xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c (revision 5c8d5e2619f7d2985adfe45608dc942ca8151aa3)
1 /*
2  * Copyright 2018 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/debugfs.h>
25 #include <linux/list.h>
26 #include <linux/module.h>
27 #include <linux/uaccess.h>
28 #include <linux/reboot.h>
29 #include <linux/syscalls.h>
30 #include <linux/pm_runtime.h>
31 #include <linux/list_sort.h>
32 
33 #include "amdgpu.h"
34 #include "amdgpu_ras.h"
35 #include "amdgpu_atomfirmware.h"
36 #include "amdgpu_xgmi.h"
37 #include "ivsrcid/nbio/irqsrcs_nbif_7_4.h"
38 #include "nbio_v4_3.h"
39 #include "nbif_v6_3_1.h"
40 #include "nbio_v7_9.h"
41 #include "atom.h"
42 #include "amdgpu_reset.h"
43 #include "amdgpu_psp.h"
44 
45 #ifdef CONFIG_X86_MCE_AMD
46 #include <asm/mce.h>
47 
48 static bool notifier_registered;
49 #endif
50 static const char *RAS_FS_NAME = "ras";
51 
52 const char *ras_error_string[] = {
53 	"none",
54 	"parity",
55 	"single_correctable",
56 	"multi_uncorrectable",
57 	"poison",
58 };
59 
60 const char *ras_block_string[] = {
61 	"umc",
62 	"sdma",
63 	"gfx",
64 	"mmhub",
65 	"athub",
66 	"pcie_bif",
67 	"hdp",
68 	"xgmi_wafl",
69 	"df",
70 	"smn",
71 	"sem",
72 	"mp0",
73 	"mp1",
74 	"fuse",
75 	"mca",
76 	"vcn",
77 	"jpeg",
78 	"ih",
79 	"mpio",
80 	"mmsch",
81 };
82 
83 const char *ras_mca_block_string[] = {
84 	"mca_mp0",
85 	"mca_mp1",
86 	"mca_mpio",
87 	"mca_iohc",
88 };
89 
90 struct amdgpu_ras_block_list {
91 	/* ras block link */
92 	struct list_head node;
93 
94 	struct amdgpu_ras_block_object *ras_obj;
95 };
96 
97 const char *get_ras_block_str(struct ras_common_if *ras_block)
98 {
99 	if (!ras_block)
100 		return "NULL";
101 
102 	if (ras_block->block >= AMDGPU_RAS_BLOCK_COUNT ||
103 	    ras_block->block >= ARRAY_SIZE(ras_block_string))
104 		return "OUT OF RANGE";
105 
106 	if (ras_block->block == AMDGPU_RAS_BLOCK__MCA)
107 		return ras_mca_block_string[ras_block->sub_block_index];
108 
109 	return ras_block_string[ras_block->block];
110 }
111 
112 #define ras_block_str(_BLOCK_) \
113 	(((_BLOCK_) < ARRAY_SIZE(ras_block_string)) ? ras_block_string[_BLOCK_] : "Out Of Range")
114 
115 #define ras_err_str(i) (ras_error_string[ffs(i)])
116 
117 #define RAS_DEFAULT_FLAGS (AMDGPU_RAS_FLAG_INIT_BY_VBIOS)
118 
119 /* inject address is 52 bits */
120 #define	RAS_UMC_INJECT_ADDR_LIMIT	(0x1ULL << 52)
121 
122 /* typical ECC bad page rate is 1 bad page per 100MB VRAM */
123 #define RAS_BAD_PAGE_COVER              (100 * 1024 * 1024ULL)
124 
125 #define MAX_UMC_POISON_POLLING_TIME_ASYNC  300  //ms
126 
127 #define AMDGPU_RAS_RETIRE_PAGE_INTERVAL 100  //ms
128 
129 #define MAX_FLUSH_RETIRE_DWORK_TIMES  100
130 
131 #define BYPASS_ALLOCATED_ADDRESS        0x0
132 #define BYPASS_INITIALIZATION_ADDRESS   0x1
133 
134 enum amdgpu_ras_retire_page_reservation {
135 	AMDGPU_RAS_RETIRE_PAGE_RESERVED,
136 	AMDGPU_RAS_RETIRE_PAGE_PENDING,
137 	AMDGPU_RAS_RETIRE_PAGE_FAULT,
138 };
139 
140 atomic_t amdgpu_ras_in_intr = ATOMIC_INIT(0);
141 
142 static int amdgpu_ras_check_bad_page_unlock(struct amdgpu_ras *con,
143 				uint64_t addr);
144 static int amdgpu_ras_check_bad_page(struct amdgpu_device *adev,
145 				uint64_t addr);
146 
147 static void amdgpu_ras_critical_region_init(struct amdgpu_device *adev);
148 static void amdgpu_ras_critical_region_fini(struct amdgpu_device *adev);
149 
150 #ifdef CONFIG_X86_MCE_AMD
151 static void amdgpu_register_bad_pages_mca_notifier(struct amdgpu_device *adev);
152 struct mce_notifier_adev_list {
153 	struct amdgpu_device *devs[MAX_GPU_INSTANCE];
154 	int num_gpu;
155 };
156 static struct mce_notifier_adev_list mce_adev_list;
157 #endif
158 
159 void amdgpu_ras_set_error_query_ready(struct amdgpu_device *adev, bool ready)
160 {
161 	if (adev && amdgpu_ras_get_context(adev))
162 		amdgpu_ras_get_context(adev)->error_query_ready = ready;
163 }
164 
165 static bool amdgpu_ras_get_error_query_ready(struct amdgpu_device *adev)
166 {
167 	if (adev && amdgpu_ras_get_context(adev))
168 		return amdgpu_ras_get_context(adev)->error_query_ready;
169 
170 	return false;
171 }
172 
173 static int amdgpu_reserve_page_direct(struct amdgpu_device *adev, uint64_t address)
174 {
175 	struct ras_err_data err_data;
176 	struct eeprom_table_record err_rec;
177 	int ret;
178 
179 	ret = amdgpu_ras_check_bad_page(adev, address);
180 	if (ret == -EINVAL) {
181 		dev_warn(adev->dev,
182 			"RAS WARN: input address 0x%llx is invalid.\n",
183 			address);
184 		return -EINVAL;
185 	} else if (ret == 1) {
186 		dev_warn(adev->dev,
187 			"RAS WARN: 0x%llx has already been marked as bad page!\n",
188 			address);
189 		return 0;
190 	}
191 
192 	ret = amdgpu_ras_error_data_init(&err_data);
193 	if (ret)
194 		return ret;
195 
196 	memset(&err_rec, 0x0, sizeof(struct eeprom_table_record));
197 	err_data.err_addr = &err_rec;
198 	amdgpu_umc_fill_error_record(&err_data, address, address, 0, 0);
199 
200 	if (amdgpu_bad_page_threshold != 0) {
201 		amdgpu_ras_add_bad_pages(adev, err_data.err_addr,
202 					 err_data.err_addr_cnt, false);
203 		amdgpu_ras_save_bad_pages(adev, NULL);
204 	}
205 
206 	amdgpu_ras_error_data_fini(&err_data);
207 
208 	dev_warn(adev->dev, "WARNING: THIS IS ONLY FOR TEST PURPOSES AND WILL CORRUPT RAS EEPROM\n");
209 	dev_warn(adev->dev, "Clear EEPROM:\n");
210 	dev_warn(adev->dev, "    echo 1 > /sys/kernel/debug/dri/0/ras/ras_eeprom_reset\n");
211 
212 	return 0;
213 }
214 
215 static int amdgpu_check_address_validity(struct amdgpu_device *adev,
216 			uint64_t address, uint64_t flags)
217 {
218 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
219 	struct amdgpu_vram_block_info blk_info;
220 	uint64_t page_pfns[32] = {0};
221 	int i, ret, count;
222 
223 	if (amdgpu_ip_version(adev, UMC_HWIP, 0) < IP_VERSION(12, 0, 0))
224 		return 0;
225 
226 	if ((address >= adev->gmc.mc_vram_size) ||
227 	    (address >= RAS_UMC_INJECT_ADDR_LIMIT))
228 		return -EFAULT;
229 
230 	count = amdgpu_umc_lookup_bad_pages_in_a_row(adev,
231 				address, page_pfns, ARRAY_SIZE(page_pfns));
232 	if (count <= 0)
233 		return -EPERM;
234 
235 	for (i = 0; i < count; i++) {
236 		memset(&blk_info, 0, sizeof(blk_info));
237 		ret = amdgpu_vram_mgr_query_address_block_info(&adev->mman.vram_mgr,
238 					page_pfns[i] << AMDGPU_GPU_PAGE_SHIFT, &blk_info);
239 		if (!ret) {
240 			/* The input address that needs to be checked is allocated by
241 			 * current calling process, so it is necessary to exclude
242 			 * the calling process.
243 			 */
244 			if ((flags == BYPASS_ALLOCATED_ADDRESS) &&
245 			    ((blk_info.task.pid != task_pid_nr(current)) ||
246 				strncmp(blk_info.task.comm, current->comm, TASK_COMM_LEN)))
247 				return -EACCES;
248 			else if ((flags == BYPASS_INITIALIZATION_ADDRESS) &&
249 				(blk_info.task.pid == con->init_task_pid) &&
250 				!strncmp(blk_info.task.comm, con->init_task_comm, TASK_COMM_LEN))
251 				return -EACCES;
252 		}
253 	}
254 
255 	return 0;
256 }
257 
258 static ssize_t amdgpu_ras_debugfs_read(struct file *f, char __user *buf,
259 					size_t size, loff_t *pos)
260 {
261 	struct ras_manager *obj = (struct ras_manager *)file_inode(f)->i_private;
262 	struct ras_query_if info = {
263 		.head = obj->head,
264 	};
265 	ssize_t s;
266 	char val[128];
267 
268 	if (amdgpu_ras_query_error_status(obj->adev, &info))
269 		return -EINVAL;
270 
271 	/* Hardware counter will be reset automatically after the query on Vega20 and Arcturus */
272 	if (amdgpu_ip_version(obj->adev, MP0_HWIP, 0) != IP_VERSION(11, 0, 2) &&
273 	    amdgpu_ip_version(obj->adev, MP0_HWIP, 0) != IP_VERSION(11, 0, 4)) {
274 		if (amdgpu_ras_reset_error_status(obj->adev, info.head.block))
275 			dev_warn(obj->adev->dev, "Failed to reset error counter and error status");
276 	}
277 
278 	s = snprintf(val, sizeof(val), "%s: %lu\n%s: %lu\n",
279 			"ue", info.ue_count,
280 			"ce", info.ce_count);
281 	if (*pos >= s)
282 		return 0;
283 
284 	s -= *pos;
285 	s = min_t(u64, s, size);
286 
287 
288 	if (copy_to_user(buf, &val[*pos], s))
289 		return -EINVAL;
290 
291 	*pos += s;
292 
293 	return s;
294 }
295 
296 static const struct file_operations amdgpu_ras_debugfs_ops = {
297 	.owner = THIS_MODULE,
298 	.read = amdgpu_ras_debugfs_read,
299 	.write = NULL,
300 	.llseek = default_llseek
301 };
302 
303 static int amdgpu_ras_find_block_id_by_name(const char *name, int *block_id)
304 {
305 	int i;
306 
307 	for (i = 0; i < ARRAY_SIZE(ras_block_string); i++) {
308 		*block_id = i;
309 		if (strcmp(name, ras_block_string[i]) == 0)
310 			return 0;
311 	}
312 	return -EINVAL;
313 }
314 
315 static int amdgpu_ras_debugfs_ctrl_parse_data(struct file *f,
316 		const char __user *buf, size_t size,
317 		loff_t *pos, struct ras_debug_if *data)
318 {
319 	ssize_t s = min_t(u64, 64, size);
320 	char str[65];
321 	char block_name[33];
322 	char err[9] = "ue";
323 	int op = -1;
324 	int block_id;
325 	uint32_t sub_block;
326 	u64 address, value;
327 	/* default value is 0 if the mask is not set by user */
328 	u32 instance_mask = 0;
329 
330 	if (*pos)
331 		return -EINVAL;
332 	*pos = size;
333 
334 	memset(str, 0, sizeof(str));
335 	memset(data, 0, sizeof(*data));
336 
337 	if (copy_from_user(str, buf, s))
338 		return -EINVAL;
339 
340 	if (sscanf(str, "disable %32s", block_name) == 1)
341 		op = 0;
342 	else if (sscanf(str, "enable %32s %8s", block_name, err) == 2)
343 		op = 1;
344 	else if (sscanf(str, "inject %32s %8s", block_name, err) == 2)
345 		op = 2;
346 	else if (strstr(str, "retire_page") != NULL)
347 		op = 3;
348 	else if (strstr(str, "check_address") != NULL)
349 		op = 4;
350 	else if (str[0] && str[1] && str[2] && str[3])
351 		/* ascii string, but commands are not matched. */
352 		return -EINVAL;
353 
354 	if (op != -1) {
355 		if (op == 3) {
356 			if (sscanf(str, "%*s 0x%llx", &address) != 1 &&
357 			    sscanf(str, "%*s %llu", &address) != 1)
358 				return -EINVAL;
359 
360 			data->op = op;
361 			data->inject.address = address;
362 
363 			return 0;
364 		} else if (op == 4) {
365 			if (sscanf(str, "%*s 0x%llx 0x%llx", &address, &value) != 2 &&
366 			    sscanf(str, "%*s %llu %llu", &address, &value) != 2)
367 				return -EINVAL;
368 
369 			data->op = op;
370 			data->inject.address = address;
371 			data->inject.value = value;
372 			return 0;
373 		}
374 
375 		if (amdgpu_ras_find_block_id_by_name(block_name, &block_id))
376 			return -EINVAL;
377 
378 		data->head.block = block_id;
379 		/* only ue, ce and poison errors are supported */
380 		if (!memcmp("ue", err, 2))
381 			data->head.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
382 		else if (!memcmp("ce", err, 2))
383 			data->head.type = AMDGPU_RAS_ERROR__SINGLE_CORRECTABLE;
384 		else if (!memcmp("poison", err, 6))
385 			data->head.type = AMDGPU_RAS_ERROR__POISON;
386 		else
387 			return -EINVAL;
388 
389 		data->op = op;
390 
391 		if (op == 2) {
392 			if (sscanf(str, "%*s %*s %*s 0x%x 0x%llx 0x%llx 0x%x",
393 				   &sub_block, &address, &value, &instance_mask) != 4 &&
394 			    sscanf(str, "%*s %*s %*s %u %llu %llu %u",
395 				   &sub_block, &address, &value, &instance_mask) != 4 &&
396 				sscanf(str, "%*s %*s %*s 0x%x 0x%llx 0x%llx",
397 				   &sub_block, &address, &value) != 3 &&
398 			    sscanf(str, "%*s %*s %*s %u %llu %llu",
399 				   &sub_block, &address, &value) != 3)
400 				return -EINVAL;
401 			data->head.sub_block_index = sub_block;
402 			data->inject.address = address;
403 			data->inject.value = value;
404 			data->inject.instance_mask = instance_mask;
405 		}
406 	} else {
407 		if (size < sizeof(*data))
408 			return -EINVAL;
409 
410 		if (copy_from_user(data, buf, sizeof(*data)))
411 			return -EINVAL;
412 	}
413 
414 	return 0;
415 }
416 
417 static void amdgpu_ras_instance_mask_check(struct amdgpu_device *adev,
418 				struct ras_debug_if *data)
419 {
420 	int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;
421 	uint32_t mask, inst_mask = data->inject.instance_mask;
422 
423 	/* no need to set instance mask if there is only one instance */
424 	if (num_xcc <= 1 && inst_mask) {
425 		data->inject.instance_mask = 0;
426 		dev_dbg(adev->dev,
427 			"RAS inject mask(0x%x) isn't supported and force it to 0.\n",
428 			inst_mask);
429 
430 		return;
431 	}
432 
433 	switch (data->head.block) {
434 	case AMDGPU_RAS_BLOCK__GFX:
435 		mask = GENMASK(num_xcc - 1, 0);
436 		break;
437 	case AMDGPU_RAS_BLOCK__SDMA:
438 		mask = GENMASK(adev->sdma.num_instances - 1, 0);
439 		break;
440 	case AMDGPU_RAS_BLOCK__VCN:
441 	case AMDGPU_RAS_BLOCK__JPEG:
442 		mask = GENMASK(adev->vcn.num_vcn_inst - 1, 0);
443 		break;
444 	default:
445 		mask = inst_mask;
446 		break;
447 	}
448 
449 	/* remove invalid bits in instance mask */
450 	data->inject.instance_mask &= mask;
451 	if (inst_mask != data->inject.instance_mask)
452 		dev_dbg(adev->dev,
453 			"Adjust RAS inject mask 0x%x to 0x%x\n",
454 			inst_mask, data->inject.instance_mask);
455 }
456 
457 /**
458  * DOC: AMDGPU RAS debugfs control interface
459  *
460  * The control interface accepts struct ras_debug_if which has two members.
461  *
462  * First member: ras_debug_if::head or ras_debug_if::inject.
463  *
464  * head is used to indicate which IP block will be under control.
465  *
466  * head has four members, they are block, type, sub_block_index, name.
467  * block: which IP will be under control.
468  * type: what kind of error will be enabled/disabled/injected.
469  * sub_block_index: some IPs have subcomponets. say, GFX, sDMA.
470  * name: the name of IP.
471  *
472  * inject has three more members than head, they are address, value and mask.
473  * As their names indicate, inject operation will write the
474  * value to the address.
475  *
476  * The second member: struct ras_debug_if::op.
477  * It has three kinds of operations.
478  *
479  * - 0: disable RAS on the block. Take ::head as its data.
480  * - 1: enable RAS on the block. Take ::head as its data.
481  * - 2: inject errors on the block. Take ::inject as its data.
482  *
483  * How to use the interface?
484  *
485  * In a program
486  *
487  * Copy the struct ras_debug_if in your code and initialize it.
488  * Write the struct to the control interface.
489  *
490  * From shell
491  *
492  * .. code-block:: bash
493  *
494  *	echo "disable <block>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
495  *	echo "enable  <block> <error>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
496  *	echo "inject  <block> <error> <sub-block> <address> <value> <mask>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
497  *
498  * Where N, is the card which you want to affect.
499  *
500  * "disable" requires only the block.
501  * "enable" requires the block and error type.
502  * "inject" requires the block, error type, address, and value.
503  *
504  * The block is one of: umc, sdma, gfx, etc.
505  *	see ras_block_string[] for details
506  *
507  * The error type is one of: ue, ce and poison where,
508  *	ue is multi-uncorrectable
509  *	ce is single-correctable
510  *	poison is poison
511  *
512  * The sub-block is a the sub-block index, pass 0 if there is no sub-block.
513  * The address and value are hexadecimal numbers, leading 0x is optional.
514  * The mask means instance mask, is optional, default value is 0x1.
515  *
516  * For instance,
517  *
518  * .. code-block:: bash
519  *
520  *	echo inject umc ue 0x0 0x0 0x0 > /sys/kernel/debug/dri/0/ras/ras_ctrl
521  *	echo inject umc ce 0 0 0 3 > /sys/kernel/debug/dri/0/ras/ras_ctrl
522  *	echo disable umc > /sys/kernel/debug/dri/0/ras/ras_ctrl
523  *
524  * How to check the result of the operation?
525  *
526  * To check disable/enable, see "ras" features at,
527  * /sys/class/drm/card[0/1/2...]/device/ras/features
528  *
529  * To check inject, see the corresponding error count at,
530  * /sys/class/drm/card[0/1/2...]/device/ras/[gfx|sdma|umc|...]_err_count
531  *
532  * .. note::
533  *	Operations are only allowed on blocks which are supported.
534  *	Check the "ras" mask at /sys/module/amdgpu/parameters/ras_mask
535  *	to see which blocks support RAS on a particular asic.
536  *
537  */
538 static ssize_t amdgpu_ras_debugfs_ctrl_write(struct file *f,
539 					     const char __user *buf,
540 					     size_t size, loff_t *pos)
541 {
542 	struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
543 	struct ras_debug_if data;
544 	int ret = 0;
545 
546 	if (!amdgpu_ras_get_error_query_ready(adev)) {
547 		dev_warn(adev->dev, "RAS WARN: error injection "
548 				"currently inaccessible\n");
549 		return size;
550 	}
551 
552 	ret = amdgpu_ras_debugfs_ctrl_parse_data(f, buf, size, pos, &data);
553 	if (ret)
554 		return ret;
555 
556 	if (data.op == 3) {
557 		ret = amdgpu_reserve_page_direct(adev, data.inject.address);
558 		if (!ret)
559 			return size;
560 		else
561 			return ret;
562 	} else if (data.op == 4) {
563 		ret = amdgpu_check_address_validity(adev, data.inject.address, data.inject.value);
564 		return ret ? ret : size;
565 	}
566 
567 	if (!amdgpu_ras_is_supported(adev, data.head.block))
568 		return -EINVAL;
569 
570 	switch (data.op) {
571 	case 0:
572 		ret = amdgpu_ras_feature_enable(adev, &data.head, 0);
573 		break;
574 	case 1:
575 		ret = amdgpu_ras_feature_enable(adev, &data.head, 1);
576 		break;
577 	case 2:
578 		/* umc ce/ue error injection for a bad page is not allowed */
579 		if (data.head.block == AMDGPU_RAS_BLOCK__UMC)
580 			ret = amdgpu_ras_check_bad_page(adev, data.inject.address);
581 		if (ret == -EINVAL) {
582 			dev_warn(adev->dev, "RAS WARN: input address 0x%llx is invalid.",
583 					data.inject.address);
584 			break;
585 		} else if (ret == 1) {
586 			dev_warn(adev->dev, "RAS WARN: inject: 0x%llx has already been marked as bad!\n",
587 					data.inject.address);
588 			break;
589 		}
590 
591 		amdgpu_ras_instance_mask_check(adev, &data);
592 
593 		/* data.inject.address is offset instead of absolute gpu address */
594 		ret = amdgpu_ras_error_inject(adev, &data.inject);
595 		break;
596 	default:
597 		ret = -EINVAL;
598 		break;
599 	}
600 
601 	if (ret)
602 		return ret;
603 
604 	return size;
605 }
606 
607 /**
608  * DOC: AMDGPU RAS debugfs EEPROM table reset interface
609  *
610  * Some boards contain an EEPROM which is used to persistently store a list of
611  * bad pages which experiences ECC errors in vram.  This interface provides
612  * a way to reset the EEPROM, e.g., after testing error injection.
613  *
614  * Usage:
615  *
616  * .. code-block:: bash
617  *
618  *	echo 1 > ../ras/ras_eeprom_reset
619  *
620  * will reset EEPROM table to 0 entries.
621  *
622  */
623 static ssize_t amdgpu_ras_debugfs_eeprom_write(struct file *f,
624 					       const char __user *buf,
625 					       size_t size, loff_t *pos)
626 {
627 	struct amdgpu_device *adev =
628 		(struct amdgpu_device *)file_inode(f)->i_private;
629 	int ret;
630 
631 	ret = amdgpu_ras_eeprom_reset_table(
632 		&(amdgpu_ras_get_context(adev)->eeprom_control));
633 
634 	if (!ret) {
635 		/* Something was written to EEPROM.
636 		 */
637 		amdgpu_ras_get_context(adev)->flags = RAS_DEFAULT_FLAGS;
638 		return size;
639 	} else {
640 		return ret;
641 	}
642 }
643 
644 static const struct file_operations amdgpu_ras_debugfs_ctrl_ops = {
645 	.owner = THIS_MODULE,
646 	.read = NULL,
647 	.write = amdgpu_ras_debugfs_ctrl_write,
648 	.llseek = default_llseek
649 };
650 
651 static const struct file_operations amdgpu_ras_debugfs_eeprom_ops = {
652 	.owner = THIS_MODULE,
653 	.read = NULL,
654 	.write = amdgpu_ras_debugfs_eeprom_write,
655 	.llseek = default_llseek
656 };
657 
658 /**
659  * DOC: AMDGPU RAS sysfs Error Count Interface
660  *
661  * It allows the user to read the error count for each IP block on the gpu through
662  * /sys/class/drm/card[0/1/2...]/device/ras/[gfx/sdma/...]_err_count
663  *
664  * It outputs the multiple lines which report the uncorrected (ue) and corrected
665  * (ce) error counts.
666  *
667  * The format of one line is below,
668  *
669  * [ce|ue]: count
670  *
671  * Example:
672  *
673  * .. code-block:: bash
674  *
675  *	ue: 0
676  *	ce: 1
677  *
678  */
679 static ssize_t amdgpu_ras_sysfs_read(struct device *dev,
680 		struct device_attribute *attr, char *buf)
681 {
682 	struct ras_manager *obj = container_of(attr, struct ras_manager, sysfs_attr);
683 	struct ras_query_if info = {
684 		.head = obj->head,
685 	};
686 
687 	if (!amdgpu_ras_get_error_query_ready(obj->adev))
688 		return sysfs_emit(buf, "Query currently inaccessible\n");
689 
690 	if (amdgpu_ras_query_error_status(obj->adev, &info))
691 		return -EINVAL;
692 
693 	if (amdgpu_ip_version(obj->adev, MP0_HWIP, 0) != IP_VERSION(11, 0, 2) &&
694 	    amdgpu_ip_version(obj->adev, MP0_HWIP, 0) != IP_VERSION(11, 0, 4)) {
695 		if (amdgpu_ras_reset_error_status(obj->adev, info.head.block))
696 			dev_warn(obj->adev->dev, "Failed to reset error counter and error status");
697 	}
698 
699 	if (info.head.block == AMDGPU_RAS_BLOCK__UMC)
700 		return sysfs_emit(buf, "%s: %lu\n%s: %lu\n%s: %lu\n", "ue", info.ue_count,
701 				"ce", info.ce_count, "de", info.de_count);
702 	else
703 		return sysfs_emit(buf, "%s: %lu\n%s: %lu\n", "ue", info.ue_count,
704 				"ce", info.ce_count);
705 }
706 
707 /* obj begin */
708 
709 #define get_obj(obj) do { (obj)->use++; } while (0)
710 #define alive_obj(obj) ((obj)->use)
711 
712 static inline void put_obj(struct ras_manager *obj)
713 {
714 	if (obj && (--obj->use == 0)) {
715 		list_del(&obj->node);
716 		amdgpu_ras_error_data_fini(&obj->err_data);
717 	}
718 
719 	if (obj && (obj->use < 0))
720 		DRM_ERROR("RAS ERROR: Unbalance obj(%s) use\n", get_ras_block_str(&obj->head));
721 }
722 
723 /* make one obj and return it. */
724 static struct ras_manager *amdgpu_ras_create_obj(struct amdgpu_device *adev,
725 		struct ras_common_if *head)
726 {
727 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
728 	struct ras_manager *obj;
729 
730 	if (!adev->ras_enabled || !con)
731 		return NULL;
732 
733 	if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
734 		return NULL;
735 
736 	if (head->block == AMDGPU_RAS_BLOCK__MCA) {
737 		if (head->sub_block_index >= AMDGPU_RAS_MCA_BLOCK__LAST)
738 			return NULL;
739 
740 		obj = &con->objs[AMDGPU_RAS_BLOCK__LAST + head->sub_block_index];
741 	} else
742 		obj = &con->objs[head->block];
743 
744 	/* already exist. return obj? */
745 	if (alive_obj(obj))
746 		return NULL;
747 
748 	if (amdgpu_ras_error_data_init(&obj->err_data))
749 		return NULL;
750 
751 	obj->head = *head;
752 	obj->adev = adev;
753 	list_add(&obj->node, &con->head);
754 	get_obj(obj);
755 
756 	return obj;
757 }
758 
759 /* return an obj equal to head, or the first when head is NULL */
760 struct ras_manager *amdgpu_ras_find_obj(struct amdgpu_device *adev,
761 		struct ras_common_if *head)
762 {
763 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
764 	struct ras_manager *obj;
765 	int i;
766 
767 	if (!adev->ras_enabled || !con)
768 		return NULL;
769 
770 	if (head) {
771 		if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
772 			return NULL;
773 
774 		if (head->block == AMDGPU_RAS_BLOCK__MCA) {
775 			if (head->sub_block_index >= AMDGPU_RAS_MCA_BLOCK__LAST)
776 				return NULL;
777 
778 			obj = &con->objs[AMDGPU_RAS_BLOCK__LAST + head->sub_block_index];
779 		} else
780 			obj = &con->objs[head->block];
781 
782 		if (alive_obj(obj))
783 			return obj;
784 	} else {
785 		for (i = 0; i < AMDGPU_RAS_BLOCK_COUNT + AMDGPU_RAS_MCA_BLOCK_COUNT; i++) {
786 			obj = &con->objs[i];
787 			if (alive_obj(obj))
788 				return obj;
789 		}
790 	}
791 
792 	return NULL;
793 }
794 /* obj end */
795 
796 /* feature ctl begin */
797 static int amdgpu_ras_is_feature_allowed(struct amdgpu_device *adev,
798 					 struct ras_common_if *head)
799 {
800 	return adev->ras_hw_enabled & BIT(head->block);
801 }
802 
803 static int amdgpu_ras_is_feature_enabled(struct amdgpu_device *adev,
804 		struct ras_common_if *head)
805 {
806 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
807 
808 	return con->features & BIT(head->block);
809 }
810 
811 /*
812  * if obj is not created, then create one.
