Skip to contentSkip to conversation

accel/ivpu: use kmem_cache for IPC rx_msg allocations

PATCH In review Karol Wachowski 3 participants 5 messages Jun 11 – Jun 15

Subsystems (2)
lore
Share & tools
lore

Karol Wachowski

author maintainer
#0 lore raw

PATCH 1/1 accel/ivpu: use kmem_cache for IPC rx_msg allocations

Use a dedicated kmem_cache for ivpu_ipc_rx_msg allocations instead of the generic kmalloc. This gives each CPU its own hot freelist for this object type, reducing allocator overhead in the IRQ handler and threaded IRQ path where rx_msg objects are allocated and freed at high frequency.

---
 drivers/accel/ivpu/ivpu_ipc.c | 17 ++++++++++++++---
 drivers/accel/ivpu/ivpu_ipc.h |  1 +
 2 files changed, 15 insertions(+), 3 deletions(-)
drivers/accel/ivpu/ivpu_ipc.c +14 -3
index 64864e9d8a0a..c5f11be86c8c 100644
@@ -146,7 +146,7 @@ ivpu_ipc_rx_msg_add(struct ivpu_device *vdev, struct ivpu_ipc_consumer *cons,
 
 	lockdep_assert_held(&ipc->cons_lock);
 
-	rx_msg = kzalloc_obj(*rx_msg, GFP_ATOMIC);
+	rx_msg = kmem_cache_zalloc(ipc->rx_msg_cache, GFP_ATOMIC);
 	if (!rx_msg) {
 		ivpu_ipc_rx_mark_free(vdev, ipc_hdr, jsm_msg);
 		return;
@@ -174,7 +174,7 @@ ivpu_ipc_rx_msg_del(struct ivpu_device *vdev, struct ivpu_ipc_rx_msg *rx_msg)
 	list_del(&rx_msg->link);
 	ivpu_ipc_rx_mark_free(vdev, rx_msg->ipc_hdr, rx_msg->jsm_msg);
 	atomic_dec(&vdev->ipc->rx_msg_count);
-	kfree(rx_msg);
+	kmem_cache_free(vdev->ipc->rx_msg_cache, rx_msg);
 }
 
 void ivpu_ipc_consumer_add(struct ivpu_device *vdev, struct ivpu_ipc_consumer *cons,
@@ -490,10 +490,18 @@ int ivpu_ipc_init(struct ivpu_device *vdev)
 	struct ivpu_ipc_info *ipc = vdev->ipc;
 	int ret;
 
+	ipc->rx_msg_cache = kmem_cache_create("ivpu_ipc_rx_msg", sizeof(struct ivpu_ipc_rx_msg), 0,
+					      SLAB_HWCACHE_ALIGN, NULL);
+	if (!ipc->rx_msg_cache) {
+		ivpu_err(vdev, "Failed to create rx_msg_cache\n");
+		return -ENOMEM;
+	}
+
 	ipc->mem_tx = ivpu_bo_create_global(vdev, SZ_16K, DRM_IVPU_BO_WC | DRM_IVPU_BO_MAPPABLE);
 	if (!ipc->mem_tx) {
 		ivpu_err(vdev, "Failed to allocate mem_tx\n");
-		return -ENOMEM;
+		ret = -ENOMEM;
+		goto err_destroy_cache;
 	}
 
 	ipc->mem_rx = ivpu_bo_create_global(vdev, SZ_16K, DRM_IVPU_BO_WC | DRM_IVPU_BO_MAPPABLE);
@@ -532,6 +540,8 @@ int ivpu_ipc_init(struct ivpu_device *vdev)
 	ivpu_bo_free(ipc->mem_rx);
 err_free_tx:
 	ivpu_bo_free(ipc->mem_tx);
+err_destroy_cache:
+	kmem_cache_destroy(ipc->rx_msg_cache);
 	return ret;
 }
 
