Commit | Line | Data |
---|---|---|
c030f2e4 | 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> | |
f867723b | 27 | #include <linux/uaccess.h> |
7c6e68c7 AG |
28 | #include <linux/reboot.h> |
29 | #include <linux/syscalls.h> | |
05adfd80 | 30 | #include <linux/pm_runtime.h> |
f867723b | 31 | |
c030f2e4 | 32 | #include "amdgpu.h" |
33 | #include "amdgpu_ras.h" | |
b404ae82 | 34 | #include "amdgpu_atomfirmware.h" |
19744f5f | 35 | #include "amdgpu_xgmi.h" |
4e644fff | 36 | #include "ivsrcid/nbio/irqsrcs_nbif_7_4.h" |
f50160cf | 37 | #include "atom.h" |
25a2b22e AG |
38 | #include "amdgpu_reset.h" |
39 | ||
12b2cab7 MJ |
40 | #ifdef CONFIG_X86_MCE_AMD |
41 | #include <asm/mce.h> | |
c030f2e4 | 42 | |
12b2cab7 MJ |
43 | static bool notifier_registered; |
44 | #endif | |
eb0c3cd4 GC |
45 | static const char *RAS_FS_NAME = "ras"; |
46 | ||
c030f2e4 | 47 | const char *ras_error_string[] = { |
48 | "none", | |
49 | "parity", | |
50 | "single_correctable", | |
51 | "multi_uncorrectable", | |
52 | "poison", | |
53 | }; | |
54 | ||
55 | const char *ras_block_string[] = { | |
56 | "umc", | |
57 | "sdma", | |
58 | "gfx", | |
59 | "mmhub", | |
60 | "athub", | |
61 | "pcie_bif", | |
62 | "hdp", | |
63 | "xgmi_wafl", | |
64 | "df", | |
65 | "smn", | |
66 | "sem", | |
67 | "mp0", | |
68 | "mp1", | |
69 | "fuse", | |
640ae42e | 70 | "mca", |
a3d63c62 MZZ |
71 | "vcn", |
72 | "jpeg", | |
c030f2e4 | 73 | }; |
74 | ||
640ae42e JC |
75 | const char *ras_mca_block_string[] = { |
76 | "mca_mp0", | |
77 | "mca_mp1", | |
78 | "mca_mpio", | |
79 | "mca_iohc", | |
80 | }; | |
81 | ||
d5e8ff5f | 82 | struct amdgpu_ras_block_list { |
83 | /* ras block link */ | |
84 | struct list_head node; | |
85 | ||
86 | struct amdgpu_ras_block_object *ras_obj; | |
87 | }; | |
88 | ||
640ae42e JC |
89 | const char *get_ras_block_str(struct ras_common_if *ras_block) |
90 | { | |
91 | if (!ras_block) | |
92 | return "NULL"; | |
93 | ||
94 | if (ras_block->block >= AMDGPU_RAS_BLOCK_COUNT) | |
95 | return "OUT OF RANGE"; | |
96 | ||
97 | if (ras_block->block == AMDGPU_RAS_BLOCK__MCA) | |
98 | return ras_mca_block_string[ras_block->sub_block_index]; | |
99 | ||
100 | return ras_block_string[ras_block->block]; | |
101 | } | |
102 | ||
954ea6aa | 103 | #define ras_block_str(_BLOCK_) \ |
104 | (((_BLOCK_) < ARRAY_SIZE(ras_block_string)) ? ras_block_string[_BLOCK_] : "Out Of Range") | |
8b0fb0e9 | 105 | |
c030f2e4 | 106 | #define ras_err_str(i) (ras_error_string[ffs(i)]) |
c030f2e4 | 107 | |
108c6a63 | 108 | #define RAS_DEFAULT_FLAGS (AMDGPU_RAS_FLAG_INIT_BY_VBIOS) |
109 | ||
7cdc2ee3 TZ |
110 | /* inject address is 52 bits */ |
111 | #define RAS_UMC_INJECT_ADDR_LIMIT (0x1ULL << 52) | |
112 | ||
e4e6a589 LT |
113 | /* typical ECC bad page rate is 1 bad page per 100MB VRAM */ |
114 | #define RAS_BAD_PAGE_COVER (100 * 1024 * 1024ULL) | |
c84d4670 | 115 | |
52dd95f2 GC |
116 | enum amdgpu_ras_retire_page_reservation { |
117 | AMDGPU_RAS_RETIRE_PAGE_RESERVED, | |
118 | AMDGPU_RAS_RETIRE_PAGE_PENDING, | |
119 | AMDGPU_RAS_RETIRE_PAGE_FAULT, | |
120 | }; | |
7c6e68c7 AG |
121 | |
122 | atomic_t amdgpu_ras_in_intr = ATOMIC_INIT(0); | |
123 | ||
676deb38 DL |
124 | static bool amdgpu_ras_check_bad_page_unlock(struct amdgpu_ras *con, |
125 | uint64_t addr); | |
6e4be987 TZ |
126 | static bool amdgpu_ras_check_bad_page(struct amdgpu_device *adev, |
127 | uint64_t addr); | |
12b2cab7 | 128 | #ifdef CONFIG_X86_MCE_AMD |
91a1a52d MJ |
129 | static void amdgpu_register_bad_pages_mca_notifier(struct amdgpu_device *adev); |
130 | struct mce_notifier_adev_list { | |
131 | struct amdgpu_device *devs[MAX_GPU_INSTANCE]; | |
132 | int num_gpu; | |
133 | }; | |
134 | static struct mce_notifier_adev_list mce_adev_list; | |
12b2cab7 | 135 | #endif |
6e4be987 | 136 | |
61380faa JC |
137 | void amdgpu_ras_set_error_query_ready(struct amdgpu_device *adev, bool ready) |
138 | { | |
a9d82d2f | 139 | if (adev && amdgpu_ras_get_context(adev)) |
61380faa JC |
140 | amdgpu_ras_get_context(adev)->error_query_ready = ready; |
141 | } | |
142 | ||
f3167919 | 143 | static bool amdgpu_ras_get_error_query_ready(struct amdgpu_device *adev) |
61380faa | 144 | { |
a9d82d2f | 145 | if (adev && amdgpu_ras_get_context(adev)) |
61380faa JC |
146 | return amdgpu_ras_get_context(adev)->error_query_ready; |
147 | ||
148 | return false; | |
149 | } | |
150 | ||
cbb8f989 JC |
151 | static int amdgpu_reserve_page_direct(struct amdgpu_device *adev, uint64_t address) |
152 | { | |
153 | struct ras_err_data err_data = {0, 0, 0, NULL}; | |
154 | struct eeprom_table_record err_rec; | |
155 | ||
156 | if ((address >= adev->gmc.mc_vram_size) || | |
157 | (address >= RAS_UMC_INJECT_ADDR_LIMIT)) { | |
158 | dev_warn(adev->dev, | |
159 | "RAS WARN: input address 0x%llx is invalid.\n", | |
160 | address); | |
161 | return -EINVAL; | |
162 | } | |
163 | ||
164 | if (amdgpu_ras_check_bad_page(adev, address)) { | |
165 | dev_warn(adev->dev, | |
80b0cd0f | 166 | "RAS WARN: 0x%llx has already been marked as bad page!\n", |
cbb8f989 JC |
167 | address); |
168 | return 0; | |
169 | } | |
170 | ||
171 | memset(&err_rec, 0x0, sizeof(struct eeprom_table_record)); | |
cbb8f989 | 172 | err_data.err_addr = &err_rec; |
400013b2 TZ |
173 | amdgpu_umc_fill_error_record(&err_data, address, |
174 | (address >> AMDGPU_GPU_PAGE_SHIFT), 0, 0); | |
cbb8f989 JC |
175 | |
176 | if (amdgpu_bad_page_threshold != 0) { | |
177 | amdgpu_ras_add_bad_pages(adev, err_data.err_addr, | |
178 | err_data.err_addr_cnt); | |
179 | amdgpu_ras_save_bad_pages(adev); | |
180 | } | |
181 | ||
182 | dev_warn(adev->dev, "WARNING: THIS IS ONLY FOR TEST PURPOSES AND WILL CORRUPT RAS EEPROM\n"); | |
183 | dev_warn(adev->dev, "Clear EEPROM:\n"); | |
184 | dev_warn(adev->dev, " echo 1 > /sys/kernel/debug/dri/0/ras/ras_eeprom_reset\n"); | |
185 | ||
186 | return 0; | |
187 | } | |
188 | ||
c030f2e4 | 189 | static ssize_t amdgpu_ras_debugfs_read(struct file *f, char __user *buf, |
190 | size_t size, loff_t *pos) | |
191 | { | |
192 | struct ras_manager *obj = (struct ras_manager *)file_inode(f)->i_private; | |
193 | struct ras_query_if info = { | |
194 | .head = obj->head, | |
195 | }; | |
196 | ssize_t s; | |
197 | char val[128]; | |
198 | ||
761d86d3 | 199 | if (amdgpu_ras_query_error_status(obj->adev, &info)) |
c030f2e4 | 200 | return -EINVAL; |
201 | ||
2a460963 CL |
202 | /* Hardware counter will be reset automatically after the query on Vega20 and Arcturus */ |
203 | if (obj->adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 2) && | |
204 | obj->adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 4)) { | |
205 | if (amdgpu_ras_reset_error_status(obj->adev, info.head.block)) | |
206 | dev_warn(obj->adev->dev, "Failed to reset error counter and error status"); | |
207 | } | |
208 | ||
c030f2e4 | 209 | s = snprintf(val, sizeof(val), "%s: %lu\n%s: %lu\n", |
210 | "ue", info.ue_count, | |
211 | "ce", info.ce_count); | |
212 | if (*pos >= s) | |
213 | return 0; | |
214 | ||
215 | s -= *pos; | |
216 | s = min_t(u64, s, size); | |
217 | ||
218 | ||
219 | if (copy_to_user(buf, &val[*pos], s)) | |
220 | return -EINVAL; | |
221 | ||
222 | *pos += s; | |
223 | ||
224 | return s; | |
225 | } | |
226 | ||
c030f2e4 | 227 | static const struct file_operations amdgpu_ras_debugfs_ops = { |
228 | .owner = THIS_MODULE, | |
229 | .read = amdgpu_ras_debugfs_read, | |
190211ab | 230 | .write = NULL, |
c030f2e4 | 231 | .llseek = default_llseek |
232 | }; | |
233 | ||
96ebb307 | 234 | static int amdgpu_ras_find_block_id_by_name(const char *name, int *block_id) |
235 | { | |
236 | int i; | |
237 | ||
238 | for (i = 0; i < ARRAY_SIZE(ras_block_string); i++) { | |
239 | *block_id = i; | |
640ae42e | 240 | if (strcmp(name, ras_block_string[i]) == 0) |
96ebb307 | 241 | return 0; |
242 | } | |
243 | return -EINVAL; | |
244 | } | |
245 | ||
246 | static int amdgpu_ras_debugfs_ctrl_parse_data(struct file *f, | |
247 | const char __user *buf, size_t size, | |
248 | loff_t *pos, struct ras_debug_if *data) | |
249 | { | |
250 | ssize_t s = min_t(u64, 64, size); | |
251 | char str[65]; | |
252 | char block_name[33]; | |
253 | char err[9] = "ue"; | |
254 | int op = -1; | |
255 | int block_id; | |
44494f96 | 256 | uint32_t sub_block; |
96ebb307 | 257 | u64 address, value; |
258 | ||
259 | if (*pos) | |
260 | return -EINVAL; | |
261 | *pos = size; | |
262 | ||
263 | memset(str, 0, sizeof(str)); | |
264 | memset(data, 0, sizeof(*data)); | |
265 | ||
266 | if (copy_from_user(str, buf, s)) | |
267 | return -EINVAL; | |
268 | ||
269 | if (sscanf(str, "disable %32s", block_name) == 1) | |
270 | op = 0; | |
271 | else if (sscanf(str, "enable %32s %8s", block_name, err) == 2) | |
272 | op = 1; | |
273 | else if (sscanf(str, "inject %32s %8s", block_name, err) == 2) | |
274 | op = 2; | |
6df23f4c | 275 | else if (strstr(str, "retire_page") != NULL) |
cbb8f989 | 276 | op = 3; |
b076296b | 277 | else if (str[0] && str[1] && str[2] && str[3]) |
96ebb307 | 278 | /* ascii string, but commands are not matched. */ |
279 | return -EINVAL; | |
280 | ||
281 | if (op != -1) { | |
cbb8f989 | 282 | if (op == 3) { |
546aa546 LT |
283 | if (sscanf(str, "%*s 0x%llx", &address) != 1 && |
284 | sscanf(str, "%*s %llu", &address) != 1) | |
6cb7a1d4 | 285 | return -EINVAL; |
cbb8f989 JC |
286 | |
287 | data->op = op; | |
288 | data->inject.address = address; | |
289 | ||
290 | return 0; | |
291 | } | |
292 | ||
96ebb307 | 293 | if (amdgpu_ras_find_block_id_by_name(block_name, &block_id)) |
294 | return -EINVAL; | |
295 | ||
296 | data->head.block = block_id; | |
e1063493 TZ |
297 | /* only ue and ce errors are supported */ |
298 | if (!memcmp("ue", err, 2)) | |
299 | data->head.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE; | |
300 | else if (!memcmp("ce", err, 2)) | |
301 | data->head.type = AMDGPU_RAS_ERROR__SINGLE_CORRECTABLE; | |
302 | else | |
303 | return -EINVAL; | |
304 | ||
96ebb307 | 305 | data->op = op; |
306 | ||
307 | if (op == 2) { | |
546aa546 LT |
308 | if (sscanf(str, "%*s %*s %*s 0x%x 0x%llx 0x%llx", |
309 | &sub_block, &address, &value) != 3 && | |
310 | sscanf(str, "%*s %*s %*s %u %llu %llu", | |
6cb7a1d4 LT |
311 | &sub_block, &address, &value) != 3) |
312 | return -EINVAL; | |
44494f96 | 313 | data->head.sub_block_index = sub_block; |
96ebb307 | 314 | data->inject.address = address; |
315 | data->inject.value = value; | |
316 | } | |
317 | } else { | |
73aa8e1a | 318 | if (size < sizeof(*data)) |
96ebb307 | 319 | return -EINVAL; |
320 | ||
321 | if (copy_from_user(data, buf, sizeof(*data))) | |
322 | return -EINVAL; | |
323 | } | |
324 | ||
325 | return 0; | |
326 | } | |
7c6e68c7 | 327 | |
74abc221 TSD |
328 | /** |
329 | * DOC: AMDGPU RAS debugfs control interface | |
36ea1bd2 | 330 | * |
737c375b | 331 | * The control interface accepts struct ras_debug_if which has two members. |
36ea1bd2 | 332 | * |
333 | * First member: ras_debug_if::head or ras_debug_if::inject. | |
96ebb307 | 334 | * |
335 | * head is used to indicate which IP block will be under control. | |
36ea1bd2 | 336 | * |
337 | * head has four members, they are block, type, sub_block_index, name. | |
338 | * block: which IP will be under control. | |
339 | * type: what kind of error will be enabled/disabled/injected. | |
340 | * sub_block_index: some IPs have subcomponets. say, GFX, sDMA. | |
341 | * name: the name of IP. | |
342 | * | |
343 | * inject has two more members than head, they are address, value. | |
344 | * As their names indicate, inject operation will write the | |
345 | * value to the address. | |
346 | * | |
ef177d11 | 347 | * The second member: struct ras_debug_if::op. |
c688a06b | 348 | * It has three kinds of operations. |
879e723d AZ |
349 | * |
350 | * - 0: disable RAS on the block. Take ::head as its data. | |
351 | * - 1: enable RAS on the block. Take ::head as its data. | |
352 | * - 2: inject errors on the block. Take ::inject as its data. | |
36ea1bd2 | 353 | * |
96ebb307 | 354 | * How to use the interface? |
ef177d11 | 355 | * |
737c375b | 356 | * In a program |
ef177d11 | 357 | * |
737c375b LT |
358 | * Copy the struct ras_debug_if in your code and initialize it. |
359 | * Write the struct to the control interface. | |
ef177d11 | 360 | * |
737c375b | 361 | * From shell |
96ebb307 | 362 | * |
879e723d AZ |
363 | * .. code-block:: bash |
364 | * | |
737c375b LT |
365 | * echo "disable <block>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl |
366 | * echo "enable <block> <error>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl | |
367 | * echo "inject <block> <error> <sub-block> <address> <value> > /sys/kernel/debug/dri/<N>/ras/ras_ctrl | |
879e723d | 368 | * |
737c375b | 369 | * Where N, is the card which you want to affect. |
ef177d11 | 370 | * |
737c375b LT |
371 | * "disable" requires only the block. |
372 | * "enable" requires the block and error type. | |
373 | * "inject" requires the block, error type, address, and value. | |
c666bbf0 | 374 | * |
737c375b | 375 | * The block is one of: umc, sdma, gfx, etc. |
879e723d | 376 | * see ras_block_string[] for details |
c666bbf0 | 377 | * |
737c375b LT |
378 | * The error type is one of: ue, ce, where, |
379 | * ue is multi-uncorrectable | |
380 | * ce is single-correctable | |
c666bbf0 | 381 | * |
737c375b LT |
382 | * The sub-block is a the sub-block index, pass 0 if there is no sub-block. |
383 | * The address and value are hexadecimal numbers, leading 0x is optional. | |
879e723d | 384 | * |
737c375b | 385 | * For instance, |
879e723d AZ |
386 | * |
387 | * .. code-block:: bash | |
96ebb307 | 388 | * |
44494f96 TZ |
389 | * echo inject umc ue 0x0 0x0 0x0 > /sys/kernel/debug/dri/0/ras/ras_ctrl |
390 | * echo inject umc ce 0 0 0 > /sys/kernel/debug/dri/0/ras/ras_ctrl | |
96ebb307 | 391 | * echo disable umc > /sys/kernel/debug/dri/0/ras/ras_ctrl |
392 | * | |
737c375b | 393 | * How to check the result of the operation? |
36ea1bd2 | 394 | * |
737c375b | 395 | * To check disable/enable, see "ras" features at, |
36ea1bd2 | 396 | * /sys/class/drm/card[0/1/2...]/device/ras/features |
397 | * | |
737c375b LT |
398 | * To check inject, see the corresponding error count at, |
399 | * /sys/class/drm/card[0/1/2...]/device/ras/[gfx|sdma|umc|...]_err_count | |
36ea1bd2 | 400 | * |
879e723d | 401 | * .. note:: |
ef177d11 | 402 | * Operations are only allowed on blocks which are supported. |
737c375b | 403 | * Check the "ras" mask at /sys/module/amdgpu/parameters/ras_mask |
