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 <linux/debugfs.h>
31 #include <linux/uaccess.h>
33 #define EEPROM_I2C_MADDR_VEGA20 0x0
34 #define EEPROM_I2C_MADDR_ARCTURUS 0x40000
35 #define EEPROM_I2C_MADDR_ARCTURUS_D342 0x0
36 #define EEPROM_I2C_MADDR_SIENNA_CICHLID 0x0
37 #define EEPROM_I2C_MADDR_ALDEBARAN 0x0
40 * The 2 macros bellow represent the actual size in bytes that
41 * those entities occupy in the EEPROM memory.
42 * RAS_TABLE_RECORD_SIZE is different than sizeof(eeprom_table_record) which
43 * uses uint64 to store 6b fields such as retired_page.
45 #define RAS_TABLE_HEADER_SIZE 20
46 #define RAS_TABLE_RECORD_SIZE 24
48 /* Table hdr is 'AMDR' */
49 #define RAS_TABLE_HDR_VAL 0x414d4452
50 #define RAS_TABLE_VER 0x00010000
52 /* Bad GPU tag ‘BADG’ */
53 #define RAS_TABLE_HDR_BAD 0x42414447
55 /* Assume 2-Mbit size EEPROM and take up the whole space. */
56 #define RAS_TBL_SIZE_BYTES (256 * 1024)
57 #define RAS_TABLE_START 0
58 #define RAS_HDR_START RAS_TABLE_START
59 #define RAS_RECORD_START (RAS_HDR_START + RAS_TABLE_HEADER_SIZE)
60 #define RAS_MAX_RECORD_COUNT ((RAS_TBL_SIZE_BYTES - RAS_TABLE_HEADER_SIZE) \
61 / RAS_TABLE_RECORD_SIZE)
63 /* Given a zero-based index of an EEPROM RAS record, yields the EEPROM
64 * offset off of RAS_TABLE_START. That is, this is something you can
65 * add to control->i2c_address, and then tell I2C layer to read
66 * from/write to there. _N is the so called absolute index,
67 * because it starts right after the table header.
69 #define RAS_INDEX_TO_OFFSET(_C, _N) ((_C)->ras_record_offset + \
70 (_N) * RAS_TABLE_RECORD_SIZE)
72 #define RAS_OFFSET_TO_INDEX(_C, _O) (((_O) - \
73 (_C)->ras_record_offset) / RAS_TABLE_RECORD_SIZE)
75 /* Given a 0-based relative record index, 0, 1, 2, ..., etc., off
76 * of "fri", return the absolute record index off of the end of
79 #define RAS_RI_TO_AI(_C, _I) (((_I) + (_C)->ras_fri) % \
80 (_C)->ras_max_record_count)
82 #define RAS_NUM_RECS(_tbl_hdr) (((_tbl_hdr)->tbl_size - \
83 RAS_TABLE_HEADER_SIZE) / RAS_TABLE_RECORD_SIZE)
85 #define to_amdgpu_device(x) (container_of(x, struct amdgpu_ras, eeprom_control))->adev
87 static bool __is_ras_eeprom_supported(struct amdgpu_device *adev)
89 return adev->asic_type == CHIP_VEGA20 ||
90 adev->asic_type == CHIP_ARCTURUS ||
91 adev->asic_type == CHIP_SIENNA_CICHLID ||
92 adev->asic_type == CHIP_ALDEBARAN;
95 static bool __get_eeprom_i2c_addr_arct(struct amdgpu_device *adev,
96 struct amdgpu_ras_eeprom_control *control)
98 struct atom_context *atom_ctx = adev->mode_info.atom_context;
100 if (!control || !atom_ctx)
103 if (strnstr(atom_ctx->vbios_version,
105 sizeof(atom_ctx->vbios_version)))
106 control->i2c_address = EEPROM_I2C_MADDR_ARCTURUS_D342;
108 control->i2c_address = EEPROM_I2C_MADDR_ARCTURUS;
