drm/amd: Update MEC HQD loading code for KFD

Various bug fixes and improvements that accumulated over the last two
years.

Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Oded Gabbay <oded.gabbay@gmail.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>

11 files changed, 322 insertions, 69 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index b8802a561cbd..8d689ab7e429 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -26,6 +26,7 @@
 #define AMDGPU_AMDKFD_H_INCLUDED
 
 #include <linux/types.h>
+#include <linux/mmu_context.h>
 #include <kgd_kfd_interface.h>
 
 struct amdgpu_device;
@@ -60,4 +61,19 @@ uint64_t get_gpu_clock_counter(struct kgd_dev *kgd);
 
 uint32_t get_max_engine_clock_in_mhz(struct kgd_dev *kgd);
 
+#define read_user_wptr(mmptr, wptr, dst)				\
+	({								\
+		bool valid = false;					\
+		if ((mmptr) && (wptr)) {				\
+			if ((mmptr) == current->mm) {			\
+				valid = !get_user((dst), (wptr));	\
+			} else if (current->mm == NULL) {		\
+				use_mm(mmptr);				\
+				valid = !get_user((dst), (wptr));	\
+				unuse_mm(mmptr);			\
+			}						\
+		}							\
+		valid;							\
+	})
+
 #endif /* AMDGPU_AMDKFD_H_INCLUDED */
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c
index dcd90e86d044..d1719becb4df 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c
@@ -39,6 +39,12 @@
 #include "gmc/gmc_7_1_sh_mask.h"
 #include "cik_structs.h"
 
+enum hqd_dequeue_request_type {
+	NO_ACTION = 0,
+	DRAIN_PIPE,
+	RESET_WAVES
+};
+
 enum {
 	MAX_TRAPID = 8,		/* 3 bits in the bitfield. */
 	MAX_WATCH_ADDRESSES = 4
@@ -96,12 +102,15 @@ static int kgd_init_pipeline(struct kgd_dev *kgd, uint32_t pipe_id,
 				uint32_t hpd_size, uint64_t hpd_gpu_addr);
 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id);
 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
-			uint32_t queue_id, uint32_t __user *wptr);
+			uint32_t queue_id, uint32_t __user *wptr,
+			uint32_t wptr_shift, uint32_t wptr_mask,
+			struct mm_struct *mm);
 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd);
 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
 				uint32_t pipe_id, uint32_t queue_id);
 
-static int kgd_hqd_destroy(struct kgd_dev *kgd, uint32_t reset_type,
+static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
+				enum kfd_preempt_type reset_type,
 				unsigned int utimeout, uint32_t pipe_id,
 				uint32_t queue_id);
 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd);
@@ -290,20 +299,38 @@ static inline struct cik_sdma_rlc_registers *get_sdma_mqd(void *mqd)
 }
 
 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
-			uint32_t queue_id, uint32_t __user *wptr)
+			uint32_t queue_id, uint32_t __user *wptr,
+			uint32_t wptr_shift, uint32_t wptr_mask,
+			struct mm_struct *mm)
 {
 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
-	uint32_t wptr_shadow, is_wptr_shadow_valid;
 	struct cik_mqd *m;
+	uint32_t *mqd_hqd;
+	uint32_t reg, wptr_val, data;
 
 	m = get_mqd(mqd);
 
-	is_wptr_shadow_valid = !get_user(wptr_shadow, wptr);
-	if (is_wptr_shadow_valid)
-		m->cp_hqd_pq_wptr = wptr_shadow;
-
 	acquire_queue(kgd, pipe_id, queue_id);
-	gfx_v7_0_mqd_commit(adev, m);
+
+	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_MQD_CONTROL. */
+	mqd_hqd = &m->cp_mqd_base_addr_lo;
+
+	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_MQD_CONTROL; reg++)
+		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
+
+	/* Copy userspace write pointer value to register.
+	 * Activate doorbell logic to monitor subsequent changes.
+	 */
+	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
+			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
+	WREG32(mmCP_HQD_PQ_DOORBELL_CONTROL, data);
+
+	if (read_user_wptr(mm, wptr, wptr_val))
+		WREG32(mmCP_HQD_PQ_WPTR, (wptr_val << wptr_shift) & wptr_mask);
+
+	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
+	WREG32(mmCP_HQD_ACTIVE, data);
+
 	release_queue(kgd);
 
 	return 0;
@@ -382,30 +409,99 @@ static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
 	return false;
 }
 
