[linux-2.6-block.git] / drivers / gpu / drm / amd / amdkfd / kfd_device.c

/*
 * 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 <linux/amd-iommu.h>
#include <linux/bsearch.h>
#include <linux/pci.h>
#include <linux/slab.h>
#include "kfd_priv.h"
#include "kfd_device_queue_manager.h"
#include "kfd_pm4_headers.h"

#define MQD_SIZE_ALIGNED 768

static const struct kfd_device_info kaveri_device_info = {
	.max_pasid_bits = 16,
	.ih_ring_entry_size = 4 * sizeof(uint32_t),
	.num_of_watch_points = 4,
	.mqd_size_aligned = MQD_SIZE_ALIGNED
};

struct kfd_deviceid {
	unsigned short did;
	const struct kfd_device_info *device_info;
};

/* Please keep this sorted by increasing device id. */
static const struct kfd_deviceid supported_devices[] = {
	{ 0x1304, &kaveri_device_info },	/* Kaveri */
	{ 0x1305, &kaveri_device_info },	/* Kaveri */
	{ 0x1306, &kaveri_device_info },	/* Kaveri */
	{ 0x1307, &kaveri_device_info },	/* Kaveri */
	{ 0x1309, &kaveri_device_info },	/* Kaveri */
	{ 0x130A, &kaveri_device_info },	/* Kaveri */
	{ 0x130B, &kaveri_device_info },	/* Kaveri */
	{ 0x130C, &kaveri_device_info },	/* Kaveri */
	{ 0x130D, &kaveri_device_info },	/* Kaveri */
	{ 0x130E, &kaveri_device_info },	/* Kaveri */
	{ 0x130F, &kaveri_device_info },	/* Kaveri */
	{ 0x1310, &kaveri_device_info },	/* Kaveri */
	{ 0x1311, &kaveri_device_info },	/* Kaveri */
	{ 0x1312, &kaveri_device_info },	/* Kaveri */
	{ 0x1313, &kaveri_device_info },	/* Kaveri */
	{ 0x1315, &kaveri_device_info },	/* Kaveri */
	{ 0x1316, &kaveri_device_info },	/* Kaveri */
	{ 0x1317, &kaveri_device_info },	/* Kaveri */
	{ 0x1318, &kaveri_device_info },	/* Kaveri */
	{ 0x131B, &kaveri_device_info },	/* Kaveri */
	{ 0x131C, &kaveri_device_info },	/* Kaveri */
	{ 0x131D, &kaveri_device_info },	/* Kaveri */
};

static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
				unsigned int chunk_size);
static void kfd_gtt_sa_fini(struct kfd_dev *kfd);

static const struct kfd_device_info *lookup_device_info(unsigned short did)
{
	size_t i;

	for (i = 0; i < ARRAY_SIZE(supported_devices); i++) {
		if (supported_devices[i].did == did) {
			BUG_ON(supported_devices[i].device_info == NULL);
			return supported_devices[i].device_info;
		}
	}

	return NULL;
}

struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev)
{
	struct kfd_dev *kfd;

	const struct kfd_device_info *device_info =
					lookup_device_info(pdev->device);

	if (!device_info)
		return NULL;

	kfd = kzalloc(sizeof(*kfd), GFP_KERNEL);
	if (!kfd)
		return NULL;

	kfd->kgd = kgd;
	kfd->device_info = device_info;
	kfd->pdev = pdev;
	kfd->init_complete = false;

	return kfd;
}

static bool device_iommu_pasid_init(struct kfd_dev *kfd)
{
	const u32 required_iommu_flags = AMD_IOMMU_DEVICE_FLAG_ATS_SUP |
					AMD_IOMMU_DEVICE_FLAG_PRI_SUP |
					AMD_IOMMU_DEVICE_FLAG_PASID_SUP;

	struct amd_iommu_device_info iommu_info;
	unsigned int pasid_limit;
	int err;

	err = amd_iommu_device_info(kfd->pdev, &iommu_info);
	if (err < 0) {
		dev_err(kfd_device,
			"error getting iommu info. is the iommu enabled?\n");
		return false;
	}

	if ((iommu_info.flags & required_iommu_flags) != required_iommu_flags) {
		dev_err(kfd_device, "error required iommu flags ats(%i), pri(%i), pasid(%i)\n",
		       (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_ATS_SUP) != 0,
		       (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PRI_SUP) != 0,
		       (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PASID_SUP) != 0);
		return false;
	}

	pasid_limit = min_t(unsigned int,
			(unsigned int)1 << kfd->device_info->max_pasid_bits,
			iommu_info.max_pasids);
	/*
	 * last pasid is used for kernel queues doorbells
	 * in the future the last pasid might be used for a kernel thread.
	 */
	pasid_limit = min_t(unsigned int,
				pasid_limit,
				kfd->doorbell_process_limit - 1);

