1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2020-2024 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_ms.h" 22 #include "ivpu_pm.h" 23 24 static bool ivpu_disable_recovery; 25 module_param_named_unsafe(disable_recovery, ivpu_disable_recovery, bool, 0644); 26 MODULE_PARM_DESC(disable_recovery, "Disables recovery when NPU hang is detected"); 27 28 static unsigned long ivpu_tdr_timeout_ms; 29 module_param_named(tdr_timeout_ms, ivpu_tdr_timeout_ms, ulong, 0644); 30 MODULE_PARM_DESC(tdr_timeout_ms, "Timeout for device hang detection, in milliseconds, 0 - default"); 31 32 #define PM_RESCHEDULE_LIMIT 5 33 34 static void ivpu_pm_prepare_cold_boot(struct ivpu_device *vdev) 35 { 36 struct ivpu_fw_info *fw = vdev->fw; 37 38 ivpu_cmdq_reset_all_contexts(vdev); 39 ivpu_ipc_reset(vdev); 40 ivpu_fw_load(vdev); 41 fw->entry_point = fw->cold_boot_entry_point; 42 } 43 44 static void ivpu_pm_prepare_warm_boot(struct ivpu_device *vdev) 45 { 46 struct ivpu_fw_info *fw = vdev->fw; 47 struct vpu_boot_params *bp = ivpu_bo_vaddr(fw->mem); 48 49 if (!bp->save_restore_ret_address) { 50 ivpu_pm_prepare_cold_boot(vdev); 51 return; 52 } 53 54 ivpu_dbg(vdev, FW_BOOT, "Save/restore entry point %llx", bp->save_restore_ret_address); 55 fw->entry_point = bp->save_restore_ret_address; 56 } 57 58 static int ivpu_suspend(struct ivpu_device *vdev) 59 { 60 int ret; 61 62 ivpu_prepare_for_reset(vdev); 63 64 ret = ivpu_shutdown(vdev); 65 if (ret) 66 ivpu_err(vdev, "Failed to shutdown NPU: %d\n", ret); 67 68 return ret; 69 } 70 71 static int ivpu_resume(struct ivpu_device *vdev) 72 { 73 int ret; 74 75 retry: 76 pci_restore_state(to_pci_dev(vdev->drm.dev)); 77 pci_set_power_state(to_pci_dev(vdev->drm.dev), PCI_D0); 78 79 ret = ivpu_hw_power_up(vdev); 80 if (ret) { 81 ivpu_err(vdev, "Failed to power up HW: %d\n", ret); 82 goto err_power_down; 83 } 84 85 ret = ivpu_mmu_enable(vdev); 86 if (ret) { 87 ivpu_err(vdev, "Failed to resume MMU: %d\n", ret); 88 goto err_power_down; 89 } 90 91 ret = ivpu_boot(vdev); 92 if (ret) 93 goto err_mmu_disable; 94 95 return 0; 96 97 err_mmu_disable: 98 ivpu_mmu_disable(vdev); 99 err_power_down: 100 ivpu_hw_power_down(vdev); 101 pci_set_power_state(to_pci_dev(vdev->drm.dev), PCI_D3hot); 102 103 if (!ivpu_fw_is_cold_boot(vdev)) { 104 ivpu_pm_prepare_cold_boot(vdev); 105 goto retry; 106 } else { 107 ivpu_err(vdev, "Failed to resume the FW: %d\n", ret); 108 } 109 110 return ret; 111 } 112 113 static void ivpu_pm_recovery_work(struct work_struct *work) 114 { 115 struct ivpu_pm_info *pm = container_of(work, struct ivpu_pm_info, recovery_work); 116 struct ivpu_device *vdev = pm->vdev; 117 char *evt[2] = {"IVPU_PM_EVENT=IVPU_RECOVER", NULL}; 118 int ret; 119 120 ivpu_err(vdev, "Recovering the NPU (reset #%d)\n", atomic_read(&vdev->pm->reset_counter)); 121 122 ret = pm_runtime_resume_and_get(vdev->drm.dev); 123 if (ret) 124 ivpu_err(vdev, "Failed to resume NPU: %d\n", ret); 125 126 ivpu_fw_log_dump(vdev); 127 128 atomic_inc(&vdev->pm->reset_counter); 129 atomic_set(&vdev->pm->reset_pending, 1); 130 down_write(&vdev->pm->reset_lock); 131 132 ivpu_suspend(vdev); 133 ivpu_pm_prepare_cold_boot(vdev); 134 ivpu_jobs_abort_all(vdev); 135 ivpu_ms_cleanup_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 atomic_inc(&vdev->pm->reset_counter); 268 ivpu_fw_log_dump(vdev); 269 ivpu_pm_prepare_cold_boot(vdev); 270 } else { 271 ivpu_pm_prepare_warm_boot(vdev); 272 } 273 274 vdev->pm->suspend_reschedule_counter = PM_RESCHEDULE_LIMIT; 275 276 ivpu_dbg(vdev, PM, "Runtime suspend done.