ubsan, mips: excessive stack usage with -fsanitize=alignment

[arndb]

Vincent got a complaint from a bot about a change he made in the scheduler:

https://groups.google.com/d/msgid/clang-built-linux/CAKfTPtB9B9jv_uSk0fS7uOTtH5FCwLcAOrPjtXzRuURLUMFnoA%40mail.gmail.com

I had a look at this, and it seems that clang ends up excessively spilling local variables when building with -fsanitize=alignment on mips32, causing hundreds of warnings like:

../fs/select.c:621:5: warning: stack frame size of 1048 bytes in function 'core_sys_select' [-Wframe-larger-than=]
int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
    ^
1 warning generated.
../drivers/bus/mhi/core/pm.c:740:6: warning: stack frame size of 1400 bytes in function 'mhi_pm_st_worker' [-Wframe-larger-than=]
void mhi_pm_st_worker(struct work_struct *work)
     ^
1 warning generated.
../kernel/sched/fair.c:8990:23: warning: unused variable 'child' [-Wunused-variable]
        struct sched_domain *child = env->sd->child;
                             ^
../kernel/sched/fair.c:8527:22: warning: unused function 'update_sd_pick_busiest' [-Wunused-function]
static noinline bool update_sd_pick_busiest(struct lb_env *env,
                     ^
2 warnings generated.
../security/apparmor/policy_unpack.c:1163:5: warning: stack frame size of 1152 bytes in function 'aa_unpack' [-Wframe-larger-than=]
int aa_unpack(struct aa_loaddata *udata, struct list_head *lh,
    ^
1 warning generated.
../kernel/acct.c:426:32: warning: implicit conversion from 'int' to 'char' changes value from 131 to -125 [-Wconstant-conversion]
        ac->ac_version = ACCT_VERSION | ACCT_BYTEORDER;
                       ~ ~~~~~~~~~~~~~^~~~~~~~~~~~~~~~
1 warning generated.
../drivers/dma/dmatest.c:568:12: warning: stack frame size of 1464 bytes in function 'dmatest_func' [-Wframe-larger-than=]
static int dmatest_func(void *data)
           ^

From the assembler code I can see that it checks every pointer for correct alignment before derefencing it, and that apparently lets the register allocator run out of registers.

I tried the same thing on armv5, which also doesn't have native alignment support, but this did not show the same problem, so I guess it's something about the mips backend.

[arndb]

And now I found that there is already a workaround in the kernel for it in Kconfig:

config UBSAN_ALIGNMENT
        bool "Perform checking for misaligned pointer usage"
        default !HAVE_EFFICIENT_UNALIGNED_ACCESS
        depends on !UBSAN_TRAP && !COMPILE_TEST
        depends on $(cc-option,-fsanitize=alignment)

The build bot that reported it did not set COMPILE_TEST, otherwise it should not have been able to turn this option on in the absence of CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS for this target.

[nathanchance]

Seems like this might be a problem for 32-bit PowerPC too? https://lore.kernel.org/r/202106100743.JgPkQBD1-lkp@intel.com/

# Enable CONFIG_UBSAN and CONFIG_UBSAN_ALIGNMENT in menuconfig
$ make -skj"$(nproc)" ARCH=powerpc CROSS_COMPILE=powerpc-linux-gnu- LLVM=1 distclean mpc83xx_defconfig menuconfig drivers/crypto/talitos.o
...
drivers/crypto/talitos.c:3328:12: warning: stack frame size of 1040 bytes in function 'talitos_probe' [-Wframe-larger-than=]
static int talitos_probe(struct platform_device *ofdev)
           ^
...