Rev 5271 | Go to most recent revision | Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
5078 | serge | 1 | /* |
2 | * Copyright 2008 Jerome Glisse. |
||
3 | * All Rights Reserved. |
||
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 (including the next |
||
13 | * paragraph) shall be included in all copies or substantial portions of the |
||
14 | * 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 | * PRECISION INSIGHT AND/OR ITS SUPPLIERS 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 OTHER |
||
22 | * DEALINGS IN THE SOFTWARE. |
||
23 | * |
||
24 | * Authors: |
||
25 | * Jerome Glisse |
||
26 | */ |
||
27 | #include |
||
28 | #include |
||
29 | #include |
||
30 | #include "radeon_reg.h" |
||
31 | #include "radeon.h" |
||
32 | #include "radeon_trace.h" |
||
33 | |||
34 | #define RADEON_CS_MAX_PRIORITY 32u |
||
35 | #define RADEON_CS_NUM_BUCKETS (RADEON_CS_MAX_PRIORITY + 1) |
||
36 | |||
37 | static inline unsigned long |
||
38 | copy_from_user(void *to, const void __user *from, unsigned long n) |
||
39 | { |
||
40 | memcpy(to, from, n); |
||
41 | return n; |
||
42 | } |
||
43 | |||
44 | /* This is based on the bucket sort with O(n) time complexity. |
||
45 | * An item with priority "i" is added to bucket[i]. The lists are then |
||
46 | * concatenated in descending order. |
||
47 | */ |
||
48 | struct radeon_cs_buckets { |
||
49 | struct list_head bucket[RADEON_CS_NUM_BUCKETS]; |
||
50 | }; |
||
51 | |||
52 | static void radeon_cs_buckets_init(struct radeon_cs_buckets *b) |
||
53 | { |
||
54 | unsigned i; |
||
55 | |||
56 | for (i = 0; i < RADEON_CS_NUM_BUCKETS; i++) |
||
57 | INIT_LIST_HEAD(&b->bucket[i]); |
||
58 | } |
||
59 | |||
60 | static void radeon_cs_buckets_add(struct radeon_cs_buckets *b, |
||
61 | struct list_head *item, unsigned priority) |
||
62 | { |
||
63 | /* Since buffers which appear sooner in the relocation list are |
||
64 | * likely to be used more often than buffers which appear later |
||
65 | * in the list, the sort mustn't change the ordering of buffers |
||
66 | * with the same priority, i.e. it must be stable. |
||
67 | */ |
||
68 | list_add_tail(item, &b->bucket[min(priority, RADEON_CS_MAX_PRIORITY)]); |
||
69 | } |
||
70 | |||
71 | static void radeon_cs_buckets_get_list(struct radeon_cs_buckets *b, |
||
72 | struct list_head *out_list) |
||
73 | { |
||
74 | unsigned i; |
||
75 | |||
76 | /* Connect the sorted buckets in the output list. */ |
||
77 | for (i = 0; i < RADEON_CS_NUM_BUCKETS; i++) { |
||
78 | list_splice(&b->bucket[i], out_list); |
||
79 | } |
||
80 | } |
||
81 | |||
82 | static int radeon_cs_parser_relocs(struct radeon_cs_parser *p) |
||
83 | { |
||
84 | struct drm_device *ddev = p->rdev->ddev; |
||
85 | struct radeon_cs_chunk *chunk; |
||
86 | struct radeon_cs_buckets buckets; |
||
87 | unsigned i, j; |
||
88 | bool duplicate; |
||
89 | |||
90 | if (p->chunk_relocs_idx == -1) { |
||
91 | return 0; |
||
92 | } |
||
93 | chunk = &p->chunks[p->chunk_relocs_idx]; |
||
94 | p->dma_reloc_idx = 0; |
||
95 | /* FIXME: we assume that each relocs use 4 dwords */ |
||
96 | p->nrelocs = chunk->length_dw / 4; |
||
97 | p->relocs_ptr = kcalloc(p->nrelocs, sizeof(void *), GFP_KERNEL); |
||
98 | if (p->relocs_ptr == NULL) { |
||
99 | return -ENOMEM; |
||
100 | } |
||
101 | p->relocs = kcalloc(p->nrelocs, sizeof(struct radeon_cs_reloc), GFP_KERNEL); |
||
102 | if (p->relocs == NULL) { |
||
103 | return -ENOMEM; |
||
104 | } |
||
105 | |||
106 | radeon_cs_buckets_init(&buckets); |
||
107 | |||
108 | for (i = 0; i < p->nrelocs; i++) { |
||
109 | struct drm_radeon_cs_reloc *r; |
||
110 | unsigned priority; |
||
111 | |||
112 | duplicate = false; |
||
