[PATCH 3/4] drm/amdgpu: add MCA smu cache support

Yang Wang kevinyang.wang at amd.com
Tue Apr 23 08:26:39 UTC 2024


v1:
because SMU CE valid mca bank will be cleared after reading,
this patch adds mca cache at the driver level to ensure that the mca bank is not lost.

v2:
refine amdgpu_mca_init/fini/reset() function name.

v3:
add mca_cache.lock support
only add CE bank to mca bank cache.

Signed-off-by: Yang Wang <kevinyang.wang at amd.com>
Reviewed-by: Tao Zhou <tao.zhou1 at amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mca.c | 95 ++++++++++++++++++++++++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_mca.h | 19 +++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c |  9 +++
 3 files changed, 116 insertions(+), 7 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mca.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_mca.c
index 859d594c02a5..264f56fd4f66 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mca.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mca.c
@@ -183,6 +183,29 @@ static int amdgpu_mca_bank_set_add_entry(struct mca_bank_set *mca_set, struct mc
 	return 0;
 }
 
+static int amdgpu_mca_bank_set_merge(struct mca_bank_set *mca_set, struct mca_bank_set *new)
+{
+	struct mca_bank_node *node;
+
+	list_for_each_entry(node, &new->list, node)
+		amdgpu_mca_bank_set_add_entry(mca_set, &node->entry);
+
+	return 0;
+}
+
+static int amdgpu_mca_bank_set_remove_node(struct mca_bank_set *mca_set, struct mca_bank_node *node)
+{
+	if (!node)
+		return -EINVAL;
+
+	list_del(&node->node);
+	kvfree(node);
+
+	mca_set->nr_entries--;
+
+	return 0;
+}
+
 static void amdgpu_mca_bank_set_release(struct mca_bank_set *mca_set)
 {
 	struct mca_bank_node *node, *tmp;
@@ -200,6 +223,41 @@ void amdgpu_mca_smu_init_funcs(struct amdgpu_device *adev, const struct amdgpu_m
 	mca->mca_funcs = mca_funcs;
 }
 
+int amdgpu_mca_init(struct amdgpu_device *adev)
+{
+	struct amdgpu_mca *mca = &adev->mca;
+	struct mca_bank_cache *mca_cache;
+	int i;
+
+	for (i = 0; i < ARRAY_SIZE(mca->mca_caches); i++) {
+		mca_cache = &mca->mca_caches[i];
+		mutex_init(&mca_cache->lock);
+		amdgpu_mca_bank_set_init(&mca_cache->mca_set);
+	}
+
+	return 0;
+}
+
+void amdgpu_mca_fini(struct amdgpu_device *adev)
+{
+	struct amdgpu_mca *mca = &adev->mca;
+	struct mca_bank_cache *mca_cache;
+	int i;
+
+	for (i = 0; i < ARRAY_SIZE(mca->mca_caches); i++) {
+		mca_cache = &mca->mca_caches[i];
+		amdgpu_mca_bank_set_release(&mca_cache->mca_set);
+		mutex_destroy(&mca_cache->lock);
+	}
+}
+
+int amdgpu_mca_reset(struct amdgpu_device *adev)
+{
+	amdgpu_mca_fini(adev);
+
+	return amdgpu_mca_init(adev);
+}
+
 int amdgpu_mca_smu_set_debug_mode(struct amdgpu_device *adev, bool enable)
 {
 	const struct amdgpu_mca_smu_funcs *mca_funcs = adev->mca.mca_funcs;
@@ -314,7 +372,7 @@ static int amdgpu_mca_dispatch_mca_set(struct amdgpu_device *adev, enum amdgpu_r
 {
 	struct ras_err_addr err_addr;
 	struct amdgpu_smuio_mcm_config_info mcm_info;
-	struct mca_bank_node *node;
+	struct mca_bank_node *node, *tmp;
 	struct mca_bank_entry *entry;
 	uint32_t count;
 	int ret;
@@ -325,7 +383,7 @@ static int amdgpu_mca_dispatch_mca_set(struct amdgpu_device *adev, enum amdgpu_r
 	if (!mca_set->nr_entries)
 		return 0;
 
-	list_for_each_entry(node, &mca_set->list, node) {
+	list_for_each_entry_safe(node, tmp, &mca_set->list, node) {
 		entry = &node->entry;
 
 		count = 0;
@@ -359,15 +417,30 @@ static int amdgpu_mca_dispatch_mca_set(struct amdgpu_device *adev, enum amdgpu_r
 				amdgpu_ras_error_statistic_ce_count(err_data,
 								    &mcm_info, &err_addr, (uint64_t)count);
 		}
+
+		amdgpu_mca_bank_set_remove_node(mca_set, node);
 	}
 
