Add back improved affinity_hint handling

The new balancer can now deal with affintiy hinting again, this time in a
reasonably sane manner.  Whereas before having an affintiy hint caused
irqbalance to just assign that hint as the affinity, we now have a policy based
operation, controlled by the hintpolicy option.  The policy can be one of:

exact - affinity_hint is applied for that irq without balancing consideration
subset - balancing takes place, but assigned affinity will be a subset of the
	 hint
ignore - affinity_hint is ignored entirely

activate.c

@@ -36,6 +36,7 @@ static void activate_mapping(struct irq_info *info, void *data __attribute__((un
 {
 	char buf[PATH_MAX];
 	FILE *file;
+	cpumask_t applied_mask;
 
 	/*
  	 * only activate mappings for irqs that have moved
@@ -52,7 +53,13 @@ static void activate_mapping(struct irq_info *info, void *data __attribute__((un
 	if (!file)
 		return;
 
-	cpumask_scnprintf(buf, PATH_MAX, info->assigned_obj->mask);
+	if ((hint_policy == HINT_POLICY_EXACT) &&
+	    (!cpus_empty(info->affinity_hint)))
+		applied_mask = info->affinity_hint;
+	else
+		applied_mask = info->assigned_obj->mask;
+
+	cpumask_scnprintf(buf, PATH_MAX, applied_mask);
 	fprintf(file, "%s", buf);
 	fclose(file);
 	info->moved = 0; /*migration is done*/

classify.c

@@ -25,8 +25,6 @@ int map_class_to_level[7] =
 { BALANCE_PACKAGE, BALANCE_CACHE, BALANCE_CACHE, BALANCE_NONE, BALANCE_CORE, BALANCE_CORE, BALANCE_CORE };
 
 
-int class_counts[7];
-
 #define MAX_CLASS 0x12
 /*
  * Class codes lifted from pci spec, appendix D.

irqbalance.1

@@ -41,6 +41,19 @@ Causes irqbalance to be run once, after which the daemon exits
 .B --debug
 Causes irqbalance to run in the foreground and extra debug information to be printed
 
+.TP
+.B --hintpolicy=[exact | subset | ignore]
+Set the policy for how irq kernel affinity hinting is treated.  Can be one of:
+.P
+.I exact
+irq affinity hint is applied unilaterally and never violated
+.P
+.I subset
+irq is balanced, but the assigned object will be a subset of the affintiy hint
+.P
+.I ignore
+irq affinity hint value is completely ignored
+
 .SH "ENVIRONMENT VARIABLES"
 .TP
 .B IRQBALANCE_ONESHOT

irqbalance.c

@@ -38,12 +38,10 @@
 int one_shot_mode;
 int debug_mode;
 int numa_avail;
-
 int need_cpu_rescan;
-
 extern cpumask_t banned_cpus;
-
 static int counter;
+enum hp_e hint_policy = HINT_POLICY_SUBSET;
 
 
 void sleep_approx(int seconds)
@@ -64,12 +62,13 @@ void sleep_approx(int seconds)
 struct option lopts[] = {
 	{"oneshot", 0, NULL, 'o'},
 	{"debug", 0, NULL, 'd'},
+	{"hintpolicy", 1, NULL, 'h'},
 	{0, 0, 0, 0}
 };
 
 static void usage(void)
 {
-	printf("irqbalance [--oneshot | -o] [--debug | -d]");
+	printf("irqbalance [--oneshot | -o] [--debug | -d] [--hintpolicy= | -h [exact|subset|ignore]]");
 }
 
 static void parse_command_line(int argc, char **argv)
@@ -78,7 +77,7 @@ static void parse_command_line(int argc, char **argv)
 	int longind;
 
 	while ((opt = getopt_long(argc, argv,
-		"",
+		"odh:",
 		lopts, &longind)) != -1) {
 
 		switch(opt) {
@@ -88,6 +87,18 @@ static void parse_command_line(int argc, char **argv)
 			case 'd':
 				debug_mode=1;
 				break;
+			case 'h':
+				if (!strncmp(optarg, "exact", strlen(optarg)))
+					hint_policy = HINT_POLICY_EXACT;
+				else if (!strncmp(optarg, "subset", strlen(optarg)))
+					hint_policy = HINT_POLICY_SUBSET;
+				else if (!strncmp(optarg, "ignore", strlen(optarg)))
+					hint_policy = HINT_POLICY_IGNORE;
+				else {
+					usage();
+					exit(1);
+				}
+				break;
 			case 'o':
 				one_shot_mode=1;
 				break;

irqbalance.h

@@ -16,12 +16,6 @@ extern int package_count;
 extern int cache_domain_count;
 extern int core_count;
 extern char *classes[];
-extern int map_class_to_level[7];
-extern int class_counts[7];
-extern int debug_mode;
-extern int power_mode;
-extern int need_cpu_rescan;
-extern int one_shot_mode;
 
 extern void parse_cpu_tree(void);
 extern void clear_work_stats(void);
@@ -46,6 +40,18 @@ void clear_cpu_tree(void);
 void pci_numa_scan(void);
 
 /*===================NEW BALANCER FUNCTIONS============================*/
+enum hp_e {
+	HINT_POLICY_IGNORE,
+	HINT_POLICY_SUBSET,
+	HINT_POLICY_EXACT
+};
+
+extern int debug_mode;
+extern int one_shot_mode;
+extern int power_mode;
+extern int need_cpu_rescan;
+extern enum hp_e hint_policy;
+
 /*
  * Numa node access routines
  */

irqlist.c

@@ -68,6 +68,13 @@ static void move_candidate_irqs(struct irq_info *info, void *data)
 {
 	int *remaining_deviation = (int *)data;
 
+	/* never move an irq that has an afinity hint when 
+ 	 * hint_policy is HINT_POLICY_EXACT 
+ 	 */
+	if (hint_policy == HINT_POLICY_EXACT)
+		if (!cpus_empty(info->affinity_hint))
+			return;
+
 	/* Don't rebalance irqs that don't want it */
 	if (info->level == BALANCE_NONE)
 		return;

placement.c

@@ -43,6 +43,20 @@ static void find_best_object(struct common_obj_data *d, void *data)
 {
 	struct obj_placement *best = (struct obj_placement *)data;
 	uint64_t newload;
+	cpumask_t subset;
+
+	/*
+ 	 * If the hint policy is subset, then we only want 
+ 	 * to consider objects that are within the irqs hint, but
+ 	 * only if that irq in fact has published a hint
+ 	 */
+	if (hint_policy == HINT_POLICY_SUBSET) {
+		if (!cpus_empty(best->info->affinity_hint)) {
+			cpus_and(subset, best->info->affinity_hint, d->mask);
+			if (cpus_empty(subset))
+				return;
+		}
+	}
 
 	newload = d->load;
 	if (newload < best->best_cost) {

types.h

@@ -26,7 +26,6 @@
 #define IRQ_TYPE_MSI	1
 #define IRQ_TYPE_MSIX	2
 
-
 struct common_obj_data {
 	uint64_t load;
 	int number;