[dm-devel] [PATCHv2 0/4] blk-mq support for dm multipath

[Keith Busch]

No comments this time. Did I totally botch this in the wrong direction,
or was the timing bad with the linux con/plumbers that week?

On Fri, 17 Oct 2014, Keith Busch wrote:
> This series makes device mapper's multipath work with both blk-mq and
> the older request_queue based block devices.
>
> As before, I tested with a slightly modified version of Matias' blk-mq
> nvme driver with dual ported nvme controller. I'll post the nvme changes
> to that list once the blk-mq mpath is good to go.
>
> I've tried to split this into logically separate patches to prep and
> support this request_queue type.
>
> Patch [2/4] is a bit of a departure from what was done before, but I
> hope something like this is alright.
>
> v1 -> v2:
>
> Micro-optimization setting the cloned rq->bio to null from dm-mpath
> instead of blk_mq_rq_ctx_init.
>
> There's no way around getting a request without calling
> blk_get_request. Making this API safe in an interrupt disabled context is
> troublesome, so submitting the dm request to the target in a different
> thread so it is NOT inline with dm's request_fn. We can use GFP_KERNEL
> now since the allocation occurs in this new context. I believe this is
> okay in both request queue types since neither allocate from general
> purpose memory.
>
> In v1, the target would allocate the request and dm would release it. This
> duality is somewhat changed: the target's map_rq is still responsible
> to allocate since only it knows what request queue to allocate from,
> and I've added a new target type function (unmap_rq) to handle releasing.
>
> I fixed the error handling; it was completely untested before and quite
> broken. To test, I synthesized 4 different errors: no requests available,
> no memory available to the clone bio, invalid target, and completing
> requests in low-level driver with an error. All worked as expected on
> blk-mq devices.
>
> Keith Busch (4):
>  dm: prep initialized requests
>  dm: Submit stacked requests in irq enabled context
>  dm: Move request allocation to dm_target type
>  block: blk-mq support for cloned requests
>
> block/blk-core.c              |    7 +-
> drivers/md/dm-mpath.c         |   22 ++++--
> drivers/md/dm-target.c        |    9 ++-
> drivers/md/dm.c               |  155 +++++++++++++++++++++--------------------
> include/linux/device-mapper.h |    7 +-
> 5 files changed, 112 insertions(+), 88 deletions(-)
>
> -- 
> 1.7.10.4