| Message ID | 20250221-mm-selftests-v2-8-28c4d66383c5@google.com |
|---|---|
| State | Superseded |
| Headers | show |
| Series | selftests/mm: Some cleanups from trying to run them | expand |
On Fri, 21 Feb 2025 at 19:26, Brendan Jackman <jackmanb@google.com> wrote: > > It seems that 9pfs does not allow truncating unlinked files, Mark Brown > has noted that NFS may also behave this way. I have not investigated at all but I _think_ over the weekend I saw ftruncate() failure on a QEMU guest where /tmp was tmpfs. Most likely explanation is user probably error (like /tmp wasn't actually tmpfs or the tmpfile() did not actually come from /tmp). But still, maybe just waving my hands and going "buggy filesystem, ignore" wasn't so smart. I should actually understand the problem before switching off tests. So, I think this and the other "ignore on weird filesystem" patch 9/9 should be ignored for the moment.
On Mon, 24 Feb 2025 at 11:25, Brendan Jackman <jackmanb@google.com> wrote: > > On Fri, 21 Feb 2025 at 19:26, Brendan Jackman <jackmanb@google.com> wrote: > > > > It seems that 9pfs does not allow truncating unlinked files, Mark Brown > > has noted that NFS may also behave this way. > > I have not investigated at all but I _think_ over the weekend I saw > ftruncate() failure on a QEMU guest where /tmp was tmpfs. > > Most likely explanation is user probably error (like /tmp wasn't > actually tmpfs or the tmpfile() did not actually come from /tmp). OK I double checked. It was in fact a 9p filesystem, turns out virtme-ng uses that unconditionally for its --rwdir/--rodir mounts even if the root is virtiofsd, and the tests operate on files in the CWD as well as in /tmp. I am still pondering ways to tackle this kinda problem more systematically but for now I think just skipping these tests is fine. > + ksft_exit_skip("ftruncate(fileno(tmpfile())) gave ENOENT, weird filesystem?"); This is missing a newline though, ditto in the subsequent patch. I will wait and see if any reviews come in before I send a fixup.
diff --git a/tools/testing/selftests/mm/map_populate.c b/tools/testing/selftests/mm/map_populate.c index 5c8a53869b1bd287b09a250edf628a66c25c2439..cdd88531dd9fe98f2d25edf4878659767c919e8e 100644 --- a/tools/testing/selftests/mm/map_populate.c +++ b/tools/testing/selftests/mm/map_populate.c @@ -87,6 +87,13 @@ int main(int argc, char **argv) BUG_ON(!ftmp, "tmpfile()"); ret = ftruncate(fileno(ftmp), MMAP_SZ); + if (ret < 0 && errno == ENOENT) { + /* + * This probably means tmpfile() made a file on a filesystem + * that doesn't handle temporary files the way we want. + */ + ksft_exit_skip("ftruncate(fileno(tmpfile())) gave ENOENT, weird filesystem?"); + } BUG_ON(ret, "ftruncate()"); smap = mmap(0, MMAP_SZ, PROT_READ | PROT_WRITE,
It seems that 9pfs does not allow truncating unlinked files, Mark Brown has noted that NFS may also behave this way. It doesn't seem quite right to call this a "bug" but it's probably a special enough case that it makes sense for the test to just SKIP if it happens. Signed-off-by: Brendan Jackman <jackmanb@google.com> --- tools/testing/selftests/mm/map_populate.c | 7 +++++++ 1 file changed, 7 insertions(+)