Re: [LSF/MM TOPIC] do we really need PG_error at all?
From: Jeff Layton
Date: Mon Feb 27 2017 - 20:01:52 EST
On Mon, 2017-02-27 at 15:51 -0700, Andreas Dilger wrote:
> On Feb 27, 2017, at 8:07 AM, Jeff Layton <jlayton@xxxxxxxxxx> wrote:
> >
> > On Mon, 2017-02-27 at 11:27 +1100, NeilBrown wrote:
> > > On Sun, Feb 26 2017, James Bottomley wrote:
> > >
> > > > On Mon, 2017-02-27 at 08:03 +1100, NeilBrown wrote:
> > > > > On Sun, Feb 26 2017, James Bottomley wrote:
> > > > >
> > > > > > [added linux-scsi and linux-block because this is part of our error
> > > > > > handling as well]
> > > > > > On Sun, 2017-02-26 at 09:42 -0500, Jeff Layton wrote:
> > > > > > > Proposing this as a LSF/MM TOPIC, but it may turn out to be me
> > > > > > > just not understanding the semantics here.
> > > > > > >
> > > > > > > As I was looking into -ENOSPC handling in cephfs, I noticed that
> > > > > > > PG_error is only ever tested in one place [1]
> > > > > > > __filemap_fdatawait_range, which does this:
> > > > > > >
> > > > > > > if (TestClearPageError(page))
> > > > > > > ret = -EIO;
> > > > > > >
> > > > > > > This error code will override any AS_* error that was set in the
> > > > > > > mapping. Which makes me wonder...why don't we just set this error
> > > > > > > in the mapping and not bother with a per-page flag? Could we
> > > > > > > potentially free up a page flag by eliminating this?
> > > > > >
> > > > > > Note that currently the AS_* codes are only set for write errors
> > > > > > not for reads and we have no mapping error handling at all for swap
> > > > > > pages, but I'm sure this is fixable.
> > > > >
> > > > > How is a read error different from a failure to set PG_uptodate?
> > > > > Does PG_error suppress retries?
> > > >
> > > > We don't do any retries in the code above the block layer (or at least
> > > > we shouldn't).
> > >
> > > I was wondering about what would/should happen if a read request was
> > > re-issued for some reason. Should the error flag on the page cause an
> > > immediate failure, or should it try again.
> > > If read-ahead sees a read-error on some future page, is it necessary to
> > > record the error so subsequent read-aheads don't notice the page is
> > > missing and repeatedly try to re-load it?
> > > When the application eventually gets to the faulty page, should a read
> > > be tried then, or is the read-ahead failure permanent?
> > >
> > >
> > >
> > > >
> > > > > >
> > > > > > From the I/O layer point of view we take great pains to try to
> > > > > > pinpoint the error exactly to the sector. We reflect this up by
> > > > > > setting the PG_error flag on the page where the error occurred. If
> > > > > > we only set the error on the mapping, we lose that granularity,
> > > > > > because the mapping is mostly at the file level (or VMA level for
> > > > > > anon pages).
> > > > >
> > > > > Are you saying that the IO layer finds the page in the bi_io_vec and
> > > > > explicitly sets PG_error,
> > > >
> > > > I didn't say anything about the mechanism. I think the function you're
> > > > looking for is fs/mpage.c:mpage_end_io(). layers below block indicate
> > > > the position in the request. Block maps the position to bio and the
> > > > bio completion maps to page. So the actual granularity seen in the
> > > > upper layer depends on how the page to bio mapping is done.
> > >
> > > If the block layer is just returning the status at a per-bio level (which
> > > makes perfect sense), then this has nothing directly to do with the
> > > PG_error flag.
> > >
> > > The page cache needs to do something with bi_error, but it isn't
> > > immediately clear that it needs to set PG_error.
> > >
> > > > :q
> > > > > rather than just passing an error indication to bi_end_io ?? That
> > > > > would seem to be wrong as the page may not be in the page cache.
> > > >
> > > > Usually pages in the mpage_end_io path are pinned, I think.
