Re: [PATCH] btrfs: free device if we fail to open it

On Thu, Dec 02, 2021 at 03:09:38PM +0800, Anand Jain wrote:

On 02/12/2021 05:18, Josef Bacik wrote:

quoted

We've been having transient failures of btrfs/197 because sometimes we
don't show a missing device.

quoted

This turned out to be because I use LVM for my devices, and sometimes we
race with udev and get the "/dev/dm-#" device registered as a device in
the fs_devices list instead of "/dev/mapper/vg0-lv#" device.
 Thus when
the test adds a device to the existing mount it doesn't find the old
device in the original fs_devices list and remove it properly.

I think most of your confusion is because you don't know what btrfs/197 does, so
I'll explain that and then answer your questions below.

DEV=/dev/vg0/lv0
RAID_DEVS=/dev/vg0/lv1 /dev/vg0/lv2 /dev/vg0/vl3 /dev/vg0/lv4

# First we create a single fs and mount it
mkfs.btrfs -f $DEV
mount $DEV /mnt/test

# Now we create the RAID fs
mkfs.btrfs -f -d raid10 -m raid10 $RAID_DEVS

# Now we add one of the raid devs to the single mount above
btrfs device add /dev/vg0/lv2 /mnt/test

# /dev/vg0/lv2 is no longer part of the fs it was made on, it's part of the fs
# that's mounted at /mnt/test

# Mount degraded with the raid setup
mount -o degraded /dev/vg0/lv1 /mnt/scratch

# Now we shouldn't have found /dev/vg0/lv2, because it was wiped and is no
# longer part of the fs_devices for this thing, except it is because it wasn't
# removed, so when we do the following it doesn't show as missing
btrfs filesystem show /mnt/scratch

 The above para is confusing. It can go. IMHO. The device path shouldn't
 matter as we depend on the bdev to compare in the device add thread.

2637         bdev = blkdev_get_by_path(device_path, FMODE_WRITE |
FMODE_EXCL,
2638                                   fs_info->bdev_holder);
::
2657         list_for_each_entry_rcu(device, &fs_devices->devices, dev_list)
{
2658                 if (device->bdev == bdev) {
2659                         ret = -EEXIST;
2660                         rcu_read_unlock();
2661                         goto error;
2662                 }
2663         }

This is on the init thread, this is just checking the fs_devices of /mnt/test,
not the fs_devices of the RAID setup that we created, so this doesn't error out
(nor should it) because we're adding it to our mounted fs.

quoted

This is fine in general, because when we open the devices we check the
UUID, and properly skip using the device that we added to the other file
system.  However we do not remove it from our fs_devices,

Hmm, context/thread is missing. Like, is it during device add or during
mkfs/dev-scan?

AFAIK btrfs_free_stale_devices() checks and handles the removing of stale
devices in the fs_devices in both the contexts/threads device-add, mkfs
(device-scan).

 For example:

$ alias cnt_free_stale_devices="bpftrace -e 'kprobe:btrfs_free_stale_devices
{ @ = count(); }'"

New FSID on 2 devices, we call free_stale_devices():

$ cnt_free_stale_devices -c 'mkfs.btrfs -fq -draid1 -mraid1 /dev/vg/scratch0
/dev/vg/scratch1'
Attaching 1 probe...

@: 2

 We do it only when there is a new fsid/device added to the fs_devices.

For example:

Clean up the fs_devices:
$ cnt_free_stale_devices -c 'btrfs dev scan --forget'
Attaching 1 probe...

@: 1

Now mounting devices are new to the fs_devices list, so call
free_stale_devices().

$ cnt_free_stale_devices -c 'mount -o device=/dev/vg/scratch0
/dev/vg/scratch1 /btrfs'
Attaching 1 probe...

@: 2

$ umount /btrfs

Below we didn't call free_stale_devices() because these two devices are
already in the appropriate fs_devices.

$ cnt_free_stale_devices -c 'mount -o device=/dev/vg/scratch0
/dev/vg/scratch1 /btrfs'
Attaching 1 probe...

@: 0

$

To me, it looks to be working correctly.

