xref: /linux/arch/arm/common/mcpm_entry.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
1 /*
2  * arch/arm/common/mcpm_entry.c -- entry point for multi-cluster PM
3  *
4  * Created by:  Nicolas Pitre, March 2012
5  * Copyright:   (C) 2012-2013  Linaro Limited
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License version 2 as
9  * published by the Free Software Foundation.
10  */
11 
12 #include <linux/kernel.h>
13 #include <linux/init.h>
14 #include <linux/irqflags.h>
15 #include <linux/cpu_pm.h>
16 
17 #include <asm/mcpm.h>
18 #include <asm/cacheflush.h>
19 #include <asm/idmap.h>
20 #include <asm/cputype.h>
21 #include <asm/suspend.h>
22 
23 /*
24  * The public API for this code is documented in arch/arm/include/asm/mcpm.h.
25  * For a comprehensive description of the main algorithm used here, please
26  * see Documentation/arm/cluster-pm-race-avoidance.txt.
27  */
28 
29 struct sync_struct mcpm_sync;
30 
31 /*
32  * __mcpm_cpu_going_down: Indicates that the cpu is being torn down.
33  *    This must be called at the point of committing to teardown of a CPU.
34  *    The CPU cache (SCTRL.C bit) is expected to still be active.
35  */
36 static void __mcpm_cpu_going_down(unsigned int cpu, unsigned int cluster)
37 {
38 	mcpm_sync.clusters[cluster].cpus[cpu].cpu = CPU_GOING_DOWN;
39 	sync_cache_w(&mcpm_sync.clusters[cluster].cpus[cpu].cpu);
40 }
41 
42 /*
43  * __mcpm_cpu_down: Indicates that cpu teardown is complete and that the
44  *    cluster can be torn down without disrupting this CPU.
45  *    To avoid deadlocks, this must be called before a CPU is powered down.
46  *    The CPU cache (SCTRL.C bit) is expected to be off.
47  *    However L2 cache might or might not be active.
48  */
49 static void __mcpm_cpu_down(unsigned int cpu, unsigned int cluster)
50 {
51 	dmb();
52 	mcpm_sync.clusters[cluster].cpus[cpu].cpu = CPU_DOWN;
53 	sync_cache_w(&mcpm_sync.clusters[cluster].cpus[cpu].cpu);
54 	sev();
55 }
56 
57 /*
58  * __mcpm_outbound_leave_critical: Leave the cluster teardown critical section.
59  * @state: the final state of the cluster:
60  *     CLUSTER_UP: no destructive teardown was done and the cluster has been
61  *         restored to the previous state (CPU cache still active); or
62  *     CLUSTER_DOWN: the cluster has been torn-down, ready for power-off
63  *         (CPU cache disabled, L2 cache either enabled or disabled).
64  */
65 static void __mcpm_outbound_leave_critical(unsigned int cluster, int state)
66 {
67 	dmb();
68 	mcpm_sync.clusters[cluster].cluster = state;
69 	sync_cache_w(&mcpm_sync.clusters[cluster].cluster);
70 	sev();
71 }
72 
73 /*
74  * __mcpm_outbound_enter_critical: Enter the cluster teardown critical section.
75  * This function should be called by the last man, after local CPU teardown
76  * is complete.  CPU cache expected to be active.
77  *
78  * Returns:
79  *     false: the critical section was not entered because an inbound CPU was
80  *         observed, or the cluster is already being set up;
81  *     true: the critical section was entered: it is now safe to tear down the
82  *         cluster.
83  */
84 static bool __mcpm_outbound_enter_critical(unsigned int cpu, unsigned int cluster)
85 {
86 	unsigned int i;
87 	struct mcpm_sync_struct *c = &mcpm_sync.clusters[cluster];
88 
89 	/* Warn inbound CPUs that the cluster is being torn down: */
90 	c->cluster = CLUSTER_GOING_DOWN;
91 	sync_cache_w(&c->cluster);
92 
93 	/* Back out if the inbound cluster is already in the critical region: */
94 	sync_cache_r(&c->inbound);
95 	if (c->inbound == INBOUND_COMING_UP)
96 		goto abort;
97 
98 	/*
99 	 * Wait for all CPUs to get out of the GOING_DOWN state, so that local
100 	 * teardown is complete on each CPU before tearing down the cluster.
101 	 *
102 	 * If any CPU has been woken up again from the DOWN state, then we
103 	 * shouldn't be taking the cluster down at all: abort in that case.
