1 files changed, 243 insertions, 152 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c
index 540e0f066b26..3f0b0e9af4f3 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c
@@ -24,6 +24,7 @@
 #include <linux/sort.h>
 #include "amdgpu.h"
 #include "umc_v6_7.h"
+#include "amdgpu_ras_mgr.h"
 #define MAX_UMC_POISON_POLLING_TIME_SYNC   20  //ms
 
 #define MAX_UMC_HASH_STRING_SIZE  256
@@ -78,7 +79,7 @@ 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);
 	}
 
@@ -96,67 +97,96 @@ void amdgpu_umc_handle_bad_pages(struct amdgpu_device *adev,
 {
 	struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status;
 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
+	struct amdgpu_ras_eeprom_control *control = &con->eeprom_control;
 	unsigned int error_query_mode;
 	int ret = 0;
 	unsigned long err_count;
 
 	amdgpu_ras_get_error_query_mode(adev, &error_query_mode);
 
+	err_data->err_addr =
+		kcalloc(adev->umc.max_ras_err_cnt_per_query,
+			sizeof(struct eeprom_table_record), GFP_KERNEL);
+
+	/* still call query_ras_error_address to clear error status
+	 * even NOMEM error is encountered
+	 */
+	if (!err_data->err_addr)
+		dev_warn(adev->dev,
+			"Failed to alloc memory for umc error address record!\n");
+	else
+		err_data->err_addr_len = adev->umc.max_ras_err_cnt_per_query;
+
 	mutex_lock(&con->page_retirement_lock);
-	ret = amdgpu_dpm_get_ecc_info(adev, (void *)&(con->umc_ecc));
-	if (ret == -EOPNOTSUPP &&
-	    error_query_mode == AMDGPU_RAS_DIRECT_ERROR_QUERY) {
-		if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
-		    adev->umc.ras->ras_block.hw_ops->query_ras_error_count)
-		    adev->umc.ras->ras_block.hw_ops->query_ras_error_count(adev, ras_error_status);
-
-		if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
-		    adev->umc.ras->ras_block.hw_ops->query_ras_error_address &&
-		    adev->umc.max_ras_err_cnt_per_query) {
-			err_data->err_addr =
-				kcalloc(adev->umc.max_ras_err_cnt_per_query,
-					sizeof(struct eeprom_table_record), GFP_KERNEL);
-
-			/* still call query_ras_error_address to clear error status
-			 * even NOMEM error is encountered
-			 */
-			if(!err_data->err_addr)
-				dev_warn(adev->dev, "Failed to alloc memory for "
-						"umc error address record!\n");
-			else
-				err_data->err_addr_len = adev->umc.max_ras_err_cnt_per_query;
-
-			/* umc query_ras_error_address is also responsible for clearing
-			 * error status
-			 */
-			adev->umc.ras->ras_block.hw_ops->query_ras_error_address(adev, ras_error_status);
+	if (!amdgpu_ras_smu_eeprom_supported(adev)) {
+		ret = amdgpu_dpm_get_ecc_info(adev, (void *)&(con->umc_ecc));
+		if (ret == -EOPNOTSUPP &&
+		    error_query_mode == AMDGPU_RAS_DIRECT_ERROR_QUERY) {
+			if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
+			    adev->umc.ras->ras_block.hw_ops->query_ras_error_count)
+				adev->umc.ras->ras_block.hw_ops->query_ras_error_count(adev,
+								ras_error_status);
+
+			if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
+			    adev->umc.ras->ras_block.hw_ops->query_ras_error_address &&
+			    adev->umc.max_ras_err_cnt_per_query) {
+				err_data->err_addr =
+					kcalloc(adev->umc.max_ras_err_cnt_per_query,
+						sizeof(struct eeprom_table_record), GFP_KERNEL);
+
+				/* still call query_ras_error_address to clear error status
+				 * even NOMEM error is encountered
+				 */
+				if (!err_data->err_addr)
+					dev_warn(adev->dev,
+						"Failed to alloc memory for umc error address record!\n");
+				else
+					err_data->err_addr_len =
+						adev->umc.max_ras_err_cnt_per_query;
+
+				/* umc query_ras_error_address is also responsible for clearing
+				 * error status
+				 */
+				adev->umc.ras->ras_block.hw_ops->query_ras_error_address(adev,
+								ras_error_status);
+			}
+		} else if (error_query_mode == AMDGPU_RAS_FIRMWARE_ERROR_QUERY ||
+		    (!ret && error_query_mode == AMDGPU_RAS_DIRECT_ERROR_QUERY)) {
+			if (adev->umc.ras &&
+			    adev->umc.ras->ecc_info_query_ras_error_count)
+				adev->umc.ras->ecc_info_query_ras_error_count(adev,
+								ras_error_status);
+
+			if (adev->umc.ras &&
+			    adev->umc.ras->ecc_info_query_ras_error_address &&
+			    adev->umc.max_ras_err_cnt_per_query) {
+				err_data->err_addr =
+					kcalloc(adev->umc.max_ras_err_cnt_per_query,
+						sizeof(struct eeprom_table_record), GFP_KERNEL);
+
+				/* still call query_ras_error_address to clear error status
+				 * even NOMEM error is encountered
+				 */
+				if (!err_data->err_addr)
+					dev_warn(adev->dev,
+						"Failed to alloc memory for umc error address record!\n");
+				else
+					err_data->err_addr_len =
+						adev->umc.max_ras_err_cnt_per_query;
+
+				/* umc query_ras_error_address is also responsible for clearing
+				 * error status
+				 */
+				adev->umc.ras->ecc_info_query_ras_error_address(adev,
+								ras_error_status);
+			}
 		}
-	} else if (error_query_mode == AMDGPU_RAS_FIRMWARE_ERROR_QUERY ||
-	    (!ret && error_query_mode == AMDGPU_RAS_DIRECT_ERROR_QUERY)) {
-		if (adev->umc.ras &&
-		    adev->umc.ras->ecc_info_query_ras_error_count)
-		    adev->umc.ras->ecc_info_query_ras_error_count(adev, ras_error_status);
-
-		if (adev->umc.ras &&
-		    adev->umc.ras->ecc_info_query_ras_error_address &&
-		    adev->umc.max_ras_err_cnt_per_query) {
-			err_data->err_addr =
-				kcalloc(adev->umc.max_ras_err_cnt_per_query,
-					sizeof(struct eeprom_table_record), GFP_KERNEL);
-
-			/* still call query_ras_error_address to clear error status
-			 * even NOMEM error is encountered
-			 */
-			if(!err_data->err_addr)
-				dev_warn(adev->dev, "Failed to alloc memory for "
-						"umc error address record!\n");
-			else
-				err_data->err_addr_len = adev->umc.max_ras_err_cnt_per_query;
-
-			/* umc query_ras_error_address is also responsible for clearing
-			 * error status
-			 */
-			adev->umc.ras->ecc_info_query_ras_error_address(adev, ras_error_status);
+	} else {
+		if (!amdgpu_ras_eeprom_update_record_num(control)) {
+			err_data->err_addr_cnt = err_data->de_count =
+				control->ras_num_recs -	control->ras_num_recs_old;
+			amdgpu_ras_eeprom_read_idx(control, err_data->err_addr,
+				control->ras_num_recs_old, err_data->de_count);
 		}
 	}
 
