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 #include "amdgpu_reset.h"
36 /* These are memory addresses as would be seen by one or more EEPROM
37 * chips strung on the I2C bus, usually by manipulating pins 1-3 of a
38 * set of EEPROM devices. They form a continuous memory space.
40 * The I2C device address includes the device type identifier, 1010b,
41 * which is a reserved value and indicates that this is an I2C EEPROM
42 * device. It also includes the top 3 bits of the 19 bit EEPROM memory
43 * address, namely bits 18, 17, and 16. This makes up the 7 bit
44 * address sent on the I2C bus with bit 0 being the direction bit,
45 * which is not represented here, and sent by the hardware directly.
48 * 50h = 1010000b => device type identifier 1010b, bits 18:16 = 000b, address 0.
49 * 54h = 1010100b => --"--, bits 18:16 = 100b, address 40000h.
50 * 56h = 1010110b => --"--, bits 18:16 = 110b, address 60000h.
51 * Depending on the size of the I2C EEPROM device(s), bits 18:16 may
52 * address memory in a device or a device on the I2C bus, depending on
53 * the status of pins 1-3. See top of amdgpu_eeprom.c.
55 * The RAS table lives either at address 0 or address 40000h of EEPROM.
57 #define EEPROM_I2C_MADDR_0 0x0
58 #define EEPROM_I2C_MADDR_4 0x40000
61 * The 2 macros bellow represent the actual size in bytes that
62 * those entities occupy in the EEPROM memory.
63 * RAS_TABLE_RECORD_SIZE is different than sizeof(eeprom_table_record) which
64 * uses uint64 to store 6b fields such as retired_page.
66 #define RAS_TABLE_HEADER_SIZE 20
67 #define RAS_TABLE_RECORD_SIZE 24
69 /* Table hdr is 'AMDR' */
70 #define RAS_TABLE_HDR_VAL 0x414d4452
71 #define RAS_TABLE_VER 0x00010000
73 /* Bad GPU tag ‘BADG’ */
74 #define RAS_TABLE_HDR_BAD 0x42414447
76 /* Assume 2-Mbit size EEPROM and take up the whole space. */
77 #define RAS_TBL_SIZE_BYTES (256 * 1024)
78 #define RAS_TABLE_START 0
79 #define RAS_HDR_START RAS_TABLE_START
80 #define RAS_RECORD_START (RAS_HDR_START + RAS_TABLE_HEADER_SIZE)
81 #define RAS_MAX_RECORD_COUNT ((RAS_TBL_SIZE_BYTES - RAS_TABLE_HEADER_SIZE) \
82 / RAS_TABLE_RECORD_SIZE)
84 /* Given a zero-based index of an EEPROM RAS record, yields the EEPROM
85 * offset off of RAS_TABLE_START. That is, this is something you can
86 * add to control->i2c_address, and then tell I2C layer to read
87 * from/write to there. _N is the so called absolute index,
88 * because it starts right after the table header.
90 #define RAS_INDEX_TO_OFFSET(_C, _N) ((_C)->ras_record_offset + \
91 (_N) * RAS_TABLE_RECORD_SIZE)
93 #define RAS_OFFSET_TO_INDEX(_C, _O) (((_O) - \
94 (_C)->ras_record_offset) / RAS_TABLE_RECORD_SIZE)
96 /* Given a 0-based relative record index, 0, 1, 2, ..., etc., off
97 * of "fri", return the absolute record index off of the end of
100 #define RAS_RI_TO_AI(_C, _I) (((_I) + (_C)->ras_fri) % \
101 (_C)->ras_max_record_count)
103 #define RAS_NUM_RECS(_tbl_hdr) (((_tbl_hdr)->tbl_size - \
104 RAS_TABLE_HEADER_SIZE) / RAS_TABLE_RECORD_SIZE)
106 #define to_amdgpu_device(x) (container_of(x, struct amdgpu_ras, eeprom_control))->adev
108 static bool __is_ras_eeprom_supported(struct amdgpu_device *adev)
110 switch (adev->ip_versions[MP1_HWIP][0]) {
111 case IP_VERSION(11, 0, 2): /* VEGA20 and ARCTURUS */
112 case IP_VERSION(11, 0, 7): /* Sienna cichlid */
113 case IP_VERSION(13, 0, 0):
114 case IP_VERSION(13, 0, 2): /* Aldebaran */
115 case IP_VERSION(13, 0, 10):
122 static bool __get_eeprom_i2c_addr(struct amdgpu_device *adev,
123 struct amdgpu_ras_eeprom_control *control)
125 struct atom_context *atom_ctx = adev->mode_info.atom_context;
131 if (amdgpu_atomfirmware_ras_rom_addr(adev, &i2c_addr)) {
132 /* The address given by VBIOS is an 8-bit, wire-format
133 * address, i.e. the most significant byte.
