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 (amdgpu_ip_version(adev, 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):
163 case IP_VERSION(13, 0, 6):
164 return (adev->gmc.is_app_apu) ? false : true;
170 static bool __get_eeprom_i2c_addr(struct amdgpu_device *adev,
171 struct amdgpu_ras_eeprom_control *control)
173 struct atom_context *atom_ctx = adev->mode_info.atom_context;
179 if (amdgpu_atomfirmware_ras_rom_addr(adev, &i2c_addr)) {
180 /* The address given by VBIOS is an 8-bit, wire-format
181 * address, i.e. the most significant byte.
183 * Normalize it to a 19-bit EEPROM address. Remove the
184 * device type identifier and make it a 7-bit address;
185 * then make it a 19-bit EEPROM address. See top of
188 i2c_addr = (i2c_addr & 0x0F) >> 1;
189 control->i2c_address = ((u32) i2c_addr) << 16;
194 switch (amdgpu_ip_version(adev, MP1_HWIP, 0)) {
195 case IP_VERSION(11, 0, 2):
196 /* VEGA20 and ARCTURUS */
197 if (adev->asic_type == CHIP_VEGA20)
198 control->i2c_address = EEPROM_I2C_MADDR_0;
199 else if (strnstr(atom_ctx->vbios_pn,
201 sizeof(atom_ctx->vbios_pn)))
202 control->i2c_address = EEPROM_I2C_MADDR_0;
204 control->i2c_address = EEPROM_I2C_MADDR_4;
206 case IP_VERSION(11, 0, 7):
207 control->i2c_address = EEPROM_I2C_MADDR_0;
209 case IP_VERSION(13, 0, 2):
210 if (strnstr(atom_ctx->vbios_pn, "D673",
211 sizeof(atom_ctx->vbios_pn)))
212 control->i2c_address = EEPROM_I2C_MADDR_4;
214 control->i2c_address = EEPROM_I2C_MADDR_0;
216 case IP_VERSION(13, 0, 0):
217 if (strnstr(atom_ctx->vbios_pn, "D707",
218 sizeof(atom_ctx->vbios_pn)))
219 control->i2c_address = EEPROM_I2C_MADDR_0;
221 control->i2c_address = EEPROM_I2C_MADDR_4;
223 case IP_VERSION(13, 0, 6):
224 case IP_VERSION(13, 0, 10):
225 control->i2c_address = EEPROM_I2C_MADDR_4;
233 __encode_table_header_to_buf(struct amdgpu_ras_eeprom_table_header *hdr,
236 u32 *pp = (uint32_t *)buf;
238 pp[0] = cpu_to_le32(hdr->header);
239 pp[1] = cpu_to_le32(hdr->version);
240 pp[2] = cpu_to_le32(hdr->first_rec_offset);
241 pp[3] = cpu_to_le32(hdr->tbl_size);
242 pp[4] = cpu_to_le32(hdr->checksum);
246 __decode_table_header_from_buf(struct amdgpu_ras_eeprom_table_header *hdr,
249 u32 *pp = (uint32_t *)buf;
251 hdr->header = le32_to_cpu(pp[0]);
252 hdr->version = le32_to_cpu(pp[1]);
253 hdr->first_rec_offset = le32_to_cpu(pp[2]);
254 hdr->tbl_size = le32_to_cpu(pp[3]);
255 hdr->checksum = le32_to_cpu(pp[4]);
258 static int __write_table_header(struct amdgpu_ras_eeprom_control *control)
260 u8 buf[RAS_TABLE_HEADER_SIZE];
261 struct amdgpu_device *adev = to_amdgpu_device(control);
264 memset(buf, 0, sizeof(buf));
265 __encode_table_header_to_buf(&control->tbl_hdr, buf);
267 /* i2c may be unstable in gpu reset */
268 down_read(&adev->reset_domain->sem);
269 res = amdgpu_eeprom_write(adev->pm.ras_eeprom_i2c_bus,
270 control->i2c_address +
271 control->ras_header_offset,
272 buf, RAS_TABLE_HEADER_SIZE);
273 up_read(&adev->reset_domain->sem);
276 DRM_ERROR("Failed to write EEPROM table header:%d", res);
277 } else if (res < RAS_TABLE_HEADER_SIZE) {
278 DRM_ERROR("Short write:%d out of %d\n",
279 res, RAS_TABLE_HEADER_SIZE);
289 __encode_table_ras_info_to_buf(struct amdgpu_ras_eeprom_table_ras_info *rai,
292 u32 *pp = (uint32_t *)buf;
295 tmp = ((uint32_t)(rai->rma_status) & 0xFF) |
296 (((uint32_t)(rai->health_percent) << 8) & 0xFF00) |
297 (((uint32_t)(rai->ecc_page_threshold) << 16) & 0xFFFF0000);
298 pp[0] = cpu_to_le32(tmp);
