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 
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 *
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 
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
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 
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 
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
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 amdgpu_device *tmp_adev = NULL;
152 	struct list_head reset_device_list;
153 	int r = 0;
154 
155 	dev_dbg(adev->dev, "aldebaran perform hw reset\n");
156 	if (adev->ip_versions[MP1_HWIP][0] == IP_VERSION(13, 0, 2) &&
157 	    reset_context->hive == NULL) {
158 		/* Wrong context, return error */
159 		return -EINVAL;
160 	}
161 
162 	INIT_LIST_HEAD(&reset_device_list);
163 	if (reset_context->hive) {
164 		list_for_each_entry (tmp_adev,
165 				     &reset_context->hive->device_list,
166 				     gmc.xgmi.head)
167 			list_add_tail(&tmp_adev->reset_list,
168 				      &reset_device_list);
169 	} else {
170 		list_add_tail(&reset_context->reset_req_dev->reset_list,
171 			      &reset_device_list);
172 	}
173 
174 	list_for_each_entry (tmp_adev, &reset_device_list, reset_list) {
175 		mutex_lock(&tmp_adev->reset_cntl->reset_lock);
176 		tmp_adev->reset_cntl->active_reset = AMD_RESET_METHOD_MODE2;
177 	}
178 	/*
179 	 * Mode2 reset doesn't need any sync between nodes in XGMI hive, instead launch
180 	 * them together so that they can be completed asynchronously on multiple nodes
181 	 */
182 	list_for_each_entry (tmp_adev, &reset_device_list, reset_list) {
183 		/* For XGMI run all resets in parallel to speed up the process */
184 		if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
185 			if (!queue_work(system_unbound_wq,
186 					&tmp_adev->reset_cntl->reset_work))
187 				r = -EALREADY;
188 		} else
189 			r = aldebaran_mode2_reset(tmp_adev);
190 		if (r) {
191 			dev_err(tmp_adev->dev,
192 				"ASIC reset failed with error, %d for drm dev, %s",
193 				r, adev_to_drm(tmp_adev)->unique);
194 			break;
195 		}
196 	}
197 
198 	/* For XGMI wait for all resets to complete before proceed */
199 	if (!r) {
200 		list_for_each_entry (tmp_adev, &reset_device_list, reset_list) {
201 			if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
202 				flush_work(&tmp_adev->reset_cntl->reset_work);
203 				r = tmp_adev->asic_reset_res;
204 				if (r)
205 					break;
206 			}
207 		}
208 	}
209 
210 	list_for_each_entry (tmp_adev, &reset_device_list, reset_list) {
211 		mutex_unlock(&tmp_adev->reset_cntl->reset_lock);
212 		tmp_adev->reset_cntl->active_reset = AMD_RESET_METHOD_NONE;
213 	}
214 
215 	return r;
216 }
217 
218 static int aldebaran_mode2_restore_ip(struct amdgpu_device *adev)
219 {
220 	struct amdgpu_firmware_info *ucode_list[AMDGPU_UCODE_ID_MAXIMUM];
221 	struct amdgpu_firmware_info *ucode;
222 	struct amdgpu_ip_block *cmn_block;
223 	int ucode_count = 0;
224 	int i, r;
225 
226 	dev_dbg(adev->dev, "Reloading ucodes after reset\n");
227 	for (i = 0; i < adev->firmware.max_ucodes; i++) {
228 		ucode = &adev->firmware.ucode[i];
229 		if (!ucode->fw)
230 			continue;
231 		switch (ucode->ucode_id) {
