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>
29 #include "amdgpu_eeprom.h"
30 #include "amdgpu_atomfirmware.h"
31 #include <linux/debugfs.h>
32 #include <linux/uaccess.h>
34 #define EEPROM_I2C_MADDR_VEGA20 0x0
35 #define EEPROM_I2C_MADDR_ARCTURUS 0x40000
36 #define EEPROM_I2C_MADDR_ARCTURUS_D342 0x0
37 #define EEPROM_I2C_MADDR_SIENNA_CICHLID 0x0
38 #define EEPROM_I2C_MADDR_ALDEBARAN 0x0
41 * The 2 macros bellow represent the actual size in bytes that
42 * those entities occupy in the EEPROM memory.
43 * RAS_TABLE_RECORD_SIZE is different than sizeof(eeprom_table_record) which
44 * uses uint64 to store 6b fields such as retired_page.
46 #define RAS_TABLE_HEADER_SIZE 20
47 #define RAS_TABLE_RECORD_SIZE 24
49 /* Table hdr is 'AMDR' */
50 #define RAS_TABLE_HDR_VAL 0x414d4452
51 #define RAS_TABLE_VER 0x00010000
53 /* Bad GPU tag ‘BADG’ */
54 #define RAS_TABLE_HDR_BAD 0x42414447
56 /* Assume 2-Mbit size EEPROM and take up the whole space. */
57 #define RAS_TBL_SIZE_BYTES (256 * 1024)
58 #define RAS_TABLE_START 0
59 #define RAS_HDR_START RAS_TABLE_START
60 #define RAS_RECORD_START (RAS_HDR_START + RAS_TABLE_HEADER_SIZE)
61 #define RAS_MAX_RECORD_COUNT ((RAS_TBL_SIZE_BYTES - RAS_TABLE_HEADER_SIZE) \
62 / RAS_TABLE_RECORD_SIZE)
64 /* Given a zero-based index of an EEPROM RAS record, yields the EEPROM
65 * offset off of RAS_TABLE_START. That is, this is something you can
66 * add to control->i2c_address, and then tell I2C layer to read
67 * from/write to there. _N is the so called absolute index,
68 * because it starts right after the table header.
70 #define RAS_INDEX_TO_OFFSET(_C, _N) ((_C)->ras_record_offset + \
71 (_N) * RAS_TABLE_RECORD_SIZE)
73 #define RAS_OFFSET_TO_INDEX(_C, _O) (((_O) - \
74 (_C)->ras_record_offset) / RAS_TABLE_RECORD_SIZE)
76 /* Given a 0-based relative record index, 0, 1, 2, ..., etc., off
77 * of "fri", return the absolute record index off of the end of
80 #define RAS_RI_TO_AI(_C, _I) (((_I) + (_C)->ras_fri) % \
81 (_C)->ras_max_record_count)
83 #define RAS_NUM_RECS(_tbl_hdr) (((_tbl_hdr)->tbl_size - \
84 RAS_TABLE_HEADER_SIZE) / RAS_TABLE_RECORD_SIZE)
86 #define to_amdgpu_device(x) (container_of(x, struct amdgpu_ras, eeprom_control))->adev
88 static bool __is_ras_eeprom_supported(struct amdgpu_device *adev)
90 return adev->asic_type == CHIP_VEGA20 ||
91 adev->asic_type == CHIP_ARCTURUS ||
92 adev->asic_type == CHIP_SIENNA_CICHLID ||
93 adev->asic_type == CHIP_ALDEBARAN;
96 static bool __get_eeprom_i2c_addr_arct(struct amdgpu_device *adev,
97 struct amdgpu_ras_eeprom_control *control)
99 struct atom_context *atom_ctx = adev->mode_info.atom_context;
101 if (!control || !atom_ctx)
104 if (strnstr(atom_ctx->vbios_version,
106 sizeof(atom_ctx->vbios_version)))
107 control->i2c_address = EEPROM_I2C_MADDR_ARCTURUS_D342;
109 control->i2c_address = EEPROM_I2C_MADDR_ARCTURUS;
114 static bool __get_eeprom_i2c_addr(struct amdgpu_device *adev,
115 struct amdgpu_ras_eeprom_control *control)
