Re: [PATCH V2 1/9] perf tools: Fix symbol annotation for relocated kernel

[Ingo Molnar <mingo@kernel.org>]

* Adrian Hunter <adrian.hunter@intel.com> wrote:

> On 30/01/14 10:59, Ingo Molnar wrote:
> > 
> > * Adrian Hunter <adrian.hunter@intel.com> wrote:
> > 
> >> On 29/01/14 20:57, Arnaldo Carvalho de Melo wrote:
> >>> Em Wed, Jan 29, 2014 at 04:14:36PM +0200, Adrian Hunter escreveu:
> >>>> Kernel maps map memory addresses to file offsets.
> >>>> For symbol annotation, objdump needs the object VMA
> >>>> addresses.  For an unrelocated kernel, that is the
> >>>> same as the memory address.
> >>>>
> >>>> The addresses passed to objdump for symbol annotation
> >>>> did not take into account kernel relocation.  This
> >>>> patch fixes that.
> >>>
> >>> Question: To fix the problem reported by Linus, i.e. the very minimal
> >>> fix, we only need this patch, right?
> >>
> >> Yes but the other fixes are needed too.
> > 
> > So, for the specific case of kernel address layout randomization, how 
> > does this fix Linus's bug with KASLR enabled? How does the code 
> > recover the random, runtime offset of the relocated kernel, which 
> > varies from boot to boot?
> 
> By comparing the address of a symbol ("_text" or "_stext")
> in /proc/kallsyms (or perf.data - see below) with the same
> symbol in vmlinux.
> 
> perf tools call this the ref_reloc_sym and stores it in
> perf.data hidden in the synthesized kernel mmap record.
> e.g.
> 
> 0xd8 [0x50]: event: 1
> .
> . ... raw event: size 80 bytes
> .  0000:  01 00 00 00 01 00 50 00 ff ff ff ff 00 00 00 00  ......P.........
> .  0010:  00 00 00 17 00 00 00 00 ff ff ff a8 ff ff ff ff  ................
> .  0020:  c8 01 00 98 ff ff ff ff 5b 6b 65 72 6e 65 6c 2e  ........[kernel.
> .  0030:  6b 61 6c 6c 73 79 6d 73 5d 5f 73 74 65 78 74 00  kallsyms]_stext.
> .  0040:  00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
> .
> 0 0xd8 [0x50]: PERF_RECORD_MMAP -1/0: [0x17000000(0xffffffffa8ffffff) @ 0xffffffff980001c8]: x [kernel.kallsyms]_stext
> 
> That tells perf tools that _stext was 0xffffffff980001c8.
> Compare to vmlinux:
> 
> $ objdump -t vmlinux | grep _stext
> ffffffff810001c8 g       .text	0000000000000000 _stext
> 
> So the relocation is 0xffffffff980001c8 - 0xffffffff810001c8
> = 0x17000000

Ok, cool, thanks!

I'd suggest the whole fix series if perf/urgent material.

Thanks,

	Ingo