-static int kgd_hqd_destroy(struct kgd_dev *kgd, uint32_t reset_type,
+static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
+				enum kfd_preempt_type reset_type,
 				unsigned int utimeout, uint32_t pipe_id,
 				uint32_t queue_id)
 {
 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
 	uint32_t temp;
-	int timeout = utimeout;
+	enum hqd_dequeue_request_type type;
+	unsigned long flags, end_jiffies;
+	int retry;
 
 	acquire_queue(kgd, pipe_id, queue_id);
 	WREG32(mmCP_HQD_PQ_DOORBELL_CONTROL, 0);
 
-	WREG32(mmCP_HQD_DEQUEUE_REQUEST, reset_type);
+	switch (reset_type) {
+	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
+		type = DRAIN_PIPE;
+		break;
+	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
+		type = RESET_WAVES;
+		break;
+	default:
+		type = DRAIN_PIPE;
+		break;
+	}
 
+	/* Workaround: If IQ timer is active and the wait time is close to or
+	 * equal to 0, dequeueing is not safe. Wait until either the wait time
+	 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
+	 * cleared before continuing. Also, ensure wait times are set to at
+	 * least 0x3.
+	 */
+	local_irq_save(flags);
+	preempt_disable();
+	retry = 5000; /* wait for 500 usecs at maximum */
+	while (true) {
+		temp = RREG32(mmCP_HQD_IQ_TIMER);
+		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
+			pr_debug("HW is processing IQ\n");
+			goto loop;
+		}
+		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
+			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
+					== 3) /* SEM-rearm is safe */
+				break;
+			/* Wait time 3 is safe for CP, but our MMIO read/write
+			 * time is close to 1 microsecond, so check for 10 to
+			 * leave more buffer room
+			 */
+			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
+					>= 10)
+				break;
+			pr_debug("IQ timer is active\n");
+		} else
+			break;
+loop:
+		if (!retry) {
+			pr_err("CP HQD IQ timer status time out\n");
+			break;
+		}
+		ndelay(100);
+		--retry;
+	}
+	retry = 1000;
+	while (true) {
+		temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
+		if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
+			break;
+		pr_debug("Dequeue request is pending\n");
+
+		if (!retry) {
+			pr_err("CP HQD dequeue request time out\n");
+			break;
+		}
+		ndelay(100);
+		--retry;
+	}
+	local_irq_restore(flags);
+	preempt_enable();
+
+	WREG32(mmCP_HQD_DEQUEUE_REQUEST, type);
+
+	end_jiffies = (utimeout * HZ / 1000) + jiffies;
 	while (true) {
 		temp = RREG32(mmCP_HQD_ACTIVE);
-		if (temp & CP_HQD_ACTIVE__ACTIVE_MASK)
+		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
 			break;
-		if (timeout <= 0) {
-			pr_err("kfd: cp queue preemption time out.\n");
+		if (time_after(jiffies, end_jiffies)) {
+			pr_err("cp queue preemption time out\n");
 			release_queue(kgd);
 			return -ETIME;
 		}
-		msleep(20);
-		timeout -= 20;
+		usleep_range(500, 1000);
 	}
 
 	release_queue(kgd);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c
index 0fccd30c9f74..29a6f5d9de8e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c
@@ -39,6 +39,12 @@
 #include "vi_structs.h"
 #include "vid.h"
 
+enum hqd_dequeue_request_type {
+	NO_ACTION = 0,
+	DRAIN_PIPE,
+	RESET_WAVES
+};
+
 struct cik_sdma_rlc_registers;
 
 /*
@@ -55,12 +61,15 @@ static int kgd_init_pipeline(struct kgd_dev *kgd, uint32_t pipe_id,
 		uint32_t hpd_size, uint64_t hpd_gpu_addr);
 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id);
 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
-		uint32_t queue_id, uint32_t __user *wptr);
+			uint32_t queue_id, uint32_t __user *wptr,
+			uint32_t wptr_shift, uint32_t wptr_mask,
+			struct mm_struct *mm);
 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd);
 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
 		uint32_t pipe_id, uint32_t queue_id);
 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd);
-static int kgd_hqd_destroy(struct kgd_dev *kgd, uint32_t reset_type,
+static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
+				enum kfd_preempt_type reset_type,
 				unsigned int utimeout, uint32_t pipe_id,
 				uint32_t queue_id);
 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
@@ -244,20 +253,67 @@ static inline struct cik_sdma_rlc_registers *get_sdma_mqd(void *mqd)
 }
 
