Linux Audio

Check our new training course

Loading...
v5.14.15
   1/*
   2 * Copyright © 2008-2015 Intel Corporation
   3 *
   4 * Permission is hereby granted, free of charge, to any person obtaining a
   5 * copy of this software and associated documentation files (the "Software"),
   6 * to deal in the Software without restriction, including without limitation
   7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8 * and/or sell copies of the Software, and to permit persons to whom the
   9 * Software is furnished to do so, subject to the following conditions:
  10 *
  11 * The above copyright notice and this permission notice (including the next
  12 * paragraph) shall be included in all copies or substantial portions of the
  13 * Software.
  14 *
  15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21 * IN THE SOFTWARE.
  22 *
  23 * Authors:
  24 *    Eric Anholt <eric@anholt.net>
  25 *
  26 */
  27
  28#include <drm/drm_vma_manager.h>
  29#include <linux/dma-fence-array.h>
  30#include <linux/kthread.h>
  31#include <linux/dma-resv.h>
  32#include <linux/shmem_fs.h>
  33#include <linux/slab.h>
  34#include <linux/stop_machine.h>
  35#include <linux/swap.h>
  36#include <linux/pci.h>
  37#include <linux/dma-buf.h>
  38#include <linux/mman.h>
  39
  40#include "display/intel_display.h"
  41#include "display/intel_frontbuffer.h"
  42
  43#include "gem/i915_gem_clflush.h"
  44#include "gem/i915_gem_context.h"
  45#include "gem/i915_gem_ioctls.h"
  46#include "gem/i915_gem_mman.h"
 
 
  47#include "gem/i915_gem_region.h"
  48#include "gt/intel_engine_user.h"
  49#include "gt/intel_gt.h"
  50#include "gt/intel_gt_pm.h"
  51#include "gt/intel_workarounds.h"
  52
  53#include "i915_drv.h"
 
  54#include "i915_trace.h"
  55#include "i915_vgpu.h"
  56
  57#include "intel_pm.h"
  58
  59static int
  60insert_mappable_node(struct i915_ggtt *ggtt, struct drm_mm_node *node, u32 size)
  61{
  62	int err;
  63
  64	err = mutex_lock_interruptible(&ggtt->vm.mutex);
  65	if (err)
  66		return err;
  67
  68	memset(node, 0, sizeof(*node));
  69	err = drm_mm_insert_node_in_range(&ggtt->vm.mm, node,
  70					  size, 0, I915_COLOR_UNEVICTABLE,
  71					  0, ggtt->mappable_end,
  72					  DRM_MM_INSERT_LOW);
  73
  74	mutex_unlock(&ggtt->vm.mutex);
  75
  76	return err;
  77}
  78
  79static void
  80remove_mappable_node(struct i915_ggtt *ggtt, struct drm_mm_node *node)
  81{
  82	mutex_lock(&ggtt->vm.mutex);
  83	drm_mm_remove_node(node);
  84	mutex_unlock(&ggtt->vm.mutex);
  85}
  86
  87int
  88i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
  89			    struct drm_file *file)
  90{
  91	struct i915_ggtt *ggtt = &to_i915(dev)->ggtt;
 
  92	struct drm_i915_gem_get_aperture *args = data;
  93	struct i915_vma *vma;
  94	u64 pinned;
  95
  96	if (mutex_lock_interruptible(&ggtt->vm.mutex))
  97		return -EINTR;
  98
  99	pinned = ggtt->vm.reserved;
 100	list_for_each_entry(vma, &ggtt->vm.bound_list, vm_link)
 101		if (i915_vma_is_pinned(vma))
 102			pinned += vma->node.size;
 103
 104	mutex_unlock(&ggtt->vm.mutex);
 105
 106	args->aper_size = ggtt->vm.total;
 107	args->aper_available_size = args->aper_size - pinned;
 108
 109	return 0;
 110}
 111
 112int i915_gem_object_unbind(struct drm_i915_gem_object *obj,
 113			   unsigned long flags)
 114{
 115	struct intel_runtime_pm *rpm = &to_i915(obj->base.dev)->runtime_pm;
 
 116	LIST_HEAD(still_in_list);
 117	intel_wakeref_t wakeref;
 118	struct i915_vma *vma;
 119	int ret;
 120
 
 
 121	if (list_empty(&obj->vma.list))
 122		return 0;
 123
 124	/*
 125	 * As some machines use ACPI to handle runtime-resume callbacks, and
 126	 * ACPI is quite kmalloc happy, we cannot resume beneath the vm->mutex
 127	 * as they are required by the shrinker. Ergo, we wake the device up
 128	 * first just in case.
 129	 */
 130	wakeref = intel_runtime_pm_get(rpm);
 131
 132try_again:
 133	ret = 0;
 134	spin_lock(&obj->vma.lock);
 135	while (!ret && (vma = list_first_entry_or_null(&obj->vma.list,
 136						       struct i915_vma,
 137						       obj_link))) {
 138		struct i915_address_space *vm = vma->vm;
 139
 140		list_move_tail(&vma->obj_link, &still_in_list);
 141		if (!i915_vma_is_bound(vma, I915_VMA_BIND_MASK))
 142			continue;
 143
 144		if (flags & I915_GEM_OBJECT_UNBIND_TEST) {
 145			ret = -EBUSY;
 146			break;
 147		}
 148
 
 
 
 
 
 
 
 149		ret = -EAGAIN;
 150		if (!i915_vm_tryopen(vm))
 151			break;
 152
 153		/* Prevent vma being freed by i915_vma_parked as we unbind */
 154		vma = __i915_vma_get(vma);
 155		spin_unlock(&obj->vma.lock);
 156
 157		if (vma) {
 158			ret = -EBUSY;
 159			if (flags & I915_GEM_OBJECT_UNBIND_ACTIVE ||
 160			    !i915_vma_is_active(vma)) {
 161				if (flags & I915_GEM_OBJECT_UNBIND_VM_TRYLOCK) {
 162					if (mutex_trylock(&vma->vm->mutex)) {
 163						ret = __i915_vma_unbind(vma);
 164						mutex_unlock(&vma->vm->mutex);
 165					} else {
 166						ret = -EBUSY;
 167					}
 168				} else {
 169					ret = i915_vma_unbind(vma);
 
