Blackfin: make deferred hardware errors more exact

Hardware errors on the Blackfin architecture are queued by nature of the
hardware design.  Things that could generate a hardware level queue up at
the system interface and might not process until much later, at which
point the system would send a notification back to the core.

As such, it is possible for user space code to do something that would
trigger a hardware error, but have it delay long enough for the process
context to switch.  So when the hardware error does signal, we mistakenly
evaluate it as a different process or as kernel context and panic (erp!).
This makes it pretty difficult to find the offending context.  But wait,
there is good news somewhere.

By forcing a SSYNC in the interrupt entry, we force all pending queues at
the system level to be processed and all hardware errors to be signaled.
Then we check the current interrupt state to see if the hardware error is
now signaled.  If so, we re-queue the current interrupt and return thus
allowing the higher priority hardware error interrupt to process properly.
Since we haven't done any other context processing yet, the right context
will be selected and killed.  There is still the possibility that the
exact offending instruction will be unknown, but at least we'll have a
much better idea of where to look.

The downside of course is that this causes system-wide syncs at every
interrupt point which results in significant performance degradation.
Since this situation should not occur in any properly configured system
(as hardware errors are triggered by things like bad pointers), make it a
debug configuration option and disable it by default.

Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>

2 files changed, 68 insertions, 18 deletions

diff --git a/arch/blackfin/mach-common/entry.S b/arch/blackfin/mach-common/entry.S
index f0636fdcb353..da0558ad1b1a 100644
--- a/arch/blackfin/mach-common/entry.S
+++ b/arch/blackfin/mach-common/entry.S
@@ -200,7 +200,18 @@ ENTRY(_ex_single_step)
 	cc = r7 == 0;
 	if !cc jump 1f;
 #endif
-
+#ifdef CONFIG_EXACT_HWERR
+	/* Read the ILAT, and to check to see if the process we are
+	 * single stepping caused a previous hardware error
+	 * If so, do not single step, (which lowers to IRQ5, and makes
+	 * us miss the error).
+	 */
+	p5.l = lo(ILAT);
+	p5.h = hi(ILAT);
+	r7 = [p5];
+	cc = bittst(r7, EVT_IVHW_P);
+	if cc jump 1f;
+#endif
 	/* Single stepping only a single instruction, so clear the trace
 	 * bit here.  */
 	r7 = syscfg;
@@ -262,15 +273,6 @@ ENTRY(_bfin_return_from_exception)
 	r6 = 0x25;
 	CC = R7 == R6;
 	if CC JUMP _double_fault;
-
-	/* Did we cause a HW error? */
-	p5.l = lo(ILAT);
-	p5.h = hi(ILAT);
-	r6 = [p5];
-	r7 = 0x20;		/* Did I just cause anther HW error? */
-	r6 = r7 & r6;
-	CC = R7 == R6;
-	if CC JUMP _double_fault;
 #endif
 
 	(R7:6,P5:4) = [sp++];
@@ -472,6 +474,16 @@ ENTRY(_trap) /* Exception: 4th entry into system event table(supervisor mode)*/
 	[--sp] = ASTAT;
 	[--sp] = (R7:6,P5:4);
 
+#ifdef CONFIG_EXACT_HWERR
+	/* Make sure all pending read/writes complete. This will ensure any
+	 * accesses which could cause hardware errors completes, and signal
+	 * the the hardware before we do something silly, like crash the
+	 * kernel. We don't need to work around anomaly 05000312, since
+	 * we are already atomic
+	 */
+	ssync;
+#endif
+
 #if ANOMALY_05000283 || ANOMALY_05000315
 	cc = r7 == r7;
 	p5.h = HI(CHIPID);
@@ -854,7 +866,7 @@ ENTRY(_ret_from_exception)
 	p1.h = _schedule_and_signal;
 	[p0] = p1;
 	csync;
-	raise 15;		/* raise evt14 to do signal or reschedule */
+	raise 15;		/* raise evt15 to do signal or reschedule */
 4:
 	r0 = syscfg;
 	bitclr(r0, 0);
@@ -915,7 +927,7 @@ ENTRY(_return_from_int)
 	p1.h = _schedule_and_signal_from_int;
 	[p0] = p1;
 	csync;
-#if ANOMALY_05000281
+#if ANOMALY_05000281 || ANOMALY_05000461
 	r0.l = lo(SAFE_USER_INSTRUCTION);
 	r0.h = hi(SAFE_USER_INSTRUCTION);
 	reti = r0;
@@ -929,18 +941,27 @@ ENTRY(_return_from_int)
 ENDPROC(_return_from_int)
 
