bb1c068bff19f3e9c24605257ea094221fc1ea75
[muen/linux.git] / drivers / firmware / psci_checker.c
1 /*
2  * This program is free software; you can redistribute it and/or modify
3  * it under the terms of the GNU General Public License version 2 as
4  * published by the Free Software Foundation.
5  *
6  * This program is distributed in the hope that it will be useful,
7  * but WITHOUT ANY WARRANTY; without even the implied warranty of
8  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
9  * GNU General Public License for more details.
10  *
11  * Copyright (C) 2016 ARM Limited
12  */
13
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
16 #include <linux/atomic.h>
17 #include <linux/completion.h>
18 #include <linux/cpu.h>
19 #include <linux/cpuidle.h>
20 #include <linux/cpu_pm.h>
21 #include <linux/kernel.h>
22 #include <linux/kthread.h>
23 #include <uapi/linux/sched/types.h>
24 #include <linux/module.h>
25 #include <linux/preempt.h>
26 #include <linux/psci.h>
27 #include <linux/slab.h>
28 #include <linux/tick.h>
29 #include <linux/topology.h>
30
31 #include <asm/cpuidle.h>
32
33 #include <uapi/linux/psci.h>
34
35 #define NUM_SUSPEND_CYCLE (10)
36
37 static unsigned int nb_available_cpus;
38 static int tos_resident_cpu = -1;
39
40 static atomic_t nb_active_threads;
41 static struct completion suspend_threads_started =
42         COMPLETION_INITIALIZER(suspend_threads_started);
43 static struct completion suspend_threads_done =
44         COMPLETION_INITIALIZER(suspend_threads_done);
45
46 /*
47  * We assume that PSCI operations are used if they are available. This is not
48  * necessarily true on arm64, since the decision is based on the
49  * "enable-method" property of each CPU in the DT, but given that there is no
50  * arch-specific way to check this, we assume that the DT is sensible.
51  */
52 static int psci_ops_check(void)
53 {
54         int migrate_type = -1;
55         int cpu;
56
57         if (!(psci_ops.cpu_off && psci_ops.cpu_on && psci_ops.cpu_suspend)) {
58                 pr_warn("Missing PSCI operations, aborting tests\n");
59                 return -EOPNOTSUPP;
60         }
61
62         if (psci_ops.migrate_info_type)
63                 migrate_type = psci_ops.migrate_info_type();
64
65         if (migrate_type == PSCI_0_2_TOS_UP_MIGRATE ||
66             migrate_type == PSCI_0_2_TOS_UP_NO_MIGRATE) {
67                 /* There is a UP Trusted OS, find on which core it resides. */
68                 for_each_online_cpu(cpu)
69                         if (psci_tos_resident_on(cpu)) {
70                                 tos_resident_cpu = cpu;
71                                 break;
72                         }
73                 if (tos_resident_cpu == -1)
74                         pr_warn("UP Trusted OS resides on no online CPU\n");
75         }
76
77         return 0;
78 }
79
80 static int find_cpu_groups(const struct cpumask *cpus,
81                            const struct cpumask **cpu_groups)
82 {
83         unsigned int nb = 0;
84         cpumask_var_t tmp;
85
86         if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
87                 return -ENOMEM;
88         cpumask_copy(tmp, cpus);
89
90         while (!cpumask_empty(tmp)) {
91                 const struct cpumask *cpu_group =
92                         topology_core_cpumask(cpumask_any(tmp));
93
94                 cpu_groups[nb++] = cpu_group;
95                 cpumask_andnot(tmp, tmp, cpu_group);
96         }
97
98         free_cpumask_var(tmp);
99         return nb;
100 }
101
102 /*
103  * offlined_cpus is a temporary array but passing it as an argument avoids
104  * multiple allocations.
105  */
106 static unsigned int down_and_up_cpus(const struct cpumask *cpus,
107                                      struct cpumask *offlined_cpus)
108 {
109         int cpu;
110         int err = 0;
111
112         cpumask_clear(offlined_cpus);
113
114         /* Try to power down all CPUs in the mask. */
115         for_each_cpu(cpu, cpus) {
116                 int ret = cpu_down(cpu);
117
118                 /*
119                  * cpu_down() checks the number of online CPUs before the TOS
120                  * resident CPU.
