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