113 | r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4]; |
||
114 | for (j = 0; j < i; j++) { |
||
115 | if (r->handle == p->relocs[j].handle) { |
||
116 | p->relocs_ptr[i] = &p->relocs[j]; |
||
117 | duplicate = true; |
||
118 | break; |
||
119 | } |
||
120 | } |
||
121 | if (duplicate) { |
||
122 | p->relocs[i].handle = 0; |
||
123 | continue; |
||
124 | } |
||
125 | |||
126 | p->relocs[i].gobj = drm_gem_object_lookup(ddev, p->filp, |
||
127 | r->handle); |
||
128 | if (p->relocs[i].gobj == NULL) { |
||
129 | DRM_ERROR("gem object lookup failed 0x%x\n", |
||
130 | r->handle); |
||
131 | return -ENOENT; |
||
132 | } |
||
133 | p->relocs_ptr[i] = &p->relocs[i]; |
||
134 | p->relocs[i].robj = gem_to_radeon_bo(p->relocs[i].gobj); |
||
135 | |||
136 | /* The userspace buffer priorities are from 0 to 15. A higher |
||
137 | * number means the buffer is more important. |
||
138 | * Also, the buffers used for write have a higher priority than |
||
139 | * the buffers used for read only, which doubles the range |
||
140 | * to 0 to 31. 32 is reserved for the kernel driver. |
||
141 | */ |
||
142 | priority = (r->flags & RADEON_RELOC_PRIO_MASK) * 2 |
||
143 | + !!r->write_domain; |
||
144 | |||
145 | /* the first reloc of an UVD job is the msg and that must be in |
||
146 | VRAM, also but everything into VRAM on AGP cards to avoid |
||
147 | image corruptions */ |
||
148 | if (p->ring == R600_RING_TYPE_UVD_INDEX && |
||
149 | (i == 0 || drm_pci_device_is_agp(p->rdev->ddev))) { |
||
150 | /* TODO: is this still needed for NI+ ? */ |
||
151 | p->relocs[i].prefered_domains = |
||
152 | RADEON_GEM_DOMAIN_VRAM; |
||
153 | |||
154 | p->relocs[i].allowed_domains = |
||
155 | RADEON_GEM_DOMAIN_VRAM; |
||
156 | |||
157 | /* prioritize this over any other relocation */ |
||
158 | priority = RADEON_CS_MAX_PRIORITY; |
||
159 | } else { |
||
160 | uint32_t domain = r->write_domain ? |
||
161 | r->write_domain : r->read_domains; |
||
162 | |||
163 | if (domain & RADEON_GEM_DOMAIN_CPU) { |
||
164 | DRM_ERROR("RADEON_GEM_DOMAIN_CPU is not valid " |
||
165 | "for command submission\n"); |
||
166 | return -EINVAL; |
||
167 | } |
||
168 | |||
169 | p->relocs[i].prefered_domains = domain; |
||
170 | if (domain == RADEON_GEM_DOMAIN_VRAM) |
||
171 | domain |= RADEON_GEM_DOMAIN_GTT; |
||
172 | p->relocs[i].allowed_domains = domain; |
||
173 | } |
||
174 | |||
175 | p->relocs[i].tv.bo = &p->relocs[i].robj->tbo; |
||
176 | p->relocs[i].handle = r->handle; |
||
177 | |||
178 | radeon_cs_buckets_add(&buckets, &p->relocs[i].tv.head, |
||
179 | priority); |
||
180 | } |
||
181 | |||
182 | radeon_cs_buckets_get_list(&buckets, &p->validated); |
||
183 | |||
184 | if (p->cs_flags & RADEON_CS_USE_VM) |
||
185 | p->vm_bos = radeon_vm_get_bos(p->rdev, p->ib.vm, |
||
186 | &p->validated); |
||
187 | |||
188 | return radeon_bo_list_validate(p->rdev, &p->ticket, &p->validated, p->ring); |
||
189 | } |
||
190 | |||
191 | static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority) |
||
192 | { |
||
193 | p->priority = priority; |
||
194 | |||
195 | switch (ring) { |
||
196 | default: |
||
197 | DRM_ERROR("unknown ring id: %d\n", ring); |
||
198 | return -EINVAL; |
||
199 | case RADEON_CS_RING_GFX: |
||
200 | p->ring = RADEON_RING_TYPE_GFX_INDEX; |
||
201 | break; |
||
202 | case RADEON_CS_RING_COMPUTE: |
||
203 | if (p->rdev->family >= CHIP_TAHITI) { |
||
204 | if (p->priority > 0) |
||
205 | p->ring = CAYMAN_RING_TYPE_CP1_INDEX; |
||
206 | else |
||
207 | p->ring = CAYMAN_RING_TYPE_CP2_INDEX; |
||
208 | } else |
||
209 | p->ring = RADEON_RING_TYPE_GFX_INDEX; |
||
210 | break; |
||
211 | case RADEON_CS_RING_DMA: |
||
