forked from ~ljy/RK356X_SDK_RELEASE

blame | history | patch | commit | commitdiff
hc
2024-12-19 9370bb92b2d16684ee45cf24e879c93c509162da 
 kernel/arch/powerpc/kernel/smp.c
....@@ -1,3 +1,4 @@
1
+// SPDX-License-Identifier: GPL-2.0-or-later
12 /*
23  * SMP support for ppc.
34  *
....@@ -8,11 +9,6 @@
89  *
910  * PowerPC-64 Support added by Dave Engebretsen, Peter Bergner, and
1011  * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
11
- *
12
- *      This program is free software; you can redistribute it and/or
13
- *      modify it under the terms of the GNU General Public License
14
- *      as published by the Free Software Foundation; either version
15
- *      2 of the License, or (at your option) any later version.
1612  */
1713 
1814 #undef DEBUG
....@@ -20,6 +16,7 @@
2016 #include <linux/kernel.h>
2117 #include <linux/export.h>
2218 #include <linux/sched/mm.h>
19
+#include <linux/sched/task_stack.h>
2320 #include <linux/sched/topology.h>
2421 #include <linux/smp.h>
2522 #include <linux/interrupt.h>
....@@ -34,6 +31,9 @@
3431 #include <linux/topology.h>
3532 #include <linux/profile.h>
3633 #include <linux/processor.h>
34
+#include <linux/random.h>
35
+#include <linux/stackprotector.h>
36
+#include <linux/pgtable.h>
3737 
3838 #include <asm/ptrace.h>
3939 #include <linux/atomic.h>
....@@ -42,7 +42,6 @@
4242 #include <asm/kvm_ppc.h>
4343 #include <asm/dbell.h>
4444 #include <asm/page.h>
45
-#include <asm/pgtable.h>
4645 #include <asm/prom.h>
4746 #include <asm/smp.h>
4847 #include <asm/time.h>
....@@ -60,6 +59,8 @@
6059 #include <asm/asm-prototypes.h>
6160 #include <asm/cpu_has_feature.h>
6261 #include <asm/ftrace.h>
62
+#include <asm/kup.h>
63
+#include <asm/fadump.h>
6364 
6465 #ifdef DEBUG
6566 #include <asm/udbg.h>
....@@ -73,15 +74,44 @@
7374 static DEFINE_PER_CPU(int, cpu_state) = { 0 };
7475 #endif
7576 
76
-struct thread_info *secondary_ti;
77
+struct task_struct *secondary_current;
78
+bool has_big_cores;
79
+bool coregroup_enabled;
7780 
7881 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
82
+DEFINE_PER_CPU(cpumask_var_t, cpu_smallcore_map);
7983 DEFINE_PER_CPU(cpumask_var_t, cpu_l2_cache_map);
8084 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
85
+DEFINE_PER_CPU(cpumask_var_t, cpu_coregroup_map);
8186 
8287 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
8388 EXPORT_PER_CPU_SYMBOL(cpu_l2_cache_map);
8489 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
90
+EXPORT_SYMBOL_GPL(has_big_cores);
91
+
92
+enum {
93
+#ifdef CONFIG_SCHED_SMT
94
+	smt_idx,
95
+#endif
96
+	cache_idx,
97
+	mc_idx,
98
+	die_idx,
99
+};
100
+
101
+#define MAX_THREAD_LIST_SIZE	8
102
+#define THREAD_GROUP_SHARE_L1   1
103
+struct thread_groups {
104
+	unsigned int property;
105
+	unsigned int nr_groups;
106
+	unsigned int threads_per_group;
107
+	unsigned int thread_list[MAX_THREAD_LIST_SIZE];
108
+};
109
+
110
+/*
111
+ * On big-cores system, cpu_l1_cache_map for each CPU corresponds to
112
+ * the set its siblings that share the L1-cache.
113
+ */
114
+DEFINE_PER_CPU(cpumask_var_t, cpu_l1_cache_map);
85115 
86116 /* SMP operations for this machine */
87117 struct smp_ops_t *smp_ops;
....@@ -442,7 +472,8 @@
442472  * - delay_us > 0 is the delay before giving up waiting for targets to
443473  *   begin executing the handler, == 0 specifies indefinite delay.
444474  */
445
-int __smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *), u64 delay_us, bool safe)
475
+static int __smp_send_nmi_ipi(int cpu, void (*fn)(struct pt_regs *),
476
+				u64 delay_us, bool safe)
446477 {
447478 	unsigned long flags;
448479 	int me = raw_smp_processor_id();
....@@ -582,6 +613,15 @@
582613 {
583614 	static bool stopped = false;
584615 
616
+	/*
617
+	 * In case of fadump, register data for all CPUs is captured by f/w
618
+	 * on ibm,os-term rtas call. Skip IPI callbacks to other CPUs before
619
+	 * this rtas call to avoid tricky post processing of those CPUs'
620
+	 * backtraces.