121                  */
122                 if (cpumask_weight(offlined_cpus) + 1 == nb_available_cpus) {
123                         if (ret != -EBUSY) {
124                                 pr_err("Unexpected return code %d while trying "
125                                        "to power down last online CPU %d\n",
126                                        ret, cpu);
127                                 ++err;
128                         }
129                 } else if (cpu == tos_resident_cpu) {
130                         if (ret != -EPERM) {
131                                 pr_err("Unexpected return code %d while trying "
132                                        "to power down TOS resident CPU %d\n",
133                                        ret, cpu);
134                                 ++err;
135                         }
136                 } else if (ret != 0) {
137                         pr_err("Error occurred (%d) while trying "
138                                "to power down CPU %d\n", ret, cpu);
139                         ++err;
140                 }
141
142                 if (ret == 0)
143                         cpumask_set_cpu(cpu, offlined_cpus);
144         }
145
146         /* Try to power up all the CPUs that have been offlined. */
147         for_each_cpu(cpu, offlined_cpus) {
148                 int ret = cpu_up(cpu);
149
150                 if (ret != 0) {
151                         pr_err("Error occurred (%d) while trying "
152                                "to power up CPU %d\n", ret, cpu);
153                         ++err;
154                 } else {
155                         cpumask_clear_cpu(cpu, offlined_cpus);
156                 }
157         }
158
159         /*
160          * Something went bad at some point and some CPUs could not be turned
161          * back on.
162          */
163         WARN_ON(!cpumask_empty(offlined_cpus) ||
164                 num_online_cpus() != nb_available_cpus);
165
166         return err;
167 }
168
169 static int hotplug_tests(void)
170 {
171         int err;
172         cpumask_var_t offlined_cpus;
173         int i, nb_cpu_group;
174         const struct cpumask **cpu_groups;
175         char *page_buf;
176
177         err = -ENOMEM;
178         if (!alloc_cpumask_var(&offlined_cpus, GFP_KERNEL))
179                 return err;
180         /* We may have up to nb_available_cpus cpu_groups. */
181         cpu_groups = kmalloc_array(nb_available_cpus, sizeof(*cpu_groups),
182                                    GFP_KERNEL);
183         if (!cpu_groups)
184                 goto out_free_cpus;
185         page_buf = (char *)__get_free_page(GFP_KERNEL);
186         if (!page_buf)
187                 goto out_free_cpu_groups;
188
189         err = 0;
190         nb_cpu_group = find_cpu_groups(cpu_online_mask, cpu_groups);
191
192         /*
193          * Of course the last CPU cannot be powered down and cpu_down() should
194          * refuse doing that.
195          */
196         pr_info("Trying to turn off and on again all CPUs\n");
197         err += down_and_up_cpus(cpu_online_mask, offlined_cpus);
198
199         /*
200          * Take down CPUs by cpu group this time. When the last CPU is turned
201          * off, the cpu group itself should shut down.
202          */
203         for (i = 0; i < nb_cpu_group; ++i) {
204                 ssize_t len = cpumap_print_to_pagebuf(true, page_buf,
205                                                       cpu_groups[i]);
206                 /* Remove trailing newline. */
207                 page_buf[len - 1] = '\0';
208                 pr_info("Trying to turn off and on again group %d (CPUs %s)\n",
209                         i, page_buf);
210                 err += down_and_up_cpus(cpu_groups[i], offlined_cpus);
211         }
212
213         free_page((unsigned long)page_buf);
214 out_free_cpu_groups:
215         kfree(cpu_groups);
216 out_free_cpus:
217         free_cpumask_var(offlined_cpus);
218         return err;
219 }
220
221 static void dummy_callback(struct timer_list *unused) {}
222
223 static int suspend_cpu(int index, bool broadcast)
224 {
225         int ret;
226
227         arch_cpu_idle_enter();
228
229         if (broadcast) {
230                 /*
231                  * The local timer will be shut down, we need to enter tick
232                  * broadcast.
233                  */
234                 ret = tick_broadcast_enter();
235                 if (ret) {
236                         /*
237                          * In the absence of hardware broadcast mechanism,
238                          * this CPU might be used to broadcast wakeups, which
239                          * may be why entering tick broadcast has failed.
240                          * There is little the kernel can do to work around
241                          * that, so enter WFI instead (idle state 0).
