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);
408 * amdgpu_ras_eeprom_reset_table -- Reset the RAS EEPROM table
409 * @control: pointer to control structure
411 * Reset the contents of the header of the RAS EEPROM table.
412 * Return 0 on success, -errno on error.
414 int amdgpu_ras_eeprom_reset_table(struct amdgpu_ras_eeprom_control *control)
416 struct amdgpu_device *adev = to_amdgpu_device(control);
417 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
418 struct amdgpu_ras_eeprom_table_ras_info *rai = &control->tbl_rai;
419 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
423 mutex_lock(&control->ras_tbl_mutex);
425 hdr->header = RAS_TABLE_HDR_VAL;
427 adev->umc.ras->set_eeprom_table_version)
428 adev->umc.ras->set_eeprom_table_version(hdr);
430 hdr->version = RAS_TABLE_VER_V1;
432 if (hdr->version == RAS_TABLE_VER_V2_1) {
433 hdr->first_rec_offset = RAS_RECORD_START_V2_1;
434 hdr->tbl_size = RAS_TABLE_HEADER_SIZE +
435 RAS_TABLE_V2_1_INFO_SIZE;
436 rai->rma_status = GPU_HEALTH_USABLE;
438 * GPU health represented as a percentage.
439 * 0 means worst health, 100 means fully health.
441 rai->health_percent = 100;
442 /* ecc_page_threshold = 0 means disable bad page retirement */
443 rai->ecc_page_threshold = con->bad_page_cnt_threshold;
445 hdr->first_rec_offset = RAS_RECORD_START;
446 hdr->tbl_size = RAS_TABLE_HEADER_SIZE;
449 csum = __calc_hdr_byte_sum(control);
450 if (hdr->version == RAS_TABLE_VER_V2_1)
451 csum += __calc_ras_info_byte_sum(control);
453 hdr->checksum = csum;
454 res = __write_table_header(control);
455 if (!res && hdr->version > RAS_TABLE_VER_V1)
456 res = __write_table_ras_info(control);
458 control->ras_num_recs = 0;
459 control->ras_fri = 0;
461 amdgpu_dpm_send_hbm_bad_pages_num(adev, control->ras_num_recs);
463 control->bad_channel_bitmap = 0;
464 amdgpu_dpm_send_hbm_bad_channel_flag(adev, control->bad_channel_bitmap);
465 con->update_channel_flag = false;
467 amdgpu_ras_debugfs_set_ret_size(control);
469 mutex_unlock(&control->ras_tbl_mutex);
475 __encode_table_record_to_buf(struct amdgpu_ras_eeprom_control *control,
476 struct eeprom_table_record *record,
482 /* Next are all record fields according to EEPROM page spec in LE foramt */
483 buf[i++] = record->err_type;
485 buf[i++] = record->bank;
487 tmp = cpu_to_le64(record->ts);
488 memcpy(buf + i, &tmp, 8);
491 tmp = cpu_to_le64((record->offset & 0xffffffffffff));
492 memcpy(buf + i, &tmp, 6);
495 buf[i++] = record->mem_channel;
496 buf[i++] = record->mcumc_id;
498 tmp = cpu_to_le64((record->retired_page & 0xffffffffffff));
499 memcpy(buf + i, &tmp, 6);
503 __decode_table_record_from_buf(struct amdgpu_ras_eeprom_control *control,
504 struct eeprom_table_record *record,
510 /* Next are all record fields according to EEPROM page spec in LE foramt */
511 record->err_type = buf[i++];
513 record->bank = buf[i++];
515 memcpy(&tmp, buf + i, 8);
516 record->ts = le64_to_cpu(tmp);
519 memcpy(&tmp, buf + i, 6);
520 record->offset = (le64_to_cpu(tmp) & 0xffffffffffff);
523 record->mem_channel = buf[i++];
524 record->mcumc_id = buf[i++];
526 memcpy(&tmp, buf + i, 6);
527 record->retired_page = (le64_to_cpu(tmp) & 0xffffffffffff);
530 bool amdgpu_ras_eeprom_check_err_threshold(struct amdgpu_device *adev)
532 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
534 if (!__is_ras_eeprom_supported(adev) ||
535 !amdgpu_bad_page_threshold)
