2 * Copyright 2019 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
24 #include "amdgpu_ras_eeprom.h"
26 #include "amdgpu_ras.h"
27 #include <linux/bits.h>
30 #define EEPROM_I2C_TARGET_ADDR_VEGA20 0xA0
31 #define EEPROM_I2C_TARGET_ADDR_ARCTURUS 0xA8
32 #define EEPROM_I2C_TARGET_ADDR_ARCTURUS_D342 0xA0
33 #define EEPROM_I2C_TARGET_ADDR_SIENNA_CICHLID 0xA0
34 #define EEPROM_I2C_TARGET_ADDR_ALDEBARAN 0xA0
37 * The 2 macros bellow represent the actual size in bytes that
38 * those entities occupy in the EEPROM memory.
39 * EEPROM_TABLE_RECORD_SIZE is different than sizeof(eeprom_table_record) which
40 * uses uint64 to store 6b fields such as retired_page.
42 #define EEPROM_TABLE_HEADER_SIZE 20
43 #define EEPROM_TABLE_RECORD_SIZE 24
45 #define EEPROM_ADDRESS_SIZE 0x2
47 /* Table hdr is 'AMDR' */
48 #define EEPROM_TABLE_HDR_VAL 0x414d4452
49 #define EEPROM_TABLE_VER 0x00010000
51 /* Bad GPU tag ‘BADG’ */
52 #define EEPROM_TABLE_HDR_BAD 0x42414447
54 /* Assume 2 Mbit size */
55 #define EEPROM_SIZE_BYTES 256000
56 #define EEPROM_PAGE__SIZE_BYTES 256
57 #define EEPROM_HDR_START 0
58 #define EEPROM_RECORD_START (EEPROM_HDR_START + EEPROM_TABLE_HEADER_SIZE)
59 #define EEPROM_MAX_RECORD_NUM ((EEPROM_SIZE_BYTES - EEPROM_TABLE_HEADER_SIZE) / EEPROM_TABLE_RECORD_SIZE)
60 #define EEPROM_ADDR_MSB_MASK GENMASK(17, 8)
62 #define to_amdgpu_device(x) (container_of(x, struct amdgpu_ras, eeprom_control))->adev
64 static bool __is_ras_eeprom_supported(struct amdgpu_device *adev)
66 if ((adev->asic_type == CHIP_VEGA20) ||
67 (adev->asic_type == CHIP_ARCTURUS) ||
68 (adev->asic_type == CHIP_SIENNA_CICHLID) ||
69 (adev->asic_type == CHIP_ALDEBARAN))
75 static bool __get_eeprom_i2c_addr_arct(struct amdgpu_device *adev,
78 struct atom_context *atom_ctx = adev->mode_info.atom_context;
80 if (!i2c_addr || !atom_ctx)
83 if (strnstr(atom_ctx->vbios_version,
85 sizeof(atom_ctx->vbios_version)))
86 *i2c_addr = EEPROM_I2C_TARGET_ADDR_ARCTURUS_D342;
88 *i2c_addr = EEPROM_I2C_TARGET_ADDR_ARCTURUS;
93 static bool __get_eeprom_i2c_addr(struct amdgpu_device *adev,
99 switch (adev->asic_type) {
101 *i2c_addr = EEPROM_I2C_TARGET_ADDR_VEGA20;
105 return __get_eeprom_i2c_addr_arct(adev, i2c_addr);
107 case CHIP_SIENNA_CICHLID:
108 *i2c_addr = EEPROM_I2C_TARGET_ADDR_SIENNA_CICHLID;
112 *i2c_addr = EEPROM_I2C_TARGET_ADDR_ALDEBARAN;
122 static void __encode_table_header_to_buff(struct amdgpu_ras_eeprom_table_header *hdr,
125 uint32_t *pp = (uint32_t *) buff;
127 pp[0] = cpu_to_le32(hdr->header);
128 pp[1] = cpu_to_le32(hdr->version);
129 pp[2] = cpu_to_le32(hdr->first_rec_offset);
130 pp[3] = cpu_to_le32(hdr->tbl_size);
131 pp[4] = cpu_to_le32(hdr->checksum);
134 static void __decode_table_header_from_buff(struct amdgpu_ras_eeprom_table_header *hdr,
137 uint32_t *pp = (uint32_t *)buff;
139 hdr->header = le32_to_cpu(pp[0]);
140 hdr->version = le32_to_cpu(pp[1]);
141 hdr->first_rec_offset = le32_to_cpu(pp[2]);
142 hdr->tbl_size = le32_to_cpu(pp[3]);
143 hdr->checksum = le32_to_cpu(pp[4]);
