Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.1.
  1// SPDX-License-Identifier: MIT
  2/*
  3 * Copyright © 2022 Intel Corporation
  4 */
  5
  6#include "xe_pm.h"
  7
  8#include <linux/pm_runtime.h>
  9
 10#include <drm/drm_managed.h>
 11#include <drm/ttm/ttm_placement.h>
 12
 13#include "xe_bo.h"
 14#include "xe_bo_evict.h"
 15#include "xe_device.h"
 16#include "xe_device_sysfs.h"
 17#include "xe_display.h"
 18#include "xe_ggtt.h"
 19#include "xe_gt.h"
 20#include "xe_guc.h"
 21#include "xe_irq.h"
 22#include "xe_pcode.h"
 23#include "xe_wa.h"
 24
 25/**
 26 * DOC: Xe Power Management
 27 *
 28 * Xe PM shall be guided by the simplicity.
 29 * Use the simplest hook options whenever possible.
 30 * Let's not reinvent the runtime_pm references and hooks.
 31 * Shall have a clear separation of display and gt underneath this component.
 32 *
 33 * What's next:
 34 *
 35 * For now s2idle and s3 are only working in integrated devices. The next step
 36 * is to iterate through all VRAM's BO backing them up into the system memory
 37 * before allowing the system suspend.
 38 *
 39 * Also runtime_pm needs to be here from the beginning.
 40 *
 41 * RC6/RPS are also critical PM features. Let's start with GuCRC and GuC SLPC
 42 * and no wait boost. Frequency optimizations should come on a next stage.
 43 */
 44
 45/**
 46 * xe_pm_suspend - Helper for System suspend, i.e. S0->S3 / S0->S2idle
 47 * @xe: xe device instance
 48 *
 49 * Return: 0 on success
 50 */
 51int xe_pm_suspend(struct xe_device *xe)
 52{
 53	struct xe_gt *gt;
 54	u8 id;
 55	int err;
 56
 57	for_each_gt(gt, xe, id)
 58		xe_gt_suspend_prepare(gt);
 59
 60	/* FIXME: Super racey... */
 61	err = xe_bo_evict_all(xe);
 62	if (err)
 63		return err;
 64
 65	xe_display_pm_suspend(xe);
 66
 67	for_each_gt(gt, xe, id) {
 68		err = xe_gt_suspend(gt);
 69		if (err) {
 70			xe_display_pm_resume(xe);
 71			return err;
 72		}
 73	}
 74
 75	xe_irq_suspend(xe);
 76
 77	xe_display_pm_suspend_late(xe);
 78
 79	return 0;
 80}
 81
 82/**
 83 * xe_pm_resume - Helper for System resume S3->S0 / S2idle->S0
 84 * @xe: xe device instance
 85 *
 86 * Return: 0 on success
 87 */
 88int xe_pm_resume(struct xe_device *xe)
 89{
 90	struct xe_tile *tile;
 91	struct xe_gt *gt;
 92	u8 id;
 93	int err;
 94
 95	for_each_tile(tile, xe, id)
 96		xe_wa_apply_tile_workarounds(tile);
 97
 98	for_each_gt(gt, xe, id) {
 99		err = xe_pcode_init(gt);
100		if (err)
101			return err;
102	}
103
104	xe_display_pm_resume_early(xe);
105
106	/*
107	 * This only restores pinned memory which is the memory required for the
108	 * GT(s) to resume.
109	 */
110	err = xe_bo_restore_kernel(xe);
111	if (err)
112		return err;
113
114	xe_irq_resume(xe);
115
116	xe_display_pm_resume(xe);
117
118	for_each_gt(gt, xe, id)
119		xe_gt_resume(gt);
120
121	err = xe_bo_restore_user(xe);
122	if (err)
123		return err;
124
125	return 0;
126}
127
128static bool xe_pm_pci_d3cold_capable(struct pci_dev *pdev)
129{
130	struct pci_dev *root_pdev;
131
132	root_pdev = pcie_find_root_port(pdev);
133	if (!root_pdev)
134		return false;
135
136	/* D3Cold requires PME capability and _PR3 power resource */
137	if (!pci_pme_capable(root_pdev, PCI_D3cold) || !pci_pr3_present(root_pdev))
138		return false;
139
140	return true;
141}
142
143static void xe_pm_runtime_init(struct xe_device *xe)
144{
145	struct device *dev = xe->drm.dev;
146
147	/*
148	 * Disable the system suspend direct complete optimization.
149	 * We need to ensure that the regular device suspend/resume functions
150	 * are called since our runtime_pm cannot guarantee local memory
151	 * eviction for d3cold.
