First attempt at reducing mutex contention in the unix VFS by providing

a separate mutex for each unixInodeInfo object.

FossilOrigin-Name: f69afaf00a659cd768a863c5f4a1d527f01187e037240da0ada527cb52c5afe8

diff --git a/manifest b/manifest
index 6af55b184ca42dcd53f02b9dfbcce34e28a1c90e..572f99e74782b831aedbdb9de2ace4d0b6b03329 100644 (file)
--- a/manifest
+++ b/manifest
@@ -1,5 +1,5 @@
-C Fix\stypo\sof\s'SQLITE_OMIT_WINDOWFUNC'.
-D 2018-07-23T18:53:49.829
+C First\sattempt\sat\sreducing\smutex\scontention\sin\sthe\sunix\sVFS\sby\sproviding\na\sseparate\smutex\sfor\seach\sunixInodeInfo\sobject.
+D 2018-07-23T21:10:37.802
 F .fossil-settings/empty-dirs dbb81e8fc0401ac46a1491ab34a7f2c7c0452f2f06b54ebb845d024ca8283ef1
 F .fossil-settings/ignore-glob 35175cdfcf539b2318cb04a9901442804be81cd677d8b889fcc9149c21f239ea
 F Makefile.in 0a3a6c81e6fcb969ff9106e882f0a08547014ba463cb6beca4c4efaecc924ee6
@@ -481,7 +481,7 @@ F src/os.c 8aeb0b0f40f8f5b0da03fe49706695adaf42d2f516ab95abc72e86c245e119de
 F src/os.h 48388821692e87da174ea198bf96b1b2d9d83be5dfc908f673ee21fafbe0d432
 F src/os_common.h b2f4707a603e36811d9b1a13278bffd757857b85
 F src/os_setup.h 0dbaea40a7d36bf311613d31342e0b99e2536586
-F src/os_unix.c c230a7a24766320d8414afd087edcd43e499fb45e86361f6f4f464f343d965a9
+F src/os_unix.c c275a32d1112973457908bee494ae0444468c0b96c07057913ade6d69781dab3
 F src/os_win.c 070cdbb400097c6cda54aa005356095afdc2f3ee691d17192c54724ef146a971
 F src/os_win.h 7b073010f1451abe501be30d12f6bc599824944a
 F src/pager.c 76d29b8a960dcb8b67210f095899d91e4a90673a6674ea58cfd1115b705a7fb9
@@ -1750,7 +1750,10 @@ F vsixtest/vsixtest.tcl 6a9a6ab600c25a91a7acc6293828957a386a8a93
 F vsixtest/vsixtest.vcxproj.data 2ed517e100c66dc455b492e1a33350c1b20fbcdc
 F vsixtest/vsixtest.vcxproj.filters 37e51ffedcdb064aad6ff33b6148725226cd608e
 F vsixtest/vsixtest_TemporaryKey.pfx e5b1b036facdb453873e7084e1cae9102ccc67a0
-P 46dd076af5880e6e9a7de974758b26bc43d0dc074f8143234038fa9cdcc726ee
-R 386d9db3da1c3cf3e08fdb519ac6201a
-U mistachkin
-Z 0f7169e4827257e93fc7e2c2fe7438e0
+P 8607b84a5c53d8c9218c90802160a3000f1138b1f188e411a3af46253e0a69cc
+R acece63ae9d3f4b29a8a0d5614555a1e
+T *branch * separate-lock-mutex
+T *sym-separate-lock-mutex *
+T -sym-trunk *
+U drh
+Z c02c3e5d8cc1425b437d3294ad5c5e34

diff --git a/manifest.uuid b/manifest.uuid
index 800ba6509cb1cafec25ee17fcf8fb891d3bf8a48..de8e1e71e2640ab435328059a6d0c436b3dd97c9 100644 (file)
--- a/manifest.uuid
+++ b/manifest.uuid
@@ -1 +1 @@
-8607b84a5c53d8c9218c90802160a3000f1138b1f188e411a3af46253e0a69cc
\ No newline at end of file
+f69afaf00a659cd768a863c5f4a1d527f01187e037240da0ada527cb52c5afe8
\ No newline at end of file

