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 <drm/i915_drm.h>
  30#include <linux/dma-fence-array.h>
  31#include <linux/kthread.h>
  32#include <linux/dma-resv.h>
  33#include <linux/shmem_fs.h>
  34#include <linux/slab.h>
  35#include <linux/stop_machine.h>
  36#include <linux/swap.h>
  37#include <linux/pci.h>
  38#include <linux/dma-buf.h>
  39#include <linux/mman.h>
  40
 
 
 
  41#include "display/intel_display.h"
  42#include "display/intel_frontbuffer.h"
  43
  44#include "gem/i915_gem_clflush.h"
  45#include "gem/i915_gem_context.h"
  46#include "gem/i915_gem_ioctls.h"
 
  47#include "gem/i915_gem_pm.h"
  48#include "gem/i915_gemfs.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_mocs.h"
  53#include "gt/intel_reset.h"
  54#include "gt/intel_renderstate.h"
  55#include "gt/intel_workarounds.h"
  56
  57#include "i915_drv.h"
  58#include "i915_scatterlist.h"
  59#include "i915_trace.h"
  60#include "i915_vgpu.h"
  61
  62#include "intel_pm.h"
  63
  64static int
  65insert_mappable_node(struct i915_ggtt *ggtt,
  66                     struct drm_mm_node *node, u32 size)
  67{
 
 
 
 
 
 
  68	memset(node, 0, sizeof(*node));
  69	return 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
  75static void
  76remove_mappable_node(struct drm_mm_node *node)
  77{
 
  78	drm_mm_remove_node(node);
 
  79}
  80
  81int
  82i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
  83			    struct drm_file *file)
  84{
  85	struct i915_ggtt *ggtt = &to_i915(dev)->ggtt;
 
  86	struct drm_i915_gem_get_aperture *args = data;
  87	struct i915_vma *vma;
  88	u64 pinned;
  89
  90	mutex_lock(&ggtt->vm.mutex);
 
  91
  92	pinned = ggtt->vm.reserved;
  93	list_for_each_entry(vma, &ggtt->vm.bound_list, vm_link)
  94		if (i915_vma_is_pinned(vma))
  95			pinned += vma->node.size;
  96
  97	mutex_unlock(&ggtt->vm.mutex);
  98
  99	args->aper_size = ggtt->vm.total;
 100	args->aper_available_size = args->aper_size - pinned;
 101
 102	return 0;
 103}
 104
 105int i915_gem_object_unbind(struct drm_i915_gem_object *obj,
 106			   unsigned long flags)
 107{
 108	struct i915_vma *vma;
 
 109	LIST_HEAD(still_in_list);
 110	int ret = 0;
 
 
 
 
 111
 112	lockdep_assert_held(&obj->base.dev->struct_mutex);
 
 113
 
 
 
 
 
 
 
 
 
 
 114	spin_lock(&obj->vma.lock);
 115	while (!ret && (vma = list_first_entry_or_null(&obj->vma.list,
 116						       struct i915_vma,
 117						       obj_link))) {
 118		list_move_tail(&vma->obj_link, &still_in_list);
 119		spin_unlock(&obj->vma.lock);
 120
 121		ret = -EBUSY;
 122		if (flags & I915_GEM_OBJECT_UNBIND_ACTIVE ||
 123		    !i915_vma_is_active(vma))
 124			ret = i915_vma_unbind(vma);
 125
 126		spin_lock(&obj->vma.lock);
 127	}
 128	list_splice(&still_in_list, &obj->vma.list);
 129	spin_unlock(&obj->vma.lock);
 130
 131	return ret;
 132}
 133
 134static int
 135i915_gem_phys_pwrite(struct drm_i915_gem_object *obj,
 136		     struct drm_i915_gem_pwrite *args,
 137		     struct drm_file *file)
 138{
 139	void *vaddr = obj->phys_handle->vaddr + args->offset;
 140	char __user *user_data = u64_to_user_ptr(args->data_ptr);
 141
 142	/*
 143	 * We manually control the domain here and pretend that it
 144	 * remains coherent i.e. in the GTT domain, like shmem_pwrite.
 145	 */
 146	intel_frontbuffer_invalidate(obj->frontbuffer, ORIGIN_CPU);
 147
 148	if (copy_from_user(vaddr, user_data, args->size))
 149		return -EFAULT;
 150
 151	drm_clflush_virt_range(vaddr, args->size);
 152	intel_gt_chipset_flush(&to_i915(obj->base.dev)->gt);
 
 
 153
 154	intel_frontbuffer_flush(obj->frontbuffer, ORIGIN_CPU);
 155	return 0;
 156}
 
 
 
 
 
 
 
 157
 158static int
 159i915_gem_create(struct drm_file *file,
 160		struct drm_i915_private *dev_priv,
 161		u64 *size_p,
 162		u32 *handle_p)
 163{
 164	struct drm_i915_gem_object *obj;
 165	u32 handle;
 166	u64 size;
 167	int ret;
 168
 169	size = round_up(*size_p, PAGE_SIZE);
 170	if (size == 0)
 171		return -EINVAL;
 172
 173	/* Allocate the new object */
 174	obj = i915_gem_object_create_shmem(dev_priv, size);
 175	if (IS_ERR(obj))
 176		return PTR_ERR(obj);
 177
 178	ret = drm_gem_handle_create(file, &obj->base, &handle);
 179	/* drop reference from allocate - handle holds it now */
 180	i915_gem_object_put(obj);
 181	if (ret)
 182		return ret;
 183
 184	*handle_p = handle;
 185	*size_p = size;
 186	return 0;
 187}
 
 
 
 
 
 
 
 188
 189int
 190i915_gem_dumb_create(struct drm_file *file,
 191		     struct drm_device *dev,
 192		     struct drm_mode_create_dumb *args)
 193{
 194	int cpp = DIV_ROUND_UP(args->bpp, 8);
 195	u32 format;
 196
 197	switch (cpp) {
 198	case 1:
 199		format = DRM_FORMAT_C8;
 200		break;
 201	case 2:
 202		format = DRM_FORMAT_RGB565;
 203		break;
 204	case 4:
 205		format = DRM_FORMAT_XRGB8888;
 206		break;
 207	default:
 208		return -EINVAL;
 209	}
 
 
 210
 211	/* have to work out size/pitch and return them */
 212	args->pitch = ALIGN(args->width * cpp, 64);
 213
 214	/* align stride to page size so that we can remap */
 215	if (args->pitch > intel_plane_fb_max_stride(to_i915(dev), format,
 216						    DRM_FORMAT_MOD_LINEAR))
 217		args->pitch = ALIGN(args->pitch, 4096);
 218
 219	args->size = args->pitch * args->height;
 220	return i915_gem_create(file, to_i915(dev),
 221			       &args->size, &args->handle);
 222}
 223
 224/**
 225 * Creates a new mm object and returns a handle to it.
 226 * @dev: drm device pointer
 227 * @data: ioctl data blob
 228 * @file: drm file pointer
 229 */
 230int
 231i915_gem_create_ioctl(struct drm_device *dev, void *data,
 232		      struct drm_file *file)
 233{
 234	struct drm_i915_private *dev_priv = to_i915(dev);
 235	struct drm_i915_gem_create *args = data;
 236
 237	i915_gem_flush_free_objects(dev_priv);
 238
 239	return i915_gem_create(file, dev_priv,
 240			       &args->size, &args->handle);
 241}
 242
 243static int
 244shmem_pread(struct page *page, int offset, int len, char __user *user_data,
 245	    bool needs_clflush)
 246{
 247	char *vaddr;
 248	int ret;
 249
 250	vaddr = kmap(page);
 251
 252	if (needs_clflush)
 253		drm_clflush_virt_range(vaddr + offset, len);
 254
 255	ret = __copy_to_user(user_data, vaddr + offset, len);
 256
 257	kunmap(page);
 258
 259	return ret ? -EFAULT : 0;
 260}
 261
 262static int
 263i915_gem_shmem_pread(struct drm_i915_gem_object *obj,
 264		     struct drm_i915_gem_pread *args)
 265{
 266	unsigned int needs_clflush;
 267	unsigned int idx, offset;
 268	struct dma_fence *fence;
 269	char __user *user_data;
 270	u64 remain;
 271	int ret;
 272
 273	ret = i915_gem_object_prepare_read(obj, &needs_clflush);
 274	if (ret)
 275		return ret;
 276
 277	fence = i915_gem_object_lock_fence(obj);
 
