On Mon, Mar 7, 2011 at 8:24 AM, Linus Torvalds
<torva...@linux-foundation.org> wrote:
>
> and I think one reason why the race is hard to get rid of is simply
> that system call return is _the_ common point of signal handling in
> UNIX (technically, obviously any return to user space, but there are
> no appreciable interrupts etc going on, and there _are_ a lot of
> system calls). The above trace is one that my patch would have handled
> correctly (it has no EINTR).
Linux also ends up making this race easier to see probably because all
system calls that are interruptible by signals are all "greedy": they
try to do as much real work as possible, rather than return EINTR.
That means that if there is work pending (like characters in a tty
buffer for "read()", or a child that has exited for "wait()"), then
system calls under Linux (and likely all other Unixes too, but Linux
is the one I can guarantee works this way) will always do as much real
work as possible, and return that real work rather than return with
EINTR.
So the _common_ case will be:
- the system call returns with success ("read a few characters" or
"found this child")
- but the signal handler will be executed immediately at the return
point, so user space won't really even "see" the success before the
signal handler is executed.
In other words, when you do
waiting_for_child++;
pid = WAITPID (-1, &status, waitpid_flags);
waiting_for_child--;
even if the "waiting_for_child--" were to be compiled to be one single
instruction, and at the exact return point of the system call, the
signal handler would still happen right in between the system call
return and that instruction.
Linus