302 __decode_table_ras_info_from_buf(struct amdgpu_ras_eeprom_table_ras_info *rai,
305 u32 *pp = (uint32_t *)buf;
308 tmp = le32_to_cpu(pp[0]);
309 rai->rma_status = tmp & 0xFF;
310 rai->health_percent = (tmp >> 8) & 0xFF;
311 rai->ecc_page_threshold = (tmp >> 16) & 0xFFFF;
314 static int __write_table_ras_info(struct amdgpu_ras_eeprom_control *control)
316 struct amdgpu_device *adev = to_amdgpu_device(control);
320 buf = kzalloc(RAS_TABLE_V2_1_INFO_SIZE, GFP_KERNEL);
322 DRM_ERROR("Failed to alloc buf to write table ras info\n");
326 __encode_table_ras_info_to_buf(&control->tbl_rai, buf);
328 /* i2c may be unstable in gpu reset */
329 down_read(&adev->reset_domain->sem);
330 res = amdgpu_eeprom_write(adev->pm.ras_eeprom_i2c_bus,
331 control->i2c_address +
332 control->ras_info_offset,
333 buf, RAS_TABLE_V2_1_INFO_SIZE);
334 up_read(&adev->reset_domain->sem);
337 DRM_ERROR("Failed to write EEPROM table ras info:%d", res);
338 } else if (res < RAS_TABLE_V2_1_INFO_SIZE) {
339 DRM_ERROR("Short write:%d out of %d\n",
340 res, RAS_TABLE_V2_1_INFO_SIZE);
351 static u8 __calc_hdr_byte_sum(const struct amdgpu_ras_eeprom_control *control)
357 /* Header checksum, skip checksum field in the calculation */
358 sz = sizeof(control->tbl_hdr) - sizeof(control->tbl_hdr.checksum);
359 pp = (u8 *) &control->tbl_hdr;
361 for (ii = 0; ii < sz; ii++, pp++)
367 static u8 __calc_ras_info_byte_sum(const struct amdgpu_ras_eeprom_control *control)
373 sz = sizeof(control->tbl_rai);
374 pp = (u8 *) &control->tbl_rai;
376 for (ii = 0; ii < sz; ii++, pp++)
382 static int amdgpu_ras_eeprom_correct_header_tag(
383 struct amdgpu_ras_eeprom_control *control,
386 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
391 csum = -hdr->checksum;
393 hh = (void *) &hdr->header;
394 csum -= (hh[0] + hh[1] + hh[2] + hh[3]);
395 hh = (void *) &header;
396 csum += hh[0] + hh[1] + hh[2] + hh[3];
398 mutex_lock(&control->ras_tbl_mutex);
399 hdr->header = header;
400 hdr->checksum = csum;
401 res = __write_table_header(control);
402 mutex_unlock(&control->ras_tbl_mutex);
407 static void amdgpu_ras_set_eeprom_table_version(struct amdgpu_ras_eeprom_control *control)
409 struct amdgpu_device *adev = to_amdgpu_device(control);
410 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
412 switch (amdgpu_ip_version(adev, UMC_HWIP, 0)) {
413 case IP_VERSION(8, 10, 0):
414 case IP_VERSION(12, 0, 0):
415 hdr->version = RAS_TABLE_VER_V2_1;
418 hdr->version = RAS_TABLE_VER_V1;
424 * amdgpu_ras_eeprom_reset_table -- Reset the RAS EEPROM table
425 * @control: pointer to control structure
427 * Reset the contents of the header of the RAS EEPROM table.
428 * Return 0 on success, -errno on error.
430 int amdgpu_ras_eeprom_reset_table(struct amdgpu_ras_eeprom_control *control)
432 struct amdgpu_device *adev = to_amdgpu_device(control);
433 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
434 struct amdgpu_ras_eeprom_table_ras_info *rai = &control->tbl_rai;
435 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
439 mutex_lock(&control->ras_tbl_mutex);
441 hdr->header = RAS_TABLE_HDR_VAL;
442 amdgpu_ras_set_eeprom_table_version(control);
444 if (hdr->version == RAS_TABLE_VER_V2_1) {
445 hdr->first_rec_offset = RAS_RECORD_START_V2_1;
446 hdr->tbl_size = RAS_TABLE_HEADER_SIZE +
447 RAS_TABLE_V2_1_INFO_SIZE;
448 rai->rma_status = GPU_HEALTH_USABLE;
450 * GPU health represented as a percentage.
451 * 0 means worst health, 100 means fully health.
