetEntryKey(de);
- minttl = t;
- }
+ int j, k, freed = 0;
+
+ if (tryFreeOneObjectFromFreelist() == REDIS_OK) continue;
+ for (j = 0; j < server.dbnum; j++) {
+ int minttl = -1;
+ robj *minkey = NULL;
+ struct dictEntry *de;
+
+ if (dictSize(server.db[j].expires)) {
+ freed = 1;
+ /* From a sample of three keys drop the one nearest to
+ * the natural expire */
+ for (k = 0; k < 3; k++) {
+ time_t t;
+
+ de = dictGetRandomKey(server.db[j].expires);
+ t = (time_t) dictGetEntryVal(de);
+ if (minttl == -1 || t < minttl) {
+ minkey = dictGetEntryKey(de);
+ minttl = t;
}
- deleteKey(server.db+j,minkey);
}
+ deleteKey(server.db+j,minkey);
}
- if (!freed) return; /* nothing to free... */
}
+ if (!freed) return; /* nothing to free... */
}
}
struct redisClient *fakeClient;
FILE *fp = fopen(filename,"r");
struct redis_stat sb;
+ unsigned long long loadedkeys = 0;
if (redis_fstat(fileno(fp),&sb) != -1 && sb.st_size == 0)
return REDIS_ERR;
/* Clean up, ready for the next command */
for (j = 0; j < argc; j++) decrRefCount(argv[j]);
zfree(argv);
+ /* Handle swapping while loading big datasets when VM is on */
+ loadedkeys++;
+ if (server.vm_enabled && (loadedkeys % 5000) == 0) {
+ while (zmalloc_used_memory() > server.vm_max_memory) {
+ if (vmSwapOneObjectBlocking() == REDIS_ERR) break;
+ }
+ }
}
fclose(fp);
freeFakeClient(fakeClient);
/* Write an object into a file in the bulk format $<count>\r\n<payload>\r\n */
static int fwriteBulk(FILE *fp, robj *obj) {
char buf[128];
- obj = getDecodedObject(obj);
+ int decrrc = 0;
+
+ /* Avoid the incr/decr ref count business if possible to help
+ * copy-on-write (we are often in a child process when this function
+ * is called).
+ * Also makes sure that key objects don't get incrRefCount-ed when VM
+ * is enabled */
+ if (obj->encoding != REDIS_ENCODING_RAW) {
+ obj = getDecodedObject(obj);
+ decrrc = 1;
+ }
snprintf(buf,sizeof(buf),"$%ld\r\n",(long)sdslen(obj->ptr));
if (fwrite(buf,strlen(buf),1,fp) == 0) goto err;
if (sdslen(obj->ptr) && fwrite(obj->ptr,sdslen(obj->ptr),1,fp) == 0)
goto err;
if (fwrite("\r\n",2,1,fp) == 0) goto err;
- decrRefCount(obj);
+ if (decrrc) decrRefCount(obj);
return 1;
err:
- decrRefCount(obj);
+ if (decrrc) decrRefCount(obj);
return 0;
}
/* Iterate this DB writing every entry */
while((de = dictNext(di)) != NULL) {
- robj *key = dictGetEntryKey(de);
- robj *o = dictGetEntryVal(de);
- time_t expiretime = getExpire(db,key);
+ robj *key, *o;
+ time_t expiretime;
+ int swapped;
+
+ key = dictGetEntryKey(de);
+ /* If the value for this key is swapped, load a preview in memory.
