Loading...
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 <linux/devcoredump.h>
25#include <generated/utsrelease.h>
26
27#include "amdgpu_reset.h"
28#include "aldebaran.h"
29#include "sienna_cichlid.h"
30#include "smu_v13_0_10.h"
31
32int amdgpu_reset_init(struct amdgpu_device *adev)
33{
34 int ret = 0;
35
36 switch (amdgpu_ip_version(adev, MP1_HWIP, 0)) {
37 case IP_VERSION(13, 0, 2):
38 case IP_VERSION(13, 0, 6):
39 ret = aldebaran_reset_init(adev);
40 break;
41 case IP_VERSION(11, 0, 7):
42 ret = sienna_cichlid_reset_init(adev);
43 break;
44 case IP_VERSION(13, 0, 10):
45 ret = smu_v13_0_10_reset_init(adev);
46 break;
47 default:
48 break;
49 }
50
51 return ret;
52}
53
54int amdgpu_reset_fini(struct amdgpu_device *adev)
55{
56 int ret = 0;
57
58 switch (amdgpu_ip_version(adev, MP1_HWIP, 0)) {
59 case IP_VERSION(13, 0, 2):
60 case IP_VERSION(13, 0, 6):
61 ret = aldebaran_reset_fini(adev);
62 break;
63 case IP_VERSION(11, 0, 7):
64 ret = sienna_cichlid_reset_fini(adev);
65 break;
66 case IP_VERSION(13, 0, 10):
67 ret = smu_v13_0_10_reset_fini(adev);
68 break;
69 default:
70 break;
71 }
72
73 return ret;
74}
75
76int amdgpu_reset_prepare_hwcontext(struct amdgpu_device *adev,
77 struct amdgpu_reset_context *reset_context)
78{
79 struct amdgpu_reset_handler *reset_handler = NULL;
80
81 if (adev->reset_cntl && adev->reset_cntl->get_reset_handler)
82 reset_handler = adev->reset_cntl->get_reset_handler(
83 adev->reset_cntl, reset_context);
84 if (!reset_handler)
85 return -EOPNOTSUPP;
86
87 return reset_handler->prepare_hwcontext(adev->reset_cntl,
88 reset_context);
89}
90
91int amdgpu_reset_perform_reset(struct amdgpu_device *adev,
92 struct amdgpu_reset_context *reset_context)
93{
94 int ret;
95 struct amdgpu_reset_handler *reset_handler = NULL;
96
97 if (adev->reset_cntl)
98 reset_handler = adev->reset_cntl->get_reset_handler(
99 adev->reset_cntl, reset_context);
100 if (!reset_handler)
101 return -EOPNOTSUPP;
102
103 ret = reset_handler->perform_reset(adev->reset_cntl, reset_context);
104 if (ret)
105 return ret;
106
107 return reset_handler->restore_hwcontext(adev->reset_cntl,
108 reset_context);
109}
110
111
112void amdgpu_reset_destroy_reset_domain(struct kref *ref)
113{
114 struct amdgpu_reset_domain *reset_domain = container_of(ref,
115 struct amdgpu_reset_domain,
116 refcount);
117 if (reset_domain->wq)
118 destroy_workqueue(reset_domain->wq);
119
120 kvfree(reset_domain);
121}
122
123struct amdgpu_reset_domain *amdgpu_reset_create_reset_domain(enum amdgpu_reset_domain_type type,
124 char *wq_name)
125{
126 struct amdgpu_reset_domain *reset_domain;
127
128 reset_domain = kvzalloc(sizeof(struct amdgpu_reset_domain), GFP_KERNEL);
129 if (!reset_domain) {
130 DRM_ERROR("Failed to allocate amdgpu_reset_domain!");
131 return NULL;
132 }
133
134 reset_domain->type = type;
135 kref_init(&reset_domain->refcount);
136
137 reset_domain->wq = create_singlethread_workqueue(wq_name);
138 if (!reset_domain->wq) {
139 DRM_ERROR("Failed to allocate wq for amdgpu_reset_domain!");
140 amdgpu_reset_put_reset_domain(reset_domain);
141 return NULL;
