1 /*
2  * Copyright © 2008-2015 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  * Authors:
24  *    Eric Anholt <eric@anholt.net>
25  *
26  */
27 
28 #include <drm/drm_vma_manager.h>
29 #include <linux/dma-fence-array.h>
30 #include <linux/kthread.h>
31 #include <linux/dma-resv.h>
32 #include <linux/shmem_fs.h>
33 #include <linux/slab.h>
34 #include <linux/stop_machine.h>
35 #include <linux/swap.h>
36 #include <linux/pci.h>
37 #include <linux/dma-buf.h>
38 #include <linux/mman.h>
39 
40 #include "display/intel_display.h"
41 #include "display/intel_frontbuffer.h"
42 
43 #include "gem/i915_gem_clflush.h"
44 #include "gem/i915_gem_context.h"
45 #include "gem/i915_gem_ioctls.h"
46 #include "gem/i915_gem_mman.h"
47 #include "gem/i915_gem_region.h"
48 #include "gt/intel_engine_user.h"
49 #include "gt/intel_gt.h"
50 #include "gt/intel_gt_pm.h"
51 #include "gt/intel_workarounds.h"
52 
53 #include "i915_drv.h"
54 #include "i915_trace.h"
55 #include "i915_vgpu.h"
56 
57 #include "intel_pm.h"
58 
59 static int
insert_mappable_node(struct i915_ggtt * ggtt,struct drm_mm_node * node,u32 size)60 insert_mappable_node(struct i915_ggtt *ggtt, struct drm_mm_node *node, u32 size)
61 {
62 	int err;
63 
64 	err = mutex_lock_interruptible(&ggtt->vm.mutex);
65 	if (err)
66 		return err;
67 
68 	memset(node, 0, sizeof(*node));
69 	err = drm_mm_insert_node_in_range(&ggtt->vm.mm, node,
70 					  size, 0, I915_COLOR_UNEVICTABLE,
71 					  0, ggtt->mappable_end,
72 					  DRM_MM_INSERT_LOW);
73 
74 	mutex_unlock(&ggtt->vm.mutex);
75 
76 	return err;
77 }
78 
79 static void
remove_mappable_node(struct i915_ggtt * ggtt,struct drm_mm_node * node)80 remove_mappable_node(struct i915_ggtt *ggtt, struct drm_mm_node *node)
81 {
82 	mutex_lock(&ggtt->vm.mutex);
83 	drm_mm_remove_node(node);
84 	mutex_unlock(&ggtt->vm.mutex);
85 }
86 
87 int
i915_gem_get_aperture_ioctl(struct drm_device * dev,void * data,struct drm_file * file)88 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
89 			    struct drm_file *file)
90 {
91 	struct i915_ggtt *ggtt = &to_i915(dev)->ggtt;
92 	struct drm_i915_gem_get_aperture *args = data;
93 	struct i915_vma *vma;
94 	u64 pinned;
95 
96 	if (mutex_lock_interruptible(&ggtt->vm.mutex))
97 		return -EINTR;
98 
99 	pinned = ggtt->vm.reserved;
100 	list_for_each_entry(vma, &ggtt->vm.bound_list, vm_link)
101 		if (i915_vma_is_pinned(vma))
102 			pinned += vma->node.size;
103 
104 	mutex_unlock(&ggtt->vm.mutex);
105 
106 	args->aper_size = ggtt->vm.total;
107 	args->aper_available_size = args->aper_size - pinned;
108 
109 	return 0;
110 }
111 
i915_gem_object_unbind(struct drm_i915_gem_object * obj,unsigned long flags)112 int i915_gem_object_unbind(struct drm_i915_gem_object *obj,
113 			   unsigned long flags)
114 {
115 	struct intel_runtime_pm *rpm = &to_i915(obj->base.dev)->runtime_pm;
116 	LIST_HEAD(still_in_list);
117 	intel_wakeref_t wakeref;
118 	struct i915_vma *vma;
119 	int ret;
120 
121 	if (list_empty(&obj->vma.list))
122 		return 0;
123 
124 	/*
125 	 * As some machines use ACPI to handle runtime-resume callbacks, and
126 	 * ACPI is quite kmalloc happy, we cannot resume beneath the vm->mutex
127 	 * as they are required by the shrinker. Ergo, we wake the device up
128 	 * first just in case.
