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
72 /* Bad GPU tag ‘BADG’ */
73 #define RAS_TABLE_HDR_BAD 0x42414447
76 * EEPROM Table structure v1
77 * ---------------------------------
79 * | EEPROM TABLE HEADER |
80 * | ( size 20 Bytes ) |
82 * ---------------------------------
84 * | BAD PAGE RECORD AREA |
86 * ---------------------------------
89 /* Assume 2-Mbit size EEPROM and take up the whole space. */
90 #define RAS_TBL_SIZE_BYTES (256 * 1024)
91 #define RAS_TABLE_START 0
92 #define RAS_HDR_START RAS_TABLE_START
93 #define RAS_RECORD_START (RAS_HDR_START + RAS_TABLE_HEADER_SIZE)
94 #define RAS_MAX_RECORD_COUNT ((RAS_TBL_SIZE_BYTES - RAS_TABLE_HEADER_SIZE) \
95 / RAS_TABLE_RECORD_SIZE)
98 * EEPROM Table structrue v2.1
99 * ---------------------------------
101 * | EEPROM TABLE HEADER |
102 * | ( size 20 Bytes ) |
104 * ---------------------------------
106 * | EEPROM TABLE RAS INFO |
107 * | (available info size 4 Bytes) |
108 * | ( reserved size 252 Bytes ) |
110 * ---------------------------------
112 * | BAD PAGE RECORD AREA |
114 * ---------------------------------
117 /* EEPROM Table V2_1 */
118 #define RAS_TABLE_V2_1_INFO_SIZE 256
119 #define RAS_TABLE_V2_1_INFO_START RAS_TABLE_HEADER_SIZE
120 #define RAS_RECORD_START_V2_1 (RAS_HDR_START + RAS_TABLE_HEADER_SIZE + \
121 RAS_TABLE_V2_1_INFO_SIZE)
122 #define RAS_MAX_RECORD_COUNT_V2_1 ((RAS_TBL_SIZE_BYTES - RAS_TABLE_HEADER_SIZE - \
123 RAS_TABLE_V2_1_INFO_SIZE) \
124 / RAS_TABLE_RECORD_SIZE)
126 /* Given a zero-based index of an EEPROM RAS record, yields the EEPROM
127 * offset off of RAS_TABLE_START. That is, this is something you can
128 * add to control->i2c_address, and then tell I2C layer to read
129 * from/write to there. _N is the so called absolute index,
130 * because it starts right after the table header.
132 #define RAS_INDEX_TO_OFFSET(_C, _N) ((_C)->ras_record_offset + \
133 (_N) * RAS_TABLE_RECORD_SIZE)
135 #define RAS_OFFSET_TO_INDEX(_C, _O) (((_O) - \
136 (_C)->ras_record_offset) / RAS_TABLE_RECORD_SIZE)
138 /* Given a 0-based relative record index, 0, 1, 2, ..., etc., off
139 * of "fri", return the absolute record index off of the end of
142 #define RAS_RI_TO_AI(_C, _I) (((_I) + (_C)->ras_fri) % \
143 (_C)->ras_max_record_count)
145 #define RAS_NUM_RECS(_tbl_hdr) (((_tbl_hdr)->tbl_size - \
146 RAS_TABLE_HEADER_SIZE) / RAS_TABLE_RECORD_SIZE)
148 #define RAS_NUM_RECS_V2_1(_tbl_hdr) (((_tbl_hdr)->tbl_size - \
149 RAS_TABLE_HEADER_SIZE - \
150 RAS_TABLE_V2_1_INFO_SIZE) / RAS_TABLE_RECORD_SIZE)
152 #define to_amdgpu_device(x) (container_of(x, struct amdgpu_ras, eeprom_control))->adev
154 static bool __is_ras_eeprom_supported(struct amdgpu_device *adev)
156 switch (adev->ip_versions[MP1_HWIP][0]) {
157 case IP_VERSION(11, 0, 2): /* VEGA20 and ARCTURUS */
158 case IP_VERSION(11, 0, 7): /* Sienna cichlid */
159 case IP_VERSION(13, 0, 0):
160 case IP_VERSION(13, 0, 2): /* Aldebaran */
161 case IP_VERSION(13, 0, 10):
168 static bool __get_eeprom_i2c_addr(struct amdgpu_device *adev,
169 struct amdgpu_ras_eeprom_control *control)
171 struct atom_context *atom_ctx = adev->mode_info.atom_context;
177 if (amdgpu_atomfirmware_ras_rom_addr(adev, &i2c_addr)) {
178 /* The address given by VBIOS is an 8-bit, wire-format
179 * address, i.e. the most significant byte.