142
143 }
144
145 atomic_set(&reset_domain->in_gpu_reset, 0);
146 atomic_set(&reset_domain->reset_res, 0);
147 init_rwsem(&reset_domain->sem);
148
149 return reset_domain;
150}
151
152void amdgpu_device_lock_reset_domain(struct amdgpu_reset_domain *reset_domain)
153{
154 atomic_set(&reset_domain->in_gpu_reset, 1);
155 down_write(&reset_domain->sem);
156}
157
158
159void amdgpu_device_unlock_reset_domain(struct amdgpu_reset_domain *reset_domain)
160{
161 atomic_set(&reset_domain->in_gpu_reset, 0);
162 up_write(&reset_domain->sem);
163}
164
165#ifndef CONFIG_DEV_COREDUMP
166void amdgpu_coredump(struct amdgpu_device *adev, bool vram_lost,
167 struct amdgpu_reset_context *reset_context)
168{
169}
170#else
171static ssize_t
172amdgpu_devcoredump_read(char *buffer, loff_t offset, size_t count,
173 void *data, size_t datalen)
174{
175 struct drm_printer p;
176 struct amdgpu_coredump_info *coredump = data;
177 struct drm_print_iterator iter;
178 int i;
179
180 iter.data = buffer;
181 iter.offset = 0;
182 iter.start = offset;
183 iter.remain = count;
184
185 p = drm_coredump_printer(&iter);
186
187 drm_printf(&p, "**** AMDGPU Device Coredump ****\n");
188 drm_printf(&p, "version: " AMDGPU_COREDUMP_VERSION "\n");
189 drm_printf(&p, "kernel: " UTS_RELEASE "\n");
190 drm_printf(&p, "module: " KBUILD_MODNAME "\n");
191 drm_printf(&p, "time: %lld.%09ld\n", coredump->reset_time.tv_sec,
192 coredump->reset_time.tv_nsec);
193
194 if (coredump->reset_task_info.pid)
195 drm_printf(&p, "process_name: %s PID: %d\n",
196 coredump->reset_task_info.process_name,
197 coredump->reset_task_info.pid);
198
199 if (coredump->ring) {
200 drm_printf(&p, "\nRing timed out details\n");
201 drm_printf(&p, "IP Type: %d Ring Name: %s\n",
202 coredump->ring->funcs->type,
203 coredump->ring->name);
204 }
205
206 if (coredump->reset_vram_lost)
207 drm_printf(&p, "VRAM is lost due to GPU reset!\n");
208 if (coredump->adev->reset_info.num_regs) {
209 drm_printf(&p, "AMDGPU register dumps:\nOffset: Value:\n");
210
211 for (i = 0; i < coredump->adev->reset_info.num_regs; i++)
212 drm_printf(&p, "0x%08x: 0x%08x\n",
213 coredump->adev->reset_info.reset_dump_reg_list[i],
214 coredump->adev->reset_info.reset_dump_reg_value[i]);
215 }
216
217 return count - iter.remain;
218}
219
220static void amdgpu_devcoredump_free(void *data)
221{
222 kfree(data);
223}
224
225void amdgpu_coredump(struct amdgpu_device *adev, bool vram_lost,
226 struct amdgpu_reset_context *reset_context)
227{
228 struct amdgpu_coredump_info *coredump;
229 struct drm_device *dev = adev_to_drm(adev);
230 struct amdgpu_job *job = reset_context->job;
231 struct drm_sched_job *s_job;
232
233 coredump = kzalloc(sizeof(*coredump), GFP_NOWAIT);
234
235 if (!coredump) {
236 DRM_ERROR("%s: failed to allocate memory for coredump\n", __func__);
237 return;
238 }
239
240 coredump->reset_vram_lost = vram_lost;
241
242 if (reset_context->job && reset_context->job->vm) {
243 struct amdgpu_task_info *ti;
244 struct amdgpu_vm *vm = reset_context->job->vm;
245
246 ti = amdgpu_vm_get_task_info_vm(vm);
247 if (ti) {
248 coredump->reset_task_info = *ti;