	err = amd_iommu_init_device(kfd->pdev, pasid_limit);
	if (err < 0) {
		dev_err(kfd_device, "error initializing iommu device\n");
		return false;
	}

	if (!kfd_set_pasid_limit(pasid_limit)) {
		dev_err(kfd_device, "error setting pasid limit\n");
		amd_iommu_free_device(kfd->pdev);
		return false;
	}

	return true;
}

static void iommu_pasid_shutdown_callback(struct pci_dev *pdev, int pasid)
{
	struct kfd_dev *dev = kfd_device_by_pci_dev(pdev);

	if (dev)
		kfd_unbind_process_from_device(dev, pasid);
}

bool kgd2kfd_device_init(struct kfd_dev *kfd,
			 const struct kgd2kfd_shared_resources *gpu_resources)
{
	unsigned int size;

	kfd->shared_resources = *gpu_resources;

	/* calculate max size of mqds needed for queues */
	size = max_num_of_processes *
		max_num_of_queues_per_process *
		kfd->device_info->mqd_size_aligned;

	/*
	 * calculate max size of runlist packet.
	 * There can be only 2 packets at once
	 */
	size += (max_num_of_processes * sizeof(struct pm4_map_process) +
		max_num_of_processes * max_num_of_queues_per_process *
		sizeof(struct pm4_map_queues) + sizeof(struct pm4_runlist)) * 2;

	/* Add size of HIQ & DIQ */
	size += KFD_KERNEL_QUEUE_SIZE * 2;

	/* add another 512KB for all other allocations on gart (HPD, fences) */
	size += 512 * 1024;

	if (kfd2kgd->init_gtt_mem_allocation(kfd->kgd, size, &kfd->gtt_mem,
			&kfd->gtt_start_gpu_addr, &kfd->gtt_start_cpu_ptr)) {
		dev_err(kfd_device,
			"Could not allocate %d bytes for device (%x:%x)\n",
			size, kfd->pdev->vendor, kfd->pdev->device);
		goto out;
	}

	dev_info(kfd_device,
		"Allocated %d bytes on gart for device(%x:%x)\n",
		size, kfd->pdev->vendor, kfd->pdev->device);

	/* Initialize GTT sa with 512 byte chunk size */
	if (kfd_gtt_sa_init(kfd, size, 512) != 0) {
		dev_err(kfd_device,
			"Error initializing gtt sub-allocator\n");
		goto kfd_gtt_sa_init_error;
	}

	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;
	}

	if (kfd_interrupt_init(kfd)) {
		dev_err(kfd_device,
			"Error initializing interrupts for device (%x:%x)\n",
			kfd->pdev->vendor, kfd->pdev->device);
		goto kfd_interrupt_error;
	}

	if (!device_iommu_pasid_init(kfd)) {
		dev_err(kfd_device,
			"Error initializing iommuv2 for device (%x:%x)\n",
			kfd->pdev->vendor, kfd->pdev->device);
		goto device_iommu_pasid_error;
	}
	amd_iommu_set_invalidate_ctx_cb(kfd->pdev,
						iommu_pasid_shutdown_callback);

	kfd->dqm = device_queue_manager_init(kfd);
	if (!kfd->dqm) {
		dev_err(kfd_device,
			"Error initializing queue manager for device (%x:%x)\n",
			kfd->pdev->vendor, kfd->pdev->device);
		goto device_queue_manager_error;
	}

	if (kfd->dqm->start(kfd->dqm) != 0) {
		dev_err(kfd_device,
			"Error starting queuen manager for device (%x:%x)\n",
			kfd->pdev->vendor, kfd->pdev->device);
		goto dqm_start_error;
	}

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

	pr_debug("kfd: Starting kfd with the following scheduling policy %d\n",
		sched_policy);

	goto out;

dqm_start_error:
	device_queue_manager_uninit(kfd->dqm);
device_queue_manager_error:
	amd_iommu_free_device(kfd->pdev);
device_iommu_pasid_error:
	kfd_interrupt_exit(kfd);
kfd_interrupt_error:
	kfd_topology_remove_device(kfd);
kfd_topology_add_device_error:
	kfd_gtt_sa_fini(kfd);
kfd_gtt_sa_init_error:
	kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem);
	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)
{
	if (kfd->init_complete) {
		device_queue_manager_uninit(kfd->dqm);
		amd_iommu_free_device(kfd->pdev);
		kfd_interrupt_exit(kfd);
		kfd_topology_remove_device(kfd);
		kfd_gtt_sa_fini(kfd);
		kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem);
	}

	kfree(kfd);
}

void kgd2kfd_suspend(struct kfd_dev *kfd)
{
	BUG_ON(kfd == NULL);

	if (kfd->init_complete) {
		kfd->dqm->stop(kfd->dqm);
		amd_iommu_set_invalidate_ctx_cb(kfd->pdev, NULL);
		amd_iommu_free_device(kfd->pdev);
	}
}

int kgd2kfd_resume(struct kfd_dev *kfd)
{
	unsigned int pasid_limit;
	int err;