 
 
 
 
 170				}
 
 
 171			}
 172
 173			__i915_vma_put(vma);
 174		}
 175
 176		i915_vm_close(vm);
 177		spin_lock(&obj->vma.lock);
 178	}
 179	list_splice_init(&still_in_list, &obj->vma.list);
 180	spin_unlock(&obj->vma.lock);
 181
 182	if (ret == -EAGAIN && flags & I915_GEM_OBJECT_UNBIND_BARRIER) {
 183		rcu_barrier(); /* flush the i915_vm_release() */
 184		goto try_again;
 185	}
 186
 187	intel_runtime_pm_put(rpm, wakeref);
 188
 189	return ret;
 190}
 191
 192static int
 193shmem_pread(struct page *page, int offset, int len, char __user *user_data,
 194	    bool needs_clflush)
 195{
 196	char *vaddr;
 197	int ret;
 198
 199	vaddr = kmap(page);
 200
 201	if (needs_clflush)
 202		drm_clflush_virt_range(vaddr + offset, len);
 203
 204	ret = __copy_to_user(user_data, vaddr + offset, len);
 205
 206	kunmap(page);
 207
 208	return ret ? -EFAULT : 0;
 209}
 210
 211static int
 212i915_gem_shmem_pread(struct drm_i915_gem_object *obj,
 213		     struct drm_i915_gem_pread *args)
 214{
 215	unsigned int needs_clflush;
 216	unsigned int idx, offset;
 217	char __user *user_data;
 
 
 218	u64 remain;
 219	int ret;
 220
 221	ret = i915_gem_object_lock_interruptible(obj, NULL);
 222	if (ret)
 223		return ret;
 224
 225	ret = i915_gem_object_pin_pages(obj);
 226	if (ret)
 227		goto err_unlock;
 228
 229	ret = i915_gem_object_prepare_read(obj, &needs_clflush);
 230	if (ret)
 231		goto err_unpin;
 232
 233	i915_gem_object_finish_access(obj);
 234	i915_gem_object_unlock(obj);
 235
 236	remain = args->size;
 237	user_data = u64_to_user_ptr(args->data_ptr);
 238	offset = offset_in_page(args->offset);
 239	for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
 240		struct page *page = i915_gem_object_get_page(obj, idx);
 241		unsigned int length = min_t(u64, remain, PAGE_SIZE - offset);
 242
 243		ret = shmem_pread(page, offset, length, user_data,
 244				  needs_clflush);
 245		if (ret)
 246			break;
 247
 248		remain -= length;
 249		user_data += length;
 250		offset = 0;
 251	}
 252
 253	i915_gem_object_unpin_pages(obj);
 254	return ret;
 255
 256err_unpin:
 257	i915_gem_object_unpin_pages(obj);
 258err_unlock:
 259	i915_gem_object_unlock(obj);
 260	return ret;
 261}
 262
 263static inline bool
 264gtt_user_read(struct io_mapping *mapping,
 265	      loff_t base, int offset,
 266	      char __user *user_data, int length)
 267{
 268	void __iomem *vaddr;
 269	unsigned long unwritten;
 270
 271	/* We can use the cpu mem copy function because this is X86. */
 272	vaddr = io_mapping_map_atomic_wc(mapping, base);
 273	unwritten = __copy_to_user_inatomic(user_data,
 274					    (void __force *)vaddr + offset,
 275					    length);
 276	io_mapping_unmap_atomic(vaddr);
 277	if (unwritten) {
 278		vaddr = io_mapping_map_wc(mapping, base, PAGE_SIZE);
 279		unwritten = copy_to_user(user_data,
 280					 (void __force *)vaddr + offset,
 281					 length);
 282		io_mapping_unmap(vaddr);
 283	}
 284	return unwritten;
 285}
 286
 287static struct i915_vma *i915_gem_gtt_prepare(struct drm_i915_gem_object *obj,
 288					     struct drm_mm_node *node,
 289					     bool write)
 290{
 291	struct drm_i915_private *i915 = to_i915(obj->base.dev);
 292	struct i915_ggtt *ggtt = &i915->ggtt;
 293	struct i915_vma *vma;
 294	struct i915_gem_ww_ctx ww;
 295	int ret;
 296
 297	i915_gem_ww_ctx_init(&ww, true);
 298retry:
 299	vma = ERR_PTR(-ENODEV);
 300	ret = i915_gem_object_lock(obj, &ww);
 301	if (ret)
 302		goto err_ww;
 303
 304	ret = i915_gem_object_set_to_gtt_domain(obj, write);
 305	if (ret)
 306		goto err_ww;
 307
 308	if (!i915_gem_object_is_tiled(obj))
 309		vma = i915_gem_object_ggtt_pin_ww(obj, &ww, NULL, 0, 0,
 310						  PIN_MAPPABLE |
 311						  PIN_NONBLOCK /* NOWARN */ |
 312						  PIN_NOEVICT);
 313	if (vma == ERR_PTR(-EDEADLK)) {
 314		ret = -EDEADLK;
 315		goto err_ww;
 316	} else if (!IS_ERR(vma)) {
 317		node->start = i915_ggtt_offset(vma);
 318		node->flags = 0;
 319	} else {
 320		ret = insert_mappable_node(ggtt, node, PAGE_SIZE);
 321		if (ret)
 322			goto err_ww;
 323		GEM_BUG_ON(!drm_mm_node_allocated(node));
 324		vma = NULL;
 325	}
 326
 327	ret = i915_gem_object_pin_pages(obj);
 328	if (ret) {
 329		if (drm_mm_node_allocated(node)) {
 330			ggtt->vm.clear_range(&ggtt->vm, node->start, node->size);
 331			remove_mappable_node(ggtt, node);
 332		} else {
 333			i915_vma_unpin(vma);
 334		}
 335	}
 336
 337err_ww:
 338	if (ret == -EDEADLK) {
 339		ret = i915_gem_ww_ctx_backoff(&ww);
 340		if (!ret)
 341			goto retry;
 342	}
 343	i915_gem_ww_ctx_fini(&ww);
 344
 345	return ret ? ERR_PTR(ret) : vma;
 346}
 347
 348static void i915_gem_gtt_cleanup(struct drm_i915_gem_object *obj,
 349				 struct drm_mm_node *node,
 350				 struct i915_vma *vma)
 351{
 352	struct drm_i915_private *i915 = to_i915(obj->base.dev);
 353	struct i915_ggtt *ggtt = &i915->ggtt;
 354
 355	i915_gem_object_unpin_pages(obj);
 356	if (drm_mm_node_allocated(node)) {
 357		ggtt->vm.clear_range(&ggtt->vm, node->start, node->size);
 358		remove_mappable_node(ggtt, node);
 359	} else {
 360		i915_vma_unpin(vma);
 361	}
 362}
 363
 364static int
 365i915_gem_gtt_pread(struct drm_i915_gem_object *obj,
 366		   const struct drm_i915_gem_pread *args)
 367{
 368	struct drm_i915_private *i915 = to_i915(obj->base.dev);
 369	struct i915_ggtt *ggtt = &i915->ggtt;
 