249 amdgpu_vm_put_task_info(ti);
250 }
251 }
252
253 if (job) {
254 s_job = &job->base;
255 coredump->ring = to_amdgpu_ring(s_job->sched);
256 }
257
258 coredump->adev = adev;
259
260 ktime_get_ts64(&coredump->reset_time);
261
262 dev_coredumpm(dev->dev, THIS_MODULE, coredump, 0, GFP_NOWAIT,
263 amdgpu_devcoredump_read, amdgpu_devcoredump_free);
264}
265#endif
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
29static int amdgpu_reset_xgmi_reset_on_init_suspend(struct amdgpu_device *adev)
30{
31 int i;
32
33 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
34 if (!adev->ip_blocks[i].status.valid)
35 continue;
36 if (!adev->ip_blocks[i].status.hw)
37 continue;
38 /* displays are handled in phase1 */
39 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
40 continue;
41
42 /* XXX handle errors */
43 amdgpu_ip_block_suspend(&adev->ip_blocks[i]);
44 adev->ip_blocks[i].status.hw = false;
45 }
46
47 /* VCN FW shared region is in frambuffer, there are some flags
48 * initialized in that region during sw_init. Make sure the region is
49 * backed up.
50 */
51 amdgpu_vcn_save_vcpu_bo(adev);
52
53 return 0;
54}
55
56static int amdgpu_reset_xgmi_reset_on_init_prep_hwctxt(
57 struct amdgpu_reset_control *reset_ctl,
58 struct amdgpu_reset_context *reset_context)
59{
60 struct list_head *reset_device_list = reset_context->reset_device_list;
61 struct amdgpu_device *tmp_adev;
62 int r;
63
64 list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
65 amdgpu_unregister_gpu_instance(tmp_adev);
66 r = amdgpu_reset_xgmi_reset_on_init_suspend(tmp_adev);
67 if (r) {
68 dev_err(tmp_adev->dev,
69 "xgmi reset on init: prepare for reset failed");
70 return r;
71 }
72 }
73
74 return r;
75}
76
77static int amdgpu_reset_xgmi_reset_on_init_restore_hwctxt(
78 struct amdgpu_reset_control *reset_ctl,
79 struct amdgpu_reset_context *reset_context)
80{
81 struct list_head *reset_device_list = reset_context->reset_device_list;
82 struct amdgpu_device *tmp_adev = NULL;
83 int r;
84
85 r = amdgpu_device_reinit_after_reset(reset_context);
86 if (r)
87 return r;
88 list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
89 if (!tmp_adev->kfd.init_complete) {
90 kgd2kfd_init_zone_device(tmp_adev);
91 amdgpu_amdkfd_device_init(tmp_adev);
92 amdgpu_amdkfd_drm_client_create(tmp_adev);
93 }
94 }
95
96 return r;
97}
98
99static int amdgpu_reset_xgmi_reset_on_init_perform_reset(
100 struct amdgpu_reset_control *reset_ctl,
101 struct amdgpu_reset_context *reset_context)
102{
103 struct amdgpu_device *adev = (struct amdgpu_device *)reset_ctl->handle;
104 struct list_head *reset_device_list = reset_context->reset_device_list;
105 struct amdgpu_device *tmp_adev = NULL;
106 int r;
107
108 dev_dbg(adev->dev, "xgmi roi - hw reset\n");
109
110 list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
111 mutex_lock(&tmp_adev->reset_cntl->reset_lock);
112 tmp_adev->reset_cntl->active_reset =
113 amdgpu_asic_reset_method(adev);
114 }
115 r = 0;
116 /* Mode1 reset needs to be triggered on all devices together */
117 list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
