amdkfd: Add basic modules to amdkfd (19f6d2a6) · Commits · e / devices / android_kernel_oneplus_sm7250

drivers/gpu/drm/amd/amdkfd/Makefile

+3 −1

Original line number	Original line	Diff line number	Diff line
	@@ -4,6 +4,8 @@

	ccflags-y := -Iinclude/drm -Idrivers/gpu/drm/amd/include/		ccflags-y := -Iinclude/drm -Idrivers/gpu/drm/amd/include/

	amdkfd-y := kfd_module.o kfd_device.o kfd_chardev.o kfd_topology.o		amdkfd-y := kfd_module.o kfd_device.o kfd_chardev.o kfd_topology.o \
			kfd_pasid.o kfd_doorbell.o kfd_flat_memory.o \
			kfd_process.o

	obj-$(CONFIG_HSA_AMD) += amdkfd.o		obj-$(CONFIG_HSA_AMD) += amdkfd.o

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

+29 −2

Original line number	Original line	Diff line number	Diff line
	@@ -38,6 +38,7 @@

	static long kfd_ioctl(struct file *, unsigned int, unsigned long);		static long kfd_ioctl(struct file *, unsigned int, unsigned long);
	static int kfd_open(struct inode , struct file );		static int kfd_open(struct inode , struct file );
			static int kfd_mmap(struct file , struct vm_area_struct );

	static const char kfd_dev_name[] = "kfd";		static const char kfd_dev_name[] = "kfd";

	@@ -46,6 +47,7 @@ static const struct file_operations kfd_fops = {
	.unlocked_ioctl = kfd_ioctl,		.unlocked_ioctl = kfd_ioctl,
	.compat_ioctl = kfd_ioctl,		.compat_ioctl = kfd_ioctl,
	.open = kfd_open,		.open = kfd_open,
			.mmap = kfd_mmap,
	};		};

	static int kfd_char_dev_major = -1;		static int kfd_char_dev_major = -1;
	@@ -98,9 +100,22 @@ struct device *kfd_chardev(void)

	static int kfd_open(struct inode inode, struct file filep)		static int kfd_open(struct inode inode, struct file filep)
	{		{
			struct kfd_process *process;

	if (iminor(inode) != 0)		if (iminor(inode) != 0)
	return -ENODEV;		return -ENODEV;

			process = kfd_create_process(current);
			if (IS_ERR(process))
			return PTR_ERR(process);

			process->is_32bit_user_mode = is_compat_task();

			dev_dbg(kfd_device, "process %d opened, compat mode (32 bit) - %d\n",
			process->pasid, process->is_32bit_user_mode);

			kfd_init_apertures(process);

	return 0;		return 0;
	}		}

	@@ -156,8 +171,9 @@ static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)
	"ioctl cmd 0x%x (#%d), arg 0x%lx\n",		"ioctl cmd 0x%x (#%d), arg 0x%lx\n",
	cmd, _IOC_NR(cmd), arg);		cmd, _IOC_NR(cmd), arg);

	/* TODO: add function that retrieves process */		process = kfd_get_process(current);
	process = NULL;		if (IS_ERR(process))
			return PTR_ERR(process);

	switch (cmd) {		switch (cmd) {
	case KFD_IOC_GET_VERSION:		case KFD_IOC_GET_VERSION:
	@@ -208,3 +224,14 @@ static long kfd_ioctl(struct file *filep, unsigned int cmd, unsigned long arg)

	return err;		return err;
	}		}

			static int kfd_mmap(struct file filp, struct vm_area_struct vma)
			{
			struct kfd_process *process;

			process = kfd_get_process(current);
			if (IS_ERR(process))
			return PTR_ERR(process);

			return kfd_doorbell_mmap(process, vma);
			}

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

+40 −6

Original line number	Original line	Diff line number	Diff line
	@@ -26,8 +26,11 @@
	#include <linux/slab.h>		#include <linux/slab.h>
	#include "kfd_priv.h"		#include "kfd_priv.h"