813  * set feature enable flag.
814  */
815 static int __amdgpu_ras_feature_enable(struct amdgpu_device *adev,
816 		struct ras_common_if *head, int enable)
817 {
818 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
819 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
820 
821 	/* If hardware does not support ras, then do not create obj.
822 	 * But if hardware support ras, we can create the obj.
823 	 * Ras framework checks con->hw_supported to see if it need do
824 	 * corresponding initialization.
825 	 * IP checks con->support to see if it need disable ras.
826 	 */
827 	if (!amdgpu_ras_is_feature_allowed(adev, head))
828 		return 0;
829 
830 	if (enable) {
831 		if (!obj) {
832 			obj = amdgpu_ras_create_obj(adev, head);
833 			if (!obj)
834 				return -EINVAL;
835 		} else {
836 			/* In case we create obj somewhere else */
837 			get_obj(obj);
838 		}
839 		con->features |= BIT(head->block);
840 	} else {
841 		if (obj && amdgpu_ras_is_feature_enabled(adev, head)) {
842 			con->features &= ~BIT(head->block);
843 			put_obj(obj);
844 		}
845 	}
846 
847 	return 0;
848 }
849 
850 /* wrapper of psp_ras_enable_features */
851 int amdgpu_ras_feature_enable(struct amdgpu_device *adev,
852 		struct ras_common_if *head, bool enable)
853 {
854 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
855 	union ta_ras_cmd_input *info;
856 	int ret;
857 
858 	if (!con)
859 		return -EINVAL;
860 
861 	/* For non-gfx ip, do not enable ras feature if it is not allowed */
862 	/* For gfx ip, regardless of feature support status, */
863 	/* Force issue enable or disable ras feature commands */
864 	if (head->block != AMDGPU_RAS_BLOCK__GFX &&
865 	    !amdgpu_ras_is_feature_allowed(adev, head))
866 		return 0;
867 
868 	/* Only enable gfx ras feature from host side */
869 	if (head->block == AMDGPU_RAS_BLOCK__GFX &&
870 	    !amdgpu_sriov_vf(adev) &&
871 	    !amdgpu_ras_intr_triggered()) {
872 		info = kzalloc(sizeof(union ta_ras_cmd_input), GFP_KERNEL);
873 		if (!info)
874 			return -ENOMEM;
875 
876 		if (!enable) {
877 			info->disable_features = (struct ta_ras_disable_features_input) {
878 				.block_id =  amdgpu_ras_block_to_ta(head->block),
879 				.error_type = amdgpu_ras_error_to_ta(head->type),
880 			};
881 		} else {
882 			info->enable_features = (struct ta_ras_enable_features_input) {
883 				.block_id =  amdgpu_ras_block_to_ta(head->block),
884 				.error_type = amdgpu_ras_error_to_ta(head->type),
885 			};
886 		}
887 
888 		ret = psp_ras_enable_features(&adev->psp, info, enable);
889 		if (ret) {
890 			dev_err(adev->dev, "ras %s %s failed poison:%d ret:%d\n",
891 				enable ? "enable":"disable",
892 				get_ras_block_str(head),
893 				amdgpu_ras_is_poison_mode_supported(adev), ret);
894 			kfree(info);
895 			return ret;
896 		}
897 
898 		kfree(info);
899 	}
900 
901 	/* setup the obj */
902 	__amdgpu_ras_feature_enable(adev, head, enable);
903 
904 	return 0;
905 }
906 
907 /* Only used in device probe stage and called only once. */
908 int amdgpu_ras_feature_enable_on_boot(struct amdgpu_device *adev,
909 		struct ras_common_if *head, bool enable)
910 {
911 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
912 	int ret;
913 
914 	if (!con)
915 		return -EINVAL;
916 
917 	if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
918 		if (enable) {
919 			/* There is no harm to issue a ras TA cmd regardless of
920 			 * the currecnt ras state.
921 			 * If current state == target state, it will do nothing
922 			 * But sometimes it requests driver to reset and repost
923 			 * with error code -EAGAIN.
924 			 */
925 			ret = amdgpu_ras_feature_enable(adev, head, 1);
926 			/* With old ras TA, we might fail to enable ras.
927 			 * Log it and just setup the object.
928 			 * TODO need remove this WA in the future.
929 			 */
930 			if (ret == -EINVAL) {
931 				ret = __amdgpu_ras_feature_enable(adev, head, 1);
932 				if (!ret)
933 					dev_info(adev->dev,
934 						"RAS INFO: %s setup object\n",
935 						get_ras_block_str(head));
936 			}
937 		} else {
938 			/* setup the object then issue a ras TA disable cmd.*/
939 			ret = __amdgpu_ras_feature_enable(adev, head, 1);
940 			if (ret)
941 				return ret;
942 
943 			/* gfx block ras disable cmd must send to ras-ta */
944 			if (head->block == AMDGPU_RAS_BLOCK__GFX)
945 				con->features |= BIT(head->block);
946 
947 			ret = amdgpu_ras_feature_enable(adev, head, 0);
948 
949 			/* clean gfx block ras features flag */
950 			if (adev->ras_enabled && head->block == AMDGPU_RAS_BLOCK__GFX)
951 				con->features &= ~BIT(head->block);
952 		}
953 	} else
954 		ret = amdgpu_ras_feature_enable(adev, head, enable);
955 
956 	return ret;
957 }
958 
959 static int amdgpu_ras_disable_all_features(struct amdgpu_device *adev,
960 		bool bypass)
961 {
962 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
963 	struct ras_manager *obj, *tmp;
964 
965 	list_for_each_entry_safe(obj, tmp, &con->head, node) {
966 		/* bypass psp.
967 		 * aka just release the obj and corresponding flags
968 		 */
969 		if (bypass) {
970 			if (__amdgpu_ras_feature_enable(adev, &obj->head, 0))
971 				break;
972 		} else {
973 			if (amdgpu_ras_feature_enable(adev, &obj->head, 0))
974 				break;
975 		}
976 	}
977 
978 	return con->features;
979 }
980 
981 static int amdgpu_ras_enable_all_features(struct amdgpu_device *adev,
982 		bool bypass)
983 {
984 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
985 	int i;
986 	const enum amdgpu_ras_error_type default_ras_type = AMDGPU_RAS_ERROR__NONE;
987 
988 	for (i = 0; i < AMDGPU_RAS_BLOCK_COUNT; i++) {
989 		struct ras_common_if head = {
990 			.block = i,
991 			.type = default_ras_type,
992 			.sub_block_index = 0,
993 		};
994 
995 		if (i == AMDGPU_RAS_BLOCK__MCA)
996 			continue;
997 
998 		if (bypass) {
999 			/*
1000 			 * bypass psp. vbios enable ras for us.
1001 			 * so just create the obj
1002 			 */
1003 			if (__amdgpu_ras_feature_enable(adev, &head, 1))
1004 				break;
1005 		} else {
1006 			if (amdgpu_ras_feature_enable(adev, &head, 1))
1007 				break;
1008 		}
1009 	}
1010 
1011 	for (i = 0; i < AMDGPU_RAS_MCA_BLOCK_COUNT; i++) {
1012 		struct ras_common_if head = {
1013 			.block = AMDGPU_RAS_BLOCK__MCA,
1014 			.type = default_ras_type,
1015 			.sub_block_index = i,
1016 		};
1017 
1018 		if (bypass) {
1019 			/*
1020 			 * bypass psp. vbios enable ras for us.
1021 			 * so just create the obj
1022 			 */
1023 			if (__amdgpu_ras_feature_enable(adev, &head, 1))
1024 				break;
1025 		} else {
1026 			if (amdgpu_ras_feature_enable(adev, &head, 1))
1027 				break;
1028 		}
1029 	}
1030 
1031 	return con->features;
1032 }
1033 /* feature ctl end */
1034 
1035 static int amdgpu_ras_block_match_default(struct amdgpu_ras_block_object *block_obj,
1036 		enum amdgpu_ras_block block)
1037 {
1038 	if (!block_obj)
1039 		return -EINVAL;
1040 
1041 	if (block_obj->ras_comm.block == block)
1042 		return 0;
1043 
1044 	return -EINVAL;
1045 }
1046 
1047 static struct amdgpu_ras_block_object *amdgpu_ras_get_ras_block(struct amdgpu_device *adev,
1048 					enum amdgpu_ras_block block, uint32_t sub_block_index)
1049 {
1050 	struct amdgpu_ras_block_list *node, *tmp;
1051 	struct amdgpu_ras_block_object *obj;
1052 
1053 	if (block >= AMDGPU_RAS_BLOCK__LAST)
1054 		return NULL;
1055 
1056 	list_for_each_entry_safe(node, tmp, &adev->ras_list, node) {
1057 		if (!node->ras_obj) {
1058 			dev_warn(adev->dev, "Warning: abnormal ras list node.\n");
1059 			continue;
1060 		}
1061 
1062 		obj = node->ras_obj;
1063 		if (obj->ras_block_match) {
1064 			if (obj->ras_block_match(obj, block, sub_block_index) == 0)
1065 				return obj;
1066 		} else {
1067 			if (amdgpu_ras_block_match_default(obj, block) == 0)
1068 				return obj;
1069 		}
1070 	}
1071 
1072 	return NULL;
1073 }
1074 
1075 static void amdgpu_ras_get_ecc_info(struct amdgpu_device *adev, struct ras_err_data *err_data)
1076 {
1077 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1078 	int ret = 0;
1079 
1080 	/*
1081 	 * choosing right query method according to
1082 	 * whether smu support query error information
1083 	 */
1084 	ret = amdgpu_dpm_get_ecc_info(adev, (void *)&(ras->umc_ecc));
1085 	if (ret == -EOPNOTSUPP) {
1086 		if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
1087 			adev->umc.ras->ras_block.hw_ops->query_ras_error_count)
1088 			adev->umc.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data);
1089 
1090 		/* umc query_ras_error_address is also responsible for clearing
1091 		 * error status
1092 		 */
1093 		if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
1094 		    adev->umc.ras->ras_block.hw_ops->query_ras_error_address)
1095 			adev->umc.ras->ras_block.hw_ops->query_ras_error_address(adev, err_data);
1096 	} else if (!ret) {
1097 		if (adev->umc.ras &&
1098 			adev->umc.ras->ecc_info_query_ras_error_count)
1099 			adev->umc.ras->ecc_info_query_ras_error_count(adev, err_data);
1100 
1101 		if (adev->umc.ras &&
1102 			adev->umc.ras->ecc_info_query_ras_error_address)
1103 			adev->umc.ras->ecc_info_query_ras_error_address(adev, err_data);
1104 	}
1105 }
1106 
1107 static void amdgpu_ras_error_print_error_data(struct amdgpu_device *adev,
1108 					      struct ras_manager *ras_mgr,
1109 					      struct ras_err_data *err_data,
1110 					      struct ras_query_context *qctx,
1111 					      const char *blk_name,
1112 					      bool is_ue,
1113 					      bool is_de)
1114 {
1115 	struct amdgpu_smuio_mcm_config_info *mcm_info;
1116 	struct ras_err_node *err_node;
1117 	struct ras_err_info *err_info;
1118 	u64 event_id = qctx->evid.event_id;
1119 
1120 	if (is_ue) {
1121 		for_each_ras_error(err_node, err_data) {
1122 			err_info = &err_node->err_info;
1123 			mcm_info = &err_info->mcm_info;
1124 			if (err_info->ue_count) {
1125 				RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d, "
1126 					      "%lld new uncorrectable hardware errors detected in %s block\n",
1127 					      mcm_info->socket_id,
1128 					      mcm_info->die_id,
1129 					      err_info->ue_count,
1130 					      blk_name);
1131 			}
1132 		}
1133 
1134 		for_each_ras_error(err_node, &ras_mgr->err_data) {
1135 			err_info = &err_node->err_info;
1136 			mcm_info = &err_info->mcm_info;
1137 			RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d, "
1138 				      "%lld uncorrectable hardware errors detected in total in %s block\n",
1139 				      mcm_info->socket_id, mcm_info->die_id, err_info->ue_count, blk_name);
1140 		}
1141 
1142 	} else {
1143 		if (is_de) {
1144 			for_each_ras_error(err_node, err_data) {
1145 				err_info = &err_node->err_info;
1146 				mcm_info = &err_info->mcm_info;
1147 				if (err_info->de_count) {
1148 					RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d, "
1149 						      "%lld new deferred hardware errors detected in %s block\n",
1150 						      mcm_info->socket_id,
1151 						      mcm_info->die_id,
1152 						      err_info->de_count,
1153 						      blk_name);
1154 				}
1155 			}
1156 
1157 			for_each_ras_error(err_node, &ras_mgr->err_data) {
1158 				err_info = &err_node->err_info;
1159 				mcm_info = &err_info->mcm_info;
1160 				RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d, "
1161 					      "%lld deferred hardware errors detected in total in %s block\n",
1162 					      mcm_info->socket_id, mcm_info->die_id,
1163 					      err_info->de_count, blk_name);
1164 			}
1165 		} else {
1166 			if (adev->debug_disable_ce_logs)
1167 				return;
1168 
1169 			for_each_ras_error(err_node, err_data) {
1170 				err_info = &err_node->err_info;
1171 				mcm_info = &err_info->mcm_info;
1172 				if (err_info->ce_count) {
1173 					RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d, "
1174 						      "%lld new correctable hardware errors detected in %s block\n",
1175 						      mcm_info->socket_id,
1176 						      mcm_info->die_id,
1177 						      err_info->ce_count,
1178 						      blk_name);
1179 				}
1180 			}
1181 
1182 			for_each_ras_error(err_node, &ras_mgr->err_data) {
1183 				err_info = &err_node->err_info;
1184 				mcm_info = &err_info->mcm_info;
1185 				RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d, "
1186 					      "%lld correctable hardware errors detected in total in %s block\n",
1187 					      mcm_info->socket_id, mcm_info->die_id,
1188 					      err_info->ce_count, blk_name);
1189 			}
1190 		}
1191 	}
1192 }
1193 
1194 static inline bool err_data_has_source_info(struct ras_err_data *data)
1195 {
1196 	return !list_empty(&data->err_node_list);
1197 }
1198 
1199 static void amdgpu_ras_error_generate_report(struct amdgpu_device *adev,
1200 					     struct ras_query_if *query_if,
1201 					     struct ras_err_data *err_data,
1202 					     struct ras_query_context *qctx)
1203 {
1204 	struct ras_manager *ras_mgr = amdgpu_ras_find_obj(adev, &query_if->head);
1205 	const char *blk_name = get_ras_block_str(&query_if->head);
1206 	u64 event_id = qctx->evid.event_id;
1207 
1208 	if (err_data->ce_count) {
1209 		if (err_data_has_source_info(err_data)) {
1210 			amdgpu_ras_error_print_error_data(adev, ras_mgr, err_data, qctx,
1211 							  blk_name, false, false);
1212 		} else if (!adev->aid_mask &&
1213 			   adev->smuio.funcs &&
1214 			   adev->smuio.funcs->get_socket_id &&
1215 			   adev->smuio.funcs->get_die_id) {
1216 			RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d "
1217 				      "%ld correctable hardware errors "
1218 				      "detected in %s block\n",
1219 				      adev->smuio.funcs->get_socket_id(adev),
1220 				      adev->smuio.funcs->get_die_id(adev),
1221 				      ras_mgr->err_data.ce_count,
1222 				      blk_name);
1223 		} else {
1224 			RAS_EVENT_LOG(adev, event_id, "%ld correctable hardware errors "
1225 				      "detected in %s block\n",
1226 				      ras_mgr->err_data.ce_count,
1227 				      blk_name);
1228 		}
1229 	}
1230 
1231 	if (err_data->ue_count) {
1232 		if (err_data_has_source_info(err_data)) {
1233 			amdgpu_ras_error_print_error_data(adev, ras_mgr, err_data, qctx,
1234 							  blk_name, true, false);
1235 		} else if (!adev->aid_mask &&
1236 			   adev->smuio.funcs &&
1237 			   adev->smuio.funcs->get_socket_id &&
1238 			   adev->smuio.funcs->get_die_id) {
1239 			RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d "
1240 				      "%ld uncorrectable hardware errors "
1241 				      "detected in %s block\n",
1242 				      adev->smuio.funcs->get_socket_id(adev),
1243 				      adev->smuio.funcs->get_die_id(adev),
1244 				      ras_mgr->err_data.ue_count,
1245 				      blk_name);
1246 		} else {
1247 			RAS_EVENT_LOG(adev, event_id, "%ld uncorrectable hardware errors "
1248 				      "detected in %s block\n",
1249 				      ras_mgr->err_data.ue_count,
1250 				      blk_name);
1251 		}
1252 	}
1253 
1254 	if (err_data->de_count) {
1255 		if (err_data_has_source_info(err_data)) {
1256 			amdgpu_ras_error_print_error_data(adev, ras_mgr, err_data, qctx,
1257 							  blk_name, false, true);
1258 		} else if (!adev->aid_mask &&
1259 			   adev->smuio.funcs &&
1260 			   adev->smuio.funcs->get_socket_id &&
1261 			   adev->smuio.funcs->get_die_id) {
1262 			RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d "
1263 				      "%ld deferred hardware errors "
1264 				      "detected in %s block\n",
1265 				      adev->smuio.funcs->get_socket_id(adev),
1266 				      adev->smuio.funcs->get_die_id(adev),
1267 				      ras_mgr->err_data.de_count,
1268 				      blk_name);
1269 		} else {
1270 			RAS_EVENT_LOG(adev, event_id, "%ld deferred hardware errors "
1271 				      "detected in %s block\n",
1272 				      ras_mgr->err_data.de_count,
1273 				      blk_name);
1274 		}
1275 	}
1276 }
1277 
1278 static void amdgpu_ras_virt_error_generate_report(struct amdgpu_device *adev,
1279 						  struct ras_query_if *query_if,
1280 						  struct ras_err_data *err_data,
1281 						  struct ras_query_context *qctx)
1282 {
1283 	unsigned long new_ue, new_ce, new_de;
1284 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &query_if->head);
1285 	const char *blk_name = get_ras_block_str(&query_if->head);
1286 	u64 event_id = qctx->evid.event_id;
1287 
1288 	new_ce = err_data->ce_count - obj->err_data.ce_count;
1289 	new_ue = err_data->ue_count - obj->err_data.ue_count;
1290 	new_de = err_data->de_count - obj->err_data.de_count;
1291 
1292 	if (new_ce) {
1293 		RAS_EVENT_LOG(adev, event_id, "%lu correctable hardware errors "
1294 			      "detected in %s block\n",
1295 			      new_ce,
1296 			      blk_name);
1297 	}
1298 
1299 	if (new_ue) {
1300 		RAS_EVENT_LOG(adev, event_id, "%lu uncorrectable hardware errors "
1301 			      "detected in %s block\n",
1302 			      new_ue,
1303 			      blk_name);
1304 	}
1305 
1306 	if (new_de) {
1307 		RAS_EVENT_LOG(adev, event_id, "%lu deferred hardware errors "
1308 			      "detected in %s block\n",
1309 			      new_de,
1310 			      blk_name);
1311 	}
1312 }
1313 
1314 static void amdgpu_rasmgr_error_data_statistic_update(struct ras_manager *obj, struct ras_err_data *err_data)
1315 {
1316 	struct ras_err_node *err_node;
1317 	struct ras_err_info *err_info;
1318 
1319 	if (err_data_has_source_info(err_data)) {
1320 		for_each_ras_error(err_node, err_data) {
1321 			err_info = &err_node->err_info;
1322 			amdgpu_ras_error_statistic_de_count(&obj->err_data,
1323 					&err_info->mcm_info, err_info->de_count);
1324 			amdgpu_ras_error_statistic_ce_count(&obj->err_data,
1325 					&err_info->mcm_info, err_info->ce_count);
1326 			amdgpu_ras_error_statistic_ue_count(&obj->err_data,
1327 					&err_info->mcm_info, err_info->ue_count);
1328 		}
1329 	} else {
1330 		/* for legacy asic path which doesn't has error source info */
1331 		obj->err_data.ue_count += err_data->ue_count;
1332 		obj->err_data.ce_count += err_data->ce_count;
1333 		obj->err_data.de_count += err_data->de_count;
1334 	}
1335 }
1336 
1337 static void amdgpu_ras_mgr_virt_error_data_statistics_update(struct ras_manager *obj,
1338 							     struct ras_err_data *err_data)
1339 {
1340 	/* Host reports absolute counts */
1341 	obj->err_data.ue_count = err_data->ue_count;
1342 	obj->err_data.ce_count = err_data->ce_count;
1343 	obj->err_data.de_count = err_data->de_count;
1344 }
1345 
1346 static struct ras_manager *get_ras_manager(struct amdgpu_device *adev, enum amdgpu_ras_block blk)
1347 {
1348 	struct ras_common_if head;
1349 
1350 	memset(&head, 0, sizeof(head));
1351 	head.block = blk;
1352 
1353 	return amdgpu_ras_find_obj(adev, &head);
1354 }
1355 
1356 int amdgpu_ras_bind_aca(struct amdgpu_device *adev, enum amdgpu_ras_block blk,
1357 			const struct aca_info *aca_info, void *data)
1358 {
1359 	struct ras_manager *obj;
1360 
1361 	/* in resume phase, no need to create aca fs node */
1362 	if (adev->in_suspend || amdgpu_reset_in_recovery(adev))
1363 		return 0;
1364 
1365 	obj = get_ras_manager(adev, blk);
1366 	if (!obj)
1367 		return -EINVAL;
1368 
1369 	return amdgpu_aca_add_handle(adev, &obj->aca_handle, ras_block_str(blk), aca_info, data);
1370 }
1371 
1372 int amdgpu_ras_unbind_aca(struct amdgpu_device *adev, enum amdgpu_ras_block blk)
1373 {
1374 	struct ras_manager *obj;
1375 
1376 	obj = get_ras_manager(adev, blk);
1377 	if (!obj)
1378 		return -EINVAL;
1379 
1380 	amdgpu_aca_remove_handle(&obj->aca_handle);
1381 
1382 	return 0;
1383 }
1384 
1385 static int amdgpu_aca_log_ras_error_data(struct amdgpu_device *adev, enum amdgpu_ras_block blk,
1386 					 enum aca_error_type type, struct ras_err_data *err_data,
1387 					 struct ras_query_context *qctx)
1388 {
1389 	struct ras_manager *obj;
1390 
1391 	obj = get_ras_manager(adev, blk);
1392 	if (!obj)
1393 		return -EINVAL;
1394 
1395 	return amdgpu_aca_get_error_data(adev, &obj->aca_handle, type, err_data, qctx);
1396 }
1397 
1398 ssize_t amdgpu_ras_aca_sysfs_read(struct device *dev, struct device_attribute *attr,
1399 				  struct aca_handle *handle, char *buf, void *data)
1400 {
1401 	struct ras_manager *obj = container_of(handle, struct ras_manager, aca_handle);
1402 	struct ras_query_if info = {
1403 		.head = obj->head,
1404 	};
1405 
1406 	if (!amdgpu_ras_get_error_query_ready(obj->adev))
1407 		return sysfs_emit(buf, "Query currently inaccessible\n");
1408 
1409 	if (amdgpu_ras_query_error_status(obj->adev, &info))
1410 		return -EINVAL;
1411 
1412 	return sysfs_emit(buf, "%s: %lu\n%s: %lu\n%s: %lu\n", "ue", info.ue_count,
1413 			  "ce", info.ce_count, "de", info.de_count);
1414 }
1415 
1416 static int amdgpu_ras_query_error_status_helper(struct amdgpu_device *adev,
1417 						struct ras_query_if *info,
1418 						struct ras_err_data *err_data,
1419 						struct ras_query_context *qctx,
1420 						unsigned int error_query_mode)
1421 {
1422 	enum amdgpu_ras_block blk = info ? info->head.block : AMDGPU_RAS_BLOCK_COUNT;
1423 	struct amdgpu_ras_block_object *block_obj = NULL;
1424 	int ret;
1425 
1426 	if (blk == AMDGPU_RAS_BLOCK_COUNT)
1427 		return -EINVAL;
1428 
1429 	if (error_query_mode == AMDGPU_RAS_INVALID_ERROR_QUERY)
1430 		return -EINVAL;
1431 
1432 	if (error_query_mode == AMDGPU_RAS_VIRT_ERROR_COUNT_QUERY) {
1433 		return amdgpu_virt_req_ras_err_count(adev, blk, err_data);
1434 	} else if (error_query_mode == AMDGPU_RAS_DIRECT_ERROR_QUERY) {
1435 		if (info->head.block == AMDGPU_RAS_BLOCK__UMC) {
1436 			amdgpu_ras_get_ecc_info(adev, err_data);
1437 		} else {
1438 			block_obj = amdgpu_ras_get_ras_block(adev, info->head.block, 0);
1439 			if (!block_obj || !block_obj->hw_ops) {
1440 				dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1441 					     get_ras_block_str(&info->head));
1442 				return -EINVAL;
1443 			}
1444 
1445 			if (block_obj->hw_ops->query_ras_error_count)
1446 				block_obj->hw_ops->query_ras_error_count(adev, err_data);
1447 
1448 			if ((info->head.block == AMDGPU_RAS_BLOCK__SDMA) ||
1449 			    (info->head.block == AMDGPU_RAS_BLOCK__GFX) ||
1450 			    (info->head.block == AMDGPU_RAS_BLOCK__MMHUB)) {
1451 				if (block_obj->hw_ops->query_ras_error_status)
1452 					block_obj->hw_ops->query_ras_error_status(adev);
1453 			}
1454 		}
1455 	} else {
1456 		if (amdgpu_aca_is_enabled(adev)) {
1457 			ret = amdgpu_aca_log_ras_error_data(adev, blk, ACA_ERROR_TYPE_UE, err_data, qctx);
1458 			if (ret)
1459 				return ret;
1460 
1461 			ret = amdgpu_aca_log_ras_error_data(adev, blk, ACA_ERROR_TYPE_CE, err_data, qctx);
1462 			if (ret)
1463 				return ret;
1464 
1465 			ret = amdgpu_aca_log_ras_error_data(adev, blk, ACA_ERROR_TYPE_DEFERRED, err_data, qctx);
1466 			if (ret)
1467 				return ret;
1468 		} else {
1469 			/* FIXME: add code to check return value later */
1470 			amdgpu_mca_smu_log_ras_error(adev, blk, AMDGPU_MCA_ERROR_TYPE_UE, err_data, qctx);
1471 			amdgpu_mca_smu_log_ras_error(adev, blk, AMDGPU_MCA_ERROR_TYPE_CE, err_data, qctx);
1472 		}
1473 	}
1474 
1475 	return 0;
1476 }
1477 
1478 /* query/inject/cure begin */
1479 static int amdgpu_ras_query_error_status_with_event(struct amdgpu_device *adev,
1480 						    struct ras_query_if *info,
1481 						    enum ras_event_type type)
1482 {
1483 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1484 	struct ras_err_data err_data;
1485 	struct ras_query_context qctx;
1486 	unsigned int error_query_mode;
1487 	int ret;
1488 
1489 	if (!obj)
1490 		return -EINVAL;
1491 
1492 	ret = amdgpu_ras_error_data_init(&err_data);
1493 	if (ret)
1494 		return ret;
1495 
1496 	if (!amdgpu_ras_get_error_query_mode(adev, &error_query_mode))
1497 		return -EINVAL;
1498 
1499 	memset(&qctx, 0, sizeof(qctx));
1500 	qctx.evid.type = type;
1501 	qctx.evid.event_id = amdgpu_ras_acquire_event_id(adev, type);
1502 
1503 	if (!down_read_trylock(&adev->reset_domain->sem)) {
1504 		ret = -EIO;
1505 		goto out_fini_err_data;
1506 	}
1507 
1508 	ret = amdgpu_ras_query_error_status_helper(adev, info,
1509 						   &err_data,
1510 						   &qctx,
1511 						   error_query_mode);
1512 	up_read(&adev->reset_domain->sem);
1513 	if (ret)
1514 		goto out_fini_err_data;
1515 
1516 	if (error_query_mode != AMDGPU_RAS_VIRT_ERROR_COUNT_QUERY) {
1517 		amdgpu_rasmgr_error_data_statistic_update(obj, &err_data);
1518 		amdgpu_ras_error_generate_report(adev, info, &err_data, &qctx);
1519 	} else {
1520 		/* Host provides absolute error counts. First generate the report
1521 		 * using the previous VF internal count against new host count.
