Skip to contentSkip to conversation

drm/amdkfd: fix debug watchpoints for logical devices

PATCH In review Jonathan Kim 2 participants 4 messages Jul 22, 2024 – Jul 31, 2024

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/amdkfd: fix debug watchpoints for logical devices

The number of watchpoints should be set and constrained per logical partition device, not by the socket device.

---
 drivers/gpu/drm/amd/amdkfd/kfd_debug.c  | 20 ++++++++++----------
 drivers/gpu/drm/amd/amdkfd/kfd_device.c |  4 ++--
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h   |  8 ++++----
 3 files changed, 16 insertions(+), 16 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +10 -10
index 34a282540c7e..312dfa84f29f 100644
@@ -365,47 +365,47 @@ static int kfd_dbg_get_dev_watch_id(struct kfd_process_device *pdd, int *watch_i
 
 	*watch_id = KFD_DEBUGGER_INVALID_WATCH_POINT_ID;
 
-	spin_lock(&pdd->dev->kfd->watch_points_lock);
+	spin_lock(&pdd->dev->watch_points_lock);
 
 	for (i = 0; i < MAX_WATCH_ADDRESSES; i++) {
 		/* device watchpoint in use so skip */
-		if ((pdd->dev->kfd->alloc_watch_ids >> i) & 0x1)
+		if ((pdd->dev->alloc_watch_ids >> i) & 0x1)
 			continue;
 
 		pdd->alloc_watch_ids |= 0x1 << i;
-		pdd->dev->kfd->alloc_watch_ids |= 0x1 << i;
+		pdd->dev->alloc_watch_ids |= 0x1 << i;
 		*watch_id = i;
-		spin_unlock(&pdd->dev->kfd->watch_points_lock);
+		spin_unlock(&pdd->dev->watch_points_lock);
 		return 0;
 	}
 
-	spin_unlock(&pdd->dev->kfd->watch_points_lock);
+	spin_unlock(&pdd->dev->watch_points_lock);
 
 	return -ENOMEM;
 }
 
 static void kfd_dbg_clear_dev_watch_id(struct kfd_process_device *pdd, int watch_id)
 {
-	spin_lock(&pdd->dev->kfd->watch_points_lock);
+	spin_lock(&pdd->dev->watch_points_lock);
 
 	/* process owns device watch point so safe to clear */
 	if ((pdd->alloc_watch_ids >> watch_id) & 0x1) {
 		pdd->alloc_watch_ids &= ~(0x1 << watch_id);
-		pdd->dev->kfd->alloc_watch_ids &= ~(0x1 << watch_id);
+		pdd->dev->alloc_watch_ids &= ~(0x1 << watch_id);
 	}
 
-	spin_unlock(&pdd->dev->kfd->watch_points_lock);
+	spin_unlock(&pdd->dev->watch_points_lock);
 }
 
 static bool kfd_dbg_owns_dev_watch_id(struct kfd_process_device *pdd, int watch_id)
 {
 	bool owns_watch_id = false;
 
-	spin_lock(&pdd->dev->kfd->watch_points_lock);
+	spin_lock(&pdd->dev->watch_points_lock);
 	owns_watch_id = watch_id < MAX_WATCH_ADDRESSES &&
 			((pdd->alloc_watch_ids >> watch_id) & 0x1);
 
-	spin_unlock(&pdd->dev->kfd->watch_points_lock);
+	spin_unlock(&pdd->dev->watch_points_lock);
 
 	return owns_watch_id;
 }
drivers/gpu/drm/amd/amdkfd/kfd_device.c +2 -2
index f4d20adaa068..f91a9b6ce3fb 100644
@@ -885,12 +885,12 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd,
 			goto node_init_error;
 		}
 		kfd->nodes[i] = node;
+
+		spin_lock_init(&kfd->nodes[i]->watch_points_lock);
 	}
 
 	svm_range_set_max_pages(kfd->adev);
 