181 * Normalize it to a 19-bit EEPROM address. Remove the
182 * device type identifier and make it a 7-bit address;
183 * then make it a 19-bit EEPROM address. See top of
186 i2c_addr = (i2c_addr & 0x0F) >> 1;
187 control->i2c_address = ((u32) i2c_addr) << 16;
192 switch (adev->ip_versions[MP1_HWIP][0]) {
193 case IP_VERSION(11, 0, 2):
194 /* VEGA20 and ARCTURUS */
195 if (adev->asic_type == CHIP_VEGA20)
196 control->i2c_address = EEPROM_I2C_MADDR_0;
197 else if (strnstr(atom_ctx->vbios_version,
199 sizeof(atom_ctx->vbios_version)))
200 control->i2c_address = EEPROM_I2C_MADDR_0;
202 control->i2c_address = EEPROM_I2C_MADDR_4;
204 case IP_VERSION(11, 0, 7):
205 control->i2c_address = EEPROM_I2C_MADDR_0;
207 case IP_VERSION(13, 0, 2):
208 if (strnstr(atom_ctx->vbios_version, "D673",
209 sizeof(atom_ctx->vbios_version)))
210 control->i2c_address = EEPROM_I2C_MADDR_4;
212 control->i2c_address = EEPROM_I2C_MADDR_0;
214 case IP_VERSION(13, 0, 0):
215 case IP_VERSION(13, 0, 10):
216 control->i2c_address = EEPROM_I2C_MADDR_4;
224 __encode_table_header_to_buf(struct amdgpu_ras_eeprom_table_header *hdr,
227 u32 *pp = (uint32_t *)buf;
229 pp[0] = cpu_to_le32(hdr->header);
230 pp[1] = cpu_to_le32(hdr->version);
231 pp[2] = cpu_to_le32(hdr->first_rec_offset);
232 pp[3] = cpu_to_le32(hdr->tbl_size);
233 pp[4] = cpu_to_le32(hdr->checksum);
237 __decode_table_header_from_buf(struct amdgpu_ras_eeprom_table_header *hdr,
240 u32 *pp = (uint32_t *)buf;
242 hdr->header = le32_to_cpu(pp[0]);
243 hdr->version = le32_to_cpu(pp[1]);
244 hdr->first_rec_offset = le32_to_cpu(pp[2]);
245 hdr->tbl_size = le32_to_cpu(pp[3]);
246 hdr->checksum = le32_to_cpu(pp[4]);
249 static int __write_table_header(struct amdgpu_ras_eeprom_control *control)
251 u8 buf[RAS_TABLE_HEADER_SIZE];
252 struct amdgpu_device *adev = to_amdgpu_device(control);
255 memset(buf, 0, sizeof(buf));
256 __encode_table_header_to_buf(&control->tbl_hdr, buf);
258 /* i2c may be unstable in gpu reset */
259 down_read(&adev->reset_domain->sem);
260 res = amdgpu_eeprom_write(adev->pm.ras_eeprom_i2c_bus,
261 control->i2c_address +
262 control->ras_header_offset,
263 buf, RAS_TABLE_HEADER_SIZE);
264 up_read(&adev->reset_domain->sem);
267 DRM_ERROR("Failed to write EEPROM table header:%d", res);
268 } else if (res < RAS_TABLE_HEADER_SIZE) {
269 DRM_ERROR("Short write:%d out of %d\n",
270 res, RAS_TABLE_HEADER_SIZE);
280 __encode_table_ras_info_to_buf(struct amdgpu_ras_eeprom_table_ras_info *rai,
283 u32 *pp = (uint32_t *)buf;
286 tmp = ((uint32_t)(rai->rma_status) & 0xFF) |
287 (((uint32_t)(rai->health_percent) << 8) & 0xFF00) |
288 (((uint32_t)(rai->ecc_page_threshold) << 16) & 0xFFFF0000);
289 pp[0] = cpu_to_le32(tmp);
293 __decode_table_ras_info_from_buf(struct amdgpu_ras_eeprom_table_ras_info *rai,
296 u32 *pp = (uint32_t *)buf;
299 tmp = le32_to_cpu(pp[0]);
300 rai->rma_status = tmp & 0xFF;
301 rai->health_percent = (tmp >> 8) & 0xFF;
302 rai->ecc_page_threshold = (tmp >> 16) & 0xFFFF;
305 static int __write_table_ras_info(struct amdgpu_ras_eeprom_control *control)
307 struct amdgpu_device *adev = to_amdgpu_device(control);
311 buf = kzalloc(RAS_TABLE_V2_1_INFO_SIZE, GFP_KERNEL);
313 DRM_ERROR("Failed to alloc buf to write table ras info\n");
317 __encode_table_ras_info_to_buf(&control->tbl_rai, buf);
319 /* i2c may be unstable in gpu reset */
320 down_read(&adev->reset_domain->sem);
321 res = amdgpu_eeprom_write(adev->pm.ras_eeprom_i2c_bus,
322 control->i2c_address +
323 control->ras_info_offset,
324 buf, RAS_TABLE_V2_1_INFO_SIZE);
325 up_read(&adev->reset_domain->sem);
328 DRM_ERROR("Failed to write EEPROM table ras info:%d", res);
329 } else if (res < RAS_TABLE_V2_1_INFO_SIZE) {
330 DRM_ERROR("Short write:%d out of %d\n",
331 res, RAS_TABLE_V2_1_INFO_SIZE);
342 static u8 __calc_hdr_byte_sum(const struct amdgpu_ras_eeprom_control *control)
348 /* Header checksum, skip checksum field in the calculation */
349 sz = sizeof(control->tbl_hdr) - sizeof(control->tbl_hdr.checksum);
350 pp = (u8 *) &control->tbl_hdr;
352 for (ii = 0; ii < sz; ii++, pp++)
358 static u8 __calc_ras_info_byte_sum(const struct amdgpu_ras_eeprom_control *control)
364 sz = sizeof(control->tbl_rai);
365 pp = (u8 *) &control->tbl_rai;
367 for (ii = 0; ii < sz; ii++, pp++)
373 static int amdgpu_ras_eeprom_correct_header_tag(
374 struct amdgpu_ras_eeprom_control *control,
377 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
382 csum = -hdr->checksum;
384 hh = (void *) &hdr->header;
385 csum -= (hh[0] + hh[1] + hh[2] + hh[3]);
386 hh = (void *) &header;
387 csum += hh[0] + hh[1] + hh[2] + hh[3];
389 mutex_lock(&control->ras_tbl_mutex);
390 hdr->header = header;
391 hdr->checksum = csum;
392 res = __write_table_header(control);
393 mutex_unlock(&control->ras_tbl_mutex);
399 * amdgpu_ras_eeprom_reset_table -- Reset the RAS EEPROM table
400 * @control: pointer to control structure
402 * Reset the contents of the header of the RAS EEPROM table.
