On Wed, Oct 26, 2016 at 9:32 AM, Linus Torvalds
[off-list ref] wrote:

Quite frankly, I think the solution is to just rip out all the insane
zone crap.

IOW, something like the attached.

Advantage:

 - just look at the number of garbage lines removed!  21
insertions(+), 182 deletions(-)

 - it will actually speed up even the current case for all common
situations: no idiotic extra indirections that will take extra cache
misses

 - because the bit_wait_table array is now denser (256 entries is
about 6kB of data on 64-bit with no spinlock debugging, so ~100
cachelines), maybe it gets fewer cache misses too

 - we know how to handle the page_waitqueue contention issue, and it
has nothing to do with the stupid NUMA zones

The only case you actually get real page wait activity is IO, and I
suspect that hashing it out over ~100 cachelines will be more than
sufficient to avoid excessive contention, plus it's a cache-miss vs an
IO, so nobody sane cares.

The only reason it did that insane per-zone thing in the first place
that right now we access those wait-queues even when we damn well
shouldn't, and we have the solution for that.

Guys, holler if you hate this, but I think it's realistically the only
sane solution to the "wait queue on stack" issue.

Oh, and the patch is obviously entirely untested. I wouldn't want to
ruin my reputation by *testing* the patches I send out. What would be
the fun in that?

             Linus