113 static bool __get_eeprom_i2c_addr(struct amdgpu_device *adev,
114 struct amdgpu_ras_eeprom_control *control)
119 switch (adev->asic_type) {
121 control->i2c_address = EEPROM_I2C_MADDR_VEGA20;
125 return __get_eeprom_i2c_addr_arct(adev, control);
127 case CHIP_SIENNA_CICHLID:
128 control->i2c_address = EEPROM_I2C_MADDR_SIENNA_CICHLID;
132 control->i2c_address = EEPROM_I2C_MADDR_ALDEBARAN;
143 __encode_table_header_to_buf(struct amdgpu_ras_eeprom_table_header *hdr,
146 u32 *pp = (uint32_t *)buf;
148 pp[0] = cpu_to_le32(hdr->header);
149 pp[1] = cpu_to_le32(hdr->version);
150 pp[2] = cpu_to_le32(hdr->first_rec_offset);
151 pp[3] = cpu_to_le32(hdr->tbl_size);
152 pp[4] = cpu_to_le32(hdr->checksum);
156 __decode_table_header_from_buf(struct amdgpu_ras_eeprom_table_header *hdr,
159 u32 *pp = (uint32_t *)buf;
161 hdr->header = le32_to_cpu(pp[0]);
162 hdr->version = le32_to_cpu(pp[1]);
163 hdr->first_rec_offset = le32_to_cpu(pp[2]);
164 hdr->tbl_size = le32_to_cpu(pp[3]);
165 hdr->checksum = le32_to_cpu(pp[4]);
168 static int __write_table_header(struct amdgpu_ras_eeprom_control *control)
170 u8 buf[RAS_TABLE_HEADER_SIZE];
171 struct amdgpu_device *adev = to_amdgpu_device(control);
174 memset(buf, 0, sizeof(buf));
175 __encode_table_header_to_buf(&control->tbl_hdr, buf);
177 /* i2c may be unstable in gpu reset */
178 down_read(&adev->reset_sem);
179 res = amdgpu_eeprom_write(&adev->pm.smu_i2c,
180 control->i2c_address +
181 control->ras_header_offset,
182 buf, RAS_TABLE_HEADER_SIZE);
183 up_read(&adev->reset_sem);
186 DRM_ERROR("Failed to write EEPROM table header:%d", res);
187 } else if (res < RAS_TABLE_HEADER_SIZE) {
188 DRM_ERROR("Short write:%d out of %d\n",
189 res, RAS_TABLE_HEADER_SIZE);
198 static u8 __calc_hdr_byte_sum(const struct amdgpu_ras_eeprom_control *control)
204 /* Header checksum, skip checksum field in the calculation */
205 sz = sizeof(control->tbl_hdr) - sizeof(control->tbl_hdr.checksum);
206 pp = (u8 *) &control->tbl_hdr;
208 for (ii = 0; ii < sz; ii++, pp++)
214 static int amdgpu_ras_eeprom_correct_header_tag(
215 struct amdgpu_ras_eeprom_control *control,
218 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
223 csum = -hdr->checksum;
225 hh = (void *) &hdr->header;
226 csum -= (hh[0] + hh[1] + hh[2] + hh[3]);
227 hh = (void *) &header;
228 csum += hh[0] + hh[1] + hh[2] + hh[3];
230 mutex_lock(&control->ras_tbl_mutex);
231 hdr->header = header;
232 hdr->checksum = csum;
233 res = __write_table_header(control);
234 mutex_unlock(&control->ras_tbl_mutex);
240 * amdgpu_ras_eeprom_reset_table -- Reset the RAS EEPROM table
241 * @control: pointer to control structure
243 * Reset the contents of the header of the RAS EEPROM table.
244 * Return 0 on success, -errno on error.