129 	 */
130 	wakeref = intel_runtime_pm_get(rpm);
131 
132 try_again:
133 	ret = 0;
134 	spin_lock(&obj->vma.lock);
135 	while (!ret && (vma = list_first_entry_or_null(&obj->vma.list,
136 						       struct i915_vma,
137 						       obj_link))) {
138 		struct i915_address_space *vm = vma->vm;
139 
140 		list_move_tail(&vma->obj_link, &still_in_list);
141 		if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK))
142 			continue;
143 
144 		if (flags & I915_GEM_OBJECT_UNBIND_TEST) {
145 			ret = -EBUSY;
146 			break;
147 		}
148 
149 		ret = -EAGAIN;
150 		if (!i915_vm_tryopen(vm))
151 			break;
152 
153 		/* Prevent vma being freed by i915_vma_parked as we unbind */
154 		vma = __i915_vma_get(vma);
155 		spin_unlock(&obj->vma.lock);
156 
157 		if (vma) {
158 			ret = -EBUSY;
159 			if (flags & I915_GEM_OBJECT_UNBIND_ACTIVE ||
160 			    !i915_vma_is_active(vma)) {
161 				if (flags & I915_GEM_OBJECT_UNBIND_VM_TRYLOCK) {
162 					if (mutex_trylock(&vma->vm->mutex)) {
163 						ret = __i915_vma_unbind(vma);
164 						mutex_unlock(&vma->vm->mutex);
165 					} else {
166 						ret = -EBUSY;
167 					}
168 				} else {
169 					ret = i915_vma_unbind(vma);
170 				}
171 			}
172 
173 			__i915_vma_put(vma);
174 		}
175 
176 		i915_vm_close(vm);
177 		spin_lock(&obj->vma.lock);
178 	}
179 	list_splice_init(&still_in_list, &obj->vma.list);
180 	spin_unlock(&obj->vma.lock);
181 
182 	if (ret == -EAGAIN && flags & I915_GEM_OBJECT_UNBIND_BARRIER) {
183 		rcu_barrier(); /* flush the i915_vm_release() */
184 		goto try_again;
185 	}
186 
187 	intel_runtime_pm_put(rpm, wakeref);
188 
189 	return ret;
190 }
191 
192 static int
shmem_pread(struct page * page,int offset,int len,char __user * user_data,bool needs_clflush)193 shmem_pread(struct page *page, int offset, int len, char __user *user_data,
194 	    bool needs_clflush)
195 {
196 	char *vaddr;
197 	int ret;
198 
199 	vaddr = kmap(page);
200 
201 	if (needs_clflush)
202 		drm_clflush_virt_range(vaddr + offset, len);
203 
204 	ret = __copy_to_user(user_data, vaddr + offset, len);
205 
206 	kunmap(page);
207 
208 	return ret ? -EFAULT : 0;
209 }
210 
211 static int
i915_gem_shmem_pread(struct drm_i915_gem_object * obj,struct drm_i915_gem_pread * args)212 i915_gem_shmem_pread(struct drm_i915_gem_object *obj,
213 		     struct drm_i915_gem_pread *args)
214 {
215 	unsigned int needs_clflush;
216 	unsigned int idx, offset;
217 	char __user *user_data;
218 	u64 remain;
219 	int ret;
220 
221 	ret = i915_gem_object_lock_interruptible(obj, NULL);
222 	if (ret)
223 		return ret;
224 
225 	ret = i915_gem_object_pin_pages(obj);
226 	if (ret)
227 		goto err_unlock;
228 
229 	ret = i915_gem_object_prepare_read(obj, &needs_clflush);
230 	if (ret)
231 		goto err_unpin;
232 
233 	i915_gem_object_finish_access(obj);
234 	i915_gem_object_unlock(obj);
235 
236 	remain = args->size;
237 	user_data = u64_to_user_ptr(args->data_ptr);
238 	offset = offset_in_page(args->offset);
239 	for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
240 		struct page *page = i915_gem_object_get_page(obj, idx);
241 		unsigned int length = min_t(u64, remain, PAGE_SIZE - offset);
242 
243 		ret = shmem_pread(page, offset, length, user_data,
244 				  needs_clflush);
245 		if (ret)
246 			break;
247 
248 		remain -= length;
249 		user_data += length;
250 		offset = 0;
251 	}
252 
253 	i915_gem_object_unpin_pages(obj);
254 	return ret;
255 
256 err_unpin:
257 	i915_gem_object_unpin_pages(obj);
258 err_unlock:
259 	i915_gem_object_unlock(obj);
260 	return ret;
261 }
262 
263 static inline bool
gtt_user_read(struct io_mapping * mapping,loff_t base,int offset,char __user * user_data,int length)264 gtt_user_read(struct io_mapping *mapping,
265 	      loff_t base, int offset,
266 	      char __user *user_data, int length)
267 {
268 	void __iomem *vaddr;
269 	unsigned long unwritten;
270 
271 	/* We can use the cpu mem copy function because this is X86. */
272 	vaddr = io_mapping_map_atomic_wc(mapping, base);
273 	unwritten = __copy_to_user_inatomic(user_data,
274 					    (void __force *)vaddr + offset,
275 					    length);
276 	io_mapping_unmap_atomic(vaddr);
277 	if (unwritten) {
278 		vaddr = io_mapping_map_wc(mapping, base, PAGE_SIZE);
279 		unwritten = copy_to_user(user_data,
280 					 (void __force *)vaddr + offset,
281 					 length);
282 		io_mapping_unmap(vaddr);
283 	}
284 	return unwritten;
285 }
286 
i915_gem_gtt_prepare(struct drm_i915_gem_object * obj,struct drm_mm_node * node,bool write)287 static struct i915_vma *i915_gem_gtt_prepare(struct drm_i915_gem_object *obj,
288 					     struct drm_mm_node *node,
289 					     bool write)
290 {
291 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
292 	struct i915_ggtt *ggtt = &i915->ggtt;
293 	struct i915_vma *vma;
294 	struct i915_gem_ww_ctx ww;
295 	int ret;
296 
297 	i915_gem_ww_ctx_init(&ww, true);
298 retry:
299 	vma = ERR_PTR(-ENODEV);
300 	ret = i915_gem_object_lock(obj, &ww);
301 	if (ret)
302 		goto err_ww;
303 
304 	ret = i915_gem_object_set_to_gtt_domain(obj, write);
305 	if (ret)
306 		goto err_ww;
307 
308 	if (!i915_gem_object_is_tiled(obj))
309 		vma = i915_gem_object_ggtt_pin_ww(obj, &ww, NULL, 0, 0,
310 						  PIN_MAPPABLE |
311 						  PIN_NONBLOCK /* NOWARN */ |
312 						  PIN_NOEVICT);
313 	if (vma == ERR_PTR(-EDEADLK)) {
314 		ret = -EDEADLK;
315 		goto err_ww;
316 	} else if (!IS_ERR(vma)) {
317 		node->start = i915_ggtt_offset(vma);
318 		node->flags = 0;
319 	} else {
320 		ret = insert_mappable_node(ggtt, node, PAGE_SIZE);
321 		if (ret)
322 			goto err_ww;
323 		GEM_BUG_ON(!drm_mm_node_allocated(node));
324 		vma = NULL;
325 	}
326 
327 	ret = i915_gem_object_pin_pages(obj);
