[dm-devel] [PATCH 3/3] block: use a driver-specific handler for the "inflight" value

[Mikulas Patocka]

On Fri, 16 Nov 2018, Jens Axboe wrote:

> On 11/16/18 6:55 AM, Mike Snitzer wrote:
> > On Fri, Nov 16 2018 at  4:11am -0500,
> > Christoph Hellwig <hch at infradead.org> wrote:
> > 
> >> On Fri, Nov 16, 2018 at 01:04:19AM +0100, Mikulas Patocka wrote:
> >>> Device mapper was converted to percpu inflight counters. In order to
> >>> display the correct values in the "inflight" sysfs file and in
> >>> /proc/diskstats, we need a custom callback that sums the percpu counters.
> >>>
> >>> The function part_round_stats calculates the number of in-flight I/Os
> >>> every jiffy and uses this to calculate the counters time_in_queue and
> >>> io_ticks. In order to avoid excessive memory traffic on systems with high
> >>> number of CPUs, this functionality is disabled when percpu inflight values
> >>> are used and the values time_in_queue and io_ticks are calculated
> >>> differently - the result is less precise.
> >>
> >> And none of that is device mapper specific.  Please submit this code
> >> to the block layer for use by all make_request based drivers.  Depending
> >> on what Jens as the maintainers thinkgs of the tradeoffs we can discuss
> >> if the summing should be on or off by default, or if we maybe even
> >> need the current code as a fallback.
> > 
> > I agree.
> > 
> > Mikulas, we discussed that the changes would be made to block core.  I
> > know that makes you less comfortable (I assume because you need to
> > consider more than DM) but it is the right way forward.
> > 
> > Now that the legacy IO path is gone we have less to consider; these
> > counters are only impacting bio-based.
> 
> Agree, either the new code is good enough to be used in general, and
> then it should be generally used, or it should not exist in the first
> place. We've always worked very hard to provide the most efficient
> helpers and infrastructure we can in the block layer itself, so that
> drivers don't have to reinvent the wheel.
> 
> -- 
> Jens Axboe

I have generalized the per-cpu changes (so that all bio-based block 
devices use per-cpu in_flight counters) and I've made patches for the 
for-4.21/block branch in your git. I'm sending them.

Mikulas