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 if (adev->asic_type == CHIP_IP_DISCOVERY) {
111 switch (adev->ip_versions[MP1_HWIP][0]) {
112 case IP_VERSION(13, 0, 0):
113 case IP_VERSION(13, 0, 10):
120 return adev->asic_type == CHIP_VEGA20 ||
121 adev->asic_type == CHIP_ARCTURUS ||
122 adev->asic_type == CHIP_SIENNA_CICHLID ||
123 adev->asic_type == CHIP_ALDEBARAN;
126 static bool __get_eeprom_i2c_addr_arct(struct amdgpu_device *adev,
127 struct amdgpu_ras_eeprom_control *control)
129 struct atom_context *atom_ctx = adev->mode_info.atom_context;
131 if (!control || !atom_ctx)
134 if (strnstr(atom_ctx->vbios_version,
136 sizeof(atom_ctx->vbios_version)))
137 control->i2c_address = EEPROM_I2C_MADDR_0;
139 control->i2c_address = EEPROM_I2C_MADDR_4;
144 static bool __get_eeprom_i2c_addr_ip_discovery(struct amdgpu_device *adev,
145 struct amdgpu_ras_eeprom_control *control)
147 switch (adev->ip_versions[MP1_HWIP][0]) {
148 case IP_VERSION(13, 0, 0):
149 case IP_VERSION(13, 0, 10):
150 control->i2c_address = EEPROM_I2C_MADDR_4;
157 static bool __get_eeprom_i2c_addr(struct amdgpu_device *adev,
158 struct amdgpu_ras_eeprom_control *control)
160 struct atom_context *atom_ctx = adev->mode_info.atom_context;
166 if (amdgpu_atomfirmware_ras_rom_addr(adev, &i2c_addr)) {
167 /* The address given by VBIOS is an 8-bit, wire-format
168 * address, i.e. the most significant byte.
170 * Normalize it to a 19-bit EEPROM address. Remove the
171 * device type identifier and make it a 7-bit address;
172 * then make it a 19-bit EEPROM address. See top of
175 i2c_addr = (i2c_addr & 0x0F) >> 1;
176 control->i2c_address = ((u32) i2c_addr) << 16;
181 switch (adev->asic_type) {
183 control->i2c_address = EEPROM_I2C_MADDR_0;
187 return __get_eeprom_i2c_addr_arct(adev, control);
189 case CHIP_SIENNA_CICHLID:
190 control->i2c_address = EEPROM_I2C_MADDR_0;
194 if (strnstr(atom_ctx->vbios_version, "D673",
195 sizeof(atom_ctx->vbios_version)))
196 control->i2c_address = EEPROM_I2C_MADDR_4;
198 control->i2c_address = EEPROM_I2C_MADDR_0;
201 case CHIP_IP_DISCOVERY:
202 return __get_eeprom_i2c_addr_ip_discovery(adev, control);
210 __encode_table_header_to_buf(struct amdgpu_ras_eeprom_table_header *hdr,
213 u32 *pp = (uint32_t *)buf;
215 pp[0] = cpu_to_le32(hdr->header);
216 pp[1] = cpu_to_le32(hdr->version);
217 pp[2] = cpu_to_le32(hdr->first_rec_offset);
218 pp[3] = cpu_to_le32(hdr->tbl_size);
219 pp[4] = cpu_to_le32(hdr->checksum);
223 __decode_table_header_from_buf(struct amdgpu_ras_eeprom_table_header *hdr,
226 u32 *pp = (uint32_t *)buf;
228 hdr->header = le32_to_cpu(pp[0]);
229 hdr->version = le32_to_cpu(pp[1]);
230 hdr->first_rec_offset = le32_to_cpu(pp[2]);
231 hdr->tbl_size = le32_to_cpu(pp[3]);
232 hdr->checksum = le32_to_cpu(pp[4]);
235 static int __write_table_header(struct amdgpu_ras_eeprom_control *control)
237 u8 buf[RAS_TABLE_HEADER_SIZE];
238 struct amdgpu_device *adev = to_amdgpu_device(control);
241 memset(buf, 0, sizeof(buf));
242 __encode_table_header_to_buf(&control->tbl_hdr, buf);
244 /* i2c may be unstable in gpu reset */
245 down_read(&adev->reset_domain->sem);
246 res = amdgpu_eeprom_write(adev->pm.ras_eeprom_i2c_bus,
247 control->i2c_address +
248 control->ras_header_offset,
249 buf, RAS_TABLE_HEADER_SIZE);