328 	if (ret) {
329 		if (drm_mm_node_allocated(node)) {
330 			ggtt->vm.clear_range(&ggtt->vm, node->start, node->size);
331 			remove_mappable_node(ggtt, node);
332 		} else {
333 			i915_vma_unpin(vma);
334 		}
335 	}
336 
337 err_ww:
338 	if (ret == -EDEADLK) {
339 		ret = i915_gem_ww_ctx_backoff(&ww);
340 		if (!ret)
341 			goto retry;
342 	}
343 	i915_gem_ww_ctx_fini(&ww);
344 
345 	return ret ? ERR_PTR(ret) : vma;
346 }
347 
i915_gem_gtt_cleanup(struct drm_i915_gem_object * obj,struct drm_mm_node * node,struct i915_vma * vma)348 static void i915_gem_gtt_cleanup(struct drm_i915_gem_object *obj,
349 				 struct drm_mm_node *node,
350 				 struct i915_vma *vma)
351 {
352 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
353 	struct i915_ggtt *ggtt = &i915->ggtt;
354 
355 	i915_gem_object_unpin_pages(obj);
356 	if (drm_mm_node_allocated(node)) {
357 		ggtt->vm.clear_range(&ggtt->vm, node->start, node->size);
358 		remove_mappable_node(ggtt, node);
359 	} else {
360 		i915_vma_unpin(vma);
361 	}
362 }
363 
364 static int
i915_gem_gtt_pread(struct drm_i915_gem_object * obj,const struct drm_i915_gem_pread * args)365 i915_gem_gtt_pread(struct drm_i915_gem_object *obj,
366 		   const struct drm_i915_gem_pread *args)
367 {
368 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
369 	struct i915_ggtt *ggtt = &i915->ggtt;
370 	intel_wakeref_t wakeref;
371 	struct drm_mm_node node;
372 	void __user *user_data;
373 	struct i915_vma *vma;
374 	u64 remain, offset;
375 	int ret = 0;
376 
377 	wakeref = intel_runtime_pm_get(&i915->runtime_pm);
378 
379 	vma = i915_gem_gtt_prepare(obj, &node, false);
380 	if (IS_ERR(vma)) {
381 		ret = PTR_ERR(vma);
382 		goto out_rpm;
383 	}
384 
385 	user_data = u64_to_user_ptr(args->data_ptr);
386 	remain = args->size;
387 	offset = args->offset;
388 
389 	while (remain > 0) {
390 		/* Operation in this page
391 		 *
392 		 * page_base = page offset within aperture
393 		 * page_offset = offset within page
394 		 * page_length = bytes to copy for this page
395 		 */
396 		u32 page_base = node.start;
397 		unsigned page_offset = offset_in_page(offset);
398 		unsigned page_length = PAGE_SIZE - page_offset;
399 		page_length = remain < page_length ? remain : page_length;
400 		if (drm_mm_node_allocated(&node)) {
401 			ggtt->vm.insert_page(&ggtt->vm,
402 					     i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT),
403 					     node.start, I915_CACHE_NONE, 0);
404 		} else {
405 			page_base += offset & PAGE_MASK;
406 		}
407 
408 		if (gtt_user_read(&ggtt->iomap, page_base, page_offset,
409 				  user_data, page_length)) {
410 			ret = -EFAULT;
411 			break;
412 		}
413 
414 		remain -= page_length;
415 		user_data += page_length;
416 		offset += page_length;
417 	}
418 
419 	i915_gem_gtt_cleanup(obj, &node, vma);
420 out_rpm:
421 	intel_runtime_pm_put(&i915->runtime_pm, wakeref);
422 	return ret;
423 }
424 
425 /**
426  * Reads data from the object referenced by handle.
427  * @dev: drm device pointer
428  * @data: ioctl data blob
429  * @file: drm file pointer
430  *
431  * On error, the contents of *data are undefined.
432  */
433 int
i915_gem_pread_ioctl(struct drm_device * dev,void * data,struct drm_file * file)434 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
435 		     struct drm_file *file)
436 {
437 	struct drm_i915_private *i915 = to_i915(dev);
438 	struct drm_i915_gem_pread *args = data;
439 	struct drm_i915_gem_object *obj;
440 	int ret;
441 
442 	/* PREAD is disallowed for all platforms after TGL-LP.  This also
443 	 * covers all platforms with local memory.
444 	 */
445 	if (GRAPHICS_VER(i915) >= 12 && !IS_TIGERLAKE(i915))
446 		return -EOPNOTSUPP;
447 
448 	if (args->size == 0)
449 		return 0;
450 
451 	if (!access_ok(u64_to_user_ptr(args->data_ptr),
452 		       args->size))
453 		return -EFAULT;
454 
455 	obj = i915_gem_object_lookup(file, args->handle);
456 	if (!obj)
457 		return -ENOENT;
458 
459 	/* Bounds check source.  */
460 	if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) {
461 		ret = -EINVAL;
462 		goto out;
463 	}
464 
465 	trace_i915_gem_object_pread(obj, args->offset, args->size);
466 	ret = -ENODEV;
467 	if (obj->ops->pread)
468 		ret = obj->ops->pread(obj, args);
469 	if (ret != -ENODEV)
470 		goto out;
471 
472 	ret = i915_gem_object_wait(obj,
473 				   I915_WAIT_INTERRUPTIBLE,
474 				   MAX_SCHEDULE_TIMEOUT);
475 	if (ret)
476 		goto out;
477 
478 	ret = i915_gem_shmem_pread(obj, args);
479 	if (ret == -EFAULT || ret == -ENODEV)
480 		ret = i915_gem_gtt_pread(obj, args);
481 
482 out:
483 	i915_gem_object_put(obj);
484 	return ret;
485 }
486 
487 /* This is the fast write path which cannot handle
488  * page faults in the source data
489  */
490 
491 static inline bool
ggtt_write(struct io_mapping * mapping,loff_t base,int offset,char __user * user_data,int length)492 ggtt_write(struct io_mapping *mapping,
493 	   loff_t base, int offset,
494 	   char __user *user_data, int length)
495 {
496 	void __iomem *vaddr;
497 	unsigned long unwritten;
498 
499 	/* We can use the cpu mem copy function because this is X86. */
500 	vaddr = io_mapping_map_atomic_wc(mapping, base);
501 	unwritten = __copy_from_user_inatomic_nocache((void __force *)vaddr + offset,
502 						      user_data, length);
503 	io_mapping_unmap_atomic(vaddr);
504 	if (unwritten) {
505 		vaddr = io_mapping_map_wc(mapping, base, PAGE_SIZE);
506 		unwritten = copy_from_user((void __force *)vaddr + offset,
507 					   user_data, length);
508 		io_mapping_unmap(vaddr);
509 	}
510 
511 	return unwritten;
512 }
513 
514 /**
515  * This is the fast pwrite path, where we copy the data directly from the
516  * user into the GTT, uncached.
