Skip to contentSkip to conversation

write-back: fix nr_to_write counter

PATCH In review Artem Bityutskiy 2 participants 6 messages Feb 3, 2009

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 write-back: fix nr_to_write counter

Hi,

commit 05fe478dd04e02fa230c305ab9b5616669821dd3

Author: Nick Piggin <[email protected]>
Date:   Tue Jan 6 14:39:08 2009 -0800

mm: write_cache_pages integrity fix

broke wbc->nr_to_write handling. Here is the fix.

I'm not 100% sure I got things right, because I am far not expert in the area. Please, review it. The patch fixes my UBIFS issues, which are caused by the fact that wbc->nr_to_write is not updated.
======================================================================

From: Artem Bityutskiy <[email protected]>
Date: Mon, 2 Feb 2009 18:33:49 +0200
Subject: [PATCH] write-back: fix nr_to_write counter

Commit 05fe478dd04e02fa230c305ab9b5616669821dd3 broke @wbc->nr_to_write. 'write_cache_pages()' changes it in the loop, but restores the original value from @nr_to_write at the end, because of this code:

       if (!wbc->no_nrwrite_index_update) {
                if (wbc->range_cyclic || (range_whole && nr_to_write > 0))
                        mapping->writeback_index = done_index;
                wbc->nr_to_write = nr_to_write;
        }

Well, in case of @wbc->no_nrwrite_index_update != 0, we do change wbc->nr_to_write, while we should not. This patch fixes this behavior.

Also, I think wbc->nr_to_write should be changed in all cases, not only when wbc->sync_mode == WB_SYNC_NONE.

Also, I add a comment explaining why we do not stop writing back.

---
 mm/page-writeback.c |   21 +++++++++++++++------
 1 files changed, 15 insertions(+), 6 deletions(-)
mm/page-writeback.c +15 -6
index b493db7..13a2b8e 100644
@@ -1051,13 +1051,22 @@ continue_unlock:
 				}
  			}
 
-			if (wbc->sync_mode == WB_SYNC_NONE) {
-				wbc->nr_to_write--;
-				if (wbc->nr_to_write <= 0) {
-					done = 1;
-					break;
-				}
+			if (nr_to_write > 0)
+				nr_to_write--;
+			else if (wbc->sync_mode == WB_SYNC_NONE) {
+				/*
+				 * We stop writing back only if we are not
+				 * doing integrity sync. In case of integrity
+				 * sync we have to keep going because someone
+				 * may be concurrently dirtying pages, and we
+				 * might have synced a lot of newly appeared
+				 * dirty pages, bud have not synced all of the
+				 * old dirty pages.
+				 */
+				done = 1;
+				break;
 			}
+
 			if (wbc->nonblocking && bdi_write_congested(bdi)) {
 				wbc->encountered_congestion = 1;
 				done = 1;
Signature · 5 lines
1.6.0.6

-- 
Best regards,
Artem Bityutskiy (Битюцкий Артём)
#1 lore raw
Quoted 23 lines · On Tuesday 03 February 2009 19:42:22 Artem Bityutskiy wrote:

Hi,

commit 05fe478dd04e02fa230c305ab9b5616669821dd3

Author: Nick Piggin <[email protected]>
Date:   Tue Jan 6 14:39:08 2009 -0800

mm: write_cache_pages integrity fix

broke wbc->nr_to_write handling. Here is the fix.

I'm not 100% sure I got things right, because I am far not expert in the area. Please, review it. The patch fixes my UBIFS issues, which are caused by the fact that wbc->nr_to_write is not updated.
======================================================================

From: Artem Bityutskiy <[email protected]>
Date: Mon, 2 Feb 2009 18:33:49 +0200
Subject: [PATCH] write-back: fix nr_to_write counter

Commit 05fe478dd04e02fa230c305ab9b5616669821dd3 broke @wbc->nr_to_write. 'write_cache_pages()' changes it in the loop, but restores the original value from @nr_to_write at the end, because of this code:

Quoted 5 lines
       if (!wbc->no_nrwrite_index_update) {
                if (wbc->range_cyclic || (range_whole && nr_to_write > 0))
                        mapping->writeback_index = done_index;
                wbc->nr_to_write = nr_to_write;
        }

The commit you quote only moves nr_to_write to not take effect for WB_SYNC_ALL (ie. data integrity) writeout. And makes no other change to write_cache_pages.