212 | if (p->rdev->family >= CHIP_CAYMAN) { |
||
213 | if (p->priority > 0) |
||
214 | p->ring = R600_RING_TYPE_DMA_INDEX; |
||
215 | else |
||
216 | p->ring = CAYMAN_RING_TYPE_DMA1_INDEX; |
||
217 | } else if (p->rdev->family >= CHIP_RV770) { |
||
218 | p->ring = R600_RING_TYPE_DMA_INDEX; |
||
219 | } else { |
||
220 | return -EINVAL; |
||
221 | } |
||
222 | break; |
||
223 | case RADEON_CS_RING_UVD: |
||
224 | p->ring = R600_RING_TYPE_UVD_INDEX; |
||
225 | break; |
||
226 | case RADEON_CS_RING_VCE: |
||
227 | /* TODO: only use the low priority ring for now */ |
||
228 | p->ring = TN_RING_TYPE_VCE1_INDEX; |
||
229 | break; |
||
230 | } |
||
231 | return 0; |
||
232 | } |
||
233 | |||
234 | static void radeon_cs_sync_rings(struct radeon_cs_parser *p) |
||
235 | { |
||
236 | int i; |
||
237 | |||
238 | for (i = 0; i < p->nrelocs; i++) { |
||
239 | if (!p->relocs[i].robj) |
||
240 | continue; |
||
241 | |||
242 | radeon_semaphore_sync_to(p->ib.semaphore, |
||
243 | p->relocs[i].robj->tbo.sync_obj); |
||
244 | } |
||
245 | } |
||
246 | |||
247 | /* XXX: note that this is called from the legacy UMS CS ioctl as well */ |
||
248 | int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data) |
||
249 | { |
||
250 | struct drm_radeon_cs *cs = data; |
||
251 | uint64_t *chunk_array_ptr; |
||
252 | unsigned size, i; |
||
253 | u32 ring = RADEON_CS_RING_GFX; |
||
254 | s32 priority = 0; |
||
255 | |||
256 | if (!cs->num_chunks) { |
||
257 | return 0; |
||
258 | } |
||
259 | /* get chunks */ |
||
260 | INIT_LIST_HEAD(&p->validated); |
||
261 | p->idx = 0; |
||
262 | p->ib.sa_bo = NULL; |
||
263 | p->ib.semaphore = NULL; |
||
264 | p->const_ib.sa_bo = NULL; |
||
265 | p->const_ib.semaphore = NULL; |
||
266 | p->chunk_ib_idx = -1; |
||
267 | p->chunk_relocs_idx = -1; |
||
268 | p->chunk_flags_idx = -1; |
||
269 | p->chunk_const_ib_idx = -1; |
||
270 | p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL); |
||
271 | if (p->chunks_array == NULL) { |
||
272 | return -ENOMEM; |
||
273 | } |
||
274 | chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks); |
||
275 | if (copy_from_user(p->chunks_array, chunk_array_ptr, |
||
276 | sizeof(uint64_t)*cs->num_chunks)) { |
||
277 | return -EFAULT; |
||
278 | } |
||
279 | p->cs_flags = 0; |
||
280 | p->nchunks = cs->num_chunks; |
||
281 | p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL); |
||
282 | if (p->chunks == NULL) { |
||
283 | return -ENOMEM; |
||
284 | } |
||
285 | for (i = 0; i < p->nchunks; i++) { |
||
286 | struct drm_radeon_cs_chunk __user **chunk_ptr = NULL; |
||
287 | struct drm_radeon_cs_chunk user_chunk; |
||
288 | uint32_t __user *cdata; |
||
289 | |||
290 | chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i]; |
||
291 | if (copy_from_user(&user_chunk, chunk_ptr, |
||
292 | sizeof(struct drm_radeon_cs_chunk))) { |
||
293 | return -EFAULT; |
||
294 | } |
||
295 | p->chunks[i].length_dw = user_chunk.length_dw; |
||
296 | p->chunks[i].chunk_id = user_chunk.chunk_id; |
||
297 | if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) { |
||
298 | p->chunk_relocs_idx = i; |
||
299 | } |
||
300 | if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) { |
||
301 | p->chunk_ib_idx = i; |
||
302 | /* zero length IB isn't useful */ |
||
303 | if (p->chunks[i].length_dw == 0) |
||
304 | return -EINVAL; |
||
305 | } |
||
306 | if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) { |
||
307 | p->chunk_const_ib_idx = i; |
||
308 | /* zero length CONST IB isn't useful */ |
||
309 | if (p->chunks[i].length_dw == 0) |
||
310 | return -EINVAL; |
||
311 | } |
||
312 | if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) { |
||
313 | p->chunk_flags_idx = i; |