 
 
 
 
 
 
 278	i915_gem_object_finish_access(obj);
 279	if (!fence)
 280		return -ENOMEM;
 281
 282	remain = args->size;
 283	user_data = u64_to_user_ptr(args->data_ptr);
 284	offset = offset_in_page(args->offset);
 285	for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
 286		struct page *page = i915_gem_object_get_page(obj, idx);
 287		unsigned int length = min_t(u64, remain, PAGE_SIZE - offset);
 288
 289		ret = shmem_pread(page, offset, length, user_data,
 290				  needs_clflush);
 291		if (ret)
 292			break;
 293
 294		remain -= length;
 295		user_data += length;
 296		offset = 0;
 297	}
 298
 299	i915_gem_object_unlock_fence(obj, fence);
 
 
 
 
 
 
 300	return ret;
 301}
 302
 303static inline bool
 304gtt_user_read(struct io_mapping *mapping,
 305	      loff_t base, int offset,
 306	      char __user *user_data, int length)
 307{
 308	void __iomem *vaddr;
 309	unsigned long unwritten;
 310
 311	/* We can use the cpu mem copy function because this is X86. */
 312	vaddr = io_mapping_map_atomic_wc(mapping, base);
 313	unwritten = __copy_to_user_inatomic(user_data,
 314					    (void __force *)vaddr + offset,
 315					    length);
 316	io_mapping_unmap_atomic(vaddr);
 317	if (unwritten) {
 318		vaddr = io_mapping_map_wc(mapping, base, PAGE_SIZE);
 319		unwritten = copy_to_user(user_data,
 320					 (void __force *)vaddr + offset,
 321					 length);
 322		io_mapping_unmap(vaddr);
 323	}
 324	return unwritten;
 325}
 326
 327static int
 328i915_gem_gtt_pread(struct drm_i915_gem_object *obj,
 329		   const struct drm_i915_gem_pread *args)
 330{
 331	struct drm_i915_private *i915 = to_i915(obj->base.dev);
 332	struct i915_ggtt *ggtt = &i915->ggtt;
 333	intel_wakeref_t wakeref;
 334	struct drm_mm_node node;
 335	struct dma_fence *fence;
 336	void __user *user_data;
 337	struct i915_vma *vma;
 338	u64 remain, offset;
 339	int ret;
 340
 341	ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
 
 
 
 342	if (ret)
 343		return ret;
 
 
 
 
 344
 345	wakeref = intel_runtime_pm_get(&i915->runtime_pm);
 346	vma = ERR_PTR(-ENODEV);
 347	if (!i915_gem_object_is_tiled(obj))
 348		vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
 349					       PIN_MAPPABLE |
 350					       PIN_NONBLOCK /* NOWARN */ |
 351					       PIN_NOEVICT);
 352	if (!IS_ERR(vma)) {
 353		node.start = i915_ggtt_offset(vma);
 354		node.allocated = false;
 
 
 
 355	} else {
 356		ret = insert_mappable_node(ggtt, &node, PAGE_SIZE);
 357		if (ret)
 358			goto out_unlock;
 359		GEM_BUG_ON(!node.allocated);
 
 360	}
 361
 362	mutex_unlock(&i915->drm.struct_mutex);
 
 
 
 
 
 
 
 
 363
 364	ret = i915_gem_object_lock_interruptible(obj);
 365	if (ret)
 366		goto out_unpin;
 
 
 
 
 367
 368	ret = i915_gem_object_set_to_gtt_domain(obj, false);
 369	if (ret) {
 370		i915_gem_object_unlock(obj);
 371		goto out_unpin;
 
 
 
 
 
 
 
 
 
 
 
 
 372	}
 
 373
 374	fence = i915_gem_object_lock_fence(obj);
 375	i915_gem_object_unlock(obj);
 376	if (!fence) {
 377		ret = -ENOMEM;
 378		goto out_unpin;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 379	}
 380
 381	user_data = u64_to_user_ptr(args->data_ptr);
 382	remain = args->size;
 383	offset = args->offset;
 384
 385	while (remain > 0) {
 386		/* Operation in this page
 387		 *
 388		 * page_base = page offset within aperture
 389		 * page_offset = offset within page
 390		 * page_length = bytes to copy for this page
 391		 */
 392		u32 page_base = node.start;
 393		unsigned page_offset = offset_in_page(offset);
 394		unsigned page_length = PAGE_SIZE - page_offset;
 395		page_length = remain < page_length ? remain : page_length;
 396		if (node.allocated) {
 397			ggtt->vm.insert_page(&ggtt->vm,
 398					     i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT),
 399					     node.start, I915_CACHE_NONE, 0);
 400		} else {
 401			page_base += offset & PAGE_MASK;
 402		}
 403
 404		if (gtt_user_read(&ggtt->iomap, page_base, page_offset,
 405				  user_data, page_length)) {
 406			ret = -EFAULT;
 407			break;
 408		}
 409
 410		remain -= page_length;
 411		user_data += page_length;
 412		offset += page_length;
 413	}
 414
 415	i915_gem_object_unlock_fence(obj, fence);
 416out_unpin:
 417	mutex_lock(&i915->drm.struct_mutex);
 418	if (node.allocated) {
 419		ggtt->vm.clear_range(&ggtt->vm, node.start, node.size);
 420		remove_mappable_node(&node);
 421	} else {
 422		i915_vma_unpin(vma);
 423	}
 424out_unlock:
 425	intel_runtime_pm_put(&i915->runtime_pm, wakeref);
 426	mutex_unlock(&i915->drm.struct_mutex);
 427
 428	return ret;
 429}
 430
 431/**
 432 * Reads data from the object referenced by handle.
 433 * @dev: drm device pointer
 434 * @data: ioctl data blob
 435 * @file: drm file pointer
 436 *
 437 * On error, the contents of *data are undefined.
 438 */
 439int
 440i915_gem_pread_ioctl(struct drm_device *dev, void *data,
 441		     struct drm_file *file)
 442{
 
 443	struct drm_i915_gem_pread *args = data;
 444	struct drm_i915_gem_object *obj;
 445	int ret;
 446
 
 
 
 
 
 
 447	if (args->size == 0)
 448		return 0;
 449
 450	if (!access_ok(u64_to_user_ptr(args->data_ptr),
 451		       args->size))
 452		return -EFAULT;
 453
 454	obj = i915_gem_object_lookup(file, args->handle);
 455	if (!obj)
 456		return -ENOENT;
 457
 458	/* Bounds check source.  */
 459	if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) {
 460		ret = -EINVAL;
 461		goto out;
 462	}
 463
 464	trace_i915_gem_object_pread(obj, args->offset, args->size);
 
 
 
 
 