-	spin_lock_init(&kfd->watch_points_lock);
-
 	kfd->init_complete = true;
 	dev_info(kfd_device, "added device %x:%x\n", kfd->adev->pdev->vendor,
 		 kfd->adev->pdev->device);
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +4 -4
index 2b3ec92981e8..653e1f934107 100644
@@ -310,6 +310,10 @@ struct kfd_node {
 	struct kfd_local_mem_info local_mem_info;
 
 	struct kfd_dev *kfd;
+
+	/* Track per device allocated watch points */
+	uint32_t alloc_watch_ids;
+	spinlock_t watch_points_lock;
 };
 
 struct kfd_dev {
@@ -362,10 +366,6 @@ struct kfd_dev {
 	struct kfd_node *nodes[MAX_KFD_NODES];
 	unsigned int num_nodes;
 
-	/* Track per device allocated watch points */
-	uint32_t alloc_watch_ids;
-	spinlock_t watch_points_lock;
-
 	/* Kernel doorbells for KFD device */
 	struct amdgpu_bo *doorbells;
 
Signature · 1 line
2.34.1
#1 lore raw

[Public]

Ping on review.

Thanks,

Jon

Quoted 134 lines

-----Original Message-----
From: Kim, Jonathan <[email protected]>
Sent: Monday, July 22, 2024 2:57 PM
To: [email protected]
Cc: Kim, Jonathan <[email protected]>; Kim, Jonathan <[email protected]>
Subject: [PATCH] drm/amdkfd: fix debug watchpoints for logical devices

The number of watchpoints should be set and constrained per logical partition device, not by the socket device.

---
 drivers/gpu/drm/amd/amdkfd/kfd_debug.c  | 20 ++++++++++----------
 drivers/gpu/drm/amd/amdkfd/kfd_device.c |  4 ++--
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h   |  8 ++++----
 3 files changed, 16 insertions(+), 16 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +0 -0
index 34a282540c7e..312dfa84f29f 100644
@@ -365,47 +365,47 @@ static int kfd_dbg_get_dev_watch_id(struct

kfd_process_device *pdd, int *watch_i

      *watch_id = KFD_DEBUGGER_INVALID_WATCH_POINT_ID;
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +5 -5
-     spin_lock(&pdd->dev->kfd->watch_points_lock);
+     spin_lock(&pdd->dev->watch_points_lock);
 
      for (i = 0; i < MAX_WATCH_ADDRESSES; i++) {
              /* device watchpoint in use so skip */
-             if ((pdd->dev->kfd->alloc_watch_ids >> i) & 0x1)
+             if ((pdd->dev->alloc_watch_ids >> i) & 0x1)
                      continue;
 
              pdd->alloc_watch_ids |= 0x1 << i;
-             pdd->dev->kfd->alloc_watch_ids |= 0x1 << i;
+             pdd->dev->alloc_watch_ids |= 0x1 << i;
              *watch_id = i;
-             spin_unlock(&pdd->dev->kfd->watch_points_lock);
+             spin_unlock(&pdd->dev->watch_points_lock);
              return 0;
      }
 
-     spin_unlock(&pdd->dev->kfd->watch_points_lock);
+     spin_unlock(&pdd->dev->watch_points_lock);
 
      return -ENOMEM;
 }
 
 static void kfd_dbg_clear_dev_watch_id(struct kfd_process_device *pdd, int

watch_id)
{

drivers/gpu/drm/amd/amdkfd/kfd_debug.c +3 -3
-     spin_lock(&pdd->dev->kfd->watch_points_lock);
+     spin_lock(&pdd->dev->watch_points_lock);
 
      /* process owns device watch point so safe to clear */
      if ((pdd->alloc_watch_ids >> watch_id) & 0x1) {
              pdd->alloc_watch_ids &= ~(0x1 << watch_id);
-             pdd->dev->kfd->alloc_watch_ids &= ~(0x1 << watch_id);
+             pdd->dev->alloc_watch_ids &= ~(0x1 << watch_id);
      }
 
-     spin_unlock(&pdd->dev->kfd->watch_points_lock);
+     spin_unlock(&pdd->dev->watch_points_lock);
 }
 
 static bool kfd_dbg_owns_dev_watch_id(struct kfd_process_device *pdd, int

watch_id)

 {
      bool owns_watch_id = false;
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +2 -2
-     spin_lock(&pdd->dev->kfd->watch_points_lock);
+     spin_lock(&pdd->dev->watch_points_lock);
      owns_watch_id = watch_id < MAX_WATCH_ADDRESSES &&
                      ((pdd->alloc_watch_ids >> watch_id) & 0x1);
 
