bcache: Fix writeback_thread never writing back incomplete stripes.

Fix writeback_thread never finishing writing back all dirty data in bcache when partial_stripes_expensive is set, and spinning, consuming 100% of CPU instead.

Signed-off-by Vojtech Pavlik <[email protected]>

---

This is a fix for the current upstream bcache, not the devel branch.

If partial_stripes_expensive is set for a cache set, then writeback_thread always attempts to write full stripes out back to the backing device first. However, since it does that based on a bitmap and not a simple linear search, like the rest of the code of refill_dirty(), it changes the last_scanned pointer so that never points to 'end'. refill_dirty() then never tries to scan from 'start', resulting in the writeback_thread looping, consuming 100% of CPU, but never making progress in writing out the incomplete dirty stripes in the cache.

Scanning the tree after not finding enough full stripes fixes the issue.

Incomplete dirty stripes are written to the backing device, the device eventually reaches a clean state if there is nothing dirtying data and writeback_thread sleeps. This also fixes the problem of the cache device not being possible to detach in the partial_stripes_expensive scenario.

It may be more efficient to separate the last_scanned field for normal and stripe scans instead.

Quoted 17 lines · Vojtech Pavlik

Fix writeback_thread never finishing writing back all dirty data in bcache when partial_stripes_expensive is set, and spinning, consuming 100% of CPU instead.

Signed-off-by Vojtech Pavlik <[email protected]>

---

This is a fix for the current upstream bcache, not the devel branch.

If partial_stripes_expensive is set for a cache set, then writeback_thread always attempts to write full stripes out back to the backing device first. However, since it does that based on a bitmap and not a simple linear search, like the rest of the code of refill_dirty(), it changes the last_scanned pointer so that never points to 'end'. refill_dirty() then never tries to scan from 'start', resulting in the writeback_thread looping, consuming 100% of CPU, but never making progress in writing out the incomplete dirty stripes in the cache.

Quoted 6 lines

Scanning the tree after not finding enough full stripes fixes the issue.

Incomplete dirty stripes are written to the backing device, the device eventually reaches a clean state if there is nothing dirtying data and writeback_thread sleeps. This also fixes the problem of the cache device not being possible to detach in the partial_stripes_expensive scenario.

Good catch!

Quoted 2 lines

It may be more efficient to separate the last_scanned field for normal and stripe scans instead.

Actually, I think I like your approach - I think it gives us the behaviour we want, although it took some thinking to figure out why.

One of the reasons for last_scanned is just to do writeback in LBA order and avoid making the disks seek around - so, if refill_full_stripes() did just queue up some IO at last_scanned, we do want to keep scanning from that position for non full stripes.

But since (as you noted) last_scanned is never going to be >= end after calling refill_full_strips() (if there were any full stripes) - really the correct thing to do is loop around to the start if necessary so that we can successfully scan everything.

The only inefficiency I see with your approach is that on the second scan, after we've looped, we don't need to scan to the end of the disk - we only need to scan up to where we initially started.

But, another observation - if we change refill_dirty() so that it always scans the entire keyspace if necessary, regardless of where last_scanned was - well, that really isn't specific to the refill_full_stripes() case - we can just always do that.

Can you give this patch a try?

-- >8 --
Subject: [PATCH] bcache: Change refill_dirty() to always scan entire disk if necessary

Previously, it would only scan the entire disk if it was starting from the very start of the disk - i.e. if the previous scan got to the end.

This was broken by refill_full_stripes(), which updates last_scanned so that refill_dirty was never triggering the searched_from_start path.

But if we change refill_dirty() to always scan the entire disk if necessary, regardless of what last_scanned was, the code gets cleaner and we fix that bug too.

