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 /*
35 * Each device exposes a doorbell aperture, a PCI MMIO aperture that
36 * receives 32-bit writes that are passed to queues as wptr values.
37 * The doorbells are intended to be written by applications as part
38 * of queueing work on user-mode queues.
39 * We assign doorbells to applications in PAGE_SIZE-sized and aligned chunks.
40 * We map the doorbell address space into user-mode when a process creates
41 * its first queue on each device.
42 * Although the mapping is done by KFD, it is equivalent to an mmap of
43 * the /dev/kfd with the particular device encoded in the mmap offset.
44 * There will be other uses for mmap of /dev/kfd, so only a range of
45 * offsets (KFD_MMAP_DOORBELL_START-END) is used for doorbells.
46 */
47
48 /* # of doorbell bytes allocated for each process. */
kfd_doorbell_process_slice(struct kfd_dev * kfd)49 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd)
50 {
51 return roundup(kfd->device_info->doorbell_size *
52 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
53 PAGE_SIZE);
54 }
55
56 /* Doorbell calculations for device init. */
kfd_doorbell_init(struct kfd_dev * kfd)57 int kfd_doorbell_init(struct kfd_dev *kfd)
58 {
59 size_t doorbell_start_offset;
60 size_t doorbell_aperture_size;
61 size_t doorbell_process_limit;
62
63 /*
64 * We start with calculations in bytes because the input data might
65 * only be byte-aligned.
66 * Only after we have done the rounding can we assume any alignment.
67 */
68
69 doorbell_start_offset =
70 roundup(kfd->shared_resources.doorbell_start_offset,
71 kfd_doorbell_process_slice(kfd));
72
73 doorbell_aperture_size =
74 rounddown(kfd->shared_resources.doorbell_aperture_size,
75 kfd_doorbell_process_slice(kfd));
76
77 if (doorbell_aperture_size > doorbell_start_offset)
78 doorbell_process_limit =
79 (doorbell_aperture_size - doorbell_start_offset) /
80 kfd_doorbell_process_slice(kfd);
81 else
82 return -ENOSPC;
83
84 if (!kfd->max_doorbell_slices ||
85 doorbell_process_limit < kfd->max_doorbell_slices)
86 kfd->max_doorbell_slices = doorbell_process_limit;
87
88 kfd->doorbell_base = kfd->shared_resources.doorbell_physical_address +
89 doorbell_start_offset;
90
91 kfd->doorbell_base_dw_offset = doorbell_start_offset / sizeof(u32);
92
93 kfd->doorbell_kernel_ptr = ioremap(kfd->doorbell_base,
94 kfd_doorbell_process_slice(kfd));
95
96 if (!kfd->doorbell_kernel_ptr)
97 return -ENOMEM;
98
99 pr_debug("Doorbell initialization:\n");
100 pr_debug("doorbell base == 0x%08lX\n",
101 (uintptr_t)kfd->doorbell_base);
102
103 pr_debug("doorbell_base_dw_offset == 0x%08lX\n",
104 kfd->doorbell_base_dw_offset);
105
106 pr_debug("doorbell_process_limit == 0x%08lX\n",
107 doorbell_process_limit);
108
109 pr_debug("doorbell_kernel_offset == 0x%08lX\n",
110 (uintptr_t)kfd->doorbell_base);
111
112 pr_debug("doorbell aperture size == 0x%08lX\n",
113 kfd->shared_resources.doorbell_aperture_size);
114
115 pr_debug("doorbell kernel address == %p\n", kfd->doorbell_kernel_ptr);
116
117 return 0;
118 }
119
kfd_doorbell_fini(struct kfd_dev * kfd)120 void kfd_doorbell_fini(struct kfd_dev *kfd)
121 {
122 if (kfd->doorbell_kernel_ptr)
123 iounmap(kfd->doorbell_kernel_ptr);
124 }
125
kfd_doorbell_mmap(struct kfd_dev * dev,struct kfd_process * process,struct vm_area_struct * vma)126 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process,
127 struct vm_area_struct *vma)
128 {
129 phys_addr_t address;
130 struct kfd_process_device *pdd;
131
132 /*
133 * For simplicitly we only allow mapping of the entire doorbell
134 * allocation of a single device & process.