135 * Normalize it to a 19-bit EEPROM address. Remove the
136 * device type identifier and make it a 7-bit address;
137 * then make it a 19-bit EEPROM address. See top of
140 i2c_addr = (i2c_addr & 0x0F) >> 1;
141 control->i2c_address = ((u32) i2c_addr) << 16;
146 switch (adev->ip_versions[MP1_HWIP][0]) {
147 case IP_VERSION(11, 0, 2):
148 /* VEGA20 and ARCTURUS */
149 if (adev->asic_type == CHIP_VEGA20)
150 control->i2c_address = EEPROM_I2C_MADDR_0;
151 else if (strnstr(atom_ctx->vbios_version,
153 sizeof(atom_ctx->vbios_version)))
154 control->i2c_address = EEPROM_I2C_MADDR_0;
156 control->i2c_address = EEPROM_I2C_MADDR_4;
158 case IP_VERSION(11, 0, 7):
159 control->i2c_address = EEPROM_I2C_MADDR_0;
161 case IP_VERSION(13, 0, 2):
162 if (strnstr(atom_ctx->vbios_version, "D673",
163 sizeof(atom_ctx->vbios_version)))
164 control->i2c_address = EEPROM_I2C_MADDR_4;
166 control->i2c_address = EEPROM_I2C_MADDR_0;
168 case IP_VERSION(13, 0, 0):
169 case IP_VERSION(13, 0, 10):
170 control->i2c_address = EEPROM_I2C_MADDR_4;
178 __encode_table_header_to_buf(struct amdgpu_ras_eeprom_table_header *hdr,
181 u32 *pp = (uint32_t *)buf;
183 pp[0] = cpu_to_le32(hdr->header);
184 pp[1] = cpu_to_le32(hdr->version);
185 pp[2] = cpu_to_le32(hdr->first_rec_offset);
186 pp[3] = cpu_to_le32(hdr->tbl_size);
187 pp[4] = cpu_to_le32(hdr->checksum);
191 __decode_table_header_from_buf(struct amdgpu_ras_eeprom_table_header *hdr,
194 u32 *pp = (uint32_t *)buf;
196 hdr->header = le32_to_cpu(pp[0]);
197 hdr->version = le32_to_cpu(pp[1]);
198 hdr->first_rec_offset = le32_to_cpu(pp[2]);
199 hdr->tbl_size = le32_to_cpu(pp[3]);
200 hdr->checksum = le32_to_cpu(pp[4]);
203 static int __write_table_header(struct amdgpu_ras_eeprom_control *control)
205 u8 buf[RAS_TABLE_HEADER_SIZE];
206 struct amdgpu_device *adev = to_amdgpu_device(control);
209 memset(buf, 0, sizeof(buf));
210 __encode_table_header_to_buf(&control->tbl_hdr, buf);
212 /* i2c may be unstable in gpu reset */
213 down_read(&adev->reset_domain->sem);
214 res = amdgpu_eeprom_write(adev->pm.ras_eeprom_i2c_bus,
215 control->i2c_address +
216 control->ras_header_offset,
217 buf, RAS_TABLE_HEADER_SIZE);
218 up_read(&adev->reset_domain->sem);
221 DRM_ERROR("Failed to write EEPROM table header:%d", res);
222 } else if (res < RAS_TABLE_HEADER_SIZE) {
223 DRM_ERROR("Short write:%d out of %d\n",
224 res, RAS_TABLE_HEADER_SIZE);
233 static u8 __calc_hdr_byte_sum(const struct amdgpu_ras_eeprom_control *control)
239 /* Header checksum, skip checksum field in the calculation */
240 sz = sizeof(control->tbl_hdr) - sizeof(control->tbl_hdr.checksum);
241 pp = (u8 *) &control->tbl_hdr;
243 for (ii = 0; ii < sz; ii++, pp++)
249 static int amdgpu_ras_eeprom_correct_header_tag(
250 struct amdgpu_ras_eeprom_control *control,
253 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
258 csum = -hdr->checksum;
260 hh = (void *) &hdr->header;
261 csum -= (hh[0] + hh[1] + hh[2] + hh[3]);
262 hh = (void *) &header;
263 csum += hh[0] + hh[1] + hh[2] + hh[3];
265 mutex_lock(&control->ras_tbl_mutex);
266 hdr->header = header;
267 hdr->checksum = csum;
268 res = __write_table_header(control);
269 mutex_unlock(&control->ras_tbl_mutex);
275 * amdgpu_ras_eeprom_reset_table -- Reset the RAS EEPROM table
276 * @control: pointer to control structure
278 * Reset the contents of the header of the RAS EEPROM table.