Signed-off-by Kent Overstreet <[email protected]>

---
 drivers/md/bcache/writeback.c | 24 ++++++++++++++++--------
 1 file changed, 16 insertions(+), 8 deletions(-)

drivers/md/bcache/writeback.c +16 -9

index cdde0f32f0..08a52db38b 100644
@@ -359,11 +359,13 @@ next:
        }
 }
 
+/*
+ * Returns true if we scanned the entire disk
+ */
 static bool refill_dirty(struct cached_dev *dc)
 {
        struct keybuf *buf = &dc->writeback_keys;
-       struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
-       bool searched_from_start = false;
+       struct bkey start_pos, end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
 
        if (dc->partial_stripes_expensive) {
                refill_full_stripes(dc);

@@ -371,14 +373,20 @@ static bool refill_dirty(struct cached_dev *dc)
                        return false;
        }
 
-       if (bkey_cmp(&buf->last_scanned, &end) >= 0) {
-               buf->last_scanned = KEY(dc->disk.id, 0, 0);
-               searched_from_start = true;
-       }
-
+       start_pos = buf->last_scanned;
        bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
 
-       return bkey_cmp(&buf->last_scanned, &end) >= 0 && searched_from_start;
+       if (bkey_cmp(&buf->last_scanned, &end) < 0)
+               return false;
+
+       /*
+        * If we get to the end start scanning again from the beginning, and
+        * only scan up to where we initially started scanning from:
+        */
+       buf->last_scanned = KEY(dc->disk.id, 0, 0);
+       bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
+
+       return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
 }
 
 static void bch_writeback(struct cached_dev *dc)
--

2.5.1

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

Hi Kent, Vojtech, list

Your fix works perfectly, I can finally remove my patch that disables the partial_stripes_expensive branch and unleash the full bcache performance on my RAID-6 array. I was aware of this problem for some time now, but never got to learning the bcache codebase enough to find out why this happens with partial stripes write-back enabled, so I just disabled it to avoid CPU spinning. Thanks a lot to Vojtech for finding the reason and to you for the patch. I have quite a collection of stability patches for the mainline bcache. Would you please be kind to review them when convenient and merge upstream the ones you think are worth merging. It will be 4.4 merge window, I believe. Please find all the patches attached. All of them are rebased on mainline 4.2 kernel. If somebody needs the 4.1.x versions, please let me know.

Quoted 122 lines · Kent Overstreet

Quoted 17 lines · Vojtech Pavlik

Fix writeback_thread never finishing writing back all dirty data in bcache when partial_stripes_expensive is set, and spinning, consuming 100% of CPU instead.

Signed-off-by Vojtech Pavlik <[email protected]>

---

This is a fix for the current upstream bcache, not the devel branch.

If partial_stripes_expensive is set for a cache set, then writeback_thread always attempts to write full stripes out back to the backing device first. However, since it does that based on a bitmap and not a simple linear search, like the rest of the code of refill_dirty(), it changes the last_scanned pointer so that never points to 'end'. refill_dirty() then never tries to scan from 'start', resulting in the writeback_thread looping, consuming 100% of CPU, but never making progress in writing out the incomplete dirty stripes in the cache.

Quoted 6 lines

Scanning the tree after not finding enough full stripes fixes the issue.

Incomplete dirty stripes are written to the backing device, the device eventually reaches a clean state if there is nothing dirtying data and writeback_thread sleeps. This also fixes the problem of the cache device not being possible to detach in the partial_stripes_expensive scenario.

Good catch!

Quoted 2 lines

It may be more efficient to separate the last_scanned field for normal and stripe scans instead.

Actually, I think I like your approach - I think it gives us the behaviour we want, although it took some thinking to figure out why.

One of the reasons for last_scanned is just to do writeback in LBA order and avoid making the disks seek around - so, if refill_full_stripes() did just queue up some IO at last_scanned, we do want to keep scanning from that position for non full stripes.

But since (as you noted) last_scanned is never going to be >= end after calling refill_full_strips() (if there were any full stripes) - really the correct thing to do is loop around to the start if necessary so that we can successfully scan everything.

The only inefficiency I see with your approach is that on the second scan, after we've looped, we don't need to scan to the end of the disk - we only need to scan up to where we initially started.

But, another observation - if we change refill_dirty() so that it always scans the entire keyspace if necessary, regardless of where last_scanned was - well, that really isn't specific to the refill_full_stripes() case - we can just always do that.

Can you give this patch a try?

-- >8 --
Subject: [PATCH] bcache: Change refill_dirty() to always scan entire disk if necessary

Previously, it would only scan the entire disk if it was starting from the very start of the disk - i.e. if the previous scan got to the end.

This was broken by refill_full_stripes(), which updates last_scanned so that refill_dirty was never triggering the searched_from_start path.