517  * @obj: i915 GEM object
518  * @args: pwrite arguments structure
519  */
520 static int
i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object * obj,const struct drm_i915_gem_pwrite * args)521 i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
522 			 const struct drm_i915_gem_pwrite *args)
523 {
524 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
525 	struct i915_ggtt *ggtt = &i915->ggtt;
526 	struct intel_runtime_pm *rpm = &i915->runtime_pm;
527 	intel_wakeref_t wakeref;
528 	struct drm_mm_node node;
529 	struct i915_vma *vma;
530 	u64 remain, offset;
531 	void __user *user_data;
532 	int ret = 0;
533 
534 	if (i915_gem_object_has_struct_page(obj)) {
535 		/*
536 		 * Avoid waking the device up if we can fallback, as
537 		 * waking/resuming is very slow (worst-case 10-100 ms
538 		 * depending on PCI sleeps and our own resume time).
539 		 * This easily dwarfs any performance advantage from
540 		 * using the cache bypass of indirect GGTT access.
541 		 */
542 		wakeref = intel_runtime_pm_get_if_in_use(rpm);
543 		if (!wakeref)
544 			return -EFAULT;
545 	} else {
546 		/* No backing pages, no fallback, we must force GGTT access */
547 		wakeref = intel_runtime_pm_get(rpm);
548 	}
549 
550 	vma = i915_gem_gtt_prepare(obj, &node, true);
551 	if (IS_ERR(vma)) {
552 		ret = PTR_ERR(vma);
553 		goto out_rpm;
554 	}
555 
556 	i915_gem_object_invalidate_frontbuffer(obj, ORIGIN_CPU);
557 
558 	user_data = u64_to_user_ptr(args->data_ptr);
559 	offset = args->offset;
560 	remain = args->size;
561 	while (remain) {
562 		/* Operation in this page
563 		 *
564 		 * page_base = page offset within aperture
565 		 * page_offset = offset within page
566 		 * page_length = bytes to copy for this page
567 		 */
568 		u32 page_base = node.start;
569 		unsigned int page_offset = offset_in_page(offset);
570 		unsigned int page_length = PAGE_SIZE - page_offset;
571 		page_length = remain < page_length ? remain : page_length;
572 		if (drm_mm_node_allocated(&node)) {
573 			/* flush the write before we modify the GGTT */
574 			intel_gt_flush_ggtt_writes(ggtt->vm.gt);
575 			ggtt->vm.insert_page(&ggtt->vm,
576 					     i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT),
577 					     node.start, I915_CACHE_NONE, 0);
578 			wmb(); /* flush modifications to the GGTT (insert_page) */
579 		} else {
580 			page_base += offset & PAGE_MASK;
581 		}
582 		/* If we get a fault while copying data, then (presumably) our
583 		 * source page isn't available.  Return the error and we'll
584 		 * retry in the slow path.
585 		 * If the object is non-shmem backed, we retry again with the
586 		 * path that handles page fault.
587 		 */
588 		if (ggtt_write(&ggtt->iomap, page_base, page_offset,
589 			       user_data, page_length)) {
590 			ret = -EFAULT;
591 			break;
592 		}
593 
594 		remain -= page_length;
595 		user_data += page_length;
596 		offset += page_length;
597 	}
598 
599 	intel_gt_flush_ggtt_writes(ggtt->vm.gt);
600 	i915_gem_object_flush_frontbuffer(obj, ORIGIN_CPU);
601 
602 	i915_gem_gtt_cleanup(obj, &node, vma);
603 out_rpm:
604 	intel_runtime_pm_put(rpm, wakeref);
605 	return ret;
606 }
607 
608 /* Per-page copy function for the shmem pwrite fastpath.
609  * Flushes invalid cachelines before writing to the target if
610  * needs_clflush_before is set and flushes out any written cachelines after
611  * writing if needs_clflush is set.