120 control->i2c_address = 0;
122 if (amdgpu_atomfirmware_ras_rom_addr(adev, (uint8_t*)&control->i2c_address))
124 if (control->i2c_address == 0xA0)
125 control->i2c_address = 0;
126 else if (control->i2c_address == 0xA8)
127 control->i2c_address = 0x40000;
129 dev_warn(adev->dev, "RAS EEPROM I2C address not supported");
136 switch (adev->asic_type) {
138 control->i2c_address = EEPROM_I2C_MADDR_VEGA20;
142 return __get_eeprom_i2c_addr_arct(adev, control);
144 case CHIP_SIENNA_CICHLID:
145 control->i2c_address = EEPROM_I2C_MADDR_SIENNA_CICHLID;
149 control->i2c_address = EEPROM_I2C_MADDR_ALDEBARAN;
160 __encode_table_header_to_buf(struct amdgpu_ras_eeprom_table_header *hdr,
163 u32 *pp = (uint32_t *)buf;
165 pp[0] = cpu_to_le32(hdr->header);
166 pp[1] = cpu_to_le32(hdr->version);
167 pp[2] = cpu_to_le32(hdr->first_rec_offset);
168 pp[3] = cpu_to_le32(hdr->tbl_size);
169 pp[4] = cpu_to_le32(hdr->checksum);
173 __decode_table_header_from_buf(struct amdgpu_ras_eeprom_table_header *hdr,
176 u32 *pp = (uint32_t *)buf;
178 hdr->header = le32_to_cpu(pp[0]);
179 hdr->version = le32_to_cpu(pp[1]);
180 hdr->first_rec_offset = le32_to_cpu(pp[2]);
181 hdr->tbl_size = le32_to_cpu(pp[3]);
182 hdr->checksum = le32_to_cpu(pp[4]);
185 static int __write_table_header(struct amdgpu_ras_eeprom_control *control)
187 u8 buf[RAS_TABLE_HEADER_SIZE];
188 struct amdgpu_device *adev = to_amdgpu_device(control);
191 memset(buf, 0, sizeof(buf));
192 __encode_table_header_to_buf(&control->tbl_hdr, buf);
194 /* i2c may be unstable in gpu reset */
195 down_read(&adev->reset_sem);
196 res = amdgpu_eeprom_write(&adev->pm.smu_i2c,
197 control->i2c_address +
198 control->ras_header_offset,
199 buf, RAS_TABLE_HEADER_SIZE);
200 up_read(&adev->reset_sem);
203 DRM_ERROR("Failed to write EEPROM table header:%d", res);
204 } else if (res < RAS_TABLE_HEADER_SIZE) {
205 DRM_ERROR("Short write:%d out of %d\n",
206 res, RAS_TABLE_HEADER_SIZE);
215 static u8 __calc_hdr_byte_sum(const struct amdgpu_ras_eeprom_control *control)
221 /* Header checksum, skip checksum field in the calculation */
222 sz = sizeof(control->tbl_hdr) - sizeof(control->tbl_hdr.checksum);
223 pp = (u8 *) &control->tbl_hdr;
225 for (ii = 0; ii < sz; ii++, pp++)
231 static int amdgpu_ras_eeprom_correct_header_tag(
232 struct amdgpu_ras_eeprom_control *control,
235 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
240 csum = -hdr->checksum;
242 hh = (void *) &hdr->header;
243 csum -= (hh[0] + hh[1] + hh[2] + hh[3]);
244 hh = (void *) &header;
245 csum += hh[0] + hh[1] + hh[2] + hh[3];
247 mutex_lock(&control->ras_tbl_mutex);
248 hdr->header = header;
249 hdr->checksum = csum;
250 res = __write_table_header(control);
251 mutex_unlock(&control->ras_tbl_mutex);
257 * amdgpu_ras_eeprom_reset_table -- Reset the RAS EEPROM table
258 * @control: pointer to control structure
260 * Reset the contents of the header of the RAS EEPROM table.
261 * Return 0 on success, -errno on error.