-     spin_unlock(&pdd->dev->kfd->watch_points_lock);
+     spin_unlock(&pdd->dev->watch_points_lock);
 
      return owns_watch_id;
 }
drivers/gpu/drm/amd/amdkfd/kfd_device.c +2 -2
index f4d20adaa068..f91a9b6ce3fb 100644
@@ -885,12 +885,12 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd,
                      goto node_init_error;
              }
              kfd->nodes[i] = node;
+
+             spin_lock_init(&kfd->nodes[i]->watch_points_lock);
      }
 
      svm_range_set_max_pages(kfd->adev);
 
-     spin_lock_init(&kfd->watch_points_lock);
-
      kfd->init_complete = true;
      dev_info(kfd_device, "added device %x:%x\n", kfd->adev->pdev-
 >vendor,
               kfd->adev->pdev->device);
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +4 -5
index 2b3ec92981e8..653e1f934107 100644
@@ -310,6 +310,10 @@ struct kfd_node {
      struct kfd_local_mem_info local_mem_info;
 
      struct kfd_dev *kfd;
+
+     /* Track per device allocated watch points */
+     uint32_t alloc_watch_ids;
+     spinlock_t watch_points_lock;
 };
 
 struct kfd_dev {
@@ -362,10 +366,6 @@ struct kfd_dev {
      struct kfd_node *nodes[MAX_KFD_NODES];
      unsigned int num_nodes;
 
-     /* Track per device allocated watch points */
-     uint32_t alloc_watch_ids;
-     spinlock_t watch_points_lock;
-
      /* Kernel doorbells for KFD device */
      struct amdgpu_bo *doorbells;
 
--

2.34.1

#2 lore raw
In reply to Jonathan Kim

[AMD Official Use Only - AMD Internal Distribution Only]

One minor comment b. Either way Reviewed-by: Harish Kasiviswanathan <[email protected]>

-----Original Message-----
From: amd-gfx <[email protected]> On Behalf Of Jonathan Kim
Sent: Monday, July 22, 2024 2:57 PM
To: [email protected]
Cc: Kim, Jonathan <[email protected]>; Kim, Jonathan <[email protected]>
Subject: [PATCH] drm/amdkfd: fix debug watchpoints for logical devices

The number of watchpoints should be set and constrained per logical partition device, not by the socket device.

---
 drivers/gpu/drm/amd/amdkfd/kfd_debug.c  | 20 ++++++++++----------
 drivers/gpu/drm/amd/amdkfd/kfd_device.c |  4 ++--
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h   |  8 ++++----
 3 files changed, 16 insertions(+), 16 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +10 -10
index 34a282540c7e..312dfa84f29f 100644
@@ -365,47 +365,47 @@ static int kfd_dbg_get_dev_watch_id(struct kfd_process_device *pdd, int *watch_i
 
        *watch_id = KFD_DEBUGGER_INVALID_WATCH_POINT_ID;
 
-       spin_lock(&pdd->dev->kfd->watch_points_lock);
+       spin_lock(&pdd->dev->watch_points_lock);
 
        for (i = 0; i < MAX_WATCH_ADDRESSES; i++) {
                /* device watchpoint in use so skip */
-               if ((pdd->dev->kfd->alloc_watch_ids >> i) & 0x1)
+               if ((pdd->dev->alloc_watch_ids >> i) & 0x1)
                        continue;
 
                pdd->alloc_watch_ids |= 0x1 << i;
-               pdd->dev->kfd->alloc_watch_ids |= 0x1 << i;
+               pdd->dev->alloc_watch_ids |= 0x1 << i;
                *watch_id = i;
-               spin_unlock(&pdd->dev->kfd->watch_points_lock);
+               spin_unlock(&pdd->dev->watch_points_lock);
                return 0;
        }
 
-       spin_unlock(&pdd->dev->kfd->watch_points_lock);
+       spin_unlock(&pdd->dev->watch_points_lock);
 
        return -ENOMEM;
 }
 
 static void kfd_dbg_clear_dev_watch_id(struct kfd_process_device *pdd, int watch_id)
 {
-       spin_lock(&pdd->dev->kfd->watch_points_lock);
+       spin_lock(&pdd->dev->watch_points_lock);
 
