From 4e7812e2372bc02809ca6eed285096baffee0ebf Mon Sep 17 00:00:00 2001 From: Tao Zhou Date: Thu, 17 Oct 2024 15:51:54 +0800 Subject: drm/amdgpu: make convert_ras_err_addr visible outside UMC block And change some UMC v12 specific functions to generic version, so the code can be shared. Signed-off-by: Tao Zhou Reviewed-by: Hawking Zhang Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c | 63 +++++++++++++++++++++++++++++++++ 1 file changed, 63 insertions(+) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c index 896f3609b0ee..fe423253ac8c 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c @@ -444,3 +444,66 @@ int amdgpu_umc_logs_ecc_err(struct amdgpu_device *adev, return ret; } + +int amdgpu_umc_lookup_bad_pages_in_a_row(struct amdgpu_device *adev, + uint64_t pa_addr, uint64_t *pfns, int len) +{ + uint32_t i, ret = 0, pos = 0; + struct ta_ras_query_address_output addr_out; + struct ras_err_data err_data; + + err_data.err_addr = + kcalloc(adev->umc.max_ras_err_cnt_per_query, + sizeof(struct eeprom_table_record), GFP_KERNEL); + if (!err_data.err_addr) { + dev_warn(adev->dev, "Failed to alloc memory in bad page lookup!\n"); + return 0; + } + + addr_out.pa.pa = pa_addr; + + if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) + adev->umc.ras->convert_ras_err_addr(adev, &err_data, NULL, + &addr_out, false); + else + goto out; + + for (i = 0; i < adev->umc.max_ras_err_cnt_per_query; i++) { + if (pos >= len) + goto out; + + pfns[pos] = err_data.err_addr[pos].retired_page; + pos++; + } + ret = pos; + +out: + kfree(err_data.err_addr); + return ret; +} + +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, + uint64_t *addr, bool dump_addr) +{ + struct ta_ras_query_address_input addr_in; + struct ta_ras_query_address_output addr_out; + + memset(&addr_in, 0, sizeof(addr_in)); + addr_in.ma.err_addr = err_addr; + addr_in.ma.ch_inst = ch; + addr_in.ma.umc_inst = umc; + addr_in.ma.node_inst = node; + addr_in.ma.socket_id = socket; + + if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) + adev->umc.ras->convert_ras_err_addr(adev, NULL, &addr_in, + &addr_out, dump_addr); + else + return 0; + + *addr = addr_out.pa.pa; + + return 0; +} -- cgit v1.2.3 From 76723fbc5f8632570c6079a2f497c2f2132bce98 Mon Sep 17 00:00:00 2001 From: Tao Zhou Date: Fri, 18 Oct 2024 16:21:26 +0800 Subject: drm/amdgpu: reduce memory usage for umc_lookup_bad_pages_in_a_row The function handles one page in one time, allocating umc.retire_unit bad page records is enough. Signed-off-by: Tao Zhou Reviewed-by: Hawking Zhang Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c index fe423253ac8c..5dd011a73ace 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c @@ -452,8 +452,7 @@ int amdgpu_umc_lookup_bad_pages_in_a_row(struct amdgpu_device *adev, struct ta_ras_query_address_output addr_out; struct ras_err_data err_data; - err_data.err_addr = - kcalloc(adev->umc.max_ras_err_cnt_per_query, + err_data.err_addr = kcalloc(adev->umc.retire_unit, sizeof(struct eeprom_table_record), GFP_KERNEL); if (!err_data.err_addr) { dev_warn(adev->dev, "Failed to alloc memory in bad page lookup!