blk-mq: fix wrong queue mapping for kdump kernel
Subsystems (9)
v2 cover blk-mq: fix wrong queue mapping for kdump kernel
Hi,
On arm and ppc64, 'maxcpus=1' is required for kdump kernel, see Documentation/admin-guide/kdump/kdump.rst, so num_possible_cpus() still returns all CPUs because 'maxcpus=1' just bring up one single cpu core during booting.
blk-mq sees single queue in kdump kernel, and in driver's viewpoint there are still multiple queues, this inconsistency causes driver to apply wrong queue mapping for handling IO, and IO timeout is triggered.
This issue is only triggered on managed irq in case of multiple hw queues. Some drivers takes online cpus into account for nr_hw_queues, and don't have such issue, such as nvme rdma/tcp.
Meantime, single queue makes much less resource utilization, and reduce risk of kernel failure.
V2:
- add helper of scsi_max_nr_hw_queues() for avoiding potential build failure because scsi driver often doesn't deal with blk-mq directly
- apply scsi_max_nr_hw_queues() for all scsi changes
- move lpfc's change into managed irq code path
Thanks,
Ming
Ming Lei (9):
blk-mq: add blk_mq_max_nr_hw_queues()
nvme-pci: use blk_mq_max_nr_hw_queues() to calculate io queues
scsi: core: add helper of scsi_max_nr_hw_queues()
scsi: lpfc: use blk_mq_max_nr_hw_queues() to calculate io vectors
scsi: hisi: take blk_mq_max_nr_hw_queues() into account for
calculating io vectors
scsi: mpi3mr: take blk_mq_max_nr_hw_queues() into account for
calculating io vectors
scsi: megaraid: take blk_mq_max_nr_hw_queues() into account for
calculating io vectors
scsi: mpt3sas: take blk_mq_max_nr_hw_queues() into account for
calculating io vectors
scsi: pm8001: take blk_mq_max_nr_hw_queues() into account for
calculating io vectors
block/blk-mq.c | 16 ++++++++++++++++ drivers/nvme/host/pci.c | 2 +- drivers/scsi/hisi_sas/hisi_sas_v3_hw.c | 3 +++ drivers/scsi/lpfc/lpfc_init.c | 2 ++ drivers/scsi/megaraid/megaraid_sas_base.c | 6 +++++- drivers/scsi/mpi3mr/mpi3mr_fw.c | 3 +++ drivers/scsi/mpt3sas/mpt3sas_base.c | 4 ++-- drivers/scsi/pm8001/pm8001_init.c | 4 +++- include/linux/blk-mq.h | 1 + include/scsi/scsi_host.h | 5 +++++ 10 files changed, 41 insertions(+), 5 deletions(-)
Signature · 1 line
2.40.1
PATCH v2 1/9 blk-mq: add blk_mq_max_nr_hw_queues()
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.
--- block/blk-mq.c | 16 ++++++++++++++++ include/linux/blk-mq.h | 1 + 2 files changed, 17 insertions(+)
block/blk-mq.c +16 -0
@@ -140,6 +140,22 @@ void blk_mq_freeze_queue_wait(struct request_queue *q) }EXPORT_SYMBOL_GPL (blk_mq_freeze_queue_wait); +/* +* Return the max supported nr_hw_queues for each hw queue type +* +* blk_mq_alloc_tag_set() may change nr_hw_queues for kdump kernel, so +* driver has to take blk-mq max supported nr_hw_queues into account +* when figuring out nr_hw_queues from hardware info, for avoiding +* inconsistency between driver and blk-mq. +*/ +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) {
include/linux/blk-mq.h +1 -0
@@ -711,6 +711,7 @@ 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,
Signature · 1 line
2.40.1
Re: [PATCH V2 1/9] blk-mq: add blk_mq_max_nr_hw_queues()
On 26/07/2023 10:40, Ming Lei wrote:
Hi Ming,
Quoted 4 lines
blk_mq_alloc_tag_set()may overrideset->nr_hw_queuesas 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,
"the only online CPU may not be 1 for kdump kernel, in which 'maxcpus=1' ..." - this seems inconsistent with the cover letter, where we have "'maxcpus=1' just bring up one single cpu core during booting."
Quoted 9 lines
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_queuesknowledge into account when calculating io queues.
Could you alternatively solve in blk_mq_pci_map_queues() by fixing the mappings in case of kdump to have all per-CPU mappings point at the HW queue associated with cpu0 (which I assume would be active)? It ain't pretty ...
Quoted 19 lines
Signed-off-by Ming Lei <[email protected]>--- block/blk-mq.c | 16 ++++++++++++++++ include/linux/blk-mq.h | 1 + 2 files changed, 17 insertions(+)block/blk-mq.c +4
-0@@ -140,6 +140,22 @@ void blk_mq_freeze_queue_wait(struct request_queue *q) }EXPORT_SYMBOL_GPL (blk_mq_freeze_queue_wait); +/* +* Return the max supported nr_hw_queues for each hw queue type +* +* blk_mq_alloc_tag_set() may change nr_hw_queues for kdump kernel, so
block/blk-mq.c void blk_mq_freeze_queue_wait(struct request_queue *q) · line 156
Quoted 10 lines
block/blk-mq.c +10
-0+* driver has to take blk-mq max supported nr_hw_queues into account +* when figuring out nr_hw_queues from hardware info, for avoiding +* inconsistency between driver and blk-mq. +*/ +unsigned int blk_mq_max_nr_hw_queues (void ) +{ +if (is_kdump_kernel ()) +return 1 ; +return nr_cpu_ids; +}
We rely on the driver having set->nr_hw_queues == 1 for kdump, right?
If so, how about enforcing it then, like:
block/blk-mq.c +2 -1
@@ -4426,7 +4426,8 @@ int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)* 64 tags to prevent using too much memory. */ if (is_kdump_kernel ()) { - set->nr_hw_queues =1 ; +if (set->nr_hw_queues !=1 ) +return -EINVAL ; set->nr_maps =1 ; set->queue_depth =min (64U , set->queue_depth); }
Thanks,
John
Re: [PATCH V2 1/9] blk-mq: add blk_mq_max_nr_hw_queues()
Quoted 4 lines · John Garry
On 26/07/2023 10:40, Ming Lei wrote:
Hi Ming,
Quoted 8 lines
Quoted 4 lines
blk_mq_alloc_tag_set()may overrideset->nr_hw_queuesas 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,"the only online CPU may not be 1 for kdump kernel, in which 'maxcpus=1' ..." - this seems inconsistent with the cover letter, where we have "'maxcpus=1' just bring up one single cpu core during booting."
