[PATCH v2 1/4] drm/amdgpu: Add a helper macro to align mqd size

[PATCH v2 1/4] drm/amdgpu: Add a helper macro to align mqd size

[Lang Yu]

MES FW uses address(mqd_addr + sizeof(struct mqd) + 3*sizeof(uint32_t))
as fence address and writes a 32 bit fence value to this address. Driver
needs to allocate some extra memory(at least 4 DWs) in addition to
sizeof(struct mqd) as mqd memory.

For gfx11/12, sizeof(struct mqd) < PAGE_SIZE, KGD allocates mqd memory with
PAGE_SIZE aligned works. For gfx12.1, sizeof(struct mqd) == PAGE_SIZE,
it doesn't work.

KFD mqd manager hardcodes mqd size to PAGE_SIZE/MQD_SIZE across different
IP versions to solve this issue.

To avoid hardcoding in differnet places and across different IP versions.
Let's use AMDGPU_MQD_SIZE_ALIGN instead. It is used in two places.

1. mqd memory alloction
2. mqd stride handling for multi xcc config

v2: Use AMDGPU_GPU_PAGE_ALIGN. (Mukul)

Signed-off-by: Lang Yu <lang.yu@amd.com>
Reviewed-by: David Belanger <david.belanger@amd.com> (v1)
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index 9c11535c44c6..8c6edb0f58dd 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -1239,6 +1239,14 @@ struct amdgpu_device {
 	struct amdgpu_kfd_dev		kfd;
 };
 
+/*
+ * MES FW uses address(mqd_addr + sizeof(struct mqd) + 3*sizeof(uint32_t))
+ * as fence address and writes a 32 bit fence value to this address.
+ * Driver needs to allocate at least 4 DWs extra memory in addition to
+ * sizeof(struct mqd). Add 8 DWs and align to AMDGPU_GPU_PAGE_SIZE for safety.
+ */
+#define AMDGPU_MQD_SIZE_ALIGN(mqd_size) AMDGPU_GPU_PAGE_ALIGN(((mqd_size) + 32))
+
 static inline uint32_t amdgpu_ip_version(const struct amdgpu_device *adev,
 					 uint8_t ip, uint8_t inst)
 {
-- 
2.34.1

[PATCH 2/4] drm/amdgpu: Use AMDGPU_MQD_SIZE_ALIGN in KGD

[Lang Yu]

Use AMDGPU_MQD_SIZE_ALIGN for both kernel and user queue.

Signed-off-by: Lang Yu <lang.yu@amd.com>
Reviewed-by: David Belanger <david.belanger@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c    | 18 ++++++++++--------
 drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c     |  2 +-
 drivers/gpu/drm/amd/amdgpu/mes_userqueue.c |  2 +-
 3 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
index 6abe5103a78d..4fede701beb8 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
@@ -385,6 +385,8 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
 	struct amdgpu_ring *ring = &kiq->ring;
 	u32 domain = AMDGPU_GEM_DOMAIN_GTT;
+	u32 gfx_mqd_size = max(adev->mqds[AMDGPU_HW_IP_GFX].mqd_size, mqd_size);
+	u32 compute_mqd_size = max(adev->mqds[AMDGPU_HW_IP_COMPUTE].mqd_size, mqd_size);
 
 #if !defined(CONFIG_ARM) && !defined(CONFIG_ARM64)
 	/* Only enable on gfx10 and 11 for now to avoid changing behavior on older chips */
@@ -424,17 +426,17 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
 		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
 			ring = &adev->gfx.gfx_ring[i];
 			if (!ring->mqd_obj) {
-				r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
-							    domain, &ring->mqd_obj,
+				r = amdgpu_bo_create_kernel(adev, AMDGPU_MQD_SIZE_ALIGN(gfx_mqd_size),
+								PAGE_SIZE, domain, &ring->mqd_obj,
 							    &ring->mqd_gpu_addr, &ring->mqd_ptr);
 				if (r) {
 					dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
 					return r;
 				}
 
-				ring->mqd_size = mqd_size;
+				ring->mqd_size = gfx_mqd_size;
 				/* prepare MQD backup */
-				adev->gfx.me.mqd_backup[i] = kzalloc(mqd_size, GFP_KERNEL);
+				adev->gfx.me.mqd_backup[i] = kzalloc(gfx_mqd_size, GFP_KERNEL);
 				if (!adev->gfx.me.mqd_backup[i]) {
 					dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
 					return -ENOMEM;
@@ -448,17 +450,17 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
 		j = i + xcc_id * adev->gfx.num_compute_rings;
 		ring = &adev->gfx.compute_ring[j];
 		if (!ring->mqd_obj) {
-			r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
-						    domain, &ring->mqd_obj,
+			r = amdgpu_bo_create_kernel(adev, AMDGPU_MQD_SIZE_ALIGN(compute_mqd_size),
+							PAGE_SIZE, domain, &ring->mqd_obj,
 						    &ring->mqd_gpu_addr, &ring->mqd_ptr);
 			if (r) {
 				dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
 				return r;
 			}
 
