powerpc: Fix stack overflow crash in resume_kernel when ftracing
authorMichael Ellerman <michael@ellerman.id.au>
Thu, 13 Jun 2013 11:04:56 +0000 (21:04 +1000)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Sat, 15 Jun 2013 02:21:57 +0000 (12:21 +1000)
It's possible for us to crash when running with ftrace enabled, eg:

  Bad kernel stack pointer bffffd12 at c00000000000a454
  cpu 0x3: Vector: 300 (Data Access) at [c00000000ffe3d40]
      pc: c00000000000a454: resume_kernel+0x34/0x60
      lr: c00000000000335c: performance_monitor_common+0x15c/0x180
      sp: bffffd12
     msr: 8000000000001032
     dar: bffffd12
   dsisr: 42000000

If we look at current's stack (paca->__current->stack) we see it is
equal to c0000002ecab0000. Our stack is 16K, and comparing to
paca->kstack (c0000002ecab3e30) we can see that we have overflowed our
kernel stack. This leads to us writing over our struct thread_info, and
in this case we have corrupted thread_info->flags and set
_TIF_EMULATE_STACK_STORE.

Dumping the stack we see:

  3:mon> t c0000002ecab0000
  [c0000002ecab0000c00000000002131c .performance_monitor_exception+0x5c/0x70
  [c0000002ecab0080c00000000000335c performance_monitor_common+0x15c/0x180
  --- Exception: f01 (Performance Monitor) at c0000000000fb2ec .trace_hardirqs_off+0x1c/0x30
  [c0000002ecab0370c00000000016fdb0 .trace_graph_entry+0xb0/0x280 (unreliable)
  [c0000002ecab0410c00000000003d038 .prepare_ftrace_return+0x98/0x130
  [c0000002ecab04b0c00000000000a920 .ftrace_graph_caller+0x14/0x28
  [c0000002ecab0520c0000000000d6b58 .idle_cpu+0x18/0x90
  [c0000002ecab05a0c00000000000a934 .return_to_handler+0x0/0x34
  [c0000002ecab0620c00000000001e660 .timer_interrupt+0x160/0x300
  [c0000002ecab06d0c0000000000025dc decrementer_common+0x15c/0x180
  --- Exception: 901 (Decrementer) at c0000000000104d4 .arch_local_irq_restore+0x74/0xa0
  [c0000002ecab09c0c0000000000fe044 .trace_hardirqs_on+0x14/0x30 (unreliable)
  [c0000002ecab0fb0c00000000016fe3c .trace_graph_entry+0x13c/0x280
  [c0000002ecab1050c00000000003d038 .prepare_ftrace_return+0x98/0x130
  [c0000002ecab10f0c00000000000a920 .ftrace_graph_caller+0x14/0x28
  [c0000002ecab1160c0000000000161f0 .__ppc64_runlatch_on+0x10/0x40
  [c0000002ecab11d0c00000000000a934 .return_to_handler+0x0/0x34
  --- Exception: 901 (Decrementer) at c0000000000104d4 .arch_local_irq_restore+0x74/0xa0

  ... and so on

__ppc64_runlatch_on() is called from RUNLATCH_ON in the exception entry
path. At that point the irq state is not consistent, ie. interrupts are
hard disabled (by the exception entry), but the paca soft-enabled flag
may be out of sync.

This leads to the local_irq_restore() in trace_graph_entry() actually
enabling interrupts, which we do not want. Because we have not yet
reprogrammed the decrementer we immediately take another decrementer
exception, and recurse.

The fix is twofold. Firstly make sure we call DISABLE_INTS before
calling RUNLATCH_ON. The badly named DISABLE_INTS actually reconciles
the irq state in the paca with the hardware, making it safe again to
call local_irq_save/restore().

Although that should be sufficient to fix the bug, we also mark the
runlatch routines as notrace. They are called very early in the
exception entry and we are asking for trouble tracing them. They are
also fairly uninteresting and tracing them just adds unnecessary
overhead.

[ This regression was introduced by fe1952fc0afb9a2e4c79f103c08aef5d13db1873
  "powerpc: Rework runlatch code" by myself --BenH
]

CC: <stable@vger.kernel.org> [v3.4+]
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/include/asm/exception-64s.h
arch/powerpc/kernel/process.c

index 8e5fae8beaf6ad0958e8188d8094ded008033425..46793b58a761d549d7bf69814530c128352f26a5 100644 (file)
@@ -513,7 +513,7 @@ label##_common:                                                     \
  */
 #define STD_EXCEPTION_COMMON_ASYNC(trap, label, hdlr)            \
        EXCEPTION_COMMON(trap, label, hdlr, ret_from_except_lite, \
-                        FINISH_NAP;RUNLATCH_ON;DISABLE_INTS)
+                        FINISH_NAP;DISABLE_INTS;RUNLATCH_ON)
 
 /*
  * When the idle code in power4_idle puts the CPU into NAP mode,
index b0f3e3f77e72830f465bfd00739fd5e33df22870..076d1242507a7fdcf32c374b66b3b005eb4e0db9 100644 (file)
@@ -1369,7 +1369,7 @@ void show_stack(struct task_struct *tsk, unsigned long *stack)
 
 #ifdef CONFIG_PPC64
 /* Called with hard IRQs off */
-void __ppc64_runlatch_on(void)
+void notrace __ppc64_runlatch_on(void)
 {
        struct thread_info *ti = current_thread_info();
        unsigned long ctrl;
@@ -1382,7 +1382,7 @@ void __ppc64_runlatch_on(void)
 }
 
 /* Called with hard IRQs off */
-void __ppc64_runlatch_off(void)
+void notrace __ppc64_runlatch_off(void)
 {
        struct thread_info *ti = current_thread_info();
        unsigned long ctrl;