On Tue, Aug 30, 2016 at 5:13 PM, Chris Murphy [off-list ref] wrote:

On Tue, Aug 30, 2016 at 4:22 AM, ojab // [off-list ref] wrote:

quoted

On Mon, Aug 29, 2016 at 9:05 PM, Chris Murphy [off-list ref] wrote:

quoted

On Mon, Aug 29, 2016 at 10:04 AM, ojab // [off-list ref] wrote:
What do you get for 'btrfs fi us <mp>'

$ sudo btrfs fi us /mnt/xxx/
Overall:
    Device size:                  3.64TiB
    Device allocated:             1.82TiB
    Device unallocated:           1.82TiB
    Device missing:                 0.00B
    Used:                     1.81TiB
    Free (estimated):             1.83TiB         (min: 943.55GiB)
    Data ratio:                  1.00
    Metadata ratio:                  2.00
    Global reserve:             512.00MiB         (used: 0.00B)

Data,RAID0: Size:1.81TiB, Used:1.80TiB
   /dev/sdb1        928.48GiB
   /dev/sdc1        928.48GiB

Metadata,RAID1: Size:3.00GiB, Used:2.15GiB
   /dev/sdb1          3.00GiB
   /dev/sdc1          3.00GiB

System,RAID1: Size:32.00MiB, Used:176.00KiB
   /dev/sdb1         32.00MiB
   /dev/sdc1         32.00MiB

Unallocated:
   /dev/sdb1          1.01MiB
   /dev/sdc1          1.00MiB
   /dev/sdd1          1.82TiB

The confusion is understandable because sdd1 is bigger than sdc1, so
why can't everything on sdc1 be moved to sdd1? Well, dev add > dev del
doesn't really do that, it's going to end up rewriting metadata to
sdb1 also, and there isn't enough space. Yes, there's 800MiB of unused
space in metadata chunks on sdb1 and sdc1, it should be enough (?) but
clearly it wants more than this for whatever reason. You could argue
it's a bug or some suboptimal behavior, but because this is a 99% full
file system, I'm willing to be it's a low priority bug. Because this
is raid0 you really need to add two devices, not just one.

quoted

I don't quite understand what exactly btrfs is trying to do: I assume
that block groups should be relocated to the new/empty drive,

There is a scant chance 'btrfs replace' will work better here. But
still the real problem remains, even if you replace sdc1 with sdd1,
sdb1 is still 99% full which in effect makes the file system 99% full
because it can't do anymore raid0 on sdb1, and it's not possible to do
raid0 chunks on a single sdd1 device.

If you can't add a 4th drive, you're going to have to convert to
single profile. Keep all three drives attached, 'btrfs balance start
-dconvert=single' and then once that's complete you should be able to
remove /dev/sdc1, although this will take a while because first
conversion will use space on all three drives, and then the removable
of sdc1 will have to copy chunks off before it can be removed.

quoted

but
during the delete `btrfs fi us` shows
Unallocated:
/dev/sdc1         16.00EiB

Known bug, also happens when resizing and conversions.

quoted

so deleted partition is counted as maximum possible empty drive and
blocks are relocated to it instead of new/empty drive? (kernel-4.7.2 &
btrfs-progs-4.7.1 here)
Is there any way to see where and why block groups are relocated
during `delete`?

The two reasons this isn't working is a.) it's 99% full already and
b.) it's raid0, so merely adding one device isn't sufficient. It's
probably too full even to do a 3 device balance to restripe raid0
across 3 devices, which is still inefficient because it would leave
50% of the space on sdd as unusable. To do this with uneven devices
and use all the space, you're going to have to use single profile.



--
Chris Murphy

Ah, thanks for the elaboration, it makes things much more meaningful now!

//wbr ojab