263 int amdgpu_ras_eeprom_reset_table(struct amdgpu_ras_eeprom_control *control)
265 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
269 mutex_lock(&control->ras_tbl_mutex);
271 hdr->header = RAS_TABLE_HDR_VAL;
272 hdr->version = RAS_TABLE_VER;
273 hdr->first_rec_offset = RAS_RECORD_START;
274 hdr->tbl_size = RAS_TABLE_HEADER_SIZE;
276 csum = __calc_hdr_byte_sum(control);
278 hdr->checksum = csum;
279 res = __write_table_header(control);
281 control->ras_num_recs = 0;
282 control->ras_fri = 0;
284 amdgpu_ras_debugfs_set_ret_size(control);
286 mutex_unlock(&control->ras_tbl_mutex);
292 __encode_table_record_to_buf(struct amdgpu_ras_eeprom_control *control,
293 struct eeprom_table_record *record,
299 /* Next are all record fields according to EEPROM page spec in LE foramt */
300 buf[i++] = record->err_type;
302 buf[i++] = record->bank;
304 tmp = cpu_to_le64(record->ts);
305 memcpy(buf + i, &tmp, 8);
308 tmp = cpu_to_le64((record->offset & 0xffffffffffff));
309 memcpy(buf + i, &tmp, 6);
312 buf[i++] = record->mem_channel;
313 buf[i++] = record->mcumc_id;
315 tmp = cpu_to_le64((record->retired_page & 0xffffffffffff));
316 memcpy(buf + i, &tmp, 6);
320 __decode_table_record_from_buf(struct amdgpu_ras_eeprom_control *control,
321 struct eeprom_table_record *record,
327 /* Next are all record fields according to EEPROM page spec in LE foramt */
328 record->err_type = buf[i++];
330 record->bank = buf[i++];
332 memcpy(&tmp, buf + i, 8);
333 record->ts = le64_to_cpu(tmp);
336 memcpy(&tmp, buf + i, 6);
337 record->offset = (le64_to_cpu(tmp) & 0xffffffffffff);
340 record->mem_channel = buf[i++];
341 record->mcumc_id = buf[i++];
343 memcpy(&tmp, buf + i, 6);
344 record->retired_page = (le64_to_cpu(tmp) & 0xffffffffffff);
347 bool amdgpu_ras_eeprom_check_err_threshold(struct amdgpu_device *adev)
349 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
351 if (!__is_ras_eeprom_supported(adev))
354 /* skip check eeprom table for VEGA20 Gaming */
358 if (!(con->features & BIT(AMDGPU_RAS_BLOCK__UMC)))
361 if (con->eeprom_control.tbl_hdr.header == RAS_TABLE_HDR_BAD) {
362 dev_warn(adev->dev, "This GPU is in BAD status.");
363 dev_warn(adev->dev, "Please retire it or set a larger "
364 "threshold value when reloading driver.\n");
372 * __amdgpu_ras_eeprom_write -- write indexed from buffer to EEPROM
373 * @control: pointer to control structure
374 * @buf: pointer to buffer containing data to write
375 * @fri: start writing at this index
376 * @num: number of records to write
378 * The caller must hold the table mutex in @control.
379 * Return 0 on success, -errno otherwise.
381 static int __amdgpu_ras_eeprom_write(struct amdgpu_ras_eeprom_control *control,
382 u8 *buf, const u32 fri, const u32 num)
384 struct amdgpu_device *adev = to_amdgpu_device(control);
388 /* i2c may be unstable in gpu reset */
389 down_read(&adev->reset_sem);
390 buf_size = num * RAS_TABLE_RECORD_SIZE;
391 res = amdgpu_eeprom_write(&adev->pm.smu_i2c,
392 control->i2c_address +
393 RAS_INDEX_TO_OFFSET(control, fri),
395 up_read(&adev->reset_sem);
397 DRM_ERROR("Writing %d EEPROM table records error:%d",
399 } else if (res < buf_size) {
400 /* Short write, return error.
402 DRM_ERROR("Wrote %d records out of %d",
403 res / RAS_TABLE_RECORD_SIZE, num);
413 amdgpu_ras_eeprom_append_table(struct amdgpu_ras_eeprom_control *control,
414 struct eeprom_table_record *record,
421 buf = kcalloc(num, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
425 /* Encode all of them in one go.