538 /* skip check eeprom table for VEGA20 Gaming */
542 if (!(con->features & BIT(AMDGPU_RAS_BLOCK__UMC)))
545 if (con->eeprom_control.tbl_hdr.header == RAS_TABLE_HDR_BAD) {
546 if (amdgpu_bad_page_threshold == -1) {
547 dev_warn(adev->dev, "RAS records:%d exceed threshold:%d",
548 con->eeprom_control.ras_num_recs, con->bad_page_cnt_threshold);
550 "But GPU can be operated due to bad_page_threshold = -1.\n");
553 dev_warn(adev->dev, "This GPU is in BAD status.");
554 dev_warn(adev->dev, "Please retire it or set a larger "
555 "threshold value when reloading driver.\n");
564 * __amdgpu_ras_eeprom_write -- write indexed from buffer to EEPROM
565 * @control: pointer to control structure
566 * @buf: pointer to buffer containing data to write
567 * @fri: start writing at this index
568 * @num: number of records to write
570 * The caller must hold the table mutex in @control.
571 * Return 0 on success, -errno otherwise.
573 static int __amdgpu_ras_eeprom_write(struct amdgpu_ras_eeprom_control *control,
574 u8 *buf, const u32 fri, const u32 num)
576 struct amdgpu_device *adev = to_amdgpu_device(control);
580 /* i2c may be unstable in gpu reset */
581 down_read(&adev->reset_domain->sem);
582 buf_size = num * RAS_TABLE_RECORD_SIZE;
583 res = amdgpu_eeprom_write(adev->pm.ras_eeprom_i2c_bus,
584 control->i2c_address +
585 RAS_INDEX_TO_OFFSET(control, fri),
587 up_read(&adev->reset_domain->sem);
589 DRM_ERROR("Writing %d EEPROM table records error:%d",
591 } else if (res < buf_size) {
592 /* Short write, return error.
594 DRM_ERROR("Wrote %d records out of %d",
595 res / RAS_TABLE_RECORD_SIZE, num);
605 amdgpu_ras_eeprom_append_table(struct amdgpu_ras_eeprom_control *control,
606 struct eeprom_table_record *record,
609 struct amdgpu_ras *con = amdgpu_ras_get_context(to_amdgpu_device(control));
614 buf = kcalloc(num, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
618 /* Encode all of them in one go.
621 for (i = 0; i < num; i++, pp += RAS_TABLE_RECORD_SIZE) {
622 __encode_table_record_to_buf(control, &record[i], pp);
624 /* update bad channel bitmap */
625 if ((record[i].mem_channel < BITS_PER_TYPE(control->bad_channel_bitmap)) &&
626 !(control->bad_channel_bitmap & (1 << record[i].mem_channel))) {
627 control->bad_channel_bitmap |= 1 << record[i].mem_channel;
628 con->update_channel_flag = true;
632 /* a, first record index to write into.
633 * b, last record index to write into.
634 * a = first index to read (fri) + number of records in the table,
636 * Let N = control->ras_max_num_record_count, then we have,
637 * case 0: 0 <= a <= b < N,
638 * just append @num records starting at a;
639 * case 1: 0 <= a < N <= b,
640 * append (N - a) records starting at a, and
641 * append the remainder, b % N + 1, starting at 0.
642 * case 2: 0 <= fri < N <= a <= b, then modulo N we get two subcases,
643 * case 2a: 0 <= a <= b < N
644 * append num records starting at a; and fix fri if b overwrote it,
645 * and since a <= b, if b overwrote it then a must've also,
646 * and if b didn't overwrite it, then a didn't also.
647 * case 2b: 0 <= b < a < N
648 * write num records starting at a, which wraps around 0=N
649 * and overwrite fri unconditionally. Now from case 2a,
650 * this means that b eclipsed fri to overwrite it and wrap
651 * around 0 again, i.e. b = 2N+r pre modulo N, so we unconditionally
652 * set fri = b + 1 (mod N).
