Skip to contentSkip to conversation

drm/amdgpu: Move xgmi ras initialization from .late_init to .early_init

PATCH In review yipechai 3 participants 7 messages Jan 20, 2022

Subsystems (2)
lore
Share & tools
lore

yipechai

author
#0 lore raw

PATCH v2 1/2 drm/amdgpu: Move xgmi ras initialization from .late_init to .early_init

Move xgmi ras initialization from .late_init to .early_init, which let xgmi ras can be initialized only once.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c | 15 ++++++++++-----
 drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h |  1 +
 drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c  |  5 +++++
 drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c   |  5 +++++
 4 files changed, 21 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c +10 -5
index 3483a82f5734..788c0257832d 100644
@@ -436,6 +436,16 @@ void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
 	} while (fault->timestamp < tmp);
 }
 
+int amdgpu_gmc_ras_early_init(struct amdgpu_device *adev)
+{
+	if (!adev->gmc.xgmi.connected_to_cpu) {
+		adev->gmc.xgmi.ras = &xgmi_ras;
+		amdgpu_ras_register_ras_block(adev, &adev->gmc.xgmi.ras->ras_block);
+	}
+
+	return 0;
+}
+
 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
 {
 	int r;
@@ -452,11 +462,6 @@ int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
 			return r;
 	}
 
-	if (!adev->gmc.xgmi.connected_to_cpu) {
-		adev->gmc.xgmi.ras = &xgmi_ras;
-		amdgpu_ras_register_ras_block(adev, &adev->gmc.xgmi.ras->ras_block);
-	}
-
 	if (adev->gmc.xgmi.ras && adev->gmc.xgmi.ras->ras_block.ras_late_init) {
 		r = adev->gmc.xgmi.ras->ras_block.ras_late_init(adev, NULL);
 		if (r)
drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h +1 -0
index 0001631cfedb..ac4c0e50b45c 100644
@@ -318,6 +318,7 @@ bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev,
 			      uint16_t pasid, uint64_t timestamp);
 void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
 				     uint16_t pasid);
+int amdgpu_gmc_ras_early_init(struct amdgpu_device *adev);
 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev);
 void amdgpu_gmc_ras_fini(struct amdgpu_device *adev);
 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev);
drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +5 -0
index 4f8d356f8432..7a6ad5d467b2 100644
@@ -719,6 +719,7 @@ static void gmc_v10_0_set_gfxhub_funcs(struct amdgpu_device *adev)
 
 static int gmc_v10_0_early_init(void *handle)
 {
+	int r;
 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 
 	gmc_v10_0_set_mmhub_funcs(adev);
@@ -734,6 +735,10 @@ static int gmc_v10_0_early_init(void *handle)
 	adev->gmc.private_aperture_end =
 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
 
+	r = amdgpu_gmc_ras_early_init(adev);
+	if (r)
+		return r;
+
 	return 0;
 }
 
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +5 -0
index c76ffd1a70cd..3cdd3d459d51 100644
@@ -1318,6 +1318,7 @@ static void gmc_v9_0_set_mca_funcs(struct amdgpu_device *adev)
 
 static int gmc_v9_0_early_init(void *handle)
 {
+	int r;
 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 
 	/* ARCT and VEGA20 don't have XGMI defined in their IP discovery tables */
@@ -1347,6 +1348,10 @@ static int gmc_v9_0_early_init(void *handle)
 	adev->gmc.private_aperture_end =
 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
 
+	r = amdgpu_gmc_ras_early_init(adev);
+	if (r)
+		return r;
+
 	return 0;
 }
 
Signature · 1 line
2.25.1
author
#1 lore raw

PATCH v2 2/2 Revert "drm/amdgpu: No longer insert ras blocks into ras_list if it already exists in ras_list"

This reverts commit 48e175f7476c6deb7ccf1f10d081322d52830a17.