But if we change refill_dirty() to always scan the entire disk if necessary, regardless of what last_scanned was, the code gets cleaner and we fix that bug too.

Signed-off-by Kent Overstreet <[email protected]>

---
 drivers/md/bcache/writeback.c | 24 ++++++++++++++++--------
 1 file changed, 16 insertions(+), 8 deletions(-)

drivers/md/bcache/writeback.c +16 -9

index cdde0f32f0..08a52db38b 100644
@@ -359,11 +359,13 @@ next:
        }
 }
 
+/*
+ * Returns true if we scanned the entire disk
+ */
 static bool refill_dirty(struct cached_dev *dc)
 {
        struct keybuf *buf = &dc->writeback_keys;
-       struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
-       bool searched_from_start = false;
+       struct bkey start_pos, end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
 
        if (dc->partial_stripes_expensive) {
                refill_full_stripes(dc);

@@ -371,14 +373,20 @@ static bool refill_dirty(struct cached_dev *dc)
                        return false;
        }
 
-       if (bkey_cmp(&buf->last_scanned, &end) >= 0) {
-               buf->last_scanned = KEY(dc->disk.id, 0, 0);
-               searched_from_start = true;
-       }
-
+       start_pos = buf->last_scanned;
        bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
 
-       return bkey_cmp(&buf->last_scanned, &end) >= 0 && searched_from_start;
+       if (bkey_cmp(&buf->last_scanned, &end) < 0)
+               return false;
+
+       /*
+        * If we get to the end start scanning again from the beginning, and
+        * only scan up to where we initially started scanning from:
+        */
+       buf->last_scanned = KEY(dc->disk.id, 0, 0);
+       bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
+
+       return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
 }
 
 static void bch_writeback(struct cached_dev *dc)
--

2.5.1

Quoted 4 lines

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

--

Denis

Well, it turns out my celebration was a bit premature.

PLEASE, DO NOT APPLY THE PATCH POSTED BY KENT (not the one Vojtech posted) ON A PRODUCTION SYSTEM, IT CAUSES DATA CORRUPTION.

The interesting thing is that it somehow damaged the partition that was not supposed to receive any writes (the file system was mounted read-only), so my guess is that the patch causes the blocks residing in the write-back cache to flush to the wrong blocks on the backing device.
Everything was going great until I rebooted and saw this in the log:

[   19.639082] attempt to access beyond end of device
[   19.643984] md1p2: rw=1, want=75497520, limit=62914560
[   19.659033] attempt to access beyond end of device
[   19.663929] md1p2: rw=1, want=75497624, limit=62914560
[   19.669447] attempt to access beyond end of device
[   19.674338] md1p2: rw=1, want=75497752, limit=62914560
[   19.679195] attempt to access beyond end of device
[   19.679199] md1p2: rw=1, want=75498080, limit=62914560
[   19.689007] attempt to access beyond end of device
[   19.689011] md1p2: rw=1, want=75563376, limit=62914560
[   19.699055] attempt to access beyond end of device
[   19.699059] md1p2: rw=1, want=79691816, limit=62914560
[   19.719246] attempt to access beyond end of device
[   19.724144] md1p2: rw=1, want=79691928, limit=62914560

......
(it's a small example, the list was much longer)
And the next thing I found out the super block on my 10-Tb XFS RAID was gone. :) Oh well, it's a good thing I have backups.
I knew what I was doing when trying the untested patches. I should have made the RAID md partition read-only, not the file system. I kind of expected that something could have gone wrong with the file system I was testing, just did not expect it would fire nukes at the innocent bystanders.

Quoted 26 lines · Denis Bychkov

Well, it turns out my celebration was a bit premature.

PLEASE, DO NOT APPLY THE PATCH POSTED BY KENT (not the one Vojtech posted) ON A PRODUCTION SYSTEM, IT CAUSES DATA CORRUPTION.

