Skip to contentSkip to conversation

block: detach bdev inode from its wb in __blkdev_put()

PATCH In review Ilya Dryomov 5 participants 12 messages Nov 20, 2015 – Dec 4, 2015 VFS

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 block: detach bdev inode from its wb in __blkdev_put()

Since 52ebea749aae ("writeback: make backing_dev_info host cgroup-specific bdi_writebacks") inode, at some point in its lifetime, gets attached to a wb (struct bdi_writeback). Detaching happens on evict, in inode_detach_wb() called from __destroy_inode(), and involves updating wb.

However, detaching an internal bdev inode from its wb in __destroy_inode() is too late. Its bdi and by extension root wb are embedded into struct request_queue, which has different lifetime rules and can be freed long before the final bdput() is called (can be from __fput() of a corresponding /dev inode, through dput() - evict() - bd_forget(). bdevs hold onto the underlying disk/queue pair only while opened; as soon as bdev is closed all bets are off. In fact, disk/queue can be gone before __blkdev_put() even returns:

1499 static void __blkdev_put(struct block_device *bdev, fmode_t mode, int for_part) 1500 {
...
1518 if (bdev->bd_contains == bdev) {
1519 if (disk->fops->release)
1520 disk->fops->release(disk, mode);

[ Driver puts its references to disk/queue ]

1521 }
1522 if (!bdev->bd_openers) {
1523 struct module *owner = disk->fops->owner;
1524

1525                 disk_put_part(bdev->bd_part);
1526                 bdev->bd_part = NULL;
1527                 bdev->bd_disk = NULL;
1528                 if (bdev != bdev->bd_contains)

1529 victim = bdev->bd_contains;
1530 bdev->bd_contains = NULL;
1531
1532 put_disk(disk);

[ We put ours, the queue is gone
The last bdput() would result in a write to invalid memory ]

1533 module_put(owner);
...
1539 }

Since bdev inodes are special anyway, detach them in __blkdev_put() after clearing inode's dirty bits, turning the problematic inode_detach_wb() in __destroy_inode() into a noop.

add_disk() grabs its disk->queue since 523e1d399ce0 ("block: make gendisk hold a reference to its queue"), so the old ->release comment is removed in favor of the new inode_detach_wb() comment.

Cc [email protected] # 4.2+, needs backporting
---
 fs/block_dev.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)
fs/block_dev.c +6 -3
index bb0dfb1c7af1..6f8dd407c533 100644
@@ -1509,11 +1509,14 @@ static void __blkdev_put(struct block_device *bdev, fmode_t mode, int for_part)
 		WARN_ON_ONCE(bdev->bd_holders);
 		sync_blockdev(bdev);
 		kill_bdev(bdev);
+
+		bdev_write_inode(bdev);
 		/*
-		 * ->release can cause the queue to disappear, so flush all
-		 * dirty data before.
+		 * Detaching bdev inode from its wb in __destroy_inode()
+		 * is too late: the queue which embeds its bdi (along with
+		 * root wb) can be gone as soon as we put_disk() below.
 		 */
-		bdev_write_inode(bdev);
+		inode_detach_wb(bdev->bd_inode);
 	}
 	if (bdev->bd_contains == bdev) {
 		if (disk->fops->release)
Signature · 1 line
2.4.3
Acked-by on v1 1/1
#1 lore raw
Quoted 48 lines · Ilya Dryomov

Since 52ebea749aae ("writeback: make backing_dev_info host cgroup-specific bdi_writebacks") inode, at some point in its lifetime, gets attached to a wb (struct bdi_writeback). Detaching happens on evict, in inode_detach_wb() called from __destroy_inode(), and involves updating wb.

