drm/amdkfd: allow split HQD on per-queue granularity v5

Update the KGD to KFD interface to allow sharing pipes with queue
granularity instead of pipe granularity.

This allows for more interesting pipe/queue splits.

v2: fix overflow check for res.queue_mask
v3: fix shift overflow when setting res.queue_mask
v4: fix comment in is_pipeline_enabled()
v5: clamp res.queue_mask to the first MEC only

Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c

@@ -95,14 +95,30 @@ void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev)
 
 void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
 {
+	int i;
+	int last_valid_bit;
 	if (adev->kfd) {
 		struct kgd2kfd_shared_resources gpu_resources = {
 			.compute_vmid_bitmap = 0xFF00,
-
+			.num_mec = adev->gfx.mec.num_mec,
-			.first_compute_pipe = 1,
+			.num_pipe_per_mec = adev->gfx.mec.num_pipe_per_mec,
-			.compute_pipe_count = 4 - 1,
+			.num_queue_per_pipe = adev->gfx.mec.num_queue_per_pipe
 		};
 
+		/* this is going to have a few of the MSBs set that we need to
+		 * clear */
+		bitmap_complement(gpu_resources.queue_bitmap,
+				  adev->gfx.mec.queue_bitmap,
+				  KGD_MAX_QUEUES);
+
+		/* According to linux/bitmap.h we shouldn't use bitmap_clear if
+		 * nbits is not compile time constant */
+		last_valid_bit = adev->gfx.mec.num_mec
+				* adev->gfx.mec.num_pipe_per_mec
+				* adev->gfx.mec.num_queue_per_pipe;
+		for (i = last_valid_bit; i < KGD_MAX_QUEUES; ++i)
+			clear_bit(i, gpu_resources.queue_bitmap);
+
 		amdgpu_doorbell_get_kfd_info(adev,
 				&gpu_resources.doorbell_physical_address,
 				&gpu_resources.doorbell_aperture_size,

drivers/gpu/drm/amd/amdkfd/kfd_device.c

@@ -226,6 +226,10 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd,
 
 	kfd->shared_resources = *gpu_resources;
 
+	/* We only use the first MEC */
+	if (kfd->shared_resources.num_mec > 1)
+		kfd->shared_resources.num_mec = 1;
+
 	/* calculate max size of mqds needed for queues */
 	size = max_num_of_queues_per_device *
 			kfd->device_info->mqd_size_aligned;

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c

@@ -63,21 +63,44 @@ enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type)
 	return KFD_MQD_TYPE_CP;
 }
 
-unsigned int get_first_pipe(struct device_queue_manager *dqm)
+static bool is_pipe_enabled(struct device_queue_manager *dqm, int mec, int pipe)
 {
-	BUG_ON(!dqm || !dqm->dev);
+	int i;
-	return dqm->dev->shared_resources.first_compute_pipe;
+	int pipe_offset = mec * dqm->dev->shared_resources.num_pipe_per_mec
+		+ pipe * dqm->dev->shared_resources.num_queue_per_pipe;
+
+	/* queue is available for KFD usage if bit is 1 */
+	for (i = 0; i <  dqm->dev->shared_resources.num_queue_per_pipe; ++i)
+		if (test_bit(pipe_offset + i,
+			      dqm->dev->shared_resources.queue_bitmap))
+			return true;
+	return false;
 }
 
-unsigned int get_pipes_num(struct device_queue_manager *dqm)
+unsigned int get_mec_num(struct device_queue_manager *dqm)
 {
 	BUG_ON(!dqm || !dqm->dev);
-	return dqm->dev->shared_resources.compute_pipe_count;
+
+	return dqm->dev->shared_resources.num_mec;
 }
 