146 static int __update_table_header(struct amdgpu_ras_eeprom_control *control,
150 struct amdgpu_device *adev = to_amdgpu_device(control);
151 struct i2c_msg msg = {
154 .len = EEPROM_ADDRESS_SIZE + EEPROM_TABLE_HEADER_SIZE,
159 *(uint16_t *)buff = EEPROM_HDR_START;
160 __encode_table_header_to_buff(&control->tbl_hdr, buff + EEPROM_ADDRESS_SIZE);
162 msg.addr = control->i2c_address;
164 /* i2c may be unstable in gpu reset */
165 down_read(&adev->reset_sem);
166 ret = i2c_transfer(&adev->pm.smu_i2c, &msg, 1);
167 up_read(&adev->reset_sem);
170 DRM_ERROR("Failed to write EEPROM table header, ret:%d", ret);
175 static uint32_t __calc_hdr_byte_sum(struct amdgpu_ras_eeprom_control *control)
178 uint32_t tbl_sum = 0;
180 /* Header checksum, skip checksum field in the calculation */
181 for (i = 0; i < sizeof(control->tbl_hdr) - sizeof(control->tbl_hdr.checksum); i++)
182 tbl_sum += *(((unsigned char *)&control->tbl_hdr) + i);
187 static uint32_t __calc_recs_byte_sum(struct eeprom_table_record *records,
191 uint32_t tbl_sum = 0;
193 /* Records checksum */
194 for (i = 0; i < num; i++) {
195 struct eeprom_table_record *record = &records[i];
197 for (j = 0; j < sizeof(*record); j++) {
198 tbl_sum += *(((unsigned char *)record) + j);
205 static inline uint32_t __calc_tbl_byte_sum(struct amdgpu_ras_eeprom_control *control,
206 struct eeprom_table_record *records, int num)
208 return __calc_hdr_byte_sum(control) + __calc_recs_byte_sum(records, num);
211 /* Checksum = 256 -((sum of all table entries) mod 256) */
212 static void __update_tbl_checksum(struct amdgpu_ras_eeprom_control *control,
213 struct eeprom_table_record *records, int num,
214 uint32_t old_hdr_byte_sum)
217 * This will update the table sum with new records.
219 * TODO: What happens when the EEPROM table is to be wrapped around
220 * and old records from start will get overridden.
223 /* need to recalculate updated header byte sum */
224 control->tbl_byte_sum -= old_hdr_byte_sum;
225 control->tbl_byte_sum += __calc_tbl_byte_sum(control, records, num);
227 control->tbl_hdr.checksum = 256 - (control->tbl_byte_sum % 256);
230 /* table sum mod 256 + checksum must equals 256 */
231 static bool __validate_tbl_checksum(struct amdgpu_ras_eeprom_control *control,
232 struct eeprom_table_record *records, int num)
234 control->tbl_byte_sum = __calc_tbl_byte_sum(control, records, num);
236 if (control->tbl_hdr.checksum + (control->tbl_byte_sum % 256) != 256) {
237 DRM_WARN("Checksum mismatch, checksum: %u ", control->tbl_hdr.checksum);
244 static int amdgpu_ras_eeprom_correct_header_tag(
245 struct amdgpu_ras_eeprom_control *control,
248 unsigned char buff[EEPROM_ADDRESS_SIZE + EEPROM_TABLE_HEADER_SIZE];
249 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
252 memset(buff, 0, EEPROM_ADDRESS_SIZE + EEPROM_TABLE_HEADER_SIZE);
254 mutex_lock(&control->tbl_mutex);
255 hdr->header = header;
256 ret = __update_table_header(control, buff);
257 mutex_unlock(&control->tbl_mutex);
262 int amdgpu_ras_eeprom_reset_table(struct amdgpu_ras_eeprom_control *control)