ef177d11 AD |
404 | * to see which blocks support RAS on a particular asic. |
405 | * | |
36ea1bd2 | 406 | */ |
cf696091 LT |
407 | static ssize_t amdgpu_ras_debugfs_ctrl_write(struct file *f, |
408 | const char __user *buf, | |
409 | size_t size, loff_t *pos) | |
36ea1bd2 | 410 | { |
411 | struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private; | |
412 | struct ras_debug_if data; | |
413 | int ret = 0; | |
414 | ||
61380faa | 415 | if (!amdgpu_ras_get_error_query_ready(adev)) { |
6952e99c GC |
416 | dev_warn(adev->dev, "RAS WARN: error injection " |
417 | "currently inaccessible\n"); | |
43c4d576 JC |
418 | return size; |
419 | } | |
420 | ||
96ebb307 | 421 | ret = amdgpu_ras_debugfs_ctrl_parse_data(f, buf, size, pos, &data); |
422 | if (ret) | |
cf696091 | 423 | return ret; |
36ea1bd2 | 424 | |
80b0cd0f | 425 | if (data.op == 3) { |
cbb8f989 | 426 | ret = amdgpu_reserve_page_direct(adev, data.inject.address); |
80b0cd0f | 427 | if (!ret) |
cbb8f989 JC |
428 | return size; |
429 | else | |
430 | return ret; | |
431 | } | |
432 | ||
36ea1bd2 | 433 | if (!amdgpu_ras_is_supported(adev, data.head.block)) |
434 | return -EINVAL; | |
435 | ||
436 | switch (data.op) { | |
437 | case 0: | |
438 | ret = amdgpu_ras_feature_enable(adev, &data.head, 0); | |
439 | break; | |
440 | case 1: | |
441 | ret = amdgpu_ras_feature_enable(adev, &data.head, 1); | |
442 | break; | |
443 | case 2: | |
7cdc2ee3 TZ |
444 | if ((data.inject.address >= adev->gmc.mc_vram_size) || |
445 | (data.inject.address >= RAS_UMC_INJECT_ADDR_LIMIT)) { | |
b0d4783a GC |
446 | dev_warn(adev->dev, "RAS WARN: input address " |
447 | "0x%llx is invalid.", | |
448 | data.inject.address); | |
7cdc2ee3 TZ |
449 | ret = -EINVAL; |
450 | break; | |
451 | } | |
452 | ||
6e4be987 TZ |
453 | /* umc ce/ue error injection for a bad page is not allowed */ |
454 | if ((data.head.block == AMDGPU_RAS_BLOCK__UMC) && | |
455 | amdgpu_ras_check_bad_page(adev, data.inject.address)) { | |
c65b0805 LT |
456 | dev_warn(adev->dev, "RAS WARN: inject: 0x%llx has " |
457 | "already been marked as bad!\n", | |
458 | data.inject.address); | |
6e4be987 TZ |
459 | break; |
460 | } | |
461 | ||
7cdc2ee3 | 462 | /* data.inject.address is offset instead of absolute gpu address */ |
36ea1bd2 | 463 | ret = amdgpu_ras_error_inject(adev, &data.inject); |
464 | break; | |
96ebb307 | 465 | default: |
466 | ret = -EINVAL; | |
467 | break; | |
374bf7bd | 468 | } |
36ea1bd2 | 469 | |
470 | if (ret) | |
79c04621 | 471 | return ret; |
36ea1bd2 | 472 | |
473 | return size; | |
474 | } | |
475 | ||
084fe13b AG |
476 | /** |
477 | * DOC: AMDGPU RAS debugfs EEPROM table reset interface | |
478 | * | |
f77c7109 | 479 | * Some boards contain an EEPROM which is used to persistently store a list of |
ef177d11 | 480 | * bad pages which experiences ECC errors in vram. This interface provides |
f77c7109 AD |
481 | * a way to reset the EEPROM, e.g., after testing error injection. |
482 | * | |
483 | * Usage: | |
484 | * | |
485 | * .. code-block:: bash | |
486 | * | |
487 | * echo 1 > ../ras/ras_eeprom_reset | |
488 | * | |
489 | * will reset EEPROM table to 0 entries. | |
490 | * | |
084fe13b | 491 | */ |
cf696091 LT |
492 | static ssize_t amdgpu_ras_debugfs_eeprom_write(struct file *f, |
493 | const char __user *buf, | |
494 | size_t size, loff_t *pos) | |
084fe13b | 495 | { |
bf0b91b7 GC |
496 | struct amdgpu_device *adev = |
497 | (struct amdgpu_device *)file_inode(f)->i_private; | |
084fe13b AG |
498 | int ret; |
499 | ||
bf0b91b7 | 500 | ret = amdgpu_ras_eeprom_reset_table( |
cf696091 | 501 | &(amdgpu_ras_get_context(adev)->eeprom_control)); |
084fe13b | 502 | |
63d4c081 | 503 | if (!ret) { |
cf696091 LT |
504 | /* Something was written to EEPROM. |
505 | */ | |
bf0b91b7 GC |
506 | amdgpu_ras_get_context(adev)->flags = RAS_DEFAULT_FLAGS; |
507 | return size; | |
508 | } else { | |
cf696091 | 509 | return ret; |
bf0b91b7 | 510 | } |
084fe13b AG |
511 | } |
512 | ||
36ea1bd2 | 513 | static const struct file_operations amdgpu_ras_debugfs_ctrl_ops = { |
514 | .owner = THIS_MODULE, | |
515 | .read = NULL, | |
516 | .write = amdgpu_ras_debugfs_ctrl_write, | |
517 | .llseek = default_llseek | |
518 | }; | |
519 | ||
084fe13b AG |
520 | static const struct file_operations amdgpu_ras_debugfs_eeprom_ops = { |
521 | .owner = THIS_MODULE, | |
522 | .read = NULL, | |
523 | .write = amdgpu_ras_debugfs_eeprom_write, | |
524 | .llseek = default_llseek | |
525 | }; | |
526 | ||
f77c7109 AD |
527 | /** |
528 | * DOC: AMDGPU RAS sysfs Error Count Interface | |
529 | * | |
ef177d11 | 530 | * It allows the user to read the error count for each IP block on the gpu through |
f77c7109 AD |
531 | * /sys/class/drm/card[0/1/2...]/device/ras/[gfx/sdma/...]_err_count |
532 | * | |
533 | * It outputs the multiple lines which report the uncorrected (ue) and corrected | |
534 | * (ce) error counts. | |
535 | * | |
536 | * The format of one line is below, | |
537 | * | |
538 | * [ce|ue]: count | |
539 | * | |
540 | * Example: | |
541 | * | |
542 | * .. code-block:: bash | |
543 | * | |
544 | * ue: 0 | |
545 | * ce: 1 | |
546 | * | |
547 | */ | |
c030f2e4 | 548 | static ssize_t amdgpu_ras_sysfs_read(struct device *dev, |
549 | struct device_attribute *attr, char *buf) | |
550 | { | |
551 | struct ras_manager *obj = container_of(attr, struct ras_manager, sysfs_attr); | |
552 | struct ras_query_if info = { | |
553 | .head = obj->head, | |
554 | }; | |
555 | ||
61380faa | 556 | if (!amdgpu_ras_get_error_query_ready(obj->adev)) |
36000c7a | 557 | return sysfs_emit(buf, "Query currently inaccessible\n"); |
43c4d576 | 558 | |
761d86d3 | 559 | if (amdgpu_ras_query_error_status(obj->adev, &info)) |
c030f2e4 | 560 | return -EINVAL; |
561 | ||
2a460963 CL |
562 | if (obj->adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 2) && |
563 | obj->adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 4)) { | |
1f0d8e37 | 564 | if (amdgpu_ras_reset_error_status(obj->adev, info.head.block)) |
2a460963 | 565 | dev_warn(obj->adev->dev, "Failed to reset error counter and error status"); |
1f0d8e37 MJ |
566 | } |
567 | ||
36000c7a TT |
568 | return sysfs_emit(buf, "%s: %lu\n%s: %lu\n", "ue", info.ue_count, |
569 | "ce", info.ce_count); | |
c030f2e4 | 570 | } |
571 | ||
572 | /* obj begin */ | |
573 | ||
574 | #define get_obj(obj) do { (obj)->use++; } while (0) | |
575 | #define alive_obj(obj) ((obj)->use) | |
576 | ||
577 | static inline void put_obj(struct ras_manager *obj) | |
578 | { | |
f0872686 | 579 | if (obj && (--obj->use == 0)) |
c030f2e4 | 580 | list_del(&obj->node); |
f0872686 | 581 | if (obj && (obj->use < 0)) |
640ae42e | 582 | DRM_ERROR("RAS ERROR: Unbalance obj(%s) use\n", get_ras_block_str(&obj->head)); |
c030f2e4 | 583 | } |
584 | ||
585 | /* make one obj and return it. */ | |
586 | static struct ras_manager *amdgpu_ras_create_obj(struct amdgpu_device *adev, | |
587 | struct ras_common_if *head) | |
588 | { | |
589 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
590 | struct ras_manager *obj; | |
591 | ||
8ab0d6f0 | 592 | if (!adev->ras_enabled || !con) |
c030f2e4 | 593 | return NULL; |
594 | ||
595 | if (head->block >= AMDGPU_RAS_BLOCK_COUNT) | |
596 | return NULL; | |
597 | ||
640ae42e JC |
598 | if (head->block == AMDGPU_RAS_BLOCK__MCA) { |
599 | if (head->sub_block_index >= AMDGPU_RAS_MCA_BLOCK__LAST) | |
600 | return NULL; | |
601 | ||
602 | obj = &con->objs[AMDGPU_RAS_BLOCK__LAST + head->sub_block_index]; | |
603 | } else | |
604 | obj = &con->objs[head->block]; | |
605 | ||
c030f2e4 | 606 | /* already exist. return obj? */ |
607 | if (alive_obj(obj)) | |
608 | return NULL; | |
609 | ||
610 | obj->head = *head; | |
611 | obj->adev = adev; | |
612 | list_add(&obj->node, &con->head); | |
613 | get_obj(obj); | |
614 | ||
615 | return obj; | |
616 | } | |
617 | ||
618 | /* return an obj equal to head, or the first when head is NULL */ | |
f2a79be1 | 619 | struct ras_manager *amdgpu_ras_find_obj(struct amdgpu_device *adev, |
c030f2e4 | 620 | struct ras_common_if *head) |
621 | { | |
622 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
623 | struct ras_manager *obj; | |
624 | int i; | |
625 | ||
8ab0d6f0 | 626 | if (!adev->ras_enabled || !con) |
c030f2e4 | 627 | return NULL; |
628 | ||
629 | if (head) { | |
630 | if (head->block >= AMDGPU_RAS_BLOCK_COUNT) | |
631 | return NULL; | |
632 | ||
640ae42e JC |
633 | if (head->block == AMDGPU_RAS_BLOCK__MCA) { |
634 | if (head->sub_block_index >= AMDGPU_RAS_MCA_BLOCK__LAST) | |
635 | return NULL; | |
636 | ||
637 | obj = &con->objs[AMDGPU_RAS_BLOCK__LAST + head->sub_block_index]; | |
638 | } else | |
639 | obj = &con->objs[head->block]; | |
c030f2e4 | 640 | |
640ae42e | 641 | if (alive_obj(obj)) |
c030f2e4 | 642 | return obj; |
c030f2e4 | 643 | } else { |
640ae42e | 644 | for (i = 0; i < AMDGPU_RAS_BLOCK_COUNT + AMDGPU_RAS_MCA_BLOCK_COUNT; i++) { |
c030f2e4 | 645 | obj = &con->objs[i]; |
640ae42e | 646 | if (alive_obj(obj)) |
c030f2e4 | 647 | return obj; |
c030f2e4 | 648 | } |
649 | } | |
650 | ||
651 | return NULL; | |
652 | } | |
653 | /* obj end */ | |
654 | ||
655 | /* feature ctl begin */ | |
656 | static int amdgpu_ras_is_feature_allowed(struct amdgpu_device *adev, | |
e509965e | 657 | struct ras_common_if *head) |
c030f2e4 | 658 | { |
8ab0d6f0 | 659 | return adev->ras_hw_enabled & BIT(head->block); |
c030f2e4 | 660 | } |
661 | ||
662 | static int amdgpu_ras_is_feature_enabled(struct amdgpu_device *adev, | |
663 | struct ras_common_if *head) | |
664 | { | |
665 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
666 | ||
667 | return con->features & BIT(head->block); | |
668 | } | |
669 | ||
670 | /* | |
671 | * if obj is not created, then create one. | |
672 | * set feature enable flag. | |
673 | */ | |
674 | static int __amdgpu_ras_feature_enable(struct amdgpu_device *adev, | |
675 | struct ras_common_if *head, int enable) | |
676 | { | |
677 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
678 | struct ras_manager *obj = amdgpu_ras_find_obj(adev, head); | |
679 | ||
5caf466a | 680 | /* If hardware does not support ras, then do not create obj. |
681 | * But if hardware support ras, we can create the obj. | |
682 | * Ras framework checks con->hw_supported to see if it need do | |
683 | * corresponding initialization. | |
684 | * IP checks con->support to see if it need disable ras. | |
685 | */ | |
c030f2e4 | 686 | if (!amdgpu_ras_is_feature_allowed(adev, head)) |
687 | return 0; | |
c030f2e4 | 688 | |
689 | if (enable) { | |
690 | if (!obj) { | |
691 | obj = amdgpu_ras_create_obj(adev, head); | |
692 | if (!obj) | |
693 | return -EINVAL; | |
694 | } else { | |
695 | /* In case we create obj somewhere else */ | |
696 | get_obj(obj); | |
697 | } | |
698 | con->features |= BIT(head->block); | |
699 | } else { | |
700 | if (obj && amdgpu_ras_is_feature_enabled(adev, head)) { | |
19d0dfda | 701 | con->features &= ~BIT(head->block); |
c030f2e4 | 702 | put_obj(obj); |
703 | } | |
704 | } | |
705 | ||
706 | return 0; | |
707 | } | |
708 | ||
c26cd999 SY |
709 | static int amdgpu_ras_check_feature_allowed(struct amdgpu_device *adev, |
710 | struct ras_common_if *head) | |
711 | { | |
712 | if (amdgpu_ras_is_feature_allowed(adev, head) || | |
713 | amdgpu_ras_is_poison_mode_supported(adev)) | |
714 | return 1; | |
715 | else | |
716 | return 0; | |
717 | } | |
718 | ||
c030f2e4 | 719 | /* wrapper of psp_ras_enable_features */ |
720 | int amdgpu_ras_feature_enable(struct amdgpu_device *adev, | |
721 | struct ras_common_if *head, bool enable) | |
722 | { | |
723 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
7fcffecf | 724 | union ta_ras_cmd_input *info; |
c26cd999 | 725 | int ret = 0; |
c030f2e4 | 726 | |
727 | if (!con) | |
728 | return -EINVAL; | |
729 | ||
26093ce1 SY |
730 | if (head->block == AMDGPU_RAS_BLOCK__GFX) { |
731 | info = kzalloc(sizeof(union ta_ras_cmd_input), GFP_KERNEL); | |
732 | if (!info) | |
733 | return -ENOMEM; | |
734 | ||
735 | if (!enable) { | |
736 | info->disable_features = (struct ta_ras_disable_features_input) { | |
737 | .block_id = amdgpu_ras_block_to_ta(head->block), | |
738 | .error_type = amdgpu_ras_error_to_ta(head->type), | |
739 | }; | |
740 | } else { | |
741 | info->enable_features = (struct ta_ras_enable_features_input) { | |
742 | .block_id = amdgpu_ras_block_to_ta(head->block), | |
743 | .error_type = amdgpu_ras_error_to_ta(head->type), | |
744 | }; | |
745 | } | |
c030f2e4 | 746 | } |
747 | ||
748 | /* Do not enable if it is not allowed. */ | |
c26cd999 SY |
749 | if (enable && !amdgpu_ras_check_feature_allowed(adev, head)) |
750 | goto out; | |
c030f2e4 | 751 | |
950d6425 | 752 | /* Only enable ras feature operation handle on host side */ |
26093ce1 SY |
753 | if (head->block == AMDGPU_RAS_BLOCK__GFX && |
754 | !amdgpu_sriov_vf(adev) && | |
950d6425 | 755 | !amdgpu_ras_intr_triggered()) { |
7fcffecf | 756 | ret = psp_ras_enable_features(&adev->psp, info, enable); |
bff77e86 | 757 | if (ret) { |
e4348849 | 758 | dev_err(adev->dev, "ras %s %s failed poison:%d ret:%d\n", |
011907fd | 759 | enable ? "enable":"disable", |
640ae42e | 760 | get_ras_block_str(head), |
e4348849 | 761 | amdgpu_ras_is_poison_mode_supported(adev), ret); |
7fcffecf | 762 | goto out; |
bff77e86 | 763 | } |
c030f2e4 | 764 | } |
765 | ||
766 | /* setup the obj */ | |
767 | __amdgpu_ras_feature_enable(adev, head, enable); | |
7fcffecf | 768 | out: |
26093ce1 SY |
769 | if (head->block == AMDGPU_RAS_BLOCK__GFX) |
770 | kfree(info); | |
7fcffecf | 771 | return ret; |
c030f2e4 | 772 | } |
773 | ||
77de502b | 774 | /* Only used in device probe stage and called only once. */ |
775 | int amdgpu_ras_feature_enable_on_boot(struct amdgpu_device *adev, | |
776 | struct ras_common_if *head, bool enable) | |
777 | { | |
778 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
779 | int ret; | |
780 | ||
781 | if (!con) | |
782 | return -EINVAL; | |
783 | ||
784 | if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) { | |
7af23ebe | 785 | if (enable) { |
786 | /* There is no harm to issue a ras TA cmd regardless of | |
787 | * the currecnt ras state. | |
788 | * If current state == target state, it will do nothing | |
789 | * But sometimes it requests driver to reset and repost | |
790 | * with error code -EAGAIN. | |
791 | */ | |
792 | ret = amdgpu_ras_feature_enable(adev, head, 1); | |
793 | /* With old ras TA, we might fail to enable ras. | |
