On Thu, 19 Jan 2006, Nick Piggin wrote:
>
> But actually it doesn't matter that we might touch page_count, only
> that we not clear PageLRU. So the enabler is simply moving the
> TestClearPageLRU after the get_page_testone.
One side note on your patch: the pure bit _test_ operation is very cheap,
but the "bit change" operation is very expensive (and not really any less
expensive than the "test-and-change" one).
So the patch to avoid "test_and_clear_bit()" really helps only if the test
usually results in us not doing the clear. Is that the case? Hmm..
So I _think_ that at least the case in "isolate_lru_page()", you'd
actually be better off doing the "test-and-clear" instead of separate
"test" and "clear-bit" ops, no? In that one, it would seem that 99+% of
the time, the bit is set (because we tested it just before getting the
lock).
No?
> I needed the de-skewing patch for something unrelated and it seemed that
> it opened the possibility for the following optimisations (ie. because
> we no longer touch a page after its refcount goes to zero).
>
> But actually it doesn't matter that we might touch page_count, only
> that we not clear PageLRU. So the enabler is simply moving the
> TestClearPageLRU after the get_page_testone.
Yes.
Now, that whole "we might touch the page count" thing does actually worry
me a bit. The locking rules are subtle (but they -seem- safe: before we
actually really put the page on the free-list in the freeing path, we'll
have locked the LRU list if it was on one).
But if you were to change _that_ one to a
atomic_add_unless(&page->counter, 1, -1);
I think that would be a real cleanup. And at that point I won't even
complain that "atomic_inc_test()" is faster - that "get_page_testone()"
thing is just fundamentally a bit scary, so I'd applaud it regardless.
(The difference: the "counter skewing" may be unexpected, but it's just a
simple trick. In contrast, the "touch the count after the page may be
already in the freeing stage" is a scary subtle thing. Even if I can't
see any actual bug in it, it just worries me in a way that offsetting a
counter by one does not..)
Linus
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[Index of Archives]
[Kernel Newbies]
[Netfilter]
[Bugtraq]
[Photo]
[Stuff]
[Gimp]
[Yosemite News]
[MIPS Linux]
[ARM Linux]
[Linux Security]
[Linux RAID]
[Video 4 Linux]
[Linux for the blind]
[Linux Resources]