653 * Now, since fri is updated in every case, except the trivial case 0,
654 * the number of records present in the table after writing, is,
655 * num_recs - 1 = b - fri (mod N), and we take the positive value,
656 * by adding an arbitrary multiple of N before taking the modulo N
659 a = control->ras_fri + control->ras_num_recs;
661 if (b < control->ras_max_record_count) {
662 res = __amdgpu_ras_eeprom_write(control, buf, a, num);
663 } else if (a < control->ras_max_record_count) {
666 g0 = control->ras_max_record_count - a;
667 g1 = b % control->ras_max_record_count + 1;
668 res = __amdgpu_ras_eeprom_write(control, buf, a, g0);
671 res = __amdgpu_ras_eeprom_write(control,
672 buf + g0 * RAS_TABLE_RECORD_SIZE,
676 if (g1 > control->ras_fri)
677 control->ras_fri = g1 % control->ras_max_record_count;
679 a %= control->ras_max_record_count;
680 b %= control->ras_max_record_count;
683 /* Note that, b - a + 1 = num. */
684 res = __amdgpu_ras_eeprom_write(control, buf, a, num);
687 if (b >= control->ras_fri)
688 control->ras_fri = (b + 1) % control->ras_max_record_count;
692 /* b < a, which means, we write from
693 * a to the end of the table, and from
694 * the start of the table to b.
696 g0 = control->ras_max_record_count - a;
698 res = __amdgpu_ras_eeprom_write(control, buf, a, g0);
701 res = __amdgpu_ras_eeprom_write(control,
702 buf + g0 * RAS_TABLE_RECORD_SIZE,
706 control->ras_fri = g1 % control->ras_max_record_count;
709 control->ras_num_recs = 1 + (control->ras_max_record_count + b
711 % control->ras_max_record_count;
718 amdgpu_ras_eeprom_update_header(struct amdgpu_ras_eeprom_control *control)
720 struct amdgpu_device *adev = to_amdgpu_device(control);
721 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
726 /* Modify the header if it exceeds.
728 if (amdgpu_bad_page_threshold != 0 &&
729 control->ras_num_recs >= ras->bad_page_cnt_threshold) {
731 "Saved bad pages %d reaches threshold value %d\n",
732 control->ras_num_recs, ras->bad_page_cnt_threshold);
733 control->tbl_hdr.header = RAS_TABLE_HDR_BAD;
734 if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1) {
735 control->tbl_rai.rma_status = GPU_RETIRED__ECC_REACH_THRESHOLD;
736 control->tbl_rai.health_percent = 0;
739 /* ignore the -ENOTSUPP return value */
740 amdgpu_dpm_send_rma_reason(adev);
743 if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1)
744 control->tbl_hdr.tbl_size = RAS_TABLE_HEADER_SIZE +
745 RAS_TABLE_V2_1_INFO_SIZE +
746 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
748 control->tbl_hdr.tbl_size = RAS_TABLE_HEADER_SIZE +
749 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
750 control->tbl_hdr.checksum = 0;
752 buf_size = control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
753 buf = kcalloc(control->ras_num_recs, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
755 DRM_ERROR("allocating memory for table of size %d bytes failed\n",
756 control->tbl_hdr.tbl_size);
761 down_read(&adev->reset_domain->sem);
762 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
763 control->i2c_address +
764 control->ras_record_offset,
766 up_read(&adev->reset_domain->sem);
768 DRM_ERROR("EEPROM failed reading records:%d\n",
771 } else if (res < buf_size) {
772 DRM_ERROR("EEPROM read %d out of %d bytes\n",
779 * bad page records have been stored in eeprom,
780 * now calculate gpu health percent
782 if (amdgpu_bad_page_threshold != 0 &&
783 control->tbl_hdr.version == RAS_TABLE_VER_V2_1 &&
784 control->ras_num_recs < ras->bad_page_cnt_threshold)
785 control->tbl_rai.health_percent = ((ras->bad_page_cnt_threshold -
786 control->ras_num_recs) * 100) /
787 ras->bad_page_cnt_threshold;
789 /* Recalc the checksum.
