On Mon, Jun 10, 2019 at 09:37:32AM +0300, Amir Goldstein wrote: > >Why do you think thhis is xfs_io fall back and not kernel fall back to >do_splice_direct()? Anyway, both cases allow read from swapfile >on upstream. Ah, I had assumed this was changed that was made because if you are implementing copy_file_range in terms of some kind of reflink-like mechanism, it becomes super-messy since you know have to break tons and tons of COW sharing each time the kernel swaps to the swap file. I didn't think we had (or maybe we did, and I missed it) a discussion about whether reading from a swap file should be prohibited. Personally, I think it's security theatre, and not worth the effort/overhead, but whatever.... my main complaint was with the unnecessary test failures with upstream kernels. > Trying to understand the desired flow of tests and fixes. > I agree that generic/554 failure may be a test/interface bug that > we should fix in a way that current upstream passes the test for > ext4. Unless there is objection, I will send a patch to fix the test > to only test copy *to* swapfile. > > generic/553, OTOH, is expected to fail on upstream kernel. > Are you leaving 553 in appliance build in anticipation to upstream fix? > I guess the answer is in the ext4 IS_IMMUTABLE patch that you > posted and plan to push to upstream/stable sooner than VFS patches. So I find it kind of annoying when tests land before the fixes do upstream. I still have this in my global_exclude file: # The proposed fix for generic/484, "locks: change POSIX lock # ownership on execve when files_struct is displaced" would break NFS # Jeff Layton and Eric Biederman have some ideas for how to address it # but fixing it is non-trivial generic/484 The generic/484 test landed in August 2018, and as far as I know, the issue which it is testing for *still* hasn't been fixed upstream. (There were issues raised with the proposed fix, and it looks like the people looking at the kernel change have lost interest.) The problem is that there are people who are trying to use xfstests to look for failures, and unless they start digging into the kernel archives from last year, they won't understand that generic/484 is a known failing test, and it will get fixed....someday. For people in the know, or for people who use my kvm-xfstests, gce-xfstests, it's not a big problem, since I've already blacklisted the test. But not everyone (and in fact, probably most people don't) use my front end scripts. For generic/553, I have a fix in ext4 so it will clear the failure, and that's fine, since I think we've all agreed in principle what the correct fix will be, and presumably it will get fixed soon. At that point, I might revert the commit from ext4, and rely on the VFS to catch the error, but the overhead of a few extra unlikely() tests aren't that big. But yeah, I did that mainly because unnecessary test failures because doing an ext4-specific fix didn't have many downsides, and one risk of adding tests to the global exclude file is that I then have to remember to remove it from the global exclude file when the issue is finally fixed upstream. > Do you think that should there be a different policy w.r.t timing of > merging xfstests tests that fail on upstream kernel? That's my opinion, and generic/484 is the best argument for why we should wait. Other people may have other opinions though, and I have a workaround, so I don't feel super-strong about it. (generic/454 is now the second test in my global exclude file. :-) At the very *least* there should be a comment in the test that fix is pending, and might not be applied yet, with a URL to the mailing list discussion. That will save effort when months (years?) go by, and the fix still hasn't landed the upstream kernel.... - Ted
