lkml.org 
[lkml]   [2023]   [Jan]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 02/11] filemap: Remove filemap_check_and_keep_errors()
On Mon, Jan 09, 2023 at 09:31:00AM -0500, Jeff Layton wrote:
> On Mon, 2023-01-09 at 14:02 +0000, Matthew Wilcox wrote:
> > On Mon, Jan 09, 2023 at 08:48:49AM -0500, Jeff Layton wrote:
> > > On Mon, 2023-01-09 at 05:18 +0000, Matthew Wilcox (Oracle) wrote:
> > > > Convert both callers to use the "new" errseq infrastructure.
> > >
> > > I looked at making this sort of change across the board alongside the
> > > original wb_err patches, but I backed off at the time.
> > >
> > > With the above patch, this function will no longer report a writeback
> > > error that occurs before the sample. Given that writeback can happen at
> > > any time, that seemed like it might be an undesirable change, and I
> > > didn't follow through.
> > >
> > > It is true that the existing flag-based code may miss errors too, if
> > > multiple tasks are test_and_clear'ing the bits, but I think the above is
> > > even more likely to happen, esp. under memory pressure.
> > >
> > > To do this right, we probably need to look at these callers and have
> > > them track a long-term errseq_t "since" value before they ever dirty the
> > > pages, and then continually check-and-advance vs. that.
> > >
> > > For instance, the main caller of the above function is jbd2. Would it be
> > > reasonable to add in a new errseq_t value to the jnode for tracking
> > > errors?
> >
> > Doesn't b4678df184b3 address this problem? If nobody has seen the
> > error, we return 0 instead of the current value of wb_err, ensuring
> > that somebody always sees the error.
> >
>
> I was originally thinking no, but now I think you're correct.
>
> We do initialize the "since" value to 0 if an error has never been seen,
> so that (sort of) emulates the behavior of the existing AS_EIO/AS_ENOSPC
> flags.
>
> It's still not quite as reliable as plumbing a "since" value through all
> of the callers (particularly in the case where there are multiple
> waiters), but maybe it's good enough here.

I actually think we may have the opposite problem; that for some of
these scenarios, we never mark the error as seen. ie we always end
up calling errseq_check() and never errseq_check_and_advance(). So
every time we write something, it'll remind us that we have an error.

\
 
 \ /
  Last update: 2023-03-26 23:33    [W:0.086 / U:0.256 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site