However, detaching an internal bdev inode from its wb in __destroy_inode() is too late. Its bdi and by extension root wb are embedded into struct request_queue, which has different lifetime rules and can be freed long before the final bdput() is called (can be from __fput() of a corresponding /dev inode, through dput() - evict() - bd_forget(). bdevs hold onto the underlying disk/queue pair only while opened; as soon as bdev is closed all bets are off. In fact, disk/queue can be gone before __blkdev_put() even returns:

1499 static void __blkdev_put(struct block_device *bdev, fmode_t mode, int for_part) 1500 {
...
1518 if (bdev->bd_contains == bdev) {
1519 if (disk->fops->release)
1520 disk->fops->release(disk, mode);

[ Driver puts its references to disk/queue ]

1521 }
1522 if (!bdev->bd_openers) {
1523 struct module *owner = disk->fops->owner;
1524

1525                 disk_put_part(bdev->bd_part);
1526                 bdev->bd_part = NULL;
1527                 bdev->bd_disk = NULL;
1528                 if (bdev != bdev->bd_contains)

1529 victim = bdev->bd_contains;
1530 bdev->bd_contains = NULL;
1531
1532 put_disk(disk);

[ We put ours, the queue is gone
The last bdput() would result in a write to invalid memory ]

1533 module_put(owner);
...
1539 }

Since bdev inodes are special anyway, detach them in __blkdev_put() after clearing inode's dirty bits, turning the problematic inode_detach_wb() in __destroy_inode() into a noop.

Quoted 6 lines

add_disk() grabs its disk->queue since 523e1d399ce0 ("block: make gendisk hold a reference to its queue"), so the old ->release comment is removed in favor of the new inode_detach_wb() comment.

Cc [email protected] # 4.2+, needs backporting

Thanks.

Signature · 1 line
tejun
#2 lore raw
In reply to Ilya Dryomov

The right fix is to never point from the inode to a BDI. I fixed this with a lot of effort, and the BDI writeback series put it back a little later. We need to revert that damage (not neseccarily literally, but the effect).

#3 lore raw

Hello, Christoph.

Quoted 4 lines · Christoph Hellwig

The right fix is to never point from the inode to a BDI. I fixed this with a lot of effort, and the BDI writeback series put it back a little later. We need to revert that damage (not neseccarily literally, but the effect).

Can you please explain a bit more why inode pointing to its associated bdi_writeback is a bad idea? The only alternative would be recording a key and looking up each time. We sure can do that but I don't get why we would want to.

Thanks.

Signature · 1 line
tejun
#4 lore raw
Quoted 8 lines · Tejun Heo

Hello, Christoph.

Quoted 4 lines · Christoph Hellwig

The right fix is to never point from the inode to a BDI. I fixed this with a lot of effort, and the BDI writeback series put it back a little later. We need to revert that damage (not neseccarily literally, but the effect).

Quoted 4 lines

Can you please explain a bit more why inode pointing to its associated bdi_writeback is a bad idea? The only alternative would be recording a key and looking up each time. We sure can do that but I don't get why we would want to.

Because the writeback context really is a per-super block concept (except for the magic block device inodes). Having another pointer pointer in the inode (or address_space back then) lead to all kinds of confusing bugs and life time issues, nevermind massively bloating the inode for no reason. But then again bloating the inode has been rather en vogue lately anyway.

#5 lore raw

Hello, Christoph.

Quoted 7 lines · Christoph Hellwig
Quoted 4 lines

Can you please explain a bit more why inode pointing to its associated bdi_writeback is a bad idea? The only alternative would be recording a key and looking up each time. We sure can do that but I don't get why we would want to.

Because the writeback context really is a per-super block concept (except for the magic block device inodes). Having another pointer

With cgroup writeback, inodes belonging to the same superblock can belong to different writeback domains, so it no longer is strictly per-super block concept.

Quoted 2 lines

pointer in the inode (or address_space back then) lead to all kinds of confusing bugs and life time issues, nevermind massively bloating the

The bdev case is different because a bdev's lifetime doesn't coincide with the actual object it's representing but the root cause there is lower layer objects being destroyed while higher layer objects depending on them are still around. Severing inode -> lower layer object dependency may be able to solve some issues but those approaches are usually pretty brittle. The right thing to do is letting objects getting torn down in order by maintaining proper refs.

Quoted 2 lines

inode for no reason. But then again bloating the inode has been rather en vogue lately anyway.

How is it for no reason when there are clearly purposes for them? If the extra pointer is problematic, we can try to split sb's so that we can encode writeback domain information by pointing to per-writeback-domain split sb's but I'm not sure the added complexity would be justifiable.

Thanks.

