drivers/accel/ivpu/ivpu_pm.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2020-2023 Intel Corporation
   4  */
   5
   6 #include <linux/highmem.h>
   7 #include <linux/moduleparam.h>
   8 #include <linux/pci.h>
   9 #include <linux/pm_runtime.h>
  10 #include <linux/reboot.h>
  11
  12 #include "vpu_boot_api.h"
  13 #include "ivpu_drv.h"
  14 #include "ivpu_hw.h"
  15 #include "ivpu_fw.h"
  16 #include "ivpu_fw_log.h"
  17 #include "ivpu_ipc.h"
  18 #include "ivpu_job.h"
  19 #include "ivpu_jsm_msg.h"
  20 #include "ivpu_mmu.h"
  21 #include "ivpu_pm.h"
  22
  23 static bool ivpu_disable_recovery;
  24 module_param_named_unsafe(disable_recovery, ivpu_disable_recovery, bool, 0644);
  25 MODULE_PARM_DESC(disable_recovery, "Disables recovery when NPU hang is detected");
  26
  27 static unsigned long ivpu_tdr_timeout_ms;
  28 module_param_named(tdr_timeout_ms, ivpu_tdr_timeout_ms, ulong, 0644);
  29 MODULE_PARM_DESC(tdr_timeout_ms, "Timeout for device hang detection, in milliseconds, 0 - default");
  30
  31 #define PM_RESCHEDULE_LIMIT     5
  32
  33 static void ivpu_pm_prepare_cold_boot(struct ivpu_device *vdev)
  34 {
  35         struct ivpu_fw_info *fw = vdev->fw;
  36
  37         ivpu_cmdq_reset_all_contexts(vdev);
  38         ivpu_ipc_reset(vdev);
  39         ivpu_fw_load(vdev);
  40         fw->entry_point = fw->cold_boot_entry_point;
  41 }
  42
  43 static void ivpu_pm_prepare_warm_boot(struct ivpu_device *vdev)
  44 {
  45         struct ivpu_fw_info *fw = vdev->fw;
  46         struct vpu_boot_params *bp = ivpu_bo_vaddr(fw->mem);
  47
  48         if (!bp->save_restore_ret_address) {
  49                 ivpu_pm_prepare_cold_boot(vdev);
  50                 return;
  51         }
  52
  53         ivpu_dbg(vdev, FW_BOOT, "Save/restore entry point %llx", bp->save_restore_ret_address);
  54         fw->entry_point = bp->save_restore_ret_address;
  55 }
  56
  57 static int ivpu_suspend(struct ivpu_device *vdev)
  58 {
  59         int ret;
  60
  61         /* Save PCI state before powering down as it sometimes gets corrupted if NPU hangs */
  62         pci_save_state(to_pci_dev(vdev->drm.dev));
  63
  64         ret = ivpu_shutdown(vdev);
  65         if (ret)
  66                 ivpu_err(vdev, "Failed to shutdown VPU: %d\n", ret);
  67
  68         pci_set_power_state(to_pci_dev(vdev->drm.dev), PCI_D3hot);
  69
  70         return ret;
  71 }
  72
  73 static int ivpu_resume(struct ivpu_device *vdev)
  74 {
  75         int ret;
  76
  77         pci_set_power_state(to_pci_dev(vdev->drm.dev), PCI_D0);
  78         pci_restore_state(to_pci_dev(vdev->drm.dev));
  79
  80 retry:
  81         ret = ivpu_hw_power_up(vdev);
  82         if (ret) {
  83                 ivpu_err(vdev, "Failed to power up HW: %d\n", ret);
  84                 goto err_power_down;
  85         }
  86
  87         ret = ivpu_mmu_enable(vdev);
  88         if (ret) {
  89                 ivpu_err(vdev, "Failed to resume MMU: %d\n", ret);
  90                 goto err_power_down;
  91         }
  92
  93         ret = ivpu_boot(vdev);
  94         if (ret)
  95                 goto err_mmu_disable;
  96
  97         return 0;
  98
  99 err_mmu_disable:
 100         ivpu_mmu_disable(vdev);
 101 err_power_down:
 102         ivpu_hw_power_down(vdev);
 103
 104         if (!ivpu_fw_is_cold_boot(vdev)) {
 105                 ivpu_pm_prepare_cold_boot(vdev);
 106                 goto retry;
 107         } else {
