In large VMs with multiple NUMA nodes, network performance is usually
best if network interrupts are all assigned to the same virtual NUMA
node. This patch assigns online CPU according to a numa aware policy,
local cpus are returned first, followed by non-local ones, then it wraps
around.
Signed-off-by: Saurabh Sengar <ssengar@linux.microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Link: https://lore.kernel.org/r/1667282761-11547-1-git-send-email-ssengar@linux.microsoft.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
void __iomem *shm_base;
void __iomem *db_page_base;
u32 db_page_size;
void __iomem *shm_base;
void __iomem *db_page_base;
u32 db_page_size;
/* Shared memory chanenl (used to bootstrap HWC) */
struct shm_channel shm_channel;
/* Shared memory chanenl (used to bootstrap HWC) */
struct shm_channel shm_channel;
struct gdma_context *gc = pci_get_drvdata(pdev);
struct gdma_irq_context *gic;
unsigned int max_irqs;
struct gdma_context *gc = pci_get_drvdata(pdev);
struct gdma_irq_context *gic;
unsigned int max_irqs;
+ u16 *cpus;
+ cpumask_var_t req_mask;
if (max_queues_per_port > MANA_MAX_NUM_QUEUES)
max_queues_per_port = MANA_MAX_NUM_QUEUES;
if (max_queues_per_port > MANA_MAX_NUM_QUEUES)
max_queues_per_port = MANA_MAX_NUM_QUEUES;
+ if (!zalloc_cpumask_var(&req_mask, GFP_KERNEL)) {
+ err = -ENOMEM;
+ goto free_irq;
+ }
+
+ cpus = kcalloc(nvec, sizeof(*cpus), GFP_KERNEL);
+ if (!cpus) {
+ err = -ENOMEM;
+ goto free_mask;
+ }
+ for (i = 0; i < nvec; i++)
+ cpus[i] = cpumask_local_spread(i, gc->numa_node);
+
for (i = 0; i < nvec; i++) {
for (i = 0; i < nvec; i++) {
+ cpumask_set_cpu(cpus[i], req_mask);
gic = &gc->irq_contexts[i];
gic->handler = NULL;
gic->arg = NULL;
gic = &gc->irq_contexts[i];
gic->handler = NULL;
gic->arg = NULL;
irq = pci_irq_vector(pdev, i);
if (irq < 0) {
err = irq;
irq = pci_irq_vector(pdev, i);
if (irq < 0) {
err = irq;
}
err = request_irq(irq, mana_gd_intr, 0, "mana_intr", gic);
if (err)
}
err = request_irq(irq, mana_gd_intr, 0, "mana_intr", gic);
if (err)
+ goto free_mask;
+ irq_set_affinity_and_hint(irq, req_mask);
+ cpumask_clear(req_mask);
+ free_cpumask_var(req_mask);
+ kfree(cpus);
err = mana_gd_alloc_res_map(nvec, &gc->msix_resource);
if (err)
err = mana_gd_alloc_res_map(nvec, &gc->msix_resource);
if (err)
+free_mask:
+ free_cpumask_var(req_mask);
+ kfree(cpus);
free_irq:
for (j = i - 1; j >= 0; j--) {
irq = pci_irq_vector(pdev, j);
free_irq:
for (j = i - 1; j >= 0; j--) {
irq = pci_irq_vector(pdev, j);
if (!bar0_va)
goto free_gc;
if (!bar0_va)
goto free_gc;
+ gc->numa_node = dev_to_node(&pdev->dev);
gc->is_pf = mana_is_pf(pdev->device);
gc->bar0_va = bar0_va;
gc->dev = &pdev->dev;
gc->is_pf = mana_is_pf(pdev->device);
gc->bar0_va = bar0_va;
gc->dev = &pdev->dev;