]>
Commit | Line | Data |
---|---|---|
673a394b EA |
1 | /* |
2 | * Copyright © 2008 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 | * Authors: | |
24 | * Eric Anholt <eric@anholt.net> | |
25 | * | |
26 | */ | |
27 | ||
28 | #include "drmP.h" | |
29 | #include "drm.h" | |
30 | #include "i915_drm.h" | |
31 | #include "i915_drv.h" | |
32 | #include <linux/swap.h> | |
33 | ||
28dfe52a EA |
34 | #define I915_GEM_GPU_DOMAINS (~(I915_GEM_DOMAIN_CPU | I915_GEM_DOMAIN_GTT)) |
35 | ||
673a394b EA |
36 | static int |
37 | i915_gem_object_set_domain(struct drm_gem_object *obj, | |
38 | uint32_t read_domains, | |
39 | uint32_t write_domain); | |
40 | static int | |
41 | i915_gem_object_set_domain_range(struct drm_gem_object *obj, | |
42 | uint64_t offset, | |
43 | uint64_t size, | |
44 | uint32_t read_domains, | |
45 | uint32_t write_domain); | |
46 | static int | |
47 | i915_gem_set_domain(struct drm_gem_object *obj, | |
48 | struct drm_file *file_priv, | |
49 | uint32_t read_domains, | |
50 | uint32_t write_domain); | |
51 | static int i915_gem_object_get_page_list(struct drm_gem_object *obj); | |
52 | static void i915_gem_object_free_page_list(struct drm_gem_object *obj); | |
53 | static int i915_gem_object_wait_rendering(struct drm_gem_object *obj); | |
54 | ||
6dbe2772 KP |
55 | static void |
56 | i915_gem_cleanup_ringbuffer(struct drm_device *dev); | |
57 | ||
673a394b EA |
58 | int |
59 | i915_gem_init_ioctl(struct drm_device *dev, void *data, | |
60 | struct drm_file *file_priv) | |
61 | { | |
62 | drm_i915_private_t *dev_priv = dev->dev_private; | |
63 | struct drm_i915_gem_init *args = data; | |
64 | ||
65 | mutex_lock(&dev->struct_mutex); | |
66 | ||
67 | if (args->gtt_start >= args->gtt_end || | |
68 | (args->gtt_start & (PAGE_SIZE - 1)) != 0 || | |
69 | (args->gtt_end & (PAGE_SIZE - 1)) != 0) { | |
70 | mutex_unlock(&dev->struct_mutex); | |
71 | return -EINVAL; | |
72 | } | |
73 | ||
74 | drm_mm_init(&dev_priv->mm.gtt_space, args->gtt_start, | |
75 | args->gtt_end - args->gtt_start); | |
76 | ||
77 | dev->gtt_total = (uint32_t) (args->gtt_end - args->gtt_start); | |
78 | ||
79 | mutex_unlock(&dev->struct_mutex); | |
80 | ||
81 | return 0; | |
82 | } | |
83 | ||
5a125c3c EA |
84 | int |
85 | i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data, | |
86 | struct drm_file *file_priv) | |
87 | { | |
5a125c3c | 88 | struct drm_i915_gem_get_aperture *args = data; |
5a125c3c EA |
89 | |
90 | if (!(dev->driver->driver_features & DRIVER_GEM)) | |
91 | return -ENODEV; | |
92 | ||
93 | args->aper_size = dev->gtt_total; | |
2678d9d6 KP |
94 | args->aper_available_size = (args->aper_size - |
95 | atomic_read(&dev->pin_memory)); | |
5a125c3c EA |
96 | |
97 | return 0; | |
98 | } | |
99 | ||
673a394b EA |
100 | |
101 | /** | |
102 | * Creates a new mm object and returns a handle to it. | |
103 | */ | |
104 | int | |
105 | i915_gem_create_ioctl(struct drm_device *dev, void *data, | |
106 | struct drm_file *file_priv) | |
107 | { | |
108 | struct drm_i915_gem_create *args = data; | |
109 | struct drm_gem_object *obj; | |
110 | int handle, ret; | |
111 | ||
112 | args->size = roundup(args->size, PAGE_SIZE); | |
113 | ||
114 | /* Allocate the new object */ | |
115 | obj = drm_gem_object_alloc(dev, args->size); | |
116 | if (obj == NULL) | |
117 | return -ENOMEM; | |
118 | ||
119 | ret = drm_gem_handle_create(file_priv, obj, &handle); | |
120 | mutex_lock(&dev->struct_mutex); | |
121 | drm_gem_object_handle_unreference(obj); | |
122 | mutex_unlock(&dev->struct_mutex); | |
123 | ||
124 | if (ret) | |
125 | return ret; | |
126 | ||
127 | args->handle = handle; | |
128 | ||
129 | return 0; | |
130 | } | |
131 | ||
132 | /** | |
133 | * Reads data from the object referenced by handle. | |
134 | * | |
135 | * On error, the contents of *data are undefined. | |
136 | */ | |
137 | int | |
138 | i915_gem_pread_ioctl(struct drm_device *dev, void *data, | |
139 | struct drm_file *file_priv) | |
140 | { | |
141 | struct drm_i915_gem_pread *args = data; | |
142 | struct drm_gem_object *obj; | |
143 | struct drm_i915_gem_object *obj_priv; | |
144 | ssize_t read; | |
145 | loff_t offset; | |
146 | int ret; | |
147 | ||
148 | obj = drm_gem_object_lookup(dev, file_priv, args->handle); | |
149 | if (obj == NULL) | |
150 | return -EBADF; | |
151 | obj_priv = obj->driver_private; | |
152 | ||
153 | /* Bounds check source. | |
154 | * | |
155 | * XXX: This could use review for overflow issues... | |
156 | */ | |
157 | if (args->offset > obj->size || args->size > obj->size || | |
158 | args->offset + args->size > obj->size) { | |
159 | drm_gem_object_unreference(obj); | |
160 | return -EINVAL; | |
161 | } | |
162 | ||
163 | mutex_lock(&dev->struct_mutex); | |
164 | ||
165 | ret = i915_gem_object_set_domain_range(obj, args->offset, args->size, | |
166 | I915_GEM_DOMAIN_CPU, 0); | |
167 | if (ret != 0) { | |
168 | drm_gem_object_unreference(obj); | |
169 | mutex_unlock(&dev->struct_mutex); | |
e7d22bc3 | 170 | return ret; |
673a394b EA |
171 | } |
172 | ||
173 | offset = args->offset; | |
174 | ||
175 | read = vfs_read(obj->filp, (char __user *)(uintptr_t)args->data_ptr, | |
176 | args->size, &offset); | |
177 | if (read != args->size) { | |
178 | drm_gem_object_unreference(obj); | |
179 | mutex_unlock(&dev->struct_mutex); | |
180 | if (read < 0) | |
181 | return read; | |
182 | else | |
183 | return -EINVAL; | |
184 | } | |
185 | ||
186 | drm_gem_object_unreference(obj); | |
187 | mutex_unlock(&dev->struct_mutex); | |
188 | ||
189 | return 0; | |
190 | } | |
191 | ||
0839ccb8 KP |
192 | /* This is the fast write path which cannot handle |
193 | * page faults in the source data | |
9b7530cc | 194 | */ |
0839ccb8 KP |
195 | |
196 | static inline int | |
197 | fast_user_write(struct io_mapping *mapping, | |
198 | loff_t page_base, int page_offset, | |
199 | char __user *user_data, | |
200 | int length) | |
9b7530cc | 201 | { |
9b7530cc | 202 | char *vaddr_atomic; |
0839ccb8 | 203 | unsigned long unwritten; |
9b7530cc | 204 | |
0839ccb8 KP |
205 | vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base); |
206 | unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + page_offset, | |
207 | user_data, length); | |
208 | io_mapping_unmap_atomic(vaddr_atomic); | |
209 | if (unwritten) | |
210 | return -EFAULT; | |
211 | return 0; | |
212 | } | |
213 | ||
214 | /* Here's the write path which can sleep for | |
215 | * page faults | |
216 | */ | |
217 | ||
218 | static inline int | |
219 | slow_user_write(struct io_mapping *mapping, | |
220 | loff_t page_base, int page_offset, | |
221 | char __user *user_data, | |
222 | int length) | |
223 | { | |
224 | char __iomem *vaddr; | |
225 | unsigned long unwritten; | |
226 | ||
227 | vaddr = io_mapping_map_wc(mapping, page_base); | |
228 | if (vaddr == NULL) | |
229 | return -EFAULT; | |
230 | unwritten = __copy_from_user(vaddr + page_offset, | |
231 | user_data, length); | |
232 | io_mapping_unmap(vaddr); | |
233 | if (unwritten) | |
234 | return -EFAULT; | |
9b7530cc | 235 | return 0; |
9b7530cc LT |
236 | } |
237 | ||
673a394b EA |
238 | static int |
239 | i915_gem_gtt_pwrite(struct drm_device *dev, struct drm_gem_object *obj, | |
240 | struct drm_i915_gem_pwrite *args, | |
241 | struct drm_file *file_priv) | |
242 | { | |
243 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
0839ccb8 | 244 | drm_i915_private_t *dev_priv = dev->dev_private; |
673a394b | 245 | ssize_t remain; |
0839ccb8 | 246 | loff_t offset, page_base; |
673a394b | 247 | char __user *user_data; |
0839ccb8 KP |
248 | int page_offset, page_length; |
249 | int ret; | |
673a394b EA |
250 | |
251 | user_data = (char __user *) (uintptr_t) args->data_ptr; | |
252 | remain = args->size; | |
253 | if (!access_ok(VERIFY_READ, user_data, remain)) | |
254 | return -EFAULT; | |
255 | ||
256 | ||
257 | mutex_lock(&dev->struct_mutex); | |
258 | ret = i915_gem_object_pin(obj, 0); | |
259 | if (ret) { | |
260 | mutex_unlock(&dev->struct_mutex); | |
261 | return ret; | |
262 | } | |
263 | ret = i915_gem_set_domain(obj, file_priv, | |
264 | I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT); | |
265 | if (ret) | |
266 | goto fail; | |
267 | ||
268 | obj_priv = obj->driver_private; | |
269 | offset = obj_priv->gtt_offset + args->offset; | |
270 | obj_priv->dirty = 1; | |
271 | ||
272 | while (remain > 0) { | |
273 | /* Operation in this page | |
274 | * | |
0839ccb8 KP |
275 | * page_base = page offset within aperture |
276 | * page_offset = offset within page | |
277 | * page_length = bytes to copy for this page | |
673a394b | 278 | */ |
0839ccb8 KP |
279 | page_base = (offset & ~(PAGE_SIZE-1)); |
280 | page_offset = offset & (PAGE_SIZE-1); | |
281 | page_length = remain; | |
282 | if ((page_offset + remain) > PAGE_SIZE) | |
283 | page_length = PAGE_SIZE - page_offset; | |
284 | ||
285 | ret = fast_user_write (dev_priv->mm.gtt_mapping, page_base, | |
286 | page_offset, user_data, page_length); | |
287 | ||
288 | /* If we get a fault while copying data, then (presumably) our | |
289 | * source page isn't available. In this case, use the | |
290 | * non-atomic function | |
291 | */ | |
292 | if (ret) { | |
293 | ret = slow_user_write (dev_priv->mm.gtt_mapping, | |
294 | page_base, page_offset, | |
295 | user_data, page_length); | |
296 | if (ret) | |
673a394b | 297 | goto fail; |
673a394b EA |
298 | } |
299 | ||
0839ccb8 KP |
300 | remain -= page_length; |
301 | user_data += page_length; | |
302 | offset += page_length; | |
673a394b | 303 | } |
673a394b EA |
304 | |
305 | fail: | |
306 | i915_gem_object_unpin(obj); | |
307 | mutex_unlock(&dev->struct_mutex); | |
308 | ||
309 | return ret; | |
310 | } | |
311 | ||
3043c60c | 312 | static int |
673a394b EA |
313 | i915_gem_shmem_pwrite(struct drm_device *dev, struct drm_gem_object *obj, |
314 | struct drm_i915_gem_pwrite *args, | |
315 | struct drm_file *file_priv) | |
316 | { | |
317 | int ret; | |
318 | loff_t offset; | |
319 | ssize_t written; | |
320 | ||
321 | mutex_lock(&dev->struct_mutex); | |
322 | ||
323 | ret = i915_gem_set_domain(obj, file_priv, | |
324 | I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU); | |
325 | if (ret) { | |
326 | mutex_unlock(&dev->struct_mutex); | |
327 | return ret; | |
328 | } | |
329 | ||
330 | offset = args->offset; | |
331 | ||
332 | written = vfs_write(obj->filp, | |
333 | (char __user *)(uintptr_t) args->data_ptr, | |
334 | args->size, &offset); | |
335 | if (written != args->size) { | |
336 | mutex_unlock(&dev->struct_mutex); | |
337 | if (written < 0) | |
338 | return written; | |
339 | else | |
340 | return -EINVAL; | |
341 | } | |
342 | ||
343 | mutex_unlock(&dev->struct_mutex); | |
344 | ||
345 | return 0; | |
346 | } | |
347 | ||
348 | /** | |
349 | * Writes data to the object referenced by handle. | |
350 | * | |
351 | * On error, the contents of the buffer that were to be modified are undefined. | |
352 | */ | |
353 | int | |
354 | i915_gem_pwrite_ioctl(struct drm_device *dev, void *data, | |
355 | struct drm_file *file_priv) | |
356 | { | |
357 | struct drm_i915_gem_pwrite *args = data; | |
358 | struct drm_gem_object *obj; | |
359 | struct drm_i915_gem_object *obj_priv; | |
360 | int ret = 0; | |
361 | ||
362 | obj = drm_gem_object_lookup(dev, file_priv, args->handle); | |
363 | if (obj == NULL) | |
364 | return -EBADF; | |
365 | obj_priv = obj->driver_private; | |
366 | ||
367 | /* Bounds check destination. | |
368 | * | |
369 | * XXX: This could use review for overflow issues... | |
370 | */ | |
