1 /*
2 * Copyright © 2016-2018 Intel Corporation
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 (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * 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 AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include <errno.h>
25 #include <string.h>
26 #include <unistd.h>
27 #include <sys/types.h>
28 #include <sys/mman.h>
29
30 #include "aub_mem.h"
31 #include "util/anon_file.h"
32
33 struct bo_map {
34 struct list_head link;
35 struct intel_batch_decode_bo bo;
36 bool unmap_after_use;
37 bool ppgtt;
38 };
39
40 struct ggtt_entry {
41 struct rb_node node;
42 uint64_t virt_addr;
43 uint64_t phys_addr;
44 };
45
46 struct phys_mem {
47 struct rb_node node;
48 uint64_t fd_offset;
49 uint64_t phys_addr;
50 uint8_t *data;
51 const uint8_t *aub_data;
52 };
53
54 static void
add_gtt_bo_map(struct aub_mem * mem,struct intel_batch_decode_bo bo,bool ppgtt,bool unmap_after_use)55 add_gtt_bo_map(struct aub_mem *mem, struct intel_batch_decode_bo bo, bool ppgtt, bool unmap_after_use)
56 {
57 struct bo_map *m = calloc(1, sizeof(*m));
58
59 m->ppgtt = ppgtt;
60 m->bo = bo;
61 m->unmap_after_use = unmap_after_use;
62 list_add(&m->link, &mem->maps);
63 }
64
65 void
aub_mem_clear_bo_maps(struct aub_mem * mem)66 aub_mem_clear_bo_maps(struct aub_mem *mem)
67 {
68 list_for_each_entry_safe(struct bo_map, i, &mem->maps, link) {
69 if (i->unmap_after_use)
70 munmap((void *)i->bo.map, i->bo.size);
71 list_del(&i->link);
72 free(i);
73 }
74 }
75
76 static inline struct ggtt_entry *
ggtt_entry_next(struct ggtt_entry * entry)77 ggtt_entry_next(struct ggtt_entry *entry)
78 {
79 if (!entry)
80 return NULL;
81 struct rb_node *node = rb_node_next(&entry->node);
82 if (!node)
83 return NULL;
84 return rb_node_data(struct ggtt_entry, node, node);
85 }
86
87 static inline int
cmp_uint64(uint64_t a,uint64_t b)88 cmp_uint64(uint64_t a, uint64_t b)
89 {
90 if (a < b)
91 return 1;
92 if (a > b)
93 return -1;
94 return 0;
95 }
96
97 static inline int
cmp_ggtt_entry(const struct rb_node * node,const void * addr)98 cmp_ggtt_entry(const struct rb_node *node, const void *addr)
99 {
100 struct ggtt_entry *entry = rb_node_data(struct ggtt_entry, node, node);
101 return cmp_uint64(entry->virt_addr, *(const uint64_t *)addr);
102 }
103
104 static struct ggtt_entry *
ensure_ggtt_entry(struct aub_mem * mem,uint64_t virt_addr)105 ensure_ggtt_entry(struct aub_mem *mem, uint64_t virt_addr)
106 {
107 struct rb_node *node = rb_tree_search_sloppy(&mem->ggtt, &virt_addr,
108 cmp_ggtt_entry);
109 int cmp = 0;
110 if (!node || (cmp = cmp_ggtt_entry(node, &virt_addr))) {
111 struct ggtt_entry *new_entry = calloc(1, sizeof(*new_entry));
112 new_entry->virt_addr = virt_addr;
113 rb_tree_insert_at(&mem->ggtt, node, &new_entry->node, cmp < 0);
114 node = &new_entry->node;
115 }
116
117 return rb_node_data(struct ggtt_entry, node, node);
118 }
119
120 static struct ggtt_entry *
search_ggtt_entry(struct aub_mem * mem,uint64_t virt_addr)121 search_ggtt_entry(struct aub_mem *mem, uint64_t virt_addr)
122 {
123 virt_addr &= ~0xfff;
124
125 struct rb_node *node = rb_tree_search(&mem->ggtt, &virt_addr, cmp_ggtt_entry);
126
127 if (!node)
128 return NULL;
129
130 return rb_node_data(struct ggtt_entry, node, node);
131 }
132
133 static inline int
cmp_phys_mem(const struct rb_node * node,const void * addr)134 cmp_phys_mem(const struct rb_node *node, const void *addr)
135 {
136 struct phys_mem *mem = rb_node_data(struct phys_mem, node, node);
137 return cmp_uint64(mem->phys_addr, *(uint64_t *)addr);
138 }
139
