Linux Audio

Check our new training course

Loading...
v6.13.7
   1/*
   2 * Copyright © 2008-2015 Intel Corporation
   3 *
   4 * Permission is hereby granted, free of charge, to any person obtaining a
   5 * copy of this software and associated documentation files (the "Software"),
   6 * to deal in the Software without restriction, including without limitation
   7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8 * and/or sell copies of the Software, and to permit persons to whom the
   9 * Software is furnished to do so, subject to the following conditions:
  10 *
  11 * The above copyright notice and this permission notice (including the next
  12 * paragraph) shall be included in all copies or substantial portions of the
  13 * Software.
  14 *
  15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21 * IN THE SOFTWARE.
  22 *
  23 * Authors:
  24 *    Eric Anholt <eric@anholt.net>
  25 *
  26 */
  27
  28#include <linux/dma-fence-array.h>
  29#include <linux/kthread.h>
  30#include <linux/dma-resv.h>
 
 
 
  31#include <linux/shmem_fs.h>
  32#include <linux/slab.h>
  33#include <linux/stop_machine.h>
  34#include <linux/swap.h>
  35#include <linux/pci.h>
  36#include <linux/dma-buf.h>
  37#include <linux/mman.h>
  38
  39#include <drm/drm_cache.h>
  40#include <drm/drm_vma_manager.h>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  41
  42#include "gem/i915_gem_clflush.h"
  43#include "gem/i915_gem_context.h"
  44#include "gem/i915_gem_ioctls.h"
  45#include "gem/i915_gem_mman.h"
  46#include "gem/i915_gem_object_frontbuffer.h"
  47#include "gem/i915_gem_pm.h"
  48#include "gem/i915_gem_region.h"
  49#include "gt/intel_engine_user.h"
  50#include "gt/intel_gt.h"
  51#include "gt/intel_gt_pm.h"
  52#include "gt/intel_workarounds.h"
  53
  54#include "i915_drv.h"
  55#include "i915_file_private.h"
  56#include "i915_trace.h"
  57#include "i915_vgpu.h"
  58#include "intel_clock_gating.h"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  59
  60static int
  61insert_mappable_node(struct i915_ggtt *ggtt, struct drm_mm_node *node, u32 size)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  62{
  63	int err;
 
  64
  65	err = mutex_lock_interruptible(&ggtt->vm.mutex);
  66	if (err)
  67		return err;
  68
  69	memset(node, 0, sizeof(*node));
  70	err = drm_mm_insert_node_in_range(&ggtt->vm.mm, node,
  71					  size, 0, I915_COLOR_UNEVICTABLE,
  72					  0, ggtt->mappable_end,
  73					  DRM_MM_INSERT_LOW);
  74
  75	mutex_unlock(&ggtt->vm.mutex);
 
 
  76
  77	return err;
 
 
 
  78}
  79
  80static void
  81remove_mappable_node(struct i915_ggtt *ggtt, struct drm_mm_node *node)
 
  82{
  83	mutex_lock(&ggtt->vm.mutex);
  84	drm_mm_remove_node(node);
  85	mutex_unlock(&ggtt->vm.mutex);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  86}
  87
  88int
  89i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
  90			    struct drm_file *file)
  91{
  92	struct drm_i915_private *i915 = to_i915(dev);
  93	struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
  94	struct drm_i915_gem_get_aperture *args = data;
  95	struct i915_vma *vma;
  96	u64 pinned;
  97
  98	if (mutex_lock_interruptible(&ggtt->vm.mutex))
  99		return -EINTR;
 100
 101	pinned = ggtt->vm.reserved;
 102	list_for_each_entry(vma, &ggtt->vm.bound_list, vm_link)
 103		if (i915_vma_is_pinned(vma))
 104			pinned += vma->node.size;
 105
 106	mutex_unlock(&ggtt->vm.mutex);
 
 
 
 
 
 107
 108	args->aper_size = ggtt->vm.total;
 109	args->aper_available_size = args->aper_size - pinned;
 110
 111	return 0;
 112}
 113
 114int i915_gem_object_unbind(struct drm_i915_gem_object *obj,
 115			   unsigned long flags)
 116{
 117	struct intel_runtime_pm *rpm = &to_i915(obj->base.dev)->runtime_pm;
 118	bool vm_trylock = !!(flags & I915_GEM_OBJECT_UNBIND_VM_TRYLOCK);
 119	LIST_HEAD(still_in_list);
 120	intel_wakeref_t wakeref;
 121	struct i915_vma *vma;
 122	int ret;
 123
 124	assert_object_held(obj);
 
 125
 126	if (list_empty(&obj->vma.list))
 127		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 128
 129	/*
 130	 * As some machines use ACPI to handle runtime-resume callbacks, and
 131	 * ACPI is quite kmalloc happy, we cannot resume beneath the vm->mutex
 132	 * as they are required by the shrinker. Ergo, we wake the device up
 133	 * first just in case.
 134	 */
 135	wakeref = intel_runtime_pm_get(rpm);
 136
 137try_again:
 138	ret = 0;
 139	spin_lock(&obj->vma.lock);
 140	while (!ret && (vma = list_first_entry_or_null(&obj->vma.list,
 141						       struct i915_vma,
 142						       obj_link))) {
 143		list_move_tail(&vma->obj_link, &still_in_list);
 144		if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK))
 145			continue;
 146
 147		if (flags & I915_GEM_OBJECT_UNBIND_TEST) {
 148			ret = -EBUSY;
 149			break;
 150		}
 151
 152		/*
 153		 * Requiring the vm destructor to take the object lock
 154		 * before destroying a vma would help us eliminate the
 155		 * i915_vm_tryget() here, AND thus also the barrier stuff
 156		 * at the end. That's an easy fix, but sleeping locks in
 157		 * a kthread should generally be avoided.
 158		 */
 159		ret = -EAGAIN;
 160		if (!i915_vm_tryget(vma->vm))
 161			break;
 162
 163		spin_unlock(&obj->vma.lock);
 
 164
 165		/*
 166		 * Since i915_vma_parked() takes the object lock
 167		 * before vma destruction, it won't race us here,
 168		 * and destroy the vma from under us.
 169		 */
 170
 171		ret = -EBUSY;
 172		if (flags & I915_GEM_OBJECT_UNBIND_ASYNC) {
 173			assert_object_held(vma->obj);
 174			ret = i915_vma_unbind_async(vma, vm_trylock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 175		}
 176
 177		if (ret == -EBUSY && (flags & I915_GEM_OBJECT_UNBIND_ACTIVE ||
 178				      !i915_vma_is_active(vma))) {
 179			if (vm_trylock) {
 180				if (mutex_trylock(&vma->vm->mutex)) {
 181					ret = __i915_vma_unbind(vma);
 182					mutex_unlock(&vma->vm->mutex);
 183				}
 184			} else {
 185				ret = i915_vma_unbind(vma);
 186			}
 187		}
 188
 189		i915_vm_put(vma->vm);
 190		spin_lock(&obj->vma.lock);
 191	}
 192	list_splice_init(&still_in_list, &obj->vma.list);
 193	spin_unlock(&obj->vma.lock);
 194
 195	if (ret == -EAGAIN && flags & I915_GEM_OBJECT_UNBIND_BARRIER) {
 196		rcu_barrier(); /* flush the i915_vm_release() */
 197		goto try_again;
 198	}
 199
 200	intel_runtime_pm_put(rpm, wakeref);
 201
 202	return ret;
 203}
 204
 205static int
 206shmem_pread(struct page *page, int offset, int len, char __user *user_data,
 207	    bool needs_clflush)
 
 208{
 209	char *vaddr;
 210	int ret;
 211
 212	vaddr = kmap(page);
 213
 214	if (needs_clflush)
 215		drm_clflush_virt_range(vaddr + offset, len);
 
 
 
 
 
 
 
 
 
 
 216
 217	ret = __copy_to_user(user_data, vaddr + offset, len);
 
 
 218
 219	kunmap(page);
 
 
 
 
 220
 221	return ret ? -EFAULT : 0;
 
 
 
 222}
 223
 224static int
 225i915_gem_shmem_pread(struct drm_i915_gem_object *obj,
 226		     struct drm_i915_gem_pread *args)
 
 
 227{
 228	unsigned int needs_clflush;
 229	char __user *user_data;
 230	unsigned long offset;
 231	pgoff_t idx;
 232	u64 remain;
 233	int ret;
 
 
 
 
 
 234
 235	ret = i915_gem_object_lock_interruptible(obj, NULL);
 
 
 
 
 
 
 
 236	if (ret)
 237		return ret;
 238
 239	ret = i915_gem_object_pin_pages(obj);
 240	if (ret)
 241		goto err_unlock;
 242
 243	ret = i915_gem_object_prepare_read(obj, &needs_clflush);
 244	if (ret)
 245		goto err_unpin;
 
 
 
 
 
 
 
 
 246
 247	i915_gem_object_finish_access(obj);
 248	i915_gem_object_unlock(obj);
 
 
 
 
 
 
 249
 250	remain = args->size;
 251	user_data = u64_to_user_ptr(args->data_ptr);
 252	offset = offset_in_page(args->offset);
 253	for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
 254		struct page *page = i915_gem_object_get_page(obj, idx);
 255		unsigned int length = min_t(u64, remain, PAGE_SIZE - offset);
 
 
 
 
 256
 257		ret = shmem_pread(page, offset, length, user_data,
 258				  needs_clflush);
 
 
 
 
 
 
 259		if (ret)
 260			break;
 261
 262		remain -= length;
 263		user_data += length;
 264		offset = 0;
 265	}
 266
 267	i915_gem_object_unpin_pages(obj);
 268	return ret;
 269
 270err_unpin:
 271	i915_gem_object_unpin_pages(obj);
 272err_unlock:
 273	i915_gem_object_unlock(obj);
 274	return ret;
 275}
 276
 277static inline bool
 278gtt_user_read(struct io_mapping *mapping,
 279	      loff_t base, int offset,
 280	      char __user *user_data, int length)
 281{
 282	void __iomem *vaddr;
 283	unsigned long unwritten;
 284
 285	/* We can use the cpu mem copy function because this is X86. */
 286	vaddr = io_mapping_map_atomic_wc(mapping, base);
 287	unwritten = __copy_to_user_inatomic(user_data,
 288					    (void __force *)vaddr + offset,
 289					    length);
 290	io_mapping_unmap_atomic(vaddr);
 291	if (unwritten) {
 292		vaddr = io_mapping_map_wc(mapping, base, PAGE_SIZE);
 293		unwritten = copy_to_user(user_data,
 294					 (void __force *)vaddr + offset,
 295					 length);
 296		io_mapping_unmap(vaddr);
 
 
 297	}
 298	return unwritten;
 
 299}
 300
 301static struct i915_vma *i915_gem_gtt_prepare(struct drm_i915_gem_object *obj,
 302					     struct drm_mm_node *node,
 303					     bool write)
 
 
 
 
 304{
 305	struct drm_i915_private *i915 = to_i915(obj->base.dev);
 306	struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
 307	struct i915_vma *vma;
 308	struct i915_gem_ww_ctx ww;
 309	int ret;
 310
 311	i915_gem_ww_ctx_init(&ww, true);
 312retry:
 313	vma = ERR_PTR(-ENODEV);
 314	ret = i915_gem_object_lock(obj, &ww);
 315	if (ret)
 316		goto err_ww;
 317
 318	ret = i915_gem_object_set_to_gtt_domain(obj, write);
 319	if (ret)
 320		goto err_ww;
 321
 322	if (!i915_gem_object_is_tiled(obj))
 323		vma = i915_gem_object_ggtt_pin_ww(obj, &ww, NULL, 0, 0,
 324						  PIN_MAPPABLE |
 325						  PIN_NONBLOCK /* NOWARN */ |
 326						  PIN_NOEVICT);
 327	if (vma == ERR_PTR(-EDEADLK)) {
 328		ret = -EDEADLK;
 329		goto err_ww;
 330	} else if (!IS_ERR(vma)) {
 331		node->start = i915_ggtt_offset(vma);
 332		node->flags = 0;
 333	} else {
 334		ret = insert_mappable_node(ggtt, node, PAGE_SIZE);
 335		if (ret)
 336			goto err_ww;
 337		GEM_BUG_ON(!drm_mm_node_allocated(node));
 338		vma = NULL;
 339	}
 340
 341	ret = i915_gem_object_pin_pages(obj);
 342	if (ret) {
 343		if (drm_mm_node_allocated(node)) {
 344			ggtt->vm.clear_range(&ggtt->vm, node->start, node->size);
 345			remove_mappable_node(ggtt, node);
 346		} else {
 347			i915_vma_unpin(vma);
 348		}
 349	}
 350
 351err_ww:
 352	if (ret == -EDEADLK) {
 353		ret = i915_gem_ww_ctx_backoff(&ww);
 354		if (!ret)
 355			goto retry;
 356	}
 357	i915_gem_ww_ctx_fini(&ww);
 358
 359	return ret ? ERR_PTR(ret) : vma;
 360}
 361
 362static void i915_gem_gtt_cleanup(struct drm_i915_gem_object *obj,
 363				 struct drm_mm_node *node,
 364				 struct i915_vma *vma)
 
 
 
 
 365{
 366	struct drm_i915_private *i915 = to_i915(obj->base.dev);
 367	struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
 368
 369	i915_gem_object_unpin_pages(obj);
 370	if (drm_mm_node_allocated(node)) {
 371		ggtt->vm.clear_range(&ggtt->vm, node->start, node->size);
 372		remove_mappable_node(ggtt, node);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 373	} else {
 374		i915_vma_unpin(vma);
 375	}
 
 376}
 377
 
 
 378static int
 379i915_gem_gtt_pread(struct drm_i915_gem_object *obj,
 380		   const struct drm_i915_gem_pread *args)
 
 381{
 382	struct drm_i915_private *i915 = to_i915(obj->base.dev);
 383	struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
 384	unsigned long remain, offset;
 385	intel_wakeref_t wakeref;
 386	struct drm_mm_node node;
 387	void __user *user_data;
 388	struct i915_vma *vma;
 389	int ret = 0;
 390
 391	if (overflows_type(args->size, remain) ||
 392	    overflows_type(args->offset, offset))
 393		return -EINVAL;
 
 
 
 
 
 
 
 
 
 
 
 
 394
 395	wakeref = intel_runtime_pm_get(&i915->runtime_pm);
 
 396
 397	vma = i915_gem_gtt_prepare(obj, &node, false);
 398	if (IS_ERR(vma)) {
 399		ret = PTR_ERR(vma);
 400		goto out_rpm;
 401	}
 
 
 
 
 
 
 
 
 
 402
 403	user_data = u64_to_user_ptr(args->data_ptr);
 404	remain = args->size;
 
 
 
 
 
 
 
 405	offset = args->offset;
 406
 407	while (remain > 0) {
 
 
 
 
 
 
 408		/* Operation in this page
 409		 *
 410		 * page_base = page offset within aperture
 411		 * page_offset = offset within page
 412		 * page_length = bytes to copy for this page
 413		 */
 414		u32 page_base = node.start;
 415		unsigned page_offset = offset_in_page(offset);
 416		unsigned page_length = PAGE_SIZE - page_offset;
 417		page_length = remain < page_length ? remain : page_length;
 418		if (drm_mm_node_allocated(&node)) {
 419			ggtt->vm.insert_page(&ggtt->vm,
 420					     i915_gem_object_get_dma_address(obj,
 421									     offset >> PAGE_SHIFT),
 422					     node.start,
 423					     i915_gem_get_pat_index(i915,
 424								    I915_CACHE_NONE), 0);
 425		} else {
 426			page_base += offset & PAGE_MASK;
 
 
 
 
 
 
 
 
 
 
 
 427		}
 428
 429		if (gtt_user_read(&ggtt->iomap, page_base, page_offset,
 430				  user_data, page_length)) {
 431			ret = -EFAULT;
 432			break;
 433		}
 434
 
 
 
 
 
 
 435		remain -= page_length;
 436		user_data += page_length;
 437		offset += page_length;
 438	}
 439
 440	i915_gem_gtt_cleanup(obj, &node, vma);
 441out_rpm:
 442	intel_runtime_pm_put(&i915->runtime_pm, wakeref);
 443	return ret;
 444}
 445
 446/**
 447 * i915_gem_pread_ioctl - Reads data from the object referenced by handle.
 448 * @dev: drm device pointer
 449 * @data: ioctl data blob
 450 * @file: drm file pointer
 451 *
 452 * On error, the contents of *data are undefined.
 453 */
 454int
 455i915_gem_pread_ioctl(struct drm_device *dev, void *data,
 456		     struct drm_file *file)
 457{
 458	struct drm_i915_private *i915 = to_i915(dev);
 459	struct drm_i915_gem_pread *args = data;
 460	struct drm_i915_gem_object *obj;
 461	int ret;
 462
 463	/* PREAD is disallowed for all platforms after TGL-LP.  This also
 464	 * covers all platforms with local memory.
 465	 */
 466	if (GRAPHICS_VER(i915) >= 12 && !IS_TIGERLAKE(i915))
 467		return -EOPNOTSUPP;
 468
 469	if (args->size == 0)
 470		return 0;
 471
 472	if (!access_ok(u64_to_user_ptr(args->data_ptr),
 
 473		       args->size))
 474		return -EFAULT;
 475
 476	obj = i915_gem_object_lookup(file, args->handle);
 477	if (!obj)
 478		return -ENOENT;
 
 
 
 
 
 
 479
 480	/* Bounds check source.  */
 481	if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) {
 
 482		ret = -EINVAL;
 483		goto out;
 484	}
 485
 486	trace_i915_gem_object_pread(obj, args->offset, args->size);
 487	ret = -ENODEV;
 488	if (obj->ops->pread)
 489		ret = obj->ops->pread(obj, args);
 490	if (ret != -ENODEV)
 491		goto out;
 
 492
 493	ret = i915_gem_object_wait(obj,
 494				   I915_WAIT_INTERRUPTIBLE,
 495				   MAX_SCHEDULE_TIMEOUT);
 496	if (ret)
 497		goto out;
 498
 499	ret = i915_gem_shmem_pread(obj, args);
 500	if (ret == -EFAULT || ret == -ENODEV)
 501		ret = i915_gem_gtt_pread(obj, args);
 502
 503out:
 504	i915_gem_object_put(obj);
 
 
 505	return ret;
 506}
 507
 508/* This is the fast write path which cannot handle
 509 * page faults in the source data
 510 */
 511
 512static inline bool
 513ggtt_write(struct io_mapping *mapping,
 514	   loff_t base, int offset,
 515	   char __user *user_data, int length)
 
 516{
 517	void __iomem *vaddr;
 
 518	unsigned long unwritten;
 519
 
 520	/* We can use the cpu mem copy function because this is X86. */
 521	vaddr = io_mapping_map_atomic_wc(mapping, base);
 522	unwritten = __copy_from_user_inatomic_nocache((void __force *)vaddr + offset,
 523						      user_data, length);
 524	io_mapping_unmap_atomic(vaddr);
 525	if (unwritten) {
 526		vaddr = io_mapping_map_wc(mapping, base, PAGE_SIZE);
 527		unwritten = copy_from_user((void __force *)vaddr + offset,
 528					   user_data, length);
 529		io_mapping_unmap(vaddr);
 530	}
 531
 532	return unwritten;
 533}
 534
 535/**
 536 * i915_gem_gtt_pwrite_fast - This is the fast pwrite path, where we copy the data directly from the
 537 * user into the GTT, uncached.
 538 * @obj: i915 GEM object
 539 * @args: pwrite arguments structure
 540 */
 541static int
 542i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
 543			 const struct drm_i915_gem_pwrite *args)
 
 
 544{
 545	struct drm_i915_private *i915 = to_i915(obj->base.dev);
 546	struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
 547	struct intel_runtime_pm *rpm = &i915->runtime_pm;
 548	unsigned long remain, offset;
 549	intel_wakeref_t wakeref;
 550	struct drm_mm_node node;
 551	struct i915_vma *vma;
 552	void __user *user_data;
 553	int ret = 0;
 554
 555	if (overflows_type(args->size, remain) ||
 556	    overflows_type(args->offset, offset))
 557		return -EINVAL;
 558
 559	if (i915_gem_object_has_struct_page(obj)) {
 560		/*
 561		 * Avoid waking the device up if we can fallback, as
 562		 * waking/resuming is very slow (worst-case 10-100 ms
 563		 * depending on PCI sleeps and our own resume time).
 564		 * This easily dwarfs any performance advantage from
 565		 * using the cache bypass of indirect GGTT access.
 566		 */
 567		wakeref = intel_runtime_pm_get_if_in_use(rpm);
 568		if (!wakeref)
 569			return -EFAULT;
 570	} else {
 571		/* No backing pages, no fallback, we must force GGTT access */
 572		wakeref = intel_runtime_pm_get(rpm);
 573	}
 574
 575	vma = i915_gem_gtt_prepare(obj, &node, true);
 576	if (IS_ERR(vma)) {
 577		ret = PTR_ERR(vma);
 578		goto out_rpm;
 579	}
 580
 581	i915_gem_object_invalidate_frontbuffer(obj, ORIGIN_CPU);
 
 
 582
 583	user_data = u64_to_user_ptr(args->data_ptr);
 584	offset = args->offset;
 585	remain = args->size;
 586	while (remain) {
 
 
 
 587		/* Operation in this page
 588		 *
 589		 * page_base = page offset within aperture
 590		 * page_offset = offset within page
 591		 * page_length = bytes to copy for this page
 592		 */
 593		u32 page_base = node.start;
 594		unsigned int page_offset = offset_in_page(offset);
 595		unsigned int page_length = PAGE_SIZE - page_offset;
 596		page_length = remain < page_length ? remain : page_length;
 597		if (drm_mm_node_allocated(&node)) {
 598			/* flush the write before we modify the GGTT */
 599			intel_gt_flush_ggtt_writes(ggtt->vm.gt);
 600			ggtt->vm.insert_page(&ggtt->vm,
 601					     i915_gem_object_get_dma_address(obj,
 602									     offset >> PAGE_SHIFT),
 603					     node.start,
 604					     i915_gem_get_pat_index(i915,
 605								    I915_CACHE_NONE), 0);
 606			wmb(); /* flush modifications to the GGTT (insert_page) */
 607		} else {
 608			page_base += offset & PAGE_MASK;
 609		}
 610		/* If we get a fault while copying data, then (presumably) our
 611		 * source page isn't available.  Return the error and we'll
 612		 * retry in the slow path.
 613		 * If the object is non-shmem backed, we retry again with the
 614		 * path that handles page fault.
 615		 */
 616		if (ggtt_write(&ggtt->iomap, page_base, page_offset,
 617			       user_data, page_length)) {
 618			ret = -EFAULT;
 619			break;
 620		}
 621
 622		remain -= page_length;
 623		user_data += page_length;
 624		offset += page_length;
 625	}
 626
 627	intel_gt_flush_ggtt_writes(ggtt->vm.gt);
 628	i915_gem_object_flush_frontbuffer(obj, ORIGIN_CPU);
 629
 630	i915_gem_gtt_cleanup(obj, &node, vma);
 631out_rpm:
 632	intel_runtime_pm_put(rpm, wakeref);
 633	return ret;
 634}
 635
 636/* Per-page copy function for the shmem pwrite fastpath.
 637 * Flushes invalid cachelines before writing to the target if
 638 * needs_clflush_before is set and flushes out any written cachelines after
 639 * writing if needs_clflush is set.
 640 */
 641static int
 642shmem_pwrite(struct page *page, int offset, int len, char __user *user_data,
 643	     bool needs_clflush_before,
 644	     bool needs_clflush_after)
 
 
 645{
 646	char *vaddr;
 647	int ret;
 648
 649	vaddr = kmap(page);
 
 650
 
 651	if (needs_clflush_before)
 652		drm_clflush_virt_range(vaddr + offset, len);
 
 
 
 
 
 
 
 
 
 
 653
 654	ret = __copy_from_user(vaddr + offset, user_data, len);
 655	if (!ret && needs_clflush_after)
 656		drm_clflush_virt_range(vaddr + offset, len);
 
 
 
 
 
 
 
 
 657
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 658	kunmap(page);
 659
 660	return ret ? -EFAULT : 0;
 661}
 662
 663static int
 664i915_gem_shmem_pwrite(struct drm_i915_gem_object *obj,
 665		      const struct drm_i915_gem_pwrite *args)
 
 
 666{
 667	unsigned int partial_cacheline_write;
 668	unsigned int needs_clflush;
 669	void __user *user_data;
 670	unsigned long offset;
 671	pgoff_t idx;
 672	u64 remain;
 673	int ret;
 
 
 
 
 
 674
 675	ret = i915_gem_object_lock_interruptible(obj, NULL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 676	if (ret)
 677		return ret;
 678
 679	ret = i915_gem_object_pin_pages(obj);
 680	if (ret)
 681		goto err_unlock;
 682
 683	ret = i915_gem_object_prepare_write(obj, &needs_clflush);
 684	if (ret)
 685		goto err_unpin;
 686
 687	i915_gem_object_finish_access(obj);
 688	i915_gem_object_unlock(obj);
 
 
 689
 690	/* If we don't overwrite a cacheline completely we need to be
 691	 * careful to have up-to-date data by first clflushing. Don't
 692	 * overcomplicate things and flush the entire patch.
 693	 */
 694	partial_cacheline_write = 0;
 695	if (needs_clflush & CLFLUSH_BEFORE)
 696		partial_cacheline_write = boot_cpu_data.x86_clflush_size - 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 697
 698	user_data = u64_to_user_ptr(args->data_ptr);
 699	remain = args->size;
 700	offset = offset_in_page(args->offset);
 701	for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
 702		struct page *page = i915_gem_object_get_page(obj, idx);
 703		unsigned int length = min_t(u64, remain, PAGE_SIZE - offset);
 704
 705		ret = shmem_pwrite(page, offset, length, user_data,
 706				   (offset | length) & partial_cacheline_write,
 707				   needs_clflush & CLFLUSH_AFTER);
 708		if (ret)
 709			break;
 710
 711		remain -= length;
 712		user_data += length;
 713		offset = 0;
 
 714	}
 715
 716	i915_gem_object_flush_frontbuffer(obj, ORIGIN_CPU);
 717
 718	i915_gem_object_unpin_pages(obj);
 719	return ret;
 720
 721err_unpin:
 722	i915_gem_object_unpin_pages(obj);
 723err_unlock:
 724	i915_gem_object_unlock(obj);
 
 
 
 
 
 
 
 
 
 
 
 
 725	return ret;
 726}
 727
 728/**
 729 * i915_gem_pwrite_ioctl - Writes data to the object referenced by handle.
 730 * @dev: drm device
 731 * @data: ioctl data blob
 732 * @file: drm file
 733 *
 734 * On error, the contents of the buffer that were to be modified are undefined.
 735 */
 736int
 737i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
 738		      struct drm_file *file)
 739{
 740	struct drm_i915_private *i915 = to_i915(dev);
 741	struct drm_i915_gem_pwrite *args = data;
 742	struct drm_i915_gem_object *obj;
 743	int ret;
 744
 745	/* PWRITE is disallowed for all platforms after TGL-LP.  This also
 746	 * covers all platforms with local memory.
 747	 */
 748	if (GRAPHICS_VER(i915) >= 12 && !IS_TIGERLAKE(i915))
 749		return -EOPNOTSUPP;
 750
 751	if (args->size == 0)
 752		return 0;
 753
 754	if (!access_ok(u64_to_user_ptr(args->data_ptr), args->size))
 