Xgmi ras initialization had been moved from .late_init to early_init, the defect of repeated calling amdgpu_ras_register_ras_block had been fixed, so revert this patch.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c | 7 -------
 1 file changed, 7 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ras.c +0 -7
index 7a1d2bac698e..c92383fe7834 100644
@@ -2765,19 +2765,12 @@ int amdgpu_ras_reset_gpu(struct amdgpu_device *adev)
 int amdgpu_ras_register_ras_block(struct amdgpu_device *adev,
 		struct amdgpu_ras_block_object *ras_block_obj)
 {
-	struct amdgpu_ras_block_object *obj, *tmp;
 	if (!adev || !ras_block_obj)
 		return -EINVAL;
 
 	if (!amdgpu_ras_asic_supported(adev))
 		return 0;
 
-	/* If the ras object is in ras_list, don't add it again */
-	list_for_each_entry_safe(obj, tmp, &adev->ras_list, node) {
-		if (obj == ras_block_obj)
-			return 0;
-	}
-
 	INIT_LIST_HEAD(&ras_block_obj->node);
 	list_add_tail(&ras_block_obj->node, &adev->ras_list);
 
Signature · 1 line
2.25.1
Reviewed-by on v2 1/2
#2 lore raw
In reply to yipechai

[AMD Official Use Only]

The series is:

Quoted 123 lines

-----Original Message-----
From: Chai, Thomas <[email protected]>
Sent: Thursday, January 20, 2022 11:19 AM
To: [email protected]
Cc: Chai, Thomas <[email protected]>; Zhang, Hawking <[email protected]>; Zhou1, Tao <[email protected]>; Clements, John <[email protected]>; Chai, Thomas <[email protected]>
Subject: [PATCH V2 1/2] drm/amdgpu: Move xgmi ras initialization from .late_init to .early_init

Move xgmi ras initialization from .late_init to .early_init, which let xgmi ras can be initialized only once.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c | 15 ++++++++++-----
drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h |  1 +
drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c  |  5 +++++
 drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c   |  5 +++++
 4 files changed, 21 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c +9 -5
index 3483a82f5734..788c0257832d 100644
@@ -436,6 +436,16 @@ void amdgpu_gmc_filter_faults_remove(struct
 amdgpu_device *adev, uint64_t addr,
 	} while (fault->timestamp < tmp);
 }
 
+int amdgpu_gmc_ras_early_init(struct amdgpu_device *adev) {
+	if (!adev->gmc.xgmi.connected_to_cpu) {
+		adev->gmc.xgmi.ras = &xgmi_ras;
+		amdgpu_ras_register_ras_block(adev, &adev->gmc.xgmi.ras-
 >ras_block);
+	}
+
+	return 0;
+}
+
 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)  {
 	int r;
@@ -452,11 +462,6 @@ int amdgpu_gmc_ras_late_init(struct amdgpu_device
 *adev)
 			return r;
 	}
 
-	if (!adev->gmc.xgmi.connected_to_cpu) {
-		adev->gmc.xgmi.ras = &xgmi_ras;
-		amdgpu_ras_register_ras_block(adev, &adev->gmc.xgmi.ras-
 >ras_block);
-	}
-
 	if (adev->gmc.xgmi.ras && adev->gmc.xgmi.ras->ras_block.ras_late_init)

{

drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c +0 -0
 		r = adev->gmc.xgmi.ras->ras_block.ras_late_init(adev, NULL);
 		if (r)
drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h +1 -0
index 0001631cfedb..ac4c0e50b45c 100644
@@ -318,6 +318,7 @@ bool amdgpu_gmc_filter_faults(struct amdgpu_device
 *adev,
 			      uint16_t pasid, uint64_t timestamp);  void
 amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
 				     uint16_t pasid);
+int amdgpu_gmc_ras_early_init(struct amdgpu_device *adev);
 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev);  void

amdgpu_gmc_ras_fini(struct amdgpu_device *adev); int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev); diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
index 4f8d356f8432..7a6ad5d467b2 100644

drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +5 -0
@@ -719,6 +719,7 @@ static void gmc_v10_0_set_gfxhub_funcs(struct
 amdgpu_device *adev)
 
 static int gmc_v10_0_early_init(void *handle)  {
+	int r;
 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 
 	gmc_v10_0_set_mmhub_funcs(adev);
@@ -734,6 +735,10 @@ static int gmc_v10_0_early_init(void *handle)
 	adev->gmc.private_aperture_end =
 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
 
