xref: /linux/drivers/gpu/drm/i915/gem/i915_gem_pm.c (revision b7e1e969c887c897947fdc3754fe9b0c24acb155)
1 /*
2  * SPDX-License-Identifier: MIT
3  *
4  * Copyright © 2019 Intel Corporation
5  */
6 
7 #include "gem/i915_gem_pm.h"
8 #include "gem/i915_gem_ttm_pm.h"
9 #include "gt/intel_gt.h"
10 #include "gt/intel_gt_pm.h"
11 #include "gt/intel_gt_requests.h"
12 
13 #include "i915_driver.h"
14 #include "i915_drv.h"
15 
16 #if defined(CONFIG_X86)
17 #include <asm/smp.h>
18 #else
19 #define wbinvd_on_all_cpus() \
20 	pr_warn(DRIVER_NAME ": Missing cache flush in %s\n", __func__)
21 #endif
22 
i915_gem_suspend(struct drm_i915_private * i915)23 void i915_gem_suspend(struct drm_i915_private *i915)
24 {
25 	struct intel_gt *gt;
26 	unsigned int i;
27 
28 	GEM_TRACE("%s\n", dev_name(i915->drm.dev));
29 
30 	intel_wakeref_auto(&i915->runtime_pm.userfault_wakeref, 0);
31 	/*
32 	 * On rare occasions, we've observed the fence completion triggers
33 	 * free_engines asynchronously via rcu_call. Ensure those are done.
34 	 * This path is only called on suspend, so it's an acceptable cost.
35 	 */
36 	rcu_barrier();
37 
38 	flush_workqueue(i915->wq);
39 
40 	/*
41 	 * We have to flush all the executing contexts to main memory so
42 	 * that they can saved in the hibernation image. To ensure the last
43 	 * context image is coherent, we have to switch away from it. That
44 	 * leaves the i915->kernel_context still active when
45 	 * we actually suspend, and its image in memory may not match the GPU
46 	 * state. Fortunately, the kernel_context is disposable and we do
47 	 * not rely on its state.
48 	 */
49 	for_each_gt(gt, i915, i)
50 		intel_gt_suspend_prepare(gt);
51 
52 	i915_gem_drain_freed_objects(i915);
53 }
54 
lmem_restore(struct drm_i915_private * i915,u32 flags)55 static int lmem_restore(struct drm_i915_private *i915, u32 flags)
56 {
57 	struct intel_memory_region *mr;
58 	int ret = 0, id;
59 
60 	for_each_memory_region(mr, i915, id) {
61 		if (mr->type == INTEL_MEMORY_LOCAL) {
62 			ret = i915_ttm_restore_region(mr, flags);
63 			if (ret)
64 				break;
65 		}
66 	}
67 
68 	return ret;
69 }
70 
lmem_suspend(struct drm_i915_private * i915,u32 flags)71 static int lmem_suspend(struct drm_i915_private *i915, u32 flags)
72 {
73 	struct intel_memory_region *mr;
74 	int ret = 0, id;
75 
76 	for_each_memory_region(mr, i915, id) {
77 		if (mr->type == INTEL_MEMORY_LOCAL) {
78 			ret = i915_ttm_backup_region(mr, flags);
79 			if (ret)
80 				break;
81 		}
82 	}
83 
84 	return ret;
85 }
86 
lmem_recover(struct drm_i915_private * i915)87 static void lmem_recover(struct drm_i915_private *i915)
88 {
89 	struct intel_memory_region *mr;
90 	int id;
91 
92 	for_each_memory_region(mr, i915, id)
93 		if (mr->type == INTEL_MEMORY_LOCAL)
94 			i915_ttm_recover_region(mr);
95 }
96 
i915_gem_backup_suspend(struct drm_i915_private * i915)97 int i915_gem_backup_suspend(struct drm_i915_private *i915)
98 {
99 	int ret;
100 
101 	/* Opportunistically try to evict unpinned objects */
102 	ret = lmem_suspend(i915, I915_TTM_BACKUP_ALLOW_GPU);
103 	if (ret)
104 		goto out_recover;
105 
106 	i915_gem_suspend(i915);
107 
108 	/*
109 	 * More objects may have become unpinned as requests were
110 	 * retired. Now try to evict again. The gt may be wedged here
111 	 * in which case we automatically fall back to memcpy.
112 	 * We allow also backing up pinned objects that have not been
113 	 * marked for early recover, and that may contain, for example,
114 	 * page-tables for the migrate context.
115 	 */
116 	ret = lmem_suspend(i915, I915_TTM_BACKUP_ALLOW_GPU |
117 			   I915_TTM_BACKUP_PINNED);
118 	if (ret)
119 		goto out_recover;
120 
121 	/*
122 	 * Remaining objects are backed up using memcpy once we've stopped
123 	 * using the migrate context.
