splice(-> FIFO) never wakes up inotify IN_MODIFY?

[Ahelenia Ziemiańska <nabijaczleweli@xxxxxxxxxxxxxxxxxx>]

Hi!

Consider the following programs:
-- >8 --
==> ino.c <==
#define _GNU_SOURCE
#include <stdio.h>
#include <sys/inotify.h>
#include <unistd.h>
int main() {
  int ino = inotify_init1(IN_CLOEXEC);
  inotify_add_watch(ino, "/dev/fd/0", IN_MODIFY);

  char buf[64 * 1024];
  struct inotify_event ev;
  while (read(ino, &ev, sizeof(ev)) > 0) {
    fprintf(stderr, "%d: mask=%x, cook=%x, len=%x, name=%.*s\n", ev.wd, ev.mask,
            ev.cookie, ev.len, (int)ev.len, ev.name);
    fprintf(stderr, "rd=%zd\n", read(0, buf, sizeof(buf)));
  }
}

==> se.c <==
#define _GNU_SOURCE
#include <stdio.h>
#include <sys/sendfile.h>
int main() {
  ssize_t rd, acc = 0;
  while ((rd = sendfile(1, 0, 0, 128 * 1024 * 1024)) > 0)
    acc += rd;
  fprintf(stderr, "se=%zd: %m\n", acc);
}

==> sp.c <==
#define _GNU_SOURCE
#include <fcntl.h>
#include <stdio.h>
int main() {
  ssize_t rd, acc = 0;
  while ((rd = splice(0, 0, 1, 0, 128 * 1024 * 1024, 0)) > 0)
    acc += rd;
  fprintf(stderr, "sp=%zd: %m\n", acc);
}
-- >8 --

By all means, ./sp | ./ino and ./se | ./ino should be equivalent,
right?

-- >8 --
$ make se sp ino
$ mkfifo fifo
$ ./ino < fifo &
[1] 230
$ echo a > fifo
$ echo a > fifo
1: mask=2, cook=0, len=0, name=
rd=4
$ echo c > fifo
1: mask=2, cook=0, len=0, name=
rd=2
$ ./se > fifo
abcdef
1: mask=2, cook=0, len=0, name=
asd
^D
se=11: Success
rd=11
1: mask=2, cook=0, len=0, name=
rd=0
$ ./sp > fifo
abcdefg
asd
dsasdadadad
sp=24: Success
$ < sp ./sp > fifo
sp=25856: Success
$ < sp ./sp > fifo
^C
$ echo sp > fifo
^C
-- >8 --

Note how in all ./sp > fifo cases, ./ino doesn't wake up!
Note also how, thus, we've managed to fill the pipe buffer with ./sp
(when it transferred 25856), and now we can't /ever/ write there again
(both splicing and normal writes block, since there's no space left in
 the pipe; ./ino hasn't seen this and will never wake up or service the
 pipe):
so we've effectively "denied service" by slickily using a different
syscall to do the write, right?

I consider this to be unexpected behaviour because (a) obviously and
(b) sendfile() sends the inotify event.

Happens on my linus checkout (v6.4-rc7-234-g547cc9be86f4) and bookworm (6.1.27-1).

Attachment: signature.asc
Description: PGP signature