428 for (i = 0; i < num; i++, pp += RAS_TABLE_RECORD_SIZE)
429 __encode_table_record_to_buf(control, &record[i], pp);
431 /* a, first record index to write into.
432 * b, last record index to write into.
433 * a = first index to read (fri) + number of records in the table,
435 * Let N = control->ras_max_num_record_count, then we have,
436 * case 0: 0 <= a <= b < N,
437 * just append @num records starting at a;
438 * case 1: 0 <= a < N <= b,
439 * append (N - a) records starting at a, and
440 * append the remainder, b % N + 1, starting at 0.
441 * case 2: 0 <= fri < N <= a <= b, then modulo N we get two subcases,
442 * case 2a: 0 <= a <= b < N
443 * append num records starting at a; and fix fri if b overwrote it,
444 * and since a <= b, if b overwrote it then a must've also,
445 * and if b didn't overwrite it, then a didn't also.
446 * case 2b: 0 <= b < a < N
447 * write num records starting at a, which wraps around 0=N
448 * and overwrite fri unconditionally. Now from case 2a,
449 * this means that b eclipsed fri to overwrite it and wrap
450 * around 0 again, i.e. b = 2N+r pre modulo N, so we unconditionally
451 * set fri = b + 1 (mod N).
452 * Now, since fri is updated in every case, except the trivial case 0,
453 * the number of records present in the table after writing, is,
454 * num_recs - 1 = b - fri (mod N), and we take the positive value,
455 * by adding an arbitrary multiple of N before taking the modulo N
458 a = control->ras_fri + control->ras_num_recs;
460 if (b < control->ras_max_record_count) {
461 res = __amdgpu_ras_eeprom_write(control, buf, a, num);
462 } else if (a < control->ras_max_record_count) {
465 g0 = control->ras_max_record_count - a;
466 g1 = b % control->ras_max_record_count + 1;
467 res = __amdgpu_ras_eeprom_write(control, buf, a, g0);
470 res = __amdgpu_ras_eeprom_write(control,
471 buf + g0 * RAS_TABLE_RECORD_SIZE,
475 if (g1 > control->ras_fri)
476 control->ras_fri = g1 % control->ras_max_record_count;
478 a %= control->ras_max_record_count;
479 b %= control->ras_max_record_count;
482 /* Note that, b - a + 1 = num. */
483 res = __amdgpu_ras_eeprom_write(control, buf, a, num);
486 if (b >= control->ras_fri)
487 control->ras_fri = (b + 1) % control->ras_max_record_count;
491 /* b < a, which means, we write from
492 * a to the end of the table, and from
493 * the start of the table to b.
495 g0 = control->ras_max_record_count - a;
497 res = __amdgpu_ras_eeprom_write(control, buf, a, g0);
500 res = __amdgpu_ras_eeprom_write(control,
501 buf + g0 * RAS_TABLE_RECORD_SIZE,
505 control->ras_fri = g1 % control->ras_max_record_count;
508 control->ras_num_recs = 1 + (control->ras_max_record_count + b
510 % control->ras_max_record_count;
517 amdgpu_ras_eeprom_update_header(struct amdgpu_ras_eeprom_control *control)
519 struct amdgpu_device *adev = to_amdgpu_device(control);
520 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
525 /* Modify the header if it exceeds.
527 if (amdgpu_bad_page_threshold != 0 &&
528 control->ras_num_recs >= ras->bad_page_cnt_threshold) {
530 "Saved bad pages %d reaches threshold value %d\n",
531 control->ras_num_recs, ras->bad_page_cnt_threshold);
532 control->tbl_hdr.header = RAS_TABLE_HDR_BAD;
535 control->tbl_hdr.version = RAS_TABLE_VER;
536 control->tbl_hdr.first_rec_offset = RAS_INDEX_TO_OFFSET(control, control->ras_fri);
537 control->tbl_hdr.tbl_size = RAS_TABLE_HEADER_SIZE + control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
538 control->tbl_hdr.checksum = 0;
540 buf_size = control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
541 buf = kcalloc(control->ras_num_recs, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
543 DRM_ERROR("allocating memory for table of size %d bytes failed\n",
544 control->tbl_hdr.tbl_size);
549 down_read(&adev->reset_sem);
550 res = amdgpu_eeprom_read(&adev->pm.smu_i2c,
551 control->i2c_address +
552 control->ras_record_offset,
554 up_read(&adev->reset_sem);
556 DRM_ERROR("EEPROM failed reading records:%d\n",
559 } else if (res < buf_size) {
560 DRM_ERROR("EEPROM read %d out of %d bytes\n",
566 /* Recalc the checksum.
