[2/2] xfs: don't preempt writeback sequence on single page wb error
diff mbox series

Message ID 20181031140155.17996-3-bfoster@redhat.com
State New
Headers show
Series
  • xfs: don't preempt writeback on page errors
Related show

Commit Message

Brian Foster Oct. 31, 2018, 2:01 p.m. UTC
xfs_do_writepage() currently returns errors directly regardless of
whether it is called via ->writepages() or ->writepage(). In the
case of ->writepages(), an xfs_do_writepage() error return breaks
the current writeback sequence in write_cache_pages(). This means
that an integrity writeback (i.e., sync), for example, returns
before all associated pages have been processed.

This can be problematic in cases like unmount. If the writeback
doesn't process all delalloc pages before unmounting, we end up
reclaiming inodes with non-zero delalloc block counts. In turn, this
breaks block accounting and leaves the fs inconsistent.

XFS explicitly discards delalloc blocks on such writepage failures
to avoid this problem. This isn't terribly useful if we allow an
integrity writeback to complete (and thus a filesystem to unmount)
without addressing the entire set of dirty pages on an inode.
Therefore, change ->writepage[s]() to track high level error state
in the xfs_writepage_ctx structure and return it from the higher
level operation callout rather than xfs_do_writepage(). This ensures
that write_cache_pages() does not exit prematurely when called via
->writepages(), but both ->writepage() and ->writepages() still
ultimately return an error for the higher level operation.

This patch introduces a subtle change in the behavior of background
writeback in the event of persistent errors. The current behavior of
returning an error preempts the background writeback. Writeback
eventually comes around again and repeats the process for a few more
pages (in practice) before it once again fails. This repeats over
and over until the entire set of dirty pages is cleaned. This
behavior results in a somewhat slower stream of "page discard"
errors in the system log and dictates that many repeated fsync calls
may be required before the entire data set is processed and mapping
error consumed. With this change in place, background writeback
executes on as many pages as necessary as if each page writeback
were successful. The pages are cleaned immediately and significantly
more page discard errors can be observed at once.

Signed-off-by: Brian Foster <bfoster@redhat.com>
---
 fs/xfs/xfs_aops.c | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)

Comments

Dave Chinner Oct. 31, 2018, 11:02 p.m. UTC | #1
On Wed, Oct 31, 2018 at 10:01:55AM -0400, Brian Foster wrote:
> xfs_do_writepage() currently returns errors directly regardless of
> whether it is called via ->writepages() or ->writepage(). In the
> case of ->writepages(), an xfs_do_writepage() error return breaks
> the current writeback sequence in write_cache_pages(). This means
> that an integrity writeback (i.e., sync), for example, returns
> before all associated pages have been processed.

That sounds like a bug in write_cache_pages(). It sends pages
one at a time to the writepage context, and it is supposed to
iterate the entire range on a data sync operation. If you look at
the code, is clearly stopping at the first error.

IOWs, every filesystem that uses write_cache_pages() for data
integrity writes is broken in the same way.

And a quick look as fs specific writepages implementations indicates
that ext4_writepages() and ibtrfs's extent_write_cache_pages() have
the same issue.

> This can be problematic in cases like unmount. If the writeback
> doesn't process all delalloc pages before unmounting, we end up
> reclaiming inodes with non-zero delalloc block counts. In turn, this
> breaks block accounting and leaves the fs inconsistent.

XFS is probably the only filesystem that leaves detectable state
around and then detects and reports it....

> XFS explicitly discards delalloc blocks on such writepage failures
> to avoid this problem. This isn't terribly useful if we allow an
> integrity writeback to complete (and thus a filesystem to unmount)
> without addressing the entire set of dirty pages on an inode.
> Therefore, change ->writepage[s]() to track high level error state
> in the xfs_writepage_ctx structure and return it from the higher
> level operation callout rather than xfs_do_writepage(). This ensures
> that write_cache_pages() does not exit prematurely when called via
> ->writepages(), but both ->writepage() and ->writepages() still
> ultimately return an error for the higher level operation.
> 
> This patch introduces a subtle change in the behavior of background
> writeback in the event of persistent errors. The current behavior of
> returning an error preempts the background writeback. Writeback
> eventually comes around again and repeats the process for a few more
> pages (in practice) before it once again fails. This repeats over
> and over until the entire set of dirty pages is cleaned. This
> behavior results in a somewhat slower stream of "page discard"
> errors in the system log and dictates that many repeated fsync calls
> may be required before the entire data set is processed and mapping
> error consumed. With this change in place, background writeback
> executes on as many pages as necessary as if each page writeback
> were successful. The pages are cleaned immediately and significantly
> more page discard errors can be observed at once.