 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
-			uint32_t queue_id, uint32_t __user *wptr)
+			uint32_t queue_id, uint32_t __user *wptr,
+			uint32_t wptr_shift, uint32_t wptr_mask,
+			struct mm_struct *mm)
 {
-	struct vi_mqd *m;
-	uint32_t shadow_wptr, valid_wptr;
 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
+	struct vi_mqd *m;
+	uint32_t *mqd_hqd;
+	uint32_t reg, wptr_val, data;
 
 	m = get_mqd(mqd);
 
-	valid_wptr = copy_from_user(&shadow_wptr, wptr, sizeof(shadow_wptr));
-	if (valid_wptr == 0)
-		m->cp_hqd_pq_wptr = shadow_wptr;
-
 	acquire_queue(kgd, pipe_id, queue_id);
-	gfx_v8_0_mqd_commit(adev, mqd);
+
+	/* HIQ is set during driver init period with vmid set to 0*/
+	if (m->cp_hqd_vmid == 0) {
+		uint32_t value, mec, pipe;
+
+		mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
+		pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
+
+		pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
+			mec, pipe, queue_id);
+		value = RREG32(mmRLC_CP_SCHEDULERS);
+		value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
+			((mec << 5) | (pipe << 3) | queue_id | 0x80));
+		WREG32(mmRLC_CP_SCHEDULERS, value);
+	}
+
+	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
+	mqd_hqd = &m->cp_mqd_base_addr_lo;
+
+	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_HQD_EOP_CONTROL; reg++)
+		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
+
+	/* Tonga errata: EOP RPTR/WPTR should be left unmodified.
+	 * This is safe since EOP RPTR==WPTR for any inactive HQD
+	 * on ASICs that do not support context-save.
+	 * EOP writes/reads can start anywhere in the ring.
+	 */
+	if (get_amdgpu_device(kgd)->asic_type != CHIP_TONGA) {
+		WREG32(mmCP_HQD_EOP_RPTR, m->cp_hqd_eop_rptr);
+		WREG32(mmCP_HQD_EOP_WPTR, m->cp_hqd_eop_wptr);
+		WREG32(mmCP_HQD_EOP_WPTR_MEM, m->cp_hqd_eop_wptr_mem);
+	}
+
+	for (reg = mmCP_HQD_EOP_EVENTS; reg <= mmCP_HQD_ERROR; reg++)
+		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
+
+	/* Copy userspace write pointer value to register.
+	 * Activate doorbell logic to monitor subsequent changes.
+	 */
+	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
+			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
+	WREG32(mmCP_HQD_PQ_DOORBELL_CONTROL, data);
+
+	if (read_user_wptr(mm, wptr, wptr_val))
+		WREG32(mmCP_HQD_PQ_WPTR, (wptr_val << wptr_shift) & wptr_mask);
+
+	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
+	WREG32(mmCP_HQD_ACTIVE, data);
+
 	release_queue(kgd);
 
 	return 0;
@@ -308,29 +364,102 @@ static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
 	return false;
 }
 
-static int kgd_hqd_destroy(struct kgd_dev *kgd, uint32_t reset_type,
+static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
+				enum kfd_preempt_type reset_type,
 				unsigned int utimeout, uint32_t pipe_id,
 				uint32_t queue_id)
 {
 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
 	uint32_t temp;
-	int timeout = utimeout;
+	enum hqd_dequeue_request_type type;
+	unsigned long flags, end_jiffies;
+	int retry;
+	struct vi_mqd *m = get_mqd(mqd);
 
 	acquire_queue(kgd, pipe_id, queue_id);
 