 108                 ivpu_err(vdev, "Failed to resume the FW: %d\n", ret);
 109         }
 110
 111         return ret;
 112 }
 113
 114 static void ivpu_pm_recovery_work(struct work_struct *work)
 115 {
 116         struct ivpu_pm_info *pm = container_of(work, struct ivpu_pm_info, recovery_work);
 117         struct ivpu_device *vdev = pm->vdev;
 118         char *evt[2] = {"IVPU_PM_EVENT=IVPU_RECOVER", NULL};
 119         int ret;
 120
 121         ivpu_err(vdev, "Recovering the NPU (reset #%d)\n", atomic_read(&vdev->pm->reset_counter));
 122
 123         ret = pm_runtime_resume_and_get(vdev->drm.dev);
 124         if (ret)
 125                 ivpu_err(vdev, "Failed to resume NPU: %d\n", ret);
 126
 127         ivpu_fw_log_dump(vdev);
 128
 129         atomic_inc(&vdev->pm->reset_counter);
 130         atomic_set(&vdev->pm->reset_pending, 1);
 131         down_write(&vdev->pm->reset_lock);
 132
 133         ivpu_suspend(vdev);
 134         ivpu_pm_prepare_cold_boot(vdev);
 135         ivpu_jobs_abort_all(vdev);
 136
 137         ret = ivpu_resume(vdev);
 138         if (ret)
 139                 ivpu_err(vdev, "Failed to resume NPU: %d\n", ret);
 140
 141         up_write(&vdev->pm->reset_lock);
 142         atomic_set(&vdev->pm->reset_pending, 0);
 143
 144         kobject_uevent_env(&vdev->drm.dev->kobj, KOBJ_CHANGE, evt);
 145         pm_runtime_mark_last_busy(vdev->drm.dev);
 146         pm_runtime_put_autosuspend(vdev->drm.dev);
 147 }
 148
 149 void ivpu_pm_trigger_recovery(struct ivpu_device *vdev, const char *reason)
 150 {
 151         ivpu_err(vdev, "Recovery triggered by %s\n", reason);
 152
 153         if (ivpu_disable_recovery) {
 154                 ivpu_err(vdev, "Recovery not available when disable_recovery param is set\n");
 155                 return;
 156         }
 157
 158         if (ivpu_is_fpga(vdev)) {
 159                 ivpu_err(vdev, "Recovery not available on FPGA\n");
 160                 return;
 161         }
 162
 163         /* Trigger recovery if it's not in progress */
 164         if (atomic_cmpxchg(&vdev->pm->reset_pending, 0, 1) == 0) {
 165                 ivpu_hw_diagnose_failure(vdev);
 166                 ivpu_hw_irq_disable(vdev); /* Disable IRQ early to protect from IRQ storm */
 167                 queue_work(system_long_wq, &vdev->pm->recovery_work);
 168         }
 169 }
 170
 171 static void ivpu_job_timeout_work(struct work_struct *work)
 172 {
 173         struct ivpu_pm_info *pm = container_of(work, struct ivpu_pm_info, job_timeout_work.work);
 174         struct ivpu_device *vdev = pm->vdev;
 175
 176         ivpu_pm_trigger_recovery(vdev, "TDR");
 177 }
 178
 179 void ivpu_start_job_timeout_detection(struct ivpu_device *vdev)
 180 {
 181         unsigned long timeout_ms = ivpu_tdr_timeout_ms ? ivpu_tdr_timeout_ms : vdev->timeout.tdr;
 182
 183         /* No-op if already queued */
 184         queue_delayed_work(system_wq, &vdev->pm->job_timeout_work, msecs_to_jiffies(timeout_ms));
 185 }
 186
 187 void ivpu_stop_job_timeout_detection(struct ivpu_device *vdev)
 188 {
 189         cancel_delayed_work_sync(&vdev->pm->job_timeout_work);
 190 }
 191
 192 int ivpu_pm_suspend_cb(struct device *dev)
 193 {
 194         struct drm_device *drm = dev_get_drvdata(dev);
 195         struct ivpu_device *vdev = to_ivpu_device(drm);
 196         unsigned long timeout;
 197
 198         ivpu_dbg(vdev, PM, "Suspend..\n");
 199