	BUG_ON(kfd == NULL);

	pasid_limit = kfd_get_pasid_limit();

	if (kfd->init_complete) {
		err = amd_iommu_init_device(kfd->pdev, pasid_limit);
		if (err < 0)
			return -ENXIO;
		amd_iommu_set_invalidate_ctx_cb(kfd->pdev,
						iommu_pasid_shutdown_callback);
		kfd->dqm->start(kfd->dqm);
	}

	return 0;
}

/* This is called directly from KGD at ISR. */
void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
{
	if (kfd->init_complete) {
		spin_lock(&kfd->interrupt_lock);

		if (kfd->interrupts_active
		    && enqueue_ih_ring_entry(kfd, ih_ring_entry))
			schedule_work(&kfd->interrupt_work);

		spin_unlock(&kfd->interrupt_lock);
	}
}

static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
				unsigned int chunk_size)
{
	unsigned int num_of_bits;

	BUG_ON(!kfd);
	BUG_ON(!kfd->gtt_mem);
	BUG_ON(buf_size < chunk_size);
	BUG_ON(buf_size == 0);
	BUG_ON(chunk_size == 0);

	kfd->gtt_sa_chunk_size = chunk_size;
	kfd->gtt_sa_num_of_chunks = buf_size / chunk_size;

	num_of_bits = kfd->gtt_sa_num_of_chunks / BITS_PER_BYTE;
	BUG_ON(num_of_bits == 0);

	kfd->gtt_sa_bitmap = kzalloc(num_of_bits, GFP_KERNEL);

	if (!kfd->gtt_sa_bitmap)
		return -ENOMEM;

	pr_debug("kfd: gtt_sa_num_of_chunks = %d, gtt_sa_bitmap = %p\n",
			kfd->gtt_sa_num_of_chunks, kfd->gtt_sa_bitmap);

	mutex_init(&kfd->gtt_sa_lock);

	return 0;

}

static void kfd_gtt_sa_fini(struct kfd_dev *kfd)
{
	mutex_destroy(&kfd->gtt_sa_lock);
	kfree(kfd->gtt_sa_bitmap);
}

static inline uint64_t kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr,
						unsigned int bit_num,
						unsigned int chunk_size)
{
	return start_addr + bit_num * chunk_size;
}

static inline uint32_t *kfd_gtt_sa_calc_cpu_addr(void *start_addr,
						unsigned int bit_num,
						unsigned int chunk_size)
{
	return (uint32_t *) ((uint64_t) start_addr + bit_num * chunk_size);
}

int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
			struct kfd_mem_obj **mem_obj)
{
	unsigned int found, start_search, cur_size;

	BUG_ON(!kfd);

	if (size == 0)
		return -EINVAL;

	if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size)
		return -ENOMEM;

	*mem_obj = kmalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
	if ((*mem_obj) == NULL)
		return -ENOMEM;

	pr_debug("kfd: allocated mem_obj = %p for size = %d\n", *mem_obj, size);

	start_search = 0;

	mutex_lock(&kfd->gtt_sa_lock);

kfd_gtt_restart_search:
	/* Find the first chunk that is free */
	found = find_next_zero_bit(kfd->gtt_sa_bitmap,
					kfd->gtt_sa_num_of_chunks,
					start_search);

	pr_debug("kfd: found = %d\n", found);

	/* If there wasn't any free chunk, bail out */
	if (found == kfd->gtt_sa_num_of_chunks)
		goto kfd_gtt_no_free_chunk;

	/* Update fields of mem_obj */
	(*mem_obj)->range_start = found;
	(*mem_obj)->range_end = found;
	(*mem_obj)->gpu_addr = kfd_gtt_sa_calc_gpu_addr(
					kfd->gtt_start_gpu_addr,
					found,
					kfd->gtt_sa_chunk_size);
	(*mem_obj)->cpu_ptr = kfd_gtt_sa_calc_cpu_addr(
					kfd->gtt_start_cpu_ptr,
					found,
					kfd->gtt_sa_chunk_size);

	pr_debug("kfd: gpu_addr = %p, cpu_addr = %p\n",
			(uint64_t *) (*mem_obj)->gpu_addr, (*mem_obj)->cpu_ptr);

	/* If we need only one chunk, mark it as allocated and get out */
	if (size <= kfd->gtt_sa_chunk_size) {
		pr_debug("kfd: single bit\n");
		set_bit(found, kfd->gtt_sa_bitmap);
		goto kfd_gtt_out;
	}

	/* Otherwise, try to see if we have enough contiguous chunks */
	cur_size = size - kfd->gtt_sa_chunk_size;
	do {
		(*mem_obj)->range_end =
			find_next_zero_bit(kfd->gtt_sa_bitmap,
					kfd->gtt_sa_num_of_chunks, ++found);
		/*
		 * If next free chunk is not contiguous than we need to
		 * restart our search from the last free chunk we found (which
		 * wasn't contiguous to the previous ones
		 */
		if ((*mem_obj)->range_end != found) {
			start_search = found;
			goto kfd_gtt_restart_search;
		}