371 | if (args->offset > obj->size || args->size > obj->size || | |
372 | args->offset + args->size > obj->size) { | |
373 | drm_gem_object_unreference(obj); | |
374 | return -EINVAL; | |
375 | } | |
376 | ||
377 | /* We can only do the GTT pwrite on untiled buffers, as otherwise | |
378 | * it would end up going through the fenced access, and we'll get | |
379 | * different detiling behavior between reading and writing. | |
380 | * pread/pwrite currently are reading and writing from the CPU | |
381 | * perspective, requiring manual detiling by the client. | |
382 | */ | |
383 | if (obj_priv->tiling_mode == I915_TILING_NONE && | |
384 | dev->gtt_total != 0) | |
385 | ret = i915_gem_gtt_pwrite(dev, obj, args, file_priv); | |
386 | else | |
387 | ret = i915_gem_shmem_pwrite(dev, obj, args, file_priv); | |
388 | ||
389 | #if WATCH_PWRITE | |
390 | if (ret) | |
391 | DRM_INFO("pwrite failed %d\n", ret); | |
392 | #endif | |
393 | ||
394 | drm_gem_object_unreference(obj); | |
395 | ||
396 | return ret; | |
397 | } | |
398 | ||
399 | /** | |
400 | * Called when user space prepares to use an object | |
401 | */ | |
402 | int | |
403 | i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, | |
404 | struct drm_file *file_priv) | |
405 | { | |
406 | struct drm_i915_gem_set_domain *args = data; | |
407 | struct drm_gem_object *obj; | |
408 | int ret; | |
409 | ||
410 | if (!(dev->driver->driver_features & DRIVER_GEM)) | |
411 | return -ENODEV; | |
412 | ||
413 | obj = drm_gem_object_lookup(dev, file_priv, args->handle); | |
414 | if (obj == NULL) | |
415 | return -EBADF; | |
416 | ||
417 | mutex_lock(&dev->struct_mutex); | |
418 | #if WATCH_BUF | |
419 | DRM_INFO("set_domain_ioctl %p(%d), %08x %08x\n", | |
420 | obj, obj->size, args->read_domains, args->write_domain); | |
421 | #endif | |
422 | ret = i915_gem_set_domain(obj, file_priv, | |
423 | args->read_domains, args->write_domain); | |
424 | drm_gem_object_unreference(obj); | |
425 | mutex_unlock(&dev->struct_mutex); | |
426 | return ret; | |
427 | } | |
428 | ||
429 | /** | |
430 | * Called when user space has done writes to this buffer | |
431 | */ | |
432 | int | |
433 | i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data, | |
434 | struct drm_file *file_priv) | |
435 | { | |
436 | struct drm_i915_gem_sw_finish *args = data; | |
437 | struct drm_gem_object *obj; | |
438 | struct drm_i915_gem_object *obj_priv; | |
439 | int ret = 0; | |
440 | ||
441 | if (!(dev->driver->driver_features & DRIVER_GEM)) | |
442 | return -ENODEV; | |
443 | ||
444 | mutex_lock(&dev->struct_mutex); | |
445 | obj = drm_gem_object_lookup(dev, file_priv, args->handle); | |
446 | if (obj == NULL) { | |
447 | mutex_unlock(&dev->struct_mutex); | |
448 | return -EBADF; | |
449 | } | |
450 | ||
451 | #if WATCH_BUF | |
452 | DRM_INFO("%s: sw_finish %d (%p %d)\n", | |
453 | __func__, args->handle, obj, obj->size); | |
454 | #endif | |
455 | obj_priv = obj->driver_private; | |
456 | ||
457 | /* Pinned buffers may be scanout, so flush the cache */ | |
458 | if ((obj->write_domain & I915_GEM_DOMAIN_CPU) && obj_priv->pin_count) { | |
459 | i915_gem_clflush_object(obj); | |
460 | drm_agp_chipset_flush(dev); | |
461 | } | |
462 | drm_gem_object_unreference(obj); | |
463 | mutex_unlock(&dev->struct_mutex); | |
464 | return ret; | |
465 | } | |
466 | ||
467 | /** | |
468 | * Maps the contents of an object, returning the address it is mapped | |
469 | * into. | |
470 | * | |
471 | * While the mapping holds a reference on the contents of the object, it doesn't | |
472 | * imply a ref on the object itself. | |
473 | */ | |
474 | int | |
475 | i915_gem_mmap_ioctl(struct drm_device *dev, void *data, | |
476 | struct drm_file *file_priv) | |
477 | { | |
478 | struct drm_i915_gem_mmap *args = data; | |
479 | struct drm_gem_object *obj; | |
480 | loff_t offset; | |
481 | unsigned long addr; | |
482 | ||
483 | if (!(dev->driver->driver_features & DRIVER_GEM)) | |
484 | return -ENODEV; | |
485 | ||
486 | obj = drm_gem_object_lookup(dev, file_priv, args->handle); | |
487 | if (obj == NULL) | |
488 | return -EBADF; | |
489 | ||
490 | offset = args->offset; | |
491 | ||
492 | down_write(¤t->mm->mmap_sem); | |
493 | addr = do_mmap(obj->filp, 0, args->size, | |
494 | PROT_READ | PROT_WRITE, MAP_SHARED, | |
495 | args->offset); | |
496 | up_write(¤t->mm->mmap_sem); | |
497 | mutex_lock(&dev->struct_mutex); | |
498 | drm_gem_object_unreference(obj); | |
499 | mutex_unlock(&dev->struct_mutex); | |
500 | if (IS_ERR((void *)addr)) | |
501 | return addr; | |
502 | ||
503 | args->addr_ptr = (uint64_t) addr; | |
504 | ||
505 | return 0; | |
506 | } | |
507 | ||
508 | static void | |
509 | i915_gem_object_free_page_list(struct drm_gem_object *obj) | |
510 | { | |
511 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
512 | int page_count = obj->size / PAGE_SIZE; | |
513 | int i; | |
514 | ||
515 | if (obj_priv->page_list == NULL) | |
516 | return; | |
517 | ||
518 | ||
519 | for (i = 0; i < page_count; i++) | |
520 | if (obj_priv->page_list[i] != NULL) { | |
521 | if (obj_priv->dirty) | |
522 | set_page_dirty(obj_priv->page_list[i]); | |
523 | mark_page_accessed(obj_priv->page_list[i]); | |
524 | page_cache_release(obj_priv->page_list[i]); | |
525 | } | |
526 | obj_priv->dirty = 0; | |
527 | ||
528 | drm_free(obj_priv->page_list, | |
529 | page_count * sizeof(struct page *), | |
530 | DRM_MEM_DRIVER); | |
531 | obj_priv->page_list = NULL; | |
532 | } | |
533 | ||
534 | static void | |
ce44b0ea | 535 | i915_gem_object_move_to_active(struct drm_gem_object *obj, uint32_t seqno) |
673a394b EA |
536 | { |
537 | struct drm_device *dev = obj->dev; | |
538 | drm_i915_private_t *dev_priv = dev->dev_private; | |
539 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
540 | ||
541 | /* Add a reference if we're newly entering the active list. */ | |
542 | if (!obj_priv->active) { | |
543 | drm_gem_object_reference(obj); | |
544 | obj_priv->active = 1; | |
545 | } | |
546 | /* Move from whatever list we were on to the tail of execution. */ | |
547 | list_move_tail(&obj_priv->list, | |
548 | &dev_priv->mm.active_list); | |
ce44b0ea | 549 | obj_priv->last_rendering_seqno = seqno; |
673a394b EA |
550 | } |
551 | ||
ce44b0ea EA |
552 | static void |
553 | i915_gem_object_move_to_flushing(struct drm_gem_object *obj) | |
554 | { | |
555 | struct drm_device *dev = obj->dev; | |
556 | drm_i915_private_t *dev_priv = dev->dev_private; | |
557 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
558 | ||
559 | BUG_ON(!obj_priv->active); | |
560 | list_move_tail(&obj_priv->list, &dev_priv->mm.flushing_list); | |
561 | obj_priv->last_rendering_seqno = 0; | |
562 | } | |
673a394b EA |
563 | |
564 | static void | |
565 | i915_gem_object_move_to_inactive(struct drm_gem_object *obj) | |
566 | { | |
567 | struct drm_device *dev = obj->dev; | |
568 | drm_i915_private_t *dev_priv = dev->dev_private; | |
569 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
570 | ||
571 | i915_verify_inactive(dev, __FILE__, __LINE__); | |
572 | if (obj_priv->pin_count != 0) | |
573 | list_del_init(&obj_priv->list); | |
574 | else | |
575 | list_move_tail(&obj_priv->list, &dev_priv->mm.inactive_list); | |
576 | ||
ce44b0ea | 577 | obj_priv->last_rendering_seqno = 0; |
673a394b EA |
578 | if (obj_priv->active) { |
579 | obj_priv->active = 0; | |
580 | drm_gem_object_unreference(obj); | |
581 | } | |
582 | i915_verify_inactive(dev, __FILE__, __LINE__); | |
583 | } | |
584 | ||
585 | /** | |
586 | * Creates a new sequence number, emitting a write of it to the status page | |
587 | * plus an interrupt, which will trigger i915_user_interrupt_handler. | |
588 | * | |
589 | * Must be called with struct_lock held. | |
590 | * | |
591 | * Returned sequence numbers are nonzero on success. | |
592 | */ | |
593 | static uint32_t | |
594 | i915_add_request(struct drm_device *dev, uint32_t flush_domains) | |
595 | { | |
596 | drm_i915_private_t *dev_priv = dev->dev_private; | |
597 | struct drm_i915_gem_request *request; | |
598 | uint32_t seqno; | |
599 | int was_empty; | |
600 | RING_LOCALS; | |
601 | ||
602 | request = drm_calloc(1, sizeof(*request), DRM_MEM_DRIVER); | |
603 | if (request == NULL) | |
604 | return 0; | |
605 | ||
606 | /* Grab the seqno we're going to make this request be, and bump the | |
607 | * next (skipping 0 so it can be the reserved no-seqno value). | |
608 | */ | |
609 | seqno = dev_priv->mm.next_gem_seqno; | |
610 | dev_priv->mm.next_gem_seqno++; | |
611 | if (dev_priv->mm.next_gem_seqno == 0) | |
612 | dev_priv->mm.next_gem_seqno++; | |
613 | ||
614 | BEGIN_LP_RING(4); | |
615 | OUT_RING(MI_STORE_DWORD_INDEX); | |
616 | OUT_RING(I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT); | |
617 | OUT_RING(seqno); | |
618 | ||
619 | OUT_RING(MI_USER_INTERRUPT); | |
620 | ADVANCE_LP_RING(); | |
621 | ||
622 | DRM_DEBUG("%d\n", seqno); | |
623 | ||
624 | request->seqno = seqno; | |
625 | request->emitted_jiffies = jiffies; | |
673a394b EA |
626 | was_empty = list_empty(&dev_priv->mm.request_list); |
627 | list_add_tail(&request->list, &dev_priv->mm.request_list); | |
628 | ||
ce44b0ea EA |
629 | /* Associate any objects on the flushing list matching the write |
630 | * domain we're flushing with our flush. | |
631 | */ | |
632 | if (flush_domains != 0) { | |
633 | struct drm_i915_gem_object *obj_priv, *next; | |
634 | ||
635 | list_for_each_entry_safe(obj_priv, next, | |
636 | &dev_priv->mm.flushing_list, list) { | |
637 | struct drm_gem_object *obj = obj_priv->obj; | |
638 | ||
639 | if ((obj->write_domain & flush_domains) == | |
640 | obj->write_domain) { | |
641 | obj->write_domain = 0; | |
642 | i915_gem_object_move_to_active(obj, seqno); | |
643 | } | |
644 | } | |
645 | ||
646 | } | |
647 | ||
6dbe2772 | 648 | if (was_empty && !dev_priv->mm.suspended) |
673a394b EA |
649 | schedule_delayed_work(&dev_priv->mm.retire_work, HZ); |
650 | return seqno; | |
651 | } | |
652 | ||
653 | /** | |
654 | * Command execution barrier | |
655 | * | |
656 | * Ensures that all commands in the ring are finished | |
657 | * before signalling the CPU | |
658 | */ | |
3043c60c | 659 | static uint32_t |
673a394b EA |
660 | i915_retire_commands(struct drm_device *dev) |
661 | { | |
662 | drm_i915_private_t *dev_priv = dev->dev_private; | |
663 | uint32_t cmd = MI_FLUSH | MI_NO_WRITE_FLUSH; | |
664 | uint32_t flush_domains = 0; | |
665 | RING_LOCALS; | |
666 | ||
667 | /* The sampler always gets flushed on i965 (sigh) */ | |
668 | if (IS_I965G(dev)) | |
669 | flush_domains |= I915_GEM_DOMAIN_SAMPLER; | |
670 | BEGIN_LP_RING(2); | |
671 | OUT_RING(cmd); | |
672 | OUT_RING(0); /* noop */ | |
673 | ADVANCE_LP_RING(); | |
674 | return flush_domains; | |
675 | } | |
676 | ||
677 | /** | |
678 | * Moves buffers associated only with the given active seqno from the active | |
679 | * to inactive list, potentially freeing them. | |
680 | */ | |
681 | static void | |
682 | i915_gem_retire_request(struct drm_device *dev, | |
683 | struct drm_i915_gem_request *request) | |
684 | { | |
685 | drm_i915_private_t *dev_priv = dev->dev_private; | |
686 | ||
687 | /* Move any buffers on the active list that are no longer referenced | |
688 | * by the ringbuffer to the flushing/inactive lists as appropriate. | |
689 | */ | |
690 | while (!list_empty(&dev_priv->mm.active_list)) { | |
691 | struct drm_gem_object *obj; | |
692 | struct drm_i915_gem_object *obj_priv; | |
693 | ||
694 | obj_priv = list_first_entry(&dev_priv->mm.active_list, | |
695 | struct drm_i915_gem_object, | |
696 | list); | |
