/** put the basic write operation into the buffer (bb) to be journaled */
static void prepBasicWrite_inlock(AlignedBuilder&bb, const WriteIntent *i, RelativePath& lastDbPath) {
size_t ofs = 1;
MongoMMF *mmf = findMMF_inlock(i->start(), /*out*/ofs);
if( unlikely(!mmf->willNeedRemap()) ) {
// tag this mmf as needed a remap of its private view later.
// usually it will already be dirty/already set, so we do the if above first
// to avoid possibility of cpu cache line contention
mmf->willNeedRemap() = true;
}
// since we have already looked up the mmf, we go ahead and remember the write view location
// so we don't have to find the MongoMMF again later in WRITETODATAFILES()
//
// this was for WRITETODATAFILES_Impl2 so commented out now
//
/*
dassert( i->w_ptr == 0 );
i->w_ptr = ((char*)mmf->view_write()) + ofs;
*/
JEntry e;
e.len = min(i->length(), (unsigned)(mmf->length() - ofs)); //dont write past end of file
assert( ofs <= 0x80000000 );
e.ofs = (unsigned) ofs;
e.setFileNo( mmf->fileSuffixNo() );
if( mmf->relativePath() == local ) {
e.setLocalDbContextBit();
}
else if( mmf->relativePath() != lastDbPath ) {
lastDbPath = mmf->relativePath();
JDbContext c;
bb.appendStruct(c);
bb.appendStr(lastDbPath.toString());
}
bb.appendStruct(e);
#if defined(_EXPERIMENTAL)
i->ofsInJournalBuffer = bb.len();
#endif
bb.appendBuf(i->start(), e.len);
if (unlikely(e.len != (unsigned)i->length())) {
log() << "journal info splitting prepBasicWrite at boundary" << endl;
// This only happens if we write to the last byte in a file and
// the fist byte in another file that is mapped adjacently. I
// think most OSs leave at least a one page gap between
// mappings, but better to be safe.
WriteIntent next ((char*)i->start() + e.len, i->length() - e.len);
prepBasicWrite_inlock(bb, &next, lastDbPath);
}
}
/** We need to remap the private views periodically. otherwise they would become very large.
Call within write lock.
*/
void REMAPPRIVATEVIEW() {
static unsigned startAt;
static unsigned long long lastRemap;
dbMutex.assertWriteLocked();
dbMutex._remapPrivateViewRequested = false;
assert( !commitJob.hasWritten() );
if( 0 ) {
log() << "TEMP remapprivateview disabled for testing - will eventually run oom in this mode if db bigger than ram" << endl;
return;
}
// we want to remap all private views about every 2 seconds. there could be ~1000 views so
// we do a little each pass; beyond the remap time, more significantly, there will be copy on write
// faults after remapping, so doing a little bit at a time will avoid big load spikes on
// remapping.
unsigned long long now = curTimeMicros64();
double fraction = (now-lastRemap)/20000000.0;
set<MongoFile*>& files = MongoFile::getAllFiles();
unsigned sz = files.size();
if( sz == 0 )
return;
unsigned ntodo = (unsigned) (sz * fraction);
if( ntodo < 1 ) ntodo = 1;
if( ntodo > sz ) ntodo = sz;
const set<MongoFile*>::iterator b = files.begin();
const set<MongoFile*>::iterator e = files.end();
set<MongoFile*>::iterator i = b;
for( unsigned x = 0; x < startAt; x++ ) {
i++;
if( i == e ) i = b;
}
startAt = (startAt + ntodo) % sz;
for( unsigned x = 0; x < ntodo; x++ ) {
dassert( i != e );
MongoMMF *mmf = dynamic_cast<MongoMMF*>(*i);
if( mmf && mmf->willNeedRemap() ) {
mmf->willNeedRemap() = false;
mmf->remapThePrivateView();
}
i++;
if( i == e ) i = b;
}
}
/** we will build an output buffer ourself and then use O_DIRECT
we could be in read lock for this
caller handles locking
*/
static void PREPLOGBUFFER() {
assert( cmdLine.dur );
AlignedBuilder& bb = commitJob._ab;
bb.reset();
unsigned lenOfs;
// JSectHeader
{
bb.appendStr("\nHH\n", false);
lenOfs = bb.skip(4);
}
// ops other than basic writes
{
for( vector< shared_ptr<DurOp> >::iterator i = commitJob.ops().begin(); i != commitJob.ops().end(); ++i ) {
(*i)->serialize(bb);
}
}
// write intents
{
scoped_lock lk(privateViews._mutex());
string lastFilePath;
for( vector<WriteIntent>::iterator i = commitJob.writes().begin(); i != commitJob.writes().end(); i++ ) {
size_t ofs;
MongoMMF *mmf = privateViews._find(i->p, ofs);
if( mmf == 0 ) {
string s = str::stream() << "view pointer cannot be resolved " << (size_t) i->p;
journalingFailure(s.c_str()); // asserts
return;
}
if( !mmf->willNeedRemap() ) {
mmf->willNeedRemap() = true; // usually it will already be dirty so don't bother writing then
}
//size_t ofs = ((char *)i->p) - ((char*)mmf->getView().p);
i->w_ptr = ((char*)mmf->view_write()) + ofs;
if( mmf->filePath() != lastFilePath ) {
lastFilePath = mmf->filePath();
JDbContext c;
bb.appendStruct(c);
bb.appendStr(lastFilePath);
}
JEntry e;
e.len = i->len;
assert( ofs <= 0x80000000 );
e.ofs = (unsigned) ofs;
e.fileNo = mmf->fileSuffixNo();
bb.appendStruct(e);
bb.appendBuf(i->p, i->len);
}
}
{
JSectFooter f(bb.buf(), bb.len());
bb.appendStruct(f);
}
{
assert( 0xffffe000 == (~(Alignment-1)) );
unsigned L = (bb.len() + Alignment-1) & (~(Alignment-1)); // fill to alignment
dassert( L >= (unsigned) bb.len() );
*((unsigned*)bb.atOfs(lenOfs)) = L;
unsigned padding = L - bb.len();
bb.skip(padding);
dassert( bb.len() % Alignment == 0 );
}
return;
}
/** We need to remap the private views periodically. otherwise they would become very large.