Yeah, this is a good change in behaviour, but I think the
implementation is wrong. write_cache_pages() needs to continue
iterating the range if WB_SYNC_ALL is set even when errors occur.

i.e. the error state should be maintained by write_cache_pages and
returned on completion, not require the filesystem to hide errors
from wcp in it's own specific writepage structure...

Cheers,

Dave.
Brian Foster Nov. 1, 2018, 2:17 p.m. UTC | #2
On Thu, Nov 01, 2018 at 10:02:25AM +1100, Dave Chinner wrote:
> On Wed, Oct 31, 2018 at 10:01:55AM -0400, Brian Foster wrote:
> > xfs_do_writepage() currently returns errors directly regardless of
> > whether it is called via ->writepages() or ->writepage(). In the
> > case of ->writepages(), an xfs_do_writepage() error return breaks
> > the current writeback sequence in write_cache_pages(). This means
> > that an integrity writeback (i.e., sync), for example, returns
> > before all associated pages have been processed.
> 
> That sounds like a bug in write_cache_pages(). It sends pages
> one at a time to the writepage context, and it is supposed to
> iterate the entire range on a data sync operation. If you look at
> the code, is clearly stopping at the first error.
> 
> IOWs, every filesystem that uses write_cache_pages() for data
> integrity writes is broken in the same way.
> 
> And a quick look as fs specific writepages implementations indicates
> that ext4_writepages() and ibtrfs's extent_write_cache_pages() have
> the same issue.
> 

I veered away from this approach initially because I didn't want to risk
breaking behavior of other filesystems. I can revisit it if more
appropriate and is broadly safe.

> > This can be problematic in cases like unmount. If the writeback
> > doesn't process all delalloc pages before unmounting, we end up
> > reclaiming inodes with non-zero delalloc block counts. In turn, this
> > breaks block accounting and leaves the fs inconsistent.
> 
> XFS is probably the only filesystem that leaves detectable state
> around and then detects and reports it....
> 
> > XFS explicitly discards delalloc blocks on such writepage failures
> > to avoid this problem. This isn't terribly useful if we allow an
> > integrity writeback to complete (and thus a filesystem to unmount)
> > without addressing the entire set of dirty pages on an inode.
> > Therefore, change ->writepage[s]() to track high level error state
> > in the xfs_writepage_ctx structure and return it from the higher
> > level operation callout rather than xfs_do_writepage(). This ensures
> > that write_cache_pages() does not exit prematurely when called via
> > ->writepages(), but both ->writepage() and ->writepages() still
> > ultimately return an error for the higher level operation.
> > 
> > This patch introduces a subtle change in the behavior of background
> > writeback in the event of persistent errors. The current behavior of
> > returning an error preempts the background writeback. Writeback
> > eventually comes around again and repeats the process for a few more
> > pages (in practice) before it once again fails. This repeats over
> > and over until the entire set of dirty pages is cleaned. This
> > behavior results in a somewhat slower stream of "page discard"
> > errors in the system log and dictates that many repeated fsync calls
> > may be required before the entire data set is processed and mapping
> > error consumed. With this change in place, background writeback
> > executes on as many pages as necessary as if each page writeback
> > were successful. The pages are cleaned immediately and significantly
> > more page discard errors can be observed at once.
> 
> Yeah, this is a good change in behaviour, but I think the
> implementation is wrong. write_cache_pages() needs to continue
> iterating the range if WB_SYNC_ALL is set even when errors occur.
> 

The other thing that comes to mind is that any time we have this kind of
iterative/callback interface, it's usually appropriate to have some kind
of means for the callback to signal the caller to stop processing. You
could argue that right now write_cache_pages() works as expected,
returning an error is that stop mechanism and XFS has the bug as it is
the only filesystem that depends on completely processing integrity wbs
for fs correctness (or we're all just using it wrong ;P, I haven't
investigated enough yet to say one way or the other).

OTOH, we could maintain the ability to interrupt the scan with a magic
error return or some such, that just may require a bit more care and
attention across the set of existing callers. Thoughts?

