[PATCH] x86-64: Fix FPU corruption
Andi Kleen [Thu, 5 Oct 2006 16:47:22 +0000 (18:47 +0200)]
This reverts an earlier patch that was found to cause FPU
state corruption. I think the corruption happens because
unlazy_fpu() can cause FPU exceptions and when it happens
after the current switch some processing would affect
the state in the wrong process.

Thanks to  Douglas Crosher and Tom Hughes for testing.

Cc: jbeulich@novell.com

Signed-off-by: Andi Kleen <ak@suse.de>

arch/x86_64/kernel/process.c

index de10cb8..5e95b25 100644 (file)
@@ -615,6 +615,9 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
                prev->gsindex = gsindex;
        }
 
+       /* Must be after DS reload */
+       unlazy_fpu(prev_p);
+
        /* 
         * Switch the PDA and FPU contexts.
         */
@@ -622,10 +625,6 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
        write_pda(oldrsp, next->userrsp); 
        write_pda(pcurrent, next_p); 
 
-       /* This must be here to ensure both math_state_restore() and
-          kernel_fpu_begin() work consistently. 
-          And the AMD workaround requires it to be after DS reload. */
-       unlazy_fpu(prev_p);
        write_pda(kernelstack,
        (unsigned long)task_stack_page(next_p) + THREAD_SIZE - PDA_STACKOFFSET);
 #ifdef CONFIG_CC_STACKPROTECTOR