403 * Return 0 on success, -errno on error.
405 int amdgpu_ras_eeprom_reset_table(struct amdgpu_ras_eeprom_control *control)
407 struct amdgpu_device *adev = to_amdgpu_device(control);
408 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
409 struct amdgpu_ras_eeprom_table_ras_info *rai = &control->tbl_rai;
410 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
414 mutex_lock(&control->ras_tbl_mutex);
416 hdr->header = RAS_TABLE_HDR_VAL;
418 adev->umc.ras->set_eeprom_table_version)
419 adev->umc.ras->set_eeprom_table_version(hdr);
421 hdr->version = RAS_TABLE_VER_V1;
423 if (hdr->version == RAS_TABLE_VER_V2_1) {
424 hdr->first_rec_offset = RAS_RECORD_START_V2_1;
425 hdr->tbl_size = RAS_TABLE_HEADER_SIZE +
426 RAS_TABLE_V2_1_INFO_SIZE;
427 rai->rma_status = GPU_HEALTH_USABLE;
429 * GPU health represented as a percentage.
430 * 0 means worst health, 100 means fully health.
432 rai->health_percent = 100;
433 /* ecc_page_threshold = 0 means disable bad page retirement */
434 rai->ecc_page_threshold = con->bad_page_cnt_threshold;
436 hdr->first_rec_offset = RAS_RECORD_START;
437 hdr->tbl_size = RAS_TABLE_HEADER_SIZE;
440 csum = __calc_hdr_byte_sum(control);
441 if (hdr->version == RAS_TABLE_VER_V2_1)
442 csum += __calc_ras_info_byte_sum(control);
444 hdr->checksum = csum;
445 res = __write_table_header(control);
446 if (!res && hdr->version > RAS_TABLE_VER_V1)
447 res = __write_table_ras_info(control);
449 control->ras_num_recs = 0;
450 control->ras_fri = 0;
452 amdgpu_dpm_send_hbm_bad_pages_num(adev, control->ras_num_recs);
454 control->bad_channel_bitmap = 0;
455 amdgpu_dpm_send_hbm_bad_channel_flag(adev, control->bad_channel_bitmap);
456 con->update_channel_flag = false;
458 amdgpu_ras_debugfs_set_ret_size(control);
460 mutex_unlock(&control->ras_tbl_mutex);
466 __encode_table_record_to_buf(struct amdgpu_ras_eeprom_control *control,
467 struct eeprom_table_record *record,
473 /* Next are all record fields according to EEPROM page spec in LE foramt */
474 buf[i++] = record->err_type;
476 buf[i++] = record->bank;
478 tmp = cpu_to_le64(record->ts);
479 memcpy(buf + i, &tmp, 8);
482 tmp = cpu_to_le64((record->offset & 0xffffffffffff));
483 memcpy(buf + i, &tmp, 6);
486 buf[i++] = record->mem_channel;
487 buf[i++] = record->mcumc_id;
489 tmp = cpu_to_le64((record->retired_page & 0xffffffffffff));
490 memcpy(buf + i, &tmp, 6);
494 __decode_table_record_from_buf(struct amdgpu_ras_eeprom_control *control,
495 struct eeprom_table_record *record,
501 /* Next are all record fields according to EEPROM page spec in LE foramt */
502 record->err_type = buf[i++];
504 record->bank = buf[i++];
506 memcpy(&tmp, buf + i, 8);
507 record->ts = le64_to_cpu(tmp);
510 memcpy(&tmp, buf + i, 6);
511 record->offset = (le64_to_cpu(tmp) & 0xffffffffffff);
514 record->mem_channel = buf[i++];
515 record->mcumc_id = buf[i++];
517 memcpy(&tmp, buf + i, 6);
518 record->retired_page = (le64_to_cpu(tmp) & 0xffffffffffff);
521 bool amdgpu_ras_eeprom_check_err_threshold(struct amdgpu_device *adev)
523 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
525 if (!__is_ras_eeprom_supported(adev) ||
526 !amdgpu_bad_page_threshold)
529 /* skip check eeprom table for VEGA20 Gaming */
533 if (!(con->features & BIT(AMDGPU_RAS_BLOCK__UMC)))
536 if (con->eeprom_control.tbl_hdr.header == RAS_TABLE_HDR_BAD) {
537 if (amdgpu_bad_page_threshold == -1) {
538 dev_warn(adev->dev, "RAS records:%d exceed threshold:%d",
539 con->eeprom_control.ras_num_recs, con->bad_page_cnt_threshold);
541 "But GPU can be operated due to bad_page_threshold = -1.\n");
544 dev_warn(adev->dev, "This GPU is in BAD status.");
545 dev_warn(adev->dev, "Please retire it or set a larger "
546 "threshold value when reloading driver.\n");
555 * __amdgpu_ras_eeprom_write -- write indexed from buffer to EEPROM
556 * @control: pointer to control structure
557 * @buf: pointer to buffer containing data to write
558 * @fri: start writing at this index
559 * @num: number of records to write
561 * The caller must hold the table mutex in @control.
