Linux Audio

Check our new training course

Loading...
v5.9
   1/*
   2 * SPDX-License-Identifier: MIT
   3 *
   4 * Copyright © 2008,2010 Intel Corporation
   5 */
   6
   7#include <linux/intel-iommu.h>
   8#include <linux/dma-resv.h>
   9#include <linux/sync_file.h>
  10#include <linux/uaccess.h>
  11
  12#include <drm/drm_syncobj.h>
  13
  14#include "display/intel_frontbuffer.h"
  15
  16#include "gem/i915_gem_ioctls.h"
  17#include "gt/intel_context.h"
 
  18#include "gt/intel_gt.h"
  19#include "gt/intel_gt_buffer_pool.h"
  20#include "gt/intel_gt_pm.h"
  21#include "gt/intel_ring.h"
  22
  23#include "i915_drv.h"
  24#include "i915_gem_clflush.h"
  25#include "i915_gem_context.h"
  26#include "i915_gem_ioctls.h"
  27#include "i915_sw_fence_work.h"
  28#include "i915_trace.h"
 
  29
  30struct eb_vma {
  31	struct i915_vma *vma;
  32	unsigned int flags;
  33
  34	/** This vma's place in the execbuf reservation list */
  35	struct drm_i915_gem_exec_object2 *exec;
  36	struct list_head bind_link;
  37	struct list_head reloc_link;
  38
  39	struct hlist_node node;
  40	u32 handle;
  41};
  42
  43struct eb_vma_array {
  44	struct kref kref;
  45	struct eb_vma vma[];
  46};
  47
  48enum {
  49	FORCE_CPU_RELOC = 1,
  50	FORCE_GTT_RELOC,
  51	FORCE_GPU_RELOC,
  52#define DBG_FORCE_RELOC 0 /* choose one of the above! */
  53};
  54
  55#define __EXEC_OBJECT_HAS_PIN		BIT(31)
  56#define __EXEC_OBJECT_HAS_FENCE		BIT(30)
  57#define __EXEC_OBJECT_NEEDS_MAP		BIT(29)
  58#define __EXEC_OBJECT_NEEDS_BIAS	BIT(28)
  59#define __EXEC_OBJECT_INTERNAL_FLAGS	(~0u << 28) /* all of the above */
 
 
 
  60
  61#define __EXEC_HAS_RELOC	BIT(31)
  62#define __EXEC_INTERNAL_FLAGS	(~0u << 31)
 
 
  63#define UPDATE			PIN_OFFSET_FIXED
  64
  65#define BATCH_OFFSET_BIAS (256*1024)
  66
  67#define __I915_EXEC_ILLEGAL_FLAGS \
  68	(__I915_EXEC_UNKNOWN_FLAGS | \
  69	 I915_EXEC_CONSTANTS_MASK  | \
  70	 I915_EXEC_RESOURCE_STREAMER)
  71
  72/* Catch emission of unexpected errors for CI! */
  73#if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
  74#undef EINVAL
  75#define EINVAL ({ \
  76	DRM_DEBUG_DRIVER("EINVAL at %s:%d\n", __func__, __LINE__); \
  77	22; \
  78})
  79#endif
  80
  81/**
  82 * DOC: User command execution
  83 *
  84 * Userspace submits commands to be executed on the GPU as an instruction
  85 * stream within a GEM object we call a batchbuffer. This instructions may
  86 * refer to other GEM objects containing auxiliary state such as kernels,
  87 * samplers, render targets and even secondary batchbuffers. Userspace does
  88 * not know where in the GPU memory these objects reside and so before the
  89 * batchbuffer is passed to the GPU for execution, those addresses in the
  90 * batchbuffer and auxiliary objects are updated. This is known as relocation,
  91 * or patching. To try and avoid having to relocate each object on the next
  92 * execution, userspace is told the location of those objects in this pass,
  93 * but this remains just a hint as the kernel may choose a new location for
  94 * any object in the future.
  95 *
  96 * At the level of talking to the hardware, submitting a batchbuffer for the
  97 * GPU to execute is to add content to a buffer from which the HW
  98 * command streamer is reading.
  99 *
 100 * 1. Add a command to load the HW context. For Logical Ring Contexts, i.e.
 101 *    Execlists, this command is not placed on the same buffer as the
 102 *    remaining items.
 103 *
 104 * 2. Add a command to invalidate caches to the buffer.
 105 *
 106 * 3. Add a batchbuffer start command to the buffer; the start command is
 107 *    essentially a token together with the GPU address of the batchbuffer
 108 *    to be executed.
 109 *
 110 * 4. Add a pipeline flush to the buffer.
 111 *
 112 * 5. Add a memory write command to the buffer to record when the GPU
 113 *    is done executing the batchbuffer. The memory write writes the
 114 *    global sequence number of the request, ``i915_request::global_seqno``;
 115 *    the i915 driver uses the current value in the register to determine
 116 *    if the GPU has completed the batchbuffer.
 117 *
 118 * 6. Add a user interrupt command to the buffer. This command instructs
 119 *    the GPU to issue an interrupt when the command, pipeline flush and
 120 *    memory write are completed.
 121 *
 122 * 7. Inform the hardware of the additional commands added to the buffer
 123 *    (by updating the tail pointer).
 124 *
 125 * Processing an execbuf ioctl is conceptually split up into a few phases.
 126 *
 127 * 1. Validation - Ensure all the pointers, handles and flags are valid.
 128 * 2. Reservation - Assign GPU address space for every object
 129 * 3. Relocation - Update any addresses to point to the final locations
 130 * 4. Serialisation - Order the request with respect to its dependencies
 131 * 5. Construction - Construct a request to execute the batchbuffer
 132 * 6. Submission (at some point in the future execution)
 133 *
 134 * Reserving resources for the execbuf is the most complicated phase. We
 135 * neither want to have to migrate the object in the address space, nor do
 136 * we want to have to update any relocations pointing to this object. Ideally,
 137 * we want to leave the object where it is and for all the existing relocations
 138 * to match. If the object is given a new address, or if userspace thinks the
 139 * object is elsewhere, we have to parse all the relocation entries and update
 140 * the addresses. Userspace can set the I915_EXEC_NORELOC flag to hint that
 141 * all the target addresses in all of its objects match the value in the
 142 * relocation entries and that they all match the presumed offsets given by the
 143 * list of execbuffer objects. Using this knowledge, we know that if we haven't
 144 * moved any buffers, all the relocation entries are valid and we can skip
 145 * the update. (If userspace is wrong, the likely outcome is an impromptu GPU
 146 * hang.) The requirement for using I915_EXEC_NO_RELOC are:
 147 *
 148 *      The addresses written in the objects must match the corresponding
 149 *      reloc.presumed_offset which in turn must match the corresponding
 150 *      execobject.offset.
 151 *
 152 *      Any render targets written to in the batch must be flagged with
 153 *      EXEC_OBJECT_WRITE.
 154 *
 155 *      To avoid stalling, execobject.offset should match the current
 156 *      address of that object within the active context.
 157 *
 158 * The reservation is done is multiple phases. First we try and keep any
 159 * object already bound in its current location - so as long as meets the
 160 * constraints imposed by the new execbuffer. Any object left unbound after the
 161 * first pass is then fitted into any available idle space. If an object does
 162 * not fit, all objects are removed from the reservation and the process rerun
 163 * after sorting the objects into a priority order (more difficult to fit
 164 * objects are tried first). Failing that, the entire VM is cleared and we try
 165 * to fit the execbuf once last time before concluding that it simply will not
 166 * fit.
 167 *
 168 * A small complication to all of this is that we allow userspace not only to
 169 * specify an alignment and a size for the object in the address space, but
 170 * we also allow userspace to specify the exact offset. This objects are
 171 * simpler to place (the location is known a priori) all we have to do is make
 172 * sure the space is available.
 173 *
 174 * Once all the objects are in place, patching up the buried pointers to point
 175 * to the final locations is a fairly simple job of walking over the relocation
 176 * entry arrays, looking up the right address and rewriting the value into
 177 * the object. Simple! ... The relocation entries are stored in user memory
 178 * and so to access them we have to copy them into a local buffer. That copy
 179 * has to avoid taking any pagefaults as they may lead back to a GEM object
 180 * requiring the struct_mutex (i.e. recursive deadlock). So once again we split
 181 * the relocation into multiple passes. First we try to do everything within an
 182 * atomic context (avoid the pagefaults) which requires that we never wait. If
 183 * we detect that we may wait, or if we need to fault, then we have to fallback
 184 * to a slower path. The slowpath has to drop the mutex. (Can you hear alarm
 185 * bells yet?) Dropping the mutex means that we lose all the state we have
 186 * built up so far for the execbuf and we must reset any global data. However,
 187 * we do leave the objects pinned in their final locations - which is a
 188 * potential issue for concurrent execbufs. Once we have left the mutex, we can
 189 * allocate and copy all the relocation entries into a large array at our
 190 * leisure, reacquire the mutex, reclaim all the objects and other state and
 191 * then proceed to update any incorrect addresses with the objects.
 192 *
 193 * As we process the relocation entries, we maintain a record of whether the
 194 * object is being written to. Using NORELOC, we expect userspace to provide
 195 * this information instead. We also check whether we can skip the relocation
 196 * by comparing the expected value inside the relocation entry with the target's
 197 * final address. If they differ, we have to map the current object and rewrite
 198 * the 4 or 8 byte pointer within.
 199 *
 200 * Serialising an execbuf is quite simple according to the rules of the GEM
 201 * ABI. Execution within each context is ordered by the order of submission.
 202 * Writes to any GEM object are in order of submission and are exclusive. Reads
 203 * from a GEM object are unordered with respect to other reads, but ordered by
 204 * writes. A write submitted after a read cannot occur before the read, and
 205 * similarly any read submitted after a write cannot occur before the write.
 206 * Writes are ordered between engines such that only one write occurs at any
 207 * time (completing any reads beforehand) - using semaphores where available
 208 * and CPU serialisation otherwise. Other GEM access obey the same rules, any
 209 * write (either via mmaps using set-domain, or via pwrite) must flush all GPU
 210 * reads before starting, and any read (either using set-domain or pread) must
 211 * flush all GPU writes before starting. (Note we only employ a barrier before,
 212 * we currently rely on userspace not concurrently starting a new execution
 213 * whilst reading or writing to an object. This may be an advantage or not
 214 * depending on how much you trust userspace not to shoot themselves in the
 215 * foot.) Serialisation may just result in the request being inserted into
 216 * a DAG awaiting its turn, but most simple is to wait on the CPU until
 217 * all dependencies are resolved.
 218 *
 219 * After all of that, is just a matter of closing the request and handing it to
 220 * the hardware (well, leaving it in a queue to be executed). However, we also
 221 * offer the ability for batchbuffers to be run with elevated privileges so
 222 * that they access otherwise hidden registers. (Used to adjust L3 cache etc.)
 223 * Before any batch is given extra privileges we first must check that it
 224 * contains no nefarious instructions, we check that each instruction is from
 225 * our whitelist and all registers are also from an allowed list. We first
 226 * copy the user's batchbuffer to a shadow (so that the user doesn't have
 227 * access to it, either by the CPU or GPU as we scan it) and then parse each
 228 * instruction. If everything is ok, we set a flag telling the hardware to run
 229 * the batchbuffer in trusted mode, otherwise the ioctl is rejected.
 230 */
 231
 
 
 
 
 
 
 
 232struct i915_execbuffer {
 233	struct drm_i915_private *i915; /** i915 backpointer */
 234	struct drm_file *file; /** per-file lookup tables and limits */
 235	struct drm_i915_gem_execbuffer2 *args; /** ioctl parameters */
 236	struct drm_i915_gem_exec_object2 *exec; /** ioctl execobj[] */
 237	struct eb_vma *vma;
 238
 239	struct intel_engine_cs *engine; /** engine to queue the request to */
 240	struct intel_context *context; /* logical state for the request */
 241	struct i915_gem_context *gem_context; /** caller's context */
 242
 243	struct i915_request *request; /** our request to build */
 244	struct eb_vma *batch; /** identity of the batch obj/vma */
 245	struct i915_vma *trampoline; /** trampoline used for chaining */
 246
 247	/** actual size of execobj[] as we may extend it for the cmdparser */
 248	unsigned int buffer_count;
 249
 250	/** list of vma not yet bound during reservation phase */
 251	struct list_head unbound;
 252
 253	/** list of vma that have execobj.relocation_count */
 254	struct list_head relocs;
 255
 
 
 256	/**
 257	 * Track the most recently used object for relocations, as we
 258	 * frequently have to perform multiple relocations within the same
 259	 * obj/page
 260	 */
 261	struct reloc_cache {
 262		struct drm_mm_node node; /** temporary GTT binding */
 263		unsigned long vaddr; /** Current kmap address */
 264		unsigned long page; /** Currently mapped page index */
 265		unsigned int gen; /** Cached value of INTEL_GEN */
 266		bool use_64bit_reloc : 1;
 267		bool has_llc : 1;
 268		bool has_fence : 1;
 269		bool needs_unfenced : 1;
 270
 271		struct i915_vma *target;
 272		struct i915_request *rq;
 273		struct i915_vma *rq_vma;
 274		u32 *rq_cmd;
 275		unsigned int rq_size;
 
 276	} reloc_cache;
 277
 
 
 
 278	u64 invalid_flags; /** Set of execobj.flags that are invalid */
 279	u32 context_flags; /** Set of execobj.flags to insert from the ctx */
 280
 
 281	u32 batch_start_offset; /** Location within object of batch */
 282	u32 batch_len; /** Length of batch within object */
 283	u32 batch_flags; /** Flags composed for emit_bb_start() */
 
 284
 285	/**
 286	 * Indicate either the size of the hastable used to resolve
 287	 * relocation handles, or if negative that we are using a direct
 288	 * index into the execobj[].
 289	 */
 290	int lut_size;
 291	struct hlist_head *buckets; /** ht for relocation handles */
 292	struct eb_vma_array *array;
 
 
 293};
 294
 
 
 
 
 
