Date: Wed, 27 Aug 2014 12:47:46 -0400 From: Rich Felker <dalias@...c.org> To: musl@...ts.openwall.com Subject: Re: Multi-threaded performance progress On Wed, Aug 27, 2014 at 11:53:10AM +0200, Jens Gustedt wrote: > Am Dienstag, den 26.08.2014, 17:36 -0400 schrieb Rich Felker: > > Fixing it should be trivial via the design I mentioned earlier: don't > > use a waiter flag like this, but instead offset the initial value of > > ref by +1 and a_dec it just before waiting. As in other places, of > > course, a wake to an invalid address is possible either way; this is > > "fixable" if necessary via FUTEX_WAKE_OP (having the kernel do the > > atomic dec after acquiring the futex bin locks). > > generally it would be nice to have such a lock functionality that > takes care of the waiters inside the int itself, this could perhaps be > used in other places Yes. My idea was to have a "__wake_store" function or similar that wraps FUTEX_WAKE_OP and does a fallback to FUTEX_WAKE if the kernel lacks FUTEX_WAKE_OP (if there are older kernels that lack it; not sure). Note that this is less powerful that what I mentioned above, but you can know before the above a_dec if it will write zero since, if the value is 1, you're the last thread to modify it. I think most places where FUTEX_WAKE_OP would be used fit this pattern -- you know in advance that you're not racing with other atomic writers (or if you are, they're using a_cas and their cas would fail with the value you're writing and with the value you're overwriting). Rich
Powered by blists - more mailing lists
Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.