118 /* For XGMI run all resets in parallel to speed up the process */
119 if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work))
120 r = -EALREADY;
121 if (r) {
122 dev_err(tmp_adev->dev,
123 "xgmi reset on init: reset failed with error, %d",
124 r);
125 break;
126 }
127 }
128
129 /* For XGMI wait for all resets to complete before proceed */
130 if (!r) {
131 list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
132 flush_work(&tmp_adev->xgmi_reset_work);
133 r = tmp_adev->asic_reset_res;
134 if (r)
135 break;
136 }
137 }
138
139 list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
140 mutex_unlock(&tmp_adev->reset_cntl->reset_lock);
141 tmp_adev->reset_cntl->active_reset = AMD_RESET_METHOD_NONE;
142 }
143
144 return r;
145}
146
147int amdgpu_reset_do_xgmi_reset_on_init(
148 struct amdgpu_reset_context *reset_context)
149{
150 struct list_head *reset_device_list = reset_context->reset_device_list;
151 struct amdgpu_device *adev;
152 int r;
153
154 if (!reset_device_list || list_empty(reset_device_list) ||
155 list_is_singular(reset_device_list))
156 return -EINVAL;
157
158 adev = list_first_entry(reset_device_list, struct amdgpu_device,
159 reset_list);
160 r = amdgpu_reset_prepare_hwcontext(adev, reset_context);
161 if (r)
162 return r;
163
164 r = amdgpu_reset_perform_reset(adev, reset_context);
165
166 return r;
167}
168
169struct amdgpu_reset_handler xgmi_reset_on_init_handler = {
170 .reset_method = AMD_RESET_METHOD_ON_INIT,
171 .prepare_env = NULL,
172 .prepare_hwcontext = amdgpu_reset_xgmi_reset_on_init_prep_hwctxt,
173 .perform_reset = amdgpu_reset_xgmi_reset_on_init_perform_reset,
174 .restore_hwcontext = amdgpu_reset_xgmi_reset_on_init_restore_hwctxt,
175 .restore_env = NULL,
176 .do_reset = NULL,
177};
178
179int amdgpu_reset_init(struct amdgpu_device *adev)
180{
181 int ret = 0;
182
183 switch (amdgpu_ip_version(adev, MP1_HWIP, 0)) {
184 case IP_VERSION(13, 0, 2):
185 case IP_VERSION(13, 0, 6):
186 case IP_VERSION(13, 0, 14):
187 ret = aldebaran_reset_init(adev);
188 break;
189 case IP_VERSION(11, 0, 7):
190 ret = sienna_cichlid_reset_init(adev);
191 break;
192 case IP_VERSION(13, 0, 10):
193 ret = smu_v13_0_10_reset_init(adev);
194 break;
195 default:
196 break;
197 }
198
199 return ret;
200}
201
202int amdgpu_reset_fini(struct amdgpu_device *adev)
203{
204 int ret = 0;
205
206 switch (amdgpu_ip_version(adev, MP1_HWIP, 0)) {
207 case IP_VERSION(13, 0, 2):
208 case IP_VERSION(13, 0, 6):
209 case IP_VERSION(13, 0, 14):
210 ret = aldebaran_reset_fini(adev);
211 break;
212 case IP_VERSION(11, 0, 7):
213 ret = sienna_cichlid_reset_fini(adev);
214 break;
215 case IP_VERSION(13, 0, 10):
216 ret = smu_v13_0_10_reset_fini(adev);
217 break;
218 default:
219 break;
220 }
221
222 return ret;
223}
224
225int amdgpu_reset_prepare_hwcontext(struct amdgpu_device *adev,
226 struct amdgpu_reset_context *reset_context)
227{
228 struct amdgpu_reset_handler *reset_handler = NULL;
229
230 if (adev->reset_cntl && adev->reset_cntl->get_reset_handler)
231 reset_handler = adev->reset_cntl->get_reset_handler(
232 adev->reset_cntl, reset_context);
