On 2022/9/15 0:28, Darrick J. Wong wrote:
On Wed, Sep 14, 2022 at 08:34:26AM -0400, Brian Foster wrote:
On Wed, Sep 14, 2022 at 05:38:02PM +0800, Yang, Xiao/杨 晓 wrote:
On 2022/9/14 14:44, Yang, Xiao/杨 晓 wrote:
On 2022/9/9 21:01, Brian Foster wrote:
Yes.. I don't recall all the internals of the tools and test, but IIRC
it relied on discard to perform zeroing between checkpoints or some such
and avoid spurious failures. The purpose of running on dm-thin was
merely to provide reliable discard zeroing behavior on the target device
and thus to allow the test to run reliably.
Hi Brian,
As far as I know, generic/470 was original designed to verify
mmap(MAP_SYNC) on the dm-log-writes device enabling DAX. Due to the
reason, we need to ensure that all underlying devices under
dm-log-writes device support DAX. However dm-thin device never supports
DAX so
running generic/470 with dm-thin device always returns "not run".
Please see the difference between old and new logic:
old logic new logic
---------------------------------------------------------------
log-writes device(DAX) log-writes device(DAX)
| |
PMEM0(DAX) + PMEM1(DAX) Thin device(non-DAX) + PMEM1(DAX)
|
PMEM0(DAX)
---------------------------------------------------------------
We think dm-thin device is not a good solution for generic/470, is there
any other solution to support both discard zero and DAX?
Hi Brian,
I have sent a patch[1] to revert your fix because I think it's not good for
generic/470 to use thin volume as my revert patch[1] describes:
[1] https://lore.kernel.org/fstests/20220914090625.32207-1-yangx.jy@xxxxxxxxxxx/T/#u
I think the history here is that generic/482 was changed over first in
commit 65cc9a235919 ("generic/482: use thin volume as data device"), and
then sometime later we realized generic/455,457,470 had the same general
flaw and were switched over. The dm/dax compatibility thing was probably
just an oversight, but I am a little curious about that because it should
It's not an oversight -- it used to work (albeit with EXPERIMENTAL
tags), and now we've broken it on fsdax as the pmem/blockdev divorce
progresses.
Hi
Do you mean that the following patch set changed the test result of
generic/470 with thin-volume? (pass => not run/failure)
https://lore.kernel.org/linux-xfs/20211129102203.2243509-1-hch@xxxxxx/
have been obvious that the change caused the test to no longer run. Did
something change after that to trigger that change in behavior?
With the revert, generic/470 can always run successfully on my environment
so I wonder how to reproduce the out-of-order replay issue on XFS v5
filesystem?
I don't quite recall the characteristics of the failures beyond that we
were seeing spurious test failures with generic/482 that were due to
essentially putting the fs/log back in time in a way that wasn't quite
accurate due to the clearing by the logwrites tool not taking place. If
you wanted to reproduce in order to revisit that, perhaps start with
generic/482 and let it run in a loop for a while and see if it
eventually triggers a failure/corruption..?
PS: I want to reproduce the issue and try to find a better solution to fix
it.
It's been a while since I looked at any of this tooling to semi-grok how
it works.
I /think/ this was the crux of the problem, back in 2019?
https://lore.kernel.org/fstests/20190227061529.GF16436@dastard/
Agreed.
Perhaps it could learn to rely on something more explicit like
zero range (instead of discard?) or fall back to manual zeroing?
AFAICT src/log-writes/ actually /can/ do zeroing, but (a) it probably
ought to be adapted to call BLKZEROOUT and (b) in the worst case it
writes zeroes to the entire device, which is/can be slow.
For a (crass) example, one of my cloudy test VMs uses 34GB partitions,
and for cost optimization purposes we're only "paying" for the cheapest
tier. Weirdly that maps to an upper limit of 6500 write iops and
48MB/s(!) but that would take about 20 minutes to zero the entire
device if the dm-thin hack wasn't in place. Frustratingly, it doesn't
support discard or write-zeroes.
Do you mean that discard zero(BLKDISCARD) is faster than both fill
zero(BLKZEROOUT) and write zero on user space?
Best Regards,
Xiao Yang
If the
eventual solution is simple and low enough overhead, it might make some
sense to replace the dmthin hack across the set of tests mentioned
above.
That said, for a *pmem* test you'd expect it to be faster than that...
--D
Brian
Best Regards,
Xiao Yang
BTW, only log-writes, stripe and linear support DAX for now.
