On Fri, 28 Mar 2025 12:15:47 +0200 Jaco Kroon <jaco@xxxxxxxxx> wrote: > Hi All, > > I've not seen feedback on this, please may I get some direction on this? The only thing I can think of is that the longer loop might affect scheduling latencies. David > > Kind regards, > Jaco > > On 2025/03/15 00:16, Jaco Kroon wrote: > > This is a follow up to the attempt made a while ago. > > > > Whist the patch worked, newer kernels have moved from pages to folios, > > which gave me the motivation to implement the mechanism based on the > > userspace buffer size patch that Miklos supplied. > > > > That patch works as is, but I note there are changes to components > > (overlayfs and exportfs) that I've got very little experience with, and > > have not tested specifically here. They do look logical. I've marked > > Miklos as the Author: here, and added my own Signed-off-by - I hope this > > is correct. > > > > The second patch in the series implements the changes to fuse's readdir > > in order to utilize the first to enable reading more than one page of > > dent structures at a time from userspace, I've included a strace from > > before and after this patch in the commit to illustrate the difference. > > > > To get the relevant performance on glusterfs improved (which was > > mentioned elsewhere in the thread) changes to glusterfs to increase the > > number of cached dentries is also required (these are pushed to github > > but not yet merged, because similar to this patch, got stalled before > > getting to the "ready for merge" phase even though it's operational). > > > > Please advise if these two patches looks good (I've only done relatively > > basic testing now, and it's not running on production systems yet) > > > > Kind regards, > > Jaco > > >