 465
 466	ret = i915_gem_object_wait(obj,
 467				   I915_WAIT_INTERRUPTIBLE,
 468				   MAX_SCHEDULE_TIMEOUT);
 469	if (ret)
 470		goto out;
 471
 472	ret = i915_gem_object_pin_pages(obj);
 473	if (ret)
 474		goto out;
 475
 476	ret = i915_gem_shmem_pread(obj, args);
 477	if (ret == -EFAULT || ret == -ENODEV)
 478		ret = i915_gem_gtt_pread(obj, args);
 479
 480	i915_gem_object_unpin_pages(obj);
 481out:
 482	i915_gem_object_put(obj);
 483	return ret;
 484}
 485
 486/* This is the fast write path which cannot handle
 487 * page faults in the source data
 488 */
 489
 490static inline bool
 491ggtt_write(struct io_mapping *mapping,
 492	   loff_t base, int offset,
 493	   char __user *user_data, int length)
 494{
 495	void __iomem *vaddr;
 496	unsigned long unwritten;
 497
 498	/* We can use the cpu mem copy function because this is X86. */
 499	vaddr = io_mapping_map_atomic_wc(mapping, base);
 500	unwritten = __copy_from_user_inatomic_nocache((void __force *)vaddr + offset,
 501						      user_data, length);
 502	io_mapping_unmap_atomic(vaddr);
 503	if (unwritten) {
 504		vaddr = io_mapping_map_wc(mapping, base, PAGE_SIZE);
 505		unwritten = copy_from_user((void __force *)vaddr + offset,
 506					   user_data, length);
 507		io_mapping_unmap(vaddr);
 508	}
 509
 510	return unwritten;
 511}
 512
 513/**
 514 * This is the fast pwrite path, where we copy the data directly from the
 515 * user into the GTT, uncached.
 516 * @obj: i915 GEM object
 517 * @args: pwrite arguments structure
 518 */
 519static int
 520i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
 521			 const struct drm_i915_gem_pwrite *args)
 522{
 523	struct drm_i915_private *i915 = to_i915(obj->base.dev);
 524	struct i915_ggtt *ggtt = &i915->ggtt;
 525	struct intel_runtime_pm *rpm = &i915->runtime_pm;
 526	intel_wakeref_t wakeref;
 527	struct drm_mm_node node;
 528	struct dma_fence *fence;
 529	struct i915_vma *vma;
 530	u64 remain, offset;
 531	void __user *user_data;
 532	int ret;
 533
 534	ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
 535	if (ret)
 536		return ret;
 537
 538	if (i915_gem_object_has_struct_page(obj)) {
 539		/*
 540		 * Avoid waking the device up if we can fallback, as
 541		 * waking/resuming is very slow (worst-case 10-100 ms
 542		 * depending on PCI sleeps and our own resume time).
 543		 * This easily dwarfs any performance advantage from
 544		 * using the cache bypass of indirect GGTT access.
 545		 */
 546		wakeref = intel_runtime_pm_get_if_in_use(rpm);
 547		if (!wakeref) {
 548			ret = -EFAULT;
 549			goto out_unlock;
 550		}
 551	} else {
 552		/* No backing pages, no fallback, we must force GGTT access */
 553		wakeref = intel_runtime_pm_get(rpm);
 554	}
 555
 556	vma = ERR_PTR(-ENODEV);
 557	if (!i915_gem_object_is_tiled(obj))
 558		vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
 559					       PIN_MAPPABLE |
 560					       PIN_NONBLOCK /* NOWARN */ |
 561					       PIN_NOEVICT);
 562	if (!IS_ERR(vma)) {
 563		node.start = i915_ggtt_offset(vma);
 564		node.allocated = false;
 565	} else {
 566		ret = insert_mappable_node(ggtt, &node, PAGE_SIZE);
 567		if (ret)
 568			goto out_rpm;
 569		GEM_BUG_ON(!node.allocated);
 570	}
 571
 572	mutex_unlock(&i915->drm.struct_mutex);
 573
 574	ret = i915_gem_object_lock_interruptible(obj);
 575	if (ret)
 576		goto out_unpin;
 577
 578	ret = i915_gem_object_set_to_gtt_domain(obj, true);
 579	if (ret) {
 580		i915_gem_object_unlock(obj);
 581		goto out_unpin;
 582	}
 583
 584	fence = i915_gem_object_lock_fence(obj);
 585	i915_gem_object_unlock(obj);
 586	if (!fence) {
 587		ret = -ENOMEM;
 588		goto out_unpin;
 589	}
 590
 591	intel_frontbuffer_invalidate(obj->frontbuffer, ORIGIN_CPU);
 592
 593	user_data = u64_to_user_ptr(args->data_ptr);
 594	offset = args->offset;
 595	remain = args->size;
 596	while (remain) {
 597		/* Operation in this page
 598		 *
 599		 * page_base = page offset within aperture
 600		 * page_offset = offset within page
 601		 * page_length = bytes to copy for this page
 602		 */
 603		u32 page_base = node.start;
 604		unsigned int page_offset = offset_in_page(offset);
 605		unsigned int page_length = PAGE_SIZE - page_offset;
 606		page_length = remain < page_length ? remain : page_length;
 607		if (node.allocated) {
 608			/* flush the write before we modify the GGTT */
 609			intel_gt_flush_ggtt_writes(ggtt->vm.gt);
 610			ggtt->vm.insert_page(&ggtt->vm,
 611					     i915_gem_object_get_dma_address(obj, offset >> PAGE_SHIFT),
 612					     node.start, I915_CACHE_NONE, 0);
 613			wmb(); /* flush modifications to the GGTT (insert_page) */
 614		} else {
 615			page_base += offset & PAGE_MASK;
 616		}
 617		/* If we get a fault while copying data, then (presumably) our
 618		 * source page isn't available.  Return the error and we'll
 619		 * retry in the slow path.
 620		 * If the object is non-shmem backed, we retry again with the
 621		 * path that handles page fault.
 622		 */
 623		if (ggtt_write(&ggtt->iomap, page_base, page_offset,
 624			       user_data, page_length)) {
 625			ret = -EFAULT;
 626			break;
 627		}
 628
 629		remain -= page_length;
 630		user_data += page_length;
 631		offset += page_length;
 632	}
 633	intel_frontbuffer_flush(obj->frontbuffer, ORIGIN_CPU);
 634
 635	i915_gem_object_unlock_fence(obj, fence);
 636out_unpin:
 637	mutex_lock(&i915->drm.struct_mutex);
 638	intel_gt_flush_ggtt_writes(ggtt->vm.gt);
 639	if (node.allocated) {
 640		ggtt->vm.clear_range(&ggtt->vm, node.start, node.size);
 641		remove_mappable_node(&node);
 642	} else {
 643		i915_vma_unpin(vma);
 644	}
 645out_rpm:
 646	intel_runtime_pm_put(rpm, wakeref);
 647out_unlock:
 648	mutex_unlock(&i915->drm.struct_mutex);
 649	return ret;
 650}
 651
 652/* Per-page copy function for the shmem pwrite fastpath.
 653 * Flushes invalid cachelines before writing to the target if
 654 * needs_clflush_before is set and flushes out any written cachelines after
 655 * writing if needs_clflush is set.
 656 */
 657static int
 658shmem_pwrite(struct page *page, int offset, int len, char __user *user_data,
 659	     bool needs_clflush_before,
 660	     bool needs_clflush_after)
 661{
 662	char *vaddr;
 663	int ret;
 664
 665	vaddr = kmap(page);
 666
 667	if (needs_clflush_before)
 668		drm_clflush_virt_range(vaddr + offset, len);
 669
 670	ret = __copy_from_user(vaddr + offset, user_data, len);
 671	if (!ret && needs_clflush_after)
 672		drm_clflush_virt_range(vaddr + offset, len);
 673
 674	kunmap(page);
 675
 676	return ret ? -EFAULT : 0;
 677}
 678
 679static int
 680i915_gem_shmem_pwrite(struct drm_i915_gem_object *obj,
 681		      const struct drm_i915_gem_pwrite *args)
 682{
 683	unsigned int partial_cacheline_write;
 684	unsigned int needs_clflush;
 685	unsigned int offset, idx;
 686	struct dma_fence *fence;
 687	void __user *user_data;
 688	u64 remain;
 689	int ret;
 690
 691	ret = i915_gem_object_prepare_write(obj, &needs_clflush);
 692	if (ret)
 693		return ret;
 694
 695	fence = i915_gem_object_lock_fence(obj);
 
 
 
 
 
 
 