453 rai->health_percent = 100;
454 /* ecc_page_threshold = 0 means disable bad page retirement */
455 rai->ecc_page_threshold = con->bad_page_cnt_threshold;
457 hdr->first_rec_offset = RAS_RECORD_START;
458 hdr->tbl_size = RAS_TABLE_HEADER_SIZE;
461 csum = __calc_hdr_byte_sum(control);
462 if (hdr->version == RAS_TABLE_VER_V2_1)
463 csum += __calc_ras_info_byte_sum(control);
465 hdr->checksum = csum;
466 res = __write_table_header(control);
467 if (!res && hdr->version > RAS_TABLE_VER_V1)
468 res = __write_table_ras_info(control);
470 control->ras_num_recs = 0;
471 control->ras_fri = 0;
473 amdgpu_dpm_send_hbm_bad_pages_num(adev, control->ras_num_recs);
475 control->bad_channel_bitmap = 0;
476 amdgpu_dpm_send_hbm_bad_channel_flag(adev, control->bad_channel_bitmap);
477 con->update_channel_flag = false;
479 amdgpu_ras_debugfs_set_ret_size(control);
481 mutex_unlock(&control->ras_tbl_mutex);
487 __encode_table_record_to_buf(struct amdgpu_ras_eeprom_control *control,
488 struct eeprom_table_record *record,
494 /* Next are all record fields according to EEPROM page spec in LE foramt */
495 buf[i++] = record->err_type;
497 buf[i++] = record->bank;
499 tmp = cpu_to_le64(record->ts);
500 memcpy(buf + i, &tmp, 8);
503 tmp = cpu_to_le64((record->offset & 0xffffffffffff));
504 memcpy(buf + i, &tmp, 6);
507 buf[i++] = record->mem_channel;
508 buf[i++] = record->mcumc_id;
510 tmp = cpu_to_le64((record->retired_page & 0xffffffffffff));
511 memcpy(buf + i, &tmp, 6);
515 __decode_table_record_from_buf(struct amdgpu_ras_eeprom_control *control,
516 struct eeprom_table_record *record,
522 /* Next are all record fields according to EEPROM page spec in LE foramt */
523 record->err_type = buf[i++];
525 record->bank = buf[i++];
527 memcpy(&tmp, buf + i, 8);
528 record->ts = le64_to_cpu(tmp);
531 memcpy(&tmp, buf + i, 6);
532 record->offset = (le64_to_cpu(tmp) & 0xffffffffffff);
535 record->mem_channel = buf[i++];
536 record->mcumc_id = buf[i++];
538 memcpy(&tmp, buf + i, 6);
539 record->retired_page = (le64_to_cpu(tmp) & 0xffffffffffff);
542 bool amdgpu_ras_eeprom_check_err_threshold(struct amdgpu_device *adev)
544 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
546 if (!__is_ras_eeprom_supported(adev) ||
547 !amdgpu_bad_page_threshold)
550 /* skip check eeprom table for VEGA20 Gaming */
554 if (!(con->features & BIT(AMDGPU_RAS_BLOCK__UMC)))
557 if (con->eeprom_control.tbl_hdr.header == RAS_TABLE_HDR_BAD) {
558 if (amdgpu_bad_page_threshold == -1) {
559 dev_warn(adev->dev, "RAS records:%d exceed threshold:%d",
560 con->eeprom_control.ras_num_recs, con->bad_page_cnt_threshold);
562 "But GPU can be operated due to bad_page_threshold = -1.\n");
565 dev_warn(adev->dev, "This GPU is in BAD status.");
566 dev_warn(adev->dev, "Please retire it or set a larger "
567 "threshold value when reloading driver.\n");
576 * __amdgpu_ras_eeprom_write -- write indexed from buffer to EEPROM
577 * @control: pointer to control structure
578 * @buf: pointer to buffer containing data to write
579 * @fri: start writing at this index
580 * @num: number of records to write
582 * The caller must hold the table mutex in @control.
583 * Return 0 on success, -errno otherwise.
585 static int __amdgpu_ras_eeprom_write(struct amdgpu_ras_eeprom_control *control,
586 u8 *buf, const u32 fri, const u32 num)
588 struct amdgpu_device *adev = to_amdgpu_device(control);
592 /* i2c may be unstable in gpu reset */
593 down_read(&adev->reset_domain->sem);
594 buf_size = num * RAS_TABLE_RECORD_SIZE;
595 res = amdgpu_eeprom_write(adev->pm.ras_eeprom_i2c_bus,
596 control->i2c_address +
597 RAS_INDEX_TO_OFFSET(control, fri),
599 up_read(&adev->reset_domain->sem);
601 DRM_ERROR("Writing %d EEPROM table records error:%d",
603 } else if (res < buf_size) {
604 /* Short write, return error.
606 DRM_ERROR("Wrote %d records out of %d",
607 res / RAS_TABLE_RECORD_SIZE, num);
617 amdgpu_ras_eeprom_append_table(struct amdgpu_ras_eeprom_control *control,
618 struct eeprom_table_record *record,
621 struct amdgpu_ras *con = amdgpu_ras_get_context(to_amdgpu_device(control));
626 buf = kcalloc(num, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
630 /* Encode all of them in one go.
633 for (i = 0; i < num; i++, pp += RAS_TABLE_RECORD_SIZE) {
634 __encode_table_record_to_buf(control, &record[i], pp);
636 /* update bad channel bitmap */
637 if ((record[i].mem_channel < BITS_PER_TYPE(control->bad_channel_bitmap)) &&
638 !(control->bad_channel_bitmap & (1 << record[i].mem_channel))) {
639 control->bad_channel_bitmap |= 1 << record[i].mem_channel;
640 con->update_channel_flag = true;
644 /* a, first record index to write into.
