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. */
doorbell_process_allocation(void)54 static 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. */
kfd_doorbell_init(struct kfd_dev * kfd)62 int 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 if (!kfd->doorbell_kernel_ptr)
99 return -ENOMEM;
100
101 pr_debug("Doorbell initialization:\n");
102 pr_debug("doorbell base == 0x%08lX\n",
103 (uintptr_t)kfd->doorbell_base);
104
105 pr_debug("doorbell_id_offset == 0x%08lX\n",
106 kfd->doorbell_id_offset);
107
108 pr_debug("doorbell_process_limit == 0x%08lX\n",
109 doorbell_process_limit);
110
111 pr_debug("doorbell_kernel_offset == 0x%08lX\n",
112 (uintptr_t)kfd->doorbell_base);
113
114 pr_debug("doorbell aperture size == 0x%08lX\n",
115 kfd->shared_resources.doorbell_aperture_size);
116
117 pr_debug("doorbell kernel address == 0x%08lX\n",
118 (uintptr_t)kfd->doorbell_kernel_ptr);
119
120 return 0;
121 }
122
kfd_doorbell_fini(struct kfd_dev * kfd)123 void kfd_doorbell_fini(struct kfd_dev *kfd)
124 {
125 if (kfd->doorbell_kernel_ptr)
126 iounmap(kfd->doorbell_kernel_ptr);
127 }
128
kfd_doorbell_mmap(struct kfd_process * process,struct vm_area_struct * vma)129 int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma)
130 {
131 phys_addr_t address;
132 struct kfd_dev *dev;
133
134 /*
135 * For simplicitly we only allow mapping of the entire doorbell
136 * allocation of a single device & process.
137 */
138 if (vma->vm_end - vma->vm_start != doorbell_process_allocation())
139 return -EINVAL;
140
141 /* Find kfd device according to gpu id */
142 dev = kfd_device_by_id(vma->vm_pgoff);
143 if (!dev)
144 return -EINVAL;
145
146 /* Calculate physical address of doorbell */
147 address = kfd_get_process_doorbells(dev, process);
148
149 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
150 VM_DONTDUMP | VM_PFNMAP;
151
152 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
153
154 pr_debug("Mapping doorbell page\n"
155 " target user address == 0x%08llX\n"
156 " physical address == 0x%08llX\n"
157 " vm_flags == 0x%04lX\n"
158 " size == 0x%04lX\n",
159 (unsigned long long) vma->vm_start, address, vma->vm_flags,
160 doorbell_process_allocation());
161
162
163 return io_remap_pfn_range(vma,
164 vma->vm_start,
165 address >> PAGE_SHIFT,
166 doorbell_process_allocation(),
167 vma->vm_page_prot);
168 }
169
170
171 /* get kernel iomem pointer for a doorbell */
kfd_get_kernel_doorbell(struct kfd_dev * kfd,unsigned int * doorbell_off)172 u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
173 unsigned int *doorbell_off)
174 {
175 u32 inx;
176
177 mutex_lock(&kfd->doorbell_mutex);
178 inx = find_first_zero_bit(kfd->doorbell_available_index,
179 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
180
181 __set_bit(inx, kfd->doorbell_available_index);
182 mutex_unlock(&kfd->doorbell_mutex);
183
184 if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
185 return NULL;
186
187 /*
188 * Calculating the kernel doorbell offset using "faked" kernel
189 * pasid that allocated for kernel queues only
190 */
191 *doorbell_off = KERNEL_DOORBELL_PASID * (doorbell_process_allocation() /
192 sizeof(u32)) + inx;
193
194 pr_debug("Get kernel queue doorbell\n"
195 " doorbell offset == 0x%08X\n"
196 " kernel address == 0x%08lX\n",
197 *doorbell_off, (uintptr_t)(kfd->doorbell_kernel_ptr + inx));
198
199 return kfd->doorbell_kernel_ptr + inx;
200 }
201
kfd_release_kernel_doorbell(struct kfd_dev * kfd,u32 __iomem * db_addr)202 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr)
203 {
204 unsigned int inx;
205
206 inx = (unsigned int)(db_addr - kfd->doorbell_kernel_ptr);
207
208 mutex_lock(&kfd->doorbell_mutex);
209 __clear_bit(inx, kfd->doorbell_available_index);
210 mutex_unlock(&kfd->doorbell_mutex);
211 }
212
write_kernel_doorbell(u32 __iomem * db,u32 value)213 inline void write_kernel_doorbell(u32 __iomem *db, u32 value)
214 {
215 if (db) {
216 writel(value, db);
217 pr_debug("Writing %d to doorbell address 0x%p\n", value, db);
218 }
219 }
220
221 /*
222 * queue_ids are in the range [0,MAX_PROCESS_QUEUES) and are mapped 1:1
223 * to doorbells with the process's doorbell page
224 */
kfd_queue_id_to_doorbell(struct kfd_dev * kfd,struct kfd_process * process,unsigned int queue_id)225 unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd,
226 struct kfd_process *process,
227 unsigned int queue_id)
228 {
229 /*
230 * doorbell_id_offset accounts for doorbells taken by KGD.
231 * pasid * doorbell_process_allocation/sizeof(u32) adjusts
232 * to the process's doorbells
233 */
234 return kfd->doorbell_id_offset +
235 process->pasid * (doorbell_process_allocation()/sizeof(u32)) +
236 queue_id;
237 }
238
kfd_get_number_elems(struct kfd_dev * kfd)239 uint64_t kfd_get_number_elems(struct kfd_dev *kfd)
240 {
241 uint64_t num_of_elems = (kfd->shared_resources.doorbell_aperture_size -
242 kfd->shared_resources.doorbell_start_offset) /
243 doorbell_process_allocation() + 1;
244
245 return num_of_elems;
246
247 }
248
kfd_get_process_doorbells(struct kfd_dev * dev,struct kfd_process * process)249 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
250 struct kfd_process *process)
251 {
252 return dev->doorbell_base +
253 process->pasid * doorbell_process_allocation();
254 }
255