697 | obj = obj_priv->obj; | |
698 | ||
699 | /* If the seqno being retired doesn't match the oldest in the | |
700 | * list, then the oldest in the list must still be newer than | |
701 | * this seqno. | |
702 | */ | |
703 | if (obj_priv->last_rendering_seqno != request->seqno) | |
704 | return; | |
705 | #if WATCH_LRU | |
706 | DRM_INFO("%s: retire %d moves to inactive list %p\n", | |
707 | __func__, request->seqno, obj); | |
708 | #endif | |
709 | ||
ce44b0ea EA |
710 | if (obj->write_domain != 0) |
711 | i915_gem_object_move_to_flushing(obj); | |
712 | else | |
673a394b | 713 | i915_gem_object_move_to_inactive(obj); |
673a394b EA |
714 | } |
715 | } | |
716 | ||
717 | /** | |
718 | * Returns true if seq1 is later than seq2. | |
719 | */ | |
720 | static int | |
721 | i915_seqno_passed(uint32_t seq1, uint32_t seq2) | |
722 | { | |
723 | return (int32_t)(seq1 - seq2) >= 0; | |
724 | } | |
725 | ||
726 | uint32_t | |
727 | i915_get_gem_seqno(struct drm_device *dev) | |
728 | { | |
729 | drm_i915_private_t *dev_priv = dev->dev_private; | |
730 | ||
731 | return READ_HWSP(dev_priv, I915_GEM_HWS_INDEX); | |
732 | } | |
733 | ||
734 | /** | |
735 | * This function clears the request list as sequence numbers are passed. | |
736 | */ | |
737 | void | |
738 | i915_gem_retire_requests(struct drm_device *dev) | |
739 | { | |
740 | drm_i915_private_t *dev_priv = dev->dev_private; | |
741 | uint32_t seqno; | |
742 | ||
743 | seqno = i915_get_gem_seqno(dev); | |
744 | ||
745 | while (!list_empty(&dev_priv->mm.request_list)) { | |
746 | struct drm_i915_gem_request *request; | |
747 | uint32_t retiring_seqno; | |
748 | ||
749 | request = list_first_entry(&dev_priv->mm.request_list, | |
750 | struct drm_i915_gem_request, | |
751 | list); | |
752 | retiring_seqno = request->seqno; | |
753 | ||
754 | if (i915_seqno_passed(seqno, retiring_seqno) || | |
755 | dev_priv->mm.wedged) { | |
756 | i915_gem_retire_request(dev, request); | |
757 | ||
758 | list_del(&request->list); | |
759 | drm_free(request, sizeof(*request), DRM_MEM_DRIVER); | |
760 | } else | |
761 | break; | |
762 | } | |
763 | } | |
764 | ||
765 | void | |
766 | i915_gem_retire_work_handler(struct work_struct *work) | |
767 | { | |
768 | drm_i915_private_t *dev_priv; | |
769 | struct drm_device *dev; | |
770 | ||
771 | dev_priv = container_of(work, drm_i915_private_t, | |
772 | mm.retire_work.work); | |
773 | dev = dev_priv->dev; | |
774 | ||
775 | mutex_lock(&dev->struct_mutex); | |
776 | i915_gem_retire_requests(dev); | |
6dbe2772 KP |
777 | if (!dev_priv->mm.suspended && |
778 | !list_empty(&dev_priv->mm.request_list)) | |
673a394b EA |
779 | schedule_delayed_work(&dev_priv->mm.retire_work, HZ); |
780 | mutex_unlock(&dev->struct_mutex); | |
781 | } | |
782 | ||
783 | /** | |
784 | * Waits for a sequence number to be signaled, and cleans up the | |
785 | * request and object lists appropriately for that event. | |
786 | */ | |
3043c60c | 787 | static int |
673a394b EA |
788 | i915_wait_request(struct drm_device *dev, uint32_t seqno) |
789 | { | |
790 | drm_i915_private_t *dev_priv = dev->dev_private; | |
791 | int ret = 0; | |
792 | ||
793 | BUG_ON(seqno == 0); | |
794 | ||
795 | if (!i915_seqno_passed(i915_get_gem_seqno(dev), seqno)) { | |
796 | dev_priv->mm.waiting_gem_seqno = seqno; | |
797 | i915_user_irq_get(dev); | |
798 | ret = wait_event_interruptible(dev_priv->irq_queue, | |
799 | i915_seqno_passed(i915_get_gem_seqno(dev), | |
800 | seqno) || | |
801 | dev_priv->mm.wedged); | |
802 | i915_user_irq_put(dev); | |
803 | dev_priv->mm.waiting_gem_seqno = 0; | |
804 | } | |
805 | if (dev_priv->mm.wedged) | |
806 | ret = -EIO; | |
807 | ||
808 | if (ret && ret != -ERESTARTSYS) | |
809 | DRM_ERROR("%s returns %d (awaiting %d at %d)\n", | |
810 | __func__, ret, seqno, i915_get_gem_seqno(dev)); | |
811 | ||
812 | /* Directly dispatch request retiring. While we have the work queue | |
813 | * to handle this, the waiter on a request often wants an associated | |
814 | * buffer to have made it to the inactive list, and we would need | |
815 | * a separate wait queue to handle that. | |
816 | */ | |
817 | if (ret == 0) | |
818 | i915_gem_retire_requests(dev); | |
819 | ||
820 | return ret; | |
821 | } | |
822 | ||
823 | static void | |
824 | i915_gem_flush(struct drm_device *dev, | |
825 | uint32_t invalidate_domains, | |
826 | uint32_t flush_domains) | |
827 | { | |
828 | drm_i915_private_t *dev_priv = dev->dev_private; | |
829 | uint32_t cmd; | |
830 | RING_LOCALS; | |
831 | ||
832 | #if WATCH_EXEC | |
833 | DRM_INFO("%s: invalidate %08x flush %08x\n", __func__, | |
834 | invalidate_domains, flush_domains); | |
835 | #endif | |
836 | ||
837 | if (flush_domains & I915_GEM_DOMAIN_CPU) | |
838 | drm_agp_chipset_flush(dev); | |
839 | ||
840 | if ((invalidate_domains | flush_domains) & ~(I915_GEM_DOMAIN_CPU | | |
841 | I915_GEM_DOMAIN_GTT)) { | |
842 | /* | |
843 | * read/write caches: | |
844 | * | |
845 | * I915_GEM_DOMAIN_RENDER is always invalidated, but is | |
846 | * only flushed if MI_NO_WRITE_FLUSH is unset. On 965, it is | |
847 | * also flushed at 2d versus 3d pipeline switches. | |
848 | * | |
849 | * read-only caches: | |
850 | * | |
851 | * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if | |
852 | * MI_READ_FLUSH is set, and is always flushed on 965. | |
853 | * | |
854 | * I915_GEM_DOMAIN_COMMAND may not exist? | |
855 | * | |
856 | * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is | |
857 | * invalidated when MI_EXE_FLUSH is set. | |
858 | * | |
859 | * I915_GEM_DOMAIN_VERTEX, which exists on 965, is | |
860 | * invalidated with every MI_FLUSH. | |
861 | * | |
862 | * TLBs: | |
863 | * | |
864 | * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND | |
865 | * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and | |
866 | * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER | |
867 | * are flushed at any MI_FLUSH. | |
868 | */ | |
869 | ||
870 | cmd = MI_FLUSH | MI_NO_WRITE_FLUSH; | |
871 | if ((invalidate_domains|flush_domains) & | |
872 | I915_GEM_DOMAIN_RENDER) | |
873 | cmd &= ~MI_NO_WRITE_FLUSH; | |
874 | if (!IS_I965G(dev)) { | |
875 | /* | |
876 | * On the 965, the sampler cache always gets flushed | |
877 | * and this bit is reserved. | |
878 | */ | |
879 | if (invalidate_domains & I915_GEM_DOMAIN_SAMPLER) | |
880 | cmd |= MI_READ_FLUSH; | |
881 | } | |
882 | if (invalidate_domains & I915_GEM_DOMAIN_INSTRUCTION) | |
883 | cmd |= MI_EXE_FLUSH; | |
884 | ||
885 | #if WATCH_EXEC | |
886 | DRM_INFO("%s: queue flush %08x to ring\n", __func__, cmd); | |
887 | #endif | |
888 | BEGIN_LP_RING(2); | |
889 | OUT_RING(cmd); | |
890 | OUT_RING(0); /* noop */ | |
891 | ADVANCE_LP_RING(); | |
892 | } | |
893 | } | |
894 | ||
895 | /** | |
896 | * Ensures that all rendering to the object has completed and the object is | |
897 | * safe to unbind from the GTT or access from the CPU. | |
898 | */ | |
899 | static int | |
900 | i915_gem_object_wait_rendering(struct drm_gem_object *obj) | |
901 | { | |
902 | struct drm_device *dev = obj->dev; | |
903 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
904 | int ret; | |
905 | ||
906 | /* If there are writes queued to the buffer, flush and | |
907 | * create a new seqno to wait for. | |
908 | */ | |
909 | if (obj->write_domain & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT)) { | |
ce44b0ea | 910 | uint32_t seqno, write_domain = obj->write_domain; |
673a394b EA |
911 | #if WATCH_BUF |
912 | DRM_INFO("%s: flushing object %p from write domain %08x\n", | |
913 | __func__, obj, write_domain); | |
914 | #endif | |
915 | i915_gem_flush(dev, 0, write_domain); | |
916 | ||
ce44b0ea EA |
917 | seqno = i915_add_request(dev, write_domain); |
918 | i915_gem_object_move_to_active(obj, seqno); | |
673a394b EA |
919 | #if WATCH_LRU |
920 | DRM_INFO("%s: flush moves to exec list %p\n", __func__, obj); | |
921 | #endif | |
922 | } | |
923 | ||
924 | /* If there is rendering queued on the buffer being evicted, wait for | |
925 | * it. | |
926 | */ | |
927 | if (obj_priv->active) { | |
928 | #if WATCH_BUF | |
929 | DRM_INFO("%s: object %p wait for seqno %08x\n", | |
930 | __func__, obj, obj_priv->last_rendering_seqno); | |
931 | #endif | |
932 | ret = i915_wait_request(dev, obj_priv->last_rendering_seqno); | |
933 | if (ret != 0) | |
934 | return ret; | |
935 | } | |
936 | ||
937 | return 0; | |
938 | } | |
939 | ||
940 | /** | |
941 | * Unbinds an object from the GTT aperture. | |
942 | */ | |
943 | static int | |
944 | i915_gem_object_unbind(struct drm_gem_object *obj) | |
945 | { | |
946 | struct drm_device *dev = obj->dev; | |
947 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
948 | int ret = 0; | |
949 | ||
950 | #if WATCH_BUF | |
951 | DRM_INFO("%s:%d %p\n", __func__, __LINE__, obj); | |
952 | DRM_INFO("gtt_space %p\n", obj_priv->gtt_space); | |
953 | #endif | |
954 | if (obj_priv->gtt_space == NULL) | |
955 | return 0; | |
956 | ||
957 | if (obj_priv->pin_count != 0) { | |
958 | DRM_ERROR("Attempting to unbind pinned buffer\n"); | |
959 | return -EINVAL; | |
960 | } | |
961 | ||
962 | /* Wait for any rendering to complete | |
963 | */ | |
964 | ret = i915_gem_object_wait_rendering(obj); | |
965 | if (ret) { | |
966 | DRM_ERROR("wait_rendering failed: %d\n", ret); | |
967 | return ret; | |
968 | } | |
969 | ||
970 | /* Move the object to the CPU domain to ensure that | |
971 | * any possible CPU writes while it's not in the GTT | |
972 | * are flushed when we go to remap it. This will | |
973 | * also ensure that all pending GPU writes are finished | |
974 | * before we unbind. | |
975 | */ | |
976 | ret = i915_gem_object_set_domain(obj, I915_GEM_DOMAIN_CPU, | |
977 | I915_GEM_DOMAIN_CPU); | |
978 | if (ret) { | |
979 | DRM_ERROR("set_domain failed: %d\n", ret); | |
980 | return ret; | |
981 | } | |
982 | ||
983 | if (obj_priv->agp_mem != NULL) { | |
984 | drm_unbind_agp(obj_priv->agp_mem); | |
985 | drm_free_agp(obj_priv->agp_mem, obj->size / PAGE_SIZE); | |
986 | obj_priv->agp_mem = NULL; | |
987 | } | |
988 | ||
989 | BUG_ON(obj_priv->active); | |
990 | ||
991 | i915_gem_object_free_page_list(obj); | |
992 | ||
993 | if (obj_priv->gtt_space) { | |
994 | atomic_dec(&dev->gtt_count); | |
995 | atomic_sub(obj->size, &dev->gtt_memory); | |
996 | ||
997 | drm_mm_put_block(obj_priv->gtt_space); | |
998 | obj_priv->gtt_space = NULL; | |
999 | } | |
1000 | ||
1001 | /* Remove ourselves from the LRU list if present. */ | |
1002 | if (!list_empty(&obj_priv->list)) | |
1003 | list_del_init(&obj_priv->list); | |
1004 | ||
1005 | return 0; | |
1006 | } | |
1007 | ||
1008 | static int | |
1009 | i915_gem_evict_something(struct drm_device *dev) | |
1010 | { | |
1011 | drm_i915_private_t *dev_priv = dev->dev_private; | |
1012 | struct drm_gem_object *obj; | |
1013 | struct drm_i915_gem_object *obj_priv; | |
1014 | int ret = 0; | |
1015 | ||
1016 | for (;;) { | |
1017 | /* If there's an inactive buffer available now, grab it | |
1018 | * and be done. | |
1019 | */ | |
1020 | if (!list_empty(&dev_priv->mm.inactive_list)) { | |
1021 | obj_priv = list_first_entry(&dev_priv->mm.inactive_list, | |
1022 | struct drm_i915_gem_object, | |
1023 | list); | |
1024 | obj = obj_priv->obj; | |
1025 | BUG_ON(obj_priv->pin_count != 0); | |
1026 | #if WATCH_LRU | |
1027 | DRM_INFO("%s: evicting %p\n", __func__, obj); | |
1028 | #endif | |
1029 | BUG_ON(obj_priv->active); | |
1030 | ||
1031 | /* Wait on the rendering and unbind the buffer. */ | |
1032 | ret = i915_gem_object_unbind(obj); | |