645 * b, last record index to write into.
646 * a = first index to read (fri) + number of records in the table,
648 * Let N = control->ras_max_num_record_count, then we have,
649 * case 0: 0 <= a <= b < N,
650 * just append @num records starting at a;
651 * case 1: 0 <= a < N <= b,
652 * append (N - a) records starting at a, and
653 * append the remainder, b % N + 1, starting at 0.
654 * case 2: 0 <= fri < N <= a <= b, then modulo N we get two subcases,
655 * case 2a: 0 <= a <= b < N
656 * append num records starting at a; and fix fri if b overwrote it,
657 * and since a <= b, if b overwrote it then a must've also,
658 * and if b didn't overwrite it, then a didn't also.
659 * case 2b: 0 <= b < a < N
660 * write num records starting at a, which wraps around 0=N
661 * and overwrite fri unconditionally. Now from case 2a,
662 * this means that b eclipsed fri to overwrite it and wrap
663 * around 0 again, i.e. b = 2N+r pre modulo N, so we unconditionally
664 * set fri = b + 1 (mod N).
665 * Now, since fri is updated in every case, except the trivial case 0,
666 * the number of records present in the table after writing, is,
667 * num_recs - 1 = b - fri (mod N), and we take the positive value,
668 * by adding an arbitrary multiple of N before taking the modulo N
671 a = control->ras_fri + control->ras_num_recs;
673 if (b < control->ras_max_record_count) {
674 res = __amdgpu_ras_eeprom_write(control, buf, a, num);
675 } else if (a < control->ras_max_record_count) {
678 g0 = control->ras_max_record_count - a;
679 g1 = b % control->ras_max_record_count + 1;
680 res = __amdgpu_ras_eeprom_write(control, buf, a, g0);
683 res = __amdgpu_ras_eeprom_write(control,
684 buf + g0 * RAS_TABLE_RECORD_SIZE,
688 if (g1 > control->ras_fri)
689 control->ras_fri = g1 % control->ras_max_record_count;
691 a %= control->ras_max_record_count;
692 b %= control->ras_max_record_count;
695 /* Note that, b - a + 1 = num. */
696 res = __amdgpu_ras_eeprom_write(control, buf, a, num);
699 if (b >= control->ras_fri)
700 control->ras_fri = (b + 1) % control->ras_max_record_count;
704 /* b < a, which means, we write from
705 * a to the end of the table, and from
706 * the start of the table to b.
708 g0 = control->ras_max_record_count - a;
710 res = __amdgpu_ras_eeprom_write(control, buf, a, g0);
713 res = __amdgpu_ras_eeprom_write(control,
714 buf + g0 * RAS_TABLE_RECORD_SIZE,
718 control->ras_fri = g1 % control->ras_max_record_count;
721 control->ras_num_recs = 1 + (control->ras_max_record_count + b
723 % control->ras_max_record_count;
730 amdgpu_ras_eeprom_update_header(struct amdgpu_ras_eeprom_control *control)
732 struct amdgpu_device *adev = to_amdgpu_device(control);
733 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
738 /* Modify the header if it exceeds.
740 if (amdgpu_bad_page_threshold != 0 &&
741 control->ras_num_recs >= ras->bad_page_cnt_threshold) {
743 "Saved bad pages %d reaches threshold value %d\n",
744 control->ras_num_recs, ras->bad_page_cnt_threshold);
745 control->tbl_hdr.header = RAS_TABLE_HDR_BAD;
746 if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1) {
747 control->tbl_rai.rma_status = GPU_RETIRED__ECC_REACH_THRESHOLD;
748 control->tbl_rai.health_percent = 0;
751 /* ignore the -ENOTSUPP return value */
752 amdgpu_dpm_send_rma_reason(adev);
755 if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1)
756 control->tbl_hdr.tbl_size = RAS_TABLE_HEADER_SIZE +
757 RAS_TABLE_V2_1_INFO_SIZE +
758 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
760 control->tbl_hdr.tbl_size = RAS_TABLE_HEADER_SIZE +
761 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
762 control->tbl_hdr.checksum = 0;
764 buf_size = control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
765 buf = kcalloc(control->ras_num_recs, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
767 DRM_ERROR("allocating memory for table of size %d bytes failed\n",
768 control->tbl_hdr.tbl_size);
773 down_read(&adev->reset_domain->sem);
774 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
775 control->i2c_address +
776 control->ras_record_offset,
778 up_read(&adev->reset_domain->sem);
780 DRM_ERROR("EEPROM failed reading records:%d\n",
783 } else if (res < buf_size) {
784 DRM_ERROR("EEPROM read %d out of %d bytes\n",
791 * bad page records have been stored in eeprom,
792 * now calculate gpu health percent
794 if (amdgpu_bad_page_threshold != 0 &&
795 control->tbl_hdr.version == RAS_TABLE_VER_V2_1 &&
796 control->ras_num_recs < ras->bad_page_cnt_threshold)
797 control->tbl_rai.health_percent = ((ras->bad_page_cnt_threshold -
798 control->ras_num_recs) * 100) /
799 ras->bad_page_cnt_threshold;
801 /* Recalc the checksum.