OK, looks I should have mentioned "the only online CPU may not be 1 for maxcpus=1" in cover letter.
Quoted 6 lines
Quoted 6 lines
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.
Quoted 7 lines
Quoted 3 lines
Prepare for fixing this kind of issue by applying the added helper, so driver can take blk-mq max
nr_hw_queuesknowledge into account when calculating io queues.Could you alternatively solve in
blk_mq_pci_map_queues()by fixing the mappings in case of kdump to have all per-CPU mappings point at the HW queue associated with cpu0 (which I assume would be active)? It ain't pretty ...
cpu0 mapping isn't correct, as I mentioned that 'maxcpus=1' doesn't mean the only online cpu is cpu0, such as, on ppc64, the only online cpu could be selected at random during booting.
Quoted 37 lines
Quoted 19 lines
Signed-off-by Ming Lei <[email protected]>--- block/blk-mq.c | 16 ++++++++++++++++ include/linux/blk-mq.h | 1 + 2 files changed, 17 insertions(+)block/blk-mq.c +4
-0@@ -140,6 +140,22 @@ void blk_mq_freeze_queue_wait(struct request_queue *q) }EXPORT_SYMBOL_GPL (blk_mq_freeze_queue_wait); +/* +* Return the max supported nr_hw_queues for each hw queue type +* +* blk_mq_alloc_tag_set() may change nr_hw_queues for kdump kernel, so block/blk-mq.c void blk_mq_freeze_queue_wait(struct request_queue *q) · line 155
Quoted 10 lines
block/blk-mq.c +10
-0+* driver has to take blk-mq max supported nr_hw_queues into account +* when figuring out nr_hw_queues from hardware info, for avoiding +* inconsistency between driver and blk-mq. +*/ +unsigned int blk_mq_max_nr_hw_queues (void ) +{ +if (is_kdump_kernel ()) +return 1 ; +return nr_cpu_ids; +}We rely on the driver having
set->nr_hw_queues== 1 for kdump, right?If so, how about enforcing it then, like:
block/blk-mq.c int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set) · line 4433
Quoted 9 lines
block/blk-mq.c +2
-1@@ -4426,7 +4426,8 @@ int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)* 64 tags to prevent using too much memory. */ if (is_kdump_kernel ()) { - set->nr_hw_queues =1 ; +if (set->nr_hw_queues !=1 ) +return -EINVAL ; set->nr_maps =1 ; set->queue_depth =min (64U , set->queue_depth); }
In theory, this way should be ideal approach, but it needs all MQ drivers to get converted with blk_mq_max_nr_hw_queues().
However, it shouldn't be one issue for non-managed irq given non-managed irq can be migrated to the only online cpu.
Thanks,
Ming
PATCH v2 2/9 nvme-pci: use blk_mq_max_nr_hw_queues() to calculate io queues
In reply to Ming LeiTake blk-mq's knowledge into account for calculating io queues.
Fix wrong queue mapping in case of kdump kernel.
On arm and ppc64, 'maxcpus=1' is passed to kdump command line, see Documentation/admin-guide/kdump/kdump.rst, so num_possible_cpus() still returns all CPUs because 'maxcpus=1' just bring up one single cpu core during booting.
blk-mq sees single queue in kdump kernel, and in driver's viewpoint there are still multiple queues, this inconsistency causes driver to apply wrong queue mapping for handling IO, and IO timeout is triggered.
Meantime, single queue makes much less resource utilization, and reduce risk of kernel failure.
--- drivers/nvme/host/pci.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/nvme/host/pci.c +1 -1
@@ -2251,7 +2251,7 @@ static unsigned int nvme_max_io_queues(struct nvme_dev *dev)*/ if (dev->ctrl.quirks &NVME_QUIRK_SHARED_TAGS )return 1 ; -return num_possible_cpus () + dev->nr_write_queues + dev->nr_poll_queues; +return blk_mq_max_nr_hw_queues () + dev->nr_write_queues + dev->nr_poll_queues; }static int nvme_setup_io_queues (struct nvme_dev *dev)
Signature · 1 line
2.40.1
PATCH v2 3/9 scsi: core: add helper of scsi_max_nr_hw_queues()
In reply to Ming Leiblk_mq_alloc_tag_set() may change nr_hw_queues for kdump kernel, so driver has to take blk-mq max supported nr_hw_queues into account when figuring out nr_hw_queues from hardware info, for avoiding inconsistency between scsi driver and blk-mq.
Add helper of scsi_max_nr_hw_queues() for avoiding nr_hw_queues inconsistency between scsi driver and blk-mq.
--- include/scsi/scsi_host.h | 5 +++++ 1 file changed, 5 insertions(+)
include/scsi/scsi_host.h +5 -0
@@ -803,6 +803,11 @@ extern int scsi_host_unblock(struct Scsi_Host *shost, int new_state);void scsi_host_busy_iter (struct Scsi_Host *,bool (*fn)(struct scsi_cmnd *,void *),void *priv); +static inline unsigned int scsi_max_nr_hw_queues (void ) +{ +return blk_mq_max_nr_hw_queues (); +} +struct class_container ;/*
Signature · 1 line
2.40.1
PATCH v2 4/9 scsi: lpfc: use blk_mq_max_nr_hw_queues() to calculate io vectors
In reply to Ming LeiTake blk-mq's knowledge into account for calculating io queues.
Fix wrong queue mapping in case of kdump kernel.
On arm and ppc64, 'maxcpus=1' is passed to kdump kernel command line, see Documentation/admin-guide/kdump/kdump.rst, so num_possible_cpus() still returns all CPUs because 'maxcpus=1' just bring up one single cpu core during booting.
blk-mq sees single queue in kdump kernel, and in driver's viewpoint there are still multiple queues, this inconsistency causes driver to apply wrong queue mapping for handling IO, and IO timeout is triggered.
Meantime, single queue makes much less resource utilization, and reduce risk of kernel failure.