-			ring->mqd_size = mqd_size;
+			ring->mqd_size = compute_mqd_size;
 			/* prepare MQD backup */
-			adev->gfx.mec.mqd_backup[j] = kzalloc(mqd_size, GFP_KERNEL);
+			adev->gfx.mec.mqd_backup[j] = kzalloc(compute_mqd_size, GFP_KERNEL);
 			if (!adev->gfx.mec.mqd_backup[j]) {
 				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
 				return -ENOMEM;
diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c b/drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c
index 86cc90a66296..4dabcd7c9f5e 100644
--- a/drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c
+++ b/drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c
@@ -2235,7 +2235,7 @@ static int gfx_v12_1_compute_mqd_init(struct amdgpu_device *adev, void *m,
 	mqd->cp_hqd_queue_priority = prop->hqd_queue_priority;
 
 	mqd->cp_mqd_stride_size = prop->mqd_stride_size ? prop->mqd_stride_size :
-		sizeof(struct v12_1_compute_mqd);
+		AMDGPU_MQD_SIZE_ALIGN(adev->mqds[AMDGPU_HW_IP_COMPUTE].mqd_size);
 
 	mqd->cp_hqd_active = prop->hqd_active;
 
diff --git a/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c b/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
index f2309d72bbe6..82bff8dc286d 100644
--- a/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
+++ b/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
@@ -289,7 +289,7 @@ static int mes_userq_mqd_create(struct amdgpu_usermode_queue *queue,
 		return -ENOMEM;
 	}
 
-	r = amdgpu_userq_create_object(uq_mgr, &queue->mqd, mqd_hw_default->mqd_size);
+	r = amdgpu_userq_create_object(uq_mgr, &queue->mqd, AMDGPU_MQD_SIZE_ALIGN(mqd_hw_default->mqd_size));
 	if (r) {
 		DRM_ERROR("Failed to create MQD object for userqueue\n");
 		goto free_props;
-- 
2.34.1

[PATCH 3/4] drm/amdkfd: Adjust parameter of allocate_mqd

[Lang Yu]

Make allocate_mqd consistent with other callbacks.
Prepare for next patch to use mqd_manager->mqd_size.

Signed-off-by: Lang Yu <lang.yu@amd.com>
Reviewed-by: David Belanger <david.belanger@amd.com>
---
 drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 4 ++--
 drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c         | 2 +-
 drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c          | 6 ++++--
 drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h          | 6 +++---
 drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c      | 3 ++-
 drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c      | 3 ++-
 drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c      | 3 ++-
 drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c      | 3 ++-
 drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c    | 3 ++-
 drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c       | 3 ++-
 drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c       | 3 ++-
 11 files changed, 24 insertions(+), 15 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
index b542de9d50d1..804851632c4c 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
@@ -676,7 +676,7 @@ static int create_queue_nocpsch(struct device_queue_manager *dqm,
 
 	/* Temporarily release dqm lock to avoid a circular lock dependency */
 	dqm_unlock(dqm);
-	q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties);
+	q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr, &q->properties);
 	dqm_lock(dqm);
 
 	if (!q->mqd_mem_obj) {
@@ -2002,7 +2002,7 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
 		dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
 	q->properties.tba_addr = qpd->tba_addr;
 	q->properties.tma_addr = qpd->tma_addr;
-	q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties);
+	q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr, &q->properties);
 	if (!q->mqd_mem_obj) {
 		retval = -ENOMEM;
 		goto out_deallocate_doorbell;
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
index 961a4b84e974..d987ff7ccfc9 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
@@ -130,7 +130,7 @@ static bool kq_initialize(struct kernel_queue *kq, struct kfd_node *dev,
 
 	kq->queue->device = dev;
 
-	kq->queue->mqd_mem_obj = kq->mqd_mgr->allocate_mqd(kq->mqd_mgr->dev,
+	kq->queue->mqd_mem_obj = kq->mqd_mgr->allocate_mqd(kq->mqd_mgr,
 					&kq->queue->properties);
 	if (!kq->queue->mqd_mem_obj)
 		goto err_allocate_mqd;
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c
index f78b249e1a41..ceb6566ff3e1 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c
@@ -46,9 +46,10 @@ int pipe_priority_map[] = {
 	KFD_PIPE_PRIORITY_CS_HIGH
 };
 
