xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c (revision 8c6a0234739e33c8be8830c2eee13a49acfd59ea)
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 #include "amdgpu_ras_mgr.h"
45 
46 #ifdef CONFIG_X86_MCE_AMD
47 #include <asm/mce.h>
48 
49 static bool notifier_registered;
50 #endif
51 static const char *RAS_FS_NAME = "ras";
52 
53 const char *ras_error_string[] = {
54 	"none",
55 	"parity",
56 	"single_correctable",
57 	"multi_uncorrectable",
58 	"poison",
59 };
60 
61 const char *ras_block_string[] = {
62 	"umc",
63 	"sdma",
64 	"gfx",
65 	"mmhub",
66 	"athub",
67 	"pcie_bif",
68 	"hdp",
69 	"xgmi_wafl",
70 	"df",
71 	"smn",
72 	"sem",
73 	"mp0",
74 	"mp1",
75 	"fuse",
76 	"mca",
77 	"vcn",
78 	"jpeg",
79 	"ih",
80 	"mpio",
81 	"mmsch",
82 };
83 
84 const char *ras_mca_block_string[] = {
85 	"mca_mp0",
86 	"mca_mp1",
87 	"mca_mpio",
88 	"mca_iohc",
89 };
90 
91 struct amdgpu_ras_block_list {
92 	/* ras block link */
93 	struct list_head node;
94 
95 	struct amdgpu_ras_block_object *ras_obj;
96 };
97 
98 const char *get_ras_block_str(struct ras_common_if *ras_block)
99 {
100 	if (!ras_block)
101 		return "NULL";
102 
103 	if (ras_block->block >= AMDGPU_RAS_BLOCK_COUNT ||
104 	    ras_block->block >= ARRAY_SIZE(ras_block_string))
105 		return "OUT OF RANGE";
106 
107 	if (ras_block->block == AMDGPU_RAS_BLOCK__MCA)
108 		return ras_mca_block_string[ras_block->sub_block_index];
109 
110 	return ras_block_string[ras_block->block];
111 }
112 
113 #define ras_block_str(_BLOCK_) \
114 	(((_BLOCK_) < ARRAY_SIZE(ras_block_string)) ? ras_block_string[_BLOCK_] : "Out Of Range")
115 
116 #define ras_err_str(i) (ras_error_string[ffs(i)])
117 
118 #define RAS_DEFAULT_FLAGS (AMDGPU_RAS_FLAG_INIT_BY_VBIOS)
119 
120 /* inject address is 52 bits */
121 #define	RAS_UMC_INJECT_ADDR_LIMIT	(0x1ULL << 52)
122 
123 /* typical ECC bad page rate is 1 bad page per 100MB VRAM */
124 #define RAS_BAD_PAGE_COVER              (100 * 1024 * 1024ULL)
125 
126 #define MAX_UMC_POISON_POLLING_TIME_ASYNC  10
127 
128 #define AMDGPU_RAS_RETIRE_PAGE_INTERVAL 100  //ms
129 
130 #define MAX_FLUSH_RETIRE_DWORK_TIMES  100
131 
132 #define BYPASS_ALLOCATED_ADDRESS        0x0
133 #define BYPASS_INITIALIZATION_ADDRESS   0x1
134 
135 enum amdgpu_ras_retire_page_reservation {
136 	AMDGPU_RAS_RETIRE_PAGE_RESERVED,
137 	AMDGPU_RAS_RETIRE_PAGE_PENDING,
138 	AMDGPU_RAS_RETIRE_PAGE_FAULT,
139 };
140 
141 atomic_t amdgpu_ras_in_intr = ATOMIC_INIT(0);
142 
143 static int amdgpu_ras_check_bad_page_unlock(struct amdgpu_ras *con,
144 				uint64_t addr);
145 static int amdgpu_ras_check_bad_page(struct amdgpu_device *adev,
146 				uint64_t addr);
147 
148 static void amdgpu_ras_critical_region_init(struct amdgpu_device *adev);
149 static void amdgpu_ras_critical_region_fini(struct amdgpu_device *adev);
150 
151 #ifdef CONFIG_X86_MCE_AMD
152 static void amdgpu_register_bad_pages_mca_notifier(struct amdgpu_device *adev);
153 struct mce_notifier_adev_list {
154 	struct amdgpu_device *devs[MAX_GPU_INSTANCE];
155 	int num_gpu;
156 };
157 static struct mce_notifier_adev_list mce_adev_list;
158 #endif
159 
160 void amdgpu_ras_set_error_query_ready(struct amdgpu_device *adev, bool ready)
161 {
162 	if (adev && amdgpu_ras_get_context(adev))
163 		amdgpu_ras_get_context(adev)->error_query_ready = ready;
164 }
165 
166 static bool amdgpu_ras_get_error_query_ready(struct amdgpu_device *adev)
167 {
168 	if (adev && amdgpu_ras_get_context(adev))
169 		return amdgpu_ras_get_context(adev)->error_query_ready;
170 
171 	return false;
172 }
173 
174 static int amdgpu_reserve_page_direct(struct amdgpu_device *adev, uint64_t address)
175 {
176 	struct ras_err_data err_data;
177 	struct eeprom_table_record err_rec;
178 	int ret;
179 
180 	ret = amdgpu_ras_check_bad_page(adev, address);
181 	if (ret == -EINVAL) {
182 		dev_warn(adev->dev,
183 			"RAS WARN: input address 0x%llx is invalid.\n",
184 			address);
185 		return -EINVAL;
186 	} else if (ret == 1) {
187 		dev_warn(adev->dev,
188 			"RAS WARN: 0x%llx has already been marked as bad page!\n",
189 			address);
190 		return 0;
191 	}
192 
193 	ret = amdgpu_ras_error_data_init(&err_data);
194 	if (ret)
195 		return ret;
196 
197 	memset(&err_rec, 0x0, sizeof(struct eeprom_table_record));
198 	err_data.err_addr = &err_rec;
199 	amdgpu_umc_fill_error_record(&err_data, address, address, 0, 0);
200 
201 	if (amdgpu_bad_page_threshold != 0) {
202 		amdgpu_ras_add_bad_pages(adev, err_data.err_addr,
203 					 err_data.err_addr_cnt, false);
204 		amdgpu_ras_save_bad_pages(adev, NULL);
205 	}
206 
207 	amdgpu_ras_error_data_fini(&err_data);
208 
209 	dev_warn(adev->dev, "WARNING: THIS IS ONLY FOR TEST PURPOSES AND WILL CORRUPT RAS EEPROM\n");
210 	dev_warn(adev->dev, "Clear EEPROM:\n");
211 	dev_warn(adev->dev, "    echo 1 > /sys/kernel/debug/dri/0/ras/ras_eeprom_reset\n");
212 
213 	return 0;
214 }
215 
216 static int amdgpu_check_address_validity(struct amdgpu_device *adev,
217 			uint64_t address, uint64_t flags)
218 {
219 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
220 	struct amdgpu_vram_block_info blk_info;
221 	uint64_t page_pfns[32] = {0};
222 	int i, ret, count;
223 	bool hit = false;
224 
225 	if (amdgpu_ip_version(adev, UMC_HWIP, 0) < IP_VERSION(12, 0, 0))
226 		return 0;
227 
228 	if (amdgpu_sriov_vf(adev)) {
229 		if (amdgpu_virt_check_vf_critical_region(adev, address, &hit))
230 			return -EPERM;
231 		return hit ? -EACCES : 0;
232 	}
233 
234 	if ((address >= adev->gmc.mc_vram_size) ||
235 	    (address >= RAS_UMC_INJECT_ADDR_LIMIT))
236 		return -EFAULT;
237 
238 	count = amdgpu_umc_lookup_bad_pages_in_a_row(adev,
239 				address, page_pfns, ARRAY_SIZE(page_pfns));
240 	if (count <= 0)
241 		return -EPERM;
242 
243 	for (i = 0; i < count; i++) {
244 		memset(&blk_info, 0, sizeof(blk_info));
245 		ret = amdgpu_vram_mgr_query_address_block_info(&adev->mman.vram_mgr,
246 					page_pfns[i] << AMDGPU_GPU_PAGE_SHIFT, &blk_info);
247 		if (!ret) {
248 			/* The input address that needs to be checked is allocated by
249 			 * current calling process, so it is necessary to exclude
250 			 * the calling process.
251 			 */
252 			if ((flags == BYPASS_ALLOCATED_ADDRESS) &&
253 			    ((blk_info.task.pid != task_pid_nr(current)) ||
254 				strncmp(blk_info.task.comm, current->comm, TASK_COMM_LEN)))
255 				return -EACCES;
256 			else if ((flags == BYPASS_INITIALIZATION_ADDRESS) &&
257 				(blk_info.task.pid == con->init_task_pid) &&
258 				!strncmp(blk_info.task.comm, con->init_task_comm, TASK_COMM_LEN))
259 				return -EACCES;
260 		}
261 	}
262 
263 	return 0;
264 }
265 
266 static ssize_t amdgpu_ras_debugfs_read(struct file *f, char __user *buf,
267 					size_t size, loff_t *pos)
268 {
269 	struct ras_manager *obj = (struct ras_manager *)file_inode(f)->i_private;
270 	struct ras_query_if info = {
271 		.head = obj->head,
272 	};
273 	ssize_t s;
274 	char val[128];
275 
276 	if (amdgpu_ras_query_error_status(obj->adev, &info))
277 		return -EINVAL;
278 
279 	/* Hardware counter will be reset automatically after the query on Vega20 and Arcturus */
280 	if (amdgpu_ip_version(obj->adev, MP0_HWIP, 0) != IP_VERSION(11, 0, 2) &&
281 	    amdgpu_ip_version(obj->adev, MP0_HWIP, 0) != IP_VERSION(11, 0, 4)) {
282 		if (amdgpu_ras_reset_error_status(obj->adev, info.head.block))
283 			dev_warn(obj->adev->dev, "Failed to reset error counter and error status");
284 	}
285 
286 	s = snprintf(val, sizeof(val), "%s: %lu\n%s: %lu\n",
287 			"ue", info.ue_count,
288 			"ce", info.ce_count);
289 	if (*pos >= s)
290 		return 0;
291 
292 	s -= *pos;
293 	s = min_t(u64, s, size);
294 
295 
296 	if (copy_to_user(buf, &val[*pos], s))
297 		return -EINVAL;
298 
299 	*pos += s;
300 
301 	return s;
302 }
303 
304 static const struct file_operations amdgpu_ras_debugfs_ops = {
305 	.owner = THIS_MODULE,
306 	.read = amdgpu_ras_debugfs_read,
307 	.write = NULL,
308 	.llseek = default_llseek
309 };
310 
311 static int amdgpu_ras_find_block_id_by_name(const char *name, int *block_id)
312 {
313 	int i;
314 
315 	for (i = 0; i < ARRAY_SIZE(ras_block_string); i++) {
316 		*block_id = i;
317 		if (strcmp(name, ras_block_string[i]) == 0)
318 			return 0;
319 	}
320 	return -EINVAL;
321 }
322 
323 static int amdgpu_ras_debugfs_ctrl_parse_data(struct file *f,
324 		const char __user *buf, size_t size,
325 		loff_t *pos, struct ras_debug_if *data)
326 {
327 	ssize_t s = min_t(u64, 64, size);
328 	char str[65];
329 	char block_name[33];
330 	char err[9] = "ue";
331 	int op = -1;
332 	int block_id;
333 	uint32_t sub_block;
334 	u64 address, value;
335 	/* default value is 0 if the mask is not set by user */
336 	u32 instance_mask = 0;
337 
338 	if (*pos)
339 		return -EINVAL;
340 	*pos = size;
341 
342 	memset(str, 0, sizeof(str));
343 	memset(data, 0, sizeof(*data));
344 
345 	if (copy_from_user(str, buf, s))
346 		return -EINVAL;
347 
348 	if (sscanf(str, "disable %32s", block_name) == 1)
349 		op = 0;
350 	else if (sscanf(str, "enable %32s %8s", block_name, err) == 2)
351 		op = 1;
352 	else if (sscanf(str, "inject %32s %8s", block_name, err) == 2)
353 		op = 2;
354 	else if (strstr(str, "retire_page") != NULL)
355 		op = 3;
356 	else if (strstr(str, "check_address") != NULL)
357 		op = 4;
358 	else if (str[0] && str[1] && str[2] && str[3])
359 		/* ascii string, but commands are not matched. */
360 		return -EINVAL;
361 
362 	if (op != -1) {
363 		if (op == 3) {
364 			if (sscanf(str, "%*s 0x%llx", &address) != 1 &&
365 			    sscanf(str, "%*s %llu", &address) != 1)
366 				return -EINVAL;
367 
368 			data->op = op;
369 			data->inject.address = address;
370 
371 			return 0;
372 		} else if (op == 4) {
373 			if (sscanf(str, "%*s 0x%llx 0x%llx", &address, &value) != 2 &&
374 			    sscanf(str, "%*s %llu %llu", &address, &value) != 2)
375 				return -EINVAL;
376 
377 			data->op = op;
378 			data->inject.address = address;
379 			data->inject.value = value;
380 			return 0;
381 		}
382 
383 		if (amdgpu_ras_find_block_id_by_name(block_name, &block_id))
384 			return -EINVAL;
385 
386 		data->head.block = block_id;
387 		/* only ue, ce and poison errors are supported */
388 		if (!memcmp("ue", err, 2))
389 			data->head.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
390 		else if (!memcmp("ce", err, 2))
391 			data->head.type = AMDGPU_RAS_ERROR__SINGLE_CORRECTABLE;
392 		else if (!memcmp("poison", err, 6))
393 			data->head.type = AMDGPU_RAS_ERROR__POISON;
394 		else
395 			return -EINVAL;
396 
397 		data->op = op;
398 
399 		if (op == 2) {
400 			if (sscanf(str, "%*s %*s %*s 0x%x 0x%llx 0x%llx 0x%x",
401 				   &sub_block, &address, &value, &instance_mask) != 4 &&
402 			    sscanf(str, "%*s %*s %*s %u %llu %llu %u",
403 				   &sub_block, &address, &value, &instance_mask) != 4 &&
404 				sscanf(str, "%*s %*s %*s 0x%x 0x%llx 0x%llx",
405 				   &sub_block, &address, &value) != 3 &&
406 			    sscanf(str, "%*s %*s %*s %u %llu %llu",
407 				   &sub_block, &address, &value) != 3)
408 				return -EINVAL;
409 			data->head.sub_block_index = sub_block;
410 			data->inject.address = address;
411 			data->inject.value = value;
412 			data->inject.instance_mask = instance_mask;
413 		}
414 	} else {
415 		if (size < sizeof(*data))
416 			return -EINVAL;
417 
418 		if (copy_from_user(data, buf, sizeof(*data)))
419 			return -EINVAL;
420 	}
421 
422 	return 0;
423 }
424 
425 static void amdgpu_ras_instance_mask_check(struct amdgpu_device *adev,
426 				struct ras_debug_if *data)
427 {
428 	int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;
429 	uint32_t mask, inst_mask = data->inject.instance_mask;
430 
431 	/* no need to set instance mask if there is only one instance */
432 	if (num_xcc <= 1 && inst_mask) {
433 		data->inject.instance_mask = 0;
434 		dev_dbg(adev->dev,
435 			"RAS inject mask(0x%x) isn't supported and force it to 0.\n",
436 			inst_mask);
437 
438 		return;
439 	}
440 
441 	switch (data->head.block) {
442 	case AMDGPU_RAS_BLOCK__GFX:
443 		mask = GENMASK(num_xcc - 1, 0);
444 		break;
445 	case AMDGPU_RAS_BLOCK__SDMA:
446 		mask = GENMASK(adev->sdma.num_instances - 1, 0);
447 		break;
448 	case AMDGPU_RAS_BLOCK__VCN:
449 	case AMDGPU_RAS_BLOCK__JPEG:
450 		mask = GENMASK(adev->vcn.num_vcn_inst - 1, 0);
451 		break;
452 	default:
453 		mask = inst_mask;
454 		break;
455 	}
456 
457 	/* remove invalid bits in instance mask */
458 	data->inject.instance_mask &= mask;
459 	if (inst_mask != data->inject.instance_mask)
460 		dev_dbg(adev->dev,
461 			"Adjust RAS inject mask 0x%x to 0x%x\n",
462 			inst_mask, data->inject.instance_mask);
463 }
464 
465 /**
466  * DOC: AMDGPU RAS debugfs control interface
467  *
468  * The control interface accepts struct ras_debug_if which has two members.
469  *
470  * First member: ras_debug_if::head or ras_debug_if::inject.
471  *
472  * head is used to indicate which IP block will be under control.
473  *
474  * head has four members, they are block, type, sub_block_index, name.
475  * block: which IP will be under control.
476  * type: what kind of error will be enabled/disabled/injected.
477  * sub_block_index: some IPs have subcomponets. say, GFX, sDMA.
478  * name: the name of IP.
479  *
480  * inject has three more members than head, they are address, value and mask.
481  * As their names indicate, inject operation will write the
482  * value to the address.
483  *
484  * The second member: struct ras_debug_if::op.
485  * It has three kinds of operations.
486  *
487  * - 0: disable RAS on the block. Take ::head as its data.
488  * - 1: enable RAS on the block. Take ::head as its data.
489  * - 2: inject errors on the block. Take ::inject as its data.
490  *
491  * How to use the interface?
492  *
493  * In a program
494  *
495  * Copy the struct ras_debug_if in your code and initialize it.
496  * Write the struct to the control interface.
497  *
498  * From shell
499  *
500  * .. code-block:: bash
501  *
502  *	echo "disable <block>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
503  *	echo "enable  <block> <error>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
504  *	echo "inject  <block> <error> <sub-block> <address> <value> <mask>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
505  *
506  * Where N, is the card which you want to affect.
507  *
508  * "disable" requires only the block.
509  * "enable" requires the block and error type.
510  * "inject" requires the block, error type, address, and value.
511  *
512  * The block is one of: umc, sdma, gfx, etc.
513  *	see ras_block_string[] for details
514  *
515  * The error type is one of: ue, ce and poison where,
516  *	ue is multi-uncorrectable
517  *	ce is single-correctable
518  *	poison is poison
519  *
520  * The sub-block is a the sub-block index, pass 0 if there is no sub-block.
521  * The address and value are hexadecimal numbers, leading 0x is optional.
522  * The mask means instance mask, is optional, default value is 0x1.
523  *
524  * For instance,
525  *
526  * .. code-block:: bash
527  *
528  *	echo inject umc ue 0x0 0x0 0x0 > /sys/kernel/debug/dri/0/ras/ras_ctrl
529  *	echo inject umc ce 0 0 0 3 > /sys/kernel/debug/dri/0/ras/ras_ctrl
530  *	echo disable umc > /sys/kernel/debug/dri/0/ras/ras_ctrl
531  *
532  * How to check the result of the operation?
533  *
534  * To check disable/enable, see "ras" features at,
535  * /sys/class/drm/card[0/1/2...]/device/ras/features
536  *
537  * To check inject, see the corresponding error count at,
538  * /sys/class/drm/card[0/1/2...]/device/ras/[gfx|sdma|umc|...]_err_count
539  *
540  * .. note::
541  *	Operations are only allowed on blocks which are supported.
542  *	Check the "ras" mask at /sys/module/amdgpu/parameters/ras_mask
543  *	to see which blocks support RAS on a particular asic.
544  *
545  */
546 static ssize_t amdgpu_ras_debugfs_ctrl_write(struct file *f,
547 					     const char __user *buf,
548 					     size_t size, loff_t *pos)
549 {
550 	struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
551 	struct ras_debug_if data;
552 	int ret = 0;
553 
554 	if (!amdgpu_ras_get_error_query_ready(adev)) {
555 		dev_warn(adev->dev, "RAS WARN: error injection "
556 				"currently inaccessible\n");
557 		return size;
558 	}
559 
560 	ret = amdgpu_ras_debugfs_ctrl_parse_data(f, buf, size, pos, &data);
561 	if (ret)
562 		return ret;
563 
564 	if (data.op == 3) {
565 		ret = amdgpu_reserve_page_direct(adev, data.inject.address);
566 		if (!ret)
567 			return size;
568 		else
569 			return ret;
570 	} else if (data.op == 4) {
571 		ret = amdgpu_check_address_validity(adev, data.inject.address, data.inject.value);
572 		return ret ? ret : size;
573 	}
574 
575 	if (!amdgpu_ras_is_supported(adev, data.head.block))
576 		return -EINVAL;
577 
578 	switch (data.op) {
579 	case 0:
580 		ret = amdgpu_ras_feature_enable(adev, &data.head, 0);
581 		break;
582 	case 1:
583 		ret = amdgpu_ras_feature_enable(adev, &data.head, 1);
584 		break;
585 	case 2:
586 		/* umc ce/ue error injection for a bad page is not allowed */
587 		if (data.head.block == AMDGPU_RAS_BLOCK__UMC)
588 			ret = amdgpu_ras_check_bad_page(adev, data.inject.address);
589 		if (ret == -EINVAL) {
590 			dev_warn(adev->dev, "RAS WARN: input address 0x%llx is invalid.",
591 					data.inject.address);
592 			break;
593 		} else if (ret == 1) {
594 			dev_warn(adev->dev, "RAS WARN: inject: 0x%llx has already been marked as bad!\n",
595 					data.inject.address);
596 			break;
597 		}
598 
599 		amdgpu_ras_instance_mask_check(adev, &data);
600 
601 		/* data.inject.address is offset instead of absolute gpu address */
602 		ret = amdgpu_ras_error_inject(adev, &data.inject);
603 		break;
604 	default:
605 		ret = -EINVAL;
606 		break;
607 	}
608 
609 	if (ret)
610 		return ret;
611 
612 	return size;
613 }
614 
615 /**
616  * DOC: AMDGPU RAS debugfs EEPROM table reset interface
617  *
618  * Some boards contain an EEPROM which is used to persistently store a list of
619  * bad pages which experiences ECC errors in vram.  This interface provides
620  * a way to reset the EEPROM, e.g., after testing error injection.
621  *
622  * Usage:
623  *
624  * .. code-block:: bash
625  *
626  *	echo 1 > ../ras/ras_eeprom_reset
627  *
628  * will reset EEPROM table to 0 entries.
629  *
630  */
631 static ssize_t amdgpu_ras_debugfs_eeprom_write(struct file *f,
632 					       const char __user *buf,
633 					       size_t size, loff_t *pos)
634 {
635 	struct amdgpu_device *adev =
636 		(struct amdgpu_device *)file_inode(f)->i_private;
637 	int ret;
638 
639 	ret = amdgpu_ras_eeprom_reset_table(
640 		&(amdgpu_ras_get_context(adev)->eeprom_control));
641 
642 	if (!ret) {
643 		/* Something was written to EEPROM.
644 		 */
645 		amdgpu_ras_get_context(adev)->flags = RAS_DEFAULT_FLAGS;
646 		return size;
647 	} else {
648 		return ret;
649 	}
650 }
651 
652 static const struct file_operations amdgpu_ras_debugfs_ctrl_ops = {
653 	.owner = THIS_MODULE,
654 	.read = NULL,
655 	.write = amdgpu_ras_debugfs_ctrl_write,
656 	.llseek = default_llseek
657 };
658 
659 static const struct file_operations amdgpu_ras_debugfs_eeprom_ops = {
660 	.owner = THIS_MODULE,
661 	.read = NULL,
662 	.write = amdgpu_ras_debugfs_eeprom_write,
663 	.llseek = default_llseek
664 };
665 
666 /**
667  * DOC: AMDGPU RAS sysfs Error Count Interface
668  *
669  * It allows the user to read the error count for each IP block on the gpu through
670  * /sys/class/drm/card[0/1/2...]/device/ras/[gfx/sdma/...]_err_count
671  *
672  * It outputs the multiple lines which report the uncorrected (ue) and corrected
673  * (ce) error counts.
674  *
675  * The format of one line is below,
676  *
677  * [ce|ue]: count
678  *
679  * Example:
680  *
681  * .. code-block:: bash
682  *
683  *	ue: 0
684  *	ce: 1
685  *
686  */
687 static ssize_t amdgpu_ras_sysfs_read(struct device *dev,
688 		struct device_attribute *attr, char *buf)
689 {
690 	struct ras_manager *obj = container_of(attr, struct ras_manager, sysfs_attr);
691 	struct ras_query_if info = {
692 		.head = obj->head,
693 	};
694 
695 	if (!amdgpu_ras_get_error_query_ready(obj->adev))
696 		return sysfs_emit(buf, "Query currently inaccessible\n");
697 
698 	if (amdgpu_ras_query_error_status(obj->adev, &info))
699 		return -EINVAL;
700 
701 	if (amdgpu_ip_version(obj->adev, MP0_HWIP, 0) != IP_VERSION(11, 0, 2) &&
702 	    amdgpu_ip_version(obj->adev, MP0_HWIP, 0) != IP_VERSION(11, 0, 4)) {
703 		if (amdgpu_ras_reset_error_status(obj->adev, info.head.block))
704 			dev_warn(obj->adev->dev, "Failed to reset error counter and error status");
705 	}
706 
707 	if (info.head.block == AMDGPU_RAS_BLOCK__UMC)
708 		return sysfs_emit(buf, "%s: %lu\n%s: %lu\n%s: %lu\n", "ue", info.ue_count,
709 				"ce", info.ce_count, "de", info.de_count);
710 	else
711 		return sysfs_emit(buf, "%s: %lu\n%s: %lu\n", "ue", info.ue_count,
712 				"ce", info.ce_count);
713 }
714 
715 /* obj begin */
716 
717 #define get_obj(obj) do { (obj)->use++; } while (0)
718 #define alive_obj(obj) ((obj)->use)
719 
720 static inline void put_obj(struct ras_manager *obj)
721 {
722 	if (obj && (--obj->use == 0)) {
723 		list_del(&obj->node);
724 		amdgpu_ras_error_data_fini(&obj->err_data);
725 	}
726 
727 	if (obj && (obj->use < 0))
728 		DRM_ERROR("RAS ERROR: Unbalance obj(%s) use\n", get_ras_block_str(&obj->head));
729 }
730 
731 /* make one obj and return it. */
732 static struct ras_manager *amdgpu_ras_create_obj(struct amdgpu_device *adev,
733 		struct ras_common_if *head)
734 {
735 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
736 	struct ras_manager *obj;
737 
738 	if (!adev->ras_enabled || !con)
739 		return NULL;
740 
741 	if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
742 		return NULL;
743 
744 	if (head->block == AMDGPU_RAS_BLOCK__MCA) {
745 		if (head->sub_block_index >= AMDGPU_RAS_MCA_BLOCK__LAST)
746 			return NULL;
747 
748 		obj = &con->objs[AMDGPU_RAS_BLOCK__LAST + head->sub_block_index];
749 	} else
750 		obj = &con->objs[head->block];
751 
752 	/* already exist. return obj? */
753 	if (alive_obj(obj))
754 		return NULL;
755 
756 	if (amdgpu_ras_error_data_init(&obj->err_data))
757 		return NULL;
758 
759 	obj->head = *head;
760 	obj->adev = adev;
761 	list_add(&obj->node, &con->head);
762 	get_obj(obj);
763 
764 	return obj;
765 }
766 
767 /* return an obj equal to head, or the first when head is NULL */
768 struct ras_manager *amdgpu_ras_find_obj(struct amdgpu_device *adev,
769 		struct ras_common_if *head)
770 {
771 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
772 	struct ras_manager *obj;
773 	int i;
774 
775 	if (!adev->ras_enabled || !con)
776 		return NULL;
777 
778 	if (head) {
779 		if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
780 			return NULL;
781 
782 		if (head->block == AMDGPU_RAS_BLOCK__MCA) {
783 			if (head->sub_block_index >= AMDGPU_RAS_MCA_BLOCK__LAST)
784 				return NULL;
785 
786 			obj = &con->objs[AMDGPU_RAS_BLOCK__LAST + head->sub_block_index];
787 		} else
788 			obj = &con->objs[head->block];
789 
790 		if (alive_obj(obj))
791 			return obj;
792 	} else {
793 		for (i = 0; i < AMDGPU_RAS_BLOCK_COUNT + AMDGPU_RAS_MCA_BLOCK_COUNT; i++) {
794 			obj = &con->objs[i];
795 			if (alive_obj(obj))
796 				return obj;
797 		}
798 	}
799 
800 	return NULL;
801 }
802 /* obj end */
803 
804 /* feature ctl begin */
805 static int amdgpu_ras_is_feature_allowed(struct amdgpu_device *adev,
806 					 struct ras_common_if *head)
807 {
808 	return adev->ras_hw_enabled & BIT(head->block);
809 }
810 
811 static int amdgpu_ras_is_feature_enabled(struct amdgpu_device *adev,
812 		struct ras_common_if *head)
813 {
814 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
815 
816 	return con->features & BIT(head->block);
817 }
818 
819 /*
820  * if obj is not created, then create one.
821  * set feature enable flag.
822  */
823 static int __amdgpu_ras_feature_enable(struct amdgpu_device *adev,
824 		struct ras_common_if *head, int enable)
825 {
826 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
827 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
828 
829 	/* If hardware does not support ras, then do not create obj.
830 	 * But if hardware support ras, we can create the obj.
831 	 * Ras framework checks con->hw_supported to see if it need do
832 	 * corresponding initialization.
833 	 * IP checks con->support to see if it need disable ras.
834 	 */
835 	if (!amdgpu_ras_is_feature_allowed(adev, head))
836 		return 0;
837 
838 	if (enable) {
839 		if (!obj) {
840 			obj = amdgpu_ras_create_obj(adev, head);
841 			if (!obj)
842 				return -EINVAL;
843 		} else {
844 			/* In case we create obj somewhere else */
845 			get_obj(obj);
846 		}
847 		con->features |= BIT(head->block);
848 	} else {
849 		if (obj && amdgpu_ras_is_feature_enabled(adev, head)) {
850 			con->features &= ~BIT(head->block);
851 			put_obj(obj);
852 		}
853 	}
854 
855 	return 0;
856 }
857 
858 /* wrapper of psp_ras_enable_features */
859 int amdgpu_ras_feature_enable(struct amdgpu_device *adev,
860 		struct ras_common_if *head, bool enable)
861 {
862 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
863 	union ta_ras_cmd_input *info;
864 	int ret;
865 
866 	if (!con)
867 		return -EINVAL;
868 
869 	/* For non-gfx ip, do not enable ras feature if it is not allowed */
870 	/* For gfx ip, regardless of feature support status, */
871 	/* Force issue enable or disable ras feature commands */
872 	if (head->block != AMDGPU_RAS_BLOCK__GFX &&
873 	    !amdgpu_ras_is_feature_allowed(adev, head))
874 		return 0;
875 
876 	/* Only enable gfx ras feature from host side */
877 	if (head->block == AMDGPU_RAS_BLOCK__GFX &&
878 	    !amdgpu_sriov_vf(adev) &&
879 	    !amdgpu_ras_intr_triggered()) {
880 		info = kzalloc(sizeof(union ta_ras_cmd_input), GFP_KERNEL);
881 		if (!info)
882 			return -ENOMEM;
883 
884 		if (!enable) {
885 			info->disable_features = (struct ta_ras_disable_features_input) {
886 				.block_id =  amdgpu_ras_block_to_ta(head->block),
887 				.error_type = amdgpu_ras_error_to_ta(head->type),
888 			};
889 		} else {
890 			info->enable_features = (struct ta_ras_enable_features_input) {
891 				.block_id =  amdgpu_ras_block_to_ta(head->block),
892 				.error_type = amdgpu_ras_error_to_ta(head->type),
893 			};
894 		}
895 
896 		ret = psp_ras_enable_features(&adev->psp, info, enable);
897 		if (ret) {
898 			dev_err(adev->dev, "ras %s %s failed poison:%d ret:%d\n",
899 				enable ? "enable":"disable",
900 				get_ras_block_str(head),
901 				amdgpu_ras_is_poison_mode_supported(adev), ret);
902 			kfree(info);
903 			return ret;
904 		}
905 
906 		kfree(info);
907 	}
908 
909 	/* setup the obj */
910 	__amdgpu_ras_feature_enable(adev, head, enable);
911 
912 	return 0;
913 }
914 
915 /* Only used in device probe stage and called only once. */
916 int amdgpu_ras_feature_enable_on_boot(struct amdgpu_device *adev,
917 		struct ras_common_if *head, bool enable)
918 {
919 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
920 	int ret;
921 
922 	if (!con)
923 		return -EINVAL;
924 
925 	if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
926 		if (enable) {
927 			/* There is no harm to issue a ras TA cmd regardless of
928 			 * the currecnt ras state.
929 			 * If current state == target state, it will do nothing
930 			 * But sometimes it requests driver to reset and repost
931 			 * with error code -EAGAIN.
932 			 */
933 			ret = amdgpu_ras_feature_enable(adev, head, 1);
934 			/* With old ras TA, we might fail to enable ras.
935 			 * Log it and just setup the object.
936 			 * TODO need remove this WA in the future.
937 			 */
938 			if (ret == -EINVAL) {
939 				ret = __amdgpu_ras_feature_enable(adev, head, 1);
940 				if (!ret)
941 					dev_info(adev->dev,
942 						"RAS INFO: %s setup object\n",
943 						get_ras_block_str(head));
944 			}
945 		} else {
946 			/* setup the object then issue a ras TA disable cmd.*/
947 			ret = __amdgpu_ras_feature_enable(adev, head, 1);
948 			if (ret)
949 				return ret;
950 
951 			/* gfx block ras disable cmd must send to ras-ta */
952 			if (head->block == AMDGPU_RAS_BLOCK__GFX)
953 				con->features |= BIT(head->block);
954 
955 			ret = amdgpu_ras_feature_enable(adev, head, 0);
956 
957 			/* clean gfx block ras features flag */
958 			if (adev->ras_enabled && head->block == AMDGPU_RAS_BLOCK__GFX)
959 				con->features &= ~BIT(head->block);
960 		}
961 	} else
962 		ret = amdgpu_ras_feature_enable(adev, head, enable);
963 
964 	return ret;
965 }
966 
967 static int amdgpu_ras_disable_all_features(struct amdgpu_device *adev,
968 		bool bypass)
969 {
970 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
971 	struct ras_manager *obj, *tmp;
972 
973 	list_for_each_entry_safe(obj, tmp, &con->head, node) {
974 		/* bypass psp.
975 		 * aka just release the obj and corresponding flags
976 		 */
977 		if (bypass) {
978 			if (__amdgpu_ras_feature_enable(adev, &obj->head, 0))
979 				break;
980 		} else {
981 			if (amdgpu_ras_feature_enable(adev, &obj->head, 0))
982 				break;
983 		}
984 	}
985 
986 	return con->features;
987 }
988 
989 static int amdgpu_ras_enable_all_features(struct amdgpu_device *adev,
990 		bool bypass)
991 {
992 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
993 	int i;
994 	const enum amdgpu_ras_error_type default_ras_type = AMDGPU_RAS_ERROR__NONE;
995 
996 	for (i = 0; i < AMDGPU_RAS_BLOCK_COUNT; i++) {
997 		struct ras_common_if head = {
998 			.block = i,
999 			.type = default_ras_type,
1000 			.sub_block_index = 0,
1001 		};
1002 
1003 		if (i == AMDGPU_RAS_BLOCK__MCA)
1004 			continue;
1005 
1006 		if (bypass) {
1007 			/*
1008 			 * bypass psp. vbios enable ras for us.
1009 			 * so just create the obj
1010 			 */
1011 			if (__amdgpu_ras_feature_enable(adev, &head, 1))
1012 				break;
1013 		} else {
1014 			if (amdgpu_ras_feature_enable(adev, &head, 1))
1015 				break;
1016 		}
1017 	}
1018 
1019 	for (i = 0; i < AMDGPU_RAS_MCA_BLOCK_COUNT; i++) {
1020 		struct ras_common_if head = {
1021 			.block = AMDGPU_RAS_BLOCK__MCA,
1022 			.type = default_ras_type,
1023 			.sub_block_index = i,
1024 		};
1025 
1026 		if (bypass) {
1027 			/*
1028 			 * bypass psp. vbios enable ras for us.
1029 			 * so just create the obj
1030 			 */
1031 			if (__amdgpu_ras_feature_enable(adev, &head, 1))
1032 				break;
1033 		} else {
1034 			if (amdgpu_ras_feature_enable(adev, &head, 1))
1035 				break;
1036 		}
1037 	}
1038 
1039 	return con->features;
1040 }
1041 /* feature ctl end */
1042 
1043 static int amdgpu_ras_block_match_default(struct amdgpu_ras_block_object *block_obj,
1044 		enum amdgpu_ras_block block)
1045 {
1046 	if (!block_obj)
1047 		return -EINVAL;
1048 
1049 	if (block_obj->ras_comm.block == block)
1050 		return 0;
1051 
1052 	return -EINVAL;
1053 }
1054 
1055 static struct amdgpu_ras_block_object *amdgpu_ras_get_ras_block(struct amdgpu_device *adev,
1056 					enum amdgpu_ras_block block, uint32_t sub_block_index)
1057 {
1058 	struct amdgpu_ras_block_list *node, *tmp;
1059 	struct amdgpu_ras_block_object *obj;
1060 
1061 	if (block >= AMDGPU_RAS_BLOCK__LAST)
1062 		return NULL;
1063 
1064 	list_for_each_entry_safe(node, tmp, &adev->ras_list, node) {
1065 		if (!node->ras_obj) {
1066 			dev_warn(adev->dev, "Warning: abnormal ras list node.\n");
1067 			continue;
1068 		}
1069 
1070 		obj = node->ras_obj;
1071 		if (obj->ras_block_match) {
1072 			if (obj->ras_block_match(obj, block, sub_block_index) == 0)
1073 				return obj;
1074 		} else {
1075 			if (amdgpu_ras_block_match_default(obj, block) == 0)
1076 				return obj;
1077 		}
1078 	}
1079 
1080 	return NULL;
1081 }
1082 
1083 static void amdgpu_ras_get_ecc_info(struct amdgpu_device *adev, struct ras_err_data *err_data)
1084 {
1085 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1086 	int ret = 0;
1087 
1088 	/*
1089 	 * choosing right query method according to
1090 	 * whether smu support query error information
1091 	 */
1092 	ret = amdgpu_dpm_get_ecc_info(adev, (void *)&(ras->umc_ecc));
1093 	if (ret == -EOPNOTSUPP) {
1094 		if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
1095 			adev->umc.ras->ras_block.hw_ops->query_ras_error_count)
1096 			adev->umc.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data);
1097 
1098 		/* umc query_ras_error_address is also responsible for clearing
1099 		 * error status
1100 		 */
1101 		if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
1102 		    adev->umc.ras->ras_block.hw_ops->query_ras_error_address)
1103 			adev->umc.ras->ras_block.hw_ops->query_ras_error_address(adev, err_data);
1104 	} else if (!ret) {
1105 		if (adev->umc.ras &&
1106 			adev->umc.ras->ecc_info_query_ras_error_count)
1107 			adev->umc.ras->ecc_info_query_ras_error_count(adev, err_data);
1108 
1109 		if (adev->umc.ras &&
1110 			adev->umc.ras->ecc_info_query_ras_error_address)
1111 			adev->umc.ras->ecc_info_query_ras_error_address(adev, err_data);
1112 	}
1113 }
1114 
1115 static void amdgpu_ras_error_print_error_data(struct amdgpu_device *adev,
1116 					      struct ras_manager *ras_mgr,
1117 					      struct ras_err_data *err_data,
1118 					      struct ras_query_context *qctx,
1119 					      const char *blk_name,
1120 					      bool is_ue,
1121 					      bool is_de)
1122 {
1123 	struct amdgpu_smuio_mcm_config_info *mcm_info;
1124 	struct ras_err_node *err_node;
1125 	struct ras_err_info *err_info;
1126 	u64 event_id = qctx->evid.event_id;
1127 
1128 	if (is_ue) {
1129 		for_each_ras_error(err_node, err_data) {
1130 			err_info = &err_node->err_info;
1131 			mcm_info = &err_info->mcm_info;
1132 			if (err_info->ue_count) {
1133 				RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d, "
1134 					      "%lld new uncorrectable hardware errors detected in %s block\n",
1135 					      mcm_info->socket_id,
1136 					      mcm_info->die_id,
1137 					      err_info->ue_count,
1138 					      blk_name);
1139 			}
1140 		}
1141 
1142 		for_each_ras_error(err_node, &ras_mgr->err_data) {
1143 			err_info = &err_node->err_info;
1144 			mcm_info = &err_info->mcm_info;
1145 			RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d, "
1146 				      "%lld uncorrectable hardware errors detected in total in %s block\n",
1147 				      mcm_info->socket_id, mcm_info->die_id, err_info->ue_count, blk_name);
1148 		}
1149 
1150 	} else {
1151 		if (is_de) {
1152 			for_each_ras_error(err_node, err_data) {
1153 				err_info = &err_node->err_info;
1154 				mcm_info = &err_info->mcm_info;
1155 				if (err_info->de_count) {
1156 					RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d, "
1157 						      "%lld new deferred hardware errors detected in %s block\n",
1158 						      mcm_info->socket_id,
1159 						      mcm_info->die_id,
1160 						      err_info->de_count,
1161 						      blk_name);
1162 				}
1163 			}
1164 
1165 			for_each_ras_error(err_node, &ras_mgr->err_data) {
1166 				err_info = &err_node->err_info;
1167 				mcm_info = &err_info->mcm_info;
1168 				RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d, "
1169 					      "%lld deferred hardware errors detected in total in %s block\n",
1170 					      mcm_info->socket_id, mcm_info->die_id,
1171 					      err_info->de_count, blk_name);
1172 			}
1173 		} else {
1174 			if (adev->debug_disable_ce_logs)
1175 				return;
1176 
1177 			for_each_ras_error(err_node, err_data) {
1178 				err_info = &err_node->err_info;
1179 				mcm_info = &err_info->mcm_info;
1180 				if (err_info->ce_count) {
1181 					RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d, "
1182 						      "%lld new correctable hardware errors detected in %s block\n",
1183 						      mcm_info->socket_id,
1184 						      mcm_info->die_id,
1185 						      err_info->ce_count,
1186 						      blk_name);
1187 				}
1188 			}
1189 
1190 			for_each_ras_error(err_node, &ras_mgr->err_data) {
1191 				err_info = &err_node->err_info;
1192 				mcm_info = &err_info->mcm_info;
1193 				RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d, "
1194 					      "%lld correctable hardware errors detected in total in %s block\n",
1195 					      mcm_info->socket_id, mcm_info->die_id,
1196 					      err_info->ce_count, blk_name);
1197 			}
1198 		}
1199 	}
1200 }
1201 
1202 static inline bool err_data_has_source_info(struct ras_err_data *data)
1203 {
1204 	return !list_empty(&data->err_node_list);
1205 }
1206 
1207 static void amdgpu_ras_error_generate_report(struct amdgpu_device *adev,
1208 					     struct ras_query_if *query_if,
1209 					     struct ras_err_data *err_data,
1210 					     struct ras_query_context *qctx)
1211 {
1212 	struct ras_manager *ras_mgr = amdgpu_ras_find_obj(adev, &query_if->head);
1213 	const char *blk_name = get_ras_block_str(&query_if->head);
1214 	u64 event_id = qctx->evid.event_id;
1215 
1216 	if (err_data->ce_count) {
1217 		if (err_data_has_source_info(err_data)) {
1218 			amdgpu_ras_error_print_error_data(adev, ras_mgr, err_data, qctx,
1219 							  blk_name, false, false);
1220 		} else if (!adev->aid_mask &&
1221 			   adev->smuio.funcs &&
1222 			   adev->smuio.funcs->get_socket_id &&
1223 			   adev->smuio.funcs->get_die_id) {
1224 			RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d "
1225 				      "%ld correctable hardware errors "
1226 				      "detected in %s block\n",
1227 				      adev->smuio.funcs->get_socket_id(adev),
1228 				      adev->smuio.funcs->get_die_id(adev),
1229 				      ras_mgr->err_data.ce_count,
1230 				      blk_name);
1231 		} else {
1232 			RAS_EVENT_LOG(adev, event_id, "%ld correctable hardware errors "
1233 				      "detected in %s block\n",
1234 				      ras_mgr->err_data.ce_count,
1235 				      blk_name);
1236 		}
1237 	}
1238 
1239 	if (err_data->ue_count) {
1240 		if (err_data_has_source_info(err_data)) {
1241 			amdgpu_ras_error_print_error_data(adev, ras_mgr, err_data, qctx,
1242 							  blk_name, true, false);
1243 		} else if (!adev->aid_mask &&
1244 			   adev->smuio.funcs &&
1245 			   adev->smuio.funcs->get_socket_id &&
1246 			   adev->smuio.funcs->get_die_id) {
1247 			RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d "
1248 				      "%ld uncorrectable hardware errors "
1249 				      "detected in %s block\n",
1250 				      adev->smuio.funcs->get_socket_id(adev),
1251 				      adev->smuio.funcs->get_die_id(adev),
1252 				      ras_mgr->err_data.ue_count,
1253 				      blk_name);
1254 		} else {
1255 			RAS_EVENT_LOG(adev, event_id, "%ld uncorrectable hardware errors "
1256 				      "detected in %s block\n",
1257 				      ras_mgr->err_data.ue_count,
1258 				      blk_name);
1259 		}
1260 	}
1261 
1262 	if (err_data->de_count) {
1263 		if (err_data_has_source_info(err_data)) {
1264 			amdgpu_ras_error_print_error_data(adev, ras_mgr, err_data, qctx,
1265 							  blk_name, false, true);
1266 		} else if (!adev->aid_mask &&
1267 			   adev->smuio.funcs &&
1268 			   adev->smuio.funcs->get_socket_id &&
1269 			   adev->smuio.funcs->get_die_id) {
1270 			RAS_EVENT_LOG(adev, event_id, "socket: %d, die: %d "
1271 				      "%ld deferred hardware errors "
1272 				      "detected in %s block\n",
1273 				      adev->smuio.funcs->get_socket_id(adev),
1274 				      adev->smuio.funcs->get_die_id(adev),
1275 				      ras_mgr->err_data.de_count,
1276 				      blk_name);
1277 		} else {
1278 			RAS_EVENT_LOG(adev, event_id, "%ld deferred hardware errors "
1279 				      "detected in %s block\n",
1280 				      ras_mgr->err_data.de_count,
1281 				      blk_name);
1282 		}
1283 	}
1284 }
1285 
1286 static void amdgpu_ras_virt_error_generate_report(struct amdgpu_device *adev,
1287 						  struct ras_query_if *query_if,
1288 						  struct ras_err_data *err_data,
1289 						  struct ras_query_context *qctx)
1290 {
1291 	unsigned long new_ue, new_ce, new_de;
1292 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &query_if->head);
1293 	const char *blk_name = get_ras_block_str(&query_if->head);
1294 	u64 event_id = qctx->evid.event_id;
1295 
1296 	new_ce = err_data->ce_count - obj->err_data.ce_count;
1297 	new_ue = err_data->ue_count - obj->err_data.ue_count;
1298 	new_de = err_data->de_count - obj->err_data.de_count;
1299 
1300 	if (new_ce) {
1301 		RAS_EVENT_LOG(adev, event_id, "%lu correctable hardware errors "
1302 			      "detected in %s block\n",
1303 			      new_ce,
1304 			      blk_name);
1305 	}
1306 
1307 	if (new_ue) {
1308 		RAS_EVENT_LOG(adev, event_id, "%lu uncorrectable hardware errors "
1309 			      "detected in %s block\n",
1310 			      new_ue,
1311 			      blk_name);
1312 	}
1313 
1314 	if (new_de) {
1315 		RAS_EVENT_LOG(adev, event_id, "%lu deferred hardware errors "
1316 			      "detected in %s block\n",
1317 			      new_de,
1318 			      blk_name);
1319 	}
1320 }
1321 
1322 static void amdgpu_rasmgr_error_data_statistic_update(struct ras_manager *obj, struct ras_err_data *err_data)
1323 {
1324 	struct ras_err_node *err_node;
1325 	struct ras_err_info *err_info;
1326 
1327 	if (err_data_has_source_info(err_data)) {
1328 		for_each_ras_error(err_node, err_data) {
1329 			err_info = &err_node->err_info;
1330 			amdgpu_ras_error_statistic_de_count(&obj->err_data,
1331 					&err_info->mcm_info, err_info->de_count);
1332 			amdgpu_ras_error_statistic_ce_count(&obj->err_data,
1333 					&err_info->mcm_info, err_info->ce_count);
1334 			amdgpu_ras_error_statistic_ue_count(&obj->err_data,
1335 					&err_info->mcm_info, err_info->ue_count);
1336 		}
1337 	} else {
1338 		/* for legacy asic path which doesn't has error source info */
1339 		obj->err_data.ue_count += err_data->ue_count;
1340 		obj->err_data.ce_count += err_data->ce_count;
1341 		obj->err_data.de_count += err_data->de_count;
1342 	}
1343 }
1344 
1345 static void amdgpu_ras_mgr_virt_error_data_statistics_update(struct ras_manager *obj,
1346 							     struct ras_err_data *err_data)
1347 {
1348 	/* Host reports absolute counts */
1349 	obj->err_data.ue_count = err_data->ue_count;
1350 	obj->err_data.ce_count = err_data->ce_count;
1351 	obj->err_data.de_count = err_data->de_count;
1352 }
1353 
1354 static struct ras_manager *get_ras_manager(struct amdgpu_device *adev, enum amdgpu_ras_block blk)
1355 {
1356 	struct ras_common_if head;
1357 
1358 	memset(&head, 0, sizeof(head));
1359 	head.block = blk;
1360 
1361 	return amdgpu_ras_find_obj(adev, &head);
1362 }
1363 
1364 int amdgpu_ras_bind_aca(struct amdgpu_device *adev, enum amdgpu_ras_block blk,
1365 			const struct aca_info *aca_info, void *data)
1366 {
1367 	struct ras_manager *obj;
1368 
1369 	/* in resume phase, no need to create aca fs node */
1370 	if (adev->in_suspend || amdgpu_reset_in_recovery(adev))
1371 		return 0;
1372 
1373 	obj = get_ras_manager(adev, blk);
1374 	if (!obj)
1375 		return -EINVAL;
1376 
1377 	return amdgpu_aca_add_handle(adev, &obj->aca_handle, ras_block_str(blk), aca_info, data);
1378 }
1379 
1380 int amdgpu_ras_unbind_aca(struct amdgpu_device *adev, enum amdgpu_ras_block blk)
1381 {
1382 	struct ras_manager *obj;
1383 
1384 	obj = get_ras_manager(adev, blk);
1385 	if (!obj)
1386 		return -EINVAL;
1387 
1388 	amdgpu_aca_remove_handle(&obj->aca_handle);
1389 
1390 	return 0;
1391 }
1392 
1393 static int amdgpu_aca_log_ras_error_data(struct amdgpu_device *adev, enum amdgpu_ras_block blk,
1394 					 enum aca_error_type type, struct ras_err_data *err_data,
1395 					 struct ras_query_context *qctx)
1396 {
1397 	struct ras_manager *obj;
1398 
1399 	obj = get_ras_manager(adev, blk);
1400 	if (!obj)
1401 		return -EINVAL;
1402 
1403 	return amdgpu_aca_get_error_data(adev, &obj->aca_handle, type, err_data, qctx);
1404 }
1405 
1406 ssize_t amdgpu_ras_aca_sysfs_read(struct device *dev, struct device_attribute *attr,
1407 				  struct aca_handle *handle, char *buf, void *data)
1408 {
1409 	struct ras_manager *obj = container_of(handle, struct ras_manager, aca_handle);
1410 	struct ras_query_if info = {
1411 		.head = obj->head,
1412 	};
1413 
1414 	if (!amdgpu_ras_get_error_query_ready(obj->adev))
1415 		return sysfs_emit(buf, "Query currently inaccessible\n");
1416 
1417 	if (amdgpu_ras_query_error_status(obj->adev, &info))
1418 		return -EINVAL;
1419 
1420 	return sysfs_emit(buf, "%s: %lu\n%s: %lu\n%s: %lu\n", "ue", info.ue_count,
1421 			  "ce", info.ce_count, "de", info.de_count);
1422 }
1423 
1424 static int amdgpu_ras_query_error_status_helper(struct amdgpu_device *adev,
1425 						struct ras_query_if *info,
1426 						struct ras_err_data *err_data,
1427 						struct ras_query_context *qctx,
1428 						unsigned int error_query_mode)
1429 {
1430 	enum amdgpu_ras_block blk = info ? info->head.block : AMDGPU_RAS_BLOCK_COUNT;
1431 	struct amdgpu_ras_block_object *block_obj = NULL;
1432 	int ret;
1433 
1434 	if (blk == AMDGPU_RAS_BLOCK_COUNT)
1435 		return -EINVAL;
1436 
1437 	if (error_query_mode == AMDGPU_RAS_INVALID_ERROR_QUERY)
1438 		return -EINVAL;
1439 
1440 	if (error_query_mode == AMDGPU_RAS_VIRT_ERROR_COUNT_QUERY) {
1441 		return amdgpu_virt_req_ras_err_count(adev, blk, err_data);
1442 	} else if (error_query_mode == AMDGPU_RAS_DIRECT_ERROR_QUERY) {
1443 		if (info->head.block == AMDGPU_RAS_BLOCK__UMC) {
1444 			amdgpu_ras_get_ecc_info(adev, err_data);
1445 		} else {
1446 			block_obj = amdgpu_ras_get_ras_block(adev, info->head.block, 0);
1447 			if (!block_obj || !block_obj->hw_ops) {
1448 				dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1449 					     get_ras_block_str(&info->head));
1450 				return -EINVAL;
1451 			}
1452 
1453 			if (block_obj->hw_ops->query_ras_error_count)
1454 				block_obj->hw_ops->query_ras_error_count(adev, err_data);
1455 
1456 			if ((info->head.block == AMDGPU_RAS_BLOCK__SDMA) ||
1457 			    (info->head.block == AMDGPU_RAS_BLOCK__GFX) ||
1458 			    (info->head.block == AMDGPU_RAS_BLOCK__MMHUB)) {
1459 				if (block_obj->hw_ops->query_ras_error_status)
1460 					block_obj->hw_ops->query_ras_error_status(adev);
1461 			}
1462 		}
1463 	} else {
1464 		if (amdgpu_aca_is_enabled(adev)) {
1465 			ret = amdgpu_aca_log_ras_error_data(adev, blk, ACA_ERROR_TYPE_UE, err_data, qctx);
1466 			if (ret)
1467 				return ret;
1468 
1469 			ret = amdgpu_aca_log_ras_error_data(adev, blk, ACA_ERROR_TYPE_CE, err_data, qctx);
1470 			if (ret)
1471 				return ret;
1472 
1473 			ret = amdgpu_aca_log_ras_error_data(adev, blk, ACA_ERROR_TYPE_DEFERRED, err_data, qctx);
1474 			if (ret)
1475 				return ret;
1476 		} else {
1477 			/* FIXME: add code to check return value later */
1478 			amdgpu_mca_smu_log_ras_error(adev, blk, AMDGPU_MCA_ERROR_TYPE_UE, err_data, qctx);
1479 			amdgpu_mca_smu_log_ras_error(adev, blk, AMDGPU_MCA_ERROR_TYPE_CE, err_data, qctx);
1480 		}
1481 	}
1482 
1483 	return 0;
1484 }
1485 
1486 /* query/inject/cure begin */
1487 static int amdgpu_ras_query_error_status_with_event(struct amdgpu_device *adev,
1488 						    struct ras_query_if *info,
1489 						    enum ras_event_type type)
1490 {
1491 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1492 	struct ras_err_data err_data;
1493 	struct ras_query_context qctx;
1494 	unsigned int error_query_mode;
1495 	int ret;
1496 
1497 	if (!obj)
1498 		return -EINVAL;
1499 
1500 	ret = amdgpu_ras_error_data_init(&err_data);
1501 	if (ret)
1502 		return ret;
1503 
1504 	if (!amdgpu_ras_get_error_query_mode(adev, &error_query_mode))
1505 		return -EINVAL;
1506 
1507 	memset(&qctx, 0, sizeof(qctx));
1508 	qctx.evid.type = type;
1509 	qctx.evid.event_id = amdgpu_ras_acquire_event_id(adev, type);
1510 
1511 	if (!down_read_trylock(&adev->reset_domain->sem)) {
1512 		ret = -EIO;
1513 		goto out_fini_err_data;
1514 	}
1515 
1516 	ret = amdgpu_ras_query_error_status_helper(adev, info,
1517 						   &err_data,
1518 						   &qctx,
1519 						   error_query_mode);
1520 	up_read(&adev->reset_domain->sem);
1521 	if (ret)
1522 		goto out_fini_err_data;
1523 
1524 	if (error_query_mode != AMDGPU_RAS_VIRT_ERROR_COUNT_QUERY) {
1525 		amdgpu_rasmgr_error_data_statistic_update(obj, &err_data);
1526 		amdgpu_ras_error_generate_report(adev, info, &err_data, &qctx);
1527 	} else {
1528 		/* Host provides absolute error counts. First generate the report
1529 		 * using the previous VF internal count against new host count.
1530 		 * Then Update VF internal count.
1531 		 */
1532 		amdgpu_ras_virt_error_generate_report(adev, info, &err_data, &qctx);
1533 		amdgpu_ras_mgr_virt_error_data_statistics_update(obj, &err_data);
1534 	}
1535 
1536 	info->ue_count = obj->err_data.ue_count;
1537 	info->ce_count = obj->err_data.ce_count;
1538 	info->de_count = obj->err_data.de_count;
1539 
1540 out_fini_err_data:
1541 	amdgpu_ras_error_data_fini(&err_data);
1542 
1543 	return ret;
1544 }
1545 
1546 int amdgpu_ras_query_error_status(struct amdgpu_device *adev, struct ras_query_if *info)
1547 {
1548 	return amdgpu_ras_query_error_status_with_event(adev, info, RAS_EVENT_TYPE_INVALID);
1549 }
1550 
1551 int amdgpu_ras_reset_error_count(struct amdgpu_device *adev,
1552 		enum amdgpu_ras_block block)
1553 {
1554 	struct amdgpu_ras_block_object *block_obj = amdgpu_ras_get_ras_block(adev, block, 0);
1555 	const struct amdgpu_mca_smu_funcs *mca_funcs = adev->mca.mca_funcs;
1556 	const struct aca_smu_funcs *smu_funcs = adev->aca.smu_funcs;
1557 
1558 	if (!block_obj || !block_obj->hw_ops) {
1559 		dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1560 				ras_block_str(block));
1561 		return -EOPNOTSUPP;
1562 	}
1563 
1564 	if (!amdgpu_ras_is_supported(adev, block) ||
1565 	    !amdgpu_ras_get_aca_debug_mode(adev))
1566 		return -EOPNOTSUPP;
1567 
1568 	if (amdgpu_sriov_vf(adev))
1569 		return -EOPNOTSUPP;
1570 
1571 	/* skip ras error reset in gpu reset */
1572 	if ((amdgpu_in_reset(adev) || amdgpu_ras_in_recovery(adev)) &&
1573 	    ((smu_funcs && smu_funcs->set_debug_mode) ||
1574 	     (mca_funcs && mca_funcs->mca_set_debug_mode)))
1575 		return -EOPNOTSUPP;
1576 
1577 	if (block_obj->hw_ops->reset_ras_error_count)
1578 		block_obj->hw_ops->reset_ras_error_count(adev);
1579 
1580 	return 0;
1581 }
1582 
1583 int amdgpu_ras_reset_error_status(struct amdgpu_device *adev,
1584 		enum amdgpu_ras_block block)
1585 {
1586 	struct amdgpu_ras_block_object *block_obj = amdgpu_ras_get_ras_block(adev, block, 0);
1587 
1588 	if (amdgpu_ras_reset_error_count(adev, block) == -EOPNOTSUPP)
1589 		return 0;
1590 
1591 	if ((block == AMDGPU_RAS_BLOCK__GFX) ||
1592 	    (block == AMDGPU_RAS_BLOCK__MMHUB)) {
1593 		if (block_obj->hw_ops->reset_ras_error_status)
1594 			block_obj->hw_ops->reset_ras_error_status(adev);
1595 	}
1596 
1597 	return 0;
1598 }
1599 
1600 /* wrapper of psp_ras_trigger_error */
1601 int amdgpu_ras_error_inject(struct amdgpu_device *adev,
1602 		struct ras_inject_if *info)
1603 {
1604 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1605 	struct ta_ras_trigger_error_input block_info = {
1606 		.block_id =  amdgpu_ras_block_to_ta(info->head.block),
1607 		.inject_error_type = amdgpu_ras_error_to_ta(info->head.type),
1608 		.sub_block_index = info->head.sub_block_index,
1609 		.address = info->address,
1610 		.value = info->value,
1611 	};
1612 	int ret = -EINVAL;
1613 	struct amdgpu_ras_block_object *block_obj = amdgpu_ras_get_ras_block(adev,
1614 							info->head.block,
1615 							info->head.sub_block_index);
1616 
1617 	/* inject on guest isn't allowed, return success directly */
1618 	if (amdgpu_sriov_vf(adev))
1619 		return 0;
1620 
1621 	if (!obj)
1622 		return -EINVAL;
1623 
1624 	if (!block_obj || !block_obj->hw_ops)	{
1625 		dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1626 			     get_ras_block_str(&info->head));
1627 		return -EINVAL;
1628 	}
1629 
1630 	/* Calculate XGMI relative offset */
1631 	if (adev->gmc.xgmi.num_physical_nodes > 1 &&
1632 	    info->head.block != AMDGPU_RAS_BLOCK__GFX) {
1633 		block_info.address =
1634 			amdgpu_xgmi_get_relative_phy_addr(adev,
1635 							  block_info.address);
1636 	}
1637 
1638 	if (block_obj->hw_ops->ras_error_inject) {
1639 		if (info->head.block == AMDGPU_RAS_BLOCK__GFX)
1640 			ret = block_obj->hw_ops->ras_error_inject(adev, info, info->instance_mask);
1641 		else /* Special ras_error_inject is defined (e.g: xgmi) */
1642 			ret = block_obj->hw_ops->ras_error_inject(adev, &block_info,
1643 						info->instance_mask);
1644 	} else {
1645 		/* default path */
1646 		ret = psp_ras_trigger_error(&adev->psp, &block_info, info->instance_mask);
1647 	}
1648 
1649 	if (ret)
1650 		dev_err(adev->dev, "ras inject %s failed %d\n",
1651 			get_ras_block_str(&info->head), ret);
1652 
1653 	return ret;
1654 }
1655 
1656 /**
1657  * amdgpu_ras_query_error_count_helper -- Get error counter for specific IP
1658  * @adev: pointer to AMD GPU device
1659  * @ce_count: pointer to an integer to be set to the count of correctible errors.
1660  * @ue_count: pointer to an integer to be set to the count of uncorrectible errors.
1661  * @query_info: pointer to ras_query_if
1662  *
1663  * Return 0 for query success or do nothing, otherwise return an error
1664  * on failures
1665  */
1666 static int amdgpu_ras_query_error_count_helper(struct amdgpu_device *adev,
1667 					       unsigned long *ce_count,
1668 					       unsigned long *ue_count,
1669 					       struct ras_query_if *query_info)
1670 {
1671 	int ret;
1672 
1673 	if (!query_info)
1674 		/* do nothing if query_info is not specified */
1675 		return 0;
1676 
1677 	ret = amdgpu_ras_query_error_status(adev, query_info);
1678 	if (ret)
1679 		return ret;
1680 
1681 	*ce_count += query_info->ce_count;
1682 	*ue_count += query_info->ue_count;
1683 
1684 	/* some hardware/IP supports read to clear
1685 	 * no need to explictly reset the err status after the query call */
1686 	if (amdgpu_ip_version(adev, MP0_HWIP, 0) != IP_VERSION(11, 0, 2) &&
1687 	    amdgpu_ip_version(adev, MP0_HWIP, 0) != IP_VERSION(11, 0, 4)) {
1688 		if (amdgpu_ras_reset_error_status(adev, query_info->head.block))
1689 			dev_warn(adev->dev,
1690 				 "Failed to reset error counter and error status\n");
1691 	}
1692 
1693 	return 0;
1694 }
1695 
1696 /**
1697  * amdgpu_ras_query_error_count -- Get error counts of all IPs or specific IP
1698  * @adev: pointer to AMD GPU device
1699  * @ce_count: pointer to an integer to be set to the count of correctible errors.
1700  * @ue_count: pointer to an integer to be set to the count of uncorrectible
1701  * errors.
1702  * @query_info: pointer to ras_query_if if the query request is only for
1703  * specific ip block; if info is NULL, then the qurey request is for
1704  * all the ip blocks that support query ras error counters/status
1705  *
1706  * If set, @ce_count or @ue_count, count and return the corresponding
1707  * error counts in those integer pointers. Return 0 if the device
1708  * supports RAS. Return -EOPNOTSUPP if the device doesn't support RAS.
1709  */
1710 int amdgpu_ras_query_error_count(struct amdgpu_device *adev,
1711 				 unsigned long *ce_count,
1712 				 unsigned long *ue_count,
1713 				 struct ras_query_if *query_info)
1714 {
1715 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1716 	struct ras_manager *obj;
1717 	unsigned long ce, ue;
1718 	int ret;
1719 
1720 	if (!adev->ras_enabled || !con)
1721 		return -EOPNOTSUPP;
1722 
1723 	/* Don't count since no reporting.
1724 	 */
1725 	if (!ce_count && !ue_count)
1726 		return 0;
1727 
1728 	ce = 0;
1729 	ue = 0;
1730 	if (!query_info) {
1731 		/* query all the ip blocks that support ras query interface */
1732 		list_for_each_entry(obj, &con->head, node) {
1733 			struct ras_query_if info = {
1734 				.head = obj->head,
1735 			};
1736 
1737 			ret = amdgpu_ras_query_error_count_helper(adev, &ce, &ue, &info);
1738 		}
1739 	} else {
1740 		/* query specific ip block */
1741 		ret = amdgpu_ras_query_error_count_helper(adev, &ce, &ue, query_info);
1742 	}
1743 
1744 	if (ret)
1745 		return ret;
1746 
1747 	if (ce_count)
1748 		*ce_count = ce;
1749 
1750 	if (ue_count)
1751 		*ue_count = ue;
1752 
1753 	return 0;
1754 }
1755 /* query/inject/cure end */
1756 
1757 
1758 /* sysfs begin */
1759 
1760 static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
1761 		struct ras_badpage **bps, unsigned int *count);
1762 
1763 static char *amdgpu_ras_badpage_flags_str(unsigned int flags)
1764 {
1765 	switch (flags) {
1766 	case AMDGPU_RAS_RETIRE_PAGE_RESERVED:
1767 		return "R";
1768 	case AMDGPU_RAS_RETIRE_PAGE_PENDING:
1769 		return "P";
1770 	case AMDGPU_RAS_RETIRE_PAGE_FAULT:
1771 	default:
1772 		return "F";
1773 	}
1774 }
1775 
1776 /**
1777  * DOC: AMDGPU RAS sysfs gpu_vram_bad_pages Interface
1778  *
1779  * It allows user to read the bad pages of vram on the gpu through
1780  * /sys/class/drm/card[0/1/2...]/device/ras/gpu_vram_bad_pages
1781  *
1782  * It outputs multiple lines, and each line stands for one gpu page.
1783  *
1784  * The format of one line is below,
1785  * gpu pfn : gpu page size : flags
1786  *
1787  * gpu pfn and gpu page size are printed in hex format.
1788  * flags can be one of below character,
1789  *
1790  * R: reserved, this gpu page is reserved and not able to use.
1791  *
1792  * P: pending for reserve, this gpu page is marked as bad, will be reserved
1793  * in next window of page_reserve.
1794  *
1795  * F: unable to reserve. this gpu page can't be reserved due to some reasons.
1796  *
1797  * Examples:
1798  *
1799  * .. code-block:: bash
1800  *
1801  *	0x00000001 : 0x00001000 : R
1802  *	0x00000002 : 0x00001000 : P
1803  *
1804  */
1805 
1806 static ssize_t amdgpu_ras_sysfs_badpages_read(struct file *f,
1807 		struct kobject *kobj, const struct bin_attribute *attr,
1808 		char *buf, loff_t ppos, size_t count)
1809 {
1810 	struct amdgpu_ras *con =
1811 		container_of(attr, struct amdgpu_ras, badpages_attr);
1812 	struct amdgpu_device *adev = con->adev;
1813 	const unsigned int element_size =
1814 		sizeof("0xabcdabcd : 0x12345678 : R\n") - 1;
1815 	unsigned int start = div64_ul(ppos + element_size - 1, element_size);
1816 	unsigned int end = div64_ul(ppos + count - 1, element_size);
1817 	ssize_t s = 0;
1818 	struct ras_badpage *bps = NULL;
1819 	unsigned int bps_count = 0;
1820 
1821 	memset(buf, 0, count);
1822 
1823 	if (amdgpu_ras_badpages_read(adev, &bps, &bps_count))
1824 		return 0;
1825 
1826 	for (; start < end && start < bps_count; start++)
1827 		s += scnprintf(&buf[s], element_size + 1,
1828 				"0x%08x : 0x%08x : %1s\n",
1829 				bps[start].bp,
1830 				bps[start].size,
1831 				amdgpu_ras_badpage_flags_str(bps[start].flags));
1832 
1833 	kfree(bps);
1834 
1835 	return s;
1836 }
1837 
1838 static ssize_t amdgpu_ras_sysfs_features_read(struct device *dev,
1839 		struct device_attribute *attr, char *buf)
1840 {
1841 	struct amdgpu_ras *con =
1842 		container_of(attr, struct amdgpu_ras, features_attr);
1843 
1844 	return sysfs_emit(buf, "feature mask: 0x%x\n", con->features);
1845 }
1846 
1847 static ssize_t amdgpu_ras_sysfs_version_show(struct device *dev,
1848 		struct device_attribute *attr, char *buf)
1849 {
1850 	struct amdgpu_ras *con =
1851 		container_of(attr, struct amdgpu_ras, version_attr);
1852 	return sysfs_emit(buf, "table version: 0x%x\n", con->eeprom_control.tbl_hdr.version);
1853 }
1854 
1855 static ssize_t amdgpu_ras_sysfs_schema_show(struct device *dev,
1856 		struct device_attribute *attr, char *buf)
1857 {
1858 	struct amdgpu_ras *con =
1859 		container_of(attr, struct amdgpu_ras, schema_attr);
1860 	return sysfs_emit(buf, "schema: 0x%x\n", con->schema);
1861 }
1862 
1863 static struct {
1864 	enum ras_event_type type;
1865 	const char *name;
1866 } dump_event[] = {
1867 	{RAS_EVENT_TYPE_FATAL, "Fatal Error"},
1868 	{RAS_EVENT_TYPE_POISON_CREATION, "Poison Creation"},
1869 	{RAS_EVENT_TYPE_POISON_CONSUMPTION, "Poison Consumption"},
1870 };
1871 
1872 static ssize_t amdgpu_ras_sysfs_event_state_show(struct device *dev,
1873 						 struct device_attribute *attr, char *buf)
1874 {
1875 	struct amdgpu_ras *con =
1876 		container_of(attr, struct amdgpu_ras, event_state_attr);
1877 	struct ras_event_manager *event_mgr = con->event_mgr;
1878 	struct ras_event_state *event_state;
1879 	int i, size = 0;
1880 
1881 	if (!event_mgr)
1882 		return -EINVAL;
1883 
1884 	size += sysfs_emit_at(buf, size, "current seqno: %llu\n", atomic64_read(&event_mgr->seqno));
1885 	for (i = 0; i < ARRAY_SIZE(dump_event); i++) {
1886 		event_state = &event_mgr->event_state[dump_event[i].type];
1887 		size += sysfs_emit_at(buf, size, "%s: count:%llu, last_seqno:%llu\n",
1888 				      dump_event[i].name,
1889 				      atomic64_read(&event_state->count),
1890 				      event_state->last_seqno);
1891 	}
1892 
1893 	return (ssize_t)size;
1894 }
1895 
1896 static void amdgpu_ras_sysfs_remove_bad_page_node(struct amdgpu_device *adev)
1897 {
1898 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1899 
1900 	if (adev->dev->kobj.sd)
1901 		sysfs_remove_file_from_group(&adev->dev->kobj,
1902 				&con->badpages_attr.attr,
1903 				RAS_FS_NAME);
1904 }
1905 
1906 static int amdgpu_ras_sysfs_remove_dev_attr_node(struct amdgpu_device *adev)
1907 {
1908 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1909 	struct attribute *attrs[] = {
1910 		&con->features_attr.attr,
1911 		&con->version_attr.attr,
1912 		&con->schema_attr.attr,
1913 		&con->event_state_attr.attr,
1914 		NULL
1915 	};
1916 	struct attribute_group group = {
1917 		.name = RAS_FS_NAME,
1918 		.attrs = attrs,
1919 	};
1920 
1921 	if (adev->dev->kobj.sd)
1922 		sysfs_remove_group(&adev->dev->kobj, &group);
1923 
1924 	return 0;
1925 }
1926 
1927 int amdgpu_ras_sysfs_create(struct amdgpu_device *adev,
1928 		struct ras_common_if *head)
1929 {
1930 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1931 
1932 	if (amdgpu_aca_is_enabled(adev))
1933 		return 0;
1934 
1935 	if (!obj || obj->attr_inuse)
1936 		return -EINVAL;
1937 
1938 	if (amdgpu_sriov_vf(adev) && !amdgpu_virt_ras_telemetry_block_en(adev, head->block))
1939 		return 0;
1940 
1941 	get_obj(obj);
1942 
1943 	snprintf(obj->fs_data.sysfs_name, sizeof(obj->fs_data.sysfs_name),
1944 		"%s_err_count", head->name);
1945 
1946 	obj->sysfs_attr = (struct device_attribute){
1947 		.attr = {
1948 			.name = obj->fs_data.sysfs_name,
1949 			.mode = S_IRUGO,
1950 		},
1951 			.show = amdgpu_ras_sysfs_read,
1952 	};
1953 	sysfs_attr_init(&obj->sysfs_attr.attr);
1954 
1955 	if (sysfs_add_file_to_group(&adev->dev->kobj,
1956 				&obj->sysfs_attr.attr,
1957 				RAS_FS_NAME)) {
1958 		put_obj(obj);
1959 		return -EINVAL;
1960 	}
1961 
1962 	obj->attr_inuse = 1;
1963 
1964 	return 0;
1965 }
1966 
1967 int amdgpu_ras_sysfs_remove(struct amdgpu_device *adev,
1968 		struct ras_common_if *head)
1969 {
1970 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1971 
1972 	if (amdgpu_aca_is_enabled(adev))
1973 		return 0;
1974 
1975 	if (!obj || !obj->attr_inuse)
1976 		return -EINVAL;
1977 
1978 	if (adev->dev->kobj.sd)
1979 		sysfs_remove_file_from_group(&adev->dev->kobj,
1980 				&obj->sysfs_attr.attr,
1981 				RAS_FS_NAME);
1982 	obj->attr_inuse = 0;
1983 	put_obj(obj);
1984 
1985 	return 0;
1986 }
1987 
1988 static int amdgpu_ras_sysfs_remove_all(struct amdgpu_device *adev)
1989 {
1990 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1991 	struct ras_manager *obj, *tmp;
1992 
1993 	list_for_each_entry_safe(obj, tmp, &con->head, node) {
1994 		amdgpu_ras_sysfs_remove(adev, &obj->head);
1995 	}
1996 
1997 	if (amdgpu_bad_page_threshold != 0)
1998 		amdgpu_ras_sysfs_remove_bad_page_node(adev);
1999 
2000 	amdgpu_ras_sysfs_remove_dev_attr_node(adev);
2001 
2002 	return 0;
2003 }
2004 /* sysfs end */
2005 
2006 /**
2007  * DOC: AMDGPU RAS Reboot Behavior for Unrecoverable Errors
2008  *
2009  * Normally when there is an uncorrectable error, the driver will reset
2010  * the GPU to recover.  However, in the event of an unrecoverable error,
2011  * the driver provides an interface to reboot the system automatically
2012  * in that event.
2013  *
2014  * The following file in debugfs provides that interface:
2015  * /sys/kernel/debug/dri/[0/1/2...]/ras/auto_reboot
2016  *
2017  * Usage:
2018  *
2019  * .. code-block:: bash
2020  *
2021  *	echo true > .../ras/auto_reboot
2022  *
2023  */
2024 /* debugfs begin */
2025 static struct dentry *amdgpu_ras_debugfs_create_ctrl_node(struct amdgpu_device *adev)
2026 {
2027 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2028 	struct amdgpu_ras_eeprom_control *eeprom = &con->eeprom_control;
2029 	struct drm_minor  *minor = adev_to_drm(adev)->primary;
2030 	struct dentry     *dir;
2031 
2032 	dir = debugfs_create_dir(RAS_FS_NAME, minor->debugfs_root);
2033 	debugfs_create_file("ras_ctrl", S_IWUGO | S_IRUGO, dir, adev,
2034 			    &amdgpu_ras_debugfs_ctrl_ops);
2035 	debugfs_create_file("ras_eeprom_reset", S_IWUGO | S_IRUGO, dir, adev,
2036 			    &amdgpu_ras_debugfs_eeprom_ops);
2037 	debugfs_create_u32("bad_page_cnt_threshold", 0444, dir,
2038 			   &con->bad_page_cnt_threshold);
2039 	debugfs_create_u32("ras_num_recs", 0444, dir, &eeprom->ras_num_recs);
2040 	debugfs_create_x32("ras_hw_enabled", 0444, dir, &adev->ras_hw_enabled);
2041 	debugfs_create_x32("ras_enabled", 0444, dir, &adev->ras_enabled);
2042 	debugfs_create_file("ras_eeprom_size", S_IRUGO, dir, adev,
2043 			    &amdgpu_ras_debugfs_eeprom_size_ops);
2044 	con->de_ras_eeprom_table = debugfs_create_file("ras_eeprom_table",
2045 						       S_IRUGO, dir, adev,
2046 						       &amdgpu_ras_debugfs_eeprom_table_ops);
2047 	amdgpu_ras_debugfs_set_ret_size(&con->eeprom_control);
2048 
2049 	/*
2050 	 * After one uncorrectable error happens, usually GPU recovery will
2051 	 * be scheduled. But due to the known problem in GPU recovery failing
2052 	 * to bring GPU back, below interface provides one direct way to
2053 	 * user to reboot system automatically in such case within
2054 	 * ERREVENT_ATHUB_INTERRUPT generated. Normal GPU recovery routine
2055 	 * will never be called.
2056 	 */
2057 	debugfs_create_bool("auto_reboot", S_IWUGO | S_IRUGO, dir, &con->reboot);
2058 
2059 	/*
2060 	 * User could set this not to clean up hardware's error count register
2061 	 * of RAS IPs during ras recovery.
2062 	 */
2063 	debugfs_create_bool("disable_ras_err_cnt_harvest", 0644, dir,
2064 			    &con->disable_ras_err_cnt_harvest);
2065 	return dir;
2066 }
2067 
2068 static void amdgpu_ras_debugfs_create(struct amdgpu_device *adev,
2069 				      struct ras_fs_if *head,
2070 				      struct dentry *dir)
2071 {
2072 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head->head);
2073 
2074 	if (!obj || !dir)
2075 		return;
2076 
2077 	get_obj(obj);
2078 
2079 	memcpy(obj->fs_data.debugfs_name,
2080 			head->debugfs_name,
2081 			sizeof(obj->fs_data.debugfs_name));
2082 
2083 	debugfs_create_file(obj->fs_data.debugfs_name, S_IWUGO | S_IRUGO, dir,
2084 			    obj, &amdgpu_ras_debugfs_ops);
2085 }
2086 
2087 static bool amdgpu_ras_aca_is_supported(struct amdgpu_device *adev)
2088 {
2089 	bool ret;
2090 
2091 	switch (amdgpu_ip_version(adev, MP0_HWIP, 0)) {
2092 	case IP_VERSION(13, 0, 6):
2093 	case IP_VERSION(13, 0, 12):
2094 	case IP_VERSION(13, 0, 14):
2095 		ret = true;
2096 		break;
2097 	default:
2098 		ret = false;
2099 		break;
2100 	}
2101 
2102 	return ret;
2103 }
2104 
2105 void amdgpu_ras_debugfs_create_all(struct amdgpu_device *adev)
2106 {
2107 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2108 	struct dentry *dir;
2109 	struct ras_manager *obj;
2110 	struct ras_fs_if fs_info;
2111 
2112 	/*
2113 	 * it won't be called in resume path, no need to check
2114 	 * suspend and gpu reset status
2115 	 */
2116 	if (!IS_ENABLED(CONFIG_DEBUG_FS) || !con)
2117 		return;
2118 
2119 	dir = amdgpu_ras_debugfs_create_ctrl_node(adev);
2120 
2121 	list_for_each_entry(obj, &con->head, node) {
2122 		if (amdgpu_ras_is_supported(adev, obj->head.block) &&
2123 			(obj->attr_inuse == 1)) {
2124 			sprintf(fs_info.debugfs_name, "%s_err_inject",
2125 					get_ras_block_str(&obj->head));
2126 			fs_info.head = obj->head;
2127 			amdgpu_ras_debugfs_create(adev, &fs_info, dir);
2128 		}
2129 	}
2130 
2131 	if (amdgpu_ras_aca_is_supported(adev)) {
2132 		if (amdgpu_aca_is_enabled(adev))
2133 			amdgpu_aca_smu_debugfs_init(adev, dir);
2134 		else
2135 			amdgpu_mca_smu_debugfs_init(adev, dir);
2136 	}
2137 }
2138 
2139 /* debugfs end */
2140 
2141 /* ras fs */
2142 static const BIN_ATTR(gpu_vram_bad_pages, S_IRUGO,
2143 		      amdgpu_ras_sysfs_badpages_read, NULL, 0);
2144 static DEVICE_ATTR(features, S_IRUGO,
2145 		amdgpu_ras_sysfs_features_read, NULL);
2146 static DEVICE_ATTR(version, 0444,
2147 		amdgpu_ras_sysfs_version_show, NULL);
2148 static DEVICE_ATTR(schema, 0444,
2149 		amdgpu_ras_sysfs_schema_show, NULL);
2150 static DEVICE_ATTR(event_state, 0444,
2151 		   amdgpu_ras_sysfs_event_state_show, NULL);
2152 static int amdgpu_ras_fs_init(struct amdgpu_device *adev)
2153 {
2154 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2155 	struct attribute_group group = {
2156 		.name = RAS_FS_NAME,
2157 	};
2158 	struct attribute *attrs[] = {
2159 		&con->features_attr.attr,
2160 		&con->version_attr.attr,
2161 		&con->schema_attr.attr,
2162 		&con->event_state_attr.attr,
2163 		NULL
2164 	};
2165 	const struct bin_attribute *bin_attrs[] = {
2166 		NULL,
2167 		NULL,
2168 	};
2169 	int r;
2170 
2171 	group.attrs = attrs;
2172 
2173 	/* add features entry */
2174 	con->features_attr = dev_attr_features;
2175 	sysfs_attr_init(attrs[0]);
2176 
2177 	/* add version entry */
2178 	con->version_attr = dev_attr_version;
2179 	sysfs_attr_init(attrs[1]);
2180 
2181 	/* add schema entry */
2182 	con->schema_attr = dev_attr_schema;
2183 	sysfs_attr_init(attrs[2]);
2184 
2185 	/* add event_state entry */
2186 	con->event_state_attr = dev_attr_event_state;
2187 	sysfs_attr_init(attrs[3]);
2188 
2189 	if (amdgpu_bad_page_threshold != 0) {
2190 		/* add bad_page_features entry */
2191 		con->badpages_attr = bin_attr_gpu_vram_bad_pages;
2192 		sysfs_bin_attr_init(&con->badpages_attr);
2193 		bin_attrs[0] = &con->badpages_attr;
2194 		group.bin_attrs = bin_attrs;
2195 	}
2196 
2197 	r = sysfs_create_group(&adev->dev->kobj, &group);
2198 	if (r)
2199 		dev_err(adev->dev, "Failed to create RAS sysfs group!");
2200 
2201 	return 0;
2202 }
2203 
2204 static int amdgpu_ras_fs_fini(struct amdgpu_device *adev)
2205 {
2206 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2207 	struct ras_manager *con_obj, *ip_obj, *tmp;
2208 
2209 	if (IS_ENABLED(CONFIG_DEBUG_FS)) {
2210 		list_for_each_entry_safe(con_obj, tmp, &con->head, node) {
2211 			ip_obj = amdgpu_ras_find_obj(adev, &con_obj->head);
2212 			if (ip_obj)
2213 				put_obj(ip_obj);
2214 		}
2215 	}
2216 
2217 	amdgpu_ras_sysfs_remove_all(adev);
2218 	return 0;
2219 }
2220 /* ras fs end */
2221 
2222 /* ih begin */
2223 
2224 /* For the hardware that cannot enable bif ring for both ras_controller_irq
2225  * and ras_err_evnet_athub_irq ih cookies, the driver has to poll status
2226  * register to check whether the interrupt is triggered or not, and properly
2227  * ack the interrupt if it is there
2228  */
2229 void amdgpu_ras_interrupt_fatal_error_handler(struct amdgpu_device *adev)
2230 {
2231 	/* Fatal error events are handled on host side */
2232 	if (amdgpu_sriov_vf(adev))
2233 		return;
2234 	/*
2235 	 * If the current interrupt is caused by a non-fatal RAS error, skip
2236 	 * check for fatal error. For fatal errors, FED status of all devices
2237 	 * in XGMI hive gets set when the first device gets fatal error
2238 	 * interrupt. The error gets propagated to other devices as well, so
2239 	 * make sure to ack the interrupt regardless of FED status.
2240 	 */
2241 	if (!amdgpu_ras_get_fed_status(adev) &&
2242 	    amdgpu_ras_is_err_state(adev, AMDGPU_RAS_BLOCK__ANY))
2243 		return;
2244 
2245 	if (amdgpu_uniras_enabled(adev)) {
2246 		amdgpu_ras_mgr_handle_fatal_interrupt(adev, NULL);
2247 		return;
2248 	}
2249 
2250 	if (adev->nbio.ras &&
2251 	    adev->nbio.ras->handle_ras_controller_intr_no_bifring)
2252 		adev->nbio.ras->handle_ras_controller_intr_no_bifring(adev);
2253 
2254 	if (adev->nbio.ras &&
2255 	    adev->nbio.ras->handle_ras_err_event_athub_intr_no_bifring)
2256 		adev->nbio.ras->handle_ras_err_event_athub_intr_no_bifring(adev);
2257 }
2258 
2259 static void amdgpu_ras_interrupt_poison_consumption_handler(struct ras_manager *obj,
2260 				struct amdgpu_iv_entry *entry)
2261 {
2262 	bool poison_stat = false;
2263 	struct amdgpu_device *adev = obj->adev;
2264 	struct amdgpu_ras_block_object *block_obj =
2265 		amdgpu_ras_get_ras_block(adev, obj->head.block, 0);
2266 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2267 	enum ras_event_type type = RAS_EVENT_TYPE_POISON_CONSUMPTION;
2268 	u64 event_id;
2269 	int ret;
2270 
2271 	if (!block_obj || !con)
2272 		return;
2273 
2274 	ret = amdgpu_ras_mark_ras_event(adev, type);
2275 	if (ret)
2276 		return;
2277 
2278 	amdgpu_ras_set_err_poison(adev, block_obj->ras_comm.block);
2279 	/* both query_poison_status and handle_poison_consumption are optional,
2280 	 * but at least one of them should be implemented if we need poison
2281 	 * consumption handler
2282 	 */
2283 	if (block_obj->hw_ops && block_obj->hw_ops->query_poison_status) {
2284 		poison_stat = block_obj->hw_ops->query_poison_status(adev);
2285 		if (!poison_stat) {
2286 			/* Not poison consumption interrupt, no need to handle it */
2287 			dev_info(adev->dev, "No RAS poison status in %s poison IH.\n",
2288 					block_obj->ras_comm.name);
2289 
2290 			return;
2291 		}
2292 	}
2293 
2294 	amdgpu_umc_poison_handler(adev, obj->head.block, 0);
2295 
2296 	if (block_obj->hw_ops && block_obj->hw_ops->handle_poison_consumption)
2297 		poison_stat = block_obj->hw_ops->handle_poison_consumption(adev);
2298 
2299 	/* gpu reset is fallback for failed and default cases.
2300 	 * For RMA case, amdgpu_umc_poison_handler will handle gpu reset.
2301 	 */
2302 	if (poison_stat && !amdgpu_ras_is_rma(adev)) {
2303 		event_id = amdgpu_ras_acquire_event_id(adev, type);
2304 		RAS_EVENT_LOG(adev, event_id,
2305 			      "GPU reset for %s RAS poison consumption is issued!\n",
2306 			      block_obj->ras_comm.name);
2307 		amdgpu_ras_reset_gpu(adev);
2308 	}
2309 
2310 	if (!poison_stat)
2311 		amdgpu_gfx_poison_consumption_handler(adev, entry);
2312 }
2313 
2314 static void amdgpu_ras_interrupt_poison_creation_handler(struct ras_manager *obj,
2315 				struct amdgpu_iv_entry *entry)
2316 {
2317 	struct amdgpu_device *adev = obj->adev;
2318 	enum ras_event_type type = RAS_EVENT_TYPE_POISON_CREATION;
2319 	u64 event_id;
2320 	int ret;
2321 
2322 	ret = amdgpu_ras_mark_ras_event(adev, type);
2323 	if (ret)
2324 		return;
2325 
2326 	event_id = amdgpu_ras_acquire_event_id(adev, type);
2327 	RAS_EVENT_LOG(adev, event_id, "Poison is created\n");
2328 
2329 	if (amdgpu_ip_version(obj->adev, UMC_HWIP, 0) >= IP_VERSION(12, 0, 0)) {
2330 		struct amdgpu_ras *con = amdgpu_ras_get_context(obj->adev);
2331 
2332 		atomic_inc(&con->page_retirement_req_cnt);
2333 		atomic_inc(&con->poison_creation_count);
2334 
2335 		wake_up(&con->page_retirement_wq);
2336 	}
2337 }
2338 
2339 static void amdgpu_ras_interrupt_umc_handler(struct ras_manager *obj,
2340 				struct amdgpu_iv_entry *entry)
2341 {
2342 	struct ras_ih_data *data = &obj->ih_data;
2343 	struct ras_err_data err_data;
2344 	int ret;
2345 
2346 	if (!data->cb)
2347 		return;
2348 
2349 	ret = amdgpu_ras_error_data_init(&err_data);
2350 	if (ret)
2351 		return;
2352 
2353 	/* Let IP handle its data, maybe we need get the output
2354 	 * from the callback to update the error type/count, etc
2355 	 */
2356 	amdgpu_ras_set_fed(obj->adev, true);
2357 	ret = data->cb(obj->adev, &err_data, entry);
2358 	/* ue will trigger an interrupt, and in that case
2359 	 * we need do a reset to recovery the whole system.
2360 	 * But leave IP do that recovery, here we just dispatch
2361 	 * the error.
2362 	 */
2363 	if (ret == AMDGPU_RAS_SUCCESS) {
2364 		/* these counts could be left as 0 if
2365 		 * some blocks do not count error number
2366 		 */
2367 		obj->err_data.ue_count += err_data.ue_count;
2368 		obj->err_data.ce_count += err_data.ce_count;
2369 		obj->err_data.de_count += err_data.de_count;
2370 	}
2371 
2372 	amdgpu_ras_error_data_fini(&err_data);
2373 }
2374 
2375 static void amdgpu_ras_interrupt_handler(struct ras_manager *obj)
2376 {
2377 	struct ras_ih_data *data = &obj->ih_data;
2378 	struct amdgpu_iv_entry entry;
2379 
2380 	while (data->rptr != data->wptr) {
2381 		rmb();
2382 		memcpy(&entry, &data->ring[data->rptr],
2383 				data->element_size);
2384 
2385 		wmb();
2386 		data->rptr = (data->aligned_element_size +
2387 				data->rptr) % data->ring_size;
2388 
2389 		if (amdgpu_ras_is_poison_mode_supported(obj->adev)) {
2390 			if (obj->head.block == AMDGPU_RAS_BLOCK__UMC)
2391 				amdgpu_ras_interrupt_poison_creation_handler(obj, &entry);
2392 			else
2393 				amdgpu_ras_interrupt_poison_consumption_handler(obj, &entry);
2394 		} else {
2395 			if (obj->head.block == AMDGPU_RAS_BLOCK__UMC)
2396 				amdgpu_ras_interrupt_umc_handler(obj, &entry);
2397 			else
2398 				dev_warn(obj->adev->dev,
2399 					"No RAS interrupt handler for non-UMC block with poison disabled.\n");
2400 		}
2401 	}
2402 }
2403 
2404 static void amdgpu_ras_interrupt_process_handler(struct work_struct *work)
2405 {
2406 	struct ras_ih_data *data =
2407 		container_of(work, struct ras_ih_data, ih_work);
2408 	struct ras_manager *obj =
2409 		container_of(data, struct ras_manager, ih_data);
2410 
2411 	amdgpu_ras_interrupt_handler(obj);
2412 }
2413 
2414 int amdgpu_ras_interrupt_dispatch(struct amdgpu_device *adev,
2415 		struct ras_dispatch_if *info)
2416 {
2417 	struct ras_manager *obj;
2418 	struct ras_ih_data *data;
2419 
2420 	if (amdgpu_uniras_enabled(adev)) {
2421 		struct ras_ih_info ih_info;
2422 
2423 		memset(&ih_info, 0, sizeof(ih_info));
2424 		ih_info.block = info->head.block;
2425 		memcpy(&ih_info.iv_entry, info->entry, sizeof(struct amdgpu_iv_entry));
2426 
2427 		return amdgpu_ras_mgr_handle_controller_interrupt(adev, &ih_info);
2428 	}
2429 
2430 	obj = amdgpu_ras_find_obj(adev, &info->head);
2431 	if (!obj)
2432 		return -EINVAL;
2433 
2434 	data = &obj->ih_data;
2435 
2436 	if (data->inuse == 0)
2437 		return 0;
2438 
2439 	/* Might be overflow... */
2440 	memcpy(&data->ring[data->wptr], info->entry,
2441 			data->element_size);
2442 
2443 	wmb();
2444 	data->wptr = (data->aligned_element_size +
2445 			data->wptr) % data->ring_size;
2446 
2447 	schedule_work(&data->ih_work);
2448 
2449 	return 0;
2450 }
2451 
2452 int amdgpu_ras_interrupt_remove_handler(struct amdgpu_device *adev,
2453 		struct ras_common_if *head)
2454 {
2455 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
2456 	struct ras_ih_data *data;
2457 
2458 	if (!obj)
2459 		return -EINVAL;
2460 
2461 	data = &obj->ih_data;
2462 	if (data->inuse == 0)
2463 		return 0;
2464 
2465 	cancel_work_sync(&data->ih_work);
2466 
2467 	kfree(data->ring);
2468 	memset(data, 0, sizeof(*data));
2469 	put_obj(obj);
2470 
2471 	return 0;
2472 }
2473 
2474 int amdgpu_ras_interrupt_add_handler(struct amdgpu_device *adev,
2475 		struct ras_common_if *head)
2476 {
2477 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
2478 	struct ras_ih_data *data;
2479 	struct amdgpu_ras_block_object *ras_obj;
2480 
2481 	if (!obj) {
2482 		/* in case we registe the IH before enable ras feature */
2483 		obj = amdgpu_ras_create_obj(adev, head);
2484 		if (!obj)
2485 			return -EINVAL;
2486 	} else
2487 		get_obj(obj);
2488 
2489 	ras_obj = container_of(head, struct amdgpu_ras_block_object, ras_comm);
2490 
2491 	data = &obj->ih_data;
2492 	/* add the callback.etc */
2493 	*data = (struct ras_ih_data) {
2494 		.inuse = 0,
2495 		.cb = ras_obj->ras_cb,
2496 		.element_size = sizeof(struct amdgpu_iv_entry),
2497 		.rptr = 0,
2498 		.wptr = 0,
2499 	};
2500 
2501 	INIT_WORK(&data->ih_work, amdgpu_ras_interrupt_process_handler);
2502 
2503 	data->aligned_element_size = ALIGN(data->element_size, 8);
2504 	/* the ring can store 64 iv entries. */
2505 	data->ring_size = 64 * data->aligned_element_size;
2506 	data->ring = kmalloc(data->ring_size, GFP_KERNEL);
2507 	if (!data->ring) {
2508 		put_obj(obj);
2509 		return -ENOMEM;
2510 	}
2511 
2512 	/* IH is ready */
2513 	data->inuse = 1;
2514 
2515 	return 0;
2516 }
2517 
2518 static int amdgpu_ras_interrupt_remove_all(struct amdgpu_device *adev)
2519 {
2520 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2521 	struct ras_manager *obj, *tmp;
2522 
2523 	list_for_each_entry_safe(obj, tmp, &con->head, node) {
2524 		amdgpu_ras_interrupt_remove_handler(adev, &obj->head);
2525 	}
2526 
2527 	return 0;
2528 }
2529 /* ih end */
2530 
2531 /* traversal all IPs except NBIO to query error counter */
2532 static void amdgpu_ras_log_on_err_counter(struct amdgpu_device *adev, enum ras_event_type type)
2533 {
2534 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2535 	struct ras_manager *obj;
2536 
2537 	if (!adev->ras_enabled || !con)
2538 		return;
2539 
2540 	list_for_each_entry(obj, &con->head, node) {
2541 		struct ras_query_if info = {
2542 			.head = obj->head,
2543 		};
2544 
2545 		/*
2546 		 * PCIE_BIF IP has one different isr by ras controller
2547 		 * interrupt, the specific ras counter query will be
2548 		 * done in that isr. So skip such block from common
2549 		 * sync flood interrupt isr calling.
2550 		 */
2551 		if (info.head.block == AMDGPU_RAS_BLOCK__PCIE_BIF)
2552 			continue;
2553 
2554 		/*
2555 		 * this is a workaround for aldebaran, skip send msg to
2556 		 * smu to get ecc_info table due to smu handle get ecc
2557 		 * info table failed temporarily.
2558 		 * should be removed until smu fix handle ecc_info table.
2559 		 */
2560 		if ((info.head.block == AMDGPU_RAS_BLOCK__UMC) &&
2561 		    (amdgpu_ip_version(adev, MP1_HWIP, 0) ==
2562 		     IP_VERSION(13, 0, 2)))
2563 			continue;
2564 
2565 		amdgpu_ras_query_error_status_with_event(adev, &info, type);
2566 
2567 		if (amdgpu_ip_version(adev, MP0_HWIP, 0) !=
2568 			    IP_VERSION(11, 0, 2) &&
2569 		    amdgpu_ip_version(adev, MP0_HWIP, 0) !=
2570 			    IP_VERSION(11, 0, 4) &&
2571 		    amdgpu_ip_version(adev, MP0_HWIP, 0) !=
2572 			    IP_VERSION(13, 0, 0)) {
2573 			if (amdgpu_ras_reset_error_status(adev, info.head.block))
2574 				dev_warn(adev->dev, "Failed to reset error counter and error status");
2575 		}
2576 	}
2577 }
2578 
2579 /* Parse RdRspStatus and WrRspStatus */
2580 static void amdgpu_ras_error_status_query(struct amdgpu_device *adev,
2581 					  struct ras_query_if *info)
2582 {
2583 	struct amdgpu_ras_block_object *block_obj;
2584 	/*
2585 	 * Only two block need to query read/write
2586 	 * RspStatus at current state
2587 	 */
2588 	if ((info->head.block != AMDGPU_RAS_BLOCK__GFX) &&
2589 		(info->head.block != AMDGPU_RAS_BLOCK__MMHUB))
2590 		return;
2591 
2592 	block_obj = amdgpu_ras_get_ras_block(adev,
2593 					info->head.block,
2594 					info->head.sub_block_index);
2595 
2596 	if (!block_obj || !block_obj->hw_ops) {
2597 		dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
2598 			     get_ras_block_str(&info->head));
2599 		return;
2600 	}
2601 
2602 	if (block_obj->hw_ops->query_ras_error_status)
2603 		block_obj->hw_ops->query_ras_error_status(adev);
2604 
2605 }
2606 
2607 static void amdgpu_ras_query_err_status(struct amdgpu_device *adev)
2608 {
2609 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2610 	struct ras_manager *obj;
2611 
2612 	if (!adev->ras_enabled || !con)
2613 		return;
2614 
2615 	list_for_each_entry(obj, &con->head, node) {
2616 		struct ras_query_if info = {
2617 			.head = obj->head,
2618 		};
2619 
2620 		amdgpu_ras_error_status_query(adev, &info);
2621 	}
2622 }
2623 
2624 /* recovery begin */
2625 
2626 /* return 0 on success.
2627  * caller need free bps.
2628  */
2629 static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
2630 		struct ras_badpage **bps, unsigned int *count)
2631 {
2632 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2633 	struct ras_err_handler_data *data;
2634 	int i = 0;
2635 	int ret = 0, status;
2636 
2637 	if (!con || !con->eh_data || !bps || !count)
2638 		return -EINVAL;
2639 
2640 	mutex_lock(&con->recovery_lock);
2641 	data = con->eh_data;
2642 	if (!data || data->count == 0) {
2643 		*bps = NULL;
2644 		ret = -EINVAL;
2645 		goto out;
2646 	}
2647 
2648 	*bps = kmalloc_array(data->count, sizeof(struct ras_badpage), GFP_KERNEL);
2649 	if (!*bps) {
2650 		ret = -ENOMEM;
2651 		goto out;
2652 	}
2653 
2654 	for (; i < data->count; i++) {
2655 		if (!data->bps[i].ts)
2656 			continue;
2657 
2658 		(*bps)[i] = (struct ras_badpage){
2659 			.bp = data->bps[i].retired_page,
2660 			.size = AMDGPU_GPU_PAGE_SIZE,
2661 			.flags = AMDGPU_RAS_RETIRE_PAGE_RESERVED,
2662 		};
2663 
2664 		if (amdgpu_ras_check_critical_address(adev,
2665 			data->bps[i].retired_page << AMDGPU_GPU_PAGE_SHIFT))
2666 			continue;
2667 
2668 		status = amdgpu_vram_mgr_query_page_status(&adev->mman.vram_mgr,
2669 				data->bps[i].retired_page << AMDGPU_GPU_PAGE_SHIFT);
2670 		if (status == -EBUSY)
2671 			(*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_PENDING;
2672 		else if (status == -ENOENT)
2673 			(*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_FAULT;
2674 	}
2675 
2676 	*count = con->bad_page_num;
2677 out:
2678 	mutex_unlock(&con->recovery_lock);
2679 	return ret;
2680 }
2681 
2682 static void amdgpu_ras_set_fed_all(struct amdgpu_device *adev,
2683 				   struct amdgpu_hive_info *hive, bool status)
2684 {
2685 	struct amdgpu_device *tmp_adev;
2686 
2687 	if (hive) {
2688 		list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head)
2689 			amdgpu_ras_set_fed(tmp_adev, status);
2690 	} else {
2691 		amdgpu_ras_set_fed(adev, status);
2692 	}
2693 }
2694 
2695 bool amdgpu_ras_in_recovery(struct amdgpu_device *adev)
2696 {
2697 	struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
2698 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
2699 	int hive_ras_recovery = 0;
2700 
2701 	if (hive) {
2702 		hive_ras_recovery = atomic_read(&hive->ras_recovery);
2703 		amdgpu_put_xgmi_hive(hive);
2704 	}
2705 
2706 	if (ras && (atomic_read(&ras->in_recovery) || hive_ras_recovery))
2707 		return true;
2708 
2709 	return false;
2710 }
2711 
2712 static enum ras_event_type amdgpu_ras_get_fatal_error_event(struct amdgpu_device *adev)
2713 {
2714 	if (amdgpu_ras_intr_triggered())
2715 		return RAS_EVENT_TYPE_FATAL;
2716 	else
2717 		return RAS_EVENT_TYPE_POISON_CONSUMPTION;
2718 }
2719 
2720 static void amdgpu_ras_do_recovery(struct work_struct *work)
2721 {
2722 	struct amdgpu_ras *ras =
2723 		container_of(work, struct amdgpu_ras, recovery_work);
2724 	struct amdgpu_device *remote_adev = NULL;
2725 	struct amdgpu_device *adev = ras->adev;
2726 	struct list_head device_list, *device_list_handle =  NULL;
2727 	struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
2728 	unsigned int error_query_mode;
2729 	enum ras_event_type type;
2730 
2731 	if (hive) {
2732 		atomic_set(&hive->ras_recovery, 1);
2733 
2734 		/* If any device which is part of the hive received RAS fatal
2735 		 * error interrupt, set fatal error status on all. This
2736 		 * condition will need a recovery, and flag will be cleared
2737 		 * as part of recovery.
2738 		 */
2739 		list_for_each_entry(remote_adev, &hive->device_list,
2740 				    gmc.xgmi.head)
2741 			if (amdgpu_ras_get_fed_status(remote_adev)) {
2742 				amdgpu_ras_set_fed_all(adev, hive, true);
2743 				break;
2744 			}
2745 	}
2746 	if (!ras->disable_ras_err_cnt_harvest) {
2747 
2748 		/* Build list of devices to query RAS related errors */
2749 		if  (hive && adev->gmc.xgmi.num_physical_nodes > 1) {
2750 			device_list_handle = &hive->device_list;
2751 		} else {
2752 			INIT_LIST_HEAD(&device_list);
2753 			list_add_tail(&adev->gmc.xgmi.head, &device_list);
2754 			device_list_handle = &device_list;
2755 		}
2756 
2757 		if (amdgpu_ras_get_error_query_mode(adev, &error_query_mode)) {
2758 			if (error_query_mode == AMDGPU_RAS_FIRMWARE_ERROR_QUERY) {
2759 				/* wait 500ms to ensure pmfw polling mca bank info done */
2760 				msleep(500);
2761 			}
2762 		}
2763 
2764 		type = amdgpu_ras_get_fatal_error_event(adev);
2765 		list_for_each_entry(remote_adev,
2766 				device_list_handle, gmc.xgmi.head) {
2767 			amdgpu_ras_query_err_status(remote_adev);
2768 			amdgpu_ras_log_on_err_counter(remote_adev, type);
2769 		}
2770 
2771 	}
2772 
2773 	if (amdgpu_device_should_recover_gpu(ras->adev)) {
2774 		struct amdgpu_reset_context reset_context;
2775 		memset(&reset_context, 0, sizeof(reset_context));
2776 
2777 		reset_context.method = AMD_RESET_METHOD_NONE;
2778 		reset_context.reset_req_dev = adev;
2779 		reset_context.src = AMDGPU_RESET_SRC_RAS;
2780 		set_bit(AMDGPU_SKIP_COREDUMP, &reset_context.flags);
2781 
2782 		/* Perform full reset in fatal error mode */
2783 		if (!amdgpu_ras_is_poison_mode_supported(ras->adev))
2784 			set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
2785 		else {
2786 			clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
2787 
2788 			if (ras->gpu_reset_flags & AMDGPU_RAS_GPU_RESET_MODE2_RESET) {
2789 				ras->gpu_reset_flags &= ~AMDGPU_RAS_GPU_RESET_MODE2_RESET;
2790 				reset_context.method = AMD_RESET_METHOD_MODE2;
2791 			}
2792 
2793 			/* Fatal error occurs in poison mode, mode1 reset is used to
2794 			 * recover gpu.
2795 			 */
2796 			if (ras->gpu_reset_flags & AMDGPU_RAS_GPU_RESET_MODE1_RESET) {
2797 				ras->gpu_reset_flags &= ~AMDGPU_RAS_GPU_RESET_MODE1_RESET;
2798 				set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
2799 
2800 				psp_fatal_error_recovery_quirk(&adev->psp);
2801 			}
2802 		}
2803 
2804 		amdgpu_device_gpu_recover(ras->adev, NULL, &reset_context);
2805 	}
2806 	atomic_set(&ras->in_recovery, 0);
2807 	if (hive) {
2808 		atomic_set(&hive->ras_recovery, 0);
2809 		amdgpu_put_xgmi_hive(hive);
2810 	}
2811 }
2812 
2813 /* alloc/realloc bps array */
2814 static int amdgpu_ras_realloc_eh_data_space(struct amdgpu_device *adev,
2815 		struct ras_err_handler_data *data, int pages)
2816 {
2817 	unsigned int old_space = data->count + data->space_left;
2818 	unsigned int new_space = old_space + pages;
2819 	unsigned int align_space = ALIGN(new_space, 512);
2820 	void *bps = kmalloc_array(align_space, sizeof(*data->bps), GFP_KERNEL);
2821 
2822 	if (!bps) {
2823 		return -ENOMEM;
2824 	}
2825 
2826 	if (data->bps) {
2827 		memcpy(bps, data->bps,
2828 				data->count * sizeof(*data->bps));
2829 		kfree(data->bps);
2830 	}
2831 
2832 	data->bps = bps;
2833 	data->space_left += align_space - old_space;
2834 	return 0;
2835 }
2836 
2837 static int amdgpu_ras_mca2pa_by_idx(struct amdgpu_device *adev,
2838 			struct eeprom_table_record *bps,
2839 			struct ras_err_data *err_data)
2840 {
2841 	struct ta_ras_query_address_input addr_in;
2842 	uint32_t socket = 0;
2843 	int ret = 0;
2844 
2845 	if (adev->smuio.funcs && adev->smuio.funcs->get_socket_id)
2846 		socket = adev->smuio.funcs->get_socket_id(adev);
2847 
2848 	/* reinit err_data */
2849 	err_data->err_addr_cnt = 0;
2850 	err_data->err_addr_len = adev->umc.retire_unit;
2851 
2852 	memset(&addr_in, 0, sizeof(addr_in));
2853 	addr_in.ma.err_addr = bps->address;
2854 	addr_in.ma.socket_id = socket;
2855 	addr_in.ma.ch_inst = bps->mem_channel;
2856 	/* tell RAS TA the node instance is not used */
2857 	addr_in.ma.node_inst = TA_RAS_INV_NODE;
2858 
2859 	if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr)
2860 		ret = adev->umc.ras->convert_ras_err_addr(adev, err_data,
2861 				&addr_in, NULL, false);
2862 
2863 	return ret;
2864 }
2865 
2866 static int amdgpu_ras_mca2pa(struct amdgpu_device *adev,
2867 			struct eeprom_table_record *bps,
2868 			struct ras_err_data *err_data)
2869 {
2870 	struct ta_ras_query_address_input addr_in;
2871 	uint32_t die_id, socket = 0;
2872 
2873 	if (adev->smuio.funcs && adev->smuio.funcs->get_socket_id)
2874 		socket = adev->smuio.funcs->get_socket_id(adev);
2875 
2876 	/* although die id is gotten from PA in nps1 mode, the id is
2877 	 * fitable for any nps mode
2878 	 */
2879 	if (adev->umc.ras && adev->umc.ras->get_die_id_from_pa)
2880 		die_id = adev->umc.ras->get_die_id_from_pa(adev, bps->address,
2881 					bps->retired_page << AMDGPU_GPU_PAGE_SHIFT);
2882 	else
2883 		return -EINVAL;
2884 
2885 	/* reinit err_data */
2886 	err_data->err_addr_cnt = 0;
2887 	err_data->err_addr_len = adev->umc.retire_unit;
2888 
2889 	memset(&addr_in, 0, sizeof(addr_in));
2890 	addr_in.ma.err_addr = bps->address;
2891 	addr_in.ma.ch_inst = bps->mem_channel;
2892 	addr_in.ma.umc_inst = bps->mcumc_id;
2893 	addr_in.ma.node_inst = die_id;
2894 	addr_in.ma.socket_id = socket;
2895 
2896 	if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr)
2897 		return adev->umc.ras->convert_ras_err_addr(adev, err_data,
2898 					&addr_in, NULL, false);
2899 	else
2900 		return  -EINVAL;
2901 }
2902 
2903 static int __amdgpu_ras_restore_bad_pages(struct amdgpu_device *adev,
2904 					struct eeprom_table_record *bps, int count)
2905 {
2906 	int j;
2907 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2908 	struct ras_err_handler_data *data = con->eh_data;
2909 
2910 	for (j = 0; j < count; j++) {
2911 		if (amdgpu_ras_check_bad_page_unlock(con,
2912 			bps[j].retired_page << AMDGPU_GPU_PAGE_SHIFT)) {
2913 			data->count++;
2914 			data->space_left--;
2915 			continue;
2916 		}
2917 
2918 		if (!data->space_left &&
2919 		    amdgpu_ras_realloc_eh_data_space(adev, data, 256)) {
2920 			return -ENOMEM;
2921 		}
2922 
2923 		amdgpu_ras_reserve_page(adev, bps[j].retired_page);
2924 
2925 		memcpy(&data->bps[data->count], &(bps[j]),
2926 				sizeof(struct eeprom_table_record));
2927 		data->count++;
2928 		data->space_left--;
2929 		con->bad_page_num++;
2930 	}
2931 
2932 	return 0;
2933 }
2934 
2935 static int __amdgpu_ras_convert_rec_array_from_rom(struct amdgpu_device *adev,
2936 				struct eeprom_table_record *bps, struct ras_err_data *err_data,
2937 				enum amdgpu_memory_partition nps)
2938 {
2939 	int i = 0;
2940 	enum amdgpu_memory_partition save_nps;
2941 
2942 	save_nps = (bps[0].retired_page >> UMC_NPS_SHIFT) & UMC_NPS_MASK;
2943 
2944 	/*old asics just have pa in eeprom*/
2945 	if (IP_VERSION_MAJ(amdgpu_ip_version(adev, UMC_HWIP, 0)) < 12) {
2946 		memcpy(err_data->err_addr, bps,
2947 			sizeof(struct eeprom_table_record) * adev->umc.retire_unit);
2948 		goto out;
2949 	}
2950 
2951 	for (i = 0; i < adev->umc.retire_unit; i++)
2952 		bps[i].retired_page &= ~(UMC_NPS_MASK << UMC_NPS_SHIFT);
2953 
2954 	if (save_nps) {
2955 		if (save_nps == nps) {
2956 			if (amdgpu_umc_pages_in_a_row(adev, err_data,
2957 					bps[0].retired_page << AMDGPU_GPU_PAGE_SHIFT))
2958 				return -EINVAL;
2959 			for (i = 0; i < adev->umc.retire_unit; i++) {
2960 				err_data->err_addr[i].address = bps[0].address;
2961 				err_data->err_addr[i].mem_channel = bps[0].mem_channel;
2962 				err_data->err_addr[i].bank = bps[0].bank;
2963 				err_data->err_addr[i].err_type = bps[0].err_type;
2964 				err_data->err_addr[i].mcumc_id = bps[0].mcumc_id;
2965 			}
2966 		} else {
2967 			if (amdgpu_ras_mca2pa_by_idx(adev, &bps[0], err_data))
2968 				return -EINVAL;
2969 		}
2970 	} else {
2971 		if (bps[0].address == 0) {
2972 			/* for specific old eeprom data, mca address is not stored,
2973 			 * calc it from pa
2974 			 */
2975 			if (amdgpu_umc_pa2mca(adev, bps[0].retired_page << AMDGPU_GPU_PAGE_SHIFT,
2976 				&(bps[0].address), AMDGPU_NPS1_PARTITION_MODE))
2977 				return -EINVAL;
2978 		}
2979 
2980 		if (amdgpu_ras_mca2pa(adev, &bps[0], err_data)) {
2981 			if (nps == AMDGPU_NPS1_PARTITION_MODE)
2982 				memcpy(err_data->err_addr, bps,
2983 					sizeof(struct eeprom_table_record) * adev->umc.retire_unit);
2984 			else
2985 				return -EOPNOTSUPP;
2986 		}
2987 	}
2988 
2989 out:
2990 	return __amdgpu_ras_restore_bad_pages(adev, err_data->err_addr, adev->umc.retire_unit);
2991 }
2992 
2993 static int __amdgpu_ras_convert_rec_from_rom(struct amdgpu_device *adev,
2994 				struct eeprom_table_record *bps, struct ras_err_data *err_data,
2995 				enum amdgpu_memory_partition nps)
2996 {
2997 	int i = 0;
2998 	enum amdgpu_memory_partition save_nps;
2999 
3000 	save_nps = (bps->retired_page >> UMC_NPS_SHIFT) & UMC_NPS_MASK;
3001 	bps->retired_page &= ~(UMC_NPS_MASK << UMC_NPS_SHIFT);
3002 
3003 	if (save_nps == nps) {
3004 		if (amdgpu_umc_pages_in_a_row(adev, err_data,
3005 				bps->retired_page << AMDGPU_GPU_PAGE_SHIFT))
3006 			return -EINVAL;
3007 		for (i = 0; i < adev->umc.retire_unit; i++) {
3008 			err_data->err_addr[i].address = bps->address;
3009 			err_data->err_addr[i].mem_channel = bps->mem_channel;
3010 			err_data->err_addr[i].bank = bps->bank;
3011 			err_data->err_addr[i].err_type = bps->err_type;
3012 			err_data->err_addr[i].mcumc_id = bps->mcumc_id;
3013 		}
3014 	} else {
3015 		if (bps->address) {
3016 			if (amdgpu_ras_mca2pa_by_idx(adev, bps, err_data))
3017 				return -EINVAL;
3018 		} else {
3019 			/* for specific old eeprom data, mca address is not stored,
3020 			 * calc it from pa
3021 			 */
3022 			if (amdgpu_umc_pa2mca(adev, bps->retired_page << AMDGPU_GPU_PAGE_SHIFT,
3023 				&(bps->address), AMDGPU_NPS1_PARTITION_MODE))
3024 				return -EINVAL;
3025 
3026 			if (amdgpu_ras_mca2pa(adev, bps, err_data))
3027 				return -EOPNOTSUPP;
3028 		}
3029 	}
3030 
3031 	return __amdgpu_ras_restore_bad_pages(adev, err_data->err_addr,
3032 									adev->umc.retire_unit);
3033 }
3034 
3035 /* it deal with vram only. */
3036 int amdgpu_ras_add_bad_pages(struct amdgpu_device *adev,
3037 		struct eeprom_table_record *bps, int pages, bool from_rom)
3038 {
3039 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3040 	struct ras_err_data err_data;
3041 	struct amdgpu_ras_eeprom_control *control =
3042 			&adev->psp.ras_context.ras->eeprom_control;
3043 	enum amdgpu_memory_partition nps = AMDGPU_NPS1_PARTITION_MODE;
3044 	int ret = 0;
3045 	uint32_t i = 0;
3046 
3047 	if (!con || !con->eh_data || !bps || pages <= 0)
3048 		return 0;
3049 
3050 	if (from_rom) {
3051 		err_data.err_addr =
3052 			kcalloc(adev->umc.retire_unit,
3053 				sizeof(struct eeprom_table_record), GFP_KERNEL);
3054 		if (!err_data.err_addr) {
3055 			dev_warn(adev->dev, "Failed to alloc UMC error address record in mca2pa conversion!\n");
3056 			return -ENOMEM;
3057 		}
3058 
3059 		if (adev->gmc.gmc_funcs->query_mem_partition_mode)
3060 			nps = adev->gmc.gmc_funcs->query_mem_partition_mode(adev);
3061 	}
3062 
3063 	mutex_lock(&con->recovery_lock);
3064 
3065 	if (from_rom) {
3066 		/* there is no pa recs in V3, so skip pa recs processing */
3067 		if (control->tbl_hdr.version < RAS_TABLE_VER_V3) {
3068 			for (i = 0; i < pages; i++) {
3069 				if (control->ras_num_recs - i >= adev->umc.retire_unit) {
3070 					if ((bps[i].address == bps[i + 1].address) &&
3071 						(bps[i].mem_channel == bps[i + 1].mem_channel)) {
3072 						/* deal with retire_unit records a time */
3073 						ret = __amdgpu_ras_convert_rec_array_from_rom(adev,
3074 										&bps[i], &err_data, nps);
3075 						if (ret)
3076 							con->bad_page_num -= adev->umc.retire_unit;
3077 						i += (adev->umc.retire_unit - 1);
3078 					} else {
3079 						break;
3080 					}
3081 				} else {
3082 					break;
3083 				}
3084 			}
3085 		}
3086 		for (; i < pages; i++) {
3087 			ret = __amdgpu_ras_convert_rec_from_rom(adev,
3088 				&bps[i], &err_data, nps);
3089 			if (ret)
3090 				con->bad_page_num -= adev->umc.retire_unit;
3091 		}
3092 
3093 		con->eh_data->count_saved = con->eh_data->count;
3094 	} else {
3095 		ret = __amdgpu_ras_restore_bad_pages(adev, bps, pages);
3096 	}
3097 
3098 	if (from_rom)
3099 		kfree(err_data.err_addr);
3100 	mutex_unlock(&con->recovery_lock);
3101 
3102 	return ret;
3103 }
3104 
3105 /*
3106  * write error record array to eeprom, the function should be
3107  * protected by recovery_lock
3108  * new_cnt: new added UE count, excluding reserved bad pages, can be NULL
3109  */
3110 int amdgpu_ras_save_bad_pages(struct amdgpu_device *adev,
3111 		unsigned long *new_cnt)
3112 {
3113 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3114 	struct ras_err_handler_data *data;
3115 	struct amdgpu_ras_eeprom_control *control;
3116 	int save_count, unit_num, i;
3117 
3118 	if (!con || !con->eh_data) {
3119 		if (new_cnt)
3120 			*new_cnt = 0;
3121 
3122 		return 0;
3123 	}
3124 
3125 	if (!con->eeprom_control.is_eeprom_valid) {
3126 		dev_warn(adev->dev,
3127 			"Failed to save EEPROM table data because of EEPROM data corruption!");
3128 		if (new_cnt)
3129 			*new_cnt = 0;
3130 
3131 		return 0;
3132 	}
3133 
3134 	mutex_lock(&con->recovery_lock);
3135 	control = &con->eeprom_control;
3136 	data = con->eh_data;
3137 	unit_num = data->count / adev->umc.retire_unit - control->ras_num_recs;
3138 	save_count = con->bad_page_num - control->ras_num_bad_pages;
3139 	mutex_unlock(&con->recovery_lock);
3140 
3141 	if (new_cnt)
3142 		*new_cnt = unit_num;
3143 
3144 	/* only new entries are saved */
3145 	if (unit_num > 0) {
3146 		/*old asics only save pa to eeprom like before*/
3147 		if (IP_VERSION_MAJ(amdgpu_ip_version(adev, UMC_HWIP, 0)) < 12) {
3148 			if (amdgpu_ras_eeprom_append(control,
3149 					&data->bps[data->count_saved], unit_num)) {
3150 				dev_err(adev->dev, "Failed to save EEPROM table data!");
3151 				return -EIO;
3152 			}
3153 		} else {
3154 			for (i = 0; i < unit_num; i++) {
3155 				if (amdgpu_ras_eeprom_append(control,
3156 						&data->bps[data->count_saved +
3157 						i * adev->umc.retire_unit], 1)) {
3158 					dev_err(adev->dev, "Failed to save EEPROM table data!");
3159 					return -EIO;
3160 				}
3161 			}
3162 		}
3163 
3164 		dev_info(adev->dev, "Saved %d pages to EEPROM table.\n", save_count);
3165 		data->count_saved = data->count;
3166 	}
3167 
3168 	return 0;
3169 }
3170 
3171 /*
3172  * read error record array in eeprom and reserve enough space for
3173  * storing new bad pages
3174  */
3175 static int amdgpu_ras_load_bad_pages(struct amdgpu_device *adev)
3176 {
3177 	struct amdgpu_ras_eeprom_control *control =
3178 		&adev->psp.ras_context.ras->eeprom_control;
3179 	struct eeprom_table_record *bps;
3180 	int ret, i = 0;
3181 
3182 	/* no bad page record, skip eeprom access */
3183 	if (control->ras_num_recs == 0 || amdgpu_bad_page_threshold == 0)
3184 		return 0;
3185 
3186 	bps = kcalloc(control->ras_num_recs, sizeof(*bps), GFP_KERNEL);
3187 	if (!bps)
3188 		return -ENOMEM;
3189 
3190 	ret = amdgpu_ras_eeprom_read(control, bps, control->ras_num_recs);
3191 	if (ret) {
3192 		dev_err(adev->dev, "Failed to load EEPROM table records!");
3193 	} else {
3194 		if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) {
3195 			/*In V3, there is no pa recs, and some cases(when address==0) may be parsed
3196 			as pa recs, so add verion check to avoid it.
3197 			*/
3198 			if (control->tbl_hdr.version < RAS_TABLE_VER_V3) {
3199 				for (i = 0; i < control->ras_num_recs; i++) {
3200 					if ((control->ras_num_recs - i) >= adev->umc.retire_unit) {
3201 						if ((bps[i].address == bps[i + 1].address) &&
3202 							(bps[i].mem_channel == bps[i + 1].mem_channel)) {
3203 							control->ras_num_pa_recs += adev->umc.retire_unit;
3204 							i += (adev->umc.retire_unit - 1);
3205 						} else {
3206 							control->ras_num_mca_recs +=
3207 										(control->ras_num_recs - i);
3208 							break;
3209 						}
3210 					} else {
3211 						control->ras_num_mca_recs += (control->ras_num_recs - i);
3212 						break;
3213 					}
3214 				}
3215 			} else {
3216 				control->ras_num_mca_recs = control->ras_num_recs;
3217 			}
3218 		}
3219 
3220 		ret = amdgpu_ras_add_bad_pages(adev, bps, control->ras_num_recs, true);
3221 		if (ret)
3222 			goto out;
3223 
3224 		ret = amdgpu_ras_eeprom_check(control);
3225 		if (ret)
3226 			goto out;
3227 
3228 		/* HW not usable */
3229 		if (amdgpu_ras_is_rma(adev))
3230 			ret = -EHWPOISON;
3231 	}
3232 
3233 out:
3234 	kfree(bps);
3235 	return ret;
3236 }
3237 
3238 static int amdgpu_ras_check_bad_page_unlock(struct amdgpu_ras *con,
3239 				uint64_t addr)
3240 {
3241 	struct ras_err_handler_data *data = con->eh_data;
3242 	struct amdgpu_device *adev = con->adev;
3243 	int i;
3244 
3245 	if ((addr >= adev->gmc.mc_vram_size &&
3246 	    adev->gmc.mc_vram_size) ||
3247 	    (addr >= RAS_UMC_INJECT_ADDR_LIMIT))
3248 		return -EINVAL;
3249 
3250 	addr >>= AMDGPU_GPU_PAGE_SHIFT;
3251 	for (i = 0; i < data->count; i++)
3252 		if (addr == data->bps[i].retired_page)
3253 			return 1;
3254 
3255 	return 0;
3256 }
3257 
3258 /*
3259  * check if an address belongs to bad page
3260  *
3261  * Note: this check is only for umc block
3262  */
3263 static int amdgpu_ras_check_bad_page(struct amdgpu_device *adev,
3264 				uint64_t addr)
3265 {
3266 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3267 	int ret = 0;
3268 
3269 	if (!con || !con->eh_data)
3270 		return ret;
3271 
3272 	mutex_lock(&con->recovery_lock);
3273 	ret = amdgpu_ras_check_bad_page_unlock(con, addr);
3274 	mutex_unlock(&con->recovery_lock);
3275 	return ret;
3276 }
3277 
3278 static void amdgpu_ras_validate_threshold(struct amdgpu_device *adev,
3279 					  uint32_t max_count)
3280 {
3281 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3282 
3283 	/*
3284 	 * amdgpu_bad_page_threshold is used to config
3285 	 * the threshold for the number of bad pages.
3286 	 * -1:  Threshold is set to default value
3287 	 *      Driver will issue a warning message when threshold is reached
3288 	 *      and continue runtime services.
3289 	 * 0:   Disable bad page retirement
3290 	 *      Driver will not retire bad pages
3291 	 *      which is intended for debugging purpose.
3292 	 * -2:  Threshold is determined by a formula
3293 	 *      that assumes 1 bad page per 100M of local memory.
3294 	 *      Driver will continue runtime services when threhold is reached.
3295 	 * 0 < threshold < max number of bad page records in EEPROM,
3296 	 *      A user-defined threshold is set
3297 	 *      Driver will halt runtime services when this custom threshold is reached.
3298 	 */
3299 	if (amdgpu_bad_page_threshold == -2) {
3300 		u64 val = adev->gmc.mc_vram_size;
3301 
3302 		do_div(val, RAS_BAD_PAGE_COVER);
3303 		con->bad_page_cnt_threshold = min(lower_32_bits(val),
3304 						  max_count);
3305 	} else if (amdgpu_bad_page_threshold == -1) {
3306 		con->bad_page_cnt_threshold = ((con->reserved_pages_in_bytes) >> 21) << 4;
3307 	} else {
3308 		con->bad_page_cnt_threshold = min_t(int, max_count,
3309 						    amdgpu_bad_page_threshold);
3310 	}
3311 }
3312 
3313 int amdgpu_ras_put_poison_req(struct amdgpu_device *adev,
3314 		enum amdgpu_ras_block block, uint16_t pasid,
3315 		pasid_notify pasid_fn, void *data, uint32_t reset)
3316 {
3317 	int ret = 0;
3318 	struct ras_poison_msg poison_msg;
3319 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3320 
3321 	memset(&poison_msg, 0, sizeof(poison_msg));
3322 	poison_msg.block = block;
3323 	poison_msg.pasid = pasid;
3324 	poison_msg.reset = reset;
3325 	poison_msg.pasid_fn = pasid_fn;
3326 	poison_msg.data = data;
3327 
3328 	ret = kfifo_put(&con->poison_fifo, poison_msg);
3329 	if (!ret) {
3330 		dev_err(adev->dev, "Poison message fifo is full!\n");
3331 		return -ENOSPC;
3332 	}
3333 
3334 	return 0;
3335 }
3336 
3337 static int amdgpu_ras_get_poison_req(struct amdgpu_device *adev,
3338 		struct ras_poison_msg *poison_msg)
3339 {
3340 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3341 
3342 	return kfifo_get(&con->poison_fifo, poison_msg);
3343 }
3344 
3345 static void amdgpu_ras_ecc_log_init(struct ras_ecc_log_info *ecc_log)
3346 {
3347 	mutex_init(&ecc_log->lock);
3348 
3349 	INIT_RADIX_TREE(&ecc_log->de_page_tree, GFP_KERNEL);
3350 	ecc_log->de_queried_count = 0;
3351 	ecc_log->consumption_q_count = 0;
3352 }
3353 
3354 static void amdgpu_ras_ecc_log_fini(struct ras_ecc_log_info *ecc_log)
3355 {
3356 	struct radix_tree_iter iter;
3357 	void __rcu **slot;
3358 	struct ras_ecc_err *ecc_err;
3359 
3360 	mutex_lock(&ecc_log->lock);
3361 	radix_tree_for_each_slot(slot, &ecc_log->de_page_tree, &iter, 0) {
3362 		ecc_err = radix_tree_deref_slot(slot);
3363 		kfree(ecc_err->err_pages.pfn);
3364 		kfree(ecc_err);
3365 		radix_tree_iter_delete(&ecc_log->de_page_tree, &iter, slot);
3366 	}
3367 	mutex_unlock(&ecc_log->lock);
3368 
3369 	mutex_destroy(&ecc_log->lock);
3370 	ecc_log->de_queried_count = 0;
3371 	ecc_log->consumption_q_count = 0;
3372 }
3373 
3374 static bool amdgpu_ras_schedule_retirement_dwork(struct amdgpu_ras *con,
3375 				uint32_t delayed_ms)
3376 {
3377 	int ret;
3378 
3379 	mutex_lock(&con->umc_ecc_log.lock);
3380 	ret = radix_tree_tagged(&con->umc_ecc_log.de_page_tree,
3381 			UMC_ECC_NEW_DETECTED_TAG);
3382 	mutex_unlock(&con->umc_ecc_log.lock);
3383 
3384 	if (ret)
3385 		schedule_delayed_work(&con->page_retirement_dwork,
3386 			msecs_to_jiffies(delayed_ms));
3387 
3388 	return ret ? true : false;
3389 }
3390 
3391 static void amdgpu_ras_do_page_retirement(struct work_struct *work)
3392 {
3393 	struct amdgpu_ras *con = container_of(work, struct amdgpu_ras,
3394 					      page_retirement_dwork.work);
3395 	struct amdgpu_device *adev = con->adev;
3396 	struct ras_err_data err_data;
3397 
3398 	/* If gpu reset is ongoing, delay retiring the bad pages */
3399 	if (amdgpu_in_reset(adev) || amdgpu_ras_in_recovery(adev)) {
3400 		amdgpu_ras_schedule_retirement_dwork(con,
3401 				AMDGPU_RAS_RETIRE_PAGE_INTERVAL * 3);
3402 		return;
3403 	}
3404 
3405 	amdgpu_ras_error_data_init(&err_data);
3406 
3407 	amdgpu_umc_handle_bad_pages(adev, &err_data);
3408 
3409 	amdgpu_ras_error_data_fini(&err_data);
3410 
3411 	amdgpu_ras_schedule_retirement_dwork(con,
3412 			AMDGPU_RAS_RETIRE_PAGE_INTERVAL);
3413 }
3414 
3415 static int amdgpu_ras_poison_creation_handler(struct amdgpu_device *adev,
3416 				uint32_t poison_creation_count)
3417 {
3418 	int ret = 0;
3419 	struct ras_ecc_log_info *ecc_log;
3420 	struct ras_query_if info;
3421 	u32 timeout = MAX_UMC_POISON_POLLING_TIME_ASYNC;
3422 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
3423 	u64 de_queried_count;
3424 	u64 consumption_q_count;
3425 	enum ras_event_type type = RAS_EVENT_TYPE_POISON_CREATION;
3426 
3427 	memset(&info, 0, sizeof(info));
3428 	info.head.block = AMDGPU_RAS_BLOCK__UMC;
3429 
3430 	ecc_log = &ras->umc_ecc_log;
3431 	ecc_log->de_queried_count = 0;
3432 	ecc_log->consumption_q_count = 0;
3433 
3434 	do {
3435 		ret = amdgpu_ras_query_error_status_with_event(adev, &info, type);
3436 		if (ret)
3437 			return ret;
3438 
3439 		de_queried_count = ecc_log->de_queried_count;
3440 		consumption_q_count = ecc_log->consumption_q_count;
3441 
3442 		if (de_queried_count && consumption_q_count)
3443 			break;
3444 
3445 		msleep(100);
3446 	} while (--timeout);
3447 
3448 	if (de_queried_count)
3449 		schedule_delayed_work(&ras->page_retirement_dwork, 0);
3450 
3451 	if (amdgpu_ras_is_rma(adev) && atomic_cmpxchg(&ras->rma_in_recovery, 0, 1) == 0)
3452 		amdgpu_ras_reset_gpu(adev);
3453 
3454 	return 0;
3455 }
3456 
3457 static void amdgpu_ras_clear_poison_fifo(struct amdgpu_device *adev)
3458 {
3459 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3460 	struct ras_poison_msg msg;
3461 	int ret;
3462 
3463 	do {
3464 		ret = kfifo_get(&con->poison_fifo, &msg);
3465 	} while (ret);
3466 }
3467 
3468 static int amdgpu_ras_poison_consumption_handler(struct amdgpu_device *adev,
3469 			uint32_t msg_count, uint32_t *gpu_reset)
3470 {
3471 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3472 	uint32_t reset_flags = 0, reset = 0;
3473 	struct ras_poison_msg msg;
3474 	int ret, i;
3475 
3476 	kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
3477 
3478 	for (i = 0; i < msg_count; i++) {
3479 		ret = amdgpu_ras_get_poison_req(adev, &msg);
3480 		if (!ret)
3481 			continue;
3482 
3483 		if (msg.pasid_fn)
3484 			msg.pasid_fn(adev, msg.pasid, msg.data);
3485 
3486 		reset_flags |= msg.reset;
3487 	}
3488 
3489 	/*
3490 	 * Try to ensure poison creation handler is completed first
3491 	 * to set rma if bad page exceed threshold.
3492 	 */
3493 	flush_delayed_work(&con->page_retirement_dwork);
3494 
3495 	/* for RMA, amdgpu_ras_poison_creation_handler will trigger gpu reset */
3496 	if (reset_flags && !amdgpu_ras_is_rma(adev)) {
3497 		if (reset_flags & AMDGPU_RAS_GPU_RESET_MODE1_RESET)
3498 			reset = AMDGPU_RAS_GPU_RESET_MODE1_RESET;
3499 		else if (reset_flags & AMDGPU_RAS_GPU_RESET_MODE2_RESET)
3500 			reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET;
3501 		else
3502 			reset = reset_flags;
3503 
3504 		con->gpu_reset_flags |= reset;
3505 		amdgpu_ras_reset_gpu(adev);
3506 
3507 		*gpu_reset = reset;
3508 
3509 		/* Wait for gpu recovery to complete */
3510 		flush_work(&con->recovery_work);
3511 	}
3512 
3513 	return 0;
3514 }
3515 
3516 static int amdgpu_ras_page_retirement_thread(void *param)
3517 {
3518 	struct amdgpu_device *adev = (struct amdgpu_device *)param;
3519 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3520 	uint32_t poison_creation_count, msg_count;
3521 	uint32_t gpu_reset;
3522 	int ret;
3523 
3524 	while (!kthread_should_stop()) {
3525 
3526 		wait_event_interruptible(con->page_retirement_wq,
3527 				kthread_should_stop() ||
3528 				atomic_read(&con->page_retirement_req_cnt));
3529 
3530 		if (kthread_should_stop())
3531 			break;
3532 
3533 		mutex_lock(&con->poison_lock);
3534 		gpu_reset = 0;
3535 
3536 		do {
3537 			poison_creation_count = atomic_read(&con->poison_creation_count);
3538 			ret = amdgpu_ras_poison_creation_handler(adev, poison_creation_count);
3539 			if (ret == -EIO)
3540 				break;
3541 
3542 			if (poison_creation_count) {
3543 				atomic_sub(poison_creation_count, &con->poison_creation_count);
3544 				atomic_sub(poison_creation_count, &con->page_retirement_req_cnt);
3545 			}
3546 		} while (atomic_read(&con->poison_creation_count) &&
3547 			!atomic_read(&con->poison_consumption_count));
3548 
3549 		if (ret != -EIO) {
3550 			msg_count = kfifo_len(&con->poison_fifo);
3551 			if (msg_count) {
3552 				ret = amdgpu_ras_poison_consumption_handler(adev,
3553 						msg_count, &gpu_reset);
3554 				if ((ret != -EIO) &&
3555 				    (gpu_reset != AMDGPU_RAS_GPU_RESET_MODE1_RESET))
3556 					atomic_sub(msg_count, &con->page_retirement_req_cnt);
3557 			}
3558 		}
3559 
3560 		if ((ret == -EIO) || (gpu_reset == AMDGPU_RAS_GPU_RESET_MODE1_RESET)) {
3561 			/* gpu mode-1 reset is ongoing or just completed ras mode-1 reset */
3562 			/* Clear poison creation request */
3563 			atomic_set(&con->poison_creation_count, 0);
3564 			atomic_set(&con->poison_consumption_count, 0);
3565 
3566 			/* Clear poison fifo */
3567 			amdgpu_ras_clear_poison_fifo(adev);
3568 
3569 			/* Clear all poison requests */
3570 			atomic_set(&con->page_retirement_req_cnt, 0);
3571 
3572 			if (ret == -EIO) {
3573 				/* Wait for mode-1 reset to complete */
3574 				down_read(&adev->reset_domain->sem);
3575 				up_read(&adev->reset_domain->sem);
3576 			}
3577 
3578 			/* Wake up work to save bad pages to eeprom */
3579 			schedule_delayed_work(&con->page_retirement_dwork, 0);
3580 		} else if (gpu_reset) {
3581 			/* gpu just completed mode-2 reset or other reset */
3582 			/* Clear poison consumption messages cached in fifo */
3583 			msg_count = kfifo_len(&con->poison_fifo);
3584 			if (msg_count) {
3585 				amdgpu_ras_clear_poison_fifo(adev);
3586 				atomic_sub(msg_count, &con->page_retirement_req_cnt);
3587 			}
3588 
3589 			atomic_set(&con->poison_consumption_count, 0);
3590 
3591 			/* Wake up work to save bad pages to eeprom */
3592 			schedule_delayed_work(&con->page_retirement_dwork, 0);
3593 		}
3594 		mutex_unlock(&con->poison_lock);
3595 	}
3596 
3597 	return 0;
3598 }
3599 
3600 int amdgpu_ras_init_badpage_info(struct amdgpu_device *adev)
3601 {
3602 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3603 	struct amdgpu_ras_eeprom_control *control;
3604 	int ret;
3605 
3606 	if (!con || amdgpu_sriov_vf(adev))
3607 		return 0;
3608 
3609 	if (amdgpu_uniras_enabled(adev))
3610 		return 0;
3611 
3612 	control = &con->eeprom_control;
3613 	ret = amdgpu_ras_eeprom_init(control);
3614 	control->is_eeprom_valid = !ret;
3615 
3616 	if (!adev->umc.ras || !adev->umc.ras->convert_ras_err_addr)
3617 		control->ras_num_pa_recs = control->ras_num_recs;
3618 
3619 	if (adev->umc.ras &&
3620 	    adev->umc.ras->get_retire_flip_bits)
3621 		adev->umc.ras->get_retire_flip_bits(adev);
3622 
3623 	if (control->ras_num_recs && control->is_eeprom_valid) {
3624 		ret = amdgpu_ras_load_bad_pages(adev);
3625 		if (ret) {
3626 			control->is_eeprom_valid = false;
3627 			return 0;
3628 		}
3629 
3630 		amdgpu_dpm_send_hbm_bad_pages_num(
3631 			adev, control->ras_num_bad_pages);
3632 
3633 		if (con->update_channel_flag == true) {
3634 			amdgpu_dpm_send_hbm_bad_channel_flag(
3635 				adev, control->bad_channel_bitmap);
3636 			con->update_channel_flag = false;
3637 		}
3638 
3639 		/* The format action is only applied to new ASICs */
3640 		if (IP_VERSION_MAJ(amdgpu_ip_version(adev, UMC_HWIP, 0)) >= 12 &&
3641 		    control->tbl_hdr.version < RAS_TABLE_VER_V3)
3642 			if (!amdgpu_ras_eeprom_reset_table(control))
3643 				if (amdgpu_ras_save_bad_pages(adev, NULL))
3644 					dev_warn(adev->dev, "Failed to format RAS EEPROM data in V3 version!\n");
3645 	}
3646 
3647 	return 0;
3648 }
3649 
3650 int amdgpu_ras_recovery_init(struct amdgpu_device *adev, bool init_bp_info)
3651 {
3652 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3653 	struct ras_err_handler_data **data;
3654 	u32  max_eeprom_records_count = 0;
3655 	int ret;
3656 
3657 	if (!con || amdgpu_sriov_vf(adev))
3658 		return 0;
3659 
3660 	/* Allow access to RAS EEPROM via debugfs, when the ASIC
3661 	 * supports RAS and debugfs is enabled, but when
3662 	 * adev->ras_enabled is unset, i.e. when "ras_enable"
3663 	 * module parameter is set to 0.
3664 	 */
3665 	con->adev = adev;
3666 
3667 	if (!adev->ras_enabled)
3668 		return 0;
3669 
3670 	data = &con->eh_data;
3671 	*data = kzalloc(sizeof(**data), GFP_KERNEL);
3672 	if (!*data) {
3673 		ret = -ENOMEM;
3674 		goto out;
3675 	}
3676 
3677 	mutex_init(&con->recovery_lock);
3678 	mutex_init(&con->poison_lock);
3679 	INIT_WORK(&con->recovery_work, amdgpu_ras_do_recovery);
3680 	atomic_set(&con->in_recovery, 0);
3681 	atomic_set(&con->rma_in_recovery, 0);
3682 	con->eeprom_control.bad_channel_bitmap = 0;
3683 
3684 	max_eeprom_records_count = amdgpu_ras_eeprom_max_record_count(&con->eeprom_control);
3685 	amdgpu_ras_validate_threshold(adev, max_eeprom_records_count);
3686 
3687 	if (init_bp_info) {
3688 		ret = amdgpu_ras_init_badpage_info(adev);
3689 		if (ret)
3690 			goto free;
3691 	}
3692 
3693 	mutex_init(&con->page_rsv_lock);
3694 	INIT_KFIFO(con->poison_fifo);
3695 	mutex_init(&con->page_retirement_lock);
3696 	init_waitqueue_head(&con->page_retirement_wq);
3697 	atomic_set(&con->page_retirement_req_cnt, 0);
3698 	atomic_set(&con->poison_creation_count, 0);
3699 	atomic_set(&con->poison_consumption_count, 0);
3700 	con->page_retirement_thread =
3701 		kthread_run(amdgpu_ras_page_retirement_thread, adev, "umc_page_retirement");
3702 	if (IS_ERR(con->page_retirement_thread)) {
3703 		con->page_retirement_thread = NULL;
3704 		dev_warn(adev->dev, "Failed to create umc_page_retirement thread!!!\n");
3705 	}
3706 
3707 	INIT_DELAYED_WORK(&con->page_retirement_dwork, amdgpu_ras_do_page_retirement);
3708 	amdgpu_ras_ecc_log_init(&con->umc_ecc_log);
3709 #ifdef CONFIG_X86_MCE_AMD
3710 	if ((adev->asic_type == CHIP_ALDEBARAN) &&
3711 	    (adev->gmc.xgmi.connected_to_cpu))
3712 		amdgpu_register_bad_pages_mca_notifier(adev);
3713 #endif
3714 	return 0;
3715 
3716 free:
3717 	kfree((*data)->bps);
3718 	kfree(*data);
3719 	con->eh_data = NULL;
3720 out:
3721 	dev_warn(adev->dev, "Failed to initialize ras recovery! (%d)\n", ret);
3722 
3723 	/*
3724 	 * Except error threshold exceeding case, other failure cases in this
3725 	 * function would not fail amdgpu driver init.
3726 	 */
3727 	if (!amdgpu_ras_is_rma(adev))
3728 		ret = 0;
3729 	else
3730 		ret = -EINVAL;
3731 
3732 	return ret;
3733 }
3734 
3735 static int amdgpu_ras_recovery_fini(struct amdgpu_device *adev)
3736 {
3737 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3738 	struct ras_err_handler_data *data = con->eh_data;
3739 	int max_flush_timeout = MAX_FLUSH_RETIRE_DWORK_TIMES;
3740 	bool ret;
3741 
3742 	/* recovery_init failed to init it, fini is useless */
3743 	if (!data)
3744 		return 0;
3745 
3746 	/* Save all cached bad pages to eeprom */
3747 	do {
3748 		flush_delayed_work(&con->page_retirement_dwork);
3749 		ret = amdgpu_ras_schedule_retirement_dwork(con, 0);
3750 	} while (ret && max_flush_timeout--);
3751 
3752 	if (con->page_retirement_thread)
3753 		kthread_stop(con->page_retirement_thread);
3754 
3755 	atomic_set(&con->page_retirement_req_cnt, 0);
3756 	atomic_set(&con->poison_creation_count, 0);
3757 
3758 	mutex_destroy(&con->page_rsv_lock);
3759 
3760 	cancel_work_sync(&con->recovery_work);
3761 
3762 	cancel_delayed_work_sync(&con->page_retirement_dwork);
3763 
3764 	amdgpu_ras_ecc_log_fini(&con->umc_ecc_log);
3765 
3766 	mutex_lock(&con->recovery_lock);
3767 	con->eh_data = NULL;
3768 	kfree(data->bps);
3769 	kfree(data);
3770 	mutex_unlock(&con->recovery_lock);
3771 
3772 	amdgpu_ras_critical_region_init(adev);
3773 
3774 	return 0;
3775 }
3776 /* recovery end */
3777 
3778 static bool amdgpu_ras_asic_supported(struct amdgpu_device *adev)
3779 {
3780 	if (amdgpu_sriov_vf(adev)) {
3781 		switch (amdgpu_ip_version(adev, MP0_HWIP, 0)) {
3782 		case IP_VERSION(13, 0, 2):
3783 		case IP_VERSION(13, 0, 6):
3784 		case IP_VERSION(13, 0, 12):
3785 		case IP_VERSION(13, 0, 14):
3786 			return true;
3787 		default:
3788 			return false;
3789 		}
3790 	}
3791 
3792 	if (adev->asic_type == CHIP_IP_DISCOVERY) {
3793 		switch (amdgpu_ip_version(adev, MP0_HWIP, 0)) {
3794 		case IP_VERSION(13, 0, 0):
3795 		case IP_VERSION(13, 0, 6):
3796 		case IP_VERSION(13, 0, 10):
3797 		case IP_VERSION(13, 0, 12):
3798 		case IP_VERSION(13, 0, 14):
3799 		case IP_VERSION(14, 0, 3):
3800 			return true;
3801 		default:
3802 			return false;
3803 		}
3804 	}
3805 
3806 	return adev->asic_type == CHIP_VEGA10 ||
3807 		adev->asic_type == CHIP_VEGA20 ||
3808 		adev->asic_type == CHIP_ARCTURUS ||
3809 		adev->asic_type == CHIP_ALDEBARAN ||
3810 		adev->asic_type == CHIP_SIENNA_CICHLID;
3811 }
3812 
3813 /*
3814  * this is workaround for vega20 workstation sku,
3815  * force enable gfx ras, ignore vbios gfx ras flag
3816  * due to GC EDC can not write
3817  */
3818 static void amdgpu_ras_get_quirks(struct amdgpu_device *adev)
3819 {
3820 	struct atom_context *ctx = adev->mode_info.atom_context;
3821 
3822 	if (!ctx)
3823 		return;
3824 
3825 	if (strnstr(ctx->vbios_pn, "D16406",
3826 		    sizeof(ctx->vbios_pn)) ||
3827 		strnstr(ctx->vbios_pn, "D36002",
3828 			sizeof(ctx->vbios_pn)))
3829 		adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX);
3830 }
3831 
3832 /* Query ras capablity via atomfirmware interface */
3833 static void amdgpu_ras_query_ras_capablity_from_vbios(struct amdgpu_device *adev)
3834 {
3835 	/* mem_ecc cap */
3836 	if (amdgpu_atomfirmware_mem_ecc_supported(adev)) {
3837 		dev_info(adev->dev, "MEM ECC is active.\n");
3838 		adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__UMC |
3839 					 1 << AMDGPU_RAS_BLOCK__DF);
3840 	} else {
3841 		dev_info(adev->dev, "MEM ECC is not presented.\n");
3842 	}
3843 
3844 	/* sram_ecc cap */
3845 	if (amdgpu_atomfirmware_sram_ecc_supported(adev)) {
3846 		dev_info(adev->dev, "SRAM ECC is active.\n");
3847 		if (!amdgpu_sriov_vf(adev))
3848 			adev->ras_hw_enabled |= ~(1 << AMDGPU_RAS_BLOCK__UMC |
3849 						  1 << AMDGPU_RAS_BLOCK__DF);
3850 		else
3851 			adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__PCIE_BIF |
3852 						 1 << AMDGPU_RAS_BLOCK__SDMA |
3853 						 1 << AMDGPU_RAS_BLOCK__GFX);
3854 
3855 		/*
3856 		 * VCN/JPEG RAS can be supported on both bare metal and
3857 		 * SRIOV environment
3858 		 */
3859 		if (amdgpu_ip_version(adev, VCN_HWIP, 0) == IP_VERSION(2, 6, 0) ||
3860 		    amdgpu_ip_version(adev, VCN_HWIP, 0) == IP_VERSION(4, 0, 0) ||
3861 		    amdgpu_ip_version(adev, VCN_HWIP, 0) == IP_VERSION(4, 0, 3) ||
3862 		    amdgpu_ip_version(adev, VCN_HWIP, 0) == IP_VERSION(5, 0, 1))
3863 			adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__VCN |
3864 						 1 << AMDGPU_RAS_BLOCK__JPEG);
3865 		else
3866 			adev->ras_hw_enabled &= ~(1 << AMDGPU_RAS_BLOCK__VCN |
3867 						  1 << AMDGPU_RAS_BLOCK__JPEG);
3868 
3869 		/*
3870 		 * XGMI RAS is not supported if xgmi num physical nodes
3871 		 * is zero
3872 		 */
3873 		if (!adev->gmc.xgmi.num_physical_nodes)
3874 			adev->ras_hw_enabled &= ~(1 << AMDGPU_RAS_BLOCK__XGMI_WAFL);
3875 	} else {
3876 		dev_info(adev->dev, "SRAM ECC is not presented.\n");
3877 	}
3878 }
3879 
3880 /* Query poison mode from umc/df IP callbacks */
3881 static void amdgpu_ras_query_poison_mode(struct amdgpu_device *adev)
3882 {
3883 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
3884 	bool df_poison, umc_poison;
3885 
3886 	/* poison setting is useless on SRIOV guest */
3887 	if (amdgpu_sriov_vf(adev) || !con)
3888 		return;
3889 
3890 	/* Init poison supported flag, the default value is false */
3891 	if (adev->gmc.xgmi.connected_to_cpu ||
3892 	    adev->gmc.is_app_apu) {
3893 		/* enabled by default when GPU is connected to CPU */
3894 		con->poison_supported = true;
3895 	} else if (adev->df.funcs &&
3896 	    adev->df.funcs->query_ras_poison_mode &&
3897 	    adev->umc.ras &&
3898 	    adev->umc.ras->query_ras_poison_mode) {
3899 		df_poison =
3900 			adev->df.funcs->query_ras_poison_mode(adev);
3901 		umc_poison =
3902 			adev->umc.ras->query_ras_poison_mode(adev);
3903 
3904 		/* Only poison is set in both DF and UMC, we can support it */
3905 		if (df_poison && umc_poison)
3906 			con->poison_supported = true;
3907 		else if (df_poison != umc_poison)
3908 			dev_warn(adev->dev,
3909 				"Poison setting is inconsistent in DF/UMC(%d:%d)!\n",
3910 				df_poison, umc_poison);
3911 	}
3912 }
3913 
3914 /*
3915  * check hardware's ras ability which will be saved in hw_supported.
3916  * if hardware does not support ras, we can skip some ras initializtion and
3917  * forbid some ras operations from IP.
3918  * if software itself, say boot parameter, limit the ras ability. We still
3919  * need allow IP do some limited operations, like disable. In such case,
3920  * we have to initialize ras as normal. but need check if operation is
3921  * allowed or not in each function.
3922  */
3923 static void amdgpu_ras_check_supported(struct amdgpu_device *adev)
3924 {
3925 	adev->ras_hw_enabled = adev->ras_enabled = 0;
3926 
3927 	if (!amdgpu_ras_asic_supported(adev))
3928 		return;
3929 
3930 	if (amdgpu_sriov_vf(adev)) {
3931 		if (amdgpu_virt_get_ras_capability(adev))
3932 			goto init_ras_enabled_flag;
3933 	}
3934 
3935 	/* query ras capability from psp */
3936 	if (amdgpu_psp_get_ras_capability(&adev->psp))
3937 		goto init_ras_enabled_flag;
3938 
3939 	/* query ras capablity from bios */
3940 	if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
3941 		amdgpu_ras_query_ras_capablity_from_vbios(adev);
3942 	} else {
3943 		/* driver only manages a few IP blocks RAS feature
3944 		 * when GPU is connected cpu through XGMI */
3945 		adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX |
3946 					   1 << AMDGPU_RAS_BLOCK__SDMA |
3947 					   1 << AMDGPU_RAS_BLOCK__MMHUB);
3948 	}
3949 
3950 	/* apply asic specific settings (vega20 only for now) */
3951 	amdgpu_ras_get_quirks(adev);
3952 
3953 	/* query poison mode from umc/df ip callback */
3954 	amdgpu_ras_query_poison_mode(adev);
3955 
3956 init_ras_enabled_flag:
3957 	/* hw_supported needs to be aligned with RAS block mask. */
3958 	adev->ras_hw_enabled &= AMDGPU_RAS_BLOCK_MASK;
3959 
3960 	adev->ras_enabled = amdgpu_ras_enable == 0 ? 0 :
3961 		adev->ras_hw_enabled & amdgpu_ras_mask;
3962 
3963 	/* aca is disabled by default except for psp v13_0_6/v13_0_12/v13_0_14 */
3964 	if (!amdgpu_sriov_vf(adev)) {
3965 		adev->aca.is_enabled =
3966 			(amdgpu_ip_version(adev, MP0_HWIP, 0) == IP_VERSION(13, 0, 6) ||
3967 			amdgpu_ip_version(adev, MP0_HWIP, 0) == IP_VERSION(13, 0, 12) ||
3968 			amdgpu_ip_version(adev, MP0_HWIP, 0) == IP_VERSION(13, 0, 14));
3969 	}
3970 
3971 	/* bad page feature is not applicable to specific app platform */
3972 	if (adev->gmc.is_app_apu &&
3973 	    amdgpu_ip_version(adev, UMC_HWIP, 0) == IP_VERSION(12, 0, 0))
3974 		amdgpu_bad_page_threshold = 0;
3975 }
3976 
3977 static void amdgpu_ras_counte_dw(struct work_struct *work)
3978 {
3979 	struct amdgpu_ras *con = container_of(work, struct amdgpu_ras,
3980 					      ras_counte_delay_work.work);
3981 	struct amdgpu_device *adev = con->adev;
3982 	struct drm_device *dev = adev_to_drm(adev);
3983 	unsigned long ce_count, ue_count;
3984 	int res;
3985 
3986 	res = pm_runtime_get_sync(dev->dev);
3987 	if (res < 0)
3988 		goto Out;
3989 
3990 	/* Cache new values.
3991 	 */
3992 	if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count, NULL) == 0) {
3993 		atomic_set(&con->ras_ce_count, ce_count);
3994 		atomic_set(&con->ras_ue_count, ue_count);
3995 	}
3996 
3997 	pm_runtime_mark_last_busy(dev->dev);
3998 Out:
3999 	pm_runtime_put_autosuspend(dev->dev);
4000 }
4001 
4002 static int amdgpu_get_ras_schema(struct amdgpu_device *adev)
4003 {
4004 	return  amdgpu_ras_is_poison_mode_supported(adev) ? AMDGPU_RAS_ERROR__POISON : 0 |
4005 			AMDGPU_RAS_ERROR__SINGLE_CORRECTABLE |
4006 			AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE |
4007 			AMDGPU_RAS_ERROR__PARITY;
4008 }
4009 
4010 static void ras_event_mgr_init(struct ras_event_manager *mgr)
4011 {
4012 	struct ras_event_state *event_state;
4013 	int i;
4014 
4015 	memset(mgr, 0, sizeof(*mgr));
4016 	atomic64_set(&mgr->seqno, 0);
4017 
4018 	for (i = 0; i < ARRAY_SIZE(mgr->event_state); i++) {
4019 		event_state = &mgr->event_state[i];
4020 		event_state->last_seqno = RAS_EVENT_INVALID_ID;
4021 		atomic64_set(&event_state->count, 0);
4022 	}
4023 }
4024 
4025 static void amdgpu_ras_event_mgr_init(struct amdgpu_device *adev)
4026 {
4027 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
4028 	struct amdgpu_hive_info *hive;
4029 
4030 	if (!ras)
4031 		return;
4032 
4033 	hive = amdgpu_get_xgmi_hive(adev);
4034 	ras->event_mgr = hive ? &hive->event_mgr : &ras->__event_mgr;
4035 
4036 	/* init event manager with node 0 on xgmi system */
4037 	if (!amdgpu_reset_in_recovery(adev)) {
4038 		if (!hive || adev->gmc.xgmi.node_id == 0)
4039 			ras_event_mgr_init(ras->event_mgr);
4040 	}
4041 
4042 	if (hive)
4043 		amdgpu_put_xgmi_hive(hive);
4044 }
4045 
4046 static void amdgpu_ras_init_reserved_vram_size(struct amdgpu_device *adev)
4047 {
4048 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4049 
4050 	if (!con || (adev->flags & AMD_IS_APU))
4051 		return;
4052 
4053 	switch (amdgpu_ip_version(adev, MP0_HWIP, 0)) {
4054 	case IP_VERSION(13, 0, 2):
4055 	case IP_VERSION(13, 0, 6):
4056 	case IP_VERSION(13, 0, 12):
4057 		con->reserved_pages_in_bytes = AMDGPU_RAS_RESERVED_VRAM_SIZE_DEFAULT;
4058 		break;
4059 	case IP_VERSION(13, 0, 14):
4060 		con->reserved_pages_in_bytes = (AMDGPU_RAS_RESERVED_VRAM_SIZE_DEFAULT << 1);
4061 		break;
4062 	default:
4063 		break;
4064 	}
4065 }
4066 
4067 int amdgpu_ras_init(struct amdgpu_device *adev)
4068 {
4069 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4070 	int r;
4071 
4072 	if (con)
4073 		return 0;
4074 
4075 	con = kzalloc(sizeof(*con) +
4076 			sizeof(struct ras_manager) * AMDGPU_RAS_BLOCK_COUNT +
4077 			sizeof(struct ras_manager) * AMDGPU_RAS_MCA_BLOCK_COUNT,
4078 			GFP_KERNEL);
4079 	if (!con)
4080 		return -ENOMEM;
4081 
4082 	con->adev = adev;
4083 	INIT_DELAYED_WORK(&con->ras_counte_delay_work, amdgpu_ras_counte_dw);
4084 	atomic_set(&con->ras_ce_count, 0);
4085 	atomic_set(&con->ras_ue_count, 0);
4086 
4087 	con->objs = (struct ras_manager *)(con + 1);
4088 
4089 	amdgpu_ras_set_context(adev, con);
4090 
4091 	amdgpu_ras_check_supported(adev);
4092 
4093 	if (!adev->ras_enabled || adev->asic_type == CHIP_VEGA10) {
4094 		/* set gfx block ras context feature for VEGA20 Gaming
4095 		 * send ras disable cmd to ras ta during ras late init.
4096 		 */
4097 		if (!adev->ras_enabled && adev->asic_type == CHIP_VEGA20) {
4098 			con->features |= BIT(AMDGPU_RAS_BLOCK__GFX);
4099 
4100 			return 0;
4101 		}
4102 
4103 		r = 0;
4104 		goto release_con;
4105 	}
4106 
4107 	con->update_channel_flag = false;
4108 	con->features = 0;
4109 	con->schema = 0;
4110 	INIT_LIST_HEAD(&con->head);
4111 	/* Might need get this flag from vbios. */
4112 	con->flags = RAS_DEFAULT_FLAGS;
4113 
4114 	/* initialize nbio ras function ahead of any other
4115 	 * ras functions so hardware fatal error interrupt
4116 	 * can be enabled as early as possible */
4117 	switch (amdgpu_ip_version(adev, NBIO_HWIP, 0)) {
4118 	case IP_VERSION(7, 4, 0):
4119 	case IP_VERSION(7, 4, 1):
4120 	case IP_VERSION(7, 4, 4):
4121 		if (!adev->gmc.xgmi.connected_to_cpu)
4122 			adev->nbio.ras = &nbio_v7_4_ras;
4123 		break;
4124 	case IP_VERSION(4, 3, 0):
4125 		if (adev->ras_hw_enabled & (1 << AMDGPU_RAS_BLOCK__DF))
4126 			/* unlike other generation of nbio ras,
4127 			 * nbio v4_3 only support fatal error interrupt
4128 			 * to inform software that DF is freezed due to
4129 			 * system fatal error event. driver should not
4130 			 * enable nbio ras in such case. Instead,
4131 			 * check DF RAS */
4132 			adev->nbio.ras = &nbio_v4_3_ras;
4133 		break;
4134 	case IP_VERSION(6, 3, 1):
4135 		if (adev->ras_hw_enabled & (1 << AMDGPU_RAS_BLOCK__DF))
4136 			/* unlike other generation of nbio ras,
4137 			 * nbif v6_3_1 only support fatal error interrupt
4138 			 * to inform software that DF is freezed due to
4139 			 * system fatal error event. driver should not
4140 			 * enable nbio ras in such case. Instead,
4141 			 * check DF RAS
4142 			 */
4143 			adev->nbio.ras = &nbif_v6_3_1_ras;
4144 		break;
4145 	case IP_VERSION(7, 9, 0):
4146 	case IP_VERSION(7, 9, 1):
4147 		if (!adev->gmc.is_app_apu)
4148 			adev->nbio.ras = &nbio_v7_9_ras;
4149 		break;
4150 	default:
4151 		/* nbio ras is not available */
4152 		break;
4153 	}
4154 
4155 	/* nbio ras block needs to be enabled ahead of other ras blocks
4156 	 * to handle fatal error */
4157 	r = amdgpu_nbio_ras_sw_init(adev);
4158 	if (r)
4159 		return r;
4160 
4161 	if (adev->nbio.ras &&
4162 	    adev->nbio.ras->init_ras_controller_interrupt) {
4163 		r = adev->nbio.ras->init_ras_controller_interrupt(adev);
4164 		if (r)
4165 			goto release_con;
4166 	}
4167 
4168 	if (adev->nbio.ras &&
4169 	    adev->nbio.ras->init_ras_err_event_athub_interrupt) {
4170 		r = adev->nbio.ras->init_ras_err_event_athub_interrupt(adev);
4171 		if (r)
4172 			goto release_con;
4173 	}
4174 
4175 	/* Packed socket_id to ras feature mask bits[31:29] */
4176 	if (adev->smuio.funcs &&
4177 	    adev->smuio.funcs->get_socket_id)
4178 		con->features |= ((adev->smuio.funcs->get_socket_id(adev)) <<
4179 					AMDGPU_RAS_FEATURES_SOCKETID_SHIFT);
4180 
4181 	/* Get RAS schema for particular SOC */
4182 	con->schema = amdgpu_get_ras_schema(adev);
4183 
4184 	amdgpu_ras_init_reserved_vram_size(adev);
4185 
4186 	if (amdgpu_ras_fs_init(adev)) {
4187 		r = -EINVAL;
4188 		goto release_con;
4189 	}
4190 
4191 	if (amdgpu_ras_aca_is_supported(adev)) {
4192 		if (amdgpu_aca_is_enabled(adev))
4193 			r = amdgpu_aca_init(adev);
4194 		else
4195 			r = amdgpu_mca_init(adev);
4196 		if (r)
4197 			goto release_con;
4198 	}
4199 
4200 	con->init_task_pid = task_pid_nr(current);
4201 	get_task_comm(con->init_task_comm, current);
4202 
4203 	mutex_init(&con->critical_region_lock);
4204 	INIT_LIST_HEAD(&con->critical_region_head);
4205 
4206 	dev_info(adev->dev, "RAS INFO: ras initialized successfully, "
4207 		 "hardware ability[%x] ras_mask[%x]\n",
4208 		 adev->ras_hw_enabled, adev->ras_enabled);
4209 
4210 	return 0;
4211 release_con:
4212 	amdgpu_ras_set_context(adev, NULL);
4213 	kfree(con);
4214 
4215 	return r;
4216 }
4217 
4218 int amdgpu_persistent_edc_harvesting_supported(struct amdgpu_device *adev)
4219 {
4220 	if (adev->gmc.xgmi.connected_to_cpu ||
4221 	    adev->gmc.is_app_apu)
4222 		return 1;
4223 	return 0;
4224 }
4225 
4226 static int amdgpu_persistent_edc_harvesting(struct amdgpu_device *adev,
4227 					struct ras_common_if *ras_block)
4228 {
4229 	struct ras_query_if info = {
4230 		.head = *ras_block,
4231 	};
4232 
4233 	if (!amdgpu_persistent_edc_harvesting_supported(adev))
4234 		return 0;
4235 
4236 	if (amdgpu_ras_query_error_status(adev, &info) != 0)
4237 		DRM_WARN("RAS init harvest failure");
4238 
4239 	if (amdgpu_ras_reset_error_status(adev, ras_block->block) != 0)
4240 		DRM_WARN("RAS init harvest reset failure");
4241 
4242 	return 0;
4243 }
4244 
4245 bool amdgpu_ras_is_poison_mode_supported(struct amdgpu_device *adev)
4246 {
4247        struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4248 
4249        if (!con)
4250                return false;
4251 
4252        return con->poison_supported;
4253 }
4254 
4255 /* helper function to handle common stuff in ip late init phase */
4256 int amdgpu_ras_block_late_init(struct amdgpu_device *adev,
4257 			 struct ras_common_if *ras_block)
4258 {
4259 	struct amdgpu_ras_block_object *ras_obj = NULL;
4260 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4261 	struct ras_query_if *query_info;
4262 	unsigned long ue_count, ce_count;
4263 	int r;
4264 
4265 	/* disable RAS feature per IP block if it is not supported */
4266 	if (!amdgpu_ras_is_supported(adev, ras_block->block)) {
4267 		amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
4268 		return 0;
4269 	}
4270 
4271 	r = amdgpu_ras_feature_enable_on_boot(adev, ras_block, 1);
4272 	if (r) {
4273 		if (adev->in_suspend || amdgpu_reset_in_recovery(adev)) {
4274 			/* in resume phase, if fail to enable ras,
4275 			 * clean up all ras fs nodes, and disable ras */
4276 			goto cleanup;
4277 		} else
4278 			return r;
4279 	}
4280 
4281 	/* check for errors on warm reset edc persisant supported ASIC */
4282 	amdgpu_persistent_edc_harvesting(adev, ras_block);
4283 
4284 	/* in resume phase, no need to create ras fs node */
4285 	if (adev->in_suspend || amdgpu_reset_in_recovery(adev))
4286 		return 0;
4287 
4288 	ras_obj = container_of(ras_block, struct amdgpu_ras_block_object, ras_comm);
4289 	if (ras_obj->ras_cb || (ras_obj->hw_ops &&
4290 	    (ras_obj->hw_ops->query_poison_status ||
4291 	    ras_obj->hw_ops->handle_poison_consumption))) {
4292 		r = amdgpu_ras_interrupt_add_handler(adev, ras_block);
4293 		if (r)
4294 			goto cleanup;
4295 	}
4296 
4297 	if (ras_obj->hw_ops &&
4298 	    (ras_obj->hw_ops->query_ras_error_count ||
4299 	     ras_obj->hw_ops->query_ras_error_status)) {
4300 		r = amdgpu_ras_sysfs_create(adev, ras_block);
4301 		if (r)
4302 			goto interrupt;
4303 
4304 		/* Those are the cached values at init.
4305 		 */
4306 		query_info = kzalloc(sizeof(*query_info), GFP_KERNEL);
4307 		if (!query_info)
4308 			return -ENOMEM;
4309 		memcpy(&query_info->head, ras_block, sizeof(struct ras_common_if));
4310 
4311 		if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count, query_info) == 0) {
4312 			atomic_set(&con->ras_ce_count, ce_count);
4313 			atomic_set(&con->ras_ue_count, ue_count);
4314 		}
4315 
4316 		kfree(query_info);
4317 	}
4318 
4319 	return 0;
4320 
4321 interrupt:
4322 	if (ras_obj->ras_cb)
4323 		amdgpu_ras_interrupt_remove_handler(adev, ras_block);
4324 cleanup:
4325 	amdgpu_ras_feature_enable(adev, ras_block, 0);
4326 	return r;
4327 }
4328 
4329 static int amdgpu_ras_block_late_init_default(struct amdgpu_device *adev,
4330 			 struct ras_common_if *ras_block)
4331 {
4332 	return amdgpu_ras_block_late_init(adev, ras_block);
4333 }
4334 
4335 /* helper function to remove ras fs node and interrupt handler */
4336 void amdgpu_ras_block_late_fini(struct amdgpu_device *adev,
4337 			  struct ras_common_if *ras_block)
4338 {
4339 	struct amdgpu_ras_block_object *ras_obj;
4340 	if (!ras_block)
4341 		return;
4342 
4343 	amdgpu_ras_sysfs_remove(adev, ras_block);
4344 
4345 	ras_obj = container_of(ras_block, struct amdgpu_ras_block_object, ras_comm);
4346 	if (ras_obj->ras_cb)
4347 		amdgpu_ras_interrupt_remove_handler(adev, ras_block);
4348 }
4349 
4350 static void amdgpu_ras_block_late_fini_default(struct amdgpu_device *adev,
4351 			  struct ras_common_if *ras_block)
4352 {
4353 	return amdgpu_ras_block_late_fini(adev, ras_block);
4354 }
4355 
4356 /* do some init work after IP late init as dependence.
4357  * and it runs in resume/gpu reset/booting up cases.
4358  */
4359 void amdgpu_ras_resume(struct amdgpu_device *adev)
4360 {
4361 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4362 	struct ras_manager *obj, *tmp;
4363 
4364 	if (!adev->ras_enabled || !con) {
4365 		/* clean ras context for VEGA20 Gaming after send ras disable cmd */
4366 		amdgpu_release_ras_context(adev);
4367 
4368 		return;
4369 	}
4370 
4371 	if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
4372 		/* Set up all other IPs which are not implemented. There is a
4373 		 * tricky thing that IP's actual ras error type should be
4374 		 * MULTI_UNCORRECTABLE, but as driver does not handle it, so
4375 		 * ERROR_NONE make sense anyway.
4376 		 */
4377 		amdgpu_ras_enable_all_features(adev, 1);
4378 
4379 		/* We enable ras on all hw_supported block, but as boot
4380 		 * parameter might disable some of them and one or more IP has
4381 		 * not implemented yet. So we disable them on behalf.
4382 		 */
4383 		list_for_each_entry_safe(obj, tmp, &con->head, node) {
4384 			if (!amdgpu_ras_is_supported(adev, obj->head.block)) {
4385 				amdgpu_ras_feature_enable(adev, &obj->head, 0);
4386 				/* there should be no any reference. */
4387 				WARN_ON(alive_obj(obj));
4388 			}
4389 		}
4390 	}
4391 }
4392 
4393 void amdgpu_ras_suspend(struct amdgpu_device *adev)
4394 {
4395 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4396 
4397 	if (!adev->ras_enabled || !con)
4398 		return;
4399 
4400 	amdgpu_ras_disable_all_features(adev, 0);
4401 	/* Make sure all ras objects are disabled. */
4402 	if (AMDGPU_RAS_GET_FEATURES(con->features))
4403 		amdgpu_ras_disable_all_features(adev, 1);
4404 }
4405 
4406 int amdgpu_ras_late_init(struct amdgpu_device *adev)
4407 {
4408 	struct amdgpu_ras_block_list *node, *tmp;
4409 	struct amdgpu_ras_block_object *obj;
4410 	int r;
4411 
4412 	amdgpu_ras_event_mgr_init(adev);
4413 
4414 	if (amdgpu_ras_aca_is_supported(adev)) {
4415 		if (amdgpu_reset_in_recovery(adev)) {
4416 			if (amdgpu_aca_is_enabled(adev))
4417 				r = amdgpu_aca_reset(adev);
4418 			else
4419 				r = amdgpu_mca_reset(adev);
4420 			if (r)
4421 				return r;
4422 		}
4423 
4424 		if (!amdgpu_sriov_vf(adev)) {
4425 			if (amdgpu_aca_is_enabled(adev))
4426 				amdgpu_ras_set_aca_debug_mode(adev, false);
4427 			else
4428 				amdgpu_ras_set_mca_debug_mode(adev, false);
4429 		}
4430 	}
4431 
4432 	/* Guest side doesn't need init ras feature */
4433 	if (amdgpu_sriov_vf(adev) && !amdgpu_sriov_ras_telemetry_en(adev))
4434 		return 0;
4435 
4436 	list_for_each_entry_safe(node, tmp, &adev->ras_list, node) {
4437 		obj = node->ras_obj;
4438 		if (!obj) {
4439 			dev_warn(adev->dev, "Warning: abnormal ras list node.\n");
4440 			continue;
4441 		}
4442 
4443 		if (!amdgpu_ras_is_supported(adev, obj->ras_comm.block))
4444 			continue;
4445 
4446 		if (obj->ras_late_init) {
4447 			r = obj->ras_late_init(adev, &obj->ras_comm);
4448 			if (r) {
4449 				dev_err(adev->dev, "%s failed to execute ras_late_init! ret:%d\n",
4450 					obj->ras_comm.name, r);
4451 				return r;
4452 			}
4453 		} else
4454 			amdgpu_ras_block_late_init_default(adev, &obj->ras_comm);
4455 	}
4456 
4457 	return 0;
4458 }
4459 
4460 /* do some fini work before IP fini as dependence */
4461 int amdgpu_ras_pre_fini(struct amdgpu_device *adev)
4462 {
4463 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4464 
4465 	if (!adev->ras_enabled || !con)
4466 		return 0;
4467 
4468 
4469 	/* Need disable ras on all IPs here before ip [hw/sw]fini */
4470 	if (AMDGPU_RAS_GET_FEATURES(con->features))
4471 		amdgpu_ras_disable_all_features(adev, 0);
4472 	amdgpu_ras_recovery_fini(adev);
4473 	return 0;
4474 }
4475 
4476 int amdgpu_ras_fini(struct amdgpu_device *adev)
4477 {
4478 	struct amdgpu_ras_block_list *ras_node, *tmp;
4479 	struct amdgpu_ras_block_object *obj = NULL;
4480 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4481 
4482 	if (!adev->ras_enabled || !con)
4483 		return 0;
4484 
4485 	amdgpu_ras_critical_region_fini(adev);
4486 	mutex_destroy(&con->critical_region_lock);
4487 
4488 	list_for_each_entry_safe(ras_node, tmp, &adev->ras_list, node) {
4489 		if (ras_node->ras_obj) {
4490 			obj = ras_node->ras_obj;
4491 			if (amdgpu_ras_is_supported(adev, obj->ras_comm.block) &&
4492 			    obj->ras_fini)
4493 				obj->ras_fini(adev, &obj->ras_comm);
4494 			else
4495 				amdgpu_ras_block_late_fini_default(adev, &obj->ras_comm);
4496 		}
4497 
4498 		/* Clear ras blocks from ras_list and free ras block list node */
4499 		list_del(&ras_node->node);
4500 		kfree(ras_node);
4501 	}
4502 
4503 	amdgpu_ras_fs_fini(adev);
4504 	amdgpu_ras_interrupt_remove_all(adev);
4505 
4506 	if (amdgpu_ras_aca_is_supported(adev)) {
4507 		if (amdgpu_aca_is_enabled(adev))
4508 			amdgpu_aca_fini(adev);
4509 		else
4510 			amdgpu_mca_fini(adev);
4511 	}
4512 
4513 	WARN(AMDGPU_RAS_GET_FEATURES(con->features), "Feature mask is not cleared");
4514 
4515 	if (AMDGPU_RAS_GET_FEATURES(con->features))
4516 		amdgpu_ras_disable_all_features(adev, 0);
4517 
4518 	cancel_delayed_work_sync(&con->ras_counte_delay_work);
4519 
4520 	amdgpu_ras_set_context(adev, NULL);
4521 	kfree(con);
4522 
4523 	return 0;
4524 }
4525 
4526 bool amdgpu_ras_get_fed_status(struct amdgpu_device *adev)
4527 {
4528 	struct amdgpu_ras *ras;
4529 
4530 	ras = amdgpu_ras_get_context(adev);
4531 	if (!ras)
4532 		return false;
4533 
4534 	return test_bit(AMDGPU_RAS_BLOCK__LAST, &ras->ras_err_state);
4535 }
4536 
4537 void amdgpu_ras_set_fed(struct amdgpu_device *adev, bool status)
4538 {
4539 	struct amdgpu_ras *ras;
4540 
4541 	ras = amdgpu_ras_get_context(adev);
4542 	if (ras) {
4543 		if (status)
4544 			set_bit(AMDGPU_RAS_BLOCK__LAST, &ras->ras_err_state);
4545 		else
4546 			clear_bit(AMDGPU_RAS_BLOCK__LAST, &ras->ras_err_state);
4547 	}
4548 }
4549 
4550 void amdgpu_ras_clear_err_state(struct amdgpu_device *adev)
4551 {
4552 	struct amdgpu_ras *ras;
4553 
4554 	ras = amdgpu_ras_get_context(adev);
4555 	if (ras) {
4556 		ras->ras_err_state = 0;
4557 		ras->gpu_reset_flags = 0;
4558 	}
4559 }
4560 
4561 void amdgpu_ras_set_err_poison(struct amdgpu_device *adev,
4562 			       enum amdgpu_ras_block block)
4563 {
4564 	struct amdgpu_ras *ras;
4565 
4566 	ras = amdgpu_ras_get_context(adev);
4567 	if (ras)
4568 		set_bit(block, &ras->ras_err_state);
4569 }
4570 
4571 bool amdgpu_ras_is_err_state(struct amdgpu_device *adev, int block)
4572 {
4573 	struct amdgpu_ras *ras;
4574 
4575 	ras = amdgpu_ras_get_context(adev);
4576 	if (ras) {
4577 		if (block == AMDGPU_RAS_BLOCK__ANY)
4578 			return (ras->ras_err_state != 0);
4579 		else
4580 			return test_bit(block, &ras->ras_err_state) ||
4581 			       test_bit(AMDGPU_RAS_BLOCK__LAST,
4582 					&ras->ras_err_state);
4583 	}
4584 
4585 	return false;
4586 }
4587 
4588 static struct ras_event_manager *__get_ras_event_mgr(struct amdgpu_device *adev)
4589 {
4590 	struct amdgpu_ras *ras;
4591 
4592 	ras = amdgpu_ras_get_context(adev);
4593 	if (!ras)
4594 		return NULL;
4595 
4596 	return ras->event_mgr;
4597 }
4598 
4599 int amdgpu_ras_mark_ras_event_caller(struct amdgpu_device *adev, enum ras_event_type type,
4600 				     const void *caller)
4601 {
4602 	struct ras_event_manager *event_mgr;
4603 	struct ras_event_state *event_state;
4604 	int ret = 0;
4605 
4606 	if (amdgpu_uniras_enabled(adev))
4607 		return 0;
4608 
4609 	if (type >= RAS_EVENT_TYPE_COUNT) {
4610 		ret = -EINVAL;
4611 		goto out;
4612 	}
4613 
4614 	event_mgr = __get_ras_event_mgr(adev);
4615 	if (!event_mgr) {
4616 		ret = -EINVAL;
4617 		goto out;
4618 	}
4619 
4620 	event_state = &event_mgr->event_state[type];
4621 	event_state->last_seqno = atomic64_inc_return(&event_mgr->seqno);
4622 	atomic64_inc(&event_state->count);
4623 
4624 out:
4625 	if (ret && caller)
4626 		dev_warn(adev->dev, "failed mark ras event (%d) in %ps, ret:%d\n",
4627 			 (int)type, caller, ret);
4628 
4629 	return ret;
4630 }
4631 
4632 u64 amdgpu_ras_acquire_event_id(struct amdgpu_device *adev, enum ras_event_type type)
4633 {
4634 	struct ras_event_manager *event_mgr;
4635 	u64 id;
4636 
4637 	if (type >= RAS_EVENT_TYPE_COUNT)
4638 		return RAS_EVENT_INVALID_ID;
4639 
4640 	switch (type) {
4641 	case RAS_EVENT_TYPE_FATAL:
4642 	case RAS_EVENT_TYPE_POISON_CREATION:
4643 	case RAS_EVENT_TYPE_POISON_CONSUMPTION:
4644 		event_mgr = __get_ras_event_mgr(adev);
4645 		if (!event_mgr)
4646 			return RAS_EVENT_INVALID_ID;
4647 
4648 		id = event_mgr->event_state[type].last_seqno;
4649 		break;
4650 	case RAS_EVENT_TYPE_INVALID:
4651 	default:
4652 		id = RAS_EVENT_INVALID_ID;
4653 		break;
4654 	}
4655 
4656 	return id;
4657 }
4658 
4659 int amdgpu_ras_global_ras_isr(struct amdgpu_device *adev)
4660 {
4661 	if (atomic_cmpxchg(&amdgpu_ras_in_intr, 0, 1) == 0) {
4662 		struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
4663 		enum ras_event_type type = RAS_EVENT_TYPE_FATAL;
4664 		u64 event_id = RAS_EVENT_INVALID_ID;
4665 
4666 		if (amdgpu_uniras_enabled(adev))
4667 			return 0;
4668 
4669 		if (!amdgpu_ras_mark_ras_event(adev, type))
4670 			event_id = amdgpu_ras_acquire_event_id(adev, type);
4671 
4672 		RAS_EVENT_LOG(adev, event_id, "uncorrectable hardware error"
4673 			      "(ERREVENT_ATHUB_INTERRUPT) detected!\n");
4674 
4675 		amdgpu_ras_set_fed(adev, true);
4676 		ras->gpu_reset_flags |= AMDGPU_RAS_GPU_RESET_MODE1_RESET;
4677 		amdgpu_ras_reset_gpu(adev);
4678 	}
4679 
4680 	return -EBUSY;
4681 }
4682 
4683 bool amdgpu_ras_need_emergency_restart(struct amdgpu_device *adev)
4684 {
4685 	if (adev->asic_type == CHIP_VEGA20 &&
4686 	    adev->pm.fw_version <= 0x283400) {
4687 		return !(amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) &&
4688 				amdgpu_ras_intr_triggered();
4689 	}
4690 
4691 	return false;
4692 }
4693 
4694 void amdgpu_release_ras_context(struct amdgpu_device *adev)
4695 {
4696 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4697 
4698 	if (!con)
4699 		return;
4700 
4701 	if (!adev->ras_enabled && con->features & BIT(AMDGPU_RAS_BLOCK__GFX)) {
4702 		con->features &= ~BIT(AMDGPU_RAS_BLOCK__GFX);
4703 		amdgpu_ras_set_context(adev, NULL);
4704 		kfree(con);
4705 	}
4706 }
4707 
4708 #ifdef CONFIG_X86_MCE_AMD
4709 static struct amdgpu_device *find_adev(uint32_t node_id)
4710 {
4711 	int i;
4712 	struct amdgpu_device *adev = NULL;
4713 
4714 	for (i = 0; i < mce_adev_list.num_gpu; i++) {
4715 		adev = mce_adev_list.devs[i];
4716 
4717 		if (adev && adev->gmc.xgmi.connected_to_cpu &&
4718 		    adev->gmc.xgmi.physical_node_id == node_id)
4719 			break;
4720 		adev = NULL;
4721 	}
4722 
4723 	return adev;
4724 }
4725 
4726 #define GET_MCA_IPID_GPUID(m)	(((m) >> 44) & 0xF)
4727 #define GET_UMC_INST(m)		(((m) >> 21) & 0x7)
4728 #define GET_CHAN_INDEX(m)	((((m) >> 12) & 0x3) | (((m) >> 18) & 0x4))
4729 #define GPU_ID_OFFSET		8
4730 
4731 static int amdgpu_bad_page_notifier(struct notifier_block *nb,
4732 				    unsigned long val, void *data)
4733 {
4734 	struct mce *m = (struct mce *)data;
4735 	struct amdgpu_device *adev = NULL;
4736 	uint32_t gpu_id = 0;
4737 	uint32_t umc_inst = 0, ch_inst = 0;
4738 
4739 	/*
4740 	 * If the error was generated in UMC_V2, which belongs to GPU UMCs,
4741 	 * and error occurred in DramECC (Extended error code = 0) then only
4742 	 * process the error, else bail out.
4743 	 */
4744 	if (!m || !((smca_get_bank_type(m->extcpu, m->bank) == SMCA_UMC_V2) &&
4745 		    (XEC(m->status, 0x3f) == 0x0)))
4746 		return NOTIFY_DONE;
4747 
4748 	/*
4749 	 * If it is correctable error, return.
4750 	 */
4751 	if (mce_is_correctable(m))
4752 		return NOTIFY_OK;
4753 
4754 	/*
4755 	 * GPU Id is offset by GPU_ID_OFFSET in MCA_IPID_UMC register.
4756 	 */
4757 	gpu_id = GET_MCA_IPID_GPUID(m->ipid) - GPU_ID_OFFSET;
4758 
4759 	adev = find_adev(gpu_id);
4760 	if (!adev) {
4761 		DRM_WARN("%s: Unable to find adev for gpu_id: %d\n", __func__,
4762 								gpu_id);
4763 		return NOTIFY_DONE;
4764 	}
4765 
4766 	/*
4767 	 * If it is uncorrectable error, then find out UMC instance and
4768 	 * channel index.
4769 	 */
4770 	umc_inst = GET_UMC_INST(m->ipid);
4771 	ch_inst = GET_CHAN_INDEX(m->ipid);
4772 
4773 	dev_info(adev->dev, "Uncorrectable error detected in UMC inst: %d, chan_idx: %d",
4774 			     umc_inst, ch_inst);
4775 
4776 	if (!amdgpu_umc_page_retirement_mca(adev, m->addr, ch_inst, umc_inst))
4777 		return NOTIFY_OK;
4778 	else
4779 		return NOTIFY_DONE;
4780 }
4781 
4782 static struct notifier_block amdgpu_bad_page_nb = {
4783 	.notifier_call  = amdgpu_bad_page_notifier,
4784 	.priority       = MCE_PRIO_UC,
4785 };
4786 
4787 static void amdgpu_register_bad_pages_mca_notifier(struct amdgpu_device *adev)
4788 {
4789 	/*
4790 	 * Add the adev to the mce_adev_list.
4791 	 * During mode2 reset, amdgpu device is temporarily
4792 	 * removed from the mgpu_info list which can cause
4793 	 * page retirement to fail.
4794 	 * Use this list instead of mgpu_info to find the amdgpu
4795 	 * device on which the UMC error was reported.
4796 	 */
4797 	mce_adev_list.devs[mce_adev_list.num_gpu++] = adev;
4798 
4799 	/*
4800 	 * Register the x86 notifier only once
4801 	 * with MCE subsystem.
4802 	 */
4803 	if (notifier_registered == false) {
4804 		mce_register_decode_chain(&amdgpu_bad_page_nb);
4805 		notifier_registered = true;
4806 	}
4807 }
4808 #endif
4809 
4810 struct amdgpu_ras *amdgpu_ras_get_context(struct amdgpu_device *adev)
4811 {
4812 	if (!adev)
4813 		return NULL;
4814 
4815 	return adev->psp.ras_context.ras;
4816 }
4817 
4818 int amdgpu_ras_set_context(struct amdgpu_device *adev, struct amdgpu_ras *ras_con)
4819 {
4820 	if (!adev)
4821 		return -EINVAL;
4822 
4823 	adev->psp.ras_context.ras = ras_con;
4824 	return 0;
4825 }
4826 
4827 /* check if ras is supported on block, say, sdma, gfx */
4828 int amdgpu_ras_is_supported(struct amdgpu_device *adev,
4829 		unsigned int block)
4830 {
4831 	int ret = 0;
4832 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
4833 
4834 	if (block >= AMDGPU_RAS_BLOCK_COUNT)
4835 		return 0;
4836 
4837 	ret = ras && (adev->ras_enabled & (1 << block));
4838 
4839 	/* For the special asic with mem ecc enabled but sram ecc
4840 	 * not enabled, even if the ras block is not supported on
4841 	 * .ras_enabled, if the asic supports poison mode and the
4842 	 * ras block has ras configuration, it can be considered
4843 	 * that the ras block supports ras function.
4844 	 */
4845 	if (!ret &&
4846 	    (block == AMDGPU_RAS_BLOCK__GFX ||
4847 	     block == AMDGPU_RAS_BLOCK__SDMA ||
4848 	     block == AMDGPU_RAS_BLOCK__VCN ||
4849 	     block == AMDGPU_RAS_BLOCK__JPEG) &&
4850 		(amdgpu_ras_mask & (1 << block)) &&
4851 	    amdgpu_ras_is_poison_mode_supported(adev) &&
4852 	    amdgpu_ras_get_ras_block(adev, block, 0))
4853 		ret = 1;
4854 
4855 	return ret;
4856 }
4857 
4858 int amdgpu_ras_reset_gpu(struct amdgpu_device *adev)
4859 {
4860 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
4861 
4862 	/* mode1 is the only selection for RMA status */
4863 	if (amdgpu_ras_is_rma(adev)) {
4864 		ras->gpu_reset_flags = 0;
4865 		ras->gpu_reset_flags |= AMDGPU_RAS_GPU_RESET_MODE1_RESET;
4866 	}
4867 
4868 	if (atomic_cmpxchg(&ras->in_recovery, 0, 1) == 0) {
4869 		struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
4870 		int hive_ras_recovery = 0;
4871 
4872 		if (hive) {
4873 			hive_ras_recovery = atomic_read(&hive->ras_recovery);
4874 			amdgpu_put_xgmi_hive(hive);
4875 		}
4876 		/* In the case of multiple GPUs, after a GPU has started
4877 		 * resetting all GPUs on hive, other GPUs do not need to
4878 		 * trigger GPU reset again.
4879 		 */
4880 		if (!hive_ras_recovery)
4881 			amdgpu_reset_domain_schedule(ras->adev->reset_domain, &ras->recovery_work);
4882 		else
4883 			atomic_set(&ras->in_recovery, 0);
4884 	} else {
4885 		flush_work(&ras->recovery_work);
4886 		amdgpu_reset_domain_schedule(ras->adev->reset_domain, &ras->recovery_work);
4887 	}
4888 
4889 	return 0;
4890 }
4891 
4892 int amdgpu_ras_set_mca_debug_mode(struct amdgpu_device *adev, bool enable)
4893 {
4894 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4895 	int ret = 0;
4896 
4897 	if (con) {
4898 		ret = amdgpu_mca_smu_set_debug_mode(adev, enable);
4899 		if (!ret)
4900 			con->is_aca_debug_mode = enable;
4901 	}
4902 
4903 	return ret;
4904 }
4905 
4906 int amdgpu_ras_set_aca_debug_mode(struct amdgpu_device *adev, bool enable)
4907 {
4908 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4909 	int ret = 0;
4910 
4911 	if (con) {
4912 		if (amdgpu_aca_is_enabled(adev))
4913 			ret = amdgpu_aca_smu_set_debug_mode(adev, enable);
4914 		else
4915 			ret = amdgpu_mca_smu_set_debug_mode(adev, enable);
4916 		if (!ret)
4917 			con->is_aca_debug_mode = enable;
4918 	}
4919 
4920 	return ret;
4921 }
4922 
4923 bool amdgpu_ras_get_aca_debug_mode(struct amdgpu_device *adev)
4924 {
4925 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4926 	const struct aca_smu_funcs *smu_funcs = adev->aca.smu_funcs;
4927 	const struct amdgpu_mca_smu_funcs *mca_funcs = adev->mca.mca_funcs;
4928 
4929 	if (!con)
4930 		return false;
4931 
4932 	if ((amdgpu_aca_is_enabled(adev) && smu_funcs && smu_funcs->set_debug_mode) ||
4933 	    (!amdgpu_aca_is_enabled(adev) && mca_funcs && mca_funcs->mca_set_debug_mode))
4934 		return con->is_aca_debug_mode;
4935 	else
4936 		return true;
4937 }
4938 
4939 bool amdgpu_ras_get_error_query_mode(struct amdgpu_device *adev,
4940 				     unsigned int *error_query_mode)
4941 {
4942 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
4943 	const struct amdgpu_mca_smu_funcs *mca_funcs = adev->mca.mca_funcs;
4944 	const struct aca_smu_funcs *smu_funcs = adev->aca.smu_funcs;
4945 
4946 	if (!con) {
4947 		*error_query_mode = AMDGPU_RAS_INVALID_ERROR_QUERY;
4948 		return false;
4949 	}
4950 
4951 	if (amdgpu_sriov_vf(adev)) {
4952 		*error_query_mode = AMDGPU_RAS_VIRT_ERROR_COUNT_QUERY;
4953 	} else if ((smu_funcs && smu_funcs->set_debug_mode) || (mca_funcs && mca_funcs->mca_set_debug_mode)) {
4954 		*error_query_mode =
4955 			(con->is_aca_debug_mode) ? AMDGPU_RAS_DIRECT_ERROR_QUERY : AMDGPU_RAS_FIRMWARE_ERROR_QUERY;
4956 	} else {
4957 		*error_query_mode = AMDGPU_RAS_DIRECT_ERROR_QUERY;
4958 	}
4959 
4960 	return true;
4961 }
4962 
4963 /* Register each ip ras block into amdgpu ras */
4964 int amdgpu_ras_register_ras_block(struct amdgpu_device *adev,
4965 		struct amdgpu_ras_block_object *ras_block_obj)
4966 {
4967 	struct amdgpu_ras_block_list *ras_node;
4968 	if (!adev || !ras_block_obj)
4969 		return -EINVAL;
4970 
4971 	ras_node = kzalloc(sizeof(*ras_node), GFP_KERNEL);
4972 	if (!ras_node)
4973 		return -ENOMEM;
4974 
4975 	INIT_LIST_HEAD(&ras_node->node);
4976 	ras_node->ras_obj = ras_block_obj;
4977 	list_add_tail(&ras_node->node, &adev->ras_list);
4978 
4979 	return 0;
4980 }
4981 
4982 void amdgpu_ras_get_error_type_name(uint32_t err_type, char *err_type_name)
4983 {
4984 	if (!err_type_name)
4985 		return;
4986 
4987 	switch (err_type) {
4988 	case AMDGPU_RAS_ERROR__SINGLE_CORRECTABLE:
4989 		sprintf(err_type_name, "correctable");
4990 		break;
4991 	case AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE:
4992 		sprintf(err_type_name, "uncorrectable");
4993 		break;
4994 	default:
4995 		sprintf(err_type_name, "unknown");
4996 		break;
4997 	}
4998 }
4999 
5000 bool amdgpu_ras_inst_get_memory_id_field(struct amdgpu_device *adev,
5001 					 const struct amdgpu_ras_err_status_reg_entry *reg_entry,
5002 					 uint32_t instance,
5003 					 uint32_t *memory_id)
5004 {
5005 	uint32_t err_status_lo_data, err_status_lo_offset;
5006 
5007 	if (!reg_entry)
5008 		return false;
5009 
5010 	err_status_lo_offset =
5011 		AMDGPU_RAS_REG_ENTRY_OFFSET(reg_entry->hwip, instance,
5012 					    reg_entry->seg_lo, reg_entry->reg_lo);
5013 	err_status_lo_data = RREG32(err_status_lo_offset);
5014 
5015 	if ((reg_entry->flags & AMDGPU_RAS_ERR_STATUS_VALID) &&
5016 	    !REG_GET_FIELD(err_status_lo_data, ERR_STATUS_LO, ERR_STATUS_VALID_FLAG))
5017 		return false;
5018 
5019 	*memory_id = REG_GET_FIELD(err_status_lo_data, ERR_STATUS_LO, MEMORY_ID);
5020 
5021 	return true;
5022 }
5023 
5024 bool amdgpu_ras_inst_get_err_cnt_field(struct amdgpu_device *adev,
5025 				       const struct amdgpu_ras_err_status_reg_entry *reg_entry,
5026 				       uint32_t instance,
5027 				       unsigned long *err_cnt)
5028 {
5029 	uint32_t err_status_hi_data, err_status_hi_offset;
5030 
5031 	if (!reg_entry)
5032 		return false;
5033 
5034 	err_status_hi_offset =
5035 		AMDGPU_RAS_REG_ENTRY_OFFSET(reg_entry->hwip, instance,
5036 					    reg_entry->seg_hi, reg_entry->reg_hi);
5037 	err_status_hi_data = RREG32(err_status_hi_offset);
5038 
5039 	if ((reg_entry->flags & AMDGPU_RAS_ERR_INFO_VALID) &&
5040 	    !REG_GET_FIELD(err_status_hi_data, ERR_STATUS_HI, ERR_INFO_VALID_FLAG))
5041 		/* keep the check here in case we need to refer to the result later */
5042 		dev_dbg(adev->dev, "Invalid err_info field\n");
5043 
5044 	/* read err count */
5045 	*err_cnt = REG_GET_FIELD(err_status_hi_data, ERR_STATUS, ERR_CNT);
5046 
5047 	return true;
5048 }
5049 
5050 void amdgpu_ras_inst_query_ras_error_count(struct amdgpu_device *adev,
5051 					   const struct amdgpu_ras_err_status_reg_entry *reg_list,
5052 					   uint32_t reg_list_size,
5053 					   const struct amdgpu_ras_memory_id_entry *mem_list,
5054 					   uint32_t mem_list_size,
5055 					   uint32_t instance,
5056 					   uint32_t err_type,
5057 					   unsigned long *err_count)
5058 {
5059 	uint32_t memory_id;
5060 	unsigned long err_cnt;
5061 	char err_type_name[16];
5062 	uint32_t i, j;
5063 
5064 	for (i = 0; i < reg_list_size; i++) {
5065 		/* query memory_id from err_status_lo */
5066 		if (!amdgpu_ras_inst_get_memory_id_field(adev, &reg_list[i],
5067 							 instance, &memory_id))
5068 			continue;
5069 
5070 		/* query err_cnt from err_status_hi */
5071 		if (!amdgpu_ras_inst_get_err_cnt_field(adev, &reg_list[i],
5072 						       instance, &err_cnt) ||
5073 		    !err_cnt)
5074 			continue;
5075 
5076 		*err_count += err_cnt;
5077 
5078 		/* log the errors */
5079 		amdgpu_ras_get_error_type_name(err_type, err_type_name);
5080 		if (!mem_list) {
5081 			/* memory_list is not supported */
5082 			dev_info(adev->dev,
5083 				 "%ld %s hardware errors detected in %s, instance: %d, memory_id: %d\n",
5084 				 err_cnt, err_type_name,
5085 				 reg_list[i].block_name,
5086 				 instance, memory_id);
5087 		} else {
5088 			for (j = 0; j < mem_list_size; j++) {
5089 				if (memory_id == mem_list[j].memory_id) {
5090 					dev_info(adev->dev,
5091 						 "%ld %s hardware errors detected in %s, instance: %d, memory block: %s\n",
5092 						 err_cnt, err_type_name,
5093 						 reg_list[i].block_name,
5094 						 instance, mem_list[j].name);
5095 					break;
5096 				}
5097 			}
5098 		}
5099 	}
5100 }
5101 
5102 void amdgpu_ras_inst_reset_ras_error_count(struct amdgpu_device *adev,
5103 					   const struct amdgpu_ras_err_status_reg_entry *reg_list,
5104 					   uint32_t reg_list_size,
5105 					   uint32_t instance)
5106 {
5107 	uint32_t err_status_lo_offset, err_status_hi_offset;
5108 	uint32_t i;
5109 
5110 	for (i = 0; i < reg_list_size; i++) {
5111 		err_status_lo_offset =
5112 			AMDGPU_RAS_REG_ENTRY_OFFSET(reg_list[i].hwip, instance,
5113 						    reg_list[i].seg_lo, reg_list[i].reg_lo);
5114 		err_status_hi_offset =
5115 			AMDGPU_RAS_REG_ENTRY_OFFSET(reg_list[i].hwip, instance,
5116 						    reg_list[i].seg_hi, reg_list[i].reg_hi);
5117 		WREG32(err_status_lo_offset, 0);
5118 		WREG32(err_status_hi_offset, 0);
5119 	}
5120 }
5121 
5122 int amdgpu_ras_error_data_init(struct ras_err_data *err_data)
5123 {
5124 	memset(err_data, 0, sizeof(*err_data));
5125 
5126 	INIT_LIST_HEAD(&err_data->err_node_list);
5127 
5128 	return 0;
5129 }
5130 
5131 static void amdgpu_ras_error_node_release(struct ras_err_node *err_node)
5132 {
5133 	if (!err_node)
5134 		return;
5135 
5136 	list_del(&err_node->node);
5137 	kvfree(err_node);
5138 }
5139 
5140 void amdgpu_ras_error_data_fini(struct ras_err_data *err_data)
5141 {
5142 	struct ras_err_node *err_node, *tmp;
5143 
5144 	list_for_each_entry_safe(err_node, tmp, &err_data->err_node_list, node)
5145 		amdgpu_ras_error_node_release(err_node);
5146 }
5147 
5148 static struct ras_err_node *amdgpu_ras_error_find_node_by_id(struct ras_err_data *err_data,
5149 							     struct amdgpu_smuio_mcm_config_info *mcm_info)
5150 {
5151 	struct ras_err_node *err_node;
5152 	struct amdgpu_smuio_mcm_config_info *ref_id;
5153 
5154 	if (!err_data || !mcm_info)
5155 		return NULL;
5156 
5157 	for_each_ras_error(err_node, err_data) {
5158 		ref_id = &err_node->err_info.mcm_info;
5159 
5160 		if (mcm_info->socket_id == ref_id->socket_id &&
5161 		    mcm_info->die_id == ref_id->die_id)
5162 			return err_node;
5163 	}
5164 
5165 	return NULL;
5166 }
5167 
5168 static struct ras_err_node *amdgpu_ras_error_node_new(void)
5169 {
5170 	struct ras_err_node *err_node;
5171 
5172 	err_node = kvzalloc(sizeof(*err_node), GFP_KERNEL);
5173 	if (!err_node)
5174 		return NULL;
5175 
5176 	INIT_LIST_HEAD(&err_node->node);
5177 
5178 	return err_node;
5179 }
5180 
5181 static int ras_err_info_cmp(void *priv, const struct list_head *a, const struct list_head *b)
5182 {
5183 	struct ras_err_node *nodea = container_of(a, struct ras_err_node, node);
5184 	struct ras_err_node *nodeb = container_of(b, struct ras_err_node, node);
5185 	struct amdgpu_smuio_mcm_config_info *infoa = &nodea->err_info.mcm_info;
5186 	struct amdgpu_smuio_mcm_config_info *infob = &nodeb->err_info.mcm_info;
5187 
5188 	if (unlikely(infoa->socket_id != infob->socket_id))
5189 		return infoa->socket_id - infob->socket_id;
5190 	else
5191 		return infoa->die_id - infob->die_id;
5192 
5193 	return 0;
5194 }
5195 
5196 static struct ras_err_info *amdgpu_ras_error_get_info(struct ras_err_data *err_data,
5197 				struct amdgpu_smuio_mcm_config_info *mcm_info)
5198 {
5199 	struct ras_err_node *err_node;
5200 
5201 	err_node = amdgpu_ras_error_find_node_by_id(err_data, mcm_info);
5202 	if (err_node)
5203 		return &err_node->err_info;
5204 
5205 	err_node = amdgpu_ras_error_node_new();
5206 	if (!err_node)
5207 		return NULL;
5208 
5209 	memcpy(&err_node->err_info.mcm_info, mcm_info, sizeof(*mcm_info));
5210 
5211 	err_data->err_list_count++;
5212 	list_add_tail(&err_node->node, &err_data->err_node_list);
5213 	list_sort(NULL, &err_data->err_node_list, ras_err_info_cmp);
5214 
5215 	return &err_node->err_info;
5216 }
5217 
5218 int amdgpu_ras_error_statistic_ue_count(struct ras_err_data *err_data,
5219 					struct amdgpu_smuio_mcm_config_info *mcm_info,
5220 					u64 count)
5221 {
5222 	struct ras_err_info *err_info;
5223 
5224 	if (!err_data || !mcm_info)
5225 		return -EINVAL;
5226 
5227 	if (!count)
5228 		return 0;
5229 
5230 	err_info = amdgpu_ras_error_get_info(err_data, mcm_info);
5231 	if (!err_info)
5232 		return -EINVAL;
5233 
5234 	err_info->ue_count += count;
5235 	err_data->ue_count += count;
5236 
5237 	return 0;
5238 }
5239 
5240 int amdgpu_ras_error_statistic_ce_count(struct ras_err_data *err_data,
5241 					struct amdgpu_smuio_mcm_config_info *mcm_info,
5242 					u64 count)
5243 {
5244 	struct ras_err_info *err_info;
5245 
5246 	if (!err_data || !mcm_info)
5247 		return -EINVAL;
5248 
5249 	if (!count)
5250 		return 0;
5251 
5252 	err_info = amdgpu_ras_error_get_info(err_data, mcm_info);
5253 	if (!err_info)
5254 		return -EINVAL;
5255 
5256 	err_info->ce_count += count;
5257 	err_data->ce_count += count;
5258 
5259 	return 0;
5260 }
5261 
5262 int amdgpu_ras_error_statistic_de_count(struct ras_err_data *err_data,
5263 					struct amdgpu_smuio_mcm_config_info *mcm_info,
5264 					u64 count)
5265 {
5266 	struct ras_err_info *err_info;
5267 
5268 	if (!err_data || !mcm_info)
5269 		return -EINVAL;
5270 
5271 	if (!count)
5272 		return 0;
5273 
5274 	err_info = amdgpu_ras_error_get_info(err_data, mcm_info);
5275 	if (!err_info)
5276 		return -EINVAL;
5277 
5278 	err_info->de_count += count;
5279 	err_data->de_count += count;
5280 
5281 	return 0;
5282 }
5283 
5284 #define mmMP0_SMN_C2PMSG_92	0x1609C
5285 #define mmMP0_SMN_C2PMSG_126	0x160BE
5286 static void amdgpu_ras_boot_time_error_reporting(struct amdgpu_device *adev,
5287 						 u32 instance)
5288 {
5289 	u32 socket_id, aid_id, hbm_id;
5290 	u32 fw_status;
5291 	u32 boot_error;
5292 	u64 reg_addr;
5293 
5294 	/* The pattern for smn addressing in other SOC could be different from
5295 	 * the one for aqua_vanjaram. We should revisit the code if the pattern
5296 	 * is changed. In such case, replace the aqua_vanjaram implementation
5297 	 * with more common helper */
5298 	reg_addr = (mmMP0_SMN_C2PMSG_92 << 2) +
5299 		   aqua_vanjaram_encode_ext_smn_addressing(instance);
5300 	fw_status = amdgpu_device_indirect_rreg_ext(adev, reg_addr);
5301 
5302 	reg_addr = (mmMP0_SMN_C2PMSG_126 << 2) +
5303 		   aqua_vanjaram_encode_ext_smn_addressing(instance);
5304 	boot_error = amdgpu_device_indirect_rreg_ext(adev, reg_addr);
5305 
5306 	socket_id = AMDGPU_RAS_GPU_ERR_SOCKET_ID(boot_error);
5307 	aid_id = AMDGPU_RAS_GPU_ERR_AID_ID(boot_error);
5308 	hbm_id = ((1 == AMDGPU_RAS_GPU_ERR_HBM_ID(boot_error)) ? 0 : 1);
5309 
5310 	if (AMDGPU_RAS_GPU_ERR_MEM_TRAINING(boot_error))
5311 		dev_info(adev->dev,
5312 			 "socket: %d, aid: %d, hbm: %d, fw_status: 0x%x, memory training failed\n",
5313 			 socket_id, aid_id, hbm_id, fw_status);
5314 
5315 	if (AMDGPU_RAS_GPU_ERR_FW_LOAD(boot_error))
5316 		dev_info(adev->dev,
5317 			 "socket: %d, aid: %d, fw_status: 0x%x, firmware load failed at boot time\n",
5318 			 socket_id, aid_id, fw_status);
5319 
5320 	if (AMDGPU_RAS_GPU_ERR_WAFL_LINK_TRAINING(boot_error))
5321 		dev_info(adev->dev,
5322 			 "socket: %d, aid: %d, fw_status: 0x%x, wafl link training failed\n",
5323 			 socket_id, aid_id, fw_status);
5324 
5325 	if (AMDGPU_RAS_GPU_ERR_XGMI_LINK_TRAINING(boot_error))
5326 		dev_info(adev->dev,
5327 			 "socket: %d, aid: %d, fw_status: 0x%x, xgmi link training failed\n",
5328 			 socket_id, aid_id, fw_status);
5329 
5330 	if (AMDGPU_RAS_GPU_ERR_USR_CP_LINK_TRAINING(boot_error))
5331 		dev_info(adev->dev,
5332 			 "socket: %d, aid: %d, fw_status: 0x%x, usr cp link training failed\n",
5333 			 socket_id, aid_id, fw_status);
5334 
5335 	if (AMDGPU_RAS_GPU_ERR_USR_DP_LINK_TRAINING(boot_error))
5336 		dev_info(adev->dev,
5337 			 "socket: %d, aid: %d, fw_status: 0x%x, usr dp link training failed\n",
5338 			 socket_id, aid_id, fw_status);
5339 
5340 	if (AMDGPU_RAS_GPU_ERR_HBM_MEM_TEST(boot_error))
5341 		dev_info(adev->dev,
5342 			 "socket: %d, aid: %d, hbm: %d, fw_status: 0x%x, hbm memory test failed\n",
5343 			 socket_id, aid_id, hbm_id, fw_status);
5344 
5345 	if (AMDGPU_RAS_GPU_ERR_HBM_BIST_TEST(boot_error))
5346 		dev_info(adev->dev,
5347 			 "socket: %d, aid: %d, hbm: %d, fw_status: 0x%x, hbm bist test failed\n",
5348 			 socket_id, aid_id, hbm_id, fw_status);
5349 
5350 	if (AMDGPU_RAS_GPU_ERR_DATA_ABORT(boot_error))
5351 		dev_info(adev->dev,
5352 			 "socket: %d, aid: %d, fw_status: 0x%x, data abort exception\n",
5353 			 socket_id, aid_id, fw_status);
5354 
5355 	if (AMDGPU_RAS_GPU_ERR_GENERIC(boot_error))
5356 		dev_info(adev->dev,
5357 			 "socket: %d, aid: %d, fw_status: 0x%x, Boot Controller Generic Error\n",
5358 			 socket_id, aid_id, fw_status);
5359 }
5360 
5361 static bool amdgpu_ras_boot_error_detected(struct amdgpu_device *adev,
5362 					   u32 instance)
5363 {
5364 	u64 reg_addr;
5365 	u32 reg_data;
5366 	int retry_loop;
5367 
5368 	reg_addr = (mmMP0_SMN_C2PMSG_92 << 2) +
5369 		   aqua_vanjaram_encode_ext_smn_addressing(instance);
5370 
5371 	for (retry_loop = 0; retry_loop < AMDGPU_RAS_BOOT_STATUS_POLLING_LIMIT; retry_loop++) {
5372 		reg_data = amdgpu_device_indirect_rreg_ext(adev, reg_addr);
5373 		if ((reg_data & AMDGPU_RAS_BOOT_STATUS_MASK) == AMDGPU_RAS_BOOT_STEADY_STATUS)
5374 			return false;
5375 		else
5376 			msleep(1);
5377 	}
5378 
5379 	return true;
5380 }
5381 
5382 void amdgpu_ras_query_boot_status(struct amdgpu_device *adev, u32 num_instances)
5383 {
5384 	u32 i;
5385 
5386 	for (i = 0; i < num_instances; i++) {
5387 		if (amdgpu_ras_boot_error_detected(adev, i))
5388 			amdgpu_ras_boot_time_error_reporting(adev, i);
5389 	}
5390 }
5391 
5392 int amdgpu_ras_reserve_page(struct amdgpu_device *adev, uint64_t pfn)
5393 {
5394 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
5395 	struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr;
5396 	uint64_t start = pfn << AMDGPU_GPU_PAGE_SHIFT;
5397 	int ret = 0;
5398 
5399 	if (amdgpu_ras_check_critical_address(adev, start))
5400 		return 0;
5401 
5402 	mutex_lock(&con->page_rsv_lock);
5403 	ret = amdgpu_vram_mgr_query_page_status(mgr, start);
5404 	if (ret == -ENOENT)
5405 		ret = amdgpu_vram_mgr_reserve_range(mgr, start, AMDGPU_GPU_PAGE_SIZE);
5406 	mutex_unlock(&con->page_rsv_lock);
5407 
5408 	return ret;
5409 }
5410 
5411 void amdgpu_ras_event_log_print(struct amdgpu_device *adev, u64 event_id,
5412 				const char *fmt, ...)
5413 {
5414 	struct va_format vaf;
5415 	va_list args;
5416 
5417 	va_start(args, fmt);
5418 	vaf.fmt = fmt;
5419 	vaf.va = &args;
5420 
5421 	if (RAS_EVENT_ID_IS_VALID(event_id))
5422 		dev_printk(KERN_INFO, adev->dev, "{%llu}%pV", event_id, &vaf);
5423 	else
5424 		dev_printk(KERN_INFO, adev->dev, "%pV", &vaf);
5425 
5426 	va_end(args);
5427 }
5428 
5429 bool amdgpu_ras_is_rma(struct amdgpu_device *adev)
5430 {
5431 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
5432 
5433 	if (amdgpu_uniras_enabled(adev))
5434 		return amdgpu_ras_mgr_is_rma(adev);
5435 
5436 	if (!con)
5437 		return false;
5438 
5439 	return con->is_rma;
5440 }
5441 
5442 int amdgpu_ras_add_critical_region(struct amdgpu_device *adev,
5443 			struct amdgpu_bo *bo)
5444 {
5445 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
5446 	struct amdgpu_vram_mgr_resource *vres;
5447 	struct ras_critical_region *region;
5448 	struct drm_buddy_block *block;
5449 	int ret = 0;
5450 
5451 	if (!bo || !bo->tbo.resource)
5452 		return -EINVAL;
5453 
5454 	vres = to_amdgpu_vram_mgr_resource(bo->tbo.resource);
5455 
5456 	mutex_lock(&con->critical_region_lock);
5457 
5458 	/* Check if the bo had been recorded */
5459 	list_for_each_entry(region, &con->critical_region_head, node)
5460 		if (region->bo == bo)
5461 			goto out;
5462 
5463 	/* Record new critical amdgpu bo */
5464 	list_for_each_entry(block, &vres->blocks, link) {
5465 		region = kzalloc(sizeof(*region), GFP_KERNEL);
5466 		if (!region) {
5467 			ret = -ENOMEM;
5468 			goto out;
5469 		}
5470 		region->bo = bo;
5471 		region->start = amdgpu_vram_mgr_block_start(block);
5472 		region->size = amdgpu_vram_mgr_block_size(block);
5473 		list_add_tail(&region->node, &con->critical_region_head);
5474 	}
5475 
5476 out:
5477 	mutex_unlock(&con->critical_region_lock);
5478 
5479 	return ret;
5480 }
5481 
5482 static void amdgpu_ras_critical_region_init(struct amdgpu_device *adev)
5483 {
5484 	amdgpu_ras_add_critical_region(adev, adev->mman.fw_reserved_memory);
5485 }
5486 
5487 static void amdgpu_ras_critical_region_fini(struct amdgpu_device *adev)
5488 {
5489 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
5490 	struct ras_critical_region *region, *tmp;
5491 
5492 	mutex_lock(&con->critical_region_lock);
5493 	list_for_each_entry_safe(region, tmp, &con->critical_region_head, node) {
5494 		list_del(&region->node);
5495 		kfree(region);
5496 	}
5497 	mutex_unlock(&con->critical_region_lock);
5498 }
5499 
5500 bool amdgpu_ras_check_critical_address(struct amdgpu_device *adev, uint64_t addr)
5501 {
5502 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
5503 	struct ras_critical_region *region;
5504 	bool ret = false;
5505 
5506 	mutex_lock(&con->critical_region_lock);
5507 	list_for_each_entry(region, &con->critical_region_head, node) {
5508 		if ((region->start <= addr) &&
5509 		    (addr < (region->start + region->size))) {
5510 			ret = true;
5511 			break;
5512 		}
5513 	}
5514 	mutex_unlock(&con->critical_region_lock);
5515 
5516 	return ret;
5517 }
5518