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 <linux/module.h>
  24 #include <linux/preempt.h>
  25 #include <linux/psci.h>
  26 #include <linux/slab.h>
  27 #include <linux/tick.h>
  28 #include <linux/topology.h>
  29
  30 #include <asm/cpuidle.h>
  31
  32 #include <uapi/linux/psci.h>
  33
  34 #define NUM_SUSPEND_CYCLE (10)
  35
  36 static unsigned int nb_available_cpus;
  37 static int tos_resident_cpu = -1;
  38
  39 static atomic_t nb_active_threads;
  40 static struct completion suspend_threads_started =
  41         COMPLETION_INITIALIZER(suspend_threads_started);
  42 static struct completion suspend_threads_done =
  43         COMPLETION_INITIALIZER(suspend_threads_done);
  44
  45 /*
  46  * We assume that PSCI operations are used if they are available. This is not
  47  * necessarily true on arm64, since the decision is based on the
  48  * "enable-method" property of each CPU in the DT, but given that there is no
  49  * arch-specific way to check this, we assume that the DT is sensible.
  50  */
  51 static int psci_ops_check(void)
  52 {
  53         int migrate_type = -1;
  54         int cpu;
  55
  56         if (!(psci_ops.cpu_off && psci_ops.cpu_on && psci_ops.cpu_suspend)) {
  57                 pr_warn("Missing PSCI operations, aborting tests\n");
  58                 return -EOPNOTSUPP;
  59         }
  60
  61         if (psci_ops.migrate_info_type)
  62                 migrate_type = psci_ops.migrate_info_type();
  63
  64         if (migrate_type == PSCI_0_2_TOS_UP_MIGRATE ||
  65             migrate_type == PSCI_0_2_TOS_UP_NO_MIGRATE) {
  66                 /* There is a UP Trusted OS, find on which core it resides. */
  67                 for_each_online_cpu(cpu)
  68                         if (psci_tos_resident_on(cpu)) {
  69                                 tos_resident_cpu = cpu;
  70                                 break;
  71                         }
  72                 if (tos_resident_cpu == -1)
  73                         pr_warn("UP Trusted OS resides on no online CPU\n");
  74         }
  75
  76         return 0;
  77 }
  78
  79 static int find_clusters(const struct cpumask *cpus,
  80                          const struct cpumask **clusters)
  81 {
  82         unsigned int nb = 0;
  83         cpumask_var_t tmp;
  84
  85         if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
  86                 return -ENOMEM;
  87         cpumask_copy(tmp, cpus);
  88
  89         while (!cpumask_empty(tmp)) {
  90                 const struct cpumask *cluster =
  91                         topology_core_cpumask(cpumask_any(tmp));
  92
  93                 clusters[nb++] = cluster;
  94                 cpumask_andnot(tmp, tmp, cluster);
  95         }
  96
  97         free_cpumask_var(tmp);
  98         return nb;
  99 }
 100
 101 /*
 102  * offlined_cpus is a temporary array but passing it as an argument avoids
 103  * multiple allocations.
 104  */
 105 static unsigned int down_and_up_cpus(const struct cpumask *cpus,
 106                                      struct cpumask *offlined_cpus)
 107 {
 108         int cpu;
 109         int err = 0;
 110
 111         cpumask_clear(offlined_cpus);
 112
 113         /* Try to power down all CPUs in the mask. */
 114         for_each_cpu(cpu, cpus) {
 115                 int ret = cpu_down(cpu);
 116
 117                 /*
 118                  * cpu_down() checks the number of online CPUs before the TOS
 119                  * resident CPU.