804 for (pp = buf; pp < buf + buf_size; pp++)
807 csum += __calc_hdr_byte_sum(control);
808 if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1)
809 csum += __calc_ras_info_byte_sum(control);
810 /* avoid sign extension when assigning to "checksum" */
812 control->tbl_hdr.checksum = csum;
813 res = __write_table_header(control);
814 if (!res && control->tbl_hdr.version > RAS_TABLE_VER_V1)
815 res = __write_table_ras_info(control);
822 * amdgpu_ras_eeprom_append -- append records to the EEPROM RAS table
823 * @control: pointer to control structure
824 * @record: array of records to append
825 * @num: number of records in @record array
827 * Append @num records to the table, calculate the checksum and write
828 * the table back to EEPROM. The maximum number of records that
829 * can be appended is between 1 and control->ras_max_record_count,
830 * regardless of how many records are already stored in the table.
832 * Return 0 on success or if EEPROM is not supported, -errno on error.
834 int amdgpu_ras_eeprom_append(struct amdgpu_ras_eeprom_control *control,
835 struct eeprom_table_record *record,
838 struct amdgpu_device *adev = to_amdgpu_device(control);
841 if (!__is_ras_eeprom_supported(adev))
845 DRM_ERROR("will not append 0 records\n");
847 } else if (num > control->ras_max_record_count) {
848 DRM_ERROR("cannot append %d records than the size of table %d\n",
849 num, control->ras_max_record_count);
853 mutex_lock(&control->ras_tbl_mutex);
855 res = amdgpu_ras_eeprom_append_table(control, record, num);
857 res = amdgpu_ras_eeprom_update_header(control);
859 amdgpu_ras_debugfs_set_ret_size(control);
861 mutex_unlock(&control->ras_tbl_mutex);
866 * __amdgpu_ras_eeprom_read -- read indexed from EEPROM into buffer
867 * @control: pointer to control structure
868 * @buf: pointer to buffer to read into
869 * @fri: first record index, start reading at this index, absolute index
870 * @num: number of records to read
872 * The caller must hold the table mutex in @control.
873 * Return 0 on success, -errno otherwise.
875 static int __amdgpu_ras_eeprom_read(struct amdgpu_ras_eeprom_control *control,
876 u8 *buf, const u32 fri, const u32 num)
878 struct amdgpu_device *adev = to_amdgpu_device(control);
882 /* i2c may be unstable in gpu reset */
883 down_read(&adev->reset_domain->sem);
884 buf_size = num * RAS_TABLE_RECORD_SIZE;
885 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
886 control->i2c_address +
887 RAS_INDEX_TO_OFFSET(control, fri),
889 up_read(&adev->reset_domain->sem);
891 DRM_ERROR("Reading %d EEPROM table records error:%d",
893 } else if (res < buf_size) {
894 /* Short read, return error.
896 DRM_ERROR("Read %d records out of %d",
897 res / RAS_TABLE_RECORD_SIZE, num);
907 * amdgpu_ras_eeprom_read -- read EEPROM
908 * @control: pointer to control structure
909 * @record: array of records to read into
910 * @num: number of records in @record
912 * Reads num records from the RAS table in EEPROM and
913 * writes the data into @record array.
915 * Returns 0 on success, -errno on error.
917 int amdgpu_ras_eeprom_read(struct amdgpu_ras_eeprom_control *control,
918 struct eeprom_table_record *record,
921 struct amdgpu_device *adev = to_amdgpu_device(control);
922 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
927 if (!__is_ras_eeprom_supported(adev))
931 DRM_ERROR("will not read 0 records\n");
933 } else if (num > control->ras_num_recs) {
934 DRM_ERROR("too many records to read:%d available:%d\n",
935 num, control->ras_num_recs);
939 buf = kcalloc(num, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
943 /* Determine how many records to read, from the first record
944 * index, fri, to the end of the table, and from the beginning
945 * of the table, such that the total number of records is
946 * @num, and we handle wrap around when fri > 0 and
947 * fri + num > RAS_MAX_RECORD_COUNT.
949 * First we compute the index of the last element
950 * which would be fetched from each region,
951 * g0 is in [fri, fri + num - 1], and
952 * g1 is in [0, RAS_MAX_RECORD_COUNT - 1].
953 * Then, if g0 < RAS_MAX_RECORD_COUNT, the index of
954 * the last element to fetch, we set g0 to _the number_
955 * of elements to fetch, @num, since we know that the last
956 * indexed to be fetched does not exceed the table.