246 int amdgpu_ras_eeprom_reset_table(struct amdgpu_ras_eeprom_control *control)
248 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
252 mutex_lock(&control->ras_tbl_mutex);
254 hdr->header = RAS_TABLE_HDR_VAL;
255 hdr->version = RAS_TABLE_VER;
256 hdr->first_rec_offset = RAS_RECORD_START;
257 hdr->tbl_size = RAS_TABLE_HEADER_SIZE;
259 csum = __calc_hdr_byte_sum(control);
261 hdr->checksum = csum;
262 res = __write_table_header(control);
264 control->ras_num_recs = 0;
265 control->ras_fri = 0;
267 amdgpu_ras_debugfs_set_ret_size(control);
269 mutex_unlock(&control->ras_tbl_mutex);
275 __encode_table_record_to_buf(struct amdgpu_ras_eeprom_control *control,
276 struct eeprom_table_record *record,
282 /* Next are all record fields according to EEPROM page spec in LE foramt */
283 buf[i++] = record->err_type;
285 buf[i++] = record->bank;
287 tmp = cpu_to_le64(record->ts);
288 memcpy(buf + i, &tmp, 8);
291 tmp = cpu_to_le64((record->offset & 0xffffffffffff));
292 memcpy(buf + i, &tmp, 6);
295 buf[i++] = record->mem_channel;
296 buf[i++] = record->mcumc_id;
298 tmp = cpu_to_le64((record->retired_page & 0xffffffffffff));
299 memcpy(buf + i, &tmp, 6);
303 __decode_table_record_from_buf(struct amdgpu_ras_eeprom_control *control,
304 struct eeprom_table_record *record,
310 /* Next are all record fields according to EEPROM page spec in LE foramt */
311 record->err_type = buf[i++];
313 record->bank = buf[i++];
315 memcpy(&tmp, buf + i, 8);
316 record->ts = le64_to_cpu(tmp);
319 memcpy(&tmp, buf + i, 6);
320 record->offset = (le64_to_cpu(tmp) & 0xffffffffffff);
323 record->mem_channel = buf[i++];
324 record->mcumc_id = buf[i++];
326 memcpy(&tmp, buf + i, 6);
327 record->retired_page = (le64_to_cpu(tmp) & 0xffffffffffff);
330 bool amdgpu_ras_eeprom_check_err_threshold(struct amdgpu_device *adev)
332 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
334 if (!__is_ras_eeprom_supported(adev))
337 /* skip check eeprom table for VEGA20 Gaming */
341 if (!(con->features & BIT(AMDGPU_RAS_BLOCK__UMC)))
344 if (con->eeprom_control.tbl_hdr.header == RAS_TABLE_HDR_BAD) {
345 dev_warn(adev->dev, "This GPU is in BAD status.");
346 dev_warn(adev->dev, "Please retire it or set a larger "
347 "threshold value when reloading driver.\n");
355 * __amdgpu_ras_eeprom_write -- write indexed from buffer to EEPROM
356 * @control: pointer to control structure
357 * @buf: pointer to buffer containing data to write
358 * @fri: start writing at this index
359 * @num: number of records to write
361 * The caller must hold the table mutex in @control.
362 * Return 0 on success, -errno otherwise.
364 static int __amdgpu_ras_eeprom_write(struct amdgpu_ras_eeprom_control *control,
365 u8 *buf, const u32 fri, const u32 num)
367 struct amdgpu_device *adev = to_amdgpu_device(control);
371 /* i2c may be unstable in gpu reset */
372 down_read(&adev->reset_sem);
373 buf_size = num * RAS_TABLE_RECORD_SIZE;
374 res = amdgpu_eeprom_write(&adev->pm.smu_i2c,
375 control->i2c_address +
376 RAS_INDEX_TO_OFFSET(control, fri),
378 up_read(&adev->reset_sem);
380 DRM_ERROR("Writing %d EEPROM table records error:%d",
382 } else if (res < buf_size) {
383 /* Short write, return error.
385 DRM_ERROR("Wrote %d records out of %d",
386 res / RAS_TABLE_RECORD_SIZE, num);
396 amdgpu_ras_eeprom_append_table(struct amdgpu_ras_eeprom_control *control,
397 struct eeprom_table_record *record,
404 buf = kcalloc(num, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
408 /* Encode all of them in one go.
411 for (i = 0; i < num; i++, pp += RAS_TABLE_RECORD_SIZE)
412 __encode_table_record_to_buf(control, &record[i], pp);
414 /* a, first record index to write into.