@@ -166,10 +196,11 @@ 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, amdgpu_ras_smu_eeprom_supported(adev));
 			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);
@@ -195,7 +226,8 @@ static int amdgpu_umc_do_page_retirement(struct amdgpu_device *adev,
 	kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
 	amdgpu_umc_handle_bad_pages(adev, ras_error_status);
 
-	if (err_data->ue_count && reset) {
+	if ((err_data->ue_count || err_data->de_count) &&
+	    (reset || amdgpu_ras_is_rma(adev))) {
 		con->gpu_reset_flags |= reset;
 		amdgpu_ras_reset_gpu(adev);
 	}
@@ -203,56 +235,6 @@ static int amdgpu_umc_do_page_retirement(struct amdgpu_device *adev,
 	return AMDGPU_RAS_SUCCESS;
 }
 
-int amdgpu_umc_bad_page_polling_timeout(struct amdgpu_device *adev,
-			uint32_t reset, uint32_t timeout_ms)
-{
-	struct ras_err_data err_data;
-	struct ras_common_if head = {
-		.block = AMDGPU_RAS_BLOCK__UMC,
-	};
-	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head);
-	uint32_t timeout = timeout_ms;
-
-	memset(&err_data, 0, sizeof(err_data));
-	amdgpu_ras_error_data_init(&err_data);
-
-	do {
-
-		amdgpu_umc_handle_bad_pages(adev, &err_data);
-
-		if (timeout && !err_data.de_count) {
-			msleep(1);
-			timeout--;
-		}
-
-	} while (timeout && !err_data.de_count);
-
-	if (!timeout)
-		dev_warn(adev->dev, "Can't find bad pages\n");
-
-	if (err_data.de_count)
-		dev_info(adev->dev, "%ld new deferred hardware errors detected\n", err_data.de_count);
-
-	if (obj) {
-		obj->err_data.ue_count += err_data.ue_count;
-		obj->err_data.ce_count += err_data.ce_count;
-		obj->err_data.de_count += err_data.de_count;
-	}
-
-	amdgpu_ras_error_data_fini(&err_data);
-
-	kgd2kfd_set_sram_ecc_flag(adev->kfd.dev);
-
-	if (reset) {
-		struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
-
-		con->gpu_reset_flags |= reset;
-		amdgpu_ras_reset_gpu(adev);
-	}
-
-	return 0;
-}
-
 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)
@@ -292,15 +274,26 @@ int amdgpu_umc_pasid_poison_handler(struct amdgpu_device *adev,
 			}
 