diff --git a/src/os_unix.c b/src/os_unix.c
index 0dd461da0a241ad2bb23327728b192e3be0baadc..48067f4cc35f9dec119027dfa243ef2bd344cf8d 100644 (file)
--- a/src/os_unix.c
+++ b/src/os_unix.c
@@ -1108,15 +1108,29 @@ struct unixFileId {
 ** A single inode can have multiple file descriptors, so each unixFile
 ** structure contains a pointer to an instance of this object and this
 ** object keeps a count of the number of unixFile pointing to it.
+**
+** Mutex rules:
+**
+**  (1) The pLockMutex mutex must be held in order to read or write
+**      any of the locking fields:
+**          nShared, nLock, eFileLock, or bProcessLock
+**
+**  (2) When nRef>0, then the following fields are unchanging and can
+**      be read (but not written) without holding any mutex:
+**          fileId, pLockMutex
+**
+**  (3) With the exceptions above, all the fields may only be read
+**      or written while holding the global unixBigLock mutex.
 */
 struct unixInodeInfo {
   struct unixFileId fileId;       /* The lookup key */
-  int nShared;                    /* Number of SHARED locks held */
-  unsigned char eFileLock;        /* One of SHARED_LOCK, RESERVED_LOCK etc. */
-  unsigned char bProcessLock;     /* An exclusive process lock is held */
+  sqlite3_mutex *pLockMutex;      /* Hold this mutex for... */
+  int nShared;                      /* Number of SHARED locks held */
+  int nLock;                        /* Number of outstanding file locks */
+  unsigned char eFileLock;          /* One of SHARED_LOCK, RESERVED_LOCK etc. */
+  unsigned char bProcessLock;       /* An exclusive process lock is held */
   int nRef;                       /* Number of pointers to this structure */
   unixShmNode *pShmNode;          /* Shared memory associated with this inode */
-  int nLock;                      /* Number of outstanding file locks */
   UnixUnusedFd *pUnused;          /* Unused file descriptors to close */
   unixInodeInfo *pNext;           /* List of all unixInodeInfo objects */
   unixInodeInfo *pPrev;           /*    .... doubly linked */
@@ -1273,6 +1287,7 @@ static void releaseInodeInfo(unixFile *pFile){
         assert( pInode->pNext->pPrev==pInode );
         pInode->pNext->pPrev = pInode->pPrev;
       }
+      sqlite3_mutex_free(pInode->pLockMutex);
       sqlite3_free(pInode);
     }
   }
@@ -1358,6 +1373,7 @@ static int findInodeInfo(
     }
     memset(pInode, 0, sizeof(*pInode));
     memcpy(&pInode->fileId, &fileId, sizeof(fileId));
+    pInode->pLockMutex = sqlite3_mutex_alloc(SQLITE_MUTEX_FAST);
     pInode->nRef = 1;
     pInode->pNext = inodeList;
     pInode->pPrev = 0;
@@ -1436,7 +1452,7 @@ static int unixCheckReservedLock(sqlite3_file *id, int *pResOut){
 
   assert( pFile );
   assert( pFile->eFileLock<=SHARED_LOCK );
-  unixEnterMutex(); /* Because pFile->pInode is shared across threads */
+  sqlite3_mutex_enter(pFile->pInode->pLockMutex);
 
   /* Check if a thread in this process holds such a lock */
   if( pFile->pInode->eFileLock>SHARED_LOCK ){
@@ -1461,7 +1477,7 @@ static int unixCheckReservedLock(sqlite3_file *id, int *pResOut){
   }
 #endif
   
-  unixLeaveMutex();
+  sqlite3_mutex_leave(pFile->pInode->pLockMutex);
   OSTRACE(("TEST WR-LOCK %d %d %d (unix)\n", pFile->h, rc, reserved));
 
   *pResOut = reserved;
@@ -1527,8 +1543,8 @@ static int osSetPosixAdvisoryLock(
 static int unixFileLock(unixFile *pFile, struct flock *pLock){
   int rc;
   unixInodeInfo *pInode = pFile->pInode;
-  assert( unixMutexHeld() );
   assert( pInode!=0 );
+  assert( sqlite3_mutex_held(pInode->pLockMutex) );
   if( (pFile->ctrlFlags & (UNIXFILE_EXCL|UNIXFILE_RDONLY))==UNIXFILE_EXCL ){
     if( pInode->bProcessLock==0 ){
       struct flock lock;
@@ -1647,8 +1663,8 @@ static int unixLock(sqlite3_file *id, int eFileLock){
 