 370	intel_wakeref_t wakeref;
 371	struct drm_mm_node node;
 372	void __user *user_data;
 373	struct i915_vma *vma;
 374	u64 remain, offset;
 375	int ret = 0;
 376
 
 
 
 
 377	wakeref = intel_runtime_pm_get(&i915->runtime_pm);
 378
 379	vma = i915_gem_gtt_prepare(obj, &node, false);
 380	if (IS_ERR(vma)) {
 381		ret = PTR_ERR(vma);
 382		goto out_rpm;
 383	}
 384
 385	user_data = u64_to_user_ptr(args->data_ptr);
 386	remain = args->size;
 387	offset = args->offset;
 388
 389	while (remain > 0) {
 390		/* Operation in this page
 391		 *
 392		 * page_base = page offset within aperture
 393		 * page_offset = offset within page
 394		 * page_length = bytes to copy for this page
 395		 */
 396		u32 page_base = node.start;
 397		unsigned page_offset = offset_in_page(offset);
 398		unsigned page_length = PAGE_SIZE - page_offset;
 399		page_length = remain < page_length ? remain : page_length;
 400		if (drm_mm_node_allocated(&node)) {
 401			ggtt->vm.insert_page(&ggtt->vm,
 402					     i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT),
 403					     node.start, I915_CACHE_NONE, 0);
 
 
 
 404		} else {
 405			page_base += offset & PAGE_MASK;
 406		}
 407
 408		if (gtt_user_read(&ggtt->iomap, page_base, page_offset,
 409				  user_data, page_length)) {
 410			ret = -EFAULT;
 411			break;
 412		}
 413
 414		remain -= page_length;
 415		user_data += page_length;
 416		offset += page_length;
 417	}
 418
 419	i915_gem_gtt_cleanup(obj, &node, vma);
 420out_rpm:
 421	intel_runtime_pm_put(&i915->runtime_pm, wakeref);
 422	return ret;
 423}
 424
 425/**
 426 * Reads data from the object referenced by handle.
 427 * @dev: drm device pointer
 428 * @data: ioctl data blob
 429 * @file: drm file pointer
 430 *
 431 * On error, the contents of *data are undefined.
 432 */
 433int
 434i915_gem_pread_ioctl(struct drm_device *dev, void *data,
 435		     struct drm_file *file)
 436{
 437	struct drm_i915_private *i915 = to_i915(dev);
 438	struct drm_i915_gem_pread *args = data;
 439	struct drm_i915_gem_object *obj;
 440	int ret;
 441
 442	/* PREAD is disallowed for all platforms after TGL-LP.  This also
 443	 * covers all platforms with local memory.
 444	 */
 445	if (GRAPHICS_VER(i915) >= 12 && !IS_TIGERLAKE(i915))
 446		return -EOPNOTSUPP;
 447
 448	if (args->size == 0)
 449		return 0;
 450
 451	if (!access_ok(u64_to_user_ptr(args->data_ptr),
 452		       args->size))
 453		return -EFAULT;
 454
 455	obj = i915_gem_object_lookup(file, args->handle);
 456	if (!obj)
 457		return -ENOENT;
 458
 459	/* Bounds check source.  */
 460	if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) {
 461		ret = -EINVAL;
 462		goto out;
 463	}
 464
 465	trace_i915_gem_object_pread(obj, args->offset, args->size);
 466	ret = -ENODEV;
 467	if (obj->ops->pread)
 468		ret = obj->ops->pread(obj, args);
 469	if (ret != -ENODEV)
 470		goto out;
 471
 472	ret = -ENODEV;
 473	if (obj->ops->pread)
 474		ret = obj->ops->pread(obj, args);
 475	if (ret != -ENODEV)
 476		goto out;
 477
 478	ret = i915_gem_object_wait(obj,
 479				   I915_WAIT_INTERRUPTIBLE,
 480				   MAX_SCHEDULE_TIMEOUT);
 481	if (ret)
 482		goto out;
 483
 484	ret = i915_gem_shmem_pread(obj, args);
 485	if (ret == -EFAULT || ret == -ENODEV)
 486		ret = i915_gem_gtt_pread(obj, args);
 487
 488out:
 489	i915_gem_object_put(obj);
 490	return ret;
 491}
 492
 493/* This is the fast write path which cannot handle
 494 * page faults in the source data
 495 */
 496
 497static inline bool
 498ggtt_write(struct io_mapping *mapping,
 499	   loff_t base, int offset,
 500	   char __user *user_data, int length)
 501{
 502	void __iomem *vaddr;
 503	unsigned long unwritten;
 504
 505	/* We can use the cpu mem copy function because this is X86. */
 506	vaddr = io_mapping_map_atomic_wc(mapping, base);
 507	unwritten = __copy_from_user_inatomic_nocache((void __force *)vaddr + offset,
 508						      user_data, length);
 509	io_mapping_unmap_atomic(vaddr);
 510	if (unwritten) {
 511		vaddr = io_mapping_map_wc(mapping, base, PAGE_SIZE);
 512		unwritten = copy_from_user((void __force *)vaddr + offset,
 513					   user_data, length);
 514		io_mapping_unmap(vaddr);
 515	}
 516
 517	return unwritten;
 518}
 519
 520/**
 521 * This is the fast pwrite path, where we copy the data directly from the
 522 * user into the GTT, uncached.
 523 * @obj: i915 GEM object
 524 * @args: pwrite arguments structure
 525 */
 526static int
 527i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
 528			 const struct drm_i915_gem_pwrite *args)
 529{
 530	struct drm_i915_private *i915 = to_i915(obj->base.dev);
 531	struct i915_ggtt *ggtt = &i915->ggtt;
 532	struct intel_runtime_pm *rpm = &i915->runtime_pm;
 
