Re: Re: Proposed Windows-specific change: Enable crash dumps (like core files)

OK, it's pretty much ready for proper testing now. If a few people are
happy with the results and think it's a good idea I'll chuck it in the
commitfest app.

As built, the crash dump handler works with a stock PostgreSQL 9.0
(release build) as shipped in EDB's installer. Just drop crashdump.dll
in lib\, optionally pop the dbghelp.dll redist in bin\, add 'crashdump'
to shared_preload_libraries, and crash some backends however you feel
like doing so.

The current build of crashdump.dll for release versions of PostgreSQL
9.0 on 32-bit Windows is here:

http://www.postnewspapers.com.au/~craig/webfiles/crashdump.dll

If folks are happy with how this works, all it needs is:

- Consideration of whether elog should be used or not. I'm inclined to
suggest using elog to tell the user about the dump, but only after
the crash dump has been written successfully.

- Comments on whether this should be left as a loadable module, or
integerated into core so it loads early in backend startup. The latter
will permit crash dumping of early backend startup problems, and will
have tiny overhead because there's no DLL to find and load. OTOH, it's
harder to pull out if somehow something breaks.

I'd want to start with loadable module in shared_preload_libraries
and if that proves useful, only then consider integrating in core.
I'm way too bad a programmer to want my code anywhere near Pg's core
without plenty of real world testing.

- (Maybe) a method to configure the crash dump type. I'm not too sure
it's necessary given the set of dump flags I've landed up with,
so I'd leave this be unless it proves to be necessary in real-world
testing.

Remember that with these crash dumps the user only has to email the
(~4MB in my tests) crash dump. They don't need debugging tools or the
ability to use them, only the recipient does.

I'm using a tweaked set of minidump flags now, to generate considerably
bigger dumps (around 4MB for the configuration I'm testing) that contain
pretty much everything except shared memory contents, the contents of
memory mapped files, and the loaded read-only code segments of the
executables and DLLs. You can see the results of using it to debug that
autovac crash at the end of this mail.

When testing the script provided by Andrea Peri, when the autovacuum
worker crashes, the message:

> 2010-10-05 22:23:44 WST 2040 WARNING: crashdump: wrote crash dump to crashdumps\postgres-2040.mdmp

is emitted before the process resumes crashing out as it would've normally.

Opening and displaying that dump in windbg shows a useful stack trace
from the crashing process, and it's possible to examine the state of
local/global variables etc.

> 0:000> .ecxr
> eax=90fffffe ebx=040af140 ecx=68f08610 edx=040af248 esi=011f64d4 edi=040b001c
> eip=015d5267 esp=0055f1cc ebp=011f5bc8 iopl=0 nv up ei pl zr na pe nc
> cs=001b ss=0023 ds=0023 es=0023 fs=003b gs=0000 efl=00010246
> postgres!pfree+0x7:
> 015d5267 8b5004 mov edx,dword ptr [eax+4] ds:0023:91000002=????????
> 0:000> ~*k
>
> . 0 Id: 7f8.7b8 Suspend: 0 Teb: 7ffdf000 Unfrozen
> ChildEBP RetAddr
> 0055e210 75b4c1ee ntdll!KiFastSystemCallRet
> 0055e250 76e7100e KERNELBASE!FindClose+0x93
> 0055e514 679160c3 kernel32!_SEH_epilog4_GS+0xa
> 0055e6cc 779734e0 dbghelp!Win32LiveSystemProvider::OpenMapping+0x2c3
> 0055e7a8 7797353a ntdll!RtlpAllocateHeap+0xab2
> 0055e828 77965edc ntdll!RtlAllocateHeap+0x23a
> 0055e840 76e7123a ntdll!ZwWriteFile+0xc
> 0055e874 67916861 kernel32!WriteFileImplementation+0x76
> 0055e8ac 67916910 dbghelp!Win32LiveSystemProvider::ReadVirtual+0x71
> 0055e8d8 67908f09 dbghelp!Win32LiveSystemProvider::ReadAllVirtual+0x30
> 0055e910 679094b4 dbghelp!WriteMemoryFromProcess+0xa9
> 0055e9a8 6790d522 dbghelp!WriteThreadList+0x184
> 0055e9c0 6790e65a dbghelp!WriteDumpData+0x102
> 0055eb58 6790e9cb dbghelp!MiniDumpProvideDump+0x3fa
> 0055ebd0 73231353 dbghelp!MiniDumpWriteDump+0x1cb
> 0055ed14 76e82c4a crashdump!crashDumpHandler+0x183 [c:\users\craig\developer\postgresql-9.0.0\contrib\crashdump\crashdump.c @ 159]
> 0055ed9c 77995af4 kernel32!UnhandledExceptionFilter+0x127
> 0055eda4 7793d964 ntdll!__RtlUserThreadStart+0x62
> 0055edb8 7793d7fc ntdll!_EH4_CallFilterFunc+0x12
> 0055ede0 779665f9 ntdll!_except_handler4+0x8e
> 0055ee04 779665cb ntdll!ExecuteHandler2+0x26
> 0055eeb4 77966457 ntdll!ExecuteHandler+0x24
> 0055eeb4 015d5267 ntdll!KiUserExceptionDispatcher+0xf
> 0055f1c8 0139eee7 postgres!pfree+0x7 [c:\pginstaller-repo\postgres.windows\src\backend\utils\mmgr\mcxt.c @ 591]
> 0055f1e0 0139f14a postgres!examine_attribute+0x207 [c:\pginstaller-repo\postgres.windows\src\backend\commands\analyze.c @ 877]
> 0055f284 0139f94c postgres!do_analyze_rel+0x24a [c:\pginstaller-repo\postgres.windows\src\backend\commands\analyze.c @ 357]
> 0055f2ac 013eb74a postgres!analyze_rel+0x1bc [c:\pginstaller-repo\postgres.windows\src\backend\commands\analyze.c @ 232]
> 0055f330 014b30c6 postgres!vacuum+0x1ea [c:\pginstaller-repo\postgres.windows\src\backend\commands\vacuum.c @ 248]
> 0055f368 014b3991 postgres!autovacuum_do_vac_analyze+0x86 [c:\pginstaller-repo\postgres.windows\src\backend\postmaster\autovacuum.c @ 2651]
> 0055f4f4 014b41c5 postgres!do_autovacuum+0x811 [c:\pginstaller-repo\postgres.windows\src\backend\postmaster\autovacuum.c @ 2231]
> 0055f588 014bed02 postgres!AutoVacWorkerMain+0x265 [c:\pginstaller-repo\postgres.windows\src\backend\postmaster\autovacuum.c @ 1611]
> 0055f7d8 01423ce8 postgres!SubPostmasterMain+0x442 [c:\pginstaller-repo\postgres.windows\src\backend\postmaster\postmaster.c @ 4093]
> 0055f7f0 015ee1ad postgres!main+0x1f8 [c:\pginstaller-repo\postgres.windows\src\backend\main\main.c @ 165]
> 0055f834 76e71194 postgres!__tmainCRTStartup+0x10f [f:\dd\vctools\crt_bld\self_x86\crt\src\crtexe.c @ 586]
> 0055f840 7797b495 kernel32!BaseThreadInitThunk+0xe
> 0055f880 7797b468 ntdll!__RtlUserThreadStart+0x70
> 0055f898 00000000 ntdll!_RtlUserThreadStart+0x1b