794 | * Log it and just setup the object. | |
795 | * TODO need remove this WA in the future. | |
796 | */ | |
797 | if (ret == -EINVAL) { | |
798 | ret = __amdgpu_ras_feature_enable(adev, head, 1); | |
799 | if (!ret) | |
6952e99c GC |
800 | dev_info(adev->dev, |
801 | "RAS INFO: %s setup object\n", | |
640ae42e | 802 | get_ras_block_str(head)); |
7af23ebe | 803 | } |
804 | } else { | |
805 | /* setup the object then issue a ras TA disable cmd.*/ | |
806 | ret = __amdgpu_ras_feature_enable(adev, head, 1); | |
807 | if (ret) | |
808 | return ret; | |
77de502b | 809 | |
970fd197 SY |
810 | /* gfx block ras dsiable cmd must send to ras-ta */ |
811 | if (head->block == AMDGPU_RAS_BLOCK__GFX) | |
812 | con->features |= BIT(head->block); | |
813 | ||
77de502b | 814 | ret = amdgpu_ras_feature_enable(adev, head, 0); |
19d0dfda SY |
815 | |
816 | /* clean gfx block ras features flag */ | |
8ab0d6f0 | 817 | if (adev->ras_enabled && head->block == AMDGPU_RAS_BLOCK__GFX) |
19d0dfda | 818 | con->features &= ~BIT(head->block); |
7af23ebe | 819 | } |
77de502b | 820 | } else |
821 | ret = amdgpu_ras_feature_enable(adev, head, enable); | |
822 | ||
823 | return ret; | |
824 | } | |
825 | ||
c030f2e4 | 826 | static int amdgpu_ras_disable_all_features(struct amdgpu_device *adev, |
827 | bool bypass) | |
828 | { | |
829 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
830 | struct ras_manager *obj, *tmp; | |
831 | ||
832 | list_for_each_entry_safe(obj, tmp, &con->head, node) { | |
833 | /* bypass psp. | |
834 | * aka just release the obj and corresponding flags | |
835 | */ | |
836 | if (bypass) { | |
837 | if (__amdgpu_ras_feature_enable(adev, &obj->head, 0)) | |
838 | break; | |
839 | } else { | |
840 | if (amdgpu_ras_feature_enable(adev, &obj->head, 0)) | |
841 | break; | |
842 | } | |
289d513b | 843 | } |
c030f2e4 | 844 | |
845 | return con->features; | |
846 | } | |
847 | ||
848 | static int amdgpu_ras_enable_all_features(struct amdgpu_device *adev, | |
849 | bool bypass) | |
850 | { | |
851 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
c030f2e4 | 852 | int i; |
640ae42e | 853 | const enum amdgpu_ras_error_type default_ras_type = AMDGPU_RAS_ERROR__NONE; |
c030f2e4 | 854 | |
640ae42e | 855 | for (i = 0; i < AMDGPU_RAS_BLOCK_COUNT; i++) { |
c030f2e4 | 856 | struct ras_common_if head = { |
857 | .block = i, | |
191051a1 | 858 | .type = default_ras_type, |
c030f2e4 | 859 | .sub_block_index = 0, |
860 | }; | |
640ae42e JC |
861 | |
862 | if (i == AMDGPU_RAS_BLOCK__MCA) | |
863 | continue; | |
864 | ||
865 | if (bypass) { | |
866 | /* | |
867 | * bypass psp. vbios enable ras for us. | |
868 | * so just create the obj | |
869 | */ | |
870 | if (__amdgpu_ras_feature_enable(adev, &head, 1)) | |
871 | break; | |
872 | } else { | |
873 | if (amdgpu_ras_feature_enable(adev, &head, 1)) | |
874 | break; | |
875 | } | |
876 | } | |
877 | ||
878 | for (i = 0; i < AMDGPU_RAS_MCA_BLOCK_COUNT; i++) { | |
879 | struct ras_common_if head = { | |
880 | .block = AMDGPU_RAS_BLOCK__MCA, | |
881 | .type = default_ras_type, | |
882 | .sub_block_index = i, | |
883 | }; | |
884 | ||
c030f2e4 | 885 | if (bypass) { |
886 | /* | |
887 | * bypass psp. vbios enable ras for us. | |
888 | * so just create the obj | |
889 | */ | |
890 | if (__amdgpu_ras_feature_enable(adev, &head, 1)) | |
891 | break; | |
892 | } else { | |
893 | if (amdgpu_ras_feature_enable(adev, &head, 1)) | |
894 | break; | |
895 | } | |
289d513b | 896 | } |
c030f2e4 | 897 | |
898 | return con->features; | |
899 | } | |
900 | /* feature ctl end */ | |
901 | ||
e3d833f4 | 902 | static int amdgpu_ras_block_match_default(struct amdgpu_ras_block_object *block_obj, |
903 | enum amdgpu_ras_block block) | |
6492e1b0 | 904 | { |
b6efdb02 | 905 | if (!block_obj) |
6492e1b0 | 906 | return -EINVAL; |
907 | ||
bdb3489c | 908 | if (block_obj->ras_comm.block == block) |
6492e1b0 | 909 | return 0; |
640ae42e | 910 | |
6492e1b0 | 911 | return -EINVAL; |
912 | } | |
913 | ||
b6efdb02 | 914 | static struct amdgpu_ras_block_object *amdgpu_ras_get_ras_block(struct amdgpu_device *adev, |
6492e1b0 | 915 | enum amdgpu_ras_block block, uint32_t sub_block_index) |
640ae42e | 916 | { |
d5e8ff5f | 917 | struct amdgpu_ras_block_list *node, *tmp; |
918 | struct amdgpu_ras_block_object *obj; | |
6492e1b0 | 919 | |
920 | if (block >= AMDGPU_RAS_BLOCK__LAST) | |
921 | return NULL; | |
922 | ||
923 | if (!amdgpu_ras_is_supported(adev, block)) | |
924 | return NULL; | |
925 | ||
d5e8ff5f | 926 | list_for_each_entry_safe(node, tmp, &adev->ras_list, node) { |
927 | if (!node->ras_obj) { | |
928 | dev_warn(adev->dev, "Warning: abnormal ras list node.\n"); | |
929 | continue; | |
930 | } | |
931 | ||
932 | obj = node->ras_obj; | |
6492e1b0 | 933 | if (obj->ras_block_match) { |
934 | if (obj->ras_block_match(obj, block, sub_block_index) == 0) | |
935 | return obj; | |
936 | } else { | |
937 | if (amdgpu_ras_block_match_default(obj, block) == 0) | |
938 | return obj; | |
939 | } | |
640ae42e | 940 | } |
6492e1b0 | 941 | |
942 | return NULL; | |
640ae42e JC |
943 | } |
944 | ||
fdcb279d SY |
945 | static void amdgpu_ras_get_ecc_info(struct amdgpu_device *adev, struct ras_err_data *err_data) |
946 | { | |
947 | struct amdgpu_ras *ras = amdgpu_ras_get_context(adev); | |
948 | int ret = 0; | |
949 | ||
950 | /* | |
951 | * choosing right query method according to | |
952 | * whether smu support query error information | |
953 | */ | |
bc143d8b | 954 | ret = amdgpu_dpm_get_ecc_info(adev, (void *)&(ras->umc_ecc)); |
fdcb279d | 955 | if (ret == -EOPNOTSUPP) { |
efe17d5a | 956 | if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops && |
957 | adev->umc.ras->ras_block.hw_ops->query_ras_error_count) | |
958 | adev->umc.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data); | |
fdcb279d SY |
959 | |
960 | /* umc query_ras_error_address is also responsible for clearing | |
961 | * error status | |
962 | */ | |
efe17d5a | 963 | if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops && |
964 | adev->umc.ras->ras_block.hw_ops->query_ras_error_address) | |
965 | adev->umc.ras->ras_block.hw_ops->query_ras_error_address(adev, err_data); | |
fdcb279d | 966 | } else if (!ret) { |
efe17d5a | 967 | if (adev->umc.ras && |
968 | adev->umc.ras->ecc_info_query_ras_error_count) | |
969 | adev->umc.ras->ecc_info_query_ras_error_count(adev, err_data); | |
fdcb279d | 970 | |
efe17d5a | 971 | if (adev->umc.ras && |
972 | adev->umc.ras->ecc_info_query_ras_error_address) | |
973 | adev->umc.ras->ecc_info_query_ras_error_address(adev, err_data); | |
fdcb279d SY |
974 | } |
975 | } | |
976 | ||
c030f2e4 | 977 | /* query/inject/cure begin */ |
761d86d3 | 978 | int amdgpu_ras_query_error_status(struct amdgpu_device *adev, |
4d9f771e | 979 | struct ras_query_if *info) |
c030f2e4 | 980 | { |
b6efdb02 | 981 | struct amdgpu_ras_block_object *block_obj = NULL; |
c030f2e4 | 982 | struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head); |
6f102dba | 983 | struct ras_err_data err_data = {0, 0, 0, NULL}; |
c030f2e4 | 984 | |
985 | if (!obj) | |
986 | return -EINVAL; | |
c030f2e4 | 987 | |
7389a5b8 | 988 | if (info->head.block == AMDGPU_RAS_BLOCK__UMC) { |
fdcb279d | 989 | amdgpu_ras_get_ecc_info(adev, &err_data); |
7389a5b8 | 990 | } else { |
991 | block_obj = amdgpu_ras_get_ras_block(adev, info->head.block, 0); | |
8b0fb0e9 | 992 | if (!block_obj || !block_obj->hw_ops) { |
afa37315 LT |
993 | dev_dbg_once(adev->dev, "%s doesn't config RAS function\n", |
994 | get_ras_block_str(&info->head)); | |
8b0fb0e9 | 995 | return -EINVAL; |
3e81ee9a | 996 | } |
761d86d3 | 997 | |
6c245386 | 998 | if (block_obj->hw_ops->query_ras_error_count) |
999 | block_obj->hw_ops->query_ras_error_count(adev, &err_data); | |
7389a5b8 | 1000 | |
1001 | if ((info->head.block == AMDGPU_RAS_BLOCK__SDMA) || | |
1002 | (info->head.block == AMDGPU_RAS_BLOCK__GFX) || | |
1003 | (info->head.block == AMDGPU_RAS_BLOCK__MMHUB)) { | |
1004 | if (block_obj->hw_ops->query_ras_error_status) | |
1005 | block_obj->hw_ops->query_ras_error_status(adev); | |
1006 | } | |
939e2258 | 1007 | } |
05a58345 TZ |
1008 | |
1009 | obj->err_data.ue_count += err_data.ue_count; | |
1010 | obj->err_data.ce_count += err_data.ce_count; | |
1011 | ||
c030f2e4 | 1012 | info->ue_count = obj->err_data.ue_count; |
1013 | info->ce_count = obj->err_data.ce_count; | |
1014 | ||
7c6e68c7 | 1015 | if (err_data.ce_count) { |
a30f1286 HZ |
1016 | if (adev->smuio.funcs && |
1017 | adev->smuio.funcs->get_socket_id && | |
1018 | adev->smuio.funcs->get_die_id) { | |
1019 | dev_info(adev->dev, "socket: %d, die: %d " | |
1020 | "%ld correctable hardware errors " | |
6952e99c GC |
1021 | "detected in %s block, no user " |
1022 | "action is needed.\n", | |
a30f1286 HZ |
1023 | adev->smuio.funcs->get_socket_id(adev), |
1024 | adev->smuio.funcs->get_die_id(adev), | |
6952e99c | 1025 | obj->err_data.ce_count, |
640ae42e | 1026 | get_ras_block_str(&info->head)); |
a30f1286 HZ |
1027 | } else { |
1028 | dev_info(adev->dev, "%ld correctable hardware errors " | |
6952e99c GC |
1029 | "detected in %s block, no user " |
1030 | "action is needed.\n", | |
1031 | obj->err_data.ce_count, | |
640ae42e | 1032 | get_ras_block_str(&info->head)); |
a30f1286 | 1033 | } |
7c6e68c7 AG |
1034 | } |
1035 | if (err_data.ue_count) { | |
a30f1286 HZ |
1036 | if (adev->smuio.funcs && |
1037 | adev->smuio.funcs->get_socket_id && | |
1038 | adev->smuio.funcs->get_die_id) { | |
1039 | dev_info(adev->dev, "socket: %d, die: %d " | |
1040 | "%ld uncorrectable hardware errors " | |
6952e99c | 1041 | "detected in %s block\n", |
a30f1286 HZ |
1042 | adev->smuio.funcs->get_socket_id(adev), |
1043 | adev->smuio.funcs->get_die_id(adev), | |
6952e99c | 1044 | obj->err_data.ue_count, |
640ae42e | 1045 | get_ras_block_str(&info->head)); |
a30f1286 HZ |
1046 | } else { |
1047 | dev_info(adev->dev, "%ld uncorrectable hardware errors " | |
6952e99c GC |
1048 | "detected in %s block\n", |
1049 | obj->err_data.ue_count, | |
640ae42e | 1050 | get_ras_block_str(&info->head)); |
a30f1286 | 1051 | } |
7c6e68c7 | 1052 | } |
05a58345 | 1053 | |
c030f2e4 | 1054 | return 0; |
1055 | } | |
1056 | ||
761d86d3 DL |
1057 | int amdgpu_ras_reset_error_status(struct amdgpu_device *adev, |
1058 | enum amdgpu_ras_block block) | |
1059 | { | |
b6efdb02 | 1060 | struct amdgpu_ras_block_object *block_obj = amdgpu_ras_get_ras_block(adev, block, 0); |
8b0fb0e9 | 1061 | |
761d86d3 DL |
1062 | if (!amdgpu_ras_is_supported(adev, block)) |
1063 | return -EINVAL; | |
1064 | ||
7389a5b8 | 1065 | if (!block_obj || !block_obj->hw_ops) { |
afa37315 LT |
1066 | dev_dbg_once(adev->dev, "%s doesn't config RAS function\n", |
1067 | ras_block_str(block)); | |
7389a5b8 | 1068 | return -EINVAL; |
761d86d3 DL |
1069 | } |
1070 | ||
7389a5b8 | 1071 | if (block_obj->hw_ops->reset_ras_error_count) |
1072 | block_obj->hw_ops->reset_ras_error_count(adev); | |
5c23e9e0 | 1073 | |
7389a5b8 | 1074 | if ((block == AMDGPU_RAS_BLOCK__GFX) || |
1075 | (block == AMDGPU_RAS_BLOCK__MMHUB)) { | |
8b0fb0e9 | 1076 | if (block_obj->hw_ops->reset_ras_error_status) |
1077 | block_obj->hw_ops->reset_ras_error_status(adev); | |
761d86d3 | 1078 | } |
5c23e9e0 | 1079 | |
761d86d3 | 1080 | return 0; |
5c23e9e0 JC |
1081 | } |
1082 | ||
c030f2e4 | 1083 | /* wrapper of psp_ras_trigger_error */ |
1084 | int amdgpu_ras_error_inject(struct amdgpu_device *adev, | |
1085 | struct ras_inject_if *info) | |
1086 | { | |
1087 | struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head); | |
1088 | struct ta_ras_trigger_error_input block_info = { | |
828cfa29 | 1089 | .block_id = amdgpu_ras_block_to_ta(info->head.block), |
1090 | .inject_error_type = amdgpu_ras_error_to_ta(info->head.type), | |
c030f2e4 | 1091 | .sub_block_index = info->head.sub_block_index, |
1092 | .address = info->address, | |
1093 | .value = info->value, | |
1094 | }; | |
ab3b9de6 YL |
1095 | int ret = -EINVAL; |
1096 | struct amdgpu_ras_block_object *block_obj = amdgpu_ras_get_ras_block(adev, | |
1097 | info->head.block, | |
1098 | info->head.sub_block_index); | |
c030f2e4 | 1099 | |
248c9635 TZ |
1100 | /* inject on guest isn't allowed, return success directly */ |
1101 | if (amdgpu_sriov_vf(adev)) | |
1102 | return 0; | |
1103 | ||
c030f2e4 | 1104 | if (!obj) |
1105 | return -EINVAL; | |
1106 | ||
22d4ba53 | 1107 | if (!block_obj || !block_obj->hw_ops) { |
afa37315 LT |
1108 | dev_dbg_once(adev->dev, "%s doesn't config RAS function\n", |
1109 | get_ras_block_str(&info->head)); | |
22d4ba53 | 1110 | return -EINVAL; |
1111 | } | |
1112 | ||
a6c44d25 JC |
1113 | /* Calculate XGMI relative offset */ |
1114 | if (adev->gmc.xgmi.num_physical_nodes > 1) { | |
19744f5f HZ |
1115 | block_info.address = |
1116 | amdgpu_xgmi_get_relative_phy_addr(adev, | |
1117 | block_info.address); | |
a6c44d25 JC |
1118 | } |
1119 | ||
22d4ba53 | 1120 | if (info->head.block == AMDGPU_RAS_BLOCK__GFX) { |
8b0fb0e9 | 1121 | if (block_obj->hw_ops->ras_error_inject) |
1122 | ret = block_obj->hw_ops->ras_error_inject(adev, info); | |
22d4ba53 | 1123 | } else { |
1124 | /* If defined special ras_error_inject(e.g: xgmi), implement special ras_error_inject */ | |
1125 | if (block_obj->hw_ops->ras_error_inject) | |
1126 | ret = block_obj->hw_ops->ras_error_inject(adev, &block_info); | |
1127 | else /*If not defined .ras_error_inject, use default ras_error_inject*/ | |
1128 | ret = psp_ras_trigger_error(&adev->psp, &block_info); | |
a5dd40ca HZ |
1129 | } |
1130 | ||
011907fd DL |
1131 | if (ret) |
1132 | dev_err(adev->dev, "ras inject %s failed %d\n", | |
640ae42e | 1133 | get_ras_block_str(&info->head), ret); |
c030f2e4 | 1134 | |
1135 | return ret; | |
1136 | } | |
1137 | ||
4d9f771e LT |
1138 | /** |
1139 | * amdgpu_ras_query_error_count -- Get error counts of all IPs | |
bbe04dec IB |
1140 | * @adev: pointer to AMD GPU device |
1141 | * @ce_count: pointer to an integer to be set to the count of correctible errors. | |
1142 | * @ue_count: pointer to an integer to be set to the count of uncorrectible | |
4d9f771e LT |
1143 | * errors. |
1144 | * | |
1145 | * If set, @ce_count or @ue_count, count and return the corresponding | |
1146 | * error counts in those integer pointers. Return 0 if the device | |
1147 | * supports RAS. Return -EOPNOTSUPP if the device doesn't support RAS. | |
1148 | */ | |
1149 | int amdgpu_ras_query_error_count(struct amdgpu_device *adev, | |