 533	intel_wakeref_t wakeref;
 534	struct drm_mm_node node;
 535	struct i915_vma *vma;
 536	u64 remain, offset;
 537	void __user *user_data;
 538	int ret = 0;
 539
 
 
 
 
 540	if (i915_gem_object_has_struct_page(obj)) {
 541		/*
 542		 * Avoid waking the device up if we can fallback, as
 543		 * waking/resuming is very slow (worst-case 10-100 ms
 544		 * depending on PCI sleeps and our own resume time).
 545		 * This easily dwarfs any performance advantage from
 546		 * using the cache bypass of indirect GGTT access.
 547		 */
 548		wakeref = intel_runtime_pm_get_if_in_use(rpm);
 549		if (!wakeref)
 550			return -EFAULT;
 551	} else {
 552		/* No backing pages, no fallback, we must force GGTT access */
 553		wakeref = intel_runtime_pm_get(rpm);
 554	}
 555
 556	vma = i915_gem_gtt_prepare(obj, &node, true);
 557	if (IS_ERR(vma)) {
 558		ret = PTR_ERR(vma);
 559		goto out_rpm;
 560	}
 561
 562	i915_gem_object_invalidate_frontbuffer(obj, ORIGIN_CPU);
 563
 564	user_data = u64_to_user_ptr(args->data_ptr);
 565	offset = args->offset;
 566	remain = args->size;
 567	while (remain) {
 568		/* Operation in this page
 569		 *
 570		 * page_base = page offset within aperture
 571		 * page_offset = offset within page
 572		 * page_length = bytes to copy for this page
 573		 */
 574		u32 page_base = node.start;
 575		unsigned int page_offset = offset_in_page(offset);
 576		unsigned int page_length = PAGE_SIZE - page_offset;
 577		page_length = remain < page_length ? remain : page_length;
 578		if (drm_mm_node_allocated(&node)) {
 579			/* flush the write before we modify the GGTT */
 580			intel_gt_flush_ggtt_writes(ggtt->vm.gt);
 581			ggtt->vm.insert_page(&ggtt->vm,
 582					     i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT),
 583					     node.start, I915_CACHE_NONE, 0);
 
 
 
 584			wmb(); /* flush modifications to the GGTT (insert_page) */
 585		} else {
 586			page_base += offset & PAGE_MASK;
 587		}
 588		/* If we get a fault while copying data, then (presumably) our
 589		 * source page isn't available.  Return the error and we'll
 590		 * retry in the slow path.
 591		 * If the object is non-shmem backed, we retry again with the
 592		 * path that handles page fault.
 593		 */
 594		if (ggtt_write(&ggtt->iomap, page_base, page_offset,
 595			       user_data, page_length)) {
 596			ret = -EFAULT;
 597			break;
 598		}
 599
 600		remain -= page_length;
 601		user_data += page_length;
 602		offset += page_length;
 603	}
 604
 605	intel_gt_flush_ggtt_writes(ggtt->vm.gt);
 606	i915_gem_object_flush_frontbuffer(obj, ORIGIN_CPU);
 607
 608	i915_gem_gtt_cleanup(obj, &node, vma);
 609out_rpm:
 610	intel_runtime_pm_put(rpm, wakeref);
 611	return ret;
 612}
 613
 614/* Per-page copy function for the shmem pwrite fastpath.
 615 * Flushes invalid cachelines before writing to the target if
 616 * needs_clflush_before is set and flushes out any written cachelines after
 617 * writing if needs_clflush is set.
 618 */
 619static int
 620shmem_pwrite(struct page *page, int offset, int len, char __user *user_data,
 621	     bool needs_clflush_before,
 622	     bool needs_clflush_after)
 623{
 624	char *vaddr;
 625	int ret;
 626
 627	vaddr = kmap(page);
 628
 629	if (needs_clflush_before)
 630		drm_clflush_virt_range(vaddr + offset, len);
 631
 632	ret = __copy_from_user(vaddr + offset, user_data, len);
 633	if (!ret && needs_clflush_after)
 634		drm_clflush_virt_range(vaddr + offset, len);
 635
 636	kunmap(page);
 637
 638	return ret ? -EFAULT : 0;
 639}
 640
 641static int
 642i915_gem_shmem_pwrite(struct drm_i915_gem_object *obj,
 643		      const struct drm_i915_gem_pwrite *args)
 644{
 645	unsigned int partial_cacheline_write;
 646	unsigned int needs_clflush;
 647	unsigned int offset, idx;
 648	void __user *user_data;
 