 696	i915_gem_object_finish_access(obj);
 697	if (!fence)
 698		return -ENOMEM;
 699
 700	/* If we don't overwrite a cacheline completely we need to be
 701	 * careful to have up-to-date data by first clflushing. Don't
 702	 * overcomplicate things and flush the entire patch.
 703	 */
 704	partial_cacheline_write = 0;
 705	if (needs_clflush & CLFLUSH_BEFORE)
 706		partial_cacheline_write = boot_cpu_data.x86_clflush_size - 1;
 707
 708	user_data = u64_to_user_ptr(args->data_ptr);
 709	remain = args->size;
 710	offset = offset_in_page(args->offset);
 711	for (idx = args->offset >> PAGE_SHIFT; remain; idx++) {
 712		struct page *page = i915_gem_object_get_page(obj, idx);
 713		unsigned int length = min_t(u64, remain, PAGE_SIZE - offset);
 714
 715		ret = shmem_pwrite(page, offset, length, user_data,
 716				   (offset | length) & partial_cacheline_write,
 717				   needs_clflush & CLFLUSH_AFTER);
 718		if (ret)
 719			break;
 720
 721		remain -= length;
 722		user_data += length;
 723		offset = 0;
 724	}
 725
 726	intel_frontbuffer_flush(obj->frontbuffer, ORIGIN_CPU);
 727	i915_gem_object_unlock_fence(obj, fence);
 728
 
 
 
 
 
 
 
 729	return ret;
 730}
 731
 732/**
 733 * Writes data to the object referenced by handle.
 734 * @dev: drm device
 735 * @data: ioctl data blob
 736 * @file: drm file
 737 *
 738 * On error, the contents of the buffer that were to be modified are undefined.
 739 */
 740int
 741i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
 742		      struct drm_file *file)
 743{
 
 744	struct drm_i915_gem_pwrite *args = data;
 745	struct drm_i915_gem_object *obj;
 746	int ret;
 747
 
 
 
 
 
 
 748	if (args->size == 0)
 749		return 0;
 750
 751	if (!access_ok(u64_to_user_ptr(args->data_ptr), args->size))
 752		return -EFAULT;
 753
 754	obj = i915_gem_object_lookup(file, args->handle);
 755	if (!obj)
 756		return -ENOENT;
 757
 758	/* Bounds check destination. */
 759	if (range_overflows_t(u64, args->offset, args->size, obj->base.size)) {
 760		ret = -EINVAL;
 761		goto err;
 762	}
 763
 764	/* Writes not allowed into this read-only object */
 765	if (i915_gem_object_is_readonly(obj)) {
 766		ret = -EINVAL;
 767		goto err;
 768	}
 769
 770	trace_i915_gem_object_pwrite(obj, args->offset, args->size);
 771
 772	ret = -ENODEV;
 773	if (obj->ops->pwrite)
 774		ret = obj->ops->pwrite(obj, args);
 775	if (ret != -ENODEV)
 776		goto err;
 777
 778	ret = i915_gem_object_wait(obj,
 779				   I915_WAIT_INTERRUPTIBLE |
 780				   I915_WAIT_ALL,
 781				   MAX_SCHEDULE_TIMEOUT);
 782	if (ret)
 783		goto err;
 784
 785	ret = i915_gem_object_pin_pages(obj);
 786	if (ret)
 787		goto err;
 788
 789	ret = -EFAULT;
 790	/* We can only do the GTT pwrite on untiled buffers, as otherwise
 791	 * it would end up going through the fenced access, and we'll get
 792	 * different detiling behavior between reading and writing.
 793	 * pread/pwrite currently are reading and writing from the CPU
 794	 * perspective, requiring manual detiling by the client.
 795	 */
 796	if (!i915_gem_object_has_struct_page(obj) ||
 797	    cpu_write_needs_clflush(obj))
 798		/* Note that the gtt paths might fail with non-page-backed user
 799		 * pointers (e.g. gtt mappings when moving data between
 800		 * textures). Fallback to the shmem path in that case.
 801		 */
 802		ret = i915_gem_gtt_pwrite_fast(obj, args);
 803
 804	if (ret == -EFAULT || ret == -ENOSPC) {
 805		if (obj->phys_handle)
 806			ret = i915_gem_phys_pwrite(obj, args, file);
 807		else
 808			ret = i915_gem_shmem_pwrite(obj, args);
 809	}
 810
 811	i915_gem_object_unpin_pages(obj);
 812err:
 813	i915_gem_object_put(obj);
 814	return ret;
 815}
 816
 817/**
 818 * Called when user space has done writes to this buffer
 819 * @dev: drm device
 820 * @data: ioctl data blob
 821 * @file: drm file
 822 */
 823int
 824i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
 825			 struct drm_file *file)
 826{
 827	struct drm_i915_gem_sw_finish *args = data;
 828	struct drm_i915_gem_object *obj;
 829
 830	obj = i915_gem_object_lookup(file, args->handle);
 831	if (!obj)
 832		return -ENOENT;
 833
 834	/*
 835	 * Proxy objects are barred from CPU access, so there is no
 836	 * need to ban sw_finish as it is a nop.
 837	 */
 838
 839	/* Pinned buffers may be scanout, so flush the cache */
 840	i915_gem_object_flush_if_display(obj);
 841	i915_gem_object_put(obj);
 842
 843	return 0;
 844}
 845
 846void i915_gem_runtime_suspend(struct drm_i915_private *i915)
 847{
 848	struct drm_i915_gem_object *obj, *on;
 849	int i;
 850
 851	/*
 852	 * Only called during RPM suspend. All users of the userfault_list
 853	 * must be holding an RPM wakeref to ensure that this can not
 854	 * run concurrently with themselves (and use the struct_mutex for
 855	 * protection between themselves).
 856	 */
 857
 858	list_for_each_entry_safe(obj, on,
 859				 &i915->ggtt.userfault_list, userfault_link)
 860		__i915_gem_object_release_mmap(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 < i915->ggtt.num_fences; i++) {
 868		struct i915_fence_reg *reg = &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 long
 891wait_for_timelines(struct drm_i915_private *i915,
 892		   unsigned int wait, long timeout)
 893{
 894	struct intel_gt_timelines *timelines = &i915->gt.timelines;
 895	struct intel_timeline *tl;
 896	unsigned long flags;
 897
 898	spin_lock_irqsave(&timelines->lock, flags);
 899	list_for_each_entry(tl, &timelines->active_list, link) {
 900		struct i915_request *rq;
 901
 902		rq = i915_active_request_get_unlocked(&tl->last_request);
 903		if (!rq)
 904			continue;
 905
 906		spin_unlock_irqrestore(&timelines->lock, flags);
 907
 908		/*
 909		 * "Race-to-idle".
 910		 *
 911		 * Switching to the kernel context is often used a synchronous
 912		 * step prior to idling, e.g. in suspend for flushing all
 913		 * current operations to memory before sleeping. These we
 914		 * want to complete as quickly as possible to avoid prolonged
 915		 * stalls, so allow the gpu to boost to maximum clocks.
 916		 */
 917		if (wait & I915_WAIT_FOR_IDLE_BOOST)
 918			gen6_rps_boost(rq);
 919
 920		timeout = i915_request_wait(rq, wait, timeout);
 921		i915_request_put(rq);
 922		if (timeout < 0)
 923			return timeout;
 924
 925		/* restart after reacquiring the lock */
 926		spin_lock_irqsave(&timelines->lock, flags);
 927		tl = list_entry(&timelines->active_list, typeof(*tl), link);
 928	}
 929	spin_unlock_irqrestore(&timelines->lock, flags);
 930
 931	return timeout;
 932}
 933
 934int i915_gem_wait_for_idle(struct drm_i915_private *i915,
 935			   unsigned int flags, long timeout)
 936{
 937	/* If the device is asleep, we have no requests outstanding */
 938	if (!intel_gt_pm_is_awake(&i915->gt))
 939		return 0;
 940
 941	GEM_TRACE("flags=%x (%s), timeout=%ld%s\n",
 942		  flags, flags & I915_WAIT_LOCKED ? "locked" : "unlocked",
 943		  timeout, timeout == MAX_SCHEDULE_TIMEOUT ? " (forever)" : "");
 944
 945	timeout = wait_for_timelines(i915, flags, timeout);
 946	if (timeout < 0)
 947		return timeout;
 948
 949	if (flags & I915_WAIT_LOCKED) {
 950		lockdep_assert_held(&i915->drm.struct_mutex);
 951
 952		i915_retire_requests(i915);
 953	}
 954
 955	return 0;
 956}
 957
 958struct i915_vma *
 959i915_gem_object_ggtt_pin(struct drm_i915_gem_object *obj,
 960			 const struct i915_ggtt_view *view,
 961			 u64 size,
 962			 u64 alignment,
 963			 u64 flags)
 964{
 965	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
 966	struct i915_address_space *vm = &dev_priv->ggtt.vm;
 967
 968	return i915_gem_object_pin(obj, vm, view, size, alignment,
 969				   flags | PIN_GLOBAL);
 970}
 971
 972struct i915_vma *
 973i915_gem_object_pin(struct drm_i915_gem_object *obj,
 974		    struct i915_address_space *vm,
 975		    const struct i915_ggtt_view *view,
 976		    u64 size,
 977		    u64 alignment,
 978		    u64 flags)
 979{
 980	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
 