104 	 */
105 	sync_cache_r(&c->cpus);
106 	for (i = 0; i < MAX_CPUS_PER_CLUSTER; i++) {
107 		int cpustate;
108 
109 		if (i == cpu)
110 			continue;
111 
112 		while (1) {
113 			cpustate = c->cpus[i].cpu;
114 			if (cpustate != CPU_GOING_DOWN)
115 				break;
116 
117 			wfe();
118 			sync_cache_r(&c->cpus[i].cpu);
119 		}
120 
121 		switch (cpustate) {
122 		case CPU_DOWN:
123 			continue;
124 
125 		default:
126 			goto abort;
127 		}
128 	}
129 
130 	return true;
131 
132 abort:
133 	__mcpm_outbound_leave_critical(cluster, CLUSTER_UP);
134 	return false;
135 }
136 
137 static int __mcpm_cluster_state(unsigned int cluster)
138 {
139 	sync_cache_r(&mcpm_sync.clusters[cluster].cluster);
140 	return mcpm_sync.clusters[cluster].cluster;
141 }
142 
143 extern unsigned long mcpm_entry_vectors[MAX_NR_CLUSTERS][MAX_CPUS_PER_CLUSTER];
144 
145 void mcpm_set_entry_vector(unsigned cpu, unsigned cluster, void *ptr)
146 {
147 	unsigned long val = ptr ? virt_to_phys(ptr) : 0;
148 	mcpm_entry_vectors[cluster][cpu] = val;
149 	sync_cache_w(&mcpm_entry_vectors[cluster][cpu]);
150 }
151 
152 extern unsigned long mcpm_entry_early_pokes[MAX_NR_CLUSTERS][MAX_CPUS_PER_CLUSTER][2];
153 
154 void mcpm_set_early_poke(unsigned cpu, unsigned cluster,
155 			 unsigned long poke_phys_addr, unsigned long poke_val)
156 {
157 	unsigned long *poke = &mcpm_entry_early_pokes[cluster][cpu][0];
158 	poke[0] = poke_phys_addr;
159 	poke[1] = poke_val;
160 	__sync_cache_range_w(poke, 2 * sizeof(*poke));
161 }
162 
163 static const struct mcpm_platform_ops *platform_ops;
164 
165 int __init mcpm_platform_register(const struct mcpm_platform_ops *ops)
166 {
167 	if (platform_ops)
168 		return -EBUSY;
169 	platform_ops = ops;
170 	return 0;
171 }
172 
173 bool mcpm_is_available(void)
174 {
175 	return (platform_ops) ? true : false;
176 }
177 
178 /*
179  * We can't use regular spinlocks. In the switcher case, it is possible
180  * for an outbound CPU to call power_down() after its inbound counterpart
181  * is already live using the same logical CPU number which trips lockdep
182  * debugging.
183  */
184 static arch_spinlock_t mcpm_lock = __ARCH_SPIN_LOCK_UNLOCKED;
185 
186 static int mcpm_cpu_use_count[MAX_NR_CLUSTERS][MAX_CPUS_PER_CLUSTER];
187 
188 static inline bool mcpm_cluster_unused(unsigned int cluster)
189 {
190 	int i, cnt;
191 	for (i = 0, cnt = 0; i < MAX_CPUS_PER_CLUSTER; i++)
192 		cnt |= mcpm_cpu_use_count[cluster][i];
193 	return !cnt;
194 }
195 
196 int mcpm_cpu_power_up(unsigned int cpu, unsigned int cluster)
197 {
198 	bool cpu_is_down, cluster_is_down;
199 	int ret = 0;
200 
201 	pr_debug("%s: cpu %u cluster %u\n", __func__, cpu, cluster);
202 	if (!platform_ops)
203 		return -EUNATCH; /* try not to shadow power_up errors */
204 	might_sleep();
205 
206 	/*
207 	 * Since this is called with IRQs enabled, and no arch_spin_lock_irq
208 	 * variant exists, we need to disable IRQs manually here.
209 	 */
210 	local_irq_disable();
211 	arch_spin_lock(&mcpm_lock);
212 
213 	cpu_is_down = !mcpm_cpu_use_count[cluster][cpu];
214 	cluster_is_down = mcpm_cluster_unused(cluster);
215 
216 	mcpm_cpu_use_count[cluster][cpu]++;
217 	/*
218 	 * The only possible values are:
219 	 * 0 = CPU down
220 	 * 1 = CPU (still) up
221 	 * 2 = CPU requested to be up before it had a chance
222 	 *     to actually make itself down.
223 	 * Any other value is a bug.
