On 09/23/2017 12:29 AM, Arnd Bergmann wrote:

We get a lot of very large stack frames using gcc-7.0.1 with the default
-fsanitize-address-use-after-scope --param asan-stack=1 options, which
can easily cause an overflow of the kernel stack, e.g.

drivers/gpu/drm/i915/gvt/handlers.c:2407:1: error: the frame size of 31216 bytes is larger than 2048 bytes
drivers/net/wireless/ralink/rt2x00/rt2800lib.c:5650:1: error: the frame size of 23632 bytes is larger than 2048 bytes
drivers/scsi/fnic/fnic_trace.c:451:1: error: the frame size of 5152 bytes is larger than 2048 bytes
fs/btrfs/relocation.c:1202:1: error: the frame size of 4256 bytes is larger than 2048 bytes
fs/fscache/stats.c:287:1: error: the frame size of 6552 bytes is larger than 2048 bytes
lib/atomic64_test.c:250:1: error: the frame size of 12616 bytes is larger than 2048 bytes
mm/vmscan.c:1367:1: error: the frame size of 5080 bytes is larger than 2048 bytes
net/wireless/nl80211.c:1905:1: error: the frame size of 4232 bytes is larger than 2048 bytes

To reduce this risk, -fsanitize-address-use-after-scope is now split
out into a separate CONFIG_KASAN_EXTRA Kconfig option, leading to stack
frames that are smaller than 2 kilobytes most of the time on x86_64. An
earlier version of this patch also prevented combining KASAN_EXTRA with
KASAN_INLINE, but that is no longer necessary with gcc-7.0.1.

A lot of warnings with KASAN_EXTRA go away if we disable KMEMCHECK,
as -fsanitize-address-use-after-scope seems to understand the builtin
memcpy, but adds checking code around an extern memcpy call. I had to work
around a circular dependency, as DEBUG_SLAB/SLUB depended on !KMEMCHECK,
while KASAN did it the other way round. Now we handle both the same way
and make KASAN and KMEMCHECK mutually exclusive.

All patches to get the frame size below 2048 bytes with CONFIG_KASAN=y
and CONFIG_KASAN_EXTRA=n have been submitted along with this patch, so
we can bring back that default now. KASAN_EXTRA=y still causes lots of
warnings but now defaults to !COMPILE_TEST to disable it in allmodconfig,
and it remains disabled in all other defconfigs since it is a new option.
I arbitrarily raise the warning limit for KASAN_EXTRA to 3072 to reduce
the noise, but an allmodconfig kernel still has around 50 warnings
on gcc-7.

I experimented a bit more with smaller stack frames and have another
follow-up series that reduces the warning limit for 64-bit architectures
to 1280 bytes (without CONFIG_KASAN).

With earlier versions of this patch series, I also had patches to
address the warnings we get with KASAN and/or KASAN_EXTRA, using a
"noinline_if_stackbloat" annotation. That annotation now got replaced with
a gcc-8 bugfix (see https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81715)
and a workaround for older compilers, which means that KASAN_EXTRA is
now just as bad as before and will lead to an instant stack overflow in
a few extreme cases.

This reverts parts of commit commit 3f181b4 ("lib/Kconfig.debug: disable
-Wframe-larger-than warnings with KASAN=y").

Signed-off-by: Arnd Bergmann <arnd@arndb.de>

Acked-by: Andrey Ryabinin <redacted>