This is the mail archive of the mailing list for the elfutils project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: PR25394 patch: debuginfod groom vs. scan race condition

Hi -

> Nice work. And I think I almost understand it. I am just a little
> confused about thread_main_fts_source_paths also being an "idler":

Yup, it's a bit subtle.

> >   while (! interrupted)
> >      {
> > -      set_metric("thread_timer", "role","traverse", rescan_timer);
> > -      // set_metric("thread_forced_total", "role","traverse", forced_rescan_count);
> > -      if (rescan_s && rescan_timer > rescan_s)
> > -        rescan_timer = 0;
> > +      sleep (1);
> > +      scanq.wait_idle(); // don't start a new traversal while scanners haven't finished the job
> > +      scanq.done_idle(); // release the hounds
> > +      if (interrupted) break;

> I guess the comment "release the hounds" gets me confused. I don't
> know precisely what that means.

> Is this needed for correctness, or is it simply to have an hint that
> the scanners have ran because you register as an idler and so at
> this point you could only run if no scanner is running?

It's because we don't want to start a new fts traversal pass while the
scanners haven't even finished the last one.  I've seen a case where
the rescan interval was so short that the scanner backlog just kept
growing and growing and growing.  (Changing the work queue to a set
from a dequeue also helps with duplicate elimination.)  So in this
way, it just returns to the timing behaviour of the pre-workqueue
implementation: we wait a certain amount of time AFTER the scanning is
ALL DONE, before starting a new scan.

> But given you immediately tell them you are done a scanner could in
> theory start their work up again while you are doing your own
> traversal?

Yes, because that's ideal for concurrency: while one thread fts(3)'s
and enumerates the results, the scanner threads can process the
already enumerated files.

- FChE

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]