1150 | unsigned long *ce_count, | |
1151 | unsigned long *ue_count) | |
c030f2e4 | 1152 | { |
1153 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
1154 | struct ras_manager *obj; | |
a46751fb | 1155 | unsigned long ce, ue; |
c030f2e4 | 1156 | |
8ab0d6f0 | 1157 | if (!adev->ras_enabled || !con) |
4d9f771e LT |
1158 | return -EOPNOTSUPP; |
1159 | ||
1160 | /* Don't count since no reporting. | |
1161 | */ | |
1162 | if (!ce_count && !ue_count) | |
1163 | return 0; | |
c030f2e4 | 1164 | |
a46751fb LT |
1165 | ce = 0; |
1166 | ue = 0; | |
c030f2e4 | 1167 | list_for_each_entry(obj, &con->head, node) { |
1168 | struct ras_query_if info = { | |
1169 | .head = obj->head, | |
1170 | }; | |
4d9f771e | 1171 | int res; |
c030f2e4 | 1172 | |
4d9f771e LT |
1173 | res = amdgpu_ras_query_error_status(adev, &info); |
1174 | if (res) | |
1175 | return res; | |
c030f2e4 | 1176 | |
2a460963 CL |
1177 | if (adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 2) && |
1178 | adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 4)) { | |
1179 | if (amdgpu_ras_reset_error_status(adev, info.head.block)) | |
1180 | dev_warn(adev->dev, "Failed to reset error counter and error status"); | |
1181 | } | |
1182 | ||
a46751fb LT |
1183 | ce += info.ce_count; |
1184 | ue += info.ue_count; | |
c030f2e4 | 1185 | } |
1186 | ||
a46751fb LT |
1187 | if (ce_count) |
1188 | *ce_count = ce; | |
1189 | ||
1190 | if (ue_count) | |
1191 | *ue_count = ue; | |
4d9f771e LT |
1192 | |
1193 | return 0; | |
c030f2e4 | 1194 | } |
1195 | /* query/inject/cure end */ | |
1196 | ||
1197 | ||
1198 | /* sysfs begin */ | |
1199 | ||
466b1793 | 1200 | static int amdgpu_ras_badpages_read(struct amdgpu_device *adev, |
1201 | struct ras_badpage **bps, unsigned int *count); | |
1202 | ||
1203 | static char *amdgpu_ras_badpage_flags_str(unsigned int flags) | |
1204 | { | |
1205 | switch (flags) { | |
52dd95f2 | 1206 | case AMDGPU_RAS_RETIRE_PAGE_RESERVED: |
466b1793 | 1207 | return "R"; |
52dd95f2 | 1208 | case AMDGPU_RAS_RETIRE_PAGE_PENDING: |
466b1793 | 1209 | return "P"; |
52dd95f2 | 1210 | case AMDGPU_RAS_RETIRE_PAGE_FAULT: |
466b1793 | 1211 | default: |
1212 | return "F"; | |
aec576f9 | 1213 | } |
466b1793 | 1214 | } |
1215 | ||
f77c7109 AD |
1216 | /** |
1217 | * DOC: AMDGPU RAS sysfs gpu_vram_bad_pages Interface | |
466b1793 | 1218 | * |
1219 | * It allows user to read the bad pages of vram on the gpu through | |
1220 | * /sys/class/drm/card[0/1/2...]/device/ras/gpu_vram_bad_pages | |
1221 | * | |
1222 | * It outputs multiple lines, and each line stands for one gpu page. | |
1223 | * | |
1224 | * The format of one line is below, | |
1225 | * gpu pfn : gpu page size : flags | |
1226 | * | |
1227 | * gpu pfn and gpu page size are printed in hex format. | |
1228 | * flags can be one of below character, | |
f77c7109 | 1229 | * |
466b1793 | 1230 | * R: reserved, this gpu page is reserved and not able to use. |
f77c7109 | 1231 | * |
466b1793 | 1232 | * P: pending for reserve, this gpu page is marked as bad, will be reserved |
f77c7109 AD |
1233 | * in next window of page_reserve. |
1234 | * | |
466b1793 | 1235 | * F: unable to reserve. this gpu page can't be reserved due to some reasons. |
1236 | * | |
f77c7109 AD |
1237 | * Examples: |
1238 | * | |
1239 | * .. code-block:: bash | |
1240 | * | |
1241 | * 0x00000001 : 0x00001000 : R | |
1242 | * 0x00000002 : 0x00001000 : P | |
1243 | * | |
466b1793 | 1244 | */ |
1245 | ||
1246 | static ssize_t amdgpu_ras_sysfs_badpages_read(struct file *f, | |
1247 | struct kobject *kobj, struct bin_attribute *attr, | |
1248 | char *buf, loff_t ppos, size_t count) | |
1249 | { | |
1250 | struct amdgpu_ras *con = | |
1251 | container_of(attr, struct amdgpu_ras, badpages_attr); | |
1252 | struct amdgpu_device *adev = con->adev; | |
1253 | const unsigned int element_size = | |
1254 | sizeof("0xabcdabcd : 0x12345678 : R\n") - 1; | |
d6ee400e SA |
1255 | unsigned int start = div64_ul(ppos + element_size - 1, element_size); |
1256 | unsigned int end = div64_ul(ppos + count - 1, element_size); | |
466b1793 | 1257 | ssize_t s = 0; |
1258 | struct ras_badpage *bps = NULL; | |
1259 | unsigned int bps_count = 0; | |
1260 | ||
1261 | memset(buf, 0, count); | |
1262 | ||
1263 | if (amdgpu_ras_badpages_read(adev, &bps, &bps_count)) | |
1264 | return 0; | |
1265 | ||
1266 | for (; start < end && start < bps_count; start++) | |
1267 | s += scnprintf(&buf[s], element_size + 1, | |
1268 | "0x%08x : 0x%08x : %1s\n", | |
1269 | bps[start].bp, | |
1270 | bps[start].size, | |
1271 | amdgpu_ras_badpage_flags_str(bps[start].flags)); | |
1272 | ||
1273 | kfree(bps); | |
1274 | ||
1275 | return s; | |
1276 | } | |
1277 | ||
c030f2e4 | 1278 | static ssize_t amdgpu_ras_sysfs_features_read(struct device *dev, |
1279 | struct device_attribute *attr, char *buf) | |
1280 | { | |
1281 | struct amdgpu_ras *con = | |
1282 | container_of(attr, struct amdgpu_ras, features_attr); | |
c030f2e4 | 1283 | |
2cffcb66 | 1284 | return sysfs_emit(buf, "feature mask: 0x%x\n", con->features); |
c030f2e4 | 1285 | } |
1286 | ||
f848159b GC |
1287 | static void amdgpu_ras_sysfs_remove_bad_page_node(struct amdgpu_device *adev) |
1288 | { | |
1289 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
1290 | ||
1291 | sysfs_remove_file_from_group(&adev->dev->kobj, | |
1292 | &con->badpages_attr.attr, | |
1293 | RAS_FS_NAME); | |
1294 | } | |
1295 | ||
c030f2e4 | 1296 | static int amdgpu_ras_sysfs_remove_feature_node(struct amdgpu_device *adev) |
1297 | { | |
1298 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
1299 | struct attribute *attrs[] = { | |
1300 | &con->features_attr.attr, | |
1301 | NULL | |
1302 | }; | |
1303 | struct attribute_group group = { | |
eb0c3cd4 | 1304 | .name = RAS_FS_NAME, |
c030f2e4 | 1305 | .attrs = attrs, |
1306 | }; | |
1307 | ||
1308 | sysfs_remove_group(&adev->dev->kobj, &group); | |
1309 | ||
1310 | return 0; | |
1311 | } | |
1312 | ||
1313 | int amdgpu_ras_sysfs_create(struct amdgpu_device *adev, | |
9252d33d | 1314 | struct ras_common_if *head) |
c030f2e4 | 1315 | { |
9252d33d | 1316 | struct ras_manager *obj = amdgpu_ras_find_obj(adev, head); |
c030f2e4 | 1317 | |
1318 | if (!obj || obj->attr_inuse) | |
1319 | return -EINVAL; | |
1320 | ||
1321 | get_obj(obj); | |
1322 | ||
9252d33d | 1323 | snprintf(obj->fs_data.sysfs_name, sizeof(obj->fs_data.sysfs_name), |
1324 | "%s_err_count", head->name); | |
c030f2e4 | 1325 | |
1326 | obj->sysfs_attr = (struct device_attribute){ | |
1327 | .attr = { | |
1328 | .name = obj->fs_data.sysfs_name, | |
1329 | .mode = S_IRUGO, | |
1330 | }, | |
1331 | .show = amdgpu_ras_sysfs_read, | |
1332 | }; | |
163def43 | 1333 | sysfs_attr_init(&obj->sysfs_attr.attr); |
c030f2e4 | 1334 | |
1335 | if (sysfs_add_file_to_group(&adev->dev->kobj, | |
1336 | &obj->sysfs_attr.attr, | |
eb0c3cd4 | 1337 | RAS_FS_NAME)) { |
c030f2e4 | 1338 | put_obj(obj); |
1339 | return -EINVAL; | |
1340 | } | |
1341 | ||
1342 | obj->attr_inuse = 1; | |
1343 | ||
1344 | return 0; | |
1345 | } | |
1346 | ||
1347 | int amdgpu_ras_sysfs_remove(struct amdgpu_device *adev, | |
1348 | struct ras_common_if *head) | |
1349 | { | |
1350 | struct ras_manager *obj = amdgpu_ras_find_obj(adev, head); | |
1351 | ||
1352 | if (!obj || !obj->attr_inuse) | |
1353 | return -EINVAL; | |
1354 | ||
1355 | sysfs_remove_file_from_group(&adev->dev->kobj, | |
1356 | &obj->sysfs_attr.attr, | |
eb0c3cd4 | 1357 | RAS_FS_NAME); |
c030f2e4 | 1358 | obj->attr_inuse = 0; |
1359 | put_obj(obj); | |
1360 | ||
1361 | return 0; | |
1362 | } | |
1363 | ||
1364 | static int amdgpu_ras_sysfs_remove_all(struct amdgpu_device *adev) | |
1365 | { | |
1366 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
1367 | struct ras_manager *obj, *tmp; | |
1368 | ||
1369 | list_for_each_entry_safe(obj, tmp, &con->head, node) { | |
1370 | amdgpu_ras_sysfs_remove(adev, &obj->head); | |
1371 | } | |
1372 | ||
f848159b GC |
1373 | if (amdgpu_bad_page_threshold != 0) |
1374 | amdgpu_ras_sysfs_remove_bad_page_node(adev); | |
1375 | ||
c030f2e4 | 1376 | amdgpu_ras_sysfs_remove_feature_node(adev); |
1377 | ||
1378 | return 0; | |
1379 | } | |
1380 | /* sysfs end */ | |
1381 | ||
ef177d11 AD |
1382 | /** |
1383 | * DOC: AMDGPU RAS Reboot Behavior for Unrecoverable Errors | |
1384 | * | |
1385 | * Normally when there is an uncorrectable error, the driver will reset | |
1386 | * the GPU to recover. However, in the event of an unrecoverable error, | |
1387 | * the driver provides an interface to reboot the system automatically | |
1388 | * in that event. | |
1389 | * | |
1390 | * The following file in debugfs provides that interface: | |
1391 | * /sys/kernel/debug/dri/[0/1/2...]/ras/auto_reboot | |
1392 | * | |
1393 | * Usage: | |
1394 | * | |
1395 | * .. code-block:: bash | |
1396 | * | |
1397 | * echo true > .../ras/auto_reboot | |
1398 | * | |
1399 | */ | |
c030f2e4 | 1400 | /* debugfs begin */ |
ea1b8c9b | 1401 | static struct dentry *amdgpu_ras_debugfs_create_ctrl_node(struct amdgpu_device *adev) |
36ea1bd2 | 1402 | { |
1403 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
ef0d7d20 LT |
1404 | struct drm_minor *minor = adev_to_drm(adev)->primary; |
1405 | struct dentry *dir; | |
36ea1bd2 | 1406 | |
88293c03 ND |
1407 | dir = debugfs_create_dir(RAS_FS_NAME, minor->debugfs_root); |
1408 | debugfs_create_file("ras_ctrl", S_IWUGO | S_IRUGO, dir, adev, | |
1409 | &amdgpu_ras_debugfs_ctrl_ops); | |
1410 | debugfs_create_file("ras_eeprom_reset", S_IWUGO | S_IRUGO, dir, adev, | |
1411 | &amdgpu_ras_debugfs_eeprom_ops); | |
7fb64071 LT |
1412 | debugfs_create_u32("bad_page_cnt_threshold", 0444, dir, |
1413 | &con->bad_page_cnt_threshold); | |
ef0d7d20 LT |
1414 | debugfs_create_x32("ras_hw_enabled", 0444, dir, &adev->ras_hw_enabled); |
1415 | debugfs_create_x32("ras_enabled", 0444, dir, &adev->ras_enabled); | |
c65b0805 LT |
1416 | debugfs_create_file("ras_eeprom_size", S_IRUGO, dir, adev, |
1417 | &amdgpu_ras_debugfs_eeprom_size_ops); | |
1418 | con->de_ras_eeprom_table = debugfs_create_file("ras_eeprom_table", | |
1419 | S_IRUGO, dir, adev, | |
1420 | &amdgpu_ras_debugfs_eeprom_table_ops); | |
1421 | amdgpu_ras_debugfs_set_ret_size(&con->eeprom_control); | |
c688a06b GC |
1422 | |
1423 | /* | |
1424 | * After one uncorrectable error happens, usually GPU recovery will | |
1425 | * be scheduled. But due to the known problem in GPU recovery failing | |
1426 | * to bring GPU back, below interface provides one direct way to | |
1427 | * user to reboot system automatically in such case within | |
1428 | * ERREVENT_ATHUB_INTERRUPT generated. Normal GPU recovery routine | |
1429 | * will never be called. | |
1430 | */ | |
88293c03 | 1431 | debugfs_create_bool("auto_reboot", S_IWUGO | S_IRUGO, dir, &con->reboot); |
66459e1d GC |
1432 | |
1433 | /* | |
1434 | * User could set this not to clean up hardware's error count register | |
1435 | * of RAS IPs during ras recovery. | |
1436 | */ | |
88293c03 ND |
1437 | debugfs_create_bool("disable_ras_err_cnt_harvest", 0644, dir, |
1438 | &con->disable_ras_err_cnt_harvest); | |
1439 | return dir; | |
36ea1bd2 | 1440 | } |
1441 | ||
cedf7884 | 1442 | static void amdgpu_ras_debugfs_create(struct amdgpu_device *adev, |
88293c03 ND |
1443 | struct ras_fs_if *head, |
1444 | struct dentry *dir) | |
c030f2e4 | 1445 | { |
c030f2e4 | 1446 | struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head->head); |
c030f2e4 | 1447 | |
88293c03 | 1448 | if (!obj || !dir) |
450f30ea | 1449 | return; |
c030f2e4 | 1450 | |
1451 | get_obj(obj); | |
1452 | ||
1453 | memcpy(obj->fs_data.debugfs_name, | |
1454 | head->debugfs_name, | |
1455 | sizeof(obj->fs_data.debugfs_name)); | |
1456 | ||
88293c03 ND |
1457 | debugfs_create_file(obj->fs_data.debugfs_name, S_IWUGO | S_IRUGO, dir, |
1458 | obj, &amdgpu_ras_debugfs_ops); | |
c030f2e4 | 1459 | } |
1460 | ||
f9317014 TZ |
1461 | void amdgpu_ras_debugfs_create_all(struct amdgpu_device *adev) |
1462 | { | |
1463 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
88293c03 | 1464 | struct dentry *dir; |
c1509f3f | 1465 | struct ras_manager *obj; |
f9317014 TZ |
1466 | struct ras_fs_if fs_info; |
1467 | ||
1468 | /* | |
1469 | * it won't be called in resume path, no need to check | |
1470 | * suspend and gpu reset status | |
1471 | */ | |
cedf7884 | 1472 | if (!IS_ENABLED(CONFIG_DEBUG_FS) || !con) |
f9317014 TZ |
1473 | return; |
1474 | ||
88293c03 | 1475 | dir = amdgpu_ras_debugfs_create_ctrl_node(adev); |
f9317014 | 1476 | |
c1509f3f | 1477 | list_for_each_entry(obj, &con->head, node) { |
f9317014 TZ |
1478 | if (amdgpu_ras_is_supported(adev, obj->head.block) && |
1479 | (obj->attr_inuse == 1)) { | |
1480 | sprintf(fs_info.debugfs_name, "%s_err_inject", | |
640ae42e | 1481 | get_ras_block_str(&obj->head)); |
f9317014 | 1482 | fs_info.head = obj->head; |
88293c03 | 1483 | amdgpu_ras_debugfs_create(adev, &fs_info, dir); |
f9317014 TZ |
1484 | } |
1485 | } | |
1486 | } | |
1487 | ||
c030f2e4 | 1488 | /* debugfs end */ |
1489 | ||
1490 | /* ras fs */ | |
c3d4d45d GC |
1491 | static BIN_ATTR(gpu_vram_bad_pages, S_IRUGO, |
1492 | amdgpu_ras_sysfs_badpages_read, NULL, 0); | |
1493 | static DEVICE_ATTR(features, S_IRUGO, | |
1494 | amdgpu_ras_sysfs_features_read, NULL); | |
c030f2e4 | 1495 | static int amdgpu_ras_fs_init(struct amdgpu_device *adev) |
1496 | { | |
c3d4d45d GC |
1497 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); |
1498 | struct attribute_group group = { | |
1499 | .name = RAS_FS_NAME, | |
1500 | }; | |
1501 | struct attribute *attrs[] = { | |
1502 | &con->features_attr.attr, | |
1503 | NULL | |
1504 | }; | |
1505 | struct bin_attribute *bin_attrs[] = { | |
1506 | NULL, | |
1507 | NULL, | |
1508 | }; | |
a069a9eb | 1509 | int r; |
c030f2e4 | 1510 | |
c3d4d45d GC |
1511 | /* add features entry */ |
1512 | con->features_attr = dev_attr_features; | |
1513 | group.attrs = attrs; | |
1514 | sysfs_attr_init(attrs[0]); | |
1515 | ||
1516 | if (amdgpu_bad_page_threshold != 0) { | |
1517 | /* add bad_page_features entry */ | |
1518 | bin_attr_gpu_vram_bad_pages.private = NULL; | |
1519 | con->badpages_attr = bin_attr_gpu_vram_bad_pages; | |
1520 | bin_attrs[0] = &con->badpages_attr; | |
1521 | group.bin_attrs = bin_attrs; | |
1522 | sysfs_bin_attr_init(bin_attrs[0]); | |
1523 | } | |
1524 | ||
a069a9eb AD |
1525 | r = sysfs_create_group(&adev->dev->kobj, &group); |
1526 | if (r) | |
1527 | dev_err(adev->dev, "Failed to create RAS sysfs group!"); | |