958 * If, however, g0 >= RAS_MAX_RECORD_COUNT, then
959 * we set g0 to the number of elements to read
960 * until the end of the table, and g1 to the number of
961 * elements to read from the beginning of the table.
963 g0 = control->ras_fri + num - 1;
964 g1 = g0 % control->ras_max_record_count;
965 if (g0 < control->ras_max_record_count) {
969 g0 = control->ras_max_record_count - control->ras_fri;
973 mutex_lock(&control->ras_tbl_mutex);
974 res = __amdgpu_ras_eeprom_read(control, buf, control->ras_fri, g0);
978 res = __amdgpu_ras_eeprom_read(control,
979 buf + g0 * RAS_TABLE_RECORD_SIZE,
987 /* Read up everything? Then transform.
990 for (i = 0; i < num; i++, pp += RAS_TABLE_RECORD_SIZE) {
991 __decode_table_record_from_buf(control, &record[i], pp);
993 /* update bad channel bitmap */
994 if ((record[i].mem_channel < BITS_PER_TYPE(control->bad_channel_bitmap)) &&
995 !(control->bad_channel_bitmap & (1 << record[i].mem_channel))) {
996 control->bad_channel_bitmap |= 1 << record[i].mem_channel;
997 con->update_channel_flag = true;
1002 mutex_unlock(&control->ras_tbl_mutex);
1007 uint32_t amdgpu_ras_eeprom_max_record_count(struct amdgpu_ras_eeprom_control *control)
1009 if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1)
1010 return RAS_MAX_RECORD_COUNT_V2_1;
1012 return RAS_MAX_RECORD_COUNT;
1016 amdgpu_ras_debugfs_eeprom_size_read(struct file *f, char __user *buf,
1017 size_t size, loff_t *pos)
1019 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
1020 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1021 struct amdgpu_ras_eeprom_control *control = ras ? &ras->eeprom_control : NULL;
1028 if (!ras || !control) {
1029 res = snprintf(data, sizeof(data), "Not supported\n");
1031 res = snprintf(data, sizeof(data), "%d bytes or %d records\n",
1032 RAS_TBL_SIZE_BYTES, control->ras_max_record_count);
1039 res = min_t(size_t, res, size);
1041 if (copy_to_user(buf, &data[*pos], res))
1049 const struct file_operations amdgpu_ras_debugfs_eeprom_size_ops = {
1050 .owner = THIS_MODULE,
1051 .read = amdgpu_ras_debugfs_eeprom_size_read,
1053 .llseek = default_llseek,
1056 static const char *tbl_hdr_str = " Signature Version FirstOffs Size Checksum\n";
1057 static const char *tbl_hdr_fmt = "0x%08X 0x%08X 0x%08X 0x%08X 0x%08X\n";
1058 #define tbl_hdr_fmt_size (5 * (2+8) + 4 + 1)
1059 static const char *rec_hdr_str = "Index Offset ErrType Bank/CU TimeStamp Offs/Addr MemChl MCUMCID RetiredPage\n";
1060 static const char *rec_hdr_fmt = "%5d 0x%05X %7s 0x%02X 0x%016llX 0x%012llX 0x%02X 0x%02X 0x%012llX\n";
1061 #define rec_hdr_fmt_size (5 + 1 + 7 + 1 + 7 + 1 + 7 + 1 + 18 + 1 + 14 + 1 + 6 + 1 + 7 + 1 + 14 + 1)
1063 static const char *record_err_type_str[AMDGPU_RAS_EEPROM_ERR_COUNT] = {
1069 static loff_t amdgpu_ras_debugfs_table_size(struct amdgpu_ras_eeprom_control *control)
1071 return strlen(tbl_hdr_str) + tbl_hdr_fmt_size +
1072 strlen(rec_hdr_str) + rec_hdr_fmt_size * control->ras_num_recs;
1075 void amdgpu_ras_debugfs_set_ret_size(struct amdgpu_ras_eeprom_control *control)
1077 struct amdgpu_ras *ras = container_of(control, struct amdgpu_ras,
1079 struct dentry *de = ras->de_ras_eeprom_table;
1082 d_inode(de)->i_size = amdgpu_ras_debugfs_table_size(control);
1085 static ssize_t amdgpu_ras_debugfs_table_read(struct file *f, char __user *buf,
1086 size_t size, loff_t *pos)
1088 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
1089 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1090 struct amdgpu_ras_eeprom_control *control = &ras->eeprom_control;
1091 const size_t orig_size = size;
1095 mutex_lock(&control->ras_tbl_mutex);
1097 /* We want *pos - data_len > 0, which means there's
1098 * bytes to be printed from data.
