diff mbox series

arm64/fpsimd: Don't flush SME register hardware state along with thread

Message ID 20230724-arm64-dont-flush-smstate-v1-1-9a8b637ace6c@kernel.org
State New
Headers
Series arm64/fpsimd: Don't flush SME register hardware state along with thread |

Commit Message

Mark Brown July 24, 2023, 1:09 p.m. UTC 
  We recently changed the fpsimd thread flush to flush the physical SME
state as well as the thread state for the current thread.  Unfortunately
this leads to intermittent corruption in interaction with the lazy
FPSIMD register switching.  When under heavy load such as can be
triggered by the startup phase of fp-stress it is possible that the
current thread may not be scheduled prior to returning to userspace, and
indeed we may end up returning to the last thread that was scheduled on
the PE without ever exiting the kernel to any other task.  If that
happens then we will not reload the register state from memory, leading
to loss of any SME register state.

Since this was purely an attempt to defensively close off potential
problems revert the change.

Fixes: af3215fd0230 Fixes: ("arm64/fpsimd: Exit streaming mode when flushing tasks")
Signed-off-by: Mark Brown <broonie@kernel.org>
---
 arch/arm64/kernel/fpsimd.c | 1 -
 1 file changed, 1 deletion(-)


---
base-commit: 6eaae198076080886b9e7d57f4ae06fa782f90ef
change-id: 20230721-arm64-dont-flush-smstate-c109f4b3f31c

Best regards,

Comments

Catalin Marinas July 26, 2023, 6:45 p.m. UTC | #1 
On Mon, 24 Jul 2023 14:09:19 +0100, Mark Brown wrote:
> We recently changed the fpsimd thread flush to flush the physical SME
> state as well as the thread state for the current thread.  Unfortunately
> this leads to intermittent corruption in interaction with the lazy
> FPSIMD register switching.  When under heavy load such as can be
> triggered by the startup phase of fp-stress it is possible that the
> current thread may not be scheduled prior to returning to userspace, and
> indeed we may end up returning to the last thread that was scheduled on
> the PE without ever exiting the kernel to any other task.  If that
> happens then we will not reload the register state from memory, leading
> to loss of any SME register state.
> 
> [...]

Applied to arm64 (for-next/fixes), thanks!

[1/1] arm64/fpsimd: Don't flush SME register hardware state along with thread
      https://git.kernel.org/arm64/c/3421ddbe6d64
diff mbox series

Patch

diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c
index 89d54a5242d1..9d7d10d60bfd 100644
--- a/arch/arm64/kernel/fpsimd.c
+++ b/arch/arm64/kernel/fpsimd.c
@@ -1666,7 +1666,6 @@  void fpsimd_flush_thread(void)
 
 		fpsimd_flush_thread_vl(ARM64_VEC_SME);
 		current->thread.svcr = 0;
-		sme_smstop();
 	}
 
 	current->thread.fp_type = FP_STATE_FPSIMD;