Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / arch / arm / common / mcpm_entry.c
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 extern unsigned long mcpm_entry_vectors[MAX_NR_CLUSTERS][MAX_CPUS_PER_CLUSTER];
24
25 void mcpm_set_entry_vector(unsigned cpu, unsigned cluster, void *ptr)
26 {
27         unsigned long val = ptr ? virt_to_phys(ptr) : 0;
28         mcpm_entry_vectors[cluster][cpu] = val;
29         sync_cache_w(&mcpm_entry_vectors[cluster][cpu]);
30 }
31
32 extern unsigned long mcpm_entry_early_pokes[MAX_NR_CLUSTERS][MAX_CPUS_PER_CLUSTER][2];
33
34 void mcpm_set_early_poke(unsigned cpu, unsigned cluster,
35                          unsigned long poke_phys_addr, unsigned long poke_val)
36 {
37         unsigned long *poke = &mcpm_entry_early_pokes[cluster][cpu][0];
38         poke[0] = poke_phys_addr;
39         poke[1] = poke_val;
40         __sync_cache_range_w(poke, 2 * sizeof(*poke));
41 }
42
43 static const struct mcpm_platform_ops *platform_ops;
44
45 int __init mcpm_platform_register(const struct mcpm_platform_ops *ops)
46 {
47         if (platform_ops)
48                 return -EBUSY;
49         platform_ops = ops;
50         return 0;
51 }
52
53 bool mcpm_is_available(void)
54 {
55         return (platform_ops) ? true : false;
56 }
57
58 /*
59  * We can't use regular spinlocks. In the switcher case, it is possible
60  * for an outbound CPU to call power_down() after its inbound counterpart
61  * is already live using the same logical CPU number which trips lockdep
62  * debugging.
63  */
64 static arch_spinlock_t mcpm_lock = __ARCH_SPIN_LOCK_UNLOCKED;
65
66 static int mcpm_cpu_use_count[MAX_NR_CLUSTERS][MAX_CPUS_PER_CLUSTER];
67
68 static inline bool mcpm_cluster_unused(unsigned int cluster)
69 {
70         int i, cnt;
71         for (i = 0, cnt = 0; i < MAX_CPUS_PER_CLUSTER; i++)
72                 cnt |= mcpm_cpu_use_count[cluster][i];
73         return !cnt;
74 }
75
76 int mcpm_cpu_power_up(unsigned int cpu, unsigned int cluster)
77 {
78         bool cpu_is_down, cluster_is_down;
79         int ret = 0;
80
81         if (!platform_ops)
82                 return -EUNATCH; /* try not to shadow power_up errors */
83         might_sleep();
84
85         /* backward compatibility callback */
86         if (platform_ops->power_up)
87                 return platform_ops->power_up(cpu, cluster);
88
89         pr_debug("%s: cpu %u cluster %u\n", __func__, cpu, cluster);
90
91         /*
92          * Since this is called with IRQs enabled, and no arch_spin_lock_irq
93          * variant exists, we need to disable IRQs manually here.
94          */
95         local_irq_disable();
96         arch_spin_lock(&mcpm_lock);
97
98         cpu_is_down = !mcpm_cpu_use_count[cluster][cpu];
99         cluster_is_down = mcpm_cluster_unused(cluster);
100
101         mcpm_cpu_use_count[cluster][cpu]++;
102         /*
103          * The only possible values are:
104          * 0 = CPU down
105          * 1 = CPU (still) up
106          * 2 = CPU requested to be up before it had a chance
107          *     to actually make itself down.
108          * Any other value is a bug.
109          */
110         BUG_ON(mcpm_cpu_use_count[cluster][cpu] != 1 &&
111                mcpm_cpu_use_count[cluster][cpu] != 2);
112
113         if (cluster_is_down)
114                 ret = platform_ops->cluster_powerup(cluster);
115         if (cpu_is_down && !ret)
116                 ret = platform_ops->cpu_powerup(cpu, cluster);
117
118         arch_spin_unlock(&mcpm_lock);
119         local_irq_enable();
120         return ret;
121 }
122
123 typedef void (*phys_reset_t)(unsigned long);
124
125 void mcpm_cpu_power_down(void)
126 {
127         unsigned int mpidr, cpu, cluster;
128         bool cpu_going_down, last_man;
129         phys_reset_t phys_reset;
130
131         if (WARN_ON_ONCE(!platform_ops))
132                return;
133         BUG_ON(!irqs_disabled());
134
135         /*
136          * Do this before calling into the power_down method,
137          * as it might not always be safe to do afterwards.
