From a7640c07cbbf4aef3c18f25e46706290e3695c99 Mon Sep 17 00:00:00 2001 From: Karol Wachowski Date: Thu, 11 Jun 2026 07:52:14 +0200 Subject: [PATCH] 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. Reviewed-by: Andrzej Kacprowski Signed-off-by: Karol Wachowski Link: https://patch.msgid.link/20260611055214.948748-1-karol.wachowski@linux.intel.com --- drivers/accel/ivpu/ivpu_ipc.c | 17 ++++++++++++++--- drivers/accel/ivpu/ivpu_ipc.h | 1 + 2 files changed, 15 insertions(+), 3 deletions(-) diff --git a/drivers/accel/ivpu/ivpu_ipc.c b/drivers/accel/ivpu/ivpu_ipc.c index 64864e9d8a0a..c5f11be86c8c 100644 --- a/drivers/accel/ivpu/ivpu_ipc.c +++ b/drivers/accel/ivpu/ivpu_ipc.c @@ -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) diff --git a/drivers/accel/ivpu/ivpu_ipc.h b/drivers/accel/ivpu/ivpu_ipc.h index 1c9cf67f1e5b..fb3e5816ca50 100644 --- a/drivers/accel/ivpu/ivpu_ipc.h +++ b/drivers/accel/ivpu/ivpu_ipc.h @@ -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;