BUILD_BUG() failed with arm64's gki_defconfig

[nathanchance]

During Android's mainline merges, I noticed a patch removing some BUILD_BUG() calls that were triggering. This can be reproduced with a ToT LLVM and mainline kernel with Android's gki_defconfig so this is probably something that should be investigated and fixed.

$ clang --version | head -n1
ClangBuiltLinux clang version 14.0.0 (https://github.com/llvm/llvm-project 9b6f264d2b09ab5e970e54f77119eb823f0fcc50)

$ git show -s --format='%h ("%s")'
5833291ab6de ("Merge tag 'pci-v5.16-fixes-1' of git://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci")

# gitlies's txt export is base64 encoded so it has to be decoded...
$ curl -LSs 'https://android.googlesource.com/kernel/common/+/refs/heads/android-mainline/arch/arm64/configs/gki_defconfig?format=TEXT' | base64 -d > .config

# Android uses full LTO, which takes forever; this bug can be reproduced with ThinLTO
$ scripts/config -d LTO_CLANG_FULL -e LTO_CLANG_THIN

$ make -skj"$(nproc)" ARCH=arm64 LLVM=1 olddefconfig vmlinux
ld.lld: error: call to __compiletime_assert_200 marked "dontcall-error": BUILD_BUG failed
...

Brief initial triage leads me to believe this is an LTO problem, as outright disabling LTO makes the problem go away:

$ curl -LSs 'https://android.googlesource.com/kernel/common/+/refs/heads/android-mainline/arch/arm64/configs/gki_defconfig?format=TEXT' | base64 -d > .config

$ scripts/config -d LTO_CLANG_FULL

$ make -skj"$(nproc)" ARCH=arm64 LLVM=1 olddefconfig vmlinux

$ echo $?
0

I will see if I can figure out exactly what configuration options tickle this then create a small reproducer and report it upstream.

[nathanchance]

There is probably a way better IR reproducer for this but this is what cvise came up with (it appears CFI related in reality):

$ cat arm_arch_timer.i
struct thread_info {
  struct {
    struct {
      int count;
    } preempt;
  };
} preempt_schedule_notrace(void);
struct arch_timer_erratum_workaround {
  long (*read_cntvct_el0)();
} timer_unstable_counter_workaround;
long arch_timer_reg_read_cp15___val, __arch_counter_get_cntvct_stable___ptr,
    erratum_set_next_event_generic_cval;
typeof(&timer_unstable_counter_workaround)
    __arch_counter_get_cntvct_stable_pscr_ret__;
long arch_timer_read_cntvct_el0();
inline long arch_timer_reg_read_cp15(int access) {
  if (access == 0)
    return ({
      asm("");
      arch_timer_reg_read_cp15___val;
    });
  if (access == 1)
    return ({
      asm("");
      arch_timer_reg_read_cp15___val;
    });
  __attribute__((__noreturn__)) void __compiletime_assert_200(void)
      __attribute__((__error__("BUILD_BUG failed")));
  __compiletime_assert_200();
}
void __arch_counter_get_cntvct_stable() {
  __asm__("");
  ({
    struct arch_timer_erratum_workaround *__wa = ({
      ({
        __arch_counter_get_cntvct_stable_pscr_ret__ = ({
          *({
            ({
              (typeof((typeof(&timer_unstable_counter_workaround) *)0))
                  __arch_counter_get_cntvct_stable___ptr;
            });
          });
        });
      });
    });
    __wa && __wa->read_cntvct_el0 ? ({
      asm("");
      __wa->read_cntvct_el0;
    })
                                  : arch_timer_read_cntvct_el0;
  })("");
  long ti_0_0 = ({
    ({
      typeof(&ti_0_0) __x = &ti_0_0;
      asm("" : : "Q"(__x));
      0;
    });
  });
  preempt_schedule_notrace();
}
void arch_counter_get_cntpct_stable() {
  typeof(&((struct thread_info *)0)->preempt.count) __x =
      &((struct thread_info *)0)->preempt.count;
  switch (((struct thread_info *)0)->preempt.count)
  case 4:
    asm("" : : "Q"(__x));
  __asm__("");
  long ti_0_0 = ({
    ({
      typeof(&ti_0_0) __x = &ti_0_0;
      asm("" : : "Q"(__x));
      0;
    });
  });
}
static void erratum_set_next_event_generic(int access) {
  arch_timer_reg_read_cp15(access);
  arch_counter_get_cntpct_stable();
  asm("msr "
      "cntp_cval_el0"
      ", %x0"
      :
      : "rZ"(0));
  __arch_counter_get_cntvct_stable();
  asm("msr "
      "cntv_cval_el0"
      ", %x0"
      :
      : "rZ"(erratum_set_next_event_generic_cval));
}
void erratum_set_next_event_virt() {
  erratum_set_next_event_generic(1);
  erratum_set_next_event_generic(0);
}

# ThinLTO + CFI has the issue
$ clang -O2 --target=aarch64-linux-gnu -flto=thin -fsplit-lto-unit -fvisibility=hidden -fsanitize=cfi -fsanitize-cfi-cross-dso -fno-sanitize-cfi-canonical-jump-tables -fno-sanitize-trap=cfi -fno-sanitize-blacklist -c -o arm_arch_
timer.o arm_arch_timer.i

$ llvm-ar cDPrST arm_arch_timer.a arm_arch_timer.o

$ ld.lld -EL -maarch64elf -z norelro -mllvm -import-instr-limit=5 -r -o /dev/null --whole-archive arm_arch_timer.a
ld.lld: error: call to __compiletime_assert_200 marked "dontcall-error": BUILD_BUG failed

# Just ThinLTO is fine
$ clang -O2 --target=aarch64-linux-gnu -flto=thin -c -o arm_arch_timer.o arm_arch_timer.i

$ llvm-ar cDPrST arm_arch_timer.a arm_arch_timer.o

$ ld.lld -EL -maarch64elf -z norelro -mllvm -import-instr-limit=5 -r -o /dev/null --whole-archive arm_arch_timer.a

$ echo $?
0

cvise files here: https://github.com/nathanchance/creduce-files/tree/1475da898a391c5eeb19eaf57c5656530d21e920/cbl-1503

[nathanchance]

It looks like this can be reproduced with just defconfig + CONFIG_LTO_CLANG_THIN=y + CONFIG_CFI_CLANG=y (our CI is now hitting this): https://github.com/ClangBuiltLinux/continuous-integration2/runs/4212005132?check_suite_focus=true

[nathanchance]

Marc Zyngier sent https://lore.kernel.org/r/[email protected]/ to resolve this.

[nathanchance]

Merged into mainline: https://git.kernel.org/linus/1edb7e74a7d3d64dc4e69e7059b4eea526d19a10