250 up_read(&adev->reset_domain->sem);
253 DRM_ERROR("Failed to write EEPROM table header:%d", res);
254 } else if (res < RAS_TABLE_HEADER_SIZE) {
255 DRM_ERROR("Short write:%d out of %d\n",
256 res, RAS_TABLE_HEADER_SIZE);
265 static u8 __calc_hdr_byte_sum(const struct amdgpu_ras_eeprom_control *control)
271 /* Header checksum, skip checksum field in the calculation */
272 sz = sizeof(control->tbl_hdr) - sizeof(control->tbl_hdr.checksum);
273 pp = (u8 *) &control->tbl_hdr;
275 for (ii = 0; ii < sz; ii++, pp++)
281 static int amdgpu_ras_eeprom_correct_header_tag(
282 struct amdgpu_ras_eeprom_control *control,
285 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
290 csum = -hdr->checksum;
292 hh = (void *) &hdr->header;
293 csum -= (hh[0] + hh[1] + hh[2] + hh[3]);
294 hh = (void *) &header;
295 csum += hh[0] + hh[1] + hh[2] + hh[3];
297 mutex_lock(&control->ras_tbl_mutex);
298 hdr->header = header;
299 hdr->checksum = csum;
300 res = __write_table_header(control);
301 mutex_unlock(&control->ras_tbl_mutex);
307 * amdgpu_ras_eeprom_reset_table -- Reset the RAS EEPROM table
308 * @control: pointer to control structure
310 * Reset the contents of the header of the RAS EEPROM table.
311 * Return 0 on success, -errno on error.
313 int amdgpu_ras_eeprom_reset_table(struct amdgpu_ras_eeprom_control *control)
315 struct amdgpu_device *adev = to_amdgpu_device(control);
316 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
317 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
321 mutex_lock(&control->ras_tbl_mutex);
323 hdr->header = RAS_TABLE_HDR_VAL;
324 hdr->version = RAS_TABLE_VER;
325 hdr->first_rec_offset = RAS_RECORD_START;
326 hdr->tbl_size = RAS_TABLE_HEADER_SIZE;
328 csum = __calc_hdr_byte_sum(control);
330 hdr->checksum = csum;
331 res = __write_table_header(control);
333 control->ras_num_recs = 0;
334 control->ras_fri = 0;
336 amdgpu_dpm_send_hbm_bad_pages_num(adev, control->ras_num_recs);
338 control->bad_channel_bitmap = 0;
339 amdgpu_dpm_send_hbm_bad_channel_flag(adev, control->bad_channel_bitmap);
340 con->update_channel_flag = false;
342 amdgpu_ras_debugfs_set_ret_size(control);
344 mutex_unlock(&control->ras_tbl_mutex);
350 __encode_table_record_to_buf(struct amdgpu_ras_eeprom_control *control,
351 struct eeprom_table_record *record,
357 /* Next are all record fields according to EEPROM page spec in LE foramt */
358 buf[i++] = record->err_type;
360 buf[i++] = record->bank;
362 tmp = cpu_to_le64(record->ts);
363 memcpy(buf + i, &tmp, 8);
366 tmp = cpu_to_le64((record->offset & 0xffffffffffff));
367 memcpy(buf + i, &tmp, 6);
370 buf[i++] = record->mem_channel;
371 buf[i++] = record->mcumc_id;
373 tmp = cpu_to_le64((record->retired_page & 0xffffffffffff));
374 memcpy(buf + i, &tmp, 6);
378 __decode_table_record_from_buf(struct amdgpu_ras_eeprom_control *control,
379 struct eeprom_table_record *record,
385 /* Next are all record fields according to EEPROM page spec in LE foramt */
386 record->err_type = buf[i++];
388 record->bank = buf[i++];
390 memcpy(&tmp, buf + i, 8);
391 record->ts = le64_to_cpu(tmp);
394 memcpy(&tmp, buf + i, 6);
395 record->offset = (le64_to_cpu(tmp) & 0xffffffffffff);
398 record->mem_channel = buf[i++];
399 record->mcumc_id = buf[i++];
401 memcpy(&tmp, buf + i, 6);
