For SCSI hosts which enable host_tagset the NUMA node returned from blk_mq_hw_queue_to_node() is NUMA_NO_NODE always. Then, since in scsi_mq_setup_tags() the default we choose for the tag_set NUMA node is NUMA_NO_NODE, we always evaluate the NUMA node as NUMA_NO_NODE in functions like blk_mq_alloc_rq_map(). The reason we get NUMA_NO_NODE from blk_mq_hw_queue_to_node() is that the hctx_idx passed is BLK_MQ_NO_HCTX_IDX - so we can't match against a (HW) queue mapping index. Improve this by defaulting the tag_set NUMA node to the same NUMA node of the SCSI host DMA dev. Signed-off-by: John Garry <john.garry@xxxxxxxxxx> --- Difference to v1: - use dev_to_node() diff --git a/drivers/scsi/hosts.c b/drivers/scsi/hosts.c index f69b77cbf538..8352f90d997d 100644 --- a/drivers/scsi/hosts.c +++ b/drivers/scsi/hosts.c @@ -229,10 +229,6 @@ int scsi_add_host_with_dma(struct Scsi_Host *shost, struct device *dev, if (error) goto fail; - error = scsi_mq_setup_tags(shost); - if (error) - goto fail; - if (!shost->shost_gendev.parent) shost->shost_gendev.parent = dev ? dev : &platform_bus; if (!dma_dev) @@ -240,6 +236,10 @@ int scsi_add_host_with_dma(struct Scsi_Host *shost, struct device *dev, shost->dma_dev = dma_dev; + error = scsi_mq_setup_tags(shost); + if (error) + goto fail; + /* * Increase usage count temporarily here so that calling * scsi_autopm_put_host() will trigger runtime idle if there is diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c index a7788184908e..e14ad193a9c8 100644 --- a/drivers/scsi/scsi_lib.c +++ b/drivers/scsi/scsi_lib.c @@ -1977,7 +1977,7 @@ int scsi_mq_setup_tags(struct Scsi_Host *shost) tag_set->nr_maps = shost->nr_maps ? : 1; tag_set->queue_depth = shost->can_queue; tag_set->cmd_size = cmd_size; - tag_set->numa_node = NUMA_NO_NODE; + tag_set->numa_node = dev_to_node(shost->dma_dev); tag_set->flags = BLK_MQ_F_SHOULD_MERGE; tag_set->flags |= BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy); -- 2.26.2