569 for (pp = buf; pp < buf + buf_size; pp++)
572 csum += __calc_hdr_byte_sum(control);
573 /* avoid sign extension when assigning to "checksum" */
575 control->tbl_hdr.checksum = csum;
576 res = __write_table_header(control);
583 * amdgpu_ras_eeprom_append -- append records to the EEPROM RAS table
584 * @control: pointer to control structure
585 * @record: array of records to append
586 * @num: number of records in @record array
588 * Append @num records to the table, calculate the checksum and write
589 * the table back to EEPROM. The maximum number of records that
590 * can be appended is between 1 and control->ras_max_record_count,
591 * regardless of how many records are already stored in the table.
593 * Return 0 on success or if EEPROM is not supported, -errno on error.
595 int amdgpu_ras_eeprom_append(struct amdgpu_ras_eeprom_control *control,
596 struct eeprom_table_record *record,
599 struct amdgpu_device *adev = to_amdgpu_device(control);
602 if (!__is_ras_eeprom_supported(adev))
606 DRM_ERROR("will not append 0 records\n");
608 } else if (num > control->ras_max_record_count) {
609 DRM_ERROR("cannot append %d records than the size of table %d\n",
610 num, control->ras_max_record_count);
614 mutex_lock(&control->ras_tbl_mutex);
616 res = amdgpu_ras_eeprom_append_table(control, record, num);
618 res = amdgpu_ras_eeprom_update_header(control);
620 amdgpu_ras_debugfs_set_ret_size(control);
622 mutex_unlock(&control->ras_tbl_mutex);
627 * __amdgpu_ras_eeprom_read -- read indexed from EEPROM into buffer
628 * @control: pointer to control structure
629 * @buf: pointer to buffer to read into
630 * @fri: first record index, start reading at this index, absolute index
631 * @num: number of records to read
633 * The caller must hold the table mutex in @control.
634 * Return 0 on success, -errno otherwise.
636 static int __amdgpu_ras_eeprom_read(struct amdgpu_ras_eeprom_control *control,
637 u8 *buf, const u32 fri, const u32 num)
639 struct amdgpu_device *adev = to_amdgpu_device(control);
643 /* i2c may be unstable in gpu reset */
644 down_read(&adev->reset_sem);
645 buf_size = num * RAS_TABLE_RECORD_SIZE;
646 res = amdgpu_eeprom_read(&adev->pm.smu_i2c,
647 control->i2c_address +
648 RAS_INDEX_TO_OFFSET(control, fri),
650 up_read(&adev->reset_sem);
652 DRM_ERROR("Reading %d EEPROM table records error:%d",
654 } else if (res < buf_size) {
655 /* Short read, return error.
657 DRM_ERROR("Read %d records out of %d",
658 res / RAS_TABLE_RECORD_SIZE, num);
668 * amdgpu_ras_eeprom_read -- read EEPROM
669 * @control: pointer to control structure
670 * @record: array of records to read into
671 * @num: number of records in @record
673 * Reads num records from the RAS table in EEPROM and
674 * writes the data into @record array.
676 * Returns 0 on success, -errno on error.
678 int amdgpu_ras_eeprom_read(struct amdgpu_ras_eeprom_control *control,
679 struct eeprom_table_record *record,
682 struct amdgpu_device *adev = to_amdgpu_device(control);
687 if (!__is_ras_eeprom_supported(adev))
691 DRM_ERROR("will not read 0 records\n");
693 } else if (num > control->ras_num_recs) {
694 DRM_ERROR("too many records to read:%d available:%d\n",
695 num, control->ras_num_recs);
699 buf = kcalloc(num, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
703 /* Determine how many records to read, from the first record
704 * index, fri, to the end of the table, and from the beginning
705 * of the table, such that the total number of records is
706 * @num, and we handle wrap around when fri > 0 and
707 * fri + num > RAS_MAX_RECORD_COUNT.