 295static inline bool eb_use_cmdparser(const struct i915_execbuffer *eb)
 296{
 297	return intel_engine_requires_cmd_parser(eb->engine) ||
 298		(intel_engine_using_cmd_parser(eb->engine) &&
 299		 eb->args->batch_len);
 300}
 301
 302static struct eb_vma_array *eb_vma_array_create(unsigned int count)
 303{
 304	struct eb_vma_array *arr;
 305
 306	arr = kvmalloc(struct_size(arr, vma, count), GFP_KERNEL | __GFP_NOWARN);
 307	if (!arr)
 308		return NULL;
 309
 310	kref_init(&arr->kref);
 311	arr->vma[0].vma = NULL;
 312
 313	return arr;
 314}
 315
 316static inline void eb_unreserve_vma(struct eb_vma *ev)
 317{
 318	struct i915_vma *vma = ev->vma;
 319
 320	if (unlikely(ev->flags & __EXEC_OBJECT_HAS_FENCE))
 321		__i915_vma_unpin_fence(vma);
 322
 323	if (ev->flags & __EXEC_OBJECT_HAS_PIN)
 324		__i915_vma_unpin(vma);
 325
 326	ev->flags &= ~(__EXEC_OBJECT_HAS_PIN |
 327		       __EXEC_OBJECT_HAS_FENCE);
 328}
 329
 330static void eb_vma_array_destroy(struct kref *kref)
 331{
 332	struct eb_vma_array *arr = container_of(kref, typeof(*arr), kref);
 333	struct eb_vma *ev = arr->vma;
 334
 335	while (ev->vma) {
 336		eb_unreserve_vma(ev);
 337		i915_vma_put(ev->vma);
 338		ev++;
 339	}
 340
 341	kvfree(arr);
 342}
 343
 344static void eb_vma_array_put(struct eb_vma_array *arr)
 345{
 346	kref_put(&arr->kref, eb_vma_array_destroy);
 347}
 348
 349static int eb_create(struct i915_execbuffer *eb)
 350{
 351	/* Allocate an extra slot for use by the command parser + sentinel */
 352	eb->array = eb_vma_array_create(eb->buffer_count + 2);
 353	if (!eb->array)
 354		return -ENOMEM;
 355
 356	eb->vma = eb->array->vma;
 357
 358	if (!(eb->args->flags & I915_EXEC_HANDLE_LUT)) {
 359		unsigned int size = 1 + ilog2(eb->buffer_count);
 360
 361		/*
 362		 * Without a 1:1 association between relocation handles and
 363		 * the execobject[] index, we instead create a hashtable.
 364		 * We size it dynamically based on available memory, starting
 365		 * first with 1:1 assocative hash and scaling back until
 366		 * the allocation succeeds.
 367		 *
 368		 * Later on we use a positive lut_size to indicate we are
 369		 * using this hashtable, and a negative value to indicate a
 370		 * direct lookup.
 371		 */
 372		do {
 373			gfp_t flags;
 374
 375			/* While we can still reduce the allocation size, don't
 376			 * raise a warning and allow the allocation to fail.
 377			 * On the last pass though, we want to try as hard
 378			 * as possible to perform the allocation and warn
 379			 * if it fails.
 380			 */
 381			flags = GFP_KERNEL;
 382			if (size > 1)
 383				flags |= __GFP_NORETRY | __GFP_NOWARN;
 384
 385			eb->buckets = kzalloc(sizeof(struct hlist_head) << size,
 386					      flags);
 387			if (eb->buckets)
 388				break;
 389		} while (--size);
 390
 391		if (unlikely(!size)) {
 392			eb_vma_array_put(eb->array);
 393			return -ENOMEM;
 394		}
 395
 396		eb->lut_size = size;
 397	} else {
 398		eb->lut_size = -eb->buffer_count;
 399	}
 400
 401	return 0;
 402}
 403
 404static bool
 405eb_vma_misplaced(const struct drm_i915_gem_exec_object2 *entry,
 406		 const struct i915_vma *vma,
 407		 unsigned int flags)
 408{
 409	if (vma->node.size < entry->pad_to_size)
 410		return true;
 411
 412	if (entry->alignment && !IS_ALIGNED(vma->node.start, entry->alignment))
 413		return true;
 414
 415	if (flags & EXEC_OBJECT_PINNED &&
 416	    vma->node.start != entry->offset)
 417		return true;
 418
 419	if (flags & __EXEC_OBJECT_NEEDS_BIAS &&
 420	    vma->node.start < BATCH_OFFSET_BIAS)
 421		return true;
 422
 423	if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) &&
 424	    (vma->node.start + vma->node.size - 1) >> 32)
 425		return true;
 426
 427	if (flags & __EXEC_OBJECT_NEEDS_MAP &&
 428	    !i915_vma_is_map_and_fenceable(vma))
 429		return true;
 430
 431	return false;
 432}
 433
 434static u64 eb_pin_flags(const struct drm_i915_gem_exec_object2 *entry,
 435			unsigned int exec_flags)
 436{
 437	u64 pin_flags = 0;
 438
 439	if (exec_flags & EXEC_OBJECT_NEEDS_GTT)
 440		pin_flags |= PIN_GLOBAL;
 441
 442	/*
 443	 * Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset,
 444	 * limit address to the first 4GBs for unflagged objects.
 445	 */
 446	if (!(exec_flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS))
 447		pin_flags |= PIN_ZONE_4G;
 448
 449	if (exec_flags & __EXEC_OBJECT_NEEDS_MAP)
 450		pin_flags |= PIN_MAPPABLE;
 451
 452	if (exec_flags & EXEC_OBJECT_PINNED)
 453		pin_flags |= entry->offset | PIN_OFFSET_FIXED;
 454	else if (exec_flags & __EXEC_OBJECT_NEEDS_BIAS)
 455		pin_flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
 456
 457	return pin_flags;
 458}
 459
 460static inline bool
 461eb_pin_vma(struct i915_execbuffer *eb,
 462	   const struct drm_i915_gem_exec_object2 *entry,
 463	   struct eb_vma *ev)
 464{
 465	struct i915_vma *vma = ev->vma;
 466	u64 pin_flags;
 
 467
 468	if (vma->node.size)
 469		pin_flags = vma->node.start;
 470	else
 471		pin_flags = entry->offset & PIN_OFFSET_MASK;
 472
 473	pin_flags |= PIN_USER | PIN_NOEVICT | PIN_OFFSET_FIXED;
 474	if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_GTT))
 475		pin_flags |= PIN_GLOBAL;
 476
 477	/* Attempt to reuse the current location if available */
 478	if (unlikely(i915_vma_pin(vma, 0, 0, pin_flags))) {
 
 
 
 
 479		if (entry->flags & EXEC_OBJECT_PINNED)
 480			return false;
 481
 482		/* Failing that pick any _free_ space if suitable */
 483		if (unlikely(i915_vma_pin(vma,
 484					  entry->pad_to_size,
 485					  entry->alignment,
 486					  eb_pin_flags(entry, ev->flags) |
 487					  PIN_USER | PIN_NOEVICT)))
 488			return false;
 
 489	}
 490
 491	if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) {
 492		if (unlikely(i915_vma_pin_fence(vma))) {
 
 493			i915_vma_unpin(vma);
 494			return false;
 495		}
 496
 497		if (vma->fence)
 498			ev->flags |= __EXEC_OBJECT_HAS_FENCE;
 499	}
 500
 501	ev->flags |= __EXEC_OBJECT_HAS_PIN;
 502	return !eb_vma_misplaced(entry, vma, ev->flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 503}
 504
 505static int
 506eb_validate_vma(struct i915_execbuffer *eb,
 507		struct drm_i915_gem_exec_object2 *entry,
 508		struct i915_vma *vma)
 509{
 
 
 
 
 
 
 
 510	if (unlikely(entry->flags & eb->invalid_flags))
 511		return -EINVAL;
 512
 513	if (unlikely(entry->alignment &&
 514		     !is_power_of_2_u64(entry->alignment)))
 515		return -EINVAL;
 516
 517	/*
 518	 * Offset can be used as input (EXEC_OBJECT_PINNED), reject
 519	 * any non-page-aligned or non-canonical addresses.
 520	 */
 521	if (unlikely(entry->flags & EXEC_OBJECT_PINNED &&
 522		     entry->offset != gen8_canonical_addr(entry->offset & I915_GTT_PAGE_MASK)))
 523		return -EINVAL;
 524
 525	/* pad_to_size was once a reserved field, so sanitize it */
 526	if (entry->flags & EXEC_OBJECT_PAD_TO_SIZE) {
 527		if (unlikely(offset_in_page(entry->pad_to_size)))
 528			return -EINVAL;
 529	} else {
 530		entry->pad_to_size = 0;
 531	}
 532	/*
 533	 * From drm_mm perspective address space is continuous,
 534	 * so from this point we're always using non-canonical
 535	 * form internally.
 536	 */
 537	entry->offset = gen8_noncanonical_addr(entry->offset);
 538
 539	if (!eb->reloc_cache.has_fence) {
 540		entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
 541	} else {
 542		if ((entry->flags & EXEC_OBJECT_NEEDS_FENCE ||
 543		     eb->reloc_cache.needs_unfenced) &&
 544		    i915_gem_object_is_tiled(vma->obj))
 545			entry->flags |= EXEC_OBJECT_NEEDS_GTT | __EXEC_OBJECT_NEEDS_MAP;
 546	}
 547
 548	if (!(entry->flags & EXEC_OBJECT_PINNED))
 549		entry->flags |= eb->context_flags;
 550
 551	return 0;
 552}
 553
 554static void
 555eb_add_vma(struct i915_execbuffer *eb,
 556	   unsigned int i, unsigned batch_idx,
 557	   struct i915_vma *vma)
 558{
 559	struct drm_i915_gem_exec_object2 *entry = &eb->exec[i];
 560	struct eb_vma *ev = &eb->vma[i];
 561
 562	GEM_BUG_ON(i915_vma_is_closed(vma));
 563
 564	ev->vma = vma;
 565	ev->exec = entry;
 566	ev->flags = entry->flags;
 567
 568	if (eb->lut_size > 0) {
 569		ev->handle = entry->handle;
 570		hlist_add_head(&ev->node,
 571			       &eb->buckets[hash_32(entry->handle,
 572						    eb->lut_size)]);
 573	}
 574
 575	if (entry->relocation_count)
 576		list_add_tail(&ev->reloc_link, &eb->relocs);
 577
 578	/*
 579	 * SNA is doing fancy tricks with compressing batch buffers, which leads
 580	 * to negative relocation deltas. Usually that works out ok since the
 581	 * relocate address is still positive, except when the batch is placed
 582	 * very low in the GTT. Ensure this doesn't happen.
 583	 *
 584	 * Note that actual hangs have only been observed on gen7, but for
 585	 * paranoia do it everywhere.
 586	 */
 587	if (i == batch_idx) {
 588		if (entry->relocation_count &&
 589		    !(ev->flags & EXEC_OBJECT_PINNED))
 590			ev->flags |= __EXEC_OBJECT_NEEDS_BIAS;
 591		if (eb->reloc_cache.has_fence)
 592			ev->flags |= EXEC_OBJECT_NEEDS_FENCE;
 593
 594		eb->batch = ev;
 595	}
 596
 597	if (eb_pin_vma(eb, entry, ev)) {
 598		if (entry->offset != vma->node.start) {
 599			entry->offset = vma->node.start | UPDATE;
 600			eb->args->flags |= __EXEC_HAS_RELOC;
 601		}
 602	} else {
 603		eb_unreserve_vma(ev);
 604		list_add_tail(&ev->bind_link, &eb->unbound);
 605	}
 606}
 607
 608static inline int use_cpu_reloc(const struct reloc_cache *cache,
 609				const struct drm_i915_gem_object *obj)
 610{
 611	if (!i915_gem_object_has_struct_page(obj))
 612		return false;
 613
 614	if (DBG_FORCE_RELOC == FORCE_CPU_RELOC)
 615		return true;
 616
 617	if (DBG_FORCE_RELOC == FORCE_GTT_RELOC)
 618		return false;
 619
 620	return (cache->has_llc ||
 621		obj->cache_dirty ||
 622		obj->cache_level != I915_CACHE_NONE);
 623}
 624
 625static int eb_reserve_vma(const struct i915_execbuffer *eb,
 626			  struct eb_vma *ev,
 627			  u64 pin_flags)
 628{
 629	struct drm_i915_gem_exec_object2 *entry = ev->exec;
 630	struct i915_vma *vma = ev->vma;
 631	int err;
 632
 633	if (drm_mm_node_allocated(&vma->node) &&
 634	    eb_vma_misplaced(entry, vma, ev->flags)) {
 635		err = i915_vma_unbind(vma);
 636		if (err)
 637			return err;
 638	}
 639
 640	err = i915_vma_pin(vma,
 641			   entry->pad_to_size, entry->alignment,
 642			   eb_pin_flags(entry, ev->flags) | pin_flags);
 643	if (err)
 644		return err;
 645
 646	if (entry->offset != vma->node.start) {
 647		entry->offset = vma->node.start | UPDATE;
 648		eb->args->flags |= __EXEC_HAS_RELOC;
 649	}
 650
 651	if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) {
 652		err = i915_vma_pin_fence(vma);
 653		if (unlikely(err)) {
 654			i915_vma_unpin(vma);
 655			return err;
 656		}
 657
 658		if (vma->fence)
 659			ev->flags |= __EXEC_OBJECT_HAS_FENCE;
 660	}
 661
 662	ev->flags |= __EXEC_OBJECT_HAS_PIN;
 663	GEM_BUG_ON(eb_vma_misplaced(entry, vma, ev->flags));
 664
 665	return 0;
 666}
 667
 668static int eb_reserve(struct i915_execbuffer *eb)
 669{
 670	const unsigned int count = eb->buffer_count;
 671	unsigned int pin_flags = PIN_USER | PIN_NONBLOCK;
 672	struct list_head last;
 673	struct eb_vma *ev;
 674	unsigned int i, pass;
 675	int err = 0;
 676
 677	/*
 678	 * Attempt to pin all of the buffers into the GTT.
 679	 * This is done in 3 phases:
 680	 *
 681	 * 1a. Unbind all objects that do not match the GTT constraints for
 682	 *     the execbuffer (fenceable, mappable, alignment etc).
 683	 * 1b. Increment pin count for already bound objects.
 684	 * 2.  Bind new objects.
 685	 * 3.  Decrement pin count.
 686	 *
 687	 * This avoid unnecessary unbinding of later objects in order to make
 688	 * room for the earlier objects *unless* we need to defragment.
 689	 */
 690
 691	if (mutex_lock_interruptible(&eb->i915->drm.struct_mutex))
 692		return -EINTR;
 693
 694	pass = 0;
 695	do {
 696		list_for_each_entry(ev, &eb->unbound, bind_link) {
 697			err = eb_reserve_vma(eb, ev, pin_flags);
 698			if (err)
 699				break;
 700		}
 701		if (!(err == -ENOSPC || err == -EAGAIN))
 702			break;
 703
 704		/* Resort *all* the objects into priority order */
 705		INIT_LIST_HEAD(&eb->unbound);
 706		INIT_LIST_HEAD(&last);
 707		for (i = 0; i < count; i++) {
 708			unsigned int flags;
 709
 710			ev = &eb->vma[i];
 711			flags = ev->flags;
 712			if (flags & EXEC_OBJECT_PINNED &&
 713			    flags & __EXEC_OBJECT_HAS_PIN)
 714				continue;
 715
 716			eb_unreserve_vma(ev);
 717
 718			if (flags & EXEC_OBJECT_PINNED)
 719				/* Pinned must have their slot */
 720				list_add(&ev->bind_link, &eb->unbound);
 721			else if (flags & __EXEC_OBJECT_NEEDS_MAP)
 722				/* Map require the lowest 256MiB (aperture) */
 723				list_add_tail(&ev->bind_link, &eb->unbound);
 724			else if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS))
 725				/* Prioritise 4GiB region for restricted bo */
 726				list_add(&ev->bind_link, &last);
 727			else
 728				list_add_tail(&ev->bind_link, &last);
 729		}
 730		list_splice_tail(&last, &eb->unbound);
 731
 732		if (err == -EAGAIN) {
 733			mutex_unlock(&eb->i915->drm.struct_mutex);
 734			flush_workqueue(eb->i915->mm.userptr_wq);
 735			mutex_lock(&eb->i915->drm.struct_mutex);
 736			continue;
 737		}
 738
 739		switch (pass++) {
 740		case 0:
 741			break;
 742
 743		case 1:
 744			/* Too fragmented, unbind everything and retry */
 745			mutex_lock(&eb->context->vm->mutex);
 746			err = i915_gem_evict_vm(eb->context->vm);
 747			mutex_unlock(&eb->context->vm->mutex);
 748			if (err)
 749				goto unlock;
 750			break;
 751
 752		default:
 753			err = -ENOSPC;
 754			goto unlock;
 755		}
 756
 757		pin_flags = PIN_USER;
 758	} while (1);
 759
 760unlock:
 761	mutex_unlock(&eb->i915->drm.struct_mutex);
 762	return err;
 763}
 764
 765static unsigned int eb_batch_index(const struct i915_execbuffer *eb)
 766{
 767	if (eb->args->flags & I915_EXEC_BATCH_FIRST)
 768		return 0;
 769	else
 770		return eb->buffer_count - 1;
 771}
 772
 773static int eb_select_context(struct i915_execbuffer *eb)
 774{
 775	struct i915_gem_context *ctx;
 776
 777	ctx = i915_gem_context_lookup(eb->file->driver_priv, eb->args->rsvd1);
 778	if (unlikely(!ctx))
 779		return -ENOENT;
 780
 781	eb->gem_context = ctx;
 782	if (rcu_access_pointer(ctx->vm))
 783		eb->invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
 784
 785	eb->context_flags = 0;
 786	if (test_bit(UCONTEXT_NO_ZEROMAP, &ctx->user_flags))
 787		eb->context_flags |= __EXEC_OBJECT_NEEDS_BIAS;
 788
 789	return 0;
 790}
 791
 792static int __eb_add_lut(struct i915_execbuffer *eb,
 793			u32 handle, struct i915_vma *vma)
 794{
 795	struct i915_gem_context *ctx = eb->gem_context;
 796	struct i915_lut_handle *lut;
 797	int err;
 798
 799	lut = i915_lut_handle_alloc();
 800	if (unlikely(!lut))
 801		return -ENOMEM;
 802
 803	i915_vma_get(vma);
 804	if (!atomic_fetch_inc(&vma->open_count))
 805		i915_vma_reopen(vma);
 806	lut->handle = handle;
 807	lut->ctx = ctx;
 808
 809	/* Check that the context hasn't been closed in the meantime */
 810	err = -EINTR;
 811	if (!mutex_lock_interruptible(&ctx->lut_mutex)) {
 812		struct i915_address_space *vm = rcu_access_pointer(ctx->vm);
 813
 814		if (unlikely(vm && vma->vm != vm))
 815			err = -EAGAIN; /* user racing with ctx set-vm */
 816		else if (likely(!i915_gem_context_is_closed(ctx)))
 817			err = radix_tree_insert(&ctx->handles_vma, handle, vma);
 818		else
 819			err = -ENOENT;
 820		if (err == 0) { /* And nor has this handle */
 821			struct drm_i915_gem_object *obj = vma->obj;
 822
 823			spin_lock(&obj->lut_lock);
 824			if (idr_find(&eb->file->object_idr, handle) == obj) {
 825				list_add(&lut->obj_link, &obj->lut_list);
 826			} else {
 827				radix_tree_delete(&ctx->handles_vma, handle);
 828				err = -ENOENT;
 829			}
 830			spin_unlock(&obj->lut_lock);
 831		}
 832		mutex_unlock(&ctx->lut_mutex);
 833	}
 834	if (unlikely(err))
 835		goto err;
 836
 837	return 0;
 838
 839err:
 840	i915_vma_close(vma);
 841	i915_vma_put(vma);
 842	i915_lut_handle_free(lut);
 843	return err;
 844}
 845
 846static struct i915_vma *eb_lookup_vma(struct i915_execbuffer *eb, u32 handle)
 847{
 848	struct i915_address_space *vm = eb->context->vm;
 849
 850	do {
 851		struct drm_i915_gem_object *obj;
 852		struct i915_vma *vma;
 853		int err;
 854
 855		rcu_read_lock();
 856		vma = radix_tree_lookup(&eb->gem_context->handles_vma, handle);
 857		if (likely(vma && vma->vm == vm))
 858			vma = i915_vma_tryget(vma);
 859		rcu_read_unlock();
 860		if (likely(vma))
 861			return vma;
 862
 863		obj = i915_gem_object_lookup(eb->file, handle);
 864		if (unlikely(!obj))
 865			return ERR_PTR(-ENOENT);
 866
 867		vma = i915_vma_instance(obj, vm, NULL);
 868		if (IS_ERR(vma)) {
 869			i915_gem_object_put(obj);
 870			return vma;
 871		}
 872
 873		err = __eb_add_lut(eb, handle, vma);
 874		if (likely(!err))
 875			return vma;
 876
 877		i915_gem_object_put(obj);
 878		if (err != -EEXIST)
 879			return ERR_PTR(err);
 880	} while (1);
 881}
 882
 883static int eb_lookup_vmas(struct i915_execbuffer *eb)
 884{
 