-struct kfd_mem_obj *allocate_hiq_mqd(struct kfd_node *dev, struct queue_properties *q)
+struct kfd_mem_obj *allocate_hiq_mqd(struct mqd_manager *mm, struct queue_properties *q)
 {
 	struct kfd_mem_obj *mqd_mem_obj;
+	struct kfd_node *dev = mm->dev;
 
 	mqd_mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
 	if (!mqd_mem_obj)
@@ -61,10 +62,11 @@ struct kfd_mem_obj *allocate_hiq_mqd(struct kfd_node *dev, struct queue_properti
 	return mqd_mem_obj;
 }
 
-struct kfd_mem_obj *allocate_sdma_mqd(struct kfd_node *dev,
+struct kfd_mem_obj *allocate_sdma_mqd(struct mqd_manager *mm,
 					struct queue_properties *q)
 {
 	struct kfd_mem_obj *mqd_mem_obj;
+	struct kfd_node *dev = mm->dev;
 	uint64_t offset;
 
 	mqd_mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h
index 17cc1f25c8d0..2429d278ef0e 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h
@@ -68,7 +68,7 @@
  */
 extern int pipe_priority_map[];
 struct mqd_manager {
-	struct kfd_mem_obj*	(*allocate_mqd)(struct kfd_node *kfd,
+	struct kfd_mem_obj*	(*allocate_mqd)(struct mqd_manager *mm,
 		struct queue_properties *q);
 
 	void	(*init_mqd)(struct mqd_manager *mm, void **mqd,
@@ -153,10 +153,10 @@ struct mqd_user_context_save_area_header {
 	uint32_t wave_state_size;
 };
 
-struct kfd_mem_obj *allocate_hiq_mqd(struct kfd_node *dev,
+struct kfd_mem_obj *allocate_hiq_mqd(struct mqd_manager *mm,
 				struct queue_properties *q);
 
-struct kfd_mem_obj *allocate_sdma_mqd(struct kfd_node *dev,
+struct kfd_mem_obj *allocate_sdma_mqd(struct mqd_manager *mm,
 					struct queue_properties *q);
 void free_mqd_hiq_sdma(struct mqd_manager *mm, void *mqd,
 				struct kfd_mem_obj *mqd_mem_obj);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c
index 05f3ac2eaef9..90ac3a30e81d 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c
@@ -73,9 +73,10 @@ static void set_priority(struct cik_mqd *m, struct queue_properties *q)
 	m->cp_hqd_queue_priority = q->priority;
 }
 
-static struct kfd_mem_obj *allocate_mqd(struct kfd_node *kfd,
+static struct kfd_mem_obj *allocate_mqd(struct mqd_manager *mm,
 					struct queue_properties *q)
 {
+	struct kfd_node *kfd = mm->dev;
 	struct kfd_mem_obj *mqd_mem_obj;
 
 	if (kfd_gtt_sa_allocate(kfd, sizeof(struct cik_mqd),
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c
index 1695dd78ede8..97055f808d4a 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c
@@ -73,9 +73,10 @@ static void set_priority(struct v10_compute_mqd *m, struct queue_properties *q)
 	m->cp_hqd_queue_priority = q->priority;
 }
 
-static struct kfd_mem_obj *allocate_mqd(struct kfd_node *kfd,
+static struct kfd_mem_obj *allocate_mqd(struct mqd_manager *mm,
 		struct queue_properties *q)
 {
+	struct kfd_node *kfd = mm->dev;
 	struct kfd_mem_obj *mqd_mem_obj;
 
 	if (kfd_gtt_sa_allocate(kfd, sizeof(struct v10_compute_mqd),
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c
index 3c0ae28c5923..5c44d0987737 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c
@@ -99,9 +99,10 @@ static void set_priority(struct v11_compute_mqd *m, struct queue_properties *q)
 	m->cp_hqd_queue_priority = q->priority;
 }
 
-static struct kfd_mem_obj *allocate_mqd(struct kfd_node *node,
+static struct kfd_mem_obj *allocate_mqd(struct mqd_manager *mm,
 		struct queue_properties *q)
 {
+	struct kfd_node *node = mm->dev;
 	struct kfd_mem_obj *mqd_mem_obj;
 	int size;
 
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c
index 729df66ebfb3..b7ac2dea8775 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c
@@ -80,9 +80,10 @@ static void set_priority(struct v12_compute_mqd *m, struct queue_properties *q)
 	m->cp_hqd_queue_priority = q->priority;
 }
 