621
+	 */
622
+	if (should_fadump_crash())
623
+		return;
624
+
585625 	if (stopped)
586626 		return;
587627 
....@@ -650,7 +690,7 @@
650690 }
651691 #endif /* CONFIG_NMI_IPI */
652692 
653
-struct thread_info *current_set[NR_CPUS];
693
+struct task_struct *current_set[NR_CPUS];
654694 
655695 static void smp_store_cpu_info(int id)
656696 {
....@@ -681,6 +721,274 @@
681721 }
682722 #endif
683723 
724
+/*
725
+ * Extends set_cpus_related. Instead of setting one CPU at a time in
726
+ * dstmask, set srcmask at oneshot. dstmask should be super set of srcmask.
727
+ */
728
+static void or_cpumasks_related(int i, int j, struct cpumask *(*srcmask)(int),
729
+				struct cpumask *(*dstmask)(int))
730
+{
731
+	struct cpumask *mask;
732
+	int k;
733
+
734
+	mask = srcmask(j);
735
+	for_each_cpu(k, srcmask(i))
736
+		cpumask_or(dstmask(k), dstmask(k), mask);
737
+
738
+	if (i == j)
739
+		return;
740
+
741
+	mask = srcmask(i);
742
+	for_each_cpu(k, srcmask(j))
743
+		cpumask_or(dstmask(k), dstmask(k), mask);
744
+}
745
+
746
+/*
747
+ * parse_thread_groups: Parses the "ibm,thread-groups" device tree
748
+ *                      property for the CPU device node @dn and stores
749
+ *                      the parsed output in the thread_groups
750
+ *                      structure @tg if the ibm,thread-groups[0]
751
+ *                      matches @property.
752
+ *
753
+ * @dn: The device node of the CPU device.
754
+ * @tg: Pointer to a thread group structure into which the parsed
755
+ *      output of "ibm,thread-groups" is stored.
756
+ * @property: The property of the thread-group that the caller is
757
+ *            interested in.
758
+ *
759
+ * ibm,thread-groups[0..N-1] array defines which group of threads in
760
+ * the CPU-device node can be grouped together based on the property.
761
+ *
762
+ * ibm,thread-groups[0] tells us the property based on which the
763
+ * threads are being grouped together. If this value is 1, it implies
764
+ * that the threads in the same group share L1, translation cache.
765
+ *
766
+ * ibm,thread-groups[1] tells us how many such thread groups exist.
767
+ *
768
+ * ibm,thread-groups[2] tells us the number of threads in each such
769
+ * group.
770
+ *
771
+ * ibm,thread-groups[3..N-1] is the list of threads identified by
772
+ * "ibm,ppc-interrupt-server#s" arranged as per their membership in
773
+ * the grouping.
774
+ *
775
+ * Example: If ibm,thread-groups = [1,2,4,5,6,7,8,9,10,11,12] it
776
+ * implies that there are 2 groups of 4 threads each, where each group
777
+ * of threads share L1, translation cache.
778
+ *
779
+ * The "ibm,ppc-interrupt-server#s" of the first group is {5,6,7,8}
780
+ * and the "ibm,ppc-interrupt-server#s" of the second group is {9, 10,
781
+ * 11, 12} structure
782
+ *
783
+ * Returns 0 on success, -EINVAL if the property does not exist,
784
+ * -ENODATA if property does not have a value, and -EOVERFLOW if the
785
+ * property data isn't large enough.
786
+ */
787
+static int parse_thread_groups(struct device_node *dn,
788
+			       struct thread_groups *tg,
789
+			       unsigned int property)
790
+{
791
+	int i;
792
+	u32 thread_group_array[3 + MAX_THREAD_LIST_SIZE];
793
+	u32 *thread_list;
794
+	size_t total_threads;
795
+	int ret;
796
+
797
+	ret = of_property_read_u32_array(dn, "ibm,thread-groups",
798
+					 thread_group_array, 3);
799
+	if (ret)
800
+		return ret;
801
+
802
+	tg->property = thread_group_array[0];
803
+	tg->nr_groups = thread_group_array[1];
804
+	tg->threads_per_group = thread_group_array[2];
805
+	if (tg->property != property ||
806
+	    tg->nr_groups < 1 ||
807
+	    tg->threads_per_group < 1)
808
+		return -ENODATA;
809
+
810
+	total_threads = tg->nr_groups * tg->threads_per_group;
811
+
812
+	ret = of_property_read_u32_array(dn, "ibm,thread-groups",
813
+					 thread_group_array,
814
+					 3 + total_threads);
815
+	if (ret)
816
+		return ret;
817
+
818
+	thread_list = &thread_group_array[3];
819
+
820
+	for (i = 0 ; i < total_threads; i++)
821
+		tg->thread_list[i] = thread_list[i];
822
+
823
+	return 0;
824
+}
825
+
826
+/*
827
+ * get_cpu_thread_group_start : Searches the thread group in tg->thread_list
828
+ *                              that @cpu belongs to.