--- drivers/scsi/lpfc/lpfc_init.c | 2 ++ 1 file changed, 2 insertions(+)
drivers/scsi/lpfc/lpfc_init.c +2 -0
@@ -13022,6 +13022,8 @@ lpfc_sli4_enable_msix(struct lpfc_hba *phba) cpu =cpumask_first (aff_mask); cpu_select =lpfc_next_online_cpu (aff_mask, cpu); }else { + vectors = min_t(unsigned int, vectors, +scsi_max_nr_hw_queues ()); flags |=PCI_IRQ_AFFINITY ; }
Signature · 1 line
2.40.1
Re: [PATCH V2 4/9] scsi: lpfc: use blk_mq_max_nr_hw_queues() to calculate io vectors
Hi Ming,
From version 1 of the patchset, I thought we had planned to put the min comparison right above pci_alloc_irq_vectors instead?
drivers/scsi/lpfc/lpfc_init.c +2 -0
@@ -13025,6 +13025,8 @@ lpfc_sli4_enable_msix(struct lpfc_hba *phba) flags |=PCI_IRQ_AFFINITY ; } + vectors = min_t(unsigned int, vectors, scsi_max_nr_hw_queues()); + rc =pci_alloc_irq_vectors (phba->pcidev,1 , vectors, flags);if (rc <0 ) { lpfc_printf_log(phba, KERN_INFO, LOG_INIT,
Thanks,
Justin
Quoted 40 lines · Ming Lei
Take blk-mq's knowledge into account for calculating io queues.
Fix wrong queue mapping in case of kdump kernel.
On arm and ppc64, 'maxcpus=1' is passed to kdump kernel command line, see
Documentation/admin-guide/kdump/kdump.rst, sonum_possible_cpus()still returns all CPUs because 'maxcpus=1' just bring up one single cpu core during booting.blk-mq sees single queue in kdump kernel, and in driver's viewpoint there are still multiple queues, this inconsistency causes driver to apply wrong queue mapping for handling IO, and IO timeout is triggered.
Meantime, single queue makes much less resource utilization, and reduce risk of kernel failure.
Cc Justin Tee <[email protected]>Cc James Smart <[email protected]>Signed-off-by Ming Lei <[email protected]>--- drivers/scsi/lpfc/lpfc_init.c | 2 ++ 1 file changed, 2 insertions(+)drivers/scsi/lpfc/lpfc_init.c +2
-1@@ -13022,6 +13022,8 @@ lpfc_sli4_enable_msix(struct lpfc_hba *phba) cpu =cpumask_first (aff_mask); cpu_select =lpfc_next_online_cpu (aff_mask, cpu); }else { + vectors = min_t(unsigned int, vectors, +scsi_max_nr_hw_queues ()); flags |=PCI_IRQ_AFFINITY ; } --2.40.1
Re: [PATCH V2 4/9] scsi: lpfc: use blk_mq_max_nr_hw_queues() to calculate io vectors
Quoted 10 lines · Justin Tee
Hi Ming,
From version 1 of the patchset, I thought we had planned to put the min comparison right above
pci_alloc_irq_vectorsinstead?drivers/scsi/lpfc/lpfc_init.c +0
-0
drivers/scsi/lpfc/lpfc_init.c lpfc_sli4_enable_msix(struct lpfc_hba *phba) · line 13028
Quoted 4 lines
drivers/scsi/lpfc/lpfc_init.c +1
-0@@ -13025,6 +13025,8 @@ lpfc_sli4_enable_msix(struct lpfc_hba *phba) flags |=PCI_IRQ_AFFINITY ; } + vectors = min_t(unsigned int, vectors, scsi_max_nr_hw_queues());
Strictly speaking, the above change is better, but non-managed irq doesn't have such issue, that is why I just apply the change on managed irq branch.
Thanks,
Ming
Justin Tee
Reviewed-byRe: [PATCH V2 4/9] scsi: lpfc: use blk_mq_max_nr_hw_queues() to calculate io vectors
Quoted 7 lines · Ming Lei
Strictly speaking, the above change is better, but non-managed irq doesn't have such issue, that is why I just apply the change on managed irq branch.
Thanks,
Ming
Sure, thanks Ming.
Regards,
Justin
PATCH v2 5/9 scsi: hisi: take blk_mq_max_nr_hw_queues() into account for calculating io vectors
In reply to Ming LeiTake blk-mq's knowledge into account for calculating io queues.
Fix wrong queue mapping in case of kdump kernel.
On arm and ppc64, 'maxcpus=1' is passed to kdump kernel command line, see Documentation/admin-guide/kdump/kdump.rst, so num_possible_cpus() still returns all CPUs because 'maxcpus=1' just bring up one single cpu core during booting.
blk-mq sees single queue in kdump kernel, and in driver's viewpoint there are still multiple queues, this inconsistency causes driver to apply wrong queue mapping for handling IO, and IO timeout is triggered.
Meantime, single queue makes much less resource utilization, and reduce risk of kernel failure.
--- drivers/scsi/hisi_sas/hisi_sas_v3_hw.c | 3 +++ 1 file changed, 3 insertions(+)
drivers/scsi/hisi_sas/hisi_sas_v3_hw.c +3 -0
@@ -2550,6 +2550,9 @@ static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba) hisi_hba->cq_nvecs = vectors -BASE_VECTORS_V3_HW - hisi_hba->iopoll_q_cnt; +if (hisi_hba->cq_nvecs >scsi_max_nr_hw_queues ()) + hisi_hba->cq_nvecs =scsi_max_nr_hw_queues (); + shost->nr_hw_queues = hisi_hba->cq_nvecs + hisi_hba->iopoll_q_cnt;return devm_add_action (&pdev->dev, hisi_sas_v3_free_vectors, pdev);
Signature · 1 line
2.40.1
Re: [PATCH V2 5/9] scsi: hisi: take blk_mq_max_nr_hw_queues() into account for calculating io vectors
Quoted 27 lines · Ming Lei
Take blk-mq's knowledge into account for calculating io queues.
Fix wrong queue mapping in case of kdump kernel.
On arm and ppc64, 'maxcpus=1' is passed to kdump kernel command line, see
Documentation/admin-guide/kdump/kdump.rst, sonum_possible_cpus()still returns all CPUs because 'maxcpus=1' just bring up one single cpu core during booting.blk-mq sees single queue in kdump kernel, and in driver's viewpoint there are still multiple queues, this inconsistency causes driver to apply wrong queue mapping for handling IO, and IO timeout is triggered.
Meantime, single queue makes much less resource utilization, and reduce risk of kernel failure.
Cc Xiang Chen <[email protected]>Signed-off-by Ming Lei <[email protected]>--- drivers/scsi/hisi_sas/hisi_sas_v3_hw.c | 3 +++ 1 file changed, 3 insertions(+)drivers/scsi/hisi_sas/hisi_sas_v3_hw.c +0
-0
drivers/scsi/hisi_sas/hisi_sas_v3_hw.c static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba) · line 2556
Quoted 7 lines
drivers/scsi/hisi_sas/hisi_sas_v3_hw.c +3
-0@@ -2550,6 +2550,9 @@ static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba) hisi_hba->cq_nvecs = vectors -BASE_VECTORS_V3_HW - hisi_hba->iopoll_q_cnt; +if (hisi_hba->cq_nvecs >scsi_max_nr_hw_queues ()) + hisi_hba->cq_nvecs =scsi_max_nr_hw_queues (); + shost->nr_hw_queues = hisi_hba->cq_nvecs + hisi_hba->iopoll_q_cnt;
For other drivers you limit the max MSI vectors which we try to allocate according to scsi_max_nr_hw_queues(), but here you continue to alloc the same max vectors but then limit the driver's completion queue count. Why not limit the max MSI vectors also here?