||
314 | /* zero length flags aren't useful */ |
||
315 | if (p->chunks[i].length_dw == 0) |
||
316 | return -EINVAL; |
||
317 | } |
||
318 | |||
319 | size = p->chunks[i].length_dw; |
||
320 | cdata = (void __user *)(unsigned long)user_chunk.chunk_data; |
||
321 | p->chunks[i].user_ptr = cdata; |
||
322 | if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) |
||
323 | continue; |
||
324 | |||
325 | if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) { |
||
326 | if (!p->rdev || !(p->rdev->flags & RADEON_IS_AGP)) |
||
327 | continue; |
||
328 | } |
||
329 | |||
330 | p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t)); |
||
331 | size *= sizeof(uint32_t); |
||
332 | if (p->chunks[i].kdata == NULL) { |
||
333 | return -ENOMEM; |
||
334 | } |
||
335 | if (copy_from_user(p->chunks[i].kdata, cdata, size)) { |
||
336 | return -EFAULT; |
||
337 | } |
||
338 | if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) { |
||
339 | p->cs_flags = p->chunks[i].kdata[0]; |
||
340 | if (p->chunks[i].length_dw > 1) |
||
341 | ring = p->chunks[i].kdata[1]; |
||
342 | if (p->chunks[i].length_dw > 2) |
||
343 | priority = (s32)p->chunks[i].kdata[2]; |
||
344 | } |
||
345 | } |
||
346 | |||
347 | /* these are KMS only */ |
||
348 | if (p->rdev) { |
||
349 | if ((p->cs_flags & RADEON_CS_USE_VM) && |
||
350 | !p->rdev->vm_manager.enabled) { |
||
351 | DRM_ERROR("VM not active on asic!\n"); |
||
352 | return -EINVAL; |
||
353 | } |
||
354 | |||
355 | if (radeon_cs_get_ring(p, ring, priority)) |
||
356 | return -EINVAL; |
||
357 | |||
358 | /* we only support VM on some SI+ rings */ |
||
359 | if ((p->cs_flags & RADEON_CS_USE_VM) == 0) { |
||
360 | if (p->rdev->asic->ring[p->ring]->cs_parse == NULL) { |
||
361 | DRM_ERROR("Ring %d requires VM!\n", p->ring); |
||
362 | return -EINVAL; |
||
363 | } |
||
364 | } else { |
||
365 | if (p->rdev->asic->ring[p->ring]->ib_parse == NULL) { |
||
366 | DRM_ERROR("VM not supported on ring %d!\n", |
||
367 | p->ring); |
||
368 | return -EINVAL; |
||
369 | } |
||
370 | } |
||
371 | } |
||
372 | |||
373 | return 0; |
||
374 | } |
||
375 | |||
376 | static int cmp_size_smaller_first(void *priv, struct list_head *a, |
||
377 | struct list_head *b) |
||
378 | { |
||
379 | struct radeon_cs_reloc *la = list_entry(a, struct radeon_cs_reloc, tv.head); |
||
380 | struct radeon_cs_reloc *lb = list_entry(b, struct radeon_cs_reloc, tv.head); |
||
381 | |||
382 | /* Sort A before B if A is smaller. */ |
||
383 | return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages; |
||
384 | } |
||
385 | |||
386 | /** |
||
387 | * cs_parser_fini() - clean parser states |
||
388 | * @parser: parser structure holding parsing context. |
||
389 | * @error: error number |
||
390 | * |
||
391 | * If error is set than unvalidate buffer, otherwise just free memory |
||
392 | * used by parsing context. |
||
393 | **/ |
||
394 | static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error, bool backoff) |
||
395 | { |
||
396 | unsigned i; |
||
397 | |||
398 | if (!error) { |
||
399 | /* Sort the buffer list from the smallest to largest buffer, |
||
400 | * which affects the order of buffers in the LRU list. |
||
401 | * This assures that the smallest buffers are added first |
||
402 | * to the LRU list, so they are likely to be later evicted |
||
403 | * first, instead of large buffers whose eviction is more |
||
404 | * expensive. |
||
405 | * |
||
406 | * This slightly lowers the number of bytes moved by TTM |
||
407 | * per frame under memory pressure. |
||
408 | */ |
||
409 | list_sort(NULL, &parser->validated, cmp_size_smaller_first); |
||
410 | |||
411 | ttm_eu_fence_buffer_objects(&parser->ticket, |
