]>
Commit | Line | Data |
---|---|---|
673a394b | 1 | /* |
be6a0376 | 2 | * Copyright © 2008-2015 Intel Corporation |
673a394b EA |
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 | ||
760285e7 | 28 | #include <drm/drmP.h> |
0de23977 | 29 | #include <drm/drm_vma_manager.h> |
760285e7 | 30 | #include <drm/i915_drm.h> |
673a394b | 31 | #include "i915_drv.h" |
57822dc6 | 32 | #include "i915_gem_clflush.h" |
eb82289a | 33 | #include "i915_vgpu.h" |
1c5d22f7 | 34 | #include "i915_trace.h" |
652c393a | 35 | #include "intel_drv.h" |
5d723d7a | 36 | #include "intel_frontbuffer.h" |
0ccdacf6 | 37 | #include "intel_mocs.h" |
6b5e90f5 | 38 | #include <linux/dma-fence-array.h> |
fe3288b5 | 39 | #include <linux/kthread.h> |
c13d87ea | 40 | #include <linux/reservation.h> |
5949eac4 | 41 | #include <linux/shmem_fs.h> |
5a0e3ad6 | 42 | #include <linux/slab.h> |
20e4933c | 43 | #include <linux/stop_machine.h> |
673a394b | 44 | #include <linux/swap.h> |
79e53945 | 45 | #include <linux/pci.h> |
1286ff73 | 46 | #include <linux/dma-buf.h> |
673a394b | 47 | |
fbbd37b3 | 48 | static void i915_gem_flush_free_objects(struct drm_i915_private *i915); |
05394f39 | 49 | static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj); |
e62b59e4 | 50 | static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj); |
61050808 | 51 | |
2c22569b CW |
52 | static bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj) |
53 | { | |
b50a5371 AS |
54 | if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) |
55 | return false; | |
56 | ||
e59dc172 | 57 | if (!i915_gem_object_is_coherent(obj)) |
2c22569b CW |
58 | return true; |
59 | ||
60 | return obj->pin_display; | |
61 | } | |
62 | ||
4f1959ee | 63 | static int |
bb6dc8d9 | 64 | insert_mappable_node(struct i915_ggtt *ggtt, |
4f1959ee AS |
65 | struct drm_mm_node *node, u32 size) |
66 | { | |
67 | memset(node, 0, sizeof(*node)); | |
4e64e553 CW |
68 | return drm_mm_insert_node_in_range(&ggtt->base.mm, node, |
69 | size, 0, I915_COLOR_UNEVICTABLE, | |
70 | 0, ggtt->mappable_end, | |
71 | DRM_MM_INSERT_LOW); | |
4f1959ee AS |
72 | } |
73 | ||
74 | static void | |
75 | remove_mappable_node(struct drm_mm_node *node) | |
76 | { | |
77 | drm_mm_remove_node(node); | |
78 | } | |
79 | ||
73aa808f CW |
80 | /* some bookkeeping */ |
81 | static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv, | |
3ef7f228 | 82 | u64 size) |
73aa808f | 83 | { |
c20e8355 | 84 | spin_lock(&dev_priv->mm.object_stat_lock); |
73aa808f CW |
85 | dev_priv->mm.object_count++; |
86 | dev_priv->mm.object_memory += size; | |
c20e8355 | 87 | spin_unlock(&dev_priv->mm.object_stat_lock); |
73aa808f CW |
88 | } |
89 | ||
90 | static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv, | |
3ef7f228 | 91 | u64 size) |
73aa808f | 92 | { |
c20e8355 | 93 | spin_lock(&dev_priv->mm.object_stat_lock); |
73aa808f CW |
94 | dev_priv->mm.object_count--; |
95 | dev_priv->mm.object_memory -= size; | |
c20e8355 | 96 | spin_unlock(&dev_priv->mm.object_stat_lock); |
73aa808f CW |
97 | } |
98 | ||
21dd3734 | 99 | static int |
33196ded | 100 | i915_gem_wait_for_error(struct i915_gpu_error *error) |
30dbf0c0 | 101 | { |
30dbf0c0 CW |
102 | int ret; |
103 | ||
4c7d62c6 CW |
104 | might_sleep(); |
105 | ||
0a6759c6 DV |
106 | /* |
107 | * Only wait 10 seconds for the gpu reset to complete to avoid hanging | |
108 | * userspace. If it takes that long something really bad is going on and | |
109 | * we should simply try to bail out and fail as gracefully as possible. | |
110 | */ | |
1f83fee0 | 111 | ret = wait_event_interruptible_timeout(error->reset_queue, |
8c185eca | 112 | !i915_reset_backoff(error), |
b52992c0 | 113 | I915_RESET_TIMEOUT); |
0a6759c6 DV |
114 | if (ret == 0) { |
115 | DRM_ERROR("Timed out waiting for the gpu reset to complete\n"); | |
116 | return -EIO; | |
117 | } else if (ret < 0) { | |
30dbf0c0 | 118 | return ret; |
d98c52cf CW |
119 | } else { |
120 | return 0; | |
0a6759c6 | 121 | } |
30dbf0c0 CW |
122 | } |
123 | ||
54cf91dc | 124 | int i915_mutex_lock_interruptible(struct drm_device *dev) |
76c1dec1 | 125 | { |
fac5e23e | 126 | struct drm_i915_private *dev_priv = to_i915(dev); |
76c1dec1 CW |
127 | int ret; |
128 | ||
33196ded | 129 | ret = i915_gem_wait_for_error(&dev_priv->gpu_error); |
76c1dec1 CW |
130 | if (ret) |
131 | return ret; | |
132 | ||
133 | ret = mutex_lock_interruptible(&dev->struct_mutex); | |
134 | if (ret) | |
135 | return ret; | |
136 | ||
76c1dec1 CW |
137 | return 0; |
138 | } | |
30dbf0c0 | 139 | |
5a125c3c EA |
140 | int |
141 | i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 142 | struct drm_file *file) |
5a125c3c | 143 | { |
72e96d64 | 144 | struct drm_i915_private *dev_priv = to_i915(dev); |
62106b4f | 145 | struct i915_ggtt *ggtt = &dev_priv->ggtt; |
72e96d64 | 146 | struct drm_i915_gem_get_aperture *args = data; |
ca1543be | 147 | struct i915_vma *vma; |
6299f992 | 148 | size_t pinned; |
5a125c3c | 149 | |
6299f992 | 150 | pinned = 0; |
73aa808f | 151 | mutex_lock(&dev->struct_mutex); |
1c7f4bca | 152 | list_for_each_entry(vma, &ggtt->base.active_list, vm_link) |
20dfbde4 | 153 | if (i915_vma_is_pinned(vma)) |
ca1543be | 154 | pinned += vma->node.size; |
1c7f4bca | 155 | list_for_each_entry(vma, &ggtt->base.inactive_list, vm_link) |
20dfbde4 | 156 | if (i915_vma_is_pinned(vma)) |
ca1543be | 157 | pinned += vma->node.size; |
73aa808f | 158 | mutex_unlock(&dev->struct_mutex); |
5a125c3c | 159 | |
72e96d64 | 160 | args->aper_size = ggtt->base.total; |
0206e353 | 161 | args->aper_available_size = args->aper_size - pinned; |
6299f992 | 162 | |
5a125c3c EA |
163 | return 0; |
164 | } | |
165 | ||
03ac84f1 | 166 | static struct sg_table * |
6a2c4232 | 167 | i915_gem_object_get_pages_phys(struct drm_i915_gem_object *obj) |
00731155 | 168 | { |
93c76a3d | 169 | struct address_space *mapping = obj->base.filp->f_mapping; |
dbb4351b | 170 | drm_dma_handle_t *phys; |
6a2c4232 CW |
171 | struct sg_table *st; |
172 | struct scatterlist *sg; | |
dbb4351b | 173 | char *vaddr; |
6a2c4232 | 174 | int i; |
00731155 | 175 | |
6a2c4232 | 176 | if (WARN_ON(i915_gem_object_needs_bit17_swizzle(obj))) |
03ac84f1 | 177 | return ERR_PTR(-EINVAL); |
6a2c4232 | 178 | |
dbb4351b CW |
179 | /* Always aligning to the object size, allows a single allocation |
180 | * to handle all possible callers, and given typical object sizes, | |
181 | * the alignment of the buddy allocation will naturally match. | |
182 | */ | |
183 | phys = drm_pci_alloc(obj->base.dev, | |
184 | obj->base.size, | |
185 | roundup_pow_of_two(obj->base.size)); | |
186 | if (!phys) | |
187 | return ERR_PTR(-ENOMEM); | |
188 | ||
189 | vaddr = phys->vaddr; | |
6a2c4232 CW |
190 | for (i = 0; i < obj->base.size / PAGE_SIZE; i++) { |
191 | struct page *page; | |
192 | char *src; | |
193 | ||
194 | page = shmem_read_mapping_page(mapping, i); | |
dbb4351b CW |
195 | if (IS_ERR(page)) { |
196 | st = ERR_CAST(page); | |
197 | goto err_phys; | |
198 | } | |
6a2c4232 CW |
199 | |
200 | src = kmap_atomic(page); | |
201 | memcpy(vaddr, src, PAGE_SIZE); | |
202 | drm_clflush_virt_range(vaddr, PAGE_SIZE); | |
203 | kunmap_atomic(src); | |
204 | ||
09cbfeaf | 205 | put_page(page); |
6a2c4232 CW |
206 | vaddr += PAGE_SIZE; |
207 | } | |
208 | ||
c033666a | 209 | i915_gem_chipset_flush(to_i915(obj->base.dev)); |
6a2c4232 CW |
210 | |
211 | st = kmalloc(sizeof(*st), GFP_KERNEL); | |
dbb4351b CW |
212 | if (!st) { |
213 | st = ERR_PTR(-ENOMEM); | |
214 | goto err_phys; | |
215 | } | |
6a2c4232 CW |
216 | |
217 | if (sg_alloc_table(st, 1, GFP_KERNEL)) { | |
218 | kfree(st); | |
dbb4351b CW |
219 | st = ERR_PTR(-ENOMEM); |
220 | goto err_phys; | |
6a2c4232 CW |
221 | } |
222 | ||
223 | sg = st->sgl; | |
224 | sg->offset = 0; | |
225 | sg->length = obj->base.size; | |
00731155 | 226 | |
dbb4351b | 227 | sg_dma_address(sg) = phys->busaddr; |
6a2c4232 CW |
228 | sg_dma_len(sg) = obj->base.size; |
229 | ||
dbb4351b CW |
230 | obj->phys_handle = phys; |
231 | return st; | |
232 | ||
233 | err_phys: | |
234 | drm_pci_free(obj->base.dev, phys); | |
03ac84f1 | 235 | return st; |
6a2c4232 CW |
236 | } |
237 | ||
238 | static void | |
2b3c8317 | 239 | __i915_gem_object_release_shmem(struct drm_i915_gem_object *obj, |
e5facdf9 CW |
240 | struct sg_table *pages, |
241 | bool needs_clflush) | |
6a2c4232 | 242 | { |
a4f5ea64 | 243 | GEM_BUG_ON(obj->mm.madv == __I915_MADV_PURGED); |
00731155 | 244 | |
a4f5ea64 CW |
245 | if (obj->mm.madv == I915_MADV_DONTNEED) |
246 | obj->mm.dirty = false; | |
6a2c4232 | 247 | |
e5facdf9 CW |
248 | if (needs_clflush && |
249 | (obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0 && | |
e59dc172 | 250 | !i915_gem_object_is_coherent(obj)) |
2b3c8317 | 251 | drm_clflush_sg(pages); |
03ac84f1 CW |
252 | |
253 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
254 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
255 | } | |
256 | ||
257 | static void | |
258 | i915_gem_object_put_pages_phys(struct drm_i915_gem_object *obj, | |
259 | struct sg_table *pages) | |
260 | { | |
e5facdf9 | 261 | __i915_gem_object_release_shmem(obj, pages, false); |
03ac84f1 | 262 | |
a4f5ea64 | 263 | if (obj->mm.dirty) { |
93c76a3d | 264 | struct address_space *mapping = obj->base.filp->f_mapping; |
6a2c4232 | 265 | char *vaddr = obj->phys_handle->vaddr; |
00731155 CW |
266 | int i; |
267 | ||
268 | for (i = 0; i < obj->base.size / PAGE_SIZE; i++) { | |
6a2c4232 CW |
269 | struct page *page; |
270 | char *dst; | |
271 | ||
272 | page = shmem_read_mapping_page(mapping, i); | |
273 | if (IS_ERR(page)) | |
274 | continue; | |
275 | ||
276 | dst = kmap_atomic(page); | |
277 | drm_clflush_virt_range(vaddr, PAGE_SIZE); | |
278 | memcpy(dst, vaddr, PAGE_SIZE); | |
279 | kunmap_atomic(dst); | |
280 | ||
281 | set_page_dirty(page); | |
a4f5ea64 | 282 | if (obj->mm.madv == I915_MADV_WILLNEED) |
00731155 | 283 | mark_page_accessed(page); |
09cbfeaf | 284 | put_page(page); |
00731155 CW |
285 | vaddr += PAGE_SIZE; |
286 | } | |
a4f5ea64 | 287 | obj->mm.dirty = false; |
00731155 CW |
288 | } |
289 | ||
03ac84f1 CW |
290 | sg_free_table(pages); |
291 | kfree(pages); | |
dbb4351b CW |
292 | |
293 | drm_pci_free(obj->base.dev, obj->phys_handle); | |
6a2c4232 CW |
294 | } |
295 | ||
296 | static void | |
297 | i915_gem_object_release_phys(struct drm_i915_gem_object *obj) | |
298 | { | |
a4f5ea64 | 299 | i915_gem_object_unpin_pages(obj); |
6a2c4232 CW |
300 | } |
301 | ||
302 | static const struct drm_i915_gem_object_ops i915_gem_phys_ops = { | |
303 | .get_pages = i915_gem_object_get_pages_phys, | |
304 | .put_pages = i915_gem_object_put_pages_phys, | |
305 | .release = i915_gem_object_release_phys, | |
306 | }; | |
307 | ||
581ab1fe CW |
308 | static const struct drm_i915_gem_object_ops i915_gem_object_ops; |
309 | ||
35a9611c | 310 | int i915_gem_object_unbind(struct drm_i915_gem_object *obj) |
aa653a68 CW |
311 | { |
312 | struct i915_vma *vma; | |
313 | LIST_HEAD(still_in_list); | |
02bef8f9 CW |
314 | int ret; |
315 | ||
316 | lockdep_assert_held(&obj->base.dev->struct_mutex); | |
aa653a68 | 317 | |
02bef8f9 CW |
318 | /* Closed vma are removed from the obj->vma_list - but they may |
319 | * still have an active binding on the object. To remove those we | |
320 | * must wait for all rendering to complete to the object (as unbinding | |
321 | * must anyway), and retire the requests. | |
aa653a68 | 322 | */ |
e95433c7 CW |
323 | ret = i915_gem_object_wait(obj, |
324 | I915_WAIT_INTERRUPTIBLE | | |
325 | I915_WAIT_LOCKED | | |
326 | I915_WAIT_ALL, | |
327 | MAX_SCHEDULE_TIMEOUT, | |
328 | NULL); | |
02bef8f9 CW |
329 | if (ret) |
330 | return ret; | |
331 | ||
332 | i915_gem_retire_requests(to_i915(obj->base.dev)); | |
333 | ||
aa653a68 CW |
334 | while ((vma = list_first_entry_or_null(&obj->vma_list, |
335 | struct i915_vma, | |
336 | obj_link))) { | |
337 | list_move_tail(&vma->obj_link, &still_in_list); | |
338 | ret = i915_vma_unbind(vma); | |
339 | if (ret) | |
340 | break; | |
341 | } | |
342 | list_splice(&still_in_list, &obj->vma_list); | |
343 | ||
344 | return ret; | |
345 | } | |
346 | ||
e95433c7 CW |
347 | static long |
348 | i915_gem_object_wait_fence(struct dma_fence *fence, | |
349 | unsigned int flags, | |
350 | long timeout, | |
351 | struct intel_rps_client *rps) | |
00e60f26 | 352 | { |
e95433c7 | 353 | struct drm_i915_gem_request *rq; |
00e60f26 | 354 | |
e95433c7 | 355 | BUILD_BUG_ON(I915_WAIT_INTERRUPTIBLE != 0x1); |
00e60f26 | 356 | |
e95433c7 CW |
357 | if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) |
358 | return timeout; | |
359 | ||
360 | if (!dma_fence_is_i915(fence)) | |
361 | return dma_fence_wait_timeout(fence, | |
362 | flags & I915_WAIT_INTERRUPTIBLE, | |
363 | timeout); | |
364 | ||
365 | rq = to_request(fence); | |
366 | if (i915_gem_request_completed(rq)) | |
367 | goto out; | |
368 | ||
369 | /* This client is about to stall waiting for the GPU. In many cases | |
370 | * this is undesirable and limits the throughput of the system, as | |
371 | * many clients cannot continue processing user input/output whilst | |
372 | * blocked. RPS autotuning may take tens of milliseconds to respond | |
373 | * to the GPU load and thus incurs additional latency for the client. | |
374 | * We can circumvent that by promoting the GPU frequency to maximum | |
375 | * before we wait. This makes the GPU throttle up much more quickly | |
376 | * (good for benchmarks and user experience, e.g. window animations), | |
377 | * but at a cost of spending more power processing the workload | |
378 | * (bad for battery). Not all clients even want their results | |
379 | * immediately and for them we should just let the GPU select its own | |
380 | * frequency to maximise efficiency. To prevent a single client from | |
381 | * forcing the clocks too high for the whole system, we only allow | |
382 | * each client to waitboost once in a busy period. | |
383 | */ | |
384 | if (rps) { | |
385 | if (INTEL_GEN(rq->i915) >= 6) | |
386 | gen6_rps_boost(rq->i915, rps, rq->emitted_jiffies); | |
387 | else | |
388 | rps = NULL; | |
00e60f26 CW |
389 | } |
390 | ||
e95433c7 CW |
391 | timeout = i915_wait_request(rq, flags, timeout); |
392 | ||
393 | out: | |
394 | if (flags & I915_WAIT_LOCKED && i915_gem_request_completed(rq)) | |
395 | i915_gem_request_retire_upto(rq); | |
396 | ||
754c9fd5 | 397 | if (rps && i915_gem_request_global_seqno(rq) == intel_engine_last_submit(rq->engine)) { |
e95433c7 CW |
398 | /* The GPU is now idle and this client has stalled. |
399 | * Since no other client has submitted a request in the | |
400 | * meantime, assume that this client is the only one | |
401 | * supplying work to the GPU but is unable to keep that | |
402 | * work supplied because it is waiting. Since the GPU is | |
403 | * then never kept fully busy, RPS autoclocking will | |
404 | * keep the clocks relatively low, causing further delays. | |
405 | * Compensate by giving the synchronous client credit for | |
406 | * a waitboost next time. | |
407 | */ | |
408 | spin_lock(&rq->i915->rps.client_lock); | |
409 | list_del_init(&rps->link); | |
410 | spin_unlock(&rq->i915->rps.client_lock); | |
411 | } | |
412 | ||
413 | return timeout; | |
414 | } | |
415 | ||
416 | static long | |
417 | i915_gem_object_wait_reservation(struct reservation_object *resv, | |
418 | unsigned int flags, | |
419 | long timeout, | |
420 | struct intel_rps_client *rps) | |
421 | { | |
e54ca977 | 422 | unsigned int seq = __read_seqcount_begin(&resv->seq); |
e95433c7 | 423 | struct dma_fence *excl; |
e54ca977 | 424 | bool prune_fences = false; |
e95433c7 CW |
425 | |
426 | if (flags & I915_WAIT_ALL) { | |
427 | struct dma_fence **shared; | |
428 | unsigned int count, i; | |
00e60f26 CW |
429 | int ret; |
430 | ||
e95433c7 CW |
431 | ret = reservation_object_get_fences_rcu(resv, |
432 | &excl, &count, &shared); | |
00e60f26 CW |
433 | if (ret) |
434 | return ret; | |
00e60f26 | 435 | |
e95433c7 CW |
436 | for (i = 0; i < count; i++) { |
437 | timeout = i915_gem_object_wait_fence(shared[i], | |
438 | flags, timeout, | |
439 | rps); | |
d892e939 | 440 | if (timeout < 0) |
e95433c7 | 441 | break; |
00e60f26 | 442 | |
e95433c7 CW |
443 | dma_fence_put(shared[i]); |
444 | } | |
445 | ||
446 | for (; i < count; i++) | |
447 | dma_fence_put(shared[i]); | |
448 | kfree(shared); | |
e54ca977 CW |
449 | |
450 | prune_fences = count && timeout >= 0; | |
e95433c7 CW |
451 | } else { |
452 | excl = reservation_object_get_excl_rcu(resv); | |
00e60f26 CW |
453 | } |
454 | ||
e54ca977 | 455 | if (excl && timeout >= 0) { |
e95433c7 | 456 | timeout = i915_gem_object_wait_fence(excl, flags, timeout, rps); |
e54ca977 CW |
457 | prune_fences = timeout >= 0; |
458 | } | |
e95433c7 CW |
459 | |
460 | dma_fence_put(excl); | |
461 | ||
03d1cac6 CW |
462 | /* Oportunistically prune the fences iff we know they have *all* been |
463 | * signaled and that the reservation object has not been changed (i.e. | |
464 | * no new fences have been added). | |
465 | */ | |
e54ca977 | 466 | if (prune_fences && !__read_seqcount_retry(&resv->seq, seq)) { |
03d1cac6 CW |
467 | if (reservation_object_trylock(resv)) { |
468 | if (!__read_seqcount_retry(&resv->seq, seq)) | |
469 | reservation_object_add_excl_fence(resv, NULL); | |
470 | reservation_object_unlock(resv); | |
471 | } | |
e54ca977 CW |
472 | } |
473 | ||
e95433c7 | 474 | return timeout; |
00e60f26 CW |
475 | } |
476 | ||
6b5e90f5 CW |
477 | static void __fence_set_priority(struct dma_fence *fence, int prio) |
478 | { | |
479 | struct drm_i915_gem_request *rq; | |
480 | struct intel_engine_cs *engine; | |
481 | ||
482 | if (!dma_fence_is_i915(fence)) | |
483 | return; | |
484 | ||
485 | rq = to_request(fence); | |
486 | engine = rq->engine; | |
487 | if (!engine->schedule) | |
488 | return; | |
489 | ||
490 | engine->schedule(rq, prio); | |
491 | } | |
492 | ||
493 | static void fence_set_priority(struct dma_fence *fence, int prio) | |
494 | { | |
495 | /* Recurse once into a fence-array */ | |
496 | if (dma_fence_is_array(fence)) { | |
497 | struct dma_fence_array *array = to_dma_fence_array(fence); | |
498 | int i; | |
499 | ||
500 | for (i = 0; i < array->num_fences; i++) | |
501 | __fence_set_priority(array->fences[i], prio); | |
502 | } else { | |
503 | __fence_set_priority(fence, prio); | |
504 | } | |
505 | } | |
506 | ||
507 | int | |
508 | i915_gem_object_wait_priority(struct drm_i915_gem_object *obj, | |
509 | unsigned int flags, | |
510 | int prio) | |
511 | { | |
512 | struct dma_fence *excl; | |
513 | ||
514 | if (flags & I915_WAIT_ALL) { | |
515 | struct dma_fence **shared; | |
516 | unsigned int count, i; | |
517 | int ret; | |
518 | ||
519 | ret = reservation_object_get_fences_rcu(obj->resv, | |
520 | &excl, &count, &shared); | |
521 | if (ret) | |
522 | return ret; | |
523 | ||
524 | for (i = 0; i < count; i++) { | |
525 | fence_set_priority(shared[i], prio); | |
526 | dma_fence_put(shared[i]); | |
527 | } | |
528 | ||
529 | kfree(shared); | |
530 | } else { | |
531 | excl = reservation_object_get_excl_rcu(obj->resv); | |
532 | } | |
533 | ||
534 | if (excl) { | |
535 | fence_set_priority(excl, prio); | |
536 | dma_fence_put(excl); | |
537 | } | |
538 | return 0; | |
539 | } | |
540 | ||
e95433c7 CW |
541 | /** |
542 | * Waits for rendering to the object to be completed | |
543 | * @obj: i915 gem object | |
544 | * @flags: how to wait (under a lock, for all rendering or just for writes etc) | |
545 | * @timeout: how long to wait | |
546 | * @rps: client (user process) to charge for any waitboosting | |
00e60f26 | 547 | */ |
e95433c7 CW |
548 | int |
549 | i915_gem_object_wait(struct drm_i915_gem_object *obj, | |
550 | unsigned int flags, | |
551 | long timeout, | |
552 | struct intel_rps_client *rps) | |
00e60f26 | 553 | { |
e95433c7 CW |
554 | might_sleep(); |
555 | #if IS_ENABLED(CONFIG_LOCKDEP) | |
556 | GEM_BUG_ON(debug_locks && | |
557 | !!lockdep_is_held(&obj->base.dev->struct_mutex) != | |
558 | !!(flags & I915_WAIT_LOCKED)); | |
559 | #endif | |
560 | GEM_BUG_ON(timeout < 0); | |
00e60f26 | 561 | |
d07f0e59 CW |
562 | timeout = i915_gem_object_wait_reservation(obj->resv, |
563 | flags, timeout, | |
564 | rps); | |
e95433c7 | 565 | return timeout < 0 ? timeout : 0; |
00e60f26 CW |
566 | } |
567 | ||
568 | static struct intel_rps_client *to_rps_client(struct drm_file *file) | |
569 | { | |
570 | struct drm_i915_file_private *fpriv = file->driver_priv; | |
571 | ||
572 | return &fpriv->rps; | |
573 | } | |
574 | ||
00731155 CW |
575 | int |
576 | i915_gem_object_attach_phys(struct drm_i915_gem_object *obj, | |
577 | int align) | |
578 | { | |
6a2c4232 | 579 | int ret; |
00731155 | 580 | |
dbb4351b CW |
581 | if (align > obj->base.size) |
582 | return -EINVAL; | |
00731155 | 583 | |
dbb4351b | 584 | if (obj->ops == &i915_gem_phys_ops) |
00731155 | 585 | return 0; |
00731155 | 586 | |
a4f5ea64 | 587 | if (obj->mm.madv != I915_MADV_WILLNEED) |
00731155 CW |
588 | return -EFAULT; |
589 | ||
590 | if (obj->base.filp == NULL) | |
591 | return -EINVAL; | |
592 | ||
4717ca9e CW |
593 | ret = i915_gem_object_unbind(obj); |
594 | if (ret) | |
595 | return ret; | |
596 | ||
548625ee | 597 | __i915_gem_object_put_pages(obj, I915_MM_NORMAL); |
03ac84f1 CW |
598 | if (obj->mm.pages) |
599 | return -EBUSY; | |
6a2c4232 | 600 | |
581ab1fe | 601 | GEM_BUG_ON(obj->ops != &i915_gem_object_ops); |
6a2c4232 CW |
602 | obj->ops = &i915_gem_phys_ops; |
603 | ||
581ab1fe CW |
604 | ret = i915_gem_object_pin_pages(obj); |
605 | if (ret) | |
606 | goto err_xfer; | |
607 | ||
608 | return 0; | |
609 | ||
610 | err_xfer: | |
611 | obj->ops = &i915_gem_object_ops; | |
612 | return ret; | |
00731155 CW |
613 | } |
614 | ||
615 | static int | |
616 | i915_gem_phys_pwrite(struct drm_i915_gem_object *obj, | |
617 | struct drm_i915_gem_pwrite *args, | |
03ac84f1 | 618 | struct drm_file *file) |
00731155 | 619 | { |
00731155 | 620 | void *vaddr = obj->phys_handle->vaddr + args->offset; |
3ed605bc | 621 | char __user *user_data = u64_to_user_ptr(args->data_ptr); |
6a2c4232 CW |
622 | |
623 | /* We manually control the domain here and pretend that it | |
624 | * remains coherent i.e. in the GTT domain, like shmem_pwrite. | |
625 | */ | |
77a0d1ca | 626 | intel_fb_obj_invalidate(obj, ORIGIN_CPU); |
10466d2a CW |
627 | if (copy_from_user(vaddr, user_data, args->size)) |
628 | return -EFAULT; | |
00731155 | 629 | |
6a2c4232 | 630 | drm_clflush_virt_range(vaddr, args->size); |
10466d2a | 631 | i915_gem_chipset_flush(to_i915(obj->base.dev)); |
063e4e6b | 632 | |
d59b21ec | 633 | intel_fb_obj_flush(obj, ORIGIN_CPU); |
10466d2a | 634 | return 0; |
00731155 CW |
635 | } |
636 | ||
187685cb | 637 | void *i915_gem_object_alloc(struct drm_i915_private *dev_priv) |
42dcedd4 | 638 | { |
efab6d8d | 639 | return kmem_cache_zalloc(dev_priv->objects, GFP_KERNEL); |
42dcedd4 CW |
640 | } |
641 | ||
642 | void i915_gem_object_free(struct drm_i915_gem_object *obj) | |
643 | { | |
fac5e23e | 644 | struct drm_i915_private *dev_priv = to_i915(obj->base.dev); |
efab6d8d | 645 | kmem_cache_free(dev_priv->objects, obj); |
42dcedd4 CW |
646 | } |
647 | ||
ff72145b DA |
648 | static int |
649 | i915_gem_create(struct drm_file *file, | |
12d79d78 | 650 | struct drm_i915_private *dev_priv, |
ff72145b DA |
651 | uint64_t size, |
652 | uint32_t *handle_p) | |
673a394b | 653 | { |
05394f39 | 654 | struct drm_i915_gem_object *obj; |
a1a2d1d3 PP |
655 | int ret; |
656 | u32 handle; | |
673a394b | 657 | |
ff72145b | 658 | size = roundup(size, PAGE_SIZE); |
8ffc0246 CW |
659 | if (size == 0) |
660 | return -EINVAL; | |
673a394b EA |
661 | |
662 | /* Allocate the new object */ | |
12d79d78 | 663 | obj = i915_gem_object_create(dev_priv, size); |
fe3db79b CW |
664 | if (IS_ERR(obj)) |
665 | return PTR_ERR(obj); | |
673a394b | 666 | |
05394f39 | 667 | ret = drm_gem_handle_create(file, &obj->base, &handle); |
202f2fef | 668 | /* drop reference from allocate - handle holds it now */ |
f0cd5182 | 669 | i915_gem_object_put(obj); |
d861e338 DV |
670 | if (ret) |
671 | return ret; | |
202f2fef | 672 | |
ff72145b | 673 | *handle_p = handle; |
673a394b EA |
674 | return 0; |
675 | } | |
676 | ||
ff72145b DA |
677 | int |
678 | i915_gem_dumb_create(struct drm_file *file, | |
679 | struct drm_device *dev, | |
680 | struct drm_mode_create_dumb *args) | |
681 | { | |
682 | /* have to work out size/pitch and return them */ | |
de45eaf7 | 683 | args->pitch = ALIGN(args->width * DIV_ROUND_UP(args->bpp, 8), 64); |
ff72145b | 684 | args->size = args->pitch * args->height; |
12d79d78 | 685 | return i915_gem_create(file, to_i915(dev), |
da6b51d0 | 686 | args->size, &args->handle); |
ff72145b DA |
687 | } |
688 | ||
ff72145b DA |
689 | /** |
690 | * Creates a new mm object and returns a handle to it. | |
14bb2c11 TU |
691 | * @dev: drm device pointer |
692 | * @data: ioctl data blob | |
693 | * @file: drm file pointer | |
ff72145b DA |
694 | */ |
695 | int | |
696 | i915_gem_create_ioctl(struct drm_device *dev, void *data, | |
697 | struct drm_file *file) | |
698 | { | |
12d79d78 | 699 | struct drm_i915_private *dev_priv = to_i915(dev); |
ff72145b | 700 | struct drm_i915_gem_create *args = data; |
63ed2cb2 | 701 | |
12d79d78 | 702 | i915_gem_flush_free_objects(dev_priv); |
fbbd37b3 | 703 | |
12d79d78 | 704 | return i915_gem_create(file, dev_priv, |
da6b51d0 | 705 | args->size, &args->handle); |
ff72145b DA |
706 | } |
707 | ||
8461d226 DV |
708 | static inline int |
709 | __copy_to_user_swizzled(char __user *cpu_vaddr, | |
710 | const char *gpu_vaddr, int gpu_offset, | |
711 | int length) | |
712 | { | |
713 | int ret, cpu_offset = 0; | |
714 | ||
715 | while (length > 0) { | |
716 | int cacheline_end = ALIGN(gpu_offset + 1, 64); | |
717 | int this_length = min(cacheline_end - gpu_offset, length); | |
718 | int swizzled_gpu_offset = gpu_offset ^ 64; | |
719 | ||
720 | ret = __copy_to_user(cpu_vaddr + cpu_offset, | |
721 | gpu_vaddr + swizzled_gpu_offset, | |
722 | this_length); | |
723 | if (ret) | |
724 | return ret + length; | |
725 | ||
726 | cpu_offset += this_length; | |
727 | gpu_offset += this_length; | |
728 | length -= this_length; | |