+	r = amdgpu_gmc_ras_early_init(adev);
+	if (r)
+		return r;
+
 	return 0;
 }
 
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +1 -0
index c76ffd1a70cd..3cdd3d459d51 100644
@@ -1318,6 +1318,7 @@ static void gmc_v9_0_set_mca_funcs(struct
 amdgpu_device *adev)
 
 static int gmc_v9_0_early_init(void *handle)  {
+	int r;
 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 
 	/* ARCT and VEGA20 don't have XGMI defined in their IP discovery

tables */ @@ -1347,6 +1348,10 @@ static int gmc_v9_0_early_init(void *handle)

drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +4 -1
 	adev->gmc.private_aperture_end =
 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
 
+	r = amdgpu_gmc_ras_early_init(adev);
+	if (r)
+		return r;
+
 	return 0;
 }
 
--

2.25.1

#3 lore raw
In reply to yipechai
Quoted 69 lines · yipechai

Move xgmi ras initialization from .late_init to .early_init, which let xgmi ras can be initialized only once.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c | 15 ++++++++++-----
  drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h |  1 +
  drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c  |  5 +++++
  drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c   |  5 +++++
  4 files changed, 21 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c +10 -5
index 3483a82f5734..788c0257832d 100644
@@ -436,6 +436,16 @@ void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
  	} while (fault->timestamp < tmp);
  }
 
+int amdgpu_gmc_ras_early_init(struct amdgpu_device *adev)
+{
+	if (!adev->gmc.xgmi.connected_to_cpu) {
+		adev->gmc.xgmi.ras = &xgmi_ras;
+		amdgpu_ras_register_ras_block(adev, &adev->gmc.xgmi.ras->ras_block);
+	}
+
+	return 0;
+}
+
  int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
  {
  	int r;
@@ -452,11 +462,6 @@ int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
  			return r;
  	}
 
-	if (!adev->gmc.xgmi.connected_to_cpu) {
-		adev->gmc.xgmi.ras = &xgmi_ras;
-		amdgpu_ras_register_ras_block(adev, &adev->gmc.xgmi.ras->ras_block);
-	}
-
  	if (adev->gmc.xgmi.ras && adev->gmc.xgmi.ras->ras_block.ras_late_init) {
  		r = adev->gmc.xgmi.ras->ras_block.ras_late_init(adev, NULL);
  		if (r)
drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h +1 -0
index 0001631cfedb..ac4c0e50b45c 100644
@@ -318,6 +318,7 @@ bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev,
  			      uint16_t pasid, uint64_t timestamp);
  void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
  				     uint16_t pasid);
+int amdgpu_gmc_ras_early_init(struct amdgpu_device *adev);
  int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev);
  void amdgpu_gmc_ras_fini(struct amdgpu_device *adev);
  int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev);
drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +1 -0
index 4f8d356f8432..7a6ad5d467b2 100644
@@ -719,6 +719,7 @@ static void gmc_v10_0_set_gfxhub_funcs(struct amdgpu_device *adev)
 
  static int gmc_v10_0_early_init(void *handle)
  {
+	int r;
  	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 
  	gmc_v10_0_set_mmhub_funcs(adev);

drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c static int gmc_v10_0_early_init(void *handle) · line 741

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +4 -0
@@ -734,6 +735,10 @@ static int gmc_v10_0_early_init(void *handle)
  	adev->gmc.private_aperture_end =
  		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
 
+	r = amdgpu_gmc_ras_early_init(adev);
+	if (r)
+		return r;
+

At this point it's unknown if RAS is applicable for the SKU. I think this failure check shouldn't be there (here and below one).

amdgpu_gmc_ras_early_init is return 0 always, that way also this check is not needed.