Yes it does work correctly, most of the time.  If you run it in a loop 500 times
it'll fail, because _sometimes_ udev goes in and does teh btrfs device scan and
changes the name of the device in the fs_devices for the RAID group.  So the
btrfs_free_stale_devices() thing doesn't find the exising device, because it's
just looking at the device->name, which is different from the device we're
adding.

We have the fs_devices for the RAID fs, and instead of /dev/vg0/lv2, we have
/dev/dm-4 or whatever.  So we do the addition of /dev/vg0/lv2, go to find it in
any other fs_devices, and don't find it because strcmp("/dev/vg0/lv2",
"/dev/dm0-4") != 0, and thus leave the device on the fs_devices for the RAID
file system.

quoted

so when we get
the device info we still see this old device and think that everything
is ok.

quoted

We have a check for -ENODATA coming back from reading the super off of
the device to catch the wipefs case, but we don't catch literally any
other error where the device is no longer valid and thus do not remove
the device.

quoted

Fix this to not special case an empty device when reading the super
block, and simply remove any device we are unable to open.

With this fix we properly print out missing devices in the test case,
and after 500 iterations I'm no longer able to reproduce the problem,
whereas I could usually reproduce within 200 iterations.

commit 7f551d969037 ("btrfs: free alien device after device add")
fixed the case we weren't freeing the stale device in the device-add
context.

However, here I don't understand the thread/context we are missing to free
the stale device here.

quoted

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
---
  fs/btrfs/disk-io.c | 2 +-
  fs/btrfs/volumes.c | 2 +-
  2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
index 5c598e124c25..fa34b8807f8d 100644
--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c

@@ -3924,7 +3924,7 @@ struct btrfs_super_block *btrfs_read_dev_one_super(struct block_device *bdev,
  	super = page_address(page);
  	if (btrfs_super_magic(super) != BTRFS_MAGIC) {
  		btrfs_release_disk_super(super);
-		return ERR_PTR(-ENODATA);
+		return ERR_PTR(-EINVAL);
  	}

 I think you are removing ENODATA because it is going away in the parent
function. And, we don't reach this condition in the test case btrfs/197.
 Instead, we fail here at line 645 below and, we return -EINVAL;

I'm changing it to be consistent, instead of special casing this one specific
failure, just treat all failures like we need to remove the device, and thus we
can just make this be EINVAL as well.

 645         if (memcmp(device->uuid, disk_super->dev_item.uuid,
BTRFS_UUID_SIZE))
 646                 goto error_free_page;
 647

 687 error_free_page:
 688         btrfs_release_disk_super(disk_super);
 689         blkdev_put(bdev, flags);
 690
 691         return -EINVAL;

function stack carries the return code to the parent open_fs_devices().

open_fs_devices()
 btrfs_open_one_device()
  btrfs_get_bdev_and_sb()
   btrfs_read_dev_super()
    btrfs_read_dev_one_super()

quoted

  	if (btrfs_super_bytenr(super) != bytenr_orig) {

diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
index f38c230111be..890153dd2a2b 100644
--- a/fs/btrfs/volumes.c
+++ b/fs/btrfs/volumes.c

@@ -1223,7 +1223,7 @@ static int open_fs_devices(struct btrfs_fs_devices *fs_devices,
  		if (ret == 0 &&
  		    (!latest_dev || device->generation > latest_dev->generation)) {
  			latest_dev = device;
-		} else if (ret == -ENODATA) {
+		} else if (ret) {
  			fs_devices->num_devices--;
  			list_del(&device->dev_list);
  			btrfs_free_device(device);

There are many other cases, for example including -EACCES (from
blkdev_get_by_path()) (I haven't checked other error codes). For which,
calling btrfs_free_device() is inappropriate.

Yeah I was a little fuzzy on this.  I think *any* failure should mean that we
remove the device from the fs_devices tho right?  So that we show we're missing
a device, since we can't actually access it?  I'm actually asking, because I
think we can go either way, but to me I think any failure sure result in the
removal of the device so we can re-scan the correct one.  Thanks,

Josef