[MIPS] IRQ Affinity Support for SMTC on Malta Platform
    
Signed-off-by: Kevin D. Kissell <kevink@mips.com>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
diff --git a/arch/mips/Kconfig b/arch/mips/Kconfig
index 3b807b4..b09eee2 100644
--- a/arch/mips/Kconfig
+++ b/arch/mips/Kconfig
@@ -1378,6 +1378,19 @@
 	  impact on interrupt service overhead. Disable it only if you know
 	  what you are doing.
 
+config MIPS_MT_SMTC_IRQAFF
+	bool "Support IRQ affinity API"
+	depends on MIPS_MT_SMTC
+	default n
+	help
+	  Enables SMP IRQ affinity API (/proc/irq/*/smp_affinity, etc.)
+	  for SMTC Linux kernel. Requires platform support, of which
+	  an example can be found in the MIPS kernel i8259 and Malta
+	  platform code.  It is recommended that MIPS_MT_SMTC_INSTANT_REPLAY
+	  be enabled if MIPS_MT_SMTC_IRQAFF is used. Adds overhead to
+	  interrupt dispatch, and should be used only if you know what
+	  you are doing.
+
 config MIPS_VPE_LOADER_TOM
 	bool "Load VPE program into memory hidden from linux"
 	depends on MIPS_VPE_LOADER
diff --git a/arch/mips/kernel/i8259.c b/arch/mips/kernel/i8259.c
index 3a2d255..4f4359b 100644
--- a/arch/mips/kernel/i8259.c
+++ b/arch/mips/kernel/i8259.c
@@ -39,6 +39,9 @@
 	.disable	= disable_8259A_irq,
 	.unmask		= enable_8259A_irq,
 	.mask_ack	= mask_and_ack_8259A,
+#ifdef CONFIG_MIPS_MT_SMTC_IRQAFF
+	.set_affinity	= plat_set_irq_affinity,
+#endif /* CONFIG_MIPS_MT_SMTC_IRQAFF */
 };
 
 /*
diff --git a/arch/mips/kernel/smtc.c b/arch/mips/kernel/smtc.c
index f094043..fe22387 100644
--- a/arch/mips/kernel/smtc.c
+++ b/arch/mips/kernel/smtc.c
@@ -606,6 +606,60 @@
 	return setup_irq(irq, new);
 }
 