Thanks,
Lijo

Quoted 27 lines
drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +0 -0
  	return 0;
  }
 
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +5 -0
index c76ffd1a70cd..3cdd3d459d51 100644
@@ -1318,6 +1318,7 @@ static void gmc_v9_0_set_mca_funcs(struct amdgpu_device *adev)
 
  static int gmc_v9_0_early_init(void *handle)
  {
+	int r;
  	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 
  	/* ARCT and VEGA20 don't have XGMI defined in their IP discovery tables */
@@ -1347,6 +1348,10 @@ static int gmc_v9_0_early_init(void *handle)
  	adev->gmc.private_aperture_end =
  		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
 
+	r = amdgpu_gmc_ras_early_init(adev);
+	if (r)
+		return r;
+
  	return 0;
  }
 
#4 lore raw

-----Original Message-----
From: Lazar, Lijo <[email protected]>
Sent: Thursday, January 20, 2022 1:49 PM
To: Chai, Thomas <[email protected]>; [email protected]
Cc: Zhou1, Tao <[email protected]>; Zhang, Hawking <[email protected]>; Clements, John <[email protected]>; Chai, Thomas <[email protected]>
Subject: Re: [PATCH V2 1/2] drm/amdgpu: Move xgmi ras initialization from .late_init to .early_init

Quoted 82 lines · yipechai

Move xgmi ras initialization from .late_init to .early_init, which let xgmi ras can be initialized only once.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c | 15 ++++++++++-----
  drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h |  1 +
  drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c  |  5 +++++
  drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c   |  5 +++++
  4 files changed, 21 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c +9 -5
index 3483a82f5734..788c0257832d 100644
@@ -436,6 +436,16 @@ void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
  	} while (fault->timestamp < tmp);
  }
 
+int amdgpu_gmc_ras_early_init(struct amdgpu_device *adev) {
+	if (!adev->gmc.xgmi.connected_to_cpu) {
+		adev->gmc.xgmi.ras = &xgmi_ras;
+		amdgpu_ras_register_ras_block(adev, &adev->gmc.xgmi.ras->ras_block);
+	}
+
+	return 0;
+}
+
  int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
  {
  	int r;
@@ -452,11 +462,6 @@ int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
  			return r;
  	}
 
-	if (!adev->gmc.xgmi.connected_to_cpu) {
-		adev->gmc.xgmi.ras = &xgmi_ras;
-		amdgpu_ras_register_ras_block(adev, &adev->gmc.xgmi.ras->ras_block);
-	}
-
  	if (adev->gmc.xgmi.ras && adev->gmc.xgmi.ras->ras_block.ras_late_init) {
  		r = adev->gmc.xgmi.ras->ras_block.ras_late_init(adev, NULL);
  		if (r)
drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h +1 -1
index 0001631cfedb..ac4c0e50b45c 100644
@@ -318,6 +318,7 @@ bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev,
  			      uint16_t pasid, uint64_t timestamp);
  void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
  				     uint16_t pasid);
+int amdgpu_gmc_ras_early_init(struct amdgpu_device *adev);
  int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev);
  void amdgpu_gmc_ras_fini(struct amdgpu_device *adev);
  int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev); diff 
--git a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c 

b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
index 4f8d356f8432..7a6ad5d467b2 100644

drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +1 -0
@@ -719,6 +719,7 @@ static void gmc_v10_0_set_gfxhub_funcs(struct 
 amdgpu_device *adev)
 
  static int gmc_v10_0_early_init(void *handle)
  {
+	int r;
  	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 
  	gmc_v10_0_set_mmhub_funcs(adev);
drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +4 -0
@@ -734,6 +735,10 @@ static int gmc_v10_0_early_init(void *handle)
  	adev->gmc.private_aperture_end =
  		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
 
+	r = amdgpu_gmc_ras_early_init(adev);
+	if (r)
+		return r;
+
 

At this point it's unknown if RAS is applicable for the SKU. I think this failure check shouldn't be there (here and below one).