1522 		 * Then Update VF internal count.
1523 		 */
1524 		amdgpu_ras_virt_error_generate_report(adev, info, &err_data, &qctx);
1525 		amdgpu_ras_mgr_virt_error_data_statistics_update(obj, &err_data);
1526 	}
1527 
1528 	info->ue_count = obj->err_data.ue_count;
1529 	info->ce_count = obj->err_data.ce_count;
1530 	info->de_count = obj->err_data.de_count;
1531 
1532 out_fini_err_data:
1533 	amdgpu_ras_error_data_fini(&err_data);
1534 
1535 	return ret;
1536 }
1537 
1538 int amdgpu_ras_query_error_status(struct amdgpu_device *adev, struct ras_query_if *info)
1539 {
1540 	return amdgpu_ras_query_error_status_with_event(adev, info, RAS_EVENT_TYPE_INVALID);
1541 }
1542 
1543 int amdgpu_ras_reset_error_count(struct amdgpu_device *adev,
1544 		enum amdgpu_ras_block block)
1545 {
1546 	struct amdgpu_ras_block_object *block_obj = amdgpu_ras_get_ras_block(adev, block, 0);
1547 	const struct amdgpu_mca_smu_funcs *mca_funcs = adev->mca.mca_funcs;
1548 	const struct aca_smu_funcs *smu_funcs = adev->aca.smu_funcs;
1549 
1550 	if (!block_obj || !block_obj->hw_ops) {
1551 		dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1552 				ras_block_str(block));
1553 		return -EOPNOTSUPP;
1554 	}
1555 
1556 	if (!amdgpu_ras_is_supported(adev, block) ||
1557 	    !amdgpu_ras_get_aca_debug_mode(adev))
1558 		return -EOPNOTSUPP;
1559 
1560 	if (amdgpu_sriov_vf(adev))
1561 		return -EOPNOTSUPP;
1562 
1563 	/* skip ras error reset in gpu reset */
1564 	if ((amdgpu_in_reset(adev) || amdgpu_ras_in_recovery(adev)) &&
1565 	    ((smu_funcs && smu_funcs->set_debug_mode) ||
1566 	     (mca_funcs && mca_funcs->mca_set_debug_mode)))
1567 		return -EOPNOTSUPP;
1568 
1569 	if (block_obj->hw_ops->reset_ras_error_count)
1570 		block_obj->hw_ops->reset_ras_error_count(adev);
1571 
1572 	return 0;
1573 }
1574 
1575 int amdgpu_ras_reset_error_status(struct amdgpu_device *adev,
1576 		enum amdgpu_ras_block block)
1577 {
1578 	struct amdgpu_ras_block_object *block_obj = amdgpu_ras_get_ras_block(adev, block, 0);
1579 
1580 	if (amdgpu_ras_reset_error_count(adev, block) == -EOPNOTSUPP)
1581 		return 0;
1582 
1583 	if ((block == AMDGPU_RAS_BLOCK__GFX) ||
1584 	    (block == AMDGPU_RAS_BLOCK__MMHUB)) {
1585 		if (block_obj->hw_ops->reset_ras_error_status)
1586 			block_obj->hw_ops->reset_ras_error_status(adev);
1587 	}
1588 
1589 	return 0;
1590 }
1591 
1592 /* wrapper of psp_ras_trigger_error */
1593 int amdgpu_ras_error_inject(struct amdgpu_device *adev,
1594 		struct ras_inject_if *info)
1595 {
1596 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1597 	struct ta_ras_trigger_error_input block_info = {
1598 		.block_id =  amdgpu_ras_block_to_ta(info->head.block),
1599 		.inject_error_type = amdgpu_ras_error_to_ta(info->head.type),
1600 		.sub_block_index = info->head.sub_block_index,
1601 		.address = info->address,
1602 		.value = info->value,
1603 	};
1604 	int ret = -EINVAL;
1605 	struct amdgpu_ras_block_object *block_obj = amdgpu_ras_get_ras_block(adev,
1606 							info->head.block,
1607 							info->head.sub_block_index);
1608 
1609 	/* inject on guest isn't allowed, return success directly */
1610 	if (amdgpu_sriov_vf(adev))
1611 		return 0;
1612 
1613 	if (!obj)
1614 		return -EINVAL;
1615 
1616 	if (!block_obj || !block_obj->hw_ops)	{
1617 		dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1618 			     get_ras_block_str(&info->head));
1619 		return -EINVAL;
1620 	}
1621 
1622 	/* Calculate XGMI relative offset */
1623 	if (adev->gmc.xgmi.num_physical_nodes > 1 &&
1624 	    info->head.block != AMDGPU_RAS_BLOCK__GFX) {
1625 		block_info.address =
1626 			amdgpu_xgmi_get_relative_phy_addr(adev,
1627 							  block_info.address);
1628 	}
1629 
1630 	if (block_obj->hw_ops->ras_error_inject) {
1631 		if (info->head.block == AMDGPU_RAS_BLOCK__GFX)
1632 			ret = block_obj->hw_ops->ras_error_inject(adev, info, info->instance_mask);
1633 		else /* Special ras_error_inject is defined (e.g: xgmi) */
1634 			ret = block_obj->hw_ops->ras_error_inject(adev, &block_info,
1635 						info->instance_mask);
1636 	} else {
1637 		/* default path */
1638 		ret = psp_ras_trigger_error(&adev->psp, &block_info, info->instance_mask);
1639 	}
1640 
1641 	if (ret)
1642 		dev_err(adev->dev, "ras inject %s failed %d\n",
1643 			get_ras_block_str(&info->head), ret);
1644 
1645 	return ret;
1646 }
1647 
1648 /**
1649  * amdgpu_ras_query_error_count_helper -- Get error counter for specific IP
1650  * @adev: pointer to AMD GPU device
1651  * @ce_count: pointer to an integer to be set to the count of correctible errors.
1652  * @ue_count: pointer to an integer to be set to the count of uncorrectible errors.
1653  * @query_info: pointer to ras_query_if
1654  *
1655  * Return 0 for query success or do nothing, otherwise return an error
1656  * on failures
1657  */
1658 static int amdgpu_ras_query_error_count_helper(struct amdgpu_device *adev,
1659 					       unsigned long *ce_count,
1660 					       unsigned long *ue_count,
1661 					       struct ras_query_if *query_info)
1662 {
1663 	int ret;
1664 
1665 	if (!query_info)
1666 		/* do nothing if query_info is not specified */
1667 		return 0;
1668 
1669 	ret = amdgpu_ras_query_error_status(adev, query_info);
1670 	if (ret)
1671 		return ret;
1672 
1673 	*ce_count += query_info->ce_count;
1674 	*ue_count += query_info->ue_count;
1675 
1676 	/* some hardware/IP supports read to clear
1677 	 * no need to explictly reset the err status after the query call */
1678 	if (amdgpu_ip_version(adev, MP0_HWIP, 0) != IP_VERSION(11, 0, 2) &&
1679 	    amdgpu_ip_version(adev, MP0_HWIP, 0) != IP_VERSION(11, 0, 4)) {
1680 		if (amdgpu_ras_reset_error_status(adev, query_info->head.block))
1681 			dev_warn(adev->dev,
1682 				 "Failed to reset error counter and error status\n");
1683 	}
1684 
1685 	return 0;
1686 }
1687 
1688 /**
1689  * amdgpu_ras_query_error_count -- Get error counts of all IPs or specific IP
1690  * @adev: pointer to AMD GPU device
1691  * @ce_count: pointer to an integer to be set to the count of correctible errors.
1692  * @ue_count: pointer to an integer to be set to the count of uncorrectible
1693  * errors.
1694  * @query_info: pointer to ras_query_if if the query request is only for
1695  * specific ip block; if info is NULL, then the qurey request is for
1696  * all the ip blocks that support query ras error counters/status
1697  *
1698  * If set, @ce_count or @ue_count, count and return the corresponding
1699  * error counts in those integer pointers. Return 0 if the device
1700  * supports RAS. Return -EOPNOTSUPP if the device doesn't support RAS.
1701  */
1702 int amdgpu_ras_query_error_count(struct amdgpu_device *adev,
1703 				 unsigned long *ce_count,
1704 				 unsigned long *ue_count,
1705 				 struct ras_query_if *query_info)
1706 {
1707 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1708 	struct ras_manager *obj;
1709 	unsigned long ce, ue;
1710 	int ret;
1711 
1712 	if (!adev->ras_enabled || !con)
1713 		return -EOPNOTSUPP;
1714 
1715 	/* Don't count since no reporting.
1716 	 */
1717 	if (!ce_count && !ue_count)
1718 		return 0;
1719 
1720 	ce = 0;
1721 	ue = 0;
1722 	if (!query_info) {
1723 		/* query all the ip blocks that support ras query interface */
1724 		list_for_each_entry(obj, &con->head, node) {
1725 			struct ras_query_if info = {
1726 				.head = obj->head,
1727 			};
1728 
1729 			ret = amdgpu_ras_query_error_count_helper(adev, &ce, &ue, &info);
1730 		}
1731 	} else {
1732 		/* query specific ip block */
1733 		ret = amdgpu_ras_query_error_count_helper(adev, &ce, &ue, query_info);
1734 	}
1735 
1736 	if (ret)
1737 		return ret;
1738 
1739 	if (ce_count)
1740 		*ce_count = ce;
1741 
1742 	if (ue_count)
1743 		*ue_count = ue;
1744 
1745 	return 0;
1746 }
1747 /* query/inject/cure end */
1748 
1749 
1750 /* sysfs begin */
1751 
1752 static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
1753 		struct ras_badpage **bps, unsigned int *count);
1754 
1755 static char *amdgpu_ras_badpage_flags_str(unsigned int flags)
1756 {
1757 	switch (flags) {
1758 	case AMDGPU_RAS_RETIRE_PAGE_RESERVED:
1759 		return "R";
1760 	case AMDGPU_RAS_RETIRE_PAGE_PENDING:
1761 		return "P";
1762 	case AMDGPU_RAS_RETIRE_PAGE_FAULT:
1763 	default:
1764 		return "F";
1765 	}
1766 }
1767 
1768 /**
1769  * DOC: AMDGPU RAS sysfs gpu_vram_bad_pages Interface
1770  *
1771  * It allows user to read the bad pages of vram on the gpu through
1772  * /sys/class/drm/card[0/1/2...]/device/ras/gpu_vram_bad_pages
1773  *
1774  * It outputs multiple lines, and each line stands for one gpu page.
1775  *
1776  * The format of one line is below,
1777  * gpu pfn : gpu page size : flags
1778  *
1779  * gpu pfn and gpu page size are printed in hex format.
1780  * flags can be one of below character,
1781  *
1782  * R: reserved, this gpu page is reserved and not able to use.
1783  *
1784  * P: pending for reserve, this gpu page is marked as bad, will be reserved
1785  * in next window of page_reserve.
1786  *
1787  * F: unable to reserve. this gpu page can't be reserved due to some reasons.
1788  *
1789  * Examples:
1790  *
1791  * .. code-block:: bash
1792  *
1793  *	0x00000001 : 0x00001000 : R
1794  *	0x00000002 : 0x00001000 : P
1795  *
1796  */
1797 
1798 static ssize_t amdgpu_ras_sysfs_badpages_read(struct file *f,
1799 		struct kobject *kobj, const struct bin_attribute *attr,
1800 		char *buf, loff_t ppos, size_t count)
1801 {
1802 	struct amdgpu_ras *con =
1803 		container_of(attr, struct amdgpu_ras, badpages_attr);
1804 	struct amdgpu_device *adev = con->adev;
1805 	const unsigned int element_size =
1806 		sizeof("0xabcdabcd : 0x12345678 : R\n") - 1;
1807 	unsigned int start = div64_ul(ppos + element_size - 1, element_size);
1808 	unsigned int end = div64_ul(ppos + count - 1, element_size);
1809 	ssize_t s = 0;
1810 	struct ras_badpage *bps = NULL;
1811 	unsigned int bps_count = 0;
1812 
1813 	memset(buf, 0, count);
1814 
1815 	if (amdgpu_ras_badpages_read(adev, &bps, &bps_count))
1816 		return 0;
1817 
1818 	for (; start < end && start < bps_count; start++)
1819 		s += scnprintf(&buf[s], element_size + 1,
1820 				"0x%08x : 0x%08x : %1s\n",
1821 				bps[start].bp,
1822 				bps[start].size,
1823 				amdgpu_ras_badpage_flags_str(bps[start].flags));
1824 
1825 	kfree(bps);
1826 
1827 	return s;
1828 }
1829 
1830 static ssize_t amdgpu_ras_sysfs_features_read(struct device *dev,
1831 		struct device_attribute *attr, char *buf)
1832 {
1833 	struct amdgpu_ras *con =
1834 		container_of(attr, struct amdgpu_ras, features_attr);
1835 
1836 	return sysfs_emit(buf, "feature mask: 0x%x\n", con->features);
1837 }
1838 
1839 static ssize_t amdgpu_ras_sysfs_version_show(struct device *dev,
1840 		struct device_attribute *attr, char *buf)
1841 {
1842 	struct amdgpu_ras *con =
1843 		container_of(attr, struct amdgpu_ras, version_attr);
1844 	return sysfs_emit(buf, "table version: 0x%x\n", con->eeprom_control.tbl_hdr.version);
1845 }
1846 
1847 static ssize_t amdgpu_ras_sysfs_schema_show(struct device *dev,
1848 		struct device_attribute *attr, char *buf)
1849 {
1850 	struct amdgpu_ras *con =
1851 		container_of(attr, struct amdgpu_ras, schema_attr);
1852 	return sysfs_emit(buf, "schema: 0x%x\n", con->schema);
1853 }
1854 
1855 static struct {
1856 	enum ras_event_type type;
1857 	const char *name;
1858 } dump_event[] = {
1859 	{RAS_EVENT_TYPE_FATAL, "Fatal Error"},
1860 	{RAS_EVENT_TYPE_POISON_CREATION, "Poison Creation"},
1861 	{RAS_EVENT_TYPE_POISON_CONSUMPTION, "Poison Consumption"},
1862 };
1863 
1864 static ssize_t amdgpu_ras_sysfs_event_state_show(struct device *dev,
1865 						 struct device_attribute *attr, char *buf)
1866 {
1867 	struct amdgpu_ras *con =
1868 		container_of(attr, struct amdgpu_ras, event_state_attr);
1869 	struct ras_event_manager *event_mgr = con->event_mgr;
1870 	struct ras_event_state *event_state;
1871 	int i, size = 0;
1872 
1873 	if (!event_mgr)
1874 		return -EINVAL;
1875 
1876 	size += sysfs_emit_at(buf, size, "current seqno: %llu\n", atomic64_read(&event_mgr->seqno));
1877 	for (i = 0; i < ARRAY_SIZE(dump_event); i++) {
1878 		event_state = &event_mgr->event_state[dump_event[i].type];
1879 		size += sysfs_emit_at(buf, size, "%s: count:%llu, last_seqno:%llu\n",
1880 				      dump_event[i].name,
1881 				      atomic64_read(&event_state->count),
1882 				      event_state->last_seqno);
1883 	}
1884 
1885 	return (ssize_t)size;
1886 }
1887 
1888 static void amdgpu_ras_sysfs_remove_bad_page_node(struct amdgpu_device *adev)
1889 {
1890 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1891 
1892 	if (adev->dev->kobj.sd)
1893 		sysfs_remove_file_from_group(&adev->dev->kobj,
1894 				&con->badpages_attr.attr,
1895 				RAS_FS_NAME);
1896 }
1897 
1898 static int amdgpu_ras_sysfs_remove_dev_attr_node(struct amdgpu_device *adev)
1899 {
1900 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1901 	struct attribute *attrs[] = {
1902 		&con->features_attr.attr,
1903 		&con->version_attr.attr,
1904 		&con->schema_attr.attr,
1905 		&con->event_state_attr.attr,
1906 		NULL
1907 	};
1908 	struct attribute_group group = {
1909 		.name = RAS_FS_NAME,
1910 		.attrs = attrs,
1911 	};
1912 
1913 	if (adev->dev->kobj.sd)
1914 		sysfs_remove_group(&adev->dev->kobj, &group);
1915 
1916 	return 0;
1917 }
1918 
1919 int amdgpu_ras_sysfs_create(struct amdgpu_device *adev,
1920 		struct ras_common_if *head)
1921 {
1922 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1923 
1924 	if (amdgpu_aca_is_enabled(adev))
1925 		return 0;
1926 
1927 	if (!obj || obj->attr_inuse)
1928 		return -EINVAL;
1929 
1930 	if (amdgpu_sriov_vf(adev) && !amdgpu_virt_ras_telemetry_block_en(adev, head->block))
1931 		return 0;
1932 
1933 	get_obj(obj);
1934 
1935 	snprintf(obj->fs_data.sysfs_name, sizeof(obj->fs_data.sysfs_name),
1936 		"%s_err_count", head->name);
1937 
1938 	obj->sysfs_attr = (struct device_attribute){
1939 		.attr = {
1940 			.name = obj->fs_data.sysfs_name,
1941 			.mode = S_IRUGO,
1942 		},
1943 			.show = amdgpu_ras_sysfs_read,
1944 	};
1945 	sysfs_attr_init(&obj->sysfs_attr.attr);
1946 
1947 	if (sysfs_add_file_to_group(&adev->dev->kobj,
1948 				&obj->sysfs_attr.attr,
1949 				RAS_FS_NAME)) {
1950 		put_obj(obj);
1951 		return -EINVAL;
1952 	}
1953 
1954 	obj->attr_inuse = 1;
1955 
1956 	return 0;
1957 }
1958 
1959 int amdgpu_ras_sysfs_remove(struct amdgpu_device *adev,
1960 		struct ras_common_if *head)
1961 {
1962 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1963 
1964 	if (amdgpu_aca_is_enabled(adev))
1965 		return 0;
1966 
1967 	if (!obj || !obj->attr_inuse)
1968 		return -EINVAL;
1969 
1970 	if (adev->dev->kobj.sd)
1971 		sysfs_remove_file_from_group(&adev->dev->kobj,
1972 				&obj->sysfs_attr.attr,
1973 				RAS_FS_NAME);
1974 	obj->attr_inuse = 0;
1975 	put_obj(obj);
1976 
1977 	return 0;
1978 }
1979 
1980 static int amdgpu_ras_sysfs_remove_all(struct amdgpu_device *adev)
1981 {
1982 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1983 	struct ras_manager *obj, *tmp;
1984 
1985 	list_for_each_entry_safe(obj, tmp, &con->head, node) {
1986 		amdgpu_ras_sysfs_remove(adev, &obj->head);
1987 	}
1988 
1989 	if (amdgpu_bad_page_threshold != 0)
1990 		amdgpu_ras_sysfs_remove_bad_page_node(adev);
1991 
1992 	amdgpu_ras_sysfs_remove_dev_attr_node(adev);
1993 
1994 	return 0;
1995 }
1996 /* sysfs end */
1997 
1998 /**
1999  * DOC: AMDGPU RAS Reboot Behavior for Unrecoverable Errors
2000  *
2001  * Normally when there is an uncorrectable error, the driver will reset
2002  * the GPU to recover.  However, in the event of an unrecoverable error,
2003  * the driver provides an interface to reboot the system automatically
2004  * in that event.
2005  *
2006  * The following file in debugfs provides that interface:
2007  * /sys/kernel/debug/dri/[0/1/2...]/ras/auto_reboot
2008  *
2009  * Usage:
2010  *
2011  * .. code-block:: bash
2012  *
2013  *	echo true > .../ras/auto_reboot
2014  *
2015  */
2016 /* debugfs begin */
2017 static struct dentry *amdgpu_ras_debugfs_create_ctrl_node(struct amdgpu_device *adev)
2018 {
2019 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2020 	struct amdgpu_ras_eeprom_control *eeprom = &con->eeprom_control;
2021 	struct drm_minor  *minor = adev_to_drm(adev)->primary;
2022 	struct dentry     *dir;
2023 
2024 	dir = debugfs_create_dir(RAS_FS_NAME, minor->debugfs_root);
2025 	debugfs_create_file("ras_ctrl", S_IWUGO | S_IRUGO, dir, adev,
2026 			    &amdgpu_ras_debugfs_ctrl_ops);
2027 	debugfs_create_file("ras_eeprom_reset", S_IWUGO | S_IRUGO, dir, adev,
2028 			    &amdgpu_ras_debugfs_eeprom_ops);
2029 	debugfs_create_u32("bad_page_cnt_threshold", 0444, dir,
2030 			   &con->bad_page_cnt_threshold);
2031 	debugfs_create_u32("ras_num_recs", 0444, dir, &eeprom->ras_num_recs);
2032 	debugfs_create_x32("ras_hw_enabled", 0444, dir, &adev->ras_hw_enabled);
2033 	debugfs_create_x32("ras_enabled", 0444, dir, &adev->ras_enabled);
2034 	debugfs_create_file("ras_eeprom_size", S_IRUGO, dir, adev,
2035 			    &amdgpu_ras_debugfs_eeprom_size_ops);
2036 	con->de_ras_eeprom_table = debugfs_create_file("ras_eeprom_table",
2037 						       S_IRUGO, dir, adev,
2038 						       &amdgpu_ras_debugfs_eeprom_table_ops);
2039 	amdgpu_ras_debugfs_set_ret_size(&con->eeprom_control);
2040 
2041 	/*
2042 	 * After one uncorrectable error happens, usually GPU recovery will
2043 	 * be scheduled. But due to the known problem in GPU recovery failing
2044 	 * to bring GPU back, below interface provides one direct way to
2045 	 * user to reboot system automatically in such case within
2046 	 * ERREVENT_ATHUB_INTERRUPT generated. Normal GPU recovery routine
2047 	 * will never be called.
2048 	 */
2049 	debugfs_create_bool("auto_reboot", S_IWUGO | S_IRUGO, dir, &con->reboot);
2050 
2051 	/*
2052 	 * User could set this not to clean up hardware's error count register
2053 	 * of RAS IPs during ras recovery.