Thanks,
John
Quoted 2 lines
drivers/scsi/hisi_sas/hisi_sas_v3_hw.c +0
-0return devm_add_action (&pdev->dev, hisi_sas_v3_free_vectors, pdev);
Re: [PATCH V2 5/9] scsi: hisi: take blk_mq_max_nr_hw_queues() into account for calculating io vectors
Quoted 23 lines · John Garry
Quoted 22 lines · Ming Lei
Take blk-mq's knowledge into account for calculating io queues.
Fix wrong queue mapping in case of kdump kernel.
On arm and ppc64, 'maxcpus=1' is passed to kdump kernel command line, see
Documentation/admin-guide/kdump/kdump.rst, sonum_possible_cpus()still returns all CPUs because 'maxcpus=1' just bring up one single cpu core during booting.blk-mq sees single queue in kdump kernel, and in driver's viewpoint there are still multiple queues, this inconsistency causes driver to apply wrong queue mapping for handling IO, and IO timeout is triggered.
Meantime, single queue makes much less resource utilization, and reduce risk of kernel failure.
Cc Xiang Chen <[email protected]>Signed-off-by Ming Lei <[email protected]>--- drivers/scsi/hisi_sas/hisi_sas_v3_hw.c | 3 +++ 1 file changed, 3 insertions(+)
Quoted 15 lines
drivers/scsi/hisi_sas/hisi_sas_v3_hw.c static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba) · line 2554
Quoted 10 lines
drivers/scsi/hisi_sas/hisi_sas_v3_hw.c +3
-0@@ -2550,6 +2550,9 @@ static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba) hisi_hba->cq_nvecs = vectors -BASE_VECTORS_V3_HW - hisi_hba->iopoll_q_cnt; +if (hisi_hba->cq_nvecs >scsi_max_nr_hw_queues ()) + hisi_hba->cq_nvecs =scsi_max_nr_hw_queues (); + shost->nr_hw_queues = hisi_hba->cq_nvecs + hisi_hba->iopoll_q_cnt;For other drivers you limit the max MSI vectors which we try to allocate according to
scsi_max_nr_hw_queues(), but here you continue to alloc the same max vectors but then limit the driver's completion queue count. Why not limit the max MSI vectors also here?
Good catch!
I guess it is because of the following line:
hisi_hba->cq_nvecs = vectors - BASE_VECTORS_V3_HW - hisi_hba->iopoll_q_cnt;
I am a bit confused why hisi_sas_v3 takes ->iopoll_q_cnt into account allocated msi vectors? ->iopoll_q_cnt supposes to not consume msi vectors, so I think we need the following fix first:
drivers/scsi/hisi_sas/hisi_sas_v3_hw.c +1 -1
@@ -2549,7 +2549,7 @@ static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba)return -ENOENT ; - hisi_hba->cq_nvecs = vectors -BASE_VECTORS_V3_HW - hisi_hba->iopoll_q_cnt; + hisi_hba->cq_nvecs = vectors -BASE_VECTORS_V3_HW ; shost->nr_hw_queues = hisi_hba->cq_nvecs + hisi_hba->iopoll_q_cnt;return devm_add_action (&pdev->dev, hisi_sas_v3_free_vectors, pdev);
Thanks,
Ming
Re: [PATCH V2 5/9] scsi: hisi: take blk_mq_max_nr_hw_queues() into account for calculating io vectors
Quoted 4 lines · Ming Lei
Quoted 4 lines
Quoted 4 lines
hisi_sas_v3_hw.c
+++b/drivers/scsi/hisi_sas/hisi_sas_v3_hw.c(quoted diff) +0
-0@@ -2550,6 +2550,9 @@ static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba) hisi_hba->cq_nvecs = vectors - BASE_VECTORS_V3_HW - hisi_hba->iopoll_q_cnt;
Quoted 8 lines
Quoted 8 lines
Review static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba) · line 2554
Quoted 4 lines
(quoted diff) +3
-0+ if (hisi_hba->cq_nvecs > scsi_max_nr_hw_queues()) + hisi_hba->cq_nvecs = scsi_max_nr_hw_queues(); + shost->nr_hw_queues = hisi_hba->cq_nvecs + hisi_hba->iopoll_q_cnt;For other drivers you limit the max MSI vectors which we try to allocate according to
scsi_max_nr_hw_queues(), but here you continue to alloc the same max vectors but then limit the driver's completion queue count. Why not limit the max MSI vectors also here?
Ah, checking again, I think that this driver always allocates maximum possible MSI due to arm interrupt controller driver bug - see comment at top of function interrupt_preinit_v3_hw(). IIRC, there was a problem if we remove and re-insert the device driver that the GIC ITS fails to allocate MSI unless all MSI were previously allocated.
Xiang Chen can confirm.
Quoted 15 lines
Good catch!
I guess it is because of the following line:
hisi_hba->cq_nvecs = vectors - BASE_VECTORS_V3_HW - hisi_hba->iopoll_q_cnt;I am a bit confused why
hisi_sas_v3takes ->iopoll_q_cnt into account allocated msi vectors? ->iopoll_q_cnt supposes to not consume msi vectors, so I think we need the following fix first:drivers/scsi/hisi_sas/hisi_sas_v3_hw.c +0
-0
drivers/scsi/hisi_sas/hisi_sas_v3_hw.c static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba) · line 2553
Quoted 6 lines
drivers/scsi/hisi_sas/hisi_sas_v3_hw.c +1
-1@@ -2549,7 +2549,7 @@ static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba)return -ENOENT ; - hisi_hba->cq_nvecs = vectors -BASE_VECTORS_V3_HW - hisi_hba->iopoll_q_cnt; + hisi_hba->cq_nvecs = vectors -BASE_VECTORS_V3_HW ; shost->nr_hw_queues = hisi_hba->cq_nvecs + hisi_hba->iopoll_q_cnt;
Thanks,
John
Re: [PATCH V2 5/9] scsi: hisi: take blk_mq_max_nr_hw_queues() into account for calculating io vectors
Quoted 14 lines · John Garry
Quoted 4 lines · Ming Lei
Quoted 4 lines
Quoted 4 lines
hisi_sas_v3_hw.c
+++b/drivers/scsi/hisi_sas/hisi_sas_v3_hw.c(quoted diff) +0
-0@@ -2550,6 +2550,9 @@ static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba) hisi_hba->cq_nvecs = vectors - BASE_VECTORS_V3_HW - hisi_hba->iopoll_q_cnt;Quoted 8 lines
Quoted 8 lines
Review static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba) · line 2554
Quoted 4 lines
(quoted diff) +3
-0+ if (hisi_hba->cq_nvecs > scsi_max_nr_hw_queues()) + hisi_hba->cq_nvecs = scsi_max_nr_hw_queues(); + shost->nr_hw_queues = hisi_hba->cq_nvecs + hisi_hba->iopoll_q_cnt;For other drivers you limit the max MSI vectors which we try to allocate according to
scsi_max_nr_hw_queues(), but here you continue to alloc the same max vectors but then limit the driver's completion queue count. Why not limit the max MSI vectors also here?