 120                  */
 121                 if (cpumask_weight(offlined_cpus) + 1 == nb_available_cpus) {
 122                         if (ret != -EBUSY) {
 123                                 pr_err("Unexpected return code %d while trying "
 124                                        "to power down last online CPU %d\n",
 125                                        ret, cpu);
 126                                 ++err;
 127                         }
 128                 } else if (cpu == tos_resident_cpu) {
 129                         if (ret != -EPERM) {
 130                                 pr_err("Unexpected return code %d while trying "
 131                                        "to power down TOS resident CPU %d\n",
 132                                        ret, cpu);
 133                                 ++err;
 134                         }
 135                 } else if (ret != 0) {
 136                         pr_err("Error occurred (%d) while trying "
 137                                "to power down CPU %d\n", ret, cpu);
 138                         ++err;
 139                 }
 140
 141                 if (ret == 0)
 142                         cpumask_set_cpu(cpu, offlined_cpus);
 143         }
 144
 145         /* Try to power up all the CPUs that have been offlined. */
 146         for_each_cpu(cpu, offlined_cpus) {
 147                 int ret = cpu_up(cpu);
 148
 149                 if (ret != 0) {
 150                         pr_err("Error occurred (%d) while trying "
 151                                "to power up CPU %d\n", ret, cpu);
 152                         ++err;
 153                 } else {
 154                         cpumask_clear_cpu(cpu, offlined_cpus);
 155                 }
 156         }
 157
 158         /*
 159          * Something went bad at some point and some CPUs could not be turned
 160          * back on.
 161          */
 162         WARN_ON(!cpumask_empty(offlined_cpus) ||
 163                 num_online_cpus() != nb_available_cpus);
 164
 165         return err;
 166 }
 167
 168 static int hotplug_tests(void)
 169 {
 170         int err;
 171         cpumask_var_t offlined_cpus;
 172         int i, nb_cluster;
 173         const struct cpumask **clusters;
 174         char *page_buf;
 175
 176         err = -ENOMEM;
 177         if (!alloc_cpumask_var(&offlined_cpus, GFP_KERNEL))
 178                 return err;
 179         /* We may have up to nb_available_cpus clusters. */
 180         clusters = kmalloc_array(nb_available_cpus, sizeof(*clusters),
 181                                  GFP_KERNEL);
 182         if (!clusters)
 183                 goto out_free_cpus;
 184         page_buf = (char *)__get_free_page(GFP_KERNEL);
 185         if (!page_buf)
 186                 goto out_free_clusters;
 187
 188         err = 0;
 189         nb_cluster = find_clusters(cpu_online_mask, clusters);
 190
 191         /*
 192          * Of course the last CPU cannot be powered down and cpu_down() should
 193          * refuse doing that.
 194          */
 195         pr_info("Trying to turn off and on again all CPUs\n");
 196         err += down_and_up_cpus(cpu_online_mask, offlined_cpus);
 197
 198         /*
 199          * Take down CPUs by cluster this time. When the last CPU is turned
 200          * off, the cluster itself should shut down.
 201          */
 202         for (i = 0; i < nb_cluster; ++i) {
 203                 int cluster_id =
 204                         topology_physical_package_id(cpumask_any(clusters[i]));
 205                 ssize_t len = cpumap_print_to_pagebuf(true, page_buf,
 206                                                       clusters[i]);
 207                 /* Remove trailing newline. */
 208                 page_buf[len - 1] = '\0';
 209                 pr_info("Trying to turn off and on again cluster %d "
 210                         "(CPUs %s)\n", cluster_id, page_buf);
 211                 err += down_and_up_cpus(clusters[i], offlined_cpus);
 212         }
 213
 214         free_page((unsigned long)page_buf);
 215 out_free_clusters:
 216         kfree(clusters);
 217 out_free_cpus:
 218         free_cpumask_var(offlined_cpus);
 219         return err;
 220 }
 221
 222 static void dummy_callback(unsigned long ignored) {}
 223
 224 static int suspend_cpu(int index, bool broadcast)
 225 {
 226         int ret;
 227
 228         arch_cpu_idle_enter();
 229
 230         if (broadcast) {
 231                 /*
 232                  * The local timer will be shut down, we need to enter tick
 233                  * broadcast.
 234                  */
 235                 ret = tick_broadcast_enter();
 236                 if (ret) {
 237                         /*
 238                          * In the absence of hardware broadcast mechanism,
 239                          * this CPU might be used to broadcast wakeups, which
 240                          * may be why entering tick broadcast has failed.
