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 "aldebaran.h"
25 #include "amdgpu_reset.h"
26 #include "amdgpu_amdkfd.h"
27 #include "amdgpu_dpm.h"
28 #include "amdgpu_job.h"
29 #include "amdgpu_ring.h"
30 #include "amdgpu_ras.h"
31 #include "amdgpu_psp.h"
32 #include "amdgpu_xgmi.h"
33
aldebaran_is_mode2_default(struct amdgpu_reset_control * reset_ctl)34 static bool aldebaran_is_mode2_default(struct amdgpu_reset_control *reset_ctl)
35 {
36 struct amdgpu_device *adev = (struct amdgpu_device *)reset_ctl->handle;
37
38 if ((adev->ip_versions[MP1_HWIP][0] == IP_VERSION(13, 0, 2) &&
39 adev->gmc.xgmi.connected_to_cpu))
40 return true;
41
42 return false;
43 }
44
45 static struct amdgpu_reset_handler *
aldebaran_get_reset_handler(struct amdgpu_reset_control * reset_ctl,struct amdgpu_reset_context * reset_context)46 aldebaran_get_reset_handler(struct amdgpu_reset_control *reset_ctl,
47 struct amdgpu_reset_context *reset_context)
48 {
49 struct amdgpu_reset_handler *handler;
50 struct amdgpu_device *adev = (struct amdgpu_device *)reset_ctl->handle;
51
52 if (reset_context->method != AMD_RESET_METHOD_NONE) {
53 dev_dbg(adev->dev, "Getting reset handler for method %d\n",
54 reset_context->method);
55 list_for_each_entry(handler, &reset_ctl->reset_handlers,
56 handler_list) {
57 if (handler->reset_method == reset_context->method)
58 return handler;
59 }
60 }
61
62 if (aldebaran_is_mode2_default(reset_ctl)) {
63 list_for_each_entry(handler, &reset_ctl->reset_handlers,
64 handler_list) {
65 if (handler->reset_method == AMD_RESET_METHOD_MODE2) {
66 reset_context->method = AMD_RESET_METHOD_MODE2;
67 return handler;
68 }
69 }
70 }
71
72 dev_dbg(adev->dev, "Reset handler not found!\n");
73
74 return NULL;
75 }
76
aldebaran_mode2_suspend_ip(struct amdgpu_device * adev)77 static int aldebaran_mode2_suspend_ip(struct amdgpu_device *adev)
78 {
79 int r, i;
80
81 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
82 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
83
84 for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
85 if (!(adev->ip_blocks[i].version->type ==
86 AMD_IP_BLOCK_TYPE_GFX ||
87 adev->ip_blocks[i].version->type ==
88 AMD_IP_BLOCK_TYPE_SDMA))
89 continue;
90
91 r = adev->ip_blocks[i].version->funcs->suspend(adev);
92
93 if (r) {
94 dev_err(adev->dev,
95 "suspend of IP block <%s> failed %d\n",
96 adev->ip_blocks[i].version->funcs->name, r);
97 return r;
98 }
99
100 adev->ip_blocks[i].status.hw = false;
101 }
102
103 return r;
104 }
105
106 static int
aldebaran_mode2_prepare_hwcontext(struct amdgpu_reset_control * reset_ctl,struct amdgpu_reset_context * reset_context)107 aldebaran_mode2_prepare_hwcontext(struct amdgpu_reset_control *reset_ctl,
108 struct amdgpu_reset_context *reset_context)
109 {
110 int r = 0;
111 struct amdgpu_device *adev = (struct amdgpu_device *)reset_ctl->handle;
112
113 dev_dbg(adev->dev, "Aldebaran prepare hw context\n");
114 /* Don't suspend on bare metal if we are not going to HW reset the ASIC */
115 if (!amdgpu_sriov_vf(adev))
116 r = aldebaran_mode2_suspend_ip(adev);
117
118 return r;
119 }
120
aldebaran_async_reset(struct work_struct * work)121 static void aldebaran_async_reset(struct work_struct *work)
122 {
123 struct amdgpu_reset_handler *handler;
124 struct amdgpu_reset_control *reset_ctl =
125 container_of(work, struct amdgpu_reset_control, reset_work);
126 struct amdgpu_device *adev = (struct amdgpu_device *)reset_ctl->handle;
127
128 list_for_each_entry(handler, &reset_ctl->reset_handlers,
129 handler_list) {