829
+ *
830
+ * @cpu : The logical CPU whose thread group is being searched.
831
+ * @tg : The thread-group structure of the CPU node which @cpu belongs
832
+ *       to.
833
+ *
834
+ * Returns the index to tg->thread_list that points to the the start
835
+ * of the thread_group that @cpu belongs to.
836
+ *
837
+ * Returns -1 if cpu doesn't belong to any of the groups pointed to by
838
+ * tg->thread_list.
839
+ */
840
+static int get_cpu_thread_group_start(int cpu, struct thread_groups *tg)
841
+{
842
+	int hw_cpu_id = get_hard_smp_processor_id(cpu);
843
+	int i, j;
844
+
845
+	for (i = 0; i < tg->nr_groups; i++) {
846
+		int group_start = i * tg->threads_per_group;
847
+
848
+		for (j = 0; j < tg->threads_per_group; j++) {
849
+			int idx = group_start + j;
850
+
851
+			if (tg->thread_list[idx] == hw_cpu_id)
852
+				return group_start;
853
+		}
854
+	}
855
+
856
+	return -1;
857
+}
858
+
859
+static int init_cpu_l1_cache_map(int cpu)
860
+
861
+{
862
+	struct device_node *dn = of_get_cpu_node(cpu, NULL);
863
+	struct thread_groups tg = {.property = 0,
864
+				   .nr_groups = 0,
865
+				   .threads_per_group = 0};
866
+	int first_thread = cpu_first_thread_sibling(cpu);
867
+	int i, cpu_group_start = -1, err = 0;
868
+
869
+	if (!dn)
870
+		return -ENODATA;
871
+
872
+	err = parse_thread_groups(dn, &tg, THREAD_GROUP_SHARE_L1);
873
+	if (err)
874
+		goto out;
875
+
876
+	cpu_group_start = get_cpu_thread_group_start(cpu, &tg);
877
+
878
+	if (unlikely(cpu_group_start == -1)) {
879
+		WARN_ON_ONCE(1);
880
+		err = -ENODATA;
881
+		goto out;
882
+	}
883
+
884
+	zalloc_cpumask_var_node(&per_cpu(cpu_l1_cache_map, cpu),
885
+				GFP_KERNEL, cpu_to_node(cpu));
886
+
887
+	for (i = first_thread; i < first_thread + threads_per_core; i++) {
888
+		int i_group_start = get_cpu_thread_group_start(i, &tg);
889
+
890
+		if (unlikely(i_group_start == -1)) {
891
+			WARN_ON_ONCE(1);
892
+			err = -ENODATA;
893
+			goto out;
894
+		}
895
+
896
+		if (i_group_start == cpu_group_start)
897
+			cpumask_set_cpu(i, per_cpu(cpu_l1_cache_map, cpu));
898
+	}
899
+
900
+out:
901
+	of_node_put(dn);
902
+	return err;
903
+}
904
+
905
+static bool shared_caches;
906
+
907
+#ifdef CONFIG_SCHED_SMT
908
+/* cpumask of CPUs with asymmetric SMT dependency */
909
+static int powerpc_smt_flags(void)
910
+{
911
+	int flags = SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES;
912
+
913
+	if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
914
+		printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
915
+		flags |= SD_ASYM_PACKING;
916
+	}
917
+	return flags;
918
+}
919
+#endif
920
+
921
+/*
922
+ * P9 has a slightly odd architecture where pairs of cores share an L2 cache.
923
+ * This topology makes it *much* cheaper to migrate tasks between adjacent cores
924
+ * since the migrated task remains cache hot. We want to take advantage of this
925
+ * at the scheduler level so an extra topology level is required.
926
+ */
927
+static int powerpc_shared_cache_flags(void)
928
+{
929
+	return SD_SHARE_PKG_RESOURCES;
930
+}
931
+
932
+/*
933
+ * We can't just pass cpu_l2_cache_mask() directly because
934
+ * returns a non-const pointer and the compiler barfs on that.
