[PATCH v4 0/4] x86/uaccess: Use pointer masking to limit uaccess speculation

From: Josh Poimboeuf
Date: Tue May 04 2021 - 23:54:53 EST


This one managed to fall through the cracks back in September. Here's a
fresh new version.

Ideally, we'd switch all access_ok() users to access_ok_mask() or
something, but that's a much bigger change.

I dropped all the ack/review tags because the rebase was significant.

Please review carefully :-)


v4 changes:

- Rebased on the latest.

- Split up into multiple logical patches.

- Renamed "force_user_ptr()" -> "mask_user_ptr()" to prevent confusing
it with '__force' casting. [based on Dan's comment]

- Instead of reusing array_index_nospec(), made a new separate inline
asm statement. Otherwise it fails the build on recent toolchains
and/or kernels because the "g" constraint in array_index_mask_nospec()
isn't big enough for TASK_SIZE_MAX. I could have changed "g" to "r",
but that would negatively impact code generation for the other users.


v3 was here:

https://lore.kernel.org/lkml/1d06ed6485b66b9f674900368b63d7ef79f666ca.1599756789.git.jpoimboe@xxxxxxxxxx/


Josh Poimboeuf (4):
uaccess: Always inline strn*_user() helper functions
uaccess: Fix __user annotations for copy_mc_to_user()
x86/uaccess: Use pointer masking to limit uaccess speculation
x86/nospec: Remove barrier_nospec()

Documentation/admin-guide/hw-vuln/spectre.rst | 6 +--
arch/x86/include/asm/barrier.h | 3 --
arch/x86/include/asm/futex.h | 5 ++
arch/x86/include/asm/uaccess.h | 48 +++++++++++++------
arch/x86/include/asm/uaccess_64.h | 12 ++---
arch/x86/kernel/cpu/sgx/virt.c | 6 ++-
arch/x86/lib/copy_mc.c | 10 ++--
arch/x86/lib/csum-wrappers_64.c | 5 +-
arch/x86/lib/getuser.S | 16 ++-----
arch/x86/lib/putuser.S | 8 ++++
arch/x86/lib/usercopy_32.c | 6 +--
arch/x86/lib/usercopy_64.c | 7 +--
lib/iov_iter.c | 2 +-
lib/strncpy_from_user.c | 6 ++-
lib/strnlen_user.c | 4 +-
15 files changed, 89 insertions(+), 55 deletions(-)

--
2.31.1