729 | } | |
730 | ||
731 | return 0; | |
732 | } | |
733 | ||
8c59967c | 734 | static inline int |
4f0c7cfb BW |
735 | __copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset, |
736 | const char __user *cpu_vaddr, | |
8c59967c DV |
737 | int length) |
738 | { | |
739 | int ret, cpu_offset = 0; | |
740 | ||
741 | while (length > 0) { | |
742 | int cacheline_end = ALIGN(gpu_offset + 1, 64); | |
743 | int this_length = min(cacheline_end - gpu_offset, length); | |
744 | int swizzled_gpu_offset = gpu_offset ^ 64; | |
745 | ||
746 | ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset, | |
747 | cpu_vaddr + cpu_offset, | |
748 | this_length); | |
749 | if (ret) | |
750 | return ret + length; | |
751 | ||
752 | cpu_offset += this_length; | |
753 | gpu_offset += this_length; | |
754 | length -= this_length; | |
755 | } | |
756 | ||
757 | return 0; | |
758 | } | |
759 | ||
4c914c0c BV |
760 | /* |
761 | * Pins the specified object's pages and synchronizes the object with | |
762 | * GPU accesses. Sets needs_clflush to non-zero if the caller should | |
763 | * flush the object from the CPU cache. | |
764 | */ | |
765 | int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj, | |
43394c7d | 766 | unsigned int *needs_clflush) |
4c914c0c BV |
767 | { |
768 | int ret; | |
769 | ||
e95433c7 | 770 | lockdep_assert_held(&obj->base.dev->struct_mutex); |
4c914c0c | 771 | |
e95433c7 | 772 | *needs_clflush = 0; |
43394c7d CW |
773 | if (!i915_gem_object_has_struct_page(obj)) |
774 | return -ENODEV; | |
4c914c0c | 775 | |
e95433c7 CW |
776 | ret = i915_gem_object_wait(obj, |
777 | I915_WAIT_INTERRUPTIBLE | | |
778 | I915_WAIT_LOCKED, | |
779 | MAX_SCHEDULE_TIMEOUT, | |
780 | NULL); | |
c13d87ea CW |
781 | if (ret) |
782 | return ret; | |
783 | ||
a4f5ea64 | 784 | ret = i915_gem_object_pin_pages(obj); |
9764951e CW |
785 | if (ret) |
786 | return ret; | |
787 | ||
7f5f95d8 CW |
788 | if (i915_gem_object_is_coherent(obj) || |
789 | !static_cpu_has(X86_FEATURE_CLFLUSH)) { | |
790 | ret = i915_gem_object_set_to_cpu_domain(obj, false); | |
791 | if (ret) | |
792 | goto err_unpin; | |
793 | else | |
794 | goto out; | |
795 | } | |
796 | ||
a314d5cb CW |
797 | i915_gem_object_flush_gtt_write_domain(obj); |
798 | ||
43394c7d CW |
799 | /* If we're not in the cpu read domain, set ourself into the gtt |
800 | * read domain and manually flush cachelines (if required). This | |
801 | * optimizes for the case when the gpu will dirty the data | |
802 | * anyway again before the next pread happens. | |
803 | */ | |
804 | if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) | |
7f5f95d8 | 805 | *needs_clflush = CLFLUSH_BEFORE; |
4c914c0c | 806 | |
7f5f95d8 | 807 | out: |
9764951e | 808 | /* return with the pages pinned */ |
43394c7d | 809 | return 0; |
9764951e CW |
810 | |
811 | err_unpin: | |
812 | i915_gem_object_unpin_pages(obj); | |
813 | return ret; | |
43394c7d CW |
814 | } |
815 | ||
816 | int i915_gem_obj_prepare_shmem_write(struct drm_i915_gem_object *obj, | |
817 | unsigned int *needs_clflush) | |
818 | { | |
819 | int ret; | |
820 | ||
e95433c7 CW |
821 | lockdep_assert_held(&obj->base.dev->struct_mutex); |
822 | ||
43394c7d CW |
823 | *needs_clflush = 0; |
824 | if (!i915_gem_object_has_struct_page(obj)) | |
825 | return -ENODEV; | |
826 | ||
e95433c7 CW |
827 | ret = i915_gem_object_wait(obj, |
828 | I915_WAIT_INTERRUPTIBLE | | |
829 | I915_WAIT_LOCKED | | |
830 | I915_WAIT_ALL, | |
831 | MAX_SCHEDULE_TIMEOUT, | |
832 | NULL); | |
43394c7d CW |
833 | if (ret) |
834 | return ret; | |
835 | ||
a4f5ea64 | 836 | ret = i915_gem_object_pin_pages(obj); |
9764951e CW |
837 | if (ret) |
838 | return ret; | |
839 | ||
7f5f95d8 CW |
840 | if (i915_gem_object_is_coherent(obj) || |
841 | !static_cpu_has(X86_FEATURE_CLFLUSH)) { | |
842 | ret = i915_gem_object_set_to_cpu_domain(obj, true); | |
843 | if (ret) | |
844 | goto err_unpin; | |
845 | else | |
846 | goto out; | |
847 | } | |
848 | ||
a314d5cb CW |
849 | i915_gem_object_flush_gtt_write_domain(obj); |
850 | ||
43394c7d CW |
851 | /* If we're not in the cpu write domain, set ourself into the |
852 | * gtt write domain and manually flush cachelines (as required). | |
853 | * This optimizes for the case when the gpu will use the data | |
854 | * right away and we therefore have to clflush anyway. | |
855 | */ | |
856 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) | |
7f5f95d8 | 857 | *needs_clflush |= CLFLUSH_AFTER; |
43394c7d CW |
858 | |
859 | /* Same trick applies to invalidate partially written cachelines read | |
860 | * before writing. | |
861 | */ | |
862 | if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) | |
7f5f95d8 | 863 | *needs_clflush |= CLFLUSH_BEFORE; |
43394c7d | 864 | |
7f5f95d8 | 865 | out: |
43394c7d | 866 | intel_fb_obj_invalidate(obj, ORIGIN_CPU); |
a4f5ea64 | 867 | obj->mm.dirty = true; |
9764951e | 868 | /* return with the pages pinned */ |
43394c7d | 869 | return 0; |
9764951e CW |
870 | |
871 | err_unpin: | |
872 | i915_gem_object_unpin_pages(obj); | |
873 | return ret; | |
4c914c0c BV |
874 | } |
875 | ||
23c18c71 DV |
876 | static void |
877 | shmem_clflush_swizzled_range(char *addr, unsigned long length, | |
878 | bool swizzled) | |
879 | { | |
e7e58eb5 | 880 | if (unlikely(swizzled)) { |
23c18c71 DV |
881 | unsigned long start = (unsigned long) addr; |
882 | unsigned long end = (unsigned long) addr + length; | |
883 | ||
884 | /* For swizzling simply ensure that we always flush both | |
885 | * channels. Lame, but simple and it works. Swizzled | |
886 | * pwrite/pread is far from a hotpath - current userspace | |
887 | * doesn't use it at all. */ | |
888 | start = round_down(start, 128); | |
889 | end = round_up(end, 128); | |
890 | ||
891 | drm_clflush_virt_range((void *)start, end - start); | |
892 | } else { | |
893 | drm_clflush_virt_range(addr, length); | |
894 | } | |
895 | ||
896 | } | |
897 | ||
d174bd64 DV |
898 | /* Only difference to the fast-path function is that this can handle bit17 |
899 | * and uses non-atomic copy and kmap functions. */ | |
900 | static int | |
bb6dc8d9 | 901 | shmem_pread_slow(struct page *page, int offset, int length, |
d174bd64 DV |
902 | char __user *user_data, |
903 | bool page_do_bit17_swizzling, bool needs_clflush) | |
904 | { | |
905 | char *vaddr; | |
906 | int ret; | |
907 | ||
908 | vaddr = kmap(page); | |
909 | if (needs_clflush) | |
bb6dc8d9 | 910 | shmem_clflush_swizzled_range(vaddr + offset, length, |
23c18c71 | 911 | page_do_bit17_swizzling); |
d174bd64 DV |
912 | |
913 | if (page_do_bit17_swizzling) | |
bb6dc8d9 | 914 | ret = __copy_to_user_swizzled(user_data, vaddr, offset, length); |
d174bd64 | 915 | else |
bb6dc8d9 | 916 | ret = __copy_to_user(user_data, vaddr + offset, length); |
d174bd64 DV |
917 | kunmap(page); |
918 | ||
f60d7f0c | 919 | return ret ? - EFAULT : 0; |
d174bd64 DV |
920 | } |
921 | ||
bb6dc8d9 CW |
922 | static int |
923 | shmem_pread(struct page *page, int offset, int length, char __user *user_data, | |
924 | bool page_do_bit17_swizzling, bool needs_clflush) | |
925 | { | |
926 | int ret; | |
927 | ||
928 | ret = -ENODEV; | |
929 | if (!page_do_bit17_swizzling) { | |
930 | char *vaddr = kmap_atomic(page); | |
931 | ||
932 | if (needs_clflush) | |
933 | drm_clflush_virt_range(vaddr + offset, length); | |
934 | ret = __copy_to_user_inatomic(user_data, vaddr + offset, length); | |
935 | kunmap_atomic(vaddr); | |
936 | } | |
937 | if (ret == 0) | |
938 | return 0; | |
939 | ||
940 | return shmem_pread_slow(page, offset, length, user_data, | |
941 | page_do_bit17_swizzling, needs_clflush); | |
942 | } | |
943 | ||
944 | static int | |
945 | i915_gem_shmem_pread(struct drm_i915_gem_object *obj, | |
946 | struct drm_i915_gem_pread *args) | |
947 | { | |
948 | char __user *user_data; | |
949 | u64 remain; | |
950 | unsigned int obj_do_bit17_swizzling; | |
951 | unsigned int needs_clflush; | |
952 | unsigned int idx, offset; | |
953 | int ret; | |
954 | ||
955 | obj_do_bit17_swizzling = 0; | |
956 | if (i915_gem_object_needs_bit17_swizzle(obj)) | |
957 | obj_do_bit17_swizzling = BIT(17); | |
958 | ||
959 | ret = mutex_lock_interruptible(&obj->base.dev->struct_mutex); | |
960 | if (ret) | |
961 | return ret; | |
962 | ||
963 | ret = i915_gem_obj_prepare_shmem_read(obj, &needs_clflush); | |
964 | mutex_unlock(&obj->base.dev->struct_mutex); | |
965 | if (ret) | |
966 | return ret; | |
967 | ||
968 | remain = args->size; | |
969 | user_data = u64_to_user_ptr(args->data_ptr); | |
970 | offset = offset_in_page(args->offset); | |
971 | for (idx = args->offset >> PAGE_SHIFT; remain; idx++) { | |
972 | struct page *page = i915_gem_object_get_page(obj, idx); | |
973 | int length; | |
974 | ||
975 | length = remain; | |
976 | if (offset + length > PAGE_SIZE) | |
977 | length = PAGE_SIZE - offset; | |
978 | ||
979 | ret = shmem_pread(page, offset, length, user_data, | |
980 | page_to_phys(page) & obj_do_bit17_swizzling, | |
981 | needs_clflush); | |
982 | if (ret) | |
983 | break; | |
984 | ||
985 | remain -= length; | |
986 | user_data += length; | |
987 | offset = 0; | |
988 | } | |
989 | ||
990 | i915_gem_obj_finish_shmem_access(obj); | |
991 | return ret; | |
992 | } | |
993 | ||
994 | static inline bool | |
995 | gtt_user_read(struct io_mapping *mapping, | |
996 | loff_t base, int offset, | |
997 | char __user *user_data, int length) | |
b50a5371 | 998 | { |
b50a5371 | 999 | void *vaddr; |
bb6dc8d9 | 1000 | unsigned long unwritten; |
b50a5371 | 1001 | |
b50a5371 | 1002 | /* We can use the cpu mem copy function because this is X86. */ |
bb6dc8d9 CW |
1003 | vaddr = (void __force *)io_mapping_map_atomic_wc(mapping, base); |
1004 | unwritten = __copy_to_user_inatomic(user_data, vaddr + offset, length); | |
1005 | io_mapping_unmap_atomic(vaddr); | |
1006 | if (unwritten) { | |
1007 | vaddr = (void __force *) | |
1008 | io_mapping_map_wc(mapping, base, PAGE_SIZE); | |
1009 | unwritten = copy_to_user(user_data, vaddr + offset, length); | |
1010 | io_mapping_unmap(vaddr); | |
1011 | } | |
b50a5371 AS |
1012 | return unwritten; |
1013 | } | |
1014 | ||
1015 | static int | |
bb6dc8d9 CW |
1016 | i915_gem_gtt_pread(struct drm_i915_gem_object *obj, |
1017 | const struct drm_i915_gem_pread *args) | |
b50a5371 | 1018 | { |
bb6dc8d9 CW |
1019 | struct drm_i915_private *i915 = to_i915(obj->base.dev); |
1020 | struct i915_ggtt *ggtt = &i915->ggtt; | |
b50a5371 | 1021 | struct drm_mm_node node; |
bb6dc8d9 CW |
1022 | struct i915_vma *vma; |
1023 | void __user *user_data; | |
1024 | u64 remain, offset; | |
b50a5371 AS |
1025 | int ret; |
1026 | ||
bb6dc8d9 CW |
1027 | ret = mutex_lock_interruptible(&i915->drm.struct_mutex); |
1028 | if (ret) | |
1029 | return ret; | |
1030 | ||
1031 | intel_runtime_pm_get(i915); | |
1032 | vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, | |
1033 | PIN_MAPPABLE | PIN_NONBLOCK); | |
18034584 CW |
1034 | if (!IS_ERR(vma)) { |
1035 | node.start = i915_ggtt_offset(vma); | |
1036 | node.allocated = false; | |
49ef5294 | 1037 | ret = i915_vma_put_fence(vma); |
18034584 CW |
1038 | if (ret) { |
1039 | i915_vma_unpin(vma); | |
1040 | vma = ERR_PTR(ret); | |
1041 | } | |
1042 | } | |
058d88c4 | 1043 | if (IS_ERR(vma)) { |
bb6dc8d9 | 1044 | ret = insert_mappable_node(ggtt, &node, PAGE_SIZE); |
b50a5371 | 1045 | if (ret) |
bb6dc8d9 CW |
1046 | goto out_unlock; |
1047 | GEM_BUG_ON(!node.allocated); | |
b50a5371 AS |
1048 | } |
1049 | ||
1050 | ret = i915_gem_object_set_to_gtt_domain(obj, false); | |
1051 | if (ret) | |
1052 | goto out_unpin; | |
1053 | ||
bb6dc8d9 | 1054 | mutex_unlock(&i915->drm.struct_mutex); |
b50a5371 | 1055 | |
bb6dc8d9 CW |
1056 | user_data = u64_to_user_ptr(args->data_ptr); |
1057 | remain = args->size; | |
1058 | offset = args->offset; | |
b50a5371 AS |
1059 | |
1060 | while (remain > 0) { | |
1061 | /* Operation in this page | |
1062 | * | |
1063 | * page_base = page offset within aperture | |
1064 | * page_offset = offset within page | |
1065 | * page_length = bytes to copy for this page | |
1066 | */ | |
1067 | u32 page_base = node.start; | |
1068 | unsigned page_offset = offset_in_page(offset); | |
1069 | unsigned page_length = PAGE_SIZE - page_offset; | |
1070 | page_length = remain < page_length ? remain : page_length; | |
1071 | if (node.allocated) { | |
1072 | wmb(); | |
1073 | ggtt->base.insert_page(&ggtt->base, | |
1074 | i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT), | |
bb6dc8d9 | 1075 | node.start, I915_CACHE_NONE, 0); |
b50a5371 AS |
1076 | wmb(); |
1077 | } else { | |
1078 | page_base += offset & PAGE_MASK; | |
1079 | } | |
bb6dc8d9 CW |
1080 | |
1081 | if (gtt_user_read(&ggtt->mappable, page_base, page_offset, | |
1082 | user_data, page_length)) { | |
b50a5371 AS |
1083 | ret = -EFAULT; |
1084 | break; | |
1085 | } | |
1086 | ||
1087 | remain -= page_length; | |
1088 | user_data += page_length; | |
1089 | offset += page_length; | |
1090 | } | |
1091 | ||
bb6dc8d9 | 1092 | mutex_lock(&i915->drm.struct_mutex); |
b50a5371 AS |
1093 | out_unpin: |
1094 | if (node.allocated) { | |
1095 | wmb(); | |
1096 | ggtt->base.clear_range(&ggtt->base, | |
4fb84d99 | 1097 | node.start, node.size); |
b50a5371 AS |
1098 | remove_mappable_node(&node); |
1099 | } else { | |
058d88c4 | 1100 | i915_vma_unpin(vma); |
b50a5371 | 1101 | } |
bb6dc8d9 CW |
1102 | out_unlock: |
1103 | intel_runtime_pm_put(i915); | |
1104 | mutex_unlock(&i915->drm.struct_mutex); | |
f60d7f0c | 1105 | |
eb01459f EA |
1106 | return ret; |
1107 | } | |
1108 | ||
673a394b EA |
1109 | /** |
1110 | * Reads data from the object referenced by handle. | |
14bb2c11 TU |
1111 | * @dev: drm device pointer |
1112 | * @data: ioctl data blob | |
1113 | * @file: drm file pointer | |
673a394b EA |
1114 | * |
1115 | * On error, the contents of *data are undefined. | |
1116 | */ | |
1117 | int | |
1118 | i915_gem_pread_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 1119 | struct drm_file *file) |
673a394b EA |
1120 | { |
1121 | struct drm_i915_gem_pread *args = data; | |
05394f39 | 1122 | struct drm_i915_gem_object *obj; |
bb6dc8d9 | 1123 | int ret; |
673a394b | 1124 | |
51311d0a CW |
1125 | if (args->size == 0) |
1126 | return 0; | |
1127 | ||
1128 | if (!access_ok(VERIFY_WRITE, | |
3ed605bc | 1129 | u64_to_user_ptr(args->data_ptr), |
51311d0a CW |
1130 | args->size)) |
1131 | return -EFAULT; | |
1132 | ||
03ac0642 | 1133 | obj = i915_gem_object_lookup(file, args->handle); |
258a5ede CW |
1134 | if (!obj) |
1135 | return -ENOENT; | |
673a394b | 1136 | |
7dcd2499 | 1137 | /* Bounds check source. */ |
966d5bf5 | 1138 | if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) { |
ce9d419d | 1139 | ret = -EINVAL; |
bb6dc8d9 | 1140 | goto out; |
ce9d419d CW |
1141 | } |
1142 | ||
db53a302 CW |
1143 | trace_i915_gem_object_pread(obj, args->offset, args->size); |
1144 | ||
e95433c7 CW |
1145 | ret = i915_gem_object_wait(obj, |
1146 | I915_WAIT_INTERRUPTIBLE, | |
1147 | MAX_SCHEDULE_TIMEOUT, | |
1148 | to_rps_client(file)); | |
258a5ede | 1149 | if (ret) |
bb6dc8d9 | 1150 | goto out; |
258a5ede | 1151 | |
bb6dc8d9 | 1152 | ret = i915_gem_object_pin_pages(obj); |
258a5ede | 1153 | if (ret) |
bb6dc8d9 | 1154 | goto out; |
673a394b | 1155 | |
bb6dc8d9 | 1156 | ret = i915_gem_shmem_pread(obj, args); |
9c870d03 | 1157 | if (ret == -EFAULT || ret == -ENODEV) |
bb6dc8d9 | 1158 | ret = i915_gem_gtt_pread(obj, args); |
b50a5371 | 1159 | |
bb6dc8d9 CW |
1160 | i915_gem_object_unpin_pages(obj); |
1161 | out: | |
f0cd5182 | 1162 | i915_gem_object_put(obj); |
eb01459f | 1163 | return ret; |
673a394b EA |
1164 | } |
1165 | ||
0839ccb8 KP |
1166 | /* This is the fast write path which cannot handle |
1167 | * page faults in the source data | |
9b7530cc | 1168 | */ |
0839ccb8 | 1169 | |
fe115628 CW |
1170 | static inline bool |
1171 | ggtt_write(struct io_mapping *mapping, | |
1172 | loff_t base, int offset, | |
1173 | char __user *user_data, int length) | |
9b7530cc | 1174 | { |
4f0c7cfb | 1175 | void *vaddr; |
0839ccb8 | 1176 | unsigned long unwritten; |
9b7530cc | 1177 | |
4f0c7cfb | 1178 | /* We can use the cpu mem copy function because this is X86. */ |
fe115628 CW |
1179 | vaddr = (void __force *)io_mapping_map_atomic_wc(mapping, base); |
1180 | unwritten = __copy_from_user_inatomic_nocache(vaddr + offset, | |
0839ccb8 | 1181 | user_data, length); |
fe115628 CW |
1182 | io_mapping_unmap_atomic(vaddr); |
1183 | if (unwritten) { | |
1184 | vaddr = (void __force *) | |
1185 | io_mapping_map_wc(mapping, base, PAGE_SIZE); | |
1186 | unwritten = copy_from_user(vaddr + offset, user_data, length); | |
1187 | io_mapping_unmap(vaddr); | |
1188 | } | |
bb6dc8d9 | 1189 | |
bb6dc8d9 CW |
1190 | return unwritten; |
1191 | } | |
1192 | ||
3de09aa3 EA |
1193 | /** |
1194 | * This is the fast pwrite path, where we copy the data directly from the | |
1195 | * user into the GTT, uncached. | |
fe115628 | 1196 | * @obj: i915 GEM object |
14bb2c11 | 1197 | * @args: pwrite arguments structure |
3de09aa3 | 1198 | */ |
673a394b | 1199 | static int |
fe115628 CW |
1200 | i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj, |
1201 | const struct drm_i915_gem_pwrite *args) | |
673a394b | 1202 | { |
fe115628 | 1203 | struct drm_i915_private *i915 = to_i915(obj->base.dev); |
4f1959ee AS |
1204 | struct i915_ggtt *ggtt = &i915->ggtt; |
1205 | struct drm_mm_node node; | |
fe115628 CW |
1206 | struct i915_vma *vma; |
1207 | u64 remain, offset; | |
1208 | void __user *user_data; | |
4f1959ee | 1209 | int ret; |
b50a5371 | 1210 | |
fe115628 CW |
1211 | ret = mutex_lock_interruptible(&i915->drm.struct_mutex); |
1212 | if (ret) | |
1213 | return ret; | |
935aaa69 | 1214 | |
9c870d03 | 1215 | intel_runtime_pm_get(i915); |
058d88c4 | 1216 | vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, |
de895082 | 1217 | PIN_MAPPABLE | PIN_NONBLOCK); |
18034584 CW |
1218 | if (!IS_ERR(vma)) { |
1219 | node.start = i915_ggtt_offset(vma); | |
1220 | node.allocated = false; | |
49ef5294 | 1221 | ret = i915_vma_put_fence(vma); |
18034584 CW |
1222 | if (ret) { |
1223 | i915_vma_unpin(vma); | |
1224 | vma = ERR_PTR(ret); | |
1225 | } | |
1226 | } | |
058d88c4 | 1227 | if (IS_ERR(vma)) { |
bb6dc8d9 | 1228 | ret = insert_mappable_node(ggtt, &node, PAGE_SIZE); |
4f1959ee | 1229 | if (ret) |
fe115628 CW |
1230 | goto out_unlock; |
1231 | GEM_BUG_ON(!node.allocated); | |
4f1959ee | 1232 | } |
935aaa69 DV |
1233 | |
1234 | ret = i915_gem_object_set_to_gtt_domain(obj, true); | |
1235 | if (ret) | |
1236 | goto out_unpin; | |
1237 | ||
fe115628 CW |
1238 | mutex_unlock(&i915->drm.struct_mutex); |
1239 | ||
b19482d7 | 1240 | intel_fb_obj_invalidate(obj, ORIGIN_CPU); |
063e4e6b | 1241 | |
4f1959ee AS |
1242 | user_data = u64_to_user_ptr(args->data_ptr); |
1243 | offset = args->offset; | |
1244 | remain = args->size; | |
1245 | while (remain) { | |
673a394b EA |
1246 | /* Operation in this page |
1247 | * | |
0839ccb8 KP |
1248 | * page_base = page offset within aperture |
1249 | * page_offset = offset within page | |
1250 | * page_length = bytes to copy for this page | |
673a394b | 1251 | */ |
4f1959ee | 1252 | u32 page_base = node.start; |
bb6dc8d9 CW |
1253 | unsigned int page_offset = offset_in_page(offset); |
1254 | unsigned int page_length = PAGE_SIZE - page_offset; | |
4f1959ee AS |
1255 | page_length = remain < page_length ? remain : page_length; |
1256 | if (node.allocated) { | |
1257 | wmb(); /* flush the write before we modify the GGTT */ | |
1258 | ggtt->base.insert_page(&ggtt->base, | |
1259 | i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT), | |
1260 | node.start, I915_CACHE_NONE, 0); | |
1261 | wmb(); /* flush modifications to the GGTT (insert_page) */ | |
1262 | } else { | |
1263 | page_base += offset & PAGE_MASK; | |
1264 | } | |
0839ccb8 | 1265 | /* If we get a fault while copying data, then (presumably) our |
3de09aa3 EA |
1266 | * source page isn't available. Return the error and we'll |
1267 | * retry in the slow path. | |
b50a5371 AS |
1268 | * If the object is non-shmem backed, we retry again with the |
1269 | * path that handles page fault. | |
0839ccb8 | 1270 | */ |
fe115628 CW |
1271 | if (ggtt_write(&ggtt->mappable, page_base, page_offset, |
1272 | user_data, page_length)) { | |
1273 | ret = -EFAULT; | |
1274 | break; | |
935aaa69 | 1275 | } |
673a394b | 1276 | |
0839ccb8 KP |
1277 | remain -= page_length; |
1278 | user_data += page_length; | |
1279 | offset += page_length; | |
673a394b | 1280 | } |
d59b21ec | 1281 | intel_fb_obj_flush(obj, ORIGIN_CPU); |
fe115628 CW |
1282 | |
1283 | mutex_lock(&i915->drm.struct_mutex); | |
935aaa69 | 1284 | out_unpin: |
4f1959ee AS |
1285 | if (node.allocated) { |
1286 | wmb(); | |
1287 | ggtt->base.clear_range(&ggtt->base, | |
4fb84d99 | 1288 | node.start, node.size); |
4f1959ee AS |
1289 | remove_mappable_node(&node); |
1290 | } else { | |
058d88c4 | 1291 | i915_vma_unpin(vma); |
4f1959ee | 1292 | } |
fe115628 | 1293 | out_unlock: |
9c870d03 | 1294 | intel_runtime_pm_put(i915); |
fe115628 | 1295 | mutex_unlock(&i915->drm.struct_mutex); |
3de09aa3 | 1296 | return ret; |
673a394b EA |
1297 | } |
1298 | ||
3043c60c | 1299 | static int |
fe115628 | 1300 | shmem_pwrite_slow(struct page *page, int offset, int length, |
d174bd64 DV |
1301 | char __user *user_data, |
1302 | bool page_do_bit17_swizzling, | |
1303 | bool needs_clflush_before, | |
1304 | bool needs_clflush_after) | |
673a394b | 1305 | { |
d174bd64 DV |
1306 | char *vaddr; |
1307 | int ret; | |
e5281ccd | 1308 | |
d174bd64 | 1309 | vaddr = kmap(page); |
e7e58eb5 | 1310 | if (unlikely(needs_clflush_before || page_do_bit17_swizzling)) |
fe115628 | 1311 | shmem_clflush_swizzled_range(vaddr + offset, length, |
23c18c71 | 1312 | page_do_bit17_swizzling); |
d174bd64 | 1313 | if (page_do_bit17_swizzling) |
fe115628 CW |
1314 | ret = __copy_from_user_swizzled(vaddr, offset, user_data, |
1315 | length); | |
d174bd64 | 1316 | else |
fe115628 | 1317 | ret = __copy_from_user(vaddr + offset, user_data, length); |
d174bd64 | 1318 | if (needs_clflush_after) |
fe115628 | 1319 | shmem_clflush_swizzled_range(vaddr + offset, length, |
23c18c71 | 1320 | page_do_bit17_swizzling); |
d174bd64 | 1321 | kunmap(page); |
40123c1f | 1322 | |
755d2218 | 1323 | return ret ? -EFAULT : 0; |
40123c1f EA |
1324 | } |
1325 | ||
fe115628 CW |
1326 | /* Per-page copy function for the shmem pwrite fastpath. |
1327 | * Flushes invalid cachelines before writing to the target if | |
1328 | * needs_clflush_before is set and flushes out any written cachelines after | |
1329 | * writing if needs_clflush is set. | |
1330 | */ | |
40123c1f | 1331 | static int |
fe115628 CW |
1332 | shmem_pwrite(struct page *page, int offset, int len, char __user *user_data, |
1333 | bool page_do_bit17_swizzling, | |
1334 | bool needs_clflush_before, | |
1335 | bool needs_clflush_after) | |
40123c1f | 1336 | { |
fe115628 CW |
1337 | int ret; |
1338 | ||
1339 | ret = -ENODEV; | |
1340 | if (!page_do_bit17_swizzling) { | |
1341 | char *vaddr = kmap_atomic(page); | |
1342 | ||
1343 | if (needs_clflush_before) | |
1344 | drm_clflush_virt_range(vaddr + offset, len); | |
1345 | ret = __copy_from_user_inatomic(vaddr + offset, user_data, len); | |
1346 | if (needs_clflush_after) | |
1347 | drm_clflush_virt_range(vaddr + offset, len); | |
1348 | ||
1349 | kunmap_atomic(vaddr); | |
1350 | } | |
1351 | if (ret == 0) | |
1352 | return ret; | |
1353 | ||
1354 | return shmem_pwrite_slow(page, offset, len, user_data, | |
1355 | page_do_bit17_swizzling, | |
1356 | needs_clflush_before, | |
1357 | needs_clflush_after); | |
1358 | } | |
1359 | ||
1360 | static int | |
1361 | i915_gem_shmem_pwrite(struct drm_i915_gem_object *obj, | |
1362 | const struct drm_i915_gem_pwrite *args) | |
1363 | { | |
1364 | struct drm_i915_private *i915 = to_i915(obj->base.dev); | |
1365 | void __user *user_data; | |
1366 | u64 remain; | |
1367 | unsigned int obj_do_bit17_swizzling; | |
1368 | unsigned int partial_cacheline_write; | |
43394c7d | 1369 | unsigned int needs_clflush; |
fe115628 CW |
1370 | unsigned int offset, idx; |
1371 | int ret; | |
40123c1f | 1372 | |
fe115628 | 1373 | ret = mutex_lock_interruptible(&i915->drm.struct_mutex); |
755d2218 CW |
1374 | if (ret) |
1375 | return ret; | |
1376 | ||
fe115628 CW |
1377 | ret = i915_gem_obj_prepare_shmem_write(obj, &needs_clflush); |
1378 | mutex_unlock(&i915->drm.struct_mutex); | |
1379 | if (ret) | |
1380 | return ret; | |
673a394b | 1381 | |
fe115628 CW |
1382 | obj_do_bit17_swizzling = 0; |
1383 | if (i915_gem_object_needs_bit17_swizzle(obj)) | |
1384 | obj_do_bit17_swizzling = BIT(17); | |
e5281ccd | 1385 | |
fe115628 CW |
1386 | /* If we don't overwrite a cacheline completely we need to be |
1387 | * careful to have up-to-date data by first clflushing. Don't | |
1388 | * overcomplicate things and flush the entire patch. | |
1389 | */ | |
1390 | partial_cacheline_write = 0; | |
1391 | if (needs_clflush & CLFLUSH_BEFORE) | |
1392 | partial_cacheline_write = boot_cpu_data.x86_clflush_size - 1; | |
9da3da66 | 1393 | |
fe115628 CW |
1394 | user_data = u64_to_user_ptr(args->data_ptr); |
1395 | remain = args->size; | |
1396 | offset = offset_in_page(args->offset); | |
1397 | for (idx = args->offset >> PAGE_SHIFT; remain; idx++) { | |
1398 | struct page *page = i915_gem_object_get_page(obj, idx); | |
1399 | int length; | |
40123c1f | 1400 | |
fe115628 CW |
1401 | length = remain; |
1402 | if (offset + length > PAGE_SIZE) | |
1403 | length = PAGE_SIZE - offset; | |
755d2218 | 1404 | |
fe115628 CW |
1405 | ret = shmem_pwrite(page, offset, length, user_data, |
1406 | page_to_phys(page) & obj_do_bit17_swizzling, | |
1407 | (offset | length) & partial_cacheline_write, | |
1408 | needs_clflush & CLFLUSH_AFTER); | |
755d2218 | 1409 | if (ret) |
fe115628 | 1410 | break; |
755d2218 | 1411 | |
fe115628 CW |
1412 | remain -= length; |
1413 | user_data += length; | |
1414 | offset = 0; | |
8c59967c | 1415 | } |
673a394b | 1416 | |
d59b21ec | 1417 | intel_fb_obj_flush(obj, ORIGIN_CPU); |
fe115628 | 1418 | i915_gem_obj_finish_shmem_access(obj); |
40123c1f | 1419 | return ret; |
673a394b EA |
1420 | } |
1421 | ||
1422 | /** | |
1423 | * Writes data to the object referenced by handle. | |
