On 12/09/2020 12:17 PM, Song Liu wrote:

Hi Matthew,

quoted

On Dec 8, 2020, at 7:46 PM, Matthew Ruffell [off-list ref] wrote:

Hello,

I recently backported the following patches into the Ubuntu stable kernels:

md: add md_submit_discard_bio() for submitting discard bio
md/raid10: extend r10bio devs to raid disks
md/raid10: pull codes that wait for blocked dev into one function
md/raid10: improve raid10 discard request
md/raid10: improve discard request for far layout
dm raid: fix discard limits for raid1 and raid10
dm raid: remove unnecessary discard limits for raid10

Thanks for the report!

Hi Xiao,

Could you please take a look at this and let me know soon? We need to fix
this before 5.10 official release.

Thanks,
Song

Hi all

The root cause is found. Now we use a similar way with raid0 to handle 
discard request
for raid10. Because the discard region is very big, we can calculate the 
start/end address
for each disk. Then we can submit the discard request to each disk. But 
for raid10, it has
copies. For near layout, if the discard request doesn't align with chunk 
size, we calculate
a start_disk_offset. Now we only use start_disk_offset for the first 
disk, but it should be
used for the near copies disks too.

[  789.709501] discard bio start : 70968, size : 191176
[  789.709507] first stripe index 69, start disk index 0, start disk 
offset 70968
[  789.709509] last stripe index 256, end disk index 0, end disk offset 
262144
[  789.709511] disk 0, dev start : 70968, dev end : 262144
[  789.709515] disk 1, dev start : 70656, dev end : 262144

For example, in this test case, it has 2 near copies. The 
start_disk_offset for the first disk is 70968.
It should use the same offset address for second disk. But it uses the 
start address of this chunk.
It discard more region. The patch in the attachment can fix this 
problem. It split the region that
doesn't align with chunk size.

There is another problem. The stripe size should be calculated 
differently for near layout and far layout.

@Song, do you want me to use a separate patch for this fix, or fix this 
in the original patch?

Merry Christmas
Xiao