xref: /linux/drivers/accel/ivpu/ivpu_pm.c (revision 3534eacb)
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_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 	ivpu_prepare_for_reset(vdev);
62 
63 	ret = ivpu_shutdown(vdev);
64 	if (ret)
65 		ivpu_err(vdev, "Failed to shutdown VPU: %d\n", ret);
66 
67 	return ret;
68 }
69 
70 static int ivpu_resume(struct ivpu_device *vdev)
71 {
72 	int ret;
73 
74 retry:
75 	pci_restore_state(to_pci_dev(vdev->drm.dev));
76 	pci_set_power_state(to_pci_dev(vdev->drm.dev), PCI_D0);
77 
78 	ret = ivpu_hw_power_up(vdev);
79 	if (ret) {
80 		ivpu_err(vdev, "Failed to power up HW: %d\n", ret);
81 		goto err_power_down;
82 	}
83 
84 	ret = ivpu_mmu_enable(vdev);
85 	if (ret) {
86 		ivpu_err(vdev, "Failed to resume MMU: %d\n", ret);
87 		goto err_power_down;
88 	}
89 
90 	ret = ivpu_boot(vdev);
91 	if (ret)
92 		goto err_mmu_disable;
93 
94 	return 0;
95 
96 err_mmu_disable:
97 	ivpu_mmu_disable(vdev);
98 err_power_down:
99 	ivpu_hw_power_down(vdev);
100 
101 	if (!ivpu_fw_is_cold_boot(vdev)) {
102 		ivpu_pm_prepare_cold_boot(vdev);
103 		goto retry;
104 	} else {
105 		ivpu_err(vdev, "Failed to resume the FW: %d\n", ret);
106 	}
107 
108 	return ret;
109 }
110 
111 static void ivpu_pm_recovery_work(struct work_struct *work)
112 {
113 	struct ivpu_pm_info *pm = container_of(work, struct ivpu_pm_info, recovery_work);
114 	struct ivpu_device *vdev = pm->vdev;
115 	char *evt[2] = {"IVPU_PM_EVENT=IVPU_RECOVER", NULL};
116 	int ret;
117 
118 	ivpu_err(vdev, "Recovering the NPU (reset #%d)\n", atomic_read(&vdev->pm->reset_counter));
119 
120 	ret = pm_runtime_resume_and_get(vdev->drm.dev);
121 	if (ret)
122 		ivpu_err(vdev, "Failed to resume NPU: %d\n", ret);
123 
124 	ivpu_fw_log_dump(vdev);
125 
126 	atomic_inc(&vdev->pm->reset_counter);
127 	atomic_set(&vdev->pm->reset_pending, 1);
128 	down_write(&vdev->pm->reset_lock);
129 
130 	ivpu_suspend(vdev);
131 	ivpu_pm_prepare_cold_boot(vdev);
132 	ivpu_jobs_abort_all(vdev);
133 
134 	ret = ivpu_resume(vdev);
135 	if (ret)
136 		ivpu_err(vdev, "Failed to resume NPU: %d\n", ret);
137 
138 	up_write(&vdev->pm->reset_lock);
139 	atomic_set(&vdev->pm->reset_pending, 0);
140 
141 	kobject_uevent_env(&vdev->drm.dev->kobj, KOBJ_CHANGE, evt);
142 	pm_runtime_mark_last_busy(vdev->drm.dev);
143 	pm_runtime_put_autosuspend(vdev->drm.dev);
144 }
145 
146 void ivpu_pm_trigger_recovery(struct ivpu_device *vdev, const char *reason)
147 {
148 	ivpu_err(vdev, "Recovery triggered by %s\n", reason);
149 
150 	if (ivpu_disable_recovery) {
151 		ivpu_err(vdev, "Recovery not available when disable_recovery param is set\n");
152 		return;
153 	}
154 
155 	if (ivpu_is_fpga(vdev)) {
156 		ivpu_err(vdev, "Recovery not available on FPGA\n");
157 		return;
158 	}
159 
160 	/* Trigger recovery if it's not in progress */
161 	if (atomic_cmpxchg(&vdev->pm->reset_pending, 0, 1) == 0) {
162 		ivpu_hw_diagnose_failure(vdev);
163 		ivpu_hw_irq_disable(vdev); /* Disable IRQ early to protect from IRQ storm */
164 		queue_work(system_long_wq, &vdev->pm->recovery_work);
165 	}
166 }
167 
168 static void ivpu_job_timeout_work(struct work_struct *work)
169 {
170 	struct ivpu_pm_info *pm = container_of(work, struct ivpu_pm_info, job_timeout_work.work);
