On 08.02.2010, at 09:53, Avi Kivity wrote: > On 02/08/2010 12:02 AM, Alexander Graf wrote: >> >>> It's not a good idea for the kernel either, if it happens all the >>> time. If a typical Gekko application uses the fpu and the emulated >>> instructions intensively, performance will suck badly (as in: qemu/tcg >>> will be faster). >>> >>> >> Yeah, I haven't really gotten far enough to run full-blown guests yet. >> So far I'm on demos and they look pretty good. >> >> But as far as intercept speed goes - I just tried running this little >> piece of code in kvmctl: >> >> .global _start >> _start: >> li r3, 42 >> mtsprg 0, r3 >> mfsprg r4, 0 >> b _start >> >> and measured the amount of exits I get on my test machine: >> >> processor : 0 >> cpu : PPC970MP, altivec supported >> clock : 2500.000000MHz >> revision : 1.1 (pvr 0044 0101) >> >> ---> >> >> exits 1811108 >> >> I have no idea how we manage to get that many exits, but apparently we >> are. So I'm less concerned about the speed of the FPU rerouting at the >> moment. >> > > That's pretty impressive (never saw x86 with this exit rate) but it's more than 1000 times slower than the hardware, assuming 1 fpu IPC (and the processor can probably do more). An fpu intensive application will slow to a crawl. True. > >> If it really gets unusably slow, I'd rather binary patch the guest on >> the fly in KVM according to rules set by the userspace client. > > Is that even possible? Do those register-pair instructions and registers map 1:1 to 970 instructions and registers? Almost. Basically all I need to do is execute 2 FPU instructions instead of one for single instructions and paired single special instructions. So if I could patch the instruction to jump to some shared memory page, it'd become fast. At least as long as I figure out how to make sure we run with FP=0 in normal code, but with FP=1 in the special page ;). > >> But we'll >> get there when it turns out to be too slow. For now I'd rather like to >> have something working at all and then improve speed :-). >> > > Well, I want to see the light at the end of the tunnel first. Adding code is easy, ripping it out later not so much. Hum, so you suggest I get some real application running properly first so we can evaluate if it's fast enough? Alex-- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