562 * Return 0 on success, -errno otherwise.
564 static int __amdgpu_ras_eeprom_write(struct amdgpu_ras_eeprom_control *control,
565 u8 *buf, const u32 fri, const u32 num)
567 struct amdgpu_device *adev = to_amdgpu_device(control);
571 /* i2c may be unstable in gpu reset */
572 down_read(&adev->reset_domain->sem);
573 buf_size = num * RAS_TABLE_RECORD_SIZE;
574 res = amdgpu_eeprom_write(adev->pm.ras_eeprom_i2c_bus,
575 control->i2c_address +
576 RAS_INDEX_TO_OFFSET(control, fri),
578 up_read(&adev->reset_domain->sem);
580 DRM_ERROR("Writing %d EEPROM table records error:%d",
582 } else if (res < buf_size) {
583 /* Short write, return error.
585 DRM_ERROR("Wrote %d records out of %d",
586 res / RAS_TABLE_RECORD_SIZE, num);
596 amdgpu_ras_eeprom_append_table(struct amdgpu_ras_eeprom_control *control,
597 struct eeprom_table_record *record,
600 struct amdgpu_ras *con = amdgpu_ras_get_context(to_amdgpu_device(control));
605 buf = kcalloc(num, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
609 /* Encode all of them in one go.
612 for (i = 0; i < num; i++, pp += RAS_TABLE_RECORD_SIZE) {
613 __encode_table_record_to_buf(control, &record[i], pp);
615 /* update bad channel bitmap */
616 if (!(control->bad_channel_bitmap & (1 << record[i].mem_channel))) {
617 control->bad_channel_bitmap |= 1 << record[i].mem_channel;
618 con->update_channel_flag = true;
622 /* a, first record index to write into.
623 * b, last record index to write into.
624 * a = first index to read (fri) + number of records in the table,
626 * Let N = control->ras_max_num_record_count, then we have,
627 * case 0: 0 <= a <= b < N,
628 * just append @num records starting at a;
629 * case 1: 0 <= a < N <= b,
630 * append (N - a) records starting at a, and
631 * append the remainder, b % N + 1, starting at 0.
632 * case 2: 0 <= fri < N <= a <= b, then modulo N we get two subcases,
633 * case 2a: 0 <= a <= b < N
634 * append num records starting at a; and fix fri if b overwrote it,
635 * and since a <= b, if b overwrote it then a must've also,
636 * and if b didn't overwrite it, then a didn't also.
637 * case 2b: 0 <= b < a < N
638 * write num records starting at a, which wraps around 0=N
639 * and overwrite fri unconditionally. Now from case 2a,
640 * this means that b eclipsed fri to overwrite it and wrap
641 * around 0 again, i.e. b = 2N+r pre modulo N, so we unconditionally
642 * set fri = b + 1 (mod N).
643 * Now, since fri is updated in every case, except the trivial case 0,
644 * the number of records present in the table after writing, is,
645 * num_recs - 1 = b - fri (mod N), and we take the positive value,
646 * by adding an arbitrary multiple of N before taking the modulo N
649 a = control->ras_fri + control->ras_num_recs;
651 if (b < control->ras_max_record_count) {
652 res = __amdgpu_ras_eeprom_write(control, buf, a, num);
653 } else if (a < control->ras_max_record_count) {
656 g0 = control->ras_max_record_count - a;
657 g1 = b % control->ras_max_record_count + 1;
658 res = __amdgpu_ras_eeprom_write(control, buf, a, g0);
661 res = __amdgpu_ras_eeprom_write(control,
662 buf + g0 * RAS_TABLE_RECORD_SIZE,
666 if (g1 > control->ras_fri)
667 control->ras_fri = g1 % control->ras_max_record_count;
669 a %= control->ras_max_record_count;
670 b %= control->ras_max_record_count;
673 /* Note that, b - a + 1 = num. */
674 res = __amdgpu_ras_eeprom_write(control, buf, a, num);
677 if (b >= control->ras_fri)
678 control->ras_fri = (b + 1) % control->ras_max_record_count;
682 /* b < a, which means, we write from
683 * a to the end of the table, and from
684 * the start of the table to b.
686 g0 = control->ras_max_record_count - a;
688 res = __amdgpu_ras_eeprom_write(control, buf, a, g0);
691 res = __amdgpu_ras_eeprom_write(control,
692 buf + g0 * RAS_TABLE_RECORD_SIZE,
696 control->ras_fri = g1 % control->ras_max_record_count;
699 control->ras_num_recs = 1 + (control->ras_max_record_count + b
701 % control->ras_max_record_count;
708 amdgpu_ras_eeprom_update_header(struct amdgpu_ras_eeprom_control *control)
710 struct amdgpu_device *adev = to_amdgpu_device(control);
711 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
716 /* Modify the header if it exceeds.