1033 | break; | |
1034 | } | |
1035 | ||
1036 | /* If we didn't get anything, but the ring is still processing | |
1037 | * things, wait for one of those things to finish and hopefully | |
1038 | * leave us a buffer to evict. | |
1039 | */ | |
1040 | if (!list_empty(&dev_priv->mm.request_list)) { | |
1041 | struct drm_i915_gem_request *request; | |
1042 | ||
1043 | request = list_first_entry(&dev_priv->mm.request_list, | |
1044 | struct drm_i915_gem_request, | |
1045 | list); | |
1046 | ||
1047 | ret = i915_wait_request(dev, request->seqno); | |
1048 | if (ret) | |
1049 | break; | |
1050 | ||
1051 | /* if waiting caused an object to become inactive, | |
1052 | * then loop around and wait for it. Otherwise, we | |
1053 | * assume that waiting freed and unbound something, | |
1054 | * so there should now be some space in the GTT | |
1055 | */ | |
1056 | if (!list_empty(&dev_priv->mm.inactive_list)) | |
1057 | continue; | |
1058 | break; | |
1059 | } | |
1060 | ||
1061 | /* If we didn't have anything on the request list but there | |
1062 | * are buffers awaiting a flush, emit one and try again. | |
1063 | * When we wait on it, those buffers waiting for that flush | |
1064 | * will get moved to inactive. | |
1065 | */ | |
1066 | if (!list_empty(&dev_priv->mm.flushing_list)) { | |
1067 | obj_priv = list_first_entry(&dev_priv->mm.flushing_list, | |
1068 | struct drm_i915_gem_object, | |
1069 | list); | |
1070 | obj = obj_priv->obj; | |
1071 | ||
1072 | i915_gem_flush(dev, | |
1073 | obj->write_domain, | |
1074 | obj->write_domain); | |
1075 | i915_add_request(dev, obj->write_domain); | |
1076 | ||
1077 | obj = NULL; | |
1078 | continue; | |
1079 | } | |
1080 | ||
1081 | DRM_ERROR("inactive empty %d request empty %d " | |
1082 | "flushing empty %d\n", | |
1083 | list_empty(&dev_priv->mm.inactive_list), | |
1084 | list_empty(&dev_priv->mm.request_list), | |
1085 | list_empty(&dev_priv->mm.flushing_list)); | |
1086 | /* If we didn't do any of the above, there's nothing to be done | |
1087 | * and we just can't fit it in. | |
1088 | */ | |
1089 | return -ENOMEM; | |
1090 | } | |
1091 | return ret; | |
1092 | } | |
1093 | ||
1094 | static int | |
1095 | i915_gem_object_get_page_list(struct drm_gem_object *obj) | |
1096 | { | |
1097 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
1098 | int page_count, i; | |
1099 | struct address_space *mapping; | |
1100 | struct inode *inode; | |
1101 | struct page *page; | |
1102 | int ret; | |
1103 | ||
1104 | if (obj_priv->page_list) | |
1105 | return 0; | |
1106 | ||
1107 | /* Get the list of pages out of our struct file. They'll be pinned | |
1108 | * at this point until we release them. | |
1109 | */ | |
1110 | page_count = obj->size / PAGE_SIZE; | |
1111 | BUG_ON(obj_priv->page_list != NULL); | |
1112 | obj_priv->page_list = drm_calloc(page_count, sizeof(struct page *), | |
1113 | DRM_MEM_DRIVER); | |
1114 | if (obj_priv->page_list == NULL) { | |
1115 | DRM_ERROR("Faled to allocate page list\n"); | |
1116 | return -ENOMEM; | |
1117 | } | |
1118 | ||
1119 | inode = obj->filp->f_path.dentry->d_inode; | |
1120 | mapping = inode->i_mapping; | |
1121 | for (i = 0; i < page_count; i++) { | |
1122 | page = read_mapping_page(mapping, i, NULL); | |
1123 | if (IS_ERR(page)) { | |
1124 | ret = PTR_ERR(page); | |
1125 | DRM_ERROR("read_mapping_page failed: %d\n", ret); | |
1126 | i915_gem_object_free_page_list(obj); | |
1127 | return ret; | |
1128 | } | |
1129 | obj_priv->page_list[i] = page; | |
1130 | } | |
1131 | return 0; | |
1132 | } | |
1133 | ||
1134 | /** | |
1135 | * Finds free space in the GTT aperture and binds the object there. | |
1136 | */ | |
1137 | static int | |
1138 | i915_gem_object_bind_to_gtt(struct drm_gem_object *obj, unsigned alignment) | |
1139 | { | |
1140 | struct drm_device *dev = obj->dev; | |
1141 | drm_i915_private_t *dev_priv = dev->dev_private; | |
1142 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
1143 | struct drm_mm_node *free_space; | |
1144 | int page_count, ret; | |
1145 | ||
1146 | if (alignment == 0) | |
1147 | alignment = PAGE_SIZE; | |
1148 | if (alignment & (PAGE_SIZE - 1)) { | |
1149 | DRM_ERROR("Invalid object alignment requested %u\n", alignment); | |
1150 | return -EINVAL; | |
1151 | } | |
1152 | ||
1153 | search_free: | |
1154 | free_space = drm_mm_search_free(&dev_priv->mm.gtt_space, | |
1155 | obj->size, alignment, 0); | |
1156 | if (free_space != NULL) { | |
1157 | obj_priv->gtt_space = drm_mm_get_block(free_space, obj->size, | |
1158 | alignment); | |
1159 | if (obj_priv->gtt_space != NULL) { | |
1160 | obj_priv->gtt_space->private = obj; | |
1161 | obj_priv->gtt_offset = obj_priv->gtt_space->start; | |
1162 | } | |
1163 | } | |
1164 | if (obj_priv->gtt_space == NULL) { | |
1165 | /* If the gtt is empty and we're still having trouble | |
1166 | * fitting our object in, we're out of memory. | |
1167 | */ | |
1168 | #if WATCH_LRU | |
1169 | DRM_INFO("%s: GTT full, evicting something\n", __func__); | |
1170 | #endif | |
1171 | if (list_empty(&dev_priv->mm.inactive_list) && | |
1172 | list_empty(&dev_priv->mm.flushing_list) && | |
1173 | list_empty(&dev_priv->mm.active_list)) { | |
1174 | DRM_ERROR("GTT full, but LRU list empty\n"); | |
1175 | return -ENOMEM; | |
1176 | } | |
1177 | ||
1178 | ret = i915_gem_evict_something(dev); | |
1179 | if (ret != 0) { | |
1180 | DRM_ERROR("Failed to evict a buffer %d\n", ret); | |
1181 | return ret; | |
1182 | } | |
1183 | goto search_free; | |
1184 | } | |
1185 | ||
1186 | #if WATCH_BUF | |
1187 | DRM_INFO("Binding object of size %d at 0x%08x\n", | |
1188 | obj->size, obj_priv->gtt_offset); | |
1189 | #endif | |
1190 | ret = i915_gem_object_get_page_list(obj); | |
1191 | if (ret) { | |
1192 | drm_mm_put_block(obj_priv->gtt_space); | |
1193 | obj_priv->gtt_space = NULL; | |
1194 | return ret; | |
1195 | } | |
1196 | ||
1197 | page_count = obj->size / PAGE_SIZE; | |
1198 | /* Create an AGP memory structure pointing at our pages, and bind it | |
1199 | * into the GTT. | |
1200 | */ | |
1201 | obj_priv->agp_mem = drm_agp_bind_pages(dev, | |
1202 | obj_priv->page_list, | |
1203 | page_count, | |
ba1eb1d8 KP |
1204 | obj_priv->gtt_offset, |
1205 | obj_priv->agp_type); | |
673a394b EA |
1206 | if (obj_priv->agp_mem == NULL) { |
1207 | i915_gem_object_free_page_list(obj); | |
1208 | drm_mm_put_block(obj_priv->gtt_space); | |
1209 | obj_priv->gtt_space = NULL; | |
1210 | return -ENOMEM; | |
1211 | } | |
1212 | atomic_inc(&dev->gtt_count); | |
1213 | atomic_add(obj->size, &dev->gtt_memory); | |
1214 | ||
1215 | /* Assert that the object is not currently in any GPU domain. As it | |
1216 | * wasn't in the GTT, there shouldn't be any way it could have been in | |
1217 | * a GPU cache | |
1218 | */ | |
1219 | BUG_ON(obj->read_domains & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT)); | |
1220 | BUG_ON(obj->write_domain & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT)); | |
1221 | ||
1222 | return 0; | |
1223 | } | |
1224 | ||
1225 | void | |
1226 | i915_gem_clflush_object(struct drm_gem_object *obj) | |
1227 | { | |
1228 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
1229 | ||
1230 | /* If we don't have a page list set up, then we're not pinned | |
1231 | * to GPU, and we can ignore the cache flush because it'll happen | |
1232 | * again at bind time. | |
1233 | */ | |
1234 | if (obj_priv->page_list == NULL) | |
1235 | return; | |
1236 | ||
1237 | drm_clflush_pages(obj_priv->page_list, obj->size / PAGE_SIZE); | |
1238 | } | |
1239 | ||
1240 | /* | |
1241 | * Set the next domain for the specified object. This | |
1242 | * may not actually perform the necessary flushing/invaliding though, | |
1243 | * as that may want to be batched with other set_domain operations | |
1244 | * | |
1245 | * This is (we hope) the only really tricky part of gem. The goal | |
1246 | * is fairly simple -- track which caches hold bits of the object | |
1247 | * and make sure they remain coherent. A few concrete examples may | |
1248 | * help to explain how it works. For shorthand, we use the notation | |
1249 | * (read_domains, write_domain), e.g. (CPU, CPU) to indicate the | |
1250 | * a pair of read and write domain masks. | |
1251 | * | |
1252 | * Case 1: the batch buffer | |
1253 | * | |
1254 | * 1. Allocated | |
1255 | * 2. Written by CPU | |
1256 | * 3. Mapped to GTT | |
1257 | * 4. Read by GPU | |
1258 | * 5. Unmapped from GTT | |
1259 | * 6. Freed | |
1260 | * | |
1261 | * Let's take these a step at a time | |
1262 | * | |
1263 | * 1. Allocated | |
1264 | * Pages allocated from the kernel may still have | |
1265 | * cache contents, so we set them to (CPU, CPU) always. | |
1266 | * 2. Written by CPU (using pwrite) | |
1267 | * The pwrite function calls set_domain (CPU, CPU) and | |
1268 | * this function does nothing (as nothing changes) | |
1269 | * 3. Mapped by GTT | |
1270 | * This function asserts that the object is not | |
1271 | * currently in any GPU-based read or write domains | |
1272 | * 4. Read by GPU | |
1273 | * i915_gem_execbuffer calls set_domain (COMMAND, 0). | |
1274 | * As write_domain is zero, this function adds in the | |
1275 | * current read domains (CPU+COMMAND, 0). | |
1276 | * flush_domains is set to CPU. | |
1277 | * invalidate_domains is set to COMMAND | |
1278 | * clflush is run to get data out of the CPU caches | |
1279 | * then i915_dev_set_domain calls i915_gem_flush to | |
1280 | * emit an MI_FLUSH and drm_agp_chipset_flush | |
1281 | * 5. Unmapped from GTT | |
1282 | * i915_gem_object_unbind calls set_domain (CPU, CPU) | |
1283 | * flush_domains and invalidate_domains end up both zero | |
1284 | * so no flushing/invalidating happens | |
1285 | * 6. Freed | |
1286 | * yay, done | |
1287 | * | |
1288 | * Case 2: The shared render buffer | |
1289 | * | |
1290 | * 1. Allocated | |
1291 | * 2. Mapped to GTT | |
1292 | * 3. Read/written by GPU | |
1293 | * 4. set_domain to (CPU,CPU) | |
1294 | * 5. Read/written by CPU | |
1295 | * 6. Read/written by GPU | |
1296 | * | |
1297 | * 1. Allocated | |
1298 | * Same as last example, (CPU, CPU) | |
1299 | * 2. Mapped to GTT | |
1300 | * Nothing changes (assertions find that it is not in the GPU) | |
1301 | * 3. Read/written by GPU | |
1302 | * execbuffer calls set_domain (RENDER, RENDER) | |
1303 | * flush_domains gets CPU | |
1304 | * invalidate_domains gets GPU | |
1305 | * clflush (obj) | |
1306 | * MI_FLUSH and drm_agp_chipset_flush | |
1307 | * 4. set_domain (CPU, CPU) | |
1308 | * flush_domains gets GPU | |
1309 | * invalidate_domains gets CPU | |
1310 | * wait_rendering (obj) to make sure all drawing is complete. | |
1311 | * This will include an MI_FLUSH to get the data from GPU | |
1312 | * to memory | |
1313 | * clflush (obj) to invalidate the CPU cache | |
1314 | * Another MI_FLUSH in i915_gem_flush (eliminate this somehow?) | |
1315 | * 5. Read/written by CPU | |
1316 | * cache lines are loaded and dirtied | |
1317 | * 6. Read written by GPU | |
1318 | * Same as last GPU access | |
1319 | * | |
1320 | * Case 3: The constant buffer | |
1321 | * | |
1322 | * 1. Allocated | |
1323 | * 2. Written by CPU | |
1324 | * 3. Read by GPU | |
1325 | * 4. Updated (written) by CPU again | |
1326 | * 5. Read by GPU | |
1327 | * | |
1328 | * 1. Allocated | |
1329 | * (CPU, CPU) | |
1330 | * 2. Written by CPU | |
1331 | * (CPU, CPU) | |
1332 | * 3. Read by GPU | |
1333 | * (CPU+RENDER, 0) | |
1334 | * flush_domains = CPU | |
1335 | * invalidate_domains = RENDER | |
1336 | * clflush (obj) | |
1337 | * MI_FLUSH | |