612  */
613 static int
shmem_pwrite(struct page * page,int offset,int len,char __user * user_data,bool needs_clflush_before,bool needs_clflush_after)614 shmem_pwrite(struct page *page, int offset, int len, char __user *user_data,
615 	     bool needs_clflush_before,
616 	     bool needs_clflush_after)
617 {
618 	char *vaddr;
619 	int ret;
620 
621 	vaddr = kmap(page);
622 
623 	if (needs_clflush_before)
624 		drm_clflush_virt_range(vaddr + offset, len);
625 
626 	ret = __copy_from_user(vaddr + offset, user_data, len);
627 	if (!ret && needs_clflush_after)
628 		drm_clflush_virt_range(vaddr + offset, len);
629 
630 	kunmap(page);
631 
632 	return ret ? -EFAULT : 0;
633 }
634 
635 static int
i915_gem_shmem_pwrite(struct drm_i915_gem_object * obj,const struct drm_i915_gem_pwrite * args)636 i915_gem_shmem_pwrite(struct drm_i915_gem_object *obj,
637 		      const struct drm_i915_gem_pwrite *args)
638 {
639 	unsigned int partial_cacheline_write;
640 	unsigned int needs_clflush;
641 	unsigned int offset, idx;
642 	void __user *user_data;
643 	u64 remain;
644 	int ret;
645 
646 	ret = i915_gem_object_lock_interruptible(obj, NULL);
647 	if (ret)
648 		return ret;
649 
650 	ret = i915_gem_object_pin_pages(obj);
651 	if (ret)
652 		goto err_unlock;
653 
654 	ret = i915_gem_object_prepare_write(obj, &needs_clflush);
655 	if (ret)
656 		goto err_unpin;
657 
658 	i915_gem_object_finish_access(obj);
659 	i915_gem_object_unlock(obj);
660 
661 	/* If we don't overwrite a cacheline completely we need to be
662 	 * careful to have up-to-date data by first clflushing. Don't
663 	 * overcomplicate things and flush the entire patch.
664 	 */
665 	partial_cacheline_write = 0;
666 	if (needs_clflush & CLFLUSH_BEFORE)
667 		partial_cacheline_write = boot_cpu_data.x86_clflush_size - 1;
668 
669 	user_data = u64_to_user_ptr(args->data_ptr);
670 	remain = args->size;
671 	offset = offset_in_page(args->offset);
672 	for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
673 		struct page *page = i915_gem_object_get_page(obj, idx);
674 		unsigned int length = min_t(u64, remain, PAGE_SIZE - offset);
675 
676 		ret = shmem_pwrite(page, offset, length, user_data,
677 				   (offset | length) & partial_cacheline_write,
678 				   needs_clflush & CLFLUSH_AFTER);
679 		if (ret)
680 			break;
681 
682 		remain -= length;
683 		user_data += length;
684 		offset = 0;
685 	}
686 
687 	i915_gem_object_flush_frontbuffer(obj, ORIGIN_CPU);
688 
689 	i915_gem_object_unpin_pages(obj);
690 	return ret;
691 
692 err_unpin:
693 	i915_gem_object_unpin_pages(obj);
694 err_unlock:
695 	i915_gem_object_unlock(obj);
696 	return ret;
697 }
698 
699 /**
700  * Writes data to the object referenced by handle.
701  * @dev: drm device
702  * @data: ioctl data blob
703  * @file: drm file
704  *
705  * On error, the contents of the buffer that were to be modified are undefined.
706  */
707 int
i915_gem_pwrite_ioctl(struct drm_device * dev,void * data,struct drm_file * file)708 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
709 		      struct drm_file *file)
710 {
711 	struct drm_i915_private *i915 = to_i915(dev);
712 	struct drm_i915_gem_pwrite *args = data;
713 	struct drm_i915_gem_object *obj;
714 	int ret;
715 
716 	/* PWRITE is disallowed for all platforms after TGL-LP.  This also
717 	 * covers all platforms with local memory.
718 	 */
719 	if (GRAPHICS_VER(i915) >= 12 && !IS_TIGERLAKE(i915))
720 		return -EOPNOTSUPP;
721 
722 	if (args->size == 0)
723 		return 0;
724 
725 	if (!access_ok(u64_to_user_ptr(args->data_ptr), args->size))
726 		return -EFAULT;
727 
728 	obj = i915_gem_object_lookup(file, args->handle);
729 	if (!obj)
730 		return -ENOENT;
731 
732 	/* Bounds check destination. */
733 	if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) {
734 		ret = -EINVAL;
735 		goto err;
736 	}
737 
738 	/* Writes not allowed into this read-only object */
739 	if (i915_gem_object_is_readonly(obj)) {
740 		ret = -EINVAL;
741 		goto err;
742 	}
743 
744 	trace_i915_gem_object_pwrite(obj, args->offset, args->size);
745 
746 	ret = -ENODEV;
747 	if (obj->ops->pwrite)
748 		ret = obj->ops->pwrite(obj, args);
749 	if (ret != -ENODEV)
750 		goto err;
751 
752 	ret = i915_gem_object_wait(obj,
753 				   I915_WAIT_INTERRUPTIBLE |
754 				   I915_WAIT_ALL,
755 				   MAX_SCHEDULE_TIMEOUT);
756 	if (ret)
757 		goto err;
758 
759 	ret = -EFAULT;
760 	/* We can only do the GTT pwrite on untiled buffers, as otherwise
761 	 * it would end up going through the fenced access, and we'll get
762 	 * different detiling behavior between reading and writing.
763 	 * pread/pwrite currently are reading and writing from the CPU
764 	 * perspective, requiring manual detiling by the client.
765 	 */
766 	if (!i915_gem_object_has_struct_page(obj) ||
767 	    cpu_write_needs_clflush(obj))
768 		/* Note that the gtt paths might fail with non-page-backed user
769 		 * pointers (e.g. gtt mappings when moving data between
770 		 * textures). Fallback to the shmem path in that case.
771 		 */
772 		ret = i915_gem_gtt_pwrite_fast(obj, args);
773 
774 	if (ret == -EFAULT || ret == -ENOSPC) {
775 		if (i915_gem_object_has_struct_page(obj))
776 			ret = i915_gem_shmem_pwrite(obj, args);
777 	}
778 
779 err:
780 	i915_gem_object_put(obj);
781 	return ret;
782 }
783 
784 /**
785  * Called when user space has done writes to this buffer
786  * @dev: drm device
787  * @data: ioctl data blob
788  * @file: drm file
789  */
790 int
i915_gem_sw_finish_ioctl(struct drm_device * dev,void * data,struct drm_file * file)791 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
792 			 struct drm_file *file)
793 {
794 	struct drm_i915_gem_sw_finish *args = data;
795 	struct drm_i915_gem_object *obj;
796 
797 	obj = i915_gem_object_lookup(file, args->handle);
798 	if (!obj)
799 		return -ENOENT;
800 
801 	/*
802 	 * Proxy objects are barred from CPU access, so there is no
803 	 * need to ban sw_finish as it is a nop.