138          */
139         setup_mm_for_reboot();
140
141         /* backward compatibility callback */
142         if (platform_ops->power_down) {
143                 platform_ops->power_down();
144                 goto not_dead;
145         }
146
147         mpidr = read_cpuid_mpidr();
148         cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
149         cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
150         pr_debug("%s: cpu %u cluster %u\n", __func__, cpu, cluster);
151
152         __mcpm_cpu_going_down(cpu, cluster);
153
154         arch_spin_lock(&mcpm_lock);
155         BUG_ON(__mcpm_cluster_state(cluster) != CLUSTER_UP);
156
157         mcpm_cpu_use_count[cluster][cpu]--;
158         BUG_ON(mcpm_cpu_use_count[cluster][cpu] != 0 &&
159                mcpm_cpu_use_count[cluster][cpu] != 1);
160         cpu_going_down = !mcpm_cpu_use_count[cluster][cpu];
161         last_man = mcpm_cluster_unused(cluster);
162
163         if (last_man && __mcpm_outbound_enter_critical(cpu, cluster)) {
164                 platform_ops->cpu_powerdown_prepare(cpu, cluster);
165                 platform_ops->cluster_powerdown_prepare(cluster);
166                 arch_spin_unlock(&mcpm_lock);
167                 platform_ops->cluster_cache_disable();
168                 __mcpm_outbound_leave_critical(cluster, CLUSTER_DOWN);
169         } else {
170                 if (cpu_going_down)
171                         platform_ops->cpu_powerdown_prepare(cpu, cluster);
172                 arch_spin_unlock(&mcpm_lock);
173                 /*
174                  * If cpu_going_down is false here, that means a power_up
175                  * request raced ahead of us.  Even if we do not want to
176                  * shut this CPU down, the caller still expects execution
177                  * to return through the system resume entry path, like
178                  * when the WFI is aborted due to a new IRQ or the like..
179                  * So let's continue with cache cleaning in all cases.
180                  */
181                 platform_ops->cpu_cache_disable();
182         }
183
184         __mcpm_cpu_down(cpu, cluster);
185
186         /* Now we are prepared for power-down, do it: */
187         if (cpu_going_down)
188                 wfi();
189
190 not_dead:
191         /*
192          * It is possible for a power_up request to happen concurrently
193          * with a power_down request for the same CPU. In this case the
194          * CPU might not be able to actually enter a powered down state
195          * with the WFI instruction if the power_up request has removed
196          * the required reset condition.  We must perform a re-entry in
197          * the kernel as if the power_up method just had deasserted reset
198          * on the CPU.
199          */
200         phys_reset = (phys_reset_t)(unsigned long)virt_to_phys(cpu_reset);
201         phys_reset(virt_to_phys(mcpm_entry_point));
202
203         /* should never get here */
204         BUG();
205 }
206
207 int mcpm_wait_for_cpu_powerdown(unsigned int cpu, unsigned int cluster)
208 {
209         int ret;
210
211         if (WARN_ON_ONCE(!platform_ops || !platform_ops->wait_for_powerdown))
212                 return -EUNATCH;
213
214         ret = platform_ops->wait_for_powerdown(cpu, cluster);
215         if (ret)
216                 pr_warn("%s: cpu %u, cluster %u failed to power down (%d)\n",
217                         __func__, cpu, cluster, ret);
218
219         return ret;
220 }
221
222 void mcpm_cpu_suspend(u64 expected_residency)
223 {
224         if (WARN_ON_ONCE(!platform_ops))
225                 return;
226
227         /* backward compatibility callback */
228         if (platform_ops->suspend) {
229                 phys_reset_t phys_reset;
230                 BUG_ON(!irqs_disabled());
231                 setup_mm_for_reboot();
232                 platform_ops->suspend(expected_residency);
233                 phys_reset = (phys_reset_t)(unsigned long)virt_to_phys(cpu_reset);
234                 phys_reset(virt_to_phys(mcpm_entry_point));
235                 BUG();
236         }
237
238         /* Some platforms might have to enable special resume modes, etc. */
239         if (platform_ops->cpu_suspend_prepare) {
240                 unsigned int mpidr = read_cpuid_mpidr();
241                 unsigned int cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
242                 unsigned int cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1); 