233 if (!reset_handler)
234 return -EOPNOTSUPP;
235
236 return reset_handler->prepare_hwcontext(adev->reset_cntl,
237 reset_context);
238}
239
240int amdgpu_reset_perform_reset(struct amdgpu_device *adev,
241 struct amdgpu_reset_context *reset_context)
242{
243 int ret;
244 struct amdgpu_reset_handler *reset_handler = NULL;
245
246 if (adev->reset_cntl)
247 reset_handler = adev->reset_cntl->get_reset_handler(
248 adev->reset_cntl, reset_context);
249 if (!reset_handler)
250 return -EOPNOTSUPP;
251
252 ret = reset_handler->perform_reset(adev->reset_cntl, reset_context);
253 if (ret)
254 return ret;
255
256 return reset_handler->restore_hwcontext(adev->reset_cntl,
257 reset_context);
258}
259
260
261void amdgpu_reset_destroy_reset_domain(struct kref *ref)
262{
263 struct amdgpu_reset_domain *reset_domain = container_of(ref,
264 struct amdgpu_reset_domain,
265 refcount);
266 if (reset_domain->wq)
267 destroy_workqueue(reset_domain->wq);
268
269 kvfree(reset_domain);
270}
271
272struct amdgpu_reset_domain *amdgpu_reset_create_reset_domain(enum amdgpu_reset_domain_type type,
273 char *wq_name)
274{
275 struct amdgpu_reset_domain *reset_domain;
276
277 reset_domain = kvzalloc(sizeof(struct amdgpu_reset_domain), GFP_KERNEL);
278 if (!reset_domain) {
279 DRM_ERROR("Failed to allocate amdgpu_reset_domain!");
280 return NULL;
281 }
282
283 reset_domain->type = type;
284 kref_init(&reset_domain->refcount);
285
286 reset_domain->wq = create_singlethread_workqueue(wq_name);
287 if (!reset_domain->wq) {
288 DRM_ERROR("Failed to allocate wq for amdgpu_reset_domain!");
289 amdgpu_reset_put_reset_domain(reset_domain);
290 return NULL;
291
292 }
293
294 atomic_set(&reset_domain->in_gpu_reset, 0);
295 atomic_set(&reset_domain->reset_res, 0);
296 init_rwsem(&reset_domain->sem);
297
298 return reset_domain;
299}
300
301void amdgpu_device_lock_reset_domain(struct amdgpu_reset_domain *reset_domain)
302{
303 atomic_set(&reset_domain->in_gpu_reset, 1);
304 down_write(&reset_domain->sem);
305}
306
307
308void amdgpu_device_unlock_reset_domain(struct amdgpu_reset_domain *reset_domain)
309{
310 atomic_set(&reset_domain->in_gpu_reset, 0);
311 up_write(&reset_domain->sem);
312}
313
314void amdgpu_reset_get_desc(struct amdgpu_reset_context *rst_ctxt, char *buf,
315 size_t len)
316{
317 if (!buf || !len)
318 return;
319
320 switch (rst_ctxt->src) {
321 case AMDGPU_RESET_SRC_JOB:
322 if (rst_ctxt->job) {
323 snprintf(buf, len, "job hang on ring:%s",
324 rst_ctxt->job->base.sched->name);
325 } else {
326 strscpy(buf, "job hang", len);
327 }
328 break;
329 case AMDGPU_RESET_SRC_RAS:
330 strscpy(buf, "RAS error", len);
331 break;
332 case AMDGPU_RESET_SRC_MES:
333 strscpy(buf, "MES hang", len);
334 break;
335 case AMDGPU_RESET_SRC_HWS:
336 strscpy(buf, "HWS hang", len);
337 break;
338 case AMDGPU_RESET_SRC_USER:
339 strscpy(buf, "user trigger", len);
340 break;
341 default:
342 strscpy(buf, "unknown", len);
343 }
344}
345
346bool amdgpu_reset_in_recovery(struct amdgpu_device *adev)
347{
348 return (adev->init_lvl->level == AMDGPU_INIT_LEVEL_RESET_RECOVERY);
349}