171 	struct ivpu_device *vdev = pm->vdev;
172 
173 	ivpu_pm_trigger_recovery(vdev, "TDR");
174 }
175 
176 void ivpu_start_job_timeout_detection(struct ivpu_device *vdev)
177 {
178 	unsigned long timeout_ms = ivpu_tdr_timeout_ms ? ivpu_tdr_timeout_ms : vdev->timeout.tdr;
179 
180 	/* No-op if already queued */
181 	queue_delayed_work(system_wq, &vdev->pm->job_timeout_work, msecs_to_jiffies(timeout_ms));
182 }
183 
184 void ivpu_stop_job_timeout_detection(struct ivpu_device *vdev)
185 {
186 	cancel_delayed_work_sync(&vdev->pm->job_timeout_work);
187 }
188 
189 int ivpu_pm_suspend_cb(struct device *dev)
190 {
191 	struct drm_device *drm = dev_get_drvdata(dev);
192 	struct ivpu_device *vdev = to_ivpu_device(drm);
193 	unsigned long timeout;
194 
195 	ivpu_dbg(vdev, PM, "Suspend..\n");
196 
197 	timeout = jiffies + msecs_to_jiffies(vdev->timeout.tdr);
198 	while (!ivpu_hw_is_idle(vdev)) {
199 		cond_resched();
200 		if (time_after_eq(jiffies, timeout)) {
201 			ivpu_err(vdev, "Failed to enter idle on system suspend\n");
202 			return -EBUSY;
203 		}
204 	}
205 
206 	ivpu_jsm_pwr_d0i3_enter(vdev);
207 
208 	ivpu_suspend(vdev);
209 	ivpu_pm_prepare_warm_boot(vdev);
210 
211 	ivpu_dbg(vdev, PM, "Suspend done.\n");
212 
213 	return 0;
214 }
215 
216 int ivpu_pm_resume_cb(struct device *dev)
217 {
218 	struct drm_device *drm = dev_get_drvdata(dev);
219 	struct ivpu_device *vdev = to_ivpu_device(drm);
220 	int ret;
221 
222 	ivpu_dbg(vdev, PM, "Resume..\n");
223 
224 	ret = ivpu_resume(vdev);
225 	if (ret)
226 		ivpu_err(vdev, "Failed to resume: %d\n", ret);
227 
228 	ivpu_dbg(vdev, PM, "Resume done.\n");
229 
230 	return ret;
231 }
232 
233 int ivpu_pm_runtime_suspend_cb(struct device *dev)
234 {
235 	struct drm_device *drm = dev_get_drvdata(dev);
236 	struct ivpu_device *vdev = to_ivpu_device(drm);
237 	bool hw_is_idle = true;
238 	int ret;
239 
240 	drm_WARN_ON(&vdev->drm, !xa_empty(&vdev->submitted_jobs_xa));
241 	drm_WARN_ON(&vdev->drm, work_pending(&vdev->pm->recovery_work));
242 
243 	ivpu_dbg(vdev, PM, "Runtime suspend..\n");
244 
245 	if (!ivpu_hw_is_idle(vdev) && vdev->pm->suspend_reschedule_counter) {
246 		ivpu_dbg(vdev, PM, "Failed to enter idle, rescheduling suspend, retries left %d\n",
247 			 vdev->pm->suspend_reschedule_counter);
248 		pm_schedule_suspend(dev, vdev->timeout.reschedule_suspend);
249 		vdev->pm->suspend_reschedule_counter--;
250 		return -EAGAIN;
251 	}
252 
253 	if (!vdev->pm->suspend_reschedule_counter)
254 		hw_is_idle = false;
255 	else if (ivpu_jsm_pwr_d0i3_enter(vdev))
256 		hw_is_idle = false;
257 
258 	ret = ivpu_suspend(vdev);
259 	if (ret)
260 		ivpu_err(vdev, "Failed to suspend NPU: %d\n", ret);
261 
262 	if (!hw_is_idle) {
263 		ivpu_err(vdev, "NPU failed to enter idle, force suspended.\n");
264 		ivpu_fw_log_dump(vdev);
265 		ivpu_pm_prepare_cold_boot(vdev);
266 	} else {
267 		ivpu_pm_prepare_warm_boot(vdev);
268 	}
269 
270 	vdev->pm->suspend_reschedule_counter = PM_RESCHEDULE_LIMIT;
271 
272 	ivpu_dbg(vdev, PM, "Runtime suspend done.\n");
273 
274 	return 0;
275 }
276 
277 int ivpu_pm_runtime_resume_cb(struct device *dev)
278 {
279 	struct drm_device *drm = dev_get_drvdata(dev);
280 	struct ivpu_device *vdev = to_ivpu_device(drm);
281 	int ret;
282 
283 	ivpu_dbg(vdev, PM, "Runtime resume..\n");