Quoted 1 line

amdgpu_gmc_ras_early_init is return 0 always, that way also this check is not needed.

[Thomas] Just like calling amdgpu_gmc_ras_late_init, checking the return status may make the code extensible.
In amdgpu_gmc_ras_early_init, the xgmi ras initialization may always return 0, but it may add functions that need to check the return status in future.

Thanks,
Lijo

Quoted 29 lines
  	return 0;
  }
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +1 -0
index c76ffd1a70cd..3cdd3d459d51 100644
@@ -1318,6 +1318,7 @@ static void gmc_v9_0_set_mca_funcs(struct 
 amdgpu_device *adev)
 
  static int gmc_v9_0_early_init(void *handle)
  {
+	int r;
  	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 

/* ARCT and VEGA20 don't have XGMI defined in their IP discovery
tables */ @@ -1347,6 +1348,10 @@ static int gmc_v9_0_early_init(void *handle)
adev->gmc.private_aperture_end =
adev->gmc.private_aperture_start + (4ULL << 30) - 1;

drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +4 -0
+	r = amdgpu_gmc_ras_early_init(adev);
+	if (r)
+		return r;
+
  	return 0;
  }
#5 lore raw
Quoted 91 lines · Thomas

-----Original Message-----
From: Lazar, Lijo <[email protected]>
Sent: Thursday, January 20, 2022 1:49 PM
To: Chai, Thomas <[email protected]>; [email protected]
Cc: Zhou1, Tao <[email protected]>; Zhang, Hawking <[email protected]>; Clements, John <[email protected]>; Chai, Thomas <[email protected]>
Subject: Re: [PATCH V2 1/2] drm/amdgpu: Move xgmi ras initialization from .late_init to .early_init

Quoted 79 lines · yipechai

Move xgmi ras initialization from .late_init to .early_init, which let xgmi ras can be initialized only once.

---
   drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c | 15 ++++++++++-----
   drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h |  1 +
   drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c  |  5 +++++
   drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c   |  5 +++++
   4 files changed, 21 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c +9 -5
index 3483a82f5734..788c0257832d 100644
@@ -436,6 +436,16 @@ void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
   	} while (fault->timestamp < tmp);
   }
 
+int amdgpu_gmc_ras_early_init(struct amdgpu_device *adev) {
+	if (!adev->gmc.xgmi.connected_to_cpu) {
+		adev->gmc.xgmi.ras = &xgmi_ras;
+		amdgpu_ras_register_ras_block(adev, &adev->gmc.xgmi.ras->ras_block);
+	}
+
+	return 0;
+}
+
   int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
   {
   	int r;
@@ -452,11 +462,6 @@ int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
   			return r;
   	}
 
-	if (!adev->gmc.xgmi.connected_to_cpu) {
-		adev->gmc.xgmi.ras = &xgmi_ras;
-		amdgpu_ras_register_ras_block(adev, &adev->gmc.xgmi.ras->ras_block);
-	}
-
   	if (adev->gmc.xgmi.ras && adev->gmc.xgmi.ras->ras_block.ras_late_init) {
   		r = adev->gmc.xgmi.ras->ras_block.ras_late_init(adev, NULL);
   		if (r)
drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h +1 -1
index 0001631cfedb..ac4c0e50b45c 100644
@@ -318,6 +318,7 @@ bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev,
   			      uint16_t pasid, uint64_t timestamp);
   void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
   				     uint16_t pasid);
+int amdgpu_gmc_ras_early_init(struct amdgpu_device *adev);
   int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev);
   void amdgpu_gmc_ras_fini(struct amdgpu_device *adev);
   int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev); diff
--git a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c