@@ -544,6 +554,7 @@ void ivpu_ipc_fini(struct ivpu_device *vdev)
 	drm_WARN_ON(&vdev->drm, atomic_read(&ipc->rx_msg_count) > 0);
 
 	ivpu_ipc_mem_fini(vdev);
+	kmem_cache_destroy(ipc->rx_msg_cache);
 }
 
 void ivpu_ipc_enable(struct ivpu_device *vdev)
drivers/accel/ivpu/ivpu_ipc.h +1 -0
index 1c9cf67f1e5b..fb3e5816ca50 100644
@@ -70,6 +70,7 @@ struct ivpu_ipc_info {
 	struct gen_pool *mm_tx;
 	struct ivpu_bo *mem_tx;
 	struct ivpu_bo *mem_rx;
+	struct kmem_cache *rx_msg_cache;
 
 	atomic_t rx_msg_count;
 
Signature · 1 line
2.43.0
#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [Medium] Duplicate kmem_cache name on multi-device systems --

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] Duplicate kmem_cache name on multi-device systems
--

commit 41c16409be6704d2ef5ab21270390fee65dcf7aa
Author: Karol Wachowski <[email protected]>

accel/ivpu: use kmem_cache for IPC rx_msg allocations

This commit introduces a dedicated kmem_cache for allocating IPC rx_msg objects to reduce allocator overhead in IRQ handlers.

Quoted 3 lines
drivers/accel/ivpu/ivpu_ipc.c +0 -0

[ ... ]

drivers/accel/ivpu/ivpu_ipc.c int ivpu_ipc_init(struct ivpu_device *vdev) · line 494

Quoted 6 lines
drivers/accel/ivpu/ivpu_ipc.c +2 -0
@@ -490,10 +490,18 @@ int ivpu_ipc_init(struct ivpu_device *vdev)
 	struct ivpu_ipc_info *ipc = vdev->ipc;
 	int ret;
 
+	ipc->rx_msg_cache = kmem_cache_create("ivpu_ipc_rx_msg", sizeof(struct ivpu_ipc_rx_msg), 0,
+					      SLAB_HWCACHE_ALIGN, NULL);

[Severity: Medium]
Will this cause a name collision if there are multiple ivpu devices in the same system?

Since this is initialized per-device in ivpu_ipc_init(), calling kmem_cache_create() multiple times with the same hardcoded string "ivpu_ipc_rx_msg" can trigger a kernel WARN in kmem_cache_sanity_check() when CONFIG_DEBUG_VM is enabled.

Furthermore, if SLUB merging is disabled via boot parameters or config, the duplicate name causes a sysfs kobject collision, leading to a sysfs warning and a failure to register the cache in sysfs.

Should this cache name be made unique per device using dev_name(), or should the cache be a global singleton?

Quoted 12 lines
drivers/accel/ivpu/ivpu_ipc.c +7 -1
+	if (!ipc->rx_msg_cache) {
+		ivpu_err(vdev, "Failed to create rx_msg_cache\n");
+		return -ENOMEM;
+	}
+
 	ipc->mem_tx = ivpu_bo_create_global(vdev, SZ_16K, DRM_IVPU_BO_WC | DRM_IVPU_BO_MAPPABLE);
 	if (!ipc->mem_tx) {
 		ivpu_err(vdev, "Failed to allocate mem_tx\n");
-		return -ENOMEM;
+		ret = -ENOMEM;
+		goto err_destroy_cache;
 	}
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
author maintainer
#2 lore raw
Quoted 22 lines · On 11-Jun-26 8:04, [email protected] wrote:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] Duplicate kmem_cache name on multi-device systems
--

commit 41c16409be6704d2ef5ab21270390fee65dcf7aa
Author: Karol Wachowski <[email protected]>

accel/ivpu: use kmem_cache for IPC rx_msg allocations

This commit introduces a dedicated kmem_cache for allocating IPC rx_msg objects to reduce allocator overhead in IRQ handlers.