415 * b, last record index to write into.
416 * a = first index to read (fri) + number of records in the table,
418 * Let N = control->ras_max_num_record_count, then we have,
419 * case 0: 0 <= a <= b < N,
420 * just append @num records starting at a;
421 * case 1: 0 <= a < N <= b,
422 * append (N - a) records starting at a, and
423 * append the remainder, b % N + 1, starting at 0.
424 * case 2: 0 <= fri < N <= a <= b, then modulo N we get two subcases,
425 * case 2a: 0 <= a <= b < N
426 * append num records starting at a; and fix fri if b overwrote it,
427 * and since a <= b, if b overwrote it then a must've also,
428 * and if b didn't overwrite it, then a didn't also.
429 * case 2b: 0 <= b < a < N
430 * write num records starting at a, which wraps around 0=N
431 * and overwrite fri unconditionally. Now from case 2a,
432 * this means that b eclipsed fri to overwrite it and wrap
433 * around 0 again, i.e. b = 2N+r pre modulo N, so we unconditionally
434 * set fri = b + 1 (mod N).
435 * Now, since fri is updated in every case, except the trivial case 0,
436 * the number of records present in the table after writing, is,
437 * num_recs - 1 = b - fri (mod N), and we take the positive value,
438 * by adding an arbitrary multiple of N before taking the modulo N
441 a = control->ras_fri + control->ras_num_recs;
443 if (b < control->ras_max_record_count) {
444 res = __amdgpu_ras_eeprom_write(control, buf, a, num);
445 } else if (a < control->ras_max_record_count) {
448 g0 = control->ras_max_record_count - a;
449 g1 = b % control->ras_max_record_count + 1;
450 res = __amdgpu_ras_eeprom_write(control, buf, a, g0);
453 res = __amdgpu_ras_eeprom_write(control,
454 buf + g0 * RAS_TABLE_RECORD_SIZE,
458 if (g1 > control->ras_fri)
459 control->ras_fri = g1 % control->ras_max_record_count;
461 a %= control->ras_max_record_count;
462 b %= control->ras_max_record_count;
465 /* Note that, b - a + 1 = num. */
466 res = __amdgpu_ras_eeprom_write(control, buf, a, num);
469 if (b >= control->ras_fri)
470 control->ras_fri = (b + 1) % control->ras_max_record_count;
474 /* b < a, which means, we write from
475 * a to the end of the table, and from
476 * the start of the table to b.
478 g0 = control->ras_max_record_count - a;
480 res = __amdgpu_ras_eeprom_write(control, buf, a, g0);
483 res = __amdgpu_ras_eeprom_write(control,
484 buf + g0 * RAS_TABLE_RECORD_SIZE,
488 control->ras_fri = g1 % control->ras_max_record_count;
491 control->ras_num_recs = 1 + (control->ras_max_record_count + b
493 % control->ras_max_record_count;
500 amdgpu_ras_eeprom_update_header(struct amdgpu_ras_eeprom_control *control)
502 struct amdgpu_device *adev = to_amdgpu_device(control);
503 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
508 /* Modify the header if it exceeds.
510 if (amdgpu_bad_page_threshold != 0 &&
511 control->ras_num_recs >= ras->bad_page_cnt_threshold) {
513 "Saved bad pages %d reaches threshold value %d\n",
514 control->ras_num_recs, ras->bad_page_cnt_threshold);
515 control->tbl_hdr.header = RAS_TABLE_HDR_BAD;
518 control->tbl_hdr.version = RAS_TABLE_VER;
519 control->tbl_hdr.first_rec_offset = RAS_INDEX_TO_OFFSET(control, control->ras_fri);
520 control->tbl_hdr.tbl_size = RAS_TABLE_HEADER_SIZE + control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
521 control->tbl_hdr.checksum = 0;
523 buf_size = control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
524 buf = kcalloc(control->ras_num_recs, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
526 DRM_ERROR("allocating memory for table of size %d bytes failed\n",
527 control->tbl_hdr.tbl_size);
532 down_read(&adev->reset_sem);
533 res = amdgpu_eeprom_read(&adev->pm.smu_i2c,
534 control->i2c_address +
535 control->ras_record_offset,
537 up_read(&adev->reset_sem);
539 DRM_ERROR("EEPROM failed reading records:%d\n",
542 } else if (res < buf_size) {
543 DRM_ERROR("EEPROM read %d out of %d bytes\n",
549 /* Recalc the checksum.
