f2fs: clear cold data flag if IO is not counted

PATCH In review Jaegeuk Kim 3 participants 11 messages Oct 10, 2018 – Oct 19, 2018 F2FS file system

lore

Share & tools

lore

Jaegeuk Kim

author maintainer

2018-10-10 21:22 UTC #0 lore raw

PATCH 1/1 f2fs: clear cold data flag if IO is not counted

If we clear the cold data flag out of the writeback flow, we can miscount -1 by end_io.

Signed-off-by Jaegeuk Kim <[email protected]>

---
 fs/f2fs/data.c | 4 ----
 1 file changed, 4 deletions(-)

fs/f2fs/data.c +0 -4

index 29a9d3b8f709..4102799b5558 100644
@@ -2636,10 +2636,6 @@ static int f2fs_set_data_page_dirty(struct page *page)
 	if (!PageUptodate(page))
 		SetPageUptodate(page);
 
-	/* don't remain PG_checked flag which was set during GC */
-	if (is_cold_data(page))
-		clear_cold_data(page);
-
 	if (f2fs_is_atomic_file(inode) && !f2fs_is_commit_atomic_write(inode)) {
 		if (!IS_ATOMIC_WRITTEN_PAGE(page)) {
 			f2fs_register_inmem_page(inode, page);

Signature · 1 line

2.19.0.605.g01d371f741-goog

Chao Yu

2018-10-15 12:38 UTC #1 lore raw

Quoted 2 lines · Jaegeuk Kim

If we clear the cold data flag out of the writeback flow, we can miscount -1 by end_io.

I didn't get it, which count do you mean?

Thanks,

Quoted 22 lines

Signed-off-by Jaegeuk Kim <[email protected]>

---
 fs/f2fs/data.c | 4 ----
 1 file changed, 4 deletions(-)

fs/f2fs/data.c +0 -4

index 29a9d3b8f709..4102799b5558 100644
@@ -2636,10 +2636,6 @@ static int f2fs_set_data_page_dirty(struct page *page)
 	if (!PageUptodate(page))
 		SetPageUptodate(page);
 
-	/* don't remain PG_checked flag which was set during GC */
-	if (is_cold_data(page))
-		clear_cold_data(page);
-
 	if (f2fs_is_atomic_file(inode) && !f2fs_is_commit_atomic_write(inode)) {
 		if (!IS_ATOMIC_WRITTEN_PAGE(page)) {
 			f2fs_register_inmem_page(inode, page);

Jaegeuk Kim

author maintainer

2018-10-15 23:08 UTC #2 lore raw

Quoted 5 lines · Chao Yu

Quoted 2 lines · Jaegeuk Kim
If we clear the cold data flag out of the writeback flow, we can miscount -1 by end_io.
I didn't get it, which count do you mean?

It's the number of dirty pages.

Balancing F2FS Async:
- IO (CP: 1, Data: -1, Flush: ( 0 0 1), Discard: ( 0 129304)) cmd: 0 undiscard: 0

Quoted 25 lines

Thanks,

Quoted 22 lines

Signed-off-by Jaegeuk Kim <[email protected]>

---
 fs/f2fs/data.c | 4 ----
 1 file changed, 4 deletions(-)

fs/f2fs/data.c +0 -4

index 29a9d3b8f709..4102799b5558 100644
@@ -2636,10 +2636,6 @@ static int f2fs_set_data_page_dirty(struct page *page)
 	if (!PageUptodate(page))
 		SetPageUptodate(page);
 
-	/* don't remain PG_checked flag which was set during GC */
-	if (is_cold_data(page))
-		clear_cold_data(page);
-
 	if (f2fs_is_atomic_file(inode) && !f2fs_is_commit_atomic_write(inode)) {
 		if (!IS_ATOMIC_WRITTEN_PAGE(page)) {
 			f2fs_register_inmem_page(inode, page);

Chao Yu

2018-10-16 01:31 UTC #3 lore raw

Quoted 5 lines · Jaegeuk Kim

Quoted 4 lines · Chao Yu
Quoted 2 lines · Jaegeuk Kim
If we clear the cold data flag out of the writeback flow, we can miscount -1 by end_io.

