Re: [RFC PATCH-tip v4 02/10] locking/rwsem: Stop active read lock ASAP

[Waiman Long]

On 10/06/2016 05:47 PM, Dave Chinner wrote:
> On Thu, Oct 06, 2016 at 11:17:18AM -0700, Davidlohr Bueso wrote:
> > On Thu, 18 Aug 2016, Waiman Long wrote:
> >
> > > Currently, when down_read() fails, the active read locking isn't undone
> > > until the rwsem_down_read_failed() function grabs the wait_lock. If the
> > > wait_lock is contended, it may takes a while to get the lock. During
> > > that period, writer lock stealing will be disabled because of the
> > > active read lock.
> > >
> > > This patch will release the active read lock ASAP so that writer lock
> > > stealing can happen sooner. The only downside is when the reader is
> > > the first one in the wait queue as it has to issue another atomic
> > > operation to update the count.
> > >
> > > On a 4-socket Haswell machine running on a 4.7-rc1 tip-based kernel,
> > > the fio test with multithreaded randrw and randwrite tests on the
> > > same file on a XFS partition on top of a NVDIMM with DAX were run,
> > > the aggregated bandwidths before and after the patch were as follows:
> > >
> > > Test      BW before patch     BW after patch  % change
> > > ----      ---------------     --------------  --------
> > > randrw        1210 MB/s          1352 MB/s      +12%
> > > randwrite     1622 MB/s          1710 MB/s      +5.4%
> > Yeah, this is really a bad workload to make decisions on locking
> > heuristics imo - if I'm thinking of the same workload. Mainly because
> > concurrent buffered io to the same file isn't very realistic and you
> > end up pathologically pounding on i_rwsem (which used to be until
> > recently i_mutex until Al's parallel lookup/readdir). Obviously write
> > lock stealing wins in this case.
> Except that it's DAX, and in 4.7-rc1 that used shared locking at the
> XFS level and never took exclusive locks.
>
> *However*, the DAX IO path locking in XFS  has changed in 4.9-rc1 to
> match the buffered IO single writer POSIX semantics - the test is a
> bad test based on the fact it exercised a path that is under heavy
> development and so can't be used as a regression test across
> multiple kernels.
>
> If you want to stress concurrent access to a single file, please
> use direct IO, not DAX or buffered IO.

Thanks for the update. I will change the test when I update this patch.

Cheers,
Longman