130 if (handler->reset_method == reset_ctl->active_reset) {
131 dev_dbg(adev->dev, "Resetting device\n");
132 handler->do_reset(adev);
133 break;
134 }
135 }
136 }
137
aldebaran_mode2_reset(struct amdgpu_device * adev)138 static int aldebaran_mode2_reset(struct amdgpu_device *adev)
139 {
140 /* disable BM */
141 pci_clear_master(adev->pdev);
142 adev->asic_reset_res = amdgpu_dpm_mode2_reset(adev);
143 return adev->asic_reset_res;
144 }
145
146 static int
aldebaran_mode2_perform_reset(struct amdgpu_reset_control * reset_ctl,struct amdgpu_reset_context * reset_context)147 aldebaran_mode2_perform_reset(struct amdgpu_reset_control *reset_ctl,
148 struct amdgpu_reset_context *reset_context)
149 {
150 struct amdgpu_device *adev = (struct amdgpu_device *)reset_ctl->handle;
151 struct list_head *reset_device_list = reset_context->reset_device_list;
152 struct amdgpu_device *tmp_adev = NULL;
153 int r = 0;
154
155 dev_dbg(adev->dev, "aldebaran perform hw reset\n");
156
157 if (reset_device_list == NULL)
158 return -EINVAL;
159
160 if (adev->ip_versions[MP1_HWIP][0] == IP_VERSION(13, 0, 2) &&
161 reset_context->hive == NULL) {
162 /* Wrong context, return error */
163 return -EINVAL;
164 }
165
166 list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
167 mutex_lock(&tmp_adev->reset_cntl->reset_lock);
168 tmp_adev->reset_cntl->active_reset = AMD_RESET_METHOD_MODE2;
169 }
170 /*
171 * Mode2 reset doesn't need any sync between nodes in XGMI hive, instead launch
172 * them together so that they can be completed asynchronously on multiple nodes
173 */
174 list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
175 /* For XGMI run all resets in parallel to speed up the process */
176 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
177 if (!queue_work(system_unbound_wq,
178 &tmp_adev->reset_cntl->reset_work))
179 r = -EALREADY;
180 } else
181 r = aldebaran_mode2_reset(tmp_adev);
182 if (r) {
183 dev_err(tmp_adev->dev,
184 "ASIC reset failed with error, %d for drm dev, %s",
185 r, adev_to_drm(tmp_adev)->unique);
186 break;
187 }
188 }
189
190 /* For XGMI wait for all resets to complete before proceed */
191 if (!r) {
192 list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
193 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
194 flush_work(&tmp_adev->reset_cntl->reset_work);
195 r = tmp_adev->asic_reset_res;
196 if (r)
197 break;
198 }
199 }
200 }
201
202 list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
203 mutex_unlock(&tmp_adev->reset_cntl->reset_lock);
204 tmp_adev->reset_cntl->active_reset = AMD_RESET_METHOD_NONE;
205 }
206
207 return r;
208 }
209
aldebaran_mode2_restore_ip(struct amdgpu_device * adev)210 static int aldebaran_mode2_restore_ip(struct amdgpu_device *adev)
211 {
212 struct amdgpu_firmware_info *ucode_list[AMDGPU_UCODE_ID_MAXIMUM];
213 struct amdgpu_firmware_info *ucode;
214 struct amdgpu_ip_block *cmn_block;
215 int ucode_count = 0;
216 int i, r;
217
218 dev_dbg(adev->dev, "Reloading ucodes after reset\n");
219 for (i = 0; i < adev->firmware.max_ucodes; i++) {
220 ucode = &adev->firmware.ucode[i];
221 if (!ucode->fw)
222 continue;
223 switch (ucode->ucode_id) {
224 case AMDGPU_UCODE_ID_SDMA0:
225 case AMDGPU_UCODE_ID_SDMA1:
226 case AMDGPU_UCODE_ID_SDMA2:
227 case AMDGPU_UCODE_ID_SDMA3:
228 case AMDGPU_UCODE_ID_SDMA4:
229 case AMDGPU_UCODE_ID_SDMA5:
230 case AMDGPU_UCODE_ID_SDMA6:
231 case AMDGPU_UCODE_ID_SDMA7:
232 case AMDGPU_UCODE_ID_CP_MEC1:
233 case AMDGPU_UCODE_ID_CP_MEC1_JT:
234 case AMDGPU_UCODE_ID_RLC_RESTORE_LIST_CNTL:
235 case AMDGPU_UCODE_ID_RLC_RESTORE_LIST_GPM_MEM:
236 case AMDGPU_UCODE_ID_RLC_RESTORE_LIST_SRM_MEM:
237 case AMDGPU_UCODE_ID_RLC_G:
238 ucode_list[ucode_count++] = ucode;
239 break;
240 default:
241 break;
242 }
243 }
244
245 /* Reinit NBIF block */
246 cmn_block =
247 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_COMMON);
248 if (unlikely(!cmn_block)) {
249 dev_err(adev->dev, "Failed to get BIF handle\n");
250 return -EINVAL;
251 }
252 r = cmn_block->version->funcs->resume(adev);
253 if (r)
254 return r;
255
256 /* Reinit GFXHUB */
257 adev->gfxhub.funcs->init(adev);
258 r = adev->gfxhub.funcs->gart_enable(adev);
259 if (r) {
260 dev_err(adev->dev, "GFXHUB gart reenable failed after reset\n");
261 return r;
262 }
263
264 /* Reload GFX firmware */
265 r = psp_load_fw_list(&adev->psp, ucode_list, ucode_count);
266 if (r) {
267 dev_err(adev->dev, "GFX ucode load failed after reset\n");
268 return r;
269 }
270
271 /* Resume RLC, FW needs RLC alive to complete reset process */
272 adev->gfx.rlc.funcs->resume(adev);
273
274 /* Wait for FW reset event complete */
275 r = amdgpu_dpm_wait_for_event(adev, SMU_EVENT_RESET_COMPLETE, 0);
276 if (r) {
277 dev_err(adev->dev,
278 "Failed to get response from firmware after reset\n");
279 return r;
280 }
281
282 for (i = 0; i < adev->num_ip_blocks; i++) {
283 if (!(adev->ip_blocks[i].version->type ==
284 AMD_IP_BLOCK_TYPE_GFX ||
285 adev->ip_blocks[i].version->type ==
286 AMD_IP_BLOCK_TYPE_SDMA))
287 continue;
288 r = adev->ip_blocks[i].version->funcs->resume(adev);
289 if (r) {
290 dev_err(adev->dev,
291 "resume of IP block <%s> failed %d\n",
292 adev->ip_blocks[i].version->funcs->name, r);
293 return r;
294 }
295
296 adev->ip_blocks[i].status.hw = true;
297 }
298
299 for (i = 0; i < adev->num_ip_blocks; i++) {
300 if (!(adev->ip_blocks[i].version->type ==
301 AMD_IP_BLOCK_TYPE_GFX ||
302 adev->ip_blocks[i].version->type ==
303 AMD_IP_BLOCK_TYPE_SDMA ||
304 adev->ip_blocks[i].version->type ==
305 AMD_IP_BLOCK_TYPE_COMMON))
306 continue;
307
308 if (adev->ip_blocks[i].version->funcs->late_init) {
309 r = adev->ip_blocks[i].version->funcs->late_init(
310 (void *)adev);
311 if (r) {
312 dev_err(adev->dev,
313 "late_init of IP block <%s> failed %d after reset\n",
314 adev->ip_blocks[i].version->funcs->name,
315 r);
316 return r;
317 }
318 }
319 adev->ip_blocks[i].status.late_initialized = true;
320 }
321
322 amdgpu_ras_set_error_query_ready(adev, true);
323
324 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
325 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
326
327 return r;
328 }
329
330 static int
aldebaran_mode2_restore_hwcontext(struct amdgpu_reset_control * reset_ctl,struct amdgpu_reset_context * reset_context)331 aldebaran_mode2_restore_hwcontext(struct amdgpu_reset_control *reset_ctl,
332 struct amdgpu_reset_context *reset_context)
333 {
334 struct list_head *reset_device_list = reset_context->reset_device_list;
335 struct amdgpu_device *tmp_adev = NULL;
336 struct amdgpu_ras *con;
337 int r;
338
339 if (reset_device_list == NULL)
340 return -EINVAL;
341
342 if (reset_context->reset_req_dev->ip_versions[MP1_HWIP][0] ==
343 IP_VERSION(13, 0, 2) &&
344 reset_context->hive == NULL) {
345 /* Wrong context, return error */
346 return -EINVAL;
347 }
348
349 list_for_each_entry(tmp_adev, reset_device_list, reset_list) {
350 dev_info(tmp_adev->dev,
351 "GPU reset succeeded, trying to resume\n");
352 r = aldebaran_mode2_restore_ip(tmp_adev);
353 if (r)
354 goto end;
355
356 /*
357 * Add this ASIC as tracked as reset was already
358 * complete successfully.