The interesting thing is that it somehow damaged the partition that was not supposed to receive any writes (the file system was mounted read-only), so my guess is that the patch causes the blocks residing in the write-back cache to flush to the wrong blocks on the backing device.
Everything was going great until I rebooted and saw this in the log:

[   19.639082] attempt to access beyond end of device
[   19.643984] md1p2: rw=1, want=75497520, limit=62914560
[   19.659033] attempt to access beyond end of device
[   19.663929] md1p2: rw=1, want=75497624, limit=62914560
[   19.669447] attempt to access beyond end of device
[   19.674338] md1p2: rw=1, want=75497752, limit=62914560
[   19.679195] attempt to access beyond end of device
[   19.679199] md1p2: rw=1, want=75498080, limit=62914560
[   19.689007] attempt to access beyond end of device
[   19.689011] md1p2: rw=1, want=75563376, limit=62914560
[   19.699055] attempt to access beyond end of device
[   19.699059] md1p2: rw=1, want=79691816, limit=62914560
[   19.719246] attempt to access beyond end of device
[   19.724144] md1p2: rw=1, want=79691928, limit=62914560

Quoted 9 lines

......
(it's a small example, the list was much longer)
And the next thing I found out the super block on my 10-Tb XFS RAID was gone. :) Oh well, it's a good thing I have backups.
I knew what I was doing when trying the untested patches. I should have made the RAID md partition read-only, not the file system. I kind of expected that something could have gone wrong with the file system I was testing, just did not expect it would fire nukes at the innocent bystanders.

Aw, shit. That's just _bizzare_.

I have a theory - it appears that last_scanned isn't getting initialized before it's used, so it's going to be all 0s the very first time... which it appears could cause it to slurp up keys from the wrong device (and if that device was bigger than the correct device, that could explain the accesses beyond the end of the device).

Currently just a theory though, and I have no clue why it would only be exposed with my patch.

Quoted 39 lines · Kent Overstreet

Quoted 26 lines · Denis Bychkov

Well, it turns out my celebration was a bit premature.

PLEASE, DO NOT APPLY THE PATCH POSTED BY KENT (not the one Vojtech posted) ON A PRODUCTION SYSTEM, IT CAUSES DATA CORRUPTION.

The interesting thing is that it somehow damaged the partition that was not supposed to receive any writes (the file system was mounted read-only), so my guess is that the patch causes the blocks residing in the write-back cache to flush to the wrong blocks on the backing device.
Everything was going great until I rebooted and saw this in the log:

[   19.639082] attempt to access beyond end of device
[   19.643984] md1p2: rw=1, want=75497520, limit=62914560
[   19.659033] attempt to access beyond end of device
[   19.663929] md1p2: rw=1, want=75497624, limit=62914560
[   19.669447] attempt to access beyond end of device
[   19.674338] md1p2: rw=1, want=75497752, limit=62914560
[   19.679195] attempt to access beyond end of device
[   19.679199] md1p2: rw=1, want=75498080, limit=62914560
[   19.689007] attempt to access beyond end of device
[   19.689011] md1p2: rw=1, want=75563376, limit=62914560
[   19.699055] attempt to access beyond end of device
[   19.699059] md1p2: rw=1, want=79691816, limit=62914560
[   19.719246] attempt to access beyond end of device
[   19.724144] md1p2: rw=1, want=79691928, limit=62914560

Quoted 9 lines

......
(it's a small example, the list was much longer)
And the next thing I found out the super block on my 10-Tb XFS RAID was gone. :) Oh well, it's a good thing I have backups.
I knew what I was doing when trying the untested patches. I should have made the RAID md partition read-only, not the file system. I kind of expected that something could have gone wrong with the file system I was testing, just did not expect it would fire nukes at the innocent bystanders.

Aw, shit. That's just _bizzare_.

Quoted 8 lines

I have a theory - it appears that last_scanned isn't getting initialized before it's used, so it's going to be all 0s the very first time... which it appears could cause it to slurp up keys from the wrong device (and if that device was bigger than the correct device, that could explain the accesses beyond the end of the device).

Currently just a theory though, and I have no clue why it would only be exposed with my patch.

Here's an updated patch that has a fix for _that_ theory, and also a new BUG_ON(). Any chance you could test it?

Oh - I didn't ask - _do_ you have multiple backing devices attached to the same cache set? Because if you don't, this isn't it at all...

-- >8 --
Subject: [PATCH] bcache: Change refill_dirty() to always scan entire disk if necessary

Previously, it would only scan the entire disk if it was starting from the very start of the disk - i.e. if the previous scan got to the end.

This was broken by refill_full_stripes(), which updates last_scanned so that refill_dirty was never triggering the searched_from_start path.

