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. */
kfd_doorbell_process_slice(struct kfd_dev * kfd)50 size_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. */
kfd_doorbell_init(struct kfd_dev * kfd)62 int 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 
kfd_doorbell_fini(struct kfd_dev * kfd)135 void kfd_doorbell_fini(struct kfd_dev *kfd)
136 {
137 	if (kfd->doorbell_kernel_ptr)
138 		iounmap(kfd->doorbell_kernel_ptr);
139 }
140 
kfd_doorbell_mmap(struct kfd_dev * dev,struct kfd_process * process,struct vm_area_struct * vma)141 int 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 */
kfd_get_kernel_doorbell(struct kfd_dev * kfd,unsigned int * doorbell_off)185 void __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 
kfd_release_kernel_doorbell(struct kfd_dev * kfd,u32 __iomem * db_addr)216 void 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 
write_kernel_doorbell(void __iomem * db,u32 value)228 void 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 
write_kernel_doorbell64(void __iomem * db,u64 value)236 void 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 
kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev * kfd,struct kfd_process_device * pdd,unsigned int doorbell_id)246 unsigned 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 
kfd_get_number_elems(struct kfd_dev * kfd)268 uint64_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 
kfd_get_process_doorbells(struct kfd_process_device * pdd)278 phys_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 
kfd_alloc_process_doorbells(struct kfd_dev * kfd,unsigned int * doorbell_index)291 int 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 
kfd_free_process_doorbells(struct kfd_dev * kfd,unsigned int doorbell_index)312 void 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 }
323