drm/amdkfd: Make doorbell size ASIC-dependent

This prepares for GFXv9 (Vega10), which has 64-bit doorbells.
Signed-off-by: NFelix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: NOded Gabbay <oded.gabbay@gmail.com>
Signed-off-by: NOded Gabbay <oded.gabbay@gmail.com>

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

@@ -41,6 +41,7 @@ static const struct kfd_device_info kaveri_device_info = {
 	.max_pasid_bits = 16,
 	/* max num of queues for KV.TODO should be a dynamic value */
 	.max_no_of_hqd	= 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -55,6 +56,7 @@ static const struct kfd_device_info carrizo_device_info = {
 	.max_pasid_bits = 16,
 	/* max num of queues for CZ.TODO should be a dynamic value */
 	.max_no_of_hqd	= 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -70,6 +72,7 @@ static const struct kfd_device_info hawaii_device_info = {
 	.max_pasid_bits = 16,
 	/* max num of queues for KV.TODO should be a dynamic value */
 	.max_no_of_hqd	= 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -83,6 +86,7 @@ static const struct kfd_device_info tonga_device_info = {
 	.asic_family = CHIP_TONGA,
 	.max_pasid_bits = 16,
 	.max_no_of_hqd  = 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -96,6 +100,7 @@ static const struct kfd_device_info tonga_vf_device_info = {
 	.asic_family = CHIP_TONGA,
 	.max_pasid_bits = 16,
 	.max_no_of_hqd  = 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -109,6 +114,7 @@ static const struct kfd_device_info fiji_device_info = {
 	.asic_family = CHIP_FIJI,
 	.max_pasid_bits = 16,
 	.max_no_of_hqd  = 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -122,6 +128,7 @@ static const struct kfd_device_info fiji_vf_device_info = {
 	.asic_family = CHIP_FIJI,
 	.max_pasid_bits = 16,
 	.max_no_of_hqd  = 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -136,6 +143,7 @@ static const struct kfd_device_info polaris10_device_info = {
 	.asic_family = CHIP_POLARIS10,
 	.max_pasid_bits = 16,
 	.max_no_of_hqd  = 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -149,6 +157,7 @@ static const struct kfd_device_info polaris10_vf_device_info = {
 	.asic_family = CHIP_POLARIS10,
 	.max_pasid_bits = 16,
 	.max_no_of_hqd  = 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,
@@ -162,6 +171,7 @@ static const struct kfd_device_info polaris11_device_info = {
 	.asic_family = CHIP_POLARIS11,
 	.max_pasid_bits = 16,
 	.max_no_of_hqd  = 24,
+	.doorbell_size  = 4,
 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
 	.event_interrupt_class = &event_interrupt_class_cik,
 	.num_of_watch_points = 4,

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

@@ -33,7 +33,6 @@
 
 static DEFINE_IDA(doorbell_ida);
 static unsigned int max_doorbell_slices;
-#define KFD_SIZE_OF_DOORBELL_IN_BYTES 4
 
 /*
  * Each device exposes a doorbell aperture, a PCI MMIO aperture that
@@ -50,9 +49,9 @@ static unsigned int max_doorbell_slices;
  */
 
 /* # of doorbell bytes allocated for each process. */
-static inline size_t doorbell_process_allocation(void)
+static size_t kfd_doorbell_process_slice(struct kfd_dev *kfd)
 {
-	return roundup(KFD_SIZE_OF_DOORBELL_IN_BYTES *
+	return roundup(kfd->device_info->doorbell_size *
 			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
 			PAGE_SIZE);
 }
@@ -72,16 +71,16 @@ int kfd_doorbell_init(struct kfd_dev *kfd)
 
 	doorbell_start_offset =
 			roundup(kfd->shared_resources.doorbell_start_offset,
-					doorbell_process_allocation());
+					kfd_doorbell_process_slice(kfd));
 
 	doorbell_aperture_size =
 			rounddown(kfd->shared_resources.doorbell_aperture_size,
-					doorbell_process_allocation());
+					kfd_doorbell_process_slice(kfd));
 
 	if (doorbell_aperture_size > doorbell_start_offset)
 		doorbell_process_limit =
 			(doorbell_aperture_size - doorbell_start_offset) /
-						doorbell_process_allocation();
+						kfd_doorbell_process_slice(kfd);
 	else
 		return -ENOSPC;
 
@@ -95,7 +94,7 @@ int kfd_doorbell_init(struct kfd_dev *kfd)
 	kfd->doorbell_id_offset = doorbell_start_offset / sizeof(u32);
 
 	kfd->doorbell_kernel_ptr = ioremap(kfd->doorbell_base,
-						doorbell_process_allocation());
+					   kfd_doorbell_process_slice(kfd));
 
 	if (!kfd->doorbell_kernel_ptr)
 		return -ENOMEM;
@@ -132,16 +131,16 @@ int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma)
 	phys_addr_t address;
 	struct kfd_dev *dev;
 
