database/mariadb-columnstore-engine
1
0
Fork 0
mirror of https://github.com/mariadb-corporation/mariadb-columnstore-engine.git synced 2025-07-30 19:23:07 +03:00
Code Activity

fix(DEC): MCOL-5602 fixing potentially endless loop in DEC (#3049)

Browse Source
This commit is contained in:
drrtuy
2023-12-05 17:39:24 +02:00
committed by GitHub
parent 63b032e3fd
commit 1d416bc6ed
3 changed files with 73 additions and 79 deletions
Download Patch File Download Diff File Expand all files Collapse all files

133
dbcon/joblist/distributedenginecomm.cpp
View File

@ -356,8 +356,6 @@ int32_t DistributedEngineComm::Setup()
fWlock.swap(newLocks);
fPmConnections.swap(newClients);
// memory barrier to prevent the pmCount assignment migrating upward
atomicops::atomicMb();
pmCount = newPmCount;
newLocks.clear();
@ -419,14 +417,14 @@ Error:
for (map_tok = fSessionMessages.begin(); map_tok != fSessionMessages.end(); ++map_tok)
{
map_tok->second->queue.clear();
(void)atomicops::atomicInc(&map_tok->second->unackedWork[0]);
++map_tok->second->unackedWork[0];
map_tok->second->queue.push(sbs);
}
lk.unlock();
if (fIsExeMgr)
{
decltype(pmCount) originalPMCount = pmCount;
const uint32_t originalPMCount = pmCount;
// Re-establish if a remote PM restarted.
std::this_thread::sleep_for(std::chrono::seconds(3));
auto rc = Setup();
@ -455,7 +453,7 @@ void DistributedEngineComm::addQueue(uint32_t key, bool sendACKs)
mqe->throttled = false;
std::lock_guard lk(fMlock);
b = fSessionMessages.insert(pair<uint32_t, boost::shared_ptr<MQE> >(key, mqe)).second;
b = fSessionMessages.insert(pair<uint32_t, boost::shared_ptr<MQE>>(key, mqe)).second;
if (!b)
{
@ -635,7 +633,7 @@ void DistributedEngineComm::sendAcks(uint32_t uniqueID, const vector<SBS>& msgs,
size_t queueSize)
{
ISMPacketHeader* ism;
uint32_t l_msgCount = msgs.size();
size_t l_msgCount = msgs.size();
/* If the current queue size > target, do nothing.
* If the original queue size > target, ACK the msgs below the target.
@ -643,14 +641,13 @@ void DistributedEngineComm::sendAcks(uint32_t uniqueID, const vector<SBS>& msgs,
if (!mqe->throttled || queueSize >= mqe->targetQueueSize)
{
/* no acks will be sent, but update unackedwork to keep the #s accurate */
uint16_t numack = 0;
uint32_t sockidx = 0;
while (l_msgCount > 0)
{
nextPMToACK(mqe, l_msgCount, &sockidx, &numack);
idbassert(numack <= l_msgCount);
l_msgCount -= numack;
size_t numAcked = subsMsgCounterAndRotatePM(mqe, l_msgCount, &sockidx);
idbassert(numAcked <= l_msgCount);
l_msgCount -= numAcked;
}
return;
@ -665,9 +662,8 @@ void DistributedEngineComm::sendAcks(uint32_t uniqueID, const vector<SBS>& msgs,
{
/* update unackedwork for the overage that will never be acked */
int64_t overage = queueSize + totalMsgSize - mqe->targetQueueSize;
uint16_t numack = 0;
uint32_t sockidx = 0;
uint32_t msgsToIgnore;
size_t msgsToIgnore;
for (msgsToIgnore = 0; overage >= 0; msgsToIgnore++)
overage -= msgs[msgsToIgnore]->lengthWithHdrOverhead();
@ -679,16 +675,15 @@ void DistributedEngineComm::sendAcks(uint32_t uniqueID, const vector<SBS>& msgs,
while (msgsToIgnore > 0)
{
nextPMToACK(mqe, msgsToIgnore, &sockidx, &numack);
idbassert(numack <= msgsToIgnore);
msgsToIgnore -= numack;
size_t numAcked = subsMsgCounterAndRotatePM(mqe, msgsToIgnore, &sockidx);