-	WREG32(mmCP_HQD_DEQUEUE_REQUEST, reset_type);
+	if (m->cp_hqd_vmid == 0)
+		WREG32_FIELD(RLC_CP_SCHEDULERS, scheduler1, 0);
+
+	switch (reset_type) {
+	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
+		type = DRAIN_PIPE;
+		break;
+	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
+		type = RESET_WAVES;
+		break;
+	default:
+		type = DRAIN_PIPE;
+		break;
+	}
+
+	/* Workaround: If IQ timer is active and the wait time is close to or
+	 * equal to 0, dequeueing is not safe. Wait until either the wait time
+	 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
+	 * cleared before continuing. Also, ensure wait times are set to at
+	 * least 0x3.
+	 */
+	local_irq_save(flags);
+	preempt_disable();
+	retry = 5000; /* wait for 500 usecs at maximum */
+	while (true) {
+		temp = RREG32(mmCP_HQD_IQ_TIMER);
+		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
+			pr_debug("HW is processing IQ\n");
+			goto loop;
+		}
+		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
+			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
+					== 3) /* SEM-rearm is safe */
+				break;
+			/* Wait time 3 is safe for CP, but our MMIO read/write
+			 * time is close to 1 microsecond, so check for 10 to
+			 * leave more buffer room
+			 */
+			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
+					>= 10)
+				break;
+			pr_debug("IQ timer is active\n");
+		} else
+			break;
+loop:
+		if (!retry) {
+			pr_err("CP HQD IQ timer status time out\n");
+			break;
+		}
+		ndelay(100);
+		--retry;
+	}
+	retry = 1000;
+	while (true) {
+		temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
+		if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
+			break;
+		pr_debug("Dequeue request is pending\n");
 
+		if (!retry) {
+			pr_err("CP HQD dequeue request time out\n");
+			break;
+		}
+		ndelay(100);
+		--retry;
+	}
+	local_irq_restore(flags);
+	preempt_enable();
+
+	WREG32(mmCP_HQD_DEQUEUE_REQUEST, type);
+
+	end_jiffies = (utimeout * HZ / 1000) + jiffies;
 	while (true) {
 		temp = RREG32(mmCP_HQD_ACTIVE);
-		if (temp & CP_HQD_ACTIVE__ACTIVE_MASK)
+		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
 			break;
-		if (timeout <= 0) {
-			pr_err("kfd: cp queue preemption time out.\n");
+		if (time_after(jiffies, end_jiffies)) {
+			pr_err("cp queue preemption time out.\n");
 			release_queue(kgd);
 			return -ETIME;
 		}
-		msleep(20);
-		timeout -= 20;
+		usleep_range(500, 1000);
 	}
 
 	release_queue(kgd);
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 e553c5e45264..618ac65b6136 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
@@ -270,8 +270,8 @@ static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
 	pr_debug("Loading mqd to hqd on pipe %d, queue %d\n",
 			q->pipe, q->queue);
 
-	retval = mqd->load_mqd(mqd, q->mqd, q->pipe,
-			q->queue, (uint32_t __user *) q->properties.write_ptr);
+	retval = mqd->load_mqd(mqd, q->mqd, q->pipe, q->queue, &q->properties,
+			       q->process->mm);
 	if (retval)
 		goto out_uninit_mqd;
 
@@ -587,8 +587,7 @@ static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
 	if (retval)
 		goto out_deallocate_sdma_queue;
 
-	retval = mqd->load_mqd(mqd, q->mqd, 0,
-				0, NULL);
+	retval = mqd->load_mqd(mqd, q->mqd, 0, 0, &q->properties, NULL);
 	if (retval)
 		goto out_uninit_mqd;
 
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
index 0e4d4a98dc2b..681b639f5133 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_kernel_queue.c
@@ -143,7 +143,8 @@ static bool initialize(struct kernel_queue *kq, struct kfd_dev *dev,
 		kq->queue->pipe = KFD_CIK_HIQ_PIPE;
 		kq->queue->queue = KFD_CIK_HIQ_QUEUE;
 		kq->mqd->load_mqd(kq->mqd, kq->queue->mqd, kq->queue->pipe,
-					kq->queue->queue, NULL);
+				  kq->queue->queue, &kq->queue->properties,
+				  NULL);
 	} else {
 		/* allocate fence for DIQ */
 