402 record->retired_page = (le64_to_cpu(tmp) & 0xffffffffffff);
405 bool amdgpu_ras_eeprom_check_err_threshold(struct amdgpu_device *adev)
407 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
409 if (!__is_ras_eeprom_supported(adev) ||
410 !amdgpu_bad_page_threshold)
413 /* skip check eeprom table for VEGA20 Gaming */
417 if (!(con->features & BIT(AMDGPU_RAS_BLOCK__UMC)))
420 if (con->eeprom_control.tbl_hdr.header == RAS_TABLE_HDR_BAD) {
421 if (amdgpu_bad_page_threshold == -1) {
422 dev_warn(adev->dev, "RAS records:%d exceed threshold:%d",
423 con->eeprom_control.ras_num_recs, con->bad_page_cnt_threshold);
425 "But GPU can be operated due to bad_page_threshold = -1.\n");
428 dev_warn(adev->dev, "This GPU is in BAD status.");
429 dev_warn(adev->dev, "Please retire it or set a larger "
430 "threshold value when reloading driver.\n");
439 * __amdgpu_ras_eeprom_write -- write indexed from buffer to EEPROM
440 * @control: pointer to control structure
441 * @buf: pointer to buffer containing data to write
442 * @fri: start writing at this index
443 * @num: number of records to write
445 * The caller must hold the table mutex in @control.
446 * Return 0 on success, -errno otherwise.
448 static int __amdgpu_ras_eeprom_write(struct amdgpu_ras_eeprom_control *control,
449 u8 *buf, const u32 fri, const u32 num)
451 struct amdgpu_device *adev = to_amdgpu_device(control);
455 /* i2c may be unstable in gpu reset */
456 down_read(&adev->reset_domain->sem);
457 buf_size = num * RAS_TABLE_RECORD_SIZE;
458 res = amdgpu_eeprom_write(adev->pm.ras_eeprom_i2c_bus,
459 control->i2c_address +
460 RAS_INDEX_TO_OFFSET(control, fri),
462 up_read(&adev->reset_domain->sem);
464 DRM_ERROR("Writing %d EEPROM table records error:%d",
466 } else if (res < buf_size) {
467 /* Short write, return error.
469 DRM_ERROR("Wrote %d records out of %d",
470 res / RAS_TABLE_RECORD_SIZE, num);
480 amdgpu_ras_eeprom_append_table(struct amdgpu_ras_eeprom_control *control,
481 struct eeprom_table_record *record,
484 struct amdgpu_ras *con = amdgpu_ras_get_context(to_amdgpu_device(control));
489 buf = kcalloc(num, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
493 /* Encode all of them in one go.
496 for (i = 0; i < num; i++, pp += RAS_TABLE_RECORD_SIZE) {
497 __encode_table_record_to_buf(control, &record[i], pp);
499 /* update bad channel bitmap */
500 if (!(control->bad_channel_bitmap & (1 << record[i].mem_channel))) {
501 control->bad_channel_bitmap |= 1 << record[i].mem_channel;
502 con->update_channel_flag = true;
506 /* a, first record index to write into.
507 * b, last record index to write into.
508 * a = first index to read (fri) + number of records in the table,
510 * Let N = control->ras_max_num_record_count, then we have,
511 * case 0: 0 <= a <= b < N,
512 * just append @num records starting at a;
513 * case 1: 0 <= a < N <= b,
514 * append (N - a) records starting at a, and
515 * append the remainder, b % N + 1, starting at 0.
516 * case 2: 0 <= fri < N <= a <= b, then modulo N we get two subcases,
517 * case 2a: 0 <= a <= b < N
518 * append num records starting at a; and fix fri if b overwrote it,
519 * and since a <= b, if b overwrote it then a must've also,
520 * and if b didn't overwrite it, then a didn't also.
