Messages in this thread | | | Date | Mon, 15 Sep 2014 08:23:20 -0600 | From | David Ahern <> | Subject | Re: [PATCH RFC] perf kvm stat live: cache mmap()ed events |
| |
On 9/15/14, 6:57 AM, Alexander Yarygin wrote: > David Ahern <dsahern@gmail.com> writes: > >> On 9/12/14, 10:27 AM, Alexander Yarygin wrote: >>> During mmap() process 'perf kvm stat live' gets a pointer to events and >>> passes them to the session queue. Events are stored in shared memory and >>> eventually they will be overwritten by the kernel. The problem is, that >>> when events come too fast, old events can be overwritten before they >>> have been processed that can lead to perf crash. >>> >>> To prevent that happening, we can copy upcoming events and pass a copy >>> to the session queue. There is a safe place to copy event: before >>> perf_evlist__mmap_consume() is executed. There are 3 places to free it: >>> when event is processed, when it's lost and on exit, if it's turned out >>> unprocessed. >> >> Did you see what I proposed a year ago: >> https://lkml.org/lkml/2013/9/6/388 >> >> The intent is to keep the copy generic and not local a command since >> conceptually other live commands need the same. >> >> David > > Hello, > > Yes, your patch works fine. But as far as I understand, right now only > the 'perf kvm stat live' is infected: other 'live' tools were fixed by > the patch "PERF: The tail position of the event buffer should only be modified > after actually use that event." and since they don't use ordered queue > they don't need a copying. That's why I came up with 'perf kvm stat > live' specific approach. Maybe I missed something...
No, you understand the problem. My point is that the key issue with the code (overwriting events in the mmap'ed buffers) really has nothing to do with perf-kvm. It might be the only command in-tree today but others could come along, so might as well fix the issue in the event processing code.
> > Anyhow, having >30K events is a quite usual situation on s390 and the > 'perf kvm stat live' command hardly works there, so it would be good to > have at least some working solution. Any ideas? :)
Sure. With nested VMs on x86 I see 100,000's of events per second. That use case is what drove me to look at copying events. I just have not had time to come back to that one.
David
| |