diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/umc_v12_0.c')
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/umc_v12_0.c | 615 |
1 files changed, 386 insertions, 229 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/umc_v12_0.c b/drivers/gpu/drm/amd/amdgpu/umc_v12_0.c index 77af4e25ff46..e590cbdd8de9 100644 --- a/drivers/gpu/drm/amd/amdgpu/umc_v12_0.c +++ b/drivers/gpu/drm/amd/amdgpu/umc_v12_0.c @@ -28,27 +28,8 @@ #include "umc/umc_12_0_0_sh_mask.h" #include "mp/mp_13_0_6_sh_mask.h" -const uint32_t - umc_v12_0_channel_idx_tbl[] - [UMC_V12_0_UMC_INSTANCE_NUM] - [UMC_V12_0_CHANNEL_INSTANCE_NUM] = { - {{3, 7, 11, 15, 2, 6, 10, 14}, {1, 5, 9, 13, 0, 4, 8, 12}, - {19, 23, 27, 31, 18, 22, 26, 30}, {17, 21, 25, 29, 16, 20, 24, 28}}, - {{47, 43, 39, 35, 46, 42, 38, 34}, {45, 41, 37, 33, 44, 40, 36, 32}, - {63, 59, 55, 51, 62, 58, 54, 50}, {61, 57, 53, 49, 60, 56, 52, 48}}, - {{79, 75, 71, 67, 78, 74, 70, 66}, {77, 73, 69, 65, 76, 72, 68, 64}, - {95, 91, 87, 83, 94, 90, 86, 82}, {93, 89, 85, 81, 92, 88, 84, 80}}, - {{99, 103, 107, 111, 98, 102, 106, 110}, {97, 101, 105, 109, 96, 100, 104, 108}, - {115, 119, 123, 127, 114, 118, 122, 126}, {113, 117, 121, 125, 112, 116, 120, 124}} - }; - -/* mapping of MCA error address to normalized address */ -static const uint32_t umc_v12_0_ma2na_mapping[] = { - 0, 5, 6, 8, 9, 14, 12, 13, - 10, 11, 15, 16, 17, 18, 19, 20, - 21, 22, 23, 24, 25, 26, 27, 28, - 24, 7, 29, 30, -}; +#define MAX_ECC_NUM_PER_RETIREMENT 32 +#define DELAYED_TIME_FOR_GPU_RESET 1000 //ms static inline uint64_t get_umc_v12_0_reg_offset(struct amdgpu_device *adev, uint32_t node_inst, @@ -91,7 +72,7 @@ static void umc_v12_0_reset_error_count(struct amdgpu_device *adev) bool umc_v12_0_is_deferred_error(struct amdgpu_device *adev, uint64_t mc_umc_status) { - dev_info(adev->dev, + dev_dbg(adev->dev, "MCA_UMC_STATUS(0x%llx): Val:%llu, Poison:%llu, Deferred:%llu, PCC:%llu, UC:%llu, TCC:%llu\n", mc_umc_status, REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, Val), @@ -104,7 +85,8 @@ bool umc_v12_0_is_deferred_error(struct amdgpu_device *adev, uint64_t mc_umc_sta return (amdgpu_ras_is_poison_mode_supported(adev) && (REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, Val) == 1) && - (REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, Deferred) == 1)); + ((REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, Deferred) == 1) || + (REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, Poison) == 1))); } bool umc_v12_0_is_uncorrectable_error(struct amdgpu_device *adev, uint64_t mc_umc_status) @@ -176,9 +158,9 @@ static int umc_v12_0_query_error_count(struct amdgpu_device *adev, umc_v12_0_query_error_count_per_type(adev, umc_reg_offset, &de_count, umc_v12_0_is_deferred_error); - amdgpu_ras_error_statistic_ue_count(err_data, &mcm_info, NULL, ue_count); - amdgpu_ras_error_statistic_ce_count(err_data, &mcm_info, NULL, ce_count); - amdgpu_ras_error_statistic_de_count(err_data, &mcm_info, NULL, de_count); + amdgpu_ras_error_statistic_ue_count(err_data, &mcm_info, ue_count); + amdgpu_ras_error_statistic_ce_count(err_data, &mcm_info, ce_count); + amdgpu_ras_error_statistic_de_count(err_data, &mcm_info, de_count); return 