diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h')
| -rw-r--r-- | drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h | 110 |
1 files changed, 101 insertions, 9 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h index e46439274f3a..28dff750c47e 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h @@ -21,7 +21,7 @@ #ifndef __AMDGPU_UMC_H__ #define __AMDGPU_UMC_H__ #include "amdgpu_ras.h" - +#include "amdgpu_mca.h" /* * (addr / 256) * 4096, the higher 26 bits in ErrorAddr * is the index of 4KB block @@ -32,6 +32,11 @@ * is the index of 8KB block */ #define ADDR_OF_8KB_BLOCK(addr) (((addr) & ~0xffULL) << 5) +/* + * (addr / 256) * 32768, the higher 26 bits in ErrorAddr + * is the index of 8KB block + */ +#define ADDR_OF_32KB_BLOCK(addr) (((addr) & ~0xffULL) << 7) /* channel index is the index of 256B block */ #define ADDR_OF_256B_BLOCK(channel_index) ((channel_index) << 8) /* offset in 256B block */ @@ -42,22 +47,74 @@ #define LOOP_UMC_INST_AND_CH(umc_inst, ch_inst) LOOP_UMC_INST((umc_inst)) LOOP_UMC_CH_INST((ch_inst)) #define LOOP_UMC_NODE_INST(node_inst) \ - for ((node_inst) = 0; (node_inst) < adev->umc.node_inst_num; (node_inst)++) + for_each_set_bit((node_inst), &(adev->umc.active_mask), adev->umc.node_inst_num) #define LOOP_UMC_EACH_NODE_INST_AND_CH(node_inst, umc_inst, ch_inst) \ LOOP_UMC_NODE_INST((node_inst)) LOOP_UMC_INST_AND_CH((umc_inst), (ch_inst)) +/* Page retirement tag */ +#define UMC_ECC_NEW_DETECTED_TAG 0x1 +/* + * a flag to indicate v2 of channel index stored in eeprom + * + * v1 (legacy way): store channel index within a umc instance in eeprom + * range in UMC v12: 0 ~ 7 + * v2: store global channel index in eeprom + * range in UMC v12: 0 ~ 127 + * + * NOTE: it's better to store it in eeprom_table_record.mem_channel, + * but there is only 8 bits in mem_channel, and the channel number may + * increase in the future, we decide to save it in + * eeprom_table_record.retired_page. retired_page is useless in v2, + * we depend on eeprom_table_record.address instead of retired_page in v2. + * Only 48 bits are saved on eeprom, use bit 47 here. + */ +#define UMC_CHANNEL_IDX_V2 BIT_ULL(47) + +/* + * save nps value to eeprom_table_record.retired_page[47:40], + * the channel index flag above will be retired. + */ +#define UMC_NPS_SHIFT 40 +#define UMC_NPS_MASK 0xffULL + +/* three column bits and one row bit in MCA address flip + * in bad page retirement + */ +#define RETIRE_FLIP_BITS_NUM 4 + +struct amdgpu_umc_flip_bits { + uint32_t flip_bits_in_pa[RETIRE_FLIP_BITS_NUM]; + uint32_t flip_row_bit; + uint32_t r13_in_pa; + uint32_t bit_num; +}; + +typedef int (*umc_func)(struct amdgpu_device *adev, uint32_t node_inst, + uint32_t umc_inst, uint32_t ch_inst, void *data); + struct amdgpu_umc_ras { struct amdgpu_ras_block_object ras_block; void (*err_cnt_init)(struct amdgpu_device *adev); bool (*query_ras_poison_mode)(struct amdgpu_device *adev); - void (*convert_ras_error_address)(struct amdgpu_device *adev, - struct ras_err_data *err_data, uint64_t err_addr, - uint32_t ch_inst, uint32_t umc_inst); void (*ecc_info_query_ras_error_count)(struct amdgpu_device *adev, void *ras_error_status); void (*ecc_info_query_ras_error_address)(struct amdgpu_device *adev, void *ras_error_status); + bool (*check_ecc_err_status)(struct amdgpu_device *adev, + enum amdgpu_mca_error_type