		/*
		 * If we reached end of buffer, bail out with error
		 */
		if (found == kfd->gtt_sa_num_of_chunks)
			goto kfd_gtt_no_free_chunk;

		/* Check if we don't need another chunk */
		if (cur_size <= kfd->gtt_sa_chunk_size)
			cur_size = 0;
		else
			cur_size -= kfd->gtt_sa_chunk_size;

	} while (cur_size > 0);

	pr_debug("kfd: range_start = %d, range_end = %d\n",
		(*mem_obj)->range_start, (*mem_obj)->range_end);

	/* Mark the chunks as allocated */
	for (found = (*mem_obj)->range_start;
		found <= (*mem_obj)->range_end;
		found++)
		set_bit(found, kfd->gtt_sa_bitmap);

kfd_gtt_out:
	mutex_unlock(&kfd->gtt_sa_lock);
	return 0;

kfd_gtt_no_free_chunk:
	pr_debug("kfd: allocation failed with mem_obj = %p\n", mem_obj);
	mutex_unlock(&kfd->gtt_sa_lock);
	kfree(mem_obj);
	return -ENOMEM;
}

int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj)
{
	unsigned int bit;

	BUG_ON(!kfd);
	BUG_ON(!mem_obj);

	pr_debug("kfd: free mem_obj = %p, range_start = %d, range_end = %d\n",
			mem_obj, mem_obj->range_start, mem_obj->range_end);

	mutex_lock(&kfd->gtt_sa_lock);

	/* Mark the chunks as free */
	for (bit = mem_obj->range_start;
		bit <= mem_obj->range_end;
		bit++)
		clear_bit(bit, kfd->gtt_sa_bitmap);

	mutex_unlock(&kfd->gtt_sa_lock);

