Hello hackers,
After fixing bug #16161 (pg_ctl inability to open just deleted
postmaster.pid) there are still some errors related to the same
Windows-only issue.
Namely, pg_upgradeCheck failures seen on
https://buildfarm.postgresql.org/cgi-bin/show_history.pl?nm=fairywren&br=REL_13_STABLE
Here pg_ls_dir_files couldn't handle the "delete pending" state too.
I've made a simple test to reproduce this behavior with pg_ls_waldir,
that fails with:
error running SQL: 'psql:<stdin>:1: ERROR: could not stat file
"pg_wal/dummy": Permission denied'
As noted in [1], a sensible solution would be putting the same "retry on
ERROR_ACCESS_DENIED" action in a wrapper for stat().
And bed90759f brought in master the _pgstat64() function, where such
error handling should be placed.
So please look at the patch (based on the previous one made to fix
bug#16161), that makes the attached test pass.
And I have a few questions.
For now genfile.c needs to include "port.h" explicitly to override
definitions in "sys/stat.h", that override "stat" redefinition for
Windows included implicitly above via "postgres.h". Without it the
pg_ls_dir_files() function just uses system' stat().
So I wonder, is there a way to make all stat-calling code use the
improved stat()?
And if so, maybe the stat() call in pgwin32_open should be replaced with
microsoft_native_stat().
And regarding released versions, are there any reasons to not backport
bed90759f (with the proposed fix or alike)?
[1] https://www.postgresql.org/message-id/396837.1605298573%40sss.pgh.pa.us
Best regards,
Alexander