Quoted 5 lines
Ah, checking again, I think that this driver always allocates maximum possible MSI due to arm interrupt controller driver bug - see comment at top of function
interrupt_preinit_v3_hw(). IIRC, there was a problem if we remove and re-insert the device driver that the GIC ITS fails to allocate MSI unless all MSI were previously allocated.
My question is actually why hisi_hba->iopoll_q_cnt is subtracted from allocated vectors since io poll queues does _not_ use msi vector.
So it isn't related with driver's msi vector allocation bug, is it?
Thanks,
Ming
Re: [PATCH V2 5/9] scsi: hisi: take blk_mq_max_nr_hw_queues() into account for calculating io vectors
Quoted 12 lines · Ming Lei
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines
Review static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba) · line 2554
Quoted 8 lines
hisi_sas_v3_hw.c
+++b/drivers/scsi/hisi_sas/hisi_sas_v3_hw.c(quoted diff) +3
-0@@ -2550,6 +2550,9 @@ static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba) hisi_hba->cq_nvecs = vectors - BASE_VECTORS_V3_HW - hisi_hba->iopoll_q_cnt; + if (hisi_hba->cq_nvecs > scsi_max_nr_hw_queues()) + hisi_hba->cq_nvecs = scsi_max_nr_hw_queues(); + shost->nr_hw_queues = hisi_hba->cq_nvecs + hisi_hba->iopoll_q_cnt;For other drivers you limit the max MSI vectors which we try to allocate according to
scsi_max_nr_hw_queues(), but here you continue to alloc the same max vectors but then limit the driver's completion queue count. Why not limit the max MSI vectors also here?
Quoted 8 lines
Quoted 5 lines
Ah, checking again, I think that this driver always allocates maximum possible MSI due to arm interrupt controller driver bug - see comment at top of function
interrupt_preinit_v3_hw(). IIRC, there was a problem if we remove and re-insert the device driver that the GIC ITS fails to allocate MSI unless all MSI were previously allocated.My question is actually why
hisi_hba->iopoll_q_cntis subtracted from allocated vectors since io poll queuesdoes_not_use msi vector.
It is subtracted as superfluous vectors were allocated.
As I mentioned, I think that the driver is forced to allocate all 32 MSI vectors, even though it really needs max of 32 - iopoll_q_cnt, i.e. we don't need an MSI vector for a HW queue assigned to polling. Then, since it gets 16x MSI for cq vectors, it subtracts iopoll_q_cnt to give the desired count in cq_nvecs.
Quoted 1 line
So it isn't related with driver's msi vector allocation bug, is it?
My deduction is this is how this currently "works" for non-zero iopoll queues:
- allocate max MSI of 32, which gives 32 vectors including 16 cq vectors. That then gives:
- cq_nvecs = 16 - iopoll_q_cnt
- shost->nr_hw_queues = 16
- 16x MSI cq vectors were spread over all CPUs
- in hisi_sas_map_queues()
- HCTX_TYPE_DEFAULT qmap->nr_queues = 16 - iopoll_q_cnt, and forblk_mq_pci_map_queues() we setup affinity for 16 - iopoll_q_cnt hw queues. This looks broken, as we originally spread 16x vectors over all CPUs, but now only setup mappings for (16 - iopoll_q_cnt) vectors, whose affinity would spread a subset of CPUs. And then qmap->mq_map[] for other CPUs is not set at all.
Thanks,
John
Re: [PATCH V2 5/9] scsi: hisi: take blk_mq_max_nr_hw_queues() into account for calculating io vectors
Quoted 22 lines · John Garry
Quoted 12 lines · Ming Lei
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines
Review static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba) · line 2554
Quoted 8 lines
hisi_sas_v3_hw.c
+++b/drivers/scsi/hisi_sas/hisi_sas_v3_hw.c(quoted diff) +3
-0@@ -2550,6 +2550,9 @@ static int interrupt_preinit_v3_hw(struct hisi_hba *hisi_hba) hisi_hba->cq_nvecs = vectors - BASE_VECTORS_V3_HW - hisi_hba->iopoll_q_cnt; + if (hisi_hba->cq_nvecs > scsi_max_nr_hw_queues()) + hisi_hba->cq_nvecs = scsi_max_nr_hw_queues(); + shost->nr_hw_queues = hisi_hba->cq_nvecs + hisi_hba->iopoll_q_cnt;For other drivers you limit the max MSI vectors which we try to allocate according to
scsi_max_nr_hw_queues(), but here you continue to alloc the same max vectors but then limit the driver's completion queue count. Why not limit the max MSI vectors also here?Quoted 8 lines
Quoted 5 lines
Ah, checking again, I think that this driver always allocates maximum possible MSI due to arm interrupt controller driver bug - see comment at top of function
interrupt_preinit_v3_hw(). IIRC, there was a problem if we remove and re-insert the device driver that the GIC ITS fails to allocate MSI unless all MSI were previously allocated.My question is actually why
hisi_hba->iopoll_q_cntis subtracted from allocated vectors since io poll queuesdoes_not_use msi vector.
Quoted 7 lines
It is subtracted as superfluous vectors were allocated.
As I mentioned, I think that the driver is forced to allocate all 32 MSI vectors, even though it really needs max of 32 -
iopoll_q_cnt, i.e. we don't need an MSI vector for a HW queue assigned to polling. Then, since it gets 16x MSI for cq vectors, it subtractsiopoll_q_cntto give the desired count incq_nvecs.
Looks the current driver wires nr_irq_queues with nr_poll_queues, which is wrong logically.