792 for (pp = buf; pp < buf + buf_size; pp++)
795 csum += __calc_hdr_byte_sum(control);
796 if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1)
797 csum += __calc_ras_info_byte_sum(control);
798 /* avoid sign extension when assigning to "checksum" */
800 control->tbl_hdr.checksum = csum;
801 res = __write_table_header(control);
802 if (!res && control->tbl_hdr.version > RAS_TABLE_VER_V1)
803 res = __write_table_ras_info(control);
810 * amdgpu_ras_eeprom_append -- append records to the EEPROM RAS table
811 * @control: pointer to control structure
812 * @record: array of records to append
813 * @num: number of records in @record array
815 * Append @num records to the table, calculate the checksum and write
816 * the table back to EEPROM. The maximum number of records that
817 * can be appended is between 1 and control->ras_max_record_count,
818 * regardless of how many records are already stored in the table.
820 * Return 0 on success or if EEPROM is not supported, -errno on error.
822 int amdgpu_ras_eeprom_append(struct amdgpu_ras_eeprom_control *control,
823 struct eeprom_table_record *record,
826 struct amdgpu_device *adev = to_amdgpu_device(control);
829 if (!__is_ras_eeprom_supported(adev))
833 DRM_ERROR("will not append 0 records\n");
835 } else if (num > control->ras_max_record_count) {
836 DRM_ERROR("cannot append %d records than the size of table %d\n",
837 num, control->ras_max_record_count);
841 mutex_lock(&control->ras_tbl_mutex);
843 res = amdgpu_ras_eeprom_append_table(control, record, num);
845 res = amdgpu_ras_eeprom_update_header(control);
847 amdgpu_ras_debugfs_set_ret_size(control);
849 mutex_unlock(&control->ras_tbl_mutex);
854 * __amdgpu_ras_eeprom_read -- read indexed from EEPROM into buffer
855 * @control: pointer to control structure
856 * @buf: pointer to buffer to read into
857 * @fri: first record index, start reading at this index, absolute index
858 * @num: number of records to read
860 * The caller must hold the table mutex in @control.
861 * Return 0 on success, -errno otherwise.
863 static int __amdgpu_ras_eeprom_read(struct amdgpu_ras_eeprom_control *control,
864 u8 *buf, const u32 fri, const u32 num)
866 struct amdgpu_device *adev = to_amdgpu_device(control);
870 /* i2c may be unstable in gpu reset */
871 down_read(&adev->reset_domain->sem);
872 buf_size = num * RAS_TABLE_RECORD_SIZE;
873 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
874 control->i2c_address +
875 RAS_INDEX_TO_OFFSET(control, fri),
877 up_read(&adev->reset_domain->sem);
879 DRM_ERROR("Reading %d EEPROM table records error:%d",
881 } else if (res < buf_size) {
882 /* Short read, return error.
884 DRM_ERROR("Read %d records out of %d",
885 res / RAS_TABLE_RECORD_SIZE, num);
895 * amdgpu_ras_eeprom_read -- read EEPROM
896 * @control: pointer to control structure
897 * @record: array of records to read into
898 * @num: number of records in @record
900 * Reads num records from the RAS table in EEPROM and
901 * writes the data into @record array.
903 * Returns 0 on success, -errno on error.
905 int amdgpu_ras_eeprom_read(struct amdgpu_ras_eeprom_control *control,
906 struct eeprom_table_record *record,
909 struct amdgpu_device *adev = to_amdgpu_device(control);
910 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
915 if (!__is_ras_eeprom_supported(adev))
919 DRM_ERROR("will not read 0 records\n");
921 } else if (num > control->ras_num_recs) {
922 DRM_ERROR("too many records to read:%d available:%d\n",
923 num, control->ras_num_recs);
927 buf = kcalloc(num, RAS_TABLE_RECORD_SIZE, GFP_KERNEL);
931 /* Determine how many records to read, from the first record
932 * index, fri, to the end of the table, and from the beginning
933 * of the table, such that the total number of records is
934 * @num, and we handle wrap around when fri > 0 and
935 * fri + num > RAS_MAX_RECORD_COUNT.
