[dm-devel] Max length limitation with BLKZEROOUT on dm device ?

[Bharata B Rao]

On Mon, Jul 1, 2013 at 6:24 PM, Mike Snitzer <snitzer at redhat.com> wrote:
> On Mon, Jul 01 2013 at  5:04am -0400,
> Bharata B Rao <bharata.rao at gmail.com> wrote:
>
>> Hi,
>>
>> While experimenting on zeroing out a dm device from a storage that
>> supports WRITE SAME, I see the below warnings from kernel (3.10.0):
>>
>> ******************
>> INFO: task a.out:21365 blocked for more than 120 seconds.
>> "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
>> a.out           D ffffffff8182d620     0 21365  19945 0x00000080
>> ffff882079321bf8 0000000000000086 ffff882079321fd8 0000000000013fc0
>> ffff882079320010 0000000000013fc0 0000000000013fc0 0000000000013fc0
>> ffff882079321fd8 0000000000013fc0 ffff882032a6eb10 ffff8810795ea0c0
>> Call Trace:
>> [<ffffffff816426f9>] schedule+0x29/0x70
>> [<ffffffff81640945>] schedule_timeout+0x195/0x220
>> [<ffffffff812ad777>] ? __blk_run_queue+0x37/0x50
>> [<ffffffff812af53f>] ? blk_queue_bio+0x2ff/0x3b0
>> [<ffffffff81064859>] ? read_tsc+0x9/0x20
>> [<ffffffff810f5ebc>] ? ktime_get_ts+0x4c/0xf0
>> [<ffffffff81641c4f>] io_schedule_timeout+0x9f/0x100
>> [<ffffffff812ae44a>] ? generic_make_request+0xca/0x100
>> [<ffffffff81642aec>] wait_for_completion_io+0xdc/0x110
>> [<ffffffff810d8fb0>] ? try_to_wake_up+0x2b0/0x2b0
>> [<ffffffff812b5c3c>] blkdev_issue_write_same+0x1bc/0x1d0
>> [<ffffffff812b5cfd>] blkdev_issue_zeroout+0xad/0x120
>> [<ffffffff812b6c68>] blkdev_ioctl+0x728/0x7b0
>> [<ffffffff8120296c>] block_ioctl+0x3c/0x40
>> [<ffffffff811ddb89>] do_vfs_ioctl+0x89/0x350
>> [<ffffffff811ddef1>] SyS_ioctl+0xa1/0xb0
>> [<ffffffff8164ca99>] system_call_fastpath+0x16/0x1b
>> ******************
>>
>> My test program looks like this:
>>
>> {
>> unsigned long arg[2] = {0, 21474836480}; /* len = 20G */
>> fd = open("/dev/mapper/ws_vg-lvol0", O_RDWR);
>> ioctl(fd, BLKZEROOUT, &arg);
>> }
>>
>> # ls -l /dev/mapper/ws_vg-lvol0
>> lrwxrwxrwx. 1 root root 8 Jul  1 12:25 /dev/mapper/ws_vg-lvol0 -> ../dm-17
>>
>> # cat /sys/block/dm-17/queue/write_same_max_bytes
>> 268435456
>>
>> I don't see such messages when len is 512M, 1G, 10G or 15G, but with
>> 20G, such messages are seen.
>>
>> Though I don't see any limitation on the number of bytes that could be
>> zeroed out from a single BLKZEROOUT ioctl call, could this be the
>> result of using a len as large as 20G in a single ioctl ?
>
> It is quite likely a function of the storage.  The 20G BLKZEROOUT
> eventually completes right?

Yes, the 20G BLKZEROOUT eventually completes, but shouldn't
blkdev_issue_write_same() limit the number of bio's submitted in a
batch instead of submitting bio's covering the entire 20G range and
then waiting for all of them at once ? Correct me I am wrong, but
that's what I understood from the code. Shouldn't
blkdev_issue_write_same() issue bio's at the rate which the device can
handle ?

>
> It could be that 15G just happens to complete before the 120s threshold,
> but 20G causes it to exceed the 120s hung task timeout?

Yes very likely.

Regards,
Bharata.