-static struct kfd_mem_obj *allocate_mqd(struct kfd_node *node,
+static struct kfd_mem_obj *allocate_mqd(struct mqd_manager *mm,
 		struct queue_properties *q)
 {
+	struct kfd_node *node = mm->dev;
 	struct kfd_mem_obj *mqd_mem_obj;
 
 	/*
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c
index 558216395a4d..0b0d802a0917 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c
@@ -145,9 +145,10 @@ static void set_priority(struct v12_1_compute_mqd *m, struct queue_properties *q
 	m->cp_hqd_queue_priority = q->priority;
 }
 
-static struct kfd_mem_obj *allocate_mqd(struct kfd_node *node,
+static struct kfd_mem_obj *allocate_mqd(struct mqd_manager *mm,
 		struct queue_properties *q)
 {
+	struct kfd_node *node = mm->dev;
 	struct kfd_mem_obj *mqd_mem_obj;
 	unsigned int size;
 
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c
index d867dccae675..dcf4bbfa641b 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c
@@ -120,10 +120,11 @@ static bool mqd_on_vram(struct amdgpu_device *adev)
 	}
 }
 
-static struct kfd_mem_obj *allocate_mqd(struct kfd_node *node,
+static struct kfd_mem_obj *allocate_mqd(struct mqd_manager *mm,
 		struct queue_properties *q)
 {
 	int retval;
+	struct kfd_node *node = mm->dev;
 	struct kfd_mem_obj *mqd_mem_obj = NULL;
 
 	/* For V9 only, due to a HW bug, the control stack of a user mode
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c
index c1fafc502515..09483f0862d4 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c
@@ -76,9 +76,10 @@ static void set_priority(struct vi_mqd *m, struct queue_properties *q)
 	m->cp_hqd_queue_priority = q->priority;
 }
 
-static struct kfd_mem_obj *allocate_mqd(struct kfd_node *kfd,
+static struct kfd_mem_obj *allocate_mqd(struct mqd_manager *mm,
 					struct queue_properties *q)
 {
+	struct kfd_node *kfd = mm->dev;
 	struct kfd_mem_obj *mqd_mem_obj;
 
 	if (kfd_gtt_sa_allocate(kfd, sizeof(struct vi_mqd),
-- 
2.34.1

[PATCH 4/4] drm/amdkfd: Use AMDGPU_MQD_SIZE_ALIGN in gfx11+ kfd mqd manager

[Lang Yu]

MES is enabled by default from gfx11+, use AMDGPU_MQD_SIZE_ALIGN
unconditionally for gfx11+.

Signed-off-by: Lang Yu <lang.yu@amd.com>
Reviewed-by: David Belanger <david.belanger@amd.com>
---
 drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c  |  3 ++
 .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c  | 22 +++-----------
 .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c  | 10 +++----
 .../drm/amd/amdkfd/kfd_mqd_manager_v12_1.c    | 29 ++++---------------
 4 files changed, 17 insertions(+), 47 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c
index ceb6566ff3e1..d88d0de58edd 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c
@@ -292,6 +292,9 @@ void kfd_get_hiq_xcc_mqd(struct kfd_node *dev, struct kfd_mem_obj *mqd_mem_obj,
 uint64_t kfd_mqd_stride(struct mqd_manager *mm,
 			struct queue_properties *q)
 {
+	if (KFD_GC_VERSION(mm->dev) >= IP_VERSION(11, 0, 0))
+		return AMDGPU_MQD_SIZE_ALIGN(mm->mqd_size);
+
 	return mm->mqd_size;
 }
 
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c
index 5c44d0987737..7e5a7ab6d0c0 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c
@@ -102,20 +102,11 @@ static void set_priority(struct v11_compute_mqd *m, struct queue_properties *q)
 static struct kfd_mem_obj *allocate_mqd(struct mqd_manager *mm,
 		struct queue_properties *q)
 {
+	u32 mqd_size = AMDGPU_MQD_SIZE_ALIGN(mm->mqd_size);
 	struct kfd_node *node = mm->dev;
 	struct kfd_mem_obj *mqd_mem_obj;
-	int size;
-
-	/*
-	 * MES write to areas beyond MQD size. So allocate
-	 * 1 PAGE_SIZE memory for MQD is MES is enabled.
-	 */
-	if (node->kfd->shared_resources.enable_mes)
-		size = PAGE_SIZE;
-	else
-		size = sizeof(struct v11_compute_mqd);
 
