]>
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" | |
1c5d22f7 | 32 | #include "i915_trace.h" |
652c393a | 33 | #include "intel_drv.h" |
5949eac4 | 34 | #include <linux/shmem_fs.h> |
5a0e3ad6 | 35 | #include <linux/slab.h> |
673a394b | 36 | #include <linux/swap.h> |
79e53945 | 37 | #include <linux/pci.h> |
673a394b | 38 | |
88241785 | 39 | static __must_check int i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object *obj); |
05394f39 CW |
40 | static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj); |
41 | static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj); | |
88241785 CW |
42 | static __must_check int i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj, |
43 | unsigned alignment, | |
44 | bool map_and_fenceable); | |
d9e86c0e CW |
45 | static void i915_gem_clear_fence_reg(struct drm_device *dev, |
46 | struct drm_i915_fence_reg *reg); | |
05394f39 CW |
47 | static int i915_gem_phys_pwrite(struct drm_device *dev, |
48 | struct drm_i915_gem_object *obj, | |
71acb5eb | 49 | struct drm_i915_gem_pwrite *args, |
05394f39 CW |
50 | struct drm_file *file); |
51 | static void i915_gem_free_object_tail(struct drm_i915_gem_object *obj); | |
673a394b | 52 | |
17250b71 | 53 | static int i915_gem_inactive_shrink(struct shrinker *shrinker, |
1495f230 | 54 | struct shrink_control *sc); |
8c59967c | 55 | static void i915_gem_object_truncate(struct drm_i915_gem_object *obj); |
31169714 | 56 | |
73aa808f CW |
57 | /* some bookkeeping */ |
58 | static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv, | |
59 | size_t size) | |
60 | { | |
61 | dev_priv->mm.object_count++; | |
62 | dev_priv->mm.object_memory += size; | |
63 | } | |
64 | ||
65 | static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv, | |
66 | size_t size) | |
67 | { | |
68 | dev_priv->mm.object_count--; | |
69 | dev_priv->mm.object_memory -= size; | |
70 | } | |
71 | ||
21dd3734 CW |
72 | static int |
73 | i915_gem_wait_for_error(struct drm_device *dev) | |
30dbf0c0 CW |
74 | { |
75 | struct drm_i915_private *dev_priv = dev->dev_private; | |
76 | struct completion *x = &dev_priv->error_completion; | |
77 | unsigned long flags; | |
78 | int ret; | |
79 | ||
80 | if (!atomic_read(&dev_priv->mm.wedged)) | |
81 | return 0; | |
82 | ||
83 | ret = wait_for_completion_interruptible(x); | |
84 | if (ret) | |
85 | return ret; | |
86 | ||
21dd3734 CW |
87 | if (atomic_read(&dev_priv->mm.wedged)) { |
88 | /* GPU is hung, bump the completion count to account for | |
89 | * the token we just consumed so that we never hit zero and | |
90 | * end up waiting upon a subsequent completion event that | |
91 | * will never happen. | |
92 | */ | |
93 | spin_lock_irqsave(&x->wait.lock, flags); | |
94 | x->done++; | |
95 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
96 | } | |
97 | return 0; | |
30dbf0c0 CW |
98 | } |
99 | ||
54cf91dc | 100 | int i915_mutex_lock_interruptible(struct drm_device *dev) |
76c1dec1 | 101 | { |
76c1dec1 CW |
102 | int ret; |
103 | ||
21dd3734 | 104 | ret = i915_gem_wait_for_error(dev); |
76c1dec1 CW |
105 | if (ret) |
106 | return ret; | |
107 | ||
108 | ret = mutex_lock_interruptible(&dev->struct_mutex); | |
109 | if (ret) | |
110 | return ret; | |
111 | ||
23bc5982 | 112 | WARN_ON(i915_verify_lists(dev)); |
76c1dec1 CW |
113 | return 0; |
114 | } | |
30dbf0c0 | 115 | |
7d1c4804 | 116 | static inline bool |
05394f39 | 117 | i915_gem_object_is_inactive(struct drm_i915_gem_object *obj) |
7d1c4804 | 118 | { |
05394f39 | 119 | return obj->gtt_space && !obj->active && obj->pin_count == 0; |
7d1c4804 CW |
120 | } |
121 | ||
79e53945 JB |
122 | int |
123 | i915_gem_init_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 124 | struct drm_file *file) |
79e53945 JB |
125 | { |
126 | struct drm_i915_gem_init *args = data; | |
2021746e CW |
127 | |
128 | if (args->gtt_start >= args->gtt_end || | |
129 | (args->gtt_end | args->gtt_start) & (PAGE_SIZE - 1)) | |
130 | return -EINVAL; | |
79e53945 | 131 | |
f534bc0b DV |
132 | /* GEM with user mode setting was never supported on ilk and later. */ |
133 | if (INTEL_INFO(dev)->gen >= 5) | |
134 | return -ENODEV; | |
135 | ||
79e53945 | 136 | mutex_lock(&dev->struct_mutex); |
644ec02b DV |
137 | i915_gem_init_global_gtt(dev, args->gtt_start, |
138 | args->gtt_end, args->gtt_end); | |
673a394b EA |
139 | mutex_unlock(&dev->struct_mutex); |
140 | ||
2021746e | 141 | return 0; |
673a394b EA |
142 | } |
143 | ||
5a125c3c EA |
144 | int |
145 | i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 146 | struct drm_file *file) |
5a125c3c | 147 | { |
73aa808f | 148 | struct drm_i915_private *dev_priv = dev->dev_private; |
5a125c3c | 149 | struct drm_i915_gem_get_aperture *args = data; |
6299f992 CW |
150 | struct drm_i915_gem_object *obj; |
151 | size_t pinned; | |
5a125c3c EA |
152 | |
153 | if (!(dev->driver->driver_features & DRIVER_GEM)) | |
154 | return -ENODEV; | |
155 | ||
6299f992 | 156 | pinned = 0; |
73aa808f | 157 | mutex_lock(&dev->struct_mutex); |
6299f992 CW |
158 | list_for_each_entry(obj, &dev_priv->mm.pinned_list, mm_list) |
159 | pinned += obj->gtt_space->size; | |
73aa808f | 160 | mutex_unlock(&dev->struct_mutex); |
5a125c3c | 161 | |
6299f992 | 162 | args->aper_size = dev_priv->mm.gtt_total; |
0206e353 | 163 | args->aper_available_size = args->aper_size - pinned; |
6299f992 | 164 | |
5a125c3c EA |
165 | return 0; |
166 | } | |
167 | ||
ff72145b DA |
168 | static int |
169 | i915_gem_create(struct drm_file *file, | |
170 | struct drm_device *dev, | |
171 | uint64_t size, | |
172 | uint32_t *handle_p) | |
673a394b | 173 | { |
05394f39 | 174 | struct drm_i915_gem_object *obj; |
a1a2d1d3 PP |
175 | int ret; |
176 | u32 handle; | |
673a394b | 177 | |
ff72145b | 178 | size = roundup(size, PAGE_SIZE); |
8ffc0246 CW |
179 | if (size == 0) |
180 | return -EINVAL; | |
673a394b EA |
181 | |
182 | /* Allocate the new object */ | |
ff72145b | 183 | obj = i915_gem_alloc_object(dev, size); |
673a394b EA |
184 | if (obj == NULL) |
185 | return -ENOMEM; | |
186 | ||
05394f39 | 187 | ret = drm_gem_handle_create(file, &obj->base, &handle); |
1dfd9754 | 188 | if (ret) { |
05394f39 CW |
189 | drm_gem_object_release(&obj->base); |
190 | i915_gem_info_remove_obj(dev->dev_private, obj->base.size); | |
202f2fef | 191 | kfree(obj); |
673a394b | 192 | return ret; |
1dfd9754 | 193 | } |
673a394b | 194 | |
202f2fef | 195 | /* drop reference from allocate - handle holds it now */ |
05394f39 | 196 | drm_gem_object_unreference(&obj->base); |
202f2fef CW |
197 | trace_i915_gem_object_create(obj); |
198 | ||
ff72145b | 199 | *handle_p = handle; |
673a394b EA |
200 | return 0; |
201 | } | |
202 | ||
ff72145b DA |
203 | int |
204 | i915_gem_dumb_create(struct drm_file *file, | |
205 | struct drm_device *dev, | |
206 | struct drm_mode_create_dumb *args) | |
207 | { | |
208 | /* have to work out size/pitch and return them */ | |
ed0291fd | 209 | args->pitch = ALIGN(args->width * ((args->bpp + 7) / 8), 64); |
ff72145b DA |
210 | args->size = args->pitch * args->height; |
211 | return i915_gem_create(file, dev, | |
212 | args->size, &args->handle); | |
213 | } | |
214 | ||
215 | int i915_gem_dumb_destroy(struct drm_file *file, | |
216 | struct drm_device *dev, | |
217 | uint32_t handle) | |
218 | { | |
219 | return drm_gem_handle_delete(file, handle); | |
220 | } | |
221 | ||
222 | /** | |
223 | * Creates a new mm object and returns a handle to it. | |
224 | */ | |
225 | int | |
226 | i915_gem_create_ioctl(struct drm_device *dev, void *data, | |
227 | struct drm_file *file) | |
228 | { | |
229 | struct drm_i915_gem_create *args = data; | |
230 | return i915_gem_create(file, dev, | |
231 | args->size, &args->handle); | |
232 | } | |
233 | ||
05394f39 | 234 | static int i915_gem_object_needs_bit17_swizzle(struct drm_i915_gem_object *obj) |
280b713b | 235 | { |
05394f39 | 236 | drm_i915_private_t *dev_priv = obj->base.dev->dev_private; |
280b713b EA |
237 | |
238 | return dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_9_10_17 && | |
05394f39 | 239 | obj->tiling_mode != I915_TILING_NONE; |
280b713b EA |
240 | } |
241 | ||
8461d226 DV |
242 | static inline int |
243 | __copy_to_user_swizzled(char __user *cpu_vaddr, | |
244 | const char *gpu_vaddr, int gpu_offset, | |
245 | int length) | |
246 | { | |
247 | int ret, cpu_offset = 0; | |
248 | ||
249 | while (length > 0) { | |
250 | int cacheline_end = ALIGN(gpu_offset + 1, 64); | |
251 | int this_length = min(cacheline_end - gpu_offset, length); | |
252 | int swizzled_gpu_offset = gpu_offset ^ 64; | |
253 | ||
254 | ret = __copy_to_user(cpu_vaddr + cpu_offset, | |
255 | gpu_vaddr + swizzled_gpu_offset, | |
256 | this_length); | |
257 | if (ret) | |
258 | return ret + length; | |
259 | ||
260 | cpu_offset += this_length; | |
261 | gpu_offset += this_length; | |
262 | length -= this_length; | |
263 | } | |
264 | ||
265 | return 0; | |
266 | } | |
267 | ||
8c59967c DV |
268 | static inline int |
269 | __copy_from_user_swizzled(char __user *gpu_vaddr, int gpu_offset, | |
270 | const char *cpu_vaddr, | |
271 | int length) | |
272 | { | |
273 | int ret, cpu_offset = 0; | |
274 | ||
275 | while (length > 0) { | |
276 | int cacheline_end = ALIGN(gpu_offset + 1, 64); | |
277 | int this_length = min(cacheline_end - gpu_offset, length); | |
278 | int swizzled_gpu_offset = gpu_offset ^ 64; | |
279 | ||
280 | ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset, | |
281 | cpu_vaddr + cpu_offset, | |
282 | this_length); | |
283 | if (ret) | |
284 | return ret + length; | |
285 | ||
286 | cpu_offset += this_length; | |
287 | gpu_offset += this_length; | |
288 | length -= this_length; | |
289 | } | |
290 | ||
291 | return 0; | |
292 | } | |
293 | ||
d174bd64 DV |
294 | /* Per-page copy function for the shmem pread fastpath. |
295 | * Flushes invalid cachelines before reading the target if | |
296 | * needs_clflush is set. */ | |
eb01459f | 297 | static int |
d174bd64 DV |
298 | shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length, |
299 | char __user *user_data, | |
300 | bool page_do_bit17_swizzling, bool needs_clflush) | |
301 | { | |
302 | char *vaddr; | |
303 | int ret; | |
304 | ||
e7e58eb5 | 305 | if (unlikely(page_do_bit17_swizzling)) |
d174bd64 DV |
306 | return -EINVAL; |
307 | ||
308 | vaddr = kmap_atomic(page); | |
309 | if (needs_clflush) | |
310 | drm_clflush_virt_range(vaddr + shmem_page_offset, | |
311 | page_length); | |
312 | ret = __copy_to_user_inatomic(user_data, | |
313 | vaddr + shmem_page_offset, | |
314 | page_length); | |
315 | kunmap_atomic(vaddr); | |
316 | ||
317 | return ret; | |
318 | } | |
319 | ||
23c18c71 DV |
320 | static void |
321 | shmem_clflush_swizzled_range(char *addr, unsigned long length, | |
322 | bool swizzled) | |
323 | { | |
e7e58eb5 | 324 | if (unlikely(swizzled)) { |
23c18c71 DV |
325 | unsigned long start = (unsigned long) addr; |
326 | unsigned long end = (unsigned long) addr + length; | |
327 | ||
328 | /* For swizzling simply ensure that we always flush both | |
329 | * channels. Lame, but simple and it works. Swizzled | |
330 | * pwrite/pread is far from a hotpath - current userspace | |
331 | * doesn't use it at all. */ | |
332 | start = round_down(start, 128); | |
333 | end = round_up(end, 128); | |
334 | ||
335 | drm_clflush_virt_range((void *)start, end - start); | |
336 | } else { | |
337 | drm_clflush_virt_range(addr, length); | |
338 | } | |
339 | ||
340 | } | |
341 | ||
d174bd64 DV |
342 | /* Only difference to the fast-path function is that this can handle bit17 |
343 | * and uses non-atomic copy and kmap functions. */ | |
344 | static int | |
345 | shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length, | |
346 | char __user *user_data, | |
347 | bool page_do_bit17_swizzling, bool needs_clflush) | |
348 | { | |
349 | char *vaddr; | |
350 | int ret; | |
351 | ||
352 | vaddr = kmap(page); | |
353 | if (needs_clflush) | |
23c18c71 DV |
354 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
355 | page_length, | |
356 | page_do_bit17_swizzling); | |
d174bd64 DV |
357 | |
358 | if (page_do_bit17_swizzling) | |
359 | ret = __copy_to_user_swizzled(user_data, | |
360 | vaddr, shmem_page_offset, | |
361 | page_length); | |
362 | else | |
363 | ret = __copy_to_user(user_data, | |
364 | vaddr + shmem_page_offset, | |
365 | page_length); | |
366 | kunmap(page); | |
367 | ||
368 | return ret; | |
369 | } | |
370 | ||
eb01459f | 371 | static int |
dbf7bff0 DV |
372 | i915_gem_shmem_pread(struct drm_device *dev, |
373 | struct drm_i915_gem_object *obj, | |
374 | struct drm_i915_gem_pread *args, | |
375 | struct drm_file *file) | |
eb01459f | 376 | { |
05394f39 | 377 | struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping; |
8461d226 | 378 | char __user *user_data; |
eb01459f | 379 | ssize_t remain; |
8461d226 | 380 | loff_t offset; |
eb2c0c81 | 381 | int shmem_page_offset, page_length, ret = 0; |
8461d226 | 382 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; |
dbf7bff0 | 383 | int hit_slowpath = 0; |
96d79b52 | 384 | int prefaulted = 0; |
8489731c | 385 | int needs_clflush = 0; |
692a576b | 386 | int release_page; |
eb01459f | 387 | |
8461d226 | 388 | user_data = (char __user *) (uintptr_t) args->data_ptr; |
eb01459f EA |
389 | remain = args->size; |
390 | ||
8461d226 | 391 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); |
eb01459f | 392 | |
8489731c DV |
393 | if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) { |
394 | /* If we're not in the cpu read domain, set ourself into the gtt | |
395 | * read domain and manually flush cachelines (if required). This | |
396 | * optimizes for the case when the gpu will dirty the data | |
397 | * anyway again before the next pread happens. */ | |
398 | if (obj->cache_level == I915_CACHE_NONE) | |
399 | needs_clflush = 1; | |
400 | ret = i915_gem_object_set_to_gtt_domain(obj, false); | |
401 | if (ret) | |
402 | return ret; | |
403 | } | |
eb01459f | 404 | |
8461d226 | 405 | offset = args->offset; |
eb01459f EA |
406 | |
407 | while (remain > 0) { | |
e5281ccd CW |
408 | struct page *page; |
409 | ||
eb01459f EA |
410 | /* Operation in this page |
411 | * | |
eb01459f | 412 | * shmem_page_offset = offset within page in shmem file |
eb01459f EA |
413 | * page_length = bytes to copy for this page |
414 | */ | |
c8cbbb8b | 415 | shmem_page_offset = offset_in_page(offset); |
eb01459f EA |
416 | page_length = remain; |
417 | if ((shmem_page_offset + page_length) > PAGE_SIZE) | |
418 | page_length = PAGE_SIZE - shmem_page_offset; | |
eb01459f | 419 | |
692a576b DV |
420 | if (obj->pages) { |
421 | page = obj->pages[offset >> PAGE_SHIFT]; | |
422 | release_page = 0; | |
423 | } else { | |
424 | page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT); | |
425 | if (IS_ERR(page)) { | |
426 | ret = PTR_ERR(page); | |
427 | goto out; | |
428 | } | |
429 | release_page = 1; | |
b65552f0 | 430 | } |
e5281ccd | 431 | |
8461d226 DV |
432 | page_do_bit17_swizzling = obj_do_bit17_swizzling && |
433 | (page_to_phys(page) & (1 << 17)) != 0; | |
434 | ||
d174bd64 DV |
435 | ret = shmem_pread_fast(page, shmem_page_offset, page_length, |
436 | user_data, page_do_bit17_swizzling, | |
437 | needs_clflush); | |
438 | if (ret == 0) | |
439 | goto next_page; | |
dbf7bff0 DV |
440 | |
441 | hit_slowpath = 1; | |
692a576b | 442 | page_cache_get(page); |
dbf7bff0 DV |
443 | mutex_unlock(&dev->struct_mutex); |
444 | ||
96d79b52 | 445 | if (!prefaulted) { |
f56f821f | 446 | ret = fault_in_multipages_writeable(user_data, remain); |
96d79b52 DV |
447 | /* Userspace is tricking us, but we've already clobbered |
448 | * its pages with the prefault and promised to write the | |
449 | * data up to the first fault. Hence ignore any errors | |
450 | * and just continue. */ | |
451 | (void)ret; | |
452 | prefaulted = 1; | |
453 | } | |
eb01459f | 454 | |
d174bd64 DV |
455 | ret = shmem_pread_slow(page, shmem_page_offset, page_length, |
456 | user_data, page_do_bit17_swizzling, | |
457 | needs_clflush); | |
eb01459f | 458 | |
dbf7bff0 | 459 | mutex_lock(&dev->struct_mutex); |
e5281ccd | 460 | page_cache_release(page); |
dbf7bff0 | 461 | next_page: |
e5281ccd | 462 | mark_page_accessed(page); |
692a576b DV |
463 | if (release_page) |
464 | page_cache_release(page); | |
e5281ccd | 465 | |
8461d226 DV |
466 | if (ret) { |
467 | ret = -EFAULT; | |
468 | goto out; | |
469 | } | |
470 | ||
eb01459f | 471 | remain -= page_length; |
8461d226 | 472 | user_data += page_length; |
eb01459f EA |
473 | offset += page_length; |
474 | } | |
475 | ||
4f27b75d | 476 | out: |
dbf7bff0 DV |
477 | if (hit_slowpath) { |
478 | /* Fixup: Kill any reinstated backing storage pages */ | |
479 | if (obj->madv == __I915_MADV_PURGED) | |
480 | i915_gem_object_truncate(obj); | |
481 | } | |
eb01459f EA |
482 | |
483 | return ret; | |
484 | } | |
485 | ||
673a394b EA |
486 | /** |
487 | * Reads data from the object referenced by handle. | |
488 | * | |
489 | * On error, the contents of *data are undefined. | |
490 | */ | |
491 | int | |
492 | i915_gem_pread_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 493 | struct drm_file *file) |
673a394b EA |
494 | { |
495 | struct drm_i915_gem_pread *args = data; | |
05394f39 | 496 | struct drm_i915_gem_object *obj; |
35b62a89 | 497 | int ret = 0; |
673a394b | 498 | |
51311d0a CW |
499 | if (args->size == 0) |
500 | return 0; | |
501 | ||
502 | if (!access_ok(VERIFY_WRITE, | |
503 | (char __user *)(uintptr_t)args->data_ptr, | |
504 | args->size)) | |
505 | return -EFAULT; | |
506 | ||
4f27b75d | 507 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 508 | if (ret) |
4f27b75d | 509 | return ret; |
673a394b | 510 | |
05394f39 | 511 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 512 | if (&obj->base == NULL) { |
1d7cfea1 CW |
513 | ret = -ENOENT; |
514 | goto unlock; | |
4f27b75d | 515 | } |
673a394b | 516 | |
7dcd2499 | 517 | /* Bounds check source. */ |
05394f39 CW |
518 | if (args->offset > obj->base.size || |
519 | args->size > obj->base.size - args->offset) { | |
ce9d419d | 520 | ret = -EINVAL; |
35b62a89 | 521 | goto out; |
ce9d419d CW |
522 | } |
523 | ||
db53a302 CW |
524 | trace_i915_gem_object_pread(obj, args->offset, args->size); |
525 | ||
dbf7bff0 | 526 | ret = i915_gem_shmem_pread(dev, obj, args, file); |
673a394b | 527 | |
35b62a89 | 528 | out: |
05394f39 | 529 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 530 | unlock: |
4f27b75d | 531 | mutex_unlock(&dev->struct_mutex); |
eb01459f | 532 | return ret; |
673a394b EA |
533 | } |
534 | ||
0839ccb8 KP |
535 | /* This is the fast write path which cannot handle |
536 | * page faults in the source data | |
9b7530cc | 537 | */ |
0839ccb8 KP |
538 | |
539 | static inline int | |
540 | fast_user_write(struct io_mapping *mapping, | |
541 | loff_t page_base, int page_offset, | |
542 | char __user *user_data, | |
543 | int length) | |
9b7530cc | 544 | { |
9b7530cc | 545 | char *vaddr_atomic; |
0839ccb8 | 546 | unsigned long unwritten; |
9b7530cc | 547 | |
3e4d3af5 | 548 | vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base); |
0839ccb8 KP |
549 | unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + page_offset, |
550 | user_data, length); | |
3e4d3af5 | 551 | io_mapping_unmap_atomic(vaddr_atomic); |
fbd5a26d | 552 | return unwritten; |
0839ccb8 KP |
553 | } |
554 | ||
3de09aa3 EA |
555 | /** |
556 | * This is the fast pwrite path, where we copy the data directly from the | |
557 | * user into the GTT, uncached. | |
558 | */ | |
673a394b | 559 | static int |
05394f39 CW |
560 | i915_gem_gtt_pwrite_fast(struct drm_device *dev, |
561 | struct drm_i915_gem_object *obj, | |
3de09aa3 | 562 | struct drm_i915_gem_pwrite *args, |
05394f39 | 563 | struct drm_file *file) |
673a394b | 564 | { |
0839ccb8 | 565 | drm_i915_private_t *dev_priv = dev->dev_private; |
673a394b | 566 | ssize_t remain; |
0839ccb8 | 567 | loff_t offset, page_base; |
673a394b | 568 | char __user *user_data; |
935aaa69 DV |
569 | int page_offset, page_length, ret; |
570 | ||
571 | ret = i915_gem_object_pin(obj, 0, true); | |
572 | if (ret) | |
573 | goto out; | |
574 | ||
575 | ret = i915_gem_object_set_to_gtt_domain(obj, true); | |
576 | if (ret) | |
