xref: /dflybsd-src/sys/dev/drm/i915/i915_gem.c (revision 31c068aaf635ad9fa72dbc4c65b32d890ff7544d)
1 /*
2  * Copyright © 2008 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  * Authors:
24  *    Eric Anholt <eric@anholt.net>
25  *
26  * Copyright (c) 2011 The FreeBSD Foundation
27  * All rights reserved.
28  *
29  * This software was developed by Konstantin Belousov under sponsorship from
30  * the FreeBSD Foundation.
31  *
32  * Redistribution and use in source and binary forms, with or without
33  * modification, are permitted provided that the following conditions
34  * are met:
35  * 1. Redistributions of source code must retain the above copyright
36  *    notice, this list of conditions and the following disclaimer.
37  * 2. Redistributions in binary form must reproduce the above copyright
38  *    notice, this list of conditions and the following disclaimer in the
39  *    documentation and/or other materials provided with the distribution.
40  *
41  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
42  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
43  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
44  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
45  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
46  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
47  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
48  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
49  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
50  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
51  * SUCH DAMAGE.
52  *
53  */
54 
55 #include <sys/resourcevar.h>
56 #include <sys/sfbuf.h>
57 
58 #include <drm/drmP.h>
59 #include <drm/i915_drm.h>
60 #include "i915_drv.h"
61 #include "intel_drv.h"
62 #include <linux/shmem_fs.h>
63 #include <linux/completion.h>
64 #include <linux/highmem.h>
65 #include <linux/jiffies.h>
66 #include <linux/time.h>
67 
68 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);
69 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj);
70 static __must_check int i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
71 						    unsigned alignment,
72 						    bool map_and_fenceable,
73 						    bool nonblocking);
74 static int i915_gem_phys_pwrite(struct drm_device *dev,
75 				struct drm_i915_gem_object *obj,
76 				struct drm_i915_gem_pwrite *args,
77 				struct drm_file *file);
78 
79 static void i915_gem_write_fence(struct drm_device *dev, int reg,
80 				 struct drm_i915_gem_object *obj);
81 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
82 					 struct drm_i915_fence_reg *fence,
83 					 bool enable);
84 
85 static long i915_gem_purge(struct drm_i915_private *dev_priv, long target);
86 
87 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj);
88 static void i915_gem_object_truncate(struct drm_i915_gem_object *obj);
89 
90 static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object *obj)
91 {
92 	if (obj->tiling_mode)
93 		i915_gem_release_mmap(obj);
94 
95 	/* As we do not have an associated fence register, we will force
96 	 * a tiling change if we ever need to acquire one.
97 	 */
98 	obj->fence_dirty = false;
99 	obj->fence_reg = I915_FENCE_REG_NONE;
100 }
101 
102 static int i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj);
103 static bool i915_gem_object_is_inactive(struct drm_i915_gem_object *obj);
104 static void i915_gem_reset_fences(struct drm_device *dev);
105 static void i915_gem_lowmem(void *arg);
106 
107 /* some bookkeeping */
108 static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
109 				  size_t size)
110 {
111 	dev_priv->mm.object_count++;
112 	dev_priv->mm.object_memory += size;
113 }
114 
115 static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
116 				     size_t size)
117 {
118 	dev_priv->mm.object_count--;
119 	dev_priv->mm.object_memory -= size;
120 }
121 
122 static int
123 i915_gem_wait_for_error(struct i915_gpu_error *error)
124 {
125 	int ret;
126 
127 #define EXIT_COND (!i915_reset_in_progress(error) || \
128 		   i915_terminally_wedged(error))
129 	if (EXIT_COND)
130 		return 0;
131 
132 	/*
133 	 * Only wait 10 seconds for the gpu reset to complete to avoid hanging
134 	 * userspace. If it takes that long something really bad is going on and
135 	 * we should simply try to bail out and fail as gracefully as possible.
136 	 */
137 	ret = wait_event_interruptible_timeout(error->reset_queue,
138 					       EXIT_COND,
139 					       10*HZ);
140 	if (ret == 0) {
141 		DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
142 		return -EIO;
143 	} else if (ret < 0) {
144 		return ret;
145 	}
146 #undef EXIT_COND
147 
148 	return 0;
149 }
150 
151 int i915_mutex_lock_interruptible(struct drm_device *dev)
152 {
153 	struct drm_i915_private *dev_priv = dev->dev_private;
154 	int ret;
155 
156 	ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
157 	if (ret)
158 		return ret;
159 
160 	ret = lockmgr(&dev->struct_mutex, LK_EXCLUSIVE|LK_SLEEPFAIL);
161 	if (ret)
162 		return -EINTR;
163 
164 	WARN_ON(i915_verify_lists(dev));
165 	return 0;
166 }
167 
168 static inline bool
169 i915_gem_object_is_inactive(struct drm_i915_gem_object *obj)
170 {
171 	return !obj->active;
172 }
173 
174 int
175 i915_gem_init_ioctl(struct drm_device *dev, void *data,
176 		    struct drm_file *file)
177 {
178 	struct drm_i915_gem_init *args = data;
179 
180 	if (drm_core_check_feature(dev, DRIVER_MODESET))
181 		return -ENODEV;
182 
183 	if (args->gtt_start >= args->gtt_end ||
184 	    (args->gtt_end | args->gtt_start) & (PAGE_SIZE - 1))
185 		return -EINVAL;
186 
187 	/* GEM with user mode setting was never supported on ilk and later. */
188 	if (INTEL_INFO(dev)->gen >= 5)
189 		return -ENODEV;
190 
191 	mutex_lock(&dev->struct_mutex);
192 	i915_gem_setup_global_gtt(dev, args->gtt_start, args->gtt_end,
193 				  args->gtt_end);
194 	mutex_unlock(&dev->struct_mutex);
195 
196 	return 0;
197 }
198 
199 int
200 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
201 			    struct drm_file *file)
202 {
203 	struct drm_i915_private *dev_priv = dev->dev_private;
204 	struct drm_i915_gem_get_aperture *args = data;
205 	struct drm_i915_gem_object *obj;
206 	size_t pinned;
207 
208 	pinned = 0;
209 	mutex_lock(&dev->struct_mutex);
210 	list_for_each_entry(obj, &dev_priv->mm.bound_list, gtt_list)
211 		if (obj->pin_count)
212 			pinned += obj->gtt_space->size;
213 	mutex_unlock(&dev->struct_mutex);
214 
215 	args->aper_size = dev_priv->gtt.total;
216 	args->aper_available_size = args->aper_size - pinned;
217 
218 	return 0;
219 }
220 
221 static int
222 i915_gem_create(struct drm_file *file,
223 		struct drm_device *dev,
224 		uint64_t size,
225 		uint32_t *handle_p)
226 {
227 	struct drm_i915_gem_object *obj;
228 	int ret;
229 	u32 handle;
230 
231 	size = roundup(size, PAGE_SIZE);
232 	if (size == 0)
233 		return -EINVAL;
234 
235 	/* Allocate the new object */
236 	obj = i915_gem_alloc_object(dev, size);
237 	if (obj == NULL)
238 		return -ENOMEM;
239 
240 	handle = 0;
241 	ret = drm_gem_handle_create(file, &obj->base, &handle);
242 	if (ret) {
243 		drm_gem_object_release(&obj->base);
244 		i915_gem_info_remove_obj(dev->dev_private, obj->base.size);
245 		drm_free(obj, M_DRM);
246 		return (-ret);
247 	}
248 
249 	/* drop reference from allocate - handle holds it now */
250 	drm_gem_object_unreference(&obj->base);
251 	*handle_p = handle;
252 	return 0;
253 }
254 
255 int
256 i915_gem_dumb_create(struct drm_file *file,
257 		     struct drm_device *dev,
258 		     struct drm_mode_create_dumb *args)
259 {
260 
261 	/* have to work out size/pitch and return them */
262 	args->pitch = roundup2(args->width * ((args->bpp + 7) / 8), 64);
263 	args->size = args->pitch * args->height;
264 	return i915_gem_create(file, dev,
265 			       args->size, &args->handle);
266 }
267 
268 int i915_gem_dumb_destroy(struct drm_file *file,
269 			  struct drm_device *dev,
270 			  uint32_t handle)
271 {
272 
273 	return drm_gem_handle_delete(file, handle);
274 }
275 
276 /**
277  * Creates a new mm object and returns a handle to it.
278  */
279 int
280 i915_gem_create_ioctl(struct drm_device *dev, void *data,
281 		      struct drm_file *file)
282 {
283 	struct drm_i915_gem_create *args = data;
284 
285 	return i915_gem_create(file, dev,
286 			       args->size, &args->handle);
287 }
288 
289 static inline void vm_page_reference(vm_page_t m)
290 {
291 	vm_page_flag_set(m, PG_REFERENCED);
292 }
293 
294 static int
295 i915_gem_shmem_pread(struct drm_device *dev,
296 		     struct drm_i915_gem_object *obj,
297 		     struct drm_i915_gem_pread *args,
298 		     struct drm_file *file)
299 {
300 	vm_object_t vm_obj;
301 	vm_page_t m;
302 	struct sf_buf *sf;
303 	vm_offset_t mkva;
304 	vm_pindex_t obj_pi;
305 	int cnt, do_bit17_swizzling, length, obj_po, ret, swizzled_po;
306 
307 	do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
308 
309 	obj->dirty = 1;
310 	vm_obj = obj->base.vm_obj;
311 	ret = 0;
312 
313 	VM_OBJECT_LOCK(vm_obj);
314 	vm_object_pip_add(vm_obj, 1);
315 	while (args->size > 0) {
316 		obj_pi = OFF_TO_IDX(args->offset);
317 		obj_po = args->offset & PAGE_MASK;
318 
319 		m = shmem_read_mapping_page(vm_obj, obj_pi);
320 		VM_OBJECT_UNLOCK(vm_obj);
321 
322 		sf = sf_buf_alloc(m);
323 		mkva = sf_buf_kva(sf);
324 		length = min(args->size, PAGE_SIZE - obj_po);
325 		while (length > 0) {
326 			if (do_bit17_swizzling &&
327 			    (VM_PAGE_TO_PHYS(m) & (1 << 17)) != 0) {
328 				cnt = roundup2(obj_po + 1, 64);
329 				cnt = min(cnt - obj_po, length);
330 				swizzled_po = obj_po ^ 64;
331 			} else {
332 				cnt = length;
333 				swizzled_po = obj_po;
334 			}
335 			ret = -copyout_nofault(
336 			    (char *)mkva + swizzled_po,
337 			    (void *)(uintptr_t)args->data_ptr, cnt);
338 			if (ret != 0)
339 				break;
340 			args->data_ptr += cnt;
341 			args->size -= cnt;
342 			length -= cnt;
343 			args->offset += cnt;
344 			obj_po += cnt;
345 		}
346 		sf_buf_free(sf);
347 		VM_OBJECT_LOCK(vm_obj);
348 		vm_page_reference(m);
349 		vm_page_busy_wait(m, FALSE, "i915gem");
350 		vm_page_unwire(m, 1);
351 		vm_page_wakeup(m);
352 
353 		if (ret != 0)
354 			break;
355 	}
356 	vm_object_pip_wakeup(vm_obj);
357 	VM_OBJECT_UNLOCK(vm_obj);
358 
359 	return (ret);
360 }
361 
362 /**
363  * Reads data from the object referenced by handle.
364  *
365  * On error, the contents of *data are undefined.
366  */
367 int
368 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
369 		     struct drm_file *file)
370 {
371 	struct drm_i915_gem_pread *args = data;
372 	struct drm_i915_gem_object *obj;
373 	int ret = 0;
374 
375 	if (args->size == 0)
376 		return 0;
377 
378 	ret = i915_mutex_lock_interruptible(dev);
379 	if (ret)
380 		return ret;
381 
382 	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
383 	if (&obj->base == NULL) {
384 		ret = -ENOENT;
385 		goto unlock;
386 	}
387 
388 	/* Bounds check source.  */
389 	if (args->offset > obj->base.size ||
390 	    args->size > obj->base.size - args->offset) {
391 		ret = -EINVAL;
392 		goto out;
393 	}
394 
395 	ret = i915_gem_shmem_pread(dev, obj, args, file);
396 out:
397 	drm_gem_object_unreference(&obj->base);
398 unlock:
399 	mutex_unlock(&dev->struct_mutex);
400 	return ret;
401 }
402 
403 #if 0
404 /* This is the fast write path which cannot handle
405  * page faults in the source data
406  */
407 
408 static inline int
409 fast_user_write(struct io_mapping *mapping,
410 		loff_t page_base, int page_offset,
411 		char __user *user_data,
412 		int length)
413 {
414 	void __iomem *vaddr_atomic;
415 	void *vaddr;
416 	unsigned long unwritten;
417 
418 	vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
419 	/* We can use the cpu mem copy function because this is X86. */
420 	vaddr = (void __force*)vaddr_atomic + page_offset;
421 	unwritten = __copy_from_user_inatomic_nocache(vaddr,
422 						      user_data, length);
423 	io_mapping_unmap_atomic(vaddr_atomic);
424 	return unwritten;
425 }
426 
427 /**
428  * This is the fast pwrite path, where we copy the data directly from the
429  * user into the GTT, uncached.
430  */
431 static int
432 i915_gem_gtt_pwrite_fast(struct drm_device *dev,
433 			 struct drm_i915_gem_object *obj,
434 			 struct drm_i915_gem_pwrite *args,
435 			 struct drm_file *file)
436 {
437 	drm_i915_private_t *dev_priv = dev->dev_private;
438 	ssize_t remain;
439 	loff_t offset, page_base;
440 	char __user *user_data;
441 	int page_offset, page_length, ret;
442 
443 	ret = i915_gem_object_pin(obj, 0, true, true);
444 	if (ret)
445 		goto out;
446 
447 	ret = i915_gem_object_set_to_gtt_domain(obj, true);
448 	if (ret)
449 		goto out_unpin;
450 
451 	ret = i915_gem_object_put_fence(obj);
452 	if (ret)
453 		goto out_unpin;
454 
455 	user_data = (char __user *) (uintptr_t) args->data_ptr;
456 	remain = args->size;
457 
458 	offset = obj->gtt_offset + args->offset;
459 
460 	while (remain > 0) {
461 		/* Operation in this page
462 		 *
463 		 * page_base = page offset within aperture
464 		 * page_offset = offset within page
465 		 * page_length = bytes to copy for this page
466 		 */
467 		page_base = offset & PAGE_MASK;
468 		page_offset = offset_in_page(offset);
469 		page_length = remain;
470 		if ((page_offset + remain) > PAGE_SIZE)
471 			page_length = PAGE_SIZE - page_offset;
472 
473 		/* If we get a fault while copying data, then (presumably) our
474 		 * source page isn't available.  Return the error and we'll
475 		 * retry in the slow path.
476 		 */
477 		if (fast_user_write(dev_priv->gtt.mappable, page_base,
478 				    page_offset, user_data, page_length)) {
479 			ret = -EFAULT;
480 			goto out_unpin;
481 		}
482 
483 		remain -= page_length;
484 		user_data += page_length;
485 		offset += page_length;
486 	}
487 
488 out_unpin:
489 	i915_gem_object_unpin(obj);
490 out:
491 	return ret;
492 }
493 #endif
494 
495 static int
496 i915_gem_gtt_write(struct drm_device *dev, struct drm_i915_gem_object *obj,
497     uint64_t data_ptr, uint64_t size, uint64_t offset, struct drm_file *file)
498 {
499 	vm_offset_t mkva;
500 	int ret;
501 
502 	/*
503 	 * Pass the unaligned physical address and size to pmap_mapdev_attr()
504 	 * so it can properly calculate whether an extra page needs to be
505 	 * mapped or not to cover the requested range.  The function will
506 	 * add the page offset into the returned mkva for us.
507 	 */
508 	mkva = (vm_offset_t)pmap_mapdev_attr(dev->agp->base + obj->gtt_offset +
509 	    offset, size, PAT_WRITE_COMBINING);
510 	ret = -copyin_nofault((void *)(uintptr_t)data_ptr, (char *)mkva, size);
511 	pmap_unmapdev(mkva, size);
512 	return ret;
513 }
514 
515 static int
516 i915_gem_shmem_pwrite(struct drm_device *dev,
517 		      struct drm_i915_gem_object *obj,
518 		      struct drm_i915_gem_pwrite *args,
519 		      struct drm_file *file)
520 {
521 	vm_object_t vm_obj;
522 	vm_page_t m;
523 	struct sf_buf *sf;
524 	vm_offset_t mkva;
525 	vm_pindex_t obj_pi;
526 	int cnt, do_bit17_swizzling, length, obj_po, ret, swizzled_po;
527 
528 	do_bit17_swizzling = 0;
529 
530 	obj->dirty = 1;
531 	vm_obj = obj->base.vm_obj;
532 	ret = 0;
533 
534 	VM_OBJECT_LOCK(vm_obj);
535 	vm_object_pip_add(vm_obj, 1);
536 	while (args->size > 0) {
537 		obj_pi = OFF_TO_IDX(args->offset);
538 		obj_po = args->offset & PAGE_MASK;
539 
540 		m = shmem_read_mapping_page(vm_obj, obj_pi);
541 		VM_OBJECT_UNLOCK(vm_obj);
542 
543 		sf = sf_buf_alloc(m);
544 		mkva = sf_buf_kva(sf);
545 		length = min(args->size, PAGE_SIZE - obj_po);
546 		while (length > 0) {
547 			if (do_bit17_swizzling &&
548 			    (VM_PAGE_TO_PHYS(m) & (1 << 17)) != 0) {
549 				cnt = roundup2(obj_po + 1, 64);
550 				cnt = min(cnt - obj_po, length);
551 				swizzled_po = obj_po ^ 64;
552 			} else {
553 				cnt = length;
554 				swizzled_po = obj_po;
555 			}
556 			ret = -copyin_nofault(
557 			    (void *)(uintptr_t)args->data_ptr,
558 			    (char *)mkva + swizzled_po, cnt);
559 			if (ret != 0)
560 				break;
561 			args->data_ptr += cnt;
562 			args->size -= cnt;
563 			length -= cnt;
564 			args->offset += cnt;
565 			obj_po += cnt;
566 		}
567 		sf_buf_free(sf);
568 		VM_OBJECT_LOCK(vm_obj);
569 		vm_page_dirty(m);
570 		vm_page_reference(m);
571 		vm_page_busy_wait(m, FALSE, "i915gem");
572 		vm_page_unwire(m, 1);
573 		vm_page_wakeup(m);
574 
575 		if (ret != 0)
576 			break;
577 	}
578 	vm_object_pip_wakeup(vm_obj);
579 	VM_OBJECT_UNLOCK(vm_obj);
580 
581 	return (ret);
582 }
583 
584 /**
585  * Writes data to the object referenced by handle.
