On 11/01/2021 11:57, Jinpu Wang wrote:
Hi John, On Tue, Jan 5, 2021 at 12:21 PM John Garry <john.garry@xxxxxxxxxx> wrote:In commit 05c6c029a44d ("scsi: pm80xx: Increase number of supported queues"), support for 80xx chip was improved by enabling multiple HW queues. In this, like other SCSI MQ HBA drivers, the HW queues were not exposed to upper layer, and instead the driver managed the queues internally. However, this management duplicates blk-mq code. In addition, the HW queue management is sub-optimal for a system where the number of CPUs exceeds the HW queues - this is because queues are selected in a round-robin fashion, when it would be better to make adjacent CPUs submit on the same queue. And finally, the affinity of the completion queue interrupts is not set to mirror the cpu<->HQ queue mapping, which is suboptimal. As such, for when MSIX is supported, expose HW queues to upper layer. Flag PCI_IRQ_AFFINITY is set for allocating the MSIX vectors to automatically assign affinity for the completion queue interrupts. Signed-off-by: John Garry <john.garry@xxxxxxxxxx> --- I sent as an RFC/RFT as I have no HW to test. In addition, since HW queue #0 is used always for internal commands (like in send_task_abort()), if all CPUs associated with HW queue #0 are offlined, the interrupt for that queue will be shutdown, and no CPUs would be available to service any internal commands completion. To solve that, we need [0] merged first and switch over to use the new API. But we can still test performance in the meantime. I assume someone else is making the change to use the request tag for IO tag management. [0] https://lore.kernel.org/linux-scsi/47ba045e-a490-198b-1744-529f97192d3b@xxxxxxx/Thanks for the patch, maybe Viswas can help to test?
That's what I am hoping for :) Thanks!
