void verifyChunkMD5(vector<File> &files) {
try {
while (true) {
Chunk * c = chunksToComputeMD5.consume();
if (files[c->parentFileIndex].matchStatus == File::Status::FAILED_TO_MATCH_REMOTE_FILE) {
// We have already marked file as a non-match, don't waste time reading more chunks from it
c->log("File status == FAILED_TO_MATCH_REMOTE_FILE, Skipping the MD5 compute...");
c->clear();
chunksSkipped.produce(c);
} else {
c->log("Computing MD5...");
string computedMD5 = c->computeMD5();
c->clear();
if (c->expectedMD5 != computedMD5) {
c->log("MISMATCH between expected MD5 '" + c->expectedMD5 + "', and computed MD5 '" + computedMD5 + "' ... marking the file as Mismatch");
files[c->parentFileIndex].matchStatus = File::Status::FAILED_TO_MATCH_REMOTE_FILE;
chunksFailed.produce(c);
} else {
c->log("Expected and computed MD5 match!");
chunksFinished.produce(c);
}
}
}
} catch (boost::thread_interrupted &ti) {
return;
}
}
void compressChunks() {
try {
while (true) {
Chunk * c = chunksToCompress.consume();
if (c->toCompress) {
c->log("Compressing...");
c->compress();
c->log("Finished compressing");
} else {
c->log("Not compressing");
}
chunksToUpload.produce(c);
// Sleep for tiny amount of time, to make sure we yield to other threads.
// Note: boost::this_thread::yield() is not a valid interruption point,
// so we have to use sleep()
boost::this_thread::sleep(boost::posix_time::microseconds(100));
}
} catch(std::bad_alloc &e) {
boost::call_once(bad_alloc_once, boost::bind(&handle_bad_alloc, e));
} catch (boost::thread_interrupted &ti) {
return;
}
}
// There is currently the possibility of a race condition if a chunk
// upload timed-out. It's possible that a second upload succeeds,
// has the chunk marked as "complete" and then the first request makes
// its way through the queue and marks the chunk as pending again.
// Since we are just about to close the file, we'll check to see if any
// chunks are marked as pending, and if so, we'll retry them.
void check_for_complete_chunks(vector<File> &files) {
for (int currCheckNum=0; currCheckNum < NUM_CHUNK_CHECKS; ++currCheckNum){
map<string, JSON> fileDescriptions;
while (!chunksFinished.empty()) {
Chunk *c = chunksFinished.consume();
// Cache file descriptions so we only have to do once per file,
// not once per chunk.
if (fileDescriptions.find(c->fileID) == fileDescriptions.end())
fileDescriptions[c->fileID] = fileDescribe(c->fileID);
if (!is_chunk_complete(c, fileDescriptions[c->fileID])) {
// After the chunk was uploaded, it was cleared, removing the data
// from the buffer. We need to reload if we're going to upload again.
chunksToRead.produce(c);
}
}
// All of the chunks were marked as complete, so let's exit and we
// should be safeish to close the file.
if(chunksToRead.size() == 0)
return;
// Set the totalChunks variable to the # of chunks we're going
// to retry now plus the number of chunks in the failed queue. The monitor
// thread will be busy until the size of chunksFinished + chunksFailed
// equals totalChunks.
DXLOG(logINFO) << "Retrying " << chunksToRead.size() << " chunks that did not complete.";
totalChunks = chunksToRead.size() + chunksFailed.size();
// Read, compress, and upload the chunks which weren't marked as complete.
createWorkerThreads(files);
boost::thread monitorThread(monitor);
monitorThread.join();
interruptWorkerThreads();
joinWorkerThreads();
}
// We have tried to upload incomplete chunks NUM_CHUNK_CHECKS times!
// Check to see if there are any chunks still not complete and if so,
// print warning.
map<string, JSON> fileDescriptions;
while (!chunksFinished.empty()) {
Chunk *c = chunksFinished.consume();
// Cache file descriptions so we only have to do once per file,
// not once per chunk.