 981	struct i915_vma *vma;
 982	int ret;
 983
 984	lockdep_assert_held(&obj->base.dev->struct_mutex);
 985
 986	if (i915_gem_object_never_bind_ggtt(obj))
 987		return ERR_PTR(-ENODEV);
 988
 989	if (flags & PIN_MAPPABLE &&
 990	    (!view || view->type == I915_GGTT_VIEW_NORMAL)) {
 991		/* If the required space is larger than the available
 
 992		 * aperture, we will not able to find a slot for the
 993		 * object and unbinding the object now will be in
 994		 * vain. Worse, doing so may cause us to ping-pong
 995		 * the object in and out of the Global GTT and
 996		 * waste a lot of cycles under the mutex.
 997		 */
 998		if (obj->base.size > dev_priv->ggtt.mappable_end)
 999			return ERR_PTR(-E2BIG);
1000
1001		/* If NONBLOCK is set the caller is optimistically
 
1002		 * trying to cache the full object within the mappable
1003		 * aperture, and *must* have a fallback in place for
1004		 * situations where we cannot bind the object. We
1005		 * can be a little more lax here and use the fallback
1006		 * more often to avoid costly migrations of ourselves
1007		 * and other objects within the aperture.
1008		 *
1009		 * Half-the-aperture is used as a simple heuristic.
1010		 * More interesting would to do search for a free
1011		 * block prior to making the commitment to unbind.
1012		 * That caters for the self-harm case, and with a
1013		 * little more heuristics (e.g. NOFAULT, NOEVICT)
1014		 * we could try to minimise harm to others.
1015		 */
1016		if (flags & PIN_NONBLOCK &&
1017		    obj->base.size > dev_priv->ggtt.mappable_end / 2)
1018			return ERR_PTR(-ENOSPC);
1019	}
1020
1021	vma = i915_vma_instance(obj, vm, view);
 
1022	if (IS_ERR(vma))
1023		return vma;
1024
1025	if (i915_vma_misplaced(vma, size, alignment, flags)) {
1026		if (flags & PIN_NONBLOCK) {
1027			if (i915_vma_is_pinned(vma) || i915_vma_is_active(vma))
1028				return ERR_PTR(-ENOSPC);
1029
 
 
 
 
 
 
 
 
 
 
1030			if (flags & PIN_MAPPABLE &&
1031			    vma->fence_size > dev_priv->ggtt.mappable_end / 2)
 
1032				return ERR_PTR(-ENOSPC);
1033		}
1034
1035		WARN(i915_vma_is_pinned(vma),
1036		     "bo is already pinned in ggtt with incorrect alignment:"
1037		     " offset=%08x, req.alignment=%llx,"
1038		     " req.map_and_fenceable=%d, vma->map_and_fenceable=%d\n",
1039		     i915_ggtt_offset(vma), alignment,
1040		     !!(flags & PIN_MAPPABLE),
1041		     i915_vma_is_map_and_fenceable(vma));
1042		ret = i915_vma_unbind(vma);
1043		if (ret)
1044			return ERR_PTR(ret);
1045	}
1046
 
 
 
 
 
1047	if (vma->fence && !i915_gem_object_is_tiled(obj)) {
1048		mutex_lock(&vma->vm->mutex);
1049		ret = i915_vma_revoke_fence(vma);
1050		mutex_unlock(&vma->vm->mutex);
1051		if (ret)
1052			return ERR_PTR(ret);
1053	}
1054
1055	ret = i915_vma_pin(vma, size, alignment, flags);
1056	if (ret)
 
1057		return ERR_PTR(ret);
 
1058
1059	return vma;
1060}
1061
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1062int
1063i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
1064		       struct drm_file *file_priv)
1065{
1066	struct drm_i915_private *i915 = to_i915(dev);
1067	struct drm_i915_gem_madvise *args = data;
1068	struct drm_i915_gem_object *obj;
1069	int err;
1070
1071	switch (args->madv) {
1072	case I915_MADV_DONTNEED:
1073	case I915_MADV_WILLNEED:
1074	    break;
1075	default:
1076	    return -EINVAL;
1077	}
1078
1079	obj = i915_gem_object_lookup(file_priv, args->handle);
1080	if (!obj)
1081		return -ENOENT;
1082
1083	err = mutex_lock_interruptible(&obj->mm.lock);
1084	if (err)
1085		goto out;
1086
1087	if (i915_gem_object_has_pages(obj) &&
1088	    i915_gem_object_is_tiled(obj) &&
1089	    i915->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
1090		if (obj->mm.madv == I915_MADV_WILLNEED) {
1091			GEM_BUG_ON(!obj->mm.quirked);
1092			__i915_gem_object_unpin_pages(obj);
1093			obj->mm.quirked = false;
1094		}
1095		if (args->madv == I915_MADV_WILLNEED) {
1096			GEM_BUG_ON(obj->mm.quirked);
1097			__i915_gem_object_pin_pages(obj);
1098			obj->mm.quirked = true;
1099		}
1100	}
1101
1102	if (obj->mm.madv != __I915_MADV_PURGED)
1103		obj->mm.madv = args->madv;
 
 
 
1104
1105	if (i915_gem_object_has_pages(obj)) {
1106		struct list_head *list;
1107
1108		if (i915_gem_object_is_shrinkable(obj)) {
1109			unsigned long flags;
1110
1111			spin_lock_irqsave(&i915->mm.obj_lock, flags);
 
 
1112
1113			if (obj->mm.madv != I915_MADV_WILLNEED)
1114				list = &i915->mm.purge_list;
1115			else
1116				list = &i915->mm.shrink_list;
1117			list_move_tail(&obj->mm.link, list);
1118
1119			spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
1120		}
 
1121	}
1122
1123	/* if the object is no longer attached, discard its backing storage */
1124	if (obj->mm.madv == I915_MADV_DONTNEED &&
1125	    !i915_gem_object_has_pages(obj))
1126		i915_gem_object_truncate(obj);
1127
1128	args->retained = obj->mm.madv != __I915_MADV_PURGED;
1129	mutex_unlock(&obj->mm.lock);
1130
 
