[PATCH 3/3] lib/core/fiber: Put watermarks into stack to track its usage

[Cyrill Gorcunov]

On Wed, Mar 13, 2019 at 04:52:04PM +0300, Vladimir Davydov wrote:
> >  
> > +static bool
> > +stack_has_wmark(void *addr)
> 
> Please add brief comments to all the functions you add.

ok

> > +
> > +	/*
> > +	 * When dropping pages make sure the page
> > +	 * containing overflow mark is untouched.
> > +	 * Same time no need to unmap the page which
> > +	 * carries "shrink" wmark, since we're updating
> > +	 * this page anyway.
> > +	 */
> > +	if (stack_direction < 0) {
> > +		end = page_align_down(fiber->stack_shrink_wmark);
> > +		start = page_align_up(fiber->stack_overflow_wmark);
> > +		madvise(start, end - start, MADV_DONTNEED);
> > +	} else {
> > +		start = page_align_up(fiber->stack_shrink_wmark);
> > +		end = page_align_down(fiber->stack_overflow_wmark);
> > +		madvise(start, end - start, MADV_DONTNEED);
> > +	}
> 
> Nit: you could move madvise() out of the if-else block.

indeed, thanks!

> > +
> > +	if (!stack_has_wmark(fiber->stack_overflow_wmark)) {
> > +		say_warn("stack usage is close to the limit of %zu bytes",
> > +			 (size_t)FIBER_STACK_SIZE_DEFAULT);
> 
> IMO warning only when we are close to the limit isn't very useful.
> Let's keep track of the max stack size instead and print it whenever
> it's increased.

Could you please elaborate. I don't understand what you mean by
keep tracking of max stack size. You propose to remember somewhere
if this mark has been vanished?

> > +	fiber->wmark_inpage_offset = ((rand() % 128) + 8) & ~7;
> 
> Please define a constant for this.

ok

> > +		fiber->stack_shrink_wmark -= 16 << 12;
> 
> Ditto.

ok

> > @@ -927,8 +1090,12 @@ cord_create(struct cord *cord, const char *name)
> >  	/* Record stack extents */
> >  	tt_pthread_attr_getstack(cord->id, &cord->sched.stack,
> >  				 &cord->sched.stack_size);
> > +	cord->sched.stack_overflow_wmark = cord->sched.stack;
> > +	cord->sched.stack_shrink_wmark = cord->sched.stack;
> 
> I don't think we need to set the watermarks for the sched fiber,
> even if ASAN is on.

Won't they left unitialized then? The only reason I put nils here
is to force them have known values even if they are unused.

	Cyrill