Messages in this thread | | | Date | Sun, 12 Apr 2009 12:54:16 +0200 | From | Ingo Molnar <> | Subject | Re: iptables very slow after commit 784544739a25c30637397ace5489eeb6e15d7d49 |
| |
* Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote:
> On Sat, Apr 11, 2009 at 09:08:54AM +0200, Ingo Molnar wrote: > > > > * Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote: > > > > > I will nevertheless suggest the following egregious hack to > > > get a consistent sample of one counter for some other CPU: > > > > > > a. Disable interrupts > > > b. Atomically exchange the bottom 32 bits of the > > > counter with the value zero. > > > c. Atomically exchange the top 32 bits of the counter > > > with the value zero. > > > d. Concatenate the values obtained in (b) and (c), which > > > is the snapshot value. > > > > Note, i have recently implemented full atomic64_t support on 32-bit > > x86, for the perfcounters code, based on the CMPXCHG8B instruction. > > > > Which, while not the lightest of instructions, is still much better > > than the sequence above. > > > > So i think a better approach would be to also add a dumb generic > > implementation for atomic64_t (using a global lock or so), and then > > generic code could just assume that atomic64_t always exists. > > > > It is far nicer - and faster as well - as the hack above, even on > > 32-bit x86. > > If the generic implementation is needed only on !SMP systems, that > could work. The architectures I would be worried about include > powerpc and ia64, which I believe support 32-bit SMP builds.
ia64 would naturally support the CMPXCHG8B instructions.
Not sure about powerpc32. Having a lock for the library implementation is not _that_ much of a problem. We obviously dont want the design of Linux to be dictated by the weakest link of all platforms, right?
Ingo
| |