Re: [PATCH] perf: Add branch stack support to perf script

[Arun Sharma <asharma@fb.com>]

On 3/15/12 1:08 PM, Arun Sharma wrote:

> The samples seem to have the sense of call/return
> inverted. If the original callgraph was a ->  b ->  c
> I get samples like:
>
> from	to
> c	b
> b	a
>
> To restore the normal sense, I'm printing them as:
>
> to =>  from

I debugged this some more and something seems to be wrong with the the 
way the kernel maps PERF_SAMPLE_BRANCH_ANY_CALL to MSR_LBR_SELECT.

perf record -aj any_call,u -F 1 --   sleep 300 &

msr[0x1c8] = 0x1d

perf record -aj any_ret,u -F 1 --   sleep 300 &

msr[0x1c8] = 0x1ad

Stephane: does this give a clue about what may be wrong? It doesn't 
match the kernel code:

static const int nhm_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX] = {
...
[PERF_SAMPLE_BRANCH_ANY_CALL] =
          LBR_REL_CALL | LBR_IND_CALL | LBR_REL_JMP | LBR_IND_JMP | LBR_FAR,
};

It'd be nice to have *_lbr_sel_map[] in the same order as
enum perf_branch_sample_type. Right now, the call and return entries are 
reversed. Which shouldn't matter in theory, since the initializer has an 
explicit index.

I also got a kernel hang running the two perf record lines above 
simultaneously (without the -F 1).

  -Arun