bool run(OperationContext* txn, const string& dbname, BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool fromRepl ) {
boost::filesystem::path p = dur::getJournalDir();
p /= "journalLatencyTest";
// remove file if already present
try {
boost::filesystem::remove(p);
}
catch(...) { }
BSONObjBuilder bb[2];
for( int pass = 0; pass < 2; pass++ ) {
LogFile f(p.string());
AlignedBuilder b(1024 * 1024);
{
Timer t;
for( int i = 0 ; i < 100; i++ ) {
f.synchronousAppend(b.buf(), 8192);
}
bb[pass].append("8KB", t.millis() / 100.0);
}
{
const int N = 50;
Timer t2;
long long x = 0;
for( int i = 0 ; i < N; i++ ) {
Timer t;
f.synchronousAppend(b.buf(), 8192);
x += t.micros();
sleepmillis(4);
}
long long y = t2.micros() - 4*N*1000;
// not really trusting the timer granularity on all platforms so whichever is higher of x and y
bb[pass].append("8KBWithPauses", max(x,y) / (N*1000.0));
}
{
Timer t;
for( int i = 0 ; i < 20; i++ ) {
f.synchronousAppend(b.buf(), 1024 * 1024);
}
bb[pass].append("1MB", t.millis() / 20.0);
}
// second time around, we are prealloced.
}
result.append("timeMillis", bb[0].obj());
result.append("timeMillisWithPrealloc", bb[1].obj());
try {
remove(p);
}
catch(...) { }
try {
result.append("onSamePartition", onSamePartition(dur::getJournalDir().string(),
storageGlobalParams.dbpath));
}
catch(...) { }
return 1;
}
/**
* The main durability thread loop. There is a single instance of this function running.
*/
static void durThread(ClockSource* cs, int64_t serverStartMs) {
Client::initThread("durability");
log() << "Durability thread started";
bool samePartition = true;
try {
const std::string dbpathDir = boost::filesystem::path(storageGlobalParams.dbpath).string();
samePartition = onSamePartition(getJournalDir().string(), dbpathDir);
} catch (...) {
}
// Spawn the journal writer thread
JournalWriter journalWriter(&commitNotify, &applyToDataFilesNotify, NumAsyncJournalWrites);
journalWriter.start();
// Used as an estimate of how much / how fast to remap
uint64_t commitCounter(0);
uint64_t estimatedPrivateMapSize(0);
uint64_t remapLastTimestamp(0);
while (shutdownRequested.loadRelaxed() == 0) {
unsigned ms = storageGlobalParams.journalCommitIntervalMs;
if (ms == 0) {
ms = samePartition ? 100 : 30;
}
// +1 so it never goes down to zero
const int64_t oneThird = (ms / 3) + 1;
// Reset the stats based on the reset interval
if (stats.curr()->getCurrentDurationMillis() > DurStatsResetIntervalMillis) {
stats.reset();
}
try {
stdx::unique_lock<stdx::mutex> lock(flushMutex);
for (unsigned i = 0; i <= 2; i++) {
if (stdx::cv_status::no_timeout ==
flushRequested.wait_for(lock, Milliseconds(oneThird).toSystemDuration())) {
// Someone forced a flush
break;
}
if (commitNotify.nWaiting()) {
// One or more getLastError j:true is pending
break;
}
if (commitJob.bytes() > UncommittedBytesLimit / 2) {
// The number of written bytes is growing
break;
}
}
// The commit logic itself
LOG(4) << "groupCommit begin";
Timer t;
const ServiceContext::UniqueOperationContext txnPtr = cc().makeOperationContext();
OperationContext& txn = *txnPtr;
AutoAcquireFlushLockForMMAPV1Commit autoFlushLock(txn.lockState());
// We need to snapshot the commitNumber after the flush lock has been obtained,
// because at this point we know that we have a stable snapshot of the data.
const CommitNotifier::When commitNumber(commitNotify.now());
LOG(4) << "Processing commit number " << commitNumber;
if (!commitJob.hasWritten()) {
// We do not need the journal lock anymore. Free it here, for the really
// unlikely possibility that the writeBuffer command below blocks.
autoFlushLock.release();
// getlasterror request could have came after the data was already committed.
// No need to call committingReset though, because we have not done any
// writes (hasWritten == false).
JournalWriter::Buffer* const buffer = journalWriter.newBuffer();
buffer->setNoop();
buffer->journalListenerToken = getJournalListener()->getToken();
journalWriter.writeBuffer(buffer, commitNumber);
} else {
// This copies all the in-memory changes into the journal writer's buffer.
JournalWriter::Buffer* const buffer = journalWriter.newBuffer();
PREPLOGBUFFER(buffer->getHeader(), buffer->getBuilder(), cs, serverStartMs);
estimatedPrivateMapSize += commitJob.bytes();
commitCounter++;
// Now that the write intents have been copied to the buffer, the commit job is
// free to be reused. We need to reset the commit job's contents while under
// the S flush lock, because otherwise someone might have done a write and this
// would wipe out their changes without ever being committed.
commitJob.committingReset();
double systemMemoryPressurePercentage =
ProcessInfo::getSystemMemoryPressurePercentage();
// Now that the in-memory modifications have been collected, we can potentially
// release the flush lock if remap is not necessary.