0; } @@ -192,143 +174,164 @@ static void umc_v12_0_query_ras_error_count(struct amdgpu_device *adev, umc_v12_0_reset_error_count(adev); } -static bool umc_v12_0_bit_wise_xor(uint32_t val) +static void umc_v12_0_get_retire_flip_bits(struct amdgpu_device *adev) { - bool result = 0; - int i; + enum amdgpu_memory_partition nps = AMDGPU_NPS1_PARTITION_MODE; + uint32_t vram_type = adev->gmc.vram_type; + struct amdgpu_umc_flip_bits *flip_bits = &(adev->umc.flip_bits); + + if (adev->gmc.gmc_funcs->query_mem_partition_mode) + nps = adev->gmc.gmc_funcs->query_mem_partition_mode(adev); + + /* default setting */ + flip_bits->flip_bits_in_pa[0] = UMC_V12_0_PA_C2_BIT; + flip_bits->flip_bits_in_pa[1] = UMC_V12_0_PA_C3_BIT; + flip_bits->flip_bits_in_pa[2] = UMC_V12_0_PA_C4_BIT; + flip_bits->flip_bits_in_pa[3] = UMC_V12_0_PA_R13_BIT; + flip_bits->flip_row_bit = 13; + flip_bits->bit_num = 4; + flip_bits->r13_in_pa = UMC_V12_0_PA_R13_BIT; + + if (nps == AMDGPU_NPS2_PARTITION_MODE) { + flip_bits->flip_bits_in_pa[0] = UMC_V12_0_PA_CH5_BIT; + flip_bits->flip_bits_in_pa[1] = UMC_V12_0_PA_C2_BIT; + flip_bits->flip_bits_in_pa[2] = UMC_V12_0_PA_B1_BIT; + flip_bits->r13_in_pa = UMC_V12_0_PA_R12_BIT; + } else if (nps == AMDGPU_NPS4_PARTITION_MODE) { + flip_bits->flip_bits_in_pa[0] = UMC_V12_0_PA_CH4_BIT; + flip_bits->flip_bits_in_pa[1] = UMC_V12_0_PA_CH5_BIT; + flip_bits->flip_bits_in_pa[2] = UMC_V12_0_PA_B0_BIT; + flip_bits->r13_in_pa = UMC_V12_0_PA_R11_BIT; + } - for (i = 0; i < 32; i++) - result = result ^ ((val >> i) & 0x1); + switch (vram_type) { + case AMDGPU_VRAM_TYPE_HBM: + /* other nps modes are taken as nps1 */ + if (nps == AMDGPU_NPS2_PARTITION_MODE) + flip_bits->flip_bits_in_pa[3] = UMC_V12_0_PA_R12_BIT; + else if (nps == AMDGPU_NPS4_PARTITION_MODE) + flip_bits->flip_bits_in_pa[3] = UMC_V12_0_PA_R11_BIT; - return result; -} + break; + case AMDGPU_VRAM_TYPE_HBM3E: + flip_bits->flip_bits_in_pa[3] = UMC_V12_0_PA_R12_BIT; + flip_bits->flip_row_bit = 12; -static void umc_v12_0_mca_addr_to_pa(struct amdgpu_device *adev, - uint64_t err_addr, uint32_t ch_inst, uint32_t umc_inst, - uint32_t node_inst, - struct ta_ras_query_address_output *addr_out) -{ - uint32_t channel_index, i; - uint64_t na, soc_pa; - uint32_t bank_hash0, bank_hash1, bank_hash2, bank_hash3, col, row; - uint32_t bank0, bank1, bank2, bank3, bank; - - bank_hash0 = (err_addr >> UMC_V12_0_MCA_B0_BIT) & 0x1ULL; - bank_hash1 = (err_addr >> UMC_V12_0_MCA_B1_BIT) & 0x1ULL; - bank_hash2 = (err_addr >> UMC_V12_0_MCA_B2_BIT) & 0x1ULL; - bank_hash3 = (err_addr >> UMC_V12_0_MCA_B3_BIT) & 0x1ULL; - col = (err_addr >> 1) & 0x1fULL; - row = (err_addr >> 10) & 0x3fffULL; - - /* apply bank hash algorithm */ - bank0 = - bank_hash0 ^ (UMC_V12_0_XOR_EN0 & - (umc_v12_0_bit_wise_xor(col & UMC_V12_0_COL_XOR0) ^ - (umc_v12_0_bit_wise_xor(row & UMC_V12_0_ROW_XOR0)))); - bank1 = - bank_hash1 ^ (UMC_V12_0_XOR_EN1 & - (umc_v12_0_bit_wise_xor(col & UMC_V12_0_COL_XOR1) ^ - (umc_v12_0_bit_wise_xor(row & UMC_V12_0_ROW_XOR1)))); - bank2 = - bank_hash2 ^ (UMC_V12_0_XOR_EN2 & - (umc_v12_0_bit_wise_xor(col & UMC_V12_0_COL_XOR2) ^ - (umc_v12_0_bit_wise_xor(row & UMC_V12_0_ROW_XOR2)))); - bank3 = - bank_hash3 ^ (UMC_V12_0_XOR_EN3 & - (umc_v12_0_bit_wise_xor(col & UMC_V12_0_COL_XOR3) ^ - (umc_v12_0_bit_wise_xor(row & UMC_V12_0_ROW_XOR3)))); - - bank = bank0 | (bank1 << 1) | (bank2 << 2) | (bank3 << 3); - err_addr &= ~0x3c0ULL; - err_addr |= (bank << UMC_V12_0_MCA_B0_BIT); - - na = 0x0; - /* convert mca error address to normalized address */ - for (i = 1; i < ARRAY_SIZE(umc_v12_0_ma2na_mapping); i++) - na |= ((err_addr >> i) & 0x1ULL) << umc_v12_0_ma2na_mapping[i]; - - channel_index = - adev->umc.channel_idx_tbl[node_inst * adev->umc.umc_inst_num * - adev->umc.channel_inst_num + - umc_inst * adev->umc.channel_inst_num + - ch_inst]; - /* translate umc channel address to soc pa, 3 parts are included */ - soc_pa = ADDR_OF_32KB_BLOCK(na) | - ADDR_OF_256B_BLOCK(channel_index) | - OFFSET_IN_256B_BLOCK(na); - - /* the umc channel bits are not original values, they are hashed */ - UMC_V12_0_SET_CHANNEL_HASH(channel_index, soc_pa); - - addr_out->pa.pa = soc_pa; - addr_out->pa.bank = bank; - addr_out->pa.channel_idx = channel_index; + if (nps == AMDGPU_NPS2_PARTITION_MODE) + flip_bits->flip_bits_in_pa[3] = UMC_V12_0_PA_R11_BIT; + else if (nps == AMDGPU_NPS4_PARTITION_MODE) + flip_bits->flip_bits_in_pa[3] = UMC_V12_0_PA_R10_BIT; + + break; + default: + dev_warn(adev->dev, + "Unknown HBM type, set RAS retire flip bits to the value in NPS1 mode.\n"); + break; + } + + adev->umc.retire_unit = 0x1 << flip_bits->bit_num; } -static void umc_v12_0_convert_error_address(struct amdgpu_device *adev, - struct ras_err_data *err_data, uint64_t err_addr, - uint32_t ch_inst, uint32_t umc_inst, - uint32_t node_inst) +static int umc_v12_0_convert_error_address(struct amdgpu_device *adev, + struct ras_err_data *err_data, + struct ta_ras_query_address_input *addr_in, + struct ta_ras_query_address_output *addr_out, + bool dump_addr) { - uint32_t col, row, row_xor, bank, channel_index; - uint64_t soc_pa, retired_page, column; - struct ta_ras_query_address_input addr_in; - struct ta_ras_query_address_output addr_out; + uint32_t col, col_lower, row, row_lower, row_high, bank; + uint32_t channel_index = 0, umc_inst = 0; + uint32_t i, bit_num, retire_unit, *flip_bits; + uint64_t soc_pa, column, err_addr; + struct ta_ras_query_address_output addr_out_tmp; + struct ta_ras_query_address_output *paddr_out; + int ret = 0; + + if (!addr_out) + paddr_out = &addr_out_tmp; + else + paddr_out = addr_out; + + err_addr = bank = 0; + if (addr_in) { + err_addr = addr_in->ma.err_addr; + addr_in->addr_type = TA_RAS_MCA_TO_PA; + ret = psp_ras_query_address(&adev->psp, addr_in, paddr_out); + if (ret) { + dev_warn(adev->dev, "Failed to query RAS physical address for 0x%llx", + err_addr); + + goto out; + } + + bank = paddr_out->pa.bank; + /* no need to care about umc inst if addr_in is NULL */ + umc_inst = addr_in->ma.umc_inst; + } + + flip_bits = adev->umc.flip_bits.flip_bits_in_pa; + bit_num = adev->umc.flip_bits.bit_num; + retire_unit = adev->umc.retire_unit; + + soc_pa = paddr_out->pa.pa; + channel_index = paddr_out->pa.channel_idx; + /* clear loop bits in soc physical address */ + for (i = 0; i < bit_num; i++) + soc_pa &= ~BIT_ULL(flip_bits[i]); + + paddr_out->pa.pa = soc_pa; + /* get column bit 0 and 1 in mca address */ + col_lower = (err_addr >> 1) & 0x3ULL; + /* extra row bit will be handled later */ + row_lower = (err_addr >> UMC_V12_0_MA_R0_BIT) & 0x1fffULL; + row_lower &= ~BIT_ULL(adev->umc.flip_bits.flip_row_bit); + + if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(9, 5, 0)) { + row_high = (soc_pa >> adev->umc.flip_bits.r13_in_pa) & 0x3ULL; + /* it's 2.25GB in each channel, from MCA address to PA + * [R14 R13] is converted if the two bits value are 0x3, + * get them from PA instead of MCA address. + */ + row_lower |= (row_high << 13); + } - addr_in.addr_type = TA_RAS_MCA_TO_PA; - addr_in.ma.err_addr = err_addr; - addr_in.ma.ch_inst = ch_inst; - addr_in.ma.umc_inst = umc_inst; - addr_in.ma.node_inst = node_inst; - - if (psp_ras_query_address(&adev->psp, &addr_in, &addr_out)) - /* fallback to old path if fail to get pa from psp */ - umc_v12_0_mca_addr_to_pa(adev, err_addr, ch_inst, umc_inst, - node_inst, &addr_out); - - soc_pa = addr_out.pa.pa; - bank = addr_out.pa.bank; - channel_index = addr_out.pa.channel_idx; - - col = (err_addr >> 1) & 0x1fULL; - row = (err_addr >> 10) & 0x3fffULL; - row_xor = row ^ (0x1ULL << 13); - /* clear [C3 C2] in soc physical address */ - soc_pa &= ~(0x3ULL << UMC_V12_0_PA_C2_BIT); - /* clear [C4] in soc physical address */ - soc_pa &= ~(0x1ULL << UMC_V12_0_PA_C4_BIT); - - /* loop for all possibilities of [C4 C3 C2] */ - for (column = 0; column < UMC_V12_0_NA_MAP_PA_NUM; column++) { - retired_page = soc_pa | ((column & 0x3) << UMC_V12_0_PA_C2_BIT); - retired_page |= (((column & 0x4) >> 2) << UMC_V12_0_PA_C4_BIT); - /* include column bit 0 and 1 */ - col &= 0x3; - col |= (column << 2); - dev_info(adev->dev, - "Error Address(PA):0x%-10llx Row:0x%-4x Col:0x%-2x Bank:0x%x Channel:0x%x\n", - retired_page, row, col, bank, channel_index); - amdgpu_umc_fill_error_record(err_data, err_addr, - retired_page, channel_index, umc_inst); - - /* shift R13 bit */ - retired_page ^= (0x1ULL << UMC_V12_0_PA_R13_BIT); - dev_info(adev->dev, - "Error Address(PA):0x%-10llx Row:0x%-4x Col:0x%-2x Bank:0x%x Channel:0x%x\n", - retired_page, row_xor, col, bank, channel_index); - amdgpu_umc_fill_error_record(err_data, err_addr, - retired_page, channel_index, umc_inst); + if (!err_data && !dump_addr) + goto out; + + /* loop for all possibilities of retired bits */ + for (column = 0; column < retire_unit; column++) { + soc_pa = paddr_out->pa.pa; + for (i = 0; i < bit_num; i++) + soc_pa |= (((column >> i) & 0x1ULL) << flip_bits[i]); + + col = ((column & 0x7) << 2) | col_lower; + /* handle extra row bit */ + if (bit_num == RETIRE_FLIP_BITS_NUM) + row = ((column >> 3) << adev->umc.flip_bits.flip_row_bit) | + row_lower; + + if (dump_addr) + dev_info(adev->dev, + "Error Address(PA):0x%-10llx Row:0x%-4x Col:0x%-2x Bank:0x%x Channel:0x%x\n", + soc_pa, row, col, bank, channel_index); + + if (err_data) + amdgpu_umc_fill_error_record(err_data, err_addr, + soc_pa, channel_index, umc_inst); } + +out: + return ret; } static int umc_v12_0_query_error_address(struct amdgpu_device *adev, uint32_t node_inst, uint32_t umc_inst, uint32_t ch_inst, void *data) { + struct ras_err_data *err_data = (struct ras_err_data *)data; + struct ta_ras_query_address_input addr_in; uint64_t mc_umc_status_addr; uint64_t mc_umc_status, err_addr; uint64_t mc_umc_addrt0; - struct ras_err_data *err_data = (struct ras_err_data *)data; uint64_t umc_reg_offset = get_umc_v12_0_reg_offset(adev, node_inst, umc_inst, ch_inst); @@ -357,8 +360,19 @@ static int umc_v12_0_query_error_address(struct amdgpu_device *adev, err_addr = REG_GET_FIELD(err_addr, MCA_UMC_UMC0_MCUMC_ADDRT0, ErrorAddr); - umc_v12_0_convert_error_address(adev, err_data, err_addr, - ch_inst, umc_inst, node_inst); + if (!adev->aid_mask && + adev->smuio.funcs && + adev->smuio.funcs->get_socket_id) + addr_in.ma.socket_id = adev->smuio.funcs->get_socket_id(adev); + else + addr_in.ma.socket_id = 0; + + addr_in.ma.err_addr = err_addr; + addr_in.ma.ch_inst = ch_inst; + addr_in.ma.umc_inst = umc_inst; + addr_in.ma.node_inst = node_inst; + + umc_v12_0_convert_error_address(adev, err_data, &addr_in, NULL, true); } /* clear umc status */ @@ -401,64 +415,6 @@ static int umc_v12_0_err_cnt_init_per_channel(struct amdgpu_device *adev, return 0; } -static void umc_v12_0_ecc_info_query_ras_error_count(struct amdgpu_device *adev, - void *ras_error_status) -{ - amdgpu_mca_smu_log_ras_error(adev, - AMDGPU_RAS_BLOCK__UMC, AMDGPU_MCA_ERROR_TYPE_CE, ras_error_status); - amdgpu_mca_smu_log_ras_error(adev, - AMDGPU_RAS_BLOCK__UMC, AMDGPU_MCA_ERROR_TYPE_UE, ras_error_status); -} - -static void umc_v12_0_ecc_info_query_ras_error_address(struct amdgpu_device *adev, - void *ras_error_status) -{ - struct ras_err_node *err_node; - uint64_t mc_umc_status; - struct ras_err_info *err_info; - struct ras_err_addr *mca_err_addr, *tmp; - struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; - - for_each_ras_error(err_node, err_data) { - err_info = &err_node->err_info; - if (list_empty(&err_info->err_addr_list)) - continue; - - list_for_each_entry_safe(mca_err_addr, tmp, &err_info->err_addr_list, node) { - mc_umc_status = mca_err_addr->err_status; - if (mc_umc_status && - (umc_v12_0_is_uncorrectable_error(adev, mc_umc_status) || - umc_v12_0_is_deferred_error(adev, mc_umc_status))) { - uint64_t mca_addr, err_addr, mca_ipid; - uint32_t InstanceIdLo; - - mca_addr = mca_err_addr->err_addr; - mca_ipid = mca_err_addr->err_ipid; - - err_addr = REG_GET_FIELD(mca_addr, - MCA_UMC_UMC0_MCUMC_ADDRT0, ErrorAddr); - InstanceIdLo = REG_GET_FIELD(mca_ipid, MCMP1_IPIDT0, InstanceIdLo); - - dev_info(adev->dev, "UMC:IPID:0x%llx, aid:%d, inst:%d, ch:%d, err_addr:0x%llx\n", - mca_ipid, - err_info->mcm_info.die_id, - MCA_IPID_LO_2_UMC_INST(InstanceIdLo), - MCA_IPID_LO_2_UMC_CH(InstanceIdLo), - err_addr); - - umc_v12_0_convert_error_address(adev, - err_data, err_addr, - MCA_IPID_LO_2_UMC_CH(InstanceIdLo), - MCA_IPID_LO_2_UMC_INST(InstanceIdLo), - err_info->mcm_info.die_id); - } - - /* Delete