Quoted 7 lines

Quoted 1 line
I didn't get it, which count do you mean?
It's the number of dirty pages.
Balancing F2FS Async:
- IO (CP: 1, Data: -1, Flush: ( 0 0 1), Discard: ( 0 129304)) cmd: 0 undiscard: 0

So I guess the race should be:

GC thread: IRQ
- move_data_page()
- set_page_dirty()
- clear_cold_data()
- f2fs_write_end_io()
- type = WB_DATA_TYPE(page); here, we get wrong type
- dec_page_count(sbi, type);
- f2fs_wait_on_page_writeback()

How about relocate wait_writeback & set_page_dirty to avoid above race:

move_data_page()

retry:

		f2fs_wait_on_page_writeback(page, DATA, true);
		set_page_dirty(page);

Thanks,

Quoted 28 lines

Quoted 17 lines

Thanks,

Quoted 14 lines

Signed-off-by Jaegeuk Kim <[email protected]>

---
 fs/f2fs/data.c | 4 ----
 1 file changed, 4 deletions(-)

fs/f2fs/data.c +0 -0

index 29a9d3b8f709..4102799b5558 100644
@@ -2636,10 +2636,6 @@ static int f2fs_set_data_page_dirty(struct page *page)
 	if (!PageUptodate(page))
 		SetPageUptodate(page);

Quoted 8 lines

fs/f2fs/data.c static int f2fs_set_data_page_dirty(struct page *page) · line 2641

Quoted 8 lines

fs/f2fs/data.c +0 -4

-	/* don't remain PG_checked flag which was set during GC */
-	if (is_cold_data(page))
-		clear_cold_data(page);
-
 	if (f2fs_is_atomic_file(inode) && !f2fs_is_commit_atomic_write(inode)) {
 		if (!IS_ATOMIC_WRITTEN_PAGE(page)) {
 			f2fs_register_inmem_page(inode, page);

Jaegeuk Kim

author maintainer

2018-10-16 03:10 UTC #4 lore raw

Quoted 26 lines · Chao Yu

Quoted 5 lines · Jaegeuk Kim
Quoted 4 lines · Chao Yu
Quoted 2 lines · Jaegeuk Kim
If we clear the cold data flag out of the writeback flow, we can miscount -1 by end_io.
Quoted 7 lines
Quoted 1 line
I didn't get it, which count do you mean?
It's the number of dirty pages.
Balancing F2FS Async:
- IO (CP: 1, Data: -1, Flush: ( 0 0 1), Discard: ( 0 129304)) cmd: 0 undiscard: 0
So I guess the race should be:
GC thread: IRQ
- move_data_page()
- set_page_dirty()
- clear_cold_data()
- f2fs_write_end_io()
- type = WB_DATA_TYPE(page); here, we get wrong type
- dec_page_count(sbi, type);
- f2fs_wait_on_page_writeback()

Quoted 1 line

How about relocate wait_writeback & set_page_dirty to avoid above race:

Well, wait_on_stable_page() doesn't guarantee its end_io. So, I'm not sure this is the only case.

Quoted 37 lines

move_data_page()

retry:

		f2fs_wait_on_page_writeback(page, DATA, true);
		set_page_dirty(page);

Thanks,

Quoted 28 lines

Quoted 17 lines

Thanks,

Quoted 14 lines

Signed-off-by Jaegeuk Kim <[email protected]>

---
 fs/f2fs/data.c | 4 ----
 1 file changed, 4 deletions(-)

fs/f2fs/data.c +0 -0

index 29a9d3b8f709..4102799b5558 100644
@@ -2636,10 +2636,6 @@ static int f2fs_set_data_page_dirty(struct page *page)
 	if (!PageUptodate(page))
 		SetPageUptodate(page);

Quoted 8 lines

fs/f2fs/data.c static int f2fs_set_data_page_dirty(struct page *page) · line 2641

Quoted 8 lines

fs/f2fs/data.c +0 -4

-	/* don't remain PG_checked flag which was set during GC */
-	if (is_cold_data(page))
-		clear_cold_data(page);
-
 	if (f2fs_is_atomic_file(inode) && !f2fs_is_commit_atomic_write(inode)) {
 		if (!IS_ATOMIC_WRITTEN_PAGE(page)) {
 			f2fs_register_inmem_page(inode, page);

Chao Yu

2018-10-17 01:30 UTC #5 lore raw

Quoted 7 lines · Jaegeuk Kim

Quoted 6 lines · Chao Yu
Quoted 5 lines · Jaegeuk Kim
Quoted 4 lines · Chao Yu
Quoted 2 lines · Jaegeuk Kim
If we clear the cold data flag out of the writeback flow, we can miscount -1 by end_io.