804 	 */
805 
806 	/* Pinned buffers may be scanout, so flush the cache */
807 	i915_gem_object_flush_if_display(obj);
808 	i915_gem_object_put(obj);
809 
810 	return 0;
811 }
812 
i915_gem_runtime_suspend(struct drm_i915_private * i915)813 void i915_gem_runtime_suspend(struct drm_i915_private *i915)
814 {
815 	struct drm_i915_gem_object *obj, *on;
816 	int i;
817 
818 	/*
819 	 * Only called during RPM suspend. All users of the userfault_list
820 	 * must be holding an RPM wakeref to ensure that this can not
821 	 * run concurrently with themselves (and use the struct_mutex for
822 	 * protection between themselves).
823 	 */
824 
825 	list_for_each_entry_safe(obj, on,
826 				 &i915->ggtt.userfault_list, userfault_link)
827 		__i915_gem_object_release_mmap_gtt(obj);
828 
829 	/*
830 	 * The fence will be lost when the device powers down. If any were
831 	 * in use by hardware (i.e. they are pinned), we should not be powering
832 	 * down! All other fences will be reacquired by the user upon waking.
833 	 */
834 	for (i = 0; i < i915->ggtt.num_fences; i++) {
835 		struct i915_fence_reg *reg = &i915->ggtt.fence_regs[i];
836 
837 		/*
838 		 * Ideally we want to assert that the fence register is not
839 		 * live at this point (i.e. that no piece of code will be
840 		 * trying to write through fence + GTT, as that both violates
841 		 * our tracking of activity and associated locking/barriers,
842 		 * but also is illegal given that the hw is powered down).
843 		 *
844 		 * Previously we used reg->pin_count as a "liveness" indicator.
845 		 * That is not sufficient, and we need a more fine-grained
846 		 * tool if we want to have a sanity check here.
847 		 */
848 
849 		if (!reg->vma)
850 			continue;
851 
852 		GEM_BUG_ON(i915_vma_has_userfault(reg->vma));
853 		reg->dirty = true;
854 	}
855 }
856 
discard_ggtt_vma(struct i915_vma * vma)857 static void discard_ggtt_vma(struct i915_vma *vma)
858 {
859 	struct drm_i915_gem_object *obj = vma->obj;
860 
861 	spin_lock(&obj->vma.lock);
862 	if (!RB_EMPTY_NODE(&vma->obj_node)) {
863 		rb_erase(&vma->obj_node, &obj->vma.tree);
864 		RB_CLEAR_NODE(&vma->obj_node);
865 	}
866 	spin_unlock(&obj->vma.lock);
867 }
868 
869 struct i915_vma *
i915_gem_object_ggtt_pin_ww(struct drm_i915_gem_object * obj,struct i915_gem_ww_ctx * ww,const struct i915_ggtt_view * view,u64 size,u64 alignment,u64 flags)870 i915_gem_object_ggtt_pin_ww(struct drm_i915_gem_object *obj,
871 			    struct i915_gem_ww_ctx *ww,
872 			    const struct i915_ggtt_view *view,
873 			    u64 size, u64 alignment, u64 flags)
874 {
875 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
876 	struct i915_ggtt *ggtt = &i915->ggtt;
877 	struct i915_vma *vma;
878 	int ret;
879 
880 	if (flags & PIN_MAPPABLE &&
881 	    (!view || view->type == I915_GGTT_VIEW_NORMAL)) {
882 		/*
883 		 * If the required space is larger than the available
884 		 * aperture, we will not able to find a slot for the
885 		 * object and unbinding the object now will be in
886 		 * vain. Worse, doing so may cause us to ping-pong
887 		 * the object in and out of the Global GTT and
888 		 * waste a lot of cycles under the mutex.
889 		 */
890 		if (obj->base.size > ggtt->mappable_end)
891 			return ERR_PTR(-E2BIG);
892 
893 		/*
894 		 * If NONBLOCK is set the caller is optimistically
895 		 * trying to cache the full object within the mappable
896 		 * aperture, and *must* have a fallback in place for
897 		 * situations where we cannot bind the object. We
898 		 * can be a little more lax here and use the fallback
899 		 * more often to avoid costly migrations of ourselves
900 		 * and other objects within the aperture.
901 		 *
902 		 * Half-the-aperture is used as a simple heuristic.
903 		 * More interesting would to do search for a free
904 		 * block prior to making the commitment to unbind.
905 		 * That caters for the self-harm case, and with a
906 		 * little more heuristics (e.g. NOFAULT, NOEVICT)
907 		 * we could try to minimise harm to others.
