genirq/affinity: Spread IRQs to all available NUMA nodes
PATCH v2 1/1 genirq/affinity: Spread IRQs to all available NUMA nodes
From: Long Li <[email protected]>
On systems with large number of NUMA nodes, there may be more NUMA nodes than the number of MSI/MSI-X interrupts that device requests for. The current code always picks up the NUMA nodes starting from the node 0, up to the number of interrupts requested. This may left some later NUMA nodes unused.
For example, if the system has 16 NUMA nodes, and the device reqeusts for 8 interrupts, NUMA node 0 to 7 are assigned for those interrupts, NUMA 8 to 15 are unused.
There are several problems with this approach:
1. Later, when those managed IRQs are allocated, they can not be assigned to NUMA 8 to 15, this may create an IRQ concentration on NUMA 0 to 7.
2. Some upper layers assume affinity mask has a complete coverage over NUMA nodes. For example, block layer use the affinity mask to decide how to map CPU queues to hardware queues, missing NUMA nodes in the masks may result in an uneven mapping of queues. For the above example of 16 NUMA nodes, CPU queues on NUMA node 0 to 7 are assigned to the hardware queues 0 to 7, respectively. But CPU queues on NUMA node 8 to 15 are all assigned to the hardware queue 0.
Fix this problem by going over all NUMA nodes and assign them round-robin to all IRQs.
Change in v2: Removed extra code for calculating "done". (Michael Kelley <[email protected]>)
--- kernel/irq/affinity.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-)
kernel/irq/affinity.c +2 -3
@@ -117,12 +117,11 @@ static int irq_build_affinity_masks(const struct irq_affinity *affd,*/ if (numvecs <= nodes) {for_each_node_mask (n, nodemsk) { -cpumask_copy (masks + curvec, node_to_cpumask[n]); -if (++done == numvecs) -break ; +cpumask_or (masks + curvec, masks + curvec, node_to_cpumask[n]);if (++curvec == last_affv) curvec = affd->pre_vectors; } + done = numvecs; goto out; }
Signature · 1 line
2.14.1
tip-bot for Long Li
merge botgenirq/affinity: Spread IRQs to all available NUMA nodes
Commit-ID: b82592199032bf7c778f861b936287e37ebc9f62 Gitweb: https://git.kernel.org/tip/b82592199032bf7c778f861b936287e37ebc9f62 Author: Long Li <[email protected]> AuthorDate: Fri, 2 Nov 2018 18:02:48 +0000 Committer: Thomas Gleixner <[email protected]> CommitDate: Mon, 5 Nov 2018 12:16:26 +0100
genirq/affinity: Spread IRQs to all available NUMA nodes
If the number of NUMA nodes exceeds the number of MSI/MSI-X interrupts which are allocated for a device, the interrupt affinity spreading code fails to spread them across all nodes.
The reason is, that the spreading code starts from node 0 and continues up to the number of interrupts requested for allocation. This leaves the nodes past the last interrupt unused.
This results in interrupt concentration on the first nodes which violates the assumption of the block layer that all nodes are covered evenly. As a consequence the NUMA nodes above the number of interrupts are all assigned to hardware queue 0 and therefore NUMA node 0, which results in bad performance and has CPU hotplug implications, because queue 0 gets shut down when the last CPU of node 0 is offlined.
Go over all NUMA nodes and assign them round-robin to all requested interrupts to solve this.
[ tglx: Massaged changelog ]
--- kernel/irq/affinity.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-)
kernel/irq/affinity.c +2 -3
@@ -117,12 +117,11 @@ static int irq_build_affinity_masks(const struct irq_affinity *affd,*/ if (numvecs <= nodes) {for_each_node_mask (n, nodemsk) { -cpumask_copy (masks + curvec, node_to_cpumask[n]); -if (++done == numvecs) -break ; +cpumask_or (masks + curvec, masks + curvec, node_to_cpumask[n]);if (++curvec == last_affv) curvec = affd->pre_vectors; } + done = numvecs; goto out; }