\n"); @@ -468,7 +467,7 @@ int amdgpu_umc_lookup_bad_pages_in_a_row(struct amdgpu_device *adev, else goto out; - for (i = 0; i < adev->umc.max_ras_err_cnt_per_query; i++) { + for (i = 0; i < adev->umc.retire_unit; i++) { if (pos >= len) goto out; -- cgit v1.2.3 From f44a30583bcf2b9c3846264515d618d349e67485 Mon Sep 17 00:00:00 2001 From: Tao Zhou Date: Fri, 18 Oct 2024 16:43:44 +0800 Subject: drm/amdgpu: add return value for convert_ras_err_addr So upper layer can return failure directly if address conversion fails. Signed-off-by: Tao Zhou Reviewed-by: Hawking Zhang Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c | 19 +++++++++++++------ drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h | 2 +- drivers/gpu/drm/amd/amdgpu/umc_v12_0.c | 12 ++++++++---- 3 files changed, 22 insertions(+), 11 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c index 5dd011a73ace..64e1a3406c99 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c @@ -461,11 +461,14 @@ int amdgpu_umc_lookup_bad_pages_in_a_row(struct amdgpu_device *adev, addr_out.pa.pa = pa_addr; - if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) - adev->umc.ras->convert_ras_err_addr(adev, &err_data, NULL, + if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) { + ret = adev->umc.ras->convert_ras_err_addr(adev, &err_data, NULL, &addr_out, false); - else + if (ret) + goto out; + } else { goto out; + } for (i = 0; i < adev->umc.retire_unit; i++) { if (pos >= len) @@ -488,6 +491,7 @@ int amdgpu_umc_mca_to_addr(struct amdgpu_device *adev, { struct ta_ras_query_address_input addr_in; struct ta_ras_query_address_output addr_out; + int ret; memset(&addr_in, 0, sizeof(addr_in)); addr_in.ma.err_addr = err_addr; @@ -496,11 +500,14 @@ int amdgpu_umc_mca_to_addr(struct amdgpu_device *adev, addr_in.ma.node_inst = node; addr_in.ma.socket_id = socket; - if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) - adev->umc.ras->convert_ras_err_addr(adev, NULL, &addr_in, + if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) { + ret = adev->umc.ras->convert_ras_err_addr(adev, NULL, &addr_in, &addr_out, dump_addr); - else + if (ret) + return ret; + } else { return 0; + } *addr = addr_out.pa.pa; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h index abde7597bda8..f45408a6ff03 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h @@ -70,7 +70,7 @@ struct amdgpu_umc_ras { 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); - void (*convert_ras_err_addr)(struct amdgpu_device *adev, + 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, diff --git a/drivers/gpu/drm/amd/amdgpu/umc_v12_0.c b/drivers/gpu/drm/amd/amdgpu/umc_v12_0.c index 9b93ff769b86..ce60fd6675ce 100644 --- a/drivers/gpu/drm/amd/amdgpu/umc_v12_0.c +++ b/drivers/gpu/drm/amd/amdgpu/umc_v12_0.c @@ -173,7 +173,7 @@ static void umc_v12_0_query_ras_error_count(struct amdgpu_device *adev, umc_v12_0_reset_error_count(adev); } -static void umc_v12_0_convert_error_address(struct amdgpu_device *adev, +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, @@ -183,6 +183,7 @@ static void umc_v12_0_convert_error_address(struct amdgpu_device *adev, uint64_t soc_pa, retired_page, 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; @@ -193,11 +194,12 @@ static void umc_v12_0_convert_error_address(struct amdgpu_device *adev, if (addr_in) { err_addr = addr_in->ma.err_addr; addr_in->addr_type = TA_RAS_MCA_TO_PA; - if (psp_ras_query_address(&adev->psp, addr_in, paddr_out)) { + 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); - return; + return ret; } bank = paddr_out->pa.bank; @@ -209,7 +211,7 @@ static void umc_v12_0_convert_error_address(struct amdgpu_device *adev, soc_pa = paddr_out->pa.pa; if (!err_data && !dump_addr) - return; + return ret; col = (err_addr >> 1) & 0x1fULL; /* clear [C3 C2] in soc physical address */ @@ -241,6 +243,8 @@ static void umc_v12_0_convert_error_address(struct amdgpu_device *adev, amdgpu_umc_fill_error_record(err_data, err_addr, retired_page, channel_index, umc_inst); } + + return ret; } static int umc_v12_0_query_error_address(struct amdgpu_device *adev, -- cgit v1.2.3 From b02ef407729d9eb26fbc74ca229e28dbe971e1c1 Mon Sep 17 00:00:00 2001 From: Tao Zhou