Quoted 6 lines

Quoted 6 lines
Quoted 6 lines
Quoted 1 line
I didn't get it, which count do you mean?
It's the number of dirty pages.
Balancing F2FS Async:
- IO (CP: 1, Data: -1, Flush: ( 0 0 1), Discard: ( 0 129304)) cmd: 0 undiscard: 0

Better to add such info in commit message. :)

Quoted 14 lines

Quoted 14 lines
Quoted 1 line
So I guess the race should be:
GC thread: IRQ
- move_data_page()
- set_page_dirty()
- clear_cold_data()
- f2fs_write_end_io()
- type = WB_DATA_TYPE(page); here, we get wrong type
- dec_page_count(sbi, type);
- f2fs_wait_on_page_writeback()

Quoted 4 lines

Quoted 1 line
How about relocate wait_writeback & set_page_dirty to avoid above race:
Well, wait_on_stable_page() doesn't guarantee its end_io. So, I'm not sure this is the only case.

Yes, you're right, I missed that case.

Can you use git-revert to generate the patch? so we can remain original commit info for better backward tracking.

How do you think pick up original patch I submitted:

https://lkml.org/lkml/2018/7/27/415

Thanks,

Quoted 41 lines

Quoted 38 lines
move_data_page()
retry:
		f2fs_wait_on_page_writeback(page, DATA, true);
		set_page_dirty(page);
Thanks,
Quoted 28 lines
Quoted 17 lines
Thanks,
Quoted 14 lines
Signed-off-by Jaegeuk Kim <[email protected]>
---
 fs/f2fs/data.c | 4 ----
 1 file changed, 4 deletions(-)
fs/f2fs/data.c +0 -0
index 29a9d3b8f709..4102799b5558 100644
@@ -2636,10 +2636,6 @@ static int f2fs_set_data_page_dirty(struct page *page)
 	if (!PageUptodate(page))
 		SetPageUptodate(page);
 
Quoted 8 lines
fs/f2fs/data.c static int f2fs_set_data_page_dirty(struct page *page) · line 2641
Quoted 8 lines
fs/f2fs/data.c +0 -4
-	/* don't remain PG_checked flag which was set during GC */
-	if (is_cold_data(page))
-		clear_cold_data(page);
-
 	if (f2fs_is_atomic_file(inode) && !f2fs_is_commit_atomic_write(inode)) {
 		if (!IS_ATOMIC_WRITTEN_PAGE(page)) {
 			f2fs_register_inmem_page(inode, page);
.
.

Jaegeuk Kim

author maintainer

2018-10-17 02:43 UTC #6 lore raw

Quoted 36 lines · Chao Yu

Quoted 7 lines · Jaegeuk Kim
Quoted 6 lines · Chao Yu
Quoted 5 lines · Jaegeuk Kim
Quoted 4 lines · Chao Yu
Quoted 2 lines · Jaegeuk Kim
If we clear the cold data flag out of the writeback flow, we can miscount -1 by end_io.
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 1 line
I didn't get it, which count do you mean?
It's the number of dirty pages.
Balancing F2FS Async:
- IO (CP: 1, Data: -1, Flush: ( 0 0 1), Discard: ( 0 129304)) cmd: 0 undiscard: 0
Better to add such info in commit message. :)
Quoted 14 lines
Quoted 14 lines
Quoted 1 line
So I guess the race should be:
GC thread: IRQ
- move_data_page()
- set_page_dirty()
- clear_cold_data()
- f2fs_write_end_io()
- type = WB_DATA_TYPE(page); here, we get wrong type
- dec_page_count(sbi, type);
- f2fs_wait_on_page_writeback()
Quoted 4 lines
Quoted 1 line
How about relocate wait_writeback & set_page_dirty to avoid above race:
Well, wait_on_stable_page() doesn't guarantee its end_io. So, I'm not sure this is the only case.

