linux-kernel - Re: [PATCH V4 6/6] blk-mq: Build default queue map via group_cpus

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <e37b3007-a9eb-ff07-2efd-467a5eb8114c@oracle.com>
Date:   Wed, 11 Jan 2023 09:58:45 +0000
From:   John Garry <john.g.garry@...cle.com>
To:     Ming Lei <ming.lei@...hat.com>,
        Thomas Gleixner <tglx@...utronix.de>,
        Jens Axboe <axboe@...nel.dk>
Cc:     linux-kernel@...r.kernel.org, linux-block@...r.kernel.org,
        Christoph Hellwig <hch@....de>
Subject: Re: [PATCH V4 6/6] blk-mq: Build default queue map via
 group_cpus_evenly()

On 27/12/2022 02:29, Ming Lei wrote:
> The default queue mapping builder of blk_mq_map_queues doesn't take NUMA
> topo into account, so the built mapping is pretty bad, since CPUs
> belonging to different NUMA node are assigned to same queue. It is
> observed that IOPS drops by ~30% when running two jobs on same hctx
> of null_blk from two CPUs belonging to two NUMA nodes compared with
> from same NUMA node.
> 
> Address the issue by reusing group_cpus_evenly() for building queue
> mapping since group_cpus_evenly() does group cpus according to CPU/NUMA
> locality.
> 
> Also performance data becomes more stable with this patchset given
> correct queue mapping is applied wrt. numa locality viewpoint, for
> example, on one two nodes arm64 machine with 160 cpus, node 0(cpu 0~79),
> node 1(cpu 80~159):
> 
> 1) modprobe null_blk nr_devices=1 submit_queues=2
> 
> 2) run 'fio(t/io_uring -p 0 -n 4 -r 20 /dev/nullb0)', and observe that
> IOPS becomes much stable on multiple tests:
> 
> - without patched: IOPS is 2.5M ~ 4.5M
> - patched: IOPS is 4.3 ~ 5M
> 
> Lots of drivers may benefit from the change, such as nvme pci poll,
> nvme tcp, ...
> 
> Reviewed-by: Christoph Hellwig <hch@....de>
> Signed-off-by: Ming Lei <ming.lei@...hat.com>

FWIW, but just a comment below:

Reviewed-by: John Garry <john.g.garry@...cle.com>

> ---
>   block/blk-mq-cpumap.c | 63 +++++++++----------------------------------
>   1 file changed, 13 insertions(+), 50 deletions(-)
> 
> diff --git a/block/blk-mq-cpumap.c b/block/blk-mq-cpumap.c
> index 9c2fce1a7b50..0c612c19feb8 100644
> --- a/block/blk-mq-cpumap.c
> +++ b/block/blk-mq-cpumap.c
> @@ -10,66 +10,29 @@
>   #include <linux/mm.h>
>   #include <linux/smp.h>
>   #include <linux/cpu.h>
> +#include <linux/group_cpus.h>
>   
>   #include <linux/blk-mq.h>
>   #include "blk.h"
>   #include "blk-mq.h"
>   
> -static int queue_index(struct blk_mq_queue_map *qmap,
> -		       unsigned int nr_queues, const int q)
> -{
> -	return qmap->queue_offset + (q % nr_queues);
> -}
> -
> -static int get_first_sibling(unsigned int cpu)
> -{
> -	unsigned int ret;
> -
> -	ret = cpumask_first(topology_sibling_cpumask(cpu));
> -	if (ret < nr_cpu_ids)
> -		return ret;
> -
> -	return cpu;
> -}
> -
>   void blk_mq_map_queues(struct blk_mq_queue_map *qmap)
>   {
> -	unsigned int *map = qmap->mq_map;
> -	unsigned int nr_queues = qmap->nr_queues;
> -	unsigned int cpu, first_sibling, q = 0;
> -
> -	for_each_possible_cpu(cpu)
> -		map[cpu] = -1;
> -
> -	/*
> -	 * Spread queues among present CPUs first for minimizing
> -	 * count of dead queues which are mapped by all un-present CPUs
> -	 */
> -	for_each_present_cpu(cpu) {
> -		if (q >= nr_queues)
> -			break;
> -		map[cpu] = queue_index(qmap, nr_queues, q++);
> +	const struct cpumask *masks;
> +	unsigned int queue, cpu;
> +
> +	masks = group_cpus_evenly(qmap->nr_queues);
> +	if (!masks) {
> +		for_each_possible_cpu(cpu)
> +			qmap->mq_map[cpu] = qmap->queue_offset;

I'm not sure if we should try something better than just assigning all 
CPUs to a single queue (which we seem to be doing), but I suppose we 
don't expect masks alloc to fail and there are bigger issues to deal 
with if it does ....

> +		return;
>   	}
>   
> -	for_each_possible_cpu(cpu) {
> -		if (map[cpu] != -1)
> -			continue;
> -		/*
> -		 * First do sequential mapping between CPUs and queues.
> -		 * In case we still have CPUs to map, and we have some number of
> -		 * threads per cores then map sibling threads to the same queue
> -		 * for performance optimizations.
> -		 */
> -		if (q < nr_queues) {
> -			map[cpu] = queue_index(qmap, nr_queues, q++);
> -		} else {
> -			first_sibling = get_first_sibling(cpu);
> -			if (first_sibling == cpu)
> -				map[cpu] = queue_index(qmap, nr_queues, q++);
> -			else
> -				map[cpu] = map[first_sibling];
> -		}
> +	for (queue = 0; queue < qmap->nr_queues; queue++) {
> +		for_each_cpu(cpu, &masks[queue])
> +			qmap->mq_map[cpu] = qmap->queue_offset + queue;
>   	}
> +	kfree(masks);
>   }
>   EXPORT_SYMBOL_GPL(blk_mq_map_queues);
>