552 for (pp = buf; pp < buf + buf_size; pp++)
555 csum += __calc_hdr_byte_sum(control);
556 /* avoid sign extension when assigning to "checksum" */
558 control->tbl_hdr.checksum = csum;
559 res = __write_table_header(control);
566 * amdgpu_ras_eeprom_append -- append records to the EEPROM RAS table
567 * @control: pointer to control structure
568 * @record: array of records to append
569 * @num: number of records in @record array
571 * Append @num records to the table, calculate the checksum and write
572 * the table back to EEPROM. The maximum number of records that
573 * can be appended is between 1 and control->ras_max_record_count,
574 * regardless of how many records are already stored in the table.
576 * Return 0 on success or if EEPROM is not supported, -errno on error.
578 int amdgpu_ras_eeprom_append(struct amdgpu_ras_eeprom_control *control,
579 struct eeprom_table_record *record,
582 struct amdgpu_device *adev = to_amdgpu_device(control);
585 if (!__is_ras_eeprom_supported(adev))
589 DRM_ERROR("will not append 0 records\n");
591 } else if (num > control->ras_max_record_count) {
592 DRM_ERROR("cannot append %d records than the size of table %d\n",
593 num, control->ras_max_record_count);
597 mutex_lock(&control->ras_tbl_mutex);
599 res = amdgpu_ras_eeprom_append_table(control, record, num);
601 res = amdgpu_ras_eeprom_update_header(control);
603 amdgpu_ras_debugfs_set_ret_size(control);
605 mutex_unlock(&control->ras_tbl_mutex);
610 * __amdgpu_ras_eeprom_read -- read indexed from EEPROM into buffer
611 * @control: pointer to control structure
612 * @buf: pointer to buffer to read into
613 * @fri: first record index, start reading at this index, absolute index
614 * @num: number of records to read
616 * The caller must hold the table mutex in @control.
617 * Return 0 on success, -errno otherwise.
619 static int __amdgpu_ras_eeprom_read(struct amdgpu_ras_eeprom_control *control,
620 u8 *buf, const u32 fri, const u32 num)
622 struct amdgpu_device *adev = to_amdgpu_device(control);
626 /* i2c may be unstable in gpu reset */
627 down_read(&adev->reset_sem);
628 buf_size = num * RAS_TABLE_RECORD_SIZE;
629 res = amdgpu_eeprom_read(&adev->pm.smu_i2c,
630 control->i2c_address +
631 RAS_INDEX_TO_OFFSET(control, fri),
633 up_read(&adev->reset_sem);
635 DRM_ERROR("Reading %d EEPROM table records error:%d",
637 } else if (res < buf_size) {
638 /* Short read, return error.
640 DRM_ERROR("Read %d records out of %d",
641 res / RAS_TABLE_RECORD_SIZE, num);
651 * amdgpu_ras_eeprom_read -- read EEPROM
652 * @control: pointer to control structure
653 * @record: array of records to read into
654 * @num: number of records in @record
656 * Reads num records from the RAS table in EEPROM and
657 * writes the data into @record array.
659 * Returns 0 on success, -errno on error.
661 int amdgpu_ras_eeprom_read(struct amdgpu_ras_eeprom_control *control,
662 struct eeprom_table_record *record,
665 struct amdgpu_device *adev = to_amdgpu_device(control);
670 if (!__is_ras_eeprom_supported(adev))
674 DRM_ERROR("will not read 0 records\n");
676 } else if (num > control->ras_num_recs) {
677 DRM_ERROR("too many records to read:%d available:%d\n",
678 num, control->ras_num_recs);
682 buf = kcalloc(num, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
686 /* Determine how many records to read, from the first record
687 * index, fri, to the end of the table, and from the beginning
688 * of the table, such that the total number of records is
689 * @num, and we handle wrap around when fri > 0 and
690 * fri + num > RAS_MAX_RECORD_COUNT.
