Re: 'perf probe' and symbols from .text.<something>

[Masami Hiramatsu]

On Tue, 23 Feb 2021 13:45:46 -0600
Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:

> On Tue, Feb 23, 2021 at 04:36:19PM +0900, Masami Hiramatsu wrote:
> > On Tue, 23 Feb 2021 10:23:31 +0900
> > Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote:
> > 
> > > On Mon, 22 Feb 2021 11:51:50 -0600
> > > Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> > > 
> > > > On Tue, Feb 23, 2021 at 12:05:08AM +0900, Masami Hiramatsu wrote:
> > > > > > Of course, one could place probes using absolute addresses of the 
> > > > > > functions but that would be less convenient.
> > > > > > 
> > > > > > This also affects many livepatch modules where the kernel code can be 
> > > > > > compiled with -ffunction-sections and each function may end up in a 
> > > > > > separate section .text.<function_name>. 'perf probe' cannot be used 
> > > > > > there, except with the absolute addresses.
> > > > > > 
> > > > > > Moreover, if FGKASLR patches are merged 
> > > > > > (https://lwn.net/Articles/832434/) and the kernel is built with FGKASLR 
> > > > > > enabled, -ffunction-sections will be used too. 'perf probe' will be 
> > > > > > unable to see the kernel functions then.
> > > > > 
> > > > > Hmm, if the FGKASLAR really randomizes the symbol address, perf-probe
> > > > > should give up "_text-relative" probe for that kernel, and must fallback
> > > > > to the "symbol-based" probe. (Are there any way to check the FGKASLR is on?)
> > > > > The problem of "symbol-based" probe is that local (static) symbols
> > > > > may share a same name sometimes. In that case, it can not find correct
> > > > > symbol. (Maybe I can find a candidate from its size.)
> > > > > Anyway, sometimes the security and usability are trade-off.
> > > > 
> > > > We had a similar issue with FGKASLR and live patching.  The proposed
> > > > solution is a new linker flag which eliminates duplicates: -z
> > > > unique-symbol.
> > > > 
> > > > https://sourceware.org/bugzilla/show_bug.cgi?id=26391
> > > 
> > > Interesting, but it might not be enough for perf-probe.
> > > Since the perf-probe has to handle both dwarf and elf, both must be
> > > changed. I think the problem is that the dwarf is generated while
> > > compiling, but this -z seems converting elf symbols in linkage.
> > > As far as I can see, this appends ".COUNT" suffix to the non-unique
> > > symbols in the linkage phase. Is that also applied to dwarf too?
> > 
> > Ah, OK. If there is an offline elf binary with symbol map, I can convert
> > DWARF symbol -> address -> offline elf symbol (unique name)-> kallsyms.
> > Currently, it directly converts address by kallsyms, so I will change it
> > to find elf-symbol and solve address by kallsyms in post processing.
> 
> DWARF sections have references to the ELF symbols, which are renamed by
> the linker.  So DWARF should automatically show the new symbol name.

OK, I'll check what elfutils provides about that information.
> 
> And kallsyms is generated after the kernel is linked.  So I'm not sure I
> understand the problem.

Actually, perf-probe currently uses subprogram DIE(Dwarf node) name for
the symbol name and post-process tries to find correct symbol name
from kallsyms by the address.
So I have to change it to find the ELF symbol name from DIE itself.

Thank you,

-- 
Masami Hiramatsu <mhiramat@xxxxxxxxxx>