709 * First we compute the index of the last element
710 * which would be fetched from each region,
711 * g0 is in [fri, fri + num - 1], and
712 * g1 is in [0, RAS_MAX_RECORD_COUNT - 1].
713 * Then, if g0 < RAS_MAX_RECORD_COUNT, the index of
714 * the last element to fetch, we set g0 to _the number_
715 * of elements to fetch, @num, since we know that the last
716 * indexed to be fetched does not exceed the table.
718 * If, however, g0 >= RAS_MAX_RECORD_COUNT, then
719 * we set g0 to the number of elements to read
720 * until the end of the table, and g1 to the number of
721 * elements to read from the beginning of the table.
723 g0 = control->ras_fri + num - 1;
724 g1 = g0 % control->ras_max_record_count;
725 if (g0 < control->ras_max_record_count) {
729 g0 = control->ras_max_record_count - control->ras_fri;
733 mutex_lock(&control->ras_tbl_mutex);
734 res = __amdgpu_ras_eeprom_read(control, buf, control->ras_fri, g0);
738 res = __amdgpu_ras_eeprom_read(control,
739 buf + g0 * RAS_TABLE_RECORD_SIZE,
747 /* Read up everything? Then transform.
750 for (i = 0; i < num; i++, pp += RAS_TABLE_RECORD_SIZE)
751 __decode_table_record_from_buf(control, &record[i], pp);
754 mutex_unlock(&control->ras_tbl_mutex);
759 inline uint32_t amdgpu_ras_eeprom_max_record_count(void)
761 return RAS_MAX_RECORD_COUNT;
765 amdgpu_ras_debugfs_eeprom_size_read(struct file *f, char __user *buf,
766 size_t size, loff_t *pos)
768 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
769 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
770 struct amdgpu_ras_eeprom_control *control = ras ? &ras->eeprom_control : NULL;
777 if (!ras || !control) {
778 res = snprintf(data, sizeof(data), "Not supported\n");
780 res = snprintf(data, sizeof(data), "%d bytes or %d records\n",
781 RAS_TBL_SIZE_BYTES, control->ras_max_record_count);
788 res = min_t(size_t, res, size);
790 if (copy_to_user(buf, &data[*pos], res))
798 const struct file_operations amdgpu_ras_debugfs_eeprom_size_ops = {
799 .owner = THIS_MODULE,
800 .read = amdgpu_ras_debugfs_eeprom_size_read,
802 .llseek = default_llseek,
805 static const char *tbl_hdr_str = " Signature Version FirstOffs Size Checksum\n";
806 static const char *tbl_hdr_fmt = "0x%08X 0x%08X 0x%08X 0x%08X 0x%08X\n";
807 #define tbl_hdr_fmt_size (5 * (2+8) + 4 + 1)
808 static const char *rec_hdr_str = "Index Offset ErrType Bank/CU TimeStamp Offs/Addr MemChl MCUMCID RetiredPage\n";
809 static const char *rec_hdr_fmt = "%5d 0x%05X %7s 0x%02X 0x%016llX 0x%012llX 0x%02X 0x%02X 0x%012llX\n";
810 #define rec_hdr_fmt_size (5 + 1 + 7 + 1 + 7 + 1 + 7 + 1 + 18 + 1 + 14 + 1 + 6 + 1 + 7 + 1 + 14 + 1)
812 static const char *record_err_type_str[AMDGPU_RAS_EEPROM_ERR_COUNT] = {
818 static loff_t amdgpu_ras_debugfs_table_size(struct amdgpu_ras_eeprom_control *control)
820 return strlen(tbl_hdr_str) + tbl_hdr_fmt_size +
821 strlen(rec_hdr_str) + rec_hdr_fmt_size * control->ras_num_recs;
824 void amdgpu_ras_debugfs_set_ret_size(struct amdgpu_ras_eeprom_control *control)
826 struct amdgpu_ras *ras = container_of(control, struct amdgpu_ras,
828 struct dentry *de = ras->de_ras_eeprom_table;
831 d_inode(de)->i_size = amdgpu_ras_debugfs_table_size(control);
834 static ssize_t amdgpu_ras_debugfs_table_read(struct file *f, char __user *buf,
835 size_t size, loff_t *pos)
837 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
838 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
839 struct amdgpu_ras_eeprom_control *control = &ras->eeprom_control;
840 const size_t orig_size = size;
844 mutex_lock(&control->ras_tbl_mutex);
846 /* We want *pos - data_len > 0, which means there's
847 * bytes to be printed from data.
