xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c (revision cb8fef4b07f71da01843457b66e579ad248a625e)
1 /*
2  * Copyright 2021 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  */
23 
24 #include "amdgpu_reset.h"
25 #include "aldebaran.h"
26 #include "sienna_cichlid.h"
27 #include "smu_v13_0_10.h"
28 #include "soc_v1_0.h"
29 
30 static int amdgpu_reset_xgmi_reset_on_init_suspend(struct amdgpu_device *adev)
31 {
32 	int i;
33 
34 	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
35 		if (!adev->ip_blocks[i].status.valid)
36 			continue;
37 		if (!adev->ip_blocks[i].status.hw)
38 			continue;
39 		/* displays are handled in phase1 */
40 		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
41 			continue;
42 
43 		/* XXX handle errors */
44 		amdgpu_ip_block_suspend(&adev->ip_blocks[i]);
45 		adev->ip_blocks[i].status.hw = false;
46 	}
47 
48 	/* VCN FW shared region is in frambuffer, there are some flags
49 	 * initialized in that region during sw_init. Make sure the region is
50 	 * backed up.
51 	 */
52 	amdgpu_vcn_save_vcpu_bo(adev);
53 
54 	return 0;
55 }
56 
57 static int amdgpu_reset_xgmi_reset_on_init_prep_hwctxt(
58 	struct amdgpu_reset_control *reset_ctl,
59 	struct amdgpu_reset_context *reset_context)
60 {
61 	struct list_head *reset_device_list = reset_context->reset_device_list;
62 	struct amdgpu_device *tmp_adev;
63 	int r;
64 
65 	list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
66 		amdgpu_unregister_gpu_instance(tmp_adev);
67 		r = amdgpu_reset_xgmi_reset_on_init_suspend(tmp_adev);
68 		if (r) {
69 			dev_err(tmp_adev->dev,
70 				"xgmi reset on init: prepare for reset failed");
71 			return r;
72 		}
73 	}
74 
75 	return r;
76 }
77 
78 static int amdgpu_reset_xgmi_reset_on_init_restore_hwctxt(
79 	struct amdgpu_reset_control *reset_ctl,
80 	struct amdgpu_reset_context *reset_context)
81 {
82 	struct list_head *reset_device_list = reset_context->reset_device_list;
83 	struct amdgpu_device *tmp_adev = NULL;
84 	int r;
85 
86 	r = amdgpu_device_reinit_after_reset(reset_context);
87 	if (r)
88 		return r;
89 	list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
90 		if (!tmp_adev->kfd.init_complete) {
91 			kgd2kfd_init_zone_device(tmp_adev);
92 			amdgpu_amdkfd_device_init(tmp_adev);
93 			amdgpu_amdkfd_drm_client_create(tmp_adev);
94 			amdgpu_ptl_sysfs_init(tmp_adev);
95 		}
96 	}
97 
98 	return r;
99 }
100 
101 static int amdgpu_reset_xgmi_reset_on_init_perform_reset(
102 	struct amdgpu_reset_control *reset_ctl,
103 	struct amdgpu_reset_context *reset_context)
104 {
105 	struct amdgpu_device *adev = (struct amdgpu_device *)reset_ctl->handle;
106 	struct list_head *reset_device_list = reset_context->reset_device_list;
107 	struct amdgpu_device *tmp_adev = NULL;
108 	int r;
109 
110 	dev_dbg(adev->dev, "xgmi roi - hw reset\n");
111 
112 	list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
113 		mutex_lock(&tmp_adev->reset_cntl->reset_lock);
114 		tmp_adev->reset_cntl->active_reset =
115 			amdgpu_asic_reset_method(adev);
116 	}
117 	r = 0;
118 	/* Mode1 reset needs to be triggered on all devices together */
119 	list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
120 		/* For XGMI run all resets in parallel to speed up the process */
121 		if (!queue_work(system_dfl_wq, &tmp_adev->xgmi_reset_work))
122 			r = -EALREADY;
123 		if (r) {