577 | goto out_unpin; | |
578 | ||
579 | ret = i915_gem_object_put_fence(obj); | |
580 | if (ret) | |
581 | goto out_unpin; | |
673a394b EA |
582 | |
583 | user_data = (char __user *) (uintptr_t) args->data_ptr; | |
584 | remain = args->size; | |
673a394b | 585 | |
05394f39 | 586 | offset = obj->gtt_offset + args->offset; |
673a394b EA |
587 | |
588 | while (remain > 0) { | |
589 | /* Operation in this page | |
590 | * | |
0839ccb8 KP |
591 | * page_base = page offset within aperture |
592 | * page_offset = offset within page | |
593 | * page_length = bytes to copy for this page | |
673a394b | 594 | */ |
c8cbbb8b CW |
595 | page_base = offset & PAGE_MASK; |
596 | page_offset = offset_in_page(offset); | |
0839ccb8 KP |
597 | page_length = remain; |
598 | if ((page_offset + remain) > PAGE_SIZE) | |
599 | page_length = PAGE_SIZE - page_offset; | |
600 | ||
0839ccb8 | 601 | /* If we get a fault while copying data, then (presumably) our |
3de09aa3 EA |
602 | * source page isn't available. Return the error and we'll |
603 | * retry in the slow path. | |
0839ccb8 | 604 | */ |
fbd5a26d | 605 | if (fast_user_write(dev_priv->mm.gtt_mapping, page_base, |
935aaa69 DV |
606 | page_offset, user_data, page_length)) { |
607 | ret = -EFAULT; | |
608 | goto out_unpin; | |
609 | } | |
673a394b | 610 | |
0839ccb8 KP |
611 | remain -= page_length; |
612 | user_data += page_length; | |
613 | offset += page_length; | |
673a394b | 614 | } |
673a394b | 615 | |
935aaa69 DV |
616 | out_unpin: |
617 | i915_gem_object_unpin(obj); | |
618 | out: | |
3de09aa3 | 619 | return ret; |
673a394b EA |
620 | } |
621 | ||
d174bd64 DV |
622 | /* Per-page copy function for the shmem pwrite fastpath. |
623 | * Flushes invalid cachelines before writing to the target if | |
624 | * needs_clflush_before is set and flushes out any written cachelines after | |
625 | * writing if needs_clflush is set. */ | |
3043c60c | 626 | static int |
d174bd64 DV |
627 | shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length, |
628 | char __user *user_data, | |
629 | bool page_do_bit17_swizzling, | |
630 | bool needs_clflush_before, | |
631 | bool needs_clflush_after) | |
673a394b | 632 | { |
d174bd64 | 633 | char *vaddr; |
673a394b | 634 | int ret; |
3de09aa3 | 635 | |
e7e58eb5 | 636 | if (unlikely(page_do_bit17_swizzling)) |
d174bd64 | 637 | return -EINVAL; |
3de09aa3 | 638 | |
d174bd64 DV |
639 | vaddr = kmap_atomic(page); |
640 | if (needs_clflush_before) | |
641 | drm_clflush_virt_range(vaddr + shmem_page_offset, | |
642 | page_length); | |
643 | ret = __copy_from_user_inatomic_nocache(vaddr + shmem_page_offset, | |
644 | user_data, | |
645 | page_length); | |
646 | if (needs_clflush_after) | |
647 | drm_clflush_virt_range(vaddr + shmem_page_offset, | |
648 | page_length); | |
649 | kunmap_atomic(vaddr); | |
3de09aa3 EA |
650 | |
651 | return ret; | |
652 | } | |
653 | ||
d174bd64 DV |
654 | /* Only difference to the fast-path function is that this can handle bit17 |
655 | * and uses non-atomic copy and kmap functions. */ | |
3043c60c | 656 | static int |
d174bd64 DV |
657 | shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length, |
658 | char __user *user_data, | |
659 | bool page_do_bit17_swizzling, | |
660 | bool needs_clflush_before, | |
661 | bool needs_clflush_after) | |
673a394b | 662 | { |
d174bd64 DV |
663 | char *vaddr; |
664 | int ret; | |
e5281ccd | 665 | |
d174bd64 | 666 | vaddr = kmap(page); |
e7e58eb5 | 667 | if (unlikely(needs_clflush_before || page_do_bit17_swizzling)) |
23c18c71 DV |
668 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
669 | page_length, | |
670 | page_do_bit17_swizzling); | |
d174bd64 DV |
671 | if (page_do_bit17_swizzling) |
672 | ret = __copy_from_user_swizzled(vaddr, shmem_page_offset, | |
e5281ccd CW |
673 | user_data, |
674 | page_length); | |
d174bd64 DV |
675 | else |
676 | ret = __copy_from_user(vaddr + shmem_page_offset, | |
677 | user_data, | |
678 | page_length); | |
679 | if (needs_clflush_after) | |
23c18c71 DV |
680 | shmem_clflush_swizzled_range(vaddr + shmem_page_offset, |
681 | page_length, | |
682 | page_do_bit17_swizzling); | |
d174bd64 | 683 | kunmap(page); |
40123c1f | 684 | |
d174bd64 | 685 | return ret; |
40123c1f EA |
686 | } |
687 | ||
40123c1f | 688 | static int |
e244a443 DV |
689 | i915_gem_shmem_pwrite(struct drm_device *dev, |
690 | struct drm_i915_gem_object *obj, | |
691 | struct drm_i915_gem_pwrite *args, | |
692 | struct drm_file *file) | |
40123c1f | 693 | { |
05394f39 | 694 | struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping; |
40123c1f | 695 | ssize_t remain; |
8c59967c DV |
696 | loff_t offset; |
697 | char __user *user_data; | |
eb2c0c81 | 698 | int shmem_page_offset, page_length, ret = 0; |
8c59967c | 699 | int obj_do_bit17_swizzling, page_do_bit17_swizzling; |
e244a443 | 700 | int hit_slowpath = 0; |
58642885 DV |
701 | int needs_clflush_after = 0; |
702 | int needs_clflush_before = 0; | |
692a576b | 703 | int release_page; |
40123c1f | 704 | |
8c59967c | 705 | user_data = (char __user *) (uintptr_t) args->data_ptr; |
40123c1f EA |
706 | remain = args->size; |
707 | ||
8c59967c | 708 | obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); |
40123c1f | 709 | |
58642885 DV |
710 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) { |
711 | /* If we're not in the cpu write domain, set ourself into the gtt | |
712 | * write domain and manually flush cachelines (if required). This | |
713 | * optimizes for the case when the gpu will use the data | |
714 | * right away and we therefore have to clflush anyway. */ | |
715 | if (obj->cache_level == I915_CACHE_NONE) | |
716 | needs_clflush_after = 1; | |
717 | ret = i915_gem_object_set_to_gtt_domain(obj, true); | |
718 | if (ret) | |
719 | return ret; | |
720 | } | |
721 | /* Same trick applies for invalidate partially written cachelines before | |
722 | * writing. */ | |
723 | if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU) | |
724 | && obj->cache_level == I915_CACHE_NONE) | |
725 | needs_clflush_before = 1; | |
726 | ||
673a394b | 727 | offset = args->offset; |
05394f39 | 728 | obj->dirty = 1; |
673a394b | 729 | |
40123c1f | 730 | while (remain > 0) { |
e5281ccd | 731 | struct page *page; |
58642885 | 732 | int partial_cacheline_write; |
e5281ccd | 733 | |
40123c1f EA |
734 | /* Operation in this page |
735 | * | |
40123c1f | 736 | * shmem_page_offset = offset within page in shmem file |
40123c1f EA |
737 | * page_length = bytes to copy for this page |
738 | */ | |
c8cbbb8b | 739 | shmem_page_offset = offset_in_page(offset); |
40123c1f EA |
740 | |
741 | page_length = remain; | |
742 | if ((shmem_page_offset + page_length) > PAGE_SIZE) | |
743 | page_length = PAGE_SIZE - shmem_page_offset; | |
40123c1f | 744 | |
58642885 DV |
745 | /* If we don't overwrite a cacheline completely we need to be |
746 | * careful to have up-to-date data by first clflushing. Don't | |
747 | * overcomplicate things and flush the entire patch. */ | |
748 | partial_cacheline_write = needs_clflush_before && | |
749 | ((shmem_page_offset | page_length) | |
750 | & (boot_cpu_data.x86_clflush_size - 1)); | |
751 | ||
692a576b DV |
752 | if (obj->pages) { |
753 | page = obj->pages[offset >> PAGE_SHIFT]; | |
754 | release_page = 0; | |
755 | } else { | |
756 | page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT); | |
757 | if (IS_ERR(page)) { | |
758 | ret = PTR_ERR(page); | |
759 | goto out; | |
760 | } | |
761 | release_page = 1; | |
e5281ccd CW |
762 | } |
763 | ||
8c59967c DV |
764 | page_do_bit17_swizzling = obj_do_bit17_swizzling && |
765 | (page_to_phys(page) & (1 << 17)) != 0; | |
766 | ||
d174bd64 DV |
767 | ret = shmem_pwrite_fast(page, shmem_page_offset, page_length, |
768 | user_data, page_do_bit17_swizzling, | |
769 | partial_cacheline_write, | |
770 | needs_clflush_after); | |
771 | if (ret == 0) | |
772 | goto next_page; | |
e244a443 DV |
773 | |
774 | hit_slowpath = 1; | |
692a576b | 775 | page_cache_get(page); |
e244a443 DV |
776 | mutex_unlock(&dev->struct_mutex); |
777 | ||
d174bd64 DV |
778 | ret = shmem_pwrite_slow(page, shmem_page_offset, page_length, |
779 | user_data, page_do_bit17_swizzling, | |
780 | partial_cacheline_write, | |
781 | needs_clflush_after); | |
40123c1f | 782 | |
e244a443 | 783 | mutex_lock(&dev->struct_mutex); |
692a576b | 784 | page_cache_release(page); |
e244a443 | 785 | next_page: |
e5281ccd CW |
786 | set_page_dirty(page); |
787 | mark_page_accessed(page); | |
692a576b DV |
788 | if (release_page) |
789 | page_cache_release(page); | |
e5281ccd | 790 | |
8c59967c DV |
791 | if (ret) { |
792 | ret = -EFAULT; | |
793 | goto out; | |
794 | } | |
795 | ||
40123c1f | 796 | remain -= page_length; |
8c59967c | 797 | user_data += page_length; |
40123c1f | 798 | offset += page_length; |
673a394b EA |
799 | } |
800 | ||
fbd5a26d | 801 | out: |
e244a443 DV |
802 | if (hit_slowpath) { |
803 | /* Fixup: Kill any reinstated backing storage pages */ | |
804 | if (obj->madv == __I915_MADV_PURGED) | |
805 | i915_gem_object_truncate(obj); | |
806 | /* and flush dirty cachelines in case the object isn't in the cpu write | |
807 | * domain anymore. */ | |
808 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) { | |
809 | i915_gem_clflush_object(obj); | |
810 | intel_gtt_chipset_flush(); | |
811 | } | |
8c59967c | 812 | } |
673a394b | 813 | |
58642885 DV |
814 | if (needs_clflush_after) |
815 | intel_gtt_chipset_flush(); | |
816 | ||
40123c1f | 817 | return ret; |
673a394b EA |
818 | } |
819 | ||
820 | /** | |
821 | * Writes data to the object referenced by handle. | |
822 | * | |
823 | * On error, the contents of the buffer that were to be modified are undefined. | |
824 | */ | |
825 | int | |
826 | i915_gem_pwrite_ioctl(struct drm_device *dev, void *data, | |
fbd5a26d | 827 | struct drm_file *file) |
673a394b EA |
828 | { |
829 | struct drm_i915_gem_pwrite *args = data; | |
05394f39 | 830 | struct drm_i915_gem_object *obj; |
51311d0a CW |
831 | int ret; |
832 | ||
833 | if (args->size == 0) | |
834 | return 0; | |
835 | ||
836 | if (!access_ok(VERIFY_READ, | |
837 | (char __user *)(uintptr_t)args->data_ptr, | |
838 | args->size)) | |
839 | return -EFAULT; | |
840 | ||
f56f821f DV |
841 | ret = fault_in_multipages_readable((char __user *)(uintptr_t)args->data_ptr, |
842 | args->size); | |
51311d0a CW |
843 | if (ret) |
844 | return -EFAULT; | |
673a394b | 845 | |
fbd5a26d | 846 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 847 | if (ret) |
fbd5a26d | 848 | return ret; |
1d7cfea1 | 849 | |
05394f39 | 850 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 851 | if (&obj->base == NULL) { |
1d7cfea1 CW |
852 | ret = -ENOENT; |
853 | goto unlock; | |
fbd5a26d | 854 | } |
673a394b | 855 | |
7dcd2499 | 856 | /* Bounds check destination. */ |
05394f39 CW |
857 | if (args->offset > obj->base.size || |
858 | args->size > obj->base.size - args->offset) { | |
ce9d419d | 859 | ret = -EINVAL; |
35b62a89 | 860 | goto out; |
ce9d419d CW |
861 | } |
862 | ||
db53a302 CW |
863 | trace_i915_gem_object_pwrite(obj, args->offset, args->size); |
864 | ||
935aaa69 | 865 | ret = -EFAULT; |
673a394b EA |
866 | /* We can only do the GTT pwrite on untiled buffers, as otherwise |
867 | * it would end up going through the fenced access, and we'll get | |
868 | * different detiling behavior between reading and writing. | |
869 | * pread/pwrite currently are reading and writing from the CPU | |
870 | * perspective, requiring manual detiling by the client. | |
871 | */ | |
5c0480f2 | 872 | if (obj->phys_obj) { |
fbd5a26d | 873 | ret = i915_gem_phys_pwrite(dev, obj, args, file); |
5c0480f2 DV |
874 | goto out; |
875 | } | |
876 | ||
877 | if (obj->gtt_space && | |
3ae53783 | 878 | obj->cache_level == I915_CACHE_NONE && |
c07496fa | 879 | obj->tiling_mode == I915_TILING_NONE && |
ffc62976 | 880 | obj->map_and_fenceable && |
5c0480f2 | 881 | obj->base.write_domain != I915_GEM_DOMAIN_CPU) { |
fbd5a26d | 882 | ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file); |
935aaa69 DV |
883 | /* Note that the gtt paths might fail with non-page-backed user |
884 | * pointers (e.g. gtt mappings when moving data between | |
885 | * textures). Fallback to the shmem path in that case. */ | |
fbd5a26d | 886 | } |
673a394b | 887 | |
5c0480f2 | 888 | if (ret == -EFAULT) |
935aaa69 | 889 | ret = i915_gem_shmem_pwrite(dev, obj, args, file); |
5c0480f2 | 890 | |
35b62a89 | 891 | out: |
05394f39 | 892 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 893 | unlock: |
fbd5a26d | 894 | mutex_unlock(&dev->struct_mutex); |
673a394b EA |
895 | return ret; |
896 | } | |
897 | ||
898 | /** | |
2ef7eeaa EA |
899 | * Called when user space prepares to use an object with the CPU, either |
900 | * through the mmap ioctl's mapping or a GTT mapping. | |
673a394b EA |
901 | */ |
902 | int | |
903 | i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 904 | struct drm_file *file) |
673a394b EA |
905 | { |
906 | struct drm_i915_gem_set_domain *args = data; | |
05394f39 | 907 | struct drm_i915_gem_object *obj; |
2ef7eeaa EA |
908 | uint32_t read_domains = args->read_domains; |
909 | uint32_t write_domain = args->write_domain; | |
673a394b EA |
910 | int ret; |
911 | ||
912 | if (!(dev->driver->driver_features & DRIVER_GEM)) | |
913 | return -ENODEV; | |
914 | ||
2ef7eeaa | 915 | /* Only handle setting domains to types used by the CPU. */ |
21d509e3 | 916 | if (write_domain & I915_GEM_GPU_DOMAINS) |
2ef7eeaa EA |
917 | return -EINVAL; |
918 | ||
21d509e3 | 919 | if (read_domains & I915_GEM_GPU_DOMAINS) |
2ef7eeaa EA |
920 | return -EINVAL; |
921 | ||
922 | /* Having something in the write domain implies it's in the read | |
923 | * domain, and only that read domain. Enforce that in the request. | |
924 | */ | |
925 | if (write_domain != 0 && read_domains != write_domain) | |
926 | return -EINVAL; | |
927 | ||
76c1dec1 | 928 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 929 | if (ret) |
76c1dec1 | 930 | return ret; |
1d7cfea1 | 931 | |
05394f39 | 932 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 933 | if (&obj->base == NULL) { |
1d7cfea1 CW |
934 | ret = -ENOENT; |
935 | goto unlock; | |
76c1dec1 | 936 | } |
673a394b | 937 | |
2ef7eeaa EA |
938 | if (read_domains & I915_GEM_DOMAIN_GTT) { |
939 | ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0); | |
02354392 EA |
940 | |
941 | /* Silently promote "you're not bound, there was nothing to do" | |
942 | * to success, since the client was just asking us to | |
943 | * make sure everything was done. | |
944 | */ | |
945 | if (ret == -EINVAL) | |
946 | ret = 0; | |
2ef7eeaa | 947 | } else { |
e47c68e9 | 948 | ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0); |
2ef7eeaa EA |
949 | } |
950 | ||
05394f39 | 951 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 952 | unlock: |
673a394b EA |
953 | mutex_unlock(&dev->struct_mutex); |
954 | return ret; | |
955 | } | |
956 | ||
957 | /** | |
958 | * Called when user space has done writes to this buffer | |
959 | */ | |
960 | int | |
961 | i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 962 | struct drm_file *file) |
673a394b EA |
963 | { |
964 | struct drm_i915_gem_sw_finish *args = data; | |
05394f39 | 965 | struct drm_i915_gem_object *obj; |
673a394b EA |
966 | int ret = 0; |
967 | ||
968 | if (!(dev->driver->driver_features & DRIVER_GEM)) | |
969 | return -ENODEV; | |
970 | ||
76c1dec1 | 971 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 972 | if (ret) |
76c1dec1 | 973 | return ret; |
1d7cfea1 | 974 | |
05394f39 | 975 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 976 | if (&obj->base == NULL) { |
1d7cfea1 CW |
977 | ret = -ENOENT; |
978 | goto unlock; | |
673a394b EA |
979 | } |
980 | ||
673a394b | 981 | /* Pinned buffers may be scanout, so flush the cache */ |
05394f39 | 982 | if (obj->pin_count) |
e47c68e9 EA |
983 | i915_gem_object_flush_cpu_write_domain(obj); |
984 | ||
05394f39 | 985 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 986 | unlock: |
673a394b EA |
987 | mutex_unlock(&dev->struct_mutex); |
988 | return ret; | |
989 | } | |
990 | ||
991 | /** | |
992 | * Maps the contents of an object, returning the address it is mapped | |
993 | * into. | |
994 | * | |
995 | * While the mapping holds a reference on the contents of the object, it doesn't | |
996 | * imply a ref on the object itself. | |
997 | */ | |
998 | int | |
999 | i915_gem_mmap_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 1000 | struct drm_file *file) |
673a394b EA |
1001 | { |
1002 | struct drm_i915_gem_mmap *args = data; | |
1003 | struct drm_gem_object *obj; | |
673a394b EA |
1004 | unsigned long addr; |
1005 | ||
1006 | if (!(dev->driver->driver_features & DRIVER_GEM)) | |
1007 | return -ENODEV; | |
1008 | ||
05394f39 | 1009 | obj = drm_gem_object_lookup(dev, file, args->handle); |
673a394b | 1010 | if (obj == NULL) |
bf79cb91 | 1011 | return -ENOENT; |
673a394b | 1012 | |
673a394b EA |
1013 | down_write(¤t->mm->mmap_sem); |
1014 | addr = do_mmap(obj->filp, 0, args->size, | |
1015 | PROT_READ | PROT_WRITE, MAP_SHARED, | |
1016 | args->offset); | |
1017 | up_write(¤t->mm->mmap_sem); | |
bc9025bd | 1018 | drm_gem_object_unreference_unlocked(obj); |
673a394b EA |
1019 | if (IS_ERR((void *)addr)) |
1020 | return addr; | |
1021 | ||
1022 | args->addr_ptr = (uint64_t) addr; | |
1023 | ||
1024 | return 0; | |
1025 | } | |
1026 | ||
de151cf6 JB |
1027 | /** |
1028 | * i915_gem_fault - fault a page into the GTT | |
1029 | * vma: VMA in question | |
1030 | * vmf: fault info | |
1031 | * | |
1032 | * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped | |
1033 | * from userspace. The fault handler takes care of binding the object to | |
1034 | * the GTT (if needed), allocating and programming a fence register (again, | |
1035 | * only if needed based on whether the old reg is still valid or the object | |
1036 | * is tiled) and inserting a new PTE into the faulting process. | |
1037 | * | |
1038 | * Note that the faulting process may involve evicting existing objects | |
1039 | * from the GTT and/or fence registers to make room. So performance may | |
1040 | * suffer if the GTT working set is large or there are few fence registers | |
1041 | * left. | |
1042 | */ | |
1043 | int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf) | |
1044 | { | |
05394f39 CW |
1045 | struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data); |
1046 | struct drm_device *dev = obj->base.dev; | |
7d1c4804 | 1047 | drm_i915_private_t *dev_priv = dev->dev_private; |
de151cf6 JB |
1048 | pgoff_t page_offset; |
1049 | unsigned long pfn; | |
1050 | int ret = 0; | |
0f973f27 | 1051 | bool write = !!(vmf->flags & FAULT_FLAG_WRITE); |
de151cf6 JB |
1052 | |
1053 | /* We don't use vmf->pgoff since that has the fake offset */ | |
1054 | page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >> | |
1055 | PAGE_SHIFT; | |
1056 | ||
d9bc7e9f CW |
1057 | ret = i915_mutex_lock_interruptible(dev); |
1058 | if (ret) | |
1059 | goto out; | |
a00b10c3 | 1060 | |
db53a302 CW |
1061 | trace_i915_gem_object_fault(obj, page_offset, true, write); |
1062 | ||
d9bc7e9f | 1063 | /* Now bind it into the GTT if needed */ |
919926ae CW |
1064 | if (!obj->map_and_fenceable) { |
1065 | ret = i915_gem_object_unbind(obj); | |
1066 | if (ret) | |
1067 | goto unlock; | |
a00b10c3 | 1068 | } |
05394f39 | 1069 | if (!obj->gtt_space) { |
75e9e915 | 1070 | ret = i915_gem_object_bind_to_gtt(obj, 0, true); |
c715089f CW |
1071 | if (ret) |
1072 | goto unlock; | |
de151cf6 | 1073 | |
e92d03bf EA |
1074 | ret = i915_gem_object_set_to_gtt_domain(obj, write); |
1075 | if (ret) | |
1076 | goto unlock; | |
1077 | } | |
4a684a41 | 1078 | |
74898d7e DV |
1079 | if (!obj->has_global_gtt_mapping) |
1080 | i915_gem_gtt_bind_object(obj, obj->cache_level); | |
1081 | ||
9a5a53b3 | 1082 | ret = i915_gem_object_get_fence(obj, NULL); |
d9e86c0e CW |
1083 | if (ret) |
1084 | goto unlock; | |
de151cf6 | 1085 | |
05394f39 CW |
1086 | if (i915_gem_object_is_inactive(obj)) |
1087 | list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list); | |
7d1c4804 | 1088 | |
6299f992 CW |
1089 | obj->fault_mappable = true; |
1090 | ||
