On 01/09/2019 01:37 PM, Waiman Long wrote: > On 01/09/2019 01:24 PM, Matthew Wilcox wrote: >> On Wed, Jan 09, 2019 at 01:03:33PM -0500, Waiman Long wrote: >>> The paragraph above may be a bit misleading. This v2 patch actually >>> touches very little on percpu accounting aspect of the IRQ counts. See >>> patches 2 and 3 for the relevant changes which is just a few line of new >>> codes. Please review the individual patches before Nak'ing. >>> >>> I could theoretically generalize them into a new set of percpu counting >>> helpers, but the idea behind it is quite different from the use cases of >>> percpu counter. So it may not be a good idea of adding it to there. >> Did you even try just using the general purpose infrastructure that's >> in place? If that shows a performance problem _then_ it's time to make >> this special snowflake just a little more special. Not before. > I have looked into the percpu counter code. There are two aspects that I > don't like to introduce to the interrupt handler's code path for > updating the counts. > > 1) There is a raw spinlock in the percpu_counter structure that may need > to be acquired in the update path. This can be a performance drag > especially if lockdep is enabled. > > 2) The percpu_counter structure is 40 bytes in size on 64-bit systems > compared with just 8 bytes for the percpu count pointer and an > additional 4 bytes that I introduced in patch 2. With thousands of irq > descriptors, it can consume quite a lot more memory. Memory consumption > was a point that you brought up in one of your previous mails. If you read patch 4, you can see that quite a bit of CPU cycles was spent looking up the radix tree to locate the IRQ descriptor for each of the interrupts. Those overhead will still be there even if I use percpu counters. So using percpu counter alone won't be as performant as this patch or my previous v1 patch. Cheers, Longman
