Front page | perl.perl5.porters |
Postings from February 2012
Zefram <zefram@fysh.org> wrote:
> static void THX_wrap_op_checker(pTHX_ Optype opcode,
> Perl_check_t new_checker, Perl_check_t *old_checker_p)
> {
> if(*old_checker_p) return;
> OP_REFCNT_LOCK;
> if(!*old_checker_p) {
> *old_checker_p = PL_check[opcode];
> PL_check[opcode] = new_checker;
> }
> OP_REFCNT_UNLOCK;
> }
Do we support any platforms where an access to a Perl_check_t might
not be atomic? (For that matter, do any such platforms exist? Since
it's a function pointer, that would presumably mean 64-bit platforms
which only do atomic memory accesses of up to 32 bits.)
If so, I think the OP_REFCNT_LOCK needs to precede the initial
test/return, to avoid this race:
1. Thread A loads the first half of *old_checker_p, which is all-bits-zero
2. Thread B takes the lock
3. Thread B writes 64 bits into *old_checker_p; the second half
happens to be all-bits-zero
4. Thread B releases the lock
5. Thread A loads the second half of the new value of *old_checker_p
(all-bits-zero again)
6. Thread A thinks the full value was 0, so continues; but it wasn't,
so it should have returned
Or am I misunderstanding something?
--
Aaron Crane ** http://aaroncrane.co.uk/
Thread Previous
|
Thread Next