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 
32 #include "amdgpu.h"
33 #include "amdgpu_ras.h"
34 #include "amdgpu_atomfirmware.h"
35 #include "amdgpu_xgmi.h"
36 #include "ivsrcid/nbio/irqsrcs_nbif_7_4.h"
37 #include "atom.h"
38 #ifdef CONFIG_X86_MCE_AMD
39 #include <asm/mce.h>
40 
41 static bool notifier_registered;
42 #endif
43 static const char *RAS_FS_NAME = "ras";
44 
45 const char *ras_error_string[] = {
46 	"none",
47 	"parity",
48 	"single_correctable",
49 	"multi_uncorrectable",
50 	"poison",
51 };
52 
53 const char *ras_block_string[] = {
54 	"umc",
55 	"sdma",
56 	"gfx",
57 	"mmhub",
58 	"athub",
59 	"pcie_bif",
60 	"hdp",
61 	"xgmi_wafl",
62 	"df",
63 	"smn",
64 	"sem",
65 	"mp0",
66 	"mp1",
67 	"fuse",
68 	"mca",
69 	"vcn",
70 	"jpeg",
71 };
72 
73 const char *ras_mca_block_string[] = {
74 	"mca_mp0",
75 	"mca_mp1",
76 	"mca_mpio",
77 	"mca_iohc",
78 };
79 
80 struct amdgpu_ras_block_list {
81 	/* ras block link */
82 	struct list_head node;
83 
84 	struct amdgpu_ras_block_object *ras_obj;
85 };
86 
87 const char *get_ras_block_str(struct ras_common_if *ras_block)
88 {
89 	if (!ras_block)
90 		return "NULL";
91 
92 	if (ras_block->block >= AMDGPU_RAS_BLOCK_COUNT)
93 		return "OUT OF RANGE";
94 
95 	if (ras_block->block == AMDGPU_RAS_BLOCK__MCA)
96 		return ras_mca_block_string[ras_block->sub_block_index];
97 
98 	return ras_block_string[ras_block->block];
99 }
100 
101 #define ras_block_str(_BLOCK_) \
102 	(((_BLOCK_) < ARRAY_SIZE(ras_block_string)) ? ras_block_string[_BLOCK_] : "Out Of Range")
103 
104 #define ras_err_str(i) (ras_error_string[ffs(i)])
105 
106 #define RAS_DEFAULT_FLAGS (AMDGPU_RAS_FLAG_INIT_BY_VBIOS)
107 
108 /* inject address is 52 bits */
109 #define	RAS_UMC_INJECT_ADDR_LIMIT	(0x1ULL << 52)
110 
111 /* typical ECC bad page rate is 1 bad page per 100MB VRAM */
112 #define RAS_BAD_PAGE_COVER              (100 * 1024 * 1024ULL)
113 
114 enum amdgpu_ras_retire_page_reservation {
115 	AMDGPU_RAS_RETIRE_PAGE_RESERVED,
116 	AMDGPU_RAS_RETIRE_PAGE_PENDING,
117 	AMDGPU_RAS_RETIRE_PAGE_FAULT,
118 };
119 
120 atomic_t amdgpu_ras_in_intr = ATOMIC_INIT(0);
121 
122 static bool amdgpu_ras_check_bad_page_unlock(struct amdgpu_ras *con,
123 				uint64_t addr);
124 static bool amdgpu_ras_check_bad_page(struct amdgpu_device *adev,
125 				uint64_t addr);
126 #ifdef CONFIG_X86_MCE_AMD
127 static void amdgpu_register_bad_pages_mca_notifier(struct amdgpu_device *adev);
128 struct mce_notifier_adev_list {
129 	struct amdgpu_device *devs[MAX_GPU_INSTANCE];
130 	int num_gpu;
131 };
132 static struct mce_notifier_adev_list mce_adev_list;
133 #endif
134 
135 void amdgpu_ras_set_error_query_ready(struct amdgpu_device *adev, bool ready)
136 {
137 	if (adev && amdgpu_ras_get_context(adev))
138 		amdgpu_ras_get_context(adev)->error_query_ready = ready;
139 }
140 
141 static bool amdgpu_ras_get_error_query_ready(struct amdgpu_device *adev)
142 {
143 	if (adev && amdgpu_ras_get_context(adev))
144 		return amdgpu_ras_get_context(adev)->error_query_ready;
145 
146 	return false;
147 }
148 
149 static int amdgpu_reserve_page_direct(struct amdgpu_device *adev, uint64_t address)
150 {
151 	struct ras_err_data err_data = {0, 0, 0, NULL};
152 	struct eeprom_table_record err_rec;
153 
154 	if ((address >= adev->gmc.mc_vram_size) ||
155 	    (address >= RAS_UMC_INJECT_ADDR_LIMIT)) {
156 		dev_warn(adev->dev,
157 		         "RAS WARN: input address 0x%llx is invalid.\n",
158 		         address);
159 		return -EINVAL;
160 	}
161 
162 	if (amdgpu_ras_check_bad_page(adev, address)) {
163 		dev_warn(adev->dev,
164 			 "RAS WARN: 0x%llx has already been marked as bad page!\n",
165 			 address);
166 		return 0;
167 	}
168 
169 	memset(&err_rec, 0x0, sizeof(struct eeprom_table_record));
170 	err_data.err_addr = &err_rec;
171 	amdgpu_umc_fill_error_record(&err_data, address,
172 			(address >> AMDGPU_GPU_PAGE_SHIFT), 0, 0);
173 
174 	if (amdgpu_bad_page_threshold != 0) {
175 		amdgpu_ras_add_bad_pages(adev, err_data.err_addr,
176 					 err_data.err_addr_cnt);
177 		amdgpu_ras_save_bad_pages(adev);
178 	}
179 
180 	dev_warn(adev->dev, "WARNING: THIS IS ONLY FOR TEST PURPOSES AND WILL CORRUPT RAS EEPROM\n");
181 	dev_warn(adev->dev, "Clear EEPROM:\n");
182 	dev_warn(adev->dev, "    echo 1 > /sys/kernel/debug/dri/0/ras/ras_eeprom_reset\n");
183 
184 	return 0;
185 }
186 
187 static ssize_t amdgpu_ras_debugfs_read(struct file *f, char __user *buf,
188 					size_t size, loff_t *pos)
189 {
190 	struct ras_manager *obj = (struct ras_manager *)file_inode(f)->i_private;
191 	struct ras_query_if info = {
192 		.head = obj->head,
193 	};
194 	ssize_t s;
195 	char val[128];
196 
197 	if (amdgpu_ras_query_error_status(obj->adev, &info))
198 		return -EINVAL;
199 
200 	s = snprintf(val, sizeof(val), "%s: %lu\n%s: %lu\n",
201 			"ue", info.ue_count,
202 			"ce", info.ce_count);
203 	if (*pos >= s)
204 		return 0;
205 
206 	s -= *pos;
207 	s = min_t(u64, s, size);
208 
209 
210 	if (copy_to_user(buf, &val[*pos], s))
211 		return -EINVAL;
212 
213 	*pos += s;
214 
215 	return s;
216 }
217 
218 static const struct file_operations amdgpu_ras_debugfs_ops = {
219 	.owner = THIS_MODULE,
220 	.read = amdgpu_ras_debugfs_read,
221 	.write = NULL,
222 	.llseek = default_llseek
223 };
224 
225 static int amdgpu_ras_find_block_id_by_name(const char *name, int *block_id)
226 {
227 	int i;
228 
229 	for (i = 0; i < ARRAY_SIZE(ras_block_string); i++) {
230 		*block_id = i;
231 		if (strcmp(name, ras_block_string[i]) == 0)
232 			return 0;
233 	}
234 	return -EINVAL;
235 }
236 
237 static int amdgpu_ras_debugfs_ctrl_parse_data(struct file *f,
238 		const char __user *buf, size_t size,
239 		loff_t *pos, struct ras_debug_if *data)
240 {
241 	ssize_t s = min_t(u64, 64, size);
242 	char str[65];
243 	char block_name[33];
244 	char err[9] = "ue";
245 	int op = -1;
246 	int block_id;
247 	uint32_t sub_block;
248 	u64 address, value;
249 
250 	if (*pos)
251 		return -EINVAL;
252 	*pos = size;
253 
254 	memset(str, 0, sizeof(str));
255 	memset(data, 0, sizeof(*data));
256 
257 	if (copy_from_user(str, buf, s))
258 		return -EINVAL;
259 
260 	if (sscanf(str, "disable %32s", block_name) == 1)
261 		op = 0;
262 	else if (sscanf(str, "enable %32s %8s", block_name, err) == 2)
263 		op = 1;
264 	else if (sscanf(str, "inject %32s %8s", block_name, err) == 2)
265 		op = 2;
266 	else if (strstr(str, "retire_page") != NULL)
267 		op = 3;
268 	else if (str[0] && str[1] && str[2] && str[3])
269 		/* ascii string, but commands are not matched. */
270 		return -EINVAL;
271 
272 	if (op != -1) {
273 		if (op == 3) {
274 			if (sscanf(str, "%*s 0x%llx", &address) != 1 &&
275 			    sscanf(str, "%*s %llu", &address) != 1)
276 				return -EINVAL;
277 
278 			data->op = op;
279 			data->inject.address = address;
280 
281 			return 0;
282 		}
283 
284 		if (amdgpu_ras_find_block_id_by_name(block_name, &block_id))
285 			return -EINVAL;
286 
287 		data->head.block = block_id;
288 		/* only ue and ce errors are supported */
289 		if (!memcmp("ue", err, 2))
290 			data->head.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
291 		else if (!memcmp("ce", err, 2))
292 			data->head.type = AMDGPU_RAS_ERROR__SINGLE_CORRECTABLE;
293 		else
294 			return -EINVAL;
295 
296 		data->op = op;
297 
298 		if (op == 2) {
299 			if (sscanf(str, "%*s %*s %*s 0x%x 0x%llx 0x%llx",
300 				   &sub_block, &address, &value) != 3 &&
301 			    sscanf(str, "%*s %*s %*s %u %llu %llu",
302 				   &sub_block, &address, &value) != 3)
303 				return -EINVAL;
304 			data->head.sub_block_index = sub_block;
305 			data->inject.address = address;
306 			data->inject.value = value;
307 		}
308 	} else {
309 		if (size < sizeof(*data))
310 			return -EINVAL;
311 
312 		if (copy_from_user(data, buf, sizeof(*data)))
313 			return -EINVAL;
314 	}
315 
316 	return 0;
317 }
318 
319 /**
320  * DOC: AMDGPU RAS debugfs control interface
321  *
322  * The control interface accepts struct ras_debug_if which has two members.
323  *
324  * First member: ras_debug_if::head or ras_debug_if::inject.
325  *
326  * head is used to indicate which IP block will be under control.
327  *
328  * head has four members, they are block, type, sub_block_index, name.
329  * block: which IP will be under control.
330  * type: what kind of error will be enabled/disabled/injected.
331  * sub_block_index: some IPs have subcomponets. say, GFX, sDMA.
332  * name: the name of IP.
333  *
334  * inject has two more members than head, they are address, value.
335  * As their names indicate, inject operation will write the
336  * value to the address.
337  *
338  * The second member: struct ras_debug_if::op.
339  * It has three kinds of operations.
340  *
341  * - 0: disable RAS on the block. Take ::head as its data.
342  * - 1: enable RAS on the block. Take ::head as its data.
343  * - 2: inject errors on the block. Take ::inject as its data.
344  *
345  * How to use the interface?
346  *
347  * In a program
348  *
349  * Copy the struct ras_debug_if in your code and initialize it.
350  * Write the struct to the control interface.
351  *
352  * From shell
353  *
354  * .. code-block:: bash
355  *
356  *	echo "disable <block>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
357  *	echo "enable  <block> <error>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
358  *	echo "inject  <block> <error> <sub-block> <address> <value> > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
359  *
360  * Where N, is the card which you want to affect.
361  *
362  * "disable" requires only the block.
363  * "enable" requires the block and error type.
364  * "inject" requires the block, error type, address, and value.
365  *
366  * The block is one of: umc, sdma, gfx, etc.
367  *	see ras_block_string[] for details
368  *
369  * The error type is one of: ue, ce, where,
370  *	ue is multi-uncorrectable
371  *	ce is single-correctable
372  *
373  * The sub-block is a the sub-block index, pass 0 if there is no sub-block.
374  * The address and value are hexadecimal numbers, leading 0x is optional.
375  *
376  * For instance,
377  *
378  * .. code-block:: bash
379  *
380  *	echo inject umc ue 0x0 0x0 0x0 > /sys/kernel/debug/dri/0/ras/ras_ctrl
381  *	echo inject umc ce 0 0 0 > /sys/kernel/debug/dri/0/ras/ras_ctrl
382  *	echo disable umc > /sys/kernel/debug/dri/0/ras/ras_ctrl
383  *
384  * How to check the result of the operation?
385  *
386  * To check disable/enable, see "ras" features at,
387  * /sys/class/drm/card[0/1/2...]/device/ras/features
388  *
389  * To check inject, see the corresponding error count at,
390  * /sys/class/drm/card[0/1/2...]/device/ras/[gfx|sdma|umc|...]_err_count
391  *
392  * .. note::
393  *	Operations are only allowed on blocks which are supported.
394  *	Check the "ras" mask at /sys/module/amdgpu/parameters/ras_mask
395  *	to see which blocks support RAS on a particular asic.
396  *
397  */
398 static ssize_t amdgpu_ras_debugfs_ctrl_write(struct file *f,
399 					     const char __user *buf,
400 					     size_t size, loff_t *pos)
401 {
402 	struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
403 	struct ras_debug_if data;
404 	int ret = 0;
405 
406 	if (!amdgpu_ras_get_error_query_ready(adev)) {
407 		dev_warn(adev->dev, "RAS WARN: error injection "
408 				"currently inaccessible\n");
409 		return size;
410 	}
411 
412 	ret = amdgpu_ras_debugfs_ctrl_parse_data(f, buf, size, pos, &data);
413 	if (ret)
414 		return ret;
415 
416 	if (data.op == 3) {
417 		ret = amdgpu_reserve_page_direct(adev, data.inject.address);
418 		if (!ret)
419 			return size;
420 		else
421 			return ret;
422 	}
423 
424 	if (!amdgpu_ras_is_supported(adev, data.head.block))
425 		return -EINVAL;
426 
427 	switch (data.op) {
428 	case 0:
429 		ret = amdgpu_ras_feature_enable(adev, &data.head, 0);
430 		break;
431 	case 1:
432 		ret = amdgpu_ras_feature_enable(adev, &data.head, 1);
433 		break;
434 	case 2:
435 		if ((data.inject.address >= adev->gmc.mc_vram_size) ||
436 		    (data.inject.address >= RAS_UMC_INJECT_ADDR_LIMIT)) {
437 			dev_warn(adev->dev, "RAS WARN: input address "
438 					"0x%llx is invalid.",
439 					data.inject.address);
440 			ret = -EINVAL;
441 			break;
442 		}
443 
444 		/* umc ce/ue error injection for a bad page is not allowed */
445 		if ((data.head.block == AMDGPU_RAS_BLOCK__UMC) &&
446 		    amdgpu_ras_check_bad_page(adev, data.inject.address)) {
447 			dev_warn(adev->dev, "RAS WARN: inject: 0x%llx has "
448 				 "already been marked as bad!\n",
449 				 data.inject.address);
450 			break;
451 		}
452 
453 		/* data.inject.address is offset instead of absolute gpu address */
454 		ret = amdgpu_ras_error_inject(adev, &data.inject);
455 		break;
456 	default:
457 		ret = -EINVAL;
458 		break;
459 	}
460 
461 	if (ret)
462 		return ret;
463 
464 	return size;
465 }
466 
467 /**
468  * DOC: AMDGPU RAS debugfs EEPROM table reset interface
469  *
470  * Some boards contain an EEPROM which is used to persistently store a list of
471  * bad pages which experiences ECC errors in vram.  This interface provides
472  * a way to reset the EEPROM, e.g., after testing error injection.
473  *
474  * Usage:
475  *
476  * .. code-block:: bash
477  *
478  *	echo 1 > ../ras/ras_eeprom_reset
479  *
480  * will reset EEPROM table to 0 entries.
481  *
482  */
483 static ssize_t amdgpu_ras_debugfs_eeprom_write(struct file *f,
484 					       const char __user *buf,
485 					       size_t size, loff_t *pos)
486 {
487 	struct amdgpu_device *adev =
488 		(struct amdgpu_device *)file_inode(f)->i_private;
489 	int ret;
490 
491 	ret = amdgpu_ras_eeprom_reset_table(
492 		&(amdgpu_ras_get_context(adev)->eeprom_control));
493 
494 	if (!ret) {
495 		/* Something was written to EEPROM.
496 		 */
497 		amdgpu_ras_get_context(adev)->flags = RAS_DEFAULT_FLAGS;
498 		return size;
499 	} else {
500 		return ret;
501 	}
502 }
503 
504 static const struct file_operations amdgpu_ras_debugfs_ctrl_ops = {
505 	.owner = THIS_MODULE,
506 	.read = NULL,
507 	.write = amdgpu_ras_debugfs_ctrl_write,
508 	.llseek = default_llseek
509 };
510 
511 static const struct file_operations amdgpu_ras_debugfs_eeprom_ops = {
512 	.owner = THIS_MODULE,
513 	.read = NULL,
514 	.write = amdgpu_ras_debugfs_eeprom_write,
515 	.llseek = default_llseek
516 };
517 
518 /**
519  * DOC: AMDGPU RAS sysfs Error Count Interface
520  *
521  * It allows the user to read the error count for each IP block on the gpu through
522  * /sys/class/drm/card[0/1/2...]/device/ras/[gfx/sdma/...]_err_count
523  *
524  * It outputs the multiple lines which report the uncorrected (ue) and corrected
525  * (ce) error counts.
526  *
527  * The format of one line is below,
528  *
529  * [ce|ue]: count
530  *
531  * Example:
532  *
533  * .. code-block:: bash
534  *
535  *	ue: 0
536  *	ce: 1
537  *
538  */
539 static ssize_t amdgpu_ras_sysfs_read(struct device *dev,
540 		struct device_attribute *attr, char *buf)
541 {
542 	struct ras_manager *obj = container_of(attr, struct ras_manager, sysfs_attr);
543 	struct ras_query_if info = {
544 		.head = obj->head,
545 	};
546 
547 	if (!amdgpu_ras_get_error_query_ready(obj->adev))
548 		return sysfs_emit(buf, "Query currently inaccessible\n");
549 
550 	if (amdgpu_ras_query_error_status(obj->adev, &info))
551 		return -EINVAL;
552 
553 	if (obj->adev->asic_type == CHIP_ALDEBARAN) {
554 		if (amdgpu_ras_reset_error_status(obj->adev, info.head.block))
555 			DRM_WARN("Failed to reset error counter and error status");
556 	}
557 
558 	return sysfs_emit(buf, "%s: %lu\n%s: %lu\n", "ue", info.ue_count,
559 			  "ce", info.ce_count);
560 }
561 
562 /* obj begin */
563 
564 #define get_obj(obj) do { (obj)->use++; } while (0)
565 #define alive_obj(obj) ((obj)->use)
566 
567 static inline void put_obj(struct ras_manager *obj)
568 {
569 	if (obj && (--obj->use == 0))
570 		list_del(&obj->node);
571 	if (obj && (obj->use < 0))
572 		DRM_ERROR("RAS ERROR: Unbalance obj(%s) use\n", get_ras_block_str(&obj->head));
573 }
574 
575 /* make one obj and return it. */
576 static struct ras_manager *amdgpu_ras_create_obj(struct amdgpu_device *adev,
577 		struct ras_common_if *head)
578 {
579 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
580 	struct ras_manager *obj;
581 
582 	if (!adev->ras_enabled || !con)
583 		return NULL;
584 
585 	if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
586 		return NULL;
587 
588 	if (head->block == AMDGPU_RAS_BLOCK__MCA) {
589 		if (head->sub_block_index >= AMDGPU_RAS_MCA_BLOCK__LAST)
590 			return NULL;
591 
592 		obj = &con->objs[AMDGPU_RAS_BLOCK__LAST + head->sub_block_index];
593 	} else
594 		obj = &con->objs[head->block];
595 
596 	/* already exist. return obj? */
597 	if (alive_obj(obj))
598 		return NULL;
599 
600 	obj->head = *head;
601 	obj->adev = adev;
602 	list_add(&obj->node, &con->head);
603 	get_obj(obj);
604 
605 	return obj;
606 }
607 
608 /* return an obj equal to head, or the first when head is NULL */
609 struct ras_manager *amdgpu_ras_find_obj(struct amdgpu_device *adev,
610 		struct ras_common_if *head)
611 {
612 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
613 	struct ras_manager *obj;
614 	int i;
615 
616 	if (!adev->ras_enabled || !con)
617 		return NULL;
618 
619 	if (head) {
620 		if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
621 			return NULL;
622 
623 		if (head->block == AMDGPU_RAS_BLOCK__MCA) {
624 			if (head->sub_block_index >= AMDGPU_RAS_MCA_BLOCK__LAST)
625 				return NULL;
626 
627 			obj = &con->objs[AMDGPU_RAS_BLOCK__LAST + head->sub_block_index];
628 		} else
629 			obj = &con->objs[head->block];
630 
631 		if (alive_obj(obj))
632 			return obj;
633 	} else {
634 		for (i = 0; i < AMDGPU_RAS_BLOCK_COUNT + AMDGPU_RAS_MCA_BLOCK_COUNT; i++) {
635 			obj = &con->objs[i];
636 			if (alive_obj(obj))
637 				return obj;
638 		}
639 	}
640 
641 	return NULL;
642 }
643 /* obj end */
644 
645 /* feature ctl begin */
646 static int amdgpu_ras_is_feature_allowed(struct amdgpu_device *adev,
647 					 struct ras_common_if *head)
648 {
649 	return adev->ras_hw_enabled & BIT(head->block);
650 }
651 
652 static int amdgpu_ras_is_feature_enabled(struct amdgpu_device *adev,
653 		struct ras_common_if *head)
654 {
655 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
656 
657 	return con->features & BIT(head->block);
658 }
659 
660 /*
661  * if obj is not created, then create one.
662  * set feature enable flag.
663  */
664 static int __amdgpu_ras_feature_enable(struct amdgpu_device *adev,
665 		struct ras_common_if *head, int enable)
666 {
667 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
668 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
669 
670 	/* If hardware does not support ras, then do not create obj.
671 	 * But if hardware support ras, we can create the obj.
672 	 * Ras framework checks con->hw_supported to see if it need do
673 	 * corresponding initialization.
674 	 * IP checks con->support to see if it need disable ras.
675 	 */
676 	if (!amdgpu_ras_is_feature_allowed(adev, head))
677 		return 0;
678 
679 	if (enable) {
680 		if (!obj) {
681 			obj = amdgpu_ras_create_obj(adev, head);
682 			if (!obj)
683 				return -EINVAL;
684 		} else {
685 			/* In case we create obj somewhere else */
686 			get_obj(obj);
687 		}
688 		con->features |= BIT(head->block);
689 	} else {
690 		if (obj && amdgpu_ras_is_feature_enabled(adev, head)) {
691 			con->features &= ~BIT(head->block);
692 			put_obj(obj);
693 		}
694 	}
695 
696 	return 0;
697 }
698 
699 /* wrapper of psp_ras_enable_features */
700 int amdgpu_ras_feature_enable(struct amdgpu_device *adev,
701 		struct ras_common_if *head, bool enable)
702 {
703 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
704 	union ta_ras_cmd_input *info;
705 	int ret;
706 
707 	if (!con)
708 		return -EINVAL;
709 
710 	info = kzalloc(sizeof(union ta_ras_cmd_input), GFP_KERNEL);
711 	if (!info)
712 		return -ENOMEM;
713 
714 	if (!enable) {
715 		info->disable_features = (struct ta_ras_disable_features_input) {
716 			.block_id =  amdgpu_ras_block_to_ta(head->block),
717 			.error_type = amdgpu_ras_error_to_ta(head->type),
718 		};
719 	} else {
720 		info->enable_features = (struct ta_ras_enable_features_input) {
721 			.block_id =  amdgpu_ras_block_to_ta(head->block),
722 			.error_type = amdgpu_ras_error_to_ta(head->type),
723 		};
724 	}
725 
726 	/* Do not enable if it is not allowed. */
727 	WARN_ON(enable && !amdgpu_ras_is_feature_allowed(adev, head));
728 
729 	if (!amdgpu_ras_intr_triggered()) {
730 		ret = psp_ras_enable_features(&adev->psp, info, enable);
731 		if (ret) {
732 			dev_err(adev->dev, "ras %s %s failed poison:%d ret:%d\n",
733 				enable ? "enable":"disable",
734 				get_ras_block_str(head),
735 				amdgpu_ras_is_poison_mode_supported(adev), ret);
736 			goto out;
737 		}
738 	}
739 
740 	/* setup the obj */
741 	__amdgpu_ras_feature_enable(adev, head, enable);
742 	ret = 0;
743 out:
744 	kfree(info);
745 	return ret;
746 }
747 
748 /* Only used in device probe stage and called only once. */
749 int amdgpu_ras_feature_enable_on_boot(struct amdgpu_device *adev,
750 		struct ras_common_if *head, bool enable)
751 {
752 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
753 	int ret;
754 
755 	if (!con)
756 		return -EINVAL;
757 
758 	if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
759 		if (enable) {
760 			/* There is no harm to issue a ras TA cmd regardless of
761 			 * the currecnt ras state.
762 			 * If current state == target state, it will do nothing
763 			 * But sometimes it requests driver to reset and repost
764 			 * with error code -EAGAIN.
765 			 */
766 			ret = amdgpu_ras_feature_enable(adev, head, 1);
767 			/* With old ras TA, we might fail to enable ras.
768 			 * Log it and just setup the object.
769 			 * TODO need remove this WA in the future.
770 			 */
771 			if (ret == -EINVAL) {
772 				ret = __amdgpu_ras_feature_enable(adev, head, 1);
773 				if (!ret)
774 					dev_info(adev->dev,
775 						"RAS INFO: %s setup object\n",
776 						get_ras_block_str(head));
777 			}
778 		} else {
779 			/* setup the object then issue a ras TA disable cmd.*/
780 			ret = __amdgpu_ras_feature_enable(adev, head, 1);
781 			if (ret)
782 				return ret;
783 
784 			/* gfx block ras dsiable cmd must send to ras-ta */
785 			if (head->block == AMDGPU_RAS_BLOCK__GFX)
786 				con->features |= BIT(head->block);
787 
788 			ret = amdgpu_ras_feature_enable(adev, head, 0);
789 
790 			/* clean gfx block ras features flag */
791 			if (adev->ras_enabled && head->block == AMDGPU_RAS_BLOCK__GFX)
792 				con->features &= ~BIT(head->block);
793 		}
794 	} else
795 		ret = amdgpu_ras_feature_enable(adev, head, enable);
796 
797 	return ret;
798 }
799 
800 static int amdgpu_ras_disable_all_features(struct amdgpu_device *adev,
801 		bool bypass)
802 {
803 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
804 	struct ras_manager *obj, *tmp;
805 
806 	list_for_each_entry_safe(obj, tmp, &con->head, node) {
807 		/* bypass psp.
808 		 * aka just release the obj and corresponding flags
809 		 */
810 		if (bypass) {
811 			if (__amdgpu_ras_feature_enable(adev, &obj->head, 0))
812 				break;
813 		} else {
814 			if (amdgpu_ras_feature_enable(adev, &obj->head, 0))
815 				break;
816 		}
817 	}
818 
819 	return con->features;
820 }
821 
822 static int amdgpu_ras_enable_all_features(struct amdgpu_device *adev,
823 		bool bypass)
824 {
825 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
826 	int i;
827 	const enum amdgpu_ras_error_type default_ras_type = AMDGPU_RAS_ERROR__NONE;
828 
829 	for (i = 0; i < AMDGPU_RAS_BLOCK_COUNT; i++) {
830 		struct ras_common_if head = {
831 			.block = i,
832 			.type = default_ras_type,
833 			.sub_block_index = 0,
834 		};
835 
836 		if (i == AMDGPU_RAS_BLOCK__MCA)
837 			continue;
838 
839 		if (bypass) {
840 			/*
841 			 * bypass psp. vbios enable ras for us.
842 			 * so just create the obj
843 			 */
844 			if (__amdgpu_ras_feature_enable(adev, &head, 1))
845 				break;
846 		} else {
847 			if (amdgpu_ras_feature_enable(adev, &head, 1))
848 				break;
849 		}
850 	}
851 
852 	for (i = 0; i < AMDGPU_RAS_MCA_BLOCK_COUNT; i++) {
853 		struct ras_common_if head = {
854 			.block = AMDGPU_RAS_BLOCK__MCA,
855 			.type = default_ras_type,
856 			.sub_block_index = i,
857 		};
858 
859 		if (bypass) {
860 			/*
861 			 * bypass psp. vbios enable ras for us.
862 			 * so just create the obj
863 			 */
864 			if (__amdgpu_ras_feature_enable(adev, &head, 1))
865 				break;
866 		} else {
867 			if (amdgpu_ras_feature_enable(adev, &head, 1))
868 				break;
869 		}
870 	}
871 
872 	return con->features;
873 }
874 /* feature ctl end */
875 
876 static int amdgpu_ras_block_match_default(struct amdgpu_ras_block_object *block_obj,
877 		enum amdgpu_ras_block block)
878 {
879 	if (!block_obj)
880 		return -EINVAL;
881 
882 	if (block_obj->ras_comm.block == block)
883 		return 0;
884 
885 	return -EINVAL;
886 }
887 
888 static struct amdgpu_ras_block_object *amdgpu_ras_get_ras_block(struct amdgpu_device *adev,
889 					enum amdgpu_ras_block block, uint32_t sub_block_index)
890 {
891 	struct amdgpu_ras_block_list *node, *tmp;
892 	struct amdgpu_ras_block_object *obj;
893 
894 	if (block >= AMDGPU_RAS_BLOCK__LAST)
895 		return NULL;
896 
897 	if (!amdgpu_ras_is_supported(adev, block))
898 		return NULL;
899 
900 	list_for_each_entry_safe(node, tmp, &adev->ras_list, node) {
901 		if (!node->ras_obj) {
902 			dev_warn(adev->dev, "Warning: abnormal ras list node.\n");
903 			continue;
904 		}
905 
906 		obj = node->ras_obj;
907 		if (obj->ras_block_match) {
908 			if (obj->ras_block_match(obj, block, sub_block_index) == 0)
909 				return obj;
910 		} else {
911 			if (amdgpu_ras_block_match_default(obj, block) == 0)
912 				return obj;
913 		}
914 	}
915 
916 	return NULL;
917 }
918 
919 static void amdgpu_ras_get_ecc_info(struct amdgpu_device *adev, struct ras_err_data *err_data)
920 {
921 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
922 	int ret = 0;
923 
924 	/*
925 	 * choosing right query method according to
926 	 * whether smu support query error information
927 	 */
928 	ret = amdgpu_dpm_get_ecc_info(adev, (void *)&(ras->umc_ecc));
929 	if (ret == -EOPNOTSUPP) {
930 		if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
931 			adev->umc.ras->ras_block.hw_ops->query_ras_error_count)
932 			adev->umc.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data);
933 
934 		/* umc query_ras_error_address is also responsible for clearing
935 		 * error status
936 		 */
937 		if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
938 		    adev->umc.ras->ras_block.hw_ops->query_ras_error_address)
939 			adev->umc.ras->ras_block.hw_ops->query_ras_error_address(adev, err_data);
940 	} else if (!ret) {
941 		if (adev->umc.ras &&
942 			adev->umc.ras->ecc_info_query_ras_error_count)
943 			adev->umc.ras->ecc_info_query_ras_error_count(adev, err_data);
944 
945 		if (adev->umc.ras &&
946 			adev->umc.ras->ecc_info_query_ras_error_address)
947 			adev->umc.ras->ecc_info_query_ras_error_address(adev, err_data);
948 	}
949 }
950 
951 /* query/inject/cure begin */
952 int amdgpu_ras_query_error_status(struct amdgpu_device *adev,
953 				  struct ras_query_if *info)
954 {
955 	struct amdgpu_ras_block_object *block_obj = NULL;
956 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
957 	struct ras_err_data err_data = {0, 0, 0, NULL};
958 
959 	if (!obj)
960 		return -EINVAL;
961 
962 	if (info->head.block == AMDGPU_RAS_BLOCK__UMC) {
963 		amdgpu_ras_get_ecc_info(adev, &err_data);
964 	} else {
965 		block_obj = amdgpu_ras_get_ras_block(adev, info->head.block, 0);
966 		if (!block_obj || !block_obj->hw_ops)   {
967 			dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
968 				     get_ras_block_str(&info->head));
969 			return -EINVAL;
970 		}
971 
972 		if (block_obj->hw_ops->query_ras_error_count)
973 			block_obj->hw_ops->query_ras_error_count(adev, &err_data);
974 
975 		if ((info->head.block == AMDGPU_RAS_BLOCK__SDMA) ||
976 		    (info->head.block == AMDGPU_RAS_BLOCK__GFX) ||
977 		    (info->head.block == AMDGPU_RAS_BLOCK__MMHUB)) {
978 				if (block_obj->hw_ops->query_ras_error_status)
979 					block_obj->hw_ops->query_ras_error_status(adev);
980 			}
981 	}
982 
983 	obj->err_data.ue_count += err_data.ue_count;
984 	obj->err_data.ce_count += err_data.ce_count;
985 
986 	info->ue_count = obj->err_data.ue_count;
987 	info->ce_count = obj->err_data.ce_count;
988 
989 	if (err_data.ce_count) {
990 		if (adev->smuio.funcs &&
991 		    adev->smuio.funcs->get_socket_id &&
992 		    adev->smuio.funcs->get_die_id) {
993 			dev_info(adev->dev, "socket: %d, die: %d "
994 					"%ld correctable hardware errors "
995 					"detected in %s block, no user "
996 					"action is needed.\n",
997 					adev->smuio.funcs->get_socket_id(adev),
998 					adev->smuio.funcs->get_die_id(adev),
999 					obj->err_data.ce_count,
1000 					get_ras_block_str(&info->head));
1001 		} else {
1002 			dev_info(adev->dev, "%ld correctable hardware errors "
1003 					"detected in %s block, no user "
1004 					"action is needed.\n",
1005 					obj->err_data.ce_count,
1006 					get_ras_block_str(&info->head));
1007 		}
1008 	}
1009 	if (err_data.ue_count) {
1010 		if (adev->smuio.funcs &&
1011 		    adev->smuio.funcs->get_socket_id &&
1012 		    adev->smuio.funcs->get_die_id) {
1013 			dev_info(adev->dev, "socket: %d, die: %d "
1014 					"%ld uncorrectable hardware errors "
1015 					"detected in %s block\n",
1016 					adev->smuio.funcs->get_socket_id(adev),
1017 					adev->smuio.funcs->get_die_id(adev),
1018 					obj->err_data.ue_count,
1019 					get_ras_block_str(&info->head));
1020 		} else {
1021 			dev_info(adev->dev, "%ld uncorrectable hardware errors "
1022 					"detected in %s block\n",
1023 					obj->err_data.ue_count,
1024 					get_ras_block_str(&info->head));
1025 		}
1026 	}
1027 
1028 	if (!amdgpu_persistent_edc_harvesting_supported(adev))
1029 		amdgpu_ras_reset_error_status(adev, info->head.block);
1030 
1031 	return 0;
1032 }
1033 
1034 int amdgpu_ras_reset_error_status(struct amdgpu_device *adev,
1035 		enum amdgpu_ras_block block)
1036 {
1037 	struct amdgpu_ras_block_object *block_obj = amdgpu_ras_get_ras_block(adev, block, 0);
1038 
1039 	if (!amdgpu_ras_is_supported(adev, block))
1040 		return -EINVAL;
1041 
1042 	if (!block_obj || !block_obj->hw_ops)   {
1043 		dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1044 			     ras_block_str(block));
1045 		return -EINVAL;
1046 	}
1047 
1048 	if (block_obj->hw_ops->reset_ras_error_count)
1049 		block_obj->hw_ops->reset_ras_error_count(adev);
1050 
1051 	if ((block == AMDGPU_RAS_BLOCK__GFX) ||
1052 	    (block == AMDGPU_RAS_BLOCK__MMHUB)) {
1053 		if (block_obj->hw_ops->reset_ras_error_status)
1054 			block_obj->hw_ops->reset_ras_error_status(adev);
1055 	}
1056 
1057 	return 0;
1058 }
1059 
1060 /* wrapper of psp_ras_trigger_error */
1061 int amdgpu_ras_error_inject(struct amdgpu_device *adev,
1062 		struct ras_inject_if *info)
1063 {
1064 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1065 	struct ta_ras_trigger_error_input block_info = {
1066 		.block_id =  amdgpu_ras_block_to_ta(info->head.block),
1067 		.inject_error_type = amdgpu_ras_error_to_ta(info->head.type),
1068 		.sub_block_index = info->head.sub_block_index,
1069 		.address = info->address,
1070 		.value = info->value,
1071 	};
1072 	int ret = -EINVAL;
1073 	struct amdgpu_ras_block_object *block_obj = amdgpu_ras_get_ras_block(adev,
1074 							info->head.block,
1075 							info->head.sub_block_index);
1076 
1077 	if (!obj)
1078 		return -EINVAL;
1079 
1080 	if (!block_obj || !block_obj->hw_ops)	{
1081 		dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1082 			     get_ras_block_str(&info->head));
1083 		return -EINVAL;
1084 	}
1085 
1086 	/* Calculate XGMI relative offset */
1087 	if (adev->gmc.xgmi.num_physical_nodes > 1) {
1088 		block_info.address =
1089 			amdgpu_xgmi_get_relative_phy_addr(adev,
1090 							  block_info.address);
1091 	}
1092 
1093 	if (info->head.block == AMDGPU_RAS_BLOCK__GFX) {
1094 		if (block_obj->hw_ops->ras_error_inject)
1095 			ret = block_obj->hw_ops->ras_error_inject(adev, info);
1096 	} else {
1097 		/* If defined special ras_error_inject(e.g: xgmi), implement special ras_error_inject */
1098 		if (block_obj->hw_ops->ras_error_inject)
1099 			ret = block_obj->hw_ops->ras_error_inject(adev, &block_info);
1100 		else  /*If not defined .ras_error_inject, use default ras_error_inject*/
1101 			ret = psp_ras_trigger_error(&adev->psp, &block_info);
1102 	}
1103 
1104 	if (ret)
1105 		dev_err(adev->dev, "ras inject %s failed %d\n",
1106 			get_ras_block_str(&info->head), ret);
1107 
1108 	return ret;
1109 }
1110 
1111 /**
1112  * amdgpu_ras_query_error_count -- Get error counts of all IPs
1113  * @adev: pointer to AMD GPU device
1114  * @ce_count: pointer to an integer to be set to the count of correctible errors.
1115  * @ue_count: pointer to an integer to be set to the count of uncorrectible
1116  * errors.
1117  *
1118  * If set, @ce_count or @ue_count, count and return the corresponding
1119  * error counts in those integer pointers. Return 0 if the device
1120  * supports RAS. Return -EOPNOTSUPP if the device doesn't support RAS.
1121  */
1122 int amdgpu_ras_query_error_count(struct amdgpu_device *adev,
1123 				 unsigned long *ce_count,
1124 				 unsigned long *ue_count)
1125 {
1126 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1127 	struct ras_manager *obj;
1128 	unsigned long ce, ue;
1129 
1130 	if (!adev->ras_enabled || !con)
1131 		return -EOPNOTSUPP;
1132 
1133 	/* Don't count since no reporting.
1134 	 */
1135 	if (!ce_count && !ue_count)
1136 		return 0;
1137 
1138 	ce = 0;
1139 	ue = 0;
1140 	list_for_each_entry(obj, &con->head, node) {
1141 		struct ras_query_if info = {
1142 			.head = obj->head,
1143 		};
1144 		int res;
1145 
1146 		res = amdgpu_ras_query_error_status(adev, &info);
1147 		if (res)
1148 			return res;
1149 
1150 		ce += info.ce_count;
1151 		ue += info.ue_count;
1152 	}
1153 
1154 	if (ce_count)
1155 		*ce_count = ce;
1156 
1157 	if (ue_count)
1158 		*ue_count = ue;
1159 
1160 	return 0;
1161 }
1162 /* query/inject/cure end */
1163 
1164 
1165 /* sysfs begin */
1166 
1167 static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
1168 		struct ras_badpage **bps, unsigned int *count);
1169 
1170 static char *amdgpu_ras_badpage_flags_str(unsigned int flags)
1171 {
1172 	switch (flags) {
1173 	case AMDGPU_RAS_RETIRE_PAGE_RESERVED:
1174 		return "R";
1175 	case AMDGPU_RAS_RETIRE_PAGE_PENDING:
1176 		return "P";
1177 	case AMDGPU_RAS_RETIRE_PAGE_FAULT:
1178 	default:
1179 		return "F";
1180 	}
1181 }
1182 
1183 /**
1184  * DOC: AMDGPU RAS sysfs gpu_vram_bad_pages Interface
1185  *
1186  * It allows user to read the bad pages of vram on the gpu through
1187  * /sys/class/drm/card[0/1/2...]/device/ras/gpu_vram_bad_pages
1188  *
1189  * It outputs multiple lines, and each line stands for one gpu page.
1190  *
1191  * The format of one line is below,
1192  * gpu pfn : gpu page size : flags
1193  *
1194  * gpu pfn and gpu page size are printed in hex format.
1195  * flags can be one of below character,
1196  *
1197  * R: reserved, this gpu page is reserved and not able to use.
1198  *
1199  * P: pending for reserve, this gpu page is marked as bad, will be reserved
1200  * in next window of page_reserve.
1201  *
1202  * F: unable to reserve. this gpu page can't be reserved due to some reasons.
1203  *
1204  * Examples:
1205  *
1206  * .. code-block:: bash
1207  *
1208  *	0x00000001 : 0x00001000 : R
1209  *	0x00000002 : 0x00001000 : P
1210  *
1211  */
1212 
1213 static ssize_t amdgpu_ras_sysfs_badpages_read(struct file *f,
1214 		struct kobject *kobj, struct bin_attribute *attr,
1215 		char *buf, loff_t ppos, size_t count)
1216 {
1217 	struct amdgpu_ras *con =
1218 		container_of(attr, struct amdgpu_ras, badpages_attr);
1219 	struct amdgpu_device *adev = con->adev;
1220 	const unsigned int element_size =
1221 		sizeof("0xabcdabcd : 0x12345678 : R\n") - 1;
1222 	unsigned int start = div64_ul(ppos + element_size - 1, element_size);
1223 	unsigned int end = div64_ul(ppos + count - 1, element_size);
1224 	ssize_t s = 0;
1225 	struct ras_badpage *bps = NULL;
1226 	unsigned int bps_count = 0;
1227 
1228 	memset(buf, 0, count);
1229 
1230 	if (amdgpu_ras_badpages_read(adev, &bps, &bps_count))
1231 		return 0;
1232 
1233 	for (; start < end && start < bps_count; start++)
1234 		s += scnprintf(&buf[s], element_size + 1,
1235 				"0x%08x : 0x%08x : %1s\n",
1236 				bps[start].bp,
1237 				bps[start].size,
1238 				amdgpu_ras_badpage_flags_str(bps[start].flags));
1239 
1240 	kfree(bps);
1241 
1242 	return s;
1243 }
1244 
1245 static ssize_t amdgpu_ras_sysfs_features_read(struct device *dev,
1246 		struct device_attribute *attr, char *buf)
1247 {
1248 	struct amdgpu_ras *con =
1249 		container_of(attr, struct amdgpu_ras, features_attr);
1250 
1251 	return scnprintf(buf, PAGE_SIZE, "feature mask: 0x%x\n", con->features);
1252 }
1253 
1254 static void amdgpu_ras_sysfs_remove_bad_page_node(struct amdgpu_device *adev)
1255 {
1256 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1257 
1258 	sysfs_remove_file_from_group(&adev->dev->kobj,
1259 				&con->badpages_attr.attr,
1260 				RAS_FS_NAME);
1261 }
1262 
1263 static int amdgpu_ras_sysfs_remove_feature_node(struct amdgpu_device *adev)
1264 {
1265 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1266 	struct attribute *attrs[] = {
1267 		&con->features_attr.attr,
1268 		NULL
1269 	};
1270 	struct attribute_group group = {
1271 		.name = RAS_FS_NAME,
1272 		.attrs = attrs,
1273 	};
1274 
1275 	sysfs_remove_group(&adev->dev->kobj, &group);
1276 
1277 	return 0;
1278 }
1279 
1280 int amdgpu_ras_sysfs_create(struct amdgpu_device *adev,
1281 		struct ras_common_if *head)
1282 {
1283 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1284 
1285 	if (!obj || obj->attr_inuse)
1286 		return -EINVAL;
1287 
1288 	get_obj(obj);
1289 
1290 	snprintf(obj->fs_data.sysfs_name, sizeof(obj->fs_data.sysfs_name),
1291 		"%s_err_count", head->name);
1292 
1293 	obj->sysfs_attr = (struct device_attribute){
1294 		.attr = {
1295 			.name = obj->fs_data.sysfs_name,
1296 			.mode = S_IRUGO,
1297 		},
1298 			.show = amdgpu_ras_sysfs_read,
1299 	};
1300 	sysfs_attr_init(&obj->sysfs_attr.attr);
1301 
1302 	if (sysfs_add_file_to_group(&adev->dev->kobj,
1303 				&obj->sysfs_attr.attr,
1304 				RAS_FS_NAME)) {
1305 		put_obj(obj);
1306 		return -EINVAL;
1307 	}
1308 
1309 	obj->attr_inuse = 1;
1310 
1311 	return 0;
1312 }
1313 
1314 int amdgpu_ras_sysfs_remove(struct amdgpu_device *adev,
1315 		struct ras_common_if *head)
1316 {
1317 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1318 
1319 	if (!obj || !obj->attr_inuse)
1320 		return -EINVAL;
1321 
1322 	sysfs_remove_file_from_group(&adev->dev->kobj,
1323 				&obj->sysfs_attr.attr,
1324 				RAS_FS_NAME);
1325 	obj->attr_inuse = 0;
1326 	put_obj(obj);
1327 
1328 	return 0;
1329 }
1330 
1331 static int amdgpu_ras_sysfs_remove_all(struct amdgpu_device *adev)
1332 {
1333 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1334 	struct ras_manager *obj, *tmp;
1335 
1336 	list_for_each_entry_safe(obj, tmp, &con->head, node) {
1337 		amdgpu_ras_sysfs_remove(adev, &obj->head);
1338 	}
1339 
1340 	if (amdgpu_bad_page_threshold != 0)
1341 		amdgpu_ras_sysfs_remove_bad_page_node(adev);
1342 
1343 	amdgpu_ras_sysfs_remove_feature_node(adev);
1344 
1345 	return 0;
1346 }
1347 /* sysfs end */
1348 
1349 /**
1350  * DOC: AMDGPU RAS Reboot Behavior for Unrecoverable Errors
1351  *
1352  * Normally when there is an uncorrectable error, the driver will reset
1353  * the GPU to recover.  However, in the event of an unrecoverable error,
1354  * the driver provides an interface to reboot the system automatically
1355  * in that event.
1356  *
1357  * The following file in debugfs provides that interface:
1358  * /sys/kernel/debug/dri/[0/1/2...]/ras/auto_reboot
1359  *
1360  * Usage:
1361  *
1362  * .. code-block:: bash
1363  *
1364  *	echo true > .../ras/auto_reboot
1365  *
1366  */
1367 /* debugfs begin */
1368 static struct dentry *amdgpu_ras_debugfs_create_ctrl_node(struct amdgpu_device *adev)
1369 {
1370 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1371 	struct drm_minor  *minor = adev_to_drm(adev)->primary;
1372 	struct dentry     *dir;
1373 
1374 	dir = debugfs_create_dir(RAS_FS_NAME, minor->debugfs_root);
1375 	debugfs_create_file("ras_ctrl", S_IWUGO | S_IRUGO, dir, adev,
1376 			    &amdgpu_ras_debugfs_ctrl_ops);
1377 	debugfs_create_file("ras_eeprom_reset", S_IWUGO | S_IRUGO, dir, adev,
1378 			    &amdgpu_ras_debugfs_eeprom_ops);
1379 	debugfs_create_u32("bad_page_cnt_threshold", 0444, dir,
1380 			   &con->bad_page_cnt_threshold);
1381 	debugfs_create_x32("ras_hw_enabled", 0444, dir, &adev->ras_hw_enabled);
1382 	debugfs_create_x32("ras_enabled", 0444, dir, &adev->ras_enabled);
1383 	debugfs_create_file("ras_eeprom_size", S_IRUGO, dir, adev,
1384 			    &amdgpu_ras_debugfs_eeprom_size_ops);
1385 	con->de_ras_eeprom_table = debugfs_create_file("ras_eeprom_table",
1386 						       S_IRUGO, dir, adev,
1387 						       &amdgpu_ras_debugfs_eeprom_table_ops);
1388 	amdgpu_ras_debugfs_set_ret_size(&con->eeprom_control);
1389 
1390 	/*
1391 	 * After one uncorrectable error happens, usually GPU recovery will
1392 	 * be scheduled. But due to the known problem in GPU recovery failing
1393 	 * to bring GPU back, below interface provides one direct way to
1394 	 * user to reboot system automatically in such case within
1395 	 * ERREVENT_ATHUB_INTERRUPT generated. Normal GPU recovery routine
1396 	 * will never be called.
1397 	 */
1398 	debugfs_create_bool("auto_reboot", S_IWUGO | S_IRUGO, dir, &con->reboot);
1399 
1400 	/*
1401 	 * User could set this not to clean up hardware's error count register
1402 	 * of RAS IPs during ras recovery.
1403 	 */
1404 	debugfs_create_bool("disable_ras_err_cnt_harvest", 0644, dir,
1405 			    &con->disable_ras_err_cnt_harvest);
1406 	return dir;
1407 }
1408 
1409 static void amdgpu_ras_debugfs_create(struct amdgpu_device *adev,
1410 				      struct ras_fs_if *head,
1411 				      struct dentry *dir)
1412 {
1413 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head->head);
1414 
1415 	if (!obj || !dir)
1416 		return;
1417 
1418 	get_obj(obj);
1419 
1420 	memcpy(obj->fs_data.debugfs_name,
1421 			head->debugfs_name,
1422 			sizeof(obj->fs_data.debugfs_name));
1423 
1424 	debugfs_create_file(obj->fs_data.debugfs_name, S_IWUGO | S_IRUGO, dir,
1425 			    obj, &amdgpu_ras_debugfs_ops);
1426 }
1427 
1428 void amdgpu_ras_debugfs_create_all(struct amdgpu_device *adev)
1429 {
1430 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1431 	struct dentry *dir;
1432 	struct ras_manager *obj;
1433 	struct ras_fs_if fs_info;
1434 
1435 	/*
1436 	 * it won't be called in resume path, no need to check
1437 	 * suspend and gpu reset status
1438 	 */
1439 	if (!IS_ENABLED(CONFIG_DEBUG_FS) || !con)
1440 		return;
1441 
1442 	dir = amdgpu_ras_debugfs_create_ctrl_node(adev);
1443 
1444 	list_for_each_entry(obj, &con->head, node) {
1445 		if (amdgpu_ras_is_supported(adev, obj->head.block) &&
1446 			(obj->attr_inuse == 1)) {
1447 			sprintf(fs_info.debugfs_name, "%s_err_inject",
1448 					get_ras_block_str(&obj->head));
1449 			fs_info.head = obj->head;
1450 			amdgpu_ras_debugfs_create(adev, &fs_info, dir);
1451 		}
1452 	}
1453 }
1454 
1455 /* debugfs end */
1456 
1457 /* ras fs */
1458 static BIN_ATTR(gpu_vram_bad_pages, S_IRUGO,
1459 		amdgpu_ras_sysfs_badpages_read, NULL, 0);
1460 static DEVICE_ATTR(features, S_IRUGO,
1461 		amdgpu_ras_sysfs_features_read, NULL);
1462 static int amdgpu_ras_fs_init(struct amdgpu_device *adev)
1463 {
1464 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1465 	struct attribute_group group = {
1466 		.name = RAS_FS_NAME,
1467 	};
1468 	struct attribute *attrs[] = {
1469 		&con->features_attr.attr,
1470 		NULL
1471 	};
1472 	struct bin_attribute *bin_attrs[] = {
1473 		NULL,
1474 		NULL,
1475 	};
1476 	int r;
1477 
1478 	/* add features entry */
1479 	con->features_attr = dev_attr_features;
1480 	group.attrs = attrs;
1481 	sysfs_attr_init(attrs[0]);
1482 
1483 	if (amdgpu_bad_page_threshold != 0) {
1484 		/* add bad_page_features entry */
1485 		bin_attr_gpu_vram_bad_pages.private = NULL;
1486 		con->badpages_attr = bin_attr_gpu_vram_bad_pages;
1487 		bin_attrs[0] = &con->badpages_attr;
1488 		group.bin_attrs = bin_attrs;
1489 		sysfs_bin_attr_init(bin_attrs[0]);
1490 	}
1491 
1492 	r = sysfs_create_group(&adev->dev->kobj, &group);
1493 	if (r)
1494 		dev_err(adev->dev, "Failed to create RAS sysfs group!");
1495 
1496 	return 0;
1497 }
1498 
1499 static int amdgpu_ras_fs_fini(struct amdgpu_device *adev)
1500 {
1501 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1502 	struct ras_manager *con_obj, *ip_obj, *tmp;
1503 
1504 	if (IS_ENABLED(CONFIG_DEBUG_FS)) {
1505 		list_for_each_entry_safe(con_obj, tmp, &con->head, node) {
1506 			ip_obj = amdgpu_ras_find_obj(adev, &con_obj->head);
1507 			if (ip_obj)
1508 				put_obj(ip_obj);
1509 		}
1510 	}
1511 
1512 	amdgpu_ras_sysfs_remove_all(adev);
1513 	return 0;
1514 }
1515 /* ras fs end */
1516 
1517 /* ih begin */
1518 
1519 /* For the hardware that cannot enable bif ring for both ras_controller_irq
1520  * and ras_err_evnet_athub_irq ih cookies, the driver has to poll status
1521  * register to check whether the interrupt is triggered or not, and properly
1522  * ack the interrupt if it is there
1523  */
1524 void amdgpu_ras_interrupt_fatal_error_handler(struct amdgpu_device *adev)
1525 {
1526 	if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__PCIE_BIF))
1527 		return;
1528 
1529 	if (adev->nbio.ras &&
1530 	    adev->nbio.ras->handle_ras_controller_intr_no_bifring)
1531 		adev->nbio.ras->handle_ras_controller_intr_no_bifring(adev);
1532 
1533 	if (adev->nbio.ras &&
1534 	    adev->nbio.ras->handle_ras_err_event_athub_intr_no_bifring)
1535 		adev->nbio.ras->handle_ras_err_event_athub_intr_no_bifring(adev);
1536 }
1537 
1538 static void amdgpu_ras_interrupt_poison_consumption_handler(struct ras_manager *obj,
1539 				struct amdgpu_iv_entry *entry)
1540 {
1541 	bool poison_stat = false;
1542 	struct amdgpu_device *adev = obj->adev;
1543 	struct ras_err_data err_data = {0, 0, 0, NULL};
1544 	struct amdgpu_ras_block_object *block_obj =
1545 		amdgpu_ras_get_ras_block(adev, obj->head.block, 0);
1546 
1547 	if (!block_obj || !block_obj->hw_ops)
1548 		return;
1549 
1550 	/* both query_poison_status and handle_poison_consumption are optional,
1551 	 * but at least one of them should be implemented if we need poison
1552 	 * consumption handler
1553 	 */
1554 	if (block_obj->hw_ops->query_poison_status) {
1555 		poison_stat = block_obj->hw_ops->query_poison_status(adev);
1556 		if (!poison_stat) {
1557 			/* Not poison consumption interrupt, no need to handle it */
1558 			dev_info(adev->dev, "No RAS poison status in %s poison IH.\n",
1559 					block_obj->ras_comm.name);
1560 
1561 			return;
1562 		}
1563 	}
1564 
1565 	if (!adev->gmc.xgmi.connected_to_cpu)
1566 		amdgpu_umc_poison_handler(adev, &err_data, false);
1567 
1568 	if (block_obj->hw_ops->handle_poison_consumption)
1569 		poison_stat = block_obj->hw_ops->handle_poison_consumption(adev);
1570 
1571 	/* gpu reset is fallback for failed and default cases */
1572 	if (poison_stat) {
1573 		dev_info(adev->dev, "GPU reset for %s RAS poison consumption is issued!\n",
1574 				block_obj->ras_comm.name);
1575 		amdgpu_ras_reset_gpu(adev);
1576 	}
1577 }
1578 
1579 static void amdgpu_ras_interrupt_poison_creation_handler(struct ras_manager *obj,
1580 				struct amdgpu_iv_entry *entry)
1581 {
1582 	dev_info(obj->adev->dev,
1583 		"Poison is created, no user action is needed.\n");
1584 }
1585 
1586 static void amdgpu_ras_interrupt_umc_handler(struct ras_manager *obj,
1587 				struct amdgpu_iv_entry *entry)
1588 {
1589 	struct ras_ih_data *data = &obj->ih_data;
1590 	struct ras_err_data err_data = {0, 0, 0, NULL};
1591 	int ret;
1592 
1593 	if (!data->cb)
1594 		return;
1595 
1596 	/* Let IP handle its data, maybe we need get the output
1597 	 * from the callback to update the error type/count, etc
1598 	 */
1599 	ret = data->cb(obj->adev, &err_data, entry);
1600 	/* ue will trigger an interrupt, and in that case
1601 	 * we need do a reset to recovery the whole system.
1602 	 * But leave IP do that recovery, here we just dispatch
1603 	 * the error.
1604 	 */
1605 	if (ret == AMDGPU_RAS_SUCCESS) {
1606 		/* these counts could be left as 0 if
1607 		 * some blocks do not count error number
1608 		 */
1609 		obj->err_data.ue_count += err_data.ue_count;
1610 		obj->err_data.ce_count += err_data.ce_count;
1611 	}
1612 }
1613 
1614 static void amdgpu_ras_interrupt_handler(struct ras_manager *obj)
1615 {
1616 	struct ras_ih_data *data = &obj->ih_data;
1617 	struct amdgpu_iv_entry entry;
1618 
1619 	while (data->rptr != data->wptr) {
1620 		rmb();
1621 		memcpy(&entry, &data->ring[data->rptr],
1622 				data->element_size);
1623 
1624 		wmb();
1625 		data->rptr = (data->aligned_element_size +
1626 				data->rptr) % data->ring_size;
1627 
1628 		if (amdgpu_ras_is_poison_mode_supported(obj->adev)) {
1629 			if (obj->head.block == AMDGPU_RAS_BLOCK__UMC)
1630 				amdgpu_ras_interrupt_poison_creation_handler(obj, &entry);
1631 			else
1632 				amdgpu_ras_interrupt_poison_consumption_handler(obj, &entry);
1633 		} else {
1634 			if (obj->head.block == AMDGPU_RAS_BLOCK__UMC)
1635 				amdgpu_ras_interrupt_umc_handler(obj, &entry);
1636 			else
1637 				dev_warn(obj->adev->dev,
1638 					"No RAS interrupt handler for non-UMC block with poison disabled.\n");
1639 		}
1640 	}
1641 }
1642 
1643 static void amdgpu_ras_interrupt_process_handler(struct work_struct *work)
1644 {
1645 	struct ras_ih_data *data =
1646 		container_of(work, struct ras_ih_data, ih_work);
1647 	struct ras_manager *obj =
1648 		container_of(data, struct ras_manager, ih_data);
1649 
1650 	amdgpu_ras_interrupt_handler(obj);
1651 }
1652 
1653 int amdgpu_ras_interrupt_dispatch(struct amdgpu_device *adev,
1654 		struct ras_dispatch_if *info)
1655 {
1656 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1657 	struct ras_ih_data *data = &obj->ih_data;
1658 
1659 	if (!obj)
1660 		return -EINVAL;
1661 
1662 	if (data->inuse == 0)
1663 		return 0;
1664 
1665 	/* Might be overflow... */
1666 	memcpy(&data->ring[data->wptr], info->entry,
1667 			data->element_size);
1668 
1669 	wmb();
1670 	data->wptr = (data->aligned_element_size +
1671 			data->wptr) % data->ring_size;
1672 
1673 	schedule_work(&data->ih_work);
1674 
1675 	return 0;
1676 }
1677 
1678 int amdgpu_ras_interrupt_remove_handler(struct amdgpu_device *adev,
1679 		struct ras_common_if *head)
1680 {
1681 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1682 	struct ras_ih_data *data;
1683 
1684 	if (!obj)
1685 		return -EINVAL;
1686 
1687 	data = &obj->ih_data;
1688 	if (data->inuse == 0)
1689 		return 0;
1690 
1691 	cancel_work_sync(&data->ih_work);
1692 
1693 	kfree(data->ring);
1694 	memset(data, 0, sizeof(*data));
1695 	put_obj(obj);
1696 
1697 	return 0;
1698 }
1699 
1700 int amdgpu_ras_interrupt_add_handler(struct amdgpu_device *adev,
1701 		struct ras_common_if *head)
1702 {
1703 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1704 	struct ras_ih_data *data;
1705 	struct amdgpu_ras_block_object *ras_obj;
1706 
1707 	if (!obj) {
1708 		/* in case we registe the IH before enable ras feature */
1709 		obj = amdgpu_ras_create_obj(adev, head);
1710 		if (!obj)
1711 			return -EINVAL;
1712 	} else
1713 		get_obj(obj);
1714 
1715 	ras_obj = container_of(head, struct amdgpu_ras_block_object, ras_comm);
1716 
1717 	data = &obj->ih_data;
1718 	/* add the callback.etc */
1719 	*data = (struct ras_ih_data) {
1720 		.inuse = 0,
1721 		.cb = ras_obj->ras_cb,
1722 		.element_size = sizeof(struct amdgpu_iv_entry),
1723 		.rptr = 0,
1724 		.wptr = 0,
1725 	};
1726 
1727 	INIT_WORK(&data->ih_work, amdgpu_ras_interrupt_process_handler);
1728 
1729 	data->aligned_element_size = ALIGN(data->element_size, 8);
1730 	/* the ring can store 64 iv entries. */
1731 	data->ring_size = 64 * data->aligned_element_size;
1732 	data->ring = kmalloc(data->ring_size, GFP_KERNEL);
1733 	if (!data->ring) {
1734 		put_obj(obj);
1735 		return -ENOMEM;
1736 	}
1737 
1738 	/* IH is ready */
1739 	data->inuse = 1;
1740 
1741 	return 0;
1742 }
1743 
1744 static int amdgpu_ras_interrupt_remove_all(struct amdgpu_device *adev)
1745 {
1746 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1747 	struct ras_manager *obj, *tmp;
1748 
1749 	list_for_each_entry_safe(obj, tmp, &con->head, node) {
1750 		amdgpu_ras_interrupt_remove_handler(adev, &obj->head);
1751 	}
1752 
1753 	return 0;
1754 }
1755 /* ih end */
1756 
1757 /* traversal all IPs except NBIO to query error counter */
1758 static void amdgpu_ras_log_on_err_counter(struct amdgpu_device *adev)
1759 {
1760 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1761 	struct ras_manager *obj;
1762 
1763 	if (!adev->ras_enabled || !con)
1764 		return;
1765 
1766 	list_for_each_entry(obj, &con->head, node) {
1767 		struct ras_query_if info = {
1768 			.head = obj->head,
1769 		};
1770 
1771 		/*
1772 		 * PCIE_BIF IP has one different isr by ras controller
1773 		 * interrupt, the specific ras counter query will be
1774 		 * done in that isr. So skip such block from common
1775 		 * sync flood interrupt isr calling.
1776 		 */
1777 		if (info.head.block == AMDGPU_RAS_BLOCK__PCIE_BIF)
1778 			continue;
1779 
1780 		/*
1781 		 * this is a workaround for aldebaran, skip send msg to
1782 		 * smu to get ecc_info table due to smu handle get ecc
1783 		 * info table failed temporarily.
1784 		 * should be removed until smu fix handle ecc_info table.
1785 		 */
1786 		if ((info.head.block == AMDGPU_RAS_BLOCK__UMC) &&
1787 			(adev->ip_versions[MP1_HWIP][0] == IP_VERSION(13, 0, 2)))
1788 			continue;
1789 
1790 		amdgpu_ras_query_error_status(adev, &info);
1791 	}
1792 }
1793 
1794 /* Parse RdRspStatus and WrRspStatus */
1795 static void amdgpu_ras_error_status_query(struct amdgpu_device *adev,
1796 					  struct ras_query_if *info)
1797 {
1798 	struct amdgpu_ras_block_object *block_obj;
1799 	/*
1800 	 * Only two block need to query read/write
1801 	 * RspStatus at current state
1802 	 */
1803 	if ((info->head.block != AMDGPU_RAS_BLOCK__GFX) &&
1804 		(info->head.block != AMDGPU_RAS_BLOCK__MMHUB))
1805 		return;
1806 
1807 	block_obj = amdgpu_ras_get_ras_block(adev,
1808 					info->head.block,
1809 					info->head.sub_block_index);
1810 
1811 	if (!block_obj || !block_obj->hw_ops) {
1812 		dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1813 			     get_ras_block_str(&info->head));
1814 		return;
1815 	}
1816 
1817 	if (block_obj->hw_ops->query_ras_error_status)
1818 		block_obj->hw_ops->query_ras_error_status(adev);
1819 
1820 }
1821 
1822 static void amdgpu_ras_query_err_status(struct amdgpu_device *adev)
1823 {
1824 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1825 	struct ras_manager *obj;
1826 
1827 	if (!adev->ras_enabled || !con)
1828 		return;
1829 
1830 	list_for_each_entry(obj, &con->head, node) {
1831 		struct ras_query_if info = {
1832 			.head = obj->head,
1833 		};
1834 
1835 		amdgpu_ras_error_status_query(adev, &info);
1836 	}
1837 }
1838 
1839 /* recovery begin */
1840 
1841 /* return 0 on success.
1842  * caller need free bps.
1843  */
1844 static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
1845 		struct ras_badpage **bps, unsigned int *count)
1846 {
1847 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1848 	struct ras_err_handler_data *data;
1849 	int i = 0;
1850 	int ret = 0, status;
1851 
1852 	if (!con || !con->eh_data || !bps || !count)
1853 		return -EINVAL;
1854 
1855 	mutex_lock(&con->recovery_lock);
1856 	data = con->eh_data;
1857 	if (!data || data->count == 0) {
1858 		*bps = NULL;
1859 		ret = -EINVAL;
1860 		goto out;
1861 	}
1862 
1863 	*bps = kmalloc(sizeof(struct ras_badpage) * data->count, GFP_KERNEL);
1864 	if (!*bps) {
1865 		ret = -ENOMEM;
1866 		goto out;
1867 	}
1868 
1869 	for (; i < data->count; i++) {
1870 		(*bps)[i] = (struct ras_badpage){
1871 			.bp = data->bps[i].retired_page,
1872 			.size = AMDGPU_GPU_PAGE_SIZE,
1873 			.flags = AMDGPU_RAS_RETIRE_PAGE_RESERVED,
1874 		};
1875 		status = amdgpu_vram_mgr_query_page_status(&adev->mman.vram_mgr,
1876 				data->bps[i].retired_page);
1877 		if (status == -EBUSY)
1878 			(*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_PENDING;
1879 		else if (status == -ENOENT)
1880 			(*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_FAULT;
1881 	}
1882 
1883 	*count = data->count;
1884 out:
1885 	mutex_unlock(&con->recovery_lock);
1886 	return ret;
1887 }
1888 
1889 static void amdgpu_ras_do_recovery(struct work_struct *work)
1890 {
1891 	struct amdgpu_ras *ras =
1892 		container_of(work, struct amdgpu_ras, recovery_work);
1893 	struct amdgpu_device *remote_adev = NULL;
1894 	struct amdgpu_device *adev = ras->adev;
1895 	struct list_head device_list, *device_list_handle =  NULL;
1896 
1897 	if (!ras->disable_ras_err_cnt_harvest) {
1898 		struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
1899 
1900 		/* Build list of devices to query RAS related errors */
1901 		if  (hive && adev->gmc.xgmi.num_physical_nodes > 1) {
1902 			device_list_handle = &hive->device_list;
1903 		} else {
1904 			INIT_LIST_HEAD(&device_list);
1905 			list_add_tail(&adev->gmc.xgmi.head, &device_list);
1906 			device_list_handle = &device_list;
1907 		}
1908 
1909 		list_for_each_entry(remote_adev,
1910 				device_list_handle, gmc.xgmi.head) {
1911 			amdgpu_ras_query_err_status(remote_adev);
1912 			amdgpu_ras_log_on_err_counter(remote_adev);
1913 		}
1914 
1915 		amdgpu_put_xgmi_hive(hive);
1916 	}
1917 
1918 	if (amdgpu_device_should_recover_gpu(ras->adev))
1919 		amdgpu_device_gpu_recover(ras->adev, NULL);
1920 	atomic_set(&ras->in_recovery, 0);
1921 }
1922 
1923 /* alloc/realloc bps array */
1924 static int amdgpu_ras_realloc_eh_data_space(struct amdgpu_device *adev,
1925 		struct ras_err_handler_data *data, int pages)
1926 {
1927 	unsigned int old_space = data->count + data->space_left;
1928 	unsigned int new_space = old_space + pages;
1929 	unsigned int align_space = ALIGN(new_space, 512);
1930 	void *bps = kmalloc(align_space * sizeof(*data->bps), GFP_KERNEL);
1931 
1932 	if (!bps) {
1933 		return -ENOMEM;
1934 	}
1935 
1936 	if (data->bps) {
1937 		memcpy(bps, data->bps,
1938 				data->count * sizeof(*data->bps));
1939 		kfree(data->bps);
1940 	}
1941 
1942 	data->bps = bps;
1943 	data->space_left += align_space - old_space;
1944 	return 0;
1945 }
1946 
1947 /* it deal with vram only. */
1948 int amdgpu_ras_add_bad_pages(struct amdgpu_device *adev,
1949 		struct eeprom_table_record *bps, int pages)
1950 {
1951 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1952 	struct ras_err_handler_data *data;
1953 	int ret = 0;
1954 	uint32_t i;
1955 
1956 	if (!con || !con->eh_data || !bps || pages <= 0)
1957 		return 0;
1958 
1959 	mutex_lock(&con->recovery_lock);
1960 	data = con->eh_data;
1961 	if (!data)
1962 		goto out;
1963 
1964 	for (i = 0; i < pages; i++) {
1965 		if (amdgpu_ras_check_bad_page_unlock(con,
1966 			bps[i].retired_page << AMDGPU_GPU_PAGE_SHIFT))
1967 			continue;
1968 
1969 		if (!data->space_left &&
1970 			amdgpu_ras_realloc_eh_data_space(adev, data, 256)) {
1971 			ret = -ENOMEM;
1972 			goto out;
1973 		}
1974 
1975 		amdgpu_vram_mgr_reserve_range(&adev->mman.vram_mgr,
1976 			bps[i].retired_page << AMDGPU_GPU_PAGE_SHIFT,
1977 			AMDGPU_GPU_PAGE_SIZE);
1978 
1979 		memcpy(&data->bps[data->count], &bps[i], sizeof(*data->bps));
1980 		data->count++;
1981 		data->space_left--;
1982 	}
1983 out:
1984 	mutex_unlock(&con->recovery_lock);
1985 
1986 	return ret;
1987 }
1988 
1989 /*
1990  * write error record array to eeprom, the function should be
1991  * protected by recovery_lock
1992  */
1993 int amdgpu_ras_save_bad_pages(struct amdgpu_device *adev)
1994 {
1995 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1996 	struct ras_err_handler_data *data;
1997 	struct amdgpu_ras_eeprom_control *control;
1998 	int save_count;
1999 
2000 	if (!con || !con->eh_data)
2001 		return 0;
2002 
2003 	mutex_lock(&con->recovery_lock);
2004 	control = &con->eeprom_control;
2005 	data = con->eh_data;
2006 	save_count = data->count - control->ras_num_recs;
2007 	mutex_unlock(&con->recovery_lock);
2008 	/* only new entries are saved */
2009 	if (save_count > 0) {
2010 		if (amdgpu_ras_eeprom_append(control,
2011 					     &data->bps[control->ras_num_recs],
2012 					     save_count)) {
2013 			dev_err(adev->dev, "Failed to save EEPROM table data!");
2014 			return -EIO;
2015 		}
2016 
2017 		dev_info(adev->dev, "Saved %d pages to EEPROM table.\n", save_count);
2018 	}
2019 
2020 	return 0;
2021 }
2022 
2023 /*
2024  * read error record array in eeprom and reserve enough space for
2025  * storing new bad pages
2026  */
2027 static int amdgpu_ras_load_bad_pages(struct amdgpu_device *adev)
2028 {
2029 	struct amdgpu_ras_eeprom_control *control =
2030 		&adev->psp.ras_context.ras->eeprom_control;
2031 	struct eeprom_table_record *bps;
2032 	int ret;
2033 
2034 	/* no bad page record, skip eeprom access */
2035 	if (control->ras_num_recs == 0 || amdgpu_bad_page_threshold == 0)
2036 		return 0;
2037 
2038 	bps = kcalloc(control->ras_num_recs, sizeof(*bps), GFP_KERNEL);
2039 	if (!bps)
2040 		return -ENOMEM;
2041 
2042 	ret = amdgpu_ras_eeprom_read(control, bps, control->ras_num_recs);
2043 	if (ret)
2044 		dev_err(adev->dev, "Failed to load EEPROM table records!");
2045 	else
2046 		ret = amdgpu_ras_add_bad_pages(adev, bps, control->ras_num_recs);
2047 
2048 	kfree(bps);
2049 	return ret;
2050 }
2051 
2052 static bool amdgpu_ras_check_bad_page_unlock(struct amdgpu_ras *con,
2053 				uint64_t addr)
2054 {
2055 	struct ras_err_handler_data *data = con->eh_data;
2056 	int i;
2057 
2058 	addr >>= AMDGPU_GPU_PAGE_SHIFT;
2059 	for (i = 0; i < data->count; i++)
2060 		if (addr == data->bps[i].retired_page)
2061 			return true;
2062 
2063 	return false;
2064 }
2065 
2066 /*
2067  * check if an address belongs to bad page
2068  *
2069  * Note: this check is only for umc block
2070  */
2071 static bool amdgpu_ras_check_bad_page(struct amdgpu_device *adev,
2072 				uint64_t addr)
2073 {
2074 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2075 	bool ret = false;
2076 
2077 	if (!con || !con->eh_data)
2078 		return ret;
2079 
2080 	mutex_lock(&con->recovery_lock);
2081 	ret = amdgpu_ras_check_bad_page_unlock(con, addr);
2082 	mutex_unlock(&con->recovery_lock);
2083 	return ret;
2084 }
2085 
2086 static void amdgpu_ras_validate_threshold(struct amdgpu_device *adev,
2087 					  uint32_t max_count)
2088 {
2089 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2090 
2091 	/*
2092 	 * Justification of value bad_page_cnt_threshold in ras structure
2093 	 *
2094 	 * Generally, -1 <= amdgpu_bad_page_threshold <= max record length
2095 	 * in eeprom, and introduce two scenarios accordingly.
2096 	 *
2097 	 * Bad page retirement enablement:
2098 	 *    - If amdgpu_bad_page_threshold = -1,
2099 	 *      bad_page_cnt_threshold = typical value by formula.
2100 	 *
2101 	 *    - When the value from user is 0 < amdgpu_bad_page_threshold <
2102 	 *      max record length in eeprom, use it directly.
2103 	 *
2104 	 * Bad page retirement disablement:
2105 	 *    - If amdgpu_bad_page_threshold = 0, bad page retirement
2106 	 *      functionality is disabled, and bad_page_cnt_threshold will
2107 	 *      take no effect.
2108 	 */
2109 
2110 	if (amdgpu_bad_page_threshold < 0) {
2111 		u64 val = adev->gmc.mc_vram_size;
2112 
2113 		do_div(val, RAS_BAD_PAGE_COVER);
2114 		con->bad_page_cnt_threshold = min(lower_32_bits(val),
2115 						  max_count);
2116 	} else {
2117 		con->bad_page_cnt_threshold = min_t(int, max_count,
2118 						    amdgpu_bad_page_threshold);
2119 	}
2120 }
2121 
2122 int amdgpu_ras_recovery_init(struct amdgpu_device *adev)
2123 {
2124 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2125 	struct ras_err_handler_data **data;
2126 	u32  max_eeprom_records_count = 0;
2127 	bool exc_err_limit = false;
2128 	int ret;
2129 
2130 	if (!con)
2131 		return 0;
2132 
2133 	/* Allow access to RAS EEPROM via debugfs, when the ASIC
2134 	 * supports RAS and debugfs is enabled, but when
2135 	 * adev->ras_enabled is unset, i.e. when "ras_enable"
2136 	 * module parameter is set to 0.
2137 	 */
2138 	con->adev = adev;
2139 
2140 	if (!adev->ras_enabled)
2141 		return 0;
2142 
2143 	data = &con->eh_data;
2144 	*data = kmalloc(sizeof(**data), GFP_KERNEL | __GFP_ZERO);
2145 	if (!*data) {
2146 		ret = -ENOMEM;
2147 		goto out;
2148 	}
2149 
2150 	mutex_init(&con->recovery_lock);
2151 	INIT_WORK(&con->recovery_work, amdgpu_ras_do_recovery);
2152 	atomic_set(&con->in_recovery, 0);
2153 	con->eeprom_control.bad_channel_bitmap = 0;
2154 
2155 	max_eeprom_records_count = amdgpu_ras_eeprom_max_record_count();
2156 	amdgpu_ras_validate_threshold(adev, max_eeprom_records_count);
2157 
2158 	/* Todo: During test the SMU might fail to read the eeprom through I2C
2159 	 * when the GPU is pending on XGMI reset during probe time
2160 	 * (Mostly after second bus reset), skip it now
2161 	 */
2162 	if (adev->gmc.xgmi.pending_reset)
2163 		return 0;
2164 	ret = amdgpu_ras_eeprom_init(&con->eeprom_control, &exc_err_limit);
2165 	/*
2166 	 * This calling fails when exc_err_limit is true or
2167 	 * ret != 0.
2168 	 */
2169 	if (exc_err_limit || ret)
2170 		goto free;
2171 
2172 	if (con->eeprom_control.ras_num_recs) {
2173 		ret = amdgpu_ras_load_bad_pages(adev);
2174 		if (ret)
2175 			goto free;
2176 
2177 		amdgpu_dpm_send_hbm_bad_pages_num(adev, con->eeprom_control.ras_num_recs);
2178 
2179 		if (con->update_channel_flag == true) {
2180 			amdgpu_dpm_send_hbm_bad_channel_flag(adev, con->eeprom_control.bad_channel_bitmap);
2181 			con->update_channel_flag = false;
2182 		}
2183 	}
2184 
2185 #ifdef CONFIG_X86_MCE_AMD
2186 	if ((adev->asic_type == CHIP_ALDEBARAN) &&
2187 	    (adev->gmc.xgmi.connected_to_cpu))
2188 		amdgpu_register_bad_pages_mca_notifier(adev);
2189 #endif
2190 	return 0;
2191 
2192 free:
2193 	kfree((*data)->bps);
2194 	kfree(*data);
2195 	con->eh_data = NULL;
2196 out:
2197 	dev_warn(adev->dev, "Failed to initialize ras recovery! (%d)\n", ret);
2198 
2199 	/*
2200 	 * Except error threshold exceeding case, other failure cases in this
2201 	 * function would not fail amdgpu driver init.
2202 	 */
2203 	if (!exc_err_limit)
2204 		ret = 0;
2205 	else
2206 		ret = -EINVAL;
2207 
2208 	return ret;
2209 }
2210 
2211 static int amdgpu_ras_recovery_fini(struct amdgpu_device *adev)
2212 {
2213 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2214 	struct ras_err_handler_data *data = con->eh_data;
2215 
2216 	/* recovery_init failed to init it, fini is useless */
2217 	if (!data)
2218 		return 0;
2219 
2220 	cancel_work_sync(&con->recovery_work);
2221 
2222 	mutex_lock(&con->recovery_lock);
2223 	con->eh_data = NULL;
2224 	kfree(data->bps);
2225 	kfree(data);
2226 	mutex_unlock(&con->recovery_lock);
2227 
2228 	return 0;
2229 }
2230 /* recovery end */
2231 
2232 static bool amdgpu_ras_asic_supported(struct amdgpu_device *adev)
2233 {
2234 	return adev->asic_type == CHIP_VEGA10 ||
2235 		adev->asic_type == CHIP_VEGA20 ||
2236 		adev->asic_type == CHIP_ARCTURUS ||
2237 		adev->asic_type == CHIP_ALDEBARAN ||
2238 		adev->asic_type == CHIP_SIENNA_CICHLID;
2239 }
2240 
2241 /*
2242  * this is workaround for vega20 workstation sku,
2243  * force enable gfx ras, ignore vbios gfx ras flag
2244  * due to GC EDC can not write
2245  */
2246 static void amdgpu_ras_get_quirks(struct amdgpu_device *adev)
2247 {
2248 	struct atom_context *ctx = adev->mode_info.atom_context;
2249 
2250 	if (!ctx)
2251 		return;
2252 
2253 	if (strnstr(ctx->vbios_version, "D16406",
2254 		    sizeof(ctx->vbios_version)) ||
2255 		strnstr(ctx->vbios_version, "D36002",
2256 			sizeof(ctx->vbios_version)))
2257 		adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX);
2258 }
2259 
2260 /*
2261  * check hardware's ras ability which will be saved in hw_supported.
2262  * if hardware does not support ras, we can skip some ras initializtion and
2263  * forbid some ras operations from IP.
2264  * if software itself, say boot parameter, limit the ras ability. We still
2265  * need allow IP do some limited operations, like disable. In such case,
2266  * we have to initialize ras as normal. but need check if operation is
2267  * allowed or not in each function.
2268  */
2269 static void amdgpu_ras_check_supported(struct amdgpu_device *adev)
2270 {
2271 	adev->ras_hw_enabled = adev->ras_enabled = 0;
2272 
2273 	if (amdgpu_sriov_vf(adev) || !adev->is_atom_fw ||
2274 	    !amdgpu_ras_asic_supported(adev))
2275 		return;
2276 
2277 	if (!adev->gmc.xgmi.connected_to_cpu) {
2278 		if (amdgpu_atomfirmware_mem_ecc_supported(adev)) {
2279 			dev_info(adev->dev, "MEM ECC is active.\n");
2280 			adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__UMC |
2281 						   1 << AMDGPU_RAS_BLOCK__DF);
2282 		} else {
2283 			dev_info(adev->dev, "MEM ECC is not presented.\n");
2284 		}
2285 
2286 		if (amdgpu_atomfirmware_sram_ecc_supported(adev)) {
2287 			dev_info(adev->dev, "SRAM ECC is active.\n");
2288 			adev->ras_hw_enabled |= ~(1 << AMDGPU_RAS_BLOCK__UMC |
2289 						    1 << AMDGPU_RAS_BLOCK__DF);
2290 
2291 			if (adev->ip_versions[VCN_HWIP][0] == IP_VERSION(2, 6, 0))
2292 				adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__VCN |
2293 						1 << AMDGPU_RAS_BLOCK__JPEG);
2294 			else
2295 				adev->ras_hw_enabled &= ~(1 << AMDGPU_RAS_BLOCK__VCN |
2296 						1 << AMDGPU_RAS_BLOCK__JPEG);
2297 		} else {
2298 			dev_info(adev->dev, "SRAM ECC is not presented.\n");
2299 		}
2300 	} else {
2301 		/* driver only manages a few IP blocks RAS feature
2302 		 * when GPU is connected cpu through XGMI */
2303 		adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX |
2304 					   1 << AMDGPU_RAS_BLOCK__SDMA |
2305 					   1 << AMDGPU_RAS_BLOCK__MMHUB);
2306 	}
2307 
2308 	amdgpu_ras_get_quirks(adev);
2309 
2310 	/* hw_supported needs to be aligned with RAS block mask. */
2311 	adev->ras_hw_enabled &= AMDGPU_RAS_BLOCK_MASK;
2312 
2313 	adev->ras_enabled = amdgpu_ras_enable == 0 ? 0 :
2314 		adev->ras_hw_enabled & amdgpu_ras_mask;
2315 }
2316 
2317 static void amdgpu_ras_counte_dw(struct work_struct *work)
2318 {
2319 	struct amdgpu_ras *con = container_of(work, struct amdgpu_ras,
2320 					      ras_counte_delay_work.work);
2321 	struct amdgpu_device *adev = con->adev;
2322 	struct drm_device *dev = adev_to_drm(adev);
2323 	unsigned long ce_count, ue_count;
2324 	int res;
2325 
2326 	res = pm_runtime_get_sync(dev->dev);
2327 	if (res < 0)
2328 		goto Out;
2329 
2330 	/* Cache new values.
2331 	 */
2332 	if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count) == 0) {
2333 		atomic_set(&con->ras_ce_count, ce_count);
2334 		atomic_set(&con->ras_ue_count, ue_count);
2335 	}
2336 
2337 	pm_runtime_mark_last_busy(dev->dev);
2338 Out:
2339 	pm_runtime_put_autosuspend(dev->dev);
2340 }
2341 
2342 int amdgpu_ras_init(struct amdgpu_device *adev)
2343 {
2344 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2345 	int r;
2346 	bool df_poison, umc_poison;
2347 
2348 	if (con)
2349 		return 0;
2350 
2351 	con = kmalloc(sizeof(struct amdgpu_ras) +
2352 			sizeof(struct ras_manager) * AMDGPU_RAS_BLOCK_COUNT +
2353 			sizeof(struct ras_manager) * AMDGPU_RAS_MCA_BLOCK_COUNT,
2354 			GFP_KERNEL|__GFP_ZERO);
2355 	if (!con)
2356 		return -ENOMEM;
2357 
2358 	con->adev = adev;
2359 	INIT_DELAYED_WORK(&con->ras_counte_delay_work, amdgpu_ras_counte_dw);
2360 	atomic_set(&con->ras_ce_count, 0);
2361 	atomic_set(&con->ras_ue_count, 0);
2362 
2363 	con->objs = (struct ras_manager *)(con + 1);
2364 
2365 	amdgpu_ras_set_context(adev, con);
2366 
2367 	amdgpu_ras_check_supported(adev);
2368 
2369 	if (!adev->ras_enabled || adev->asic_type == CHIP_VEGA10) {
2370 		/* set gfx block ras context feature for VEGA20 Gaming
2371 		 * send ras disable cmd to ras ta during ras late init.
2372 		 */
2373 		if (!adev->ras_enabled && adev->asic_type == CHIP_VEGA20) {
2374 			con->features |= BIT(AMDGPU_RAS_BLOCK__GFX);
2375 
2376 			return 0;
2377 		}
2378 
2379 		r = 0;
2380 		goto release_con;
2381 	}
2382 
2383 	con->update_channel_flag = false;
2384 	con->features = 0;
2385 	INIT_LIST_HEAD(&con->head);
2386 	/* Might need get this flag from vbios. */
2387 	con->flags = RAS_DEFAULT_FLAGS;
2388 
2389 	/* initialize nbio ras function ahead of any other
2390 	 * ras functions so hardware fatal error interrupt
2391 	 * can be enabled as early as possible */
2392 	switch (adev->asic_type) {
2393 	case CHIP_VEGA20:
2394 	case CHIP_ARCTURUS:
2395 	case CHIP_ALDEBARAN:
2396 		if (!adev->gmc.xgmi.connected_to_cpu) {
2397 			adev->nbio.ras = &nbio_v7_4_ras;
2398 			amdgpu_ras_register_ras_block(adev, &adev->nbio.ras->ras_block);
2399 			adev->nbio.ras_if = &adev->nbio.ras->ras_block.ras_comm;
2400 		}
2401 		break;
2402 	default:
2403 		/* nbio ras is not available */
2404 		break;
2405 	}
2406 
2407 	if (adev->nbio.ras &&
2408 	    adev->nbio.ras->init_ras_controller_interrupt) {
2409 		r = adev->nbio.ras->init_ras_controller_interrupt(adev);
2410 		if (r)
2411 			goto release_con;
2412 	}
2413 
2414 	if (adev->nbio.ras &&
2415 	    adev->nbio.ras->init_ras_err_event_athub_interrupt) {
2416 		r = adev->nbio.ras->init_ras_err_event_athub_interrupt(adev);
2417 		if (r)
2418 			goto release_con;
2419 	}
2420 
2421 	/* Init poison supported flag, the default value is false */
2422 	if (adev->gmc.xgmi.connected_to_cpu) {
2423 		/* enabled by default when GPU is connected to CPU */
2424 		con->poison_supported = true;
2425 	}
2426 	else if (adev->df.funcs &&
2427 	    adev->df.funcs->query_ras_poison_mode &&
2428 	    adev->umc.ras &&
2429 	    adev->umc.ras->query_ras_poison_mode) {
2430 		df_poison =
2431 			adev->df.funcs->query_ras_poison_mode(adev);
2432 		umc_poison =
2433 			adev->umc.ras->query_ras_poison_mode(adev);
2434 		/* Only poison is set in both DF and UMC, we can support it */
2435 		if (df_poison && umc_poison)
2436 			con->poison_supported = true;
2437 		else if (df_poison != umc_poison)
2438 			dev_warn(adev->dev, "Poison setting is inconsistent in DF/UMC(%d:%d)!\n",
2439 					df_poison, umc_poison);
2440 	}
2441 
2442 	if (amdgpu_ras_fs_init(adev)) {
2443 		r = -EINVAL;
2444 		goto release_con;
2445 	}
2446 
2447 	dev_info(adev->dev, "RAS INFO: ras initialized successfully, "
2448 		 "hardware ability[%x] ras_mask[%x]\n",
2449 		 adev->ras_hw_enabled, adev->ras_enabled);
2450 
2451 	return 0;
2452 release_con:
2453 	amdgpu_ras_set_context(adev, NULL);
2454 	kfree(con);
2455 
2456 	return r;
2457 }
2458 
2459 int amdgpu_persistent_edc_harvesting_supported(struct amdgpu_device *adev)
2460 {
2461 	if (adev->gmc.xgmi.connected_to_cpu)
2462 		return 1;
2463 	return 0;
2464 }
2465 
2466 static int amdgpu_persistent_edc_harvesting(struct amdgpu_device *adev,
2467 					struct ras_common_if *ras_block)
2468 {
2469 	struct ras_query_if info = {
2470 		.head = *ras_block,
2471 	};
2472 
2473 	if (!amdgpu_persistent_edc_harvesting_supported(adev))
2474 		return 0;
2475 
2476 	if (amdgpu_ras_query_error_status(adev, &info) != 0)
2477 		DRM_WARN("RAS init harvest failure");
2478 
2479 	if (amdgpu_ras_reset_error_status(adev, ras_block->block) != 0)
2480 		DRM_WARN("RAS init harvest reset failure");
2481 
2482 	return 0;
2483 }
2484 
2485 bool amdgpu_ras_is_poison_mode_supported(struct amdgpu_device *adev)
2486 {
2487        struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2488 
2489        if (!con)
2490                return false;
2491 
2492        return con->poison_supported;
2493 }
2494 
2495 /* helper function to handle common stuff in ip late init phase */
2496 int amdgpu_ras_block_late_init(struct amdgpu_device *adev,
2497 			 struct ras_common_if *ras_block)
2498 {
2499 	struct amdgpu_ras_block_object *ras_obj = NULL;
2500 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2501 	unsigned long ue_count, ce_count;
2502 	int r;
2503 
2504 	/* disable RAS feature per IP block if it is not supported */
2505 	if (!amdgpu_ras_is_supported(adev, ras_block->block)) {
2506 		amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
2507 		return 0;
2508 	}
2509 
2510 	r = amdgpu_ras_feature_enable_on_boot(adev, ras_block, 1);
2511 	if (r) {
2512 		if (adev->in_suspend || amdgpu_in_reset(adev)) {
2513 			/* in resume phase, if fail to enable ras,
2514 			 * clean up all ras fs nodes, and disable ras */
2515 			goto cleanup;
2516 		} else
2517 			return r;
2518 	}
2519 
2520 	/* check for errors on warm reset edc persisant supported ASIC */
2521 	amdgpu_persistent_edc_harvesting(adev, ras_block);
2522 
2523 	/* in resume phase, no need to create ras fs node */
2524 	if (adev->in_suspend || amdgpu_in_reset(adev))
2525 		return 0;
2526 
2527 	ras_obj = container_of(ras_block, struct amdgpu_ras_block_object, ras_comm);
2528 	if (ras_obj->ras_cb || (ras_obj->hw_ops &&
2529 	    (ras_obj->hw_ops->query_poison_status ||
2530 	    ras_obj->hw_ops->handle_poison_consumption))) {
2531 		r = amdgpu_ras_interrupt_add_handler(adev, ras_block);
2532 		if (r)
2533 			goto cleanup;
2534 	}
2535 
2536 	r = amdgpu_ras_sysfs_create(adev, ras_block);
2537 	if (r)
2538 		goto interrupt;
2539 
2540 	/* Those are the cached values at init.
2541 	 */
2542 	if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count) == 0) {
2543 		atomic_set(&con->ras_ce_count, ce_count);
2544 		atomic_set(&con->ras_ue_count, ue_count);
2545 	}
2546 
2547 	return 0;
2548 
2549 interrupt:
2550 	if (ras_obj->ras_cb)
2551 		amdgpu_ras_interrupt_remove_handler(adev, ras_block);
2552 cleanup:
2553 	amdgpu_ras_feature_enable(adev, ras_block, 0);
2554 	return r;
2555 }
2556 
2557 static int amdgpu_ras_block_late_init_default(struct amdgpu_device *adev,
2558 			 struct ras_common_if *ras_block)
2559 {
2560 	return amdgpu_ras_block_late_init(adev, ras_block);
2561 }
2562 
2563 /* helper function to remove ras fs node and interrupt handler */
2564 void amdgpu_ras_block_late_fini(struct amdgpu_device *adev,
2565 			  struct ras_common_if *ras_block)
2566 {
2567 	struct amdgpu_ras_block_object *ras_obj;
2568 	if (!ras_block)
2569 		return;
2570 
2571 	amdgpu_ras_sysfs_remove(adev, ras_block);
2572 
2573 	ras_obj = container_of(ras_block, struct amdgpu_ras_block_object, ras_comm);
2574 	if (ras_obj->ras_cb)
2575 		amdgpu_ras_interrupt_remove_handler(adev, ras_block);
2576 }
2577 
2578 static void amdgpu_ras_block_late_fini_default(struct amdgpu_device *adev,
2579 			  struct ras_common_if *ras_block)
2580 {
2581 	return amdgpu_ras_block_late_fini(adev, ras_block);
2582 }
2583 
2584 /* do some init work after IP late init as dependence.
2585  * and it runs in resume/gpu reset/booting up cases.
2586  */
2587 void amdgpu_ras_resume(struct amdgpu_device *adev)
2588 {
2589 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2590 	struct ras_manager *obj, *tmp;
2591 
2592 	if (!adev->ras_enabled || !con) {
2593 		/* clean ras context for VEGA20 Gaming after send ras disable cmd */
2594 		amdgpu_release_ras_context(adev);
2595 
2596 		return;
2597 	}
2598 
2599 	if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
2600 		/* Set up all other IPs which are not implemented. There is a
2601 		 * tricky thing that IP's actual ras error type should be
2602 		 * MULTI_UNCORRECTABLE, but as driver does not handle it, so
2603 		 * ERROR_NONE make sense anyway.
2604 		 */
2605 		amdgpu_ras_enable_all_features(adev, 1);
2606 
2607 		/* We enable ras on all hw_supported block, but as boot
2608 		 * parameter might disable some of them and one or more IP has
2609 		 * not implemented yet. So we disable them on behalf.
2610 		 */
2611 		list_for_each_entry_safe(obj, tmp, &con->head, node) {
2612 			if (!amdgpu_ras_is_supported(adev, obj->head.block)) {
2613 				amdgpu_ras_feature_enable(adev, &obj->head, 0);
2614 				/* there should be no any reference. */
2615 				WARN_ON(alive_obj(obj));
2616 			}
2617 		}
2618 	}
2619 }
2620 
2621 void amdgpu_ras_suspend(struct amdgpu_device *adev)
2622 {
2623 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2624 
2625 	if (!adev->ras_enabled || !con)
2626 		return;
2627 
2628 	amdgpu_ras_disable_all_features(adev, 0);
2629 	/* Make sure all ras objects are disabled. */
2630 	if (con->features)
2631 		amdgpu_ras_disable_all_features(adev, 1);
2632 }
2633 
2634 int amdgpu_ras_late_init(struct amdgpu_device *adev)
2635 {
2636 	struct amdgpu_ras_block_list *node, *tmp;
2637 	struct amdgpu_ras_block_object *obj;
2638 	int r;
2639 
2640 	list_for_each_entry_safe(node, tmp, &adev->ras_list, node) {
2641 		if (!node->ras_obj) {
2642 			dev_warn(adev->dev, "Warning: abnormal ras list node.\n");
2643 			continue;
2644 		}
2645 
2646 		obj = node->ras_obj;
2647 		if (obj->ras_late_init) {
2648 			r = obj->ras_late_init(adev, &obj->ras_comm);
2649 			if (r) {
2650 				dev_err(adev->dev, "%s failed to execute ras_late_init! ret:%d\n",
2651 					obj->ras_comm.name, r);
2652 				return r;
2653 			}
2654 		} else
2655 			amdgpu_ras_block_late_init_default(adev, &obj->ras_comm);
2656 	}
2657 
2658 	return 0;
2659 }
2660 
2661 /* do some fini work before IP fini as dependence */
2662 int amdgpu_ras_pre_fini(struct amdgpu_device *adev)
2663 {
2664 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2665 
2666 	if (!adev->ras_enabled || !con)
2667 		return 0;
2668 
2669 
2670 	/* Need disable ras on all IPs here before ip [hw/sw]fini */
2671 	amdgpu_ras_disable_all_features(adev, 0);
2672 	amdgpu_ras_recovery_fini(adev);
2673 	return 0;
2674 }
2675 
2676 int amdgpu_ras_fini(struct amdgpu_device *adev)
2677 {
2678 	struct amdgpu_ras_block_list *ras_node, *tmp;
2679 	struct amdgpu_ras_block_object *obj = NULL;
2680 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2681 
2682 	if (!adev->ras_enabled || !con)
2683 		return 0;
2684 
2685 	list_for_each_entry_safe(ras_node, tmp, &adev->ras_list, node) {
2686 		if (ras_node->ras_obj) {
2687 			obj = ras_node->ras_obj;
2688 			if (amdgpu_ras_is_supported(adev, obj->ras_comm.block) &&
2689 			    obj->ras_fini)
2690 				obj->ras_fini(adev, &obj->ras_comm);
2691 			else
2692 				amdgpu_ras_block_late_fini_default(adev, &obj->ras_comm);
2693 		}
2694 
2695 		/* Clear ras blocks from ras_list and free ras block list node */
2696 		list_del(&ras_node->node);
2697 		kfree(ras_node);
2698 	}
2699 
2700 	amdgpu_ras_fs_fini(adev);
2701 	amdgpu_ras_interrupt_remove_all(adev);
2702 
2703 	WARN(con->features, "Feature mask is not cleared");
2704 
2705 	if (con->features)
2706 		amdgpu_ras_disable_all_features(adev, 1);
2707 
2708 	cancel_delayed_work_sync(&con->ras_counte_delay_work);
2709 
2710 	amdgpu_ras_set_context(adev, NULL);
2711 	kfree(con);
2712 
2713 	return 0;
2714 }
2715 
2716 void amdgpu_ras_global_ras_isr(struct amdgpu_device *adev)
2717 {
2718 	amdgpu_ras_check_supported(adev);
2719 	if (!adev->ras_hw_enabled)
2720 		return;
2721 
2722 	if (atomic_cmpxchg(&amdgpu_ras_in_intr, 0, 1) == 0) {
2723 		dev_info(adev->dev, "uncorrectable hardware error"
2724 			"(ERREVENT_ATHUB_INTERRUPT) detected!\n");
2725 
2726 		amdgpu_ras_reset_gpu(adev);
2727 	}
2728 }
2729 
2730 bool amdgpu_ras_need_emergency_restart(struct amdgpu_device *adev)
2731 {
2732 	if (adev->asic_type == CHIP_VEGA20 &&
2733 	    adev->pm.fw_version <= 0x283400) {
2734 		return !(amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) &&
2735 				amdgpu_ras_intr_triggered();
2736 	}
2737 
2738 	return false;
2739 }
2740 
2741 void amdgpu_release_ras_context(struct amdgpu_device *adev)
2742 {
2743 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2744 
2745 	if (!con)
2746 		return;
2747 
2748 	if (!adev->ras_enabled && con->features & BIT(AMDGPU_RAS_BLOCK__GFX)) {
2749 		con->features &= ~BIT(AMDGPU_RAS_BLOCK__GFX);
2750 		amdgpu_ras_set_context(adev, NULL);
2751 		kfree(con);
2752 	}
2753 }
2754 
2755 #ifdef CONFIG_X86_MCE_AMD
2756 static struct amdgpu_device *find_adev(uint32_t node_id)
2757 {
2758 	int i;
2759 	struct amdgpu_device *adev = NULL;
2760 
2761 	for (i = 0; i < mce_adev_list.num_gpu; i++) {
2762 		adev = mce_adev_list.devs[i];
2763 
2764 		if (adev && adev->gmc.xgmi.connected_to_cpu &&
2765 		    adev->gmc.xgmi.physical_node_id == node_id)
2766 			break;
2767 		adev = NULL;
2768 	}
2769 
2770 	return adev;
2771 }
2772 
2773 #define GET_MCA_IPID_GPUID(m)	(((m) >> 44) & 0xF)
2774 #define GET_UMC_INST(m)		(((m) >> 21) & 0x7)
2775 #define GET_CHAN_INDEX(m)	((((m) >> 12) & 0x3) | (((m) >> 18) & 0x4))
2776 #define GPU_ID_OFFSET		8
2777 
2778 static int amdgpu_bad_page_notifier(struct notifier_block *nb,
2779 				    unsigned long val, void *data)
2780 {
2781 	struct mce *m = (struct mce *)data;
2782 	struct amdgpu_device *adev = NULL;
2783 	uint32_t gpu_id = 0;
2784 	uint32_t umc_inst = 0;
2785 	uint32_t ch_inst, channel_index = 0;
2786 	struct ras_err_data err_data = {0, 0, 0, NULL};
2787 	struct eeprom_table_record err_rec;
2788 	uint64_t retired_page;
2789 
2790 	/*
2791 	 * If the error was generated in UMC_V2, which belongs to GPU UMCs,
2792 	 * and error occurred in DramECC (Extended error code = 0) then only
2793 	 * process the error, else bail out.
2794 	 */
2795 	if (!m || !((smca_get_bank_type(m->extcpu, m->bank) == SMCA_UMC_V2) &&
2796 		    (XEC(m->status, 0x3f) == 0x0)))
2797 		return NOTIFY_DONE;
2798 
2799 	/*
2800 	 * If it is correctable error, return.
2801 	 */
2802 	if (mce_is_correctable(m))
2803 		return NOTIFY_OK;
2804 
2805 	/*
2806 	 * GPU Id is offset by GPU_ID_OFFSET in MCA_IPID_UMC register.
2807 	 */
2808 	gpu_id = GET_MCA_IPID_GPUID(m->ipid) - GPU_ID_OFFSET;
2809 
2810 	adev = find_adev(gpu_id);
2811 	if (!adev) {
2812 		DRM_WARN("%s: Unable to find adev for gpu_id: %d\n", __func__,
2813 								gpu_id);
2814 		return NOTIFY_DONE;
2815 	}
2816 
2817 	/*
2818 	 * If it is uncorrectable error, then find out UMC instance and
2819 	 * channel index.
2820 	 */
2821 	umc_inst = GET_UMC_INST(m->ipid);
2822 	ch_inst = GET_CHAN_INDEX(m->ipid);
2823 
2824 	dev_info(adev->dev, "Uncorrectable error detected in UMC inst: %d, chan_idx: %d",
2825 			     umc_inst, ch_inst);
2826 
2827 	/*
2828 	 * Translate UMC channel address to Physical address
2829 	 */
2830 	channel_index =
2831 		adev->umc.channel_idx_tbl[umc_inst * adev->umc.channel_inst_num
2832 					  + ch_inst];
2833 
2834 	retired_page = ADDR_OF_8KB_BLOCK(m->addr) |
2835 			ADDR_OF_256B_BLOCK(channel_index) |
2836 			OFFSET_IN_256B_BLOCK(m->addr);
2837 
2838 	memset(&err_rec, 0x0, sizeof(struct eeprom_table_record));
2839 	err_data.err_addr = &err_rec;
2840 	amdgpu_umc_fill_error_record(&err_data, m->addr,
2841 			retired_page, channel_index, umc_inst);
2842 
2843 	if (amdgpu_bad_page_threshold != 0) {
2844 		amdgpu_ras_add_bad_pages(adev, err_data.err_addr,
2845 						err_data.err_addr_cnt);
2846 		amdgpu_ras_save_bad_pages(adev);
2847 	}
2848 
2849 	return NOTIFY_OK;
2850 }
2851 
2852 static struct notifier_block amdgpu_bad_page_nb = {
2853 	.notifier_call  = amdgpu_bad_page_notifier,
2854 	.priority       = MCE_PRIO_UC,
2855 };
2856 
2857 static void amdgpu_register_bad_pages_mca_notifier(struct amdgpu_device *adev)
2858 {
2859 	/*
2860 	 * Add the adev to the mce_adev_list.
2861 	 * During mode2 reset, amdgpu device is temporarily
2862 	 * removed from the mgpu_info list which can cause
2863 	 * page retirement to fail.
2864 	 * Use this list instead of mgpu_info to find the amdgpu
2865 	 * device on which the UMC error was reported.
2866 	 */
2867 	mce_adev_list.devs[mce_adev_list.num_gpu++] = adev;
2868 
2869 	/*
2870 	 * Register the x86 notifier only once
2871 	 * with MCE subsystem.
2872 	 */
2873 	if (notifier_registered == false) {
2874 		mce_register_decode_chain(&amdgpu_bad_page_nb);
2875 		notifier_registered = true;
2876 	}
2877 }
2878 #endif
2879 
2880 struct amdgpu_ras *amdgpu_ras_get_context(struct amdgpu_device *adev)
2881 {
2882 	if (!adev)
2883 		return NULL;
2884 
2885 	return adev->psp.ras_context.ras;
2886 }
2887 
2888 int amdgpu_ras_set_context(struct amdgpu_device *adev, struct amdgpu_ras *ras_con)
2889 {
2890 	if (!adev)
2891 		return -EINVAL;
2892 
2893 	adev->psp.ras_context.ras = ras_con;
2894 	return 0;
2895 }
2896 
2897 /* check if ras is supported on block, say, sdma, gfx */
2898 int amdgpu_ras_is_supported(struct amdgpu_device *adev,
2899 		unsigned int block)
2900 {
2901 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
2902 
2903 	if (block >= AMDGPU_RAS_BLOCK_COUNT)
2904 		return 0;
2905 	return ras && (adev->ras_enabled & (1 << block));
2906 }
2907 
2908 int amdgpu_ras_reset_gpu(struct amdgpu_device *adev)
2909 {
2910 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
2911 
2912 	if (atomic_cmpxchg(&ras->in_recovery, 0, 1) == 0)
2913 		schedule_work(&ras->recovery_work);
2914 	return 0;
2915 }
2916 
2917 
2918 /* Register each ip ras block into amdgpu ras */
2919 int amdgpu_ras_register_ras_block(struct amdgpu_device *adev,
2920 		struct amdgpu_ras_block_object *ras_block_obj)
2921 {
2922 	struct amdgpu_ras_block_list *ras_node;
2923 	if (!adev || !ras_block_obj)
2924 		return -EINVAL;
2925 
2926 	if (!amdgpu_ras_asic_supported(adev))
2927 		return 0;
2928 
2929 	ras_node = kzalloc(sizeof(*ras_node), GFP_KERNEL);
2930 	if (!ras_node)
2931 		return -ENOMEM;
2932 
2933 	INIT_LIST_HEAD(&ras_node->node);
2934 	ras_node->ras_obj = ras_block_obj;
2935 	list_add_tail(&ras_node->node, &adev->ras_list);
2936 
2937 	return 0;
2938 }
2939