Re: [PATCH v2] x86/xen: Fix 32-bit PV guests's usage of kernel_stack

[David Vrabel <david.vrabel@citrix.com>]

On 10/04/14 17:17, Boris Ostrovsky wrote:
> Commit 198d208df4371734ac4728f69cb585c284d20a15 ("x86: Keep thread_info
> on thread stack in x86_32") made 32-bit kernels use kernel_stack to point to
> thread_info. That change missed a couple of updates needed by Xen's
> 32-bit PV guests:
> 
> 1. kernel_stack needs to be initialized for secondary CPUs
> 2. GET_THREAD_INFO() now uses %fs register which may not be the kernel's
> version when executing xen_iret().
> 
> With respect to the second issue, we don't need GET_THREAD_INFO()
> anymore: we used it as an intermediate step to get to per_cpu xen_vcpu and avoid
> referencing %fs. Now that we are going to use %fs anyway we may as well go
> directly to xen_vcpu.
> 
> Signed-off-by: Boris Ostrovsky <boris.ostrovsky@oracle.com>
> ---
>  arch/x86/xen/smp.c        |    3 ++-
>  arch/x86/xen/xen-asm_32.S |   25 +++++++++++++++++--------
>  2 files changed, 19 insertions(+), 9 deletions(-)
> 
> diff --git a/arch/x86/xen/smp.c b/arch/x86/xen/smp.c
> index a18eadd..7005974 100644
> --- a/arch/x86/xen/smp.c
> +++ b/arch/x86/xen/smp.c
> @@ -441,10 +441,11 @@ static int xen_cpu_up(unsigned int cpu, struct task_struct *idle)
>  	irq_ctx_init(cpu);
>  #else
>  	clear_tsk_thread_flag(idle, TIF_FORK);
> +#endif
>  	per_cpu(kernel_stack, cpu) =
>  		(unsigned long)task_stack_page(idle) -
>  		KERNEL_STACK_OFFSET + THREAD_SIZE;
> -#endif
> +
>  	xen_setup_runstate_info(cpu);
>  	xen_setup_timer(cpu);
>  	xen_init_lock_cpu(cpu);
> diff --git a/arch/x86/xen/xen-asm_32.S b/arch/x86/xen/xen-asm_32.S
> index 33ca6e4..fd92a64 100644
> --- a/arch/x86/xen/xen-asm_32.S
> +++ b/arch/x86/xen/xen-asm_32.S
> @@ -75,6 +75,17 @@ ENDPROC(xen_sysexit)
>   * stack state in whatever form its in, we keep things simple by only
>   * using a single register which is pushed/popped on the stack.
>   */
> +
> +.macro POP_FS
> +1:
> +	popw %fs
> +.pushsection .fixup, "ax"
> +2:	movw $0, (%esp)
> +	jmp 1b
> +.popsection
> +	_ASM_EXTABLE(1b,2b)
> +.endm
> +
>  ENTRY(xen_iret)
>  	/* test eflags for special cases */
>  	testl $(X86_EFLAGS_VM | XEN_EFLAGS_NMI), 8(%esp)
> @@ -83,15 +94,13 @@ ENTRY(xen_iret)
>  	push %eax
>  	ESP_OFFSET=4	# bytes pushed onto stack
>  
> -	/*
> -	 * Store vcpu_info pointer for easy access.  Do it this way to
> -	 * avoid having to reload %fs
> -	 */
> +	/* Store vcpu_info pointer for easy access */
>  #ifdef CONFIG_SMP
> -	GET_THREAD_INFO(%eax)
> -	movl %ss:TI_cpu(%eax), %eax
> -	movl %ss:__per_cpu_offset(,%eax,4), %eax
> -	mov %ss:xen_vcpu(%eax), %eax
> +	pushw %fs
> +	movl $(__KERNEL_PERCPU), %eax
> +	movl %eax, %fs
> +	movl %fs:xen_vcpu, %eax

How can this get the correct per-cpu xen_vcpu pointer if it doesn't ever
get the current cpu number?  Doesn't this always get VCPU#0's xen_vcpu?

David