718 if (amdgpu_bad_page_threshold != 0 &&
719 control->ras_num_recs >= ras->bad_page_cnt_threshold) {
721 "Saved bad pages %d reaches threshold value %d\n",
722 control->ras_num_recs, ras->bad_page_cnt_threshold);
723 control->tbl_hdr.header = RAS_TABLE_HDR_BAD;
724 if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1) {
725 control->tbl_rai.rma_status = GPU_RETIRED__ECC_REACH_THRESHOLD;
726 control->tbl_rai.health_percent = 0;
730 if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1)
731 control->tbl_hdr.tbl_size = RAS_TABLE_HEADER_SIZE +
732 RAS_TABLE_V2_1_INFO_SIZE +
733 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
735 control->tbl_hdr.tbl_size = RAS_TABLE_HEADER_SIZE +
736 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
737 control->tbl_hdr.checksum = 0;
739 buf_size = control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
740 buf = kcalloc(control->ras_num_recs, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
742 DRM_ERROR("allocating memory for table of size %d bytes failed\n",
743 control->tbl_hdr.tbl_size);
748 down_read(&adev->reset_domain->sem);
749 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
750 control->i2c_address +
751 control->ras_record_offset,
753 up_read(&adev->reset_domain->sem);
755 DRM_ERROR("EEPROM failed reading records:%d\n",
758 } else if (res < buf_size) {
759 DRM_ERROR("EEPROM read %d out of %d bytes\n",
766 * bad page records have been stored in eeprom,
767 * now calculate gpu health percent
769 if (amdgpu_bad_page_threshold != 0 &&
770 control->tbl_hdr.version == RAS_TABLE_VER_V2_1 &&
771 control->ras_num_recs < ras->bad_page_cnt_threshold)
772 control->tbl_rai.health_percent = ((ras->bad_page_cnt_threshold -
773 control->ras_num_recs) * 100) /
774 ras->bad_page_cnt_threshold;
776 /* Recalc the checksum.
779 for (pp = buf; pp < buf + buf_size; pp++)
782 csum += __calc_hdr_byte_sum(control);
783 if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1)
784 csum += __calc_ras_info_byte_sum(control);
785 /* avoid sign extension when assigning to "checksum" */
787 control->tbl_hdr.checksum = csum;
788 res = __write_table_header(control);
789 if (!res && control->tbl_hdr.version > RAS_TABLE_VER_V1)
790 res = __write_table_ras_info(control);
797 * amdgpu_ras_eeprom_append -- append records to the EEPROM RAS table
798 * @control: pointer to control structure
799 * @record: array of records to append
800 * @num: number of records in @record array
802 * Append @num records to the table, calculate the checksum and write
803 * the table back to EEPROM. The maximum number of records that
804 * can be appended is between 1 and control->ras_max_record_count,
805 * regardless of how many records are already stored in the table.
807 * Return 0 on success or if EEPROM is not supported, -errno on error.
809 int amdgpu_ras_eeprom_append(struct amdgpu_ras_eeprom_control *control,
810 struct eeprom_table_record *record,
813 struct amdgpu_device *adev = to_amdgpu_device(control);
816 if (!__is_ras_eeprom_supported(adev))
820 DRM_ERROR("will not append 0 records\n");
822 } else if (num > control->ras_max_record_count) {
823 DRM_ERROR("cannot append %d records than the size of table %d\n",
824 num, control->ras_max_record_count);
828 mutex_lock(&control->ras_tbl_mutex);
830 res = amdgpu_ras_eeprom_append_table(control, record, num);
832 res = amdgpu_ras_eeprom_update_header(control);
834 amdgpu_ras_debugfs_set_ret_size(control);
836 mutex_unlock(&control->ras_tbl_mutex);
841 * __amdgpu_ras_eeprom_read -- read indexed from EEPROM into buffer
842 * @control: pointer to control structure
843 * @buf: pointer to buffer to read into
844 * @fri: first record index, start reading at this index, absolute index
845 * @num: number of records to read
847 * The caller must hold the table mutex in @control.
848 * Return 0 on success, -errno otherwise.
850 static int __amdgpu_ras_eeprom_read(struct amdgpu_ras_eeprom_control *control,
851 u8 *buf, const u32 fri, const u32 num)
853 struct amdgpu_device *adev = to_amdgpu_device(control);
857 /* i2c may be unstable in gpu reset */
858 down_read(&adev->reset_domain->sem);
859 buf_size = num * RAS_TABLE_RECORD_SIZE;
860 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
861 control->i2c_address +
862 RAS_INDEX_TO_OFFSET(control, fri),
864 up_read(&adev->reset_domain->sem);
866 DRM_ERROR("Reading %d EEPROM table records error:%d",
868 } else if (res < buf_size) {
869 /* Short read, return error.
871 DRM_ERROR("Read %d records out of %d",
872 res / RAS_TABLE_RECORD_SIZE, num);
882 * amdgpu_ras_eeprom_read -- read EEPROM
883 * @control: pointer to control structure
884 * @record: array of records to read into
885 * @num: number of records in @record
887 * Reads num records from the RAS table in EEPROM and
888 * writes the data into @record array.
890 * Returns 0 on success, -errno on error.
