On 10/03/2017 10:06 AM, Matthew Wilcox wrote: > On Fri, Sep 29, 2017 at 02:17:32AM +0200, Jens Axboe wrote: >> On 09/28/2017 11:44 PM, Linus Torvalds wrote: >>> On Thu, Sep 28, 2017 at 2:41 PM, Andrew Morton >>> <akpm@xxxxxxxxxxxxxxxxxxxx> wrote: >>>> >>>> test_and_set_bit()? >>> >>> If there aren't any atomicity concerns (either because of higher-level >>> locking, or because racing and having two people set the bit is fine), >>> it can be better to do them separately if the test_bit() is the common >>> case and you can avoid dirtying a cacheline that way. >>> >>> But yeah, if that is the case, it might be worth documenting, because >>> test_and_set_bit() is the more obviously appropriate "there can be >>> only one" model. >> >> It is documented though, but maybe not well enough... >> >> I've actually had to document/explain it enough times now, that it >> might be worth making a general construct. Though it has to be >> used carefully, so perhaps it's better contained as separate use >> cases. > > test_and_test_and_set_bit()? It's an unusual name, so when either > reading it or writing it, people are going to say "something unusual > here", rather than "That Jens Axboe is such a n00b, he doesn't know how > to use test_and_set_bit()". There are a few references out on the web > to test-and-test-and-set already, so it's not entirely unique to Linux. > > Plus, some architectures might be able to optimise that, particularly > those which are ll/sc based. It might be exactly the same as their > test_and_set(). I like that suggestion, but would suggest we make it test_then_test_and_set_bit() since the 'then' naming would work for having similar test_then_clear_bit() and not clash with test_and_set_bit(). And yes, some archs would be able to optimize this nicely. All worth it if I never have to explain it or add special comments about it again :-) -- Jens Axboe
