On 10/04/2023 22:17, Sagar.Biradar@microchip.com wrote:
On 28/03/2023 22:41, Sagar Biradar wrote:
Fix the IO hang that arises because of MSIx vector not having a mapped online CPU upon receiving completion.
What about if the CPU targeted goes offline while the IO is in-flight?
This patch sets up a reply queue mapping to CPUs based on the IRQ affinity retrieved using pci_irq_get_affinity() API.
blk-mq already does what you want here, including handling for the case I mention above. It maintains a CPU -> HW queue mapping, and using a reply map in the LLD is the old way of doing this.
Could you instead follow the example in commit 664f0dce2058 ("scsi: mpt3sas: Add support for shared host tagset for CPU hotplug"), and expose the HW queues to the upper layer? You can alternatively check the example of any SCSI driver which sets shost->host_tagset for this.
Thanks, John [Sagar Biradar]
***What about if the CPU targeted goes offline while the IO is in-flight? We ran multiple random cases with the IO's running in parallel and disabling load-bearing CPU's. We saw that the load was transferred to the other online CPUs successfully every time. The same was tested at vendor and their customer site - they did not see any issues too.
You need to ensure that all CPUs associated with the HW queue are offline and stay offline until any IO may timeout, which would be 30 seconds according to SCSI sd default timeout. I am not sure if you were doing that exactly.
***blk-mq already does what you want here, including handling for the case I mention above. It maintains a CPU -> HW queue mapping, and using a reply map in the LLD is the old way of doing this. We also tried implementing the blk-mq mechanism in the driver and we saw command timeouts. The firmware has limitation of fixed number of queues per vector and the blk-mq changes would saturate that limit. That answers the possible command timeout.
Also this is EOL product and there will be no firmware code changes. Given this, we have decided to stick to the reply_map mechanism. (https://urldefense.com/v3/__https://storage.microsemi.com/en-us/support/seri... )
Thank you for your review comments and we hope you will reconsider the original patch.
I've been checking the driver a bit more and this drivers uses some "reserved" commands, right? That would be internal commands which the driver sends to the adapter which does not have a scsi_cmnd associated. If so, it gets a bit more tricky to use blk-mq support for HW queues, as we need to manually find a HW queue for those "reserved commands", like https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/driv...
Anyway, it's not up to me ...
Thanks, John