937 * First we compute the index of the last element
938 * which would be fetched from each region,
939 * g0 is in [fri, fri + num - 1], and
940 * g1 is in [0, RAS_MAX_RECORD_COUNT - 1].
941 * Then, if g0 < RAS_MAX_RECORD_COUNT, the index of
942 * the last element to fetch, we set g0 to _the number_
943 * of elements to fetch, @num, since we know that the last
944 * indexed to be fetched does not exceed the table.
946 * If, however, g0 >= RAS_MAX_RECORD_COUNT, then
947 * we set g0 to the number of elements to read
948 * until the end of the table, and g1 to the number of
949 * elements to read from the beginning of the table.
951 g0 = control->ras_fri + num - 1;
952 g1 = g0 % control->ras_max_record_count;
953 if (g0 < control->ras_max_record_count) {
957 g0 = control->ras_max_record_count - control->ras_fri;
961 mutex_lock(&control->ras_tbl_mutex);
962 res = __amdgpu_ras_eeprom_read(control, buf, control->ras_fri, g0);
966 res = __amdgpu_ras_eeprom_read(control,
967 buf + g0 * RAS_TABLE_RECORD_SIZE,
975 /* Read up everything? Then transform.
978 for (i = 0; i < num; i++, pp += RAS_TABLE_RECORD_SIZE) {
979 __decode_table_record_from_buf(control, &record[i], pp);
981 /* update bad channel bitmap */
982 if ((record[i].mem_channel < BITS_PER_TYPE(control->bad_channel_bitmap)) &&
983 !(control->bad_channel_bitmap & (1 << record[i].mem_channel))) {
984 control->bad_channel_bitmap |= 1 << record[i].mem_channel;
985 con->update_channel_flag = true;
990 mutex_unlock(&control->ras_tbl_mutex);
995 uint32_t amdgpu_ras_eeprom_max_record_count(struct amdgpu_ras_eeprom_control *control)
997 if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1)
998 return RAS_MAX_RECORD_COUNT_V2_1;
1000 return RAS_MAX_RECORD_COUNT;
1004 amdgpu_ras_debugfs_eeprom_size_read(struct file *f, char __user *buf,
1005 size_t size, loff_t *pos)
1007 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
1008 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1009 struct amdgpu_ras_eeprom_control *control = ras ? &ras->eeprom_control : NULL;
1016 if (!ras || !control) {
1017 res = snprintf(data, sizeof(data), "Not supported\n");
1019 res = snprintf(data, sizeof(data), "%d bytes or %d records\n",
1020 RAS_TBL_SIZE_BYTES, control->ras_max_record_count);
1027 res = min_t(size_t, res, size);
1029 if (copy_to_user(buf, &data[*pos], res))
1037 const struct file_operations amdgpu_ras_debugfs_eeprom_size_ops = {
1038 .owner = THIS_MODULE,
1039 .read = amdgpu_ras_debugfs_eeprom_size_read,
1041 .llseek = default_llseek,
1044 static const char *tbl_hdr_str = " Signature Version FirstOffs Size Checksum\n";
1045 static const char *tbl_hdr_fmt = "0x%08X 0x%08X 0x%08X 0x%08X 0x%08X\n";
1046 #define tbl_hdr_fmt_size (5 * (2+8) + 4 + 1)
1047 static const char *rec_hdr_str = "Index Offset ErrType Bank/CU TimeStamp Offs/Addr MemChl MCUMCID RetiredPage\n";
1048 static const char *rec_hdr_fmt = "%5d 0x%05X %7s 0x%02X 0x%016llX 0x%012llX 0x%02X 0x%02X 0x%012llX\n";
1049 #define rec_hdr_fmt_size (5 + 1 + 7 + 1 + 7 + 1 + 7 + 1 + 18 + 1 + 14 + 1 + 6 + 1 + 7 + 1 + 14 + 1)
1051 static const char *record_err_type_str[AMDGPU_RAS_EEPROM_ERR_COUNT] = {
1057 static loff_t amdgpu_ras_debugfs_table_size(struct amdgpu_ras_eeprom_control *control)
1059 return strlen(tbl_hdr_str) + tbl_hdr_fmt_size +
1060 strlen(rec_hdr_str) + rec_hdr_fmt_size * control->ras_num_recs;
1063 void amdgpu_ras_debugfs_set_ret_size(struct amdgpu_ras_eeprom_control *control)
1065 struct amdgpu_ras *ras = container_of(control, struct amdgpu_ras,
