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>

1 files changed, 89 insertions, 3 deletions

diff --git a/arch/blackfin/include/asm/entry.h b/arch/blackfin/include/asm/entry.h
index b30a296..ec58efc 100644
--- a/arch/blackfin/include/asm/entry.h
+++ b/arch/blackfin/include/asm/entry.h
@@ -35,21 +35,39 @@
 #else
 # define LOAD_IPIPE_IPEND
 #endif
+
+#ifndef CONFIG_EXACT_HWERR
+/* As a debugging aid - we save IPEND when DEBUG_KERNEL is on,
+ * otherwise it is a waste of cycles.
+ */
+# ifndef CONFIG_DEBUG_KERNEL
+#define INTERRUPT_ENTRY(N)						\
+    [--sp] = SYSCFG;							\
+    [--sp] = P0;	/*orig_p0*/					\
+    [--sp] = R0;	/*orig_r0*/					\
+    [--sp] = (R7:0,P5:0);						\
+    R0 = (N);								\
+    LOAD_IPIPE_IPEND							\
+    jump __common_int_entry;
+# else /* CONFIG_DEBUG_KERNEL */
 #define INTERRUPT_ENTRY(N)						\
     [--sp] = SYSCFG;							\
-									\
     [--sp] = P0;	/*orig_p0*/					\
     [--sp] = R0;	/*orig_r0*/					\
     [--sp] = (R7:0,P5:0);						\
+    p0.l = lo(IPEND);							\
+    p0.h = hi(IPEND);							\
+    r1 = [p0];								\
     R0 = (N);								\
     LOAD_IPIPE_IPEND							\
     jump __common_int_entry;
+# endif /* CONFIG_DEBUG_KERNEL */
 
 /* For timer interrupts, we need to save IPEND, since the user_mode
-	   macro accesses it to determine where to account time.  */
+ *macro accesses it to determine where to account time.
+ */
 #define TIMER_INTERRUPT_ENTRY(N)					\
     [--sp] = SYSCFG;							\
-									\
     [--sp] = P0;	/*orig_p0*/					\
     [--sp] = R0;	/*orig_r0*/					\
     [--sp] = (R7:0,P5:0);						\
@@ -58,6 +76,74 @@
     r1 = [p0];								\
     R0 = (N);								\
     jump __common_int_entry;
+#else /* CONFIG_EXACT_HWERR is defined */
+
+/* if we want hardware error to be exact, we need to do a SSYNC (which forces
+ * read/writes to complete to the memory controllers), and check to see that
+ * caused a pending HW error condition. If so, we assume it was caused by user
+ * space, by setting the same interrupt that we are in (so it goes off again)
+ * and context restore, and a RTI (without servicing anything). This should
+ * cause the pending HWERR to fire, and when that is done, this interrupt will
+ * be re-serviced properly.
+ * As you can see by the code - we actually need to do two SSYNCS - one to
+ * make sure the read/writes complete, and another to make sure the hardware
+ * error is recognized by the core.
+ */
+#define INTERRUPT_ENTRY(N)						\
+    SSYNC;								\
+    SSYNC;								\
+    [--sp] = SYSCFG;							\
+    [--sp] = P0;	/*orig_p0*/					\
+    [--sp] = R0;	/*orig_r0*/					\
+    [--sp] = (R7:0,P5:0);						\
+    R1 = ASTAT;								\
+    P0.L = LO(ILAT);							\
+    P0.H = HI(ILAT);							\
+    R0 = [P0];								\
+    CC = BITTST(R0, EVT_IVHW_P);					\
+    IF CC JUMP 1f;							\
+    ASTAT = R1;								\
+    p0.l = lo(IPEND);							\
+    p0.h = hi(IPEND);							\
+    r1 = [p0];								\
+    R0 = (N);								\
+    LOAD_IPIPE_IPEND							\
+    jump __common_int_entry;						\
+1:  ASTAT = R1;								\
+    RAISE N;								\
+    (R7:0, P5:0) = [SP++];						\
+    SP += 0x8;								\
+    SYSCFG = [SP++];							\
+    CSYNC;								\
+    RTI;
+
+#define TIMER_INTERRUPT_ENTRY(N)					\
+    SSYNC;								\
+    SSYNC;								\
+    [--sp] = SYSCFG;							\
+    [--sp] = P0;	/*orig_p0*/					\
+    [--sp] = R0;	/*orig_r0*/					\
+    [--sp] = (R7:0,P5:0);						\
+    R1 = ASTAT;								\
+    P0.L = LO(ILAT);							\
+    P0.H = HI(ILAT);							\
+    R0 = [P0];								\
+    CC = BITTST(R0, EVT_IVHW_P);					\
+    IF CC JUMP 1f;							\
+    ASTAT = R1;								\
+    p0.l = lo(IPEND);							\
+    p0.h = hi(IPEND);							\
+    r1 = [p0];								\
+    R0 = (N);								\
+    jump __common_int_entry;						\
+1:  ASTAT = R1;								\
+    RAISE N;								\
+    (R7:0, P5:0) = [SP++];						\
+    SP += 0x8;								\
+    SYSCFG = [SP++];							\
+    CSYNC;								\
+    RTI;
+#endif	/* CONFIG_EXACT_HWERR */
 
 /* This one pushes RETI without using CLI.  Interrupts are enabled.  */
 #define SAVE_CONTEXT_SYSCALL	save_context_syscall