splice07 fails on kernels <5.10

[edliaw]

splice07 seems to depend on torvalds/linux@36e2c74 ("fs: don't allow splice read/write without explicit ops") to pass. I've compiled a list of failing combinations on 5.4, but I'm not sure what's the right approach to modifying the test:

in_fd	out_fd	error
TST_FD_PIPE_READ	TST_FD_EPOLL	hangs
TST_FD_PIPE_READ	TST_FD_EVENTFD	hangs
TST_FD_PIPE_READ	TST_FD_SIGNALFD	hangs
TST_FD_PIPE_READ	TST_FD_TIMERFD	hangs
TST_FD_PIPE_READ	TST_FD_PIDFD	hangs
TST_FD_PIPE_READ	TST_FD_PERF_EVENT	hangs
TST_FD_PIPE_READ	TST_FD_IO_URING	hangs
TST_FD_PIPE_READ	TST_FD_BPF_MAP	hangs
TST_FD_PIPE_READ	TST_FD_FSOPEN	hangs
TST_FD_PIPE_READ	TST_FD_FSPICK	hangs
TST_FD_INOTIFY	TST_FD_PIPE_WRITE	hangs
TST_FD_DIR	TST_FD_PIPE_WRITE	EISDIR
TST_FD_PERF_EVENT	TST_FD_PIPE_WRITE	ENODATA
TST_FD_FSOPEN	TST_FD_PIPE_WRITE	ENODATA
TST_FD_FSPICK	TST_FD_PIPE_WRITE	ENODATA

Should it just be gated by kernel version or is there a preferred way to modify it so that it is backwards compatible?

[pevik]

Thanks for a detailed table.

@metan-ucw @mdoucha know more than me. AFAIK we noticed at least incorrect errno EISDIR (obviously there is also ENODATA) and hang on TST_FD_EPOLL (I did not know about the following hangs), although AFAIK we did not reported it to the kernel mainline ML (lore).

splice07 seems to depend on torvalds/linux@36e2c74 ("fs: don't allow splice read/write without explicit ops") to pass.

Yes (https://lore.kernel.org/ltp/ZbO-Pl9S7KH2cKkb@yuki/).

FYI before the latest release we whitelisted 2 cases: (c042185, e5970b2) and had a plan to add tests which actually test
splice() on /dev/zero and /proc/self/maps for kernels between v4.4 and v5.3 where it was working (https://lore.kernel.org/ltp/20240126132046.GA508599@pevik/).

[mdoucha]

The right approach would be to fix the kernel since all these failures are kernel bugs (maybe except for inotify -> writable pipe). It does not make sense to modify the test, not even to add a kernel version check.

[pevik]

The right approach would be to fix the kernel since all these failures are kernel bugs (maybe except for inotify -> writable pipe). It does not make sense to modify the test, not even to add a kernel version check.

5.10 got fixed by 36e2c7421f02 ("fs: don't allow splice read/write without explicit ops"). I managed to backport necessary commits on the top of 5.5, but I was not sure if I did not introduce any regression, thus I did not post it.

I was thinking to backport necessary changes to 5.4 LTS kernel, but that would require quite a lot of work (much more commits to backport because the code changed a lot). That is the reason why nobody thought about backporting the code (also some would consider it as a new feature not a fix).