521 * case 2b: 0 <= b < a < N
522 * write num records starting at a, which wraps around 0=N
523 * and overwrite fri unconditionally. Now from case 2a,
524 * this means that b eclipsed fri to overwrite it and wrap
525 * around 0 again, i.e. b = 2N+r pre modulo N, so we unconditionally
526 * set fri = b + 1 (mod N).
527 * Now, since fri is updated in every case, except the trivial case 0,
528 * the number of records present in the table after writing, is,
529 * num_recs - 1 = b - fri (mod N), and we take the positive value,
530 * by adding an arbitrary multiple of N before taking the modulo N
533 a = control->ras_fri + control->ras_num_recs;
535 if (b < control->ras_max_record_count) {
536 res = __amdgpu_ras_eeprom_write(control, buf, a, num);
537 } else if (a < control->ras_max_record_count) {
540 g0 = control->ras_max_record_count - a;
541 g1 = b % control->ras_max_record_count + 1;
542 res = __amdgpu_ras_eeprom_write(control, buf, a, g0);
545 res = __amdgpu_ras_eeprom_write(control,
546 buf + g0 * RAS_TABLE_RECORD_SIZE,
550 if (g1 > control->ras_fri)
551 control->ras_fri = g1 % control->ras_max_record_count;
553 a %= control->ras_max_record_count;
554 b %= control->ras_max_record_count;
557 /* Note that, b - a + 1 = num. */
558 res = __amdgpu_ras_eeprom_write(control, buf, a, num);
561 if (b >= control->ras_fri)
562 control->ras_fri = (b + 1) % control->ras_max_record_count;
566 /* b < a, which means, we write from
567 * a to the end of the table, and from
568 * the start of the table to b.
570 g0 = control->ras_max_record_count - a;
572 res = __amdgpu_ras_eeprom_write(control, buf, a, g0);
575 res = __amdgpu_ras_eeprom_write(control,
576 buf + g0 * RAS_TABLE_RECORD_SIZE,
580 control->ras_fri = g1 % control->ras_max_record_count;
583 control->ras_num_recs = 1 + (control->ras_max_record_count + b
585 % control->ras_max_record_count;
592 amdgpu_ras_eeprom_update_header(struct amdgpu_ras_eeprom_control *control)
594 struct amdgpu_device *adev = to_amdgpu_device(control);
595 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
600 /* Modify the header if it exceeds.
602 if (amdgpu_bad_page_threshold != 0 &&
603 control->ras_num_recs >= ras->bad_page_cnt_threshold) {
605 "Saved bad pages %d reaches threshold value %d\n",
606 control->ras_num_recs, ras->bad_page_cnt_threshold);
607 control->tbl_hdr.header = RAS_TABLE_HDR_BAD;
610 control->tbl_hdr.version = RAS_TABLE_VER;
611 control->tbl_hdr.first_rec_offset = RAS_INDEX_TO_OFFSET(control, control->ras_fri);
612 control->tbl_hdr.tbl_size = RAS_TABLE_HEADER_SIZE + control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
613 control->tbl_hdr.checksum = 0;
615 buf_size = control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
616 buf = kcalloc(control->ras_num_recs, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
618 DRM_ERROR("allocating memory for table of size %d bytes failed\n",
619 control->tbl_hdr.tbl_size);
624 down_read(&adev->reset_domain->sem);
625 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
626 control->i2c_address +
627 control->ras_record_offset,
629 up_read(&adev->reset_domain->sem);
631 DRM_ERROR("EEPROM failed reading records:%d\n",
634 } else if (res < buf_size) {
635 DRM_ERROR("EEPROM read %d out of %d bytes\n",
641 /* Recalc the checksum.
644 for (pp = buf; pp < buf + buf_size; pp++)
647 csum += __calc_hdr_byte_sum(control);
648 /* avoid sign extension when assigning to "checksum" */
650 control->tbl_hdr.checksum = csum;
651 res = __write_table_header(control);
658 * amdgpu_ras_eeprom_append -- append records to the EEPROM RAS table
659 * @control: pointer to control structure
660 * @record: array of records to append
661 * @num: number of records in @record array
663 * Append @num records to the table, calculate the checksum and write
664 * the table back to EEPROM. The maximum number of records that
665 * can be appended is between 1 and control->ras_max_record_count,
666 * regardless of how many records are already stored in the table.