-	if (kfd_gtt_sa_allocate(node, size, &mqd_mem_obj))
+	if (kfd_gtt_sa_allocate(node, mqd_size, &mqd_mem_obj))
 		return NULL;
 
 	return mqd_mem_obj;
@@ -127,18 +118,13 @@ static void init_mqd(struct mqd_manager *mm, void **mqd,
 {
 	uint64_t addr;
 	struct v11_compute_mqd *m;
-	int size;
+	u32 mqd_size = AMDGPU_MQD_SIZE_ALIGN(mm->mqd_size);
 	uint32_t wa_mask = q->is_dbg_wa ? 0xffff : 0xffffffff;
 
 	m = (struct v11_compute_mqd *) mqd_mem_obj->cpu_ptr;
 	addr = mqd_mem_obj->gpu_addr;
 
-	if (mm->dev->kfd->shared_resources.enable_mes)
-		size = PAGE_SIZE;
-	else
-		size = sizeof(struct v11_compute_mqd);
-
-	memset(m, 0, size);
+	memset(m, 0, mqd_size);
 
 	m->header = 0xC0310800;
 	m->compute_pipelinestat_enable = 1;
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c
index b7ac2dea8775..a51f217329db 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12.c
@@ -83,14 +83,11 @@ static void set_priority(struct v12_compute_mqd *m, struct queue_properties *q)
 static struct kfd_mem_obj *allocate_mqd(struct mqd_manager *mm,
 		struct queue_properties *q)
 {
+	u32 mqd_size = AMDGPU_MQD_SIZE_ALIGN(mm->mqd_size);
 	struct kfd_node *node = mm->dev;
 	struct kfd_mem_obj *mqd_mem_obj;
 
-	/*
-	 * Allocate one PAGE_SIZE memory for MQD as MES writes to areas beyond
-	 * struct MQD size.
-	 */
-	if (kfd_gtt_sa_allocate(node, PAGE_SIZE, &mqd_mem_obj))
+	if (kfd_gtt_sa_allocate(node, mqd_size, &mqd_mem_obj))
 		return NULL;
 
 	return mqd_mem_obj;
@@ -102,11 +99,12 @@ static void init_mqd(struct mqd_manager *mm, void **mqd,
 {
 	uint64_t addr;
 	struct v12_compute_mqd *m;
+	u32 mqd_size = AMDGPU_MQD_SIZE_ALIGN(mm->mqd_size);
 
 	m = (struct v12_compute_mqd *) mqd_mem_obj->cpu_ptr;
 	addr = mqd_mem_obj->gpu_addr;
 
-	memset(m, 0, PAGE_SIZE);
+	memset(m, 0, mqd_size);
 
 	m->header = 0xC0310800;
 	m->compute_pipelinestat_enable = 1;
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c
index 0b0d802a0917..d0776ba2cc99 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c
@@ -32,17 +32,6 @@
 #include "amdgpu_amdkfd.h"
 #include "kfd_device_queue_manager.h"
 
-#define MQD_SIZE	(2 * PAGE_SIZE)
-
-static uint64_t mqd_stride_v12_1(struct mqd_manager *mm,
-				struct queue_properties *q)
-{
-	if (q->type == KFD_QUEUE_TYPE_COMPUTE)
-		return MQD_SIZE;
-	else
-		return PAGE_SIZE;
-}
-
 static inline struct v12_1_compute_mqd *get_mqd(void *mqd)
 {
 	return (struct v12_1_compute_mqd *)mqd;
@@ -148,21 +137,14 @@ static void set_priority(struct v12_1_compute_mqd *m, struct queue_properties *q
 static struct kfd_mem_obj *allocate_mqd(struct mqd_manager *mm,
 		struct queue_properties *q)
 {
+	u32 mqd_size = AMDGPU_MQD_SIZE_ALIGN(mm->mqd_size);
 	struct kfd_node *node = mm->dev;
 	struct kfd_mem_obj *mqd_mem_obj;
-	unsigned int size;
 
-	/*
-	 * Allocate two PAGE_SIZE memory for Compute MQD as MES writes to areas beyond
-	 * struct MQD size. Size of the Compute MQD is 1 PAGE_SIZE.
-	 * For SDMA MQD, we allocate 1 Page_size.
-	 */
 	if (q->type == KFD_QUEUE_TYPE_COMPUTE)
-		size = MQD_SIZE * NUM_XCC(node->xcc_mask);
-	else
-		size = PAGE_SIZE;
+		mqd_size *= NUM_XCC(node->xcc_mask);
 