   /* This mutex is needed because pFile->pInode is shared across threads
   */
-  unixEnterMutex();
   pInode = pFile->pInode;
+  sqlite3_mutex_enter(pInode->pLockMutex);
 
   /* If some thread using this PID has a lock via a different unixFile*
   ** handle that precludes the requested lock, return BUSY.
@@ -1791,7 +1807,7 @@ static int unixLock(sqlite3_file *id, int eFileLock){
   }
 
 end_lock:
-  unixLeaveMutex();
+  sqlite3_mutex_leave(pInode->pLockMutex);
   OSTRACE(("LOCK    %d %s %s (unix)\n", pFile->h, azFileLock(eFileLock), 
       rc==SQLITE_OK ? "ok" : "failed"));
   return rc;
@@ -1839,8 +1855,8 @@ static int posixUnlock(sqlite3_file *id, int eFileLock, int handleNFSUnlock){
   if( pFile->eFileLock<=eFileLock ){
     return SQLITE_OK;
   }
-  unixEnterMutex();
   pInode = pFile->pInode;
+  sqlite3_mutex_enter(pInode->pLockMutex);
   assert( pInode->nShared!=0 );
   if( pFile->eFileLock>SHARED_LOCK ){
     assert( pInode->eFileLock==pFile->eFileLock );
@@ -1972,7 +1988,7 @@ static int posixUnlock(sqlite3_file *id, int eFileLock, int handleNFSUnlock){
   }
 
 end_unlock:
-  unixLeaveMutex();
+  sqlite3_mutex_leave(pInode->pLockMutex);
   if( rc==SQLITE_OK ) pFile->eFileLock = eFileLock;
   return rc;
 }
@@ -2771,8 +2787,7 @@ static int afpCheckReservedLock(sqlite3_file *id, int *pResOut){
     *pResOut = 1;
     return SQLITE_OK;
   }
-  unixEnterMutex(); /* Because pFile->pInode is shared across threads */
-  
+  sqlite3_mutex_enter(pFile->pInode->pLockMutex);
   /* Check if a thread in this process holds such a lock */
   if( pFile->pInode->eFileLock>SHARED_LOCK ){
     reserved = 1;
@@ -2796,7 +2811,7 @@ static int afpCheckReservedLock(sqlite3_file *id, int *pResOut){
     }
   }
   
-  unixLeaveMutex();
+  sqlite3_mutex_leave(pFile->pInode->pLockMutex);
   OSTRACE(("TEST WR-LOCK %d %d %d (afp)\n", pFile->h, rc, reserved));
   
   *pResOut = reserved;
@@ -2859,8 +2874,8 @@ static int afpLock(sqlite3_file *id, int eFileLock){
   
   /* This mutex is needed because pFile->pInode is shared across threads
   */
-  unixEnterMutex();
   pInode = pFile->pInode;
+  sqlite3_mutex_enter(pInode->pLockMutex);
 
   /* If some thread using this PID has a lock via a different unixFile*
   ** handle that precludes the requested lock, return BUSY.
@@ -2996,7 +3011,7 @@ static int afpLock(sqlite3_file *id, int eFileLock){
   }
   
 afp_end_lock:
-  unixLeaveMutex();
+  sqlite3_mutex_leave(pInode->pLockMutex);
   OSTRACE(("LOCK    %d %s %s (afp)\n", pFile->h, azFileLock(eFileLock), 
          rc==SQLITE_OK ? "ok" : "failed"));
   return rc;
@@ -3028,8 +3043,8 @@ static int afpUnlock(sqlite3_file *id, int eFileLock) {
   if( pFile->eFileLock<=eFileLock ){
     return SQLITE_OK;
   }
-  unixEnterMutex();
   pInode = pFile->pInode;
+  sqlite3_mutex_enter(pInode->pLockMutex);
   assert( pInode->nShared!=0 );
   if( pFile->eFileLock>SHARED_LOCK ){
     assert( pInode->eFileLock==pFile->eFileLock );
@@ -3104,7 +3119,7 @@ static int afpUnlock(sqlite3_file *id, int eFileLock) {
     }
   }
   
-  unixLeaveMutex();
+  sqlite3_mutex_leave(pInode->pLockMutex);
   if( rc==SQLITE_OK ) pFile->eFileLock = eFileLock;
   return rc;
 }