On Mon, Aug 10, 2020 at 04:52:17PM +0900, Tetsuo Handa wrote:

On 2020/08/10 12:19, Ming Lei wrote:

quoted

Block layer usually doesn't support or allow zero-length bvec. Since
commit 1bdc76aea115 ("iov_iter: use bvec iterator to implement
iterate_bvec()"), iterate_bvec() switches to bvec iterator. However,
Al mentioned that 'Zero-length segments are not disallowed' in iov_iter.

Fixes for_each_bvec() so that it can move on after seeing one zero
length bvec.

Signed-off-by: Ming Lei <redacted>
Link: https://www.mail-archive.com/linux-kernel@vger.kernel.org/msg2262077.html
Fixes: 1bdc76aea115 ("iov_iter: use bvec iterator to implement iterate_bvec()")

Is this Fixes: correct? That commit should be in RHEL8's 4.18 kernel but that kernel
does not hit this bug.

Yeah, it is correct, see the following link:

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?h=v5.8&id=1bdc76aea1159a750846c2fc98e404403eb7d51c

Commit 1bdc76aea115 was merged to v4.8, so it is definitely in both RHEL8's
4.18 based kernel and upstream kernel.

Moreover, maybe nobody cares, but behavior of splice() differs when there are only 
zero-length pages. With this fix, splice() returns 0 despite there is still pipe writers.

It is another new issue, which isn't related with Commit 1bdc76aea115,
see below.

Man page seems to say that splice() returns 0 when there is no pipe writers...

    A return value of 0 means end of input.  If fd_in refers to a pipe,
    then this means that there was no data to transfer, and it would not
    make sense to block because there are no writers connected to the
    write end of the pipe.

----- test case -----
#define _GNU_SOURCE
#include <stdio.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <unistd.h>
#include <string.h>

int main(int argc, char *argv[])
{
        static char buffer[4096];
        const int fd = open("/tmp/testfile", O_WRONLY | O_CREAT, 0600);
        int pipe_fd[2] = { EOF, EOF };
        pipe(pipe_fd);
        write(pipe_fd[1], NULL, sizeof(buffer));
        write(pipe_fd[1], NULL, sizeof(buffer));
        memset(buffer, 'a', sizeof(buffer));
        //write(pipe_fd[1], buffer, sizeof(buffer));
        write(pipe_fd[1], NULL, sizeof(buffer));
        write(pipe_fd[1], NULL, sizeof(buffer));
        memset(buffer, 'b', sizeof(buffer));
        //write(pipe_fd[1], buffer, sizeof(buffer));
        write(pipe_fd[1], NULL, sizeof(buffer));
        write(pipe_fd[1], NULL, sizeof(buffer));
        memset(buffer, 'c', sizeof(buffer));
        //write(pipe_fd[1], buffer, sizeof(buffer));
        write(pipe_fd[1], NULL, sizeof(buffer));
        write(pipe_fd[1], NULL, sizeof(buffer));
        memset(buffer, 'd', sizeof(buffer));
        //write(pipe_fd[1], buffer, sizeof(buffer));
        write(pipe_fd[1], NULL, sizeof(buffer));
        write(pipe_fd[1], NULL, sizeof(buffer));
        splice(pipe_fd[0], NULL, fd, NULL, 65536, 0);
        return 0;
}

The above test doesn't trigger the reported lockup issue, so this patch
isn't related with the new issue you described.

----- 4.18.0-193.14.2.el8_2.x86_64 -----
openat(AT_FDCWD, "/tmp/testfile", O_WRONLY|O_CREAT, 0600) = 3
pipe([4, 5])                            = 0
write(5, NULL, 4096)                    = -1 EFAULT (Bad address)
write(5, NULL, 4096)                    = -1 EFAULT (Bad address)
write(5, NULL, 4096)                    = -1 EFAULT (Bad address)
write(5, NULL, 4096)                    = -1 EFAULT (Bad address)
write(5, NULL, 4096)                    = -1 EFAULT (Bad address)
write(5, NULL, 4096)                    = -1 EFAULT (Bad address)
write(5, NULL, 4096)                    = -1 EFAULT (Bad address)
write(5, NULL, 4096)                    = -1 EFAULT (Bad address)
write(5, NULL, 4096)                    = -1 EFAULT (Bad address)
write(5, NULL, 4096)                    = -1 EFAULT (Bad address)
splice(4, NULL, 3, NULL, 65536, 0

^C)      = ? ERESTARTSYS (To be restarted if SA_RESTART is set)
strace: Process 1486 detached

The same behavior can be observed on v4.8 too, both v4.8 and v4.18
includes 1bdc76aea115. If you apply the fix against v4.8, you can
observe the same behavior too.

----- linux.git + this fix -----

It should have been linux.git, :-)

I think this new issue may be introduced between v4.18 and v5.8.

OK.


Thanks,
Ming