 
 755		return -EFAULT;
 756
 757	obj = i915_gem_object_lookup(file, args->handle);
 758	if (!obj)
 759		return -ENOENT;
 
 
 
 
 
 
 
 
 
 
 
 
 
 760
 761	/* Bounds check destination. */
 762	if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) {
 
 763		ret = -EINVAL;
 764		goto err;
 765	}
 766
 767	/* Writes not allowed into this read-only object */
 768	if (i915_gem_object_is_readonly(obj)) {
 
 
 769		ret = -EINVAL;
 770		goto err;
 771	}
 772
 773	trace_i915_gem_object_pwrite(obj, args->offset, args->size);
 774
 775	ret = -ENODEV;
 776	if (obj->ops->pwrite)
 777		ret = obj->ops->pwrite(obj, args);
 778	if (ret != -ENODEV)
 779		goto err;
 780
 781	ret = i915_gem_object_wait(obj,
 782				   I915_WAIT_INTERRUPTIBLE |
 783				   I915_WAIT_ALL,
 784				   MAX_SCHEDULE_TIMEOUT);
 785	if (ret)
 786		goto err;
 787
 788	ret = -EFAULT;
 789	/* We can only do the GTT pwrite on untiled buffers, as otherwise
 790	 * it would end up going through the fenced access, and we'll get
 791	 * different detiling behavior between reading and writing.
 792	 * pread/pwrite currently are reading and writing from the CPU
 793	 * perspective, requiring manual detiling by the client.
 794	 */
 795	if (!i915_gem_object_has_struct_page(obj) ||
 796	    i915_gem_cpu_write_needs_clflush(obj))
 
 
 
 
 
 
 
 797		/* Note that the gtt paths might fail with non-page-backed user
 798		 * pointers (e.g. gtt mappings when moving data between
 799		 * textures). Fallback to the shmem path in that case.
 800		 */
 801		ret = i915_gem_gtt_pwrite_fast(obj, args);
 
 
 802
 803	if (ret == -EFAULT || ret == -ENOSPC) {
 804		if (i915_gem_object_has_struct_page(obj))
 805			ret = i915_gem_shmem_pwrite(obj, args);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 806	}
 807
 808err:
 809	i915_gem_object_put(obj);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 810	return ret;
 811}
 812
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 813/**
 814 * i915_gem_sw_finish_ioctl - Called when user space has done writes to this buffer
 815 * @dev: drm device
 816 * @data: ioctl data blob
 817 * @file: drm file
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 818 */
 819int
 820i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
 821			 struct drm_file *file)
 822{
 823	struct drm_i915_gem_sw_finish *args = data;
 824	struct drm_i915_gem_object *obj;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 825
 826	obj = i915_gem_object_lookup(file, args->handle);
 827	if (!obj)
 828		return -ENOENT;
 829
 830	/*
 831	 * Proxy objects are barred from CPU access, so there is no
 832	 * need to ban sw_finish as it is a nop.
 833	 */
 
 
 
 
 834
 835	/* Pinned buffers may be scanout, so flush the cache */
 836	i915_gem_object_flush_if_display(obj);
 837	i915_gem_object_put(obj);
 
 
 
 
 
 838
 839	return 0;
 840}
 841
 842void i915_gem_runtime_suspend(struct drm_i915_private *i915)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 843{
 844	struct drm_i915_gem_object *obj, *on;
 845	int i;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 846
 847	/*
 848	 * Only called during RPM suspend. All users of the userfault_list
 849	 * must be holding an RPM wakeref to ensure that this can not
 850	 * run concurrently with themselves (and use the struct_mutex for
 851	 * protection between themselves).
 852	 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 853
 854	list_for_each_entry_safe(obj, on,
 855				 &to_gt(i915)->ggtt->userfault_list, userfault_link)
 856		__i915_gem_object_release_mmap_gtt(obj);
 857
 858	list_for_each_entry_safe(obj, on,
 859				 &i915->runtime_pm.lmem_userfault_list, userfault_link)
 860		i915_gem_object_runtime_pm_release_mmap_offset(obj);
 861
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 862	/*
 863	 * The fence will be lost when the device powers down. If any were
 864	 * in use by hardware (i.e. they are pinned), we should not be powering
 865	 * down! All other fences will be reacquired by the user upon waking.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 866	 */
 867	for (i = 0; i < to_gt(i915)->ggtt->num_fences; i++) {
 868		struct i915_fence_reg *reg = &to_gt(i915)->ggtt->fence_regs[i];
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 869
 870		/*
 871		 * Ideally we want to assert that the fence register is not
 872		 * live at this point (i.e. that no piece of code will be
 873		 * trying to write through fence + GTT, as that both violates
 874		 * our tracking of activity and associated locking/barriers,
 875		 * but also is illegal given that the hw is powered down).
 876		 *
 877		 * Previously we used reg->pin_count as a "liveness" indicator.
 878		 * That is not sufficient, and we need a more fine-grained
 879		 * tool if we want to have a sanity check here.
 
 
 
 
 
 
 
 
 
 880		 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 881
 882		if (!reg->vma)
 
 
 
 
 
 
 
 883			continue;
 
 
 
 
 
 
 
 
 
 
 
 884
 885		GEM_BUG_ON(i915_vma_has_userfault(reg->vma));
 886		reg->dirty = true;
 887	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 888}
 889
 890static void discard_ggtt_vma(struct i915_vma *vma)
 
 
 
 
 
 
 
 
 891{
 892	struct drm_i915_gem_object *obj = vma->obj;
 
 
 893
 894	spin_lock(&obj->vma.lock);
 895	if (!RB_EMPTY_NODE(&vma->obj_node)) {
 896		rb_erase(&vma->obj_node, &obj->vma.tree);
 897		RB_CLEAR_NODE(&vma->obj_node);
 898	}
 899	spin_unlock(&obj->vma.lock);
 900}
 901
 902struct i915_vma *
 903i915_gem_object_ggtt_pin_ww(struct drm_i915_gem_object *obj,
 904			    struct i915_gem_ww_ctx *ww,
 905			    const struct i915_gtt_view *view,
 906			    u64 size, u64 alignment, u64 flags)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 907{
 908	struct drm_i915_private *i915 = to_i915(obj->base.dev);
 909	struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
 910	struct i915_vma *vma;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 911	int ret;
 912
 913	GEM_WARN_ON(!ww);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 914
 915	if (flags & PIN_MAPPABLE &&
 916	    (!view || view->type == I915_GTT_VIEW_NORMAL)) {
 917		/*
 918		 * If the required space is larger than the available
 919		 * aperture, we will not able to find a slot for the
 920		 * object and unbinding the object now will be in
 921		 * vain. Worse, doing so may cause us to ping-pong
 922		 * the object in and out of the Global GTT and
 923		 * waste a lot of cycles under the mutex.
 924		 */
 925		if (obj->base.size > ggtt->mappable_end)
 926			return ERR_PTR(-E2BIG);
 
 
 
 
 
 
 927
 928		/*
 929		 * If NONBLOCK is set the caller is optimistically
 930		 * trying to cache the full object within the mappable
 931		 * aperture, and *must* have a fallback in place for
 932		 * situations where we cannot bind the object. We
 933		 * can be a little more lax here and use the fallback
 934		 * more often to avoid costly migrations of ourselves
 935		 * and other objects within the aperture.
 936		 *
 937		 * Half-the-aperture is used as a simple heuristic.
 938		 * More interesting would to do search for a free
 939		 * block prior to making the commitment to unbind.
 940		 * That caters for the self-harm case, and with a
 941		 * little more heuristics (e.g. NOFAULT, NOEVICT)
 942		 * we could try to minimise harm to others.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 943		 */
 944		if (flags & PIN_NONBLOCK &&
 945		    obj->base.size > ggtt->mappable_end / 2)
 946			return ERR_PTR(-ENOSPC);
 947	}
 948
 949new_vma:
 950	vma = i915_vma_instance(obj, &ggtt->vm, view);
 951	if (IS_ERR(vma))
 952		return vma;
 
 953
 954	if (i915_vma_misplaced(vma, size, alignment, flags)) {
 955		if (flags & PIN_NONBLOCK) {
 956			if (i915_vma_is_pinned(vma) || i915_vma_is_active(vma))
 957				return ERR_PTR(-ENOSPC);
 958
 959			/*
 960			 * If this misplaced vma is too big (i.e, at-least
 961			 * half the size of aperture) or hasn't been pinned
 962			 * mappable before, we ignore the misplacement when
 963			 * PIN_NONBLOCK is set in order to avoid the ping-pong
 964			 * issue described above. In other words, we try to
 965			 * avoid the costly operation of unbinding this vma
 966			 * from the GGTT and rebinding it back because there
 967			 * may not be enough space for this vma in the aperture.
 968			 */
 969			if (flags & PIN_MAPPABLE &&
 970			    (vma->fence_size > ggtt->mappable_end / 2 ||
 971			    !i915_vma_is_map_and_fenceable(vma)))
 972				return ERR_PTR(-ENOSPC);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 973		}
 
 
 
 
 
 
 
 
 
 
 
 974
 975		if (i915_vma_is_pinned(vma) || i915_vma_is_active(vma)) {
 976			discard_ggtt_vma(vma);
 977			goto new_vma;
 978		}
 
 
 979
 980		ret = i915_vma_unbind(vma);
 981		if (ret)
 982			return ERR_PTR(ret);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 983	}
 984
 985	ret = i915_vma_pin_ww(vma, ww, size, alignment, flags | PIN_GLOBAL);
 986
 
 
 
 
 
 
 
 
 
 
 
 
 987	if (ret)
 988		return ERR_PTR(ret);
 989
 990	if (vma->fence && !i915_gem_object_is_tiled(obj)) {
 991		mutex_lock(&ggtt->vm.mutex);
 992		i915_vma_revoke_fence(vma);
 993		mutex_unlock(&ggtt->vm.mutex);
 994	}
 995
 996	ret = i915_vma_wait_for_bind(vma);
 
 
 
 
 
 
 
 
 
 
 997	if (ret) {
 998		i915_vma_unpin(vma);
 999		return ERR_PTR(ret);
 
 
 
 
 
 
 
 
 
 
 
1000	}
1001
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1002	return vma;
1003}
1004
1005struct i915_vma * __must_check
1006i915_gem_object_ggtt_pin(struct drm_i915_gem_object *obj,
1007			 const struct i915_gtt_view *view,
1008			 u64 size, u64 alignment, u64 flags)
1009{
1010	struct i915_gem_ww_ctx ww;
1011	struct i915_vma *ret;
1012	int err;
1013
1014	for_i915_gem_ww(&ww, err, true) {
1015		err = i915_gem_object_lock(obj, &ww);
1016		if (err)
1017			continue;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1018
1019		ret = i915_gem_object_ggtt_pin_ww(obj, &ww, view, size,
1020						  alignment, flags);
1021		if (IS_ERR(ret))
1022			err = PTR_ERR(ret);
1023	}
 
 
 
 
 
 
 
 
 
1024
1025	return err ? ERR_PTR(err) : ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1026}
1027
1028int
1029i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
1030		       struct drm_file *file_priv)
1031{
1032	struct drm_i915_private *i915 = to_i915(dev);
1033	struct drm_i915_gem_madvise *args = data;
1034	struct drm_i915_gem_object *obj;
1035	int err;
1036
1037	switch (args->madv) {
1038	case I915_MADV_DONTNEED:
1039	case I915_MADV_WILLNEED:
1040	    break;
1041	default:
1042	    return -EINVAL;
1043	}
1044
1045	obj = i915_gem_object_lookup(file_priv, args->handle);
1046	if (!obj)
1047		return -ENOENT;
1048
1049	err = i915_gem_object_lock_interruptible(obj, NULL);
1050	if (err)
 
 
 
 
 
 
1051		goto out;
 
1052
1053	if (i915_gem_object_has_pages(obj) &&
1054	    i915_gem_object_is_tiled(obj) &&
1055	    i915->gem_quirks & GEM_QUIRK_PIN_SWIZZLED_PAGES) {
1056		if (obj->mm.madv == I915_MADV_WILLNEED) {
1057			GEM_BUG_ON(!i915_gem_object_has_tiling_quirk(obj));
1058			i915_gem_object_clear_tiling_quirk(obj);
1059			i915_gem_object_make_shrinkable(obj);
1060		}
1061		if (args->madv == I915_MADV_WILLNEED) {
1062			GEM_BUG_ON(i915_gem_object_has_tiling_quirk(obj));
1063			i915_gem_object_make_unshrinkable(obj);
1064			i915_gem_object_set_tiling_quirk(obj);
1065		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1066	}
1067
1068	if (obj->mm.madv != __I915_MADV_PURGED) {
1069		obj->mm.madv = args->madv;
1070		if (obj->ops->adjust_lru)
1071			obj->ops->adjust_lru(obj);
 
1072	}
1073
1074	if (i915_gem_object_has_pages(obj) ||
1075	    i915_gem_object_has_self_managed_shrink_list(obj)) {
1076		unsigned long flags;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1077
1078		spin_lock_irqsave(&i915->mm.obj_lock, flags);
1079		if (!list_empty(&obj->mm.link)) {
1080			struct list_head *list;
1081
1082			if (obj->mm.madv != I915_MADV_WILLNEED)
1083				list = &i915->mm.purge_list;
1084			else
1085				list = &i915->mm.shrink_list;
1086			list_move_tail(&obj->mm.link, list);
1087
 
 
 
 
 
 
 
 
 
 
 
1088		}
1089		spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
1090	}
1091
1092	/* if the object is no longer attached, discard its backing storage */
1093	if (obj->mm.madv == I915_MADV_DONTNEED &&
1094	    !i915_gem_object_has_pages(obj))
1095		i915_gem_object_truncate(obj);
1096
1097	args->retained = obj->mm.madv != __I915_MADV_PURGED;
 
 
 
1098
1099	i915_gem_object_unlock(obj);
1100out:
1101	i915_gem_object_put(obj);
1102	return err;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1103}
1104
1105/*
1106 * A single pass should suffice to release all the freed objects (along most
1107 * call paths), but be a little more paranoid in that freeing the objects does
1108 * take a little amount of time, during which the rcu callbacks could have added
1109 * new objects into the freed list, and armed the work again.
1110 */
1111void i915_gem_drain_freed_objects(struct drm_i915_private *i915)
1112{
1113	while (atomic_read(&i915->mm.free_count)) {
1114		flush_work(&i915->mm.free_work);
1115		drain_workqueue(i915->bdev.wq);
1116		rcu_barrier();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1117	}
 
 
 
 
1118}
1119
1120/*
1121 * Similar to objects above (see i915_gem_drain_freed-objects), in general we
1122 * have workers that are armed by RCU and then rearm themselves in their
1123 * callbacks. To be paranoid, we need to drain the workqueue a second time after
1124 * waiting for the RCU grace period so that we catch work queued via RCU from
1125 * the first pass. As neither drain_workqueue() nor flush_workqueue() report a
1126 * result, we make an assumption that we only don't require more than 3 passes
1127 * to catch all _recursive_ RCU delayed work.
1128 */
1129void i915_gem_drain_workqueue(struct drm_i915_private *i915)
1130{
1131	int i;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1132
1133	for (i = 0; i < 3; i++) {
1134		flush_workqueue(i915->wq);
1135		rcu_barrier();
1136		i915_gem_drain_freed_objects(i915);
 
 
 
 
 
 
 
1137	}
1138
1139	drain_workqueue(i915->wq);
1140}
1141
1142int i915_gem_init(struct drm_i915_private *dev_priv)
1143{
1144	struct intel_gt *gt;
1145	unsigned int i;
1146	int ret;
1147
1148	/*
1149	 * In the proccess of replacing cache_level with pat_index a tricky
1150	 * dependency is created on the definition of the enum i915_cache_level.
1151	 * in case this enum is changed, PTE encode would be broken.
1152	 * Add a WARNING here. And remove when we completely quit using this
1153	 * enum
1154	 */
1155	BUILD_BUG_ON(I915_CACHE_NONE != 0 ||
1156		     I915_CACHE_LLC != 1 ||
1157		     I915_CACHE_L3_LLC != 2 ||
1158		     I915_CACHE_WT != 3 ||
1159		     I915_MAX_CACHE_LEVEL != 4);
1160
1161	/* We need to fallback to 4K pages if host doesn't support huge gtt. */
1162	if (intel_vgpu_active(dev_priv) && !intel_vgpu_has_huge_gtt(dev_priv))
1163		RUNTIME_INFO(dev_priv)->page_sizes = I915_GTT_PAGE_SIZE_4K;
1164
1165	for_each_gt(gt, dev_priv, i) {
1166		intel_uc_fetch_firmwares(&gt->uc);
1167		intel_wopcm_init(&gt->wopcm);
1168		if (GRAPHICS_VER(dev_priv) >= 8)
1169			setup_private_pat(gt);
1170	}
1171
1172	ret = i915_init_ggtt(dev_priv);
1173	if (ret) {
1174		GEM_BUG_ON(ret == -EIO);
1175		goto err_unlock;
1176	}
1177
1178	/*
1179	 * Despite its name intel_clock_gating_init applies both display
1180	 * clock gating workarounds; GT mmio workarounds and the occasional
1181	 * GT power context workaround. Worse, sometimes it includes a context
1182	 * register workaround which we need to apply before we record the
1183	 * default HW state for all contexts.
1184	 *
1185	 * FIXME: break up the workarounds and apply them at the right time!
1186	 */
1187	intel_clock_gating_init(dev_priv);
1188
1189	for_each_gt(gt, dev_priv, i) {
1190		ret = intel_gt_init(gt);
1191		if (ret)
1192			goto err_unlock;
1193	}
1194
1195	/*
1196	 * Register engines early to ensure the engine list is in its final
1197	 * rb-tree form, lowering the amount of code that has to deal with
1198	 * the intermediate llist state.
1199	 */
1200	intel_engines_driver_register(dev_priv);
1201
1202	return 0;
1203
1204	/*
1205	 * Unwinding is complicated by that we want to handle -EIO to mean
1206	 * disable GPU submission but keep KMS alive. We want to mark the
1207	 * HW as irrevisibly wedged, but keep enough state around that the
1208	 * driver doesn't explode during runtime.
1209	 */
1210err_unlock:
1211	i915_gem_drain_workqueue(dev_priv);
 
 
 
 
 
 
 
 
 
1212
1213	if (ret != -EIO) {
1214		for_each_gt(gt, dev_priv, i) {
1215			intel_gt_driver_remove(gt);
1216			intel_gt_driver_release(gt);
1217			intel_uc_cleanup_firmwares(&gt->uc);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1218		}
1219	}
1220
1221	if (ret == -EIO) {
1222		/*
1223		 * Allow engines or uC initialisation to fail by marking the GPU
1224		 * as wedged. But we only want to do this when the GPU is angry,
1225		 * for all other failure, such as an allocation failure, bail.
1226		 */
1227		for_each_gt(gt, dev_priv, i) {
1228			if (!intel_gt_is_wedged(gt)) {
1229				i915_probe_error(dev_priv,
1230						 "Failed to initialize GPU, declaring it wedged!\n");
1231				intel_gt_set_wedged(gt);
1232			}
1233		}
1234
1235		/* Minimal basic recovery for KMS */
1236		ret = i915_ggtt_enable_hw(dev_priv);
1237		i915_ggtt_resume(to_gt(dev_priv)->ggtt);
1238		intel_clock_gating_init(dev_priv);
 
 
 
 
 
 
 
 
 
 
1239	}
1240
1241	i915_gem_drain_freed_objects(dev_priv);
1242
 
 
1243	return ret;
1244}
1245
1246void i915_gem_driver_register(struct drm_i915_private *i915)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1247{
1248	i915_gem_driver_register__shrinker(i915);
 
 
 
 
 
1249}
1250
1251void i915_gem_driver_unregister(struct drm_i915_private *i915)
 
 
1252{
1253	i915_gem_driver_unregister__shrinker(i915);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1254}
1255
1256void i915_gem_driver_remove(struct drm_i915_private *dev_priv)
 
 
1257{
1258	struct intel_gt *gt;
1259	unsigned int i;
1260
1261	i915_gem_suspend_late(dev_priv);
1262	for_each_gt(gt, dev_priv, i)
1263		intel_gt_driver_remove(gt);
1264	dev_priv->uabi_engines = RB_ROOT;
1265
1266	/* Flush any outstanding unpin_work. */
1267	i915_gem_drain_workqueue(dev_priv);
1268}
1269
1270void i915_gem_driver_release(struct drm_i915_private *dev_priv)
 
1271{
1272	struct intel_gt *gt;
1273	unsigned int i;
1274
1275	for_each_gt(gt, dev_priv, i) {
1276		intel_gt_driver_release(gt);
1277		intel_uc_cleanup_firmwares(&gt->uc);
1278	}
1279
1280	/* Flush any outstanding work, including i915_gem_context.release_work. */
1281	i915_gem_drain_workqueue(dev_priv);
 
 
 
 
 
 
 
 
 
1282
1283	drm_WARN_ON(&dev_priv->drm, !list_empty(&dev_priv->gem.contexts.list));
 
 
 
 
 
 
 
 
 
1284}
1285
1286static void i915_gem_init__mm(struct drm_i915_private *i915)
 
1287{
1288	spin_lock_init(&i915->mm.obj_lock);
 
1289
1290	init_llist_head(&i915->mm.free_list);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1291
1292	INIT_LIST_HEAD(&i915->mm.purge_list);
1293	INIT_LIST_HEAD(&i915->mm.shrink_list);
1294
1295	i915_gem_init__objects(i915);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1296}
1297
1298void i915_gem_init_early(struct drm_i915_private *dev_priv)
1299{
1300	i915_gem_init__mm(dev_priv);
1301	i915_gem_init__contexts(dev_priv);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1302}
1303
1304void i915_gem_cleanup_early(struct drm_i915_private *dev_priv)
 
1305{
1306	i915_gem_drain_workqueue(dev_priv);
1307	GEM_BUG_ON(!llist_empty(&dev_priv->mm.free_list));
1308	GEM_BUG_ON(atomic_read(&dev_priv->mm.free_count));
1309	drm_WARN_ON(&dev_priv->drm, dev_priv->mm.shrink_count);
1310}
1311
1312int i915_gem_open(struct drm_i915_private *i915, struct drm_file *file)
1313{
1314	struct drm_i915_file_private *file_priv;
1315	struct i915_drm_client *client;
1316	int ret = -ENOMEM;
1317
1318	drm_dbg(&i915->drm, "\n");
1319
1320	file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL);
1321	if (!file_priv)
1322		goto err_alloc;
1323
1324	client = i915_drm_client_alloc();
1325	if (!client)
1326		goto err_client;
1327
1328	file->driver_priv = file_priv;
1329	file_priv->i915 = i915;
1330	file_priv->file = file;
1331	file_priv->client = client;
1332
1333	file_priv->bsd_engine = -1;
1334	file_priv->hang_timestamp = jiffies;
 
 
1335
1336	ret = i915_gem_context_open(i915, file);
1337	if (ret)
1338		goto err_context;
1339
1340	return 0;
1341
1342err_context:
1343	i915_drm_client_put(client);
1344err_client:
1345	kfree(file_priv);
1346err_alloc:
1347	return ret;
1348}
1349
1350#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1351#include "selftests/mock_gem_device.c"
1352#include "selftests/i915_gem.c"
 
 
 
 
 
 
 
1353#endif
v3.15
   1/*
   2 * Copyright © 2008 Intel Corporation
   3 *
   4 * Permission is hereby granted, free of charge, to any person obtaining a
   5 * copy of this software and associated documentation files (the "Software"),
   6 * to deal in the Software without restriction, including without limitation
   7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8 * and/or sell copies of the Software, and to permit persons to whom the
   9 * Software is furnished to do so, subject to the following conditions:
  10 *
  11 * The above copyright notice and this permission notice (including the next
  12 * paragraph) shall be included in all copies or substantial portions of the
  13 * Software.
  14 *
  15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21 * IN THE SOFTWARE.
  22 *
  23 * Authors:
  24 *    Eric Anholt <eric@anholt.net>
  25 *
  26 */
  27
  28#include <drm/drmP.h>
  29#include <drm/drm_vma_manager.h>
  30#include <drm/i915_drm.h>
  31#include "i915_drv.h"
  32#include "i915_trace.h"
  33#include "intel_drv.h"
  34#include <linux/shmem_fs.h>
  35#include <linux/slab.h>
 
  36#include <linux/swap.h>
  37#include <linux/pci.h>
  38#include <linux/dma-buf.h>
 
  39
  40static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);
  41static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj,
  42						   bool force);
  43static __must_check int
  44i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
  45			       bool readonly);
  46
  47static void i915_gem_write_fence(struct drm_device *dev, int reg,
  48				 struct drm_i915_gem_object *obj);
  49static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
  50					 struct drm_i915_fence_reg *fence,
  51					 bool enable);
  52
  53static unsigned long i915_gem_inactive_count(struct shrinker *shrinker,
  54					     struct shrink_control *sc);
  55static unsigned long i915_gem_inactive_scan(struct shrinker *shrinker,
  56					    struct shrink_control *sc);
  57static unsigned long i915_gem_purge(struct drm_i915_private *dev_priv, long target);
  58static unsigned long i915_gem_shrink_all(struct drm_i915_private *dev_priv);
  59static void i915_gem_object_truncate(struct drm_i915_gem_object *obj);
  60static void i915_gem_retire_requests_ring(struct intel_ring_buffer *ring);
  61
  62static bool cpu_cache_is_coherent(struct drm_device *dev,
  63				  enum i915_cache_level level)
  64{
  65	return HAS_LLC(dev) || level != I915_CACHE_NONE;
  66}
 
 
 
 
 
 
  67
  68static bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj)
  69{
  70	if (!cpu_cache_is_coherent(obj->base.dev, obj->cache_level))
  71		return true;
  72
  73	return obj->pin_display;
  74}
  75
  76static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object *obj)
  77{
  78	if (obj->tiling_mode)
  79		i915_gem_release_mmap(obj);
  80
  81	/* As we do not have an associated fence register, we will force
  82	 * a tiling change if we ever need to acquire one.
  83	 */
  84	obj->fence_dirty = false;
  85	obj->fence_reg = I915_FENCE_REG_NONE;
  86}
  87
  88/* some bookkeeping */
  89static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
  90				  size_t size)
  91{
  92	spin_lock(&dev_priv->mm.object_stat_lock);
  93	dev_priv->mm.object_count++;
  94	dev_priv->mm.object_memory += size;
  95	spin_unlock(&dev_priv->mm.object_stat_lock);
  96}
  97
  98static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
  99				     size_t size)
 100{
 101	spin_lock(&dev_priv->mm.object_stat_lock);
 102	dev_priv->mm.object_count--;
 103	dev_priv->mm.object_memory -= size;
 104	spin_unlock(&dev_priv->mm.object_stat_lock);
 105}
 106
 107static int
 108i915_gem_wait_for_error(struct i915_gpu_error *error)
 109{
 110	int ret;
 111
 112#define EXIT_COND (!i915_reset_in_progress(error) || \
 113		   i915_terminally_wedged(error))
 114	if (EXIT_COND)
 115		return 0;
 116
 117	/*
 118	 * Only wait 10 seconds for the gpu reset to complete to avoid hanging
 119	 * userspace. If it takes that long something really bad is going on and
 120	 * we should simply try to bail out and fail as gracefully as possible.
 121	 */
 122	ret = wait_event_interruptible_timeout(error->reset_queue,
 123					       EXIT_COND,
 124					       10*HZ);
 125	if (ret == 0) {
 126		DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
 127		return -EIO;
 128	} else if (ret < 0) {
 129		return ret;
 130	}
 131#undef EXIT_COND
 132
 133	return 0;
 134}
 135
 136int i915_mutex_lock_interruptible(struct drm_device *dev)
 137{
 138	struct drm_i915_private *dev_priv = dev->dev_private;
 139	int ret;
 140
 141	ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
 142	if (ret)
 143		return ret;
 144
 145	ret = mutex_lock_interruptible(&dev->struct_mutex);
 146	if (ret)
 147		return ret;
 