1131out:
1132	i915_gem_object_put(obj);
1133	return err;
1134}
1135
1136void i915_gem_sanitize(struct drm_i915_private *i915)
1137{
1138	intel_wakeref_t wakeref;
1139
1140	GEM_TRACE("\n");
1141
1142	wakeref = intel_runtime_pm_get(&i915->runtime_pm);
1143	intel_uncore_forcewake_get(&i915->uncore, FORCEWAKE_ALL);
1144
1145	/*
1146	 * As we have just resumed the machine and woken the device up from
1147	 * deep PCI sleep (presumably D3_cold), assume the HW has been reset
1148	 * back to defaults, recovering from whatever wedged state we left it
1149	 * in and so worth trying to use the device once more.
1150	 */
1151	if (intel_gt_is_wedged(&i915->gt))
1152		intel_gt_unset_wedged(&i915->gt);
1153
1154	/*
1155	 * If we inherit context state from the BIOS or earlier occupants
1156	 * of the GPU, the GPU may be in an inconsistent state when we
1157	 * try to take over. The only way to remove the earlier state
1158	 * is by resetting. However, resetting on earlier gen is tricky as
1159	 * it may impact the display and we are uncertain about the stability
1160	 * of the reset, so this could be applied to even earlier gen.
1161	 */
1162	intel_gt_sanitize(&i915->gt, false);
1163
1164	intel_uncore_forcewake_put(&i915->uncore, FORCEWAKE_ALL);
1165	intel_runtime_pm_put(&i915->runtime_pm, wakeref);
1166}
1167
1168static void init_unused_ring(struct intel_gt *gt, u32 base)
1169{
1170	struct intel_uncore *uncore = gt->uncore;
1171
1172	intel_uncore_write(uncore, RING_CTL(base), 0);
1173	intel_uncore_write(uncore, RING_HEAD(base), 0);
1174	intel_uncore_write(uncore, RING_TAIL(base), 0);
1175	intel_uncore_write(uncore, RING_START(base), 0);
1176}
1177
1178static void init_unused_rings(struct intel_gt *gt)
1179{
1180	struct drm_i915_private *i915 = gt->i915;
1181
1182	if (IS_I830(i915)) {
1183		init_unused_ring(gt, PRB1_BASE);
1184		init_unused_ring(gt, SRB0_BASE);
1185		init_unused_ring(gt, SRB1_BASE);
1186		init_unused_ring(gt, SRB2_BASE);
1187		init_unused_ring(gt, SRB3_BASE);
1188	} else if (IS_GEN(i915, 2)) {
1189		init_unused_ring(gt, SRB0_BASE);
1190		init_unused_ring(gt, SRB1_BASE);
1191	} else if (IS_GEN(i915, 3)) {
1192		init_unused_ring(gt, PRB1_BASE);
1193		init_unused_ring(gt, PRB2_BASE);
1194	}
1195}
1196
1197int i915_gem_init_hw(struct drm_i915_private *i915)
1198{
1199	struct intel_uncore *uncore = &i915->uncore;
1200	struct intel_gt *gt = &i915->gt;
1201	int ret;
1202
1203	BUG_ON(!i915->kernel_context);
1204	ret = intel_gt_terminally_wedged(gt);
1205	if (ret)
1206		return ret;
1207
1208	gt->last_init_time = ktime_get();
1209
1210	/* Double layer security blanket, see i915_gem_init() */
1211	intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
1212
1213	if (HAS_EDRAM(i915) && INTEL_GEN(i915) < 9)
1214		intel_uncore_rmw(uncore, HSW_IDICR, 0, IDIHASHMSK(0xf));
1215
1216	if (IS_HASWELL(i915))
1217		intel_uncore_write(uncore,
1218				   MI_PREDICATE_RESULT_2,
1219				   IS_HSW_GT3(i915) ?
1220				   LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED);
1221
1222	/* Apply the GT workarounds... */
1223	intel_gt_apply_workarounds(gt);
1224	/* ...and determine whether they are sticking. */
1225	intel_gt_verify_workarounds(gt, "init");
1226
1227	intel_gt_init_swizzling(gt);
1228
1229	/*
1230	 * At least 830 can leave some of the unused rings
1231	 * "active" (ie. head != tail) after resume which
1232	 * will prevent c3 entry. Makes sure all unused rings
1233	 * are totally idle.
1234	 */
1235	init_unused_rings(gt);
1236
1237	ret = i915_ppgtt_init_hw(gt);
1238	if (ret) {
1239		DRM_ERROR("Enabling PPGTT failed (%d)\n", ret);
1240		goto out;
1241	}
1242
1243	/* We can't enable contexts until all firmware is loaded */
1244	ret = intel_uc_init_hw(&gt->uc);
1245	if (ret) {
1246		i915_probe_error(i915, "Enabling uc failed (%d)\n", ret);
1247		goto out;
1248	}
1249
1250	intel_mocs_init(gt);
1251
1252out:
1253	intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
1254	return ret;
1255}
1256
1257static int __intel_engines_record_defaults(struct drm_i915_private *i915)
1258{
1259	struct i915_request *requests[I915_NUM_ENGINES] = {};
1260	struct intel_engine_cs *engine;
1261	enum intel_engine_id id;
1262	int err = 0;
1263
1264	/*
1265	 * As we reset the gpu during very early sanitisation, the current
1266	 * register state on the GPU should reflect its defaults values.
1267	 * We load a context onto the hw (with restore-inhibit), then switch
1268	 * over to a second context to save that default register state. We
1269	 * can then prime every new context with that state so they all start
1270	 * from the same default HW values.
1271	 */
1272
1273	for_each_engine(engine, i915, id) {
1274		struct intel_context *ce;
1275		struct i915_request *rq;
1276
1277		/* We must be able to switch to something! */
1278		GEM_BUG_ON(!engine->kernel_context);
1279		engine->serial++; /* force the kernel context switch */
1280
1281		ce = intel_context_create(i915->kernel_context, engine);
1282		if (IS_ERR(ce)) {
1283			err = PTR_ERR(ce);
1284			goto out;
1285		}
1286
1287		rq = intel_context_create_request(ce);
1288		if (IS_ERR(rq)) {
1289			err = PTR_ERR(rq);
1290			intel_context_put(ce);
1291			goto out;
1292		}
1293
1294		err = intel_engine_emit_ctx_wa(rq);
1295		if (err)
1296			goto err_rq;
1297
1298		/*
1299		 * Failing to program the MOCS is non-fatal.The system will not
1300		 * run at peak performance. So warn the user and carry on.
1301		 */
1302		err = intel_mocs_emit(rq);
1303		if (err)
1304			dev_notice(i915->drm.dev,
1305				   "Failed to program MOCS registers; expect performance issues.\n");
1306
1307		err = intel_renderstate_emit(rq);
1308		if (err)
1309			goto err_rq;
1310
1311err_rq:
1312		requests[id] = i915_request_get(rq);
1313		i915_request_add(rq);
1314		if (err)
1315			goto out;
1316	}
1317
1318	/* Flush the default context image to memory, and enable powersaving. */
1319	if (!i915_gem_load_power_context(i915)) {
1320		err = -EIO;
1321		goto out;
1322	}
1323
1324	for (id = 0; id < ARRAY_SIZE(requests); id++) {
1325		struct i915_request *rq;
1326		struct i915_vma *state;
1327		void *vaddr;
1328
1329		rq = requests[id];
1330		if (!rq)
1331			continue;
1332
1333		/* We want to be able to unbind the state from the GGTT */
1334		GEM_BUG_ON(intel_context_is_pinned(rq->hw_context));
1335
1336		state = rq->hw_context->state;
1337		if (!state)
1338			continue;
1339
1340		/*
1341		 * As we will hold a reference to the logical state, it will
1342		 * not be torn down with the context, and importantly the
1343		 * object will hold onto its vma (making it possible for a
1344		 * stray GTT write to corrupt our defaults). Unmap the vma
1345		 * from the GTT to prevent such accidents and reclaim the
1346		 * space.
1347		 */
1348		err = i915_vma_unbind(state);
1349		if (err)
1350			goto out;
1351
1352		i915_gem_object_lock(state->obj);
1353		err = i915_gem_object_set_to_cpu_domain(state->obj, false);
1354		i915_gem_object_unlock(state->obj);
1355		if (err)
1356			goto out;
1357
1358		i915_gem_object_set_cache_coherency(state->obj, I915_CACHE_LLC);
1359
1360		/* Check we can acquire the image of the context state */
1361		vaddr = i915_gem_object_pin_map(state->obj, I915_MAP_FORCE_WB);
1362		if (IS_ERR(vaddr)) {
1363			err = PTR_ERR(vaddr);
1364			goto out;
1365		}
1366
1367		rq->engine->default_state = i915_gem_object_get(state->obj);
1368		i915_gem_object_unpin_map(state->obj);
1369	}
1370
1371out:
1372	/*
1373	 * If we have to abandon now, we expect the engines to be idle
1374	 * and ready to be torn-down. The quickest way we can accomplish
1375	 * this is by declaring ourselves wedged.
1376	 */
1377	if (err)
1378		intel_gt_set_wedged(&i915->gt);
1379
1380	for (id = 0; id < ARRAY_SIZE(requests); id++) {
1381		struct intel_context *ce;
1382		struct i915_request *rq;
1383
1384		rq = requests[id];
1385		if (!rq)
1386			continue;
1387
1388		ce = rq->hw_context;
1389		i915_request_put(rq);
1390		intel_context_put(ce);
1391	}
1392	return err;
1393}
1394
1395static int
1396i915_gem_init_scratch(struct drm_i915_private *i915, unsigned int size)
1397{
1398	return intel_gt_init_scratch(&i915->gt, size);
1399}
1400
1401static void i915_gem_fini_scratch(struct drm_i915_private *i915)
1402{
1403	intel_gt_fini_scratch(&i915->gt);
1404}
1405
1406static int intel_engines_verify_workarounds(struct drm_i915_private *i915)
 