Quoted 3 lines
drivers/accel/ivpu/ivpu_ipc.c +0 -0

[ ... ]

Quoted 4 lines
drivers/accel/ivpu/ivpu_ipc.c +0 -0
@@ -490,10 +490,18 @@ int ivpu_ipc_init(struct ivpu_device *vdev)
  	struct ivpu_ipc_info *ipc = vdev->ipc;
  	int ret;
 
Quoted 6 lines

drivers/accel/ivpu/ivpu_ipc.c int ivpu_ipc_init(struct ivpu_device *vdev) · line 494

Quoted 2 lines
drivers/accel/ivpu/ivpu_ipc.c +2 -0
+	ipc->rx_msg_cache = kmem_cache_create("ivpu_ipc_rx_msg", sizeof(struct ivpu_ipc_rx_msg), 0,
+					      SLAB_HWCACHE_ALIGN, NULL);

[Severity: Medium]
Will this cause a name collision if there are multiple ivpu devices in the same system?

In practice, no:
1. There is a single NPU instance per system, so ivpu_ipc_init()
runs once and the cache is created once

2. Even with multiple instances it's safe, because the cache is
mergeable (only SLAB_HWCACHE_ALIGN flag). On the merge path the name is registered only as a sysfs alias via sysfs_slab_alias(). So the CONFIG_DEBUG_VM duplicate-name check (which scans slab_caches by name) shouldn't fire.

The WARN and sysfs collision require slab_nomerge and more than one
device which doesn't occur for this driver.

Quoted 26 lines

Since this is initialized per-device in ivpu_ipc_init(), calling kmem_cache_create() multiple times with the same hardcoded string "ivpu_ipc_rx_msg" can trigger a kernel WARN in kmem_cache_sanity_check() when CONFIG_DEBUG_VM is enabled.

Furthermore, if SLUB merging is disabled via boot parameters or config, the duplicate name causes a sysfs kobject collision, leading to a sysfs warning and a failure to register the cache in sysfs.

Should this cache name be made unique per device using dev_name(), or should the cache be a global singleton?

Quoted 12 lines
drivers/accel/ivpu/ivpu_ipc.c +7 -1
+	if (!ipc->rx_msg_cache) {
+		ivpu_err(vdev, "Failed to create rx_msg_cache\n");
+		return -ENOMEM;
+	}
+
  	ipc->mem_tx = ivpu_bo_create_global(vdev, SZ_16K, DRM_IVPU_BO_WC | DRM_IVPU_BO_MAPPABLE);
  	if (!ipc->mem_tx) {
  		ivpu_err(vdev, "Failed to allocate mem_tx\n");
-		return -ENOMEM;
+		ret = -ENOMEM;
+		goto err_destroy_cache;
  	}
maintainer Reviewed-by on v1 1/1
#3 lore raw
In reply to Karol Wachowski
Quoted 75 lines · Karol Wachowski

Use a dedicated kmem_cache for ivpu_ipc_rx_msg allocations instead of the generic kmalloc. This gives each CPU its own hot freelist for this object type, reducing allocator overhead in the IRQ handler and threaded IRQ path where rx_msg objects are allocated and freed at high frequency.

---
  drivers/accel/ivpu/ivpu_ipc.c | 17 ++++++++++++++---
  drivers/accel/ivpu/ivpu_ipc.h |  1 +
  2 files changed, 15 insertions(+), 3 deletions(-)
drivers/accel/ivpu/ivpu_ipc.c +14 -3
index 64864e9d8a0a..c5f11be86c8c 100644
@@ -146,7 +146,7 @@ ivpu_ipc_rx_msg_add(struct ivpu_device *vdev, struct ivpu_ipc_consumer *cons,
 
  	lockdep_assert_held(&ipc->cons_lock);
 
