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 /*
  81  * offlined_cpus is a temporary array but passing it as an argument avoids
  82  * multiple allocations.
  83  */
  84 static unsigned int down_and_up_cpus(const struct cpumask *cpus,
  85                                      struct cpumask *offlined_cpus)
  86 {
  87         int cpu;
  88         int err = 0;
  89
  90         cpumask_clear(offlined_cpus);
  91
  92         /* Try to power down all CPUs in the mask. */
  93         for_each_cpu(cpu, cpus) {
  94                 int ret = cpu_down(cpu);
  95
  96                 /*
  97                  * cpu_down() checks the number of online CPUs before the TOS
  98                  * resident CPU.
  99                  */
 100                 if (cpumask_weight(offlined_cpus) + 1 == nb_available_cpus) {
 101                         if (ret != -EBUSY) {
 102                                 pr_err("Unexpected return code %d while trying "
 103                                        "to power down last online CPU %d\n",
 104                                        ret, cpu);
 105                                 ++err;
 106                         }
 107                 } else if (cpu == tos_resident_cpu) {
 108                         if (ret != -EPERM) {
 109                                 pr_err("Unexpected return code %d while trying "
 110                                        "to power down TOS resident CPU %d\n",
 111                                        ret, cpu);
 112                                 ++err;
 113                         }
 114                 } else if (ret != 0) {
 115                         pr_err("Error occurred (%d) while trying "
 116                                "to power down CPU %d\n", ret, cpu);
 117                         ++err;
 118                 }
 119
 120                 if (ret == 0)
 121                         cpumask_set_cpu(cpu, offlined_cpus);
 122         }
 123
 124         /* Try to power up all the CPUs that have been offlined. */
 125         for_each_cpu(cpu, offlined_cpus) {
 126                 int ret = cpu_up(cpu);
 127
 128                 if (ret != 0) {
 129                         pr_err("Error occurred (%d) while trying "
 130                                "to power up CPU %d\n", ret, cpu);
 131                         ++err;
 132                 } else {
 133                         cpumask_clear_cpu(cpu, offlined_cpus);
 134                 }
 135         }
 136
 137         /*
 138          * Something went bad at some point and some CPUs could not be turned
 139          * back on.
 140          */
 141         WARN_ON(!cpumask_empty(offlined_cpus) ||
 142                 num_online_cpus() != nb_available_cpus);
 143
 144         return err;
 145 }
 146
 147 static void free_cpu_groups(int num, cpumask_var_t **pcpu_groups)
 148 {
 149         int i;
 150         cpumask_var_t *cpu_groups = *pcpu_groups;
 151
 152         for (i = 0; i < num; ++i)
 153                 free_cpumask_var(cpu_groups[i]);
 154         kfree(cpu_groups);
 155 }
 156
 157 static int alloc_init_cpu_groups(cpumask_var_t **pcpu_groups)
 158 {
 159         int num_groups = 0;
 160         cpumask_var_t tmp, *cpu_groups;
 161
 162         if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
 163                 return -ENOMEM;
 164
 165         cpu_groups = kcalloc(nb_available_cpus, sizeof(*cpu_groups),
 166                              GFP_KERNEL);
 167         if (!cpu_groups) {
 168                 free_cpumask_var(tmp);
 169                 return -ENOMEM;
 170         }
 171
 172         cpumask_copy(tmp, cpu_online_mask);
 173
 174         while (!cpumask_empty(tmp)) {
 175                 const struct cpumask *cpu_group =
 176                         topology_core_cpumask(cpumask_any(tmp));
 177
 178                 if (!alloc_cpumask_var(&cpu_groups[num_groups], GFP_KERNEL)) {
 179                         free_cpumask_var(tmp);
 180                         free_cpu_groups(num_groups, &cpu_groups);
 181                         return -ENOMEM;
 182                 }
 183                 cpumask_copy(cpu_groups[num_groups++], cpu_group);
 184                 cpumask_andnot(tmp, tmp, cpu_group);
 185         }
 186
 187         free_cpumask_var(tmp);
 188         *pcpu_groups = cpu_groups;
 189
 190         return num_groups;
 191 }
 192
 193 static int hotplug_tests(void)
 194 {
 195         int i, nb_cpu_group, err = -ENOMEM;
 196         cpumask_var_t offlined_cpus, *cpu_groups;
 197         char *page_buf;
 198
 199         if (!alloc_cpumask_var(&offlined_cpus, GFP_KERNEL))
 200                 return err;
 201
 202         nb_cpu_group = alloc_init_cpu_groups(&cpu_groups);
 203         if (nb_cpu_group < 0)
 204                 goto out_free_cpus;
 205         page_buf = (char *)__get_free_page(GFP_KERNEL);
 206         if (!page_buf)
 207                 goto out_free_cpu_groups;
 208
 209         err = 0;
 210         /*
 211          * Of course the last CPU cannot be powered down and cpu_down() should
 212          * refuse doing that.