264 unsigned char buff[EEPROM_ADDRESS_SIZE + EEPROM_TABLE_HEADER_SIZE] = { 0 };
265 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
268 mutex_lock(&control->tbl_mutex);
270 hdr->header = EEPROM_TABLE_HDR_VAL;
271 hdr->version = EEPROM_TABLE_VER;
272 hdr->first_rec_offset = EEPROM_RECORD_START;
273 hdr->tbl_size = EEPROM_TABLE_HEADER_SIZE;
275 control->tbl_byte_sum = 0;
276 __update_tbl_checksum(control, NULL, 0, 0);
277 control->next_addr = EEPROM_RECORD_START;
279 ret = __update_table_header(control, buff);
281 mutex_unlock(&control->tbl_mutex);
287 int amdgpu_ras_eeprom_init(struct amdgpu_ras_eeprom_control *control,
288 bool *exceed_err_limit)
291 struct amdgpu_device *adev = to_amdgpu_device(control);
292 unsigned char buff[EEPROM_ADDRESS_SIZE + EEPROM_TABLE_HEADER_SIZE] = { 0 };
293 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
294 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
295 struct i2c_msg msg = {
298 .len = EEPROM_ADDRESS_SIZE + EEPROM_TABLE_HEADER_SIZE,
302 *exceed_err_limit = false;
304 if (!__is_ras_eeprom_supported(adev))
307 /* Verify i2c adapter is initialized */
308 if (!adev->pm.smu_i2c.algo)
311 if (!__get_eeprom_i2c_addr(adev, &control->i2c_address))
314 mutex_init(&control->tbl_mutex);
316 msg.addr = control->i2c_address;
317 /* Read/Create table header from EEPROM address 0 */
318 ret = i2c_transfer(&adev->pm.smu_i2c, &msg, 1);
320 DRM_ERROR("Failed to read EEPROM table header, ret:%d", ret);
324 __decode_table_header_from_buff(hdr, &buff[2]);
326 if (hdr->header == EEPROM_TABLE_HDR_VAL) {
327 control->num_recs = (hdr->tbl_size - EEPROM_TABLE_HEADER_SIZE) /
328 EEPROM_TABLE_RECORD_SIZE;
329 control->tbl_byte_sum = __calc_hdr_byte_sum(control);
330 control->next_addr = EEPROM_RECORD_START;
332 DRM_DEBUG_DRIVER("Found existing EEPROM table with %d records",
335 } else if ((hdr->header == EEPROM_TABLE_HDR_BAD) &&
336 (amdgpu_bad_page_threshold != 0)) {
337 if (ras->bad_page_cnt_threshold > control->num_recs) {
338 dev_info(adev->dev, "Using one valid bigger bad page "
339 "threshold and correcting eeprom header tag.\n");
340 ret = amdgpu_ras_eeprom_correct_header_tag(control,
341 EEPROM_TABLE_HDR_VAL);
343 *exceed_err_limit = true;
344 dev_err(adev->dev, "Exceeding the bad_page_threshold parameter, "
345 "disabling the GPU.\n");
348 DRM_INFO("Creating new EEPROM table");
350 ret = amdgpu_ras_eeprom_reset_table(control);
353 return ret == 1 ? 0 : -EIO;
356 static void __encode_table_record_to_buff(struct amdgpu_ras_eeprom_control *control,
357 struct eeprom_table_record *record,
363 /* Next are all record fields according to EEPROM page spec in LE foramt */
364 buff[i++] = record->err_type;
366 buff[i++] = record->bank;
368 tmp = cpu_to_le64(record->ts);
369 memcpy(buff + i, &tmp, 8);
372 tmp = cpu_to_le64((record->offset & 0xffffffffffff));
373 memcpy(buff + i, &tmp, 6);
376 buff[i++] = record->mem_channel;
377 buff[i++] = record->mcumc_id;
379 tmp = cpu_to_le64((record->retired_page & 0xffffffffffff));
380 memcpy(buff + i, &tmp, 6);