type, void *ras_error_status); + int (*update_ecc_status)(struct amdgpu_device *adev, + uint64_t status, uint64_t ipid, uint64_t addr); + int (*convert_ras_err_addr)(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 (*get_die_id_from_pa)(struct amdgpu_device *adev, + uint64_t mca_addr, uint64_t retired_page); + void (*get_retire_flip_bits)(struct amdgpu_device *adev); + void (*mca_ipid_parse)(struct amdgpu_device *adev, uint64_t ipid, + uint32_t *did, uint32_t *ch, uint32_t *umc_inst, uint32_t *sid); }; struct amdgpu_umc_funcs { @@ -72,27 +129,39 @@ struct amdgpu_umc { /* number of umc instance with memory map register access */ uint32_t umc_inst_num; - /*number of umc node instance with memory map register access*/ + /* Total number of umc node instance including harvest one */ uint32_t node_inst_num; /* UMC regiser per channel offset */ uint32_t channel_offs; + /* how many pages are retired in one UE */ + uint32_t retire_unit; /* channel index table of interleaved memory */ const uint32_t *channel_idx_tbl; struct ras_common_if *ras_if; const struct amdgpu_umc_funcs *funcs; struct amdgpu_umc_ras *ras; + + /* active mask for umc node instance */ + unsigned long active_mask; + + struct amdgpu_umc_flip_bits flip_bits; + + unsigned long err_addr_cnt; }; +int amdgpu_umc_ras_sw_init(struct amdgpu_device *adev); int amdgpu_umc_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block); int amdgpu_umc_poison_handler(struct amdgpu_device *adev, - void *ras_error_status, - bool reset); + enum amdgpu_ras_block block, uint32_t reset); +int amdgpu_umc_pasid_poison_handler(struct amdgpu_device *adev, + enum amdgpu_ras_block block, uint16_t pasid, + pasid_notify pasid_fn, void *data, uint32_t reset); int amdgpu_umc_process_ecc_irq(struct amdgpu_device *adev, struct amdgpu_irq_src *source, struct amdgpu_iv_entry *entry); -void amdgpu_umc_fill_error_record(struct ras_err_data *err_data, +int amdgpu_umc_fill_error_record(struct ras_err_data *err_data, uint64_t err_addr, uint64_t retired_page, uint32_t channel_index, @@ -101,4 +170,27 @@ void amdgpu_umc_fill_error_record(struct ras_err_data *err_data, int amdgpu_umc_process_ras_data_cb(struct amdgpu_device *adev, void *ras_error_status, struct amdgpu_iv_entry *entry); +int amdgpu_umc_page_retirement_mca(struct amdgpu_device *adev, + uint64_t err_addr, uint32_t ch_inst, uint32_t umc_inst); + +int amdgpu_umc_loop_channels(struct amdgpu_device *adev, + umc_func func, void *data); + +int amdgpu_umc_update_ecc_status(struct amdgpu_device *adev, + uint64_t status, uint64_t ipid, uint64_t addr); +int amdgpu_umc_logs_ecc_err(struct amdgpu_device *adev, + struct radix_tree_root *ecc_tree, struct ras_ecc_err *ecc_err); + +void amdgpu_umc_handle_bad_pages(struct amdgpu_device *adev, + void *ras_error_status); +int amdgpu_umc_pages_in_a_row(struct amdgpu_device *adev, + struct ras_err_data *err_data, uint64_t pa_addr); +int amdgpu_umc_lookup_bad_pages_in_a_row(struct amdgpu_device *adev, + uint64_t pa_addr, uint64_t *pfns, int len); +int amdgpu_umc_mca_to_addr(struct amdgpu_device *adev, + uint64_t err_addr, uint32_t ch, uint32_t umc, + uint32_t node, uint32_t socket, + struct ta_ras_query_address_output *addr_out, bool dump_addr); +int amdgpu_umc_pa2mca(struct amdgpu_device *adev, + uint64_t pa, uint64_t *mca, enum amdgpu_memory_partition nps); #endif |