135 */
136 if (vma->vm_end - vma->vm_start != kfd_doorbell_process_slice(dev))
137 return -EINVAL;
138
139 pdd = kfd_get_process_device_data(dev, process);
140 if (!pdd)
141 return -EINVAL;
142
143 /* Calculate physical address of doorbell */
144 address = kfd_get_process_doorbells(pdd);
145 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE |
146 VM_DONTDUMP | VM_PFNMAP;
147
148 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
149
150 pr_debug("Mapping doorbell page\n"
151 " target user address == 0x%08llX\n"
152 " physical address == 0x%08llX\n"
153 " vm_flags == 0x%04lX\n"
154 " size == 0x%04lX\n",
155 (unsigned long long) vma->vm_start, address, vma->vm_flags,
156 kfd_doorbell_process_slice(dev));
157
158
159 return io_remap_pfn_range(vma,
160 vma->vm_start,
161 address >> PAGE_SHIFT,
162 kfd_doorbell_process_slice(dev),
163 vma->vm_page_prot);
164 }
165
166
167 /* get kernel iomem pointer for a doorbell */
kfd_get_kernel_doorbell(struct kfd_dev * kfd,unsigned int * doorbell_off)168 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
169 unsigned int *doorbell_off)
170 {
171 u32 inx;
172
173 mutex_lock(&kfd->doorbell_mutex);
174 inx = find_first_zero_bit(kfd->doorbell_available_index,
175 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
176
177 __set_bit(inx, kfd->doorbell_available_index);
178 mutex_unlock(&kfd->doorbell_mutex);
179
180 if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
181 return NULL;
182
183 inx *= kfd->device_info->doorbell_size / sizeof(u32);
184
185 /*
186 * Calculating the kernel doorbell offset using the first
187 * doorbell page.
188 */
189 *doorbell_off = kfd->doorbell_base_dw_offset + inx;
190
191 pr_debug("Get kernel queue doorbell\n"
192 " doorbell offset == 0x%08X\n"
193 " doorbell index == 0x%x\n",
194 *doorbell_off, inx);
195
196 return kfd->doorbell_kernel_ptr + inx;
197 }
198
kfd_release_kernel_doorbell(struct kfd_dev * kfd,u32 __iomem * db_addr)199 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr)
200 {
201 unsigned int inx;
202
203 inx = (unsigned int)(db_addr - kfd->doorbell_kernel_ptr)
204 * sizeof(u32) / kfd->device_info->doorbell_size;
205
206 mutex_lock(&kfd->doorbell_mutex);
207 __clear_bit(inx, kfd->doorbell_available_index);
208 mutex_unlock(&kfd->doorbell_mutex);
209 }
210
write_kernel_doorbell(void __iomem * db,u32 value)211 void write_kernel_doorbell(void __iomem *db, u32 value)
212 {
213 if (db) {
214 writel(value, db);
215 pr_debug("Writing %d to doorbell address %p\n", value, db);
216 }
217 }
218
write_kernel_doorbell64(void __iomem * db,u64 value)219 void write_kernel_doorbell64(void __iomem *db, u64 value)
220 {
221 if (db) {
222 WARN(((unsigned long)db & 7) != 0,
223 "Unaligned 64-bit doorbell");
224 writeq(value, (u64 __iomem *)db);
225 pr_debug("writing %llu to doorbell address %p\n", value, db);
226 }
227 }
228
kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev * kfd,struct kfd_process_device * pdd,unsigned int doorbell_id)229 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd,
230 struct kfd_process_device *pdd,
231 unsigned int doorbell_id)
232 {
233 /*
234 * doorbell_base_dw_offset accounts for doorbells taken by KGD.
235 * index * kfd_doorbell_process_slice/sizeof(u32) adjusts to
236 * the process's doorbells. The offset returned is in dword
237 * units regardless of the ASIC-dependent doorbell size.
238 */
239 return kfd->doorbell_base_dw_offset +
240 pdd->doorbell_index
241 * kfd_doorbell_process_slice(kfd) / sizeof(u32) +
242 doorbell_id * kfd->device_info->doorbell_size / sizeof(u32);
243 }
244
kfd_get_number_elems(struct kfd_dev * kfd)245 uint64_t kfd_get_number_elems(struct kfd_dev *kfd)
246 {
247 uint64_t num_of_elems = (kfd->shared_resources.doorbell_aperture_size -
248 kfd->shared_resources.doorbell_start_offset) /
249 kfd_doorbell_process_slice(kfd) + 1;
250
251 return num_of_elems;
252
253 }
254
kfd_get_process_doorbells(struct kfd_process_device * pdd)255 phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd)
256 {
257 return pdd->dev->doorbell_base +
258 pdd->doorbell_index * kfd_doorbell_process_slice(pdd->dev);
259 }
260
kfd_alloc_process_doorbells(struct kfd_dev * kfd,unsigned int * doorbell_index)261 int kfd_alloc_process_doorbells(struct kfd_dev *kfd, unsigned int *doorbell_index)
262 {
263 int r = ida_simple_get(&kfd->doorbell_ida, 1, kfd->max_doorbell_slices,
264 GFP_KERNEL);
265 if (r > 0)
266 *doorbell_index = r;
267
268 return r;
269 }
270
kfd_free_process_doorbells(struct kfd_dev * kfd,unsigned int doorbell_index)271 void kfd_free_process_doorbells(struct kfd_dev *kfd, unsigned int doorbell_index)
272 {
273 if (doorbell_index)
274 ida_simple_remove(&kfd->doorbell_ida, doorbell_index);
275 }
276