xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c (revision 3a39d672e7f48b8d6b91a09afa4b55352773b4b5)
1 /*
2  * Copyright 2019 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/sort.h>
25 #include "amdgpu.h"
26 #include "umc_v6_7.h"
27 #define MAX_UMC_POISON_POLLING_TIME_SYNC   20  //ms
28 
29 #define MAX_UMC_HASH_STRING_SIZE  256
30 
amdgpu_umc_convert_error_address(struct amdgpu_device * adev,struct ras_err_data * err_data,uint64_t err_addr,uint32_t ch_inst,uint32_t umc_inst)31 static int amdgpu_umc_convert_error_address(struct amdgpu_device *adev,
32 				    struct ras_err_data *err_data, uint64_t err_addr,
33 				    uint32_t ch_inst, uint32_t umc_inst)
34 {
35 	switch (amdgpu_ip_version(adev, UMC_HWIP, 0)) {
36 	case IP_VERSION(6, 7, 0):
37 		umc_v6_7_convert_error_address(adev,
38 				err_data, err_addr, ch_inst, umc_inst);
39 		break;
40 	default:
41 		dev_warn(adev->dev,
42 			 "UMC address to Physical address translation is not supported\n");
43 		return AMDGPU_RAS_FAIL;
44 	}
45 
46 	return AMDGPU_RAS_SUCCESS;
47 }
48 
amdgpu_umc_page_retirement_mca(struct amdgpu_device * adev,uint64_t err_addr,uint32_t ch_inst,uint32_t umc_inst)49 int amdgpu_umc_page_retirement_mca(struct amdgpu_device *adev,
50 			uint64_t err_addr, uint32_t ch_inst, uint32_t umc_inst)
51 {
52 	struct ras_err_data err_data;
53 	int ret;
54 
55 	ret = amdgpu_ras_error_data_init(&err_data);
56 	if (ret)
57 		return ret;
58 
59 	err_data.err_addr =
60 		kcalloc(adev->umc.max_ras_err_cnt_per_query,
61 			sizeof(struct eeprom_table_record), GFP_KERNEL);
62 	if (!err_data.err_addr) {
63 		dev_warn(adev->dev,
64 			"Failed to alloc memory for umc error record in MCA notifier!\n");
65 		ret = AMDGPU_RAS_FAIL;
66 		goto out_fini_err_data;
67 	}
68 
69 	err_data.err_addr_len = adev->umc.max_ras_err_cnt_per_query;
70 
71 	/*
72 	 * Translate UMC channel address to Physical address
73 	 */
74 	ret = amdgpu_umc_convert_error_address(adev, &err_data, err_addr,
75 					ch_inst, umc_inst);
76 	if (ret)
77 		goto out_free_err_addr;
78 
79 	if (amdgpu_bad_page_threshold != 0) {
80 		amdgpu_ras_add_bad_pages(adev, err_data.err_addr,
81 						err_data.err_addr_cnt);
82 		amdgpu_ras_save_bad_pages(adev, NULL);
83 	}
84 
85 out_free_err_addr:
86 	kfree(err_data.err_addr);
87 
88 out_fini_err_data:
89 	amdgpu_ras_error_data_fini(&err_data);
90 
91 	return ret;
92 }
93 
amdgpu_umc_handle_bad_pages(struct amdgpu_device * adev,void * ras_error_status)94 void amdgpu_umc_handle_bad_pages(struct amdgpu_device *adev,
95 			void *ras_error_status)
96 {
97 	struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status;
98 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
99 	unsigned int error_query_mode;
100 	int ret = 0;
101 	unsigned long err_count;
102 
103 	amdgpu_ras_get_error_query_mode(adev, &error_query_mode);
104 
105 	mutex_lock(&con->page_retirement_lock);
106 	ret = amdgpu_dpm_get_ecc_info(adev, (void *)&(con->umc_ecc));
107 	if (ret == -EOPNOTSUPP &&
108 	    error_query_mode == AMDGPU_RAS_DIRECT_ERROR_QUERY) {
109 		if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
110 		    adev->umc.ras->ras_block.hw_ops->query_ras_error_count)
111 		    adev->umc.ras->ras_block.hw_ops->query_ras_error_count(adev, ras_error_status);
112 
113 		if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
114 		    adev->umc.ras->ras_block.hw_ops->query_ras_error_address &&
115 		    adev->umc.max_ras_err_cnt_per_query) {
116 			err_data->err_addr =
117 				kcalloc(adev->umc.max_ras_err_cnt_per_query,
118 					sizeof(struct eeprom_table_record), GFP_KERNEL);
119 
120 			/* still call query_ras_error_address to clear error status
121 			 * even NOMEM error is encountered
122 			 */