 213          */
 214         pr_info("Trying to turn off and on again all CPUs\n");
 215         err += down_and_up_cpus(cpu_online_mask, offlined_cpus);
 216
 217         /*
 218          * Take down CPUs by cpu group this time. When the last CPU is turned
 219          * off, the cpu group itself should shut down.
 220          */
 221         for (i = 0; i < nb_cpu_group; ++i) {
 222                 ssize_t len = cpumap_print_to_pagebuf(true, page_buf,
 223                                                       cpu_groups[i]);
 224                 /* Remove trailing newline. */
 225                 page_buf[len - 1] = '\0';
 226                 pr_info("Trying to turn off and on again group %d (CPUs %s)\n",
 227                         i, page_buf);
 228                 err += down_and_up_cpus(cpu_groups[i], offlined_cpus);
 229         }
 230
 231         free_page((unsigned long)page_buf);
 232 out_free_cpu_groups:
 233         free_cpu_groups(nb_cpu_group, &cpu_groups);
 234 out_free_cpus:
 235         free_cpumask_var(offlined_cpus);
 236         return err;
 237 }
 238
 239 static void dummy_callback(struct timer_list *unused) {}
 240
 241 static int suspend_cpu(int index, bool broadcast)
 242 {
 243         int ret;
 244
 245         arch_cpu_idle_enter();
 246
 247         if (broadcast) {
 248                 /*
 249                  * The local timer will be shut down, we need to enter tick
 250                  * broadcast.
 251                  */
 252                 ret = tick_broadcast_enter();
 253                 if (ret) {
 254                         /*
 255                          * In the absence of hardware broadcast mechanism,
 256                          * this CPU might be used to broadcast wakeups, which
 257                          * may be why entering tick broadcast has failed.
 258                          * There is little the kernel can do to work around
 259                          * that, so enter WFI instead (idle state 0).
 260                          */
 261                         cpu_do_idle();
 262                         ret = 0;
 263                         goto out_arch_exit;
 264                 }
 265         }
 266
 267         /*
 268          * Replicate the common ARM cpuidle enter function
 269          * (arm_enter_idle_state).
 270          */
 271         ret = CPU_PM_CPU_IDLE_ENTER(arm_cpuidle_suspend, index);
 272
 273         if (broadcast)
 274                 tick_broadcast_exit();
 275
 276 out_arch_exit:
 277         arch_cpu_idle_exit();
 278
 279         return ret;
 280 }
 281
 282 static int suspend_test_thread(void *arg)
 283 {
 284         int cpu = (long)arg;
 285         int i, nb_suspend = 0, nb_shallow_sleep = 0, nb_err = 0;
 286         struct sched_param sched_priority = { .sched_priority = MAX_RT_PRIO-1 };
 287         struct cpuidle_device *dev;
 288         struct cpuidle_driver *drv;
 289         /* No need for an actual callback, we just want to wake up the CPU. */
 290         struct timer_list wakeup_timer;
 291
 292         /* Wait for the main thread to give the start signal. */
 293         wait_for_completion(&suspend_threads_started);
 294
 295         /* Set maximum priority to preempt all other threads on this CPU. */
 296         if (sched_setscheduler_nocheck(current, SCHED_FIFO, &sched_priority))
 297                 pr_warn("Failed to set suspend thread scheduler on CPU %d\n",
 298                         cpu);
 299
 300         dev = this_cpu_read(cpuidle_devices);
 301         drv = cpuidle_get_cpu_driver(dev);
 302
 303         pr_info("CPU %d entering suspend cycles, states 1 through %d\n",
 304                 cpu, drv->state_count - 1);
 305
 306         timer_setup_on_stack(&wakeup_timer, dummy_callback, 0);
 307         for (i = 0; i < NUM_SUSPEND_CYCLE; ++i) {
 308                 int index;
 309                 /*
 310                  * Test all possible states, except 0 (which is usually WFI and
 311                  * doesn't use PSCI).