279 * Return 0 on success, -errno on error.
281 int amdgpu_ras_eeprom_reset_table(struct amdgpu_ras_eeprom_control *control)
283 struct amdgpu_device *adev = to_amdgpu_device(control);
284 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
285 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
289 mutex_lock(&control->ras_tbl_mutex);
291 hdr->header = RAS_TABLE_HDR_VAL;
292 hdr->version = RAS_TABLE_VER;
293 hdr->first_rec_offset = RAS_RECORD_START;
294 hdr->tbl_size = RAS_TABLE_HEADER_SIZE;
296 csum = __calc_hdr_byte_sum(control);
298 hdr->checksum = csum;
299 res = __write_table_header(control);
301 control->ras_num_recs = 0;
302 control->ras_fri = 0;
304 amdgpu_dpm_send_hbm_bad_pages_num(adev, control->ras_num_recs);
306 control->bad_channel_bitmap = 0;
307 amdgpu_dpm_send_hbm_bad_channel_flag(adev, control->bad_channel_bitmap);
308 con->update_channel_flag = false;
310 amdgpu_ras_debugfs_set_ret_size(control);
312 mutex_unlock(&control->ras_tbl_mutex);
318 __encode_table_record_to_buf(struct amdgpu_ras_eeprom_control *control,
319 struct eeprom_table_record *record,
325 /* Next are all record fields according to EEPROM page spec in LE foramt */
326 buf[i++] = record->err_type;
328 buf[i++] = record->bank;
330 tmp = cpu_to_le64(record->ts);
331 memcpy(buf + i, &tmp, 8);
334 tmp = cpu_to_le64((record->offset & 0xffffffffffff));
335 memcpy(buf + i, &tmp, 6);
338 buf[i++] = record->mem_channel;
339 buf[i++] = record->mcumc_id;
341 tmp = cpu_to_le64((record->retired_page & 0xffffffffffff));
342 memcpy(buf + i, &tmp, 6);
346 __decode_table_record_from_buf(struct amdgpu_ras_eeprom_control *control,
347 struct eeprom_table_record *record,
353 /* Next are all record fields according to EEPROM page spec in LE foramt */
354 record->err_type = buf[i++];
356 record->bank = buf[i++];
358 memcpy(&tmp, buf + i, 8);
359 record->ts = le64_to_cpu(tmp);
362 memcpy(&tmp, buf + i, 6);
363 record->offset = (le64_to_cpu(tmp) & 0xffffffffffff);
366 record->mem_channel = buf[i++];
367 record->mcumc_id = buf[i++];
369 memcpy(&tmp, buf + i, 6);
370 record->retired_page = (le64_to_cpu(tmp) & 0xffffffffffff);
373 bool amdgpu_ras_eeprom_check_err_threshold(struct amdgpu_device *adev)
375 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
377 if (!__is_ras_eeprom_supported(adev) ||
378 !amdgpu_bad_page_threshold)
381 /* skip check eeprom table for VEGA20 Gaming */
385 if (!(con->features & BIT(AMDGPU_RAS_BLOCK__UMC)))
388 if (con->eeprom_control.tbl_hdr.header == RAS_TABLE_HDR_BAD) {
389 if (amdgpu_bad_page_threshold == -1) {
390 dev_warn(adev->dev, "RAS records:%d exceed threshold:%d",
391 con->eeprom_control.ras_num_recs, con->bad_page_cnt_threshold);
393 "But GPU can be operated due to bad_page_threshold = -1.\n");
396 dev_warn(adev->dev, "This GPU is in BAD status.");
397 dev_warn(adev->dev, "Please retire it or set a larger "
398 "threshold value when reloading driver.\n");
407 * __amdgpu_ras_eeprom_write -- write indexed from buffer to EEPROM
408 * @control: pointer to control structure
409 * @buf: pointer to buffer containing data to write
410 * @fri: start writing at this index
411 * @num: number of records to write
413 * The caller must hold the table mutex in @control.
414 * Return 0 on success, -errno otherwise.