586  *
587  * On error, the contents of the buffer that were to be modified are undefined.
588  */
589 int
590 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
591 		      struct drm_file *file)
592 {
593 	struct drm_i915_gem_pwrite *args = data;
594 	struct drm_i915_gem_object *obj;
595 	vm_page_t *ma;
596 	vm_offset_t start, end;
597 	int npages, ret;
598 
599 	if (args->size == 0)
600 		return 0;
601 
602 	start = trunc_page(args->data_ptr);
603 	end = round_page(args->data_ptr + args->size);
604 	npages = howmany(end - start, PAGE_SIZE);
605 	ma = kmalloc(npages * sizeof(vm_page_t), M_DRM, M_WAITOK |
606 	    M_ZERO);
607 	npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
608 	    (vm_offset_t)args->data_ptr, args->size,
609 	    VM_PROT_READ, ma, npages);
610 	if (npages == -1) {
611 		ret = -EFAULT;
612 		goto free_ma;
613 	}
614 
615 	ret = i915_mutex_lock_interruptible(dev);
616 	if (ret != 0)
617 		goto unlocked;
618 
619 	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
620 	if (&obj->base == NULL) {
621 		ret = -ENOENT;
622 		goto unlock;
623 	}
624 
625 	/* Bounds check destination. */
626 	if (args->offset > obj->base.size ||
627 	    args->size > obj->base.size - args->offset) {
628 		ret = -EINVAL;
629 		goto out;
630 	}
631 
632 	if (obj->phys_obj) {
633 		ret = i915_gem_phys_pwrite(dev, obj, args, file);
634 	} else if (obj->gtt_space &&
635 		    obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
636 		ret = i915_gem_object_pin(obj, 0, true, false);
637 		if (ret != 0)
638 			goto out;
639 		ret = i915_gem_object_set_to_gtt_domain(obj, true);
640 		if (ret != 0)
641 			goto out_unpin;
642 		ret = i915_gem_object_put_fence(obj);
643 		if (ret != 0)
644 			goto out_unpin;
645 		ret = i915_gem_gtt_write(dev, obj, args->data_ptr, args->size,
646 		    args->offset, file);
647 out_unpin:
648 		i915_gem_object_unpin(obj);
649 	} else {
650 		ret = i915_gem_object_set_to_cpu_domain(obj, true);
651 		if (ret != 0)
652 			goto out;
653 		ret = i915_gem_shmem_pwrite(dev, obj, args, file);
654 	}
655 out:
656 	drm_gem_object_unreference(&obj->base);
657 unlock:
658 	mutex_unlock(&dev->struct_mutex);
659 unlocked:
660 	vm_page_unhold_pages(ma, npages);
661 free_ma:
662 	drm_free(ma, M_DRM);
663 	return ret;
664 }
665 
666 int
667 i915_gem_check_wedge(struct i915_gpu_error *error,
668 		     bool interruptible)
669 {
670 	if (i915_reset_in_progress(error)) {
671 		/* Non-interruptible callers can't handle -EAGAIN, hence return
672 		 * -EIO unconditionally for these. */
673 		if (!interruptible)
674 			return -EIO;
675 
676 		/* Recovery complete, but the reset failed ... */
677 		if (i915_terminally_wedged(error))
678 			return -EIO;
679 
680 		return -EAGAIN;
681 	}
682 
683 	return 0;
684 }
685 
686 /*
687  * Compare seqno against outstanding lazy request. Emit a request if they are
688  * equal.
689  */
690 static int
691 i915_gem_check_olr(struct intel_ring_buffer *ring, u32 seqno)
692 {
693 	int ret;
694 
695 	DRM_LOCK_ASSERT(ring->dev);
696 
697 	ret = 0;
698 	if (seqno == ring->outstanding_lazy_request)
699 		ret = i915_add_request(ring, NULL, NULL);
700 
701 	return ret;
702 }
703 
704 /**
705  * __wait_seqno - wait until execution of seqno has finished
706  * @ring: the ring expected to report seqno
707  * @seqno: duh!
708  * @reset_counter: reset sequence associated with the given seqno
709  * @interruptible: do an interruptible wait (normally yes)
710  * @timeout: in - how long to wait (NULL forever); out - how much time remaining
711  *
712  * Note: It is of utmost importance that the passed in seqno and reset_counter
713  * values have been read by the caller in an smp safe manner. Where read-side
714  * locks are involved, it is sufficient to read the reset_counter before
715  * unlocking the lock that protects the seqno. For lockless tricks, the
716  * reset_counter _must_ be read before, and an appropriate smp_rmb must be
717  * inserted.
718  *
719  * Returns 0 if the seqno was found within the alloted time. Else returns the
720  * errno with remaining time filled in timeout argument.
721  */
722 static int __wait_seqno(struct intel_ring_buffer *ring, u32 seqno,
723 			unsigned reset_counter,
724 			bool interruptible, struct timespec *timeout)
725 {
726 	drm_i915_private_t *dev_priv = ring->dev->dev_private;
727 	struct timespec before, now, wait_time={1,0};
728 	unsigned long timeout_jiffies;
729 	long end;
730 	bool wait_forever = true;
731 	int ret;
732 
733 	if (i915_seqno_passed(ring->get_seqno(ring, true), seqno))
734 		return 0;
735 
736 	if (timeout != NULL) {
737 		wait_time = *timeout;
738 		wait_forever = false;
739 	}
740 
741 	timeout_jiffies = timespec_to_jiffies(&wait_time);
742 
743 	if (WARN_ON(!ring->irq_get(ring)))
744 		return -ENODEV;
745 
746 	/* Record current time in case interrupted by signal, or wedged * */
747 	getrawmonotonic(&before);
748 
749 #define EXIT_COND \
750 	(i915_seqno_passed(ring->get_seqno(ring, false), seqno) || \
751 	 i915_reset_in_progress(&dev_priv->gpu_error) || \
752 	 reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter))
753 	do {
754 		if (interruptible)
755 			end = wait_event_interruptible_timeout(ring->irq_queue,
756 							       EXIT_COND,
757 							       timeout_jiffies);
758 		else
759 			end = wait_event_timeout(ring->irq_queue, EXIT_COND,
760 						 timeout_jiffies);
761 
762 		/* We need to check whether any gpu reset happened in between
763 		 * the caller grabbing the seqno and now ... */
764 		if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter))
765 			end = -EAGAIN;
766 
767 		/* ... but upgrade the -EGAIN to an -EIO if the gpu is truely
768 		 * gone. */
769 		ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
770 		if (ret)
771 			end = ret;
772 	} while (end == 0 && wait_forever);
773 
774 	getrawmonotonic(&now);
775 
776 	ring->irq_put(ring);
777 #undef EXIT_COND
778 
779 	if (timeout) {
780 		struct timespec sleep_time = timespec_sub(now, before);
781 		*timeout = timespec_sub(*timeout, sleep_time);
782 	}
783 
784 	switch (end) {
785 	case -EIO:
786 	case -EAGAIN: /* Wedged */
787 	case -ERESTARTSYS: /* Signal */
788 		return (int)end;
789 	case 0: /* Timeout */
790 		if (timeout)
791 			set_normalized_timespec(timeout, 0, 0);
792 		return -ETIMEDOUT;	/* -ETIME on Linux */
793 	default: /* Completed */
794 		WARN_ON(end < 0); /* We're not aware of other errors */
795 		return 0;
796 	}
797 }
798 
799 /**
800  * Waits for a sequence number to be signaled, and cleans up the
801  * request and object lists appropriately for that event.
802  */
803 int
804 i915_wait_seqno(struct intel_ring_buffer *ring, uint32_t seqno)
805 {
806 	struct drm_device *dev = ring->dev;
807 	struct drm_i915_private *dev_priv = dev->dev_private;
808 	bool interruptible = dev_priv->mm.interruptible;
809 	int ret;
810 
811 	DRM_LOCK_ASSERT(dev);
812 	BUG_ON(seqno == 0);
813 
814 	ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
815 	if (ret)
816 		return ret;
817 
818 	ret = i915_gem_check_olr(ring, seqno);
819 	if (ret)
820 		return ret;
821 
822 	return __wait_seqno(ring, seqno,
823 			    atomic_read(&dev_priv->gpu_error.reset_counter),
824 			    interruptible, NULL);
825 }
826 
827 /**
828  * Ensures that all rendering to the object has completed and the object is
829  * safe to unbind from the GTT or access from the CPU.
830  */
831 static __must_check int
832 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
833 			       bool readonly)
834 {
835 	struct intel_ring_buffer *ring = obj->ring;
836 	u32 seqno;
837 	int ret;
838 
839 	seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno;
840 	if (seqno == 0)
841 		return 0;
842 
843 	ret = i915_wait_seqno(ring, seqno);
844 	if (ret)
845 		return ret;
846 
847 	i915_gem_retire_requests_ring(ring);
848 
849 	/* Manually manage the write flush as we may have not yet
850 	 * retired the buffer.
851 	 */
852 	if (obj->last_write_seqno &&
853 	    i915_seqno_passed(seqno, obj->last_write_seqno)) {
854 		obj->last_write_seqno = 0;
855 		obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
856 	}
857 
858 	return 0;
859 }
860 
861 /* A nonblocking variant of the above wait. This is a highly dangerous routine
862  * as the object state may change during this call.
863  */
864 static __must_check int
865 i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
866 					    bool readonly)
867 {
868 	struct drm_device *dev = obj->base.dev;
869 	struct drm_i915_private *dev_priv = dev->dev_private;
870 	struct intel_ring_buffer *ring = obj->ring;
871 	unsigned reset_counter;
872 	u32 seqno;
873 	int ret;
874 
875 	DRM_LOCK_ASSERT(dev);
876 	BUG_ON(!dev_priv->mm.interruptible);
877 
878 	seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno;
879 	if (seqno == 0)
880 		return 0;
881 
882 	ret = i915_gem_check_wedge(&dev_priv->gpu_error, true);
883 	if (ret)
884 		return ret;
885 
886 	ret = i915_gem_check_olr(ring, seqno);
887 	if (ret)
888 		return ret;
889 
890 	reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
891 	mutex_unlock(&dev->struct_mutex);
892 	ret = __wait_seqno(ring, seqno, reset_counter, true, NULL);
893 	mutex_lock(&dev->struct_mutex);
894 
895 	i915_gem_retire_requests_ring(ring);
896 
897 	/* Manually manage the write flush as we may have not yet
898 	 * retired the buffer.
899 	 */
900 	if (obj->last_write_seqno &&
901 	    i915_seqno_passed(seqno, obj->last_write_seqno)) {
902 		obj->last_write_seqno = 0;
903 		obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
904 	}
905 
906 	return ret;
907 }
908 
909 /**
910  * Called when user space prepares to use an object with the CPU, either
911  * through the mmap ioctl's mapping or a GTT mapping.
912  */
913 int
914 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
915 			  struct drm_file *file)
916 {
917 	struct drm_i915_gem_set_domain *args = data;
918 	struct drm_i915_gem_object *obj;
919 	uint32_t read_domains = args->read_domains;
920 	uint32_t write_domain = args->write_domain;
921 	int ret;
922 
923 	/* Only handle setting domains to types used by the CPU. */
924 	if (write_domain & I915_GEM_GPU_DOMAINS)
925 		return -EINVAL;
926 
927 	if (read_domains & I915_GEM_GPU_DOMAINS)
928 		return -EINVAL;
929 
930 	/* Having something in the write domain implies it's in the read
931 	 * domain, and only that read domain.  Enforce that in the request.
932 	 */
933 	if (write_domain != 0 && read_domains != write_domain)
934 		return -EINVAL;
935 
936 	ret = i915_mutex_lock_interruptible(dev);
937 	if (ret)
938 		return ret;
939 
940 	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
941 	if (&obj->base == NULL) {
942 		ret = -ENOENT;
943 		goto unlock;
944 	}
945 
946 	/* Try to flush the object off the GPU without holding the lock.
947 	 * We will repeat the flush holding the lock in the normal manner
948 	 * to catch cases where we are gazumped.
949 	 */
950 	ret = i915_gem_object_wait_rendering__nonblocking(obj, !write_domain);
951 	if (ret)
952 		goto unref;
953 
954 	if (read_domains & I915_GEM_DOMAIN_GTT) {
955 		ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
956 
957 		/* Silently promote "you're not bound, there was nothing to do"
958 		 * to success, since the client was just asking us to
959 		 * make sure everything was done.
960 		 */
961 		if (ret == -EINVAL)
962 			ret = 0;
963 	} else {
964 		ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
965 	}
966 
967 unref:
968 	drm_gem_object_unreference(&obj->base);
969 unlock:
970 	mutex_unlock(&dev->struct_mutex);
971 	return ret;
972 }
973 
974 /**
975  * Called when user space has done writes to this buffer
976  */
977 int
978 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
979 			 struct drm_file *file)
980 {
981 	struct drm_i915_gem_sw_finish *args = data;
982 	struct drm_i915_gem_object *obj;
983 	int ret = 0;
984 
985 	ret = i915_mutex_lock_interruptible(dev);
986 	if (ret)
987 		return ret;
988 	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
989 	if (&obj->base == NULL) {
990 		ret = -ENOENT;
991 		goto unlock;
992 	}
993 
994 	/* Pinned buffers may be scanout, so flush the cache */
995 	if (obj->pin_count)
996 		i915_gem_object_flush_cpu_write_domain(obj);
997 
998 	drm_gem_object_unreference(&obj->base);
999 unlock:
1000 	mutex_unlock(&dev->struct_mutex);
1001 	return ret;
1002 }
1003 
1004 /**
1005  * Maps the contents of an object, returning the address it is mapped
1006  * into.
1007  *
1008  * While the mapping holds a reference on the contents of the object, it doesn't
1009  * imply a ref on the object itself.
1010  */
1011 int
1012 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1013 		    struct drm_file *file)
1014 {
1015 	struct drm_i915_gem_mmap *args = data;
1016 	struct drm_gem_object *obj;
1017 	struct proc *p = curproc;
1018 	vm_map_t map = &p->p_vmspace->vm_map;
1019 	vm_offset_t addr;
1020 	vm_size_t size;
1021 	int error = 0, rv;
1022 
1023 	obj = drm_gem_object_lookup(dev, file, args->handle);
1024 	if (obj == NULL)
1025 		return -ENOENT;
1026 
1027 	if (args->size == 0)
1028 		goto out;
1029 
1030 	size = round_page(args->size);
1031 	if (map->size + size > p->p_rlimit[RLIMIT_VMEM].rlim_cur) {
1032 		error = ENOMEM;
1033 		goto out;
1034 	}
1035 
1036 	addr = 0;
1037 	vm_object_hold(obj->vm_obj);
1038 	vm_object_reference_locked(obj->vm_obj);
1039 	vm_object_drop(obj->vm_obj);
1040 	rv = vm_map_find(map, obj->vm_obj, NULL,
1041 			 args->offset, &addr, args->size,
1042 			 PAGE_SIZE, /* align */
1043 			 TRUE, /* fitit */
1044 			 VM_MAPTYPE_NORMAL, /* maptype */
1045 			 VM_PROT_READ | VM_PROT_WRITE, /* prot */
1046 			 VM_PROT_READ | VM_PROT_WRITE, /* max */
1047 			 MAP_SHARED /* cow */);
1048 	if (rv != KERN_SUCCESS) {
1049 		vm_object_deallocate(obj->vm_obj);
1050 		error = -vm_mmap_to_errno(rv);
1051 	} else {
1052 		args->addr_ptr = (uint64_t)addr;
1053 	}
1054 out:
1055 	drm_gem_object_unreference(obj);
1056 	return (error);
1057 }
1058 
1059 /**
1060  * i915_gem_fault - fault a page into the GTT
1061  * vma: VMA in question
1062  * vmf: fault info
1063  *
1064  * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1065  * from userspace.  The fault handler takes care of binding the object to
1066  * the GTT (if needed), allocating and programming a fence register (again,
1067  * only if needed based on whether the old reg is still valid or the object
1068  * is tiled) and inserting a new PTE into the faulting process.
1069  *
1070  * Note that the faulting process may involve evicting existing objects
1071  * from the GTT and/or fence registers to make room.  So performance may
1072  * suffer if the GTT working set is large or there are few fence registers
1073  * left.
1074  */
1075 #if 0
1076 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1077 {
1078 	struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1079 	struct drm_device *dev = obj->base.dev;
1080 	drm_i915_private_t *dev_priv = dev->dev_private;
1081 	pgoff_t page_offset;
1082 	unsigned long pfn;
1083 	int ret = 0;
1084 	bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1085 
1086 	/* We don't use vmf->pgoff since that has the fake offset */
1087 	page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1088 		PAGE_SHIFT;
1089 
1090 	ret = i915_mutex_lock_interruptible(dev);
1091 	if (ret)
1092 		goto out;
1093 
1094 	trace_i915_gem_object_fault(obj, page_offset, true, write);
1095 
1096 	/* Access to snoopable pages through the GTT is incoherent. */
1097 	if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) {
1098 		ret = -EINVAL;
1099 		goto unlock;
1100 	}
1101 
1102 	/* Now bind it into the GTT if needed */
1103 	ret = i915_gem_object_pin(obj, 0, true, false);
1104 	if (ret)
1105 		goto unlock;
1106 
1107 	ret = i915_gem_object_set_to_gtt_domain(obj, write);
1108 	if (ret)
1109 		goto unpin;
1110 
1111 	ret = i915_gem_object_get_fence(obj);
1112 	if (ret)
1113 		goto unpin;
1114 
1115 	obj->fault_mappable = true;
1116 
1117 	pfn = ((dev_priv->gtt.mappable_base + obj->gtt_offset) >> PAGE_SHIFT) +
1118 		page_offset;
1119 
1120 	/* Finally, remap it using the new GTT offset */
1121 	ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1122 unpin:
1123 	i915_gem_object_unpin(obj);
1124 unlock:
1125 	mutex_unlock(&dev->struct_mutex);
1126 out:
1127 	switch (ret) {
1128 	case -EIO:
1129 		/* If this -EIO is due to a gpu hang, give the reset code a
1130 		 * chance to clean up the mess. Otherwise return the proper
1131 		 * SIGBUS. */
1132 		if (i915_terminally_wedged(&dev_priv->gpu_error))
1133 			return VM_FAULT_SIGBUS;
1134 	case -EAGAIN:
1135 		/* Give the error handler a chance to run and move the
1136 		 * objects off the GPU active list. Next time we service the
1137 		 * fault, we should be able to transition the page into the
1138 		 * GTT without touching the GPU (and so avoid further
1139 		 * EIO/EGAIN). If the GPU is wedged, then there is no issue
1140 		 * with coherency, just lost writes.