232 		case AMDGPU_UCODE_ID_SDMA0:
233 		case AMDGPU_UCODE_ID_SDMA1:
234 		case AMDGPU_UCODE_ID_SDMA2:
235 		case AMDGPU_UCODE_ID_SDMA3:
236 		case AMDGPU_UCODE_ID_SDMA4:
237 		case AMDGPU_UCODE_ID_SDMA5:
238 		case AMDGPU_UCODE_ID_SDMA6:
239 		case AMDGPU_UCODE_ID_SDMA7:
240 		case AMDGPU_UCODE_ID_CP_MEC1:
241 		case AMDGPU_UCODE_ID_CP_MEC1_JT:
242 		case AMDGPU_UCODE_ID_RLC_RESTORE_LIST_CNTL:
243 		case AMDGPU_UCODE_ID_RLC_RESTORE_LIST_GPM_MEM:
244 		case AMDGPU_UCODE_ID_RLC_RESTORE_LIST_SRM_MEM:
245 		case AMDGPU_UCODE_ID_RLC_G:
246 			ucode_list[ucode_count++] = ucode;
247 			break;
248 		default:
249 			break;
250 		}
251 	}
252 
253 	/* Reinit NBIF block */
254 	cmn_block =
255 		amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_COMMON);
256 	if (unlikely(!cmn_block)) {
257 		dev_err(adev->dev, "Failed to get BIF handle\n");
258 		return -EINVAL;
259 	}
260 	r = cmn_block->version->funcs->resume(adev);
261 	if (r)
262 		return r;
263 
264 	/* Reinit GFXHUB */
265 	adev->gfxhub.funcs->init(adev);
266 	r = adev->gfxhub.funcs->gart_enable(adev);
267 	if (r) {
268 		dev_err(adev->dev, "GFXHUB gart reenable failed after reset\n");
269 		return r;
270 	}
271 
272 	/* Reload GFX firmware */
273 	r = psp_load_fw_list(&adev->psp, ucode_list, ucode_count);
274 	if (r) {
275 		dev_err(adev->dev, "GFX ucode load failed after reset\n");
276 		return r;
277 	}
278 
279 	/* Resume RLC, FW needs RLC alive to complete reset process */
280 	adev->gfx.rlc.funcs->resume(adev);
281 
282 	/* Wait for FW reset event complete */
283 	r = amdgpu_dpm_wait_for_event(adev, SMU_EVENT_RESET_COMPLETE, 0);
284 	if (r) {
285 		dev_err(adev->dev,
286 			"Failed to get response from firmware after reset\n");
287 		return r;
288 	}
289 
290 	for (i = 0; i < adev->num_ip_blocks; i++) {
291 		if (!(adev->ip_blocks[i].version->type ==
292 			      AMD_IP_BLOCK_TYPE_GFX ||
293 		      adev->ip_blocks[i].version->type ==
294 			      AMD_IP_BLOCK_TYPE_SDMA))
295 			continue;
296 		r = adev->ip_blocks[i].version->funcs->resume(adev);
297 		if (r) {
298 			dev_err(adev->dev,
299 				"resume of IP block <%s> failed %d\n",
300 				adev->ip_blocks[i].version->funcs->name, r);
301 			return r;
302 		}
303 
304 		adev->ip_blocks[i].status.hw = true;
305 	}
306 
307 	for (i = 0; i < adev->num_ip_blocks; i++) {
308 		if (!(adev->ip_blocks[i].version->type ==
309 			      AMD_IP_BLOCK_TYPE_GFX ||
310 		      adev->ip_blocks[i].version->type ==
311 			      AMD_IP_BLOCK_TYPE_SDMA ||
312 		      adev->ip_blocks[i].version->type ==
313 			      AMD_IP_BLOCK_TYPE_COMMON))
314 			continue;
315 
316 		if (adev->ip_blocks[i].version->funcs->late_init) {
317 			r = adev->ip_blocks[i].version->funcs->late_init(
318 				(void *)adev);
319 			if (r) {
320 				dev_err(adev->dev,
321 					"late_init of IP block <%s> failed %d after reset\n",
322 					adev->ip_blocks[i].version->funcs->name,
323 					r);
324 				return r;
325 			}
326 		}
327 		adev->ip_blocks[i].status.late_initialized = true;
328 	}
329 
330 	amdgpu_ras_set_error_query_ready(adev, true);
331 
332 	amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
333 	amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
334 