-	if (kfd_gtt_sa_allocate(node, size, &mqd_mem_obj))
+	if (kfd_gtt_sa_allocate(node, mqd_size, &mqd_mem_obj))
 		return NULL;
 
 	return mqd_mem_obj;
@@ -174,11 +156,12 @@ static void init_mqd(struct mqd_manager *mm, void **mqd,
 {
 	uint64_t addr;
 	struct v12_1_compute_mqd *m;
+	u32 mqd_size = AMDGPU_MQD_SIZE_ALIGN(mm->mqd_size);
 
 	m = (struct v12_1_compute_mqd *) mqd_mem_obj->cpu_ptr;
 	addr = mqd_mem_obj->gpu_addr;
 
-	memset(m, 0, MQD_SIZE);
+	memset(m, 0, mqd_size);
 
 	m->header = 0xC0310800;
 	m->compute_pipelinestat_enable = 1;
@@ -681,7 +664,7 @@ struct mqd_manager *mqd_manager_init_v12_1(enum KFD_MQD_TYPE type,
 		mqd->is_occupied = kfd_is_occupied_cp;
 		mqd->mqd_size = sizeof(struct v12_1_compute_mqd);
 		mqd->get_wave_state = get_wave_state_v12_1;
-		mqd->mqd_stride = mqd_stride_v12_1;
+		mqd->mqd_stride = kfd_mqd_stride;
 #if defined(CONFIG_DEBUG_FS)
 		mqd->debugfs_show_mqd = debugfs_show_mqd;
 #endif
-- 
2.34.1

RE: [PATCH v2 1/4] drm/amdgpu: Add a helper macro to align mqd size

[Zhang, Hawking]

[AMD Official Use Only - AMD Internal Distribution Only]

Series is

Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>

Regards,
Hawking
-----Original Message-----
From: Yu, Lang <Lang.Yu@amd.com>
Sent: Wednesday, January 28, 2026 10:21
To: amd-gfx@lists.freedesktop.org
Cc: Zhang, Hawking <Hawking.Zhang@amd.com>; Deucher, Alexander <Alexander.Deucher@amd.com>; Joshi, Mukul <Mukul.Joshi@amd.com>; Belanger, David <David.Belanger@amd.com>; Yu, Lang <Lang.Yu@amd.com>
Subject: [PATCH v2 1/4] drm/amdgpu: Add a helper macro to align mqd size

MES FW uses address(mqd_addr + sizeof(struct mqd) + 3*sizeof(uint32_t)) as fence address and writes a 32 bit fence value to this address. Driver needs to allocate some extra memory(at least 4 DWs) in addition to sizeof(struct mqd) as mqd memory.

For gfx11/12, sizeof(struct mqd) < PAGE_SIZE, KGD allocates mqd memory with PAGE_SIZE aligned works. For gfx12.1, sizeof(struct mqd) == PAGE_SIZE, it doesn't work.

KFD mqd manager hardcodes mqd size to PAGE_SIZE/MQD_SIZE across different IP versions to solve this issue.

To avoid hardcoding in differnet places and across different IP versions.
Let's use AMDGPU_MQD_SIZE_ALIGN instead. It is used in two places.

1. mqd memory alloction
2. mqd stride handling for multi xcc config

v2: Use AMDGPU_GPU_PAGE_ALIGN. (Mukul)

Signed-off-by: Lang Yu <lang.yu@amd.com>
Reviewed-by: David Belanger <david.belanger@amd.com> (v1)
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index 9c11535c44c6..8c6edb0f58dd 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -1239,6 +1239,14 @@ struct amdgpu_device {
        struct amdgpu_kfd_dev           kfd;
 };

+/*
+ * MES FW uses address(mqd_addr + sizeof(struct mqd) +
+3*sizeof(uint32_t))
+ * as fence address and writes a 32 bit fence value to this address.
+ * Driver needs to allocate at least 4 DWs extra memory in addition to
+ * sizeof(struct mqd). Add 8 DWs and align to AMDGPU_GPU_PAGE_SIZE for safety.
+ */
+#define AMDGPU_MQD_SIZE_ALIGN(mqd_size)
+AMDGPU_GPU_PAGE_ALIGN(((mqd_size) + 32))
+
 static inline uint32_t amdgpu_ip_version(const struct amdgpu_device *adev,
                                         uint8_t ip, uint8_t inst)
 {
--
2.34.1

RE: [PATCH v2 1/4] drm/amdgpu: Add a helper macro to align mqd size

[Joshi, Mukul]

[AMD Official Use Only - AMD Internal Distribution Only]