05394f39 | 1091 | pfn = ((dev->agp->base + obj->gtt_offset) >> PAGE_SHIFT) + |
de151cf6 JB |
1092 | page_offset; |
1093 | ||
1094 | /* Finally, remap it using the new GTT offset */ | |
1095 | ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn); | |
c715089f | 1096 | unlock: |
de151cf6 | 1097 | mutex_unlock(&dev->struct_mutex); |
d9bc7e9f | 1098 | out: |
de151cf6 | 1099 | switch (ret) { |
d9bc7e9f | 1100 | case -EIO: |
045e769a | 1101 | case -EAGAIN: |
d9bc7e9f CW |
1102 | /* Give the error handler a chance to run and move the |
1103 | * objects off the GPU active list. Next time we service the | |
1104 | * fault, we should be able to transition the page into the | |
1105 | * GTT without touching the GPU (and so avoid further | |
1106 | * EIO/EGAIN). If the GPU is wedged, then there is no issue | |
1107 | * with coherency, just lost writes. | |
1108 | */ | |
045e769a | 1109 | set_need_resched(); |
c715089f CW |
1110 | case 0: |
1111 | case -ERESTARTSYS: | |
bed636ab | 1112 | case -EINTR: |
c715089f | 1113 | return VM_FAULT_NOPAGE; |
de151cf6 | 1114 | case -ENOMEM: |
de151cf6 | 1115 | return VM_FAULT_OOM; |
de151cf6 | 1116 | default: |
c715089f | 1117 | return VM_FAULT_SIGBUS; |
de151cf6 JB |
1118 | } |
1119 | } | |
1120 | ||
901782b2 CW |
1121 | /** |
1122 | * i915_gem_release_mmap - remove physical page mappings | |
1123 | * @obj: obj in question | |
1124 | * | |
af901ca1 | 1125 | * Preserve the reservation of the mmapping with the DRM core code, but |
901782b2 CW |
1126 | * relinquish ownership of the pages back to the system. |
1127 | * | |
1128 | * It is vital that we remove the page mapping if we have mapped a tiled | |
1129 | * object through the GTT and then lose the fence register due to | |
1130 | * resource pressure. Similarly if the object has been moved out of the | |
1131 | * aperture, than pages mapped into userspace must be revoked. Removing the | |
1132 | * mapping will then trigger a page fault on the next user access, allowing | |
1133 | * fixup by i915_gem_fault(). | |
1134 | */ | |
d05ca301 | 1135 | void |
05394f39 | 1136 | i915_gem_release_mmap(struct drm_i915_gem_object *obj) |
901782b2 | 1137 | { |
6299f992 CW |
1138 | if (!obj->fault_mappable) |
1139 | return; | |
901782b2 | 1140 | |
f6e47884 CW |
1141 | if (obj->base.dev->dev_mapping) |
1142 | unmap_mapping_range(obj->base.dev->dev_mapping, | |
1143 | (loff_t)obj->base.map_list.hash.key<<PAGE_SHIFT, | |
1144 | obj->base.size, 1); | |
fb7d516a | 1145 | |
6299f992 | 1146 | obj->fault_mappable = false; |
901782b2 CW |
1147 | } |
1148 | ||
92b88aeb | 1149 | static uint32_t |
e28f8711 | 1150 | i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode) |
92b88aeb | 1151 | { |
e28f8711 | 1152 | uint32_t gtt_size; |
92b88aeb CW |
1153 | |
1154 | if (INTEL_INFO(dev)->gen >= 4 || | |
e28f8711 CW |
1155 | tiling_mode == I915_TILING_NONE) |
1156 | return size; | |
92b88aeb CW |
1157 | |
1158 | /* Previous chips need a power-of-two fence region when tiling */ | |
1159 | if (INTEL_INFO(dev)->gen == 3) | |
e28f8711 | 1160 | gtt_size = 1024*1024; |
92b88aeb | 1161 | else |
e28f8711 | 1162 | gtt_size = 512*1024; |
92b88aeb | 1163 | |
e28f8711 CW |
1164 | while (gtt_size < size) |
1165 | gtt_size <<= 1; | |
92b88aeb | 1166 | |
e28f8711 | 1167 | return gtt_size; |
92b88aeb CW |
1168 | } |
1169 | ||
de151cf6 JB |
1170 | /** |
1171 | * i915_gem_get_gtt_alignment - return required GTT alignment for an object | |
1172 | * @obj: object to check | |
1173 | * | |
1174 | * Return the required GTT alignment for an object, taking into account | |
5e783301 | 1175 | * potential fence register mapping. |
de151cf6 JB |
1176 | */ |
1177 | static uint32_t | |
e28f8711 CW |
1178 | i915_gem_get_gtt_alignment(struct drm_device *dev, |
1179 | uint32_t size, | |
1180 | int tiling_mode) | |
de151cf6 | 1181 | { |
de151cf6 JB |
1182 | /* |
1183 | * Minimum alignment is 4k (GTT page size), but might be greater | |
1184 | * if a fence register is needed for the object. | |
1185 | */ | |
a00b10c3 | 1186 | if (INTEL_INFO(dev)->gen >= 4 || |
e28f8711 | 1187 | tiling_mode == I915_TILING_NONE) |
de151cf6 JB |
1188 | return 4096; |
1189 | ||
a00b10c3 CW |
1190 | /* |
1191 | * Previous chips need to be aligned to the size of the smallest | |
1192 | * fence register that can contain the object. | |
1193 | */ | |
e28f8711 | 1194 | return i915_gem_get_gtt_size(dev, size, tiling_mode); |
a00b10c3 CW |
1195 | } |
1196 | ||
5e783301 DV |
1197 | /** |
1198 | * i915_gem_get_unfenced_gtt_alignment - return required GTT alignment for an | |
1199 | * unfenced object | |
e28f8711 CW |
1200 | * @dev: the device |
1201 | * @size: size of the object | |
1202 | * @tiling_mode: tiling mode of the object | |
5e783301 DV |
1203 | * |
1204 | * Return the required GTT alignment for an object, only taking into account | |
1205 | * unfenced tiled surface requirements. | |
1206 | */ | |
467cffba | 1207 | uint32_t |
e28f8711 CW |
1208 | i915_gem_get_unfenced_gtt_alignment(struct drm_device *dev, |
1209 | uint32_t size, | |
1210 | int tiling_mode) | |
5e783301 | 1211 | { |
5e783301 DV |
1212 | /* |
1213 | * Minimum alignment is 4k (GTT page size) for sane hw. | |
1214 | */ | |
1215 | if (INTEL_INFO(dev)->gen >= 4 || IS_G33(dev) || | |
e28f8711 | 1216 | tiling_mode == I915_TILING_NONE) |
5e783301 DV |
1217 | return 4096; |
1218 | ||
e28f8711 CW |
1219 | /* Previous hardware however needs to be aligned to a power-of-two |
1220 | * tile height. The simplest method for determining this is to reuse | |
1221 | * the power-of-tile object size. | |
5e783301 | 1222 | */ |
e28f8711 | 1223 | return i915_gem_get_gtt_size(dev, size, tiling_mode); |
5e783301 DV |
1224 | } |
1225 | ||
de151cf6 | 1226 | int |
ff72145b DA |
1227 | i915_gem_mmap_gtt(struct drm_file *file, |
1228 | struct drm_device *dev, | |
1229 | uint32_t handle, | |
1230 | uint64_t *offset) | |
de151cf6 | 1231 | { |
da761a6e | 1232 | struct drm_i915_private *dev_priv = dev->dev_private; |
05394f39 | 1233 | struct drm_i915_gem_object *obj; |
de151cf6 JB |
1234 | int ret; |
1235 | ||
1236 | if (!(dev->driver->driver_features & DRIVER_GEM)) | |
1237 | return -ENODEV; | |
1238 | ||
76c1dec1 | 1239 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 1240 | if (ret) |
76c1dec1 | 1241 | return ret; |
de151cf6 | 1242 | |
ff72145b | 1243 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle)); |
c8725226 | 1244 | if (&obj->base == NULL) { |
1d7cfea1 CW |
1245 | ret = -ENOENT; |
1246 | goto unlock; | |
1247 | } | |
de151cf6 | 1248 | |
05394f39 | 1249 | if (obj->base.size > dev_priv->mm.gtt_mappable_end) { |
da761a6e | 1250 | ret = -E2BIG; |
ff56b0bc | 1251 | goto out; |
da761a6e CW |
1252 | } |
1253 | ||
05394f39 | 1254 | if (obj->madv != I915_MADV_WILLNEED) { |
ab18282d | 1255 | DRM_ERROR("Attempting to mmap a purgeable buffer\n"); |
1d7cfea1 CW |
1256 | ret = -EINVAL; |
1257 | goto out; | |
ab18282d CW |
1258 | } |
1259 | ||
05394f39 | 1260 | if (!obj->base.map_list.map) { |
b464e9a2 | 1261 | ret = drm_gem_create_mmap_offset(&obj->base); |
1d7cfea1 CW |
1262 | if (ret) |
1263 | goto out; | |
de151cf6 JB |
1264 | } |
1265 | ||
ff72145b | 1266 | *offset = (u64)obj->base.map_list.hash.key << PAGE_SHIFT; |
de151cf6 | 1267 | |
1d7cfea1 | 1268 | out: |
05394f39 | 1269 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 1270 | unlock: |
de151cf6 | 1271 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 1272 | return ret; |
de151cf6 JB |
1273 | } |
1274 | ||
ff72145b DA |
1275 | /** |
1276 | * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing | |
1277 | * @dev: DRM device | |
1278 | * @data: GTT mapping ioctl data | |
1279 | * @file: GEM object info | |
1280 | * | |
1281 | * Simply returns the fake offset to userspace so it can mmap it. | |
1282 | * The mmap call will end up in drm_gem_mmap(), which will set things | |
1283 | * up so we can get faults in the handler above. | |
1284 | * | |
1285 | * The fault handler will take care of binding the object into the GTT | |
1286 | * (since it may have been evicted to make room for something), allocating | |
1287 | * a fence register, and mapping the appropriate aperture address into | |
1288 | * userspace. | |
1289 | */ | |
1290 | int | |
1291 | i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data, | |
1292 | struct drm_file *file) | |
1293 | { | |
1294 | struct drm_i915_gem_mmap_gtt *args = data; | |
1295 | ||
1296 | if (!(dev->driver->driver_features & DRIVER_GEM)) | |
1297 | return -ENODEV; | |
1298 | ||
1299 | return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset); | |
1300 | } | |
1301 | ||
1302 | ||
e5281ccd | 1303 | static int |
05394f39 | 1304 | i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj, |
e5281ccd CW |
1305 | gfp_t gfpmask) |
1306 | { | |
e5281ccd CW |
1307 | int page_count, i; |
1308 | struct address_space *mapping; | |
1309 | struct inode *inode; | |
1310 | struct page *page; | |
1311 | ||
1312 | /* Get the list of pages out of our struct file. They'll be pinned | |
1313 | * at this point until we release them. | |
1314 | */ | |
05394f39 CW |
1315 | page_count = obj->base.size / PAGE_SIZE; |
1316 | BUG_ON(obj->pages != NULL); | |
1317 | obj->pages = drm_malloc_ab(page_count, sizeof(struct page *)); | |
1318 | if (obj->pages == NULL) | |
e5281ccd CW |
1319 | return -ENOMEM; |
1320 | ||
05394f39 | 1321 | inode = obj->base.filp->f_path.dentry->d_inode; |
e5281ccd | 1322 | mapping = inode->i_mapping; |
5949eac4 HD |
1323 | gfpmask |= mapping_gfp_mask(mapping); |
1324 | ||
e5281ccd | 1325 | for (i = 0; i < page_count; i++) { |
5949eac4 | 1326 | page = shmem_read_mapping_page_gfp(mapping, i, gfpmask); |
e5281ccd CW |
1327 | if (IS_ERR(page)) |
1328 | goto err_pages; | |
1329 | ||
05394f39 | 1330 | obj->pages[i] = page; |
e5281ccd CW |
1331 | } |
1332 | ||
6dacfd2f | 1333 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
e5281ccd CW |
1334 | i915_gem_object_do_bit_17_swizzle(obj); |
1335 | ||
1336 | return 0; | |
1337 | ||
1338 | err_pages: | |
1339 | while (i--) | |
05394f39 | 1340 | page_cache_release(obj->pages[i]); |
e5281ccd | 1341 | |
05394f39 CW |
1342 | drm_free_large(obj->pages); |
1343 | obj->pages = NULL; | |
e5281ccd CW |
1344 | return PTR_ERR(page); |
1345 | } | |
1346 | ||
5cdf5881 | 1347 | static void |
05394f39 | 1348 | i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj) |
673a394b | 1349 | { |
05394f39 | 1350 | int page_count = obj->base.size / PAGE_SIZE; |
673a394b EA |
1351 | int i; |
1352 | ||
05394f39 | 1353 | BUG_ON(obj->madv == __I915_MADV_PURGED); |
673a394b | 1354 | |
6dacfd2f | 1355 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
280b713b EA |
1356 | i915_gem_object_save_bit_17_swizzle(obj); |
1357 | ||
05394f39 CW |
1358 | if (obj->madv == I915_MADV_DONTNEED) |
1359 | obj->dirty = 0; | |
3ef94daa CW |
1360 | |
1361 | for (i = 0; i < page_count; i++) { | |
05394f39 CW |
1362 | if (obj->dirty) |
1363 | set_page_dirty(obj->pages[i]); | |
3ef94daa | 1364 | |
05394f39 CW |
1365 | if (obj->madv == I915_MADV_WILLNEED) |
1366 | mark_page_accessed(obj->pages[i]); | |
3ef94daa | 1367 | |
05394f39 | 1368 | page_cache_release(obj->pages[i]); |
3ef94daa | 1369 | } |
05394f39 | 1370 | obj->dirty = 0; |
673a394b | 1371 | |
05394f39 CW |
1372 | drm_free_large(obj->pages); |
1373 | obj->pages = NULL; | |
673a394b EA |
1374 | } |
1375 | ||
54cf91dc | 1376 | void |
05394f39 | 1377 | i915_gem_object_move_to_active(struct drm_i915_gem_object *obj, |
1ec14ad3 CW |
1378 | struct intel_ring_buffer *ring, |
1379 | u32 seqno) | |
673a394b | 1380 | { |
05394f39 | 1381 | struct drm_device *dev = obj->base.dev; |
69dc4987 | 1382 | struct drm_i915_private *dev_priv = dev->dev_private; |
617dbe27 | 1383 | |
852835f3 | 1384 | BUG_ON(ring == NULL); |
05394f39 | 1385 | obj->ring = ring; |
673a394b EA |
1386 | |
1387 | /* Add a reference if we're newly entering the active list. */ | |
05394f39 CW |
1388 | if (!obj->active) { |
1389 | drm_gem_object_reference(&obj->base); | |
1390 | obj->active = 1; | |
673a394b | 1391 | } |
e35a41de | 1392 | |
673a394b | 1393 | /* Move from whatever list we were on to the tail of execution. */ |
05394f39 CW |
1394 | list_move_tail(&obj->mm_list, &dev_priv->mm.active_list); |
1395 | list_move_tail(&obj->ring_list, &ring->active_list); | |
caea7476 | 1396 | |
05394f39 | 1397 | obj->last_rendering_seqno = seqno; |
caea7476 | 1398 | |
7dd49065 | 1399 | if (obj->fenced_gpu_access) { |
caea7476 CW |
1400 | obj->last_fenced_seqno = seqno; |
1401 | obj->last_fenced_ring = ring; | |
1402 | ||
7dd49065 CW |
1403 | /* Bump MRU to take account of the delayed flush */ |
1404 | if (obj->fence_reg != I915_FENCE_REG_NONE) { | |
1405 | struct drm_i915_fence_reg *reg; | |
1406 | ||
1407 | reg = &dev_priv->fence_regs[obj->fence_reg]; | |
1408 | list_move_tail(®->lru_list, | |
1409 | &dev_priv->mm.fence_list); | |
1410 | } | |
caea7476 CW |
1411 | } |
1412 | } | |
1413 | ||
1414 | static void | |
1415 | i915_gem_object_move_off_active(struct drm_i915_gem_object *obj) | |
1416 | { | |
1417 | list_del_init(&obj->ring_list); | |
1418 | obj->last_rendering_seqno = 0; | |
15a13bbd | 1419 | obj->last_fenced_seqno = 0; |
673a394b EA |
1420 | } |
1421 | ||
ce44b0ea | 1422 | static void |
05394f39 | 1423 | i915_gem_object_move_to_flushing(struct drm_i915_gem_object *obj) |
ce44b0ea | 1424 | { |
05394f39 | 1425 | struct drm_device *dev = obj->base.dev; |
ce44b0ea | 1426 | drm_i915_private_t *dev_priv = dev->dev_private; |
ce44b0ea | 1427 | |
05394f39 CW |
1428 | BUG_ON(!obj->active); |
1429 | list_move_tail(&obj->mm_list, &dev_priv->mm.flushing_list); | |
caea7476 CW |
1430 | |
1431 | i915_gem_object_move_off_active(obj); | |
1432 | } | |
1433 | ||
1434 | static void | |
1435 | i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj) | |
1436 | { | |
1437 | struct drm_device *dev = obj->base.dev; | |
1438 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1439 | ||
1440 | if (obj->pin_count != 0) | |
1441 | list_move_tail(&obj->mm_list, &dev_priv->mm.pinned_list); | |
1442 | else | |
1443 | list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list); | |
1444 | ||
1445 | BUG_ON(!list_empty(&obj->gpu_write_list)); | |
1446 | BUG_ON(!obj->active); | |
1447 | obj->ring = NULL; | |
15a13bbd | 1448 | obj->last_fenced_ring = NULL; |
caea7476 CW |
1449 | |
1450 | i915_gem_object_move_off_active(obj); | |
1451 | obj->fenced_gpu_access = false; | |
caea7476 CW |
1452 | |
1453 | obj->active = 0; | |
87ca9c8a | 1454 | obj->pending_gpu_write = false; |
caea7476 CW |
1455 | drm_gem_object_unreference(&obj->base); |
1456 | ||
1457 | WARN_ON(i915_verify_lists(dev)); | |
ce44b0ea | 1458 | } |
673a394b | 1459 | |
963b4836 CW |
1460 | /* Immediately discard the backing storage */ |
1461 | static void | |
05394f39 | 1462 | i915_gem_object_truncate(struct drm_i915_gem_object *obj) |
963b4836 | 1463 | { |
bb6baf76 | 1464 | struct inode *inode; |
963b4836 | 1465 | |
ae9fed6b CW |
1466 | /* Our goal here is to return as much of the memory as |
1467 | * is possible back to the system as we are called from OOM. | |
1468 | * To do this we must instruct the shmfs to drop all of its | |
e2377fe0 | 1469 | * backing pages, *now*. |
ae9fed6b | 1470 | */ |
05394f39 | 1471 | inode = obj->base.filp->f_path.dentry->d_inode; |
e2377fe0 | 1472 | shmem_truncate_range(inode, 0, (loff_t)-1); |
bb6baf76 | 1473 | |
a14917ee CW |
1474 | if (obj->base.map_list.map) |
1475 | drm_gem_free_mmap_offset(&obj->base); | |
1476 | ||
05394f39 | 1477 | obj->madv = __I915_MADV_PURGED; |
963b4836 CW |
1478 | } |
1479 | ||
1480 | static inline int | |
05394f39 | 1481 | i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj) |
963b4836 | 1482 | { |
05394f39 | 1483 | return obj->madv == I915_MADV_DONTNEED; |
963b4836 CW |
1484 | } |
1485 | ||
63560396 | 1486 | static void |
db53a302 CW |
1487 | i915_gem_process_flushing_list(struct intel_ring_buffer *ring, |
1488 | uint32_t flush_domains) | |
63560396 | 1489 | { |
05394f39 | 1490 | struct drm_i915_gem_object *obj, *next; |
63560396 | 1491 | |
05394f39 | 1492 | list_for_each_entry_safe(obj, next, |
64193406 | 1493 | &ring->gpu_write_list, |
63560396 | 1494 | gpu_write_list) { |
05394f39 CW |
1495 | if (obj->base.write_domain & flush_domains) { |
1496 | uint32_t old_write_domain = obj->base.write_domain; | |
63560396 | 1497 | |
05394f39 CW |
1498 | obj->base.write_domain = 0; |
1499 | list_del_init(&obj->gpu_write_list); | |
1ec14ad3 | 1500 | i915_gem_object_move_to_active(obj, ring, |
db53a302 | 1501 | i915_gem_next_request_seqno(ring)); |
63560396 | 1502 | |
63560396 | 1503 | trace_i915_gem_object_change_domain(obj, |
05394f39 | 1504 | obj->base.read_domains, |
63560396 DV |
1505 | old_write_domain); |
1506 | } | |
1507 | } | |
1508 | } | |
8187a2b7 | 1509 | |
53d227f2 DV |
1510 | static u32 |
1511 | i915_gem_get_seqno(struct drm_device *dev) | |
1512 | { | |
1513 | drm_i915_private_t *dev_priv = dev->dev_private; | |
1514 | u32 seqno = dev_priv->next_seqno; | |
1515 | ||
1516 | /* reserve 0 for non-seqno */ | |
1517 | if (++dev_priv->next_seqno == 0) | |
1518 | dev_priv->next_seqno = 1; | |
1519 | ||
1520 | return seqno; | |
1521 | } | |
1522 | ||
1523 | u32 | |
1524 | i915_gem_next_request_seqno(struct intel_ring_buffer *ring) | |
1525 | { | |
1526 | if (ring->outstanding_lazy_request == 0) | |
1527 | ring->outstanding_lazy_request = i915_gem_get_seqno(ring->dev); | |
1528 | ||
1529 | return ring->outstanding_lazy_request; | |
1530 | } | |
1531 | ||
3cce469c | 1532 | int |
db53a302 | 1533 | i915_add_request(struct intel_ring_buffer *ring, |
f787a5f5 | 1534 | struct drm_file *file, |
db53a302 | 1535 | struct drm_i915_gem_request *request) |
673a394b | 1536 | { |
db53a302 | 1537 | drm_i915_private_t *dev_priv = ring->dev->dev_private; |
673a394b | 1538 | uint32_t seqno; |
a71d8d94 | 1539 | u32 request_ring_position; |
673a394b | 1540 | int was_empty; |
3cce469c CW |
1541 | int ret; |
1542 | ||
1543 | BUG_ON(request == NULL); | |
53d227f2 | 1544 | seqno = i915_gem_next_request_seqno(ring); |
673a394b | 1545 | |
a71d8d94 CW |
1546 | /* Record the position of the start of the request so that |
1547 | * should we detect the updated seqno part-way through the | |
1548 | * GPU processing the request, we never over-estimate the | |
1549 | * position of the head. | |
1550 | */ | |
1551 | request_ring_position = intel_ring_get_tail(ring); | |
1552 | ||
3cce469c CW |
1553 | ret = ring->add_request(ring, &seqno); |
1554 | if (ret) | |
1555 | return ret; | |
673a394b | 1556 | |
db53a302 | 1557 | trace_i915_gem_request_add(ring, seqno); |
673a394b EA |
1558 | |
1559 | request->seqno = seqno; | |
852835f3 | 1560 | request->ring = ring; |
a71d8d94 | 1561 | request->tail = request_ring_position; |
673a394b | 1562 | request->emitted_jiffies = jiffies; |
852835f3 ZN |
1563 | was_empty = list_empty(&ring->request_list); |
1564 | list_add_tail(&request->list, &ring->request_list); | |
1565 | ||
db53a302 CW |
1566 | if (file) { |
1567 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
1568 | ||
1c25595f | 1569 | spin_lock(&file_priv->mm.lock); |
f787a5f5 | 1570 | request->file_priv = file_priv; |
b962442e | 1571 | list_add_tail(&request->client_list, |
f787a5f5 | 1572 | &file_priv->mm.request_list); |
1c25595f | 1573 | spin_unlock(&file_priv->mm.lock); |
b962442e | 1574 | } |
673a394b | 1575 | |
5391d0cf | 1576 | ring->outstanding_lazy_request = 0; |
db53a302 | 1577 | |
f65d9421 | 1578 | if (!dev_priv->mm.suspended) { |
3e0dc6b0 BW |
1579 | if (i915_enable_hangcheck) { |
1580 | mod_timer(&dev_priv->hangcheck_timer, | |
1581 | jiffies + | |
1582 | msecs_to_jiffies(DRM_I915_HANGCHECK_PERIOD)); | |
1583 | } | |
f65d9421 | 1584 | if (was_empty) |
b3b079db CW |
1585 | queue_delayed_work(dev_priv->wq, |
1586 | &dev_priv->mm.retire_work, HZ); | |
f65d9421 | 1587 | } |
3cce469c | 1588 | return 0; |
673a394b EA |
1589 | } |
1590 | ||
f787a5f5 CW |
1591 | static inline void |
1592 | i915_gem_request_remove_from_client(struct drm_i915_gem_request *request) | |
673a394b | 1593 | { |
1c25595f | 1594 | struct drm_i915_file_private *file_priv = request->file_priv; |
673a394b | 1595 | |
1c25595f CW |
1596 | if (!file_priv) |
1597 | return; | |
1c5d22f7 | 1598 | |