416 static int __amdgpu_ras_eeprom_write(struct amdgpu_ras_eeprom_control *control,
417 u8 *buf, const u32 fri, const u32 num)
419 struct amdgpu_device *adev = to_amdgpu_device(control);
423 /* i2c may be unstable in gpu reset */
424 down_read(&adev->reset_domain->sem);
425 buf_size = num * RAS_TABLE_RECORD_SIZE;
426 res = amdgpu_eeprom_write(adev->pm.ras_eeprom_i2c_bus,
427 control->i2c_address +
428 RAS_INDEX_TO_OFFSET(control, fri),
430 up_read(&adev->reset_domain->sem);
432 DRM_ERROR("Writing %d EEPROM table records error:%d",
434 } else if (res < buf_size) {
435 /* Short write, return error.
437 DRM_ERROR("Wrote %d records out of %d",
438 res / RAS_TABLE_RECORD_SIZE, num);
448 amdgpu_ras_eeprom_append_table(struct amdgpu_ras_eeprom_control *control,
449 struct eeprom_table_record *record,
452 struct amdgpu_ras *con = amdgpu_ras_get_context(to_amdgpu_device(control));
457 buf = kcalloc(num, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
461 /* Encode all of them in one go.
464 for (i = 0; i < num; i++, pp += RAS_TABLE_RECORD_SIZE) {
465 __encode_table_record_to_buf(control, &record[i], pp);
467 /* update bad channel bitmap */
468 if (!(control->bad_channel_bitmap & (1 << record[i].mem_channel))) {
469 control->bad_channel_bitmap |= 1 << record[i].mem_channel;
470 con->update_channel_flag = true;
474 /* a, first record index to write into.
475 * b, last record index to write into.
476 * a = first index to read (fri) + number of records in the table,
478 * Let N = control->ras_max_num_record_count, then we have,
479 * case 0: 0 <= a <= b < N,
480 * just append @num records starting at a;
481 * case 1: 0 <= a < N <= b,
482 * append (N - a) records starting at a, and
483 * append the remainder, b % N + 1, starting at 0.
484 * case 2: 0 <= fri < N <= a <= b, then modulo N we get two subcases,
485 * case 2a: 0 <= a <= b < N
486 * append num records starting at a; and fix fri if b overwrote it,
487 * and since a <= b, if b overwrote it then a must've also,
488 * and if b didn't overwrite it, then a didn't also.
489 * case 2b: 0 <= b < a < N
490 * write num records starting at a, which wraps around 0=N
491 * and overwrite fri unconditionally. Now from case 2a,
492 * this means that b eclipsed fri to overwrite it and wrap
493 * around 0 again, i.e. b = 2N+r pre modulo N, so we unconditionally
494 * set fri = b + 1 (mod N).
495 * Now, since fri is updated in every case, except the trivial case 0,
496 * the number of records present in the table after writing, is,
497 * num_recs - 1 = b - fri (mod N), and we take the positive value,
498 * by adding an arbitrary multiple of N before taking the modulo N
501 a = control->ras_fri + control->ras_num_recs;
503 if (b < control->ras_max_record_count) {
504 res = __amdgpu_ras_eeprom_write(control, buf, a, num);
505 } else if (a < control->ras_max_record_count) {
508 g0 = control->ras_max_record_count - a;
509 g1 = b % control->ras_max_record_count + 1;
510 res = __amdgpu_ras_eeprom_write(control, buf, a, g0);
513 res = __amdgpu_ras_eeprom_write(control,
514 buf + g0 * RAS_TABLE_RECORD_SIZE,
518 if (g1 > control->ras_fri)
519 control->ras_fri = g1 % control->ras_max_record_count;
521 a %= control->ras_max_record_count;
522 b %= control->ras_max_record_count;
525 /* Note that, b - a + 1 = num. */
526 res = __amdgpu_ras_eeprom_write(control, buf, a, num);
529 if (b >= control->ras_fri)
530 control->ras_fri = (b + 1) % control->ras_max_record_count;
534 /* b < a, which means, we write from
535 * a to the end of the table, and from
536 * the start of the table to b.
538 g0 = control->ras_max_record_count - a;
540 res = __amdgpu_ras_eeprom_write(control, buf, a, g0);
543 res = __amdgpu_ras_eeprom_write(control,
544 buf + g0 * RAS_TABLE_RECORD_SIZE,
548 control->ras_fri = g1 % control->ras_max_record_count;
551 control->ras_num_recs = 1 + (control->ras_max_record_count + b
553 % control->ras_max_record_count;
560 amdgpu_ras_eeprom_update_header(struct amdgpu_ras_eeprom_control *control)
562 struct amdgpu_device *adev = to_amdgpu_device(control);
563 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
568 /* Modify the header if it exceeds.