But if we change refill_dirty() to always scan the entire disk if necessary, regardless of what last_scanned was, the code gets cleaner and we fix that bug too.

Signed-off-by Kent Overstreet <[email protected]>

---
 drivers/md/bcache/writeback.c | 37 ++++++++++++++++++++++++++++++-------
 1 file changed, 30 insertions(+), 7 deletions(-)

drivers/md/bcache/writeback.c +30 -8

index cdde0f32f0..d383024247 100644
@@ -310,6 +310,10 @@ void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned inode,
 
 static bool dirty_pred(struct keybuf *buf, struct bkey *k)
 {
+       struct cached_dev *dc = container_of(buf, struct cached_dev, writeback_keys);
+
+       BUG_ON(KEY_INODE(k) != dc->disk.id);
+
        return KEY_DIRTY(k);
 }
 

@@ -359,11 +363,24 @@ next:
        }
 }
 
+/*
+ * Returns true if we scanned the entire disk
+ */
 static bool refill_dirty(struct cached_dev *dc)
 {
        struct keybuf *buf = &dc->writeback_keys;
+       struct bkey start = KEY(dc->disk.id, 0, 0);
        struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
-       bool searched_from_start = false;
+       struct bkey start_pos;
+
+       /*
+        * make sure keybuf pos is inside the range for this disk - at bringup
+        * we might not be attached yet so this disk's inode nr isn't
+        * initialized then
+        */
+       if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
+           bkey_cmp(&buf->last_scanned, &end) > 0)
+               buf->last_scanned = start;
 
        if (dc->partial_stripes_expensive) {
                refill_full_stripes(dc);

@@ -371,14 +388,20 @@ static bool refill_dirty(struct cached_dev *dc)
                        return false;
        }
 
-       if (bkey_cmp(&buf->last_scanned, &end) >= 0) {
-               buf->last_scanned = KEY(dc->disk.id, 0, 0);
-               searched_from_start = true;
-       }
-
+       start_pos = buf->last_scanned;
        bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
 
-       return bkey_cmp(&buf->last_scanned, &end) >= 0 && searched_from_start;
+       if (bkey_cmp(&buf->last_scanned, &end) < 0)
+               return false;
+
+       /*
+        * If we get to the end start scanning again from the beginning, and
+        * only scan up to where we initially started scanning from:
+        */
+       buf->last_scanned = start;
+       bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
+
+       return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
 }
 
 static void bch_writeback(struct cached_dev *dc)
--

2.5.1

Quoted 39 lines · Kent Overstreet

Quoted 26 lines · Denis Bychkov

Well, it turns out my celebration was a bit premature.

PLEASE, DO NOT APPLY THE PATCH POSTED BY KENT (not the one Vojtech posted) ON A PRODUCTION SYSTEM, IT CAUSES DATA CORRUPTION.

The interesting thing is that it somehow damaged the partition that was not supposed to receive any writes (the file system was mounted read-only), so my guess is that the patch causes the blocks residing in the write-back cache to flush to the wrong blocks on the backing device.
Everything was going great until I rebooted and saw this in the log:

[   19.639082] attempt to access beyond end of device
[   19.643984] md1p2: rw=1, want=75497520, limit=62914560
[   19.659033] attempt to access beyond end of device
[   19.663929] md1p2: rw=1, want=75497624, limit=62914560
[   19.669447] attempt to access beyond end of device
[   19.674338] md1p2: rw=1, want=75497752, limit=62914560
[   19.679195] attempt to access beyond end of device
[   19.679199] md1p2: rw=1, want=75498080, limit=62914560
[   19.689007] attempt to access beyond end of device
[   19.689011] md1p2: rw=1, want=75563376, limit=62914560
[   19.699055] attempt to access beyond end of device
[   19.699059] md1p2: rw=1, want=79691816, limit=62914560
[   19.719246] attempt to access beyond end of device
[   19.724144] md1p2: rw=1, want=79691928, limit=62914560

Quoted 9 lines

......
(it's a small example, the list was much longer)
And the next thing I found out the super block on my 10-Tb XFS RAID was gone. :) Oh well, it's a good thing I have backups.
I knew what I was doing when trying the untested patches. I should have made the RAID md partition read-only, not the file system. I kind of expected that something could have gone wrong with the file system I was testing, just did not expect it would fire nukes at the innocent bystanders.