892 int amdgpu_ras_eeprom_read(struct amdgpu_ras_eeprom_control *control,
893 struct eeprom_table_record *record,
896 struct amdgpu_device *adev = to_amdgpu_device(control);
897 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
902 if (!__is_ras_eeprom_supported(adev))
906 DRM_ERROR("will not read 0 records\n");
908 } else if (num > control->ras_num_recs) {
909 DRM_ERROR("too many records to read:%d available:%d\n",
910 num, control->ras_num_recs);
914 buf = kcalloc(num, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
918 /* Determine how many records to read, from the first record
919 * index, fri, to the end of the table, and from the beginning
920 * of the table, such that the total number of records is
921 * @num, and we handle wrap around when fri > 0 and
922 * fri + num > RAS_MAX_RECORD_COUNT.
924 * First we compute the index of the last element
925 * which would be fetched from each region,
926 * g0 is in [fri, fri + num - 1], and
927 * g1 is in [0, RAS_MAX_RECORD_COUNT - 1].
928 * Then, if g0 < RAS_MAX_RECORD_COUNT, the index of
929 * the last element to fetch, we set g0 to _the number_
930 * of elements to fetch, @num, since we know that the last
931 * indexed to be fetched does not exceed the table.
933 * If, however, g0 >= RAS_MAX_RECORD_COUNT, then
934 * we set g0 to the number of elements to read
935 * until the end of the table, and g1 to the number of
936 * elements to read from the beginning of the table.
938 g0 = control->ras_fri + num - 1;
939 g1 = g0 % control->ras_max_record_count;
940 if (g0 < control->ras_max_record_count) {
944 g0 = control->ras_max_record_count - control->ras_fri;
948 mutex_lock(&control->ras_tbl_mutex);
949 res = __amdgpu_ras_eeprom_read(control, buf, control->ras_fri, g0);
953 res = __amdgpu_ras_eeprom_read(control,
954 buf + g0 * RAS_TABLE_RECORD_SIZE,
962 /* Read up everything? Then transform.
965 for (i = 0; i < num; i++, pp += RAS_TABLE_RECORD_SIZE) {
966 __decode_table_record_from_buf(control, &record[i], pp);
968 /* update bad channel bitmap */
969 if (!(control->bad_channel_bitmap & (1 << record[i].mem_channel))) {
970 control->bad_channel_bitmap |= 1 << record[i].mem_channel;
971 con->update_channel_flag = true;
976 mutex_unlock(&control->ras_tbl_mutex);
981 uint32_t amdgpu_ras_eeprom_max_record_count(struct amdgpu_ras_eeprom_control *control)
983 if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1)
984 return RAS_MAX_RECORD_COUNT_V2_1;
986 return RAS_MAX_RECORD_COUNT;
990 amdgpu_ras_debugfs_eeprom_size_read(struct file *f, char __user *buf,
991 size_t size, loff_t *pos)
993 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
994 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
995 struct amdgpu_ras_eeprom_control *control = ras ? &ras->eeprom_control : NULL;
1002 if (!ras || !control) {
1003 res = snprintf(data, sizeof(data), "Not supported\n");
1005 res = snprintf(data, sizeof(data), "%d bytes or %d records\n",
1006 RAS_TBL_SIZE_BYTES, control->ras_max_record_count);
1013 res = min_t(size_t, res, size);
1015 if (copy_to_user(buf, &data[*pos], res))
1023 const struct file_operations amdgpu_ras_debugfs_eeprom_size_ops = {
1024 .owner = THIS_MODULE,
1025 .read = amdgpu_ras_debugfs_eeprom_size_read,
1027 .llseek = default_llseek,
1030 static const char *tbl_hdr_str = " Signature Version FirstOffs Size Checksum\n";
1031 static const char *tbl_hdr_fmt = "0x%08X 0x%08X 0x%08X 0x%08X 0x%08X\n";
1032 #define tbl_hdr_fmt_size (5 * (2+8) + 4 + 1)
1033 static const char *rec_hdr_str = "Index Offset ErrType Bank/CU TimeStamp Offs/Addr MemChl MCUMCID RetiredPage\n";
1034 static const char *rec_hdr_fmt = "%5d 0x%05X %7s 0x%02X 0x%016llX 0x%012llX 0x%02X 0x%02X 0x%012llX\n";
1035 #define rec_hdr_fmt_size (5 + 1 + 7 + 1 + 7 + 1 + 7 + 1 + 18 + 1 + 14 + 1 + 6 + 1 + 7 + 1 + 14 + 1)
1037 static const char *record_err_type_str[AMDGPU_RAS_EEPROM_ERR_COUNT] = {
1043 static loff_t amdgpu_ras_debugfs_table_size(struct amdgpu_ras_eeprom_control *control)
1045 return strlen(tbl_hdr_str) + tbl_hdr_fmt_size +
1046 strlen(rec_hdr_str) + rec_hdr_fmt_size * control->ras_num_recs;
1049 void amdgpu_ras_debugfs_set_ret_size(struct amdgpu_ras_eeprom_control *control)
1051 struct amdgpu_ras *ras = container_of(control, struct amdgpu_ras,
1053 struct dentry *de = ras->de_ras_eeprom_table;
1056 d_inode(de)->i_size = amdgpu_ras_debugfs_table_size(control);
1059 static ssize_t amdgpu_ras_debugfs_table_read(struct file *f, char __user *buf,
1060 size_t size, loff_t *pos)
1062 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
1063 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1064 struct amdgpu_ras_eeprom_control *control = &ras->eeprom_control;
1065 const size_t orig_size = size;
1069 mutex_lock(&control->ras_tbl_mutex);
1071 /* We want *pos - data_len > 0, which means there's
1072 * bytes to be printed from data.