I thought your problem might have been that you were calling this with WB_SYNC_ALL and expecting it to heed nr_to_write, however...

Quoted 2 lines

Also, I think wbc->nr_to_write should be changed in all cases, not only when wbc->sync_mode == WB_SYNC_NONE.

... you mention this here like it is an *additional* issue on top of your problem. So I fail to see how my commit could have caused this problem?

Quoted 2 lines

Well, in case of @wbc->no_nrwrite_index_update != 0, we do change wbc->nr_to_write, while we should not. This patch fixes this behavior.

And I don't know what you mean by this because the patch doesn't fix any problem there AFAIKS.

Anyway, I did probably not pay enough attention to ubifs when making this change, and if it wants wbc->nr_to_write updated even for data integrity syncs, I don't see the harm in that. So I don't have any objection to your patch. Thanks.

Can you cc [email protected] when a final version gets merged upstream please?

Quoted 43 lines

Also, I add a comment explaining why we do not stop writing back.

---
 mm/page-writeback.c |   21 +++++++++++++++------
 1 files changed, 15 insertions(+), 6 deletions(-)
mm/page-writeback.c +15 -7
index b493db7..13a2b8e 100644
@@ -1051,13 +1051,22 @@ continue_unlock:
 				}
  			}
 
-			if (wbc->sync_mode == WB_SYNC_NONE) {
-				wbc->nr_to_write--;
-				if (wbc->nr_to_write <= 0) {
-					done = 1;
-					break;
-				}
+			if (nr_to_write > 0)
+				nr_to_write--;
+			else if (wbc->sync_mode == WB_SYNC_NONE) {
+				/*
+				 * We stop writing back only if we are not
+				 * doing integrity sync. In case of integrity
+				 * sync we have to keep going because someone
+				 * may be concurrently dirtying pages, and we
+				 * might have synced a lot of newly appeared
+				 * dirty pages, bud have not synced all of the
+				 * old dirty pages.
+				 */
+				done = 1;
+				break;
 			}
+
 			if (wbc->nonblocking && bdi_write_congested(bdi)) {
 				wbc->encountered_congestion = 1;
 				done = 1;
--

1.6.0.6

#2 lore raw
Quoted 50 lines · On Tuesday 03 February 2009 20:13:56 Nick Piggin wrote:
Quoted 23 lines · On Tuesday 03 February 2009 19:42:22 Artem Bityutskiy wrote:

Hi,

commit 05fe478dd04e02fa230c305ab9b5616669821dd3

Author: Nick Piggin <[email protected]>
Date:   Tue Jan 6 14:39:08 2009 -0800

mm: write_cache_pages integrity fix

broke wbc->nr_to_write handling. Here is the fix.

I'm not 100% sure I got things right, because I am far not expert in the area. Please, review it. The patch fixes my UBIFS issues, which are caused by the fact that wbc->nr_to_write is not updated.
======================================================================

From: Artem Bityutskiy <[email protected]>
Date: Mon, 2 Feb 2009 18:33:49 +0200
Subject: [PATCH] write-back: fix nr_to_write counter

Commit 05fe478dd04e02fa230c305ab9b5616669821dd3 broke @wbc->nr_to_write. 'write_cache_pages()' changes it in the loop, but restores the original value from @nr_to_write at the end, because of this code:

Quoted 5 lines
       if (!wbc->no_nrwrite_index_update) {
                if (wbc->range_cyclic || (range_whole && nr_to_write >
0)) mapping->writeback_index = done_index;
                wbc->nr_to_write = nr_to_write;
        }

The commit you quote only moves nr_to_write to not take effect for WB_SYNC_ALL (ie. data integrity) writeout. And makes no other change to write_cache_pages.

I thought your problem might have been that you were calling this with WB_SYNC_ALL and expecting it to heed nr_to_write, however...