 
 148
 149	WARN_ON(i915_verify_lists(dev));
 150	return 0;
 151}
 152
 153static inline bool
 154i915_gem_object_is_inactive(struct drm_i915_gem_object *obj)
 155{
 156	return i915_gem_obj_bound_any(obj) && !obj->active;
 157}
 158
 159int
 160i915_gem_init_ioctl(struct drm_device *dev, void *data,
 161		    struct drm_file *file)
 162{
 163	struct drm_i915_private *dev_priv = dev->dev_private;
 164	struct drm_i915_gem_init *args = data;
 165
 166	if (drm_core_check_feature(dev, DRIVER_MODESET))
 167		return -ENODEV;
 168
 169	if (args->gtt_start >= args->gtt_end ||
 170	    (args->gtt_end | args->gtt_start) & (PAGE_SIZE - 1))
 171		return -EINVAL;
 172
 173	/* GEM with user mode setting was never supported on ilk and later. */
 174	if (INTEL_INFO(dev)->gen >= 5)
 175		return -ENODEV;
 176
 177	mutex_lock(&dev->struct_mutex);
 178	i915_gem_setup_global_gtt(dev, args->gtt_start, args->gtt_end,
 179				  args->gtt_end);
 180	dev_priv->gtt.mappable_end = args->gtt_end;
 181	mutex_unlock(&dev->struct_mutex);
 182
 183	return 0;
 184}
 185
 186int
 187i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
 188			    struct drm_file *file)
 189{
 190	struct drm_i915_private *dev_priv = dev->dev_private;
 
 191	struct drm_i915_gem_get_aperture *args = data;
 192	struct drm_i915_gem_object *obj;
 193	size_t pinned;
 
 
 
 
 
 
 
 
 194
 195	pinned = 0;
 196	mutex_lock(&dev->struct_mutex);
 197	list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list)
 198		if (i915_gem_obj_is_pinned(obj))
 199			pinned += i915_gem_obj_ggtt_size(obj);
 200	mutex_unlock(&dev->struct_mutex);
 201
 202	args->aper_size = dev_priv->gtt.base.total;
 203	args->aper_available_size = args->aper_size - pinned;
 204
 205	return 0;
 206}
 207
 208static void i915_gem_object_detach_phys(struct drm_i915_gem_object *obj)
 
 209{
 210	drm_dma_handle_t *phys = obj->phys_handle;
 
 
 
 
 
 211
 212	if (!phys)
 213		return;
 214
 215	if (obj->madv == I915_MADV_WILLNEED) {
 216		struct address_space *mapping = file_inode(obj->base.filp)->i_mapping;
 217		char *vaddr = phys->vaddr;
 218		int i;
 219
 220		for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
 221			struct page *page = shmem_read_mapping_page(mapping, i);
 222			if (!IS_ERR(page)) {
 223				char *dst = kmap_atomic(page);
 224				memcpy(dst, vaddr, PAGE_SIZE);
 225				drm_clflush_virt_range(dst, PAGE_SIZE);
 226				kunmap_atomic(dst);
 227
 228				set_page_dirty(page);
 229				mark_page_accessed(page);
 230				page_cache_release(page);
 231			}
 232			vaddr += PAGE_SIZE;
 233		}
 234		i915_gem_chipset_flush(obj->base.dev);
 235	}
 236
 237#ifdef CONFIG_X86
 238	set_memory_wb((unsigned long)phys->vaddr, phys->size / PAGE_SIZE);
 239#endif
 240	drm_pci_free(obj->base.dev, phys);
 241	obj->phys_handle = NULL;
 242}
 
 243
 244int
 245i915_gem_object_attach_phys(struct drm_i915_gem_object *obj,
 246			    int align)
 247{
 248	drm_dma_handle_t *phys;
 249	struct address_space *mapping;
 250	char *vaddr;
 251	int i;
 
 252
 253	if (obj->phys_handle) {
 254		if ((unsigned long)obj->phys_handle->vaddr & (align -1))
 255			return -EBUSY;
 
 256
 257		return 0;
 258	}
 
 
 
 
 
 
 
 
 259
 260	if (obj->madv != I915_MADV_WILLNEED)
 261		return -EFAULT;
 262
 263	if (obj->base.filp == NULL)
 264		return -EINVAL;
 
 
 
 265
 266	/* create a new object */
 267	phys = drm_pci_alloc(obj->base.dev, obj->base.size, align);
 268	if (!phys)
 269		return -ENOMEM;
 270
 271	vaddr = phys->vaddr;
 272#ifdef CONFIG_X86
 273	set_memory_wc((unsigned long)vaddr, phys->size / PAGE_SIZE);
 274#endif
 275	mapping = file_inode(obj->base.filp)->i_mapping;
 276	for (i = 0; i < obj->base.size / PAGE_SIZE; i++) {
 277		struct page *page;
 278		char *src;
 279
 280		page = shmem_read_mapping_page(mapping, i);
 281		if (IS_ERR(page)) {
 282#ifdef CONFIG_X86
 283			set_memory_wb((unsigned long)phys->vaddr, phys->size / PAGE_SIZE);
 284#endif
 285			drm_pci_free(obj->base.dev, phys);
 286			return PTR_ERR(page);
 287		}
 288
 289		src = kmap_atomic(page);
 290		memcpy(vaddr, src, PAGE_SIZE);
 291		kunmap_atomic(src);
 
 
 
 
 
 
 
 
 292
 293		mark_page_accessed(page);
 294		page_cache_release(page);
 
 
 
 295
 296		vaddr += PAGE_SIZE;
 
 
 297	}
 298
 299	obj->phys_handle = phys;
 300	return 0;
 
 301}
 302
 303static int
 304i915_gem_phys_pwrite(struct drm_i915_gem_object *obj,
 305		     struct drm_i915_gem_pwrite *args,
 306		     struct drm_file *file_priv)
 307{
 308	struct drm_device *dev = obj->base.dev;
 309	void *vaddr = obj->phys_handle->vaddr + args->offset;
 310	char __user *user_data = to_user_ptr(args->data_ptr);
 311
 312	if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
 313		unsigned long unwritten;
 314
 315		/* The physical object once assigned is fixed for the lifetime
 316		 * of the obj, so we can safely drop the lock and continue
 317		 * to access vaddr.
 318		 */
 319		mutex_unlock(&dev->struct_mutex);
 320		unwritten = copy_from_user(vaddr, user_data, args->size);
 321		mutex_lock(&dev->struct_mutex);
 322		if (unwritten)
 323			return -EFAULT;
 324	}
 325
 326	i915_gem_chipset_flush(dev);
 327	return 0;
 328}
 329
 330void *i915_gem_object_alloc(struct drm_device *dev)
 331{
 332	struct drm_i915_private *dev_priv = dev->dev_private;
 333	return kmem_cache_zalloc(dev_priv->slab, GFP_KERNEL);
 334}
 335
 336void i915_gem_object_free(struct drm_i915_gem_object *obj)
 337{
 338	struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
 339	kmem_cache_free(dev_priv->slab, obj);
 340}
 341
 342static int
 343i915_gem_create(struct drm_file *file,
 344		struct drm_device *dev,
 345		uint64_t size,
 346		uint32_t *handle_p)
 347{
 348	struct drm_i915_gem_object *obj;
 
 
 
 
 349	int ret;
 350	u32 handle;
 351
 352	size = roundup(size, PAGE_SIZE);
 353	if (size == 0)
 354		return -EINVAL;
 355
 356	/* Allocate the new object */
 357	obj = i915_gem_alloc_object(dev, size);
 358	if (obj == NULL)
 359		return -ENOMEM;
 360
 361	ret = drm_gem_handle_create(file, &obj->base, &handle);
 362	/* drop reference from allocate - handle holds it now */
 363	drm_gem_object_unreference_unlocked(&obj->base);
 364	if (ret)
 365		return ret;
 366
 367	*handle_p = handle;
 368	return 0;
 369}
 370
 371int
 372i915_gem_dumb_create(struct drm_file *file,
 373		     struct drm_device *dev,
 374		     struct drm_mode_create_dumb *args)
 375{
 376	/* have to work out size/pitch and return them */
 377	args->pitch = ALIGN(args->width * DIV_ROUND_UP(args->bpp, 8), 64);
 378	args->size = args->pitch * args->height;
 379	return i915_gem_create(file, dev,
 380			       args->size, &args->handle);
 381}
 382
 383/**
 384 * Creates a new mm object and returns a handle to it.
 385 */
 386int
 387i915_gem_create_ioctl(struct drm_device *dev, void *data,
 388		      struct drm_file *file)
 389{
 390	struct drm_i915_gem_create *args = data;
 391
 392	return i915_gem_create(file, dev,
 393			       args->size, &args->handle);
 394}
 395
 396static inline int
 397__copy_to_user_swizzled(char __user *cpu_vaddr,
 398			const char *gpu_vaddr, int gpu_offset,
 399			int length)
 400{
 401	int ret, cpu_offset = 0;
 402
 403	while (length > 0) {
 404		int cacheline_end = ALIGN(gpu_offset + 1, 64);
 405		int this_length = min(cacheline_end - gpu_offset, length);
 406		int swizzled_gpu_offset = gpu_offset ^ 64;
 407
 408		ret = __copy_to_user(cpu_vaddr + cpu_offset,
 409				     gpu_vaddr + swizzled_gpu_offset,
 410				     this_length);
 411		if (ret)
 412			return ret + length;
 413
 414		cpu_offset += this_length;
 415		gpu_offset += this_length;
 416		length -= this_length;
 417	}
 418
 419	return 0;
 
 
 
 
 
 
 
 420}
 421
 422static inline int
 423__copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset,
 424			  const char __user *cpu_vaddr,
 425			  int length)
 426{
 427	int ret, cpu_offset = 0;
 
 428
 429	while (length > 0) {
 430		int cacheline_end = ALIGN(gpu_offset + 1, 64);
 431		int this_length = min(cacheline_end - gpu_offset, length);
 432		int swizzled_gpu_offset = gpu_offset ^ 64;
 433
 434		ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset,
 435				       cpu_vaddr + cpu_offset,
 436				       this_length);
 437		if (ret)
 438			return ret + length;
 439
 440		cpu_offset += this_length;
 441		gpu_offset += this_length;
 442		length -= this_length;
 443	}
 444
 445	return 0;
 446}
 447
 448/*
 449 * Pins the specified object's pages and synchronizes the object with
 450 * GPU accesses. Sets needs_clflush to non-zero if the caller should
 451 * flush the object from the CPU cache.
 452 */
 453int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj,
 454				    int *needs_clflush)
 455{
 
 
 
 
 456	int ret;
 457
 458	*needs_clflush = 0;
 
 
 
 
 
 459
 460	if (!obj->base.filp)
 461		return -EINVAL;
 
 462
 463	if (!(obj->base.read_domains & I915_GEM_DOMAIN_CPU)) {
 464		/* If we're not in the cpu read domain, set ourself into the gtt
 465		 * read domain and manually flush cachelines (if required). This
 466		 * optimizes for the case when the gpu will dirty the data
 467		 * anyway again before the next pread happens. */
 468		*needs_clflush = !cpu_cache_is_coherent(obj->base.dev,
 469							obj->cache_level);
 470		ret = i915_gem_object_wait_rendering(obj, true);
 
 
 
 
 
 471		if (ret)
 472			return ret;
 
 
 473	}
 474
 475	ret = i915_gem_object_get_pages(obj);
 476	if (ret)
 477		return ret;
 
 
 
 
 
 
 478
 479	i915_gem_object_pin_pages(obj);
 
 
 
 
 
 
 480
 481	return ret;
 482}
 483
 484/* Per-page copy function for the shmem pread fastpath.
 485 * Flushes invalid cachelines before reading the target if
 486 * needs_clflush is set. */
 487static int
 488shmem_pread_fast(struct page *page, int shmem_page_offset, int page_length,
 489		 char __user *user_data,
 490		 bool page_do_bit17_swizzling, bool needs_clflush)
 491{
 492	char *vaddr;
 493	int ret;
 494
 495	if (unlikely(page_do_bit17_swizzling))
 496		return -EINVAL;
 497
 498	vaddr = kmap_atomic(page);
 499	if (needs_clflush)
 500		drm_clflush_virt_range(vaddr + shmem_page_offset,
 501				       page_length);
 502	ret = __copy_to_user_inatomic(user_data,
 503				      vaddr + shmem_page_offset,
 504				      page_length);
 505	kunmap_atomic(vaddr);
 506
 507	return ret ? -EFAULT : 0;
 508}
 509
 510static void
 511shmem_clflush_swizzled_range(char *addr, unsigned long length,
 512			     bool swizzled)
 513{
 514	if (unlikely(swizzled)) {
 515		unsigned long start = (unsigned long) addr;
 516		unsigned long end = (unsigned long) addr + length;
 517
 518		/* For swizzling simply ensure that we always flush both
 519		 * channels. Lame, but simple and it works. Swizzled
 520		 * pwrite/pread is far from a hotpath - current userspace
 521		 * doesn't use it at all. */
 522		start = round_down(start, 128);
 523		end = round_up(end, 128);
 524
 525		drm_clflush_virt_range((void *)start, end - start);
 526	} else {
 527		drm_clflush_virt_range(addr, length);
 528	}
 529
 530}
 531
 532/* Only difference to the fast-path function is that this can handle bit17
 533 * and uses non-atomic copy and kmap functions. */
 534static int
 535shmem_pread_slow(struct page *page, int shmem_page_offset, int page_length,
 536		 char __user *user_data,
 537		 bool page_do_bit17_swizzling, bool needs_clflush)
 538{
 539	char *vaddr;
 540	int ret;
 
 
 
 
 
 
 541
 542	vaddr = kmap(page);
 543	if (needs_clflush)
 544		shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
 545					     page_length,
 546					     page_do_bit17_swizzling);
 547
 548	if (page_do_bit17_swizzling)
 549		ret = __copy_to_user_swizzled(user_data,
 550					      vaddr, shmem_page_offset,
 551					      page_length);
 552	else
 553		ret = __copy_to_user(user_data,
 554				     vaddr + shmem_page_offset,
 555				     page_length);
 556	kunmap(page);
 557
 558	return ret ? - EFAULT : 0;
 559}
 560
 561static int
 562i915_gem_shmem_pread(struct drm_device *dev,
 563		     struct drm_i915_gem_object *obj,
 564		     struct drm_i915_gem_pread *args,
 565		     struct drm_file *file)
 566{
 567	char __user *user_data;
 568	ssize_t remain;
 569	loff_t offset;
 570	int shmem_page_offset, page_length, ret = 0;
 571	int obj_do_bit17_swizzling, page_do_bit17_swizzling;
 572	int prefaulted = 0;
 573	int needs_clflush = 0;
 574	struct sg_page_iter sg_iter;
 575
 576	user_data = to_user_ptr(args->data_ptr);
 577	remain = args->size;
 578
 579	obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
 580
 581	ret = i915_gem_obj_prepare_shmem_read(obj, &needs_clflush);
 582	if (ret)
 583		return ret;
 584
 585	offset = args->offset;
 586
 587	for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
 588			 offset >> PAGE_SHIFT) {
 589		struct page *page = sg_page_iter_page(&sg_iter);
 590
 591		if (remain <= 0)
 592			break;
 593
 594		/* Operation in this page
 595		 *
 596		 * shmem_page_offset = offset within page in shmem file
 
 597		 * page_length = bytes to copy for this page
 598		 */
 599		shmem_page_offset = offset_in_page(offset);
 600		page_length = remain;
 601		if ((shmem_page_offset + page_length) > PAGE_SIZE)
 602			page_length = PAGE_SIZE - shmem_page_offset;
 603
 604		page_do_bit17_swizzling = obj_do_bit17_swizzling &&
 605			(page_to_phys(page) & (1 << 17)) != 0;
 606
 607		ret = shmem_pread_fast(page, shmem_page_offset, page_length,
 608				       user_data, page_do_bit17_swizzling,
 609				       needs_clflush);
 610		if (ret == 0)
 611			goto next_page;
 612
 613		mutex_unlock(&dev->struct_mutex);
 614
 615		if (likely(!i915.prefault_disable) && !prefaulted) {
 616			ret = fault_in_multipages_writeable(user_data, remain);
 617			/* Userspace is tricking us, but we've already clobbered
 618			 * its pages with the prefault and promised to write the
 619			 * data up to the first fault. Hence ignore any errors
 620			 * and just continue. */
 621			(void)ret;
 622			prefaulted = 1;
 623		}
 624
 625		ret = shmem_pread_slow(page, shmem_page_offset, page_length,
 626				       user_data, page_do_bit17_swizzling,
 627				       needs_clflush);
 
 
 628
 629		mutex_lock(&dev->struct_mutex);
 630
 631		if (ret)
 632			goto out;
 633
 634next_page:
 635		remain -= page_length;
 636		user_data += page_length;
 637		offset += page_length;
 638	}
 639
 640out:
 641	i915_gem_object_unpin_pages(obj);
 642
 643	return ret;
 644}
 645
 646/**
 647 * Reads data from the object referenced by handle.
 
 
 
 648 *
 649 * On error, the contents of *data are undefined.
 650 */
 651int
 652i915_gem_pread_ioctl(struct drm_device *dev, void *data,
 653		     struct drm_file *file)
 654{
 
 655	struct drm_i915_gem_pread *args = data;
 656	struct drm_i915_gem_object *obj;
 657	int ret = 0;
 
 
 
 
 
 
 658
 659	if (args->size == 0)
 660		return 0;
 661
 662	if (!access_ok(VERIFY_WRITE,
 663		       to_user_ptr(args->data_ptr),
 664		       args->size))
 665		return -EFAULT;
 666
 667	ret = i915_mutex_lock_interruptible(dev);
 668	if (ret)
 669		return ret;
 670
 671	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
 672	if (&obj->base == NULL) {
 673		ret = -ENOENT;
 674		goto unlock;
 675	}
 676
 677	/* Bounds check source.  */
 678	if (args->offset > obj->base.size ||
 679	    args->size > obj->base.size - args->offset) {
 680		ret = -EINVAL;
 681		goto out;
 682	}
 683
 684	/* prime objects have no backing filp to GEM pread/pwrite
 685	 * pages from.
 686	 */
 687	if (!obj->base.filp) {
 688		ret = -EINVAL;
 689		goto out;
 690	}
 691
 692	trace_i915_gem_object_pread(obj, args->offset, args->size);
 
 
 
 
 693
 694	ret = i915_gem_shmem_pread(dev, obj, args, file);
 
 
 695
 696out:
 697	drm_gem_object_unreference(&obj->base);
 698unlock:
 699	mutex_unlock(&dev->struct_mutex);
 700	return ret;
 701}
 702
 703/* This is the fast write path which cannot handle
 704 * page faults in the source data
 705 */
 706
 707static inline int
 708fast_user_write(struct io_mapping *mapping,
 709		loff_t page_base, int page_offset,
 710		char __user *user_data,
 711		int length)
 712{
 713	void __iomem *vaddr_atomic;
 714	void *vaddr;
 715	unsigned long unwritten;
 716
 717	vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
 718	/* We can use the cpu mem copy function because this is X86. */
 719	vaddr = (void __force*)vaddr_atomic + page_offset;
 720	unwritten = __copy_from_user_inatomic_nocache(vaddr,
 721						      user_data, length);
 722	io_mapping_unmap_atomic(vaddr_atomic);
 
 
 
 
 
 
 
 723	return unwritten;
 724}
 725
 726/**
 727 * This is the fast pwrite path, where we copy the data directly from the
 728 * user into the GTT, uncached.
 
 
 729 */
 730static int
 731i915_gem_gtt_pwrite_fast(struct drm_device *dev,
 732			 struct drm_i915_gem_object *obj,
 733			 struct drm_i915_gem_pwrite *args,
 734			 struct drm_file *file)
 735{
 736	struct drm_i915_private *dev_priv = dev->dev_private;
 737	ssize_t remain;
 738	loff_t offset, page_base;
 739	char __user *user_data;
 740	int page_offset, page_length, ret;
 
 
 
 
 
 
 
 
 741
 742	ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE | PIN_NONBLOCK);
 743	if (ret)
 744		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 745
 746	ret = i915_gem_object_set_to_gtt_domain(obj, true);
 747	if (ret)
 748		goto out_unpin;
 
 
 749
 750	ret = i915_gem_object_put_fence(obj);
 751	if (ret)
 752		goto out_unpin;
 753
 754	user_data = to_user_ptr(args->data_ptr);
 
 755	remain = args->size;
 756
 757	offset = i915_gem_obj_ggtt_offset(obj) + args->offset;
 758
 759	while (remain > 0) {
 760		/* Operation in this page
 761		 *
 762		 * page_base = page offset within aperture
 763		 * page_offset = offset within page
 764		 * page_length = bytes to copy for this page
 765		 */
 766		page_base = offset & PAGE_MASK;
 767		page_offset = offset_in_page(offset);
 768		page_length = remain;
 769		if ((page_offset + remain) > PAGE_SIZE)
 770			page_length = PAGE_SIZE - page_offset;
 771
 
 
 
 
 
 
 
 
 
 
 
 772		/* If we get a fault while copying data, then (presumably) our
 773		 * source page isn't available.  Return the error and we'll
 774		 * retry in the slow path.
 
 
 775		 */
 776		if (fast_user_write(dev_priv->gtt.mappable, page_base,
 777				    page_offset, user_data, page_length)) {
 778			ret = -EFAULT;
 779			goto out_unpin;
 780		}
 781
 782		remain -= page_length;
 783		user_data += page_length;
 784		offset += page_length;
 785	}
 786
 787out_unpin:
 788	i915_gem_object_ggtt_unpin(obj);
 789out:
 
 
 
 790	return ret;
 791}
 792
 793/* Per-page copy function for the shmem pwrite fastpath.
 794 * Flushes invalid cachelines before writing to the target if
 795 * needs_clflush_before is set and flushes out any written cachelines after
 796 * writing if needs_clflush is set. */
 
 797static int
 798shmem_pwrite_fast(struct page *page, int shmem_page_offset, int page_length,
 799		  char __user *user_data,
 800		  bool page_do_bit17_swizzling,
 801		  bool needs_clflush_before,
 802		  bool needs_clflush_after)
 803{
 804	char *vaddr;
 805	int ret;
 806
 807	if (unlikely(page_do_bit17_swizzling))
 808		return -EINVAL;
 809
 810	vaddr = kmap_atomic(page);
 811	if (needs_clflush_before)
 812		drm_clflush_virt_range(vaddr + shmem_page_offset,
 813				       page_length);
 814	ret = __copy_from_user_inatomic(vaddr + shmem_page_offset,
 815					user_data, page_length);
 816	if (needs_clflush_after)
 817		drm_clflush_virt_range(vaddr + shmem_page_offset,
 818				       page_length);
 819	kunmap_atomic(vaddr);
 820
 821	return ret ? -EFAULT : 0;
 822}
 823
 824/* Only difference to the fast-path function is that this can handle bit17
 825 * and uses non-atomic copy and kmap functions. */
 826static int
 827shmem_pwrite_slow(struct page *page, int shmem_page_offset, int page_length,
 828		  char __user *user_data,
 829		  bool page_do_bit17_swizzling,
 830		  bool needs_clflush_before,
 831		  bool needs_clflush_after)
 832{
 833	char *vaddr;
 834	int ret;
 835
 836	vaddr = kmap(page);
 837	if (unlikely(needs_clflush_before || page_do_bit17_swizzling))
 838		shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
 839					     page_length,
 840					     page_do_bit17_swizzling);
 841	if (page_do_bit17_swizzling)
 842		ret = __copy_from_user_swizzled(vaddr, shmem_page_offset,
 843						user_data,
 844						page_length);
 845	else
 846		ret = __copy_from_user(vaddr + shmem_page_offset,
 847				       user_data,
 848				       page_length);
 849	if (needs_clflush_after)
 850		shmem_clflush_swizzled_range(vaddr + shmem_page_offset,
 851					     page_length,
 852					     page_do_bit17_swizzling);
 853	kunmap(page);
 854
 855	return ret ? -EFAULT : 0;
 856}
 857
 858static int
 859i915_gem_shmem_pwrite(struct drm_device *dev,
 860		      struct drm_i915_gem_object *obj,
 861		      struct drm_i915_gem_pwrite *args,
 862		      struct drm_file *file)
 863{
 864	ssize_t remain;
 865	loff_t offset;
 866	char __user *user_data;
 867	int shmem_page_offset, page_length, ret = 0;
 868	int obj_do_bit17_swizzling, page_do_bit17_swizzling;
 869	int hit_slowpath = 0;
 870	int needs_clflush_after = 0;
 871	int needs_clflush_before = 0;
 872	struct sg_page_iter sg_iter;
 873
 874	user_data = to_user_ptr(args->data_ptr);
 875	remain = args->size;
 876
 877	obj_do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
 878
 879	if (obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
 880		/* If we're not in the cpu write domain, set ourself into the gtt
 881		 * write domain and manually flush cachelines (if required). This
 882		 * optimizes for the case when the gpu will use the data
 883		 * right away and we therefore have to clflush anyway. */
 884		needs_clflush_after = cpu_write_needs_clflush(obj);
 885		ret = i915_gem_object_wait_rendering(obj, false);
 886		if (ret)
 887			return ret;
 888	}
 889	/* Same trick applies to invalidate partially written cachelines read
 890	 * before writing. */
 891	if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0)
 892		needs_clflush_before =
 893			!cpu_cache_is_coherent(dev, obj->cache_level);
 894
 895	ret = i915_gem_object_get_pages(obj);
 896	if (ret)
 897		return ret;
 898
 899	i915_gem_object_pin_pages(obj);
 
 
 900
 901	offset = args->offset;
 902	obj->dirty = 1;
 
 903
 904	for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
 905			 offset >> PAGE_SHIFT) {
 906		struct page *page = sg_page_iter_page(&sg_iter);
 907		int partial_cacheline_write;
 908
 909		if (remain <= 0)
 910			break;
 911
 912		/* Operation in this page
 913		 *
 914		 * shmem_page_offset = offset within page in shmem file
 915		 * page_length = bytes to copy for this page
 916		 */
 917		shmem_page_offset = offset_in_page(offset);
 918
 919		page_length = remain;
 920		if ((shmem_page_offset + page_length) > PAGE_SIZE)
 921			page_length = PAGE_SIZE - shmem_page_offset;
 922
 923		/* If we don't overwrite a cacheline completely we need to be
 924		 * careful to have up-to-date data by first clflushing. Don't
 925		 * overcomplicate things and flush the entire patch. */
 926		partial_cacheline_write = needs_clflush_before &&
 927			((shmem_page_offset | page_length)
 928				& (boot_cpu_data.x86_clflush_size - 1));
 929
 930		page_do_bit17_swizzling = obj_do_bit17_swizzling &&
 931			(page_to_phys(page) & (1 << 17)) != 0;
 932
 933		ret = shmem_pwrite_fast(page, shmem_page_offset, page_length,
 934					user_data, page_do_bit17_swizzling,
 935					partial_cacheline_write,
 936					needs_clflush_after);
 937		if (ret == 0)
 938			goto next_page;
 939
 940		hit_slowpath = 1;
 941		mutex_unlock(&dev->struct_mutex);
 942		ret = shmem_pwrite_slow(page, shmem_page_offset, page_length,
 943					user_data, page_do_bit17_swizzling,
 944					partial_cacheline_write,
 945					needs_clflush_after);
 946
 947		mutex_lock(&dev->struct_mutex);
 948
 
 
 
 
 
 
 
 
 
 
 949		if (ret)
 950			goto out;
 951
 952next_page:
 953		remain -= page_length;
 954		user_data += page_length;
 955		offset += page_length;
 956	}
 957
 958out:
 
 959	i915_gem_object_unpin_pages(obj);
 
 960
 961	if (hit_slowpath) {
 962		/*
 963		 * Fixup: Flush cpu caches in case we didn't flush the dirty
 964		 * cachelines in-line while writing and the object moved
 965		 * out of the cpu write domain while we've dropped the lock.
 966		 */
 967		if (!needs_clflush_after &&
 968		    obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
 969			if (i915_gem_clflush_object(obj, obj->pin_display))
 970				i915_gem_chipset_flush(dev);
 971		}
 972	}
 973
 974	if (needs_clflush_after)
 975		i915_gem_chipset_flush(dev);
 976
 977	return ret;
 978}
 979
 980/**
 981 * Writes data to the object referenced by handle.
 