849 data_len = strlen(tbl_hdr_str);
850 if (*pos < data_len) {
852 data_len = min_t(size_t, data_len, size);
853 if (copy_to_user(buf, &tbl_hdr_str[*pos], data_len))
860 data_len = strlen(tbl_hdr_str) + tbl_hdr_fmt_size;
861 if (*pos < data_len && size > 0) {
862 u8 data[tbl_hdr_fmt_size + 1];
865 snprintf(data, sizeof(data), tbl_hdr_fmt,
866 control->tbl_hdr.header,
867 control->tbl_hdr.version,
868 control->tbl_hdr.first_rec_offset,
869 control->tbl_hdr.tbl_size,
870 control->tbl_hdr.checksum);
873 data_len = min_t(size_t, data_len, size);
874 lpos = *pos - strlen(tbl_hdr_str);
875 if (copy_to_user(buf, &data[lpos], data_len))
882 data_len = strlen(tbl_hdr_str) + tbl_hdr_fmt_size + strlen(rec_hdr_str);
883 if (*pos < data_len && size > 0) {
887 data_len = min_t(size_t, data_len, size);
888 lpos = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size;
889 if (copy_to_user(buf, &rec_hdr_str[lpos], data_len))
896 data_len = amdgpu_ras_debugfs_table_size(control);
897 if (*pos < data_len && size > 0) {
898 u8 dare[RAS_TABLE_RECORD_SIZE];
899 u8 data[rec_hdr_fmt_size + 1];
900 struct eeprom_table_record record;
903 /* Find the starting record index
905 s = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size -
907 s = s / rec_hdr_fmt_size;
908 r = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size -
910 r = r % rec_hdr_fmt_size;
912 for ( ; size > 0 && s < control->ras_num_recs; s++) {
913 u32 ai = RAS_RI_TO_AI(control, s);
914 /* Read a single record
916 res = __amdgpu_ras_eeprom_read(control, dare, ai, 1);
919 __decode_table_record_from_buf(control, &record, dare);
920 snprintf(data, sizeof(data), rec_hdr_fmt,
922 RAS_INDEX_TO_OFFSET(control, ai),
923 record_err_type_str[record.err_type],
929 record.retired_page);
931 data_len = min_t(size_t, rec_hdr_fmt_size - r, size);
932 if (copy_to_user(buf, &data[r], data_len)) {
944 mutex_unlock(&control->ras_tbl_mutex);
945 return res < 0 ? res : orig_size - size;
949 amdgpu_ras_debugfs_eeprom_table_read(struct file *f, char __user *buf,
950 size_t size, loff_t *pos)
952 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
953 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
954 struct amdgpu_ras_eeprom_control *control = ras ? &ras->eeprom_control : NULL;
961 if (!ras || !control) {
962 res = snprintf(data, sizeof(data), "Not supported\n");
967 res = min_t(size_t, res, size);
969 if (copy_to_user(buf, &data[*pos], res))
976 return amdgpu_ras_debugfs_table_read(f, buf, size, pos);
980 const struct file_operations amdgpu_ras_debugfs_eeprom_table_ops = {
981 .owner = THIS_MODULE,
982 .read = amdgpu_ras_debugfs_eeprom_table_read,
984 .llseek = default_llseek,
988 * __verify_ras_table_checksum -- verify the RAS EEPROM table checksum
989 * @control: pointer to control structure
991 * Check the checksum of the stored in EEPROM RAS table.