        /* process owns device watch point so safe to clear */
        if ((pdd->alloc_watch_ids >> watch_id) & 0x1) {
                pdd->alloc_watch_ids &= ~(0x1 << watch_id);
-               pdd->dev->kfd->alloc_watch_ids &= ~(0x1 << watch_id);
+               pdd->dev->alloc_watch_ids &= ~(0x1 << watch_id);
        }
 
-       spin_unlock(&pdd->dev->kfd->watch_points_lock);
+       spin_unlock(&pdd->dev->watch_points_lock);
 }
 
 static bool kfd_dbg_owns_dev_watch_id(struct kfd_process_device *pdd, int watch_id)
 {
        bool owns_watch_id = false;
 
-       spin_lock(&pdd->dev->kfd->watch_points_lock);
+       spin_lock(&pdd->dev->watch_points_lock);
        owns_watch_id = watch_id < MAX_WATCH_ADDRESSES &&
                        ((pdd->alloc_watch_ids >> watch_id) & 0x1);
 
-       spin_unlock(&pdd->dev->kfd->watch_points_lock);
+       spin_unlock(&pdd->dev->watch_points_lock);
 
        return owns_watch_id;
 }
drivers/gpu/drm/amd/amdkfd/kfd_device.c +0 -0
index f4d20adaa068..f91a9b6ce3fb 100644
@@ -885,12 +885,12 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd,
                        goto node_init_error;
                }
 

Wouldn't it better to initialize node->watch_points_lock here. spin_lock_init(&node->watch_points_lock);

                kfd->nodes[i] = node;
+
+               spin_lock_init(&kfd->nodes[i]->watch_points_lock);
        }

        svm_range_set_max_pages(kfd->adev);

- spin_lock_init(&kfd->watch_points_lock);
-

        kfd->init_complete = true;
        dev_info(kfd_device, "added device %x:%x\n", kfd->adev->pdev->vendor,
                 kfd->adev->pdev->device);
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +4 -4
index 2b3ec92981e8..653e1f934107 100644
@@ -310,6 +310,10 @@ struct kfd_node {
        struct kfd_local_mem_info local_mem_info;
 
        struct kfd_dev *kfd;
+
+       /* Track per device allocated watch points */
+       uint32_t alloc_watch_ids;
+       spinlock_t watch_points_lock;
 };
 
 struct kfd_dev {
@@ -362,10 +366,6 @@ struct kfd_dev {
        struct kfd_node *nodes[MAX_KFD_NODES];
        unsigned int num_nodes;
 
-       /* Track per device allocated watch points */
-       uint32_t alloc_watch_ids;
-       spinlock_t watch_points_lock;
-
        /* Kernel doorbells for KFD device */
        struct amdgpu_bo *doorbells;
 

--
2.34.1

#3 lore raw

[Public]

Quoted 107 lines

-----Original Message-----
From: Kasiviswanathan, Harish <[email protected]>
Sent: Tuesday, July 30, 2024 12:23 PM
To: Kim, Jonathan <[email protected]>; [email protected]
Cc: Kim, Jonathan <[email protected]>; Kim, Jonathan <[email protected]>
Subject: RE: [PATCH] drm/amdkfd: fix debug watchpoints for logical devices

[AMD Official Use Only - AMD Internal Distribution Only]

One minor comment b. Either way Reviewed-by: Harish Kasiviswanathan <[email protected]>

-----Original Message-----
From: amd-gfx <[email protected]> On Behalf Of Jonathan Kim
Sent: Monday, July 22, 2024 2:57 PM
To: [email protected]
Cc: Kim, Jonathan <[email protected]>; Kim, Jonathan <[email protected]>
Subject: [PATCH] drm/amdkfd: fix debug watchpoints for logical devices

The number of watchpoints should be set and constrained per logical partition device, not by the socket device.

---
 drivers/gpu/drm/amd/amdkfd/kfd_debug.c  | 20 ++++++++++----------
 drivers/gpu/drm/amd/amdkfd/kfd_device.c |  4 ++--
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h   |  8 ++++----
 3 files changed, 16 insertions(+), 16 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +0 -0
index 34a282540c7e..312dfa84f29f 100644
@@ -365,47 +365,47 @@ static int kfd_dbg_get_dev_watch_id(struct

kfd_process_device *pdd, int *watch_i

        *watch_id = KFD_DEBUGGER_INVALID_WATCH_POINT_ID;
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +5 -5
-       spin_lock(&pdd->dev->kfd->watch_points_lock);
+       spin_lock(&pdd->dev->watch_points_lock);
 
