Slow discard on RAID1 with --write-mostly: split into 1MB requests

[Roman Mamedov <rm@romanrm.net>]

Hello,

I found that if an mdadm RAID1 is created with a write-mostly device, each
discard request to the array ends up being split into 1MB pieces. My SSD is
capable of processing at most 200 of these per second, so as a result
discarding a 512GB array takes 42 minutes.

The kernel is 6.1.9.

In both cases:

# grep . /sys/block/md0/queue/discard_*
/sys/block/md0/queue/discard_granularity:512
/sys/block/md0/queue/discard_max_bytes:2199023255040
/sys/block/md0/queue/discard_max_hw_bytes:2199023255040
/sys/block/md0/queue/discard_zeroes_data:0

Steps to reproduce:

# wipefs -a /dev/nvme1n1p2 /dev/nvme2n1p2
/dev/nvme1n1p2: 4 bytes were erased at offset 0x00001000 (linux_raid_member): fc 4e 2b a9
/dev/nvme2n1p2: 4 bytes were erased at offset 0x00001000 (linux_raid_member): fc 4e 2b a9

# mdadm --create --level 1 -n2 --assume-clean --metadata=1.2 /dev/md0 /dev/nvme1n1p2 /dev/nvme2n1p2
mdadm: array /dev/md0 started.

# mdadm --grow --array-size 10G /dev/md0 # (so that the test doesn't take forever)

# time blkdiscard -f /dev/md0

real	0m0.335s
user	0m0.000s
sys	0m0.003s

# mdadm --stop /dev/md0 
mdadm: stopped /dev/md0

# wipefs -a /dev/nvme1n1p2 /dev/nvme2n1p2
/dev/nvme1n1p2: 4 bytes were erased at offset 0x00001000 (linux_raid_member): fc 4e 2b a9
/dev/nvme2n1p2: 4 bytes were erased at offset 0x00001000 (linux_raid_member): fc 4e 2b a9

# mdadm --create --level 1 -n2 --assume-clean --metadata=1.2 /dev/md0 /dev/nvme1n1p2 --write-mostly /dev/nvme2n1p2
mdadm: array /dev/md0 started.

# mdadm --grow --array-size 10G /dev/md0

# time blkdiscard -f /dev/md0

real	0m48.744s
user	0m0.000s
sys	0m0.019s

-- 
With respect,
Roman