||
412 | &parser->validated, |
||
413 | parser->ib.fence); |
||
414 | } else if (backoff) { |
||
415 | ttm_eu_backoff_reservation(&parser->ticket, |
||
416 | &parser->validated); |
||
417 | } |
||
418 | |||
419 | if (parser->relocs != NULL) { |
||
420 | for (i = 0; i < parser->nrelocs; i++) { |
||
421 | if (parser->relocs[i].gobj) |
||
422 | drm_gem_object_unreference_unlocked(parser->relocs[i].gobj); |
||
423 | } |
||
424 | } |
||
425 | kfree(parser->track); |
||
426 | kfree(parser->relocs); |
||
427 | kfree(parser->relocs_ptr); |
||
428 | kfree(parser->vm_bos); |
||
429 | for (i = 0; i < parser->nchunks; i++) |
||
430 | drm_free_large(parser->chunks[i].kdata); |
||
431 | kfree(parser->chunks); |
||
432 | kfree(parser->chunks_array); |
||
433 | radeon_ib_free(parser->rdev, &parser->ib); |
||
434 | radeon_ib_free(parser->rdev, &parser->const_ib); |
||
435 | } |
||
436 | |||
437 | static int radeon_cs_ib_chunk(struct radeon_device *rdev, |
||
438 | struct radeon_cs_parser *parser) |
||
439 | { |
||
440 | int r; |
||
441 | |||
442 | if (parser->chunk_ib_idx == -1) |
||
443 | return 0; |
||
444 | |||
445 | if (parser->cs_flags & RADEON_CS_USE_VM) |
||
446 | return 0; |
||
447 | |||
448 | r = radeon_cs_parse(rdev, parser->ring, parser); |
||
449 | if (r || parser->parser_error) { |
||
450 | DRM_ERROR("Invalid command stream !\n"); |
||
451 | return r; |
||
452 | } |
||
453 | |||
454 | if (parser->ring == R600_RING_TYPE_UVD_INDEX) |
||
455 | radeon_uvd_note_usage(rdev); |
||
456 | else if ((parser->ring == TN_RING_TYPE_VCE1_INDEX) || |
||
457 | (parser->ring == TN_RING_TYPE_VCE2_INDEX)) |
||
458 | radeon_vce_note_usage(rdev); |
||
459 | |||
460 | radeon_cs_sync_rings(parser); |
||
461 | r = radeon_ib_schedule(rdev, &parser->ib, NULL, true); |
||
462 | if (r) { |
||
463 | DRM_ERROR("Failed to schedule IB !\n"); |
||
464 | } |
||
465 | return r; |
||
466 | } |
||
467 | |||
468 | static int radeon_bo_vm_update_pte(struct radeon_cs_parser *p, |
||
469 | struct radeon_vm *vm) |
||
470 | { |
||
471 | struct radeon_device *rdev = p->rdev; |
||
472 | struct radeon_bo_va *bo_va; |
||
473 | int i, r; |
||
474 | |||
475 | r = radeon_vm_update_page_directory(rdev, vm); |
||
476 | if (r) |
||
477 | return r; |
||
478 | |||
479 | r = radeon_vm_clear_freed(rdev, vm); |
||
480 | if (r) |
||
481 | return r; |
||
482 | |||
483 | if (vm->ib_bo_va == NULL) { |
||
484 | DRM_ERROR("Tmp BO not in VM!\n"); |
||
485 | return -EINVAL; |
||
486 | } |
||
487 | |||
488 | r = radeon_vm_bo_update(rdev, vm->ib_bo_va, |
||
489 | &rdev->ring_tmp_bo.bo->tbo.mem); |
||
490 | if (r) |
||
491 | return r; |
||
492 | |||
493 | for (i = 0; i < p->nrelocs; i++) { |
||
494 | struct radeon_bo *bo; |
||
495 | |||
496 | /* ignore duplicates */ |
||
497 | if (p->relocs_ptr[i] != &p->relocs[i]) |
||
498 | continue; |
||
499 | |||
500 | bo = p->relocs[i].robj; |
||
501 | bo_va = radeon_vm_bo_find(vm, bo); |
||
502 | if (bo_va == NULL) { |
||
503 | dev_err(rdev->dev, "bo %p not in vm %p\n", bo, vm); |
||
504 | return -EINVAL; |
||
505 | } |
||
506 | |||
507 | r = radeon_vm_bo_update(rdev, bo_va, &bo->tbo.mem); |
||
508 | if (r) |
||
509 | return r; |
||
510 | } |
||
511 | |||
512 | return radeon_vm_clear_invalids(rdev, vm); |
||
513 | } |
||
514 | |||
515 | static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev, |
||
516 | struct radeon_cs_parser *parser) |
||
517 | { |
||
518 | struct radeon_fpriv *fpriv = parser->filp->driver_priv; |
||
519 | struct radeon_vm *vm = &fpriv->vm; |
||
520 | int r; |
||
521 | |||
522 | if (parser->chunk_ib_idx == -1) |
||
523 | return 0; |
||
524 | if ((parser->cs_flags & RADEON_CS_USE_VM) == 0) |