 
 
 982 *
 983 * On error, the contents of the buffer that were to be modified are undefined.
 984 */
 985int
 986i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
 987		      struct drm_file *file)
 988{
 
 989	struct drm_i915_gem_pwrite *args = data;
 990	struct drm_i915_gem_object *obj;
 991	int ret;
 992
 
 
 
 
 
 
 993	if (args->size == 0)
 994		return 0;
 995
 996	if (!access_ok(VERIFY_READ,
 997		       to_user_ptr(args->data_ptr),
 998		       args->size))
 999		return -EFAULT;
1000
1001	if (likely(!i915.prefault_disable)) {
1002		ret = fault_in_multipages_readable(to_user_ptr(args->data_ptr),
1003						   args->size);
1004		if (ret)
1005			return -EFAULT;
1006	}
1007
1008	ret = i915_mutex_lock_interruptible(dev);
1009	if (ret)
1010		return ret;
1011
1012	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1013	if (&obj->base == NULL) {
1014		ret = -ENOENT;
1015		goto unlock;
1016	}
1017
1018	/* Bounds check destination. */
1019	if (args->offset > obj->base.size ||
1020	    args->size > obj->base.size - args->offset) {
1021		ret = -EINVAL;
1022		goto out;
1023	}
1024
1025	/* prime objects have no backing filp to GEM pread/pwrite
1026	 * pages from.
1027	 */
1028	if (!obj->base.filp) {
1029		ret = -EINVAL;
1030		goto out;
1031	}
1032
1033	trace_i915_gem_object_pwrite(obj, args->offset, args->size);
1034
 
 
 
 
 
 
 
 
 
 
 
 
 
1035	ret = -EFAULT;
1036	/* We can only do the GTT pwrite on untiled buffers, as otherwise
1037	 * it would end up going through the fenced access, and we'll get
1038	 * different detiling behavior between reading and writing.
1039	 * pread/pwrite currently are reading and writing from the CPU
1040	 * perspective, requiring manual detiling by the client.
1041	 */
1042	if (obj->phys_handle) {
1043		ret = i915_gem_phys_pwrite(obj, args, file);
1044		goto out;
1045	}
1046
1047	if (obj->tiling_mode == I915_TILING_NONE &&
1048	    obj->base.write_domain != I915_GEM_DOMAIN_CPU &&
1049	    cpu_write_needs_clflush(obj)) {
1050		ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
1051		/* Note that the gtt paths might fail with non-page-backed user
1052		 * pointers (e.g. gtt mappings when moving data between
1053		 * textures). Fallback to the shmem path in that case. */
1054	}
1055
1056	if (ret == -EFAULT || ret == -ENOSPC)
1057		ret = i915_gem_shmem_pwrite(dev, obj, args, file);
1058
1059out:
1060	drm_gem_object_unreference(&obj->base);
1061unlock:
1062	mutex_unlock(&dev->struct_mutex);
1063	return ret;
1064}
1065
1066int
1067i915_gem_check_wedge(struct i915_gpu_error *error,
1068		     bool interruptible)
1069{
1070	if (i915_reset_in_progress(error)) {
1071		/* Non-interruptible callers can't handle -EAGAIN, hence return
1072		 * -EIO unconditionally for these. */
1073		if (!interruptible)
1074			return -EIO;
1075
1076		/* Recovery complete, but the reset failed ... */
1077		if (i915_terminally_wedged(error))
1078			return -EIO;
1079
1080		return -EAGAIN;
1081	}
1082
1083	return 0;
1084}
1085
1086/*
1087 * Compare seqno against outstanding lazy request. Emit a request if they are
1088 * equal.
1089 */
1090static int
1091i915_gem_check_olr(struct intel_ring_buffer *ring, u32 seqno)
1092{
1093	int ret;
1094
1095	BUG_ON(!mutex_is_locked(&ring->dev->struct_mutex));
1096
1097	ret = 0;
1098	if (seqno == ring->outstanding_lazy_seqno)
1099		ret = i915_add_request(ring, NULL);
1100
1101	return ret;
1102}
1103
1104static void fake_irq(unsigned long data)
1105{
1106	wake_up_process((struct task_struct *)data);
1107}
1108
1109static bool missed_irq(struct drm_i915_private *dev_priv,
1110		       struct intel_ring_buffer *ring)
1111{
1112	return test_bit(ring->id, &dev_priv->gpu_error.missed_irq_rings);
1113}
1114
1115static bool can_wait_boost(struct drm_i915_file_private *file_priv)
1116{
1117	if (file_priv == NULL)
1118		return true;
1119
1120	return !atomic_xchg(&file_priv->rps_wait_boost, true);
1121}
1122
1123/**
1124 * __wait_seqno - wait until execution of seqno has finished
1125 * @ring: the ring expected to report seqno
1126 * @seqno: duh!
1127 * @reset_counter: reset sequence associated with the given seqno
1128 * @interruptible: do an interruptible wait (normally yes)
1129 * @timeout: in - how long to wait (NULL forever); out - how much time remaining
1130 *
1131 * Note: It is of utmost importance that the passed in seqno and reset_counter
1132 * values have been read by the caller in an smp safe manner. Where read-side
1133 * locks are involved, it is sufficient to read the reset_counter before
1134 * unlocking the lock that protects the seqno. For lockless tricks, the
1135 * reset_counter _must_ be read before, and an appropriate smp_rmb must be
1136 * inserted.
1137 *
1138 * Returns 0 if the seqno was found within the alloted time. Else returns the
1139 * errno with remaining time filled in timeout argument.
1140 */
1141static int __wait_seqno(struct intel_ring_buffer *ring, u32 seqno,
1142			unsigned reset_counter,
1143			bool interruptible,
1144			struct timespec *timeout,
1145			struct drm_i915_file_private *file_priv)
1146{
1147	struct drm_device *dev = ring->dev;
1148	struct drm_i915_private *dev_priv = dev->dev_private;
1149	const bool irq_test_in_progress =
1150		ACCESS_ONCE(dev_priv->gpu_error.test_irq_rings) & intel_ring_flag(ring);
1151	struct timespec before, now;
1152	DEFINE_WAIT(wait);
1153	unsigned long timeout_expire;
1154	int ret;
1155
1156	WARN(dev_priv->pm.irqs_disabled, "IRQs disabled\n");
1157
1158	if (i915_seqno_passed(ring->get_seqno(ring, true), seqno))
1159		return 0;
1160
1161	timeout_expire = timeout ? jiffies + timespec_to_jiffies_timeout(timeout) : 0;
1162
1163	if (INTEL_INFO(dev)->gen >= 6 && can_wait_boost(file_priv)) {
1164		gen6_rps_boost(dev_priv);
1165		if (file_priv)
1166			mod_delayed_work(dev_priv->wq,
1167					 &file_priv->mm.idle_work,
1168					 msecs_to_jiffies(100));
1169	}
1170
1171	if (!irq_test_in_progress && WARN_ON(!ring->irq_get(ring)))
1172		return -ENODEV;
1173
1174	/* Record current time in case interrupted by signal, or wedged */
1175	trace_i915_gem_request_wait_begin(ring, seqno);
1176	getrawmonotonic(&before);
1177	for (;;) {
1178		struct timer_list timer;
1179
1180		prepare_to_wait(&ring->irq_queue, &wait,
1181				interruptible ? TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE);
1182
1183		/* We need to check whether any gpu reset happened in between
1184		 * the caller grabbing the seqno and now ... */
1185		if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) {
1186			/* ... but upgrade the -EAGAIN to an -EIO if the gpu
1187			 * is truely gone. */
1188			ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1189			if (ret == 0)
1190				ret = -EAGAIN;
1191			break;
1192		}
1193
1194		if (i915_seqno_passed(ring->get_seqno(ring, false), seqno)) {
1195			ret = 0;
1196			break;
1197		}
1198
1199		if (interruptible && signal_pending(current)) {
1200			ret = -ERESTARTSYS;
1201			break;
1202		}
1203
1204		if (timeout && time_after_eq(jiffies, timeout_expire)) {
1205			ret = -ETIME;
1206			break;
1207		}
1208
1209		timer.function = NULL;
1210		if (timeout || missed_irq(dev_priv, ring)) {
1211			unsigned long expire;
1212
1213			setup_timer_on_stack(&timer, fake_irq, (unsigned long)current);
1214			expire = missed_irq(dev_priv, ring) ? jiffies + 1 : timeout_expire;
1215			mod_timer(&timer, expire);
1216		}
1217
1218		io_schedule();
1219
1220		if (timer.function) {
1221			del_singleshot_timer_sync(&timer);
1222			destroy_timer_on_stack(&timer);
1223		}
1224	}
1225	getrawmonotonic(&now);
1226	trace_i915_gem_request_wait_end(ring, seqno);
1227
1228	if (!irq_test_in_progress)
1229		ring->irq_put(ring);
1230
1231	finish_wait(&ring->irq_queue, &wait);
1232
1233	if (timeout) {
1234		struct timespec sleep_time = timespec_sub(now, before);
1235		*timeout = timespec_sub(*timeout, sleep_time);
1236		if (!timespec_valid(timeout)) /* i.e. negative time remains */
1237			set_normalized_timespec(timeout, 0, 0);
1238	}
1239
1240	return ret;
1241}
1242
1243/**
1244 * Waits for a sequence number to be signaled, and cleans up the
1245 * request and object lists appropriately for that event.
1246 */
1247int
1248i915_wait_seqno(struct intel_ring_buffer *ring, uint32_t seqno)
1249{
1250	struct drm_device *dev = ring->dev;
1251	struct drm_i915_private *dev_priv = dev->dev_private;
1252	bool interruptible = dev_priv->mm.interruptible;
1253	int ret;
1254
1255	BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1256	BUG_ON(seqno == 0);
1257
1258	ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
1259	if (ret)
1260		return ret;
1261
1262	ret = i915_gem_check_olr(ring, seqno);
1263	if (ret)
1264		return ret;
1265
1266	return __wait_seqno(ring, seqno,
1267			    atomic_read(&dev_priv->gpu_error.reset_counter),
1268			    interruptible, NULL, NULL);
1269}
1270
1271static int
1272i915_gem_object_wait_rendering__tail(struct drm_i915_gem_object *obj,
1273				     struct intel_ring_buffer *ring)
1274{
1275	i915_gem_retire_requests_ring(ring);
1276
1277	/* Manually manage the write flush as we may have not yet
1278	 * retired the buffer.
1279	 *
1280	 * Note that the last_write_seqno is always the earlier of
1281	 * the two (read/write) seqno, so if we haved successfully waited,
1282	 * we know we have passed the last write.
1283	 */
1284	obj->last_write_seqno = 0;
1285	obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
1286
1287	return 0;
1288}
1289
1290/**
1291 * Ensures that all rendering to the object has completed and the object is
1292 * safe to unbind from the GTT or access from the CPU.
1293 */
1294static __must_check int
1295i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
1296			       bool readonly)
1297{
1298	struct intel_ring_buffer *ring = obj->ring;
1299	u32 seqno;
1300	int ret;
1301
1302	seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno;
1303	if (seqno == 0)
1304		return 0;
1305
1306	ret = i915_wait_seqno(ring, seqno);
1307	if (ret)
1308		return ret;
1309
1310	return i915_gem_object_wait_rendering__tail(obj, ring);
1311}
1312
1313/* A nonblocking variant of the above wait. This is a highly dangerous routine
1314 * as the object state may change during this call.
1315 */
1316static __must_check int
1317i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
1318					    struct drm_i915_file_private *file_priv,
1319					    bool readonly)
1320{
1321	struct drm_device *dev = obj->base.dev;
1322	struct drm_i915_private *dev_priv = dev->dev_private;
1323	struct intel_ring_buffer *ring = obj->ring;
1324	unsigned reset_counter;
1325	u32 seqno;
1326	int ret;
1327
1328	BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1329	BUG_ON(!dev_priv->mm.interruptible);
1330
1331	seqno = readonly ? obj->last_write_seqno : obj->last_read_seqno;
1332	if (seqno == 0)
1333		return 0;
1334
1335	ret = i915_gem_check_wedge(&dev_priv->gpu_error, true);
1336	if (ret)
1337		return ret;
1338
1339	ret = i915_gem_check_olr(ring, seqno);
1340	if (ret)
1341		return ret;
1342
1343	reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
1344	mutex_unlock(&dev->struct_mutex);
1345	ret = __wait_seqno(ring, seqno, reset_counter, true, NULL, file_priv);
1346	mutex_lock(&dev->struct_mutex);
1347	if (ret)
1348		return ret;
1349
1350	return i915_gem_object_wait_rendering__tail(obj, ring);
1351}
1352
1353/**
1354 * Called when user space prepares to use an object with the CPU, either
1355 * through the mmap ioctl's mapping or a GTT mapping.
1356 */
1357int
1358i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1359			  struct drm_file *file)
1360{
1361	struct drm_i915_gem_set_domain *args = data;
1362	struct drm_i915_gem_object *obj;
1363	uint32_t read_domains = args->read_domains;
1364	uint32_t write_domain = args->write_domain;
1365	int ret;
1366
1367	/* Only handle setting domains to types used by the CPU. */
1368	if (write_domain & I915_GEM_GPU_DOMAINS)
1369		return -EINVAL;
1370
1371	if (read_domains & I915_GEM_GPU_DOMAINS)
1372		return -EINVAL;
1373
1374	/* Having something in the write domain implies it's in the read
1375	 * domain, and only that read domain.  Enforce that in the request.
1376	 */
1377	if (write_domain != 0 && read_domains != write_domain)
1378		return -EINVAL;
1379
1380	ret = i915_mutex_lock_interruptible(dev);
1381	if (ret)
1382		return ret;
1383
1384	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1385	if (&obj->base == NULL) {
1386		ret = -ENOENT;
1387		goto unlock;
1388	}
1389
1390	/* Try to flush the object off the GPU without holding the lock.
1391	 * We will repeat the flush holding the lock in the normal manner
1392	 * to catch cases where we are gazumped.
1393	 */
1394	ret = i915_gem_object_wait_rendering__nonblocking(obj,
1395							  file->driver_priv,
1396							  !write_domain);
1397	if (ret)
1398		goto unref;
1399
1400	if (read_domains & I915_GEM_DOMAIN_GTT) {
1401		ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1402
1403		/* Silently promote "you're not bound, there was nothing to do"
1404		 * to success, since the client was just asking us to
1405		 * make sure everything was done.
1406		 */
1407		if (ret == -EINVAL)
1408			ret = 0;
1409	} else {
1410		ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1411	}
1412
1413unref:
1414	drm_gem_object_unreference(&obj->base);
1415unlock:
1416	mutex_unlock(&dev->struct_mutex);
1417	return ret;
1418}
1419
1420/**
1421 * Called when user space has done writes to this buffer
1422 */
1423int
1424i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1425			 struct drm_file *file)
1426{
1427	struct drm_i915_gem_sw_finish *args = data;
1428	struct drm_i915_gem_object *obj;
1429	int ret = 0;
1430
1431	ret = i915_mutex_lock_interruptible(dev);
1432	if (ret)
1433		return ret;
1434
1435	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1436	if (&obj->base == NULL) {
1437		ret = -ENOENT;
1438		goto unlock;
1439	}
1440
1441	/* Pinned buffers may be scanout, so flush the cache */
1442	if (obj->pin_display)
1443		i915_gem_object_flush_cpu_write_domain(obj, true);
1444
1445	drm_gem_object_unreference(&obj->base);
1446unlock:
1447	mutex_unlock(&dev->struct_mutex);
1448	return ret;
1449}
1450
1451/**
1452 * Maps the contents of an object, returning the address it is mapped
1453 * into.
1454 *
1455 * While the mapping holds a reference on the contents of the object, it doesn't
1456 * imply a ref on the object itself.
1457 */
1458int
1459i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1460		    struct drm_file *file)
1461{
1462	struct drm_i915_gem_mmap *args = data;
1463	struct drm_gem_object *obj;
1464	unsigned long addr;
1465
1466	obj = drm_gem_object_lookup(dev, file, args->handle);
1467	if (obj == NULL)
1468		return -ENOENT;
1469
1470	/* prime objects have no backing filp to GEM mmap
1471	 * pages from.
 
1472	 */
1473	if (!obj->filp) {
1474		drm_gem_object_unreference_unlocked(obj);
1475		return -EINVAL;
1476	}
1477
1478	addr = vm_mmap(obj->filp, 0, args->size,
1479		       PROT_READ | PROT_WRITE, MAP_SHARED,
1480		       args->offset);
1481	drm_gem_object_unreference_unlocked(obj);
1482	if (IS_ERR((void *)addr))
1483		return addr;
1484
1485	args->addr_ptr = (uint64_t) addr;
1486
1487	return 0;
1488}
1489
1490/**
1491 * i915_gem_fault - fault a page into the GTT
1492 * vma: VMA in question
1493 * vmf: fault info
1494 *
1495 * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1496 * from userspace.  The fault handler takes care of binding the object to
1497 * the GTT (if needed), allocating and programming a fence register (again,
1498 * only if needed based on whether the old reg is still valid or the object
1499 * is tiled) and inserting a new PTE into the faulting process.
1500 *
1501 * Note that the faulting process may involve evicting existing objects
1502 * from the GTT and/or fence registers to make room.  So performance may
1503 * suffer if the GTT working set is large or there are few fence registers
1504 * left.
1505 */
1506int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1507{
1508	struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1509	struct drm_device *dev = obj->base.dev;
1510	struct drm_i915_private *dev_priv = dev->dev_private;
1511	pgoff_t page_offset;
1512	unsigned long pfn;
1513	int ret = 0;
1514	bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1515
1516	intel_runtime_pm_get(dev_priv);
1517
1518	/* We don't use vmf->pgoff since that has the fake offset */
1519	page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1520		PAGE_SHIFT;
1521
1522	ret = i915_mutex_lock_interruptible(dev);
1523	if (ret)
1524		goto out;
1525
1526	trace_i915_gem_object_fault(obj, page_offset, true, write);
1527
1528	/* Try to flush the object off the GPU first without holding the lock.
1529	 * Upon reacquiring the lock, we will perform our sanity checks and then
1530	 * repeat the flush holding the lock in the normal manner to catch cases
1531	 * where we are gazumped.
1532	 */
1533	ret = i915_gem_object_wait_rendering__nonblocking(obj, NULL, !write);
1534	if (ret)
1535		goto unlock;
1536
1537	/* Access to snoopable pages through the GTT is incoherent. */
1538	if (obj->cache_level != I915_CACHE_NONE && !HAS_LLC(dev)) {
1539		ret = -EINVAL;
1540		goto unlock;
1541	}
1542
1543	/* Now bind it into the GTT if needed */
1544	ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE);
1545	if (ret)
1546		goto unlock;
1547
1548	ret = i915_gem_object_set_to_gtt_domain(obj, write);
1549	if (ret)
1550		goto unpin;
1551
1552	ret = i915_gem_object_get_fence(obj);
1553	if (ret)
1554		goto unpin;
1555
1556	obj->fault_mappable = true;
1557
1558	pfn = dev_priv->gtt.mappable_base + i915_gem_obj_ggtt_offset(obj);
1559	pfn >>= PAGE_SHIFT;
1560	pfn += page_offset;
1561
1562	/* Finally, remap it using the new GTT offset */
1563	ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1564unpin:
1565	i915_gem_object_ggtt_unpin(obj);
1566unlock:
1567	mutex_unlock(&dev->struct_mutex);
1568out:
1569	switch (ret) {
1570	case -EIO:
1571		/* If this -EIO is due to a gpu hang, give the reset code a
1572		 * chance to clean up the mess. Otherwise return the proper
1573		 * SIGBUS. */
1574		if (i915_terminally_wedged(&dev_priv->gpu_error)) {
1575			ret = VM_FAULT_SIGBUS;
1576			break;
1577		}
1578	case -EAGAIN:
1579		/*
1580		 * EAGAIN means the gpu is hung and we'll wait for the error
1581		 * handler to reset everything when re-faulting in
1582		 * i915_mutex_lock_interruptible.
1583		 */
1584	case 0:
1585	case -ERESTARTSYS:
1586	case -EINTR:
1587	case -EBUSY:
1588		/*
1589		 * EBUSY is ok: this just means that another thread
1590		 * already did the job.
1591		 */
1592		ret = VM_FAULT_NOPAGE;
1593		break;
1594	case -ENOMEM:
1595		ret = VM_FAULT_OOM;
1596		break;
1597	case -ENOSPC:
1598	case -EFAULT:
1599		ret = VM_FAULT_SIGBUS;
1600		break;
1601	default:
1602		WARN_ONCE(ret, "unhandled error in i915_gem_fault: %i\n", ret);
1603		ret = VM_FAULT_SIGBUS;
1604		break;
1605	}
1606
1607	intel_runtime_pm_put(dev_priv);
1608	return ret;
1609}
1610
1611void i915_gem_release_all_mmaps(struct drm_i915_private *dev_priv)
1612{
1613	struct i915_vma *vma;
1614
1615	/*
1616	 * Only the global gtt is relevant for gtt memory mappings, so restrict
1617	 * list traversal to objects bound into the global address space. Note
1618	 * that the active list should be empty, but better safe than sorry.
 
1619	 */
1620	WARN_ON(!list_empty(&dev_priv->gtt.base.active_list));
1621	list_for_each_entry(vma, &dev_priv->gtt.base.active_list, mm_list)
1622		i915_gem_release_mmap(vma->obj);
1623	list_for_each_entry(vma, &dev_priv->gtt.base.inactive_list, mm_list)
1624		i915_gem_release_mmap(vma->obj);
1625}
1626
1627/**
1628 * i915_gem_release_mmap - remove physical page mappings
1629 * @obj: obj in question
1630 *
1631 * Preserve the reservation of the mmapping with the DRM core code, but
1632 * relinquish ownership of the pages back to the system.
1633 *
1634 * It is vital that we remove the page mapping if we have mapped a tiled
1635 * object through the GTT and then lose the fence register due to
1636 * resource pressure. Similarly if the object has been moved out of the
1637 * aperture, than pages mapped into userspace must be revoked. Removing the
1638 * mapping will then trigger a page fault on the next user access, allowing
1639 * fixup by i915_gem_fault().
1640 */
1641void
1642i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1643{
1644	if (!obj->fault_mappable)
1645		return;
1646
1647	drm_vma_node_unmap(&obj->base.vma_node,
1648			   obj->base.dev->anon_inode->i_mapping);
1649	obj->fault_mappable = false;
1650}
 
 
 