123 			if(!err_data->err_addr)
124 				dev_warn(adev->dev, "Failed to alloc memory for "
125 						"umc error address record!\n");
126 			else
127 				err_data->err_addr_len = adev->umc.max_ras_err_cnt_per_query;
128 
129 			/* umc query_ras_error_address is also responsible for clearing
130 			 * error status
131 			 */
132 			adev->umc.ras->ras_block.hw_ops->query_ras_error_address(adev, ras_error_status);
133 		}
134 	} else if (error_query_mode == AMDGPU_RAS_FIRMWARE_ERROR_QUERY ||
135 	    (!ret && error_query_mode == AMDGPU_RAS_DIRECT_ERROR_QUERY)) {
136 		if (adev->umc.ras &&
137 		    adev->umc.ras->ecc_info_query_ras_error_count)
138 		    adev->umc.ras->ecc_info_query_ras_error_count(adev, ras_error_status);
139 
140 		if (adev->umc.ras &&
141 		    adev->umc.ras->ecc_info_query_ras_error_address &&
142 		    adev->umc.max_ras_err_cnt_per_query) {
143 			err_data->err_addr =
144 				kcalloc(adev->umc.max_ras_err_cnt_per_query,
145 					sizeof(struct eeprom_table_record), GFP_KERNEL);
146 
147 			/* still call query_ras_error_address to clear error status
148 			 * even NOMEM error is encountered
149 			 */
150 			if(!err_data->err_addr)
151 				dev_warn(adev->dev, "Failed to alloc memory for "
152 						"umc error address record!\n");
153 			else
154 				err_data->err_addr_len = adev->umc.max_ras_err_cnt_per_query;
155 
156 			/* umc query_ras_error_address is also responsible for clearing
157 			 * error status
158 			 */
159 			adev->umc.ras->ecc_info_query_ras_error_address(adev, ras_error_status);
160 		}
161 	}
162 
163 	/* only uncorrectable error needs gpu reset */
164 	if (err_data->ue_count || err_data->de_count) {
165 		err_count = err_data->ue_count + err_data->de_count;
166 		if ((amdgpu_bad_page_threshold != 0) &&
167 			err_data->err_addr_cnt) {
168 			amdgpu_ras_add_bad_pages(adev, err_data->err_addr,
169 						err_data->err_addr_cnt);
170 			amdgpu_ras_save_bad_pages(adev, &err_count);
171 
172 			amdgpu_dpm_send_hbm_bad_pages_num(adev, con->eeprom_control.ras_num_recs);
173 
174 			if (con->update_channel_flag == true) {
175 				amdgpu_dpm_send_hbm_bad_channel_flag(adev, con->eeprom_control.bad_channel_bitmap);
176 				con->update_channel_flag = false;
177 			}
178 		}
179 	}
180 
181 	kfree(err_data->err_addr);
182 	err_data->err_addr = NULL;
183 
184 	mutex_unlock(&con->page_retirement_lock);
185 }
186 
amdgpu_umc_do_page_retirement(struct amdgpu_device * adev,void * ras_error_status,struct amdgpu_iv_entry * entry,uint32_t reset)187 static int amdgpu_umc_do_page_retirement(struct amdgpu_device *adev,
188 		void *ras_error_status,
189 		struct amdgpu_iv_entry *entry,
190 		uint32_t reset)
191 {
192 	struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status;
193 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
194 
195 	kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
196 	amdgpu_umc_handle_bad_pages(adev, ras_error_status);
197 
198 	if ((err_data->ue_count || err_data->de_count) &&
199 	    (reset || amdgpu_ras_is_rma(adev))) {
200 		con->gpu_reset_flags |= reset;
201 		amdgpu_ras_reset_gpu(adev);
202 	}
203 
204 	return AMDGPU_RAS_SUCCESS;
205 }
206 
amdgpu_umc_pasid_poison_handler(struct amdgpu_device * adev,enum amdgpu_ras_block block,uint16_t pasid,pasid_notify pasid_fn,void * data,uint32_t reset)207 int amdgpu_umc_pasid_poison_handler(struct amdgpu_device *adev,
208 			enum amdgpu_ras_block block, uint16_t pasid,
209 			pasid_notify pasid_fn, void *data, uint32_t reset)
210 {
211 	int ret = AMDGPU_RAS_SUCCESS;
212 
213 	if (adev->gmc.xgmi.connected_to_cpu ||
214 		adev->gmc.is_app_apu) {
215 		if (reset) {
216 			/* MCA poison handler is only responsible for GPU reset,
217 			 * let MCA notifier do page retirement.