f848159b | 1528 | |
c030f2e4 | 1529 | return 0; |
1530 | } | |
1531 | ||
1532 | static int amdgpu_ras_fs_fini(struct amdgpu_device *adev) | |
1533 | { | |
88293c03 ND |
1534 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); |
1535 | struct ras_manager *con_obj, *ip_obj, *tmp; | |
1536 | ||
1537 | if (IS_ENABLED(CONFIG_DEBUG_FS)) { | |
1538 | list_for_each_entry_safe(con_obj, tmp, &con->head, node) { | |
1539 | ip_obj = amdgpu_ras_find_obj(adev, &con_obj->head); | |
1540 | if (ip_obj) | |
1541 | put_obj(ip_obj); | |
1542 | } | |
1543 | } | |
1544 | ||
c030f2e4 | 1545 | amdgpu_ras_sysfs_remove_all(adev); |
1546 | return 0; | |
1547 | } | |
1548 | /* ras fs end */ | |
1549 | ||
1550 | /* ih begin */ | |
b3c76814 TZ |
1551 | |
1552 | /* For the hardware that cannot enable bif ring for both ras_controller_irq | |
1553 | * and ras_err_evnet_athub_irq ih cookies, the driver has to poll status | |
1554 | * register to check whether the interrupt is triggered or not, and properly | |
1555 | * ack the interrupt if it is there | |
1556 | */ | |
1557 | void amdgpu_ras_interrupt_fatal_error_handler(struct amdgpu_device *adev) | |
1558 | { | |
950d6425 SY |
1559 | /* Fatal error events are handled on host side */ |
1560 | if (amdgpu_sriov_vf(adev) || | |
1561 | !amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__PCIE_BIF)) | |
b3c76814 TZ |
1562 | return; |
1563 | ||
1564 | if (adev->nbio.ras && | |
1565 | adev->nbio.ras->handle_ras_controller_intr_no_bifring) | |
1566 | adev->nbio.ras->handle_ras_controller_intr_no_bifring(adev); | |
1567 | ||
1568 | if (adev->nbio.ras && | |
1569 | adev->nbio.ras->handle_ras_err_event_athub_intr_no_bifring) | |
1570 | adev->nbio.ras->handle_ras_err_event_athub_intr_no_bifring(adev); | |
1571 | } | |
1572 | ||
66f87949 TZ |
1573 | static void amdgpu_ras_interrupt_poison_consumption_handler(struct ras_manager *obj, |
1574 | struct amdgpu_iv_entry *entry) | |
1575 | { | |
b63ac5d3 | 1576 | bool poison_stat = false; |
66f87949 | 1577 | struct amdgpu_device *adev = obj->adev; |
66f87949 TZ |
1578 | struct amdgpu_ras_block_object *block_obj = |
1579 | amdgpu_ras_get_ras_block(adev, obj->head.block, 0); | |
1580 | ||
b63ac5d3 TZ |
1581 | if (!block_obj || !block_obj->hw_ops) |
1582 | return; | |
66f87949 | 1583 | |
b63ac5d3 TZ |
1584 | /* both query_poison_status and handle_poison_consumption are optional, |
1585 | * but at least one of them should be implemented if we need poison | |
1586 | * consumption handler | |
1587 | */ | |
1588 | if (block_obj->hw_ops->query_poison_status) { | |
1589 | poison_stat = block_obj->hw_ops->query_poison_status(adev); | |
1590 | if (!poison_stat) { | |
1591 | /* Not poison consumption interrupt, no need to handle it */ | |
1592 | dev_info(adev->dev, "No RAS poison status in %s poison IH.\n", | |
1593 | block_obj->ras_comm.name); | |
1594 | ||
1595 | return; | |
66f87949 TZ |
1596 | } |
1597 | } | |
1598 | ||
b63ac5d3 | 1599 | if (!adev->gmc.xgmi.connected_to_cpu) |
1ed0e176 | 1600 | amdgpu_umc_poison_handler(adev, false); |
b63ac5d3 TZ |
1601 | |
1602 | if (block_obj->hw_ops->handle_poison_consumption) | |
1603 | poison_stat = block_obj->hw_ops->handle_poison_consumption(adev); | |
1604 | ||
1605 | /* gpu reset is fallback for failed and default cases */ | |
1606 | if (poison_stat) { | |
1607 | dev_info(adev->dev, "GPU reset for %s RAS poison consumption is issued!\n", | |
1608 | block_obj->ras_comm.name); | |
66f87949 | 1609 | amdgpu_ras_reset_gpu(adev); |
b63ac5d3 | 1610 | } |
66f87949 TZ |
1611 | } |
1612 | ||
50a7d025 TZ |
1613 | static void amdgpu_ras_interrupt_poison_creation_handler(struct ras_manager *obj, |
1614 | struct amdgpu_iv_entry *entry) | |
1615 | { | |
1616 | dev_info(obj->adev->dev, | |
1617 | "Poison is created, no user action is needed.\n"); | |
1618 | } | |
1619 | ||
1620 | static void amdgpu_ras_interrupt_umc_handler(struct ras_manager *obj, | |
1621 | struct amdgpu_iv_entry *entry) | |
1622 | { | |
1623 | struct ras_ih_data *data = &obj->ih_data; | |
1624 | struct ras_err_data err_data = {0, 0, 0, NULL}; | |
1625 | int ret; | |
1626 | ||
1627 | if (!data->cb) | |
1628 | return; | |
1629 | ||
1630 | /* Let IP handle its data, maybe we need get the output | |
1631 | * from the callback to update the error type/count, etc | |
1632 | */ | |
1633 | ret = data->cb(obj->adev, &err_data, entry); | |
1634 | /* ue will trigger an interrupt, and in that case | |
1635 | * we need do a reset to recovery the whole system. | |
1636 | * But leave IP do that recovery, here we just dispatch | |
1637 | * the error. | |
1638 | */ | |
1639 | if (ret == AMDGPU_RAS_SUCCESS) { | |
1640 | /* these counts could be left as 0 if | |
1641 | * some blocks do not count error number | |
1642 | */ | |
1643 | obj->err_data.ue_count += err_data.ue_count; | |
1644 | obj->err_data.ce_count += err_data.ce_count; | |
1645 | } | |
1646 | } | |
1647 | ||
c030f2e4 | 1648 | static void amdgpu_ras_interrupt_handler(struct ras_manager *obj) |
1649 | { | |
1650 | struct ras_ih_data *data = &obj->ih_data; | |
1651 | struct amdgpu_iv_entry entry; | |
c030f2e4 | 1652 | |
1653 | while (data->rptr != data->wptr) { | |
1654 | rmb(); | |
1655 | memcpy(&entry, &data->ring[data->rptr], | |
1656 | data->element_size); | |
1657 | ||
1658 | wmb(); | |
1659 | data->rptr = (data->aligned_element_size + | |
1660 | data->rptr) % data->ring_size; | |
1661 | ||
50a7d025 TZ |
1662 | if (amdgpu_ras_is_poison_mode_supported(obj->adev)) { |
1663 | if (obj->head.block == AMDGPU_RAS_BLOCK__UMC) | |
1664 | amdgpu_ras_interrupt_poison_creation_handler(obj, &entry); | |
66f87949 TZ |
1665 | else |
1666 | amdgpu_ras_interrupt_poison_consumption_handler(obj, &entry); | |
50a7d025 TZ |
1667 | } else { |
1668 | if (obj->head.block == AMDGPU_RAS_BLOCK__UMC) | |
1669 | amdgpu_ras_interrupt_umc_handler(obj, &entry); | |
1670 | else | |
1671 | dev_warn(obj->adev->dev, | |
1672 | "No RAS interrupt handler for non-UMC block with poison disabled.\n"); | |
c030f2e4 | 1673 | } |
1674 | } | |
1675 | } | |
1676 | ||
1677 | static void amdgpu_ras_interrupt_process_handler(struct work_struct *work) | |
1678 | { | |
1679 | struct ras_ih_data *data = | |
1680 | container_of(work, struct ras_ih_data, ih_work); | |
1681 | struct ras_manager *obj = | |
1682 | container_of(data, struct ras_manager, ih_data); | |
1683 | ||
1684 | amdgpu_ras_interrupt_handler(obj); | |
1685 | } | |
1686 | ||
1687 | int amdgpu_ras_interrupt_dispatch(struct amdgpu_device *adev, | |
1688 | struct ras_dispatch_if *info) | |
1689 | { | |
1690 | struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head); | |
1691 | struct ras_ih_data *data = &obj->ih_data; | |
1692 | ||
1693 | if (!obj) | |
1694 | return -EINVAL; | |
1695 | ||
1696 | if (data->inuse == 0) | |
1697 | return 0; | |
1698 | ||
1699 | /* Might be overflow... */ | |
1700 | memcpy(&data->ring[data->wptr], info->entry, | |
1701 | data->element_size); | |
1702 | ||
1703 | wmb(); | |
1704 | data->wptr = (data->aligned_element_size + | |
1705 | data->wptr) % data->ring_size; | |
1706 | ||
1707 | schedule_work(&data->ih_work); | |
1708 | ||
1709 | return 0; | |
1710 | } | |
1711 | ||
1712 | int amdgpu_ras_interrupt_remove_handler(struct amdgpu_device *adev, | |
9252d33d | 1713 | struct ras_common_if *head) |
c030f2e4 | 1714 | { |
9252d33d | 1715 | struct ras_manager *obj = amdgpu_ras_find_obj(adev, head); |
c030f2e4 | 1716 | struct ras_ih_data *data; |
1717 | ||
1718 | if (!obj) | |
1719 | return -EINVAL; | |
1720 | ||
1721 | data = &obj->ih_data; | |
1722 | if (data->inuse == 0) | |
1723 | return 0; | |
1724 | ||
1725 | cancel_work_sync(&data->ih_work); | |
1726 | ||
1727 | kfree(data->ring); | |
1728 | memset(data, 0, sizeof(*data)); | |
1729 | put_obj(obj); | |
1730 | ||
1731 | return 0; | |
1732 | } | |
1733 | ||
1734 | int amdgpu_ras_interrupt_add_handler(struct amdgpu_device *adev, | |
9252d33d | 1735 | struct ras_common_if *head) |
c030f2e4 | 1736 | { |
9252d33d | 1737 | struct ras_manager *obj = amdgpu_ras_find_obj(adev, head); |
c030f2e4 | 1738 | struct ras_ih_data *data; |
9252d33d | 1739 | struct amdgpu_ras_block_object *ras_obj; |
c030f2e4 | 1740 | |
1741 | if (!obj) { | |
1742 | /* in case we registe the IH before enable ras feature */ | |
9252d33d | 1743 | obj = amdgpu_ras_create_obj(adev, head); |
c030f2e4 | 1744 | if (!obj) |
1745 | return -EINVAL; | |
1746 | } else | |
1747 | get_obj(obj); | |
1748 | ||
9252d33d | 1749 | ras_obj = container_of(head, struct amdgpu_ras_block_object, ras_comm); |
1750 | ||
c030f2e4 | 1751 | data = &obj->ih_data; |
1752 | /* add the callback.etc */ | |
1753 | *data = (struct ras_ih_data) { | |
1754 | .inuse = 0, | |
9252d33d | 1755 | .cb = ras_obj->ras_cb, |
c030f2e4 | 1756 | .element_size = sizeof(struct amdgpu_iv_entry), |
1757 | .rptr = 0, | |
1758 | .wptr = 0, | |
1759 | }; | |
1760 | ||
1761 | INIT_WORK(&data->ih_work, amdgpu_ras_interrupt_process_handler); | |
1762 | ||
1763 | data->aligned_element_size = ALIGN(data->element_size, 8); | |
1764 | /* the ring can store 64 iv entries. */ | |
1765 | data->ring_size = 64 * data->aligned_element_size; | |
1766 | data->ring = kmalloc(data->ring_size, GFP_KERNEL); | |
1767 | if (!data->ring) { | |
1768 | put_obj(obj); | |
1769 | return -ENOMEM; | |
1770 | } | |
1771 | ||
1772 | /* IH is ready */ | |
1773 | data->inuse = 1; | |
1774 | ||
1775 | return 0; | |
1776 | } | |
1777 | ||
1778 | static int amdgpu_ras_interrupt_remove_all(struct amdgpu_device *adev) | |
1779 | { | |
1780 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
1781 | struct ras_manager *obj, *tmp; | |
1782 | ||
1783 | list_for_each_entry_safe(obj, tmp, &con->head, node) { | |
9252d33d | 1784 | amdgpu_ras_interrupt_remove_handler(adev, &obj->head); |
c030f2e4 | 1785 | } |
1786 | ||
1787 | return 0; | |
1788 | } | |
1789 | /* ih end */ | |
1790 | ||
313c8fd3 GC |
1791 | /* traversal all IPs except NBIO to query error counter */ |
1792 | static void amdgpu_ras_log_on_err_counter(struct amdgpu_device *adev) | |
1793 | { | |
1794 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
1795 | struct ras_manager *obj; | |
1796 | ||
8ab0d6f0 | 1797 | if (!adev->ras_enabled || !con) |
313c8fd3 GC |
1798 | return; |
1799 | ||
1800 | list_for_each_entry(obj, &con->head, node) { | |
1801 | struct ras_query_if info = { | |
1802 | .head = obj->head, | |
1803 | }; | |
1804 | ||
1805 | /* | |
1806 | * PCIE_BIF IP has one different isr by ras controller | |
1807 | * interrupt, the specific ras counter query will be | |
1808 | * done in that isr. So skip such block from common | |
1809 | * sync flood interrupt isr calling. | |
1810 | */ | |
1811 | if (info.head.block == AMDGPU_RAS_BLOCK__PCIE_BIF) | |
1812 | continue; | |
1813 | ||
cf63b702 SY |
1814 | /* |
1815 | * this is a workaround for aldebaran, skip send msg to | |
1816 | * smu to get ecc_info table due to smu handle get ecc | |
1817 | * info table failed temporarily. | |
1818 | * should be removed until smu fix handle ecc_info table. | |
1819 | */ | |
1820 | if ((info.head.block == AMDGPU_RAS_BLOCK__UMC) && | |
1821 | (adev->ip_versions[MP1_HWIP][0] == IP_VERSION(13, 0, 2))) | |
1822 | continue; | |
1823 | ||
761d86d3 | 1824 | amdgpu_ras_query_error_status(adev, &info); |
2a460963 CL |
1825 | |
1826 | if (adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 2) && | |
6da15a23 CL |
1827 | adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 4) && |
1828 | adev->ip_versions[MP0_HWIP][0] != IP_VERSION(13, 0, 0)) { | |
2a460963 CL |
1829 | if (amdgpu_ras_reset_error_status(adev, info.head.block)) |
1830 | dev_warn(adev->dev, "Failed to reset error counter and error status"); | |
1831 | } | |
313c8fd3 GC |
1832 | } |
1833 | } | |
1834 | ||
3f975d0f | 1835 | /* Parse RdRspStatus and WrRspStatus */ |
cd92df93 LJ |
1836 | static void amdgpu_ras_error_status_query(struct amdgpu_device *adev, |
1837 | struct ras_query_if *info) | |
3f975d0f | 1838 | { |
8eb53bb2 | 1839 | struct amdgpu_ras_block_object *block_obj; |
3f975d0f SY |
1840 | /* |
1841 | * Only two block need to query read/write | |
1842 | * RspStatus at current state | |
1843 | */ | |
5e67bba3 | 1844 | if ((info->head.block != AMDGPU_RAS_BLOCK__GFX) && |
1845 | (info->head.block != AMDGPU_RAS_BLOCK__MMHUB)) | |
b6efdb02 | 1846 | return; |
1847 | ||
1848 | block_obj = amdgpu_ras_get_ras_block(adev, | |
1849 | info->head.block, | |
1850 | info->head.sub_block_index); | |
5e67bba3 | 1851 | |
5e67bba3 | 1852 | if (!block_obj || !block_obj->hw_ops) { |
afa37315 LT |
1853 | dev_dbg_once(adev->dev, "%s doesn't config RAS function\n", |
1854 | get_ras_block_str(&info->head)); | |
b6efdb02 | 1855 | return; |
3f975d0f | 1856 | } |
5e67bba3 | 1857 | |
1858 | if (block_obj->hw_ops->query_ras_error_status) | |
ab3b9de6 | 1859 | block_obj->hw_ops->query_ras_error_status(adev); |
5e67bba3 | 1860 | |
3f975d0f SY |
1861 | } |
1862 | ||
1863 | static void amdgpu_ras_query_err_status(struct amdgpu_device *adev) | |
1864 | { | |
1865 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
1866 | struct ras_manager *obj; | |
1867 | ||
8ab0d6f0 | 1868 | if (!adev->ras_enabled || !con) |
3f975d0f SY |
1869 | return; |
1870 | ||
1871 | list_for_each_entry(obj, &con->head, node) { | |
1872 | struct ras_query_if info = { | |
1873 | .head = obj->head, | |
1874 | }; | |
1875 | ||
1876 | amdgpu_ras_error_status_query(adev, &info); | |
1877 | } | |
1878 | } | |
1879 | ||
c030f2e4 | 1880 | /* recovery begin */ |
466b1793 | 1881 | |
1882 | /* return 0 on success. | |
1883 | * caller need free bps. | |
1884 | */ | |
1885 | static int amdgpu_ras_badpages_read(struct amdgpu_device *adev, | |
1886 | struct ras_badpage **bps, unsigned int *count) | |
1887 | { | |
1888 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
1889 | struct ras_err_handler_data *data; | |
1890 | int i = 0; | |
732f2a30 | 1891 | int ret = 0, status; |
466b1793 | 1892 | |
1893 | if (!con || !con->eh_data || !bps || !count) | |
1894 | return -EINVAL; | |
1895 | ||
1896 | mutex_lock(&con->recovery_lock); | |
1897 | data = con->eh_data; | |
1898 | if (!data || data->count == 0) { | |
1899 | *bps = NULL; | |
46cf2fec | 1900 | ret = -EINVAL; |
466b1793 | 1901 | goto out; |
1902 | } | |
1903 | ||
1904 | *bps = kmalloc(sizeof(struct ras_badpage) * data->count, GFP_KERNEL); | |
1905 | if (!*bps) { | |
1906 | ret = -ENOMEM; | |
1907 | goto out; | |
1908 | } | |
1909 | ||
1910 | for (; i < data->count; i++) { | |
