On 12/16/24 12:23 PM, Damien Le Moal wrote:
On 2024/12/16 11:24, Bart Van Assche wrote:
If 'qd=1' is changed into 'qd=2' in tests/zbd/012 then this test fails
against all kernel versions I tried, including kernel version 6.9. Do
you agree that this test should pass? If you agree with this, do you
agree that the only solution is to postpone error handling of zoned
writes until all pending zoned writes have completed and only to
resubmit failed writes after all pending writes have completed?
Well, of course: if one write fails, the target zone write pointer will not
advance as it should have, so all writes for the same zone after the failed one
will be unaligned and fail. Is that what you are talking about ?
With the fixes applied to rc3, the automatic error recovery in the block layer
is gone. So it is up to the user (FS, DM or application) to do the right thing.
Hi Damien,
For non-zoned storage the BLK_STS_RESOURCE status is not reported to the
I/O submitter (filesystem). The BLK_STS_RESOURCE status causes the block
layer to retry a request. For zoned storage if the block driver reports
the BLK_STS_RESOURCE status and if QD > 1 then the submitter
(filesystem) has to retry the I/O. Isn't that inconsistent? Solving this
inconsistency is one reason why I would like to postpone handling of
zoned write errors until all pending I/O has either completed or failed.
Another reason is that this behavior change is an essential step towards
supporting write pipelining. If multiple zoned writes are outstanding,
and the block driver postpones execution of any of these writes (unit
attention, BLK_STS_RESOURCE, ...) then any zoned writes must only be
resubmitted after all pending zoned writes have either completed or failed.
Thanks,
Bart.