idbassert(numAcked <= msgsToIgnore);
msgsToIgnore -= numAcked;
}
}
if (l_msgCount > 0)
{
SBS msg(new ByteStream(sizeof(ISMPacketHeader)));
uint16_t* toAck;
vector<bool> pmAcked(pmCount, false);
ism = (ISMPacketHeader*)msg->getInputPtr();
@ -698,7 +693,6 @@ void DistributedEngineComm::sendAcks(uint32_t uniqueID, const vector<SBS>& msgs,
ism->Interleave = uniqueID;
ism->Command = BATCH_PRIMITIVE_ACK;
toAck = &ism->Size;
msg->advanceInputPtr(sizeof(ISMPacketHeader));
// There must be only one local connection here.
@ -710,9 +704,20 @@ void DistributedEngineComm::sendAcks(uint32_t uniqueID, const vector<SBS>& msgs,
/* This will reset the ACK field in the Bytestream directly, and nothing
* else needs to change if multiple msgs are sent. */
nextPMToACK(mqe, l_msgCount, &sockIndex, toAck);
idbassert(*toAck <= l_msgCount);
l_msgCount -= *toAck;
size_t numAcked = subsMsgCounterAndRotatePM(mqe, l_msgCount, &sockIndex);
// Potential truncation here from size_t to 16 bits.
ism->Size = static_cast<uint16_t>(
std::min(static_cast<size_t>(std::numeric_limits<uint16_t>::max()), numAcked));
idbassert(numAcked <= l_msgCount);
l_msgCount -= numAcked;
// We get some logs if the above mentioned truncation happens.
if (numAcked > std::numeric_limits<uint16_t>::max())
{
cerr << "DEC::sendAcks(): msgs acked number [" << numAcked
<< "] overflows PrimMsg protocol header field capacity 2^16." << std::endl;
}
if (sockIndex == localConnectionId_ && fIsExeMgr)
{
sendToLocal = true;
@ -739,7 +744,7 @@ void DistributedEngineComm::sendAcks(uint32_t uniqueID, const vector<SBS>& msgs,
if (totalUnackedWork == 0)
{
*toAck = 1;
ism->Size = 1; // size of the batch acked
for (uint32_t i = 0; i < pmCount; ++i)
{
@ -761,8 +766,8 @@ void DistributedEngineComm::sendAcks(uint32_t uniqueID, const vector<SBS>& msgs,
}
}
void DistributedEngineComm::nextPMToACK(boost::shared_ptr<MQE> mqe, uint32_t maxAck, uint32_t* sockIndex,
uint16_t* numToAck)
size_t DistributedEngineComm::subsMsgCounterAndRotatePM(boost::shared_ptr<MQE> mqe, const size_t maxAck,
uint32_t* sockIndex)
{
uint32_t& nextIndex = mqe->ackSocketIndex;
@ -770,56 +775,57 @@ void DistributedEngineComm::nextPMToACK(boost::shared_ptr<MQE> mqe, uint32_t max
* the locking env, mqe->unackedWork can only grow; whatever gets latched in this fcn
* is a safe minimum at the point of use. */
if (mqe->unackedWork[nextIndex] >= maxAck)
if (mqe->unackedWork[nextIndex].load() >= maxAck)
{
(void)atomicops::atomicSub(&mqe->unackedWork[nextIndex], maxAck);
atomicops::atomicSubstitute(mqe->unackedWork[nextIndex], maxAck);
*sockIndex = nextIndex;
// FIXME: we're going to truncate here from 32 to 16 bits. Hopefully this will always fit...
*numToAck = maxAck;
if (pmCount > 0)
nextIndex = (nextIndex + 1) % pmCount;
return;
return maxAck;
}
else
{
for (int i = pmCount - 1; i >= 0; --i)
{
uint32_t curVal = mqe->unackedWork[nextIndex];
uint32_t unackedWork = (curVal > maxAck ? maxAck : curVal);
const size_t curVal = mqe->unackedWork[nextIndex].load();
const size_t unackedWork = std::min(curVal, maxAck);
if (unackedWork > 0)
{
(void)atomicops::atomicSub(&mqe->unackedWork[nextIndex], unackedWork);
atomicops::atomicSubstitute(mqe->unackedWork[nextIndex], unackedWork);
*sockIndex = nextIndex;
*numToAck = unackedWork;
if (pmCount > 0)