Brian

> i.e. the error state should be maintained by write_cache_pages and
> returned on completion, not require the filesystem to hide errors
> from wcp in it's own specific writepage structure...
> 
> Cheers,
> 
> Dave.
> -- 
> Dave Chinner
> david@fromorbit.com
Dave Chinner Nov. 1, 2018, 9:37 p.m. UTC | #3
On Thu, Nov 01, 2018 at 10:17:26AM -0400, Brian Foster wrote:
> On Thu, Nov 01, 2018 at 10:02:25AM +1100, Dave Chinner wrote:
> > On Wed, Oct 31, 2018 at 10:01:55AM -0400, Brian Foster wrote:
> > > xfs_do_writepage() currently returns errors directly regardless of
> > > whether it is called via ->writepages() or ->writepage(). In the
> > > case of ->writepages(), an xfs_do_writepage() error return breaks
> > > the current writeback sequence in write_cache_pages(). This means
> > > that an integrity writeback (i.e., sync), for example, returns
> > > before all associated pages have been processed.
> > 
> > That sounds like a bug in write_cache_pages(). It sends pages
> > one at a time to the writepage context, and it is supposed to
> > iterate the entire range on a data sync operation. If you look at
> > the code, is clearly stopping at the first error.
> > 
> > IOWs, every filesystem that uses write_cache_pages() for data
> > integrity writes is broken in the same way.
> > 
> > And a quick look as fs specific writepages implementations indicates
> > that ext4_writepages() and ibtrfs's extent_write_cache_pages() have
> > the same issue.
> > 
> 
> I veered away from this approach initially because I didn't want to risk
> breaking behavior of other filesystems. I can revisit it if more
> appropriate and is broadly safe.

We should fix problems in the generic code if that's where the
problem lies. You can ignore filesystems with their own custom code,
though, because you are more likely to break them that generic
code...

> > > This can be problematic in cases like unmount. If the writeback
> > > doesn't process all delalloc pages before unmounting, we end up
> > > reclaiming inodes with non-zero delalloc block counts. In turn, this
> > > breaks block accounting and leaves the fs inconsistent.
> > 
> > XFS is probably the only filesystem that leaves detectable state
> > around and then detects and reports it....
> > 
> > > XFS explicitly discards delalloc blocks on such writepage failures
> > > to avoid this problem. This isn't terribly useful if we allow an
> > > integrity writeback to complete (and thus a filesystem to unmount)
> > > without addressing the entire set of dirty pages on an inode.
> > > Therefore, change ->writepage[s]() to track high level error state
> > > in the xfs_writepage_ctx structure and return it from the higher
> > > level operation callout rather than xfs_do_writepage(). This ensures
> > > that write_cache_pages() does not exit prematurely when called via
> > > ->writepages(), but both ->writepage() and ->writepages() still
> > > ultimately return an error for the higher level operation.
> > > 
> > > This patch introduces a subtle change in the behavior of background
> > > writeback in the event of persistent errors. The current behavior of
> > > returning an error preempts the background writeback. Writeback
> > > eventually comes around again and repeats the process for a few more
> > > pages (in practice) before it once again fails. This repeats over
> > > and over until the entire set of dirty pages is cleaned. This
> > > behavior results in a somewhat slower stream of "page discard"
> > > errors in the system log and dictates that many repeated fsync calls
> > > may be required before the entire data set is processed and mapping
> > > error consumed. With this change in place, background writeback
> > > executes on as many pages as necessary as if each page writeback
> > > were successful. The pages are cleaned immediately and significantly
> > > more page discard errors can be observed at once.
> > 
> > Yeah, this is a good change in behaviour, but I think the
> > implementation is wrong. write_cache_pages() needs to continue
> > iterating the range if WB_SYNC_ALL is set even when errors occur.
> > 
> 
> The other thing that comes to mind is that any time we have this kind of
> iterative/callback interface, it's usually appropriate to have some kind
> of means for the callback to signal the caller to stop processing. You
> could argue that right now write_cache_pages() works as expected,
> returning an error is that stop mechanism and XFS has the bug as it is
> the only filesystem that depends on completely processing integrity wbs
> for fs correctness (or we're all just using it wrong ;P, I haven't
> investigated enough yet to say one way or the other).