error address node from list and free memory */ - amdgpu_ras_del_mca_err_addr(err_info, mca_err_addr); - } - } -} - static bool umc_v12_0_check_ecc_err_status(struct amdgpu_device *adev, enum amdgpu_mca_error_type type, void *ras_error_status) { @@ -498,43 +454,54 @@ const struct amdgpu_ras_block_hw_ops umc_v12_0_ras_hw_ops = { .query_ras_error_address = umc_v12_0_query_ras_error_address, }; -static int umc_v12_0_aca_bank_generate_report(struct aca_handle *handle, struct aca_bank *bank, enum aca_error_type type, - struct aca_bank_report *report, void *data) +static int umc_v12_0_aca_bank_parser(struct aca_handle *handle, struct aca_bank *bank, + enum aca_smu_type type, void *data) { struct amdgpu_device *adev = handle->adev; - u64 status; + struct aca_bank_info info; + enum aca_error_type err_type; + u64 status, count; + u32 ext_error_code; int ret; - ret = aca_bank_info_decode(bank, &report->info); + status = bank->regs[ACA_REG_IDX_STATUS]; + if (umc_v12_0_is_deferred_error(adev, status)) + err_type = ACA_ERROR_TYPE_DEFERRED; + else if (umc_v12_0_is_uncorrectable_error(adev, status)) + err_type = ACA_ERROR_TYPE_UE; + else if (umc_v12_0_is_correctable_error(adev, status)) + err_type = ACA_ERROR_TYPE_CE; + else + return 0; + bank->aca_err_type = err_type; + + ret = aca_bank_info_decode(bank, &info); if (ret) return ret; - status = bank->regs[ACA_REG_IDX_STATUS]; - switch (type) { - case ACA_ERROR_TYPE_UE: - if (umc_v12_0_is_uncorrectable_error(adev, status)) { - report->count[type] = 1; - } - break; - case ACA_ERROR_TYPE_CE: - if (umc_v12_0_is_correctable_error(adev, status)) { - report->count[type] = 1; - } - break; - default: - return -EINVAL; - } + amdgpu_umc_update_ecc_status(adev, + bank->regs[ACA_REG_IDX_STATUS], + bank->regs[ACA_REG_IDX_IPID], + bank->regs[ACA_REG_IDX_ADDR]); - return 0; + ext_error_code = ACA_REG__STATUS__ERRORCODEEXT(status); + if (umc_v12_0_is_deferred_error(adev, status)) + count = ext_error_code == 0 ? + adev->umc.err_addr_cnt / adev->umc.retire_unit : 1ULL; + else + count = ext_error_code == 0 ? + ACA_REG__MISC0__ERRCNT(bank->regs[ACA_REG_IDX_MISC0]) : 1ULL; + + return aca_error_cache_log_bank_error(handle, &info, err_type, count); } static const struct aca_bank_ops umc_v12_0_aca_bank_ops = { - .aca_bank_generate_report = umc_v12_0_aca_bank_generate_report, + .aca_bank_parser = umc_v12_0_aca_bank_parser, }; const struct aca_info umc_v12_0_aca_info = { .hwip = ACA_HWIP_TYPE_UMC, - .mask = ACA_ERROR_UE_MASK | ACA_ERROR_CE_MASK, + .mask = ACA_ERROR_UE_MASK | ACA_ERROR_CE_MASK | ACA_ERROR_DEFERRED_MASK, .bank_ops = &umc_v12_0_aca_bank_ops, }; @@ -554,6 +521,193 @@ static int umc_v12_0_ras_late_init(struct amdgpu_device *adev, struct ras_common return 0; } +static int umc_v12_0_update_ecc_status(struct amdgpu_device *adev, + uint64_t status, uint64_t ipid, uint64_t addr) +{ + struct amdgpu_ras *con = amdgpu_ras_get_context(adev); + uint16_t hwid, mcatype; + uint64_t page_pfn[UMC_V12_0_BAD_PAGE_NUM_PER_CHANNEL]; + uint64_t err_addr, pa_addr = 0; + struct ras_ecc_err *ecc_err; + struct ta_ras_query_address_output addr_out; + uint32_t shift_bit = adev->umc.flip_bits.flip_bits_in_pa[2]; + int count, ret, i; + + hwid = REG_GET_FIELD(ipid, MCMP1_IPIDT0, HardwareID); + mcatype = REG_GET_FIELD(ipid, MCMP1_IPIDT0, McaType); + + if ((hwid != MCA_UMC_HWID_V12_0) || (mcatype != MCA_UMC_MCATYPE_V12_0)) + return 0; + + if (!status) + return 0; + + if (!umc_v12_0_is_deferred_error(adev, status)) + return 0; + + err_addr = REG_GET_FIELD(addr, + MCA_UMC_UMC0_MCUMC_ADDRT0, ErrorAddr); + + dev_dbg(adev->dev, + "UMC:IPID:0x%llx, socket:%llu, aid:%llu, inst:%llu, ch:%llu, err_addr:0x%llx\n", + ipid, + MCA_IPID_2_SOCKET_ID(ipid), + MCA_IPID_2_DIE_ID(ipid), + MCA_IPID_2_UMC_INST(ipid), + MCA_IPID_2_UMC_CH(ipid), + err_addr); + + ret = amdgpu_umc_mca_to_addr(adev, + err_addr, MCA_IPID_2_UMC_CH(ipid), + MCA_IPID_2_UMC_INST(ipid), MCA_IPID_2_DIE_ID(ipid), + MCA_IPID_2_SOCKET_ID(ipid), &addr_out, true); + if (ret) + return ret; + + ecc_err = kzalloc(sizeof(*ecc_err), GFP_KERNEL); + if (!ecc_err) + return -ENOMEM; + + pa_addr = addr_out.pa.pa; + ecc_err->status = status; + ecc_err->ipid = ipid; + ecc_err->addr = addr; + ecc_err->pa_pfn = pa_addr >> AMDGPU_GPU_PAGE_SHIFT; + ecc_err->channel_idx = addr_out.pa.channel_idx; + + /* If converted pa_pfn is 0, use pa C4 pfn. */ + if (!ecc_err->pa_pfn) + ecc_err->pa_pfn = BIT_ULL(shift_bit) >> AMDGPU_GPU_PAGE_SHIFT; + + ret = amdgpu_umc_logs_ecc_err(adev, &con->umc_ecc_log.de_page_tree, ecc_err); + if (ret) { + if (ret == -EEXIST) + con->umc_ecc_log.de_queried_count++; + else + dev_err(adev->dev, "Fail to log ecc error! ret:%d\n", ret); + + kfree(ecc_err); + return ret; + } + + con->umc_ecc_log.de_queried_count++; + + memset(page_pfn, 0, sizeof(page_pfn)); + count = amdgpu_umc_lookup_bad_pages_in_a_row(adev, + pa_addr, + page_pfn, ARRAY_SIZE(page_pfn)); + if (count <= 0) { + dev_warn(adev->dev, "Fail to convert error address! count:%d\n", count); + return 0; + } + + /* Reserve memory */ + for (i = 0; i < count; i++) + amdgpu_ras_reserve_page(adev, page_pfn[i]); + + /* The problem case is as follows: + * 1. GPU A triggers a gpu ras reset, and GPU A drives + * GPU B to also perform a gpu ras reset. + * 2. After gpu B ras reset started, gpu B queried a DE + * data. Since the DE data was queried in the ras reset + * thread instead of the page retirement thread, bad + * page retirement work would not be triggered. Then + * even if all gpu resets are completed, the bad pages + * will be cached in RAM until GPU B's bad page retirement + * work is triggered again and then saved to eeprom. + * Trigger delayed work to save the bad pages to eeprom in time + * after gpu ras reset is completed. + */ + if (amdgpu_ras_in_recovery(adev)) + schedule_delayed_work(&con->page_retirement_dwork, + msecs_to_jiffies(DELAYED_TIME_FOR_GPU_RESET)); + + return 0; +} + +static int