+ * We use a "swapped" flag to remember if we need to free the
+ * value object instead to just increment the ref count anyway
+ * in order to avoid copy-on-write of pages if we are forked() */
+ if (!server.vm_enabled || key->storage == REDIS_VM_MEMORY ||
+ key->storage == REDIS_VM_SWAPPING) {
+ o = dictGetEntryVal(de);
+ swapped = 0;
+ } else {
+ o = vmPreviewObject(key);
+ swapped = 1;
+ }
+ expiretime = getExpire(db,key);
/* Save the key and associated value */
if (o->type == REDIS_STRING) {
/* Emit the RPUSHes needed to rebuild the list */
list *list = o->ptr;
listNode *ln;
+ listIter li;
- listRewind(list);
- while((ln = listYield(list))) {
+ listRewind(list,&li);
+ while((ln = listNext(&li))) {
char cmd[]="*3\r\n$5\r\nRPUSH\r\n";
robj *eleobj = listNodeValue(ln);
if (fwriteBulk(fp,key) == 0) goto werr;
if (fwriteBulkLong(fp,expiretime) == 0) goto werr;
}
+ if (swapped) decrRefCount(o);
}
dictReleaseIterator(di);
}
pid_t childpid;
if (server.bgrewritechildpid != -1) return REDIS_ERR;
+ if (server.vm_enabled) waitZeroActiveThreads();
if ((childpid = fork()) == 0) {
/* Child */
char tmpfile[256];
unlink(tmpfile);
}
-/* =============================== Virtual Memory =========================== */
+/* Virtual Memory is composed mainly of two subsystems:
+ * - Blocking Virutal Memory
+ * - Threaded Virtual Memory I/O
+ * The two parts are not fully decoupled, but functions are split among two
+ * different sections of the source code (delimited by comments) in order to
+ * make more clear what functionality is about the blocking VM and what about
+ * the threaded (not blocking) VM.
+ *
+ * Redis VM design:
+ *
+ * Redis VM is a blocking VM (one that blocks reading swapped values from
+ * disk into memory when a value swapped out is needed in memory) that is made
+ * unblocking by trying to examine the command argument vector in order to
+ * load in background values that will likely be needed in order to exec
+ * the command. The command is executed only once all the relevant keys
+ * are loaded into memory.
+ *
+ * This basically is almost as simple of a blocking VM, but almost as parallel
+ * as a fully non-blocking VM.
+ */
+
+/* =================== Virtual Memory - Blocking Side ====================== */
static void vmInit(void) {
off_t totsize;
+ int pipefds[2];
+ size_t stacksize;
+
+ if (server.vm_max_threads != 0)
+ zmalloc_enable_thread_safeness(); /* we need thread safe zmalloc() */
server.vm_fp = fopen("/tmp/redisvm","w+b");
if (server.vm_fp == NULL) {
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_NOTICE,"Swap file allocated with success");
}
server.vm_bitmap = zmalloc((server.vm_pages+7)/8);
- redisLog(REDIS_DEBUG,"Allocated %lld bytes page table for %lld pages",
+ 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");
+
+ /* Initialize threaded I/O (used by Virtual Memory) */
+ server.io_newjobs = listCreate();
+ server.io_processing = listCreate();
+ server.io_processed = listCreate();
+ server.io_clients = listCreate();
+ pthread_mutex_init(&server.io_mutex,NULL);
+ pthread_mutex_init(&server.obj_freelist_mutex,NULL);
+ pthread_mutex_init(&server.io_swapfile_mutex,NULL);
+ server.io_active_threads = 0;
+ if (pipe(pipefds) == -1) {
+ redisLog(REDIS_WARNING,"Unable to intialized VM: pipe(2): %s. Exiting."
