Re: [PATCH] perf: Fix oops when kthread execs user process

[Will Deacon]

On Wed, May 29, 2019 at 06:19:55PM +0200, Peter Zijlstra wrote:
> On Wed, May 29, 2019 at 03:35:10PM +0100, Will Deacon wrote:
> > On Wed, May 29, 2019 at 03:25:15PM +0200, Peter Zijlstra wrote:
> > > On Wed, May 29, 2019 at 02:05:21PM +0100, Will Deacon wrote:
> > > > On Wed, May 29, 2019 at 02:55:57PM +0200, Peter Zijlstra wrote:
> > > 
> > > > >  	if (user_mode(regs)) {
> > > > 
> > > > Hmm, so it just occurred to me that Mark's observation is that the regs
> > > > can be junk in some cases. In which case, should we be checking for
> > > > kthreads first?
> > > 
> > > task_pt_regs() can return garbage, but @regs is the exception (or
> > > perf_arch_fetch_caller_regs()) regs, and for those user_mode() had
> > > better be correct.
> > 
> > So what should we report for the idle task?
> 
> If an interrupt hits the idle task, @regs would be !user_mode(regs),
> we'll find current->flags & PF_KTHREAD (idle not having passed through
> exec()) and therefore we'll take ABI_NONE for the user regs.
> 
> Or am I not getting it?

Sorry, I'm not trying to catch you out! Just trying to understand what the
semantics are supposed to be.

I do find the concept of user_mode(regs) bizarre for the idle task. By the
above, we definitely have a bug on arm64 (user_mode(regs) tends to be
true for the idle task), and I couldn't figure out how you avoided it on
x86. I guess it happens to work because the stack is zero-initialised or
something?

Will