-static inline unsigned int get_pipes_num_cpsch(void)
+unsigned int get_queues_num(struct device_queue_manager *dqm)
 {
-	return PIPE_PER_ME_CP_SCHEDULING;
+	BUG_ON(!dqm || !dqm->dev);
+	return bitmap_weight(dqm->dev->shared_resources.queue_bitmap,
+				KGD_MAX_QUEUES);
+}
+
+unsigned int get_queues_per_pipe(struct device_queue_manager *dqm)
+{
+	BUG_ON(!dqm || !dqm->dev);
+	return dqm->dev->shared_resources.num_queue_per_pipe;
+}
+
+unsigned int get_pipes_per_mec(struct device_queue_manager *dqm)
+{
+	BUG_ON(!dqm || !dqm->dev);
+	return dqm->dev->shared_resources.num_pipe_per_mec;
 }
 
 void program_sh_mem_settings(struct device_queue_manager *dqm,
@@ -200,12 +223,16 @@ static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
 
 	set = false;
 
-	for (pipe = dqm->next_pipe_to_allocate, i = 0; i < get_pipes_num(dqm);
+	for (pipe = dqm->next_pipe_to_allocate, i = 0; i < get_pipes_per_mec(dqm);
-			pipe = ((pipe + 1) % get_pipes_num(dqm)), ++i) {
+			pipe = ((pipe + 1) % get_pipes_per_mec(dqm)), ++i) {
+
+		if (!is_pipe_enabled(dqm, 0, pipe))
+			continue;
+
 		if (dqm->allocated_queues[pipe] != 0) {
 			bit = find_first_bit(
 				(unsigned long *)&dqm->allocated_queues[pipe],
-				QUEUES_PER_PIPE);
+				get_queues_per_pipe(dqm));
 
 			clear_bit(bit,
 				(unsigned long *)&dqm->allocated_queues[pipe]);
@@ -222,7 +249,7 @@ static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
 	pr_debug("kfd: DQM %s hqd slot - pipe (%d) queue(%d)\n",
 				__func__, q->pipe, q->queue);
 	/* horizontal hqd allocation */
-	dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_num(dqm);
+	dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_per_mec(dqm);
 
 	return 0;
 }
@@ -469,36 +496,25 @@ set_pasid_vmid_mapping(struct device_queue_manager *dqm, unsigned int pasid,
 						vmid);
 }
 
-int init_pipelines(struct device_queue_manager *dqm,
-			unsigned int pipes_num, unsigned int first_pipe)
-{
-	BUG_ON(!dqm || !dqm->dev);
-
-	pr_debug("kfd: In func %s\n", __func__);
-
-	return 0;
-}
-
 static void init_interrupts(struct device_queue_manager *dqm)
 {
 	unsigned int i;
 
 	BUG_ON(dqm == NULL);
 
-	for (i = 0 ; i < get_pipes_num(dqm) ; i++)
+	for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++)
-		dqm->dev->kfd2kgd->init_interrupts(dqm->dev->kgd,
+		if (is_pipe_enabled(dqm, 0, i))
-				i + get_first_pipe(dqm));
+			dqm->dev->kfd2kgd->init_interrupts(dqm->dev->kgd, i);
 }
 
 static int init_scheduler(struct device_queue_manager *dqm)
 {
-	int retval;
+	int retval = 0;
 
 	BUG_ON(!dqm);
 
 	pr_debug("kfd: In %s\n", __func__);
 
-	retval = init_pipelines(dqm, get_pipes_num(dqm), get_first_pipe(dqm));
 	return retval;
 }
 
@@ -509,21 +525,21 @@ static int initialize_nocpsch(struct device_queue_manager *dqm)
 	BUG_ON(!dqm);
 
 	pr_debug("kfd: In func %s num of pipes: %d\n",
-			__func__, get_pipes_num(dqm));
+			__func__, get_pipes_per_mec(dqm));
 
 	mutex_init(&dqm->lock);
 	INIT_LIST_HEAD(&dqm->queues);
 	dqm->queue_count = dqm->next_pipe_to_allocate = 0;
 	dqm->sdma_queue_count = 0;
-	dqm->allocated_queues = kcalloc(get_pipes_num(dqm),
+	dqm->allocated_queues = kcalloc(get_pipes_per_mec(dqm),
 					sizeof(unsigned int), GFP_KERNEL);
 	if (!dqm->allocated_queues) {
 		mutex_destroy(&dqm->lock);
 		return -ENOMEM;
 	}
 