383 static void __decode_table_record_from_buff(struct amdgpu_ras_eeprom_control *control,
384 struct eeprom_table_record *record,
390 /* Next are all record fields according to EEPROM page spec in LE foramt */
391 record->err_type = buff[i++];
393 record->bank = buff[i++];
395 memcpy(&tmp, buff + i, 8);
396 record->ts = le64_to_cpu(tmp);
399 memcpy(&tmp, buff + i, 6);
400 record->offset = (le64_to_cpu(tmp) & 0xffffffffffff);
403 record->mem_channel = buff[i++];
404 record->mcumc_id = buff[i++];
406 memcpy(&tmp, buff + i, 6);
407 record->retired_page = (le64_to_cpu(tmp) & 0xffffffffffff);
411 * When reaching end of EEPROM memory jump back to 0 record address
412 * When next record access will go beyond EEPROM page boundary modify bits A17/A8
413 * in I2C selector to go to next page
415 static uint32_t __correct_eeprom_dest_address(uint32_t curr_address)
417 uint32_t next_address = curr_address + EEPROM_TABLE_RECORD_SIZE;
419 /* When all EEPROM memory used jump back to 0 address */
420 if (next_address > EEPROM_SIZE_BYTES) {
421 DRM_INFO("Reached end of EEPROM memory, jumping to 0 "
422 "and overriding old record");
423 return EEPROM_RECORD_START;
427 * To check if we overflow page boundary compare next address with
428 * current and see if bits 17/8 of the EEPROM address will change
429 * If they do start from the next 256b page
431 * https://www.st.com/resource/en/datasheet/m24m02-dr.pdf sec. 5.1.2
433 if ((curr_address & EEPROM_ADDR_MSB_MASK) != (next_address & EEPROM_ADDR_MSB_MASK)) {
434 DRM_DEBUG_DRIVER("Reached end of EEPROM memory page, jumping to next: %lx",
435 (next_address & EEPROM_ADDR_MSB_MASK));
437 return (next_address & EEPROM_ADDR_MSB_MASK);
443 bool amdgpu_ras_eeprom_check_err_threshold(struct amdgpu_device *adev)
445 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
447 if (!__is_ras_eeprom_supported(adev))
450 /* skip check eeprom table for VEGA20 Gaming */
454 if (!(con->features & BIT(AMDGPU_RAS_BLOCK__UMC)))
457 if (con->eeprom_control.tbl_hdr.header == EEPROM_TABLE_HDR_BAD) {
458 dev_warn(adev->dev, "This GPU is in BAD status.");
459 dev_warn(adev->dev, "Please retire it or setting one bigger "
460 "threshold value when reloading driver.\n");
467 int amdgpu_ras_eeprom_process_recods(struct amdgpu_ras_eeprom_control *control,
468 struct eeprom_table_record *records,
473 struct i2c_msg *msgs, *msg;
474 unsigned char *buffs, *buff;
475 struct eeprom_table_record *record;
476 struct amdgpu_device *adev = to_amdgpu_device(control);
477 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
479 if (!__is_ras_eeprom_supported(adev))
482 buffs = kcalloc(num, EEPROM_ADDRESS_SIZE + EEPROM_TABLE_RECORD_SIZE,
487 mutex_lock(&control->tbl_mutex);
489 msgs = kcalloc(num, sizeof(*msgs), GFP_KERNEL);
496 * If saved bad pages number exceeds the bad page threshold for
497 * the whole VRAM, update table header to mark the BAD GPU tag
498 * and schedule one ras recovery after eeprom write is done,
499 * this can avoid the missing for latest records.