14bb2c11 TU |
1424 | * @dev: drm device |
1425 | * @data: ioctl data blob | |
1426 | * @file: drm file | |
673a394b EA |
1427 | * |
1428 | * On error, the contents of the buffer that were to be modified are undefined. | |
1429 | */ | |
1430 | int | |
1431 | i915_gem_pwrite_ioctl(struct drm_device *dev, void *data, | |
fbd5a26d | 1432 | struct drm_file *file) |
673a394b EA |
1433 | { |
1434 | struct drm_i915_gem_pwrite *args = data; | |
05394f39 | 1435 | struct drm_i915_gem_object *obj; |
51311d0a CW |
1436 | int ret; |
1437 | ||
1438 | if (args->size == 0) | |
1439 | return 0; | |
1440 | ||
1441 | if (!access_ok(VERIFY_READ, | |
3ed605bc | 1442 | u64_to_user_ptr(args->data_ptr), |
51311d0a CW |
1443 | args->size)) |
1444 | return -EFAULT; | |
1445 | ||
03ac0642 | 1446 | obj = i915_gem_object_lookup(file, args->handle); |
258a5ede CW |
1447 | if (!obj) |
1448 | return -ENOENT; | |
673a394b | 1449 | |
7dcd2499 | 1450 | /* Bounds check destination. */ |
966d5bf5 | 1451 | if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) { |
ce9d419d | 1452 | ret = -EINVAL; |
258a5ede | 1453 | goto err; |
ce9d419d CW |
1454 | } |
1455 | ||
db53a302 CW |
1456 | trace_i915_gem_object_pwrite(obj, args->offset, args->size); |
1457 | ||
7c55e2c5 CW |
1458 | ret = -ENODEV; |
1459 | if (obj->ops->pwrite) | |
1460 | ret = obj->ops->pwrite(obj, args); | |
1461 | if (ret != -ENODEV) | |
1462 | goto err; | |
1463 | ||
e95433c7 CW |
1464 | ret = i915_gem_object_wait(obj, |
1465 | I915_WAIT_INTERRUPTIBLE | | |
1466 | I915_WAIT_ALL, | |
1467 | MAX_SCHEDULE_TIMEOUT, | |
1468 | to_rps_client(file)); | |
258a5ede CW |
1469 | if (ret) |
1470 | goto err; | |
1471 | ||
fe115628 | 1472 | ret = i915_gem_object_pin_pages(obj); |
258a5ede | 1473 | if (ret) |
fe115628 | 1474 | goto err; |
258a5ede | 1475 | |
935aaa69 | 1476 | ret = -EFAULT; |
673a394b EA |
1477 | /* We can only do the GTT pwrite on untiled buffers, as otherwise |
1478 | * it would end up going through the fenced access, and we'll get | |
1479 | * different detiling behavior between reading and writing. | |
1480 | * pread/pwrite currently are reading and writing from the CPU | |
1481 | * perspective, requiring manual detiling by the client. | |
1482 | */ | |
6eae0059 | 1483 | if (!i915_gem_object_has_struct_page(obj) || |
9c870d03 | 1484 | cpu_write_needs_clflush(obj)) |
935aaa69 DV |
1485 | /* Note that the gtt paths might fail with non-page-backed user |
1486 | * pointers (e.g. gtt mappings when moving data between | |
9c870d03 CW |
1487 | * textures). Fallback to the shmem path in that case. |
1488 | */ | |
fe115628 | 1489 | ret = i915_gem_gtt_pwrite_fast(obj, args); |
673a394b | 1490 | |
d1054ee4 | 1491 | if (ret == -EFAULT || ret == -ENOSPC) { |
6a2c4232 CW |
1492 | if (obj->phys_handle) |
1493 | ret = i915_gem_phys_pwrite(obj, args, file); | |
b50a5371 | 1494 | else |
fe115628 | 1495 | ret = i915_gem_shmem_pwrite(obj, args); |
6a2c4232 | 1496 | } |
5c0480f2 | 1497 | |
fe115628 | 1498 | i915_gem_object_unpin_pages(obj); |
258a5ede | 1499 | err: |
f0cd5182 | 1500 | i915_gem_object_put(obj); |
258a5ede | 1501 | return ret; |
673a394b EA |
1502 | } |
1503 | ||
d243ad82 | 1504 | static inline enum fb_op_origin |
aeecc969 CW |
1505 | write_origin(struct drm_i915_gem_object *obj, unsigned domain) |
1506 | { | |
50349247 CW |
1507 | return (domain == I915_GEM_DOMAIN_GTT ? |
1508 | obj->frontbuffer_ggtt_origin : ORIGIN_CPU); | |
aeecc969 CW |
1509 | } |
1510 | ||
40e62d5d CW |
1511 | static void i915_gem_object_bump_inactive_ggtt(struct drm_i915_gem_object *obj) |
1512 | { | |
1513 | struct drm_i915_private *i915; | |
1514 | struct list_head *list; | |
1515 | struct i915_vma *vma; | |
1516 | ||
1517 | list_for_each_entry(vma, &obj->vma_list, obj_link) { | |
1518 | if (!i915_vma_is_ggtt(vma)) | |
28f412e0 | 1519 | break; |
40e62d5d CW |
1520 | |
1521 | if (i915_vma_is_active(vma)) | |
1522 | continue; | |
1523 | ||
1524 | if (!drm_mm_node_allocated(&vma->node)) | |
1525 | continue; | |
1526 | ||
1527 | list_move_tail(&vma->vm_link, &vma->vm->inactive_list); | |
1528 | } | |
1529 | ||
1530 | i915 = to_i915(obj->base.dev); | |
1531 | list = obj->bind_count ? &i915->mm.bound_list : &i915->mm.unbound_list; | |
56cea323 | 1532 | list_move_tail(&obj->global_link, list); |
40e62d5d CW |
1533 | } |
1534 | ||
673a394b | 1535 | /** |
2ef7eeaa EA |
1536 | * Called when user space prepares to use an object with the CPU, either |
1537 | * through the mmap ioctl's mapping or a GTT mapping. | |
14bb2c11 TU |
1538 | * @dev: drm device |
1539 | * @data: ioctl data blob | |
1540 | * @file: drm file | |
673a394b EA |
1541 | */ |
1542 | int | |
1543 | i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 1544 | struct drm_file *file) |
673a394b EA |
1545 | { |
1546 | struct drm_i915_gem_set_domain *args = data; | |
05394f39 | 1547 | struct drm_i915_gem_object *obj; |
2ef7eeaa EA |
1548 | uint32_t read_domains = args->read_domains; |
1549 | uint32_t write_domain = args->write_domain; | |
40e62d5d | 1550 | int err; |
673a394b | 1551 | |
2ef7eeaa | 1552 | /* Only handle setting domains to types used by the CPU. */ |
b8f9096d | 1553 | if ((write_domain | read_domains) & I915_GEM_GPU_DOMAINS) |
2ef7eeaa EA |
1554 | return -EINVAL; |
1555 | ||
1556 | /* Having something in the write domain implies it's in the read | |
1557 | * domain, and only that read domain. Enforce that in the request. | |
1558 | */ | |
1559 | if (write_domain != 0 && read_domains != write_domain) | |
1560 | return -EINVAL; | |
1561 | ||
03ac0642 | 1562 | obj = i915_gem_object_lookup(file, args->handle); |
b8f9096d CW |
1563 | if (!obj) |
1564 | return -ENOENT; | |
673a394b | 1565 | |
3236f57a CW |
1566 | /* Try to flush the object off the GPU without holding the lock. |
1567 | * We will repeat the flush holding the lock in the normal manner | |
1568 | * to catch cases where we are gazumped. | |
1569 | */ | |
40e62d5d | 1570 | err = i915_gem_object_wait(obj, |
e95433c7 CW |
1571 | I915_WAIT_INTERRUPTIBLE | |
1572 | (write_domain ? I915_WAIT_ALL : 0), | |
1573 | MAX_SCHEDULE_TIMEOUT, | |
1574 | to_rps_client(file)); | |
40e62d5d | 1575 | if (err) |
f0cd5182 | 1576 | goto out; |
b8f9096d | 1577 | |
40e62d5d CW |
1578 | /* Flush and acquire obj->pages so that we are coherent through |
1579 | * direct access in memory with previous cached writes through | |
1580 | * shmemfs and that our cache domain tracking remains valid. | |
1581 | * For example, if the obj->filp was moved to swap without us | |
1582 | * being notified and releasing the pages, we would mistakenly | |
1583 | * continue to assume that the obj remained out of the CPU cached | |
1584 | * domain. | |
1585 | */ | |
1586 | err = i915_gem_object_pin_pages(obj); | |
1587 | if (err) | |
f0cd5182 | 1588 | goto out; |
40e62d5d CW |
1589 | |
1590 | err = i915_mutex_lock_interruptible(dev); | |
1591 | if (err) | |
f0cd5182 | 1592 | goto out_unpin; |
3236f57a | 1593 | |
43566ded | 1594 | if (read_domains & I915_GEM_DOMAIN_GTT) |
40e62d5d | 1595 | err = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0); |
43566ded | 1596 | else |
40e62d5d | 1597 | err = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0); |
2ef7eeaa | 1598 | |
40e62d5d CW |
1599 | /* And bump the LRU for this access */ |
1600 | i915_gem_object_bump_inactive_ggtt(obj); | |
031b698a | 1601 | |
673a394b | 1602 | mutex_unlock(&dev->struct_mutex); |
b8f9096d | 1603 | |
40e62d5d CW |
1604 | if (write_domain != 0) |
1605 | intel_fb_obj_invalidate(obj, write_origin(obj, write_domain)); | |
1606 | ||
f0cd5182 | 1607 | out_unpin: |
40e62d5d | 1608 | i915_gem_object_unpin_pages(obj); |
f0cd5182 CW |
1609 | out: |
1610 | i915_gem_object_put(obj); | |
40e62d5d | 1611 | return err; |
673a394b EA |
1612 | } |
1613 | ||
1614 | /** | |
1615 | * Called when user space has done writes to this buffer | |
14bb2c11 TU |
1616 | * @dev: drm device |
1617 | * @data: ioctl data blob | |
1618 | * @file: drm file | |
673a394b EA |
1619 | */ |
1620 | int | |
1621 | i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 1622 | struct drm_file *file) |
673a394b EA |
1623 | { |
1624 | struct drm_i915_gem_sw_finish *args = data; | |
05394f39 | 1625 | struct drm_i915_gem_object *obj; |
1d7cfea1 | 1626 | |
03ac0642 | 1627 | obj = i915_gem_object_lookup(file, args->handle); |
c21724cc CW |
1628 | if (!obj) |
1629 | return -ENOENT; | |
673a394b | 1630 | |
673a394b | 1631 | /* Pinned buffers may be scanout, so flush the cache */ |
5a97bcc6 | 1632 | i915_gem_object_flush_if_display(obj); |
f0cd5182 | 1633 | i915_gem_object_put(obj); |
5a97bcc6 CW |
1634 | |
1635 | return 0; | |
673a394b EA |
1636 | } |
1637 | ||
1638 | /** | |
14bb2c11 TU |
1639 | * i915_gem_mmap_ioctl - Maps the contents of an object, returning the address |
1640 | * it is mapped to. | |
1641 | * @dev: drm device | |
1642 | * @data: ioctl data blob | |
1643 | * @file: drm file | |
673a394b EA |
1644 | * |
1645 | * While the mapping holds a reference on the contents of the object, it doesn't | |
1646 | * imply a ref on the object itself. | |
34367381 DV |
1647 | * |
1648 | * IMPORTANT: | |
1649 | * | |
1650 | * DRM driver writers who look a this function as an example for how to do GEM | |
1651 | * mmap support, please don't implement mmap support like here. The modern way | |
1652 | * to implement DRM mmap support is with an mmap offset ioctl (like | |
1653 | * i915_gem_mmap_gtt) and then using the mmap syscall on the DRM fd directly. | |
1654 | * That way debug tooling like valgrind will understand what's going on, hiding | |
1655 | * the mmap call in a driver private ioctl will break that. The i915 driver only | |
1656 | * does cpu mmaps this way because we didn't know better. | |
673a394b EA |
1657 | */ |
1658 | int | |
1659 | i915_gem_mmap_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 1660 | struct drm_file *file) |
673a394b EA |
1661 | { |
1662 | struct drm_i915_gem_mmap *args = data; | |
03ac0642 | 1663 | struct drm_i915_gem_object *obj; |
673a394b EA |
1664 | unsigned long addr; |
1665 | ||
1816f923 AG |
1666 | if (args->flags & ~(I915_MMAP_WC)) |
1667 | return -EINVAL; | |
1668 | ||
568a58e5 | 1669 | if (args->flags & I915_MMAP_WC && !boot_cpu_has(X86_FEATURE_PAT)) |
1816f923 AG |
1670 | return -ENODEV; |
1671 | ||
03ac0642 CW |
1672 | obj = i915_gem_object_lookup(file, args->handle); |
1673 | if (!obj) | |
bf79cb91 | 1674 | return -ENOENT; |
673a394b | 1675 | |
1286ff73 DV |
1676 | /* prime objects have no backing filp to GEM mmap |
1677 | * pages from. | |
1678 | */ | |
03ac0642 | 1679 | if (!obj->base.filp) { |
f0cd5182 | 1680 | i915_gem_object_put(obj); |
1286ff73 DV |
1681 | return -EINVAL; |
1682 | } | |
1683 | ||
03ac0642 | 1684 | addr = vm_mmap(obj->base.filp, 0, args->size, |
673a394b EA |
1685 | PROT_READ | PROT_WRITE, MAP_SHARED, |
1686 | args->offset); | |
1816f923 AG |
1687 | if (args->flags & I915_MMAP_WC) { |
1688 | struct mm_struct *mm = current->mm; | |
1689 | struct vm_area_struct *vma; | |
1690 | ||
80a89a5e | 1691 | if (down_write_killable(&mm->mmap_sem)) { |
f0cd5182 | 1692 | i915_gem_object_put(obj); |
80a89a5e MH |
1693 | return -EINTR; |
1694 | } | |
1816f923 AG |
1695 | vma = find_vma(mm, addr); |
1696 | if (vma) | |
1697 | vma->vm_page_prot = | |
1698 | pgprot_writecombine(vm_get_page_prot(vma->vm_flags)); | |
1699 | else | |
1700 | addr = -ENOMEM; | |
1701 | up_write(&mm->mmap_sem); | |
aeecc969 CW |
1702 | |
1703 | /* This may race, but that's ok, it only gets set */ | |
50349247 | 1704 | WRITE_ONCE(obj->frontbuffer_ggtt_origin, ORIGIN_CPU); |
1816f923 | 1705 | } |
f0cd5182 | 1706 | i915_gem_object_put(obj); |
673a394b EA |
1707 | if (IS_ERR((void *)addr)) |
1708 | return addr; | |
1709 | ||
1710 | args->addr_ptr = (uint64_t) addr; | |
1711 | ||
1712 | return 0; | |
1713 | } | |
1714 | ||
03af84fe CW |
1715 | static unsigned int tile_row_pages(struct drm_i915_gem_object *obj) |
1716 | { | |
6649a0b6 | 1717 | return i915_gem_object_get_tile_row_size(obj) >> PAGE_SHIFT; |
03af84fe CW |
1718 | } |
1719 | ||
4cc69075 CW |
1720 | /** |
1721 | * i915_gem_mmap_gtt_version - report the current feature set for GTT mmaps | |
1722 | * | |
1723 | * A history of the GTT mmap interface: | |
1724 | * | |
1725 | * 0 - Everything had to fit into the GTT. Both parties of a memcpy had to | |
1726 | * aligned and suitable for fencing, and still fit into the available | |
1727 | * mappable space left by the pinned display objects. A classic problem | |
1728 | * we called the page-fault-of-doom where we would ping-pong between | |
1729 | * two objects that could not fit inside the GTT and so the memcpy | |
1730 | * would page one object in at the expense of the other between every | |
1731 | * single byte. | |
1732 | * | |
1733 | * 1 - Objects can be any size, and have any compatible fencing (X Y, or none | |
1734 | * as set via i915_gem_set_tiling() [DRM_I915_GEM_SET_TILING]). If the | |
1735 | * object is too large for the available space (or simply too large | |
1736 | * for the mappable aperture!), a view is created instead and faulted | |
1737 | * into userspace. (This view is aligned and sized appropriately for | |
1738 | * fenced access.) | |
1739 | * | |
1740 | * Restrictions: | |
1741 | * | |
1742 | * * snoopable objects cannot be accessed via the GTT. It can cause machine | |
1743 | * hangs on some architectures, corruption on others. An attempt to service | |
1744 | * a GTT page fault from a snoopable object will generate a SIGBUS. | |
1745 | * | |
1746 | * * the object must be able to fit into RAM (physical memory, though no | |
1747 | * limited to the mappable aperture). | |
1748 | * | |
1749 | * | |
1750 | * Caveats: | |
1751 | * | |
1752 | * * a new GTT page fault will synchronize rendering from the GPU and flush | |
1753 | * all data to system memory. Subsequent access will not be synchronized. | |
1754 | * | |
1755 | * * all mappings are revoked on runtime device suspend. | |
1756 | * | |
1757 | * * there are only 8, 16 or 32 fence registers to share between all users | |
1758 | * (older machines require fence register for display and blitter access | |
1759 | * as well). Contention of the fence registers will cause the previous users | |
1760 | * to be unmapped and any new access will generate new page faults. | |
1761 | * | |
1762 | * * running out of memory while servicing a fault may generate a SIGBUS, | |
1763 | * rather than the expected SIGSEGV. | |
1764 | */ | |
1765 | int i915_gem_mmap_gtt_version(void) | |
1766 | { | |
1767 | return 1; | |
1768 | } | |
1769 | ||
2d4281bb CW |
1770 | static inline struct i915_ggtt_view |
1771 | compute_partial_view(struct drm_i915_gem_object *obj, | |
2d4281bb CW |
1772 | pgoff_t page_offset, |
1773 | unsigned int chunk) | |
1774 | { | |
1775 | struct i915_ggtt_view view; | |
1776 | ||
1777 | if (i915_gem_object_is_tiled(obj)) | |
1778 | chunk = roundup(chunk, tile_row_pages(obj)); | |
1779 | ||
2d4281bb | 1780 | view.type = I915_GGTT_VIEW_PARTIAL; |
8bab1193 CW |
1781 | view.partial.offset = rounddown(page_offset, chunk); |
1782 | view.partial.size = | |
2d4281bb | 1783 | min_t(unsigned int, chunk, |
8bab1193 | 1784 | (obj->base.size >> PAGE_SHIFT) - view.partial.offset); |
2d4281bb CW |
1785 | |
1786 | /* If the partial covers the entire object, just create a normal VMA. */ | |
1787 | if (chunk >= obj->base.size >> PAGE_SHIFT) | |
1788 | view.type = I915_GGTT_VIEW_NORMAL; | |
1789 | ||
1790 | return view; | |
1791 | } | |
1792 | ||
de151cf6 JB |
1793 | /** |
1794 | * i915_gem_fault - fault a page into the GTT | |
d9072a3e | 1795 | * @vmf: fault info |
de151cf6 JB |
1796 | * |
1797 | * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped | |
1798 | * from userspace. The fault handler takes care of binding the object to | |
1799 | * the GTT (if needed), allocating and programming a fence register (again, | |
1800 | * only if needed based on whether the old reg is still valid or the object | |
1801 | * is tiled) and inserting a new PTE into the faulting process. | |
1802 | * | |
1803 | * Note that the faulting process may involve evicting existing objects | |
1804 | * from the GTT and/or fence registers to make room. So performance may | |
1805 | * suffer if the GTT working set is large or there are few fence registers | |
1806 | * left. | |
4cc69075 CW |
1807 | * |
1808 | * The current feature set supported by i915_gem_fault() and thus GTT mmaps | |
1809 | * is exposed via I915_PARAM_MMAP_GTT_VERSION (see i915_gem_mmap_gtt_version). | |
de151cf6 | 1810 | */ |
11bac800 | 1811 | int i915_gem_fault(struct vm_fault *vmf) |
de151cf6 | 1812 | { |
03af84fe | 1813 | #define MIN_CHUNK_PAGES ((1 << 20) >> PAGE_SHIFT) /* 1 MiB */ |
11bac800 | 1814 | struct vm_area_struct *area = vmf->vma; |
058d88c4 | 1815 | struct drm_i915_gem_object *obj = to_intel_bo(area->vm_private_data); |
05394f39 | 1816 | struct drm_device *dev = obj->base.dev; |
72e96d64 JL |
1817 | struct drm_i915_private *dev_priv = to_i915(dev); |
1818 | struct i915_ggtt *ggtt = &dev_priv->ggtt; | |
b8f9096d | 1819 | bool write = !!(vmf->flags & FAULT_FLAG_WRITE); |
058d88c4 | 1820 | struct i915_vma *vma; |
de151cf6 | 1821 | pgoff_t page_offset; |
82118877 | 1822 | unsigned int flags; |
b8f9096d | 1823 | int ret; |
f65c9168 | 1824 | |
de151cf6 | 1825 | /* We don't use vmf->pgoff since that has the fake offset */ |
1a29d85e | 1826 | page_offset = (vmf->address - area->vm_start) >> PAGE_SHIFT; |
de151cf6 | 1827 | |
db53a302 CW |
1828 | trace_i915_gem_object_fault(obj, page_offset, true, write); |
1829 | ||
6e4930f6 | 1830 | /* Try to flush the object off the GPU first without holding the lock. |
b8f9096d | 1831 | * Upon acquiring the lock, we will perform our sanity checks and then |
6e4930f6 CW |
1832 | * repeat the flush holding the lock in the normal manner to catch cases |
1833 | * where we are gazumped. | |
1834 | */ | |
e95433c7 CW |
1835 | ret = i915_gem_object_wait(obj, |
1836 | I915_WAIT_INTERRUPTIBLE, | |
1837 | MAX_SCHEDULE_TIMEOUT, | |
1838 | NULL); | |
6e4930f6 | 1839 | if (ret) |
b8f9096d CW |
1840 | goto err; |
1841 | ||
40e62d5d CW |
1842 | ret = i915_gem_object_pin_pages(obj); |
1843 | if (ret) | |
1844 | goto err; | |
1845 | ||
b8f9096d CW |
1846 | intel_runtime_pm_get(dev_priv); |
1847 | ||
1848 | ret = i915_mutex_lock_interruptible(dev); | |
1849 | if (ret) | |
1850 | goto err_rpm; | |
6e4930f6 | 1851 | |
eb119bd6 | 1852 | /* Access to snoopable pages through the GTT is incoherent. */ |
0031fb96 | 1853 | if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev_priv)) { |
ddeff6ee | 1854 | ret = -EFAULT; |
b8f9096d | 1855 | goto err_unlock; |
eb119bd6 CW |
1856 | } |
1857 | ||
82118877 CW |
1858 | /* If the object is smaller than a couple of partial vma, it is |
1859 | * not worth only creating a single partial vma - we may as well | |
1860 | * clear enough space for the full object. | |
1861 | */ | |
1862 | flags = PIN_MAPPABLE; | |
1863 | if (obj->base.size > 2 * MIN_CHUNK_PAGES << PAGE_SHIFT) | |
1864 | flags |= PIN_NONBLOCK | PIN_NONFAULT; | |
1865 | ||
a61007a8 | 1866 | /* Now pin it into the GTT as needed */ |
82118877 | 1867 | vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, flags); |
a61007a8 | 1868 | if (IS_ERR(vma)) { |
a61007a8 | 1869 | /* Use a partial view if it is bigger than available space */ |
2d4281bb | 1870 | struct i915_ggtt_view view = |
8201c1fa | 1871 | compute_partial_view(obj, page_offset, MIN_CHUNK_PAGES); |
aa136d9d | 1872 | |
50349247 CW |
1873 | /* Userspace is now writing through an untracked VMA, abandon |
1874 | * all hope that the hardware is able to track future writes. | |
1875 | */ | |
1876 | obj->frontbuffer_ggtt_origin = ORIGIN_CPU; | |
1877 | ||
a61007a8 CW |
1878 | vma = i915_gem_object_ggtt_pin(obj, &view, 0, 0, PIN_MAPPABLE); |
1879 | } | |
058d88c4 CW |
1880 | if (IS_ERR(vma)) { |
1881 | ret = PTR_ERR(vma); | |
b8f9096d | 1882 | goto err_unlock; |
058d88c4 | 1883 | } |
4a684a41 | 1884 | |
c9839303 CW |
1885 | ret = i915_gem_object_set_to_gtt_domain(obj, write); |
1886 | if (ret) | |
b8f9096d | 1887 | goto err_unpin; |
74898d7e | 1888 | |
49ef5294 | 1889 | ret = i915_vma_get_fence(vma); |
d9e86c0e | 1890 | if (ret) |
b8f9096d | 1891 | goto err_unpin; |
7d1c4804 | 1892 | |
275f039d | 1893 | /* Mark as being mmapped into userspace for later revocation */ |
9c870d03 | 1894 | assert_rpm_wakelock_held(dev_priv); |
275f039d CW |
1895 | if (list_empty(&obj->userfault_link)) |
1896 | list_add(&obj->userfault_link, &dev_priv->mm.userfault_list); | |
275f039d | 1897 | |
b90b91d8 | 1898 | /* Finally, remap it using the new GTT offset */ |
c58305af | 1899 | ret = remap_io_mapping(area, |
8bab1193 | 1900 | area->vm_start + (vma->ggtt_view.partial.offset << PAGE_SHIFT), |
c58305af CW |
1901 | (ggtt->mappable_base + vma->node.start) >> PAGE_SHIFT, |
1902 | min_t(u64, vma->size, area->vm_end - area->vm_start), | |
1903 | &ggtt->mappable); | |
a61007a8 | 1904 | |
b8f9096d | 1905 | err_unpin: |
058d88c4 | 1906 | __i915_vma_unpin(vma); |
b8f9096d | 1907 | err_unlock: |
de151cf6 | 1908 | mutex_unlock(&dev->struct_mutex); |
b8f9096d CW |
1909 | err_rpm: |
1910 | intel_runtime_pm_put(dev_priv); | |
40e62d5d | 1911 | i915_gem_object_unpin_pages(obj); |
b8f9096d | 1912 | err: |
de151cf6 | 1913 | switch (ret) { |
d9bc7e9f | 1914 | case -EIO: |
2232f031 DV |
1915 | /* |
1916 | * We eat errors when the gpu is terminally wedged to avoid | |
1917 | * userspace unduly crashing (gl has no provisions for mmaps to | |
1918 | * fail). But any other -EIO isn't ours (e.g. swap in failure) | |
1919 | * and so needs to be reported. | |
1920 | */ | |
1921 | if (!i915_terminally_wedged(&dev_priv->gpu_error)) { | |
f65c9168 PZ |
1922 | ret = VM_FAULT_SIGBUS; |
1923 | break; | |
1924 | } | |
045e769a | 1925 | case -EAGAIN: |
571c608d DV |
1926 | /* |
1927 | * EAGAIN means the gpu is hung and we'll wait for the error | |
1928 | * handler to reset everything when re-faulting in | |
1929 | * i915_mutex_lock_interruptible. | |
d9bc7e9f | 1930 | */ |
c715089f CW |
1931 | case 0: |
1932 | case -ERESTARTSYS: | |
bed636ab | 1933 | case -EINTR: |
e79e0fe3 DR |
1934 | case -EBUSY: |
1935 | /* | |
1936 | * EBUSY is ok: this just means that another thread | |
1937 | * already did the job. | |
1938 | */ | |
f65c9168 PZ |
1939 | ret = VM_FAULT_NOPAGE; |
1940 | break; | |
de151cf6 | 1941 | case -ENOMEM: |
f65c9168 PZ |
1942 | ret = VM_FAULT_OOM; |
1943 | break; | |
a7c2e1aa | 1944 | case -ENOSPC: |
45d67817 | 1945 | case -EFAULT: |
f65c9168 PZ |
1946 | ret = VM_FAULT_SIGBUS; |
1947 | break; | |
de151cf6 | 1948 | default: |
a7c2e1aa | 1949 | WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret); |
f65c9168 PZ |
1950 | ret = VM_FAULT_SIGBUS; |
1951 | break; | |
de151cf6 | 1952 | } |
f65c9168 | 1953 | return ret; |
de151cf6 JB |
1954 | } |
1955 | ||
901782b2 CW |
1956 | /** |
1957 | * i915_gem_release_mmap - remove physical page mappings | |
1958 | * @obj: obj in question | |
1959 | * | |
af901ca1 | 1960 | * Preserve the reservation of the mmapping with the DRM core code, but |
901782b2 CW |
1961 | * relinquish ownership of the pages back to the system. |
1962 | * | |
1963 | * It is vital that we remove the page mapping if we have mapped a tiled | |
1964 | * object through the GTT and then lose the fence register due to | |
1965 | * resource pressure. Similarly if the object has been moved out of the | |
1966 | * aperture, than pages mapped into userspace must be revoked. Removing the | |
1967 | * mapping will then trigger a page fault on the next user access, allowing | |
1968 | * fixup by i915_gem_fault(). | |
1969 | */ | |
d05ca301 | 1970 | void |
05394f39 | 1971 | i915_gem_release_mmap(struct drm_i915_gem_object *obj) |
901782b2 | 1972 | { |
275f039d | 1973 | struct drm_i915_private *i915 = to_i915(obj->base.dev); |
275f039d | 1974 | |
349f2ccf CW |
1975 | /* Serialisation between user GTT access and our code depends upon |
1976 | * revoking the CPU's PTE whilst the mutex is held. The next user | |
1977 | * pagefault then has to wait until we release the mutex. | |
9c870d03 CW |
1978 | * |
1979 | * Note that RPM complicates somewhat by adding an additional | |
1980 | * requirement that operations to the GGTT be made holding the RPM | |
1981 | * wakeref. | |
349f2ccf | 1982 | */ |
275f039d | 1983 | lockdep_assert_held(&i915->drm.struct_mutex); |
9c870d03 | 1984 | intel_runtime_pm_get(i915); |
349f2ccf | 1985 | |
3594a3e2 | 1986 | if (list_empty(&obj->userfault_link)) |
9c870d03 | 1987 | goto out; |
901782b2 | 1988 | |
3594a3e2 | 1989 | list_del_init(&obj->userfault_link); |
6796cb16 DH |
1990 | drm_vma_node_unmap(&obj->base.vma_node, |
1991 | obj->base.dev->anon_inode->i_mapping); | |
349f2ccf CW |
1992 | |
1993 | /* Ensure that the CPU's PTE are revoked and there are not outstanding | |
1994 | * memory transactions from userspace before we return. The TLB | |
1995 | * flushing implied above by changing the PTE above *should* be | |
1996 | * sufficient, an extra barrier here just provides us with a bit | |
1997 | * of paranoid documentation about our requirement to serialise | |
1998 | * memory writes before touching registers / GSM. | |
1999 | */ | |
2000 | wmb(); | |
9c870d03 CW |
2001 | |
2002 | out: | |
2003 | intel_runtime_pm_put(i915); | |
901782b2 CW |
2004 | } |
2005 | ||
7c108fd8 | 2006 | void i915_gem_runtime_suspend(struct drm_i915_private *dev_priv) |
eedd10f4 | 2007 | { |
3594a3e2 | 2008 | struct drm_i915_gem_object *obj, *on; |
7c108fd8 | 2009 | int i; |
eedd10f4 | 2010 | |
3594a3e2 CW |
2011 | /* |
2012 | * Only called during RPM suspend. All users of the userfault_list | |
2013 | * must be holding an RPM wakeref to ensure that this can not | |
2014 | * run concurrently with themselves (and use the struct_mutex for | |
2015 | * protection between themselves). | |
2016 | */ | |
275f039d | 2017 | |
3594a3e2 CW |
2018 | list_for_each_entry_safe(obj, on, |
2019 | &dev_priv->mm.userfault_list, userfault_link) { | |
2020 | list_del_init(&obj->userfault_link); | |
275f039d CW |
2021 | drm_vma_node_unmap(&obj->base.vma_node, |
2022 | obj->base.dev->anon_inode->i_mapping); | |
275f039d | 2023 | } |