Here:
1) hisi_hba->queue_count means the max supported queue count(irq queues
+ poll queues)
2) max vectors(32) means the max supported irq queues, but actual nr_irq_queues can be less than allocated vectors because of the irq allocation bug
3) so the max supported poll queues should be (hisi_hba->queue_count - nr_irq_queues).
What I meant is that nr_poll_queues shouldn't be related with allocated msi vectors.
Quoted 9 lines
Quoted 1 line
So it isn't related with driver's msi vector allocation bug, is it?
My deduction is this is how this currently "works" for non-zero iopoll queues:
- allocate max MSI of 32, which gives 32 vectors including 16 cq vectors. That then gives:
-cq_nvecs= 16 -iopoll_q_cnt
-shost->nr_hw_queues= 16
- 16x MSI cq vectors were spread over all CPUs
It should be that cq_nvecs vectors spread over all CPUs, and iopoll_q_cnt are spread over all CPUs too.
For each queue type, nr_queues of this type are spread over all CPUs.
Quoted 8 lines
- in
hisi_sas_map_queues()
-HCTX_TYPE_DEFAULTqmap->nr_queues= 16 -iopoll_q_cnt, and forblk_mq_pci_map_queues()we setup affinity for 16 -iopoll_q_cnthw queues. This looks broken, as we originally spread 16x vectors over all CPUs, but now only setup mappings for (16 -iopoll_q_cnt) vectors, whose affinity would spread a subset of CPUs. And then qmap->mq_map[] for other CPUs is not set at all.
That isn't true, please see my above comment.
Thanks,
Ming
Re: [PATCH V2 5/9] scsi: hisi: take blk_mq_max_nr_hw_queues() into account for calculating io vectors
Quoted 5 lines · Ming Lei
Quoted 5 lines
As I mentioned, I think that the driver is forced to allocate all 32 MSI vectors, even though it really needs max of 32 -
iopoll_q_cnt, i.e. we don't need an MSI vector for a HW queue assigned to polling. Then, since it gets 16x MSI for cq vectors, it subtractsiopoll_q_cntto give the desired count incq_nvecs.
Quoted 8 lines
Looks the current driver wires
nr_irq_queueswithnr_poll_queues, which is wrong logically.Here:
1)
hisi_hba->queue_countmeans the max supported queue count(irq queues
+ poll queues)
JFYI, from https://github.com/tianocore/edk2-platforms/blob/master/Silicon/Hisilicon/Hi1620/Hi1620AcpiTables/Dsdt/Hi1620Pci.asl#L587C25-L587C41, hisi_hba->queue_count is fixed at 16.
Quoted 1 line
2) max vectors(32) means the max supported irq queues,
I'll just mention some details of this HW, in case missed.
For this HW, max vectors is 32, but not all are for completion queue interrupts.
interrupts 0-15 are for housekeeping, like phy up notification interrupts 16-31 are for completion queue interrupts
That is a fixed assignment. irq #16 is for HW queue #0 interrupt, irq #17 is for HW queue #1 interrupt, and so on.
Quoted 9 lines
but actual
nr_irq_queuescan be less than allocated vectors because of the irq allocation bug3) so the max supported poll queues should be (
hisi_hba->queue_count-nr_irq_queues).What I meant is that
nr_poll_queuesshouldn't be related with allocated msi vectors.
Sure, I agree with that. nr_poll_queues is set in that driver as a module param.
I am just saying that we have a fixed number of HW queues (16), each of which may be used for interrupt or polling mode. And since we always allocate max number of MSI, then number of interrupt queues available will be 16 - nr_poll_queues.
Quoted 10 lines
Quoted 10 lines
Quoted 3 lines
So it isn't related with driver's msi vector allocation bug, is it?
My deduction is this is how this currently "works" for non-zero iopoll queues:
- allocate max MSI of 32, which gives 32 vectors including 16 cq vectors. That then gives:
-cq_nvecs= 16 -iopoll_q_cnt
-shost->nr_hw_queues= 16
- 16x MSI cq vectors were spread over all CPUs
Quoted 2 lines
It should be that
cq_nvecsvectors spread over all CPUs, andiopoll_q_cntare spread over all CPUs too.
I agree, it should be, but I don't think that it is for HCTX_TYPE_DEFAULT, as below.
Quoted 10 lines
For each queue type,
nr_queuesof this type are spread over all CPUs. >> - inhisi_sas_map_queues()Quoted 6 lines
-
HCTX_TYPE_DEFAULTqmap->nr_queues= 16 -iopoll_q_cnt, and forblk_mq_pci_map_queues()we setup affinity for 16 -iopoll_q_cnthw queues. This looks broken, as we originally spread 16x vectors over all CPUs, but now only setup mappings for (16 -iopoll_q_cnt) vectors, whose affinity would spread a subset of CPUs. And then qmap->mq_map[] for other CPUs is not set at all.That isn't true, please see my above comment.
I am just basing that on what I mention above, so please let me know my inaccuracy there.
Thanks,
John
Re: [PATCH V2 5/9] scsi: hisi: take blk_mq_max_nr_hw_queues() into account for calculating io vectors
Quoted 40 lines · John Garry
Quoted 5 lines · Ming Lei
Quoted 5 lines
As I mentioned, I think that the driver is forced to allocate all 32 MSI vectors, even though it really needs max of 32 -
iopoll_q_cnt, i.e. we don't need an MSI vector for a HW queue assigned to polling. Then, since it gets 16x MSI for cq vectors, it subtractsiopoll_q_cntto give the desired count incq_nvecs.Quoted 8 lines
Looks the current driver wires
nr_irq_queueswithnr_poll_queues, which is wrong logically.Here:
1)
hisi_hba->queue_countmeans the max supported queue count(irq queues
+ poll queues)JFYI, from https://github.com/tianocore/edk2-platforms/blob/master/Silicon/Hisilicon/Hi1620/Hi1620AcpiTables/Dsdt/Hi1620Pci.asl#L587C25-L587C41,
hisi_hba->queue_countis fixed at 16.Quoted 1 line
2) max vectors(32) means the max supported irq queues,
I'll just mention some details of this HW, in case missed.
For this HW, max vectors is 32, but not all are for completion queue interrupts.
interrupts 0-15 are for housekeeping, like phy up notification interrupts 16-31 are for completion queue interrupts
That is a fixed assignment. irq #16 is for HW queue #0 interrupt, irq #17 is for HW queue #1 interrupt, and so on.
Quoted 9 lines
but actual
nr_irq_queuescan be less than allocated vectors because of the irq allocation bug3) so the max supported poll queues should be (
hisi_hba->queue_count-nr_irq_queues).What I meant is that
nr_poll_queuesshouldn't be related with allocated msi vectors.