 885	unsigned int batch = eb_batch_index(eb);
 886	unsigned int i;
 887	int err = 0;
 888
 889	INIT_LIST_HEAD(&eb->relocs);
 890	INIT_LIST_HEAD(&eb->unbound);
 891
 892	for (i = 0; i < eb->buffer_count; i++) {
 893		struct i915_vma *vma;
 894
 895		vma = eb_lookup_vma(eb, eb->exec[i].handle);
 896		if (IS_ERR(vma)) {
 897			err = PTR_ERR(vma);
 898			break;
 899		}
 900
 901		err = eb_validate_vma(eb, &eb->exec[i], vma);
 902		if (unlikely(err)) {
 903			i915_vma_put(vma);
 904			break;
 905		}
 906
 907		eb_add_vma(eb, i, batch, vma);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 908	}
 909
 
 
 
 910	eb->vma[i].vma = NULL;
 911	return err;
 912}
 913
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 914static struct eb_vma *
 915eb_get_vma(const struct i915_execbuffer *eb, unsigned long handle)
 916{
 917	if (eb->lut_size < 0) {
 918		if (handle >= -eb->lut_size)
 919			return NULL;
 920		return &eb->vma[handle];
 921	} else {
 922		struct hlist_head *head;
 923		struct eb_vma *ev;
 924
 925		head = &eb->buckets[hash_32(handle, eb->lut_size)];
 926		hlist_for_each_entry(ev, head, node) {
 927			if (ev->handle == handle)
 928				return ev;
 929		}
 930		return NULL;
 931	}
 932}
 933
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 934static void eb_destroy(const struct i915_execbuffer *eb)
 935{
 936	GEM_BUG_ON(eb->reloc_cache.rq);
 937
 938	if (eb->array)
 939		eb_vma_array_put(eb->array);
 940
 941	if (eb->lut_size > 0)
 942		kfree(eb->buckets);
 943}
 944
 945static inline u64
 946relocation_target(const struct drm_i915_gem_relocation_entry *reloc,
 947		  const struct i915_vma *target)
 948{
 949	return gen8_canonical_addr((int)reloc->delta + target->node.start);
 950}
 951
 
 
 
 
 
 
 
 
 952static void reloc_cache_init(struct reloc_cache *cache,
 953			     struct drm_i915_private *i915)
 954{
 955	cache->page = -1;
 956	cache->vaddr = 0;
 957	/* Must be a variable in the struct to allow GCC to unroll. */
 958	cache->gen = INTEL_GEN(i915);
 959	cache->has_llc = HAS_LLC(i915);
 960	cache->use_64bit_reloc = HAS_64BIT_RELOC(i915);
 961	cache->has_fence = cache->gen < 4;
 962	cache->needs_unfenced = INTEL_INFO(i915)->unfenced_needs_alignment;
 963	cache->node.flags = 0;
 964	cache->rq = NULL;
 965	cache->target = NULL;
 966}
 967
 968static inline void *unmask_page(unsigned long p)
 969{
 970	return (void *)(uintptr_t)(p & PAGE_MASK);
 971}
 972
 973static inline unsigned int unmask_flags(unsigned long p)
 974{
 975	return p & ~PAGE_MASK;
 976}
 977
 978#define KMAP 0x4 /* after CLFLUSH_FLAGS */
 979
 980static inline struct i915_ggtt *cache_to_ggtt(struct reloc_cache *cache)
 981{
 982	struct drm_i915_private *i915 =
 983		container_of(cache, struct i915_execbuffer, reloc_cache)->i915;
 984	return &i915->ggtt;
 985}
 986
 987#define RELOC_TAIL 4
 988
 989static int reloc_gpu_chain(struct reloc_cache *cache)
 990{
 991	struct intel_gt_buffer_pool_node *pool;
 992	struct i915_request *rq = cache->rq;
 993	struct i915_vma *batch;
 994	u32 *cmd;
 995	int err;
 996
 997	pool = intel_gt_get_buffer_pool(rq->engine->gt, PAGE_SIZE);
 998	if (IS_ERR(pool))
 999		return PTR_ERR(pool);
1000
1001	batch = i915_vma_instance(pool->obj, rq->context->vm, NULL);
1002	if (IS_ERR(batch)) {
1003		err = PTR_ERR(batch);
1004		goto out_pool;
1005	}
1006
1007	err = i915_vma_pin(batch, 0, 0, PIN_USER | PIN_NONBLOCK);
1008	if (err)
1009		goto out_pool;
1010
1011	GEM_BUG_ON(cache->rq_size + RELOC_TAIL > PAGE_SIZE  / sizeof(u32));
1012	cmd = cache->rq_cmd + cache->rq_size;
1013	*cmd++ = MI_ARB_CHECK;
1014	if (cache->gen >= 8)
1015		*cmd++ = MI_BATCH_BUFFER_START_GEN8;
1016	else if (cache->gen >= 6)
1017		*cmd++ = MI_BATCH_BUFFER_START;
1018	else
1019		*cmd++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT;
1020	*cmd++ = lower_32_bits(batch->node.start);
1021	*cmd++ = upper_32_bits(batch->node.start); /* Always 0 for gen<8 */
1022	i915_gem_object_flush_map(cache->rq_vma->obj);
1023	i915_gem_object_unpin_map(cache->rq_vma->obj);
1024	cache->rq_vma = NULL;
1025
1026	err = intel_gt_buffer_pool_mark_active(pool, rq);
1027	if (err == 0) {
1028		i915_vma_lock(batch);
1029		err = i915_request_await_object(rq, batch->obj, false);
1030		if (err == 0)
1031			err = i915_vma_move_to_active(batch, rq, 0);
1032		i915_vma_unlock(batch);
1033	}
1034	i915_vma_unpin(batch);
1035	if (err)
1036		goto out_pool;
1037
1038	cmd = i915_gem_object_pin_map(batch->obj,
1039				      cache->has_llc ?
1040				      I915_MAP_FORCE_WB :
1041				      I915_MAP_FORCE_WC);
1042	if (IS_ERR(cmd)) {
1043		err = PTR_ERR(cmd);
1044		goto out_pool;
1045	}
1046
1047	/* Return with batch mapping (cmd) still pinned */
1048	cache->rq_cmd = cmd;
1049	cache->rq_size = 0;
1050	cache->rq_vma = batch;
1051
1052out_pool:
1053	intel_gt_buffer_pool_put(pool);
1054	return err;
1055}
1056
1057static unsigned int reloc_bb_flags(const struct reloc_cache *cache)
1058{
1059	return cache->gen > 5 ? 0 : I915_DISPATCH_SECURE;
 
 
 
 
 
 
1060}
1061
1062static int reloc_gpu_flush(struct reloc_cache *cache)
1063{
1064	struct i915_request *rq;
1065	int err;
1066
1067	rq = fetch_and_zero(&cache->rq);
1068	if (!rq)
1069		return 0;
1070
1071	if (cache->rq_vma) {
1072		struct drm_i915_gem_object *obj = cache->rq_vma->obj;
1073
1074		GEM_BUG_ON(cache->rq_size >= obj->base.size / sizeof(u32));
1075		cache->rq_cmd[cache->rq_size++] = MI_BATCH_BUFFER_END;
1076
1077		__i915_gem_object_flush_map(obj,
1078					    0, sizeof(u32) * cache->rq_size);
1079		i915_gem_object_unpin_map(obj);
1080	}
1081
1082	err = 0;
1083	if (rq->engine->emit_init_breadcrumb)
1084		err = rq->engine->emit_init_breadcrumb(rq);
1085	if (!err)
1086		err = rq->engine->emit_bb_start(rq,
1087						rq->batch->node.start,
1088						PAGE_SIZE,
1089						reloc_bb_flags(cache));
1090	if (err)
1091		i915_request_set_error_once(rq, err);
1092
1093	intel_gt_chipset_flush(rq->engine->gt);
1094	i915_request_add(rq);
1095
1096	return err;
1097}
1098
1099static void reloc_cache_reset(struct reloc_cache *cache)
1100{
1101	void *vaddr;
1102
 
 
 
1103	if (!cache->vaddr)
1104		return;
1105
1106	vaddr = unmask_page(cache->vaddr);
1107	if (cache->vaddr & KMAP) {
 
 
1108		if (cache->vaddr & CLFLUSH_AFTER)
1109			mb();
1110
1111		kunmap_atomic(vaddr);
1112		i915_gem_object_finish_access((struct drm_i915_gem_object *)cache->node.mm);
1113	} else {
1114		struct i915_ggtt *ggtt = cache_to_ggtt(cache);
1115
1116		intel_gt_flush_ggtt_writes(ggtt->vm.gt);
1117		io_mapping_unmap_atomic((void __iomem *)vaddr);
1118
1119		if (drm_mm_node_allocated(&cache->node)) {
1120			ggtt->vm.clear_range(&ggtt->vm,
1121					     cache->node.start,
1122					     cache->node.size);
1123			mutex_lock(&ggtt->vm.mutex);
1124			drm_mm_remove_node(&cache->node);
1125			mutex_unlock(&ggtt->vm.mutex);
1126		} else {
1127			i915_vma_unpin((struct i915_vma *)cache->node.mm);
1128		}
1129	}
1130
1131	cache->vaddr = 0;
1132	cache->page = -1;
1133}
1134
1135static void *reloc_kmap(struct drm_i915_gem_object *obj,
1136			struct reloc_cache *cache,
1137			unsigned long page)
1138{
1139	void *vaddr;
 
1140
1141	if (cache->vaddr) {
1142		kunmap_atomic(unmask_page(cache->vaddr));
1143	} else {
1144		unsigned int flushes;
1145		int err;
1146
1147		err = i915_gem_object_prepare_write(obj, &flushes);
1148		if (err)
1149			return ERR_PTR(err);
1150
1151		BUILD_BUG_ON(KMAP & CLFLUSH_FLAGS);
1152		BUILD_BUG_ON((KMAP | CLFLUSH_FLAGS) & PAGE_MASK);
1153
1154		cache->vaddr = flushes | KMAP;
1155		cache->node.mm = (void *)obj;
1156		if (flushes)
1157			mb();
1158	}
1159
1160	vaddr = kmap_atomic(i915_gem_object_get_dirty_page(obj, page));
 
 
 
 
1161	cache->vaddr = unmask_flags(cache->vaddr) | (unsigned long)vaddr;
1162	cache->page = page;
1163
1164	return vaddr;
1165}
1166
1167static void *reloc_iomap(struct drm_i915_gem_object *obj,
1168			 struct reloc_cache *cache,
1169			 unsigned long page)
1170{
 
1171	struct i915_ggtt *ggtt = cache_to_ggtt(cache);
1172	unsigned long offset;
1173	void *vaddr;
1174
1175	if (cache->vaddr) {
1176		intel_gt_flush_ggtt_writes(ggtt->vm.gt);
1177		io_mapping_unmap_atomic((void __force __iomem *) unmask_page(cache->vaddr));
1178	} else {
1179		struct i915_vma *vma;
1180		int err;
1181
1182		if (i915_gem_object_is_tiled(obj))
1183			return ERR_PTR(-EINVAL);
1184
1185		if (use_cpu_reloc(cache, obj))
1186			return NULL;
1187
1188		i915_gem_object_lock(obj);
1189		err = i915_gem_object_set_to_gtt_domain(obj, true);
1190		i915_gem_object_unlock(obj);
1191		if (err)
1192			return ERR_PTR(err);
1193
1194		vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
1195					       PIN_MAPPABLE |
1196					       PIN_NONBLOCK /* NOWARN */ |
1197					       PIN_NOEVICT);
 
 
 
1198		if (IS_ERR(vma)) {
1199			memset(&cache->node, 0, sizeof(cache->node));
1200			mutex_lock(&ggtt->vm.mutex);
1201			err = drm_mm_insert_node_in_range
1202				(&ggtt->vm.mm, &cache->node,
1203				 PAGE_SIZE, 0, I915_COLOR_UNEVICTABLE,
1204				 0, ggtt->mappable_end,
1205				 DRM_MM_INSERT_LOW);
1206			mutex_unlock(&ggtt->vm.mutex);
1207			if (err) /* no inactive aperture space, use cpu reloc */
1208				return NULL;
1209		} else {
1210			cache->node.start = vma->node.start;
1211			cache->node.mm = (void *)vma;
1212		}
1213	}
1214
1215	offset = cache->node.start;
1216	if (drm_mm_node_allocated(&cache->node)) {
1217		ggtt->vm.insert_page(&ggtt->vm,
1218				     i915_gem_object_get_dma_address(obj, page),
1219				     offset, I915_CACHE_NONE, 0);
1220	} else {
1221		offset += page << PAGE_SHIFT;
1222	}
1223
1224	vaddr = (void __force *)io_mapping_map_atomic_wc(&ggtt->iomap,
1225							 offset);
1226	cache->page = page;
1227	cache->vaddr = (unsigned long)vaddr;
1228
1229	return vaddr;
1230}
1231
1232static void *reloc_vaddr(struct drm_i915_gem_object *obj,
1233			 struct reloc_cache *cache,
1234			 unsigned long page)
1235{
 