			#define MQD_SIZE_ALIGNED 768

	static const struct kfd_device_info kaveri_device_info = {		static const struct kfd_device_info kaveri_device_info = {
	.max_pasid_bits = 16,		.max_pasid_bits = 16,
			.mqd_size_aligned = MQD_SIZE_ALIGNED
	};		};

	struct kfd_deviceid {		struct kfd_deviceid {
	@@ -92,6 +95,7 @@ struct kfd_dev kgd2kfd_probe(struct kgd_dev kgd, struct pci_dev *pdev)
	kfd->kgd = kgd;		kfd->kgd = kgd;
	kfd->device_info = device_info;		kfd->device_info = device_info;
	kfd->pdev = pdev;		kfd->pdev = pdev;
			kfd->init_complete = false;

	return kfd;		return kfd;
	}		}
	@@ -99,23 +103,53 @@ struct kfd_dev kgd2kfd_probe(struct kgd_dev kgd, struct pci_dev *pdev)
	bool kgd2kfd_device_init(struct kfd_dev *kfd,		bool kgd2kfd_device_init(struct kfd_dev *kfd,
	const struct kgd2kfd_shared_resources *gpu_resources)		const struct kgd2kfd_shared_resources *gpu_resources)
	{		{
			unsigned int size;

	kfd->shared_resources = *gpu_resources;		kfd->shared_resources = *gpu_resources;

	if (kfd_topology_add_device(kfd) != 0)		/* calculate max size of mqds needed for queues */
	return false;		size = max_num_of_processes *
			max_num_of_queues_per_process *
			kfd->device_info->mqd_size_aligned;

			/* add another 512KB for all other allocations on gart */
			size += 512 * 1024;

			if (kfd2kgd->init_sa_manager(kfd->kgd, size)) {
			dev_err(kfd_device,
			"Error initializing sa manager for device (%x:%x)\n",
			kfd->pdev->vendor, kfd->pdev->device);
			goto out;
			}

			kfd_doorbell_init(kfd);

			if (kfd_topology_add_device(kfd) != 0) {
			dev_err(kfd_device,
			"Error adding device (%x:%x) to topology\n",
			kfd->pdev->vendor, kfd->pdev->device);
			goto kfd_topology_add_device_error;
			}


	kfd->init_complete = true;		kfd->init_complete = true;
	dev_info(kfd_device, "added device (%x:%x)\n", kfd->pdev->vendor,		dev_info(kfd_device, "added device (%x:%x)\n", kfd->pdev->vendor,
	kfd->pdev->device);		kfd->pdev->device);

	return true;		goto out;

			kfd_topology_add_device_error:
			kfd2kgd->fini_sa_manager(kfd->kgd);
			dev_err(kfd_device,
			"device (%x:%x) NOT added due to errors\n",
			kfd->pdev->vendor, kfd->pdev->device);
			out:
			return kfd->init_complete;
	}		}

	void kgd2kfd_device_exit(struct kfd_dev *kfd)		void kgd2kfd_device_exit(struct kfd_dev *kfd)
	{		{
	int err = kfd_topology_remove_device(kfd);		kfd_topology_remove_device(kfd);

	BUG_ON(err != 0);

	kfree(kfd);		kfree(kfd);
	}		}

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

0 → 100644

+255 −0

Original line number	Original line	Diff line number	Diff line
			/*
			* Copyright 2014 Advanced Micro Devices, Inc.
			*
			* Permission is hereby granted, free of charge, to any person obtaining a
			* copy of this software and associated documentation files (the "Software"),
			* to deal in the Software without restriction, including without limitation
			* the rights to use, copy, modify, merge, publish, distribute, sublicense,
			* and/or sell copies of the Software, and to permit persons to whom the
			* Software is furnished to do so, subject to the following conditions:
			*
			* The above copyright notice and this permission notice shall be included in
			* all copies or substantial portions of the Software.
			*
			* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
			* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
			* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
			* THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
			* OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
			* ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
			* OTHER DEALINGS IN THE SOFTWARE.
			*/
			#include "kfd_priv.h"
			#include <linux/mm.h>
			#include <linux/mman.h>
			#include <linux/slab.h>