1651
1652uint32_t
1653i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1654{
1655	uint32_t gtt_size;
1656
1657	if (INTEL_INFO(dev)->gen >= 4 ||
1658	    tiling_mode == I915_TILING_NONE)
1659		return size;
1660
1661	/* Previous chips need a power-of-two fence region when tiling */
1662	if (INTEL_INFO(dev)->gen == 3)
1663		gtt_size = 1024*1024;
1664	else
1665		gtt_size = 512*1024;
1666
1667	while (gtt_size < size)
1668		gtt_size <<= 1;
1669
1670	return gtt_size;
1671}
1672
1673/**
1674 * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1675 * @obj: object to check
1676 *
1677 * Return the required GTT alignment for an object, taking into account
1678 * potential fence register mapping.
1679 */
1680uint32_t
1681i915_gem_get_gtt_alignment(struct drm_device *dev, uint32_t size,
1682			   int tiling_mode, bool fenced)
1683{
1684	/*
1685	 * Minimum alignment is 4k (GTT page size), but might be greater
1686	 * if a fence register is needed for the object.
1687	 */
1688	if (INTEL_INFO(dev)->gen >= 4 || (!fenced && IS_G33(dev)) ||
1689	    tiling_mode == I915_TILING_NONE)
1690		return 4096;
1691
1692	/*
1693	 * Previous chips need to be aligned to the size of the smallest
1694	 * fence register that can contain the object.
1695	 */
1696	return i915_gem_get_gtt_size(dev, size, tiling_mode);
1697}
1698
1699static int i915_gem_object_create_mmap_offset(struct drm_i915_gem_object *obj)
1700{
1701	struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
1702	int ret;
1703
1704	if (drm_vma_node_has_offset(&obj->base.vma_node))
1705		return 0;
1706
1707	dev_priv->mm.shrinker_no_lock_stealing = true;
1708
1709	ret = drm_gem_create_mmap_offset(&obj->base);
1710	if (ret != -ENOSPC)
1711		goto out;
1712
1713	/* Badly fragmented mmap space? The only way we can recover
1714	 * space is by destroying unwanted objects. We can't randomly release
1715	 * mmap_offsets as userspace expects them to be persistent for the
1716	 * lifetime of the objects. The closest we can is to release the
1717	 * offsets on purgeable objects by truncating it and marking it purged,
1718	 * which prevents userspace from ever using that object again.
1719	 */
1720	i915_gem_purge(dev_priv, obj->base.size >> PAGE_SHIFT);
1721	ret = drm_gem_create_mmap_offset(&obj->base);
1722	if (ret != -ENOSPC)
1723		goto out;
1724
1725	i915_gem_shrink_all(dev_priv);
1726	ret = drm_gem_create_mmap_offset(&obj->base);
1727out:
1728	dev_priv->mm.shrinker_no_lock_stealing = false;
1729
1730	return ret;
1731}
1732
1733static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)
1734{
1735	drm_gem_free_mmap_offset(&obj->base);
1736}
1737
1738int
1739i915_gem_mmap_gtt(struct drm_file *file,
1740		  struct drm_device *dev,
1741		  uint32_t handle,
1742		  uint64_t *offset)
1743{
1744	struct drm_i915_private *dev_priv = dev->dev_private;
1745	struct drm_i915_gem_object *obj;
1746	int ret;
1747
1748	ret = i915_mutex_lock_interruptible(dev);
1749	if (ret)
1750		return ret;
1751
1752	obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
1753	if (&obj->base == NULL) {
1754		ret = -ENOENT;
1755		goto unlock;
1756	}
1757
1758	if (obj->base.size > dev_priv->gtt.mappable_end) {
1759		ret = -E2BIG;
1760		goto out;
1761	}
1762
1763	if (obj->madv != I915_MADV_WILLNEED) {
1764		DRM_DEBUG("Attempting to mmap a purgeable buffer\n");
1765		ret = -EFAULT;
1766		goto out;
1767	}
1768
1769	ret = i915_gem_object_create_mmap_offset(obj);
1770	if (ret)
1771		goto out;
1772
1773	*offset = drm_vma_node_offset_addr(&obj->base.vma_node);
1774
1775out:
1776	drm_gem_object_unreference(&obj->base);
1777unlock:
1778	mutex_unlock(&dev->struct_mutex);
1779	return ret;
1780}
1781
1782/**
1783 * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1784 * @dev: DRM device
1785 * @data: GTT mapping ioctl data
1786 * @file: GEM object info
1787 *
1788 * Simply returns the fake offset to userspace so it can mmap it.
1789 * The mmap call will end up in drm_gem_mmap(), which will set things
1790 * up so we can get faults in the handler above.
1791 *
1792 * The fault handler will take care of binding the object into the GTT
1793 * (since it may have been evicted to make room for something), allocating
1794 * a fence register, and mapping the appropriate aperture address into
1795 * userspace.
1796 */
1797int
1798i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1799			struct drm_file *file)
1800{
1801	struct drm_i915_gem_mmap_gtt *args = data;
1802
1803	return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
1804}
1805
1806/* Immediately discard the backing storage */
1807static void
1808i915_gem_object_truncate(struct drm_i915_gem_object *obj)
1809{
1810	struct inode *inode;
1811
1812	i915_gem_object_free_mmap_offset(obj);
1813
1814	if (obj->base.filp == NULL)
1815		return;
1816
1817	/* Our goal here is to return as much of the memory as
1818	 * is possible back to the system as we are called from OOM.
1819	 * To do this we must instruct the shmfs to drop all of its
1820	 * backing pages, *now*.
1821	 */
1822	inode = file_inode(obj->base.filp);
1823	shmem_truncate_range(inode, 0, (loff_t)-1);
1824
1825	obj->madv = __I915_MADV_PURGED;
1826}
1827
1828static inline int
1829i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj)
1830{
1831	return obj->madv == I915_MADV_DONTNEED;
1832}
1833
1834static void
1835i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
1836{
1837	struct sg_page_iter sg_iter;
1838	int ret;
1839
1840	BUG_ON(obj->madv == __I915_MADV_PURGED);
1841
1842	ret = i915_gem_object_set_to_cpu_domain(obj, true);
1843	if (ret) {
1844		/* In the event of a disaster, abandon all caches and
1845		 * hope for the best.
1846		 */
1847		WARN_ON(ret != -EIO);
1848		i915_gem_clflush_object(obj, true);
1849		obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
1850	}
1851
1852	if (i915_gem_object_needs_bit17_swizzle(obj))
1853		i915_gem_object_save_bit_17_swizzle(obj);
1854
1855	if (obj->madv == I915_MADV_DONTNEED)
1856		obj->dirty = 0;
1857
1858	for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
1859		struct page *page = sg_page_iter_page(&sg_iter);
1860
1861		if (obj->dirty)
1862			set_page_dirty(page);
1863
1864		if (obj->madv == I915_MADV_WILLNEED)
1865			mark_page_accessed(page);
1866
1867		page_cache_release(page);
1868	}
1869	obj->dirty = 0;
1870
1871	sg_free_table(obj->pages);
1872	kfree(obj->pages);
1873}
1874
1875int
1876i915_gem_object_put_pages(struct drm_i915_gem_object *obj)
1877{
1878	const struct drm_i915_gem_object_ops *ops = obj->ops;
1879
1880	if (obj->pages == NULL)
1881		return 0;
1882
1883	if (obj->pages_pin_count)
1884		return -EBUSY;
1885
1886	BUG_ON(i915_gem_obj_bound_any(obj));
1887
1888	/* ->put_pages might need to allocate memory for the bit17 swizzle
1889	 * array, hence protect them from being reaped by removing them from gtt
1890	 * lists early. */
1891	list_del(&obj->global_list);
1892
1893	ops->put_pages(obj);
1894	obj->pages = NULL;
1895
1896	if (i915_gem_object_is_purgeable(obj))
1897		i915_gem_object_truncate(obj);
1898
1899	return 0;
1900}
1901
1902static unsigned long
1903__i915_gem_shrink(struct drm_i915_private *dev_priv, long target,
1904		  bool purgeable_only)
1905{
1906	struct list_head still_bound_list;
1907	struct drm_i915_gem_object *obj, *next;
1908	unsigned long count = 0;
1909
1910	list_for_each_entry_safe(obj, next,
1911				 &dev_priv->mm.unbound_list,
1912				 global_list) {
1913		if ((i915_gem_object_is_purgeable(obj) || !purgeable_only) &&
1914		    i915_gem_object_put_pages(obj) == 0) {
1915			count += obj->base.size >> PAGE_SHIFT;
1916			if (count >= target)
1917				return count;
1918		}
1919	}
1920
1921	/*
1922	 * As we may completely rewrite the bound list whilst unbinding
1923	 * (due to retiring requests) we have to strictly process only
1924	 * one element of the list at the time, and recheck the list
1925	 * on every iteration.
1926	 */
1927	INIT_LIST_HEAD(&still_bound_list);
1928	while (count < target && !list_empty(&dev_priv->mm.bound_list)) {
1929		struct i915_vma *vma, *v;
1930
1931		obj = list_first_entry(&dev_priv->mm.bound_list,
1932				       typeof(*obj), global_list);
1933		list_move_tail(&obj->global_list, &still_bound_list);
1934
1935		if (!i915_gem_object_is_purgeable(obj) && purgeable_only)
1936			continue;
1937
1938		/*
1939		 * Hold a reference whilst we unbind this object, as we may
1940		 * end up waiting for and retiring requests. This might
1941		 * release the final reference (held by the active list)
1942		 * and result in the object being freed from under us.
1943		 * in this object being freed.
1944		 *
1945		 * Note 1: Shrinking the bound list is special since only active
1946		 * (and hence bound objects) can contain such limbo objects, so
1947		 * we don't need special tricks for shrinking the unbound list.
1948		 * The only other place where we have to be careful with active
1949		 * objects suddenly disappearing due to retiring requests is the
1950		 * eviction code.
1951		 *
1952		 * Note 2: Even though the bound list doesn't hold a reference
1953		 * to the object we can safely grab one here: The final object
1954		 * unreferencing and the bound_list are both protected by the
1955		 * dev->struct_mutex and so we won't ever be able to observe an
1956		 * object on the bound_list with a reference count equals 0.
1957		 */
1958		drm_gem_object_reference(&obj->base);
1959
1960		list_for_each_entry_safe(vma, v, &obj->vma_list, vma_link)
1961			if (i915_vma_unbind(vma))
1962				break;
1963
1964		if (i915_gem_object_put_pages(obj) == 0)
1965			count += obj->base.size >> PAGE_SHIFT;
1966
1967		drm_gem_object_unreference(&obj->base);
1968	}
1969	list_splice(&still_bound_list, &dev_priv->mm.bound_list);
1970
1971	return count;
1972}
1973
1974static unsigned long
1975i915_gem_purge(struct drm_i915_private *dev_priv, long target)
1976{
1977	return __i915_gem_shrink(dev_priv, target, true);
1978}
1979
1980static unsigned long
1981i915_gem_shrink_all(struct drm_i915_private *dev_priv)
1982{
1983	struct drm_i915_gem_object *obj, *next;
1984	long freed = 0;
1985
1986	i915_gem_evict_everything(dev_priv->dev);
1987
1988	list_for_each_entry_safe(obj, next, &dev_priv->mm.unbound_list,
1989				 global_list) {
1990		if (i915_gem_object_put_pages(obj) == 0)
1991			freed += obj->base.size >> PAGE_SHIFT;
1992	}
1993	return freed;
1994}
1995
1996static int
1997i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
1998{
1999	struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2000	int page_count, i;
2001	struct address_space *mapping;
2002	struct sg_table *st;
2003	struct scatterlist *sg;
2004	struct sg_page_iter sg_iter;
2005	struct page *page;
2006	unsigned long last_pfn = 0;	/* suppress gcc warning */
2007	gfp_t gfp;
2008
2009	/* Assert that the object is not currently in any GPU domain. As it
2010	 * wasn't in the GTT, there shouldn't be any way it could have been in
2011	 * a GPU cache
2012	 */
2013	BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
2014	BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
2015
2016	st = kmalloc(sizeof(*st), GFP_KERNEL);
2017	if (st == NULL)
2018		return -ENOMEM;
2019
2020	page_count = obj->base.size / PAGE_SIZE;
2021	if (sg_alloc_table(st, page_count, GFP_KERNEL)) {
2022		kfree(st);
2023		return -ENOMEM;
2024	}
2025
2026	/* Get the list of pages out of our struct file.  They'll be pinned
2027	 * at this point until we release them.
2028	 *
2029	 * Fail silently without starting the shrinker
2030	 */
2031	mapping = file_inode(obj->base.filp)->i_mapping;
2032	gfp = mapping_gfp_mask(mapping);
2033	gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD;
2034	gfp &= ~(__GFP_IO | __GFP_WAIT);
2035	sg = st->sgl;
2036	st->nents = 0;
2037	for (i = 0; i < page_count; i++) {
2038		page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2039		if (IS_ERR(page)) {
2040			i915_gem_purge(dev_priv, page_count);
2041			page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2042		}
2043		if (IS_ERR(page)) {
2044			/* We've tried hard to allocate the memory by reaping
2045			 * our own buffer, now let the real VM do its job and
2046			 * go down in flames if truly OOM.
2047			 */
2048			gfp &= ~(__GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD);
2049			gfp |= __GFP_IO | __GFP_WAIT;
2050
2051			i915_gem_shrink_all(dev_priv);
2052			page = shmem_read_mapping_page_gfp(mapping, i, gfp);
2053			if (IS_ERR(page))
2054				goto err_pages;
2055
2056			gfp |= __GFP_NORETRY | __GFP_NOWARN | __GFP_NO_KSWAPD;
2057			gfp &= ~(__GFP_IO | __GFP_WAIT);
2058		}
2059#ifdef CONFIG_SWIOTLB
2060		if (swiotlb_nr_tbl()) {
2061			st->nents++;
2062			sg_set_page(sg, page, PAGE_SIZE, 0);
2063			sg = sg_next(sg);
2064			continue;
2065		}
2066#endif
2067		if (!i || page_to_pfn(page) != last_pfn + 1) {
2068			if (i)
2069				sg = sg_next(sg);
2070			st->nents++;
2071			sg_set_page(sg, page, PAGE_SIZE, 0);
2072		} else {
2073			sg->length += PAGE_SIZE;
2074		}
2075		last_pfn = page_to_pfn(page);
2076
2077		/* Check that the i965g/gm workaround works. */
2078		WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
2079	}
2080#ifdef CONFIG_SWIOTLB
2081	if (!swiotlb_nr_tbl())
2082#endif
2083		sg_mark_end(sg);
2084	obj->pages = st;
2085
2086	if (i915_gem_object_needs_bit17_swizzle(obj))
2087		i915_gem_object_do_bit_17_swizzle(obj);
2088
2089	return 0;
2090
2091err_pages:
2092	sg_mark_end(sg);
2093	for_each_sg_page(st->sgl, &sg_iter, st->nents, 0)
2094		page_cache_release(sg_page_iter_page(&sg_iter));
2095	sg_free_table(st);
2096	kfree(st);
2097	return PTR_ERR(page);
2098}
2099
2100/* Ensure that the associated pages are gathered from the backing storage
2101 * and pinned into our object. i915_gem_object_get_pages() may be called
2102 * multiple times before they are released by a single call to
2103 * i915_gem_object_put_pages() - once the pages are no longer referenced
2104 * either as a result of memory pressure (reaping pages under the shrinker)
2105 * or as the object is itself released.
2106 */
2107int
2108i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
2109{
2110	struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2111	const struct drm_i915_gem_object_ops *ops = obj->ops;
2112	int ret;
2113
2114	if (obj->pages)
2115		return 0;
2116
2117	if (obj->madv != I915_MADV_WILLNEED) {
2118		DRM_DEBUG("Attempting to obtain a purgeable object\n");
2119		return -EFAULT;
2120	}
2121
2122	BUG_ON(obj->pages_pin_count);
2123
2124	ret = ops->get_pages(obj);
2125	if (ret)
2126		return ret;
2127
2128	list_add_tail(&obj->global_list, &dev_priv->mm.unbound_list);
2129	return 0;
2130}
2131
2132static void
2133i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
2134			       struct intel_ring_buffer *ring)
2135{
2136	struct drm_device *dev = obj->base.dev;
2137	struct drm_i915_private *dev_priv = dev->dev_private;
2138	u32 seqno = intel_ring_get_seqno(ring);
2139
2140	BUG_ON(ring == NULL);
2141	if (obj->ring != ring && obj->last_write_seqno) {
2142		/* Keep the seqno relative to the current ring */
2143		obj->last_write_seqno = seqno;
2144	}
2145	obj->ring = ring;
2146
2147	/* Add a reference if we're newly entering the active list. */
2148	if (!obj->active) {
2149		drm_gem_object_reference(&obj->base);
2150		obj->active = 1;
2151	}
2152
2153	list_move_tail(&obj->ring_list, &ring->active_list);
2154
2155	obj->last_read_seqno = seqno;
2156
2157	if (obj->fenced_gpu_access) {
2158		obj->last_fenced_seqno = seqno;
2159
2160		/* Bump MRU to take account of the delayed flush */
2161		if (obj->fence_reg != I915_FENCE_REG_NONE) {
2162			struct drm_i915_fence_reg *reg;
2163
2164			reg = &dev_priv->fence_regs[obj->fence_reg];
2165			list_move_tail(&reg->lru_list,
2166				       &dev_priv->mm.fence_list);
2167		}
2168	}
2169}
2170
2171void i915_vma_move_to_active(struct i915_vma *vma,
2172			     struct intel_ring_buffer *ring)
2173{
2174	list_move_tail(&vma->mm_list, &vma->vm->active_list);
2175	return i915_gem_object_move_to_active(vma->obj, ring);
2176}
2177
2178static void
2179i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
2180{
2181	struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2182	struct i915_address_space *vm;
2183	struct i915_vma *vma;
2184
2185	BUG_ON(obj->base.write_domain & ~I915_GEM_GPU_DOMAINS);
2186	BUG_ON(!obj->active);
2187
2188	list_for_each_entry(vm, &dev_priv->vm_list, global_link) {
2189		vma = i915_gem_obj_to_vma(obj, vm);
2190		if (vma && !list_empty(&vma->mm_list))
2191			list_move_tail(&vma->mm_list, &vm->inactive_list);
2192	}
2193
2194	list_del_init(&obj->ring_list);
2195	obj->ring = NULL;
2196
2197	obj->last_read_seqno = 0;
2198	obj->last_write_seqno = 0;
2199	obj->base.write_domain = 0;
2200
2201	obj->last_fenced_seqno = 0;
2202	obj->fenced_gpu_access = false;
2203
2204	obj->active = 0;
2205	drm_gem_object_unreference(&obj->base);
2206
2207	WARN_ON(i915_verify_lists(dev));
2208}
2209
2210static int
2211i915_gem_init_seqno(struct drm_device *dev, u32 seqno)
2212{
2213	struct drm_i915_private *dev_priv = dev->dev_private;
2214	struct intel_ring_buffer *ring;
2215	int ret, i, j;
2216
2217	/* Carefully retire all requests without writing to the rings */
2218	for_each_ring(ring, dev_priv, i) {
2219		ret = intel_ring_idle(ring);
2220		if (ret)
2221			return ret;
2222	}
2223	i915_gem_retire_requests(dev);
2224
2225	/* Finally reset hw state */
2226	for_each_ring(ring, dev_priv, i) {
2227		intel_ring_init_seqno(ring, seqno);
2228
2229		for (j = 0; j < ARRAY_SIZE(ring->sync_seqno); j++)
2230			ring->sync_seqno[j] = 0;
2231	}
2232
2233	return 0;
2234}
2235
2236int i915_gem_set_seqno(struct drm_device *dev, u32 seqno)
2237{
2238	struct drm_i915_private *dev_priv = dev->dev_private;
2239	int ret;
2240
2241	if (seqno == 0)
2242		return -EINVAL;
2243
2244	/* HWS page needs to be set less than what we
2245	 * will inject to ring
2246	 */
2247	ret = i915_gem_init_seqno(dev, seqno - 1);
2248	if (ret)
2249		return ret;
2250
2251	/* Carefully set the last_seqno value so that wrap
2252	 * detection still works
2253	 */
2254	dev_priv->next_seqno = seqno;
2255	dev_priv->last_seqno = seqno - 1;
2256	if (dev_priv->last_seqno == 0)
2257		dev_priv->last_seqno--;
2258
2259	return 0;
2260}
2261
2262int
2263i915_gem_get_seqno(struct drm_device *dev, u32 *seqno)
2264{
2265	struct drm_i915_private *dev_priv = dev->dev_private;
2266
2267	/* reserve 0 for non-seqno */
2268	if (dev_priv->next_seqno == 0) {
2269		int ret = i915_gem_init_seqno(dev, 0);
2270		if (ret)
2271			return ret;
2272
2273		dev_priv->next_seqno = 1;
2274	}
2275
2276	*seqno = dev_priv->last_seqno = dev_priv->next_seqno++;
2277	return 0;
2278}
2279
2280int __i915_add_request(struct intel_ring_buffer *ring,
2281		       struct drm_file *file,
2282		       struct drm_i915_gem_object *obj,
2283		       u32 *out_seqno)
2284{
2285	struct drm_i915_private *dev_priv = ring->dev->dev_private;
2286	struct drm_i915_gem_request *request;
2287	u32 request_ring_position, request_start;
2288	int ret;
2289
2290	request_start = intel_ring_get_tail(ring);
2291	/*
2292	 * Emit any outstanding flushes - execbuf can fail to emit the flush
2293	 * after having emitted the batchbuffer command. Hence we need to fix
2294	 * things up similar to emitting the lazy request. The difference here
2295	 * is that the flush _must_ happen before the next request, no matter
2296	 * what.
2297	 */
2298	ret = intel_ring_flush_all_caches(ring);
2299	if (ret)
2300		return ret;
2301
2302	request = ring->preallocated_lazy_request;
2303	if (WARN_ON(request == NULL))
2304		return -ENOMEM;
2305
2306	/* Record the position of the start of the request so that
2307	 * should we detect the updated seqno part-way through the
2308	 * GPU processing the request, we never over-estimate the
2309	 * position of the head.
2310	 */
2311	request_ring_position = intel_ring_get_tail(ring);
2312
2313	ret = ring->add_request(ring);
2314	if (ret)
2315		return ret;
2316
2317	request->seqno = intel_ring_get_seqno(ring);
2318	request->ring = ring;
2319	request->head = request_start;
2320	request->tail = request_ring_position;
2321
2322	/* Whilst this request exists, batch_obj will be on the
2323	 * active_list, and so will hold the active reference. Only when this
2324	 * request is retired will the the batch_obj be moved onto the
2325	 * inactive_list and lose its active reference. Hence we do not need
2326	 * to explicitly hold another reference here.
2327	 */
2328	request->batch_obj = obj;
2329
2330	/* Hold a reference to the current context so that we can inspect
2331	 * it later in case a hangcheck error event fires.
2332	 */
2333	request->ctx = ring->last_context;
2334	if (request->ctx)
2335		i915_gem_context_reference(request->ctx);
2336
2337	request->emitted_jiffies = jiffies;
2338	list_add_tail(&request->list, &ring->request_list);
2339	request->file_priv = NULL;
2340
2341	if (file) {
2342		struct drm_i915_file_private *file_priv = file->driver_priv;
2343
2344		spin_lock(&file_priv->mm.lock);
2345		request->file_priv = file_priv;
2346		list_add_tail(&request->client_list,
2347			      &file_priv->mm.request_list);
2348		spin_unlock(&file_priv->mm.lock);
2349	}
2350
2351	trace_i915_gem_request_add(ring, request->seqno);
2352	ring->outstanding_lazy_seqno = 0;
2353	ring->preallocated_lazy_request = NULL;
2354
2355	if (!dev_priv->ums.mm_suspended) {
2356		i915_queue_hangcheck(ring->dev);
2357
2358		cancel_delayed_work_sync(&dev_priv->mm.idle_work);
2359		queue_delayed_work(dev_priv->wq,
2360				   &dev_priv->mm.retire_work,
2361				   round_jiffies_up_relative(HZ));
2362		intel_mark_busy(dev_priv->dev);
2363	}
2364
2365	if (out_seqno)
2366		*out_seqno = request->seqno;
2367	return 0;
2368}
2369
2370static inline void
2371i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
2372{
2373	struct drm_i915_file_private *file_priv = request->file_priv;
2374
2375	if (!file_priv)
2376		return;
2377
2378	spin_lock(&file_priv->mm.lock);
2379	list_del(&request->client_list);
2380	request->file_priv = NULL;
2381	spin_unlock(&file_priv->mm.lock);
2382}
2383
2384static bool i915_context_is_banned(struct drm_i915_private *dev_priv,
2385				   const struct i915_hw_context *ctx)
2386{
2387	unsigned long elapsed;
2388
2389	elapsed = get_seconds() - ctx->hang_stats.guilty_ts;
2390
2391	if (ctx->hang_stats.banned)
2392		return true;
2393
2394	if (elapsed <= DRM_I915_CTX_BAN_PERIOD) {
2395		if (!i915_gem_context_is_default(ctx)) {
2396			DRM_DEBUG("context hanging too fast, banning!\n");
2397			return true;
2398		} else if (dev_priv->gpu_error.stop_rings == 0) {
2399			DRM_ERROR("gpu hanging too fast, banning!\n");
2400			return true;
2401		}
2402	}
2403
2404	return false;
2405}
2406
2407static void i915_set_reset_status(struct drm_i915_private *dev_priv,
2408				  struct i915_hw_context *ctx,
2409				  const bool guilty)
2410{
2411	struct i915_ctx_hang_stats *hs;
2412
2413	if (WARN_ON(!ctx))
2414		return;
2415
2416	hs = &ctx->hang_stats;
2417
2418	if (guilty) {
2419		hs->banned = i915_context_is_banned(dev_priv, ctx);
2420		hs->batch_active++;
2421		hs->guilty_ts = get_seconds();
2422	} else {
2423		hs->batch_pending++;
2424	}
2425}
2426
2427static void i915_gem_free_request(struct drm_i915_gem_request *request)
2428{
2429	list_del(&request->list);
2430	i915_gem_request_remove_from_client(request);
2431
2432	if (request->ctx)
2433		i915_gem_context_unreference(request->ctx);
2434
2435	kfree(request);
2436}
2437
2438struct drm_i915_gem_request *
2439i915_gem_find_active_request(struct intel_ring_buffer *ring)
2440{
2441	struct drm_i915_gem_request *request;
2442	u32 completed_seqno;
2443
2444	completed_seqno = ring->get_seqno(ring, false);
2445
2446	list_for_each_entry(request, &ring->request_list, list) {
2447		if (i915_seqno_passed(completed_seqno, request->seqno))
2448			continue;
2449
2450		return request;
2451	}
2452
2453	return NULL;
2454}
2455
2456static void i915_gem_reset_ring_status(struct drm_i915_private *dev_priv,
2457				       struct intel_ring_buffer *ring)
2458{
2459	struct drm_i915_gem_request *request;
2460	bool ring_hung;
2461
2462	request = i915_gem_find_active_request(ring);
2463
2464	if (request == NULL)
2465		return;
2466
2467	ring_hung = ring->hangcheck.score >= HANGCHECK_SCORE_RING_HUNG;
2468
2469	i915_set_reset_status(dev_priv, request->ctx, ring_hung);
2470
2471	list_for_each_entry_continue(request, &ring->request_list, list)
2472		i915_set_reset_status(dev_priv, request->ctx, false);
2473}
2474
2475static void i915_gem_reset_ring_cleanup(struct drm_i915_private *dev_priv,
2476					struct intel_ring_buffer *ring)
2477{
2478	while (!list_empty(&ring->active_list)) {
2479		struct drm_i915_gem_object *obj;
2480
2481		obj = list_first_entry(&ring->active_list,
2482				       struct drm_i915_gem_object,
2483				       ring_list);
2484
2485		i915_gem_object_move_to_inactive(obj);
2486	}
2487
2488	/*
2489	 * We must free the requests after all the corresponding objects have
2490	 * been moved off active lists. Which is the same order as the normal
2491	 * retire_requests function does. This is important if object hold
2492	 * implicit references on things like e.g. ppgtt address spaces through
2493	 * the request.
2494	 */
2495	while (!list_empty(&ring->request_list)) {
2496		struct drm_i915_gem_request *request;
2497
2498		request = list_first_entry(&ring->request_list,
2499					   struct drm_i915_gem_request,
2500					   list);
2501
2502		i915_gem_free_request(request);
2503	}
2504}
2505
2506void i915_gem_restore_fences(struct drm_device *dev)
2507{
2508	struct drm_i915_private *dev_priv = dev->dev_private;
2509	int i;
2510
2511	for (i = 0; i < dev_priv->num_fence_regs; i++) {
2512		struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
2513
 
 
2514		/*
2515		 * Commit delayed tiling changes if we have an object still
2516		 * attached to the fence, otherwise just clear the fence.
 