218 			 */
219 			kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
220 			amdgpu_ras_reset_gpu(adev);
221 		}
222 		return ret;
223 	}
224 
225 	if (!amdgpu_sriov_vf(adev)) {
226 		if (amdgpu_ip_version(adev, UMC_HWIP, 0) < IP_VERSION(12, 0, 0)) {
227 			struct ras_err_data err_data;
228 			struct ras_common_if head = {
229 				.block = AMDGPU_RAS_BLOCK__UMC,
230 			};
231 			struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head);
232 
233 			ret = amdgpu_ras_error_data_init(&err_data);
234 			if (ret)
235 				return ret;
236 
237 			ret = amdgpu_umc_do_page_retirement(adev, &err_data, NULL, reset);
238 
239 			if (ret == AMDGPU_RAS_SUCCESS && obj) {
240 				obj->err_data.ue_count += err_data.ue_count;
241 				obj->err_data.ce_count += err_data.ce_count;
242 				obj->err_data.de_count += err_data.de_count;
243 			}
244 
245 			amdgpu_ras_error_data_fini(&err_data);
246 		} else {
247 			struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
248 			int ret;
249 
250 			ret = amdgpu_ras_put_poison_req(adev,
251 				block, pasid, pasid_fn, data, reset);
252 			if (!ret) {
253 				atomic_inc(&con->page_retirement_req_cnt);
254 				wake_up(&con->page_retirement_wq);
255 			}
256 		}
257 	} else {
258 		if (adev->virt.ops && adev->virt.ops->ras_poison_handler)
259 			adev->virt.ops->ras_poison_handler(adev, block);
260 		else
261 			dev_warn(adev->dev,
262 				"No ras_poison_handler interface in SRIOV!\n");
263 	}
264 
265 	return ret;
266 }
267 
amdgpu_umc_poison_handler(struct amdgpu_device * adev,enum amdgpu_ras_block block,uint32_t reset)268 int amdgpu_umc_poison_handler(struct amdgpu_device *adev,
269 			enum amdgpu_ras_block block, uint32_t reset)
270 {
271 	return amdgpu_umc_pasid_poison_handler(adev,
272 				block, 0, NULL, NULL, reset);
273 }
274 
amdgpu_umc_process_ras_data_cb(struct amdgpu_device * adev,void * ras_error_status,struct amdgpu_iv_entry * entry)275 int amdgpu_umc_process_ras_data_cb(struct amdgpu_device *adev,
276 		void *ras_error_status,
277 		struct amdgpu_iv_entry *entry)
278 {
279 	return amdgpu_umc_do_page_retirement(adev, ras_error_status, entry,
280 				AMDGPU_RAS_GPU_RESET_MODE1_RESET);
281 }
282 
amdgpu_umc_ras_sw_init(struct amdgpu_device * adev)283 int amdgpu_umc_ras_sw_init(struct amdgpu_device *adev)
284 {
285 	int err;
286 	struct amdgpu_umc_ras *ras;
287 
288 	if (!adev->umc.ras)
289 		return 0;
290 
291 	ras = adev->umc.ras;
292 
293 	err = amdgpu_ras_register_ras_block(adev, &ras->ras_block);
294 	if (err) {
295 		dev_err(adev->dev, "Failed to register umc ras block!\n");
296 		return err;
297 	}
298 
299 	strcpy(adev->umc.ras->ras_block.ras_comm.name, "umc");
300 	ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__UMC;
301 	ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
302 	adev->umc.ras_if = &ras->ras_block.ras_comm;
303 
304 	if (!ras->ras_block.ras_late_init)
305 		ras->ras_block.ras_late_init = amdgpu_umc_ras_late_init;
306 
307 	if (!ras->ras_block.ras_cb)
308 		ras->ras_block.ras_cb = amdgpu_umc_process_ras_data_cb;
309 
310 	return 0;
311 }
312 
amdgpu_umc_ras_late_init(struct amdgpu_device * adev,struct ras_common_if * ras_block)313 int amdgpu_umc_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block)
314 {
315 	int r;
316 
317 	r = amdgpu_ras_block_late_init(adev, ras_block);
318 	if (r)
319 		return r;
320 
321 	if (amdgpu_ras_is_supported(adev, ras_block->block)) {
322 		r = amdgpu_irq_get(adev, &adev->gmc.ecc_irq, 0);
323 		if (r)
324 			goto late_fini;
325 	}
326 
327 	/* ras init of specific umc version */
328 	if (adev->umc.ras &&
329 	    adev->umc.ras->err_cnt_init)
330 		adev->umc.ras->err_cnt_init(adev);
