Re: [PATCH v4 3/3] btrfs: free alien device due to device add

[David Sterba <dsterba@suse.cz>]

On Tue, May 05, 2020 at 02:58:26AM +0800, Anand Jain wrote:
> When the old device has new fsid through btrfs device add -f <dev> our
> fs_devices list has an alien device in one of the fs_devices. So this is
> a trigger and not the root cause of the issue. This patch fixes the trigger.
> 
> By having an alien device in fs_devices, we have two issues so far
> 
> 1. missing device is not shows as missing in the userland
> 
> Which is due to cracks in the function btrfs_open_one_device() and
> hardened by the pending patches in the ml.
>  btrfs: remove identified alien device in open_fs_devices
>  btrfs: remove identified alien btrfs device in open_fs_devices

Such notes do not belong to changelog, the subjects don't say enough
about the problem or the fix.

> 2. mount of a degraded fs_devices fails
> 
> Which is due to cracks in the function btrfs_free_extra_devids() and
> hardened by patch (included in the set).
>  btrfs: include non-missing as a qualifier for the latest_bdev

Same.

> The trigger for both of this issue is that there is an alien (does not
> contain the intended fsid/btrfs_magic) device in the fs_devices.
> 
> We know a device can be scanned/added through
> btrfs-control::BTRFS_IOC_SCAN_DEV|BTRFS_IOC_DEVICES_READY
> or by
> ioctl::BTRFS_IOC_ADD_DEV
> 
> And device coming through btrfs-control is checked against the all other
> devices in btrfs kernel but not coming through BTRFS_IOC_ADD_DEV.
> 
> This patch checks if the device add is alienating any other scanned
> device and deletes it.
> 
> In fact, this patch fixes both the issues 1 and 2 (above) by eliminating
> the trigger of the issue, but still they have their own patch as well
> because its the right way to harden the functions and fill the cracks.
> 
> Signed-off-by: Anand Jain <anand.jain@oracle.com>
> Reviewed-by: Josef Bacik <josef@toxicpanda.com>

So I'm not sure how much the review holds when the code changes between
iterations. The right way is to drop the rev-by and CC the person for
any non-trivial or maybe for all changes.

I also rewrote the changelog as it's sometimes hard to understand what
you mean, I had most problems with the use of the term 'trigger'.

> --- a/fs/btrfs/volumes.c
> +++ b/fs/btrfs/volumes.c
> @@ -2664,6 +2664,15 @@ int btrfs_init_new_device(struct btrfs_fs_info *fs_info, const char *device_path
>  
>  	/* Update ctime/mtime for libblkid */
>  	update_dev_time(device_path);
> +
> +	/*
> +	 * Now that we have written a new sb into this device, check all other
> +	 * fs_devices list if device_path alienates any other scanned device.
> +	 * Ignore the return as we are successfull in the core task - add
> +	 * the device.
> +	 */
> +	btrfs_forget_devices(device_path);

Actually this should go before update_dev_time, as this is the point
when the file change could be detected by udev and rescanned, which
increases the window between commit and removal from fs_devices.