xref: /openbmc/linux/arch/s390/kernel/topology.c (revision 8cb5d748)
1 /*
2  *    Copyright IBM Corp. 2007, 2011
3  *    Author(s): Heiko Carstens <heiko.carstens@de.ibm.com>
4  */
5 
6 #define KMSG_COMPONENT "cpu"
7 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
8 
9 #include <linux/workqueue.h>
10 #include <linux/bootmem.h>
11 #include <linux/uaccess.h>
12 #include <linux/sysctl.h>
13 #include <linux/cpuset.h>
14 #include <linux/device.h>
15 #include <linux/export.h>
16 #include <linux/kernel.h>
17 #include <linux/sched.h>
18 #include <linux/sched/topology.h>
19 #include <linux/delay.h>
20 #include <linux/init.h>
21 #include <linux/slab.h>
22 #include <linux/cpu.h>
23 #include <linux/smp.h>
24 #include <linux/mm.h>
25 #include <linux/nodemask.h>
26 #include <linux/node.h>
27 #include <asm/sysinfo.h>
28 #include <asm/numa.h>
29 
30 #define PTF_HORIZONTAL	(0UL)
31 #define PTF_VERTICAL	(1UL)
32 #define PTF_CHECK	(2UL)
33 
34 enum {
35 	TOPOLOGY_MODE_HW,
36 	TOPOLOGY_MODE_SINGLE,
37 	TOPOLOGY_MODE_PACKAGE,
38 	TOPOLOGY_MODE_UNINITIALIZED
39 };
40 
41 struct mask_info {
42 	struct mask_info *next;
43 	unsigned char id;
44 	cpumask_t mask;
45 };
46 
47 static int topology_mode = TOPOLOGY_MODE_UNINITIALIZED;
48 static void set_topology_timer(void);
49 static void topology_work_fn(struct work_struct *work);
50 static struct sysinfo_15_1_x *tl_info;
51 
52 static DECLARE_WORK(topology_work, topology_work_fn);
53 
54 /*
55  * Socket/Book linked lists and cpu_topology updates are
56  * protected by "sched_domains_mutex".
57  */
58 static struct mask_info socket_info;
59 static struct mask_info book_info;
60 static struct mask_info drawer_info;
61 
62 struct cpu_topology_s390 cpu_topology[NR_CPUS];
63 EXPORT_SYMBOL_GPL(cpu_topology);
64 
65 cpumask_t cpus_with_topology;
66 
67 static cpumask_t cpu_group_map(struct mask_info *info, unsigned int cpu)
68 {
69 	cpumask_t mask;
70 
71 	cpumask_copy(&mask, cpumask_of(cpu));
72 	switch (topology_mode) {
73 	case TOPOLOGY_MODE_HW:
74 		while (info) {
75 			if (cpumask_test_cpu(cpu, &info->mask)) {
76 				mask = info->mask;
77 				break;
78 			}
79 			info = info->next;
80 		}
81 		if (cpumask_empty(&mask))
82 			cpumask_copy(&mask, cpumask_of(cpu));
83 		break;
84 	case TOPOLOGY_MODE_PACKAGE:
85 		cpumask_copy(&mask, cpu_present_mask);
86 		break;
87 	default:
88 		/* fallthrough */
89 	case TOPOLOGY_MODE_SINGLE:
90 		cpumask_copy(&mask, cpumask_of(cpu));
91 		break;
92 	}
93 	return mask;
94 }
95 
96 static cpumask_t cpu_thread_map(unsigned int cpu)
97 {
98 	cpumask_t mask;
99 	int i;
100 
101 	cpumask_copy(&mask, cpumask_of(cpu));
102 	if (topology_mode != TOPOLOGY_MODE_HW)
103 		return mask;
104 	cpu -= cpu % (smp_cpu_mtid + 1);
105 	for (i = 0; i <= smp_cpu_mtid; i++)
106 		if (cpu_present(cpu + i))
107 			cpumask_set_cpu(cpu + i, &mask);
108 	return mask;
109 }
110 
111 #define TOPOLOGY_CORE_BITS	64
112 
113 static void add_cpus_to_mask(struct topology_core *tl_core,
114 			     struct mask_info *drawer,
