Merge pull request #386 from EQ-Alpha/fix_rdb_hang

add readwrite lock for forking

Former-commit-id: dabb81960f6ccc7f62c53648127924fc2fef7cdc

src/module.cpp

@@ -365,11 +365,7 @@ typedef struct RedisModuleCommandFilter {
 static list *moduleCommandFilters;
 
 /* Module GIL Variables */
-static int s_cAcquisitionsServer = 0;
+static readWriteLock s_moduleGIL;
-static int s_cAcquisitionsModule = 0;
-static std::mutex s_mutex;
-static std::condition_variable s_cv;
-static std::recursive_mutex s_mutexModule;
 thread_local bool g_fModuleThread = false;
 
 typedef void (*RedisModuleForkDoneHandler) (int exitcode, int bysignal, void *user_data);
@@ -5969,95 +5965,58 @@ void RM_ThreadSafeContextUnlock(RedisModuleCtx *ctx) {
 //  as the server thread acquisition is sufficient.  If we did try to lock we would deadlock
 static bool FModuleCallBackLock(bool fServerThread)
 {
-    return !fServerThread && aeThreadOwnsLock() && !g_fModuleThread && s_cAcquisitionsServer > 0;
+    return !fServerThread && aeThreadOwnsLock() && !g_fModuleThread && s_moduleGIL.hasReader();
 }
 void moduleAcquireGIL(int fServerThread, int fExclusive) {
-    std::unique_lock<std::mutex> lock(s_mutex);
-    int *pcheck = fServerThread ? &s_cAcquisitionsModule : &s_cAcquisitionsServer;
-
     if (FModuleCallBackLock(fServerThread)) {
         return;
     }
 
-    while (*pcheck > 0)
-        s_cv.wait(lock);
-
     if (fServerThread)
     {
-        ++s_cAcquisitionsServer;
+        s_moduleGIL.acquireRead();
     }
     else
     {
-        // only try to acquire the mutexModule in exclusive mode 
+        s_moduleGIL.acquireWrite(fExclusive);
-        if (fExclusive){
-            // It is possible that another module thread holds the GIL (and s_mutexModule as a result). 
-            // When said thread goes to release the GIL, it will wait for s_mutex, which this thread owns. 
-            // This thread is however waiting for the GIL (and s_mutexModule) that the other thread owns.
-            // As a result, a deadlock has occured. 
-            // We release the lock on s_mutex and wait until we are able to safely acquire the GIL 
-            // in order to prevent this deadlock from occuring. 
-            while (!s_mutexModule.try_lock())
-                s_cv.wait(lock);     
-        }    
-        ++s_cAcquisitionsModule;
-        fModuleGILWlocked++;
     }
 }
 
 int moduleTryAcquireGIL(bool fServerThread, int fExclusive) {
-    std::unique_lock<std::mutex> lock(s_mutex, std::defer_lock);
-    if (!lock.try_lock())
-        return 1;
-    int *pcheck = fServerThread ? &s_cAcquisitionsModule : &s_cAcquisitionsServer;
-
     if (FModuleCallBackLock(fServerThread)) {
         return 0;
     }
 
-    if (*pcheck > 0)
-        return 1;
-
     if (fServerThread)
     {
-        ++s_cAcquisitionsServer;
+        if (!s_moduleGIL.tryAcquireRead())
+            return 1;
     }
     else
     {
-        // only try to acquire the mutexModule in exclusive mode 
+        if (!s_moduleGIL.tryAcquireWrite(fExclusive))
-        if (fExclusive){
-            if (!s_mutexModule.try_lock())
             return 1;
     }
-        ++s_cAcquisitionsModule;
-        fModuleGILWlocked++;
-    }
     return 0;
 }
 
 void moduleReleaseGIL(int fServerThread, int fExclusive) {
-    std::unique_lock<std::mutex> lock(s_mutex);
-
     if (FModuleCallBackLock(fServerThread)) {
         return;
     }
     
     if (fServerThread)
     {
-        --s_cAcquisitionsServer;
+        s_moduleGIL.releaseRead();
     }
     else
     {
-        // only try to release the mutexModule in exclusive mode
+        s_moduleGIL.releaseWrite(fExclusive);
-        if (fExclusive)
-            s_mutexModule.unlock();
-        --s_cAcquisitionsModule;
-        fModuleGILWlocked--;
     }
-    s_cv.notify_all();
 }
 
 int moduleGILAcquiredByModule(void) {
-    return fModuleGILWlocked > 0;
+    return s_moduleGIL.hasWriter();
 }
 