||
525 | return 0; |
||
526 | |||
527 | if (parser->const_ib.length_dw) { |
||
528 | r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib); |
||
529 | if (r) { |
||
530 | return r; |
||
531 | } |
||
532 | } |
||
533 | |||
534 | r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib); |
||
535 | if (r) { |
||
536 | return r; |
||
537 | } |
||
538 | |||
539 | if (parser->ring == R600_RING_TYPE_UVD_INDEX) |
||
540 | radeon_uvd_note_usage(rdev); |
||
541 | |||
542 | mutex_lock(&vm->mutex); |
||
543 | r = radeon_bo_vm_update_pte(parser, vm); |
||
544 | if (r) { |
||
545 | goto out; |
||
546 | } |
||
547 | radeon_cs_sync_rings(parser); |
||
548 | radeon_semaphore_sync_to(parser->ib.semaphore, vm->fence); |
||
549 | |||
550 | if ((rdev->family >= CHIP_TAHITI) && |
||
551 | (parser->chunk_const_ib_idx != -1)) { |
||
552 | r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib, true); |
||
553 | } else { |
||
554 | r = radeon_ib_schedule(rdev, &parser->ib, NULL, true); |
||
555 | } |
||
556 | |||
557 | out: |
||
558 | mutex_unlock(&vm->mutex); |
||
559 | return r; |
||
560 | } |
||
561 | |||
562 | static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r) |
||
563 | { |
||
564 | if (r == -EDEADLK) { |
||
565 | r = radeon_gpu_reset(rdev); |
||
566 | if (!r) |
||
567 | r = -EAGAIN; |
||
568 | } |
||
569 | return r; |
||
570 | } |
||
571 | |||
572 | static int radeon_cs_ib_fill(struct radeon_device *rdev, struct radeon_cs_parser *parser) |
||
573 | { |
||
574 | struct radeon_cs_chunk *ib_chunk; |
||
575 | struct radeon_vm *vm = NULL; |
||
576 | int r; |
||
577 | |||
578 | if (parser->chunk_ib_idx == -1) |
||
579 | return 0; |
||
580 | |||
581 | if (parser->cs_flags & RADEON_CS_USE_VM) { |
||
582 | struct radeon_fpriv *fpriv = parser->filp->driver_priv; |
||
583 | vm = &fpriv->vm; |
||
584 | |||
585 | if ((rdev->family >= CHIP_TAHITI) && |
||
586 | (parser->chunk_const_ib_idx != -1)) { |
||
587 | ib_chunk = &parser->chunks[parser->chunk_const_ib_idx]; |
||
588 | if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) { |
||
589 | DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw); |
||
590 | return -EINVAL; |
||
591 | } |
||
592 | r = radeon_ib_get(rdev, parser->ring, &parser->const_ib, |
||
593 | vm, ib_chunk->length_dw * 4); |
||
594 | if (r) { |
||
595 | DRM_ERROR("Failed to get const ib !\n"); |
||
596 | return r; |
||
597 | } |
||
598 | parser->const_ib.is_const_ib = true; |
||
599 | parser->const_ib.length_dw = ib_chunk->length_dw; |
||
600 | if (copy_from_user(parser->const_ib.ptr, |
||
601 | ib_chunk->user_ptr, |
||
602 | ib_chunk->length_dw * 4)) |
||
603 | return -EFAULT; |
||
604 | } |
||
605 | |||
606 | ib_chunk = &parser->chunks[parser->chunk_ib_idx]; |
||
607 | if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) { |
||
608 | DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw); |
||
609 | return -EINVAL; |
||
610 | } |
||
611 | } |
||
612 | ib_chunk = &parser->chunks[parser->chunk_ib_idx]; |
||
613 | |||
614 | r = radeon_ib_get(rdev, parser->ring, &parser->ib, |
||
615 | vm, ib_chunk->length_dw * 4); |
||
616 | if (r) { |
||
617 | DRM_ERROR("Failed to get ib !\n"); |
||
618 | return r; |
||
619 | } |
||
620 | parser->ib.length_dw = ib_chunk->length_dw; |
||
621 | if (ib_chunk->kdata) |
||
622 | memcpy(parser->ib.ptr, ib_chunk->kdata, ib_chunk->length_dw * 4); |
||
623 | else if (copy_from_user(parser->ib.ptr, ib_chunk->user_ptr, ib_chunk->length_dw * 4)) |
||
624 | return -EFAULT; |
||
625 | return 0; |
||
626 | } |
||
627 | |||
628 | int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp) |
||
629 | { |
||
630 | struct radeon_device *rdev = dev->dev_private; |
