Re: [bug fix] postgres.exe crashes with access violation on Windows while starting up

On Wed, Nov 1, 2017 at 12:07 AM, Tsunakawa, Takayuki
<tsunakawa(dot)takay(at)jp(dot)fujitsu(dot)com> wrote:
> From: Michael Paquier [mailto:michael(dot)paquier(at)gmail(dot)com]
>> On Tue, Oct 31, 2017 at 6:59 AM, Tsunakawa, Takayuki
>> <tsunakawa(dot)takay(at)jp(dot)fujitsu(dot)com> wrote:
>> > When CurrentMemoryContext is NULL, the message is logged with
>> ReportEventA(). This is similar to write_console().
>>
>> My point is that as Postgres is running as a service, isn't it wrong to
>> write a message to stderr as a fallback if the memory context is not set?
>> You would lose a message. It seems to me that for an operation that can
>> happen at a low-level like the postmaster startup, we should really use
>> a low-level operation as well.
>
> I'm sorry I may not have been clear. With this patch, write_eventlog() outputs the message to event log with ReportEventA() when CurrentMemoryContext is NULL. It doesn't write to stderr. So the message won't be lost.

Oh, yes. Sorry. I got confused a bit by write_eventlog(), which is
already doing what your patch is adding for write_eventlog(). I am
switching the patch as ready for committer, I definitely agree that
you are taking the write approach here.

I am also adding Noah to get some input on this issue, as he is the
author and committer of 5f538ad0 which has improved the handling of
non-ASCII characters in this code path, and more importantly has
tweaked 43adc7a7 to handle properly transaction contexts in
pgwin32_message_to_UTF16() which is where the palloc calls happen. I
would be the one in the pool of committers who would most likely
commit your patch.
--
Michael