935
+ */
936
+static const struct cpumask *shared_cache_mask(int cpu)
937
+{
938
+	return per_cpu(cpu_l2_cache_map, cpu);
939
+}
940
+
941
+#ifdef CONFIG_SCHED_SMT
942
+static const struct cpumask *smallcore_smt_mask(int cpu)
943
+{
944
+	return cpu_smallcore_mask(cpu);
945
+}
946
+#endif
947
+
948
+static struct cpumask *cpu_coregroup_mask(int cpu)
949
+{
950
+	return per_cpu(cpu_coregroup_map, cpu);
951
+}
952
+
953
+static bool has_coregroup_support(void)
954
+{
955
+	return coregroup_enabled;
956
+}
957
+
958
+static const struct cpumask *cpu_mc_mask(int cpu)
959
+{
960
+	return cpu_coregroup_mask(cpu);
961
+}
962
+
963
+static struct sched_domain_topology_level powerpc_topology[] = {
964
+#ifdef CONFIG_SCHED_SMT
965
+	{ cpu_smt_mask, powerpc_smt_flags, SD_INIT_NAME(SMT) },
966
+#endif
967
+	{ shared_cache_mask, powerpc_shared_cache_flags, SD_INIT_NAME(CACHE) },
968
+	{ cpu_mc_mask, SD_INIT_NAME(MC) },
969
+	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
970
+	{ NULL, },
971
+};
972
+
973
+static int __init init_big_cores(void)
974
+{
975
+	int cpu;
976
+
977
+	for_each_possible_cpu(cpu) {
978
+		int err = init_cpu_l1_cache_map(cpu);
979
+
980
+		if (err)
981
+			return err;
982
+
983
+		zalloc_cpumask_var_node(&per_cpu(cpu_smallcore_map, cpu),
984
+					GFP_KERNEL,
985
+					cpu_to_node(cpu));
986
+	}
987
+
988
+	has_big_cores = true;
989
+	return 0;
990
+}
991
+
684992 void __init smp_prepare_cpus(unsigned int max_cpus)
685993 {
686994 	unsigned int cpu;
....@@ -704,6 +1012,11 @@
7041012 					GFP_KERNEL, cpu_to_node(cpu));
7051013 		zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu),
7061014 					GFP_KERNEL, cpu_to_node(cpu));
1015
+		if (has_coregroup_support())
1016
+			zalloc_cpumask_var_node(&per_cpu(cpu_coregroup_map, cpu),
1017
+						GFP_KERNEL, cpu_to_node(cpu));
1018
+
1019
+#ifdef CONFIG_NEED_MULTIPLE_NODES
7071020 		/*
7081021 		 * numa_node_id() works after this.
7091022 		 */
....@@ -712,12 +1025,22 @@
7121025 			set_cpu_numa_mem(cpu,
7131026 				local_memory_node(numa_cpu_lookup_table[cpu]));
7141027 		}
1028
+#endif
7151029 	}
7161030 
7171031 	/* Init the cpumasks so the boot CPU is related to itself */
7181032 	cpumask_set_cpu(boot_cpuid, cpu_sibling_mask(boot_cpuid));
7191033 	cpumask_set_cpu(boot_cpuid, cpu_l2_cache_mask(boot_cpuid));
7201034 	cpumask_set_cpu(boot_cpuid, cpu_core_mask(boot_cpuid));
1035
+
1036
+	if (has_coregroup_support())
1037
+		cpumask_set_cpu(boot_cpuid, cpu_coregroup_mask(boot_cpuid));
1038
+
1039
+	init_big_cores();
1040
+	if (has_big_cores) {
1041
+		cpumask_set_cpu(boot_cpuid,
1042
+				cpu_smallcore_mask(boot_cpuid));
1043
+	}
7211044 
7221045 	if (smp_ops && smp_ops->probe)
7231046 		smp_ops->probe();
....@@ -730,7 +1053,7 @@
7301053 	paca_ptrs[boot_cpuid]->__current = current;
7311054 #endif
7321055 	set_numa_node(numa_cpu_lookup_table[boot_cpuid]);
733
-	current_set[boot_cpuid] = task_thread_info(current);
1056
+	current_set[boot_cpuid] = current;
7341057 }
7351058 
7361059 #ifdef CONFIG_HOTPLUG_CPU
....@@ -815,14 +1138,13 @@
8151138 
8161139 static void cpu_idle_thread_init(unsigned int cpu, struct task_struct *idle)
