On 1/23/24 11:47 AM, Bart Van Assche wrote:
> On 1/23/24 09:34, Jens Axboe wrote:
>> Based on the similar patch for mq-deadline, this uses separate
>> insertion lists so we can defer touching dd->lock until dispatch
> ^^^^^^^^
> bfqd->lock?
Thanks
>> with ~30% lock contention and 14.5% sys time, by applying the lessons
>> learnt with scaling mq-deadline. Patch needs to be split, but it:
>
> Is the last sentence above perhaps incomplete?
It should just be removed, it's outdated as it doesn't need splitting
anymore, I already did that work.
>> diff --git a/block/bfq-iosched.h b/block/bfq-iosched.h
>> index 56ff69f22163..f44f5d4ec2f4 100644
>> --- a/block/bfq-iosched.h
>> +++ b/block/bfq-iosched.h
>> @@ -516,10 +516,14 @@ enum {
>> struct bfq_data {
>> struct {
>> spinlock_t lock;
>> + spinlock_t insert_lock;
>> } ____cacheline_aligned_in_smp;
>
> Can lock contention be reduced further by applying
> ____cacheline_aligned_in_smp to each spinlock instead of the
> surrounding struct?
Same as deadline, it doesn't really matter in my testing.
--
Jens Axboe
