On Fri, Sep 07, 2018 at 04:54:59PM +0200, David Sterba wrote: > The series peels off the custom locking that's used for dev-replace and > uses read-write semaphore in the end. > > I've mainly focused on correctness and haven't measured the performance > effects. There should be none as the blocking and waiting was merely > open coding what the rw semaphore does, but without the fairness. The > overall number of locks taken is low, there's a lots of IO in between so > even if new scheme is slower, I don't expect any dramatic change.
Mixed results. The btrfs/011 is a test that usually takes long time (around 1000 sec on my test box) and I have a long list of run times to compare. There this patchset does not show any problems. However on a VM with 8 CPUs, this goes from similar times (1000 sec) to several hours, without apparent reason. The is other low activity on the system, but so this is with testing other patchsets and the times are not that bad. So, I'm going to merge the first part of the patchset that does not switch the locking primitives to the semaphores, more analysis needed.