8171140 {
818
-	struct thread_info *ti = task_thread_info(idle);
819
-
8201141 #ifdef CONFIG_PPC64
8211142 	paca_ptrs[cpu]->__current = idle;
822
-	paca_ptrs[cpu]->kstack = (unsigned long)ti + THREAD_SIZE - STACK_FRAME_OVERHEAD;
1143
+	paca_ptrs[cpu]->kstack = (unsigned long)task_stack_page(idle) +
1144
+				 THREAD_SIZE - STACK_FRAME_OVERHEAD;
8231145 #endif
824
-	ti->cpu = cpu;
825
-	secondary_ti = current_set[cpu] = ti;
1146
+	idle->cpu = cpu;
1147
+	secondary_current = current_set[cpu] = idle;
8261148 }
8271149 
8281150 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
....@@ -964,26 +1286,46 @@
9641286 	return cache;
9651287 }
9661288 
967
-static bool update_mask_by_l2(int cpu, struct cpumask *(*mask_fn)(int))
1289
+static bool update_mask_by_l2(int cpu, cpumask_var_t *mask)
9681290 {
1291
+	struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
9691292 	struct device_node *l2_cache, *np;
9701293 	int i;
9711294 
972
-	l2_cache = cpu_to_l2cache(cpu);
973
-	if (!l2_cache)
974
-		return false;
1295
+	if (has_big_cores)
1296
+		submask_fn = cpu_smallcore_mask;
9751297 
976
-	for_each_cpu(i, cpu_online_mask) {
1298
+	l2_cache = cpu_to_l2cache(cpu);
1299
+	if (!l2_cache || !*mask) {
1300
+		/* Assume only core siblings share cache with this CPU */
1301
+		for_each_cpu(i, submask_fn(cpu))
1302
+			set_cpus_related(cpu, i, cpu_l2_cache_mask);
1303
+
1304
+		return false;
1305
+	}
1306
+
1307
+	cpumask_and(*mask, cpu_online_mask, cpu_cpu_mask(cpu));
1308
+
1309
+	/* Update l2-cache mask with all the CPUs that are part of submask */
1310
+	or_cpumasks_related(cpu, cpu, submask_fn, cpu_l2_cache_mask);
1311
+
1312
+	/* Skip all CPUs already part of current CPU l2-cache mask */
1313
+	cpumask_andnot(*mask, *mask, cpu_l2_cache_mask(cpu));
1314
+
1315
+	for_each_cpu(i, *mask) {
9771316 		/*
9781317 		 * when updating the marks the current CPU has not been marked
9791318 		 * online, but we need to update the cache masks
9801319 		 */
9811320 		np = cpu_to_l2cache(i);
982
-		if (!np)
983
-			continue;
9841321 
985
-		if (np == l2_cache)
986
-			set_cpus_related(cpu, i, mask_fn);
1322
+		/* Skip all CPUs already part of current CPU l2-cache */
1323
+		if (np == l2_cache) {
1324
+			or_cpumasks_related(cpu, i, submask_fn, cpu_l2_cache_mask);
1325
+			cpumask_andnot(*mask, *mask, submask_fn(i));
1326
+		} else {
1327
+			cpumask_andnot(*mask, *mask, cpu_l2_cache_mask(i));
1328
+		}
9871329 
9881330 		of_node_put(np);
9891331 	}
....@@ -995,21 +1337,87 @@
9951337 #ifdef CONFIG_HOTPLUG_CPU
9961338 static void remove_cpu_from_masks(int cpu)
9971339 {
1340
+	struct cpumask *(*mask_fn)(int) = cpu_sibling_mask;
9981341 	int i;
9991342 
1000
-	/* NB: cpu_core_mask is a superset of the others */
1001
-	for_each_cpu(i, cpu_core_mask(cpu)) {
1002
-		set_cpus_unrelated(cpu, i, cpu_core_mask);
1343
+	if (shared_caches)
1344
+		mask_fn = cpu_l2_cache_mask;
1345
+
1346
+	for_each_cpu(i, mask_fn(cpu)) {
10031347 		set_cpus_unrelated(cpu, i, cpu_l2_cache_mask);
10041348 		set_cpus_unrelated(cpu, i, cpu_sibling_mask);
1349
+		if (has_big_cores)
1350
+			set_cpus_unrelated(cpu, i, cpu_smallcore_mask);