b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
index 4f8d356f8432..7a6ad5d467b2 100644

drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +1 -0
@@ -719,6 +719,7 @@ static void gmc_v10_0_set_gfxhub_funcs(struct
 amdgpu_device *adev)
 
   static int gmc_v10_0_early_init(void *handle)
   {
+	int r;
   	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 
   	gmc_v10_0_set_mmhub_funcs(adev);
drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +4 -0
@@ -734,6 +735,10 @@ static int gmc_v10_0_early_init(void *handle)
   	adev->gmc.private_aperture_end =
   		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
 
+	r = amdgpu_gmc_ras_early_init(adev);
+	if (r)
+		return r;
+
Quoted 7 lines
Quoted 3 lines

At this point it's unknown if RAS is applicable for the SKU. I think this failure check shouldn't be there (here and below one).

amdgpu_gmc_ras_early_init is return 0 always, that way also this check is not needed.

[Thomas] Just like calling amdgpu_gmc_ras_late_init, checking the return status may make the code extensible.
In amdgpu_gmc_ras_early_init, the xgmi ras initialization may always return 0, but it may add functions that need to check the return status in future.

At this point, it's unknown

1) If the device is part of XGMI hive or not.
2) If the device supports RAS.

For such devices, it doesn't make any sense to fail here based on this function.

Quoted 32 lines

Thanks,
Lijo

Quoted 29 lines
   	return 0;
   }
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +1 -0
index c76ffd1a70cd..3cdd3d459d51 100644
@@ -1318,6 +1318,7 @@ static void gmc_v9_0_set_mca_funcs(struct
 amdgpu_device *adev)
 
   static int gmc_v9_0_early_init(void *handle)
   {
+	int r;
   	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 
   	/* ARCT and VEGA20 don't have XGMI defined in their IP discovery

tables */ @@ -1347,6 +1348,10 @@ static int gmc_v9_0_early_init(void *handle)
adev->gmc.private_aperture_end =
adev->gmc.private_aperture_start + (4ULL << 30) - 1;

drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +4 -0
+	r = amdgpu_gmc_ras_early_init(adev);
+	if (r)
+		return r;
+
   	return 0;
   }
#6 lore raw

-----Original Message-----
From: Lazar, Lijo <[email protected]>
Sent: Thursday, January 20, 2022 3:32 PM
To: Chai, Thomas <[email protected]>; [email protected]
Cc: Zhou1, Tao <[email protected]>; Zhang, Hawking <[email protected]>; Clements, John <[email protected]>
Subject: Re: [PATCH V2 1/2] drm/amdgpu: Move xgmi ras initialization from .late_init to .early_init

Quoted 102 lines · Thomas

-----Original Message-----
From: Lazar, Lijo <[email protected]>
Sent: Thursday, January 20, 2022 1:49 PM
To: Chai, Thomas <[email protected]>; [email protected]
Cc: Zhou1, Tao <[email protected]>; Zhang, Hawking <[email protected]>; Clements, John <[email protected]>; Chai, Thomas <[email protected]>
Subject: Re: [PATCH V2 1/2] drm/amdgpu: Move xgmi ras initialization from .late_init to .early_init

Quoted 82 lines · yipechai

Move xgmi ras initialization from .late_init to .early_init, which let xgmi ras can be initialized only once.