 241                          * There is little the kernel can do to work around
 242                          * that, so enter WFI instead (idle state 0).
 243                          */
 244                         cpu_do_idle();
 245                         ret = 0;
 246                         goto out_arch_exit;
 247                 }
 248         }
 249
 250         /*
 251          * Replicate the common ARM cpuidle enter function
 252          * (arm_enter_idle_state).
 253          */
 254         ret = CPU_PM_CPU_IDLE_ENTER(arm_cpuidle_suspend, index);
 255
 256         if (broadcast)
 257                 tick_broadcast_exit();
 258
 259 out_arch_exit:
 260         arch_cpu_idle_exit();
 261
 262         return ret;
 263 }
 264
 265 static int suspend_test_thread(void *arg)
 266 {
 267         int cpu = (long)arg;
 268         int i, nb_suspend = 0, nb_shallow_sleep = 0, nb_err = 0;
 269         struct sched_param sched_priority = { .sched_priority = MAX_RT_PRIO-1 };
 270         struct cpuidle_device *dev;
 271         struct cpuidle_driver *drv;
 272         /* No need for an actual callback, we just want to wake up the CPU. */
 273         struct timer_list wakeup_timer;
 274
 275         /* Wait for the main thread to give the start signal. */
 276         wait_for_completion(&suspend_threads_started);
 277
 278         /* Set maximum priority to preempt all other threads on this CPU. */
 279         if (sched_setscheduler_nocheck(current, SCHED_FIFO, &sched_priority))
 280                 pr_warn("Failed to set suspend thread scheduler on CPU %d\n",
 281                         cpu);
 282
 283         dev = this_cpu_read(cpuidle_devices);
 284         drv = cpuidle_get_cpu_driver(dev);
 285
 286         pr_info("CPU %d entering suspend cycles, states 1 through %d\n",
 287                 cpu, drv->state_count - 1);
 288
 289         setup_timer_on_stack(&wakeup_timer, dummy_callback, 0);
 290         for (i = 0; i < NUM_SUSPEND_CYCLE; ++i) {
 291                 int index;
 292                 /*
 293                  * Test all possible states, except 0 (which is usually WFI and
 294                  * doesn't use PSCI).
 295                  */
 296                 for (index = 1; index < drv->state_count; ++index) {
 297                         struct cpuidle_state *state = &drv->states[index];
 298                         bool broadcast = state->flags & CPUIDLE_FLAG_TIMER_STOP;
 299                         int ret;
 300
 301                         /*
 302                          * Set the timer to wake this CPU up in some time (which
 303                          * should be largely sufficient for entering suspend).
 304                          * If the local tick is disabled when entering suspend,
 305                          * suspend_cpu() takes care of switching to a broadcast
 306                          * tick, so the timer will still wake us up.
 307                          */
 308                         mod_timer(&wakeup_timer, jiffies +
 309                                   usecs_to_jiffies(state->target_residency));
 310
 311                         /* IRQs must be disabled during suspend operations. */
 312                         local_irq_disable();
 313
 314                         ret = suspend_cpu(index, broadcast);
 315
 316                         /*
 317                          * We have woken up. Re-enable IRQs to handle any
 318                          * pending interrupt, do not wait until the end of the
 319                          * loop.
 320                          */
 321                         local_irq_enable();
 322
 323                         if (ret == index) {
 324                                 ++nb_suspend;
 325                         } else if (ret >= 0) {
 326                                 /* We did not enter the expected state. */
 327                                 ++nb_shallow_sleep;
 328                         } else {
 329                                 pr_err("Failed to suspend CPU %d: error %d "
 330                                        "(requested state %d, cycle %d)\n",
 331                                        cpu, ret, index, i);
 332                                 ++nb_err;
 333                         }
 334                 }
 335         }
 336
 337         /*
 338          * Disable the timer to make sure that the timer will not trigger
 339          * later.
 340          */
 341         del_timer(&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);