 312                  */
 313                 for (index = 1; index < drv->state_count; ++index) {
 314                         struct cpuidle_state *state = &drv->states[index];
 315                         bool broadcast = state->flags & CPUIDLE_FLAG_TIMER_STOP;
 316                         int ret;
 317
 318                         /*
 319                          * Set the timer to wake this CPU up in some time (which
 320                          * should be largely sufficient for entering suspend).
 321                          * If the local tick is disabled when entering suspend,
 322                          * suspend_cpu() takes care of switching to a broadcast
 323                          * tick, so the timer will still wake us up.
 324                          */
 325                         mod_timer(&wakeup_timer, jiffies +
 326                                   usecs_to_jiffies(state->target_residency));
 327
 328                         /* IRQs must be disabled during suspend operations. */
 329                         local_irq_disable();
 330
 331                         ret = suspend_cpu(index, broadcast);
 332
 333                         /*
 334                          * We have woken up. Re-enable IRQs to handle any
 335                          * pending interrupt, do not wait until the end of the
 336                          * loop.
 337                          */
 338                         local_irq_enable();
 339
 340                         if (ret == index) {
 341                                 ++nb_suspend;
 342                         } else if (ret >= 0) {
 343                                 /* We did not enter the expected state. */
 344                                 ++nb_shallow_sleep;
 345                         } else {
 346                                 pr_err("Failed to suspend CPU %d: error %d "
 347                                        "(requested state %d, cycle %d)\n",
 348                                        cpu, ret, index, i);
 349                                 ++nb_err;
 350                         }
 351                 }
 352         }
 353
 354         /*
 355          * Disable the timer to make sure that the timer will not trigger
 356          * later.
 357          */
 358         del_timer(&wakeup_timer);
 359         destroy_timer_on_stack(&wakeup_timer);
 360
 361         if (atomic_dec_return_relaxed(&nb_active_threads) == 0)
 362                 complete(&suspend_threads_done);
 363
 364         /* Give up on RT scheduling and wait for termination. */
 365         sched_priority.sched_priority = 0;
 366         if (sched_setscheduler_nocheck(current, SCHED_NORMAL, &sched_priority))
 367                 pr_warn("Failed to set suspend thread scheduler on CPU %d\n",
 368                         cpu);
 369         for (;;) {
 370                 /* Needs to be set first to avoid missing a wakeup. */
 371                 set_current_state(TASK_INTERRUPTIBLE);
 372                 if (kthread_should_park())
 373                         break;
 374                 schedule();
 375         }
 376
 377         pr_info("CPU %d suspend test results: success %d, shallow states %d, errors %d\n",
 378                 cpu, nb_suspend, nb_shallow_sleep, nb_err);
 379
 380         kthread_parkme();
 381
 382         return nb_err;
 383 }
 384
 385 static int suspend_tests(void)
 386 {
 387         int i, cpu, err = 0;
 388         struct task_struct **threads;
 389         int nb_threads = 0;
 390
 391         threads = kmalloc_array(nb_available_cpus, sizeof(*threads),
 392                                 GFP_KERNEL);
 393         if (!threads)
 394                 return -ENOMEM;
 395
 396         /*
 397          * Stop cpuidle to prevent the idle tasks from entering a deep sleep
 398          * mode, as it might interfere with the suspend threads on other CPUs.