7c108fd8 CW |
2024 | |
2025 | /* The fence will be lost when the device powers down. If any were | |
2026 | * in use by hardware (i.e. they are pinned), we should not be powering | |
2027 | * down! All other fences will be reacquired by the user upon waking. | |
2028 | */ | |
2029 | for (i = 0; i < dev_priv->num_fence_regs; i++) { | |
2030 | struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i]; | |
2031 | ||
e0ec3ec6 CW |
2032 | /* Ideally we want to assert that the fence register is not |
2033 | * live at this point (i.e. that no piece of code will be | |
2034 | * trying to write through fence + GTT, as that both violates | |
2035 | * our tracking of activity and associated locking/barriers, | |
2036 | * but also is illegal given that the hw is powered down). | |
2037 | * | |
2038 | * Previously we used reg->pin_count as a "liveness" indicator. | |
2039 | * That is not sufficient, and we need a more fine-grained | |
2040 | * tool if we want to have a sanity check here. | |
2041 | */ | |
7c108fd8 CW |
2042 | |
2043 | if (!reg->vma) | |
2044 | continue; | |
2045 | ||
2046 | GEM_BUG_ON(!list_empty(®->vma->obj->userfault_link)); | |
2047 | reg->dirty = true; | |
2048 | } | |
eedd10f4 CW |
2049 | } |
2050 | ||
d8cb5086 CW |
2051 | static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj) |
2052 | { | |
fac5e23e | 2053 | struct drm_i915_private *dev_priv = to_i915(obj->base.dev); |
f3f6184c | 2054 | int err; |
da494d7c | 2055 | |
f3f6184c | 2056 | err = drm_gem_create_mmap_offset(&obj->base); |
b42a13d9 | 2057 | if (likely(!err)) |
f3f6184c | 2058 | return 0; |
d8cb5086 | 2059 | |
b42a13d9 CW |
2060 | /* Attempt to reap some mmap space from dead objects */ |
2061 | do { | |
2062 | err = i915_gem_wait_for_idle(dev_priv, I915_WAIT_INTERRUPTIBLE); | |
2063 | if (err) | |
2064 | break; | |
f3f6184c | 2065 | |
b42a13d9 | 2066 | i915_gem_drain_freed_objects(dev_priv); |
f3f6184c | 2067 | err = drm_gem_create_mmap_offset(&obj->base); |
b42a13d9 CW |
2068 | if (!err) |
2069 | break; | |
2070 | ||
2071 | } while (flush_delayed_work(&dev_priv->gt.retire_work)); | |
da494d7c | 2072 | |
f3f6184c | 2073 | return err; |
d8cb5086 CW |
2074 | } |
2075 | ||
2076 | static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj) | |
2077 | { | |
d8cb5086 CW |
2078 | drm_gem_free_mmap_offset(&obj->base); |
2079 | } | |
2080 | ||
da6b51d0 | 2081 | int |
ff72145b DA |
2082 | i915_gem_mmap_gtt(struct drm_file *file, |
2083 | struct drm_device *dev, | |
da6b51d0 | 2084 | uint32_t handle, |
ff72145b | 2085 | uint64_t *offset) |
de151cf6 | 2086 | { |
05394f39 | 2087 | struct drm_i915_gem_object *obj; |
de151cf6 JB |
2088 | int ret; |
2089 | ||
03ac0642 | 2090 | obj = i915_gem_object_lookup(file, handle); |
f3f6184c CW |
2091 | if (!obj) |
2092 | return -ENOENT; | |
ab18282d | 2093 | |
d8cb5086 | 2094 | ret = i915_gem_object_create_mmap_offset(obj); |
f3f6184c CW |
2095 | if (ret == 0) |
2096 | *offset = drm_vma_node_offset_addr(&obj->base.vma_node); | |
de151cf6 | 2097 | |
f0cd5182 | 2098 | i915_gem_object_put(obj); |
1d7cfea1 | 2099 | return ret; |
de151cf6 JB |
2100 | } |
2101 | ||
ff72145b DA |
2102 | /** |
2103 | * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing | |
2104 | * @dev: DRM device | |
2105 | * @data: GTT mapping ioctl data | |
2106 | * @file: GEM object info | |
2107 | * | |
2108 | * Simply returns the fake offset to userspace so it can mmap it. | |
2109 | * The mmap call will end up in drm_gem_mmap(), which will set things | |
2110 | * up so we can get faults in the handler above. | |
2111 | * | |
2112 | * The fault handler will take care of binding the object into the GTT | |
2113 | * (since it may have been evicted to make room for something), allocating | |
2114 | * a fence register, and mapping the appropriate aperture address into | |
2115 | * userspace. | |
2116 | */ | |
2117 | int | |
2118 | i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data, | |
2119 | struct drm_file *file) | |
2120 | { | |
2121 | struct drm_i915_gem_mmap_gtt *args = data; | |
2122 | ||
da6b51d0 | 2123 | return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset); |
ff72145b DA |
2124 | } |
2125 | ||
225067ee DV |
2126 | /* Immediately discard the backing storage */ |
2127 | static void | |
2128 | i915_gem_object_truncate(struct drm_i915_gem_object *obj) | |
e5281ccd | 2129 | { |
4d6294bf | 2130 | i915_gem_object_free_mmap_offset(obj); |
1286ff73 | 2131 | |
4d6294bf CW |
2132 | if (obj->base.filp == NULL) |
2133 | return; | |
e5281ccd | 2134 | |
225067ee DV |
2135 | /* Our goal here is to return as much of the memory as |
2136 | * is possible back to the system as we are called from OOM. | |
2137 | * To do this we must instruct the shmfs to drop all of its | |
2138 | * backing pages, *now*. | |
2139 | */ | |
5537252b | 2140 | shmem_truncate_range(file_inode(obj->base.filp), 0, (loff_t)-1); |
a4f5ea64 | 2141 | obj->mm.madv = __I915_MADV_PURGED; |
4e5462ee | 2142 | obj->mm.pages = ERR_PTR(-EFAULT); |
225067ee | 2143 | } |
e5281ccd | 2144 | |
5537252b | 2145 | /* Try to discard unwanted pages */ |
03ac84f1 | 2146 | void __i915_gem_object_invalidate(struct drm_i915_gem_object *obj) |
225067ee | 2147 | { |
5537252b CW |
2148 | struct address_space *mapping; |
2149 | ||
1233e2db CW |
2150 | lockdep_assert_held(&obj->mm.lock); |
2151 | GEM_BUG_ON(obj->mm.pages); | |
2152 | ||
a4f5ea64 | 2153 | switch (obj->mm.madv) { |
5537252b CW |
2154 | case I915_MADV_DONTNEED: |
2155 | i915_gem_object_truncate(obj); | |
2156 | case __I915_MADV_PURGED: | |
2157 | return; | |
2158 | } | |
2159 | ||
2160 | if (obj->base.filp == NULL) | |
2161 | return; | |
2162 | ||
93c76a3d | 2163 | mapping = obj->base.filp->f_mapping, |
5537252b | 2164 | invalidate_mapping_pages(mapping, 0, (loff_t)-1); |
e5281ccd CW |
2165 | } |
2166 | ||
5cdf5881 | 2167 | static void |
03ac84f1 CW |
2168 | i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj, |
2169 | struct sg_table *pages) | |
673a394b | 2170 | { |
85d1225e DG |
2171 | struct sgt_iter sgt_iter; |
2172 | struct page *page; | |
1286ff73 | 2173 | |
e5facdf9 | 2174 | __i915_gem_object_release_shmem(obj, pages, true); |
673a394b | 2175 | |
03ac84f1 | 2176 | i915_gem_gtt_finish_pages(obj, pages); |
e2273302 | 2177 | |
6dacfd2f | 2178 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
03ac84f1 | 2179 | i915_gem_object_save_bit_17_swizzle(obj, pages); |
280b713b | 2180 | |
03ac84f1 | 2181 | for_each_sgt_page(page, sgt_iter, pages) { |
a4f5ea64 | 2182 | if (obj->mm.dirty) |
9da3da66 | 2183 | set_page_dirty(page); |
3ef94daa | 2184 | |
a4f5ea64 | 2185 | if (obj->mm.madv == I915_MADV_WILLNEED) |
9da3da66 | 2186 | mark_page_accessed(page); |
3ef94daa | 2187 | |
09cbfeaf | 2188 | put_page(page); |
3ef94daa | 2189 | } |
a4f5ea64 | 2190 | obj->mm.dirty = false; |
673a394b | 2191 | |
03ac84f1 CW |
2192 | sg_free_table(pages); |
2193 | kfree(pages); | |
37e680a1 | 2194 | } |
6c085a72 | 2195 | |
96d77634 CW |
2196 | static void __i915_gem_object_reset_page_iter(struct drm_i915_gem_object *obj) |
2197 | { | |
2198 | struct radix_tree_iter iter; | |
2199 | void **slot; | |
2200 | ||
a4f5ea64 CW |
2201 | radix_tree_for_each_slot(slot, &obj->mm.get_page.radix, &iter, 0) |
2202 | radix_tree_delete(&obj->mm.get_page.radix, iter.index); | |
96d77634 CW |
2203 | } |
2204 | ||
548625ee CW |
2205 | void __i915_gem_object_put_pages(struct drm_i915_gem_object *obj, |
2206 | enum i915_mm_subclass subclass) | |
37e680a1 | 2207 | { |
03ac84f1 | 2208 | struct sg_table *pages; |
37e680a1 | 2209 | |
a4f5ea64 | 2210 | if (i915_gem_object_has_pinned_pages(obj)) |
03ac84f1 | 2211 | return; |
a5570178 | 2212 | |
15717de2 | 2213 | GEM_BUG_ON(obj->bind_count); |
1233e2db CW |
2214 | if (!READ_ONCE(obj->mm.pages)) |
2215 | return; | |
2216 | ||
2217 | /* May be called by shrinker from within get_pages() (on another bo) */ | |
548625ee | 2218 | mutex_lock_nested(&obj->mm.lock, subclass); |
1233e2db CW |
2219 | if (unlikely(atomic_read(&obj->mm.pages_pin_count))) |
2220 | goto unlock; | |
3e123027 | 2221 | |
a2165e31 CW |
2222 | /* ->put_pages might need to allocate memory for the bit17 swizzle |
2223 | * array, hence protect them from being reaped by removing them from gtt | |
2224 | * lists early. */ | |
03ac84f1 CW |
2225 | pages = fetch_and_zero(&obj->mm.pages); |
2226 | GEM_BUG_ON(!pages); | |
a2165e31 | 2227 | |
a4f5ea64 | 2228 | if (obj->mm.mapping) { |
4b30cb23 CW |
2229 | void *ptr; |
2230 | ||
a4f5ea64 | 2231 | ptr = ptr_mask_bits(obj->mm.mapping); |
4b30cb23 CW |
2232 | if (is_vmalloc_addr(ptr)) |
2233 | vunmap(ptr); | |
fb8621d3 | 2234 | else |
4b30cb23 CW |
2235 | kunmap(kmap_to_page(ptr)); |
2236 | ||
a4f5ea64 | 2237 | obj->mm.mapping = NULL; |
0a798eb9 CW |
2238 | } |
2239 | ||
96d77634 CW |
2240 | __i915_gem_object_reset_page_iter(obj); |
2241 | ||
4e5462ee CW |
2242 | if (!IS_ERR(pages)) |
2243 | obj->ops->put_pages(obj, pages); | |
2244 | ||
1233e2db CW |
2245 | unlock: |
2246 | mutex_unlock(&obj->mm.lock); | |
6c085a72 CW |
2247 | } |
2248 | ||
935a2f77 | 2249 | static bool i915_sg_trim(struct sg_table *orig_st) |
0c40ce13 TU |
2250 | { |
2251 | struct sg_table new_st; | |
2252 | struct scatterlist *sg, *new_sg; | |
2253 | unsigned int i; | |
2254 | ||
2255 | if (orig_st->nents == orig_st->orig_nents) | |
935a2f77 | 2256 | return false; |
0c40ce13 | 2257 | |
8bfc478f | 2258 | if (sg_alloc_table(&new_st, orig_st->nents, GFP_KERNEL | __GFP_NOWARN)) |
935a2f77 | 2259 | return false; |
0c40ce13 TU |
2260 | |
2261 | new_sg = new_st.sgl; | |
2262 | for_each_sg(orig_st->sgl, sg, orig_st->nents, i) { | |
2263 | sg_set_page(new_sg, sg_page(sg), sg->length, 0); | |
2264 | /* called before being DMA mapped, no need to copy sg->dma_* */ | |
2265 | new_sg = sg_next(new_sg); | |
2266 | } | |
c2dc6cc9 | 2267 | GEM_BUG_ON(new_sg); /* Should walk exactly nents and hit the end */ |
0c40ce13 TU |
2268 | |
2269 | sg_free_table(orig_st); | |
2270 | ||
2271 | *orig_st = new_st; | |
935a2f77 | 2272 | return true; |
0c40ce13 TU |
2273 | } |
2274 | ||
03ac84f1 | 2275 | static struct sg_table * |
6c085a72 | 2276 | i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj) |
e5281ccd | 2277 | { |
fac5e23e | 2278 | struct drm_i915_private *dev_priv = to_i915(obj->base.dev); |
d766ef53 CW |
2279 | const unsigned long page_count = obj->base.size / PAGE_SIZE; |
2280 | unsigned long i; | |
e5281ccd | 2281 | struct address_space *mapping; |
9da3da66 CW |
2282 | struct sg_table *st; |
2283 | struct scatterlist *sg; | |
85d1225e | 2284 | struct sgt_iter sgt_iter; |
e5281ccd | 2285 | struct page *page; |
90797e6d | 2286 | unsigned long last_pfn = 0; /* suppress gcc warning */ |
4ff340f0 | 2287 | unsigned int max_segment; |
e2273302 | 2288 | int ret; |
6c085a72 | 2289 | gfp_t gfp; |
e5281ccd | 2290 | |
6c085a72 CW |
2291 | /* Assert that the object is not currently in any GPU domain. As it |
2292 | * wasn't in the GTT, there shouldn't be any way it could have been in | |
2293 | * a GPU cache | |
2294 | */ | |
03ac84f1 CW |
2295 | GEM_BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS); |
2296 | GEM_BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS); | |
6c085a72 | 2297 | |
7453c549 | 2298 | max_segment = swiotlb_max_segment(); |
871dfbd6 | 2299 | if (!max_segment) |
4ff340f0 | 2300 | max_segment = rounddown(UINT_MAX, PAGE_SIZE); |
871dfbd6 | 2301 | |
9da3da66 CW |
2302 | st = kmalloc(sizeof(*st), GFP_KERNEL); |
2303 | if (st == NULL) | |
03ac84f1 | 2304 | return ERR_PTR(-ENOMEM); |
9da3da66 | 2305 | |
d766ef53 | 2306 | rebuild_st: |
9da3da66 | 2307 | if (sg_alloc_table(st, page_count, GFP_KERNEL)) { |
9da3da66 | 2308 | kfree(st); |
03ac84f1 | 2309 | return ERR_PTR(-ENOMEM); |
9da3da66 | 2310 | } |
e5281ccd | 2311 | |
9da3da66 CW |
2312 | /* Get the list of pages out of our struct file. They'll be pinned |
2313 | * at this point until we release them. | |
2314 | * | |
2315 | * Fail silently without starting the shrinker | |
2316 | */ | |
93c76a3d | 2317 | mapping = obj->base.filp->f_mapping; |
c62d2555 | 2318 | gfp = mapping_gfp_constraint(mapping, ~(__GFP_IO | __GFP_RECLAIM)); |
d0164adc | 2319 | gfp |= __GFP_NORETRY | __GFP_NOWARN; |
90797e6d ID |
2320 | sg = st->sgl; |
2321 | st->nents = 0; | |
2322 | for (i = 0; i < page_count; i++) { | |
6c085a72 | 2323 | page = shmem_read_mapping_page_gfp(mapping, i, gfp); |
24f8e00a | 2324 | if (unlikely(IS_ERR(page))) { |
21ab4e74 CW |
2325 | i915_gem_shrink(dev_priv, |
2326 | page_count, | |
2327 | I915_SHRINK_BOUND | | |
2328 | I915_SHRINK_UNBOUND | | |
2329 | I915_SHRINK_PURGEABLE); | |
6c085a72 CW |
2330 | page = shmem_read_mapping_page_gfp(mapping, i, gfp); |
2331 | } | |
24f8e00a CW |
2332 | if (unlikely(IS_ERR(page))) { |
2333 | gfp_t reclaim; | |
2334 | ||
6c085a72 CW |
2335 | /* We've tried hard to allocate the memory by reaping |
2336 | * our own buffer, now let the real VM do its job and | |
2337 | * go down in flames if truly OOM. | |
24f8e00a CW |
2338 | * |
2339 | * However, since graphics tend to be disposable, | |
2340 | * defer the oom here by reporting the ENOMEM back | |
2341 | * to userspace. | |
6c085a72 | 2342 | */ |
24f8e00a CW |
2343 | reclaim = mapping_gfp_constraint(mapping, 0); |
2344 | reclaim |= __GFP_NORETRY; /* reclaim, but no oom */ | |
2345 | ||
40149f00 | 2346 | page = shmem_read_mapping_page_gfp(mapping, i, reclaim); |
e2273302 ID |
2347 | if (IS_ERR(page)) { |
2348 | ret = PTR_ERR(page); | |
b17993b7 | 2349 | goto err_sg; |
e2273302 | 2350 | } |
6c085a72 | 2351 | } |
871dfbd6 CW |
2352 | if (!i || |
2353 | sg->length >= max_segment || | |
2354 | page_to_pfn(page) != last_pfn + 1) { | |
90797e6d ID |
2355 | if (i) |
2356 | sg = sg_next(sg); | |
2357 | st->nents++; | |
2358 | sg_set_page(sg, page, PAGE_SIZE, 0); | |
2359 | } else { | |
2360 | sg->length += PAGE_SIZE; | |
2361 | } | |
2362 | last_pfn = page_to_pfn(page); | |
3bbbe706 DV |
2363 | |
2364 | /* Check that the i965g/gm workaround works. */ | |
2365 | WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL)); | |
e5281ccd | 2366 | } |
871dfbd6 | 2367 | if (sg) /* loop terminated early; short sg table */ |
426729dc | 2368 | sg_mark_end(sg); |
74ce6b6c | 2369 | |
0c40ce13 TU |
2370 | /* Trim unused sg entries to avoid wasting memory. */ |
2371 | i915_sg_trim(st); | |
2372 | ||
03ac84f1 | 2373 | ret = i915_gem_gtt_prepare_pages(obj, st); |
d766ef53 CW |
2374 | if (ret) { |
2375 | /* DMA remapping failed? One possible cause is that | |
2376 | * it could not reserve enough large entries, asking | |
2377 | * for PAGE_SIZE chunks instead may be helpful. | |
2378 | */ | |
2379 | if (max_segment > PAGE_SIZE) { | |
2380 | for_each_sgt_page(page, sgt_iter, st) | |
2381 | put_page(page); | |
2382 | sg_free_table(st); | |
2383 | ||
2384 | max_segment = PAGE_SIZE; | |
2385 | goto rebuild_st; | |
2386 | } else { | |
2387 | dev_warn(&dev_priv->drm.pdev->dev, | |
2388 | "Failed to DMA remap %lu pages\n", | |
2389 | page_count); | |
2390 | goto err_pages; | |
2391 | } | |
2392 | } | |
e2273302 | 2393 | |
6dacfd2f | 2394 | if (i915_gem_object_needs_bit17_swizzle(obj)) |
03ac84f1 | 2395 | i915_gem_object_do_bit_17_swizzle(obj, st); |
e5281ccd | 2396 | |
03ac84f1 | 2397 | return st; |
e5281ccd | 2398 | |
b17993b7 | 2399 | err_sg: |
90797e6d | 2400 | sg_mark_end(sg); |
b17993b7 | 2401 | err_pages: |
85d1225e DG |
2402 | for_each_sgt_page(page, sgt_iter, st) |
2403 | put_page(page); | |
9da3da66 CW |
2404 | sg_free_table(st); |
2405 | kfree(st); | |
0820baf3 CW |
2406 | |
2407 | /* shmemfs first checks if there is enough memory to allocate the page | |
2408 | * and reports ENOSPC should there be insufficient, along with the usual | |
2409 | * ENOMEM for a genuine allocation failure. | |
2410 | * | |
2411 | * We use ENOSPC in our driver to mean that we have run out of aperture | |
2412 | * space and so want to translate the error from shmemfs back to our | |
2413 | * usual understanding of ENOMEM. | |
2414 | */ | |
e2273302 ID |
2415 | if (ret == -ENOSPC) |
2416 | ret = -ENOMEM; | |
2417 | ||
03ac84f1 CW |
2418 | return ERR_PTR(ret); |
2419 | } | |
2420 | ||
2421 | void __i915_gem_object_set_pages(struct drm_i915_gem_object *obj, | |
2422 | struct sg_table *pages) | |
2423 | { | |
1233e2db | 2424 | lockdep_assert_held(&obj->mm.lock); |
03ac84f1 CW |
2425 | |
2426 | obj->mm.get_page.sg_pos = pages->sgl; | |
2427 | obj->mm.get_page.sg_idx = 0; | |
2428 | ||
2429 | obj->mm.pages = pages; | |
2c3a3f44 CW |
2430 | |
2431 | if (i915_gem_object_is_tiled(obj) && | |
2432 | to_i915(obj->base.dev)->quirks & QUIRK_PIN_SWIZZLED_PAGES) { | |
2433 | GEM_BUG_ON(obj->mm.quirked); | |
2434 | __i915_gem_object_pin_pages(obj); | |
2435 | obj->mm.quirked = true; | |
2436 | } | |
03ac84f1 CW |
2437 | } |
2438 | ||
2439 | static int ____i915_gem_object_get_pages(struct drm_i915_gem_object *obj) | |
2440 | { | |
2441 | struct sg_table *pages; | |
2442 | ||
2c3a3f44 CW |
2443 | GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj)); |
2444 | ||
03ac84f1 CW |
2445 | if (unlikely(obj->mm.madv != I915_MADV_WILLNEED)) { |
2446 | DRM_DEBUG("Attempting to obtain a purgeable object\n"); | |
2447 | return -EFAULT; | |
2448 | } | |
2449 | ||
2450 | pages = obj->ops->get_pages(obj); | |
2451 | if (unlikely(IS_ERR(pages))) | |
2452 | return PTR_ERR(pages); | |
2453 | ||
2454 | __i915_gem_object_set_pages(obj, pages); | |
2455 | return 0; | |
673a394b EA |
2456 | } |
2457 | ||
37e680a1 | 2458 | /* Ensure that the associated pages are gathered from the backing storage |
1233e2db | 2459 | * and pinned into our object. i915_gem_object_pin_pages() may be called |
37e680a1 | 2460 | * multiple times before they are released by a single call to |
1233e2db | 2461 | * i915_gem_object_unpin_pages() - once the pages are no longer referenced |
37e680a1 CW |
2462 | * either as a result of memory pressure (reaping pages under the shrinker) |
2463 | * or as the object is itself released. | |
2464 | */ | |
a4f5ea64 | 2465 | int __i915_gem_object_get_pages(struct drm_i915_gem_object *obj) |
37e680a1 | 2466 | { |
03ac84f1 | 2467 | int err; |
37e680a1 | 2468 | |
1233e2db CW |
2469 | err = mutex_lock_interruptible(&obj->mm.lock); |
2470 | if (err) | |
2471 | return err; | |
4c7d62c6 | 2472 | |
4e5462ee | 2473 | if (unlikely(IS_ERR_OR_NULL(obj->mm.pages))) { |
2c3a3f44 CW |
2474 | err = ____i915_gem_object_get_pages(obj); |
2475 | if (err) | |
2476 | goto unlock; | |
37e680a1 | 2477 | |
2c3a3f44 CW |
2478 | smp_mb__before_atomic(); |
2479 | } | |
2480 | atomic_inc(&obj->mm.pages_pin_count); | |
ee286370 | 2481 | |
1233e2db CW |
2482 | unlock: |
2483 | mutex_unlock(&obj->mm.lock); | |
03ac84f1 | 2484 | return err; |
673a394b EA |
2485 | } |
2486 | ||
dd6034c6 | 2487 | /* The 'mapping' part of i915_gem_object_pin_map() below */ |
d31d7cb1 CW |
2488 | static void *i915_gem_object_map(const struct drm_i915_gem_object *obj, |
2489 | enum i915_map_type type) | |
dd6034c6 DG |
2490 | { |
2491 | unsigned long n_pages = obj->base.size >> PAGE_SHIFT; | |
a4f5ea64 | 2492 | struct sg_table *sgt = obj->mm.pages; |
85d1225e DG |
2493 | struct sgt_iter sgt_iter; |
2494 | struct page *page; | |
b338fa47 DG |
2495 | struct page *stack_pages[32]; |
2496 | struct page **pages = stack_pages; | |
dd6034c6 | 2497 | unsigned long i = 0; |
d31d7cb1 | 2498 | pgprot_t pgprot; |
dd6034c6 DG |
2499 | void *addr; |
2500 | ||
2501 | /* A single page can always be kmapped */ | |
d31d7cb1 | 2502 | if (n_pages == 1 && type == I915_MAP_WB) |
dd6034c6 DG |
2503 | return kmap(sg_page(sgt->sgl)); |
2504 | ||
b338fa47 DG |
2505 | if (n_pages > ARRAY_SIZE(stack_pages)) { |
2506 | /* Too big for stack -- allocate temporary array instead */ | |
2507 | pages = drm_malloc_gfp(n_pages, sizeof(*pages), GFP_TEMPORARY); | |
2508 | if (!pages) | |
2509 | return NULL; | |
2510 | } | |
dd6034c6 | 2511 | |
85d1225e DG |
2512 | for_each_sgt_page(page, sgt_iter, sgt) |
2513 | pages[i++] = page; | |
dd6034c6 DG |
2514 | |
2515 | /* Check that we have the expected number of pages */ | |
2516 | GEM_BUG_ON(i != n_pages); | |
2517 | ||
d31d7cb1 CW |
2518 | switch (type) { |
2519 | case I915_MAP_WB: | |
2520 | pgprot = PAGE_KERNEL; | |
2521 | break; | |
2522 | case I915_MAP_WC: | |
2523 | pgprot = pgprot_writecombine(PAGE_KERNEL_IO); | |
2524 | break; | |
2525 | } | |
2526 | addr = vmap(pages, n_pages, 0, pgprot); | |
dd6034c6 | 2527 | |
b338fa47 DG |
2528 | if (pages != stack_pages) |
2529 | drm_free_large(pages); | |
dd6034c6 DG |
2530 | |
2531 | return addr; | |
2532 | } | |
2533 | ||
2534 | /* get, pin, and map the pages of the object into kernel space */ | |
d31d7cb1 CW |
2535 | void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj, |
2536 | enum i915_map_type type) | |
0a798eb9 | 2537 | { |
d31d7cb1 CW |
2538 | enum i915_map_type has_type; |
2539 | bool pinned; | |
2540 | void *ptr; | |
0a798eb9 CW |
2541 | int ret; |
2542 | ||
d31d7cb1 | 2543 | GEM_BUG_ON(!i915_gem_object_has_struct_page(obj)); |
0a798eb9 | 2544 | |
1233e2db | 2545 | ret = mutex_lock_interruptible(&obj->mm.lock); |
0a798eb9 CW |
2546 | if (ret) |
2547 | return ERR_PTR(ret); | |
2548 | ||
1233e2db CW |
2549 | pinned = true; |
2550 | if (!atomic_inc_not_zero(&obj->mm.pages_pin_count)) { | |
4e5462ee | 2551 | if (unlikely(IS_ERR_OR_NULL(obj->mm.pages))) { |
2c3a3f44 CW |
2552 | ret = ____i915_gem_object_get_pages(obj); |
2553 | if (ret) | |
2554 | goto err_unlock; | |
1233e2db | 2555 | |
2c3a3f44 CW |
2556 | smp_mb__before_atomic(); |
2557 | } | |
2558 | atomic_inc(&obj->mm.pages_pin_count); | |
1233e2db CW |
2559 | pinned = false; |
2560 | } | |
2561 | GEM_BUG_ON(!obj->mm.pages); | |
0a798eb9 | 2562 | |
a4f5ea64 | 2563 | ptr = ptr_unpack_bits(obj->mm.mapping, has_type); |
d31d7cb1 CW |
2564 | if (ptr && has_type != type) { |
2565 | if (pinned) { | |
2566 | ret = -EBUSY; | |
1233e2db | 2567 | goto err_unpin; |
0a798eb9 | 2568 | } |
d31d7cb1 CW |
2569 | |
2570 | if (is_vmalloc_addr(ptr)) | |
2571 | vunmap(ptr); | |
2572 | else | |
2573 | kunmap(kmap_to_page(ptr)); | |
2574 | ||
a4f5ea64 | 2575 | ptr = obj->mm.mapping = NULL; |
0a798eb9 CW |
2576 | } |
2577 | ||
d31d7cb1 CW |
2578 | if (!ptr) { |
2579 | ptr = i915_gem_object_map(obj, type); | |
2580 | if (!ptr) { | |
2581 | ret = -ENOMEM; | |
1233e2db | 2582 | goto err_unpin; |
d31d7cb1 CW |
2583 | } |
2584 | ||
a4f5ea64 | 2585 | obj->mm.mapping = ptr_pack_bits(ptr, type); |
d31d7cb1 CW |
2586 | } |
2587 | ||
1233e2db CW |
2588 | out_unlock: |
2589 | mutex_unlock(&obj->mm.lock); | |
d31d7cb1 CW |
2590 | return ptr; |
2591 | ||
1233e2db CW |
2592 | err_unpin: |
2593 | atomic_dec(&obj->mm.pages_pin_count); | |
2594 | err_unlock: | |
2595 | ptr = ERR_PTR(ret); | |
2596 | goto out_unlock; | |
0a798eb9 CW |
2597 | } |
2598 | ||
7c55e2c5 CW |
2599 | static int |
2600 | i915_gem_object_pwrite_gtt(struct drm_i915_gem_object *obj, | |
2601 | const struct drm_i915_gem_pwrite *arg) | |
2602 | { | |
2603 | struct address_space *mapping = obj->base.filp->f_mapping; | |
2604 | char __user *user_data = u64_to_user_ptr(arg->data_ptr); | |
2605 | u64 remain, offset; | |
2606 | unsigned int pg; | |
2607 | ||
2608 | /* Before we instantiate/pin the backing store for our use, we | |
2609 | * can prepopulate the shmemfs filp efficiently using a write into | |
2610 | * the pagecache. We avoid the penalty of instantiating all the | |
2611 | * pages, important if the user is just writing to a few and never | |
2612 | * uses the object on the GPU, and using a direct write into shmemfs | |
2613 | * allows it to avoid the cost of retrieving a page (either swapin | |
2614 | * or clearing-before-use) before it is overwritten. | |
2615 | */ | |
2616 | if (READ_ONCE(obj->mm.pages)) | |
2617 | return -ENODEV; | |
2618 | ||
2619 | /* Before the pages are instantiated the object is treated as being | |
2620 | * in the CPU domain. The pages will be clflushed as required before | |
2621 | * use, and we can freely write into the pages directly. If userspace | |
2622 | * races pwrite with any other operation; corruption will ensue - | |
2623 | * that is userspace's prerogative! | |
2624 | */ | |
2625 | ||
2626 | remain = arg->size; | |
2627 | offset = arg->offset; | |
2628 | pg = offset_in_page(offset); | |
2629 | ||
2630 | do { | |
2631 | unsigned int len, unwritten; | |
2632 | struct page *page; | |
2633 | void *data, *vaddr; | |
2634 | int err; | |
2635 | ||
2636 | len = PAGE_SIZE - pg; | |
2637 | if (len > remain) | |
2638 | len = remain; | |
2639 | ||
2640 | err = pagecache_write_begin(obj->base.filp, mapping, | |
2641 | offset, len, 0, | |
2642 | &page, &data); | |
2643 | if (err < 0) | |
2644 | return err; | |
2645 | ||
2646 | vaddr = kmap(page); | |
2647 | unwritten = copy_from_user(vaddr + pg, user_data, len); | |
2648 | kunmap(page); | |
2649 | ||
2650 | err = pagecache_write_end(obj->base.filp, mapping, | |
2651 | offset, len, len - unwritten, | |
2652 | page, data); | |
2653 | if (err < 0) | |
2654 | return err; | |
2655 | ||
2656 | if (unwritten) | |
2657 | return -EFAULT; | |
2658 | ||
2659 | remain -= len; | |
2660 | user_data += len; | |
2661 | offset += len; | |
2662 | pg = 0; | |
2663 | } while (remain); | |
2664 | ||
2665 | return 0; | |
2666 | } | |
2667 | ||
6095868a | 2668 | static bool ban_context(const struct i915_gem_context *ctx) |
be62acb4 | 2669 | { |
6095868a CW |
2670 | return (i915_gem_context_is_bannable(ctx) && |
2671 | ctx->ban_score >= CONTEXT_SCORE_BAN_THRESHOLD); | |
be62acb4 MK |
2672 | } |
2673 | ||
e5e1fc47 | 2674 | static void i915_gem_context_mark_guilty(struct i915_gem_context *ctx) |
aa60c664 | 2675 | { |
bc1d53c6 | 2676 | ctx->guilty_count++; |
6095868a CW |
2677 | ctx->ban_score += CONTEXT_SCORE_GUILTY; |
2678 | if (ban_context(ctx)) | |
2679 | i915_gem_context_set_banned(ctx); | |
b083a087 MK |
2680 | |
2681 | DRM_DEBUG_DRIVER("context %s marked guilty (score %d) banned? %s\n", | |
bc1d53c6 | 2682 | ctx->name, ctx->ban_score, |
6095868a | 2683 | yesno(i915_gem_context_is_banned(ctx))); |
b083a087 | 2684 | |
6095868a | 2685 | if (!i915_gem_context_is_banned(ctx) || IS_ERR_OR_NULL(ctx->file_priv)) |
b083a087 MK |
2686 | return; |
2687 | ||
d9e9da64 CW |
2688 | ctx->file_priv->context_bans++; |
