On Thu, Jun 30, 2016 at 6:01 AM, Alvaro Herrera
<alvherre@2ndquadrant.com> wrote:
> Alvaro Herrera wrote:
>
>> I propose to push this patch, closing the open item, and you can rework
>> on top -- I suppose you would completely remove the original conninfo
>> from shared memory and instead only copy the obfuscated version there
>> (and probably also remove the ready_to_display flag). I think we'd need
>> to see the patch before deciding whether we want it in 9.6 or not,
>> keeping in mind that having the conninfo in shared memory is a
>> pre-existing problem, unrelated to the pgstats view new in 9.6.
>
> Pushed this.
Thanks for pushing the patch!
But I found two problems in the patch you pushed.
(1)
ready_to_display flag must be reset to false when walreceiver dies.
Otherwise, pg_stat_wal_receiver can report the password (i.e.,
the problem that I reported upthread can happen) when walreceiver restarts
because ready_to_display flag is true from the beginning in that case.
But you forgot to reset the flag to false when walreceiver dies.
(2)
+retry:
+ SpinLockAcquire(&walrcv->mutex);
+ if (!walrcv->ready_to_display)
+ {
+ SpinLockRelease(&walrcv->mutex);
+ CHECK_FOR_INTERRUPTS();
+ pg_usleep(1000);
+ goto retry;
+ }
+ SpinLockRelease(&walrcv->mutex);
ISTM that we will never be able to get out of this loop if walreceiver
fails to connect to the master (e.g., password is wrong) after we enter
this loop.
Regards,
--
Fujii Masao