+ ,strerror(errno));
+ exit(1);
+ }
+ server.io_ready_pipe_read = pipefds[0];
+ server.io_ready_pipe_write = pipefds[1];
+ redisAssert(anetNonBlock(NULL,server.io_ready_pipe_read) != ANET_ERR);
+ /* LZF requires a lot of stack */
+ pthread_attr_init(&server.io_threads_attr);
+ pthread_attr_getstacksize(&server.io_threads_attr, &stacksize);
+ while (stacksize < REDIS_THREAD_STACK_SIZE) stacksize *= 2;
+ pthread_attr_setstacksize(&server.io_threads_attr, stacksize);
+ /* Listen for events in the threaded I/O pipe */
+ if (aeCreateFileEvent(server.el, server.io_ready_pipe_read, AE_READABLE,
+ vmThreadedIOCompletedJob, NULL) == AE_ERR)
+ oom("creating file event");
}
/* Mark the page as used */
off_t byte = page/8;
int bit = page&7;
server.vm_bitmap[byte] |= 1<<bit;
- printf("Mark used: %lld (byte:%lld bit:%d)\n", (long long)page,
- (long long)byte, 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. */
for (j = 0; j < count; j++)
vmMarkPageUsed(page+j);
+ server.vm_stats_used_pages += count;
}
/* Mark the page as free */
for (j = 0; j < count; j++)
vmMarkPageFree(page+j);
+ server.vm_stats_used_pages -= count;
}
/* Test if the page is free */
* 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) {
+static int vmFindContiguousPages(off_t *first, off_t n) {
off_t base, offset = 0, since_jump = 0, numfree = 0;
if (server.vm_near_pages == REDIS_VM_MAX_NEAR_PAGES) {
while(offset < server.vm_pages) {
off_t this = base+offset;
- printf("THIS: %lld (%c)\n", (long long) this, vmFreePage(this) ? 'F' : 'X');
+ 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;
return REDIS_ERR;
}
+/* Write the specified object at the specified page of the swap file */
+static int vmWriteObjectOnSwap(robj *o, off_t page) {
+ if (server.vm_enabled) pthread_mutex_lock(&server.io_swapfile_mutex);
+ if (fseeko(server.vm_fp,page*server.vm_page_size,SEEK_SET) == -1) {
+ if (server.vm_enabled) pthread_mutex_unlock(&server.io_swapfile_mutex);
+ redisLog(REDIS_WARNING,
+ "Critical VM problem in vmSwapObjectBlocking(): can't seek: %s",
+ strerror(errno));
+ return REDIS_ERR;
+ }
+ rdbSaveObject(server.vm_fp,o);
+ if (server.vm_enabled) pthread_mutex_unlock(&server.io_swapfile_mutex);
+ return REDIS_OK;
+}
+
/* 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);
+static int vmSwapObjectBlocking(robj *key, robj *val) {
+ off_t pages = rdbSavedObjectPages(val,NULL);
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);
+ if (vmWriteObjectOnSwap(val,page) == REDIS_ERR) return REDIS_ERR;
key->vm.page = page;
key->vm.usedpages = pages;
key->storage = REDIS_VM_SWAPPED;
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++;
+ fflush(server.vm_fp);
return REDIS_OK;
}
-/* Load the value object relative to the 'key' object from swap to memory.
- * The newly allocated object is returned. */
-static robj *vmLoadObject(robj *key) {
- robj *val;
+static robj *vmReadObjectFromSwap(off_t page, int type) {
+ robj *o;
- assert(key->storage == REDIS_VM_SWAPPED);
- if (fseeko(server.vm_fp,key->vm.page*server.vm_page_size,SEEK_SET) == -1) {
+ if (server.vm_enabled) pthread_mutex_lock(&server.io_swapfile_mutex);
+ if (fseeko(server.vm_fp,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) {
+ o = rdbLoadObject(type,server.vm_fp);
+ if (o == NULL) {
redisLog(REDIS_WARNING, "Unrecoverable VM problem in vmLoadObject(): can't load object from swap file: %s", strerror(errno));
exit(1);
}
- 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);
+ if (server.vm_enabled) pthread_mutex_unlock(&server.io_swapfile_mutex);
+ return o;
+}
+
+/* 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);
+ val = vmReadObjectFromSwap(key->vm.page,key->vtype);
+ 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) {
+ /* If we are loading the object in background, stop it, we
+ * need to load this object synchronously ASAP. */
+ if (key->storage == REDIS_VM_LOADING)
+ vmCancelThreadedIOJob(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.