 200         timeout = jiffies + msecs_to_jiffies(vdev->timeout.tdr);
 201         while (!ivpu_hw_is_idle(vdev)) {
 202                 cond_resched();
 203                 if (time_after_eq(jiffies, timeout)) {
 204                         ivpu_err(vdev, "Failed to enter idle on system suspend\n");
 205                         return -EBUSY;
 206                 }
 207         }
 208
 209         ivpu_jsm_pwr_d0i3_enter(vdev);
 210
 211         ivpu_suspend(vdev);
 212         ivpu_pm_prepare_warm_boot(vdev);
 213
 214         ivpu_dbg(vdev, PM, "Suspend done.\n");
 215
 216         return 0;
 217 }
 218
 219 int ivpu_pm_resume_cb(struct device *dev)
 220 {
 221         struct drm_device *drm = dev_get_drvdata(dev);
 222         struct ivpu_device *vdev = to_ivpu_device(drm);
 223         int ret;
 224
 225         ivpu_dbg(vdev, PM, "Resume..\n");
 226
 227         ret = ivpu_resume(vdev);
 228         if (ret)
 229                 ivpu_err(vdev, "Failed to resume: %d\n", ret);
 230
 231         ivpu_dbg(vdev, PM, "Resume done.\n");
 232
 233         return ret;
 234 }
 235
 236 int ivpu_pm_runtime_suspend_cb(struct device *dev)
 237 {
 238         struct drm_device *drm = dev_get_drvdata(dev);
 239         struct ivpu_device *vdev = to_ivpu_device(drm);
 240         bool hw_is_idle = true;
 241         int ret;
 242
 243         drm_WARN_ON(&vdev->drm, !xa_empty(&vdev->submitted_jobs_xa));
 244         drm_WARN_ON(&vdev->drm, work_pending(&vdev->pm->recovery_work));
 245
 246         ivpu_dbg(vdev, PM, "Runtime suspend..\n");
 247
 248         if (!ivpu_hw_is_idle(vdev) && vdev->pm->suspend_reschedule_counter) {
 249                 ivpu_dbg(vdev, PM, "Failed to enter idle, rescheduling suspend, retries left %d\n",
 250                          vdev->pm->suspend_reschedule_counter);
 251                 pm_schedule_suspend(dev, vdev->timeout.reschedule_suspend);
 252                 vdev->pm->suspend_reschedule_counter--;
 253                 return -EAGAIN;
 254         }
 255
 256         if (!vdev->pm->suspend_reschedule_counter)
 257                 hw_is_idle = false;
 258         else if (ivpu_jsm_pwr_d0i3_enter(vdev))
 259                 hw_is_idle = false;
 260
 261         ret = ivpu_suspend(vdev);
 262         if (ret)
 263                 ivpu_err(vdev, "Failed to suspend NPU: %d\n", ret);
 264
 265         if (!hw_is_idle) {
 266                 ivpu_err(vdev, "NPU failed to enter idle, force suspended.\n");
 267                 ivpu_fw_log_dump(vdev);
 268                 ivpu_pm_prepare_cold_boot(vdev);
 269         } else {
 270                 ivpu_pm_prepare_warm_boot(vdev);
 271         }
 272
 273         vdev->pm->suspend_reschedule_counter = PM_RESCHEDULE_LIMIT;
 274
 275         ivpu_dbg(vdev, PM, "Runtime suspend done.\n");
 276
 277         return 0;
 278 }
 279
 280 int ivpu_pm_runtime_resume_cb(struct device *dev)
 281 {
 282         struct drm_device *drm = dev_get_drvdata(dev);
 283         struct ivpu_device *vdev = to_ivpu_device(drm);
 284         int ret;
 285
 286         ivpu_dbg(vdev, PM, "Runtime resume..\n");
 287
 288         ret = ivpu_resume(vdev);
 289         if (ret)
 290                 ivpu_err(vdev, "Failed to set RESUME state: %d\n", ret);
 291
 292         ivpu_dbg(vdev, PM, "Runtime resume done.\n");
 293
 294         return ret;
 295 }
 296
 297 int ivpu_rpm_get(struct ivpu_device *vdev)
 298 {
 299         int ret;
 300
 301         ret = pm_runtime_resume_and_get(vdev->drm.dev);