The comment on wcp error handling is enlightening:

		ret = (*writepage)(page, wbc, data);
		if (unlikely(ret)) {
			if (ret == AOP_WRITEPAGE_ACTIVATE) {
				unlock_page(page);
				ret = 0;
			} else {
				/*
				 * done_index is set past this page,
				 * so media errors will not choke
				 * background writeout for the entire
				 * file. This has consequences for
				 * range_cyclic semantics (ie. it may
				 * not be suitable for data integrity
				 * writeout).
				 */
				done_index = page->index + 1;
				done = 1;
				break;
			}
		}

"ie it may not be suitable for data integrity writeout"

In a nutshell. The current error handling behaviour is not suitable
for data integrity writeout, as your proposed patch has demonstrated.


> OTOH, we could maintain the ability to interrupt the scan with a magic
> error return or some such, that just may require a bit more care and
> attention across the set of existing callers. Thoughts?

I'm not sure why we'd want "stop on error" for data integrity
writeback given the problems you've pointed out.  For WB_SYNC_NONE
writeback, the current behaviour (stop on error) is fine, so i don't
see a big need to make unnecessary API changes here.

Cheers,

Dave.
Brian Foster Nov. 2, 2018, 11:42 a.m. UTC | #4
On Fri, Nov 02, 2018 at 08:37:40AM +1100, Dave Chinner wrote:
> On Thu, Nov 01, 2018 at 10:17:26AM -0400, Brian Foster wrote:
> > On Thu, Nov 01, 2018 at 10:02:25AM +1100, Dave Chinner wrote:
> > > On Wed, Oct 31, 2018 at 10:01:55AM -0400, Brian Foster wrote:
> > > > xfs_do_writepage() currently returns errors directly regardless of
> > > > whether it is called via ->writepages() or ->writepage(). In the
> > > > case of ->writepages(), an xfs_do_writepage() error return breaks
> > > > the current writeback sequence in write_cache_pages(). This means
> > > > that an integrity writeback (i.e., sync), for example, returns
> > > > before all associated pages have been processed.
> > > 
> > > That sounds like a bug in write_cache_pages(). It sends pages
> > > one at a time to the writepage context, and it is supposed to
> > > iterate the entire range on a data sync operation. If you look at
> > > the code, is clearly stopping at the first error.
> > > 
> > > IOWs, every filesystem that uses write_cache_pages() for data
> > > integrity writes is broken in the same way.
> > > 
> > > And a quick look as fs specific writepages implementations indicates
> > > that ext4_writepages() and ibtrfs's extent_write_cache_pages() have
> > > the same issue.
> > > 
> > 
> > I veered away from this approach initially because I didn't want to risk
> > breaking behavior of other filesystems. I can revisit it if more
> > appropriate and is broadly safe.
> 
> We should fix problems in the generic code if that's where the
> problem lies. You can ignore filesystems with their own custom code,
> though, because you are more likely to break them that generic
> code...
> 
> > > > This can be problematic in cases like unmount. If the writeback
> > > > doesn't process all delalloc pages before unmounting, we end up
> > > > reclaiming inodes with non-zero delalloc block counts. In turn, this
> > > > breaks block accounting and leaves the fs inconsistent.
> > > 
> > > XFS is probably the only filesystem that leaves detectable state
> > > around and then detects and reports it....
> > > 
> > > > XFS explicitly discards delalloc blocks on such writepage failures
> > > > to avoid this problem. This isn't terribly useful if we allow an
> > > > integrity writeback to complete (and thus a filesystem to unmount)
> > > > without addressing the entire set of dirty pages on an inode.
> > > > Therefore, change ->writepage[s]() to track high level error state
> > > > in the xfs_writepage_ctx structure and return it from the higher
> > > > level operation callout rather than xfs_do_writepage(). This ensures
> > > > that write_cache_pages() does not exit prematurely when called via
> > > > ->writepages(), but both ->writepage() and ->writepages() still
> > > > ultimately return an error for the higher level operation.
> > > > 
> > > > This patch introduces a subtle change in the behavior of background
> > > > writeback in the event of persistent errors. The current behavior of
> > > > returning an error preempts the background writeback. Writeback
> > > > eventually comes around again and repeats the process for a few more
> > > > pages (in practice) before it once again fails. This repeats over
> > > > and over until the entire set of dirty pages is cleaned. This
> > > > behavior results in a somewhat slower stream of "page discard"
> > > > errors in the system log and dictates that many repeated fsync calls
> > > > may be required before the entire data set is processed and mapping
> > > > error consumed. With this change in place, background writeback
> > > > executes on as many pages as necessary as if each page writeback
> > > > were successful. The pages are cleaned immediately and significantly
> > > > more page discard errors can be observed at once.
> > > 
> > > Yeah, this is a good change in behaviour, but I think the
> > > implementation is wrong. write_cache_pages() needs to continue
> > > iterating the range if WB_SYNC_ALL is set even when errors occur.
> > > 
> > 
> > The other thing that comes to mind is that any time we have this kind of
> > iterative/callback interface, it's usually appropriate to have some kind
> > of means for the callback to signal the caller to stop processing. You
> > could argue that right now write_cache_pages() works as expected,
> > returning an error is that stop mechanism and XFS has the bug as it is
> > the only filesystem that depends on completely processing integrity wbs
> > for fs correctness (or we're all just using it wrong ;P, I haven't
> > investigated enough yet to say one way or the other).
> 
> The comment on wcp error handling is enlightening:
> 
> 		ret = (*writepage)(page, wbc, data);
> 		if (unlikely(ret)) {
> 			if (ret == AOP_WRITEPAGE_ACTIVATE) {
> 				unlock_page(page);
> 				ret = 0;
> 			} else {
> 				/*
> 				 * done_index is set past this page,
> 				 * so media errors will not choke
> 				 * background writeout for the entire
> 				 * file. This has consequences for
> 				 * range_cyclic semantics (ie. it may
> 				 * not be suitable for data integrity
> 				 * writeout).
> 				 */
> 				done_index = page->index + 1;
> 				done = 1;
> 				break;
> 			}
> 		}
> 
> "ie it may not be suitable for data integrity writeout"
> 