	kfree(mem_obj);
	return 0;
}
Commit	Line	Data
4a488a7a OG	1	/*
	2	* Copyright 2014 Advanced Micro Devices, Inc.
	3	*
	4	* Permission is hereby granted, free of charge, to any person obtaining a
	5	* copy of this software and associated documentation files (the "Software"),
	6	* to deal in the Software without restriction, including without limitation
	7	* the rights to use, copy, modify, merge, publish, distribute, sublicense,
	8	* and/or sell copies of the Software, and to permit persons to whom the
	9	* Software is furnished to do so, subject to the following conditions:
	10	*
	11	* The above copyright notice and this permission notice shall be included in
	12	* all copies or substantial portions of the Software.
	13	*
	14	* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
	15	* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
	16	* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
	17	* THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
	18	* OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
	19	* ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
	20	* OTHER DEALINGS IN THE SOFTWARE.
	21	*/
	22
	23	#include <linux/amd-iommu.h>
	24	#include <linux/bsearch.h>
	25	#include <linux/pci.h>
	26	#include <linux/slab.h>
	27	#include "kfd_priv.h"
64c7f8cf	28	#include "kfd_device_queue_manager.h"
e18e794e	29	#include "kfd_pm4_headers.h"
4a488a7a	30
19f6d2a6 OG	31	#define MQD_SIZE_ALIGNED 768
19f6d2a6 OG	32
4a488a7a OG	33	static const struct kfd_device_info kaveri_device_info = {
4a488a7a OG	34	.max_pasid_bits = 16,
b3f5e6b4	35	.ih_ring_entry_size = 4 * sizeof(uint32_t),
f7c826ad	36	.num_of_watch_points = 4,
19f6d2a6	37	.mqd_size_aligned = MQD_SIZE_ALIGNED
4a488a7a OG	38	};
	39
	40	struct kfd_deviceid {
	41	unsigned short did;
	42	const struct kfd_device_info *device_info;
	43	};
	44
	45	/* Please keep this sorted by increasing device id. */
	46	static const struct kfd_deviceid supported_devices[] = {
	47	{ 0x1304, &kaveri_device_info }, /* Kaveri */
	48	{ 0x1305, &kaveri_device_info }, /* Kaveri */
	49	{ 0x1306, &kaveri_device_info }, /* Kaveri */
	50	{ 0x1307, &kaveri_device_info }, /* Kaveri */
	51	{ 0x1309, &kaveri_device_info }, /* Kaveri */
	52	{ 0x130A, &kaveri_device_info }, /* Kaveri */
	53	{ 0x130B, &kaveri_device_info }, /* Kaveri */
	54	{ 0x130C, &kaveri_device_info }, /* Kaveri */
	55	{ 0x130D, &kaveri_device_info }, /* Kaveri */
	56	{ 0x130E, &kaveri_device_info }, /* Kaveri */
	57	{ 0x130F, &kaveri_device_info }, /* Kaveri */
	58	{ 0x1310, &kaveri_device_info }, /* Kaveri */
	59	{ 0x1311, &kaveri_device_info }, /* Kaveri */
	60	{ 0x1312, &kaveri_device_info }, /* Kaveri */
	61	{ 0x1313, &kaveri_device_info }, /* Kaveri */
	62	{ 0x1315, &kaveri_device_info }, /* Kaveri */
	63	{ 0x1316, &kaveri_device_info }, /* Kaveri */
	64	{ 0x1317, &kaveri_device_info }, /* Kaveri */
	65	{ 0x1318, &kaveri_device_info }, /* Kaveri */
	66	{ 0x131B, &kaveri_device_info }, /* Kaveri */
	67	{ 0x131C, &kaveri_device_info }, /* Kaveri */
	68	{ 0x131D, &kaveri_device_info }, /* Kaveri */
	69	};
	70
6e81090b OG	71	static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
	72	unsigned int chunk_size);
	73	static void kfd_gtt_sa_fini(struct kfd_dev *kfd);
	74
4a488a7a OG	75	static const struct kfd_device_info *lookup_device_info(unsigned short did)
	76	{
	77	size_t i;
	78
	79	for (i = 0; i < ARRAY_SIZE(supported_devices); i++) {
	80	if (supported_devices[i].did == did) {
	81	BUG_ON(supported_devices[i].device_info == NULL);
	82	return supported_devices[i].device_info;
	83	}
	84	}
	85
	86	return NULL;
	87	}
	88
	89	struct kfd_dev kgd2kfd_probe(struct kgd_dev kgd, struct pci_dev *pdev)
	90	{
	91	struct kfd_dev *kfd;
	92
	93	const struct kfd_device_info *device_info =
	94	lookup_device_info(pdev->device);
	95
	96	if (!device_info)
	97	return NULL;
	98
	99	kfd = kzalloc(sizeof(*kfd), GFP_KERNEL);
	100	if (!kfd)
	101	return NULL;
	102
	103	kfd->kgd = kgd;
	104	kfd->device_info = device_info;
	105	kfd->pdev = pdev;
19f6d2a6	106	kfd->init_complete = false;
4a488a7a OG	107
	108	return kfd;
	109	}
	110
b17f068a OG	111	static bool device_iommu_pasid_init(struct kfd_dev *kfd)
	112	{