1100 data_len = strlen(tbl_hdr_str);
1101 if (*pos < data_len) {
1103 data_len = min_t(size_t, data_len, size);
1104 if (copy_to_user(buf, &tbl_hdr_str[*pos], data_len))
1111 data_len = strlen(tbl_hdr_str) + tbl_hdr_fmt_size;
1112 if (*pos < data_len && size > 0) {
1113 u8 data[tbl_hdr_fmt_size + 1];
1116 snprintf(data, sizeof(data), tbl_hdr_fmt,
1117 control->tbl_hdr.header,
1118 control->tbl_hdr.version,
1119 control->tbl_hdr.first_rec_offset,
1120 control->tbl_hdr.tbl_size,
1121 control->tbl_hdr.checksum);
1124 data_len = min_t(size_t, data_len, size);
1125 lpos = *pos - strlen(tbl_hdr_str);
1126 if (copy_to_user(buf, &data[lpos], data_len))
1133 data_len = strlen(tbl_hdr_str) + tbl_hdr_fmt_size + strlen(rec_hdr_str);
1134 if (*pos < data_len && size > 0) {
1138 data_len = min_t(size_t, data_len, size);
1139 lpos = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size;
1140 if (copy_to_user(buf, &rec_hdr_str[lpos], data_len))
1147 data_len = amdgpu_ras_debugfs_table_size(control);
1148 if (*pos < data_len && size > 0) {
1149 u8 dare[RAS_TABLE_RECORD_SIZE];
1150 u8 data[rec_hdr_fmt_size + 1];
1151 struct eeprom_table_record record;
1154 /* Find the starting record index
1156 s = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size -
1157 strlen(rec_hdr_str);
1158 s = s / rec_hdr_fmt_size;
1159 r = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size -
1160 strlen(rec_hdr_str);
1161 r = r % rec_hdr_fmt_size;
1163 for ( ; size > 0 && s < control->ras_num_recs; s++) {
1164 u32 ai = RAS_RI_TO_AI(control, s);
1165 /* Read a single record
1167 res = __amdgpu_ras_eeprom_read(control, dare, ai, 1);
1170 __decode_table_record_from_buf(control, &record, dare);
1171 snprintf(data, sizeof(data), rec_hdr_fmt,
1173 RAS_INDEX_TO_OFFSET(control, ai),
1174 record_err_type_str[record.err_type],
1180 record.retired_page);
1182 data_len = min_t(size_t, rec_hdr_fmt_size - r, size);
1183 if (copy_to_user(buf, &data[r], data_len)) {
1195 mutex_unlock(&control->ras_tbl_mutex);
1196 return res < 0 ? res : orig_size - size;
1200 amdgpu_ras_debugfs_eeprom_table_read(struct file *f, char __user *buf,
1201 size_t size, loff_t *pos)
1203 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
1204 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1205 struct amdgpu_ras_eeprom_control *control = ras ? &ras->eeprom_control : NULL;
1212 if (!ras || !control) {
1213 res = snprintf(data, sizeof(data), "Not supported\n");
1218 res = min_t(size_t, res, size);
1220 if (copy_to_user(buf, &data[*pos], res))
1227 return amdgpu_ras_debugfs_table_read(f, buf, size, pos);
1231 const struct file_operations amdgpu_ras_debugfs_eeprom_table_ops = {
1232 .owner = THIS_MODULE,
1233 .read = amdgpu_ras_debugfs_eeprom_table_read,
1235 .llseek = default_llseek,
1239 * __verify_ras_table_checksum -- verify the RAS EEPROM table checksum
1240 * @control: pointer to control structure
1242 * Check the checksum of the stored in EEPROM RAS table.
1244 * Return 0 if the checksum is correct,
1245 * positive if it is not correct, and
1246 * -errno on I/O error.
1248 static int __verify_ras_table_checksum(struct amdgpu_ras_eeprom_control *control)
1250 struct amdgpu_device *adev = to_amdgpu_device(control);
1254 if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1)
1255 buf_size = RAS_TABLE_HEADER_SIZE +
1256 RAS_TABLE_V2_1_INFO_SIZE +
1257 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
1259 buf_size = RAS_TABLE_HEADER_SIZE +
1260 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
1262 buf = kzalloc(buf_size, GFP_KERNEL);
1264 DRM_ERROR("Out of memory checking RAS table checksum.\n");
1268 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
1269 control->i2c_address +
1270 control->ras_header_offset,
1272 if (res < buf_size) {
1273 DRM_ERROR("Partial read for checksum, res:%d\n", res);
1274 /* On partial reads, return -EIO.