570 if (amdgpu_bad_page_threshold != 0 &&
571 control->ras_num_recs >= ras->bad_page_cnt_threshold) {
573 "Saved bad pages %d reaches threshold value %d\n",
574 control->ras_num_recs, ras->bad_page_cnt_threshold);
575 control->tbl_hdr.header = RAS_TABLE_HDR_BAD;
578 control->tbl_hdr.version = RAS_TABLE_VER;
579 control->tbl_hdr.first_rec_offset = RAS_INDEX_TO_OFFSET(control, control->ras_fri);
580 control->tbl_hdr.tbl_size = RAS_TABLE_HEADER_SIZE + control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
581 control->tbl_hdr.checksum = 0;
583 buf_size = control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
584 buf = kcalloc(control->ras_num_recs, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
586 DRM_ERROR("allocating memory for table of size %d bytes failed\n",
587 control->tbl_hdr.tbl_size);
592 down_read(&adev->reset_domain->sem);
593 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
594 control->i2c_address +
595 control->ras_record_offset,
597 up_read(&adev->reset_domain->sem);
599 DRM_ERROR("EEPROM failed reading records:%d\n",
602 } else if (res < buf_size) {
603 DRM_ERROR("EEPROM read %d out of %d bytes\n",
609 /* Recalc the checksum.
612 for (pp = buf; pp < buf + buf_size; pp++)
615 csum += __calc_hdr_byte_sum(control);
616 /* avoid sign extension when assigning to "checksum" */
618 control->tbl_hdr.checksum = csum;
619 res = __write_table_header(control);
626 * amdgpu_ras_eeprom_append -- append records to the EEPROM RAS table
627 * @control: pointer to control structure
628 * @record: array of records to append
629 * @num: number of records in @record array
631 * Append @num records to the table, calculate the checksum and write
632 * the table back to EEPROM. The maximum number of records that
633 * can be appended is between 1 and control->ras_max_record_count,
634 * regardless of how many records are already stored in the table.
636 * Return 0 on success or if EEPROM is not supported, -errno on error.
638 int amdgpu_ras_eeprom_append(struct amdgpu_ras_eeprom_control *control,
639 struct eeprom_table_record *record,
642 struct amdgpu_device *adev = to_amdgpu_device(control);
645 if (!__is_ras_eeprom_supported(adev))
649 DRM_ERROR("will not append 0 records\n");
651 } else if (num > control->ras_max_record_count) {
652 DRM_ERROR("cannot append %d records than the size of table %d\n",
653 num, control->ras_max_record_count);
657 mutex_lock(&control->ras_tbl_mutex);
659 res = amdgpu_ras_eeprom_append_table(control, record, num);
661 res = amdgpu_ras_eeprom_update_header(control);
663 amdgpu_ras_debugfs_set_ret_size(control);
665 mutex_unlock(&control->ras_tbl_mutex);
670 * __amdgpu_ras_eeprom_read -- read indexed from EEPROM into buffer
671 * @control: pointer to control structure
672 * @buf: pointer to buffer to read into
673 * @fri: first record index, start reading at this index, absolute index
674 * @num: number of records to read
676 * The caller must hold the table mutex in @control.
677 * Return 0 on success, -errno otherwise.
679 static int __amdgpu_ras_eeprom_read(struct amdgpu_ras_eeprom_control *control,
680 u8 *buf, const u32 fri, const u32 num)
682 struct amdgpu_device *adev = to_amdgpu_device(control);
686 /* i2c may be unstable in gpu reset */
687 down_read(&adev->reset_domain->sem);
688 buf_size = num * RAS_TABLE_RECORD_SIZE;
689 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
690 control->i2c_address +
691 RAS_INDEX_TO_OFFSET(control, fri),
693 up_read(&adev->reset_domain->sem);
695 DRM_ERROR("Reading %d EEPROM table records error:%d",
697 } else if (res < buf_size) {
698 /* Short read, return error.
700 DRM_ERROR("Read %d records out of %d",
701 res / RAS_TABLE_RECORD_SIZE, num);
711 * amdgpu_ras_eeprom_read -- read EEPROM
712 * @control: pointer to control structure
713 * @record: array of records to read into
714 * @num: number of records in @record
716 * Reads num records from the RAS table in EEPROM and
717 * writes the data into @record array.
719 * Returns 0 on success, -errno on error.