	113	const u32 required_iommu_flags = AMD_IOMMU_DEVICE_FLAG_ATS_SUP \|
	114	AMD_IOMMU_DEVICE_FLAG_PRI_SUP \|
	115	AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
	116
	117	struct amd_iommu_device_info iommu_info;
	118	unsigned int pasid_limit;
	119	int err;
	120
	121	err = amd_iommu_device_info(kfd->pdev, &iommu_info);
	122	if (err < 0) {
	123	dev_err(kfd_device,
	124	"error getting iommu info. is the iommu enabled?\n");
	125	return false;
	126	}
	127
	128	if ((iommu_info.flags & required_iommu_flags) != required_iommu_flags) {
	129	dev_err(kfd_device, "error required iommu flags ats(%i), pri(%i), pasid(%i)\n",
	130	(iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_ATS_SUP) != 0,
	131	(iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PRI_SUP) != 0,
	132	(iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PASID_SUP) != 0);
	133	return false;
	134	}
	135
	136	pasid_limit = min_t(unsigned int,
	137	(unsigned int)1 << kfd->device_info->max_pasid_bits,
	138	iommu_info.max_pasids);
	139	/*
	140	* last pasid is used for kernel queues doorbells
	141	* in the future the last pasid might be used for a kernel thread.
	142	*/
	143	pasid_limit = min_t(unsigned int,
	144	pasid_limit,
	145	kfd->doorbell_process_limit - 1);
	146
	147	err = amd_iommu_init_device(kfd->pdev, pasid_limit);
	148	if (err < 0) {
	149	dev_err(kfd_device, "error initializing iommu device\n");
	150	return false;
	151	}
	152
	153	if (!kfd_set_pasid_limit(pasid_limit)) {
	154	dev_err(kfd_device, "error setting pasid limit\n");
	155	amd_iommu_free_device(kfd->pdev);
	156	return false;
	157	}
	158
	159	return true;
	160	}
	161
	162	static void iommu_pasid_shutdown_callback(struct pci_dev *pdev, int pasid)
	163	{
	164	struct kfd_dev *dev = kfd_device_by_pci_dev(pdev);
	165
	166	if (dev)
	167	kfd_unbind_process_from_device(dev, pasid);
	168	}
	169
4a488a7a OG	170	bool kgd2kfd_device_init(struct kfd_dev *kfd,
	171	const struct kgd2kfd_shared_resources *gpu_resources)
	172	{
19f6d2a6 OG	173	unsigned int size;
19f6d2a6 OG	174
4a488a7a OG	175	kfd->shared_resources = *gpu_resources;
4a488a7a OG	176
19f6d2a6 OG	177	/* calculate max size of mqds needed for queues */
	178	size = max_num_of_processes *
	179	max_num_of_queues_per_process *
	180	kfd->device_info->mqd_size_aligned;
	181
e18e794e OG	182	/*
	183	* calculate max size of runlist packet.
	184	* There can be only 2 packets at once
	185	*/
	186	size += (max_num_of_processes * sizeof(struct pm4_map_process) +
	187	max_num_of_processes * max_num_of_queues_per_process *
	188	sizeof(struct pm4_map_queues) + sizeof(struct pm4_runlist)) * 2;
	189
	190	/* Add size of HIQ & DIQ */
	191	size += KFD_KERNEL_QUEUE_SIZE * 2;
	192
	193	/* add another 512KB for all other allocations on gart (HPD, fences) */
19f6d2a6 OG	194	size += 512 * 1024;
19f6d2a6 OG	195
73a1da0b OG	196	if (kfd2kgd->init_gtt_mem_allocation(kfd->kgd, size, &kfd->gtt_mem,
73a1da0b OG	197	&kfd->gtt_start_gpu_addr, &kfd->gtt_start_cpu_ptr)) {
19f6d2a6	198	dev_err(kfd_device,
e18e794e OG	199	"Could not allocate %d bytes for device (%x:%x)\n",
e18e794e OG	200	size, kfd->pdev->vendor, kfd->pdev->device);
19f6d2a6 OG	201	goto out;
	202	}
	203
e18e794e OG	204	dev_info(kfd_device,
	205	"Allocated %d bytes on gart for device(%x:%x)\n",
	206	size, kfd->pdev->vendor, kfd->pdev->device);
	207
73a1da0b OG	208	/* Initialize GTT sa with 512 byte chunk size */
	209	if (kfd_gtt_sa_init(kfd, size, 512) != 0) {
	210	dev_err(kfd_device,
	211	"Error initializing gtt sub-allocator\n");
	212	goto kfd_gtt_sa_init_error;
	213	}
	214
19f6d2a6 OG	215	kfd_doorbell_init(kfd);
	216
	217	if (kfd_topology_add_device(kfd) != 0) {
	218	dev_err(kfd_device,
	219	"Error adding device (%x:%x) to topology\n",
	220	kfd->pdev->vendor, kfd->pdev->device);
	221	goto kfd_topology_add_device_error;
	222	}
	223
b3f5e6b4 AL	224	if (kfd_interrupt_init(kfd)) {
	225	dev_err(kfd_device,
	226	"Error initializing interrupts for device (%x:%x)\n",
	227	kfd->pdev->vendor, kfd->pdev->device);
	228	goto kfd_interrupt_error;
	229	}
	230
b17f068a OG	231	if (!device_iommu_pasid_init(kfd)) {
	232	dev_err(kfd_device,
	233	"Error initializing iommuv2 for device (%x:%x)\n",
	234	kfd->pdev->vendor, kfd->pdev->device);
