drm/amdgpu: umc v12_0 logs ecc errors

1. umc v12_0 logs ecc errors.
2. Reserve newly detected ecc error pages.
3. Add tag for bad pages, so that they can
   be retired later.

Signed-off-by: YiPeng Chai <[email protected]>
Reviewed-by: Tao Zhou <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

Author: YiPeng Chai

drivers/gpu/drm/amd/amdgpu/amdgpu_umc.c

@@ -21,10 +21,13 @@
  *
  */
 
+#include <linux/sort.h>
 #include "amdgpu.h"
 #include "umc_v6_7.h"
 #define MAX_UMC_POISON_POLLING_TIME_SYNC   20  //ms
 
+#define MAX_UMC_HASH_STRING_SIZE  256
+
 static int amdgpu_umc_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)
@@ -446,3 +449,67 @@ int amdgpu_umc_update_ecc_status(struct amdgpu_device *adev,
 					status, ipid, addr);
 	return 0;
 }
+
+static int amdgpu_umc_uint64_cmp(const void *a, const void *b)
+{
+	uint64_t *addr_a = (uint64_t *)a;
+	uint64_t *addr_b = (uint64_t *)b;
+
+	if (*addr_a > *addr_b)
+		return 1;
+	else if (*addr_a < *addr_b)
+		return -1;
+	else
+		return 0;
+}
+
+/* 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)
+{
+	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;
+
+	if (!pfns || !len)
+		return -EINVAL;
+
+	sort(pfns, len, sizeof(uint64_t), amdgpu_umc_uint64_cmp, NULL);
+
+	for (i = 0; i < len; i++)
+		offset += snprintf(&buf[offset], sizeof(buf) - offset, "%llx", pfns[i]);
+
+	hash_val = siphash(buf, offset, &con->umc_ecc_log.ecc_key);
+
+	*val = hash_val;
+
+	return 0;
+}
+
+int amdgpu_umc_logs_ecc_err(struct amdgpu_device *adev,
+		struct radix_tree_root *ecc_tree, struct ras_ecc_err *ecc_err)
+{
+	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
+	struct ras_ecc_log_info *ecc_log;
+	int ret;
+
+	ecc_log = &con->umc_ecc_log;
+
+	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;
+
+		/* Reserve memory */
+		for (i = 0; i < err_pages->count; i++)
+			amdgpu_ras_reserve_page(adev, err_pages->pfn[i]);
+
+		radix_tree_tag_set(ecc_tree,
+			ecc_err->hash_index, UMC_ECC_NEW_DETECTED_TAG);
+	}
+	mutex_unlock(&ecc_log->lock);
+
+	return ret;
+}

drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h

@@ -52,6 +52,8 @@
 #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
 
 typedef int (*umc_func)(struct amdgpu_device *adev, uint32_t node_inst,
 			uint32_t umc_inst, uint32_t ch_inst, void *data);
@@ -127,5 +129,8 @@ int amdgpu_umc_bad_page_polling_timeout(struct amdgpu_device *adev,
 
 int amdgpu_umc_update_ecc_status(struct amdgpu_device *adev,
 				uint64_t status, uint64_t ipid, uint64_t addr);
-
+int amdgpu_umc_build_pages_hash(struct amdgpu_device *adev,
+		uint64_t *pfns, int len, uint64_t *val);
+int amdgpu_umc_logs_ecc_err(struct amdgpu_device *adev,
+		struct radix_tree_root *ecc_tree, struct ras_ecc_err *ecc_err);
 #endif

drivers/gpu/drm/amd/amdgpu/umc_v12_0.c

@@ -546,8 +546,10 @@ static int umc_v12_0_update_ecc_status(struct amdgpu_device *adev,
 	uint16_t hwid, mcatype;
 	struct ta_ras_query_address_input addr_in;
 	uint64_t page_pfn[UMC_V12_0_BAD_PAGE_NUM_PER_CHANNEL];
-	uint64_t err_addr;
+	uint64_t err_addr, hash_val = 0;
+	struct ras_ecc_err *ecc_err;
 	int count;
+	int ret;
 
 	hwid = REG_GET_FIELD(ipid, MCMP1_IPIDT0, HardwareID);
 	mcatype = REG_GET_FIELD(ipid, MCMP1_IPIDT0, McaType);
@@ -589,6 +591,43 @@ static int umc_v12_0_update_ecc_status(struct amdgpu_device *adev,
 		return 0;
 	}
 
+	ret = amdgpu_umc_build_pages_hash(adev,
+			page_pfn, count, &hash_val);
+	if (ret) {
+		dev_err(adev->dev, "Fail to build error pages hash\n");
+		return ret;
+	}
+
+	ecc_err = kzalloc(sizeof(*ecc_err), GFP_KERNEL);
+	if (!ecc_err)
+		return -ENOMEM;
+
+	ecc_err->err_pages.pfn = kcalloc(count, sizeof(*ecc_err->err_pages.pfn), GFP_KERNEL);
+	if (!ecc_err->err_pages.pfn) {
+		kfree(ecc_err);
+		return -ENOMEM;
+	}
+
+	memcpy(ecc_err->err_pages.pfn, page_pfn, count * sizeof(*ecc_err->err_pages.pfn));
+	ecc_err->err_pages.count = count;
+
+	ecc_err->hash_index = hash_val;
+	ecc_err->status = status;
+	ecc_err->ipid = ipid;
+	ecc_err->addr = addr;
+
+	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_updated = true;
+		else
+			dev_err(adev->dev, "Fail to log ecc error! ret:%d\n", ret);
+
+		kfree(ecc_err->err_pages.pfn);
+		kfree(ecc_err);
+		return ret;
+	}
+
 	con->umc_ecc_log.de_updated = true;
 
 	return 0;