Date: Thu, 24 Oct 2024 14:46:19 +0800 Subject: drm/amdgpu: add function to find all memory pages in one physical row And the function can be reused across amdgpu driver. Signed-off-by: Tao Zhou Reviewed-by: Hawking Zhang Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c | 38 ++++++++++++++++++++------------- drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h | 2 ++ 2 files changed, 25 insertions(+), 15 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c index 64e1a3406c99..984af815fb38 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c @@ -445,11 +445,27 @@ int amdgpu_umc_logs_ecc_err(struct amdgpu_device *adev, return ret; } +int amdgpu_umc_pages_in_a_row(struct amdgpu_device *adev, + struct ras_err_data *err_data, uint64_t pa_addr) +{ + struct ta_ras_query_address_output addr_out; + + /* reinit err_data */ + err_data->err_addr_cnt = 0; + err_data->err_addr_len = adev->umc.retire_unit; + + addr_out.pa.pa = pa_addr; + if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) + return adev->umc.ras->convert_ras_err_addr(adev, err_data, NULL, + &addr_out, false); + else + return -EINVAL; +} + int amdgpu_umc_lookup_bad_pages_in_a_row(struct amdgpu_device *adev, uint64_t pa_addr, uint64_t *pfns, int len) { - uint32_t i, ret = 0, pos = 0; - struct ta_ras_query_address_output addr_out; + int i, ret; struct ras_err_data err_data; err_data.err_addr = kcalloc(adev->umc.retire_unit, @@ -459,25 +475,17 @@ int amdgpu_umc_lookup_bad_pages_in_a_row(struct amdgpu_device *adev, return 0; } - addr_out.pa.pa = pa_addr; - - if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) { - ret = adev->umc.ras->convert_ras_err_addr(adev, &err_data, NULL, - &addr_out, false); - if (ret) - goto out; - } else { + ret = amdgpu_umc_pages_in_a_row(adev, &err_data, pa_addr); + if (ret) goto out; - } for (i = 0; i < adev->umc.retire_unit; i++) { - if (pos >= len) + if (i >= len) goto out; - pfns[pos] = err_data.err_addr[pos].retired_page; - pos++; + pfns[i] = err_data.err_addr[i].retired_page; } - ret = pos; + ret = i; out: kfree(err_data.err_addr); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h index f45408a6ff03..ce1e4fb385b5 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h @@ -139,6 +139,8 @@ int amdgpu_umc_logs_ecc_err(struct amdgpu_device *adev, 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, -- cgit v1.2.3 From 71a0e9630027f77d7646c5b750593c9ecfaa27d3 Mon Sep 17 00:00:00 2001 From: Tao Zhou Date: Tue, 29 Oct 2024 19:46:44 +0800 Subject: drm/amdgpu: save UMC global channel index to eeprom Save the global channel index returned by RAS TA to eeprom. We can get memory physical address by MCA address and channel index. Signed-off-by: Tao Zhou Reviewed-by: Hawking Zhang Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h | 2 ++ drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c | 7 ++----- drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h | 2 +- drivers/gpu/drm/amd/amdgpu/umc_v12_0.c | 13 ++++++++----- 4 files changed, 13 insertions(+), 11 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h index b13debcf48ee..a6473c7cdeb6 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h @@ -483,6 +483,8 @@ struct ras_ecc_err { uint64_t ipid; uint64_t addr; uint64_t pa_pfn; + /* save global channel index across all UMC instances */ + uint32_t channel_idx; struct ras_err_pages err_pages; }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c index 984af815fb38..4107b78d9dda 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c @@ -495,10 +495,9 @@ out: 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, - uint64_t *addr, bool dump_addr) + struct ta_ras_query_address_output *addr_out, bool dump_addr) { struct ta_ras_query_address_input addr_in; - struct ta_ras_query_address_output addr_out; int ret; memset(&addr_in, 0, sizeof(addr_in)); @@ -510,14 +509,12 @@ int amdgpu_umc_mca_to_addr(struct amdgpu_device *adev, if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) { ret = adev->umc.ras->convert_ras_err_addr(adev, NULL, &addr_in, - &addr_out, dump_addr); + addr_out, dump_addr); if (ret) return ret; } else { return 0; } - *addr = addr_out.pa.pa; - return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h index ce1e4fb385b5..2f71194d5da8 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h @@ -146,5 +146,5 @@ int amdgpu_umc_lookup_bad_pages_in_a_row(struct amdgpu_device *adev, 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, - uint64_t *addr, bool dump_addr); + struct ta_ras_query_address_output *addr_out, bool dump_addr); #endif diff --git a/drivers/gpu/drm/amd/amdgpu/umc_v12_0.c b/drivers/gpu/drm/amd/amdgpu/umc_v12_0.c index 17ef9a6743f5..cce93b4ffb58 100644 --- a/drivers/gpu/drm/amd/amdgpu/umc_v12_0.c +++ b/drivers/gpu/drm/amd/amdgpu/umc_v12_0.c @@ -180,7 +180,7 @@ static int umc_v12_0_convert_error_address(struct amdgpu_device *adev, bool dump_addr) { uint32_t col, col_lower, row, row_lower, bank; - uint32_t channel_index, umc_inst = 0; + uint32_t channel_index = 0, umc_inst = 0; uint32_t i, loop_bits[UMC_V12_0_RETIRE_LOOP_BITS]; uint64_t soc_pa, column, err_addr; struct ta_ras_query_address_output addr_out_tmp; @@ -193,7 +193,7 @@ static int umc_v12_0_convert_error_address(struct amdgpu_device *adev, else paddr_out = addr_out; - err_addr = bank = channel_index = 0; + err_addr = bank = 0; if (addr_in) { err_addr = addr_in->ma.err_addr; addr_in->addr_type = TA_RAS_MCA_TO_PA; @@ -206,7 +206,6 @@ static int umc_v12_0_convert_error_address(struct amdgpu_device *adev, } bank = paddr_out->pa.bank; - channel_index = paddr_out->pa.channel_idx; /* no need to care about umc inst if addr_in is NULL */ umc_inst = addr_in->ma.umc_inst; } @@ -228,6 +227,7 @@ static int umc_v12_0_convert_error_address(struct amdgpu_device *adev, } soc_pa = paddr_out->pa.pa; + channel_index = paddr_out->pa.channel_idx; /* clear loop bits in soc physical address */ for (i = 0; i < UMC_V12_0_RETIRE_LOOP_BITS; i++) soc_pa &= ~BIT_ULL(loop_bits[i]); @@ -466,6 +466,7 @@ static int umc_v12_0_update_ecc_status(struct amdgpu_device *adev, 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; int count, ret, i; hwid = REG_GET_FIELD(ipid, MCMP1_IPIDT0, HardwareID); @@ -495,7 +496,7 @@ static int umc_v12_0_update_ecc_status(struct amdgpu_device *adev, 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), &pa_addr, true); + MCA_IPID_2_SOCKET_ID(ipid), &addr_out, true); if (ret) return ret; @@ -503,10 +504,12 @@ static int umc_v12_0_update_ecc_status(struct amdgpu_device *adev, 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) @@ -577,7 +580,7 @@ static int umc_v12_0_fill_error_record(struct amdgpu_device *adev, ret = amdgpu_umc_fill_error_record(err_data, ecc_err->addr, page_pfn[i] << AMDGPU_GPU_PAGE_SHIFT, - MCA_IPID_2_UMC_CH(ecc_err->ipid), + ecc_err->channel_idx, MCA_IPID_2_UMC_INST(ecc_err->ipid)); if (ret) break; -- cgit v1.2.3 From a8d133e625ceb147a173b6cafc862a9bd4312894 Mon Sep 17 00:00:00 2001 From: Tao Zhou Date: Thu, 31 Oct 2024 15:48:10 +0800 Subject: drm/amdgpu: parse legacy RAS bad page mixed with new data in various NPS modes All legacy RAS bad pages are generated in NPS1 mode, but new bad page can be generated in any NPS mode, so we can't use retired_page stored on eeprom directly in non-nps1 mode even for legacy data. We need to take