Aw, shit. That's just _bizzare_.

Quoted 8 lines

I have a theory - it appears that last_scanned isn't getting initialized before it's used, so it's going to be all 0s the very first time... which it appears could cause it to slurp up keys from the wrong device (and if that device was bigger than the correct device, that could explain the accesses beyond the end of the device).

Currently just a theory though, and I have no clue why it would only be exposed with my patch.

Here's an updated patch that has a fix for _that_ theory, and also a new BUG_ON(). Any chance you could test it?

Oh - I didn't ask - _do_ you have multiple backing devices attached to the same cache set? Because if you don't, this isn't it at all...

-- >8 --
Subject: [PATCH] bcache: Change refill_dirty() to always scan entire disk if necessary

Previously, it would only scan the entire disk if it was starting from the very start of the disk - i.e. if the previous scan got to the end.

This was broken by refill_full_stripes(), which updates last_scanned so that refill_dirty was never triggering the searched_from_start path.

But if we change refill_dirty() to always scan the entire disk if necessary, regardless of what last_scanned was, the code gets cleaner and we fix that bug too.

Signed-off-by Kent Overstreet <[email protected]>

---
 drivers/md/bcache/writeback.c | 37 ++++++++++++++++++++++++++++++-------
 1 file changed, 30 insertions(+), 7 deletions(-)

drivers/md/bcache/writeback.c +30 -8

index cdde0f32f0..d383024247 100644
@@ -310,6 +310,10 @@ void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned inode,
 
 static bool dirty_pred(struct keybuf *buf, struct bkey *k)
 {
+       struct cached_dev *dc = container_of(buf, struct cached_dev, writeback_keys);
+
+       BUG_ON(KEY_INODE(k) != dc->disk.id);
+
        return KEY_DIRTY(k);
 }
 

@@ -359,11 +363,24 @@ next:
        }
 }
 
+/*
+ * Returns true if we scanned the entire disk
+ */
 static bool refill_dirty(struct cached_dev *dc)
 {
        struct keybuf *buf = &dc->writeback_keys;
+       struct bkey start = KEY(dc->disk.id, 0, 0);
        struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
-       bool searched_from_start = false;
+       struct bkey start_pos;
+
+       /*
+        * make sure keybuf pos is inside the range for this disk - at bringup
+        * we might not be attached yet so this disk's inode nr isn't
+        * initialized then
+        */
+       if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
+           bkey_cmp(&buf->last_scanned, &end) > 0)
+               buf->last_scanned = start;
 
        if (dc->partial_stripes_expensive) {
                refill_full_stripes(dc);

@@ -371,14 +388,20 @@ static bool refill_dirty(struct cached_dev *dc)
                        return false;
        }
 
-       if (bkey_cmp(&buf->last_scanned, &end) >= 0) {
-               buf->last_scanned = KEY(dc->disk.id, 0, 0);
-               searched_from_start = true;
-       }
-
+       start_pos = buf->last_scanned;
        bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
 
-       return bkey_cmp(&buf->last_scanned, &end) >= 0 && searched_from_start;
+       if (bkey_cmp(&buf->last_scanned, &end) < 0)
+               return false;
+
+       /*
+        * If we get to the end start scanning again from the beginning, and
+        * only scan up to where we initially started scanning from:
+        */
+       buf->last_scanned = start;
+       bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
+
+       return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
 }
 
 static void bch_writeback(struct cached_dev *dc)
--

2.5.1

Here's an updated patch that has a fix for _that_ theory, and also a new BUG_ON(). Any chance you could test it?

Hi Kent, Vojtech, list

Your fix works perfectly, I can finally remove my patch that disables the partial_stripes_expensive branch and unleash the full bcache performance on my RAID-6 array. I was aware of this problem for some time now, but never got to learning the bcache codebase enough to find out why this happens with partial stripes write-back enabled, so I just disabled it to avoid CPU spinning. Thanks a lot to Vojtech for finding the reason and to you for the patch. I have quite a collection of stability patches for the mainline bcache. Would you please be kind to review them when convenient and merge upstream the ones you think are worth merging. It will be 4.4 merge window, I believe. Please find all the patches attached. All of them are rebased on mainline 4.2 kernel. If somebody needs the 4.1.x versions, please let me know.

