From: Avi Kivity on 25 Oct 2009 08:30 On 09/16/2009 03:45 PM, Avi Kivity wrote: > The various x86 syscall related MSRs (MSR_LSTAR and friends, EFER when SCE > needs to be updated) are fairly expensive to read or write. Since different > operating systems can set different values for these MSRs, KVM needs to reload > them when switching to a different guest or to the host. > > Switching on every guest entry/exit is too expensive, so KVM reloads on > guest preemption, which is a lot rarer. Even so, preemption based reload > is suboptimal: > > - if we're switching to a kernel thread and back, there's no need to reload > the MSRs. Examples of kernel threads we're likely to switch to are: > > - the idle task > - a threaded interrupt handler > - a kernel-mode virtio server (vhost-net) > > - if we're switching to a guest running the same OS, the MSRs will have the > same values and there's no need to reload them > > - if the guest and host run the same OS, again the MSRs need not be reloaded. > > This patchset implements just-in-time reloads to defer them to the last > possible instant. When we do reload, we check whether the values have in > fact changed and reload conditionally. > > For the just-in-time reloads the first patch implements "user return > notifiers", a callback invoked just before return to userspace. This has > been written so that there is no code impact if KVM is not configured, and > no runtime impact if KVM is not running. > > The patchset improves guest/idle/guest switches by about 2000 cycles. > > I've applied this to kvm.git master. -- error compiling committee.c: too many arguments to function -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
|
Pages: 1 Prev: Compiling 2.6.31/31.5 Next: Broken ondemand scheduler in Linux 2.6.30+ on Pentium IVs |