||
631 | struct radeon_cs_parser parser; |
||
632 | int r; |
||
633 | |||
634 | // down_read(&rdev->exclusive_lock); |
||
635 | if (!rdev->accel_working) { |
||
636 | // up_read(&rdev->exclusive_lock); |
||
637 | return -EBUSY; |
||
638 | } |
||
639 | /* initialize parser */ |
||
640 | memset(&parser, 0, sizeof(struct radeon_cs_parser)); |
||
641 | parser.filp = filp; |
||
642 | parser.rdev = rdev; |
||
643 | parser.dev = rdev->dev; |
||
644 | parser.family = rdev->family; |
||
645 | r = radeon_cs_parser_init(&parser, data); |
||
646 | if (r) { |
||
647 | DRM_ERROR("Failed to initialize parser !\n"); |
||
648 | radeon_cs_parser_fini(&parser, r, false); |
||
649 | // up_read(&rdev->exclusive_lock); |
||
650 | r = radeon_cs_handle_lockup(rdev, r); |
||
651 | return r; |
||
652 | } |
||
653 | |||
654 | r = radeon_cs_ib_fill(rdev, &parser); |
||
655 | if (!r) { |
||
656 | r = radeon_cs_parser_relocs(&parser); |
||
657 | if (r && r != -ERESTARTSYS) |
||
658 | DRM_ERROR("Failed to parse relocation %d!\n", r); |
||
659 | } |
||
660 | |||
661 | if (r) { |
||
662 | radeon_cs_parser_fini(&parser, r, false); |
||
663 | // up_read(&rdev->exclusive_lock); |
||
664 | r = radeon_cs_handle_lockup(rdev, r); |
||
665 | return r; |
||
666 | } |
||
667 | |||
668 | trace_radeon_cs(&parser); |
||
669 | |||
670 | r = radeon_cs_ib_chunk(rdev, &parser); |
||
671 | if (r) { |
||
672 | goto out; |
||
673 | } |
||
674 | r = radeon_cs_ib_vm_chunk(rdev, &parser); |
||
675 | if (r) { |
||
676 | goto out; |
||
677 | } |
||
678 | out: |
||
679 | radeon_cs_parser_fini(&parser, r, true); |
||
680 | // up_read(&rdev->exclusive_lock); |
||
681 | r = radeon_cs_handle_lockup(rdev, r); |
||
682 | return r; |
||
683 | } |
||
684 | |||
685 | /** |
||
686 | * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet |
||
687 | * @parser: parser structure holding parsing context. |
||
688 | * @pkt: where to store packet information |
||
689 | * |
||
690 | * Assume that chunk_ib_index is properly set. Will return -EINVAL |
||
691 | * if packet is bigger than remaining ib size. or if packets is unknown. |
||
692 | **/ |
||
693 | int radeon_cs_packet_parse(struct radeon_cs_parser *p, |
||
694 | struct radeon_cs_packet *pkt, |
||
695 | unsigned idx) |
||
696 | { |
||
697 | struct radeon_cs_chunk *ib_chunk = &p->chunks[p->chunk_ib_idx]; |
||
698 | struct radeon_device *rdev = p->rdev; |
||
699 | uint32_t header; |
||
700 | |||
701 | if (idx >= ib_chunk->length_dw) { |
||
702 | DRM_ERROR("Can not parse packet at %d after CS end %d !\n", |
||
703 | idx, ib_chunk->length_dw); |
||
704 | return -EINVAL; |
||
705 | } |
||
706 | header = radeon_get_ib_value(p, idx); |
||
707 | pkt->idx = idx; |
||
708 | pkt->type = RADEON_CP_PACKET_GET_TYPE(header); |
||
709 | pkt->count = RADEON_CP_PACKET_GET_COUNT(header); |
||
710 | pkt->one_reg_wr = 0; |
||
711 | switch (pkt->type) { |
||
712 | case RADEON_PACKET_TYPE0: |
||
713 | if (rdev->family < CHIP_R600) { |
||
714 | pkt->reg = R100_CP_PACKET0_GET_REG(header); |
||
715 | pkt->one_reg_wr = |
||
716 | RADEON_CP_PACKET0_GET_ONE_REG_WR(header); |
||
717 | } else |
||
718 | pkt->reg = R600_CP_PACKET0_GET_REG(header); |
||
719 | break; |
||
720 | case RADEON_PACKET_TYPE3: |
||
721 | pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header); |
||
722 | break; |
||
723 | case RADEON_PACKET_TYPE2: |
||
724 | pkt->count = -1; |
||
725 | break; |
||
726 | default: |
||
727 | DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx); |
||
728 | return -EINVAL; |
||
729 | } |
||
730 | if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) { |
||
731 | DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n", |
||