2054 	 */
2055 	debugfs_create_bool("disable_ras_err_cnt_harvest", 0644, dir,
2056 			    &con->disable_ras_err_cnt_harvest);
2057 	return dir;
2058 }
2059 
2060 static void amdgpu_ras_debugfs_create(struct amdgpu_device *adev,
2061 				      struct ras_fs_if *head,
2062 				      struct dentry *dir)
2063 {
2064 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head->head);
2065 
2066 	if (!obj || !dir)
2067 		return;
2068 
2069 	get_obj(obj);
2070 
2071 	memcpy(obj->fs_data.debugfs_name,
2072 			head->debugfs_name,
2073 			sizeof(obj->fs_data.debugfs_name));
2074 
2075 	debugfs_create_file(obj->fs_data.debugfs_name, S_IWUGO | S_IRUGO, dir,
2076 			    obj, &amdgpu_ras_debugfs_ops);
2077 }
2078 
2079 static bool amdgpu_ras_aca_is_supported(struct amdgpu_device *adev)
2080 {
2081 	bool ret;
2082 
2083 	switch (amdgpu_ip_version(adev, MP0_HWIP, 0)) {
2084 	case IP_VERSION(13, 0, 6):
2085 	case IP_VERSION(13, 0, 12):
2086 	case IP_VERSION(13, 0, 14):
2087 		ret = true;
2088 		break;
2089 	default:
2090 		ret = false;
2091 		break;
2092 	}
2093 
2094 	return ret;
2095 }
2096 
2097 void amdgpu_ras_debugfs_create_all(struct amdgpu_device *adev)
2098 {
2099 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2100 	struct dentry *dir;
2101 	struct ras_manager *obj;
2102 	struct ras_fs_if fs_info;
2103 
2104 	/*
2105 	 * it won't be called in resume path, no need to check
2106 	 * suspend and gpu reset status
2107 	 */
2108 	if (!IS_ENABLED(CONFIG_DEBUG_FS) || !con)
2109 		return;
2110 
2111 	dir = amdgpu_ras_debugfs_create_ctrl_node(adev);
2112 
2113 	list_for_each_entry(obj, &con->head, node) {
2114 		if (amdgpu_ras_is_supported(adev, obj->head.block) &&
2115 			(obj->attr_inuse == 1)) {
2116 			sprintf(fs_info.debugfs_name, "%s_err_inject",
2117 					get_ras_block_str(&obj->head));
2118 			fs_info.head = obj->head;
2119 			amdgpu_ras_debugfs_create(adev, &fs_info, dir);
2120 		}
2121 	}
2122 
2123 	if (amdgpu_ras_aca_is_supported(adev)) {
2124 		if (amdgpu_aca_is_enabled(adev))
2125 			amdgpu_aca_smu_debugfs_init(adev, dir);
2126 		else
2127 			amdgpu_mca_smu_debugfs_init(adev, dir);
2128 	}
2129 }
2130 
2131 /* debugfs end */
2132 
2133 /* ras fs */
2134 static const BIN_ATTR(gpu_vram_bad_pages, S_IRUGO,
2135 		      amdgpu_ras_sysfs_badpages_read, NULL, 0);
2136 static DEVICE_ATTR(features, S_IRUGO,
2137 		amdgpu_ras_sysfs_features_read, NULL);
2138 static DEVICE_ATTR(version, 0444,
2139 		amdgpu_ras_sysfs_version_show, NULL);
2140 static DEVICE_ATTR(schema, 0444,
2141 		amdgpu_ras_sysfs_schema_show, NULL);
2142 static DEVICE_ATTR(event_state, 0444,
2143 		   amdgpu_ras_sysfs_event_state_show, NULL);
2144 static int amdgpu_ras_fs_init(struct amdgpu_device *adev)
2145 {
2146 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2147 	struct attribute_group group = {
2148 		.name = RAS_FS_NAME,
2149 	};
2150 	struct attribute *attrs[] = {
2151 		&con->features_attr.attr,
2152 		&con->version_attr.attr,
2153 		&con->schema_attr.attr,
2154 		&con->event_state_attr.attr,
2155 		NULL
2156 	};
2157 	const struct bin_attribute *bin_attrs[] = {
2158 		NULL,
2159 		NULL,
2160 	};
2161 	int r;
2162 
2163 	group.attrs = attrs;
2164 
2165 	/* add features entry */
2166 	con->features_attr = dev_attr_features;
2167 	sysfs_attr_init(attrs[0]);
2168 
2169 	/* add version entry */
2170 	con->version_attr = dev_attr_version;
2171 	sysfs_attr_init(attrs[1]);
2172 
2173 	/* add schema entry */
2174 	con->schema_attr = dev_attr_schema;
2175 	sysfs_attr_init(attrs[2]);
2176 
2177 	/* add event_state entry */
2178 	con->event_state_attr = dev_attr_event_state;
2179 	sysfs_attr_init(attrs[3]);
2180 
2181 	if (amdgpu_bad_page_threshold != 0) {
2182 		/* add bad_page_features entry */
2183 		con->badpages_attr = bin_attr_gpu_vram_bad_pages;
2184 		sysfs_bin_attr_init(&con->badpages_attr);
2185 		bin_attrs[0] = &con->badpages_attr;
2186 		group.bin_attrs_new = bin_attrs;
2187 	}
2188 
2189 	r = sysfs_create_group(&adev->dev->kobj, &group);
2190 	if (r)
2191 		dev_err(adev->dev, "Failed to create RAS sysfs group!");
2192 
2193 	return 0;
2194 }
2195 
2196 static int amdgpu_ras_fs_fini(struct amdgpu_device *adev)
2197 {
2198 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2199 	struct ras_manager *con_obj, *ip_obj, *tmp;
2200 
2201 	if (IS_ENABLED(CONFIG_DEBUG_FS)) {
2202 		list_for_each_entry_safe(con_obj, tmp, &con->head, node) {
2203 			ip_obj = amdgpu_ras_find_obj(adev, &con_obj->head);
2204 			if (ip_obj)
2205 				put_obj(ip_obj);
2206 		}
2207 	}
2208 
2209 	amdgpu_ras_sysfs_remove_all(adev);
2210 	return 0;
2211 }
2212 /* ras fs end */
2213 
2214 /* ih begin */
2215 
2216 /* For the hardware that cannot enable bif ring for both ras_controller_irq
2217  * and ras_err_evnet_athub_irq ih cookies, the driver has to poll status
2218  * register to check whether the interrupt is triggered or not, and properly
2219  * ack the interrupt if it is there
2220  */
2221 void amdgpu_ras_interrupt_fatal_error_handler(struct amdgpu_device *adev)
2222 {
2223 	/* Fatal error events are handled on host side */
2224 	if (amdgpu_sriov_vf(adev))
2225 		return;
2226 	/*
2227 	 * If the current interrupt is caused by a non-fatal RAS error, skip
2228 	 * check for fatal error. For fatal errors, FED status of all devices
2229 	 * in XGMI hive gets set when the first device gets fatal error
2230 	 * interrupt. The error gets propagated to other devices as well, so
2231 	 * make sure to ack the interrupt regardless of FED status.
2232 	 */
2233 	if (!amdgpu_ras_get_fed_status(adev) &&
2234 	    amdgpu_ras_is_err_state(adev, AMDGPU_RAS_BLOCK__ANY))
2235 		return;
2236 
2237 	if (adev->nbio.ras &&
2238 	    adev->nbio.ras->handle_ras_controller_intr_no_bifring)
2239 		adev->nbio.ras->handle_ras_controller_intr_no_bifring(adev);
2240 
2241 	if (adev->nbio.ras &&
2242 	    adev->nbio.ras->handle_ras_err_event_athub_intr_no_bifring)
2243 		adev->nbio.ras->handle_ras_err_event_athub_intr_no_bifring(adev);
2244 }
2245 
2246 static void amdgpu_ras_interrupt_poison_consumption_handler(struct ras_manager *obj,
2247 				struct amdgpu_iv_entry *entry)
2248 {
2249 	bool poison_stat = false;
2250 	struct amdgpu_device *adev = obj->adev;
2251 	struct amdgpu_ras_block_object *block_obj =
2252 		amdgpu_ras_get_ras_block(adev, obj->head.block, 0);
2253 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2254 	enum ras_event_type type = RAS_EVENT_TYPE_POISON_CONSUMPTION;
2255 	u64 event_id;
2256 	int ret;
2257 
2258 	if (!block_obj || !con)
2259 		return;
2260 
2261 	ret = amdgpu_ras_mark_ras_event(adev, type);
2262 	if (ret)
2263 		return;
2264 
2265 	amdgpu_ras_set_err_poison(adev, block_obj->ras_comm.block);
2266 	/* both query_poison_status and handle_poison_consumption are optional,
2267 	 * but at least one of them should be implemented if we need poison
2268 	 * consumption handler
2269 	 */
2270 	if (block_obj->hw_ops && block_obj->hw_ops->query_poison_status) {
2271 		poison_stat = block_obj->hw_ops->query_poison_status(adev);
2272 		if (!poison_stat) {
2273 			/* Not poison consumption interrupt, no need to handle it */
2274 			dev_info(adev->dev, "No RAS poison status in %s poison IH.\n",
2275 					block_obj->ras_comm.name);
2276 
2277 			return;
2278 		}
2279 	}
2280 
2281 	amdgpu_umc_poison_handler(adev, obj->head.block, 0);
2282 
2283 	if (block_obj->hw_ops && block_obj->hw_ops->handle_poison_consumption)
2284 		poison_stat = block_obj->hw_ops->handle_poison_consumption(adev);
2285 
2286 	/* gpu reset is fallback for failed and default cases.
2287 	 * For RMA case, amdgpu_umc_poison_handler will handle gpu reset.
2288 	 */
2289 	if (poison_stat && !amdgpu_ras_is_rma(adev)) {
2290 		event_id = amdgpu_ras_acquire_event_id(adev, type);
2291 		RAS_EVENT_LOG(adev, event_id,
2292 			      "GPU reset for %s RAS poison consumption is issued!\n",
2293 			      block_obj->ras_comm.name);
2294 		amdgpu_ras_reset_gpu(adev);
2295 	}
2296 
2297 	if (!poison_stat)
2298 		amdgpu_gfx_poison_consumption_handler(adev, entry);
2299 }
2300 
2301 static void amdgpu_ras_interrupt_poison_creation_handler(struct ras_manager *obj,
2302 				struct amdgpu_iv_entry *entry)
2303 {
2304 	struct amdgpu_device *adev = obj->adev;
2305 	enum ras_event_type type = RAS_EVENT_TYPE_POISON_CREATION;
2306 	u64 event_id;
2307 	int ret;
2308 
2309 	ret = amdgpu_ras_mark_ras_event(adev, type);
2310 	if (ret)
2311 		return;
2312 
2313 	event_id = amdgpu_ras_acquire_event_id(adev, type);
2314 	RAS_EVENT_LOG(adev, event_id, "Poison is created\n");
2315 
2316 	if (amdgpu_ip_version(obj->adev, UMC_HWIP, 0) >= IP_VERSION(12, 0, 0)) {
2317 		struct amdgpu_ras *con = amdgpu_ras_get_context(obj->adev);
2318 
2319 		atomic_inc(&con->page_retirement_req_cnt);
2320 		atomic_inc(&con->poison_creation_count);
2321 
2322 		wake_up(&con->page_retirement_wq);
2323 	}
2324 }
2325 
2326 static void amdgpu_ras_interrupt_umc_handler(struct ras_manager *obj,
2327 				struct amdgpu_iv_entry *entry)
2328 {
2329 	struct ras_ih_data *data = &obj->ih_data;
2330 	struct ras_err_data err_data;
2331 	int ret;
2332 
2333 	if (!data->cb)
2334 		return;
2335 
2336 	ret = amdgpu_ras_error_data_init(&err_data);
2337 	if (ret)
2338 		return;
2339 
2340 	/* Let IP handle its data, maybe we need get the output
2341 	 * from the callback to update the error type/count, etc
2342 	 */
2343 	amdgpu_ras_set_fed(obj->adev, true);
2344 	ret = data->cb(obj->adev, &err_data, entry);
2345 	/* ue will trigger an interrupt, and in that case
2346 	 * we need do a reset to recovery the whole system.
2347 	 * But leave IP do that recovery, here we just dispatch
2348 	 * the error.
2349 	 */
2350 	if (ret == AMDGPU_RAS_SUCCESS) {
2351 		/* these counts could be left as 0 if
2352 		 * some blocks do not count error number
2353 		 */
2354 		obj->err_data.ue_count += err_data.ue_count;
2355 		obj->err_data.ce_count += err_data.ce_count;
2356 		obj->err_data.de_count += err_data.de_count;
2357 	}
2358 
2359 	amdgpu_ras_error_data_fini(&err_data);
2360 }
2361 
2362 static void amdgpu_ras_interrupt_handler(struct ras_manager *obj)
2363 {
2364 	struct ras_ih_data *data = &obj->ih_data;
2365 	struct amdgpu_iv_entry entry;
2366 
2367 	while (data->rptr != data->wptr) {
2368 		rmb();
2369 		memcpy(&entry, &data->ring[data->rptr],
2370 				data->element_size);
2371 
2372 		wmb();
2373 		data->rptr = (data->aligned_element_size +
2374 				data->rptr) % data->ring_size;
2375 
2376 		if (amdgpu_ras_is_poison_mode_supported(obj->adev)) {
2377 			if (obj->head.block == AMDGPU_RAS_BLOCK__UMC)
2378 				amdgpu_ras_interrupt_poison_creation_handler(obj, &entry);
2379 			else
2380 				amdgpu_ras_interrupt_poison_consumption_handler(obj, &entry);
2381 		} else {
2382 			if (obj->head.block == AMDGPU_RAS_BLOCK__UMC)
2383 				amdgpu_ras_interrupt_umc_handler(obj, &entry);
2384 			else
2385 				dev_warn(obj->adev->dev,
2386 					"No RAS interrupt handler for non-UMC block with poison disabled.\n");
2387 		}
2388 	}
2389 }
2390 
2391 static void amdgpu_ras_interrupt_process_handler(struct work_struct *work)
2392 {
2393 	struct ras_ih_data *data =
2394 		container_of(work, struct ras_ih_data, ih_work);
2395 	struct ras_manager *obj =
2396 		container_of(data, struct ras_manager, ih_data);
2397 
2398 	amdgpu_ras_interrupt_handler(obj);
2399 }
2400 
2401 int amdgpu_ras_interrupt_dispatch(struct amdgpu_device *adev,
2402 		struct ras_dispatch_if *info)
2403 {
2404 	struct ras_manager *obj;
2405 	struct ras_ih_data *data;
2406 
2407 	obj = amdgpu_ras_find_obj(adev, &info->head);
2408 	if (!obj)
2409 		return -EINVAL;
2410 
2411 	data = &obj->ih_data;
2412 
2413 	if (data->inuse == 0)
2414 		return 0;
2415 
2416 	/* Might be overflow... */
2417 	memcpy(&data->ring[data->wptr], info->entry,
2418 			data->element_size);
2419 
2420 	wmb();
2421 	data->wptr = (data->aligned_element_size +
2422 			data->wptr) % data->ring_size;
2423 
2424 	schedule_work(&data->ih_work);
2425 
2426 	return 0;
2427 }
2428 
2429 int amdgpu_ras_interrupt_remove_handler(struct amdgpu_device *adev,
2430 		struct ras_common_if *head)
2431 {
2432 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
2433 	struct ras_ih_data *data;
2434 
2435 	if (!obj)
2436 		return -EINVAL;
2437 
2438 	data = &obj->ih_data;
2439 	if (data->inuse == 0)
2440 		return 0;
2441 
2442 	cancel_work_sync(&data->ih_work);
2443 
2444 	kfree(data->ring);
2445 	memset(data, 0, sizeof(*data));
2446 	put_obj(obj);
2447 
2448 	return 0;
2449 }
2450 
2451 int amdgpu_ras_interrupt_add_handler(struct amdgpu_device *adev,
2452 		struct ras_common_if *head)
2453 {
2454 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
2455 	struct ras_ih_data *data;
2456 	struct amdgpu_ras_block_object *ras_obj;
2457 
2458 	if (!obj) {
2459 		/* in case we registe the IH before enable ras feature */
2460 		obj = amdgpu_ras_create_obj(adev, head);
2461 		if (!obj)
2462 			return -EINVAL;
2463 	} else
2464 		get_obj(obj);
2465 
2466 	ras_obj = container_of(head, struct amdgpu_ras_block_object, ras_comm);
2467 
2468 	data = &obj->ih_data;
2469 	/* add the callback.etc */
2470 	*data = (struct ras_ih_data) {
2471 		.inuse = 0,
2472 		.cb = ras_obj->ras_cb,
2473 		.element_size = sizeof(struct amdgpu_iv_entry),
2474 		.rptr = 0,
2475 		.wptr = 0,
2476 	};
2477 
2478 	INIT_WORK(&data->ih_work, amdgpu_ras_interrupt_process_handler);
2479 
2480 	data->aligned_element_size = ALIGN(data->element_size, 8);
2481 	/* the ring can store 64 iv entries. */
2482 	data->ring_size = 64 * data->aligned_element_size;
2483 	data->ring = kmalloc(data->ring_size, GFP_KERNEL);
2484 	if (!data->ring) {
2485 		put_obj(obj);
2486 		return -ENOMEM;
2487 	}
2488 
2489 	/* IH is ready */
2490 	data->inuse = 1;
2491 
2492 	return 0;
2493 }
2494 
2495 static int amdgpu_ras_interrupt_remove_all(struct amdgpu_device *adev)
2496 {
2497 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2498 	struct ras_manager *obj, *tmp;
2499 
2500 	list_for_each_entry_safe(obj, tmp, &con->head, node) {
2501 		amdgpu_ras_interrupt_remove_handler(adev, &obj->head);
2502 	}
2503 
2504 	return 0;
2505 }
2506 /* ih end */
2507 
2508 /* traversal all IPs except NBIO to query error counter */
2509 static void amdgpu_ras_log_on_err_counter(struct amdgpu_device *adev, enum ras_event_type type)
2510 {
2511 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2512 	struct ras_manager *obj;
2513 
2514 	if (!adev->ras_enabled || !con)
2515 		return;
2516 
2517 	list_for_each_entry(obj, &con->head, node) {
2518 		struct ras_query_if info = {
2519 			.head = obj->head,
2520 		};
2521 
2522 		/*
2523 		 * PCIE_BIF IP has one different isr by ras controller
2524 		 * interrupt, the specific ras counter query will be
2525 		 * done in that isr. So skip such block from common
2526 		 * sync flood interrupt isr calling.
2527 		 */
2528 		if (info.head.block == AMDGPU_RAS_BLOCK__PCIE_BIF)
2529 			continue;
2530 
2531 		/*
2532 		 * this is a workaround for aldebaran, skip send msg to
2533 		 * smu to get ecc_info table due to smu handle get ecc
2534 		 * info table failed temporarily.
2535 		 * should be removed until smu fix handle ecc_info table.
2536 		 */
2537 		if ((info.head.block == AMDGPU_RAS_BLOCK__UMC) &&
2538 		    (amdgpu_ip_version(adev, MP1_HWIP, 0) ==
2539 		     IP_VERSION(13, 0, 2)))
2540 			continue;
2541 
2542 		amdgpu_ras_query_error_status_with_event(adev, &info, type);
2543 
2544 		if (amdgpu_ip_version(adev, MP0_HWIP, 0) !=
2545 			    IP_VERSION(11, 0, 2) &&
2546 		    amdgpu_ip_version(adev, MP0_HWIP, 0) !=
2547 			    IP_VERSION(11, 0, 4) &&
2548 		    amdgpu_ip_version(adev, MP0_HWIP, 0) !=
2549 			    IP_VERSION(13, 0, 0)) {
2550 			if (amdgpu_ras_reset_error_status(adev, info.head.block))
2551 				dev_warn(adev->dev, "Failed to reset error counter and error status");
2552 		}
2553 	}
2554 }
2555 
2556 /* Parse RdRspStatus and WrRspStatus */
2557 static void amdgpu_ras_error_status_query(struct amdgpu_device *adev,
2558 					  struct ras_query_if *info)
2559 {
2560 	struct amdgpu_ras_block_object *block_obj;
2561 	/*
2562 	 * Only two block need to query read/write
2563 	 * RspStatus at current state
2564 	 */
2565 	if ((info->head.block != AMDGPU_RAS_BLOCK__GFX) &&
2566 		(info->head.block != AMDGPU_RAS_BLOCK__MMHUB))
2567 		return;
2568 
2569 	block_obj = amdgpu_ras_get_ras_block(adev,
2570 					info->head.block,
2571 					info->head.sub_block_index);
2572 
2573 	if (!block_obj || !block_obj->hw_ops) {
2574 		dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
2575 			     get_ras_block_str(&info->head));
2576 		return;
2577 	}
2578 
2579 	if (block_obj->hw_ops->query_ras_error_status)
2580 		block_obj->hw_ops->query_ras_error_status(adev);
2581 
2582 }
2583 
2584 static void amdgpu_ras_query_err_status(struct amdgpu_device *adev)
2585 {
2586 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2587 	struct ras_manager *obj;
2588 
2589 	if (!adev->ras_enabled || !con)
2590 		return;
2591 
2592 	list_for_each_entry(obj, &con->head, node) {
2593 		struct ras_query_if info = {
2594 			.head = obj->head,
2595 		};
2596 
2597 		amdgpu_ras_error_status_query(adev, &info);
2598 	}
2599 }
2600 
2601 /* recovery begin */
2602 
2603 /* return 0 on success.
2604  * caller need free bps.
2605  */
2606 static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
2607 		struct ras_badpage **bps, unsigned int *count)
2608 {
2609 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2610 	struct ras_err_handler_data *data;
2611 	int i = 0;
2612 	int ret = 0, status;
2613 
2614 	if (!con || !con->eh_data || !bps || !count)
2615 		return -EINVAL;
2616 
2617 	mutex_lock(&con->recovery_lock);
2618 	data = con->eh_data;
2619 	if (!data || data->count == 0) {
2620 		*bps = NULL;
2621 		ret = -EINVAL;
2622 		goto out;
2623 	}
2624 
2625 	*bps = kmalloc_array(data->count, sizeof(struct ras_badpage), GFP_KERNEL);
2626 	if (!*bps) {
2627 		ret = -ENOMEM;
2628 		goto out;
2629 	}
2630 
2631 	for (; i < data->count; i++) {
2632 		if (!data->bps[i].ts)
2633 			continue;
2634 
2635 		(*bps)[i] = (struct ras_badpage){
2636 			.bp = data->bps[i].retired_page,
2637 			.size = AMDGPU_GPU_PAGE_SIZE,
2638 			.flags = AMDGPU_RAS_RETIRE_PAGE_RESERVED,
2639 		};
2640 
2641 		if (amdgpu_ras_check_critical_address(adev,
2642 			data->bps[i].retired_page << AMDGPU_GPU_PAGE_SHIFT))
2643 			continue;
2644 
2645 		status = amdgpu_vram_mgr_query_page_status(&adev->mman.vram_mgr,
2646 				data->bps[i].retired_page << AMDGPU_GPU_PAGE_SHIFT);
2647 		if (status == -EBUSY)
2648 			(*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_PENDING;
2649 		else if (status == -ENOENT)
2650 			(*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_FAULT;
2651 	}
2652 
2653 	*count = con->bad_page_num;
2654 out:
2655 	mutex_unlock(&con->recovery_lock);
2656 	return ret;
2657 }
2658 
2659 static void amdgpu_ras_set_fed_all(struct amdgpu_device *adev,
2660 				   struct amdgpu_hive_info *hive, bool status)
2661 {
2662 	struct amdgpu_device *tmp_adev;
2663 
2664 	if (hive) {
2665 		list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head)
2666 			amdgpu_ras_set_fed(tmp_adev, status);
2667 	} else {
2668 		amdgpu_ras_set_fed(adev, status);
2669 	}
2670 }
2671 
2672 bool amdgpu_ras_in_recovery(struct amdgpu_device *adev)
2673 {
2674 	struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
2675 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
2676 	int hive_ras_recovery = 0;
2677 
2678 	if (hive) {
2679 		hive_ras_recovery = atomic_read(&hive->ras_recovery);
2680 		amdgpu_put_xgmi_hive(hive);
2681 	}
2682 
2683 	if (ras && (atomic_read(&ras->in_recovery) || hive_ras_recovery))
2684 		return true;
2685 
2686 	return false;
2687 }
2688 
2689 static enum ras_event_type amdgpu_ras_get_fatal_error_event(struct amdgpu_device *adev)
2690 {
2691 	if (amdgpu_ras_intr_triggered())
2692 		return RAS_EVENT_TYPE_FATAL;
2693 	else
2694 		return RAS_EVENT_TYPE_POISON_CONSUMPTION;
2695 }
2696 
2697 static void amdgpu_ras_do_recovery(struct work_struct *work)
2698 {
2699 	struct amdgpu_ras *ras =
2700 		container_of(work, struct amdgpu_ras, recovery_work);
2701 	struct amdgpu_device *remote_adev = NULL;
2702 	struct amdgpu_device *adev = ras->adev;
2703 	struct list_head device_list, *device_list_handle =  NULL;
2704 	struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
2705 	enum ras_event_type type;
2706 
2707 	if (hive) {
2708 		atomic_set(&hive->ras_recovery, 1);
2709 
2710 		/* If any device which is part of the hive received RAS fatal
2711 		 * error interrupt, set fatal error status on all. This
2712 		 * condition will need a recovery, and flag will be cleared
2713 		 * as part of recovery.
2714 		 */
2715 		list_for_each_entry(remote_adev, &hive->device_list,
2716 				    gmc.xgmi.head)
2717 			if (amdgpu_ras_get_fed_status(remote_adev)) {
2718 				amdgpu_ras_set_fed_all(adev, hive, true);
2719 				break;
2720 			}
2721 	}
2722 	if (!ras->disable_ras_err_cnt_harvest) {
2723 
2724 		/* Build list of devices to query RAS related errors */
2725 		if  (hive && adev->gmc.xgmi.num_physical_nodes > 1) {
2726 			device_list_handle = &hive->device_list;
2727 		} else {
2728 			INIT_LIST_HEAD(&device_list);
2729 			list_add_tail(&adev->gmc.xgmi.head, &device_list);
2730 			device_list_handle = &device_list;
2731 		}
2732 
2733 		type = amdgpu_ras_get_fatal_error_event(adev);
2734 		list_for_each_entry(remote_adev,
2735 				device_list_handle, gmc.xgmi.head) {
2736 			amdgpu_ras_query_err_status(remote_adev);
2737 			amdgpu_ras_log_on_err_counter(remote_adev, type);
2738 		}
2739 
2740 	}
2741 
2742 	if (amdgpu_device_should_recover_gpu(ras->adev)) {
2743 		struct amdgpu_reset_context reset_context;
2744 		memset(&reset_context, 0, sizeof(reset_context));
2745 
2746 		reset_context.method = AMD_RESET_METHOD_NONE;
2747 		reset_context.reset_req_dev = adev;
2748 		reset_context.src = AMDGPU_RESET_SRC_RAS;
2749 		set_bit(AMDGPU_SKIP_COREDUMP, &reset_context.flags);
2750 
2751 		/* Perform full reset in fatal error mode */
2752 		if (!amdgpu_ras_is_poison_mode_supported(ras->adev))
2753 			set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
2754 		else {
2755 			clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
2756 
2757 			if (ras->gpu_reset_flags & AMDGPU_RAS_GPU_RESET_MODE2_RESET) {
2758 				ras->gpu_reset_flags &= ~AMDGPU_RAS_GPU_RESET_MODE2_RESET;
2759 				reset_context.method = AMD_RESET_METHOD_MODE2;
2760 			}
2761 
2762 			/* Fatal error occurs in poison mode, mode1 reset is used to
2763 			 * recover gpu.