if (fileDescriptions.find(c->fileID) == fileDescriptions.end())
fileDescriptions[c->fileID] = fileDescribe(c->fileID);
if (!is_chunk_complete(c, fileDescriptions[c->fileID])) {
cerr << "Chunk " << c->index << " of file " << c->fileID << " did not complete. This file will not be accessible. PLease try to upload this file again." << endl;
}
}
}
void readChunks(const vector<File> &files) {
try {
while (true) {
Chunk * c = chunksToRead.consume();
if (files[c->parentFileIndex].matchStatus == File::Status::FAILED_TO_MATCH_REMOTE_FILE) {
// We have already marked file as a non-match, don't waste time reading more chunks from it
c->log("File status == FAILED_TO_MATCH_REMOTE_FILE, Skipping the read...");
chunksSkipped.produce(c);
} else {
c->log("Reading...");
c->read();
c->log("Finished reading");
chunksToComputeMD5.produce(c);
}
}
} catch (boost::thread_interrupted &ti) {
return;
}
}
unsigned int File::createChunks(BlockingQueue<Chunk *> &queue) {
using namespace dx;
// For creating chunks:
// 1) We sort the part IDs in increasing order (by pushing them all in a std::map after converting to int)
// 2) We start creating file chunks
if (matchStatus == Status::FAILED_TO_MATCH_REMOTE_FILE) {
return 0; // we have already marked the file as a non-match, no need to create chunks
}
// We will do it in 2 pass .. first create all "keys" in the map .. then in second pass we will add their sizes etc
map<int, JSON> chunkInfo;
// Note: We are not asserting for "parts" JSON structure, since it has been already done in init()
for (JSON::object_iterator it = parts.object_begin(); it != parts.object_end(); ++it) {
chunkInfo[boost::lexical_cast<int>(it->first)] = JSON::parse("{\"md5\": \"" + it->second["md5"].get<string>() + "\"}");
}
// second pass: Add start and end location for chunks in the local file
int64_t start = 0;
for (map<int, JSON>::iterator it = chunkInfo.begin(); it != chunkInfo.end(); ++it) {
it->second["start"] = start;
start = it->second["end"] = (start + parts[boost::lexical_cast<string>(it->first)]["size"].get<int64_t>());
}
// TODO: Sanity check that this works for empty file as well
LOG << "Creating chunks:" << endl;
fs::path p(localFile);
int actualChunksCreated = 0;
for (map<int, JSON>::iterator it = chunkInfo.begin(); it != chunkInfo.end(); ++it) {
Chunk *c = new Chunk(localFile, it->second["md5"].get<string>(), it->second["start"].get<int64_t>(), it->second["end"].get<int64_t>(), fileIndex);
c->log("created");
queue.produce(c);
actualChunksCreated++;
}
return actualChunksCreated;
}
void uploadChunks(vector<File> &files) {
try {
while (true) {
Chunk * c = chunksToUpload.consume();
c->log("Uploading...");
bool uploaded = false;
try {
c->upload(opt);
uploaded = true;
} catch (runtime_error &e) {
ostringstream msg;
msg << "Upload failed: " << e.what();
c->log(msg.str(), logERROR);
}
if (uploaded) {
c->log("Upload succeeded!");
int64_t size_of_chunk = c->data.size(); // this can be different than (c->end - c->start) because of compression
c->clear();
chunksFinished.produce(c);
// Update number of bytes uploaded in parent file object
boost::mutex::scoped_lock boLock(bytesUploadedMutex);
files[c->parentFileIndex].bytesUploaded += (c->end - c->start);
files[c->parentFileIndex].atleastOnePartDone = true;
bytesUploadedSinceStart += size_of_chunk;
boLock.unlock();
} else if (c->triesLeft > 0) {
int numTry = NUMTRIES_g - c->triesLeft + 1; // find out which try is it
int timeout = (numTry > 6) ? 256 : 4 << numTry; // timeout is always between [8, 256] seconds
c->log("Will retry reading and uploading this chunks in " + boost::lexical_cast<string>(timeout) + " seconds", logWARNING);
if (!opt.noRoundRobinDNS) {
boost::mutex::scoped_lock forceRefreshLock(forceRefreshDNSMutex);
c->log("Setting forceRefreshDNS = true in main.cpp:uploadChunks()");
forceRefreshDNS = true; // refresh the DNS list in next call to getRandomIP()
}
--(c->triesLeft);
c->clear(); // we will read & compress data again
boost::this_thread::sleep(boost::posix_time::milliseconds(timeout * 1000));
// We push the chunk to retry to "chunksToRead" and not "chunksToUpload"
// Since chunksToUpload queue is bounded, and chunksToUpload.produce() can block,
// thus giving rise to deadlock
chunksToRead.produce(c);
} else {
c->log("Not retrying", logERROR);
// TODO: Should we print it on stderr or DXLOG (verbose only) ??
cerr << "\nFailed to upload Chunk [" << c->start << " - " << c->end << "] for local file ("
<< files[c->parentFileIndex].localFile << "). APIServer response for last try: '" << c->respData << "'" << endl;
c->clear();
chunksFailed.produce(c);
}
// Sleep for tiny amount of time, to make sure we yield to other threads.
// Note: boost::this_thread::yield() is not a valid interruption point,
// so we have to use sleep()
boost::this_thread::sleep(boost::posix_time::microseconds(100));
}
} catch(std::bad_alloc &e) {
boost::call_once(bad_alloc_once, boost::bind(&handle_bad_alloc, e));
} catch (boost::thread_interrupted &ti) {
return;
}
}