721 int amdgpu_ras_eeprom_read(struct amdgpu_ras_eeprom_control *control,
722 struct eeprom_table_record *record,
725 struct amdgpu_device *adev = to_amdgpu_device(control);
726 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
731 if (!__is_ras_eeprom_supported(adev))
735 DRM_ERROR("will not read 0 records\n");
737 } else if (num > control->ras_num_recs) {
738 DRM_ERROR("too many records to read:%d available:%d\n",
739 num, control->ras_num_recs);
743 buf = kcalloc(num, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
747 /* Determine how many records to read, from the first record
748 * index, fri, to the end of the table, and from the beginning
749 * of the table, such that the total number of records is
750 * @num, and we handle wrap around when fri > 0 and
751 * fri + num > RAS_MAX_RECORD_COUNT.
753 * First we compute the index of the last element
754 * which would be fetched from each region,
755 * g0 is in [fri, fri + num - 1], and
756 * g1 is in [0, RAS_MAX_RECORD_COUNT - 1].
757 * Then, if g0 < RAS_MAX_RECORD_COUNT, the index of
758 * the last element to fetch, we set g0 to _the number_
759 * of elements to fetch, @num, since we know that the last
760 * indexed to be fetched does not exceed the table.
762 * If, however, g0 >= RAS_MAX_RECORD_COUNT, then
763 * we set g0 to the number of elements to read
764 * until the end of the table, and g1 to the number of
765 * elements to read from the beginning of the table.
767 g0 = control->ras_fri + num - 1;
768 g1 = g0 % control->ras_max_record_count;
769 if (g0 < control->ras_max_record_count) {
773 g0 = control->ras_max_record_count - control->ras_fri;
777 mutex_lock(&control->ras_tbl_mutex);
778 res = __amdgpu_ras_eeprom_read(control, buf, control->ras_fri, g0);
782 res = __amdgpu_ras_eeprom_read(control,
783 buf + g0 * RAS_TABLE_RECORD_SIZE,
791 /* Read up everything? Then transform.
794 for (i = 0; i < num; i++, pp += RAS_TABLE_RECORD_SIZE) {
795 __decode_table_record_from_buf(control, &record[i], pp);
797 /* update bad channel bitmap */
798 if (!(control->bad_channel_bitmap & (1 << record[i].mem_channel))) {
799 control->bad_channel_bitmap |= 1 << record[i].mem_channel;
800 con->update_channel_flag = true;
805 mutex_unlock(&control->ras_tbl_mutex);
810 uint32_t amdgpu_ras_eeprom_max_record_count(void)
812 return RAS_MAX_RECORD_COUNT;
816 amdgpu_ras_debugfs_eeprom_size_read(struct file *f, char __user *buf,
817 size_t size, loff_t *pos)
819 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
820 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
821 struct amdgpu_ras_eeprom_control *control = ras ? &ras->eeprom_control : NULL;
828 if (!ras || !control) {
829 res = snprintf(data, sizeof(data), "Not supported\n");
831 res = snprintf(data, sizeof(data), "%d bytes or %d records\n",
832 RAS_TBL_SIZE_BYTES, control->ras_max_record_count);
839 res = min_t(size_t, res, size);
841 if (copy_to_user(buf, &data[*pos], res))
849 const struct file_operations amdgpu_ras_debugfs_eeprom_size_ops = {
850 .owner = THIS_MODULE,
851 .read = amdgpu_ras_debugfs_eeprom_size_read,
853 .llseek = default_llseek,
856 static const char *tbl_hdr_str = " Signature Version FirstOffs Size Checksum\n";
857 static const char *tbl_hdr_fmt = "0x%08X 0x%08X 0x%08X 0x%08X 0x%08X\n";
858 #define tbl_hdr_fmt_size (5 * (2+8) + 4 + 1)
859 static const char *rec_hdr_str = "Index Offset ErrType Bank/CU TimeStamp Offs/Addr MemChl MCUMCID RetiredPage\n";
860 static const char *rec_hdr_fmt = "%5d 0x%05X %7s 0x%02X 0x%016llX 0x%012llX 0x%02X 0x%02X 0x%012llX\n";
861 #define rec_hdr_fmt_size (5 + 1 + 7 + 1 + 7 + 1 + 7 + 1 + 18 + 1 + 14 + 1 + 6 + 1 + 7 + 1 + 14 + 1)
863 static const char *record_err_type_str[AMDGPU_RAS_EEPROM_ERR_COUNT] = {
869 static loff_t amdgpu_ras_debugfs_table_size(struct amdgpu_ras_eeprom_control *control)
871 return strlen(tbl_hdr_str) + tbl_hdr_fmt_size +
872 strlen(rec_hdr_str) + rec_hdr_fmt_size * control->ras_num_recs;
875 void amdgpu_ras_debugfs_set_ret_size(struct amdgpu_ras_eeprom_control *control)
877 struct amdgpu_ras *ras = container_of(control, struct amdgpu_ras,
879 struct dentry *de = ras->de_ras_eeprom_table;
882 d_inode(de)->i_size = amdgpu_ras_debugfs_table_size(control);
885 static ssize_t amdgpu_ras_debugfs_table_read(struct file *f, char __user *buf,
886 size_t size, loff_t *pos)
888 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
889 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
890 struct amdgpu_ras_eeprom_control *control = &ras->eeprom_control;
891 const size_t orig_size = size;
895 mutex_lock(&control->ras_tbl_mutex);
897 /* We want *pos - data_len > 0, which means there's
898 * bytes to be printed from data.
