On Tue, Oct 25, 2022 at 03:29:42AM +0000, Thomas Munro wrote:
> Fix unlink() for STATUS_DELETE_PENDING on Windows.
>
> Commit f357233c assumed that it was OK to return ENOENT directly if
> lstat() failed that way. If we got STATUS_DELETE_PENDING while trying
> to unlink a file that we had already unlinked successfully once before
> but someone else still had open (on a kernel version that has "pending"
> unlinks by default), then we would no longer reach the retry loop in
> pgunlink(). That loop claims to be only for handling sharing violations
> (a different phenomenon), but the errno is the same.
>
> Restore that behavior with an explicit check, to see if it fixes the
> occasional 'directory not empty' failures seen in the pg_upgrade tests
> on CI. Further improvements are possible with proposed upgrades to
> modern Windows APIs that would replace this convoluted code.
>
> Reported-by: Justin Pryzby <pryzby@telsasoft.com>
> Reviewed-by: Michael Paquier <michael@paquier.xyz>
> Discussion: https://postgr.es/m/20220920013122.GA31833%40telsasoft.com
> Discussion: https://postgr.es/m/CA%2BhUKG%2BajSQ_8eu2AogTncOnZ5me2D-Cn66iN_-wZnRjLN%2Bicg%40mail.gmail.com
...
> Details
> -------
> https://git.postgresql.org/pg/commitdiff/e109e43921d21d069c03f18d7c9d8f4e5cb6a0c3
If I'm not wrong, this didn't fix the issue you said it fixed.
cfbot says this ran 1h ago, and its HEAD^3 is e109e43.
https://cirrus-ci.com/task/5939314583404544