1141 		 */
1142 		set_need_resched();
1143 	case 0:
1144 	case -ERESTARTSYS:
1145 	case -EINTR:
1146 	case -EBUSY:
1147 		/*
1148 		 * EBUSY is ok: this just means that another thread
1149 		 * already did the job.
1150 		 */
1151 		return VM_FAULT_NOPAGE;
1152 	case -ENOMEM:
1153 		return VM_FAULT_OOM;
1154 	case -ENOSPC:
1155 		return VM_FAULT_SIGBUS;
1156 	default:
1157 		WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret);
1158 		return VM_FAULT_SIGBUS;
1159 	}
1160 }
1161 #endif
1162 
1163 /**
1164  * i915_gem_release_mmap - remove physical page mappings
1165  * @obj: obj in question
1166  *
1167  * Preserve the reservation of the mmapping with the DRM core code, but
1168  * relinquish ownership of the pages back to the system.
1169  *
1170  * It is vital that we remove the page mapping if we have mapped a tiled
1171  * object through the GTT and then lose the fence register due to
1172  * resource pressure. Similarly if the object has been moved out of the
1173  * aperture, than pages mapped into userspace must be revoked. Removing the
1174  * mapping will then trigger a page fault on the next user access, allowing
1175  * fixup by i915_gem_fault().
1176  */
1177 void
1178 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1179 {
1180 	vm_object_t devobj;
1181 	vm_page_t m;
1182 	int i, page_count;
1183 
1184 	if (!obj->fault_mappable)
1185 		return;
1186 
1187 	devobj = cdev_pager_lookup(obj);
1188 	if (devobj != NULL) {
1189 		page_count = OFF_TO_IDX(obj->base.size);
1190 
1191 		VM_OBJECT_LOCK(devobj);
1192 		for (i = 0; i < page_count; i++) {
1193 			m = vm_page_lookup_busy_wait(devobj, i, TRUE, "915unm");
1194 			if (m == NULL)
1195 				continue;
1196 			cdev_pager_free_page(devobj, m);
1197 		}
1198 		VM_OBJECT_UNLOCK(devobj);
1199 		vm_object_deallocate(devobj);
1200 	}
1201 
1202 	obj->fault_mappable = false;
1203 }
1204 
1205 uint32_t
1206 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1207 {
1208 	uint32_t gtt_size;
1209 
1210 	if (INTEL_INFO(dev)->gen >= 4 ||
1211 	    tiling_mode == I915_TILING_NONE)
1212 		return size;
1213 
1214 	/* Previous chips need a power-of-two fence region when tiling */
1215 	if (INTEL_INFO(dev)->gen == 3)
1216 		gtt_size = 1024*1024;
1217 	else
1218 		gtt_size = 512*1024;
1219 
1220 	while (gtt_size < size)
1221 		gtt_size <<= 1;
1222 
1223 	return gtt_size;
1224 }
1225 
1226 /**
1227  * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1228  * @obj: object to check
1229  *
1230  * Return the required GTT alignment for an object, taking into account
1231  * potential fence register mapping.
1232  */
1233 uint32_t
1234 i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size,
1235 			   int tiling_mode, bool fenced)
1236 {
1237 
1238 	/*
1239 	 * Minimum alignment is 4k (GTT page size), but might be greater
1240 	 * if a fence register is needed for the object.
1241 	 */
1242 	if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) ||
1243 	    tiling_mode == I915_TILING_NONE)
1244 		return 4096;
1245 
1246 	/*
1247 	 * Previous chips need to be aligned to the size of the smallest
1248 	 * fence register that can contain the object.
1249 	 */
1250 	return i915_gem_get_gtt_size(dev, size, tiling_mode);
1251 }
1252 
1253 int
1254 i915_gem_mmap_gtt(struct drm_file *file,
1255 		  struct drm_device *dev,
1256 		  uint32_t handle,
1257 		  uint64_t *offset)
1258 {
1259 	struct drm_i915_private *dev_priv = dev->dev_private;
1260 	struct drm_i915_gem_object *obj;
1261 	int ret;
1262 
1263 	ret = i915_mutex_lock_interruptible(dev);
1264 	if (ret)
1265 		return ret;
1266 
1267 	obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
1268 	if (&obj->base == NULL) {
1269 		ret = -ENOENT;
1270 		goto unlock;
1271 	}
1272 
1273 	if (obj->base.size > dev_priv->gtt.mappable_end) {
1274 		ret = -E2BIG;
1275 		goto out;
1276 	}
1277 
1278 	if (obj->madv != I915_MADV_WILLNEED) {
1279 		DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1280 		ret = -EINVAL;
1281 		goto out;
1282 	}
1283 
1284 	ret = drm_gem_create_mmap_offset(&obj->base);
1285 	if (ret)
1286 		goto out;
1287 
1288 	*offset = DRM_GEM_MAPPING_OFF(obj->base.map_list.key) |
1289 	    DRM_GEM_MAPPING_KEY;
1290 out:
1291 	drm_gem_object_unreference(&obj->base);
1292 unlock:
1293 	mutex_unlock(&dev->struct_mutex);
1294 	return ret;
1295 }
1296 
1297 /**
1298  * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1299  * @dev: DRM device
1300  * @data: GTT mapping ioctl data
1301  * @file: GEM object info
1302  *
1303  * Simply returns the fake offset to userspace so it can mmap it.
1304  * The mmap call will end up in drm_gem_mmap(), which will set things
1305  * up so we can get faults in the handler above.
1306  *
1307  * The fault handler will take care of binding the object into the GTT
1308  * (since it may have been evicted to make room for something), allocating
1309  * a fence register, and mapping the appropriate aperture address into
1310  * userspace.
1311  */
1312 int
1313 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1314 			struct drm_file *file)
1315 {
1316 	struct drm_i915_gem_mmap_gtt *args = data;
1317 
1318 	return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
1319 }
1320 
1321 /* Immediately discard the backing storage */
1322 static void
1323 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
1324 {
1325 	vm_object_t vm_obj;
1326 
1327 	vm_obj = obj->base.vm_obj;
1328 	VM_OBJECT_LOCK(vm_obj);
1329 	vm_object_page_remove(vm_obj, 0, 0, false);
1330 	VM_OBJECT_UNLOCK(vm_obj);
1331 	obj->madv = __I915_MADV_PURGED;
1332 }
1333 
1334 static inline int
1335 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj)
1336 {
1337 	return obj->madv == I915_MADV_DONTNEED;
1338 }
1339 
1340 static void
1341 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
1342 {
1343 	vm_page_t m;
1344 	int page_count, i;
1345 
1346 	BUG_ON(obj->madv == __I915_MADV_PURGED);
1347 
1348 	if (obj->tiling_mode != I915_TILING_NONE)
1349 		i915_gem_object_save_bit_17_swizzle(obj);
1350 	if (obj->madv == I915_MADV_DONTNEED)
1351 		obj->dirty = 0;
1352 	page_count = obj->base.size / PAGE_SIZE;
1353 	VM_OBJECT_LOCK(obj->base.vm_obj);
1354 #if GEM_PARANOID_CHECK_GTT
1355 	i915_gem_assert_pages_not_mapped(obj->base.dev, obj->pages, page_count);
1356 #endif
1357 	for (i = 0; i < page_count; i++) {
1358 		m = obj->pages[i];
1359 		if (obj->dirty)
1360 			vm_page_dirty(m);
1361 		if (obj->madv == I915_MADV_WILLNEED)
1362 			vm_page_reference(m);
1363 		vm_page_busy_wait(obj->pages[i], FALSE, "i915gem");
1364 		vm_page_unwire(obj->pages[i], 1);
1365 		vm_page_wakeup(obj->pages[i]);
1366 	}
1367 	VM_OBJECT_UNLOCK(obj->base.vm_obj);
1368 	obj->dirty = 0;
1369 	drm_free(obj->pages, M_DRM);
1370 	obj->pages = NULL;
1371 }
1372 
1373 int
1374 i915_gem_object_put_pages(struct drm_i915_gem_object *obj)
1375 {
1376 	const struct drm_i915_gem_object_ops *ops = obj->ops;
1377 
1378 	if (obj->pages == NULL)
1379 		return 0;
1380 
1381 	BUG_ON(obj->gtt_space);
1382 
1383 	if (obj->pages_pin_count)
1384 		return -EBUSY;
1385 
1386 	/* ->put_pages might need to allocate memory for the bit17 swizzle
1387 	 * array, hence protect them from being reaped by removing them from gtt
1388 	 * lists early. */
1389 	list_del(&obj->gtt_list);
1390 
1391 	ops->put_pages(obj);
1392 	obj->pages = NULL;
1393 
1394 	if (i915_gem_object_is_purgeable(obj))
1395 		i915_gem_object_truncate(obj);
1396 
1397 	return 0;
1398 }
1399 
1400 static int
1401 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
1402 {
1403 	struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1404 	struct drm_device *dev;
1405 	vm_object_t vm_obj;
1406 	int page_count, i, j;
1407 	struct vm_page *page;
1408 
1409 	dev = obj->base.dev;
1410 	KASSERT(obj->pages == NULL, ("Obj already has pages"));
1411 	page_count = obj->base.size / PAGE_SIZE;
1412 	obj->pages = kmalloc(page_count * sizeof(vm_page_t), M_DRM,
1413 	    M_WAITOK);
1414 
1415 	vm_obj = obj->base.vm_obj;
1416 	VM_OBJECT_LOCK(vm_obj);
1417 
1418 	for (i = 0; i < page_count; i++) {
1419 		page = shmem_read_mapping_page(vm_obj, i);
1420 		if (IS_ERR(page)) {
1421 			i915_gem_purge(dev_priv, page_count);
1422 			goto err_pages;
1423 		}
1424 
1425 		obj->pages[i] = page;
1426 	}
1427 
1428 	VM_OBJECT_UNLOCK(vm_obj);
1429 	if (i915_gem_object_needs_bit17_swizzle(obj))
1430 		i915_gem_object_do_bit_17_swizzle(obj);
1431 
1432 	return 0;
1433 
1434 err_pages:
1435 	for (j = 0; j < i; j++) {
1436 		page = obj->pages[j];
1437 		vm_page_busy_wait(page, FALSE, "i915gem");
1438 		vm_page_unwire(page, 0);
1439 		vm_page_wakeup(page);
1440 	}
1441 	VM_OBJECT_UNLOCK(vm_obj);
1442 	drm_free(obj->pages, M_DRM);
1443 	obj->pages = NULL;
1444 	return (-EIO);
1445 }
1446 
1447 /* Ensure that the associated pages are gathered from the backing storage
1448  * and pinned into our object. i915_gem_object_get_pages() may be called
1449  * multiple times before they are released by a single call to
1450  * i915_gem_object_put_pages() - once the pages are no longer referenced
1451  * either as a result of memory pressure (reaping pages under the shrinker)
1452  * or as the object is itself released.
1453  */
1454 int
1455 i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
1456 {
1457 	struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1458 	const struct drm_i915_gem_object_ops *ops = obj->ops;
1459 	int ret;
1460 
1461 	if (obj->pages)
1462 		return 0;
1463 
1464 	if (obj->madv != I915_MADV_WILLNEED) {
1465 		DRM_ERROR("Attempting to obtain a purgeable object\n");
1466 		return -EINVAL;
1467 	}
1468 
1469 	BUG_ON(obj->pages_pin_count);
1470 
1471 	ret = ops->get_pages(obj);
1472 	if (ret)
1473 		return ret;
1474 
1475 	list_add_tail(&obj->gtt_list, &dev_priv->mm.unbound_list);
1476 	return 0;
1477 }
1478 
1479 void
1480 i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
1481 			       struct intel_ring_buffer *ring)
1482 {
1483 	struct drm_device *dev = obj->base.dev;
1484 	struct drm_i915_private *dev_priv = dev->dev_private;
1485 	u32 seqno = intel_ring_get_seqno(ring);
1486 
1487 	BUG_ON(ring == NULL);
1488 	obj->ring = ring;
1489 
1490 	/* Add a reference if we're newly entering the active list. */
1491 	if (!obj->active) {
1492 		drm_gem_object_reference(&obj->base);
1493 		obj->active = 1;
1494 	}
1495 
1496 	/* Move from whatever list we were on to the tail of execution. */
1497 	list_move_tail(&obj->mm_list, &dev_priv->mm.active_list);
1498 	list_move_tail(&obj->ring_list, &ring->active_list);
1499 
1500 	obj->last_read_seqno = seqno;
1501 
1502 	if (obj->fenced_gpu_access) {
1503 		obj->last_fenced_seqno = seqno;
1504 
1505 		/* Bump MRU to take account of the delayed flush */
1506 		if (obj->fence_reg != I915_FENCE_REG_NONE) {
1507 			struct drm_i915_fence_reg *reg;
1508 
1509 			reg = &dev_priv->fence_regs[obj->fence_reg];
1510 			list_move_tail(&reg->lru_list,
1511 				       &dev_priv->mm.fence_list);
1512 		}
1513 	}
1514 }
1515 
1516 static void
1517 i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
1518 {
1519 	struct drm_device *dev = obj->base.dev;
1520 	struct drm_i915_private *dev_priv = dev->dev_private;
1521 
1522 	BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS);
1523 	BUG_ON(!obj->active);
1524 
1525 	list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1526 
1527 	list_del_init(&obj->ring_list);
1528 	obj->ring = NULL;
1529 
1530 	obj->last_read_seqno = 0;
1531 	obj->last_write_seqno = 0;
1532 	obj->base.write_domain = 0;
1533 
1534 	obj->last_fenced_seqno = 0;
1535 	obj->fenced_gpu_access = false;
1536 
1537 	obj->active = 0;
1538 	drm_gem_object_unreference(&obj->base);
1539 
1540 	WARN_ON(i915_verify_lists(dev));
1541 }
1542 
1543 static int
1544 i915_gem_init_seqno(struct drm_device *dev, u32 seqno)
1545 {
1546 	struct drm_i915_private *dev_priv = dev->dev_private;
1547 	struct intel_ring_buffer *ring;
1548 	int ret, i, j;
1549 
1550 	/* Carefully retire all requests without writing to the rings */
1551 	for_each_ring(ring, dev_priv, i) {
1552 		ret = intel_ring_idle(ring);
1553 		if (ret)
1554 			return ret;
1555 	}
1556 	i915_gem_retire_requests(dev);
1557 
1558 	/* Finally reset hw state */
1559 	for_each_ring(ring, dev_priv, i) {
1560 		intel_ring_init_seqno(ring, seqno);
1561 
1562 		for (j = 0; j < ARRAY_SIZE(ring->sync_seqno); j++)
1563 			ring->sync_seqno[j] = 0;
1564 	}
1565 
1566 	return 0;
1567 }
1568 
1569 int i915_gem_set_seqno(struct drm_device *dev, u32 seqno)
1570 {
1571 	struct drm_i915_private *dev_priv = dev->dev_private;
1572 	int ret;
1573 
1574 	if (seqno == 0)
1575 		return -EINVAL;
1576 
1577 	/* HWS page needs to be set less than what we
1578 	 * will inject to ring
1579 	 */
1580 	ret = i915_gem_init_seqno(dev, seqno - 1);
1581 	if (ret)
1582 		return ret;
1583 
1584 	/* Carefully set the last_seqno value so that wrap
1585 	 * detection still works
1586 	 */
1587 	dev_priv->next_seqno = seqno;
1588 	dev_priv->last_seqno = seqno - 1;
1589 	if (dev_priv->last_seqno == 0)
1590 		dev_priv->last_seqno--;
1591 
1592 	return 0;
1593 }
1594 
1595 int
1596 i915_gem_get_seqno(struct drm_device *dev, u32 *seqno)
1597 {
1598 	struct drm_i915_private *dev_priv = dev->dev_private;
1599 
1600 	/* reserve 0 for non-seqno */
1601 	if (dev_priv->next_seqno == 0) {
1602 		int ret = i915_gem_init_seqno(dev, 0);
1603 		if (ret)
1604 			return ret;
1605 
1606 		dev_priv->next_seqno = 1;
1607 	}
1608 
1609 	*seqno = dev_priv->last_seqno = dev_priv->next_seqno++;
1610 	return 0;
1611 }
1612 
1613 int
1614 i915_add_request(struct intel_ring_buffer *ring,
1615 		 struct drm_file *file,
1616 		 u32 *out_seqno)
1617 {
1618 	drm_i915_private_t *dev_priv = ring->dev->dev_private;
1619 	struct drm_i915_gem_request *request;
1620 	u32 request_ring_position;
1621 	int was_empty;
1622 	int ret;
1623 
1624 	/*
1625 	 * Emit any outstanding flushes - execbuf can fail to emit the flush
1626 	 * after having emitted the batchbuffer command. Hence we need to fix
1627 	 * things up similar to emitting the lazy request. The difference here
1628 	 * is that the flush _must_ happen before the next request, no matter
1629 	 * what.
1630 	 */
1631 	ret = intel_ring_flush_all_caches(ring);
1632 	if (ret)
1633 		return ret;
1634 
1635 	request = kmalloc(sizeof(*request), M_DRM, M_WAITOK | M_ZERO);
1636 	if (request == NULL)
1637 		return -ENOMEM;
1638 
1639 
1640 	/* Record the position of the start of the request so that
1641 	 * should we detect the updated seqno part-way through the
1642 	 * GPU processing the request, we never over-estimate the
1643 	 * position of the head.