Series is :
Reviewed-by: Mukul Joshi <mukul.joshi@amd.com>


> -----Original Message-----
> From: Yu, Lang <Lang.Yu@amd.com>
> Sent: Tuesday, January 27, 2026 9:21 PM
> To: amd-gfx@lists.freedesktop.org
> Cc: Zhang, Hawking <Hawking.Zhang@amd.com>; Deucher, Alexander
> <Alexander.Deucher@amd.com>; Joshi, Mukul <Mukul.Joshi@amd.com>;
> Belanger, David <David.Belanger@amd.com>; Yu, Lang <Lang.Yu@amd.com>
> Subject: [PATCH v2 1/4] drm/amdgpu: Add a helper macro to align mqd size
>
> MES FW uses address(mqd_addr + sizeof(struct mqd) + 3*sizeof(uint32_t)) as
> fence address and writes a 32 bit fence value to this address. Driver needs to
> allocate some extra memory(at least 4 DWs) in addition to sizeof(struct mqd)
> as mqd memory.
>
> For gfx11/12, sizeof(struct mqd) < PAGE_SIZE, KGD allocates mqd memory
> with PAGE_SIZE aligned works. For gfx12.1, sizeof(struct mqd) == PAGE_SIZE,
> it doesn't work.
>
> KFD mqd manager hardcodes mqd size to PAGE_SIZE/MQD_SIZE across
> different IP versions to solve this issue.
>
> To avoid hardcoding in differnet places and across different IP versions.
> Let's use AMDGPU_MQD_SIZE_ALIGN instead. It is used in two places.
>
> 1. mqd memory alloction
> 2. mqd stride handling for multi xcc config
>
> v2: Use AMDGPU_GPU_PAGE_ALIGN. (Mukul)
>
> Signed-off-by: Lang Yu <lang.yu@amd.com>
> Reviewed-by: David Belanger <david.belanger@amd.com> (v1)
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu.h | 8 ++++++++
>  1 file changed, 8 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> index 9c11535c44c6..8c6edb0f58dd 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> @@ -1239,6 +1239,14 @@ struct amdgpu_device {
>       struct amdgpu_kfd_dev           kfd;
>  };
>
> +/*
> + * MES FW uses address(mqd_addr + sizeof(struct mqd) +
> +3*sizeof(uint32_t))
> + * as fence address and writes a 32 bit fence value to this address.
> + * Driver needs to allocate at least 4 DWs extra memory in addition to
> + * sizeof(struct mqd). Add 8 DWs and align to AMDGPU_GPU_PAGE_SIZE for
> safety.
> + */
> +#define AMDGPU_MQD_SIZE_ALIGN(mqd_size)
> +AMDGPU_GPU_PAGE_ALIGN(((mqd_size) + 32))
> +
>  static inline uint32_t amdgpu_ip_version(const struct amdgpu_device *adev,
>                                        uint8_t ip, uint8_t inst)
>  {
> --
> 2.34.1

[PATCH 1/4] drm/amdgpu: Add a helper macro to align mqd size

[Lang Yu]

MES FW uses addr(mqd_addr + sizeof(struct mqd) + 3*sizeof(uint32_t))
as fence address and writes a 32 bit fence value to this address.

Driver needs to allocate some extra memory(at least 4 DWs) in addition to
sizeof(struct mqd) as mqd memory.

For gfx11/12, sizeof(struct mqd) < PAGE_SIZE, allocate mqd memory with
PAGE_SIZE aligned works. For gfx12.1, sizeof(struct mqd) == PAGE_SIZE,
it doesn't work.

KFD mqd manager hardcodes mqd size to PAGE_SIZE/MQD_SIZE.

Let's use AMDGPU_MQD_SIZE_ALIGN to avoid hardcoding in differnet place
and across different IP version. It is used in two place.
1. mqd memory alloction
2. mqd stride initialization

Signed-off-by: Lang Yu <lang.yu@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index 9c11535c44c6..41f32ed39113 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -1239,6 +1239,11 @@ struct amdgpu_device {
 	struct amdgpu_kfd_dev		kfd;
 };
 
+/*
+ * MES will use memory beyond struct MQD size, 5 DWs currently
+ */
+#define AMDGPU_MQD_SIZE_ALIGN(mqd_size) ALIGN(((mqd_size) + 20), PAGE_SIZE)
+
 static inline uint32_t amdgpu_ip_version(const struct amdgpu_device *adev,
 					 uint8_t ip, uint8_t inst)
 {
-- 
2.34.1

[PATCH 2/4] drm/amdgpu: Use AMDGPU_MQD_SIZE_ALIGN in KGD

[Lang Yu]

Use AMDGPU_MQD_SIZE_ALIGN for both kernel and user queue.