908 		 */
909 		if (flags & PIN_NONBLOCK &&
910 		    obj->base.size > ggtt->mappable_end / 2)
911 			return ERR_PTR(-ENOSPC);
912 	}
913 
914 new_vma:
915 	vma = i915_vma_instance(obj, &ggtt->vm, view);
916 	if (IS_ERR(vma))
917 		return vma;
918 
919 	if (i915_vma_misplaced(vma, size, alignment, flags)) {
920 		if (flags & PIN_NONBLOCK) {
921 			if (i915_vma_is_pinned(vma) || i915_vma_is_active(vma))
922 				return ERR_PTR(-ENOSPC);
923 
924 			if (flags & PIN_MAPPABLE &&
925 			    vma->fence_size > ggtt->mappable_end / 2)
926 				return ERR_PTR(-ENOSPC);
927 		}
928 
929 		if (i915_vma_is_pinned(vma) || i915_vma_is_active(vma)) {
930 			discard_ggtt_vma(vma);
931 			goto new_vma;
932 		}
933 
934 		ret = i915_vma_unbind(vma);
935 		if (ret)
936 			return ERR_PTR(ret);
937 	}
938 
939 	if (ww)
940 		ret = i915_vma_pin_ww(vma, ww, size, alignment, flags | PIN_GLOBAL);
941 	else
942 		ret = i915_vma_pin(vma, size, alignment, flags | PIN_GLOBAL);
943 
944 	if (ret)
945 		return ERR_PTR(ret);
946 
947 	if (vma->fence && !i915_gem_object_is_tiled(obj)) {
948 		mutex_lock(&ggtt->vm.mutex);
949 		i915_vma_revoke_fence(vma);
950 		mutex_unlock(&ggtt->vm.mutex);
951 	}
952 
953 	ret = i915_vma_wait_for_bind(vma);
954 	if (ret) {
955 		i915_vma_unpin(vma);
956 		return ERR_PTR(ret);
957 	}
958 
959 	return vma;
960 }
961 
962 int
i915_gem_madvise_ioctl(struct drm_device * dev,void * data,struct drm_file * file_priv)963 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
964 		       struct drm_file *file_priv)
965 {
966 	struct drm_i915_private *i915 = to_i915(dev);
967 	struct drm_i915_gem_madvise *args = data;
968 	struct drm_i915_gem_object *obj;
969 	int err;
970 
971 	switch (args->madv) {
972 	case I915_MADV_DONTNEED:
973 	case I915_MADV_WILLNEED:
974 	    break;
975 	default:
976 	    return -EINVAL;
977 	}
978 
979 	obj = i915_gem_object_lookup(file_priv, args->handle);
980 	if (!obj)
981 		return -ENOENT;
982 
983 	err = i915_gem_object_lock_interruptible(obj, NULL);
984 	if (err)
985 		goto out;
986 
987 	if (i915_gem_object_has_pages(obj) &&
988 	    i915_gem_object_is_tiled(obj) &&
989 	    i915->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
990 		if (obj->mm.madv == I915_MADV_WILLNEED) {
991 			GEM_BUG_ON(!i915_gem_object_has_tiling_quirk(obj));
992 			i915_gem_object_clear_tiling_quirk(obj);
993 			i915_gem_object_make_shrinkable(obj);
994 		}
995 		if (args->madv == I915_MADV_WILLNEED) {
996 			GEM_BUG_ON(i915_gem_object_has_tiling_quirk(obj));
997 			i915_gem_object_make_unshrinkable(obj);
998 			i915_gem_object_set_tiling_quirk(obj);
999 		}
1000 	}
1001 
1002 	if (obj->mm.madv != __I915_MADV_PURGED) {
1003 		obj->mm.madv = args->madv;
1004 		if (obj->ops->adjust_lru)
1005 			obj->ops->adjust_lru(obj);
1006 	}
1007 
1008 	if (i915_gem_object_has_pages(obj)) {
1009 		unsigned long flags;
1010 
1011 		spin_lock_irqsave(&i915->mm.obj_lock, flags);
1012 		if (!list_empty(&obj->mm.link)) {
1013 			struct list_head *list;
1014 
1015 			if (obj->mm.madv != I915_MADV_WILLNEED)
1016 				list = &i915->mm.purge_list;
1017 			else
1018 				list = &i915->mm.shrink_list;
1019 			list_move_tail(&obj->mm.link, list);
1020 
1021 		}
1022 		spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
1023 	}
1024 
1025 	/* if the object is no longer attached, discard its backing storage */
1026 	if (obj->mm.madv == I915_MADV_DONTNEED &&
1027 	    !i915_gem_object_has_pages(obj))
1028 		i915_gem_object_truncate(obj);
1029 
1030 	args->retained = obj->mm.madv != __I915_MADV_PURGED;
1031 
1032 	i915_gem_object_unlock(obj);
1033 out:
1034 	i915_gem_object_put(obj);
1035 	return err;
1036 }
1037 
i915_gem_init(struct drm_i915_private * dev_priv)1038 int i915_gem_init(struct drm_i915_private *dev_priv)
1039 {
1040 	int ret;
1041 
1042 	/* We need to fallback to 4K pages if host doesn't support huge gtt. */
1043 	if (intel_vgpu_active(dev_priv) && !intel_vgpu_has_huge_gtt(dev_priv))
1044 		mkwrite_device_info(dev_priv)->page_sizes =
1045 			I915_GTT_PAGE_SIZE_4K;
1046 
1047 	ret = i915_gem_init_userptr(dev_priv);
1048 	if (ret)
1049 		return ret;
1050 
1051 	intel_uc_fetch_firmwares(&dev_priv->gt.uc);
1052 	intel_wopcm_init(&dev_priv->wopcm);
1053 
1054 	ret = i915_init_ggtt(dev_priv);
1055 	if (ret) {
1056 		GEM_BUG_ON(ret == -EIO);
1057 		goto err_unlock;
1058 	}
1059 
1060 	/*
1061 	 * Despite its name intel_init_clock_gating applies both display
1062 	 * clock gating workarounds; GT mmio workarounds and the occasional
1063 	 * GT power context workaround. Worse, sometimes it includes a context
1064 	 * register workaround which we need to apply before we record the
1065 	 * default HW state for all contexts.
1066 	 *
1067 	 * FIXME: break up the workarounds and apply them at the right time!