1644 	 */
1645 	request_ring_position = intel_ring_get_tail(ring);
1646 
1647 	ret = ring->add_request(ring);
1648 	if (ret) {
1649 		kfree(request, M_DRM);
1650 		return ret;
1651 	}
1652 
1653 	request->seqno = intel_ring_get_seqno(ring);
1654 	request->ring = ring;
1655 	request->tail = request_ring_position;
1656 	request->emitted_jiffies = jiffies;
1657 	was_empty = list_empty(&ring->request_list);
1658 	list_add_tail(&request->list, &ring->request_list);
1659 	request->file_priv = NULL;
1660 
1661 	if (file) {
1662 		struct drm_i915_file_private *file_priv = file->driver_priv;
1663 
1664 		spin_lock(&file_priv->mm.lock);
1665 		request->file_priv = file_priv;
1666 		list_add_tail(&request->client_list,
1667 			      &file_priv->mm.request_list);
1668 		spin_unlock(&file_priv->mm.lock);
1669 	}
1670 
1671 	ring->outstanding_lazy_request = 0;
1672 
1673 	if (!dev_priv->mm.suspended) {
1674 		if (i915_enable_hangcheck) {
1675 			mod_timer(&dev_priv->gpu_error.hangcheck_timer,
1676 				  round_jiffies_up(jiffies + DRM_I915_HANGCHECK_JIFFIES));
1677 		}
1678 		if (was_empty) {
1679 			queue_delayed_work(dev_priv->wq,
1680 					   &dev_priv->mm.retire_work,
1681 					   round_jiffies_up_relative(hz));
1682 			intel_mark_busy(dev_priv->dev);
1683 		}
1684 	}
1685 
1686 	if (out_seqno)
1687 		*out_seqno = request->seqno;
1688 	return 0;
1689 }
1690 
1691 static inline void
1692 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
1693 {
1694 	struct drm_i915_file_private *file_priv = request->file_priv;
1695 
1696 	if (!file_priv)
1697 		return;
1698 
1699 	spin_lock(&file_priv->mm.lock);
1700 	if (request->file_priv) {
1701 		list_del(&request->client_list);
1702 		request->file_priv = NULL;
1703 	}
1704 	spin_unlock(&file_priv->mm.lock);
1705 }
1706 
1707 static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv,
1708 				      struct intel_ring_buffer *ring)
1709 {
1710 	while (!list_empty(&ring->request_list)) {
1711 		struct drm_i915_gem_request *request;
1712 
1713 		request = list_first_entry(&ring->request_list,
1714 					   struct drm_i915_gem_request,
1715 					   list);
1716 
1717 		list_del(&request->list);
1718 		i915_gem_request_remove_from_client(request);
1719 		drm_free(request, M_DRM);
1720 	}
1721 
1722 	while (!list_empty(&ring->active_list)) {
1723 		struct drm_i915_gem_object *obj;
1724 
1725 		obj = list_first_entry(&ring->active_list,
1726 				       struct drm_i915_gem_object,
1727 				       ring_list);
1728 
1729 		i915_gem_object_move_to_inactive(obj);
1730 	}
1731 }
1732 
1733 static void i915_gem_reset_fences(struct drm_device *dev)
1734 {
1735 	struct drm_i915_private *dev_priv = dev->dev_private;
1736 	int i;
1737 
1738 	for (i = 0; i < dev_priv->num_fence_regs; i++) {
1739 		struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
1740 
1741 		i915_gem_write_fence(dev, i, NULL);
1742 
1743 		if (reg->obj)
1744 			i915_gem_object_fence_lost(reg->obj);
1745 
1746 		reg->pin_count = 0;
1747 		reg->obj = NULL;
1748 		INIT_LIST_HEAD(&reg->lru_list);
1749 	}
1750 
1751 	INIT_LIST_HEAD(&dev_priv->mm.fence_list);
1752 }
1753 
1754 void i915_gem_reset(struct drm_device *dev)
1755 {
1756 	struct drm_i915_private *dev_priv = dev->dev_private;
1757 	struct drm_i915_gem_object *obj;
1758 	struct intel_ring_buffer *ring;
1759 	int i;
1760 
1761 	for_each_ring(ring, dev_priv, i)
1762 		i915_gem_reset_ring_lists(dev_priv, ring);
1763 
1764 	/* Move everything out of the GPU domains to ensure we do any
1765 	 * necessary invalidation upon reuse.
1766 	 */
1767 	list_for_each_entry(obj,
1768 			    &dev_priv->mm.inactive_list,
1769 			    mm_list)
1770 	{
1771 		obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
1772 	}
1773 
1774 	/* The fence registers are invalidated so clear them out */
1775 	i915_gem_reset_fences(dev);
1776 }
1777 
1778 /**
1779  * This function clears the request list as sequence numbers are passed.
1780  */
1781 void
1782 i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
1783 {
1784 	uint32_t seqno;
1785 
1786 	if (list_empty(&ring->request_list))
1787 		return;
1788 
1789 	WARN_ON(i915_verify_lists(ring->dev));
1790 
1791 	seqno = ring->get_seqno(ring, true);
1792 
1793 	while (!list_empty(&ring->request_list)) {
1794 		struct drm_i915_gem_request *request;
1795 
1796 		request = list_first_entry(&ring->request_list,
1797 					   struct drm_i915_gem_request,
1798 					   list);
1799 
1800 		if (!i915_seqno_passed(seqno, request->seqno))
1801 			break;
1802 
1803 		/* We know the GPU must have read the request to have
1804 		 * sent us the seqno + interrupt, so use the position
1805 		 * of tail of the request to update the last known position
1806 		 * of the GPU head.
1807 		 */
1808 		ring->last_retired_head = request->tail;
1809 
1810 		list_del(&request->list);
1811 		i915_gem_request_remove_from_client(request);
1812 		kfree(request, M_DRM);
1813 	}
1814 
1815 	/* Move any buffers on the active list that are no longer referenced
1816 	 * by the ringbuffer to the flushing/inactive lists as appropriate.
1817 	 */
1818 	while (!list_empty(&ring->active_list)) {
1819 		struct drm_i915_gem_object *obj;
1820 
1821 		obj = list_first_entry(&ring->active_list,
1822 				      struct drm_i915_gem_object,
1823 				      ring_list);
1824 
1825 		if (!i915_seqno_passed(seqno, obj->last_read_seqno))
1826 			break;
1827 
1828 		i915_gem_object_move_to_inactive(obj);
1829 	}
1830 
1831 	if (unlikely(ring->trace_irq_seqno &&
1832 		     i915_seqno_passed(seqno, ring->trace_irq_seqno))) {
1833 		ring->irq_put(ring);
1834 		ring->trace_irq_seqno = 0;
1835 	}
1836 
1837 }
1838 
1839 void
1840 i915_gem_retire_requests(struct drm_device *dev)
1841 {
1842 	drm_i915_private_t *dev_priv = dev->dev_private;
1843 	struct intel_ring_buffer *ring;
1844 	int i;
1845 
1846 	for_each_ring(ring, dev_priv, i)
1847 		i915_gem_retire_requests_ring(ring);
1848 }
1849 
1850 static long
1851 __i915_gem_shrink(struct drm_i915_private *dev_priv, long target,
1852 		  bool purgeable_only)
1853 {
1854 	struct drm_i915_gem_object *obj, *next;
1855 	long count = 0;
1856 
1857 	list_for_each_entry_safe(obj, next,
1858 				 &dev_priv->mm.unbound_list,
1859 				 gtt_list) {
1860 #if 0
1861 		if ((i915_gem_object_is_purgeable(obj) || !purgeable_only) &&
1862 		    i915_gem_object_put_pages(obj) == 0) {
1863 			count += obj->base.size >> PAGE_SHIFT;
1864 			if (count >= target)
1865 				return count;
1866 		}
1867 #endif
1868 	}
1869 
1870 	list_for_each_entry_safe(obj, next,
1871 				 &dev_priv->mm.inactive_list,
1872 				 mm_list) {
1873 #if 0
1874 		if ((i915_gem_object_is_purgeable(obj) || !purgeable_only) &&
1875 		    i915_gem_object_unbind(obj) == 0 &&
1876 		    i915_gem_object_put_pages(obj) == 0) {
1877 			count += obj->base.size >> PAGE_SHIFT;
1878 			if (count >= target)
1879 				return count;
1880 		}
1881 #endif
1882 	}
1883 
1884 	return count;
1885 }
1886 
1887 static long
1888 i915_gem_purge(struct drm_i915_private *dev_priv, long target)
1889 {
1890 	return __i915_gem_shrink(dev_priv, target, true);
1891 }
1892 
1893 static void
1894 i915_gem_retire_work_handler(struct work_struct *work)
1895 {
1896 	drm_i915_private_t *dev_priv;
1897 	struct drm_device *dev;
1898 	struct intel_ring_buffer *ring;
1899 	bool idle;
1900 	int i;
1901 
1902 	dev_priv = container_of(work, drm_i915_private_t,
1903 				mm.retire_work.work);
1904 	dev = dev_priv->dev;
1905 
1906 	/* Come back later if the device is busy... */
1907 	if (lockmgr(&dev->struct_mutex, LK_EXCLUSIVE|LK_NOWAIT)) {
1908 		queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
1909 				   round_jiffies_up_relative(hz));
1910 		return;
1911 	}
1912 
1913 	i915_gem_retire_requests(dev);
1914 
1915 	/* Send a periodic flush down the ring so we don't hold onto GEM
1916 	 * objects indefinitely.
1917 	 */
1918 	idle = true;
1919 	for_each_ring(ring, dev_priv, i) {
1920 		if (ring->gpu_caches_dirty)
1921 			i915_add_request(ring, NULL, NULL);
1922 
1923 		idle &= list_empty(&ring->request_list);
1924 	}
1925 
1926 	if (!dev_priv->mm.suspended && !idle)
1927 		queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
1928 				   round_jiffies_up_relative(hz));
1929 	if (idle)
1930 		intel_mark_idle(dev);
1931 
1932 	mutex_unlock(&dev->struct_mutex);
1933 }
1934 /**
1935  * Ensures that an object will eventually get non-busy by flushing any required
1936  * write domains, emitting any outstanding lazy request and retiring and
1937  * completed requests.
1938  */
1939 static int
1940 i915_gem_object_flush_active(struct drm_i915_gem_object *obj)
1941 {
1942 	int ret;
1943 
1944 	if (obj->active) {
1945 		ret = i915_gem_check_olr(obj->ring, obj->last_read_seqno);
1946 		if (ret)
1947 			return ret;
1948 
1949 		i915_gem_retire_requests_ring(obj->ring);
1950 	}
1951 
1952 	return 0;
1953 }
1954 
1955 /**
1956  * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
1957  * @DRM_IOCTL_ARGS: standard ioctl arguments
1958  *
1959  * Returns 0 if successful, else an error is returned with the remaining time in
1960  * the timeout parameter.
1961  *  -ETIME: object is still busy after timeout
1962  *  -ERESTARTSYS: signal interrupted the wait
1963  *  -ENONENT: object doesn't exist
1964  * Also possible, but rare:
1965  *  -EAGAIN: GPU wedged
1966  *  -ENOMEM: damn
1967  *  -ENODEV: Internal IRQ fail
1968  *  -E?: The add request failed
1969  *
1970  * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
1971  * non-zero timeout parameter the wait ioctl will wait for the given number of
1972  * nanoseconds on an object becoming unbusy. Since the wait itself does so
1973  * without holding struct_mutex the object may become re-busied before this
1974  * function completes. A similar but shorter * race condition exists in the busy
1975  * ioctl
1976  */
1977 int
1978 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
1979 {
1980 	drm_i915_private_t *dev_priv = dev->dev_private;
1981 	struct drm_i915_gem_wait *args = data;
1982 	struct drm_i915_gem_object *obj;
1983 	struct intel_ring_buffer *ring = NULL;
1984 	struct timespec timeout_stack, *timeout = NULL;
1985 	unsigned reset_counter;
1986 	u32 seqno = 0;
1987 	int ret = 0;
1988 
1989 	if (args->timeout_ns >= 0) {
1990 		timeout_stack = ns_to_timespec(args->timeout_ns);
1991 		timeout = &timeout_stack;
1992 	}
1993 
1994 	ret = i915_mutex_lock_interruptible(dev);
1995 	if (ret)
1996 		return ret;
1997 
1998 	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle));
1999 	if (&obj->base == NULL) {
2000 		mutex_unlock(&dev->struct_mutex);
2001 		return -ENOENT;
2002 	}
2003 
2004 	/* Need to make sure the object gets inactive eventually. */
2005 	ret = i915_gem_object_flush_active(obj);
2006 	if (ret)
2007 		goto out;
2008 
2009 	if (obj->active) {
2010 		seqno = obj->last_read_seqno;
2011 		ring = obj->ring;
2012 	}
2013 
2014 	if (seqno == 0)
2015 		 goto out;
2016 
2017 	/* Do this after OLR check to make sure we make forward progress polling
2018 	 * on this IOCTL with a 0 timeout (like busy ioctl)
2019 	 */
2020 	if (!args->timeout_ns) {
2021 		ret = -ETIMEDOUT;
2022 		goto out;
2023 	}
2024 
2025 	drm_gem_object_unreference(&obj->base);
2026 	reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
2027 	mutex_unlock(&dev->struct_mutex);
2028 
2029 	ret = __wait_seqno(ring, seqno, reset_counter, true, timeout);
2030 	if (timeout) {
2031 		WARN_ON(!timespec_valid(timeout));
2032 		args->timeout_ns = timespec_to_ns(timeout);
2033 	}
2034 	return ret;
2035 
2036 out:
2037 	drm_gem_object_unreference(&obj->base);
2038 	mutex_unlock(&dev->struct_mutex);
2039 	return ret;
2040 }
2041 
2042 /**
2043  * i915_gem_object_sync - sync an object to a ring.
2044  *
2045  * @obj: object which may be in use on another ring.
2046  * @to: ring we wish to use the object on. May be NULL.
2047  *
2048  * This code is meant to abstract object synchronization with the GPU.
2049  * Calling with NULL implies synchronizing the object with the CPU
2050  * rather than a particular GPU ring.
2051  *
2052  * Returns 0 if successful, else propagates up the lower layer error.
2053  */
2054 int
2055 i915_gem_object_sync(struct drm_i915_gem_object *obj,
2056 		     struct intel_ring_buffer *to)
2057 {
2058 	struct intel_ring_buffer *from = obj->ring;
2059 	u32 seqno;
2060 	int ret, idx;
2061 
2062 	if (from == NULL || to == from)
2063 		return 0;
2064 
2065 	if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev))
2066 		return i915_gem_object_wait_rendering(obj, false);
2067 
2068 	idx = intel_ring_sync_index(from, to);
2069 
2070 	seqno = obj->last_read_seqno;
2071 	if (seqno <= from->sync_seqno[idx])
2072 		return 0;
2073 
2074 	ret = i915_gem_check_olr(obj->ring, seqno);
2075 	if (ret)
2076 		return ret;
2077 
2078 	ret = to->sync_to(to, from, seqno);
2079 	if (!ret)
2080 		/* We use last_read_seqno because sync_to()
2081 		 * might have just caused seqno wrap under
2082 		 * the radar.
2083 		 */
2084 		from->sync_seqno[idx] = obj->last_read_seqno;
2085 
2086 	return ret;
2087 }
2088 
2089 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
2090 {
2091 	u32 old_write_domain, old_read_domains;
2092 
2093 	/* Force a pagefault for domain tracking on next user access */
2094 	i915_gem_release_mmap(obj);
2095 
2096 	if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
2097 		return;
2098 
2099 	/* Wait for any direct GTT access to complete */
2100 	cpu_mfence();
2101 
2102 	old_read_domains = obj->base.read_domains;
2103 	old_write_domain = obj->base.write_domain;
2104 
2105 	obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
2106 	obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
2107 
2108 }
2109 
2110 /**
2111  * Unbinds an object from the GTT aperture.
2112  */
2113 int
2114 i915_gem_object_unbind(struct drm_i915_gem_object *obj)
2115 {
2116 	drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
2117 	int ret;
2118 
2119 	if (obj->gtt_space == NULL)
2120 		return 0;
2121 
2122 	if (obj->pin_count)
2123 		return -EBUSY;
2124 
2125 	BUG_ON(obj->pages == NULL);
2126 
2127 	ret = i915_gem_object_finish_gpu(obj);
2128 	if (ret)
2129 		return ret;
2130 	/* Continue on if we fail due to EIO, the GPU is hung so we
2131 	 * should be safe and we need to cleanup or else we might
2132 	 * cause memory corruption through use-after-free.
2133 	 */
2134 
2135 	i915_gem_object_finish_gtt(obj);
2136 
2137 	/* Move the object to the CPU domain to ensure that
2138 	 * any possible CPU writes while it's not in the GTT
2139 	 * are flushed when we go to remap it.
2140 	 */
2141 	if (ret == 0)
2142 		ret = i915_gem_object_set_to_cpu_domain(obj, 1);
2143 	if (ret == -ERESTARTSYS)
2144 		return ret;
2145 	if (ret) {
2146 		/* In the event of a disaster, abandon all caches and
2147 		 * hope for the best.