2764 			 */
2765 			if (ras->gpu_reset_flags & AMDGPU_RAS_GPU_RESET_MODE1_RESET) {
2766 				ras->gpu_reset_flags &= ~AMDGPU_RAS_GPU_RESET_MODE1_RESET;
2767 				set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
2768 
2769 				psp_fatal_error_recovery_quirk(&adev->psp);
2770 			}
2771 		}
2772 
2773 		amdgpu_device_gpu_recover(ras->adev, NULL, &reset_context);
2774 	}
2775 	atomic_set(&ras->in_recovery, 0);
2776 	if (hive) {
2777 		atomic_set(&hive->ras_recovery, 0);
2778 		amdgpu_put_xgmi_hive(hive);
2779 	}
2780 }
2781 
2782 /* alloc/realloc bps array */
2783 static int amdgpu_ras_realloc_eh_data_space(struct amdgpu_device *adev,
2784 		struct ras_err_handler_data *data, int pages)
2785 {
2786 	unsigned int old_space = data->count + data->space_left;
2787 	unsigned int new_space = old_space + pages;
2788 	unsigned int align_space = ALIGN(new_space, 512);
2789 	void *bps = kmalloc_array(align_space, sizeof(*data->bps), GFP_KERNEL);
2790 
2791 	if (!bps) {
2792 		return -ENOMEM;
2793 	}
2794 
2795 	if (data->bps) {
2796 		memcpy(bps, data->bps,
2797 				data->count * sizeof(*data->bps));
2798 		kfree(data->bps);
2799 	}
2800 
2801 	data->bps = bps;
2802 	data->space_left += align_space - old_space;
2803 	return 0;
2804 }
2805 
2806 static int amdgpu_ras_mca2pa_by_idx(struct amdgpu_device *adev,
2807 			struct eeprom_table_record *bps,
2808 			struct ras_err_data *err_data)
2809 {
2810 	struct ta_ras_query_address_input addr_in;
2811 	uint32_t socket = 0;
2812 	int ret = 0;
2813 
2814 	if (adev->smuio.funcs && adev->smuio.funcs->get_socket_id)
2815 		socket = adev->smuio.funcs->get_socket_id(adev);
2816 
2817 	/* reinit err_data */
2818 	err_data->err_addr_cnt = 0;
2819 	err_data->err_addr_len = adev->umc.retire_unit;
2820 
2821 	memset(&addr_in, 0, sizeof(addr_in));
2822 	addr_in.ma.err_addr = bps->address;
2823 	addr_in.ma.socket_id = socket;
2824 	addr_in.ma.ch_inst = bps->mem_channel;
2825 	/* tell RAS TA the node instance is not used */
2826 	addr_in.ma.node_inst = TA_RAS_INV_NODE;
2827 
2828 	if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr)
2829 		ret = adev->umc.ras->convert_ras_err_addr(adev, err_data,
2830 				&addr_in, NULL, false);
2831 
2832 	return ret;
2833 }
2834 
2835 static int amdgpu_ras_mca2pa(struct amdgpu_device *adev,
2836 			struct eeprom_table_record *bps,
2837 			struct ras_err_data *err_data)
2838 {
2839 	struct ta_ras_query_address_input addr_in;
2840 	uint32_t die_id, socket = 0;
2841 
2842 	if (adev->smuio.funcs && adev->smuio.funcs->get_socket_id)
2843 		socket = adev->smuio.funcs->get_socket_id(adev);
2844 
2845 	/* although die id is gotten from PA in nps1 mode, the id is
2846 	 * fitable for any nps mode
2847 	 */
2848 	if (adev->umc.ras && adev->umc.ras->get_die_id_from_pa)
2849 		die_id = adev->umc.ras->get_die_id_from_pa(adev, bps->address,
2850 					bps->retired_page << AMDGPU_GPU_PAGE_SHIFT);
2851 	else
2852 		return -EINVAL;
2853 
2854 	/* reinit err_data */
2855 	err_data->err_addr_cnt = 0;
2856 	err_data->err_addr_len = adev->umc.retire_unit;
2857 
2858 	memset(&addr_in, 0, sizeof(addr_in));
2859 	addr_in.ma.err_addr = bps->address;
2860 	addr_in.ma.ch_inst = bps->mem_channel;
2861 	addr_in.ma.umc_inst = bps->mcumc_id;
2862 	addr_in.ma.node_inst = die_id;
2863 	addr_in.ma.socket_id = socket;
2864 
2865 	if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr)
2866 		return adev->umc.ras->convert_ras_err_addr(adev, err_data,
2867 					&addr_in, NULL, false);
2868 	else
2869 		return  -EINVAL;
2870 }
2871 
2872 static int __amdgpu_ras_restore_bad_pages(struct amdgpu_device *adev,
2873 					struct eeprom_table_record *bps, int count)
2874 {
2875 	int j;
2876 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2877 	struct ras_err_handler_data *data = con->eh_data;
2878 
2879 	for (j = 0; j < count; j++) {
2880 		if (amdgpu_ras_check_bad_page_unlock(con,
2881 			bps[j].retired_page << AMDGPU_GPU_PAGE_SHIFT)) {
2882 			data->count++;
2883 			data->space_left--;
2884 			continue;
2885 		}
2886 
2887 		if (!data->space_left &&
2888 		    amdgpu_ras_realloc_eh_data_space(adev, data, 256)) {
2889 			return -ENOMEM;
2890 		}
2891 
2892 		amdgpu_ras_reserve_page(adev, bps[j].retired_page);
2893 
2894 		memcpy(&data->bps[data->count], &(bps[j]),
2895 				sizeof(struct eeprom_table_record));
2896 		data->count++;
2897 		data->space_left--;
2898 		con->bad_page_num++;
2899 	}
2900 
2901 	return 0;
2902 }
2903 
2904 static int __amdgpu_ras_convert_rec_array_from_rom(struct amdgpu_device *adev,
2905 				struct eeprom_table_record *bps, struct ras_err_data *err_data,
2906 				enum amdgpu_memory_partition nps)
2907 {
2908 	int i = 0;
2909 	enum amdgpu_memory_partition save_nps;
2910 
2911 	save_nps = (bps[0].retired_page >> UMC_NPS_SHIFT) & UMC_NPS_MASK;
2912 
2913 	/*old asics just have pa in eeprom*/
2914 	if (IP_VERSION_MAJ(amdgpu_ip_version(adev, UMC_HWIP, 0)) < 12) {
2915 		memcpy(err_data->err_addr, bps,
2916 			sizeof(struct eeprom_table_record) * adev->umc.retire_unit);
2917 		goto out;
2918 	}
2919 
2920 	for (i = 0; i < adev->umc.retire_unit; i++)
2921 		bps[i].retired_page &= ~(UMC_NPS_MASK << UMC_NPS_SHIFT);
2922 
2923 	if (save_nps) {
2924 		if (save_nps == nps) {
2925 			if (amdgpu_umc_pages_in_a_row(adev, err_data,
2926 					bps[0].retired_page << AMDGPU_GPU_PAGE_SHIFT))
2927 				return -EINVAL;
2928 			for (i = 0; i < adev->umc.retire_unit; i++) {
2929 				err_data->err_addr[i].address = bps[0].address;
2930 				err_data->err_addr[i].mem_channel = bps[0].mem_channel;
2931 				err_data->err_addr[i].bank = bps[0].bank;
2932 				err_data->err_addr[i].err_type = bps[0].err_type;
2933 				err_data->err_addr[i].mcumc_id = bps[0].mcumc_id;
2934 			}
2935 		} else {
2936 			if (amdgpu_ras_mca2pa_by_idx(adev, &bps[0], err_data))
2937 				return -EINVAL;
2938 		}
2939 	} else {
2940 		if (bps[0].address == 0) {
2941 			/* for specific old eeprom data, mca address is not stored,
2942 			 * calc it from pa
2943 			 */
2944 			if (amdgpu_umc_pa2mca(adev, bps[0].retired_page << AMDGPU_GPU_PAGE_SHIFT,
2945 				&(bps[0].address), AMDGPU_NPS1_PARTITION_MODE))
2946 				return -EINVAL;
2947 		}
2948 
2949 		if (amdgpu_ras_mca2pa(adev, &bps[0], err_data)) {
2950 			if (nps == AMDGPU_NPS1_PARTITION_MODE)
2951 				memcpy(err_data->err_addr, bps,
2952 					sizeof(struct eeprom_table_record) * adev->umc.retire_unit);
2953 			else
2954 				return -EOPNOTSUPP;
2955 		}
2956 	}
2957 
2958 out:
2959 	return __amdgpu_ras_restore_bad_pages(adev, err_data->err_addr, adev->umc.retire_unit);
2960 }
2961 
2962 static int __amdgpu_ras_convert_rec_from_rom(struct amdgpu_device *adev,
2963 				struct eeprom_table_record *bps, struct ras_err_data *err_data,
2964 				enum amdgpu_memory_partition nps)
2965 {
2966 	int i = 0;
2967 	enum amdgpu_memory_partition save_nps;
2968 
2969 	save_nps = (bps->retired_page >> UMC_NPS_SHIFT) & UMC_NPS_MASK;
2970 	bps->retired_page &= ~(UMC_NPS_MASK << UMC_NPS_SHIFT);
2971 
2972 	if (save_nps == nps) {
2973 		if (amdgpu_umc_pages_in_a_row(adev, err_data,
2974 				bps->retired_page << AMDGPU_GPU_PAGE_SHIFT))
2975 			return -EINVAL;
2976 		for (i = 0; i < adev->umc.retire_unit; i++) {
2977 			err_data->err_addr[i].address = bps->address;
2978 			err_data->err_addr[i].mem_channel = bps->mem_channel;
2979 			err_data->err_addr[i].bank = bps->bank;
2980 			err_data->err_addr[i].err_type = bps->err_type;
2981 			err_data->err_addr[i].mcumc_id = bps->mcumc_id;
2982 		}
2983 	} else {
2984 		if (bps->address) {
2985 			if (amdgpu_ras_mca2pa_by_idx(adev, bps, err_data))
2986 				return -EINVAL;
2987 		} else {
2988 			/* for specific old eeprom data, mca address is not stored,
2989 			 * calc it from pa
2990 			 */
2991 			if (amdgpu_umc_pa2mca(adev, bps->retired_page << AMDGPU_GPU_PAGE_SHIFT,
2992 				&(bps->address), AMDGPU_NPS1_PARTITION_MODE))
2993 				return -EINVAL;
2994 
2995 			if (amdgpu_ras_mca2pa(adev, bps, err_data))
2996 				return -EOPNOTSUPP;
2997 		}
2998 	}
2999 
3000 	return __amdgpu_ras_restore_bad_pages(adev, err_data->err_addr,
3001 									adev->umc.retire_unit);
3002 }
3003 
3004 /* it deal with vram only. */
3005 int amdgpu_ras_add_bad_pages(struct amdgpu_device *adev,
3006 		struct eeprom_table_record *bps, int pages, bool from_rom)
3007 {
3008 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3009 	struct ras_err_data err_data;
3010 	struct amdgpu_ras_eeprom_control *control =
3011 			&adev->psp.ras_context.ras->eeprom_control;
3012 	enum amdgpu_memory_partition nps = AMDGPU_NPS1_PARTITION_MODE;
3013 	int ret = 0;
3014 	uint32_t i = 0;
3015 
3016 	if (!con || !con->eh_data || !bps || pages <= 0)
3017 		return 0;
3018 
3019 	if (from_rom) {
3020 		err_data.err_addr =
3021 			kcalloc(adev->umc.retire_unit,
3022 				sizeof(struct eeprom_table_record), GFP_KERNEL);
3023 		if (!err_data.err_addr) {
3024 			dev_warn(adev->dev, "Failed to alloc UMC error address record in mca2pa conversion!\n");
3025 			return -ENOMEM;
3026 		}
3027 
3028 		if (adev->gmc.gmc_funcs->query_mem_partition_mode)
3029 			nps = adev->gmc.gmc_funcs->query_mem_partition_mode(adev);
3030 	}
3031 
3032 	mutex_lock(&con->recovery_lock);
3033 
3034 	if (from_rom) {
3035 		/* there is no pa recs in V3, so skip pa recs processing */
3036 		if (control->tbl_hdr.version < RAS_TABLE_VER_V3) {
3037 			for (i = 0; i < pages; i++) {
3038 				if (control->ras_num_recs - i >= adev->umc.retire_unit) {
3039 					if ((bps[i].address == bps[i + 1].address) &&
3040 						(bps[i].mem_channel == bps[i + 1].mem_channel)) {
3041 						/* deal with retire_unit records a time */
3042 						ret = __amdgpu_ras_convert_rec_array_from_rom(adev,
3043 										&bps[i], &err_data, nps);
3044 						if (ret)
3045 							con->bad_page_num -= adev->umc.retire_unit;
3046 						i += (adev->umc.retire_unit - 1);
3047 					} else {
3048 						break;
3049 					}
3050 				} else {
3051 					break;
3052 				}
3053 			}
3054 		}
3055 		for (; i < pages; i++) {
3056 			ret = __amdgpu_ras_convert_rec_from_rom(adev,
3057 				&bps[i], &err_data, nps);
3058 			if (ret)
3059 				con->bad_page_num -= adev->umc.retire_unit;
3060 		}
3061 
3062 		con->eh_data->count_saved = con->eh_data->count;
3063 	} else {
3064 		ret = __amdgpu_ras_restore_bad_pages(adev, bps, pages);
3065 	}
3066 
3067 	if (from_rom)
3068 		kfree(err_data.err_addr);
3069 	mutex_unlock(&con->recovery_lock);
3070 
3071 	return ret;
3072 }
3073 
3074 /*
3075  * write error record array to eeprom, the function should be
3076  * protected by recovery_lock
3077  * new_cnt: new added UE count, excluding reserved bad pages, can be NULL
3078  */
3079 int amdgpu_ras_save_bad_pages(struct amdgpu_device *adev,
3080 		unsigned long *new_cnt)
3081 {
3082 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3083 	struct ras_err_handler_data *data;
3084 	struct amdgpu_ras_eeprom_control *control;
3085 	int save_count, unit_num, i;
3086 
3087 	if (!con || !con->eh_data) {
3088 		if (new_cnt)
3089 			*new_cnt = 0;
3090 
3091 		return 0;
3092 	}
3093 
3094 	if (!con->eeprom_control.is_eeprom_valid) {
3095 		dev_warn(adev->dev,
3096 			"Failed to save EEPROM table data because of EEPROM data corruption!");
3097 		if (new_cnt)
3098 			*new_cnt = 0;
3099 
3100 		return 0;
3101 	}
3102 
3103 	mutex_lock(&con->recovery_lock);
3104 	control = &con->eeprom_control;
3105 	data = con->eh_data;
3106 	unit_num = data->count / adev->umc.retire_unit - control->ras_num_recs;
3107 	save_count = con->bad_page_num - control->ras_num_bad_pages;
3108 	mutex_unlock(&con->recovery_lock);
3109 
3110 	if (new_cnt)
3111 		*new_cnt = unit_num;
3112 
3113 	/* only new entries are saved */
3114 	if (unit_num > 0) {
3115 		/*old asics only save pa to eeprom like before*/
3116 		if (IP_VERSION_MAJ(amdgpu_ip_version(adev, UMC_HWIP, 0)) < 12) {
3117 			if (amdgpu_ras_eeprom_append(control,
3118 					&data->bps[data->count_saved], unit_num)) {
3119 				dev_err(adev->dev, "Failed to save EEPROM table data!");
3120 				return -EIO;
3121 			}
3122 		} else {
3123 			for (i = 0; i < unit_num; i++) {
3124 				if (amdgpu_ras_eeprom_append(control,
3125 						&data->bps[data->count_saved +
3126 						i * adev->umc.retire_unit], 1)) {
3127 					dev_err(adev->dev, "Failed to save EEPROM table data!");
3128 					return -EIO;
3129 				}
3130 			}
3131 		}
3132 
3133 		dev_info(adev->dev, "Saved %d pages to EEPROM table.\n", save_count);
3134 		data->count_saved = data->count;
3135 	}
3136 
3137 	return 0;
3138 }
3139 
3140 /*
3141  * read error record array in eeprom and reserve enough space for
3142  * storing new bad pages
3143  */
3144 static int amdgpu_ras_load_bad_pages(struct amdgpu_device *adev)
3145 {
3146 	struct amdgpu_ras_eeprom_control *control =
3147 		&adev->psp.ras_context.ras->eeprom_control;
3148 	struct eeprom_table_record *bps;
3149 	int ret, i = 0;
3150 
3151 	/* no bad page record, skip eeprom access */
3152 	if (control->ras_num_recs == 0 || amdgpu_bad_page_threshold == 0)
3153 		return 0;
3154 
3155 	bps = kcalloc(control->ras_num_recs, sizeof(*bps), GFP_KERNEL);
3156 	if (!bps)
3157 		return -ENOMEM;
3158 
3159 	ret = amdgpu_ras_eeprom_read(control, bps, control->ras_num_recs);
3160 	if (ret) {
3161 		dev_err(adev->dev, "Failed to load EEPROM table records!");
3162 	} else {
3163 		if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) {
3164 			/*In V3, there is no pa recs, and some cases(when address==0) may be parsed
3165 			as pa recs, so add verion check to avoid it.
3166 			*/
3167 			if (control->tbl_hdr.version < RAS_TABLE_VER_V3) {
3168 				for (i = 0; i < control->ras_num_recs; i++) {
3169 					if ((control->ras_num_recs - i) >= adev->umc.retire_unit) {
3170 						if ((bps[i].address == bps[i + 1].address) &&
3171 							(bps[i].mem_channel == bps[i + 1].mem_channel)) {
3172 							control->ras_num_pa_recs += adev->umc.retire_unit;
3173 							i += (adev->umc.retire_unit - 1);
3174 						} else {
3175 							control->ras_num_mca_recs +=
3176 										(control->ras_num_recs - i);
3177 							break;
3178 						}
3179 					} else {
3180 						control->ras_num_mca_recs += (control->ras_num_recs - i);
3181 						break;
3182 					}
3183 				}
3184 			} else {
3185 				control->ras_num_mca_recs = control->ras_num_recs;
3186 			}
3187 		}
3188 
3189 		ret = amdgpu_ras_add_bad_pages(adev, bps, control->ras_num_recs, true);
3190 		if (ret)
3191 			goto out;
3192 
3193 		ret = amdgpu_ras_eeprom_check(control);
3194 		if (ret)
3195 			goto out;
3196 
3197 		/* HW not usable */
3198 		if (amdgpu_ras_is_rma(adev))
3199 			ret = -EHWPOISON;
3200 	}
3201 
3202 out:
3203 	kfree(bps);
3204 	return ret;
3205 }
3206 
3207 static int amdgpu_ras_check_bad_page_unlock(struct amdgpu_ras *con,
3208 				uint64_t addr)
3209 {
3210 	struct ras_err_handler_data *data = con->eh_data;
3211 	struct amdgpu_device *adev = con->adev;
3212 	int i;
3213 
3214 	if ((addr >= adev->gmc.mc_vram_size &&
3215 	    adev->gmc.mc_vram_size) ||
3216 	    (addr >= RAS_UMC_INJECT_ADDR_LIMIT))
3217 		return -EINVAL;
3218 
3219 	addr >>= AMDGPU_GPU_PAGE_SHIFT;
3220 	for (i = 0; i < data->count; i++)
3221 		if (addr == data->bps[i].retired_page)
3222 			return 1;
3223 
3224 	return 0;
3225 }
3226 
3227 /*
3228  * check if an address belongs to bad page
3229  *
3230  * Note: this check is only for umc block
3231  */
3232 static int amdgpu_ras_check_bad_page(struct amdgpu_device *adev,
3233 				uint64_t addr)
3234 {
3235 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3236 	int ret = 0;
3237 
3238 	if (!con || !con->eh_data)
3239 		return ret;
3240 
3241 	mutex_lock(&con->recovery_lock);
3242 	ret = amdgpu_ras_check_bad_page_unlock(con, addr);
3243 	mutex_unlock(&con->recovery_lock);
3244 	return ret;
3245 }
3246 
3247 static void amdgpu_ras_validate_threshold(struct amdgpu_device *adev,
3248 					  uint32_t max_count)
3249 {
3250 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3251 
3252 	/*
3253 	 * amdgpu_bad_page_threshold is used to config
3254 	 * the threshold for the number of bad pages.
3255 	 * -1:  Threshold is set to default value
3256 	 *      Driver will issue a warning message when threshold is reached
3257 	 *      and continue runtime services.
3258 	 * 0:   Disable bad page retirement
3259 	 *      Driver will not retire bad pages
3260 	 *      which is intended for debugging purpose.
3261 	 * -2:  Threshold is determined by a formula
3262 	 *      that assumes 1 bad page per 100M of local memory.
3263 	 *      Driver will continue runtime services when threhold is reached.
3264 	 * 0 < threshold < max number of bad page records in EEPROM,
3265 	 *      A user-defined threshold is set
3266 	 *      Driver will halt runtime services when this custom threshold is reached.
3267 	 */
3268 	if (amdgpu_bad_page_threshold == -2) {
3269 		u64 val = adev->gmc.mc_vram_size;
3270 
3271 		do_div(val, RAS_BAD_PAGE_COVER);
3272 		con->bad_page_cnt_threshold = min(lower_32_bits(val),
3273 						  max_count);
3274 	} else if (amdgpu_bad_page_threshold == -1) {
3275 		con->bad_page_cnt_threshold = ((con->reserved_pages_in_bytes) >> 21) << 4;
3276 	} else {
3277 		con->bad_page_cnt_threshold = min_t(int, max_count,
3278 						    amdgpu_bad_page_threshold);
3279 	}
3280 }
3281 
3282 int amdgpu_ras_put_poison_req(struct amdgpu_device *adev,
3283 		enum amdgpu_ras_block block, uint16_t pasid,
3284 		pasid_notify pasid_fn, void *data, uint32_t reset)
3285 {
3286 	int ret = 0;
3287 	struct ras_poison_msg poison_msg;
3288 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3289 
3290 	memset(&poison_msg, 0, sizeof(poison_msg));
3291 	poison_msg.block = block;
3292 	poison_msg.pasid = pasid;
3293 	poison_msg.reset = reset;
3294 	poison_msg.pasid_fn = pasid_fn;
3295 	poison_msg.data = data;
3296 
3297 	ret = kfifo_put(&con->poison_fifo, poison_msg);
3298 	if (!ret) {
3299 		dev_err(adev->dev, "Poison message fifo is full!\n");
3300 		return -ENOSPC;
3301 	}
3302 
3303 	return 0;
3304 }
3305 
3306 static int amdgpu_ras_get_poison_req(struct amdgpu_device *adev,
3307 		struct ras_poison_msg *poison_msg)
3308 {
3309 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3310 
3311 	return kfifo_get(&con->poison_fifo, poison_msg);
3312 }
3313 
3314 static void amdgpu_ras_ecc_log_init(struct ras_ecc_log_info *ecc_log)
3315 {
3316 	mutex_init(&ecc_log->lock);
3317 
3318 	INIT_RADIX_TREE(&ecc_log->de_page_tree, GFP_KERNEL);
3319 	ecc_log->de_queried_count = 0;
3320 	ecc_log->prev_de_queried_count = 0;
3321 }
3322 
3323 static void amdgpu_ras_ecc_log_fini(struct ras_ecc_log_info *ecc_log)
3324 {
3325 	struct radix_tree_iter iter;
3326 	void __rcu **slot;
3327 	struct ras_ecc_err *ecc_err;
3328 
3329 	mutex_lock(&ecc_log->lock);
3330 	radix_tree_for_each_slot(slot, &ecc_log->de_page_tree, &iter, 0) {
3331 		ecc_err = radix_tree_deref_slot(slot);
3332 		kfree(ecc_err->err_pages.pfn);
3333 		kfree(ecc_err);
3334 		radix_tree_iter_delete(&ecc_log->de_page_tree, &iter, slot);
3335 	}
3336 	mutex_unlock(&ecc_log->lock);
3337 
3338 	mutex_destroy(&ecc_log->lock);
3339 	ecc_log->de_queried_count = 0;
3340 	ecc_log->prev_de_queried_count = 0;
3341 }
3342 
3343 static bool amdgpu_ras_schedule_retirement_dwork(struct amdgpu_ras *con,
3344 				uint32_t delayed_ms)
3345 {
3346 	int ret;
3347 
3348 	mutex_lock(&con->umc_ecc_log.lock);
3349 	ret = radix_tree_tagged(&con->umc_ecc_log.de_page_tree,
3350 			UMC_ECC_NEW_DETECTED_TAG);
3351 	mutex_unlock(&con->umc_ecc_log.lock);
3352 
3353 	if (ret)
3354 		schedule_delayed_work(&con->page_retirement_dwork,
3355 			msecs_to_jiffies(delayed_ms));
3356 
3357 	return ret ? true : false;
3358 }
3359 
3360 static void amdgpu_ras_do_page_retirement(struct work_struct *work)
3361 {
3362 	struct amdgpu_ras *con = container_of(work, struct amdgpu_ras,
3363 					      page_retirement_dwork.work);
3364 	struct amdgpu_device *adev = con->adev;
3365 	struct ras_err_data err_data;
3366 
3367 	/* If gpu reset is ongoing, delay retiring the bad pages */
3368 	if (amdgpu_in_reset(adev) || amdgpu_ras_in_recovery(adev)) {
3369 		amdgpu_ras_schedule_retirement_dwork(con,
3370 				AMDGPU_RAS_RETIRE_PAGE_INTERVAL * 3);
3371 		return;
3372 	}
3373 
3374 	amdgpu_ras_error_data_init(&err_data);
3375 
3376 	amdgpu_umc_handle_bad_pages(adev, &err_data);
3377 
3378 	amdgpu_ras_error_data_fini(&err_data);
3379 
3380 	amdgpu_ras_schedule_retirement_dwork(con,
3381 			AMDGPU_RAS_RETIRE_PAGE_INTERVAL);
3382 }
3383 
3384 static int amdgpu_ras_poison_creation_handler(struct amdgpu_device *adev,
3385 				uint32_t poison_creation_count)
3386 {
3387 	int ret = 0;
3388 	struct ras_ecc_log_info *ecc_log;
3389 	struct ras_query_if info;
3390 	uint32_t timeout = 0;
3391 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
3392 	uint64_t de_queried_count;
3393 	uint32_t new_detect_count, total_detect_count;
3394 	uint32_t need_query_count = poison_creation_count;
3395 	enum ras_event_type type = RAS_EVENT_TYPE_POISON_CREATION;
3396 
3397 	memset(&info, 0, sizeof(info));
3398 	info.head.block = AMDGPU_RAS_BLOCK__UMC;
3399 
3400 	ecc_log = &ras->umc_ecc_log;
3401 	total_detect_count = 0;
3402 	do {
3403 		ret = amdgpu_ras_query_error_status_with_event(adev, &info, type);
3404 		if (ret)
3405 			return ret;
3406 
3407 		de_queried_count = ecc_log->de_queried_count;
3408 		if (de_queried_count > ecc_log->prev_de_queried_count) {
3409 			new_detect_count = de_queried_count - ecc_log->prev_de_queried_count;
3410 			ecc_log->prev_de_queried_count = de_queried_count;
3411 			timeout = 0;
3412 		} else {
3413 			new_detect_count = 0;
3414 		}
3415 
3416 		if (new_detect_count) {
3417 			total_detect_count += new_detect_count;
3418 		} else {
3419 			if (!timeout && need_query_count)
3420 				timeout = MAX_UMC_POISON_POLLING_TIME_ASYNC;
3421 
3422 			if (timeout) {
3423 				if (!--timeout)
3424 					break;
3425 				msleep(1);
3426 			}
3427 		}
3428 	} while (total_detect_count < need_query_count);
3429 
3430 	if (total_detect_count)
3431 		schedule_delayed_work(&ras->page_retirement_dwork, 0);
3432 
3433 	if (amdgpu_ras_is_rma(adev) && atomic_cmpxchg(&ras->rma_in_recovery, 0, 1) == 0)
3434 		amdgpu_ras_reset_gpu(adev);
3435 
3436 	return 0;
3437 }
3438 
3439 static void amdgpu_ras_clear_poison_fifo(struct amdgpu_device *adev)
3440 {
3441 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3442 	struct ras_poison_msg msg;
3443 	int ret;
3444 
3445 	do {
3446 		ret = kfifo_get(&con->poison_fifo, &msg);
3447 	} while (ret);
3448 }
3449 
3450 static int amdgpu_ras_poison_consumption_handler(struct amdgpu_device *adev,
3451 			uint32_t msg_count, uint32_t *gpu_reset)
3452 {
3453 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3454 	uint32_t reset_flags = 0, reset = 0;
3455 	struct ras_poison_msg msg;
3456 	int ret, i;
3457 
3458 	kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
3459 
3460 	for (i = 0; i < msg_count; i++) {
3461 		ret = amdgpu_ras_get_poison_req(adev, &msg);
3462 		if (!ret)
3463 			continue;
3464 
3465 		if (msg.pasid_fn)
3466 			msg.pasid_fn(adev, msg.pasid, msg.data);
3467 
3468 		reset_flags |= msg.reset;
3469 	}
3470 
3471 	/*
3472 	 * Try to ensure poison creation handler is completed first
3473 	 * to set rma if bad page exceed threshold.