1067 struct dentry *de = ras->de_ras_eeprom_table;
1070 d_inode(de)->i_size = amdgpu_ras_debugfs_table_size(control);
1073 static ssize_t amdgpu_ras_debugfs_table_read(struct file *f, char __user *buf,
1074 size_t size, loff_t *pos)
1076 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
1077 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1078 struct amdgpu_ras_eeprom_control *control = &ras->eeprom_control;
1079 const size_t orig_size = size;
1083 mutex_lock(&control->ras_tbl_mutex);
1085 /* We want *pos - data_len > 0, which means there's
1086 * bytes to be printed from data.
1088 data_len = strlen(tbl_hdr_str);
1089 if (*pos < data_len) {
1091 data_len = min_t(size_t, data_len, size);
1092 if (copy_to_user(buf, &tbl_hdr_str[*pos], data_len))
1099 data_len = strlen(tbl_hdr_str) + tbl_hdr_fmt_size;
1100 if (*pos < data_len && size > 0) {
1101 u8 data[tbl_hdr_fmt_size + 1];
1104 snprintf(data, sizeof(data), tbl_hdr_fmt,
1105 control->tbl_hdr.header,
1106 control->tbl_hdr.version,
1107 control->tbl_hdr.first_rec_offset,
1108 control->tbl_hdr.tbl_size,
1109 control->tbl_hdr.checksum);
1112 data_len = min_t(size_t, data_len, size);
1113 lpos = *pos - strlen(tbl_hdr_str);
1114 if (copy_to_user(buf, &data[lpos], data_len))
1121 data_len = strlen(tbl_hdr_str) + tbl_hdr_fmt_size + strlen(rec_hdr_str);
1122 if (*pos < data_len && size > 0) {
1126 data_len = min_t(size_t, data_len, size);
1127 lpos = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size;
1128 if (copy_to_user(buf, &rec_hdr_str[lpos], data_len))
1135 data_len = amdgpu_ras_debugfs_table_size(control);
1136 if (*pos < data_len && size > 0) {
1137 u8 dare[RAS_TABLE_RECORD_SIZE];
1138 u8 data[rec_hdr_fmt_size + 1];
1139 struct eeprom_table_record record;
1142 /* Find the starting record index
1144 s = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size -
1145 strlen(rec_hdr_str);
1146 s = s / rec_hdr_fmt_size;
1147 r = *pos - strlen(tbl_hdr_str) - tbl_hdr_fmt_size -
1148 strlen(rec_hdr_str);
1149 r = r % rec_hdr_fmt_size;
1151 for ( ; size > 0 && s < control->ras_num_recs; s++) {
1152 u32 ai = RAS_RI_TO_AI(control, s);
1153 /* Read a single record
1155 res = __amdgpu_ras_eeprom_read(control, dare, ai, 1);
1158 __decode_table_record_from_buf(control, &record, dare);
1159 snprintf(data, sizeof(data), rec_hdr_fmt,
1161 RAS_INDEX_TO_OFFSET(control, ai),
1162 record_err_type_str[record.err_type],
1168 record.retired_page);
1170 data_len = min_t(size_t, rec_hdr_fmt_size - r, size);
1171 if (copy_to_user(buf, &data[r], data_len)) {
1183 mutex_unlock(&control->ras_tbl_mutex);
1184 return res < 0 ? res : orig_size - size;
1188 amdgpu_ras_debugfs_eeprom_table_read(struct file *f, char __user *buf,
1189 size_t size, loff_t *pos)
1191 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
1192 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1193 struct amdgpu_ras_eeprom_control *control = ras ? &ras->eeprom_control : NULL;
1200 if (!ras || !control) {
1201 res = snprintf(data, sizeof(data), "Not supported\n");
1206 res = min_t(size_t, res, size);
1208 if (copy_to_user(buf, &data[*pos], res))
1215 return amdgpu_ras_debugfs_table_read(f, buf, size, pos);
1219 const struct file_operations amdgpu_ras_debugfs_eeprom_table_ops = {
1220 .owner = THIS_MODULE,
1221 .read = amdgpu_ras_debugfs_eeprom_table_read,
1223 .llseek = default_llseek,
1227 * __verify_ras_table_checksum -- verify the RAS EEPROM table checksum
1228 * @control: pointer to control structure
1230 * Check the checksum of the stored in EEPROM RAS table.