124 			dev_err(tmp_adev->dev,
125 				"xgmi reset on init: reset failed with error, %d",
126 				r);
127 			break;
128 		}
129 	}
130 
131 	/* For XGMI wait for all resets to complete before proceed */
132 	if (!r) {
133 		list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
134 			flush_work(&tmp_adev->xgmi_reset_work);
135 			r = tmp_adev->asic_reset_res;
136 			if (r)
137 				break;
138 		}
139 	}
140 
141 	list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
142 		mutex_unlock(&tmp_adev->reset_cntl->reset_lock);
143 		tmp_adev->reset_cntl->active_reset = AMD_RESET_METHOD_NONE;
144 	}
145 
146 	return r;
147 }
148 
149 int amdgpu_reset_do_xgmi_reset_on_init(
150 	struct amdgpu_reset_context *reset_context)
151 {
152 	struct list_head *reset_device_list = reset_context->reset_device_list;
153 	struct amdgpu_device *adev;
154 	int r;
155 
156 	if (!reset_device_list || list_empty(reset_device_list) ||
157 	    list_is_singular(reset_device_list))
158 		return -EINVAL;
159 
160 	adev = list_first_entry(reset_device_list, struct amdgpu_device,
161 				reset_list);
162 	r = amdgpu_reset_prepare_hwcontext(adev, reset_context);
163 	if (r)
164 		return r;
165 
166 	r = amdgpu_reset_perform_reset(adev, reset_context);
167 
168 	return r;
169 }
170 
171 struct amdgpu_reset_handler xgmi_reset_on_init_handler = {
172 	.reset_method = AMD_RESET_METHOD_ON_INIT,
173 	.prepare_env = NULL,
174 	.prepare_hwcontext = amdgpu_reset_xgmi_reset_on_init_prep_hwctxt,
175 	.perform_reset = amdgpu_reset_xgmi_reset_on_init_perform_reset,
176 	.restore_hwcontext = amdgpu_reset_xgmi_reset_on_init_restore_hwctxt,
177 	.restore_env = NULL,
178 	.do_reset = NULL,
179 };
180 
181 int amdgpu_reset_init(struct amdgpu_device *adev)
182 {
183 	int ret = 0;
184 
185 	switch (amdgpu_ip_version(adev, MP1_HWIP, 0)) {
186 	case IP_VERSION(13, 0, 2):
187 	case IP_VERSION(13, 0, 6):
188 	case IP_VERSION(13, 0, 12):
189 	case IP_VERSION(13, 0, 14):
190 		ret = aldebaran_reset_init(adev);
191 		break;
192 	case IP_VERSION(11, 0, 7):
193 		ret = sienna_cichlid_reset_init(adev);
194 		break;
195 	case IP_VERSION(13, 0, 10):
196 		ret = smu_v13_0_10_reset_init(adev);
197 		break;
198 	case IP_VERSION(15, 0, 8):
199 		ret = soc_v1_0_reset_init(adev);
200 		break;
201 	default:
202 		break;
203 	}
204 
205 	return ret;
206 }
207 
208 int amdgpu_reset_fini(struct amdgpu_device *adev)
209 {
210 	int ret = 0;
211 
212 	switch (amdgpu_ip_version(adev, MP1_HWIP, 0)) {
213 	case IP_VERSION(13, 0, 2):
214 	case IP_VERSION(13, 0, 6):
215 	case IP_VERSION(13, 0, 12):
216 	case IP_VERSION(13, 0, 14):
217 		ret = aldebaran_reset_fini(adev);
218 		break;
219 	case IP_VERSION(11, 0, 7):
220 		ret = sienna_cichlid_reset_fini(adev);
221 		break;
222 	case IP_VERSION(13, 0, 10):
223 		ret = smu_v13_0_10_reset_fini(adev);
224 		break;
225 	case IP_VERSION(15, 0, 8):
226 		ret = soc_v1_0_reset_fini(adev);
227 		break;
228 	default:
229 		break;
230 	}
231 
232 	return ret;
233 }
234 
235 int amdgpu_reset_prepare_hwcontext(struct amdgpu_device *adev,
236 				   struct amdgpu_reset_context *reset_context)
237 {
238 	struct amdgpu_reset_handler *reset_handler = NULL;
239 
240 	if (adev->reset_cntl && adev->reset_cntl->get_reset_handler)
241 		reset_handler = adev->reset_cntl->get_reset_handler(
242 			adev->reset_cntl, reset_context);
