| Message ID | 20230712125455.1986455-2-ming.lei@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | blk-mq: fix wrong queue mapping for kdump kernel | expand |
On Wed, Jul 12, 2023 at 08:54:48PM +0800, Ming Lei wrote: > +/* Max nr_hw_queues for each hw queue type */ > +unsigned int blk_mq_max_nr_hw_queues(void) > +{ > + if (is_kdump_kernel()) > + return 1; > + return nr_cpu_ids; Again, these is_kdump_kernel hacks don't make any sense. The amount of maximum available CPU needs to come through a proper API, and we need to use it, not add hacks like this. The only thing that makes sense here is to find the last CPU in cpu_possible_mask, and for kdump kernels to ensure that number is 1 or whatever low value they want.
On Wed, Jul 12, 2023 at 03:00:17PM +0200, Christoph Hellwig wrote: > On Wed, Jul 12, 2023 at 08:54:48PM +0800, Ming Lei wrote: > > +/* Max nr_hw_queues for each hw queue type */ > > +unsigned int blk_mq_max_nr_hw_queues(void) > > +{ > > + if (is_kdump_kernel()) > > + return 1; > > + return nr_cpu_ids; > > Again, these is_kdump_kernel hacks don't make any sense. The amount > of maximum available CPU needs to come through a proper API, and we > need to use it, not add hacks like this. > > The only thing that makes sense here is to find the last CPU > in cpu_possible_mask, and for kdump kernels to ensure that number > is 1 or whatever low value they want. It doesn't matter how many cpus are available, given at least one cpu is online. The problem is that blk_mq_alloc_tag_set() forces to set nr_hw_queues as 1 for kdump kernel, that is why blk_mq_max_nr_hw_queues() has to return 1 for kdump kernel. We have to tell driver that blk-mq only supports 1 queue for kdump kernel. Or: Thomas, can we disable managed irq for kdump kernel and switch to non-managed irq? Then we can avoid driver's change. I'd suggest this way if it is possible. Thanks, Ming
On Wed, Jul 12, 2023 at 09:16:11PM +0800, Ming Lei wrote: > The problem is that blk_mq_alloc_tag_set() forces to set nr_hw_queues > as 1 for kdump kernel, that is why blk_mq_max_nr_hw_queues() has to > return 1 for kdump kernel. Well, let's fix that first and work from there. Same argument against that deep magic applies there as well. > Thomas, can we disable managed irq for kdump kernel and switch to > non-managed irq? Then we can avoid driver's change. I'd suggest > this way if it is possible. Why the heck would we?
On Wed, Jul 12, 2023 at 03:19:25PM +0200, Christoph Hellwig wrote: > On Wed, Jul 12, 2023 at 09:16:11PM +0800, Ming Lei wrote: > > The problem is that blk_mq_alloc_tag_set() forces to set nr_hw_queues > > as 1 for kdump kernel, that is why blk_mq_max_nr_hw_queues() has to > > return 1 for kdump kernel. > > Well, let's fix that first and work from there. Same argument against > that deep magic applies there as well. In short, driver needs to figure out nr_hw_queues first by hardware info, then pass it to blk_mq_alloc_tag_set(), but blk_mq_alloc_tag_set() changes it, so inconsistency is caused. The only solution in this way is to tell driver the max supported number from the beginning, that is what this patchset is doing. > > > Thomas, can we disable managed irq for kdump kernel and switch to > > non-managed irq? Then we can avoid driver's change. I'd suggest > > this way if it is possible. > > Why the heck would we? IMO irq kernel doesn't make sense in kdump kernel, which is very resource limited and has to be reliable. PCI_IRQ_AFFINITY can be just one hint, pci_alloc_irq_vectors_affinity() still allocates affinity in managed way, then queue mapping can work just fine, and the only difference is that genirq handles this irqs as non-manged wrt. migration. This way should solve queue mapping issue, but driver still allocates lots of queues, which take resource useless. So looks we still have to fix drivers. Thanks, Ming
diff --git a/block/blk-mq.c b/block/blk-mq.c index 5504719b970d..b764da69a416 100644 --- a/block/blk-mq.c +++ b/block/blk-mq.c @@ -140,6 +140,15 @@ void blk_mq_freeze_queue_wait(struct request_queue *q) } EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait); +/* Max nr_hw_queues for each hw queue type */ +unsigned int blk_mq_max_nr_hw_queues(void) +{ + if (is_kdump_kernel()) + return 1; + return nr_cpu_ids; +} +EXPORT_SYMBOL_GPL(blk_mq_max_nr_hw_queues); + int blk_mq_freeze_queue_wait_timeout(struct request_queue *q, unsigned long timeout) { diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h index 2b7fb8e87793..2407978fbc30 100644 --- a/include/linux/blk-mq.h +++ b/include/linux/blk-mq.h @@ -713,6 +713,7 @@ int blk_mq_alloc_sq_tag_set(struct blk_mq_tag_set *set, const struct blk_mq_ops *ops, unsigned int queue_depth, unsigned int set_flags); void blk_mq_free_tag_set(struct blk_mq_tag_set *set); +unsigned int blk_mq_max_nr_hw_queues(void); void blk_mq_free_request(struct request *rq); int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,
blk_mq_alloc_tag_set() may override set->nr_hw_queues as 1 in case of kdump kernel. This way causes trouble for driver, because blk-mq and driver see different queue mapping. Especially the only online CPU may not be 1 for kdump kernel, in which 'maxcpus=1' is passed from kernel command line, then driver may map hctx0 into one inactive real hw queue which cpu affinity is 0(offline). The issue exists on all drivers which use managed irq and support multiple hw queue. Prepare for fixing this kind of issue by applying the added helper, so driver can take blk-mq max nr_hw_queues knowledge into account when calculating io queues. Signed-off-by: Ming Lei <ming.lei@redhat.com> --- block/blk-mq.c | 9 +++++++++ include/linux/blk-mq.h | 1 + 2 files changed, 10 insertions(+)