Re: [PATCH] x86/fpu: Delay instruction pointer fixup until after after warning

[Dave Hansen]

On 6/18/25 12:51, Alison Schofield wrote:
>> Do not fix up RIP until after printing the warning.
> How was this found and how is the change verified?

Good questions.

I found it from an Intel-internal bug report. It's not clear what's
causing the underlying XRSTOR #GP. But I spent some time scratching my
head about how RIP got pointing to the wrong place. I was blaming the
simulator at first.

I validated the fix using the attached patch. It waits until there's a
program named "dave" running, then corrupts the XSAVE buffer in a way
that will cause XRSTOR to #GP, triggering the warning that was off by an
instruction.


---

 b/arch/x86/kernel/fpu/core.c |    3 +++
 1 file changed, 3 insertions(+)

diff -puN arch/x86/kernel/fpu/core.c~os_rstor_fun arch/x86/kernel/fpu/core.c
--- a/arch/x86/kernel/fpu/core.c~os_rstor_fun	2025-06-18 11:22:58.583510842 -0700
+++ b/arch/x86/kernel/fpu/core.c	2025-06-18 11:23:46.626730032 -0700
@@ -202,6 +202,9 @@ void restore_fpregs_from_fpstate(struct
 		 */
 		mask = fpu_kernel_cfg.max_features & mask;
 
+		if (!strncmp(current->comm, "dave", 4))
+			fpstate->regs.xsave.header.xcomp_bv = 0;
+
 		os_xrstor(fpstate, mask);
 	} else {
 		if (use_fxsr())
_