[dm-devel] Re: [RFC PATCH 1/3] Implement generic freeze feature

[Takashi Sato]

Hi,

>> +    if (test_and_set_bit(BD_FREEZE_OP, &bdev->bd_state))
>> +        return ERR_PTR(-EBUSY);
>> +
>> +    sb = get_super_without_lock(bdev);
>> +
>> +    /* If super_block has been already frozen, return. */
>> +    if (sb && sb->s_frozen != SB_UNFROZEN) {
>> +        put_super(sb);
>> +       clear_bit(BD_FREEZE_OP, &bdev->bd_state);
>> +       return sb;
>> +    }
>
> The BD_FREEZE_OP flag in the block_device already prevents multiple
> freezes for a singe block device, so there is no need for this
> additional check and the get_super_without_lock helper.

I think the above check is needed because BD_FREEZE_OP is used to
guarantee that the following two operations are done atomically.
1. Check if the filesystem has already been frozen.
   (sb->s_frozen != SB_UNFROZEN)
2. Do the filesystem specific freeze operation
   (write_super_lockfs)

The freeze operation can be done again for the frozen filesystem
unless the filesystem is checked.
And if write_super_lockfs doesn't consider that the filesystem
has already been frozen, a problem might occur.

>>  down(&bdev->bd_mount_sem);
> 
> And with that flag bd_mount_sem is also obsolete for preventing the
> multiple freeze operations.  

bd_mount_sem is needed because it is used to guarantee that the freeze
operation never run while mounting (get_sb_bdev()).

> We still need investigate what
> synchronization we need vs unmount which also takes bd_mount_sem without
> every having document what it exactly protects.

And s_umount (semaphore) is acquired in get_user() (called by freeze_bdev())
to guarantee that unmount never run while the filesystem is frozen.

Cheers, Takashi