1911 | (*bps)[i] = (struct ras_badpage){ | |
9dc23a63 | 1912 | .bp = data->bps[i].retired_page, |
466b1793 | 1913 | .size = AMDGPU_GPU_PAGE_SIZE, |
52dd95f2 | 1914 | .flags = AMDGPU_RAS_RETIRE_PAGE_RESERVED, |
466b1793 | 1915 | }; |
ec6aae97 | 1916 | status = amdgpu_vram_mgr_query_page_status(&adev->mman.vram_mgr, |
676deb38 | 1917 | data->bps[i].retired_page); |
732f2a30 | 1918 | if (status == -EBUSY) |
52dd95f2 | 1919 | (*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_PENDING; |
732f2a30 | 1920 | else if (status == -ENOENT) |
52dd95f2 | 1921 | (*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_FAULT; |
466b1793 | 1922 | } |
1923 | ||
1924 | *count = data->count; | |
1925 | out: | |
1926 | mutex_unlock(&con->recovery_lock); | |
1927 | return ret; | |
1928 | } | |
1929 | ||
c030f2e4 | 1930 | static void amdgpu_ras_do_recovery(struct work_struct *work) |
1931 | { | |
1932 | struct amdgpu_ras *ras = | |
1933 | container_of(work, struct amdgpu_ras, recovery_work); | |
b3dbd6d3 JC |
1934 | struct amdgpu_device *remote_adev = NULL; |
1935 | struct amdgpu_device *adev = ras->adev; | |
1936 | struct list_head device_list, *device_list_handle = NULL; | |
b3dbd6d3 | 1937 | |
f75e94d8 | 1938 | if (!ras->disable_ras_err_cnt_harvest) { |
d95e8e97 DL |
1939 | struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev); |
1940 | ||
f75e94d8 GC |
1941 | /* Build list of devices to query RAS related errors */ |
1942 | if (hive && adev->gmc.xgmi.num_physical_nodes > 1) { | |
1943 | device_list_handle = &hive->device_list; | |
1944 | } else { | |
1945 | INIT_LIST_HEAD(&device_list); | |
1946 | list_add_tail(&adev->gmc.xgmi.head, &device_list); | |
1947 | device_list_handle = &device_list; | |
1948 | } | |
c030f2e4 | 1949 | |
f75e94d8 | 1950 | list_for_each_entry(remote_adev, |
3f975d0f SY |
1951 | device_list_handle, gmc.xgmi.head) { |
1952 | amdgpu_ras_query_err_status(remote_adev); | |
f75e94d8 | 1953 | amdgpu_ras_log_on_err_counter(remote_adev); |
3f975d0f | 1954 | } |
d95e8e97 DL |
1955 | |
1956 | amdgpu_put_xgmi_hive(hive); | |
b3dbd6d3 | 1957 | } |
313c8fd3 | 1958 | |
f1549c09 LG |
1959 | if (amdgpu_device_should_recover_gpu(ras->adev)) { |
1960 | struct amdgpu_reset_context reset_context; | |
1961 | memset(&reset_context, 0, sizeof(reset_context)); | |
1962 | ||
1963 | reset_context.method = AMD_RESET_METHOD_NONE; | |
1964 | reset_context.reset_req_dev = adev; | |
1a11a65d YC |
1965 | |
1966 | /* Perform full reset in fatal error mode */ | |
1967 | if (!amdgpu_ras_is_poison_mode_supported(ras->adev)) | |
1968 | set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags); | |
1969 | else | |
1970 | clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags); | |
f1549c09 LG |
1971 | |
1972 | amdgpu_device_gpu_recover(ras->adev, NULL, &reset_context); | |
1973 | } | |
c030f2e4 | 1974 | atomic_set(&ras->in_recovery, 0); |
1975 | } | |
1976 | ||
c030f2e4 | 1977 | /* alloc/realloc bps array */ |
1978 | static int amdgpu_ras_realloc_eh_data_space(struct amdgpu_device *adev, | |
1979 | struct ras_err_handler_data *data, int pages) | |
1980 | { | |
1981 | unsigned int old_space = data->count + data->space_left; | |
1982 | unsigned int new_space = old_space + pages; | |
9dc23a63 TZ |
1983 | unsigned int align_space = ALIGN(new_space, 512); |
1984 | void *bps = kmalloc(align_space * sizeof(*data->bps), GFP_KERNEL); | |
9dc23a63 | 1985 | |
676deb38 | 1986 | if (!bps) { |
c030f2e4 | 1987 | return -ENOMEM; |
9dc23a63 | 1988 | } |
c030f2e4 | 1989 | |
1990 | if (data->bps) { | |
9dc23a63 | 1991 | memcpy(bps, data->bps, |
c030f2e4 | 1992 | data->count * sizeof(*data->bps)); |
1993 | kfree(data->bps); | |
1994 | } | |
1995 | ||
9dc23a63 | 1996 | data->bps = bps; |
c030f2e4 | 1997 | data->space_left += align_space - old_space; |
1998 | return 0; | |
1999 | } | |
2000 | ||
2001 | /* it deal with vram only. */ | |
2002 | int amdgpu_ras_add_bad_pages(struct amdgpu_device *adev, | |
9dc23a63 | 2003 | struct eeprom_table_record *bps, int pages) |
c030f2e4 | 2004 | { |
2005 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
73aa8e1a | 2006 | struct ras_err_handler_data *data; |
c030f2e4 | 2007 | int ret = 0; |
676deb38 | 2008 | uint32_t i; |
c030f2e4 | 2009 | |
73aa8e1a | 2010 | if (!con || !con->eh_data || !bps || pages <= 0) |
c030f2e4 | 2011 | return 0; |
2012 | ||
2013 | mutex_lock(&con->recovery_lock); | |
73aa8e1a | 2014 | data = con->eh_data; |
c030f2e4 | 2015 | if (!data) |
2016 | goto out; | |
2017 | ||
676deb38 DL |
2018 | for (i = 0; i < pages; i++) { |
2019 | if (amdgpu_ras_check_bad_page_unlock(con, | |
2020 | bps[i].retired_page << AMDGPU_GPU_PAGE_SHIFT)) | |
2021 | continue; | |
2022 | ||
2023 | if (!data->space_left && | |
2024 | amdgpu_ras_realloc_eh_data_space(adev, data, 256)) { | |
c030f2e4 | 2025 | ret = -ENOMEM; |
2026 | goto out; | |
2027 | } | |
2028 | ||
ec6aae97 | 2029 | amdgpu_vram_mgr_reserve_range(&adev->mman.vram_mgr, |
676deb38 DL |
2030 | bps[i].retired_page << AMDGPU_GPU_PAGE_SHIFT, |
2031 | AMDGPU_GPU_PAGE_SIZE); | |
9dc23a63 | 2032 | |
676deb38 DL |
2033 | memcpy(&data->bps[data->count], &bps[i], sizeof(*data->bps)); |
2034 | data->count++; | |
2035 | data->space_left--; | |
2036 | } | |
c030f2e4 | 2037 | out: |
2038 | mutex_unlock(&con->recovery_lock); | |
2039 | ||
2040 | return ret; | |
2041 | } | |
2042 | ||
78ad00c9 TZ |
2043 | /* |
2044 | * write error record array to eeprom, the function should be | |
2045 | * protected by recovery_lock | |
2046 | */ | |
22503d80 | 2047 | int amdgpu_ras_save_bad_pages(struct amdgpu_device *adev) |
78ad00c9 TZ |
2048 | { |
2049 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
2050 | struct ras_err_handler_data *data; | |
8a3e801f | 2051 | struct amdgpu_ras_eeprom_control *control; |
78ad00c9 TZ |
2052 | int save_count; |
2053 | ||
2054 | if (!con || !con->eh_data) | |
2055 | return 0; | |
2056 | ||
d9a69fe5 | 2057 | mutex_lock(&con->recovery_lock); |
8a3e801f | 2058 | control = &con->eeprom_control; |
78ad00c9 | 2059 | data = con->eh_data; |
0686627b | 2060 | save_count = data->count - control->ras_num_recs; |
d9a69fe5 | 2061 | mutex_unlock(&con->recovery_lock); |
78ad00c9 | 2062 | /* only new entries are saved */ |
b1628425 | 2063 | if (save_count > 0) { |
63d4c081 LT |
2064 | if (amdgpu_ras_eeprom_append(control, |
2065 | &data->bps[control->ras_num_recs], | |
2066 | save_count)) { | |
6952e99c | 2067 | dev_err(adev->dev, "Failed to save EEPROM table data!"); |
78ad00c9 TZ |
2068 | return -EIO; |
2069 | } | |
2070 | ||
b1628425 GC |
2071 | dev_info(adev->dev, "Saved %d pages to EEPROM table.\n", save_count); |
2072 | } | |
2073 | ||
78ad00c9 TZ |
2074 | return 0; |
2075 | } | |
2076 | ||
2077 | /* | |
2078 | * read error record array in eeprom and reserve enough space for | |
2079 | * storing new bad pages | |
2080 | */ | |
2081 | static int amdgpu_ras_load_bad_pages(struct amdgpu_device *adev) | |
2082 | { | |
2083 | struct amdgpu_ras_eeprom_control *control = | |
6457205c | 2084 | &adev->psp.ras_context.ras->eeprom_control; |
e4e6a589 LT |
2085 | struct eeprom_table_record *bps; |
2086 | int ret; | |
78ad00c9 TZ |
2087 | |
2088 | /* no bad page record, skip eeprom access */ | |
0686627b | 2089 | if (control->ras_num_recs == 0 || amdgpu_bad_page_threshold == 0) |
e4e6a589 | 2090 | return 0; |
78ad00c9 | 2091 | |
0686627b | 2092 | bps = kcalloc(control->ras_num_recs, sizeof(*bps), GFP_KERNEL); |
78ad00c9 TZ |
2093 | if (!bps) |
2094 | return -ENOMEM; | |
2095 | ||
0686627b | 2096 | ret = amdgpu_ras_eeprom_read(control, bps, control->ras_num_recs); |
e4e6a589 | 2097 | if (ret) |
6952e99c | 2098 | dev_err(adev->dev, "Failed to load EEPROM table records!"); |
e4e6a589 | 2099 | else |
0686627b | 2100 | ret = amdgpu_ras_add_bad_pages(adev, bps, control->ras_num_recs); |
78ad00c9 | 2101 | |
78ad00c9 TZ |
2102 | kfree(bps); |
2103 | return ret; | |
2104 | } | |
2105 | ||
676deb38 DL |
2106 | static bool amdgpu_ras_check_bad_page_unlock(struct amdgpu_ras *con, |
2107 | uint64_t addr) | |
2108 | { | |
2109 | struct ras_err_handler_data *data = con->eh_data; | |
2110 | int i; | |
2111 | ||
2112 | addr >>= AMDGPU_GPU_PAGE_SHIFT; | |
2113 | for (i = 0; i < data->count; i++) | |
2114 | if (addr == data->bps[i].retired_page) | |
2115 | return true; | |
2116 | ||
2117 | return false; | |
2118 | } | |
2119 | ||
6e4be987 TZ |
2120 | /* |
2121 | * check if an address belongs to bad page | |
2122 | * | |
2123 | * Note: this check is only for umc block | |
2124 | */ | |
2125 | static bool amdgpu_ras_check_bad_page(struct amdgpu_device *adev, | |
2126 | uint64_t addr) | |
2127 | { | |
2128 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
6e4be987 TZ |
2129 | bool ret = false; |
2130 | ||
2131 | if (!con || !con->eh_data) | |
2132 | return ret; | |
2133 | ||
2134 | mutex_lock(&con->recovery_lock); | |
676deb38 | 2135 | ret = amdgpu_ras_check_bad_page_unlock(con, addr); |
6e4be987 TZ |
2136 | mutex_unlock(&con->recovery_lock); |
2137 | return ret; | |
2138 | } | |
2139 | ||
e5c04edf | 2140 | static void amdgpu_ras_validate_threshold(struct amdgpu_device *adev, |
e4e6a589 | 2141 | uint32_t max_count) |
c84d4670 | 2142 | { |
e5c04edf | 2143 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); |
c84d4670 GC |
2144 | |
2145 | /* | |
2146 | * Justification of value bad_page_cnt_threshold in ras structure | |
2147 | * | |
2148 | * Generally, -1 <= amdgpu_bad_page_threshold <= max record length | |
2149 | * in eeprom, and introduce two scenarios accordingly. | |
2150 | * | |
2151 | * Bad page retirement enablement: | |
2152 | * - If amdgpu_bad_page_threshold = -1, | |
2153 | * bad_page_cnt_threshold = typical value by formula. | |
2154 | * | |
2155 | * - When the value from user is 0 < amdgpu_bad_page_threshold < | |
2156 | * max record length in eeprom, use it directly. | |
2157 | * | |
2158 | * Bad page retirement disablement: | |
2159 | * - If amdgpu_bad_page_threshold = 0, bad page retirement | |
2160 | * functionality is disabled, and bad_page_cnt_threshold will | |
2161 | * take no effect. | |
2162 | */ | |
2163 | ||
e4e6a589 LT |
2164 | if (amdgpu_bad_page_threshold < 0) { |
2165 | u64 val = adev->gmc.mc_vram_size; | |
c84d4670 | 2166 | |
e4e6a589 | 2167 | do_div(val, RAS_BAD_PAGE_COVER); |
e5c04edf | 2168 | con->bad_page_cnt_threshold = min(lower_32_bits(val), |
e4e6a589 | 2169 | max_count); |
e5c04edf | 2170 | } else { |
e4e6a589 LT |
2171 | con->bad_page_cnt_threshold = min_t(int, max_count, |
2172 | amdgpu_bad_page_threshold); | |
c84d4670 GC |
2173 | } |
2174 | } | |
2175 | ||
1a6fc071 | 2176 | int amdgpu_ras_recovery_init(struct amdgpu_device *adev) |
c030f2e4 | 2177 | { |
2178 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
4d1337d2 | 2179 | struct ras_err_handler_data **data; |
e4e6a589 | 2180 | u32 max_eeprom_records_count = 0; |
b82e65a9 | 2181 | bool exc_err_limit = false; |
78ad00c9 | 2182 | int ret; |
c030f2e4 | 2183 | |
e0e146d5 | 2184 | if (!con || amdgpu_sriov_vf(adev)) |
1d9d2ca8 LT |
2185 | return 0; |
2186 | ||
2187 | /* Allow access to RAS EEPROM via debugfs, when the ASIC | |
2188 | * supports RAS and debugfs is enabled, but when | |
2189 | * adev->ras_enabled is unset, i.e. when "ras_enable" | |
2190 | * module parameter is set to 0. | |
2191 | */ | |
2192 | con->adev = adev; | |
2193 | ||
2194 | if (!adev->ras_enabled) | |
4d1337d2 AG |
2195 | return 0; |
2196 | ||
1d9d2ca8 | 2197 | data = &con->eh_data; |
1a6fc071 TZ |
2198 | *data = kmalloc(sizeof(**data), GFP_KERNEL | __GFP_ZERO); |
2199 | if (!*data) { | |
2200 | ret = -ENOMEM; | |
2201 | goto out; | |
2202 | } | |
c030f2e4 | 2203 | |
2204 | mutex_init(&con->recovery_lock); | |
2205 | INIT_WORK(&con->recovery_work, amdgpu_ras_do_recovery); | |
2206 | atomic_set(&con->in_recovery, 0); | |
69691c82 | 2207 | con->eeprom_control.bad_channel_bitmap = 0; |
c030f2e4 | 2208 | |
e4e6a589 LT |
2209 | max_eeprom_records_count = amdgpu_ras_eeprom_max_record_count(); |
2210 | amdgpu_ras_validate_threshold(adev, max_eeprom_records_count); | |
c84d4670 | 2211 | |
e5086659 | 2212 | /* Todo: During test the SMU might fail to read the eeprom through I2C |
2213 | * when the GPU is pending on XGMI reset during probe time | |
2214 | * (Mostly after second bus reset), skip it now | |
2215 | */ | |
2216 | if (adev->gmc.xgmi.pending_reset) | |
2217 | return 0; | |
b82e65a9 GC |
2218 | ret = amdgpu_ras_eeprom_init(&con->eeprom_control, &exc_err_limit); |
2219 | /* | |
2220 | * This calling fails when exc_err_limit is true or | |
2221 | * ret != 0. | |
2222 | */ | |
2223 | if (exc_err_limit || ret) | |
1a6fc071 | 2224 | goto free; |
78ad00c9 | 2225 | |
0686627b | 2226 | if (con->eeprom_control.ras_num_recs) { |
78ad00c9 TZ |
2227 | ret = amdgpu_ras_load_bad_pages(adev); |
2228 | if (ret) | |
1a6fc071 | 2229 | goto free; |
513befa6 | 2230 | |
bc143d8b | 2231 | amdgpu_dpm_send_hbm_bad_pages_num(adev, con->eeprom_control.ras_num_recs); |
69691c82 SY |
2232 | |
2233 | if (con->update_channel_flag == true) { | |
2234 | amdgpu_dpm_send_hbm_bad_channel_flag(adev, con->eeprom_control.bad_channel_bitmap); | |
2235 | con->update_channel_flag = false; | |
2236 | } | |
78ad00c9 | 2237 | } |
c030f2e4 | 2238 | |
12b2cab7 MJ |
2239 | #ifdef CONFIG_X86_MCE_AMD |
2240 | if ((adev->asic_type == CHIP_ALDEBARAN) && | |
2241 | (adev->gmc.xgmi.connected_to_cpu)) | |
91a1a52d | 2242 | amdgpu_register_bad_pages_mca_notifier(adev); |
12b2cab7 | 2243 | #endif |
c030f2e4 | 2244 | return 0; |
1a6fc071 | 2245 | |
1a6fc071 | 2246 | free: |
1a6fc071 | 2247 | kfree((*data)->bps); |
1a6fc071 | 2248 | kfree(*data); |
1995b3a3 | 2249 | con->eh_data = NULL; |
1a6fc071 | 2250 | out: |
cf696091 | 2251 | dev_warn(adev->dev, "Failed to initialize ras recovery! (%d)\n", ret); |
1a6fc071 | 2252 | |
b82e65a9 GC |
2253 | /* |
2254 | * Except error threshold exceeding case, other failure cases in this | |
2255 | * function would not fail amdgpu driver init. | |
2256 | */ | |
2257 | if (!exc_err_limit) | |
2258 | ret = 0; | |
2259 | else | |
2260 | ret = -EINVAL; | |
2261 | ||
1a6fc071 | 2262 | return ret; |
c030f2e4 | 2263 | } |
2264 | ||
2265 | static int amdgpu_ras_recovery_fini(struct amdgpu_device *adev) | |
2266 | { | |
2267 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
2268 | struct ras_err_handler_data *data = con->eh_data; | |
2269 | ||
1a6fc071 TZ |
2270 | /* recovery_init failed to init it, fini is useless */ |