2148 		 */
2149 		i915_gem_clflush_object(obj);
2150 		obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
2151 	}
2152 
2153 	/* release the fence reg _after_ flushing */
2154 	ret = i915_gem_object_put_fence(obj);
2155 	if (ret)
2156 		return ret;
2157 
2158 	if (obj->has_global_gtt_mapping)
2159 		i915_gem_gtt_unbind_object(obj);
2160 	if (obj->has_aliasing_ppgtt_mapping) {
2161 		i915_ppgtt_unbind_object(dev_priv->mm.aliasing_ppgtt, obj);
2162 		obj->has_aliasing_ppgtt_mapping = 0;
2163 	}
2164 	i915_gem_gtt_finish_object(obj);
2165 
2166 	i915_gem_object_put_pages_gtt(obj);
2167 
2168 	list_del_init(&obj->gtt_list);
2169 	list_del_init(&obj->mm_list);
2170 	/* Avoid an unnecessary call to unbind on rebind. */
2171 	obj->map_and_fenceable = true;
2172 
2173 	drm_mm_put_block(obj->gtt_space);
2174 	obj->gtt_space = NULL;
2175 	obj->gtt_offset = 0;
2176 
2177 	if (i915_gem_object_is_purgeable(obj))
2178 		i915_gem_object_truncate(obj);
2179 
2180 	return ret;
2181 }
2182 
2183 int i915_gpu_idle(struct drm_device *dev)
2184 {
2185 	drm_i915_private_t *dev_priv = dev->dev_private;
2186 	struct intel_ring_buffer *ring;
2187 	int ret, i;
2188 
2189 	/* Flush everything onto the inactive list. */
2190 	for_each_ring(ring, dev_priv, i) {
2191 		ret = i915_switch_context(ring, NULL, DEFAULT_CONTEXT_ID);
2192 		if (ret)
2193 			return ret;
2194 
2195 		ret = intel_ring_idle(ring);
2196 		if (ret)
2197 			return ret;
2198 	}
2199 
2200 	return 0;
2201 }
2202 
2203 static void i965_write_fence_reg(struct drm_device *dev, int reg,
2204 				 struct drm_i915_gem_object *obj)
2205 {
2206 	drm_i915_private_t *dev_priv = dev->dev_private;
2207 	int fence_reg;
2208 	int fence_pitch_shift;
2209 	uint64_t val;
2210 
2211 	if (INTEL_INFO(dev)->gen >= 6) {
2212 		fence_reg = FENCE_REG_SANDYBRIDGE_0;
2213 		fence_pitch_shift = SANDYBRIDGE_FENCE_PITCH_SHIFT;
2214 	} else {
2215 		fence_reg = FENCE_REG_965_0;
2216 		fence_pitch_shift = I965_FENCE_PITCH_SHIFT;
2217 	}
2218 
2219 	if (obj) {
2220 		u32 size = obj->gtt_space->size;
2221 
2222 		val = (uint64_t)((obj->gtt_offset + size - 4096) &
2223 				 0xfffff000) << 32;
2224 		val |= obj->gtt_offset & 0xfffff000;
2225 		val |= (uint64_t)((obj->stride / 128) - 1) << fence_pitch_shift;
2226 		if (obj->tiling_mode == I915_TILING_Y)
2227 			val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2228 		val |= I965_FENCE_REG_VALID;
2229 	} else
2230 		val = 0;
2231 
2232 	fence_reg += reg * 8;
2233 	I915_WRITE64(fence_reg, val);
2234 	POSTING_READ(fence_reg);
2235 }
2236 
2237 static void i915_write_fence_reg(struct drm_device *dev, int reg,
2238 				 struct drm_i915_gem_object *obj)
2239 {
2240 	drm_i915_private_t *dev_priv = dev->dev_private;
2241 	u32 val;
2242 
2243 	if (obj) {
2244 		u32 size = obj->gtt_space->size;
2245 		int pitch_val;
2246 		int tile_width;
2247 
2248 		WARN((obj->gtt_offset & ~I915_FENCE_START_MASK) ||
2249 		     (size & -size) != size ||
2250 		     (obj->gtt_offset & (size - 1)),
2251 		     "object 0x%08x [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
2252 		     obj->gtt_offset, obj->map_and_fenceable, size);
2253 
2254 		if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
2255 			tile_width = 128;
2256 		else
2257 			tile_width = 512;
2258 
2259 		/* Note: pitch better be a power of two tile widths */
2260 		pitch_val = obj->stride / tile_width;
2261 		pitch_val = ffs(pitch_val) - 1;
2262 
2263 		val = obj->gtt_offset;
2264 		if (obj->tiling_mode == I915_TILING_Y)
2265 			val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2266 		val |= I915_FENCE_SIZE_BITS(size);
2267 		val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2268 		val |= I830_FENCE_REG_VALID;
2269 	} else
2270 		val = 0;
2271 
2272 	if (reg < 8)
2273 		reg = FENCE_REG_830_0 + reg * 4;
2274 	else
2275 		reg = FENCE_REG_945_8 + (reg - 8) * 4;
2276 
2277 	I915_WRITE(reg, val);
2278 	POSTING_READ(reg);
2279 }
2280 
2281 static void i830_write_fence_reg(struct drm_device *dev, int reg,
2282 				struct drm_i915_gem_object *obj)
2283 {
2284 	drm_i915_private_t *dev_priv = dev->dev_private;
2285 	uint32_t val;
2286 
2287 	if (obj) {
2288 		u32 size = obj->gtt_space->size;
2289 		uint32_t pitch_val;
2290 
2291 		WARN((obj->gtt_offset & ~I830_FENCE_START_MASK) ||
2292 		     (size & -size) != size ||
2293 		     (obj->gtt_offset & (size - 1)),
2294 		     "object 0x%08x not 512K or pot-size 0x%08x aligned\n",
2295 		     obj->gtt_offset, size);
2296 
2297 		pitch_val = obj->stride / 128;
2298 		pitch_val = ffs(pitch_val) - 1;
2299 
2300 		val = obj->gtt_offset;
2301 		if (obj->tiling_mode == I915_TILING_Y)
2302 			val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2303 		val |= I830_FENCE_SIZE_BITS(size);
2304 		val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2305 		val |= I830_FENCE_REG_VALID;
2306 	} else
2307 		val = 0;
2308 
2309 	I915_WRITE(FENCE_REG_830_0 + reg * 4, val);
2310 	POSTING_READ(FENCE_REG_830_0 + reg * 4);
2311 }
2312 
2313 inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object *obj)
2314 {
2315 	return obj && obj->base.read_domains & I915_GEM_DOMAIN_GTT;
2316 }
2317 
2318 static void i915_gem_write_fence(struct drm_device *dev, int reg,
2319 				 struct drm_i915_gem_object *obj)
2320 {
2321 	struct drm_i915_private *dev_priv = dev->dev_private;
2322 
2323 	/* Ensure that all CPU reads are completed before installing a fence
2324 	 * and all writes before removing the fence.
2325 	 */
2326 	if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj))
2327 		cpu_mfence();
2328 
2329 	switch (INTEL_INFO(dev)->gen) {
2330 	case 7:
2331 	case 6:
2332 	case 5:
2333 	case 4: i965_write_fence_reg(dev, reg, obj); break;
2334 	case 3: i915_write_fence_reg(dev, reg, obj); break;
2335 	case 2: i830_write_fence_reg(dev, reg, obj); break;
2336 	default: BUG();
2337 	}
2338 
2339 	/* And similarly be paranoid that no direct access to this region
2340 	 * is reordered to before the fence is installed.
2341 	 */
2342 	if (i915_gem_object_needs_mb(obj))
2343 		cpu_mfence();
2344 }
2345 
2346 static inline int fence_number(struct drm_i915_private *dev_priv,
2347 			       struct drm_i915_fence_reg *fence)
2348 {
2349 	return fence - dev_priv->fence_regs;
2350 }
2351 
2352 static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
2353 					 struct drm_i915_fence_reg *fence,
2354 					 bool enable)
2355 {
2356 	struct drm_device *dev = obj->base.dev;
2357 	struct drm_i915_private *dev_priv = dev->dev_private;
2358 	int fence_reg = fence_number(dev_priv, fence);
2359 
2360 	/* In order to fully serialize access to the fenced region and
2361 	 * the update to the fence register we need to take extreme
2362 	 * measures on SNB+. In theory, the write to the fence register
2363 	 * flushes all memory transactions before, and coupled with the
2364 	 * mb() placed around the register write we serialise all memory
2365 	 * operations with respect to the changes in the tiler. Yet, on
2366 	 * SNB+ we need to take a step further and emit an explicit wbinvd()
2367 	 * on each processor in order to manually flush all memory
2368 	 * transactions before updating the fence register.
2369 	 */
2370 	if (HAS_LLC(obj->base.dev))
2371 		cpu_wbinvd_on_all_cpus();
2372 	i915_gem_write_fence(dev, fence_reg, enable ? obj : NULL);
2373 
2374 	if (enable) {
2375 		obj->fence_reg = fence_reg;
2376 		fence->obj = obj;
2377 		list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list);
2378 	} else {
2379 		obj->fence_reg = I915_FENCE_REG_NONE;
2380 		fence->obj = NULL;
2381 		list_del_init(&fence->lru_list);
2382 	}
2383 }
2384 
2385 static int
2386 i915_gem_object_wait_fence(struct drm_i915_gem_object *obj)
2387 {
2388 	if (obj->last_fenced_seqno) {
2389 		int ret = i915_wait_seqno(obj->ring, obj->last_fenced_seqno);
2390 		if (ret)
2391 			return ret;
2392 
2393 		obj->last_fenced_seqno = 0;
2394 	}
2395 
2396 	obj->fenced_gpu_access = false;
2397 	return 0;
2398 }
2399 
2400 int
2401 i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
2402 {
2403 	struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2404 	int ret;
2405 
2406 	ret = i915_gem_object_wait_fence(obj);
2407 	if (ret)
2408 		return ret;
2409 
2410 	if (obj->fence_reg == I915_FENCE_REG_NONE)
2411 		return 0;
2412 
2413 	i915_gem_object_update_fence(obj,
2414 				     &dev_priv->fence_regs[obj->fence_reg],
2415 				     false);
2416 	i915_gem_object_fence_lost(obj);
2417 
2418 	return 0;
2419 }
2420 
2421 static struct drm_i915_fence_reg *
2422 i915_find_fence_reg(struct drm_device *dev)
2423 {
2424 	struct drm_i915_private *dev_priv = dev->dev_private;
2425 	struct drm_i915_fence_reg *reg, *avail;
2426 	int i;
2427 
2428 	/* First try to find a free reg */
2429 	avail = NULL;
2430 	for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2431 		reg = &dev_priv->fence_regs[i];
2432 		if (!reg->obj)
2433 			return reg;
2434 
2435 		if (!reg->pin_count)
2436 			avail = reg;
2437 	}
2438 
2439 	if (avail == NULL)
2440 		return NULL;
2441 
2442 	/* None available, try to steal one or wait for a user to finish */
2443 	list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
2444 		if (reg->pin_count)
2445 			continue;
2446 
2447 		return reg;
2448 	}
2449 
2450 	return NULL;
2451 }
2452 
2453 /**
2454  * i915_gem_object_get_fence - set up fencing for an object
2455  * @obj: object to map through a fence reg
2456  *
2457  * When mapping objects through the GTT, userspace wants to be able to write
2458  * to them without having to worry about swizzling if the object is tiled.
2459  * This function walks the fence regs looking for a free one for @obj,
2460  * stealing one if it can't find any.
2461  *
2462  * It then sets up the reg based on the object's properties: address, pitch
2463  * and tiling format.
2464  *
2465  * For an untiled surface, this removes any existing fence.
2466  */
2467 int
2468 i915_gem_object_get_fence(struct drm_i915_gem_object *obj)
2469 {
2470 	struct drm_device *dev = obj->base.dev;
2471 	struct drm_i915_private *dev_priv = dev->dev_private;
2472 	bool enable = obj->tiling_mode != I915_TILING_NONE;
2473 	struct drm_i915_fence_reg *reg;
2474 	int ret;
2475 
2476 	/* Have we updated the tiling parameters upon the object and so
2477 	 * will need to serialise the write to the associated fence register?
2478 	 */
2479 	if (obj->fence_dirty) {
2480 		ret = i915_gem_object_wait_fence(obj);
2481 		if (ret)
2482 			return ret;
2483 	}
2484 
2485 	/* Just update our place in the LRU if our fence is getting reused. */
2486 	if (obj->fence_reg != I915_FENCE_REG_NONE) {
2487 		reg = &dev_priv->fence_regs[obj->fence_reg];
2488 		if (!obj->fence_dirty) {
2489 			list_move_tail(&reg->lru_list,
2490 				       &dev_priv->mm.fence_list);
2491 			return 0;
2492 		}
2493 	} else if (enable) {
2494 		reg = i915_find_fence_reg(dev);
2495 		if (reg == NULL)
2496 			return -EDEADLK;
2497 
2498 		if (reg->obj) {
2499 			struct drm_i915_gem_object *old = reg->obj;
2500 
2501 			ret = i915_gem_object_wait_fence(old);
2502 			if (ret)
2503 				return ret;
2504 
2505 			i915_gem_object_fence_lost(old);
2506 		}
2507 	} else
2508 		return 0;
2509 
2510 	i915_gem_object_update_fence(obj, reg, enable);
2511 	obj->fence_dirty = false;
2512 
2513 	return 0;
2514 }
2515 
2516 static bool i915_gem_valid_gtt_space(struct drm_device *dev,
2517 				     struct drm_mm_node *gtt_space,
2518 				     unsigned long cache_level)
2519 {
2520 	struct drm_mm_node *other;
2521 
2522 	/* On non-LLC machines we have to be careful when putting differing
2523 	 * types of snoopable memory together to avoid the prefetcher
2524 	 * crossing memory domains and dying.
2525 	 */
2526 	if (HAS_LLC(dev))
2527 		return true;
2528 
2529 	if (gtt_space == NULL)
2530 		return true;
2531 
2532 	if (list_empty(&gtt_space->node_list))
2533 		return true;
2534 
2535 	other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list);
2536 	if (other->allocated && !other->hole_follows && other->color != cache_level)
2537 		return false;
2538 
2539 	other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list);
2540 	if (other->allocated && !gtt_space->hole_follows && other->color != cache_level)
2541 		return false;
2542 
2543 	return true;
2544 }
2545 
2546 static void i915_gem_verify_gtt(struct drm_device *dev)
2547 {
2548 #if WATCH_GTT
2549 	struct drm_i915_private *dev_priv = dev->dev_private;
2550 	struct drm_i915_gem_object *obj;
2551 	int err = 0;
2552 
2553 	list_for_each_entry(obj, &dev_priv->mm.gtt_list, gtt_list) {
2554 		if (obj->gtt_space == NULL) {
2555 			printk(KERN_ERR "object found on GTT list with no space reserved\n");
2556 			err++;
2557 			continue;
2558 		}
2559 
2560 		if (obj->cache_level != obj->gtt_space->color) {
2561 			printk(KERN_ERR "object reserved space [%08lx, %08lx] with wrong color, cache_level=%x, color=%lx\n",
2562 			       obj->gtt_space->start,
2563 			       obj->gtt_space->start + obj->gtt_space->size,
2564 			       obj->cache_level,
2565 			       obj->gtt_space->color);
2566 			err++;
2567 			continue;
2568 		}
2569 
2570 		if (!i915_gem_valid_gtt_space(dev,
2571 					      obj->gtt_space,
2572 					      obj->cache_level)) {
2573 			printk(KERN_ERR "invalid GTT space found at [%08lx, %08lx] - color=%x\n",
2574 			       obj->gtt_space->start,
2575 			       obj->gtt_space->start + obj->gtt_space->size,
2576 			       obj->cache_level);
2577 			err++;
2578 			continue;
2579 		}
2580 	}
2581 
2582 	WARN_ON(err);
2583 #endif
2584 }
2585 
2586 /**
2587  * Finds free space in the GTT aperture and binds the object there.
2588  */
2589 static int
2590 i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
2591 			    unsigned alignment,
2592 			    bool map_and_fenceable,
2593 			    bool nonblocking)
2594 {
2595 	struct drm_device *dev = obj->base.dev;
2596 	drm_i915_private_t *dev_priv = dev->dev_private;
2597 	struct drm_mm_node *free_space;
2598 	uint32_t size, fence_size, fence_alignment, unfenced_alignment;
2599 	bool mappable, fenceable;
2600 	int ret;
2601 
2602 	fence_size = i915_gem_get_gtt_size(dev,
2603 					   obj->base.size,
2604 					   obj->tiling_mode);
2605 	fence_alignment = i915_gem_get_gtt_alignment(dev,
2606 						     obj->base.size,
2607 						     obj->tiling_mode, true);
2608 	unfenced_alignment =
2609 		i915_gem_get_gtt_alignment(dev,
2610 						    obj->base.size,
2611 						    obj->tiling_mode, false);
2612 
2613 	if (alignment == 0)
2614 		alignment = map_and_fenceable ? fence_alignment :
2615 						unfenced_alignment;
2616 	if (map_and_fenceable && alignment & (fence_alignment - 1)) {
2617 		DRM_ERROR("Invalid object alignment requested %u\n", alignment);
2618 		return -EINVAL;
2619 	}
2620 
2621 	size = map_and_fenceable ? fence_size : obj->base.size;
2622 
2623 	/* If the object is bigger than the entire aperture, reject it early
2624 	 * before evicting everything in a vain attempt to find space.
2625 	 */
2626 	if (obj->base.size >
2627 	    (map_and_fenceable ? dev_priv->gtt.mappable_end : dev_priv->gtt.total)) {
2628 		DRM_ERROR("Attempting to bind an object larger than the aperture\n");
2629 		return -E2BIG;
2630 	}
2631 
2632  search_free:
2633 	if (map_and_fenceable)
2634 		free_space =
2635 			drm_mm_search_free_in_range_color(&dev_priv->mm.gtt_space,
2636 							  size, alignment, obj->cache_level,
2637 							  0, dev_priv->gtt.mappable_end,
2638 							  false);
2639 	else
2640 		free_space = drm_mm_search_free_color(&dev_priv->mm.gtt_space,
2641 						      size, alignment, obj->cache_level,
2642 						      false);
2643 	if (free_space != NULL) {
2644 		if (map_and_fenceable)
2645 			obj->gtt_space =
2646 				drm_mm_get_block_range_generic(free_space,
2647 							       size, alignment, obj->cache_level,
2648 							       0, dev_priv->gtt.mappable_end,
2649 							       false);
2650 		else
2651 			obj->gtt_space =
2652 				drm_mm_get_block_generic(free_space,
2653 							 size, alignment, obj->cache_level,
2654 							 false);
2655 	}
2656 	if (obj->gtt_space == NULL) {
2657 		ret = i915_gem_evict_something(dev, size, alignment,
2658 					       obj->cache_level,
2659 					       map_and_fenceable,
2660 					       nonblocking);
2661 		if (ret)
2662 			return ret;
2663 
2664 		goto search_free;
2665 	}
2666 
2667 	/*
2668 	 * NOTE: i915_gem_object_get_pages_gtt() cannot
2669 	 *	 return ENOMEM, since we used VM_ALLOC_RETRY.
2670 	 */
2671 	ret = i915_gem_object_get_pages_gtt(obj);
2672 	if (ret != 0) {
2673 		drm_mm_put_block(obj->gtt_space);
2674 		obj->gtt_space = NULL;
2675 		return ret;
2676 	}
2677 
2678 	i915_gem_gtt_bind_object(obj, obj->cache_level);
2679 	if (ret != 0) {
2680 		i915_gem_object_put_pages_gtt(obj);
2681 		drm_mm_put_block(obj->gtt_space);
2682 		obj->gtt_space = NULL;
2683 		if (i915_gem_evict_everything(dev))
2684 			return (ret);
2685 		goto search_free;
2686 	}
2687 
2688 	list_add_tail(&obj->gtt_list, &dev_priv->mm.bound_list);
2689 	list_add_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
2690 
2691 	obj->gtt_offset = obj->gtt_space->start;
2692 
2693 	fenceable =
2694 		obj->gtt_space->size == fence_size &&
2695 		(obj->gtt_space->start & (fence_alignment - 1)) == 0;
2696 
2697 	mappable =
2698 		obj->gtt_offset + obj->base.size <= dev_priv->gtt.mappable_end;
2699 
2700 	obj->map_and_fenceable = mappable && fenceable;
2701 
2702 	i915_gem_verify_gtt(dev);
2703 	return 0;
2704 }
2705 
2706 void
2707 i915_gem_clflush_object(struct drm_i915_gem_object *obj)
2708 {
2709 
2710 	/* If we don't have a page list set up, then we're not pinned
2711 	 * to GPU, and we can ignore the cache flush because it'll happen
2712 	 * again at bind time.
2713 	 */
2714 	if (obj->pages == NULL)
2715 		return;
2716 
2717 	/*
2718 	 * Stolen memory is always coherent with the GPU as it is explicitly
2719 	 * marked as wc by the system, or the system is cache-coherent.
2720 	 */
2721 	if (obj->stolen)
2722 		return;
2723 
2724 	/* If the GPU is snooping the contents of the CPU cache,
2725 	 * we do not need to manually clear the CPU cache lines.  However,
2726 	 * the caches are only snooped when the render cache is
2727 	 * flushed/invalidated.  As we always have to emit invalidations
2728 	 * and flushes when moving into and out of the RENDER domain, correct
2729 	 * snooping behaviour occurs naturally as the result of our domain
2730 	 * tracking.