115 			     struct mask_info *book,
116 			     struct mask_info *socket)
117 {
118 	struct cpu_topology_s390 *topo;
119 	unsigned int core;
120 
121 	for_each_set_bit(core, &tl_core->mask, TOPOLOGY_CORE_BITS) {
122 		unsigned int rcore;
123 		int lcpu, i;
124 
125 		rcore = TOPOLOGY_CORE_BITS - 1 - core + tl_core->origin;
126 		lcpu = smp_find_processor_id(rcore << smp_cpu_mt_shift);
127 		if (lcpu < 0)
128 			continue;
129 		for (i = 0; i <= smp_cpu_mtid; i++) {
130 			topo = &cpu_topology[lcpu + i];
131 			topo->drawer_id = drawer->id;
132 			topo->book_id = book->id;
133 			topo->socket_id = socket->id;
134 			topo->core_id = rcore;
135 			topo->thread_id = lcpu + i;
136 			cpumask_set_cpu(lcpu + i, &drawer->mask);
137 			cpumask_set_cpu(lcpu + i, &book->mask);
138 			cpumask_set_cpu(lcpu + i, &socket->mask);
139 			cpumask_set_cpu(lcpu + i, &cpus_with_topology);
140 			smp_cpu_set_polarization(lcpu + i, tl_core->pp);
141 		}
142 	}
143 }
144 
145 static void clear_masks(void)
146 {
147 	struct mask_info *info;
148 
149 	info = &socket_info;
150 	while (info) {
151 		cpumask_clear(&info->mask);
152 		info = info->next;
153 	}
154 	info = &book_info;
155 	while (info) {
156 		cpumask_clear(&info->mask);
157 		info = info->next;
158 	}
159 	info = &drawer_info;
160 	while (info) {
161 		cpumask_clear(&info->mask);
162 		info = info->next;
163 	}
164 }
165 
166 static union topology_entry *next_tle(union topology_entry *tle)
167 {
168 	if (!tle->nl)
169 		return (union topology_entry *)((struct topology_core *)tle + 1);
170 	return (union topology_entry *)((struct topology_container *)tle + 1);
171 }
172 
173 static void tl_to_masks(struct sysinfo_15_1_x *info)
174 {
175 	struct mask_info *socket = &socket_info;
176 	struct mask_info *book = &book_info;
177 	struct mask_info *drawer = &drawer_info;
178 	union topology_entry *tle, *end;
179 
180 	clear_masks();
181 	tle = info->tle;
182 	end = (union topology_entry *)((unsigned long)info + info->length);
183 	while (tle < end) {
184 		switch (tle->nl) {
185 		case 3:
186 			drawer = drawer->next;
187 			drawer->id = tle->container.id;
188 			break;
189 		case 2:
190 			book = book->next;
191 			book->id = tle->container.id;
192 			break;
193 		case 1:
194 			socket = socket->next;
195 			socket->id = tle->container.id;
196 			break;
197 		case 0:
198 			add_cpus_to_mask(&tle->cpu, drawer, book, socket);
199 			break;
200 		default:
201 			clear_masks();
202 			return;
203 		}
204 		tle = next_tle(tle);
205 	}
206 }
207 
208 static void topology_update_polarization_simple(void)
209 {
210 	int cpu;
211 
212 	for_each_possible_cpu(cpu)
213 		smp_cpu_set_polarization(cpu, POLARIZATION_HRZ);
214 }
215 
216 static int ptf(unsigned long fc)
217 {
218 	int rc;
219 
220 	asm volatile(
221 		"	.insn	rre,0xb9a20000,%1,%1\n"
222 		"	ipm	%0\n"
223 		"	srl	%0,28\n"
224 		: "=d" (rc)
225 		: "d" (fc)  : "cc");
226 	return rc;
227 }
228 
229 int topology_set_cpu_management(int fc)
230 {
231 	int cpu, rc;
232 
233 	if (!MACHINE_HAS_TOPOLOGY)