	235	goto device_iommu_pasid_error;
	236	}
	237	amd_iommu_set_invalidate_ctx_cb(kfd->pdev,
	238	iommu_pasid_shutdown_callback);
5b5c4e40	239
64c7f8cf BG	240	kfd->dqm = device_queue_manager_init(kfd);
	241	if (!kfd->dqm) {
	242	dev_err(kfd_device,
	243	"Error initializing queue manager for device (%x:%x)\n",
	244	kfd->pdev->vendor, kfd->pdev->device);
	245	goto device_queue_manager_error;
	246	}
	247
	248	if (kfd->dqm->start(kfd->dqm) != 0) {
	249	dev_err(kfd_device,
	250	"Error starting queuen manager for device (%x:%x)\n",
	251	kfd->pdev->vendor, kfd->pdev->device);
	252	goto dqm_start_error;
	253	}
	254
4a488a7a OG	255	kfd->init_complete = true;
	256	dev_info(kfd_device, "added device (%x:%x)\n", kfd->pdev->vendor,
	257	kfd->pdev->device);
	258
64c7f8cf BG	259	pr_debug("kfd: Starting kfd with the following scheduling policy %d\n",
	260	sched_policy);
	261
19f6d2a6 OG	262	goto out;
19f6d2a6 OG	263
64c7f8cf BG	264	dqm_start_error:
	265	device_queue_manager_uninit(kfd->dqm);
	266	device_queue_manager_error:
	267	amd_iommu_free_device(kfd->pdev);
b17f068a	268	device_iommu_pasid_error:
b3f5e6b4 AL	269	kfd_interrupt_exit(kfd);
b3f5e6b4 AL	270	kfd_interrupt_error:
b17f068a	271	kfd_topology_remove_device(kfd);
19f6d2a6	272	kfd_topology_add_device_error:
73a1da0b OG	273	kfd_gtt_sa_fini(kfd);
	274	kfd_gtt_sa_init_error:
	275	kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem);
19f6d2a6 OG	276	dev_err(kfd_device,
	277	"device (%x:%x) NOT added due to errors\n",
	278	kfd->pdev->vendor, kfd->pdev->device);
	279	out:
	280	return kfd->init_complete;
4a488a7a OG	281	}
	282
	283	void kgd2kfd_device_exit(struct kfd_dev *kfd)
	284	{
b17f068a	285	if (kfd->init_complete) {
64c7f8cf	286	device_queue_manager_uninit(kfd->dqm);
b17f068a	287	amd_iommu_free_device(kfd->pdev);
b3f5e6b4	288	kfd_interrupt_exit(kfd);
b17f068a	289	kfd_topology_remove_device(kfd);
73a1da0b OG	290	kfd_gtt_sa_fini(kfd);
73a1da0b OG	291	kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem);
b17f068a	292	}
5b5c4e40	293
4a488a7a OG	294	kfree(kfd);
	295	}
	296
	297	void kgd2kfd_suspend(struct kfd_dev *kfd)
	298	{
	299	BUG_ON(kfd == NULL);
b17f068a	300
64c7f8cf BG	301	if (kfd->init_complete) {
64c7f8cf BG	302	kfd->dqm->stop(kfd->dqm);
abc9d3e3	303	amd_iommu_set_invalidate_ctx_cb(kfd->pdev, NULL);
b17f068a	304	amd_iommu_free_device(kfd->pdev);
64c7f8cf	305	}
4a488a7a OG	306	}
	307
	308	int kgd2kfd_resume(struct kfd_dev *kfd)
	309	{
b17f068a OG	310	unsigned int pasid_limit;
	311	int err;
	312
4a488a7a OG	313	BUG_ON(kfd == NULL);
4a488a7a OG	314
b17f068a OG	315	pasid_limit = kfd_get_pasid_limit();
	316
	317	if (kfd->init_complete) {
	318	err = amd_iommu_init_device(kfd->pdev, pasid_limit);
	319	if (err < 0)
	320	return -ENXIO;
	321	amd_iommu_set_invalidate_ctx_cb(kfd->pdev,
	322	iommu_pasid_shutdown_callback);
64c7f8cf	323	kfd->dqm->start(kfd->dqm);
b17f068a OG	324	}
b17f068a OG	325
4a488a7a OG	326	return 0;
	327	}
	328
b3f5e6b4 AL	329	/* This is called directly from KGD at ISR. */
b3f5e6b4 AL	330	void kgd2kfd_interrupt(struct kfd_dev kfd, const void ih_ring_entry)
4a488a7a	331	{
b3f5e6b4 AL	332	if (kfd->init_complete) {
	333	spin_lock(&kfd->interrupt_lock);
	334
	335	if (kfd->interrupts_active
	336	&& enqueue_ih_ring_entry(kfd, ih_ring_entry))
	337	schedule_work(&kfd->interrupt_work);
	338
	339	spin_unlock(&kfd->interrupt_lock);
	340	}
4a488a7a	341	}
6e81090b OG	342
	343	static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
	344	unsigned int chunk_size)
	345	{
	346	unsigned int num_of_bits;
	347
	348	BUG_ON(!kfd);
	349	BUG_ON(!kfd->gtt_mem);
	350	BUG_ON(buf_size < chunk_size);
	351	BUG_ON(buf_size == 0);
	352	BUG_ON(chunk_size == 0);
	353
	354	kfd->gtt_sa_chunk_size = chunk_size;
	355	kfd->gtt_sa_num_of_chunks = buf_size / chunk_size;
	356
	357	num_of_bits = kfd->gtt_sa_num_of_chunks / BITS_PER_BYTE;
	358	BUG_ON(num_of_bits == 0);
	359
	360	kfd->gtt_sa_bitmap = kzalloc(num_of_bits, GFP_KERNEL);
	361
	362	if (!kfd->gtt_sa_bitmap)
	363	return -ENOMEM;
	364
	365	pr_debug("kfd: gtt_sa_num_of_chunks = %d, gtt_sa_bitmap = %p\n",
	366	kfd->gtt_sa_num_of_chunks, kfd->gtt_sa_bitmap);
	367
	368	mutex_init(&kfd->gtt_sa_lock);
	369
	370	return 0;
	371