668 * Return 0 on success or if EEPROM is not supported, -errno on error.
670 int amdgpu_ras_eeprom_append(struct amdgpu_ras_eeprom_control *control,
671 struct eeprom_table_record *record,
674 struct amdgpu_device *adev = to_amdgpu_device(control);
677 if (!__is_ras_eeprom_supported(adev))
681 DRM_ERROR("will not append 0 records\n");
683 } else if (num > control->ras_max_record_count) {
684 DRM_ERROR("cannot append %d records than the size of table %d\n",
685 num, control->ras_max_record_count);
689 mutex_lock(&control->ras_tbl_mutex);
691 res = amdgpu_ras_eeprom_append_table(control, record, num);
693 res = amdgpu_ras_eeprom_update_header(control);
695 amdgpu_ras_debugfs_set_ret_size(control);
697 mutex_unlock(&control->ras_tbl_mutex);
702 * __amdgpu_ras_eeprom_read -- read indexed from EEPROM into buffer
703 * @control: pointer to control structure
704 * @buf: pointer to buffer to read into
705 * @fri: first record index, start reading at this index, absolute index
706 * @num: number of records to read
708 * The caller must hold the table mutex in @control.
709 * Return 0 on success, -errno otherwise.
711 static int __amdgpu_ras_eeprom_read(struct amdgpu_ras_eeprom_control *control,
712 u8 *buf, const u32 fri, const u32 num)
714 struct amdgpu_device *adev = to_amdgpu_device(control);
718 /* i2c may be unstable in gpu reset */
719 down_read(&adev->reset_domain->sem);
720 buf_size = num * RAS_TABLE_RECORD_SIZE;
721 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
722 control->i2c_address +
723 RAS_INDEX_TO_OFFSET(control, fri),
725 up_read(&adev->reset_domain->sem);
727 DRM_ERROR("Reading %d EEPROM table records error:%d",
729 } else if (res < buf_size) {
730 /* Short read, return error.
732 DRM_ERROR("Read %d records out of %d",
733 res / RAS_TABLE_RECORD_SIZE, num);
743 * amdgpu_ras_eeprom_read -- read EEPROM
744 * @control: pointer to control structure
745 * @record: array of records to read into
746 * @num: number of records in @record
748 * Reads num records from the RAS table in EEPROM and
749 * writes the data into @record array.
751 * Returns 0 on success, -errno on error.
753 int amdgpu_ras_eeprom_read(struct amdgpu_ras_eeprom_control *control,
754 struct eeprom_table_record *record,
757 struct amdgpu_device *adev = to_amdgpu_device(control);
758 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
763 if (!__is_ras_eeprom_supported(adev))
767 DRM_ERROR("will not read 0 records\n");
769 } else if (num > control->ras_num_recs) {
770 DRM_ERROR("too many records to read:%d available:%d\n",
771 num, control->ras_num_recs);
775 buf = kcalloc(num, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
779 /* Determine how many records to read, from the first record
780 * index, fri, to the end of the table, and from the beginning
781 * of the table, such that the total number of records is
782 * @num, and we handle wrap around when fri > 0 and
783 * fri + num > RAS_MAX_RECORD_COUNT.
785 * First we compute the index of the last element
786 * which would be fetched from each region,
787 * g0 is in [fri, fri + num - 1], and
788 * g1 is in [0, RAS_MAX_RECORD_COUNT - 1].
789 * Then, if g0 < RAS_MAX_RECORD_COUNT, the index of
790 * the last element to fetch, we set g0 to _the number_
791 * of elements to fetch, @num, since we know that the last
792 * indexed to be fetched does not exceed the table.
794 * If, however, g0 >= RAS_MAX_RECORD_COUNT, then
795 * we set g0 to the number of elements to read
796 * until the end of the table, and g1 to the number of
797 * elements to read from the beginning of the table.