 
 649	u64 remain;
 650	int ret;
 651
 652	ret = i915_gem_object_lock_interruptible(obj, NULL);
 653	if (ret)
 654		return ret;
 655
 656	ret = i915_gem_object_pin_pages(obj);
 657	if (ret)
 658		goto err_unlock;
 659
 660	ret = i915_gem_object_prepare_write(obj, &needs_clflush);
 661	if (ret)
 662		goto err_unpin;
 663
 664	i915_gem_object_finish_access(obj);
 665	i915_gem_object_unlock(obj);
 666
 667	/* If we don't overwrite a cacheline completely we need to be
 668	 * careful to have up-to-date data by first clflushing. Don't
 669	 * overcomplicate things and flush the entire patch.
 670	 */
 671	partial_cacheline_write = 0;
 672	if (needs_clflush & CLFLUSH_BEFORE)
 673		partial_cacheline_write = boot_cpu_data.x86_clflush_size - 1;
 674
 675	user_data = u64_to_user_ptr(args->data_ptr);
 676	remain = args->size;
 677	offset = offset_in_page(args->offset);
 678	for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
 679		struct page *page = i915_gem_object_get_page(obj, idx);
 680		unsigned int length = min_t(u64, remain, PAGE_SIZE - offset);
 681
 682		ret = shmem_pwrite(page, offset, length, user_data,
 683				   (offset | length) & partial_cacheline_write,
 684				   needs_clflush & CLFLUSH_AFTER);
 685		if (ret)
 686			break;
 687
 688		remain -= length;
 689		user_data += length;
 690		offset = 0;
 691	}
 692
 693	i915_gem_object_flush_frontbuffer(obj, ORIGIN_CPU);
 694
 695	i915_gem_object_unpin_pages(obj);
 696	return ret;
 697
 698err_unpin:
 699	i915_gem_object_unpin_pages(obj);
 700err_unlock:
 701	i915_gem_object_unlock(obj);
 702	return ret;
 703}
 704
 705/**
 706 * Writes data to the object referenced by handle.
 707 * @dev: drm device
 708 * @data: ioctl data blob
 709 * @file: drm file
 710 *
 711 * On error, the contents of the buffer that were to be modified are undefined.
 712 */
 713int
 714i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
 715		      struct drm_file *file)
 716{
 717	struct drm_i915_private *i915 = to_i915(dev);
 718	struct drm_i915_gem_pwrite *args = data;
 719	struct drm_i915_gem_object *obj;
 720	int ret;
 721
 722	/* PWRITE is disallowed for all platforms after TGL-LP.  This also
 723	 * covers all platforms with local memory.
 724	 */
 725	if (GRAPHICS_VER(i915) >= 12 && !IS_TIGERLAKE(i915))
 726		return -EOPNOTSUPP;
 727
 728	if (args->size == 0)
 729		return 0;
 730
 731	if (!access_ok(u64_to_user_ptr(args->data_ptr), args->size))
 732		return -EFAULT;
 733
 734	obj = i915_gem_object_lookup(file, args->handle);
 735	if (!obj)
 736		return -ENOENT;
 737
 738	/* Bounds check destination. */
 739	if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) {
 740		ret = -EINVAL;
 741		goto err;
 742	}
 743
 744	/* Writes not allowed into this read-only object */
 745	if (i915_gem_object_is_readonly(obj)) {
 746		ret = -EINVAL;
 747		goto err;
 748	}
 749
 750	trace_i915_gem_object_pwrite(obj, args->offset, args->size);
 751
 752	ret = -ENODEV;
 753	if (obj->ops->pwrite)
 754		ret = obj->ops->pwrite(obj, args);
 755	if (ret != -ENODEV)
 756		goto err;
 757
 758	ret = i915_gem_object_wait(obj,
 759				   I915_WAIT_INTERRUPTIBLE |
 760				   I915_WAIT_ALL,
 761				   MAX_SCHEDULE_TIMEOUT);
 762	if (ret)
 763		goto err;
 764
 765	ret = -EFAULT;
 766	/* We can only do the GTT pwrite on untiled buffers, as otherwise
 767	 * it would end up going through the fenced access, and we'll get
 768	 * different detiling behavior between reading and writing.
 769	 * pread/pwrite currently are reading and writing from the CPU
 770	 * perspective, requiring manual detiling by the client.
 771	 */
 772	if (!i915_gem_object_has_struct_page(obj) ||
 773	    cpu_write_needs_clflush(obj))
 774		/* Note that the gtt paths might fail with non-page-backed user
 775		 * pointers (e.g. gtt mappings when moving data between
 776		 * textures). Fallback to the shmem path in that case.
 777		 */
 778		ret = i915_gem_gtt_pwrite_fast(obj, args);
 779
 780	if (ret == -EFAULT || ret == -ENOSPC) {
 781		if (i915_gem_object_has_struct_page(obj))
 782			ret = i915_gem_shmem_pwrite(obj, args);
 783	}
 784
 785err:
 786	i915_gem_object_put(obj);
 787	return ret;
 788}
 789
 790/**
 791 * Called when user space has done writes to this buffer
 792 * @dev: drm device
 793 * @data: ioctl data blob
 794 * @file: drm file
 795 */
 796int
 797i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
 798			 struct drm_file *file)
 799{
 800	struct drm_i915_gem_sw_finish *args = data;
 801	struct drm_i915_gem_object *obj;
 802
 803	obj = i915_gem_object_lookup(file, args->handle);
 804	if (!obj)
 805		return -ENOENT;
 806
 807	/*
 808	 * Proxy objects are barred from CPU access, so there is no
 809	 * need to ban sw_finish as it is a nop.
 810	 */
 811
 812	/* Pinned buffers may be scanout, so flush the cache */
 813	i915_gem_object_flush_if_display(obj);
 814	i915_gem_object_put(obj);
 815
 816	return 0;
 817}
 818
 819void i915_gem_runtime_suspend(struct drm_i915_private *i915)
 820{
 821	struct drm_i915_gem_object *obj, *on;
 822	int i;
 823
 824	/*
 825	 * Only called during RPM suspend. All users of the userfault_list
 826	 * must be holding an RPM wakeref to ensure that this can not
 827	 * run concurrently with themselves (and use the struct_mutex for
 828	 * protection between themselves).
 829	 */
 830
 831	list_for_each_entry_safe(obj, on,
 832				 &i915->ggtt.userfault_list, userfault_link)
 833		__i915_gem_object_release_mmap_gtt(obj);
 834
 