1236	void *vaddr;
1237
1238	if (cache->page == page) {
1239		vaddr = unmask_page(cache->vaddr);
1240	} else {
1241		vaddr = NULL;
1242		if ((cache->vaddr & KMAP) == 0)
1243			vaddr = reloc_iomap(obj, cache, page);
1244		if (!vaddr)
1245			vaddr = reloc_kmap(obj, cache, page);
1246	}
1247
1248	return vaddr;
1249}
1250
1251static void clflush_write32(u32 *addr, u32 value, unsigned int flushes)
1252{
1253	if (unlikely(flushes & (CLFLUSH_BEFORE | CLFLUSH_AFTER))) {
1254		if (flushes & CLFLUSH_BEFORE) {
1255			clflushopt(addr);
1256			mb();
1257		}
1258
1259		*addr = value;
1260
1261		/*
1262		 * Writes to the same cacheline are serialised by the CPU
1263		 * (including clflush). On the write path, we only require
1264		 * that it hits memory in an orderly fashion and place
1265		 * mb barriers at the start and end of the relocation phase
1266		 * to ensure ordering of clflush wrt to the system.
1267		 */
1268		if (flushes & CLFLUSH_AFTER)
1269			clflushopt(addr);
1270	} else
1271		*addr = value;
1272}
1273
1274static int reloc_move_to_gpu(struct i915_request *rq, struct i915_vma *vma)
1275{
1276	struct drm_i915_gem_object *obj = vma->obj;
1277	int err;
1278
1279	i915_vma_lock(vma);
1280
1281	if (obj->cache_dirty & ~obj->cache_coherent)
1282		i915_gem_clflush_object(obj, 0);
1283	obj->write_domain = 0;
1284
1285	err = i915_request_await_object(rq, vma->obj, true);
1286	if (err == 0)
1287		err = i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE);
1288
1289	i915_vma_unlock(vma);
1290
1291	return err;
1292}
1293
1294static int __reloc_gpu_alloc(struct i915_execbuffer *eb,
1295			     struct intel_engine_cs *engine,
 
1296			     unsigned int len)
1297{
1298	struct reloc_cache *cache = &eb->reloc_cache;
1299	struct intel_gt_buffer_pool_node *pool;
1300	struct i915_request *rq;
1301	struct i915_vma *batch;
1302	u32 *cmd;
1303	int err;
1304
1305	pool = intel_gt_get_buffer_pool(engine->gt, PAGE_SIZE);
1306	if (IS_ERR(pool))
1307		return PTR_ERR(pool);
1308
1309	cmd = i915_gem_object_pin_map(pool->obj,
1310				      cache->has_llc ?
1311				      I915_MAP_FORCE_WB :
1312				      I915_MAP_FORCE_WC);
 
 
 
 
 
 
 
1313	if (IS_ERR(cmd)) {
1314		err = PTR_ERR(cmd);
1315		goto out_pool;
1316	}
 
 
 
1317
1318	batch = i915_vma_instance(pool->obj, eb->context->vm, NULL);
1319	if (IS_ERR(batch)) {
1320		err = PTR_ERR(batch);
1321		goto err_unmap;
1322	}
1323
1324	err = i915_vma_pin(batch, 0, 0, PIN_USER | PIN_NONBLOCK);
1325	if (err)
1326		goto err_unmap;
1327
1328	if (engine == eb->context->engine) {
1329		rq = i915_request_create(eb->context);
1330	} else {
1331		struct intel_context *ce;
1332
1333		ce = intel_context_create(engine);
1334		if (IS_ERR(ce)) {
1335			err = PTR_ERR(ce);
1336			goto err_unpin;
 
 
 
 
 
 
1337		}
1338
1339		i915_vm_put(ce->vm);
1340		ce->vm = i915_vm_get(eb->context->vm);
 
1341
1342		rq = intel_context_create_request(ce);
1343		intel_context_put(ce);
1344	}
1345	if (IS_ERR(rq)) {
1346		err = PTR_ERR(rq);
1347		goto err_unpin;
1348	}
1349
1350	err = intel_gt_buffer_pool_mark_active(pool, rq);
1351	if (err)
1352		goto err_request;
1353
1354	i915_vma_lock(batch);
 
 
 
 
 
 
 
 
 
 
1355	err = i915_request_await_object(rq, batch->obj, false);
1356	if (err == 0)
1357		err = i915_vma_move_to_active(batch, rq, 0);
1358	i915_vma_unlock(batch);
1359	if (err)
1360		goto skip_request;
1361
1362	rq->batch = batch;
1363	i915_vma_unpin(batch);
1364
1365	cache->rq = rq;
1366	cache->rq_cmd = cmd;
1367	cache->rq_size = 0;
1368	cache->rq_vma = batch;
1369
1370	/* Return with batch mapping (cmd) still pinned */
1371	goto out_pool;
1372
1373skip_request:
1374	i915_request_set_error_once(rq, err);
1375err_request:
1376	i915_request_add(rq);
1377err_unpin:
1378	i915_vma_unpin(batch);
1379err_unmap:
1380	i915_gem_object_unpin_map(pool->obj);
1381out_pool:
1382	intel_gt_buffer_pool_put(pool);
1383	return err;
1384}
1385
1386static bool reloc_can_use_engine(const struct intel_engine_cs *engine)
1387{
1388	return engine->class != VIDEO_DECODE_CLASS || !IS_GEN(engine->i915, 6);
1389}
1390
1391static u32 *reloc_gpu(struct i915_execbuffer *eb,
1392		      struct i915_vma *vma,
1393		      unsigned int len)
1394{
1395	struct reloc_cache *cache = &eb->reloc_cache;
1396	u32 *cmd;
1397	int err;
 
 
1398
1399	if (unlikely(!cache->rq)) {
 
1400		struct intel_engine_cs *engine = eb->engine;
1401
 
 
 
 
1402		if (!reloc_can_use_engine(engine)) {
1403			engine = engine->gt->engine_class[COPY_ENGINE_CLASS][0];
1404			if (!engine)
1405				return ERR_PTR(-ENODEV);
1406		}
1407
1408		err = __reloc_gpu_alloc(eb, engine, len);
1409		if (unlikely(err))
1410			return ERR_PTR(err);
1411	}
1412
1413	if (vma != cache->target) {
1414		err = reloc_move_to_gpu(cache->rq, vma);
1415		if (unlikely(err)) {
1416			i915_request_set_error_once(cache->rq, err);
1417			return ERR_PTR(err);
1418		}
1419
1420		cache->target = vma;
1421	}
1422
1423	if (unlikely(cache->rq_size + len >
1424		     PAGE_SIZE / sizeof(u32) - RELOC_TAIL)) {
1425		err = reloc_gpu_chain(cache);
1426		if (unlikely(err)) {
1427			i915_request_set_error_once(cache->rq, err);
1428			return ERR_PTR(err);
1429		}
1430	}
1431
1432	GEM_BUG_ON(cache->rq_size + len >= PAGE_SIZE  / sizeof(u32));
1433	cmd = cache->rq_cmd + cache->rq_size;
1434	cache->rq_size += len;
1435
1436	return cmd;
1437}
1438
1439static inline bool use_reloc_gpu(struct i915_vma *vma)
1440{
1441	if (DBG_FORCE_RELOC == FORCE_GPU_RELOC)
1442		return true;
1443
1444	if (DBG_FORCE_RELOC)
1445		return false;
1446
1447	return !dma_resv_test_signaled_rcu(vma->resv, true);
1448}
1449
1450static unsigned long vma_phys_addr(struct i915_vma *vma, u32 offset)
1451{
1452	struct page *page;
1453	unsigned long addr;
1454
1455	GEM_BUG_ON(vma->pages != vma->obj->mm.pages);
1456
1457	page = i915_gem_object_get_page(vma->obj, offset >> PAGE_SHIFT);
1458	addr = PFN_PHYS(page_to_pfn(page));
1459	GEM_BUG_ON(overflows_type(addr, u32)); /* expected dma32 */
1460
1461	return addr + offset_in_page(offset);
1462}
1463
1464static bool __reloc_entry_gpu(struct i915_execbuffer *eb,
1465			      struct i915_vma *vma,
1466			      u64 offset,
1467			      u64 target_addr)
1468{
1469	const unsigned int gen = eb->reloc_cache.gen;
1470	unsigned int len;
1471	u32 *batch;
1472	u64 addr;
1473
1474	if (gen >= 8)
1475		len = offset & 7 ? 8 : 5;
1476	else if (gen >= 4)
1477		len = 4;
1478	else
1479		len = 3;
1480
1481	batch = reloc_gpu(eb, vma, len);
1482	if (IS_ERR(batch))
 
 
1483		return false;
1484
1485	addr = gen8_canonical_addr(vma->node.start + offset);
1486	if (gen >= 8) {
1487		if (offset & 7) {
1488			*batch++ = MI_STORE_DWORD_IMM_GEN4;
1489			*batch++ = lower_32_bits(addr);
1490			*batch++ = upper_32_bits(addr);
1491			*batch++ = lower_32_bits(target_addr);
1492
1493			addr = gen8_canonical_addr(addr + 4);
1494
1495			*batch++ = MI_STORE_DWORD_IMM_GEN4;
1496			*batch++ = lower_32_bits(addr);
1497			*batch++ = upper_32_bits(addr);
1498			*batch++ = upper_32_bits(target_addr);
1499		} else {
1500			*batch++ = (MI_STORE_DWORD_IMM_GEN4 | (1 << 21)) + 1;
1501			*batch++ = lower_32_bits(addr);
1502			*batch++ = upper_32_bits(addr);
1503			*batch++ = lower_32_bits(target_addr);
1504			*batch++ = upper_32_bits(target_addr);
1505		}
1506	} else if (gen >= 6) {
1507		*batch++ = MI_STORE_DWORD_IMM_GEN4;
1508		*batch++ = 0;
1509		*batch++ = addr;
1510		*batch++ = target_addr;
1511	} else if (IS_I965G(eb->i915)) {
1512		*batch++ = MI_STORE_DWORD_IMM_GEN4;
1513		*batch++ = 0;
1514		*batch++ = vma_phys_addr(vma, offset);
1515		*batch++ = target_addr;
1516	} else if (gen >= 4) {
1517		*batch++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
1518		*batch++ = 0;
1519		*batch++ = addr;
1520		*batch++ = target_addr;
1521	} else if (gen >= 3 &&
1522		   !(IS_I915G(eb->i915) || IS_I915GM(eb->i915))) {
1523		*batch++ = MI_STORE_DWORD_IMM | MI_MEM_VIRTUAL;
1524		*batch++ = addr;
1525		*batch++ = target_addr;
1526	} else {
1527		*batch++ = MI_STORE_DWORD_IMM;
1528		*batch++ = vma_phys_addr(vma, offset);
1529		*batch++ = target_addr;
1530	}
1531
1532	return true;
1533}
1534
1535static bool reloc_entry_gpu(struct i915_execbuffer *eb,
1536			    struct i915_vma *vma,
1537			    u64 offset,
1538			    u64 target_addr)
1539{
1540	if (eb->reloc_cache.vaddr)
1541		return false;
1542
1543	if (!use_reloc_gpu(vma))
1544		return false;
1545
1546	return __reloc_entry_gpu(eb, vma, offset, target_addr);
1547}
1548
1549static u64
1550relocate_entry(struct i915_vma *vma,
1551	       const struct drm_i915_gem_relocation_entry *reloc,
1552	       struct i915_execbuffer *eb,
1553	       const struct i915_vma *target)
1554{
1555	u64 target_addr = relocation_target(reloc, target);
1556	u64 offset = reloc->offset;
 
 
 
 
1557
1558	if (!reloc_entry_gpu(eb, vma, offset, target_addr)) {
1559		bool wide = eb->reloc_cache.use_64bit_reloc;
1560		void *vaddr;
1561
1562repeat:
1563		vaddr = reloc_vaddr(vma->obj,
1564				    &eb->reloc_cache,
1565				    offset >> PAGE_SHIFT);
1566		if (IS_ERR(vaddr))
1567			return PTR_ERR(vaddr);
1568
1569		GEM_BUG_ON(!IS_ALIGNED(offset, sizeof(u32)));
1570		clflush_write32(vaddr + offset_in_page(offset),
1571				lower_32_bits(target_addr),
1572				eb->reloc_cache.vaddr);
1573
1574		if (wide) {
1575			offset += sizeof(u32);
1576			target_addr >>= 32;
1577			wide = false;
1578			goto repeat;
1579		}
1580	}
1581
1582	return target->node.start | UPDATE;
1583}
1584
1585static u64
1586eb_relocate_entry(struct i915_execbuffer *eb,
1587		  struct eb_vma *ev,
1588		  const struct drm_i915_gem_relocation_entry *reloc)
1589{
1590	struct drm_i915_private *i915 = eb->i915;
1591	struct eb_vma *target;
1592	int err;
1593
1594	/* we've already hold a reference to all valid objects */
1595	target = eb_get_vma(eb, reloc->target_handle);
1596	if (unlikely(!target))
1597		return -ENOENT;
1598
1599	/* Validate that the target is in a valid r/w GPU domain */
1600	if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
1601		drm_dbg(&i915->drm, "reloc with multiple write domains: "
1602			  "target %d offset %d "
1603			  "read %08x write %08x",
1604			  reloc->target_handle,
1605			  (int) reloc->offset,
1606			  reloc->read_domains,
1607			  reloc->write_domain);
1608		return -EINVAL;
1609	}
1610	if (unlikely((reloc->write_domain | reloc->read_domains)
1611		     & ~I915_GEM_GPU_DOMAINS)) {
1612		drm_dbg(&i915->drm, "reloc with read/write non-GPU domains: "
1613			  "target %d offset %d "
1614			  "read %08x write %08x",
1615			  reloc->target_handle,
1616			  (int) reloc->offset,
1617			  reloc->read_domains,
1618			  reloc->write_domain);
1619		return -EINVAL;
1620	}
1621
1622	if (reloc->write_domain) {
1623		target->flags |= EXEC_OBJECT_WRITE;
1624
1625		/*
1626		 * Sandybridge PPGTT errata: We need a global gtt mapping
1627		 * for MI and pipe_control writes because the gpu doesn't
1628		 * properly redirect them through the ppgtt for non_secure
1629		 * batchbuffers.
1630		 */
1631		if (reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
1632		    IS_GEN(eb->i915, 6)) {
1633			err = i915_vma_bind(target->vma,
1634					    target->vma->obj->cache_level,
1635					    PIN_GLOBAL, NULL);
1636			if (err)
1637				return err;
1638		}
1639	}
1640
1641	/*
1642	 * If the relocation already has the right value in it, no
1643	 * more work needs to be done.
1644	 */
1645	if (!DBG_FORCE_RELOC &&
1646	    gen8_canonical_addr(target->vma->node.start) == reloc->presumed_offset)
1647		return 0;
1648
1649	/* Check that the relocation address is valid... */
1650	if (unlikely(reloc->offset >
1651		     ev->vma->size - (eb->reloc_cache.use_64bit_reloc ? 8 : 4))) {
1652		drm_dbg(&i915->drm, "Relocation beyond object bounds: "
1653			  "target %d offset %d size %d.\n",
1654			  reloc->target_handle,
1655			  (int)reloc->offset,
1656			  (int)ev->vma->size);
1657		return -EINVAL;
1658	}
1659	if (unlikely(reloc->offset & 3)) {
1660		drm_dbg(&i915->drm, "Relocation not 4-byte aligned: "
1661			  "target %d offset %d.\n",
1662			  reloc->target_handle,
1663			  (int)reloc->offset);
1664		return -EINVAL;
1665	}
1666
1667	/*
1668	 * If we write into the object, we need to force the synchronisation
1669	 * barrier, either with an asynchronous clflush or if we executed the
1670	 * patching using the GPU (though that should be serialised by the
1671	 * timeline). To be completely sure, and since we are required to
1672	 * do relocations we are already stalling, disable the user's opt
1673	 * out of our synchronisation.
1674	 */
1675	ev->flags &= ~EXEC_OBJECT_ASYNC;
1676
1677	/* and update the user's relocation entry */
1678	return relocate_entry(ev->vma, reloc, eb, target->vma);
1679}
1680
1681static int eb_relocate_vma(struct i915_execbuffer *eb, struct eb_vma *ev)
1682{
1683#define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
1684	struct drm_i915_gem_relocation_entry stack[N_RELOC(512)];
1685	const struct drm_i915_gem_exec_object2 *entry = ev->exec;
1686	struct drm_i915_gem_relocation_entry __user *urelocs =
1687		u64_to_user_ptr(entry->relocs_ptr);
1688	unsigned long remain = entry->relocation_count;
1689
1690	if (unlikely(remain > N_RELOC(ULONG_MAX)))
1691		return -EINVAL;
1692
1693	/*
1694	 * We must check that the entire relocation array is safe
1695	 * to read. However, if the array is not writable the user loses
1696	 * the updated relocation values.
1697	 */
1698	if (unlikely(!access_ok(urelocs, remain * sizeof(*urelocs))))
1699		return -EFAULT;
1700
1701	do {
1702		struct drm_i915_gem_relocation_entry *r = stack;
1703		unsigned int count =
1704			min_t(unsigned long, remain, ARRAY_SIZE(stack));
1705		unsigned int copied;
1706
1707		/*
1708		 * This is the fast path and we cannot handle a pagefault
1709		 * whilst holding the struct mutex lest the user pass in the
1710		 * relocations contained within a mmaped bo. For in such a case
1711		 * we, the page fault handler would call i915_gem_fault() and
1712		 * we would try to acquire the struct mutex again. Obviously
1713		 * this is bad and so lockdep complains vehemently.
1714		 */
1715		copied = __copy_from_user(r, urelocs, count * sizeof(r[0]));
 