1c25595f | 1599 | spin_lock(&file_priv->mm.lock); |
09bfa517 HRK |
1600 | if (request->file_priv) { |
1601 | list_del(&request->client_list); | |
1602 | request->file_priv = NULL; | |
1603 | } | |
1c25595f | 1604 | spin_unlock(&file_priv->mm.lock); |
673a394b | 1605 | } |
673a394b | 1606 | |
dfaae392 CW |
1607 | static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv, |
1608 | struct intel_ring_buffer *ring) | |
9375e446 | 1609 | { |
dfaae392 CW |
1610 | while (!list_empty(&ring->request_list)) { |
1611 | struct drm_i915_gem_request *request; | |
673a394b | 1612 | |
dfaae392 CW |
1613 | request = list_first_entry(&ring->request_list, |
1614 | struct drm_i915_gem_request, | |
1615 | list); | |
de151cf6 | 1616 | |
dfaae392 | 1617 | list_del(&request->list); |
f787a5f5 | 1618 | i915_gem_request_remove_from_client(request); |
dfaae392 CW |
1619 | kfree(request); |
1620 | } | |
673a394b | 1621 | |
dfaae392 | 1622 | while (!list_empty(&ring->active_list)) { |
05394f39 | 1623 | struct drm_i915_gem_object *obj; |
9375e446 | 1624 | |
05394f39 CW |
1625 | obj = list_first_entry(&ring->active_list, |
1626 | struct drm_i915_gem_object, | |
1627 | ring_list); | |
9375e446 | 1628 | |
05394f39 CW |
1629 | obj->base.write_domain = 0; |
1630 | list_del_init(&obj->gpu_write_list); | |
1631 | i915_gem_object_move_to_inactive(obj); | |
673a394b EA |
1632 | } |
1633 | } | |
1634 | ||
312817a3 CW |
1635 | static void i915_gem_reset_fences(struct drm_device *dev) |
1636 | { | |
1637 | struct drm_i915_private *dev_priv = dev->dev_private; | |
1638 | int i; | |
1639 | ||
4b9de737 | 1640 | for (i = 0; i < dev_priv->num_fence_regs; i++) { |
312817a3 | 1641 | struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i]; |
7d2cb39c CW |
1642 | struct drm_i915_gem_object *obj = reg->obj; |
1643 | ||
1644 | if (!obj) | |
1645 | continue; | |
1646 | ||
1647 | if (obj->tiling_mode) | |
1648 | i915_gem_release_mmap(obj); | |
1649 | ||
d9e86c0e CW |
1650 | reg->obj->fence_reg = I915_FENCE_REG_NONE; |
1651 | reg->obj->fenced_gpu_access = false; | |
1652 | reg->obj->last_fenced_seqno = 0; | |
1653 | reg->obj->last_fenced_ring = NULL; | |
1654 | i915_gem_clear_fence_reg(dev, reg); | |
312817a3 CW |
1655 | } |
1656 | } | |
1657 | ||
069efc1d | 1658 | void i915_gem_reset(struct drm_device *dev) |
673a394b | 1659 | { |
77f01230 | 1660 | struct drm_i915_private *dev_priv = dev->dev_private; |
05394f39 | 1661 | struct drm_i915_gem_object *obj; |
1ec14ad3 | 1662 | int i; |
673a394b | 1663 | |
1ec14ad3 CW |
1664 | for (i = 0; i < I915_NUM_RINGS; i++) |
1665 | i915_gem_reset_ring_lists(dev_priv, &dev_priv->ring[i]); | |
dfaae392 CW |
1666 | |
1667 | /* Remove anything from the flushing lists. The GPU cache is likely | |
1668 | * to be lost on reset along with the data, so simply move the | |
1669 | * lost bo to the inactive list. | |
1670 | */ | |
1671 | while (!list_empty(&dev_priv->mm.flushing_list)) { | |
0206e353 | 1672 | obj = list_first_entry(&dev_priv->mm.flushing_list, |
05394f39 CW |
1673 | struct drm_i915_gem_object, |
1674 | mm_list); | |
dfaae392 | 1675 | |
05394f39 CW |
1676 | obj->base.write_domain = 0; |
1677 | list_del_init(&obj->gpu_write_list); | |
1678 | i915_gem_object_move_to_inactive(obj); | |
dfaae392 CW |
1679 | } |
1680 | ||
1681 | /* Move everything out of the GPU domains to ensure we do any | |
1682 | * necessary invalidation upon reuse. | |
1683 | */ | |
05394f39 | 1684 | list_for_each_entry(obj, |
77f01230 | 1685 | &dev_priv->mm.inactive_list, |
69dc4987 | 1686 | mm_list) |
77f01230 | 1687 | { |
05394f39 | 1688 | obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS; |
77f01230 | 1689 | } |
069efc1d CW |
1690 | |
1691 | /* The fence registers are invalidated so clear them out */ | |
312817a3 | 1692 | i915_gem_reset_fences(dev); |
673a394b EA |
1693 | } |
1694 | ||
1695 | /** | |
1696 | * This function clears the request list as sequence numbers are passed. | |
1697 | */ | |
a71d8d94 | 1698 | void |
db53a302 | 1699 | i915_gem_retire_requests_ring(struct intel_ring_buffer *ring) |
673a394b | 1700 | { |
673a394b | 1701 | uint32_t seqno; |
1ec14ad3 | 1702 | int i; |
673a394b | 1703 | |
db53a302 | 1704 | if (list_empty(&ring->request_list)) |
6c0594a3 KW |
1705 | return; |
1706 | ||
db53a302 | 1707 | WARN_ON(i915_verify_lists(ring->dev)); |
673a394b | 1708 | |
78501eac | 1709 | seqno = ring->get_seqno(ring); |
1ec14ad3 | 1710 | |
076e2c0e | 1711 | for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++) |
1ec14ad3 CW |
1712 | if (seqno >= ring->sync_seqno[i]) |
1713 | ring->sync_seqno[i] = 0; | |
1714 | ||
852835f3 | 1715 | while (!list_empty(&ring->request_list)) { |
673a394b | 1716 | struct drm_i915_gem_request *request; |
673a394b | 1717 | |
852835f3 | 1718 | request = list_first_entry(&ring->request_list, |
673a394b EA |
1719 | struct drm_i915_gem_request, |
1720 | list); | |
673a394b | 1721 | |
dfaae392 | 1722 | if (!i915_seqno_passed(seqno, request->seqno)) |
b84d5f0c CW |
1723 | break; |
1724 | ||
db53a302 | 1725 | trace_i915_gem_request_retire(ring, request->seqno); |
a71d8d94 CW |
1726 | /* We know the GPU must have read the request to have |
1727 | * sent us the seqno + interrupt, so use the position | |
1728 | * of tail of the request to update the last known position | |
1729 | * of the GPU head. | |
1730 | */ | |
1731 | ring->last_retired_head = request->tail; | |
b84d5f0c CW |
1732 | |
1733 | list_del(&request->list); | |
f787a5f5 | 1734 | i915_gem_request_remove_from_client(request); |
b84d5f0c CW |
1735 | kfree(request); |
1736 | } | |
673a394b | 1737 | |
b84d5f0c CW |
1738 | /* Move any buffers on the active list that are no longer referenced |
1739 | * by the ringbuffer to the flushing/inactive lists as appropriate. | |
1740 | */ | |
1741 | while (!list_empty(&ring->active_list)) { | |
05394f39 | 1742 | struct drm_i915_gem_object *obj; |
b84d5f0c | 1743 | |
0206e353 | 1744 | obj = list_first_entry(&ring->active_list, |
05394f39 CW |
1745 | struct drm_i915_gem_object, |
1746 | ring_list); | |
673a394b | 1747 | |
05394f39 | 1748 | if (!i915_seqno_passed(seqno, obj->last_rendering_seqno)) |
673a394b | 1749 | break; |
b84d5f0c | 1750 | |
05394f39 | 1751 | if (obj->base.write_domain != 0) |
b84d5f0c CW |
1752 | i915_gem_object_move_to_flushing(obj); |
1753 | else | |
1754 | i915_gem_object_move_to_inactive(obj); | |
673a394b | 1755 | } |
9d34e5db | 1756 | |
db53a302 CW |
1757 | if (unlikely(ring->trace_irq_seqno && |
1758 | i915_seqno_passed(seqno, ring->trace_irq_seqno))) { | |
1ec14ad3 | 1759 | ring->irq_put(ring); |
db53a302 | 1760 | ring->trace_irq_seqno = 0; |
9d34e5db | 1761 | } |
23bc5982 | 1762 | |
db53a302 | 1763 | WARN_ON(i915_verify_lists(ring->dev)); |
673a394b EA |
1764 | } |
1765 | ||
b09a1fec CW |
1766 | void |
1767 | i915_gem_retire_requests(struct drm_device *dev) | |
1768 | { | |
1769 | drm_i915_private_t *dev_priv = dev->dev_private; | |
1ec14ad3 | 1770 | int i; |
b09a1fec | 1771 | |
be72615b | 1772 | if (!list_empty(&dev_priv->mm.deferred_free_list)) { |
05394f39 | 1773 | struct drm_i915_gem_object *obj, *next; |
be72615b CW |
1774 | |
1775 | /* We must be careful that during unbind() we do not | |
1776 | * accidentally infinitely recurse into retire requests. | |
1777 | * Currently: | |
1778 | * retire -> free -> unbind -> wait -> retire_ring | |
1779 | */ | |
05394f39 | 1780 | list_for_each_entry_safe(obj, next, |
be72615b | 1781 | &dev_priv->mm.deferred_free_list, |
69dc4987 | 1782 | mm_list) |
05394f39 | 1783 | i915_gem_free_object_tail(obj); |
be72615b CW |
1784 | } |
1785 | ||
1ec14ad3 | 1786 | for (i = 0; i < I915_NUM_RINGS; i++) |
db53a302 | 1787 | i915_gem_retire_requests_ring(&dev_priv->ring[i]); |
b09a1fec CW |
1788 | } |
1789 | ||
75ef9da2 | 1790 | static void |
673a394b EA |
1791 | i915_gem_retire_work_handler(struct work_struct *work) |
1792 | { | |
1793 | drm_i915_private_t *dev_priv; | |
1794 | struct drm_device *dev; | |
0a58705b CW |
1795 | bool idle; |
1796 | int i; | |
673a394b EA |
1797 | |
1798 | dev_priv = container_of(work, drm_i915_private_t, | |
1799 | mm.retire_work.work); | |
1800 | dev = dev_priv->dev; | |
1801 | ||
891b48cf CW |
1802 | /* Come back later if the device is busy... */ |
1803 | if (!mutex_trylock(&dev->struct_mutex)) { | |
1804 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ); | |
1805 | return; | |
1806 | } | |
1807 | ||
b09a1fec | 1808 | i915_gem_retire_requests(dev); |
d1b851fc | 1809 | |
0a58705b CW |
1810 | /* Send a periodic flush down the ring so we don't hold onto GEM |
1811 | * objects indefinitely. | |
1812 | */ | |
1813 | idle = true; | |
1814 | for (i = 0; i < I915_NUM_RINGS; i++) { | |
1815 | struct intel_ring_buffer *ring = &dev_priv->ring[i]; | |
1816 | ||
1817 | if (!list_empty(&ring->gpu_write_list)) { | |
1818 | struct drm_i915_gem_request *request; | |
1819 | int ret; | |
1820 | ||
db53a302 CW |
1821 | ret = i915_gem_flush_ring(ring, |
1822 | 0, I915_GEM_GPU_DOMAINS); | |
0a58705b CW |
1823 | request = kzalloc(sizeof(*request), GFP_KERNEL); |
1824 | if (ret || request == NULL || | |
db53a302 | 1825 | i915_add_request(ring, NULL, request)) |
0a58705b CW |
1826 | kfree(request); |
1827 | } | |
1828 | ||
1829 | idle &= list_empty(&ring->request_list); | |
1830 | } | |
1831 | ||
1832 | if (!dev_priv->mm.suspended && !idle) | |
9c9fe1f8 | 1833 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ); |
0a58705b | 1834 | |
673a394b EA |
1835 | mutex_unlock(&dev->struct_mutex); |
1836 | } | |
1837 | ||
db53a302 CW |
1838 | /** |
1839 | * Waits for a sequence number to be signaled, and cleans up the | |
1840 | * request and object lists appropriately for that event. | |
1841 | */ | |
5a5a0c64 | 1842 | int |
db53a302 | 1843 | i915_wait_request(struct intel_ring_buffer *ring, |
b93f9cf1 BW |
1844 | uint32_t seqno, |
1845 | bool do_retire) | |
673a394b | 1846 | { |
db53a302 | 1847 | drm_i915_private_t *dev_priv = ring->dev->dev_private; |
802c7eb6 | 1848 | u32 ier; |
673a394b EA |
1849 | int ret = 0; |
1850 | ||
1851 | BUG_ON(seqno == 0); | |
1852 | ||
d9bc7e9f CW |
1853 | if (atomic_read(&dev_priv->mm.wedged)) { |
1854 | struct completion *x = &dev_priv->error_completion; | |
1855 | bool recovery_complete; | |
1856 | unsigned long flags; | |
1857 | ||
1858 | /* Give the error handler a chance to run. */ | |
1859 | spin_lock_irqsave(&x->wait.lock, flags); | |
1860 | recovery_complete = x->done > 0; | |
1861 | spin_unlock_irqrestore(&x->wait.lock, flags); | |
1862 | ||
1863 | return recovery_complete ? -EIO : -EAGAIN; | |
1864 | } | |
30dbf0c0 | 1865 | |
5d97eb69 | 1866 | if (seqno == ring->outstanding_lazy_request) { |
3cce469c CW |
1867 | struct drm_i915_gem_request *request; |
1868 | ||
1869 | request = kzalloc(sizeof(*request), GFP_KERNEL); | |
1870 | if (request == NULL) | |
e35a41de | 1871 | return -ENOMEM; |
3cce469c | 1872 | |
db53a302 | 1873 | ret = i915_add_request(ring, NULL, request); |
3cce469c CW |
1874 | if (ret) { |
1875 | kfree(request); | |
1876 | return ret; | |
1877 | } | |
1878 | ||
1879 | seqno = request->seqno; | |
e35a41de | 1880 | } |
ffed1d09 | 1881 | |
78501eac | 1882 | if (!i915_seqno_passed(ring->get_seqno(ring), seqno)) { |
db53a302 | 1883 | if (HAS_PCH_SPLIT(ring->dev)) |
036a4a7d | 1884 | ier = I915_READ(DEIER) | I915_READ(GTIER); |
23e3f9b3 JB |
1885 | else if (IS_VALLEYVIEW(ring->dev)) |
1886 | ier = I915_READ(GTIER) | I915_READ(VLV_IER); | |
036a4a7d ZW |
1887 | else |
1888 | ier = I915_READ(IER); | |
802c7eb6 JB |
1889 | if (!ier) { |
1890 | DRM_ERROR("something (likely vbetool) disabled " | |
1891 | "interrupts, re-enabling\n"); | |
f01c22fd CW |
1892 | ring->dev->driver->irq_preinstall(ring->dev); |
1893 | ring->dev->driver->irq_postinstall(ring->dev); | |
802c7eb6 JB |
1894 | } |
1895 | ||
db53a302 | 1896 | trace_i915_gem_request_wait_begin(ring, seqno); |
1c5d22f7 | 1897 | |
b2223497 | 1898 | ring->waiting_seqno = seqno; |
b13c2b96 | 1899 | if (ring->irq_get(ring)) { |
ce453d81 | 1900 | if (dev_priv->mm.interruptible) |
b13c2b96 CW |
1901 | ret = wait_event_interruptible(ring->irq_queue, |
1902 | i915_seqno_passed(ring->get_seqno(ring), seqno) | |
1903 | || atomic_read(&dev_priv->mm.wedged)); | |
1904 | else | |
1905 | wait_event(ring->irq_queue, | |
1906 | i915_seqno_passed(ring->get_seqno(ring), seqno) | |
1907 | || atomic_read(&dev_priv->mm.wedged)); | |
1908 | ||
1909 | ring->irq_put(ring); | |
e959b5db EA |
1910 | } else if (wait_for_atomic(i915_seqno_passed(ring->get_seqno(ring), |
1911 | seqno) || | |
1912 | atomic_read(&dev_priv->mm.wedged), 3000)) | |
b5ba177d | 1913 | ret = -EBUSY; |
b2223497 | 1914 | ring->waiting_seqno = 0; |
1c5d22f7 | 1915 | |
db53a302 | 1916 | trace_i915_gem_request_wait_end(ring, seqno); |
673a394b | 1917 | } |
ba1234d1 | 1918 | if (atomic_read(&dev_priv->mm.wedged)) |
30dbf0c0 | 1919 | ret = -EAGAIN; |
673a394b | 1920 | |
673a394b EA |
1921 | /* Directly dispatch request retiring. While we have the work queue |
1922 | * to handle this, the waiter on a request often wants an associated | |
1923 | * buffer to have made it to the inactive list, and we would need | |
1924 | * a separate wait queue to handle that. | |
1925 | */ | |
b93f9cf1 | 1926 | if (ret == 0 && do_retire) |
db53a302 | 1927 | i915_gem_retire_requests_ring(ring); |
673a394b EA |
1928 | |
1929 | return ret; | |
1930 | } | |
1931 | ||
673a394b EA |
1932 | /** |
1933 | * Ensures that all rendering to the object has completed and the object is | |
1934 | * safe to unbind from the GTT or access from the CPU. | |
1935 | */ | |
54cf91dc | 1936 | int |
ce453d81 | 1937 | i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj) |
673a394b | 1938 | { |
673a394b EA |
1939 | int ret; |
1940 | ||
e47c68e9 EA |
1941 | /* This function only exists to support waiting for existing rendering, |
1942 | * not for emitting required flushes. | |
673a394b | 1943 | */ |
05394f39 | 1944 | BUG_ON((obj->base.write_domain & I915_GEM_GPU_DOMAINS) != 0); |
673a394b EA |
1945 | |
1946 | /* If there is rendering queued on the buffer being evicted, wait for | |
1947 | * it. | |
1948 | */ | |
05394f39 | 1949 | if (obj->active) { |
b93f9cf1 BW |
1950 | ret = i915_wait_request(obj->ring, obj->last_rendering_seqno, |
1951 | true); | |
2cf34d7b | 1952 | if (ret) |
673a394b EA |
1953 | return ret; |
1954 | } | |
1955 | ||
1956 | return 0; | |
1957 | } | |
1958 | ||
5816d648 BW |
1959 | /** |
1960 | * i915_gem_object_sync - sync an object to a ring. | |
1961 | * | |
1962 | * @obj: object which may be in use on another ring. | |
1963 | * @to: ring we wish to use the object on. May be NULL. | |
1964 | * | |
1965 | * This code is meant to abstract object synchronization with the GPU. | |
1966 | * Calling with NULL implies synchronizing the object with the CPU | |
1967 | * rather than a particular GPU ring. | |
1968 | * | |
1969 | * Returns 0 if successful, else propagates up the lower layer error. | |
1970 | */ | |
2911a35b BW |
1971 | int |
1972 | i915_gem_object_sync(struct drm_i915_gem_object *obj, | |
1973 | struct intel_ring_buffer *to) | |
1974 | { | |
1975 | struct intel_ring_buffer *from = obj->ring; | |
1976 | u32 seqno; | |
1977 | int ret, idx; | |
1978 | ||
1979 | if (from == NULL || to == from) | |
1980 | return 0; | |
1981 | ||
5816d648 | 1982 | if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev)) |
2911a35b BW |
1983 | return i915_gem_object_wait_rendering(obj); |
1984 | ||
1985 | idx = intel_ring_sync_index(from, to); | |
1986 | ||
1987 | seqno = obj->last_rendering_seqno; | |
1988 | if (seqno <= from->sync_seqno[idx]) | |
1989 | return 0; | |
1990 | ||
1991 | if (seqno == from->outstanding_lazy_request) { | |
1992 | struct drm_i915_gem_request *request; | |
1993 | ||
1994 | request = kzalloc(sizeof(*request), GFP_KERNEL); | |
1995 | if (request == NULL) | |
1996 | return -ENOMEM; | |
1997 | ||
1998 | ret = i915_add_request(from, NULL, request); | |
1999 | if (ret) { | |
2000 | kfree(request); | |
2001 | return ret; | |
2002 | } | |
2003 | ||
2004 | seqno = request->seqno; | |
2005 | } | |
2006 | ||
2911a35b | 2007 | |
1500f7ea | 2008 | ret = to->sync_to(to, from, seqno); |
e3a5a225 BW |
2009 | if (!ret) |
2010 | from->sync_seqno[idx] = seqno; | |
2911a35b | 2011 | |
e3a5a225 | 2012 | return ret; |
2911a35b BW |
2013 | } |
2014 | ||
b5ffc9bc CW |
2015 | static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj) |
2016 | { | |
2017 | u32 old_write_domain, old_read_domains; | |
2018 | ||
b5ffc9bc CW |
2019 | /* Act a barrier for all accesses through the GTT */ |
2020 | mb(); | |
2021 | ||
2022 | /* Force a pagefault for domain tracking on next user access */ | |
2023 | i915_gem_release_mmap(obj); | |
2024 | ||
b97c3d9c KP |
2025 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) |
2026 | return; | |
2027 | ||
b5ffc9bc CW |
2028 | old_read_domains = obj->base.read_domains; |
2029 | old_write_domain = obj->base.write_domain; | |
2030 | ||
2031 | obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT; | |
2032 | obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT; | |
2033 | ||
2034 | trace_i915_gem_object_change_domain(obj, | |
2035 | old_read_domains, | |
2036 | old_write_domain); | |
2037 | } | |
2038 | ||
673a394b EA |
2039 | /** |
2040 | * Unbinds an object from the GTT aperture. | |
2041 | */ | |
0f973f27 | 2042 | int |
05394f39 | 2043 | i915_gem_object_unbind(struct drm_i915_gem_object *obj) |
673a394b | 2044 | { |
7bddb01f | 2045 | drm_i915_private_t *dev_priv = obj->base.dev->dev_private; |
673a394b EA |
2046 | int ret = 0; |
2047 | ||
05394f39 | 2048 | if (obj->gtt_space == NULL) |
673a394b EA |
2049 | return 0; |
2050 | ||
05394f39 | 2051 | if (obj->pin_count != 0) { |
673a394b EA |
2052 | DRM_ERROR("Attempting to unbind pinned buffer\n"); |
2053 | return -EINVAL; | |
2054 | } | |
2055 | ||
a8198eea CW |
2056 | ret = i915_gem_object_finish_gpu(obj); |
2057 | if (ret == -ERESTARTSYS) | |
2058 | return ret; | |
2059 | /* Continue on if we fail due to EIO, the GPU is hung so we | |
2060 | * should be safe and we need to cleanup or else we might | |
2061 | * cause memory corruption through use-after-free. | |
2062 | */ | |
2063 | ||
b5ffc9bc | 2064 | i915_gem_object_finish_gtt(obj); |
5323fd04 | 2065 | |
673a394b EA |
2066 | /* Move the object to the CPU domain to ensure that |
2067 | * any possible CPU writes while it's not in the GTT | |
a8198eea | 2068 | * are flushed when we go to remap it. |
673a394b | 2069 | */ |
a8198eea CW |
2070 | if (ret == 0) |
2071 | ret = i915_gem_object_set_to_cpu_domain(obj, 1); | |
8dc1775d | 2072 | if (ret == -ERESTARTSYS) |
673a394b | 2073 | return ret; |
812ed492 | 2074 | if (ret) { |
a8198eea CW |
2075 | /* In the event of a disaster, abandon all caches and |
2076 | * hope for the best. | |
2077 | */ | |
812ed492 | 2078 | i915_gem_clflush_object(obj); |
05394f39 | 2079 | obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
812ed492 | 2080 | } |
673a394b | 2081 | |
96b47b65 | 2082 | /* release the fence reg _after_ flushing */ |
d9e86c0e CW |
2083 | ret = i915_gem_object_put_fence(obj); |
2084 | if (ret == -ERESTARTSYS) | |
2085 | return ret; | |
96b47b65 | 2086 | |
db53a302 CW |
2087 | trace_i915_gem_object_unbind(obj); |
2088 | ||
74898d7e DV |
2089 | if (obj->has_global_gtt_mapping) |
2090 | i915_gem_gtt_unbind_object(obj); | |
7bddb01f DV |
2091 | if (obj->has_aliasing_ppgtt_mapping) { |
2092 | i915_ppgtt_unbind_object(dev_priv->mm.aliasing_ppgtt, obj); | |
2093 | obj->has_aliasing_ppgtt_mapping = 0; | |
2094 | } | |
74163907 | 2095 | i915_gem_gtt_finish_object(obj); |
7bddb01f | 2096 | |
e5281ccd | 2097 | i915_gem_object_put_pages_gtt(obj); |
673a394b | 2098 | |
6299f992 | 2099 | list_del_init(&obj->gtt_list); |
05394f39 | 2100 | list_del_init(&obj->mm_list); |
75e9e915 | 2101 | /* Avoid an unnecessary call to unbind on rebind. */ |
05394f39 | 2102 | obj->map_and_fenceable = true; |
673a394b | 2103 | |
05394f39 CW |
2104 | drm_mm_put_block(obj->gtt_space); |
2105 | obj->gtt_space = NULL; | |
2106 | obj->gtt_offset = 0; | |
673a394b | 2107 | |
05394f39 | 2108 | if (i915_gem_object_is_purgeable(obj)) |
963b4836 CW |
2109 | i915_gem_object_truncate(obj); |
2110 | ||
8dc1775d | 2111 | return ret; |