152	 * TODO: Check HDA audio dependencies claimed by i915, and then enforce
153	 *       this option to integrated graphics as well.
154	 */
155	if (IS_DGFX(xe))
156		dev_pm_set_driver_flags(dev, DPM_FLAG_NO_DIRECT_COMPLETE);
157
158	pm_runtime_use_autosuspend(dev);
159	pm_runtime_set_autosuspend_delay(dev, 1000);
160	pm_runtime_set_active(dev);
161	pm_runtime_allow(dev);
162	pm_runtime_mark_last_busy(dev);
163	pm_runtime_put(dev);
164}
165
166void xe_pm_init(struct xe_device *xe)
167{
168	struct pci_dev *pdev = to_pci_dev(xe->drm.dev);
169
170	/* For now suspend/resume is only allowed with GuC */
171	if (!xe_device_uc_enabled(xe))
172		return;
173
174	drmm_mutex_init(&xe->drm, &xe->d3cold.lock);
175
176	xe->d3cold.capable = xe_pm_pci_d3cold_capable(pdev);
177
178	if (xe->d3cold.capable) {
179		xe_device_sysfs_init(xe);
180		xe_pm_set_vram_threshold(xe, DEFAULT_VRAM_THRESHOLD);
181	}
182
183	xe_pm_runtime_init(xe);
184}
185
186void xe_pm_runtime_fini(struct xe_device *xe)
187{
188	struct device *dev = xe->drm.dev;
189
190	pm_runtime_get_sync(dev);
191	pm_runtime_forbid(dev);
192}
193
194static void xe_pm_write_callback_task(struct xe_device *xe,
195				      struct task_struct *task)
196{
197	WRITE_ONCE(xe->pm_callback_task, task);
198
199	/*
200	 * Just in case it's somehow possible for our writes to be reordered to
201	 * the extent that something else re-uses the task written in
202	 * pm_callback_task. For example after returning from the callback, but
203	 * before the reordered write that resets pm_callback_task back to NULL.
204	 */
205	smp_mb(); /* pairs with xe_pm_read_callback_task */
206}
207
208struct task_struct *xe_pm_read_callback_task(struct xe_device *xe)
209{
210	smp_mb(); /* pairs with xe_pm_write_callback_task */
211
212	return READ_ONCE(xe->pm_callback_task);
213}
214
215int xe_pm_runtime_suspend(struct xe_device *xe)
216{
217	struct xe_gt *gt;
218	u8 id;
219	int err = 0;
220
221	if (xe->d3cold.allowed && xe_device_mem_access_ongoing(xe))
222		return -EBUSY;
223
224	/* Disable access_ongoing asserts and prevent recursive pm calls */
225	xe_pm_write_callback_task(xe, current);
226
227	/*
228	 * The actual xe_device_mem_access_put() is always async underneath, so
229	 * exactly where that is called should makes no difference to us. However
230	 * we still need to be very careful with the locks that this callback
231	 * acquires and the locks that are acquired and held by any callers of
232	 * xe_device_mem_access_get(). We already have the matching annotation
233	 * on that side, but we also need it here. For example lockdep should be
234	 * able to tell us if the following scenario is in theory possible:
235	 *
236	 * CPU0                          | CPU1 (kworker)
237	 * lock(A)                       |
238	 *                               | xe_pm_runtime_suspend()
239	 *                               |      lock(A)
240	 * xe_device_mem_access_get()    |
241	 *
242	 * This will clearly deadlock since rpm core needs to wait for
243	 * xe_pm_runtime_suspend() to complete, but here we are holding lock(A)
244	 * on CPU0 which prevents CPU1 making forward progress.  With the
245	 * annotation here and in xe_device_mem_access_get() lockdep will see
246	 * the potential lock inversion and give us a nice splat.
247	 */
248	lock_map_acquire(&xe_device_mem_access_lockdep_map);
249
250	if (xe->d3cold.allowed) {
251		err = xe_bo_evict_all(xe);
252		if (err)
253			goto out;
254	}
255
256	for_each_gt(gt, xe, id) {
257		err = xe_gt_suspend(gt);
258		if (err)
259			goto out;
260	}
261
262	xe_irq_suspend(xe);
263out:
264	lock_map_release(&xe_device_mem_access_lockdep_map);
265	xe_pm_write_callback_task(xe, NULL);
266	return err;
267}
268
269int xe_pm_runtime_resume(struct xe_device *xe)
270{
271	struct xe_gt *gt;
272	u8 id;
273	int err = 0;
274
275	/* Disable access_ongoing asserts and prevent recursive pm calls */
276	xe_pm_write_callback_task(xe, current);
277
278	lock_map_acquire(&xe_device_mem_access_lockdep_map);
279
280	/*
281	 * It can be possible that xe has allowed d3cold but other pcie devices
282	 * in gfx card soc would have blocked d3cold, therefore card has not
283	 * really lost power. Detecting primary Gt power is sufficient.