---
   drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c | 15 ++++++++++-----
   drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h |  1 +
   drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c  |  5 +++++
   drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c   |  5 +++++
   4 files changed, 21 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c +9 -5
index 3483a82f5734..788c0257832d 100644
@@ -436,6 +436,16 @@ void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
   	} while (fault->timestamp < tmp);
   }
 
+int amdgpu_gmc_ras_early_init(struct amdgpu_device *adev) {
+	if (!adev->gmc.xgmi.connected_to_cpu) {
+		adev->gmc.xgmi.ras = &xgmi_ras;
+		amdgpu_ras_register_ras_block(adev, &adev->gmc.xgmi.ras->ras_block);
+	}
+
+	return 0;
+}
+
   int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
   {
   	int r;
@@ -452,11 +462,6 @@ int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
   			return r;
   	}
 
-	if (!adev->gmc.xgmi.connected_to_cpu) {
-		adev->gmc.xgmi.ras = &xgmi_ras;
-		amdgpu_ras_register_ras_block(adev, &adev->gmc.xgmi.ras->ras_block);
-	}
-
   	if (adev->gmc.xgmi.ras && adev->gmc.xgmi.ras->ras_block.ras_late_init) {
   		r = adev->gmc.xgmi.ras->ras_block.ras_late_init(adev, NULL);
   		if (r)
drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h +1 -0
index 0001631cfedb..ac4c0e50b45c 100644
@@ -318,6 +318,7 @@ bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev,
   			      uint16_t pasid, uint64_t timestamp);
   void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
   				     uint16_t pasid);
+int amdgpu_gmc_ras_early_init(struct amdgpu_device *adev);
   int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev);
   void amdgpu_gmc_ras_fini(struct amdgpu_device *adev);
   int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev); 
drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +1 -0
index 4f8d356f8432..7a6ad5d467b2 100644
@@ -719,6 +719,7 @@ static void gmc_v10_0_set_gfxhub_funcs(struct 
 amdgpu_device *adev)
 
   static int gmc_v10_0_early_init(void *handle)
   {
+	int r;
   	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 
   	gmc_v10_0_set_mmhub_funcs(adev);
drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +4 -0
@@ -734,6 +735,10 @@ static int gmc_v10_0_early_init(void *handle)
   	adev->gmc.private_aperture_end =
   		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
 
+	r = amdgpu_gmc_ras_early_init(adev);
+	if (r)
+		return r;
+
 

At this point it's unknown if RAS is applicable for the SKU. I think this failure check shouldn't be there (here and below one).

Quoted 1 line

amdgpu_gmc_ras_early_init is return 0 always, that way also this check is not needed.

[Thomas] Just like calling amdgpu_gmc_ras_late_init, checking the return status may make the code extensible.
In amdgpu_gmc_ras_early_init, the xgmi ras initialization may always return 0, but it may add functions that need to check the return status in future.

Quoted 6 lines

At this point, it's unknown

1) If the device is part of XGMI hive or not.
2) If the device supports RAS.

For such devices, it doesn't make any sense to fail here based on this function.

[Thomas] The current code in amdgpu_gmc_ras_early_init has no effect for the devices, whether the device has xgmi hive or not, whether it support RAS or not.
Checking the return status is just for amdgpu_gmc_ras_early_init to make it easy to add future functions that need to check the return status.

Quoted 32 lines

Thanks,
Lijo

Quoted 29 lines
   	return 0;
   }
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +1 -0
index c76ffd1a70cd..3cdd3d459d51 100644
@@ -1318,6 +1318,7 @@ static void gmc_v9_0_set_mca_funcs(struct 
 amdgpu_device *adev)
 
   static int gmc_v9_0_early_init(void *handle)
   {
+	int r;
   	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 
   	/* ARCT and VEGA20 don't have XGMI defined in their IP discovery 

tables */ @@ -1347,6 +1348,10 @@ static int gmc_v9_0_early_init(void *handle)
adev->gmc.private_aperture_end =
adev->gmc.private_aperture_start + (4ULL << 30) - 1;

drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +4 -0
+	r = amdgpu_gmc_ras_early_init(adev);
+	if (r)
+		return r;
+
   	return 0;
   }