 
 
 
 
 
 
 
 
1407{
1408	struct intel_engine_cs *engine;
1409	enum intel_engine_id id;
1410	int err = 0;
1411
1412	if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
1413		return 0;
1414
1415	for_each_engine(engine, i915, id) {
1416		if (intel_engine_verify_workarounds(engine, "load"))
1417			err = -EIO;
 
1418	}
1419
1420	return err;
1421}
1422
1423int i915_gem_init(struct drm_i915_private *dev_priv)
1424{
 
 
1425	int ret;
1426
1427	/* We need to fallback to 4K pages if host doesn't support huge gtt. */
1428	if (intel_vgpu_active(dev_priv) && !intel_vgpu_has_huge_gtt(dev_priv))
1429		mkwrite_device_info(dev_priv)->page_sizes =
1430			I915_GTT_PAGE_SIZE_4K;
1431
1432	intel_timelines_init(dev_priv);
1433
1434	ret = i915_gem_init_userptr(dev_priv);
1435	if (ret)
1436		return ret;
1437
1438	intel_uc_fetch_firmwares(&dev_priv->gt.uc);
1439	intel_wopcm_init(&dev_priv->wopcm);
1440
1441	/* This is just a security blanket to placate dragons.
1442	 * On some systems, we very sporadically observe that the first TLBs
1443	 * used by the CS may be stale, despite us poking the TLB reset. If
1444	 * we hold the forcewake during initialisation these problems
1445	 * just magically go away.
1446	 */
1447	mutex_lock(&dev_priv->drm.struct_mutex);
1448	intel_uncore_forcewake_get(&dev_priv->uncore, FORCEWAKE_ALL);
1449
1450	ret = i915_init_ggtt(dev_priv);
1451	if (ret) {
1452		GEM_BUG_ON(ret == -EIO);
1453		goto err_unlock;
1454	}
1455
1456	ret = i915_gem_init_scratch(dev_priv,
1457				    IS_GEN(dev_priv, 2) ? SZ_256K : PAGE_SIZE);
1458	if (ret) {
1459		GEM_BUG_ON(ret == -EIO);
1460		goto err_ggtt;
1461	}
1462
1463	ret = intel_engines_setup(dev_priv);
1464	if (ret) {
1465		GEM_BUG_ON(ret == -EIO);
1466		goto err_unlock;
1467	}
1468
1469	ret = i915_gem_contexts_init(dev_priv);
1470	if (ret) {
1471		GEM_BUG_ON(ret == -EIO);
1472		goto err_scratch;
1473	}
1474
1475	ret = intel_engines_init(dev_priv);
1476	if (ret) {
1477		GEM_BUG_ON(ret == -EIO);
1478		goto err_context;
1479	}
1480
1481	intel_init_gt_powersave(dev_priv);
1482
1483	intel_uc_init(&dev_priv->gt.uc);
1484
1485	ret = i915_gem_init_hw(dev_priv);
1486	if (ret)
1487		goto err_uc_init;
1488
1489	/* Only when the HW is re-initialised, can we replay the requests */
1490	ret = intel_gt_resume(&dev_priv->gt);
1491	if (ret)
1492		goto err_init_hw;
1493
1494	/*
1495	 * Despite its name intel_init_clock_gating applies both display
1496	 * clock gating workarounds; GT mmio workarounds and the occasional
1497	 * GT power context workaround. Worse, sometimes it includes a context
1498	 * register workaround which we need to apply before we record the
1499	 * default HW state for all contexts.
1500	 *
1501	 * FIXME: break up the workarounds and apply them at the right time!
1502	 */
1503	intel_init_clock_gating(dev_priv);
1504
1505	ret = intel_engines_verify_workarounds(dev_priv);
1506	if (ret)
1507		goto err_gt;
1508
1509	ret = __intel_engines_record_defaults(dev_priv);
1510	if (ret)
1511		goto err_gt;
1512
1513	ret = i915_inject_load_error(dev_priv, -ENODEV);
1514	if (ret)
1515		goto err_gt;
1516
1517	ret = i915_inject_load_error(dev_priv, -EIO);
1518	if (ret)
1519		goto err_gt;
1520
1521	intel_uncore_forcewake_put(&dev_priv->uncore, FORCEWAKE_ALL);
1522	mutex_unlock(&dev_priv->drm.struct_mutex);
1523
1524	return 0;
1525
1526	/*
1527	 * Unwinding is complicated by that we want to handle -EIO to mean
1528	 * disable GPU submission but keep KMS alive. We want to mark the
1529	 * HW as irrevisibly wedged, but keep enough state around that the
1530	 * driver doesn't explode during runtime.
1531	 */
1532err_gt:
1533	mutex_unlock(&dev_priv->drm.struct_mutex);
1534
1535	intel_gt_set_wedged(&dev_priv->gt);
1536	i915_gem_suspend(dev_priv);
1537	i915_gem_suspend_late(dev_priv);
1538
1539	i915_gem_drain_workqueue(dev_priv);
1540
1541	mutex_lock(&dev_priv->drm.struct_mutex);
1542err_init_hw:
1543	intel_uc_fini_hw(&dev_priv->gt.uc);
1544err_uc_init:
1545	if (ret != -EIO) {
1546		intel_uc_fini(&dev_priv->gt.uc);
1547		intel_cleanup_gt_powersave(dev_priv);
1548		intel_engines_cleanup(dev_priv);
1549	}
1550err_context:
1551	if (ret != -EIO)
1552		i915_gem_contexts_fini(dev_priv);
1553err_scratch:
1554	i915_gem_fini_scratch(dev_priv);
1555err_ggtt:
1556err_unlock:
1557	intel_uncore_forcewake_put(&dev_priv->uncore, FORCEWAKE_ALL);
1558	mutex_unlock(&dev_priv->drm.struct_mutex);
1559
1560	if (ret != -EIO) {
1561		intel_uc_cleanup_firmwares(&dev_priv->gt.uc);
1562		i915_gem_cleanup_userptr(dev_priv);
1563		intel_timelines_fini(dev_priv);
 
 
1564	}
1565
1566	if (ret == -EIO) {
1567		mutex_lock(&dev_priv->drm.struct_mutex);
1568
1569		/*
1570		 * Allow engines or uC initialisation to fail by marking the GPU
1571		 * as wedged. But we only want to do this when the GPU is angry,
1572		 * for all other failure, such as an allocation failure, bail.
1573		 */
1574		if (!intel_gt_is_wedged(&dev_priv->gt)) {
1575			i915_probe_error(dev_priv,
1576					 "Failed to initialize GPU, declaring it wedged!\n");
1577			intel_gt_set_wedged(&dev_priv->gt);
 