 

src/networking.cpp

@@ -3863,37 +3863,10 @@ int checkClientPauseTimeoutAndReturnIfPaused(void) {
  *
  * The function returns the total number of events processed. */
 void processEventsWhileBlocked(int iel) {
-    serverAssert(GlobalLocksAcquired());
+
+    int eventsCount = 0;
+    executeWithoutGlobalLock([&](){
         int iterations = 4; /* See the function top-comment. */
-
-    std::vector<client*> vecclients;
-    listIter li;
-    listNode *ln;
-    listRewind(g_pserver->clients, &li);
-
-    // All client locks must be acquired *after* the global lock is reacquired to prevent deadlocks
-    //  so unlock here, and save them for reacquisition later
-    while ((ln = listNext(&li)) != nullptr)
-    {
-        client *c = (client*)listNodeValue(ln);
-        if (c->lock.fOwnLock()) {
-            serverAssert(c->flags & CLIENT_PROTECTED);  // If the client is not protected we have no gurantee they won't be free'd in the event loop
-            c->lock.unlock();
-            vecclients.push_back(c);
-        }
-    }
-    
-    /* Since we're about to release our lock we need to flush the repl backlog queue */
-    bool fReplBacklog = g_pserver->repl_batch_offStart >= 0;
-    if (fReplBacklog) {
-        flushReplBacklogToClients();
-        g_pserver->repl_batch_idxStart = -1;
-        g_pserver->repl_batch_offStart = -1;
-    }
-
-    long long eventsCount = 0;
-    aeReleaseLock();
-    serverAssert(!GlobalLocksAcquired());
         try
         {
             ProcessingEventsWhileBlocked = 1;
@@ -3912,33 +3885,15 @@ void processEventsWhileBlocked(int iel) {
         }
         catch (...)
         {
-        // Caller expects us to be locked so fix and rethrow
             ProcessingEventsWhileBlocked = 0;
-        AeLocker locker;
-        locker.arm(nullptr);
-        locker.release();
-        for (client *c : vecclients)
-            c->lock.lock();
             throw;
         }
-    
+    });
-    AeLocker locker;
-    locker.arm(nullptr);
-    locker.release();
 
     g_pserver->events_processed_while_blocked += eventsCount;
 
     whileBlockedCron();
 
-    // Restore it so the calling code is not confused
-    if (fReplBacklog) {
-        g_pserver->repl_batch_idxStart = g_pserver->repl_backlog_idx;
-        g_pserver->repl_batch_offStart = g_pserver->master_repl_offset;
-    }
-
-    for (client *c : vecclients)
-        c->lock.lock();
-
     // If a different thread processed the shutdown we need to abort the lua command or we will hang
     if (serverTL->el->stop)
         throw ShutdownException();

