In the CONFIG_CC_HAS_ASM_GOTO_OUTPUT=y version of __get_mem_asm(), we
incorrectly use _ASM_EXTABLE_##type##ACCESS_ERR() such that upon a fault
the extable fixup handler writes -EFAULT into "%w0", which is the
register containing 'x' (the result of the load).

This was a thinko in commit:

  86a6a68f ("arm64: start using 'asm goto' for get_user() when available")

Prior to that commit _ASM_EXTABLE_##type##ACCESS_ERR_ZERO() was used
such that the extable fixup handler wrote -EFAULT into "%w0" (the
register containing 'err'), and zero into "%w1" (the register containing
'x'). When the 'err' variable was removed, the extable entry was updated
incorrectly.

Writing -EFAULT to the value register is unnecessary but benign:

* We never want -EFAULT in the value register, and previously this would
  have been zeroed in the extable fixup handler.

* In __get_user_error() the value is overwritten with zero explicitly in
  the error path.

* The asm goto outputs cannot be used when the goto label is taken, as
  older compilers (e.g. clang < 16.0.0) do not guarantee that asm goto
  outputs are usable in this path and may use a stale value rather than
  the value in an output register. Consequently, zeroing in the extable
  fixup handler is insufficient to ensure callers see zero in the error
  path.

* The expected usage of unsafe_get_user() and get_kernel_nofault()
  requires that the value is not consumed in the error path.

Some versions of GCC would mis-compile asm goto with outputs, and
erroneously omit subsequent assignments, breaking the error path
handling in __get_user_error(). This was discussed at:

  https://lore.kernel.org/lkml/ZpfxLrJAOF2YNqCk@J2N7QTR9R3.cambridge.arm.com/

... and was fixed by removing support for asm goto with outputs on those
broken compilers in commit:

  f2f6a8e8 ("init/Kconfig: remove CONFIG_GCC_ASM_GOTO_OUTPUT_WORKAROUND")

With that out of the way, we can safely replace the usage of
_ASM_EXTABLE_##type##ACCESS_ERR() with _ASM_EXTABLE_##type##ACCESS(),
leaving the value register unchanged in the case a fault is taken, as
was originally intended. This matches other architectures and matches
our __put_mem_asm().
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Will Deacon <will@kernel.org>
Link: https://lore.kernel.org/r/20240807103731.2498893-1-mark.rutland@arm.comSigned-off-by: Catalin Marinas <catalin.marinas@arm.com>