3474 	 */
3475 	flush_delayed_work(&con->page_retirement_dwork);
3476 
3477 	/* for RMA, amdgpu_ras_poison_creation_handler will trigger gpu reset */
3478 	if (reset_flags && !amdgpu_ras_is_rma(adev)) {
3479 		if (reset_flags & AMDGPU_RAS_GPU_RESET_MODE1_RESET)
3480 			reset = AMDGPU_RAS_GPU_RESET_MODE1_RESET;
3481 		else if (reset_flags & AMDGPU_RAS_GPU_RESET_MODE2_RESET)
3482 			reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET;
3483 		else
3484 			reset = reset_flags;
3485 
3486 		con->gpu_reset_flags |= reset;
3487 		amdgpu_ras_reset_gpu(adev);
3488 
3489 		*gpu_reset = reset;
3490 
3491 		/* Wait for gpu recovery to complete */
3492 		flush_work(&con->recovery_work);
3493 	}
3494 
3495 	return 0;
3496 }
3497 
3498 static int amdgpu_ras_page_retirement_thread(void *param)
3499 {
3500 	struct amdgpu_device *adev = (struct amdgpu_device *)param;
3501 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3502 	uint32_t poison_creation_count, msg_count;
3503 	uint32_t gpu_reset;
3504 	int ret;
3505 
3506 	while (!kthread_should_stop()) {
3507 
3508 		wait_event_interruptible(con->page_retirement_wq,
3509 				kthread_should_stop() ||
3510 				atomic_read(&con->page_retirement_req_cnt));
3511 
3512 		if (kthread_should_stop())
3513 			break;
3514 
3515 		gpu_reset = 0;
3516 
3517 		do {
3518 			poison_creation_count = atomic_read(&con->poison_creation_count);
3519 			ret = amdgpu_ras_poison_creation_handler(adev, poison_creation_count);
3520 			if (ret == -EIO)
3521 				break;
3522 
3523 			if (poison_creation_count) {
3524 				atomic_sub(poison_creation_count, &con->poison_creation_count);
3525 				atomic_sub(poison_creation_count, &con->page_retirement_req_cnt);
3526 			}
3527 		} while (atomic_read(&con->poison_creation_count));
3528 
3529 		if (ret != -EIO) {
3530 			msg_count = kfifo_len(&con->poison_fifo);
3531 			if (msg_count) {
3532 				ret = amdgpu_ras_poison_consumption_handler(adev,
3533 						msg_count, &gpu_reset);
3534 				if ((ret != -EIO) &&
3535 				    (gpu_reset != AMDGPU_RAS_GPU_RESET_MODE1_RESET))
3536 					atomic_sub(msg_count, &con->page_retirement_req_cnt);
3537 			}
3538 		}
3539 
3540 		if ((ret == -EIO) || (gpu_reset == AMDGPU_RAS_GPU_RESET_MODE1_RESET)) {
3541 			/* gpu mode-1 reset is ongoing or just completed ras mode-1 reset */
3542 			/* Clear poison creation request */
3543 			atomic_set(&con->poison_creation_count, 0);
3544 
3545 			/* Clear poison fifo */
3546 			amdgpu_ras_clear_poison_fifo(adev);
3547 
3548 			/* Clear all poison requests */
3549 			atomic_set(&con->page_retirement_req_cnt, 0);
3550 
3551 			if (ret == -EIO) {
3552 				/* Wait for mode-1 reset to complete */
3553 				down_read(&adev->reset_domain->sem);
3554 				up_read(&adev->reset_domain->sem);
3555 			}
3556 
3557 			/* Wake up work to save bad pages to eeprom */
3558 			schedule_delayed_work(&con->page_retirement_dwork, 0);
3559 		} else if (gpu_reset) {
3560 			/* gpu just completed mode-2 reset or other reset */
3561 			/* Clear poison consumption messages cached in fifo */
3562 			msg_count = kfifo_len(&con->poison_fifo);
3563 			if (msg_count) {
3564 				amdgpu_ras_clear_poison_fifo(adev);
3565 				atomic_sub(msg_count, &con->page_retirement_req_cnt);
3566 			}
3567 
3568 			/* Wake up work to save bad pages to eeprom */
3569 			schedule_delayed_work(&con->page_retirement_dwork, 0);
3570 		}
3571 	}
3572 
3573 	return 0;
3574 }
3575 
3576 int amdgpu_ras_init_badpage_info(struct amdgpu_device *adev)
3577 {
3578 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3579 	struct amdgpu_ras_eeprom_control *control;
3580 	int ret;
3581 
3582 	if (!con || amdgpu_sriov_vf(adev))
3583 		return 0;
3584 
3585 	control = &con->eeprom_control;
3586 	ret = amdgpu_ras_eeprom_init(control);
3587 	control->is_eeprom_valid = !ret;
3588 
3589 	if (!adev->umc.ras || !adev->umc.ras->convert_ras_err_addr)
3590 		control->ras_num_pa_recs = control->ras_num_recs;
3591 
3592 	if (adev->umc.ras &&
3593 	    adev->umc.ras->get_retire_flip_bits)
3594 		adev->umc.ras->get_retire_flip_bits(adev);
3595 
3596 	if (control->ras_num_recs && control->is_eeprom_valid) {
3597 		ret = amdgpu_ras_load_bad_pages(adev);
3598 		if (ret) {
3599 			control->is_eeprom_valid = false;
3600 			return 0;
3601 		}
3602 
3603 		amdgpu_dpm_send_hbm_bad_pages_num(
3604 			adev, control->ras_num_bad_pages);
3605 
3606 		if (con->update_channel_flag == true) {
3607 			amdgpu_dpm_send_hbm_bad_channel_flag(
3608 				adev, control->bad_channel_bitmap);
3609 			con->update_channel_flag = false;
3610 		}
3611 
3612 		/* The format action is only applied to new ASICs */
3613 		if (IP_VERSION_MAJ(amdgpu_ip_version(adev, UMC_HWIP, 0)) >= 12 &&
3614 		    control->tbl_hdr.version < RAS_TABLE_VER_V3)
3615 			if (!amdgpu_ras_eeprom_reset_table(control))
3616 				if (amdgpu_ras_save_bad_pages(adev, NULL))
3617 					dev_warn(adev->dev, "Failed to format RAS EEPROM data in V3 version!\n");
3618 	}
3619 
3620 	return 0;
3621 }
3622 
3623 int amdgpu_ras_recovery_init(struct amdgpu_device *adev, bool init_bp_info)
3624 {
3625 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3626 	struct ras_err_handler_data **data;
3627 	u32  max_eeprom_records_count = 0;
3628 	int ret;
3629 
3630 	if (!con || amdgpu_sriov_vf(adev))
3631 		return 0;
3632 
3633 	/* Allow access to RAS EEPROM via debugfs, when the ASIC
3634 	 * supports RAS and debugfs is enabled, but when
3635 	 * adev->ras_enabled is unset, i.e. when "ras_enable"
3636 	 * module parameter is set to 0.
3637 	 */
3638 	con->adev = adev;
3639 
3640 	if (!adev->ras_enabled)
3641 		return 0;
3642 
3643 	data = &con->eh_data;
3644 	*data = kzalloc(sizeof(**data), GFP_KERNEL);
3645 	if (!*data) {
3646 		ret = -ENOMEM;
3647 		goto out;
3648 	}
3649 
3650 	mutex_init(&con->recovery_lock);
3651 	INIT_WORK(&con->recovery_work, amdgpu_ras_do_recovery);
3652 	atomic_set(&con->in_recovery, 0);
3653 	atomic_set(&con->rma_in_recovery, 0);
3654 	con->eeprom_control.bad_channel_bitmap = 0;
3655 
3656 	max_eeprom_records_count = amdgpu_ras_eeprom_max_record_count(&con->eeprom_control);
3657 	amdgpu_ras_validate_threshold(adev, max_eeprom_records_count);
3658 
3659 	if (init_bp_info) {
3660 		ret = amdgpu_ras_init_badpage_info(adev);
3661 		if (ret)
3662 			goto free;
3663 	}
3664 
3665 	mutex_init(&con->page_rsv_lock);
3666 	INIT_KFIFO(con->poison_fifo);
3667 	mutex_init(&con->page_retirement_lock);
3668 	init_waitqueue_head(&con->page_retirement_wq);
3669 	atomic_set(&con->page_retirement_req_cnt, 0);
3670 	atomic_set(&con->poison_creation_count, 0);
3671 	con->page_retirement_thread =
3672 		kthread_run(amdgpu_ras_page_retirement_thread, adev, "umc_page_retirement");
3673 	if (IS_ERR(con->page_retirement_thread)) {
3674 		con->page_retirement_thread = NULL;
3675 		dev_warn(adev->dev, "Failed to create umc_page_retirement thread!!!\n");
3676 	}
3677 
3678 	INIT_DELAYED_WORK(&con->page_retirement_dwork, amdgpu_ras_do_page_retirement);
3679 	amdgpu_ras_ecc_log_init(&con->umc_ecc_log);
3680 #ifdef CONFIG_X86_MCE_AMD
3681 	if ((adev->asic_type == CHIP_ALDEBARAN) &&
3682 	    (adev->gmc.xgmi.connected_to_cpu))
3683 		amdgpu_register_bad_pages_mca_notifier(adev);
3684 #endif
3685 	return 0;
3686 
3687 free:
3688 	kfree((*data)->bps);
3689 	kfree(*data);
3690 	con->eh_data = NULL;
3691 out:
3692 	dev_warn(adev->dev, "Failed to initialize ras recovery! (%d)\n", ret);
3693 
3694 	/*
3695 	 * Except error threshold exceeding case, other failure cases in this
3696 	 * function would not fail amdgpu driver init.
3697 	 */
3698 	if (!amdgpu_ras_is_rma(adev))
3699 		ret = 0;
3700 	else
3701 		ret = -EINVAL;
3702 
3703 	return ret;
3704 }
3705 
3706 static int amdgpu_ras_recovery_fini(struct amdgpu_device *adev)
3707 {
3708 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3709 	struct ras_err_handler_data *data = con->eh_data;
3710 	int max_flush_timeout = MAX_FLUSH_RETIRE_DWORK_TIMES;
3711 	bool ret;
3712 
3713 	/* recovery_init failed to init it, fini is useless */
3714 	if (!data)
3715 		return 0;
3716 
3717 	/* Save all cached bad pages to eeprom */
3718 	do {
3719 		flush_delayed_work(&con->page_retirement_dwork);
3720 		ret = amdgpu_ras_schedule_retirement_dwork(con, 0);
3721 	} while (ret && max_flush_timeout--);
3722 
3723 	if (con->page_retirement_thread)
3724 		kthread_stop(con->page_retirement_thread);
3725 
3726 	atomic_set(&con->page_retirement_req_cnt, 0);
3727 	atomic_set(&con->poison_creation_count, 0);
3728 
3729 	mutex_destroy(&con->page_rsv_lock);
3730 
3731 	cancel_work_sync(&con->recovery_work);
3732 
3733 	cancel_delayed_work_sync(&con->page_retirement_dwork);
3734 
3735 	amdgpu_ras_ecc_log_fini(&con->umc_ecc_log);
3736 
3737 	mutex_lock(&con->recovery_lock);
3738 	con->eh_data = NULL;
3739 	kfree(data->bps);
3740 	kfree(data);
3741 	mutex_unlock(&con->recovery_lock);
3742 
3743 	amdgpu_ras_critical_region_init(adev);
3744 
3745 	return 0;
3746 }
3747 /* recovery end */
3748 
3749 static bool amdgpu_ras_asic_supported(struct amdgpu_device *adev)
3750 {
3751 	if (amdgpu_sriov_vf(adev)) {
3752 		switch (amdgpu_ip_version(adev, MP0_HWIP, 0)) {
3753 		case IP_VERSION(13, 0, 2):
3754 		case IP_VERSION(13, 0, 6):
3755 		case IP_VERSION(13, 0, 12):
3756 		case IP_VERSION(13, 0, 14):
3757 			return true;
3758 		default:
3759 			return false;
3760 		}
3761 	}
3762 
3763 	if (adev->asic_type == CHIP_IP_DISCOVERY) {
3764 		switch (amdgpu_ip_version(adev, MP0_HWIP, 0)) {
3765 		case IP_VERSION(13, 0, 0):
3766 		case IP_VERSION(13, 0, 6):
3767 		case IP_VERSION(13, 0, 10):
3768 		case IP_VERSION(13, 0, 12):
3769 		case IP_VERSION(13, 0, 14):
3770 		case IP_VERSION(14, 0, 3):
3771 			return true;
3772 		default:
3773 			return false;
3774 		}
3775 	}
3776 
3777 	return adev->asic_type == CHIP_VEGA10 ||
3778 		adev->asic_type == CHIP_VEGA20 ||
3779 		adev->asic_type == CHIP_ARCTURUS ||
3780 		adev->asic_type == CHIP_ALDEBARAN ||
3781 		adev->asic_type == CHIP_SIENNA_CICHLID;
3782 }
3783 
3784 /*
3785  * this is workaround for vega20 workstation sku,
3786  * force enable gfx ras, ignore vbios gfx ras flag
3787  * due to GC EDC can not write
3788  */
3789 static void amdgpu_ras_get_quirks(struct amdgpu_device *adev)
3790 {
3791 	struct atom_context *ctx = adev->mode_info.atom_context;
3792 
3793 	if (!ctx)
3794 		return;
3795 
3796 	if (strnstr(ctx->vbios_pn, "D16406",
3797 		    sizeof(ctx->vbios_pn)) ||
3798 		strnstr(ctx->vbios_pn, "D36002",
3799 			sizeof(ctx->vbios_pn)))
3800 		adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX);
3801 }
3802 
3803 /* Query ras capablity via atomfirmware interface */
3804 static void amdgpu_ras_query_ras_capablity_from_vbios(struct amdgpu_device *adev)
3805 {
3806 	/* mem_ecc cap */
3807 	if (amdgpu_atomfirmware_mem_ecc_supported(adev)) {
3808 		dev_info(adev->dev, "MEM ECC is active.\n");
3809 		adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__UMC |
3810 					 1 << AMDGPU_RAS_BLOCK__DF);
3811 	} else {
3812 		dev_info(adev->dev, "MEM ECC is not presented.\n");
3813 	}
3814 
3815 	/* sram_ecc cap */
3816 	if (amdgpu_atomfirmware_sram_ecc_supported(adev)) {
3817 		dev_info(adev->dev, "SRAM ECC is active.\n");
3818 		if (!amdgpu_sriov_vf(adev))
3819 			adev->ras_hw_enabled |= ~(1 << AMDGPU_RAS_BLOCK__UMC |
3820 						  1 << AMDGPU_RAS_BLOCK__DF);
3821 		else
3822 			adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__PCIE_BIF |
3823 						 1 << AMDGPU_RAS_BLOCK__SDMA |
3824 						 1 << AMDGPU_RAS_BLOCK__GFX);
3825 
3826 		/*
3827 		 * VCN/JPEG RAS can be supported on both bare metal and
3828 		 * SRIOV environment
3829 		 */
3830 		if (amdgpu_ip_version(adev, VCN_HWIP, 0) == IP_VERSION(2, 6, 0) ||
3831 		    amdgpu_ip_version(adev, VCN_HWIP, 0) == IP_VERSION(4, 0, 0) ||
3832 		    amdgpu_ip_version(adev, VCN_HWIP, 0) == IP_VERSION(4, 0, 3) ||
3833 		    amdgpu_ip_version(adev, VCN_HWIP, 0) == IP_VERSION(5, 0, 1))
3834 			adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__VCN |
3835 						 1 << AMDGPU_RAS_BLOCK__JPEG);
3836 		else
3837 			adev->ras_hw_enabled &= ~(1 << AMDGPU_RAS_BLOCK__VCN |
3838 						  1 << AMDGPU_RAS_BLOCK__JPEG);
3839 
3840 		/*
3841 		 * XGMI RAS is not supported if xgmi num physical nodes
3842 		 * is zero
3843 		 */
3844 		if (!adev->gmc.xgmi.num_physical_nodes)
3845 			adev->ras_hw_enabled &= ~(1 << AMDGPU_RAS_BLOCK__XGMI_WAFL);
3846 	} else {
3847 		dev_info(adev->dev, "SRAM ECC is not presented.\n");
3848 	}
3849 }
3850 
3851 /* Query poison mode from umc/df IP callbacks */
3852 static void amdgpu_ras_query_poison_mode(struct amdgpu_device *adev)
3853 {
3854 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3855 	bool df_poison, umc_poison;
3856 
3857 	/* poison setting is useless on SRIOV guest */
3858 	if (amdgpu_sriov_vf(adev) || !con)
3859 		return;
3860 
3861 	/* Init poison supported flag, the default value is false */
3862 	if (adev->gmc.xgmi.connected_to_cpu ||
3863 	    adev->gmc.is_app_apu) {
3864 		/* enabled by default when GPU is connected to CPU */
3865 		con->poison_supported = true;
3866 	} else if (adev->df.funcs &&
3867 	    adev->df.funcs->query_ras_poison_mode &&
3868 	    adev->umc.ras &&
3869 	    adev->umc.ras->query_ras_poison_mode) {
3870 		df_poison =
3871 			adev->df.funcs->query_ras_poison_mode(adev);
3872 		umc_poison =
3873 			adev->umc.ras->query_ras_poison_mode(adev);
3874 
3875 		/* Only poison is set in both DF and UMC, we can support it */
3876 		if (df_poison && umc_poison)
3877 			con->poison_supported = true;
3878 		else if (df_poison != umc_poison)
3879 			dev_warn(adev->dev,
3880 				"Poison setting is inconsistent in DF/UMC(%d:%d)!\n",
3881 				df_poison, umc_poison);
3882 	}
3883 }
3884 
3885 /*
3886  * check hardware's ras ability which will be saved in hw_supported.
3887  * if hardware does not support ras, we can skip some ras initializtion and
3888  * forbid some ras operations from IP.
3889  * if software itself, say boot parameter, limit the ras ability. We still
3890  * need allow IP do some limited operations, like disable. In such case,
3891  * we have to initialize ras as normal. but need check if operation is
3892  * allowed or not in each function.
3893  */
3894 static void amdgpu_ras_check_supported(struct amdgpu_device *adev)
3895 {
3896 	adev->ras_hw_enabled = adev->ras_enabled = 0;
3897 
3898 	if (!amdgpu_ras_asic_supported(adev))
3899 		return;
3900 
3901 	if (amdgpu_sriov_vf(adev)) {
3902 		if (amdgpu_virt_get_ras_capability(adev))
3903 			goto init_ras_enabled_flag;
3904 	}
3905 
3906 	/* query ras capability from psp */
3907 	if (amdgpu_psp_get_ras_capability(&adev->psp))
3908 		goto init_ras_enabled_flag;
3909 
3910 	/* query ras capablity from bios */
3911 	if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
3912 		amdgpu_ras_query_ras_capablity_from_vbios(adev);
3913 	} else {
3914 		/* driver only manages a few IP blocks RAS feature
3915 		 * when GPU is connected cpu through XGMI */
3916 		adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX |
3917 					   1 << AMDGPU_RAS_BLOCK__SDMA |
3918 					   1 << AMDGPU_RAS_BLOCK__MMHUB);
3919 	}
3920 
3921 	/* apply asic specific settings (vega20 only for now) */
3922 	amdgpu_ras_get_quirks(adev);
3923 
3924 	/* query poison mode from umc/df ip callback */
3925 	amdgpu_ras_query_poison_mode(adev);
3926 
3927 init_ras_enabled_flag:
3928 	/* hw_supported needs to be aligned with RAS block mask. */
3929 	adev->ras_hw_enabled &= AMDGPU_RAS_BLOCK_MASK;
3930 
3931 	adev->ras_enabled = amdgpu_ras_enable == 0 ? 0 :
3932 		adev->ras_hw_enabled & amdgpu_ras_mask;
3933 
3934 	/* aca is disabled by default except for psp v13_0_6/v13_0_12/v13_0_14 */
3935 	if (!amdgpu_sriov_vf(adev)) {
3936 		adev->aca.is_enabled =
3937 			(amdgpu_ip_version(adev, MP0_HWIP, 0) == IP_VERSION(13, 0, 6) ||
3938 			amdgpu_ip_version(adev, MP0_HWIP, 0) == IP_VERSION(13, 0, 12) ||
3939 			amdgpu_ip_version(adev, MP0_HWIP, 0) == IP_VERSION(13, 0, 14));
3940 	}
3941 
3942 	/* bad page feature is not applicable to specific app platform */
3943 	if (adev->gmc.is_app_apu &&
3944 	    amdgpu_ip_version(adev, UMC_HWIP, 0) == IP_VERSION(12, 0, 0))
3945 		amdgpu_bad_page_threshold = 0;
3946 }
3947 
3948 static void amdgpu_ras_counte_dw(struct work_struct *work)
3949 {
3950 	struct amdgpu_ras *con = container_of(work, struct amdgpu_ras,
3951 					      ras_counte_delay_work.work);
3952 	struct amdgpu_device *adev = con->adev;
3953 	struct drm_device *dev = adev_to_drm(adev);
3954 	unsigned long ce_count, ue_count;
3955 	int res;
3956 
3957 	res = pm_runtime_get_sync(dev->dev);
3958 	if (res < 0)
3959 		goto Out;
3960 
3961 	/* Cache new values.
3962 	 */
3963 	if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count, NULL) == 0) {
3964 		atomic_set(&con->ras_ce_count, ce_count);
3965 		atomic_set(&con->ras_ue_count, ue_count);
3966 	}
3967 
3968 	pm_runtime_mark_last_busy(dev->dev);
3969 Out:
3970 	pm_runtime_put_autosuspend(dev->dev);
3971 }
3972 
3973 static int amdgpu_get_ras_schema(struct amdgpu_device *adev)
3974 {
3975 	return  amdgpu_ras_is_poison_mode_supported(adev) ? AMDGPU_RAS_ERROR__POISON : 0 |
3976 			AMDGPU_RAS_ERROR__SINGLE_CORRECTABLE |
3977 			AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE |
3978 			AMDGPU_RAS_ERROR__PARITY;
3979 }
3980 
3981 static void ras_event_mgr_init(struct ras_event_manager *mgr)
3982 {
3983 	struct ras_event_state *event_state;
3984 	int i;
3985 
3986 	memset(mgr, 0, sizeof(*mgr));
3987 	atomic64_set(&mgr->seqno, 0);
3988 
3989 	for (i = 0; i < ARRAY_SIZE(mgr->event_state); i++) {
3990 		event_state = &mgr->event_state[i];
3991 		event_state->last_seqno = RAS_EVENT_INVALID_ID;
3992 		atomic64_set(&event_state->count, 0);
3993 	}
3994 }
3995 
3996 static void amdgpu_ras_event_mgr_init(struct amdgpu_device *adev)
3997 {
3998 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
3999 	struct amdgpu_hive_info *hive;
4000 
4001 	if (!ras)
4002 		return;
4003 
4004 	hive = amdgpu_get_xgmi_hive(adev);
4005 	ras->event_mgr = hive ? &hive->event_mgr : &ras->__event_mgr;
4006 
4007 	/* init event manager with node 0 on xgmi system */
4008 	if (!amdgpu_reset_in_recovery(adev)) {
4009 		if (!hive || adev->gmc.xgmi.node_id == 0)
4010 			ras_event_mgr_init(ras->event_mgr);
4011 	}
4012 
4013 	if (hive)
4014 		amdgpu_put_xgmi_hive(hive);
4015 }
4016 
4017 static void amdgpu_ras_init_reserved_vram_size(struct amdgpu_device *adev)
4018 {
4019 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4020 
4021 	if (!con || (adev->flags & AMD_IS_APU))
4022 		return;
4023 
4024 	switch (amdgpu_ip_version(adev, MP0_HWIP, 0)) {
4025 	case IP_VERSION(13, 0, 2):
4026 	case IP_VERSION(13, 0, 6):
4027 	case IP_VERSION(13, 0, 12):
4028 		con->reserved_pages_in_bytes = AMDGPU_RAS_RESERVED_VRAM_SIZE_DEFAULT;
4029 		break;
4030 	case IP_VERSION(13, 0, 14):
4031 		con->reserved_pages_in_bytes = (AMDGPU_RAS_RESERVED_VRAM_SIZE_DEFAULT << 1);
4032 		break;
4033 	default:
4034 		break;
4035 	}
4036 }
4037 
4038 int amdgpu_ras_init(struct amdgpu_device *adev)
4039 {
4040 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4041 	int r;
4042 
4043 	if (con)
4044 		return 0;
4045 
4046 	con = kzalloc(sizeof(*con) +
4047 			sizeof(struct ras_manager) * AMDGPU_RAS_BLOCK_COUNT +
4048 			sizeof(struct ras_manager) * AMDGPU_RAS_MCA_BLOCK_COUNT,
4049 			GFP_KERNEL);
4050 	if (!con)
4051 		return -ENOMEM;
4052 
4053 	con->adev = adev;
4054 	INIT_DELAYED_WORK(&con->ras_counte_delay_work, amdgpu_ras_counte_dw);
4055 	atomic_set(&con->ras_ce_count, 0);
4056 	atomic_set(&con->ras_ue_count, 0);
4057 
4058 	con->objs = (struct ras_manager *)(con + 1);
4059 
4060 	amdgpu_ras_set_context(adev, con);
4061 
4062 	amdgpu_ras_check_supported(adev);
4063 
4064 	if (!adev->ras_enabled || adev->asic_type == CHIP_VEGA10) {
4065 		/* set gfx block ras context feature for VEGA20 Gaming
4066 		 * send ras disable cmd to ras ta during ras late init.
