Messages in this thread | | | Date | Thu, 22 Dec 2016 11:28:31 -0800 (PST) | From | Hugh Dickins <> | Subject | Re: [RFC][PATCH] make global bitlock waitqueues per-node |
| |
On Wed, 21 Dec 2016, Linus Torvalds wrote: > On Wed, Dec 21, 2016 at 12:32 AM, Peter Zijlstra <peterz@infradead.org> wrote: > > > > FWIW, here's mine.. compiles and boots on a NUMA x86_64 machine. > > So I like how your patch is smaller, but your patch is also broken. > > First off, the whole contention bit is *not* NUMA-specific. It should > help non-NUMA too, by avoiding the stupid extra cache miss. > > Secondly, CONFIG_NUMA is a broken thing to test anyway, since adding a > bit for the NUMA case can overflow the page flags as far as I can tell > (MIPS seems to support NUMA on 32-bit, for example, but I didn't > really check the Kconfig details). Making it dependent on 64-bit might > be ok (and would fix the issue above - I don't think we really need to > care too much about 32-bit any more) > > But making it conditional at all means that now you have those two > different cases for this, which is a maintenance nightmare. So don't > do it even if we could say "screw 32-bit". > > Anyway, the conditional thing could be fixed by just taking Nick's > patch 1/2, and your patch (with the conditional bits stripped out).
Yup.
> > I do think your approach of just re-using the existing bit waiting > with just a page-specific waiting function is nicer than Nick's "let's > just roll new waiting functions" approach. It also avoids the extra > initcall. > > Nick, comments? > > Hugh - mind testing PeterZ's patch too? My comments about the > conditional PG_waiters bit and page bit overflow are not relevant for > your particular scenario, so you can ignore that part, and just take > PaterZ's patch directly.
Right, I put them both through some loads yesterday and overnight: Peter's patch and Nick's patch each work fine here, no issues seen with either (but I didn't attempt to compare them, aesthetically nor in performance).
Hugh
| |