359 */
360 amdgpu_register_gpu_instance(tmp_adev);
361
362 /* Resume RAS, ecc_irq */
363 con = amdgpu_ras_get_context(tmp_adev);
364 if (!amdgpu_sriov_vf(tmp_adev) && con) {
365 if (tmp_adev->sdma.ras &&
366 tmp_adev->sdma.ras->ras_block.ras_late_init) {
367 r = tmp_adev->sdma.ras->ras_block.ras_late_init(tmp_adev,
368 &tmp_adev->sdma.ras->ras_block.ras_comm);
369 if (r) {
370 dev_err(tmp_adev->dev, "SDMA failed to execute ras_late_init! ret:%d\n", r);
371 goto end;
372 }
373 }
374
375 if (tmp_adev->gfx.ras &&
376 tmp_adev->gfx.ras->ras_block.ras_late_init) {
377 r = tmp_adev->gfx.ras->ras_block.ras_late_init(tmp_adev,
378 &tmp_adev->gfx.ras->ras_block.ras_comm);
379 if (r) {
380 dev_err(tmp_adev->dev, "GFX failed to execute ras_late_init! ret:%d\n", r);
381 goto end;
382 }
383 }
384 }
385
386 amdgpu_ras_resume(tmp_adev);
387
388 /* Update PSP FW topology after reset */
389 if (reset_context->hive &&
390 tmp_adev->gmc.xgmi.num_physical_nodes > 1)
391 r = amdgpu_xgmi_update_topology(reset_context->hive,
392 tmp_adev);
393
394 if (!r) {
395 amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
396
397 r = amdgpu_ib_ring_tests(tmp_adev);
398 if (r) {
399 dev_err(tmp_adev->dev,
400 "ib ring test failed (%d).\n", r);
401 r = -EAGAIN;
402 tmp_adev->asic_reset_res = r;
403 goto end;
404 }
405 }
406 }
407
408 end:
409 return r;
410 }
411
412 static struct amdgpu_reset_handler aldebaran_mode2_handler = {
413 .reset_method = AMD_RESET_METHOD_MODE2,
414 .prepare_env = NULL,
415 .prepare_hwcontext = aldebaran_mode2_prepare_hwcontext,
416 .perform_reset = aldebaran_mode2_perform_reset,
417 .restore_hwcontext = aldebaran_mode2_restore_hwcontext,
418 .restore_env = NULL,
419 .do_reset = aldebaran_mode2_reset,
420 };
421
aldebaran_reset_init(struct amdgpu_device * adev)422 int aldebaran_reset_init(struct amdgpu_device *adev)
423 {
424 struct amdgpu_reset_control *reset_ctl;
425
426 reset_ctl = kzalloc(sizeof(*reset_ctl), GFP_KERNEL);
427 if (!reset_ctl)
428 return -ENOMEM;
429
430 reset_ctl->handle = adev;
431 reset_ctl->async_reset = aldebaran_async_reset;
432 reset_ctl->active_reset = AMD_RESET_METHOD_NONE;
433 reset_ctl->get_reset_handler = aldebaran_get_reset_handler;
434
435 INIT_LIST_HEAD(&reset_ctl->reset_handlers);
436 INIT_WORK(&reset_ctl->reset_work, reset_ctl->async_reset);
437 /* Only mode2 is handled through reset control now */
438 amdgpu_reset_add_handler(reset_ctl, &aldebaran_mode2_handler);
439
440 adev->reset_cntl = reset_ctl;
441
442 return 0;
443 }
444
aldebaran_reset_fini(struct amdgpu_device * adev)445 int aldebaran_reset_fini(struct amdgpu_device *adev)
446 {
447 kfree(adev->reset_cntl);
448 adev->reset_cntl = NULL;
449 return 0;
450 }
451