900 data_len = strlen(tbl_hdr_str);
901 if (*pos < data_len) {
903 data_len = min_t(size_t, data_len, size);
904 if (copy_to_user(buf, &tbl_hdr_str[*pos], data_len))
911 data_len = strlen(tbl_hdr_str) + tbl_hdr_fmt_size;
912 if (*pos < data_len && size > 0) {
913 u8 data[tbl_hdr_fmt_size + 1];
916 snprintf(data, sizeof(data), tbl_hdr_fmt,
917 control->tbl_hdr.header,
918 control->tbl_hdr.version,
919 control->tbl_hdr.first_rec_offset,
920 control->tbl_hdr.tbl_size,
921 control->tbl_hdr.checksum);
924 data_len = min_t(size_t, data_len, size);
925 lpos = *pos - strlen(tbl_hdr_str);
926 if (copy_to_user(buf, &data[lpos], data_len))
933 data_len = strlen(tbl_hdr_str) + tbl_hdr_fmt_size + strlen(rec_hdr_str);
934 if (*pos < data_len && size > 0) {
938 data_len = min_t(size_t, data_len, size);
939 lpos = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size;
940 if (copy_to_user(buf, &rec_hdr_str[lpos], data_len))
947 data_len = amdgpu_ras_debugfs_table_size(control);
948 if (*pos < data_len && size > 0) {
949 u8 dare[RAS_TABLE_RECORD_SIZE];
950 u8 data[rec_hdr_fmt_size + 1];
951 struct eeprom_table_record record;
954 /* Find the starting record index
956 s = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size -
958 s = s / rec_hdr_fmt_size;
959 r = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size -
961 r = r % rec_hdr_fmt_size;
963 for ( ; size > 0 && s < control->ras_num_recs; s++) {
964 u32 ai = RAS_RI_TO_AI(control, s);
965 /* Read a single record
967 res = __amdgpu_ras_eeprom_read(control, dare, ai, 1);
970 __decode_table_record_from_buf(control, &record, dare);
971 snprintf(data, sizeof(data), rec_hdr_fmt,
973 RAS_INDEX_TO_OFFSET(control, ai),
974 record_err_type_str[record.err_type],
980 record.retired_page);
982 data_len = min_t(size_t, rec_hdr_fmt_size - r, size);
983 if (copy_to_user(buf, &data[r], data_len)) {
995 mutex_unlock(&control->ras_tbl_mutex);
996 return res < 0 ? res : orig_size - size;
1000 amdgpu_ras_debugfs_eeprom_table_read(struct file *f, char __user *buf,
1001 size_t size, loff_t *pos)
1003 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
1004 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1005 struct amdgpu_ras_eeprom_control *control = ras ? &ras->eeprom_control : NULL;
1012 if (!ras || !control) {
1013 res = snprintf(data, sizeof(data), "Not supported\n");
1018 res = min_t(size_t, res, size);
1020 if (copy_to_user(buf, &data[*pos], res))
1027 return amdgpu_ras_debugfs_table_read(f, buf, size, pos);
1031 const struct file_operations amdgpu_ras_debugfs_eeprom_table_ops = {
1032 .owner = THIS_MODULE,
1033 .read = amdgpu_ras_debugfs_eeprom_table_read,
1035 .llseek = default_llseek,
1039 * __verify_ras_table_checksum -- verify the RAS EEPROM table checksum
1040 * @control: pointer to control structure
1042 * Check the checksum of the stored in EEPROM RAS table.
1044 * Return 0 if the checksum is correct,
1045 * positive if it is not correct, and
1046 * -errno on I/O error.