*
/* 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) {
+ * to swap any object at all.
+ *
+ * If 'usethreaded' is true, Redis will try to swap the object in background
+ * using I/O threads. */
+static int vmSwapOneObject(int usethreads) {
int j, i;
struct dictEntry *best = NULL;
double best_swappability = 0;
+ redisDb *best_db = NULL;
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) continue;
+ /* Only swap objects that are currently in memory.
+ *
+ * Also don't swap shared objects if threaded VM is on, as we
+ * try to ensure that the main thread does not touch the
+ * object while the I/O thread is using it, but we can't
+ * control other keys without adding additional mutex. */
+ if (key->storage != REDIS_VM_MEMORY ||
+ (server.vm_max_threads != 0 && val->refcount != 1)) {
+ if (maxtries) i--; /* don't count this try */
+ continue;
+ }
swappability = computeObjectSwappability(val);
if (!best || swappability > best_swappability) {
best = de;
best_swappability = swappability;
+ best_db = db;
}
}
}
- if (best == NULL) return REDIS_ERR;
+ 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\n",
+ redisLog(REDIS_DEBUG,"Key with best swappability: %s, %f",
key->ptr, best_swappability);
/* Unshare the key if needed */
key = dictGetEntryKey(best) = newkey;
}
/* Swap it */
- if (vmSwapObject(key,val) == REDIS_OK) {
- dictGetEntryVal(best) = NULL;
+ if (usethreads) {
+ vmSwapObjectThreaded(key,val,best_db);
return REDIS_OK;
} else {
- return REDIS_ERR;
+ if (vmSwapObjectBlocking(key,val) == REDIS_OK) {
+ dictGetEntryVal(best) = NULL;
+ return REDIS_OK;
+ } else {
+ return REDIS_ERR;
+ }
}
}
+static int vmSwapOneObjectBlocking() {
+ return vmSwapOneObject(0);
+}
+
+static int vmSwapOneObjectThreaded() {
+ return vmSwapOneObject(1);
+}
+
+/* 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;
+}
+
+/* =================== Virtual Memory - Threaded I/O ======================= */
+
+static void freeIOJob(iojob *j) {
+ if (j->type == REDIS_IOJOB_PREPARE_SWAP ||
+ j->type == REDIS_IOJOB_DO_SWAP)
+ 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;
+ int processed = 0;
+ 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);
+ 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) {
+ /* 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++;
+ freeIOJob(j);
+ } 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 (vmFindContiguousPages(&j->page,j->pages) == REDIS_ERR) {
+ /* Ooops... no space! */
+ freeIOJob(j);
+ } 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 (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) break;
+ }
+ }
+ }
+ processed++;
+ if (processed == REDIS_MAX_COMPLETED_JOBS_PROCESSED) 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) (LIST ID %d)\n",
+ (void*)job, (char*)o->ptr, i);
+ /* Mark the pages as free since the swap didn't happened
+ * or happened but is now discarded. */
+ if (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, and
+ * probably with the object if this is a
+ * PREPARE_SWAP or DO_SWAP job. Better to wait for the
+ * job to move into the next queue... */
+ if (job->type != REDIS_IOJOB_LOAD) {
+ /* Yes, 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;
+ } else {
+ job->canceled = 1;
+ break;
+ }
+ 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) {
+#ifdef REDIS_HELGRIND_FRIENDLY
+ /* No new jobs? Wait and retry, because to be Helgrind
+ * (valgrind --tool=helgrind) what's needed is to take