2731 	 */
2732 	if (obj->cache_level != I915_CACHE_NONE)
2733 		return;
2734 
2735 	drm_clflush_pages(obj->pages, obj->base.size / PAGE_SIZE);
2736 }
2737 
2738 /** Flushes the GTT write domain for the object if it's dirty. */
2739 static void
2740 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
2741 {
2742 	uint32_t old_write_domain;
2743 
2744 	if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
2745 		return;
2746 
2747 	/* No actual flushing is required for the GTT write domain.  Writes
2748 	 * to it immediately go to main memory as far as we know, so there's
2749 	 * no chipset flush.  It also doesn't land in render cache.
2750 	 *
2751 	 * However, we do have to enforce the order so that all writes through
2752 	 * the GTT land before any writes to the device, such as updates to
2753 	 * the GATT itself.
2754 	 */
2755 	cpu_sfence();
2756 
2757 	old_write_domain = obj->base.write_domain;
2758 	obj->base.write_domain = 0;
2759 }
2760 
2761 /** Flushes the CPU write domain for the object if it's dirty. */
2762 static void
2763 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj)
2764 {
2765 	uint32_t old_write_domain;
2766 
2767 	if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
2768 		return;
2769 
2770 	i915_gem_clflush_object(obj);
2771 	i915_gem_chipset_flush(obj->base.dev);
2772 	old_write_domain = obj->base.write_domain;
2773 	obj->base.write_domain = 0;
2774 }
2775 
2776 /**
2777  * Moves a single object to the GTT read, and possibly write domain.
2778  *
2779  * This function returns when the move is complete, including waiting on
2780  * flushes to occur.
2781  */
2782 int
2783 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
2784 {
2785 	drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
2786 	uint32_t old_write_domain, old_read_domains;
2787 	int ret;
2788 
2789 	/* Not valid to be called on unbound objects. */
2790 	if (obj->gtt_space == NULL)
2791 		return -EINVAL;
2792 
2793 	if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
2794 		return 0;
2795 
2796 	ret = i915_gem_object_wait_rendering(obj, !write);
2797 	if (ret)
2798 		return ret;
2799 
2800 	i915_gem_object_flush_cpu_write_domain(obj);
2801 
2802 	/* Serialise direct access to this object with the barriers for
2803 	 * coherent writes from the GPU, by effectively invalidating the
2804 	 * GTT domain upon first access.
2805 	 */
2806 	if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
2807 		cpu_mfence();
2808 
2809 	old_write_domain = obj->base.write_domain;
2810 	old_read_domains = obj->base.read_domains;
2811 
2812 	/* It should now be out of any other write domains, and we can update
2813 	 * the domain values for our changes.
2814 	 */
2815 	BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2816 	obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
2817 	if (write) {
2818 		obj->base.read_domains = I915_GEM_DOMAIN_GTT;
2819 		obj->base.write_domain = I915_GEM_DOMAIN_GTT;
2820 		obj->dirty = 1;
2821 	}
2822 
2823 	/* And bump the LRU for this access */
2824 	if (i915_gem_object_is_inactive(obj))
2825 		list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
2826 
2827 	return 0;
2828 }
2829 
2830 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
2831 				    enum i915_cache_level cache_level)
2832 {
2833 	struct drm_device *dev = obj->base.dev;
2834 	drm_i915_private_t *dev_priv = dev->dev_private;
2835 	int ret;
2836 
2837 	if (obj->cache_level == cache_level)
2838 		return 0;
2839 
2840 	if (obj->pin_count) {
2841 		DRM_DEBUG("can not change the cache level of pinned objects\n");
2842 		return -EBUSY;
2843 	}
2844 
2845 	if (!i915_gem_valid_gtt_space(dev, obj->gtt_space, cache_level)) {
2846 		ret = i915_gem_object_unbind(obj);
2847 		if (ret)
2848 			return ret;
2849 	}
2850 
2851 	if (obj->gtt_space) {
2852 		ret = i915_gem_object_finish_gpu(obj);
2853 		if (ret)
2854 			return ret;
2855 
2856 		i915_gem_object_finish_gtt(obj);
2857 
2858 		/* Before SandyBridge, you could not use tiling or fence
2859 		 * registers with snooped memory, so relinquish any fences
2860 		 * currently pointing to our region in the aperture.
2861 		 */
2862 		if (INTEL_INFO(dev)->gen < 6) {
2863 			ret = i915_gem_object_put_fence(obj);
2864 			if (ret)
2865 				return ret;
2866 		}
2867 
2868 		if (obj->has_global_gtt_mapping)
2869 			i915_gem_gtt_bind_object(obj, cache_level);
2870 		if (obj->has_aliasing_ppgtt_mapping)
2871 			i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt,
2872 					       obj, cache_level);
2873 
2874 		obj->gtt_space->color = cache_level;
2875 	}
2876 
2877 	if (cache_level == I915_CACHE_NONE) {
2878 		u32 old_read_domains, old_write_domain;
2879 
2880 		/* If we're coming from LLC cached, then we haven't
2881 		 * actually been tracking whether the data is in the
2882 		 * CPU cache or not, since we only allow one bit set
2883 		 * in obj->write_domain and have been skipping the clflushes.
2884 		 * Just set it to the CPU cache for now.
2885 		 */
2886 		WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU);
2887 		WARN_ON(obj->base.read_domains & ~I915_GEM_DOMAIN_CPU);
2888 
2889 		old_read_domains = obj->base.read_domains;
2890 		old_write_domain = obj->base.write_domain;
2891 
2892 		obj->base.read_domains = I915_GEM_DOMAIN_CPU;
2893 		obj->base.write_domain = I915_GEM_DOMAIN_CPU;
2894 
2895 	}
2896 
2897 	obj->cache_level = cache_level;
2898 	i915_gem_verify_gtt(dev);
2899 	return 0;
2900 }
2901 
2902 int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data,
2903 			       struct drm_file *file)
2904 {
2905 	struct drm_i915_gem_caching *args = data;
2906 	struct drm_i915_gem_object *obj;
2907 	int ret;
2908 
2909 	ret = i915_mutex_lock_interruptible(dev);
2910 	if (ret)
2911 		return ret;
2912 
2913 	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
2914 	if (&obj->base == NULL) {
2915 		ret = -ENOENT;
2916 		goto unlock;
2917 	}
2918 
2919 	args->caching = obj->cache_level != I915_CACHE_NONE;
2920 
2921 	drm_gem_object_unreference(&obj->base);
2922 unlock:
2923 	mutex_unlock(&dev->struct_mutex);
2924 	return ret;
2925 }
2926 
2927 int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
2928 			       struct drm_file *file)
2929 {
2930 	struct drm_i915_gem_caching *args = data;
2931 	struct drm_i915_gem_object *obj;
2932 	enum i915_cache_level level;
2933 	int ret;
2934 
2935 	switch (args->caching) {
2936 	case I915_CACHING_NONE:
2937 		level = I915_CACHE_NONE;
2938 		break;
2939 	case I915_CACHING_CACHED:
2940 		level = I915_CACHE_LLC;
2941 		break;
2942 	default:
2943 		return -EINVAL;
2944 	}
2945 
2946 	ret = i915_mutex_lock_interruptible(dev);
2947 	if (ret)
2948 		return ret;
2949 
2950 	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
2951 	if (&obj->base == NULL) {
2952 		ret = -ENOENT;
2953 		goto unlock;
2954 	}
2955 
2956 	ret = i915_gem_object_set_cache_level(obj, level);
2957 
2958 	drm_gem_object_unreference(&obj->base);
2959 unlock:
2960 	mutex_unlock(&dev->struct_mutex);
2961 	return ret;
2962 }
2963 
2964 /*
2965  * Prepare buffer for display plane (scanout, cursors, etc).
2966  * Can be called from an uninterruptible phase (modesetting) and allows
2967  * any flushes to be pipelined (for pageflips).
2968  */
2969 int
2970 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
2971 				     u32 alignment,
2972 				     struct intel_ring_buffer *pipelined)
2973 {
2974 	u32 old_read_domains, old_write_domain;
2975 	int ret;
2976 
2977 	if (pipelined != obj->ring) {
2978 		ret = i915_gem_object_sync(obj, pipelined);
2979 		if (ret)
2980 			return ret;
2981 	}
2982 
2983 	/* The display engine is not coherent with the LLC cache on gen6.  As
2984 	 * a result, we make sure that the pinning that is about to occur is
2985 	 * done with uncached PTEs. This is lowest common denominator for all
2986 	 * chipsets.
2987 	 *
2988 	 * However for gen6+, we could do better by using the GFDT bit instead
2989 	 * of uncaching, which would allow us to flush all the LLC-cached data
2990 	 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
2991 	 */
2992 	ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE);
2993 	if (ret)
2994 		return ret;
2995 
2996 	/* As the user may map the buffer once pinned in the display plane
2997 	 * (e.g. libkms for the bootup splash), we have to ensure that we
2998 	 * always use map_and_fenceable for all scanout buffers.
2999 	 */
3000 	ret = i915_gem_object_pin(obj, alignment, true, false);
3001 	if (ret)
3002 		return ret;
3003 
3004 	i915_gem_object_flush_cpu_write_domain(obj);
3005 
3006 	old_write_domain = obj->base.write_domain;
3007 	old_read_domains = obj->base.read_domains;
3008 
3009 	/* It should now be out of any other write domains, and we can update
3010 	 * the domain values for our changes.
3011 	 */
3012 	obj->base.write_domain = 0;
3013 	obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3014 
3015 	return 0;
3016 }
3017 
3018 int
3019 i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
3020 {
3021 	int ret;
3022 
3023 	if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
3024 		return 0;
3025 
3026 	ret = i915_gem_object_wait_rendering(obj, false);
3027 	if (ret)
3028 		return ret;
3029 
3030 	/* Ensure that we invalidate the GPU's caches and TLBs. */
3031 	obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
3032 	return 0;
3033 }
3034 
3035 /**
3036  * Moves a single object to the CPU read, and possibly write domain.
3037  *
3038  * This function returns when the move is complete, including waiting on
3039  * flushes to occur.
3040  */
3041 int
3042 i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
3043 {
3044 	uint32_t old_write_domain, old_read_domains;
3045 	int ret;
3046 
3047 	if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
3048 		return 0;
3049 
3050 	ret = i915_gem_object_wait_rendering(obj, !write);
3051 	if (ret)
3052 		return ret;
3053 
3054 	i915_gem_object_flush_gtt_write_domain(obj);
3055 
3056 	old_write_domain = obj->base.write_domain;
3057 	old_read_domains = obj->base.read_domains;
3058 
3059 	/* Flush the CPU cache if it's still invalid. */
3060 	if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
3061 		i915_gem_clflush_object(obj);
3062 
3063 		obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3064 	}
3065 
3066 	/* It should now be out of any other write domains, and we can update
3067 	 * the domain values for our changes.
3068 	 */
3069 	BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3070 
3071 	/* If we're writing through the CPU, then the GPU read domains will
3072 	 * need to be invalidated at next use.
3073 	 */
3074 	if (write) {
3075 		obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3076 		obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3077 	}
3078 
3079 	return 0;
3080 }
3081 
3082 /* Throttle our rendering by waiting until the ring has completed our requests
3083  * emitted over 20 msec ago.
3084  *
3085  * Note that if we were to use the current jiffies each time around the loop,
3086  * we wouldn't escape the function with any frames outstanding if the time to
3087  * render a frame was over 20ms.
3088  *
3089  * This should get us reasonable parallelism between CPU and GPU but also
3090  * relatively low latency when blocking on a particular request to finish.
3091  */
3092 static int
3093 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
3094 {
3095 	struct drm_i915_private *dev_priv = dev->dev_private;
3096 	struct drm_i915_file_private *file_priv = file->driver_priv;
3097 	unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3098 	struct drm_i915_gem_request *request;
3099 	struct intel_ring_buffer *ring = NULL;
3100 	unsigned reset_counter;
3101 	u32 seqno = 0;
3102 	int ret;
3103 
3104 	ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
3105 	if (ret)
3106 		return ret;
3107 
3108 	ret = i915_gem_check_wedge(&dev_priv->gpu_error, false);
3109 	if (ret)
3110 		return ret;
3111 
3112 	spin_lock(&file_priv->mm.lock);
3113 	list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
3114 		if (time_after_eq(request->emitted_jiffies, recent_enough))
3115 			break;
3116 
3117 		ring = request->ring;
3118 		seqno = request->seqno;
3119 	}
3120 	reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
3121 	spin_unlock(&file_priv->mm.lock);
3122 
3123 	if (seqno == 0)
3124 		return 0;
3125 
3126 	ret = __wait_seqno(ring, seqno, reset_counter, true, NULL);
3127 	if (ret == 0)
3128 		queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
3129 
3130 	return ret;
3131 }
3132 
3133 int
3134 i915_gem_object_pin(struct drm_i915_gem_object *obj,
3135 		    uint32_t alignment,
3136 		    bool map_and_fenceable,
3137 		    bool nonblocking)
3138 {
3139 	int ret;
3140 
3141 	if (WARN_ON(obj->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT))
3142 		return -EBUSY;
3143 
3144 	if (obj->gtt_space != NULL) {
3145 		if ((alignment && obj->gtt_offset & (alignment - 1)) ||
3146 		    (map_and_fenceable && !obj->map_and_fenceable)) {
3147 			WARN(obj->pin_count,
3148 			     "bo is already pinned with incorrect alignment:"
3149 			     " offset=%x, req.alignment=%x, req.map_and_fenceable=%d,"
3150 			     " obj->map_and_fenceable=%d\n",
3151 			     obj->gtt_offset, alignment,
3152 			     map_and_fenceable,
3153 			     obj->map_and_fenceable);
3154 			ret = i915_gem_object_unbind(obj);
3155 			if (ret)
3156 				return ret;
3157 		}
3158 	}
3159 
3160 	if (obj->gtt_space == NULL) {
3161 		struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3162 
3163 		ret = i915_gem_object_bind_to_gtt(obj, alignment,
3164 						  map_and_fenceable,
3165 						  nonblocking);
3166 		if (ret)
3167 			return ret;
3168 
3169 		if (!dev_priv->mm.aliasing_ppgtt)
3170 			i915_gem_gtt_bind_object(obj, obj->cache_level);
3171 	}
3172 
3173 	if (!obj->has_global_gtt_mapping && map_and_fenceable)
3174 		i915_gem_gtt_bind_object(obj, obj->cache_level);
3175 
3176 	obj->pin_count++;
3177 	obj->pin_mappable |= map_and_fenceable;
3178 
3179 	return 0;
3180 }
3181 
3182 void
3183 i915_gem_object_unpin(struct drm_i915_gem_object *obj)
3184 {
3185 	BUG_ON(obj->pin_count == 0);
3186 	BUG_ON(obj->gtt_space == NULL);
3187 
3188 	if (--obj->pin_count == 0)
3189 		obj->pin_mappable = false;
3190 }
3191 
3192 int
3193 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
3194 		   struct drm_file *file)
3195 {
3196 	struct drm_i915_gem_pin *args = data;
3197 	struct drm_i915_gem_object *obj;
3198 	int ret;
3199 
3200 	ret = i915_mutex_lock_interruptible(dev);
3201 	if (ret)
3202 		return ret;
3203 
3204 	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3205 	if (&obj->base == NULL) {
3206 		ret = -ENOENT;
3207 		goto unlock;
3208 	}
3209 
3210 	if (obj->madv != I915_MADV_WILLNEED) {
3211 		DRM_ERROR("Attempting to pin a purgeable buffer\n");
3212 		ret = -EINVAL;
3213 		goto out;
3214 	}
3215 
3216 	if (obj->pin_filp != NULL && obj->pin_filp != file) {
3217 		DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
3218 			  args->handle);
3219 		ret = -EINVAL;
3220 		goto out;
3221 	}
3222 
3223 	if (obj->user_pin_count == 0) {
3224 		ret = i915_gem_object_pin(obj, args->alignment, true, false);
3225 		if (ret)
3226 			goto out;
3227 	}
3228 
3229 	obj->user_pin_count++;
3230 	obj->pin_filp = file;
3231 
3232 	/* XXX - flush the CPU caches for pinned objects
3233 	 * as the X server doesn't manage domains yet
3234 	 */
3235 	i915_gem_object_flush_cpu_write_domain(obj);
3236 	args->offset = obj->gtt_offset;
3237 out:
3238 	drm_gem_object_unreference(&obj->base);
3239 unlock:
3240 	mutex_unlock(&dev->struct_mutex);
3241 	return ret;
3242 }
3243 
3244 int
3245 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
3246 		     struct drm_file *file)
3247 {
3248 	struct drm_i915_gem_pin *args = data;
3249 	struct drm_i915_gem_object *obj;
3250 	int ret;
3251 
3252 	ret = i915_mutex_lock_interruptible(dev);
3253 	if (ret)
3254 		return ret;
3255 
3256 	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3257 	if (&obj->base == NULL) {
3258 		ret = -ENOENT;
3259 		goto unlock;
3260 	}
3261 
3262 	if (obj->pin_filp != file) {
3263 		DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
3264 			  args->handle);
3265 		ret = -EINVAL;
3266 		goto out;
3267 	}
3268 	obj->user_pin_count--;
3269 	if (obj->user_pin_count == 0) {
3270 		obj->pin_filp = NULL;
3271 		i915_gem_object_unpin(obj);
3272 	}
3273 
3274 out:
3275 	drm_gem_object_unreference(&obj->base);
3276 unlock:
3277 	mutex_unlock(&dev->struct_mutex);
3278 	return ret;
3279 }
3280 
3281 int
3282 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
3283 		    struct drm_file *file)
3284 {
3285 	struct drm_i915_gem_busy *args = data;
3286 	struct drm_i915_gem_object *obj;
3287 	int ret;
3288 
3289 	ret = i915_mutex_lock_interruptible(dev);
3290 	if (ret)
3291 		return ret;
3292 
3293 	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3294 	if (&obj->base == NULL) {
3295 		ret = -ENOENT;
3296 		goto unlock;
3297 	}
3298 
3299 	/* Count all active objects as busy, even if they are currently not used
3300 	 * by the gpu. Users of this interface expect objects to eventually
3301 	 * become non-busy without any further actions, therefore emit any
3302 	 * necessary flushes here.