242                          */
243                         cpu_do_idle();
244                         ret = 0;
245                         goto out_arch_exit;
246                 }
247         }
248
249         /*
250          * Replicate the common ARM cpuidle enter function
251          * (arm_enter_idle_state).
252          */
253         ret = CPU_PM_CPU_IDLE_ENTER(arm_cpuidle_suspend, index);
254
255         if (broadcast)
256                 tick_broadcast_exit();
257
258 out_arch_exit:
259         arch_cpu_idle_exit();
260
261         return ret;
262 }
263
264 static int suspend_test_thread(void *arg)
265 {
266         int cpu = (long)arg;
267         int i, nb_suspend = 0, nb_shallow_sleep = 0, nb_err = 0;
268         struct sched_param sched_priority = { .sched_priority = MAX_RT_PRIO-1 };
269         struct cpuidle_device *dev;
270         struct cpuidle_driver *drv;
271         /* No need for an actual callback, we just want to wake up the CPU. */
272         struct timer_list wakeup_timer;
273
274         /* Wait for the main thread to give the start signal. */
275         wait_for_completion(&suspend_threads_started);
276
277         /* Set maximum priority to preempt all other threads on this CPU. */
278         if (sched_setscheduler_nocheck(current, SCHED_FIFO, &sched_priority))
279                 pr_warn("Failed to set suspend thread scheduler on CPU %d\n",
280                         cpu);
281
282         dev = this_cpu_read(cpuidle_devices);
283         drv = cpuidle_get_cpu_driver(dev);
284
285         pr_info("CPU %d entering suspend cycles, states 1 through %d\n",
286                 cpu, drv->state_count - 1);
287
288         timer_setup_on_stack(&wakeup_timer, dummy_callback, 0);
289         for (i = 0; i < NUM_SUSPEND_CYCLE; ++i) {
290                 int index;
291                 /*
292                  * Test all possible states, except 0 (which is usually WFI and
293                  * doesn't use PSCI).
294                  */
295                 for (index = 1; index < drv->state_count; ++index) {
296                         struct cpuidle_state *state = &drv->states[index];
297                         bool broadcast = state->flags & CPUIDLE_FLAG_TIMER_STOP;
298                         int ret;
299
300                         /*
301                          * Set the timer to wake this CPU up in some time (which
302                          * should be largely sufficient for entering suspend).
303                          * If the local tick is disabled when entering suspend,
304                          * suspend_cpu() takes care of switching to a broadcast
305                          * tick, so the timer will still wake us up.
306                          */
307                         mod_timer(&wakeup_timer, jiffies +
308                                   usecs_to_jiffies(state->target_residency));
309
310                         /* IRQs must be disabled during suspend operations. */
311                         local_irq_disable();
312
313                         ret = suspend_cpu(index, broadcast);
314
315                         /*
316                          * We have woken up. Re-enable IRQs to handle any
317                          * pending interrupt, do not wait until the end of the
318                          * loop.
319                          */
320                         local_irq_enable();
321
322                         if (ret == index) {
323                                 ++nb_suspend;
324                         } else if (ret >= 0) {
325                                 /* We did not enter the expected state. */
326                                 ++nb_shallow_sleep;
327                         } else {
328                                 pr_err("Failed to suspend CPU %d: error %d "
329                                        "(requested state %d, cycle %d)\n",
330                                        cpu, ret, index, i);
331                                 ++nb_err;
332                         }
333                 }
334         }
335
336         /*
337          * Disable the timer to make sure that the timer will not trigger
338          * later.
339          */
340         del_timer(&wakeup_timer);
341         destroy_timer_on_stack(&wakeup_timer);
342
343         if (atomic_dec_return_relaxed(&nb_active_threads) == 0)
344                 complete(&suspend_threads_done);
345
346         /* Give up on RT scheduling and wait for termination. */
347         sched_priority.sched_priority = 0;
348         if (sched_setscheduler_nocheck(current, SCHED_NORMAL, &sched_priority))
349                 pr_warn("Failed to set suspend thread scheduler on CPU %d\n",
350                         cpu);
351         for (;;) {
352                 /* Needs to be set first to avoid missing a wakeup. */
353                 set_current_state(TASK_INTERRUPTIBLE);
354                 if (kthread_should_stop()) {
355                         __set_current_state(TASK_RUNNING);
356                         break;
357                 }
358                 schedule();
359         }
360
361         pr_info("CPU %d suspend test results: success %d, shallow states %d, errors %d\n",
362                 cpu, nb_suspend, nb_shallow_sleep, nb_err);
363
364         return nb_err;
365 }
366
367 static int suspend_tests(void)
368 {
369         int i, cpu, err = 0;
370         struct task_struct **threads;
371         int nb_threads = 0;
372
373         threads = kmalloc_array(nb_available_cpus, sizeof(*threads),
374                                 GFP_KERNEL);
375         if (!threads)
376                 return -ENOMEM;
377
378         /*
379          * Stop cpuidle to prevent the idle tasks from entering a deep sleep
380          * mode, as it might interfere with the suspend threads on other CPUs.