2689 | DRM_DEBUG_DRIVER("client %s has had %d context banned\n", | |
2690 | ctx->name, ctx->file_priv->context_bans); | |
e5e1fc47 MK |
2691 | } |
2692 | ||
2693 | static void i915_gem_context_mark_innocent(struct i915_gem_context *ctx) | |
2694 | { | |
bc1d53c6 | 2695 | ctx->active_count++; |
aa60c664 MK |
2696 | } |
2697 | ||
8d9fc7fd | 2698 | struct drm_i915_gem_request * |
0bc40be8 | 2699 | i915_gem_find_active_request(struct intel_engine_cs *engine) |
9375e446 | 2700 | { |
754c9fd5 CW |
2701 | struct drm_i915_gem_request *request, *active = NULL; |
2702 | unsigned long flags; | |
4db080f9 | 2703 | |
f69a02c9 CW |
2704 | /* We are called by the error capture and reset at a random |
2705 | * point in time. In particular, note that neither is crucially | |
2706 | * ordered with an interrupt. After a hang, the GPU is dead and we | |
2707 | * assume that no more writes can happen (we waited long enough for | |
2708 | * all writes that were in transaction to be flushed) - adding an | |
2709 | * extra delay for a recent interrupt is pointless. Hence, we do | |
2710 | * not need an engine->irq_seqno_barrier() before the seqno reads. | |
2711 | */ | |
754c9fd5 | 2712 | spin_lock_irqsave(&engine->timeline->lock, flags); |
73cb9701 | 2713 | list_for_each_entry(request, &engine->timeline->requests, link) { |
754c9fd5 CW |
2714 | if (__i915_gem_request_completed(request, |
2715 | request->global_seqno)) | |
4db080f9 | 2716 | continue; |
aa60c664 | 2717 | |
36193acd | 2718 | GEM_BUG_ON(request->engine != engine); |
c00122f3 CW |
2719 | GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, |
2720 | &request->fence.flags)); | |
754c9fd5 CW |
2721 | |
2722 | active = request; | |
2723 | break; | |
4db080f9 | 2724 | } |
754c9fd5 | 2725 | spin_unlock_irqrestore(&engine->timeline->lock, flags); |
b6b0fac0 | 2726 | |
754c9fd5 | 2727 | return active; |
b6b0fac0 MK |
2728 | } |
2729 | ||
bf2f0436 MK |
2730 | static bool engine_stalled(struct intel_engine_cs *engine) |
2731 | { | |
2732 | if (!engine->hangcheck.stalled) | |
2733 | return false; | |
2734 | ||
2735 | /* Check for possible seqno movement after hang declaration */ | |
2736 | if (engine->hangcheck.seqno != intel_engine_get_seqno(engine)) { | |
2737 | DRM_DEBUG_DRIVER("%s pardoned\n", engine->name); | |
2738 | return false; | |
2739 | } | |
2740 | ||
2741 | return true; | |
2742 | } | |
2743 | ||
0e178aef | 2744 | int i915_gem_reset_prepare(struct drm_i915_private *dev_priv) |
4c965543 CW |
2745 | { |
2746 | struct intel_engine_cs *engine; | |
2747 | enum intel_engine_id id; | |
0e178aef | 2748 | int err = 0; |
4c965543 CW |
2749 | |
2750 | /* Ensure irq handler finishes, and not run again. */ | |
0e178aef CW |
2751 | for_each_engine(engine, dev_priv, id) { |
2752 | struct drm_i915_gem_request *request; | |
2753 | ||
fe3288b5 CW |
2754 | /* Prevent the signaler thread from updating the request |
2755 | * state (by calling dma_fence_signal) as we are processing | |
2756 | * the reset. The write from the GPU of the seqno is | |
2757 | * asynchronous and the signaler thread may see a different | |
2758 | * value to us and declare the request complete, even though | |
2759 | * the reset routine have picked that request as the active | |
2760 | * (incomplete) request. This conflict is not handled | |
2761 | * gracefully! | |
2762 | */ | |
2763 | kthread_park(engine->breadcrumbs.signaler); | |
2764 | ||
1f7b847d CW |
2765 | /* Prevent request submission to the hardware until we have |
2766 | * completed the reset in i915_gem_reset_finish(). If a request | |
2767 | * is completed by one engine, it may then queue a request | |
2768 | * to a second via its engine->irq_tasklet *just* as we are | |
2769 | * calling engine->init_hw() and also writing the ELSP. | |
2770 | * Turning off the engine->irq_tasklet until the reset is over | |
2771 | * prevents the race. | |
2772 | */ | |
4c965543 | 2773 | tasklet_kill(&engine->irq_tasklet); |
1d309634 | 2774 | tasklet_disable(&engine->irq_tasklet); |
4c965543 | 2775 | |
8c12d121 CW |
2776 | if (engine->irq_seqno_barrier) |
2777 | engine->irq_seqno_barrier(engine); | |
2778 | ||
0e178aef CW |
2779 | if (engine_stalled(engine)) { |
2780 | request = i915_gem_find_active_request(engine); | |
2781 | if (request && request->fence.error == -EIO) | |
2782 | err = -EIO; /* Previous reset failed! */ | |
2783 | } | |
2784 | } | |
2785 | ||
4c965543 | 2786 | i915_gem_revoke_fences(dev_priv); |
0e178aef CW |
2787 | |
2788 | return err; | |
4c965543 CW |
2789 | } |
2790 | ||
36193acd | 2791 | static void skip_request(struct drm_i915_gem_request *request) |
821ed7df CW |
2792 | { |
2793 | void *vaddr = request->ring->vaddr; | |
2794 | u32 head; | |
2795 | ||
2796 | /* As this request likely depends on state from the lost | |
2797 | * context, clear out all the user operations leaving the | |
2798 | * breadcrumb at the end (so we get the fence notifications). | |
2799 | */ | |
2800 | head = request->head; | |
2801 | if (request->postfix < head) { | |
2802 | memset(vaddr + head, 0, request->ring->size - head); | |
2803 | head = 0; | |
2804 | } | |
2805 | memset(vaddr + head, 0, request->postfix - head); | |
c0d5f32c CW |
2806 | |
2807 | dma_fence_set_error(&request->fence, -EIO); | |
821ed7df CW |
2808 | } |
2809 | ||
36193acd MK |
2810 | static void engine_skip_context(struct drm_i915_gem_request *request) |
2811 | { | |
2812 | struct intel_engine_cs *engine = request->engine; | |
2813 | struct i915_gem_context *hung_ctx = request->ctx; | |
2814 | struct intel_timeline *timeline; | |
2815 | unsigned long flags; | |
2816 | ||
2817 | timeline = i915_gem_context_lookup_timeline(hung_ctx, engine); | |
2818 | ||
2819 | spin_lock_irqsave(&engine->timeline->lock, flags); | |
2820 | spin_lock(&timeline->lock); | |
2821 | ||
2822 | list_for_each_entry_continue(request, &engine->timeline->requests, link) | |
2823 | if (request->ctx == hung_ctx) | |
2824 | skip_request(request); | |
2825 | ||
2826 | list_for_each_entry(request, &timeline->requests, link) | |
2827 | skip_request(request); | |
2828 | ||
2829 | spin_unlock(&timeline->lock); | |
2830 | spin_unlock_irqrestore(&engine->timeline->lock, flags); | |
2831 | } | |
2832 | ||
61da5362 MK |
2833 | /* Returns true if the request was guilty of hang */ |
2834 | static bool i915_gem_reset_request(struct drm_i915_gem_request *request) | |
2835 | { | |
2836 | /* Read once and return the resolution */ | |
2837 | const bool guilty = engine_stalled(request->engine); | |
2838 | ||
71895a08 MK |
2839 | /* The guilty request will get skipped on a hung engine. |
2840 | * | |
2841 | * Users of client default contexts do not rely on logical | |
2842 | * state preserved between batches so it is safe to execute | |
2843 | * queued requests following the hang. Non default contexts | |
2844 | * rely on preserved state, so skipping a batch loses the | |
2845 | * evolution of the state and it needs to be considered corrupted. | |
2846 | * Executing more queued batches on top of corrupted state is | |
2847 | * risky. But we take the risk by trying to advance through | |
2848 | * the queued requests in order to make the client behaviour | |
2849 | * more predictable around resets, by not throwing away random | |
2850 | * amount of batches it has prepared for execution. Sophisticated | |
2851 | * clients can use gem_reset_stats_ioctl and dma fence status | |
2852 | * (exported via sync_file info ioctl on explicit fences) to observe | |
2853 | * when it loses the context state and should rebuild accordingly. | |
2854 | * | |
2855 | * The context ban, and ultimately the client ban, mechanism are safety | |
2856 | * valves if client submission ends up resulting in nothing more than | |
2857 | * subsequent hangs. | |
2858 | */ | |
2859 | ||
61da5362 MK |
2860 | if (guilty) { |
2861 | i915_gem_context_mark_guilty(request->ctx); | |
2862 | skip_request(request); | |
2863 | } else { | |
2864 | i915_gem_context_mark_innocent(request->ctx); | |
2865 | dma_fence_set_error(&request->fence, -EAGAIN); | |
2866 | } | |
2867 | ||
2868 | return guilty; | |
2869 | } | |
2870 | ||
821ed7df | 2871 | static void i915_gem_reset_engine(struct intel_engine_cs *engine) |
b6b0fac0 MK |
2872 | { |
2873 | struct drm_i915_gem_request *request; | |
b6b0fac0 | 2874 | |
0bc40be8 | 2875 | request = i915_gem_find_active_request(engine); |
c0dcb203 CW |
2876 | if (request && i915_gem_reset_request(request)) { |
2877 | DRM_DEBUG_DRIVER("resetting %s to restart from tail of request 0x%x\n", | |
2878 | engine->name, request->global_seqno); | |
821ed7df | 2879 | |
c0dcb203 CW |
2880 | /* If this context is now banned, skip all pending requests. */ |
2881 | if (i915_gem_context_is_banned(request->ctx)) | |
2882 | engine_skip_context(request); | |
2883 | } | |
821ed7df CW |
2884 | |
2885 | /* Setup the CS to resume from the breadcrumb of the hung request */ | |
2886 | engine->reset_hw(engine, request); | |
4db080f9 | 2887 | } |
aa60c664 | 2888 | |
d8027093 | 2889 | void i915_gem_reset(struct drm_i915_private *dev_priv) |
4db080f9 | 2890 | { |
821ed7df | 2891 | struct intel_engine_cs *engine; |
3b3f1650 | 2892 | enum intel_engine_id id; |
608c1a52 | 2893 | |
4c7d62c6 CW |
2894 | lockdep_assert_held(&dev_priv->drm.struct_mutex); |
2895 | ||
821ed7df CW |
2896 | i915_gem_retire_requests(dev_priv); |
2897 | ||
2ae55738 CW |
2898 | for_each_engine(engine, dev_priv, id) { |
2899 | struct i915_gem_context *ctx; | |
2900 | ||
821ed7df | 2901 | i915_gem_reset_engine(engine); |
2ae55738 CW |
2902 | ctx = fetch_and_zero(&engine->last_retired_context); |
2903 | if (ctx) | |
2904 | engine->context_unpin(engine, ctx); | |
2905 | } | |
821ed7df | 2906 | |
4362f4f6 | 2907 | i915_gem_restore_fences(dev_priv); |
f2a91d1a CW |
2908 | |
2909 | if (dev_priv->gt.awake) { | |
2910 | intel_sanitize_gt_powersave(dev_priv); | |
2911 | intel_enable_gt_powersave(dev_priv); | |
2912 | if (INTEL_GEN(dev_priv) >= 6) | |
2913 | gen6_rps_busy(dev_priv); | |
2914 | } | |
821ed7df CW |
2915 | } |
2916 | ||
d8027093 CW |
2917 | void i915_gem_reset_finish(struct drm_i915_private *dev_priv) |
2918 | { | |
1f7b847d CW |
2919 | struct intel_engine_cs *engine; |
2920 | enum intel_engine_id id; | |
2921 | ||
d8027093 | 2922 | lockdep_assert_held(&dev_priv->drm.struct_mutex); |
1f7b847d | 2923 | |
fe3288b5 | 2924 | for_each_engine(engine, dev_priv, id) { |
1f7b847d | 2925 | tasklet_enable(&engine->irq_tasklet); |
fe3288b5 CW |
2926 | kthread_unpark(engine->breadcrumbs.signaler); |
2927 | } | |
d8027093 CW |
2928 | } |
2929 | ||
821ed7df CW |
2930 | static void nop_submit_request(struct drm_i915_gem_request *request) |
2931 | { | |
3cd9442f | 2932 | dma_fence_set_error(&request->fence, -EIO); |
3dcf93f7 CW |
2933 | i915_gem_request_submit(request); |
2934 | intel_engine_init_global_seqno(request->engine, request->global_seqno); | |
821ed7df CW |
2935 | } |
2936 | ||
2a20d6f8 | 2937 | static void engine_set_wedged(struct intel_engine_cs *engine) |
821ed7df | 2938 | { |
3cd9442f CW |
2939 | struct drm_i915_gem_request *request; |
2940 | unsigned long flags; | |
2941 | ||
20e4933c CW |
2942 | /* We need to be sure that no thread is running the old callback as |
2943 | * we install the nop handler (otherwise we would submit a request | |
2944 | * to hardware that will never complete). In order to prevent this | |
2945 | * race, we wait until the machine is idle before making the swap | |
2946 | * (using stop_machine()). | |
2947 | */ | |
821ed7df | 2948 | engine->submit_request = nop_submit_request; |
70c2a24d | 2949 | |
3cd9442f CW |
2950 | /* Mark all executing requests as skipped */ |
2951 | spin_lock_irqsave(&engine->timeline->lock, flags); | |
2952 | list_for_each_entry(request, &engine->timeline->requests, link) | |
2953 | dma_fence_set_error(&request->fence, -EIO); | |
2954 | spin_unlock_irqrestore(&engine->timeline->lock, flags); | |
2955 | ||
c4b0930b CW |
2956 | /* Mark all pending requests as complete so that any concurrent |
2957 | * (lockless) lookup doesn't try and wait upon the request as we | |
2958 | * reset it. | |
2959 | */ | |
73cb9701 | 2960 | intel_engine_init_global_seqno(engine, |
cb399eab | 2961 | intel_engine_last_submit(engine)); |
c4b0930b | 2962 | |
dcb4c12a OM |
2963 | /* |
2964 | * Clear the execlists queue up before freeing the requests, as those | |
2965 | * are the ones that keep the context and ringbuffer backing objects | |
2966 | * pinned in place. | |
2967 | */ | |
dcb4c12a | 2968 | |
7de1691a | 2969 | if (i915.enable_execlists) { |
663f71e7 CW |
2970 | unsigned long flags; |
2971 | ||
2972 | spin_lock_irqsave(&engine->timeline->lock, flags); | |
2973 | ||
70c2a24d CW |
2974 | i915_gem_request_put(engine->execlist_port[0].request); |
2975 | i915_gem_request_put(engine->execlist_port[1].request); | |
2976 | memset(engine->execlist_port, 0, sizeof(engine->execlist_port)); | |
20311bd3 CW |
2977 | engine->execlist_queue = RB_ROOT; |
2978 | engine->execlist_first = NULL; | |
663f71e7 CW |
2979 | |
2980 | spin_unlock_irqrestore(&engine->timeline->lock, flags); | |
dcb4c12a | 2981 | } |
673a394b EA |
2982 | } |
2983 | ||
20e4933c | 2984 | static int __i915_gem_set_wedged_BKL(void *data) |
673a394b | 2985 | { |
20e4933c | 2986 | struct drm_i915_private *i915 = data; |
e2f80391 | 2987 | struct intel_engine_cs *engine; |
3b3f1650 | 2988 | enum intel_engine_id id; |
673a394b | 2989 | |
20e4933c | 2990 | for_each_engine(engine, i915, id) |
2a20d6f8 | 2991 | engine_set_wedged(engine); |
20e4933c CW |
2992 | |
2993 | return 0; | |
2994 | } | |
2995 | ||
2996 | void i915_gem_set_wedged(struct drm_i915_private *dev_priv) | |
2997 | { | |
821ed7df CW |
2998 | lockdep_assert_held(&dev_priv->drm.struct_mutex); |
2999 | set_bit(I915_WEDGED, &dev_priv->gpu_error.flags); | |
4db080f9 | 3000 | |
2c170af7 CW |
3001 | /* Retire completed requests first so the list of inflight/incomplete |
3002 | * requests is accurate and we don't try and mark successful requests | |
3003 | * as in error during __i915_gem_set_wedged_BKL(). | |
3004 | */ | |
3005 | i915_gem_retire_requests(dev_priv); | |
3006 | ||
20e4933c | 3007 | stop_machine(__i915_gem_set_wedged_BKL, dev_priv, NULL); |
dfaae392 | 3008 | |
20e4933c | 3009 | i915_gem_context_lost(dev_priv); |
20e4933c CW |
3010 | |
3011 | mod_delayed_work(dev_priv->wq, &dev_priv->gt.idle_work, 0); | |
673a394b EA |
3012 | } |
3013 | ||
2e8f9d32 CW |
3014 | bool i915_gem_unset_wedged(struct drm_i915_private *i915) |
3015 | { | |
3016 | struct i915_gem_timeline *tl; | |
3017 | int i; | |
3018 | ||
3019 | lockdep_assert_held(&i915->drm.struct_mutex); | |
3020 | if (!test_bit(I915_WEDGED, &i915->gpu_error.flags)) | |
3021 | return true; | |
3022 | ||
3023 | /* Before unwedging, make sure that all pending operations | |
3024 | * are flushed and errored out - we may have requests waiting upon | |
3025 | * third party fences. We marked all inflight requests as EIO, and | |
3026 | * every execbuf since returned EIO, for consistency we want all | |
3027 | * the currently pending requests to also be marked as EIO, which | |
3028 | * is done inside our nop_submit_request - and so we must wait. | |
3029 | * | |
3030 | * No more can be submitted until we reset the wedged bit. | |
3031 | */ | |
3032 | list_for_each_entry(tl, &i915->gt.timelines, link) { | |
3033 | for (i = 0; i < ARRAY_SIZE(tl->engine); i++) { | |
3034 | struct drm_i915_gem_request *rq; | |
3035 | ||
3036 | rq = i915_gem_active_peek(&tl->engine[i].last_request, | |
3037 | &i915->drm.struct_mutex); | |
3038 | if (!rq) | |
3039 | continue; | |
3040 | ||
3041 | /* We can't use our normal waiter as we want to | |
3042 | * avoid recursively trying to handle the current | |
3043 | * reset. The basic dma_fence_default_wait() installs | |
3044 | * a callback for dma_fence_signal(), which is | |
3045 | * triggered by our nop handler (indirectly, the | |
3046 | * callback enables the signaler thread which is | |
3047 | * woken by the nop_submit_request() advancing the seqno | |
3048 | * and when the seqno passes the fence, the signaler | |
3049 | * then signals the fence waking us up). | |
3050 | */ | |
3051 | if (dma_fence_default_wait(&rq->fence, true, | |
3052 | MAX_SCHEDULE_TIMEOUT) < 0) | |
3053 | return false; | |
3054 | } | |
3055 | } | |
3056 | ||
3057 | /* Undo nop_submit_request. We prevent all new i915 requests from | |
3058 | * being queued (by disallowing execbuf whilst wedged) so having | |
3059 | * waited for all active requests above, we know the system is idle | |
3060 | * and do not have to worry about a thread being inside | |
3061 | * engine->submit_request() as we swap over. So unlike installing | |
3062 | * the nop_submit_request on reset, we can do this from normal | |
3063 | * context and do not require stop_machine(). | |
3064 | */ | |
3065 | intel_engines_reset_default_submission(i915); | |
3066 | ||
3067 | smp_mb__before_atomic(); /* complete takeover before enabling execbuf */ | |
3068 | clear_bit(I915_WEDGED, &i915->gpu_error.flags); | |
3069 | ||
3070 | return true; | |
3071 | } | |
3072 | ||
75ef9da2 | 3073 | static void |
673a394b EA |
3074 | i915_gem_retire_work_handler(struct work_struct *work) |
3075 | { | |
b29c19b6 | 3076 | struct drm_i915_private *dev_priv = |
67d97da3 | 3077 | container_of(work, typeof(*dev_priv), gt.retire_work.work); |
91c8a326 | 3078 | struct drm_device *dev = &dev_priv->drm; |
673a394b | 3079 | |
891b48cf | 3080 | /* Come back later if the device is busy... */ |
b29c19b6 | 3081 | if (mutex_trylock(&dev->struct_mutex)) { |
67d97da3 | 3082 | i915_gem_retire_requests(dev_priv); |
b29c19b6 | 3083 | mutex_unlock(&dev->struct_mutex); |
673a394b | 3084 | } |
67d97da3 CW |
3085 | |
3086 | /* Keep the retire handler running until we are finally idle. | |
3087 | * We do not need to do this test under locking as in the worst-case | |
3088 | * we queue the retire worker once too often. | |
3089 | */ | |
c9615613 CW |
3090 | if (READ_ONCE(dev_priv->gt.awake)) { |
3091 | i915_queue_hangcheck(dev_priv); | |
67d97da3 CW |
3092 | queue_delayed_work(dev_priv->wq, |
3093 | &dev_priv->gt.retire_work, | |
bcb45086 | 3094 | round_jiffies_up_relative(HZ)); |
c9615613 | 3095 | } |
b29c19b6 | 3096 | } |
0a58705b | 3097 | |
b29c19b6 CW |
3098 | static void |
3099 | i915_gem_idle_work_handler(struct work_struct *work) | |
3100 | { | |
3101 | struct drm_i915_private *dev_priv = | |
67d97da3 | 3102 | container_of(work, typeof(*dev_priv), gt.idle_work.work); |
91c8a326 | 3103 | struct drm_device *dev = &dev_priv->drm; |
b4ac5afc | 3104 | struct intel_engine_cs *engine; |
3b3f1650 | 3105 | enum intel_engine_id id; |
67d97da3 CW |
3106 | bool rearm_hangcheck; |
3107 | ||
3108 | if (!READ_ONCE(dev_priv->gt.awake)) | |
3109 | return; | |
3110 | ||
0cb5670b ID |
3111 | /* |
3112 | * Wait for last execlists context complete, but bail out in case a | |
3113 | * new request is submitted. | |
3114 | */ | |
8490ae20 | 3115 | wait_for(intel_engines_are_idle(dev_priv), 10); |
28176ef4 | 3116 | if (READ_ONCE(dev_priv->gt.active_requests)) |
67d97da3 CW |
3117 | return; |
3118 | ||
3119 | rearm_hangcheck = | |
3120 | cancel_delayed_work_sync(&dev_priv->gpu_error.hangcheck_work); | |
3121 | ||
3122 | if (!mutex_trylock(&dev->struct_mutex)) { | |
3123 | /* Currently busy, come back later */ | |
3124 | mod_delayed_work(dev_priv->wq, | |
3125 | &dev_priv->gt.idle_work, | |
3126 | msecs_to_jiffies(50)); | |
3127 | goto out_rearm; | |
3128 | } | |
3129 | ||
93c97dc1 ID |
3130 | /* |
3131 | * New request retired after this work handler started, extend active | |
3132 | * period until next instance of the work. | |
3133 | */ | |
3134 | if (work_pending(work)) | |
3135 | goto out_unlock; | |
3136 | ||
28176ef4 | 3137 | if (dev_priv->gt.active_requests) |
67d97da3 | 3138 | goto out_unlock; |
b29c19b6 | 3139 | |
05425249 | 3140 | if (wait_for(intel_engines_are_idle(dev_priv), 10)) |
0cb5670b ID |
3141 | DRM_ERROR("Timeout waiting for engines to idle\n"); |
3142 | ||
67b807a8 CW |
3143 | for_each_engine(engine, dev_priv, id) { |
3144 | intel_engine_disarm_breadcrumbs(engine); | |
67d97da3 | 3145 | i915_gem_batch_pool_fini(&engine->batch_pool); |
67b807a8 | 3146 | } |
35c94185 | 3147 | |
67d97da3 CW |
3148 | GEM_BUG_ON(!dev_priv->gt.awake); |
3149 | dev_priv->gt.awake = false; | |
3150 | rearm_hangcheck = false; | |
30ecad77 | 3151 | |
67d97da3 CW |
3152 | if (INTEL_GEN(dev_priv) >= 6) |
3153 | gen6_rps_idle(dev_priv); | |
3154 | intel_runtime_pm_put(dev_priv); | |
3155 | out_unlock: | |
3156 | mutex_unlock(&dev->struct_mutex); | |
b29c19b6 | 3157 | |
67d97da3 CW |
3158 | out_rearm: |
3159 | if (rearm_hangcheck) { | |
3160 | GEM_BUG_ON(!dev_priv->gt.awake); | |
3161 | i915_queue_hangcheck(dev_priv); | |
35c94185 | 3162 | } |
673a394b EA |
3163 | } |
3164 | ||
b1f788c6 CW |
3165 | void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file) |
3166 | { | |
3167 | struct drm_i915_gem_object *obj = to_intel_bo(gem); | |
3168 | struct drm_i915_file_private *fpriv = file->driver_priv; | |
3169 | struct i915_vma *vma, *vn; | |
3170 | ||
3171 | mutex_lock(&obj->base.dev->struct_mutex); | |
3172 | list_for_each_entry_safe(vma, vn, &obj->vma_list, obj_link) | |
3173 | if (vma->vm->file == fpriv) | |
3174 | i915_vma_close(vma); | |
f8a7fde4 CW |
3175 | |
3176 | if (i915_gem_object_is_active(obj) && | |
3177 | !i915_gem_object_has_active_reference(obj)) { | |
3178 | i915_gem_object_set_active_reference(obj); | |
3179 | i915_gem_object_get(obj); | |
3180 | } | |
b1f788c6 CW |
3181 | mutex_unlock(&obj->base.dev->struct_mutex); |
3182 | } | |
3183 | ||
e95433c7 CW |
3184 | static unsigned long to_wait_timeout(s64 timeout_ns) |
3185 | { | |
3186 | if (timeout_ns < 0) | |
3187 | return MAX_SCHEDULE_TIMEOUT; | |
3188 | ||
3189 | if (timeout_ns == 0) | |
3190 | return 0; | |
3191 | ||
3192 | return nsecs_to_jiffies_timeout(timeout_ns); | |
3193 | } | |
3194 | ||
23ba4fd0 BW |
3195 | /** |
3196 | * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT | |
14bb2c11 TU |
3197 | * @dev: drm device pointer |
3198 | * @data: ioctl data blob | |
3199 | * @file: drm file pointer | |
23ba4fd0 BW |
3200 | * |
3201 | * Returns 0 if successful, else an error is returned with the remaining time in | |
3202 | * the timeout parameter. | |
3203 | * -ETIME: object is still busy after timeout | |
3204 | * -ERESTARTSYS: signal interrupted the wait | |
3205 | * -ENONENT: object doesn't exist | |
3206 | * Also possible, but rare: | |
3207 | * -EAGAIN: GPU wedged | |
3208 | * -ENOMEM: damn | |
3209 | * -ENODEV: Internal IRQ fail | |
3210 | * -E?: The add request failed | |
3211 | * | |
3212 | * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any | |
3213 | * non-zero timeout parameter the wait ioctl will wait for the given number of | |
3214 | * nanoseconds on an object becoming unbusy. Since the wait itself does so | |
3215 | * without holding struct_mutex the object may become re-busied before this | |
3216 | * function completes. A similar but shorter * race condition exists in the busy | |
3217 | * ioctl | |
3218 | */ | |
3219 | int | |
3220 | i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file) | |
3221 | { | |
3222 | struct drm_i915_gem_wait *args = data; | |
3223 | struct drm_i915_gem_object *obj; | |
e95433c7 CW |
3224 | ktime_t start; |
3225 | long ret; | |
23ba4fd0 | 3226 | |
11b5d511 DV |
3227 | if (args->flags != 0) |
3228 | return -EINVAL; | |
3229 | ||
03ac0642 | 3230 | obj = i915_gem_object_lookup(file, args->bo_handle); |
033d549b | 3231 | if (!obj) |
23ba4fd0 | 3232 | return -ENOENT; |
23ba4fd0 | 3233 | |
e95433c7 CW |
3234 | start = ktime_get(); |
3235 | ||
3236 | ret = i915_gem_object_wait(obj, | |
3237 | I915_WAIT_INTERRUPTIBLE | I915_WAIT_ALL, | |
3238 | to_wait_timeout(args->timeout_ns), | |
3239 | to_rps_client(file)); | |
3240 | ||
3241 | if (args->timeout_ns > 0) { | |
3242 | args->timeout_ns -= ktime_to_ns(ktime_sub(ktime_get(), start)); | |
3243 | if (args->timeout_ns < 0) | |
3244 | args->timeout_ns = 0; | |
c1d2061b CW |
3245 | |
3246 | /* | |
3247 | * Apparently ktime isn't accurate enough and occasionally has a | |
3248 | * bit of mismatch in the jiffies<->nsecs<->ktime loop. So patch | |
3249 | * things up to make the test happy. We allow up to 1 jiffy. | |
3250 | * | |
3251 | * This is a regression from the timespec->ktime conversion. | |
3252 | */ | |
3253 | if (ret == -ETIME && !nsecs_to_jiffies(args->timeout_ns)) | |
3254 | args->timeout_ns = 0; | |
b4716185 CW |
3255 | } |
3256 | ||
f0cd5182 | 3257 | i915_gem_object_put(obj); |
ff865885 | 3258 | return ret; |
23ba4fd0 BW |
3259 | } |
3260 | ||
73cb9701 | 3261 | static int wait_for_timeline(struct i915_gem_timeline *tl, unsigned int flags) |
4df2faf4 | 3262 | { |
73cb9701 | 3263 | int ret, i; |
4df2faf4 | 3264 | |
73cb9701 CW |
3265 | for (i = 0; i < ARRAY_SIZE(tl->engine); i++) { |
3266 | ret = i915_gem_active_wait(&tl->engine[i].last_request, flags); | |
3267 | if (ret) | |
3268 | return ret; | |
3269 | } | |
62e63007 | 3270 | |
73cb9701 CW |
3271 | return 0; |
3272 | } | |
3273 | ||
3274 | int i915_gem_wait_for_idle(struct drm_i915_private *i915, unsigned int flags) | |
3275 | { | |
73cb9701 CW |
3276 | int ret; |
3277 | ||
9caa34aa CW |
3278 | if (flags & I915_WAIT_LOCKED) { |
3279 | struct i915_gem_timeline *tl; | |
3280 | ||
3281 | lockdep_assert_held(&i915->drm.struct_mutex); | |
3282 | ||
3283 | list_for_each_entry(tl, &i915->gt.timelines, link) { | |
3284 | ret = wait_for_timeline(tl, flags); | |
3285 | if (ret) | |
3286 | return ret; | |
3287 | } | |
3288 | } else { | |
3289 | ret = wait_for_timeline(&i915->gt.global_timeline, flags); | |
1ec14ad3 CW |
3290 | if (ret) |
3291 | return ret; | |
3292 | } | |
4df2faf4 | 3293 | |
8a1a49f9 | 3294 | return 0; |
4df2faf4 DV |
3295 | } |
3296 | ||
e47c68e9 EA |
3297 | /** Flushes the GTT write domain for the object if it's dirty. */ |
3298 | static void | |
05394f39 | 3299 | i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj) |
e47c68e9 | 3300 | { |
3b5724d7 | 3301 | struct drm_i915_private *dev_priv = to_i915(obj->base.dev); |
1c5d22f7 | 3302 | |
05394f39 | 3303 | if (obj->base.write_domain != I915_GEM_DOMAIN_GTT) |