2271 | if (!data) | |
2272 | return 0; | |
2273 | ||
c030f2e4 | 2274 | cancel_work_sync(&con->recovery_work); |
c030f2e4 | 2275 | |
2276 | mutex_lock(&con->recovery_lock); | |
2277 | con->eh_data = NULL; | |
2278 | kfree(data->bps); | |
2279 | kfree(data); | |
2280 | mutex_unlock(&con->recovery_lock); | |
2281 | ||
2282 | return 0; | |
2283 | } | |
2284 | /* recovery end */ | |
2285 | ||
084e2640 | 2286 | static bool amdgpu_ras_asic_supported(struct amdgpu_device *adev) |
5436ab94 | 2287 | { |
82835055 YC |
2288 | if (amdgpu_sriov_vf(adev)) { |
2289 | switch (adev->ip_versions[MP0_HWIP][0]) { | |
2290 | case IP_VERSION(13, 0, 2): | |
2291 | return true; | |
2292 | default: | |
2293 | return false; | |
2294 | } | |
2295 | } | |
2296 | ||
073285ef YC |
2297 | if (adev->asic_type == CHIP_IP_DISCOVERY) { |
2298 | switch (adev->ip_versions[MP0_HWIP][0]) { | |
2299 | case IP_VERSION(13, 0, 0): | |
2300 | case IP_VERSION(13, 0, 10): | |
2301 | return true; | |
2302 | default: | |
2303 | return false; | |
2304 | } | |
2305 | } | |
2306 | ||
084e2640 LT |
2307 | return adev->asic_type == CHIP_VEGA10 || |
2308 | adev->asic_type == CHIP_VEGA20 || | |
2309 | adev->asic_type == CHIP_ARCTURUS || | |
75f06251 | 2310 | adev->asic_type == CHIP_ALDEBARAN || |
084e2640 | 2311 | adev->asic_type == CHIP_SIENNA_CICHLID; |
5436ab94 SY |
2312 | } |
2313 | ||
f50160cf SY |
2314 | /* |
2315 | * this is workaround for vega20 workstation sku, | |
2316 | * force enable gfx ras, ignore vbios gfx ras flag | |
2317 | * due to GC EDC can not write | |
2318 | */ | |
e509965e | 2319 | static void amdgpu_ras_get_quirks(struct amdgpu_device *adev) |
f50160cf SY |
2320 | { |
2321 | struct atom_context *ctx = adev->mode_info.atom_context; | |
2322 | ||
2323 | if (!ctx) | |
2324 | return; | |
2325 | ||
2326 | if (strnstr(ctx->vbios_version, "D16406", | |
e11d5e0d SY |
2327 | sizeof(ctx->vbios_version)) || |
2328 | strnstr(ctx->vbios_version, "D36002", | |
2329 | sizeof(ctx->vbios_version))) | |
8ab0d6f0 | 2330 | adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX); |
f50160cf SY |
2331 | } |
2332 | ||
5caf466a | 2333 | /* |
2334 | * check hardware's ras ability which will be saved in hw_supported. | |
2335 | * if hardware does not support ras, we can skip some ras initializtion and | |
2336 | * forbid some ras operations from IP. | |
2337 | * if software itself, say boot parameter, limit the ras ability. We still | |
2338 | * need allow IP do some limited operations, like disable. In such case, | |
2339 | * we have to initialize ras as normal. but need check if operation is | |
2340 | * allowed or not in each function. | |
2341 | */ | |
e509965e | 2342 | static void amdgpu_ras_check_supported(struct amdgpu_device *adev) |
c030f2e4 | 2343 | { |
8ab0d6f0 | 2344 | adev->ras_hw_enabled = adev->ras_enabled = 0; |
c030f2e4 | 2345 | |
950d6425 | 2346 | if (!adev->is_atom_fw || |
084e2640 | 2347 | !amdgpu_ras_asic_supported(adev)) |
5caf466a | 2348 | return; |
b404ae82 | 2349 | |
75f06251 HZ |
2350 | if (!adev->gmc.xgmi.connected_to_cpu) { |
2351 | if (amdgpu_atomfirmware_mem_ecc_supported(adev)) { | |
2352 | dev_info(adev->dev, "MEM ECC is active.\n"); | |
8ab0d6f0 | 2353 | adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__UMC | |
e509965e | 2354 | 1 << AMDGPU_RAS_BLOCK__DF); |
75f06251 HZ |
2355 | } else { |
2356 | dev_info(adev->dev, "MEM ECC is not presented.\n"); | |
2357 | } | |
88474cca | 2358 | |
75f06251 HZ |
2359 | if (amdgpu_atomfirmware_sram_ecc_supported(adev)) { |
2360 | dev_info(adev->dev, "SRAM ECC is active.\n"); | |
3189501e | 2361 | if (!amdgpu_sriov_vf(adev)) |
950d6425 SY |
2362 | adev->ras_hw_enabled |= ~(1 << AMDGPU_RAS_BLOCK__UMC | |
2363 | 1 << AMDGPU_RAS_BLOCK__DF); | |
3189501e | 2364 | else |
950d6425 SY |
2365 | adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__PCIE_BIF | |
2366 | 1 << AMDGPU_RAS_BLOCK__SDMA | | |
2367 | 1 << AMDGPU_RAS_BLOCK__GFX); | |
3189501e TZ |
2368 | |
2369 | /* VCN/JPEG RAS can be supported on both bare metal and | |
2370 | * SRIOV environment | |
2371 | */ | |
2372 | if (adev->ip_versions[VCN_HWIP][0] == IP_VERSION(2, 6, 0) || | |
2373 | adev->ip_versions[VCN_HWIP][0] == IP_VERSION(4, 0, 0)) | |
2374 | adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__VCN | | |
2375 | 1 << AMDGPU_RAS_BLOCK__JPEG); | |
2376 | else | |
2377 | adev->ras_hw_enabled &= ~(1 << AMDGPU_RAS_BLOCK__VCN | | |
2378 | 1 << AMDGPU_RAS_BLOCK__JPEG); | |
75f06251 HZ |
2379 | } else { |
2380 | dev_info(adev->dev, "SRAM ECC is not presented.\n"); | |
2381 | } | |
2382 | } else { | |
2383 | /* driver only manages a few IP blocks RAS feature | |
2384 | * when GPU is connected cpu through XGMI */ | |
8ab0d6f0 | 2385 | adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX | |
e509965e LT |
2386 | 1 << AMDGPU_RAS_BLOCK__SDMA | |
2387 | 1 << AMDGPU_RAS_BLOCK__MMHUB); | |
75f06251 | 2388 | } |
88474cca | 2389 | |
e509965e | 2390 | amdgpu_ras_get_quirks(adev); |
f50160cf | 2391 | |
88474cca | 2392 | /* hw_supported needs to be aligned with RAS block mask. */ |
8ab0d6f0 | 2393 | adev->ras_hw_enabled &= AMDGPU_RAS_BLOCK_MASK; |
b404ae82 | 2394 | |
8ab0d6f0 LT |
2395 | adev->ras_enabled = amdgpu_ras_enable == 0 ? 0 : |
2396 | adev->ras_hw_enabled & amdgpu_ras_mask; | |
c030f2e4 | 2397 | } |
2398 | ||
05adfd80 LT |
2399 | static void amdgpu_ras_counte_dw(struct work_struct *work) |
2400 | { | |
2401 | struct amdgpu_ras *con = container_of(work, struct amdgpu_ras, | |
2402 | ras_counte_delay_work.work); | |
2403 | struct amdgpu_device *adev = con->adev; | |
a3fbb0d8 | 2404 | struct drm_device *dev = adev_to_drm(adev); |
05adfd80 LT |
2405 | unsigned long ce_count, ue_count; |
2406 | int res; | |
2407 | ||
2408 | res = pm_runtime_get_sync(dev->dev); | |
2409 | if (res < 0) | |
2410 | goto Out; | |
2411 | ||
2412 | /* Cache new values. | |
2413 | */ | |
4d9f771e LT |
2414 | if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count) == 0) { |
2415 | atomic_set(&con->ras_ce_count, ce_count); | |
2416 | atomic_set(&con->ras_ue_count, ue_count); | |
2417 | } | |
05adfd80 LT |
2418 | |
2419 | pm_runtime_mark_last_busy(dev->dev); | |
2420 | Out: | |
2421 | pm_runtime_put_autosuspend(dev->dev); | |
2422 | } | |
2423 | ||
2dd9032b TZ |
2424 | static void amdgpu_ras_query_poison_mode(struct amdgpu_device *adev) |
2425 | { | |
2426 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
2427 | bool df_poison, umc_poison; | |
2428 | ||
2429 | /* poison setting is useless on SRIOV guest */ | |
2430 | if (amdgpu_sriov_vf(adev) || !con) | |
2431 | return; | |
2432 | ||
2433 | /* Init poison supported flag, the default value is false */ | |
2434 | if (adev->gmc.xgmi.connected_to_cpu) { | |
2435 | /* enabled by default when GPU is connected to CPU */ | |
2436 | con->poison_supported = true; | |
2437 | } else if (adev->df.funcs && | |
2438 | adev->df.funcs->query_ras_poison_mode && | |
2439 | adev->umc.ras && | |
2440 | adev->umc.ras->query_ras_poison_mode) { | |
2441 | df_poison = | |
2442 | adev->df.funcs->query_ras_poison_mode(adev); | |
2443 | umc_poison = | |
2444 | adev->umc.ras->query_ras_poison_mode(adev); | |
2445 | ||
2446 | /* Only poison is set in both DF and UMC, we can support it */ | |
2447 | if (df_poison && umc_poison) | |
2448 | con->poison_supported = true; | |
2449 | else if (df_poison != umc_poison) | |
2450 | dev_warn(adev->dev, | |
2451 | "Poison setting is inconsistent in DF/UMC(%d:%d)!\n", | |
2452 | df_poison, umc_poison); | |
2453 | } | |
2454 | } | |
2455 | ||
c030f2e4 | 2456 | int amdgpu_ras_init(struct amdgpu_device *adev) |
2457 | { | |
2458 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
4e644fff | 2459 | int r; |
c030f2e4 | 2460 | |
b404ae82 | 2461 | if (con) |
c030f2e4 | 2462 | return 0; |
2463 | ||
2464 | con = kmalloc(sizeof(struct amdgpu_ras) + | |
640ae42e JC |
2465 | sizeof(struct ras_manager) * AMDGPU_RAS_BLOCK_COUNT + |
2466 | sizeof(struct ras_manager) * AMDGPU_RAS_MCA_BLOCK_COUNT, | |
c030f2e4 | 2467 | GFP_KERNEL|__GFP_ZERO); |
2468 | if (!con) | |
2469 | return -ENOMEM; | |
2470 | ||
05adfd80 LT |
2471 | con->adev = adev; |
2472 | INIT_DELAYED_WORK(&con->ras_counte_delay_work, amdgpu_ras_counte_dw); | |
2473 | atomic_set(&con->ras_ce_count, 0); | |
2474 | atomic_set(&con->ras_ue_count, 0); | |
2475 | ||
c030f2e4 | 2476 | con->objs = (struct ras_manager *)(con + 1); |
2477 | ||
2478 | amdgpu_ras_set_context(adev, con); | |
2479 | ||
e509965e LT |
2480 | amdgpu_ras_check_supported(adev); |
2481 | ||
7ddd9770 | 2482 | if (!adev->ras_enabled || adev->asic_type == CHIP_VEGA10) { |
970fd197 SY |
2483 | /* set gfx block ras context feature for VEGA20 Gaming |
2484 | * send ras disable cmd to ras ta during ras late init. | |
2485 | */ | |
8ab0d6f0 | 2486 | if (!adev->ras_enabled && adev->asic_type == CHIP_VEGA20) { |
970fd197 SY |
2487 | con->features |= BIT(AMDGPU_RAS_BLOCK__GFX); |
2488 | ||
2489 | return 0; | |
2490 | } | |
2491 | ||
5e91160a | 2492 | r = 0; |
5436ab94 | 2493 | goto release_con; |
fb2a3607 HZ |
2494 | } |
2495 | ||
69691c82 | 2496 | con->update_channel_flag = false; |
c030f2e4 | 2497 | con->features = 0; |
2498 | INIT_LIST_HEAD(&con->head); | |
108c6a63 | 2499 | /* Might need get this flag from vbios. */ |
2500 | con->flags = RAS_DEFAULT_FLAGS; | |
c030f2e4 | 2501 | |
6e36f231 HZ |
2502 | /* initialize nbio ras function ahead of any other |
2503 | * ras functions so hardware fatal error interrupt | |
2504 | * can be enabled as early as possible */ | |
2505 | switch (adev->asic_type) { | |
2506 | case CHIP_VEGA20: | |
2507 | case CHIP_ARCTURUS: | |
2508 | case CHIP_ALDEBARAN: | |
2e54fe5d | 2509 | if (!adev->gmc.xgmi.connected_to_cpu) { |
2510 | adev->nbio.ras = &nbio_v7_4_ras; | |
2511 | amdgpu_ras_register_ras_block(adev, &adev->nbio.ras->ras_block); | |
80ed77f9 | 2512 | adev->nbio.ras_if = &adev->nbio.ras->ras_block.ras_comm; |
2e54fe5d | 2513 | } |
6e36f231 HZ |
2514 | break; |
2515 | default: | |
2516 | /* nbio ras is not available */ | |
2517 | break; | |
2518 | } | |
2519 | ||
2e54fe5d | 2520 | if (adev->nbio.ras && |
2521 | adev->nbio.ras->init_ras_controller_interrupt) { | |
2522 | r = adev->nbio.ras->init_ras_controller_interrupt(adev); | |
4e644fff | 2523 | if (r) |
5436ab94 | 2524 | goto release_con; |
4e644fff HZ |
2525 | } |
2526 | ||
2e54fe5d | 2527 | if (adev->nbio.ras && |
2528 | adev->nbio.ras->init_ras_err_event_athub_interrupt) { | |
2529 | r = adev->nbio.ras->init_ras_err_event_athub_interrupt(adev); | |
4e644fff | 2530 | if (r) |
5436ab94 | 2531 | goto release_con; |
4e644fff HZ |
2532 | } |
2533 | ||
2dd9032b | 2534 | amdgpu_ras_query_poison_mode(adev); |
e4348849 | 2535 | |
5e91160a GC |
2536 | if (amdgpu_ras_fs_init(adev)) { |
2537 | r = -EINVAL; | |
5436ab94 | 2538 | goto release_con; |
5e91160a | 2539 | } |
c030f2e4 | 2540 | |
6952e99c | 2541 | dev_info(adev->dev, "RAS INFO: ras initialized successfully, " |
e509965e | 2542 | "hardware ability[%x] ras_mask[%x]\n", |
8ab0d6f0 | 2543 | adev->ras_hw_enabled, adev->ras_enabled); |
e509965e | 2544 | |
c030f2e4 | 2545 | return 0; |
5436ab94 | 2546 | release_con: |
c030f2e4 | 2547 | amdgpu_ras_set_context(adev, NULL); |
2548 | kfree(con); | |
2549 | ||
5e91160a | 2550 | return r; |
c030f2e4 | 2551 | } |
2552 | ||
8f6368a9 | 2553 | int amdgpu_persistent_edc_harvesting_supported(struct amdgpu_device *adev) |
134d16d5 JC |
2554 | { |
2555 | if (adev->gmc.xgmi.connected_to_cpu) | |
2556 | return 1; | |
2557 | return 0; | |
2558 | } | |
2559 | ||
2560 | static int amdgpu_persistent_edc_harvesting(struct amdgpu_device *adev, | |
2561 | struct ras_common_if *ras_block) | |
2562 | { | |
2563 | struct ras_query_if info = { | |
2564 | .head = *ras_block, | |
2565 | }; | |
2566 | ||
2567 | if (!amdgpu_persistent_edc_harvesting_supported(adev)) | |
2568 | return 0; | |
2569 | ||
2570 | if (amdgpu_ras_query_error_status(adev, &info) != 0) | |
2571 | DRM_WARN("RAS init harvest failure"); | |
2572 | ||
2573 | if (amdgpu_ras_reset_error_status(adev, ras_block->block) != 0) | |
2574 | DRM_WARN("RAS init harvest reset failure"); | |
2575 | ||
2576 | return 0; | |
2577 | } | |
2578 | ||
e4348849 TZ |
2579 | bool amdgpu_ras_is_poison_mode_supported(struct amdgpu_device *adev) |
2580 | { | |
2581 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
2582 | ||
2583 | if (!con) | |
2584 | return false; | |
2585 | ||
2586 | return con->poison_supported; | |
2587 | } | |
2588 | ||
b293e891 | 2589 | /* helper function to handle common stuff in ip late init phase */ |
563285c8 | 2590 | int amdgpu_ras_block_late_init(struct amdgpu_device *adev, |
2591 | struct ras_common_if *ras_block) | |
b293e891 | 2592 | { |
29c9b6cd | 2593 | struct amdgpu_ras_block_object *ras_obj = NULL; |
05adfd80 LT |
2594 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); |
2595 | unsigned long ue_count, ce_count; | |
b293e891 HZ |
2596 | int r; |
2597 | ||
2598 | /* disable RAS feature per IP block if it is not supported */ | |
2599 | if (!amdgpu_ras_is_supported(adev, ras_block->block)) { | |
2600 | amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0); | |
2601 | return 0; | |
2602 | } | |
2603 | ||
2604 | r = amdgpu_ras_feature_enable_on_boot(adev, ras_block, 1); | |
2605 | if (r) { | |
9080a18f | 2606 | if (adev->in_suspend || amdgpu_in_reset(adev)) { |
b293e891 HZ |
2607 | /* in resume phase, if fail to enable ras, |
2608 | * clean up all ras fs nodes, and disable ras */ | |
2609 | goto cleanup; | |
2610 | } else | |
2611 | return r; | |
2612 | } | |
2613 | ||
134d16d5 JC |
2614 | /* check for errors on warm reset edc persisant supported ASIC */ |
2615 | amdgpu_persistent_edc_harvesting(adev, ras_block); | |
2616 | ||
b293e891 | 2617 | /* in resume phase, no need to create ras fs node */ |
53b3f8f4 | 2618 | if (adev->in_suspend || amdgpu_in_reset(adev)) |
b293e891 HZ |
2619 | return 0; |
2620 | ||
563285c8 | 2621 | ras_obj = container_of(ras_block, struct amdgpu_ras_block_object, ras_comm); |
36780606 TZ |
2622 | if (ras_obj->ras_cb || (ras_obj->hw_ops && |
2623 | (ras_obj->hw_ops->query_poison_status || | |
2624 | ras_obj->hw_ops->handle_poison_consumption))) { | |
9252d33d | 2625 | r = amdgpu_ras_interrupt_add_handler(adev, ras_block); |
b293e891 | 2626 | if (r) |
779596ce | 2627 | goto cleanup; |
b293e891 HZ |
2628 | } |
2629 | ||
9252d33d | 2630 | r = amdgpu_ras_sysfs_create(adev, ras_block); |
b293e891 | 2631 | if (r) |
779596ce | 2632 | goto interrupt; |
