Re: [Qemu-devel] [PATCH] exec: Rename and fix trace events for tracing I/O port access.

[Paolo Bonzini <pbonzini@redhat.com>]

On 29/03/2016 17:02, Richard W.M. Jones wrote:
> Back in the day you used to be able to set DEBUG_IOPORT in ioport.c
> and get qemu to dump what (x86) I/O ports were being accessed by the
> guest.  This was rather useful for finding out what closed source
> device drivers were up to.
> 
> Now you're supposed to use cpu_in/cpu_out tracepoints instead.
> However for the majority of guests these tracepoints will never be
> called.

For what it's worth, the breakage dated to way before DEBUG_IOPORT was
converted to tracepoints.  (This is why putting tracing code behind #if
0 is bad).

It happened when I/O ports stopped being special snowflakes, and started
being just like any other MMIO target (except they live in another
address space).

> It turns out this is still not particularly useful for debugging
> because (a) it creates massive amounts of log messages and (b) there's
> no way to select a range of addresses or a device of interest.  For
> example, if you have a serial port, everything else gets swamped by
> I/O access to the serial port.

Am I wrong that this is not any better/worse than "-d ioport" used to be?

> Maybe using a different tracing backend (eg. stap) would help?

You are actually logging every memory write (except stuff that it's
DMA'd from/to block devices directly by passing a guest memory pointer
to preadv/pwritev).  That's way more than just I/O ports.

However I understand that it's useful to treat them as special
snowflakes for the purpose of tracing.  To fix the problem with your
patch, I suggest to:

- consolidate the six tracepoints in 2 (cpu_in and cpu_out) that take
the size as an extra argument

- keep the existing calls to the tracepoints, and add more calls in
kvm_handle_io and in the helpers in target-i386/misc_helper.c.

Paolo