1232 * Return 0 if the checksum is correct,
1233 * positive if it is not correct, and
1234 * -errno on I/O error.
1236 static int __verify_ras_table_checksum(struct amdgpu_ras_eeprom_control *control)
1238 struct amdgpu_device *adev = to_amdgpu_device(control);
1242 if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1)
1243 buf_size = RAS_TABLE_HEADER_SIZE +
1244 RAS_TABLE_V2_1_INFO_SIZE +
1245 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
1247 buf_size = RAS_TABLE_HEADER_SIZE +
1248 control->ras_num_recs * RAS_TABLE_RECORD_SIZE;
1250 buf = kzalloc(buf_size, GFP_KERNEL);
1252 DRM_ERROR("Out of memory checking RAS table checksum.\n");
1256 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
1257 control->i2c_address +
1258 control->ras_header_offset,
1260 if (res < buf_size) {
1261 DRM_ERROR("Partial read for checksum, res:%d\n", res);
1262 /* On partial reads, return -EIO.
1270 for (pp = buf; pp < buf + buf_size; pp++)
1274 return res < 0 ? res : csum;
1277 static int __read_table_ras_info(struct amdgpu_ras_eeprom_control *control)
1279 struct amdgpu_ras_eeprom_table_ras_info *rai = &control->tbl_rai;
1280 struct amdgpu_device *adev = to_amdgpu_device(control);
1284 buf = kzalloc(RAS_TABLE_V2_1_INFO_SIZE, GFP_KERNEL);
1286 DRM_ERROR("Failed to alloc buf to read EEPROM table ras info\n");
1291 * EEPROM table V2_1 supports ras info,
1292 * read EEPROM table ras info
1294 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
1295 control->i2c_address + control->ras_info_offset,
1296 buf, RAS_TABLE_V2_1_INFO_SIZE);
1297 if (res < RAS_TABLE_V2_1_INFO_SIZE) {
1298 DRM_ERROR("Failed to read EEPROM table ras info, res:%d", res);
1299 res = res >= 0 ? -EIO : res;
1303 __decode_table_ras_info_from_buf(rai, buf);
1307 return res == RAS_TABLE_V2_1_INFO_SIZE ? 0 : res;
1310 int amdgpu_ras_eeprom_init(struct amdgpu_ras_eeprom_control *control,
1311 bool *exceed_err_limit)
1313 struct amdgpu_device *adev = to_amdgpu_device(control);
1314 unsigned char buf[RAS_TABLE_HEADER_SIZE] = { 0 };
1315 struct amdgpu_ras_eeprom_table_header *hdr = &control->tbl_hdr;
1316 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1319 *exceed_err_limit = false;
1321 if (!__is_ras_eeprom_supported(adev))
1324 /* Verify i2c adapter is initialized */
1325 if (!adev->pm.ras_eeprom_i2c_bus || !adev->pm.ras_eeprom_i2c_bus->algo)
1328 if (!__get_eeprom_i2c_addr(adev, control))
1331 control->ras_header_offset = RAS_HDR_START;
1332 control->ras_info_offset = RAS_TABLE_V2_1_INFO_START;
1333 mutex_init(&control->ras_tbl_mutex);
1335 /* Read the table header from EEPROM address */
1336 res = amdgpu_eeprom_read(adev->pm.ras_eeprom_i2c_bus,
1337 control->i2c_address + control->ras_header_offset,