nextIndex = (nextIndex + 1) % pmCount;
return;
return unackedWork;
}
if (pmCount > 0)
nextIndex = (nextIndex + 1) % pmCount;
}
cerr << "DEC::nextPMToACK(): Couldn't find a PM to ACK! ";
cerr << "DEC::subsMsgCounterAndRotatePM(): Couldn't find a PM to ACK! ";
for (int i = pmCount - 1; i >= 0; --i)
cerr << mqe->unackedWork[i] << " ";
cerr << " max: " << maxAck;
cerr << endl;
cerr << " max: " << maxAck << endl;
// make sure the returned vars are legitimate
*sockIndex = nextIndex;
*numToAck = maxAck / pmCount;
if (pmCount > 0)
nextIndex = (nextIndex + 1) % pmCount;
// This will potentially results in invalid pmCount but it's better than crashing on 0 div.
const uint32_t localPmCount = pmCount.load();
if (localPmCount > 0)
{
nextIndex = (nextIndex + 1) % localPmCount;
return maxAck / localPmCount;
}
return;
cerr << "DEC::subsMsgCounterAndRotatePM(): The number of PMs is 0.";
return maxAck;
}
}
@ -970,7 +976,7 @@ void DistributedEngineComm::addDataToOutput(SBS sbs, uint32_t connIndex, Stats*
if (pmCount > 0)
{
(void)atomicops::atomicInc(&mqe->unackedWork[connIndex % pmCount]);
++mqe->unackedWork[connIndex % pmCount];
}
TSQSize_t queueSize = mqe->queue.push(sbs);
@ -1092,7 +1098,7 @@ int DistributedEngineComm::writeToClient(size_t aPMIndex, const SBS& bs, uint32_
for (map_tok = fSessionMessages.begin(); map_tok != fSessionMessages.end(); ++map_tok)
{
map_tok->second->queue.clear();
(void)atomicops::atomicInc(&map_tok->second->unackedWork[0]);
++map_tok->second->unackedWork[0];
map_tok->second->queue.push(sbs);
}
@ -1120,37 +1126,6 @@ int DistributedEngineComm::writeToClient(size_t aPMIndex, const SBS& bs, uint32_
// Connection was established.
return 1;
/*
// reconfig the connection array
ClientList tempConns;
{
//cout << "WARNING: DEC WRITE BROKEN PIPE " <<
fPmConnections[index]->otherEnd()<< endl; boost::mutex::scoped_lock onErrLock(fOnErrMutex); string
moduleName = fPmConnections[index]->moduleName();
//cout << "module name = " << moduleName << endl;
if (index >= fPmConnections.size()) return 0;
for (uint32_t i = 0; i < fPmConnections.size(); i++)
{
if (moduleName != fPmConnections[i]->moduleName())
tempConns.push_back(fPmConnections[i]);
}
if (tempConns.size() == fPmConnections.size()) return 0;
fPmConnections.swap(tempConns);
pmCount = (pmCount == 0 ? 0 : pmCount - 1);
}
// send alarm
ALARMManager alarmMgr;
string alarmItem("UNKNOWN");
if (index < fPmConnections.size())
{
alarmItem = fPmConnections[index]->addr2String();
}
alarmItem.append(" PrimProc");
alarmMgr.sendAlarmReport(alarmItem.c_str(), oam::CONN_FAILURE, SET);
*/
}
return 0;
}
@ -1206,9 +1181,15 @@ DistributedEngineComm::MQE::MQE(const uint32_t pCount, const uint32_t initialInt
const uint64_t flowControlEnableBytesThresh)
: ackSocketIndex(0), pmCount(pCount), hasBigMsgs(false), targetQueueSize(flowControlEnableBytesThresh)
{
unackedWork.reset(new volatile uint32_t[pmCount]);
interleaver.reset(new uint32_t[pmCount]);
memset((void*)unackedWork.get(), 0, pmCount * sizeof(uint32_t));
const uint32_t localPmCount = pmCount;
AtomicSizeVec s(localPmCount);
for (size_t i = 0; i < localPmCount; ++i)
{
s[i].store(0);
}
unackedWork.swap(s);
uint32_t interleaverValue = initialInterleaverValue;
initialConnectionId = initialInterleaverValue;
for (size_t pmId = 0; pmId < pmCount; ++pmId)

