Change in splice() behaviour after 5.10?

Change in splice() behaviour after 5.10?

[Terry Tritton]

Hi,
We've found a change in behaviour while testing the splice07 LTP test.
In versions before 5.10 the test will hang on certain combinations but after
5.10 the splice call will return.
I bisected the change to the following commit:
    36e2c7421f02a22f71c9283e55fdb672a9eb58e7
    fs: don't allow splice read/write without explicit ops

There has been some discussion on the LTP github page already:
    https://github.com/linux-test-project/ltp/issues/1156

From the github link these combinations fail on 5.4:
|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 |


PoC below, this program will hang before 36e2c7421f and complete after it.

#define _GNU_SOURCE
#include <fcntl.h>
#include <stdio.h>
#include <sys/epoll.h>
#include <unistd.h>

int main(){
    int fd_in[2];
    int fd_out;

    pipe(fd_in);
    fd_out = epoll_create(1);

    splice(fd_in[0], NULL, fd_out, NULL, 1, 0);

    printf("Should not hang!\n");

    return 0;
}

Is this change expected?

Re: Change in splice() behaviour after 5.10?

[Christoph Hellwig]

On Mon, Jun 03, 2024 at 10:59:15AM +0100, Terry Tritton wrote:
> Hi,
> We've found a change in behaviour while testing the splice07 LTP test.
> In versions before 5.10 the test will hang on certain combinations but after
> 5.10 the splice call will return.
> I bisected the change to the following commit:
>     36e2c7421f02a22f71c9283e55fdb672a9eb58e7
>     fs: don't allow splice read/write without explicit ops
> 
> There has been some discussion on the LTP github page already:
>     https://github.com/linux-test-project/ltp/issues/1156

In that case the return probably is an error because epoll doesn't
support read_iter/write_iter and thus completely expected.

If the underlying bug hasn't been fix in the mean time that probably
means it will be back if Jens' conversion of all misc file operations
to the iter based ones every gets merged.

If you are interested more in this please discuss it on the relevant
mailing lists instead of in private mail.

Re: Change in splice() behaviour after 5.10?

[Kees Cook]

On June 3, 2024 9:50:30 PM PDT, Christoph Hellwig <hch@lst.de> wrote:
>On Mon, Jun 03, 2024 at 10:59:15AM +0100, Terry Tritton wrote:
>> Hi,
>> We've found a change in behaviour while testing the splice07 LTP test.
>> In versions before 5.10 the test will hang on certain combinations but after
>> 5.10 the splice call will return.
>> I bisected the change to the following commit:
>>     36e2c7421f02a22f71c9283e55fdb672a9eb58e7
>>     fs: don't allow splice read/write without explicit ops
>> 
>> There has been some discussion on the LTP github page already:
>>     https://github.com/linux-test-project/ltp/issues/1156
>
>In that case the return probably is an error because epoll doesn't
>support read_iter/write_iter and thus completely expected.
>
>If the underlying bug hasn't been fix in the mean time that probably
>means it will be back if Jens' conversion of all misc file operations
>to the iter based ones every gets merged.
>
>If you are interested more in this please discuss it on the relevant
>mailing lists instead of in private mail.

Eh? LKML is in CC...

I've added fsdevel and Jens now too, though. 

-Kees

-- 
Kees Cook