1338 | * drm_agp_chipset_flush | |
1339 | * 4. Updated (written) by CPU again | |
1340 | * (CPU, CPU) | |
1341 | * flush_domains = 0 (no previous write domain) | |
1342 | * invalidate_domains = 0 (no new read domains) | |
1343 | * 5. Read by GPU | |
1344 | * (CPU+RENDER, 0) | |
1345 | * flush_domains = CPU | |
1346 | * invalidate_domains = RENDER | |
1347 | * clflush (obj) | |
1348 | * MI_FLUSH | |
1349 | * drm_agp_chipset_flush | |
1350 | */ | |
1351 | static int | |
1352 | i915_gem_object_set_domain(struct drm_gem_object *obj, | |
1353 | uint32_t read_domains, | |
1354 | uint32_t write_domain) | |
1355 | { | |
1356 | struct drm_device *dev = obj->dev; | |
1357 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
1358 | uint32_t invalidate_domains = 0; | |
1359 | uint32_t flush_domains = 0; | |
1360 | int ret; | |
1361 | ||
1362 | #if WATCH_BUF | |
1363 | DRM_INFO("%s: object %p read %08x -> %08x write %08x -> %08x\n", | |
1364 | __func__, obj, | |
1365 | obj->read_domains, read_domains, | |
1366 | obj->write_domain, write_domain); | |
1367 | #endif | |
1368 | /* | |
1369 | * If the object isn't moving to a new write domain, | |
1370 | * let the object stay in multiple read domains | |
1371 | */ | |
1372 | if (write_domain == 0) | |
1373 | read_domains |= obj->read_domains; | |
1374 | else | |
1375 | obj_priv->dirty = 1; | |
1376 | ||
1377 | /* | |
1378 | * Flush the current write domain if | |
1379 | * the new read domains don't match. Invalidate | |
1380 | * any read domains which differ from the old | |
1381 | * write domain | |
1382 | */ | |
1383 | if (obj->write_domain && obj->write_domain != read_domains) { | |
1384 | flush_domains |= obj->write_domain; | |
1385 | invalidate_domains |= read_domains & ~obj->write_domain; | |
1386 | } | |
1387 | /* | |
1388 | * Invalidate any read caches which may have | |
1389 | * stale data. That is, any new read domains. | |
1390 | */ | |
1391 | invalidate_domains |= read_domains & ~obj->read_domains; | |
1392 | if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_CPU) { | |
1393 | #if WATCH_BUF | |
1394 | DRM_INFO("%s: CPU domain flush %08x invalidate %08x\n", | |
1395 | __func__, flush_domains, invalidate_domains); | |
1396 | #endif | |
1397 | /* | |
1398 | * If we're invaliding the CPU cache and flushing a GPU cache, | |
1399 | * then pause for rendering so that the GPU caches will be | |
1400 | * flushed before the cpu cache is invalidated | |
1401 | */ | |
1402 | if ((invalidate_domains & I915_GEM_DOMAIN_CPU) && | |
1403 | (flush_domains & ~(I915_GEM_DOMAIN_CPU | | |
1404 | I915_GEM_DOMAIN_GTT))) { | |
1405 | ret = i915_gem_object_wait_rendering(obj); | |
1406 | if (ret) | |
1407 | return ret; | |
1408 | } | |
1409 | i915_gem_clflush_object(obj); | |
1410 | } | |
1411 | ||
1412 | if ((write_domain | flush_domains) != 0) | |
1413 | obj->write_domain = write_domain; | |
1414 | ||
1415 | /* If we're invalidating the CPU domain, clear the per-page CPU | |
1416 | * domain list as well. | |
1417 | */ | |
1418 | if (obj_priv->page_cpu_valid != NULL && | |
1419 | (write_domain != 0 || | |
1420 | read_domains & I915_GEM_DOMAIN_CPU)) { | |
1421 | drm_free(obj_priv->page_cpu_valid, obj->size / PAGE_SIZE, | |
1422 | DRM_MEM_DRIVER); | |
1423 | obj_priv->page_cpu_valid = NULL; | |
1424 | } | |
1425 | obj->read_domains = read_domains; | |
1426 | ||
1427 | dev->invalidate_domains |= invalidate_domains; | |
1428 | dev->flush_domains |= flush_domains; | |
1429 | #if WATCH_BUF | |
1430 | DRM_INFO("%s: read %08x write %08x invalidate %08x flush %08x\n", | |
1431 | __func__, | |
1432 | obj->read_domains, obj->write_domain, | |
1433 | dev->invalidate_domains, dev->flush_domains); | |
1434 | #endif | |
1435 | return 0; | |
1436 | } | |
1437 | ||
1438 | /** | |
1439 | * Set the read/write domain on a range of the object. | |
1440 | * | |
1441 | * Currently only implemented for CPU reads, otherwise drops to normal | |
1442 | * i915_gem_object_set_domain(). | |
1443 | */ | |
1444 | static int | |
1445 | i915_gem_object_set_domain_range(struct drm_gem_object *obj, | |
1446 | uint64_t offset, | |
1447 | uint64_t size, | |
1448 | uint32_t read_domains, | |
1449 | uint32_t write_domain) | |
1450 | { | |
1451 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
1452 | int ret, i; | |
1453 | ||
1454 | if (obj->read_domains & I915_GEM_DOMAIN_CPU) | |
1455 | return 0; | |
1456 | ||
1457 | if (read_domains != I915_GEM_DOMAIN_CPU || | |
1458 | write_domain != 0) | |
1459 | return i915_gem_object_set_domain(obj, | |
1460 | read_domains, write_domain); | |
1461 | ||
1462 | /* Wait on any GPU rendering to the object to be flushed. */ | |
6a47baa6 OT |
1463 | ret = i915_gem_object_wait_rendering(obj); |
1464 | if (ret) | |
1465 | return ret; | |
673a394b EA |
1466 | |
1467 | if (obj_priv->page_cpu_valid == NULL) { | |
1468 | obj_priv->page_cpu_valid = drm_calloc(1, obj->size / PAGE_SIZE, | |
1469 | DRM_MEM_DRIVER); | |
1470 | } | |
1471 | ||
1472 | /* Flush the cache on any pages that are still invalid from the CPU's | |
1473 | * perspective. | |
1474 | */ | |
1475 | for (i = offset / PAGE_SIZE; i <= (offset + size - 1) / PAGE_SIZE; i++) { | |
1476 | if (obj_priv->page_cpu_valid[i]) | |
1477 | continue; | |
1478 | ||
1479 | drm_clflush_pages(obj_priv->page_list + i, 1); | |
1480 | ||
1481 | obj_priv->page_cpu_valid[i] = 1; | |
1482 | } | |
1483 | ||
1484 | return 0; | |
1485 | } | |
1486 | ||
1487 | /** | |
1488 | * Once all of the objects have been set in the proper domain, | |
1489 | * perform the necessary flush and invalidate operations. | |
1490 | * | |
1491 | * Returns the write domains flushed, for use in flush tracking. | |
1492 | */ | |
1493 | static uint32_t | |
1494 | i915_gem_dev_set_domain(struct drm_device *dev) | |
1495 | { | |
1496 | uint32_t flush_domains = dev->flush_domains; | |
1497 | ||
1498 | /* | |
1499 | * Now that all the buffers are synced to the proper domains, | |
1500 | * flush and invalidate the collected domains | |
1501 | */ | |
1502 | if (dev->invalidate_domains | dev->flush_domains) { | |
1503 | #if WATCH_EXEC | |
1504 | DRM_INFO("%s: invalidate_domains %08x flush_domains %08x\n", | |
1505 | __func__, | |
1506 | dev->invalidate_domains, | |
1507 | dev->flush_domains); | |
1508 | #endif | |
1509 | i915_gem_flush(dev, | |
1510 | dev->invalidate_domains, | |
1511 | dev->flush_domains); | |
1512 | dev->invalidate_domains = 0; | |
1513 | dev->flush_domains = 0; | |
1514 | } | |
1515 | ||
1516 | return flush_domains; | |
1517 | } | |
1518 | ||
1519 | /** | |
1520 | * Pin an object to the GTT and evaluate the relocations landing in it. | |
1521 | */ | |
1522 | static int | |
1523 | i915_gem_object_pin_and_relocate(struct drm_gem_object *obj, | |
1524 | struct drm_file *file_priv, | |
1525 | struct drm_i915_gem_exec_object *entry) | |
1526 | { | |
1527 | struct drm_device *dev = obj->dev; | |
0839ccb8 | 1528 | drm_i915_private_t *dev_priv = dev->dev_private; |
673a394b EA |
1529 | struct drm_i915_gem_relocation_entry reloc; |
1530 | struct drm_i915_gem_relocation_entry __user *relocs; | |
1531 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
1532 | int i, ret; | |
0839ccb8 | 1533 | void __iomem *reloc_page; |
673a394b EA |
1534 | |
1535 | /* Choose the GTT offset for our buffer and put it there. */ | |
1536 | ret = i915_gem_object_pin(obj, (uint32_t) entry->alignment); | |
1537 | if (ret) | |
1538 | return ret; | |
1539 | ||
1540 | entry->offset = obj_priv->gtt_offset; | |
1541 | ||
1542 | relocs = (struct drm_i915_gem_relocation_entry __user *) | |
1543 | (uintptr_t) entry->relocs_ptr; | |
1544 | /* Apply the relocations, using the GTT aperture to avoid cache | |
1545 | * flushing requirements. | |
1546 | */ | |
1547 | for (i = 0; i < entry->relocation_count; i++) { | |
1548 | struct drm_gem_object *target_obj; | |
1549 | struct drm_i915_gem_object *target_obj_priv; | |
3043c60c EA |
1550 | uint32_t reloc_val, reloc_offset; |
1551 | uint32_t __iomem *reloc_entry; | |
673a394b EA |
1552 | |
1553 | ret = copy_from_user(&reloc, relocs + i, sizeof(reloc)); | |
1554 | if (ret != 0) { | |
1555 | i915_gem_object_unpin(obj); | |
1556 | return ret; | |
1557 | } | |
1558 | ||
1559 | target_obj = drm_gem_object_lookup(obj->dev, file_priv, | |
1560 | reloc.target_handle); | |
1561 | if (target_obj == NULL) { | |
1562 | i915_gem_object_unpin(obj); | |
1563 | return -EBADF; | |
1564 | } | |
1565 | target_obj_priv = target_obj->driver_private; | |
1566 | ||
1567 | /* The target buffer should have appeared before us in the | |
1568 | * exec_object list, so it should have a GTT space bound by now. | |
1569 | */ | |
1570 | if (target_obj_priv->gtt_space == NULL) { | |
1571 | DRM_ERROR("No GTT space found for object %d\n", | |
1572 | reloc.target_handle); | |
1573 | drm_gem_object_unreference(target_obj); | |
1574 | i915_gem_object_unpin(obj); | |
1575 | return -EINVAL; | |
1576 | } | |
1577 | ||
1578 | if (reloc.offset > obj->size - 4) { | |
1579 | DRM_ERROR("Relocation beyond object bounds: " | |
1580 | "obj %p target %d offset %d size %d.\n", | |
1581 | obj, reloc.target_handle, | |
1582 | (int) reloc.offset, (int) obj->size); | |
1583 | drm_gem_object_unreference(target_obj); | |
1584 | i915_gem_object_unpin(obj); | |
1585 | return -EINVAL; | |
1586 | } | |
1587 | if (reloc.offset & 3) { | |
1588 | DRM_ERROR("Relocation not 4-byte aligned: " | |
1589 | "obj %p target %d offset %d.\n", | |
1590 | obj, reloc.target_handle, | |
1591 | (int) reloc.offset); | |
1592 | drm_gem_object_unreference(target_obj); | |
1593 | i915_gem_object_unpin(obj); | |
1594 | return -EINVAL; | |
1595 | } | |
1596 | ||
1597 | if (reloc.write_domain && target_obj->pending_write_domain && | |
1598 | reloc.write_domain != target_obj->pending_write_domain) { | |
1599 | DRM_ERROR("Write domain conflict: " | |
1600 | "obj %p target %d offset %d " | |
1601 | "new %08x old %08x\n", | |
1602 | obj, reloc.target_handle, | |
1603 | (int) reloc.offset, | |
1604 | reloc.write_domain, | |
1605 | target_obj->pending_write_domain); | |
1606 | drm_gem_object_unreference(target_obj); | |
1607 | i915_gem_object_unpin(obj); | |
1608 | return -EINVAL; | |
1609 | } | |
1610 | ||
1611 | #if WATCH_RELOC | |
1612 | DRM_INFO("%s: obj %p offset %08x target %d " | |
1613 | "read %08x write %08x gtt %08x " | |
1614 | "presumed %08x delta %08x\n", | |
1615 | __func__, | |
1616 | obj, | |
1617 | (int) reloc.offset, | |
1618 | (int) reloc.target_handle, | |
1619 | (int) reloc.read_domains, | |
1620 | (int) reloc.write_domain, | |
1621 | (int) target_obj_priv->gtt_offset, | |
1622 | (int) reloc.presumed_offset, | |
1623 | reloc.delta); | |
1624 | #endif | |
1625 | ||
1626 | target_obj->pending_read_domains |= reloc.read_domains; | |
1627 | target_obj->pending_write_domain |= reloc.write_domain; | |
1628 | ||
1629 | /* If the relocation already has the right value in it, no | |
1630 | * more work needs to be done. | |
1631 | */ | |
1632 | if (target_obj_priv->gtt_offset == reloc.presumed_offset) { | |
1633 | drm_gem_object_unreference(target_obj); | |
1634 | continue; | |
1635 | } | |
1636 | ||
1637 | /* Now that we're going to actually write some data in, | |
1638 | * make sure that any rendering using this buffer's contents | |
1639 | * is completed. | |
1640 | */ | |
1641 | i915_gem_object_wait_rendering(obj); | |
1642 | ||
1643 | /* As we're writing through the gtt, flush | |
1644 | * any CPU writes before we write the relocations | |
1645 | */ | |
1646 | if (obj->write_domain & I915_GEM_DOMAIN_CPU) { | |
1647 | i915_gem_clflush_object(obj); | |
1648 | drm_agp_chipset_flush(dev); | |
1649 | obj->write_domain = 0; | |
1650 | } | |
1651 | ||
1652 | /* Map the page containing the relocation we're going to | |
1653 | * perform. | |
1654 | */ | |
1655 | reloc_offset = obj_priv->gtt_offset + reloc.offset; | |
0839ccb8 KP |
1656 | reloc_page = io_mapping_map_atomic_wc(dev_priv->mm.gtt_mapping, |
1657 | (reloc_offset & | |
1658 | ~(PAGE_SIZE - 1))); | |
3043c60c | 1659 | reloc_entry = (uint32_t __iomem *)(reloc_page + |
0839ccb8 | 1660 | (reloc_offset & (PAGE_SIZE - 1))); |
673a394b EA |
1661 | reloc_val = target_obj_priv->gtt_offset + reloc.delta; |
1662 | ||
1663 | #if WATCH_BUF | |
1664 | DRM_INFO("Applied relocation: %p@0x%08x %08x -> %08x\n", | |
1665 | obj, (unsigned int) reloc.offset, | |
1666 | readl(reloc_entry), reloc_val); | |
1667 | #endif | |
1668 | writel(reloc_val, reloc_entry); | |
0839ccb8 | 1669 | io_mapping_unmap_atomic(reloc_page); |
673a394b EA |
1670 | |
1671 | /* Write the updated presumed offset for this entry back out | |
1672 | * to the user. | |
1673 | */ | |
1674 | reloc.presumed_offset = target_obj_priv->gtt_offset; | |
1675 | ret = copy_to_user(relocs + i, &reloc, sizeof(reloc)); | |
1676 | if (ret != 0) { | |
1677 | drm_gem_object_unreference(target_obj); | |
1678 | i915_gem_object_unpin(obj); | |
1679 | return ret; | |
1680 | } | |
1681 | ||
1682 | drm_gem_object_unreference(target_obj); | |
1683 | } | |
1684 | ||
673a394b EA |
1685 | #if WATCH_BUF |
1686 | if (0) | |
1687 | i915_gem_dump_object(obj, 128, __func__, ~0); | |
1688 | #endif | |
1689 | return 0; | |
1690 | } | |
1691 | ||
1692 | /** Dispatch a batchbuffer to the ring | |
1693 | */ | |
1694 | static int | |
1695 | i915_dispatch_gem_execbuffer(struct drm_device *dev, | |
1696 | struct drm_i915_gem_execbuffer *exec, | |
1697 | uint64_t exec_offset) | |
1698 | { | |
1699 | drm_i915_private_t *dev_priv = dev->dev_private; | |
1700 | struct drm_clip_rect __user *boxes = (struct drm_clip_rect __user *) | |
1701 | (uintptr_t) exec->cliprects_ptr; | |
1702 | int nbox = exec->num_cliprects; | |
1703 | int i = 0, count; | |
1704 | uint32_t exec_start, exec_len; | |
1705 | RING_LOCALS; | |
1706 | ||
1707 | exec_start = (uint32_t) exec_offset + exec->batch_start_offset; | |
1708 | exec_len = (uint32_t) exec->batch_len; | |
1709 | ||
1710 | if ((exec_start | exec_len) & 0x7) { | |
1711 | DRM_ERROR("alignment\n"); | |
1712 | return -EINVAL; | |
1713 | } | |
1714 | ||
1715 | if (!exec_start) | |
1716 | return -EINVAL; | |
1717 | ||
1718 | count = nbox ? nbox : 1; | |
1719 | ||
1720 | for (i = 0; i < count; i++) { | |
1721 | if (i < nbox) { | |
1722 | int ret = i915_emit_box(dev, boxes, i, | |
1723 | exec->DR1, exec->DR4); | |
1724 | if (ret) | |
1725 | return ret; | |
1726 | } | |
1727 | ||
1728 | if (IS_I830(dev) || IS_845G(dev)) { | |
1729 | BEGIN_LP_RING(4); | |
1730 | OUT_RING(MI_BATCH_BUFFER); | |
1731 | OUT_RING(exec_start | MI_BATCH_NON_SECURE); | |
1732 | OUT_RING(exec_start + exec_len - 4); | |
1733 | OUT_RING(0); | |
1734 | ADVANCE_LP_RING(); | |
1735 | } else { | |
1736 | BEGIN_LP_RING(2); | |
1737 | if (IS_I965G(dev)) { | |
1738 | OUT_RING(MI_BATCH_BUFFER_START | | |
1739 | (2 << 6) | | |
1740 | MI_BATCH_NON_SECURE_I965); | |
1741 | OUT_RING(exec_start); | |
1742 | } else { | |
1743 | OUT_RING(MI_BATCH_BUFFER_START | | |
1744 | (2 << 6)); | |
1745 | OUT_RING(exec_start | MI_BATCH_NON_SECURE); | |
1746 | } | |
1747 | ADVANCE_LP_RING(); | |
1748 | } | |
1749 | } | |
1750 | ||
1751 | /* XXX breadcrumb */ | |
1752 | return 0; | |
1753 | } | |
1754 | ||
1755 | /* Throttle our rendering by waiting until the ring has completed our requests | |
1756 | * emitted over 20 msec ago. | |
1757 | * | |
1758 | * This should get us reasonable parallelism between CPU and GPU but also | |
1759 | * relatively low latency when blocking on a particular request to finish. | |
1760 | */ | |
1761 | static int | |
1762 | i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file_priv) | |
1763 | { | |
1764 | struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv; | |
1765 | int ret = 0; | |
1766 | uint32_t seqno; | |
1767 | ||
1768 | mutex_lock(&dev->struct_mutex); | |
1769 | seqno = i915_file_priv->mm.last_gem_throttle_seqno; | |
1770 | i915_file_priv->mm.last_gem_throttle_seqno = | |
1771 | i915_file_priv->mm.last_gem_seqno; | |
1772 | if (seqno) | |
1773 | ret = i915_wait_request(dev, seqno); | |
1774 | mutex_unlock(&dev->struct_mutex); | |
1775 | return ret; | |
1776 | } | |
1777 | ||
1778 | int | |
1779 | i915_gem_execbuffer(struct drm_device *dev, void *data, | |
1780 | struct drm_file *file_priv) | |
1781 | { | |
1782 | drm_i915_private_t *dev_priv = dev->dev_private; | |
1783 | struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv; | |
1784 | struct drm_i915_gem_execbuffer *args = data; | |
1785 | struct drm_i915_gem_exec_object *exec_list = NULL; | |
1786 | struct drm_gem_object **object_list = NULL; | |
1787 | struct drm_gem_object *batch_obj; | |
1788 | int ret, i, pinned = 0; | |
1789 | uint64_t exec_offset; | |
1790 | uint32_t seqno, flush_domains; | |
1791 | ||
1792 | #if WATCH_EXEC | |
1793 | DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n", | |
1794 | (int) args->buffers_ptr, args->buffer_count, args->batch_len); | |
1795 | #endif | |
1796 | ||
4f481ed2 EA |
1797 | if (args->buffer_count < 1) { |
1798 | DRM_ERROR("execbuf with %d buffers\n", args->buffer_count); | |
1799 | return -EINVAL; | |
1800 | } | |
673a394b EA |
1801 | /* Copy in the exec list from userland */ |
1802 | exec_list = drm_calloc(sizeof(*exec_list), args->buffer_count, | |
1803 | DRM_MEM_DRIVER); | |
1804 | object_list = drm_calloc(sizeof(*object_list), args->buffer_count, | |
1805 | DRM_MEM_DRIVER); | |
1806 | if (exec_list == NULL || object_list == NULL) { | |
1807 | DRM_ERROR("Failed to allocate exec or object list " | |
1808 | "for %d buffers\n", | |
1809 | args->buffer_count); | |
1810 | ret = -ENOMEM; | |
1811 | goto pre_mutex_err; | |
1812 | } | |
1813 | ret = copy_from_user(exec_list, | |
1814 | (struct drm_i915_relocation_entry __user *) | |
1815 | (uintptr_t) args->buffers_ptr, | |
1816 | sizeof(*exec_list) * args->buffer_count); | |
1817 | if (ret != 0) { | |
1818 | DRM_ERROR("copy %d exec entries failed %d\n", | |
1819 | args->buffer_count, ret); | |
1820 | goto pre_mutex_err; | |
1821 | } | |
1822 | ||
1823 | mutex_lock(&dev->struct_mutex); | |
1824 | ||
1825 | i915_verify_inactive(dev, __FILE__, __LINE__); | |
1826 | ||
1827 | if (dev_priv->mm.wedged) { | |
1828 | DRM_ERROR("Execbuf while wedged\n"); | |
1829 | mutex_unlock(&dev->struct_mutex); | |
1830 | return -EIO; | |
1831 | } | |
1832 | ||
1833 | if (dev_priv->mm.suspended) { | |
1834 | DRM_ERROR("Execbuf while VT-switched.\n"); | |
1835 | mutex_unlock(&dev->struct_mutex); | |
1836 | return -EBUSY; | |
1837 | } | |
1838 | ||
1839 | /* Zero the gloabl flush/invalidate flags. These | |
1840 | * will be modified as each object is bound to the | |
1841 | * gtt | |
1842 | */ | |
1843 | dev->invalidate_domains = 0; | |
1844 | dev->flush_domains = 0; | |
1845 | ||
1846 | /* Look up object handles and perform the relocations */ | |
1847 | for (i = 0; i < args->buffer_count; i++) { | |
1848 | object_list[i] = drm_gem_object_lookup(dev, file_priv, | |
1849 | exec_list[i].handle); | |
1850 | if (object_list[i] == NULL) { | |
1851 | DRM_ERROR("Invalid object handle %d at index %d\n", | |
1852 | exec_list[i].handle, i); | |
1853 | ret = -EBADF; | |
1854 | goto err; | |
1855 | } | |
1856 | ||
1857 | object_list[i]->pending_read_domains = 0; | |
1858 | object_list[i]->pending_write_domain = 0; | |
1859 | ret = i915_gem_object_pin_and_relocate(object_list[i], | |
1860 | file_priv, | |
1861 | &exec_list[i]); | |
1862 | if (ret) { | |
1863 | DRM_ERROR("object bind and relocate failed %d\n", ret); | |
1864 | goto err; | |
1865 | } | |
1866 | pinned = i + 1; | |
1867 | } | |
1868 | ||
1869 | /* Set the pending read domains for the batch buffer to COMMAND */ | |
1870 | batch_obj = object_list[args->buffer_count-1]; | |
1871 | batch_obj->pending_read_domains = I915_GEM_DOMAIN_COMMAND; | |
1872 | batch_obj->pending_write_domain = 0; | |
1873 | ||
1874 | i915_verify_inactive(dev, __FILE__, __LINE__); | |
1875 | ||
1876 | for (i = 0; i < args->buffer_count; i++) { | |
1877 | struct drm_gem_object *obj = object_list[i]; | |
673a394b EA |
1878 | |
1879 | /* make sure all previous memory operations have passed */ | |
1880 | ret = i915_gem_object_set_domain(obj, | |
1881 | obj->pending_read_domains, | |
1882 | obj->pending_write_domain); | |
b670d815 EA |
1883 | if (ret) { |
1884 | /* As we've partially updated domains on our buffers, | |
1885 | * we have to emit the flush we've accumulated | |
1886 | * before exiting, or we'll have broken the | |
1887 | * active/flushing/inactive invariants. | |
1888 | * | |
1889 | * We'll potentially have some things marked as | |
1890 | * being in write domains that they actually aren't, | |
1891 | * but that should be merely a minor performance loss. | |
1892 | */ | |
1893 | flush_domains = i915_gem_dev_set_domain(dev); | |
1894 | (void)i915_add_request(dev, flush_domains); | |
673a394b | 1895 | goto err; |
b670d815 | 1896 | } |
673a394b EA |
1897 | } |
1898 | ||
1899 | i915_verify_inactive(dev, __FILE__, __LINE__); | |
1900 | ||
1901 | /* Flush/invalidate caches and chipset buffer */ | |
1902 | flush_domains = i915_gem_dev_set_domain(dev); | |
1903 | ||
1904 | i915_verify_inactive(dev, __FILE__, __LINE__); | |
1905 | ||
1906 | #if WATCH_COHERENCY | |
1907 | for (i = 0; i < args->buffer_count; i++) { | |
1908 | i915_gem_object_check_coherency(object_list[i], | |
1909 | exec_list[i].handle); | |
1910 | } | |
1911 | #endif | |
1912 | ||
1913 | exec_offset = exec_list[args->buffer_count - 1].offset; | |
1914 | ||
1915 | #if WATCH_EXEC | |
1916 | i915_gem_dump_object(object_list[args->buffer_count - 1], | |
1917 | args->batch_len, | |
1918 | __func__, | |
1919 | ~0); | |
1920 | #endif | |
1921 | ||
1922 | (void)i915_add_request(dev, flush_domains); | |
1923 | ||
1924 | /* Exec the batchbuffer */ | |
1925 | ret = i915_dispatch_gem_execbuffer(dev, args, exec_offset); | |
1926 | if (ret) { | |
1927 | DRM_ERROR("dispatch failed %d\n", ret); | |
1928 | goto err; | |
1929 | } | |
1930 | ||
1931 | /* | |
1932 | * Ensure that the commands in the batch buffer are | |
1933 | * finished before the interrupt fires | |
1934 | */ | |
1935 | flush_domains = i915_retire_commands(dev); | |
1936 | ||
1937 | i915_verify_inactive(dev, __FILE__, __LINE__); | |
1938 | ||
1939 | /* | |
1940 | * Get a seqno representing the execution of the current buffer, | |
1941 | * which we can wait on. We would like to mitigate these interrupts, | |
1942 | * likely by only creating seqnos occasionally (so that we have | |
1943 | * *some* interrupts representing completion of buffers that we can | |
1944 | * wait on when trying to clear up gtt space). | |
1945 | */ | |
1946 | seqno = i915_add_request(dev, flush_domains); | |
1947 | BUG_ON(seqno == 0); | |
1948 | i915_file_priv->mm.last_gem_seqno = seqno; | |
1949 | for (i = 0; i < args->buffer_count; i++) { | |
1950 | struct drm_gem_object *obj = object_list[i]; | |
673a394b | 1951 | |
ce44b0ea | 1952 | i915_gem_object_move_to_active(obj, seqno); |
673a394b EA |
1953 | #if WATCH_LRU |
1954 | DRM_INFO("%s: move to exec list %p\n", __func__, obj); | |
1955 | #endif | |
1956 | } | |
1957 | #if WATCH_LRU | |
1958 | i915_dump_lru(dev, __func__); | |
1959 | #endif | |
1960 | ||
