Re: [Qemu-devel] Tracking the VM making an IO request

[Fam Zheng <famz@redhat.com>]

On Sat, 03/12 16:17, Aarian P. Aleahmad wrote:
> Thanks for helping me. what should I do in case of using KVM?

The said method works exactly the same with or without -enable-kvm.

Fam

> 
> On Wed, Feb 10, 2016 at 4:10 PM, Paolo Bonzini <pbonzini@redhat.com> wrote:
> 
> >
> >
> > On 10/02/2016 11:23, Stefan Hajnoczi wrote:
> > > On Wed, Feb 10, 2016 at 12:35:54PM +0330, Aarian P. Aleahmad
> > > wrote:
> > >> I'm a student, engaged in a project in which QEMU is a candidate
> > >> to be used to make some studies about IO usage etc. I need to
> > >> track the IO requests made to the block devices (e.g. HDD, SSD,
> > >> etc.). I check the source code but I was confused. What I want to
> > >> know is that when an IO request is made, find out that which on
> > >> of the VMs has made that request. I'll thank you if you help me
> > >> on this issue.
> > >
> > > There are trace events that you can use.  See docs/tracing.txt and
> > > trace-events.
> > >
> > > virtio_blk_handle_write and virtio_blk_handle_read can be used if
> > > your guest has virtio-blk.
> > >
> > > The QEMU block layer also has trace events named bdrv_aio_*.
> > >
> > > Or you could use blktrace(8) in the guest or on the host, depending
> > > on how you've set up storage.
> >
> > It's the third time I've gotten the question recently which makes me
> > believe the others were friends with Aarian...
> >
> > Each QEMU process represents a single VM.  Therefore, it is simple to
> > answer the question "which VM is making the request"; the answer is
> > "the one for the QEMU process you are tracing".
> >
> > You probably want to use blktrace if you care about multiple VMs.
> > Alternatively, you can use tracing as mentioned by Stefan.  If you
> > compile QEMU with --enable-trace-backend=simple, the resulting files
> > can be parsed with Python programs (see scripts/simpletrace.py).  The
> > trace files include the pid and a timestamp based on CLOCK_MONOTONIC,
> > so it should be easy to merge the traces together.
> >
> > Paolo
> >