Date:	Wed, 5 Nov 2014 10:04:29 -0300
From:	Arnaldo Carvalho de Melo <acme@...nel.org>
To:	Masami Hiramatsu <masami.hiramatsu.pt@...achi.com>
Cc:	srikar@...ux.vnet.ibm.com, Peter Zijlstra <peterz@...radead.org>,
	Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
	Brendan Gregg <brendan.gregg@...il.com>,
	yrl.pp-manager.tt@...achi.com, namhyung@...nel.org,
	Hemant Kumar <hemant@...ux.vnet.ibm.com>,
	Ingo Molnar <mingo@...nel.org>,
	Adrian Hunter <adrian.hunter@...el.com>
Subject: Re: [PATCH perf/core 0/6] perf-probe: Bugfix and add new options for
 cache

Em Wed, Nov 05, 2014 at 01:22:46AM +0900, Masami Hiramatsu escreveu:
> (2014/11/04 23:38), Arnaldo Carvalho de Melo wrote:
> > Em Tue, Nov 04, 2014 at 01:36:31PM +0900, Masami Hiramatsu escreveu:
> >> (2014/11/04 1:19), Arnaldo Carvalho de Melo wrote:
> >>> Em Mon, Nov 03, 2014 at 09:11:18PM +0900, Masami Hiramatsu escreveu:
> >>>> (2014/10/31 21:13), Arnaldo Carvalho de Melo wrote:
> >>>>> Em Fri, Oct 31, 2014 at 02:51:29PM -0400, Masami Hiramatsu escreveu:
> >>>> Actually, kprobe event itself can reject command if the given address
> >>>> is not in the kernel text nor instruction boundary (perhaps, uprobes
> >>>> may have a problem...), so for the kernel level, it is safe.
> > 
> >>> No, it is not necessarily safe.
> > 
> >>> What if you specify function foo() that has address 0x1234 for kernel
> >>> v3.16 and then run the cached probe on kernel v3.18 and on that kernel
> >>> the function foo() maps to address 0x2345 and function bar() instead
> >>> maps to address 0x1234? Oops.

> >> In that case user just trace bar() instead of foo(). Of course it's
> >> not correct, but shouldn't break the kernel (if the kernel is broken,
> >> it is a bug of kprobes).

> > I.e. no crashes, just misleading information :-\
 
> Right.

The kernel doesn't crashes, just the user, after scratching his head
trying to make sense of wrong information :-P

Digressing: is there some kprobe_tracer_fuzzer out there?
 
> >>> The build-id was designed to uniquely identify a DSO, we need to use it.

> >>> I think that at some point not using it should be left to a, in
> >>> systemtap parlance, "guru" mode, with tooling warning profusely when
> >>> build ids are not available and requiring even more forcing when it
> >>> doesn't matches.

> >> But it is not necessarily everyone uses perf probe to set up the probe
> >> events(because it is a part of ftrace), as we can see in the Brendan's
> >> scripts.

> > Right, If I implied that some particular tool should be used, sorry
> > about that, what I wanted to get accross was that the information that
> > allows users or tools to make sure there is no mismatch between the
> > cached probes and the target kernel is collected at cached probe
> > creating time and available at target use time.
 
> Yes, and if user setting probes via perf, the perf must ensure that it
> picks up the correct cache by using build-id. If someone wants to use
> other tools, he/she must ensure it. We just give a information how to
> check that :)
 
> >> I think, at least what we need is clarifying how can they ensure
> >> build-id before setting the probe events. I'd like to give them options
> >> with knowledge instead of forcing by tools.

> > Right, so we need to have the build-id as part of the cache format,
> > perhaps as the first line, starting with a comment (#), that way the
> > user can use whatever way he has at its disposal to check that the
> > running kernel build-id is the same as the one on that comment. Using
> > that script you provided, that uses just things that are on the machine
> > (od, /sys/kernel/notes).
 
> Ah, that's a good idea :)
> So, with such build-id comment line, would you think we can have an
> --output option? Or we'd better moving onto the .debug/ cache file?

Well, these are two separate things: The ./debug cache file is for
repeated use of the same probe on the same machine, say, across reboots
or in tools that script using perf probe to add some probe, then remove
the probes at exit, using it multiple times would provide a after
caching, as no ELF DWARF parsing would be involved after we cached it.

As well for when using 'perf archive' to reuse the probe definitions
using 'perf probe' on the target machine, where no -debuginfo packages
(or binaries with DWARF) would be available.

I.e. both having the comment on the cache file with the build id and
having it stored in ~/.debug/ have values for different workloads or
different ways of doing the same thing.
 
> What I'm thinking about this feature is to make a compact and reduced
> function-entry level probe cache while building the kernel (as a part
> of kbuild), so that we can deploy the stripped kernel and the cache
> to remote machines.

