Re: [patch 0/4] Port KVM-trace to tracepoints

[fche@redhat.com (Frank Ch. Eigler)]

Avi Kivity <avi@qumranet.com> writes:

> [...]
> kvm tracepoints are heavily tied into the implementation; and making
> them harder to write means we will have less information.  In fact, I
> am contemplating moving in another direction (when looking at the
> pgprintk()s scattered around arch/x86/kvm/mmu.c:
>
>   kvm_trace("pfentry", "page_fault entry addr %lx error code %x\n",
> cr2, error_code);
>
> Unlike printk()s, no actual formatting would occur during runtime.

Have you considered using trace_mark() directly - eliminating the
KVM_TRACEN() middlemen?

> Instead, at initialization time all the strings would be parsed into
> a data structure that describes the data types, and the runtime
> would simply consult this structure and copy the arguments into
> trace records.  User space would also be able to pull this structure
> and so recreate the formatted string.

If one really wanted to, one could build such a mechanism on top of
marker-based callbacks.

> The advantages I see to this are:
>
>  - easy to add traces; the most important advantage
>  - when the code changes, obsolete traces are completely removed
>  - good performance

Ditto.

>  - no need to have a formats file in userspace (which is tied to the
> kernel version)

OTOH, you'd have the kernel collecting compact binary records
containing just the parameters, which are at least as tied to kernel
version.

>  - can also send printk()s along, for synchronization with other kvm
> and kernel events

Ditto.  It is elementary to attach a printk-generating marker callback.


- FChE