284	 */
285	gt = xe_device_get_gt(xe, 0);
286	xe->d3cold.power_lost = xe_guc_in_reset(&gt->uc.guc);
287
288	if (xe->d3cold.allowed && xe->d3cold.power_lost) {
289		for_each_gt(gt, xe, id) {
290			err = xe_pcode_init(gt);
291			if (err)
292				goto out;
293		}
294
295		/*
296		 * This only restores pinned memory which is the memory
297		 * required for the GT(s) to resume.
298		 */
299		err = xe_bo_restore_kernel(xe);
300		if (err)
301			goto out;
302	}
303
304	xe_irq_resume(xe);
305
306	for_each_gt(gt, xe, id)
307		xe_gt_resume(gt);
308
309	if (xe->d3cold.allowed && xe->d3cold.power_lost) {
310		err = xe_bo_restore_user(xe);
311		if (err)
312			goto out;
313	}
314out:
315	lock_map_release(&xe_device_mem_access_lockdep_map);
316	xe_pm_write_callback_task(xe, NULL);
317	return err;
318}
319
320int xe_pm_runtime_get(struct xe_device *xe)
321{
322	return pm_runtime_get_sync(xe->drm.dev);
323}
324
325int xe_pm_runtime_put(struct xe_device *xe)
326{
327	pm_runtime_mark_last_busy(xe->drm.dev);
328	return pm_runtime_put(xe->drm.dev);
329}
330
331int xe_pm_runtime_get_if_active(struct xe_device *xe)
332{
333	return pm_runtime_get_if_active(xe->drm.dev, true);
334}
335
336void xe_pm_assert_unbounded_bridge(struct xe_device *xe)
337{
338	struct pci_dev *pdev = to_pci_dev(xe->drm.dev);
339	struct pci_dev *bridge = pci_upstream_bridge(pdev);
340
341	if (!bridge)
342		return;
343
344	if (!bridge->driver) {
345		drm_warn(&xe->drm, "unbounded parent pci bridge, device won't support any PM support.\n");
346		device_set_pm_not_required(&pdev->dev);
347	}
348}
349
350int xe_pm_set_vram_threshold(struct xe_device *xe, u32 threshold)
351{
352	struct ttm_resource_manager *man;
353	u32 vram_total_mb = 0;
354	int i;
355
356	for (i = XE_PL_VRAM0; i <= XE_PL_VRAM1; ++i) {
357		man = ttm_manager_type(&xe->ttm, i);
358		if (man)
359			vram_total_mb += DIV_ROUND_UP_ULL(man->size, 1024 * 1024);
360	}
361
362	drm_dbg(&xe->drm, "Total vram %u mb\n", vram_total_mb);
363
364	if (threshold > vram_total_mb)
365		return -EINVAL;
366
367	mutex_lock(&xe->d3cold.lock);
368	xe->d3cold.vram_threshold = threshold;
369	mutex_unlock(&xe->d3cold.lock);
370
371	return 0;
372}
373
374void xe_pm_d3cold_allowed_toggle(struct xe_device *xe)
375{
376	struct ttm_resource_manager *man;
377	u32 total_vram_used_mb = 0;
378	u64 vram_used;
379	int i;
380
381	if (!xe->d3cold.capable) {
382		xe->d3cold.allowed = false;
383		return;
384	}
385
386	for (i = XE_PL_VRAM0; i <= XE_PL_VRAM1; ++i) {
387		man = ttm_manager_type(&xe->ttm, i);
388		if (man) {
389			vram_used = ttm_resource_manager_usage(man);
390			total_vram_used_mb += DIV_ROUND_UP_ULL(vram_used, 1024 * 1024);
391		}
392	}
393
394	mutex_lock(&xe->d3cold.lock);
395
396	if (total_vram_used_mb < xe->d3cold.vram_threshold)
397		xe->d3cold.allowed = true;
398	else
399		xe->d3cold.allowed = false;
400
401	mutex_unlock(&xe->d3cold.lock);
402
403	drm_dbg(&xe->drm,
404		"d3cold: allowed=%s\n", str_yes_no(xe->d3cold.allowed));
405}