 			amdgpu_ras_error_data_fini(&err_data);
+		} else if (amdgpu_uniras_enabled(adev)) {
+			struct ras_ih_info ih_info = {0};
+
+			ih_info.block = block;
+			ih_info.pasid = pasid;
+			ih_info.reset = reset;
+			ih_info.pasid_fn = pasid_fn;
+			ih_info.data = data;
+			amdgpu_ras_mgr_handle_consumer_interrupt(adev, &ih_info);
 		} else {
-				struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
-
-				amdgpu_ras_put_poison_req(adev,
-					block, pasid, pasid_fn, data, reset);
+			struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
+			int ret;
 
+			ret = amdgpu_ras_put_poison_req(adev,
+				block, pasid, pasid_fn, data, reset);
+			if (!ret) {
 				atomic_inc(&con->page_retirement_req_cnt);
-
+				atomic_inc(&con->poison_consumption_count);
 				wake_up(&con->page_retirement_wq);
+			}
 		}
 	} else {
 		if (adev->virt.ops && adev->virt.ops->ras_poison_handler)
@@ -366,6 +359,9 @@ int amdgpu_umc_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *r
 	if (r)
 		return r;
 
+	if (amdgpu_sriov_vf(adev))
+		return r;
+
 	if (amdgpu_ras_is_supported(adev, ras_block->block)) {
 		r = amdgpu_irq_get(adev, &adev->gmc.ecc_irq, 0);
 		if (r)
@@ -431,6 +427,45 @@ int amdgpu_umc_fill_error_record(struct ras_err_data *err_data,
 	return 0;
 }
 
+static int amdgpu_umc_loop_all_aid(struct amdgpu_device *adev, umc_func func,
+				   void *data)
+{
+	uint32_t umc_node_inst;
+	uint32_t node_inst;
+	uint32_t umc_inst;
+	uint32_t ch_inst;
+	int ret;
+
+	/*
+	 * This loop is done based on the following -
+	 * umc.active mask = mask of active umc instances across all nodes
+	 * umc.umc_inst_num = maximum number of umc instancess per node
+	 * umc.node_inst_num = maximum number of node instances
+	 * Channel instances are not assumed to be harvested.
+	 */
+	dev_dbg(adev->dev, "active umcs :%lx umc_inst per node: %d",
+		adev->umc.active_mask, adev->umc.umc_inst_num);
+	for_each_set_bit(umc_node_inst, &(adev->umc.active_mask),
+			 adev->umc.node_inst_num * adev->umc.umc_inst_num) {
+		node_inst = umc_node_inst / adev->umc.umc_inst_num;
+		umc_inst = umc_node_inst % adev->umc.umc_inst_num;
+		LOOP_UMC_CH_INST(ch_inst) {
+			dev_dbg(adev->dev,
+				"node_inst :%d umc_inst: %d ch_inst: %d",
+				node_inst, umc_inst, ch_inst);
+			ret = func(adev, node_inst, umc_inst, ch_inst, data);
+			if (ret) {
+				dev_err(adev->dev,
+					"Node %d umc %d ch %d func returns %d\n",
+					node_inst, umc_inst, ch_inst, ret);
+				return ret;
+			}
+		}
+	}
+
+	return 0;
+}
+
 int amdgpu_umc_loop_channels(struct amdgpu_device *adev,
 			umc_func func, void *data)
 {
@@ -439,6 +474,9 @@ int amdgpu_umc_loop_channels(struct amdgpu_device *adev,
 	uint32_t ch_inst         = 0;
 	int ret = 0;
 
+	if (adev->aid_mask)
+		return amdgpu_umc_loop_all_aid(adev, func, data);
+
 	if (adev->umc.node_inst_num) {
 		LOOP_UMC_EACH_NODE_INST_AND_CH(node_inst, umc_inst, ch_inst) {
 			ret = func(adev, node_inst, umc_inst, ch_inst, data);
@@ -471,66 +509,119 @@ int amdgpu_umc_update_ecc_status(struct amdgpu_device *adev,
 	return 0;
 }
 
-static int amdgpu_umc_uint64_cmp(const void *a, const void *b)
+int amdgpu_umc_logs_ecc_err(struct amdgpu_device *adev,
+		struct radix_tree_root *ecc_tree, struct ras_ecc_err *ecc_err)
 {
-	uint64_t *addr_a = (uint64_t *)a;
-	uint64_t *addr_b = (uint64_t *)b;
+	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
+	struct ras_ecc_log_info *ecc_log;
+	int ret;
 