1048 static int __verify_ras_table_checksum(struct amdgpu_ras_eeprom_control *control)
1050 struct amdgpu_device *adev = to_amdgpu_device(control);
1054 buf_size = RAS_TABLE_HEADER_SIZE +
1055 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
1056 buf = kzalloc(buf_size, GFP_KERNEL);
1058 DRM_ERROR("Out of memory checking RAS table checksum.\n");
1062 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
1063 control->i2c_address +
1064 control->ras_header_offset,
1066 if (res < buf_size) {
1067 DRM_ERROR("Partial read for checksum, res:%d\n", res);
1068 /* On partial reads, return -EIO.
1076 for (pp = buf; pp < buf + buf_size; pp++)
1080 return res < 0 ? res : csum;
1083 int amdgpu_ras_eeprom_init(struct amdgpu_ras_eeprom_control *control,
1084 bool *exceed_err_limit)
1086 struct amdgpu_device *adev = to_amdgpu_device(control);
1087 unsigned char buf[RAS_TABLE_HEADER_SIZE] = { 0 };
1088 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
1089 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1092 *exceed_err_limit = false;
1094 if (!__is_ras_eeprom_supported(adev))
1097 /* Verify i2c adapter is initialized */
1098 if (!adev->pm.ras_eeprom_i2c_bus || !adev->pm.ras_eeprom_i2c_bus->algo)
1101 if (!__get_eeprom_i2c_addr(adev, control))
1104 control->ras_header_offset = RAS_HDR_START;
1105 control->ras_record_offset = RAS_RECORD_START;
1106 control->ras_max_record_count = RAS_MAX_RECORD_COUNT;
1107 mutex_init(&control->ras_tbl_mutex);
1109 /* Read the table header from EEPROM address */
1110 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
1111 control->i2c_address + control->ras_header_offset,
1112 buf, RAS_TABLE_HEADER_SIZE);
1113 if (res < RAS_TABLE_HEADER_SIZE) {
1114 DRM_ERROR("Failed to read EEPROM table header, res:%d", res);
1115 return res >= 0 ? -EIO : res;
1118 __decode_table_header_from_buf(hdr, buf);
1120 control->ras_num_recs = RAS_NUM_RECS(hdr);
1121 control->ras_fri = RAS_OFFSET_TO_INDEX(control, hdr->first_rec_offset);
1123 if (hdr->header == RAS_TABLE_HDR_VAL) {
1124 DRM_DEBUG_DRIVER("Found existing EEPROM table with %d records",
1125 control->ras_num_recs);
1126 res = __verify_ras_table_checksum(control);
1128 DRM_ERROR("RAS table incorrect checksum or error:%d\n",
1131 /* Warn if we are at 90% of the threshold or above
1133 if (10 * control->ras_num_recs >= 9 * ras->bad_page_cnt_threshold)
1134 dev_warn(adev->dev, "RAS records:%u exceeds 90%% of threshold:%d",
1135 control->ras_num_recs,
1136 ras->bad_page_cnt_threshold);
1137 } else if (hdr->header == RAS_TABLE_HDR_BAD &&
1138 amdgpu_bad_page_threshold != 0) {
1139 res = __verify_ras_table_checksum(control);
1141 DRM_ERROR("RAS Table incorrect checksum or error:%d\n",
1143 if (ras->bad_page_cnt_threshold > control->ras_num_recs) {
1144 /* This means that, the threshold was increased since
1145 * the last time the system was booted, and now,
1146 * ras->bad_page_cnt_threshold - control->num_recs > 0,
1147 * so that at least one more record can be saved,
1148 * before the page count threshold is reached.
1151 "records:%d threshold:%d, resetting "
1152 "RAS table header signature",
1153 control->ras_num_recs,
1154 ras->bad_page_cnt_threshold);
1155 res = amdgpu_ras_eeprom_correct_header_tag(control,
1158 dev_err(adev->dev, "RAS records:%d exceed threshold:%d",
1159 control->ras_num_recs, ras->bad_page_cnt_threshold);
1160 if (amdgpu_bad_page_threshold == -1) {
1161 dev_warn(adev->dev, "GPU will be initialized due to bad_page_threshold = -1.");
1164 *exceed_err_limit = true;
1166 "RAS records:%d exceed threshold:%d, "
1167 "GPU will not be initialized. Replace this GPU or increase the threshold",
1168 control->ras_num_recs, ras->bad_page_cnt_threshold);
1172 DRM_INFO("Creating a new EEPROM table");
1174 res = amdgpu_ras_eeprom_reset_table(control);
1177 return res < 0 ? res : 0;