powerpc/trace: Add a dummy stack frame for trace_hardirqs_off
authorKevin Hao <kexin.hao@windriver.com>
Thu, 10 Nov 2011 16:04:17 +0000 (16:04 +0000)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Wed, 16 Nov 2011 03:47:54 +0000 (14:47 +1100)
The trace_hardirqs_off will use CALLER_ADDR0 and CALLER_ADDR1.
If an exception occurs in user mode, there is only one stack frame
on the stack and accessing the CALLER_ADDR1 will causes the following
call trace. So we create a dummy stack frame to make
trace_hardirqs_off happy.

WARNING: at kernel/smp.c:459
Modules linked in:
NIP: c0093280 LR: c00930a0 CTR: c0010780
REGS: edb87ae0 TRAP: 0700   Not tainted  (3.1.0)
MSR: 00021002 <ME,CE>  CR: 28002888  XER: 00000000
TASK = edce2ac0[17658] 'mthread-lock-on' THREAD: edb86000 CPU: 5
GPR00: 00000001 edb87b90 edce2ac0 00000005 c0019594 edb87bd8 00000001 00000fe3
GPR08: 00041000 c084138c 4e20120d edb87b90 48002888 1001aa7c 00000000 00000000
GPR16: 48830000 10012a8c 00000000 10000af4 00000001 c0810000 00000000 00000000
GPR24: ee9aa920 c0816a18 00000000 00000005 c0019594 edb87bd8 ee20178c edb87b90
NIP [c0093280] smp_call_function_many+0x214/0x2b4
LR [c00930a0] smp_call_function_many+0x34/0x2b4
Call Trace:
[edb87b90] [c00930a0] smp_call_function_many+0x34/0x2b4 (unreliable)
[edb87bd0] [c00194ec] __flush_tlb_page+0xac/0x100
[edb87c00] [c001957c] flush_tlb_page+0x3c/0x54
[edb87c10] [c00180ac] ptep_set_access_flags+0x74/0x12c
[edb87c40] [c0128068] handle_pte_fault+0x2f0/0x9ac
[edb87cb0] [c0128c3c] handle_mm_fault+0x104/0x1dc
[edb87ce0] [c05f40f4] do_page_fault+0x2dc/0x630
[edb87e50] [c001078c] handle_page_fault+0xc/0x80

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/kernel/entry_32.S

index 56212bc0ab087c4b40bb21df3496c89849a01f06..4f80cf1ce77b84c30e905bcb8cb8d660c005c22b 100644 (file)
@@ -215,7 +215,22 @@ reenable_mmu:                              /* re-enable mmu so we can */
        stw     r9,8(r1)
        stw     r11,12(r1)
        stw     r3,ORIG_GPR3(r1)
+       /*
+        * The trace_hardirqs_off will use CALLER_ADDR0 and CALLER_ADDR1.
+        * If from user mode there is only one stack frame on the stack, and
+        * accessing CALLER_ADDR1 will cause oops. So we need create a dummy
+        * stack frame to make trace_hardirqs_off happy.
+        */
+       andi.   r12,r12,MSR_PR
+       beq     11f
+       stwu    r1,-16(r1)
+       bl      trace_hardirqs_off
+       addi    r1,r1,16
+       b       12f
+
+11:
        bl      trace_hardirqs_off
+12:
        lwz     r0,GPR0(r1)
        lwz     r3,ORIG_GPR3(r1)
        lwz     r4,GPR4(r1)