-	if (*addr_a > *addr_b)
-		return 1;
-	else if (*addr_a < *addr_b)
-		return -1;
-	else
-		return 0;
+	ecc_log = &con->umc_ecc_log;
+
+	mutex_lock(&ecc_log->lock);
+	ret = radix_tree_insert(ecc_tree, ecc_err->pa_pfn, ecc_err);
+	if (!ret)
+		radix_tree_tag_set(ecc_tree,
+			ecc_err->pa_pfn, UMC_ECC_NEW_DETECTED_TAG);
+	mutex_unlock(&ecc_log->lock);
+
+	return ret;
 }
 
-/* Use string hash to avoid logging the same bad pages repeatedly */
-int amdgpu_umc_build_pages_hash(struct amdgpu_device *adev,
-		uint64_t *pfns, int len, uint64_t *val)
+int amdgpu_umc_pages_in_a_row(struct amdgpu_device *adev,
+			struct ras_err_data *err_data, uint64_t pa_addr)
 {
-	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
-	char buf[MAX_UMC_HASH_STRING_SIZE] = {0};
-	int offset = 0, i = 0;
-	uint64_t hash_val;
+	struct ta_ras_query_address_output addr_out;
 
-	if (!pfns || !len)
+	/* 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;
+}
 
-	sort(pfns, len, sizeof(uint64_t), amdgpu_umc_uint64_cmp, NULL);
+int amdgpu_umc_lookup_bad_pages_in_a_row(struct amdgpu_device *adev,
+			uint64_t pa_addr, uint64_t *pfns, int len)
+{
+	int i, ret;
+	struct ras_err_data err_data;
 
-	for (i = 0; i < len; i++)
-		offset += snprintf(&buf[offset], sizeof(buf) - offset, "%llx", pfns[i]);
+	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");
+		return 0;
+	}
 
-	hash_val = siphash(buf, offset, &con->umc_ecc_log.ecc_key);
+	ret = amdgpu_umc_pages_in_a_row(adev, &err_data, pa_addr);
+	if (ret)
+		goto out;
 
-	*val = hash_val;
+	for (i = 0; i < adev->umc.retire_unit; i++) {
+		if (i >= len)
+			goto out;
 
-	return 0;
+		pfns[i] = err_data.err_addr[i].retired_page;
+	}
+	ret = i;
+	adev->umc.err_addr_cnt = err_data.err_addr_cnt;
+
+out:
+	kfree(err_data.err_addr);
+	return ret;
 }
 
-int amdgpu_umc_logs_ecc_err(struct amdgpu_device *adev,
-		struct radix_tree_root *ecc_tree, struct ras_ecc_err *ecc_err)
+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)
 {
-	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
-	struct ras_ecc_log_info *ecc_log;
+	struct ta_ras_query_address_input addr_in;
 	int ret;
 
-	ecc_log = &con->umc_ecc_log;
+	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) {
+		ret = adev->umc.ras->convert_ras_err_addr(adev, NULL, &addr_in,
+				addr_out, dump_addr);
+		if (ret)
+			return ret;
+	} else {
+		return 0;
+	}
 
-	mutex_lock(&ecc_log->lock);
-	ret = radix_tree_insert(ecc_tree, ecc_err->hash_index, ecc_err);
-	if (!ret) {
-		struct ras_err_pages *err_pages = &ecc_err->err_pages;
-		int i;
+	return 0;
+}
 
-		/* Reserve memory */
-		for (i = 0; i < err_pages->count; i++)
-			amdgpu_ras_reserve_page(adev, err_pages->pfn[i]);
+int amdgpu_umc_pa2mca(struct amdgpu_device *adev,
+		uint64_t pa, uint64_t *mca, enum amdgpu_memory_partition nps)
+{
+	struct ta_ras_query_address_input addr_in;
+	struct ta_ras_query_address_output addr_out;
+	int ret;
 
-		radix_tree_tag_set(ecc_tree,
-			ecc_err->hash_index, UMC_ECC_NEW_DETECTED_TAG);
+	/* nps: the pa belongs to */
+	addr_in.pa.pa = pa | ((uint64_t)nps << 58);
+	addr_in.addr_type = TA_RAS_PA_TO_MCA;
+	ret = psp_ras_query_address(&adev->psp, &addr_in, &addr_out);
+	if (ret) {
+		dev_warn(adev->dev, "Failed to query RAS MCA address for 0x%llx",
+			pa);
+
+		return ret;
 	}
-	mutex_unlock(&ecc_log->lock);
 
-	return ret;
+	*mca = addr_out.ma.err_addr;
+
+	return 0;
 }