243                 arch_spin_lock(&mcpm_lock);
244                 platform_ops->cpu_suspend_prepare(cpu, cluster);
245                 arch_spin_unlock(&mcpm_lock);
246         }
247         mcpm_cpu_power_down();
248 }
249
250 int mcpm_cpu_powered_up(void)
251 {
252         unsigned int mpidr, cpu, cluster;
253         bool cpu_was_down, first_man;
254         unsigned long flags;
255
256         if (!platform_ops)
257                 return -EUNATCH;
258
259         /* backward compatibility callback */
260         if (platform_ops->powered_up) {
261                 platform_ops->powered_up();
262                 return 0;
263         }
264
265         mpidr = read_cpuid_mpidr();
266         cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
267         cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
268         local_irq_save(flags);
269         arch_spin_lock(&mcpm_lock);
270
271         cpu_was_down = !mcpm_cpu_use_count[cluster][cpu];
272         first_man = mcpm_cluster_unused(cluster);
273
274         if (first_man && platform_ops->cluster_is_up)
275                 platform_ops->cluster_is_up(cluster);
276         if (cpu_was_down)
277                 mcpm_cpu_use_count[cluster][cpu] = 1;
278         if (platform_ops->cpu_is_up)
279                 platform_ops->cpu_is_up(cpu, cluster);
280
281         arch_spin_unlock(&mcpm_lock);
282         local_irq_restore(flags);
283
284         return 0;
285 }
286
287 #ifdef CONFIG_ARM_CPU_SUSPEND
288
289 static int __init nocache_trampoline(unsigned long _arg)
290 {
291         void (*cache_disable)(void) = (void *)_arg;
292         unsigned int mpidr = read_cpuid_mpidr();
293         unsigned int cpu = MPIDR_AFFINITY_LEVEL(mpidr, 0);
294         unsigned int cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
295         phys_reset_t phys_reset;
296
297         mcpm_set_entry_vector(cpu, cluster, cpu_resume);
298         setup_mm_for_reboot();
299
300         __mcpm_cpu_going_down(cpu, cluster);
301         BUG_ON(!__mcpm_outbound_enter_critical(cpu, cluster));
302         cache_disable();
303         __mcpm_outbound_leave_critical(cluster, CLUSTER_DOWN);
304         __mcpm_cpu_down(cpu, cluster);
305
306         phys_reset = (phys_reset_t)(unsigned long)virt_to_phys(cpu_reset);
307         phys_reset(virt_to_phys(mcpm_entry_point));
308         BUG();
309 }
310
311 int __init mcpm_loopback(void (*cache_disable)(void))
312 {
313         int ret;
314
315         /*
316          * We're going to soft-restart the current CPU through the
317          * low-level MCPM code by leveraging the suspend/resume
318          * infrastructure. Let's play it safe by using cpu_pm_enter()
319          * in case the CPU init code path resets the VFP or similar.
320          */
321         local_irq_disable();
322         local_fiq_disable();
323         ret = cpu_pm_enter();
324         if (!ret) {
325                 ret = cpu_suspend((unsigned long)cache_disable, nocache_trampoline);
326                 cpu_pm_exit();
327         }
328         local_fiq_enable();
329         local_irq_enable();
330         if (ret)
331                 pr_err("%s returned %d\n", __func__, ret);
332         return ret;
333 }
334
335 #endif
336
337 struct sync_struct mcpm_sync;
338
339 /*
340  * __mcpm_cpu_going_down: Indicates that the cpu is being torn down.
341  *    This must be called at the point of committing to teardown of a CPU.
342  *    The CPU cache (SCTRL.C bit) is expected to still be active.
343  */
344 void __mcpm_cpu_going_down(unsigned int cpu, unsigned int cluster)
345 {
346         mcpm_sync.clusters[cluster].cpus[cpu].cpu = CPU_GOING_DOWN;
347         sync_cache_w(&mcpm_sync.clusters[cluster].cpus[cpu].cpu);
348 }
349
350 /*
351  * __mcpm_cpu_down: Indicates that cpu teardown is complete and that the
352  *    cluster can be torn down without disrupting this CPU.
353  *    To avoid deadlocks, this must be called before a CPU is powered down.
354  *    The CPU cache (SCTRL.C bit) is expected to be off.
355  *    However L2 cache might or might not be active.
356  */
357 void __mcpm_cpu_down(unsigned int cpu, unsigned int cluster)
358 {
359         dmb();
360         mcpm_sync.clusters[cluster].cpus[cpu].cpu = CPU_DOWN;
361         sync_cache_w(&mcpm_sync.clusters[cluster].cpus[cpu].cpu);
362         sev();
363 }
364
365 /*
366  * __mcpm_outbound_leave_critical: Leave the cluster teardown critical section.
