[lkml]   [2010]   [Jan]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [Patch 1/1] Introduce register_user_hbp_by_pid() and unregister_user_hbp_by_pid()
On Fri, Jan 01, 2010 at 12:18:04AM +0530, K.Prasad wrote:
> On Wed, Dec 30, 2009 at 11:28:39PM +0100, Frederic Weisbecker wrote:
> > On Tue, Dec 22, 2009 at 12:16:31AM +0530, K.Prasad wrote:
> > > On Fri, Dec 18, 2009 at 09:47:48PM +0100, Frederic Weisbecker wrote:
> >
> > > > And this function needs rcu too.
> > > >
> > > > I don't see any in-kernel user for this new feature.
> > > > That would be required to integrate it.
> > > >
> > >
> > > The proposed interfaces, as obvious, are mere wrappers over existing
> > > (un)register_user_* interfaces, and don't do anything vastly different
> > > in order to demonstrate them separately.
> > >
> > > I can get a sample kernel module ready - that consumes pid and user-space
> > > address to track write accesses, if you prefer it.
> >
> >
> > Ok. The code looks good and useful.
> >
> > But the usual philosophy in the kernel is to not add code
> > that is left unused upstream. And samples don't substitute a user.
> > I'm not sure this is a good idea to merge this.
> >
> Back to the old trick!...How about an ftrace plugin that accepts pid,
> user-space address and memory access type and traces all the IP
> addresses that caused access?
> echo <pid>:<user_addr><access_type> > usym_trace_filter
> echo 567:0x1234567:rw- > usym_trace_filter
> Breakpoint IP
> ------------ ---------
> 567:0x1234567 0x0abcdef
> I'm unsure if it sounds interesting at all, but I suspect it wouldn't be
> as easy as above to gather the shown information through any existing
> tools.

That's a good idea to trace userspace breakpoints but:

I think the perf interface to use breakpoints is much more powerful
than the breakpoint ftrace plugin, which is somehow deprecated now.

The fact is that ftrace plugins in general (I mean the plugins based
on struct tracer, not the trace events) are mostly deprecated in favor
of trace events.

There are still some areas where the plugins are necessary though, such as
function tracing, and some other tracers. But these are exceptions.
And breakpoints interface was one of them, but now we have
a much more powerful existing interface for it in perf tools.

When it comes to think about improving or adding a new ftrace plugin,
if we know an existing interface that is proven better, let's rather
improve the latter (that's why I'll try to convert the function graph
tracer into a trace event...not an easy task).

The breakpoint ftrace plugin can only trace the whole kernel, has no
per-cpu or per task (+inheritance) granularity, backtraces,
or whatever the perf tools can already offer.

To sum-up, sure we could improve it but:

- we already have better, as a unified and easier to improve interface.
Extending this ftrace plugin would make it even harder to maintain.

- ftrace plugins are deprecated, except for particular cases

So, concerning breakpoints, I really suggest to focus on the perf tools
and deprecate this breakpoint ftrace plugin.

Also, I'm pretty sure that this would already work:

./perf record -e mem:@addr_in_ls:rw ls /usr
./perf record -e mem:@addr_in_ls:rw --pid $(pid_of_a_running_ls)

I've never tested it, but if that doesn't work, that's probably because
of a guardian inside the kernel that only accepts userspace breakpoints
from ptraced processes. I should check that. But if it doesn't work yet,
that would require very few changes for it to work.

 \ /
  Last update: 2010-01-10 05:05    [W:0.047 / U:1.528 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site