 
 
 
 835	/*
 836	 * The fence will be lost when the device powers down. If any were
 837	 * in use by hardware (i.e. they are pinned), we should not be powering
 838	 * down! All other fences will be reacquired by the user upon waking.
 839	 */
 840	for (i = 0; i < i915->ggtt.num_fences; i++) {
 841		struct i915_fence_reg *reg = &i915->ggtt.fence_regs[i];
 842
 843		/*
 844		 * Ideally we want to assert that the fence register is not
 845		 * live at this point (i.e. that no piece of code will be
 846		 * trying to write through fence + GTT, as that both violates
 847		 * our tracking of activity and associated locking/barriers,
 848		 * but also is illegal given that the hw is powered down).
 849		 *
 850		 * Previously we used reg->pin_count as a "liveness" indicator.
 851		 * That is not sufficient, and we need a more fine-grained
 852		 * tool if we want to have a sanity check here.
 853		 */
 854
 855		if (!reg->vma)
 856			continue;
 857
 858		GEM_BUG_ON(i915_vma_has_userfault(reg->vma));
 859		reg->dirty = true;
 860	}
 861}
 862
 863static void discard_ggtt_vma(struct i915_vma *vma)
 864{
 865	struct drm_i915_gem_object *obj = vma->obj;
 866
 867	spin_lock(&obj->vma.lock);
 868	if (!RB_EMPTY_NODE(&vma->obj_node)) {
 869		rb_erase(&vma->obj_node, &obj->vma.tree);
 870		RB_CLEAR_NODE(&vma->obj_node);
 871	}
 872	spin_unlock(&obj->vma.lock);
 873}
 874
 875struct i915_vma *
 876i915_gem_object_ggtt_pin_ww(struct drm_i915_gem_object *obj,
 877			    struct i915_gem_ww_ctx *ww,
 878			    const struct i915_ggtt_view *view,
 879			    u64 size, u64 alignment, u64 flags)
 880{
 881	struct drm_i915_private *i915 = to_i915(obj->base.dev);
 882	struct i915_ggtt *ggtt = &i915->ggtt;
 883	struct i915_vma *vma;
 884	int ret;
 885
 
 
 886	if (flags & PIN_MAPPABLE &&
 887	    (!view || view->type == I915_GGTT_VIEW_NORMAL)) {
 888		/*
 889		 * If the required space is larger than the available
 890		 * aperture, we will not able to find a slot for the
 891		 * object and unbinding the object now will be in
 892		 * vain. Worse, doing so may cause us to ping-pong
 893		 * the object in and out of the Global GTT and
 894		 * waste a lot of cycles under the mutex.
 895		 */
 896		if (obj->base.size > ggtt->mappable_end)
 897			return ERR_PTR(-E2BIG);
 898
 899		/*
 900		 * If NONBLOCK is set the caller is optimistically
 901		 * trying to cache the full object within the mappable
 902		 * aperture, and *must* have a fallback in place for
 903		 * situations where we cannot bind the object. We
 904		 * can be a little more lax here and use the fallback
 905		 * more often to avoid costly migrations of ourselves
 906		 * and other objects within the aperture.
 907		 *
 908		 * Half-the-aperture is used as a simple heuristic.
 909		 * More interesting would to do search for a free
 910		 * block prior to making the commitment to unbind.
 911		 * That caters for the self-harm case, and with a
 912		 * little more heuristics (e.g. NOFAULT, NOEVICT)
 913		 * we could try to minimise harm to others.
 914		 */
 915		if (flags & PIN_NONBLOCK &&
 916		    obj->base.size > ggtt->mappable_end / 2)
 917			return ERR_PTR(-ENOSPC);
 918	}
 919
 920new_vma:
 921	vma = i915_vma_instance(obj, &ggtt->vm, view);
 922	if (IS_ERR(vma))
 923		return vma;
 924
 925	if (i915_vma_misplaced(vma, size, alignment, flags)) {
 926		if (flags & PIN_NONBLOCK) {
 927			if (i915_vma_is_pinned(vma) || i915_vma_is_active(vma))
 928				return ERR_PTR(-ENOSPC);
 929
 
 
 
 
 
 
 
 
 
 
 930			if (flags & PIN_MAPPABLE &&
 931			    vma->fence_size > ggtt->mappable_end / 2)
 