993 * Return 0 if the checksum is correct,
994 * positive if it is not correct, and
995 * -errno on I/O error.
997 static int __verify_ras_table_checksum(struct amdgpu_ras_eeprom_control *control)
999 struct amdgpu_device *adev = to_amdgpu_device(control);
1003 buf_size = RAS_TABLE_HEADER_SIZE +
1004 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
1005 buf = kzalloc(buf_size, GFP_KERNEL);
1007 DRM_ERROR("Out of memory checking RAS table checksum.\n");
1011 res = amdgpu_eeprom_read(&adev->pm.smu_i2c,
1012 control->i2c_address +
1013 control->ras_header_offset,
1015 if (res < buf_size) {
1016 DRM_ERROR("Partial read for checksum, res:%d\n", res);
1017 /* On partial reads, return -EIO.
1025 for (pp = buf; pp < buf + buf_size; pp++)
1029 return res < 0 ? res : csum;
1032 int amdgpu_ras_eeprom_init(struct amdgpu_ras_eeprom_control *control,
1033 bool *exceed_err_limit)
1035 struct amdgpu_device *adev = to_amdgpu_device(control);
1036 unsigned char buf[RAS_TABLE_HEADER_SIZE] = { 0 };
1037 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
1038 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1041 *exceed_err_limit = false;
1043 if (!__is_ras_eeprom_supported(adev))
1046 /* Verify i2c adapter is initialized */
1047 if (!adev->pm.smu_i2c.algo)
1050 if (!__get_eeprom_i2c_addr(adev, control))
1053 control->ras_header_offset = RAS_HDR_START;
1054 control->ras_record_offset = RAS_RECORD_START;
1055 control->ras_max_record_count = RAS_MAX_RECORD_COUNT;
1056 mutex_init(&control->ras_tbl_mutex);
1058 /* Read the table header from EEPROM address */
1059 res = amdgpu_eeprom_read(&adev->pm.smu_i2c,
1060 control->i2c_address + control->ras_header_offset,
1061 buf, RAS_TABLE_HEADER_SIZE);
1062 if (res < RAS_TABLE_HEADER_SIZE) {
1063 DRM_ERROR("Failed to read EEPROM table header, res:%d", res);
1064 return res >= 0 ? -EIO : res;
1067 __decode_table_header_from_buf(hdr, buf);
1069 control->ras_num_recs = RAS_NUM_RECS(hdr);
1070 control->ras_fri = RAS_OFFSET_TO_INDEX(control, hdr->first_rec_offset);
1072 if (hdr->header == RAS_TABLE_HDR_VAL) {
1073 DRM_DEBUG_DRIVER("Found existing EEPROM table with %d records",
1074 control->ras_num_recs);
1075 res = __verify_ras_table_checksum(control);
1077 DRM_ERROR("RAS table incorrect checksum or error:%d\n",
1079 } else if (hdr->header == RAS_TABLE_HDR_BAD &&
1080 amdgpu_bad_page_threshold != 0) {
1081 res = __verify_ras_table_checksum(control);
1083 DRM_ERROR("RAS Table incorrect checksum or error:%d\n",
1085 if (ras->bad_page_cnt_threshold > control->ras_num_recs) {
1086 /* This means that, the threshold was increased since
1087 * the last time the system was booted, and now,
1088 * ras->bad_page_cnt_threshold - control->num_recs > 0,
1089 * so that at least one more record can be saved,
1090 * before the page count threshold is reached.
1093 "records:%d threshold:%d, resetting "
1094 "RAS table header signature",
1095 control->ras_num_recs,
1096 ras->bad_page_cnt_threshold);
1097 res = amdgpu_ras_eeprom_correct_header_tag(control,
1100 *exceed_err_limit = true;
1102 "RAS records:%d exceed threshold:%d, "
1103 "maybe retire this GPU?",
1104 control->ras_num_recs, ras->bad_page_cnt_threshold);
1107 DRM_INFO("Creating a new EEPROM table");
1109 res = amdgpu_ras_eeprom_reset_table(control);
1112 return res < 0 ? res : 0;