-	for (i = 0; i < get_pipes_num(dqm); i++)
+	for (i = 0; i < get_pipes_per_mec(dqm); i++)
-		dqm->allocated_queues[i] = (1 << QUEUES_PER_PIPE) - 1;
+		dqm->allocated_queues[i] = (1 << get_queues_per_pipe(dqm)) - 1;
 
 	dqm->vmid_bitmap = (1 << VMID_PER_DEVICE) - 1;
 	dqm->sdma_bitmap = (1 << CIK_SDMA_QUEUES) - 1;
@@ -630,18 +646,38 @@ static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
 
 static int set_sched_resources(struct device_queue_manager *dqm)
 {
+	int i, mec;
 	struct scheduling_resources res;
-	unsigned int queue_num, queue_mask;
 
 	BUG_ON(!dqm);
 
 	pr_debug("kfd: In func %s\n", __func__);
 
-	queue_num = get_pipes_num_cpsch() * QUEUES_PER_PIPE;
-	queue_mask = (1 << queue_num) - 1;
 	res.vmid_mask = (1 << VMID_PER_DEVICE) - 1;
 	res.vmid_mask <<= KFD_VMID_START_OFFSET;
-	res.queue_mask = queue_mask << (get_first_pipe(dqm) * QUEUES_PER_PIPE);
+
+	res.queue_mask = 0;
+	for (i = 0; i < KGD_MAX_QUEUES; ++i) {
+		mec = (i / dqm->dev->shared_resources.num_queue_per_pipe)
+			/ dqm->dev->shared_resources.num_pipe_per_mec;
+
+		if (!test_bit(i, dqm->dev->shared_resources.queue_bitmap))
+			continue;
+
+		/* only acquire queues from the first MEC */
+		if (mec > 0)
+			continue;
+
+		/* This situation may be hit in the future if a new HW
+		 * generation exposes more than 64 queues. If so, the
+		 * definition of res.queue_mask needs updating */
+		if (WARN_ON(i > (sizeof(res.queue_mask)*8))) {
+			pr_err("Invalid queue enabled by amdgpu: %d\n", i);
+			break;
+		}
+
+		res.queue_mask |= (1ull << i);
+	}
 	res.gws_mask = res.oac_mask = res.gds_heap_base =
 						res.gds_heap_size = 0;
 
@@ -660,7 +696,7 @@ static int initialize_cpsch(struct device_queue_manager *dqm)
 	BUG_ON(!dqm);
 
 	pr_debug("kfd: In func %s num of pipes: %d\n",
-			__func__, get_pipes_num_cpsch());
+			__func__, get_pipes_per_mec(dqm));
 
 	mutex_init(&dqm->lock);
 	INIT_LIST_HEAD(&dqm->queues);

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h

@@ -30,8 +30,6 @@
 #include "kfd_mqd_manager.h"
 
 #define QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS	(500)
-#define QUEUES_PER_PIPE				(8)
-#define PIPE_PER_ME_CP_SCHEDULING		(3)
 #define CIK_VMID_NUM				(8)
 #define KFD_VMID_START_OFFSET			(8)
 #define VMID_PER_DEVICE				CIK_VMID_NUM
@@ -182,10 +180,10 @@ void device_queue_manager_init_cik(struct device_queue_manager_asic_ops *ops);
 void device_queue_manager_init_vi(struct device_queue_manager_asic_ops *ops);
 void program_sh_mem_settings(struct device_queue_manager *dqm,
 					struct qcm_process_device *qpd);
-int init_pipelines(struct device_queue_manager *dqm,
+unsigned int get_mec_num(struct device_queue_manager *dqm);
-		unsigned int pipes_num, unsigned int first_pipe);
+unsigned int get_queues_num(struct device_queue_manager *dqm);
-unsigned int get_first_pipe(struct device_queue_manager *dqm);
+unsigned int get_queues_per_pipe(struct device_queue_manager *dqm);
-unsigned int get_pipes_num(struct device_queue_manager *dqm);
+unsigned int get_pipes_per_mec(struct device_queue_manager *dqm);
 
 static inline unsigned int get_sh_mem_bases_32(struct kfd_process_device *pdd)
 {

