Loading...
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("mapping doorbell page:\n");
146 pr_debug(" target user address == 0x%08llX\n",
147 (unsigned long long) vma->vm_start);
148 pr_debug(" physical address == 0x%08llX\n", address);
149 pr_debug(" vm_flags == 0x%04lX\n", vma->vm_flags);
150 pr_debug(" size == 0x%04lX\n",
151 doorbell_process_allocation());
152
153 return io_remap_pfn_range(vma,
154 vma->vm_start,
155 address >> PAGE_SHIFT,
156 doorbell_process_allocation(),
157 vma->vm_page_prot);
158}
159
160
161/* get kernel iomem pointer for a doorbell */
162u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
163 unsigned int *doorbell_off)
164{
165 u32 inx;
166
167 BUG_ON(!kfd || !doorbell_off);
168
169 mutex_lock(&kfd->doorbell_mutex);
170 inx = find_first_zero_bit(kfd->doorbell_available_index,
171 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
172
173 __set_bit(inx, kfd->doorbell_available_index);
174 mutex_unlock(&kfd->doorbell_mutex);
175
176 if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
177 return NULL;
178
179 /*
180 * Calculating the kernel doorbell offset using "faked" kernel
181 * pasid that allocated for kernel queues only
182 */
183 *doorbell_off = KERNEL_DOORBELL_PASID * (doorbell_process_allocation() /
184 sizeof(u32)) + inx;
185
186 pr_debug("kfd: get kernel queue doorbell\n"
187 " doorbell offset == 0x%08d\n"
188 " kernel address == 0x%08lX\n",
189 *doorbell_off, (uintptr_t)(kfd->doorbell_kernel_ptr + inx));
190
191 return kfd->doorbell_kernel_ptr + inx;
192}
193
194void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr)
195{
196 unsigned int inx;
197
198 BUG_ON(!kfd || !db_addr);
199
200 inx = (unsigned int)(db_addr - kfd->doorbell_kernel_ptr);
201
202 mutex_lock(&kfd->doorbell_mutex);
203 __clear_bit(inx, kfd->doorbell_available_index);
204 mutex_unlock(&kfd->doorbell_mutex);
205}
206
207inline void write_kernel_doorbell(u32 __iomem *db, u32 value)
208{
209 if (db) {
210 writel(value, db);
211 pr_debug("writing %d to doorbell address 0x%p\n", value, db);
212 }
213}
214
215/*
216 * queue_ids are in the range [0,MAX_PROCESS_QUEUES) and are mapped 1:1
217 * to doorbells with the process's doorbell page
218 */
219unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd,
220 struct kfd_process *process,
221 unsigned int queue_id)
222{
223 /*
224 * doorbell_id_offset accounts for doorbells taken by KGD.
225 * pasid * doorbell_process_allocation/sizeof(u32) adjusts
226 * to the process's doorbells
227 */
228 return kfd->doorbell_id_offset +
229 process->pasid * (doorbell_process_allocation()/sizeof(u32)) +
230 queue_id;
231}
232
233uint64_t kfd_get_number_elems(struct kfd_dev *kfd)
234{
235 uint64_t num_of_elems = (kfd->shared_resources.doorbell_aperture_size -
236 kfd->shared_resources.doorbell_start_offset) /
237 doorbell_process_allocation() + 1;
238
239 return num_of_elems;
240
241}
242
243phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
244 struct kfd_process *process)
245{
246 return dev->doorbell_base +
247 process->pasid * doorbell_process_allocation();
248}
1// SPDX-License-Identifier: GPL-2.0 OR MIT
2/*
3 * Copyright 2014-2022 Advanced Micro Devices, Inc.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice shall be included in
13 * all copies or substantial portions of the Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21 * OTHER DEALINGS IN THE SOFTWARE.
22 */
23#include "kfd_priv.h"
24#include <linux/mm.h>
25#include <linux/mman.h>
26#include <linux/slab.h>
27#include <linux/io.h>
28#include <linux/idr.h>
29
30/*
31 * This extension supports a kernel level doorbells management for the
32 * kernel queues using the first doorbell page reserved for the kernel.
33 */
34
35/*
36 * Each device exposes a doorbell aperture, a PCI MMIO aperture that
37 * receives 32-bit writes that are passed to queues as wptr values.
38 * The doorbells are intended to be written by applications as part
39 * of queueing work on user-mode queues.
40 * We assign doorbells to applications in PAGE_SIZE-sized and aligned chunks.
41 * We map the doorbell address space into user-mode when a process creates
42 * its first queue on each device.
43 * Although the mapping is done by KFD, it is equivalent to an mmap of
44 * the /dev/kfd with the particular device encoded in the mmap offset.
45 * There will be other uses for mmap of /dev/kfd, so only a range of
46 * offsets (KFD_MMAP_DOORBELL_START-END) is used for doorbells.
47 */
48
49/* # of doorbell bytes allocated for each process. */
50size_t kfd_doorbell_process_slice(struct kfd_dev *kfd)
51{
52 if (!kfd->shared_resources.enable_mes)
53 return roundup(kfd->device_info.doorbell_size *
54 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
55 PAGE_SIZE);
56 else
57 return amdgpu_mes_doorbell_process_slice(
58 (struct amdgpu_device *)kfd->adev);
59}
60
61/* Doorbell calculations for device init. */
62int 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 * With MES enabled, just set the doorbell base as it is needed
70 * to calculate doorbell physical address.
71 */
72 if (kfd->shared_resources.enable_mes) {
73 kfd->doorbell_base =
74 kfd->shared_resources.doorbell_physical_address;
75 return 0;
76 }
77
78 /*
79 * We start with calculations in bytes because the input data might
80 * only be byte-aligned.
81 * Only after we have done the rounding can we assume any alignment.
82 */
83
84 doorbell_start_offset =
85 roundup(kfd->shared_resources.doorbell_start_offset,
86 kfd_doorbell_process_slice(kfd));
87
88 doorbell_aperture_size =
89 rounddown(kfd->shared_resources.doorbell_aperture_size,
90 kfd_doorbell_process_slice(kfd));
91
92 if (doorbell_aperture_size > doorbell_start_offset)
93 doorbell_process_limit =
94 (doorbell_aperture_size - doorbell_start_offset) /
95 kfd_doorbell_process_slice(kfd);
96 else
97 return -ENOSPC;
98
99 if (!kfd->max_doorbell_slices ||
100 doorbell_process_limit < kfd->max_doorbell_slices)
101 kfd->max_doorbell_slices = doorbell_process_limit;
102
103 kfd->doorbell_base = kfd->shared_resources.doorbell_physical_address +
104 doorbell_start_offset;
105
106 kfd->doorbell_base_dw_offset = doorbell_start_offset / sizeof(u32);
107
108 kfd->doorbell_kernel_ptr = ioremap(kfd->doorbell_base,
109 kfd_doorbell_process_slice(kfd));
110
111 if (!kfd->doorbell_kernel_ptr)
112 return -ENOMEM;
113
114 pr_debug("Doorbell initialization:\n");
115 pr_debug("doorbell base == 0x%08lX\n",
116 (uintptr_t)kfd->doorbell_base);
117
118 pr_debug("doorbell_base_dw_offset == 0x%08lX\n",
119 kfd->doorbell_base_dw_offset);
120
121 pr_debug("doorbell_process_limit == 0x%08lX\n",
122 doorbell_process_limit);
123
124 pr_debug("doorbell_kernel_offset == 0x%08lX\n",
125 (uintptr_t)kfd->doorbell_base);
126
127 pr_debug("doorbell aperture size == 0x%08lX\n",
128 kfd->shared_resources.doorbell_aperture_size);
129
130 pr_debug("doorbell kernel address == %p\n", kfd->doorbell_kernel_ptr);
131
132 return 0;
133}
134
135void kfd_doorbell_fini(struct kfd_dev *kfd)
136{
137 if (kfd->doorbell_kernel_ptr)
138 iounmap(kfd->doorbell_kernel_ptr);
139}
140
141int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process,
142 struct vm_area_struct *vma)
143{
144 phys_addr_t address;
145 struct kfd_process_device *pdd;
146
147 /*
148 * For simplicitly we only allow mapping of the entire doorbell
149 * allocation of a single device & process.
150 */
151 if (vma->vm_end - vma->vm_start != kfd_doorbell_process_slice(dev))
152 return -EINVAL;
153
154 pdd = kfd_get_process_device_data(dev, process);
155 if (!pdd)
156 return -EINVAL;
157
158 /* Calculate physical address of doorbell */
159 address = kfd_get_process_doorbells(pdd);
160 if (!address)
161 return -ENOMEM;
162 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
163 VM_DONTDUMP | VM_PFNMAP;
164
165 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
166
167 pr_debug("Mapping doorbell page\n"
168 " target user address == 0x%08llX\n"
169 " physical address == 0x%08llX\n"
170 " vm_flags == 0x%04lX\n"
171 " size == 0x%04lX\n",
172 (unsigned long long) vma->vm_start, address, vma->vm_flags,
173 kfd_doorbell_process_slice(dev));
174
175
176 return io_remap_pfn_range(vma,
177 vma->vm_start,
178 address >> PAGE_SHIFT,
179 kfd_doorbell_process_slice(dev),
180 vma->vm_page_prot);
181}
182
183
184/* get kernel iomem pointer for a doorbell */
185void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
186 unsigned int *doorbell_off)
187{
188 u32 inx;
189
190 mutex_lock(&kfd->doorbell_mutex);
191 inx = find_first_zero_bit(kfd->doorbell_available_index,
192 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
193
194 __set_bit(inx, kfd->doorbell_available_index);
195 mutex_unlock(&kfd->doorbell_mutex);
196
197 if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
198 return NULL;
199
200 inx *= kfd->device_info.doorbell_size / sizeof(u32);
201
202 /*
203 * Calculating the kernel doorbell offset using the first
204 * doorbell page.
205 */
206 *doorbell_off = kfd->doorbell_base_dw_offset + inx;
207
208 pr_debug("Get kernel queue doorbell\n"
209 " doorbell offset == 0x%08X\n"
210 " doorbell index == 0x%x\n",
211 *doorbell_off, inx);
212
213 return kfd->doorbell_kernel_ptr + inx;
214}
215
216void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr)
217{
218 unsigned int inx;
219
220 inx = (unsigned int)(db_addr - kfd->doorbell_kernel_ptr)
221 * sizeof(u32) / kfd->device_info.doorbell_size;
222
223 mutex_lock(&kfd->doorbell_mutex);
224 __clear_bit(inx, kfd->doorbell_available_index);
225 mutex_unlock(&kfd->doorbell_mutex);
226}
227
228void write_kernel_doorbell(void __iomem *db, u32 value)
229{
230 if (db) {
231 writel(value, db);
232 pr_debug("Writing %d to doorbell address %p\n", value, db);
233 }
234}
235
236void write_kernel_doorbell64(void __iomem *db, u64 value)
237{
238 if (db) {
239 WARN(((unsigned long)db & 7) != 0,
240 "Unaligned 64-bit doorbell");
241 writeq(value, (u64 __iomem *)db);
242 pr_debug("writing %llu to doorbell address %p\n", value, db);
243 }
244}
245
246unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd,
247 struct kfd_process_device *pdd,
248 unsigned int doorbell_id)
249{
250 /*
251 * doorbell_base_dw_offset accounts for doorbells taken by KGD.
252 * index * kfd_doorbell_process_slice/sizeof(u32) adjusts to
253 * the process's doorbells. The offset returned is in dword
254 * units regardless of the ASIC-dependent doorbell size.
255 */
256 if (!kfd->shared_resources.enable_mes)
257 return kfd->doorbell_base_dw_offset +
258 pdd->doorbell_index
259 * kfd_doorbell_process_slice(kfd) / sizeof(u32) +
260 doorbell_id *
261 kfd->device_info.doorbell_size / sizeof(u32);
262 else
263 return amdgpu_mes_get_doorbell_dw_offset_in_bar(
264 (struct amdgpu_device *)kfd->adev,
265 pdd->doorbell_index, doorbell_id);
266}
267
268uint64_t kfd_get_number_elems(struct kfd_dev *kfd)
269{
270 uint64_t num_of_elems = (kfd->shared_resources.doorbell_aperture_size -
271 kfd->shared_resources.doorbell_start_offset) /
272 kfd_doorbell_process_slice(kfd) + 1;
273
274 return num_of_elems;
275
276}
277
278phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd)
279{
280 if (!pdd->doorbell_index) {
281 int r = kfd_alloc_process_doorbells(pdd->dev,
282 &pdd->doorbell_index);
283 if (r)
284 return 0;
285 }
286
287 return pdd->dev->doorbell_base +
288 pdd->doorbell_index * kfd_doorbell_process_slice(pdd->dev);
289}
290
291int kfd_alloc_process_doorbells(struct kfd_dev *kfd, unsigned int *doorbell_index)
292{
293 int r = 0;
294
295 if (!kfd->shared_resources.enable_mes)
296 r = ida_simple_get(&kfd->doorbell_ida, 1,
297 kfd->max_doorbell_slices, GFP_KERNEL);
298 else
299 r = amdgpu_mes_alloc_process_doorbells(
300 (struct amdgpu_device *)kfd->adev,
301 doorbell_index);
302
303 if (r > 0)
304 *doorbell_index = r;
305
306 if (r < 0)
307 pr_err("Failed to allocate process doorbells\n");
308
309 return r;
310}
311
312void kfd_free_process_doorbells(struct kfd_dev *kfd, unsigned int doorbell_index)
313{
314 if (doorbell_index) {
315 if (!kfd->shared_resources.enable_mes)
316 ida_simple_remove(&kfd->doorbell_ida, doorbell_index);
317 else
318 amdgpu_mes_free_process_doorbells(
319 (struct amdgpu_device *)kfd->adev,
320 doorbell_index);
321 }
322}