drivers/firmware/psci/psci_checker.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *
   4  * Copyright (C) 2016 ARM Limited
   5  */
   6
   7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   8
   9 #include <linux/atomic.h>
  10 #include <linux/completion.h>
  11 #include <linux/cpu.h>
  12 #include <linux/cpuidle.h>
  13 #include <linux/cpu_pm.h>
  14 #include <linux/kernel.h>
  15 #include <linux/kthread.h>
  16 #include <uapi/linux/sched/types.h>
  17 #include <linux/module.h>
  18 #include <linux/preempt.h>
  19 #include <linux/psci.h>
  20 #include <linux/slab.h>
  21 #include <linux/tick.h>
  22 #include <linux/topology.h>
  23
  24 #include <asm/cpuidle.h>
  25
  26 #include <uapi/linux/psci.h>
  27
  28 #define NUM_SUSPEND_CYCLE (10)
  29
  30 static unsigned int nb_available_cpus;
  31 static int tos_resident_cpu = -1;
  32
  33 static atomic_t nb_active_threads;
  34 static struct completion suspend_threads_started =
  35         COMPLETION_INITIALIZER(suspend_threads_started);
  36 static struct completion suspend_threads_done =
  37         COMPLETION_INITIALIZER(suspend_threads_done);
  38
  39 /*
  40  * We assume that PSCI operations are used if they are available. This is not
  41  * necessarily true on arm64, since the decision is based on the
  42  * "enable-method" property of each CPU in the DT, but given that there is no
  43  * arch-specific way to check this, we assume that the DT is sensible.
  44  */
  45 static int psci_ops_check(void)
  46 {
  47         int migrate_type = -1;
  48         int cpu;
  49
  50         if (!(psci_ops.cpu_off && psci_ops.cpu_on && psci_ops.cpu_suspend)) {
  51                 pr_warn("Missing PSCI operations, aborting tests\n");
  52                 return -EOPNOTSUPP;
  53         }
  54
  55         if (psci_ops.migrate_info_type)
  56                 migrate_type = psci_ops.migrate_info_type();
  57
  58         if (migrate_type == PSCI_0_2_TOS_UP_MIGRATE ||
  59             migrate_type == PSCI_0_2_TOS_UP_NO_MIGRATE) {
  60                 /* There is a UP Trusted OS, find on which core it resides. */
  61                 for_each_online_cpu(cpu)
  62                         if (psci_tos_resident_on(cpu)) {
  63                                 tos_resident_cpu = cpu;
  64                                 break;
  65                         }
  66                 if (tos_resident_cpu == -1)
  67                         pr_warn("UP Trusted OS resides on no online CPU\n");
  68         }
  69
  70         return 0;
  71 }
  72
  73 /*
  74  * offlined_cpus is a temporary array but passing it as an argument avoids
  75  * multiple allocations.
  76  */
  77 static unsigned int down_and_up_cpus(const struct cpumask *cpus,
  78                                      struct cpumask *offlined_cpus)
  79 {
  80         int cpu;
  81         int err = 0;
  82
  83         cpumask_clear(offlined_cpus);
  84
  85         /* Try to power down all CPUs in the mask. */
  86         for_each_cpu(cpu, cpus) {
  87                 int ret = cpu_down(cpu);
  88
  89                 /*
  90                  * cpu_down() checks the number of online CPUs before the TOS
  91                  * resident CPU.
  92                  */
  93                 if (cpumask_weight(offlined_cpus) + 1 == nb_available_cpus) {
  94                         if (ret != -EBUSY) {
  95                                 pr_err("Unexpected return code %d while trying "
  96                                        "to power down last online CPU %d\n",
  97                                        ret, cpu);
  98                                 ++err;
  99                         }
 100                 } else if (cpu == tos_resident_cpu) {
 101                         if (ret != -EPERM) {
 102                                 pr_err("Unexpected return code %d while trying "
 103                                        "to power down TOS resident CPU %d\n",
 104                                        ret, cpu);
 105                                 ++err;
 106                         }
 107                 } else if (ret != 0) {
 108                         pr_err("Error occurred (%d) while trying "
 109                                "to power down CPU %d\n", ret, cpu);
 110                         ++err;
 111                 }
 112
 113                 if (ret == 0)
 114                         cpumask_set_cpu(cpu, offlined_cpus);
 115         }
 116
 117         /* Try to power up all the CPUs that have been offlined. */
 118         for_each_cpu(cpu, offlined_cpus) {
 119                 int ret = cpu_up(cpu);
 120
 121                 if (ret != 0) {
 122                         pr_err("Error occurred (%d) while trying "
 123                                "to power up CPU %d\n", ret, cpu);
 124                         ++err;
 125                 } else {
 126                         cpumask_clear_cpu(cpu, offlined_cpus);
 127                 }
 128         }
 129
 130         /*
 131          * Something went bad at some point and some CPUs could not be turned
 132          * back on.
 133          */
 134         WARN_ON(!cpumask_empty(offlined_cpus) ||
 135                 num_online_cpus() != nb_available_cpus);
 136
 137         return err;
 138 }
 139
 140 static void free_cpu_groups(int num, cpumask_var_t **pcpu_groups)
 141 {
 142         int i;
 143         cpumask_var_t *cpu_groups = *pcpu_groups;
 144
 145         for (i = 0; i < num; ++i)
 146                 free_cpumask_var(cpu_groups[i]);
 147         kfree(cpu_groups);
 148 }
 149
 150 static int alloc_init_cpu_groups(cpumask_var_t **pcpu_groups)
 151 {
 152         int num_groups = 0;
 153         cpumask_var_t tmp, *cpu_groups;
 154
 155         if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
 156                 return -ENOMEM;
 157
 158         cpu_groups = kcalloc(nb_available_cpus, sizeof(cpu_groups),
 159                              GFP_KERNEL);
 160         if (!cpu_groups)
 161                 return -ENOMEM;
 162
 163         cpumask_copy(tmp, cpu_online_mask);
 164
 165         while (!cpumask_empty(tmp)) {
 166                 const struct cpumask *cpu_group =
 167                         topology_core_cpumask(cpumask_any(tmp));
 168
 169                 if (!alloc_cpumask_var(&cpu_groups[num_groups], GFP_KERNEL)) {
 170                         free_cpu_groups(num_groups, &cpu_groups);
 171                         return -ENOMEM;
 172                 }
 173                 cpumask_copy(cpu_groups[num_groups++], cpu_group);
 174                 cpumask_andnot(tmp, tmp, cpu_group);
 175         }
 176
 177         free_cpumask_var(tmp);
 178         *pcpu_groups = cpu_groups;
 179
 180         return num_groups;
 181 }
 182
 183 static int hotplug_tests(void)
 184 {
 185         int i, nb_cpu_group, err = -ENOMEM;
 186         cpumask_var_t offlined_cpus, *cpu_groups;
 187         char *page_buf;
 188
 189         if (!alloc_cpumask_var(&offlined_cpus, GFP_KERNEL))
 190                 return err;
 191
 192         nb_cpu_group = alloc_init_cpu_groups(&cpu_groups);
 193         if (nb_cpu_group < 0)
 194                 goto out_free_cpus;
 195         page_buf = (char *)__get_free_page(GFP_KERNEL);
 196         if (!page_buf)
 197                 goto out_free_cpu_groups;
 198
 199         err = 0;
 200         /*
 201          * Of course the last CPU cannot be powered down and cpu_down() should
 202          * refuse doing that.
 203          */
 204         pr_info("Trying to turn off and on again all CPUs\n");
 205         err += down_and_up_cpus(cpu_online_mask, offlined_cpus);
 206
 207         /*
 208          * Take down CPUs by cpu group this time. When the last CPU is turned
 209          * off, the cpu group itself should shut down.
 210          */
 211         for (i = 0; i < nb_cpu_group; ++i) {
 212                 ssize_t len = cpumap_print_to_pagebuf(true, page_buf,
 213                                                       cpu_groups[i]);
 214                 /* Remove trailing newline. */
 215                 page_buf[len - 1] = '\0';
 216                 pr_info("Trying to turn off and on again group %d (CPUs %s)\n",
 217                         i, page_buf);
 218                 err += down_and_up_cpus(cpu_groups[i], offlined_cpus);
 219         }
 220
 221         free_page((unsigned long)page_buf);
 222 out_free_cpu_groups:
 223         free_cpu_groups(nb_cpu_group, &cpu_groups);
 224 out_free_cpus:
 225         free_cpumask_var(offlined_cpus);
 226         return err;
 227 }
 228
 229 static void dummy_callback(struct timer_list *unused) {}
 230
 231 static int suspend_cpu(int index, bool broadcast)
 232 {
 233         int ret;
 234
 235         arch_cpu_idle_enter();
 236
 237         if (broadcast) {
 238                 /*
 239                  * The local timer will be shut down, we need to enter tick
 240                  * broadcast.
 241                  */
 242                 ret = tick_broadcast_enter();
 243                 if (ret) {
 244                         /*
 245                          * In the absence of hardware broadcast mechanism,
 246                          * this CPU might be used to broadcast wakeups, which
 247                          * may be why entering tick broadcast has failed.
 248                          * There is little the kernel can do to work around
 249                          * that, so enter WFI instead (idle state 0).
 250                          */
 251                         cpu_do_idle();
 252                         ret = 0;
 253                         goto out_arch_exit;
 254                 }
 255         }
 256
 257         /*
 258          * Replicate the common ARM cpuidle enter function
 259          * (arm_enter_idle_state).
 260          */
 261         ret = CPU_PM_CPU_IDLE_ENTER(arm_cpuidle_suspend, index);
 262
 263         if (broadcast)
 264                 tick_broadcast_exit();
 265
 266 out_arch_exit:
 267         arch_cpu_idle_exit();
 268
 269         return ret;
 270 }
 271
 272 static int suspend_test_thread(void *arg)
 273 {
 274         int cpu = (long)arg;
 275         int i, nb_suspend = 0, nb_shallow_sleep = 0, nb_err = 0;
 276         struct sched_param sched_priority = { .sched_priority = MAX_RT_PRIO-1 };
 277         struct cpuidle_device *dev;
 278         struct cpuidle_driver *drv;
 279         /* No need for an actual callback, we just want to wake up the CPU. */
 280         struct timer_list wakeup_timer;
 281
 282         /* Wait for the main thread to give the start signal. */
 283         wait_for_completion(&suspend_threads_started);
 284
 285         /* Set maximum priority to preempt all other threads on this CPU. */
 286         if (sched_setscheduler_nocheck(current, SCHED_FIFO, &sched_priority))
 287                 pr_warn("Failed to set suspend thread scheduler on CPU %d\n",
 288                         cpu);
 289
 290         dev = this_cpu_read(cpuidle_devices);
 291         drv = cpuidle_get_cpu_driver(dev);
 292
 293         pr_info("CPU %d entering suspend cycles, states 1 through %d\n",
 294                 cpu, drv->state_count - 1);
 295
 296         timer_setup_on_stack(&wakeup_timer, dummy_callback, 0);
 297         for (i = 0; i < NUM_SUSPEND_CYCLE; ++i) {
 298                 int index;
 299                 /*
 300                  * Test all possible states, except 0 (which is usually WFI and
 301                  * doesn't use PSCI).
 302                  */
 303                 for (index = 1; index < drv->state_count; ++index) {
 304                         struct cpuidle_state *state = &drv->states[index];
 305                         bool broadcast = state->flags & CPUIDLE_FLAG_TIMER_STOP;
 306                         int ret;
 307
 308                         /*
 309                          * Set the timer to wake this CPU up in some time (which
 310                          * should be largely sufficient for entering suspend).
 311                          * If the local tick is disabled when entering suspend,
 312                          * suspend_cpu() takes care of switching to a broadcast
 313                          * tick, so the timer will still wake us up.
 314                          */
 315                         mod_timer(&wakeup_timer, jiffies +
 316                                   usecs_to_jiffies(state->target_residency));
 317
 318                         /* IRQs must be disabled during suspend operations. */
 319                         local_irq_disable();
 320
 321                         ret = suspend_cpu(index, broadcast);
 322
 323                         /*
 324                          * We have woken up. Re-enable IRQs to handle any
 325                          * pending interrupt, do not wait until the end of the
 326                          * loop.
 327                          */
 328                         local_irq_enable();
 329
 330                         if (ret == index) {
 331                                 ++nb_suspend;
 332                         } else if (ret >= 0) {
 333                                 /* We did not enter the expected state. */
 334                                 ++nb_shallow_sleep;
 335                         } else {
 336                                 pr_err("Failed to suspend CPU %d: error %d "
 337                                        "(requested state %d, cycle %d)\n",
 338                                        cpu, ret, index, i);
 339                                 ++nb_err;
 340                         }
 341                 }
 342         }
 343
 344         /*
 345          * Disable the timer to make sure that the timer will not trigger
 346          * later.
 347          */
 348         del_timer(&wakeup_timer);
 349         destroy_timer_on_stack(&wakeup_timer);
 350
 351         if (atomic_dec_return_relaxed(&nb_active_threads) == 0)
 352                 complete(&suspend_threads_done);
 353
 354         /* Give up on RT scheduling and wait for termination. */
 355         sched_priority.sched_priority = 0;
 356         if (sched_setscheduler_nocheck(current, SCHED_NORMAL, &sched_priority))
 357                 pr_warn("Failed to set suspend thread scheduler on CPU %d\n",
 358                         cpu);
 359         for (;;) {
 360                 /* Needs to be set first to avoid missing a wakeup. */
 361                 set_current_state(TASK_INTERRUPTIBLE);
 362                 if (kthread_should_park())
 363                         break;
 364                 schedule();
 365         }
 366
 367         pr_info("CPU %d suspend test results: success %d, shallow states %d, errors %d\n",
 368                 cpu, nb_suspend, nb_shallow_sleep, nb_err);
 369
 370         kthread_parkme();
 371
 372         return nb_err;
 373 }
 374
 375 static int suspend_tests(void)
 376 {
 377         int i, cpu, err = 0;
 378         struct task_struct **threads;
 379         int nb_threads = 0;
 380
 381         threads = kmalloc_array(nb_available_cpus, sizeof(*threads),
 382                                 GFP_KERNEL);
 383         if (!threads)
 384                 return -ENOMEM;
 385
 386         /*
 387          * Stop cpuidle to prevent the idle tasks from entering a deep sleep
 388          * mode, as it might interfere with the suspend threads on other CPUs.