+ * the same threads running instead to create/destroy threads
+ * as needed (otherwise valgrind will fail) */
+ unlockThreadedIO();
+ usleep(1); /* Give some time for the I/O thread to work. */
+ continue;
+#endif
+ /* No new jobs in queue, exit. */
+ redisLog(REDIS_DEBUG,"Thread %lld exiting, nothing to do",
+ (long 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 %lld got a new job (type %d): %p about key '%s'",
+ (long long) pthread_self(), j->type, (void*)j, (char*)j->key->ptr);
+
+ /* Process the Job */
+ if (j->type == REDIS_IOJOB_LOAD) {
+ } 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 %lld completed the job: %p (key %s)",
+ (long 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;
+
+ pthread_create(&thread,&server.io_threads_attr,IOThreadEntryPoint,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 waitZeroActiveThreads(void) {
+ while(1) {
+ lockThreadedIO();
+ if (server.io_active_threads == 0) {
+ unlockThreadedIO();
+ return;
+ }
+ unlockThreadedIO();
+ usleep(10000); /* 10 milliseconds */
+ }
+}
+
+/* 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;
+}
+
/* ================================= Debugging ============================== */
static void debugCommand(redisClient *c) {
}
key = dictGetEntryKey(de);
val = dictGetEntryVal(de);
- addReplySds(c,sdscatprintf(sdsempty(),
- "+Key at:%p refcount:%d, value at:%p refcount:%d encoding:%d serializedlength:%lld\r\n",
+ if (server.vm_enabled && (key->storage == REDIS_VM_MEMORY ||
+ key->storage == REDIS_VM_SWAPPING)) {
+ addReplySds(c,sdscatprintf(sdsempty(),
+ "+Key at:%p refcount:%d, value at:%p refcount:%d "
+ "encoding:%d serializedlength:%lld\r\n",
(void*)key, key->refcount, (void*)val, val->refcount,
- val->encoding, rdbSavedObjectLen(val)));
+ val->encoding, rdbSavedObjectLen(val,NULL)));
+ } else {
+ addReplySds(c,sdscatprintf(sdsempty(),
+ "+Key at:%p refcount:%d, value swapped at: page %llu "
+ "using %llu pages\r\n",
+ (void*)key, key->refcount, (unsigned long long) key->vm.page,
+ (unsigned long long) key->vm.usedpages));
+ }
} else if (!strcasecmp(c->argv[1]->ptr,"swapout") && c->argc == 3) {
dictEntry *de = dictFind(c->db->dict,c->argv[2]);
robj *key, *val;
/* Swap it */
if (key->storage != REDIS_VM_MEMORY) {
addReplySds(c,sdsnew("-ERR This key is not in memory\r\n"));
- } else if (vmSwapObject(key,val) == REDIS_OK) {
+ } else if (vmSwapObjectBlocking(key,val) == REDIS_OK) {
dictGetEntryVal(de) = NULL;
addReply(c,shared.ok);
} else {
}
}
-static void _redisAssert(char *estr) {
+static void _redisAssert(char *estr, char *file, int line) {
redisLog(REDIS_WARNING,"=== ASSERTION FAILED ===");
- redisLog(REDIS_WARNING,"==> %s\n",estr);
+ redisLog(REDIS_WARNING,"==> %s:%d '%s' is not true\n",file,line,estr);
#ifdef HAVE_BACKTRACE
redisLog(REDIS_WARNING,"(forcing SIGSEGV in order to print the stack trace)");
*((char*)-1) = 'x';
FILE *fp;
if (fork() != 0) exit(0); /* parent exits */
- printf("New pid: %d\n", getpid());
setsid(); /* create a new session */
/* Every output goes to /dev/null. If Redis is daemonized but
if (rdbLoad(server.dbfilename) == REDIS_OK)
redisLog(REDIS_NOTICE,"DB loaded from disk");
}
- if (aeCreateFileEvent(server.el, server.fd, AE_READABLE,
- acceptHandler, NULL) == AE_ERR) oom("creating file event");
redisLog(REDIS_NOTICE,"The server is now ready to accept connections on port %d", server.port);
aeMain(server.el);
aeDeleteEventLoop(server.el);
projects / redis.git / blobdiff