different actions for different data, new data can be identified from old data by UMC_CHANNEL_IDX_V2 flag. Signed-off-by: Tao Zhou Reviewed-by: Hawking Zhang Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 96 +++++++++++++++++++++++++++------ drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c | 4 +- 3 files changed, 84 insertions(+), 18 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c index d41a0664aa27..81713888d073 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c @@ -192,7 +192,7 @@ static int amdgpu_reserve_page_direct(struct amdgpu_device *adev, uint64_t addre if (amdgpu_bad_page_threshold != 0) { amdgpu_ras_add_bad_pages(adev, err_data.err_addr, - err_data.err_addr_cnt); + err_data.err_addr_cnt, false); amdgpu_ras_save_bad_pages(adev, NULL); } @@ -2728,7 +2728,7 @@ static int amdgpu_ras_realloc_eh_data_space(struct amdgpu_device *adev, return 0; } -static int amdgpu_ras_mca2pa(struct amdgpu_device *adev, +static int amdgpu_ras_mca2pa_by_idx(struct amdgpu_device *adev, struct eeprom_table_record *bps, struct ras_err_data *err_data) { @@ -2757,9 +2757,46 @@ static int amdgpu_ras_mca2pa(struct amdgpu_device *adev, return ret; } +static int amdgpu_ras_mca2pa(struct amdgpu_device *adev, + struct eeprom_table_record *bps, + struct ras_err_data *err_data) +{ + struct ta_ras_query_address_input addr_in; + uint32_t die_id, socket = 0; + + if (adev->smuio.funcs && adev->smuio.funcs->get_socket_id) + socket = adev->smuio.funcs->get_socket_id(adev); + + /* although die id is gotten from PA in nps1 mode, the id is + * fitable for any nps mode + */ + if (adev->umc.ras && adev->umc.ras->get_die_id_from_pa) + die_id = adev->umc.ras->get_die_id_from_pa(adev, bps->address, + bps->retired_page << AMDGPU_GPU_PAGE_SHIFT); + else + return -EINVAL; + + /* reinit err_data */ + err_data->err_addr_cnt = 0; + err_data->err_addr_len = adev->umc.retire_unit; + + memset(&addr_in, 0, sizeof(addr_in)); + addr_in.ma.err_addr = bps->address; + addr_in.ma.ch_inst = bps->mem_channel; + addr_in.ma.umc_inst = bps->mcumc_id; + addr_in.ma.node_inst = die_id; + addr_in.ma.socket_id = socket; + + if (adev->umc.ras && adev->umc.ras->convert_ras_err_addr) + return adev->umc.ras->convert_ras_err_addr(adev, err_data, + &addr_in, NULL, false); + else + return -EINVAL; +} + /* it deal with vram only. */ int amdgpu_ras_add_bad_pages(struct amdgpu_device *adev, - struct eeprom_table_record *bps, int pages) + struct eeprom_table_record *bps, int pages, bool from_rom) { struct amdgpu_ras *con = amdgpu_ras_get_context(adev); struct ras_err_handler_data *data; @@ -2782,12 +2819,7 @@ int amdgpu_ras_add_bad_pages(struct amdgpu_device *adev, is_mca_add = false; } - mutex_lock(&con->recovery_lock); - data = con->eh_data; - if (!data) - goto out; - - if (is_mca_add) { + if (from_rom) { err_data.err_addr = kcalloc(adev->umc.retire_unit, sizeof(struct eeprom_table_record), GFP_KERNEL); @@ -2797,15 +2829,21 @@ int amdgpu_ras_add_bad_pages(struct amdgpu_device *adev, goto out; } + err_rec = err_data.err_addr; loop_cnt = adev->umc.retire_unit; if (adev->gmc.gmc_funcs->query_mem_partition_mode) nps = adev->gmc.gmc_funcs->query_mem_partition_mode(adev); } + mutex_lock(&con->recovery_lock); + data = con->eh_data; + if (!data) + goto free; + for (i = 0; i < pages; i++) { if (is_mca_add) { if (!find_pages_per_pa) { - if (amdgpu_ras_mca2pa(adev, &bps[i], &err_data)) { + if (amdgpu_ras_mca2pa_by_idx(adev, &bps[i], &err_data)) { if (!i && nps == AMDGPU_NPS1_PARTITION_MODE) { /* may use old RAS TA, use PA to find pages in * one row @@ -2825,10 +2863,38 @@ int amdgpu_ras_add_bad_pages(struct amdgpu_device *adev, bps[i].retired_page << AMDGPU_GPU_PAGE_SHIFT)) goto free; } - - err_rec = err_data.err_addr; } else { - err_rec = &bps[i]; + if (from_rom && !find_pages_per_pa) { + if (bps[i].retired_page & UMC_CHANNEL_IDX_V2) { + /* bad page in any NPS mode in eeprom */ + if (amdgpu_ras_mca2pa_by_idx(adev, &bps[i], &err_data)) + goto free; + } else { + /* legacy bad page in eeprom, generated only in + * NPS1 mode + */ + if (amdgpu_ras_mca2pa(adev, &bps[i], &err_data)) { + /* old RAS TA or ASICs which don't support to + * convert addrss via mca address + */ + if (!i && nps == AMDGPU_NPS1_PARTITION_MODE) { + find_pages_per_pa = true; + err_rec = &bps[i]; + loop_cnt = 1; + } else { + /* non-nps1 mode, old RAS TA + * can't support it + */ + goto free; + } + } + } + + if (!find_pages_per_pa) + i += (adev->umc.retire_unit - 1); + } else { + err_rec = &bps[i]; + } } for (j = 0; j < loop_cnt; j++) { @@ -2852,7 +2918,7 @@ int amdgpu_ras_add_bad_pages(struct amdgpu_device *adev, } free: - if (is_mca_add) + if (from_rom) kfree(err_data.err_addr); out: mutex_unlock(&con->recovery_lock); @@ -2955,7 +3021,7 @@ static int amdgpu_ras_load_bad_pages(struct amdgpu_device *adev) control->rec_type = AMDGPU_RAS_EEPROM_REC_MCA; } - ret = amdgpu_ras_add_bad_pages(adev, bps, control->ras_num_recs); + ret = amdgpu_ras_add_bad_pages(adev, bps, control->ras_num_recs, true); } kfree(bps); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h index a6473c7cdeb6..82db986c36a0 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h @@ -753,7 +753,7 @@ int amdgpu_ras_query_error_count(struct amdgpu_device *adev, /* error handling functions */ int amdgpu_ras_add_bad_pages(struct amdgpu_device *adev, - struct eeprom_table_record *bps, int pages); + struct eeprom_table_record *bps, int pages, bool from_rom); int amdgpu_ras_save_bad_pages(struct amdgpu_device *adev, unsigned long *new_cnt); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c index 4107b78d9dda..06145c37e1dd 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c @@ -78,7 +78,7 @@ int amdgpu_umc_page_retirement_mca(struct amdgpu_device *adev, if (amdgpu_bad_page_threshold != 0) { amdgpu_ras_add_bad_pages(adev, err_data.err_addr, - err_data.err_addr_cnt); + err_data.err_addr_cnt, false); amdgpu_ras_save_bad_pages(adev, NULL); } @@ -166,7 +166,7 @@ void amdgpu_umc_handle_bad_pages(struct amdgpu_device *adev, if ((amdgpu_bad_page_threshold != 0) && err_data->err_addr_cnt) { amdgpu_ras_add_bad_pages(adev, err_data->err_addr, - err_data->err_addr_cnt); + err_data->err_addr_cnt, false); amdgpu_ras_save_bad_pages(adev, &err_count); amdgpu_dpm_send_hbm_bad_pages_num(adev, con->eeprom_control.ras_num_recs); -- cgit v1.2.3 From ae756cd8536923841b45340f7ee369272c9bf105 Mon Sep 17 00:00:00 2001 From: Tao Zhou Date: Fri, 29 Nov 2024 16:52:41 +0800 Subject: drm/amdgpu: correct the calculation of RAS bad page After the introduction of NPS RAS, one bad page record on eeprom may be related to 1 or 16 bad pages, so the bad page record and bad page are two different concepts, define a new variable to store bad page number. Signed-off-by: Tao Zhou Reviewed-by: Hawking Zhang Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 10 ++----- drivers/gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.c | 40 +++++++++++++++++--------- drivers/gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.h | 5 ++++ drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c | 3 +- 4 files changed, 36 insertions(+), 22 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c index 738a645867ef..1de934cd5764 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c @@ -2943,13 +2943,7 @@ int amdgpu_ras_save_bad_pages(struct