381          * This does not prevent the suspend threads from using cpuidle (only
382          * the idle tasks check this status). Take the idle lock so that
383          * the cpuidle driver and device look-up can be carried out safely.
384          */
385         cpuidle_pause_and_lock();
386
387         for_each_online_cpu(cpu) {
388                 struct task_struct *thread;
389                 /* Check that cpuidle is available on that CPU. */
390                 struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
391                 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
392
393                 if (!dev || !drv) {
394                         pr_warn("cpuidle not available on CPU %d, ignoring\n",
395                                 cpu);
396                         continue;
397                 }
398
399                 thread = kthread_create_on_cpu(suspend_test_thread,
400                                                (void *)(long)cpu, cpu,
401                                                "psci_suspend_test");
402                 if (IS_ERR(thread))
403                         pr_err("Failed to create kthread on CPU %d\n", cpu);
404                 else
405                         threads[nb_threads++] = thread;
406         }
407
408         if (nb_threads < 1) {
409                 err = -ENODEV;
410                 goto out;
411         }
412
413         atomic_set(&nb_active_threads, nb_threads);
414
415         /*
416          * Wake up the suspend threads. To avoid the main thread being preempted
417          * before all the threads have been unparked, the suspend threads will
418          * wait for the completion of suspend_threads_started.
419          */
420         for (i = 0; i < nb_threads; ++i)
421                 wake_up_process(threads[i]);
422         complete_all(&suspend_threads_started);
423
424         wait_for_completion(&suspend_threads_done);
425
426
427         /* Stop and destroy all threads, get return status. */
428         for (i = 0; i < nb_threads; ++i)
429                 err += kthread_stop(threads[i]);
430  out:
431         cpuidle_resume_and_unlock();
432         kfree(threads);
433         return err;
434 }
435
436 static int __init psci_checker(void)
437 {
438         int ret;
439
440         /*
441          * Since we're in an initcall, we assume that all the CPUs that all
442          * CPUs that can be onlined have been onlined.
443          *
444          * The tests assume that hotplug is enabled but nobody else is using it,
445          * otherwise the results will be unpredictable. However, since there
446          * is no userspace yet in initcalls, that should be fine, as long as
447          * no torture test is running at the same time (see Kconfig).
448          */
449         nb_available_cpus = num_online_cpus();
450
451         /* Check PSCI operations are set up and working. */
452         ret = psci_ops_check();
453         if (ret)
454                 return ret;
455
456         pr_info("PSCI checker started using %u CPUs\n", nb_available_cpus);
457
458         pr_info("Starting hotplug tests\n");
459         ret = hotplug_tests();
460         if (ret == 0)
461                 pr_info("Hotplug tests passed OK\n");
462         else if (ret > 0)
463                 pr_err("%d error(s) encountered in hotplug tests\n", ret);
464         else {
465                 pr_err("Out of memory\n");
466                 return ret;
467         }
468
469         pr_info("Starting suspend tests (%d cycles per state)\n",
470                 NUM_SUSPEND_CYCLE);
471         ret = suspend_tests();
472         if (ret == 0)
473                 pr_info("Suspend tests passed OK\n");
474         else if (ret > 0)
475                 pr_err("%d error(s) encountered in suspend tests\n", ret);
476         else {
477                 switch (ret) {
478                 case -ENOMEM:
479                         pr_err("Out of memory\n");
480                         break;
481                 case -ENODEV:
482                         pr_warn("Could not start suspend tests on any CPU\n");
483                         break;
484                 }
485         }
486
487         pr_info("PSCI checker completed\n");
488         return ret < 0 ? ret : 0;
489 }
490 late_initcall(psci_checker);