Re: [PATCH 4/4] perf tools: Add support for breakpoint events in perf tools

[Frederic Weisbecker <fweisbec@gmail.com>]

On Mon, Nov 23, 2009 at 11:06:01PM +0530, K.Prasad wrote:
> On Mon, Nov 23, 2009 at 03:42:35PM +0100, Frederic Weisbecker wrote:
> > Add the breakpoint events support with this new sysnopsis:
> > 
> > mem:addr[:access]
> > 
> > Where addr is a raw addr value in the kernel and access can be
> > either [r][w][x]
> > 
> > Example to profile tasklist_lock:
> > 
> > 	$ grep tasklist_lock /proc/kallsyms
> > 	ffffffff8189c000 D tasklist_lock
> > 
> > 	$ perf record -e mem:0xffffffff8189c000:rw -a -f -c 1
> > 	$ perf report
> 
> The problem in obtaining just the breakpoint address is that there can
> be a variety of breakpoint lengths that can be associated with them -
> assigning the smallest possible length (1-Byte) may cause loss of
> exceptions and a higher length would lead to stray exceptions (such
> dilemmas led to the support of symbol-only breakpoint in ksym_tracer and
> perf-tools in my patchset...the default 1-Byte breakpoint length being a
> temporary fix).


Right.

 
> With kernel symbols as input it would be possible to derive the bkpt
> length based on the symbol-size, say using
> kallsyms_lookup_size_offset() (although the corresponding length may not
> be available on the host processor such requests can be failed or
> over-ridden by the user using a smaller length), but for addresses I think
> it is vital to know what breakpoint length is desired by the user.



Yeah. I guess we first need a way to manually add this length, may
be:

	mem:addr/len:access

And as you said, finding it automatically for symbols. But still, passing
symbols to perf attr leads to confusion and complexity if we want to profile
in userspace.

I think we should find this symbol length from userspace.
I'm not sure how yet, probably using Dwarf. Arnaldo, do you have an
idea about that?

> This comes at the cost of exposing the user to variations in
> breakpoint implementation across architectures and demand processor-specific
> knowledge, but specifying a kernel-space address would anyway require the
> user to penetrate beyond the normal abstraction provided by the
> interface/tool...so I presume it must be acceptable.



Yeah. We'll probably need to write a quick sum-up about
such variations to facilitate perf uses.


> On a related note, the supported breakpoint length for PPC64 is a fixed
> 8-Byte length (which means all extraneous exceptions must be filtered by
> the breakpoint architecture) and Book-E Power processors matching
> addresses against a bitmask; for S390 it can be practically anything
> (bound by a set of start and end addresses)...and you see what a
> quandary it leads to!
> 
> Thanks,
> K.Prasad


Yep :)

Thanks.