XrdBuffManager leaks memory

[gbitzes]

Hi,

There's this note in the destructor of XrdBuffManager:

           ~XrdBuffManager() {} // The buffmanager is never deleted

My unit tests for quarkdb involve lots of request parsing, which uses XrdBuffManager internally for memory recycling. Every unit test is independent of each other, meaning it sets up its own XrdBuffManager which gets deleted at the end, but the buffers are not freed.. This makes running a leak detector difficult, since at the end I get a ton of such messages:

Direct leak of 432 byte(s) in 18 object(s) allocated from:
    #0 0x7fbc8611bb62 in operator new(unsigned long) (/home/geobijes/dev/quarkdb/build/src/libXrdRedis.so+0x240b62)
    #1 0x7fbc85cb08be in XrdBuffManager::Obtain(int) /home/geobijes/dev/xrootd/src/Xrd/XrdBuffer.cc:169
    #2 0x7fbc8606d1e0 in quarkdb::BufferedReader::BufferedReader(quarkdb::Link*, XrdBuffManager*, unsigned long) /home/geobijes/dev/quarkdb/src/BufferedReader.cc:33
    #3 0x7fbc8606ecc2 in quarkdb::RedisParser::RedisParser(quarkdb::Link*, XrdBuffManager*) /home/geobijes/dev/quarkdb/src/RedisParser.cc:28
    #4 0x7fbc86085dd7 in quarkdb::Poller::worker(int, quarkdb::Dispatcher*) /home/geobijes/dev/quarkdb/src/Poller.cc:60
    #5 0x7fbc86089e43 in void std::_Mem_fn_base<void (quarkdb::Poller::*)(int, quarkdb::Dispatcher*), true>::operator()<int, quarkdb::Dispatcher*, void>(quarkdb::Poller*, int&&, quarkdb::Dispatcher*&&) const /usr/include/c++/5/functional:600
    #6 0x7fbc86089ca2 in void std::_Bind_simple<std::_Mem_fn<void (quarkdb::Poller::*)(int, quarkdb::Dispatcher*)> (quarkdb::Poller*, int, quarkdb::Dispatcher*)>::_M_invoke<0ul, 1ul, 2ul>(std::_Index_tuple<0ul, 1ul, 2ul>) /usr/include/c++/5/functional:1531
    #7 0x7fbc86089a8f in std::_Bind_simple<std::_Mem_fn<void (quarkdb::Poller::*)(int, quarkdb::Dispatcher*)> (quarkdb::Poller*, int, quarkdb::Dispatcher*)>::operator()() /usr/include/c++/5/functional:1520
    #8 0x7fbc860899bf in std::thread::_Impl<std::_Bind_simple<std::_Mem_fn<void (quarkdb::Poller::*)(int, quarkdb::Dispatcher*)> (quarkdb::Poller*, int, quarkdb::Dispatcher*)> >::_M_run() /usr/include/c++/5/thread:115
    #9 0x7fbc85767c7f  (/usr/lib/x86_64-linux-gnu/libstdc++.so.6+0xb8c7f)

This might be worth fixing, if only to make the use of leak detectors more pleasant.

[abh3]

We could do that but it still wouldn't be able to free any buffers that
were handed out (and not yet returned) so you will still get lost memory
messages at the end. Is that still acceptable?

Andy

On Wed, 12 Oct 2016, Georgios Bitzes wrote:

Hi,

There's this note in the destructor of XrdBuffManager:

          ~XrdBuffManager() {} // The buffmanager is never deleted

My unit tests for quarkdb involve lots of request parsing, which uses XrdBuffManager internally for memory recycling. Every unit test is independent of each other, meaning it sets up its own XrdBuffManager which gets deleted at the end, but the buffers are not freed.. This makes running a leak detector difficult, since at the end I get a ton of such messages:

Direct leak of 432 byte(s) in 18 object(s) allocated from:
   #0 0x7fbc8611bb62 in operator new(unsigned long) (/home/geobijes/dev/quarkdb/build/src/libXrdRedis.so+0x240b62)
   #1 0x7fbc85cb08be in XrdBuffManager::Obtain(int) /home/geobijes/dev/xrootd/src/Xrd/XrdBuffer.cc:169
   #2 0x7fbc8606d1e0 in quarkdb::BufferedReader::BufferedReader(quarkdb::Link*, XrdBuffManager*, unsigned long) /home/geobijes/dev/quarkdb/src/BufferedReader.cc:33
   #3 0x7fbc8606ecc2 in quarkdb::RedisParser::RedisParser(quarkdb::Link*, XrdBuffManager*) /home/geobijes/dev/quarkdb/src/RedisParser.cc:28
   #4 0x7fbc86085dd7 in quarkdb::Poller::worker(int, quarkdb::Dispatcher*) /home/geobijes/dev/quarkdb/src/Poller.cc:60
   #5 0x7fbc86089e43 in void std::_Mem_fn_base<void (quarkdb::Poller::*)(int, quarkdb::Dispatcher*), true>::operator()<int, quarkdb::Dispatcher*, void>(quarkdb::Poller*, int&&, quarkdb::Dispatcher*&&) const /usr/include/c++/5/functional:600
   #6 0x7fbc86089ca2 in void std::_Bind_simple<std::_Mem_fn<void (quarkdb::Poller::*)(int, quarkdb::Dispatcher*)> (quarkdb::Poller*, int, quarkdb::Dispatcher*)>::_M_invoke<0ul, 1ul, 2ul>(std::_Index_tuple<0ul, 1ul, 2ul>) /usr/include/c++/5/functional:1531
   #7 0x7fbc86089a8f in std::_Bind_simple<std::_Mem_fn<void (quarkdb::Poller::*)(int, quarkdb::Dispatcher*)> (quarkdb::Poller*, int, quarkdb::Dispatcher*)>::operator()() /usr/include/c++/5/functional:1520
   #8 0x7fbc860899bf in std::thread::_Impl<std::_Bind_simple<std::_Mem_fn<void (quarkdb::Poller::*)(int, quarkdb::Dispatcher*)> (quarkdb::Poller*, int, quarkdb::Dispatcher*)> >::_M_run() /usr/include/c++/5/thread:115
   #9 0x7fbc85767c7f  (/usr/lib/x86_64-linux-gnu/libstdc++.so.6+0xb8c7f)

This might be worth fixing, if only to make the use of leak detectors more pleasant.

You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub:
#424

[gbitzes]

Hi Andy,

Yes, of course that's acceptable. If the buffers are not eventually released back to the pool, and not free'd manually, that's probably a bug of the user of the class, and a valid case of a memory leak to be reported.

[abh3]

Fixed in 4.6.