+	/* Find kfd device according to gpu id */
+	dev = kfd_device_by_id(vma->vm_pgoff);
+	if (!dev)
+		return -EINVAL;
+
 	/*
 	 * For simplicitly we only allow mapping of the entire doorbell
 	 * allocation of a single device & process.
 	 */
-	if (vma->vm_end - vma->vm_start != doorbell_process_allocation())
-		return -EINVAL;
-
-	/* Find kfd device according to gpu id */
-	dev = kfd_device_by_id(vma->vm_pgoff);
-	if (!dev)
+	if (vma->vm_end - vma->vm_start != kfd_doorbell_process_slice(dev))
 		return -EINVAL;
 
 	/* Calculate physical address of doorbell */
@@ -158,19 +157,19 @@ int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma)
 		 "     vm_flags            == 0x%04lX\n"
 		 "     size                == 0x%04lX\n",
 		 (unsigned long long) vma->vm_start, address, vma->vm_flags,
-		 doorbell_process_allocation());
+		 kfd_doorbell_process_slice(dev));
 
 
 	return io_remap_pfn_range(vma,
 				vma->vm_start,
 				address >> PAGE_SHIFT,
-				doorbell_process_allocation(),
+				kfd_doorbell_process_slice(dev),
 				vma->vm_page_prot);
 }
 
 
 /* get kernel iomem pointer for a doorbell */
-u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
+void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
 					unsigned int *doorbell_off)
 {
 	u32 inx;
@@ -185,6 +184,8 @@ u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
 	if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
 		return NULL;
 
+	inx *= kfd->device_info->doorbell_size / sizeof(u32);
+
 	/*
 	 * Calculating the kernel doorbell offset using the first
 	 * doorbell page.
@@ -210,7 +211,7 @@ void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr)
 	mutex_unlock(&kfd->doorbell_mutex);
 }
 
-inline void write_kernel_doorbell(u32 __iomem *db, u32 value)
+void write_kernel_doorbell(void __iomem *db, u32 value)
 {
 	if (db) {
 		writel(value, db);
@@ -228,20 +229,21 @@ unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd,
 {
 	/*
 	 * doorbell_id_offset accounts for doorbells taken by KGD.
-	 * index * doorbell_process_allocation/sizeof(u32) adjusts to
-	 * the process's doorbells.
+	 * index * kfd_doorbell_process_slice/sizeof(u32) adjusts to
+	 * the process's doorbells. The offset returned is in dword
+	 * units regardless of the ASIC-dependent doorbell size.
 	 */
 	return kfd->doorbell_id_offset +
 		process->doorbell_index
-		* doorbell_process_allocation() / sizeof(u32) +
-		queue_id;
+		* kfd_doorbell_process_slice(kfd) / sizeof(u32) +
+		queue_id * kfd->device_info->doorbell_size / sizeof(u32);
 }
 
 uint64_t kfd_get_number_elems(struct kfd_dev *kfd)
 {
 	uint64_t num_of_elems = (kfd->shared_resources.doorbell_aperture_size -
 				kfd->shared_resources.doorbell_start_offset) /
-					doorbell_process_allocation() + 1;
+					kfd_doorbell_process_slice(kfd) + 1;
 
 	return num_of_elems;
 
@@ -251,7 +253,7 @@ phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
 					struct kfd_process *process)
 {
 	return dev->doorbell_base +
-		process->doorbell_index * doorbell_process_allocation();
+		process->doorbell_index * kfd_doorbell_process_slice(dev);
 }
 
 int kfd_alloc_process_doorbells(struct kfd_process *process)

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

@@ -160,6 +160,7 @@ struct kfd_device_info {
 	const struct kfd_event_interrupt_class *event_interrupt_class;
 	unsigned int max_pasid_bits;
 	unsigned int max_no_of_hqd;
+	unsigned int doorbell_size;
 	size_t ih_ring_entry_size;
 	uint8_t num_of_watch_points;
 	uint16_t mqd_size_aligned;
@@ -364,7 +365,7 @@ struct queue_properties {
 	uint32_t queue_percent;
 	uint32_t *read_ptr;
 	uint32_t *write_ptr;
-	uint32_t __iomem *doorbell_ptr;
+	void __iomem *doorbell_ptr;
 	uint32_t doorbell_off;
 	bool is_interop;
 	bool is_evicted;
@@ -728,11 +729,11 @@ void kfd_pasid_free(unsigned int pasid);
 int kfd_doorbell_init(struct kfd_dev *kfd);
 void kfd_doorbell_fini(struct kfd_dev *kfd);
 int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma);
-u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
+void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
 					unsigned int *doorbell_off);
 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
 u32 read_kernel_doorbell(u32 __iomem *db);
-void write_kernel_doorbell(u32 __iomem *db, u32 value);
+void write_kernel_doorbell(void __iomem *db, u32 value);
 unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd,
 					struct kfd_process *process,
 					unsigned int queue_id);