140 static void
check_mmap_result(const void * res)141 check_mmap_result(const void *res)
142 {
143 if (res != MAP_FAILED)
144 return;
145
146 if (errno == ENOMEM) {
147 fprintf(stderr,
148 "Not enough memory available or maximum number of mappings reached. "
149 "Consider increasing sysctl vm.max_map_count.\n");
150 } else {
151 perror("mmap");
152 }
153
154 abort();
155 }
156
157 static struct phys_mem *
ensure_phys_mem(struct aub_mem * mem,uint64_t phys_addr)158 ensure_phys_mem(struct aub_mem *mem, uint64_t phys_addr)
159 {
160 struct rb_node *node = rb_tree_search_sloppy(&mem->mem, &phys_addr, cmp_phys_mem);
161 int cmp = 0;
162 if (!node || (cmp = cmp_phys_mem(node, &phys_addr))) {
163 struct phys_mem *new_mem = calloc(1, sizeof(*new_mem));
164 new_mem->phys_addr = phys_addr;
165 new_mem->fd_offset = mem->mem_fd_len;
166
167 ASSERTED int ftruncate_res = ftruncate(mem->mem_fd, mem->mem_fd_len += 4096);
168 assert(ftruncate_res == 0);
169
170 new_mem->data = mmap(NULL, 4096, PROT_READ | PROT_WRITE, MAP_SHARED,
171 mem->mem_fd, new_mem->fd_offset);
172 check_mmap_result(new_mem->data);
173
174 rb_tree_insert_at(&mem->mem, node, &new_mem->node, cmp < 0);
175 node = &new_mem->node;
176 }
177
178 return rb_node_data(struct phys_mem, node, node);
179 }
180
181 static struct phys_mem *
search_phys_mem(struct aub_mem * mem,uint64_t phys_addr)182 search_phys_mem(struct aub_mem *mem, uint64_t phys_addr)
183 {
184 phys_addr &= ~0xfff;
185
186 struct rb_node *node = rb_tree_search(&mem->mem, &phys_addr, cmp_phys_mem);
187
188 if (!node)
189 return NULL;
190
191 return rb_node_data(struct phys_mem, node, node);
192 }
193
194 void
aub_mem_local_write(void * _mem,uint64_t address,const void * data,uint32_t size)195 aub_mem_local_write(void *_mem, uint64_t address,
196 const void *data, uint32_t size)
197 {
198 struct aub_mem *mem = _mem;
199 struct intel_batch_decode_bo bo = {
200 .map = data,
201 .addr = address,
202 .size = size,
203 };
204 add_gtt_bo_map(mem, bo, false, false);
205 }
206
207 void
aub_mem_ggtt_entry_write(void * _mem,uint64_t address,const void * _data,uint32_t _size)208 aub_mem_ggtt_entry_write(void *_mem, uint64_t address,
209 const void *_data, uint32_t _size)
210 {
211 struct aub_mem *mem = _mem;
212 uint64_t virt_addr = (address / sizeof(uint64_t)) << 12;
213 const uint64_t *data = _data;
214 size_t size = _size / sizeof(*data);
215 for (const uint64_t *entry = data;
216 entry < data + size;
217 entry++, virt_addr += 4096) {
218 struct ggtt_entry *pt = ensure_ggtt_entry(mem, virt_addr);
219 pt->phys_addr = *entry;
220 }
221 }
222
223 void
aub_mem_phys_write(void * _mem,uint64_t phys_address,const void * data,uint32_t size)224 aub_mem_phys_write(void *_mem, uint64_t phys_address,
225 const void *data, uint32_t size)
226 {
227 struct aub_mem *mem = _mem;
228 uint32_t to_write = size;
229 for (uint64_t page = phys_address & ~0xfff; page < phys_address + size; page += 4096) {
230 struct phys_mem *pmem = ensure_phys_mem(mem, page);
231 uint64_t offset = MAX2(page, phys_address) - page;
232 uint32_t size_this_page = MIN2(to_write, 4096 - offset);
233 to_write -= size_this_page;
234 memcpy(pmem->data + offset, data, size_this_page);
235 pmem->aub_data = data - offset;
236 data = (const uint8_t *)data + size_this_page;
237 }
238 }
239
240 void
aub_mem_ggtt_write(void * _mem,uint64_t virt_address,const void * data,uint32_t size)241 aub_mem_ggtt_write(void *_mem, uint64_t virt_address,
242 const void *data, uint32_t size)
243 {
244 struct aub_mem *mem = _mem;
245 uint32_t to_write = size;
246 for (uint64_t page = virt_address & ~0xfff; page < virt_address + size; page += 4096) {
247 struct ggtt_entry *entry = search_ggtt_entry(mem, page);
248 assert(entry && entry->phys_addr & 0x1);
249
250 uint64_t offset = MAX2(page, virt_address) - page;
251 uint32_t size_this_page = MIN2(to_write, 4096 - offset);
252 to_write -= size_this_page;
253
254 uint64_t phys_page = entry->phys_addr & ~0xfff; /* Clear the validity bits. */
255 aub_mem_phys_write(mem, phys_page + offset, data, size_this_page);
256 data = (const uint8_t *)data + size_this_page;
257 }
258 }
259
260 struct intel_batch_decode_bo
aub_mem_get_ggtt_bo(void * _mem,uint64_t address)261 aub_mem_get_ggtt_bo(void *_mem, uint64_t address)
262 {
263 struct aub_mem *mem = _mem;
264 struct intel_batch_decode_bo bo = {0};
265
266 list_for_each_entry(struct bo_map, i, &mem->maps, link)
267 if (!i->ppgtt && i->bo.addr <= address && i->bo.addr + i->bo.size > address)
268 return i->bo;
269
270 address &= ~0xfff;
271
272 struct ggtt_entry *start =
273 (struct ggtt_entry *)rb_tree_search_sloppy(&mem->ggtt, &address,
274 cmp_ggtt_entry);
275 if (start && start->virt_addr < address)
276 start = ggtt_entry_next(start);
277 if (!start)
278 return bo;
279
280 struct ggtt_entry *last = start;
281 for (struct ggtt_entry *i = ggtt_entry_next(last);
282 i && last->virt_addr + 4096 == i->virt_addr;
283 last = i, i = ggtt_entry_next(last))
284 ;
285
286 bo.addr = MIN2(address, start->virt_addr);
287 bo.size = last->virt_addr - bo.addr + 4096;
288 bo.map = mmap(NULL, bo.size, PROT_READ, MAP_SHARED | MAP_ANONYMOUS, -1, 0);
289 check_mmap_result(bo.map);
290
291 for (struct ggtt_entry *i = start;
292 i;
293 i = i == last ? NULL : ggtt_entry_next(i)) {
294 uint64_t phys_addr = i->phys_addr & ~0xfff;
295 struct phys_mem *phys_mem = search_phys_mem(mem, phys_addr);
296
297 if (!phys_mem)
298 continue;
299
300 uint32_t map_offset = i->virt_addr - address;
301 void *res = mmap((uint8_t *)bo.map + map_offset, 4096, PROT_READ,
302 MAP_SHARED | MAP_FIXED, mem->mem_fd, phys_mem->fd_offset);
303 check_mmap_result(res);
304 }
305
306 add_gtt_bo_map(mem, bo, false, true);
307
308 return bo;
309 }
310
311 static struct phys_mem *
ppgtt_walk(struct aub_mem * mem,uint64_t pml4,uint64_t address)312 ppgtt_walk(struct aub_mem *mem, uint64_t pml4, uint64_t address)
313 {
314 uint64_t shift = 39;
315 uint64_t addr = pml4;
316 for (int level = 4; level > 0; level--) {
317 struct phys_mem *table = search_phys_mem(mem, addr);
318 if (!table)
319 return NULL;
320 int index = (address >> shift) & 0x1ff;
321 uint64_t entry = ((uint64_t *)table->data)[index];
322 if (!(entry & 1))
323 return NULL;
324 addr = entry & ~0xfff;
325 shift -= 9;
326 }
327 return search_phys_mem(mem, addr);
328 }
329
330 static bool
ppgtt_mapped(struct aub_mem * mem,uint64_t pml4,uint64_t address)331 ppgtt_mapped(struct aub_mem *mem, uint64_t pml4, uint64_t address)
332 {
333 return ppgtt_walk(mem, pml4, address) != NULL;
334 }
335
336 struct intel_batch_decode_bo
aub_mem_get_ppgtt_bo(void * _mem,uint64_t address)337 aub_mem_get_ppgtt_bo(void *_mem, uint64_t address)
338 {
339 struct aub_mem *mem = _mem;
340 struct intel_batch_decode_bo bo = {0};
341
342 list_for_each_entry(struct bo_map, i, &mem->maps, link)
343 if (i->ppgtt && i->bo.addr <= address && i->bo.addr + i->bo.size > address)
344 return i->bo;
345
346 address &= ~0xfff;
347
348 if (!ppgtt_mapped(mem, mem->pml4, address))
349 return bo;
350
351 /* Map everything until the first gap since we don't know how much the
352 * decoder actually needs.
