[PATCH] ARM: enable_irq before ret_fast_syscall tracing

[Josh Stone]

On 06/22/2015 05:08 PM, Josh Stone wrote:
> When reached via the slow path __sys_trace, __sys_trace_return and its
> callees usually have interrupts still enabled.  This is important if any
> will schedule, like for a ptrace syscall-exit-stop.
> 
> In the rarer case where tracing was not enabled on syscall entry, and
> then ret_fast_syscall sees tracing was enabled mid-syscall, then it
> also ought to branch to __sys_trace_return with interrupts enabled.

Side note -- I haven't actually found any hard evidence that disabled
interrupts here are a problem, even though it seems obviously bad.

I used kprobes to confirm that I am indeed reaching this case, and
modified my original testcase to spam this scenario in a loop.  But I
never encountered any instability or debug messages about interrupts.

Still, it also runs cleanly with this patch, and I think this is more
correct.  Please correct me if I'm wrong!

> Cc: Russell King <rmk+kernel at arm.linux.org.uk>
> Signed-off-by: Josh Stone <jistone at redhat.com>
> ---
>  arch/arm/kernel/entry-common.S | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/arm/kernel/entry-common.S b/arch/arm/kernel/entry-common.S
> index 4e7f40c577e6..5d8eb11b8571 100644
> --- a/arch/arm/kernel/entry-common.S
> +++ b/arch/arm/kernel/entry-common.S
> @@ -35,7 +35,7 @@ ret_fast_syscall:
>  	disable_irq				@ disable interrupts
>  	ldr	r1, [tsk, #TI_FLAGS]		@ re-check for syscall tracing
>  	tst	r1, #_TIF_SYSCALL_WORK
> -	bne	__sys_trace_return
> +	bne	ret_fast_syscall_trace
>  	tst	r1, #_TIF_WORK_MASK
>  	bne	fast_work_pending
>  	asm_trace_hardirqs_on
> @@ -45,6 +45,10 @@ ret_fast_syscall:
>  	ct_user_enter
>  
>  	restore_user_regs fast = 1, offset = S_OFF
> +
> +ret_fast_syscall_trace:
> +	enable_irq				@ enable interrupts
> +	b	__sys_trace_return
>   UNWIND(.fnend		)
>  
>  /*
>