234 		return -EOPNOTSUPP;
235 	if (fc)
236 		rc = ptf(PTF_VERTICAL);
237 	else
238 		rc = ptf(PTF_HORIZONTAL);
239 	if (rc)
240 		return -EBUSY;
241 	for_each_possible_cpu(cpu)
242 		smp_cpu_set_polarization(cpu, POLARIZATION_UNKNOWN);
243 	return rc;
244 }
245 
246 static void update_cpu_masks(void)
247 {
248 	struct cpu_topology_s390 *topo;
249 	int cpu, id;
250 
251 	for_each_possible_cpu(cpu) {
252 		topo = &cpu_topology[cpu];
253 		topo->thread_mask = cpu_thread_map(cpu);
254 		topo->core_mask = cpu_group_map(&socket_info, cpu);
255 		topo->book_mask = cpu_group_map(&book_info, cpu);
256 		topo->drawer_mask = cpu_group_map(&drawer_info, cpu);
257 		if (topology_mode != TOPOLOGY_MODE_HW) {
258 			id = topology_mode == TOPOLOGY_MODE_PACKAGE ? 0 : cpu;
259 			topo->thread_id = cpu;
260 			topo->core_id = cpu;
261 			topo->socket_id = id;
262 			topo->book_id = id;
263 			topo->drawer_id = id;
264 			if (cpu_present(cpu))
265 				cpumask_set_cpu(cpu, &cpus_with_topology);
266 		}
267 	}
268 	numa_update_cpu_topology();
269 }
270 
271 void store_topology(struct sysinfo_15_1_x *info)
272 {
273 	stsi(info, 15, 1, topology_mnest_limit());
274 }
275 
276 static int __arch_update_cpu_topology(void)
277 {
278 	struct sysinfo_15_1_x *info = tl_info;
279 	int rc = 0;
280 
281 	mutex_lock(&smp_cpu_state_mutex);
282 	cpumask_clear(&cpus_with_topology);
283 	if (MACHINE_HAS_TOPOLOGY) {
284 		rc = 1;
285 		store_topology(info);
286 		tl_to_masks(info);
287 	}
288 	update_cpu_masks();
289 	if (!MACHINE_HAS_TOPOLOGY)
290 		topology_update_polarization_simple();
291 	mutex_unlock(&smp_cpu_state_mutex);
292 	return rc;
293 }
294 
295 int arch_update_cpu_topology(void)
296 {
297 	struct device *dev;
298 	int cpu, rc;
299 
300 	rc = __arch_update_cpu_topology();
301 	for_each_online_cpu(cpu) {
302 		dev = get_cpu_device(cpu);
303 		kobject_uevent(&dev->kobj, KOBJ_CHANGE);
304 	}
305 	return rc;
306 }
307 
308 static void topology_work_fn(struct work_struct *work)
309 {
310 	rebuild_sched_domains();
311 }
312 
313 void topology_schedule_update(void)
314 {
315 	schedule_work(&topology_work);
316 }
317 
318 static void topology_flush_work(void)
319 {
320 	flush_work(&topology_work);
321 }
322 
323 static void topology_timer_fn(unsigned long ignored)
324 {
325 	if (ptf(PTF_CHECK))
326 		topology_schedule_update();
327 	set_topology_timer();
328 }
329 
330 static struct timer_list topology_timer =
331 	TIMER_DEFERRED_INITIALIZER(topology_timer_fn, 0, 0);
332 
333 static atomic_t topology_poll = ATOMIC_INIT(0);
334 
335 static void set_topology_timer(void)
336 {
337 	if (atomic_add_unless(&topology_poll, -1, 0))
338 		mod_timer(&topology_timer, jiffies + HZ / 10);
339 	else
340 		mod_timer(&topology_timer, jiffies + HZ * 60);
341 }
342 
343 void topology_expect_change(void)
344 {
345 	if (!MACHINE_HAS_TOPOLOGY)
346 		return;
347 	/* This is racy, but it doesn't matter since it is just a heuristic.
348 	 * Worst case is that we poll in a higher frequency for a bit longer.
