Re: [PATCH] reduce large do_mount stack usage with noinlines

[Andrew Morton]

On Wed, 06 Feb 2008 17:34:39 -0600
Eric Sandeen <sandeen@xxxxxxxxxx> wrote:

> Andrew Morton wrote:
> > On Wed, 06 Feb 2008 17:11:38 -0600
> > Eric Sandeen <sandeen@xxxxxxxxxx> wrote:
> > 
> >>  /*
> >>   * recursively change the type of the mountpoint.
> >> + * noinline this do_mount helper to save do_mount stack space.
> >>   */
> >> -static int do_change_type(struct nameidata *nd, int flag)
> >> +static noinline int do_change_type(struct nameidata *nd, int flag)
> > 
> > What we could do here is defined a new noinline_because_of_stack_suckiness
> > and use that.  Reasons:
> > 
> > - self-documenting, so we don't need to comment each site
> > 
> > - can be made a no-op for suitable __GNUC__ values if gcc ever fixes this
> > 
> > - our children we can go through and delete them all when nobody is using
> >   the offending gcc versions any more.
> > 
> > what thinkest thou?
> 
> Yes, sounds very good to me.  I'm sure there are more places which could
> use this.
> 
> Or perhaps there are some magic flags to gcc so that it doesn't inline
> so aggressively in situations like this?  IOW is it gcc breakage, or
> design - but maybe with defaults that don't fit well in the limited
> stack... (well, I suppose the "for N mutually exclusive helper functions
> each with stack usage S, use about N*S stack when inlining them all" bit
> probably isn't a feature.)

The auto inlining is OK.  The problem is that when gcc handles this:

static inline foo()
{
	char a[10];
}

static inline bar()
{
	char a[10];
}

zot()
{
	foo();
	bar();
}

it will use 20 bytes of stack instead of using the same 10 bytes for both
"calls".  It doesn't need to do that, and other compilers avoid it, iirc.


Now, it _used_ to be the case that when presented with this:

foo()
{
	{
		char a[10];

		bar(a);
	}
	{
		char a[10];

		bar(a);
	}
}

gcc would also consume 20 bytes of stack.  But I see that this is fixed in
gcc-4.0.3.

These two things are equivalent and hopefully gcc will soon fix the
inlined-functions-use-too-much-stack thing.  Once that happens, we don't
need your patch.


> FWIW the XFS team finally just wrested control from GCC.
> http://oss.sgi.com/archives/xfs/2006-11/msg00195.html
> 

yup.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/