| Message ID | 20240430193157.419425-1-seanjc@google.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | KVM: Fold kvm_arch_sched_in() into kvm_arch_vcpu_load() | expand |
On Tue, Apr 30, 2024 at 12:31:53PM -0700, Sean Christopherson wrote: > Drop kvm_arch_sched_in() and instead pass a @sched_in boolean to > kvm_arch_vcpu_load(). > > While fiddling with an idea for optimizing state management on AMD CPUs, > I wanted to skip re-saving certain host state when a vCPU is scheduled back > in, as the state (theoretically) shouldn't change for the task while it's > scheduled out. Actually doing that was annoying and unnecessarily brittle > due to having a separate API for the kvm_sched_in() case (the state save > needed to be in kvm_arch_vcpu_load() for the common path). > > E.g. I could have set a "temporary"-ish flag somewhere in kvm_vcpu, but (a) > that's gross and (b) it would rely on the arbitrary ordering between > sched_in() and vcpu_load() staying the same. Another option would be to change the rules around kvm_arch_sched_in() where the callee is expected to load the vCPU context. The default implementation could just call kvm_arch_vcpu_load() directly and the x86 implementation can order things the way it wants before kvm_arch_vcpu_load(). I say this because ... > The only real downside I see is that arm64 and riscv end up having to pass > "false" for their direct usage of kvm_arch_vcpu_load(), and passing boolean > literals isn't ideal. But that can be solved by adding an inner helper that > omits the @sched_in param (I almost added a patch to do that, but I couldn't > convince myself it was necessary). Needing to pass @sched_in for other usage of kvm_arch_vcpu_load() hurts readability, especially when no other architecture besides x86 cares about it.
On Wed, May 01, 2024, Oliver Upton wrote: > On Tue, Apr 30, 2024 at 12:31:53PM -0700, Sean Christopherson wrote: > > Drop kvm_arch_sched_in() and instead pass a @sched_in boolean to > > kvm_arch_vcpu_load(). > > > > While fiddling with an idea for optimizing state management on AMD CPUs, > > I wanted to skip re-saving certain host state when a vCPU is scheduled back > > in, as the state (theoretically) shouldn't change for the task while it's > > scheduled out. Actually doing that was annoying and unnecessarily brittle > > due to having a separate API for the kvm_sched_in() case (the state save > > needed to be in kvm_arch_vcpu_load() for the common path). > > > > E.g. I could have set a "temporary"-ish flag somewhere in kvm_vcpu, but (a) > > that's gross and (b) it would rely on the arbitrary ordering between > > sched_in() and vcpu_load() staying the same. > > Another option would be to change the rules around kvm_arch_sched_in() > where the callee is expected to load the vCPU context. > > The default implementation could just call kvm_arch_vcpu_load() directly > and the x86 implementation can order things the way it wants before > kvm_arch_vcpu_load(). > > I say this because ... > > > The only real downside I see is that arm64 and riscv end up having to pass > > "false" for their direct usage of kvm_arch_vcpu_load(), and passing boolean > > literals isn't ideal. But that can be solved by adding an inner helper that > > omits the @sched_in param (I almost added a patch to do that, but I couldn't > > convince myself it was necessary). > > Needing to pass @sched_in for other usage of kvm_arch_vcpu_load() hurts > readability, especially when no other architecture besides x86 cares > about it. Yeah, that bothers me too. I tried your suggestion of having x86's kvm_arch_sched_in() do kvm_arch_vcpu_load(), and even with an added kvm_arch_sched_out() to provide symmetry, the x86 code is kludgy, and even the common code is a bit confusing as it's not super obvious that kvm_sched_{in,out}() is really just kvm_arch_vcpu_{load,put}(). Staring a bit more at the vCPU flags we have, adding a "bool scheduled_out" isn't terribly gross if it's done in common code and persists across load() and put(), i.e. isn't so blatantly a temporary field. And because it's easy, it could be set with WRITE_ONCE() so that if it can be read cross-task if there's ever a reason to do so. The x86 code ends up