xfs
[Top] [All Lists]

Re: [PATCH] xfs: add readahead bufs to lru early to prevent post-unmount

To: Dave Chinner <david@xxxxxxxxxxxxx>
Subject: Re: [PATCH] xfs: add readahead bufs to lru early to prevent post-unmount panic
From: Brian Foster <bfoster@xxxxxxxxxx>
Date: Tue, 12 Jul 2016 08:03:15 -0400
Cc: xfs@xxxxxxxxxxx
Delivered-to: xfs@xxxxxxxxxxx
In-reply-to: <20160711224451.GF1922@dastard>
References: <1467291229-13548-1-git-send-email-bfoster@xxxxxxxxxx> <20160630224457.GT12670@dastard> <20160701223011.GA28130@xxxxxxxxxxxxxxx> <20160705164552.GA6317@xxxxxxxxxxxxxxx> <20160711052057.GE1922@dastard> <20160711135251.GA32896@xxxxxxxxxxxxxxx> <20160711152921.GB32896@xxxxxxxxxxxxxxx> <20160711224451.GF1922@dastard>
User-agent: Mutt/1.6.1 (2016-04-27)
On Tue, Jul 12, 2016 at 08:44:51AM +1000, Dave Chinner wrote:
> On Mon, Jul 11, 2016 at 11:29:22AM -0400, Brian Foster wrote:
> > On Mon, Jul 11, 2016 at 09:52:52AM -0400, Brian Foster wrote:
> > ...
> > > So what is your preference out of the possible approaches here? AFAICS,
> > > we have the following options:
> > > 
> > > 1.) The original "add readahead to LRU early" approach.
> > >   Pros: simple one-liner
> > >   Cons: bit of a hack, only covers readahead scenario
> > > 2.) Defer I/O count decrement to buffer release (this patch).
> > >   Pros: should cover all cases (reads/writes)
> > >   Cons: more complex (requires per-buffer accounting, etc.)
> > > 3.) Raw (buffer or bio?) I/O count (no defer to buffer release)
> > >   Pros: eliminates some complexity from #2
> > >   Cons: still more complex than #1, racy in that decrement does
> > >   not serialize against LRU addition (requires drain_workqueue(),
> > >   which still doesn't cover error conditions)
> > > 
> > > As noted above, option #3 also allows for either a buffer based count or
> > > bio based count, the latter of which might simplify things a bit further
> > > (TBD). Thoughts?
> 
> Pretty good summary :P
> 
> > FWIW, the following is a slightly cleaned up version of my initial
> > approach (option #3 above). Note that the flag is used to help deal with
> > varying ioend behavior. E.g., xfs_buf_ioend() is called once for some
> > buffers, multiple times for others with an iodone callback, that
> > behavior changes in some cases when an error is set, etc. (I'll add
> > comments before an official post.)
> 
> The approach looks good - I think there's a couple of things we can
> do to clean it up and make it robust. Comments inline.
> 
> > diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
> > index 4665ff6..45d3ddd 100644
> > --- a/fs/xfs/xfs_buf.c
> > +++ b/fs/xfs/xfs_buf.c
> > @@ -1018,7 +1018,10 @@ xfs_buf_ioend(
> >  
> >     trace_xfs_buf_iodone(bp, _RET_IP_);
> >  
> > -   bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD);
> > +   if (bp->b_flags & XBF_IN_FLIGHT)
> > +           percpu_counter_dec(&bp->b_target->bt_io_count);
> > +
> > +   bp->b_flags &= ~(XBF_READ | XBF_WRITE | XBF_READ_AHEAD | XBF_IN_FLIGHT);
> >  
> >     /*
> >      * Pull in IO completion errors now. We are guaranteed to be running
> 
> I think the XBF_IN_FLIGHT can be moved to the final xfs_buf_rele()
> processing if:
> 
> > @@ -1341,6 +1344,11 @@ xfs_buf_submit(
> >      * xfs_buf_ioend too early.
> >      */
> >     atomic_set(&bp->b_io_remaining, 1);
> > +   if (bp->b_flags & XBF_ASYNC) {
> > +           percpu_counter_inc(&bp->b_target->bt_io_count);
> > +           bp->b_flags |= XBF_IN_FLIGHT;
> > +   }
> 
> You change this to:
> 
>       if (!(bp->b_flags & XBF_IN_FLIGHT)) {
>               percpu_counter_inc(&bp->b_target->bt_io_count);
>               bp->b_flags |= XBF_IN_FLIGHT;
>       }
> 

Ok, so use the flag to cap the I/O count and defer the decrement to
release. I think that should work and addresses the raciness issue. I'll
give it a try.

> We shouldn't have to check for XBF_ASYNC in xfs_buf_submit() - it is
> the path taken for async IO submission, so we should probably
> ASSERT(bp->b_flags & XBF_ASYNC) in this function to ensure that is
> the case.
> 

Yeah, that's unnecessary. There's already such an assert in
xfs_buf_submit(), actually.

> [Thinking aloud - __test_and_set_bit() might make this code a bit
> cleaner]
> 

On a quick try, this complains about b_flags being an unsigned int. I
think I'll leave the set bit as is and use a helper for the release,
which also provides a location to explain how the count works.

> > diff --git a/fs/xfs/xfs_buf.h b/fs/xfs/xfs_buf.h
> > index 8bfb974..e1f95e0 100644
> > --- a/fs/xfs/xfs_buf.h
> > +++ b/fs/xfs/xfs_buf.h
> > @@ -43,6 +43,7 @@ typedef enum {
> >  #define XBF_READ    (1 << 0) /* buffer intended for reading from device */
> >  #define XBF_WRITE   (1 << 1) /* buffer intended for writing to device */
> >  #define XBF_READ_AHEAD      (1 << 2) /* asynchronous read-ahead */
> > +#define XBF_IN_FLIGHT       (1 << 3)
> 
> Hmmm - it's an internal flag, so probably should be prefixed with an
> "_" and moved down to the section with _XBF_KMEM and friends.
> 

Indeed, thanks.

Brian

> Thoughts?
> 
> Cheers,
> 
> Dave.
> -- 
> Dave Chinner
> david@xxxxxxxxxxxxx
> 
> _______________________________________________
> xfs mailing list
> xfs@xxxxxxxxxxx
> http://oss.sgi.com/mailman/listinfo/xfs

<Prev in Thread] Current Thread [Next in Thread>