b293e891 | 2633 | |
05adfd80 LT |
2634 | /* Those are the cached values at init. |
2635 | */ | |
4d9f771e LT |
2636 | if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count) == 0) { |
2637 | atomic_set(&con->ras_ce_count, ce_count); | |
2638 | atomic_set(&con->ras_ue_count, ue_count); | |
2639 | } | |
05adfd80 | 2640 | |
b293e891 | 2641 | return 0; |
779596ce TR |
2642 | |
2643 | interrupt: | |
563285c8 | 2644 | if (ras_obj->ras_cb) |
9252d33d | 2645 | amdgpu_ras_interrupt_remove_handler(adev, ras_block); |
779596ce | 2646 | cleanup: |
b293e891 HZ |
2647 | amdgpu_ras_feature_enable(adev, ras_block, 0); |
2648 | return r; | |
2649 | } | |
2650 | ||
d41ff22a | 2651 | static int amdgpu_ras_block_late_init_default(struct amdgpu_device *adev, |
418abce2 | 2652 | struct ras_common_if *ras_block) |
2653 | { | |
2654 | return amdgpu_ras_block_late_init(adev, ras_block); | |
2655 | } | |
2656 | ||
b293e891 | 2657 | /* helper function to remove ras fs node and interrupt handler */ |
bdb3489c | 2658 | void amdgpu_ras_block_late_fini(struct amdgpu_device *adev, |
2659 | struct ras_common_if *ras_block) | |
2660 | { | |
563285c8 | 2661 | struct amdgpu_ras_block_object *ras_obj; |
bdb3489c | 2662 | if (!ras_block) |
2663 | return; | |
2664 | ||
563285c8 | 2665 | amdgpu_ras_sysfs_remove(adev, ras_block); |
bdb3489c | 2666 | |
563285c8 | 2667 | ras_obj = container_of(ras_block, struct amdgpu_ras_block_object, ras_comm); |
2668 | if (ras_obj->ras_cb) | |
2669 | amdgpu_ras_interrupt_remove_handler(adev, ras_block); | |
bdb3489c | 2670 | } |
2671 | ||
80e0c2cb | 2672 | static void amdgpu_ras_block_late_fini_default(struct amdgpu_device *adev, |
2673 | struct ras_common_if *ras_block) | |
2674 | { | |
2675 | return amdgpu_ras_block_late_fini(adev, ras_block); | |
2676 | } | |
2677 | ||
a564808e | 2678 | /* do some init work after IP late init as dependence. |
511fdbc3 | 2679 | * and it runs in resume/gpu reset/booting up cases. |
a564808e | 2680 | */ |
511fdbc3 | 2681 | void amdgpu_ras_resume(struct amdgpu_device *adev) |
108c6a63 | 2682 | { |
2683 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
2684 | struct ras_manager *obj, *tmp; | |
2685 | ||
8ab0d6f0 | 2686 | if (!adev->ras_enabled || !con) { |
970fd197 SY |
2687 | /* clean ras context for VEGA20 Gaming after send ras disable cmd */ |
2688 | amdgpu_release_ras_context(adev); | |
2689 | ||
108c6a63 | 2690 | return; |
970fd197 | 2691 | } |
108c6a63 | 2692 | |
108c6a63 | 2693 | if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) { |
191051a1 | 2694 | /* Set up all other IPs which are not implemented. There is a |
2695 | * tricky thing that IP's actual ras error type should be | |
2696 | * MULTI_UNCORRECTABLE, but as driver does not handle it, so | |
2697 | * ERROR_NONE make sense anyway. | |
2698 | */ | |
2699 | amdgpu_ras_enable_all_features(adev, 1); | |
2700 | ||
2701 | /* We enable ras on all hw_supported block, but as boot | |
2702 | * parameter might disable some of them and one or more IP has | |
2703 | * not implemented yet. So we disable them on behalf. | |
2704 | */ | |
108c6a63 | 2705 | list_for_each_entry_safe(obj, tmp, &con->head, node) { |
2706 | if (!amdgpu_ras_is_supported(adev, obj->head.block)) { | |
2707 | amdgpu_ras_feature_enable(adev, &obj->head, 0); | |
2708 | /* there should be no any reference. */ | |
2709 | WARN_ON(alive_obj(obj)); | |
2710 | } | |
191051a1 | 2711 | } |
108c6a63 | 2712 | } |
2713 | } | |
2714 | ||
511fdbc3 | 2715 | void amdgpu_ras_suspend(struct amdgpu_device *adev) |
2716 | { | |
2717 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
2718 | ||
8ab0d6f0 | 2719 | if (!adev->ras_enabled || !con) |
511fdbc3 | 2720 | return; |
2721 | ||
2722 | amdgpu_ras_disable_all_features(adev, 0); | |
2723 | /* Make sure all ras objects are disabled. */ | |
2724 | if (con->features) | |
2725 | amdgpu_ras_disable_all_features(adev, 1); | |
2726 | } | |
2727 | ||
867e24ca | 2728 | int amdgpu_ras_late_init(struct amdgpu_device *adev) |
2729 | { | |
2730 | struct amdgpu_ras_block_list *node, *tmp; | |
2731 | struct amdgpu_ras_block_object *obj; | |
2732 | int r; | |
2733 | ||
950d6425 SY |
2734 | /* Guest side doesn't need init ras feature */ |
2735 | if (amdgpu_sriov_vf(adev)) | |
2736 | return 0; | |
2737 | ||
867e24ca | 2738 | list_for_each_entry_safe(node, tmp, &adev->ras_list, node) { |
2739 | if (!node->ras_obj) { | |
2740 | dev_warn(adev->dev, "Warning: abnormal ras list node.\n"); | |
2741 | continue; | |
2742 | } | |
418abce2 | 2743 | |
867e24ca | 2744 | obj = node->ras_obj; |
2745 | if (obj->ras_late_init) { | |
2746 | r = obj->ras_late_init(adev, &obj->ras_comm); | |
2747 | if (r) { | |
2748 | dev_err(adev->dev, "%s failed to execute ras_late_init! ret:%d\n", | |
2749 | obj->ras_comm.name, r); | |
2750 | return r; | |
2751 | } | |
418abce2 | 2752 | } else |
2753 | amdgpu_ras_block_late_init_default(adev, &obj->ras_comm); | |
867e24ca | 2754 | } |
2755 | ||
2756 | return 0; | |
2757 | } | |
2758 | ||
c030f2e4 | 2759 | /* do some fini work before IP fini as dependence */ |
2760 | int amdgpu_ras_pre_fini(struct amdgpu_device *adev) | |
2761 | { | |
2762 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
2763 | ||
8ab0d6f0 | 2764 | if (!adev->ras_enabled || !con) |
c030f2e4 | 2765 | return 0; |
2766 | ||
72c8c97b | 2767 | |
c030f2e4 | 2768 | /* Need disable ras on all IPs here before ip [hw/sw]fini */ |
642c0401 YC |
2769 | if (con->features) |
2770 | amdgpu_ras_disable_all_features(adev, 0); | |
c030f2e4 | 2771 | amdgpu_ras_recovery_fini(adev); |
2772 | return 0; | |
2773 | } | |
2774 | ||
2775 | int amdgpu_ras_fini(struct amdgpu_device *adev) | |
2776 | { | |
d5e8ff5f | 2777 | struct amdgpu_ras_block_list *ras_node, *tmp; |
1f211a82 | 2778 | struct amdgpu_ras_block_object *obj = NULL; |
c030f2e4 | 2779 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); |
2780 | ||
8ab0d6f0 | 2781 | if (!adev->ras_enabled || !con) |
c030f2e4 | 2782 | return 0; |
2783 | ||
1f211a82 | 2784 | list_for_each_entry_safe(ras_node, tmp, &adev->ras_list, node) { |
2785 | if (ras_node->ras_obj) { | |
2786 | obj = ras_node->ras_obj; | |
2787 | if (amdgpu_ras_is_supported(adev, obj->ras_comm.block) && | |
2788 | obj->ras_fini) | |
2789 | obj->ras_fini(adev, &obj->ras_comm); | |
80e0c2cb | 2790 | else |
2791 | amdgpu_ras_block_late_fini_default(adev, &obj->ras_comm); | |
1f211a82 | 2792 | } |
2793 | ||
2794 | /* Clear ras blocks from ras_list and free ras block list node */ | |
2795 | list_del(&ras_node->node); | |
2796 | kfree(ras_node); | |
2797 | } | |
2798 | ||
c030f2e4 | 2799 | amdgpu_ras_fs_fini(adev); |
2800 | amdgpu_ras_interrupt_remove_all(adev); | |
2801 | ||
2802 | WARN(con->features, "Feature mask is not cleared"); | |
2803 | ||
2804 | if (con->features) | |
2805 | amdgpu_ras_disable_all_features(adev, 1); | |
2806 | ||
05adfd80 LT |
2807 | cancel_delayed_work_sync(&con->ras_counte_delay_work); |
2808 | ||
c030f2e4 | 2809 | amdgpu_ras_set_context(adev, NULL); |
2810 | kfree(con); | |
2811 | ||
2812 | return 0; | |
2813 | } | |
7c6e68c7 AG |
2814 | |
2815 | void amdgpu_ras_global_ras_isr(struct amdgpu_device *adev) | |
2816 | { | |
e509965e | 2817 | amdgpu_ras_check_supported(adev); |
8ab0d6f0 | 2818 | if (!adev->ras_hw_enabled) |
ed606f8a AG |
2819 | return; |
2820 | ||
7c6e68c7 | 2821 | if (atomic_cmpxchg(&amdgpu_ras_in_intr, 0, 1) == 0) { |
6952e99c GC |
2822 | dev_info(adev->dev, "uncorrectable hardware error" |
2823 | "(ERREVENT_ATHUB_INTERRUPT) detected!\n"); | |
d5ea093e | 2824 | |
61934624 | 2825 | amdgpu_ras_reset_gpu(adev); |
7c6e68c7 AG |
2826 | } |
2827 | } | |
bb5c7235 WS |
2828 | |
2829 | bool amdgpu_ras_need_emergency_restart(struct amdgpu_device *adev) | |
2830 | { | |
2831 | if (adev->asic_type == CHIP_VEGA20 && | |
2832 | adev->pm.fw_version <= 0x283400) { | |
2833 | return !(amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) && | |
2834 | amdgpu_ras_intr_triggered(); | |
2835 | } | |
2836 | ||
2837 | return false; | |
2838 | } | |
970fd197 SY |
2839 | |
2840 | void amdgpu_release_ras_context(struct amdgpu_device *adev) | |
2841 | { | |
2842 | struct amdgpu_ras *con = amdgpu_ras_get_context(adev); | |
2843 | ||
2844 | if (!con) | |
2845 | return; | |
2846 | ||
8ab0d6f0 | 2847 | if (!adev->ras_enabled && con->features & BIT(AMDGPU_RAS_BLOCK__GFX)) { |
970fd197 SY |
2848 | con->features &= ~BIT(AMDGPU_RAS_BLOCK__GFX); |
2849 | amdgpu_ras_set_context(adev, NULL); | |
2850 | kfree(con); | |
2851 | } | |
2852 | } | |
12b2cab7 MJ |
2853 | |
2854 | #ifdef CONFIG_X86_MCE_AMD | |
2855 | static struct amdgpu_device *find_adev(uint32_t node_id) | |
2856 | { | |
12b2cab7 MJ |
2857 | int i; |
2858 | struct amdgpu_device *adev = NULL; | |
2859 | ||
91a1a52d MJ |
2860 | for (i = 0; i < mce_adev_list.num_gpu; i++) { |
2861 | adev = mce_adev_list.devs[i]; | |
12b2cab7 | 2862 | |
91a1a52d | 2863 | if (adev && adev->gmc.xgmi.connected_to_cpu && |
12b2cab7 MJ |
2864 | adev->gmc.xgmi.physical_node_id == node_id) |
2865 | break; | |
2866 | adev = NULL; | |
2867 | } | |
2868 | ||
12b2cab7 MJ |
2869 | return adev; |
2870 | } | |
2871 | ||
2872 | #define GET_MCA_IPID_GPUID(m) (((m) >> 44) & 0xF) | |
2873 | #define GET_UMC_INST(m) (((m) >> 21) & 0x7) | |
2874 | #define GET_CHAN_INDEX(m) ((((m) >> 12) & 0x3) | (((m) >> 18) & 0x4)) | |
2875 | #define GPU_ID_OFFSET 8 | |
2876 | ||
2877 | static int amdgpu_bad_page_notifier(struct notifier_block *nb, | |
2878 | unsigned long val, void *data) | |
2879 | { | |
2880 | struct mce *m = (struct mce *)data; | |
2881 | struct amdgpu_device *adev = NULL; | |
2882 | uint32_t gpu_id = 0; | |
cd4c99f1 | 2883 | uint32_t umc_inst = 0, ch_inst = 0; |
12b2cab7 MJ |
2884 | |
2885 | /* | |
2886 | * If the error was generated in UMC_V2, which belongs to GPU UMCs, | |
2887 | * and error occurred in DramECC (Extended error code = 0) then only | |
2888 | * process the error, else bail out. | |
2889 | */ | |
91f75eb4 | 2890 | if (!m || !((smca_get_bank_type(m->extcpu, m->bank) == SMCA_UMC_V2) && |
12b2cab7 MJ |
2891 | (XEC(m->status, 0x3f) == 0x0))) |
2892 | return NOTIFY_DONE; | |
2893 | ||
2894 | /* | |
2895 | * If it is correctable error, return. | |
2896 | */ | |
2897 | if (mce_is_correctable(m)) | |
2898 | return NOTIFY_OK; | |
2899 | ||
2900 | /* | |
2901 | * GPU Id is offset by GPU_ID_OFFSET in MCA_IPID_UMC register. | |
2902 | */ | |
2903 | gpu_id = GET_MCA_IPID_GPUID(m->ipid) - GPU_ID_OFFSET; | |
2904 | ||
2905 | adev = find_adev(gpu_id); | |
2906 | if (!adev) { | |
2907 | DRM_WARN("%s: Unable to find adev for gpu_id: %d\n", __func__, | |
2908 | gpu_id); | |
2909 | return NOTIFY_DONE; | |
2910 | } | |
2911 | ||
2912 | /* | |
2913 | * If it is uncorrectable error, then find out UMC instance and | |
2914 | * channel index. | |
2915 | */ | |
2916 | umc_inst = GET_UMC_INST(m->ipid); | |
2917 | ch_inst = GET_CHAN_INDEX(m->ipid); | |
2918 | ||
2919 | dev_info(adev->dev, "Uncorrectable error detected in UMC inst: %d, chan_idx: %d", | |
2920 | umc_inst, ch_inst); | |
2921 | ||
24b82292 TZ |
2922 | if (!amdgpu_umc_page_retirement_mca(adev, m->addr, ch_inst, umc_inst)) |
2923 | return NOTIFY_OK; | |
2924 | else | |
6c0ca748 | 2925 | return NOTIFY_DONE; |
12b2cab7 MJ |
2926 | } |
2927 | ||
2928 | static struct notifier_block amdgpu_bad_page_nb = { | |
2929 | .notifier_call = amdgpu_bad_page_notifier, | |
2930 | .priority = MCE_PRIO_UC, | |
2931 | }; | |
2932 | ||
91a1a52d | 2933 | static void amdgpu_register_bad_pages_mca_notifier(struct amdgpu_device *adev) |
12b2cab7 | 2934 | { |
91a1a52d MJ |
2935 | /* |
2936 | * Add the adev to the mce_adev_list. | |
2937 | * During mode2 reset, amdgpu device is temporarily | |
2938 | * removed from the mgpu_info list which can cause | |
2939 | * page retirement to fail. | |
2940 | * Use this list instead of mgpu_info to find the amdgpu | |
2941 | * device on which the UMC error was reported. | |
2942 | */ | |
2943 | mce_adev_list.devs[mce_adev_list.num_gpu++] = adev; | |
2944 | ||
12b2cab7 MJ |
2945 | /* |
2946 | * Register the x86 notifier only once | |
2947 | * with MCE subsystem. | |
2948 | */ | |
2949 | if (notifier_registered == false) { | |
2950 | mce_register_decode_chain(&amdgpu_bad_page_nb); | |
2951 | notifier_registered = true; | |
2952 | } | |
2953 | } | |
2954 | #endif | |
7cab2124 | 2955 | |
b6efdb02 | 2956 | struct amdgpu_ras *amdgpu_ras_get_context(struct amdgpu_device *adev) |
7cab2124 | 2957 | { |
2958 | if (!adev) | |
2959 | return NULL; | |
2960 | ||
2961 | return adev->psp.ras_context.ras; | |
2962 | } | |
2963 | ||
b6efdb02 | 2964 | int amdgpu_ras_set_context(struct amdgpu_device *adev, struct amdgpu_ras *ras_con) |
7cab2124 | 2965 | { |
2966 | if (!adev) | |
69f91d32 | 2967 | return -EINVAL; |
7cab2124 | 2968 | |
2969 | adev->psp.ras_context.ras = ras_con; | |
2970 | return 0; | |
2971 | } | |
2972 | ||
2973 | /* check if ras is supported on block, say, sdma, gfx */ | |
2974 | int amdgpu_ras_is_supported(struct amdgpu_device *adev, | |
2975 | unsigned int block) | |
2976 | { | |
2977 | struct amdgpu_ras *ras = amdgpu_ras_get_context(adev); | |
2978 | ||
2979 | if (block >= AMDGPU_RAS_BLOCK_COUNT) | |
2980 | return 0; | |
2981 | return ras && (adev->ras_enabled & (1 << block)); | |
2982 | } | |
2983 | ||
2984 | int amdgpu_ras_reset_gpu(struct amdgpu_device *adev) | |
2985 | { | |
2986 | struct amdgpu_ras *ras = amdgpu_ras_get_context(adev); | |
2987 | ||
2988 | if (atomic_cmpxchg(&ras->in_recovery, 0, 1) == 0) | |
25a2b22e | 2989 | amdgpu_reset_domain_schedule(ras->adev->reset_domain, &ras->recovery_work); |
7cab2124 | 2990 | return 0; |
2991 | } | |
2992 | ||
2993 | ||
6492e1b0 | 2994 | /* Register each ip ras block into amdgpu ras */ |
2995 | int amdgpu_ras_register_ras_block(struct amdgpu_device *adev, | |
b6efdb02 | 2996 | struct amdgpu_ras_block_object *ras_block_obj) |
6492e1b0 | 2997 | { |
d5e8ff5f | 2998 | struct amdgpu_ras_block_list *ras_node; |
6492e1b0 | 2999 | if (!adev || !ras_block_obj) |
3000 | return -EINVAL; | |
3001 | ||
df01fe73 | 3002 | if (!amdgpu_ras_asic_supported(adev)) |
3003 | return 0; | |
3004 | ||
d5e8ff5f | 3005 | ras_node = kzalloc(sizeof(*ras_node), GFP_KERNEL); |
3006 | if (!ras_node) | |
3007 | return -ENOMEM; | |
3008 | ||
3009 | INIT_LIST_HEAD(&ras_node->node); | |
3010 | ras_node->ras_obj = ras_block_obj; | |
3011 | list_add_tail(&ras_node->node, &adev->ras_list); | |
6492e1b0 | 3012 | |
3013 | return 0; | |
3014 | } |