			/*
			* This extension supports a kernel level doorbells management for
			* the kernel queues.
			* Basically the last doorbells page is devoted to kernel queues
			* and that's assures that any user process won't get access to the
			* kernel doorbells page
			*/
			static DEFINE_MUTEX(doorbell_mutex);
			static unsigned long doorbell_available_index[
			DIV_ROUND_UP(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS, BITS_PER_LONG)] = { 0 };

			#define KERNEL_DOORBELL_PASID 1
			#define KFD_SIZE_OF_DOORBELL_IN_BYTES 4

			/*
			* Each device exposes a doorbell aperture, a PCI MMIO aperture that
			* receives 32-bit writes that are passed to queues as wptr values.
			* The doorbells are intended to be written by applications as part
			* of queueing work on user-mode queues.
			* We assign doorbells to applications in PAGE_SIZE-sized and aligned chunks.
			* We map the doorbell address space into user-mode when a process creates
			* its first queue on each device.
			* Although the mapping is done by KFD, it is equivalent to an mmap of
			* the /dev/kfd with the particular device encoded in the mmap offset.
			* There will be other uses for mmap of /dev/kfd, so only a range of
			* offsets (KFD_MMAP_DOORBELL_START-END) is used for doorbells.
			*/

			/* # of doorbell bytes allocated for each process. */
			static inline size_t doorbell_process_allocation(void)
			{
			return roundup(KFD_SIZE_OF_DOORBELL_IN_BYTES *
			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
			PAGE_SIZE);
			}

			/* Doorbell calculations for device init. */
			void kfd_doorbell_init(struct kfd_dev *kfd)
			{
			size_t doorbell_start_offset;
			size_t doorbell_aperture_size;
			size_t doorbell_process_limit;

			/*
			* We start with calculations in bytes because the input data might
			* only be byte-aligned.
			* Only after we have done the rounding can we assume any alignment.
			*/

			doorbell_start_offset =
			roundup(kfd->shared_resources.doorbell_start_offset,
			doorbell_process_allocation());

			doorbell_aperture_size =
			rounddown(kfd->shared_resources.doorbell_aperture_size,
			doorbell_process_allocation());

			if (doorbell_aperture_size > doorbell_start_offset)
			doorbell_process_limit =
			(doorbell_aperture_size - doorbell_start_offset) /
			doorbell_process_allocation();
			else
			doorbell_process_limit = 0;

			kfd->doorbell_base = kfd->shared_resources.doorbell_physical_address +
			doorbell_start_offset;

			kfd->doorbell_id_offset = doorbell_start_offset / sizeof(u32);
			kfd->doorbell_process_limit = doorbell_process_limit - 1;

			kfd->doorbell_kernel_ptr = ioremap(kfd->doorbell_base,
			doorbell_process_allocation());

			BUG_ON(!kfd->doorbell_kernel_ptr);

			pr_debug("kfd: doorbell initialization:\n");
			pr_debug("kfd: doorbell base == 0x%08lX\n",
			(uintptr_t)kfd->doorbell_base);

			pr_debug("kfd: doorbell_id_offset == 0x%08lX\n",
			kfd->doorbell_id_offset);

			pr_debug("kfd: doorbell_process_limit == 0x%08lX\n",
			doorbell_process_limit);

			pr_debug("kfd: doorbell_kernel_offset == 0x%08lX\n",
			(uintptr_t)kfd->doorbell_base);

			pr_debug("kfd: doorbell aperture size == 0x%08lX\n",
			kfd->shared_resources.doorbell_aperture_size);

			pr_debug("kfd: doorbell kernel address == 0x%08lX\n",
			(uintptr_t)kfd->doorbell_kernel_ptr);
			}

			int kfd_doorbell_mmap(struct kfd_process process, struct vm_area_struct vma)
			{
			phys_addr_t address;
			struct kfd_dev *dev;

