1 /* $NetBSD: kfd_doorbell.c,v 1.3 2021/12/18 23:44:59 riastradh Exp $ */
2
3 /*
4 * Copyright 2014 Advanced Micro Devices, Inc.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
23 */
24 #include <sys/cdefs.h>
25 __KERNEL_RCSID(0, "$NetBSD: kfd_doorbell.c,v 1.3 2021/12/18 23:44:59 riastradh Exp $");
26
27 #include "kfd_priv.h"
28 #include <linux/mm.h>
29 #include <linux/mman.h>
30 #include <linux/slab.h>
31 #include <linux/io.h>
32 #include <linux/idr.h>
33
34 /*
35 * This extension supports a kernel level doorbells management for the
36 * kernel queues using the first doorbell page reserved for the kernel.
37 */
38
39 static DEFINE_IDA(doorbell_ida);
40 static unsigned int max_doorbell_slices;
41
42 /*
43 * Each device exposes a doorbell aperture, a PCI MMIO aperture that
44 * receives 32-bit writes that are passed to queues as wptr values.
45 * The doorbells are intended to be written by applications as part
46 * of queueing work on user-mode queues.
47 * We assign doorbells to applications in PAGE_SIZE-sized and aligned chunks.
48 * We map the doorbell address space into user-mode when a process creates
49 * its first queue on each device.
50 * Although the mapping is done by KFD, it is equivalent to an mmap of
51 * the /dev/kfd with the particular device encoded in the mmap offset.
52 * There will be other uses for mmap of /dev/kfd, so only a range of
53 * offsets (KFD_MMAP_DOORBELL_START-END) is used for doorbells.
54 */
55
56 /* # of doorbell bytes allocated for each process. */
kfd_doorbell_process_slice(struct kfd_dev * kfd)57 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd)
58 {
59 return roundup(kfd->device_info->doorbell_size *
60 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
61 PAGE_SIZE);
62 }
63
64 /* Doorbell calculations for device init. */
kfd_doorbell_init(struct kfd_dev * kfd)65 int kfd_doorbell_init(struct kfd_dev *kfd)
66 {
67 size_t doorbell_start_offset;
68 size_t doorbell_aperture_size;
69 size_t doorbell_process_limit;
70
71 /*
72 * We start with calculations in bytes because the input data might
73 * only be byte-aligned.
74 * Only after we have done the rounding can we assume any alignment.
75 */
76
77 doorbell_start_offset =
78 roundup(kfd->shared_resources.doorbell_start_offset,
79 kfd_doorbell_process_slice(kfd));
80
81 doorbell_aperture_size =
82 rounddown(kfd->shared_resources.doorbell_aperture_size,
83 kfd_doorbell_process_slice(kfd));
84
85 if (doorbell_aperture_size > doorbell_start_offset)
86 doorbell_process_limit =
87 (doorbell_aperture_size - doorbell_start_offset) /
88 kfd_doorbell_process_slice(kfd);
89 else
90 return -ENOSPC;
91
92 if (!max_doorbell_slices ||
93 doorbell_process_limit < max_doorbell_slices)
94 max_doorbell_slices = doorbell_process_limit;
95
96 kfd->doorbell_base = kfd->shared_resources.doorbell_physical_address +
97 doorbell_start_offset;
98
99 kfd->doorbell_base_dw_offset = doorbell_start_offset / sizeof(u32);
100
101 kfd->doorbell_kernel_ptr = ioremap(kfd->doorbell_base,
102 kfd_doorbell_process_slice(kfd));
103
104 if (!kfd->doorbell_kernel_ptr)
105 return -ENOMEM;
106
107 pr_debug("Doorbell initialization:\n");
108 pr_debug("doorbell base == 0x%08lX\n",
109 (uintptr_t)kfd->doorbell_base);
110
111 pr_debug("doorbell_base_dw_offset == 0x%08lX\n",
112 kfd->doorbell_base_dw_offset);
113
114 pr_debug("doorbell_process_limit == 0x%08lX\n",
115 doorbell_process_limit);
116
117 pr_debug("doorbell_kernel_offset == 0x%08lX\n",
118 (uintptr_t)kfd->doorbell_base);
119
120 pr_debug("doorbell aperture size == 0x%08lX\n",
121 kfd->shared_resources.doorbell_aperture_size);
122
123 pr_debug("doorbell kernel address == %p\n", kfd->doorbell_kernel_ptr);
124
125 return 0;
126 }
127
kfd_doorbell_fini(struct kfd_dev * kfd)128 void kfd_doorbell_fini(struct kfd_dev *kfd)
129 {
130 if (kfd->doorbell_kernel_ptr)
131 iounmap(kfd->doorbell_kernel_ptr);
132 }
133
kfd_doorbell_mmap(struct kfd_dev * dev,struct kfd_process * process,struct vm_area_struct * vma)134 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process,
135 struct vm_area_struct *vma)
136 {
137 phys_addr_t address;
138
139 /*
140 * For simplicitly we only allow mapping of the entire doorbell
141 * allocation of a single device & process.
