void DurableImpl::setNoJournal(void *dst, void *src, unsigned len) {
// we are at least read locked, so we need not worry about REMAPPRIVATEVIEW herein.
DEV dbMutex.assertAtLeastReadLocked();
MemoryMappedFile::makeWritable(dst, len);
// we enter the RecoveryJob mutex here, so that if WRITETODATAFILES is happening we do not
// conflict with it
scoped_lock lk1( RecoveryJob::get()._mx );
// we stay in this mutex for everything to work with DurParanoid/validateSingleMapMatches
//
// either of these mutexes also makes setNoJournal threadsafe, which is good as we call it from a read
// (not a write) lock in class SlaveTracking
//
scoped_lock lk( privateViews._mutex() );
size_t ofs;
MongoMMF *f = privateViews.find_inlock(dst, ofs);
assert(f);
void *w = (((char *)f->view_write())+ofs);
// first write it to the writable (file) view
memcpy(w, src, len);
if( memcmp(w, dst, len) ) {
// if we get here, a copy-on-write had previously occurred. so write it to the private view too
// to keep them in sync. we do this as we do not want to cause a copy on write unnecessarily.
memcpy(dst, src, len);
}
}
/** (SLOW) diagnostic to check that the private view and the non-private view are in sync.
*/
static void debugValidateMapsMatch() {
if( !DebugValidateMapsMatch )
return;
Timer t;
set<MongoFile*>& files = MongoFile::getAllFiles();
for( set<MongoFile*>::iterator i = files.begin(); i != files.end(); i++ ) {
MongoFile *mf = *i;
if( mf->isMongoMMF() ) {
MongoMMF *mmf = (MongoMMF*) mf;
const char *p = (const char *) mmf->getView();
const char *w = (const char *) mmf->view_write();
unsigned low = 0xffffffff;
unsigned high = 0;
for( unsigned i = 0; i < mmf->length(); i++ ) {
if( p[i] != w[i] ) {
log() << i << '\t' << (int) p[i] << '\t' << (int) w[i] << endl;
if( i < low ) low = i;
if( i > high ) high = i;
}
}
if( low != 0xffffffff ) {
std::stringstream ss;
ss << "dur error warning views mismatch " << mmf->filename() << ' ' << (hex) << low << ".." << high << " len:" << high-low+1;
log() << ss.str() << endl;
log() << "priv loc: " << (void*)(p+low) << endl;
vector<WriteIntent>& w = commitJob.writes();
(void)w; // mark as unused. Useful for inspection in debugger
breakpoint();
}
}
}
log() << "debugValidateMapsMatch " << t.millis() << "ms " << endl;
}
/** caller handles locking */
static bool PREPLOGBUFFER(AlignedBuilder& bb) {
bb.reset();
unsigned *lenInBlockHeader;
{
// JSectHeader
bb.appendStr("\nHH\n", false);
lenInBlockHeader = (unsigned *) bb.skip(4);
}
string lastFilePath;
{
scoped_lock lk(privateViews._mutex());
for( vector<WriteIntent>::iterator i = wi._writes.begin(); i != wi._writes.end(); i++ ) {
size_t ofs;
MongoMMF *mmf = privateViews._find(i->p, ofs);
if( mmf == 0 ) {
journalingFailure("view pointer cannot be resolved");
}
else {
if( !mmf->dirty() )
mmf->dirty() = 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;
e.fileNo = mmf->fileSuffixNo();
bb.appendStruct(e);
bb.appendBuf(i->p, i->len);
}
}
}
{
JSectFooter f;
f.hash = 0;
bb.appendStruct(f);
}
{
unsigned L = (bb.len() + 8191) & 0xffffe000; // fill to alignment
dassert( L >= (unsigned) bb.len() );
*lenInBlockHeader = L;
unsigned padding = L - bb.len();
bb.skip(padding);
dassert( bb.len() % 8192 == 0 );
}
return true;
}
void operator () (MongoFile *mf) {
if( mf->isMongoMMF() ) {
MongoMMF *mmf = (MongoMMF*) mf;
const unsigned char *p = (const unsigned char *) mmf->getView();
const unsigned char *w = (const unsigned char *) mmf->view_write();
if (!p || !w) return; // File not fully opened yet
_bytes += mmf->length();
assert( mmf->length() == (unsigned) mmf->length() );
{
scoped_lock lk( privateViews._mutex() ); // see setNoJournal
if (memcmp(p, w, (unsigned) mmf->length()) == 0)
return; // next file
}
unsigned low = 0xffffffff;
unsigned high = 0;
log() << "DurParanoid mismatch in " << mmf->filename() << endl;
int logged = 0;
unsigned lastMismatch = 0xffffffff;
for( unsigned i = 0; i < mmf->length(); i++ ) {
if( p[i] != w[i] ) {
if( lastMismatch != 0xffffffff && lastMismatch+1 != i )
log() << endl; // separate blocks of mismatches
lastMismatch= i;
if( ++logged < 60 ) {
if( logged == 1 )
log() << "ofs % 628 = 0x" << hex << (i%628) << endl; // for .ns files to find offset in record
stringstream ss;
ss << "mismatch ofs:" << hex << i << "\tfilemap:" << setw(2) << (unsigned) w[i] << "\tprivmap:" << setw(2) << (unsigned) p[i];
if( p[i] > 32 && p[i] <= 126 )
ss << '\t' << p[i];
log() << ss.str() << endl;
}
if( logged == 60 )
log() << "..." << endl;
if( i < low ) low = i;
if( i > high ) high = i;
}
}
if( low != 0xffffffff ) {
std::stringstream ss;
ss << "journal error warning views mismatch " << mmf->filename() << ' ' << (hex) << low << ".." << high << " len:" << high-low+1;
log() << ss.str() << endl;
log() << "priv loc: " << (void*)(p+low) << ' ' << endl;
set<WriteIntent>& b = commitJob.writes();
(void)b; // mark as unused. Useful for inspection in debugger
// should we abort() here so this isn't unnoticed in some circumstances?