 
1716		if (unlikely(copied)) {
1717			remain = -EFAULT;
1718			goto out;
1719		}
1720
1721		remain -= count;
1722		do {
1723			u64 offset = eb_relocate_entry(eb, ev, r);
1724
1725			if (likely(offset == 0)) {
1726			} else if ((s64)offset < 0) {
1727				remain = (int)offset;
1728				goto out;
1729			} else {
1730				/*
1731				 * Note that reporting an error now
1732				 * leaves everything in an inconsistent
1733				 * state as we have *already* changed
1734				 * the relocation value inside the
1735				 * object. As we have not changed the
1736				 * reloc.presumed_offset or will not
1737				 * change the execobject.offset, on the
1738				 * call we may not rewrite the value
1739				 * inside the object, leaving it
1740				 * dangling and causing a GPU hang. Unless
1741				 * userspace dynamically rebuilds the
1742				 * relocations on each execbuf rather than
1743				 * presume a static tree.
1744				 *
1745				 * We did previously check if the relocations
1746				 * were writable (access_ok), an error now
1747				 * would be a strange race with mprotect,
1748				 * having already demonstrated that we
1749				 * can read from this userspace address.
1750				 */
1751				offset = gen8_canonical_addr(offset & ~UPDATE);
1752				__put_user(offset,
1753					   &urelocs[r - stack].presumed_offset);
1754			}
1755		} while (r++, --count);
1756		urelocs += ARRAY_SIZE(stack);
1757	} while (remain);
1758out:
1759	reloc_cache_reset(&eb->reloc_cache);
1760	return remain;
1761}
1762
1763static int eb_relocate(struct i915_execbuffer *eb)
 
1764{
 
 
 
 
1765	int err;
1766
1767	err = eb_lookup_vmas(eb);
1768	if (err)
1769		return err;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1770
1771	if (!list_empty(&eb->unbound)) {
1772		err = eb_reserve(eb);
 
 
 
 
 
 
1773		if (err)
1774			return err;
1775	}
 
 
1776
1777	/* The objects are in their final locations, apply the relocations. */
1778	if (eb->args->flags & __EXEC_HAS_RELOC) {
1779		struct eb_vma *ev;
1780		int flush;
 
 
1781
1782		list_for_each_entry(ev, &eb->relocs, reloc_link) {
1783			err = eb_relocate_vma(eb, ev);
1784			if (err)
1785				break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1786		}
1787
1788		flush = reloc_gpu_flush(&eb->reloc_cache);
1789		if (!err)
1790			err = flush;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1791	}
1792
 
 
 
 
 
 
 
 
 
 
 
 
 
1793	return err;
1794}
1795
1796static int eb_move_to_gpu(struct i915_execbuffer *eb)
1797{
1798	const unsigned int count = eb->buffer_count;
1799	struct ww_acquire_ctx acquire;
1800	unsigned int i;
1801	int err = 0;
1802
1803	ww_acquire_init(&acquire, &reservation_ww_class);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1804
1805	for (i = 0; i < count; i++) {
1806		struct eb_vma *ev = &eb->vma[i];
1807		struct i915_vma *vma = ev->vma;
1808
1809		err = ww_mutex_lock_interruptible(&vma->resv->lock, &acquire);
1810		if (err == -EDEADLK) {
1811			GEM_BUG_ON(i == 0);
1812			do {
1813				int j = i - 1;
1814
1815				ww_mutex_unlock(&eb->vma[j].vma->resv->lock);
 
 
1816
1817				swap(eb->vma[i],  eb->vma[j]);
1818			} while (--i);
 
 
 
1819
1820			err = ww_mutex_lock_slow_interruptible(&vma->resv->lock,
1821							       &acquire);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1822		}
1823		if (err)
1824			break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1825	}
1826	ww_acquire_done(&acquire);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1827
1828	while (i--) {
1829		struct eb_vma *ev = &eb->vma[i];
1830		struct i915_vma *vma = ev->vma;
1831		unsigned int flags = ev->flags;
1832		struct drm_i915_gem_object *obj = vma->obj;
1833
1834		assert_vma_held(vma);
1835
1836		if (flags & EXEC_OBJECT_CAPTURE) {
1837			struct i915_capture_list *capture;
1838
1839			capture = kmalloc(sizeof(*capture), GFP_KERNEL);
1840			if (capture) {
1841				capture->next = eb->request->capture_list;
1842				capture->vma = vma;
1843				eb->request->capture_list = capture;
1844			}
1845		}
1846
1847		/*
1848		 * If the GPU is not _reading_ through the CPU cache, we need
1849		 * to make sure that any writes (both previous GPU writes from
1850		 * before a change in snooping levels and normal CPU writes)
1851		 * caught in that cache are flushed to main memory.
1852		 *
1853		 * We want to say
1854		 *   obj->cache_dirty &&
1855		 *   !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ)
1856		 * but gcc's optimiser doesn't handle that as well and emits
1857		 * two jumps instead of one. Maybe one day...
1858		 */
1859		if (unlikely(obj->cache_dirty & ~obj->cache_coherent)) {
1860			if (i915_gem_clflush_object(obj, 0))
1861				flags &= ~EXEC_OBJECT_ASYNC;
1862		}
1863
1864		if (err == 0 && !(flags & EXEC_OBJECT_ASYNC)) {
1865			err = i915_request_await_object
1866				(eb->request, obj, flags & EXEC_OBJECT_WRITE);
1867		}
1868
1869		if (err == 0)
1870			err = i915_vma_move_to_active(vma, eb->request, flags);
1871
1872		i915_vma_unlock(vma);
1873		eb_unreserve_vma(ev);
1874	}
1875	ww_acquire_fini(&acquire);
1876
1877	eb_vma_array_put(fetch_and_zero(&eb->array));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1878
1879	if (unlikely(err))
1880		goto err_skip;
1881
1882	/* Unconditionally flush any chipset caches (for streaming writes). */
1883	intel_gt_chipset_flush(eb->engine->gt);
1884	return 0;
1885
1886err_skip:
1887	i915_request_set_error_once(eb->request, err);
1888	return err;
1889}
1890
1891static int i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
1892{
1893	if (exec->flags & __I915_EXEC_ILLEGAL_FLAGS)
1894		return -EINVAL;
1895
1896	/* Kernel clipping was a DRI1 misfeature */
1897	if (!(exec->flags & I915_EXEC_FENCE_ARRAY)) {
 
1898		if (exec->num_cliprects || exec->cliprects_ptr)
1899			return -EINVAL;
1900	}
1901
1902	if (exec->DR4 == 0xffffffff) {
1903		DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
1904		exec->DR4 = 0;
1905	}
1906	if (exec->DR1 || exec->DR4)
1907		return -EINVAL;
1908
1909	if ((exec->batch_start_offset | exec->batch_len) & 0x7)
1910		return -EINVAL;
1911
1912	return 0;
1913}
1914
1915static int i915_reset_gen7_sol_offsets(struct i915_request *rq)
1916{
1917	u32 *cs;
1918	int i;
1919
1920	if (!IS_GEN(rq->engine->i915, 7) || rq->engine->id != RCS0) {
1921		drm_dbg(&rq->engine->i915->drm, "sol reset is gen7/rcs only\n");
1922		return -EINVAL;
1923	}
1924
1925	cs = intel_ring_begin(rq, 4 * 2 + 2);
1926	if (IS_ERR(cs))
1927		return PTR_ERR(cs);
1928
1929	*cs++ = MI_LOAD_REGISTER_IMM(4);
1930	for (i = 0; i < 4; i++) {
1931		*cs++ = i915_mmio_reg_offset(GEN7_SO_WRITE_OFFSET(i));
1932		*cs++ = 0;
1933	}
1934	*cs++ = MI_NOOP;
1935	intel_ring_advance(rq, cs);
1936
1937	return 0;
1938}
1939
1940static struct i915_vma *
1941shadow_batch_pin(struct drm_i915_gem_object *obj,
 
1942		 struct i915_address_space *vm,
1943		 unsigned int flags)
1944{
1945	struct i915_vma *vma;
1946	int err;
1947
1948	vma = i915_vma_instance(obj, vm, NULL);
1949	if (IS_ERR(vma))
1950		return vma;
1951
1952	err = i915_vma_pin(vma, 0, 0, flags);
1953	if (err)
1954		return ERR_PTR(err);
1955
1956	return vma;
1957}
1958
1959struct eb_parse_work {
1960	struct dma_fence_work base;
1961	struct intel_engine_cs *engine;
1962	struct i915_vma *batch;
1963	struct i915_vma *shadow;
1964	struct i915_vma *trampoline;
1965	unsigned int batch_offset;
1966	unsigned int batch_length;
1967};
1968
1969static int __eb_parse(struct dma_fence_work *work)
1970{
1971	struct eb_parse_work *pw = container_of(work, typeof(*pw), base);
1972
1973	return intel_engine_cmd_parser(pw->engine,
1974				       pw->batch,
1975				       pw->batch_offset,
1976				       pw->batch_length,
1977				       pw->shadow,
1978				       pw->trampoline);
1979}
1980
1981static void __eb_parse_release(struct dma_fence_work *work)
1982{
1983	struct eb_parse_work *pw = container_of(work, typeof(*pw), base);
1984
1985	if (pw->trampoline)
1986		i915_active_release(&pw->trampoline->active);
1987	i915_active_release(&pw->shadow->active);
1988	i915_active_release(&pw->batch->active);
1989}
1990
1991static const struct dma_fence_work_ops eb_parse_ops = {
1992	.name = "eb_parse",
1993	.work = __eb_parse,
1994	.release = __eb_parse_release,
1995};
1996
1997static inline int
1998__parser_mark_active(struct i915_vma *vma,
1999		     struct intel_timeline *tl,
2000		     struct dma_fence *fence)
2001{
2002	struct intel_gt_buffer_pool_node *node = vma->private;
2003
2004	return i915_active_ref(&node->active, tl, fence);
2005}
2006
2007static int
2008parser_mark_active(struct eb_parse_work *pw, struct intel_timeline *tl)
2009{
2010	int err;
2011
2012	mutex_lock(&tl->mutex);
2013
2014	err = __parser_mark_active(pw->shadow, tl, &pw->base.dma);
2015	if (err)
2016		goto unlock;
2017
2018	if (pw->trampoline) {
2019		err = __parser_mark_active(pw->trampoline, tl, &pw->base.dma);
2020		if (err)
2021			goto unlock;
2022	}
2023
2024unlock:
2025	mutex_unlock(&tl->mutex);
2026	return err;
2027}
2028
2029static int eb_parse_pipeline(struct i915_execbuffer *eb,
2030			     struct i915_vma *shadow,
2031			     struct i915_vma *trampoline)
2032{
2033	struct eb_parse_work *pw;
2034	int err;
2035
2036	pw = kzalloc(sizeof(*pw), GFP_KERNEL);
2037	if (!pw)
2038		return -ENOMEM;
2039
2040	err = i915_active_acquire(&eb->batch->vma->active);
2041	if (err)
2042		goto err_free;
2043
2044	err = i915_active_acquire(&shadow->active);
2045	if (err)
2046		goto err_batch;
2047
2048	if (trampoline) {
2049		err = i915_active_acquire(&trampoline->active);
2050		if (err)
2051			goto err_shadow;
2052	}
2053
2054	dma_fence_work_init(&pw->base, &eb_parse_ops);
2055
2056	pw->engine = eb->engine;
2057	pw->batch = eb->batch->vma;
2058	pw->batch_offset = eb->batch_start_offset;
2059	pw->batch_length = eb->batch_len;
2060	pw->shadow = shadow;
2061	pw->trampoline = trampoline;
2062
2063	/* Mark active refs early for this worker, in case we get interrupted */
2064	err = parser_mark_active(pw, eb->context->timeline);
2065	if (err)
2066		goto err_commit;
2067
2068	err = dma_resv_lock_interruptible(pw->batch->resv, NULL);
2069	if (err)
2070		goto err_commit;
2071
2072	err = dma_resv_reserve_shared(pw->batch->resv, 1);
2073	if (err)
2074		goto err_commit_unlock;
2075
2076	/* Wait for all writes (and relocs) into the batch to complete */
2077	err = i915_sw_fence_await_reservation(&pw->base.chain,
2078					      pw->batch->resv, NULL, false,
2079					      0, I915_FENCE_GFP);
2080	if (err < 0)
2081		goto err_commit_unlock;
2082
2083	/* Keep the batch alive and unwritten as we parse */
2084	dma_resv_add_shared_fence(pw->batch->resv, &pw->base.dma);
2085
2086	dma_resv_unlock(pw->batch->resv);
2087
2088	/* Force execution to wait for completion of the parser */
2089	dma_resv_lock(shadow->resv, NULL);
2090	dma_resv_add_excl_fence(shadow->resv, &pw->base.dma);
2091	dma_resv_unlock(shadow->resv);
2092
2093	dma_fence_work_commit_imm(&pw->base);
2094	return 0;
2095
2096err_commit_unlock:
2097	dma_resv_unlock(pw->batch->resv);
2098err_commit:
2099	i915_sw_fence_set_error_once(&pw->base.chain, err);
2100	dma_fence_work_commit_imm(&pw->base);
2101	return err;
2102
2103err_shadow:
2104	i915_active_release(&shadow->active);
2105err_batch:
2106	i915_active_release(&eb->batch->vma->active);
2107err_free:
2108	kfree(pw);
2109	return err;
2110}
2111
2112static int eb_parse(struct i915_execbuffer *eb)
2113{
2114	struct drm_i915_private *i915 = eb->i915;
2115	struct intel_gt_buffer_pool_node *pool;
2116	struct i915_vma *shadow, *trampoline;
2117	unsigned int len;
2118	int err;
2119
2120	if (!eb_use_cmdparser(eb))
2121		return 0;
 
 
 
 
 
2122
2123	len = eb->batch_len;
2124	if (!CMDPARSER_USES_GGTT(eb->i915)) {
2125		/*
2126		 * ppGTT backed shadow buffers must be mapped RO, to prevent
2127		 * post-scan tampering
2128		 */
2129		if (!eb->context->vm->has_read_only) {
2130			drm_dbg(&i915->drm,
2131				"Cannot prevent post-scan tampering without RO capable vm\n");
2132			return -EINVAL;
2133		}
2134	} else {
2135		len += I915_CMD_PARSER_TRAMPOLINE_SIZE;
2136	}
 
 
2137
2138	pool = intel_gt_get_buffer_pool(eb->engine->gt, len);
2139	if (IS_ERR(pool))
2140		return PTR_ERR(pool);
 
 
 
 
2141
2142	shadow = shadow_batch_pin(pool->obj, eb->context->vm, PIN_USER);
 
 
 
