At Fri, 20 Dec 2019 12:16:32 +0900 (JST), Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote in
> At Thu, 19 Dec 2019 15:09:45 -0500, Tom Lane <tgl@sss.pgh.pa.us> wrote in
> > Alexander Lakhin <exclusion@gmail.com> writes:
> > > It seems that the check for ERROR_DELETE_PENDING was added to
> > > pgwin32_safestat() blindly, the issue wasn't reproduced at that time:
> > > https://www.postgresql.org/message-id/CAB7nPqRJV6trFta-Qzgi6j2feuYR2ZC%2BKHvWdHnbpDG2scTrxw%40mail.gmail.com
> >
> > Hmm, makes one wonder whether that's actually live code.
>
> Even if it is actually dead code, it seems reasonable as it stands
> since it is intending to read status of an existing file and the
> caller is assumed not to be knowing of ERROR_ACCESS_DENIED. In our
> case, pgwin32_open should be conscious of the state so I think calling
> pgwin32_safestat does not fit. (Or pgwin32_open and pgwin32_safestat
> are on the same level of API.)
>
> Maybe we could just loop back without extra stat'ing. With the
> following change, fopen(pidfile, 'r') immediately returns if the file
> is being deleted. Postmaster waits for the file gone (it would be
> already gone at the first try in most cases). The sleep on
> ERROR_ACCESS_DENEID moves from pg_ctl to postmaster in that case,
> but I think it doesn't matter.
It is wrong. It hides real "access denied".. Sorry for the noise.
> while (CreateFile())
> {
> ...
> if (err == ERROR_ACCESS_DENIED)
> {
> if ((fileFlags & O_CREAT) == 0)
> <retun with ENOENT>
>
> pg_usleep(...);
> loops++;
> continue;
regards.
--
Kyotaro Horiguchi
NTT Open Source Software Center