Call within write lock.
*/
void _REMAPPRIVATEVIEW() {
// todo: Consider using ProcessInfo herein and watching for getResidentSize to drop. that could be a way
// to assure very good behavior here.
static unsigned startAt;
static unsigned long long lastRemap;
dbMutex.assertWriteLocked();
dbMutex._remapPrivateViewRequested = false;
assert( !commitJob.hasWritten() );
// we want to remap all private views about every 2 seconds. there could be ~1000 views so
// we do a little each pass; beyond the remap time, more significantly, there will be copy on write
// faults after remapping, so doing a little bit at a time will avoid big load spikes on
// remapping.
unsigned long long now = curTimeMicros64();
double fraction = (now-lastRemap)/2000000.0;
if( cmdLine.durOptions & CmdLine::DurAlwaysRemap )
fraction = 1;
lastRemap = now;
RWLockRecursive::Shared lk(MongoFile::mmmutex);
set<MongoFile*>& files = MongoFile::getAllFiles();
unsigned sz = files.size();
if( sz == 0 )
return;
{
// be careful not to use too much memory if the write rate is
// extremely high
double f = privateMapBytes / ((double)UncommittedBytesLimit);
if( f > fraction ) {
fraction = f;
}
privateMapBytes = 0;
}
unsigned ntodo = (unsigned) (sz * fraction);
if( ntodo < 1 ) ntodo = 1;
if( ntodo > sz ) ntodo = sz;
const set<MongoFile*>::iterator b = files.begin();
const set<MongoFile*>::iterator e = files.end();
set<MongoFile*>::iterator i = b;
// skip to our starting position
for( unsigned x = 0; x < startAt; x++ ) {
i++;
if( i == e ) i = b;
}
startAt = (startAt + ntodo) % sz; // mark where to start next time
for( unsigned x = 0; x < ntodo; x++ ) {
dassert( i != e );
if( (*i)->isMongoMMF() ) {
MongoMMF *mmf = (MongoMMF*) *i;
assert(mmf);
if( mmf->willNeedRemap() ) {
mmf->willNeedRemap() = false;
mmf->remapThePrivateView();
}
i++;
if( i == e ) i = b;
}
}
}
static void _REMAPPRIVATEVIEW() {
// todo: Consider using ProcessInfo herein and watching for getResidentSize to drop. that could be a way
// to assure very good behavior here.
static unsigned startAt;
static unsigned long long lastRemap;
LOG(4) << "journal REMAPPRIVATEVIEW" << endl;
verify( Lock::isW() );
verify( !commitJob.hasWritten() );
// we want to remap all private views about every 2 seconds. there could be ~1000 views so
// we do a little each pass; beyond the remap time, more significantly, there will be copy on write
// faults after remapping, so doing a little bit at a time will avoid big load spikes on
// remapping.
unsigned long long now = curTimeMicros64();
double fraction = (now-lastRemap)/2000000.0;
if( cmdLine.durOptions & CmdLine::DurAlwaysRemap )
fraction = 1;
lastRemap = now;
#if defined(_WIN32)
// Note that this negatively affects performance.
// We must grab the exclusive lock here because remapThePrivateView() on Windows needs
// to grab it as well, due to the lack of a non-atomic way to remap a memory mapped file.
// See SERVER-5723 for performance improvement.
// See SERVER-5680 to see why this code is necessary.
LockMongoFilesExclusive lk;
#else
LockMongoFilesShared lk;
#endif
set<MongoFile*>& files = MongoFile::getAllFiles();
unsigned sz = files.size();
if( sz == 0 )
return;
{
// be careful not to use too much memory if the write rate is
// extremely high
double f = privateMapBytes / ((double)UncommittedBytesLimit);
if( f > fraction ) {
fraction = f;
}
privateMapBytes = 0;
}
unsigned ntodo = (unsigned) (sz * fraction);
if( ntodo < 1 ) ntodo = 1;
if( ntodo > sz ) ntodo = sz;
const set<MongoFile*>::iterator b = files.begin();
const set<MongoFile*>::iterator e = files.end();
set<MongoFile*>::iterator i = b;
// skip to our starting position
for( unsigned x = 0; x < startAt; x++ ) {
i++;
if( i == e ) i = b;
}
unsigned startedAt = startAt;
startAt = (startAt + ntodo) % sz; // mark where to start next time
Timer t;
for( unsigned x = 0; x < ntodo; x++ ) {
dassert( i != e );
if( (*i)->isMongoMMF() ) {
MongoMMF *mmf = (MongoMMF*) *i;
verify(mmf);
if( mmf->willNeedRemap() ) {
mmf->willNeedRemap() = false;
mmf->remapThePrivateView();
}
i++;
if( i == e ) i = b;
}
}
LOG(2) << "journal REMAPPRIVATEVIEW done startedAt: " << startedAt << " n:" << ntodo << ' ' << t.millis() << "ms" << endl;
}