> > > >
> > > > > So I guess I misunderstand you.
> > > > >
> > > > > >
> > > > > > So I think the question for filesystem people from us would be do
> > > > > > you care about this accuracy? If it's OK just to know an error
> > > > > > occurred somewhere in this file, then perhaps we don't need it.
> > > > >
> > > > > I had always assumed that a bio would either succeed or fail, and
> > > > > that no finer granularity could be available.
> > > >
> > > > It does ... but a bio can be as small as a single page.
> > > >
> > > > > I think the question here is: Do filesystems need the pagecache to
> > > > > record which pages have seen an IO error?
> > > >
> > > > It's not just filesystems. The partition code uses PageError() ... the
> > > > metadata code might as well (those are things with no mapping). I'm
> > > > not saying we can't remove PG_error; I am saying it's not going to be
> > > > quite as simple as using the AS_ flags.
> > >
> > > The partition code could use PageUptodate().
> > > mpage_end_io() calls page_endio() on each page, and on read error that
> > > calls:
> > >
> > > ClearPageUptodate(page);
> > > SetPageError(page);
> > >
> > > are both of these necessary?
> > >
> > > fs/buffer.c can use several bios to read a single page.
> > > If any one returns an error, PG_error is set. When all of them have
> > > completed, if PG_error is clear, PG_uptodate is then set.
> > > This is an opportunistic use of PG_error, rather than an essential use.
> > > It could be "fixed", and would need to be fixed if we were to deprecate
> > > use of PG_error for read errors.
> > > There are probably other usages like this.
> > >
> >
> > Ok, I think I get it (somewhat):
> >
> > The tricky part there is how to handle the PageError check in
> > read_dev_sector if you don't use SetPageError in the result handler.
> >
> > If we can count on read_pagecache_sector and read_dax_sector reliably
> > returning an error when the page is considered to be in the cache
> > (PageUpToDate) but had a read error, then that would work. I'm not sure
> > how you'd indicate that without something like PG_error though if you
> > don't want to retry on every attempt.
> >
> > OTOH, if we want to always retry to read in pages that have had read
> > errors when someone requests them, then we can simply not set
> > PageUpToDate when readahead fails.
> >
> > To chip away at the edges of this, what may make sense is to get this
> > flag out of the writeback code as much as we can. When a write fails and
> > SetPageError is called, we should also mark the mapping with an error.
> > Then, we should be able to stop overriding the mapping error with -EIO
> > in that codepath. Maybe call ClearPageError, or maybe leave it alone
> > there?
>
> My thought is that PG_error is definitely useful for applications to get
> correct errors back when doing write()/sync_file_range() so that they know
> there is an error in the data that _they_ wrote, rather than receiving an
> error for data that may have been written by another thread, and in turn
> clearing the error from another thread so it *doesn't* know it had a write
> error.
>
Right, that was my point about sync_file_range. Today I think you can
call sync_file_range and if your range didn't hit errors (none of you
PG_error bits are set), then you might get back 0 (iff the mapping had
no error flagged on it).
The question I have is: is that the semantics that sync_file_range is
supposed to have? It's not clear from the manpage whether errors are
supposed to be that granular or not.
It's also the case that not all writepage implementations set the
PG_error flag. There is a lot of variation here, and so we end up with
different semantics on different filesystems. That's less than ideal.
> As for stray sync() clearing PG_error from underneath an application, that
> shouldn't happen since filemap_fdatawait_keep_errors() doesn't clear errors
> and is used by device flushing code (fdatawait_one_bdev(), wait_sb_inodes()).
>
>
It sure looks like it does to me. Am I missing something?
filemap_fdatawait_keep_errors calls __filemap_fdatawait_range, which
finds pages and waits on their writeback bit to clear. Once it does
that, it calls TestClearPageError and sets the return to -EIO if the bit
was set. That should result in all of the PG_error bits being cleared on
a sync() call, right?
--
Jeff Layton <jlayton@xxxxxxxxxx>