243 	if (!reset_handler)
244 		return -EOPNOTSUPP;
245 
246 	return reset_handler->prepare_hwcontext(adev->reset_cntl,
247 						reset_context);
248 }
249 
250 int amdgpu_reset_perform_reset(struct amdgpu_device *adev,
251 			       struct amdgpu_reset_context *reset_context)
252 {
253 	int ret;
254 	struct amdgpu_reset_handler *reset_handler = NULL;
255 
256 	if (adev->reset_cntl)
257 		reset_handler = adev->reset_cntl->get_reset_handler(
258 			adev->reset_cntl, reset_context);
259 	if (!reset_handler)
260 		return -EOPNOTSUPP;
261 
262 	ret = reset_handler->perform_reset(adev->reset_cntl, reset_context);
263 	if (ret)
264 		return ret;
265 
266 	return reset_handler->restore_hwcontext(adev->reset_cntl,
267 						reset_context);
268 }
269 
270 
271 void amdgpu_reset_destroy_reset_domain(struct kref *ref)
272 {
273 	struct amdgpu_reset_domain *reset_domain = container_of(ref,
274 								struct amdgpu_reset_domain,
275 								refcount);
276 	if (reset_domain->wq)
277 		destroy_workqueue(reset_domain->wq);
278 
279 	kvfree(reset_domain);
280 }
281 
282 struct amdgpu_reset_domain *amdgpu_reset_create_reset_domain(enum amdgpu_reset_domain_type type,
283 							     char *wq_name)
284 {
285 	struct amdgpu_reset_domain *reset_domain;
286 
287 	reset_domain = kvzalloc_obj(struct amdgpu_reset_domain);
288 	if (!reset_domain) {
289 		DRM_ERROR("Failed to allocate amdgpu_reset_domain!");
290 		return NULL;
291 	}
292 
293 	reset_domain->type = type;
294 	kref_init(&reset_domain->refcount);
295 
296 	reset_domain->wq = create_singlethread_workqueue(wq_name);
297 	if (!reset_domain->wq) {
298 		DRM_ERROR("Failed to allocate wq for amdgpu_reset_domain!");
299 		amdgpu_reset_put_reset_domain(reset_domain);
300 		return NULL;
301 
302 	}
303 
304 	atomic_set(&reset_domain->in_gpu_reset, 0);
305 	atomic_set(&reset_domain->reset_res, 0);
306 	init_rwsem(&reset_domain->sem);
307 
308 	return reset_domain;
309 }
310 
311 void amdgpu_device_lock_reset_domain(struct amdgpu_reset_domain *reset_domain)
312 {
313 	atomic_set(&reset_domain->in_gpu_reset, 1);
314 	down_write(&reset_domain->sem);
315 }
316 
317 
318 void amdgpu_device_unlock_reset_domain(struct amdgpu_reset_domain *reset_domain)
319 {
320 	atomic_set(&reset_domain->in_gpu_reset, 0);
321 	up_write(&reset_domain->sem);
322 }
323 
324 void amdgpu_reset_get_desc(struct amdgpu_reset_context *rst_ctxt, char *buf,
325 			   size_t len)
326 {
327 	if (!buf || !len)
328 		return;
329 
330 	switch (rst_ctxt->src) {
331 	case AMDGPU_RESET_SRC_JOB:
332 		if (rst_ctxt->job) {
333 			snprintf(buf, len, "job hang on ring:%s",
334 				 rst_ctxt->job->base.sched->name);
335 		} else {
336 			strscpy(buf, "job hang", len);
337 		}
338 		break;
339 	case AMDGPU_RESET_SRC_RAS:
340 		strscpy(buf, "RAS error", len);
341 		break;
342 	case AMDGPU_RESET_SRC_MES:
343 		strscpy(buf, "MES hang", len);
344 		break;
345 	case AMDGPU_RESET_SRC_HWS:
346 		strscpy(buf, "HWS hang", len);
347 		break;
348 	case AMDGPU_RESET_SRC_USER:
349 		strscpy(buf, "user trigger", len);
350 		break;
351 	case AMDGPU_RESET_SRC_USERQ:
352 		strscpy(buf, "user queue trigger", len);
353 		break;
354 	default:
355 		strscpy(buf, "unknown", len);
356 	}
357 }
358 
359 bool amdgpu_reset_in_recovery(struct amdgpu_device *adev)
360 {
361 	return (adev->init_lvl->level == AMDGPU_INIT_LEVEL_RESET_RECOVERY);
362 }
363