 932				return ERR_PTR(-ENOSPC);
 933		}
 934
 935		if (i915_vma_is_pinned(vma) || i915_vma_is_active(vma)) {
 936			discard_ggtt_vma(vma);
 937			goto new_vma;
 938		}
 939
 940		ret = i915_vma_unbind(vma);
 941		if (ret)
 942			return ERR_PTR(ret);
 943	}
 944
 945	if (ww)
 946		ret = i915_vma_pin_ww(vma, ww, size, alignment, flags | PIN_GLOBAL);
 947	else
 948		ret = i915_vma_pin(vma, size, alignment, flags | PIN_GLOBAL);
 949
 950	if (ret)
 951		return ERR_PTR(ret);
 952
 953	if (vma->fence && !i915_gem_object_is_tiled(obj)) {
 954		mutex_lock(&ggtt->vm.mutex);
 955		i915_vma_revoke_fence(vma);
 956		mutex_unlock(&ggtt->vm.mutex);
 957	}
 958
 959	ret = i915_vma_wait_for_bind(vma);
 960	if (ret) {
 961		i915_vma_unpin(vma);
 962		return ERR_PTR(ret);
 963	}
 964
 965	return vma;
 966}
 967
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 968int
 969i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
 970		       struct drm_file *file_priv)
 971{
 972	struct drm_i915_private *i915 = to_i915(dev);
 973	struct drm_i915_gem_madvise *args = data;
 974	struct drm_i915_gem_object *obj;
 975	int err;
 976
 977	switch (args->madv) {
 978	case I915_MADV_DONTNEED:
 979	case I915_MADV_WILLNEED:
 980	    break;
 981	default:
 982	    return -EINVAL;
 983	}
 984
 985	obj = i915_gem_object_lookup(file_priv, args->handle);
 986	if (!obj)
 987		return -ENOENT;
 988
 989	err = i915_gem_object_lock_interruptible(obj, NULL);
 990	if (err)
 991		goto out;
 992
 993	if (i915_gem_object_has_pages(obj) &&
 994	    i915_gem_object_is_tiled(obj) &&
 995	    i915->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
 996		if (obj->mm.madv == I915_MADV_WILLNEED) {
 997			GEM_BUG_ON(!i915_gem_object_has_tiling_quirk(obj));
 998			i915_gem_object_clear_tiling_quirk(obj);
 999			i915_gem_object_make_shrinkable(obj);
1000		}
1001		if (args->madv == I915_MADV_WILLNEED) {
1002			GEM_BUG_ON(i915_gem_object_has_tiling_quirk(obj));
1003			i915_gem_object_make_unshrinkable(obj);
1004			i915_gem_object_set_tiling_quirk(obj);
1005		}
1006	}
1007
1008	if (obj->mm.madv != __I915_MADV_PURGED)
1009		obj->mm.madv = args->madv;
 
 
 
1010
1011	if (i915_gem_object_has_pages(obj)) {
 
1012		unsigned long flags;
1013
1014		spin_lock_irqsave(&i915->mm.obj_lock, flags);
1015		if (!list_empty(&obj->mm.link)) {
1016			struct list_head *list;
1017
1018			if (obj->mm.madv != I915_MADV_WILLNEED)
1019				list = &i915->mm.purge_list;
1020			else
1021				list = &i915->mm.shrink_list;
1022			list_move_tail(&obj->mm.link, list);
1023
1024		}
1025		spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
1026	}
1027
1028	/* if the object is no longer attached, discard its backing storage */
1029	if (obj->mm.madv == I915_MADV_DONTNEED &&
1030	    !i915_gem_object_has_pages(obj))
1031		i915_gem_object_truncate(obj);
1032
1033	args->retained = obj->mm.madv != __I915_MADV_PURGED;
1034
1035	i915_gem_object_unlock(obj);
1036out:
1037	i915_gem_object_put(obj);
1038	return err;
1039}
1040
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1041int i915_gem_init(struct drm_i915_private *dev_priv)
1042{
 
 
1043	int ret;
1044
 
 
 
 
 
 
 
 
 
 
 
 
 
1045	/* We need to fallback to 4K pages if host doesn't support huge gtt. */
1046	if (intel_vgpu_active(dev_priv) && !intel_vgpu_has_huge_gtt(dev_priv))
1047		mkwrite_device_info(dev_priv)->page_sizes =
1048			I915_GTT_PAGE_SIZE_4K;
1049
1050	ret = i915_gem_init_userptr(dev_priv);
1051	if (ret)
1052		return ret;
1053
1054	intel_uc_fetch_firmwares(&dev_priv->gt.uc);
1055	intel_wopcm_init(&dev_priv->wopcm);
 
 
 
 
1056
1057	ret = i915_init_ggtt(dev_priv);
1058	if (ret) {
1059		GEM_BUG_ON(ret == -EIO);
1060		goto err_unlock;
1061	}
1062
1063	/*
1064	 * Despite its name intel_init_clock_gating applies both display
1065	 * clock gating workarounds; GT mmio workarounds and the occasional
1066	 * GT power context workaround. Worse, sometimes it includes a context
1067	 * register workaround which we need to apply before we record the
1068	 * default HW state for all contexts.
1069	 *
1070	 * FIXME: break up the workarounds and apply them at the right time!
1071	 */
1072	intel_init_clock_gating(dev_priv);
1073
1074	ret = intel_gt_init(&dev_priv->gt);
1075	if (ret)
1076		goto err_unlock;
 
 
 
 
 
 
 
 
 
1077
1078	return 0;
1079
1080	/*
1081	 * Unwinding is complicated by that we want to handle -EIO to mean
1082	 * disable GPU submission but keep KMS alive. We want to mark the
1083	 * HW as irrevisibly wedged, but keep enough state around that the
1084	 * driver doesn't explode during runtime.
1085	 */
1086err_unlock:
1087	i915_gem_drain_workqueue(dev_priv);
1088
1089	if (ret != -EIO)
1090		intel_uc_cleanup_firmwares(&dev_priv->gt.uc);
 
 
 
 
 
1091
1092	if (ret == -EIO) {
1093		/*
1094		 * Allow engines or uC initialisation to fail by marking the GPU
1095		 * as wedged. But we only want to do this when the GPU is angry,
1096		 * for all other failure, such as an allocation failure, bail.
1097		 */
1098		if (!intel_gt_is_wedged(&dev_priv->gt)) {
1099			i915_probe_error(dev_priv,
1100					 "Failed to initialize GPU, declaring it wedged!\n");
1101			intel_gt_set_wedged(&dev_priv->gt);
 