1351
+	}
1352
+
1353
+	for_each_cpu(i, cpu_core_mask(cpu))
1354
+		set_cpus_unrelated(cpu, i, cpu_core_mask);
1355
+
1356
+	if (has_coregroup_support()) {
1357
+		for_each_cpu(i, cpu_coregroup_mask(cpu))
1358
+			set_cpus_unrelated(cpu, i, cpu_coregroup_mask);
10051359 	}
10061360 }
10071361 #endif
10081362 
1363
+static inline void add_cpu_to_smallcore_masks(int cpu)
1364
+{
1365
+	int i;
1366
+
1367
+	if (!has_big_cores)
1368
+		return;
1369
+
1370
+	cpumask_set_cpu(cpu, cpu_smallcore_mask(cpu));
1371
+
1372
+	for_each_cpu(i, per_cpu(cpu_l1_cache_map, cpu)) {
1373
+		if (cpu_online(i))
1374
+			set_cpus_related(i, cpu, cpu_smallcore_mask);
1375
+	}
1376
+}
1377
+
1378
+static void update_coregroup_mask(int cpu, cpumask_var_t *mask)
1379
+{
1380
+	struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
1381
+	int coregroup_id = cpu_to_coregroup_id(cpu);
1382
+	int i;
1383
+
1384
+	if (shared_caches)
1385
+		submask_fn = cpu_l2_cache_mask;
1386
+
1387
+	if (!*mask) {
1388
+		/* Assume only siblings are part of this CPU's coregroup */
1389
+		for_each_cpu(i, submask_fn(cpu))
1390
+			set_cpus_related(cpu, i, cpu_coregroup_mask);
1391
+
1392
+		return;
1393
+	}
1394
+
1395
+	cpumask_and(*mask, cpu_online_mask, cpu_cpu_mask(cpu));
1396
+
1397
+	/* Update coregroup mask with all the CPUs that are part of submask */
1398
+	or_cpumasks_related(cpu, cpu, submask_fn, cpu_coregroup_mask);
1399
+
1400
+	/* Skip all CPUs already part of coregroup mask */
1401
+	cpumask_andnot(*mask, *mask, cpu_coregroup_mask(cpu));
1402
+
1403
+	for_each_cpu(i, *mask) {
1404
+		/* Skip all CPUs not part of this coregroup */
1405
+		if (coregroup_id == cpu_to_coregroup_id(i)) {
1406
+			or_cpumasks_related(cpu, i, submask_fn, cpu_coregroup_mask);
1407
+			cpumask_andnot(*mask, *mask, submask_fn(i));
1408
+		} else {
1409
+			cpumask_andnot(*mask, *mask, cpu_coregroup_mask(i));
1410
+		}
1411
+	}
1412
+}
1413
+
10091414 static void add_cpu_to_masks(int cpu)
10101415 {
1416
+	struct cpumask *(*submask_fn)(int) = cpu_sibling_mask;
10111417 	int first_thread = cpu_first_thread_sibling(cpu);
1012
-	int chipid = cpu_to_chip_id(cpu);
1418
+	int chip_id = cpu_to_chip_id(cpu);
1419
+	cpumask_var_t mask;
1420
+	bool ret;
10131421 	int i;
10141422 
10151423 	/*
....@@ -1017,47 +1425,57 @@
10171425 	 * add it to it's own thread sibling mask.
10181426 	 */
10191427 	cpumask_set_cpu(cpu, cpu_sibling_mask(cpu));
1428
+	cpumask_set_cpu(cpu, cpu_core_mask(cpu));
10201429 
10211430 	for (i = first_thread; i < first_thread + threads_per_core; i++)
10221431 		if (cpu_online(i))
10231432 			set_cpus_related(i, cpu, cpu_sibling_mask);
10241433 
1025
-	/*
1026
-	 * Copy the thread sibling mask into the cache sibling mask
1027
-	 * and mark any CPUs that share an L2 with this CPU.
1028
-	 */
1029
-	for_each_cpu(i, cpu_sibling_mask(cpu))
1030
-		set_cpus_related(cpu, i, cpu_l2_cache_mask);
1031
-	update_mask_by_l2(cpu, cpu_l2_cache_mask);
1434
+	add_cpu_to_smallcore_masks(cpu);
10321435 
1033
-	/*
1034
-	 * Copy the cache sibling mask into core sibling mask and mark
1035
-	 * any CPUs on the same chip as this CPU.
