Обсуждение: pg_receivewal.exe unhandled exception in zlib1.dll

Hi,

On 2022-02-11 16:33:11 +0100, Juan José Santamaría Flecha wrote:
> If you execute pg_receivewal.exe with the option "--compression-method
> gzip" it will fail with no error. You can see an error in the db log:
> 
> 2022-02-10 11:46:32.725 CET [11664][walsender] [pg_receivewal][3/0:0] LOG:
> could not receive data from client: An existing connection was forcibly
> closed by the remote host.
> 
> Tracing the execution you can see:
> 
> Unhandled exception at 0x00007FFEA78C1208 (ucrtbase.dll) in An invalid
> parameter was passed to a function that considers invalid parameters fatal.
>   ucrtbase.dll!00007ff8e8ae36a2() Unknown
> > zlib1.dll!gz_comp(gz_state * state, int flush) Line 111 C
>   zlib1.dll!gz_write(gz_state * state, const void * buf, unsigned __int64
> len) Line 235 C
>   pg_receivewal.exe!dir_write(void * f, const void * buf, unsigned __int64
> count) Line 300 C
>   pg_receivewal.exe!ProcessXLogDataMsg(pg_conn * conn, StreamCtl * stream,
> char * copybuf, int len, unsigned __int64 * blockpos) Line 1150 C
>   pg_receivewal.exe!HandleCopyStream(pg_conn * conn, StreamCtl * stream,
> unsigned __int64 * stoppos) Line 850 C
>   pg_receivewal.exe!ReceiveXlogStream(pg_conn * conn, StreamCtl * stream)
> Line 605 C
>   pg_receivewal.exe!StreamLog() Line 636 C
>   pg_receivewal.exe!main(int argc, char * * argv) Line 1005 C
> 
> The problem comes from the file descriptor passed to gzdopen() in
> 'src/bin/pg_basebackup/walmethods.c'. Using gzopen() instead, solves the
> issue without ifdefing for WIN32. Please find attached a patch for so.

I hit this as well. The problem is really caused by using a debug build of
postgres vs a production build of zlib. The use different C runtimes, with
different file descriptors. A lot of resources in the windows world are
unfortunately tied to the C runtime and that there can multiple C runtimes in
a single process.

Greetings,

Andres Freund

On Fri, Feb 11, 2022 at 07:50:44AM -0800, Andres Freund wrote:
> On 2022-02-11 16:33:11 +0100, Juan José Santamaría Flecha wrote:
>> The problem comes from the file descriptor passed to gzdopen() in
>> 'src/bin/pg_basebackup/walmethods.c'. Using gzopen() instead, solves the
>> issue without ifdefing for WIN32. Please find attached a patch for so.

This looks wrong to me as gzclose() would close() the file descriptor
we pass in via gzdopen(), and some code paths of walmethods.c rely on
this assumption so this would leak fds on repeated errors.

> I hit this as well. The problem is really caused by using a debug build of
> postgres vs a production build of zlib. The use different C runtimes, with
> different file descriptors. A lot of resources in the windows world are
> unfortunately tied to the C runtime and that there can multiple C runtimes in
> a single process.

It may be worth warning about that at build time, or just document the
matter perhaps?  I don't recall seeing any MSIs related to zlib that
have compile with the debug options.
--
Michael

Hi,

On 2022-02-15 11:21:42 +0100, Juan José Santamaría Flecha wrote:
> Building with a mix of debug and release components is not a good practice
> for issues like this, but is not something that MSVC forbids.

It's not a problem if you never need share memory, file descriptors, locales,
... state across DLL boundaries...


> If this is something we want to discard altogether, we can check at build
> time with 'dumpbin' to ensure that all dlls share the same vcruntime.dll.

I think these days it's mostly a question of ucrtbase.dll vs ucrtbased.dll,
right?

FWIW, I've looked at using either vcpkg or conan to more easily install /
build dependencies on windows, in the right debug / release mode.

The former was easier, but unfortunately the zlib, lz4, ... packages don't
include the gzip, lz4 etc binaries right now. That might be easy enough to fix
with an option. It does require building the dependencies locally.

Conan seemed to be a nicer architecture, but there's no builtin way to update
to newer dependency versions, and non-versioned dependencies don't actually
work. It has pre-built dependencies for the common cases.

Greetings,

Andres Freund