335 	return r;
336 }
337 
338 static int
339 aldebaran_mode2_restore_hwcontext(struct amdgpu_reset_control *reset_ctl,
340 				  struct amdgpu_reset_context *reset_context)
341 {
342 	struct amdgpu_device *tmp_adev = NULL;
343 	struct list_head reset_device_list;
344 	int r;
345 
346 	if (reset_context->reset_req_dev->ip_versions[MP1_HWIP][0] ==
347 		    IP_VERSION(13, 0, 2) &&
348 	    reset_context->hive == NULL) {
349 		/* Wrong context, return error */
350 		return -EINVAL;
351 	}
352 
353 	INIT_LIST_HEAD(&reset_device_list);
354 	if (reset_context->hive) {
355 		list_for_each_entry (tmp_adev,
356 				     &reset_context->hive->device_list,
357 				     gmc.xgmi.head)
358 			list_add_tail(&tmp_adev->reset_list,
359 				      &reset_device_list);
360 	} else {
361 		list_add_tail(&reset_context->reset_req_dev->reset_list,
362 			      &reset_device_list);
363 	}
364 
365 	list_for_each_entry (tmp_adev, &reset_device_list, reset_list) {
366 		dev_info(tmp_adev->dev,
367 			 "GPU reset succeeded, trying to resume\n");
368 		r = aldebaran_mode2_restore_ip(tmp_adev);
369 		if (r)
370 			goto end;
371 
372 		/*
373 		 * Add this ASIC as tracked as reset was already
374 		 * complete successfully.
375 		 */
376 		amdgpu_register_gpu_instance(tmp_adev);
377 
378 		/* Resume RAS */
379 		amdgpu_ras_resume(tmp_adev);
380 
381 		/* Update PSP FW topology after reset */
382 		if (reset_context->hive &&
383 		    tmp_adev->gmc.xgmi.num_physical_nodes > 1)
384 			r = amdgpu_xgmi_update_topology(reset_context->hive,
385 							tmp_adev);
386 
387 		if (!r) {
388 			amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
389 
390 			r = amdgpu_ib_ring_tests(tmp_adev);
391 			if (r) {
392 				dev_err(tmp_adev->dev,
393 					"ib ring test failed (%d).\n", r);
394 				r = -EAGAIN;
395 				tmp_adev->asic_reset_res = r;
396 				goto end;
397 			}
398 		}
399 	}
400 
401 end:
402 	return r;
403 }
404 
405 static struct amdgpu_reset_handler aldebaran_mode2_handler = {
406 	.reset_method		= AMD_RESET_METHOD_MODE2,
407 	.prepare_env		= NULL,
408 	.prepare_hwcontext	= aldebaran_mode2_prepare_hwcontext,
409 	.perform_reset		= aldebaran_mode2_perform_reset,
410 	.restore_hwcontext	= aldebaran_mode2_restore_hwcontext,
411 	.restore_env		= NULL,
412 	.do_reset		= aldebaran_mode2_reset,
413 };
414 
415 int aldebaran_reset_init(struct amdgpu_device *adev)
416 {
417 	struct amdgpu_reset_control *reset_ctl;
418 
419 	reset_ctl = kzalloc(sizeof(*reset_ctl), GFP_KERNEL);
420 	if (!reset_ctl)
421 		return -ENOMEM;
422 
423 	reset_ctl->handle = adev;
424 	reset_ctl->async_reset = aldebaran_async_reset;
425 	reset_ctl->active_reset = AMD_RESET_METHOD_NONE;
426 	reset_ctl->get_reset_handler = aldebaran_get_reset_handler;
427 
428 	INIT_LIST_HEAD(&reset_ctl->reset_handlers);
429 	INIT_WORK(&reset_ctl->reset_work, reset_ctl->async_reset);
430 	/* Only mode2 is handled through reset control now */
431 	amdgpu_reset_add_handler(reset_ctl, &aldebaran_mode2_handler);
432 
433 	adev->reset_cntl = reset_ctl;
434 
435 	return 0;
436 }
437 
438 int aldebaran_reset_fini(struct amdgpu_device *adev)
439 {
440 	kfree(adev->reset_cntl);
441 	adev->reset_cntl = NULL;
442 	return 0;
443 }
444