[Blackfin] arch: fix bug - when using trace buffer with CONFIG_MPU enabled.

There were a couple of problems with the way the trace buffer state
is saved/restored in assembly.  The DEBUG_HWTRACE_SAVE/RESTORE macros
save a value to the stack, which is not immediately obvious; the CPLB
exception code needed changes to load the correct value of the stack
pointer.  The other problem is that the SAVE/RESTORE macros weren't
pushing and popping the value downwards on the stack, but rather moving
it _upwards_, which is of course completely broken.

We also need to make sure there's a matching DEBUG_HWTRACE_RESTORE in
the error case of the CPLB handler.

Signed-off-by: Bernd Schmidt <bernds_cb1@t-online.de>
Signed-off-by: Bryan Wu <cooloney@kernel.org>

diff --git a/arch/blackfin/mach-common/entry.S b/arch/blackfin/mach-common/entry.S
index cee54ce..a504c65 100644
--- a/arch/blackfin/mach-common/entry.S
+++ b/arch/blackfin/mach-common/entry.S
@@ -121,10 +121,14 @@
 	(R7:6,P5:4) = [sp++];
 	ASTAT = [sp++];
 	SAVE_ALL_SYS
-	DEBUG_HWTRACE_SAVE(p5, r7)
 #ifdef CONFIG_MPU
+	/* We must load R1 here, _before_ DEBUG_HWTRACE_SAVE, since that
+	 * will change the stack pointer.  */
 	R0 = SEQSTAT;
 	R1 = SP;
+#endif
+	DEBUG_HWTRACE_SAVE(p5, r7)
+#ifdef CONFIG_MPU
 	sp += -12;
 	call _cplb_hdr;
 	sp += 12;
@@ -191,6 +195,7 @@
 ENDPROC(_bfin_return_from_exception)
 
 ENTRY(_handle_bad_cplb)
+	DEBUG_HWTRACE_RESTORE(p5, r7)
 	/* To get here, we just tried and failed to change a CPLB
 	 * so, handle things in trap_c (C code), by lowering to
 	 * IRQ5, just like we normally do. Since this is not a