224 	 */
225 	BUG_ON(mcpm_cpu_use_count[cluster][cpu] != 1 &&
226 	       mcpm_cpu_use_count[cluster][cpu] != 2);
227 
228 	if (cluster_is_down)
229 		ret = platform_ops->cluster_powerup(cluster);
230 	if (cpu_is_down && !ret)
231 		ret = platform_ops->cpu_powerup(cpu, cluster);
232 
233 	arch_spin_unlock(&mcpm_lock);
234 	local_irq_enable();
235 	return ret;
236 }
237 
238 typedef void (*phys_reset_t)(unsigned long);
239 
240 void mcpm_cpu_power_down(void)
241 {
242 	unsigned int mpidr, cpu, cluster;
243 	bool cpu_going_down, last_man;
244 	phys_reset_t phys_reset;
245 
246 	mpidr = read_cpuid_mpidr();
247 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
248 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
249 	pr_debug("%s: cpu %u cluster %u\n", __func__, cpu, cluster);
250 	if (WARN_ON_ONCE(!platform_ops))
251 	       return;
252 	BUG_ON(!irqs_disabled());
253 
254 	setup_mm_for_reboot();
255 
256 	__mcpm_cpu_going_down(cpu, cluster);
257 	arch_spin_lock(&mcpm_lock);
258 	BUG_ON(__mcpm_cluster_state(cluster) != CLUSTER_UP);
259 
260 	mcpm_cpu_use_count[cluster][cpu]--;
261 	BUG_ON(mcpm_cpu_use_count[cluster][cpu] != 0 &&
262 	       mcpm_cpu_use_count[cluster][cpu] != 1);
263 	cpu_going_down = !mcpm_cpu_use_count[cluster][cpu];
264 	last_man = mcpm_cluster_unused(cluster);
265 
266 	if (last_man && __mcpm_outbound_enter_critical(cpu, cluster)) {
267 		platform_ops->cpu_powerdown_prepare(cpu, cluster);
268 		platform_ops->cluster_powerdown_prepare(cluster);
269 		arch_spin_unlock(&mcpm_lock);
270 		platform_ops->cluster_cache_disable();
271 		__mcpm_outbound_leave_critical(cluster, CLUSTER_DOWN);
272 	} else {
273 		if (cpu_going_down)
274 			platform_ops->cpu_powerdown_prepare(cpu, cluster);
275 		arch_spin_unlock(&mcpm_lock);
276 		/*
277 		 * If cpu_going_down is false here, that means a power_up
278 		 * request raced ahead of us.  Even if we do not want to
279 		 * shut this CPU down, the caller still expects execution
280 		 * to return through the system resume entry path, like
281 		 * when the WFI is aborted due to a new IRQ or the like..
282 		 * So let's continue with cache cleaning in all cases.
283 		 */
284 		platform_ops->cpu_cache_disable();
285 	}
286 
287 	__mcpm_cpu_down(cpu, cluster);
288 
289 	/* Now we are prepared for power-down, do it: */
290 	if (cpu_going_down)
291 		wfi();
292 
293 	/*
294 	 * It is possible for a power_up request to happen concurrently
295 	 * with a power_down request for the same CPU. In this case the
296 	 * CPU might not be able to actually enter a powered down state
297 	 * with the WFI instruction if the power_up request has removed
298 	 * the required reset condition.  We must perform a re-entry in
299 	 * the kernel as if the power_up method just had deasserted reset
300 	 * on the CPU.
301 	 */
302 	phys_reset = (phys_reset_t)(unsigned long)virt_to_phys(cpu_reset);
303 	phys_reset(virt_to_phys(mcpm_entry_point));
304 
305 	/* should never get here */
306 	BUG();
307 }
308 
309 int mcpm_wait_for_cpu_powerdown(unsigned int cpu, unsigned int cluster)
310 {
311 	int ret;
312 
313 	if (WARN_ON_ONCE(!platform_ops || !platform_ops->wait_for_powerdown))
314 		return -EUNATCH;
315 
316 	ret = platform_ops->wait_for_powerdown(cpu, cluster);
317 	if (ret)
318 		pr_warn("%s: cpu %u, cluster %u failed to power down (%d)\n",
319 			__func__, cpu, cluster, ret);
320 
321 	return ret;
322 }
323 
324 void mcpm_cpu_suspend(void)
325 {
326 	if (WARN_ON_ONCE(!platform_ops))
327 		return;
328 
329 	/* Some platforms might have to enable special resume modes, etc. */
330 	if (platform_ops->cpu_suspend_prepare) {
331 		unsigned int mpidr = read_cpuid_mpidr();
332 		unsigned int cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
333 		unsigned int cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
334 		arch_spin_lock(&mcpm_lock);