331 
332 	return 0;
333 
334 late_fini:
335 	amdgpu_ras_block_late_fini(adev, ras_block);
336 	return r;
337 }
338 
amdgpu_umc_process_ecc_irq(struct amdgpu_device * adev,struct amdgpu_irq_src * source,struct amdgpu_iv_entry * entry)339 int amdgpu_umc_process_ecc_irq(struct amdgpu_device *adev,
340 		struct amdgpu_irq_src *source,
341 		struct amdgpu_iv_entry *entry)
342 {
343 	struct ras_common_if *ras_if = adev->umc.ras_if;
344 	struct ras_dispatch_if ih_data = {
345 		.entry = entry,
346 	};
347 
348 	if (!ras_if)
349 		return 0;
350 
351 	ih_data.head = *ras_if;
352 
353 	amdgpu_ras_interrupt_dispatch(adev, &ih_data);
354 	return 0;
355 }
356 
amdgpu_umc_fill_error_record(struct ras_err_data * err_data,uint64_t err_addr,uint64_t retired_page,uint32_t channel_index,uint32_t umc_inst)357 int amdgpu_umc_fill_error_record(struct ras_err_data *err_data,
358 		uint64_t err_addr,
359 		uint64_t retired_page,
360 		uint32_t channel_index,
361 		uint32_t umc_inst)
362 {
363 	struct eeprom_table_record *err_rec;
364 
365 	if (!err_data ||
366 	    !err_data->err_addr ||
367 	    (err_data->err_addr_cnt >= err_data->err_addr_len))
368 		return -EINVAL;
369 
370 	err_rec = &err_data->err_addr[err_data->err_addr_cnt];
371 
372 	err_rec->address = err_addr;
373 	/* page frame address is saved */
374 	err_rec->retired_page = retired_page >> AMDGPU_GPU_PAGE_SHIFT;
375 	err_rec->ts = (uint64_t)ktime_get_real_seconds();
376 	err_rec->err_type = AMDGPU_RAS_EEPROM_ERR_NON_RECOVERABLE;
377 	err_rec->cu = 0;
378 	err_rec->mem_channel = channel_index;
379 	err_rec->mcumc_id = umc_inst;
380 
381 	err_data->err_addr_cnt++;
382 
383 	return 0;
384 }
385 
amdgpu_umc_loop_channels(struct amdgpu_device * adev,umc_func func,void * data)386 int amdgpu_umc_loop_channels(struct amdgpu_device *adev,
387 			umc_func func, void *data)
388 {
389 	uint32_t node_inst       = 0;
390 	uint32_t umc_inst        = 0;
391 	uint32_t ch_inst         = 0;
392 	int ret = 0;
393 
394 	if (adev->umc.node_inst_num) {
395 		LOOP_UMC_EACH_NODE_INST_AND_CH(node_inst, umc_inst, ch_inst) {
396 			ret = func(adev, node_inst, umc_inst, ch_inst, data);
397 			if (ret) {
398 				dev_err(adev->dev, "Node %d umc %d ch %d func returns %d\n",
399 					node_inst, umc_inst, ch_inst, ret);
400 				return ret;
401 			}
402 		}
403 	} else {
404 		LOOP_UMC_INST_AND_CH(umc_inst, ch_inst) {
405 			ret = func(adev, 0, umc_inst, ch_inst, data);
406 			if (ret) {
407 				dev_err(adev->dev, "Umc %d ch %d func returns %d\n",
408 					umc_inst, ch_inst, ret);
409 				return ret;
410 			}
411 		}
412 	}
413 
414 	return 0;
415 }
416 
amdgpu_umc_update_ecc_status(struct amdgpu_device * adev,uint64_t status,uint64_t ipid,uint64_t addr)417 int amdgpu_umc_update_ecc_status(struct amdgpu_device *adev,
418 				uint64_t status, uint64_t ipid, uint64_t addr)
419 {
420 	if (adev->umc.ras->update_ecc_status)
421 		return adev->umc.ras->update_ecc_status(adev,
422 					status, ipid, addr);
423 	return 0;
424 }
425 
amdgpu_umc_logs_ecc_err(struct amdgpu_device * adev,struct radix_tree_root * ecc_tree,struct ras_ecc_err * ecc_err)426 int amdgpu_umc_logs_ecc_err(struct amdgpu_device *adev,
427 		struct radix_tree_root *ecc_tree, struct ras_ecc_err *ecc_err)
428 {
429 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
430 	struct ras_ecc_log_info *ecc_log;
431 	int ret;
432 
433 	ecc_log = &con->umc_ecc_log;
434 
435 	mutex_lock(&ecc_log->lock);
436 	ret = radix_tree_insert(ecc_tree, ecc_err->pa_pfn, ecc_err);
437 	if (!ret)
438 		radix_tree_tag_set(ecc_tree,
439 			ecc_err->pa_pfn, UMC_ECC_NEW_DETECTED_TAG);
440 	mutex_unlock(&ecc_log->lock);
441 
442 	return ret;
443 }
444