Quoted 124 lines · Kent Overstreet

Quoted 17 lines · Vojtech Pavlik

Fix writeback_thread never finishing writing back all dirty data in bcache when partial_stripes_expensive is set, and spinning, consuming 100% of CPU instead.

Signed-off-by Vojtech Pavlik <[email protected]>

---

This is a fix for the current upstream bcache, not the devel branch.

If partial_stripes_expensive is set for a cache set, then writeback_thread always attempts to write full stripes out back to the backing device first. However, since it does that based on a bitmap and not a simple linear search, like the rest of the code of refill_dirty(), it changes the last_scanned pointer so that never points to 'end'. refill_dirty() then never tries to scan from 'start', resulting in the writeback_thread looping, consuming 100% of CPU, but never making progress in writing out the incomplete dirty stripes in the cache.

Quoted 6 lines

Scanning the tree after not finding enough full stripes fixes the issue.

Incomplete dirty stripes are written to the backing device, the device eventually reaches a clean state if there is nothing dirtying data and writeback_thread sleeps. This also fixes the problem of the cache device not being possible to detach in the partial_stripes_expensive scenario.

Good catch!

Quoted 2 lines

It may be more efficient to separate the last_scanned field for normal and stripe scans instead.

Actually, I think I like your approach - I think it gives us the behaviour we want, although it took some thinking to figure out why.

One of the reasons for last_scanned is just to do writeback in LBA order and avoid making the disks seek around - so, if refill_full_stripes() did just queue up some IO at last_scanned, we do want to keep scanning from that position for non full stripes.

But since (as you noted) last_scanned is never going to be >= end after calling refill_full_strips() (if there were any full stripes) - really the correct thing to do is loop around to the start if necessary so that we can successfully scan everything.

The only inefficiency I see with your approach is that on the second scan, after we've looped, we don't need to scan to the end of the disk - we only need to scan up to where we initially started.

But, another observation - if we change refill_dirty() so that it always scans the entire keyspace if necessary, regardless of where last_scanned was - well, that really isn't specific to the refill_full_stripes() case - we can just always do that.

Can you give this patch a try?

-- >8 --
Subject: [PATCH] bcache: Change refill_dirty() to always scan entire disk if necessary

Previously, it would only scan the entire disk if it was starting from the very start of the disk - i.e. if the previous scan got to the end.

This was broken by refill_full_stripes(), which updates last_scanned so that refill_dirty was never triggering the searched_from_start path.

But if we change refill_dirty() to always scan the entire disk if necessary, regardless of what last_scanned was, the code gets cleaner and we fix that bug too.

Signed-off-by Kent Overstreet <[email protected]>

---
  drivers/md/bcache/writeback.c | 24 ++++++++++++++++--------
  1 file changed, 16 insertions(+), 8 deletions(-)

drivers/md/bcache/writeback.c +16 -9

index cdde0f32f0..08a52db38b 100644
@@ -359,11 +359,13 @@ next:
         }
  }
 
+/*
+ * Returns true if we scanned the entire disk
+ */
  static bool refill_dirty(struct cached_dev *dc)
  {
         struct keybuf *buf = &dc->writeback_keys;
-       struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
-       bool searched_from_start = false;
+       struct bkey start_pos, end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
 
         if (dc->partial_stripes_expensive) {
                 refill_full_stripes(dc);

@@ -371,14 +373,20 @@ static bool refill_dirty(struct cached_dev *dc)
                         return false;
         }
 
-       if (bkey_cmp(&buf->last_scanned, &end) >= 0) {
-               buf->last_scanned = KEY(dc->disk.id, 0, 0);
-               searched_from_start = true;
-       }
-
+       start_pos = buf->last_scanned;
         bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
 
-       return bkey_cmp(&buf->last_scanned, &end) >= 0 && searched_from_start;
+       if (bkey_cmp(&buf->last_scanned, &end) < 0)
+               return false;
+
+       /*
+        * If we get to the end start scanning again from the beginning, and
+        * only scan up to where we initially started scanning from:
+        */
+       buf->last_scanned = KEY(dc->disk.id, 0, 0);
+       bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
+
+       return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
  }
 
  static void bch_writeback(struct cached_dev *dc)
--

2.5.1

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