692 * First we compute the index of the last element
693 * which would be fetched from each region,
694 * g0 is in [fri, fri + num - 1], and
695 * g1 is in [0, RAS_MAX_RECORD_COUNT - 1].
696 * Then, if g0 < RAS_MAX_RECORD_COUNT, the index of
697 * the last element to fetch, we set g0 to _the number_
698 * of elements to fetch, @num, since we know that the last
699 * indexed to be fetched does not exceed the table.
701 * If, however, g0 >= RAS_MAX_RECORD_COUNT, then
702 * we set g0 to the number of elements to read
703 * until the end of the table, and g1 to the number of
704 * elements to read from the beginning of the table.
706 g0 = control->ras_fri + num - 1;
707 g1 = g0 % control->ras_max_record_count;
708 if (g0 < control->ras_max_record_count) {
712 g0 = control->ras_max_record_count - control->ras_fri;
716 mutex_lock(&control->ras_tbl_mutex);
717 res = __amdgpu_ras_eeprom_read(control, buf, control->ras_fri, g0);
721 res = __amdgpu_ras_eeprom_read(control,
722 buf + g0 * RAS_TABLE_RECORD_SIZE,
730 /* Read up everything? Then transform.
733 for (i = 0; i < num; i++, pp += RAS_TABLE_RECORD_SIZE)
734 __decode_table_record_from_buf(control, &record[i], pp);
737 mutex_unlock(&control->ras_tbl_mutex);
742 inline uint32_t amdgpu_ras_eeprom_max_record_count(void)
744 return RAS_MAX_RECORD_COUNT;
748 amdgpu_ras_debugfs_eeprom_size_read(struct file *f, char __user *buf,
749 size_t size, loff_t *pos)
751 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
752 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
753 struct amdgpu_ras_eeprom_control *control = ras ? &ras->eeprom_control : NULL;
760 if (!ras || !control) {
761 res = snprintf(data, sizeof(data), "Not supported\n");
763 res = snprintf(data, sizeof(data), "%d bytes or %d records\n",
764 RAS_TBL_SIZE_BYTES, control->ras_max_record_count);
771 res = min_t(size_t, res, size);
773 if (copy_to_user(buf, &data[*pos], res))
781 const struct file_operations amdgpu_ras_debugfs_eeprom_size_ops = {
782 .owner = THIS_MODULE,
783 .read = amdgpu_ras_debugfs_eeprom_size_read,
785 .llseek = default_llseek,
788 static const char *tbl_hdr_str = " Signature Version FirstOffs Size Checksum\n";
789 static const char *tbl_hdr_fmt = "0x%08X 0x%08X 0x%08X 0x%08X 0x%08X\n";
790 #define tbl_hdr_fmt_size (5 * (2+8) + 4 + 1)
791 static const char *rec_hdr_str = "Index Offset ErrType Bank/CU TimeStamp Offs/Addr MemChl MCUMCID RetiredPage\n";
792 static const char *rec_hdr_fmt = "%5d 0x%05X %7s 0x%02X 0x%016llX 0x%012llX 0x%02X 0x%02X 0x%012llX\n";
793 #define rec_hdr_fmt_size (5 + 1 + 7 + 1 + 7 + 1 + 7 + 1 + 18 + 1 + 14 + 1 + 6 + 1 + 7 + 1 + 14 + 1)
795 static const char *record_err_type_str[AMDGPU_RAS_EEPROM_ERR_COUNT] = {
801 static loff_t amdgpu_ras_debugfs_table_size(struct amdgpu_ras_eeprom_control *control)
803 return strlen(tbl_hdr_str) + tbl_hdr_fmt_size +
804 strlen(rec_hdr_str) + rec_hdr_fmt_size * control->ras_num_recs;
807 void amdgpu_ras_debugfs_set_ret_size(struct amdgpu_ras_eeprom_control *control)
809 struct amdgpu_ras *ras = container_of(control, struct amdgpu_ras,
811 struct dentry *de = ras->de_ras_eeprom_table;
814 d_inode(de)->i_size = amdgpu_ras_debugfs_table_size(control);
817 static ssize_t amdgpu_ras_debugfs_table_read(struct file *f, char __user *buf,
818 size_t size, loff_t *pos)
820 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
821 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
822 struct amdgpu_ras_eeprom_control *control = &ras->eeprom_control;
823 const size_t orig_size = size;
827 mutex_lock(&control->ras_tbl_mutex);
829 /* We want *pos - data_len > 0, which means there's
830 * bytes to be printed from data.