Signature · 1 line
tejun
Tested-by on v1 1/1
#6 lore raw
In reply to Ilya Dryomov
Quoted 48 lines · Ilya Dryomov

Since 52ebea749aae ("writeback: make backing_dev_info host cgroup-specific bdi_writebacks") inode, at some point in its lifetime, gets attached to a wb (struct bdi_writeback). Detaching happens on evict, in inode_detach_wb() called from __destroy_inode(), and involves updating wb.

However, detaching an internal bdev inode from its wb in __destroy_inode() is too late. Its bdi and by extension root wb are embedded into struct request_queue, which has different lifetime rules and can be freed long before the final bdput() is called (can be from __fput() of a corresponding /dev inode, through dput() - evict() - bd_forget(). bdevs hold onto the underlying disk/queue pair only while opened; as soon as bdev is closed all bets are off. In fact, disk/queue can be gone before __blkdev_put() even returns:

1499 static void __blkdev_put(struct block_device *bdev, fmode_t mode, int for_part) 1500 {
...
1518 if (bdev->bd_contains == bdev) {
1519 if (disk->fops->release)
1520 disk->fops->release(disk, mode);

[ Driver puts its references to disk/queue ]

1521 }
1522 if (!bdev->bd_openers) {
1523 struct module *owner = disk->fops->owner;
1524

1525                 disk_put_part(bdev->bd_part);
1526                 bdev->bd_part = NULL;
1527                 bdev->bd_disk = NULL;
1528                 if (bdev != bdev->bd_contains)

1529 victim = bdev->bd_contains;
1530 bdev->bd_contains = NULL;
1531
1532 put_disk(disk);

[ We put ours, the queue is gone
The last bdput() would result in a write to invalid memory ]

1533 module_put(owner);
...
1539 }

Since bdev inodes are special anyway, detach them in __blkdev_put() after clearing inode's dirty bits, turning the problematic inode_detach_wb() in __destroy_inode() into a noop.

Quoted 7 lines

add_disk() grabs its disk->queue since 523e1d399ce0 ("block: make gendisk hold a reference to its queue"), so the old ->release comment is removed in favor of the new inode_detach_wb() comment.

Cc [email protected] # 4.2+, needs backporting

---

Feel free to add

I was facing bad memory access problem while creating thousands of containers. With this patch I am able to create more than 10k containers without hitting the problem.
I had reported the problem here: https://lkml.org/lkml/2015/11/19/149

#7 lore raw
Quoted 57 lines · Raghavendra K T
Quoted 48 lines · Ilya Dryomov

Since 52ebea749aae ("writeback: make backing_dev_info host cgroup-specific bdi_writebacks") inode, at some point in its lifetime, gets attached to a wb (struct bdi_writeback). Detaching happens on evict, in inode_detach_wb() called from __destroy_inode(), and involves updating wb.

However, detaching an internal bdev inode from its wb in __destroy_inode() is too late. Its bdi and by extension root wb are embedded into struct request_queue, which has different lifetime rules and can be freed long before the final bdput() is called (can be from __fput() of a corresponding /dev inode, through dput() - evict() - bd_forget(). bdevs hold onto the underlying disk/queue pair only while opened; as soon as bdev is closed all bets are off. In fact, disk/queue can be gone before __blkdev_put() even returns:

1499 static void __blkdev_put(struct block_device *bdev, fmode_t mode, int for_part) 1500 {
...
1518 if (bdev->bd_contains == bdev) {
1519 if (disk->fops->release)
1520 disk->fops->release(disk, mode);

[ Driver puts its references to disk/queue ]

1521 }
1522 if (!bdev->bd_openers) {
1523 struct module *owner = disk->fops->owner;
1524

1525                 disk_put_part(bdev->bd_part);
1526                 bdev->bd_part = NULL;
1527                 bdev->bd_disk = NULL;
1528                 if (bdev != bdev->bd_contains)

1529 victim = bdev->bd_contains;
1530 bdev->bd_contains = NULL;
1531
1532 put_disk(disk);

[ We put ours, the queue is gone
The last bdput() would result in a write to invalid memory ]

1533 module_put(owner);
...
1539 }

Since bdev inodes are special anyway, detach them in __blkdev_put() after clearing inode's dirty bits, turning the problematic inode_detach_wb() in __destroy_inode() into a noop.

