Re: [patch 15/14] x86/dumpstack/64: Speedup in_exception_stack()

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Apr 02, 2019 at 10:43:29AM -0500, Josh Poimboeuf wrote:
> On Tue, Apr 02, 2019 at 12:19:46PM +0200, Thomas Gleixner wrote:
> >  static bool in_exception_stack(unsigned long *stack, struct stack_info *info)
> >  {
> > -	unsigned long estacks, begin, end, stk = (unsigned long)stack;
> > +	unsigned long begin, end, stk = (unsigned long)stack;
> > +	const struct estack_pages *ep;
> >  	struct pt_regs *regs;
> >  	unsigned int k;
> >  
> >  	BUILD_BUG_ON(N_EXCEPTION_STACKS != 4);
> >  
> > -	estacks = (unsigned long)__this_cpu_read(cea_exception_stacks);
> > -
> > -	for (k = 0; k < N_EXCEPTION_STACKS; k++) {
> > -		begin = estacks + layout[k].begin;
> > -		end   = estacks + layout[k].end;
> > -		regs  = (struct pt_regs *)end - 1;
> > -
> > -		if (stk <= begin || stk >= end)
> > -			continue;
> > -
> > -		info->type	= STACK_TYPE_EXCEPTION + k;
> > -		info->begin	= (unsigned long *)begin;
> > -		info->end	= (unsigned long *)end;
> > -		info->next_sp	= (unsigned long *)regs->sp;
> > -
> > -		return true;
> > -	}
> > -
> > -	return false;
> > +	begin = (unsigned long)__this_cpu_read(cea_exception_stacks);
> > +	end = begin + sizeof(struct cea_exception_stacks);
> > +	/* Bail if @stack is outside the exception stack area. */
> > +	if (stk <= begin || stk >= end)
> > +		return false;
> 
> This check is the most important piece.  Exception stack dumps are quite
> rare, so this ensures an early exit in most cases regardless of whether
> there's a loop below.
> 
> > +
> > +	/* Calc page offset from start of exception stacks */
> > +	k = (stk - begin) >> PAGE_SHIFT;
> > +	/* Lookup the page descriptor */
> > +	ep = &estack_pages[k];
> > +	/* Guard page? */
> > +	if (unlikely(!ep->size))
> > +		return false;
> > +
> > +	begin += (unsigned long)ep->offs;
> > +	end = begin + (unsigned long)ep->size;
> > +	regs = (struct pt_regs *)end - 1;
> > +
> > +	info->type	= ep->type;
> > +	info->begin	= (unsigned long *)begin;
> > +	info->end	= (unsigned long *)end;
> > +	info->next_sp	= (unsigned long *)regs->sp;
> > +	return true;
> 
> With the above "(stk <= begin || stk >= end)" check, removing the loop
> becomes not all that important since exception stack dumps are quite
> rare and not performance sensitive.  With all the macros this code
> becomes a little more obtuse, so I'm not sure whether removal of the
> loop is a net positive.

Are you sure; perf does a lot of stack dumps from NMI context.