 399          * This does not prevent the suspend threads from using cpuidle (only
 400          * the idle tasks check this status). Take the idle lock so that
 401          * the cpuidle driver and device look-up can be carried out safely.
 402          */
 403         cpuidle_pause_and_lock();
 404
 405         for_each_online_cpu(cpu) {
 406                 struct task_struct *thread;
 407                 /* Check that cpuidle is available on that CPU. */
 408                 struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 409                 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
 410
 411                 if (!dev || !drv) {
 412                         pr_warn("cpuidle not available on CPU %d, ignoring\n",
 413                                 cpu);
 414                         continue;
 415                 }
 416
 417                 thread = kthread_create_on_cpu(suspend_test_thread,
 418                                                (void *)(long)cpu, cpu,
 419                                                "psci_suspend_test");
 420                 if (IS_ERR(thread))
 421                         pr_err("Failed to create kthread on CPU %d\n", cpu);
 422                 else
 423                         threads[nb_threads++] = thread;
 424         }
 425
 426         if (nb_threads < 1) {
 427                 err = -ENODEV;
 428                 goto out;
 429         }
 430
 431         atomic_set(&nb_active_threads, nb_threads);
 432
 433         /*
 434          * Wake up the suspend threads. To avoid the main thread being preempted
 435          * before all the threads have been unparked, the suspend threads will
 436          * wait for the completion of suspend_threads_started.
 437          */
 438         for (i = 0; i < nb_threads; ++i)
 439                 wake_up_process(threads[i]);
 440         complete_all(&suspend_threads_started);
 441
 442         wait_for_completion(&suspend_threads_done);
 443
 444
 445         /* Stop and destroy all threads, get return status. */
 446         for (i = 0; i < nb_threads; ++i) {
 447                 err += kthread_park(threads[i]);
 448                 err += kthread_stop(threads[i]);
 449         }
 450  out:
 451         cpuidle_resume_and_unlock();
 452         kfree(threads);
 453         return err;
 454 }
 455
 456 static int __init psci_checker(void)
 457 {
 458         int ret;
 459
 460         /*
 461          * Since we're in an initcall, we assume that all the CPUs that all
 462          * CPUs that can be onlined have been onlined.
 463          *
 464          * The tests assume that hotplug is enabled but nobody else is using it,
 465          * otherwise the results will be unpredictable. However, since there
 466          * is no userspace yet in initcalls, that should be fine, as long as
 467          * no torture test is running at the same time (see Kconfig).
 468          */
 469         nb_available_cpus = num_online_cpus();
 470
 471         /* Check PSCI operations are set up and working. */
 472         ret = psci_ops_check();
 473         if (ret)
 474                 return ret;
 475
 476         pr_info("PSCI checker started using %u CPUs\n", nb_available_cpus);
 477
 478         pr_info("Starting hotplug tests\n");
 479         ret = hotplug_tests();
 480         if (ret == 0)
 481                 pr_info("Hotplug tests passed OK\n");
 482         else if (ret > 0)
 483                 pr_err("%d error(s) encountered in hotplug tests\n", ret);
 484         else {
 485                 pr_err("Out of memory\n");
 486                 return ret;
 487         }
 488
 489         pr_info("Starting suspend tests (%d cycles per state)\n",
 490                 NUM_SUSPEND_CYCLE);
 491         ret = suspend_tests();
 492         if (ret == 0)
 493                 pr_info("Suspend tests passed OK\n");
 494         else if (ret > 0)
 495                 pr_err("%d error(s) encountered in suspend tests\n", ret);
 496         else {
 497                 switch (ret) {
 498                 case -ENOMEM:
 499                         pr_err("Out of memory\n");
 500                         break;
 501                 case -ENODEV:
 502                         pr_warn("Could not start suspend tests on any CPU\n");
 503                         break;
 504                 }
 505         }
 506
 507         pr_info("PSCI checker completed\n");
 508         return ret < 0 ? ret : 0;
 509 }
 510 late_initcall(psci_checker);