massert(13599, "Written data does not match in-memory view. Missing WriteIntent?", false);
}
}
}
void DurableImpl::debugCheckLastDeclaredWrite() {
if( !DebugCheckLastDeclaredWrite )
return;
if( testIntent )
return;
static int n;
++n;
assert(debug && cmdLine.dur);
vector<WriteIntent>& w = commitJob.writes();
if( w.size() == 0 )
return;
const WriteIntent &i = w[w.size()-1];
size_t ofs;
MongoMMF *mmf = privateViews.find(i.p, ofs);
if( mmf == 0 )
return;
size_t past = ofs + i.len;
if( mmf->length() < past + 8 )
return; // too close to end of view
char *priv = (char *) mmf->getView();
char *writ = (char *) mmf->view_write();
unsigned long long *a = (unsigned long long *) (priv+past);
unsigned long long *b = (unsigned long long *) (writ+past);
if( *a != *b ) {
for( unsigned z = 0; z < w.size() - 1; z++ ) {
const WriteIntent& wi = w[z];
char *r1 = (char*) wi.p;
char *r2 = r1 + wi.len;
if( r1 <= (((char*)a)+8) && r2 > (char*)a ) {
//log() << "it's ok " << wi.p << ' ' << wi.len << endl;
return;
}
}
log() << "dur data after write area " << i.p << " does not agree" << endl;
log() << " was: " << ((void*)b) << " " << hexdump((char*)b, 8) << endl;
log() << " now: " << ((void*)a) << " " << hexdump((char*)a, 8) << endl;
log() << " n: " << n << endl;
log() << endl;
}
}
void t() {
for( int i = 0; i < 20; i++ ) {
sleepmillis(21);
string fn = "/tmp/t1";
MongoMMF f;
unsigned long long len = 1 * 1024 * 1024;
assert( f.create(fn, len, /*sequential*/rand()%2==0) );
{
char *p = (char *) f.getView();
assert(p);
// write something to the private view as a test
strcpy(p, "hello");
}
if( cmdLine.dur ) {
char *w = (char *) f.view_write();
strcpy(w + 6, "world");
}
MongoFileFinder ff;
ASSERT( ff.findByPath(fn) );
}
}
void DurableImpl::debugCheckLastDeclaredWrite() {
static int n;
++n;
assert(debug && cmdLine.dur);
if (commitJob.writes().empty())
return;
const WriteIntent &i = commitJob.lastWrite();
size_t ofs;
MongoMMF *mmf = privateViews.find(i.start(), ofs);
if( mmf == 0 )
return;
size_t past = ofs + i.length();
if( mmf->length() < past + 8 )
return; // too close to end of view
char *priv = (char *) mmf->getView();
char *writ = (char *) mmf->view_write();
unsigned long long *a = (unsigned long long *) (priv+past);
unsigned long long *b = (unsigned long long *) (writ+past);
if( *a != *b ) {
for( set<WriteIntent>::iterator it(commitJob.writes().begin()), end((commitJob.writes().begin())); it != end; ++it ) {
const WriteIntent& wi = *it;
char *r1 = (char*) wi.start();
char *r2 = (char*) wi.end();
if( r1 <= (((char*)a)+8) && r2 > (char*)a ) {
//log() << "it's ok " << wi.p << ' ' << wi.len << endl;
return;
}
}
log() << "journal data after write area " << i.start() << " does not agree" << endl;
log() << " was: " << ((void*)b) << " " << hexdump((char*)b, 8) << endl;
log() << " now: " << ((void*)a) << " " << hexdump((char*)a, 8) << endl;
log() << " n: " << n << endl;
log() << endl;
}
}
/** 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;
}
void run() {
try { boost::filesystem::remove(fn); }
catch(...) { }
writelock lk;
{
MongoMMF f;
unsigned long long len = 256 * 1024 * 1024;
assert( f.create(fn, len, /*sequential*/false) );
{
char *p = (char *) f.getView();
assert(p);
// write something to the private view as a test
if( cmdLine.dur )
MemoryMappedFile::makeWritable(p, 6);
strcpy(p, "hello");
}
if( cmdLine.dur ) {
char *w = (char *) f.view_write();
strcpy(w + 6, "world");
}
MongoFileFinder ff;
ASSERT( ff.findByPath(fn) );
ASSERT( ff.findByPath("asdf") == 0 );
}
{
MongoFileFinder ff;
ASSERT( ff.findByPath(fn) == 0 );
}
int N = 10000;
#if !defined(_WIN32) && !defined(__linux__)
// seems this test is slow on OS X.
N = 100;
#endif
// we make a lot here -- if we were leaking, presumably it would fail doing this many.
Timer t;
for( int i = 0; i < N; i++ ) {
MongoMMF f;
assert( f.open(fn, i%4==1) );
{
char *p = (char *) f.getView();
assert(p);
if( cmdLine.dur )
MemoryMappedFile::makeWritable(p, 4);
strcpy(p, "zzz");
}
if( cmdLine.dur ) {
char *w = (char *) f.view_write();
if( i % 2 == 0 )
++(*w);
assert( w[6] == 'w' );
}
}
if( t.millis() > 10000 ) {
log() << "warning: MMap LeakTest is unusually slow N:" << N << ' ' << t.millis() << "ms" << endl;
}
}