Messages in this thread |  | | From | Andy Lutomirski <> | | Date | Fri, 8 Apr 2016 08:58:27 -0700 | | Subject | Re: [RFC PATCH 0/3] restartable sequences v2: fast user-space percpu critical sections |
| |
On Apr 7, 2016 11:41 PM, "Peter Zijlstra" <peterz@infradead.org> wrote: > > On Thu, Apr 07, 2016 at 09:43:33AM -0700, Andy Lutomirski wrote: > > enter the critical section: > > 1: > > movq %[cpu], %%r12 > > movq {address of counter for our cpu}, %%r13 > > movq {some fresh value}, (%%r13) > > cmpq %[cpu], %%r12 > > jne 1b > > This is inherently racy; your forgot the detail of 'some fresh value', > but since you want to avoid collisions you really want an increment. > > But load-store archs cannot do that. Or rather, they need to do: > > load Rn, $event > add Rn, Rn, 1 > store $event, Rn > > But if they're preempted in the middle, two threads will collide and > generate the _same_ increment. Comparing CPU numbers will not fix that.
Even on x86 this won't work -- we have no actual guarantee we're on the right CPU, so we'd have to use an atomic.
I was thinking we'd allocate from a per-thread pool (say 24 bits of thread ID and the rest being a nonce). On load-store architectures this wouldn't be async-signal-safe, though. Hmm.
--Andy
|  |