+#ifdef CONFIG_MIPS_MT_SMTC_IRQAFF
+/*
+ * Support for IRQ affinity to TCs
+ */
+
+void smtc_set_irq_affinity(unsigned int irq, cpumask_t affinity)
+{
+	/*
+	 * If a "fast path" cache of quickly decodable affinity state
+	 * is maintained, this is where it gets done, on a call up
+	 * from the platform affinity code.
+	 */
+}
+
+void smtc_forward_irq(unsigned int irq)
+{
+	int target;
+
+	/*
+	 * OK wise guy, now figure out how to get the IRQ
+	 * to be serviced on an authorized "CPU".
+	 *
+	 * Ideally, to handle the situation where an IRQ has multiple
+	 * eligible CPUS, we would maintain state per IRQ that would
+	 * allow a fair distribution of service requests.  Since the
+	 * expected use model is any-or-only-one, for simplicity
+	 * and efficiency, we just pick the easiest one to find.
+	 */
+
+	target = first_cpu(irq_desc[irq].affinity);
+
+	/*
+	 * We depend on the platform code to have correctly processed
+	 * IRQ affinity change requests to ensure that the IRQ affinity
+	 * mask has been purged of bits corresponding to nonexistent and
+	 * offline "CPUs", and to TCs bound to VPEs other than the VPE
+	 * connected to the physical interrupt input for the interrupt
+	 * in question.  Otherwise we have a nasty problem with interrupt
+	 * mask management.  This is best handled in non-performance-critical
+	 * platform IRQ affinity setting code,  to minimize interrupt-time
+	 * checks.
+	 */
+
+	/* If no one is eligible, service locally */
+	if (target >= NR_CPUS) {
+		do_IRQ_no_affinity(irq);
+		return;
+	}
+
+	smtc_send_ipi(target, IRQ_AFFINITY_IPI, irq);
+}
+
+#endif /* CONFIG_MIPS_MT_SMTC_IRQAFF */
+
 /*
  * IPI model for SMTC is tricky, because interrupts aren't TC-specific.
  * Within a VPE one TC can interrupt another by different approaches.
@@ -830,6 +884,15 @@
 			break;
 		}
 		break;
+#ifdef CONFIG_MIPS_MT_SMTC_IRQAFF
+	case IRQ_AFFINITY_IPI:
+		/*
+		 * Accept a "forwarded" interrupt that was initially
+		 * taken by a TC who doesn't have affinity for the IRQ.
+		 */
+		do_IRQ_no_affinity((int)arg_copy);
+		break;
+#endif /* CONFIG_MIPS_MT_SMTC_IRQAFF */
 	default:
 		printk("Impossible SMTC IPI Type 0x%x\n", type_copy);
 		break;
diff --git a/arch/mips/mips-boards/malta/malta_smtc.c b/arch/mips/mips-boards/malta/malta_smtc.c
index ae05d05..5c980f4 100644
--- a/arch/mips/mips-boards/malta/malta_smtc.c
+++ b/arch/mips/mips-boards/malta/malta_smtc.c
@@ -88,3 +88,53 @@
 void prom_cpus_done(void)
 {
 }
+
+#ifdef CONFIG_MIPS_MT_SMTC_IRQAFF
+/*
+ * IRQ affinity hook
+ */
+
+
+void plat_set_irq_affinity(unsigned int irq, cpumask_t affinity)
+{
+	cpumask_t tmask = affinity;
+	int cpu = 0;
+	void smtc_set_irq_affinity(unsigned int irq, cpumask_t aff);
+
+	/*
+	 * On the legacy Malta development board, all I/O interrupts
+	 * are routed through the 8259 and combined in a single signal
+	 * to the CPU daughterboard, and on the CoreFPGA2/3 34K models,
+	 * that signal is brought to IP2 of both VPEs. To avoid racing
+	 * concurrent interrupt service events, IP2 is enabled only on
+	 * one VPE, by convention VPE0.  So long as no bits are ever
+	 * cleared in the affinity mask, there will never be any
+	 * interrupt forwarding.  But as soon as a program or operator
+	 * sets affinity for one of the related IRQs, we need to make
+	 * sure that we don't ever try to forward across the VPE boundry,
+	 * at least not until we engineer a system where the interrupt
+	 * _ack() or _end() function can somehow know that it corresponds
+	 * to an interrupt taken on another VPE, and perform the appropriate
+	 * restoration of Status.IM state using MFTR/MTTR instead of the
+	 * normal local behavior. We also ensure that no attempt will
+	 * be made to forward to an offline "CPU".
+	 */
+
+	for_each_cpu_mask(cpu, affinity) {
+		if ((cpu_data[cpu].vpe_id != 0) || !cpu_online(cpu))
+			cpu_clear(cpu, tmask);
+	}
+	irq_desc[irq].affinity = tmask;
+
+	if (cpus_empty(tmask))
+		/*
+		 * We could restore a default mask here, but the
+		 * runtime code can anyway deal with the null set
+		 */
+		printk(KERN_WARNING
+			"IRQ affinity leaves no legal CPU for IRQ %d\n", irq);
+
+	/* Do any generic SMTC IRQ affinity setup */
+	smtc_set_irq_affinity(irq, tmask);
+}
+#endif /* CONFIG_MIPS_MT_SMTC_IRQAFF */
diff --git a/include/asm-mips/irq.h b/include/asm-mips/irq.h
index 2cb52cf..a58f0ee 100644
--- a/include/asm-mips/irq.h
+++ b/include/asm-mips/irq.h
@@ -46,6 +46,38 @@
 
