vcmap concurrent put returns out of memory

[karczex]

ISSUE: vcmap concurrent put returns OOM:

Environment Information

• pmemkv version(s): 1.0 - 1.2
• libpmemobj-cpp version(s): built from source: 39895391980be1b18a946556cafb98dbf7276e4b
• PMDK (libpmem/libpmemobj) package version(s): 1.9
• OS(es) version(s): ubuntu 20.04, fedora 32
• compiler, libraries, packaging and other related tools version(s): gcc and clang

and possibly:

• TBB version(s):
• memkind version(s): 1.10

Please provide a reproduction of the bug:

#include <cassert>
#include <cstdlib>
#include <iostream>
#include <libpmemkv.hpp>
#include <sstream>

#include <numeric>
#include <thread>
#include <vector>

#define LOG(msg) std::cout << msg << std::endl

using namespace pmem::kv;

static inline pmem::kv::string_view uint64_to_strv(uint64_t &key)
{
        return pmem::kv::string_view((char *)&key, sizeof(uint64_t));
}

template <typename Function>
void parallel_exec(size_t threads_number, Function f)
{
        std::vector<std::thread> threads;
        threads.reserve(threads_number);

        for (size_t i = 0; i < threads_number; ++i) {
                threads.emplace_back(f, i);
        }

        for (auto &t : threads) {
                t.join();
        }
}

const uint64_t SIZE_16M = 16UL * 1024UL * 1024UL;
/*
 * /opt/workspace/pmemkv_upstream/build/tests/concurrent_put_get_remove_single_op_params
 * vcmap;{"path":"/dev/shm/vcmap__concurrent_put_get_remove_single_op_params__default_1000_0_none","size":104857600};1000
 */
int main(int argc, char *argv[])
{

        LOG("Creating config");
        config cfg;

        status s = cfg.put_string("path", "/dev/shm");
        assert(s == status::OK);
        s = cfg.put_uint64("size", SIZE_16M);
        assert(s == status::OK);

        const int threads_number = 16;

        LOG("test with " << threads_number << " threads");

        LOG("Opening pmemkv database with 'vcmap' engine");
        db *kv = new db();
        assert(kv != nullptr);
        s = kv->open("vcmap", std::move(cfg));
        assert(s == status::OK);

        std::vector<uint64_t> keys(threads_number, 0);
        std::iota(keys.begin(), keys.end(), 0);

        parallel_exec(threads_number, [&](size_t thread_id) {
                LOG("thread: " << thread_id);
                auto s = kv->put(uint64_to_strv(keys[thread_id]),
                                 uint64_to_strv(keys[thread_id]));
                if (static_cast<int>(s) != 0) {
                        LOG("status: " << static_cast<int>(s) << " ERROR occured");
                }
        });

        LOG("Closing database");
        delete kv;

        return 0;
}

dynamically linked libraries:

  NEEDED               libpmemkv.so.1
  NEEDED               libstdc++.so.6
  NEEDED               libm.so.6
  NEEDED               libgcc_s.so.1
  NEEDED               libpthread.so.0
  NEEDED               libc.so.6

How often bug is revealed:

often.

Actual behavior:

For some threads put returns out of memory caused by bad_alloc (libmemkind::pmem::allocator::allocate())

Details

#623 may be somehow related to this issue, however libunwind is not needed to trigger it.

[kilobyte]

100% reproducible for me.