umc_v12_0_fill_error_record(struct amdgpu_device *adev, + struct ras_ecc_err *ecc_err, void *ras_error_status) +{ + struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; + uint64_t page_pfn[UMC_V12_0_BAD_PAGE_NUM_PER_CHANNEL]; + int ret, i, count; + + if (!err_data || !ecc_err) + return -EINVAL; + + memset(page_pfn, 0, sizeof(page_pfn)); + count = amdgpu_umc_lookup_bad_pages_in_a_row(adev, + ecc_err->pa_pfn << AMDGPU_GPU_PAGE_SHIFT, + page_pfn, ARRAY_SIZE(page_pfn)); + + for (i = 0; i < count; i++) { + ret = amdgpu_umc_fill_error_record(err_data, + ecc_err->addr, + page_pfn[i] << AMDGPU_GPU_PAGE_SHIFT, + ecc_err->channel_idx, + MCA_IPID_2_UMC_INST(ecc_err->ipid)); + if (ret) + break; + } + + err_data->de_count++; + + return ret; +} + +static void umc_v12_0_query_ras_ecc_err_addr(struct amdgpu_device *adev, + void *ras_error_status) +{ + struct amdgpu_ras *con = amdgpu_ras_get_context(adev); + struct ras_ecc_err *entries[MAX_ECC_NUM_PER_RETIREMENT]; + struct radix_tree_root *ecc_tree; + int new_detected, ret, i; + + ecc_tree = &con->umc_ecc_log.de_page_tree; + + mutex_lock(&con->umc_ecc_log.lock); + new_detected = radix_tree_gang_lookup_tag(ecc_tree, (void **)entries, + 0, ARRAY_SIZE(entries), UMC_ECC_NEW_DETECTED_TAG); + for (i = 0; i < new_detected; i++) { + if (!entries[i]) + continue; + + ret = umc_v12_0_fill_error_record(adev, entries[i], ras_error_status); + if (ret) { + dev_err(adev->dev, "Fail to fill umc error record, ret:%d\n", ret); + break; + } + radix_tree_tag_clear(ecc_tree, + entries[i]->pa_pfn, UMC_ECC_NEW_DETECTED_TAG); + } + mutex_unlock(&con->umc_ecc_log.lock); +} + +static uint32_t umc_v12_0_get_die_id(struct amdgpu_device *adev, + uint64_t mca_addr, uint64_t retired_page) +{ + uint32_t die = 0; + + /* we only calculate die id for nps1 mode right now */ + die += ((((retired_page >> 12) & 0x1ULL)^ + ((retired_page >> 20) & 0x1ULL) ^ + ((retired_page >> 27) & 0x1ULL) ^ + ((retired_page >> 34) & 0x1ULL) ^ + ((retired_page >> 41) & 0x1ULL)) << 0); + + /* the original PA_C4 and PA_R13 may be cleared in retired_page, so + * get them from mca_addr. + */ + die += ((((retired_page >> 13) & 0x1ULL) ^ + ((mca_addr >> 5) & 0x1ULL) ^ + ((retired_page >> 28) & 0x1ULL) ^ + ((mca_addr >> 23) & 0x1ULL) ^ + ((retired_page >> 42) & 0x1ULL)) << 1); + die &= 3; + + return die; +} + struct amdgpu_umc_ras umc_v12_0_ras = { .ras_block = { .hw_ops = &umc_v12_0_ras_hw_ops, @@ -561,8 +715,11 @@ struct amdgpu_umc_ras umc_v12_0_ras = { }, .err_cnt_init = umc_v12_0_err_cnt_init, .query_ras_poison_mode = umc_v12_0_query_ras_poison_mode, - .ecc_info_query_ras_error_count = umc_v12_0_ecc_info_query_ras_error_count, - .ecc_info_query_ras_error_address = umc_v12_0_ecc_info_query_ras_error_address, + .ecc_info_query_ras_error_address = umc_v12_0_query_ras_ecc_err_addr, .check_ecc_err_status = umc_v12_0_check_ecc_err_status, + .update_ecc_status = umc_v12_0_update_ecc_status, + .convert_ras_err_addr = umc_v12_0_convert_error_address, + .get_die_id_from_pa = umc_v12_0_get_die_id, + .get_retire_flip_bits = umc_v12_0_get_retire_flip_bits, }; |