Use PIPE_BUF over some magic number 1024 in terminate_handler, since
according to pipe(7):
PIPE_BUF
POSIX.1 says that write(2)s of less than PIPE_BUF bytes must be atomic
Also note that 1024, is too small, especially for C++ stacktraces (and
especially for debug builds, that contains lots of non-inlined helpers
for various ptrs).
Even after #20464 it was still possible, for example [1].
2021.02.19 11:40:21.886191 [ 68373 ] {} <Trace> DynamicQueryHandler: Request URI: /?database=test_ds2d6y&log_comment=/usr/share/clickhouse-test/queries/0_stateless/01302_aggregate_state_exception_memory_leak.sh&enable_http_compression=1&http_zlib_compression_level=1
<snip>
2021.02.19 11:41:35.289940 [ 365 ] {} <Fatal> BaseDaemon: (version 21.3.1.6058, build id: 8D46D65205E2C8B7FE408A0B4EC76CA0483F9E92) (from thread 68373) Terminate called for uncaught exception:
Code: 24, e.displayText() = DB::Exception: Cannot write to ostream at offset 262568, Stack trace (when copying this message, always include the lines below):
0. ./obj-x86_64-linux-gnu/../contrib/libcxx/include/exception:0: Poco::Exception::Exception(std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> > const&, int) @ 0x15b3c7db in /usr/bin/clickhouse
1. ./obj-x86_64-linux-gnu/../src/Common/Exception.cpp:56: DB::Exception::Exception(std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> > const&, int, bool) @ 0x8aba66e in /usr/bin/clickhouse
2. ./obj-x86_64-linux-gnu/../src/IO/WriteBufferFromOStream.cpp:0: DB::WriteBufferFromOStream::nextImpl() @ 0x8b8c105 in /usr/bin/clickhouse
3. ./obj-x86_64-linux-gnu/../src/IO/BufferBase.h:39: DB::WriteBufferFromOStream::~WriteBufferFromOStream() @ 0x8b8c537 in /usr/bin/clickhouse
4. ./obj-x86_64-linux-gnu/../src/IO/WriteBufferFromOStream.cpp:44: DB::Write
[1]: https://clickhouse-test-reports.s3.yandex.net/16481/5d150cce4778dd14f58dcff67435bdec1efa155b/stress_test_(thread).html#fail1
And according to this partial stacktrace it seems that the dtor of
WriteBufferFromOStream was called from
WriteBufferFromHTTPServerResponse, since the class name starts from
DB::Write*
The problem is that if first time WriteBufferFromOStream::next() fails,
it will reset position to make next write no-op, however
WriteBufferFromHTTPServerResponse::next() will set position to available
buffer back, and next() will throw again, but this time it can be from
dtor.
* Refactoring: part 1
* Refactoring: part 2
* Handle request using ReadBuffer interface
* Struggles with ReadBuffer's
* Fix URI parsing
* Implement parsing of multipart/form-data
* Check HTTP_LENGTH_REQUIRED before eof() or will hang
* Fix HTTPChunkedReadBuffer
* Fix build and style
* Fix test
* Resist double-eof
* Fix arcadian build
When distributed_group_by_no_merge=2 is used (or when
optimize_distributed_group_by_sharding_key takes place), remote servers
will do full ORDER BY, so initiator can skip this step and do only merge
of ordered blocks.