Which is a cool feature as well! :-)
 
> [snip]
> >> OK, I agree using .debug/.buildid/ to store caches.
> >> Here is what I'm thinking,
> >  
> >>   # this makes caches for given pattern instead of adding probes.
> >>   perf probe --cache '* $params'
> > 
> >>   # the cache is stored in .debug/.buildid/<buildid>.probe
> >>   # the cache entry can be queried by buildid and eventname
> > 
> > To follow the existing standard this would instead go to:
> > 
> >>   # the cache is stored in .debug/probes/path/to/dso/name/buildid
> >>   # And can be found via its buildid link .debug/.buildid/bu/ildid -> ../../probes/path/to/dso/name/buildid
> 
> Ah, I see. so you meant adding a top-level .debug/probes/ dir.
> But in that case, shouldn't we change .debug/.buildid/bu/ildid to
> .debug/probes/.buildid/bu/ildid ?

Humm, understood, perhaps we should have a subdir for probes, i.e.:

# And can be found via its buildid link .debug/.buildid/probes/bu/ildid -> ../../probes/path/to/dso/name/probes/buildid

So as not to clash in both cases with the ELF file.

Which leads to a problem when we have both vmlinux and kallsyms, which,
right now, is not possible, i.e. if we add both a vmlinux and a kallsyms
file for a kernel, the one that comes last will have the

.debug/.buildid/bu/ildid

link, have to fix that so that we first look for the ELF file in the
cache, and if it fails, fallback to the kallsyms file, if available,
i.e. give preference to the richer option, and in some cases, like for
annotation, the only one that matters.

> >>   perf probe --query ${remote_buildid}:do_fork
> >>   p:probe/do_fork _text+298722 clone_flags=%di:u64 stack_start=%si:u64 stack_size=%dx:u64 parent_tidptr=%cx:u64 child_tidptr=%r8:u64
> > 
> >>   # or perf can set it up directly to local
> >>   perf probe --query-add do_fork
> > 
> > You missed the build id above, no? I.e. it would be:
> > 
> >>   # or perf can set it up directly to local
> >>   perf probe --query-add ${remote_buildid}:do_fork
 
> No, since this command set the event to local machine, perf-probe
> should check the local build-id and query the appropriate event
> from the cache.
> # BTW, maybe we'd better use perf probe --add '$do_fork' (calls
> # "cache of do_fork")  instead of long --query-add. :)

Ok, understood.
 
> > 
> > Right? and that would be the equivalent to:
> > 
> >  perf probe --query ${remote_buildid}:do_fork > /sys/kernel/debug/tracing/kprobe_events
> > 
> > No?
> > 
> > And do you intend to generate that script above that checks the build
> > id, tees into /sys/kernel/debug/tracing/kprobe_events, etc as part of
> > this process?
> 
> No, not yet. But that sounds nice :) Thanks!
> 
> > 
> > I.e. the process would be, as you describe above plus telling the user
> > that he/she then should unconpress the script and run it.
> > 
> > Scripts like Brendan's would do whatever they want with it of course,
> > but the default end result of 'perf probe' would be usable straight away
> > and doing the build id checking without requiring any extra tooling to
> > be available at the target machine.
> 
> Agreed, with .debug/ archive, perf-probe should pick and set event
> correctly without any other tools.
> 
> Thank you!
> 
> > 
> >>   Added new event:
> >>     probe:do_fork        (on do_fork with clone_flags satck_start stack_size parent_tidptr child_tidptr)
> > 
> >>   You can now use it in all perf tools, such as:
> > 
> >>         perf record -e probe:do_fork -aR sleep 1
> > 
> >> What would you think about this? :)
> > 
> > Cool feature! :-)
> > 
> >>>>> Then, later, one would use 'perf archive' passing some keys (or a
> >>>>> perf.data file, like done nowadays to pick the files in ~/.debug for
> >>>>> dsos that had hits on the specified perf.data file) to get the cached
> >>>>> values to use on some other machine, to avoid having to use the
> >>>>> debuginfo files there.
> > 
> >>>> Yeah, querying it from the BUILDID database by using a pair of remote
> >>>> build-id and the binary path is a good feature.
> > 
> >>>>> I.e. in summary I think that the format is ok, but we need to have this
> >>>>> inside the ~/.debug hierarchy so that we can make sure that we use the
> >>>>> right probe definition, one that matches the DSOs being used (the kernel
> >>>>> or some other userspace binary).
> >>>
> >>>> OK, perhaps, that is also good to SDT series at last.
> >>>
> >>> Sure thing!
> >>>
> >>
> >> Thank you!
> > 
> > You're welcome!
> > 
> > - Arnaldo
> > 
> 
> 
> -- 
> Masami HIRAMATSU
> Software Platform Research Dept. Linux Technology Research Center
> Hitachi, Ltd., Yokohama Research Laboratory
> E-mail: masami.hiramatsu.pt@...achi.com
> 
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists