Loading...
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 int size = PAGE_SIZE;
65 int r;
66
67 /*
68 * Todo: KFD kernel level operations need only one doorbell for
69 * ring test/HWS. So instead of reserving a whole page here for
70 * kernel, reserve and consume a doorbell from existing KGD kernel
71 * doorbell page.
72 */
73
74 /* Bitmap to dynamically allocate doorbells from kernel page */
75 kfd->doorbell_bitmap = bitmap_zalloc(size / sizeof(u32), GFP_KERNEL);
76 if (!kfd->doorbell_bitmap) {
77 DRM_ERROR("Failed to allocate kernel doorbell bitmap\n");
78 return -ENOMEM;
79 }
80
81 /* Alloc a doorbell page for KFD kernel usages */
82 r = amdgpu_bo_create_kernel(kfd->adev,
83 size,
84 PAGE_SIZE,
85 AMDGPU_GEM_DOMAIN_DOORBELL,
86 &kfd->doorbells,
87 NULL,
88 (void **)&kfd->doorbell_kernel_ptr);
89 if (r) {
90 pr_err("failed to allocate kernel doorbells\n");
91 bitmap_free(kfd->doorbell_bitmap);
92 return r;
93 }
94
95 pr_debug("Doorbell kernel address == %p\n", kfd->doorbell_kernel_ptr);
96 return 0;
97}
98
99void kfd_doorbell_fini(struct kfd_dev *kfd)
100{
101 bitmap_free(kfd->doorbell_bitmap);
102 amdgpu_bo_free_kernel(&kfd->doorbells, NULL,
103 (void **)&kfd->doorbell_kernel_ptr);
104}
105
106int kfd_doorbell_mmap(struct kfd_node *dev, struct kfd_process *process,
107 struct vm_area_struct *vma)
108{
109 phys_addr_t address;
110 struct kfd_process_device *pdd;
111
112 /*
113 * For simplicitly we only allow mapping of the entire doorbell
114 * allocation of a single device & process.
115 */
116 if (vma->vm_end - vma->vm_start != kfd_doorbell_process_slice(dev->kfd))
117 return -EINVAL;
118
119 pdd = kfd_get_process_device_data(dev, process);
120 if (!pdd)
121 return -EINVAL;
122
123 /* Calculate physical address of doorbell */
124 address = kfd_get_process_doorbells(pdd);
125 if (!address)
126 return -ENOMEM;
127 vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
128 VM_DONTDUMP | VM_PFNMAP);
129
130 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
131
132 pr_debug("Mapping doorbell page\n"
133 " target user address == 0x%08llX\n"
134 " physical address == 0x%08llX\n"
135 " vm_flags == 0x%04lX\n"
136 " size == 0x%04lX\n",
137 (unsigned long long) vma->vm_start, address, vma->vm_flags,
138 kfd_doorbell_process_slice(dev->kfd));
139
140
141 return io_remap_pfn_range(vma,
142 vma->vm_start,
143 address >> PAGE_SHIFT,
144 kfd_doorbell_process_slice(dev->kfd),
145 vma->vm_page_prot);
146}
147
148
149/* get kernel iomem pointer for a doorbell */
150void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
151 unsigned int *doorbell_off)
152{
153 u32 inx;
154
155 mutex_lock(&kfd->doorbell_mutex);
156 inx = find_first_zero_bit(kfd->doorbell_bitmap, PAGE_SIZE / sizeof(u32));
157
158 __set_bit(inx, kfd->doorbell_bitmap);
159 mutex_unlock(&kfd->doorbell_mutex);
160
161 if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
162 return NULL;
163
164 *doorbell_off = amdgpu_doorbell_index_on_bar(kfd->adev,
165 kfd->doorbells,
166 inx,
167 kfd->device_info.doorbell_size);
168 inx *= 2;
169
170 pr_debug("Get kernel queue doorbell\n"
171 " doorbell offset == 0x%08X\n"
172 " doorbell index == 0x%x\n",
173 *doorbell_off, inx);
174
175 return kfd->doorbell_kernel_ptr + inx;
176}
177
178void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr)
179{
180 unsigned int inx;
181
182 inx = (unsigned int)(db_addr - kfd->doorbell_kernel_ptr);
183 inx /= 2;
184
185 mutex_lock(&kfd->doorbell_mutex);
186 __clear_bit(inx, kfd->doorbell_bitmap);
187 mutex_unlock(&kfd->doorbell_mutex);
188}
189
190void write_kernel_doorbell(void __iomem *db, u32 value)
191{
192 if (db) {
193 writel(value, db);
194 pr_debug("Writing %d to doorbell address %p\n", value, db);
195 }
196}
197
198void write_kernel_doorbell64(void __iomem *db, u64 value)
199{
200 if (db) {
201 WARN(((unsigned long)db & 7) != 0,
202 "Unaligned 64-bit doorbell");
203 writeq(value, (u64 __iomem *)db);
204 pr_debug("writing %llu to doorbell address %p\n", value, db);
205 }
206}
207
208static int init_doorbell_bitmap(struct qcm_process_device *qpd,
209 struct kfd_dev *dev)
210{
211 unsigned int i;
212 int range_start = dev->shared_resources.non_cp_doorbells_start;
213 int range_end = dev->shared_resources.non_cp_doorbells_end;
214
215 if (!KFD_IS_SOC15(dev))
216 return 0;
217
218 /* Mask out doorbells reserved for SDMA, IH, and VCN on SOC15. */
219 pr_debug("reserved doorbell 0x%03x - 0x%03x\n", range_start, range_end);
220 pr_debug("reserved doorbell 0x%03x - 0x%03x\n",
221 range_start + KFD_QUEUE_DOORBELL_MIRROR_OFFSET,
222 range_end + KFD_QUEUE_DOORBELL_MIRROR_OFFSET);
223
224 for (i = 0; i < KFD_MAX_NUM_OF_QUEUES_PER_PROCESS / 2; i++) {
225 if (i >= range_start && i <= range_end) {
226 __set_bit(i, qpd->doorbell_bitmap);
227 __set_bit(i + KFD_QUEUE_DOORBELL_MIRROR_OFFSET,
228 qpd->doorbell_bitmap);
229 }
230 }
231
232 return 0;
233}
234
235phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd)
236{
237 struct amdgpu_device *adev = pdd->dev->adev;
238 uint32_t first_db_index;
239
240 if (!pdd->qpd.proc_doorbells) {
241 if (kfd_alloc_process_doorbells(pdd->dev->kfd, pdd))
242 /* phys_addr_t 0 is error */
243 return 0;
244 }
245
246 first_db_index = amdgpu_doorbell_index_on_bar(adev,
247 pdd->qpd.proc_doorbells,
248 0,
249 pdd->dev->kfd->device_info.doorbell_size);
250 return adev->doorbell.base + first_db_index * sizeof(uint32_t);
251}
252
253int kfd_alloc_process_doorbells(struct kfd_dev *kfd, struct kfd_process_device *pdd)
254{
255 int r;
256 struct qcm_process_device *qpd = &pdd->qpd;
257
258 /* Allocate bitmap for dynamic doorbell allocation */
259 qpd->doorbell_bitmap = bitmap_zalloc(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
260 GFP_KERNEL);
261 if (!qpd->doorbell_bitmap) {
262 DRM_ERROR("Failed to allocate process doorbell bitmap\n");
263 return -ENOMEM;
264 }
265
266 r = init_doorbell_bitmap(&pdd->qpd, kfd);
267 if (r) {
268 DRM_ERROR("Failed to initialize process doorbells\n");
269 r = -ENOMEM;
270 goto err;
271 }
272
273 /* Allocate doorbells for this process */
274 r = amdgpu_bo_create_kernel(kfd->adev,
275 kfd_doorbell_process_slice(kfd),
276 PAGE_SIZE,
277 AMDGPU_GEM_DOMAIN_DOORBELL,
278 &qpd->proc_doorbells,
279 NULL,
280 NULL);
281 if (r) {
282 DRM_ERROR("Failed to allocate process doorbells\n");
283 goto err;
284 }
285 return 0;
286
287err:
288 bitmap_free(qpd->doorbell_bitmap);
289 qpd->doorbell_bitmap = NULL;
290 return r;
291}
292
293void kfd_free_process_doorbells(struct kfd_dev *kfd, struct kfd_process_device *pdd)
294{
295 struct qcm_process_device *qpd = &pdd->qpd;
296
297 if (qpd->doorbell_bitmap) {
298 bitmap_free(qpd->doorbell_bitmap);
299 qpd->doorbell_bitmap = NULL;
300 }
301
302 amdgpu_bo_free_kernel(&qpd->proc_doorbells, NULL, NULL);
303}
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}