Re: [PATCH 04/15] perf: Add ability to dump user regs

[Jiri Olsa <jolsa@redhat.com>]

On Wed, Mar 28, 2012 at 10:01:15AM -0400, Frank Ch. Eigler wrote:
> Hi, Jiri -
> 
> On Wed, Mar 28, 2012 at 02:35:47PM +0200, Jiri Olsa wrote:
> > [...]
> > The register value here are those of the user space context as
> > it was before the user entered the kernel for whatever reason
> > (syscall, irq, exception, or a PMI happening in userspace).
> > [...]
> 
> As I understand the situation, there is a complication here that you
> haven't accounted for.  Upon a normal syscall entry to the kernel, not
> all user registers are saved explicitly for such easy retrieval.  The
> others may be spilled to the stack by gcc during the various sys_*
> functions or elsewhere.  It turns out that some of these saved
> registers are sometimes necessary to accomplish a user-space unwind.

Are you reffering to x86_64 where only portion of registers
is stored by SAVE_ARGS macro? Seems like 32 bits stores the
whole pt_regs.

Generally you could need all the registers to start the unwind,
but I was assuming that for most cases the stack pointer and
instruction pointer should be enough.. but I might be wrong here.

> 
> To recover these registers at run time, we found that the kernel stack
> itself has to be partially unwound - and not via frame pointers, but
> the full dwarf unwind/cfi machinery.  This RFC code does not appear
> aware of the difference between the explicitly saved and the
> incidentally-spilled registers, and thus may accidentally pass garbage
> data to perf userspace.  Correcting this could require a kernel-space
> libunwind.

AFAIK not going to happen any time soon ;)

thanks,
jirka