1961 | i915_verify_inactive(dev, __FILE__, __LINE__); | |
1962 | ||
1963 | /* Copy the new buffer offsets back to the user's exec list. */ | |
1964 | ret = copy_to_user((struct drm_i915_relocation_entry __user *) | |
1965 | (uintptr_t) args->buffers_ptr, | |
1966 | exec_list, | |
1967 | sizeof(*exec_list) * args->buffer_count); | |
1968 | if (ret) | |
1969 | DRM_ERROR("failed to copy %d exec entries " | |
1970 | "back to user (%d)\n", | |
1971 | args->buffer_count, ret); | |
1972 | err: | |
1973 | if (object_list != NULL) { | |
1974 | for (i = 0; i < pinned; i++) | |
1975 | i915_gem_object_unpin(object_list[i]); | |
1976 | ||
1977 | for (i = 0; i < args->buffer_count; i++) | |
1978 | drm_gem_object_unreference(object_list[i]); | |
1979 | } | |
1980 | mutex_unlock(&dev->struct_mutex); | |
1981 | ||
1982 | pre_mutex_err: | |
1983 | drm_free(object_list, sizeof(*object_list) * args->buffer_count, | |
1984 | DRM_MEM_DRIVER); | |
1985 | drm_free(exec_list, sizeof(*exec_list) * args->buffer_count, | |
1986 | DRM_MEM_DRIVER); | |
1987 | ||
1988 | return ret; | |
1989 | } | |
1990 | ||
1991 | int | |
1992 | i915_gem_object_pin(struct drm_gem_object *obj, uint32_t alignment) | |
1993 | { | |
1994 | struct drm_device *dev = obj->dev; | |
1995 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
1996 | int ret; | |
1997 | ||
1998 | i915_verify_inactive(dev, __FILE__, __LINE__); | |
1999 | if (obj_priv->gtt_space == NULL) { | |
2000 | ret = i915_gem_object_bind_to_gtt(obj, alignment); | |
2001 | if (ret != 0) { | |
2002 | DRM_ERROR("Failure to bind: %d", ret); | |
2003 | return ret; | |
2004 | } | |
2005 | } | |
2006 | obj_priv->pin_count++; | |
2007 | ||
2008 | /* If the object is not active and not pending a flush, | |
2009 | * remove it from the inactive list | |
2010 | */ | |
2011 | if (obj_priv->pin_count == 1) { | |
2012 | atomic_inc(&dev->pin_count); | |
2013 | atomic_add(obj->size, &dev->pin_memory); | |
2014 | if (!obj_priv->active && | |
2015 | (obj->write_domain & ~(I915_GEM_DOMAIN_CPU | | |
2016 | I915_GEM_DOMAIN_GTT)) == 0 && | |
2017 | !list_empty(&obj_priv->list)) | |
2018 | list_del_init(&obj_priv->list); | |
2019 | } | |
2020 | i915_verify_inactive(dev, __FILE__, __LINE__); | |
2021 | ||
2022 | return 0; | |
2023 | } | |
2024 | ||
2025 | void | |
2026 | i915_gem_object_unpin(struct drm_gem_object *obj) | |
2027 | { | |
2028 | struct drm_device *dev = obj->dev; | |
2029 | drm_i915_private_t *dev_priv = dev->dev_private; | |
2030 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
2031 | ||
2032 | i915_verify_inactive(dev, __FILE__, __LINE__); | |
2033 | obj_priv->pin_count--; | |
2034 | BUG_ON(obj_priv->pin_count < 0); | |
2035 | BUG_ON(obj_priv->gtt_space == NULL); | |
2036 | ||
2037 | /* If the object is no longer pinned, and is | |
2038 | * neither active nor being flushed, then stick it on | |
2039 | * the inactive list | |
2040 | */ | |
2041 | if (obj_priv->pin_count == 0) { | |
2042 | if (!obj_priv->active && | |
2043 | (obj->write_domain & ~(I915_GEM_DOMAIN_CPU | | |
2044 | I915_GEM_DOMAIN_GTT)) == 0) | |
2045 | list_move_tail(&obj_priv->list, | |
2046 | &dev_priv->mm.inactive_list); | |
2047 | atomic_dec(&dev->pin_count); | |
2048 | atomic_sub(obj->size, &dev->pin_memory); | |
2049 | } | |
2050 | i915_verify_inactive(dev, __FILE__, __LINE__); | |
2051 | } | |
2052 | ||
2053 | int | |
2054 | i915_gem_pin_ioctl(struct drm_device *dev, void *data, | |
2055 | struct drm_file *file_priv) | |
2056 | { | |
2057 | struct drm_i915_gem_pin *args = data; | |
2058 | struct drm_gem_object *obj; | |
2059 | struct drm_i915_gem_object *obj_priv; | |
2060 | int ret; | |
2061 | ||
2062 | mutex_lock(&dev->struct_mutex); | |
2063 | ||
2064 | obj = drm_gem_object_lookup(dev, file_priv, args->handle); | |
2065 | if (obj == NULL) { | |
2066 | DRM_ERROR("Bad handle in i915_gem_pin_ioctl(): %d\n", | |
2067 | args->handle); | |
2068 | mutex_unlock(&dev->struct_mutex); | |
2069 | return -EBADF; | |
2070 | } | |
2071 | obj_priv = obj->driver_private; | |
2072 | ||
2073 | ret = i915_gem_object_pin(obj, args->alignment); | |
2074 | if (ret != 0) { | |
2075 | drm_gem_object_unreference(obj); | |
2076 | mutex_unlock(&dev->struct_mutex); | |
2077 | return ret; | |
2078 | } | |
2079 | ||
2080 | /* XXX - flush the CPU caches for pinned objects | |
2081 | * as the X server doesn't manage domains yet | |
2082 | */ | |
2083 | if (obj->write_domain & I915_GEM_DOMAIN_CPU) { | |
2084 | i915_gem_clflush_object(obj); | |
2085 | drm_agp_chipset_flush(dev); | |
2086 | obj->write_domain = 0; | |
2087 | } | |
2088 | args->offset = obj_priv->gtt_offset; | |
2089 | drm_gem_object_unreference(obj); | |
2090 | mutex_unlock(&dev->struct_mutex); | |
2091 | ||
2092 | return 0; | |
2093 | } | |
2094 | ||
2095 | int | |
2096 | i915_gem_unpin_ioctl(struct drm_device *dev, void *data, | |
2097 | struct drm_file *file_priv) | |
2098 | { | |
2099 | struct drm_i915_gem_pin *args = data; | |
2100 | struct drm_gem_object *obj; | |
2101 | ||
2102 | mutex_lock(&dev->struct_mutex); | |
2103 | ||
2104 | obj = drm_gem_object_lookup(dev, file_priv, args->handle); | |
2105 | if (obj == NULL) { | |
2106 | DRM_ERROR("Bad handle in i915_gem_unpin_ioctl(): %d\n", | |
2107 | args->handle); | |
2108 | mutex_unlock(&dev->struct_mutex); | |
2109 | return -EBADF; | |
2110 | } | |
2111 | ||
2112 | i915_gem_object_unpin(obj); | |
2113 | ||
2114 | drm_gem_object_unreference(obj); | |
2115 | mutex_unlock(&dev->struct_mutex); | |
2116 | return 0; | |
2117 | } | |
2118 | ||
2119 | int | |
2120 | i915_gem_busy_ioctl(struct drm_device *dev, void *data, | |
2121 | struct drm_file *file_priv) | |
2122 | { | |
2123 | struct drm_i915_gem_busy *args = data; | |
2124 | struct drm_gem_object *obj; | |
2125 | struct drm_i915_gem_object *obj_priv; | |
2126 | ||
2127 | mutex_lock(&dev->struct_mutex); | |
2128 | obj = drm_gem_object_lookup(dev, file_priv, args->handle); | |
2129 | if (obj == NULL) { | |
2130 | DRM_ERROR("Bad handle in i915_gem_busy_ioctl(): %d\n", | |
2131 | args->handle); | |
2132 | mutex_unlock(&dev->struct_mutex); | |
2133 | return -EBADF; | |
2134 | } | |
2135 | ||
2136 | obj_priv = obj->driver_private; | |
2137 | args->busy = obj_priv->active; | |
2138 | ||
2139 | drm_gem_object_unreference(obj); | |
2140 | mutex_unlock(&dev->struct_mutex); | |
2141 | return 0; | |
2142 | } | |
2143 | ||
2144 | int | |
2145 | i915_gem_throttle_ioctl(struct drm_device *dev, void *data, | |
2146 | struct drm_file *file_priv) | |
2147 | { | |
2148 | return i915_gem_ring_throttle(dev, file_priv); | |
2149 | } | |
2150 | ||
2151 | int i915_gem_init_object(struct drm_gem_object *obj) | |
2152 | { | |
2153 | struct drm_i915_gem_object *obj_priv; | |
2154 | ||
2155 | obj_priv = drm_calloc(1, sizeof(*obj_priv), DRM_MEM_DRIVER); | |
2156 | if (obj_priv == NULL) | |
2157 | return -ENOMEM; | |
2158 | ||
2159 | /* | |
2160 | * We've just allocated pages from the kernel, | |
2161 | * so they've just been written by the CPU with | |
2162 | * zeros. They'll need to be clflushed before we | |
2163 | * use them with the GPU. | |
2164 | */ | |
2165 | obj->write_domain = I915_GEM_DOMAIN_CPU; | |
2166 | obj->read_domains = I915_GEM_DOMAIN_CPU; | |
2167 | ||
ba1eb1d8 KP |
2168 | obj_priv->agp_type = AGP_USER_MEMORY; |
2169 | ||
673a394b EA |
2170 | obj->driver_private = obj_priv; |
2171 | obj_priv->obj = obj; | |
2172 | INIT_LIST_HEAD(&obj_priv->list); | |
2173 | return 0; | |
2174 | } | |
2175 | ||
2176 | void i915_gem_free_object(struct drm_gem_object *obj) | |
2177 | { | |
2178 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
2179 | ||
2180 | while (obj_priv->pin_count > 0) | |
2181 | i915_gem_object_unpin(obj); | |
2182 | ||
2183 | i915_gem_object_unbind(obj); | |
2184 | ||
2185 | drm_free(obj_priv->page_cpu_valid, 1, DRM_MEM_DRIVER); | |
2186 | drm_free(obj->driver_private, 1, DRM_MEM_DRIVER); | |
2187 | } | |
2188 | ||
2189 | static int | |
2190 | i915_gem_set_domain(struct drm_gem_object *obj, | |
2191 | struct drm_file *file_priv, | |
2192 | uint32_t read_domains, | |
2193 | uint32_t write_domain) | |
2194 | { | |
2195 | struct drm_device *dev = obj->dev; | |
2196 | int ret; | |
2197 | uint32_t flush_domains; | |
2198 | ||
2199 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); | |
2200 | ||
2201 | ret = i915_gem_object_set_domain(obj, read_domains, write_domain); | |
2202 | if (ret) | |
2203 | return ret; | |
2204 | flush_domains = i915_gem_dev_set_domain(obj->dev); | |
2205 | ||
2206 | if (flush_domains & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT)) | |
2207 | (void) i915_add_request(dev, flush_domains); | |
2208 | ||
2209 | return 0; | |
2210 | } | |
2211 | ||
2212 | /** Unbinds all objects that are on the given buffer list. */ | |
2213 | static int | |
2214 | i915_gem_evict_from_list(struct drm_device *dev, struct list_head *head) | |
2215 | { | |
2216 | struct drm_gem_object *obj; | |
2217 | struct drm_i915_gem_object *obj_priv; | |
2218 | int ret; | |
2219 | ||
2220 | while (!list_empty(head)) { | |
2221 | obj_priv = list_first_entry(head, | |
2222 | struct drm_i915_gem_object, | |
2223 | list); | |
2224 | obj = obj_priv->obj; | |
2225 | ||
2226 | if (obj_priv->pin_count != 0) { | |
2227 | DRM_ERROR("Pinned object in unbind list\n"); | |
2228 | mutex_unlock(&dev->struct_mutex); | |
2229 | return -EINVAL; | |
2230 | } | |
2231 | ||
2232 | ret = i915_gem_object_unbind(obj); | |
2233 | if (ret != 0) { | |
2234 | DRM_ERROR("Error unbinding object in LeaveVT: %d\n", | |
2235 | ret); | |
2236 | mutex_unlock(&dev->struct_mutex); | |
2237 | return ret; | |
2238 | } | |
2239 | } | |
2240 | ||
2241 | ||
2242 | return 0; | |
2243 | } | |
2244 | ||
2245 | static int | |
2246 | i915_gem_idle(struct drm_device *dev) | |
2247 | { | |
2248 | drm_i915_private_t *dev_priv = dev->dev_private; | |
2249 | uint32_t seqno, cur_seqno, last_seqno; | |
2250 | int stuck, ret; | |
2251 | ||
6dbe2772 KP |
2252 | mutex_lock(&dev->struct_mutex); |
2253 | ||
2254 | if (dev_priv->mm.suspended || dev_priv->ring.ring_obj == NULL) { | |
2255 | mutex_unlock(&dev->struct_mutex); | |
673a394b | 2256 | return 0; |
6dbe2772 | 2257 | } |
673a394b EA |
2258 | |
2259 | /* Hack! Don't let anybody do execbuf while we don't control the chip. | |
2260 | * We need to replace this with a semaphore, or something. | |
2261 | */ | |
2262 | dev_priv->mm.suspended = 1; | |
2263 | ||
6dbe2772 KP |
2264 | /* Cancel the retire work handler, wait for it to finish if running |
2265 | */ | |
2266 | mutex_unlock(&dev->struct_mutex); | |
2267 | cancel_delayed_work_sync(&dev_priv->mm.retire_work); | |
2268 | mutex_lock(&dev->struct_mutex); | |
2269 | ||
673a394b EA |
2270 | i915_kernel_lost_context(dev); |
2271 | ||
2272 | /* Flush the GPU along with all non-CPU write domains | |
2273 | */ | |
2274 | i915_gem_flush(dev, ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT), | |
2275 | ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT)); | |
2276 | seqno = i915_add_request(dev, ~(I915_GEM_DOMAIN_CPU | | |
2277 | I915_GEM_DOMAIN_GTT)); | |
2278 | ||
2279 | if (seqno == 0) { | |
2280 | mutex_unlock(&dev->struct_mutex); | |
2281 | return -ENOMEM; | |
2282 | } | |
2283 | ||
2284 | dev_priv->mm.waiting_gem_seqno = seqno; | |
2285 | last_seqno = 0; | |
2286 | stuck = 0; | |
2287 | for (;;) { | |
2288 | cur_seqno = i915_get_gem_seqno(dev); | |
2289 | if (i915_seqno_passed(cur_seqno, seqno)) | |
2290 | break; | |
2291 | if (last_seqno == cur_seqno) { | |
2292 | if (stuck++ > 100) { | |
2293 | DRM_ERROR("hardware wedged\n"); | |