353 */
354 uint64_t end = address;
355 while (ppgtt_mapped(mem, mem->pml4, end))
356 end += 4096;
357
358 bo.addr = address;
359 bo.size = end - address;
360 bo.map = mmap(NULL, bo.size, PROT_READ, MAP_SHARED | MAP_ANONYMOUS, -1, 0);
361 assert(bo.map != MAP_FAILED);
362
363 for (uint64_t page = address; page < end; page += 4096) {
364 struct phys_mem *phys_mem = ppgtt_walk(mem, mem->pml4, page);
365
366 void *res = mmap((uint8_t *)bo.map + (page - bo.addr), 4096, PROT_READ,
367 MAP_SHARED | MAP_FIXED, mem->mem_fd, phys_mem->fd_offset);
368 check_mmap_result(res);
369 }
370
371 add_gtt_bo_map(mem, bo, true, true);
372
373 return bo;
374 }
375
376 bool
aub_mem_init(struct aub_mem * mem)377 aub_mem_init(struct aub_mem *mem)
378 {
379 memset(mem, 0, sizeof(*mem));
380
381 list_inithead(&mem->maps);
382
383 mem->mem_fd = os_create_anonymous_file(0, "phys memory");
384
385 return mem->mem_fd != -1;
386 }
387
388 void
aub_mem_fini(struct aub_mem * mem)389 aub_mem_fini(struct aub_mem *mem)
390 {
391 if (mem->mem_fd == -1)
392 return;
393
394 aub_mem_clear_bo_maps(mem);
395
396
397 rb_tree_foreach_safe(struct ggtt_entry, entry, &mem->ggtt, node) {
398 rb_tree_remove(&mem->ggtt, &entry->node);
399 free(entry);
400 }
401 rb_tree_foreach_safe(struct phys_mem, entry, &mem->mem, node) {
402 rb_tree_remove(&mem->mem, &entry->node);
403 free(entry);
404 }
405
406 close(mem->mem_fd);
407 mem->mem_fd = -1;
408 }
409
410 struct intel_batch_decode_bo
aub_mem_get_phys_addr_data(struct aub_mem * mem,uint64_t phys_addr)411 aub_mem_get_phys_addr_data(struct aub_mem *mem, uint64_t phys_addr)
412 {
413 struct phys_mem *page = search_phys_mem(mem, phys_addr);
414 return page ?
415 (struct intel_batch_decode_bo) { .map = page->data, .addr = page->phys_addr, .size = 4096 } :
416 (struct intel_batch_decode_bo) {};
417 }
418
419 struct intel_batch_decode_bo
aub_mem_get_ppgtt_addr_data(struct aub_mem * mem,uint64_t virt_addr)420 aub_mem_get_ppgtt_addr_data(struct aub_mem *mem, uint64_t virt_addr)
421 {
422 struct phys_mem *page = ppgtt_walk(mem, mem->pml4, virt_addr);
423 return page ?
424 (struct intel_batch_decode_bo) { .map = page->data, .addr = virt_addr & ~((1ULL << 12) - 1), .size = 4096 } :
425 (struct intel_batch_decode_bo) {};
426 }
427
428 struct intel_batch_decode_bo
aub_mem_get_ppgtt_addr_aub_data(struct aub_mem * mem,uint64_t virt_addr)429 aub_mem_get_ppgtt_addr_aub_data(struct aub_mem *mem, uint64_t virt_addr)
430 {
431 struct phys_mem *page = ppgtt_walk(mem, mem->pml4, virt_addr);
432 return page ?
433 (struct intel_batch_decode_bo) { .map = page->aub_data, .addr = virt_addr & ~((1ULL << 12) - 1), .size = 4096 } :
434 (struct intel_batch_decode_bo) {};
435 }
436