Signed-off-by: Lang Yu <lang.yu@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c    | 18 ++++++++++--------
 drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c     |  2 +-
 drivers/gpu/drm/amd/amdgpu/mes_userqueue.c |  3 ++-
 3 files changed, 13 insertions(+), 10 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
index 6abe5103a78d..4fede701beb8 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
@@ -385,6 +385,8 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_id];
 	struct amdgpu_ring *ring = &kiq->ring;
 	u32 domain = AMDGPU_GEM_DOMAIN_GTT;
+	u32 gfx_mqd_size = max(adev->mqds[AMDGPU_HW_IP_GFX].mqd_size, mqd_size);
+	u32 compute_mqd_size = max(adev->mqds[AMDGPU_HW_IP_COMPUTE].mqd_size, mqd_size);
 
 #if !defined(CONFIG_ARM) && !defined(CONFIG_ARM64)
 	/* Only enable on gfx10 and 11 for now to avoid changing behavior on older chips */
@@ -424,17 +426,17 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
 		for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
 			ring = &adev->gfx.gfx_ring[i];
 			if (!ring->mqd_obj) {
-				r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
-							    domain, &ring->mqd_obj,
+				r = amdgpu_bo_create_kernel(adev, AMDGPU_MQD_SIZE_ALIGN(gfx_mqd_size),
+								PAGE_SIZE, domain, &ring->mqd_obj,
 							    &ring->mqd_gpu_addr, &ring->mqd_ptr);
 				if (r) {
 					dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
 					return r;
 				}
 
-				ring->mqd_size = mqd_size;
+				ring->mqd_size = gfx_mqd_size;
 				/* prepare MQD backup */
-				adev->gfx.me.mqd_backup[i] = kzalloc(mqd_size, GFP_KERNEL);
+				adev->gfx.me.mqd_backup[i] = kzalloc(gfx_mqd_size, GFP_KERNEL);
 				if (!adev->gfx.me.mqd_backup[i]) {
 					dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
 					return -ENOMEM;
@@ -448,17 +450,17 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
 		j = i + xcc_id * adev->gfx.num_compute_rings;
 		ring = &adev->gfx.compute_ring[j];
 		if (!ring->mqd_obj) {
-			r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
-						    domain, &ring->mqd_obj,
+			r = amdgpu_bo_create_kernel(adev, AMDGPU_MQD_SIZE_ALIGN(compute_mqd_size),
+							PAGE_SIZE, domain, &ring->mqd_obj,
 						    &ring->mqd_gpu_addr, &ring->mqd_ptr);
 			if (r) {
 				dev_warn(adev->dev, "failed to create ring mqd bo (%d)", r);
 				return r;
 			}
 
-			ring->mqd_size = mqd_size;
+			ring->mqd_size = compute_mqd_size;
 			/* prepare MQD backup */
-			adev->gfx.mec.mqd_backup[j] = kzalloc(mqd_size, GFP_KERNEL);
+			adev->gfx.mec.mqd_backup[j] = kzalloc(compute_mqd_size, GFP_KERNEL);
 			if (!adev->gfx.mec.mqd_backup[j]) {
 				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
 				return -ENOMEM;
diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c b/drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c
index 86cc90a66296..4dabcd7c9f5e 100644
--- a/drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c
+++ b/drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c
@@ -2235,7 +2235,7 @@ static int gfx_v12_1_compute_mqd_init(struct amdgpu_device *adev, void *m,
 	mqd->cp_hqd_queue_priority = prop->hqd_queue_priority;
 
 	mqd->cp_mqd_stride_size = prop->mqd_stride_size ? prop->mqd_stride_size :
-		sizeof(struct v12_1_compute_mqd);
+		AMDGPU_MQD_SIZE_ALIGN(adev->mqds[AMDGPU_HW_IP_COMPUTE].mqd_size);
 
 	mqd->cp_hqd_active = prop->hqd_active;
 
diff --git a/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c b/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
index f2309d72bbe6..0319c40c7f7a 100644
--- a/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
+++ b/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
@@ -289,7 +289,8 @@ static int mes_userq_mqd_create(struct amdgpu_usermode_queue *queue,
 		return -ENOMEM;
 	}
 
-	r = amdgpu_userq_create_object(uq_mgr, &queue->mqd, mqd_hw_default->mqd_size);
+	r = amdgpu_userq_create_object(uq_mgr, &queue->mqd,
+		AMDGPU_MQD_SIZE_ALIGN(mqd_hw_default->mqd_size));
 	if (r) {
 		DRM_ERROR("Failed to create MQD object for userqueue\n");
 		goto free_props;
-- 
2.34.1