1338 buf, RAS_TABLE_HEADER_SIZE);
1339 if (res < RAS_TABLE_HEADER_SIZE) {
1340 DRM_ERROR("Failed to read EEPROM table header, res:%d", res);
1341 return res >= 0 ? -EIO : res;
1344 __decode_table_header_from_buf(hdr, buf);
1346 if (hdr->version == RAS_TABLE_VER_V2_1) {
1347 control->ras_num_recs = RAS_NUM_RECS_V2_1(hdr);
1348 control->ras_record_offset = RAS_RECORD_START_V2_1;
1349 control->ras_max_record_count = RAS_MAX_RECORD_COUNT_V2_1;
1351 control->ras_num_recs = RAS_NUM_RECS(hdr);
1352 control->ras_record_offset = RAS_RECORD_START;
1353 control->ras_max_record_count = RAS_MAX_RECORD_COUNT;
1355 control->ras_fri = RAS_OFFSET_TO_INDEX(control, hdr->first_rec_offset);
1357 if (hdr->header == RAS_TABLE_HDR_VAL) {
1358 DRM_DEBUG_DRIVER("Found existing EEPROM table with %d records",
1359 control->ras_num_recs);
1361 if (hdr->version == RAS_TABLE_VER_V2_1) {
1362 res = __read_table_ras_info(control);
1367 res = __verify_ras_table_checksum(control);
1369 DRM_ERROR("RAS table incorrect checksum or error:%d\n",
1372 /* Warn if we are at 90% of the threshold or above
1374 if (10 * control->ras_num_recs >= 9 * ras->bad_page_cnt_threshold)
1375 dev_warn(adev->dev, "RAS records:%u exceeds 90%% of threshold:%d",
1376 control->ras_num_recs,
1377 ras->bad_page_cnt_threshold);
1378 } else if (hdr->header == RAS_TABLE_HDR_BAD &&
1379 amdgpu_bad_page_threshold != 0) {
1380 if (hdr->version == RAS_TABLE_VER_V2_1) {
1381 res = __read_table_ras_info(control);
1386 res = __verify_ras_table_checksum(control);
1388 DRM_ERROR("RAS Table incorrect checksum or error:%d\n",
1390 if (ras->bad_page_cnt_threshold > control->ras_num_recs) {
1391 /* This means that, the threshold was increased since
1392 * the last time the system was booted, and now,
1393 * ras->bad_page_cnt_threshold - control->num_recs > 0,
1394 * so that at least one more record can be saved,
1395 * before the page count threshold is reached.
1398 "records:%d threshold:%d, resetting "
1399 "RAS table header signature",
1400 control->ras_num_recs,
1401 ras->bad_page_cnt_threshold);
1402 res = amdgpu_ras_eeprom_correct_header_tag(control,
1405 dev_err(adev->dev, "RAS records:%d exceed threshold:%d",
1406 control->ras_num_recs, ras->bad_page_cnt_threshold);
1407 if (amdgpu_bad_page_threshold == -1) {
1408 dev_warn(adev->dev, "GPU will be initialized due to bad_page_threshold = -1.");
1411 *exceed_err_limit = true;
1413 "RAS records:%d exceed threshold:%d, "
1414 "GPU will not be initialized. Replace this GPU or increase the threshold",
1415 control->ras_num_recs, ras->bad_page_cnt_threshold);
1419 DRM_INFO("Creating a new EEPROM table");
1421 res = amdgpu_ras_eeprom_reset_table(control);
1424 return res < 0 ? res : 0;