RE: [PATCH] netlink: move nla_put_{u8,u16,u32} out of line

From: David Laight
Date: Fri Feb 10 2017 - 07:07:37 EST


From: David Miller
> Sent: 09 February 2017 21:31
> From: Arnd Bergmann <arnd@xxxxxxxx>
> Date: Wed, 8 Feb 2017 22:18:26 +0100
>
> > When CONFIG_KASAN is enabled, the "--param asan-stack=1" causes rather large
> > stack frames in some functions. This goes unnoticed normally because
> > CONFIG_FRAME_WARN is disabled with CONFIG_KASAN by default as of commit
> > 3f181b4d8652 ("lib/Kconfig.debug: disable -Wframe-larger-than warnings with
> > KASAN=y").
> >
> > The kernelci.org build bot however has the warning enabled and that led
> > me to investigate it a little further, as every build produces these warnings:
> >
> > net/wireless/nl80211.c:4389:1: warning: the frame size of 2240 bytes is larger than 2048 bytes [-
...
> >
> > It turns out that there is a relatively simple workaround for the netlink
> > users that currently use a local variable in order to do the type conversion:
> > Moving the three functions (for each of the typical sizes) to lib/nlattr.c
> > avoids using local variables in the caller, which drastically reduces the
> > stack usage for nl80211 and br_netlink.
> >
> > It would be good if we could enable the frame size check after that again,
> > but that should be a separate patch and it requires some more testing
> > to see which the largest acceptable frame size should be.
...
> You should only extern these things when KASAN is enabled.
>
> The reason is that uninlining these routines makes attribute emission
> more expensive and for some applications performance of this matters.

If performance of nla_put() matters, then adding 1, 2 and 4 byte attributes
ought to be doable without writing the values to memory and later doing
(I presume) a memcpy().

I also can't help feeling that the gcc KASAN stuff needs some way to
annotate an 'extern' to say that a value passed by reference isn't
treated as an array.
Otherwise I suspect you get a lot of bloat all over the place.

David