 ENTRY(_lower_to_irq14)
-#if ANOMALY_05000281
+#if ANOMALY_05000281 || ANOMALY_05000461
 	r0.l = lo(SAFE_USER_INSTRUCTION);
 	r0.h = hi(SAFE_USER_INSTRUCTION);
 	reti = r0;
 #endif
-	r0 = 0x401f;
+
+#ifdef CONFIG_DEBUG_HWERR
+	/* enable irq14 & hwerr interrupt, until we transition to _evt14_softirq */
+	r0 = (EVT_IVG14 | EVT_IVHW | EVT_IRPTEN | EVT_EVX | EVT_NMI | EVT_RST | EVT_EMU);
+#else
+	/* Only enable irq14 interrupt, until we transition to _evt14_softirq */
+	r0 = (EVT_IVG14 | EVT_IRPTEN | EVT_EVX | EVT_NMI | EVT_RST | EVT_EMU);
+#endif
 	sti r0;
 	raise 14;
 	rti;
+ENDPROC(_lower_to_irq14)
+
 ENTRY(_evt14_softirq)
 #ifdef CONFIG_DEBUG_HWERR
-	r0 = 0x3f;
+	r0 = (EVT_IVHW | EVT_IRPTEN | EVT_EVX | EVT_NMI | EVT_RST | EVT_EMU);
 	sti r0;
 #else
 	cli r0;
@@ -948,8 +969,9 @@ ENTRY(_evt14_softirq)
 	[--sp] = RETI;
 	SP += 4;
 	rts;
+ENDPROC(_evt14_softirq)
 
-_schedule_and_signal_from_int:
+ENTRY(_schedule_and_signal_from_int)
 	/* To end up here, vector 15 was changed - so we have to change it
 	 * back.
 	 */
@@ -982,8 +1004,9 @@ _schedule_and_signal_from_int:
 	call _finish_atomic_sections;
 	sp += 12;
 	jump.s .Lresume_userspace;
+ENDPROC(_schedule_and_signal_from_int)
 
-_schedule_and_signal:
+ENTRY(_schedule_and_signal)
 	SAVE_CONTEXT_SYSCALL
 	/* To end up here, vector 15 was changed - so we have to change it
 	 * back.
@@ -1001,7 +1024,7 @@ _schedule_and_signal:
 1:
 	RESTORE_CONTEXT
 	rti;
-ENDPROC(_lower_to_irq14)
+ENDPROC(_schedule_and_signal)
 
 /* We handle this 100% in exception space - to reduce overhead
  * Only potiential problem is if the software buffer gets swapped out of the
diff --git a/arch/blackfin/mach-common/interrupt.S b/arch/blackfin/mach-common/interrupt.S
index 0069c2dd4625..9c46680186e4 100644
--- a/arch/blackfin/mach-common/interrupt.S
+++ b/arch/blackfin/mach-common/interrupt.S
@@ -145,6 +145,14 @@ __common_int_entry:
 
 /* interrupt routine for ivhw - 5 */
 ENTRY(_evt_ivhw)
+	/* In case a single action kicks off multiple memory transactions, (like
+	 * a cache line fetch, - this can cause multiple hardware errors, let's
+	 * catch them all. First - make sure all the actions are complete, and
+	 * the core sees the hardware errors.
+	 */
+	SSYNC;
+	SSYNC;
+
 	SAVE_ALL_SYS
 #ifdef CONFIG_FRAME_POINTER
 	fp = 0;
@@ -159,6 +167,25 @@ ENTRY(_evt_ivhw)
 1:
 #endif
 
+	/* Handle all stacked hardware errors
+	 * To make sure we don't hang forever, only do it 10 times
+	 */
+	R0 = 0;
+	R2 = 10;
+1:
+	P0.L = LO(ILAT);
+	P0.H = HI(ILAT);
+	R1 = [P0];
+	CC = BITTST(R1, EVT_IVHW_P);
+	IF ! CC JUMP 2f;
+	/* OK a hardware error is pending - clear it */
+	R1 = EVT_IVHW_P;
+	[P0] = R1;
+	R0 += 1;
+	CC = R1 == R2;
+	if CC JUMP 2f;
+	JUMP 1b;
+2:
 	# We are going to dump something out, so make sure we print IPEND properly
 	p2.l = lo(IPEND);
 	p2.h = hi(IPEND);