        for (i = 0; i < MAX_WATCH_ADDRESSES; i++) {
                /* device watchpoint in use so skip */
-               if ((pdd->dev->kfd->alloc_watch_ids >> i) & 0x1)
+               if ((pdd->dev->alloc_watch_ids >> i) & 0x1)
                        continue;
 
                pdd->alloc_watch_ids |= 0x1 << i;
-               pdd->dev->kfd->alloc_watch_ids |= 0x1 << i;
+               pdd->dev->alloc_watch_ids |= 0x1 << i;
                *watch_id = i;
-               spin_unlock(&pdd->dev->kfd->watch_points_lock);
+               spin_unlock(&pdd->dev->watch_points_lock);
                return 0;
        }
 
-       spin_unlock(&pdd->dev->kfd->watch_points_lock);
+       spin_unlock(&pdd->dev->watch_points_lock);
 
        return -ENOMEM;
 }
 
 static void kfd_dbg_clear_dev_watch_id(struct kfd_process_device *pdd, int

watch_id)
{

drivers/gpu/drm/amd/amdkfd/kfd_debug.c +3 -3
-       spin_lock(&pdd->dev->kfd->watch_points_lock);
+       spin_lock(&pdd->dev->watch_points_lock);
 
        /* process owns device watch point so safe to clear */
        if ((pdd->alloc_watch_ids >> watch_id) & 0x1) {
                pdd->alloc_watch_ids &= ~(0x1 << watch_id);
-               pdd->dev->kfd->alloc_watch_ids &= ~(0x1 << watch_id);
+               pdd->dev->alloc_watch_ids &= ~(0x1 << watch_id);
        }
 
-       spin_unlock(&pdd->dev->kfd->watch_points_lock);
+       spin_unlock(&pdd->dev->watch_points_lock);
 }
 
 static bool kfd_dbg_owns_dev_watch_id(struct kfd_process_device *pdd, int

watch_id)

 {
        bool owns_watch_id = false;
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +2 -2
-       spin_lock(&pdd->dev->kfd->watch_points_lock);
+       spin_lock(&pdd->dev->watch_points_lock);
        owns_watch_id = watch_id < MAX_WATCH_ADDRESSES &&
                        ((pdd->alloc_watch_ids >> watch_id) & 0x1);
 
-       spin_unlock(&pdd->dev->kfd->watch_points_lock);
+       spin_unlock(&pdd->dev->watch_points_lock);
 
        return owns_watch_id;
 }
drivers/gpu/drm/amd/amdkfd/kfd_device.c +0 -0
index f4d20adaa068..f91a9b6ce3fb 100644
@@ -885,12 +885,12 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd,
                        goto node_init_error;
                }
 
Quoted 2 lines

Wouldn't it better to initialize node->watch_points_lock here. spin_lock_init(&node->watch_points_lock);

Done. Thanks for the review.

Jon

Quoted 43 lines
                kfd->nodes[i] = node;
drivers/gpu/drm/amd/amdkfd/kfd_device.c +2 -2
+
+               spin_lock_init(&kfd->nodes[i]->watch_points_lock);
        }
 
        svm_range_set_max_pages(kfd->adev);
 
-       spin_lock_init(&kfd->watch_points_lock);
-
        kfd->init_complete = true;
        dev_info(kfd_device, "added device %x:%x\n", kfd->adev->pdev->vendor,
                 kfd->adev->pdev->device);
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +4 -5
index 2b3ec92981e8..653e1f934107 100644
@@ -310,6 +310,10 @@ struct kfd_node {
        struct kfd_local_mem_info local_mem_info;
 
        struct kfd_dev *kfd;
+
+       /* Track per device allocated watch points */
+       uint32_t alloc_watch_ids;
+       spinlock_t watch_points_lock;
 };
 
 struct kfd_dev {
@@ -362,10 +366,6 @@ struct kfd_dev {
        struct kfd_node *nodes[MAX_KFD_NODES];
        unsigned int num_nodes;
 
-       /* Track per device allocated watch points */
-       uint32_t alloc_watch_ids;
-       spinlock_t watch_points_lock;
-
        /* Kernel doorbells for KFD device */
        struct amdgpu_bo *doorbells;
 
--

2.34.1