e47c68e9 EA |
3304 | return; |
3305 | ||
63256ec5 | 3306 | /* No actual flushing is required for the GTT write domain. Writes |
3b5724d7 | 3307 | * to it "immediately" go to main memory as far as we know, so there's |
e47c68e9 | 3308 | * no chipset flush. It also doesn't land in render cache. |
63256ec5 CW |
3309 | * |
3310 | * However, we do have to enforce the order so that all writes through | |
3311 | * the GTT land before any writes to the device, such as updates to | |
3312 | * the GATT itself. | |
3b5724d7 CW |
3313 | * |
3314 | * We also have to wait a bit for the writes to land from the GTT. | |
3315 | * An uncached read (i.e. mmio) seems to be ideal for the round-trip | |
3316 | * timing. This issue has only been observed when switching quickly | |
3317 | * between GTT writes and CPU reads from inside the kernel on recent hw, | |
3318 | * and it appears to only affect discrete GTT blocks (i.e. on LLC | |
3319 | * system agents we cannot reproduce this behaviour). | |
e47c68e9 | 3320 | */ |
63256ec5 | 3321 | wmb(); |
54ec12af | 3322 | if (INTEL_GEN(dev_priv) >= 6 && !HAS_LLC(dev_priv)) { |
e2a2aa36 CW |
3323 | if (intel_runtime_pm_get_if_in_use(dev_priv)) { |
3324 | spin_lock_irq(&dev_priv->uncore.lock); | |
3325 | POSTING_READ_FW(RING_ACTHD(dev_priv->engine[RCS]->mmio_base)); | |
3326 | spin_unlock_irq(&dev_priv->uncore.lock); | |
3327 | intel_runtime_pm_put(dev_priv); | |
3328 | } | |
54ec12af | 3329 | } |
63256ec5 | 3330 | |
d59b21ec | 3331 | intel_fb_obj_flush(obj, write_origin(obj, I915_GEM_DOMAIN_GTT)); |
f99d7069 | 3332 | |
b0dc465f | 3333 | obj->base.write_domain = 0; |
e47c68e9 EA |
3334 | } |
3335 | ||
3336 | /** Flushes the CPU write domain for the object if it's dirty. */ | |
3337 | static void | |
e62b59e4 | 3338 | i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj) |
e47c68e9 | 3339 | { |
05394f39 | 3340 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) |
e47c68e9 EA |
3341 | return; |
3342 | ||
57822dc6 | 3343 | i915_gem_clflush_object(obj, I915_CLFLUSH_SYNC); |
b0dc465f | 3344 | obj->base.write_domain = 0; |
e47c68e9 EA |
3345 | } |
3346 | ||
5a97bcc6 CW |
3347 | static void __i915_gem_object_flush_for_display(struct drm_i915_gem_object *obj) |
3348 | { | |
3349 | if (obj->base.write_domain != I915_GEM_DOMAIN_CPU && !obj->cache_dirty) | |
3350 | return; | |
3351 | ||
57822dc6 | 3352 | i915_gem_clflush_object(obj, I915_CLFLUSH_FORCE); |
5a97bcc6 CW |
3353 | obj->base.write_domain = 0; |
3354 | } | |
3355 | ||
3356 | void i915_gem_object_flush_if_display(struct drm_i915_gem_object *obj) | |
3357 | { | |
3358 | if (!READ_ONCE(obj->pin_display)) | |
3359 | return; | |
3360 | ||
3361 | mutex_lock(&obj->base.dev->struct_mutex); | |
3362 | __i915_gem_object_flush_for_display(obj); | |
3363 | mutex_unlock(&obj->base.dev->struct_mutex); | |
3364 | } | |
3365 | ||
2ef7eeaa EA |
3366 | /** |
3367 | * Moves a single object to the GTT read, and possibly write domain. | |
14bb2c11 TU |
3368 | * @obj: object to act on |
3369 | * @write: ask for write access or read only | |
2ef7eeaa EA |
3370 | * |
3371 | * This function returns when the move is complete, including waiting on | |
3372 | * flushes to occur. | |
3373 | */ | |
79e53945 | 3374 | int |
2021746e | 3375 | i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write) |
2ef7eeaa | 3376 | { |
e47c68e9 | 3377 | int ret; |
2ef7eeaa | 3378 | |
e95433c7 | 3379 | lockdep_assert_held(&obj->base.dev->struct_mutex); |
4c7d62c6 | 3380 | |
e95433c7 CW |
3381 | ret = i915_gem_object_wait(obj, |
3382 | I915_WAIT_INTERRUPTIBLE | | |
3383 | I915_WAIT_LOCKED | | |
3384 | (write ? I915_WAIT_ALL : 0), | |
3385 | MAX_SCHEDULE_TIMEOUT, | |
3386 | NULL); | |
88241785 CW |
3387 | if (ret) |
3388 | return ret; | |
3389 | ||
c13d87ea CW |
3390 | if (obj->base.write_domain == I915_GEM_DOMAIN_GTT) |
3391 | return 0; | |
3392 | ||
43566ded CW |
3393 | /* Flush and acquire obj->pages so that we are coherent through |
3394 | * direct access in memory with previous cached writes through | |
3395 | * shmemfs and that our cache domain tracking remains valid. | |
3396 | * For example, if the obj->filp was moved to swap without us | |
3397 | * being notified and releasing the pages, we would mistakenly | |
3398 | * continue to assume that the obj remained out of the CPU cached | |
3399 | * domain. | |
3400 | */ | |
a4f5ea64 | 3401 | ret = i915_gem_object_pin_pages(obj); |
43566ded CW |
3402 | if (ret) |
3403 | return ret; | |
3404 | ||
e62b59e4 | 3405 | i915_gem_object_flush_cpu_write_domain(obj); |
1c5d22f7 | 3406 | |
d0a57789 CW |
3407 | /* Serialise direct access to this object with the barriers for |
3408 | * coherent writes from the GPU, by effectively invalidating the | |
3409 | * GTT domain upon first access. | |
3410 | */ | |
3411 | if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0) | |
3412 | mb(); | |
3413 | ||
e47c68e9 EA |
3414 | /* It should now be out of any other write domains, and we can update |
3415 | * the domain values for our changes. | |
3416 | */ | |
40e62d5d | 3417 | GEM_BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0); |
05394f39 | 3418 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; |
e47c68e9 | 3419 | if (write) { |
05394f39 CW |
3420 | obj->base.read_domains = I915_GEM_DOMAIN_GTT; |
3421 | obj->base.write_domain = I915_GEM_DOMAIN_GTT; | |
a4f5ea64 | 3422 | obj->mm.dirty = true; |
2ef7eeaa EA |
3423 | } |
3424 | ||
a4f5ea64 | 3425 | i915_gem_object_unpin_pages(obj); |
e47c68e9 EA |
3426 | return 0; |
3427 | } | |
3428 | ||
ef55f92a CW |
3429 | /** |
3430 | * Changes the cache-level of an object across all VMA. | |
14bb2c11 TU |
3431 | * @obj: object to act on |
3432 | * @cache_level: new cache level to set for the object | |
ef55f92a CW |
3433 | * |
3434 | * After this function returns, the object will be in the new cache-level | |
3435 | * across all GTT and the contents of the backing storage will be coherent, | |
3436 | * with respect to the new cache-level. In order to keep the backing storage | |
3437 | * coherent for all users, we only allow a single cache level to be set | |
3438 | * globally on the object and prevent it from being changed whilst the | |
3439 | * hardware is reading from the object. That is if the object is currently | |
3440 | * on the scanout it will be set to uncached (or equivalent display | |
3441 | * cache coherency) and all non-MOCS GPU access will also be uncached so | |
3442 | * that all direct access to the scanout remains coherent. | |
3443 | */ | |
e4ffd173 CW |
3444 | int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj, |
3445 | enum i915_cache_level cache_level) | |
3446 | { | |
aa653a68 | 3447 | struct i915_vma *vma; |
a6a7cc4b | 3448 | int ret; |
e4ffd173 | 3449 | |
4c7d62c6 CW |
3450 | lockdep_assert_held(&obj->base.dev->struct_mutex); |
3451 | ||
e4ffd173 | 3452 | if (obj->cache_level == cache_level) |
a6a7cc4b | 3453 | return 0; |
e4ffd173 | 3454 | |
ef55f92a CW |
3455 | /* Inspect the list of currently bound VMA and unbind any that would |
3456 | * be invalid given the new cache-level. This is principally to | |
3457 | * catch the issue of the CS prefetch crossing page boundaries and | |
3458 | * reading an invalid PTE on older architectures. | |
3459 | */ | |
aa653a68 CW |
3460 | restart: |
3461 | list_for_each_entry(vma, &obj->vma_list, obj_link) { | |
ef55f92a CW |
3462 | if (!drm_mm_node_allocated(&vma->node)) |
3463 | continue; | |
3464 | ||
20dfbde4 | 3465 | if (i915_vma_is_pinned(vma)) { |
ef55f92a CW |
3466 | DRM_DEBUG("can not change the cache level of pinned objects\n"); |
3467 | return -EBUSY; | |
3468 | } | |
3469 | ||
aa653a68 CW |
3470 | if (i915_gem_valid_gtt_space(vma, cache_level)) |
3471 | continue; | |
3472 | ||
3473 | ret = i915_vma_unbind(vma); | |
3474 | if (ret) | |
3475 | return ret; | |
3476 | ||
3477 | /* As unbinding may affect other elements in the | |
3478 | * obj->vma_list (due to side-effects from retiring | |
3479 | * an active vma), play safe and restart the iterator. | |
3480 | */ | |
3481 | goto restart; | |
42d6ab48 CW |
3482 | } |
3483 | ||
ef55f92a CW |
3484 | /* We can reuse the existing drm_mm nodes but need to change the |
3485 | * cache-level on the PTE. We could simply unbind them all and | |
3486 | * rebind with the correct cache-level on next use. However since | |
3487 | * we already have a valid slot, dma mapping, pages etc, we may as | |
3488 | * rewrite the PTE in the belief that doing so tramples upon less | |
3489 | * state and so involves less work. | |
3490 | */ | |
15717de2 | 3491 | if (obj->bind_count) { |
ef55f92a CW |
3492 | /* Before we change the PTE, the GPU must not be accessing it. |
3493 | * If we wait upon the object, we know that all the bound | |
3494 | * VMA are no longer active. | |
3495 | */ | |
e95433c7 CW |
3496 | ret = i915_gem_object_wait(obj, |
3497 | I915_WAIT_INTERRUPTIBLE | | |
3498 | I915_WAIT_LOCKED | | |
3499 | I915_WAIT_ALL, | |
3500 | MAX_SCHEDULE_TIMEOUT, | |
3501 | NULL); | |
e4ffd173 CW |
3502 | if (ret) |
3503 | return ret; | |
3504 | ||
0031fb96 TU |
3505 | if (!HAS_LLC(to_i915(obj->base.dev)) && |
3506 | cache_level != I915_CACHE_NONE) { | |
ef55f92a CW |
3507 | /* Access to snoopable pages through the GTT is |
3508 | * incoherent and on some machines causes a hard | |
3509 | * lockup. Relinquish the CPU mmaping to force | |
3510 | * userspace to refault in the pages and we can | |
3511 | * then double check if the GTT mapping is still | |
3512 | * valid for that pointer access. | |
3513 | */ | |
3514 | i915_gem_release_mmap(obj); | |
3515 | ||
3516 | /* As we no longer need a fence for GTT access, | |
3517 | * we can relinquish it now (and so prevent having | |
3518 | * to steal a fence from someone else on the next | |
3519 | * fence request). Note GPU activity would have | |
3520 | * dropped the fence as all snoopable access is | |
3521 | * supposed to be linear. | |
3522 | */ | |
49ef5294 CW |
3523 | list_for_each_entry(vma, &obj->vma_list, obj_link) { |
3524 | ret = i915_vma_put_fence(vma); | |
3525 | if (ret) | |
3526 | return ret; | |
3527 | } | |
ef55f92a CW |
3528 | } else { |
3529 | /* We either have incoherent backing store and | |
3530 | * so no GTT access or the architecture is fully | |
3531 | * coherent. In such cases, existing GTT mmaps | |
3532 | * ignore the cache bit in the PTE and we can | |
3533 | * rewrite it without confusing the GPU or having | |
3534 | * to force userspace to fault back in its mmaps. | |
3535 | */ | |
e4ffd173 CW |
3536 | } |
3537 | ||
1c7f4bca | 3538 | list_for_each_entry(vma, &obj->vma_list, obj_link) { |
ef55f92a CW |
3539 | if (!drm_mm_node_allocated(&vma->node)) |
3540 | continue; | |
3541 | ||
3542 | ret = i915_vma_bind(vma, cache_level, PIN_UPDATE); | |
3543 | if (ret) | |
3544 | return ret; | |
3545 | } | |
e4ffd173 CW |
3546 | } |
3547 | ||
a6a7cc4b | 3548 | if (obj->base.write_domain == I915_GEM_DOMAIN_CPU && |
e59dc172 | 3549 | i915_gem_object_is_coherent(obj)) |
a6a7cc4b CW |
3550 | obj->cache_dirty = true; |
3551 | ||
1c7f4bca | 3552 | list_for_each_entry(vma, &obj->vma_list, obj_link) |
2c22569b CW |
3553 | vma->node.color = cache_level; |
3554 | obj->cache_level = cache_level; | |
3555 | ||
e4ffd173 CW |
3556 | return 0; |
3557 | } | |
3558 | ||
199adf40 BW |
3559 | int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data, |
3560 | struct drm_file *file) | |
e6994aee | 3561 | { |
199adf40 | 3562 | struct drm_i915_gem_caching *args = data; |
e6994aee | 3563 | struct drm_i915_gem_object *obj; |
fbbd37b3 | 3564 | int err = 0; |
e6994aee | 3565 | |
fbbd37b3 CW |
3566 | rcu_read_lock(); |
3567 | obj = i915_gem_object_lookup_rcu(file, args->handle); | |
3568 | if (!obj) { | |
3569 | err = -ENOENT; | |
3570 | goto out; | |
3571 | } | |
e6994aee | 3572 | |
651d794f CW |
3573 | switch (obj->cache_level) { |
3574 | case I915_CACHE_LLC: | |
3575 | case I915_CACHE_L3_LLC: | |
3576 | args->caching = I915_CACHING_CACHED; | |
3577 | break; | |
3578 | ||
4257d3ba CW |
3579 | case I915_CACHE_WT: |
3580 | args->caching = I915_CACHING_DISPLAY; | |
3581 | break; | |
3582 | ||
651d794f CW |
3583 | default: |
3584 | args->caching = I915_CACHING_NONE; | |
3585 | break; | |
3586 | } | |
fbbd37b3 CW |
3587 | out: |
3588 | rcu_read_unlock(); | |
3589 | return err; | |
e6994aee CW |
3590 | } |
3591 | ||
199adf40 BW |
3592 | int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data, |
3593 | struct drm_file *file) | |
e6994aee | 3594 | { |
9c870d03 | 3595 | struct drm_i915_private *i915 = to_i915(dev); |
199adf40 | 3596 | struct drm_i915_gem_caching *args = data; |
e6994aee CW |
3597 | struct drm_i915_gem_object *obj; |
3598 | enum i915_cache_level level; | |
d65415df | 3599 | int ret = 0; |
e6994aee | 3600 | |
199adf40 BW |
3601 | switch (args->caching) { |
3602 | case I915_CACHING_NONE: | |
e6994aee CW |
3603 | level = I915_CACHE_NONE; |
3604 | break; | |
199adf40 | 3605 | case I915_CACHING_CACHED: |
e5756c10 ID |
3606 | /* |
3607 | * Due to a HW issue on BXT A stepping, GPU stores via a | |
3608 | * snooped mapping may leave stale data in a corresponding CPU | |
3609 | * cacheline, whereas normally such cachelines would get | |
3610 | * invalidated. | |
3611 | */ | |
9c870d03 | 3612 | if (!HAS_LLC(i915) && !HAS_SNOOP(i915)) |
e5756c10 ID |
3613 | return -ENODEV; |
3614 | ||
e6994aee CW |
3615 | level = I915_CACHE_LLC; |
3616 | break; | |
4257d3ba | 3617 | case I915_CACHING_DISPLAY: |
9c870d03 | 3618 | level = HAS_WT(i915) ? I915_CACHE_WT : I915_CACHE_NONE; |
4257d3ba | 3619 | break; |
e6994aee CW |
3620 | default: |
3621 | return -EINVAL; | |
3622 | } | |
3623 | ||
d65415df CW |
3624 | obj = i915_gem_object_lookup(file, args->handle); |
3625 | if (!obj) | |
3626 | return -ENOENT; | |
3627 | ||
3628 | if (obj->cache_level == level) | |
3629 | goto out; | |
3630 | ||
3631 | ret = i915_gem_object_wait(obj, | |
3632 | I915_WAIT_INTERRUPTIBLE, | |
3633 | MAX_SCHEDULE_TIMEOUT, | |
3634 | to_rps_client(file)); | |
3bc2913e | 3635 | if (ret) |
d65415df | 3636 | goto out; |
3bc2913e | 3637 | |
d65415df CW |
3638 | ret = i915_mutex_lock_interruptible(dev); |
3639 | if (ret) | |
3640 | goto out; | |
e6994aee CW |
3641 | |
3642 | ret = i915_gem_object_set_cache_level(obj, level); | |
e6994aee | 3643 | mutex_unlock(&dev->struct_mutex); |
d65415df CW |
3644 | |
3645 | out: | |
3646 | i915_gem_object_put(obj); | |
e6994aee CW |
3647 | return ret; |
3648 | } | |
3649 | ||
b9241ea3 | 3650 | /* |
2da3b9b9 CW |
3651 | * Prepare buffer for display plane (scanout, cursors, etc). |
3652 | * Can be called from an uninterruptible phase (modesetting) and allows | |
3653 | * any flushes to be pipelined (for pageflips). | |
b9241ea3 | 3654 | */ |
058d88c4 | 3655 | struct i915_vma * |
2da3b9b9 CW |
3656 | i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj, |
3657 | u32 alignment, | |
e6617330 | 3658 | const struct i915_ggtt_view *view) |
b9241ea3 | 3659 | { |
058d88c4 | 3660 | struct i915_vma *vma; |
b9241ea3 ZW |
3661 | int ret; |
3662 | ||
4c7d62c6 CW |
3663 | lockdep_assert_held(&obj->base.dev->struct_mutex); |
3664 | ||
cc98b413 CW |
3665 | /* Mark the pin_display early so that we account for the |
3666 | * display coherency whilst setting up the cache domains. | |
3667 | */ | |
8a0c39b1 | 3668 | obj->pin_display++; |
cc98b413 | 3669 | |
a7ef0640 EA |
3670 | /* The display engine is not coherent with the LLC cache on gen6. As |
3671 | * a result, we make sure that the pinning that is about to occur is | |
3672 | * done with uncached PTEs. This is lowest common denominator for all | |
3673 | * chipsets. | |
3674 | * | |
3675 | * However for gen6+, we could do better by using the GFDT bit instead | |
3676 | * of uncaching, which would allow us to flush all the LLC-cached data | |
3677 | * with that bit in the PTE to main memory with just one PIPE_CONTROL. | |
3678 | */ | |
651d794f | 3679 | ret = i915_gem_object_set_cache_level(obj, |
8652744b TU |
3680 | HAS_WT(to_i915(obj->base.dev)) ? |
3681 | I915_CACHE_WT : I915_CACHE_NONE); | |
058d88c4 CW |
3682 | if (ret) { |
3683 | vma = ERR_PTR(ret); | |
cc98b413 | 3684 | goto err_unpin_display; |
058d88c4 | 3685 | } |
a7ef0640 | 3686 | |
2da3b9b9 CW |
3687 | /* As the user may map the buffer once pinned in the display plane |
3688 | * (e.g. libkms for the bootup splash), we have to ensure that we | |
2efb813d CW |
3689 | * always use map_and_fenceable for all scanout buffers. However, |
3690 | * it may simply be too big to fit into mappable, in which case | |
3691 | * put it anyway and hope that userspace can cope (but always first | |
3692 | * try to preserve the existing ABI). | |
2da3b9b9 | 3693 | */ |
2efb813d | 3694 | vma = ERR_PTR(-ENOSPC); |
47a8e3f6 | 3695 | if (!view || view->type == I915_GGTT_VIEW_NORMAL) |
2efb813d CW |
3696 | vma = i915_gem_object_ggtt_pin(obj, view, 0, alignment, |
3697 | PIN_MAPPABLE | PIN_NONBLOCK); | |
767a222e CW |
3698 | if (IS_ERR(vma)) { |
3699 | struct drm_i915_private *i915 = to_i915(obj->base.dev); | |
3700 | unsigned int flags; | |
3701 | ||
3702 | /* Valleyview is definitely limited to scanning out the first | |
3703 | * 512MiB. Lets presume this behaviour was inherited from the | |
3704 | * g4x display engine and that all earlier gen are similarly | |
3705 | * limited. Testing suggests that it is a little more | |
3706 | * complicated than this. For example, Cherryview appears quite | |
3707 | * happy to scanout from anywhere within its global aperture. | |
3708 | */ | |
3709 | flags = 0; | |
3710 | if (HAS_GMCH_DISPLAY(i915)) | |
3711 | flags = PIN_MAPPABLE; | |
3712 | vma = i915_gem_object_ggtt_pin(obj, view, 0, alignment, flags); | |
3713 | } | |
058d88c4 | 3714 | if (IS_ERR(vma)) |
cc98b413 | 3715 | goto err_unpin_display; |
2da3b9b9 | 3716 | |
d8923dcf CW |
3717 | vma->display_alignment = max_t(u64, vma->display_alignment, alignment); |
3718 | ||
a6a7cc4b | 3719 | /* Treat this as an end-of-frame, like intel_user_framebuffer_dirty() */ |
5a97bcc6 | 3720 | __i915_gem_object_flush_for_display(obj); |
d59b21ec | 3721 | intel_fb_obj_flush(obj, ORIGIN_DIRTYFB); |
b118c1e3 | 3722 | |
2da3b9b9 CW |
3723 | /* It should now be out of any other write domains, and we can update |
3724 | * the domain values for our changes. | |
3725 | */ | |
05394f39 | 3726 | obj->base.read_domains |= I915_GEM_DOMAIN_GTT; |
b9241ea3 | 3727 | |
058d88c4 | 3728 | return vma; |
cc98b413 CW |
3729 | |
3730 | err_unpin_display: | |
8a0c39b1 | 3731 | obj->pin_display--; |
058d88c4 | 3732 | return vma; |
cc98b413 CW |
3733 | } |
3734 | ||
3735 | void | |
058d88c4 | 3736 | i915_gem_object_unpin_from_display_plane(struct i915_vma *vma) |
cc98b413 | 3737 | { |
49d73912 | 3738 | lockdep_assert_held(&vma->vm->i915->drm.struct_mutex); |
4c7d62c6 | 3739 | |
058d88c4 | 3740 | if (WARN_ON(vma->obj->pin_display == 0)) |
8a0c39b1 TU |
3741 | return; |
3742 | ||
d8923dcf | 3743 | if (--vma->obj->pin_display == 0) |
f51455d4 | 3744 | vma->display_alignment = I915_GTT_MIN_ALIGNMENT; |
e6617330 | 3745 | |
383d5823 | 3746 | /* Bump the LRU to try and avoid premature eviction whilst flipping */ |
befedbb7 | 3747 | i915_gem_object_bump_inactive_ggtt(vma->obj); |
383d5823 | 3748 | |
058d88c4 | 3749 | i915_vma_unpin(vma); |
b9241ea3 ZW |
3750 | } |
3751 | ||
e47c68e9 EA |
3752 | /** |
3753 | * Moves a single object to the CPU read, and possibly write domain. | |
14bb2c11 TU |
3754 | * @obj: object to act on |
3755 | * @write: requesting write or read-only access | |
e47c68e9 EA |
3756 | * |
3757 | * This function returns when the move is complete, including waiting on | |
3758 | * flushes to occur. | |
3759 | */ | |
dabdfe02 | 3760 | int |
919926ae | 3761 | i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write) |
e47c68e9 | 3762 | { |
e47c68e9 EA |
3763 | int ret; |
3764 | ||
e95433c7 | 3765 | lockdep_assert_held(&obj->base.dev->struct_mutex); |
4c7d62c6 | 3766 | |
e95433c7 CW |
3767 | ret = i915_gem_object_wait(obj, |
3768 | I915_WAIT_INTERRUPTIBLE | | |
3769 | I915_WAIT_LOCKED | | |
3770 | (write ? I915_WAIT_ALL : 0), | |
3771 | MAX_SCHEDULE_TIMEOUT, | |
3772 | NULL); | |
88241785 CW |
3773 | if (ret) |
3774 | return ret; | |
3775 | ||
c13d87ea CW |
3776 | if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) |
3777 | return 0; | |
3778 | ||
e47c68e9 | 3779 | i915_gem_object_flush_gtt_write_domain(obj); |
2ef7eeaa | 3780 | |
e47c68e9 | 3781 | /* Flush the CPU cache if it's still invalid. */ |
05394f39 | 3782 | if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) { |
57822dc6 | 3783 | i915_gem_clflush_object(obj, I915_CLFLUSH_SYNC); |
05394f39 | 3784 | obj->base.read_domains |= I915_GEM_DOMAIN_CPU; |
2ef7eeaa EA |
3785 | } |
3786 | ||
3787 | /* It should now be out of any other write domains, and we can update | |
3788 | * the domain values for our changes. | |
3789 | */ | |
40e62d5d | 3790 | GEM_BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0); |
e47c68e9 EA |
3791 | |
3792 | /* If we're writing through the CPU, then the GPU read domains will | |
3793 | * need to be invalidated at next use. | |
3794 | */ | |
3795 | if (write) { | |
05394f39 CW |
3796 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; |
3797 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
e47c68e9 | 3798 | } |
2ef7eeaa EA |
3799 | |
3800 | return 0; | |
3801 | } | |
3802 | ||
673a394b EA |
3803 | /* Throttle our rendering by waiting until the ring has completed our requests |
3804 | * emitted over 20 msec ago. | |
3805 | * | |
b962442e EA |
3806 | * Note that if we were to use the current jiffies each time around the loop, |
3807 | * we wouldn't escape the function with any frames outstanding if the time to | |
3808 | * render a frame was over 20ms. | |
3809 | * | |
673a394b EA |
3810 | * This should get us reasonable parallelism between CPU and GPU but also |
3811 | * relatively low latency when blocking on a particular request to finish. | |
3812 | */ | |
40a5f0de | 3813 | static int |
f787a5f5 | 3814 | i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file) |
40a5f0de | 3815 | { |
fac5e23e | 3816 | struct drm_i915_private *dev_priv = to_i915(dev); |
f787a5f5 | 3817 | struct drm_i915_file_private *file_priv = file->driver_priv; |
d0bc54f2 | 3818 | unsigned long recent_enough = jiffies - DRM_I915_THROTTLE_JIFFIES; |
54fb2411 | 3819 | struct drm_i915_gem_request *request, *target = NULL; |
e95433c7 | 3820 | long ret; |
93533c29 | 3821 | |
f4457ae7 CW |
3822 | /* ABI: return -EIO if already wedged */ |
3823 | if (i915_terminally_wedged(&dev_priv->gpu_error)) | |
3824 | return -EIO; | |
e110e8d6 | 3825 | |
1c25595f | 3826 | spin_lock(&file_priv->mm.lock); |
c8659efa | 3827 | list_for_each_entry(request, &file_priv->mm.request_list, client_link) { |
b962442e EA |
3828 | if (time_after_eq(request->emitted_jiffies, recent_enough)) |
3829 | break; | |
40a5f0de | 3830 | |
c8659efa CW |
3831 | if (target) { |
3832 | list_del(&target->client_link); | |
3833 | target->file_priv = NULL; | |
3834 | } | |
fcfa423c | 3835 | |
54fb2411 | 3836 | target = request; |
b962442e | 3837 | } |
ff865885 | 3838 | if (target) |
e8a261ea | 3839 | i915_gem_request_get(target); |
1c25595f | 3840 | spin_unlock(&file_priv->mm.lock); |
40a5f0de | 3841 | |
54fb2411 | 3842 | if (target == NULL) |
f787a5f5 | 3843 | return 0; |
2bc43b5c | 3844 | |
e95433c7 CW |
3845 | ret = i915_wait_request(target, |
3846 | I915_WAIT_INTERRUPTIBLE, | |
3847 | MAX_SCHEDULE_TIMEOUT); | |
e8a261ea | 3848 | i915_gem_request_put(target); |
ff865885 | 3849 | |
e95433c7 | 3850 | return ret < 0 ? ret : 0; |
40a5f0de EA |
3851 | } |
3852 | ||
058d88c4 | 3853 | struct i915_vma * |
ec7adb6e JL |
3854 | i915_gem_object_ggtt_pin(struct drm_i915_gem_object *obj, |
3855 | const struct i915_ggtt_view *view, | |
91b2db6f | 3856 | u64 size, |
2ffffd0f CW |
3857 | u64 alignment, |
3858 | u64 flags) | |
ec7adb6e | 3859 | { |
ad16d2ed CW |
3860 | struct drm_i915_private *dev_priv = to_i915(obj->base.dev); |
3861 | struct i915_address_space *vm = &dev_priv->ggtt.base; | |
59bfa124 CW |
3862 | struct i915_vma *vma; |
3863 | int ret; | |
72e96d64 | 3864 | |
4c7d62c6 CW |
3865 | lockdep_assert_held(&obj->base.dev->struct_mutex); |
3866 | ||
718659a6 | 3867 | vma = i915_vma_instance(obj, vm, view); |
e0216b76 | 3868 | if (unlikely(IS_ERR(vma))) |
058d88c4 | 3869 | return vma; |
59bfa124 CW |
3870 | |
3871 | if (i915_vma_misplaced(vma, size, alignment, flags)) { | |
3872 | if (flags & PIN_NONBLOCK && | |
3873 | (i915_vma_is_pinned(vma) || i915_vma_is_active(vma))) | |
058d88c4 | 3874 | return ERR_PTR(-ENOSPC); |
59bfa124 | 3875 | |
ad16d2ed | 3876 | if (flags & PIN_MAPPABLE) { |
ad16d2ed CW |
3877 | /* If the required space is larger than the available |
3878 | * aperture, we will not able to find a slot for the | |
3879 | * object and unbinding the object now will be in | |
3880 | * vain. Worse, doing so may cause us to ping-pong | |
3881 | * the object in and out of the Global GTT and | |
3882 | * waste a lot of cycles under the mutex. | |
3883 | */ | |
944397f0 | 3884 | if (vma->fence_size > dev_priv->ggtt.mappable_end) |
ad16d2ed CW |
3885 | return ERR_PTR(-E2BIG); |
3886 | ||
3887 | /* If NONBLOCK is set the caller is optimistically | |
3888 | * trying to cache the full object within the mappable | |
3889 | * aperture, and *must* have a fallback in place for | |
3890 | * situations where we cannot bind the object. We | |
3891 | * can be a little more lax here and use the fallback | |
3892 | * more often to avoid costly migrations of ourselves | |
3893 | * and other objects within the aperture. | |
3894 | * | |
3895 | * Half-the-aperture is used as a simple heuristic. | |
3896 | * More interesting would to do search for a free | |
3897 | * block prior to making the commitment to unbind. | |
3898 | * That caters for the self-harm case, and with a | |
3899 | * little more heuristics (e.g. NOFAULT, NOEVICT) | |
3900 | * we could try to minimise harm to others. | |
3901 | */ | |
3902 | if (flags & PIN_NONBLOCK && | |
944397f0 | 3903 | vma->fence_size > dev_priv->ggtt.mappable_end / 2) |
ad16d2ed CW |
3904 | return ERR_PTR(-ENOSPC); |
3905 | } | |
3906 | ||
59bfa124 CW |
3907 | WARN(i915_vma_is_pinned(vma), |