Quoted 2 lines

Also, I think wbc->nr_to_write should be changed in all cases, not only when wbc->sync_mode == WB_SYNC_NONE.

... you mention this here like it is an *additional* issue on top of your problem. So I fail to see how my commit could have caused this problem?

Quoted 2 lines

Well, in case of @wbc->no_nrwrite_index_update != 0, we do change wbc->nr_to_write, while we should not. This patch fixes this behavior.

And I don't know what you mean by this because the patch doesn't fix any problem there AFAIKS.

Quoted 7 lines

Anyway, I did probably not pay enough attention to ubifs when making this change, and if it wants wbc->nr_to_write updated even for data integrity syncs, I don't see the harm in that. So I don't have any objection to your patch. Thanks.

Can you cc [email protected] when a final version gets merged upstream please?

What I mean to say is that while I think the patch is not a problem, and actually should be merged if it is required for ubifs, I would like to just see the changelog clarified...

Also if you add a nice comment there, maybe mention that nr_to_write must reflect the number of pages written even for the WB_SYNC_ALL case (arguably the data integrity issue which you comment on is more obvious in comparison to the nr_to_write requirement of callers).

#3 lore raw
In reply to Nick Piggin
Quoted 30 lines · Nick Piggin
Quoted 23 lines · On Tuesday 03 February 2009 19:42:22 Artem Bityutskiy wrote:

Hi,

commit 05fe478dd04e02fa230c305ab9b5616669821dd3

Author: Nick Piggin <[email protected]>
Date:   Tue Jan 6 14:39:08 2009 -0800

mm: write_cache_pages integrity fix

broke wbc->nr_to_write handling. Here is the fix.

I'm not 100% sure I got things right, because I am far not expert in the area. Please, review it. The patch fixes my UBIFS issues, which are caused by the fact that wbc->nr_to_write is not updated.
======================================================================

From: Artem Bityutskiy <[email protected]>
Date: Mon, 2 Feb 2009 18:33:49 +0200
Subject: [PATCH] write-back: fix nr_to_write counter

Commit 05fe478dd04e02fa230c305ab9b5616669821dd3 broke @wbc->nr_to_write. 'write_cache_pages()' changes it in the loop, but restores the original value from @nr_to_write at the end, because of this code:

Quoted 5 lines
       if (!wbc->no_nrwrite_index_update) {
                if (wbc->range_cyclic || (range_whole && nr_to_write > 0))
                        mapping->writeback_index = done_index;
                wbc->nr_to_write = nr_to_write;
        }
Quoted 3 lines

The commit you quote only moves nr_to_write to not take effect for WB_SYNC_ALL (ie. data integrity) writeout. And makes no other change to write_cache_pages.

Nick, I'm sorry if my e-mail looked like I'm blame you, I referred the commit because git-bisect pointed to it and I though it for me :-)

And I apologize if I write stupid things.

Here is the commit 05fe478dd04e02fa230c305ab9b5616669821dd3:

mm/filemap.c +1 -1
index f9d8818..9c5e623 100644
@@ -210,7 +210,7 @@ int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
        int ret;
        struct writeback_control wbc = {
                .sync_mode = sync_mode,
-               .nr_to_write = mapping->nrpages * 2,
+               .nr_to_write = LONG_MAX,
                .range_start = start,
                .range_end = end,
        };
mm/page-writeback.c +4 -2
index 2e847cd..5edca67 100644
@@ -963,8 +963,10 @@ retry:
                                }
                        }
 
-                       if (--nr_to_write <= 0)
-                               done = 1;
+                       if (wbc->sync_mode == WB_SYNC_NONE) {
+                               if (--wbc->nr_to_write <= 0)
+                                       done = 1;
+                       }
                        if (wbc->nonblocking && bdi_write_congested(bdi)) {
                                wbc->encountered_congestion = 1;
                                done = 1;

It makes the following changes:
1. Decrement wbc->nr_to_write instead of nr_to_write
2. Decrement wbc->nr_to_write _only_ if wbc->sync_mode == WB_SYNC_NONE
3. If synced nr_to_write pages, stop only if if wbc->sync_mode == WB_SYNC_NONE, otherwise keep going.

