[EXPORTER] OTLP file: use thread-safe file/io

[owent]

Fixes #2674

Changes

• Use fwrite, fflush to keep it thread-safety.
• Add OPENTELEMETRY_ATTRIBUTE_LIFETIME_BOUND , OPENTELEMETRY_SANITIZER_NO_MEMORY , OPENTELEMETRY_SANITIZER_NO_THREAD, OPENTELEMETRY_SANITIZER_NO_ADDRESS, OPENTELEMETRY_HAVE_BUILTIN, OPENTELEMETRY_HAVE_FEATURE, OPENTELEMETRY_HAVE_ATTRIBUTE, OPENTELEMETRY_HAVE_CPP_ATTRIBUTE

For significant contributions please make sure you have completed the following items:

• CHANGELOG.md updated for non-trivial changes
• Unit tests have been added
• Changes in public API reviewed

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 87.67%. Comparing base (497eaf4) to head (93ec062).
Report is 77 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2675      +/-   ##
==========================================
+ Coverage   87.12%   87.67%   +0.55%     
==========================================
  Files         200      190      -10     
  Lines        6109     5851     -258     
==========================================
- Hits         5322     5129     -193     
+ Misses        787      722      -65     

Files	Coverage Δ
api/include/opentelemetry/nostd/string_view.h	97.96% <100.00%> (ø)

... and 75 files with indirect coverage changes

[marcalff]

Thanks for the PR.

The macros and annotations for sanitizers look good to me.

For the thread safety issue with file I/O,
I don't understand the initial problem, and the precise root cause.

In my understanding:

• file iostream operations are thread safe, when each thread uses its own stream
• file iostream operations are not thread safe when multiple threads uses the same stream
• file io operations are thread safe, when each thread uses its own FILE handle
• file io operations are not thread safe (in theory), when multiple threads uses the same FILE handle. In practive, they may appear to be safe if the implementation serializes calls.

Changing iostream to file io does not look like a proper fix for thread safety:

• If several threads can write to the same handle, do we need a mutex to serialize operations to write completely a record then ?

In particular, with:

  fwrite(data.data(), 1, data.size(), out.get());
  fputc('\n', out.get());

This does not look safe, even when fwrite() and fputc() are safe, if there are really several threads executing this code concurrently for the same handle (both records will be intertwined: 2 data, then 2 end of line).

Changing file stream with plain file io may be ok (even desirable) otherwise, but I don't understand how this is a fix for thread safety: please clarify.

[owent]

• file iostream operations are not thread safe when multiple threads uses the same stream

file iostream operations are not thread safe when multiple threads uses the same stream(we got data race from std::basic_streambuf<CharT,Traits>::xsputn when using tsan). But fwrite and fflush are thread safe when multiple threads uses the same FILE handle.

You are right, fwrite(data.data(), 1, data.size(), out.get()); and fputc('\n', out.get()); still have concurrency problems here. I have move the codes to append EOL in the caller.

[marcalff]

LGTM, thanks for the fix.

[marcalff]

@esigo Hi Ehsan. Any comments for the code review ?

[esigo]

LGTM
thanks for the PR :)