SERVER-14129: Use fsync instead of msync for Solaris

3 files changed, 56 insertions, 4 deletions

diff --git a/src/mongo/util/mmap_posix.cpp b/src/mongo/util/mmap_posix.cpp
index 52c64249f50..20b221a885c 100644
--- a/src/mongo/util/mmap_posix.cpp
+++ b/src/mongo/util/mmap_posix.cpp
@@ -252,9 +252,14 @@ namespace mongo {
     void MemoryMappedFile::flush(bool sync) {
         if ( views.empty() || fd == 0 )
             return;
-        if ( msync(viewForFlushing(), len, sync ? MS_SYNC : MS_ASYNC) ) {
+
+        bool useFsync = sync && !ProcessInfo::preferMsyncOverFSync();
+
+        if ( useFsync ?
+            fsync(fd) != 0 :
+            msync(viewForFlushing(), len, sync ? MS_SYNC : MS_ASYNC) ) {
             // msync failed, this is very bad
-            log() << "msync failed: " << errnoWithDescription()
+            log() << (useFsync ? "fsync failed: " : "msync failed: ") << errnoWithDescription()
                   << " file: " << filename() << endl;
             dataSyncFailedHandler();
         }
@@ -270,8 +275,11 @@ namespace mongo {
             if ( _view == NULL || _fd == 0 )
                 return;
 
-            if ( msync(_view, _len, MS_SYNC ) == 0 )
+            if ( ProcessInfo::preferMsyncOverFSync() ?
+                msync(_view, _len, MS_SYNC ) == 0 :
+                fsync(_fd) == 0 ) {
                 return;
+            }
 
             if ( errno == EBADF ) {
                 // ok, we were unlocked, so this file was closed
diff --git a/src/mongo/util/processinfo.h b/src/mongo/util/processinfo.h
index d41b085a6eb..d415bf2eea7 100644
--- a/src/mongo/util/processinfo.h
+++ b/src/mongo/util/processinfo.h
@@ -103,6 +103,11 @@ namespace mongo {
         static bool isDataFileZeroingNeeded() { return systemInfo->fileZeroNeeded;  }
 
         /**
+         * Determine if we need to workaround slow msync performance on Illumos/Solaris
+         */
+        static bool preferMsyncOverFSync() { return systemInfo->preferMsyncOverFSync;  }
+
+        /**
          * Get extra system stats
          */
         void appendSystemDetails( BSONObjBuilder& details ) const {
@@ -162,13 +167,21 @@ namespace mongo {
             //
             bool fileZeroNeeded;
 
+            // On non-Solaris (ie, Linux, Darwin, *BSD) kernels, prefer msync.
+            // Illumos kernels do O(N) scans in memory of the page table during msync which
+            // causes high CPU, Oracle Solaris 11.2 and later modified ZFS to workaround mongodb
+            // Oracle Solaris Bug: 
+            //  18658199 Speed up msync() on ZFS by 90000x with this one weird trick
+            bool preferMsyncOverFSync;
+
             SystemInfo() :
                     addrSize( 0 ),
                     memSize( 0 ),
                     numCores( 0 ),
                     pageSize( 0 ),
                     hasNuma( false ),
-                    fileZeroNeeded (false) { 
+                    fileZeroNeeded (false), 
+                    preferMsyncOverFSync (true) { 
                 // populate SystemInfo during construction
                 collectSystemInfo();
             }
diff --git a/src/mongo/util/processinfo_sunos5.cpp b/src/mongo/util/processinfo_sunos5.cpp
index 37629051b20..4b989dbea91 100644
--- a/src/mongo/util/processinfo_sunos5.cpp
+++ b/src/mongo/util/processinfo_sunos5.cpp
@@ -44,6 +44,7 @@
 #include "mongo/util/mongoutils/str.h"
 #include "mongo/util/processinfo.h"
 #include "mongo/util/scopeguard.h"
+#include "mongo/util/stringutils.h"
 
 namespace mongo {
 
@@ -145,6 +146,36 @@ namespace mongo {
         cpuArch = unameData.machine;
         hasNuma = checkNumaEnabled();
 
+        // We prefer FSync over msync, when:
+        // 1. Pre-Oracle Solaris 11.2 releases
+        // 2. Illumos kernel releases (which is all non Oracle Solaris releases)
+        preferMsyncOverFSync = false;
+
+        if (mongoutils::str::startsWith(osName, "Oracle Solaris")) {
+
+            std::vector<std::string> versionComponents;
+            splitStringDelim(osVersion, &versionComponents, '.');
+
+            if (versionComponents.size() > 1) {
+                unsigned majorInt, minorInt;
+                Status majorStatus =
+                    parseNumberFromString<unsigned>(versionComponents[0], &majorInt);
+
+                Status minorStatus =
+                    parseNumberFromString<unsigned>(versionComponents[1], &minorInt);
+
+                if (!majorStatus.Ok() || !minorStatus.Ok()) {
+                    warning() << "Could not parse OS version numbers from uname: " << osVersion;
+                }
+                else if ((majorInt == 11 && minorInt >= 2) || major > 11) {
+                    preferMsyncOverFSync = true;
+                }
+            }
+            else {
+                warning() << "Could not parse OS version string from uname: " << osVersion;
+            }
+        }
+
         BSONObjBuilder bExtra;
         bExtra.append("kernelVersion", unameData.release);
         bExtra.append("pageSize", static_cast<long long>(pageSize));