amdgpu_device *adev, mutex_lock(&con->recovery_lock); control = &con->eeprom_control; data = con->eh_data; - bad_page_num = control->ras_num_recs; - /* one record on eeprom stands for all pages in one memory row - * in this mode - */ - if (control->rec_type == AMDGPU_RAS_EEPROM_REC_MCA) - bad_page_num = control->ras_num_recs * adev->umc.retire_unit; - + bad_page_num = control->ras_num_bad_pages; save_count = data->count - bad_page_num; mutex_unlock(&con->recovery_lock); @@ -3433,7 +3427,7 @@ int amdgpu_ras_init_badpage_info(struct amdgpu_device *adev) return ret; amdgpu_dpm_send_hbm_bad_pages_num( - adev, control->ras_num_recs); + adev, control->ras_num_bad_pages); if (con->update_channel_flag == true) { amdgpu_dpm_send_hbm_bad_channel_flag( diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.c index 0db6d1bd2002..0d824f016916 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.c @@ -470,9 +470,10 @@ int amdgpu_ras_eeprom_reset_table(struct amdgpu_ras_eeprom_control *control) res = __write_table_ras_info(control); control->ras_num_recs = 0; + control->ras_num_bad_pages = 0; control->ras_fri = 0; - amdgpu_dpm_send_hbm_bad_pages_num(adev, control->ras_num_recs); + amdgpu_dpm_send_hbm_bad_pages_num(adev, control->ras_num_bad_pages); control->bad_channel_bitmap = 0; amdgpu_dpm_send_hbm_bad_channel_flag(adev, control->bad_channel_bitmap); @@ -559,7 +560,7 @@ bool amdgpu_ras_eeprom_check_err_threshold(struct amdgpu_device *adev) if (con->eeprom_control.tbl_hdr.header == RAS_TABLE_HDR_BAD) { if (amdgpu_bad_page_threshold == -1) { dev_warn(adev->dev, "RAS records:%d exceed threshold:%d", - con->eeprom_control.ras_num_recs, con->bad_page_cnt_threshold); + con->eeprom_control.ras_num_bad_pages, con->bad_page_cnt_threshold); dev_warn(adev->dev, "But GPU can be operated due to bad_page_threshold = -1.\n"); return false; @@ -621,6 +622,7 @@ amdgpu_ras_eeprom_append_table(struct amdgpu_ras_eeprom_control *control, const u32 num) { struct amdgpu_ras *con = amdgpu_ras_get_context(to_amdgpu_device(control)); + struct amdgpu_device *adev = to_amdgpu_device(control); u32 a, b, i; u8 *buf, *pp; int res; @@ -723,6 +725,12 @@ amdgpu_ras_eeprom_append_table(struct amdgpu_ras_eeprom_control *control, control->ras_num_recs = 1 + (control->ras_max_record_count + b - control->ras_fri) % control->ras_max_record_count; + + if (control->rec_type == AMDGPU_RAS_EEPROM_REC_PA) + control->ras_num_bad_pages = control->ras_num_recs; + else + control->ras_num_bad_pages = + control->ras_num_recs * adev->umc.retire_unit; Out: kfree(buf); return res; @@ -740,10 +748,10 @@ amdgpu_ras_eeprom_update_header(struct amdgpu_ras_eeprom_control *control) /* Modify the header if it exceeds. */ if (amdgpu_bad_page_threshold != 0 && - control->ras_num_recs >= ras->bad_page_cnt_threshold) { + control->ras_num_bad_pages >= ras->bad_page_cnt_threshold) { dev_warn(adev->dev, "Saved bad pages %d reaches threshold value %d\n", - control->ras_num_recs, ras->bad_page_cnt_threshold); + control->ras_num_bad_pages, ras->bad_page_cnt_threshold); control->tbl_hdr.header = RAS_TABLE_HDR_BAD; if (control->tbl_hdr.version == RAS_TABLE_VER_V2_1) { control->tbl_rai.rma_status = GPU_RETIRED__ECC_REACH_THRESHOLD; @@ -798,9 +806,9 @@ amdgpu_ras_eeprom_update_header(struct amdgpu_ras_eeprom_control *control) */ if (amdgpu_bad_page_threshold != 0 && control->tbl_hdr.version == RAS_TABLE_VER_V2_1 && - control->ras_num_recs < ras->bad_page_cnt_threshold) + control->ras_num_bad_pages < ras->bad_page_cnt_threshold) control->tbl_rai.health_percent = ((ras->bad_page_cnt_threshold - - control->ras_num_recs) * 100) / + control->ras_num_bad_pages) * 100) / ras->bad_page_cnt_threshold; /* Recalc the checksum. @@ -1402,9 +1410,15 @@ int amdgpu_ras_eeprom_check(struct amdgpu_ras_eeprom_control *control) if (!