3303 	 */
3304 	ret = i915_gem_object_flush_active(obj);
3305 
3306 	args->busy = obj->active;
3307 	if (obj->ring) {
3308 		args->busy |= intel_ring_flag(obj->ring) << 16;
3309 	}
3310 
3311 	drm_gem_object_unreference(&obj->base);
3312 unlock:
3313 	mutex_unlock(&dev->struct_mutex);
3314 	return ret;
3315 }
3316 
3317 int
3318 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
3319 			struct drm_file *file_priv)
3320 {
3321 	return i915_gem_ring_throttle(dev, file_priv);
3322 }
3323 
3324 int
3325 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
3326 		       struct drm_file *file_priv)
3327 {
3328 	struct drm_i915_gem_madvise *args = data;
3329 	struct drm_i915_gem_object *obj;
3330 	int ret;
3331 
3332 	switch (args->madv) {
3333 	case I915_MADV_DONTNEED:
3334 	case I915_MADV_WILLNEED:
3335 	    break;
3336 	default:
3337 	    return -EINVAL;
3338 	}
3339 
3340 	ret = i915_mutex_lock_interruptible(dev);
3341 	if (ret)
3342 		return ret;
3343 
3344 	obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
3345 	if (&obj->base == NULL) {
3346 		ret = -ENOENT;
3347 		goto unlock;
3348 	}
3349 
3350 	if (obj->pin_count) {
3351 		ret = -EINVAL;
3352 		goto out;
3353 	}
3354 
3355 	if (obj->madv != __I915_MADV_PURGED)
3356 		obj->madv = args->madv;
3357 
3358 	/* if the object is no longer attached, discard its backing storage */
3359 	if (i915_gem_object_is_purgeable(obj) && obj->pages == NULL)
3360 		i915_gem_object_truncate(obj);
3361 
3362 	args->retained = obj->madv != __I915_MADV_PURGED;
3363 
3364 out:
3365 	drm_gem_object_unreference(&obj->base);
3366 unlock:
3367 	mutex_unlock(&dev->struct_mutex);
3368 	return ret;
3369 }
3370 
3371 void i915_gem_object_init(struct drm_i915_gem_object *obj,
3372 			  const struct drm_i915_gem_object_ops *ops)
3373 {
3374 	INIT_LIST_HEAD(&obj->mm_list);
3375 	INIT_LIST_HEAD(&obj->gtt_list);
3376 	INIT_LIST_HEAD(&obj->ring_list);
3377 	INIT_LIST_HEAD(&obj->exec_list);
3378 
3379 	obj->ops = ops;
3380 
3381 	obj->fence_reg = I915_FENCE_REG_NONE;
3382 	obj->madv = I915_MADV_WILLNEED;
3383 	/* Avoid an unnecessary call to unbind on the first bind. */
3384 	obj->map_and_fenceable = true;
3385 
3386 	i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size);
3387 }
3388 
3389 static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
3390 	.get_pages = i915_gem_object_get_pages_gtt,
3391 	.put_pages = i915_gem_object_put_pages_gtt,
3392 };
3393 
3394 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
3395 						  size_t size)
3396 {
3397 	struct drm_i915_gem_object *obj;
3398 #if 0
3399 	struct address_space *mapping;
3400 	u32 mask;
3401 #endif
3402 
3403 	obj = kmalloc(sizeof(*obj), M_DRM, M_WAITOK | M_ZERO);
3404 	if (obj == NULL)
3405 		return NULL;
3406 
3407 	if (drm_gem_object_init(dev, &obj->base, size) != 0) {
3408 		kfree(obj, M_DRM);
3409 		return NULL;
3410 	}
3411 
3412 #if 0
3413 	mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
3414 	if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) {
3415 		/* 965gm cannot relocate objects above 4GiB. */
3416 		mask &= ~__GFP_HIGHMEM;
3417 		mask |= __GFP_DMA32;
3418 	}
3419 
3420 	mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
3421 	mapping_set_gfp_mask(mapping, mask);
3422 #endif
3423 
3424 	i915_gem_object_init(obj, &i915_gem_object_ops);
3425 
3426 	obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3427 	obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3428 
3429 	if (HAS_LLC(dev)) {
3430 		/* On some devices, we can have the GPU use the LLC (the CPU
3431 		 * cache) for about a 10% performance improvement
3432 		 * compared to uncached.  Graphics requests other than
3433 		 * display scanout are coherent with the CPU in
3434 		 * accessing this cache.  This means in this mode we
3435 		 * don't need to clflush on the CPU side, and on the
3436 		 * GPU side we only need to flush internal caches to
3437 		 * get data visible to the CPU.
3438 		 *
3439 		 * However, we maintain the display planes as UC, and so
3440 		 * need to rebind when first used as such.
3441 		 */
3442 		obj->cache_level = I915_CACHE_LLC;
3443 	} else
3444 		obj->cache_level = I915_CACHE_NONE;
3445 
3446 	return obj;
3447 }
3448 
3449 int i915_gem_init_object(struct drm_gem_object *obj)
3450 {
3451 	BUG();
3452 
3453 	return 0;
3454 }
3455 
3456 void i915_gem_free_object(struct drm_gem_object *gem_obj)
3457 {
3458 	struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
3459 	struct drm_device *dev = obj->base.dev;
3460 	drm_i915_private_t *dev_priv = dev->dev_private;
3461 
3462 	if (obj->phys_obj)
3463 		i915_gem_detach_phys_object(dev, obj);
3464 
3465 	obj->pin_count = 0;
3466 	if (WARN_ON(i915_gem_object_unbind(obj) == -ERESTARTSYS)) {
3467 		bool was_interruptible;
3468 
3469 		was_interruptible = dev_priv->mm.interruptible;
3470 		dev_priv->mm.interruptible = false;
3471 
3472 		WARN_ON(i915_gem_object_unbind(obj));
3473 
3474 		dev_priv->mm.interruptible = was_interruptible;
3475 	}
3476 
3477 	drm_gem_free_mmap_offset(&obj->base);
3478 
3479 	drm_gem_object_release(&obj->base);
3480 	i915_gem_info_remove_obj(dev_priv, obj->base.size);
3481 
3482 	drm_free(obj->bit_17, M_DRM);
3483 	drm_free(obj, M_DRM);
3484 }
3485 
3486 int
3487 i915_gem_idle(struct drm_device *dev)
3488 {
3489 	drm_i915_private_t *dev_priv = dev->dev_private;
3490 	int ret;
3491 
3492 	mutex_lock(&dev->struct_mutex);
3493 
3494 	if (dev_priv->mm.suspended) {
3495 		mutex_unlock(&dev->struct_mutex);
3496 		return 0;
3497 	}
3498 
3499 	ret = i915_gpu_idle(dev);
3500 	if (ret) {
3501 		mutex_unlock(&dev->struct_mutex);
3502 		return ret;
3503 	}
3504 	i915_gem_retire_requests(dev);
3505 
3506 	/* Under UMS, be paranoid and evict. */
3507 	if (!drm_core_check_feature(dev, DRIVER_MODESET))
3508 		i915_gem_evict_everything(dev);
3509 
3510 	i915_gem_reset_fences(dev);
3511 
3512 	/* Hack!  Don't let anybody do execbuf while we don't control the chip.
3513 	 * We need to replace this with a semaphore, or something.
3514 	 * And not confound mm.suspended!
3515 	 */
3516 	dev_priv->mm.suspended = 1;
3517 	del_timer_sync(&dev_priv->gpu_error.hangcheck_timer);
3518 
3519 	i915_kernel_lost_context(dev);
3520 	i915_gem_cleanup_ringbuffer(dev);
3521 
3522 	mutex_unlock(&dev->struct_mutex);
3523 
3524 	/* Cancel the retire work handler, which should be idle now. */
3525 	cancel_delayed_work_sync(&dev_priv->mm.retire_work);
3526 
3527 	return 0;
3528 }
3529 
3530 void i915_gem_l3_remap(struct drm_device *dev)
3531 {
3532 	drm_i915_private_t *dev_priv = dev->dev_private;
3533 	u32 misccpctl;
3534 	int i;
3535 
3536 	if (!HAS_L3_GPU_CACHE(dev))
3537 		return;
3538 
3539 	if (!dev_priv->l3_parity.remap_info)
3540 		return;
3541 
3542 	misccpctl = I915_READ(GEN7_MISCCPCTL);
3543 	I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
3544 	POSTING_READ(GEN7_MISCCPCTL);
3545 
3546 	for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) {
3547 		u32 remap = I915_READ(GEN7_L3LOG_BASE + i);
3548 		if (remap && remap != dev_priv->l3_parity.remap_info[i/4])
3549 			DRM_DEBUG("0x%x was already programmed to %x\n",
3550 				  GEN7_L3LOG_BASE + i, remap);
3551 		if (remap && !dev_priv->l3_parity.remap_info[i/4])
3552 			DRM_DEBUG_DRIVER("Clearing remapped register\n");
3553 		I915_WRITE(GEN7_L3LOG_BASE + i, dev_priv->l3_parity.remap_info[i/4]);
3554 	}
3555 
3556 	/* Make sure all the writes land before disabling dop clock gating */
3557 	POSTING_READ(GEN7_L3LOG_BASE);
3558 
3559 	I915_WRITE(GEN7_MISCCPCTL, misccpctl);
3560 }
3561 
3562 void i915_gem_init_swizzling(struct drm_device *dev)
3563 {
3564 	drm_i915_private_t *dev_priv = dev->dev_private;
3565 
3566 	if (INTEL_INFO(dev)->gen < 5 ||
3567 	    dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
3568 		return;
3569 
3570 	I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
3571 				 DISP_TILE_SURFACE_SWIZZLING);
3572 
3573 	if (IS_GEN5(dev))
3574 		return;
3575 
3576 	I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
3577 	if (IS_GEN6(dev))
3578 		I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
3579 	else if (IS_GEN7(dev))
3580 		I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
3581 	else
3582 		BUG();
3583 }
3584 
3585 static bool
3586 intel_enable_blt(struct drm_device *dev)
3587 {
3588 	int revision;
3589 
3590 	if (!HAS_BLT(dev))
3591 		return false;
3592 
3593 	/* The blitter was dysfunctional on early prototypes */
3594 	revision = pci_read_config(dev->dev, PCIR_REVID, 1);
3595 	if (IS_GEN6(dev) && revision < 8) {
3596 		DRM_INFO("BLT not supported on this pre-production hardware;"
3597 			 " graphics performance will be degraded.\n");
3598 		return false;
3599 	}
3600 
3601 	return true;
3602 }
3603 
3604 static int i915_gem_init_rings(struct drm_device *dev)
3605 {
3606 	struct drm_i915_private *dev_priv = dev->dev_private;
3607 	int ret;
3608 
3609 	ret = intel_init_render_ring_buffer(dev);
3610 	if (ret)
3611 		return ret;
3612 
3613 	if (HAS_BSD(dev)) {
3614 		ret = intel_init_bsd_ring_buffer(dev);
3615 		if (ret)
3616 			goto cleanup_render_ring;
3617 	}
3618 
3619 	if (intel_enable_blt(dev)) {
3620 		ret = intel_init_blt_ring_buffer(dev);
3621 		if (ret)
3622 			goto cleanup_bsd_ring;
3623 	}
3624 
3625 	ret = i915_gem_set_seqno(dev, ((u32)~0 - 0x1000));
3626 	if (ret)
3627 		goto cleanup_blt_ring;
3628 
3629 	return 0;
3630 
3631 cleanup_blt_ring:
3632 	intel_cleanup_ring_buffer(&dev_priv->ring[BCS]);
3633 cleanup_bsd_ring:
3634 	intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
3635 cleanup_render_ring:
3636 	intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
3637 
3638 	return ret;
3639 }
3640 
3641 int
3642 i915_gem_init_hw(struct drm_device *dev)
3643 {
3644 	drm_i915_private_t *dev_priv = dev->dev_private;
3645 	int ret;
3646 
3647 #if 0
3648 	if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt())
3649 		return -EIO;
3650 #endif
3651 
3652 	if (IS_HASWELL(dev) && (I915_READ(0x120010) == 1))
3653 		I915_WRITE(0x9008, I915_READ(0x9008) | 0xf0000);
3654 
3655 	i915_gem_l3_remap(dev);
3656 
3657 	i915_gem_init_swizzling(dev);
3658 
3659 	ret = i915_gem_init_rings(dev);
3660 	if (ret)
3661 		return ret;
3662 
3663 	/*
3664 	 * XXX: There was some w/a described somewhere suggesting loading
3665 	 * contexts before PPGTT.
3666 	 */
3667 	i915_gem_context_init(dev);
3668 	i915_gem_init_ppgtt(dev);
3669 
3670 	return 0;
3671 }
3672 
3673 int i915_gem_init(struct drm_device *dev)
3674 {
3675 	struct drm_i915_private *dev_priv = dev->dev_private;
3676 	int ret;
3677 
3678 	mutex_lock(&dev->struct_mutex);
3679 	i915_gem_init_global_gtt(dev);
3680 	ret = i915_gem_init_hw(dev);
3681 	mutex_unlock(&dev->struct_mutex);
3682 	if (ret) {
3683 		i915_gem_cleanup_aliasing_ppgtt(dev);
3684 		return ret;
3685 	}
3686 
3687 	/* Allow hardware batchbuffers unless told otherwise, but not for KMS. */
3688 	if (!drm_core_check_feature(dev, DRIVER_MODESET))
3689 		dev_priv->dri1.allow_batchbuffer = 1;
3690 	return 0;
3691 }
3692 
3693 void
3694 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
3695 {
3696 	drm_i915_private_t *dev_priv = dev->dev_private;
3697 	struct intel_ring_buffer *ring;
3698 	int i;
3699 
3700 	for_each_ring(ring, dev_priv, i)
3701 		intel_cleanup_ring_buffer(ring);
3702 }
3703 
3704 int
3705 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
3706 		       struct drm_file *file_priv)
3707 {
3708 	drm_i915_private_t *dev_priv = dev->dev_private;
3709 	int ret;
3710 
3711 	if (drm_core_check_feature(dev, DRIVER_MODESET))
3712 		return 0;
3713 
3714 	if (i915_reset_in_progress(&dev_priv->gpu_error)) {
3715 		DRM_ERROR("Reenabling wedged hardware, good luck\n");
3716 		atomic_set(&dev_priv->gpu_error.reset_counter, 0);
3717 	}
3718 
3719 	mutex_lock(&dev->struct_mutex);
3720 	dev_priv->mm.suspended = 0;
3721 
3722 	ret = i915_gem_init_hw(dev);
3723 	if (ret != 0) {
3724 		mutex_unlock(&dev->struct_mutex);
3725 		return ret;
3726 	}
3727 
3728 	KASSERT(list_empty(&dev_priv->mm.active_list), ("active list"));
3729 	mutex_unlock(&dev->struct_mutex);
3730 
3731 	ret = drm_irq_install(dev);
3732 	if (ret)
3733 		goto cleanup_ringbuffer;
3734 
3735 	return 0;
3736 
3737 cleanup_ringbuffer:
3738 	mutex_lock(&dev->struct_mutex);
3739 	i915_gem_cleanup_ringbuffer(dev);
3740 	dev_priv->mm.suspended = 1;
3741 	mutex_unlock(&dev->struct_mutex);
3742 
3743 	return ret;
3744 }
3745 
3746 int
3747 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
3748 		       struct drm_file *file_priv)
3749 {
3750 	if (drm_core_check_feature(dev, DRIVER_MODESET))
3751 		return 0;
3752 
3753 	drm_irq_uninstall(dev);
3754 	return i915_gem_idle(dev);
3755 }
3756 
3757 void
3758 i915_gem_lastclose(struct drm_device *dev)
3759 {
3760 	int ret;
3761 
3762 	if (drm_core_check_feature(dev, DRIVER_MODESET))
3763 		return;
3764 
3765 	ret = i915_gem_idle(dev);
3766 	if (ret)
3767 		DRM_ERROR("failed to idle hardware: %d\n", ret);
3768 }
3769 
3770 static void
3771 init_ring_lists(struct intel_ring_buffer *ring)
3772 {
3773 	INIT_LIST_HEAD(&ring->active_list);
3774 	INIT_LIST_HEAD(&ring->request_list);
3775 }
3776 
3777 void
3778 i915_gem_load(struct drm_device *dev)
3779 {
3780 	int i;
3781 	drm_i915_private_t *dev_priv = dev->dev_private;
3782 
3783 	INIT_LIST_HEAD(&dev_priv->mm.active_list);
3784 	INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
3785 	INIT_LIST_HEAD(&dev_priv->mm.unbound_list);
3786 	INIT_LIST_HEAD(&dev_priv->mm.bound_list);
3787 	INIT_LIST_HEAD(&dev_priv->mm.fence_list);
3788 	for (i = 0; i < I915_NUM_RINGS; i++)
3789 		init_ring_lists(&dev_priv->ring[i]);
3790 	for (i = 0; i < I915_MAX_NUM_FENCES; i++)
3791 		INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
3792 	INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
3793 			  i915_gem_retire_work_handler);
3794 	init_waitqueue_head(&dev_priv->gpu_error.reset_queue);
3795 
3796 	/* On GEN3 we really need to make sure the ARB C3 LP bit is set */
3797 	if (IS_GEN3(dev)) {
3798 		I915_WRITE(MI_ARB_STATE,
3799 			   _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
3800 	}
3801 
3802 	dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
3803 
3804 	/* Old X drivers will take 0-2 for front, back, depth buffers */
3805 	if (!drm_core_check_feature(dev, DRIVER_MODESET))
3806 		dev_priv->fence_reg_start = 3;
3807 
3808 	if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
3809 		dev_priv->num_fence_regs = 16;
3810 	else
3811 		dev_priv->num_fence_regs = 8;
3812 
3813 	/* Initialize fence registers to zero */
3814 	i915_gem_reset_fences(dev);
3815 
3816 	i915_gem_detect_bit_6_swizzle(dev);
3817 	init_waitqueue_head(&dev_priv->pending_flip_queue);
3818 
3819 	dev_priv->mm.interruptible = true;
