On 03/10/22 10:54, Yury Norov wrote:
> On Mon, Oct 03, 2022 at 04:34:16PM +0100, Valentin Schneider wrote:
>> A recent commit made cpumask_next*() trigger a warning when passed
>> n = nr_cpu_ids - 1. This means extra care must be taken when feeding CPU
>> numbers back into cpumask_next*().
>>
>> The warning occurs nearly every boot on QEMU:
>
> [...]
>
>> Fixes: 78e5a3399421 ("cpumask: fix checking valid cpu range")
>
> No! It fixes blk-mq bug, which has been revealed after 78e5a3399421.
>
>> Suggested-by: Yury Norov <yury.norov@xxxxxxxxx>
>
> OK, maybe I suggested something like this. But after looking into the code
> of blk_mq_hctx_next_cpu() code for more, I have a feeling that this should
> be overridden deeper.
>
> Can you check - did this warning raise because hctx->next_cpu, or
> because cpumask_next_and() was called twice after jumping on
> select_cpu label?
>
It seems to always happen when hctx->next_cpu == nr_cpu_ids-1 at the start
of the function - no jumping involved.
>> Signed-off-by: Valentin Schneider <vschneid@xxxxxxxxxx>
>> ---
>> block/blk-mq.c | 9 +++++++--
>> 1 file changed, 7 insertions(+), 2 deletions(-)
>>
>> diff --git a/block/blk-mq.c b/block/blk-mq.c
>> index c96c8c4f751b..30ae51eda95e 100644
>> --- a/block/blk-mq.c
>> +++ b/block/blk-mq.c
>> @@ -2046,8 +2046,13 @@ static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
>>
>> if (--hctx->next_cpu_batch <= 0) {
>> select_cpu:
>
> Because we have backward looking goto, I have a strong feeling that the
> code should be reorganized.
>
>> - next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
>> - cpu_online_mask);
>> + if (next_cpu == nr_cpu_ids - 1)
>> + next_cpu = nr_cpu_ids;
>> + else
>> + next_cpu = cpumask_next_and(next_cpu,
>> + hctx->cpumask,
>> + cpu_online_mask);
>> +
>> if (next_cpu >= nr_cpu_ids)
>> next_cpu = blk_mq_first_mapped_cpu(hctx);
>
> This simply means 'let's start from the beginning', and should be
> replaced with cpumask_next_and_wrap().
I hadn't looked in depth there, but that's a strange behaviour.
If we get to the end of the cpumask, blk_mq_first_mapped_cpu() does:
int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
if (cpu >= nr_cpu_ids)
cpu = cpumask_first(hctx->cpumask);
return cpu;
That if branch means the returned CPU is offline, which then triggers:
if (!cpu_online(next_cpu)) {
if (!tried) {
tried = true;
goto select_cpu;
}
but going back to select_cpu doesn't make much sense, we've already checked
that hctx->cpumask and cpu_online_mask were disjoint.
>
>> hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
>
>
> Maybe something like this would work?
>
> if (--hctx->next_cpu_batch > 0 && cpu_online(next_cpu)) {
> hctx->next_cpu = next_cpu;
> return next_cpu;
> }
>
> next_cpu = cpumask_next_and_wrap(next_cpu, hctx->cpumask, cpu_online_mask)
> if (next_cpu < nr_cpu_ids) {
> hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
> hctx->next_cpu = next_cpu;
> return next_cpu;
> }
>
> /*
> * Make sure to re-select CPU next time once after CPUs
> * in hctx->cpumask become online again.
> */
> hctx->next_cpu = next_cpu;
> hctx->next_cpu_batch = 1;
> return WORK_CPU_UNBOUND;
>
> I didn't test it and likely screwed some corner case. I'm just
> trying to say that picking next cpu should be an easier thing.
>
Agreed, your suggestion looks sane, let me play with that a bit.
> Thanks,
> Yury