Quoted 8 lines

Yes, you're right, I missed that case.
Can you use git-revert to generate the patch? so we can remain original commit info for better backward tracking.
How do you think pick up original patch I submitted:
https://lkml.org/lkml/2018/7/27/415

Seems that works.

Thanks,

Quoted 44 lines

Thanks,
Quoted 41 lines
Quoted 38 lines
move_data_page()
retry:
		f2fs_wait_on_page_writeback(page, DATA, true);
		set_page_dirty(page);
Thanks,
Quoted 28 lines
Quoted 17 lines
Thanks,
Quoted 14 lines
Signed-off-by Jaegeuk Kim <[email protected]>
---
 fs/f2fs/data.c | 4 ----
 1 file changed, 4 deletions(-)
fs/f2fs/data.c +0 -0
index 29a9d3b8f709..4102799b5558 100644
@@ -2636,10 +2636,6 @@ static int f2fs_set_data_page_dirty(struct page *page)
 	if (!PageUptodate(page))
 		SetPageUptodate(page);
 
Quoted 8 lines
fs/f2fs/data.c static int f2fs_set_data_page_dirty(struct page *page) · line 2641
Quoted 8 lines
fs/f2fs/data.c +0 -4
-	/* don't remain PG_checked flag which was set during GC */
-	if (is_cold_data(page))
-		clear_cold_data(page);
-
 	if (f2fs_is_atomic_file(inode) && !f2fs_is_commit_atomic_write(inode)) {
 		if (!IS_ATOMIC_WRITTEN_PAGE(page)) {
 			f2fs_register_inmem_page(inode, page);
.
.

Jaegeuk Kim

author maintainer

2018-10-17 02:34 UTC #7 lore raw

In reply to Jaegeuk Kim

This reverts commit 66110abc4c931f879d70e83e1281f891699364bf.

If we clear the cold data flag out of the writeback flow, we can miscount -1 by end_io.

Balancing F2FS Async:
- IO (CP: 1, Data: -1, Flush: ( 0 0 1), Discard: ( ...

Cc <[email protected]>

Signed-off-by Jaegeuk Kim <[email protected]>

---
 fs/f2fs/data.c | 4 ----
 1 file changed, 4 deletions(-)

fs/f2fs/data.c +0 -4

index 3f272c18fb61..0d0b4dd55b04 100644
@@ -2650,10 +2650,6 @@ static int f2fs_set_data_page_dirty(struct page *page)
 	if (!PageUptodate(page))
 		SetPageUptodate(page);
 
-	/* don't remain PG_checked flag which was set during GC */
-	if (is_cold_data(page))
-		clear_cold_data(page);
-
 	if (f2fs_is_atomic_file(inode) && !f2fs_is_commit_atomic_write(inode)) {
 		if (!IS_ATOMIC_WRITTEN_PAGE(page)) {
 			f2fs_register_inmem_page(inode, page);

Signature · 1 line

2.19.0.605.g01d371f741-goog

Chao Yu

Reviewed-by

2018-10-17 02:51 UTC #8 lore raw

Quoted 18 lines · Jaegeuk Kim

This reverts commit 66110abc4c931f879d70e83e1281f891699364bf.
If we clear the cold data flag out of the writeback flow, we can miscount -1 by end_io.
Balancing F2FS Async:
- IO (CP: 1, Data: -1, Flush: ( 0 0 1), Discard: ( ...
GC thread: IRQ
- move_data_page()
- set_page_dirty()
- clear_cold_data()
- f2fs_write_end_io()
- type = WB_DATA_TYPE(page); here, we get wrong type
- dec_page_count(sbi, type);
- f2fs_wait_on_page_writeback()

Quoted 2 lines

Cc <[email protected]>
Signed-off-by Jaegeuk Kim <[email protected]>

Reviewed-by Chao Yu <[email protected]>

Thanks,

Ju Hyung Park

Tested-by

2018-10-19 18:16 UTC #9 lore raw

In reply to Jaegeuk Kim

This does seem to fix the problem I've been having.

I was encountering all I/Os to be blocked on my Android device almost on a daily basis, but after this patch I'm rock solid for almost a week.

