Re: [PATCH] xfs: clear XBF_ASYNC if we attach an iodone callback

[Carlos Maiolino <cmaiolino@redhat.com>]

Hi,

I really apologize for my late reply, March was a busy month and this thread got
lost on my mailbox.

I'll try to come back to this fix as soon as possible,

> > > > 
> > > >                 return;
> > > >         }
> > > It's likely you'll end up with silent on disk corruption if you do
> > > that because it's just tossing the error state away.
> > 
> > I think I'm getting a handle on this, but I wonder if we don't have
> > this problem currently?  We get an error on a buffer, we retry it and
> > it fails again.  If we don't have XFS_ERR_RETRY_FOREVER set then we
> > error out and do the normal callbacks as if nothing ever happened, and
> > we get the silent corruption without shutting down the file system.
> >  How is the code as it stands without XFS_ERR_RETRY_FOREVER safe?
> >  Thanks,
> > 
> 
> We shouldn't get into a situation where we run the callbacks when the
> I/O has failed and the fs isn't shut down. IOWs, we do run the callbacks
> only when the I/O succeeds, or it fails and the fs is shutdown as a
> result (never when the I/O fails but we haven't shut down, which is the
> problematic case the patch above introduces).
> 
> That tunable error handling logic (XFS_ERR_RETRY_FOREVER) basically
> determines when an I/O error is considered permanent, which means
> retries are futile and we have no choice but to shutdown the fs. E.g.,
> if an error is deemed permanent, we jump to permanent_error:
> 
>         /*
>          * Permanent error - we need to trigger a shutdown if we haven't already
>          * to indicate that inconsistency will result from this action.
>          */
> permanent_error:
>         xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
> out_stale:
>         xfs_buf_stale(bp);
>         bp->b_flags |= XBF_DONE;
>         trace_xfs_buf_error_relse(bp, _RET_IP_);
>         return false;
> }
> 
> ... which shuts down the filesystem before we go any farther.
> 
> Brian
> 
> > Josef

Josef, did you have any success in avoiding this lockup problem with the
fail/retry mechanism mentioned? Or do you keep seeing it?


-- 
Carlos