src/readwritelock.h

@@ -0,0 +1,113 @@
+#pragma once
+#include <condition_variable>
+
+class readWriteLock {
+    std::mutex m_readLock;
+    std::recursive_mutex m_writeLock;
+    std::condition_variable m_cv;
+    int m_readCount = 0;
+    int m_writeCount = 0;
+    bool m_writeWaiting = false;
+public:
+    void acquireRead() {
+        std::unique_lock<std::mutex> rm(m_readLock);
+        while (m_writeCount > 0 || m_writeWaiting)
+            m_cv.wait(rm);
+        m_readCount++;
+    }
+    
+    bool tryAcquireRead() {
+        std::unique_lock<std::mutex> rm(m_readLock, std::defer_lock);
+        if (!rm.try_lock())
+            return false;
+        if (m_writeCount > 0 || m_writeWaiting)
+            return false;
+        m_readCount++;
+        return true;
+    }
+
+    void acquireWrite(bool exclusive = true) {
+        std::unique_lock<std::mutex> rm(m_readLock);
+        m_writeWaiting = true;
+        while (m_readCount > 0)
+            m_cv.wait(rm);
+        if (exclusive) {
+            /* Another thread might have the write lock while we have the read lock
+               but won't be able to release it until they can acquire the read lock
+               so release the read lock and try again instead of waiting to avoid deadlock */
+            while(!m_writeLock.try_lock())
+                m_cv.wait(rm);
+        }
+        m_writeCount++;
+        m_writeWaiting = false;
+    }
+
+    void upgradeWrite(bool exclusive = true) {
+        std::unique_lock<std::mutex> rm(m_readLock);
+        m_writeWaiting = true;
+        while (m_readCount > 1)
+            m_cv.wait(rm);
+        if (exclusive) {
+            /* Another thread might have the write lock while we have the read lock
+               but won't be able to release it until they can acquire the read lock
+               so release the read lock and try again instead of waiting to avoid deadlock */
+            while(!m_writeLock.try_lock())
+                m_cv.wait(rm);
+        }
+        m_writeCount++;
+        m_readCount--;
+        m_writeWaiting = false;
+    }
+
+    bool tryAcquireWrite(bool exclusive = true) {
+        std::unique_lock<std::mutex> rm(m_readLock, std::defer_lock);
+        if (!rm.try_lock())
+            return false;
+        if (m_readCount > 0)
+            return false;
+        if (exclusive)
+            if (!m_writeLock.try_lock())
+                return false;
+        m_writeCount++;
+        return true;
+    }
+
+    void releaseRead() {
+        std::unique_lock<std::mutex> rm(m_readLock);
+        serverAssert(m_readCount > 0);
+        m_readCount--;
+        m_cv.notify_all();
+    }
+
+    void releaseWrite(bool exclusive = true) {
+        std::unique_lock<std::mutex> rm(m_readLock);
+        serverAssert(m_writeCount > 0);
+        if (exclusive)
+            m_writeLock.unlock();
+        m_writeCount--;
+        m_cv.notify_all();
+    }
+
+    void downgradeWrite(bool exclusive = true) {
+        std::unique_lock<std::mutex> rm(m_readLock);
+        serverAssert(m_writeCount > 0);
+        if (exclusive)
+            m_writeLock.unlock();
+        m_writeCount--;
+        while (m_writeCount > 0 || m_writeWaiting)
+            m_cv.wait(rm);
+        m_readCount++;
+    }
+
+    bool hasReader() {
+        return m_readCount > 0;
+    }
+
+    bool hasWriter() {
+        return m_writeCount > 0;
+    }
+
+    bool writeWaiting() {
+        return m_writeWaiting;
+    }
+};

src/server.cpp

@@ -67,6 +67,7 @@
 #include "aelocker.h"
 #include "motd.h"
 #include "t_nhash.h"
+#include "readwritelock.h"
 #ifdef __linux__
 #include <sys/prctl.h>
 #include <sys/mman.h>
@@ -91,8 +92,10 @@ double R_Zero, R_PosInf, R_NegInf, R_Nan;
 /* Global vars */
 namespace GlobalHidden {
 struct redisServer server; /* Server global state */
+readWriteLock forkLock;
 }
 redisServer *g_pserver = &GlobalHidden::server;
+readWriteLock *g_forkLock = &GlobalHidden::forkLock;
 struct redisServerConst cserver;
 __thread struct redisServerThreadVars *serverTL = NULL;   // thread local server vars
 std::mutex time_thread_mutex;
@@ -2649,8 +2652,8 @@ void afterSleep(struct aeEventLoop *eventLoop) {
        Don't check here that modules are enabled, rather use the result from beforeSleep
        Otherwise you may double acquire the GIL and cause deadlocks in the module */
     if (!ProcessingEventsWhileBlocked) {
-        wakeTimeThread();
         if (serverTL->modulesEnabledThisAeLoop) moduleAcquireGIL(TRUE /*fServerThread*/);
+        wakeTimeThread();
     }
 }
 
@@ -6218,6 +6221,63 @@ void closeChildUnusedResourceAfterFork() {
     cserver.pidfile = NULL;
 }
 
+void executeWithoutGlobalLock(std::function<void()> func) {
+    serverAssert(GlobalLocksAcquired());
+
+    std::vector<client*> vecclients;
+    listIter li;
+    listNode *ln;
+    listRewind(g_pserver->clients, &li);
+
+    // All client locks must be acquired *after* the global lock is reacquired to prevent deadlocks
+    //  so unlock here, and save them for reacquisition later
+    while ((ln = listNext(&li)) != nullptr)
+    {
+        client *c = (client*)listNodeValue(ln);
+        if (c->lock.fOwnLock()) {
+            serverAssert(c->flags & CLIENT_PROTECTED || c->flags & CLIENT_EXECUTING_COMMAND);  // If the client is not protected we have no gurantee they won't be free'd in the event loop
+            c->lock.unlock();
+            vecclients.push_back(c);
+        }
+    }
+    
+    /* Since we're about to release our lock we need to flush the repl backlog queue */
+    bool fReplBacklog = g_pserver->repl_batch_offStart >= 0;
+    if (fReplBacklog) {
+        flushReplBacklogToClients();
+        g_pserver->repl_batch_idxStart = -1;
+        g_pserver->repl_batch_offStart = -1;
+    }
+
+    aeReleaseLock();
+    serverAssert(!GlobalLocksAcquired());
+    try {
+        func();
+    }
+    catch (...) {
+        // Caller expects us to be locked so fix and rethrow
+        AeLocker locker;
+        locker.arm(nullptr);
+        locker.release();
+        for (client *c : vecclients)
+            c->lock.lock();
+        throw;
+    }
+    
+    AeLocker locker;
+    locker.arm(nullptr);
+    locker.release();
+
+    // Restore it so the calling code is not confused
+    if (fReplBacklog) {
+        g_pserver->repl_batch_idxStart = g_pserver->repl_backlog_idx;
+        g_pserver->repl_batch_offStart = g_pserver->master_repl_offset;
+    }
+
+    for (client *c : vecclients)
+        c->lock.lock();
+}
+
 /* purpose is one of CHILD_TYPE_ types */
 int redisFork(int purpose) {
     int childpid;
@@ -6229,7 +6289,9 @@ int redisFork(int purpose) {
 