1282 for (pp = buf; pp < buf + buf_size; pp++)
1286 return res < 0 ? res : csum;
1289 static int __read_table_ras_info(struct amdgpu_ras_eeprom_control *control)
1291 struct amdgpu_ras_eeprom_table_ras_info *rai = &control->tbl_rai;
1292 struct amdgpu_device *adev = to_amdgpu_device(control);
1296 buf = kzalloc(RAS_TABLE_V2_1_INFO_SIZE, GFP_KERNEL);
1298 DRM_ERROR("Failed to alloc buf to read EEPROM table ras info\n");
1303 * EEPROM table V2_1 supports ras info,
1304 * read EEPROM table ras info
1306 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
1307 control->i2c_address + control->ras_info_offset,
1308 buf, RAS_TABLE_V2_1_INFO_SIZE);
1309 if (res < RAS_TABLE_V2_1_INFO_SIZE) {
1310 DRM_ERROR("Failed to read EEPROM table ras info, res:%d", res);
1311 res = res >= 0 ? -EIO : res;
1315 __decode_table_ras_info_from_buf(rai, buf);
1319 return res == RAS_TABLE_V2_1_INFO_SIZE ? 0 : res;
1322 int amdgpu_ras_eeprom_init(struct amdgpu_ras_eeprom_control *control,
1323 bool *exceed_err_limit)
1325 struct amdgpu_device *adev = to_amdgpu_device(control);
1326 unsigned char buf[RAS_TABLE_HEADER_SIZE] = { 0 };
1327 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
1328 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1331 *exceed_err_limit = false;
1333 if (!__is_ras_eeprom_supported(adev))
1336 /* Verify i2c adapter is initialized */
1337 if (!adev->pm.ras_eeprom_i2c_bus || !adev->pm.ras_eeprom_i2c_bus->algo)
1340 if (!__get_eeprom_i2c_addr(adev, control))
1343 control->ras_header_offset = RAS_HDR_START;
1344 control->ras_info_offset = RAS_TABLE_V2_1_INFO_START;
1345 mutex_init(&control->ras_tbl_mutex);
1347 /* Read the table header from EEPROM address */
1348 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
1349 control->i2c_address + control->ras_header_offset,
1350 buf, RAS_TABLE_HEADER_SIZE);
1351 if (res < RAS_TABLE_HEADER_SIZE) {
1352 DRM_ERROR("Failed to read EEPROM table header, res:%d", res);
1353 return res >= 0 ? -EIO : res;
1356 __decode_table_header_from_buf(hdr, buf);
1358 if (hdr->version == RAS_TABLE_VER_V2_1) {
1359 control->ras_num_recs = RAS_NUM_RECS_V2_1(hdr);
1360 control->ras_record_offset = RAS_RECORD_START_V2_1;
1361 control->ras_max_record_count = RAS_MAX_RECORD_COUNT_V2_1;
1363 control->ras_num_recs = RAS_NUM_RECS(hdr);
1364 control->ras_record_offset = RAS_RECORD_START;
1365 control->ras_max_record_count = RAS_MAX_RECORD_COUNT;
1367 control->ras_fri = RAS_OFFSET_TO_INDEX(control, hdr->first_rec_offset);
1369 if (hdr->header == RAS_TABLE_HDR_VAL) {
1370 DRM_DEBUG_DRIVER("Found existing EEPROM table with %d records",
1371 control->ras_num_recs);
1373 if (hdr->version == RAS_TABLE_VER_V2_1) {
1374 res = __read_table_ras_info(control);
1379 res = __verify_ras_table_checksum(control);
1381 DRM_ERROR("RAS table incorrect checksum or error:%d\n",
1384 /* Warn if we are at 90% of the threshold or above
1386 if (10 * control->ras_num_recs >= 9 * ras->bad_page_cnt_threshold)
1387 dev_warn(adev->dev, "RAS records:%u exceeds 90%% of threshold:%d",
1388 control->ras_num_recs,
1389 ras->bad_page_cnt_threshold);
1390 } else if (hdr->header == RAS_TABLE_HDR_BAD &&
1391 amdgpu_bad_page_threshold != 0) {
1392 if (hdr->version == RAS_TABLE_VER_V2_1) {
1393 res = __read_table_ras_info(control);
1398 res = __verify_ras_table_checksum(control);
1400 DRM_ERROR("RAS Table incorrect checksum or error:%d\n",
1402 if (ras->bad_page_cnt_threshold > control->ras_num_recs) {
1403 /* This means that, the threshold was increased since
1404 * the last time the system was booted, and now,
1405 * ras->bad_page_cnt_threshold - control->num_recs > 0,
1406 * so that at least one more record can be saved,
1407 * before the page count threshold is reached.
1410 "records:%d threshold:%d, resetting "
1411 "RAS table header signature",
1412 control->ras_num_recs,
1413 ras->bad_page_cnt_threshold);
1414 res = amdgpu_ras_eeprom_correct_header_tag(control,
1417 dev_err(adev->dev, "RAS records:%d exceed threshold:%d",
1418 control->ras_num_recs, ras->bad_page_cnt_threshold);
1419 if (amdgpu_bad_page_threshold == -1) {
1420 dev_warn(adev->dev, "GPU will be initialized due to bad_page_threshold = -1.");
1423 *exceed_err_limit = true;
1425 "RAS records:%d exceed threshold:%d, "
1426 "GPU will not be initialized. Replace this GPU or increase the threshold",
1427 control->ras_num_recs, ras->bad_page_cnt_threshold);
1431 DRM_INFO("Creating a new EEPROM table");
1433 res = amdgpu_ras_eeprom_reset_table(control);
1436 return res < 0 ? res : 0;