	372	}
	373
	374	static void kfd_gtt_sa_fini(struct kfd_dev *kfd)
	375	{
	376	mutex_destroy(&kfd->gtt_sa_lock);
	377	kfree(kfd->gtt_sa_bitmap);
	378	}
	379
	380	static inline uint64_t kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr,
	381	unsigned int bit_num,
	382	unsigned int chunk_size)
	383	{
	384	return start_addr + bit_num * chunk_size;
	385	}
	386
	387	static inline uint32_t kfd_gtt_sa_calc_cpu_addr(void start_addr,
	388	unsigned int bit_num,
	389	unsigned int chunk_size)
	390	{
	391	return (uint32_t ) ((uint64_t) start_addr + bit_num chunk_size);
	392	}
	393
	394	int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
	395	struct kfd_mem_obj **mem_obj)
	396	{
	397	unsigned int found, start_search, cur_size;
	398
	399	BUG_ON(!kfd);
	400
	401	if (size == 0)
	402	return -EINVAL;
	403
	404	if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size)
	405	return -ENOMEM;
406
407	*mem_obj = kmalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
408	if ((*mem_obj) == NULL)
409	return -ENOMEM;
410
411	pr_debug("kfd: allocated mem_obj = %p for size = %d\n", *mem_obj, size);
412
413	start_search = 0;
414
415	mutex_lock(&kfd->gtt_sa_lock);
416
417	kfd_gtt_restart_search:
418	/* Find the first chunk that is free */
419	found = find_next_zero_bit(kfd->gtt_sa_bitmap,
420	kfd->gtt_sa_num_of_chunks,
421	start_search);
422
423	pr_debug("kfd: found = %d\n", found);
424
425	/* If there wasn't any free chunk, bail out */
426	if (found == kfd->gtt_sa_num_of_chunks)
427	goto kfd_gtt_no_free_chunk;
428
429	/* Update fields of mem_obj */
430	(*mem_obj)->range_start = found;
431	(*mem_obj)->range_end = found;
432	(*mem_obj)->gpu_addr = kfd_gtt_sa_calc_gpu_addr(
433	kfd->gtt_start_gpu_addr,
434	found,
435	kfd->gtt_sa_chunk_size);
436	(*mem_obj)->cpu_ptr = kfd_gtt_sa_calc_cpu_addr(
437	kfd->gtt_start_cpu_ptr,
438	found,
439	kfd->gtt_sa_chunk_size);
440
441	pr_debug("kfd: gpu_addr = %p, cpu_addr = %p\n",
442	(uint64_t ) (mem_obj)->gpu_addr, (*mem_obj)->cpu_ptr);
443
444	/* If we need only one chunk, mark it as allocated and get out */
445	if (size <= kfd->gtt_sa_chunk_size) {
446	pr_debug("kfd: single bit\n");
447	set_bit(found, kfd->gtt_sa_bitmap);
448	goto kfd_gtt_out;
449	}
450
451	/* Otherwise, try to see if we have enough contiguous chunks */
452	cur_size = size - kfd->gtt_sa_chunk_size;
453	do {
454	(*mem_obj)->range_end =
455	find_next_zero_bit(kfd->gtt_sa_bitmap,
456	kfd->gtt_sa_num_of_chunks, ++found);
457	/*
458	* If next free chunk is not contiguous than we need to
459	* restart our search from the last free chunk we found (which
460	* wasn't contiguous to the previous ones
461	*/
462	if ((*mem_obj)->range_end != found) {
463	start_search = found;
464	goto kfd_gtt_restart_search;
465	}
466
467	/*
468	* If we reached end of buffer, bail out with error
469	*/
470	if (found == kfd->gtt_sa_num_of_chunks)
471	goto kfd_gtt_no_free_chunk;
472
473	/* Check if we don't need another chunk */
474	if (cur_size <= kfd->gtt_sa_chunk_size)
475	cur_size = 0;
476	else
477	cur_size -= kfd->gtt_sa_chunk_size;
478
479	} while (cur_size > 0);
480
481	pr_debug("kfd: range_start = %d, range_end = %d\n",
482	(mem_obj)->range_start, (mem_obj)->range_end);
483
484	/* Mark the chunks as allocated */
485	for (found = (*mem_obj)->range_start;
486	found <= (*mem_obj)->range_end;
487	found++)
488	set_bit(found, kfd->gtt_sa_bitmap);
489
490	kfd_gtt_out:
491	mutex_unlock(&kfd->gtt_sa_lock);
492	return 0;
493
494	kfd_gtt_no_free_chunk:
495	pr_debug("kfd: allocation failed with mem_obj = %p\n", mem_obj);
496	mutex_unlock(&kfd->gtt_sa_lock);
497	kfree(mem_obj);
498	return -ENOMEM;
499	}
500
501	int kfd_gtt_sa_free(struct kfd_dev kfd, struct kfd_mem_obj mem_obj)
502	{
503	unsigned int bit;
504
505	BUG_ON(!kfd);
506	BUG_ON(!mem_obj);
507
508	pr_debug("kfd: free mem_obj = %p, range_start = %d, range_end = %d\n",
509	mem_obj, mem_obj->range_start, mem_obj->range_end);
510
511	mutex_lock(&kfd->gtt_sa_lock);
512
513	/* Mark the chunks as free */
514	for (bit = mem_obj->range_start;
515	bit <= mem_obj->range_end;
516	bit++)
517	clear_bit(bit, kfd->gtt_sa_bitmap);
518
519	mutex_unlock(&kfd->gtt_sa_lock);
520
521	kfree(mem_obj);
522	return 0;
523	}