Quoted 7 lines

add_disk() grabs its disk->queue since 523e1d399ce0 ("block: make gendisk hold a reference to its queue"), so the old ->release comment is removed in favor of the new inode_detach_wb() comment.

Cc [email protected] # 4.2+, needs backporting

---

Quoted 7 lines

Feel free to add

I was facing bad memory access problem while creating thousands of containers. With this patch I am able to create more than 10k containers without hitting the problem.
I had reported the problem here: https://lkml.org/lkml/2015/11/19/149

Great! Christoph's concern is with ->i_wb as a whole, not this particular patch - Al, this one is marked for stable, can we get it merged into -rc4? Or should it go through Jens' tree, as cgroup writeback patches did?

Thanks,

Ilya

#8 lore raw
Quoted 67 lines · Ilya Dryomov
Quoted 57 lines · Raghavendra K T
Quoted 48 lines · Ilya Dryomov

Since 52ebea749aae ("writeback: make backing_dev_info host cgroup-specific bdi_writebacks") inode, at some point in its lifetime, gets attached to a wb (struct bdi_writeback). Detaching happens on evict, in inode_detach_wb() called from __destroy_inode(), and involves updating wb.

However, detaching an internal bdev inode from its wb in __destroy_inode() is too late. Its bdi and by extension root wb are embedded into struct request_queue, which has different lifetime rules and can be freed long before the final bdput() is called (can be from __fput() of a corresponding /dev inode, through dput() - evict() - bd_forget(). bdevs hold onto the underlying disk/queue pair only while opened; as soon as bdev is closed all bets are off. In fact, disk/queue can be gone before __blkdev_put() even returns:

1499 static void __blkdev_put(struct block_device *bdev, fmode_t mode, int for_part) 1500 {
...
1518 if (bdev->bd_contains == bdev) {
1519 if (disk->fops->release)
1520 disk->fops->release(disk, mode);

[ Driver puts its references to disk/queue ]

1521 }
1522 if (!bdev->bd_openers) {
1523 struct module *owner = disk->fops->owner;
1524

1525                 disk_put_part(bdev->bd_part);
1526                 bdev->bd_part = NULL;
1527                 bdev->bd_disk = NULL;
1528                 if (bdev != bdev->bd_contains)

1529 victim = bdev->bd_contains;
1530 bdev->bd_contains = NULL;
1531
1532 put_disk(disk);

[ We put ours, the queue is gone
The last bdput() would result in a write to invalid memory ]

1533 module_put(owner);
...
1539 }

Since bdev inodes are special anyway, detach them in __blkdev_put() after clearing inode's dirty bits, turning the problematic inode_detach_wb() in __destroy_inode() into a noop.

Quoted 7 lines

add_disk() grabs its disk->queue since 523e1d399ce0 ("block: make gendisk hold a reference to its queue"), so the old ->release comment is removed in favor of the new inode_detach_wb() comment.

Cc [email protected] # 4.2+, needs backporting

---

Quoted 7 lines

Feel free to add

I was facing bad memory access problem while creating thousands of containers. With this patch I am able to create more than 10k containers without hitting the problem.
I had reported the problem here: https://lkml.org/lkml/2015/11/19/149

Quoted 4 lines

Great! Christoph's concern is with ->i_wb as a whole, not this particular patch - Al, this one is marked for stable, can we get it merged into -rc4? Or should it go through Jens' tree, as cgroup writeback patches did?

Ping?

Thanks,

Ilya

#9 lore raw
Quoted 68 lines · Ilya Dryomov
Quoted 67 lines · Ilya Dryomov
Quoted 64 lines · Raghavendra K T
Quoted 48 lines · Ilya Dryomov

Since 52ebea749aae ("writeback: make backing_dev_info host cgroup-specific bdi_writebacks") inode, at some point in its lifetime, gets attached to a wb (struct bdi_writeback). Detaching happens on evict, in inode_detach_wb() called from __destroy_inode(), and involves updating wb.