142 */
143 if (vma->vm_end - vma->vm_start != kfd_doorbell_process_slice(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 kfd_doorbell_process_slice(dev));
161
162
163 return io_remap_pfn_range(vma,
164 vma->vm_start,
165 address >> PAGE_SHIFT,
166 kfd_doorbell_process_slice(dev),
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 void __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 inx *= kfd->device_info->doorbell_size / sizeof(u32);
188
189 /*
190 * Calculating the kernel doorbell offset using the first
191 * doorbell page.
192 */
193 *doorbell_off = kfd->doorbell_base_dw_offset + inx;
194
195 pr_debug("Get kernel queue doorbell\n"
196 " doorbell offset == 0x%08X\n"
197 " doorbell index == 0x%x\n",
198 *doorbell_off, inx);
199
200 return kfd->doorbell_kernel_ptr + inx;
201 }
202
kfd_release_kernel_doorbell(struct kfd_dev * kfd,u32 __iomem * db_addr)203 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr)
204 {
205 unsigned int inx;
206
207 inx = (unsigned int)(db_addr - kfd->doorbell_kernel_ptr)
208 * sizeof(u32) / kfd->device_info->doorbell_size;
209
210 mutex_lock(&kfd->doorbell_mutex);
211 __clear_bit(inx, kfd->doorbell_available_index);
212 mutex_unlock(&kfd->doorbell_mutex);
213 }
214
write_kernel_doorbell(void __iomem * db,u32 value)215 void write_kernel_doorbell(void __iomem *db, u32 value)
216 {
217 if (db) {
218 writel(value, db);
219 pr_debug("Writing %d to doorbell address %p\n", value, db);
220 }
221 }
222
write_kernel_doorbell64(void __iomem * db,u64 value)223 void write_kernel_doorbell64(void __iomem *db, u64 value)
224 {
225 if (db) {
226 WARN(((unsigned long)db & 7) != 0,
227 "Unaligned 64-bit doorbell");
228 writeq(value, (u64 __iomem *)db);
229 pr_debug("writing %llu to doorbell address %p\n", value, db);
230 }
231 }
232
kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev * kfd,struct kfd_process * process,unsigned int doorbell_id)233 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd,
234 struct kfd_process *process,
235 unsigned int doorbell_id)
236 {
237 /*
238 * doorbell_base_dw_offset accounts for doorbells taken by KGD.
239 * index * kfd_doorbell_process_slice/sizeof(u32) adjusts to
240 * the process's doorbells. The offset returned is in dword
241 * units regardless of the ASIC-dependent doorbell size.
242 */
243 return kfd->doorbell_base_dw_offset +
244 process->doorbell_index
245 * kfd_doorbell_process_slice(kfd) / sizeof(u32) +
246 doorbell_id * kfd->device_info->doorbell_size / sizeof(u32);
247 }
248
kfd_get_number_elems(struct kfd_dev * kfd)249 uint64_t kfd_get_number_elems(struct kfd_dev *kfd)
250 {
251 uint64_t num_of_elems = (kfd->shared_resources.doorbell_aperture_size -
252 kfd->shared_resources.doorbell_start_offset) /
253 kfd_doorbell_process_slice(kfd) + 1;
254
255 return num_of_elems;
256
257 }
258
kfd_get_process_doorbells(struct kfd_dev * dev,struct kfd_process * process)259 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
260 struct kfd_process *process)
261 {
262 return dev->doorbell_base +
263 process->doorbell_index * kfd_doorbell_process_slice(dev);
264 }
265
kfd_alloc_process_doorbells(struct kfd_process * process)266 int kfd_alloc_process_doorbells(struct kfd_process *process)
267 {
268 int r = ida_simple_get(&doorbell_ida, 1, max_doorbell_slices,
269 GFP_KERNEL);
270 if (r > 0)
271 process->doorbell_index = r;
272
273 return r;
274 }
275
kfd_free_process_doorbells(struct kfd_process * process)276 void kfd_free_process_doorbells(struct kfd_process *process)
277 {
278 if (process->doorbell_index)
279 ida_simple_remove(&doorbell_ida, process->doorbell_index);
280 }
281