 
 
 
2517		 */
2518		if (reg->obj) {
2519			i915_gem_object_update_fence(reg->obj, reg,
2520						     reg->obj->tiling_mode);
2521		} else {
2522			i915_gem_write_fence(dev, i, NULL);
2523		}
2524	}
2525}
2526
2527void i915_gem_reset(struct drm_device *dev)
2528{
2529	struct drm_i915_private *dev_priv = dev->dev_private;
2530	struct intel_ring_buffer *ring;
2531	int i;
2532
2533	/*
2534	 * Before we free the objects from the requests, we need to inspect
2535	 * them for finding the guilty party. As the requests only borrow
2536	 * their reference to the objects, the inspection must be done first.
2537	 */
2538	for_each_ring(ring, dev_priv, i)
2539		i915_gem_reset_ring_status(dev_priv, ring);
2540
2541	for_each_ring(ring, dev_priv, i)
2542		i915_gem_reset_ring_cleanup(dev_priv, ring);
2543
2544	i915_gem_cleanup_ringbuffer(dev);
2545
2546	i915_gem_context_reset(dev);
2547
2548	i915_gem_restore_fences(dev);
2549}
2550
2551/**
2552 * This function clears the request list as sequence numbers are passed.
2553 */
2554static void
2555i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
2556{
2557	uint32_t seqno;
2558
2559	if (list_empty(&ring->request_list))
2560		return;
2561
2562	WARN_ON(i915_verify_lists(ring->dev));
2563
2564	seqno = ring->get_seqno(ring, true);
2565
2566	/* Move any buffers on the active list that are no longer referenced
2567	 * by the ringbuffer to the flushing/inactive lists as appropriate,
2568	 * before we free the context associated with the requests.
2569	 */
2570	while (!list_empty(&ring->active_list)) {
2571		struct drm_i915_gem_object *obj;
2572
2573		obj = list_first_entry(&ring->active_list,
2574				      struct drm_i915_gem_object,
2575				      ring_list);
2576
2577		if (!i915_seqno_passed(seqno, obj->last_read_seqno))
2578			break;
2579
2580		i915_gem_object_move_to_inactive(obj);
2581	}
2582
2583
2584	while (!list_empty(&ring->request_list)) {
2585		struct drm_i915_gem_request *request;
2586
2587		request = list_first_entry(&ring->request_list,
2588					   struct drm_i915_gem_request,
2589					   list);
2590
2591		if (!i915_seqno_passed(seqno, request->seqno))
2592			break;
2593
2594		trace_i915_gem_request_retire(ring, request->seqno);
2595		/* We know the GPU must have read the request to have
2596		 * sent us the seqno + interrupt, so use the position
2597		 * of tail of the request to update the last known position
2598		 * of the GPU head.
2599		 */
2600		ring->last_retired_head = request->tail;
2601
2602		i915_gem_free_request(request);
2603	}
2604
2605	if (unlikely(ring->trace_irq_seqno &&
2606		     i915_seqno_passed(seqno, ring->trace_irq_seqno))) {
2607		ring->irq_put(ring);
2608		ring->trace_irq_seqno = 0;
2609	}
2610
2611	WARN_ON(i915_verify_lists(ring->dev));
2612}
2613
2614bool
2615i915_gem_retire_requests(struct drm_device *dev)
2616{
2617	struct drm_i915_private *dev_priv = dev->dev_private;
2618	struct intel_ring_buffer *ring;
2619	bool idle = true;
2620	int i;
2621
2622	for_each_ring(ring, dev_priv, i) {
2623		i915_gem_retire_requests_ring(ring);
2624		idle &= list_empty(&ring->request_list);
2625	}
2626
2627	if (idle)
2628		mod_delayed_work(dev_priv->wq,
2629				   &dev_priv->mm.idle_work,
2630				   msecs_to_jiffies(100));
2631
2632	return idle;
2633}
2634
2635static void
2636i915_gem_retire_work_handler(struct work_struct *work)
2637{
2638	struct drm_i915_private *dev_priv =
2639		container_of(work, typeof(*dev_priv), mm.retire_work.work);
2640	struct drm_device *dev = dev_priv->dev;
2641	bool idle;
2642
2643	/* Come back later if the device is busy... */
2644	idle = false;
2645	if (mutex_trylock(&dev->struct_mutex)) {
2646		idle = i915_gem_retire_requests(dev);
2647		mutex_unlock(&dev->struct_mutex);
2648	}
2649	if (!idle)
2650		queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
2651				   round_jiffies_up_relative(HZ));
2652}
2653
2654static void
2655i915_gem_idle_work_handler(struct work_struct *work)
2656{
2657	struct drm_i915_private *dev_priv =
2658		container_of(work, typeof(*dev_priv), mm.idle_work.work);
2659
2660	intel_mark_idle(dev_priv->dev);
2661}
2662
2663/**
2664 * Ensures that an object will eventually get non-busy by flushing any required
2665 * write domains, emitting any outstanding lazy request and retiring and
2666 * completed requests.
2667 */
2668static int
2669i915_gem_object_flush_active(struct drm_i915_gem_object *obj)
2670{
2671	int ret;
2672
2673	if (obj->active) {
2674		ret = i915_gem_check_olr(obj->ring, obj->last_read_seqno);
2675		if (ret)
2676			return ret;
2677
2678		i915_gem_retire_requests_ring(obj->ring);
2679	}
2680
2681	return 0;
2682}
2683
2684/**
2685 * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
2686 * @DRM_IOCTL_ARGS: standard ioctl arguments
2687 *
2688 * Returns 0 if successful, else an error is returned with the remaining time in
2689 * the timeout parameter.
2690 *  -ETIME: object is still busy after timeout
2691 *  -ERESTARTSYS: signal interrupted the wait
2692 *  -ENONENT: object doesn't exist
2693 * Also possible, but rare:
2694 *  -EAGAIN: GPU wedged
2695 *  -ENOMEM: damn
2696 *  -ENODEV: Internal IRQ fail
2697 *  -E?: The add request failed
2698 *
2699 * The wait ioctl with a timeout of 0 reimplements the busy ioctl. With any
2700 * non-zero timeout parameter the wait ioctl will wait for the given number of
2701 * nanoseconds on an object becoming unbusy. Since the wait itself does so
2702 * without holding struct_mutex the object may become re-busied before this
2703 * function completes. A similar but shorter * race condition exists in the busy
2704 * ioctl
2705 */
2706int
2707i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
2708{
2709	struct drm_i915_private *dev_priv = dev->dev_private;
2710	struct drm_i915_gem_wait *args = data;
2711	struct drm_i915_gem_object *obj;
2712	struct intel_ring_buffer *ring = NULL;
2713	struct timespec timeout_stack, *timeout = NULL;
2714	unsigned reset_counter;
2715	u32 seqno = 0;
2716	int ret = 0;
2717
2718	if (args->timeout_ns >= 0) {
2719		timeout_stack = ns_to_timespec(args->timeout_ns);
2720		timeout = &timeout_stack;
2721	}
2722
2723	ret = i915_mutex_lock_interruptible(dev);
2724	if (ret)
2725		return ret;
2726
2727	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->bo_handle));
2728	if (&obj->base == NULL) {
2729		mutex_unlock(&dev->struct_mutex);
2730		return -ENOENT;
2731	}
2732
2733	/* Need to make sure the object gets inactive eventually. */
2734	ret = i915_gem_object_flush_active(obj);
2735	if (ret)
2736		goto out;
2737
2738	if (obj->active) {
2739		seqno = obj->last_read_seqno;
2740		ring = obj->ring;
2741	}
2742
2743	if (seqno == 0)
2744		 goto out;
2745
2746	/* Do this after OLR check to make sure we make forward progress polling
2747	 * on this IOCTL with a 0 timeout (like busy ioctl)
2748	 */
2749	if (!args->timeout_ns) {
2750		ret = -ETIME;
2751		goto out;
2752	}
2753
2754	drm_gem_object_unreference(&obj->base);
2755	reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
2756	mutex_unlock(&dev->struct_mutex);
2757
2758	ret = __wait_seqno(ring, seqno, reset_counter, true, timeout, file->driver_priv);
2759	if (timeout)
2760		args->timeout_ns = timespec_to_ns(timeout);
2761	return ret;
2762
2763out:
2764	drm_gem_object_unreference(&obj->base);
2765	mutex_unlock(&dev->struct_mutex);
2766	return ret;
2767}
2768
2769/**
2770 * i915_gem_object_sync - sync an object to a ring.
2771 *
2772 * @obj: object which may be in use on another ring.
2773 * @to: ring we wish to use the object on. May be NULL.
2774 *
2775 * This code is meant to abstract object synchronization with the GPU.
2776 * Calling with NULL implies synchronizing the object with the CPU
2777 * rather than a particular GPU ring.
2778 *
2779 * Returns 0 if successful, else propagates up the lower layer error.
2780 */
2781int
2782i915_gem_object_sync(struct drm_i915_gem_object *obj,
2783		     struct intel_ring_buffer *to)
2784{
2785	struct intel_ring_buffer *from = obj->ring;
2786	u32 seqno;
2787	int ret, idx;
2788
2789	if (from == NULL || to == from)
2790		return 0;
2791
2792	if (to == NULL || !i915_semaphore_is_enabled(obj->base.dev))
2793		return i915_gem_object_wait_rendering(obj, false);
2794
2795	idx = intel_ring_sync_index(from, to);
2796
2797	seqno = obj->last_read_seqno;
2798	if (seqno <= from->sync_seqno[idx])
2799		return 0;
2800
2801	ret = i915_gem_check_olr(obj->ring, seqno);
2802	if (ret)
2803		return ret;
2804
2805	trace_i915_gem_ring_sync_to(from, to, seqno);
2806	ret = to->sync_to(to, from, seqno);
2807	if (!ret)
2808		/* We use last_read_seqno because sync_to()
2809		 * might have just caused seqno wrap under
2810		 * the radar.
2811		 */
2812		from->sync_seqno[idx] = obj->last_read_seqno;
2813
2814	return ret;
2815}
2816
2817static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
2818{
2819	u32 old_write_domain, old_read_domains;
2820
2821	/* Force a pagefault for domain tracking on next user access */
2822	i915_gem_release_mmap(obj);
2823
2824	if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
2825		return;
2826
2827	/* Wait for any direct GTT access to complete */
2828	mb();
2829
2830	old_read_domains = obj->base.read_domains;
2831	old_write_domain = obj->base.write_domain;
2832
2833	obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
2834	obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
2835
2836	trace_i915_gem_object_change_domain(obj,
2837					    old_read_domains,
2838					    old_write_domain);
2839}
2840
2841int i915_vma_unbind(struct i915_vma *vma)
2842{
2843	struct drm_i915_gem_object *obj = vma->obj;
2844	struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2845	int ret;
2846
2847	if (list_empty(&vma->vma_link))
2848		return 0;
2849
2850	if (!drm_mm_node_allocated(&vma->node)) {
2851		i915_gem_vma_destroy(vma);
2852		return 0;
2853	}
2854
2855	if (vma->pin_count)
2856		return -EBUSY;
2857
2858	BUG_ON(obj->pages == NULL);
2859
2860	ret = i915_gem_object_finish_gpu(obj);
2861	if (ret)
2862		return ret;
2863	/* Continue on if we fail due to EIO, the GPU is hung so we
2864	 * should be safe and we need to cleanup or else we might
2865	 * cause memory corruption through use-after-free.
2866	 */
2867
2868	i915_gem_object_finish_gtt(obj);
2869
2870	/* release the fence reg _after_ flushing */
2871	ret = i915_gem_object_put_fence(obj);
2872	if (ret)
2873		return ret;
2874
2875	trace_i915_vma_unbind(vma);
2876
2877	vma->unbind_vma(vma);
2878
2879	i915_gem_gtt_finish_object(obj);
2880
2881	list_del_init(&vma->mm_list);
2882	/* Avoid an unnecessary call to unbind on rebind. */
2883	if (i915_is_ggtt(vma->vm))
2884		obj->map_and_fenceable = true;
2885
2886	drm_mm_remove_node(&vma->node);
2887	i915_gem_vma_destroy(vma);
2888
2889	/* Since the unbound list is global, only move to that list if
2890	 * no more VMAs exist. */
2891	if (list_empty(&obj->vma_list))
2892		list_move_tail(&obj->global_list, &dev_priv->mm.unbound_list);
2893
2894	/* And finally now the object is completely decoupled from this vma,
2895	 * we can drop its hold on the backing storage and allow it to be
2896	 * reaped by the shrinker.
2897	 */
2898	i915_gem_object_unpin_pages(obj);
2899
2900	return 0;
2901}
2902
2903int i915_gpu_idle(struct drm_device *dev)
2904{
2905	struct drm_i915_private *dev_priv = dev->dev_private;
2906	struct intel_ring_buffer *ring;
2907	int ret, i;
2908
2909	/* Flush everything onto the inactive list. */
2910	for_each_ring(ring, dev_priv, i) {
2911		ret = i915_switch_context(ring, ring->default_context);
2912		if (ret)
2913			return ret;
2914
2915		ret = intel_ring_idle(ring);
2916		if (ret)
2917			return ret;
2918	}
2919
2920	return 0;
2921}
2922
2923static void i965_write_fence_reg(struct drm_device *dev, int reg,
2924				 struct drm_i915_gem_object *obj)
2925{
2926	struct drm_i915_private *dev_priv = dev->dev_private;
2927	int fence_reg;
2928	int fence_pitch_shift;
2929
2930	if (INTEL_INFO(dev)->gen >= 6) {
2931		fence_reg = FENCE_REG_SANDYBRIDGE_0;
2932		fence_pitch_shift = SANDYBRIDGE_FENCE_PITCH_SHIFT;
2933	} else {
2934		fence_reg = FENCE_REG_965_0;
2935		fence_pitch_shift = I965_FENCE_PITCH_SHIFT;
2936	}
2937
2938	fence_reg += reg * 8;
2939
2940	/* To w/a incoherency with non-atomic 64-bit register updates,
2941	 * we split the 64-bit update into two 32-bit writes. In order
2942	 * for a partial fence not to be evaluated between writes, we
2943	 * precede the update with write to turn off the fence register,
2944	 * and only enable the fence as the last step.
2945	 *
2946	 * For extra levels of paranoia, we make sure each step lands
2947	 * before applying the next step.
2948	 */
2949	I915_WRITE(fence_reg, 0);
2950	POSTING_READ(fence_reg);
2951
2952	if (obj) {
2953		u32 size = i915_gem_obj_ggtt_size(obj);
2954		uint64_t val;
2955
2956		val = (uint64_t)((i915_gem_obj_ggtt_offset(obj) + size - 4096) &
2957				 0xfffff000) << 32;
2958		val |= i915_gem_obj_ggtt_offset(obj) & 0xfffff000;
2959		val |= (uint64_t)((obj->stride / 128) - 1) << fence_pitch_shift;
2960		if (obj->tiling_mode == I915_TILING_Y)
2961			val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2962		val |= I965_FENCE_REG_VALID;
2963
2964		I915_WRITE(fence_reg + 4, val >> 32);
2965		POSTING_READ(fence_reg + 4);
2966
2967		I915_WRITE(fence_reg + 0, val);
2968		POSTING_READ(fence_reg);
2969	} else {
2970		I915_WRITE(fence_reg + 4, 0);
2971		POSTING_READ(fence_reg + 4);
2972	}
2973}
2974
2975static void i915_write_fence_reg(struct drm_device *dev, int reg,
2976				 struct drm_i915_gem_object *obj)
2977{
2978	struct drm_i915_private *dev_priv = dev->dev_private;
2979	u32 val;
2980
2981	if (obj) {
2982		u32 size = i915_gem_obj_ggtt_size(obj);
2983		int pitch_val;
2984		int tile_width;
2985
2986		WARN((i915_gem_obj_ggtt_offset(obj) & ~I915_FENCE_START_MASK) ||
2987		     (size & -size) != size ||
2988		     (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
2989		     "object 0x%08lx [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
2990		     i915_gem_obj_ggtt_offset(obj), obj->map_and_fenceable, size);
2991
2992		if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
2993			tile_width = 128;
2994		else
2995			tile_width = 512;
2996
2997		/* Note: pitch better be a power of two tile widths */
2998		pitch_val = obj->stride / tile_width;
2999		pitch_val = ffs(pitch_val) - 1;
3000
3001		val = i915_gem_obj_ggtt_offset(obj);
3002		if (obj->tiling_mode == I915_TILING_Y)
3003			val |= 1 << I830_FENCE_TILING_Y_SHIFT;
3004		val |= I915_FENCE_SIZE_BITS(size);
3005		val |= pitch_val << I830_FENCE_PITCH_SHIFT;
3006		val |= I830_FENCE_REG_VALID;
3007	} else
3008		val = 0;
3009
3010	if (reg < 8)
3011		reg = FENCE_REG_830_0 + reg * 4;
3012	else
3013		reg = FENCE_REG_945_8 + (reg - 8) * 4;
3014
3015	I915_WRITE(reg, val);
3016	POSTING_READ(reg);
3017}
3018
3019static void i830_write_fence_reg(struct drm_device *dev, int reg,
3020				struct drm_i915_gem_object *obj)
3021{
3022	struct drm_i915_private *dev_priv = dev->dev_private;
3023	uint32_t val;
3024
3025	if (obj) {
3026		u32 size = i915_gem_obj_ggtt_size(obj);
3027		uint32_t pitch_val;
3028
3029		WARN((i915_gem_obj_ggtt_offset(obj) & ~I830_FENCE_START_MASK) ||
3030		     (size & -size) != size ||
3031		     (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
3032		     "object 0x%08lx not 512K or pot-size 0x%08x aligned\n",
3033		     i915_gem_obj_ggtt_offset(obj), size);
3034
3035		pitch_val = obj->stride / 128;
3036		pitch_val = ffs(pitch_val) - 1;
3037
3038		val = i915_gem_obj_ggtt_offset(obj);
3039		if (obj->tiling_mode == I915_TILING_Y)
3040			val |= 1 << I830_FENCE_TILING_Y_SHIFT;
3041		val |= I830_FENCE_SIZE_BITS(size);
3042		val |= pitch_val << I830_FENCE_PITCH_SHIFT;
3043		val |= I830_FENCE_REG_VALID;
3044	} else
3045		val = 0;
3046
3047	I915_WRITE(FENCE_REG_830_0 + reg * 4, val);
3048	POSTING_READ(FENCE_REG_830_0 + reg * 4);
3049}
3050
3051inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object *obj)
3052{
3053	return obj && obj->base.read_domains & I915_GEM_DOMAIN_GTT;
3054}
3055
3056static void i915_gem_write_fence(struct drm_device *dev, int reg,
3057				 struct drm_i915_gem_object *obj)
3058{
3059	struct drm_i915_private *dev_priv = dev->dev_private;
3060
3061	/* Ensure that all CPU reads are completed before installing a fence
3062	 * and all writes before removing the fence.
3063	 */
3064	if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj))
3065		mb();
3066
3067	WARN(obj && (!obj->stride || !obj->tiling_mode),
3068	     "bogus fence setup with stride: 0x%x, tiling mode: %i\n",
3069	     obj->stride, obj->tiling_mode);
3070
3071	switch (INTEL_INFO(dev)->gen) {
3072	case 8:
3073	case 7:
3074	case 6:
3075	case 5:
3076	case 4: i965_write_fence_reg(dev, reg, obj); break;
3077	case 3: i915_write_fence_reg(dev, reg, obj); break;
3078	case 2: i830_write_fence_reg(dev, reg, obj); break;
3079	default: BUG();
3080	}
3081
3082	/* And similarly be paranoid that no direct access to this region
3083	 * is reordered to before the fence is installed.
3084	 */
3085	if (i915_gem_object_needs_mb(obj))
3086		mb();
3087}
3088
3089static inline int fence_number(struct drm_i915_private *dev_priv,
3090			       struct drm_i915_fence_reg *fence)
3091{
3092	return fence - dev_priv->fence_regs;
3093}
3094
3095static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
3096					 struct drm_i915_fence_reg *fence,
3097					 bool enable)
3098{
3099	struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3100	int reg = fence_number(dev_priv, fence);
3101
3102	i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL);
3103
3104	if (enable) {
3105		obj->fence_reg = reg;
3106		fence->obj = obj;
3107		list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list);
3108	} else {
3109		obj->fence_reg = I915_FENCE_REG_NONE;
3110		fence->obj = NULL;
3111		list_del_init(&fence->lru_list);
3112	}
3113	obj->fence_dirty = false;
3114}
3115
3116static int
3117i915_gem_object_wait_fence(struct drm_i915_gem_object *obj)
3118{
3119	if (obj->last_fenced_seqno) {
3120		int ret = i915_wait_seqno(obj->ring, obj->last_fenced_seqno);
3121		if (ret)
3122			return ret;
3123
3124		obj->last_fenced_seqno = 0;
3125	}
3126
3127	obj->fenced_gpu_access = false;
3128	return 0;
3129}
3130
3131int
3132i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
3133{
3134	struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3135	struct drm_i915_fence_reg *fence;
3136	int ret;
3137
3138	ret = i915_gem_object_wait_fence(obj);
3139	if (ret)
3140		return ret;
3141
3142	if (obj->fence_reg == I915_FENCE_REG_NONE)
3143		return 0;
3144
3145	fence = &dev_priv->fence_regs[obj->fence_reg];
3146
3147	i915_gem_object_fence_lost(obj);
3148	i915_gem_object_update_fence(obj, fence, false);
3149
3150	return 0;
3151}
3152
3153static struct drm_i915_fence_reg *
3154i915_find_fence_reg(struct drm_device *dev)
3155{
3156	struct drm_i915_private *dev_priv = dev->dev_private;
3157	struct drm_i915_fence_reg *reg, *avail;
3158	int i;
3159
3160	/* First try to find a free reg */
3161	avail = NULL;
3162	for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
3163		reg = &dev_priv->fence_regs[i];
3164		if (!reg->obj)
3165			return reg;
3166
3167		if (!reg->pin_count)
3168			avail = reg;
3169	}
3170
3171	if (avail == NULL)
3172		goto deadlock;
3173
3174	/* None available, try to steal one or wait for a user to finish */
3175	list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
3176		if (reg->pin_count)
3177			continue;
3178
3179		return reg;
3180	}
3181
3182deadlock:
3183	/* Wait for completion of pending flips which consume fences */
3184	if (intel_has_pending_fb_unpin(dev))
3185		return ERR_PTR(-EAGAIN);
3186
3187	return ERR_PTR(-EDEADLK);
3188}
3189
3190/**
3191 * i915_gem_object_get_fence - set up fencing for an object
3192 * @obj: object to map through a fence reg
3193 *
3194 * When mapping objects through the GTT, userspace wants to be able to write
3195 * to them without having to worry about swizzling if the object is tiled.
3196 * This function walks the fence regs looking for a free one for @obj,
3197 * stealing one if it can't find any.
3198 *
3199 * It then sets up the reg based on the object's properties: address, pitch
3200 * and tiling format.
3201 *
3202 * For an untiled surface, this removes any existing fence.
3203 */
3204int
3205i915_gem_object_get_fence(struct drm_i915_gem_object *obj)
3206{
3207	struct drm_device *dev = obj->base.dev;
3208	struct drm_i915_private *dev_priv = dev->dev_private;
3209	bool enable = obj->tiling_mode != I915_TILING_NONE;
3210	struct drm_i915_fence_reg *reg;
3211	int ret;
3212
3213	/* Have we updated the tiling parameters upon the object and so
3214	 * will need to serialise the write to the associated fence register?
3215	 */
3216	if (obj->fence_dirty) {
3217		ret = i915_gem_object_wait_fence(obj);
3218		if (ret)
3219			return ret;
3220	}
3221
3222	/* Just update our place in the LRU if our fence is getting reused. */
3223	if (obj->fence_reg != I915_FENCE_REG_NONE) {
3224		reg = &dev_priv->fence_regs[obj->fence_reg];
3225		if (!obj->fence_dirty) {
3226			list_move_tail(&reg->lru_list,
3227				       &dev_priv->mm.fence_list);
3228			return 0;
3229		}
3230	} else if (enable) {
3231		reg = i915_find_fence_reg(dev);
3232		if (IS_ERR(reg))
3233			return PTR_ERR(reg);
3234
3235		if (reg->obj) {
3236			struct drm_i915_gem_object *old = reg->obj;
3237
3238			ret = i915_gem_object_wait_fence(old);
3239			if (ret)
3240				return ret;
3241
3242			i915_gem_object_fence_lost(old);
 
 
3243		}
3244	} else
3245		return 0;
3246
3247	i915_gem_object_update_fence(obj, reg, enable);
3248
3249	return 0;
3250}
3251
3252static bool i915_gem_valid_gtt_space(struct drm_device *dev,
3253				     struct drm_mm_node *gtt_space,
3254				     unsigned long cache_level)
3255{
3256	struct drm_mm_node *other;
3257
3258	/* On non-LLC machines we have to be careful when putting differing
3259	 * types of snoopable memory together to avoid the prefetcher
3260	 * crossing memory domains and dying.
3261	 */
3262	if (HAS_LLC(dev))
3263		return true;
3264
3265	if (!drm_mm_node_allocated(gtt_space))
3266		return true;
3267
3268	if (list_empty(&gtt_space->node_list))
3269		return true;
3270
3271	other = list_entry(gtt_space->node_list.prev, struct drm_mm_node, node_list);
3272	if (other->allocated && !other->hole_follows && other->color != cache_level)
3273		return false;
3274
3275	other = list_entry(gtt_space->node_list.next, struct drm_mm_node, node_list);
3276	if (other->allocated && !gtt_space->hole_follows && other->color != cache_level)
3277		return false;
3278
3279	return true;
3280}
3281
3282static void i915_gem_verify_gtt(struct drm_device *dev)
3283{
3284#if WATCH_GTT
3285	struct drm_i915_private *dev_priv = dev->dev_private;
3286	struct drm_i915_gem_object *obj;
3287	int err = 0;
3288
3289	list_for_each_entry(obj, &dev_priv->mm.gtt_list, global_list) {
3290		if (obj->gtt_space == NULL) {
3291			printk(KERN_ERR "object found on GTT list with no space reserved\n");
3292			err++;
3293			continue;
3294		}
3295
3296		if (obj->cache_level != obj->gtt_space->color) {
3297			printk(KERN_ERR "object reserved space [%08lx, %08lx] with wrong color, cache_level=%x, color=%lx\n",
3298			       i915_gem_obj_ggtt_offset(obj),
3299			       i915_gem_obj_ggtt_offset(obj) + i915_gem_obj_ggtt_size(obj),
3300			       obj->cache_level,
3301			       obj->gtt_space->color);
3302			err++;
3303			continue;
3304		}
3305
3306		if (!i915_gem_valid_gtt_space(dev,
3307					      obj->gtt_space,
3308					      obj->cache_level)) {
3309			printk(KERN_ERR "invalid GTT space found at [%08lx, %08lx] - color=%x\n",
3310			       i915_gem_obj_ggtt_offset(obj),
3311			       i915_gem_obj_ggtt_offset(obj) + i915_gem_obj_ggtt_size(obj),
3312			       obj->cache_level);
3313			err++;
3314			continue;
3315		}
3316	}
3317
3318	WARN_ON(err);
3319#endif
3320}
3321
3322/**
3323 * Finds free space in the GTT aperture and binds the object there.
3324 */
3325static struct i915_vma *
3326i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj,
3327			   struct i915_address_space *vm,
3328			   unsigned alignment,
3329			   uint64_t flags)
3330{
3331	struct drm_device *dev = obj->base.dev;
3332	struct drm_i915_private *dev_priv = dev->dev_private;
3333	u32 size, fence_size, fence_alignment, unfenced_alignment;
3334	unsigned long start =
3335		flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
3336	unsigned long end =
3337		flags & PIN_MAPPABLE ? dev_priv->gtt.mappable_end : vm->total;
3338	struct i915_vma *vma;
3339	int ret;
3340
3341	fence_size = i915_gem_get_gtt_size(dev,
3342					   obj->base.size,
3343					   obj->tiling_mode);
3344	fence_alignment = i915_gem_get_gtt_alignment(dev,
3345						     obj->base.size,
3346						     obj->tiling_mode, true);
3347	unfenced_alignment =
3348		i915_gem_get_gtt_alignment(dev,
3349					   obj->base.size,
3350					   obj->tiling_mode, false);
3351
3352	if (alignment == 0)
3353		alignment = flags & PIN_MAPPABLE ? fence_alignment :
3354						unfenced_alignment;
3355	if (flags & PIN_MAPPABLE && alignment & (fence_alignment - 1)) {
3356		DRM_DEBUG("Invalid object alignment requested %u\n", alignment);
3357		return ERR_PTR(-EINVAL);
3358	}
3359
3360	size = flags & PIN_MAPPABLE ? fence_size : obj->base.size;
3361
3362	/* If the object is bigger than the entire aperture, reject it early
3363	 * before evicting everything in a vain attempt to find space.
3364	 */
3365	if (obj->base.size > end) {
3366		DRM_DEBUG("Attempting to bind an object larger than the aperture: object=%zd > %s aperture=%lu\n",
3367			  obj->base.size,
3368			  flags & PIN_MAPPABLE ? "mappable" : "total",
3369			  end);
3370		return ERR_PTR(-E2BIG);
3371	}
3372
3373	ret = i915_gem_object_get_pages(obj);
3374	if (ret)
3375		return ERR_PTR(ret);
3376
3377	i915_gem_object_pin_pages(obj);
 
 
 
 
3378
3379	vma = i915_gem_obj_lookup_or_create_vma(obj, vm);
3380	if (IS_ERR(vma))
3381		goto err_unpin;
3382
3383search_free:
3384	ret = drm_mm_insert_node_in_range_generic(&vm->mm, &vma->node,
3385						  size, alignment,
3386						  obj->cache_level,
3387						  start, end,
3388						  DRM_MM_SEARCH_DEFAULT,
3389						  DRM_MM_CREATE_DEFAULT);
3390	if (ret) {
3391		ret = i915_gem_evict_something(dev, vm, size, alignment,
3392					       obj->cache_level,
3393					       start, end,
3394					       flags);
3395		if (ret == 0)
3396			goto search_free;
3397
3398		goto err_free_vma;
3399	}
3400	if (WARN_ON(!i915_gem_valid_gtt_space(dev, &vma->node,
3401					      obj->cache_level))) {
3402		ret = -EINVAL;
3403		goto err_remove_node;
3404	}
3405
3406	ret = i915_gem_gtt_prepare_object(obj);
3407	if (ret)
3408		goto err_remove_node;
3409
3410	list_move_tail(&obj->global_list, &dev_priv->mm.bound_list);
3411	list_add_tail(&vma->mm_list, &vm->inactive_list);
3412
3413	if (i915_is_ggtt(vm)) {
3414		bool mappable, fenceable;
3415
3416		fenceable = (vma->node.size == fence_size &&
3417			     (vma->node.start & (fence_alignment - 1)) == 0);
3418
3419		mappable = (vma->node.start + obj->base.size <=
3420			    dev_priv->gtt.mappable_end);
3421
3422		obj->map_and_fenceable = mappable && fenceable;
3423	}
3424
3425	WARN_ON(flags & PIN_MAPPABLE && !obj->map_and_fenceable);
3426
3427	trace_i915_vma_bind(vma, flags);
3428	vma->bind_vma(vma, obj->cache_level,
3429		      flags & (PIN_MAPPABLE | PIN_GLOBAL) ? GLOBAL_BIND : 0);
3430
3431	i915_gem_verify_gtt(dev);
3432	return vma;
3433
3434err_remove_node:
3435	drm_mm_remove_node(&vma->node);
3436err_free_vma:
3437	i915_gem_vma_destroy(vma);
3438	vma = ERR_PTR(ret);
3439err_unpin:
3440	i915_gem_object_unpin_pages(obj);
3441	return vma;
3442}
3443
3444bool
3445i915_gem_clflush_object(struct drm_i915_gem_object *obj,
3446			bool force)
3447{
3448	/* If we don't have a page list set up, then we're not pinned
3449	 * to GPU, and we can ignore the cache flush because it'll happen
3450	 * again at bind time.
3451	 */
3452	if (obj->pages == NULL)
3453		return false;
3454
3455	/*
3456	 * Stolen memory is always coherent with the GPU as it is explicitly
3457	 * marked as wc by the system, or the system is cache-coherent.
3458	 */
3459	if (obj->stolen)
3460		return false;
3461
3462	/* If the GPU is snooping the contents of the CPU cache,
3463	 * we do not need to manually clear the CPU cache lines.  However,
3464	 * the caches are only snooped when the render cache is
3465	 * flushed/invalidated.  As we always have to emit invalidations
3466	 * and flushes when moving into and out of the RENDER domain, correct
3467	 * snooping behaviour occurs naturally as the result of our domain
3468	 * tracking.
3469	 */
3470	if (!force && cpu_cache_is_coherent(obj->base.dev, obj->cache_level))
3471		return false;
3472
3473	trace_i915_gem_object_clflush(obj);
3474	drm_clflush_sg(obj->pages);
3475
3476	return true;
3477}
3478
3479/** Flushes the GTT write domain for the object if it's dirty. */
3480static void
3481i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
3482{
3483	uint32_t old_write_domain;
3484
3485	if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
3486		return;
3487
3488	/* No actual flushing is required for the GTT write domain.  Writes
3489	 * to it immediately go to main memory as far as we know, so there's
3490	 * no chipset flush.  It also doesn't land in render cache.
3491	 *
3492	 * However, we do have to enforce the order so that all writes through
3493	 * the GTT land before any writes to the device, such as updates to
3494	 * the GATT itself.
3495	 */
3496	wmb();
3497
3498	old_write_domain = obj->base.write_domain;
3499	obj->base.write_domain = 0;
3500
3501	trace_i915_gem_object_change_domain(obj,
3502					    obj->base.read_domains,
3503					    old_write_domain);
3504}
3505
3506/** Flushes the CPU write domain for the object if it's dirty. */
3507static void
3508i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj,
3509				       bool force)
3510{
3511	uint32_t old_write_domain;
3512
3513	if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
3514		return;
3515
3516	if (i915_gem_clflush_object(obj, force))
3517		i915_gem_chipset_flush(obj->base.dev);
3518
3519	old_write_domain = obj->base.write_domain;
3520	obj->base.write_domain = 0;
3521
3522	trace_i915_gem_object_change_domain(obj,
3523					    obj->base.read_domains,
3524					    old_write_domain);
3525}
3526
3527/**
3528 * Moves a single object to the GTT read, and possibly write domain.
3529 *
3530 * This function returns when the move is complete, including waiting on
3531 * flushes to occur.
3532 */
3533int
3534i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
3535{
3536	struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
3537	uint32_t old_write_domain, old_read_domains;
3538	int ret;
3539
3540	/* Not valid to be called on unbound objects. */
3541	if (!i915_gem_obj_bound_any(obj))
3542		return -EINVAL;
3543
3544	if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
3545		return 0;
3546
3547	ret = i915_gem_object_wait_rendering(obj, !write);
3548	if (ret)
3549		return ret;
3550
3551	i915_gem_object_flush_cpu_write_domain(obj, false);
3552
3553	/* Serialise direct access to this object with the barriers for
3554	 * coherent writes from the GPU, by effectively invalidating the
3555	 * GTT domain upon first access.
3556	 */
3557	if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
3558		mb();
3559
3560	old_write_domain = obj->base.write_domain;
3561	old_read_domains = obj->base.read_domains;
3562
3563	/* It should now be out of any other write domains, and we can update
3564	 * the domain values for our changes.
3565	 */
3566	BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3567	obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3568	if (write) {
3569		obj->base.read_domains = I915_GEM_DOMAIN_GTT;
3570		obj->base.write_domain = I915_GEM_DOMAIN_GTT;
3571		obj->dirty = 1;
3572	}
3573
3574	trace_i915_gem_object_change_domain(obj,
3575					    old_read_domains,
3576					    old_write_domain);
3577
3578	/* And bump the LRU for this access */
3579	if (i915_gem_object_is_inactive(obj)) {
3580		struct i915_vma *vma = i915_gem_obj_to_ggtt(obj);
3581		if (vma)
3582			list_move_tail(&vma->mm_list,
3583				       &dev_priv->gtt.base.inactive_list);
3584
3585	}
3586
3587	return 0;
3588}
3589
3590int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
3591				    enum i915_cache_level cache_level)
3592{
3593	struct drm_device *dev = obj->base.dev;
3594	struct i915_vma *vma, *next;
3595	int ret;
3596
3597	if (obj->cache_level == cache_level)
3598		return 0;
3599
3600	if (i915_gem_obj_is_pinned(obj)) {
3601		DRM_DEBUG("can not change the cache level of pinned objects\n");
3602		return -EBUSY;
3603	}
3604
3605	list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) {
3606		if (!i915_gem_valid_gtt_space(dev, &vma->node, cache_level)) {
3607			ret = i915_vma_unbind(vma);
3608			if (ret)
3609				return ret;
3610		}
3611	}
3612
3613	if (i915_gem_obj_bound_any(obj)) {
3614		ret = i915_gem_object_finish_gpu(obj);
3615		if (ret)
3616			return ret;
3617
3618		i915_gem_object_finish_gtt(obj);
3619
3620		/* Before SandyBridge, you could not use tiling or fence
3621		 * registers with snooped memory, so relinquish any fences
3622		 * currently pointing to our region in the aperture.
3623		 */
3624		if (INTEL_INFO(dev)->gen < 6) {
3625			ret = i915_gem_object_put_fence(obj);
3626			if (ret)
3627				return ret;
3628		}
3629
3630		list_for_each_entry(vma, &obj->vma_list, vma_link)
3631			if (drm_mm_node_allocated(&vma->node))
3632				vma->bind_vma(vma, cache_level,
3633					      obj->has_global_gtt_mapping ? GLOBAL_BIND : 0);
3634	}
3635
3636	list_for_each_entry(vma, &obj->vma_list, vma_link)
3637		vma->node.color = cache_level;
3638	obj->cache_level = cache_level;
3639
3640	if (cpu_write_needs_clflush(obj)) {
3641		u32 old_read_domains, old_write_domain;
3642
3643		/* If we're coming from LLC cached, then we haven't
3644		 * actually been tracking whether the data is in the
3645		 * CPU cache or not, since we only allow one bit set
3646		 * in obj->write_domain and have been skipping the clflushes.
3647		 * Just set it to the CPU cache for now.
3648		 */
3649		WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU);
3650
3651		old_read_domains = obj->base.read_domains;
3652		old_write_domain = obj->base.write_domain;
3653
3654		obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3655		obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3656
3657		trace_i915_gem_object_change_domain(obj,
3658						    old_read_domains,
3659						    old_write_domain);
3660	}
3661
3662	i915_gem_verify_gtt(dev);
3663	return 0;
3664}
3665
3666int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data,
3667			       struct drm_file *file)
3668{
3669	struct drm_i915_gem_caching *args = data;
3670	struct drm_i915_gem_object *obj;
3671	int ret;
3672
3673	ret = i915_mutex_lock_interruptible(dev);
3674	if (ret)
3675		return ret;
3676
3677	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3678	if (&obj->base == NULL) {
3679		ret = -ENOENT;
3680		goto unlock;
3681	}
3682
3683	switch (obj->cache_level) {
3684	case I915_CACHE_LLC:
3685	case I915_CACHE_L3_LLC:
3686		args->caching = I915_CACHING_CACHED;
3687		break;
3688
3689	case I915_CACHE_WT:
3690		args->caching = I915_CACHING_DISPLAY;
3691		break;
3692
3693	default:
3694		args->caching = I915_CACHING_NONE;
3695		break;
3696	}
3697
3698	drm_gem_object_unreference(&obj->base);
3699unlock:
3700	mutex_unlock(&dev->struct_mutex);
3701	return ret;
3702}
3703
3704int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
3705			       struct drm_file *file)
3706{
3707	struct drm_i915_gem_caching *args = data;
3708	struct drm_i915_gem_object *obj;
3709	enum i915_cache_level level;
3710	int ret;
3711
3712	switch (args->caching) {
3713	case I915_CACHING_NONE:
3714		level = I915_CACHE_NONE;
3715		break;
3716	case I915_CACHING_CACHED:
3717		level = I915_CACHE_LLC;
3718		break;
3719	case I915_CACHING_DISPLAY:
3720		level = HAS_WT(dev) ? I915_CACHE_WT : I915_CACHE_NONE;
3721		break;
3722	default:
3723		return -EINVAL;
3724	}
3725
3726	ret = i915_mutex_lock_interruptible(dev);
3727	if (ret)
3728		return ret;
3729
3730	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
3731	if (&obj->base == NULL) {
3732		ret = -ENOENT;
3733		goto unlock;
3734	}
3735
3736	ret = i915_gem_object_set_cache_level(obj, level);
3737
3738	drm_gem_object_unreference(&obj->base);
3739unlock:
3740	mutex_unlock(&dev->struct_mutex);
3741	return ret;
3742}
3743
3744static bool is_pin_display(struct drm_i915_gem_object *obj)
3745{
3746	/* There are 3 sources that pin objects:
3747	 *   1. The display engine (scanouts, sprites, cursors);
3748	 *   2. Reservations for execbuffer;
3749	 *   3. The user.
3750	 *
3751	 * We can ignore reservations as we hold the struct_mutex and
3752	 * are only called outside of the reservation path.  The user
3753	 * can only increment pin_count once, and so if after
3754	 * subtracting the potential reference by the user, any pin_count
3755	 * remains, it must be due to another use by the display engine.
3756	 */
3757	return i915_gem_obj_to_ggtt(obj)->pin_count - !!obj->user_pin_count;
3758}
3759
3760/*
3761 * Prepare buffer for display plane (scanout, cursors, etc).
3762 * Can be called from an uninterruptible phase (modesetting) and allows
3763 * any flushes to be pipelined (for pageflips).
3764 */
3765int
3766i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
3767				     u32 alignment,
3768				     struct intel_ring_buffer *pipelined)
3769{
3770	u32 old_read_domains, old_write_domain;
3771	int ret;
3772
3773	if (pipelined != obj->ring) {
3774		ret = i915_gem_object_sync(obj, pipelined);
3775		if (ret)
3776			return ret;
3777	}
3778
3779	/* Mark the pin_display early so that we account for the
3780	 * display coherency whilst setting up the cache domains.
3781	 */
3782	obj->pin_display = true;
3783
3784	/* The display engine is not coherent with the LLC cache on gen6.  As
3785	 * a result, we make sure that the pinning that is about to occur is
3786	 * done with uncached PTEs. This is lowest common denominator for all
3787	 * chipsets.
3788	 *
3789	 * However for gen6+, we could do better by using the GFDT bit instead
3790	 * of uncaching, which would allow us to flush all the LLC-cached data
3791	 * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3792	 */
3793	ret = i915_gem_object_set_cache_level(obj,
3794					      HAS_WT(obj->base.dev) ? I915_CACHE_WT : I915_CACHE_NONE);
3795	if (ret)
3796		goto err_unpin_display;
3797
3798	/* As the user may map the buffer once pinned in the display plane
3799	 * (e.g. libkms for the bootup splash), we have to ensure that we
3800	 * always use map_and_fenceable for all scanout buffers.
3801	 */
3802	ret = i915_gem_obj_ggtt_pin(obj, alignment, PIN_MAPPABLE);
3803	if (ret)
3804		goto err_unpin_display;
3805
3806	i915_gem_object_flush_cpu_write_domain(obj, true);
3807
3808	old_write_domain = obj->base.write_domain;
3809	old_read_domains = obj->base.read_domains;
3810
3811	/* It should now be out of any other write domains, and we can update
3812	 * the domain values for our changes.
3813	 */
3814	obj->base.write_domain = 0;
3815	obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3816
3817	trace_i915_gem_object_change_domain(obj,
3818					    old_read_domains,
3819					    old_write_domain);
3820
3821	return 0;
3822
3823err_unpin_display:
3824	obj->pin_display = is_pin_display(obj);
3825	return ret;
3826}
3827
3828void
3829i915_gem_object_unpin_from_display_plane(struct drm_i915_gem_object *obj)
3830{
3831	i915_gem_object_ggtt_unpin(obj);
3832	obj->pin_display = is_pin_display(obj);
3833}
3834
3835int
3836i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
3837{
3838	int ret;
3839
3840	if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
3841		return 0;
3842
3843	ret = i915_gem_object_wait_rendering(obj, false);
3844	if (ret)
3845		return ret;
3846
3847	/* Ensure that we invalidate the GPU's caches and TLBs. */
3848	obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
3849	return 0;
3850}
3851
3852/**
3853 * Moves a single object to the CPU read, and possibly write domain.
3854 *
3855 * This function returns when the move is complete, including waiting on
3856 * flushes to occur.
3857 */
3858int
3859i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
3860{
3861	uint32_t old_write_domain, old_read_domains;
3862	int ret;
3863
3864	if (obj->base.write_domain == I915_GEM_DOMAIN_CPU)
3865		return 0;
3866
3867	ret = i915_gem_object_wait_rendering(obj, !write);
3868	if (ret)
3869		return ret;
3870
3871	i915_gem_object_flush_gtt_write_domain(obj);
3872
3873	old_write_domain = obj->base.write_domain;
3874	old_read_domains = obj->base.read_domains;
3875
3876	/* Flush the CPU cache if it's still invalid. */
3877	if ((obj->base.read_domains & I915_GEM_DOMAIN_CPU) == 0) {
3878		i915_gem_clflush_object(obj, false);
3879
3880		obj->base.read_domains |= I915_GEM_DOMAIN_CPU;
3881	}
3882
3883	/* It should now be out of any other write domains, and we can update
3884	 * the domain values for our changes.
3885	 */
3886	BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3887
3888	/* If we're writing through the CPU, then the GPU read domains will
3889	 * need to be invalidated at next use.
3890	 */
3891	if (write) {
3892		obj->base.read_domains = I915_GEM_DOMAIN_CPU;
3893		obj->base.write_domain = I915_GEM_DOMAIN_CPU;
3894	}
3895
3896	trace_i915_gem_object_change_domain(obj,
3897					    old_read_domains,
3898					    old_write_domain);
3899
3900	return 0;
3901}
3902
3903/* Throttle our rendering by waiting until the ring has completed our requests
3904 * emitted over 20 msec ago.
3905 *
3906 * Note that if we were to use the current jiffies each time around the loop,
3907 * we wouldn't escape the function with any frames outstanding if the time to
3908 * render a frame was over 20ms.
3909 *
3910 * This should get us reasonable parallelism between CPU and GPU but also
3911 * relatively low latency when blocking on a particular request to finish.
3912 */
3913static int
3914i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
3915{
3916	struct drm_i915_private *dev_priv = dev->dev_private;
3917	struct drm_i915_file_private *file_priv = file->driver_priv;
3918	unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3919	struct drm_i915_gem_request *request;
3920	struct intel_ring_buffer *ring = NULL;
3921	unsigned reset_counter;
3922	u32 seqno = 0;
3923	int ret;
3924
3925	ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
3926	if (ret)
3927		return ret;
3928
3929	ret = i915_gem_check_wedge(&dev_priv->gpu_error, false);
3930	if (ret)
3931		return ret;
3932
3933	spin_lock(&file_priv->mm.lock);
3934	list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
3935		if (time_after_eq(request->emitted_jiffies, recent_enough))
3936			break;
3937
3938		ring = request->ring;
3939		seqno = request->seqno;
 
 
3940	}
3941	reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
3942	spin_unlock(&file_priv->mm.lock);
3943
3944	if (seqno == 0)
3945		return 0;
3946
3947	ret = __wait_seqno(ring, seqno, reset_counter, true, NULL, NULL);
3948	if (ret == 0)
3949		queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
3950
3951	return ret;
3952}
3953
3954static bool
3955i915_vma_misplaced(struct i915_vma *vma, uint32_t alignment, uint64_t flags)
3956{
3957	struct drm_i915_gem_object *obj = vma->obj;
3958
3959	if (alignment &&
3960	    vma->node.start & (alignment - 1))
3961		return true;
3962
3963	if (flags & PIN_MAPPABLE && !obj->map_and_fenceable)
3964		return true;
3965
3966	if (flags & PIN_OFFSET_BIAS &&
3967	    vma->node.start < (flags & PIN_OFFSET_MASK))
3968		return true;
3969
3970	return false;
3971}
3972
3973int
3974i915_gem_object_pin(struct drm_i915_gem_object *obj,
3975		    struct i915_address_space *vm,
3976		    uint32_t alignment,
3977		    uint64_t flags)
3978{
3979	struct i915_vma *vma;
3980	int ret;
3981
3982	if (WARN_ON(flags & (PIN_GLOBAL | PIN_MAPPABLE) && !i915_is_ggtt(vm)))
3983		return -EINVAL;
3984
3985	vma = i915_gem_obj_to_vma(obj, vm);
3986	if (vma) {
3987		if (WARN_ON(vma->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT))
3988			return -EBUSY;
3989
3990		if (i915_vma_misplaced(vma, alignment, flags)) {
3991			WARN(vma->pin_count,
3992			     "bo is already pinned with incorrect alignment:"
3993			     " offset=%lx, req.alignment=%x, req.map_and_fenceable=%d,"
3994			     " obj->map_and_fenceable=%d\n",
3995			     i915_gem_obj_offset(obj, vm), alignment,
3996			     !!(flags & PIN_MAPPABLE),
3997			     obj->map_and_fenceable);
3998			ret = i915_vma_unbind(vma);
3999			if (ret)
4000				return ret;
4001
4002			vma = NULL;
4003		}
4004	}
4005
4006	if (vma == NULL || !drm_mm_node_allocated(&vma->node)) {
4007		vma = i915_gem_object_bind_to_vm(obj, vm, alignment, flags);
4008		if (IS_ERR(vma))
4009			return PTR_ERR(vma);
4010	}
4011
4012	if (flags & PIN_GLOBAL && !obj->has_global_gtt_mapping)
4013		vma->bind_vma(vma, obj->cache_level, GLOBAL_BIND);
4014
4015	vma->pin_count++;
4016	if (flags & PIN_MAPPABLE)
4017		obj->pin_mappable |= true;
4018
4019	return 0;
4020}
4021
4022void
4023i915_gem_object_ggtt_unpin(struct drm_i915_gem_object *obj)
4024{
4025	struct i915_vma *vma = i915_gem_obj_to_ggtt(obj);
4026
4027	BUG_ON(!vma);
4028	BUG_ON(vma->pin_count == 0);
4029	BUG_ON(!i915_gem_obj_ggtt_bound(obj));
4030
4031	if (--vma->pin_count == 0)
4032		obj->pin_mappable = false;
4033}
4034
4035int
4036i915_gem_pin_ioctl(struct drm_device *dev, void *data,
4037		   struct drm_file *file)
4038{
4039	struct drm_i915_gem_pin *args = data;
4040	struct drm_i915_gem_object *obj;
4041	int ret;
4042
4043	if (INTEL_INFO(dev)->gen >= 6)
4044		return -ENODEV;
4045
4046	ret = i915_mutex_lock_interruptible(dev);
4047	if (ret)
4048		return ret;
4049
4050	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
4051	if (&obj->base == NULL) {
4052		ret = -ENOENT;
4053		goto unlock;
4054	}
4055
4056	if (obj->madv != I915_MADV_WILLNEED) {
4057		DRM_DEBUG("Attempting to pin a purgeable buffer\n");
4058		ret = -EFAULT;
4059		goto out;
4060	}
4061
4062	if (obj->pin_filp != NULL && obj->pin_filp != file) {
4063		DRM_DEBUG("Already pinned in i915_gem_pin_ioctl(): %d\n",
4064			  args->handle);
4065		ret = -EINVAL;
4066		goto out;
4067	}
4068
4069	if (obj->user_pin_count == ULONG_MAX) {
4070		ret = -EBUSY;
4071		goto out;
4072	}
4073
4074	if (obj->user_pin_count == 0) {
4075		ret = i915_gem_obj_ggtt_pin(obj, args->alignment, PIN_MAPPABLE);
4076		if (ret)
4077			goto out;
4078	}
4079
4080	obj->user_pin_count++;
4081	obj->pin_filp = file;
4082
4083	args->offset = i915_gem_obj_ggtt_offset(obj);
4084out:
4085	drm_gem_object_unreference(&obj->base);
4086unlock:
4087	mutex_unlock(&dev->struct_mutex);
4088	return ret;
4089}
4090
4091int
4092i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
4093		     struct drm_file *file)
4094{
4095	struct drm_i915_gem_pin *args = data;
4096	struct drm_i915_gem_object *obj;
4097	int ret;
4098
4099	ret = i915_mutex_lock_interruptible(dev);
4100	if (ret)
4101		return ret;
4102
4103	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
4104	if (&obj->base == NULL) {
4105		ret = -ENOENT;
4106		goto unlock;
4107	}
4108
4109	if (obj->pin_filp != file) {
4110		DRM_DEBUG("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
4111			  args->handle);
4112		ret = -EINVAL;
4113		goto out;
4114	}
4115	obj->user_pin_count--;
4116	if (obj->user_pin_count == 0) {
4117		obj->pin_filp = NULL;
4118		i915_gem_object_ggtt_unpin(obj);
4119	}
4120
4121out:
4122	drm_gem_object_unreference(&obj->base);
4123unlock:
4124	mutex_unlock(&dev->struct_mutex);
4125	return ret;
4126}
4127
4128int
4129i915_gem_busy_ioctl(struct drm_device *dev, void *data,
4130		    struct drm_file *file)
4131{
4132	struct drm_i915_gem_busy *args = data;
4133	struct drm_i915_gem_object *obj;
4134	int ret;
4135
4136	ret = i915_mutex_lock_interruptible(dev);
4137	if (ret)
4138		return ret;
4139
4140	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
4141	if (&obj->base == NULL) {
4142		ret = -ENOENT;
4143		goto unlock;
4144	}
4145
4146	/* Count all active objects as busy, even if they are currently not used
4147	 * by the gpu. Users of this interface expect objects to eventually
4148	 * become non-busy without any further actions, therefore emit any
4149	 * necessary flushes here.
4150	 */
4151	ret = i915_gem_object_flush_active(obj);
4152
4153	args->busy = obj->active;
4154	if (obj->ring) {
4155		BUILD_BUG_ON(I915_NUM_RINGS > 16);
4156		args->busy |= intel_ring_flag(obj->ring) << 16;
4157	}
4158
4159	drm_gem_object_unreference(&obj->base);
4160unlock:
4161	mutex_unlock(&dev->struct_mutex);
4162	return ret;
4163}
4164
4165int
4166i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
4167			struct drm_file *file_priv)
4168{
4169	return i915_gem_ring_throttle(dev, file_priv);
4170}
4171
4172int
4173i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
4174		       struct drm_file *file_priv)
4175{
 
