lkml.org 
[lkml]   [2003]   [Sep]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] fix IO hangs
On Fri, Sep 05 2003, Nick Piggin wrote:
>
>
> Jens Axboe wrote:
>
> >On Fri, Sep 05 2003, Nick Piggin wrote:
> >
> >>Hi, sorry for the hangs, everyone. I think I have it worked out, but
> >>testers and an ack from Jens would be good.
> >>
> >>The insert_here code now does as advertised. The big difference being
> >>that regular blk_fs_requests will be subject to it (required for SCSI
> >>requeue). Unfortunately ll_rw_blk.c misuses it and will sometimes try
> >>to insert at requests which are not on the dispatch list, causing the
> >>badness.
> >>
> >>It looks like the code was maybe used to provide an insertion hint
> >>for the elevator. The RB tree has now eliminated that requirement even
> >>if the code did work. Which it doesn't.
> >>
> >>I can't reproduce the hangs with this patch. Please test.
> >>
> >>
> >>Aside, insert_here really seems to be quite dangerous to me. I think
> >>combination of barriers and an "insert at start/end" flag would be
> >>enough.
> >>
> >
> >Please just kill insert_here, it's exceeded its life expectancy. The 2.2
> >io scheduler did a merge scan followed by an insertion scan, 2.3
> >collapsed them into one scan as an optimization. Performance oriented io
> >schedulers need to use better data structures.
> >
> >Best would be to change it to pass a request back even for the NO_MERGE
> >case, if it has found a good insertion point. It's still a good idea to
> >be able to pass hints back like this, as it could still be a viable
> >optimization for _other_ io schedulers.
> >
>
> OK, that is sort of an ACK! Pending wider testing this patch needs
> to get in.

Fold it in with your elv-insertion fix and let it test in -mm?

> Jens, if insert_here is dead, there is no point to passing back a hint
> because it can't get back to the elevator anyway.
>
> I'd very much like to kill insert_here and be done with it. If another
> io scheduler comes along with a good use for it then the writers can
> come up with an elegant solution ;) Hey, if they know a NO_MERGE return
> means an insert will soon happen under the same lock, they could keep
> it cached privately.

Agree, lets just kill it off.

--
Jens Axboe

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-03-22 13:48    [W:0.363 / U:0.272 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site