1036
-	 */
1037
-	for_each_cpu(i, cpu_l2_cache_mask(cpu))
1038
-		set_cpus_related(cpu, i, cpu_core_mask);
1436
+	/* In CPU-hotplug path, hence use GFP_ATOMIC */
1437
+	ret = alloc_cpumask_var_node(&mask, GFP_ATOMIC, cpu_to_node(cpu));
1438
+	update_mask_by_l2(cpu, &mask);
10391439 
1040
-	if (chipid == -1)
1041
-		return;
1440
+	if (has_coregroup_support())
1441
+		update_coregroup_mask(cpu, &mask);
10421442 
1043
-	for_each_cpu(i, cpu_online_mask)
1044
-		if (cpu_to_chip_id(i) == chipid)
1045
-			set_cpus_related(cpu, i, cpu_core_mask);
1443
+	if (shared_caches)
1444
+		submask_fn = cpu_l2_cache_mask;
1445
+
1446
+	/* Update core_mask with all the CPUs that are part of submask */
1447
+	or_cpumasks_related(cpu, cpu, submask_fn, cpu_core_mask);
1448
+
1449
+	/* Skip all CPUs already part of current CPU core mask */
1450
+	cpumask_andnot(mask, cpu_online_mask, cpu_core_mask(cpu));
1451
+
1452
+	/* If chip_id is -1; limit the cpu_core_mask to within DIE*/
1453
+	if (chip_id == -1)
1454
+		cpumask_and(mask, mask, cpu_cpu_mask(cpu));
1455
+
1456
+	for_each_cpu(i, mask) {
1457
+		if (chip_id == cpu_to_chip_id(i)) {
1458
+			or_cpumasks_related(cpu, i, submask_fn, cpu_core_mask);
1459
+			cpumask_andnot(mask, mask, submask_fn(i));
1460
+		} else {
1461
+			cpumask_andnot(mask, mask, cpu_core_mask(i));
1462
+		}
1463
+	}
1464
+
1465
+	free_cpumask_var(mask);
10461466 }
1047
-
1048
-static bool shared_caches;
10491467 
10501468 /* Activate a secondary processor. */
10511469 void start_secondary(void *unused)
10521470 {
1053
-	unsigned int cpu = smp_processor_id();
1471
+	unsigned int cpu = raw_smp_processor_id();
10541472 
10551473 	mmgrab(&init_mm);
10561474 	current->active_mm = &init_mm;
10571475 
10581476 	smp_store_cpu_info(cpu);
10591477 	set_dec(tb_ticks_per_jiffy);
1060
-	preempt_disable();
1478
+	rcu_cpu_starting(cpu);
10611479 	cpu_callin_map[cpu] = 1;
10621480 
10631481 	if (smp_ops->setup_cpu)
....@@ -1083,12 +1501,22 @@
10831501 	 * Check for any shared caches. Note that this must be done on a
10841502 	 * per-core basis because one core in the pair might be disabled.
10851503 	 */
1086
-	if (!cpumask_equal(cpu_l2_cache_mask(cpu), cpu_sibling_mask(cpu)))
1087
-		shared_caches = true;
1504
+	if (!shared_caches) {
1505
+		struct cpumask *(*sibling_mask)(int) = cpu_sibling_mask;
1506
+		struct cpumask *mask = cpu_l2_cache_mask(cpu);
1507
+
1508
+		if (has_big_cores)
1509
+			sibling_mask = cpu_smallcore_mask;
1510
+
1511
+		if (cpumask_weight(mask) > cpumask_weight(sibling_mask(cpu)))
1512
+			shared_caches = true;
1513
+	}
10881514 
10891515 	smp_wmb();
10901516 	notify_cpu_starting(cpu);
10911517 	set_cpu_online(cpu, true);
1518
+
1519
+	boot_init_stack_canary();
10921520 
10931521 	local_irq_enable();
10941522 
....@@ -1107,56 +1535,44 @@
11071535 }
11081536 #endif
11091537 
1110
-#ifdef CONFIG_SCHED_SMT
1111
-/* cpumask of CPUs with asymetric SMT dependancy */
1112
-static int powerpc_smt_flags(void)
1538
+static void fixup_topology(void)
11131539 {
1114
-	int flags = SD_SHARE_CPUCAPACITY | SD_SHARE_PKG_RESOURCES;
1540
+	int i;
11151541 
1116
-	if (cpu_has_feature(CPU_FTR_ASYM_SMT)) {
1117
-		printk_once(KERN_INFO "Enabling Asymmetric SMT scheduling\n");
1118
-		flags |= SD_ASYM_PACKING;
1542
+#ifdef CONFIG_SCHED_SMT
1543
+	if (has_big_cores) {
1544
+		pr_info("Big cores detected but using small core scheduling\n");
1545
+		powerpc_topology[smt_idx].mask = smallcore_smt_mask;
11191546 	}
1120
-	return flags;
1121
-}
11221547 #endif
11231548 
1124
-static struct sched_domain_topology_level powerpc_topology[] = {
1125
-#ifdef CONFIG_SCHED_SMT
1126
-	{ cpu_smt_mask, powerpc_smt_flags, SD_INIT_NAME(SMT) },
1549
+	if (!has_coregroup_support())
1550
+		powerpc_topology[mc_idx].mask = powerpc_topology[cache_idx].mask;
1551
+
1552
+	/*
1553
+	 * Try to consolidate topology levels here instead of
1554
+	 * allowing scheduler to degenerate.