799 g0 = control->ras_fri + num - 1;
800 g1 = g0 % control->ras_max_record_count;
801 if (g0 < control->ras_max_record_count) {
805 g0 = control->ras_max_record_count - control->ras_fri;
809 mutex_lock(&control->ras_tbl_mutex);
810 res = __amdgpu_ras_eeprom_read(control, buf, control->ras_fri, g0);
814 res = __amdgpu_ras_eeprom_read(control,
815 buf + g0 * RAS_TABLE_RECORD_SIZE,
823 /* Read up everything? Then transform.
826 for (i = 0; i < num; i++, pp += RAS_TABLE_RECORD_SIZE) {
827 __decode_table_record_from_buf(control, &record[i], pp);
829 /* update bad channel bitmap */
830 if (!(control->bad_channel_bitmap & (1 << record[i].mem_channel))) {
831 control->bad_channel_bitmap |= 1 << record[i].mem_channel;
832 con->update_channel_flag = true;
837 mutex_unlock(&control->ras_tbl_mutex);
842 uint32_t amdgpu_ras_eeprom_max_record_count(void)
844 return RAS_MAX_RECORD_COUNT;
848 amdgpu_ras_debugfs_eeprom_size_read(struct file *f, char __user *buf,
849 size_t size, loff_t *pos)
851 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
852 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
853 struct amdgpu_ras_eeprom_control *control = ras ? &ras->eeprom_control : NULL;
860 if (!ras || !control) {
861 res = snprintf(data, sizeof(data), "Not supported\n");
863 res = snprintf(data, sizeof(data), "%d bytes or %d records\n",
864 RAS_TBL_SIZE_BYTES, control->ras_max_record_count);
871 res = min_t(size_t, res, size);
873 if (copy_to_user(buf, &data[*pos], res))
881 const struct file_operations amdgpu_ras_debugfs_eeprom_size_ops = {
882 .owner = THIS_MODULE,
883 .read = amdgpu_ras_debugfs_eeprom_size_read,
885 .llseek = default_llseek,
888 static const char *tbl_hdr_str = " Signature Version FirstOffs Size Checksum\n";
889 static const char *tbl_hdr_fmt = "0x%08X 0x%08X 0x%08X 0x%08X 0x%08X\n";
890 #define tbl_hdr_fmt_size (5 * (2+8) + 4 + 1)
891 static const char *rec_hdr_str = "Index Offset ErrType Bank/CU TimeStamp Offs/Addr MemChl MCUMCID RetiredPage\n";
892 static const char *rec_hdr_fmt = "%5d 0x%05X %7s 0x%02X 0x%016llX 0x%012llX 0x%02X 0x%02X 0x%012llX\n";
893 #define rec_hdr_fmt_size (5 + 1 + 7 + 1 + 7 + 1 + 7 + 1 + 18 + 1 + 14 + 1 + 6 + 1 + 7 + 1 + 14 + 1)
895 static const char *record_err_type_str[AMDGPU_RAS_EEPROM_ERR_COUNT] = {
901 static loff_t amdgpu_ras_debugfs_table_size(struct amdgpu_ras_eeprom_control *control)
903 return strlen(tbl_hdr_str) + tbl_hdr_fmt_size +
904 strlen(rec_hdr_str) + rec_hdr_fmt_size * control->ras_num_recs;
907 void amdgpu_ras_debugfs_set_ret_size(struct amdgpu_ras_eeprom_control *control)
909 struct amdgpu_ras *ras = container_of(control, struct amdgpu_ras,
911 struct dentry *de = ras->de_ras_eeprom_table;
914 d_inode(de)->i_size = amdgpu_ras_debugfs_table_size(control);
917 static ssize_t amdgpu_ras_debugfs_table_read(struct file *f, char __user *buf,
918 size_t size, loff_t *pos)
920 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
921 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
922 struct amdgpu_ras_eeprom_control *control = &ras->eeprom_control;
923 const size_t orig_size = size;
927 mutex_lock(&control->ras_tbl_mutex);
929 /* We want *pos - data_len > 0, which means there's
930 * bytes to be printed from data.