732 | pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw); |
||
733 | return -EINVAL; |
||
734 | } |
||
735 | return 0; |
||
736 | } |
||
737 | |||
738 | /** |
||
739 | * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP |
||
740 | * @p: structure holding the parser context. |
||
741 | * |
||
742 | * Check if the next packet is NOP relocation packet3. |
||
743 | **/ |
||
744 | bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p) |
||
745 | { |
||
746 | struct radeon_cs_packet p3reloc; |
||
747 | int r; |
||
748 | |||
749 | r = radeon_cs_packet_parse(p, &p3reloc, p->idx); |
||
750 | if (r) |
||
751 | return false; |
||
752 | if (p3reloc.type != RADEON_PACKET_TYPE3) |
||
753 | return false; |
||
754 | if (p3reloc.opcode != RADEON_PACKET3_NOP) |
||
755 | return false; |
||
756 | return true; |
||
757 | } |
||
758 | |||
759 | /** |
||
760 | * radeon_cs_dump_packet() - dump raw packet context |
||
761 | * @p: structure holding the parser context. |
||
762 | * @pkt: structure holding the packet. |
||
763 | * |
||
764 | * Used mostly for debugging and error reporting. |
||
765 | **/ |
||
766 | void radeon_cs_dump_packet(struct radeon_cs_parser *p, |
||
767 | struct radeon_cs_packet *pkt) |
||
768 | { |
||
769 | volatile uint32_t *ib; |
||
770 | unsigned i; |
||
771 | unsigned idx; |
||
772 | |||
773 | ib = p->ib.ptr; |
||
774 | idx = pkt->idx; |
||
775 | for (i = 0; i <= (pkt->count + 1); i++, idx++) |
||
776 | DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]); |
||
777 | } |
||
778 | |||
779 | /** |
||
780 | * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet |
||
781 | * @parser: parser structure holding parsing context. |
||
782 | * @data: pointer to relocation data |
||
783 | * @offset_start: starting offset |
||
784 | * @offset_mask: offset mask (to align start offset on) |
||
785 | * @reloc: reloc informations |
||
786 | * |
||
787 | * Check if next packet is relocation packet3, do bo validation and compute |
||
788 | * GPU offset using the provided start. |
||
789 | **/ |
||
790 | int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p, |
||
791 | struct radeon_cs_reloc **cs_reloc, |
||
792 | int nomm) |
||
793 | { |
||
794 | struct radeon_cs_chunk *relocs_chunk; |
||
795 | struct radeon_cs_packet p3reloc; |
||
796 | unsigned idx; |
||
797 | int r; |
||
798 | |||
799 | if (p->chunk_relocs_idx == -1) { |
||
800 | DRM_ERROR("No relocation chunk !\n"); |
||
801 | return -EINVAL; |
||
802 | } |
||
803 | *cs_reloc = NULL; |
||
804 | relocs_chunk = &p->chunks[p->chunk_relocs_idx]; |
||
805 | r = radeon_cs_packet_parse(p, &p3reloc, p->idx); |
||
806 | if (r) |
||
807 | return r; |
||
808 | p->idx += p3reloc.count + 2; |
||
809 | if (p3reloc.type != RADEON_PACKET_TYPE3 || |
||
810 | p3reloc.opcode != RADEON_PACKET3_NOP) { |
||
811 | DRM_ERROR("No packet3 for relocation for packet at %d.\n", |
||
812 | p3reloc.idx); |
||
813 | radeon_cs_dump_packet(p, &p3reloc); |
||
814 | return -EINVAL; |
||
815 | } |
||
816 | idx = radeon_get_ib_value(p, p3reloc.idx + 1); |
||
817 | if (idx >= relocs_chunk->length_dw) { |
||
818 | DRM_ERROR("Relocs at %d after relocations chunk end %d !\n", |
||
819 | idx, relocs_chunk->length_dw); |
||
820 | radeon_cs_dump_packet(p, &p3reloc); |
||
821 | return -EINVAL; |
||
822 | } |
||
823 | /* FIXME: we assume reloc size is 4 dwords */ |
||
824 | if (nomm) { |
||
825 | *cs_reloc = p->relocs; |
||
826 | (*cs_reloc)->gpu_offset = |
||
827 | (u64)relocs_chunk->kdata[idx + 3] << 32; |
||
828 | (*cs_reloc)->gpu_offset |= relocs_chunk->kdata[idx + 0]; |
||
829 | } else |
||
830 | *cs_reloc = p->relocs_ptr[(idx / 4)]; |
||
831 | return 0; |
||
832 | }><>=>>>>>>>>>>> |