673a394b EA |
2112 | } |
2113 | ||
88241785 | 2114 | int |
db53a302 | 2115 | i915_gem_flush_ring(struct intel_ring_buffer *ring, |
54cf91dc CW |
2116 | uint32_t invalidate_domains, |
2117 | uint32_t flush_domains) | |
2118 | { | |
88241785 CW |
2119 | int ret; |
2120 | ||
36d527de CW |
2121 | if (((invalidate_domains | flush_domains) & I915_GEM_GPU_DOMAINS) == 0) |
2122 | return 0; | |
2123 | ||
db53a302 CW |
2124 | trace_i915_gem_ring_flush(ring, invalidate_domains, flush_domains); |
2125 | ||
88241785 CW |
2126 | ret = ring->flush(ring, invalidate_domains, flush_domains); |
2127 | if (ret) | |
2128 | return ret; | |
2129 | ||
36d527de CW |
2130 | if (flush_domains & I915_GEM_GPU_DOMAINS) |
2131 | i915_gem_process_flushing_list(ring, flush_domains); | |
2132 | ||
88241785 | 2133 | return 0; |
54cf91dc CW |
2134 | } |
2135 | ||
b93f9cf1 | 2136 | static int i915_ring_idle(struct intel_ring_buffer *ring, bool do_retire) |
a56ba56c | 2137 | { |
88241785 CW |
2138 | int ret; |
2139 | ||
395b70be | 2140 | if (list_empty(&ring->gpu_write_list) && list_empty(&ring->active_list)) |
64193406 CW |
2141 | return 0; |
2142 | ||
88241785 | 2143 | if (!list_empty(&ring->gpu_write_list)) { |
db53a302 | 2144 | ret = i915_gem_flush_ring(ring, |
0ac74c6b | 2145 | I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS); |
88241785 CW |
2146 | if (ret) |
2147 | return ret; | |
2148 | } | |
2149 | ||
b93f9cf1 BW |
2150 | return i915_wait_request(ring, i915_gem_next_request_seqno(ring), |
2151 | do_retire); | |
a56ba56c CW |
2152 | } |
2153 | ||
b93f9cf1 | 2154 | int i915_gpu_idle(struct drm_device *dev, bool do_retire) |
4df2faf4 DV |
2155 | { |
2156 | drm_i915_private_t *dev_priv = dev->dev_private; | |
1ec14ad3 | 2157 | int ret, i; |
4df2faf4 | 2158 | |
4df2faf4 | 2159 | /* Flush everything onto the inactive list. */ |
1ec14ad3 | 2160 | for (i = 0; i < I915_NUM_RINGS; i++) { |
b93f9cf1 | 2161 | ret = i915_ring_idle(&dev_priv->ring[i], do_retire); |
1ec14ad3 CW |
2162 | if (ret) |
2163 | return ret; | |
2164 | } | |
4df2faf4 | 2165 | |
8a1a49f9 | 2166 | return 0; |
4df2faf4 DV |
2167 | } |
2168 | ||
c6642782 DV |
2169 | static int sandybridge_write_fence_reg(struct drm_i915_gem_object *obj, |
2170 | struct intel_ring_buffer *pipelined) | |
4e901fdc | 2171 | { |
05394f39 | 2172 | struct drm_device *dev = obj->base.dev; |
4e901fdc | 2173 | drm_i915_private_t *dev_priv = dev->dev_private; |
05394f39 CW |
2174 | u32 size = obj->gtt_space->size; |
2175 | int regnum = obj->fence_reg; | |
4e901fdc EA |
2176 | uint64_t val; |
2177 | ||
05394f39 | 2178 | val = (uint64_t)((obj->gtt_offset + size - 4096) & |
c6642782 | 2179 | 0xfffff000) << 32; |
05394f39 CW |
2180 | val |= obj->gtt_offset & 0xfffff000; |
2181 | val |= (uint64_t)((obj->stride / 128) - 1) << | |
4e901fdc EA |
2182 | SANDYBRIDGE_FENCE_PITCH_SHIFT; |
2183 | ||
05394f39 | 2184 | if (obj->tiling_mode == I915_TILING_Y) |
4e901fdc EA |
2185 | val |= 1 << I965_FENCE_TILING_Y_SHIFT; |
2186 | val |= I965_FENCE_REG_VALID; | |
2187 | ||
c6642782 DV |
2188 | if (pipelined) { |
2189 | int ret = intel_ring_begin(pipelined, 6); | |
2190 | if (ret) | |
2191 | return ret; | |
2192 | ||
2193 | intel_ring_emit(pipelined, MI_NOOP); | |
2194 | intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(2)); | |
2195 | intel_ring_emit(pipelined, FENCE_REG_SANDYBRIDGE_0 + regnum*8); | |
2196 | intel_ring_emit(pipelined, (u32)val); | |
2197 | intel_ring_emit(pipelined, FENCE_REG_SANDYBRIDGE_0 + regnum*8 + 4); | |
2198 | intel_ring_emit(pipelined, (u32)(val >> 32)); | |
2199 | intel_ring_advance(pipelined); | |
2200 | } else | |
2201 | I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + regnum * 8, val); | |
2202 | ||
2203 | return 0; | |
4e901fdc EA |
2204 | } |
2205 | ||
c6642782 DV |
2206 | static int i965_write_fence_reg(struct drm_i915_gem_object *obj, |
2207 | struct intel_ring_buffer *pipelined) | |
de151cf6 | 2208 | { |
05394f39 | 2209 | struct drm_device *dev = obj->base.dev; |
de151cf6 | 2210 | drm_i915_private_t *dev_priv = dev->dev_private; |
05394f39 CW |
2211 | u32 size = obj->gtt_space->size; |
2212 | int regnum = obj->fence_reg; | |
de151cf6 JB |
2213 | uint64_t val; |
2214 | ||
05394f39 | 2215 | val = (uint64_t)((obj->gtt_offset + size - 4096) & |
de151cf6 | 2216 | 0xfffff000) << 32; |
05394f39 CW |
2217 | val |= obj->gtt_offset & 0xfffff000; |
2218 | val |= ((obj->stride / 128) - 1) << I965_FENCE_PITCH_SHIFT; | |
2219 | if (obj->tiling_mode == I915_TILING_Y) | |
de151cf6 JB |
2220 | val |= 1 << I965_FENCE_TILING_Y_SHIFT; |
2221 | val |= I965_FENCE_REG_VALID; | |
2222 | ||
c6642782 DV |
2223 | if (pipelined) { |
2224 | int ret = intel_ring_begin(pipelined, 6); | |
2225 | if (ret) | |
2226 | return ret; | |
2227 | ||
2228 | intel_ring_emit(pipelined, MI_NOOP); | |
2229 | intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(2)); | |
2230 | intel_ring_emit(pipelined, FENCE_REG_965_0 + regnum*8); | |
2231 | intel_ring_emit(pipelined, (u32)val); | |
2232 | intel_ring_emit(pipelined, FENCE_REG_965_0 + regnum*8 + 4); | |
2233 | intel_ring_emit(pipelined, (u32)(val >> 32)); | |
2234 | intel_ring_advance(pipelined); | |
2235 | } else | |
2236 | I915_WRITE64(FENCE_REG_965_0 + regnum * 8, val); | |
2237 | ||
2238 | return 0; | |
de151cf6 JB |
2239 | } |
2240 | ||
c6642782 DV |
2241 | static int i915_write_fence_reg(struct drm_i915_gem_object *obj, |
2242 | struct intel_ring_buffer *pipelined) | |
de151cf6 | 2243 | { |
05394f39 | 2244 | struct drm_device *dev = obj->base.dev; |
de151cf6 | 2245 | drm_i915_private_t *dev_priv = dev->dev_private; |
05394f39 | 2246 | u32 size = obj->gtt_space->size; |
c6642782 | 2247 | u32 fence_reg, val, pitch_val; |
0f973f27 | 2248 | int tile_width; |
de151cf6 | 2249 | |
c6642782 DV |
2250 | if (WARN((obj->gtt_offset & ~I915_FENCE_START_MASK) || |
2251 | (size & -size) != size || | |
2252 | (obj->gtt_offset & (size - 1)), | |
2253 | "object 0x%08x [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n", | |
2254 | obj->gtt_offset, obj->map_and_fenceable, size)) | |
2255 | return -EINVAL; | |
de151cf6 | 2256 | |
c6642782 | 2257 | if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev)) |
0f973f27 | 2258 | tile_width = 128; |
de151cf6 | 2259 | else |
0f973f27 JB |
2260 | tile_width = 512; |
2261 | ||
2262 | /* Note: pitch better be a power of two tile widths */ | |
05394f39 | 2263 | pitch_val = obj->stride / tile_width; |
0f973f27 | 2264 | pitch_val = ffs(pitch_val) - 1; |
de151cf6 | 2265 | |
05394f39 CW |
2266 | val = obj->gtt_offset; |
2267 | if (obj->tiling_mode == I915_TILING_Y) | |
de151cf6 | 2268 | val |= 1 << I830_FENCE_TILING_Y_SHIFT; |
a00b10c3 | 2269 | val |= I915_FENCE_SIZE_BITS(size); |
de151cf6 JB |
2270 | val |= pitch_val << I830_FENCE_PITCH_SHIFT; |
2271 | val |= I830_FENCE_REG_VALID; | |
2272 | ||
05394f39 | 2273 | fence_reg = obj->fence_reg; |
a00b10c3 CW |
2274 | if (fence_reg < 8) |
2275 | fence_reg = FENCE_REG_830_0 + fence_reg * 4; | |
dc529a4f | 2276 | else |
a00b10c3 | 2277 | fence_reg = FENCE_REG_945_8 + (fence_reg - 8) * 4; |
c6642782 DV |
2278 | |
2279 | if (pipelined) { | |
2280 | int ret = intel_ring_begin(pipelined, 4); | |
2281 | if (ret) | |
2282 | return ret; | |
2283 | ||
2284 | intel_ring_emit(pipelined, MI_NOOP); | |
2285 | intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(1)); | |
2286 | intel_ring_emit(pipelined, fence_reg); | |
2287 | intel_ring_emit(pipelined, val); | |
2288 | intel_ring_advance(pipelined); | |
2289 | } else | |
2290 | I915_WRITE(fence_reg, val); | |
2291 | ||
2292 | return 0; | |
de151cf6 JB |
2293 | } |
2294 | ||
c6642782 DV |
2295 | static int i830_write_fence_reg(struct drm_i915_gem_object *obj, |
2296 | struct intel_ring_buffer *pipelined) | |
de151cf6 | 2297 | { |
05394f39 | 2298 | struct drm_device *dev = obj->base.dev; |
de151cf6 | 2299 | drm_i915_private_t *dev_priv = dev->dev_private; |
05394f39 CW |
2300 | u32 size = obj->gtt_space->size; |
2301 | int regnum = obj->fence_reg; | |
de151cf6 JB |
2302 | uint32_t val; |
2303 | uint32_t pitch_val; | |
2304 | ||
c6642782 DV |
2305 | if (WARN((obj->gtt_offset & ~I830_FENCE_START_MASK) || |
2306 | (size & -size) != size || | |
2307 | (obj->gtt_offset & (size - 1)), | |
2308 | "object 0x%08x not 512K or pot-size 0x%08x aligned\n", | |
2309 | obj->gtt_offset, size)) | |
2310 | return -EINVAL; | |
de151cf6 | 2311 | |
05394f39 | 2312 | pitch_val = obj->stride / 128; |
e76a16de | 2313 | pitch_val = ffs(pitch_val) - 1; |
e76a16de | 2314 | |
05394f39 CW |
2315 | val = obj->gtt_offset; |
2316 | if (obj->tiling_mode == I915_TILING_Y) | |
de151cf6 | 2317 | val |= 1 << I830_FENCE_TILING_Y_SHIFT; |
c6642782 | 2318 | val |= I830_FENCE_SIZE_BITS(size); |
de151cf6 JB |
2319 | val |= pitch_val << I830_FENCE_PITCH_SHIFT; |
2320 | val |= I830_FENCE_REG_VALID; | |
2321 | ||
c6642782 DV |
2322 | if (pipelined) { |
2323 | int ret = intel_ring_begin(pipelined, 4); | |
2324 | if (ret) | |
2325 | return ret; | |
2326 | ||
2327 | intel_ring_emit(pipelined, MI_NOOP); | |
2328 | intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(1)); | |
2329 | intel_ring_emit(pipelined, FENCE_REG_830_0 + regnum*4); | |
2330 | intel_ring_emit(pipelined, val); | |
2331 | intel_ring_advance(pipelined); | |
2332 | } else | |
2333 | I915_WRITE(FENCE_REG_830_0 + regnum * 4, val); | |
2334 | ||
2335 | return 0; | |
de151cf6 JB |
2336 | } |
2337 | ||
d9e86c0e CW |
2338 | static bool ring_passed_seqno(struct intel_ring_buffer *ring, u32 seqno) |
2339 | { | |
2340 | return i915_seqno_passed(ring->get_seqno(ring), seqno); | |
2341 | } | |
2342 | ||
2343 | static int | |
2344 | i915_gem_object_flush_fence(struct drm_i915_gem_object *obj, | |
ce453d81 | 2345 | struct intel_ring_buffer *pipelined) |
d9e86c0e CW |
2346 | { |
2347 | int ret; | |
2348 | ||
2349 | if (obj->fenced_gpu_access) { | |
88241785 | 2350 | if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) { |
db53a302 | 2351 | ret = i915_gem_flush_ring(obj->last_fenced_ring, |
88241785 CW |
2352 | 0, obj->base.write_domain); |
2353 | if (ret) | |
2354 | return ret; | |
2355 | } | |
d9e86c0e CW |
2356 | |
2357 | obj->fenced_gpu_access = false; | |
2358 | } | |
2359 | ||
2360 | if (obj->last_fenced_seqno && pipelined != obj->last_fenced_ring) { | |
2361 | if (!ring_passed_seqno(obj->last_fenced_ring, | |
2362 | obj->last_fenced_seqno)) { | |
db53a302 | 2363 | ret = i915_wait_request(obj->last_fenced_ring, |
b93f9cf1 BW |
2364 | obj->last_fenced_seqno, |
2365 | true); | |
d9e86c0e CW |
2366 | if (ret) |
2367 | return ret; | |
2368 | } | |
2369 | ||
2370 | obj->last_fenced_seqno = 0; | |
2371 | obj->last_fenced_ring = NULL; | |
2372 | } | |
2373 | ||
63256ec5 CW |
2374 | /* Ensure that all CPU reads are completed before installing a fence |
2375 | * and all writes before removing the fence. | |
2376 | */ | |
2377 | if (obj->base.read_domains & I915_GEM_DOMAIN_GTT) | |
2378 | mb(); | |
2379 | ||
d9e86c0e CW |
2380 | return 0; |
2381 | } | |
2382 | ||
2383 | int | |
2384 | i915_gem_object_put_fence(struct drm_i915_gem_object *obj) | |
2385 | { | |
2386 | int ret; | |
2387 | ||
2388 | if (obj->tiling_mode) | |
2389 | i915_gem_release_mmap(obj); | |
2390 | ||
ce453d81 | 2391 | ret = i915_gem_object_flush_fence(obj, NULL); |
d9e86c0e CW |
2392 | if (ret) |
2393 | return ret; | |
2394 | ||
2395 | if (obj->fence_reg != I915_FENCE_REG_NONE) { | |
2396 | struct drm_i915_private *dev_priv = obj->base.dev->dev_private; | |
1690e1eb CW |
2397 | |
2398 | WARN_ON(dev_priv->fence_regs[obj->fence_reg].pin_count); | |
d9e86c0e CW |
2399 | i915_gem_clear_fence_reg(obj->base.dev, |
2400 | &dev_priv->fence_regs[obj->fence_reg]); | |
2401 | ||
2402 | obj->fence_reg = I915_FENCE_REG_NONE; | |
2403 | } | |
2404 | ||
2405 | return 0; | |
2406 | } | |
2407 | ||
2408 | static struct drm_i915_fence_reg * | |
2409 | i915_find_fence_reg(struct drm_device *dev, | |
2410 | struct intel_ring_buffer *pipelined) | |
ae3db24a | 2411 | { |
ae3db24a | 2412 | struct drm_i915_private *dev_priv = dev->dev_private; |
d9e86c0e CW |
2413 | struct drm_i915_fence_reg *reg, *first, *avail; |
2414 | int i; | |
ae3db24a DV |
2415 | |
2416 | /* First try to find a free reg */ | |
d9e86c0e | 2417 | avail = NULL; |
ae3db24a DV |
2418 | for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) { |
2419 | reg = &dev_priv->fence_regs[i]; | |
2420 | if (!reg->obj) | |
d9e86c0e | 2421 | return reg; |
ae3db24a | 2422 | |
1690e1eb | 2423 | if (!reg->pin_count) |
d9e86c0e | 2424 | avail = reg; |
ae3db24a DV |
2425 | } |
2426 | ||
d9e86c0e CW |
2427 | if (avail == NULL) |
2428 | return NULL; | |
ae3db24a DV |
2429 | |
2430 | /* None available, try to steal one or wait for a user to finish */ | |
d9e86c0e CW |
2431 | avail = first = NULL; |
2432 | list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) { | |
1690e1eb | 2433 | if (reg->pin_count) |
ae3db24a DV |
2434 | continue; |
2435 | ||
d9e86c0e CW |
2436 | if (first == NULL) |
2437 | first = reg; | |
2438 | ||
2439 | if (!pipelined || | |
2440 | !reg->obj->last_fenced_ring || | |
2441 | reg->obj->last_fenced_ring == pipelined) { | |
2442 | avail = reg; | |
2443 | break; | |
2444 | } | |
ae3db24a DV |
2445 | } |
2446 | ||
d9e86c0e CW |
2447 | if (avail == NULL) |
2448 | avail = first; | |
ae3db24a | 2449 | |
a00b10c3 | 2450 | return avail; |
ae3db24a DV |
2451 | } |
2452 | ||
de151cf6 | 2453 | /** |
9a5a53b3 | 2454 | * i915_gem_object_get_fence - set up fencing for an object |
de151cf6 | 2455 | * @obj: object to map through a fence reg |
d9e86c0e | 2456 | * @pipelined: ring on which to queue the change, or NULL for CPU access |
de151cf6 JB |
2457 | * |
2458 | * When mapping objects through the GTT, userspace wants to be able to write | |
2459 | * to them without having to worry about swizzling if the object is tiled. | |
de151cf6 JB |
2460 | * This function walks the fence regs looking for a free one for @obj, |
2461 | * stealing one if it can't find any. | |
2462 | * | |
2463 | * It then sets up the reg based on the object's properties: address, pitch | |
2464 | * and tiling format. | |
9a5a53b3 CW |
2465 | * |
2466 | * For an untiled surface, this removes any existing fence. | |
de151cf6 | 2467 | */ |
8c4b8c3f | 2468 | int |
d9e86c0e | 2469 | i915_gem_object_get_fence(struct drm_i915_gem_object *obj, |
ce453d81 | 2470 | struct intel_ring_buffer *pipelined) |
de151cf6 | 2471 | { |
05394f39 | 2472 | struct drm_device *dev = obj->base.dev; |
79e53945 | 2473 | struct drm_i915_private *dev_priv = dev->dev_private; |
d9e86c0e | 2474 | struct drm_i915_fence_reg *reg; |
ae3db24a | 2475 | int ret; |
de151cf6 | 2476 | |
9a5a53b3 CW |
2477 | if (obj->tiling_mode == I915_TILING_NONE) |
2478 | return i915_gem_object_put_fence(obj); | |
2479 | ||
6bda10d1 CW |
2480 | /* XXX disable pipelining. There are bugs. Shocking. */ |
2481 | pipelined = NULL; | |
2482 | ||
d9e86c0e | 2483 | /* Just update our place in the LRU if our fence is getting reused. */ |
05394f39 CW |
2484 | if (obj->fence_reg != I915_FENCE_REG_NONE) { |
2485 | reg = &dev_priv->fence_regs[obj->fence_reg]; | |
007cc8ac | 2486 | list_move_tail(®->lru_list, &dev_priv->mm.fence_list); |
d9e86c0e | 2487 | |
29c5a587 CW |
2488 | if (obj->tiling_changed) { |
2489 | ret = i915_gem_object_flush_fence(obj, pipelined); | |
2490 | if (ret) | |
2491 | return ret; | |
2492 | ||
2493 | if (!obj->fenced_gpu_access && !obj->last_fenced_seqno) | |
2494 | pipelined = NULL; | |
2495 | ||
2496 | if (pipelined) { | |
2497 | reg->setup_seqno = | |
2498 | i915_gem_next_request_seqno(pipelined); | |
2499 | obj->last_fenced_seqno = reg->setup_seqno; | |
2500 | obj->last_fenced_ring = pipelined; | |
2501 | } | |
2502 | ||
2503 | goto update; | |
2504 | } | |
d9e86c0e CW |
2505 | |
2506 | if (!pipelined) { | |
2507 | if (reg->setup_seqno) { | |
2508 | if (!ring_passed_seqno(obj->last_fenced_ring, | |
2509 | reg->setup_seqno)) { | |
db53a302 | 2510 | ret = i915_wait_request(obj->last_fenced_ring, |
b93f9cf1 BW |
2511 | reg->setup_seqno, |
2512 | true); | |
d9e86c0e CW |
2513 | if (ret) |
2514 | return ret; | |
2515 | } | |
2516 | ||
2517 | reg->setup_seqno = 0; | |
2518 | } | |
2519 | } else if (obj->last_fenced_ring && | |
2520 | obj->last_fenced_ring != pipelined) { | |
ce453d81 | 2521 | ret = i915_gem_object_flush_fence(obj, pipelined); |
d9e86c0e CW |
2522 | if (ret) |
2523 | return ret; | |
d9e86c0e CW |
2524 | } |
2525 | ||
a09ba7fa EA |
2526 | return 0; |
2527 | } | |
2528 | ||
d9e86c0e CW |
2529 | reg = i915_find_fence_reg(dev, pipelined); |
2530 | if (reg == NULL) | |
39965b37 | 2531 | return -EDEADLK; |
de151cf6 | 2532 | |
ce453d81 | 2533 | ret = i915_gem_object_flush_fence(obj, pipelined); |
d9e86c0e | 2534 | if (ret) |
ae3db24a | 2535 | return ret; |
de151cf6 | 2536 | |
d9e86c0e CW |
2537 | if (reg->obj) { |
2538 | struct drm_i915_gem_object *old = reg->obj; | |
2539 | ||
2540 | drm_gem_object_reference(&old->base); | |
2541 | ||
2542 | if (old->tiling_mode) | |
2543 | i915_gem_release_mmap(old); | |
2544 | ||
ce453d81 | 2545 | ret = i915_gem_object_flush_fence(old, pipelined); |
d9e86c0e CW |
2546 | if (ret) { |
2547 | drm_gem_object_unreference(&old->base); | |
2548 | return ret; | |
2549 | } | |
2550 | ||
2551 | if (old->last_fenced_seqno == 0 && obj->last_fenced_seqno == 0) | |
2552 | pipelined = NULL; | |
2553 | ||
2554 | old->fence_reg = I915_FENCE_REG_NONE; | |
2555 | old->last_fenced_ring = pipelined; | |
2556 | old->last_fenced_seqno = | |
db53a302 | 2557 | pipelined ? i915_gem_next_request_seqno(pipelined) : 0; |
d9e86c0e CW |
2558 | |
2559 | drm_gem_object_unreference(&old->base); | |
2560 | } else if (obj->last_fenced_seqno == 0) | |
2561 | pipelined = NULL; | |
a09ba7fa | 2562 | |
de151cf6 | 2563 | reg->obj = obj; |
d9e86c0e CW |
2564 | list_move_tail(®->lru_list, &dev_priv->mm.fence_list); |
2565 | obj->fence_reg = reg - dev_priv->fence_regs; | |
2566 | obj->last_fenced_ring = pipelined; | |
de151cf6 | 2567 | |
d9e86c0e | 2568 | reg->setup_seqno = |
db53a302 | 2569 | pipelined ? i915_gem_next_request_seqno(pipelined) : 0; |
d9e86c0e CW |
2570 | obj->last_fenced_seqno = reg->setup_seqno; |
2571 | ||
2572 | update: | |
2573 | obj->tiling_changed = false; | |
e259befd | 2574 | switch (INTEL_INFO(dev)->gen) { |
25aebfc3 | 2575 | case 7: |
e259befd | 2576 | case 6: |
c6642782 | 2577 | ret = sandybridge_write_fence_reg(obj, pipelined); |
e259befd CW |
2578 | break; |
2579 | case 5: | |
2580 | case 4: | |
c6642782 | 2581 | ret = i965_write_fence_reg(obj, pipelined); |
e259befd CW |
2582 | break; |
2583 | case 3: | |
c6642782 | 2584 | ret = i915_write_fence_reg(obj, pipelined); |
e259befd CW |
2585 | break; |
2586 | case 2: | |
c6642782 | 2587 | ret = i830_write_fence_reg(obj, pipelined); |
e259befd CW |
2588 | break; |
2589 | } | |
d9ddcb96 | 2590 | |
c6642782 | 2591 | return ret; |
de151cf6 JB |
2592 | } |
2593 | ||
2594 | /** | |
2595 | * i915_gem_clear_fence_reg - clear out fence register info | |
2596 | * @obj: object to clear | |
2597 | * | |
2598 | * Zeroes out the fence register itself and clears out the associated | |
05394f39 | 2599 | * data structures in dev_priv and obj. |
de151cf6 JB |
2600 | */ |
2601 | static void | |
d9e86c0e CW |
2602 | i915_gem_clear_fence_reg(struct drm_device *dev, |
2603 | struct drm_i915_fence_reg *reg) | |
de151cf6 | 2604 | { |
79e53945 | 2605 | drm_i915_private_t *dev_priv = dev->dev_private; |
d9e86c0e | 2606 | uint32_t fence_reg = reg - dev_priv->fence_regs; |
de151cf6 | 2607 | |
e259befd | 2608 | switch (INTEL_INFO(dev)->gen) { |
25aebfc3 | 2609 | case 7: |
e259befd | 2610 | case 6: |
d9e86c0e | 2611 | I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + fence_reg*8, 0); |
e259befd CW |
2612 | break; |
2613 | case 5: | |
2614 | case 4: | |
d9e86c0e | 2615 | I915_WRITE64(FENCE_REG_965_0 + fence_reg*8, 0); |
e259befd CW |
2616 | break; |
2617 | case 3: | |
d9e86c0e CW |
2618 | if (fence_reg >= 8) |
2619 | fence_reg = FENCE_REG_945_8 + (fence_reg - 8) * 4; | |
dc529a4f | 2620 | else |
e259befd | 2621 | case 2: |
d9e86c0e | 2622 | fence_reg = FENCE_REG_830_0 + fence_reg * 4; |
dc529a4f EA |
2623 | |
2624 | I915_WRITE(fence_reg, 0); | |
e259befd | 2625 | break; |
dc529a4f | 2626 | } |
de151cf6 | 2627 | |
007cc8ac | 2628 | list_del_init(®->lru_list); |
d9e86c0e CW |
2629 | reg->obj = NULL; |
2630 | reg->setup_seqno = 0; | |
1690e1eb | 2631 | reg->pin_count = 0; |
52dc7d32 CW |
2632 | } |
2633 | ||
673a394b EA |
2634 | /** |
2635 | * Finds free space in the GTT aperture and binds the object there. | |
2636 | */ | |
2637 | static int | |
05394f39 | 2638 | i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj, |
920afa77 | 2639 | unsigned alignment, |
75e9e915 | 2640 | bool map_and_fenceable) |
673a394b | 2641 | { |
05394f39 | 2642 | struct drm_device *dev = obj->base.dev; |
673a394b | 2643 | drm_i915_private_t *dev_priv = dev->dev_private; |
673a394b | 2644 | struct drm_mm_node *free_space; |
a00b10c3 | 2645 | gfp_t gfpmask = __GFP_NORETRY | __GFP_NOWARN; |
5e783301 | 2646 | u32 size, fence_size, fence_alignment, unfenced_alignment; |
75e9e915 | 2647 | bool mappable, fenceable; |
07f73f69 | 2648 | int ret; |
673a394b | 2649 | |
05394f39 | 2650 | if (obj->madv != I915_MADV_WILLNEED) { |
3ef94daa CW |
2651 | DRM_ERROR("Attempting to bind a purgeable object\n"); |
2652 | return -EINVAL; | |
2653 | } | |
2654 | ||
e28f8711 CW |
2655 | fence_size = i915_gem_get_gtt_size(dev, |
2656 | obj->base.size, | |
2657 | obj->tiling_mode); | |
2658 | fence_alignment = i915_gem_get_gtt_alignment(dev, | |
2659 | obj->base.size, | |
2660 | obj->tiling_mode); | |
2661 | unfenced_alignment = | |
2662 | i915_gem_get_unfenced_gtt_alignment(dev, | |
2663 | obj->base.size, | |
2664 | obj->tiling_mode); | |
a00b10c3 | 2665 | |
673a394b | 2666 | if (alignment == 0) |
5e783301 DV |
2667 | alignment = map_and_fenceable ? fence_alignment : |
2668 | unfenced_alignment; | |
75e9e915 | 2669 | if (map_and_fenceable && alignment & (fence_alignment - 1)) { |
673a394b EA |
2670 | DRM_ERROR("Invalid object alignment requested %u\n", alignment); |
2671 | return -EINVAL; | |
2672 | } | |
2673 | ||
05394f39 | 2674 | size = map_and_fenceable ? fence_size : obj->base.size; |
a00b10c3 | 2675 | |
654fc607 CW |
2676 | /* If the object is bigger than the entire aperture, reject it early |
2677 | * before evicting everything in a vain attempt to find space. | |
2678 | */ | |
05394f39 | 2679 | if (obj->base.size > |
75e9e915 | 2680 | (map_and_fenceable ? dev_priv->mm.gtt_mappable_end : dev_priv->mm.gtt_total)) { |
654fc607 CW |
2681 | DRM_ERROR("Attempting to bind an object larger than the aperture\n"); |
2682 | return -E2BIG; | |
2683 | } | |
2684 | ||
673a394b | 2685 | search_free: |
75e9e915 | 2686 | if (map_and_fenceable) |
920afa77 DV |
2687 | free_space = |
2688 | drm_mm_search_free_in_range(&dev_priv->mm.gtt_space, | |
a00b10c3 | 2689 | size, alignment, 0, |
920afa77 DV |
2690 | dev_priv->mm.gtt_mappable_end, |
2691 | 0); | |
2692 | else | |
2693 | free_space = drm_mm_search_free(&dev_priv->mm.gtt_space, | |
a00b10c3 | 2694 | size, alignment, 0); |
920afa77 DV |
2695 | |
2696 | if (free_space != NULL) { | |
75e9e915 | 2697 | if (map_and_fenceable) |
05394f39 | 2698 | obj->gtt_space = |
920afa77 | 2699 | drm_mm_get_block_range_generic(free_space, |
a00b10c3 | 2700 | size, alignment, 0, |
920afa77 DV |
2701 | dev_priv->mm.gtt_mappable_end, |
2702 | 0); | |
2703 | else | |
05394f39 | 2704 | obj->gtt_space = |
a00b10c3 | 2705 | drm_mm_get_block(free_space, size, alignment); |
920afa77 | 2706 | } |
05394f39 | 2707 | if (obj->gtt_space == NULL) { |
673a394b EA |
2708 | /* If the gtt is empty and we're still having trouble |
2709 | * fitting our object in, we're out of memory. | |
2710 | */ | |
75e9e915 DV |
2711 | ret = i915_gem_evict_something(dev, size, alignment, |
2712 | map_and_fenceable); | |
9731129c | 2713 | if (ret) |
673a394b | 2714 | return ret; |
9731129c | 2715 | |
673a394b EA |
2716 | goto search_free; |
2717 | } | |
2718 | ||
e5281ccd | 2719 | ret = i915_gem_object_get_pages_gtt(obj, gfpmask); |
673a394b | 2720 | if (ret) { |
05394f39 CW |
2721 | drm_mm_put_block(obj->gtt_space); |
2722 | obj->gtt_space = NULL; | |
07f73f69 CW |
2723 | |
2724 | if (ret == -ENOMEM) { | |
809b6334 CW |
2725 | /* first try to reclaim some memory by clearing the GTT */ |
2726 | ret = i915_gem_evict_everything(dev, false); | |
07f73f69 | 2727 | if (ret) { |
07f73f69 | 2728 | /* now try to shrink everyone else */ |
4bdadb97 CW |
2729 | if (gfpmask) { |
2730 | gfpmask = 0; | |
2731 | goto search_free; | |
07f73f69 CW |
2732 | } |
2733 | ||
809b6334 | 2734 | return -ENOMEM; |
07f73f69 CW |
2735 | } |
2736 | ||
2737 | goto search_free; | |
2738 | } | |
2739 | ||
673a394b EA |
2740 | return ret; |
2741 | } | |
2742 | ||
74163907 | 2743 | ret = i915_gem_gtt_prepare_object(obj); |
7c2e6fdf | 2744 | if (ret) { |
e5281ccd | 2745 | i915_gem_object_put_pages_gtt(obj); |
05394f39 CW |
2746 | drm_mm_put_block(obj->gtt_space); |
2747 | obj->gtt_space = NULL; | |
07f73f69 | 2748 | |
809b6334 | 2749 | if (i915_gem_evict_everything(dev, false)) |
07f73f69 | 2750 | return ret; |
07f73f69 CW |
2751 | |
2752 | goto search_free; | |
673a394b | 2753 | } |
673a394b | 2754 | |
0ebb9829 DV |
2755 | if (!dev_priv->mm.aliasing_ppgtt) |
2756 | i915_gem_gtt_bind_object(obj, obj->cache_level); | |
673a394b | 2757 | |
6299f992 | 2758 | list_add_tail(&obj->gtt_list, &dev_priv->mm.gtt_list); |
05394f39 | 2759 | list_add_tail(&obj->mm_list, &dev_priv->mm.inactive_list); |
bf1a1092 | 2760 | |
673a394b EA |
2761 | /* Assert that the object is not currently in any GPU domain. As it |
2762 | * wasn't in the GTT, there shouldn't be any way it could have been in | |
2763 | * a GPU cache | |
2764 | */ | |
05394f39 CW |
2765 | BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS); |
2766 | BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS); | |
673a394b | 2767 | |
6299f992 | 2768 | obj->gtt_offset = obj->gtt_space->start; |
1c5d22f7 | 2769 | |
75e9e915 | 2770 | fenceable = |
05394f39 | 2771 | obj->gtt_space->size == fence_size && |
0206e353 | 2772 | (obj->gtt_space->start & (fence_alignment - 1)) == 0; |
a00b10c3 | 2773 | |
75e9e915 | 2774 | mappable = |
05394f39 | 2775 | obj->gtt_offset + obj->base.size <= dev_priv->mm.gtt_mappable_end; |
a00b10c3 | 2776 | |
05394f39 | 2777 | obj->map_and_fenceable = mappable && fenceable; |
75e9e915 | 2778 | |
db53a302 | 2779 | trace_i915_gem_object_bind(obj, map_and_fenceable); |
673a394b EA |
2780 | return 0; |
2781 | } | |
2782 | ||
2783 | void | |
05394f39 | 2784 | i915_gem_clflush_object(struct drm_i915_gem_object *obj) |
673a394b | 2785 | { |
673a394b EA |
2786 | /* If we don't have a page list set up, then we're not pinned |
2787 | * to GPU, and we can ignore the cache flush because it'll happen | |
2788 | * again at bind time. | |
2789 | */ | |
05394f39 | 2790 | if (obj->pages == NULL) |
673a394b EA |
2791 | return; |
2792 | ||
9c23f7fc CW |
2793 | /* If the GPU is snooping the contents of the CPU cache, |
2794 | * we do not need to manually clear the CPU cache lines. However, | |
2795 | * the caches are only snooped when the render cache is | |
2796 | * flushed/invalidated. As we always have to emit invalidations | |
2797 | * and flushes when moving into and out of the RENDER domain, correct | |
2798 | * snooping behaviour occurs naturally as the result of our domain | |
2799 | * tracking. | |
2800 | */ | |
2801 | if (obj->cache_level != I915_CACHE_NONE) | |
2802 | return; | |
2803 | ||
1c5d22f7 | 2804 | trace_i915_gem_object_clflush(obj); |
cfa16a0d | 2805 | |
05394f39 | 2806 | drm_clflush_pages(obj->pages, obj->base.size / PAGE_SIZE); |
673a394b EA |
2807 | } |
2808 | ||
e47c68e9 | 2809 | /** Flushes any GPU write domain for the object if it's dirty. */ |
88241785 | 2810 | static int |
3619df03 | 2811 | i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object *obj) |
e47c68e9 | 2812 | { |
05394f39 | 2813 | if ((obj->base.write_domain & I915_GEM_GPU_DOMAINS) == 0) |
88241785 | 2814 | return 0; |
e47c68e9 EA |
2815 | |
2816 | /* Queue the GPU write cache flushing we need. */ | |
db53a302 | 2817 | return i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain); |
e47c68e9 EA |
2818 | } |
2819 | ||
2820 | /** Flushes the GTT write domain for the object if it's dirty. */ | |
2821 | static void | |
05394f39 | 2822 | i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj) |
e47c68e9 | 2823 | { |
1c5d22f7 CW |
2824 | uint32_t old_write_domain; |
2825 | ||
05394f39 | 2826 | if (obj->base.write_domain != I915_GEM_DOMAIN_GTT) |
e47c68e9 EA |
2827 | return; |
2828 | ||
63256ec5 | 2829 | /* No actual flushing is required for the GTT write domain. Writes |
e47c68e9 EA |
2830 | * to it immediately go to main memory as far as we know, so there's |
2831 | * no chipset flush. It also doesn't land in render cache. | |
63256ec5 CW |
2832 | * |
2833 | * However, we do have to enforce the order so that all writes through | |
2834 | * the GTT land before any writes to the device, such as updates to | |
2835 | * the GATT itself. | |
e47c68e9 | 2836 | */ |
63256ec5 CW |
2837 | wmb(); |
2838 | ||
05394f39 CW |
2839 | old_write_domain = obj->base.write_domain; |
2840 | obj->base.write_domain = 0; | |
1c5d22f7 CW |
2841 | |
2842 | trace_i915_gem_object_change_domain(obj, | |
05394f39 | 2843 | obj->base.read_domains, |
1c5d22f7 | 2844 | old_write_domain); |
e47c68e9 EA |
2845 | } |
2846 | ||
2847 | /** Flushes the CPU write domain for the object if it's dirty. */ | |
2848 | static void | |
05394f39 | 2849 | i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj) |
e47c68e9 | 2850 | { |
1c5d22f7 | 2851 | uint32_t old_write_domain; |
e47c68e9 | 2852 | |
05394f39 | 2853 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) |
e47c68e9 EA |
2854 | return; |
2855 | ||
2856 | i915_gem_clflush_object(obj); | |
40ce6575 | 2857 | intel_gtt_chipset_flush(); |
05394f39 CW |
2858 | old_write_domain = obj->base.write_domain; |
2859 | obj->base.write_domain = 0; | |
1c5d22f7 CW |
2860 | |
2861 | trace_i915_gem_object_change_domain(obj, | |
05394f39 | 2862 | obj->base.read_domains, |
1c5d22f7 | 2863 | old_write_domain); |
e47c68e9 EA |
2864 | } |
2865 | ||
2ef7eeaa EA |
2866 | /** |
2867 | * Moves a single object to the GTT read, and possibly write domain. | |
2868 | * | |
2869 | * This function returns when the move is complete, including waiting on | |
2870 | * flushes to occur. | |
2871 | */ | |
79e53945 | 2872 | int |
2021746e | 2873 | i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write) |
2ef7eeaa | 2874 | { |
1c5d22f7 | 2875 | uint32_t old_write_domain, old_read_domains; |
e47c68e9 | 2876 | int ret; |
2ef7eeaa | 2877 | |
02354392 | 2878 | /* Not valid to be called on unbound objects. */ |
05394f39 | 2879 | if (obj->gtt_space == NULL) |
02354392 EA |
2880 | return -EINVAL; |
2881 | ||
8d7e3de1 CW |
2882 | if (obj->base.write_domain == I915_GEM_DOMAIN_GTT) |
2883 | return 0; | |
2884 | ||
88241785 CW |
2885 | ret = i915_gem_object_flush_gpu_write_domain(obj); |
2886 | if (ret) | |
2887 | return ret; | |
2888 | ||
87ca9c8a | 2889 | if (obj->pending_gpu_write || write) { |
ce453d81 | 2890 | ret = i915_gem_object_wait_rendering(obj); |
87ca9c8a CW |
2891 | if (ret) |
2892 | return ret; | |
2893 | } | |
2dafb1e0 | 2894 | |
7213342d | 2895 | i915_gem_object_flush_cpu_write_domain(obj); |
1c5d22f7 | 2896 | |
05394f39 CW |
2897 | old_write_domain = obj->base.write_domain; |
2898 | old_read_domains = obj->base.read_domains; | |
1c5d22f7 | 2899 | |
e47c68e9 EA |
2900 | /* It should now be out of any other write domains, and we can update |
2901 | * the domain values for our changes. | |
2902 | */ | |
05394f39 CW |
2903 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0); |
2904 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; | |
e47c68e9 | 2905 | if (write) { |
05394f39 CW |
2906 | obj->base.read_domains = I915_GEM_DOMAIN_GTT; |
2907 | obj->base.write_domain = I915_GEM_DOMAIN_GTT; | |
2908 | obj->dirty = 1; | |
2ef7eeaa EA |
2909 | } |
2910 | ||
1c5d22f7 CW |
2911 | trace_i915_gem_object_change_domain(obj, |
2912 | old_read_domains, | |
2913 | old_write_domain); | |
2914 | ||
e47c68e9 EA |
2915 | return 0; |
2916 | } | |
2917 | ||
e4ffd173 CW |
2918 | int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj, |
2919 | enum i915_cache_level cache_level) | |
2920 | { | |
7bddb01f DV |
2921 | struct drm_device *dev = obj->base.dev; |
2922 | drm_i915_private_t *dev_priv = dev->dev_private; | |
e4ffd173 CW |
2923 | int ret; |
2924 | ||
2925 | if (obj->cache_level == cache_level) | |
2926 | return 0; | |
2927 | ||
2928 | if (obj->pin_count) { | |
2929 | DRM_DEBUG("can not change the cache level of pinned objects\n"); | |
2930 | return -EBUSY; | |
2931 | } | |
2932 | ||
2933 | if (obj->gtt_space) { | |
2934 | ret = i915_gem_object_finish_gpu(obj); | |
2935 | if (ret) | |
2936 | return ret; | |
2937 | ||
2938 | i915_gem_object_finish_gtt(obj); | |
2939 | ||
2940 | /* Before SandyBridge, you could not use tiling or fence | |
2941 | * registers with snooped memory, so relinquish any fences | |
2942 | * currently pointing to our region in the aperture. | |
2943 | */ | |
2944 | if (INTEL_INFO(obj->base.dev)->gen < 6) { | |
2945 | ret = i915_gem_object_put_fence(obj); | |
2946 | if (ret) | |
2947 | return ret; | |
2948 | } | |
2949 | ||
74898d7e DV |
2950 | if (obj->has_global_gtt_mapping) |
2951 | i915_gem_gtt_bind_object(obj, cache_level); | |
7bddb01f DV |
2952 | if (obj->has_aliasing_ppgtt_mapping) |
2953 | i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt, | |
2954 | obj, cache_level); | |
e4ffd173 CW |
2955 | } |
2956 | ||
2957 | if (cache_level == I915_CACHE_NONE) { | |
2958 | u32 old_read_domains, old_write_domain; | |
2959 | ||
2960 | /* If we're coming from LLC cached, then we haven't | |
2961 | * actually been tracking whether the data is in the | |
2962 | * CPU cache or not, since we only allow one bit set | |
2963 | * in obj->write_domain and have been skipping the clflushes. | |
2964 | * Just set it to the CPU cache for now. | |
2965 | */ | |
2966 | WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU); | |
2967 | WARN_ON(obj->base.read_domains & ~I915_GEM_DOMAIN_CPU); | |
2968 | ||
2969 | old_read_domains = obj->base.read_domains; | |
2970 | old_write_domain = obj->base.write_domain; | |
2971 | ||
2972 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
2973 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
2974 | ||
2975 | trace_i915_gem_object_change_domain(obj, | |
2976 | old_read_domains, | |
2977 | old_write_domain); | |
2978 | } | |
2979 | ||
2980 | obj->cache_level = cache_level; | |
2981 | return 0; | |
2982 | } | |
2983 | ||
b9241ea3 | 2984 | /* |
2da3b9b9 CW |
2985 | * Prepare buffer for display plane (scanout, cursors, etc). |
2986 | * Can be called from an uninterruptible phase (modesetting) and allows | |
2987 | * any flushes to be pipelined (for pageflips). | |
b9241ea3 ZW |
2988 | */ |
2989 | int | |
2da3b9b9 CW |
2990 | i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj, |
2991 | u32 alignment, | |
919926ae | 2992 | struct intel_ring_buffer *pipelined) |
b9241ea3 | 2993 | { |
2da3b9b9 | 2994 | u32 old_read_domains, old_write_domain; |
b9241ea3 ZW |
2995 | int ret; |
2996 | ||
88241785 CW |
2997 | ret = i915_gem_object_flush_gpu_write_domain(obj); |
2998 | if (ret) | |
2999 | return ret; | |
3000 | ||
0be73284 | 3001 | if (pipelined != obj->ring) { |
2911a35b BW |
3002 | ret = i915_gem_object_sync(obj, pipelined); |
3003 | if (ret) | |
b9241ea3 ZW |
3004 | return ret; |
3005 | } | |
3006 | ||
a7ef0640 EA |
3007 | /* The display engine is not coherent with the LLC cache on gen6. As |
3008 | * a result, we make sure that the pinning that is about to occur is | |
3009 | * done with uncached PTEs. This is lowest common denominator for all | |
3010 | * chipsets. | |
3011 | * | |
3012 | * However for gen6+, we could do better by using the GFDT bit instead | |
3013 | * of uncaching, which would allow us to flush all the LLC-cached data | |
3014 | * with that bit in the PTE to main memory with just one PIPE_CONTROL. | |
3015 | */ | |
3016 | ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE); | |
3017 | if (ret) | |
3018 | return ret; | |
3019 | ||
2da3b9b9 CW |
3020 | /* As the user may map the buffer once pinned in the display plane |
3021 | * (e.g. libkms for the bootup splash), we have to ensure that we | |
3022 | * always use map_and_fenceable for all scanout buffers. | |
3023 | */ | |
3024 | ret = i915_gem_object_pin(obj, alignment, true); | |
3025 | if (ret) | |
3026 | return ret; | |
3027 | ||
b118c1e3 CW |
3028 | i915_gem_object_flush_cpu_write_domain(obj); |
3029 | ||
2da3b9b9 | 3030 | old_write_domain = obj->base.write_domain; |
05394f39 | 3031 | old_read_domains = obj->base.read_domains; |
2da3b9b9 CW |
3032 | |
3033 | /* It should now be out of any other write domains, and we can update | |
3034 | * the domain values for our changes. | |
3035 | */ | |
3036 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0); | |
05394f39 | 3037 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; |
b9241ea3 ZW |
3038 | |
3039 | trace_i915_gem_object_change_domain(obj, | |
3040 | old_read_domains, | |
2da3b9b9 | 3041 | old_write_domain); |
b9241ea3 ZW |
3042 | |
3043 | return 0; | |
3044 | } | |
3045 | ||
85345517 | 3046 | int |
a8198eea | 3047 | i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj) |
85345517 | 3048 | { |
88241785 CW |
3049 | int ret; |
3050 | ||
a8198eea | 3051 | if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0) |
85345517 CW |
3052 | return 0; |
3053 | ||
88241785 | 3054 | if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) { |
db53a302 | 3055 | ret = i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain); |
88241785 CW |
3056 | if (ret) |
3057 | return ret; | |
3058 | } | |
85345517 | 3059 | |
c501ae7f CW |
3060 | ret = i915_gem_object_wait_rendering(obj); |
3061 | if (ret) | |
3062 | return ret; | |
3063 | ||
a8198eea CW |
3064 | /* Ensure that we invalidate the GPU's caches and TLBs. */ |
3065 | obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS; | |
c501ae7f | 3066 | return 0; |
85345517 CW |
3067 | } |
3068 | ||
e47c68e9 EA |
3069 | /** |
3070 | * Moves a single object to the CPU read, and possibly write domain. | |
3071 | * | |
3072 | * This function returns when the move is complete, including waiting on | |
3073 | * flushes to occur. | |
3074 | */ | |
dabdfe02 | 3075 | int |
919926ae | 3076 | i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write) |
e47c68e9 | 3077 | { |
1c5d22f7 | 3078 | uint32_t old_write_domain, old_read_domains; |
e47c68e9 EA |
3079 | int ret; |
3080 | ||
8d7e3de1 CW |
3081 | if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) |
3082 | return 0; | |
3083 | ||
88241785 CW |
3084 | ret = i915_gem_object_flush_gpu_write_domain(obj); |
3085 | if (ret) | |
3086 | return ret; | |
3087 | ||
f8413190 CW |
3088 | if (write || obj->pending_gpu_write) { |
3089 | ret = i915_gem_object_wait_rendering(obj); | |
3090 | if (ret) | |
3091 | return ret; | |
3092 | } | |
2ef7eeaa | 3093 | |
e47c68e9 | 3094 | i915_gem_object_flush_gtt_write_domain(obj); |
2ef7eeaa | 3095 | |
05394f39 CW |
3096 | old_write_domain = obj->base.write_domain; |
3097 | old_read_domains = obj->base.read_domains; | |
1c5d22f7 | 3098 | |
e47c68e9 | 3099 | /* Flush the CPU cache if it's still invalid. */ |
05394f39 | 3100 | if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) { |
2ef7eeaa | 3101 | i915_gem_clflush_object(obj); |
2ef7eeaa | 3102 | |
05394f39 | 3103 | obj->base.read_domains |= I915_GEM_DOMAIN_CPU; |
2ef7eeaa EA |
3104 | } |
3105 | ||
3106 | /* It should now be out of any other write domains, and we can update | |
3107 | * the domain values for our changes. | |
3108 | */ | |
05394f39 | 3109 | BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0); |
e47c68e9 EA |
3110 | |
3111 | /* If we're writing through the CPU, then the GPU read domains will | |
3112 | * need to be invalidated at next use. | |
3113 | */ | |
3114 | if (write) { | |
05394f39 CW |
3115 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; |
3116 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
e47c68e9 | 3117 | } |
2ef7eeaa | 3118 | |
1c5d22f7 CW |
3119 | trace_i915_gem_object_change_domain(obj, |
3120 | old_read_domains, | |
