On 23.01.2019 00:11, Paolo Bonzini wrote: > On 22/01/19 21:48, Greg Kroah-Hartman wrote: >>> This also brings the question: shall we move these counters out of debugfs into something else? >> If you have code that relies on debugfs, yes, you need to move that out >> of debugfs because more and more systems are trying to disable it due to >> the obvious problems with it (i.e. leaking tons of debugging >> information). >> >> debugfs is for DEBUG information, not for "statistics about how my VM is >> working". That sounds like something you need to rely on, so debugfs is >> not the place for it. > > Yes, we know that and tracepoints are already one replacement. However, > they are slower that just a lock-free "vcpu->stats.foo_happened++". Yes, the tracepoints are not a proper replacement for the counters (especially the capability to get numbers after-the-fact. So I would really like to keep both. > Another idea that Steven Rostedt and I discussed a while ago is some > kind of "statfs" which would already provide some code, similar to the > one that KVM uses to accumulate statistics from multiple VMs or multiple > VCPUs into a single counter. I think that would make a lot of sense to have a common filesystem to avoid code duplication bugs.
