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#include "kfd_priv.h"
23#include <linux/mm.h>
24#include <linux/mman.h>
25#include <linux/slab.h>
26#include <linux/io.h>
27
28/*
29 * This extension supports a kernel level doorbells management for
30 * the kernel queues.
31 * Basically the last doorbells page is devoted to kernel queues
32 * and that's assures that any user process won't get access to the
33 * kernel doorbells page
34 */
35
36#define KERNEL_DOORBELL_PASID 1
37#define KFD_SIZE_OF_DOORBELL_IN_BYTES 4
38
39/*
40 * Each device exposes a doorbell aperture, a PCI MMIO aperture that
41 * receives 32-bit writes that are passed to queues as wptr values.
42 * The doorbells are intended to be written by applications as part
43 * of queueing work on user-mode queues.
44 * We assign doorbells to applications in PAGE_SIZE-sized and aligned chunks.
45 * We map the doorbell address space into user-mode when a process creates
46 * its first queue on each device.
47 * Although the mapping is done by KFD, it is equivalent to an mmap of
48 * the /dev/kfd with the particular device encoded in the mmap offset.
49 * There will be other uses for mmap of /dev/kfd, so only a range of
50 * offsets (KFD_MMAP_DOORBELL_START-END) is used for doorbells.
51 */
52
53/* # of doorbell bytes allocated for each process. */
54static inline size_t doorbell_process_allocation(void)
55{
56	return roundup(KFD_SIZE_OF_DOORBELL_IN_BYTES *
57			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
58			PAGE_SIZE);
59}
60
61/* Doorbell calculations for device init. */
62void kfd_doorbell_init(struct kfd_dev *kfd)
63{
64	size_t doorbell_start_offset;
65	size_t doorbell_aperture_size;
66	size_t doorbell_process_limit;
67
68	/*
69	 * We start with calculations in bytes because the input data might
70	 * only be byte-aligned.
71	 * Only after we have done the rounding can we assume any alignment.
72	 */
73
74	doorbell_start_offset =
75			roundup(kfd->shared_resources.doorbell_start_offset,
76					doorbell_process_allocation());
77
78	doorbell_aperture_size =
79			rounddown(kfd->shared_resources.doorbell_aperture_size,
80					doorbell_process_allocation());
81
82	if (doorbell_aperture_size > doorbell_start_offset)
83		doorbell_process_limit =
84			(doorbell_aperture_size - doorbell_start_offset) /
85						doorbell_process_allocation();
86	else
87		doorbell_process_limit = 0;
88
89	kfd->doorbell_base = kfd->shared_resources.doorbell_physical_address +
90				doorbell_start_offset;
91
92	kfd->doorbell_id_offset = doorbell_start_offset / sizeof(u32);
93	kfd->doorbell_process_limit = doorbell_process_limit - 1;
94
95	kfd->doorbell_kernel_ptr = ioremap(kfd->doorbell_base,
96						doorbell_process_allocation());
97
98	BUG_ON(!kfd->doorbell_kernel_ptr);
99
100	pr_debug("kfd: doorbell initialization:\n");
101	pr_debug("kfd: doorbell base           == 0x%08lX\n",
102			(uintptr_t)kfd->doorbell_base);
103
104	pr_debug("kfd: doorbell_id_offset      == 0x%08lX\n",
105			kfd->doorbell_id_offset);
106
107	pr_debug("kfd: doorbell_process_limit  == 0x%08lX\n",
108			doorbell_process_limit);
109
110	pr_debug("kfd: doorbell_kernel_offset  == 0x%08lX\n",
111			(uintptr_t)kfd->doorbell_base);
112
113	pr_debug("kfd: doorbell aperture size  == 0x%08lX\n",
114			kfd->shared_resources.doorbell_aperture_size);
115
116	pr_debug("kfd: doorbell kernel address == 0x%08lX\n",
117			(uintptr_t)kfd->doorbell_kernel_ptr);
118}
119
120int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma)
121{
122	phys_addr_t address;
123	struct kfd_dev *dev;
124
125	/*
126	 * For simplicitly we only allow mapping of the entire doorbell
127	 * allocation of a single device & process.
128	 */
129	if (vma->vm_end - vma->vm_start != doorbell_process_allocation())
130		return -EINVAL;
131
132	/* Find kfd device according to gpu id */
133	dev = kfd_device_by_id(vma->vm_pgoff);
134	if (dev == NULL)
135		return -EINVAL;
136
137	/* Calculate physical address of doorbell */
138	address = kfd_get_process_doorbells(dev, process);
139
140	vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
141				VM_DONTDUMP | VM_PFNMAP;
142
143	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
144
145	pr_debug("kfd: mapping doorbell page in kfd_doorbell_mmap\n"
146		 "     target user address == 0x%08llX\n"
147		 "     physical address    == 0x%08llX\n"
148		 "     vm_flags            == 0x%04lX\n"
149		 "     size                == 0x%04lX\n",
150		 (unsigned long long) vma->vm_start, address, vma->vm_flags,
151		 doorbell_process_allocation());
152
153
154	return io_remap_pfn_range(vma,
155				vma->vm_start,
156				address >> PAGE_SHIFT,
157				doorbell_process_allocation(),
158				vma->vm_page_prot);
159}
160
161
162/* get kernel iomem pointer for a doorbell */
163u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
164					unsigned int *doorbell_off)
165{
166	u32 inx;
167
168	BUG_ON(!kfd || !doorbell_off);
169
170	mutex_lock(&kfd->doorbell_mutex);
171	inx = find_first_zero_bit(kfd->doorbell_available_index,
172					KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
173
174	__set_bit(inx, kfd->doorbell_available_index);
175	mutex_unlock(&kfd->doorbell_mutex);
176
177	if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
178		return NULL;
179
180	/*
181	 * Calculating the kernel doorbell offset using "faked" kernel
182	 * pasid that allocated for kernel queues only
183	 */
184	*doorbell_off = KERNEL_DOORBELL_PASID * (doorbell_process_allocation() /
185							sizeof(u32)) + inx;
186
187	pr_debug("kfd: get kernel queue doorbell\n"
188			 "     doorbell offset   == 0x%08d\n"
189			 "     kernel address    == 0x%08lX\n",
190		*doorbell_off, (uintptr_t)(kfd->doorbell_kernel_ptr + inx));
191
192	return kfd->doorbell_kernel_ptr + inx;
193}
194
195void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr)
196{
197	unsigned int inx;
198
199	BUG_ON(!kfd || !db_addr);
200
201	inx = (unsigned int)(db_addr - kfd->doorbell_kernel_ptr);
202
203	mutex_lock(&kfd->doorbell_mutex);
204	__clear_bit(inx, kfd->doorbell_available_index);
205	mutex_unlock(&kfd->doorbell_mutex);
206}
207
208inline void write_kernel_doorbell(u32 __iomem *db, u32 value)
209{
210	if (db) {
211		writel(value, db);
212		pr_debug("writing %d to doorbell address 0x%p\n", value, db);
213	}
214}
215
216/*
217 * queue_ids are in the range [0,MAX_PROCESS_QUEUES) and are mapped 1:1
218 * to doorbells with the process's doorbell page
219 */
220unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd,
221					struct kfd_process *process,
222					unsigned int queue_id)
223{
224	/*
225	 * doorbell_id_offset accounts for doorbells taken by KGD.
226	 * pasid * doorbell_process_allocation/sizeof(u32) adjusts
227	 * to the process's doorbells
228	 */
229	return kfd->doorbell_id_offset +
230		process->pasid * (doorbell_process_allocation()/sizeof(u32)) +
231		queue_id;
232}
233
234uint64_t kfd_get_number_elems(struct kfd_dev *kfd)
235{
236	uint64_t num_of_elems = (kfd->shared_resources.doorbell_aperture_size -
237				kfd->shared_resources.doorbell_start_offset) /
238					doorbell_process_allocation() + 1;
239
240	return num_of_elems;
241
242}
243
244phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
245					struct kfd_process *process)
246{
247	return dev->doorbell_base +
248		process->pasid * doorbell_process_allocation();
249}
250