-	rx_msg = kzalloc_obj(*rx_msg, GFP_ATOMIC);
+	rx_msg = kmem_cache_zalloc(ipc->rx_msg_cache, GFP_ATOMIC);
  	if (!rx_msg) {
  		ivpu_ipc_rx_mark_free(vdev, ipc_hdr, jsm_msg);
  		return;
@@ -174,7 +174,7 @@ ivpu_ipc_rx_msg_del(struct ivpu_device *vdev, struct ivpu_ipc_rx_msg *rx_msg)
  	list_del(&rx_msg->link);
  	ivpu_ipc_rx_mark_free(vdev, rx_msg->ipc_hdr, rx_msg->jsm_msg);
  	atomic_dec(&vdev->ipc->rx_msg_count);
-	kfree(rx_msg);
+	kmem_cache_free(vdev->ipc->rx_msg_cache, rx_msg);
  }
 
  void ivpu_ipc_consumer_add(struct ivpu_device *vdev, struct ivpu_ipc_consumer *cons,
@@ -490,10 +490,18 @@ int ivpu_ipc_init(struct ivpu_device *vdev)
  	struct ivpu_ipc_info *ipc = vdev->ipc;
  	int ret;
 
+	ipc->rx_msg_cache = kmem_cache_create("ivpu_ipc_rx_msg", sizeof(struct ivpu_ipc_rx_msg), 0,
+					      SLAB_HWCACHE_ALIGN, NULL);
+	if (!ipc->rx_msg_cache) {
+		ivpu_err(vdev, "Failed to create rx_msg_cache\n");
+		return -ENOMEM;
+	}
+
  	ipc->mem_tx = ivpu_bo_create_global(vdev, SZ_16K, DRM_IVPU_BO_WC | DRM_IVPU_BO_MAPPABLE);
  	if (!ipc->mem_tx) {
  		ivpu_err(vdev, "Failed to allocate mem_tx\n");
-		return -ENOMEM;
+		ret = -ENOMEM;
+		goto err_destroy_cache;
  	}
 
  	ipc->mem_rx = ivpu_bo_create_global(vdev, SZ_16K, DRM_IVPU_BO_WC | DRM_IVPU_BO_MAPPABLE);
@@ -532,6 +540,8 @@ int ivpu_ipc_init(struct ivpu_device *vdev)
  	ivpu_bo_free(ipc->mem_rx);
  err_free_tx:
  	ivpu_bo_free(ipc->mem_tx);
+err_destroy_cache:
+	kmem_cache_destroy(ipc->rx_msg_cache);
  	return ret;
  }
 
@@ -544,6 +554,7 @@ void ivpu_ipc_fini(struct ivpu_device *vdev)
  	drm_WARN_ON(&vdev->drm, atomic_read(&ipc->rx_msg_count) > 0);
 
  	ivpu_ipc_mem_fini(vdev);
+	kmem_cache_destroy(ipc->rx_msg_cache);
  }
 
  void ivpu_ipc_enable(struct ivpu_device *vdev)
drivers/accel/ivpu/ivpu_ipc.h +0 -0
index 1c9cf67f1e5b..fb3e5816ca50 100644

drivers/accel/ivpu/ivpu_ipc.h struct ivpu_ipc_info { · line 76

Quoted 7 lines
drivers/accel/ivpu/ivpu_ipc.h +1 -0
@@ -70,6 +70,7 @@ struct ivpu_ipc_info {
  	struct gen_pool *mm_tx;
  	struct ivpu_bo *mem_tx;
  	struct ivpu_bo *mem_rx;
+	struct kmem_cache *rx_msg_cache;
 
  	atomic_t rx_msg_count;
 
author maintainer Applied
#4 lore raw
Quoted 69 lines · Andrzej Kacprowski
Quoted 68 lines · Karol Wachowski

Use a dedicated kmem_cache for ivpu_ipc_rx_msg allocations instead of the generic kmalloc. This gives each CPU its own hot freelist for this object type, reducing allocator overhead in the IRQ handler and threaded IRQ path where rx_msg objects are allocated and freed at high frequency.

