Re: public-inbox-convert hangs on systems using musl libc

[Eric Wong <e@80x24.org>]

Chris Brannon <chris@the-brannons.com> wrote:
> Eric Wong <e@80x24.org> writes:
> 
> > Do you know which pipes are which?  "lsof -p $PID +E" can help
> > with connectivity checking, as can script/dtas-graph in
> > https://80x24.org/dtas.git if you have Graph::Easy
> 
> Yes.  I'm attaching my lsof output and a typescript.
> 
> The processes of interest here are 4849 public-inbox-convert and 4879
> git cat-file.
> PID 4849's FD 11 is the write end of a pipe, with 4879's stdin as the
> read end.
> PID 4849's FD 12 is the read end of a pipe, with 4879's stdout as the
> write end.  At the point of the hang, 4849 is trying to write a SHA1 to
> FD 11, while 4879 is writing an email message to its stdout.

OK.  That is strange, because the current values are sized
conservatively for Linux (which has larger-than-required
PIPE_BUF size).

> > Some shots in the dark:
> >
> > 2. Tweak $PIPE_BUFSIZ and/or MAX_INFLIGHT to smaller values.  e.g.
> >
> > diff --git a/lib/PublicInbox/Git.pm b/lib/PublicInbox/Git.pm
> > index 882a9a4a..ec40edd7 100644
> > --- a/lib/PublicInbox/Git.pm
> > +++ b/lib/PublicInbox/Git.pm
> > @@ -23,13 +23,12 @@ use Carp qw(croak carp);
> >  use Digest::SHA ();
> >  use PublicInbox::DS qw(dwaitpid);
> >  our @EXPORT_OK = qw(git_unquote git_quote);
> > -our $PIPE_BUFSIZ = 65536; # Linux default
> > +our $PIPE_BUFSIZ = 4096; # Linux default
> >  our $in_cleanup;
> >  our $RDTIMEO = 60_000; # milliseconds
> >  our $async_warn; # true in read-only daemons
> >  
> > -use constant MAX_INFLIGHT => (POSIX::PIPE_BUF * 3) /
> > -	65; # SHA-256 hex size + "\n" in preparation for git using non-SHA1
> > +use constant MAX_INFLIGHT => 4;
> 
> This right here seems to have fixed it, when testing locally.

Are you able to isolate whether $PIPE_BUFSIZ or MAX_INFLIGHT on
it's own fixes it?

And can you confirm the ->blocking(0) change had no effect on
it's own?

Capping MAX_INFLIGHT to a smaller value is probably fine (maybe
10 can work).

The old MAX_INFLIGHT value ((512 * 3)/65 = 23) is actually
extremely conservative for Linux, since the smallest possible
PIPE_BUF is 4096 (so (4096 * 3)/65 = 189), but giant values
don't help (and hurt interruptibility).

However, making $PIPE_BUFSIZ smaller would increase syscalls
made and hurt performance on systems with expensive syscalls.
So I want to keep $PIPE_BUFSIZ as big as reasonable.  Setting
`$PIPE_BUFSIZ = 1024 ** 2' ought to work on a system with giant
pipes, even; but the default Linux pipe size is 65536 unless
it's low on memory.

I'm honestly at a loss on how to explain what went wrong for you
because the existing values are OS-independent.

I also do wonder if you've hit a kernel bug under low memory
conditions.  I've certainly encountered problems with TCP
traffic hanging processes due to memory compaction.

> PS.  Thank you for that lsof command.  I've never used lsof in that way;
> I'll have to add that to my *nix debugging toolbelt.

np :>