349 	 */
350 	if (atomic_read(&topology_poll) > 60)
351 		return;
352 	atomic_add(60, &topology_poll);
353 	set_topology_timer();
354 }
355 
356 static int cpu_management;
357 
358 static ssize_t dispatching_show(struct device *dev,
359 				struct device_attribute *attr,
360 				char *buf)
361 {
362 	ssize_t count;
363 
364 	mutex_lock(&smp_cpu_state_mutex);
365 	count = sprintf(buf, "%d\n", cpu_management);
366 	mutex_unlock(&smp_cpu_state_mutex);
367 	return count;
368 }
369 
370 static ssize_t dispatching_store(struct device *dev,
371 				 struct device_attribute *attr,
372 				 const char *buf,
373 				 size_t count)
374 {
375 	int val, rc;
376 	char delim;
377 
378 	if (sscanf(buf, "%d %c", &val, &delim) != 1)
379 		return -EINVAL;
380 	if (val != 0 && val != 1)
381 		return -EINVAL;
382 	rc = 0;
383 	get_online_cpus();
384 	mutex_lock(&smp_cpu_state_mutex);
385 	if (cpu_management == val)
386 		goto out;
387 	rc = topology_set_cpu_management(val);
388 	if (rc)
389 		goto out;
390 	cpu_management = val;
391 	topology_expect_change();
392 out:
393 	mutex_unlock(&smp_cpu_state_mutex);
394 	put_online_cpus();
395 	return rc ? rc : count;
396 }
397 static DEVICE_ATTR(dispatching, 0644, dispatching_show,
398 			 dispatching_store);
399 
400 static ssize_t cpu_polarization_show(struct device *dev,
401 				     struct device_attribute *attr, char *buf)
402 {
403 	int cpu = dev->id;
404 	ssize_t count;
405 
406 	mutex_lock(&smp_cpu_state_mutex);
407 	switch (smp_cpu_get_polarization(cpu)) {
408 	case POLARIZATION_HRZ:
409 		count = sprintf(buf, "horizontal\n");
410 		break;
411 	case POLARIZATION_VL:
412 		count = sprintf(buf, "vertical:low\n");
413 		break;
414 	case POLARIZATION_VM:
415 		count = sprintf(buf, "vertical:medium\n");
416 		break;
417 	case POLARIZATION_VH:
418 		count = sprintf(buf, "vertical:high\n");
419 		break;
420 	default:
421 		count = sprintf(buf, "unknown\n");
422 		break;
423 	}
424 	mutex_unlock(&smp_cpu_state_mutex);
425 	return count;
426 }
427 static DEVICE_ATTR(polarization, 0444, cpu_polarization_show, NULL);
428 
429 static struct attribute *topology_cpu_attrs[] = {
430 	&dev_attr_polarization.attr,
431 	NULL,
432 };
433 
434 static struct attribute_group topology_cpu_attr_group = {
435 	.attrs = topology_cpu_attrs,
436 };
437 
438 int topology_cpu_init(struct cpu *cpu)
439 {
440 	return sysfs_create_group(&cpu->dev.kobj, &topology_cpu_attr_group);
441 }
442 
443 static const struct cpumask *cpu_thread_mask(int cpu)
444 {
445 	return &cpu_topology[cpu].thread_mask;
446 }
447 
448 
449 const struct cpumask *cpu_coregroup_mask(int cpu)
450 {
451 	return &cpu_topology[cpu].core_mask;
452 }
453 
454 static const struct cpumask *cpu_book_mask(int cpu)
455 {
456 	return &cpu_topology[cpu].book_mask;
457 }
458 
459 static const struct cpumask *cpu_drawer_mask(int cpu)
460 {
461 	return &cpu_topology[cpu].drawer_mask;
462 }
463 
464 static struct sched_domain_topology_level s390_topology[] = {
465 	{ cpu_thread_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
466 	{ cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
467 	{ cpu_book_mask, SD_INIT_NAME(BOOK) },
468 	{ cpu_drawer_mask, SD_INIT_NAME(DRAWER) },
469 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
470 	{ NULL, },
471 };
472 
473 static void __init alloc_masks(struct sysinfo_15_1_x *info,
474 			       struct mask_info *mask, int offset)
475 {
476 	int i, nr_masks;
477 
478 	nr_masks = info->mag[TOPOLOGY_NR_MAG - offset];