1068 	 */
1069 	intel_init_clock_gating(dev_priv);
1070 
1071 	ret = intel_gt_init(&dev_priv->gt);
1072 	if (ret)
1073 		goto err_unlock;
1074 
1075 	return 0;
1076 
1077 	/*
1078 	 * Unwinding is complicated by that we want to handle -EIO to mean
1079 	 * disable GPU submission but keep KMS alive. We want to mark the
1080 	 * HW as irrevisibly wedged, but keep enough state around that the
1081 	 * driver doesn't explode during runtime.
1082 	 */
1083 err_unlock:
1084 	i915_gem_drain_workqueue(dev_priv);
1085 
1086 	if (ret != -EIO)
1087 		intel_uc_cleanup_firmwares(&dev_priv->gt.uc);
1088 
1089 	if (ret == -EIO) {
1090 		/*
1091 		 * Allow engines or uC initialisation to fail by marking the GPU
1092 		 * as wedged. But we only want to do this when the GPU is angry,
1093 		 * for all other failure, such as an allocation failure, bail.
1094 		 */
1095 		if (!intel_gt_is_wedged(&dev_priv->gt)) {
1096 			i915_probe_error(dev_priv,
1097 					 "Failed to initialize GPU, declaring it wedged!\n");
1098 			intel_gt_set_wedged(&dev_priv->gt);
1099 		}
1100 
1101 		/* Minimal basic recovery for KMS */
1102 		ret = i915_ggtt_enable_hw(dev_priv);
1103 		i915_ggtt_resume(&dev_priv->ggtt);
1104 		intel_init_clock_gating(dev_priv);
1105 	}
1106 
1107 	i915_gem_drain_freed_objects(dev_priv);
1108 
1109 	return ret;
1110 }
1111 
i915_gem_driver_register(struct drm_i915_private * i915)1112 void i915_gem_driver_register(struct drm_i915_private *i915)
1113 {
1114 	i915_gem_driver_register__shrinker(i915);
1115 
1116 	intel_engines_driver_register(i915);
1117 }
1118 
i915_gem_driver_unregister(struct drm_i915_private * i915)1119 void i915_gem_driver_unregister(struct drm_i915_private *i915)
1120 {
1121 	i915_gem_driver_unregister__shrinker(i915);
1122 }
1123 
i915_gem_driver_remove(struct drm_i915_private * dev_priv)1124 void i915_gem_driver_remove(struct drm_i915_private *dev_priv)
1125 {
1126 	intel_wakeref_auto_fini(&dev_priv->ggtt.userfault_wakeref);
1127 
1128 	i915_gem_suspend_late(dev_priv);
1129 	intel_gt_driver_remove(&dev_priv->gt);
1130 	dev_priv->uabi_engines = RB_ROOT;
1131 
1132 	/* Flush any outstanding unpin_work. */
1133 	i915_gem_drain_workqueue(dev_priv);
1134 
1135 	i915_gem_drain_freed_objects(dev_priv);
1136 }
1137 
i915_gem_driver_release(struct drm_i915_private * dev_priv)1138 void i915_gem_driver_release(struct drm_i915_private *dev_priv)
1139 {
1140 	intel_gt_driver_release(&dev_priv->gt);
1141 
1142 	intel_uc_cleanup_firmwares(&dev_priv->gt.uc);
1143 
1144 	i915_gem_drain_freed_objects(dev_priv);
1145 
1146 	drm_WARN_ON(&dev_priv->drm, !list_empty(&dev_priv->gem.contexts.list));
1147 }
1148 
i915_gem_init__mm(struct drm_i915_private * i915)1149 static void i915_gem_init__mm(struct drm_i915_private *i915)
1150 {
1151 	spin_lock_init(&i915->mm.obj_lock);
1152 
1153 	init_llist_head(&i915->mm.free_list);
1154 
1155 	INIT_LIST_HEAD(&i915->mm.purge_list);
1156 	INIT_LIST_HEAD(&i915->mm.shrink_list);
1157 
1158 	i915_gem_init__objects(i915);
1159 }
1160 
i915_gem_init_early(struct drm_i915_private * dev_priv)1161 void i915_gem_init_early(struct drm_i915_private *dev_priv)
1162 {
1163 	i915_gem_init__mm(dev_priv);
1164 	i915_gem_init__contexts(dev_priv);
1165 
1166 	spin_lock_init(&dev_priv->fb_tracking.lock);
1167 }
1168 
i915_gem_cleanup_early(struct drm_i915_private * dev_priv)1169 void i915_gem_cleanup_early(struct drm_i915_private *dev_priv)
1170 {
1171 	i915_gem_drain_freed_objects(dev_priv);
1172 	GEM_BUG_ON(!llist_empty(&dev_priv->mm.free_list));
1173 	GEM_BUG_ON(atomic_read(&dev_priv->mm.free_count));
1174 	drm_WARN_ON(&dev_priv->drm, dev_priv->mm.shrink_count);
1175 }
1176 
i915_gem_open(struct drm_i915_private * i915,struct drm_file * file)1177 int i915_gem_open(struct drm_i915_private *i915, struct drm_file *file)
1178 {
1179 	struct drm_i915_file_private *file_priv;
1180 	int ret;
1181 
1182 	DRM_DEBUG("\n");
1183 
1184 	file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL);
1185 	if (!file_priv)
1186 		return -ENOMEM;
1187 
1188 	file->driver_priv = file_priv;
1189 	file_priv->dev_priv = i915;
1190 	file_priv->file = file;
1191 
1192 	file_priv->bsd_engine = -1;
1193 	file_priv->hang_timestamp = jiffies;
1194 
1195 	ret = i915_gem_context_open(i915, file);
1196 	if (ret)
1197 		kfree(file_priv);
1198 
1199 	return ret;
1200 }
1201 
1202 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1203 #include "selftests/mock_gem_device.c"
1204 #include "selftests/i915_gem.c"
1205 #endif
1206