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h
index 213a71e0b6c7..1f3a6ba7eed2 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h
@@ -67,7 +67,8 @@ struct mqd_manager {
 
 	int	(*load_mqd)(struct mqd_manager *mm, void *mqd,
 				uint32_t pipe_id, uint32_t queue_id,
-				uint32_t __user *wptr);
+				struct queue_properties *p,
+				struct mm_struct *mms);
 
 	int	(*update_mqd)(struct mqd_manager *mm, void *mqd,
 				struct queue_properties *q);
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 7e0ec6bb1637..44ffd23348fc 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c
@@ -144,15 +144,21 @@ static void uninit_mqd_sdma(struct mqd_manager *mm, void *mqd,
 }
 
 static int load_mqd(struct mqd_manager *mm, void *mqd, uint32_t pipe_id,
-			uint32_t queue_id, uint32_t __user *wptr)
+		    uint32_t queue_id, struct queue_properties *p,
+		    struct mm_struct *mms)
 {
-	return mm->dev->kfd2kgd->hqd_load
-		(mm->dev->kgd, mqd, pipe_id, queue_id, wptr);
+	/* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */
+	uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0);
+	uint32_t wptr_mask = (uint32_t)((p->queue_size / sizeof(uint32_t)) - 1);
+
+	return mm->dev->kfd2kgd->hqd_load(mm->dev->kgd, mqd, pipe_id, queue_id,
+					  (uint32_t __user *)p->write_ptr,
+					  wptr_shift, wptr_mask, mms);
 }
 
 static int load_mqd_sdma(struct mqd_manager *mm, void *mqd,
-			uint32_t pipe_id, uint32_t queue_id,
-			uint32_t __user *wptr)
+			 uint32_t pipe_id, uint32_t queue_id,
+			 struct queue_properties *p, struct mm_struct *mms)
 {
 	return mm->dev->kfd2kgd->hqd_sdma_load(mm->dev->kgd, mqd);
 }
@@ -176,20 +182,17 @@ static int update_mqd(struct mqd_manager *mm, void *mqd,
 	m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8);
 	m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
 	m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
-	m->cp_hqd_pq_doorbell_control = DOORBELL_EN |
-					DOORBELL_OFFSET(q->doorbell_off);
+	m->cp_hqd_pq_doorbell_control = DOORBELL_OFFSET(q->doorbell_off);
 
 	m->cp_hqd_vmid = q->vmid;
 
 	if (q->format == KFD_QUEUE_FORMAT_AQL)
 		m->cp_hqd_pq_control |= NO_UPDATE_RPTR;
 
-	m->cp_hqd_active = 0;
 	q->is_active = false;
 	if (q->queue_size > 0 &&
 			q->queue_address != 0 &&
 			q->queue_percent > 0) {
-		m->cp_hqd_active = 1;
 		q->is_active = true;
 	}
 
@@ -239,7 +242,7 @@ static int destroy_mqd(struct mqd_manager *mm, void *mqd,
 			unsigned int timeout, uint32_t pipe_id,
 			uint32_t queue_id)
 {
-	return mm->dev->kfd2kgd->hqd_destroy(mm->dev->kgd, type, timeout,
+	return mm->dev->kfd2kgd->hqd_destroy(mm->dev->kgd, mqd, type, timeout,
 					pipe_id, queue_id);
 }
 
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 98a930e9720f..73cbfe186dd2 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c
@@ -94,10 +94,15 @@ static int init_mqd(struct mqd_manager *mm, void **mqd,
 
 static int load_mqd(struct mqd_manager *mm, void *mqd,
 			uint32_t pipe_id, uint32_t queue_id,
-			uint32_t __user *wptr)
+			struct queue_properties *p, struct mm_struct *mms)
 {
-	return mm->dev->kfd2kgd->hqd_load
-		(mm->dev->kgd, mqd, pipe_id, queue_id, wptr);
+	/* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */
+	uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0);
+	uint32_t wptr_mask = (uint32_t)((p->queue_size / sizeof(uint32_t)) - 1);
+
+	return mm->dev->kfd2kgd->hqd_load(mm->dev->kgd, mqd, pipe_id, queue_id,
+					  (uint32_t __user *)p->write_ptr,
+					  wptr_shift, wptr_mask, mms);
 }
 
 static int __update_mqd(struct mqd_manager *mm, void *mqd,
@@ -122,7 +127,6 @@ static int __update_mqd(struct mqd_manager *mm, void *mqd,
 	m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
 
 	m->cp_hqd_pq_doorbell_control =
-		1 << CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_EN__SHIFT |
 		q->doorbell_off <<
 			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
 	pr_debug("cp_hqd_pq_doorbell_control 0x%x\n",
@@ -159,12 +163,10 @@ static int __update_mqd(struct mqd_manager *mm, void *mqd,
 				2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT;
 	}
 
-	m->cp_hqd_active = 0;
 	q->is_active = false;
 	if (q->queue_size > 0 &&
 			q->queue_address != 0 &&
 			q->queue_percent > 0) {
-		m->cp_hqd_active = 1;
 		q->is_active = true;
 	}
 