4067 		 */
4068 		if (!adev->ras_enabled && adev->asic_type == CHIP_VEGA20) {
4069 			con->features |= BIT(AMDGPU_RAS_BLOCK__GFX);
4070 
4071 			return 0;
4072 		}
4073 
4074 		r = 0;
4075 		goto release_con;
4076 	}
4077 
4078 	con->update_channel_flag = false;
4079 	con->features = 0;
4080 	con->schema = 0;
4081 	INIT_LIST_HEAD(&con->head);
4082 	/* Might need get this flag from vbios. */
4083 	con->flags = RAS_DEFAULT_FLAGS;
4084 
4085 	/* initialize nbio ras function ahead of any other
4086 	 * ras functions so hardware fatal error interrupt
4087 	 * can be enabled as early as possible */
4088 	switch (amdgpu_ip_version(adev, NBIO_HWIP, 0)) {
4089 	case IP_VERSION(7, 4, 0):
4090 	case IP_VERSION(7, 4, 1):
4091 	case IP_VERSION(7, 4, 4):
4092 		if (!adev->gmc.xgmi.connected_to_cpu)
4093 			adev->nbio.ras = &nbio_v7_4_ras;
4094 		break;
4095 	case IP_VERSION(4, 3, 0):
4096 		if (adev->ras_hw_enabled & (1 << AMDGPU_RAS_BLOCK__DF))
4097 			/* unlike other generation of nbio ras,
4098 			 * nbio v4_3 only support fatal error interrupt
4099 			 * to inform software that DF is freezed due to
4100 			 * system fatal error event. driver should not
4101 			 * enable nbio ras in such case. Instead,
4102 			 * check DF RAS */
4103 			adev->nbio.ras = &nbio_v4_3_ras;
4104 		break;
4105 	case IP_VERSION(6, 3, 1):
4106 		if (adev->ras_hw_enabled & (1 << AMDGPU_RAS_BLOCK__DF))
4107 			/* unlike other generation of nbio ras,
4108 			 * nbif v6_3_1 only support fatal error interrupt
4109 			 * to inform software that DF is freezed due to
4110 			 * system fatal error event. driver should not
4111 			 * enable nbio ras in such case. Instead,
4112 			 * check DF RAS
4113 			 */
4114 			adev->nbio.ras = &nbif_v6_3_1_ras;
4115 		break;
4116 	case IP_VERSION(7, 9, 0):
4117 	case IP_VERSION(7, 9, 1):
4118 		if (!adev->gmc.is_app_apu)
4119 			adev->nbio.ras = &nbio_v7_9_ras;
4120 		break;
4121 	default:
4122 		/* nbio ras is not available */
4123 		break;
4124 	}
4125 
4126 	/* nbio ras block needs to be enabled ahead of other ras blocks
4127 	 * to handle fatal error */
4128 	r = amdgpu_nbio_ras_sw_init(adev);
4129 	if (r)
4130 		return r;
4131 
4132 	if (adev->nbio.ras &&
4133 	    adev->nbio.ras->init_ras_controller_interrupt) {
4134 		r = adev->nbio.ras->init_ras_controller_interrupt(adev);
4135 		if (r)
4136 			goto release_con;
4137 	}
4138 
4139 	if (adev->nbio.ras &&
4140 	    adev->nbio.ras->init_ras_err_event_athub_interrupt) {
4141 		r = adev->nbio.ras->init_ras_err_event_athub_interrupt(adev);
4142 		if (r)
4143 			goto release_con;
4144 	}
4145 
4146 	/* Packed socket_id to ras feature mask bits[31:29] */
4147 	if (adev->smuio.funcs &&
4148 	    adev->smuio.funcs->get_socket_id)
4149 		con->features |= ((adev->smuio.funcs->get_socket_id(adev)) <<
4150 					AMDGPU_RAS_FEATURES_SOCKETID_SHIFT);
4151 
4152 	/* Get RAS schema for particular SOC */
4153 	con->schema = amdgpu_get_ras_schema(adev);
4154 
4155 	amdgpu_ras_init_reserved_vram_size(adev);
4156 
4157 	if (amdgpu_ras_fs_init(adev)) {
4158 		r = -EINVAL;
4159 		goto release_con;
4160 	}
4161 
4162 	if (amdgpu_ras_aca_is_supported(adev)) {
4163 		if (amdgpu_aca_is_enabled(adev))
4164 			r = amdgpu_aca_init(adev);
4165 		else
4166 			r = amdgpu_mca_init(adev);
4167 		if (r)
4168 			goto release_con;
4169 	}
4170 
4171 	con->init_task_pid = task_pid_nr(current);
4172 	get_task_comm(con->init_task_comm, current);
4173 
4174 	mutex_init(&con->critical_region_lock);
4175 	INIT_LIST_HEAD(&con->critical_region_head);
4176 
4177 	dev_info(adev->dev, "RAS INFO: ras initialized successfully, "
4178 		 "hardware ability[%x] ras_mask[%x]\n",
4179 		 adev->ras_hw_enabled, adev->ras_enabled);
4180 
4181 	return 0;
4182 release_con:
4183 	amdgpu_ras_set_context(adev, NULL);
4184 	kfree(con);
4185 
4186 	return r;
4187 }
4188 
4189 int amdgpu_persistent_edc_harvesting_supported(struct amdgpu_device *adev)
4190 {
4191 	if (adev->gmc.xgmi.connected_to_cpu ||
4192 	    adev->gmc.is_app_apu)
4193 		return 1;
4194 	return 0;
4195 }
4196 
4197 static int amdgpu_persistent_edc_harvesting(struct amdgpu_device *adev,
4198 					struct ras_common_if *ras_block)
4199 {
4200 	struct ras_query_if info = {
4201 		.head = *ras_block,
4202 	};
4203 
4204 	if (!amdgpu_persistent_edc_harvesting_supported(adev))
4205 		return 0;
4206 
4207 	if (amdgpu_ras_query_error_status(adev, &info) != 0)
4208 		DRM_WARN("RAS init harvest failure");
4209 
4210 	if (amdgpu_ras_reset_error_status(adev, ras_block->block) != 0)
4211 		DRM_WARN("RAS init harvest reset failure");
4212 
4213 	return 0;
4214 }
4215 
4216 bool amdgpu_ras_is_poison_mode_supported(struct amdgpu_device *adev)
4217 {
4218        struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4219 
4220        if (!con)
4221                return false;
4222 
4223        return con->poison_supported;
4224 }
4225 
4226 /* helper function to handle common stuff in ip late init phase */
4227 int amdgpu_ras_block_late_init(struct amdgpu_device *adev,
4228 			 struct ras_common_if *ras_block)
4229 {
4230 	struct amdgpu_ras_block_object *ras_obj = NULL;
4231 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4232 	struct ras_query_if *query_info;
4233 	unsigned long ue_count, ce_count;
4234 	int r;
4235 
4236 	/* disable RAS feature per IP block if it is not supported */
4237 	if (!amdgpu_ras_is_supported(adev, ras_block->block)) {
4238 		amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
4239 		return 0;
4240 	}
4241 
4242 	r = amdgpu_ras_feature_enable_on_boot(adev, ras_block, 1);
4243 	if (r) {
4244 		if (adev->in_suspend || amdgpu_reset_in_recovery(adev)) {
4245 			/* in resume phase, if fail to enable ras,
4246 			 * clean up all ras fs nodes, and disable ras */
4247 			goto cleanup;
4248 		} else
4249 			return r;
4250 	}
4251 
4252 	/* check for errors on warm reset edc persisant supported ASIC */
4253 	amdgpu_persistent_edc_harvesting(adev, ras_block);
4254 
4255 	/* in resume phase, no need to create ras fs node */
4256 	if (adev->in_suspend || amdgpu_reset_in_recovery(adev))
4257 		return 0;
4258 
4259 	ras_obj = container_of(ras_block, struct amdgpu_ras_block_object, ras_comm);
4260 	if (ras_obj->ras_cb || (ras_obj->hw_ops &&
4261 	    (ras_obj->hw_ops->query_poison_status ||
4262 	    ras_obj->hw_ops->handle_poison_consumption))) {
4263 		r = amdgpu_ras_interrupt_add_handler(adev, ras_block);
4264 		if (r)
4265 			goto cleanup;
4266 	}
4267 
4268 	if (ras_obj->hw_ops &&
4269 	    (ras_obj->hw_ops->query_ras_error_count ||
4270 	     ras_obj->hw_ops->query_ras_error_status)) {
4271 		r = amdgpu_ras_sysfs_create(adev, ras_block);
4272 		if (r)
4273 			goto interrupt;
4274 
4275 		/* Those are the cached values at init.
4276 		 */
4277 		query_info = kzalloc(sizeof(*query_info), GFP_KERNEL);
4278 		if (!query_info)
4279 			return -ENOMEM;
4280 		memcpy(&query_info->head, ras_block, sizeof(struct ras_common_if));
4281 
4282 		if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count, query_info) == 0) {
4283 			atomic_set(&con->ras_ce_count, ce_count);
4284 			atomic_set(&con->ras_ue_count, ue_count);
4285 		}
4286 
4287 		kfree(query_info);
4288 	}
4289 
4290 	return 0;
4291 
4292 interrupt:
4293 	if (ras_obj->ras_cb)
4294 		amdgpu_ras_interrupt_remove_handler(adev, ras_block);
4295 cleanup:
4296 	amdgpu_ras_feature_enable(adev, ras_block, 0);
4297 	return r;
4298 }
4299 
4300 static int amdgpu_ras_block_late_init_default(struct amdgpu_device *adev,
4301 			 struct ras_common_if *ras_block)
4302 {
4303 	return amdgpu_ras_block_late_init(adev, ras_block);
4304 }
4305 
4306 /* helper function to remove ras fs node and interrupt handler */
4307 void amdgpu_ras_block_late_fini(struct amdgpu_device *adev,
4308 			  struct ras_common_if *ras_block)
4309 {
4310 	struct amdgpu_ras_block_object *ras_obj;
4311 	if (!ras_block)
4312 		return;
4313 
4314 	amdgpu_ras_sysfs_remove(adev, ras_block);
4315 
4316 	ras_obj = container_of(ras_block, struct amdgpu_ras_block_object, ras_comm);
4317 	if (ras_obj->ras_cb)
4318 		amdgpu_ras_interrupt_remove_handler(adev, ras_block);
4319 }
4320 
4321 static void amdgpu_ras_block_late_fini_default(struct amdgpu_device *adev,
4322 			  struct ras_common_if *ras_block)
4323 {
4324 	return amdgpu_ras_block_late_fini(adev, ras_block);
4325 }
4326 
4327 /* do some init work after IP late init as dependence.
4328  * and it runs in resume/gpu reset/booting up cases.
4329  */
4330 void amdgpu_ras_resume(struct amdgpu_device *adev)
4331 {
4332 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4333 	struct ras_manager *obj, *tmp;
4334 
4335 	if (!adev->ras_enabled || !con) {
4336 		/* clean ras context for VEGA20 Gaming after send ras disable cmd */
4337 		amdgpu_release_ras_context(adev);
4338 
4339 		return;
4340 	}
4341 
4342 	if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
4343 		/* Set up all other IPs which are not implemented. There is a
4344 		 * tricky thing that IP's actual ras error type should be
4345 		 * MULTI_UNCORRECTABLE, but as driver does not handle it, so
4346 		 * ERROR_NONE make sense anyway.
4347 		 */
4348 		amdgpu_ras_enable_all_features(adev, 1);
4349 
4350 		/* We enable ras on all hw_supported block, but as boot
4351 		 * parameter might disable some of them and one or more IP has
4352 		 * not implemented yet. So we disable them on behalf.
4353 		 */
4354 		list_for_each_entry_safe(obj, tmp, &con->head, node) {
4355 			if (!amdgpu_ras_is_supported(adev, obj->head.block)) {
4356 				amdgpu_ras_feature_enable(adev, &obj->head, 0);
4357 				/* there should be no any reference. */
4358 				WARN_ON(alive_obj(obj));
4359 			}
4360 		}
4361 	}
4362 }
4363 
4364 void amdgpu_ras_suspend(struct amdgpu_device *adev)
4365 {
4366 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4367 
4368 	if (!adev->ras_enabled || !con)
4369 		return;
4370 
4371 	amdgpu_ras_disable_all_features(adev, 0);
4372 	/* Make sure all ras objects are disabled. */
4373 	if (AMDGPU_RAS_GET_FEATURES(con->features))
4374 		amdgpu_ras_disable_all_features(adev, 1);
4375 }
4376 
4377 int amdgpu_ras_late_init(struct amdgpu_device *adev)
4378 {
4379 	struct amdgpu_ras_block_list *node, *tmp;
4380 	struct amdgpu_ras_block_object *obj;
4381 	int r;
4382 
4383 	amdgpu_ras_event_mgr_init(adev);
4384 
4385 	if (amdgpu_ras_aca_is_supported(adev)) {
4386 		if (amdgpu_reset_in_recovery(adev)) {
4387 			if (amdgpu_aca_is_enabled(adev))
4388 				r = amdgpu_aca_reset(adev);
4389 			else
4390 				r = amdgpu_mca_reset(adev);
4391 			if (r)
4392 				return r;
4393 		}
4394 
4395 		if (!amdgpu_sriov_vf(adev)) {
4396 			if (amdgpu_aca_is_enabled(adev))
4397 				amdgpu_ras_set_aca_debug_mode(adev, false);
4398 			else
4399 				amdgpu_ras_set_mca_debug_mode(adev, false);
4400 		}
4401 	}
4402 
4403 	/* Guest side doesn't need init ras feature */
4404 	if (amdgpu_sriov_vf(adev) && !amdgpu_sriov_ras_telemetry_en(adev))
4405 		return 0;
4406 
4407 	list_for_each_entry_safe(node, tmp, &adev->ras_list, node) {
4408 		obj = node->ras_obj;
4409 		if (!obj) {
4410 			dev_warn(adev->dev, "Warning: abnormal ras list node.\n");
4411 			continue;
4412 		}
4413 
4414 		if (!amdgpu_ras_is_supported(adev, obj->ras_comm.block))
4415 			continue;
4416 
4417 		if (obj->ras_late_init) {
4418 			r = obj->ras_late_init(adev, &obj->ras_comm);
4419 			if (r) {
4420 				dev_err(adev->dev, "%s failed to execute ras_late_init! ret:%d\n",
4421 					obj->ras_comm.name, r);
4422 				return r;
4423 			}
4424 		} else
4425 			amdgpu_ras_block_late_init_default(adev, &obj->ras_comm);
4426 	}
4427 
4428 	return 0;
4429 }
4430 
4431 /* do some fini work before IP fini as dependence */
4432 int amdgpu_ras_pre_fini(struct amdgpu_device *adev)
4433 {
4434 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4435 
4436 	if (!adev->ras_enabled || !con)
4437 		return 0;
4438 
4439 
4440 	/* Need disable ras on all IPs here before ip [hw/sw]fini */
4441 	if (AMDGPU_RAS_GET_FEATURES(con->features))
4442 		amdgpu_ras_disable_all_features(adev, 0);
4443 	amdgpu_ras_recovery_fini(adev);
4444 	return 0;
4445 }
4446 
4447 int amdgpu_ras_fini(struct amdgpu_device *adev)
4448 {
4449 	struct amdgpu_ras_block_list *ras_node, *tmp;
4450 	struct amdgpu_ras_block_object *obj = NULL;
4451 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4452 
4453 	if (!adev->ras_enabled || !con)
4454 		return 0;
4455 
4456 	amdgpu_ras_critical_region_fini(adev);
4457 	mutex_destroy(&con->critical_region_lock);
4458 
4459 	list_for_each_entry_safe(ras_node, tmp, &adev->ras_list, node) {
4460 		if (ras_node->ras_obj) {
4461 			obj = ras_node->ras_obj;
4462 			if (amdgpu_ras_is_supported(adev, obj->ras_comm.block) &&
4463 			    obj->ras_fini)
4464 				obj->ras_fini(adev, &obj->ras_comm);
4465 			else
4466 				amdgpu_ras_block_late_fini_default(adev, &obj->ras_comm);
4467 		}
4468 
4469 		/* Clear ras blocks from ras_list and free ras block list node */
4470 		list_del(&ras_node->node);
4471 		kfree(ras_node);
4472 	}
4473 
4474 	amdgpu_ras_fs_fini(adev);
4475 	amdgpu_ras_interrupt_remove_all(adev);
4476 
4477 	if (amdgpu_ras_aca_is_supported(adev)) {
4478 		if (amdgpu_aca_is_enabled(adev))
4479 			amdgpu_aca_fini(adev);
4480 		else
4481 			amdgpu_mca_fini(adev);
4482 	}
4483 
4484 	WARN(AMDGPU_RAS_GET_FEATURES(con->features), "Feature mask is not cleared");
4485 
4486 	if (AMDGPU_RAS_GET_FEATURES(con->features))
4487 		amdgpu_ras_disable_all_features(adev, 0);
4488 
4489 	cancel_delayed_work_sync(&con->ras_counte_delay_work);
4490 
4491 	amdgpu_ras_set_context(adev, NULL);
4492 	kfree(con);
4493 
4494 	return 0;
4495 }
4496 
4497 bool amdgpu_ras_get_fed_status(struct amdgpu_device *adev)
4498 {
4499 	struct amdgpu_ras *ras;
4500 
4501 	ras = amdgpu_ras_get_context(adev);
4502 	if (!ras)
4503 		return false;
4504 
4505 	return test_bit(AMDGPU_RAS_BLOCK__LAST, &ras->ras_err_state);
4506 }
4507 
4508 void amdgpu_ras_set_fed(struct amdgpu_device *adev, bool status)
4509 {
4510 	struct amdgpu_ras *ras;
4511 
4512 	ras = amdgpu_ras_get_context(adev);
4513 	if (ras) {
4514 		if (status)
4515 			set_bit(AMDGPU_RAS_BLOCK__LAST, &ras->ras_err_state);
4516 		else
4517 			clear_bit(AMDGPU_RAS_BLOCK__LAST, &ras->ras_err_state);
4518 	}
4519 }
4520 
4521 void amdgpu_ras_clear_err_state(struct amdgpu_device *adev)
4522 {
4523 	struct amdgpu_ras *ras;
4524 
4525 	ras = amdgpu_ras_get_context(adev);
4526 	if (ras) {
4527 		ras->ras_err_state = 0;
4528 		ras->gpu_reset_flags = 0;
4529 	}
4530 }
4531 
4532 void amdgpu_ras_set_err_poison(struct amdgpu_device *adev,
4533 			       enum amdgpu_ras_block block)
4534 {
4535 	struct amdgpu_ras *ras;
4536 
4537 	ras = amdgpu_ras_get_context(adev);
4538 	if (ras)
4539 		set_bit(block, &ras->ras_err_state);
4540 }
4541 
4542 bool amdgpu_ras_is_err_state(struct amdgpu_device *adev, int block)
4543 {
4544 	struct amdgpu_ras *ras;
4545 
4546 	ras = amdgpu_ras_get_context(adev);
4547 	if (ras) {
4548 		if (block == AMDGPU_RAS_BLOCK__ANY)
4549 			return (ras->ras_err_state != 0);
4550 		else
4551 			return test_bit(block, &ras->ras_err_state) ||
4552 			       test_bit(AMDGPU_RAS_BLOCK__LAST,
4553 					&ras->ras_err_state);
4554 	}
4555 
4556 	return false;
4557 }
4558 
4559 static struct ras_event_manager *__get_ras_event_mgr(struct amdgpu_device *adev)
4560 {
4561 	struct amdgpu_ras *ras;
4562 
4563 	ras = amdgpu_ras_get_context(adev);
4564 	if (!ras)
4565 		return NULL;
4566 
4567 	return ras->event_mgr;
4568 }
4569 
4570 int amdgpu_ras_mark_ras_event_caller(struct amdgpu_device *adev, enum ras_event_type type,
4571 				     const void *caller)
4572 {
4573 	struct ras_event_manager *event_mgr;
4574 	struct ras_event_state *event_state;
4575 	int ret = 0;
4576 
4577 	if (type >= RAS_EVENT_TYPE_COUNT) {
4578 		ret = -EINVAL;
4579 		goto out;
4580 	}
4581 
4582 	event_mgr = __get_ras_event_mgr(adev);
4583 	if (!event_mgr) {
4584 		ret = -EINVAL;
4585 		goto out;
4586 	}
4587 
4588 	event_state = &event_mgr->event_state[type];
4589 	event_state->last_seqno = atomic64_inc_return(&event_mgr->seqno);
4590 	atomic64_inc(&event_state->count);
4591 
4592 out:
4593 	if (ret && caller)
4594 		dev_warn(adev->dev, "failed mark ras event (%d) in %ps, ret:%d\n",
4595 			 (int)type, caller, ret);
4596 
4597 	return ret;
4598 }
4599 
4600 u64 amdgpu_ras_acquire_event_id(struct amdgpu_device *adev, enum ras_event_type type)
4601 {
4602 	struct ras_event_manager *event_mgr;
4603 	u64 id;
4604 
4605 	if (type >= RAS_EVENT_TYPE_COUNT)
4606 		return RAS_EVENT_INVALID_ID;
4607 
4608 	switch (type) {
4609 	case RAS_EVENT_TYPE_FATAL:
4610 	case RAS_EVENT_TYPE_POISON_CREATION:
4611 	case RAS_EVENT_TYPE_POISON_CONSUMPTION:
4612 		event_mgr = __get_ras_event_mgr(adev);
4613 		if (!event_mgr)
4614 			return RAS_EVENT_INVALID_ID;
4615 
4616 		id = event_mgr->event_state[type].last_seqno;
4617 		break;
4618 	case RAS_EVENT_TYPE_INVALID:
4619 	default:
4620 		id = RAS_EVENT_INVALID_ID;
4621 		break;
4622 	}
4623 
4624 	return id;
4625 }
4626 
4627 void amdgpu_ras_global_ras_isr(struct amdgpu_device *adev)
4628 {
4629 	if (atomic_cmpxchg(&amdgpu_ras_in_intr, 0, 1) == 0) {
4630 		struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
4631 		enum ras_event_type type = RAS_EVENT_TYPE_FATAL;
4632 		u64 event_id;
4633 
4634 		if (amdgpu_ras_mark_ras_event(adev, type)) {
4635 			dev_err(adev->dev,
4636 				"uncorrectable hardware error (ERREVENT_ATHUB_INTERRUPT) detected!\n");
4637 			return;
4638 		}
4639 
4640 		event_id = amdgpu_ras_acquire_event_id(adev, type);
4641 
4642 		RAS_EVENT_LOG(adev, event_id, "uncorrectable hardware error"
4643 			      "(ERREVENT_ATHUB_INTERRUPT) detected!\n");
4644 
4645 		amdgpu_ras_set_fed(adev, true);
4646 		ras->gpu_reset_flags |= AMDGPU_RAS_GPU_RESET_MODE1_RESET;
4647 		amdgpu_ras_reset_gpu(adev);
4648 	}
4649 }
4650 
4651 bool amdgpu_ras_need_emergency_restart(struct amdgpu_device *adev)
4652 {
4653 	if (adev->asic_type == CHIP_VEGA20 &&
4654 	    adev->pm.fw_version <= 0x283400) {
4655 		return !(amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) &&
4656 				amdgpu_ras_intr_triggered();
4657 	}
4658 
4659 	return false;
4660 }
4661 
4662 void amdgpu_release_ras_context(struct amdgpu_device *adev)
4663 {
4664 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4665 
4666 	if (!con)
4667 		return;
4668 
4669 	if (!adev->ras_enabled && con->features & BIT(AMDGPU_RAS_BLOCK__GFX)) {
4670 		con->features &= ~BIT(AMDGPU_RAS_BLOCK__GFX);
4671 		amdgpu_ras_set_context(adev, NULL);
4672 		kfree(con);
4673 	}
4674 }
4675 
4676 #ifdef CONFIG_X86_MCE_AMD
4677 static struct amdgpu_device *find_adev(uint32_t node_id)
4678 {
4679 	int i;
4680 	struct amdgpu_device *adev = NULL;
4681 
4682 	for (i = 0; i < mce_adev_list.num_gpu; i++) {
4683 		adev = mce_adev_list.devs[i];
4684 
4685 		if (adev && adev->gmc.xgmi.connected_to_cpu &&
4686 		    adev->gmc.xgmi.physical_node_id == node_id)
4687 			break;
4688 		adev = NULL;
4689 	}
4690 
4691 	return adev;
4692 }
4693 
4694 #define GET_MCA_IPID_GPUID(m)	(((m) >> 44) & 0xF)
4695 #define GET_UMC_INST(m)		(((m) >> 21) & 0x7)
4696 #define GET_CHAN_INDEX(m)	((((m) >> 12) & 0x3) | (((m) >> 18) & 0x4))
4697 #define GPU_ID_OFFSET		8
4698 
4699 static int amdgpu_bad_page_notifier(struct notifier_block *nb,
4700 				    unsigned long val, void *data)
4701 {
4702 	struct mce *m = (struct mce *)data;
4703 	struct amdgpu_device *adev = NULL;
4704 	uint32_t gpu_id = 0;
4705 	uint32_t umc_inst = 0, ch_inst = 0;
4706 
4707 	/*
4708 	 * If the error was generated in UMC_V2, which belongs to GPU UMCs,
4709 	 * and error occurred in DramECC (Extended error code = 0) then only
4710 	 * process the error, else bail out.
4711 	 */
4712 	if (!m || !((smca_get_bank_type(m->extcpu, m->bank) == SMCA_UMC_V2) &&
4713 		    (XEC(m->status, 0x3f) == 0x0)))
4714 		return NOTIFY_DONE;
4715 
4716 	/*
4717 	 * If it is correctable error, return.
4718 	 */
4719 	if (mce_is_correctable(m))
4720 		return NOTIFY_OK;
4721 
4722 	/*
4723 	 * GPU Id is offset by GPU_ID_OFFSET in MCA_IPID_UMC register.
4724 	 */
4725 	gpu_id = GET_MCA_IPID_GPUID(m->ipid) - GPU_ID_OFFSET;
4726 
4727 	adev = find_adev(gpu_id);
4728 	if (!adev) {
4729 		DRM_WARN("%s: Unable to find adev for gpu_id: %d\n", __func__,
4730 								gpu_id);
4731 		return NOTIFY_DONE;
4732 	}
4733 
4734 	/*
4735 	 * If it is uncorrectable error, then find out UMC instance and
4736 	 * channel index.
4737 	 */
4738 	umc_inst = GET_UMC_INST(m->ipid);
4739 	ch_inst = GET_CHAN_INDEX(m->ipid);
4740 
4741 	dev_info(adev->dev, "Uncorrectable error detected in UMC inst: %d, chan_idx: %d",
4742 			     umc_inst, ch_inst);
4743 
4744 	if (!amdgpu_umc_page_retirement_mca(adev, m->addr, ch_inst, umc_inst))
4745 		return NOTIFY_OK;
4746 	else
4747 		return NOTIFY_DONE;
4748 }
4749 
4750 static struct notifier_block amdgpu_bad_page_nb = {
4751 	.notifier_call  = amdgpu_bad_page_notifier,
4752 	.priority       = MCE_PRIO_UC,
4753 };
4754 
4755 static void amdgpu_register_bad_pages_mca_notifier(struct amdgpu_device *adev)
4756 {
4757 	/*
4758 	 * Add the adev to the mce_adev_list.
4759 	 * During mode2 reset, amdgpu device is temporarily
4760 	 * removed from the mgpu_info list which can cause
4761 	 * page retirement to fail.
4762 	 * Use this list instead of mgpu_info to find the amdgpu
4763 	 * device on which the UMC error was reported.
4764 	 */
4765 	mce_adev_list.devs[mce_adev_list.num_gpu++] = adev;
4766 
4767 	/*
4768 	 * Register the x86 notifier only once
4769 	 * with MCE subsystem.
4770 	 */
4771 	if (notifier_registered == false) {
4772 		mce_register_decode_chain(&amdgpu_bad_page_nb);
4773 		notifier_registered = true;
4774 	}
4775 }
4776 #endif
4777 
4778 struct amdgpu_ras *amdgpu_ras_get_context(struct amdgpu_device *adev)
4779 {
4780 	if (!adev)
4781 		return NULL;
4782 
4783 	return adev->psp.ras_context.ras;
4784 }
4785 
4786 int amdgpu_ras_set_context(struct amdgpu_device *adev, struct amdgpu_ras *ras_con)
4787 {
4788 	if (!adev)
4789 		return -EINVAL;
4790 
4791 	adev->psp.ras_context.ras = ras_con;
4792 	return 0;
4793 }
4794 
4795 /* check if ras is supported on block, say, sdma, gfx */
4796 int amdgpu_ras_is_supported(struct amdgpu_device *adev,
4797 		unsigned int block)
4798 {
4799 	int ret = 0;
4800 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
4801 
4802 	if (block >= AMDGPU_RAS_BLOCK_COUNT)
4803 		return 0;
4804 
4805 	ret = ras && (adev->ras_enabled & (1 << block));
4806 
4807 	/* For the special asic with mem ecc enabled but sram ecc
4808 	 * not enabled, even if the ras block is not supported on
4809 	 * .ras_enabled, if the asic supports poison mode and the
4810 	 * ras block has ras configuration, it can be considered
4811 	 * that the ras block supports ras function.
