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 #include <linux/idr.h>
28
29 /*
30 * This extension supports a kernel level doorbells management for the
31 * kernel queues using the first doorbell page reserved for the kernel.
32 */
33
34 static DEFINE_IDA(doorbell_ida);
35 static unsigned int max_doorbell_slices;
36
37 /*
38 * Each device exposes a doorbell aperture, a PCI MMIO aperture that
39 * receives 32-bit writes that are passed to queues as wptr values.
40 * The doorbells are intended to be written by applications as part
41 * of queueing work on user-mode queues.
42 * We assign doorbells to applications in PAGE_SIZE-sized and aligned chunks.
43 * We map the doorbell address space into user-mode when a process creates
44 * its first queue on each device.
45 * Although the mapping is done by KFD, it is equivalent to an mmap of
46 * the /dev/kfd with the particular device encoded in the mmap offset.
47 * There will be other uses for mmap of /dev/kfd, so only a range of
48 * offsets (KFD_MMAP_DOORBELL_START-END) is used for doorbells.
49 */
50
51 /* # of doorbell bytes allocated for each process. */
kfd_doorbell_process_slice(struct kfd_dev * kfd)52 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd)
53 {
54 return roundup(kfd->device_info->doorbell_size *
55 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
56 PAGE_SIZE);
57 }
58
59 /* Doorbell calculations for device init. */
kfd_doorbell_init(struct kfd_dev * kfd)60 int kfd_doorbell_init(struct kfd_dev *kfd)
61 {
62 size_t doorbell_start_offset;
63 size_t doorbell_aperture_size;
64 size_t doorbell_process_limit;
65
66 /*
67 * We start with calculations in bytes because the input data might
68 * only be byte-aligned.
69 * Only after we have done the rounding can we assume any alignment.
70 */
71
72 doorbell_start_offset =
73 roundup(kfd->shared_resources.doorbell_start_offset,
74 kfd_doorbell_process_slice(kfd));
75
76 doorbell_aperture_size =
77 rounddown(kfd->shared_resources.doorbell_aperture_size,
78 kfd_doorbell_process_slice(kfd));
79
80 if (doorbell_aperture_size > doorbell_start_offset)
81 doorbell_process_limit =
82 (doorbell_aperture_size - doorbell_start_offset) /
83 kfd_doorbell_process_slice(kfd);
84 else
85 return -ENOSPC;
86
87 if (!max_doorbell_slices ||
88 doorbell_process_limit < max_doorbell_slices)
89 max_doorbell_slices = doorbell_process_limit;
90
91 kfd->doorbell_base = kfd->shared_resources.doorbell_physical_address +
92 doorbell_start_offset;
93
94 kfd->doorbell_id_offset = doorbell_start_offset / sizeof(u32);
95
96 kfd->doorbell_kernel_ptr = ioremap(kfd->doorbell_base,
97 kfd_doorbell_process_slice(kfd));
98
99 if (!kfd->doorbell_kernel_ptr)
100 return -ENOMEM;
101
102 pr_debug("Doorbell initialization:\n");
103 pr_debug("doorbell base == 0x%08lX\n",
104 (uintptr_t)kfd->doorbell_base);
105
106 pr_debug("doorbell_id_offset == 0x%08lX\n",
107 kfd->doorbell_id_offset);
108
109 pr_debug("doorbell_process_limit == 0x%08lX\n",
110 doorbell_process_limit);
111
112 pr_debug("doorbell_kernel_offset == 0x%08lX\n",
113 (uintptr_t)kfd->doorbell_base);
114
115 pr_debug("doorbell aperture size == 0x%08lX\n",
116 kfd->shared_resources.doorbell_aperture_size);
117
118 pr_debug("doorbell kernel address == %p\n", 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_dev * dev,struct kfd_process * process,struct vm_area_struct * vma)129 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process,
130 struct vm_area_struct *vma)
131 {
132 phys_addr_t address;
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 != kfd_doorbell_process_slice(dev))
139 return -EINVAL;
140
141 /* Calculate physical address of doorbell */
142 address = kfd_get_process_doorbells(dev, process);
143
144 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
145 VM_DONTDUMP | VM_PFNMAP;
146
147 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
148
149 pr_debug("Mapping doorbell page\n"
150 " target user address == 0x%08llX\n"
151 " physical address == 0x%08llX\n"
152 " vm_flags == 0x%04lX\n"
153 " size == 0x%04lX\n",
154 (unsigned long long) vma->vm_start, address, vma->vm_flags,
155 kfd_doorbell_process_slice(dev));
156
157
158 return io_remap_pfn_range(vma,
159 vma->vm_start,
160 address >> PAGE_SHIFT,
161 kfd_doorbell_process_slice(dev),
162 vma->vm_page_prot);
163 }
164
165
166 /* get kernel iomem pointer for a doorbell */
kfd_get_kernel_doorbell(struct kfd_dev * kfd,unsigned int * doorbell_off)167 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
168 unsigned int *doorbell_off)
169 {
170 u32 inx;
171
172 mutex_lock(&kfd->doorbell_mutex);
173 inx = find_first_zero_bit(kfd->doorbell_available_index,
174 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
175
176 __set_bit(inx, kfd->doorbell_available_index);
177 mutex_unlock(&kfd->doorbell_mutex);
178
179 if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
180 return NULL;
181
182 inx *= kfd->device_info->doorbell_size / sizeof(u32);
183
184 /*
185 * Calculating the kernel doorbell offset using the first
186 * doorbell page.
187 */
188 *doorbell_off = kfd->doorbell_id_offset + inx;
189
190 pr_debug("Get kernel queue doorbell\n"
191 " doorbell offset == 0x%08X\n"
192 " doorbell index == 0x%x\n",
193 *doorbell_off, inx);
194
195 return kfd->doorbell_kernel_ptr + inx;
196 }
197
kfd_release_kernel_doorbell(struct kfd_dev * kfd,u32 __iomem * db_addr)198 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr)
199 {
200 unsigned int inx;
201
202 inx = (unsigned int)(db_addr - kfd->doorbell_kernel_ptr)
203 * sizeof(u32) / kfd->device_info->doorbell_size;
204
205 mutex_lock(&kfd->doorbell_mutex);
206 __clear_bit(inx, kfd->doorbell_available_index);
207 mutex_unlock(&kfd->doorbell_mutex);
208 }
209
write_kernel_doorbell(void __iomem * db,u32 value)210 void write_kernel_doorbell(void __iomem *db, u32 value)
211 {
212 if (db) {
213 writel(value, db);
214 pr_debug("Writing %d to doorbell address %p\n", value, db);
215 }
216 }
217
write_kernel_doorbell64(void __iomem * db,u64 value)218 void write_kernel_doorbell64(void __iomem *db, u64 value)
219 {
220 if (db) {
221 WARN(((unsigned long)db & 7) != 0,
222 "Unaligned 64-bit doorbell");
223 writeq(value, (u64 __iomem *)db);
224 pr_debug("writing %llu to doorbell address %p\n", value, db);
225 }
226 }
227
kfd_doorbell_id_to_offset(struct kfd_dev * kfd,struct kfd_process * process,unsigned int doorbell_id)228 unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd,
229 struct kfd_process *process,
230 unsigned int doorbell_id)
231 {
232 /*
233 * doorbell_id_offset accounts for doorbells taken by KGD.
234 * index * kfd_doorbell_process_slice/sizeof(u32) adjusts to
235 * the process's doorbells. The offset returned is in dword
236 * units regardless of the ASIC-dependent doorbell size.
237 */
238 return kfd->doorbell_id_offset +
239 process->doorbell_index
240 * kfd_doorbell_process_slice(kfd) / sizeof(u32) +
241 doorbell_id * kfd->device_info->doorbell_size / sizeof(u32);
242 }
243
kfd_get_number_elems(struct kfd_dev * kfd)244 uint64_t kfd_get_number_elems(struct kfd_dev *kfd)
245 {
246 uint64_t num_of_elems = (kfd->shared_resources.doorbell_aperture_size -
247 kfd->shared_resources.doorbell_start_offset) /
248 kfd_doorbell_process_slice(kfd) + 1;
249
250 return num_of_elems;
251
252 }
253
kfd_get_process_doorbells(struct kfd_dev * dev,struct kfd_process * process)254 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
255 struct kfd_process *process)
256 {
257 return dev->doorbell_base +
258 process->doorbell_index * kfd_doorbell_process_slice(dev);
259 }
260
kfd_alloc_process_doorbells(struct kfd_process * process)261 int kfd_alloc_process_doorbells(struct kfd_process *process)
262 {
263 int r = ida_simple_get(&doorbell_ida, 1, max_doorbell_slices,
264 GFP_KERNEL);
265 if (r > 0)
266 process->doorbell_index = r;
267
268 return r;
269 }
270
kfd_free_process_doorbells(struct kfd_process * process)271 void kfd_free_process_doorbells(struct kfd_process *process)
272 {
273 if (process->doorbell_index)
274 ida_simple_remove(&doorbell_ida, process->doorbell_index);
275 }
276