Re: How to replace a missing device with a smaller one

[Anand Jain <anandsuveer@gmail.com>]

On 11/20/19 8:05 AM, Zygo Blaxell wrote:
> On Wed, Nov 20, 2019 at 08:02:43AM +0800, Qu Wenruo wrote:
>>
>>
>> On 2019/11/19 下午10:38, David Sterba wrote:
>>> On Mon, Nov 18, 2019 at 03:08:00PM +0800, Qu Wenruo wrote:
>>>> On 2019/11/18 下午1:32, Qu Wenruo wrote:
>>>>> On 2019/11/18 上午10:09, Nathan Dehnel wrote:
>>>>>> I have a 10-disk raid10 with a missing device I'm trying to replace. I
>>>>>> get this error when doing it though:
>>>>>>
>>>>>> btrfs replace start 1 /dev/bcache0 /mnt
>>>>>> ERROR: target device smaller than source device (required 1000203091968 bytes)
>>>>>>
>>>>>> I see that people recommend resizing a disk before replacing it, which
>>>>>> isn't an option for me because it's gone.
>>>>>
>>>>> Oh, that's indeed a problem.
>>>>>
>>>>> We should allow to change missing device's size.
>>>>
>>>> I have CCed you with a patch to allow user to *shrink* the missing device.
>>>>
>>>> You can also get the patch from patchwork:
>>>> https://patchwork.kernel.org/patch/11249009/
>>>>
>>>> Please give a try, since the device size is pretty small, I believe with
>>>> that patch, we can go quick shrink, that means "btrfs fi resize" command
>>>> should return immediately.
>>>
>>> So it can be recteated eg. on loop devices, where some of them are
>>> slightly smaller, then go missing and replace is started, right?
>>>
>>
>> Replace will still be rejected, but we can do resize of that missing
>> dev, then replace, as a workaround.
>>
>> I haven't do the auto-resize for replace yet, since I'm not sure if that
>> could make cases like replacing 1T device with 10G driver happening.
> 
> That case would be much more tolerable if device replace/resize/delete would
> stop to check for fatal signals at least between block groups, if not more
> often.  Currently if you pick the wrong size, the only way to make it stop
> is to reboot.

  I agree. IMO too
     btrfs replace --<option to skip-check|auto-resize>
  is better as commented in the patch thread.

  Already if you comment out the size checks[1], the replace with
  a smaller device works fine. But But you have to manually estimate
  the disk size for the actual consumption. So the effort here should
  be auto pre-check.

[1]

diff --git a/cmds/replace.c b/cmds/replace.c
index 2321aa156fe2..1a4155d360d6 100644
--- a/cmds/replace.c
+++ b/cmds/replace.c
@@ -246,11 +246,13 @@ static int cmd_replace_start(const struct 
cmd_struct *cmd,
                 goto leave_with_error;

         dstdev_size = get_partition_size(dstdev);
+/*
         if (srcdev_size > dstdev_size) {
                 error("target device smaller than source device 
(required %llu bytes)",
                         srcdev_size);
                 goto leave_with_error;
         }
+*/

         fddstdev = open(dstdev, O_RDWR);
         if (fddstdev < 0) {


diff --git a/fs/btrfs/dev-replace.c b/fs/btrfs/dev-replace.c
index 9a29d6de9017..86fc0f8653be 100644
--- a/fs/btrfs/dev-replace.c
+++ b/fs/btrfs/dev-replace.c
@@ -214,7 +214,7 @@ static int btrfs_init_dev_replace_tgtdev(struct 
btrfs_fs_info *fs_info,
                 }
         }

-
+/*
         if (i_size_read(bdev->bd_inode) <
             btrfs_device_get_total_bytes(srcdev)) {
                 btrfs_err(fs_info,
@@ -222,7 +222,7 @@ static int btrfs_init_dev_replace_tgtdev(struct 
btrfs_fs_info *fs_info,
                 ret = -EINVAL;
                 goto error;
         }
-
+*/

         device = btrfs_alloc_device(NULL, &devid, NULL);
         if (IS_ERR(device)) {





>> Thanks,
>> Qu
>>
> 
> 
>