Hi Sagi,
On Fri, Nov 15, 2019 at 02:38:44PM -0800, Sagi Grimberg wrote:
Hi,
Hey Ming,
Use blk_mq_alloc_request() for allocating NVMe loop, fc, rdma and tcp's
connect request, and selecting transport queue runtime for connect
request.
Then kill blk_mq_alloc_request_hctx().
Is it really so wrong to have an API to allocate a tag that belongs to
a specific queue? Why must the tags allocation always correlate to the
running cpu? Its true that NVMe is the only consumer of this at the
moment, but does this mean that the interface should be removed because
it has one (or rather four) consumer(s)?
Now blk-mq takes a static queue mapping between CPU and hw queues, given
CPU hotplug may happen any time, so the specified hw queue may become
inactive any time.
Queue mapping from CPU to hw queue is one core model of blk-mq which
relies a lot on the fact that hw queue active or not depends on
request's submission CPU. And we always can retrieve one active hw
queue if there is at least one online CPU.
IO request is always mapped to the proper hw queue via the submission
CPU and queue type.
So blk_mq_alloc_request_hctx() is really weird from the above blk-mq's
model.
Actually the 4 consumer is just one single type of usage for submitting
connect command, seems no one explain this special usage before. And the
driver can handle well enough without this interface just like this
patch, can't it?
Does removing the cpumask_and with cpu_online_mask fix your test?
this check is wrong to begin with because it can not be true right after
the check.
This is a much simpler fix that does not create this churn local to
every driver. Also, I don't like the assumptions about tag reservations
that the drivers is taking locally (that the connect will have tag 0
for example). All this makes this look like a hack.
There is also the question of what happens when we want to make connects
parallel, which is not the case at the moment...
I would instead suggest to simply remove the constraint that
blk_mq_alloc_request_hctx() will fail if the first cpu in the mask
is not on the cpu_online_mask.. The caller of this would know and
be able to handle it.
Of course, this usage is very special, which is different with normal
IO or passthrough request.
The caller of this still needs to rely on blk-mq for dispatching this
request:
1) blk-mq needs to run hw queue in round-robin style, and different
CPU is selected from active CPU masks for running the hw queue.
2) Most of blk-mq drivers have switched to managed IRQ, which may be
shutdown even though there is still in-flight requests not completed
on the hw queue. We need to fix this issue. One solution is to free
the request and remap the bios into proper active hw queue according to
the new submission CPU.
3) warning will be caused when dispatching one request on inactive hw queue
If we are going to support this special usage, lots of blk-mq needs to
be changed for covering the so special case.
I'm starting to think we maybe need to get the connect out of the block
layer execution if its such a big problem... Its a real shame if that is
the case...
To me it feels like this approach is fundamentally wrong. IMO, having
the driver select a different queue than the tag naturally belongs to
feels like a backwards design.
The root cause is that the connection command needs to be submitted via
one specified queue, which is fundamentally not compatible with blk-mq's
queue mapping model.
Given the connect command is so special for nvme, I'd suggest to let driver
handle it completely, since blk-mq is supposed to serve generic IO function.
Its one thing to handling it locally, and to hack queue_rq to queue on a
different queue than what was determined by the block layer... If this
model is fundamentally broken with how the block layer dispatch
requests, then we need a different solution.