 
2143	if (IS_ERR(shadow)) {
2144		err = PTR_ERR(shadow);
2145		goto err;
2146	}
 
2147	i915_gem_object_set_readonly(shadow->obj);
2148	shadow->private = pool;
2149
2150	trampoline = NULL;
2151	if (CMDPARSER_USES_GGTT(eb->i915)) {
2152		trampoline = shadow;
2153
2154		shadow = shadow_batch_pin(pool->obj,
2155					  &eb->engine->gt->ggtt->vm,
2156					  PIN_GLOBAL);
2157		if (IS_ERR(shadow)) {
2158			err = PTR_ERR(shadow);
2159			shadow = trampoline;
2160			goto err_shadow;
2161		}
2162		shadow->private = pool;
2163
2164		eb->batch_flags |= I915_DISPATCH_SECURE;
2165	}
2166
2167	err = eb_parse_pipeline(eb, shadow, trampoline);
 
 
 
 
 
 
2168	if (err)
2169		goto err_trampoline;
2170
2171	eb->vma[eb->buffer_count].vma = i915_vma_get(shadow);
2172	eb->vma[eb->buffer_count].flags = __EXEC_OBJECT_HAS_PIN;
 
 
 
 
 
 
2173	eb->batch = &eb->vma[eb->buffer_count++];
2174	eb->vma[eb->buffer_count].vma = NULL;
 
2175
2176	eb->trampoline = trampoline;
2177	eb->batch_start_offset = 0;
2178
 
 
 
 
 
 
2179	return 0;
2180
 
 
 
2181err_trampoline:
2182	if (trampoline)
2183		i915_vma_unpin(trampoline);
2184err_shadow:
2185	i915_vma_unpin(shadow);
2186err:
2187	intel_gt_buffer_pool_put(pool);
2188	return err;
2189}
2190
2191static void
2192add_to_client(struct i915_request *rq, struct drm_file *file)
2193{
2194	struct drm_i915_file_private *file_priv = file->driver_priv;
2195
2196	rq->file_priv = file_priv;
2197
2198	spin_lock(&file_priv->mm.lock);
2199	list_add_tail(&rq->client_link, &file_priv->mm.request_list);
2200	spin_unlock(&file_priv->mm.lock);
2201}
2202
2203static int eb_submit(struct i915_execbuffer *eb, struct i915_vma *batch)
2204{
2205	int err;
2206
 
 
 
2207	err = eb_move_to_gpu(eb);
2208	if (err)
2209		return err;
2210
2211	if (eb->args->flags & I915_EXEC_GEN7_SOL_RESET) {
2212		err = i915_reset_gen7_sol_offsets(eb->request);
2213		if (err)
2214			return err;
2215	}
2216
2217	/*
2218	 * After we completed waiting for other engines (using HW semaphores)
2219	 * then we can signal that this request/batch is ready to run. This
2220	 * allows us to determine if the batch is still waiting on the GPU
2221	 * or actually running by checking the breadcrumb.
2222	 */
2223	if (eb->engine->emit_init_breadcrumb) {
2224		err = eb->engine->emit_init_breadcrumb(eb->request);
2225		if (err)
2226			return err;
2227	}
2228
2229	err = eb->engine->emit_bb_start(eb->request,
2230					batch->node.start +
2231					eb->batch_start_offset,
2232					eb->batch_len,
2233					eb->batch_flags);
2234	if (err)
2235		return err;
2236
2237	if (eb->trampoline) {
2238		GEM_BUG_ON(eb->batch_start_offset);
2239		err = eb->engine->emit_bb_start(eb->request,
2240						eb->trampoline->node.start +
2241						eb->batch_len,
2242						0, 0);
2243		if (err)
2244			return err;
2245	}
2246
2247	if (intel_context_nopreempt(eb->context))
2248		__set_bit(I915_FENCE_FLAG_NOPREEMPT, &eb->request->fence.flags);
2249
2250	return 0;
2251}
2252
2253static int num_vcs_engines(const struct drm_i915_private *i915)
2254{
2255	return hweight64(VDBOX_MASK(&i915->gt));
2256}
2257
2258/*
2259 * Find one BSD ring to dispatch the corresponding BSD command.
2260 * The engine index is returned.
2261 */
2262static unsigned int
2263gen8_dispatch_bsd_engine(struct drm_i915_private *dev_priv,
2264			 struct drm_file *file)
2265{
2266	struct drm_i915_file_private *file_priv = file->driver_priv;
2267
2268	/* Check whether the file_priv has already selected one ring. */
2269	if ((int)file_priv->bsd_engine < 0)
2270		file_priv->bsd_engine =
2271			get_random_int() % num_vcs_engines(dev_priv);
2272
2273	return file_priv->bsd_engine;
2274}
2275
2276static const enum intel_engine_id user_ring_map[] = {
2277	[I915_EXEC_DEFAULT]	= RCS0,
2278	[I915_EXEC_RENDER]	= RCS0,
2279	[I915_EXEC_BLT]		= BCS0,
2280	[I915_EXEC_BSD]		= VCS0,
2281	[I915_EXEC_VEBOX]	= VECS0
2282};
2283
2284static struct i915_request *eb_throttle(struct intel_context *ce)
2285{
2286	struct intel_ring *ring = ce->ring;
2287	struct intel_timeline *tl = ce->timeline;
2288	struct i915_request *rq;
2289
2290	/*
2291	 * Completely unscientific finger-in-the-air estimates for suitable
2292	 * maximum user request size (to avoid blocking) and then backoff.
2293	 */
2294	if (intel_ring_update_space(ring) >= PAGE_SIZE)
2295		return NULL;
2296
2297	/*
2298	 * Find a request that after waiting upon, there will be at least half
2299	 * the ring available. The hysteresis allows us to compete for the
2300	 * shared ring and should mean that we sleep less often prior to
2301	 * claiming our resources, but not so long that the ring completely
2302	 * drains before we can submit our next request.
2303	 */
2304	list_for_each_entry(rq, &tl->requests, link) {
2305		if (rq->ring != ring)
2306			continue;
2307
2308		if (__intel_ring_space(rq->postfix,
2309				       ring->emit, ring->size) > ring->size / 2)
2310			break;
2311	}
2312	if (&rq->link == &tl->requests)
2313		return NULL; /* weird, we will check again later for real */
2314
2315	return i915_request_get(rq);
2316}
2317
2318static int __eb_pin_engine(struct i915_execbuffer *eb, struct intel_context *ce)
2319{
 
2320	struct intel_timeline *tl;
2321	struct i915_request *rq;
2322	int err;
2323
2324	/*
2325	 * ABI: Before userspace accesses the GPU (e.g. execbuffer), report
2326	 * EIO if the GPU is already wedged.
2327	 */
2328	err = intel_gt_terminally_wedged(ce->engine->gt);
2329	if (err)
2330		return err;
2331
2332	if (unlikely(intel_context_is_banned(ce)))
2333		return -EIO;
2334
2335	/*
2336	 * Pinning the contexts may generate requests in order to acquire
2337	 * GGTT space, so do this first before we reserve a seqno for
2338	 * ourselves.
2339	 */
2340	err = intel_context_pin(ce);
2341	if (err)
2342		return err;
2343
2344	/*
2345	 * Take a local wakeref for preparing to dispatch the execbuf as
2346	 * we expect to access the hardware fairly frequently in the
2347	 * process, and require the engine to be kept awake between accesses.
2348	 * Upon dispatch, we acquire another prolonged wakeref that we hold
2349	 * until the timeline is idle, which in turn releases the wakeref
2350	 * taken on the engine, and the parent device.
2351	 */
2352	tl = intel_context_timeline_lock(ce);
2353	if (IS_ERR(tl)) {
2354		err = PTR_ERR(tl);
2355		goto err_unpin;
2356	}
2357
2358	intel_context_enter(ce);
2359	rq = eb_throttle(ce);
2360
2361	intel_context_timeline_unlock(tl);
2362
2363	if (rq) {
2364		bool nonblock = eb->file->filp->f_flags & O_NONBLOCK;
2365		long timeout;
2366
2367		timeout = MAX_SCHEDULE_TIMEOUT;
2368		if (nonblock)
2369			timeout = 0;
2370
2371		timeout = i915_request_wait(rq,
2372					    I915_WAIT_INTERRUPTIBLE,
2373					    timeout);
2374		i915_request_put(rq);
2375
2376		if (timeout < 0) {
2377			err = nonblock ? -EWOULDBLOCK : timeout;
2378			goto err_exit;
2379		}
2380	}
2381
2382	eb->engine = ce->engine;
2383	eb->context = ce;
2384	return 0;
2385
2386err_exit:
2387	mutex_lock(&tl->mutex);
2388	intel_context_exit(ce);
2389	intel_context_timeline_unlock(tl);
2390err_unpin:
2391	intel_context_unpin(ce);
2392	return err;
2393}
2394
2395static void eb_unpin_engine(struct i915_execbuffer *eb)
2396{
2397	struct intel_context *ce = eb->context;
2398	struct intel_timeline *tl = ce->timeline;
2399
 
 
 
 
 
2400	mutex_lock(&tl->mutex);
2401	intel_context_exit(ce);
2402	mutex_unlock(&tl->mutex);
2403
2404	intel_context_unpin(ce);
2405}
2406
2407static unsigned int
2408eb_select_legacy_ring(struct i915_execbuffer *eb,
2409		      struct drm_file *file,
2410		      struct drm_i915_gem_execbuffer2 *args)
2411{
2412	struct drm_i915_private *i915 = eb->i915;
 
2413	unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK;
2414
2415	if (user_ring_id != I915_EXEC_BSD &&
2416	    (args->flags & I915_EXEC_BSD_MASK)) {
2417		drm_dbg(&i915->drm,
2418			"execbuf with non bsd ring but with invalid "
2419			"bsd dispatch flags: %d\n", (int)(args->flags));
2420		return -1;
2421	}
2422
2423	if (user_ring_id == I915_EXEC_BSD && num_vcs_engines(i915) > 1) {
2424		unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK;
2425
2426		if (bsd_idx == I915_EXEC_BSD_DEFAULT) {
2427			bsd_idx = gen8_dispatch_bsd_engine(i915, file);
2428		} else if (bsd_idx >= I915_EXEC_BSD_RING1 &&
2429			   bsd_idx <= I915_EXEC_BSD_RING2) {
2430			bsd_idx >>= I915_EXEC_BSD_SHIFT;
2431			bsd_idx--;
2432		} else {
2433			drm_dbg(&i915->drm,
2434				"execbuf with unknown bsd ring: %u\n",
2435				bsd_idx);
2436			return -1;
2437		}
2438
2439		return _VCS(bsd_idx);
2440	}
2441
2442	if (user_ring_id >= ARRAY_SIZE(user_ring_map)) {
2443		drm_dbg(&i915->drm, "execbuf with unknown ring: %u\n",
2444			user_ring_id);
2445		return -1;
2446	}
2447
2448	return user_ring_map[user_ring_id];
2449}
2450
2451static int
2452eb_pin_engine(struct i915_execbuffer *eb,
2453	      struct drm_file *file,
2454	      struct drm_i915_gem_execbuffer2 *args)
2455{
2456	struct intel_context *ce;
2457	unsigned int idx;
2458	int err;
2459
2460	if (i915_gem_context_user_engines(eb->gem_context))
2461		idx = args->flags & I915_EXEC_RING_MASK;
2462	else
2463		idx = eb_select_legacy_ring(eb, file, args);
2464
2465	ce = i915_gem_context_get_engine(eb->gem_context, idx);
2466	if (IS_ERR(ce))
2467		return PTR_ERR(ce);
2468
2469	err = __eb_pin_engine(eb, ce);
2470	intel_context_put(ce);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2471
 
 
 
2472	return err;
2473}
2474
2475static void
2476__free_fence_array(struct drm_syncobj **fences, unsigned int n)
2477{
2478	while (n--)
2479		drm_syncobj_put(ptr_mask_bits(fences[n], 2));
 
 
 
 
 
 
 
 
 
 
2480	kvfree(fences);
2481}
2482
2483static struct drm_syncobj **
2484get_fence_array(struct drm_i915_gem_execbuffer2 *args,
2485		struct drm_file *file)
2486{
2487	const unsigned long nfences = args->num_cliprects;
2488	struct drm_i915_gem_exec_fence __user *user;
2489	struct drm_syncobj **fences;
2490	unsigned long n;
2491	int err;
2492
2493	if (!(args->flags & I915_EXEC_FENCE_ARRAY))
2494		return NULL;
 
2495
2496	/* Check multiplication overflow for access_ok() and kvmalloc_array() */
2497	BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long));
2498	if (nfences > min_t(unsigned long,
2499			    ULONG_MAX / sizeof(*user),
2500			    SIZE_MAX / sizeof(*fences)))
2501		return ERR_PTR(-EINVAL);
2502
2503	user = u64_to_user_ptr(args->cliprects_ptr);
2504	if (!access_ok(user, nfences * sizeof(*user)))
2505		return ERR_PTR(-EFAULT);
 
 
 
 
 
 
 
 
 
 
2506
2507	fences = kvmalloc_array(nfences, sizeof(*fences),
2508				__GFP_NOWARN | GFP_KERNEL);
2509	if (!fences)
2510		return ERR_PTR(-ENOMEM);
2511
2512	for (n = 0; n < nfences; n++) {
2513		struct drm_i915_gem_exec_fence fence;
 
 
 
2514		struct drm_syncobj *syncobj;
 
 
2515
2516		if (__copy_from_user(&fence, user++, sizeof(fence))) {
2517			err = -EFAULT;
2518			goto err;
 
 
 
 
 
 
 
 
 
 
 
 
2519		}
2520
2521		if (fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS) {
2522			err = -EINVAL;
2523			goto err;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2524		}
2525
2526		syncobj = drm_syncobj_find(file, fence.handle);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2527		if (!syncobj) {
2528			DRM_DEBUG("Invalid syncobj handle provided\n");
2529			err = -ENOENT;
2530			goto err;
 
 
 
 
 
 
 
 
2531		}
2532
2533		BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) &
2534			     ~__I915_EXEC_FENCE_UNKNOWN_FLAGS);
2535
2536		fences[n] = ptr_pack_bits(syncobj, fence.flags, 2);
 
 
 
 
 
2537	}
2538
2539	return fences;
2540
2541err:
2542	__free_fence_array(fences, n);
2543	return ERR_PTR(err);
2544}
2545
2546static void
2547put_fence_array(struct drm_i915_gem_execbuffer2 *args,
2548		struct drm_syncobj **fences)
2549{
2550	if (fences)
2551		__free_fence_array(fences, args->num_cliprects);
2552}
2553
2554static int
2555await_fence_array(struct i915_execbuffer *eb,
2556		  struct drm_syncobj **fences)
2557{
2558	const unsigned int nfences = eb->args->num_cliprects;
2559	unsigned int n;
2560	int err;
2561
2562	for (n = 0; n < nfences; n++) {
2563		struct drm_syncobj *syncobj;
2564		struct dma_fence *fence;
2565		unsigned int flags;
2566
2567		syncobj = ptr_unpack_bits(fences[n], &flags, 2);
2568		if (!(flags & I915_EXEC_FENCE_WAIT))
2569			continue;
2570
2571		fence = drm_syncobj_fence_get(syncobj);
2572		if (!fence)
2573			return -EINVAL;
2574
2575		err = i915_request_await_dma_fence(eb->request, fence);
2576		dma_fence_put(fence);
2577		if (err < 0)
2578			return err;
2579	}
2580
2581	return 0;
2582}
2583
2584static void
2585signal_fence_array(struct i915_execbuffer *eb,
2586		   struct drm_syncobj **fences)
2587{
2588	const unsigned int nfences = eb->args->num_cliprects;
2589	struct dma_fence * const fence = &eb->request->fence;
2590	unsigned int n;
2591
2592	for (n = 0; n < nfences; n++) {
2593		struct drm_syncobj *syncobj;
2594		unsigned int flags;
2595
2596		syncobj = ptr_unpack_bits(fences[n], &flags, 2);
2597		if (!(flags & I915_EXEC_FENCE_SIGNAL))
2598			continue;
2599
2600		drm_syncobj_replace_fence(syncobj, fence);
 
 
 
 
 
 
 
 
 
 
 
 
2601	}
2602}
2603
 
 
 
 
 
 
 
 
 
 
 