3820 
3821 #if 0
3822 	dev_priv->mm.inactive_shrinker.shrink = i915_gem_inactive_shrink;
3823 	dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS;
3824 	register_shrinker(&dev_priv->mm.inactive_shrinker);
3825 #else
3826 	dev_priv->mm.inactive_shrinker = EVENTHANDLER_REGISTER(vm_lowmem,
3827 	    i915_gem_lowmem, dev, EVENTHANDLER_PRI_ANY);
3828 #endif
3829 }
3830 
3831 /*
3832  * Create a physically contiguous memory object for this object
3833  * e.g. for cursor + overlay regs
3834  */
3835 static int i915_gem_init_phys_object(struct drm_device *dev,
3836 				     int id, int size, int align)
3837 {
3838 	drm_i915_private_t *dev_priv = dev->dev_private;
3839 	struct drm_i915_gem_phys_object *phys_obj;
3840 	int ret;
3841 
3842 	if (dev_priv->mm.phys_objs[id - 1] || !size)
3843 		return 0;
3844 
3845 	phys_obj = kmalloc(sizeof(struct drm_i915_gem_phys_object), M_DRM,
3846 	    M_WAITOK | M_ZERO);
3847 	if (!phys_obj)
3848 		return -ENOMEM;
3849 
3850 	phys_obj->id = id;
3851 
3852 	phys_obj->handle = drm_pci_alloc(dev, size, align);
3853 	if (!phys_obj->handle) {
3854 		ret = -ENOMEM;
3855 		goto kfree_obj;
3856 	}
3857 	pmap_change_attr((vm_offset_t)phys_obj->handle->vaddr,
3858 	    size / PAGE_SIZE, PAT_WRITE_COMBINING);
3859 
3860 	dev_priv->mm.phys_objs[id - 1] = phys_obj;
3861 
3862 	return 0;
3863 
3864 kfree_obj:
3865 	drm_free(phys_obj, M_DRM);
3866 	return ret;
3867 }
3868 
3869 static void i915_gem_free_phys_object(struct drm_device *dev, int id)
3870 {
3871 	drm_i915_private_t *dev_priv = dev->dev_private;
3872 	struct drm_i915_gem_phys_object *phys_obj;
3873 
3874 	if (!dev_priv->mm.phys_objs[id - 1])
3875 		return;
3876 
3877 	phys_obj = dev_priv->mm.phys_objs[id - 1];
3878 	if (phys_obj->cur_obj) {
3879 		i915_gem_detach_phys_object(dev, phys_obj->cur_obj);
3880 	}
3881 
3882 	drm_pci_free(dev, phys_obj->handle);
3883 	drm_free(phys_obj, M_DRM);
3884 	dev_priv->mm.phys_objs[id - 1] = NULL;
3885 }
3886 
3887 void i915_gem_free_all_phys_object(struct drm_device *dev)
3888 {
3889 	int i;
3890 
3891 	for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++)
3892 		i915_gem_free_phys_object(dev, i);
3893 }
3894 
3895 void i915_gem_detach_phys_object(struct drm_device *dev,
3896 				 struct drm_i915_gem_object *obj)
3897 {
3898 	struct vm_object *mapping = obj->base.vm_obj;
3899 	char *vaddr;
3900 	int i;
3901 	int page_count;
3902 
3903 	if (!obj->phys_obj)
3904 		return;
3905 	vaddr = obj->phys_obj->handle->vaddr;
3906 
3907 	page_count = obj->base.size / PAGE_SIZE;
3908 	VM_OBJECT_LOCK(obj->base.vm_obj);
3909 	for (i = 0; i < page_count; i++) {
3910 		struct vm_page *page = shmem_read_mapping_page(mapping, i);
3911 		if (!IS_ERR(page)) {
3912 			VM_OBJECT_UNLOCK(obj->base.vm_obj);
3913 			char *dst = kmap_atomic(page);
3914 			memcpy(dst, vaddr + i*PAGE_SIZE, PAGE_SIZE);
3915 			kunmap_atomic(dst);
3916 
3917 			drm_clflush_pages(&page, 1);
3918 
3919 #if 0
3920 			set_page_dirty(page);
3921 			mark_page_accessed(page);
3922 			page_cache_release(page);
3923 #endif
3924 			VM_OBJECT_LOCK(obj->base.vm_obj);
3925 			vm_page_reference(page);
3926 			vm_page_dirty(page);
3927 			vm_page_busy_wait(page, FALSE, "i915gem");
3928 			vm_page_unwire(page, 0);
3929 			vm_page_wakeup(page);
3930 		}
3931 	}
3932 	VM_OBJECT_UNLOCK(obj->base.vm_obj);
3933 	intel_gtt_chipset_flush();
3934 
3935 	obj->phys_obj->cur_obj = NULL;
3936 	obj->phys_obj = NULL;
3937 }
3938 
3939 int
3940 i915_gem_attach_phys_object(struct drm_device *dev,
3941 			    struct drm_i915_gem_object *obj,
3942 			    int id,
3943 			    int align)
3944 {
3945 	struct vm_object *mapping = obj->base.vm_obj;
3946 	drm_i915_private_t *dev_priv = dev->dev_private;
3947 	int ret = 0;
3948 	int page_count;
3949 	int i;
3950 
3951 	if (id > I915_MAX_PHYS_OBJECT)
3952 		return -EINVAL;
3953 
3954 	if (obj->phys_obj) {
3955 		if (obj->phys_obj->id == id)
3956 			return 0;
3957 		i915_gem_detach_phys_object(dev, obj);
3958 	}
3959 
3960 	/* create a new object */
3961 	if (!dev_priv->mm.phys_objs[id - 1]) {
3962 		ret = i915_gem_init_phys_object(dev, id,
3963 						obj->base.size, align);
3964 		if (ret) {
3965 			DRM_ERROR("failed to init phys object %d size: %zu\n",
3966 				  id, obj->base.size);
3967 			return ret;
3968 		}
3969 	}
3970 
3971 	/* bind to the object */
3972 	obj->phys_obj = dev_priv->mm.phys_objs[id - 1];
3973 	obj->phys_obj->cur_obj = obj;
3974 
3975 	page_count = obj->base.size / PAGE_SIZE;
3976 
3977 	VM_OBJECT_LOCK(obj->base.vm_obj);
3978 	for (i = 0; i < page_count; i++) {
3979 		struct vm_page *page;
3980 		char *dst, *src;
3981 
3982 		page = shmem_read_mapping_page(mapping, i);
3983 		VM_OBJECT_UNLOCK(obj->base.vm_obj);
3984 		if (IS_ERR(page))
3985 			return PTR_ERR(page);
3986 
3987 		src = kmap_atomic(page);
3988 		dst = (char*)obj->phys_obj->handle->vaddr + (i * PAGE_SIZE);
3989 		memcpy(dst, src, PAGE_SIZE);
3990 		kunmap_atomic(src);
3991 
3992 #if 0
3993 		mark_page_accessed(page);
3994 		page_cache_release(page);
3995 #endif
3996 		VM_OBJECT_LOCK(obj->base.vm_obj);
3997 		vm_page_reference(page);
3998 		vm_page_busy_wait(page, FALSE, "i915gem");
3999 		vm_page_unwire(page, 0);
4000 		vm_page_wakeup(page);
4001 	}
4002 	VM_OBJECT_UNLOCK(obj->base.vm_obj);
4003 
4004 	return 0;
4005 }
4006 
4007 static int
4008 i915_gem_phys_pwrite(struct drm_device *dev,
4009 		     struct drm_i915_gem_object *obj,
4010 		     struct drm_i915_gem_pwrite *args,
4011 		     struct drm_file *file_priv)
4012 {
4013 	void *vaddr = (char *)obj->phys_obj->handle->vaddr + args->offset;
4014 	char __user *user_data = (char __user *) (uintptr_t) args->data_ptr;
4015 
4016 	if (copyin_nofault(user_data, vaddr, args->size) != 0) {
4017 		unsigned long unwritten;
4018 
4019 		/* The physical object once assigned is fixed for the lifetime
4020 		 * of the obj, so we can safely drop the lock and continue
4021 		 * to access vaddr.
4022 		 */
4023 		mutex_unlock(&dev->struct_mutex);
4024 		unwritten = copy_from_user(vaddr, user_data, args->size);
4025 		mutex_lock(&dev->struct_mutex);
4026 		if (unwritten)
4027 			return -EFAULT;
4028 	}
4029 
4030 	i915_gem_chipset_flush(dev);
4031 	return 0;
4032 }
4033 
4034 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
4035 {
4036 	struct drm_i915_file_private *file_priv = file->driver_priv;
4037 
4038 	/* Clean up our request list when the client is going away, so that
4039 	 * later retire_requests won't dereference our soon-to-be-gone
4040 	 * file_priv.
4041 	 */
4042 	spin_lock(&file_priv->mm.lock);
4043 	while (!list_empty(&file_priv->mm.request_list)) {
4044 		struct drm_i915_gem_request *request;
4045 
4046 		request = list_first_entry(&file_priv->mm.request_list,
4047 					   struct drm_i915_gem_request,
4048 					   client_list);
4049 		list_del(&request->client_list);
4050 		request->file_priv = NULL;
4051 	}
4052 	spin_unlock(&file_priv->mm.lock);
4053 }
4054 
4055 static int
4056 i915_gem_pager_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot,
4057     vm_ooffset_t foff, struct ucred *cred, u_short *color)
4058 {
4059 
4060 	*color = 0; /* XXXKIB */
4061 	return (0);
4062 }
4063 
4064 int i915_intr_pf;
4065 
4066 static int
4067 i915_gem_pager_fault(vm_object_t vm_obj, vm_ooffset_t offset, int prot,
4068     vm_page_t *mres)
4069 {
4070 	struct drm_gem_object *gem_obj;
4071 	struct drm_i915_gem_object *obj;
4072 	struct drm_device *dev;
4073 	drm_i915_private_t *dev_priv;
4074 	vm_page_t m, oldm;
4075 	int cause, ret;
4076 	bool write;
4077 
4078 	gem_obj = vm_obj->handle;
4079 	obj = to_intel_bo(gem_obj);
4080 	dev = obj->base.dev;
4081 	dev_priv = dev->dev_private;
4082 #if 0
4083 	write = (prot & VM_PROT_WRITE) != 0;
4084 #else
4085 	write = true;
4086 #endif
4087 	vm_object_pip_add(vm_obj, 1);
4088 
4089 	/*
4090 	 * Remove the placeholder page inserted by vm_fault() from the
4091 	 * object before dropping the object lock. If
4092 	 * i915_gem_release_mmap() is active in parallel on this gem
4093 	 * object, then it owns the drm device sx and might find the
4094 	 * placeholder already. Then, since the page is busy,
4095 	 * i915_gem_release_mmap() sleeps waiting for the busy state
4096 	 * of the page cleared. We will be not able to acquire drm
4097 	 * device lock until i915_gem_release_mmap() is able to make a
4098 	 * progress.
4099 	 */
4100 	if (*mres != NULL) {
4101 		oldm = *mres;
4102 		vm_page_remove(oldm);
4103 		*mres = NULL;
4104 	} else
4105 		oldm = NULL;
4106 retry:
4107 	VM_OBJECT_UNLOCK(vm_obj);
4108 unlocked_vmobj:
4109 	cause = ret = 0;
4110 	m = NULL;
4111 
4112 	if (i915_intr_pf) {
4113 		ret = i915_mutex_lock_interruptible(dev);
4114 		if (ret != 0) {
4115 			cause = 10;
4116 			goto out;
4117 		}
4118 	} else
4119 		mutex_lock(&dev->struct_mutex);
4120 
4121 	/*
4122 	 * Since the object lock was dropped, other thread might have
4123 	 * faulted on the same GTT address and instantiated the
4124 	 * mapping for the page.  Recheck.
4125 	 */
4126 	VM_OBJECT_LOCK(vm_obj);
4127 	m = vm_page_lookup(vm_obj, OFF_TO_IDX(offset));
4128 	if (m != NULL) {
4129 		if ((m->flags & PG_BUSY) != 0) {
4130 			mutex_unlock(&dev->struct_mutex);
4131 #if 0 /* XXX */
4132 			vm_page_sleep(m, "915pee");
4133 #endif
4134 			goto retry;
4135 		}
4136 		goto have_page;
4137 	} else
4138 		VM_OBJECT_UNLOCK(vm_obj);
4139 
4140 	/* Access to snoopable pages through the GTT is incoherent. */
4141 	if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) {
4142 		ret = -EINVAL;
4143 		goto unlock;
4144 	}
4145 
4146 	/* Now bind it into the GTT if needed */
4147 	if (!obj->map_and_fenceable) {
4148 		ret = i915_gem_object_unbind(obj);
4149 		if (ret != 0) {
4150 			cause = 20;
4151 			goto unlock;
4152 		}
4153 	}
4154 	if (!obj->gtt_space) {
4155 		ret = i915_gem_object_bind_to_gtt(obj, 0, true, false);
4156 		if (ret != 0) {
4157 			cause = 30;
4158 			goto unlock;
4159 		}
4160 
4161 		ret = i915_gem_object_set_to_gtt_domain(obj, write);
4162 		if (ret != 0) {
4163 			cause = 40;
4164 			goto unlock;
4165 		}
4166 	}
4167 
4168 	if (obj->tiling_mode == I915_TILING_NONE)
4169 		ret = i915_gem_object_put_fence(obj);
4170 	else
4171 		ret = i915_gem_object_get_fence(obj);
4172 	if (ret != 0) {
4173 		cause = 50;
4174 		goto unlock;
4175 	}
4176 
4177 	if (i915_gem_object_is_inactive(obj))
4178 		list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
4179 
4180 	obj->fault_mappable = true;
4181 	VM_OBJECT_LOCK(vm_obj);
4182 	m = vm_phys_fictitious_to_vm_page(dev->agp->base + obj->gtt_offset +
4183 	    offset);
4184 	if (m == NULL) {
4185 		cause = 60;
4186 		ret = -EFAULT;
4187 		goto unlock;
4188 	}
4189 	KASSERT((m->flags & PG_FICTITIOUS) != 0,
4190 	    ("not fictitious %p", m));
4191 	KASSERT(m->wire_count == 1, ("wire_count not 1 %p", m));
4192 
4193 	if ((m->flags & PG_BUSY) != 0) {
4194 		mutex_unlock(&dev->struct_mutex);
4195 #if 0 /* XXX */
4196 		vm_page_sleep(m, "915pbs");
4197 #endif
4198 		goto retry;
4199 	}
4200 	m->valid = VM_PAGE_BITS_ALL;
4201 	vm_page_insert(m, vm_obj, OFF_TO_IDX(offset));
4202 have_page:
4203 	*mres = m;
4204 	vm_page_busy_try(m, false);
4205 
4206 	mutex_unlock(&dev->struct_mutex);
4207 	if (oldm != NULL) {
4208 		vm_page_free(oldm);
4209 	}
4210 	vm_object_pip_wakeup(vm_obj);
4211 	return (VM_PAGER_OK);
4212 
4213 unlock:
4214 	mutex_unlock(&dev->struct_mutex);
4215 out:
4216 	KASSERT(ret != 0, ("i915_gem_pager_fault: wrong return"));
4217 	if (ret == -EAGAIN || ret == -EIO || ret == -EINTR) {
4218 		goto unlocked_vmobj;
4219 	}
4220 	VM_OBJECT_LOCK(vm_obj);
4221 	vm_object_pip_wakeup(vm_obj);
4222 	return (VM_PAGER_ERROR);
4223 }
4224 
4225 static void
4226 i915_gem_pager_dtor(void *handle)
4227 {
4228 	struct drm_gem_object *obj;
4229 	struct drm_device *dev;
4230 
4231 	obj = handle;
4232 	dev = obj->dev;
4233 
4234 	mutex_lock(&dev->struct_mutex);
4235 	drm_gem_free_mmap_offset(obj);
4236 	i915_gem_release_mmap(to_intel_bo(obj));
4237 	drm_gem_object_unreference(obj);
4238 	mutex_unlock(&dev->struct_mutex);
4239 }
4240 
4241 struct cdev_pager_ops i915_gem_pager_ops = {
4242 	.cdev_pg_fault	= i915_gem_pager_fault,
4243 	.cdev_pg_ctor	= i915_gem_pager_ctor,
4244 	.cdev_pg_dtor	= i915_gem_pager_dtor
4245 };
4246 
4247 #define	GEM_PARANOID_CHECK_GTT 0
4248 #if GEM_PARANOID_CHECK_GTT
4249 static void
4250 i915_gem_assert_pages_not_mapped(struct drm_device *dev, vm_page_t *ma,
4251     int page_count)
4252 {
4253 	struct drm_i915_private *dev_priv;
4254 	vm_paddr_t pa;
4255 	unsigned long start, end;
4256 	u_int i;
4257 	int j;
4258 
4259 	dev_priv = dev->dev_private;
4260 	start = OFF_TO_IDX(dev_priv->mm.gtt_start);
4261 	end = OFF_TO_IDX(dev_priv->mm.gtt_end);
4262 	for (i = start; i < end; i++) {
4263 		pa = intel_gtt_read_pte_paddr(i);
4264 		for (j = 0; j < page_count; j++) {
4265 			if (pa == VM_PAGE_TO_PHYS(ma[j])) {
4266 				panic("Page %p in GTT pte index %d pte %x",
4267 				    ma[i], i, intel_gtt_read_pte(i));
4268 			}
4269 		}
4270 	}
4271 }
4272 #endif
4273 
4274 static int
4275 i915_gpu_is_active(struct drm_device *dev)
4276 {
4277 	drm_i915_private_t *dev_priv = dev->dev_private;
4278 
4279 	return !list_empty(&dev_priv->mm.active_list);
4280 }
4281 
4282 static void
4283 i915_gem_lowmem(void *arg)
4284 {
4285 	struct drm_device *dev;
4286 	struct drm_i915_private *dev_priv;
4287 	struct drm_i915_gem_object *obj, *next;
4288 	int cnt, cnt_fail, cnt_total;
4289 
4290 	dev = arg;
4291 	dev_priv = dev->dev_private;
4292 
4293 	if (lockmgr(&dev->struct_mutex, LK_EXCLUSIVE|LK_NOWAIT))
4294 		return;
4295 
4296 rescan:
4297 	/* first scan for clean buffers */
4298 	i915_gem_retire_requests(dev);
4299 
4300 	cnt_total = cnt_fail = cnt = 0;
4301 
4302 	list_for_each_entry_safe(obj, next, &dev_priv->mm.inactive_list,
4303 	    mm_list) {
4304 		if (i915_gem_object_is_purgeable(obj)) {
4305 			if (i915_gem_object_unbind(obj) != 0)
4306 				cnt_total++;
4307 		} else
4308 			cnt_total++;
4309 	}
4310 
4311 	/* second pass, evict/count anything still on the inactive list */
4312 	list_for_each_entry_safe(obj, next, &dev_priv->mm.inactive_list,
4313 	    mm_list) {
4314 		if (i915_gem_object_unbind(obj) == 0)
4315 			cnt++;
4316 		else
4317 			cnt_fail++;
4318 	}
4319 
4320 	if (cnt_fail > cnt_total / 100 && i915_gpu_is_active(dev)) {
4321 		/*
4322 		 * We are desperate for pages, so as a last resort, wait
4323 		 * for the GPU to finish and discard whatever we can.
4324 		 * This has a dramatic impact to reduce the number of
4325 		 * OOM-killer events whilst running the GPU aggressively.
4326 		 */
4327 		if (i915_gpu_idle(dev) == 0)
4328 			goto rescan;
4329 	}
4330 	mutex_unlock(&dev->struct_mutex);
4331 }
4332