Quoted 7 lines
Sure, I agree with that.
nr_poll_queuesis set in that driver as a module param.I am just saying that we have a fixed number of HW queues (16), each of which may be used for interrupt or polling mode. And since we always allocate max number of MSI, then number of interrupt queues available will be 16 -
nr_poll_queues.
No.
queue_count is fixed at 16, but pci_alloc_irq_vectors_affinity() still may return less vectors, which is one system-wide resource, and queue count is device resource.
So when less vectors are allocated, you should have been capable of using more poll queues, unfortunately the current code can't support that.
Even worse, hisi_hba->cq_nvecs can become negative if less vectors are returned.
Quoted 26 lines
Quoted 11 lines
Quoted 11 lines
Quoted 4 lines
So it isn't related with driver's msi vector allocation bug, is it?
My deduction is this is how this currently "works" for non-zero iopoll queues:
- allocate max MSI of 32, which gives 32 vectors including 16 cq vectors. That then gives:
-cq_nvecs= 16 -iopoll_q_cnt
-shost->nr_hw_queues= 16
- 16x MSI cq vectors were spread over all CPUsQuoted 2 lines
It should be that
cq_nvecsvectors spread over all CPUs, andiopoll_q_cntare spread over all CPUs too.I agree, it should be, but I don't think that it is for
HCTX_TYPE_DEFAULT, as below.Quoted 9 lines
For each queue type,
nr_queuesof this type are spread over all CPUs. >> - inhisi_sas_map_queues()Quoted 6 lines
-
HCTX_TYPE_DEFAULTqmap->nr_queues= 16 -iopoll_q_cnt, and forblk_mq_pci_map_queues()we setup affinity for 16 -iopoll_q_cnthw queues. This looks broken, as we originally spread 16x vectors over all CPUs, but now only setup mappings for (16 -iopoll_q_cnt) vectors, whose affinity would spread a subset of CPUs. And then qmap->mq_map[] for other CPUs is not set at all.
Quoted 4 lines
Quoted 1 line
That isn't true, please see my above comment.
I am just basing that on what I mention above, so please let me know my inaccuracy there.
You said queue mapping for HCTX_TYPE_DEFAULT is broken, but it isn't.
You said 'we originally spread 16x vectors over all CPUs', which isn't true. Again, '16 - iopoll_q_cnt' vectors are spread on all CPUs, and same with iopoll_q_cnt vectors.
Since both blk_mq_map_queues() and blk_mq_pci_map_queues() does spread map->nr_queues over all CPUs, so there isn't spread a subset of CPUs.
Thanks,
Ming
Re: [PATCH V2 5/9] scsi: hisi: take blk_mq_max_nr_hw_queues() into account for calculating io vectors
Quoted 9 lines
Quoted 6 lines
I am just saying that we have a fixed number of HW queues (16), each of which may be used for interrupt or polling mode. And since we always allocate max number of MSI, then number of interrupt queues available will be 16 -
nr_poll_queues.No.
Quoted 8 lines
queue_countis fixed at 16, butpci_alloc_irq_vectors_affinity()still may return less vectors, which is one system-wide resource, and queue count is device resource.So when less vectors are allocated, you should have been capable of using more poll queues, unfortunately the current code can't support that.
Even worse,
hisi_hba->cq_nvecscan become negative if less vectors are returned.
OK, I see what you mean here. I thought that we were only considering case of vectors allocated was max vectors requested.
Yes, I see how allocating less than max can cause an issue. I am not sure if increasing iopoll_q_cnt over driver module param value is proper then, but obviously we don't want cq_nvecs to become negative.
Quoted 29 lines
Quoted 29 lines
Quoted 12 lines
Quoted 4 lines
Quoted 8 lines
Quoted 1 line
So it isn't related with driver's msi vector allocation bug, is it?
My deduction is this is how this currently "works" for non-zero iopoll queues:
- allocate max MSI of 32, which gives 32 vectors including 16 cq vectors. That then gives:
-cq_nvecs= 16 -iopoll_q_cnt
-shost->nr_hw_queues= 16
- 16x MSI cq vectors were spread over all CPUsQuoted 2 lines
It should be that
cq_nvecsvectors spread over all CPUs, andiopoll_q_cntare spread over all CPUs too.I agree, it should be, but I don't think that it is for
HCTX_TYPE_DEFAULT, as below.Quoted 10 lines
For each queue type,
nr_queuesof this type are spread over all CPUs. >> - inhisi_sas_map_queues()Quoted 6 lines
-
HCTX_TYPE_DEFAULTqmap->nr_queues= 16 -iopoll_q_cnt, and forblk_mq_pci_map_queues()we setup affinity for 16 -iopoll_q_cnthw queues. This looks broken, as we originally spread 16x vectors over all CPUs, but now only setup mappings for (16 -iopoll_q_cnt) vectors, whose affinity would spread a subset of CPUs. And then qmap->mq_map[] for other CPUs is not set at all.That isn't true, please see my above comment.
Quoted 7 lines
Quoted 2 lines
I am just basing that on what I mention above, so please let me know my inaccuracy there.
You said queue mapping for
HCTX_TYPE_DEFAULTis broken, but it isn't.You said 'we originally spread 16x vectors over all CPUs', which isn't true.
Are you talking about the case of allocating less then max requested vectors, as above?
If we have min_msi = 17, max_msi = 32, affinity_desc = {16, 0}, and we allocate 32 vectors from pci_alloc_irq_vectors_affinity(), then I would have thought that the affinity for the 16x cq vectors is spread over all CPUs. Is that wrong?
Quoted 6 lines
Again, '16 -
iopoll_q_cnt' vectors are spread on all CPUs, and same withiopoll_q_cntvectors.Since both
blk_mq_map_queues()andblk_mq_pci_map_queues()does spreadmap->nr_queuesover all CPUs, so there isn't spread a subset of CPUs.
Thanks,
John
PATCH v2 6/9 scsi: mpi3mr: take blk_mq_max_nr_hw_queues() into account for calculating io vectors
In reply to Ming LeiTake blk-mq's knowledge into account for calculating io queues.
Fix wrong queue mapping in case of kdump kernel.
On arm and ppc64, 'maxcpus=1' is passed to kdump kernel command line, see Documentation/admin-guide/kdump/kdump.rst, so num_possible_cpus() still returns all CPUs because 'maxcpus=1' just bring up one single cpu core during booting.
blk-mq sees single queue in kdump kernel, and in driver's viewpoint there are still multiple queues, this inconsistency causes driver to apply wrong queue mapping for handling IO, and IO timeout is triggered.
Meantime, single queue makes much less resource utilization, and reduce risk of kernel failure.