__get_eeprom_i2c_addr(adev, control)) return -EINVAL; + if (control->rec_type == AMDGPU_RAS_EEPROM_REC_PA) + control->ras_num_bad_pages = control->ras_num_recs; + else + control->ras_num_bad_pages = + control->ras_num_recs * adev->umc.retire_unit; + if (hdr->header == RAS_TABLE_HDR_VAL) { DRM_DEBUG_DRIVER("Found existing EEPROM table with %d records", - control->ras_num_recs); + control->ras_num_bad_pages); if (hdr->version == RAS_TABLE_VER_V2_1) { res = __read_table_ras_info(control); @@ -1419,9 +1433,9 @@ int amdgpu_ras_eeprom_check(struct amdgpu_ras_eeprom_control *control) /* Warn if we are at 90% of the threshold or above */ - if (10 * control->ras_num_recs >= 9 * ras->bad_page_cnt_threshold) + if (10 * control->ras_num_bad_pages >= 9 * ras->bad_page_cnt_threshold) dev_warn(adev->dev, "RAS records:%u exceeds 90%% of threshold:%d", - control->ras_num_recs, + control->ras_num_bad_pages, ras->bad_page_cnt_threshold); } else if (hdr->header == RAS_TABLE_HDR_BAD && amdgpu_bad_page_threshold != 0) { @@ -1437,7 +1451,7 @@ int amdgpu_ras_eeprom_check(struct amdgpu_ras_eeprom_control *control) res); return -EINVAL; } - if (ras->bad_page_cnt_threshold > control->ras_num_recs) { + if (ras->bad_page_cnt_threshold > control->ras_num_bad_pages) { /* This means that, the threshold was increased since * the last time the system was booted, and now, * ras->bad_page_cnt_threshold - control->num_recs > 0, @@ -1447,13 +1461,13 @@ int amdgpu_ras_eeprom_check(struct amdgpu_ras_eeprom_control *control) dev_info(adev->dev, "records:%d threshold:%d, resetting " "RAS table header signature", - control->ras_num_recs, + control->ras_num_bad_pages, ras->bad_page_cnt_threshold); res = amdgpu_ras_eeprom_correct_header_tag(control, RAS_TABLE_HDR_VAL); } else { dev_err(adev->dev, "RAS records:%d exceed threshold:%d", - control->ras_num_recs, ras->bad_page_cnt_threshold); + control->ras_num_bad_pages, ras->bad_page_cnt_threshold); if (amdgpu_bad_page_threshold == -1) { dev_warn(adev->dev, "GPU will be initialized due to bad_page_threshold = -1."); res = 0; @@ -1462,7 +1476,7 @@ int amdgpu_ras_eeprom_check(struct amdgpu_ras_eeprom_control *control) dev_err(adev->dev, "RAS records:%d exceed threshold:%d, " "GPU will not be initialized. Replace this GPU or increase the threshold", - control->ras_num_recs, ras->bad_page_cnt_threshold); + control->ras_num_bad_pages, ras->bad_page_cnt_threshold); } } } else { diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.h index b87422df52fd..81d55cb7b397 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.h @@ -95,6 +95,11 @@ struct amdgpu_ras_eeprom_control { */ u32 ras_num_recs; + /* the bad page number is ras_num_recs or + * ras_num_recs * umc.retire_unit + */ + u32 ras_num_bad_pages; + /* First record index to read, 0-based. * Range is [0, num_recs-1]. This is * an absolute index, starting right after diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c index 06145c37e1dd..eafe20d8fe0b 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c @@ -169,7 +169,8 @@ void amdgpu_umc_handle_bad_pages(struct amdgpu_device *adev, err_data->err_addr_cnt, false); amdgpu_ras_save_bad_pages(adev, &err_count); - amdgpu_dpm_send_hbm_bad_pages_num(adev, con->eeprom_control.ras_num_recs); + amdgpu_dpm_send_hbm_bad_pages_num(adev, + con->eeprom_control.ras_num_bad_pages); if (con->update_channel_flag == true) { amdgpu_dpm_send_hbm_bad_channel_flag(adev, con->eeprom_control.bad_channel_bitmap); -- cgit v1.2.3