[RFC PATCH 0/7] Add support for monitoring guest TLB operations

[Punit Agrawal]

Paolo Bonzini <pbonzini at redhat.com> writes:

> On 16/08/2016 12:45, Punit Agrawal wrote:
>> Hi,
>> 
>> ARMv8 supports trapping guest TLB maintenance operations to the
>> hypervisor. This trapping mechanism can be used to monitor the use of
>> guest TLB instructions.
>> 
>> As taking a trap for every TLB operation can have significant
>> overhead, trapping should only be enabled -
>> 
>> * on user request
>> * for the VM of interest
>> 
>> This patchset adds support to listen to perf trace event state change
>> notifications. The notifications and associated context are then used
>> to enable trapping of guest TLB operations when requested by the
>> user. The trap handling generates trace events (kvm_tlb_invalidate)
>> which can already be counted using existing perf trace
>> functionality.
>> 
>> Trapping of guest TLB operations is disabled when not being monitored
>> (reducing profiling overhead).
>> 
>> I would appreciate feedback on the approach to tie the control of TLB
>> monitoring with perf trace events (Patch 1) especially if there are
>> any suggestions on avoiding (or reducing) the overhead of "perf trace"
>> notifications.
>> 
>> I looked at using regfunc/unregfunc tracepoint hooks but they don't
>> include the event context. But the bigger problem was that the
>> callbacks are only called on the first instance of simultaneously
>> executing perf stat invocations.
>> 
>> The patchset is based on v4.8-rc2 and adds support for monitoring
>> guest TLB operations on 64bit hosts. If the approach taken in the
>> patches is acceptable, I'll add 32bit host support as well.
>> 
>> With this patchset, 'perf' tool when attached to a VM process can be
>> used to monitor the TLB operations. E.g., to monitor a VM with process
>> id 4166 -
>> 
>> # perf stat -e "kvm:kvm_tlb_invalidate" -p 4166
>> 
>> Perform some operations in VM (running 'make -j 7' on the kernel
>> sources in this instance). Breaking out of perf shows -
>> 
>> Performance counter stats for process id '4166':
>> 
>>          7,471,974      kvm:kvm_tlb_invalidate
>> 
>>      374.235405282 seconds time elapsed
>> 
>> All feedback welcome.
>
> Can you explain what this is used for?  In other words, why would this
> be used instead of just running perf in the guest?

As TLB maintenance operations are synchronised in hardware, they can
impact performance beyond the guest. The operations generate traffic on
the interconnect and depending on the implementation, they can also
affect the remote TLB's translation bandwidth.

These patches are useful on systems where the host and guest are
controlled by different users - the guest could be running arbitrary
software. 

Having the ability to monitor the usage of guest TLB invalidations in
the host can be useful to diagnose performance issues on such systems.

>
> Thanks,
>
> Paolo
> _______________________________________________
> kvmarm mailing list
> kvmarm at lists.cs.columbia.edu
> https://lists.cs.columbia.edu/mailman/listinfo/kvmarm