On 2/19/19 3:19 AM, Ming Lei wrote:
On Tue, Feb 5, 2019 at 11:30 PM Hannes Reinecke <hare@xxxxxxx> wrote:
Hi all,
this came up during discussion on the mailing list (cf thread "Question
on handling managed IRQs when hotplugging CPUs").
The problem is that with managed IRQs and block-mq I/O will be routed to
individual CPUs, and the response will be send to the IRQ assigned to
that CPU.
If now a CPU hotplug event occurs when I/O is still in-flight the IRQ
will _still_ be assigned to the CPU, causing any pending interrupt to be
lost.
Hence the driver will never notice that an interrupt has happened, and
an I/O timeout occurs.
Lots of driver's timeout handler only returns BLK_EH_RESET_TIMER,
and this situation can't be covered by IO timeout for these devices.
For example, we have see IO hang issue on HPSA, megaraid_sas
before when wrong msi vector is set on IO command. Even one such
issue on aacraid isn't fixed yet.
Precisely.
One proposal was to quiesce the device when a CPU hotplug event occurs,
and only allow for CPU hotplugging once it's fully quiesced.
That is the original solution, but big problem is that queue dependency
exists, such as loop/DM's queue depends on underlying's queue, NVMe
IO queue depends on its admin queue.
Well, obviously we would have to wait for _all_ queues to be quiesced.
And for stacked devices we will need to take the I/O stack into account,
true.
While this would be working, it will be introducing quite some system
stall, and it actually a rather big impact in the system.
Another possiblity would be to have the driver abort the requests
itself, but this requires specific callbacks into the driver, and, of
course, the driver having the ability to actually do so.
I would like to discuss at LSF/MM how these issues can be addressed best.
One related topic is that the current static queue mapping without CPU hotplug
handler involved may waste lots of IRQ vectors[1], and how to deal
with this problem?
[1] http://lists.infradead.org/pipermail/linux-nvme/2019-January/021961.html
Good point. Let's do it.
Cheers,
Hannes