However, detaching an internal bdev inode from its wb in __destroy_inode() is too late. Its bdi and by extension root wb are embedded into struct request_queue, which has different lifetime rules and can be freed long before the final bdput() is called (can be from __fput() of a corresponding /dev inode, through dput() - evict() - bd_forget(). bdevs hold onto the underlying disk/queue pair only while opened; as soon as bdev is closed all bets are off. In fact, disk/queue can be gone before __blkdev_put() even returns:

1499 static void __blkdev_put(struct block_device *bdev, fmode_t mode, int for_part) 1500 {
...
1518 if (bdev->bd_contains == bdev) {
1519 if (disk->fops->release)
1520 disk->fops->release(disk, mode);

[ Driver puts its references to disk/queue ]

1521 }
1522 if (!bdev->bd_openers) {
1523 struct module *owner = disk->fops->owner;
1524

1525                 disk_put_part(bdev->bd_part);
1526                 bdev->bd_part = NULL;
1527                 bdev->bd_disk = NULL;
1528                 if (bdev != bdev->bd_contains)

1529 victim = bdev->bd_contains;
1530 bdev->bd_contains = NULL;
1531
1532 put_disk(disk);

[ We put ours, the queue is gone
The last bdput() would result in a write to invalid memory ]

1533 module_put(owner);
...
1539 }

Since bdev inodes are special anyway, detach them in __blkdev_put() after clearing inode's dirty bits, turning the problematic inode_detach_wb() in __destroy_inode() into a noop.

Quoted 7 lines

add_disk() grabs its disk->queue since 523e1d399ce0 ("block: make gendisk hold a reference to its queue"), so the old ->release comment is removed in favor of the new inode_detach_wb() comment.

Cc [email protected] # 4.2+, needs backporting

---

Feel free to add

I was facing bad memory access problem while creating thousands of containers. With this patch I am able to create more than 10k containers without hitting the problem.
I had reported the problem here: https://lkml.org/lkml/2015/11/19/149

Quoted 6 lines
Quoted 4 lines

Great! Christoph's concern is with ->i_wb as a whole, not this particular patch - Al, this one is marked for stable, can we get it merged into -rc4? Or should it go through Jens' tree, as cgroup writeback patches did?

Ping?

Was holding off, but I think we should get the simple fix in for 4.4. I've applied it for this series.

Signature · 1 line
Jens Axboe
#10 lore raw
Quoted 83 lines · Jens Axboe
Quoted 82 lines · Ilya Dryomov
Quoted 78 lines · Ilya Dryomov
Quoted 61 lines · Raghavendra K T
Quoted 49 lines · Ilya Dryomov

Since 52ebea749aae ("writeback: make backing_dev_info host cgroup-specific bdi_writebacks") inode, at some point in its lifetime, gets attached to a wb (struct bdi_writeback). Detaching happens on evict, in inode_detach_wb() called from __destroy_inode(), and involves updating wb.

However, detaching an internal bdev inode from its wb in __destroy_inode() is too late. Its bdi and by extension root wb are embedded into struct request_queue, which has different lifetime rules and can be freed long before the final bdput() is called (can be from __fput() of a corresponding /dev inode, through dput() - evict() - bd_forget(). bdevs hold onto the underlying disk/queue pair only while opened; as soon as bdev is closed all bets are off. In fact, disk/queue can be gone before __blkdev_put() even returns:

1499 static void __blkdev_put(struct block_device *bdev, fmode_t mode, int for_part)
1500 {
...
1518 if (bdev->bd_contains == bdev) {
1519 if (disk->fops->release)
1520 disk->fops->release(disk, mode);

[ Driver puts its references to disk/queue ]

1521 }
1522 if (!bdev->bd_openers) {
1523 struct module *owner = disk->fops->owner;
1524

1525                 disk_put_part(bdev->bd_part);
1526                 bdev->bd_part = NULL;
1527                 bdev->bd_disk = NULL;
1528                 if (bdev != bdev->bd_contains)

1529 victim = bdev->bd_contains;
1530 bdev->bd_contains = NULL;
1531
1532 put_disk(disk);

[ We put ours, the queue is gone
The last bdput() would result in a write to invalid memory ]

1533 module_put(owner);
...
1539 }

Since bdev inodes are special anyway, detach them in __blkdev_put() after clearing inode's dirty bits, turning the problematic inode_detach_wb() in __destroy_inode() into a noop.

Quoted 7 lines

add_disk() grabs its disk->queue since 523e1d399ce0 ("block: make gendisk hold a reference to its queue"), so the old ->release comment is removed in favor of the new inode_detach_wb() comment.