 	return 0;
 }
 
+static int amdgpu_mca_add_mca_set_to_cache(struct amdgpu_device *adev, enum amdgpu_mca_error_type type, struct mca_bank_set *new)
+{
+	struct mca_bank_cache *mca_cache = &adev->mca.mca_caches[type];
+	int ret;
+
+	mutex_lock(&mca_cache->lock);
+	ret = amdgpu_mca_bank_set_merge(&mca_cache->mca_set, new);
+	mutex_unlock(&mca_cache->lock);
+
+	return ret;
+}
+
 int amdgpu_mca_smu_log_ras_error(struct amdgpu_device *adev, enum amdgpu_ras_block blk, enum amdgpu_mca_error_type type,
 				 struct ras_err_data *err_data, struct ras_query_context *qctx)
 {
 	struct mca_bank_set mca_set;
+	struct mca_bank_cache *mca_cache = &adev->mca.mca_caches[type];
 	int ret;
 
 	amdgpu_mca_bank_set_init(&mca_set);
@@ -377,6 +450,21 @@ int amdgpu_mca_smu_log_ras_error(struct amdgpu_device *adev, enum amdgpu_ras_blo
 		goto out_mca_release;
 
 	ret = amdgpu_mca_dispatch_mca_set(adev, blk, type, &mca_set, err_data);
+	if (ret)
+		goto out_mca_release;
+
+	/* add remain mca bank to mca cache */
+	if (mca_set.nr_entries) {
+		ret = amdgpu_mca_add_mca_set_to_cache(adev, type, &mca_set);
+		if (ret)
+			goto out_mca_release;
+	}
+
+	/* dispatch mca set again if mca cache has valid data */
+	mutex_lock(&mca_cache->lock);
+	if (mca_cache->mca_set.nr_entries)
+		ret = amdgpu_mca_dispatch_mca_set(adev, blk, type, &mca_cache->mca_set, err_data);
+	mutex_unlock(&mca_cache->lock);
 
 out_mca_release:
 	amdgpu_mca_bank_set_release(&mca_set);
@@ -443,6 +531,9 @@ static int mca_dump_show(struct seq_file *m, enum amdgpu_mca_error_type type)
 	list_for_each_entry(node, &mca_set.list, node)
 		mca_dump_entry(m, &node->entry);
 
+	/* add mca bank to mca bank cache */
+	ret = amdgpu_mca_add_mca_set_to_cache(adev, type, &mca_set);
+
 err_free_mca_set:
 	amdgpu_mca_bank_set_release(&mca_set);
 
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mca.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_mca.h
index 4d0a0f91c375..9b97cfa28e05 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mca.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mca.h
@@ -77,11 +77,22 @@ struct amdgpu_mca_ras {
 	struct amdgpu_mca_ras_block *ras;
 };
 
+struct mca_bank_set {
+	int nr_entries;
+	struct list_head list;
+};
+
+struct mca_bank_cache {
+	struct mca_bank_set mca_set;
+	struct mutex lock;
+};
+
 struct amdgpu_mca {
 	struct amdgpu_mca_ras mp0;
 	struct amdgpu_mca_ras mp1;
 	struct amdgpu_mca_ras mpio;
 	const struct amdgpu_mca_smu_funcs *mca_funcs;
+	struct mca_bank_cache mca_caches[AMDGPU_MCA_ERROR_TYPE_DE];
 };
 
 enum mca_reg_idx {
@@ -113,11 +124,6 @@ struct mca_bank_node {
 	struct list_head node;
 };
 
-struct mca_bank_set {
-	int nr_entries;
-	struct list_head list;
-};
-
 struct amdgpu_mca_smu_funcs {
 	int max_ue_count;
 	int max_ce_count;
@@ -149,6 +155,9 @@ int amdgpu_mca_mp1_ras_sw_init(struct amdgpu_device *adev);
 int amdgpu_mca_mpio_ras_sw_init(struct amdgpu_device *adev);
 
 void amdgpu_mca_smu_init_funcs(struct amdgpu_device *adev, const struct amdgpu_mca_smu_funcs *mca_funcs);
+int amdgpu_mca_init(struct amdgpu_device *adev);
+void amdgpu_mca_fini(struct amdgpu_device *adev);
+int amdgpu_mca_reset(struct amdgpu_device *adev);
 int amdgpu_mca_smu_set_debug_mode(struct amdgpu_device *adev, bool enable);
 int amdgpu_mca_smu_get_mca_set_error_count(struct amdgpu_device *adev, enum amdgpu_ras_block blk,
 					   enum amdgpu_mca_error_type type, uint32_t *total);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c
index 46b7f0c5cd8a..23766bdffe5a 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c
@@ -3426,6 +3426,13 @@ int amdgpu_ras_late_init(struct amdgpu_device *adev)
 
 		amdgpu_ras_set_aca_debug_mode(adev, false);
 	} else {
+		if (amdgpu_in_reset(adev))
+			r = amdgpu_mca_reset(adev);
+		else
+			r = amdgpu_mca_init(adev);
+		if (r)
+			return r;
+
 		amdgpu_ras_set_mca_debug_mode(adev, false);
 	}
 
@@ -3498,6 +3505,8 @@ int amdgpu_ras_fini(struct amdgpu_device *adev)
 
 	if (amdgpu_aca_is_enabled(adev))
 		amdgpu_aca_fini(adev);
+	else
+		amdgpu_mca_fini(adev);
 
 	WARN(AMDGPU_RAS_GET_FEATURES(con->features), "Feature mask is not cleared");
 
-- 
2.34.1



More information about the amd-gfx mailing list