Re: [RFC 09/23] x86_64: Store userspace rsp in system_call fastpath

[Frederic Weisbecker <fweisbec@gmail.com>]

On Fri, Jun 29, 2012 at 10:03:27AM +0200, Jiri Olsa wrote:
> On Thu, Jun 28, 2012 at 02:08:49PM +0200, Frederic Weisbecker wrote:
> > On Tue, Jun 19, 2012 at 05:48:00PM +0200, Jiri Olsa wrote:
> > > hi,
> > > I'd need help with this change.. basically it works, but I guess
> > > I could alter the FIXUP_TOP_OF_STACK macro as well, so the rsp is
> > > not initialized twice in slowpath.
> > > 
> > > But it seems quite complex.. so not really sure at the moment ;)
> > > 
> > > ideas?
> > > 
> > > thanks,
> > > jirka
> > > 
> > > ---
> > > Storing the userspace rsp into the pt_regs struct for the
> > > system_call fastpath (syscall instruction handler).
> > > 
> > > Following part of the pt_regs is allocated on stack:
> > > (via KERNEL_STACK_OFFSET)
> > > 
> > >     unsigned long ip;
> > >     unsigned long cs;
> > >     unsigned long flags;
> > >     unsigned long sp;
> > >     unsigned long ss;
> > > 
> > > but only ip is actually saved for fastpath.
> > > 
> > > For perf post unwind we need at least ip and sp to be able to
> > > start the unwind, so storing the old_rsp value to the sp.
> > > 
> > > Signed-off-by: Jiri Olsa <jolsa@redhat.com>
> > > Cc: Borislav Petkov <bp@alien8.de>
> > > Cc: H. Peter Anvin <hpa@zytor.com>
> > > Cc: Roland McGrath <roland@hack.frob.com>
> > > ---
> > >  arch/x86/kernel/entry_64.S |    5 +++++
> > >  1 files changed, 5 insertions(+), 0 deletions(-)
> > > 
> > > diff --git a/arch/x86/kernel/entry_64.S b/arch/x86/kernel/entry_64.S
> > > index 111f6bb..0444917 100644
> > > --- a/arch/x86/kernel/entry_64.S
> > > +++ b/arch/x86/kernel/entry_64.S
> > > @@ -516,6 +516,11 @@ GLOBAL(system_call_after_swapgs)
> > >  	SAVE_ARGS 8,0
> > >  	movq  %rax,ORIG_RAX-ARGOFFSET(%rsp)
> > >  	movq  %rcx,RIP-ARGOFFSET(%rsp)
> > > +#ifdef CONFIG_PERF_EVENTS
> > > +	/* We need rsp in fast path for perf post unwind. */
> > > +	movq  PER_CPU_VAR(old_rsp), %rcx
> > > +	movq  %rcx,RSP-ARGOFFSET(%rsp)
> > > +#endif
> > 
> > Another solution is to set/unset some TIF flag in perf_event_sched_in/out
> > such that we take the syscall slow path (tracesys) which records every non-scratch
> > registers. I can see that old_rsp is not saved in pt_regs by SAVE_REST so this may be
> > something to add in tracesys.
> > 
> > This way we don't bloat the syscall fastpath with a feature only used by some
> > developers.
> 
> ok, this could work for task related events
> 
> but will need to think about how to do this for cpu related events which
> hit the same issue and we dont have a task to flag.. maybe do it the same
> way as for TIF_SYSCALL_TRACEPOINT and flag everyone ;)

You can also hook into tasks sched_in/out on cpu wide events.
Just add something like:

diff --git a/kernel/events/core.c b/kernel/events/core.c
index 5b06cbb..be8f18a 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -5987,7 +5987,8 @@ done:
 	}
 
 	if (!event->parent) {
-		if (event->attach_state & PERF_ATTACH_TASK)
+		if (event->attach_state & PERF_ATTACH_TASK ||
+		    event->attr.sample_type & PERF_SAMPLE_USER_REGS)
 			static_key_slow_inc(&perf_sched_events.key);
 		if (event->attr.mmap || event->attr.mmap_data)
 			atomic_inc(&nr_mmap_events);