 
2604static void retire_requests(struct intel_timeline *tl, struct i915_request *end)
2605{
2606	struct i915_request *rq, *rn;
2607
2608	list_for_each_entry_safe(rq, rn, &tl->requests, link)
2609		if (rq == end || !i915_request_retire(rq))
2610			break;
2611}
2612
2613static void eb_request_add(struct i915_execbuffer *eb)
2614{
2615	struct i915_request *rq = eb->request;
2616	struct intel_timeline * const tl = i915_request_timeline(rq);
2617	struct i915_sched_attr attr = {};
2618	struct i915_request *prev;
2619
2620	lockdep_assert_held(&tl->mutex);
2621	lockdep_unpin_lock(&tl->mutex, rq->cookie);
2622
2623	trace_i915_request_add(rq);
2624
2625	prev = __i915_request_commit(rq);
2626
2627	/* Check that the context wasn't destroyed before submission */
2628	if (likely(!intel_context_is_closed(eb->context))) {
2629		attr = eb->gem_context->sched;
2630	} else {
2631		/* Serialise with context_close via the add_to_timeline */
2632		i915_request_set_error_once(rq, -ENOENT);
2633		__i915_request_skip(rq);
 
2634	}
2635
2636	__i915_request_queue(rq, &attr);
2637
2638	/* Try to clean up the client's timeline after submitting the request */
2639	if (prev)
2640		retire_requests(tl, prev);
2641
2642	mutex_unlock(&tl->mutex);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2643}
2644
2645static int
2646i915_gem_do_execbuffer(struct drm_device *dev,
2647		       struct drm_file *file,
2648		       struct drm_i915_gem_execbuffer2 *args,
2649		       struct drm_i915_gem_exec_object2 *exec,
2650		       struct drm_syncobj **fences)
2651{
2652	struct drm_i915_private *i915 = to_i915(dev);
2653	struct i915_execbuffer eb;
2654	struct dma_fence *in_fence = NULL;
2655	struct sync_file *out_fence = NULL;
2656	struct i915_vma *batch;
2657	int out_fence_fd = -1;
2658	int err;
2659
2660	BUILD_BUG_ON(__EXEC_INTERNAL_FLAGS & ~__I915_EXEC_ILLEGAL_FLAGS);
2661	BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS &
2662		     ~__EXEC_OBJECT_UNKNOWN_FLAGS);
2663
2664	eb.i915 = i915;
2665	eb.file = file;
2666	eb.args = args;
2667	if (DBG_FORCE_RELOC || !(args->flags & I915_EXEC_NO_RELOC))
2668		args->flags |= __EXEC_HAS_RELOC;
2669
2670	eb.exec = exec;
 
 
 
 
2671
2672	eb.invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
2673	reloc_cache_init(&eb.reloc_cache, eb.i915);
2674
2675	eb.buffer_count = args->buffer_count;
2676	eb.batch_start_offset = args->batch_start_offset;
2677	eb.batch_len = args->batch_len;
2678	eb.trampoline = NULL;
2679
 
 
 
2680	eb.batch_flags = 0;
2681	if (args->flags & I915_EXEC_SECURE) {
2682		if (INTEL_GEN(i915) >= 11)
2683			return -ENODEV;
2684
2685		/* Return -EPERM to trigger fallback code on old binaries. */
2686		if (!HAS_SECURE_BATCHES(i915))
2687			return -EPERM;
2688
2689		if (!drm_is_current_master(file) || !capable(CAP_SYS_ADMIN))
2690			return -EPERM;
2691
2692		eb.batch_flags |= I915_DISPATCH_SECURE;
2693	}
2694	if (args->flags & I915_EXEC_IS_PINNED)
2695		eb.batch_flags |= I915_DISPATCH_PINNED;
2696
 
 
 
 
 
 
 
 
2697#define IN_FENCES (I915_EXEC_FENCE_IN | I915_EXEC_FENCE_SUBMIT)
2698	if (args->flags & IN_FENCES) {
2699		if ((args->flags & IN_FENCES) == IN_FENCES)
2700			return -EINVAL;
2701
2702		in_fence = sync_file_get_fence(lower_32_bits(args->rsvd2));
2703		if (!in_fence)
2704			return -EINVAL;
 
 
2705	}
2706#undef IN_FENCES
2707
2708	if (args->flags & I915_EXEC_FENCE_OUT) {
2709		out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
2710		if (out_fence_fd < 0) {
2711			err = out_fence_fd;
2712			goto err_in_fence;
2713		}
2714	}
2715
2716	err = eb_create(&eb);
2717	if (err)
2718		goto err_out_fence;
2719
2720	GEM_BUG_ON(!eb.lut_size);
2721
2722	err = eb_select_context(&eb);
2723	if (unlikely(err))
2724		goto err_destroy;
2725
2726	err = eb_pin_engine(&eb, file, args);
2727	if (unlikely(err))
2728		goto err_context;
2729
2730	err = eb_relocate(&eb);
 
 
 
 
 
 
 
 
2731	if (err) {
2732		/*
2733		 * If the user expects the execobject.offset and
2734		 * reloc.presumed_offset to be an exact match,
2735		 * as for using NO_RELOC, then we cannot update
2736		 * the execobject.offset until we have completed
2737		 * relocation.
2738		 */
2739		args->flags &= ~__EXEC_HAS_RELOC;
2740		goto err_vma;
2741	}
2742
2743	if (unlikely(eb.batch->flags & EXEC_OBJECT_WRITE)) {
2744		drm_dbg(&i915->drm,
2745			"Attempting to use self-modifying batch buffer\n");
2746		err = -EINVAL;
2747		goto err_vma;
2748	}
2749
2750	if (range_overflows_t(u64,
2751			      eb.batch_start_offset, eb.batch_len,
2752			      eb.batch->vma->size)) {
2753		drm_dbg(&i915->drm, "Attempting to use out-of-bounds batch\n");
2754		err = -EINVAL;
2755		goto err_vma;
2756	}
2757
2758	if (eb.batch_len == 0)
2759		eb.batch_len = eb.batch->vma->size - eb.batch_start_offset;
2760
2761	err = eb_parse(&eb);
2762	if (err)
2763		goto err_vma;
2764
2765	/*
2766	 * snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
2767	 * batch" bit. Hence we need to pin secure batches into the global gtt.
2768	 * hsw should have this fixed, but bdw mucks it up again. */
2769	batch = eb.batch->vma;
2770	if (eb.batch_flags & I915_DISPATCH_SECURE) {
2771		struct i915_vma *vma;
2772
2773		/*
2774		 * So on first glance it looks freaky that we pin the batch here
2775		 * outside of the reservation loop. But:
2776		 * - The batch is already pinned into the relevant ppgtt, so we
2777		 *   already have the backing storage fully allocated.
2778		 * - No other BO uses the global gtt (well contexts, but meh),
2779		 *   so we don't really have issues with multiple objects not
2780		 *   fitting due to fragmentation.
2781		 * So this is actually safe.
2782		 */
2783		vma = i915_gem_object_ggtt_pin(batch->obj, NULL, 0, 0, 0);
2784		if (IS_ERR(vma)) {
2785			err = PTR_ERR(vma);
2786			goto err_parse;
2787		}
2788
2789		batch = vma;
2790	}
2791
2792	/* All GPU relocation batches must be submitted prior to the user rq */
2793	GEM_BUG_ON(eb.reloc_cache.rq);
2794
2795	/* Allocate a request for this batch buffer nice and early. */
2796	eb.request = i915_request_create(eb.context);
2797	if (IS_ERR(eb.request)) {
2798		err = PTR_ERR(eb.request);
2799		goto err_batch_unpin;
2800	}
2801
2802	if (in_fence) {
2803		if (args->flags & I915_EXEC_FENCE_SUBMIT)
2804			err = i915_request_await_execution(eb.request,
2805							   in_fence,
2806							   eb.engine->bond_execute);
2807		else
2808			err = i915_request_await_dma_fence(eb.request,
2809							   in_fence);
2810		if (err < 0)
2811			goto err_request;
2812	}
2813
2814	if (fences) {
2815		err = await_fence_array(&eb, fences);
2816		if (err)
2817			goto err_request;
2818	}
2819
2820	if (out_fence_fd != -1) {
2821		out_fence = sync_file_create(&eb.request->fence);
2822		if (!out_fence) {
2823			err = -ENOMEM;
2824			goto err_request;
2825		}
2826	}
2827
2828	/*
2829	 * Whilst this request exists, batch_obj will be on the
2830	 * active_list, and so will hold the active reference. Only when this
2831	 * request is retired will the the batch_obj be moved onto the
2832	 * inactive_list and lose its active reference. Hence we do not need
2833	 * to explicitly hold another reference here.
2834	 */
2835	eb.request->batch = batch;
2836	if (batch->private)
2837		intel_gt_buffer_pool_mark_active(batch->private, eb.request);
2838
2839	trace_i915_request_queue(eb.request, eb.batch_flags);
2840	err = eb_submit(&eb, batch);
 
2841err_request:
2842	add_to_client(eb.request, file);
2843	i915_request_get(eb.request);
2844	eb_request_add(&eb);
2845
2846	if (fences)
2847		signal_fence_array(&eb, fences);
2848
2849	if (out_fence) {
2850		if (err == 0) {
2851			fd_install(out_fence_fd, out_fence->file);
2852			args->rsvd2 &= GENMASK_ULL(31, 0); /* keep in-fence */
2853			args->rsvd2 |= (u64)out_fence_fd << 32;
2854			out_fence_fd = -1;
2855		} else {
2856			fput(out_fence->file);
2857		}
2858	}
2859	i915_request_put(eb.request);
2860
2861err_batch_unpin:
2862	if (eb.batch_flags & I915_DISPATCH_SECURE)
2863		i915_vma_unpin(batch);
2864err_parse:
2865	if (batch->private)
2866		intel_gt_buffer_pool_put(batch->private);
2867err_vma:
 
2868	if (eb.trampoline)
2869		i915_vma_unpin(eb.trampoline);
2870	eb_unpin_engine(&eb);
 
 
 
 
 
 
 
 
 
 
2871err_context:
2872	i915_gem_context_put(eb.gem_context);
2873err_destroy:
2874	eb_destroy(&eb);
2875err_out_fence:
2876	if (out_fence_fd != -1)
2877		put_unused_fd(out_fence_fd);
2878err_in_fence:
2879	dma_fence_put(in_fence);
 
 
2880	return err;
2881}
2882
2883static size_t eb_element_size(void)
2884{
2885	return sizeof(struct drm_i915_gem_exec_object2);
2886}
2887
2888static bool check_buffer_count(size_t count)
2889{
2890	const size_t sz = eb_element_size();
2891
2892	/*
2893	 * When using LUT_HANDLE, we impose a limit of INT_MAX for the lookup
2894	 * array size (see eb_create()). Otherwise, we can accept an array as
2895	 * large as can be addressed (though use large arrays at your peril)!
2896	 */
2897
2898	return !(count < 1 || count > INT_MAX || count > SIZE_MAX / sz - 1);
2899}
2900
2901/*
2902 * Legacy execbuffer just creates an exec2 list from the original exec object
2903 * list array and passes it to the real function.
2904 */
2905int
2906i915_gem_execbuffer_ioctl(struct drm_device *dev, void *data,
2907			  struct drm_file *file)
2908{
2909	struct drm_i915_private *i915 = to_i915(dev);
2910	struct drm_i915_gem_execbuffer *args = data;
2911	struct drm_i915_gem_execbuffer2 exec2;
2912	struct drm_i915_gem_exec_object *exec_list = NULL;
2913	struct drm_i915_gem_exec_object2 *exec2_list = NULL;
2914	const size_t count = args->buffer_count;
2915	unsigned int i;
2916	int err;
2917
2918	if (!check_buffer_count(count)) {
2919		drm_dbg(&i915->drm, "execbuf2 with %zd buffers\n", count);
2920		return -EINVAL;
2921	}
2922
2923	exec2.buffers_ptr = args->buffers_ptr;
2924	exec2.buffer_count = args->buffer_count;
2925	exec2.batch_start_offset = args->batch_start_offset;
2926	exec2.batch_len = args->batch_len;
2927	exec2.DR1 = args->DR1;
2928	exec2.DR4 = args->DR4;
2929	exec2.num_cliprects = args->num_cliprects;
2930	exec2.cliprects_ptr = args->cliprects_ptr;
2931	exec2.flags = I915_EXEC_RENDER;
2932	i915_execbuffer2_set_context_id(exec2, 0);
2933
2934	err = i915_gem_check_execbuffer(&exec2);
2935	if (err)
2936		return err;
2937
2938	/* Copy in the exec list from userland */
2939	exec_list = kvmalloc_array(count, sizeof(*exec_list),
2940				   __GFP_NOWARN | GFP_KERNEL);
2941	exec2_list = kvmalloc_array(count, eb_element_size(),
2942				    __GFP_NOWARN | GFP_KERNEL);
2943	if (exec_list == NULL || exec2_list == NULL) {
2944		drm_dbg(&i915->drm,
2945			"Failed to allocate exec list for %d buffers\n",
2946			args->buffer_count);
2947		kvfree(exec_list);
2948		kvfree(exec2_list);
2949		return -ENOMEM;
2950	}
2951	err = copy_from_user(exec_list,
2952			     u64_to_user_ptr(args->buffers_ptr),
2953			     sizeof(*exec_list) * count);
2954	if (err) {
2955		drm_dbg(&i915->drm, "copy %d exec entries failed %d\n",
2956			args->buffer_count, err);
2957		kvfree(exec_list);
2958		kvfree(exec2_list);
2959		return -EFAULT;
2960	}
2961
2962	for (i = 0; i < args->buffer_count; i++) {
2963		exec2_list[i].handle = exec_list[i].handle;
2964		exec2_list[i].relocation_count = exec_list[i].relocation_count;
2965		exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
2966		exec2_list[i].alignment = exec_list[i].alignment;
2967		exec2_list[i].offset = exec_list[i].offset;
2968		if (INTEL_GEN(to_i915(dev)) < 4)
2969			exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
2970		else
2971			exec2_list[i].flags = 0;
2972	}
2973
2974	err = i915_gem_do_execbuffer(dev, file, &exec2, exec2_list, NULL);
2975	if (exec2.flags & __EXEC_HAS_RELOC) {
2976		struct drm_i915_gem_exec_object __user *user_exec_list =
2977			u64_to_user_ptr(args->buffers_ptr);
2978
2979		/* Copy the new buffer offsets back to the user's exec list. */
2980		for (i = 0; i < args->buffer_count; i++) {
2981			if (!(exec2_list[i].offset & UPDATE))
2982				continue;
2983
2984			exec2_list[i].offset =
2985				gen8_canonical_addr(exec2_list[i].offset & PIN_OFFSET_MASK);
2986			exec2_list[i].offset &= PIN_OFFSET_MASK;
2987			if (__copy_to_user(&user_exec_list[i].offset,
2988					   &exec2_list[i].offset,
2989					   sizeof(user_exec_list[i].offset)))
2990				break;
2991		}
2992	}
2993
2994	kvfree(exec_list);
2995	kvfree(exec2_list);
2996	return err;
2997}
2998
2999int
3000i915_gem_execbuffer2_ioctl(struct drm_device *dev, void *data,
3001			   struct drm_file *file)
3002{
3003	struct drm_i915_private *i915 = to_i915(dev);
3004	struct drm_i915_gem_execbuffer2 *args = data;
3005	struct drm_i915_gem_exec_object2 *exec2_list;
3006	struct drm_syncobj **fences = NULL;
3007	const size_t count = args->buffer_count;
3008	int err;
3009
3010	if (!check_buffer_count(count)) {
3011		drm_dbg(&i915->drm, "execbuf2 with %zd buffers\n", count);
3012		return -EINVAL;
3013	}
3014
3015	err = i915_gem_check_execbuffer(args);
3016	if (err)
3017		return err;
3018
3019	exec2_list = kvmalloc_array(count, eb_element_size(),
 
