Re: [PATCH v2 0/9] tracing: Improbe BTF support on probe events

[Alan Maguire <alan.maguire@oracle.com>]

On 17/07/2023 16:23, Masami Hiramatsu (Google) wrote:
> Hi,
> 
> Here is the 2nd version of series to improve the BTF support on probe events.
> The previous series is here:
> 
> https://lore.kernel.org/linux-trace-kernel/168699521817.528797.13179901018528120324.stgit@mhiramat.roam.corp.google.com/
> 
> In this version, I added a NULL check fix patch [1/9] (which will go to
> fixes branch) and move BTF related API to kernel/bpf/btf.c [2/9] and add
> a new BTF API [3/9] so that anyone can reuse it.
> Also I decided to use '$retval' directly instead of 'retval' pseudo BTF
> variable for field access at [5/9] because I introduced an idea to choose
> function 'exit' event automatically if '$retval' is used [7/9]. With that
> change, we can not use 'retval' because if a function has 'retval'
> argument, we can not decide 'f func retval' is function exit or entry.

this is fantastic work! (FWIW I ran into the retval argument issue with
ksnoop as well; I got around it by using "return" to signify the return
value since as a reserved word it won't clash with a variable name.
However in the trace subsystem context retval is used extensively so
makes sense to stick with that).

One thing we should probably figure out is a common approach to handling
ambiguous static functions that will work across ftrace and BPF.  A few
edge cases that are worth figuring out:

1. a static function with the same name exists in multiple modules,
either with different or identical function signatures
2. a static function has .isra.0 and other gcc suffixes applied to
static functions during optimization

As Alexei mentioned, we're still working on 1, so it would be good
to figure out a naming scheme that works well in both ftrace and BPF
contexts. There are a few hundred of these ambiguous functions. My
reading of the fprobe docs seems to suggest that there is no mechanism
to specify a specific module for a given symbol (as in ftrace filters),
is that right?

Jiri led a session on this topic at LSF/MM/BPF ; perhaps we should
carve out some time at Plumbers to discuss this?

With respect to 2, pahole v1.25 will generate representations for these
"."-suffixed functions in BTF via --btf_gen_optimized [1]. (BTF
representation is skipped if the optimizations impact on the registers
used for function arguments; if these don't match calling conventions
due to optimized-out params, we don't represent the function in BTF,
as the tracing expectations are violated).

However the BTF function name - in line with DWARF representation -
will not have the .isra suffix. So the thing to bear in mind is if
you use the function name with suffix as the fprobe function name,
a BTF lookup of that exact ("foo.isra.0") name will not find anything,
while a lookup of "foo" will succeed. I'll add some specifics in your
patch doing the lookups, but just wanted to highlight the issue at
the top-level.

Thanks!

Alan

[1]
https://lore.kernel.org/bpf/1675790102-23037-1-git-send-email-alan.maguire@oracle.com/

> Selftest test case [8/9] and document [9/9] are also updated according to
> those changes.
> 
> This series can be applied on top of "v6.5-rc2" kernel.
> 
> You can also get this series from:
> 
> git://git.kernel.org/pub/scm/linux/kernel/git/mhiramat/linux.git topic/fprobe-event-ext
> 
> 
> Thank you,
> 
> ---
> 
> Masami Hiramatsu (Google) (9):
>       tracing/probes: Fix to add NULL check for BTF APIs
>       bpf/btf: tracing: Move finding func-proto API and getting func-param API to BTF
>       bpf/btf: Add a function to search a member of a struct/union
>       tracing/probes: Support BTF based data structure field access
>       tracing/probes: Support BTF field access from $retval
>       tracing/probes: Add string type check with BTF
>       tracing/fprobe-event: Assume fprobe is a return event by $retval
>       selftests/ftrace: Add BTF fields access testcases
>       Documentation: tracing: Update fprobe event example with BTF field
> 
> 
>  Documentation/trace/fprobetrace.rst                |   50 ++
>  include/linux/btf.h                                |    7 
>  kernel/bpf/btf.c                                   |   83 ++++
>  kernel/trace/trace_fprobe.c                        |   58 ++-
>  kernel/trace/trace_probe.c                         |  402 +++++++++++++++-----
>  kernel/trace/trace_probe.h                         |   12 +
>  .../ftrace/test.d/dynevent/add_remove_btfarg.tc    |   11 +
>  .../ftrace/test.d/dynevent/fprobe_syntax_errors.tc |    6 
>  8 files changed, 503 insertions(+), 126 deletions(-)
> 
> --
> Masami Hiramatsu (Google) <mhiramat@kernel.org>
>