Hello,
This is an updated patch set of dm_queue feature posted on June 23, 2006.
It can be applied on top of 2.6.18-rc1-mm1.
Please review and apply.
Description of the patches:
1/4: add dm_queue and generic worker thread
2/4: add noflush feature to dm_queue
3/4: dm-mpath to enable noflush
4/4: add API to libdevmapper
This is the update since the last posting:
o Change the word which describes suspend without queue flush
from "fast-suspend" to "noflush" based on Alasdair's feedback.
(Among all patches)
o noflush support for I/O across some targets.
To implement this, added "pending_tio" to struct mapped_device.
It is the counter for pending splitted bios.
So when the "pending_tio" and the number of queued bios become same,
the mapped_device is quiet. Then, suspend can be done.
(Done in the patch 2/4)
Test results:
Following functional tests are done with/without using 'noflush' feature
in multipath target and found no problem.
o suspend/resume without table swap in no valid path situation.
o path addition and table swap in no valid path situation.
o above testing for I/O across some multipath targets.
I/Os are returned without error when using 'noflush' feature
and returned with error when not using it.
Any comments are welcome.
Thanks,
Kiyoshi Ueda
On Fri, 23 Jun 2006 17:45:38 -0400 (EDT), Kiyoshi Ueda <k-ueda@xxxxxxxxxxxxx> wrote:
> Hello,
>
> I'm working on moving I/O queue of targets to dm core so that
> the table can swap/suspend without the queue flushing.
> -----------------------------------------------------------------------
> BACKGROUND
>
> Currently, table swap is needed when we add additional path to
> existing multipath map. To swap table, suspend is needed and
> all I/Os are flushed at the suspention time.
> If all paths in the multipath map are failed and there are queued
> I/Os in the multipath target, these I/Os will return as error
> when the I/O flushing.
> In such case, user should expect that these I/Os are issued
> without error through the new additional path.
>
> To solve the issue, suspend must be done without queue flushing
> and the queued I/Os must be handed over to the new table when table
> swapping.
>
>
> DESIGN
>
> The structure of the queue and the queue processing kernel thread
> which are used by the multipath target can be generalized so that
> other targets can also use it.
> So the queue and the thread are placed in dm-core, and targets use it
> by using interface functions.
> ------------------------------------------------------------------------
>
> The following patch set is a draft to solve the issue above.
> It is for 2.6.17-rc6-mm1 + Alasdair's patches which was sent
> to this ML (Subject: "Next set of device-mapper patches").
>
> The patches are still rough edges and I'd like to get comments
> from dm developpers on this approach.
> For example,
> 1). Queues for a target are placed in "struct dm_target".
> 2). Unmap target function is added for fast-suspend feature.
> 3). Is the interface sufficient for other target drivers?
> (Current design considers multipath target only.)
>
> Below is TODO list which I'm working on.
> -----------------------------------------------------------------------
> TODO
>
> o If the I/O is splitted across some targets, fast-suspend feature
> doesn't work because the total of queue size doesn't match with
> the number of pending original bios.
>
> o Accounting issue when withdrawing queued bios.
>
> o The dm_queue daemon works for only bio queue.
> Other works should be handled? (e.g. trigger_event of multipath)
> -----------------------------------------------------------------------
>
> Any other comments are also very welcome!
>
> Regards and thanks for your time,
> Kiyoshi Ueda
--
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel
