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_clusters(const struct cpumask *cpus,
  81                          const struct cpumask **clusters)
  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 *cluster =
  92                         topology_core_cpumask(cpumask_any(tmp));
  93
  94                 clusters[nb++] = cluster;
  95                 cpumask_andnot(tmp, tmp, cluster);
  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_cluster;
 174         const struct cpumask **clusters;
 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 clusters. */
 181         clusters = kmalloc_array(nb_available_cpus, sizeof(*clusters),
 182                                  GFP_KERNEL);
 183         if (!clusters)
 184                 goto out_free_cpus;
 185         page_buf = (char *)__get_free_page(GFP_KERNEL);
 186         if (!page_buf)
 187                 goto out_free_clusters;
 188
 189         err = 0;
 190         nb_cluster = find_clusters(cpu_online_mask, clusters);
 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 cluster this time. When the last CPU is turned
 201          * off, the cluster itself should shut down.
 202          */
 203         for (i = 0; i < nb_cluster; ++i) {
 204                 int cluster_id =
 205                         topology_physical_package_id(cpumask_any(clusters[i]));
 206                 ssize_t len = cpumap_print_to_pagebuf(true, page_buf,
 207                                                       clusters[i]);
 208                 /* Remove trailing newline. */
 209                 page_buf[len - 1] = '\0';
 210                 pr_info("Trying to turn off and on again cluster %d "
 211                         "(CPUs %s)\n", cluster_id, page_buf);
 212                 err += down_and_up_cpus(clusters[i], offlined_cpus);
 213         }
 214
 215         free_page((unsigned long)page_buf);
 216 out_free_clusters:
 217         kfree(clusters);
 218 out_free_cpus:
 219         free_cpumask_var(offlined_cpus);
 220         return err;
 221 }
 222
 223 static void dummy_callback(struct timer_list *unused) {}
 224
 225 static int suspend_cpu(int index, bool broadcast)
 226 {
 227         int ret;
 228
 229         arch_cpu_idle_enter();
 230
 231         if (broadcast) {
 232                 /*
 233                  * The local timer will be shut down, we need to enter tick
 234                  * broadcast.
 235                  */
 236                 ret = tick_broadcast_enter();
 237                 if (ret) {
 238                         /*
 239                          * In the absence of hardware broadcast mechanism,
 240                          * this CPU might be used to broadcast wakeups, which
 241                          * may be why entering tick broadcast has failed.
 242                          * There is little the kernel can do to work around
 243                          * that, so enter WFI instead (idle state 0).
 244                          */
 245                         cpu_do_idle();
 246                         ret = 0;
 247                         goto out_arch_exit;
 248                 }
 249         }
 250
 251         /*
 252          * Replicate the common ARM cpuidle enter function
 253          * (arm_enter_idle_state).
 254          */
 255         ret = CPU_PM_CPU_IDLE_ENTER(arm_cpuidle_suspend, index);
 256
 257         if (broadcast)
 258                 tick_broadcast_exit();
 259
 260 out_arch_exit:
 261         arch_cpu_idle_exit();
 262
 263         return ret;
 264 }
 265
 266 static int suspend_test_thread(void *arg)
 267 {
 268         int cpu = (long)arg;
 269         int i, nb_suspend = 0, nb_shallow_sleep = 0, nb_err = 0;
 270         struct sched_param sched_priority = { .sched_priority = MAX_RT_PRIO-1 };
 271         struct cpuidle_device *dev;
 272         struct cpuidle_driver *drv;
 273         /* No need for an actual callback, we just want to wake up the CPU. */
 274         struct timer_list wakeup_timer;
 275
 276         /* Wait for the main thread to give the start signal. */
 277         wait_for_completion(&suspend_threads_started);
 278
 279         /* Set maximum priority to preempt all other threads on this CPU. */
 280         if (sched_setscheduler_nocheck(current, SCHED_FIFO, &sched_priority))
 281                 pr_warn("Failed to set suspend thread scheduler on CPU %d\n",
 282                         cpu);
 283
 284         dev = this_cpu_read(cpuidle_devices);
 285         drv = cpuidle_get_cpu_driver(dev);
 286
 287         pr_info("CPU %d entering suspend cycles, states 1 through %d\n",
 288                 cpu, drv->state_count - 1);
 289
 290         timer_setup_on_stack(&wakeup_timer, dummy_callback, 0);
 291         for (i = 0; i < NUM_SUSPEND_CYCLE; ++i) {
 292                 int index;
 293                 /*
 294                  * Test all possible states, except 0 (which is usually WFI and
 295                  * doesn't use PSCI).
 296                  */
 297                 for (index = 1; index < drv->state_count; ++index) {
 298                         struct cpuidle_state *state = &drv->states[index];
 299                         bool broadcast = state->flags & CPUIDLE_FLAG_TIMER_STOP;
 300                         int ret;
 301
 302                         /*
 303                          * Set the timer to wake this CPU up in some time (which
 304                          * should be largely sufficient for entering suspend).
 305                          * If the local tick is disabled when entering suspend,
 306                          * suspend_cpu() takes care of switching to a broadcast
 307                          * tick, so the timer will still wake us up.
 308                          */
 309                         mod_timer(&wakeup_timer, jiffies +
 310                                   usecs_to_jiffies(state->target_residency));
 311
 312                         /* IRQs must be disabled during suspend operations. */
 313                         local_irq_disable();
 314
 315                         ret = suspend_cpu(index, broadcast);
 316
 317                         /*
 318                          * We have woken up. Re-enable IRQs to handle any
 319                          * pending interrupt, do not wait until the end of the
 320                          * loop.
 321                          */
 322                         local_irq_enable();
 323
 324                         if (ret == index) {
 325                                 ++nb_suspend;
 326                         } else if (ret >= 0) {
 327                                 /* We did not enter the expected state. */
 328                                 ++nb_shallow_sleep;
 329                         } else {
 330                                 pr_err("Failed to suspend CPU %d: error %d "
 331                                        "(requested state %d, cycle %d)\n",
 332                                        cpu, ret, index, i);
 333                                 ++nb_err;
 334                         }
 335                 }
 336         }
 337
 338         /*
 339          * Disable the timer to make sure that the timer will not trigger
 340          * later.
 341          */
 342         del_timer(&wakeup_timer);
 343         destroy_timer_on_stack(&wakeup_timer);
 344
 345         if (atomic_dec_return_relaxed(&nb_active_threads) == 0)
 346                 complete(&suspend_threads_done);
 347
 348         /* Give up on RT scheduling and wait for termination. */
 349         sched_priority.sched_priority = 0;
 350         if (sched_setscheduler_nocheck(current, SCHED_NORMAL, &sched_priority))
 351                 pr_warn("Failed to set suspend thread scheduler on CPU %d\n",
 352                         cpu);
 353         for (;;) {
 354                 /* Needs to be set first to avoid missing a wakeup. */
 355                 set_current_state(TASK_INTERRUPTIBLE);
 356                 if (kthread_should_stop()) {
 357                         __set_current_state(TASK_RUNNING);
 358                         break;
 359                 }
 360                 schedule();
 361         }
 362
 363         pr_info("CPU %d suspend test results: success %d, shallow states %d, errors %d\n",
 364                 cpu, nb_suspend, nb_shallow_sleep, nb_err);
 365
 366         return nb_err;
 367 }
 368
 369 static int suspend_tests(void)
 370 {
 371         int i, cpu, err = 0;
 372         struct task_struct **threads;
 373         int nb_threads = 0;
 374
 375         threads = kmalloc_array(nb_available_cpus, sizeof(*threads),
 376                                 GFP_KERNEL);
 377         if (!threads)
 378                 return -ENOMEM;
 379
 380         /*
 381          * Stop cpuidle to prevent the idle tasks from entering a deep sleep
 382          * mode, as it might interfere with the suspend threads on other CPUs.
 383          * This does not prevent the suspend threads from using cpuidle (only
 384          * the idle tasks check this status). Take the idle lock so that
 385          * the cpuidle driver and device look-up can be carried out safely.
 386          */
 387         cpuidle_pause_and_lock();
 388
 389         for_each_online_cpu(cpu) {
 390                 struct task_struct *thread;
 391                 /* Check that cpuidle is available on that CPU. */
 392                 struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
 393                 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
 394
 395                 if (!dev || !drv) {
 396                         pr_warn("cpuidle not available on CPU %d, ignoring\n",
 397                                 cpu);
 398                         continue;
 399                 }
 400
 401                 thread = kthread_create_on_cpu(suspend_test_thread,
 402                                                (void *)(long)cpu, cpu,
 403                                                "psci_suspend_test");
 404                 if (IS_ERR(thread))
 405                         pr_err("Failed to create kthread on CPU %d\n", cpu);
 406                 else
 407                         threads[nb_threads++] = thread;
 408         }
 409
 410         if (nb_threads < 1) {
 411                 err = -ENODEV;
 412                 goto out;
 413         }
 414
 415         atomic_set(&nb_active_threads, nb_threads);
 416
 417         /*
 418          * Wake up the suspend threads. To avoid the main thread being preempted
 419          * before all the threads have been unparked, the suspend threads will
 420          * wait for the completion of suspend_threads_started.
 421          */
 422         for (i = 0; i < nb_threads; ++i)
 423                 wake_up_process(threads[i]);
 424         complete_all(&suspend_threads_started);
 425
 426         wait_for_completion(&suspend_threads_done);
 427
 428
 429         /* Stop and destroy all threads, get return status. */
 430         for (i = 0; i < nb_threads; ++i)
 431                 err += kthread_stop(threads[i]);
 432  out:
 433         cpuidle_resume_and_unlock();
 434         kfree(threads);
 435         return err;
 436 }
 437
 438 static int __init psci_checker(void)
 439 {
 440         int ret;
 441
 442         /*
 443          * Since we're in an initcall, we assume that all the CPUs that all
 444          * CPUs that can be onlined have been onlined.
 445          *
 446          * The tests assume that hotplug is enabled but nobody else is using it,
 447          * otherwise the results will be unpredictable. However, since there
 448          * is no userspace yet in initcalls, that should be fine, as long as
 449          * no torture test is running at the same time (see Kconfig).
 450          */
 451         nb_available_cpus = num_online_cpus();
 452
 453         /* Check PSCI operations are set up and working. */
 454         ret = psci_ops_check();
 455         if (ret)
 456                 return ret;
 457
 458         pr_info("PSCI checker started using %u CPUs\n", nb_available_cpus);
 459
 460         pr_info("Starting hotplug tests\n");
 461         ret = hotplug_tests();
 462         if (ret == 0)
 463                 pr_info("Hotplug tests passed OK\n");
 464         else if (ret > 0)
 465                 pr_err("%d error(s) encountered in hotplug tests\n", ret);
 466         else {
 467                 pr_err("Out of memory\n");
 468                 return ret;
 469         }
 470
 471         pr_info("Starting suspend tests (%d cycles per state)\n",
 472                 NUM_SUSPEND_CYCLE);
 473         ret = suspend_tests();
 474         if (ret == 0)
 475                 pr_info("Suspend tests passed OK\n");
 476         else if (ret > 0)
 477                 pr_err("%d error(s) encountered in suspend tests\n", ret);
 478         else {
 479                 switch (ret) {
 480                 case -ENOMEM:
 481                         pr_err("Out of memory\n");
 482                         break;
 483                 case -ENODEV:
 484                         pr_warn("Could not start suspend tests on any CPU\n");
 485                         break;
 486                 }
 487         }
 488
 489         pr_info("PSCI checker completed\n");
 490         return ret < 0 ? ret : 0;
 491 }
 492 late_initcall(psci_checker);