Re: Re: [COMMITTERS] pgsql: Perform only one ReadControlFile() during startup.

On 2017-09-19 13:15:28 -0400, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > On 2017-09-19 13:00:33 -0400, Robert Haas wrote:
> >> You mean, in the postmaster?
>
> > Yes. We try to avoid touch shmem there, but it's not like we're
> > succeeding fully. See e.g. the pgstat_get_crashed_backend_activity()
> > calls (which do rely on shmem being ok to some extent), pmsignal,
> > BackgroundWorkerStateChange(), ...
>
> Well, the point is to avoid touching data structures that could be
> corrupted enough to confuse the postmaster.  I don't have any problem with
> adding some more functionality to pmsignal, say.

Given that we're ok with reading pgstat shared memory entries, I think
adding a carefully coded variant of SendProcSignal() should be doable in
a safe manner.

Something roughly like

int
PostmasterSendProcSignal(pid_t pid, ProcSignalReason reason)
{
    volatile ProcSignalSlot *slot;

    /*
     * As this is running in postmaster, be careful not to dereference
     * any pointers from shared memory that could be corrupted, and to
     * not to throw errors.
     */

    for (i = 0; i < NumProcSignalSlots; i++)
    {
        slot = &ProcSignalSlots[i];

        if (slot->pss_pid == pid)
        {
            /*
             * The note about race conditions in SendProcSignal applies
             * here, too
             */

            /* Atomically set the proper flag */
            slot->pss_signalFlags[reason] = true;
            /* Send signal */
            return kill(pid, SIGUSR1);
        }
    }

    errno = ESRCH;
    return -1;
}

As all the memory offsets are computed based on postmaster process-local
variables, this should be safe.

I'd rather like to avoid a copy of the procsignal infrastructure if we
don't need it...

Greetings,

Andres Freund