479 	for (i = 0; i < info->mnest - offset; i++)
480 		nr_masks *= info->mag[TOPOLOGY_NR_MAG - offset - 1 - i];
481 	nr_masks = max(nr_masks, 1);
482 	for (i = 0; i < nr_masks; i++) {
483 		mask->next = memblock_virt_alloc(sizeof(*mask->next), 8);
484 		mask = mask->next;
485 	}
486 }
487 
488 void __init topology_init_early(void)
489 {
490 	struct sysinfo_15_1_x *info;
491 
492 	set_sched_topology(s390_topology);
493 	if (topology_mode == TOPOLOGY_MODE_UNINITIALIZED) {
494 		if (MACHINE_HAS_TOPOLOGY)
495 			topology_mode = TOPOLOGY_MODE_HW;
496 		else
497 			topology_mode = TOPOLOGY_MODE_SINGLE;
498 	}
499 	if (!MACHINE_HAS_TOPOLOGY)
500 		goto out;
501 	tl_info = memblock_virt_alloc(PAGE_SIZE, PAGE_SIZE);
502 	info = tl_info;
503 	store_topology(info);
504 	pr_info("The CPU configuration topology of the machine is: %d %d %d %d %d %d / %d\n",
505 		info->mag[0], info->mag[1], info->mag[2], info->mag[3],
506 		info->mag[4], info->mag[5], info->mnest);
507 	alloc_masks(info, &socket_info, 1);
508 	alloc_masks(info, &book_info, 2);
509 	alloc_masks(info, &drawer_info, 3);
510 out:
511 	__arch_update_cpu_topology();
512 }
513 
514 static inline int topology_get_mode(int enabled)
515 {
516 	if (!enabled)
517 		return TOPOLOGY_MODE_SINGLE;
518 	return MACHINE_HAS_TOPOLOGY ? TOPOLOGY_MODE_HW : TOPOLOGY_MODE_PACKAGE;
519 }
520 
521 static inline int topology_is_enabled(void)
522 {
523 	return topology_mode != TOPOLOGY_MODE_SINGLE;
524 }
525 
526 static int __init topology_setup(char *str)
527 {
528 	bool enabled;
529 	int rc;
530 
531 	rc = kstrtobool(str, &enabled);
532 	if (rc)
533 		return rc;
534 	topology_mode = topology_get_mode(enabled);
535 	return 0;
536 }
537 early_param("topology", topology_setup);
538 
539 static int topology_ctl_handler(struct ctl_table *ctl, int write,
540 				void __user *buffer, size_t *lenp, loff_t *ppos)
541 {
542 	unsigned int len;
543 	int new_mode;
544 	char buf[2];
545 
546 	if (!*lenp || *ppos) {
547 		*lenp = 0;
548 		return 0;
549 	}
550 	if (!write) {
551 		strncpy(buf, topology_is_enabled() ? "1\n" : "0\n",
552 			ARRAY_SIZE(buf));
553 		len = strnlen(buf, ARRAY_SIZE(buf));
554 		if (len > *lenp)
555 			len = *lenp;
556 		if (copy_to_user(buffer, buf, len))
557 			return -EFAULT;
558 		goto out;
559 	}
560 	len = *lenp;
561 	if (copy_from_user(buf, buffer, len > sizeof(buf) ? sizeof(buf) : len))
562 		return -EFAULT;
563 	if (buf[0] != '0' && buf[0] != '1')
564 		return -EINVAL;
565 	mutex_lock(&smp_cpu_state_mutex);
566 	new_mode = topology_get_mode(buf[0] == '1');
567 	if (topology_mode != new_mode) {
568 		topology_mode = new_mode;
569 		topology_schedule_update();
570 	}
571 	mutex_unlock(&smp_cpu_state_mutex);
572 	topology_flush_work();
573 out:
574 	*lenp = len;
575 	*ppos += len;
576 	return 0;
577 }
578 
579 static struct ctl_table topology_ctl_table[] = {
580 	{
581 		.procname	= "topology",
582 		.mode		= 0644,
583 		.proc_handler	= topology_ctl_handler,
584 	},
585 	{ },
586 };
587 
588 static struct ctl_table topology_dir_table[] = {
589 	{
590 		.procname	= "s390",
591 		.maxlen		= 0,
592 		.mode		= 0555,
593 		.child		= topology_ctl_table,
594 	},
595 	{ },
596 };
597 
598 static int __init topology_init(void)
599 {
600 	if (MACHINE_HAS_TOPOLOGY)
601 		set_topology_timer();
602 	else
603 		topology_update_polarization_simple();
604 	register_sysctl_table(topology_dir_table);
605 	return device_create_file(cpu_subsys.dev_root, &dev_attr_dispatching);
606 }
607 device_initcall(topology_init);
608