@@ -184,7 +186,7 @@ static int destroy_mqd(struct mqd_manager *mm, void *mqd,
 			uint32_t queue_id)
 {
 	return mm->dev->kfd2kgd->hqd_destroy
-		(mm->dev->kgd, type, timeout,
+		(mm->dev->kgd, mqd, type, timeout,
 		pipe_id, queue_id);
 }
 
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index f0d55cc050a5..30ce92c6e6a1 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -239,11 +239,6 @@ enum kfd_preempt_type_filter {
 	KFD_PREEMPT_TYPE_FILTER_BY_PASID
 };
 
-enum kfd_preempt_type {
-	KFD_PREEMPT_TYPE_WAVEFRONT,
-	KFD_PREEMPT_TYPE_WAVEFRONT_RESET
-};
-
 /**
  * enum kfd_queue_type
  *
diff --git a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
index 36f376677a53..ffafda0ec620 100644
--- a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
+++ b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
@@ -41,6 +41,11 @@ struct kgd_dev;
 
 struct kgd_mem;
 
+enum kfd_preempt_type {
+	KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN = 0,
+	KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
+};
+
 enum kgd_memory_pool {
 	KGD_POOL_SYSTEM_CACHEABLE = 1,
 	KGD_POOL_SYSTEM_WRITECOMBINE = 2,
@@ -153,14 +158,16 @@ struct kfd2kgd_calls {
 	int (*init_interrupts)(struct kgd_dev *kgd, uint32_t pipe_id);
 
 	int (*hqd_load)(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
-			uint32_t queue_id, uint32_t __user *wptr);
+			uint32_t queue_id, uint32_t __user *wptr,
+			uint32_t wptr_shift, uint32_t wptr_mask,
+			struct mm_struct *mm);
 
 	int (*hqd_sdma_load)(struct kgd_dev *kgd, void *mqd);
 
 	bool (*hqd_is_occupied)(struct kgd_dev *kgd, uint64_t queue_address,
 				uint32_t pipe_id, uint32_t queue_id);
 
-	int (*hqd_destroy)(struct kgd_dev *kgd, uint32_t reset_type,
+	int (*hqd_destroy)(struct kgd_dev *kgd, void *mqd, uint32_t reset_type,
 				unsigned int timeout, uint32_t pipe_id,
 				uint32_t queue_id);
 
diff --git a/drivers/gpu/drm/radeon/radeon_kfd.c b/drivers/gpu/drm/radeon/radeon_kfd.c
index cab6c0385040..f6578c96925c 100644
--- a/drivers/gpu/drm/radeon/radeon_kfd.c
+++ b/drivers/gpu/drm/radeon/radeon_kfd.c
@@ -75,12 +75,14 @@ static int kgd_init_pipeline(struct kgd_dev *kgd, uint32_t pipe_id,
 				uint32_t hpd_size, uint64_t hpd_gpu_addr);
 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id);
 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
-			uint32_t queue_id, uint32_t __user *wptr);
+			uint32_t queue_id, uint32_t __user *wptr,
+			uint32_t wptr_shift, uint32_t wptr_mask,
+			struct mm_struct *mm);
 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd);
 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
 				uint32_t pipe_id, uint32_t queue_id);
 
-static int kgd_hqd_destroy(struct kgd_dev *kgd, uint32_t reset_type,
+static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd, uint32_t reset_type,
 				unsigned int timeout, uint32_t pipe_id,
 				uint32_t queue_id);
 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd);
@@ -482,7 +484,9 @@ static inline struct cik_sdma_rlc_registers *get_sdma_mqd(void *mqd)
 }
 
 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
-			uint32_t queue_id, uint32_t __user *wptr)
+			uint32_t queue_id, uint32_t __user *wptr,
+			uint32_t wptr_shift, uint32_t wptr_mask,
+			struct mm_struct *mm)
 {
 	uint32_t wptr_shadow, is_wptr_shadow_valid;
 	struct cik_mqd *m;
@@ -636,7 +640,7 @@ static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
 	return false;
 }
 
-static int kgd_hqd_destroy(struct kgd_dev *kgd, uint32_t reset_type,
+static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd, uint32_t reset_type,
 				unsigned int timeout, uint32_t pipe_id,
 				uint32_t queue_id)
 {