         openChildInfoPipe();
     }
-    
+    long long startWriteLock = ustime();
+    g_forkLock->acquireWrite();
+    latencyAddSampleIfNeeded("fork-lock",(ustime()-startWriteLock)/1000);
     if ((childpid = fork()) == 0) {
         /* Child */
         g_pserver->in_fork_child = purpose;
@@ -6238,6 +6300,7 @@ int redisFork(int purpose) {
         closeChildUnusedResourceAfterFork();
     } else {
         /* Parent */
+        g_forkLock->releaseWrite();
         g_pserver->stat_total_forks++;
         g_pserver->stat_fork_time = ustime()-start;
         g_pserver->stat_fork_rate = (double) zmalloc_used_memory() * 1000000 / g_pserver->stat_fork_time / (1024*1024*1024); /* GB per second. */
@@ -6554,20 +6617,32 @@ void *timeThreadMain(void*) {
     timespec delay;
     delay.tv_sec = 0;
     delay.tv_nsec = 100;
+    int cycle_count = 0;
+    g_forkLock->acquireRead();
     while (true) {
         {
             std::unique_lock<std::mutex> lock(time_thread_mutex);
             if (sleeping_threads >= cserver.cthreads) {
+                g_forkLock->releaseRead();
                 time_thread_cv.wait(lock);
+                g_forkLock->acquireRead();
+                cycle_count = 0;
             }
         }
         updateCachedTime();
+        if (cycle_count == MAX_CYCLES_TO_HOLD_FORK_LOCK) {
+            g_forkLock->releaseRead();
+            g_forkLock->acquireRead();
+            cycle_count = 0;
+        }
 #if defined(__APPLE__)
         nanosleep(&delay, nullptr);
 #else
         clock_nanosleep(CLOCK_MONOTONIC, 0, &delay, NULL);
 #endif
+        cycle_count++;
     }
+    g_forkLock->releaseRead();
 }
 
 void *workerThreadMain(void *parg)

src/server.h

@@ -96,6 +96,7 @@ typedef long long ustime_t; /* microsecond time type. */
 #include "connection.h" /* Connection abstraction */
 #include "serverassert.h"
 #include "expire.h"
+#include "readwritelock.h"
 
 #define REDISMODULE_CORE 1
 #include "redismodule.h"    /* Redis modules API defines. */
@@ -733,6 +734,9 @@ typedef enum {
 #define REDISMODULE_AUX_BEFORE_RDB (1<<0)
 #define REDISMODULE_AUX_AFTER_RDB (1<<1)
 
+/* Number of cycles before time thread gives up fork lock */
+#define MAX_CYCLES_TO_HOLD_FORK_LOCK 10
+
 struct RedisModule;
 struct RedisModuleIO;
 struct RedisModuleDigest;
@@ -2113,6 +2117,7 @@ typedef struct {
 
 //extern struct redisServer server;
 extern redisServer *g_pserver;
+extern readWriteLock *g_forkLock;
 extern struct redisServerConst cserver;
 extern __thread struct redisServerThreadVars *serverTL;   // thread local server vars
 extern struct sharedObjectsStruct shared;
@@ -2496,6 +2501,7 @@ void sendChildInfo(childInfoType info_type, size_t keys, const char *pname);
 void receiveChildInfo(void);
 
 /* Fork helpers */
+void executeWithoutGlobalLock(std::function<void()> func);
 int redisFork(int type);
 int hasActiveChildProcess();
 void resetChildState();