\n"); 277 278 return 0; 279 } 280 281 int ivpu_pm_runtime_resume_cb(struct device *dev) 282 { 283 struct drm_device *drm = dev_get_drvdata(dev); 284 struct ivpu_device *vdev = to_ivpu_device(drm); 285 int ret; 286 287 ivpu_dbg(vdev, PM, "Runtime resume..\n"); 288 289 ret = ivpu_resume(vdev); 290 if (ret) 291 ivpu_err(vdev, "Failed to set RESUME state: %d\n", ret); 292 293 ivpu_dbg(vdev, PM, "Runtime resume done.\n"); 294 295 return ret; 296 } 297 298 int ivpu_rpm_get(struct ivpu_device *vdev) 299 { 300 int ret; 301 302 ret = pm_runtime_resume_and_get(vdev->drm.dev); 303 if (!drm_WARN_ON(&vdev->drm, ret < 0)) 304 vdev->pm->suspend_reschedule_counter = PM_RESCHEDULE_LIMIT; 305 306 return ret; 307 } 308 309 int ivpu_rpm_get_if_active(struct ivpu_device *vdev) 310 { 311 int ret; 312 313 ret = pm_runtime_get_if_in_use(vdev->drm.dev); 314 drm_WARN_ON(&vdev->drm, ret < 0); 315 316 return ret; 317 } 318 319 void ivpu_rpm_put(struct ivpu_device *vdev) 320 { 321 pm_runtime_mark_last_busy(vdev->drm.dev); 322 pm_runtime_put_autosuspend(vdev->drm.dev); 323 } 324 325 void ivpu_pm_reset_prepare_cb(struct pci_dev *pdev) 326 { 327 struct ivpu_device *vdev = pci_get_drvdata(pdev); 328 329 ivpu_dbg(vdev, PM, "Pre-reset..\n"); 330 atomic_inc(&vdev->pm->reset_counter); 331 atomic_set(&vdev->pm->reset_pending, 1); 332 333 pm_runtime_get_sync(vdev->drm.dev); 334 down_write(&vdev->pm->reset_lock); 335 ivpu_prepare_for_reset(vdev); 336 ivpu_hw_reset(vdev); 337 ivpu_pm_prepare_cold_boot(vdev); 338 ivpu_jobs_abort_all(vdev); 339 ivpu_ms_cleanup_all(vdev); 340 341 ivpu_dbg(vdev, PM, "Pre-reset done.\n"); 342 } 343 344 void ivpu_pm_reset_done_cb(struct pci_dev *pdev) 345 { 346 struct ivpu_device *vdev = pci_get_drvdata(pdev); 347 int ret; 348 349 ivpu_dbg(vdev, PM, "Post-reset..\n"); 350 ret = ivpu_resume(vdev); 351 if (ret) 352 ivpu_err(vdev, "Failed to set RESUME state: %d\n", ret); 353 up_write(&vdev->pm->reset_lock); 354 atomic_set(&vdev->pm->reset_pending, 0); 355 ivpu_dbg(vdev, PM, "Post-reset done.\n"); 356 357 pm_runtime_mark_last_busy(vdev->drm.dev); 358 pm_runtime_put_autosuspend(vdev->drm.dev); 359 } 360 361 void ivpu_pm_init(struct ivpu_device *vdev) 362 { 363 struct device *dev = vdev->drm.dev; 364 struct ivpu_pm_info *pm = vdev->pm; 365 int delay; 366 367 pm->vdev = vdev; 368 pm->suspend_reschedule_counter = PM_RESCHEDULE_LIMIT; 369 370 init_rwsem(&pm->reset_lock); 371 atomic_set(&pm->reset_pending, 0); 372 atomic_set(&pm->reset_counter, 0); 373 374 INIT_WORK(&pm->recovery_work, ivpu_pm_recovery_work); 375 INIT_DELAYED_WORK(&pm->job_timeout_work, ivpu_job_timeout_work); 376 377 if (ivpu_disable_recovery) 378 delay = -1; 379 else 380 delay = vdev->timeout.autosuspend; 381 382 pm_runtime_use_autosuspend(dev); 383 pm_runtime_set_autosuspend_delay(dev, delay); 384 385 ivpu_dbg(vdev, PM, "Autosuspend delay = %d\n", delay); 386 } 387 388 void ivpu_pm_cancel_recovery(struct ivpu_device *vdev) 389 { 390 drm_WARN_ON(&vdev->drm, delayed_work_pending(&vdev->pm->job_timeout_work)); 391 cancel_work_sync(&vdev->pm->recovery_work); 392 } 393 394 void ivpu_pm_enable(struct ivpu_device *vdev) 395 { 396 struct device *dev = vdev->drm.dev; 397 398 pm_runtime_set_active(dev); 399 pm_runtime_allow(dev); 400 pm_runtime_mark_last_busy(dev); 401 pm_runtime_put_autosuspend(dev); 402 } 403 404 void ivpu_pm_disable(struct ivpu_device *vdev) 405 { 406 pm_runtime_get_noresume(vdev->drm.dev); 407 pm_runtime_forbid(vdev->drm.dev); 408 } 409