being less ugly, and adding future arch/vendor code for sched_in() *or* sched_out() requires minimal churn, e.g. arch code doesn't need to override kvm_arch_sched_in(). The only weird part is that vcpu->preempted and vcpu->ready have slightly different behavior, as they are cleared before kvm_arch_vcpu_load(). But the weirdness is really with those flags no having symmetry, not with scheduled_out itself. Thoughts? static void kvm_sched_in(struct preempt_notifier *pn, int cpu) { struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn); WRITE_ONCE(vcpu->preempted, false); WRITE_ONCE(vcpu->ready, false); __this_cpu_write(kvm_running_vcpu, vcpu); kvm_arch_vcpu_load(vcpu, cpu); WRITE_ONCE(vcpu->scheduled_out, false); } static void kvm_sched_out(struct preempt_notifier *pn, struct task_struct *next) { struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn); WRITE_ONCE(vcpu->scheduled_out, true); if (current->on_rq) { WRITE_ONCE(vcpu->preempted, true); WRITE_ONCE(vcpu->ready, true); } kvm_arch_vcpu_put(vcpu); __this_cpu_write(kvm_running_vcpu, NULL); }
On Wed, May 01, 2024 at 07:28:21AM -0700, Sean Christopherson wrote: > On Wed, May 01, 2024, Oliver Upton wrote: > > On Tue, Apr 30, 2024 at 12:31:53PM -0700, Sean Christopherson wrote: > > > Drop kvm_arch_sched_in() and instead pass a @sched_in boolean to > > > kvm_arch_vcpu_load(). > > > > > > While fiddling with an idea for optimizing state management on AMD CPUs, > > > I wanted to skip re-saving certain host state when a vCPU is scheduled back > > > in, as the state (theoretically) shouldn't change for the task while it's > > > scheduled out. Actually doing that was annoying and unnecessarily brittle > > > due to having a separate API for the kvm_sched_in() case (the state save > > > needed to be in kvm_arch_vcpu_load() for the common path). > > > > > > E.g. I could have set a "temporary"-ish flag somewhere in kvm_vcpu, but (a) > > > that's gross and (b) it would rely on the arbitrary ordering between > > > sched_in() and vcpu_load() staying the same. > > > > Another option would be to change the rules around kvm_arch_sched_in() > > where the callee is expected to load the vCPU context. > > > > The default implementation could just call kvm_arch_vcpu_load() directly > > and the x86 implementation can order things the way it wants before > > kvm_arch_vcpu_load(). > > > > I say this because ... > > > > > The only real downside I see is that arm64 and riscv end up having to pass > > > "false" for their direct usage of kvm_arch_vcpu_load(), and passing boolean > > > literals isn't ideal. But that can be solved by adding an inner helper that > > > omits the @sched_in param (I almost added a patch to do that, but I couldn't > > > convince myself it was necessary). > > > > Needing to pass @sched_in for other usage of kvm_arch_vcpu_load() hurts > > readability, especially when no other architecture besides x86 cares > > about it. > > Yeah, that bothers me too. > > I tried your suggestion of having x86's kvm_arch_sched_in() do kvm_arch_vcpu_load(), > and even with an added kvm_arch_sched_out() to provide symmetry, the x86 code is > kludgy, and even the common code is a bit confusing as it's not super obvious > that kvm_sched_{in,out}() is really just kvm_arch_vcpu_{load,put}(). > > Staring a bit more at the vCPU flags we have, adding a "bool scheduled_out" isn't > terribly gross if it's done in common code and persists across load() and put(), > i.e. isn't so blatantly a temporary field. And because it's easy, it could be > set with WRITE_ONCE() so that if it can be read cross-task if there's ever a > reason to do so. > > The x86 code ends up being less ugly, and adding future arch/vendor code for > sched_in() *or* sched_out() requires minimal churn, e.g. arch code doesn't need > to override kvm_arch_sched_in(). > > The only weird part is that vcpu->preempted and vcpu->ready have slightly > different behavior, as they are cleared before kvm_arch_vcpu_load(). But the > weirdness is really with those flags no having symmetry, not with scheduled_out > itself. > > Thoughts? Yeah, this seems reasonable. Perhaps scheduled_out could be a nice hint for guardrails / sanity checks in the future.