284 
285 	ret = ivpu_resume(vdev);
286 	if (ret)
287 		ivpu_err(vdev, "Failed to set RESUME state: %d\n", ret);
288 
289 	ivpu_dbg(vdev, PM, "Runtime resume done.\n");
290 
291 	return ret;
292 }
293 
294 int ivpu_rpm_get(struct ivpu_device *vdev)
295 {
296 	int ret;
297 
298 	ret = pm_runtime_resume_and_get(vdev->drm.dev);
299 	if (!drm_WARN_ON(&vdev->drm, ret < 0))
300 		vdev->pm->suspend_reschedule_counter = PM_RESCHEDULE_LIMIT;
301 
302 	return ret;
303 }
304 
305 int ivpu_rpm_get_if_active(struct ivpu_device *vdev)
306 {
307 	int ret;
308 
309 	ret = pm_runtime_get_if_in_use(vdev->drm.dev);
310 	drm_WARN_ON(&vdev->drm, ret < 0);
311 
312 	return ret;
313 }
314 
315 void ivpu_rpm_put(struct ivpu_device *vdev)
316 {
317 	pm_runtime_mark_last_busy(vdev->drm.dev);
318 	pm_runtime_put_autosuspend(vdev->drm.dev);
319 }
320 
321 void ivpu_pm_reset_prepare_cb(struct pci_dev *pdev)
322 {
323 	struct ivpu_device *vdev = pci_get_drvdata(pdev);
324 
325 	ivpu_dbg(vdev, PM, "Pre-reset..\n");
326 	atomic_inc(&vdev->pm->reset_counter);
327 	atomic_set(&vdev->pm->reset_pending, 1);
328 
329 	pm_runtime_get_sync(vdev->drm.dev);
330 	down_write(&vdev->pm->reset_lock);
331 	ivpu_prepare_for_reset(vdev);
332 	ivpu_hw_reset(vdev);
333 	ivpu_pm_prepare_cold_boot(vdev);
334 	ivpu_jobs_abort_all(vdev);
335 	ivpu_dbg(vdev, PM, "Pre-reset done.\n");
336 }
337 
338 void ivpu_pm_reset_done_cb(struct pci_dev *pdev)
339 {
340 	struct ivpu_device *vdev = pci_get_drvdata(pdev);
341 	int ret;
342 
343 	ivpu_dbg(vdev, PM, "Post-reset..\n");
344 	ret = ivpu_resume(vdev);
345 	if (ret)
346 		ivpu_err(vdev, "Failed to set RESUME state: %d\n", ret);
347 	up_write(&vdev->pm->reset_lock);
348 	atomic_set(&vdev->pm->reset_pending, 0);
349 	ivpu_dbg(vdev, PM, "Post-reset done.\n");
350 
351 	pm_runtime_mark_last_busy(vdev->drm.dev);
352 	pm_runtime_put_autosuspend(vdev->drm.dev);
353 }
354 
355 void ivpu_pm_init(struct ivpu_device *vdev)
356 {
357 	struct device *dev = vdev->drm.dev;
358 	struct ivpu_pm_info *pm = vdev->pm;
359 	int delay;
360 
361 	pm->vdev = vdev;
362 	pm->suspend_reschedule_counter = PM_RESCHEDULE_LIMIT;
363 
364 	init_rwsem(&pm->reset_lock);
365 	atomic_set(&pm->reset_pending, 0);
366 	atomic_set(&pm->reset_counter, 0);
367 
368 	INIT_WORK(&pm->recovery_work, ivpu_pm_recovery_work);
369 	INIT_DELAYED_WORK(&pm->job_timeout_work, ivpu_job_timeout_work);
370 
371 	if (ivpu_disable_recovery)
372 		delay = -1;
373 	else
374 		delay = vdev->timeout.autosuspend;
375 
376 	pm_runtime_use_autosuspend(dev);
377 	pm_runtime_set_autosuspend_delay(dev, delay);
378 
379 	ivpu_dbg(vdev, PM, "Autosuspend delay = %d\n", delay);
380 }
381 
382 void ivpu_pm_cancel_recovery(struct ivpu_device *vdev)
383 {
384 	drm_WARN_ON(&vdev->drm, delayed_work_pending(&vdev->pm->job_timeout_work));
385 	cancel_work_sync(&vdev->pm->recovery_work);
386 }
387 
388 void ivpu_pm_enable(struct ivpu_device *vdev)
389 {
390 	struct device *dev = vdev->drm.dev;
391 
392 	pm_runtime_set_active(dev);
393 	pm_runtime_allow(dev);
394 	pm_runtime_mark_last_busy(dev);
395 	pm_runtime_put_autosuspend(dev);
396 }
397 
398 void ivpu_pm_disable(struct ivpu_device *vdev)
399 {
400 	pm_runtime_get_noresume(vdev->drm.dev);
401 	pm_runtime_forbid(vdev->drm.dev);
402 }
403