Re: [RFA][PATCH 2/5] ftrace/x86: One more missing sync after fixup of function modification failure

[Steven Rostedt <rostedt@goodmis.org>]

On Thu, 27 Feb 2014 18:19:37 +0100
Frederic Weisbecker <fweisbec@gmail.com> wrote:

> On Thu, Feb 27, 2014 at 12:00:14PM -0500, Steven Rostedt wrote:
> > On Thu, 27 Feb 2014 17:37:32 +0100
> > Frederic Weisbecker <fweisbec@gmail.com> wrote:
> > 
> > > On Thu, Feb 27, 2014 at 10:46:18AM -0500, Steven Rostedt wrote:
> > > > [Request for Ack]
> > > > 
> > > > From: Petr Mladek <pmladek@suse.cz>
> > > > 
> > > > If a failure occurs while modifying ftrace function, it bails out and will
> > > > remove the tracepoints to be back to what the code originally was.
> > > > 
> > > > There is missing the final sync run across the CPUs after the fix up is done
> > > > and before the ftrace int3 handler flag is reset.
> > > 
> > > So IIUC the risk is that other CPUs may spuriously ignore non-ftrace traps if we don't sync the
> > > other cores after reverting the int3 before decrementing the modifying_ftrace_code counter?
> > 
> > Actually, the bug is that they will not ignore the ftrace traps after
> > we decrement modifying_ftrace_code counter. Here's the race:
> > 
> > 	CPU0				CPU1
> > 	----				----
> >   remove_breakpoint();
> >   modifying_ftrace_code = 0;
> > 
> > 				[still sees breakpoint]
> > 				<takes trap>
> > 				[sees modifying_ftrace_code as zero]
> > 				[no breakpoint handler]
> > 				[goto failed case]
> > 				[trap exception - kernel breakpoint, no
> > 				 handler]
> > 				BUG()
> > 
> > 
> > Even if we had a smp_wmb() after removing the breakpoint and clearing
> > the modifying_ftrace_code, we still need the smp_rmb() on the other
> > CPUS. The run_sync() does a IPI on all CPUs doing the smp_rmb().
> 
> Ah ok. My understanding was indeed that it doesn't ignore the ftrace trap,
> but I thought the consequence was that we return immediately from the trap
> handler.

I'll add my above cpu race diagram (is that what we call it?). That
should make this change more understandable.


> Ok but what I meant is to do this instead:
> 
>  fail_update:
>     probe_kernel_write((void *)ip, &old_code[0], 1);
> +   run_sync()
>     goto out;
> 
> Because with the current patch we also call run_sync() on add_break() failure.

Ah ok (my turn to understand). Yeah, if the add_break() fails, then we
don't need to do the run_sync().

But this is just for now, to prevent the add_update_code() error from
crashing. I have more patches that clean this up further. But they are
for 3.15.

-- Steve