 302         if (!drm_WARN_ON(&vdev->drm, ret < 0))
 303                 vdev->pm->suspend_reschedule_counter = PM_RESCHEDULE_LIMIT;
 304
 305         return ret;
 306 }
 307
 308 int ivpu_rpm_get_if_active(struct ivpu_device *vdev)
 309 {
 310         int ret;
 311
 312         ret = pm_runtime_get_if_in_use(vdev->drm.dev);
 313         drm_WARN_ON(&vdev->drm, ret < 0);
 314
 315         return ret;
 316 }
 317
 318 void ivpu_rpm_put(struct ivpu_device *vdev)
 319 {
 320         pm_runtime_mark_last_busy(vdev->drm.dev);
 321         pm_runtime_put_autosuspend(vdev->drm.dev);
 322 }
 323
 324 void ivpu_pm_reset_prepare_cb(struct pci_dev *pdev)
 325 {
 326         struct ivpu_device *vdev = pci_get_drvdata(pdev);
 327
 328         ivpu_dbg(vdev, PM, "Pre-reset..\n");
 329         atomic_inc(&vdev->pm->reset_counter);
 330         atomic_set(&vdev->pm->reset_pending, 1);
 331
 332         pm_runtime_get_sync(vdev->drm.dev);
 333         down_write(&vdev->pm->reset_lock);
 334         ivpu_prepare_for_reset(vdev);
 335         ivpu_hw_reset(vdev);
 336         ivpu_pm_prepare_cold_boot(vdev);
 337         ivpu_jobs_abort_all(vdev);
 338         ivpu_dbg(vdev, PM, "Pre-reset done.\n");
 339 }
 340
 341 void ivpu_pm_reset_done_cb(struct pci_dev *pdev)
 342 {
 343         struct ivpu_device *vdev = pci_get_drvdata(pdev);
 344         int ret;
 345
 346         ivpu_dbg(vdev, PM, "Post-reset..\n");
 347         ret = ivpu_resume(vdev);
 348         if (ret)
 349                 ivpu_err(vdev, "Failed to set RESUME state: %d\n", ret);
 350         up_write(&vdev->pm->reset_lock);
 351         atomic_set(&vdev->pm->reset_pending, 0);
 352         ivpu_dbg(vdev, PM, "Post-reset done.\n");
 353
 354         pm_runtime_mark_last_busy(vdev->drm.dev);
 355         pm_runtime_put_autosuspend(vdev->drm.dev);
 356 }
 357
 358 void ivpu_pm_init(struct ivpu_device *vdev)
 359 {
 360         struct device *dev = vdev->drm.dev;
 361         struct ivpu_pm_info *pm = vdev->pm;
 362         int delay;
 363
 364         pm->vdev = vdev;
 365         pm->suspend_reschedule_counter = PM_RESCHEDULE_LIMIT;
 366
 367         init_rwsem(&pm->reset_lock);
 368         atomic_set(&pm->reset_pending, 0);
 369         atomic_set(&pm->reset_counter, 0);
 370
 371         INIT_WORK(&pm->recovery_work, ivpu_pm_recovery_work);
 372         INIT_DELAYED_WORK(&pm->job_timeout_work, ivpu_job_timeout_work);
 373
 374         if (ivpu_disable_recovery)
 375                 delay = -1;
 376         else
 377                 delay = vdev->timeout.autosuspend;
 378
 379         pm_runtime_use_autosuspend(dev);
 380         pm_runtime_set_autosuspend_delay(dev, delay);
 381
 382         ivpu_dbg(vdev, PM, "Autosuspend delay = %d\n", delay);
 383 }
 384
 385 void ivpu_pm_cancel_recovery(struct ivpu_device *vdev)
 386 {
 387         drm_WARN_ON(&vdev->drm, delayed_work_pending(&vdev->pm->job_timeout_work));
 388         cancel_work_sync(&vdev->pm->recovery_work);
 389 }
 390
 391 void ivpu_pm_enable(struct ivpu_device *vdev)
 392 {
 393         struct device *dev = vdev->drm.dev;
 394
 395         pm_runtime_set_active(dev);
 396         pm_runtime_allow(dev);
 397         pm_runtime_mark_last_busy(dev);
 398         pm_runtime_put_autosuspend(dev);
 399 }
 400
 401 void ivpu_pm_disable(struct ivpu_device *vdev)
 402 {
 403         pm_runtime_get_noresume(vdev->drm.dev);
 404         pm_runtime_forbid(vdev->drm.dev);
 405 }