---
  drivers/accel/ivpu/ivpu_ipc.c | 17 ++++++++++++++--- drivers/accel/ivpu/ivpu_ipc.h |  1 +

  2 files changed, 15 insertions(+), 3 deletions(-)
drivers/accel/ivpu/ivpu_ipc.c +11 -3
index 64864e9d8a0a..c5f11be86c8c 100644
@@ -146,7 +146,7 @@ ivpu_ipc_rx_msg_add(struct ivpu_device *vdev, 
 struct ivpu_ipc_consumer *cons,
       lockdep_assert_held(&ipc->cons_lock);
-    rx_msg = kzalloc_obj(*rx_msg, GFP_ATOMIC);
+    rx_msg = kmem_cache_zalloc(ipc->rx_msg_cache, GFP_ATOMIC);
       if (!rx_msg) {
           ivpu_ipc_rx_mark_free(vdev, ipc_hdr, jsm_msg);
           return;
@@ -174,7 +174,7 @@ ivpu_ipc_rx_msg_del(struct ivpu_device *vdev, 
 struct ivpu_ipc_rx_msg *rx_msg)
       list_del(&rx_msg->link);
       ivpu_ipc_rx_mark_free(vdev, rx_msg->ipc_hdr, rx_msg->jsm_msg);
       atomic_dec(&vdev->ipc->rx_msg_count);
-    kfree(rx_msg);
+    kmem_cache_free(vdev->ipc->rx_msg_cache, rx_msg);
   }
   void ivpu_ipc_consumer_add(struct ivpu_device *vdev, struct ivpu_ipc_consumer *cons,
@@ -490,10 +490,18 @@ int ivpu_ipc_init(struct ivpu_device *vdev)
       struct ivpu_ipc_info *ipc = vdev->ipc;
       int ret;
+    ipc->rx_msg_cache = kmem_cache_create("ivpu_ipc_rx_msg", 
 sizeof(struct ivpu_ipc_rx_msg), 0,
+                          SLAB_HWCACHE_ALIGN, NULL);
+    if (!ipc->rx_msg_cache) {
+        ivpu_err(vdev, "Failed to create rx_msg_cache\n");
+        return -ENOMEM;
+    }
+
       ipc->mem_tx = ivpu_bo_create_global(vdev, SZ_16K, DRM_IVPU_BO_WC | DRM_IVPU_BO_MAPPABLE);
       if (!ipc->mem_tx) {
           ivpu_err(vdev, "Failed to allocate mem_tx\n");
-        return -ENOMEM;
+        ret = -ENOMEM;
+        goto err_destroy_cache;
       }
       ipc->mem_rx = ivpu_bo_create_global(vdev, SZ_16K, DRM_IVPU_BO_WC | DRM_IVPU_BO_MAPPABLE);
@@ -532,6 +540,8 @@ int ivpu_ipc_init(struct ivpu_device *vdev)
       ivpu_bo_free(ipc->mem_rx);

  err_free_tx:
      ivpu_bo_free(ipc->mem_tx);
+err_destroy_cache:
+    kmem_cache_destroy(ipc->rx_msg_cache);
      return ret;
  }

drivers/accel/ivpu/ivpu_ipc.c +1 -0
@@ -544,6 +554,7 @@ void ivpu_ipc_fini(struct ivpu_device *vdev)
       drm_WARN_ON(&vdev->drm, atomic_read(&ipc->rx_msg_count) > 0);
       ivpu_ipc_mem_fini(vdev);
+    kmem_cache_destroy(ipc->rx_msg_cache);
   }
   void ivpu_ipc_enable(struct ivpu_device *vdev)
drivers/accel/ivpu/ ivpu_ipc.h renamed from drivers/accel/ivpu/ivpu_ipc.h +0 -1
index 1c9cf67f1e5b..fb3e5816ca50 100644
--- a/drivers/accel/ivpu/ivpu_ipc.h
Quoted 8 lines

drivers/accel/ivpu/ ivpu_ipc.h struct ivpu_ipc_info { · line 74

Quoted 7 lines

+++ b/drivers/accel/ivpu/ivpu_ipc.h

drivers/accel/ivpu/ ivpu_ipc.h renamed from drivers/accel/ivpu/ivpu_ipc.h +1 -0
@@ -70,6 +70,7 @@ struct ivpu_ipc_info {
       struct gen_pool *mm_tx;
       struct ivpu_bo *mem_tx;
       struct ivpu_bo *mem_rx;
+    struct kmem_cache *rx_msg_cache;
       atomic_t rx_msg_count;

Pushed to drm-misc-next.