832 data_len = strlen(tbl_hdr_str);
833 if (*pos < data_len) {
835 data_len = min_t(size_t, data_len, size);
836 if (copy_to_user(buf, &tbl_hdr_str[*pos], data_len))
843 data_len = strlen(tbl_hdr_str) + tbl_hdr_fmt_size;
844 if (*pos < data_len && size > 0) {
845 u8 data[tbl_hdr_fmt_size + 1];
848 snprintf(data, sizeof(data), tbl_hdr_fmt,
849 control->tbl_hdr.header,
850 control->tbl_hdr.version,
851 control->tbl_hdr.first_rec_offset,
852 control->tbl_hdr.tbl_size,
853 control->tbl_hdr.checksum);
856 data_len = min_t(size_t, data_len, size);
857 lpos = *pos - strlen(tbl_hdr_str);
858 if (copy_to_user(buf, &data[lpos], data_len))
865 data_len = strlen(tbl_hdr_str) + tbl_hdr_fmt_size + strlen(rec_hdr_str);
866 if (*pos < data_len && size > 0) {
870 data_len = min_t(size_t, data_len, size);
871 lpos = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size;
872 if (copy_to_user(buf, &rec_hdr_str[lpos], data_len))
879 data_len = amdgpu_ras_debugfs_table_size(control);
880 if (*pos < data_len && size > 0) {
881 u8 dare[RAS_TABLE_RECORD_SIZE];
882 u8 data[rec_hdr_fmt_size + 1];
883 struct eeprom_table_record record;
886 /* Find the starting record index
888 s = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size -
890 s = s / rec_hdr_fmt_size;
891 r = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size -
893 r = r % rec_hdr_fmt_size;
895 for ( ; size > 0 && s < control->ras_num_recs; s++) {
896 u32 ai = RAS_RI_TO_AI(control, s);
897 /* Read a single record
899 res = __amdgpu_ras_eeprom_read(control, dare, ai, 1);
902 __decode_table_record_from_buf(control, &record, dare);
903 snprintf(data, sizeof(data), rec_hdr_fmt,
905 RAS_INDEX_TO_OFFSET(control, ai),
906 record_err_type_str[record.err_type],
912 record.retired_page);
914 data_len = min_t(size_t, rec_hdr_fmt_size - r, size);
915 if (copy_to_user(buf, &data[r], data_len)) {
927 mutex_unlock(&control->ras_tbl_mutex);
928 return res < 0 ? res : orig_size - size;
932 amdgpu_ras_debugfs_eeprom_table_read(struct file *f, char __user *buf,
933 size_t size, loff_t *pos)
935 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
936 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
937 struct amdgpu_ras_eeprom_control *control = ras ? &ras->eeprom_control : NULL;
944 if (!ras || !control) {
945 res = snprintf(data, sizeof(data), "Not supported\n");
950 res = min_t(size_t, res, size);
952 if (copy_to_user(buf, &data[*pos], res))
959 return amdgpu_ras_debugfs_table_read(f, buf, size, pos);
963 const struct file_operations amdgpu_ras_debugfs_eeprom_table_ops = {
964 .owner = THIS_MODULE,
965 .read = amdgpu_ras_debugfs_eeprom_table_read,
967 .llseek = default_llseek,
971 * __verify_ras_table_checksum -- verify the RAS EEPROM table checksum
972 * @control: pointer to control structure
974 * Check the checksum of the stored in EEPROM RAS table.