335 		platform_ops->cpu_suspend_prepare(cpu, cluster);
336 		arch_spin_unlock(&mcpm_lock);
337 	}
338 	mcpm_cpu_power_down();
339 }
340 
341 int mcpm_cpu_powered_up(void)
342 {
343 	unsigned int mpidr, cpu, cluster;
344 	bool cpu_was_down, first_man;
345 	unsigned long flags;
346 
347 	if (!platform_ops)
348 		return -EUNATCH;
349 
350 	mpidr = read_cpuid_mpidr();
351 	cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
352 	cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
353 	local_irq_save(flags);
354 	arch_spin_lock(&mcpm_lock);
355 
356 	cpu_was_down = !mcpm_cpu_use_count[cluster][cpu];
357 	first_man = mcpm_cluster_unused(cluster);
358 
359 	if (first_man && platform_ops->cluster_is_up)
360 		platform_ops->cluster_is_up(cluster);
361 	if (cpu_was_down)
362 		mcpm_cpu_use_count[cluster][cpu] = 1;
363 	if (platform_ops->cpu_is_up)
364 		platform_ops->cpu_is_up(cpu, cluster);
365 
366 	arch_spin_unlock(&mcpm_lock);
367 	local_irq_restore(flags);
368 
369 	return 0;
370 }
371 
372 #ifdef CONFIG_ARM_CPU_SUSPEND
373 
374 static int __init nocache_trampoline(unsigned long _arg)
375 {
376 	void (*cache_disable)(void) = (void *)_arg;
377 	unsigned int mpidr = read_cpuid_mpidr();
378 	unsigned int cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
379 	unsigned int cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
380 	phys_reset_t phys_reset;
381 
382 	mcpm_set_entry_vector(cpu, cluster, cpu_resume);
383 	setup_mm_for_reboot();
384 
385 	__mcpm_cpu_going_down(cpu, cluster);
386 	BUG_ON(!__mcpm_outbound_enter_critical(cpu, cluster));
387 	cache_disable();
388 	__mcpm_outbound_leave_critical(cluster, CLUSTER_DOWN);
389 	__mcpm_cpu_down(cpu, cluster);
390 
391 	phys_reset = (phys_reset_t)(unsigned long)virt_to_phys(cpu_reset);
392 	phys_reset(virt_to_phys(mcpm_entry_point));
393 	BUG();
394 }
395 
396 int __init mcpm_loopback(void (*cache_disable)(void))
397 {
398 	int ret;
399 
400 	/*
401 	 * We're going to soft-restart the current CPU through the
402 	 * low-level MCPM code by leveraging the suspend/resume
403 	 * infrastructure. Let's play it safe by using cpu_pm_enter()
404 	 * in case the CPU init code path resets the VFP or similar.
405 	 */
406 	local_irq_disable();
407 	local_fiq_disable();
408 	ret = cpu_pm_enter();
409 	if (!ret) {
410 		ret = cpu_suspend((unsigned long)cache_disable, nocache_trampoline);
411 		cpu_pm_exit();
412 	}
413 	local_fiq_enable();
414 	local_irq_enable();
415 	if (ret)
416 		pr_err("%s returned %d\n", __func__, ret);
417 	return ret;
418 }
419 
420 #endif
421 
422 extern unsigned long mcpm_power_up_setup_phys;
423 
424 int __init mcpm_sync_init(
425 	void (*power_up_setup)(unsigned int affinity_level))
426 {
427 	unsigned int i, j, mpidr, this_cluster;
428 
429 	BUILD_BUG_ON(MCPM_SYNC_CLUSTER_SIZE * MAX_NR_CLUSTERS != sizeof mcpm_sync);
430 	BUG_ON((unsigned long)&mcpm_sync & (__CACHE_WRITEBACK_GRANULE - 1));
431 
432 	/*
433 	 * Set initial CPU and cluster states.
434 	 * Only one cluster is assumed to be active at this point.
435 	 */
436 	for (i = 0; i < MAX_NR_CLUSTERS; i++) {
437 		mcpm_sync.clusters[i].cluster = CLUSTER_DOWN;
438 		mcpm_sync.clusters[i].inbound = INBOUND_NOT_COMING_UP;
439 		for (j = 0; j < MAX_CPUS_PER_CLUSTER; j++)
440 			mcpm_sync.clusters[i].cpus[j].cpu = CPU_DOWN;
441 	}
442 	mpidr = read_cpuid_mpidr();
443 	this_cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
444 	for_each_online_cpu(i) {
445 		mcpm_cpu_use_count[this_cluster][i] = 1;
446 		mcpm_sync.clusters[this_cluster].cpus[i].cpu = CPU_UP;
447 	}
448 	mcpm_sync.clusters[this_cluster].cluster = CLUSTER_UP;
449 	sync_cache_w(&mcpm_sync);
450 
451 	if (power_up_setup) {
452 		mcpm_power_up_setup_phys = virt_to_phys(power_up_setup);
453 		sync_cache_w(&mcpm_power_up_setup_phys);
454 	}
455 
456 	return 0;
457 }
458