arm64: fix missing syscall trace exit
authorJosh Stone <jistone@redhat.com>
Fri, 5 Jun 2015 21:28:03 +0000 (14:28 -0700)
committerCatalin Marinas <catalin.marinas@arm.com>
Mon, 8 Jun 2015 17:34:21 +0000 (18:34 +0100)
If a syscall is entered without TIF_SYSCALL_TRACE set, then it goes on
the fast path.  It's then possible to have TIF_SYSCALL_TRACE added in
the middle of the syscall, but ret_fast_syscall doesn't check this flag
again.  This causes a ptrace syscall-exit-stop to be missed.

For instance, from a PTRACE_EVENT_FORK reported during do_fork, the
tracer might resume with PTRACE_SYSCALL, setting TIF_SYSCALL_TRACE.
Now the completion of the fork should have a syscall-exit-stop.

Russell King fixed this on arm by re-checking _TIF_SYSCALL_WORK in the
fast exit path.  Do the same on arm64.

Reviewed-by: Will Deacon <will.deacon@arm.com>
Cc: Russell King <rmk+kernel@arm.linux.org.uk>
Signed-off-by: Josh Stone <jistone@redhat.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
arch/arm64/kernel/entry.S

index d369ccff67e2f4acd1afd1fd70c8788aa98c51ab..0eeb1b95c65a57739f548e2d43ed372ca59c5427 100644 (file)
@@ -611,11 +611,16 @@ ENDPROC(cpu_switch_to)
  */
 ret_fast_syscall:
        disable_irq                             // disable interrupts
-       ldr     x1, [tsk, #TI_FLAGS]
+       ldr     x1, [tsk, #TI_FLAGS]            // re-check for syscall tracing
+       and     x2, x1, #_TIF_SYSCALL_WORK
+       cbnz    x2, ret_fast_syscall_trace
        and     x2, x1, #_TIF_WORK_MASK
        cbnz    x2, fast_work_pending
        enable_step_tsk x1, x2
        kernel_exit 0, ret = 1
+ret_fast_syscall_trace:
+       enable_irq                              // enable interrupts
+       b       __sys_trace_return
 
 /*
  * Ok, we need to do extra processing, enter the slow path.