Messages in this thread | | | Subject | Re: [PATCH v1 2/5] KVM: arm/arm64: Adjust entry/exit and trap related tracepoints | From | Zenghui Yu <> | Date | Thu, 13 Jun 2019 19:28:10 +0800 |
| |
Hi James,
On 2019/6/12 20:49, James Morse wrote: > Hi, > > On 12/06/2019 10:08, Zenghui Yu wrote: >> Currently, we use trace_kvm_exit() to report exception type (e.g., >> "IRQ", "TRAP") and exception class (ESR_ELx's bit[31:26]) together. > > (They both caused an exit!) > > >> But hardware only saves the exit class to ESR_ELx on synchronous > > EC is the 'Exception Class'. Exit is KVM/Linux's terminology. Yes, a stupid mistake ;-)
>> exceptions, not on asynchronous exceptions. When the guest exits >> due to external interrupts, we will get tracing output like: >> >> "kvm_exit: IRQ: HSR_EC: 0x0000 (UNKNOWN), PC: 0xffff87259e30" >> >> Obviously, "HSR_EC" here is meaningless. > > I assume we do it this way so there is only one guest-exit tracepoint that catches all exits. > I don't think its a problem if user-space has to know the EC isn't set for asynchronous > exceptions, this is a property of the architecture and anything using these trace-points > is already arch specific. Actually, *no* problem in current implementation, and I'm OK to still keep the EC in trace_kvm_exit(). What I really want to do is adding the EC in trace_trap_enter (the new tracepoint), will explain it later.
>> This patch splits "exit" and "trap" events by adding two tracepoints >> explicitly in handle_trap_exceptions(). Let trace_kvm_exit() report VM >> exit events, and trace_kvm_trap_exit() report VM trap events. >> >> These tracepoints are adjusted also in preparation for supporting >> 'perf kvm stat' on arm64. > > Because the existing tracepoints are ABI, I don't think we can change them. > > We can add new ones if there is something that a user reasonably needs to trace, and can't > be done any other way. > > What can't 'perf kvm stat' do with the existing trace points? (A good question! I should have made it clear in the commit message, forgive me.)
First, how does 'perf kvm stat' interact with tracepoints?
We have three handlers for a specific event (e.g., "VM-EXIT") -- "is_begin_event", "is_end_event", "decode_key". The first two handlers make use of two existing tracepoints ("kvm:kvm_exit" & "kvm:kvm_entry") to check when the VM-EXIT events started/ended, thus the time difference stats, event start/end time etc. can be calculated. "is_begin_event" handler gets a *key* from the "ret" field (exit_code) of "kvm:kvm_exit" payload, and "decode_key" handler makes use of the *key* to find out the reason for the VM-EXIT event. Of course we should maintain the mapping between exit_code and exit_reason in userspace. These are all what *patch #4* had done, #4 is a simple patch to review! Oh, we can also set "vcpu_id_str" to achieve per vcpu event record, but currently, we only have the "vcpu_pc" field in "kvm:kvm_entry", without something like "vcpu_id".
perf people must have a much deeper understanding of this.
OK, next comes the more important question - what should/can we do to the tracepoints in preparation of 'perf kvm stat' on arm64?
From the article you've provided, it's clear that we can't remove the EC from trace_kvm_exit(). But can we add something like "vcpu_id" into (at least) trace_kvm_entry(), just like what this patch has done? If not, which means we have to keep the existing tracepoints totally unchanged, then 'perf kvm stat' will have no way to record/report per vcpu VM-EXIT events (other arch like X86, powerpc, s390 etc. have this capability, if I understand it correctly).
As for TRAP events, should we consider adding two new tracepoints -- "kvm_trap_enter" and "kvm_trap_exit", to keep tracking of the trap handling process? We should also record the EC in "kvm_trap_enter", which will be used as *key* in TRAP event's "is_begin_event" handler. Patch #5 tells us the whole story, it's simple too.
What do you suggest?
>> diff --git a/arch/arm64/kvm/handle_exit.c b/arch/arm64/kvm/handle_exit.c >> index 516aead..af3c732 100644 >> --- a/arch/arm64/kvm/handle_exit.c >> +++ b/arch/arm64/kvm/handle_exit.c >> @@ -264,7 +264,10 @@ static int handle_trap_exceptions(struct kvm_vcpu *vcpu, struct kvm_run *run) >> exit_handle_fn exit_handler; >> >> exit_handler = kvm_get_exit_handler(vcpu); >> + trace_kvm_trap_enter(vcpu->vcpu_id, >> + kvm_vcpu_trap_get_class(vcpu)); >> handled = exit_handler(vcpu, run); >> + trace_kvm_trap_exit(vcpu->vcpu_id); >> } > > Why are there two? Are you using this to benchmark the exit_handler()? Almostly yes. Let perf know when the TRAP handling event start/end, and ...
> As we can't remove the EC from the exit event, I don't think this tells us anything new. As explained above, this EC is for 'perf kvm stat'.
>> diff --git a/virt/kvm/arm/arm.c b/virt/kvm/arm/arm.c >> index 90cedeb..9f63fd9 100644 >> --- a/virt/kvm/arm/arm.c >> +++ b/virt/kvm/arm/arm.c >> @@ -758,7 +758,7 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *run) >> /************************************************************** >> * Enter the guest >> */ >> - trace_kvm_entry(*vcpu_pc(vcpu)); >> + trace_kvm_entry(vcpu->vcpu_id, *vcpu_pc(vcpu)); > > Why do you need the PC? It was exported on exit. > (its mostly junk for user-space anyway, you can't infer anything from it) (I mainly wanted to add the "vcpu->vcpu_id" here.) It seems that we can't just remove the PC, which will cause ABI change?
Thanks for your reviewing!
zenghui
.
| |