7
dbcon/joblist/distributedenginecomm.h
View File

@ -225,6 +225,7 @@ class DistributedEngineComm
// A queue of ByteStreams coming in from PrimProc heading for a JobStep
typedef ThreadSafeQueue<messageqcpp::SBS> StepMsgQueue;
using AtomicSizeVec = std::vector<std::atomic<size_t>>;
/* To keep some state associated with the connection. These aren't copyable. */
struct MQE : public boost::noncopyable
@ -235,7 +236,7 @@ class DistributedEngineComm
messageqcpp::Stats stats;
StepMsgQueue queue;
uint32_t ackSocketIndex;
boost::scoped_array<volatile uint32_t> unackedWork;
AtomicSizeVec unackedWork;
boost::scoped_array<uint32_t> interleaver;
uint32_t initialConnectionId;
uint32_t pmCount;
@ -283,7 +284,7 @@ class DistributedEngineComm
std::mutex fMlock; // sessionMessages mutex
std::vector<std::shared_ptr<std::mutex>> fWlock; // PrimProc socket write mutexes
bool fBusy;
volatile uint32_t pmCount;
std::atomic<uint32_t> pmCount;
boost::mutex fOnErrMutex; // to lock function scope to reset pmconnections under error condition
boost::mutex fSetupMutex;
@ -305,7 +306,7 @@ class DistributedEngineComm
void sendAcks(uint32_t uniqueID, const std::vector<messageqcpp::SBS>& msgs, boost::shared_ptr<MQE> mqe,
size_t qSize);
void nextPMToACK(boost::shared_ptr<MQE> mqe, uint32_t maxAck, uint32_t* sockIndex, uint16_t* numToAck);
size_t subsMsgCounterAndRotatePM(boost::shared_ptr<MQE> mqe, const size_t maxAck, uint32_t* sockIndex);
void setFlowControl(bool enable, uint32_t uniqueID, boost::shared_ptr<MQE> mqe);
void doHasBigMsgs(boost::shared_ptr<MQE> mqe, uint64_t targetSize);
boost::mutex ackLock;

12
utils/common/atomicops.h
View File

@ -22,6 +22,7 @@
#include <unistd.h>
#include <stdint.h>
#include <sched.h>
#include <atomic>
/*
This is an attempt to wrap the differneces between Windows and Linux around atomic ops.
@ -92,4 +93,15 @@ inline void atomicYield()
sched_yield();
}
// This f assumes decrement is smaller than minuend.
template <typename T>
inline void atomicSubstitute(typename std::atomic<T>& minuend, T decrement)
{
T expected = minuend.load();
do
{
expected = minuend.load();
} while (!minuend.compare_exchange_weak(expected, expected - decrement));
}
} // namespace atomicops