 
1578		}
1579
1580		/* Minimal basic recovery for KMS */
1581		ret = i915_ggtt_enable_hw(dev_priv);
1582		i915_gem_restore_gtt_mappings(dev_priv);
1583		i915_gem_restore_fences(dev_priv);
1584		intel_init_clock_gating(dev_priv);
1585
1586		mutex_unlock(&dev_priv->drm.struct_mutex);
1587	}
1588
1589	i915_gem_drain_freed_objects(dev_priv);
 
1590	return ret;
1591}
1592
1593void i915_gem_driver_register(struct drm_i915_private *i915)
1594{
1595	i915_gem_driver_register__shrinker(i915);
1596
1597	intel_engines_driver_register(i915);
1598}
1599
1600void i915_gem_driver_unregister(struct drm_i915_private *i915)
1601{
1602	i915_gem_driver_unregister__shrinker(i915);
1603}
1604
1605void i915_gem_driver_remove(struct drm_i915_private *dev_priv)
1606{
1607	GEM_BUG_ON(dev_priv->gt.awake);
1608
1609	intel_wakeref_auto_fini(&dev_priv->ggtt.userfault_wakeref);
1610
1611	i915_gem_suspend_late(dev_priv);
1612	intel_disable_gt_powersave(dev_priv);
 
 
1613
1614	/* Flush any outstanding unpin_work. */
1615	i915_gem_drain_workqueue(dev_priv);
1616
1617	mutex_lock(&dev_priv->drm.struct_mutex);
1618	intel_uc_fini_hw(&dev_priv->gt.uc);
1619	intel_uc_fini(&dev_priv->gt.uc);
1620	mutex_unlock(&dev_priv->drm.struct_mutex);
1621
1622	i915_gem_drain_freed_objects(dev_priv);
1623}
1624
1625void i915_gem_driver_release(struct drm_i915_private *dev_priv)
1626{
1627	mutex_lock(&dev_priv->drm.struct_mutex);
1628	intel_engines_cleanup(dev_priv);
1629	i915_gem_contexts_fini(dev_priv);
1630	i915_gem_fini_scratch(dev_priv);
1631	mutex_unlock(&dev_priv->drm.struct_mutex);
1632
1633	intel_wa_list_free(&dev_priv->gt_wa_list);
1634
1635	intel_cleanup_gt_powersave(dev_priv);
1636
1637	intel_uc_cleanup_firmwares(&dev_priv->gt.uc);
1638	i915_gem_cleanup_userptr(dev_priv);
1639	intel_timelines_fini(dev_priv);
1640
1641	i915_gem_drain_freed_objects(dev_priv);
 
 
 
1642
1643	WARN_ON(!list_empty(&dev_priv->contexts.list));
1644}
1645
1646void i915_gem_init_mmio(struct drm_i915_private *i915)
1647{
1648	i915_gem_sanitize(i915);
1649}
1650
1651static void i915_gem_init__mm(struct drm_i915_private *i915)
1652{
1653	spin_lock_init(&i915->mm.obj_lock);
1654
1655	init_llist_head(&i915->mm.free_list);
1656
1657	INIT_LIST_HEAD(&i915->mm.purge_list);
1658	INIT_LIST_HEAD(&i915->mm.shrink_list);
1659
1660	i915_gem_init__objects(i915);
1661}
1662
1663int i915_gem_init_early(struct drm_i915_private *dev_priv)
1664{
1665	int err;
1666
1667	i915_gem_init__mm(dev_priv);
1668	i915_gem_init__pm(dev_priv);
1669
1670	spin_lock_init(&dev_priv->fb_tracking.lock);
1671
1672	err = i915_gemfs_init(dev_priv);
1673	if (err)
1674		DRM_NOTE("Unable to create a private tmpfs mount, hugepage support will be disabled(%d).\n", err);
1675
1676	return 0;
1677}
1678
1679void i915_gem_cleanup_early(struct drm_i915_private *dev_priv)
1680{
1681	i915_gem_drain_freed_objects(dev_priv);
1682	GEM_BUG_ON(!llist_empty(&dev_priv->mm.free_list));
1683	GEM_BUG_ON(atomic_read(&dev_priv->mm.free_count));
1684	WARN_ON(dev_priv->mm.shrink_count);
1685
1686	i915_gemfs_fini(dev_priv);
1687}
1688
1689int i915_gem_freeze(struct drm_i915_private *dev_priv)
1690{
1691	/* Discard all purgeable objects, let userspace recover those as
1692	 * required after resuming.
1693	 */
1694	i915_gem_shrink_all(dev_priv);
1695
1696	return 0;
1697}
1698
1699int i915_gem_freeze_late(struct drm_i915_private *i915)
1700{
1701	struct drm_i915_gem_object *obj;
1702	intel_wakeref_t wakeref;
1703
1704	/*
1705	 * Called just before we write the hibernation image.
1706	 *
1707	 * We need to update the domain tracking to reflect that the CPU
1708	 * will be accessing all the pages to create and restore from the
1709	 * hibernation, and so upon restoration those pages will be in the
1710	 * CPU domain.
1711	 *
1712	 * To make sure the hibernation image contains the latest state,
1713	 * we update that state just before writing out the image.
1714	 *
1715	 * To try and reduce the hibernation image, we manually shrink
1716	 * the objects as well, see i915_gem_freeze()
1717	 */
1718
1719	wakeref = intel_runtime_pm_get(&i915->runtime_pm);
1720
1721	i915_gem_shrink(i915, -1UL, NULL, ~0);
1722	i915_gem_drain_freed_objects(i915);
1723
1724	list_for_each_entry(obj, &i915->mm.shrink_list, mm.link) {
1725		i915_gem_object_lock(obj);
1726		WARN_ON(i915_gem_object_set_to_cpu_domain(obj, true));
1727		i915_gem_object_unlock(obj);
1728	}
1729
1730	intel_runtime_pm_put(&i915->runtime_pm, wakeref);
1731
1732	return 0;
1733}
1734
1735void i915_gem_release(struct drm_device *dev, struct drm_file *file)
1736{
1737	struct drm_i915_file_private *file_priv = file->driver_priv;
1738	struct i915_request *request;
1739
1740	/* Clean up our request list when the client is going away, so that
1741	 * later retire_requests won't dereference our soon-to-be-gone
1742	 * file_priv.
1743	 */
1744	spin_lock(&file_priv->mm.lock);
1745	list_for_each_entry(request, &file_priv->mm.request_list, client_link)
1746		request->file_priv = NULL;
1747	spin_unlock(&file_priv->mm.lock);
1748}
1749
1750int i915_gem_open(struct drm_i915_private *i915, struct drm_file *file)
1751{
1752	struct drm_i915_file_private *file_priv;
1753	int ret;
 
1754
1755	DRM_DEBUG("\n");
1756
1757	file_priv = kzalloc(sizeof(*file_priv), GFP_KERNEL);
1758	if (!file_priv)
1759		return -ENOMEM;
 
 
 
 
 
 
1760
1761	file->driver_priv = file_priv;
1762	file_priv->dev_priv = i915;
1763	file_priv->file = file;
1764
1765	spin_lock_init(&file_priv->mm.lock);
1766	INIT_LIST_HEAD(&file_priv->mm.request_list);
1767
1768	file_priv->bsd_engine = -1;
1769	file_priv->hang_timestamp = jiffies;
1770
1771	ret = i915_gem_context_open(i915, file);
1772	if (ret)
1773		kfree(file_priv);
1774
 
 
 
 
 
 
 
1775	return ret;
1776}
1777
1778#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1779#include "selftests/mock_gem_device.c"
1780#include "selftests/i915_gem.c"
1781#endif