3908 | "bo is already pinned in ggtt with incorrect alignment:" | |
05a20d09 CW |
3909 | " offset=%08x, req.alignment=%llx," |
3910 | " req.map_and_fenceable=%d, vma->map_and_fenceable=%d\n", | |
3911 | i915_ggtt_offset(vma), alignment, | |
59bfa124 | 3912 | !!(flags & PIN_MAPPABLE), |
05a20d09 | 3913 | i915_vma_is_map_and_fenceable(vma)); |
59bfa124 CW |
3914 | ret = i915_vma_unbind(vma); |
3915 | if (ret) | |
058d88c4 | 3916 | return ERR_PTR(ret); |
59bfa124 CW |
3917 | } |
3918 | ||
058d88c4 CW |
3919 | ret = i915_vma_pin(vma, size, alignment, flags | PIN_GLOBAL); |
3920 | if (ret) | |
3921 | return ERR_PTR(ret); | |
ec7adb6e | 3922 | |
058d88c4 | 3923 | return vma; |
673a394b EA |
3924 | } |
3925 | ||
edf6b76f | 3926 | static __always_inline unsigned int __busy_read_flag(unsigned int id) |
3fdc13c7 CW |
3927 | { |
3928 | /* Note that we could alias engines in the execbuf API, but | |
3929 | * that would be very unwise as it prevents userspace from | |
3930 | * fine control over engine selection. Ahem. | |
3931 | * | |
3932 | * This should be something like EXEC_MAX_ENGINE instead of | |
3933 | * I915_NUM_ENGINES. | |
3934 | */ | |
3935 | BUILD_BUG_ON(I915_NUM_ENGINES > 16); | |
3936 | return 0x10000 << id; | |
3937 | } | |
3938 | ||
3939 | static __always_inline unsigned int __busy_write_id(unsigned int id) | |
3940 | { | |
70cb472c CW |
3941 | /* The uABI guarantees an active writer is also amongst the read |
3942 | * engines. This would be true if we accessed the activity tracking | |
3943 | * under the lock, but as we perform the lookup of the object and | |
3944 | * its activity locklessly we can not guarantee that the last_write | |
3945 | * being active implies that we have set the same engine flag from | |
3946 | * last_read - hence we always set both read and write busy for | |
3947 | * last_write. | |
3948 | */ | |
3949 | return id | __busy_read_flag(id); | |
3fdc13c7 CW |
3950 | } |
3951 | ||
edf6b76f | 3952 | static __always_inline unsigned int |
d07f0e59 | 3953 | __busy_set_if_active(const struct dma_fence *fence, |
3fdc13c7 CW |
3954 | unsigned int (*flag)(unsigned int id)) |
3955 | { | |
d07f0e59 | 3956 | struct drm_i915_gem_request *rq; |
3fdc13c7 | 3957 | |
d07f0e59 CW |
3958 | /* We have to check the current hw status of the fence as the uABI |
3959 | * guarantees forward progress. We could rely on the idle worker | |
3960 | * to eventually flush us, but to minimise latency just ask the | |
3961 | * hardware. | |
1255501d | 3962 | * |
d07f0e59 | 3963 | * Note we only report on the status of native fences. |
1255501d | 3964 | */ |
d07f0e59 CW |
3965 | if (!dma_fence_is_i915(fence)) |
3966 | return 0; | |
3967 | ||
3968 | /* opencode to_request() in order to avoid const warnings */ | |
3969 | rq = container_of(fence, struct drm_i915_gem_request, fence); | |
3970 | if (i915_gem_request_completed(rq)) | |
3971 | return 0; | |
3972 | ||
3973 | return flag(rq->engine->exec_id); | |
3fdc13c7 CW |
3974 | } |
3975 | ||
edf6b76f | 3976 | static __always_inline unsigned int |
d07f0e59 | 3977 | busy_check_reader(const struct dma_fence *fence) |
3fdc13c7 | 3978 | { |
d07f0e59 | 3979 | return __busy_set_if_active(fence, __busy_read_flag); |
3fdc13c7 CW |
3980 | } |
3981 | ||
edf6b76f | 3982 | static __always_inline unsigned int |
d07f0e59 | 3983 | busy_check_writer(const struct dma_fence *fence) |
3fdc13c7 | 3984 | { |
d07f0e59 CW |
3985 | if (!fence) |
3986 | return 0; | |
3987 | ||
3988 | return __busy_set_if_active(fence, __busy_write_id); | |
3fdc13c7 CW |
3989 | } |
3990 | ||
673a394b EA |
3991 | int |
3992 | i915_gem_busy_ioctl(struct drm_device *dev, void *data, | |
05394f39 | 3993 | struct drm_file *file) |
673a394b EA |
3994 | { |
3995 | struct drm_i915_gem_busy *args = data; | |
05394f39 | 3996 | struct drm_i915_gem_object *obj; |
d07f0e59 CW |
3997 | struct reservation_object_list *list; |
3998 | unsigned int seq; | |
fbbd37b3 | 3999 | int err; |
673a394b | 4000 | |
d07f0e59 | 4001 | err = -ENOENT; |
fbbd37b3 CW |
4002 | rcu_read_lock(); |
4003 | obj = i915_gem_object_lookup_rcu(file, args->handle); | |
d07f0e59 | 4004 | if (!obj) |
fbbd37b3 | 4005 | goto out; |
d1b851fc | 4006 | |
d07f0e59 CW |
4007 | /* A discrepancy here is that we do not report the status of |
4008 | * non-i915 fences, i.e. even though we may report the object as idle, | |
4009 | * a call to set-domain may still stall waiting for foreign rendering. | |
4010 | * This also means that wait-ioctl may report an object as busy, | |
4011 | * where busy-ioctl considers it idle. | |
4012 | * | |
4013 | * We trade the ability to warn of foreign fences to report on which | |
4014 | * i915 engines are active for the object. | |
4015 | * | |
4016 | * Alternatively, we can trade that extra information on read/write | |
4017 | * activity with | |
4018 | * args->busy = | |
4019 | * !reservation_object_test_signaled_rcu(obj->resv, true); | |
4020 | * to report the overall busyness. This is what the wait-ioctl does. | |
4021 | * | |
4022 | */ | |
4023 | retry: | |
4024 | seq = raw_read_seqcount(&obj->resv->seq); | |
426960be | 4025 | |
d07f0e59 CW |
4026 | /* Translate the exclusive fence to the READ *and* WRITE engine */ |
4027 | args->busy = busy_check_writer(rcu_dereference(obj->resv->fence_excl)); | |
3fdc13c7 | 4028 | |
d07f0e59 CW |
4029 | /* Translate shared fences to READ set of engines */ |
4030 | list = rcu_dereference(obj->resv->fence); | |
4031 | if (list) { | |
4032 | unsigned int shared_count = list->shared_count, i; | |
3fdc13c7 | 4033 | |
d07f0e59 CW |
4034 | for (i = 0; i < shared_count; ++i) { |
4035 | struct dma_fence *fence = | |
4036 | rcu_dereference(list->shared[i]); | |
4037 | ||
4038 | args->busy |= busy_check_reader(fence); | |
4039 | } | |
426960be | 4040 | } |
673a394b | 4041 | |
d07f0e59 CW |
4042 | if (args->busy && read_seqcount_retry(&obj->resv->seq, seq)) |
4043 | goto retry; | |
4044 | ||
4045 | err = 0; | |
fbbd37b3 CW |
4046 | out: |
4047 | rcu_read_unlock(); | |
4048 | return err; | |
673a394b EA |
4049 | } |
4050 | ||
4051 | int | |
4052 | i915_gem_throttle_ioctl(struct drm_device *dev, void *data, | |
4053 | struct drm_file *file_priv) | |
4054 | { | |
0206e353 | 4055 | return i915_gem_ring_throttle(dev, file_priv); |
673a394b EA |
4056 | } |
4057 | ||
3ef94daa CW |
4058 | int |
4059 | i915_gem_madvise_ioctl(struct drm_device *dev, void *data, | |
4060 | struct drm_file *file_priv) | |
4061 | { | |
fac5e23e | 4062 | struct drm_i915_private *dev_priv = to_i915(dev); |
3ef94daa | 4063 | struct drm_i915_gem_madvise *args = data; |
05394f39 | 4064 | struct drm_i915_gem_object *obj; |
1233e2db | 4065 | int err; |
3ef94daa CW |
4066 | |
4067 | switch (args->madv) { | |
4068 | case I915_MADV_DONTNEED: | |
4069 | case I915_MADV_WILLNEED: | |
4070 | break; | |
4071 | default: | |
4072 | return -EINVAL; | |
4073 | } | |
4074 | ||
03ac0642 | 4075 | obj = i915_gem_object_lookup(file_priv, args->handle); |
1233e2db CW |
4076 | if (!obj) |
4077 | return -ENOENT; | |
4078 | ||
4079 | err = mutex_lock_interruptible(&obj->mm.lock); | |
4080 | if (err) | |
4081 | goto out; | |
3ef94daa | 4082 | |
a4f5ea64 | 4083 | if (obj->mm.pages && |
3e510a8e | 4084 | i915_gem_object_is_tiled(obj) && |
656bfa3a | 4085 | dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES) { |
bc0629a7 CW |
4086 | if (obj->mm.madv == I915_MADV_WILLNEED) { |
4087 | GEM_BUG_ON(!obj->mm.quirked); | |
a4f5ea64 | 4088 | __i915_gem_object_unpin_pages(obj); |
bc0629a7 CW |
4089 | obj->mm.quirked = false; |
4090 | } | |
4091 | if (args->madv == I915_MADV_WILLNEED) { | |
2c3a3f44 | 4092 | GEM_BUG_ON(obj->mm.quirked); |
a4f5ea64 | 4093 | __i915_gem_object_pin_pages(obj); |
bc0629a7 CW |
4094 | obj->mm.quirked = true; |
4095 | } | |
656bfa3a DV |
4096 | } |
4097 | ||
a4f5ea64 CW |
4098 | if (obj->mm.madv != __I915_MADV_PURGED) |
4099 | obj->mm.madv = args->madv; | |
3ef94daa | 4100 | |
6c085a72 | 4101 | /* if the object is no longer attached, discard its backing storage */ |
a4f5ea64 | 4102 | if (obj->mm.madv == I915_MADV_DONTNEED && !obj->mm.pages) |
2d7ef395 CW |
4103 | i915_gem_object_truncate(obj); |
4104 | ||
a4f5ea64 | 4105 | args->retained = obj->mm.madv != __I915_MADV_PURGED; |
1233e2db | 4106 | mutex_unlock(&obj->mm.lock); |
bb6baf76 | 4107 | |
1233e2db | 4108 | out: |
f8c417cd | 4109 | i915_gem_object_put(obj); |
1233e2db | 4110 | return err; |
3ef94daa CW |
4111 | } |
4112 | ||
5b8c8aec CW |
4113 | static void |
4114 | frontbuffer_retire(struct i915_gem_active *active, | |
4115 | struct drm_i915_gem_request *request) | |
4116 | { | |
4117 | struct drm_i915_gem_object *obj = | |
4118 | container_of(active, typeof(*obj), frontbuffer_write); | |
4119 | ||
d59b21ec | 4120 | intel_fb_obj_flush(obj, ORIGIN_CS); |
5b8c8aec CW |
4121 | } |
4122 | ||
37e680a1 CW |
4123 | void i915_gem_object_init(struct drm_i915_gem_object *obj, |
4124 | const struct drm_i915_gem_object_ops *ops) | |
0327d6ba | 4125 | { |
1233e2db CW |
4126 | mutex_init(&obj->mm.lock); |
4127 | ||
56cea323 | 4128 | INIT_LIST_HEAD(&obj->global_link); |
275f039d | 4129 | INIT_LIST_HEAD(&obj->userfault_link); |
b25cb2f8 | 4130 | INIT_LIST_HEAD(&obj->obj_exec_link); |
2f633156 | 4131 | INIT_LIST_HEAD(&obj->vma_list); |
8d9d5744 | 4132 | INIT_LIST_HEAD(&obj->batch_pool_link); |
0327d6ba | 4133 | |
37e680a1 CW |
4134 | obj->ops = ops; |
4135 | ||
d07f0e59 CW |
4136 | reservation_object_init(&obj->__builtin_resv); |
4137 | obj->resv = &obj->__builtin_resv; | |
4138 | ||
50349247 | 4139 | obj->frontbuffer_ggtt_origin = ORIGIN_GTT; |
5b8c8aec | 4140 | init_request_active(&obj->frontbuffer_write, frontbuffer_retire); |
a4f5ea64 CW |
4141 | |
4142 | obj->mm.madv = I915_MADV_WILLNEED; | |
4143 | INIT_RADIX_TREE(&obj->mm.get_page.radix, GFP_KERNEL | __GFP_NOWARN); | |
4144 | mutex_init(&obj->mm.get_page.lock); | |
0327d6ba | 4145 | |
f19ec8cb | 4146 | i915_gem_info_add_obj(to_i915(obj->base.dev), obj->base.size); |
0327d6ba CW |
4147 | } |
4148 | ||
37e680a1 | 4149 | static const struct drm_i915_gem_object_ops i915_gem_object_ops = { |
3599a91c TU |
4150 | .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE | |
4151 | I915_GEM_OBJECT_IS_SHRINKABLE, | |
7c55e2c5 | 4152 | |
37e680a1 CW |
4153 | .get_pages = i915_gem_object_get_pages_gtt, |
4154 | .put_pages = i915_gem_object_put_pages_gtt, | |
7c55e2c5 CW |
4155 | |
4156 | .pwrite = i915_gem_object_pwrite_gtt, | |
37e680a1 CW |
4157 | }; |
4158 | ||
b4bcbe2a | 4159 | struct drm_i915_gem_object * |
12d79d78 | 4160 | i915_gem_object_create(struct drm_i915_private *dev_priv, u64 size) |
ac52bc56 | 4161 | { |
c397b908 | 4162 | struct drm_i915_gem_object *obj; |
5949eac4 | 4163 | struct address_space *mapping; |
1a240d4d | 4164 | gfp_t mask; |
fe3db79b | 4165 | int ret; |
ac52bc56 | 4166 | |
b4bcbe2a CW |
4167 | /* There is a prevalence of the assumption that we fit the object's |
4168 | * page count inside a 32bit _signed_ variable. Let's document this and | |
4169 | * catch if we ever need to fix it. In the meantime, if you do spot | |
4170 | * such a local variable, please consider fixing! | |
4171 | */ | |
4172 | if (WARN_ON(size >> PAGE_SHIFT > INT_MAX)) | |
4173 | return ERR_PTR(-E2BIG); | |
4174 | ||
4175 | if (overflows_type(size, obj->base.size)) | |
4176 | return ERR_PTR(-E2BIG); | |
4177 | ||
187685cb | 4178 | obj = i915_gem_object_alloc(dev_priv); |
c397b908 | 4179 | if (obj == NULL) |
fe3db79b | 4180 | return ERR_PTR(-ENOMEM); |
673a394b | 4181 | |
12d79d78 | 4182 | ret = drm_gem_object_init(&dev_priv->drm, &obj->base, size); |
fe3db79b CW |
4183 | if (ret) |
4184 | goto fail; | |
673a394b | 4185 | |
bed1ea95 | 4186 | mask = GFP_HIGHUSER | __GFP_RECLAIMABLE; |
c0f86832 | 4187 | if (IS_I965GM(dev_priv) || IS_I965G(dev_priv)) { |
bed1ea95 CW |
4188 | /* 965gm cannot relocate objects above 4GiB. */ |
4189 | mask &= ~__GFP_HIGHMEM; | |
4190 | mask |= __GFP_DMA32; | |
4191 | } | |
4192 | ||
93c76a3d | 4193 | mapping = obj->base.filp->f_mapping; |
bed1ea95 | 4194 | mapping_set_gfp_mask(mapping, mask); |
5949eac4 | 4195 | |
37e680a1 | 4196 | i915_gem_object_init(obj, &i915_gem_object_ops); |
73aa808f | 4197 | |
c397b908 DV |
4198 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; |
4199 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; | |
673a394b | 4200 | |
0031fb96 | 4201 | if (HAS_LLC(dev_priv)) { |
3d29b842 | 4202 | /* On some devices, we can have the GPU use the LLC (the CPU |
a1871112 EA |
4203 | * cache) for about a 10% performance improvement |
4204 | * compared to uncached. Graphics requests other than | |
4205 | * display scanout are coherent with the CPU in | |
4206 | * accessing this cache. This means in this mode we | |
4207 | * don't need to clflush on the CPU side, and on the | |
4208 | * GPU side we only need to flush internal caches to | |
4209 | * get data visible to the CPU. | |
4210 | * | |
4211 | * However, we maintain the display planes as UC, and so | |
4212 | * need to rebind when first used as such. | |
4213 | */ | |
4214 | obj->cache_level = I915_CACHE_LLC; | |
4215 | } else | |
4216 | obj->cache_level = I915_CACHE_NONE; | |
4217 | ||
d861e338 DV |
4218 | trace_i915_gem_object_create(obj); |
4219 | ||
05394f39 | 4220 | return obj; |
fe3db79b CW |
4221 | |
4222 | fail: | |
4223 | i915_gem_object_free(obj); | |
fe3db79b | 4224 | return ERR_PTR(ret); |
c397b908 DV |
4225 | } |
4226 | ||
340fbd8c CW |
4227 | static bool discard_backing_storage(struct drm_i915_gem_object *obj) |
4228 | { | |
4229 | /* If we are the last user of the backing storage (be it shmemfs | |
4230 | * pages or stolen etc), we know that the pages are going to be | |
4231 | * immediately released. In this case, we can then skip copying | |
4232 | * back the contents from the GPU. | |
4233 | */ | |
4234 | ||
a4f5ea64 | 4235 | if (obj->mm.madv != I915_MADV_WILLNEED) |
340fbd8c CW |
4236 | return false; |
4237 | ||
4238 | if (obj->base.filp == NULL) | |
4239 | return true; | |
4240 | ||
4241 | /* At first glance, this looks racy, but then again so would be | |
4242 | * userspace racing mmap against close. However, the first external | |
4243 | * reference to the filp can only be obtained through the | |
4244 | * i915_gem_mmap_ioctl() which safeguards us against the user | |
4245 | * acquiring such a reference whilst we are in the middle of | |
4246 | * freeing the object. | |
4247 | */ | |
4248 | return atomic_long_read(&obj->base.filp->f_count) == 1; | |
4249 | } | |
4250 | ||
fbbd37b3 CW |
4251 | static void __i915_gem_free_objects(struct drm_i915_private *i915, |
4252 | struct llist_node *freed) | |
673a394b | 4253 | { |
fbbd37b3 | 4254 | struct drm_i915_gem_object *obj, *on; |
673a394b | 4255 | |
fbbd37b3 CW |
4256 | mutex_lock(&i915->drm.struct_mutex); |
4257 | intel_runtime_pm_get(i915); | |
4258 | llist_for_each_entry(obj, freed, freed) { | |
4259 | struct i915_vma *vma, *vn; | |
4260 | ||
4261 | trace_i915_gem_object_destroy(obj); | |
4262 | ||
4263 | GEM_BUG_ON(i915_gem_object_is_active(obj)); | |
4264 | list_for_each_entry_safe(vma, vn, | |
4265 | &obj->vma_list, obj_link) { | |
4266 | GEM_BUG_ON(!i915_vma_is_ggtt(vma)); | |
4267 | GEM_BUG_ON(i915_vma_is_active(vma)); | |
4268 | vma->flags &= ~I915_VMA_PIN_MASK; | |
4269 | i915_vma_close(vma); | |
4270 | } | |
db6c2b41 CW |
4271 | GEM_BUG_ON(!list_empty(&obj->vma_list)); |
4272 | GEM_BUG_ON(!RB_EMPTY_ROOT(&obj->vma_tree)); | |
fbbd37b3 | 4273 | |
56cea323 | 4274 | list_del(&obj->global_link); |
fbbd37b3 CW |
4275 | } |
4276 | intel_runtime_pm_put(i915); | |
4277 | mutex_unlock(&i915->drm.struct_mutex); | |
4278 | ||
4279 | llist_for_each_entry_safe(obj, on, freed, freed) { | |
4280 | GEM_BUG_ON(obj->bind_count); | |
4281 | GEM_BUG_ON(atomic_read(&obj->frontbuffer_bits)); | |
4282 | ||
4283 | if (obj->ops->release) | |
4284 | obj->ops->release(obj); | |
f65c9168 | 4285 | |
fbbd37b3 CW |
4286 | if (WARN_ON(i915_gem_object_has_pinned_pages(obj))) |
4287 | atomic_set(&obj->mm.pages_pin_count, 0); | |
548625ee | 4288 | __i915_gem_object_put_pages(obj, I915_MM_NORMAL); |
fbbd37b3 CW |
4289 | GEM_BUG_ON(obj->mm.pages); |
4290 | ||
4291 | if (obj->base.import_attach) | |
4292 | drm_prime_gem_destroy(&obj->base, NULL); | |
4293 | ||
d07f0e59 | 4294 | reservation_object_fini(&obj->__builtin_resv); |
fbbd37b3 CW |
4295 | drm_gem_object_release(&obj->base); |
4296 | i915_gem_info_remove_obj(i915, obj->base.size); | |
4297 | ||
4298 | kfree(obj->bit_17); | |
4299 | i915_gem_object_free(obj); | |
4300 | } | |
4301 | } | |
4302 | ||
4303 | static void i915_gem_flush_free_objects(struct drm_i915_private *i915) | |
4304 | { | |
4305 | struct llist_node *freed; | |
4306 | ||
4307 | freed = llist_del_all(&i915->mm.free_list); | |
4308 | if (unlikely(freed)) | |
4309 | __i915_gem_free_objects(i915, freed); | |
4310 | } | |
4311 | ||
4312 | static void __i915_gem_free_work(struct work_struct *work) | |
4313 | { | |
4314 | struct drm_i915_private *i915 = | |
4315 | container_of(work, struct drm_i915_private, mm.free_work); | |
4316 | struct llist_node *freed; | |
26e12f89 | 4317 | |
b1f788c6 CW |
4318 | /* All file-owned VMA should have been released by this point through |
4319 | * i915_gem_close_object(), or earlier by i915_gem_context_close(). | |
4320 | * However, the object may also be bound into the global GTT (e.g. | |
4321 | * older GPUs without per-process support, or for direct access through | |
4322 | * the GTT either for the user or for scanout). Those VMA still need to | |
4323 | * unbound now. | |
4324 | */ | |
1488fc08 | 4325 | |
fbbd37b3 CW |
4326 | while ((freed = llist_del_all(&i915->mm.free_list))) |
4327 | __i915_gem_free_objects(i915, freed); | |
4328 | } | |
a071fa00 | 4329 | |
fbbd37b3 CW |
4330 | static void __i915_gem_free_object_rcu(struct rcu_head *head) |
4331 | { | |
4332 | struct drm_i915_gem_object *obj = | |
4333 | container_of(head, typeof(*obj), rcu); | |
4334 | struct drm_i915_private *i915 = to_i915(obj->base.dev); | |
4335 | ||
4336 | /* We can't simply use call_rcu() from i915_gem_free_object() | |
4337 | * as we need to block whilst unbinding, and the call_rcu | |
4338 | * task may be called from softirq context. So we take a | |
4339 | * detour through a worker. | |
4340 | */ | |
4341 | if (llist_add(&obj->freed, &i915->mm.free_list)) | |
4342 | schedule_work(&i915->mm.free_work); | |
4343 | } | |
656bfa3a | 4344 | |
fbbd37b3 CW |
4345 | void i915_gem_free_object(struct drm_gem_object *gem_obj) |
4346 | { | |
4347 | struct drm_i915_gem_object *obj = to_intel_bo(gem_obj); | |
a4f5ea64 | 4348 | |
bc0629a7 CW |
4349 | if (obj->mm.quirked) |
4350 | __i915_gem_object_unpin_pages(obj); | |
4351 | ||
340fbd8c | 4352 | if (discard_backing_storage(obj)) |
a4f5ea64 | 4353 | obj->mm.madv = I915_MADV_DONTNEED; |
de151cf6 | 4354 | |
fbbd37b3 CW |
4355 | /* Before we free the object, make sure any pure RCU-only |
4356 | * read-side critical sections are complete, e.g. | |
4357 | * i915_gem_busy_ioctl(). For the corresponding synchronized | |
4358 | * lookup see i915_gem_object_lookup_rcu(). | |
4359 | */ | |
4360 | call_rcu(&obj->rcu, __i915_gem_free_object_rcu); | |
673a394b EA |
4361 | } |
4362 | ||
f8a7fde4 CW |
4363 | void __i915_gem_object_release_unless_active(struct drm_i915_gem_object *obj) |
4364 | { | |
4365 | lockdep_assert_held(&obj->base.dev->struct_mutex); | |
4366 | ||
4367 | GEM_BUG_ON(i915_gem_object_has_active_reference(obj)); | |
4368 | if (i915_gem_object_is_active(obj)) | |
4369 | i915_gem_object_set_active_reference(obj); | |
4370 | else | |
4371 | i915_gem_object_put(obj); | |
4372 | } | |
4373 | ||
3033acab CW |
4374 | static void assert_kernel_context_is_current(struct drm_i915_private *dev_priv) |
4375 | { | |
4376 | struct intel_engine_cs *engine; | |
4377 | enum intel_engine_id id; | |
4378 | ||
4379 | for_each_engine(engine, dev_priv, id) | |
f131e356 CW |
4380 | GEM_BUG_ON(engine->last_retired_context && |
4381 | !i915_gem_context_is_kernel(engine->last_retired_context)); | |
3033acab CW |
4382 | } |
4383 | ||
24145517 CW |
4384 | void i915_gem_sanitize(struct drm_i915_private *i915) |
4385 | { | |
4386 | /* | |
4387 | * If we inherit context state from the BIOS or earlier occupants | |
4388 | * of the GPU, the GPU may be in an inconsistent state when we | |
4389 | * try to take over. The only way to remove the earlier state | |
4390 | * is by resetting. However, resetting on earlier gen is tricky as | |
4391 | * it may impact the display and we are uncertain about the stability | |
4392 | * of the reset, so we only reset recent machines with logical | |
4393 | * context support (that must be reset to remove any stray contexts). | |
4394 | */ | |
4395 | if (HAS_HW_CONTEXTS(i915)) { | |
4396 | int reset = intel_gpu_reset(i915, ALL_ENGINES); | |
4397 | WARN_ON(reset && reset != -ENODEV); | |
4398 | } | |
4399 | } | |
4400 | ||
bf9e8429 | 4401 | int i915_gem_suspend(struct drm_i915_private *dev_priv) |
29105ccc | 4402 | { |
bf9e8429 | 4403 | struct drm_device *dev = &dev_priv->drm; |
dcff85c8 | 4404 | int ret; |
28dfe52a | 4405 | |
c998e8a0 | 4406 | intel_runtime_pm_get(dev_priv); |
54b4f68f CW |
4407 | intel_suspend_gt_powersave(dev_priv); |
4408 | ||
45c5f202 | 4409 | mutex_lock(&dev->struct_mutex); |
5ab57c70 CW |
4410 | |
4411 | /* We have to flush all the executing contexts to main memory so | |
4412 | * that they can saved in the hibernation image. To ensure the last | |
4413 | * context image is coherent, we have to switch away from it. That | |
4414 | * leaves the dev_priv->kernel_context still active when | |
4415 | * we actually suspend, and its image in memory may not match the GPU | |
4416 | * state. Fortunately, the kernel_context is disposable and we do | |
4417 | * not rely on its state. | |
4418 | */ | |
4419 | ret = i915_gem_switch_to_kernel_context(dev_priv); | |
4420 | if (ret) | |
c998e8a0 | 4421 | goto err_unlock; |
5ab57c70 | 4422 | |
22dd3bb9 CW |
4423 | ret = i915_gem_wait_for_idle(dev_priv, |
4424 | I915_WAIT_INTERRUPTIBLE | | |
4425 | I915_WAIT_LOCKED); | |
f7403347 | 4426 | if (ret) |
c998e8a0 | 4427 | goto err_unlock; |
f7403347 | 4428 | |
c033666a | 4429 | i915_gem_retire_requests(dev_priv); |
28176ef4 | 4430 | GEM_BUG_ON(dev_priv->gt.active_requests); |
673a394b | 4431 | |
3033acab | 4432 | assert_kernel_context_is_current(dev_priv); |
b2e862d0 | 4433 | i915_gem_context_lost(dev_priv); |
45c5f202 CW |
4434 | mutex_unlock(&dev->struct_mutex); |
4435 | ||
737b1506 | 4436 | cancel_delayed_work_sync(&dev_priv->gpu_error.hangcheck_work); |
67d97da3 | 4437 | cancel_delayed_work_sync(&dev_priv->gt.retire_work); |
bdeb9785 CW |
4438 | |
4439 | /* As the idle_work is rearming if it detects a race, play safe and | |
4440 | * repeat the flush until it is definitely idle. | |
4441 | */ | |
4442 | while (flush_delayed_work(&dev_priv->gt.idle_work)) | |
4443 | ; | |
4444 | ||
4445 | i915_gem_drain_freed_objects(dev_priv); | |
29105ccc | 4446 | |
bdcf120b CW |
4447 | /* Assert that we sucessfully flushed all the work and |
4448 | * reset the GPU back to its idle, low power state. | |
4449 | */ | |
67d97da3 | 4450 | WARN_ON(dev_priv->gt.awake); |
05425249 | 4451 | WARN_ON(!intel_engines_are_idle(dev_priv)); |
bdcf120b | 4452 | |
1c777c5d ID |
4453 | /* |
4454 | * Neither the BIOS, ourselves or any other kernel | |
4455 | * expects the system to be in execlists mode on startup, | |
4456 | * so we need to reset the GPU back to legacy mode. And the only | |
4457 | * known way to disable logical contexts is through a GPU reset. | |
4458 | * | |
4459 | * So in order to leave the system in a known default configuration, | |
4460 | * always reset the GPU upon unload and suspend. Afterwards we then | |
4461 | * clean up the GEM state tracking, flushing off the requests and | |
4462 | * leaving the system in a known idle state. | |
4463 | * | |
4464 | * Note that is of the upmost importance that the GPU is idle and | |
4465 | * all stray writes are flushed *before* we dismantle the backing | |
4466 | * storage for the pinned objects. | |
4467 | * | |
4468 | * However, since we are uncertain that resetting the GPU on older | |
4469 | * machines is a good idea, we don't - just in case it leaves the | |
4470 | * machine in an unusable condition. | |
4471 | */ | |
24145517 | 4472 | i915_gem_sanitize(dev_priv); |
c998e8a0 | 4473 | goto out_rpm_put; |
1c777c5d | 4474 | |
c998e8a0 | 4475 | err_unlock: |
45c5f202 | 4476 | mutex_unlock(&dev->struct_mutex); |
c998e8a0 CW |
4477 | out_rpm_put: |
4478 | intel_runtime_pm_put(dev_priv); | |
45c5f202 | 4479 | return ret; |
673a394b EA |
4480 | } |
4481 | ||
bf9e8429 | 4482 | void i915_gem_resume(struct drm_i915_private *dev_priv) |
5ab57c70 | 4483 | { |
bf9e8429 | 4484 | struct drm_device *dev = &dev_priv->drm; |
5ab57c70 | 4485 | |
31ab49ab ID |
4486 | WARN_ON(dev_priv->gt.awake); |
4487 | ||
5ab57c70 | 4488 | mutex_lock(&dev->struct_mutex); |
275a991c | 4489 | i915_gem_restore_gtt_mappings(dev_priv); |
5ab57c70 CW |
4490 | |
4491 | /* As we didn't flush the kernel context before suspend, we cannot | |
4492 | * guarantee that the context image is complete. So let's just reset | |
4493 | * it and start again. | |
4494 | */ | |
821ed7df | 4495 | dev_priv->gt.resume(dev_priv); |
5ab57c70 CW |
4496 | |
4497 | mutex_unlock(&dev->struct_mutex); | |
4498 | } | |
4499 | ||
c6be607a | 4500 | void i915_gem_init_swizzling(struct drm_i915_private *dev_priv) |
f691e2f4 | 4501 | { |
c6be607a | 4502 | if (INTEL_GEN(dev_priv) < 5 || |
f691e2f4 DV |
4503 | dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE) |
4504 | return; | |
4505 | ||
4506 | I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) | | |
4507 | DISP_TILE_SURFACE_SWIZZLING); | |
4508 | ||
5db94019 | 4509 | if (IS_GEN5(dev_priv)) |
11782b02 DV |
4510 | return; |
4511 | ||
f691e2f4 | 4512 | I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL); |
5db94019 | 4513 | if (IS_GEN6(dev_priv)) |
6b26c86d | 4514 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB)); |
5db94019 | 4515 | else if (IS_GEN7(dev_priv)) |
6b26c86d | 4516 | I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB)); |
5db94019 | 4517 | else if (IS_GEN8(dev_priv)) |
31a5336e | 4518 | I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW)); |
8782e26c BW |
4519 | else |
4520 | BUG(); | |
f691e2f4 | 4521 | } |
e21af88d | 4522 | |
50a0bc90 | 4523 | static void init_unused_ring(struct drm_i915_private *dev_priv, u32 base) |
81e7f200 | 4524 | { |
81e7f200 VS |
4525 | I915_WRITE(RING_CTL(base), 0); |
4526 | I915_WRITE(RING_HEAD(base), 0); | |
4527 | I915_WRITE(RING_TAIL(base), 0); | |
4528 | I915_WRITE(RING_START(base), 0); | |
4529 | } | |
4530 | ||
50a0bc90 | 4531 | static void init_unused_rings(struct drm_i915_private *dev_priv) |
81e7f200 | 4532 | { |
50a0bc90 TU |
4533 | if (IS_I830(dev_priv)) { |
4534 | init_unused_ring(dev_priv, PRB1_BASE); | |
4535 | init_unused_ring(dev_priv, SRB0_BASE); | |
4536 | init_unused_ring(dev_priv, SRB1_BASE); | |
4537 | init_unused_ring(dev_priv, SRB2_BASE); | |
4538 | init_unused_ring(dev_priv, SRB3_BASE); | |
4539 | } else if (IS_GEN2(dev_priv)) { | |
4540 | init_unused_ring(dev_priv, SRB0_BASE); | |
4541 | init_unused_ring(dev_priv, SRB1_BASE); | |
4542 | } else if (IS_GEN3(dev_priv)) { | |
4543 | init_unused_ring(dev_priv, PRB1_BASE); | |
4544 | init_unused_ring(dev_priv, PRB2_BASE); | |
81e7f200 VS |
4545 | } |
4546 | } | |
4547 | ||
20a8a74a | 4548 | static int __i915_gem_restart_engines(void *data) |
4fc7c971 | 4549 | { |
20a8a74a | 4550 | struct drm_i915_private *i915 = data; |
e2f80391 | 4551 | struct intel_engine_cs *engine; |
3b3f1650 | 4552 | enum intel_engine_id id; |
20a8a74a CW |
4553 | int err; |
4554 | ||
4555 | for_each_engine(engine, i915, id) { | |
4556 | err = engine->init_hw(engine); | |
4557 | if (err) | |
4558 | return err; | |
4559 | } | |
4560 | ||
4561 | return 0; | |
4562 | } | |
4563 | ||
4564 | int i915_gem_init_hw(struct drm_i915_private *dev_priv) | |
4565 | { | |
d200cda6 | 4566 | int ret; |
4fc7c971 | 4567 | |
de867c20 CW |
4568 | dev_priv->gt.last_init_time = ktime_get(); |
4569 | ||
5e4f5189 CW |
4570 | /* Double layer security blanket, see i915_gem_init() */ |
4571 | intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL); | |
4572 | ||
0031fb96 | 4573 | if (HAS_EDRAM(dev_priv) && INTEL_GEN(dev_priv) < 9) |
05e21cc4 | 4574 | I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf)); |
4fc7c971 | 4575 | |
772c2a51 | 4576 | if (IS_HASWELL(dev_priv)) |
50a0bc90 | 4577 | I915_WRITE(MI_PREDICATE_RESULT_2, IS_HSW_GT3(dev_priv) ? |
0bf21347 | 4578 | LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED); |
9435373e | 4579 | |
6e266956 | 4580 | if (HAS_PCH_NOP(dev_priv)) { |
fd6b8f43 | 4581 | if (IS_IVYBRIDGE(dev_priv)) { |
6ba844b0 DV |
4582 | u32 temp = I915_READ(GEN7_MSG_CTL); |
4583 | temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK); | |
4584 | I915_WRITE(GEN7_MSG_CTL, temp); | |
c6be607a | 4585 | } else if (INTEL_GEN(dev_priv) >= 7) { |
6ba844b0 DV |
4586 | u32 temp = I915_READ(HSW_NDE_RSTWRN_OPT); |
4587 | temp &= ~RESET_PCH_HANDSHAKE_ENABLE; | |
4588 | I915_WRITE(HSW_NDE_RSTWRN_OPT, temp); | |
4589 | } | |
88a2b2a3 BW |
4590 | } |
4591 | ||
c6be607a | 4592 | i915_gem_init_swizzling(dev_priv); |
4fc7c971 | 4593 | |
d5abdfda DV |
4594 | /* |
4595 | * At least 830 can leave some of the unused rings | |
4596 | * "active" (ie. head != tail) after resume which | |
4597 | * will prevent c3 entry. Makes sure all unused rings | |
4598 | * are totally idle. | |
4599 | */ | |
50a0bc90 | 4600 | init_unused_rings(dev_priv); |
d5abdfda | 4601 | |
ed54c1a1 | 4602 | BUG_ON(!dev_priv->kernel_context); |
90638cc1 | 4603 | |
c6be607a | 4604 | ret = i915_ppgtt_init_hw(dev_priv); |
4ad2fd88 JH |
4605 | if (ret) { |
4606 | DRM_ERROR("PPGTT enable HW failed %d\n", ret); | |
4607 | goto out; | |
4608 | } | |
4609 | ||
4610 | /* Need to do basic initialisation of all rings first: */ | |
20a8a74a CW |
4611 | ret = __i915_gem_restart_engines(dev_priv); |
4612 | if (ret) | |
4613 | goto out; | |
99433931 | 4614 | |
bf9e8429 | 4615 | intel_mocs_init_l3cc_table(dev_priv); |
0ccdacf6 | 4616 | |
b8991403 OM |
4617 | /* We can't enable contexts until all firmware is loaded */ |
4618 | ret = intel_uc_init_hw(dev_priv); | |
4619 | if (ret) | |
4620 | goto out; | |
33a732f4 | 4621 | |
5e4f5189 CW |
4622 | out: |
4623 | intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL); | |
2fa48d8d | 4624 | return ret; |
8187a2b7 ZN |
4625 | } |
4626 | ||
39df9190 CW |
4627 | bool intel_sanitize_semaphores(struct drm_i915_private *dev_priv, int value) |
4628 | { | |
4629 | if (INTEL_INFO(dev_priv)->gen < 6) | |
4630 | return false; | |
4631 | ||
4632 | /* TODO: make semaphores and Execlists play nicely together */ | |
4633 | if (i915.enable_execlists) | |
4634 | return false; | |
4635 | ||
4636 | if (value >= 0) | |
4637 | return value; | |
4638 | ||
4639 | #ifdef CONFIG_INTEL_IOMMU | |
4640 | /* Enable semaphores on SNB when IO remapping is off */ | |
4641 | if (INTEL_INFO(dev_priv)->gen == 6 && intel_iommu_gfx_mapped) | |
4642 | return false; | |
4643 | #endif | |
4644 | ||
4645 | return true; | |
4646 | } | |
4647 | ||
bf9e8429 | 4648 | int i915_gem_init(struct drm_i915_private *dev_priv) |
1070a42b | 4649 | { |
1070a42b CW |
4650 | int ret; |
4651 | ||
bf9e8429 | 4652 | mutex_lock(&dev_priv->drm.struct_mutex); |
d62b4892 | 4653 | |
57822dc6 CW |
4654 | i915_gem_clflush_init(dev_priv); |
4655 | ||
a83014d3 | 4656 | if (!i915.enable_execlists) { |
821ed7df | 4657 | dev_priv->gt.resume = intel_legacy_submission_resume; |
7e37f889 | 4658 | dev_priv->gt.cleanup_engine = intel_engine_cleanup; |
454afebd | 4659 | } else { |
821ed7df | 4660 | dev_priv->gt.resume = intel_lr_context_resume; |
117897f4 | 4661 | dev_priv->gt.cleanup_engine = intel_logical_ring_cleanup; |
a83014d3 OM |
4662 | } |
4663 | ||
5e4f5189 CW |
4664 | /* This is just a security blanket to placate dragons. |
4665 | * On some systems, we very sporadically observe that the first TLBs | |
4666 | * used by the CS may be stale, despite us poking the TLB reset. If | |
4667 | * we hold the forcewake during initialisation these problems | |
4668 | * just magically go away. | |
4669 | */ | |
4670 | intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL); | |
4671 | ||
72778cb2 | 4672 | i915_gem_init_userptr(dev_priv); |
f6b9d5ca CW |
4673 | |
4674 | ret = i915_gem_init_ggtt(dev_priv); | |
4675 | if (ret) | |
4676 | goto out_unlock; | |
d62b4892 | 4677 | |
bf9e8429 | 4678 | ret = i915_gem_context_init(dev_priv); |
7bcc3777 JN |
4679 | if (ret) |
4680 | goto out_unlock; | |
2fa48d8d | 4681 | |
bf9e8429 | 4682 | ret = intel_engines_init(dev_priv); |
35a57ffb | 4683 | if (ret) |
7bcc3777 | 4684 | goto out_unlock; |
2fa48d8d | 4685 | |
bf9e8429 | 4686 | ret = i915_gem_init_hw(dev_priv); |
60990320 | 4687 | if (ret == -EIO) { |
7e21d648 | 4688 | /* Allow engine initialisation to fail by marking the GPU as |
60990320 CW |
4689 | * wedged. But we only want to do this where the GPU is angry, |
4690 | * for all other failure, such as an allocation failure, bail. | |
4691 | */ | |
4692 | DRM_ERROR("Failed to initialize GPU, declaring it wedged\n"); | |
821ed7df | 4693 | i915_gem_set_wedged(dev_priv); |
60990320 | 4694 | ret = 0; |
1070a42b | 4695 | } |
7bcc3777 JN |
4696 | |
4697 | out_unlock: | |
5e4f5189 | 4698 | intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL); |
bf9e8429 | 4699 | mutex_unlock(&dev_priv->drm.struct_mutex); |
1070a42b | 4700 | |
60990320 | 4701 | return ret; |
1070a42b CW |
4702 | } |
4703 | ||
24145517 CW |
4704 | void i915_gem_init_mmio(struct drm_i915_private *i915) |
4705 | { | |
4706 | i915_gem_sanitize(i915); | |
4707 | } | |
4708 | ||
8187a2b7 | 4709 | void |
cb15d9f8 | 4710 | i915_gem_cleanup_engines(struct drm_i915_private *dev_priv) |
8187a2b7 | 4711 | { |
e2f80391 | 4712 | struct intel_engine_cs *engine; |
3b3f1650 | 4713 | enum intel_engine_id id; |
8187a2b7 | 4714 | |
3b3f1650 | 4715 | for_each_engine(engine, dev_priv, id) |
117897f4 | 4716 | dev_priv->gt.cleanup_engine(engine); |
8187a2b7 ZN |
4717 | } |
4718 | ||
40ae4e16 ID |
4719 | void |
4720 | i915_gem_load_init_fences(struct drm_i915_private *dev_priv) | |
4721 | { | |
49ef5294 | 4722 | int i; |
40ae4e16 ID |
4723 | |
4724 | if (INTEL_INFO(dev_priv)->gen >= 7 && !IS_VALLEYVIEW(dev_priv) && | |
4725 | !IS_CHERRYVIEW(dev_priv)) | |
4726 | dev_priv->num_fence_regs = 32; | |
73f67aa8 JN |
4727 | else if (INTEL_INFO(dev_priv)->gen >= 4 || |
4728 | IS_I945G(dev_priv) || IS_I945GM(dev_priv) || | |
4729 | IS_G33(dev_priv) || IS_PINEVIEW(dev_priv)) | |
40ae4e16 ID |
4730 | dev_priv->num_fence_regs = 16; |
4731 | else | |
4732 | dev_priv->num_fence_regs = 8; | |
4733 | ||
c033666a | 4734 | if (intel_vgpu_active(dev_priv)) |
40ae4e16 ID |
4735 | dev_priv->num_fence_regs = |
4736 | I915_READ(vgtif_reg(avail_rs.fence_num)); | |
4737 | ||
4738 | /* Initialize fence registers to zero */ | |
49ef5294 CW |
4739 | for (i = 0; i < dev_priv->num_fence_regs; i++) { |
4740 | struct drm_i915_fence_reg *fence = &dev_priv->fence_regs[i]; | |
4741 | ||
4742 | fence->i915 = dev_priv; | |
4743 | fence->id = i; | |
4744 | list_add_tail(&fence->link, &dev_priv->mm.fence_list); | |
4745 | } | |
4362f4f6 | 4746 | i915_gem_restore_fences(dev_priv); |
40ae4e16 | 4747 | |
4362f4f6 | 4748 | i915_gem_detect_bit_6_swizzle(dev_priv); |
40ae4e16 ID |
4749 | } |
4750 | ||
73cb9701 | 4751 | int |
cb15d9f8 | 4752 | i915_gem_load_init(struct drm_i915_private *dev_priv) |
673a394b | 4753 | { |
a933568e | 4754 | int err = -ENOMEM; |
42dcedd4 | 4755 | |
a933568e TU |
4756 | dev_priv->objects = KMEM_CACHE(drm_i915_gem_object, SLAB_HWCACHE_ALIGN); |
4757 | if (!dev_priv->objects) | |
73cb9701 | 4758 | goto err_out; |
73cb9701 | 4759 | |
a933568e TU |
4760 | dev_priv->vmas = KMEM_CACHE(i915_vma, SLAB_HWCACHE_ALIGN); |
4761 | if (!dev_priv->vmas) | |
73cb9701 | 4762 | goto err_objects; |
73cb9701 | 4763 | |
a933568e TU |
4764 | dev_priv->requests = KMEM_CACHE(drm_i915_gem_request, |
4765 | SLAB_HWCACHE_ALIGN | | |
4766 | SLAB_RECLAIM_ACCOUNT | | |
4767 | SLAB_DESTROY_BY_RCU); | |
4768 | if (!dev_priv->requests) | |
73cb9701 | 4769 | goto err_vmas; |
73cb9701 | 4770 | |
52e54209 CW |
4771 | dev_priv->dependencies = KMEM_CACHE(i915_dependency, |
4772 | SLAB_HWCACHE_ALIGN | | |
4773 | SLAB_RECLAIM_ACCOUNT); | |
4774 | if (!dev_priv->dependencies) | |
4775 | goto err_requests; | |
4776 | ||
73cb9701 CW |
4777 | mutex_lock(&dev_priv->drm.struct_mutex); |
4778 | INIT_LIST_HEAD(&dev_priv->gt.timelines); | |
bb89485e | 4779 | err = i915_gem_timeline_init__global(dev_priv); |
73cb9701 CW |
4780 | mutex_unlock(&dev_priv->drm.struct_mutex); |
4781 | if (err) | |
52e54209 | 4782 | goto err_dependencies; |
673a394b | 4783 | |
a33afea5 | 4784 | INIT_LIST_HEAD(&dev_priv->context_list); |
fbbd37b3 CW |
4785 | INIT_WORK(&dev_priv->mm.free_work, __i915_gem_free_work); |
4786 | init_llist_head(&dev_priv->mm.free_list); | |
6c085a72 CW |
4787 | INIT_LIST_HEAD(&dev_priv->mm.unbound_list); |
4788 | INIT_LIST_HEAD(&dev_priv->mm.bound_list); | |
a09ba7fa | 4789 | INIT_LIST_HEAD(&dev_priv->mm.fence_list); |
275f039d | 4790 | INIT_LIST_HEAD(&dev_priv->mm.userfault_list); |
67d97da3 | 4791 | INIT_DELAYED_WORK(&dev_priv->gt.retire_work, |
673a394b | 4792 | i915_gem_retire_work_handler); |
67d97da3 | 4793 | INIT_DELAYED_WORK(&dev_priv->gt.idle_work, |
b29c19b6 | 4794 | i915_gem_idle_work_handler); |
1f15b76f | 4795 | init_waitqueue_head(&dev_priv->gpu_error.wait_queue); |
1f83fee0 | 4796 | init_waitqueue_head(&dev_priv->gpu_error.reset_queue); |
31169714 | 4797 | |
6b95a207 | 4798 | init_waitqueue_head(&dev_priv->pending_flip_queue); |
17250b71 | 4799 | |
ce453d81 CW |
4800 | dev_priv->mm.interruptible = true; |
4801 | ||
6f633402 JL |
4802 | atomic_set(&dev_priv->mm.bsd_engine_dispatch_index, 0); |
4803 | ||
b5add959 | 4804 | spin_lock_init(&dev_priv->fb_tracking.lock); |
73cb9701 CW |
4805 | |
4806 | return 0; | |
4807 | ||
52e54209 CW |
4808 | err_dependencies: |
4809 | kmem_cache_destroy(dev_priv->dependencies); | |
73cb9701 CW |
4810 | err_requests: |
4811 | kmem_cache_destroy(dev_priv->requests); | |
4812 | err_vmas: | |
4813 | kmem_cache_destroy(dev_priv->vmas); | |
4814 | err_objects: | |
4815 | kmem_cache_destroy(dev_priv->objects); | |
4816 | err_out: | |
4817 | return err; | |
673a394b | 4818 | } |
71acb5eb | 4819 | |
cb15d9f8 | 4820 | void i915_gem_load_cleanup(struct drm_i915_private *dev_priv) |
d64aa096 | 4821 | { |
c4d4c1c6 | 4822 | i915_gem_drain_freed_objects(dev_priv); |
7d5d59e5 | 4823 | WARN_ON(!llist_empty(&dev_priv->mm.free_list)); |
c4d4c1c6 | 4824 | WARN_ON(dev_priv->mm.object_count); |
7d5d59e5 | 4825 | |
ea84aa77 MA |
4826 | mutex_lock(&dev_priv->drm.struct_mutex); |
4827 | i915_gem_timeline_fini(&dev_priv->gt.global_timeline); | |
4828 | WARN_ON(!list_empty(&dev_priv->gt.timelines)); | |
4829 | mutex_unlock(&dev_priv->drm.struct_mutex); | |
4830 | ||
52e54209 | 4831 | kmem_cache_destroy(dev_priv->dependencies); |
d64aa096 ID |
4832 | kmem_cache_destroy(dev_priv->requests); |
4833 | kmem_cache_destroy(dev_priv->vmas); | |
4834 | kmem_cache_destroy(dev_priv->objects); | |
0eafec6d CW |
4835 | |
4836 | /* And ensure that our DESTROY_BY_RCU slabs are truly destroyed */ | |
4837 | rcu_barrier(); | |
d64aa096 ID |
4838 | } |
4839 | ||
6a800eab CW |
4840 | int i915_gem_freeze(struct drm_i915_private *dev_priv) |
4841 | { | |
6a800eab CW |
4842 | mutex_lock(&dev_priv->drm.struct_mutex); |
4843 | i915_gem_shrink_all(dev_priv); | |
4844 | mutex_unlock(&dev_priv->drm.struct_mutex); | |
4845 | ||
6a800eab CW |
4846 | return 0; |
4847 | } | |
4848 | ||
461fb99c CW |
4849 | int i915_gem_freeze_late(struct drm_i915_private *dev_priv) |
4850 | { | |
4851 | struct drm_i915_gem_object *obj; | |
7aab2d53 CW |
4852 | struct list_head *phases[] = { |
4853 | &dev_priv->mm.unbound_list, | |
4854 | &dev_priv->mm.bound_list, | |
4855 | NULL | |
4856 | }, **p; | |
461fb99c CW |
4857 | |
4858 | /* Called just before we write the hibernation image. | |
4859 | * | |
4860 | * We need to update the domain tracking to reflect that the CPU | |
4861 | * will be accessing all the pages to create and restore from the | |
4862 | * hibernation, and so upon restoration those pages will be in the | |
4863 | * CPU domain. | |
4864 | * | |
4865 | * To make sure the hibernation image contains the latest state, | |
4866 | * we update that state just before writing out the image. | |
7aab2d53 CW |
4867 | * |
4868 | * To try and reduce the hibernation image, we manually shrink | |
4869 | * the objects as well. | |
461fb99c CW |
4870 | */ |
4871 | ||
6a800eab CW |
4872 | mutex_lock(&dev_priv->drm.struct_mutex); |
4873 | i915_gem_shrink(dev_priv, -1UL, I915_SHRINK_UNBOUND); | |
461fb99c | 4874 | |
7aab2d53 | 4875 | for (p = phases; *p; p++) { |
56cea323 | 4876 | list_for_each_entry(obj, *p, global_link) { |
7aab2d53 CW |
4877 | obj->base.read_domains = I915_GEM_DOMAIN_CPU; |
4878 | obj->base.write_domain = I915_GEM_DOMAIN_CPU; | |
4879 | } | |
461fb99c | 4880 | } |
6a800eab | 4881 | mutex_unlock(&dev_priv->drm.struct_mutex); |
461fb99c CW |
4882 | |
4883 | return 0; | |
4884 | } | |
4885 | ||
f787a5f5 | 4886 | void i915_gem_release(struct drm_device *dev, struct drm_file *file) |
b962442e | 4887 | { |
f787a5f5 | 4888 | struct drm_i915_file_private *file_priv = file->driver_priv; |
15f7bbc7 | 4889 | struct drm_i915_gem_request *request; |
b962442e EA |
4890 | |
4891 | /* Clean up our request list when the client is going away, so that | |
4892 | * later retire_requests won't dereference our soon-to-be-gone | |
4893 | * file_priv. | |
4894 | */ | |
1c25595f | 4895 | spin_lock(&file_priv->mm.lock); |
c8659efa | 4896 | list_for_each_entry(request, &file_priv->mm.request_list, client_link) |
f787a5f5 | 4897 | request->file_priv = NULL; |
1c25595f | 4898 | spin_unlock(&file_priv->mm.lock); |
b29c19b6 | 4899 | |
2e1b8730 | 4900 | if (!list_empty(&file_priv->rps.link)) { |
8d3afd7d | 4901 | spin_lock(&to_i915(dev)->rps.client_lock); |
2e1b8730 | 4902 | list_del(&file_priv->rps.link); |
8d3afd7d | 4903 | spin_unlock(&to_i915(dev)->rps.client_lock); |
1854d5ca | 4904 | } |
b29c19b6 CW |
4905 | } |
4906 | ||
4907 | int i915_gem_open(struct drm_device *dev, struct drm_file *file) | |
4908 | { | |
4909 | struct drm_i915_file_private *file_priv; | |
e422b888 | 4910 | int ret; |
b29c19b6 | 4911 | |
c4c29d7b | 4912 | DRM_DEBUG("\n"); |
b29c19b6 CW |
4913 | |
4914 | file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL); | |
4915 | if (!file_priv) | |
4916 | return -ENOMEM; | |
4917 | ||
4918 | file->driver_priv = file_priv; | |
f19ec8cb | 4919 | file_priv->dev_priv = to_i915(dev); |
ab0e7ff9 | 4920 | file_priv->file = file; |
2e1b8730 | 4921 | INIT_LIST_HEAD(&file_priv->rps.link); |
b29c19b6 CW |
4922 | |
4923 | spin_lock_init(&file_priv->mm.lock); | |
4924 | INIT_LIST_HEAD(&file_priv->mm.request_list); | |
b29c19b6 | 4925 | |
c80ff16e | 4926 | file_priv->bsd_engine = -1; |
de1add36 | 4927 | |
e422b888 BW |
4928 | ret = i915_gem_context_open(dev, file); |
4929 | if (ret) | |
4930 | kfree(file_priv); | |
b29c19b6 | 4931 | |
e422b888 | 4932 | return ret; |
b29c19b6 CW |
4933 | } |
4934 | ||
b680c37a DV |
4935 | /** |
4936 | * i915_gem_track_fb - update frontbuffer tracking | |
d9072a3e GT |
4937 | * @old: current GEM buffer for the frontbuffer slots |
4938 | * @new: new GEM buffer for the frontbuffer slots | |
4939 | * @frontbuffer_bits: bitmask of frontbuffer slots | |
b680c37a DV |
4940 | * |
4941 | * This updates the frontbuffer tracking bits @frontbuffer_bits by clearing them | |
4942 | * from @old and setting them in @new. Both @old and @new can be NULL. | |
4943 | */ | |
a071fa00 DV |
4944 | void i915_gem_track_fb(struct drm_i915_gem_object *old, |
4945 | struct drm_i915_gem_object *new, | |
4946 | unsigned frontbuffer_bits) | |
4947 | { | |
faf5bf0a CW |
4948 | /* Control of individual bits within the mask are guarded by |
4949 | * the owning plane->mutex, i.e. we can never see concurrent | |
4950 | * manipulation of individual bits. But since the bitfield as a whole | |
4951 | * is updated using RMW, we need to use atomics in order to update | |
4952 | * the bits. | |
4953 | */ | |
4954 | BUILD_BUG_ON(INTEL_FRONTBUFFER_BITS_PER_PIPE * I915_MAX_PIPES > | |
4955 | sizeof(atomic_t) * BITS_PER_BYTE); | |
4956 | ||
a071fa00 | 4957 | if (old) { |
faf5bf0a CW |
4958 | WARN_ON(!(atomic_read(&old->frontbuffer_bits) & frontbuffer_bits)); |
4959 | atomic_andnot(frontbuffer_bits, &old->frontbuffer_bits); | |
a071fa00 DV |
4960 | } |
4961 | ||
4962 | if (new) { | |
faf5bf0a CW |
4963 | WARN_ON(atomic_read(&new->frontbuffer_bits) & frontbuffer_bits); |
4964 | atomic_or(frontbuffer_bits, &new->frontbuffer_bits); | |
a071fa00 DV |
4965 | } |
4966 | } | |
4967 | ||
ea70299d DG |
4968 | /* Allocate a new GEM object and fill it with the supplied data */ |
4969 | struct drm_i915_gem_object * | |
12d79d78 | 4970 | i915_gem_object_create_from_data(struct drm_i915_private *dev_priv, |
ea70299d DG |
4971 | const void *data, size_t size) |
4972 | { | |
4973 | struct drm_i915_gem_object *obj; | |
be062fa4 CW |
4974 | struct file *file; |
4975 | size_t offset; | |
4976 | int err; | |
ea70299d | 4977 | |
12d79d78 | 4978 | obj = i915_gem_object_create(dev_priv, round_up(size, PAGE_SIZE)); |
fe3db79b | 4979 | if (IS_ERR(obj)) |
ea70299d DG |
4980 | return obj; |
4981 | ||
ce8ff099 | 4982 | GEM_BUG_ON(obj->base.write_domain != I915_GEM_DOMAIN_CPU); |
ea70299d | 4983 | |
be062fa4 CW |
4984 | file = obj->base.filp; |
4985 | offset = 0; | |
4986 | do { | |
4987 | unsigned int len = min_t(typeof(size), size, PAGE_SIZE); | |
4988 | struct page *page; | |
4989 | void *pgdata, *vaddr; | |
ea70299d | 4990 | |
be062fa4 CW |
4991 | err = pagecache_write_begin(file, file->f_mapping, |
4992 | offset, len, 0, | |
4993 | &page, &pgdata); | |
4994 | if (err < 0) | |
4995 | goto fail; | |
ea70299d | 4996 | |
be062fa4 CW |
4997 | vaddr = kmap(page); |
4998 | memcpy(vaddr, data, len); | |
4999 | kunmap(page); | |
5000 | ||
5001 | err = pagecache_write_end(file, file->f_mapping, | |
5002 | offset, len, len, | |
5003 | page, pgdata); | |
5004 | if (err < 0) | |
5005 | goto fail; | |
5006 | ||
5007 | size -= len; | |
5008 | data += len; | |
5009 | offset += len; | |
5010 | } while (size); | |
ea70299d DG |
5011 | |
5012 | return obj; | |
5013 | ||
5014 | fail: | |
f8c417cd | 5015 | i915_gem_object_put(obj); |
be062fa4 | 5016 | return ERR_PTR(err); |
ea70299d | 5017 | } |
96d77634 CW |
5018 | |
5019 | struct scatterlist * | |
5020 | i915_gem_object_get_sg(struct drm_i915_gem_object *obj, | |
5021 | unsigned int n, | |
5022 | unsigned int *offset) | |
5023 | { | |
a4f5ea64 | 5024 | struct i915_gem_object_page_iter *iter = &obj->mm.get_page; |
96d77634 CW |
5025 | struct scatterlist *sg; |
5026 | unsigned int idx, count; | |
5027 | ||
5028 | might_sleep(); | |
5029 | GEM_BUG_ON(n >= obj->base.size >> PAGE_SHIFT); | |
a4f5ea64 | 5030 | GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj)); |
96d77634 CW |
5031 | |
5032 | /* As we iterate forward through the sg, we record each entry in a | |
5033 | * radixtree for quick repeated (backwards) lookups. If we have seen | |
5034 | * this index previously, we will have an entry for it. | |
5035 | * | |
5036 | * Initial lookup is O(N), but this is amortized to O(1) for | |
5037 | * sequential page access (where each new request is consecutive | |
5038 | * to the previous one). Repeated lookups are O(lg(obj->base.size)), | |
5039 | * i.e. O(1) with a large constant! | |
5040 | */ | |
5041 | if (n < READ_ONCE(iter->sg_idx)) | |
5042 | goto lookup; | |
5043 | ||
5044 | mutex_lock(&iter->lock); | |
5045 | ||
5046 | /* We prefer to reuse the last sg so that repeated lookup of this | |
5047 | * (or the subsequent) sg are fast - comparing against the last | |
5048 | * sg is faster than going through the radixtree. | |
5049 | */ | |
5050 | ||
5051 | sg = iter->sg_pos; | |
5052 | idx = iter->sg_idx; | |
5053 | count = __sg_page_count(sg); | |
5054 | ||
5055 | while (idx + count <= n) { | |
5056 | unsigned long exception, i; | |
5057 | int ret; | |
5058 | ||
5059 | /* If we cannot allocate and insert this entry, or the | |
5060 | * individual pages from this range, cancel updating the | |
5061 | * sg_idx so that on this lookup we are forced to linearly | |
5062 | * scan onwards, but on future lookups we will try the | |
5063 | * insertion again (in which case we need to be careful of | |
5064 | * the error return reporting that we have already inserted | |
5065 | * this index). | |
5066 | */ | |
5067 | ret = radix_tree_insert(&iter->radix, idx, sg); | |
5068 | if (ret && ret != -EEXIST) | |
5069 | goto scan; | |
5070 | ||
5071 | exception = | |
5072 | RADIX_TREE_EXCEPTIONAL_ENTRY | | |
5073 | idx << RADIX_TREE_EXCEPTIONAL_SHIFT; | |
5074 | for (i = 1; i < count; i++) { | |
5075 | ret = radix_tree_insert(&iter->radix, idx + i, | |
5076 | (void *)exception); | |
5077 | if (ret && ret != -EEXIST) | |
5078 | goto scan; | |
5079 | } | |
5080 | ||
5081 | idx += count; | |
5082 | sg = ____sg_next(sg); | |
5083 | count = __sg_page_count(sg); | |
5084 | } | |
5085 | ||
5086 | scan: | |
5087 | iter->sg_pos = sg; | |
5088 | iter->sg_idx = idx; | |
5089 | ||
5090 | mutex_unlock(&iter->lock); | |
5091 | ||
5092 | if (unlikely(n < idx)) /* insertion completed by another thread */ | |
5093 | goto lookup; | |
5094 | ||
5095 | /* In case we failed to insert the entry into the radixtree, we need | |
5096 | * to look beyond the current sg. | |
5097 | */ | |
5098 | while (idx + count <= n) { | |
5099 | idx += count; | |
5100 | sg = ____sg_next(sg); | |
5101 | count = __sg_page_count(sg); | |
5102 | } | |
5103 | ||
5104 | *offset = n - idx; | |
5105 | return sg; | |
5106 | ||
5107 | lookup: | |
5108 | rcu_read_lock(); | |
5109 | ||
5110 | sg = radix_tree_lookup(&iter->radix, n); | |
5111 | GEM_BUG_ON(!sg); | |
5112 | ||
5113 | /* If this index is in the middle of multi-page sg entry, | |
5114 | * the radixtree will contain an exceptional entry that points | |
5115 | * to the start of that range. We will return the pointer to | |
5116 | * the base page and the offset of this page within the | |
5117 | * sg entry's range. | |
5118 | */ | |
5119 | *offset = 0; | |
5120 | if (unlikely(radix_tree_exception(sg))) { | |
5121 | unsigned long base = | |
5122 | (unsigned long)sg >> RADIX_TREE_EXCEPTIONAL_SHIFT; | |
5123 | ||
5124 | sg = radix_tree_lookup(&iter->radix, base); | |
5125 | GEM_BUG_ON(!sg); | |
5126 | ||
5127 | *offset = n - base; | |
5128 | } | |
5129 | ||
5130 | rcu_read_unlock(); | |
5131 | ||
5132 | return sg; | |
5133 | } | |
5134 | ||
5135 | struct page * | |
5136 | i915_gem_object_get_page(struct drm_i915_gem_object *obj, unsigned int n) | |
5137 | { | |
5138 | struct scatterlist *sg; | |
5139 | unsigned int offset; | |
5140 | ||
5141 | GEM_BUG_ON(!i915_gem_object_has_struct_page(obj)); | |
5142 | ||
5143 | sg = i915_gem_object_get_sg(obj, n, &offset); | |
5144 | return nth_page(sg_page(sg), offset); | |
5145 | } | |
5146 | ||
5147 | /* Like i915_gem_object_get_page(), but mark the returned page dirty */ | |
5148 | struct page * | |
5149 | i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj, | |
5150 | unsigned int n) | |
5151 | { | |
5152 | struct page *page; | |
5153 | ||
5154 | page = i915_gem_object_get_page(obj, n); | |
a4f5ea64 | 5155 | if (!obj->mm.dirty) |
96d77634 CW |
5156 | set_page_dirty(page); |
5157 | ||
5158 | return page; | |
5159 | } | |
5160 | ||
5161 | dma_addr_t | |
5162 | i915_gem_object_get_dma_address(struct drm_i915_gem_object *obj, | |
5163 | unsigned long n) | |
5164 | { | |
5165 | struct scatterlist *sg; | |
5166 | unsigned int offset; | |
5167 | ||
5168 | sg = i915_gem_object_get_sg(obj, n, &offset); | |
5169 | return sg_dma_address(sg) + (offset << PAGE_SHIFT); | |
5170 | } | |
935a2f77 CW |
5171 | |
5172 | #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) | |
5173 | #include "selftests/scatterlist.c" | |
66d9cb5d | 5174 | #include "selftests/mock_gem_device.c" |
44653988 | 5175 | #include "selftests/huge_gem_object.c" |
8335fd65 | 5176 | #include "selftests/i915_gem_object.c" |
17059450 | 5177 | #include "selftests/i915_gem_coherency.c" |
935a2f77 | 5178 | #endif |