On Tue, Jun 09, 2020 at 06:38:08PM -0700, Matthew Wilcox wrote: > On Tue, Jun 09, 2020 at 08:10:36PM -0400, Kent Overstreet wrote: > > Convert generic_file_buffered_read() to get pages to read from in > > batches, and then copy data to userspace from many pages at once - in > > particular, we now don't touch any cachelines that might be contended > > while we're in the loop to copy data to userspace. > > > > This is is a performance improvement on workloads that do buffered reads > > with large blocksizes, and a very large performance improvement if that > > file is also being accessed concurrently by different threads. > > Hey, you're stealing my performance improvements! :) > Granted, I haven't got to doing performance optimisations (certainly > not in this function), but this is one of the places where THP in the > page cache will have a useful performance improvement. > > I'm not opposed to putting this in, but I may back it out as part of > the THP work because the THPs will get the same performance improvements > that you're seeing here with less code. I'm an _enthusiastic_ supporter of the THP stuff (as you know), but my feeling is that it's going to be a long time before hugepages are everywhere - and I think even with the pagevec stuff generic_file_buffered_read() is somewhat easier to read and deal with after this series than before it. Though I could see the pagevec stuff making hugepage support a pain, so there is that. Eh.