I'd appreciate it if you could add the following tags:

Reported-by Park Ju Hyung <[email protected]>

Tested-by Park Ju Hyung <[email protected]>

Also, a little additional note would be helpful
for those who are unfamiliar with f2fs:
This fixes a deadlock causing all I/Os to be blocked during heavy GC.

Thanks,

Quoted 49 lines · Jaegeuk Kim

This reverts commit 66110abc4c931f879d70e83e1281f891699364bf.
If we clear the cold data flag out of the writeback flow, we can miscount -1 by end_io.
Balancing F2FS Async:
- IO (CP: 1, Data: -1, Flush: ( 0 0 1), Discard: ( ...
GC thread: IRQ
- move_data_page()
- set_page_dirty()
- clear_cold_data()
- f2fs_write_end_io()
- type = WB_DATA_TYPE(page); here, we get wrong type
- dec_page_count(sbi, type);
- f2fs_wait_on_page_writeback()
Cc <[email protected]>
Signed-off-by Jaegeuk Kim <[email protected]>
---
 fs/f2fs/data.c | 4 ----
 1 file changed, 4 deletions(-)
fs/f2fs/data.c +0 -5
index 3f272c18fb61..0d0b4dd55b04 100644
@@ -2650,10 +2650,6 @@ static int f2fs_set_data_page_dirty(struct page *page)
        if (!PageUptodate(page))
                SetPageUptodate(page);
 
-       /* don't remain PG_checked flag which was set during GC */
-       if (is_cold_data(page))
-               clear_cold_data(page);
-
        if (f2fs_is_atomic_file(inode) && !f2fs_is_commit_atomic_write(inode)) {
                if (!IS_ATOMIC_WRITTEN_PAGE(page)) {
                        f2fs_register_inmem_page(inode, page);
--
2.19.0.605.g01d371f741-goog
_______________________________________________
Linux-f2fs-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel

Jaegeuk Kim

author maintainer

2018-10-19 22:58 UTC #10 lore raw

Quoted 5 lines · Ju Hyung Park

This does seem to fix the problem I've been having.
I was encountering all I/Os to be blocked on my Android device almost on a daily basis, but after this patch I'm rock solid for almost a week.

Quoted 7 lines

I'd appreciate it if you could add the following tags:
Reported-by Park Ju Hyung <[email protected]>
Tested-by Park Ju Hyung <[email protected]>
Also, a little additional note would be helpful
for those who are unfamiliar with f2fs:
This fixes a deadlock causing all I/Os to be blocked during heavy GC.

Cool. Please check the below. :P

https://git.kernel.org/pub/scm/linux/kernel/git/jaegeuk/f2fs.git/commit/?h=dev&id=b921ec4e7fb5df13f805bc04375ed529faa095ae

Thanks,

Quoted 52 lines

Thanks,
Quoted 49 lines · Jaegeuk Kim
This reverts commit 66110abc4c931f879d70e83e1281f891699364bf.
If we clear the cold data flag out of the writeback flow, we can miscount -1 by end_io.
Balancing F2FS Async:
- IO (CP: 1, Data: -1, Flush: ( 0 0 1), Discard: ( ...
GC thread: IRQ
- move_data_page()
- set_page_dirty()
- clear_cold_data()
- f2fs_write_end_io()
- type = WB_DATA_TYPE(page); here, we get wrong type
- dec_page_count(sbi, type);
- f2fs_wait_on_page_writeback()
Cc <[email protected]>
Signed-off-by Jaegeuk Kim <[email protected]>
---
 fs/f2fs/data.c | 4 ----
 1 file changed, 4 deletions(-)
fs/f2fs/data.c +0 -5
index 3f272c18fb61..0d0b4dd55b04 100644
@@ -2650,10 +2650,6 @@ static int f2fs_set_data_page_dirty(struct page *page)
        if (!PageUptodate(page))
                SetPageUptodate(page);
 
-       /* don't remain PG_checked flag which was set during GC */
-       if (is_cold_data(page))
-               clear_cold_data(page);
-
        if (f2fs_is_atomic_file(inode) && !f2fs_is_commit_atomic_write(inode)) {
                if (!IS_ATOMIC_WRITTEN_PAGE(page)) {
                        f2fs_register_inmem_page(inode, page);
--
2.19.0.605.g01d371f741-goog
_______________________________________________
Linux-f2fs-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel