xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c (revision da1d9caf95def6f0320819cf941c9fd1069ba9e1)
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 	/* Hardware counter will be reset automatically after the query on Vega20 and Arcturus */
201 	if (obj->adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 2) &&
202 	    obj->adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 4)) {
203 		if (amdgpu_ras_reset_error_status(obj->adev, info.head.block))
204 			dev_warn(obj->adev->dev, "Failed to reset error counter and error status");
205 	}
206 
207 	s = snprintf(val, sizeof(val), "%s: %lu\n%s: %lu\n",
208 			"ue", info.ue_count,
209 			"ce", info.ce_count);
210 	if (*pos >= s)
211 		return 0;
212 
213 	s -= *pos;
214 	s = min_t(u64, s, size);
215 
216 
217 	if (copy_to_user(buf, &val[*pos], s))
218 		return -EINVAL;
219 
220 	*pos += s;
221 
222 	return s;
223 }
224 
225 static const struct file_operations amdgpu_ras_debugfs_ops = {
226 	.owner = THIS_MODULE,
227 	.read = amdgpu_ras_debugfs_read,
228 	.write = NULL,
229 	.llseek = default_llseek
230 };
231 
232 static int amdgpu_ras_find_block_id_by_name(const char *name, int *block_id)
233 {
234 	int i;
235 
236 	for (i = 0; i < ARRAY_SIZE(ras_block_string); i++) {
237 		*block_id = i;
238 		if (strcmp(name, ras_block_string[i]) == 0)
239 			return 0;
240 	}
241 	return -EINVAL;
242 }
243 
244 static int amdgpu_ras_debugfs_ctrl_parse_data(struct file *f,
245 		const char __user *buf, size_t size,
246 		loff_t *pos, struct ras_debug_if *data)
247 {
248 	ssize_t s = min_t(u64, 64, size);
249 	char str[65];
250 	char block_name[33];
251 	char err[9] = "ue";
252 	int op = -1;
253 	int block_id;
254 	uint32_t sub_block;
255 	u64 address, value;
256 
257 	if (*pos)
258 		return -EINVAL;
259 	*pos = size;
260 
261 	memset(str, 0, sizeof(str));
262 	memset(data, 0, sizeof(*data));
263 
264 	if (copy_from_user(str, buf, s))
265 		return -EINVAL;
266 
267 	if (sscanf(str, "disable %32s", block_name) == 1)
268 		op = 0;
269 	else if (sscanf(str, "enable %32s %8s", block_name, err) == 2)
270 		op = 1;
271 	else if (sscanf(str, "inject %32s %8s", block_name, err) == 2)
272 		op = 2;
273 	else if (strstr(str, "retire_page") != NULL)
274 		op = 3;
275 	else if (str[0] && str[1] && str[2] && str[3])
276 		/* ascii string, but commands are not matched. */
277 		return -EINVAL;
278 
279 	if (op != -1) {
280 		if (op == 3) {
281 			if (sscanf(str, "%*s 0x%llx", &address) != 1 &&
282 			    sscanf(str, "%*s %llu", &address) != 1)
283 				return -EINVAL;
284 
285 			data->op = op;
286 			data->inject.address = address;
287 
288 			return 0;
289 		}
290 
291 		if (amdgpu_ras_find_block_id_by_name(block_name, &block_id))
292 			return -EINVAL;
293 
294 		data->head.block = block_id;
295 		/* only ue and ce errors are supported */
296 		if (!memcmp("ue", err, 2))
297 			data->head.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
298 		else if (!memcmp("ce", err, 2))
299 			data->head.type = AMDGPU_RAS_ERROR__SINGLE_CORRECTABLE;
300 		else
301 			return -EINVAL;
302 
303 		data->op = op;
304 
305 		if (op == 2) {
306 			if (sscanf(str, "%*s %*s %*s 0x%x 0x%llx 0x%llx",
307 				   &sub_block, &address, &value) != 3 &&
308 			    sscanf(str, "%*s %*s %*s %u %llu %llu",
309 				   &sub_block, &address, &value) != 3)
310 				return -EINVAL;
311 			data->head.sub_block_index = sub_block;
312 			data->inject.address = address;
313 			data->inject.value = value;
314 		}
315 	} else {
316 		if (size < sizeof(*data))
317 			return -EINVAL;
318 
319 		if (copy_from_user(data, buf, sizeof(*data)))
320 			return -EINVAL;
321 	}
322 
323 	return 0;
324 }
325 
326 /**
327  * DOC: AMDGPU RAS debugfs control interface
328  *
329  * The control interface accepts struct ras_debug_if which has two members.
330  *
331  * First member: ras_debug_if::head or ras_debug_if::inject.
332  *
333  * head is used to indicate which IP block will be under control.
334  *
335  * head has four members, they are block, type, sub_block_index, name.
336  * block: which IP will be under control.
337  * type: what kind of error will be enabled/disabled/injected.
338  * sub_block_index: some IPs have subcomponets. say, GFX, sDMA.
339  * name: the name of IP.
340  *
341  * inject has two more members than head, they are address, value.
342  * As their names indicate, inject operation will write the
343  * value to the address.
344  *
345  * The second member: struct ras_debug_if::op.
346  * It has three kinds of operations.
347  *
348  * - 0: disable RAS on the block. Take ::head as its data.
349  * - 1: enable RAS on the block. Take ::head as its data.
350  * - 2: inject errors on the block. Take ::inject as its data.
351  *
352  * How to use the interface?
353  *
354  * In a program
355  *
356  * Copy the struct ras_debug_if in your code and initialize it.
357  * Write the struct to the control interface.
358  *
359  * From shell
360  *
361  * .. code-block:: bash
362  *
363  *	echo "disable <block>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
364  *	echo "enable  <block> <error>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
365  *	echo "inject  <block> <error> <sub-block> <address> <value> > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
366  *
367  * Where N, is the card which you want to affect.
368  *
369  * "disable" requires only the block.
370  * "enable" requires the block and error type.
371  * "inject" requires the block, error type, address, and value.
372  *
373  * The block is one of: umc, sdma, gfx, etc.
374  *	see ras_block_string[] for details
375  *
376  * The error type is one of: ue, ce, where,
377  *	ue is multi-uncorrectable
378  *	ce is single-correctable
379  *
380  * The sub-block is a the sub-block index, pass 0 if there is no sub-block.
381  * The address and value are hexadecimal numbers, leading 0x is optional.
382  *
383  * For instance,
384  *
385  * .. code-block:: bash
386  *
387  *	echo inject umc ue 0x0 0x0 0x0 > /sys/kernel/debug/dri/0/ras/ras_ctrl
388  *	echo inject umc ce 0 0 0 > /sys/kernel/debug/dri/0/ras/ras_ctrl
389  *	echo disable umc > /sys/kernel/debug/dri/0/ras/ras_ctrl
390  *
391  * How to check the result of the operation?
392  *
393  * To check disable/enable, see "ras" features at,
394  * /sys/class/drm/card[0/1/2...]/device/ras/features
395  *
396  * To check inject, see the corresponding error count at,
397  * /sys/class/drm/card[0/1/2...]/device/ras/[gfx|sdma|umc|...]_err_count
398  *
399  * .. note::
400  *	Operations are only allowed on blocks which are supported.
401  *	Check the "ras" mask at /sys/module/amdgpu/parameters/ras_mask
402  *	to see which blocks support RAS on a particular asic.
403  *
404  */
405 static ssize_t amdgpu_ras_debugfs_ctrl_write(struct file *f,
406 					     const char __user *buf,
407 					     size_t size, loff_t *pos)
408 {
409 	struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
410 	struct ras_debug_if data;
411 	int ret = 0;
412 
413 	if (!amdgpu_ras_get_error_query_ready(adev)) {
414 		dev_warn(adev->dev, "RAS WARN: error injection "
415 				"currently inaccessible\n");
416 		return size;
417 	}
418 
419 	ret = amdgpu_ras_debugfs_ctrl_parse_data(f, buf, size, pos, &data);
420 	if (ret)
421 		return ret;
422 
423 	if (data.op == 3) {
424 		ret = amdgpu_reserve_page_direct(adev, data.inject.address);
425 		if (!ret)
426 			return size;
427 		else
428 			return ret;
429 	}
430 
431 	if (!amdgpu_ras_is_supported(adev, data.head.block))
432 		return -EINVAL;
433 
434 	switch (data.op) {
435 	case 0:
436 		ret = amdgpu_ras_feature_enable(adev, &data.head, 0);
437 		break;
438 	case 1:
439 		ret = amdgpu_ras_feature_enable(adev, &data.head, 1);
440 		break;
441 	case 2:
442 		if ((data.inject.address >= adev->gmc.mc_vram_size) ||
443 		    (data.inject.address >= RAS_UMC_INJECT_ADDR_LIMIT)) {
444 			dev_warn(adev->dev, "RAS WARN: input address "
445 					"0x%llx is invalid.",
446 					data.inject.address);
447 			ret = -EINVAL;
448 			break;
449 		}
450 
451 		/* umc ce/ue error injection for a bad page is not allowed */
452 		if ((data.head.block == AMDGPU_RAS_BLOCK__UMC) &&
453 		    amdgpu_ras_check_bad_page(adev, data.inject.address)) {
454 			dev_warn(adev->dev, "RAS WARN: inject: 0x%llx has "
455 				 "already been marked as bad!\n",
456 				 data.inject.address);
457 			break;
458 		}
459 
460 		/* data.inject.address is offset instead of absolute gpu address */
461 		ret = amdgpu_ras_error_inject(adev, &data.inject);
462 		break;
463 	default:
464 		ret = -EINVAL;
465 		break;
466 	}
467 
468 	if (ret)
469 		return ret;
470 
471 	return size;
472 }
473 
474 /**
475  * DOC: AMDGPU RAS debugfs EEPROM table reset interface
476  *
477  * Some boards contain an EEPROM which is used to persistently store a list of
478  * bad pages which experiences ECC errors in vram.  This interface provides
479  * a way to reset the EEPROM, e.g., after testing error injection.
480  *
481  * Usage:
482  *
483  * .. code-block:: bash
484  *
485  *	echo 1 > ../ras/ras_eeprom_reset
486  *
487  * will reset EEPROM table to 0 entries.
488  *
489  */
490 static ssize_t amdgpu_ras_debugfs_eeprom_write(struct file *f,
491 					       const char __user *buf,
492 					       size_t size, loff_t *pos)
493 {
494 	struct amdgpu_device *adev =
495 		(struct amdgpu_device *)file_inode(f)->i_private;
496 	int ret;
497 
498 	ret = amdgpu_ras_eeprom_reset_table(
499 		&(amdgpu_ras_get_context(adev)->eeprom_control));
500 
501 	if (!ret) {
502 		/* Something was written to EEPROM.
503 		 */
504 		amdgpu_ras_get_context(adev)->flags = RAS_DEFAULT_FLAGS;
505 		return size;
506 	} else {
507 		return ret;
508 	}
509 }
510 
511 static const struct file_operations amdgpu_ras_debugfs_ctrl_ops = {
512 	.owner = THIS_MODULE,
513 	.read = NULL,
514 	.write = amdgpu_ras_debugfs_ctrl_write,
515 	.llseek = default_llseek
516 };
517 
518 static const struct file_operations amdgpu_ras_debugfs_eeprom_ops = {
519 	.owner = THIS_MODULE,
520 	.read = NULL,
521 	.write = amdgpu_ras_debugfs_eeprom_write,
522 	.llseek = default_llseek
523 };
524 
525 /**
526  * DOC: AMDGPU RAS sysfs Error Count Interface
527  *
528  * It allows the user to read the error count for each IP block on the gpu through
529  * /sys/class/drm/card[0/1/2...]/device/ras/[gfx/sdma/...]_err_count
530  *
531  * It outputs the multiple lines which report the uncorrected (ue) and corrected
532  * (ce) error counts.
533  *
534  * The format of one line is below,
535  *
536  * [ce|ue]: count
537  *
538  * Example:
539  *
540  * .. code-block:: bash
541  *
542  *	ue: 0
543  *	ce: 1
544  *
545  */
546 static ssize_t amdgpu_ras_sysfs_read(struct device *dev,
547 		struct device_attribute *attr, char *buf)
548 {
549 	struct ras_manager *obj = container_of(attr, struct ras_manager, sysfs_attr);
550 	struct ras_query_if info = {
551 		.head = obj->head,
552 	};
553 
554 	if (!amdgpu_ras_get_error_query_ready(obj->adev))
555 		return sysfs_emit(buf, "Query currently inaccessible\n");
556 
557 	if (amdgpu_ras_query_error_status(obj->adev, &info))
558 		return -EINVAL;
559 
560 	if (obj->adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 2) &&
561 	    obj->adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 4)) {
562 		if (amdgpu_ras_reset_error_status(obj->adev, info.head.block))
563 			dev_warn(obj->adev->dev, "Failed to reset error counter and error status");
564 	}
565 
566 	return sysfs_emit(buf, "%s: %lu\n%s: %lu\n", "ue", info.ue_count,
567 			  "ce", info.ce_count);
568 }
569 
570 /* obj begin */
571 
572 #define get_obj(obj) do { (obj)->use++; } while (0)
573 #define alive_obj(obj) ((obj)->use)
574 
575 static inline void put_obj(struct ras_manager *obj)
576 {
577 	if (obj && (--obj->use == 0))
578 		list_del(&obj->node);
579 	if (obj && (obj->use < 0))
580 		DRM_ERROR("RAS ERROR: Unbalance obj(%s) use\n", get_ras_block_str(&obj->head));
581 }
582 
583 /* make one obj and return it. */
584 static struct ras_manager *amdgpu_ras_create_obj(struct amdgpu_device *adev,
585 		struct ras_common_if *head)
586 {
587 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
588 	struct ras_manager *obj;
589 
590 	if (!adev->ras_enabled || !con)
591 		return NULL;
592 
593 	if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
594 		return NULL;
595 
596 	if (head->block == AMDGPU_RAS_BLOCK__MCA) {
597 		if (head->sub_block_index >= AMDGPU_RAS_MCA_BLOCK__LAST)
598 			return NULL;
599 
600 		obj = &con->objs[AMDGPU_RAS_BLOCK__LAST + head->sub_block_index];
601 	} else
602 		obj = &con->objs[head->block];
603 
604 	/* already exist. return obj? */
605 	if (alive_obj(obj))
606 		return NULL;
607 
608 	obj->head = *head;
609 	obj->adev = adev;
610 	list_add(&obj->node, &con->head);
611 	get_obj(obj);
612 
613 	return obj;
614 }
615 
616 /* return an obj equal to head, or the first when head is NULL */
617 struct ras_manager *amdgpu_ras_find_obj(struct amdgpu_device *adev,
618 		struct ras_common_if *head)
619 {
620 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
621 	struct ras_manager *obj;
622 	int i;
623 
624 	if (!adev->ras_enabled || !con)
625 		return NULL;
626 
627 	if (head) {
628 		if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
629 			return NULL;
630 
631 		if (head->block == AMDGPU_RAS_BLOCK__MCA) {
632 			if (head->sub_block_index >= AMDGPU_RAS_MCA_BLOCK__LAST)
633 				return NULL;
634 
635 			obj = &con->objs[AMDGPU_RAS_BLOCK__LAST + head->sub_block_index];
636 		} else
637 			obj = &con->objs[head->block];
638 
639 		if (alive_obj(obj))
640 			return obj;
641 	} else {
642 		for (i = 0; i < AMDGPU_RAS_BLOCK_COUNT + AMDGPU_RAS_MCA_BLOCK_COUNT; i++) {
643 			obj = &con->objs[i];
644 			if (alive_obj(obj))
645 				return obj;
646 		}
647 	}
648 
649 	return NULL;
650 }
651 /* obj end */
652 
653 /* feature ctl begin */
654 static int amdgpu_ras_is_feature_allowed(struct amdgpu_device *adev,
655 					 struct ras_common_if *head)
656 {
657 	return adev->ras_hw_enabled & BIT(head->block);
658 }
659 
660 static int amdgpu_ras_is_feature_enabled(struct amdgpu_device *adev,
661 		struct ras_common_if *head)
662 {
663 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
664 
665 	return con->features & BIT(head->block);
666 }
667 
668 /*
669  * if obj is not created, then create one.
670  * set feature enable flag.
671  */
672 static int __amdgpu_ras_feature_enable(struct amdgpu_device *adev,
673 		struct ras_common_if *head, int enable)
674 {
675 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
676 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
677 
678 	/* If hardware does not support ras, then do not create obj.
679 	 * But if hardware support ras, we can create the obj.
680 	 * Ras framework checks con->hw_supported to see if it need do
681 	 * corresponding initialization.
682 	 * IP checks con->support to see if it need disable ras.
683 	 */
684 	if (!amdgpu_ras_is_feature_allowed(adev, head))
685 		return 0;
686 
687 	if (enable) {
688 		if (!obj) {
689 			obj = amdgpu_ras_create_obj(adev, head);
690 			if (!obj)
691 				return -EINVAL;
692 		} else {
693 			/* In case we create obj somewhere else */
694 			get_obj(obj);
695 		}
696 		con->features |= BIT(head->block);
697 	} else {
698 		if (obj && amdgpu_ras_is_feature_enabled(adev, head)) {
699 			con->features &= ~BIT(head->block);
700 			put_obj(obj);
701 		}
702 	}
703 
704 	return 0;
705 }
706 
707 /* wrapper of psp_ras_enable_features */
708 int amdgpu_ras_feature_enable(struct amdgpu_device *adev,
709 		struct ras_common_if *head, bool enable)
710 {
711 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
712 	union ta_ras_cmd_input *info;
713 	int ret;
714 
715 	if (!con)
716 		return -EINVAL;
717 
718 	info = kzalloc(sizeof(union ta_ras_cmd_input), GFP_KERNEL);
719 	if (!info)
720 		return -ENOMEM;
721 
722 	if (!enable) {
723 		info->disable_features = (struct ta_ras_disable_features_input) {
724 			.block_id =  amdgpu_ras_block_to_ta(head->block),
725 			.error_type = amdgpu_ras_error_to_ta(head->type),
726 		};
727 	} else {
728 		info->enable_features = (struct ta_ras_enable_features_input) {
729 			.block_id =  amdgpu_ras_block_to_ta(head->block),
730 			.error_type = amdgpu_ras_error_to_ta(head->type),
731 		};
732 	}
733 
734 	/* Do not enable if it is not allowed. */
735 	WARN_ON(enable && !amdgpu_ras_is_feature_allowed(adev, head));
736 
737 	/* Only enable ras feature operation handle on host side */
738 	if (!amdgpu_sriov_vf(adev) &&
739 		!amdgpu_ras_intr_triggered()) {
740 		ret = psp_ras_enable_features(&adev->psp, info, enable);
741 		if (ret) {
742 			dev_err(adev->dev, "ras %s %s failed poison:%d ret:%d\n",
743 				enable ? "enable":"disable",
744 				get_ras_block_str(head),
745 				amdgpu_ras_is_poison_mode_supported(adev), ret);
746 			goto out;
747 		}
748 	}
749 
750 	/* setup the obj */
751 	__amdgpu_ras_feature_enable(adev, head, enable);
752 	ret = 0;
753 out:
754 	kfree(info);
755 	return ret;
756 }
757 
758 /* Only used in device probe stage and called only once. */
759 int amdgpu_ras_feature_enable_on_boot(struct amdgpu_device *adev,
760 		struct ras_common_if *head, bool enable)
761 {
762 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
763 	int ret;
764 
765 	if (!con)
766 		return -EINVAL;
767 
768 	if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
769 		if (enable) {
770 			/* There is no harm to issue a ras TA cmd regardless of
771 			 * the currecnt ras state.
772 			 * If current state == target state, it will do nothing
773 			 * But sometimes it requests driver to reset and repost
774 			 * with error code -EAGAIN.
775 			 */
776 			ret = amdgpu_ras_feature_enable(adev, head, 1);
777 			/* With old ras TA, we might fail to enable ras.
778 			 * Log it and just setup the object.
779 			 * TODO need remove this WA in the future.
780 			 */
781 			if (ret == -EINVAL) {
782 				ret = __amdgpu_ras_feature_enable(adev, head, 1);
783 				if (!ret)
784 					dev_info(adev->dev,
785 						"RAS INFO: %s setup object\n",
786 						get_ras_block_str(head));
787 			}
788 		} else {
789 			/* setup the object then issue a ras TA disable cmd.*/
790 			ret = __amdgpu_ras_feature_enable(adev, head, 1);
791 			if (ret)
792 				return ret;
793 
794 			/* gfx block ras dsiable cmd must send to ras-ta */
795 			if (head->block == AMDGPU_RAS_BLOCK__GFX)
796 				con->features |= BIT(head->block);
797 
798 			ret = amdgpu_ras_feature_enable(adev, head, 0);
799 
800 			/* clean gfx block ras features flag */
801 			if (adev->ras_enabled && head->block == AMDGPU_RAS_BLOCK__GFX)
802 				con->features &= ~BIT(head->block);
803 		}
804 	} else
805 		ret = amdgpu_ras_feature_enable(adev, head, enable);
806 
807 	return ret;
808 }
809 
810 static int amdgpu_ras_disable_all_features(struct amdgpu_device *adev,
811 		bool bypass)
812 {
813 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
814 	struct ras_manager *obj, *tmp;
815 
816 	list_for_each_entry_safe(obj, tmp, &con->head, node) {
817 		/* bypass psp.
818 		 * aka just release the obj and corresponding flags
819 		 */
820 		if (bypass) {
821 			if (__amdgpu_ras_feature_enable(adev, &obj->head, 0))
822 				break;
823 		} else {
824 			if (amdgpu_ras_feature_enable(adev, &obj->head, 0))
825 				break;
826 		}
827 	}
828 
829 	return con->features;
830 }
831 
832 static int amdgpu_ras_enable_all_features(struct amdgpu_device *adev,
833 		bool bypass)
834 {
835 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
836 	int i;
837 	const enum amdgpu_ras_error_type default_ras_type = AMDGPU_RAS_ERROR__NONE;
838 
839 	for (i = 0; i < AMDGPU_RAS_BLOCK_COUNT; i++) {
840 		struct ras_common_if head = {
841 			.block = i,
842 			.type = default_ras_type,
843 			.sub_block_index = 0,
844 		};
845 
846 		if (i == AMDGPU_RAS_BLOCK__MCA)
847 			continue;
848 
849 		if (bypass) {
850 			/*
851 			 * bypass psp. vbios enable ras for us.
852 			 * so just create the obj
853 			 */
854 			if (__amdgpu_ras_feature_enable(adev, &head, 1))
855 				break;
856 		} else {
857 			if (amdgpu_ras_feature_enable(adev, &head, 1))
858 				break;
859 		}
860 	}
861 
862 	for (i = 0; i < AMDGPU_RAS_MCA_BLOCK_COUNT; i++) {
863 		struct ras_common_if head = {
864 			.block = AMDGPU_RAS_BLOCK__MCA,
865 			.type = default_ras_type,
866 			.sub_block_index = i,
867 		};
868 
869 		if (bypass) {
870 			/*
871 			 * bypass psp. vbios enable ras for us.
872 			 * so just create the obj
873 			 */
874 			if (__amdgpu_ras_feature_enable(adev, &head, 1))
875 				break;
876 		} else {
877 			if (amdgpu_ras_feature_enable(adev, &head, 1))
878 				break;
879 		}
880 	}
881 
882 	return con->features;
883 }
884 /* feature ctl end */
885 
886 static int amdgpu_ras_block_match_default(struct amdgpu_ras_block_object *block_obj,
887 		enum amdgpu_ras_block block)
888 {
889 	if (!block_obj)
890 		return -EINVAL;
891 
892 	if (block_obj->ras_comm.block == block)
893 		return 0;
894 
895 	return -EINVAL;
896 }
897 
898 static struct amdgpu_ras_block_object *amdgpu_ras_get_ras_block(struct amdgpu_device *adev,
899 					enum amdgpu_ras_block block, uint32_t sub_block_index)
900 {
901 	struct amdgpu_ras_block_list *node, *tmp;
902 	struct amdgpu_ras_block_object *obj;
903 
904 	if (block >= AMDGPU_RAS_BLOCK__LAST)
905 		return NULL;
906 
907 	if (!amdgpu_ras_is_supported(adev, block))
908 		return NULL;
909 
910 	list_for_each_entry_safe(node, tmp, &adev->ras_list, node) {
911 		if (!node->ras_obj) {
912 			dev_warn(adev->dev, "Warning: abnormal ras list node.\n");
913 			continue;
914 		}
915 
916 		obj = node->ras_obj;
917 		if (obj->ras_block_match) {
918 			if (obj->ras_block_match(obj, block, sub_block_index) == 0)
919 				return obj;
920 		} else {
921 			if (amdgpu_ras_block_match_default(obj, block) == 0)
922 				return obj;
923 		}
924 	}
925 
926 	return NULL;
927 }
928 
929 static void amdgpu_ras_get_ecc_info(struct amdgpu_device *adev, struct ras_err_data *err_data)
930 {
931 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
932 	int ret = 0;
933 
934 	/*
935 	 * choosing right query method according to
936 	 * whether smu support query error information
937 	 */
938 	ret = amdgpu_dpm_get_ecc_info(adev, (void *)&(ras->umc_ecc));
939 	if (ret == -EOPNOTSUPP) {
940 		if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
941 			adev->umc.ras->ras_block.hw_ops->query_ras_error_count)
942 			adev->umc.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data);
943 
944 		/* umc query_ras_error_address is also responsible for clearing
945 		 * error status
946 		 */
947 		if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
948 		    adev->umc.ras->ras_block.hw_ops->query_ras_error_address)
949 			adev->umc.ras->ras_block.hw_ops->query_ras_error_address(adev, err_data);
950 	} else if (!ret) {
951 		if (adev->umc.ras &&
952 			adev->umc.ras->ecc_info_query_ras_error_count)
953 			adev->umc.ras->ecc_info_query_ras_error_count(adev, err_data);
954 
955 		if (adev->umc.ras &&
956 			adev->umc.ras->ecc_info_query_ras_error_address)
957 			adev->umc.ras->ecc_info_query_ras_error_address(adev, err_data);
958 	}
959 }
960 
961 /* query/inject/cure begin */
962 int amdgpu_ras_query_error_status(struct amdgpu_device *adev,
963 				  struct ras_query_if *info)
964 {
965 	struct amdgpu_ras_block_object *block_obj = NULL;
966 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
967 	struct ras_err_data err_data = {0, 0, 0, NULL};
968 
969 	if (!obj)
970 		return -EINVAL;
971 
972 	if (info->head.block == AMDGPU_RAS_BLOCK__UMC) {
973 		amdgpu_ras_get_ecc_info(adev, &err_data);
974 	} else {
975 		block_obj = amdgpu_ras_get_ras_block(adev, info->head.block, 0);
976 		if (!block_obj || !block_obj->hw_ops)   {
977 			dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
978 				     get_ras_block_str(&info->head));
979 			return -EINVAL;
980 		}
981 
982 		if (block_obj->hw_ops->query_ras_error_count)
983 			block_obj->hw_ops->query_ras_error_count(adev, &err_data);
984 
985 		if ((info->head.block == AMDGPU_RAS_BLOCK__SDMA) ||
986 		    (info->head.block == AMDGPU_RAS_BLOCK__GFX) ||
987 		    (info->head.block == AMDGPU_RAS_BLOCK__MMHUB)) {
988 				if (block_obj->hw_ops->query_ras_error_status)
989 					block_obj->hw_ops->query_ras_error_status(adev);
990 			}
991 	}
992 
993 	obj->err_data.ue_count += err_data.ue_count;
994 	obj->err_data.ce_count += err_data.ce_count;
995 
996 	info->ue_count = obj->err_data.ue_count;
997 	info->ce_count = obj->err_data.ce_count;
998 
999 	if (err_data.ce_count) {
1000 		if (adev->smuio.funcs &&
1001 		    adev->smuio.funcs->get_socket_id &&
1002 		    adev->smuio.funcs->get_die_id) {
1003 			dev_info(adev->dev, "socket: %d, die: %d "
1004 					"%ld correctable hardware errors "
1005 					"detected in %s block, no user "
1006 					"action is needed.\n",
1007 					adev->smuio.funcs->get_socket_id(adev),
1008 					adev->smuio.funcs->get_die_id(adev),
1009 					obj->err_data.ce_count,
1010 					get_ras_block_str(&info->head));
1011 		} else {
1012 			dev_info(adev->dev, "%ld correctable hardware errors "
1013 					"detected in %s block, no user "
1014 					"action is needed.\n",
1015 					obj->err_data.ce_count,
1016 					get_ras_block_str(&info->head));
1017 		}
1018 	}
1019 	if (err_data.ue_count) {
1020 		if (adev->smuio.funcs &&
1021 		    adev->smuio.funcs->get_socket_id &&
1022 		    adev->smuio.funcs->get_die_id) {
1023 			dev_info(adev->dev, "socket: %d, die: %d "
1024 					"%ld uncorrectable hardware errors "
1025 					"detected in %s block\n",
1026 					adev->smuio.funcs->get_socket_id(adev),
1027 					adev->smuio.funcs->get_die_id(adev),
1028 					obj->err_data.ue_count,
1029 					get_ras_block_str(&info->head));
1030 		} else {
1031 			dev_info(adev->dev, "%ld uncorrectable hardware errors "
1032 					"detected in %s block\n",
1033 					obj->err_data.ue_count,
1034 					get_ras_block_str(&info->head));
1035 		}
1036 	}
1037 
1038 	return 0;
1039 }
1040 
1041 int amdgpu_ras_reset_error_status(struct amdgpu_device *adev,
1042 		enum amdgpu_ras_block block)
1043 {
1044 	struct amdgpu_ras_block_object *block_obj = amdgpu_ras_get_ras_block(adev, block, 0);
1045 
1046 	if (!amdgpu_ras_is_supported(adev, block))
1047 		return -EINVAL;
1048 
1049 	if (!block_obj || !block_obj->hw_ops)   {
1050 		dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1051 			     ras_block_str(block));
1052 		return -EINVAL;
1053 	}
1054 
1055 	if (block_obj->hw_ops->reset_ras_error_count)
1056 		block_obj->hw_ops->reset_ras_error_count(adev);
1057 
1058 	if ((block == AMDGPU_RAS_BLOCK__GFX) ||
1059 	    (block == AMDGPU_RAS_BLOCK__MMHUB)) {
1060 		if (block_obj->hw_ops->reset_ras_error_status)
1061 			block_obj->hw_ops->reset_ras_error_status(adev);
1062 	}
1063 
1064 	return 0;
1065 }
1066 
1067 /* wrapper of psp_ras_trigger_error */
1068 int amdgpu_ras_error_inject(struct amdgpu_device *adev,
1069 		struct ras_inject_if *info)
1070 {
1071 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1072 	struct ta_ras_trigger_error_input block_info = {
1073 		.block_id =  amdgpu_ras_block_to_ta(info->head.block),
1074 		.inject_error_type = amdgpu_ras_error_to_ta(info->head.type),
1075 		.sub_block_index = info->head.sub_block_index,
1076 		.address = info->address,
1077 		.value = info->value,
1078 	};
1079 	int ret = -EINVAL;
1080 	struct amdgpu_ras_block_object *block_obj = amdgpu_ras_get_ras_block(adev,
1081 							info->head.block,
1082 							info->head.sub_block_index);
1083 
1084 	if (!obj)
1085 		return -EINVAL;
1086 
1087 	if (!block_obj || !block_obj->hw_ops)	{
1088 		dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1089 			     get_ras_block_str(&info->head));
1090 		return -EINVAL;
1091 	}
1092 
1093 	/* Calculate XGMI relative offset */
1094 	if (adev->gmc.xgmi.num_physical_nodes > 1) {
1095 		block_info.address =
1096 			amdgpu_xgmi_get_relative_phy_addr(adev,
1097 							  block_info.address);
1098 	}
1099 
1100 	if (info->head.block == AMDGPU_RAS_BLOCK__GFX) {
1101 		if (block_obj->hw_ops->ras_error_inject)
1102 			ret = block_obj->hw_ops->ras_error_inject(adev, info);
1103 	} else {
1104 		/* If defined special ras_error_inject(e.g: xgmi), implement special ras_error_inject */
1105 		if (block_obj->hw_ops->ras_error_inject)
1106 			ret = block_obj->hw_ops->ras_error_inject(adev, &block_info);
1107 		else  /*If not defined .ras_error_inject, use default ras_error_inject*/
1108 			ret = psp_ras_trigger_error(&adev->psp, &block_info);
1109 	}
1110 
1111 	if (ret)
1112 		dev_err(adev->dev, "ras inject %s failed %d\n",
1113 			get_ras_block_str(&info->head), ret);
1114 
1115 	return ret;
1116 }
1117 
1118 /**
1119  * amdgpu_ras_query_error_count -- Get error counts of all IPs
1120  * @adev: pointer to AMD GPU device
1121  * @ce_count: pointer to an integer to be set to the count of correctible errors.
1122  * @ue_count: pointer to an integer to be set to the count of uncorrectible
1123  * errors.
1124  *
1125  * If set, @ce_count or @ue_count, count and return the corresponding
1126  * error counts in those integer pointers. Return 0 if the device
1127  * supports RAS. Return -EOPNOTSUPP if the device doesn't support RAS.
1128  */
1129 int amdgpu_ras_query_error_count(struct amdgpu_device *adev,
1130 				 unsigned long *ce_count,
1131 				 unsigned long *ue_count)
1132 {
1133 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1134 	struct ras_manager *obj;
1135 	unsigned long ce, ue;
1136 
1137 	if (!adev->ras_enabled || !con)
1138 		return -EOPNOTSUPP;
1139 
1140 	/* Don't count since no reporting.
1141 	 */
1142 	if (!ce_count && !ue_count)
1143 		return 0;
1144 
1145 	ce = 0;
1146 	ue = 0;
1147 	list_for_each_entry(obj, &con->head, node) {
1148 		struct ras_query_if info = {
1149 			.head = obj->head,
1150 		};
1151 		int res;
1152 
1153 		res = amdgpu_ras_query_error_status(adev, &info);
1154 		if (res)
1155 			return res;
1156 
1157 		if (adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 2) &&
1158 		    adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 4)) {
1159 			if (amdgpu_ras_reset_error_status(adev, info.head.block))
1160 				dev_warn(adev->dev, "Failed to reset error counter and error status");
1161 		}
1162 
1163 		ce += info.ce_count;
1164 		ue += info.ue_count;
1165 	}
1166 
1167 	if (ce_count)
1168 		*ce_count = ce;
1169 
1170 	if (ue_count)
1171 		*ue_count = ue;
1172 
1173 	return 0;
1174 }
1175 /* query/inject/cure end */
1176 
1177 
1178 /* sysfs begin */
1179 
1180 static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
1181 		struct ras_badpage **bps, unsigned int *count);
1182 
1183 static char *amdgpu_ras_badpage_flags_str(unsigned int flags)
1184 {
1185 	switch (flags) {
1186 	case AMDGPU_RAS_RETIRE_PAGE_RESERVED:
1187 		return "R";
1188 	case AMDGPU_RAS_RETIRE_PAGE_PENDING:
1189 		return "P";
1190 	case AMDGPU_RAS_RETIRE_PAGE_FAULT:
1191 	default:
1192 		return "F";
1193 	}
1194 }
1195 
1196 /**
1197  * DOC: AMDGPU RAS sysfs gpu_vram_bad_pages Interface
1198  *
1199  * It allows user to read the bad pages of vram on the gpu through
1200  * /sys/class/drm/card[0/1/2...]/device/ras/gpu_vram_bad_pages
1201  *
1202  * It outputs multiple lines, and each line stands for one gpu page.
1203  *
1204  * The format of one line is below,
1205  * gpu pfn : gpu page size : flags
1206  *
1207  * gpu pfn and gpu page size are printed in hex format.
1208  * flags can be one of below character,
1209  *
1210  * R: reserved, this gpu page is reserved and not able to use.
1211  *
1212  * P: pending for reserve, this gpu page is marked as bad, will be reserved
1213  * in next window of page_reserve.
1214  *
1215  * F: unable to reserve. this gpu page can't be reserved due to some reasons.
1216  *
1217  * Examples:
1218  *
1219  * .. code-block:: bash
1220  *
1221  *	0x00000001 : 0x00001000 : R
1222  *	0x00000002 : 0x00001000 : P
1223  *
1224  */
1225 
1226 static ssize_t amdgpu_ras_sysfs_badpages_read(struct file *f,
1227 		struct kobject *kobj, struct bin_attribute *attr,
1228 		char *buf, loff_t ppos, size_t count)
1229 {
1230 	struct amdgpu_ras *con =
1231 		container_of(attr, struct amdgpu_ras, badpages_attr);
1232 	struct amdgpu_device *adev = con->adev;
1233 	const unsigned int element_size =
1234 		sizeof("0xabcdabcd : 0x12345678 : R\n") - 1;
1235 	unsigned int start = div64_ul(ppos + element_size - 1, element_size);
1236 	unsigned int end = div64_ul(ppos + count - 1, element_size);
1237 	ssize_t s = 0;
1238 	struct ras_badpage *bps = NULL;
1239 	unsigned int bps_count = 0;
1240 
1241 	memset(buf, 0, count);
1242 
1243 	if (amdgpu_ras_badpages_read(adev, &bps, &bps_count))
1244 		return 0;
1245 
1246 	for (; start < end && start < bps_count; start++)
1247 		s += scnprintf(&buf[s], element_size + 1,
1248 				"0x%08x : 0x%08x : %1s\n",
1249 				bps[start].bp,
1250 				bps[start].size,
1251 				amdgpu_ras_badpage_flags_str(bps[start].flags));
1252 
1253 	kfree(bps);
1254 
1255 	return s;
1256 }
1257 
1258 static ssize_t amdgpu_ras_sysfs_features_read(struct device *dev,
1259 		struct device_attribute *attr, char *buf)
1260 {
1261 	struct amdgpu_ras *con =
1262 		container_of(attr, struct amdgpu_ras, features_attr);
1263 
1264 	return scnprintf(buf, PAGE_SIZE, "feature mask: 0x%x\n", con->features);
1265 }
1266 
1267 static void amdgpu_ras_sysfs_remove_bad_page_node(struct amdgpu_device *adev)
1268 {
1269 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1270 
1271 	sysfs_remove_file_from_group(&adev->dev->kobj,
1272 				&con->badpages_attr.attr,
1273 				RAS_FS_NAME);
1274 }
1275 
1276 static int amdgpu_ras_sysfs_remove_feature_node(struct amdgpu_device *adev)
1277 {
1278 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1279 	struct attribute *attrs[] = {
1280 		&con->features_attr.attr,
1281 		NULL
1282 	};
1283 	struct attribute_group group = {
1284 		.name = RAS_FS_NAME,
1285 		.attrs = attrs,
1286 	};
1287 
1288 	sysfs_remove_group(&adev->dev->kobj, &group);
1289 
1290 	return 0;
1291 }
1292 
1293 int amdgpu_ras_sysfs_create(struct amdgpu_device *adev,
1294 		struct ras_common_if *head)
1295 {
1296 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1297 
1298 	if (!obj || obj->attr_inuse)
1299 		return -EINVAL;
1300 
1301 	get_obj(obj);
1302 
1303 	snprintf(obj->fs_data.sysfs_name, sizeof(obj->fs_data.sysfs_name),
1304 		"%s_err_count", head->name);
1305 
1306 	obj->sysfs_attr = (struct device_attribute){
1307 		.attr = {
1308 			.name = obj->fs_data.sysfs_name,
1309 			.mode = S_IRUGO,
1310 		},
1311 			.show = amdgpu_ras_sysfs_read,
1312 	};
1313 	sysfs_attr_init(&obj->sysfs_attr.attr);
1314 
1315 	if (sysfs_add_file_to_group(&adev->dev->kobj,
1316 				&obj->sysfs_attr.attr,
1317 				RAS_FS_NAME)) {
1318 		put_obj(obj);
1319 		return -EINVAL;
1320 	}
1321 
1322 	obj->attr_inuse = 1;
1323 
1324 	return 0;
1325 }
1326 
1327 int amdgpu_ras_sysfs_remove(struct amdgpu_device *adev,
1328 		struct ras_common_if *head)
1329 {
1330 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1331 
1332 	if (!obj || !obj->attr_inuse)
1333 		return -EINVAL;
1334 
1335 	sysfs_remove_file_from_group(&adev->dev->kobj,
1336 				&obj->sysfs_attr.attr,
1337 				RAS_FS_NAME);
1338 	obj->attr_inuse = 0;
1339 	put_obj(obj);
1340 
1341 	return 0;
1342 }
1343 
1344 static int amdgpu_ras_sysfs_remove_all(struct amdgpu_device *adev)
1345 {
1346 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1347 	struct ras_manager *obj, *tmp;
1348 
1349 	list_for_each_entry_safe(obj, tmp, &con->head, node) {
1350 		amdgpu_ras_sysfs_remove(adev, &obj->head);
1351 	}
1352 
1353 	if (amdgpu_bad_page_threshold != 0)
1354 		amdgpu_ras_sysfs_remove_bad_page_node(adev);
1355 
1356 	amdgpu_ras_sysfs_remove_feature_node(adev);
1357 
1358 	return 0;
1359 }
1360 /* sysfs end */
1361 
1362 /**
1363  * DOC: AMDGPU RAS Reboot Behavior for Unrecoverable Errors
1364  *
1365  * Normally when there is an uncorrectable error, the driver will reset
1366  * the GPU to recover.  However, in the event of an unrecoverable error,
1367  * the driver provides an interface to reboot the system automatically
1368  * in that event.
1369  *
1370  * The following file in debugfs provides that interface:
1371  * /sys/kernel/debug/dri/[0/1/2...]/ras/auto_reboot
1372  *
1373  * Usage:
1374  *
1375  * .. code-block:: bash
1376  *
1377  *	echo true > .../ras/auto_reboot
1378  *
1379  */
1380 /* debugfs begin */
1381 static struct dentry *amdgpu_ras_debugfs_create_ctrl_node(struct amdgpu_device *adev)
1382 {
1383 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1384 	struct drm_minor  *minor = adev_to_drm(adev)->primary;
1385 	struct dentry     *dir;
1386 
1387 	dir = debugfs_create_dir(RAS_FS_NAME, minor->debugfs_root);
1388 	debugfs_create_file("ras_ctrl", S_IWUGO | S_IRUGO, dir, adev,
1389 			    &amdgpu_ras_debugfs_ctrl_ops);
1390 	debugfs_create_file("ras_eeprom_reset", S_IWUGO | S_IRUGO, dir, adev,
1391 			    &amdgpu_ras_debugfs_eeprom_ops);
1392 	debugfs_create_u32("bad_page_cnt_threshold", 0444, dir,
1393 			   &con->bad_page_cnt_threshold);
1394 	debugfs_create_x32("ras_hw_enabled", 0444, dir, &adev->ras_hw_enabled);
1395 	debugfs_create_x32("ras_enabled", 0444, dir, &adev->ras_enabled);
1396 	debugfs_create_file("ras_eeprom_size", S_IRUGO, dir, adev,
1397 			    &amdgpu_ras_debugfs_eeprom_size_ops);
1398 	con->de_ras_eeprom_table = debugfs_create_file("ras_eeprom_table",
1399 						       S_IRUGO, dir, adev,
1400 						       &amdgpu_ras_debugfs_eeprom_table_ops);
1401 	amdgpu_ras_debugfs_set_ret_size(&con->eeprom_control);
1402 
1403 	/*
1404 	 * After one uncorrectable error happens, usually GPU recovery will
1405 	 * be scheduled. But due to the known problem in GPU recovery failing
1406 	 * to bring GPU back, below interface provides one direct way to
1407 	 * user to reboot system automatically in such case within
1408 	 * ERREVENT_ATHUB_INTERRUPT generated. Normal GPU recovery routine
1409 	 * will never be called.
1410 	 */
1411 	debugfs_create_bool("auto_reboot", S_IWUGO | S_IRUGO, dir, &con->reboot);
1412 
1413 	/*
1414 	 * User could set this not to clean up hardware's error count register
1415 	 * of RAS IPs during ras recovery.
1416 	 */
1417 	debugfs_create_bool("disable_ras_err_cnt_harvest", 0644, dir,
1418 			    &con->disable_ras_err_cnt_harvest);
1419 	return dir;
1420 }
1421 
1422 static void amdgpu_ras_debugfs_create(struct amdgpu_device *adev,
1423 				      struct ras_fs_if *head,
1424 				      struct dentry *dir)
1425 {
1426 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head->head);
1427 
1428 	if (!obj || !dir)
1429 		return;
1430 
1431 	get_obj(obj);
1432 
1433 	memcpy(obj->fs_data.debugfs_name,
1434 			head->debugfs_name,
1435 			sizeof(obj->fs_data.debugfs_name));
1436 
1437 	debugfs_create_file(obj->fs_data.debugfs_name, S_IWUGO | S_IRUGO, dir,
1438 			    obj, &amdgpu_ras_debugfs_ops);
1439 }
1440 
1441 void amdgpu_ras_debugfs_create_all(struct amdgpu_device *adev)
1442 {
1443 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1444 	struct dentry *dir;
1445 	struct ras_manager *obj;
1446 	struct ras_fs_if fs_info;
1447 
1448 	/*
1449 	 * it won't be called in resume path, no need to check
1450 	 * suspend and gpu reset status
1451 	 */
1452 	if (!IS_ENABLED(CONFIG_DEBUG_FS) || !con)
1453 		return;
1454 
1455 	dir = amdgpu_ras_debugfs_create_ctrl_node(adev);
1456 
1457 	list_for_each_entry(obj, &con->head, node) {
1458 		if (amdgpu_ras_is_supported(adev, obj->head.block) &&
1459 			(obj->attr_inuse == 1)) {
1460 			sprintf(fs_info.debugfs_name, "%s_err_inject",
1461 					get_ras_block_str(&obj->head));
1462 			fs_info.head = obj->head;
1463 			amdgpu_ras_debugfs_create(adev, &fs_info, dir);
1464 		}
1465 	}
1466 }
1467 
1468 /* debugfs end */
1469 
1470 /* ras fs */
1471 static BIN_ATTR(gpu_vram_bad_pages, S_IRUGO,
1472 		amdgpu_ras_sysfs_badpages_read, NULL, 0);
1473 static DEVICE_ATTR(features, S_IRUGO,
1474 		amdgpu_ras_sysfs_features_read, NULL);
1475 static int amdgpu_ras_fs_init(struct amdgpu_device *adev)
1476 {
1477 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1478 	struct attribute_group group = {
1479 		.name = RAS_FS_NAME,
1480 	};
1481 	struct attribute *attrs[] = {
1482 		&con->features_attr.attr,
1483 		NULL
1484 	};
1485 	struct bin_attribute *bin_attrs[] = {
1486 		NULL,
1487 		NULL,
1488 	};
1489 	int r;
1490 
1491 	/* add features entry */
1492 	con->features_attr = dev_attr_features;
1493 	group.attrs = attrs;
1494 	sysfs_attr_init(attrs[0]);
1495 
1496 	if (amdgpu_bad_page_threshold != 0) {
1497 		/* add bad_page_features entry */
1498 		bin_attr_gpu_vram_bad_pages.private = NULL;
1499 		con->badpages_attr = bin_attr_gpu_vram_bad_pages;
1500 		bin_attrs[0] = &con->badpages_attr;
1501 		group.bin_attrs = bin_attrs;
1502 		sysfs_bin_attr_init(bin_attrs[0]);
1503 	}
1504 
1505 	r = sysfs_create_group(&adev->dev->kobj, &group);
1506 	if (r)
1507 		dev_err(adev->dev, "Failed to create RAS sysfs group!");
1508 
1509 	return 0;
1510 }
1511 
1512 static int amdgpu_ras_fs_fini(struct amdgpu_device *adev)
1513 {
1514 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1515 	struct ras_manager *con_obj, *ip_obj, *tmp;
1516 
1517 	if (IS_ENABLED(CONFIG_DEBUG_FS)) {
1518 		list_for_each_entry_safe(con_obj, tmp, &con->head, node) {
1519 			ip_obj = amdgpu_ras_find_obj(adev, &con_obj->head);
1520 			if (ip_obj)
1521 				put_obj(ip_obj);
1522 		}
1523 	}
1524 
1525 	amdgpu_ras_sysfs_remove_all(adev);
1526 	return 0;
1527 }
1528 /* ras fs end */
1529 
1530 /* ih begin */
1531 
1532 /* For the hardware that cannot enable bif ring for both ras_controller_irq
1533  * and ras_err_evnet_athub_irq ih cookies, the driver has to poll status
1534  * register to check whether the interrupt is triggered or not, and properly
1535  * ack the interrupt if it is there
1536  */
1537 void amdgpu_ras_interrupt_fatal_error_handler(struct amdgpu_device *adev)
1538 {
1539 	/* Fatal error events are handled on host side */
1540 	if (amdgpu_sriov_vf(adev) ||
1541 		!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__PCIE_BIF))
1542 		return;
1543 
1544 	if (adev->nbio.ras &&
1545 	    adev->nbio.ras->handle_ras_controller_intr_no_bifring)
1546 		adev->nbio.ras->handle_ras_controller_intr_no_bifring(adev);
1547 
1548 	if (adev->nbio.ras &&
1549 	    adev->nbio.ras->handle_ras_err_event_athub_intr_no_bifring)
1550 		adev->nbio.ras->handle_ras_err_event_athub_intr_no_bifring(adev);
1551 }
1552 
1553 static void amdgpu_ras_interrupt_poison_consumption_handler(struct ras_manager *obj,
1554 				struct amdgpu_iv_entry *entry)
1555 {
1556 	bool poison_stat = false;
1557 	struct amdgpu_device *adev = obj->adev;
1558 	struct ras_err_data err_data = {0, 0, 0, NULL};
1559 	struct amdgpu_ras_block_object *block_obj =
1560 		amdgpu_ras_get_ras_block(adev, obj->head.block, 0);
1561 
1562 	if (!block_obj || !block_obj->hw_ops)
1563 		return;
1564 
1565 	/* both query_poison_status and handle_poison_consumption are optional,
1566 	 * but at least one of them should be implemented if we need poison
1567 	 * consumption handler
1568 	 */
1569 	if (block_obj->hw_ops->query_poison_status) {
1570 		poison_stat = block_obj->hw_ops->query_poison_status(adev);
1571 		if (!poison_stat) {
1572 			/* Not poison consumption interrupt, no need to handle it */
1573 			dev_info(adev->dev, "No RAS poison status in %s poison IH.\n",
1574 					block_obj->ras_comm.name);
1575 
1576 			return;
1577 		}
1578 	}
1579 
1580 	if (!adev->gmc.xgmi.connected_to_cpu)
1581 		amdgpu_umc_poison_handler(adev, &err_data, false);
1582 
1583 	if (block_obj->hw_ops->handle_poison_consumption)
1584 		poison_stat = block_obj->hw_ops->handle_poison_consumption(adev);
1585 
1586 	/* gpu reset is fallback for failed and default cases */
1587 	if (poison_stat) {
1588 		dev_info(adev->dev, "GPU reset for %s RAS poison consumption is issued!\n",
1589 				block_obj->ras_comm.name);
1590 		amdgpu_ras_reset_gpu(adev);
1591 	}
1592 }
1593 
1594 static void amdgpu_ras_interrupt_poison_creation_handler(struct ras_manager *obj,
1595 				struct amdgpu_iv_entry *entry)
1596 {
1597 	dev_info(obj->adev->dev,
1598 		"Poison is created, no user action is needed.\n");
1599 }
1600 
1601 static void amdgpu_ras_interrupt_umc_handler(struct ras_manager *obj,
1602 				struct amdgpu_iv_entry *entry)
1603 {
1604 	struct ras_ih_data *data = &obj->ih_data;
1605 	struct ras_err_data err_data = {0, 0, 0, NULL};
1606 	int ret;
1607 
1608 	if (!data->cb)
1609 		return;
1610 
1611 	/* Let IP handle its data, maybe we need get the output
1612 	 * from the callback to update the error type/count, etc
1613 	 */
1614 	ret = data->cb(obj->adev, &err_data, entry);
1615 	/* ue will trigger an interrupt, and in that case
1616 	 * we need do a reset to recovery the whole system.
1617 	 * But leave IP do that recovery, here we just dispatch
1618 	 * the error.
1619 	 */
1620 	if (ret == AMDGPU_RAS_SUCCESS) {
1621 		/* these counts could be left as 0 if
1622 		 * some blocks do not count error number
1623 		 */
1624 		obj->err_data.ue_count += err_data.ue_count;
1625 		obj->err_data.ce_count += err_data.ce_count;
1626 	}
1627 }
1628 
1629 static void amdgpu_ras_interrupt_handler(struct ras_manager *obj)
1630 {
1631 	struct ras_ih_data *data = &obj->ih_data;
1632 	struct amdgpu_iv_entry entry;
1633 
1634 	while (data->rptr != data->wptr) {
1635 		rmb();
1636 		memcpy(&entry, &data->ring[data->rptr],
1637 				data->element_size);
1638 
1639 		wmb();
1640 		data->rptr = (data->aligned_element_size +
1641 				data->rptr) % data->ring_size;
1642 
1643 		if (amdgpu_ras_is_poison_mode_supported(obj->adev)) {
1644 			if (obj->head.block == AMDGPU_RAS_BLOCK__UMC)
1645 				amdgpu_ras_interrupt_poison_creation_handler(obj, &entry);
1646 			else
1647 				amdgpu_ras_interrupt_poison_consumption_handler(obj, &entry);
1648 		} else {
1649 			if (obj->head.block == AMDGPU_RAS_BLOCK__UMC)
1650 				amdgpu_ras_interrupt_umc_handler(obj, &entry);
1651 			else
1652 				dev_warn(obj->adev->dev,
1653 					"No RAS interrupt handler for non-UMC block with poison disabled.\n");
1654 		}
1655 	}
1656 }
1657 
1658 static void amdgpu_ras_interrupt_process_handler(struct work_struct *work)
1659 {
1660 	struct ras_ih_data *data =
1661 		container_of(work, struct ras_ih_data, ih_work);
1662 	struct ras_manager *obj =
1663 		container_of(data, struct ras_manager, ih_data);
1664 
1665 	amdgpu_ras_interrupt_handler(obj);
1666 }
1667 
1668 int amdgpu_ras_interrupt_dispatch(struct amdgpu_device *adev,
1669 		struct ras_dispatch_if *info)
1670 {
1671 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1672 	struct ras_ih_data *data = &obj->ih_data;
1673 
1674 	if (!obj)
1675 		return -EINVAL;
1676 
1677 	if (data->inuse == 0)
1678 		return 0;
1679 
1680 	/* Might be overflow... */
1681 	memcpy(&data->ring[data->wptr], info->entry,
1682 			data->element_size);
1683 
1684 	wmb();
1685 	data->wptr = (data->aligned_element_size +
1686 			data->wptr) % data->ring_size;
1687 
1688 	schedule_work(&data->ih_work);
1689 
1690 	return 0;
1691 }
1692 
1693 int amdgpu_ras_interrupt_remove_handler(struct amdgpu_device *adev,
1694 		struct ras_common_if *head)
1695 {
1696 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1697 	struct ras_ih_data *data;
1698 
1699 	if (!obj)
1700 		return -EINVAL;
1701 
1702 	data = &obj->ih_data;
1703 	if (data->inuse == 0)
1704 		return 0;
1705 
1706 	cancel_work_sync(&data->ih_work);
1707 
1708 	kfree(data->ring);
1709 	memset(data, 0, sizeof(*data));
1710 	put_obj(obj);
1711 
1712 	return 0;
1713 }
1714 
1715 int amdgpu_ras_interrupt_add_handler(struct amdgpu_device *adev,
1716 		struct ras_common_if *head)
1717 {
1718 	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1719 	struct ras_ih_data *data;
1720 	struct amdgpu_ras_block_object *ras_obj;
1721 
1722 	if (!obj) {
1723 		/* in case we registe the IH before enable ras feature */
1724 		obj = amdgpu_ras_create_obj(adev, head);
1725 		if (!obj)
1726 			return -EINVAL;
1727 	} else
1728 		get_obj(obj);
1729 
1730 	ras_obj = container_of(head, struct amdgpu_ras_block_object, ras_comm);
1731 
1732 	data = &obj->ih_data;
1733 	/* add the callback.etc */
1734 	*data = (struct ras_ih_data) {
1735 		.inuse = 0,
1736 		.cb = ras_obj->ras_cb,
1737 		.element_size = sizeof(struct amdgpu_iv_entry),
1738 		.rptr = 0,
1739 		.wptr = 0,
1740 	};
1741 
1742 	INIT_WORK(&data->ih_work, amdgpu_ras_interrupt_process_handler);
1743 
1744 	data->aligned_element_size = ALIGN(data->element_size, 8);
1745 	/* the ring can store 64 iv entries. */
1746 	data->ring_size = 64 * data->aligned_element_size;
1747 	data->ring = kmalloc(data->ring_size, GFP_KERNEL);
1748 	if (!data->ring) {
1749 		put_obj(obj);
1750 		return -ENOMEM;
1751 	}
1752 
1753 	/* IH is ready */
1754 	data->inuse = 1;
1755 
1756 	return 0;
1757 }
1758 
1759 static int amdgpu_ras_interrupt_remove_all(struct amdgpu_device *adev)
1760 {
1761 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1762 	struct ras_manager *obj, *tmp;
1763 
1764 	list_for_each_entry_safe(obj, tmp, &con->head, node) {
1765 		amdgpu_ras_interrupt_remove_handler(adev, &obj->head);
1766 	}
1767 
1768 	return 0;
1769 }
1770 /* ih end */
1771 
1772 /* traversal all IPs except NBIO to query error counter */
1773 static void amdgpu_ras_log_on_err_counter(struct amdgpu_device *adev)
1774 {
1775 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1776 	struct ras_manager *obj;
1777 
1778 	if (!adev->ras_enabled || !con)
1779 		return;
1780 
1781 	list_for_each_entry(obj, &con->head, node) {
1782 		struct ras_query_if info = {
1783 			.head = obj->head,
1784 		};
1785 
1786 		/*
1787 		 * PCIE_BIF IP has one different isr by ras controller
1788 		 * interrupt, the specific ras counter query will be
1789 		 * done in that isr. So skip such block from common
1790 		 * sync flood interrupt isr calling.
1791 		 */
1792 		if (info.head.block == AMDGPU_RAS_BLOCK__PCIE_BIF)
1793 			continue;
1794 
1795 		/*
1796 		 * this is a workaround for aldebaran, skip send msg to
1797 		 * smu to get ecc_info table due to smu handle get ecc
1798 		 * info table failed temporarily.
1799 		 * should be removed until smu fix handle ecc_info table.
1800 		 */
1801 		if ((info.head.block == AMDGPU_RAS_BLOCK__UMC) &&
1802 			(adev->ip_versions[MP1_HWIP][0] == IP_VERSION(13, 0, 2)))
1803 			continue;
1804 
1805 		amdgpu_ras_query_error_status(adev, &info);
1806 
1807 		if (adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 2) &&
1808 		    adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 4)) {
1809 			if (amdgpu_ras_reset_error_status(adev, info.head.block))
1810 				dev_warn(adev->dev, "Failed to reset error counter and error status");
1811 		}
1812 	}
1813 }
1814 
1815 /* Parse RdRspStatus and WrRspStatus */
1816 static void amdgpu_ras_error_status_query(struct amdgpu_device *adev,
1817 					  struct ras_query_if *info)
1818 {
1819 	struct amdgpu_ras_block_object *block_obj;
1820 	/*
1821 	 * Only two block need to query read/write
1822 	 * RspStatus at current state
1823 	 */
1824 	if ((info->head.block != AMDGPU_RAS_BLOCK__GFX) &&
1825 		(info->head.block != AMDGPU_RAS_BLOCK__MMHUB))
1826 		return;
1827 
1828 	block_obj = amdgpu_ras_get_ras_block(adev,
1829 					info->head.block,
1830 					info->head.sub_block_index);
1831 
1832 	if (!block_obj || !block_obj->hw_ops) {
1833 		dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1834 			     get_ras_block_str(&info->head));
1835 		return;
1836 	}
1837 
1838 	if (block_obj->hw_ops->query_ras_error_status)
1839 		block_obj->hw_ops->query_ras_error_status(adev);
1840 
1841 }
1842 
1843 static void amdgpu_ras_query_err_status(struct amdgpu_device *adev)
1844 {
1845 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1846 	struct ras_manager *obj;
1847 
1848 	if (!adev->ras_enabled || !con)
1849 		return;
1850 
1851 	list_for_each_entry(obj, &con->head, node) {
1852 		struct ras_query_if info = {
1853 			.head = obj->head,
1854 		};
1855 
1856 		amdgpu_ras_error_status_query(adev, &info);
1857 	}
1858 }
1859 
1860 /* recovery begin */
1861 
1862 /* return 0 on success.
1863  * caller need free bps.
1864  */
1865 static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
1866 		struct ras_badpage **bps, unsigned int *count)
1867 {
1868 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1869 	struct ras_err_handler_data *data;
1870 	int i = 0;
1871 	int ret = 0, status;
1872 
1873 	if (!con || !con->eh_data || !bps || !count)
1874 		return -EINVAL;
1875 
1876 	mutex_lock(&con->recovery_lock);
1877 	data = con->eh_data;
1878 	if (!data || data->count == 0) {
1879 		*bps = NULL;
1880 		ret = -EINVAL;
1881 		goto out;
1882 	}
1883 
1884 	*bps = kmalloc(sizeof(struct ras_badpage) * data->count, GFP_KERNEL);
1885 	if (!*bps) {
1886 		ret = -ENOMEM;
1887 		goto out;
1888 	}
1889 
1890 	for (; i < data->count; i++) {
1891 		(*bps)[i] = (struct ras_badpage){
1892 			.bp = data->bps[i].retired_page,
1893 			.size = AMDGPU_GPU_PAGE_SIZE,
1894 			.flags = AMDGPU_RAS_RETIRE_PAGE_RESERVED,
1895 		};
1896 		status = amdgpu_vram_mgr_query_page_status(&adev->mman.vram_mgr,
1897 				data->bps[i].retired_page);
1898 		if (status == -EBUSY)
1899 			(*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_PENDING;
1900 		else if (status == -ENOENT)
1901 			(*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_FAULT;
1902 	}
1903 
1904 	*count = data->count;
1905 out:
1906 	mutex_unlock(&con->recovery_lock);
1907 	return ret;
1908 }
1909 
1910 static void amdgpu_ras_do_recovery(struct work_struct *work)
1911 {
1912 	struct amdgpu_ras *ras =
1913 		container_of(work, struct amdgpu_ras, recovery_work);
1914 	struct amdgpu_device *remote_adev = NULL;
1915 	struct amdgpu_device *adev = ras->adev;
1916 	struct list_head device_list, *device_list_handle =  NULL;
1917 
1918 	if (!ras->disable_ras_err_cnt_harvest) {
1919 		struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
1920 
1921 		/* Build list of devices to query RAS related errors */
1922 		if  (hive && adev->gmc.xgmi.num_physical_nodes > 1) {
1923 			device_list_handle = &hive->device_list;
1924 		} else {
1925 			INIT_LIST_HEAD(&device_list);
1926 			list_add_tail(&adev->gmc.xgmi.head, &device_list);
1927 			device_list_handle = &device_list;
1928 		}
1929 
1930 		list_for_each_entry(remote_adev,
1931 				device_list_handle, gmc.xgmi.head) {
1932 			amdgpu_ras_query_err_status(remote_adev);
1933 			amdgpu_ras_log_on_err_counter(remote_adev);
1934 		}
1935 
1936 		amdgpu_put_xgmi_hive(hive);
1937 	}
1938 
1939 	if (amdgpu_device_should_recover_gpu(ras->adev))
1940 		amdgpu_device_gpu_recover(ras->adev, NULL);
1941 	atomic_set(&ras->in_recovery, 0);
1942 }
1943 
1944 /* alloc/realloc bps array */
1945 static int amdgpu_ras_realloc_eh_data_space(struct amdgpu_device *adev,
1946 		struct ras_err_handler_data *data, int pages)
1947 {
1948 	unsigned int old_space = data->count + data->space_left;
1949 	unsigned int new_space = old_space + pages;
1950 	unsigned int align_space = ALIGN(new_space, 512);
1951 	void *bps = kmalloc(align_space * sizeof(*data->bps), GFP_KERNEL);
1952 
1953 	if (!bps) {
1954 		return -ENOMEM;
1955 	}
1956 
1957 	if (data->bps) {
1958 		memcpy(bps, data->bps,
1959 				data->count * sizeof(*data->bps));
1960 		kfree(data->bps);
1961 	}
1962 
1963 	data->bps = bps;
1964 	data->space_left += align_space - old_space;
1965 	return 0;
1966 }
1967 
1968 /* it deal with vram only. */
1969 int amdgpu_ras_add_bad_pages(struct amdgpu_device *adev,
1970 		struct eeprom_table_record *bps, int pages)
1971 {
1972 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1973 	struct ras_err_handler_data *data;
1974 	int ret = 0;
1975 	uint32_t i;
1976 
1977 	if (!con || !con->eh_data || !bps || pages <= 0)
1978 		return 0;
1979 
1980 	mutex_lock(&con->recovery_lock);
1981 	data = con->eh_data;
1982 	if (!data)
1983 		goto out;
1984 
1985 	for (i = 0; i < pages; i++) {
1986 		if (amdgpu_ras_check_bad_page_unlock(con,
1987 			bps[i].retired_page << AMDGPU_GPU_PAGE_SHIFT))
1988 			continue;
1989 
1990 		if (!data->space_left &&
1991 			amdgpu_ras_realloc_eh_data_space(adev, data, 256)) {
1992 			ret = -ENOMEM;
1993 			goto out;
1994 		}
1995 
1996 		amdgpu_vram_mgr_reserve_range(&adev->mman.vram_mgr,
1997 			bps[i].retired_page << AMDGPU_GPU_PAGE_SHIFT,
1998 			AMDGPU_GPU_PAGE_SIZE);
1999 
2000 		memcpy(&data->bps[data->count], &bps[i], sizeof(*data->bps));
2001 		data->count++;
2002 		data->space_left--;
2003 	}
2004 out:
2005 	mutex_unlock(&con->recovery_lock);
2006 
2007 	return ret;
2008 }
2009 
2010 /*
2011  * write error record array to eeprom, the function should be
2012  * protected by recovery_lock
2013  */
2014 int amdgpu_ras_save_bad_pages(struct amdgpu_device *adev)
2015 {
2016 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2017 	struct ras_err_handler_data *data;
2018 	struct amdgpu_ras_eeprom_control *control;
2019 	int save_count;
2020 
2021 	if (!con || !con->eh_data)
2022 		return 0;
2023 
2024 	mutex_lock(&con->recovery_lock);
2025 	control = &con->eeprom_control;
2026 	data = con->eh_data;
2027 	save_count = data->count - control->ras_num_recs;
2028 	mutex_unlock(&con->recovery_lock);
2029 	/* only new entries are saved */
2030 	if (save_count > 0) {
2031 		if (amdgpu_ras_eeprom_append(control,
2032 					     &data->bps[control->ras_num_recs],
2033 					     save_count)) {
2034 			dev_err(adev->dev, "Failed to save EEPROM table data!");
2035 			return -EIO;
2036 		}
2037 
2038 		dev_info(adev->dev, "Saved %d pages to EEPROM table.\n", save_count);
2039 	}
2040 
2041 	return 0;
2042 }
2043 
2044 /*
2045  * read error record array in eeprom and reserve enough space for
2046  * storing new bad pages
2047  */
2048 static int amdgpu_ras_load_bad_pages(struct amdgpu_device *adev)
2049 {
2050 	struct amdgpu_ras_eeprom_control *control =
2051 		&adev->psp.ras_context.ras->eeprom_control;
2052 	struct eeprom_table_record *bps;
2053 	int ret;
2054 
2055 	/* no bad page record, skip eeprom access */
2056 	if (control->ras_num_recs == 0 || amdgpu_bad_page_threshold == 0)
2057 		return 0;
2058 
2059 	bps = kcalloc(control->ras_num_recs, sizeof(*bps), GFP_KERNEL);
2060 	if (!bps)
2061 		return -ENOMEM;
2062 
2063 	ret = amdgpu_ras_eeprom_read(control, bps, control->ras_num_recs);
2064 	if (ret)
2065 		dev_err(adev->dev, "Failed to load EEPROM table records!");
2066 	else
2067 		ret = amdgpu_ras_add_bad_pages(adev, bps, control->ras_num_recs);
2068 
2069 	kfree(bps);
2070 	return ret;
2071 }
2072 
2073 static bool amdgpu_ras_check_bad_page_unlock(struct amdgpu_ras *con,
2074 				uint64_t addr)
2075 {
2076 	struct ras_err_handler_data *data = con->eh_data;
2077 	int i;
2078 
2079 	addr >>= AMDGPU_GPU_PAGE_SHIFT;
2080 	for (i = 0; i < data->count; i++)
2081 		if (addr == data->bps[i].retired_page)
2082 			return true;
2083 
2084 	return false;
2085 }
2086 
2087 /*
2088  * check if an address belongs to bad page
2089  *
2090  * Note: this check is only for umc block
2091  */
2092 static bool amdgpu_ras_check_bad_page(struct amdgpu_device *adev,
2093 				uint64_t addr)
2094 {
2095 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2096 	bool ret = false;
2097 
2098 	if (!con || !con->eh_data)
2099 		return ret;
2100 
2101 	mutex_lock(&con->recovery_lock);
2102 	ret = amdgpu_ras_check_bad_page_unlock(con, addr);
2103 	mutex_unlock(&con->recovery_lock);
2104 	return ret;
2105 }
2106 
2107 static void amdgpu_ras_validate_threshold(struct amdgpu_device *adev,
2108 					  uint32_t max_count)
2109 {
2110 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2111 
2112 	/*
2113 	 * Justification of value bad_page_cnt_threshold in ras structure
2114 	 *
2115 	 * Generally, -1 <= amdgpu_bad_page_threshold <= max record length
2116 	 * in eeprom, and introduce two scenarios accordingly.
2117 	 *
2118 	 * Bad page retirement enablement:
2119 	 *    - If amdgpu_bad_page_threshold = -1,
2120 	 *      bad_page_cnt_threshold = typical value by formula.
2121 	 *
2122 	 *    - When the value from user is 0 < amdgpu_bad_page_threshold <
2123 	 *      max record length in eeprom, use it directly.
2124 	 *
2125 	 * Bad page retirement disablement:
2126 	 *    - If amdgpu_bad_page_threshold = 0, bad page retirement
2127 	 *      functionality is disabled, and bad_page_cnt_threshold will
2128 	 *      take no effect.
2129 	 */
2130 
2131 	if (amdgpu_bad_page_threshold < 0) {
2132 		u64 val = adev->gmc.mc_vram_size;
2133 
2134 		do_div(val, RAS_BAD_PAGE_COVER);
2135 		con->bad_page_cnt_threshold = min(lower_32_bits(val),
2136 						  max_count);
2137 	} else {
2138 		con->bad_page_cnt_threshold = min_t(int, max_count,
2139 						    amdgpu_bad_page_threshold);
2140 	}
2141 }
2142 
2143 int amdgpu_ras_recovery_init(struct amdgpu_device *adev)
2144 {
2145 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2146 	struct ras_err_handler_data **data;
2147 	u32  max_eeprom_records_count = 0;
2148 	bool exc_err_limit = false;
2149 	int ret;
2150 
2151 	if (!con)
2152 		return 0;
2153 
2154 	/* Allow access to RAS EEPROM via debugfs, when the ASIC
2155 	 * supports RAS and debugfs is enabled, but when
2156 	 * adev->ras_enabled is unset, i.e. when "ras_enable"
2157 	 * module parameter is set to 0.
2158 	 */
2159 	con->adev = adev;
2160 
2161 	if (!adev->ras_enabled)
2162 		return 0;
2163 
2164 	data = &con->eh_data;
2165 	*data = kmalloc(sizeof(**data), GFP_KERNEL | __GFP_ZERO);
2166 	if (!*data) {
2167 		ret = -ENOMEM;
2168 		goto out;
2169 	}
2170 
2171 	mutex_init(&con->recovery_lock);
2172 	INIT_WORK(&con->recovery_work, amdgpu_ras_do_recovery);
2173 	atomic_set(&con->in_recovery, 0);
2174 	con->eeprom_control.bad_channel_bitmap = 0;
2175 
2176 	max_eeprom_records_count = amdgpu_ras_eeprom_max_record_count();
2177 	amdgpu_ras_validate_threshold(adev, max_eeprom_records_count);
2178 
2179 	/* Todo: During test the SMU might fail to read the eeprom through I2C
2180 	 * when the GPU is pending on XGMI reset during probe time
2181 	 * (Mostly after second bus reset), skip it now
2182 	 */
2183 	if (adev->gmc.xgmi.pending_reset)
2184 		return 0;
2185 	ret = amdgpu_ras_eeprom_init(&con->eeprom_control, &exc_err_limit);
2186 	/*
2187 	 * This calling fails when exc_err_limit is true or
2188 	 * ret != 0.
2189 	 */
2190 	if (exc_err_limit || ret)
2191 		goto free;
2192 
2193 	if (con->eeprom_control.ras_num_recs) {
2194 		ret = amdgpu_ras_load_bad_pages(adev);
2195 		if (ret)
2196 			goto free;
2197 
2198 		amdgpu_dpm_send_hbm_bad_pages_num(adev, con->eeprom_control.ras_num_recs);
2199 
2200 		if (con->update_channel_flag == true) {
2201 			amdgpu_dpm_send_hbm_bad_channel_flag(adev, con->eeprom_control.bad_channel_bitmap);
2202 			con->update_channel_flag = false;
2203 		}
2204 	}
2205 
2206 #ifdef CONFIG_X86_MCE_AMD
2207 	if ((adev->asic_type == CHIP_ALDEBARAN) &&
2208 	    (adev->gmc.xgmi.connected_to_cpu))
2209 		amdgpu_register_bad_pages_mca_notifier(adev);
2210 #endif
2211 	return 0;
2212 
2213 free:
2214 	kfree((*data)->bps);
2215 	kfree(*data);
2216 	con->eh_data = NULL;
2217 out:
2218 	dev_warn(adev->dev, "Failed to initialize ras recovery! (%d)\n", ret);
2219 
2220 	/*
2221 	 * Except error threshold exceeding case, other failure cases in this
2222 	 * function would not fail amdgpu driver init.
2223 	 */
2224 	if (!exc_err_limit)
2225 		ret = 0;
2226 	else
2227 		ret = -EINVAL;
2228 
2229 	return ret;
2230 }
2231 
2232 static int amdgpu_ras_recovery_fini(struct amdgpu_device *adev)
2233 {
2234 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2235 	struct ras_err_handler_data *data = con->eh_data;
2236 
2237 	/* recovery_init failed to init it, fini is useless */
2238 	if (!data)
2239 		return 0;
2240 
2241 	cancel_work_sync(&con->recovery_work);
2242 
2243 	mutex_lock(&con->recovery_lock);
2244 	con->eh_data = NULL;
2245 	kfree(data->bps);
2246 	kfree(data);
2247 	mutex_unlock(&con->recovery_lock);
2248 
2249 	return 0;
2250 }
2251 /* recovery end */
2252 
2253 static bool amdgpu_ras_asic_supported(struct amdgpu_device *adev)
2254 {
2255 	return adev->asic_type == CHIP_VEGA10 ||
2256 		adev->asic_type == CHIP_VEGA20 ||
2257 		adev->asic_type == CHIP_ARCTURUS ||
2258 		adev->asic_type == CHIP_ALDEBARAN ||
2259 		adev->asic_type == CHIP_SIENNA_CICHLID;
2260 }
2261 
2262 /*
2263  * this is workaround for vega20 workstation sku,
2264  * force enable gfx ras, ignore vbios gfx ras flag
2265  * due to GC EDC can not write
2266  */
2267 static void amdgpu_ras_get_quirks(struct amdgpu_device *adev)
2268 {
2269 	struct atom_context *ctx = adev->mode_info.atom_context;
2270 
2271 	if (!ctx)
2272 		return;
2273 
2274 	if (strnstr(ctx->vbios_version, "D16406",
2275 		    sizeof(ctx->vbios_version)) ||
2276 		strnstr(ctx->vbios_version, "D36002",
2277 			sizeof(ctx->vbios_version)))
2278 		adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX);
2279 }
2280 
2281 /*
2282  * check hardware's ras ability which will be saved in hw_supported.
2283  * if hardware does not support ras, we can skip some ras initializtion and
2284  * forbid some ras operations from IP.
2285  * if software itself, say boot parameter, limit the ras ability. We still
2286  * need allow IP do some limited operations, like disable. In such case,
2287  * we have to initialize ras as normal. but need check if operation is
2288  * allowed or not in each function.
2289  */
2290 static void amdgpu_ras_check_supported(struct amdgpu_device *adev)
2291 {
2292 	adev->ras_hw_enabled = adev->ras_enabled = 0;
2293 
2294 	if (!adev->is_atom_fw ||
2295 	    !amdgpu_ras_asic_supported(adev))
2296 		return;
2297 
2298 	/* If driver run on sriov guest side, only enable ras for aldebaran */
2299 	if (amdgpu_sriov_vf(adev) &&
2300 		adev->ip_versions[MP1_HWIP][0] != IP_VERSION(13, 0, 2))
2301 		return;
2302 
2303 	if (!adev->gmc.xgmi.connected_to_cpu) {
2304 		if (amdgpu_atomfirmware_mem_ecc_supported(adev)) {
2305 			dev_info(adev->dev, "MEM ECC is active.\n");
2306 			adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__UMC |
2307 						   1 << AMDGPU_RAS_BLOCK__DF);
2308 		} else {
2309 			dev_info(adev->dev, "MEM ECC is not presented.\n");
2310 		}
2311 
2312 		if (amdgpu_atomfirmware_sram_ecc_supported(adev)) {
2313 			dev_info(adev->dev, "SRAM ECC is active.\n");
2314 			if (!amdgpu_sriov_vf(adev)) {
2315 				adev->ras_hw_enabled |= ~(1 << AMDGPU_RAS_BLOCK__UMC |
2316 							    1 << AMDGPU_RAS_BLOCK__DF);
2317 
2318 				if (adev->ip_versions[VCN_HWIP][0] == IP_VERSION(2, 6, 0))
2319 					adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__VCN |
2320 							1 << AMDGPU_RAS_BLOCK__JPEG);
2321 				else
2322 					adev->ras_hw_enabled &= ~(1 << AMDGPU_RAS_BLOCK__VCN |
2323 							1 << AMDGPU_RAS_BLOCK__JPEG);
2324 			} else {
2325 				adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__PCIE_BIF |
2326 								1 << AMDGPU_RAS_BLOCK__SDMA |
2327 								1 << AMDGPU_RAS_BLOCK__GFX);
2328 			}
2329 		} else {
2330 			dev_info(adev->dev, "SRAM ECC is not presented.\n");
2331 		}
2332 	} else {
2333 		/* driver only manages a few IP blocks RAS feature
2334 		 * when GPU is connected cpu through XGMI */
2335 		adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX |
2336 					   1 << AMDGPU_RAS_BLOCK__SDMA |
2337 					   1 << AMDGPU_RAS_BLOCK__MMHUB);
2338 	}
2339 
2340 	amdgpu_ras_get_quirks(adev);
2341 
2342 	/* hw_supported needs to be aligned with RAS block mask. */
2343 	adev->ras_hw_enabled &= AMDGPU_RAS_BLOCK_MASK;
2344 
2345 	adev->ras_enabled = amdgpu_ras_enable == 0 ? 0 :
2346 		adev->ras_hw_enabled & amdgpu_ras_mask;
2347 }
2348 
2349 static void amdgpu_ras_counte_dw(struct work_struct *work)
2350 {
2351 	struct amdgpu_ras *con = container_of(work, struct amdgpu_ras,
2352 					      ras_counte_delay_work.work);
2353 	struct amdgpu_device *adev = con->adev;
2354 	struct drm_device *dev = adev_to_drm(adev);
2355 	unsigned long ce_count, ue_count;
2356 	int res;
2357 
2358 	res = pm_runtime_get_sync(dev->dev);
2359 	if (res < 0)
2360 		goto Out;
2361 
2362 	/* Cache new values.
2363 	 */
2364 	if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count) == 0) {
2365 		atomic_set(&con->ras_ce_count, ce_count);
2366 		atomic_set(&con->ras_ue_count, ue_count);
2367 	}
2368 
2369 	pm_runtime_mark_last_busy(dev->dev);
2370 Out:
2371 	pm_runtime_put_autosuspend(dev->dev);
2372 }
2373 
2374 int amdgpu_ras_init(struct amdgpu_device *adev)
2375 {
2376 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2377 	int r;
2378 	bool df_poison, umc_poison;
2379 
2380 	if (con)
2381 		return 0;
2382 
2383 	con = kmalloc(sizeof(struct amdgpu_ras) +
2384 			sizeof(struct ras_manager) * AMDGPU_RAS_BLOCK_COUNT +
2385 			sizeof(struct ras_manager) * AMDGPU_RAS_MCA_BLOCK_COUNT,
2386 			GFP_KERNEL|__GFP_ZERO);
2387 	if (!con)
2388 		return -ENOMEM;
2389 
2390 	con->adev = adev;
2391 	INIT_DELAYED_WORK(&con->ras_counte_delay_work, amdgpu_ras_counte_dw);
2392 	atomic_set(&con->ras_ce_count, 0);
2393 	atomic_set(&con->ras_ue_count, 0);
2394 
2395 	con->objs = (struct ras_manager *)(con + 1);
2396 
2397 	amdgpu_ras_set_context(adev, con);
2398 
2399 	amdgpu_ras_check_supported(adev);
2400 
2401 	if (!adev->ras_enabled || adev->asic_type == CHIP_VEGA10) {
2402 		/* set gfx block ras context feature for VEGA20 Gaming
2403 		 * send ras disable cmd to ras ta during ras late init.
2404 		 */
2405 		if (!adev->ras_enabled && adev->asic_type == CHIP_VEGA20) {
2406 			con->features |= BIT(AMDGPU_RAS_BLOCK__GFX);
2407 
2408 			return 0;
2409 		}
2410 
2411 		r = 0;
2412 		goto release_con;
2413 	}
2414 
2415 	con->update_channel_flag = false;
2416 	con->features = 0;
2417 	INIT_LIST_HEAD(&con->head);
2418 	/* Might need get this flag from vbios. */
2419 	con->flags = RAS_DEFAULT_FLAGS;
2420 
2421 	/* initialize nbio ras function ahead of any other
2422 	 * ras functions so hardware fatal error interrupt
2423 	 * can be enabled as early as possible */
2424 	switch (adev->asic_type) {
2425 	case CHIP_VEGA20:
2426 	case CHIP_ARCTURUS:
2427 	case CHIP_ALDEBARAN:
2428 		if (!adev->gmc.xgmi.connected_to_cpu) {
2429 			adev->nbio.ras = &nbio_v7_4_ras;
2430 			amdgpu_ras_register_ras_block(adev, &adev->nbio.ras->ras_block);
2431 			adev->nbio.ras_if = &adev->nbio.ras->ras_block.ras_comm;
2432 		}
2433 		break;
2434 	default:
2435 		/* nbio ras is not available */
2436 		break;
2437 	}
2438 
2439 	if (adev->nbio.ras &&
2440 	    adev->nbio.ras->init_ras_controller_interrupt) {
2441 		r = adev->nbio.ras->init_ras_controller_interrupt(adev);
2442 		if (r)
2443 			goto release_con;
2444 	}
2445 
2446 	if (adev->nbio.ras &&
2447 	    adev->nbio.ras->init_ras_err_event_athub_interrupt) {
2448 		r = adev->nbio.ras->init_ras_err_event_athub_interrupt(adev);
2449 		if (r)
2450 			goto release_con;
2451 	}
2452 
2453 	/* Init poison supported flag, the default value is false */
2454 	if (adev->gmc.xgmi.connected_to_cpu) {
2455 		/* enabled by default when GPU is connected to CPU */
2456 		con->poison_supported = true;
2457 	}
2458 	else if (adev->df.funcs &&
2459 	    adev->df.funcs->query_ras_poison_mode &&
2460 	    adev->umc.ras &&
2461 	    adev->umc.ras->query_ras_poison_mode) {
2462 		df_poison =
2463 			adev->df.funcs->query_ras_poison_mode(adev);
2464 		umc_poison =
2465 			adev->umc.ras->query_ras_poison_mode(adev);
2466 		/* Only poison is set in both DF and UMC, we can support it */
2467 		if (df_poison && umc_poison)
2468 			con->poison_supported = true;
2469 		else if (df_poison != umc_poison)
2470 			dev_warn(adev->dev, "Poison setting is inconsistent in DF/UMC(%d:%d)!\n",
2471 					df_poison, umc_poison);
2472 	}
2473 
2474 	if (amdgpu_ras_fs_init(adev)) {
2475 		r = -EINVAL;
2476 		goto release_con;
2477 	}
2478 
2479 	dev_info(adev->dev, "RAS INFO: ras initialized successfully, "
2480 		 "hardware ability[%x] ras_mask[%x]\n",
2481 		 adev->ras_hw_enabled, adev->ras_enabled);
2482 
2483 	return 0;
2484 release_con:
2485 	amdgpu_ras_set_context(adev, NULL);
2486 	kfree(con);
2487 
2488 	return r;
2489 }
2490 
2491 int amdgpu_persistent_edc_harvesting_supported(struct amdgpu_device *adev)
2492 {
2493 	if (adev->gmc.xgmi.connected_to_cpu)
2494 		return 1;
2495 	return 0;
2496 }
2497 
2498 static int amdgpu_persistent_edc_harvesting(struct amdgpu_device *adev,
2499 					struct ras_common_if *ras_block)
2500 {
2501 	struct ras_query_if info = {
2502 		.head = *ras_block,
2503 	};
2504 
2505 	if (!amdgpu_persistent_edc_harvesting_supported(adev))
2506 		return 0;
2507 
2508 	if (amdgpu_ras_query_error_status(adev, &info) != 0)
2509 		DRM_WARN("RAS init harvest failure");
2510 
2511 	if (amdgpu_ras_reset_error_status(adev, ras_block->block) != 0)
2512 		DRM_WARN("RAS init harvest reset failure");
2513 
2514 	return 0;
2515 }
2516 
2517 bool amdgpu_ras_is_poison_mode_supported(struct amdgpu_device *adev)
2518 {
2519        struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2520 
2521        if (!con)
2522                return false;
2523 
2524        return con->poison_supported;
2525 }
2526 
2527 /* helper function to handle common stuff in ip late init phase */
2528 int amdgpu_ras_block_late_init(struct amdgpu_device *adev,
2529 			 struct ras_common_if *ras_block)
2530 {
2531 	struct amdgpu_ras_block_object *ras_obj = NULL;
2532 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2533 	unsigned long ue_count, ce_count;
2534 	int r;
2535 
2536 	/* disable RAS feature per IP block if it is not supported */
2537 	if (!amdgpu_ras_is_supported(adev, ras_block->block)) {
2538 		amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
2539 		return 0;
2540 	}
2541 
2542 	r = amdgpu_ras_feature_enable_on_boot(adev, ras_block, 1);
2543 	if (r) {
2544 		if (adev->in_suspend || amdgpu_in_reset(adev)) {
2545 			/* in resume phase, if fail to enable ras,
2546 			 * clean up all ras fs nodes, and disable ras */
2547 			goto cleanup;
2548 		} else
2549 			return r;
2550 	}
2551 
2552 	/* check for errors on warm reset edc persisant supported ASIC */
2553 	amdgpu_persistent_edc_harvesting(adev, ras_block);
2554 
2555 	/* in resume phase, no need to create ras fs node */
2556 	if (adev->in_suspend || amdgpu_in_reset(adev))
2557 		return 0;
2558 
2559 	ras_obj = container_of(ras_block, struct amdgpu_ras_block_object, ras_comm);
2560 	if (ras_obj->ras_cb || (ras_obj->hw_ops &&
2561 	    (ras_obj->hw_ops->query_poison_status ||
2562 	    ras_obj->hw_ops->handle_poison_consumption))) {
2563 		r = amdgpu_ras_interrupt_add_handler(adev, ras_block);
2564 		if (r)
2565 			goto cleanup;
2566 	}
2567 
2568 	r = amdgpu_ras_sysfs_create(adev, ras_block);
2569 	if (r)
2570 		goto interrupt;
2571 
2572 	/* Those are the cached values at init.
2573 	 */
2574 	if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count) == 0) {
2575 		atomic_set(&con->ras_ce_count, ce_count);
2576 		atomic_set(&con->ras_ue_count, ue_count);
2577 	}
2578 
2579 	return 0;
2580 
2581 interrupt:
2582 	if (ras_obj->ras_cb)
2583 		amdgpu_ras_interrupt_remove_handler(adev, ras_block);
2584 cleanup:
2585 	amdgpu_ras_feature_enable(adev, ras_block, 0);
2586 	return r;
2587 }
2588 
2589 static int amdgpu_ras_block_late_init_default(struct amdgpu_device *adev,
2590 			 struct ras_common_if *ras_block)
2591 {
2592 	return amdgpu_ras_block_late_init(adev, ras_block);
2593 }
2594 
2595 /* helper function to remove ras fs node and interrupt handler */
2596 void amdgpu_ras_block_late_fini(struct amdgpu_device *adev,
2597 			  struct ras_common_if *ras_block)
2598 {
2599 	struct amdgpu_ras_block_object *ras_obj;
2600 	if (!ras_block)
2601 		return;
2602 
2603 	amdgpu_ras_sysfs_remove(adev, ras_block);
2604 
2605 	ras_obj = container_of(ras_block, struct amdgpu_ras_block_object, ras_comm);
2606 	if (ras_obj->ras_cb)
2607 		amdgpu_ras_interrupt_remove_handler(adev, ras_block);
2608 }
2609 
2610 static void amdgpu_ras_block_late_fini_default(struct amdgpu_device *adev,
2611 			  struct ras_common_if *ras_block)
2612 {
2613 	return amdgpu_ras_block_late_fini(adev, ras_block);
2614 }
2615 
2616 /* do some init work after IP late init as dependence.
2617  * and it runs in resume/gpu reset/booting up cases.
2618  */
2619 void amdgpu_ras_resume(struct amdgpu_device *adev)
2620 {
2621 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2622 	struct ras_manager *obj, *tmp;
2623 
2624 	if (!adev->ras_enabled || !con) {
2625 		/* clean ras context for VEGA20 Gaming after send ras disable cmd */
2626 		amdgpu_release_ras_context(adev);
2627 
2628 		return;
2629 	}
2630 
2631 	if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
2632 		/* Set up all other IPs which are not implemented. There is a
2633 		 * tricky thing that IP's actual ras error type should be
2634 		 * MULTI_UNCORRECTABLE, but as driver does not handle it, so
2635 		 * ERROR_NONE make sense anyway.
2636 		 */
2637 		amdgpu_ras_enable_all_features(adev, 1);
2638 
2639 		/* We enable ras on all hw_supported block, but as boot
2640 		 * parameter might disable some of them and one or more IP has
2641 		 * not implemented yet. So we disable them on behalf.
2642 		 */
2643 		list_for_each_entry_safe(obj, tmp, &con->head, node) {
2644 			if (!amdgpu_ras_is_supported(adev, obj->head.block)) {
2645 				amdgpu_ras_feature_enable(adev, &obj->head, 0);
2646 				/* there should be no any reference. */
2647 				WARN_ON(alive_obj(obj));
2648 			}
2649 		}
2650 	}
2651 }
2652 
2653 void amdgpu_ras_suspend(struct amdgpu_device *adev)
2654 {
2655 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2656 
2657 	if (!adev->ras_enabled || !con)
2658 		return;
2659 
2660 	amdgpu_ras_disable_all_features(adev, 0);
2661 	/* Make sure all ras objects are disabled. */
2662 	if (con->features)
2663 		amdgpu_ras_disable_all_features(adev, 1);
2664 }
2665 
2666 int amdgpu_ras_late_init(struct amdgpu_device *adev)
2667 {
2668 	struct amdgpu_ras_block_list *node, *tmp;
2669 	struct amdgpu_ras_block_object *obj;
2670 	int r;
2671 
2672 	/* Guest side doesn't need init ras feature */
2673 	if (amdgpu_sriov_vf(adev))
2674 		return 0;
2675 
2676 	list_for_each_entry_safe(node, tmp, &adev->ras_list, node) {
2677 		if (!node->ras_obj) {
2678 			dev_warn(adev->dev, "Warning: abnormal ras list node.\n");
2679 			continue;
2680 		}
2681 
2682 		obj = node->ras_obj;
2683 		if (obj->ras_late_init) {
2684 			r = obj->ras_late_init(adev, &obj->ras_comm);
2685 			if (r) {
2686 				dev_err(adev->dev, "%s failed to execute ras_late_init! ret:%d\n",
2687 					obj->ras_comm.name, r);
2688 				return r;
2689 			}
2690 		} else
2691 			amdgpu_ras_block_late_init_default(adev, &obj->ras_comm);
2692 	}
2693 
2694 	return 0;
2695 }
2696 
2697 /* do some fini work before IP fini as dependence */
2698 int amdgpu_ras_pre_fini(struct amdgpu_device *adev)
2699 {
2700 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2701 
2702 	if (!adev->ras_enabled || !con)
2703 		return 0;
2704 
2705 
2706 	/* Need disable ras on all IPs here before ip [hw/sw]fini */
2707 	amdgpu_ras_disable_all_features(adev, 0);
2708 	amdgpu_ras_recovery_fini(adev);
2709 	return 0;
2710 }
2711 
2712 int amdgpu_ras_fini(struct amdgpu_device *adev)
2713 {
2714 	struct amdgpu_ras_block_list *ras_node, *tmp;
2715 	struct amdgpu_ras_block_object *obj = NULL;
2716 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2717 
2718 	if (!adev->ras_enabled || !con)
2719 		return 0;
2720 
2721 	list_for_each_entry_safe(ras_node, tmp, &adev->ras_list, node) {
2722 		if (ras_node->ras_obj) {
2723 			obj = ras_node->ras_obj;
2724 			if (amdgpu_ras_is_supported(adev, obj->ras_comm.block) &&
2725 			    obj->ras_fini)
2726 				obj->ras_fini(adev, &obj->ras_comm);
2727 			else
2728 				amdgpu_ras_block_late_fini_default(adev, &obj->ras_comm);
2729 		}
2730 
2731 		/* Clear ras blocks from ras_list and free ras block list node */
2732 		list_del(&ras_node->node);
2733 		kfree(ras_node);
2734 	}
2735 
2736 	amdgpu_ras_fs_fini(adev);
2737 	amdgpu_ras_interrupt_remove_all(adev);
2738 
2739 	WARN(con->features, "Feature mask is not cleared");
2740 
2741 	if (con->features)
2742 		amdgpu_ras_disable_all_features(adev, 1);
2743 
2744 	cancel_delayed_work_sync(&con->ras_counte_delay_work);
2745 
2746 	amdgpu_ras_set_context(adev, NULL);
2747 	kfree(con);
2748 
2749 	return 0;
2750 }
2751 
2752 void amdgpu_ras_global_ras_isr(struct amdgpu_device *adev)
2753 {
2754 	amdgpu_ras_check_supported(adev);
2755 	if (!adev->ras_hw_enabled)
2756 		return;
2757 
2758 	if (atomic_cmpxchg(&amdgpu_ras_in_intr, 0, 1) == 0) {
2759 		dev_info(adev->dev, "uncorrectable hardware error"
2760 			"(ERREVENT_ATHUB_INTERRUPT) detected!\n");
2761 
2762 		amdgpu_ras_reset_gpu(adev);
2763 	}
2764 }
2765 
2766 bool amdgpu_ras_need_emergency_restart(struct amdgpu_device *adev)
2767 {
2768 	if (adev->asic_type == CHIP_VEGA20 &&
2769 	    adev->pm.fw_version <= 0x283400) {
2770 		return !(amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) &&
2771 				amdgpu_ras_intr_triggered();
2772 	}
2773 
2774 	return false;
2775 }
2776 
2777 void amdgpu_release_ras_context(struct amdgpu_device *adev)
2778 {
2779 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2780 
2781 	if (!con)
2782 		return;
2783 
2784 	if (!adev->ras_enabled && con->features & BIT(AMDGPU_RAS_BLOCK__GFX)) {
2785 		con->features &= ~BIT(AMDGPU_RAS_BLOCK__GFX);
2786 		amdgpu_ras_set_context(adev, NULL);
2787 		kfree(con);
2788 	}
2789 }
2790 
2791 #ifdef CONFIG_X86_MCE_AMD
2792 static struct amdgpu_device *find_adev(uint32_t node_id)
2793 {
2794 	int i;
2795 	struct amdgpu_device *adev = NULL;
2796 
2797 	for (i = 0; i < mce_adev_list.num_gpu; i++) {
2798 		adev = mce_adev_list.devs[i];
2799 
2800 		if (adev && adev->gmc.xgmi.connected_to_cpu &&
2801 		    adev->gmc.xgmi.physical_node_id == node_id)
2802 			break;
2803 		adev = NULL;
2804 	}
2805 
2806 	return adev;
2807 }
2808 
2809 #define GET_MCA_IPID_GPUID(m)	(((m) >> 44) & 0xF)
2810 #define GET_UMC_INST(m)		(((m) >> 21) & 0x7)
2811 #define GET_CHAN_INDEX(m)	((((m) >> 12) & 0x3) | (((m) >> 18) & 0x4))
2812 #define GPU_ID_OFFSET		8
2813 
2814 static int amdgpu_bad_page_notifier(struct notifier_block *nb,
2815 				    unsigned long val, void *data)
2816 {
2817 	struct mce *m = (struct mce *)data;
2818 	struct amdgpu_device *adev = NULL;
2819 	uint32_t gpu_id = 0;
2820 	uint32_t umc_inst = 0;
2821 	uint32_t ch_inst, channel_index = 0;
2822 	struct ras_err_data err_data = {0, 0, 0, NULL};
2823 	struct eeprom_table_record err_rec;
2824 	uint64_t retired_page;
2825 
2826 	/*
2827 	 * If the error was generated in UMC_V2, which belongs to GPU UMCs,
2828 	 * and error occurred in DramECC (Extended error code = 0) then only
2829 	 * process the error, else bail out.
2830 	 */
2831 	if (!m || !((smca_get_bank_type(m->extcpu, m->bank) == SMCA_UMC_V2) &&
2832 		    (XEC(m->status, 0x3f) == 0x0)))
2833 		return NOTIFY_DONE;
2834 
2835 	/*
2836 	 * If it is correctable error, return.
2837 	 */
2838 	if (mce_is_correctable(m))
2839 		return NOTIFY_OK;
2840 
2841 	/*
2842 	 * GPU Id is offset by GPU_ID_OFFSET in MCA_IPID_UMC register.
2843 	 */
2844 	gpu_id = GET_MCA_IPID_GPUID(m->ipid) - GPU_ID_OFFSET;
2845 
2846 	adev = find_adev(gpu_id);
2847 	if (!adev) {
2848 		DRM_WARN("%s: Unable to find adev for gpu_id: %d\n", __func__,
2849 								gpu_id);
2850 		return NOTIFY_DONE;
2851 	}
2852 
2853 	/*
2854 	 * If it is uncorrectable error, then find out UMC instance and
2855 	 * channel index.
2856 	 */
2857 	umc_inst = GET_UMC_INST(m->ipid);
2858 	ch_inst = GET_CHAN_INDEX(m->ipid);
2859 
2860 	dev_info(adev->dev, "Uncorrectable error detected in UMC inst: %d, chan_idx: %d",
2861 			     umc_inst, ch_inst);
2862 
2863 	/*
2864 	 * Translate UMC channel address to Physical address
2865 	 */
2866 	channel_index =
2867 		adev->umc.channel_idx_tbl[umc_inst * adev->umc.channel_inst_num
2868 					  + ch_inst];
2869 
2870 	retired_page = ADDR_OF_8KB_BLOCK(m->addr) |
2871 			ADDR_OF_256B_BLOCK(channel_index) |
2872 			OFFSET_IN_256B_BLOCK(m->addr);
2873 
2874 	memset(&err_rec, 0x0, sizeof(struct eeprom_table_record));
2875 	err_data.err_addr = &err_rec;
2876 	amdgpu_umc_fill_error_record(&err_data, m->addr,
2877 			retired_page, channel_index, umc_inst);
2878 
2879 	if (amdgpu_bad_page_threshold != 0) {
2880 		amdgpu_ras_add_bad_pages(adev, err_data.err_addr,
2881 						err_data.err_addr_cnt);
2882 		amdgpu_ras_save_bad_pages(adev);
2883 	}
2884 
2885 	return NOTIFY_OK;
2886 }
2887 
2888 static struct notifier_block amdgpu_bad_page_nb = {
2889 	.notifier_call  = amdgpu_bad_page_notifier,
2890 	.priority       = MCE_PRIO_UC,
2891 };
2892 
2893 static void amdgpu_register_bad_pages_mca_notifier(struct amdgpu_device *adev)
2894 {
2895 	/*
2896 	 * Add the adev to the mce_adev_list.
2897 	 * During mode2 reset, amdgpu device is temporarily
2898 	 * removed from the mgpu_info list which can cause
2899 	 * page retirement to fail.
2900 	 * Use this list instead of mgpu_info to find the amdgpu
2901 	 * device on which the UMC error was reported.
2902 	 */
2903 	mce_adev_list.devs[mce_adev_list.num_gpu++] = adev;
2904 
2905 	/*
2906 	 * Register the x86 notifier only once
2907 	 * with MCE subsystem.
2908 	 */
2909 	if (notifier_registered == false) {
2910 		mce_register_decode_chain(&amdgpu_bad_page_nb);
2911 		notifier_registered = true;
2912 	}
2913 }
2914 #endif
2915 
2916 struct amdgpu_ras *amdgpu_ras_get_context(struct amdgpu_device *adev)
2917 {
2918 	if (!adev)
2919 		return NULL;
2920 
2921 	return adev->psp.ras_context.ras;
2922 }
2923 
2924 int amdgpu_ras_set_context(struct amdgpu_device *adev, struct amdgpu_ras *ras_con)
2925 {
2926 	if (!adev)
2927 		return -EINVAL;
2928 
2929 	adev->psp.ras_context.ras = ras_con;
2930 	return 0;
2931 }
2932 
2933 /* check if ras is supported on block, say, sdma, gfx */
2934 int amdgpu_ras_is_supported(struct amdgpu_device *adev,
2935 		unsigned int block)
2936 {
2937 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
2938 
2939 	if (block >= AMDGPU_RAS_BLOCK_COUNT)
2940 		return 0;
2941 	return ras && (adev->ras_enabled & (1 << block));
2942 }
2943 
2944 int amdgpu_ras_reset_gpu(struct amdgpu_device *adev)
2945 {
2946 	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
2947 
2948 	if (atomic_cmpxchg(&ras->in_recovery, 0, 1) == 0)
2949 		schedule_work(&ras->recovery_work);
2950 	return 0;
2951 }
2952 
2953 
2954 /* Register each ip ras block into amdgpu ras */
2955 int amdgpu_ras_register_ras_block(struct amdgpu_device *adev,
2956 		struct amdgpu_ras_block_object *ras_block_obj)
2957 {
2958 	struct amdgpu_ras_block_list *ras_node;
2959 	if (!adev || !ras_block_obj)
2960 		return -EINVAL;
2961 
2962 	if (!amdgpu_ras_asic_supported(adev))
2963 		return 0;
2964 
2965 	ras_node = kzalloc(sizeof(*ras_node), GFP_KERNEL);
2966 	if (!ras_node)
2967 		return -ENOMEM;
2968 
2969 	INIT_LIST_HEAD(&ras_node->node);
2970 	ras_node->ras_obj = ras_block_obj;
2971 	list_add_tail(&ras_node->node, &adev->ras_list);
2972 
2973 	return 0;
2974 }
2975