4812 	 */
4813 	if (!ret &&
4814 	    (block == AMDGPU_RAS_BLOCK__GFX ||
4815 	     block == AMDGPU_RAS_BLOCK__SDMA ||
4816 	     block == AMDGPU_RAS_BLOCK__VCN ||
4817 	     block == AMDGPU_RAS_BLOCK__JPEG) &&
4818 		(amdgpu_ras_mask & (1 << block)) &&
4819 	    amdgpu_ras_is_poison_mode_supported(adev) &&
4820 	    amdgpu_ras_get_ras_block(adev, block, 0))
4821 		ret = 1;
4822 
4823 	return ret;
4824 }
4825 
4826 int amdgpu_ras_reset_gpu(struct amdgpu_device *adev)
4827 {
4828 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
4829 
4830 	/* mode1 is the only selection for RMA status */
4831 	if (amdgpu_ras_is_rma(adev)) {
4832 		ras->gpu_reset_flags = 0;
4833 		ras->gpu_reset_flags |= AMDGPU_RAS_GPU_RESET_MODE1_RESET;
4834 	}
4835 
4836 	if (atomic_cmpxchg(&ras->in_recovery, 0, 1) == 0) {
4837 		struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
4838 		int hive_ras_recovery = 0;
4839 
4840 		if (hive) {
4841 			hive_ras_recovery = atomic_read(&hive->ras_recovery);
4842 			amdgpu_put_xgmi_hive(hive);
4843 		}
4844 		/* In the case of multiple GPUs, after a GPU has started
4845 		 * resetting all GPUs on hive, other GPUs do not need to
4846 		 * trigger GPU reset again.
4847 		 */
4848 		if (!hive_ras_recovery)
4849 			amdgpu_reset_domain_schedule(ras->adev->reset_domain, &ras->recovery_work);
4850 		else
4851 			atomic_set(&ras->in_recovery, 0);
4852 	} else {
4853 		flush_work(&ras->recovery_work);
4854 		amdgpu_reset_domain_schedule(ras->adev->reset_domain, &ras->recovery_work);
4855 	}
4856 
4857 	return 0;
4858 }
4859 
4860 int amdgpu_ras_set_mca_debug_mode(struct amdgpu_device *adev, bool enable)
4861 {
4862 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4863 	int ret = 0;
4864 
4865 	if (con) {
4866 		ret = amdgpu_mca_smu_set_debug_mode(adev, enable);
4867 		if (!ret)
4868 			con->is_aca_debug_mode = enable;
4869 	}
4870 
4871 	return ret;
4872 }
4873 
4874 int amdgpu_ras_set_aca_debug_mode(struct amdgpu_device *adev, bool enable)
4875 {
4876 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4877 	int ret = 0;
4878 
4879 	if (con) {
4880 		if (amdgpu_aca_is_enabled(adev))
4881 			ret = amdgpu_aca_smu_set_debug_mode(adev, enable);
4882 		else
4883 			ret = amdgpu_mca_smu_set_debug_mode(adev, enable);
4884 		if (!ret)
4885 			con->is_aca_debug_mode = enable;
4886 	}
4887 
4888 	return ret;
4889 }
4890 
4891 bool amdgpu_ras_get_aca_debug_mode(struct amdgpu_device *adev)
4892 {
4893 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4894 	const struct aca_smu_funcs *smu_funcs = adev->aca.smu_funcs;
4895 	const struct amdgpu_mca_smu_funcs *mca_funcs = adev->mca.mca_funcs;
4896 
4897 	if (!con)
4898 		return false;
4899 
4900 	if ((amdgpu_aca_is_enabled(adev) && smu_funcs && smu_funcs->set_debug_mode) ||
4901 	    (!amdgpu_aca_is_enabled(adev) && mca_funcs && mca_funcs->mca_set_debug_mode))
4902 		return con->is_aca_debug_mode;
4903 	else
4904 		return true;
4905 }
4906 
4907 bool amdgpu_ras_get_error_query_mode(struct amdgpu_device *adev,
4908 				     unsigned int *error_query_mode)
4909 {
4910 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4911 	const struct amdgpu_mca_smu_funcs *mca_funcs = adev->mca.mca_funcs;
4912 	const struct aca_smu_funcs *smu_funcs = adev->aca.smu_funcs;
4913 
4914 	if (!con) {
4915 		*error_query_mode = AMDGPU_RAS_INVALID_ERROR_QUERY;
4916 		return false;
4917 	}
4918 
4919 	if (amdgpu_sriov_vf(adev)) {
4920 		*error_query_mode = AMDGPU_RAS_VIRT_ERROR_COUNT_QUERY;
4921 	} else if ((smu_funcs && smu_funcs->set_debug_mode) || (mca_funcs && mca_funcs->mca_set_debug_mode)) {
4922 		*error_query_mode =
4923 			(con->is_aca_debug_mode) ? AMDGPU_RAS_DIRECT_ERROR_QUERY : AMDGPU_RAS_FIRMWARE_ERROR_QUERY;
4924 	} else {
4925 		*error_query_mode = AMDGPU_RAS_DIRECT_ERROR_QUERY;
4926 	}
4927 
4928 	return true;
4929 }
4930 
4931 /* Register each ip ras block into amdgpu ras */
4932 int amdgpu_ras_register_ras_block(struct amdgpu_device *adev,
4933 		struct amdgpu_ras_block_object *ras_block_obj)
4934 {
4935 	struct amdgpu_ras_block_list *ras_node;
4936 	if (!adev || !ras_block_obj)
4937 		return -EINVAL;
4938 
4939 	ras_node = kzalloc(sizeof(*ras_node), GFP_KERNEL);
4940 	if (!ras_node)
4941 		return -ENOMEM;
4942 
4943 	INIT_LIST_HEAD(&ras_node->node);
4944 	ras_node->ras_obj = ras_block_obj;
4945 	list_add_tail(&ras_node->node, &adev->ras_list);
4946 
4947 	return 0;
4948 }
4949 
4950 void amdgpu_ras_get_error_type_name(uint32_t err_type, char *err_type_name)
4951 {
4952 	if (!err_type_name)
4953 		return;
4954 
4955 	switch (err_type) {
4956 	case AMDGPU_RAS_ERROR__SINGLE_CORRECTABLE:
4957 		sprintf(err_type_name, "correctable");
4958 		break;
4959 	case AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE:
4960 		sprintf(err_type_name, "uncorrectable");
4961 		break;
4962 	default:
4963 		sprintf(err_type_name, "unknown");
4964 		break;
4965 	}
4966 }
4967 
4968 bool amdgpu_ras_inst_get_memory_id_field(struct amdgpu_device *adev,
4969 					 const struct amdgpu_ras_err_status_reg_entry *reg_entry,
4970 					 uint32_t instance,
4971 					 uint32_t *memory_id)
4972 {
4973 	uint32_t err_status_lo_data, err_status_lo_offset;
4974 
4975 	if (!reg_entry)
4976 		return false;
4977 
4978 	err_status_lo_offset =
4979 		AMDGPU_RAS_REG_ENTRY_OFFSET(reg_entry->hwip, instance,
4980 					    reg_entry->seg_lo, reg_entry->reg_lo);
4981 	err_status_lo_data = RREG32(err_status_lo_offset);
4982 
4983 	if ((reg_entry->flags & AMDGPU_RAS_ERR_STATUS_VALID) &&
4984 	    !REG_GET_FIELD(err_status_lo_data, ERR_STATUS_LO, ERR_STATUS_VALID_FLAG))
4985 		return false;
4986 
4987 	*memory_id = REG_GET_FIELD(err_status_lo_data, ERR_STATUS_LO, MEMORY_ID);
4988 
4989 	return true;
4990 }
4991 
4992 bool amdgpu_ras_inst_get_err_cnt_field(struct amdgpu_device *adev,
4993 				       const struct amdgpu_ras_err_status_reg_entry *reg_entry,
4994 				       uint32_t instance,
4995 				       unsigned long *err_cnt)
4996 {
4997 	uint32_t err_status_hi_data, err_status_hi_offset;
4998 
4999 	if (!reg_entry)
5000 		return false;
5001 
5002 	err_status_hi_offset =
5003 		AMDGPU_RAS_REG_ENTRY_OFFSET(reg_entry->hwip, instance,
5004 					    reg_entry->seg_hi, reg_entry->reg_hi);
5005 	err_status_hi_data = RREG32(err_status_hi_offset);
5006 
5007 	if ((reg_entry->flags & AMDGPU_RAS_ERR_INFO_VALID) &&
5008 	    !REG_GET_FIELD(err_status_hi_data, ERR_STATUS_HI, ERR_INFO_VALID_FLAG))
5009 		/* keep the check here in case we need to refer to the result later */
5010 		dev_dbg(adev->dev, "Invalid err_info field\n");
5011 
5012 	/* read err count */
5013 	*err_cnt = REG_GET_FIELD(err_status_hi_data, ERR_STATUS, ERR_CNT);
5014 
5015 	return true;
5016 }
5017 
5018 void amdgpu_ras_inst_query_ras_error_count(struct amdgpu_device *adev,
5019 					   const struct amdgpu_ras_err_status_reg_entry *reg_list,
5020 					   uint32_t reg_list_size,
5021 					   const struct amdgpu_ras_memory_id_entry *mem_list,
5022 					   uint32_t mem_list_size,
5023 					   uint32_t instance,
5024 					   uint32_t err_type,
5025 					   unsigned long *err_count)
5026 {
5027 	uint32_t memory_id;
5028 	unsigned long err_cnt;
5029 	char err_type_name[16];
5030 	uint32_t i, j;
5031 
5032 	for (i = 0; i < reg_list_size; i++) {
5033 		/* query memory_id from err_status_lo */
5034 		if (!amdgpu_ras_inst_get_memory_id_field(adev, &reg_list[i],
5035 							 instance, &memory_id))
5036 			continue;
5037 
5038 		/* query err_cnt from err_status_hi */
5039 		if (!amdgpu_ras_inst_get_err_cnt_field(adev, &reg_list[i],
5040 						       instance, &err_cnt) ||
5041 		    !err_cnt)
5042 			continue;
5043 
5044 		*err_count += err_cnt;
5045 
5046 		/* log the errors */
5047 		amdgpu_ras_get_error_type_name(err_type, err_type_name);
5048 		if (!mem_list) {
5049 			/* memory_list is not supported */
5050 			dev_info(adev->dev,
5051 				 "%ld %s hardware errors detected in %s, instance: %d, memory_id: %d\n",
5052 				 err_cnt, err_type_name,
5053 				 reg_list[i].block_name,
5054 				 instance, memory_id);
5055 		} else {
5056 			for (j = 0; j < mem_list_size; j++) {
5057 				if (memory_id == mem_list[j].memory_id) {
5058 					dev_info(adev->dev,
5059 						 "%ld %s hardware errors detected in %s, instance: %d, memory block: %s\n",
5060 						 err_cnt, err_type_name,
5061 						 reg_list[i].block_name,
5062 						 instance, mem_list[j].name);
5063 					break;
5064 				}
5065 			}
5066 		}
5067 	}
5068 }
5069 
5070 void amdgpu_ras_inst_reset_ras_error_count(struct amdgpu_device *adev,
5071 					   const struct amdgpu_ras_err_status_reg_entry *reg_list,
5072 					   uint32_t reg_list_size,
5073 					   uint32_t instance)
5074 {
5075 	uint32_t err_status_lo_offset, err_status_hi_offset;
5076 	uint32_t i;
5077 
5078 	for (i = 0; i < reg_list_size; i++) {
5079 		err_status_lo_offset =
5080 			AMDGPU_RAS_REG_ENTRY_OFFSET(reg_list[i].hwip, instance,
5081 						    reg_list[i].seg_lo, reg_list[i].reg_lo);
5082 		err_status_hi_offset =
5083 			AMDGPU_RAS_REG_ENTRY_OFFSET(reg_list[i].hwip, instance,
5084 						    reg_list[i].seg_hi, reg_list[i].reg_hi);
5085 		WREG32(err_status_lo_offset, 0);
5086 		WREG32(err_status_hi_offset, 0);
5087 	}
5088 }
5089 
5090 int amdgpu_ras_error_data_init(struct ras_err_data *err_data)
5091 {
5092 	memset(err_data, 0, sizeof(*err_data));
5093 
5094 	INIT_LIST_HEAD(&err_data->err_node_list);
5095 
5096 	return 0;
5097 }
5098 
5099 static void amdgpu_ras_error_node_release(struct ras_err_node *err_node)
5100 {
5101 	if (!err_node)
5102 		return;
5103 
5104 	list_del(&err_node->node);
5105 	kvfree(err_node);
5106 }
5107 
5108 void amdgpu_ras_error_data_fini(struct ras_err_data *err_data)
5109 {
5110 	struct ras_err_node *err_node, *tmp;
5111 
5112 	list_for_each_entry_safe(err_node, tmp, &err_data->err_node_list, node)
5113 		amdgpu_ras_error_node_release(err_node);
5114 }
5115 
5116 static struct ras_err_node *amdgpu_ras_error_find_node_by_id(struct ras_err_data *err_data,
5117 							     struct amdgpu_smuio_mcm_config_info *mcm_info)
5118 {
5119 	struct ras_err_node *err_node;
5120 	struct amdgpu_smuio_mcm_config_info *ref_id;
5121 
5122 	if (!err_data || !mcm_info)
5123 		return NULL;
5124 
5125 	for_each_ras_error(err_node, err_data) {
5126 		ref_id = &err_node->err_info.mcm_info;
5127 
5128 		if (mcm_info->socket_id == ref_id->socket_id &&
5129 		    mcm_info->die_id == ref_id->die_id)
5130 			return err_node;
5131 	}
5132 
5133 	return NULL;
5134 }
5135 
5136 static struct ras_err_node *amdgpu_ras_error_node_new(void)
5137 {
5138 	struct ras_err_node *err_node;
5139 
5140 	err_node = kvzalloc(sizeof(*err_node), GFP_KERNEL);
5141 	if (!err_node)
5142 		return NULL;
5143 
5144 	INIT_LIST_HEAD(&err_node->node);
5145 
5146 	return err_node;
5147 }
5148 
5149 static int ras_err_info_cmp(void *priv, const struct list_head *a, const struct list_head *b)
5150 {
5151 	struct ras_err_node *nodea = container_of(a, struct ras_err_node, node);
5152 	struct ras_err_node *nodeb = container_of(b, struct ras_err_node, node);
5153 	struct amdgpu_smuio_mcm_config_info *infoa = &nodea->err_info.mcm_info;
5154 	struct amdgpu_smuio_mcm_config_info *infob = &nodeb->err_info.mcm_info;
5155 
5156 	if (unlikely(infoa->socket_id != infob->socket_id))
5157 		return infoa->socket_id - infob->socket_id;
5158 	else
5159 		return infoa->die_id - infob->die_id;
5160 
5161 	return 0;
5162 }
5163 
5164 static struct ras_err_info *amdgpu_ras_error_get_info(struct ras_err_data *err_data,
5165 				struct amdgpu_smuio_mcm_config_info *mcm_info)
5166 {
5167 	struct ras_err_node *err_node;
5168 
5169 	err_node = amdgpu_ras_error_find_node_by_id(err_data, mcm_info);
5170 	if (err_node)
5171 		return &err_node->err_info;
5172 
5173 	err_node = amdgpu_ras_error_node_new();
5174 	if (!err_node)
5175 		return NULL;
5176 
5177 	memcpy(&err_node->err_info.mcm_info, mcm_info, sizeof(*mcm_info));
5178 
5179 	err_data->err_list_count++;
5180 	list_add_tail(&err_node->node, &err_data->err_node_list);
5181 	list_sort(NULL, &err_data->err_node_list, ras_err_info_cmp);
5182 
5183 	return &err_node->err_info;
5184 }
5185 
5186 int amdgpu_ras_error_statistic_ue_count(struct ras_err_data *err_data,
5187 					struct amdgpu_smuio_mcm_config_info *mcm_info,
5188 					u64 count)
5189 {
5190 	struct ras_err_info *err_info;
5191 
5192 	if (!err_data || !mcm_info)
5193 		return -EINVAL;
5194 
5195 	if (!count)
5196 		return 0;
5197 
5198 	err_info = amdgpu_ras_error_get_info(err_data, mcm_info);
5199 	if (!err_info)
5200 		return -EINVAL;
5201 
5202 	err_info->ue_count += count;
5203 	err_data->ue_count += count;
5204 
5205 	return 0;
5206 }
5207 
5208 int amdgpu_ras_error_statistic_ce_count(struct ras_err_data *err_data,
5209 					struct amdgpu_smuio_mcm_config_info *mcm_info,
5210 					u64 count)
5211 {
5212 	struct ras_err_info *err_info;
5213 
5214 	if (!err_data || !mcm_info)
5215 		return -EINVAL;
5216 
5217 	if (!count)
5218 		return 0;
5219 
5220 	err_info = amdgpu_ras_error_get_info(err_data, mcm_info);
5221 	if (!err_info)
5222 		return -EINVAL;
5223 
5224 	err_info->ce_count += count;
5225 	err_data->ce_count += count;
5226 
5227 	return 0;
5228 }
5229 
5230 int amdgpu_ras_error_statistic_de_count(struct ras_err_data *err_data,
5231 					struct amdgpu_smuio_mcm_config_info *mcm_info,
5232 					u64 count)
5233 {
5234 	struct ras_err_info *err_info;
5235 
5236 	if (!err_data || !mcm_info)
5237 		return -EINVAL;
5238 
5239 	if (!count)
5240 		return 0;
5241 
5242 	err_info = amdgpu_ras_error_get_info(err_data, mcm_info);
5243 	if (!err_info)
5244 		return -EINVAL;
5245 
5246 	err_info->de_count += count;
5247 	err_data->de_count += count;
5248 
5249 	return 0;
5250 }
5251 
5252 #define mmMP0_SMN_C2PMSG_92	0x1609C
5253 #define mmMP0_SMN_C2PMSG_126	0x160BE
5254 static void amdgpu_ras_boot_time_error_reporting(struct amdgpu_device *adev,
5255 						 u32 instance)
5256 {
5257 	u32 socket_id, aid_id, hbm_id;
5258 	u32 fw_status;
5259 	u32 boot_error;
5260 	u64 reg_addr;
5261 
5262 	/* The pattern for smn addressing in other SOC could be different from
5263 	 * the one for aqua_vanjaram. We should revisit the code if the pattern
5264 	 * is changed. In such case, replace the aqua_vanjaram implementation
5265 	 * with more common helper */
5266 	reg_addr = (mmMP0_SMN_C2PMSG_92 << 2) +
5267 		   aqua_vanjaram_encode_ext_smn_addressing(instance);
5268 	fw_status = amdgpu_device_indirect_rreg_ext(adev, reg_addr);
5269 
5270 	reg_addr = (mmMP0_SMN_C2PMSG_126 << 2) +
5271 		   aqua_vanjaram_encode_ext_smn_addressing(instance);
5272 	boot_error = amdgpu_device_indirect_rreg_ext(adev, reg_addr);
5273 
5274 	socket_id = AMDGPU_RAS_GPU_ERR_SOCKET_ID(boot_error);
5275 	aid_id = AMDGPU_RAS_GPU_ERR_AID_ID(boot_error);
5276 	hbm_id = ((1 == AMDGPU_RAS_GPU_ERR_HBM_ID(boot_error)) ? 0 : 1);
5277 
5278 	if (AMDGPU_RAS_GPU_ERR_MEM_TRAINING(boot_error))
5279 		dev_info(adev->dev,
5280 			 "socket: %d, aid: %d, hbm: %d, fw_status: 0x%x, memory training failed\n",
5281 			 socket_id, aid_id, hbm_id, fw_status);
5282 
5283 	if (AMDGPU_RAS_GPU_ERR_FW_LOAD(boot_error))
5284 		dev_info(adev->dev,
5285 			 "socket: %d, aid: %d, fw_status: 0x%x, firmware load failed at boot time\n",
5286 			 socket_id, aid_id, fw_status);
5287 
5288 	if (AMDGPU_RAS_GPU_ERR_WAFL_LINK_TRAINING(boot_error))
5289 		dev_info(adev->dev,
5290 			 "socket: %d, aid: %d, fw_status: 0x%x, wafl link training failed\n",
5291 			 socket_id, aid_id, fw_status);
5292 
5293 	if (AMDGPU_RAS_GPU_ERR_XGMI_LINK_TRAINING(boot_error))
5294 		dev_info(adev->dev,
5295 			 "socket: %d, aid: %d, fw_status: 0x%x, xgmi link training failed\n",
5296 			 socket_id, aid_id, fw_status);
5297 
5298 	if (AMDGPU_RAS_GPU_ERR_USR_CP_LINK_TRAINING(boot_error))
5299 		dev_info(adev->dev,
5300 			 "socket: %d, aid: %d, fw_status: 0x%x, usr cp link training failed\n",
5301 			 socket_id, aid_id, fw_status);
5302 
5303 	if (AMDGPU_RAS_GPU_ERR_USR_DP_LINK_TRAINING(boot_error))
5304 		dev_info(adev->dev,
5305 			 "socket: %d, aid: %d, fw_status: 0x%x, usr dp link training failed\n",
5306 			 socket_id, aid_id, fw_status);
5307 
5308 	if (AMDGPU_RAS_GPU_ERR_HBM_MEM_TEST(boot_error))
5309 		dev_info(adev->dev,
5310 			 "socket: %d, aid: %d, hbm: %d, fw_status: 0x%x, hbm memory test failed\n",
5311 			 socket_id, aid_id, hbm_id, fw_status);
5312 
5313 	if (AMDGPU_RAS_GPU_ERR_HBM_BIST_TEST(boot_error))
5314 		dev_info(adev->dev,
5315 			 "socket: %d, aid: %d, hbm: %d, fw_status: 0x%x, hbm bist test failed\n",
5316 			 socket_id, aid_id, hbm_id, fw_status);
5317 
5318 	if (AMDGPU_RAS_GPU_ERR_DATA_ABORT(boot_error))
5319 		dev_info(adev->dev,
5320 			 "socket: %d, aid: %d, fw_status: 0x%x, data abort exception\n",
5321 			 socket_id, aid_id, fw_status);
5322 
5323 	if (AMDGPU_RAS_GPU_ERR_GENERIC(boot_error))
5324 		dev_info(adev->dev,
5325 			 "socket: %d, aid: %d, fw_status: 0x%x, Boot Controller Generic Error\n",
5326 			 socket_id, aid_id, fw_status);
5327 }
5328 
5329 static bool amdgpu_ras_boot_error_detected(struct amdgpu_device *adev,
5330 					   u32 instance)
5331 {
5332 	u64 reg_addr;
5333 	u32 reg_data;
5334 	int retry_loop;
5335 
5336 	reg_addr = (mmMP0_SMN_C2PMSG_92 << 2) +
5337 		   aqua_vanjaram_encode_ext_smn_addressing(instance);
5338 
5339 	for (retry_loop = 0; retry_loop < AMDGPU_RAS_BOOT_STATUS_POLLING_LIMIT; retry_loop++) {
5340 		reg_data = amdgpu_device_indirect_rreg_ext(adev, reg_addr);
5341 		if ((reg_data & AMDGPU_RAS_BOOT_STATUS_MASK) == AMDGPU_RAS_BOOT_STEADY_STATUS)
5342 			return false;
5343 		else
5344 			msleep(1);
5345 	}
5346 
5347 	return true;
5348 }
5349 
5350 void amdgpu_ras_query_boot_status(struct amdgpu_device *adev, u32 num_instances)
5351 {
5352 	u32 i;
5353 
5354 	for (i = 0; i < num_instances; i++) {
5355 		if (amdgpu_ras_boot_error_detected(adev, i))
5356 			amdgpu_ras_boot_time_error_reporting(adev, i);
5357 	}
5358 }
5359 
5360 int amdgpu_ras_reserve_page(struct amdgpu_device *adev, uint64_t pfn)
5361 {
5362 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
5363 	struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
5364 	uint64_t start = pfn << AMDGPU_GPU_PAGE_SHIFT;
5365 	int ret = 0;
5366 
5367 	if (amdgpu_ras_check_critical_address(adev, start))
5368 		return 0;
5369 
5370 	mutex_lock(&con->page_rsv_lock);
5371 	ret = amdgpu_vram_mgr_query_page_status(mgr, start);
5372 	if (ret == -ENOENT)
5373 		ret = amdgpu_vram_mgr_reserve_range(mgr, start, AMDGPU_GPU_PAGE_SIZE);
5374 	mutex_unlock(&con->page_rsv_lock);
5375 
5376 	return ret;
5377 }
5378 
5379 void amdgpu_ras_event_log_print(struct amdgpu_device *adev, u64 event_id,
5380 				const char *fmt, ...)
5381 {
5382 	struct va_format vaf;
5383 	va_list args;
5384 
5385 	va_start(args, fmt);
5386 	vaf.fmt = fmt;
5387 	vaf.va = &args;
5388 
5389 	if (RAS_EVENT_ID_IS_VALID(event_id))
5390 		dev_printk(KERN_INFO, adev->dev, "{%llu}%pV", event_id, &vaf);
5391 	else
5392 		dev_printk(KERN_INFO, adev->dev, "%pV", &vaf);
5393 
5394 	va_end(args);
5395 }
5396 
5397 bool amdgpu_ras_is_rma(struct amdgpu_device *adev)
5398 {
5399 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
5400 
5401 	if (!con)
5402 		return false;
5403 
5404 	return con->is_rma;
5405 }
5406 
5407 int amdgpu_ras_add_critical_region(struct amdgpu_device *adev,
5408 			struct amdgpu_bo *bo)
5409 {
5410 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
5411 	struct amdgpu_vram_mgr_resource *vres;
5412 	struct ras_critical_region *region;
5413 	struct drm_buddy_block *block;
5414 	int ret = 0;
5415 
5416 	if (!bo || !bo->tbo.resource)
5417 		return -EINVAL;
5418 
5419 	vres = to_amdgpu_vram_mgr_resource(bo->tbo.resource);
5420 
5421 	mutex_lock(&con->critical_region_lock);
5422 
5423 	/* Check if the bo had been recorded */
5424 	list_for_each_entry(region, &con->critical_region_head, node)
5425 		if (region->bo == bo)
5426 			goto out;
5427 
5428 	/* Record new critical amdgpu bo */
5429 	list_for_each_entry(block, &vres->blocks, link) {
5430 		region = kzalloc(sizeof(*region), GFP_KERNEL);
5431 		if (!region) {
5432 			ret = -ENOMEM;
5433 			goto out;
5434 		}
5435 		region->bo = bo;
5436 		region->start = amdgpu_vram_mgr_block_start(block);
5437 		region->size = amdgpu_vram_mgr_block_size(block);
5438 		list_add_tail(&region->node, &con->critical_region_head);
5439 	}
5440 
5441 out:
5442 	mutex_unlock(&con->critical_region_lock);
5443 
5444 	return ret;
5445 }
5446 
5447 static void amdgpu_ras_critical_region_init(struct amdgpu_device *adev)
5448 {
5449 	amdgpu_ras_add_critical_region(adev, adev->mman.fw_reserved_memory);
5450 }
5451 
5452 static void amdgpu_ras_critical_region_fini(struct amdgpu_device *adev)
5453 {
5454 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
5455 	struct ras_critical_region *region, *tmp;
5456 
5457 	mutex_lock(&con->critical_region_lock);
5458 	list_for_each_entry_safe(region, tmp, &con->critical_region_head, node) {
5459 		list_del(&region->node);
5460 		kfree(region);
5461 	}
5462 	mutex_unlock(&con->critical_region_lock);
5463 }
5464 
5465 bool amdgpu_ras_check_critical_address(struct amdgpu_device *adev, uint64_t addr)
5466 {
5467 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
5468 	struct ras_critical_region *region;
5469 	bool ret = false;
5470 
5471 	mutex_lock(&con->critical_region_lock);
5472 	list_for_each_entry(region, &con->critical_region_head, node) {
5473 		if ((region->start <= addr) &&
5474 		    (addr < (region->start + region->size))) {
5475 			ret = true;
5476 			break;
5477 		}
5478 	}
5479 	mutex_unlock(&con->critical_region_lock);
5480 
5481 	return ret;
5482 }
5483