1555
+	 * - Dont consolidate if masks are different.
1556
+	 * - Dont consolidate if sd_flags exists and are different.
1557
+	 */
1558
+	for (i = 1; i <= die_idx; i++) {
1559
+		if (powerpc_topology[i].mask != powerpc_topology[i - 1].mask)
1560
+			continue;
1561
+
1562
+		if (powerpc_topology[i].sd_flags && powerpc_topology[i - 1].sd_flags &&
1563
+				powerpc_topology[i].sd_flags != powerpc_topology[i - 1].sd_flags)
1564
+			continue;
1565
+
1566
+		if (!powerpc_topology[i - 1].sd_flags)
1567
+			powerpc_topology[i - 1].sd_flags = powerpc_topology[i].sd_flags;
1568
+
1569
+		powerpc_topology[i].mask = powerpc_topology[i + 1].mask;
1570
+		powerpc_topology[i].sd_flags = powerpc_topology[i + 1].sd_flags;
1571
+#ifdef CONFIG_SCHED_DEBUG
1572
+		powerpc_topology[i].name = powerpc_topology[i + 1].name;
11271573 #endif
1128
-	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
1129
-	{ NULL, },
1130
-};
1131
-
1132
-/*
1133
- * P9 has a slightly odd architecture where pairs of cores share an L2 cache.
1134
- * This topology makes it *much* cheaper to migrate tasks between adjacent cores
1135
- * since the migrated task remains cache hot. We want to take advantage of this
1136
- * at the scheduler level so an extra topology level is required.
1137
- */
1138
-static int powerpc_shared_cache_flags(void)
1139
-{
1140
-	return SD_SHARE_PKG_RESOURCES;
1574
+	}
11411575 }
1142
-
1143
-/*
1144
- * We can't just pass cpu_l2_cache_mask() directly because
1145
- * returns a non-const pointer and the compiler barfs on that.
1146
- */
1147
-static const struct cpumask *shared_cache_mask(int cpu)
1148
-{
1149
-	return cpu_l2_cache_mask(cpu);
1150
-}
1151
-
1152
-static struct sched_domain_topology_level power9_topology[] = {
1153
-#ifdef CONFIG_SCHED_SMT
1154
-	{ cpu_smt_mask, powerpc_smt_flags, SD_INIT_NAME(SMT) },
1155
-#endif
1156
-	{ shared_cache_mask, powerpc_shared_cache_flags, SD_INIT_NAME(CACHE) },
1157
-	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
1158
-	{ NULL, },
1159
-};
11601576 
11611577 void __init smp_cpus_done(unsigned int max_cpus)
11621578 {
....@@ -1169,24 +1585,10 @@
11691585 	if (smp_ops && smp_ops->bringup_done)
11701586 		smp_ops->bringup_done();
11711587 
1172
-	/*
1173
-	 * On a shared LPAR, associativity needs to be requested.
1174
-	 * Hence, get numa topology before dumping cpu topology
1175
-	 */
1176
-	shared_proc_topology_init();
11771588 	dump_numa_cpu_topology();
11781589 
1179
-	/*
1180
-	 * If any CPU detects that it's sharing a cache with another CPU then
1181
-	 * use the deeper topology that is aware of this sharing.
1182
-	 */
1183
-	if (shared_caches) {
1184
-		pr_info("Using shared cache scheduler topology\n");
1185
-		set_sched_topology(power9_topology);
1186
-	} else {
1187
-		pr_info("Using standard scheduler topology\n");
1188
-		set_sched_topology(powerpc_topology);
1189
-	}
1590
+	fixup_topology();
1591
+	set_sched_topology(powerpc_topology);
11901592 }
11911593 
11921594 #ifdef CONFIG_HOTPLUG_CPU
....@@ -1216,7 +1618,7 @@
12161618 		smp_ops->cpu_die(cpu);
12171619 }
12181620 
1219
-void cpu_die(void)
1621
+void arch_cpu_idle_dead(void)
12201622 {
12211623 	/*
12221624 	 * Disable on the down path. This will be re-enabled by
....@@ -1224,8 +1626,8 @@
12241626 	 */
12251627 	this_cpu_disable_ftrace();
12261628 
1227
-	if (ppc_md.cpu_die)
1228
-		ppc_md.cpu_die();
1629
+	if (smp_ops->cpu_offline_self)
1630
+		smp_ops->cpu_offline_self();
12291631 
12301632 	/* If we return, we re-enter start_secondary */
12311633 	start_secondary_resume();