The way I read that comment is that the done_index bump may not be
suitable for integrity writeback (re: the tradeoff being to not "choke"
bg writeback), but the wording is not totally clear.

> In a nutshell. The current error handling behaviour is not suitable
> for data integrity writeout, as your proposed patch has demonstrated.
> 
> 
> > OTOH, we could maintain the ability to interrupt the scan with a magic
> > error return or some such, that just may require a bit more care and
> > attention across the set of existing callers. Thoughts?
> 
> I'm not sure why we'd want "stop on error" for data integrity
> writeback given the problems you've pointed out.  For WB_SYNC_NONE
> writeback, the current behaviour (stop on error) is fine, so i don't
> see a big need to make unnecessary API changes here.
> 

Neither am I, for XFS at least, I just wasn't sure if other filesystems
depended on or expected this behavior in any way. I'll take a more
thorough look, run some broader testing and send a write_cache_pages()
patch to a wider audience.

BTW, it looks like we already have this AOP_WRITEPAGE_
positive_aop_returns enum for special return codes the callback can use
to control the caller. I suppose we could add a new one if some fs does
have a particular need to forcibly interrupt a writeback.

Brian

> Cheers,
> 
> Dave.
> -- 
> Dave Chinner
> david@fromorbit.com

Patch
diff mbox series

diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c
index 3feae3691467..438cfc66a40e 100644
--- a/fs/xfs/xfs_aops.c
+++ b/fs/xfs/xfs_aops.c
@@ -32,6 +32,7 @@  struct xfs_writepage_ctx {
 	unsigned int		io_type;
 	unsigned int		cow_seq;
 	struct xfs_ioend	*ioend;
+	int			error;
 };
 
 struct block_device *
@@ -798,7 +799,9 @@  xfs_writepage_map(
 		end_page_writeback(page);
 done:
 	mapping_set_error(page->mapping, error);
-	return error;
+	if (!wpc->error)
+		wpc->error = error;
+	return 0;
 }
 
 /*
@@ -929,8 +932,8 @@  xfs_vm_writepage(
 
 	ret = xfs_do_writepage(page, wbc, &wpc);
 	if (wpc.ioend)
-		ret = xfs_submit_ioend(wbc, wpc.ioend, ret);
-	return ret;
+		ret = xfs_submit_ioend(wbc, wpc.ioend, wpc.error);
+	return ret ? ret : wpc.error;
 }
 
 STATIC int
@@ -946,8 +949,8 @@  xfs_vm_writepages(
 	xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
 	ret = write_cache_pages(mapping, wbc, xfs_do_writepage, &wpc);
 	if (wpc.ioend)
-		ret = xfs_submit_ioend(wbc, wpc.ioend, ret);
-	return ret;
+		ret = xfs_submit_ioend(wbc, wpc.ioend, wpc.error);
+	return ret ? ret : wpc.error;
 }
 
 STATIC int