On Fri, 2019-09-13 at 17:31 +0100, Pádraig Brady wrote: > On 13/09/19 11:08, Jeff Layton wrote: > > On Thu, 2019-09-12 at 11:58 -0600, Andreas Dilger wrote: > > > Should this have a runtime fallback to stat() if statx() is not > > > implemented > > > on the running kernel, or is that already handled at another level? > > > > > > > glibc already does this fallback. > > Note we need to be portable to a lot more than glibc >
To my knowledge, only glibc has a statx wrapper so far, so others should just end up using the non-statx-enabled codepaths. I'd argue that a libc implementation that provided a statx that didn't fall back when run on an legacy kernel could be considered broken. > In general this patchset looks really useful, > and the performance testing on ceph is really appreciated. Yeah, it's always nice to see when these sorts of patches make a marked difference. While I didn't measure it, I suspect that the "ls" commands themselves also ran faster in this test, since they wouldn't have had to make as many round trips to the MDS. Note that I expect this set may also help on NFS, and Andreas said this should help Lustre too. Not sure about other netfs' yet, but I wouldn't expect this to perform any worse than the stat() variant does anywhere. Thanks, -- Jeff Layton <jlay...@kernel.org>