124 	 */
125 	ret = lmem_suspend(i915, I915_TTM_BACKUP_PINNED);
126 	if (ret)
127 		goto out_recover;
128 
129 	return 0;
130 
131 out_recover:
132 	lmem_recover(i915);
133 
134 	return ret;
135 }
136 
i915_gem_suspend_late(struct drm_i915_private * i915)137 void i915_gem_suspend_late(struct drm_i915_private *i915)
138 {
139 	struct drm_i915_gem_object *obj;
140 	struct list_head *phases[] = {
141 		&i915->mm.shrink_list,
142 		&i915->mm.purge_list,
143 		NULL
144 	}, **phase;
145 	struct intel_gt *gt;
146 	unsigned long flags;
147 	unsigned int i;
148 	bool flush = false;
149 
150 	/*
151 	 * Neither the BIOS, ourselves or any other kernel
152 	 * expects the system to be in execlists mode on startup,
153 	 * so we need to reset the GPU back to legacy mode. And the only
154 	 * known way to disable logical contexts is through a GPU reset.
155 	 *
156 	 * So in order to leave the system in a known default configuration,
157 	 * always reset the GPU upon unload and suspend. Afterwards we then
158 	 * clean up the GEM state tracking, flushing off the requests and
159 	 * leaving the system in a known idle state.
160 	 *
161 	 * Note that is of the upmost importance that the GPU is idle and
162 	 * all stray writes are flushed *before* we dismantle the backing
163 	 * storage for the pinned objects.
164 	 *
165 	 * However, since we are uncertain that resetting the GPU on older
166 	 * machines is a good idea, we don't - just in case it leaves the
167 	 * machine in an unusable condition.
168 	 */
169 
170 	/* Like i915_gem_suspend, flush tasks staged from fence triggers */
171 	rcu_barrier();
172 
173 	for_each_gt(gt, i915, i)
174 		intel_gt_suspend_late(gt);
175 
176 	spin_lock_irqsave(&i915->mm.obj_lock, flags);
177 	for (phase = phases; *phase; phase++) {
178 		list_for_each_entry(obj, *phase, mm.link) {
179 			if (!(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ))
180 				flush |= (obj->read_domains & I915_GEM_DOMAIN_CPU) == 0;
181 			__start_cpu_write(obj); /* presume auto-hibernate */
182 		}
183 	}
184 	spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
185 	if (flush)
186 		wbinvd_on_all_cpus();
187 }
188 
i915_gem_freeze(struct drm_i915_private * i915)189 int i915_gem_freeze(struct drm_i915_private *i915)
190 {
191 	/* Discard all purgeable objects, let userspace recover those as
192 	 * required after resuming.
193 	 */
194 	i915_gem_shrink_all(i915);
195 
196 	return 0;
197 }
198 
i915_gem_freeze_late(struct drm_i915_private * i915)199 int i915_gem_freeze_late(struct drm_i915_private *i915)
200 {
201 	struct drm_i915_gem_object *obj;
202 	intel_wakeref_t wakeref;
203 
204 	/*
205 	 * Called just before we write the hibernation image.
206 	 *
207 	 * We need to update the domain tracking to reflect that the CPU
208 	 * will be accessing all the pages to create and restore from the
209 	 * hibernation, and so upon restoration those pages will be in the
210 	 * CPU domain.
211 	 *
212 	 * To make sure the hibernation image contains the latest state,
213 	 * we update that state just before writing out the image.
214 	 *
215 	 * To try and reduce the hibernation image, we manually shrink
216 	 * the objects as well, see i915_gem_freeze()
217 	 */
218 
219 	with_intel_runtime_pm(&i915->runtime_pm, wakeref)
220 		i915_gem_shrink(NULL, i915, -1UL, NULL, ~0);
221 	i915_gem_drain_freed_objects(i915);
222 
223 	wbinvd_on_all_cpus();
224 	list_for_each_entry(obj, &i915->mm.shrink_list, mm.link)
225 		__start_cpu_write(obj);
226 
227 	return 0;
228 }
229 
i915_gem_resume(struct drm_i915_private * i915)230 void i915_gem_resume(struct drm_i915_private *i915)
231 {
232 	struct intel_gt *gt;
233 	int ret, i, j;
234 
235 	GEM_TRACE("%s\n", dev_name(i915->drm.dev));
236 
237 	ret = lmem_restore(i915, 0);
238 	GEM_WARN_ON(ret);
239 
240 	/*
241 	 * As we didn't flush the kernel context before suspend, we cannot
242 	 * guarantee that the context image is complete. So let's just reset
243 	 * it and start again.
244 	 */
245 	for_each_gt(gt, i915, i)
246 		if (intel_gt_resume(gt))
247 			goto err_wedged;
248 
249 	ret = lmem_restore(i915, I915_TTM_BACKUP_ALLOW_GPU);
250 	GEM_WARN_ON(ret);
251 
252 	return;
253 
254 err_wedged:
255 	for_each_gt(gt, i915, j) {
256 		if (!intel_gt_is_wedged(gt)) {
257 			dev_err(i915->drm.dev,
258 				"Failed to re-initialize GPU[%u], declaring it wedged!\n",
259 				j);
260 			intel_gt_set_wedged(gt);
261 		}
262 
263 		if (j == i)
264 			break;
265 	}
266 }
267