			/*
			* 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 == NULL)
			return -EINVAL;

			/* Find if pdd exists for combination of process and gpu id */
			if (!kfd_get_process_device_data(dev, process, 0))
			return -EINVAL;

			/* Calculate physical address of doorbell */
			address = kfd_get_process_doorbells(dev, process);

			vma->vm_flags \|= VM_IO \| VM_DONTCOPY \| VM_DONTEXPAND \| VM_NORESERVE \|
			VM_DONTDUMP \| VM_PFNMAP;

			vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);

			pr_debug("kfd: mapping doorbell page in kfd_doorbell_mmap\n"
			" target user address == 0x%08llX\n"
			" physical address == 0x%08llX\n"
			" vm_flags == 0x%04lX\n"
			" size == 0x%04lX\n",
			(unsigned long long) vma->vm_start, address, vma->vm_flags,
			doorbell_process_allocation());


			return io_remap_pfn_range(vma,
			vma->vm_start,
			address >> PAGE_SHIFT,
			doorbell_process_allocation(),
			vma->vm_page_prot);
			}


			/* get kernel iomem pointer for a doorbell */
			u32 __iomem kfd_get_kernel_doorbell(struct kfd_dev kfd,
			unsigned int *doorbell_off)
			{
			u32 inx;

			BUG_ON(!kfd \|\| !doorbell_off);

			mutex_lock(&doorbell_mutex);
			inx = find_first_zero_bit(doorbell_available_index,
			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);

			__set_bit(inx, doorbell_available_index);
			mutex_unlock(&doorbell_mutex);

			if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
			return NULL;

			/*
			* Calculating the kernel doorbell offset using "faked" kernel
			* pasid that allocated for kernel queues only
			*/
			doorbell_off = KERNEL_DOORBELL_PASID (doorbell_process_allocation() /
			sizeof(u32)) + inx;

			pr_debug("kfd: get kernel queue doorbell\n"
			" doorbell offset == 0x%08d\n"
			" kernel address == 0x%08lX\n",
			*doorbell_off, (uintptr_t)(kfd->doorbell_kernel_ptr + inx));

			return kfd->doorbell_kernel_ptr + inx;
			}

			void kfd_release_kernel_doorbell(struct kfd_dev kfd, u32 __iomem db_addr)
			{
			unsigned int inx;

			BUG_ON(!kfd \|\| !db_addr);

			inx = (unsigned int)(db_addr - kfd->doorbell_kernel_ptr);

			mutex_lock(&doorbell_mutex);
			__clear_bit(inx, doorbell_available_index);
			mutex_unlock(&doorbell_mutex);
			}

			inline void write_kernel_doorbell(u32 __iomem *db, u32 value)
			{
			if (db) {
			writel(value, db);
			pr_debug("writing %d to doorbell address 0x%p\n", value, db);
			}
			}

			/*
			* queue_ids are in the range [0,MAX_PROCESS_QUEUES) and are mapped 1:1
			* to doorbells with the process's doorbell page
			*/
			unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd,
			struct kfd_process *process,
			unsigned int queue_id)
			{
			/*
			* doorbell_id_offset accounts for doorbells taken by KGD.
			* pasid * doorbell_process_allocation/sizeof(u32) adjusts
			* to the process's doorbells
			*/
			return kfd->doorbell_id_offset +
			process->pasid * (doorbell_process_allocation()/sizeof(u32)) +
			queue_id;
			}

			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;

			return num_of_elems;

			}

			phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
			struct kfd_process *process)
			{
			return dev->doorbell_base +
			process->pasid * doorbell_process_allocation();
			}

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

0 → 100644

+355 −0

File added.

Preview size limit exceeded, changes collapsed.