1074 data_len = strlen(tbl_hdr_str);
1075 if (*pos < data_len) {
1077 data_len = min_t(size_t, data_len, size);
1078 if (copy_to_user(buf, &tbl_hdr_str[*pos], data_len))
1085 data_len = strlen(tbl_hdr_str) + tbl_hdr_fmt_size;
1086 if (*pos < data_len && size > 0) {
1087 u8 data[tbl_hdr_fmt_size + 1];
1090 snprintf(data, sizeof(data), tbl_hdr_fmt,
1091 control->tbl_hdr.header,
1092 control->tbl_hdr.version,
1093 control->tbl_hdr.first_rec_offset,
1094 control->tbl_hdr.tbl_size,
1095 control->tbl_hdr.checksum);
1098 data_len = min_t(size_t, data_len, size);
1099 lpos = *pos - strlen(tbl_hdr_str);
1100 if (copy_to_user(buf, &data[lpos], data_len))
1107 data_len = strlen(tbl_hdr_str) + tbl_hdr_fmt_size + strlen(rec_hdr_str);
1108 if (*pos < data_len && size > 0) {
1112 data_len = min_t(size_t, data_len, size);
1113 lpos = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size;
1114 if (copy_to_user(buf, &rec_hdr_str[lpos], data_len))
1121 data_len = amdgpu_ras_debugfs_table_size(control);
1122 if (*pos < data_len && size > 0) {
1123 u8 dare[RAS_TABLE_RECORD_SIZE];
1124 u8 data[rec_hdr_fmt_size + 1];
1125 struct eeprom_table_record record;
1128 /* Find the starting record index
1130 s = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size -
1131 strlen(rec_hdr_str);
1132 s = s / rec_hdr_fmt_size;
1133 r = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size -
1134 strlen(rec_hdr_str);
1135 r = r % rec_hdr_fmt_size;
1137 for ( ; size > 0 && s < control->ras_num_recs; s++) {
1138 u32 ai = RAS_RI_TO_AI(control, s);
1139 /* Read a single record
1141 res = __amdgpu_ras_eeprom_read(control, dare, ai, 1);
1144 __decode_table_record_from_buf(control, &record, dare);
1145 snprintf(data, sizeof(data), rec_hdr_fmt,
1147 RAS_INDEX_TO_OFFSET(control, ai),
1148 record_err_type_str[record.err_type],
1154 record.retired_page);
1156 data_len = min_t(size_t, rec_hdr_fmt_size - r, size);
1157 if (copy_to_user(buf, &data[r], data_len)) {
1169 mutex_unlock(&control->ras_tbl_mutex);
1170 return res < 0 ? res : orig_size - size;
1174 amdgpu_ras_debugfs_eeprom_table_read(struct file *f, char __user *buf,
1175 size_t size, loff_t *pos)
1177 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
1178 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1179 struct amdgpu_ras_eeprom_control *control = ras ? &ras->eeprom_control : NULL;
1186 if (!ras || !control) {
1187 res = snprintf(data, sizeof(data), "Not supported\n");
1192 res = min_t(size_t, res, size);
1194 if (copy_to_user(buf, &data[*pos], res))
1201 return amdgpu_ras_debugfs_table_read(f, buf, size, pos);
1205 const struct file_operations amdgpu_ras_debugfs_eeprom_table_ops = {
1206 .owner = THIS_MODULE,
1207 .read = amdgpu_ras_debugfs_eeprom_table_read,
1209 .llseek = default_llseek,
1213 * __verify_ras_table_checksum -- verify the RAS EEPROM table checksum
1214 * @control: pointer to control structure
1216 * Check the checksum of the stored in EEPROM RAS table.
1218 * Return 0 if the checksum is correct,
1219 * positive if it is not correct, and
1220 * -errno on I/O error.
1222 static int __verify_ras_table_checksum(struct amdgpu_ras_eeprom_control *control)
1224 struct amdgpu_device *adev = to_amdgpu_device(control);
1228 if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1)
1229 buf_size = RAS_TABLE_HEADER_SIZE +
1230 RAS_TABLE_V2_1_INFO_SIZE +
1231 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
1233 buf_size = RAS_TABLE_HEADER_SIZE +
1234 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
1236 buf = kzalloc(buf_size, GFP_KERNEL);
1238 DRM_ERROR("Out of memory checking RAS table checksum.\n");
1242 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
1243 control->i2c_address +
1244 control->ras_header_offset,
1246 if (res < buf_size) {
1247 DRM_ERROR("Partial read for checksum, res:%d\n", res);
1248 /* On partial reads, return -EIO.