Cc [email protected] # 4.2+, needs backporting

---

Feel free to add

Quoted 8 lines

I was facing bad memory access problem while creating thousands of containers.
With this patch I am able to create more than 10k containers without hitting
the problem.
I had reported the problem here: https://lkml.org/lkml/2015/11/19/149

Great! Christoph's concern is with ->i_wb as a whole, not this particular patch - Al, this one is marked for stable, can we get it merged into -rc4? Or should it go through Jens' tree, as cgroup writeback patches did?

Quoted 5 lines
Quoted 1 line

Ping?

Was holding off, but I think we should get the simple fix in for 4.4. I've applied it for this series.

I think you missed Raghavendra's Tested-by (I'm looking at
https://git.kernel.org/cgit/linux/kernel/git/axboe/linux-block.git/commit/?h=for-linus&id=f2148d49930497c66d38b13f137c30d78c60da3d).

Thanks,

Ilya

#11 lore raw
Quoted 84 lines · Ilya Dryomov
Quoted 83 lines · Jens Axboe
Quoted 82 lines · Ilya Dryomov
Quoted 78 lines · Ilya Dryomov
Quoted 61 lines · Raghavendra K T
Quoted 49 lines · Ilya Dryomov

Since 52ebea749aae ("writeback: make backing_dev_info host cgroup-specific bdi_writebacks") inode, at some point in its lifetime, gets attached to a wb (struct bdi_writeback). Detaching happens on evict, in inode_detach_wb() called from __destroy_inode(), and involves updating wb.

However, detaching an internal bdev inode from its wb in __destroy_inode() is too late. Its bdi and by extension root wb are embedded into struct request_queue, which has different lifetime rules and can be freed long before the final bdput() is called (can be from __fput() of a corresponding /dev inode, through dput() - evict() - bd_forget(). bdevs hold onto the underlying disk/queue pair only while opened; as soon as bdev is closed all bets are off. In fact, disk/queue can be gone before __blkdev_put() even returns:

1499 static void __blkdev_put(struct block_device *bdev, fmode_t mode, int for_part)
1500 {
...
1518 if (bdev->bd_contains == bdev) {
1519 if (disk->fops->release)
1520 disk->fops->release(disk, mode);

[ Driver puts its references to disk/queue ]

1521 }
1522 if (!bdev->bd_openers) {
1523 struct module *owner = disk->fops->owner;
1524

1525                 disk_put_part(bdev->bd_part);
1526                 bdev->bd_part = NULL;
1527                 bdev->bd_disk = NULL;
1528                 if (bdev != bdev->bd_contains)

1529 victim = bdev->bd_contains;
1530 bdev->bd_contains = NULL;
1531
1532 put_disk(disk);

[ We put ours, the queue is gone
The last bdput() would result in a write to invalid memory ]

1533 module_put(owner);
...
1539 }

Since bdev inodes are special anyway, detach them in __blkdev_put() after clearing inode's dirty bits, turning the problematic inode_detach_wb() in __destroy_inode() into a noop.

Quoted 7 lines

add_disk() grabs its disk->queue since 523e1d399ce0 ("block: make gendisk hold a reference to its queue"), so the old ->release comment is removed in favor of the new inode_detach_wb() comment.

Cc [email protected] # 4.2+, needs backporting

---

Feel free to add

Quoted 8 lines

I was facing bad memory access problem while creating thousands of containers.
With this patch I am able to create more than 10k containers without hitting
the problem.
I had reported the problem here: https://lkml.org/lkml/2015/11/19/149

Great! Christoph's concern is with ->i_wb as a whole, not this particular patch - Al, this one is marked for stable, can we get it merged into -rc4? Or should it go through Jens' tree, as cgroup writeback patches did?

Quoted 8 lines
Quoted 5 lines
Quoted 1 line

Ping?

Was holding off, but I think we should get the simple fix in for 4.4. I've applied it for this series.

I think you missed Raghavendra's Tested-by (I'm looking at
https://git.kernel.org/cgit/linux/kernel/git/axboe/linux-block.git/commit/?h=for-linus&id=f2148d49930497c66d38b13f137c30d78c60da3d).

I did, fixed.

Signature · 1 line
Jens Axboe