On Mon, Jun 17, 2019 at 9:58 PM Guoqing Jiang <gqjiang@xxxxxxxx> wrote: > > > > On 6/18/19 12:41 PM, Song Liu wrote: > > On Mon, Jun 17, 2019 at 8:41 PM Guoqing Jiang <gqjiang@xxxxxxxx> wrote: > > <snip> > > > >>>> +}; > >>> Have we measured the performance overhead of this? > >>> The linear search for every IO worries me. > >> From array's view, I think the performance will not be impacted, > >> because write IO is complete > >> after it reached all the non-writemostly devices. > >> > > Hmm... How about the cpu utilization rate? Have you got chance > > to do some simple benchmarking? > > > > I can't see the impact of cpu in simple test, because it really depends on > how slow the writemostly device is. > > And the modern multi-queue device (also flagged as writemostly) would > be fast enough to handle the write-behind IO in time, which means there > should only a few (or zero) elements in wb_list, but it is a potential issue > which need to be addressed. Yeah, this makes sense. We can keep the linear search for now. Thanks, Song