2294 | dev_priv->mm.wedged = 1; | |
2295 | DRM_WAKEUP(&dev_priv->irq_queue); | |
2296 | break; | |
2297 | } | |
2298 | } | |
2299 | msleep(10); | |
2300 | last_seqno = cur_seqno; | |
2301 | } | |
2302 | dev_priv->mm.waiting_gem_seqno = 0; | |
2303 | ||
2304 | i915_gem_retire_requests(dev); | |
2305 | ||
28dfe52a EA |
2306 | if (!dev_priv->mm.wedged) { |
2307 | /* Active and flushing should now be empty as we've | |
2308 | * waited for a sequence higher than any pending execbuffer | |
2309 | */ | |
2310 | WARN_ON(!list_empty(&dev_priv->mm.active_list)); | |
2311 | WARN_ON(!list_empty(&dev_priv->mm.flushing_list)); | |
2312 | /* Request should now be empty as we've also waited | |
2313 | * for the last request in the list | |
2314 | */ | |
2315 | WARN_ON(!list_empty(&dev_priv->mm.request_list)); | |
2316 | } | |
673a394b | 2317 | |
28dfe52a EA |
2318 | /* Empty the active and flushing lists to inactive. If there's |
2319 | * anything left at this point, it means that we're wedged and | |
2320 | * nothing good's going to happen by leaving them there. So strip | |
2321 | * the GPU domains and just stuff them onto inactive. | |
673a394b | 2322 | */ |
28dfe52a EA |
2323 | while (!list_empty(&dev_priv->mm.active_list)) { |
2324 | struct drm_i915_gem_object *obj_priv; | |
673a394b | 2325 | |
28dfe52a EA |
2326 | obj_priv = list_first_entry(&dev_priv->mm.active_list, |
2327 | struct drm_i915_gem_object, | |
2328 | list); | |
2329 | obj_priv->obj->write_domain &= ~I915_GEM_GPU_DOMAINS; | |
2330 | i915_gem_object_move_to_inactive(obj_priv->obj); | |
2331 | } | |
2332 | ||
2333 | while (!list_empty(&dev_priv->mm.flushing_list)) { | |
2334 | struct drm_i915_gem_object *obj_priv; | |
2335 | ||
151903d5 | 2336 | obj_priv = list_first_entry(&dev_priv->mm.flushing_list, |
28dfe52a EA |
2337 | struct drm_i915_gem_object, |
2338 | list); | |
2339 | obj_priv->obj->write_domain &= ~I915_GEM_GPU_DOMAINS; | |
2340 | i915_gem_object_move_to_inactive(obj_priv->obj); | |
2341 | } | |
2342 | ||
2343 | ||
2344 | /* Move all inactive buffers out of the GTT. */ | |
673a394b | 2345 | ret = i915_gem_evict_from_list(dev, &dev_priv->mm.inactive_list); |
28dfe52a | 2346 | WARN_ON(!list_empty(&dev_priv->mm.inactive_list)); |
6dbe2772 KP |
2347 | if (ret) { |
2348 | mutex_unlock(&dev->struct_mutex); | |
673a394b | 2349 | return ret; |
6dbe2772 | 2350 | } |
673a394b | 2351 | |
6dbe2772 KP |
2352 | i915_gem_cleanup_ringbuffer(dev); |
2353 | mutex_unlock(&dev->struct_mutex); | |
2354 | ||
673a394b EA |
2355 | return 0; |
2356 | } | |
2357 | ||
2358 | static int | |
2359 | i915_gem_init_hws(struct drm_device *dev) | |
2360 | { | |
2361 | drm_i915_private_t *dev_priv = dev->dev_private; | |
2362 | struct drm_gem_object *obj; | |
2363 | struct drm_i915_gem_object *obj_priv; | |
2364 | int ret; | |
2365 | ||
2366 | /* If we need a physical address for the status page, it's already | |
2367 | * initialized at driver load time. | |
2368 | */ | |
2369 | if (!I915_NEED_GFX_HWS(dev)) | |
2370 | return 0; | |
2371 | ||
2372 | obj = drm_gem_object_alloc(dev, 4096); | |
2373 | if (obj == NULL) { | |
2374 | DRM_ERROR("Failed to allocate status page\n"); | |
2375 | return -ENOMEM; | |
2376 | } | |
2377 | obj_priv = obj->driver_private; | |
ba1eb1d8 | 2378 | obj_priv->agp_type = AGP_USER_CACHED_MEMORY; |
673a394b EA |
2379 | |
2380 | ret = i915_gem_object_pin(obj, 4096); | |
2381 | if (ret != 0) { | |
2382 | drm_gem_object_unreference(obj); | |
2383 | return ret; | |
2384 | } | |
2385 | ||
2386 | dev_priv->status_gfx_addr = obj_priv->gtt_offset; | |
673a394b | 2387 | |
ba1eb1d8 KP |
2388 | dev_priv->hw_status_page = kmap(obj_priv->page_list[0]); |
2389 | if (dev_priv->hw_status_page == NULL) { | |
673a394b EA |
2390 | DRM_ERROR("Failed to map status page.\n"); |
2391 | memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map)); | |
2392 | drm_gem_object_unreference(obj); | |
2393 | return -EINVAL; | |
2394 | } | |
2395 | dev_priv->hws_obj = obj; | |
673a394b EA |
2396 | memset(dev_priv->hw_status_page, 0, PAGE_SIZE); |
2397 | I915_WRITE(HWS_PGA, dev_priv->status_gfx_addr); | |
ba1eb1d8 | 2398 | I915_READ(HWS_PGA); /* posting read */ |
673a394b EA |
2399 | DRM_DEBUG("hws offset: 0x%08x\n", dev_priv->status_gfx_addr); |
2400 | ||
2401 | return 0; | |
2402 | } | |
2403 | ||
2404 | static int | |
2405 | i915_gem_init_ringbuffer(struct drm_device *dev) | |
2406 | { | |
2407 | drm_i915_private_t *dev_priv = dev->dev_private; | |
2408 | struct drm_gem_object *obj; | |
2409 | struct drm_i915_gem_object *obj_priv; | |
2410 | int ret; | |
50aa253d | 2411 | u32 head; |
673a394b EA |
2412 | |
2413 | ret = i915_gem_init_hws(dev); | |
2414 | if (ret != 0) | |
2415 | return ret; | |
2416 | ||
2417 | obj = drm_gem_object_alloc(dev, 128 * 1024); | |
2418 | if (obj == NULL) { | |
2419 | DRM_ERROR("Failed to allocate ringbuffer\n"); | |
2420 | return -ENOMEM; | |
2421 | } | |
2422 | obj_priv = obj->driver_private; | |
2423 | ||
2424 | ret = i915_gem_object_pin(obj, 4096); | |
2425 | if (ret != 0) { | |
2426 | drm_gem_object_unreference(obj); | |
2427 | return ret; | |
2428 | } | |
2429 | ||
2430 | /* Set up the kernel mapping for the ring. */ | |
2431 | dev_priv->ring.Size = obj->size; | |
2432 | dev_priv->ring.tail_mask = obj->size - 1; | |
2433 | ||
2434 | dev_priv->ring.map.offset = dev->agp->base + obj_priv->gtt_offset; | |
2435 | dev_priv->ring.map.size = obj->size; | |
2436 | dev_priv->ring.map.type = 0; | |
2437 | dev_priv->ring.map.flags = 0; | |
2438 | dev_priv->ring.map.mtrr = 0; | |
2439 | ||
bd88ee4c | 2440 | drm_core_ioremap_wc(&dev_priv->ring.map, dev); |
673a394b EA |
2441 | if (dev_priv->ring.map.handle == NULL) { |
2442 | DRM_ERROR("Failed to map ringbuffer.\n"); | |
2443 | memset(&dev_priv->ring, 0, sizeof(dev_priv->ring)); | |
2444 | drm_gem_object_unreference(obj); | |
2445 | return -EINVAL; | |
2446 | } | |
2447 | dev_priv->ring.ring_obj = obj; | |
2448 | dev_priv->ring.virtual_start = dev_priv->ring.map.handle; | |
2449 | ||
2450 | /* Stop the ring if it's running. */ | |
2451 | I915_WRITE(PRB0_CTL, 0); | |
673a394b | 2452 | I915_WRITE(PRB0_TAIL, 0); |
50aa253d | 2453 | I915_WRITE(PRB0_HEAD, 0); |
673a394b EA |
2454 | |
2455 | /* Initialize the ring. */ | |
2456 | I915_WRITE(PRB0_START, obj_priv->gtt_offset); | |
50aa253d KP |
2457 | head = I915_READ(PRB0_HEAD) & HEAD_ADDR; |
2458 | ||
2459 | /* G45 ring initialization fails to reset head to zero */ | |
2460 | if (head != 0) { | |
2461 | DRM_ERROR("Ring head not reset to zero " | |
2462 | "ctl %08x head %08x tail %08x start %08x\n", | |
2463 | I915_READ(PRB0_CTL), | |
2464 | I915_READ(PRB0_HEAD), | |
2465 | I915_READ(PRB0_TAIL), | |
2466 | I915_READ(PRB0_START)); | |
2467 | I915_WRITE(PRB0_HEAD, 0); | |
2468 | ||
2469 | DRM_ERROR("Ring head forced to zero " | |
2470 | "ctl %08x head %08x tail %08x start %08x\n", | |
2471 | I915_READ(PRB0_CTL), | |
2472 | I915_READ(PRB0_HEAD), | |
2473 | I915_READ(PRB0_TAIL), | |
2474 | I915_READ(PRB0_START)); | |
2475 | } | |
2476 | ||
673a394b EA |
2477 | I915_WRITE(PRB0_CTL, |
2478 | ((obj->size - 4096) & RING_NR_PAGES) | | |
2479 | RING_NO_REPORT | | |
2480 | RING_VALID); | |
2481 | ||
50aa253d KP |
2482 | head = I915_READ(PRB0_HEAD) & HEAD_ADDR; |
2483 | ||
2484 | /* If the head is still not zero, the ring is dead */ | |
2485 | if (head != 0) { | |
2486 | DRM_ERROR("Ring initialization failed " | |
2487 | "ctl %08x head %08x tail %08x start %08x\n", | |
2488 | I915_READ(PRB0_CTL), | |
2489 | I915_READ(PRB0_HEAD), | |
2490 | I915_READ(PRB0_TAIL), | |
2491 | I915_READ(PRB0_START)); | |
2492 | return -EIO; | |
2493 | } | |
2494 | ||
673a394b EA |
2495 | /* Update our cache of the ring state */ |
2496 | i915_kernel_lost_context(dev); | |
2497 | ||
2498 | return 0; | |
2499 | } | |
2500 | ||
2501 | static void | |
2502 | i915_gem_cleanup_ringbuffer(struct drm_device *dev) | |
2503 | { | |
2504 | drm_i915_private_t *dev_priv = dev->dev_private; | |
2505 | ||
2506 | if (dev_priv->ring.ring_obj == NULL) | |
2507 | return; | |
2508 | ||
2509 | drm_core_ioremapfree(&dev_priv->ring.map, dev); | |
2510 | ||
2511 | i915_gem_object_unpin(dev_priv->ring.ring_obj); | |
2512 | drm_gem_object_unreference(dev_priv->ring.ring_obj); | |
2513 | dev_priv->ring.ring_obj = NULL; | |
2514 | memset(&dev_priv->ring, 0, sizeof(dev_priv->ring)); | |
2515 | ||
2516 | if (dev_priv->hws_obj != NULL) { | |
ba1eb1d8 KP |
2517 | struct drm_gem_object *obj = dev_priv->hws_obj; |
2518 | struct drm_i915_gem_object *obj_priv = obj->driver_private; | |
2519 | ||
2520 | kunmap(obj_priv->page_list[0]); | |
2521 | i915_gem_object_unpin(obj); | |
2522 | drm_gem_object_unreference(obj); | |
673a394b EA |
2523 | dev_priv->hws_obj = NULL; |
2524 | memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map)); | |
ba1eb1d8 | 2525 | dev_priv->hw_status_page = NULL; |
673a394b EA |
2526 | |
2527 | /* Write high address into HWS_PGA when disabling. */ | |
2528 | I915_WRITE(HWS_PGA, 0x1ffff000); | |
2529 | } | |
2530 | } | |
2531 | ||
2532 | int | |
2533 | i915_gem_entervt_ioctl(struct drm_device *dev, void *data, | |
2534 | struct drm_file *file_priv) | |
2535 | { | |
2536 | drm_i915_private_t *dev_priv = dev->dev_private; | |
2537 | int ret; | |
2538 | ||
2539 | if (dev_priv->mm.wedged) { | |
2540 | DRM_ERROR("Reenabling wedged hardware, good luck\n"); | |
2541 | dev_priv->mm.wedged = 0; | |
2542 | } | |
2543 | ||
2544 | ret = i915_gem_init_ringbuffer(dev); | |
2545 | if (ret != 0) | |
2546 | return ret; | |
2547 | ||
0839ccb8 KP |
2548 | dev_priv->mm.gtt_mapping = io_mapping_create_wc(dev->agp->base, |
2549 | dev->agp->agp_info.aper_size | |
2550 | * 1024 * 1024); | |
2551 | ||
673a394b EA |
2552 | mutex_lock(&dev->struct_mutex); |
2553 | BUG_ON(!list_empty(&dev_priv->mm.active_list)); | |
2554 | BUG_ON(!list_empty(&dev_priv->mm.flushing_list)); | |
2555 | BUG_ON(!list_empty(&dev_priv->mm.inactive_list)); | |
2556 | BUG_ON(!list_empty(&dev_priv->mm.request_list)); | |
2557 | dev_priv->mm.suspended = 0; | |
2558 | mutex_unlock(&dev->struct_mutex); | |
dbb19d30 KH |
2559 | |
2560 | drm_irq_install(dev); | |
2561 | ||
673a394b EA |
2562 | return 0; |
2563 | } | |
2564 | ||
2565 | int | |
2566 | i915_gem_leavevt_ioctl(struct drm_device *dev, void *data, | |
2567 | struct drm_file *file_priv) | |
2568 | { | |
0839ccb8 | 2569 | drm_i915_private_t *dev_priv = dev->dev_private; |
673a394b EA |
2570 | int ret; |
2571 | ||
673a394b | 2572 | ret = i915_gem_idle(dev); |
dbb19d30 KH |
2573 | drm_irq_uninstall(dev); |
2574 | ||
0839ccb8 | 2575 | io_mapping_free(dev_priv->mm.gtt_mapping); |
6dbe2772 | 2576 | return ret; |
673a394b EA |
2577 | } |
2578 | ||
2579 | void | |
2580 | i915_gem_lastclose(struct drm_device *dev) | |
2581 | { | |
2582 | int ret; | |
673a394b | 2583 | |
6dbe2772 KP |
2584 | ret = i915_gem_idle(dev); |
2585 | if (ret) | |
2586 | DRM_ERROR("failed to idle hardware: %d\n", ret); | |
673a394b EA |
2587 | } |
2588 | ||
2589 | void | |
2590 | i915_gem_load(struct drm_device *dev) | |
2591 | { | |
2592 | drm_i915_private_t *dev_priv = dev->dev_private; | |
2593 | ||
2594 | INIT_LIST_HEAD(&dev_priv->mm.active_list); | |
2595 | INIT_LIST_HEAD(&dev_priv->mm.flushing_list); | |
2596 | INIT_LIST_HEAD(&dev_priv->mm.inactive_list); | |
2597 | INIT_LIST_HEAD(&dev_priv->mm.request_list); | |
2598 | INIT_DELAYED_WORK(&dev_priv->mm.retire_work, | |
2599 | i915_gem_retire_work_handler); | |
2600 | dev_priv->mm.next_gem_seqno = 1; | |
2601 | ||
2602 | i915_gem_detect_bit_6_swizzle(dev); | |
2603 | } |