 #endif /* CONFIG_MIPS_MT_SMTC */
 
+#ifdef CONFIG_MIPS_MT_SMTC_IRQAFF
+#include <linux/cpumask.h>
+
+extern void plat_set_irq_affinity(unsigned int irq, cpumask_t affinity);
+extern void smtc_forward_irq(unsigned int irq);
+
+/*
+ * IRQ affinity hook invoked at the beginning of interrupt dispatch
+ * if option is enabled.
+ *
+ * Up through Linux 2.6.22 (at least) cpumask operations are very
+ * inefficient on MIPS.  Initial prototypes of SMTC IRQ affinity
+ * used a "fast path" per-IRQ-descriptor cache of affinity information
+ * to reduce latency.  As there is a project afoot to optimize the
+ * cpumask implementations, this version is optimistically assuming
+ * that cpumask.h macro overhead is reasonable during interrupt dispatch.
+ */
+#define IRQ_AFFINITY_HOOK(irq)						\
+do {									\
+    if (!cpu_isset(smp_processor_id(), irq_desc[irq].affinity)) {	\
+	smtc_forward_irq(irq);						\
+	irq_exit();							\
+	return;								\
+    }									\
+} while (0)
+
+#else /* Not doing SMTC affinity */
+
+#define IRQ_AFFINITY_HOOK(irq) do { } while (0)
+
+#endif /* CONFIG_MIPS_MT_SMTC_IRQAFF */
+
 #ifdef CONFIG_MIPS_MT_SMTC_IM_BACKSTOP
 
 /*
@@ -56,13 +88,27 @@
  */
 #define __DO_IRQ_SMTC_HOOK(irq)						\
 do {									\
+	IRQ_AFFINITY_HOOK(irq);						\
 	if (irq_hwmask[irq] & 0x0000ff00)				\
 		write_c0_tccontext(read_c0_tccontext() &		\
-		                   ~(irq_hwmask[irq] & 0x0000ff00));	\
+				   ~(irq_hwmask[irq] & 0x0000ff00));	\
 } while (0)
+
+#define __NO_AFFINITY_IRQ_SMTC_HOOK(irq)				\
+do {									\
+	if (irq_hwmask[irq] & 0x0000ff00)                               \
+		write_c0_tccontext(read_c0_tccontext() &		\
+				   ~(irq_hwmask[irq] & 0x0000ff00));	\
+} while (0)
+
 #else
 
-#define __DO_IRQ_SMTC_HOOK(irq) do { } while (0)
+#define __DO_IRQ_SMTC_HOOK(irq)						\
+do {									\
+	IRQ_AFFINITY_HOOK(irq);						\
+} while (0)
+#define __NO_AFFINITY_IRQ_SMTC_HOOK(irq) do { } while (0)
+
 #endif
 
 /*
@@ -81,6 +127,23 @@
 	irq_exit();							\
 } while (0)
 
+#ifdef CONFIG_MIPS_MT_SMTC_IRQAFF
+/*
+ * To avoid inefficient and in some cases pathological re-checking of
+ * IRQ affinity, we have this variant that skips the affinity check.
+ */
+
+
+#define do_IRQ_no_affinity(irq)						\
+do {									\
+	irq_enter();							\
+	__NO_AFFINITY_IRQ_SMTC_HOOK(irq);				\
+	generic_handle_irq(irq);					\
+	irq_exit();							\
+} while (0)
+
+#endif /* CONFIG_MIPS_MT_SMTC_IRQAFF */
+
 extern void arch_init_irq(void);
 extern void spurious_interrupt(void);
 
diff --git a/include/asm-mips/smtc_ipi.h b/include/asm-mips/smtc_ipi.h
index a52a4a7..e09131a 100644
--- a/include/asm-mips/smtc_ipi.h
+++ b/include/asm-mips/smtc_ipi.h
@@ -34,6 +34,7 @@
 
 #define LINUX_SMP_IPI 1
 #define SMTC_CLOCK_TICK 2
+#define IRQ_AFFINITY_IPI 3
 
 /*
  * A queue of IPI messages