932 data_len = strlen(tbl_hdr_str);
933 if (*pos < data_len) {
935 data_len = min_t(size_t, data_len, size);
936 if (copy_to_user(buf, &tbl_hdr_str[*pos], data_len))
943 data_len = strlen(tbl_hdr_str) + tbl_hdr_fmt_size;
944 if (*pos < data_len && size > 0) {
945 u8 data[tbl_hdr_fmt_size + 1];
948 snprintf(data, sizeof(data), tbl_hdr_fmt,
949 control->tbl_hdr.header,
950 control->tbl_hdr.version,
951 control->tbl_hdr.first_rec_offset,
952 control->tbl_hdr.tbl_size,
953 control->tbl_hdr.checksum);
956 data_len = min_t(size_t, data_len, size);
957 lpos = *pos - strlen(tbl_hdr_str);
958 if (copy_to_user(buf, &data[lpos], data_len))
965 data_len = strlen(tbl_hdr_str) + tbl_hdr_fmt_size + strlen(rec_hdr_str);
966 if (*pos < data_len && size > 0) {
970 data_len = min_t(size_t, data_len, size);
971 lpos = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size;
972 if (copy_to_user(buf, &rec_hdr_str[lpos], data_len))
979 data_len = amdgpu_ras_debugfs_table_size(control);
980 if (*pos < data_len && size > 0) {
981 u8 dare[RAS_TABLE_RECORD_SIZE];
982 u8 data[rec_hdr_fmt_size + 1];
983 struct eeprom_table_record record;
986 /* Find the starting record index
988 s = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size -
990 s = s / rec_hdr_fmt_size;
991 r = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size -
993 r = r % rec_hdr_fmt_size;
995 for ( ; size > 0 && s < control->ras_num_recs; s++) {
996 u32 ai = RAS_RI_TO_AI(control, s);
997 /* Read a single record
999 res = __amdgpu_ras_eeprom_read(control, dare, ai, 1);
1002 __decode_table_record_from_buf(control, &record, dare);
1003 snprintf(data, sizeof(data), rec_hdr_fmt,
1005 RAS_INDEX_TO_OFFSET(control, ai),
1006 record_err_type_str[record.err_type],
1012 record.retired_page);
1014 data_len = min_t(size_t, rec_hdr_fmt_size - r, size);
1015 if (copy_to_user(buf, &data[r], data_len)) {
1027 mutex_unlock(&control->ras_tbl_mutex);
1028 return res < 0 ? res : orig_size - size;
1032 amdgpu_ras_debugfs_eeprom_table_read(struct file *f, char __user *buf,
1033 size_t size, loff_t *pos)
1035 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
1036 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1037 struct amdgpu_ras_eeprom_control *control = ras ? &ras->eeprom_control : NULL;
1044 if (!ras || !control) {
1045 res = snprintf(data, sizeof(data), "Not supported\n");
1050 res = min_t(size_t, res, size);
1052 if (copy_to_user(buf, &data[*pos], res))
1059 return amdgpu_ras_debugfs_table_read(f, buf, size, pos);
1063 const struct file_operations amdgpu_ras_debugfs_eeprom_table_ops = {
1064 .owner = THIS_MODULE,
1065 .read = amdgpu_ras_debugfs_eeprom_table_read,
1067 .llseek = default_llseek,
1071 * __verify_ras_table_checksum -- verify the RAS EEPROM table checksum
1072 * @control: pointer to control structure
1074 * Check the checksum of the stored in EEPROM RAS table.
1076 * Return 0 if the checksum is correct,
1077 * positive if it is not correct, and
1078 * -errno on I/O error.
1080 static int __verify_ras_table_checksum(struct amdgpu_ras_eeprom_control *control)
1082 struct amdgpu_device *adev = to_amdgpu_device(control);
1086 buf_size = RAS_TABLE_HEADER_SIZE +
1087 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
1088 buf = kzalloc(buf_size, GFP_KERNEL);
1090 DRM_ERROR("Out of memory checking RAS table checksum.\n");
1094 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
1095 control->i2c_address +
1096 control->ras_header_offset,
1098 if (res < buf_size) {
1099 DRM_ERROR("Partial read for checksum, res:%d\n", res);
1100 /* On partial reads, return -EIO.