 
1102		}
1103
1104		/* Minimal basic recovery for KMS */
1105		ret = i915_ggtt_enable_hw(dev_priv);
1106		i915_ggtt_resume(&dev_priv->ggtt);
1107		intel_init_clock_gating(dev_priv);
1108	}
1109
1110	i915_gem_drain_freed_objects(dev_priv);
1111
1112	return ret;
1113}
1114
1115void i915_gem_driver_register(struct drm_i915_private *i915)
1116{
1117	i915_gem_driver_register__shrinker(i915);
1118
1119	intel_engines_driver_register(i915);
1120}
1121
1122void i915_gem_driver_unregister(struct drm_i915_private *i915)
1123{
1124	i915_gem_driver_unregister__shrinker(i915);
1125}
1126
1127void i915_gem_driver_remove(struct drm_i915_private *dev_priv)
1128{
1129	intel_wakeref_auto_fini(&dev_priv->ggtt.userfault_wakeref);
 
1130
1131	i915_gem_suspend_late(dev_priv);
1132	intel_gt_driver_remove(&dev_priv->gt);
 
1133	dev_priv->uabi_engines = RB_ROOT;
1134
1135	/* Flush any outstanding unpin_work. */
1136	i915_gem_drain_workqueue(dev_priv);
1137
1138	i915_gem_drain_freed_objects(dev_priv);
1139}
1140
1141void i915_gem_driver_release(struct drm_i915_private *dev_priv)
1142{
1143	intel_gt_driver_release(&dev_priv->gt);
 
1144
1145	intel_wa_list_free(&dev_priv->gt_wa_list);
1146
1147	intel_uc_cleanup_firmwares(&dev_priv->gt.uc);
 
1148
1149	i915_gem_drain_freed_objects(dev_priv);
 
1150
1151	drm_WARN_ON(&dev_priv->drm, !list_empty(&dev_priv->gem.contexts.list));
1152}
1153
1154static void i915_gem_init__mm(struct drm_i915_private *i915)
1155{
1156	spin_lock_init(&i915->mm.obj_lock);
1157
1158	init_llist_head(&i915->mm.free_list);
1159
1160	INIT_LIST_HEAD(&i915->mm.purge_list);
1161	INIT_LIST_HEAD(&i915->mm.shrink_list);
1162
1163	i915_gem_init__objects(i915);
1164}
1165
1166void i915_gem_init_early(struct drm_i915_private *dev_priv)
1167{
1168	i915_gem_init__mm(dev_priv);
1169	i915_gem_init__contexts(dev_priv);
1170
1171	spin_lock_init(&dev_priv->fb_tracking.lock);
1172}
1173
1174void i915_gem_cleanup_early(struct drm_i915_private *dev_priv)
1175{
1176	i915_gem_drain_freed_objects(dev_priv);
1177	GEM_BUG_ON(!llist_empty(&dev_priv->mm.free_list));
1178	GEM_BUG_ON(atomic_read(&dev_priv->mm.free_count));
1179	drm_WARN_ON(&dev_priv->drm, dev_priv->mm.shrink_count);
1180}
1181
1182int i915_gem_open(struct drm_i915_private *i915, struct drm_file *file)
1183{
1184	struct drm_i915_file_private *file_priv;
1185	int ret;
 
1186
1187	DRM_DEBUG("\n");
1188
1189	file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL);
1190	if (!file_priv)
1191		return -ENOMEM;
 
 
 
 
1192
1193	file->driver_priv = file_priv;
1194	file_priv->dev_priv = i915;
1195	file_priv->file = file;
 
1196
1197	file_priv->bsd_engine = -1;
1198	file_priv->hang_timestamp = jiffies;
1199
1200	ret = i915_gem_context_open(i915, file);
1201	if (ret)
1202		kfree(file_priv);
1203
1204	return ret;
1205}
1206
1207void i915_gem_ww_ctx_init(struct i915_gem_ww_ctx *ww, bool intr)
1208{
1209	ww_acquire_init(&ww->ctx, &reservation_ww_class);
1210	INIT_LIST_HEAD(&ww->obj_list);
1211	ww->intr = intr;
1212	ww->contended = NULL;
1213}
1214
1215static void i915_gem_ww_ctx_unlock_all(struct i915_gem_ww_ctx *ww)
1216{
1217	struct drm_i915_gem_object *obj;
1218
1219	while ((obj = list_first_entry_or_null(&ww->obj_list, struct drm_i915_gem_object, obj_link))) {
1220		list_del(&obj->obj_link);
1221		i915_gem_object_unlock(obj);
1222	}
1223}
1224
1225void i915_gem_ww_unlock_single(struct drm_i915_gem_object *obj)
1226{
1227	list_del(&obj->obj_link);
1228	i915_gem_object_unlock(obj);
1229}
1230
1231void i915_gem_ww_ctx_fini(struct i915_gem_ww_ctx *ww)
1232{
1233	i915_gem_ww_ctx_unlock_all(ww);
1234	WARN_ON(ww->contended);
1235	ww_acquire_fini(&ww->ctx);
1236}
1237
1238int __must_check i915_gem_ww_ctx_backoff(struct i915_gem_ww_ctx *ww)
1239{
1240	int ret = 0;
1241
1242	if (WARN_ON(!ww->contended))
1243		return -EINVAL;
1244
1245	i915_gem_ww_ctx_unlock_all(ww);
1246	if (ww->intr)
1247		ret = dma_resv_lock_slow_interruptible(ww->contended->base.resv, &ww->ctx);
1248	else
1249		dma_resv_lock_slow(ww->contended->base.resv, &ww->ctx);
1250
1251	if (!ret)
1252		list_add_tail(&ww->contended->obj_link, &ww->obj_list);
1253
1254	ww->contended = NULL;
1255
 
 
 
 
 
1256	return ret;
1257}
1258
1259#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1260#include "selftests/mock_gem_device.c"
1261#include "selftests/i915_gem.c"
1262#endif
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