--- drivers/scsi/mpi3mr/mpi3mr_fw.c | 3 +++ 1 file changed, 3 insertions(+)
drivers/scsi/mpi3mr/mpi3mr_fw.c +3 -0
@@ -815,6 +815,9 @@ static int mpi3mr_setup_isr(struct mpi3mr_ioc *mrioc, u8 setup_one) desc.post_vectors = mrioc->requested_poll_qcount; min_vec = desc.pre_vectors + desc.post_vectors; +if (max_vectors - min_vec >scsi_max_nr_hw_queues ()) + max_vectors = min_vec +scsi_max_nr_hw_queues (); + irq_flags |=PCI_IRQ_AFFINITY |PCI_IRQ_ALL_TYPES ; retval = pci_alloc_irq_vectors_affinity(mrioc->pdev,
Signature · 1 line
2.40.1
PATCH v2 7/9 scsi: megaraid: take blk_mq_max_nr_hw_queues() into account for calculating io vectors
In reply to Ming LeiTake blk-mq's knowledge into account for calculating io queues.
Fix wrong queue mapping in case of kdump kernel.
On arm and ppc64, 'maxcpus=1' is passed to kdump kernel command line, see Documentation/admin-guide/kdump/kdump.rst, so num_possible_cpus() still returns all CPUs because 'maxcpus=1' just bring up one single cpu core during booting.
blk-mq sees single queue in kdump kernel, and in driver's viewpoint there are still multiple queues, this inconsistency causes driver to apply wrong queue mapping for handling IO, and IO timeout is triggered.
Meantime, single queue makes much less resource utilization, and reduce risk of kernel failure.
--- drivers/scsi/megaraid/megaraid_sas_base.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)
drivers/scsi/megaraid/megaraid_sas_base.c +5 -1
@@ -5921,6 +5921,10 @@ __megasas_alloc_irq_vectors(struct megasas_instance *instance)int i, irq_flags;struct irq_affinity desc = { .pre_vectors = instance->low_latency_index_start };struct irq_affinity *descp = &desc; +unsigned max_vecs = instance->msix_vectors - instance->iopoll_q_count; + +if (max_vecs >scsi_max_nr_hw_queues ()) + max_vecs =scsi_max_nr_hw_queues (); irq_flags =PCI_IRQ_MSIX ;
@@ -5934,7 +5938,7 @@ __megasas_alloc_irq_vectors(struct megasas_instance *instance)*/ i =pci_alloc_irq_vectors_affinity (instance->pdev, instance->low_latency_index_start, - instance->msix_vectors - instance->iopoll_q_count, irq_flags, descp); + max_vecs, irq_flags, descp);return i; }
Signature · 1 line
2.40.1
PATCH v2 8/9 scsi: mpt3sas: take blk_mq_max_nr_hw_queues() into account for calculating io vectors
In reply to Ming LeiTake blk-mq's knowledge into account for calculating io queues.
Fix wrong queue mapping in case of kdump kernel.
On arm and ppc64, 'maxcpus=1' is passed to kdump kernel command line, see Documentation/admin-guide/kdump/kdump.rst, so num_possible_cpus() still returns all CPUs because 'maxcpus=1' just bring up one single cpu core during booting.
blk-mq sees single queue in kdump kernel, and in driver's viewpoint there are still multiple queues, this inconsistency causes driver to apply wrong queue mapping for handling IO, and IO timeout is triggered.
Meantime, single queue makes much less resource utilization, and reduce risk of kernel failure.
--- drivers/scsi/mpt3sas/mpt3sas_base.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/scsi/mpt3sas/mpt3sas_base.c +2 -2
@@ -3332,8 +3332,8 @@ _base_alloc_irq_vectors(struct MPT3SAS_ADAPTER *ioc)* Don't allocate msix vectors for poll_queues. * msix_vectors is always within a range of FW supported reply queue. */ -int nr_msix_vectors = ioc->iopoll_q_start_index; - + int nr_msix_vectors = min_t(unsigned int, ioc->iopoll_q_start_index, +scsi_max_nr_hw_queues ());if (ioc->smp_affinity_enable) irq_flags |=PCI_IRQ_AFFINITY |PCI_IRQ_ALL_TYPES ;
Signature · 1 line
2.40.1
PATCH v2 9/9 scsi: pm8001: take blk_mq_max_nr_hw_queues() into account for calculating io vectors
In reply to Ming LeiTake blk-mq's knowledge into account for calculating io queues.
Fix wrong queue mapping in case of kdump kernel.
On arm and ppc64, 'maxcpus=1' is passed to kdump kernel command line, see Documentation/admin-guide/kdump/kdump.rst, so num_possible_cpus() still returns all CPUs because 'maxcpus=1' just bring up one single cpu core during booting.
blk-mq sees single queue in kdump kernel, and in driver's viewpoint there are still multiple queues, this inconsistency causes driver to apply wrong queue mapping for handling IO, and IO timeout is triggered.
Meantime, single queue makes much less resource utilization, and reduce risk of kernel failure.
--- drivers/scsi/pm8001/pm8001_init.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
drivers/scsi/pm8001/pm8001_init.c +3 -1
@@ -965,6 +965,8 @@ static u32 pm8001_setup_msix(struct pm8001_hba_info *pm8001_ha) rc =pci_alloc_irq_vectors (pm8001_ha->pdev,1 ,1 ,PCI_IRQ_MSIX ); }else { + unsigned int max_vecs = min_t(unsigned int, PM8001_MAX_MSIX_VEC, +scsi_max_nr_hw_queues () +1 );/* * Queue index #0 is used always for housekeeping, so don't * include in the affinity spreading.
@@ -973,7 +975,7 @@ static u32 pm8001_setup_msix(struct pm8001_hba_info *pm8001_ha) .pre_vectors = 1, }; rc =pci_alloc_irq_vectors_affinity ( - pm8001_ha->pdev,2 ,PM8001_MAX_MSIX_VEC , + pm8001_ha->pdev,2 , max_vecs,PCI_IRQ_MSIX |PCI_IRQ_AFFINITY , &desc); }
Signature · 1 line
2.40.1
Christoph Hellwig
maintainerQuoted 6 lines · Ming Lei
Hi,
On arm and ppc64, 'maxcpus=1' is required for kdump kernel, see
Documentation/admin-guide/kdump/kdump.rst, sonum_possible_cpus()still returns all CPUs because 'maxcpus=1' just bring up one single cpu core during booting.
Just as said last time: Please drop the odd is_kdump_kernel checks in blk_mq_update_queue_map and blk_mq_alloc_tag_set instead of sprinkling this crap even further.