4176	struct drm_i915_gem_madvise *args = data;
4177	struct drm_i915_gem_object *obj;
4178	int ret;
4179
4180	switch (args->madv) {
4181	case I915_MADV_DONTNEED:
4182	case I915_MADV_WILLNEED:
4183	    break;
4184	default:
4185	    return -EINVAL;
4186	}
4187
4188	ret = i915_mutex_lock_interruptible(dev);
4189	if (ret)
4190		return ret;
4191
4192	obj = to_intel_bo(drm_gem_object_lookup(dev, file_priv, args->handle));
4193	if (&obj->base == NULL) {
4194		ret = -ENOENT;
4195		goto unlock;
4196	}
4197
4198	if (i915_gem_obj_is_pinned(obj)) {
4199		ret = -EINVAL;
4200		goto out;
4201	}
4202
4203	if (obj->madv != __I915_MADV_PURGED)
4204		obj->madv = args->madv;
4205
4206	/* if the object is no longer attached, discard its backing storage */
4207	if (i915_gem_object_is_purgeable(obj) && obj->pages == NULL)
4208		i915_gem_object_truncate(obj);
4209
4210	args->retained = obj->madv != __I915_MADV_PURGED;
4211
4212out:
4213	drm_gem_object_unreference(&obj->base);
4214unlock:
4215	mutex_unlock(&dev->struct_mutex);
4216	return ret;
4217}
4218
4219void i915_gem_object_init(struct drm_i915_gem_object *obj,
4220			  const struct drm_i915_gem_object_ops *ops)
4221{
4222	INIT_LIST_HEAD(&obj->global_list);
4223	INIT_LIST_HEAD(&obj->ring_list);
4224	INIT_LIST_HEAD(&obj->obj_exec_link);
4225	INIT_LIST_HEAD(&obj->vma_list);
4226
4227	obj->ops = ops;
4228
4229	obj->fence_reg = I915_FENCE_REG_NONE;
4230	obj->madv = I915_MADV_WILLNEED;
4231	/* Avoid an unnecessary call to unbind on the first bind. */
4232	obj->map_and_fenceable = true;
4233
4234	i915_gem_info_add_obj(obj->base.dev->dev_private, obj->base.size);
4235}
4236
4237static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
4238	.get_pages = i915_gem_object_get_pages_gtt,
4239	.put_pages = i915_gem_object_put_pages_gtt,
4240};
4241
4242struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
4243						  size_t size)
4244{
4245	struct drm_i915_gem_object *obj;
4246	struct address_space *mapping;
4247	gfp_t mask;
4248
4249	obj = i915_gem_object_alloc(dev);
4250	if (obj == NULL)
4251		return NULL;
4252
4253	if (drm_gem_object_init(dev, &obj->base, size) != 0) {
4254		i915_gem_object_free(obj);
4255		return NULL;
4256	}
4257
4258	mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
4259	if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) {
4260		/* 965gm cannot relocate objects above 4GiB. */
4261		mask &= ~__GFP_HIGHMEM;
4262		mask |= __GFP_DMA32;
4263	}
4264
4265	mapping = file_inode(obj->base.filp)->i_mapping;
4266	mapping_set_gfp_mask(mapping, mask);
4267
4268	i915_gem_object_init(obj, &i915_gem_object_ops);
4269
4270	obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4271	obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4272
4273	if (HAS_LLC(dev)) {
4274		/* On some devices, we can have the GPU use the LLC (the CPU
4275		 * cache) for about a 10% performance improvement
4276		 * compared to uncached.  Graphics requests other than
4277		 * display scanout are coherent with the CPU in
4278		 * accessing this cache.  This means in this mode we
4279		 * don't need to clflush on the CPU side, and on the
4280		 * GPU side we only need to flush internal caches to
4281		 * get data visible to the CPU.
4282		 *
4283		 * However, we maintain the display planes as UC, and so
4284		 * need to rebind when first used as such.
4285		 */
4286		obj->cache_level = I915_CACHE_LLC;
4287	} else
4288		obj->cache_level = I915_CACHE_NONE;
4289
4290	trace_i915_gem_object_create(obj);
4291
4292	return obj;
4293}
4294
4295void i915_gem_free_object(struct drm_gem_object *gem_obj)
4296{
4297	struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);
4298	struct drm_device *dev = obj->base.dev;
4299	struct drm_i915_private *dev_priv = dev->dev_private;
4300	struct i915_vma *vma, *next;
4301
4302	intel_runtime_pm_get(dev_priv);
4303
4304	trace_i915_gem_object_destroy(obj);
 
 
4305
4306	list_for_each_entry_safe(vma, next, &obj->vma_list, vma_link) {
4307		int ret;
 
 
 
4308
4309		vma->pin_count = 0;
4310		ret = i915_vma_unbind(vma);
4311		if (WARN_ON(ret == -ERESTARTSYS)) {
4312			bool was_interruptible;
4313
4314			was_interruptible = dev_priv->mm.interruptible;
4315			dev_priv->mm.interruptible = false;
4316
4317			WARN_ON(i915_vma_unbind(vma));
4318
4319			dev_priv->mm.interruptible = was_interruptible;
4320		}
 
4321	}
4322
4323	i915_gem_object_detach_phys(obj);
 
 
 
4324
4325	/* Stolen objects don't hold a ref, but do hold pin count. Fix that up
4326	 * before progressing. */
4327	if (obj->stolen)
4328		i915_gem_object_unpin_pages(obj);
4329
4330	if (WARN_ON(obj->pages_pin_count))
4331		obj->pages_pin_count = 0;
4332	i915_gem_object_put_pages(obj);
4333	i915_gem_object_free_mmap_offset(obj);
4334	i915_gem_object_release_stolen(obj);
4335
4336	BUG_ON(obj->pages);
4337
4338	if (obj->base.import_attach)
4339		drm_prime_gem_destroy(&obj->base, NULL);
4340
4341	drm_gem_object_release(&obj->base);
4342	i915_gem_info_remove_obj(dev_priv, obj->base.size);
4343
4344	kfree(obj->bit_17);
4345	i915_gem_object_free(obj);
4346
4347	intel_runtime_pm_put(dev_priv);
4348}
4349
4350struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj,
4351				     struct i915_address_space *vm)
4352{
4353	struct i915_vma *vma;
4354	list_for_each_entry(vma, &obj->vma_list, vma_link)
4355		if (vma->vm == vm)
4356			return vma;
4357
4358	return NULL;
4359}
4360
4361void i915_gem_vma_destroy(struct i915_vma *vma)
4362{
4363	WARN_ON(vma->node.allocated);
4364
4365	/* Keep the vma as a placeholder in the execbuffer reservation lists */
4366	if (!list_empty(&vma->exec_list))
4367		return;
4368
4369	list_del(&vma->vma_link);
4370
4371	kfree(vma);
4372}
4373
4374int
4375i915_gem_suspend(struct drm_device *dev)
4376{
4377	struct drm_i915_private *dev_priv = dev->dev_private;
4378	int ret = 0;
4379
4380	mutex_lock(&dev->struct_mutex);
4381	if (dev_priv->ums.mm_suspended)
4382		goto err;
4383
4384	ret = i915_gpu_idle(dev);
4385	if (ret)
4386		goto err;
4387
4388	i915_gem_retire_requests(dev);
4389
4390	/* Under UMS, be paranoid and evict. */
4391	if (!drm_core_check_feature(dev, DRIVER_MODESET))
4392		i915_gem_evict_everything(dev);
4393
4394	i915_kernel_lost_context(dev);
4395	i915_gem_cleanup_ringbuffer(dev);
4396
4397	/* Hack!  Don't let anybody do execbuf while we don't control the chip.
4398	 * We need to replace this with a semaphore, or something.
4399	 * And not confound ums.mm_suspended!
4400	 */
4401	dev_priv->ums.mm_suspended = !drm_core_check_feature(dev,
4402							     DRIVER_MODESET);
4403	mutex_unlock(&dev->struct_mutex);
4404
4405	del_timer_sync(&dev_priv->gpu_error.hangcheck_timer);
4406	cancel_delayed_work_sync(&dev_priv->mm.retire_work);
4407	cancel_delayed_work_sync(&dev_priv->mm.idle_work);
4408
4409	return 0;
4410
4411err:
4412	mutex_unlock(&dev->struct_mutex);
4413	return ret;
4414}
4415
4416int i915_gem_l3_remap(struct intel_ring_buffer *ring, int slice)
4417{
4418	struct drm_device *dev = ring->dev;
4419	struct drm_i915_private *dev_priv = dev->dev_private;
4420	u32 reg_base = GEN7_L3LOG_BASE + (slice * 0x200);
4421	u32 *remap_info = dev_priv->l3_parity.remap_info[slice];
4422	int i, ret;
4423
4424	if (!HAS_L3_DPF(dev) || !remap_info)
4425		return 0;
4426
4427	ret = intel_ring_begin(ring, GEN7_L3LOG_SIZE / 4 * 3);
4428	if (ret)
4429		return ret;
4430
4431	/*
4432	 * Note: We do not worry about the concurrent register cacheline hang
4433	 * here because no other code should access these registers other than
4434	 * at initialization time.
4435	 */
4436	for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) {
4437		intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
4438		intel_ring_emit(ring, reg_base + i);
4439		intel_ring_emit(ring, remap_info[i/4]);
4440	}
4441
4442	intel_ring_advance(ring);
4443
4444	return ret;
4445}
4446
4447void i915_gem_init_swizzling(struct drm_device *dev)
 
 
 
 
 
 
 
 
 
4448{
4449	struct drm_i915_private *dev_priv = dev->dev_private;
4450
4451	if (INTEL_INFO(dev)->gen < 5 ||
4452	    dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
4453		return;
4454
4455	I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
4456				 DISP_TILE_SURFACE_SWIZZLING);
4457
4458	if (IS_GEN5(dev))
4459		return;
4460
4461	I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL);
4462	if (IS_GEN6(dev))
4463		I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
4464	else if (IS_GEN7(dev))
4465		I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
4466	else if (IS_GEN8(dev))
4467		I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW));
4468	else
4469		BUG();
4470}
4471
4472static bool
4473intel_enable_blt(struct drm_device *dev)
4474{
4475	if (!HAS_BLT(dev))
4476		return false;
4477
4478	/* The blitter was dysfunctional on early prototypes */
4479	if (IS_GEN6(dev) && dev->pdev->revision < 8) {
4480		DRM_INFO("BLT not supported on this pre-production hardware;"
4481			 " graphics performance will be degraded.\n");
4482		return false;
4483	}
4484
4485	return true;
4486}
4487
4488static int i915_gem_init_rings(struct drm_device *dev)
4489{
4490	struct drm_i915_private *dev_priv = dev->dev_private;
 
4491	int ret;
4492
4493	ret = intel_init_render_ring_buffer(dev);
4494	if (ret)
4495		return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4496
4497	if (HAS_BSD(dev)) {
4498		ret = intel_init_bsd_ring_buffer(dev);
4499		if (ret)
4500			goto cleanup_render_ring;
4501	}
4502
4503	if (intel_enable_blt(dev)) {
4504		ret = intel_init_blt_ring_buffer(dev);
4505		if (ret)
4506			goto cleanup_bsd_ring;
4507	}
 
 
 
 
 
4508
4509	if (HAS_VEBOX(dev)) {
4510		ret = intel_init_vebox_ring_buffer(dev);
4511		if (ret)
4512			goto cleanup_blt_ring;
4513	}
4514
4515
4516	ret = i915_gem_set_seqno(dev, ((u32)~0 - 0x1000));
4517	if (ret)
4518		goto cleanup_vebox_ring;
 
 
4519
4520	return 0;
4521
4522cleanup_vebox_ring:
4523	intel_cleanup_ring_buffer(&dev_priv->ring[VECS]);
4524cleanup_blt_ring:
4525	intel_cleanup_ring_buffer(&dev_priv->ring[BCS]);
4526cleanup_bsd_ring:
4527	intel_cleanup_ring_buffer(&dev_priv->ring[VCS]);
4528cleanup_render_ring:
4529	intel_cleanup_ring_buffer(&dev_priv->ring[RCS]);
4530
4531	return ret;
4532}
4533
4534int
4535i915_gem_init_hw(struct drm_device *dev)
4536{
4537	struct drm_i915_private *dev_priv = dev->dev_private;
4538	int ret, i;
4539
4540	if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt())
4541		return -EIO;
4542
4543	if (dev_priv->ellc_size)
4544		I915_WRITE(HSW_IDICR, I915_READ(HSW_IDICR) | IDIHASHMSK(0xf));
4545
4546	if (IS_HASWELL(dev))
4547		I915_WRITE(MI_PREDICATE_RESULT_2, IS_HSW_GT3(dev) ?
4548			   LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED);
4549
4550	if (HAS_PCH_NOP(dev)) {
4551		if (IS_IVYBRIDGE(dev)) {
4552			u32 temp = I915_READ(GEN7_MSG_CTL);
4553			temp &= ~(WAIT_FOR_PCH_FLR_ACK | WAIT_FOR_PCH_RESET_ACK);
4554			I915_WRITE(GEN7_MSG_CTL, temp);
4555		} else if (INTEL_INFO(dev)->gen >= 7) {
4556			u32 temp = I915_READ(HSW_NDE_RSTWRN_OPT);
4557			temp &= ~RESET_PCH_HANDSHAKE_ENABLE;
4558			I915_WRITE(HSW_NDE_RSTWRN_OPT, temp);
4559		}
4560	}
4561
4562	i915_gem_init_swizzling(dev);
4563
4564	ret = i915_gem_init_rings(dev);
4565	if (ret)
4566		return ret;
 
 
 
 
 
 
 
 
4567
4568	for (i = 0; i < NUM_L3_SLICES(dev); i++)
4569		i915_gem_l3_remap(&dev_priv->ring[RCS], i);
4570
4571	/*
4572	 * XXX: Contexts should only be initialized once. Doing a switch to the
4573	 * default context switch however is something we'd like to do after
4574	 * reset or thaw (the latter may not actually be necessary for HW, but
4575	 * goes with our code better). Context switching requires rings (for
4576	 * the do_switch), but before enabling PPGTT. So don't move this.
4577	 */
4578	ret = i915_gem_context_enable(dev_priv);
4579	if (ret) {
4580		DRM_ERROR("Context enable failed %d\n", ret);
4581		goto err_out;
4582	}
4583
4584	return 0;
4585
4586err_out:
4587	i915_gem_cleanup_ringbuffer(dev);
4588	return ret;
4589}
4590
4591int i915_gem_init(struct drm_device *dev)
4592{
4593	struct drm_i915_private *dev_priv = dev->dev_private;
4594	int ret;
4595
4596	mutex_lock(&dev->struct_mutex);
4597
4598	if (IS_VALLEYVIEW(dev)) {
4599		/* VLVA0 (potential hack), BIOS isn't actually waking us */
4600		I915_WRITE(VLV_GTLC_WAKE_CTRL, 1);
4601		if (wait_for((I915_READ(VLV_GTLC_PW_STATUS) & 1) == 1, 10))
4602			DRM_DEBUG_DRIVER("allow wake ack timed out\n");
4603	}
4604
4605	i915_gem_init_global_gtt(dev);
4606
4607	ret = i915_gem_context_init(dev);
4608	if (ret) {
4609		mutex_unlock(&dev->struct_mutex);
4610		return ret;
4611	}
4612
4613	ret = i915_gem_init_hw(dev);
4614	mutex_unlock(&dev->struct_mutex);
4615	if (ret) {
4616		WARN_ON(dev_priv->mm.aliasing_ppgtt);
4617		i915_gem_context_fini(dev);
4618		drm_mm_takedown(&dev_priv->gtt.base.mm);
4619		return ret;
4620	}
4621
4622	/* Allow hardware batchbuffers unless told otherwise, but not for KMS. */
4623	if (!drm_core_check_feature(dev, DRIVER_MODESET))
4624		dev_priv->dri1.allow_batchbuffer = 1;
4625	return 0;
4626}
4627
4628void
4629i915_gem_cleanup_ringbuffer(struct drm_device *dev)
4630{
4631	struct drm_i915_private *dev_priv = dev->dev_private;
4632	struct intel_ring_buffer *ring;
4633	int i;
4634
4635	for_each_ring(ring, dev_priv, i)
4636		intel_cleanup_ring_buffer(ring);
4637}
4638
4639int
4640i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
4641		       struct drm_file *file_priv)
4642{
4643	struct drm_i915_private *dev_priv = dev->dev_private;
4644	int ret;
4645
4646	if (drm_core_check_feature(dev, DRIVER_MODESET))
4647		return 0;
4648
4649	if (i915_reset_in_progress(&dev_priv->gpu_error)) {
4650		DRM_ERROR("Reenabling wedged hardware, good luck\n");
4651		atomic_set(&dev_priv->gpu_error.reset_counter, 0);
4652	}
4653
4654	mutex_lock(&dev->struct_mutex);
4655	dev_priv->ums.mm_suspended = 0;
4656
4657	ret = i915_gem_init_hw(dev);
4658	if (ret != 0) {
4659		mutex_unlock(&dev->struct_mutex);
4660		return ret;
4661	}
4662
4663	BUG_ON(!list_empty(&dev_priv->gtt.base.active_list));
4664	mutex_unlock(&dev->struct_mutex);
4665
4666	ret = drm_irq_install(dev);
4667	if (ret)
4668		goto cleanup_ringbuffer;
4669
4670	return 0;
4671
4672cleanup_ringbuffer:
4673	mutex_lock(&dev->struct_mutex);
4674	i915_gem_cleanup_ringbuffer(dev);
4675	dev_priv->ums.mm_suspended = 1;
4676	mutex_unlock(&dev->struct_mutex);
4677
4678	return ret;
4679}
4680
4681int
4682i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
4683		       struct drm_file *file_priv)
4684{
4685	if (drm_core_check_feature(dev, DRIVER_MODESET))
4686		return 0;
4687
4688	drm_irq_uninstall(dev);
 
 
 
4689
4690	return i915_gem_suspend(dev);
 
4691}
4692
4693void
4694i915_gem_lastclose(struct drm_device *dev)
4695{
4696	int ret;
 
4697
4698	if (drm_core_check_feature(dev, DRIVER_MODESET))
4699		return;
 
 
4700
4701	ret = i915_gem_suspend(dev);
4702	if (ret)
4703		DRM_ERROR("failed to idle hardware: %d\n", ret);
4704}
4705
4706static void
4707init_ring_lists(struct intel_ring_buffer *ring)
4708{
4709	INIT_LIST_HEAD(&ring->active_list);
4710	INIT_LIST_HEAD(&ring->request_list);
4711}
4712
4713void i915_init_vm(struct drm_i915_private *dev_priv,
4714		  struct i915_address_space *vm)
4715{
4716	if (!i915_is_ggtt(vm))
4717		drm_mm_init(&vm->mm, vm->start, vm->total);
4718	vm->dev = dev_priv->dev;
4719	INIT_LIST_HEAD(&vm->active_list);
4720	INIT_LIST_HEAD(&vm->inactive_list);
4721	INIT_LIST_HEAD(&vm->global_link);
4722	list_add_tail(&vm->global_link, &dev_priv->vm_list);
4723}
4724
4725void
4726i915_gem_load(struct drm_device *dev)
4727{
4728	struct drm_i915_private *dev_priv = dev->dev_private;
4729	int i;
4730
4731	dev_priv->slab =
4732		kmem_cache_create("i915_gem_object",
4733				  sizeof(struct drm_i915_gem_object), 0,
4734				  SLAB_HWCACHE_ALIGN,
4735				  NULL);
4736
4737	INIT_LIST_HEAD(&dev_priv->vm_list);
4738	i915_init_vm(dev_priv, &dev_priv->gtt.base);
4739
4740	INIT_LIST_HEAD(&dev_priv->context_list);
4741	INIT_LIST_HEAD(&dev_priv->mm.unbound_list);
4742	INIT_LIST_HEAD(&dev_priv->mm.bound_list);
4743	INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4744	for (i = 0; i < I915_NUM_RINGS; i++)
4745		init_ring_lists(&dev_priv->ring[i]);
4746	for (i = 0; i < I915_MAX_NUM_FENCES; i++)
4747		INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
4748	INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
4749			  i915_gem_retire_work_handler);
4750	INIT_DELAYED_WORK(&dev_priv->mm.idle_work,
4751			  i915_gem_idle_work_handler);
4752	init_waitqueue_head(&dev_priv->gpu_error.reset_queue);
4753
4754	/* On GEN3 we really need to make sure the ARB C3 LP bit is set */
4755	if (IS_GEN3(dev)) {
4756		I915_WRITE(MI_ARB_STATE,
4757			   _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
4758	}
4759
4760	dev_priv->relative_constants_mode = I915_EXEC_CONSTANTS_REL_GENERAL;
 
4761
4762	/* Old X drivers will take 0-2 for front, back, depth buffers */
4763	if (!drm_core_check_feature(dev, DRIVER_MODESET))
4764		dev_priv->fence_reg_start = 3;
4765
4766	if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev))
4767		dev_priv->num_fence_regs = 32;
4768	else if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
4769		dev_priv->num_fence_regs = 16;
4770	else
4771		dev_priv->num_fence_regs = 8;
4772
4773	/* Initialize fence registers to zero */
4774	INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4775	i915_gem_restore_fences(dev);
4776
4777	i915_gem_detect_bit_6_swizzle(dev);
4778	init_waitqueue_head(&dev_priv->pending_flip_queue);
4779
4780	dev_priv->mm.interruptible = true;
4781
4782	dev_priv->mm.inactive_shrinker.scan_objects = i915_gem_inactive_scan;
4783	dev_priv->mm.inactive_shrinker.count_objects = i915_gem_inactive_count;
4784	dev_priv->mm.inactive_shrinker.seeks = DEFAULT_SEEKS;
4785	register_shrinker(&dev_priv->mm.inactive_shrinker);
4786}
4787
4788void i915_gem_release(struct drm_device *dev, struct drm_file *file)
4789{
4790	struct drm_i915_file_private *file_priv = file->driver_priv;
4791
4792	cancel_delayed_work_sync(&file_priv->mm.idle_work);
4793
4794	/* Clean up our request list when the client is going away, so that
4795	 * later retire_requests won't dereference our soon-to-be-gone
4796	 * file_priv.
4797	 */
4798	spin_lock(&file_priv->mm.lock);
4799	while (!list_empty(&file_priv->mm.request_list)) {
4800		struct drm_i915_gem_request *request;
4801
4802		request = list_first_entry(&file_priv->mm.request_list,
4803					   struct drm_i915_gem_request,
4804					   client_list);
4805		list_del(&request->client_list);
4806		request->file_priv = NULL;
4807	}
4808	spin_unlock(&file_priv->mm.lock);
4809}
4810
4811static void
4812i915_gem_file_idle_work_handler(struct work_struct *work)
4813{
4814	struct drm_i915_file_private *file_priv =
4815		container_of(work, typeof(*file_priv), mm.idle_work.work);
4816
4817	atomic_set(&file_priv->rps_wait_boost, false);
4818}
4819
4820int i915_gem_open(struct drm_device *dev, struct drm_file *file)
4821{
4822	struct drm_i915_file_private *file_priv;
4823	int ret;
 
4824
4825	DRM_DEBUG_DRIVER("\n");
4826
4827	file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL);
4828	if (!file_priv)
4829		return -ENOMEM;
 
 
 
 
4830
4831	file->driver_priv = file_priv;
4832	file_priv->dev_priv = dev->dev_private;
4833	file_priv->file = file;
 
4834
4835	spin_lock_init(&file_priv->mm.lock);
4836	INIT_LIST_HEAD(&file_priv->mm.request_list);
4837	INIT_DELAYED_WORK(&file_priv->mm.idle_work,
4838			  i915_gem_file_idle_work_handler);
4839
4840	ret = i915_gem_context_open(dev, file);
4841	if (ret)
4842		kfree(file_priv);
 
 
4843
 
 
 
 
 
4844	return ret;
4845}
4846
4847static bool mutex_is_locked_by(struct mutex *mutex, struct task_struct *task)
4848{
4849	if (!mutex_is_locked(mutex))
4850		return false;
4851
4852#if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_MUTEXES)
4853	return mutex->owner == task;
4854#else
4855	/* Since UP may be pre-empted, we cannot assume that we own the lock */
4856	return false;
4857#endif
4858}
4859
4860static unsigned long
4861i915_gem_inactive_count(struct shrinker *shrinker, struct shrink_control *sc)
4862{
4863	struct drm_i915_private *dev_priv =
4864		container_of(shrinker,
4865			     struct drm_i915_private,
4866			     mm.inactive_shrinker);
4867	struct drm_device *dev = dev_priv->dev;
4868	struct drm_i915_gem_object *obj;
4869	bool unlock = true;
4870	unsigned long count;
4871
4872	if (!mutex_trylock(&dev->struct_mutex)) {
4873		if (!mutex_is_locked_by(&dev->struct_mutex, current))
4874			return 0;
4875
4876		if (dev_priv->mm.shrinker_no_lock_stealing)
4877			return 0;
4878
4879		unlock = false;
4880	}
4881
4882	count = 0;
4883	list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list)
4884		if (obj->pages_pin_count == 0)
4885			count += obj->base.size >> PAGE_SHIFT;
4886
4887	list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
4888		if (obj->active)
4889			continue;
4890
4891		if (!i915_gem_obj_is_pinned(obj) && obj->pages_pin_count == 0)
4892			count += obj->base.size >> PAGE_SHIFT;
4893	}
4894
4895	if (unlock)
4896		mutex_unlock(&dev->struct_mutex);
4897
4898	return count;
4899}
4900
4901/* All the new VM stuff */
4902unsigned long i915_gem_obj_offset(struct drm_i915_gem_object *o,
4903				  struct i915_address_space *vm)
4904{
4905	struct drm_i915_private *dev_priv = o->base.dev->dev_private;
4906	struct i915_vma *vma;
4907
4908	if (!dev_priv->mm.aliasing_ppgtt ||
4909	    vm == &dev_priv->mm.aliasing_ppgtt->base)
4910		vm = &dev_priv->gtt.base;
4911
4912	BUG_ON(list_empty(&o->vma_list));
4913	list_for_each_entry(vma, &o->vma_list, vma_link) {
4914		if (vma->vm == vm)
4915			return vma->node.start;
4916
4917	}
4918	return -1;
4919}
4920
4921bool i915_gem_obj_bound(struct drm_i915_gem_object *o,
4922			struct i915_address_space *vm)
4923{
4924	struct i915_vma *vma;
4925
4926	list_for_each_entry(vma, &o->vma_list, vma_link)
4927		if (vma->vm == vm && drm_mm_node_allocated(&vma->node))
4928			return true;
4929
4930	return false;
4931}
4932
4933bool i915_gem_obj_bound_any(struct drm_i915_gem_object *o)
4934{
4935	struct i915_vma *vma;
4936
4937	list_for_each_entry(vma, &o->vma_list, vma_link)
4938		if (drm_mm_node_allocated(&vma->node))
4939			return true;
4940
4941	return false;
4942}
4943
4944unsigned long i915_gem_obj_size(struct drm_i915_gem_object *o,
4945				struct i915_address_space *vm)
4946{
4947	struct drm_i915_private *dev_priv = o->base.dev->dev_private;
4948	struct i915_vma *vma;
4949
4950	if (!dev_priv->mm.aliasing_ppgtt ||
4951	    vm == &dev_priv->mm.aliasing_ppgtt->base)
4952		vm = &dev_priv->gtt.base;
4953
4954	BUG_ON(list_empty(&o->vma_list));
4955
4956	list_for_each_entry(vma, &o->vma_list, vma_link)
4957		if (vma->vm == vm)
4958			return vma->node.size;
4959
4960	return 0;
4961}
4962
4963static unsigned long
4964i915_gem_inactive_scan(struct shrinker *shrinker, struct shrink_control *sc)
4965{
4966	struct drm_i915_private *dev_priv =
4967		container_of(shrinker,
4968			     struct drm_i915_private,
4969			     mm.inactive_shrinker);
4970	struct drm_device *dev = dev_priv->dev;
4971	unsigned long freed;
4972	bool unlock = true;
4973
4974	if (!mutex_trylock(&dev->struct_mutex)) {
4975		if (!mutex_is_locked_by(&dev->struct_mutex, current))
4976			return SHRINK_STOP;
4977
4978		if (dev_priv->mm.shrinker_no_lock_stealing)
4979			return SHRINK_STOP;
4980
4981		unlock = false;
4982	}
4983
4984	freed = i915_gem_purge(dev_priv, sc->nr_to_scan);
4985	if (freed < sc->nr_to_scan)
4986		freed += __i915_gem_shrink(dev_priv,
4987					   sc->nr_to_scan - freed,
4988					   false);
4989	if (freed < sc->nr_to_scan)
4990		freed += i915_gem_shrink_all(dev_priv);
4991
4992	if (unlock)
4993		mutex_unlock(&dev->struct_mutex);
4994
4995	return freed;
4996}
4997
4998struct i915_vma *i915_gem_obj_to_ggtt(struct drm_i915_gem_object *obj)
4999{
5000	struct i915_vma *vma;
5001
5002	if (WARN_ON(list_empty(&obj->vma_list)))
5003		return NULL;
5004
5005	vma = list_first_entry(&obj->vma_list, typeof(*vma), vma_link);
5006	if (vma->vm != obj_to_ggtt(obj))
5007		return NULL;
5008
5009	return vma;
5010}