From: "Magnus Hagander" <magnus@hagander.net>
> But yeah, that's probably a good idea. A quick look at the code says
we
> should at least ask people who have this problem to give it a run
with
> logging at DEBUG5 which should then log exactly what the errorcode
was.
> Or are you seeing more places that need such logging first?
I'm sorry we can't get get the Win32 error code to be displayed. I
got the following messages:
2007-01-16 09:24:48 DEBUG: checkpoint starting
2007-01-16 09:24:48 ERROR: could not open relation 1663/10819/18296:
Permission denied
2007-01-16 09:24:48 ERROR: checkpoint request failed
2007-01-16 09:24:48 HINT: Consult recent messages in the server log
for details.
2007-01-16 09:24:48 STATEMENT: checkpoint;
The reason is that src/port/open.c does not use _dosmaperr(). It
converts the Win32 error code to errno directly. EACCES is converted
from ERROR_ACCESS_DENIED only. Mmm, we may have to compromise as
Tom-san says.
BTW, why does the bgwriter try to open and write the pages of already
dropped relations? When dropping relations, DropRelFileNodeBuffers is
called to discard dirty buffers. If the relation being dropeed has
already been registered in the list of files to be fsynced, isn't it
possible to remove the file from the list before unlinking the file,
asking bgwriter in a similar way as ForwardFsyncRequest()?
#
The timestamp at the head of each message is noisy since the event
viewer has the time info, isn't it? Besides, several PostgreSQL
messages appeared as one entry of event log, separated by a LF instead
of CR LF. On Windows, CR LF should separate lines.