On Fri, 2006-09-22 at 14:32 +0200, Andi Kleen wrote: > BTW I changed my copy sorry. I redid the early PDA support > to not be in assembler. > > On Fri, Sep 22, 2006 at 09:59:45PM +1000, Rusty Russell wrote: > > This patch actually uses the gs register to implement the per-cpu > > sections. It's fairly straightforward: the gs segment starts at the > > per-cpu offset for the particular cpu (or 0, in very early boot). > > > > We also implement x86_64-inspired (via Jeremy Fitzhardinge) per-cpu > > accesses where a general lvalue isn't needed. These > > single-instruction accesses are slightly more efficient, plus (being a > > single insn) are atomic wrt. preemption so we can use them to > > implement cpu_local_inc etc. > > The problem is nobody uses cpu_local_inc() etc :/ And it is difficult > to use in generic code because of the usual preemption issues > (and being slower on other archs in many cases compared to preempt disabling > around larger block of code) True, they were primarily designed for the SNMP counters in networking (as per your comment in snmp.h). They're now dynamic per-cpu pointers, which adds a new wrinkle though 8( > Without that it is the same code as Jeremy's variant > %gs memory reference + another reference with offset as far as I can see. > > So while it looks nice I don't think it will have advantages. Or did > i miss something? Mainly that it makes more sense to use the existing per-cpu concept than introduce another kind of per-cpu var within a special structure, but it's also more efficient (see other post). Hopefully it will spark interest in making dynamic-percpu pointers use the same offset scheme, now x86 will experience the benefits. And we might even get a third user of local_t! Cheers, Rusty. -- Help! Save Australia from the worst of the DMCA: http://linux.org.au/law
