[PATCH v2] crypto: ecc - Unbreak the build on arm with CONFIG_KASAN_STACK=y

[Lukas Wunner]

Andrew reports build breakage of arm allmodconfig, reproducible with gcc
14.2.0 and 15.2.0:

  crypto/ecc.c: In function 'ecc_point_mult':
  crypto/ecc.c:1380:1: error: the frame size of 1360 bytes is larger than 1280 bytes [-Werror=frame-larger-than=]

gcc aggressively inlines functions called by ecc_point_mult() (without
there being any explicit inline declarations), which pushes stack usage
close to the limit imposed by CONFIG_FRAME_WARN.  allmodconfig implies
CONFIG_KASAN_STACK=y, which increases the stack above that limit.

In the bugzilla entry linked below, gcc maintainers explain that gcc
estimates extra stack usage caused by inlining, but ASAN instrumentation
is added in post-IPA passes and thus the inlining heuristics cannot
account for it.

It could be argued that -Werror=frame-larger-than=1280 instructs the
compiler to avoid inlining beyond that limit lest the build breaks,
which would imply gcc behaves incorrectly.  But gcc maintainers reject
this notion and believe that a warning switch should never affect code
generation, even if it is promoted to an error.

One way to unbreak the build is to limit inlining via -finline-limit=100
or by explicitly declaring some functions noinline.  However while it
does keep stack usage of individual functions below the limit, *total*
stack usage increases.

A longterm solution is to refactor ecc.c for reduced stack usage.  It
currently performs ECC point multiplication with a Montgomery ladder
which uses co-Z (conjugate) addition to trade off memory for speed.
The algorithm is susceptible to timing attacks and needs to be replaced
with a constant time Montgomery ladder, which should consume less memory
and thus resolve the stack usage issue as a side effect.

In the interim, raise the limit for ecc.c, as is already done for
several other files in the source tree.

Constrain to gcc because clang 19.1.7 does not exhibit the issue.  It
makes do with a 724 bytes stack frame even though it inlines almost the
same functions as gcc.

Link: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124949
Reported-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> # off-list
Signed-off-by: Lukas Wunner <lukas@xxxxxxxxx>
Acked-by: Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx>
---
Changes v1 -> v2:
* s/ARCH/CONFIG_ARM/, s/LLVM/CONFIG_CC_IS_GCC/ (Nathan)
* Add link to gcc bugzilla entry
* Rewrite commit message to include feedback provided by gcc maintainers
  and explain high stack usage with algorithm choice

Link to v1:
https://lore.kernel.org/r/abfaede9ab2e963d784fb70598ed74935f7f8d93.1775628469.git.lukas@xxxxxxxxx/

 crypto/Makefile | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/crypto/Makefile b/crypto/Makefile
index 1622425..c73f4d5 100644
--- a/crypto/Makefile
+++ b/crypto/Makefile
@@ -178,6 +178,11 @@ obj-$(CONFIG_CRYPTO_ZSTD) += zstd.o
 obj-$(CONFIG_CRYPTO_ECC) += ecc.o
 obj-$(CONFIG_CRYPTO_ESSIV) += essiv.o
 
+# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124949
+ifeq ($(CONFIG_ARM)$(CONFIG_KASAN_STACK)$(CONFIG_CC_IS_GCC),yyy)
+CFLAGS_ecc.o += $(call cc-option,-Wframe-larger-than=1536)
+endif
+
 ecdh_generic-y += ecdh.o
 ecdh_generic-y += ecdh_helper.o
 obj-$(CONFIG_CRYPTO_ECDH) += ecdh_generic.o
-- 
2.51.0