The commit message talks only about item 3, at leaset as I understand it. I do not quote the commit message because it is large. Thus, I thought changes 1 and 2 were not intentional.

In my patch I try to
1. Undo changes 1 and 2
2. Add a comment explaining change 3 (it very useful to have comments in _code_, not only in the commit)

Quoted 2 lines

I thought your problem might have been that you were calling this with WB_SYNC_ALL and expecting it to heed nr_to_write, however...

Err, my problem is that wbc->nr_to_write is not updated.

Quoted 6 lines
Quoted 2 lines

Also, I think wbc->nr_to_write should be changed in all cases, not only when wbc->sync_mode == WB_SYNC_NONE.

... you mention this here like it is an *additional* issue on top of your problem. So I fail to see how my commit could have caused this problem?

This is the change number 2 in your commit.

Quoted 5 lines
Quoted 2 lines

Well, in case of @wbc->no_nrwrite_index_update != 0, we do change wbc->nr_to_write, while we should not. This patch fixes this behavior.

And I don't know what you mean by this because the patch doesn't fix any problem there AFAIKS.

This is about change number 1. In the "for" loop you change wbc->nr_to_write, instead of nr_to_write. But before the function returns, it writes nr_to_write back to wbc->nr_to_write, so the result is that the caller sees wbc->nr_to_write unchanged.

Quoted 4 lines

Anyway, I did probably not pay enough attention to ubifs when making this change, and if it wants wbc->nr_to_write updated even for data integrity syncs, I don't see the harm in that. So I don't have any objection to your patch. Thanks.

It is just how things were _before_ your patch.

Quoted 2 lines

Can you cc [email protected] when a final version gets merged upstream please?

Sure, I just want to get blessing of one of MM/FS gurus. E.g. you Acked-by would be enough.

Thanks.

Signature · 7 lines
Best regards,
Artem Bityutskiy (Битюцкий Артём)

--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#4 lore raw
Quoted 32 lines · On Tuesday 03 February 2009 20:32:49 Artem Bityutskiy wrote:
Quoted 31 lines · Nick Piggin
Quoted 29 lines · On Tuesday 03 February 2009 19:42:22 Artem Bityutskiy wrote:

Hi,

commit 05fe478dd04e02fa230c305ab9b5616669821dd3

Author: Nick Piggin <[email protected]>
Date:   Tue Jan 6 14:39:08 2009 -0800

mm: write_cache_pages integrity fix

broke wbc->nr_to_write handling. Here is the fix.

I'm not 100% sure I got things right, because I am far not expert in the area. Please, review it. The patch fixes my UBIFS issues, which are caused by the fact that wbc->nr_to_write is not updated.
======================================================================

From: Artem Bityutskiy <[email protected]>
Date: Mon, 2 Feb 2009 18:33:49 +0200
Subject: [PATCH] write-back: fix nr_to_write counter

Commit 05fe478dd04e02fa230c305ab9b5616669821dd3 broke @wbc->nr_to_write. 'write_cache_pages()' changes it in the loop, but restores the original value from @nr_to_write at the end, because of this code:

       if (!wbc->no_nrwrite_index_update) {
                if (wbc->range_cyclic || (range_whole && nr_to_write >
0)) mapping->writeback_index = done_index; wbc->nr_to_write =
nr_to_write;
        }
Quoted 8 lines
Quoted 3 lines

The commit you quote only moves nr_to_write to not take effect for WB_SYNC_ALL (ie. data integrity) writeout. And makes no other change to write_cache_pages.

Nick, I'm sorry if my e-mail looked like I'm blame you, I referred the commit because git-bisect pointed to it and I though it for me :-)

And I apologize if I write stupid things.

No, you are probably right to blame me ;) but I just am not quite sure if I read your description correctly.