367  * @state: the final state of the cluster:
368  *     CLUSTER_UP: no destructive teardown was done and the cluster has been
369  *         restored to the previous state (CPU cache still active); or
370  *     CLUSTER_DOWN: the cluster has been torn-down, ready for power-off
371  *         (CPU cache disabled, L2 cache either enabled or disabled).
372  */
373 void __mcpm_outbound_leave_critical(unsigned int cluster, int state)
374 {
375         dmb();
376         mcpm_sync.clusters[cluster].cluster = state;
377         sync_cache_w(&mcpm_sync.clusters[cluster].cluster);
378         sev();
379 }
380
381 /*
382  * __mcpm_outbound_enter_critical: Enter the cluster teardown critical section.
383  * This function should be called by the last man, after local CPU teardown
384  * is complete.  CPU cache expected to be active.
385  *
386  * Returns:
387  *     false: the critical section was not entered because an inbound CPU was
388  *         observed, or the cluster is already being set up;
389  *     true: the critical section was entered: it is now safe to tear down the
390  *         cluster.
391  */
392 bool __mcpm_outbound_enter_critical(unsigned int cpu, unsigned int cluster)
393 {
394         unsigned int i;
395         struct mcpm_sync_struct *c = &mcpm_sync.clusters[cluster];
396
397         /* Warn inbound CPUs that the cluster is being torn down: */
398         c->cluster = CLUSTER_GOING_DOWN;
399         sync_cache_w(&c->cluster);
400
401         /* Back out if the inbound cluster is already in the critical region: */
402         sync_cache_r(&c->inbound);
403         if (c->inbound == INBOUND_COMING_UP)
404                 goto abort;
405
406         /*
407          * Wait for all CPUs to get out of the GOING_DOWN state, so that local
408          * teardown is complete on each CPU before tearing down the cluster.
409          *
410          * If any CPU has been woken up again from the DOWN state, then we
411          * shouldn't be taking the cluster down at all: abort in that case.
412          */
413         sync_cache_r(&c->cpus);
414         for (i = 0; i < MAX_CPUS_PER_CLUSTER; i++) {
415                 int cpustate;
416
417                 if (i == cpu)
418                         continue;
419
420                 while (1) {
421                         cpustate = c->cpus[i].cpu;
422                         if (cpustate != CPU_GOING_DOWN)
423                                 break;
424
425                         wfe();
426                         sync_cache_r(&c->cpus[i].cpu);
427                 }
428
429                 switch (cpustate) {
430                 case CPU_DOWN:
431                         continue;
432
433                 default:
434                         goto abort;
435                 }
436         }
437
438         return true;
439
440 abort:
441         __mcpm_outbound_leave_critical(cluster, CLUSTER_UP);
442         return false;
443 }
444
445 int __mcpm_cluster_state(unsigned int cluster)
446 {
447         sync_cache_r(&mcpm_sync.clusters[cluster].cluster);
448         return mcpm_sync.clusters[cluster].cluster;
449 }
450
451 extern unsigned long mcpm_power_up_setup_phys;
452
453 int __init mcpm_sync_init(
454         void (*power_up_setup)(unsigned int affinity_level))
455 {
456         unsigned int i, j, mpidr, this_cluster;
457
458         BUILD_BUG_ON(MCPM_SYNC_CLUSTER_SIZE * MAX_NR_CLUSTERS != sizeof mcpm_sync);
459         BUG_ON((unsigned long)&mcpm_sync & (__CACHE_WRITEBACK_GRANULE - 1));
460
461         /*
462          * Set initial CPU and cluster states.
463          * Only one cluster is assumed to be active at this point.
464          */
465         for (i = 0; i < MAX_NR_CLUSTERS; i++) {
466                 mcpm_sync.clusters[i].cluster = CLUSTER_DOWN;
467                 mcpm_sync.clusters[i].inbound = INBOUND_NOT_COMING_UP;
468                 for (j = 0; j < MAX_CPUS_PER_CLUSTER; j++)
469                         mcpm_sync.clusters[i].cpus[j].cpu = CPU_DOWN;
470         }
471         mpidr = read_cpuid_mpidr();
472         this_cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
473         for_each_online_cpu(i) {
474                 mcpm_cpu_use_count[this_cluster][i] = 1;
475                 mcpm_sync.clusters[this_cluster].cpus[i].cpu = CPU_UP;
476         }
477         mcpm_sync.clusters[this_cluster].cluster = CLUSTER_UP;
478         sync_cache_w(&mcpm_sync);
479
480         if (power_up_setup) {
481                 mcpm_power_up_setup_phys = virt_to_phys(power_up_setup);
482                 sync_cache_w(&mcpm_power_up_setup_phys);
483         }
484
485         return 0;
486 }