leak sanitizer reports are painfully slow under docker in some circumstances

[avikivity]

Under docker, the process can sometimes start with elevated RLIMIT_NOFILE (ulimit -n). Then, if a leak is detected (even if suppressed) this loop will spend minutes of CPU time closing billions of files:

llvm-project/compiler-rt/lib/sanitizer_common/sanitizer_posix_libcdep.cpp

Line 465 in 2e999b7

for (int fd = sysconf(_SC_OPEN_MAX); fd > 2; fd--) internal_close(fd);

Much better to use close_range() if available.

Semi-related: #48067

[MaddyGuthridge]

For anyone else encountering this issue, the workaround I'm using is to override the file limit. My OS has a limit of about 500k, so I decided to use that.

For command-line args: --ulimit nofile=524288 (source)

Inside a docker-compose.yml:

services:
  my-service:
    ulimits:
      nofile: 524288

Without limits: 0.00user 0.00system 2:01.57elapsed.

With limits: 0.00user 0.00system 0:00.09elapsed.

With a lower limit, it performs even better.