1256 for (pp = buf; pp < buf + buf_size; pp++)
1260 return res < 0 ? res : csum;
1263 static int __read_table_ras_info(struct amdgpu_ras_eeprom_control *control)
1265 struct amdgpu_ras_eeprom_table_ras_info *rai = &control->tbl_rai;
1266 struct amdgpu_device *adev = to_amdgpu_device(control);
1270 buf = kzalloc(RAS_TABLE_V2_1_INFO_SIZE, GFP_KERNEL);
1272 DRM_ERROR("Failed to alloc buf to read EEPROM table ras info\n");
1277 * EEPROM table V2_1 supports ras info,
1278 * read EEPROM table ras info
1280 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
1281 control->i2c_address + control->ras_info_offset,
1282 buf, RAS_TABLE_V2_1_INFO_SIZE);
1283 if (res < RAS_TABLE_V2_1_INFO_SIZE) {
1284 DRM_ERROR("Failed to read EEPROM table ras info, res:%d", res);
1285 res = res >= 0 ? -EIO : res;
1289 __decode_table_ras_info_from_buf(rai, buf);
1293 return res == RAS_TABLE_V2_1_INFO_SIZE ? 0 : res;
1296 int amdgpu_ras_eeprom_init(struct amdgpu_ras_eeprom_control *control,
1297 bool *exceed_err_limit)
1299 struct amdgpu_device *adev = to_amdgpu_device(control);
1300 unsigned char buf[RAS_TABLE_HEADER_SIZE] = { 0 };
1301 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
1302 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1305 *exceed_err_limit = false;
1307 if (!__is_ras_eeprom_supported(adev))
1310 /* Verify i2c adapter is initialized */
1311 if (!adev->pm.ras_eeprom_i2c_bus || !adev->pm.ras_eeprom_i2c_bus->algo)
1314 if (!__get_eeprom_i2c_addr(adev, control))
1317 control->ras_header_offset = RAS_HDR_START;
1318 control->ras_info_offset = RAS_TABLE_V2_1_INFO_START;
1319 mutex_init(&control->ras_tbl_mutex);
1321 /* Read the table header from EEPROM address */
1322 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
1323 control->i2c_address + control->ras_header_offset,
1324 buf, RAS_TABLE_HEADER_SIZE);
1325 if (res < RAS_TABLE_HEADER_SIZE) {
1326 DRM_ERROR("Failed to read EEPROM table header, res:%d", res);
1327 return res >= 0 ? -EIO : res;
1330 __decode_table_header_from_buf(hdr, buf);
1332 if (hdr->version == RAS_TABLE_VER_V2_1) {
1333 control->ras_num_recs = RAS_NUM_RECS_V2_1(hdr);
1334 control->ras_record_offset = RAS_RECORD_START_V2_1;
1335 control->ras_max_record_count = RAS_MAX_RECORD_COUNT_V2_1;
1337 control->ras_num_recs = RAS_NUM_RECS(hdr);
1338 control->ras_record_offset = RAS_RECORD_START;
1339 control->ras_max_record_count = RAS_MAX_RECORD_COUNT;
1341 control->ras_fri = RAS_OFFSET_TO_INDEX(control, hdr->first_rec_offset);
1343 if (hdr->header == RAS_TABLE_HDR_VAL) {
1344 DRM_DEBUG_DRIVER("Found existing EEPROM table with %d records",
1345 control->ras_num_recs);
1347 if (hdr->version == RAS_TABLE_VER_V2_1) {
1348 res = __read_table_ras_info(control);
1353 res = __verify_ras_table_checksum(control);
1355 DRM_ERROR("RAS table incorrect checksum or error:%d\n",
1358 /* Warn if we are at 90% of the threshold or above
1360 if (10 * control->ras_num_recs >= 9 * ras->bad_page_cnt_threshold)
1361 dev_warn(adev->dev, "RAS records:%u exceeds 90%% of threshold:%d",
1362 control->ras_num_recs,
1363 ras->bad_page_cnt_threshold);
1364 } else if (hdr->header == RAS_TABLE_HDR_BAD &&
1365 amdgpu_bad_page_threshold != 0) {
1366 if (hdr->version == RAS_TABLE_VER_V2_1) {
1367 res = __read_table_ras_info(control);
1372 res = __verify_ras_table_checksum(control);
1374 DRM_ERROR("RAS Table incorrect checksum or error:%d\n",
1376 if (ras->bad_page_cnt_threshold > control->ras_num_recs) {
1377 /* This means that, the threshold was increased since
1378 * the last time the system was booted, and now,
1379 * ras->bad_page_cnt_threshold - control->num_recs > 0,
1380 * so that at least one more record can be saved,
1381 * before the page count threshold is reached.
1384 "records:%d threshold:%d, resetting "
1385 "RAS table header signature",
1386 control->ras_num_recs,
1387 ras->bad_page_cnt_threshold);
1388 res = amdgpu_ras_eeprom_correct_header_tag(control,
1391 dev_err(adev->dev, "RAS records:%d exceed threshold:%d",
1392 control->ras_num_recs, ras->bad_page_cnt_threshold);
1393 if (amdgpu_bad_page_threshold == -1) {
1394 dev_warn(adev->dev, "GPU will be initialized due to bad_page_threshold = -1.");
1397 *exceed_err_limit = true;
1399 "RAS records:%d exceed threshold:%d, "
1400 "GPU will not be initialized. Replace this GPU or increase the threshold",
1401 control->ras_num_recs, ras->bad_page_cnt_threshold);
1405 DRM_INFO("Creating a new EEPROM table");
1407 res = amdgpu_ras_eeprom_reset_table(control);
1410 return res < 0 ? res : 0;