1108 for (pp = buf; pp < buf + buf_size; pp++)
1112 return res < 0 ? res : csum;
1115 int amdgpu_ras_eeprom_init(struct amdgpu_ras_eeprom_control *control,
1116 bool *exceed_err_limit)
1118 struct amdgpu_device *adev = to_amdgpu_device(control);
1119 unsigned char buf[RAS_TABLE_HEADER_SIZE] = { 0 };
1120 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
1121 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1124 *exceed_err_limit = false;
1126 if (!__is_ras_eeprom_supported(adev))
1129 /* Verify i2c adapter is initialized */
1130 if (!adev->pm.ras_eeprom_i2c_bus || !adev->pm.ras_eeprom_i2c_bus->algo)
1133 if (!__get_eeprom_i2c_addr(adev, control))
1136 control->ras_header_offset = RAS_HDR_START;
1137 control->ras_record_offset = RAS_RECORD_START;
1138 control->ras_max_record_count = RAS_MAX_RECORD_COUNT;
1139 mutex_init(&control->ras_tbl_mutex);
1141 /* Read the table header from EEPROM address */
1142 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
1143 control->i2c_address + control->ras_header_offset,
1144 buf, RAS_TABLE_HEADER_SIZE);
1145 if (res < RAS_TABLE_HEADER_SIZE) {
1146 DRM_ERROR("Failed to read EEPROM table header, res:%d", res);
1147 return res >= 0 ? -EIO : res;
1150 __decode_table_header_from_buf(hdr, buf);
1152 control->ras_num_recs = RAS_NUM_RECS(hdr);
1153 control->ras_fri = RAS_OFFSET_TO_INDEX(control, hdr->first_rec_offset);
1155 if (hdr->header == RAS_TABLE_HDR_VAL) {
1156 DRM_DEBUG_DRIVER("Found existing EEPROM table with %d records",
1157 control->ras_num_recs);
1158 res = __verify_ras_table_checksum(control);
1160 DRM_ERROR("RAS table incorrect checksum or error:%d\n",
1163 /* Warn if we are at 90% of the threshold or above
1165 if (10 * control->ras_num_recs >= 9 * ras->bad_page_cnt_threshold)
1166 dev_warn(adev->dev, "RAS records:%u exceeds 90%% of threshold:%d",
1167 control->ras_num_recs,
1168 ras->bad_page_cnt_threshold);
1169 } else if (hdr->header == RAS_TABLE_HDR_BAD &&
1170 amdgpu_bad_page_threshold != 0) {
1171 res = __verify_ras_table_checksum(control);
1173 DRM_ERROR("RAS Table incorrect checksum or error:%d\n",
1175 if (ras->bad_page_cnt_threshold > control->ras_num_recs) {
1176 /* This means that, the threshold was increased since
1177 * the last time the system was booted, and now,
1178 * ras->bad_page_cnt_threshold - control->num_recs > 0,
1179 * so that at least one more record can be saved,
1180 * before the page count threshold is reached.
1183 "records:%d threshold:%d, resetting "
1184 "RAS table header signature",
1185 control->ras_num_recs,
1186 ras->bad_page_cnt_threshold);
1187 res = amdgpu_ras_eeprom_correct_header_tag(control,
1190 dev_err(adev->dev, "RAS records:%d exceed threshold:%d",
1191 control->ras_num_recs, ras->bad_page_cnt_threshold);
1192 if (amdgpu_bad_page_threshold == -1) {
1193 dev_warn(adev->dev, "GPU will be initialized due to bad_page_threshold = -1.");
1196 *exceed_err_limit = true;
1198 "RAS records:%d exceed threshold:%d, "
1199 "GPU will not be initialized. Replace this GPU or increase the threshold",
1200 control->ras_num_recs, ras->bad_page_cnt_threshold);
1204 DRM_INFO("Creating a new EEPROM table");
1206 res = amdgpu_ras_eeprom_reset_table(control);
1209 return res < 0 ? res : 0;