976 * Return 0 if the checksum is correct,
977 * positive if it is not correct, and
978 * -errno on I/O error.
980 static int __verify_ras_table_checksum(struct amdgpu_ras_eeprom_control *control)
982 struct amdgpu_device *adev = to_amdgpu_device(control);
986 buf_size = RAS_TABLE_HEADER_SIZE +
987 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
988 buf = kzalloc(buf_size, GFP_KERNEL);
990 DRM_ERROR("Out of memory checking RAS table checksum.\n");
994 res = amdgpu_eeprom_read(&adev->pm.smu_i2c,
995 control->i2c_address +
996 control->ras_header_offset,
998 if (res < buf_size) {
999 DRM_ERROR("Partial read for checksum, res:%d\n", res);
1000 /* On partial reads, return -EIO.
1008 for (pp = buf; pp < buf + buf_size; pp++)
1012 return res < 0 ? res : csum;
1015 int amdgpu_ras_eeprom_init(struct amdgpu_ras_eeprom_control *control,
1016 bool *exceed_err_limit)
1018 struct amdgpu_device *adev = to_amdgpu_device(control);
1019 unsigned char buf[RAS_TABLE_HEADER_SIZE] = { 0 };
1020 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
1021 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1024 *exceed_err_limit = false;
1026 if (!__is_ras_eeprom_supported(adev))
1029 /* Verify i2c adapter is initialized */
1030 if (!adev->pm.smu_i2c.algo)
1033 if (!__get_eeprom_i2c_addr(adev, control))
1036 control->ras_header_offset = RAS_HDR_START;
1037 control->ras_record_offset = RAS_RECORD_START;
1038 control->ras_max_record_count = RAS_MAX_RECORD_COUNT;
1039 mutex_init(&control->ras_tbl_mutex);
1041 /* Read the table header from EEPROM address */
1042 res = amdgpu_eeprom_read(&adev->pm.smu_i2c,
1043 control->i2c_address + control->ras_header_offset,
1044 buf, RAS_TABLE_HEADER_SIZE);
1045 if (res < RAS_TABLE_HEADER_SIZE) {
1046 DRM_ERROR("Failed to read EEPROM table header, res:%d", res);
1047 return res >= 0 ? -EIO : res;
1050 __decode_table_header_from_buf(hdr, buf);
1052 control->ras_num_recs = RAS_NUM_RECS(hdr);
1053 control->ras_fri = RAS_OFFSET_TO_INDEX(control, hdr->first_rec_offset);
1055 if (hdr->header == RAS_TABLE_HDR_VAL) {
1056 DRM_DEBUG_DRIVER("Found existing EEPROM table with %d records",
1057 control->ras_num_recs);
1058 res = __verify_ras_table_checksum(control);
1060 DRM_ERROR("RAS table incorrect checksum or error:%d\n",
1062 } else if (hdr->header == RAS_TABLE_HDR_BAD &&
1063 amdgpu_bad_page_threshold != 0) {
1064 res = __verify_ras_table_checksum(control);
1066 DRM_ERROR("RAS Table incorrect checksum or error:%d\n",
1068 if (ras->bad_page_cnt_threshold > control->ras_num_recs) {
1069 /* This means that, the threshold was increased since
1070 * the last time the system was booted, and now,
1071 * ras->bad_page_cnt_threshold - control->num_recs > 0,
1072 * so that at least one more record can be saved,
1073 * before the page count threshold is reached.
1076 "records:%d threshold:%d, resetting "
1077 "RAS table header signature",
1078 control->ras_num_recs,
1079 ras->bad_page_cnt_threshold);
1080 res = amdgpu_ras_eeprom_correct_header_tag(control,
1083 *exceed_err_limit = true;
1085 "RAS records:%d exceed threshold:%d, "
1086 "maybe retire this GPU?",
1087 control->ras_num_recs, ras->bad_page_cnt_threshold);
1090 DRM_INFO("Creating a new EEPROM table");
1092 res = amdgpu_ras_eeprom_reset_table(control);
1095 return res < 0 ? res : 0;