501 * This new header will be picked up and checked in the bootup
502 * by ras recovery, which may break bootup process to notify
503 * user this GPU is in bad state and to retire such GPU for
506 if (write && (amdgpu_bad_page_threshold != 0) &&
507 ((control->num_recs + num) >= ras->bad_page_cnt_threshold)) {
509 "Saved bad pages(%d) reaches threshold value(%d).\n",
510 control->num_recs + num, ras->bad_page_cnt_threshold);
511 control->tbl_hdr.header = EEPROM_TABLE_HDR_BAD;
514 /* In case of overflow just start from beginning to not lose newest records */
515 if (write && (control->next_addr + EEPROM_TABLE_RECORD_SIZE * num > EEPROM_SIZE_BYTES))
516 control->next_addr = EEPROM_RECORD_START;
519 * TODO Currently makes EEPROM writes for each record, this creates
520 * internal fragmentation. Optimized the code to do full page write of
523 for (i = 0; i < num; i++) {
524 buff = &buffs[i * (EEPROM_ADDRESS_SIZE + EEPROM_TABLE_RECORD_SIZE)];
525 record = &records[i];
528 control->next_addr = __correct_eeprom_dest_address(control->next_addr);
531 * Update bits 16,17 of EEPROM address in I2C address by setting them
532 * to bits 1,2 of Device address byte
534 msg->addr = control->i2c_address |
535 ((control->next_addr & EEPROM_ADDR_MSB_MASK) >> 15);
536 msg->flags = write ? 0 : I2C_M_RD;
537 msg->len = EEPROM_ADDRESS_SIZE + EEPROM_TABLE_RECORD_SIZE;
540 /* Insert the EEPROM dest addess, bits 0-15 */
541 buff[0] = ((control->next_addr >> 8) & 0xff);
542 buff[1] = (control->next_addr & 0xff);
544 /* EEPROM table content is stored in LE format */
546 __encode_table_record_to_buff(control, record, buff + EEPROM_ADDRESS_SIZE);
549 * The destination EEPROM address might need to be corrected to account
550 * for page or entire memory wrapping
552 control->next_addr += EEPROM_TABLE_RECORD_SIZE;
555 /* i2c may be unstable in gpu reset */
556 down_read(&adev->reset_sem);
557 ret = i2c_transfer(&adev->pm.smu_i2c, msgs, num);
558 up_read(&adev->reset_sem);
561 DRM_ERROR("Failed to process EEPROM table records, ret:%d", ret);
563 /* TODO Restore prev next EEPROM address ? */
569 for (i = 0; i < num; i++) {
570 buff = &buffs[i*(EEPROM_ADDRESS_SIZE + EEPROM_TABLE_RECORD_SIZE)];
571 record = &records[i];
573 __decode_table_record_from_buff(control, record, buff + EEPROM_ADDRESS_SIZE);
578 uint32_t old_hdr_byte_sum = __calc_hdr_byte_sum(control);
581 * Update table header with size and CRC and account for table
582 * wrap around where the assumption is that we treat it as empty
585 * TODO - Check the assumption is correct
587 control->num_recs += num;
588 control->num_recs %= EEPROM_MAX_RECORD_NUM;
589 control->tbl_hdr.tbl_size += EEPROM_TABLE_RECORD_SIZE * num;
590 if (control->tbl_hdr.tbl_size > EEPROM_SIZE_BYTES)
591 control->tbl_hdr.tbl_size = EEPROM_TABLE_HEADER_SIZE +
592 control->num_recs * EEPROM_TABLE_RECORD_SIZE;
594 __update_tbl_checksum(control, records, num, old_hdr_byte_sum);
596 __update_table_header(control, buffs);
597 } else if (!__validate_tbl_checksum(control, records, num)) {
598 DRM_WARN("EEPROM Table checksum mismatch!");
599 /* TODO Uncomment when EEPROM read/write is relliable */
609 mutex_unlock(&control->tbl_mutex);
611 return ret == num ? 0 : -EIO;
614 inline uint32_t amdgpu_ras_eeprom_get_record_max_length(void)
616 return EEPROM_MAX_RECORD_NUM;
619 /* Used for testing if bugs encountered */
621 void amdgpu_ras_eeprom_test(struct amdgpu_ras_eeprom_control *control)
624 struct eeprom_table_record *recs = kcalloc(1, sizeof(*recs), GFP_KERNEL);
629 for (i = 0; i < 1 ; i++) {
630 recs[i].address = 0xdeadbeef;
631 recs[i].retired_page = i;
634 if (!amdgpu_ras_eeprom_process_recods(control, recs, true, 1)) {
636 memset(recs, 0, sizeof(*recs) * 1);
638 control->next_addr = EEPROM_RECORD_START;
640 if (!amdgpu_ras_eeprom_process_recods(control, recs, false, 1)) {
641 for (i = 0; i < 1; i++)
642 DRM_INFO("rec.address :0x%llx, rec.retired_page :%llu",
643 recs[i].address, recs[i].retired_page);
645 DRM_ERROR("Failed in reading from table");
648 DRM_ERROR("Failed in writing to table");