3121 | old_write_domain); | |
3122 | ||
2ef7eeaa EA |
3123 | return 0; |
3124 | } | |
3125 | ||
673a394b EA |
3126 | /* Throttle our rendering by waiting until the ring has completed our requests |
3127 | * emitted over 20 msec ago. | |
3128 | * | |
b962442e EA |
3129 | * Note that if we were to use the current jiffies each time around the loop, |
3130 | * we wouldn't escape the function with any frames outstanding if the time to | |
3131 | * render a frame was over 20ms. | |
3132 | * | |
673a394b EA |
3133 | * This should get us reasonable parallelism between CPU and GPU but also |
3134 | * relatively low latency when blocking on a particular request to finish. | |
3135 | */ | |
40a5f0de | 3136 | static int |
f787a5f5 | 3137 | i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file) |
40a5f0de | 3138 | { |
f787a5f5 CW |
3139 | struct drm_i915_private *dev_priv = dev->dev_private; |
3140 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
b962442e | 3141 | unsigned long recent_enough = jiffies - msecs_to_jiffies(20); |
f787a5f5 CW |
3142 | struct drm_i915_gem_request *request; |
3143 | struct intel_ring_buffer *ring = NULL; | |
3144 | u32 seqno = 0; | |
3145 | int ret; | |
93533c29 | 3146 | |
e110e8d6 CW |
3147 | if (atomic_read(&dev_priv->mm.wedged)) |
3148 | return -EIO; | |
3149 | ||
1c25595f | 3150 | spin_lock(&file_priv->mm.lock); |
f787a5f5 | 3151 | list_for_each_entry(request, &file_priv->mm.request_list, client_list) { |
b962442e EA |
3152 | if (time_after_eq(request->emitted_jiffies, recent_enough)) |
3153 | break; | |
40a5f0de | 3154 | |
f787a5f5 CW |
3155 | ring = request->ring; |
3156 | seqno = request->seqno; | |
b962442e | 3157 | } |
1c25595f | 3158 | spin_unlock(&file_priv->mm.lock); |
40a5f0de | 3159 | |
f787a5f5 CW |
3160 | if (seqno == 0) |
3161 | return 0; | |
2bc43b5c | 3162 | |
f787a5f5 | 3163 | ret = 0; |
78501eac | 3164 | if (!i915_seqno_passed(ring->get_seqno(ring), seqno)) { |
f787a5f5 CW |
3165 | /* And wait for the seqno passing without holding any locks and |
3166 | * causing extra latency for others. This is safe as the irq | |
3167 | * generation is designed to be run atomically and so is | |
3168 | * lockless. | |
3169 | */ | |
b13c2b96 CW |
3170 | if (ring->irq_get(ring)) { |
3171 | ret = wait_event_interruptible(ring->irq_queue, | |
3172 | i915_seqno_passed(ring->get_seqno(ring), seqno) | |
3173 | || atomic_read(&dev_priv->mm.wedged)); | |
3174 | ring->irq_put(ring); | |
40a5f0de | 3175 | |
b13c2b96 CW |
3176 | if (ret == 0 && atomic_read(&dev_priv->mm.wedged)) |
3177 | ret = -EIO; | |
e959b5db EA |
3178 | } else if (wait_for_atomic(i915_seqno_passed(ring->get_seqno(ring), |
3179 | seqno) || | |
7ea29b13 EA |
3180 | atomic_read(&dev_priv->mm.wedged), 3000)) { |
3181 | ret = -EBUSY; | |
b13c2b96 | 3182 | } |
40a5f0de EA |
3183 | } |
3184 | ||
f787a5f5 CW |
3185 | if (ret == 0) |
3186 | queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0); | |
40a5f0de EA |
3187 | |
3188 | return ret; | |
3189 | } | |
3190 | ||
673a394b | 3191 | int |
05394f39 CW |
3192 | i915_gem_object_pin(struct drm_i915_gem_object *obj, |
3193 | uint32_t alignment, | |
75e9e915 | 3194 | bool map_and_fenceable) |
673a394b | 3195 | { |
05394f39 | 3196 | struct drm_device *dev = obj->base.dev; |
f13d3f73 | 3197 | struct drm_i915_private *dev_priv = dev->dev_private; |
673a394b EA |
3198 | int ret; |
3199 | ||
05394f39 | 3200 | BUG_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT); |
23bc5982 | 3201 | WARN_ON(i915_verify_lists(dev)); |
ac0c6b5a | 3202 | |
05394f39 CW |
3203 | if (obj->gtt_space != NULL) { |
3204 | if ((alignment && obj->gtt_offset & (alignment - 1)) || | |
3205 | (map_and_fenceable && !obj->map_and_fenceable)) { | |
3206 | WARN(obj->pin_count, | |
ae7d49d8 | 3207 | "bo is already pinned with incorrect alignment:" |
75e9e915 DV |
3208 | " offset=%x, req.alignment=%x, req.map_and_fenceable=%d," |
3209 | " obj->map_and_fenceable=%d\n", | |
05394f39 | 3210 | obj->gtt_offset, alignment, |
75e9e915 | 3211 | map_and_fenceable, |
05394f39 | 3212 | obj->map_and_fenceable); |
ac0c6b5a CW |
3213 | ret = i915_gem_object_unbind(obj); |
3214 | if (ret) | |
3215 | return ret; | |
3216 | } | |
3217 | } | |
3218 | ||
05394f39 | 3219 | if (obj->gtt_space == NULL) { |
a00b10c3 | 3220 | ret = i915_gem_object_bind_to_gtt(obj, alignment, |
75e9e915 | 3221 | map_and_fenceable); |
9731129c | 3222 | if (ret) |
673a394b | 3223 | return ret; |
22c344e9 | 3224 | } |
76446cac | 3225 | |
74898d7e DV |
3226 | if (!obj->has_global_gtt_mapping && map_and_fenceable) |
3227 | i915_gem_gtt_bind_object(obj, obj->cache_level); | |
3228 | ||
05394f39 | 3229 | if (obj->pin_count++ == 0) { |
05394f39 CW |
3230 | if (!obj->active) |
3231 | list_move_tail(&obj->mm_list, | |
f13d3f73 | 3232 | &dev_priv->mm.pinned_list); |
673a394b | 3233 | } |
6299f992 | 3234 | obj->pin_mappable |= map_and_fenceable; |
673a394b | 3235 | |
23bc5982 | 3236 | WARN_ON(i915_verify_lists(dev)); |
673a394b EA |
3237 | return 0; |
3238 | } | |
3239 | ||
3240 | void | |
05394f39 | 3241 | i915_gem_object_unpin(struct drm_i915_gem_object *obj) |
673a394b | 3242 | { |
05394f39 | 3243 | struct drm_device *dev = obj->base.dev; |
673a394b | 3244 | drm_i915_private_t *dev_priv = dev->dev_private; |
673a394b | 3245 | |
23bc5982 | 3246 | WARN_ON(i915_verify_lists(dev)); |
05394f39 CW |
3247 | BUG_ON(obj->pin_count == 0); |
3248 | BUG_ON(obj->gtt_space == NULL); | |
673a394b | 3249 | |
05394f39 CW |
3250 | if (--obj->pin_count == 0) { |
3251 | if (!obj->active) | |
3252 | list_move_tail(&obj->mm_list, | |
673a394b | 3253 | &dev_priv->mm.inactive_list); |
6299f992 | 3254 | obj->pin_mappable = false; |
673a394b | 3255 | } |
23bc5982 | 3256 | WARN_ON(i915_verify_lists(dev)); |
673a394b EA |
3257 | } |
3258 | ||
3259 | int | |
3260 | i915_gem_pin_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 3261 | struct drm_file *file) |
673a394b EA |
3262 | { |
3263 | struct drm_i915_gem_pin *args = data; | |
05394f39 | 3264 | struct drm_i915_gem_object *obj; |
673a394b EA |
3265 | int ret; |
3266 | ||
1d7cfea1 CW |
3267 | ret = i915_mutex_lock_interruptible(dev); |
3268 | if (ret) | |
3269 | return ret; | |
673a394b | 3270 | |
05394f39 | 3271 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 3272 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3273 | ret = -ENOENT; |
3274 | goto unlock; | |
673a394b | 3275 | } |
673a394b | 3276 | |
05394f39 | 3277 | if (obj->madv != I915_MADV_WILLNEED) { |
bb6baf76 | 3278 | DRM_ERROR("Attempting to pin a purgeable buffer\n"); |
1d7cfea1 CW |
3279 | ret = -EINVAL; |
3280 | goto out; | |
3ef94daa CW |
3281 | } |
3282 | ||
05394f39 | 3283 | if (obj->pin_filp != NULL && obj->pin_filp != file) { |
79e53945 JB |
3284 | DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n", |
3285 | args->handle); | |
1d7cfea1 CW |
3286 | ret = -EINVAL; |
3287 | goto out; | |
79e53945 JB |
3288 | } |
3289 | ||
05394f39 CW |
3290 | obj->user_pin_count++; |
3291 | obj->pin_filp = file; | |
3292 | if (obj->user_pin_count == 1) { | |
75e9e915 | 3293 | ret = i915_gem_object_pin(obj, args->alignment, true); |
1d7cfea1 CW |
3294 | if (ret) |
3295 | goto out; | |
673a394b EA |
3296 | } |
3297 | ||
3298 | /* XXX - flush the CPU caches for pinned objects | |
3299 | * as the X server doesn't manage domains yet | |
3300 | */ | |
e47c68e9 | 3301 | i915_gem_object_flush_cpu_write_domain(obj); |
05394f39 | 3302 | args->offset = obj->gtt_offset; |
1d7cfea1 | 3303 | out: |
05394f39 | 3304 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 3305 | unlock: |
673a394b | 3306 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 3307 | return ret; |
673a394b EA |
3308 | } |
3309 | ||
3310 | int | |
3311 | i915_gem_unpin_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 3312 | struct drm_file *file) |
673a394b EA |
3313 | { |
3314 | struct drm_i915_gem_pin *args = data; | |
05394f39 | 3315 | struct drm_i915_gem_object *obj; |
76c1dec1 | 3316 | int ret; |
673a394b | 3317 | |
1d7cfea1 CW |
3318 | ret = i915_mutex_lock_interruptible(dev); |
3319 | if (ret) | |
3320 | return ret; | |
673a394b | 3321 | |
05394f39 | 3322 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 3323 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3324 | ret = -ENOENT; |
3325 | goto unlock; | |
673a394b | 3326 | } |
76c1dec1 | 3327 | |
05394f39 | 3328 | if (obj->pin_filp != file) { |
79e53945 JB |
3329 | DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n", |
3330 | args->handle); | |
1d7cfea1 CW |
3331 | ret = -EINVAL; |
3332 | goto out; | |
79e53945 | 3333 | } |
05394f39 CW |
3334 | obj->user_pin_count--; |
3335 | if (obj->user_pin_count == 0) { | |
3336 | obj->pin_filp = NULL; | |
79e53945 JB |
3337 | i915_gem_object_unpin(obj); |
3338 | } | |
673a394b | 3339 | |
1d7cfea1 | 3340 | out: |
05394f39 | 3341 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 3342 | unlock: |
673a394b | 3343 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 3344 | return ret; |
673a394b EA |
3345 | } |
3346 | ||
3347 | int | |
3348 | i915_gem_busy_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 3349 | struct drm_file *file) |
673a394b EA |
3350 | { |
3351 | struct drm_i915_gem_busy *args = data; | |
05394f39 | 3352 | struct drm_i915_gem_object *obj; |
30dbf0c0 CW |
3353 | int ret; |
3354 | ||
76c1dec1 | 3355 | ret = i915_mutex_lock_interruptible(dev); |
1d7cfea1 | 3356 | if (ret) |
76c1dec1 | 3357 | return ret; |
673a394b | 3358 | |
05394f39 | 3359 | obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle)); |
c8725226 | 3360 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3361 | ret = -ENOENT; |
3362 | goto unlock; | |
673a394b | 3363 | } |
d1b851fc | 3364 | |
0be555b6 CW |
3365 | /* Count all active objects as busy, even if they are currently not used |
3366 | * by the gpu. Users of this interface expect objects to eventually | |
3367 | * become non-busy without any further actions, therefore emit any | |
3368 | * necessary flushes here. | |
c4de0a5d | 3369 | */ |
05394f39 | 3370 | args->busy = obj->active; |
0be555b6 CW |
3371 | if (args->busy) { |
3372 | /* Unconditionally flush objects, even when the gpu still uses this | |
3373 | * object. Userspace calling this function indicates that it wants to | |
3374 | * use this buffer rather sooner than later, so issuing the required | |
3375 | * flush earlier is beneficial. | |
3376 | */ | |
1a1c6976 | 3377 | if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) { |
db53a302 | 3378 | ret = i915_gem_flush_ring(obj->ring, |
88241785 | 3379 | 0, obj->base.write_domain); |
1a1c6976 CW |
3380 | } else if (obj->ring->outstanding_lazy_request == |
3381 | obj->last_rendering_seqno) { | |
3382 | struct drm_i915_gem_request *request; | |
3383 | ||
7a194876 CW |
3384 | /* This ring is not being cleared by active usage, |
3385 | * so emit a request to do so. | |
3386 | */ | |
1a1c6976 | 3387 | request = kzalloc(sizeof(*request), GFP_KERNEL); |
457eafce | 3388 | if (request) { |
0206e353 | 3389 | ret = i915_add_request(obj->ring, NULL, request); |
457eafce RM |
3390 | if (ret) |
3391 | kfree(request); | |
3392 | } else | |
7a194876 CW |
3393 | ret = -ENOMEM; |
3394 | } | |
0be555b6 CW |
3395 | |
3396 | /* Update the active list for the hardware's current position. | |
3397 | * Otherwise this only updates on a delayed timer or when irqs | |
3398 | * are actually unmasked, and our working set ends up being | |
3399 | * larger than required. | |
3400 | */ | |
db53a302 | 3401 | i915_gem_retire_requests_ring(obj->ring); |
0be555b6 | 3402 | |
05394f39 | 3403 | args->busy = obj->active; |
0be555b6 | 3404 | } |
673a394b | 3405 | |
05394f39 | 3406 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 3407 | unlock: |
673a394b | 3408 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 3409 | return ret; |
673a394b EA |
3410 | } |
3411 | ||
3412 | int | |
3413 | i915_gem_throttle_ioctl(struct drm_device *dev, void *data, | |
3414 | struct drm_file *file_priv) | |
3415 | { | |
0206e353 | 3416 | return i915_gem_ring_throttle(dev, file_priv); |
673a394b EA |
3417 | } |
3418 | ||
3ef94daa CW |
3419 | int |
3420 | i915_gem_madvise_ioctl(struct drm_device *dev, void *data, | |
3421 | struct drm_file *file_priv) | |
3422 | { | |
3423 | struct drm_i915_gem_madvise *args = data; | |
05394f39 | 3424 | struct drm_i915_gem_object *obj; |
76c1dec1 | 3425 | int ret; |
3ef94daa CW |
3426 | |
3427 | switch (args->madv) { | |
3428 | case I915_MADV_DONTNEED: | |
3429 | case I915_MADV_WILLNEED: | |
3430 | break; | |
3431 | default: | |
3432 | return -EINVAL; | |
3433 | } | |
3434 | ||
1d7cfea1 CW |
3435 | ret = i915_mutex_lock_interruptible(dev); |
3436 | if (ret) | |
3437 | return ret; | |
3438 | ||
05394f39 | 3439 | obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle)); |
c8725226 | 3440 | if (&obj->base == NULL) { |
1d7cfea1 CW |
3441 | ret = -ENOENT; |
3442 | goto unlock; | |
3ef94daa | 3443 | } |
3ef94daa | 3444 | |
05394f39 | 3445 | if (obj->pin_count) { |
1d7cfea1 CW |
3446 | ret = -EINVAL; |
3447 | goto out; | |
3ef94daa CW |
3448 | } |
3449 | ||
05394f39 CW |
3450 | if (obj->madv != __I915_MADV_PURGED) |
3451 | obj->madv = args->madv; | |
3ef94daa | 3452 | |
2d7ef395 | 3453 | /* if the object is no longer bound, discard its backing storage */ |
05394f39 CW |
3454 | if (i915_gem_object_is_purgeable(obj) && |
3455 | obj->gtt_space == NULL) | |
2d7ef395 CW |
3456 | i915_gem_object_truncate(obj); |
3457 | ||
05394f39 | 3458 | args->retained = obj->madv != __I915_MADV_PURGED; |
bb6baf76 | 3459 | |
1d7cfea1 | 3460 | out: |
05394f39 | 3461 | drm_gem_object_unreference(&obj->base); |
1d7cfea1 | 3462 | unlock: |
3ef94daa | 3463 | mutex_unlock(&dev->struct_mutex); |
1d7cfea1 | 3464 | return ret; |
3ef94daa CW |
3465 | } |
3466 | ||
05394f39 CW |
3467 | struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev, |
3468 | size_t size) | |
ac52bc56 | 3469 | { |
73aa808f | 3470 | struct drm_i915_private *dev_priv = dev->dev_private; |
c397b908 | 3471 | struct drm_i915_gem_object *obj; |
5949eac4 | 3472 | struct address_space *mapping; |
ac52bc56 | 3473 | |
c397b908 DV |
3474 | obj = kzalloc(sizeof(*obj), GFP_KERNEL); |
3475 | if (obj == NULL) | |
3476 | return NULL; | |
673a394b | 3477 | |
c397b908 DV |
3478 | if (drm_gem_object_init(dev, &obj->base, size) != 0) { |
3479 | kfree(obj); | |
3480 | return NULL; | |
3481 | } | |
673a394b | 3482 | |
5949eac4 HD |
3483 | mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping; |
3484 | mapping_set_gfp_mask(mapping, GFP_HIGHUSER | __GFP_RECLAIMABLE); | |
3485 | ||
73aa808f CW |
3486 | i915_gem_info_add_obj(dev_priv, size); |
3487 | ||
c397b908 DV |
3488 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
3489 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
673a394b | 3490 | |
3d29b842 ED |
3491 | if (HAS_LLC(dev)) { |
3492 | /* On some devices, we can have the GPU use the LLC (the CPU | |
a1871112 EA |
3493 | * cache) for about a 10% performance improvement |
3494 | * compared to uncached. Graphics requests other than | |
3495 | * display scanout are coherent with the CPU in | |
3496 | * accessing this cache. This means in this mode we | |
3497 | * don't need to clflush on the CPU side, and on the | |
3498 | * GPU side we only need to flush internal caches to | |
3499 | * get data visible to the CPU. | |
3500 | * | |
3501 | * However, we maintain the display planes as UC, and so | |
3502 | * need to rebind when first used as such. | |
3503 | */ | |
3504 | obj->cache_level = I915_CACHE_LLC; | |
3505 | } else | |
3506 | obj->cache_level = I915_CACHE_NONE; | |
3507 | ||
62b8b215 | 3508 | obj->base.driver_private = NULL; |
c397b908 | 3509 | obj->fence_reg = I915_FENCE_REG_NONE; |
69dc4987 | 3510 | INIT_LIST_HEAD(&obj->mm_list); |
93a37f20 | 3511 | INIT_LIST_HEAD(&obj->gtt_list); |
69dc4987 | 3512 | INIT_LIST_HEAD(&obj->ring_list); |
432e58ed | 3513 | INIT_LIST_HEAD(&obj->exec_list); |
c397b908 | 3514 | INIT_LIST_HEAD(&obj->gpu_write_list); |
c397b908 | 3515 | obj->madv = I915_MADV_WILLNEED; |
75e9e915 DV |
3516 | /* Avoid an unnecessary call to unbind on the first bind. */ |
3517 | obj->map_and_fenceable = true; | |
de151cf6 | 3518 | |
05394f39 | 3519 | return obj; |
c397b908 DV |
3520 | } |
3521 | ||
3522 | int i915_gem_init_object(struct drm_gem_object *obj) | |
3523 | { | |
3524 | BUG(); | |
de151cf6 | 3525 | |
673a394b EA |
3526 | return 0; |
3527 | } | |
3528 | ||
05394f39 | 3529 | static void i915_gem_free_object_tail(struct drm_i915_gem_object *obj) |
673a394b | 3530 | { |
05394f39 | 3531 | struct drm_device *dev = obj->base.dev; |
be72615b | 3532 | drm_i915_private_t *dev_priv = dev->dev_private; |
be72615b | 3533 | int ret; |
673a394b | 3534 | |
be72615b CW |
3535 | ret = i915_gem_object_unbind(obj); |
3536 | if (ret == -ERESTARTSYS) { | |
05394f39 | 3537 | list_move(&obj->mm_list, |
be72615b CW |
3538 | &dev_priv->mm.deferred_free_list); |
3539 | return; | |
3540 | } | |
673a394b | 3541 | |
26e12f89 CW |
3542 | trace_i915_gem_object_destroy(obj); |
3543 | ||
05394f39 | 3544 | if (obj->base.map_list.map) |
b464e9a2 | 3545 | drm_gem_free_mmap_offset(&obj->base); |
de151cf6 | 3546 | |
05394f39 CW |
3547 | drm_gem_object_release(&obj->base); |
3548 | i915_gem_info_remove_obj(dev_priv, obj->base.size); | |
c397b908 | 3549 | |
05394f39 CW |
3550 | kfree(obj->bit_17); |
3551 | kfree(obj); | |
673a394b EA |
3552 | } |
3553 | ||
05394f39 | 3554 | void i915_gem_free_object(struct drm_gem_object *gem_obj) |
be72615b | 3555 | { |
05394f39 CW |
3556 | struct drm_i915_gem_object *obj = to_intel_bo(gem_obj); |
3557 | struct drm_device *dev = obj->base.dev; | |
be72615b | 3558 | |
05394f39 | 3559 | while (obj->pin_count > 0) |
be72615b CW |
3560 | i915_gem_object_unpin(obj); |
3561 | ||
05394f39 | 3562 | if (obj->phys_obj) |
be72615b CW |
3563 | i915_gem_detach_phys_object(dev, obj); |
3564 | ||
3565 | i915_gem_free_object_tail(obj); | |
3566 | } | |
3567 | ||
29105ccc CW |
3568 | int |
3569 | i915_gem_idle(struct drm_device *dev) | |
3570 | { | |
3571 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3572 | int ret; | |
28dfe52a | 3573 | |
29105ccc | 3574 | mutex_lock(&dev->struct_mutex); |
1c5d22f7 | 3575 | |
87acb0a5 | 3576 | if (dev_priv->mm.suspended) { |
29105ccc CW |
3577 | mutex_unlock(&dev->struct_mutex); |
3578 | return 0; | |
28dfe52a EA |
3579 | } |
3580 | ||
b93f9cf1 | 3581 | ret = i915_gpu_idle(dev, true); |
6dbe2772 KP |
3582 | if (ret) { |
3583 | mutex_unlock(&dev->struct_mutex); | |
673a394b | 3584 | return ret; |
6dbe2772 | 3585 | } |
673a394b | 3586 | |
29105ccc CW |
3587 | /* Under UMS, be paranoid and evict. */ |
3588 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) { | |
5eac3ab4 | 3589 | ret = i915_gem_evict_inactive(dev, false); |
29105ccc CW |
3590 | if (ret) { |
3591 | mutex_unlock(&dev->struct_mutex); | |
3592 | return ret; | |
3593 | } | |
3594 | } | |
3595 | ||
312817a3 CW |
3596 | i915_gem_reset_fences(dev); |
3597 | ||
29105ccc CW |
3598 | /* Hack! Don't let anybody do execbuf while we don't control the chip. |
3599 | * We need to replace this with a semaphore, or something. | |
3600 | * And not confound mm.suspended! | |
3601 | */ | |
3602 | dev_priv->mm.suspended = 1; | |
bc0c7f14 | 3603 | del_timer_sync(&dev_priv->hangcheck_timer); |
29105ccc CW |
3604 | |
3605 | i915_kernel_lost_context(dev); | |
6dbe2772 | 3606 | i915_gem_cleanup_ringbuffer(dev); |
29105ccc | 3607 | |
6dbe2772 KP |
3608 | mutex_unlock(&dev->struct_mutex); |
3609 | ||
29105ccc CW |
3610 | /* Cancel the retire work handler, which should be idle now. */ |
3611 | cancel_delayed_work_sync(&dev_priv->mm.retire_work); | |
3612 | ||
673a394b EA |
3613 | return 0; |
3614 | } | |
3615 | ||
f691e2f4 DV |
3616 | void i915_gem_init_swizzling(struct drm_device *dev) |
3617 | { | |
3618 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3619 | ||
11782b02 | 3620 | if (INTEL_INFO(dev)->gen < 5 || |
f691e2f4 DV |
3621 | dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE) |
3622 | return; | |
3623 | ||
3624 | I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) | | |
3625 | DISP_TILE_SURFACE_SWIZZLING); | |
3626 | ||
11782b02 DV |
3627 | if (IS_GEN5(dev)) |
3628 | return; | |