// When we remap due to memory pressure, we look at two criteria
// 1. If the amount of 4k pages touched exceeds 512 MB,
// a reasonable estimate of memory pressure on Linux.
// 2. Check if the amount of free memory on the machine is running low,
// since #1 is underestimates the memory pressure on Windows since
// commits in 64MB chunks.
const bool shouldRemap = (estimatedPrivateMapSize >= UncommittedBytesLimit) ||
(systemMemoryPressurePercentage > 0.0) ||
(commitCounter % NumCommitsBeforeRemap == 0) ||
(mmapv1GlobalOptions.journalOptions & MMAPV1Options::JournalAlwaysRemap);
double remapFraction = 0.0;
if (shouldRemap) {
// We want to remap all private views about every 2 seconds. There could be
// ~1000 views so we do a little each pass. There will be copy on write
// faults after remapping, so doing a little bit at a time will avoid big
// load spikes when the pages are touched.
//
// TODO: Instead of the time-based logic above, consider using ProcessInfo
// and watching for getResidentSize to drop, which is more precise.
remapFraction = (curTimeMicros64() - remapLastTimestamp) / 2000000.0;
if (mmapv1GlobalOptions.journalOptions & MMAPV1Options::JournalAlwaysRemap) {
remapFraction = 1;
} else {
// We don't want to get close to the UncommittedBytesLimit
const double remapMemFraction =
estimatedPrivateMapSize / ((double)UncommittedBytesLimit);
remapFraction = std::max(remapMemFraction, remapFraction);
remapFraction = std::max(systemMemoryPressurePercentage, remapFraction);
}
} else {
LOG(4) << "Early release flush lock";
// We will not be doing a remap so drop the flush lock. That way we will be
// doing the journal I/O outside of lock, so other threads can proceed.
invariant(!shouldRemap);
autoFlushLock.release();
}
buffer->journalListenerToken = getJournalListener()->getToken();
// Request async I/O to the journal. This may block.
journalWriter.writeBuffer(buffer, commitNumber);
// Data has now been written to the shared view. If remap was requested, we
// would still be holding the S flush lock here, so just upgrade it and
// perform the remap.
if (shouldRemap) {
// Need to wait for the previously scheduled journal writes to complete
// before any remap is attempted.
journalWriter.flush();
journalWriter.assertIdle();
// Upgrading the journal lock to flush stops all activity on the system,
// because we will be remapping memory and we don't want readers to be
// accessing it. Technically this step could be avoided on systems, which
// support atomic remap.
autoFlushLock.upgradeFlushLockToExclusive();
remapPrivateView(remapFraction);
autoFlushLock.release();
// Reset the private map estimate outside of the lock
estimatedPrivateMapSize = 0;
remapLastTimestamp = curTimeMicros64();
stats.curr()->_commitsInWriteLock++;
stats.curr()->_commitsInWriteLockMicros += t.micros();
}
}
stats.curr()->_commits++;
stats.curr()->_commitsMicros += t.micros();
LOG(4) << "groupCommit end";
} catch (DBException& e) {
severe() << "dbexception in durThread causing immediate shutdown: " << e.toString();
invariant(false);
} catch (std::ios_base::failure& e) {
severe() << "ios_base exception in durThread causing immediate shutdown: " << e.what();
invariant(false);
} catch (std::bad_alloc& e) {
severe() << "bad_alloc exception in durThread causing immediate shutdown: " << e.what();
invariant(false);
} catch (std::exception& e) {
severe() << "exception in durThread causing immediate shutdown: " << e.what();
invariant(false);
} catch (...) {
severe() << "unhandled exception in durThread causing immediate shutdown";
invariant(false);
}
}
// Stops the journal thread and ensures everything was written
invariant(!commitJob.hasWritten());
journalWriter.flush();
journalWriter.shutdown();
log() << "Durability thread stopped";
}
static void durThread() {
Client::initThread("journal");
bool samePartition = true;
try {
const std::string dbpathDir =
boost::filesystem::path(storageGlobalParams.dbpath).string();
samePartition = onSamePartition(getJournalDir().string(), dbpathDir);
}
catch(...) {
}
while (shutdownRequested.loadRelaxed() == 0) {
unsigned ms = storageGlobalParams.journalCommitInterval;
if( ms == 0 ) {
ms = samePartition ? 100 : 30;
}
unsigned oneThird = (ms / 3) + 1; // +1 so never zero
try {
stats.rotate();
boost::mutex::scoped_lock lock(flushMutex);
// commit sooner if one or more getLastError j:true is pending
for (unsigned i = 0; i <= 2; i++) {
if (flushRequested.timed_wait(lock,
Milliseconds(oneThird))) {
// Someone forced a flush
break;
}
if (commitJob._notify.nWaiting())
break;
if (commitJob.bytes() > UncommittedBytesLimit / 2)
break;
}
OperationContextImpl txn;
// Waits for all active operations to drain and won't let new ones start. This
// should be optimized to allow readers in (see SERVER-15262).
AutoAcquireFlushLockForMMAPV1Commit flushLock(txn.lockState());
groupCommit();
remapPrivateView();
}
catch(std::exception& e) {
log() << "exception in durThread causing immediate shutdown: " << e.what() << endl;
mongoAbort("exception in durThread");
}
catch (...) {
log() << "unhandled exception in durThread causing immediate shutdown" << endl;
mongoAbort("unhandled exception in durThread");
}
}
cc().shutdown();
}
