Re: [PATCH] blk-mq: fix selecting software ctx for request

[Ming Lei]

On Sun, Feb 02, 2020 at 06:20:04PM +0800, Hillf Danton wrote:
> 
> Select the current cpu if it's mapped to hardware to make helpers like
> blk_mq_rq_cpu() return correct value.
> 
> Signed-off-by: Hillf Danton <hdanton@xxxxxxxx>
> ---
> 
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -454,7 +454,10 @@ struct request *blk_mq_alloc_request_hct
>  		blk_queue_exit(q);
>  		return ERR_PTR(-EXDEV);
>  	}
> -	cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
> +	cpu = raw_smp_processor_id();
> +	if (!cpumask_test_cpu(cpu, alloc_data.hctx->cpumask))
> +		cpu = cpumask_first_and(alloc_data.hctx->cpumask,
> +						cpu_online_mask);

How can you know if there is any online CPU available for this hctx?

>  	alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
>  
>  	rq = blk_mq_get_request(q, NULL, &alloc_data);
> 

It is really one NVMe specific issue, see the following discussion:

https://lore.kernel.org/linux-block/8f4402a0-967d-f12d-2f1a-949e1dda017c@xxxxxxxxxxx/#r


Thanks,
Ming