3629 | ||
f691e2f4 DV |
3630 | I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL); |
3631 | if (IS_GEN6(dev)) | |
3632 | I915_WRITE(ARB_MODE, ARB_MODE_ENABLE(ARB_MODE_SWIZZLE_SNB)); | |
3633 | else | |
3634 | I915_WRITE(ARB_MODE, ARB_MODE_ENABLE(ARB_MODE_SWIZZLE_IVB)); | |
3635 | } | |
e21af88d DV |
3636 | |
3637 | void i915_gem_init_ppgtt(struct drm_device *dev) | |
3638 | { | |
3639 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3640 | uint32_t pd_offset; | |
3641 | struct intel_ring_buffer *ring; | |
55a254ac DV |
3642 | struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt; |
3643 | uint32_t __iomem *pd_addr; | |
3644 | uint32_t pd_entry; | |
e21af88d DV |
3645 | int i; |
3646 | ||
3647 | if (!dev_priv->mm.aliasing_ppgtt) | |
3648 | return; | |
3649 | ||
55a254ac DV |
3650 | |
3651 | pd_addr = dev_priv->mm.gtt->gtt + ppgtt->pd_offset/sizeof(uint32_t); | |
3652 | for (i = 0; i < ppgtt->num_pd_entries; i++) { | |
3653 | dma_addr_t pt_addr; | |
3654 | ||
3655 | if (dev_priv->mm.gtt->needs_dmar) | |
3656 | pt_addr = ppgtt->pt_dma_addr[i]; | |
3657 | else | |
3658 | pt_addr = page_to_phys(ppgtt->pt_pages[i]); | |
3659 | ||
3660 | pd_entry = GEN6_PDE_ADDR_ENCODE(pt_addr); | |
3661 | pd_entry |= GEN6_PDE_VALID; | |
3662 | ||
3663 | writel(pd_entry, pd_addr + i); | |
3664 | } | |
3665 | readl(pd_addr); | |
3666 | ||
3667 | pd_offset = ppgtt->pd_offset; | |
e21af88d DV |
3668 | pd_offset /= 64; /* in cachelines, */ |
3669 | pd_offset <<= 16; | |
3670 | ||
3671 | if (INTEL_INFO(dev)->gen == 6) { | |
3672 | uint32_t ecochk = I915_READ(GAM_ECOCHK); | |
3673 | I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | | |
3674 | ECOCHK_PPGTT_CACHE64B); | |
3675 | I915_WRITE(GFX_MODE, GFX_MODE_ENABLE(GFX_PPGTT_ENABLE)); | |
3676 | } else if (INTEL_INFO(dev)->gen >= 7) { | |
3677 | I915_WRITE(GAM_ECOCHK, ECOCHK_PPGTT_CACHE64B); | |
3678 | /* GFX_MODE is per-ring on gen7+ */ | |
3679 | } | |
3680 | ||
3681 | for (i = 0; i < I915_NUM_RINGS; i++) { | |
3682 | ring = &dev_priv->ring[i]; | |
3683 | ||
3684 | if (INTEL_INFO(dev)->gen >= 7) | |
3685 | I915_WRITE(RING_MODE_GEN7(ring), | |
3686 | GFX_MODE_ENABLE(GFX_PPGTT_ENABLE)); | |
3687 | ||
3688 | I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G); | |
3689 | I915_WRITE(RING_PP_DIR_BASE(ring), pd_offset); | |
3690 | } | |
3691 | } | |
3692 | ||
8187a2b7 | 3693 | int |
f691e2f4 | 3694 | i915_gem_init_hw(struct drm_device *dev) |
8187a2b7 ZN |
3695 | { |
3696 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3697 | int ret; | |
68f95ba9 | 3698 | |
f691e2f4 DV |
3699 | i915_gem_init_swizzling(dev); |
3700 | ||
5c1143bb | 3701 | ret = intel_init_render_ring_buffer(dev); |
68f95ba9 | 3702 | if (ret) |
b6913e4b | 3703 | return ret; |
68f95ba9 CW |
3704 | |
3705 | if (HAS_BSD(dev)) { | |
5c1143bb | 3706 | ret = intel_init_bsd_ring_buffer(dev); |
68f95ba9 CW |
3707 | if (ret) |
3708 | goto cleanup_render_ring; | |
d1b851fc | 3709 | } |
68f95ba9 | 3710 | |
549f7365 CW |
3711 | if (HAS_BLT(dev)) { |
3712 | ret = intel_init_blt_ring_buffer(dev); | |
3713 | if (ret) | |
3714 | goto cleanup_bsd_ring; | |
3715 | } | |
3716 | ||
6f392d54 CW |
3717 | dev_priv->next_seqno = 1; |
3718 | ||
e21af88d DV |
3719 | i915_gem_init_ppgtt(dev); |
3720 | ||
68f95ba9 CW |
3721 | return 0; |
3722 | ||
549f7365 | 3723 | cleanup_bsd_ring: |
1ec14ad3 | 3724 | intel_cleanup_ring_buffer(&dev_priv->ring[VCS]); |
68f95ba9 | 3725 | cleanup_render_ring: |
1ec14ad3 | 3726 | intel_cleanup_ring_buffer(&dev_priv->ring[RCS]); |
8187a2b7 ZN |
3727 | return ret; |
3728 | } | |
3729 | ||
3730 | void | |
3731 | i915_gem_cleanup_ringbuffer(struct drm_device *dev) | |
3732 | { | |
3733 | drm_i915_private_t *dev_priv = dev->dev_private; | |
1ec14ad3 | 3734 | int i; |
8187a2b7 | 3735 | |
1ec14ad3 CW |
3736 | for (i = 0; i < I915_NUM_RINGS; i++) |
3737 | intel_cleanup_ring_buffer(&dev_priv->ring[i]); | |
8187a2b7 ZN |
3738 | } |
3739 | ||
673a394b EA |
3740 | int |
3741 | i915_gem_entervt_ioctl(struct drm_device *dev, void *data, | |
3742 | struct drm_file *file_priv) | |
3743 | { | |
3744 | drm_i915_private_t *dev_priv = dev->dev_private; | |
1ec14ad3 | 3745 | int ret, i; |
673a394b | 3746 | |
79e53945 JB |
3747 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
3748 | return 0; | |
3749 | ||
ba1234d1 | 3750 | if (atomic_read(&dev_priv->mm.wedged)) { |
673a394b | 3751 | DRM_ERROR("Reenabling wedged hardware, good luck\n"); |
ba1234d1 | 3752 | atomic_set(&dev_priv->mm.wedged, 0); |
673a394b EA |
3753 | } |
3754 | ||
673a394b | 3755 | mutex_lock(&dev->struct_mutex); |
9bb2d6f9 EA |
3756 | dev_priv->mm.suspended = 0; |
3757 | ||
f691e2f4 | 3758 | ret = i915_gem_init_hw(dev); |
d816f6ac WF |
3759 | if (ret != 0) { |
3760 | mutex_unlock(&dev->struct_mutex); | |
9bb2d6f9 | 3761 | return ret; |
d816f6ac | 3762 | } |
9bb2d6f9 | 3763 | |
69dc4987 | 3764 | BUG_ON(!list_empty(&dev_priv->mm.active_list)); |
673a394b EA |
3765 | BUG_ON(!list_empty(&dev_priv->mm.flushing_list)); |
3766 | BUG_ON(!list_empty(&dev_priv->mm.inactive_list)); | |
1ec14ad3 CW |
3767 | for (i = 0; i < I915_NUM_RINGS; i++) { |
3768 | BUG_ON(!list_empty(&dev_priv->ring[i].active_list)); | |
3769 | BUG_ON(!list_empty(&dev_priv->ring[i].request_list)); | |
3770 | } | |
673a394b | 3771 | mutex_unlock(&dev->struct_mutex); |
dbb19d30 | 3772 | |
5f35308b CW |
3773 | ret = drm_irq_install(dev); |
3774 | if (ret) | |
3775 | goto cleanup_ringbuffer; | |
dbb19d30 | 3776 | |
673a394b | 3777 | return 0; |
5f35308b CW |
3778 | |
3779 | cleanup_ringbuffer: | |
3780 | mutex_lock(&dev->struct_mutex); | |
3781 | i915_gem_cleanup_ringbuffer(dev); | |
3782 | dev_priv->mm.suspended = 1; | |
3783 | mutex_unlock(&dev->struct_mutex); | |
3784 | ||
3785 | return ret; | |
673a394b EA |
3786 | } |
3787 | ||
3788 | int | |
3789 | i915_gem_leavevt_ioctl(struct drm_device *dev, void *data, | |
3790 | struct drm_file *file_priv) | |
3791 | { | |
79e53945 JB |
3792 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
3793 | return 0; | |
3794 | ||
dbb19d30 | 3795 | drm_irq_uninstall(dev); |
e6890f6f | 3796 | return i915_gem_idle(dev); |
673a394b EA |
3797 | } |
3798 | ||
3799 | void | |
3800 | i915_gem_lastclose(struct drm_device *dev) | |
3801 | { | |
3802 | int ret; | |
673a394b | 3803 | |
e806b495 EA |
3804 | if (drm_core_check_feature(dev, DRIVER_MODESET)) |
3805 | return; | |
3806 | ||
6dbe2772 KP |
3807 | ret = i915_gem_idle(dev); |
3808 | if (ret) | |
3809 | DRM_ERROR("failed to idle hardware: %d\n", ret); | |
673a394b EA |
3810 | } |
3811 | ||
64193406 CW |
3812 | static void |
3813 | init_ring_lists(struct intel_ring_buffer *ring) | |
3814 | { | |
3815 | INIT_LIST_HEAD(&ring->active_list); | |
3816 | INIT_LIST_HEAD(&ring->request_list); | |
3817 | INIT_LIST_HEAD(&ring->gpu_write_list); | |
3818 | } | |
3819 | ||
673a394b EA |
3820 | void |
3821 | i915_gem_load(struct drm_device *dev) | |
3822 | { | |
b5aa8a0f | 3823 | int i; |
673a394b EA |
3824 | drm_i915_private_t *dev_priv = dev->dev_private; |
3825 | ||
69dc4987 | 3826 | INIT_LIST_HEAD(&dev_priv->mm.active_list); |
673a394b EA |
3827 | INIT_LIST_HEAD(&dev_priv->mm.flushing_list); |
3828 | INIT_LIST_HEAD(&dev_priv->mm.inactive_list); | |
f13d3f73 | 3829 | INIT_LIST_HEAD(&dev_priv->mm.pinned_list); |
a09ba7fa | 3830 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); |
be72615b | 3831 | INIT_LIST_HEAD(&dev_priv->mm.deferred_free_list); |
93a37f20 | 3832 | INIT_LIST_HEAD(&dev_priv->mm.gtt_list); |
1ec14ad3 CW |
3833 | for (i = 0; i < I915_NUM_RINGS; i++) |
3834 | init_ring_lists(&dev_priv->ring[i]); | |
4b9de737 | 3835 | for (i = 0; i < I915_MAX_NUM_FENCES; i++) |
007cc8ac | 3836 | INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list); |
673a394b EA |
3837 | INIT_DELAYED_WORK(&dev_priv->mm.retire_work, |
3838 | i915_gem_retire_work_handler); | |
30dbf0c0 | 3839 | init_completion(&dev_priv->error_completion); |
31169714 | 3840 | |
94400120 DA |
3841 | /* On GEN3 we really need to make sure the ARB C3 LP bit is set */ |
3842 | if (IS_GEN3(dev)) { | |
3843 | u32 tmp = I915_READ(MI_ARB_STATE); | |
3844 | if (!(tmp & MI_ARB_C3_LP_WRITE_ENABLE)) { | |
3845 | /* arb state is a masked write, so set bit + bit in mask */ | |
3846 | tmp = MI_ARB_C3_LP_WRITE_ENABLE | (MI_ARB_C3_LP_WRITE_ENABLE << MI_ARB_MASK_SHIFT); | |
3847 | I915_WRITE(MI_ARB_STATE, tmp); | |
3848 | } | |
3849 | } | |
3850 | ||
72bfa19c CW |
3851 | dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL; |
3852 | ||
de151cf6 | 3853 | /* Old X drivers will take 0-2 for front, back, depth buffers */ |
b397c836 EA |
3854 | if (!drm_core_check_feature(dev, DRIVER_MODESET)) |
3855 | dev_priv->fence_reg_start = 3; | |
de151cf6 | 3856 | |
a6c45cf0 | 3857 | if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev)) |
de151cf6 JB |
3858 | dev_priv->num_fence_regs = 16; |
3859 | else | |
3860 | dev_priv->num_fence_regs = 8; | |
3861 | ||
b5aa8a0f | 3862 | /* Initialize fence registers to zero */ |
10ed13e4 EA |
3863 | for (i = 0; i < dev_priv->num_fence_regs; i++) { |
3864 | i915_gem_clear_fence_reg(dev, &dev_priv->fence_regs[i]); | |
b5aa8a0f | 3865 | } |
10ed13e4 | 3866 | |
673a394b | 3867 | i915_gem_detect_bit_6_swizzle(dev); |
6b95a207 | 3868 | init_waitqueue_head(&dev_priv->pending_flip_queue); |
17250b71 | 3869 | |
ce453d81 CW |
3870 | dev_priv->mm.interruptible = true; |
3871 | ||
17250b71 CW |
3872 | dev_priv->mm.inactive_shrinker.shrink = i915_gem_inactive_shrink; |
3873 | dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS; | |
3874 | register_shrinker(&dev_priv->mm.inactive_shrinker); | |
673a394b | 3875 | } |
71acb5eb DA |
3876 | |
3877 | /* | |
3878 | * Create a physically contiguous memory object for this object | |
3879 | * e.g. for cursor + overlay regs | |
3880 | */ | |
995b6762 CW |
3881 | static int i915_gem_init_phys_object(struct drm_device *dev, |
3882 | int id, int size, int align) | |
71acb5eb DA |
3883 | { |
3884 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3885 | struct drm_i915_gem_phys_object *phys_obj; | |
3886 | int ret; | |
3887 | ||
3888 | if (dev_priv->mm.phys_objs[id - 1] || !size) | |
3889 | return 0; | |
3890 | ||
9a298b2a | 3891 | phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL); |
71acb5eb DA |
3892 | if (!phys_obj) |
3893 | return -ENOMEM; | |
3894 | ||
3895 | phys_obj->id = id; | |
3896 | ||
6eeefaf3 | 3897 | phys_obj->handle = drm_pci_alloc(dev, size, align); |
71acb5eb DA |
3898 | if (!phys_obj->handle) { |
3899 | ret = -ENOMEM; | |
3900 | goto kfree_obj; | |
3901 | } | |
3902 | #ifdef CONFIG_X86 | |
3903 | set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE); | |
3904 | #endif | |
3905 | ||
3906 | dev_priv->mm.phys_objs[id - 1] = phys_obj; | |
3907 | ||
3908 | return 0; | |
3909 | kfree_obj: | |
9a298b2a | 3910 | kfree(phys_obj); |
71acb5eb DA |
3911 | return ret; |
3912 | } | |
3913 | ||
995b6762 | 3914 | static void i915_gem_free_phys_object(struct drm_device *dev, int id) |
71acb5eb DA |
3915 | { |
3916 | drm_i915_private_t *dev_priv = dev->dev_private; | |
3917 | struct drm_i915_gem_phys_object *phys_obj; | |
3918 | ||
3919 | if (!dev_priv->mm.phys_objs[id - 1]) | |
3920 | return; | |
3921 | ||
3922 | phys_obj = dev_priv->mm.phys_objs[id - 1]; | |
3923 | if (phys_obj->cur_obj) { | |
3924 | i915_gem_detach_phys_object(dev, phys_obj->cur_obj); | |
3925 | } | |
3926 | ||
3927 | #ifdef CONFIG_X86 | |
3928 | set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE); | |
3929 | #endif | |
3930 | drm_pci_free(dev, phys_obj->handle); | |
3931 | kfree(phys_obj); | |
3932 | dev_priv->mm.phys_objs[id - 1] = NULL; | |
3933 | } | |
3934 | ||
3935 | void i915_gem_free_all_phys_object(struct drm_device *dev) | |
3936 | { | |
3937 | int i; | |
3938 | ||
260883c8 | 3939 | for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++) |
71acb5eb DA |
3940 | i915_gem_free_phys_object(dev, i); |
3941 | } | |
3942 | ||
3943 | void i915_gem_detach_phys_object(struct drm_device *dev, | |
05394f39 | 3944 | struct drm_i915_gem_object *obj) |
71acb5eb | 3945 | { |
05394f39 | 3946 | struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping; |
e5281ccd | 3947 | char *vaddr; |
71acb5eb | 3948 | int i; |
71acb5eb DA |
3949 | int page_count; |
3950 | ||
05394f39 | 3951 | if (!obj->phys_obj) |
71acb5eb | 3952 | return; |
05394f39 | 3953 | vaddr = obj->phys_obj->handle->vaddr; |
71acb5eb | 3954 | |
05394f39 | 3955 | page_count = obj->base.size / PAGE_SIZE; |
71acb5eb | 3956 | for (i = 0; i < page_count; i++) { |
5949eac4 | 3957 | struct page *page = shmem_read_mapping_page(mapping, i); |
e5281ccd CW |
3958 | if (!IS_ERR(page)) { |
3959 | char *dst = kmap_atomic(page); | |
3960 | memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE); | |
3961 | kunmap_atomic(dst); | |
3962 | ||
3963 | drm_clflush_pages(&page, 1); | |
3964 | ||
3965 | set_page_dirty(page); | |
3966 | mark_page_accessed(page); | |
3967 | page_cache_release(page); | |
3968 | } | |
71acb5eb | 3969 | } |
40ce6575 | 3970 | intel_gtt_chipset_flush(); |
d78b47b9 | 3971 | |
05394f39 CW |
3972 | obj->phys_obj->cur_obj = NULL; |
3973 | obj->phys_obj = NULL; | |
71acb5eb DA |
3974 | } |
3975 | ||
3976 | int | |
3977 | i915_gem_attach_phys_object(struct drm_device *dev, | |
05394f39 | 3978 | struct drm_i915_gem_object *obj, |
6eeefaf3 CW |
3979 | int id, |
3980 | int align) | |
71acb5eb | 3981 | { |
05394f39 | 3982 | struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping; |
71acb5eb | 3983 | drm_i915_private_t *dev_priv = dev->dev_private; |
71acb5eb DA |
3984 | int ret = 0; |
3985 | int page_count; | |
3986 | int i; | |
3987 | ||
3988 | if (id > I915_MAX_PHYS_OBJECT) | |
3989 | return -EINVAL; | |
3990 | ||
05394f39 CW |
3991 | if (obj->phys_obj) { |
3992 | if (obj->phys_obj->id == id) | |
71acb5eb DA |
3993 | return 0; |
3994 | i915_gem_detach_phys_object(dev, obj); | |
3995 | } | |
3996 | ||
71acb5eb DA |
3997 | /* create a new object */ |
3998 | if (!dev_priv->mm.phys_objs[id - 1]) { | |
3999 | ret = i915_gem_init_phys_object(dev, id, | |
05394f39 | 4000 | obj->base.size, align); |
71acb5eb | 4001 | if (ret) { |
05394f39 CW |
4002 | DRM_ERROR("failed to init phys object %d size: %zu\n", |
4003 | id, obj->base.size); | |
e5281ccd | 4004 | return ret; |
71acb5eb DA |
4005 | } |
4006 | } | |
4007 | ||
4008 | /* bind to the object */ | |
05394f39 CW |
4009 | obj->phys_obj = dev_priv->mm.phys_objs[id - 1]; |
4010 | obj->phys_obj->cur_obj = obj; | |
71acb5eb | 4011 | |
05394f39 | 4012 | page_count = obj->base.size / PAGE_SIZE; |
71acb5eb DA |
4013 | |
4014 | for (i = 0; i < page_count; i++) { | |
e5281ccd CW |
4015 | struct page *page; |
4016 | char *dst, *src; | |
4017 | ||
5949eac4 | 4018 | page = shmem_read_mapping_page(mapping, i); |
e5281ccd CW |
4019 | if (IS_ERR(page)) |
4020 | return PTR_ERR(page); | |
71acb5eb | 4021 | |
ff75b9bc | 4022 | src = kmap_atomic(page); |
05394f39 | 4023 | dst = obj->phys_obj->handle->vaddr + (i * PAGE_SIZE); |
71acb5eb | 4024 | memcpy(dst, src, PAGE_SIZE); |
3e4d3af5 | 4025 | kunmap_atomic(src); |
71acb5eb | 4026 | |
e5281ccd CW |
4027 | mark_page_accessed(page); |
4028 | page_cache_release(page); | |
4029 | } | |
d78b47b9 | 4030 | |
71acb5eb | 4031 | return 0; |
71acb5eb DA |
4032 | } |
4033 | ||
4034 | static int | |
05394f39 CW |
4035 | i915_gem_phys_pwrite(struct drm_device *dev, |
4036 | struct drm_i915_gem_object *obj, | |
71acb5eb DA |
4037 | struct drm_i915_gem_pwrite *args, |
4038 | struct drm_file *file_priv) | |
4039 | { | |
05394f39 | 4040 | void *vaddr = obj->phys_obj->handle->vaddr + args->offset; |
b47b30cc | 4041 | char __user *user_data = (char __user *) (uintptr_t) args->data_ptr; |
71acb5eb | 4042 | |
b47b30cc CW |
4043 | if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) { |
4044 | unsigned long unwritten; | |
4045 | ||
4046 | /* The physical object once assigned is fixed for the lifetime | |
4047 | * of the obj, so we can safely drop the lock and continue | |
4048 | * to access vaddr. | |
4049 | */ | |
4050 | mutex_unlock(&dev->struct_mutex); | |
4051 | unwritten = copy_from_user(vaddr, user_data, args->size); | |
4052 | mutex_lock(&dev->struct_mutex); | |
4053 | if (unwritten) | |
4054 | return -EFAULT; | |
4055 | } | |
71acb5eb | 4056 | |
40ce6575 | 4057 | intel_gtt_chipset_flush(); |
71acb5eb DA |
4058 | return 0; |
4059 | } | |
b962442e | 4060 | |
f787a5f5 | 4061 | void i915_gem_release(struct drm_device *dev, struct drm_file *file) |
b962442e | 4062 | { |
f787a5f5 | 4063 | struct drm_i915_file_private *file_priv = file->driver_priv; |
b962442e EA |
4064 | |
4065 | /* Clean up our request list when the client is going away, so that | |
4066 | * later retire_requests won't dereference our soon-to-be-gone | |
4067 | * file_priv. | |
4068 | */ | |
1c25595f | 4069 | spin_lock(&file_priv->mm.lock); |
f787a5f5 CW |
4070 | while (!list_empty(&file_priv->mm.request_list)) { |
4071 | struct drm_i915_gem_request *request; | |
4072 | ||
4073 | request = list_first_entry(&file_priv->mm.request_list, | |
4074 | struct drm_i915_gem_request, | |
4075 | client_list); | |
4076 | list_del(&request->client_list); | |
4077 | request->file_priv = NULL; | |
4078 | } | |
1c25595f | 4079 | spin_unlock(&file_priv->mm.lock); |
b962442e | 4080 | } |
31169714 | 4081 | |
1637ef41 CW |
4082 | static int |
4083 | i915_gpu_is_active(struct drm_device *dev) | |
4084 | { | |
4085 | drm_i915_private_t *dev_priv = dev->dev_private; | |
4086 | int lists_empty; | |
4087 | ||
1637ef41 | 4088 | lists_empty = list_empty(&dev_priv->mm.flushing_list) && |
17250b71 | 4089 | list_empty(&dev_priv->mm.active_list); |
1637ef41 CW |
4090 | |
4091 | return !lists_empty; | |
4092 | } | |
4093 | ||
31169714 | 4094 | static int |
1495f230 | 4095 | i915_gem_inactive_shrink(struct shrinker *shrinker, struct shrink_control *sc) |
31169714 | 4096 | { |
17250b71 CW |
4097 | struct drm_i915_private *dev_priv = |
4098 | container_of(shrinker, | |
4099 | struct drm_i915_private, | |
4100 | mm.inactive_shrinker); | |
4101 | struct drm_device *dev = dev_priv->dev; | |
4102 | struct drm_i915_gem_object *obj, *next; | |
1495f230 | 4103 | int nr_to_scan = sc->nr_to_scan; |
17250b71 CW |
4104 | int cnt; |
4105 | ||
4106 | if (!mutex_trylock(&dev->struct_mutex)) | |
bbe2e11a | 4107 | return 0; |
31169714 CW |
4108 | |
4109 | /* "fast-path" to count number of available objects */ | |
4110 | if (nr_to_scan == 0) { | |
17250b71 CW |
4111 | cnt = 0; |
4112 | list_for_each_entry(obj, | |
4113 | &dev_priv->mm.inactive_list, | |
4114 | mm_list) | |
4115 | cnt++; | |
4116 | mutex_unlock(&dev->struct_mutex); | |
4117 | return cnt / 100 * sysctl_vfs_cache_pressure; | |
31169714 CW |
4118 | } |
4119 | ||
1637ef41 | 4120 | rescan: |
31169714 | 4121 | /* first scan for clean buffers */ |
17250b71 | 4122 | i915_gem_retire_requests(dev); |
31169714 | 4123 | |
17250b71 CW |
4124 | list_for_each_entry_safe(obj, next, |
4125 | &dev_priv->mm.inactive_list, | |
4126 | mm_list) { | |
4127 | if (i915_gem_object_is_purgeable(obj)) { | |
2021746e CW |
4128 | if (i915_gem_object_unbind(obj) == 0 && |
4129 | --nr_to_scan == 0) | |
17250b71 | 4130 | break; |
31169714 | 4131 | } |
31169714 CW |
4132 | } |
4133 | ||
4134 | /* second pass, evict/count anything still on the inactive list */ | |
17250b71 CW |
4135 | cnt = 0; |
4136 | list_for_each_entry_safe(obj, next, | |
4137 | &dev_priv->mm.inactive_list, | |
4138 | mm_list) { | |
2021746e CW |
4139 | if (nr_to_scan && |
4140 | i915_gem_object_unbind(obj) == 0) | |
17250b71 | 4141 | nr_to_scan--; |
2021746e | 4142 | else |
17250b71 CW |
4143 | cnt++; |
4144 | } | |
4145 | ||
4146 | if (nr_to_scan && i915_gpu_is_active(dev)) { | |
1637ef41 CW |
4147 | /* |
4148 | * We are desperate for pages, so as a last resort, wait | |
4149 | * for the GPU to finish and discard whatever we can. | |
4150 | * This has a dramatic impact to reduce the number of | |
4151 | * OOM-killer events whilst running the GPU aggressively. | |
4152 | */ | |
b93f9cf1 | 4153 | if (i915_gpu_idle(dev, true) == 0) |
1637ef41 CW |
4154 | goto rescan; |
4155 | } | |
17250b71 CW |
4156 | mutex_unlock(&dev->struct_mutex); |
4157 | return cnt / 100 * sysctl_vfs_cache_pressure; | |
31169714 | 4158 | } |