Quoted 33 lines

Here is the commit 05fe478dd04e02fa230c305ab9b5616669821dd3:

mm/filemap.c +1 -1
index f9d8818..9c5e623 100644
@@ -210,7 +210,7 @@ int __filemap_fdatawrite_range(struct address_space
 *mapping, loff_t start, int ret;
        struct writeback_control wbc = {
                .sync_mode = sync_mode,
-               .nr_to_write = mapping->nrpages * 2,
+               .nr_to_write = LONG_MAX,
                .range_start = start,
                .range_end = end,
        };
mm/page-writeback.c +4 -2
index 2e847cd..5edca67 100644
@@ -963,8 +963,10 @@ retry:
                                }
                        }
 
-                       if (--nr_to_write <= 0)
-                               done = 1;
+                       if (wbc->sync_mode == WB_SYNC_NONE) {
+                               if (--wbc->nr_to_write <= 0)
+                                       done = 1;
+                       }
                        if (wbc->nonblocking && bdi_write_congested(bdi)) {
                                wbc->encountered_congestion = 1;
                                done = 1;
Quoted 2 lines

It makes the following changes:
1. Decrement wbc->nr_to_write instead of nr_to_write

Ah, ok drat yes I missed this. Indeed this is a stupid bug :(

Quoted 7 lines

2. Decrement wbc->nr_to_write _only_ if wbc->sync_mode == WB_SYNC_NONE
3. If synced nr_to_write pages, stop only if if wbc->sync_mode == WB_SYNC_NONE, otherwise keep going.

The commit message talks only about item 3, at leaset as I understand it. I do not quote the commit message because it is large. Thus, I thought changes 1 and 2 were not intentional.

2 was more or less intentional but I glossed over ubifs. At any rate I'm happy if you restore the old behaviour.

Quoted 10 lines

In my patch I try to
1. Undo changes 1 and 2
2. Add a comment explaining change 3 (it very useful to have comments in _code_, not only in the commit)

Quoted 2 lines

I thought your problem might have been that you were calling this with WB_SYNC_ALL and expecting it to heed nr_to_write, however...

Err, my problem is that wbc->nr_to_write is not updated.

Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

Also, I think wbc->nr_to_write should be changed in all cases, not only when wbc->sync_mode == WB_SYNC_NONE.

... you mention this here like it is an *additional* issue on top of your problem. So I fail to see how my commit could have caused this problem?

This is the change number 2 in your commit.

OK, it just wasn't clear to me from the changelog (I missed noticing bug #1).

Quoted 16 lines
Quoted 5 lines
Quoted 2 lines

Well, in case of @wbc->no_nrwrite_index_update != 0, we do change wbc->nr_to_write, while we should not. This patch fixes this behavior.

And I don't know what you mean by this because the patch doesn't fix any problem there AFAIKS.

This is about change number 1. In the "for" loop you change wbc->nr_to_write, instead of nr_to_write. But before the function returns, it writes nr_to_write back to wbc->nr_to_write, so the result is that the caller sees wbc->nr_to_write unchanged.

Quoted 4 lines

Anyway, I did probably not pay enough attention to ubifs when making this change, and if it wants wbc->nr_to_write updated even for data integrity syncs, I don't see the harm in that. So I don't have any objection to your patch. Thanks.

Quoted 7 lines

It is just how things were _before_ your patch.

Quoted 2 lines

Can you cc [email protected] when a final version gets merged upstream please?

Sure, I just want to get blessing of one of MM/FS gurus. E.g. you Acked-by would be enough.

Yes please put an Acked-by: Nick Piggin <[email protected]> on there. But I found your comments in this mail to be easier to understand. Perhaps incorporate them into the changelog?

#5 lore raw
Quoted 9 lines · Nick Piggin
Quoted 5 lines
Quoted 2 lines

Can you cc [email protected] when a final version gets merged upstream please?

Sure, I just want to get blessing of one of MM/FS gurus. E.g. you Acked-by would be enough.

Yes please put an Acked-by: Nick Piggin <[email protected]> on there. But I found your comments in this mail to be easier to understand. Perhaps incorporate them into the changelog?

Will do and re-send it.

Signature · 7 lines
Best regards,
Artem Bityutskiy (Битюцкий Артём)

--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html