Peter, please take a look at unsafe_try_cmpxchg_user() to make sure it
(a) still looks right and (b) works for your use case. I added explicit
casts to play nice with clang and sparse, as well as a __chk_user_ptr()
given the use of __force.
Add uaccess macros for doing CMPXCHG on userspace addresses and use the
macros to fix KVM bugs by replacing flawed code that maps memory into the
kernel address space without proper mmu_notifier protection (or with
broken pfn calculations in one case).
Add yet another Kconfig for guarding asm_volatile_goto() to workaround a
clang-13 bug. I've verified the test passes on gcc versions of arm64,
PPC, RISC-V, and s390x that also pass the CC_HAS_ASM_GOTO_OUTPUT test.
Patches 1-4 are tagged for stable@ as patches 3 and 4 (mostly 3) need a
backportable fix, and doing CMPXCHG on the userspace address is the
simplest fix from a KVM perspective.
v2:
- Explicitly cast with (__force u<size> *) to fix clang+i386
compilation woes and sparse warnings. [kernel test robot]
- Rework i386's CMPXCHG8B to force use of ECX for the error path so that
clang doesn't run out of input/output GPRs...
- Document that gcc also has/had troubles, and note the clang and gcc
versions that (should) work with tied outputs. [Nick]
- Collect tags [Nick, Tadeusz]
v1: https://lore.kernel.org/all/20220201010838.1494405-1-seanjc@xxxxxxxxxx
Peter Zijlstra (1):
x86/uaccess: Implement macros for CMPXCHG on user addresses
Sean Christopherson (4):
Kconfig: Add option for asm goto w/ tied outputs to workaround
clang-13 bug
KVM: x86: Use __try_cmpxchg_user() to update guest PTE A/D bits
KVM: x86: Use __try_cmpxchg_user() to emulate atomic accesses
KVM: x86: Bail to userspace if emulation of atomic user access faults
arch/x86/include/asm/uaccess.h | 142 +++++++++++++++++++++++++++++++++
arch/x86/kvm/mmu/paging_tmpl.h | 45 +----------
arch/x86/kvm/x86.c | 35 ++++----
init/Kconfig | 5 ++
4 files changed, 162 insertions(+), 65 deletions(-)
base-commit: 26291c54e111ff6ba87a164d85d4a4e134b7315c
--
2.35.0.rc2.247.g8bbb082509-goog