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_cik.c

@@ -151,5 +151,5 @@ static void init_sdma_vm(struct device_queue_manager *dqm, struct queue *q,
 
 static int initialize_cpsch_cik(struct device_queue_manager *dqm)
 {
-	return init_pipelines(dqm, get_pipes_num(dqm), get_first_pipe(dqm));
+	return 0;
 }

drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c

@@ -65,8 +65,7 @@ static void pm_calc_rlib_size(struct packet_manager *pm,
 
 	/* check if there is over subscription*/
 	*over_subscription = false;
-	if ((process_count > 1) ||
+	if ((process_count > 1) || queue_count > get_queues_num(pm->dqm)) {
-		queue_count > PIPE_PER_ME_CP_SCHEDULING * QUEUES_PER_PIPE) {
 		*over_subscription = true;
 		pr_debug("kfd: over subscribed runlist\n");
 	}

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c

@@ -209,7 +209,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
 		/* check if there is over subscription */
 		if ((sched_policy == KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION) &&
 		((dev->dqm->processes_count >= VMID_PER_DEVICE) ||
-		(dev->dqm->queue_count >= PIPE_PER_ME_CP_SCHEDULING * QUEUES_PER_PIPE))) {
+		(dev->dqm->queue_count >= get_queues_num(dev->dqm)))) {
 			pr_err("kfd: over-subscription is not allowed in radeon_kfd.sched_policy == 1\n");
 			retval = -EPERM;
 			goto err_create_queue;

drivers/gpu/drm/amd/include/kgd_kfd_interface.h

@@ -29,10 +29,11 @@
 #define KGD_KFD_INTERFACE_H_INCLUDED
 
 #include <linux/types.h>
+#include <linux/bitmap.h>
 
 struct pci_dev;
 
-#define KFD_INTERFACE_VERSION 1
+#define KFD_INTERFACE_VERSION 2
 #define KGD_MAX_QUEUES 128
 
 struct kfd_dev;
@@ -62,11 +63,17 @@ struct kgd2kfd_shared_resources {
 	/* Bit n == 1 means VMID n is available for KFD. */
 	unsigned int compute_vmid_bitmap;
 
-	/* Compute pipes are counted starting from MEC0/pipe0 as 0. */
+	/* number of mec available from the hardware */
-	unsigned int first_compute_pipe;
+	uint32_t num_mec;
 
-	/* Number of MEC pipes available for KFD. */
+	/* number of pipes per mec */
-	unsigned int compute_pipe_count;
+	uint32_t num_pipe_per_mec;
+
+	/* number of queues per pipe */
+	uint32_t num_queue_per_pipe;
+
+	/* Bit n == 1 means Queue n is available for KFD */
+	DECLARE_BITMAP(queue_bitmap, KGD_MAX_QUEUES);
 
 	/* Base address of doorbell aperture. */
 	phys_addr_t doorbell_physical_address;

drivers/gpu/drm/radeon/radeon_kfd.c

@@ -179,14 +179,29 @@ void radeon_kfd_device_probe(struct radeon_device *rdev)
 
 void radeon_kfd_device_init(struct radeon_device *rdev)
 {
+	int i, queue, pipe, mec;
+
 	if (rdev->kfd) {
 		struct kgd2kfd_shared_resources gpu_resources = {
 			.compute_vmid_bitmap = 0xFF00,
-
+			.num_mec = 1,
-			.first_compute_pipe = 1,
+			.num_pipe_per_mec = 4,
-			.compute_pipe_count = 4 - 1,
+			.num_queue_per_pipe = 8
 		};
 
+		bitmap_zero(gpu_resources.queue_bitmap, KGD_MAX_QUEUES);
+
+		for (i = 0; i < KGD_MAX_QUEUES; ++i) {
+			queue = i % gpu_resources.num_queue_per_pipe;
+			pipe = (i / gpu_resources.num_queue_per_pipe)
+				% gpu_resources.num_pipe_per_mec;
+			mec = (i / gpu_resources.num_queue_per_pipe)
+				/ gpu_resources.num_pipe_per_mec;
+
+			if (mec == 0 && pipe > 0)
+				set_bit(i, gpu_resources.queue_bitmap);
+		}
+
 		radeon_doorbell_get_kfd_info(rdev,
 				&gpu_resources.doorbell_physical_address,
 				&gpu_resources.doorbell_aperture_size,