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);
}
}
}
/** (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;
}
/** 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);
}
}
__declspec(noinline) void makeChunkWritable(size_t chunkno) {
scoped_lock lk(mapViewMutex);
if( writable.get(chunkno) ) // double check lock
return;
// remap all maps in this chunk. common case is a single map, but could have more than one with smallfiles or .ns files
size_t chunkStart = chunkno * MemoryMappedFile::ChunkSize;
size_t chunkNext = chunkStart + MemoryMappedFile::ChunkSize;
scoped_lock lk2(privateViews._mutex());
map<void*,MongoMMF*>::iterator i = privateViews.finditer_inlock((void*) (chunkNext-1));
while( 1 ) {
const pair<void*,MongoMMF*> x = *(--i);
MongoMMF *mmf = x.second;
if( mmf == 0 )
break;
size_t viewStart = (size_t) x.first;
size_t viewEnd = (size_t) (viewStart + mmf->length());
if( viewEnd <= chunkStart )
break;
size_t protectStart = max(viewStart, chunkStart);
dassert(protectStart<chunkNext);
size_t protectEnd = min(viewEnd, chunkNext);
size_t protectSize = protectEnd - protectStart;
dassert(protectSize>0&&protectSize<=MemoryMappedFile::ChunkSize);
DWORD oldProtection;
bool ok = VirtualProtect( reinterpret_cast<void*>( protectStart ),
protectSize,
PAGE_WRITECOPY,
&oldProtection );
if ( !ok ) {
DWORD dosError = GetLastError();
log() << "VirtualProtect for " << mmf->filename()
<< " chunk " << chunkno
<< " failed with " << errnoWithDescription( dosError )
<< " (chunk size is " << protectSize
<< ", address is " << hex << protectStart << dec << ")"
<< " in mongo::makeChunkWritable, terminating"
<< endl;
fassertFailed( 16362 );
}
}
writable.set(chunkno);
}
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;
}
}
/** underscore version of find is for when you are already locked
@param ofs out return our offset in the view
@return the MongoMMF to which this pointer belongs
*/
MongoMMF* PointerToMMF::find_inlock(void *p, /*out*/ size_t& ofs) {
//
// .................memory..........................
// v1 p v2
// [--------------------] [-------]
//
// e.g., _find(p) == v1
//
const pair<void*,MongoMMF*> x = *(--_views.upper_bound(p));
MongoMMF *mmf = x.second;
if( mmf ) {
size_t o = ((char *)p) - ((char*)x.first);
if( o < mmf->length() ) {
ofs = o;
return mmf;
}
}
return 0;
}
/*static*/ void* MongoMMF::switchToPrivateView(void *readonly_ptr) {
void *p = readonly_ptr;
assert( durable );
assert( debug );
mutex::scoped_lock lk(our_views_mutex);
std::map< void*, MongoMMF* >::iterator i =
our_read_views.upper_bound(((char *)p)+1);
i--;
bool ok = i != our_read_views.end();
if( ok ) {
MongoMMF *mmf = i->second;
assert( mmf );
size_t ofs = ((char *)p) - ((char*)mmf->view_readonly);
if( ofs < mmf->length() ) {
return ((char *)mmf->view_private) + ofs;
}
}
if( 1 ) {
static int once;
/* temp : not using MongoMMF yet for datafiles, just .ns. more to do... */
if( once++ == 0 )
log() << "TEMP TODO _DURABLE : use mongommf for datafiles" << endl;
return p;
}
for( std::map<void*,MongoMMF*>::iterator i = our_read_views.begin(); i != our_read_views.end(); i++ ) {
char *wl = (char *) i->second->view_private;
char *wh = wl + i->second->length();
if( p >= wl && p < wh ) {
log() << "dur: perf warning p=" << p << " is already in the writable view of " << i->second->filename() << endl;
return p;
}
}
log() << "switchToPrivateView error " << p << endl;
assert( false ); // did you call writing() with a pointer that isn't into a datafile?
return 0;
}
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;
}
}