LeakSanitizer failure in ASAN mode

[wudisheng]

Describe the bug
We encountered an unexpected failure in ASAN mode when using BRPC 1.13.0. A simple testing program reports

Tracer caught signal 11: addr=0x7f2cb17fd000 pc=0x7f2cb6c9c7e8 sp=0x7f2caccf0d20
==2075==LeakSanitizer has encountered a fatal error.
==2075==HINT: For debugging, try setting environment variable LSAN_OPTIONS=verbosity=1:log_threads=1
==2075==HINT: LeakSanitizer does not work under ptrace (strace, gdb, etc)

Meanwhile, the same program runs well with BRPC 1.11.0 using exactly the same compilation and invocation commands.

To Reproduce
I have minimized the reproduction into the following code, in BRPC repository workspace, create the following files.

xxx/xxx.cc

#include <vector>

#include "bthread/bthread.h"
#include "butil/logging.h"

static void* Worker(void* argument) {
  for (int k = 0; k < 10; ++k) {
    LOG(WARNING) << "XXX: " << k;
    bthread_usleep(30000);
  }

  return nullptr;
}

int main() {
  bthread_setconcurrency(10);

  std::vector<bthread_t> workers;

  for (int k = 0; k < 3; ++k) {
    bthread_start_background(&workers.emplace_back(), nullptr, &Worker, nullptr);
  }

  for (const auto& worker : workers) {
    bthread_join(worker, nullptr);
  }
}

xxx/BUILD

cc_binary(
    name = "xxx",
    srcs = ["xxx.cc"],
    deps = [
        "//:brpc",
    ],
)

Then run with bazel run -c fastbuild --copt=-fsanitize=address --linkopt=-fsanitize=address --linkopt=-lunwind //xxx:xxx.

Note that when using BRPC 1.13.0 you may need to revise the BUILD.bazel file which has some minor BzlMod only callsites, if you are not using BzlMod. However, apparently this has nothing to do with the issue itself.

Expected behavior
Example program runs well with no ASAN or LSAN findings.

Versions
OS: Ubuntu 22.04
Compiler: gcc 11, gcc 12, clang 16, clang 18, clang 20
brpc: 1.13.0 (Problematic), 1.11.0 (Working)
protobuf: Unrelated (Example programs runs with the version each BRPC release specifies)

Additional context/screenshots
Maybe PR #2858 (released in BRPC 1.12.0) has something to do with this --- the LSAN messages disappeared when I hacked off the detection of BUTIL_USE_ASAN macro.

[chenBright]

This seems like a LeakSanitizer issue:

1. LeakSanitizer has encountered a fatal error google/sanitizers#1306
2. LeakSanitizer has encountered a fatal error google/sanitizers#764

[wudisheng]

Confirmed that it is indeed affected by /proc/sys/fs/suid_dumpable (and within a container). Thanks for your prompt reply and pointers.

[chenBright]

Confirmed that it is indeed affected by /proc/sys/fs/suid_dumpable (and within a container). Thanks for your prompt reply and pointers.

How do you solve this problem? Use docker run --privileged？

[wudisheng]

Presumably a privileged container resolves this issue, but we have no way to configure it like that. For now we just want to make sure it is a known issue and then bear with it.

[lhsoft]

I encountered this problem when using bazel test to run the test.
And cmake is ok.

[Yangfisher1]

I encountered the same problem. It's quite strange that I tried to compile another project with asan/lsan and it worked properly. However, when changed to brpc, it throw the same error. I guess maybe the linking of tcmalloc leads to this?