illumos gate

While investigating a new test failure with Python 3.10, I found that our eventfd_read() function does not return a failure when the underlying read() returns -1 due to an unsigned comparison.

Here's a simple reproducer. With a non-blocking semaphore with an initial value of 2, the third read should fail with EAGAIN. It does not.

#include <stdio.h>
#include <sys/eventfd.h>
#include <errno.h>
#include <assert.h>

int
main()
{
        int fd = eventfd(2, EFD_NONBLOCK | EFD_CLOEXEC | EFD_SEMAPHORE);
        assert(fd >= 0);

        for (uint_t i = 0; i < 3; i++) {
                eventfd_t v = 0xdeadbeef;
                int ret = eventfd_read(fd, &v);
                printf("Ret: %d (errno %d, v %p)\n", ret, errno, v);
        }
}

Here's the output:

Ret: 0 (errno 0, v 1)
Ret: 0 (errno 0, v 1)
Ret: 0 (errno 11, v deadbeef)

With a bit of dtrace, one can see that the underlying read() is returning -1, and yet eventfd_read() is not:

%  pfexec dtrace -n 'pid$target::eventfd_read:return,syscall::read:return{trace(arg1)}' -c ./eventfd
  3   1338                      read:return                 8
  3  75386              eventfd_read:return                 0
  3   1338                      read:return                 8
  3  75386              eventfd_read:return                 0
  3   1338                      read:return                -1
  3  75386              eventfd_read:return                 0