 389          * This does not prevent the suspend threads from using cpuidle (only
 390          * the idle tasks check this status). Take the idle lock so that
 391          * the cpuidle driver and device look-up can be carried out safely.
 392          */
 393         cpuidle_pause_and_lock();
 394
 395         for_each_online_cpu(cpu) {
 396                 struct task_struct *thread;
 397                 /* Check that cpuidle is available on that CPU. */
 398                 struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 399                 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
 400
 401                 if (!dev || !drv) {
 402                         pr_warn("cpuidle not available on CPU %d, ignoring\n",
 403                                 cpu);
 404                         continue;
 405                 }
 406
 407                 thread = kthread_create_on_cpu(suspend_test_thread,
 408                                                (void *)(long)cpu, cpu,
 409                                                "psci_suspend_test");
 410                 if (IS_ERR(thread))
 411                         pr_err("Failed to create kthread on CPU %d\n", cpu);
 412                 else
 413                         threads[nb_threads++] = thread;
 414         }
 415
 416         if (nb_threads < 1) {
 417                 err = -ENODEV;
 418                 goto out;
 419         }
 420
 421         atomic_set(&nb_active_threads, nb_threads);
 422
 423         /*
 424          * Wake up the suspend threads. To avoid the main thread being preempted
 425          * before all the threads have been unparked, the suspend threads will
 426          * wait for the completion of suspend_threads_started.
 427          */
 428         for (i = 0; i < nb_threads; ++i)
 429                 wake_up_process(threads[i]);
 430         complete_all(&suspend_threads_started);
 431
 432         wait_for_completion(&suspend_threads_done);
 433
 434
 435         /* Stop and destroy all threads, get return status. */
 436         for (i = 0; i < nb_threads; ++i) {
 437                 err += kthread_park(threads[i]);
 438                 err += kthread_stop(threads[i]);
 439         }
 440  out:
 441         cpuidle_resume_and_unlock();
 442         kfree(threads);
 443         return err;
 444 }
 445
 446 static int __init psci_checker(void)
 447 {
 448         int ret;
 449
 450         /*
 451          * Since we're in an initcall, we assume that all the CPUs that all
 452          * CPUs that can be onlined have been onlined.
 453          *
 454          * The tests assume that hotplug is enabled but nobody else is using it,
 455          * otherwise the results will be unpredictable. However, since there
 456          * is no userspace yet in initcalls, that should be fine, as long as
 457          * no torture test is running at the same time (see Kconfig).
 458          */
 459         nb_available_cpus = num_online_cpus();
 460
 461         /* Check PSCI operations are set up and working. */
 462         ret = psci_ops_check();
 463         if (ret)
 464                 return ret;
 465
 466         pr_info("PSCI checker started using %u CPUs\n", nb_available_cpus);
 467
 468         pr_info("Starting hotplug tests\n");
 469         ret = hotplug_tests();
 470         if (ret == 0)
 471                 pr_info("Hotplug tests passed OK\n");
 472         else if (ret > 0)
 473                 pr_err("%d error(s) encountered in hotplug tests\n", ret);
 474         else {
 475                 pr_err("Out of memory\n");
 476                 return ret;
 477         }
 478
 479         pr_info("Starting suspend tests (%d cycles per state)\n",
 480                 NUM_SUSPEND_CYCLE);
 481         ret = suspend_tests();
 482         if (ret == 0)
 483                 pr_info("Suspend tests passed OK\n");
 484         else if (ret > 0)
 485                 pr_err("%d error(s) encountered in suspend tests\n", ret);
 486         else {
 487                 switch (ret) {
 488                 case -ENOMEM:
 489                         pr_err("Out of memory\n");
 490                         break;
 491                 case -ENODEV:
 492                         pr_warn("Could not start suspend tests on any CPU\n");
 493                         break;
 494                 }
 495         }
 496
 497         pr_info("PSCI checker completed\n");
 498         return ret < 0 ? ret : 0;
 499 }
 500 late_initcall(psci_checker);