3020				    __GFP_NOWARN | GFP_KERNEL);
3021	if (exec2_list == NULL) {
3022		drm_dbg(&i915->drm, "Failed to allocate exec list for %zd buffers\n",
3023			count);
3024		return -ENOMEM;
3025	}
3026	if (copy_from_user(exec2_list,
3027			   u64_to_user_ptr(args->buffers_ptr),
3028			   sizeof(*exec2_list) * count)) {
3029		drm_dbg(&i915->drm, "copy %zd exec entries failed\n", count);
3030		kvfree(exec2_list);
3031		return -EFAULT;
3032	}
3033
3034	if (args->flags & I915_EXEC_FENCE_ARRAY) {
3035		fences = get_fence_array(args, file);
3036		if (IS_ERR(fences)) {
3037			kvfree(exec2_list);
3038			return PTR_ERR(fences);
3039		}
3040	}
3041
3042	err = i915_gem_do_execbuffer(dev, file, args, exec2_list, fences);
3043
3044	/*
3045	 * Now that we have begun execution of the batchbuffer, we ignore
3046	 * any new error after this point. Also given that we have already
3047	 * updated the associated relocations, we try to write out the current
3048	 * object locations irrespective of any error.
3049	 */
3050	if (args->flags & __EXEC_HAS_RELOC) {
3051		struct drm_i915_gem_exec_object2 __user *user_exec_list =
3052			u64_to_user_ptr(args->buffers_ptr);
3053		unsigned int i;
3054
3055		/* Copy the new buffer offsets back to the user's exec list. */
3056		/*
3057		 * Note: count * sizeof(*user_exec_list) does not overflow,
3058		 * because we checked 'count' in check_buffer_count().
3059		 *
3060		 * And this range already got effectively checked earlier
3061		 * when we did the "copy_from_user()" above.
3062		 */
3063		if (!user_write_access_begin(user_exec_list,
3064					     count * sizeof(*user_exec_list)))
3065			goto end;
3066
3067		for (i = 0; i < args->buffer_count; i++) {
3068			if (!(exec2_list[i].offset & UPDATE))
3069				continue;
3070
3071			exec2_list[i].offset =
3072				gen8_canonical_addr(exec2_list[i].offset & PIN_OFFSET_MASK);
3073			unsafe_put_user(exec2_list[i].offset,
3074					&user_exec_list[i].offset,
3075					end_user);
3076		}
3077end_user:
3078		user_write_access_end();
3079end:;
3080	}
3081
3082	args->flags &= ~__I915_EXEC_UNKNOWN_FLAGS;
3083	put_fence_array(args, fences);
3084	kvfree(exec2_list);
3085	return err;
3086}
3087
3088#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
3089#include "selftests/i915_gem_execbuffer.c"
3090#endif
v5.14.15
   1/*
   2 * SPDX-License-Identifier: MIT
   3 *
   4 * Copyright © 2008,2010 Intel Corporation
   5 */
   6
   7#include <linux/intel-iommu.h>
   8#include <linux/dma-resv.h>
   9#include <linux/sync_file.h>
  10#include <linux/uaccess.h>
  11
  12#include <drm/drm_syncobj.h>
  13
  14#include "display/intel_frontbuffer.h"
  15
  16#include "gem/i915_gem_ioctls.h"
  17#include "gt/intel_context.h"
  18#include "gt/intel_gpu_commands.h"
  19#include "gt/intel_gt.h"
  20#include "gt/intel_gt_buffer_pool.h"
  21#include "gt/intel_gt_pm.h"
  22#include "gt/intel_ring.h"
  23
  24#include "i915_drv.h"
  25#include "i915_gem_clflush.h"
  26#include "i915_gem_context.h"
  27#include "i915_gem_ioctls.h"
 
  28#include "i915_trace.h"
  29#include "i915_user_extensions.h"
  30
  31struct eb_vma {
  32	struct i915_vma *vma;
  33	unsigned int flags;
  34
  35	/** This vma's place in the execbuf reservation list */
  36	struct drm_i915_gem_exec_object2 *exec;
  37	struct list_head bind_link;
  38	struct list_head reloc_link;
  39
  40	struct hlist_node node;
  41	u32 handle;
  42};
  43
 
 
 
 
 
  44enum {
  45	FORCE_CPU_RELOC = 1,
  46	FORCE_GTT_RELOC,
  47	FORCE_GPU_RELOC,
  48#define DBG_FORCE_RELOC 0 /* choose one of the above! */
  49};
  50
  51/* __EXEC_OBJECT_NO_RESERVE is BIT(31), defined in i915_vma.h */
  52#define __EXEC_OBJECT_HAS_PIN		BIT(30)
  53#define __EXEC_OBJECT_HAS_FENCE		BIT(29)
  54#define __EXEC_OBJECT_USERPTR_INIT	BIT(28)
  55#define __EXEC_OBJECT_NEEDS_MAP		BIT(27)
  56#define __EXEC_OBJECT_NEEDS_BIAS	BIT(26)
  57#define __EXEC_OBJECT_INTERNAL_FLAGS	(~0u << 26) /* all of the above + */
  58#define __EXEC_OBJECT_RESERVED (__EXEC_OBJECT_HAS_PIN | __EXEC_OBJECT_HAS_FENCE)
  59
  60#define __EXEC_HAS_RELOC	BIT(31)
  61#define __EXEC_ENGINE_PINNED	BIT(30)
  62#define __EXEC_USERPTR_USED	BIT(29)
  63#define __EXEC_INTERNAL_FLAGS	(~0u << 29)
  64#define UPDATE			PIN_OFFSET_FIXED
  65
  66#define BATCH_OFFSET_BIAS (256*1024)
  67
  68#define __I915_EXEC_ILLEGAL_FLAGS \
  69	(__I915_EXEC_UNKNOWN_FLAGS | \
  70	 I915_EXEC_CONSTANTS_MASK  | \
  71	 I915_EXEC_RESOURCE_STREAMER)
  72
  73/* Catch emission of unexpected errors for CI! */
  74#if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
  75#undef EINVAL
  76#define EINVAL ({ \
  77	DRM_DEBUG_DRIVER("EINVAL at %s:%d\n", __func__, __LINE__); \
  78	22; \
  79})
  80#endif
  81
  82/**
  83 * DOC: User command execution
  84 *
  85 * Userspace submits commands to be executed on the GPU as an instruction
  86 * stream within a GEM object we call a batchbuffer. This instructions may
  87 * refer to other GEM objects containing auxiliary state such as kernels,
  88 * samplers, render targets and even secondary batchbuffers. Userspace does
  89 * not know where in the GPU memory these objects reside and so before the
  90 * batchbuffer is passed to the GPU for execution, those addresses in the
  91 * batchbuffer and auxiliary objects are updated. This is known as relocation,
  92 * or patching. To try and avoid having to relocate each object on the next
  93 * execution, userspace is told the location of those objects in this pass,
  94 * but this remains just a hint as the kernel may choose a new location for
  95 * any object in the future.
  96 *
  97 * At the level of talking to the hardware, submitting a batchbuffer for the
  98 * GPU to execute is to add content to a buffer from which the HW
  99 * command streamer is reading.
 100 *
 101 * 1. Add a command to load the HW context. For Logical Ring Contexts, i.e.
 102 *    Execlists, this command is not placed on the same buffer as the
 103 *    remaining items.
 104 *
 105 * 2. Add a command to invalidate caches to the buffer.
 106 *
 107 * 3. Add a batchbuffer start command to the buffer; the start command is
 108 *    essentially a token together with the GPU address of the batchbuffer
 109 *    to be executed.
 110 *
 111 * 4. Add a pipeline flush to the buffer.
 112 *
 113 * 5. Add a memory write command to the buffer to record when the GPU
 114 *    is done executing the batchbuffer. The memory write writes the
 115 *    global sequence number of the request, ``i915_request::global_seqno``;
 116 *    the i915 driver uses the current value in the register to determine
 117 *    if the GPU has completed the batchbuffer.
 118 *
 119 * 6. Add a user interrupt command to the buffer. This command instructs
 120 *    the GPU to issue an interrupt when the command, pipeline flush and
 121 *    memory write are completed.
 122 *
 123 * 7. Inform the hardware of the additional commands added to the buffer
 124 *    (by updating the tail pointer).
 125 *
 126 * Processing an execbuf ioctl is conceptually split up into a few phases.
 127 *
 128 * 1. Validation - Ensure all the pointers, handles and flags are valid.
 129 * 2. Reservation - Assign GPU address space for every object
 130 * 3. Relocation - Update any addresses to point to the final locations
 131 * 4. Serialisation - Order the request with respect to its dependencies
 132 * 5. Construction - Construct a request to execute the batchbuffer
 133 * 6. Submission (at some point in the future execution)
 134 *
 135 * Reserving resources for the execbuf is the most complicated phase. We
 136 * neither want to have to migrate the object in the address space, nor do
 137 * we want to have to update any relocations pointing to this object. Ideally,
 138 * we want to leave the object where it is and for all the existing relocations
 139 * to match. If the object is given a new address, or if userspace thinks the
 140 * object is elsewhere, we have to parse all the relocation entries and update
 141 * the addresses. Userspace can set the I915_EXEC_NORELOC flag to hint that
 142 * all the target addresses in all of its objects match the value in the
 143 * relocation entries and that they all match the presumed offsets given by the
 144 * list of execbuffer objects. Using this knowledge, we know that if we haven't
 145 * moved any buffers, all the relocation entries are valid and we can skip
 146 * the update. (If userspace is wrong, the likely outcome is an impromptu GPU
 147 * hang.) The requirement for using I915_EXEC_NO_RELOC are:
 148 *
 149 *      The addresses written in the objects must match the corresponding
 150 *      reloc.presumed_offset which in turn must match the corresponding
 151 *      execobject.offset.
 152 *
 153 *      Any render targets written to in the batch must be flagged with
 154 *      EXEC_OBJECT_WRITE.
 155 *
 156 *      To avoid stalling, execobject.offset should match the current
 157 *      address of that object within the active context.
 158 *
 159 * The reservation is done is multiple phases. First we try and keep any
 160 * object already bound in its current location - so as long as meets the
 161 * constraints imposed by the new execbuffer. Any object left unbound after the
 162 * first pass is then fitted into any available idle space. If an object does
 163 * not fit, all objects are removed from the reservation and the process rerun
 164 * after sorting the objects into a priority order (more difficult to fit
 165 * objects are tried first). Failing that, the entire VM is cleared and we try
 166 * to fit the execbuf once last time before concluding that it simply will not
 167 * fit.
 168 *
 169 * A small complication to all of this is that we allow userspace not only to
 170 * specify an alignment and a size for the object in the address space, but
 171 * we also allow userspace to specify the exact offset. This objects are
 172 * simpler to place (the location is known a priori) all we have to do is make
 173 * sure the space is available.
 174 *
 175 * Once all the objects are in place, patching up the buried pointers to point
 176 * to the final locations is a fairly simple job of walking over the relocation
 177 * entry arrays, looking up the right address and rewriting the value into
 178 * the object. Simple! ... The relocation entries are stored in user memory
 179 * and so to access them we have to copy them into a local buffer. That copy
 180 * has to avoid taking any pagefaults as they may lead back to a GEM object
 181 * requiring the struct_mutex (i.e. recursive deadlock). So once again we split
 182 * the relocation into multiple passes. First we try to do everything within an
 183 * atomic context (avoid the pagefaults) which requires that we never wait. If
 184 * we detect that we may wait, or if we need to fault, then we have to fallback
 185 * to a slower path. The slowpath has to drop the mutex. (Can you hear alarm
 186 * bells yet?) Dropping the mutex means that we lose all the state we have
 187 * built up so far for the execbuf and we must reset any global data. However,
 188 * we do leave the objects pinned in their final locations - which is a
 189 * potential issue for concurrent execbufs. Once we have left the mutex, we can
 190 * allocate and copy all the relocation entries into a large array at our
 191 * leisure, reacquire the mutex, reclaim all the objects and other state and
 192 * then proceed to update any incorrect addresses with the objects.
 193 *
 194 * As we process the relocation entries, we maintain a record of whether the
 195 * object is being written to. Using NORELOC, we expect userspace to provide
 196 * this information instead. We also check whether we can skip the relocation
 197 * by comparing the expected value inside the relocation entry with the target's
 198 * final address. If they differ, we have to map the current object and rewrite
 199 * the 4 or 8 byte pointer within.
 200 *
 201 * Serialising an execbuf is quite simple according to the rules of the GEM
 202 * ABI. Execution within each context is ordered by the order of submission.
 203 * Writes to any GEM object are in order of submission and are exclusive. Reads
 204 * from a GEM object are unordered with respect to other reads, but ordered by
 205 * writes. A write submitted after a read cannot occur before the read, and
 206 * similarly any read submitted after a write cannot occur before the write.
 207 * Writes are ordered between engines such that only one write occurs at any
 208 * time (completing any reads beforehand) - using semaphores where available
 209 * and CPU serialisation otherwise. Other GEM access obey the same rules, any
 210 * write (either via mmaps using set-domain, or via pwrite) must flush all GPU
 211 * reads before starting, and any read (either using set-domain or pread) must
 212 * flush all GPU writes before starting. (Note we only employ a barrier before,
 213 * we currently rely on userspace not concurrently starting a new execution
 214 * whilst reading or writing to an object. This may be an advantage or not
 215 * depending on how much you trust userspace not to shoot themselves in the
 216 * foot.) Serialisation may just result in the request being inserted into
 217 * a DAG awaiting its turn, but most simple is to wait on the CPU until
 218 * all dependencies are resolved.
 219 *
 220 * After all of that, is just a matter of closing the request and handing it to
 221 * the hardware (well, leaving it in a queue to be executed). However, we also
 222 * offer the ability for batchbuffers to be run with elevated privileges so
 223 * that they access otherwise hidden registers. (Used to adjust L3 cache etc.)
 224 * Before any batch is given extra privileges we first must check that it
 225 * contains no nefarious instructions, we check that each instruction is from
 226 * our whitelist and all registers are also from an allowed list. We first
 227 * copy the user's batchbuffer to a shadow (so that the user doesn't have
 228 * access to it, either by the CPU or GPU as we scan it) and then parse each
 229 * instruction. If everything is ok, we set a flag telling the hardware to run
 230 * the batchbuffer in trusted mode, otherwise the ioctl is rejected.
 231 */
 232
 233struct eb_fence {
 234	struct drm_syncobj *syncobj; /* Use with ptr_mask_bits() */
 235	struct dma_fence *dma_fence;
 236	u64 value;
 237	struct dma_fence_chain *chain_fence;
 238};
 239
 240struct i915_execbuffer {
 241	struct drm_i915_private *i915; /** i915 backpointer */
 242	struct drm_file *file; /** per-file lookup tables and limits */
 243	struct drm_i915_gem_execbuffer2 *args; /** ioctl parameters */
 244	struct drm_i915_gem_exec_object2 *exec; /** ioctl execobj[] */
 245	struct eb_vma *vma;
 246
 247	struct intel_engine_cs *engine; /** engine to queue the request to */
 248	struct intel_context *context; /* logical state for the request */
 249	struct i915_gem_context *gem_context; /** caller's context */
 250
 251	struct i915_request *request; /** our request to build */
 252	struct eb_vma *batch; /** identity of the batch obj/vma */
 253	struct i915_vma *trampoline; /** trampoline used for chaining */
 254
 255	/** actual size of execobj[] as we may extend it for the cmdparser */
 256	unsigned int buffer_count;
 257
 258	/** list of vma not yet bound during reservation phase */
 259	struct list_head unbound;
 260
 261	/** list of vma that have execobj.relocation_count */
 262	struct list_head relocs;
 263
 264	struct i915_gem_ww_ctx ww;
 265
 266	/**
 267	 * Track the most recently used object for relocations, as we
 268	 * frequently have to perform multiple relocations within the same
 269	 * obj/page
 270	 */
 271	struct reloc_cache {
 272		struct drm_mm_node node; /** temporary GTT binding */
 273		unsigned long vaddr; /** Current kmap address */
 274		unsigned long page; /** Currently mapped page index */
 275		unsigned int graphics_ver; /** Cached value of GRAPHICS_VER */
 276		bool use_64bit_reloc : 1;
 277		bool has_llc : 1;
 278		bool has_fence : 1;
 279		bool needs_unfenced : 1;
 280
 
 281		struct i915_request *rq;
 
 282		u32 *rq_cmd;
 283		unsigned int rq_size;
 284		struct intel_gt_buffer_pool_node *pool;
 285	} reloc_cache;
 286
 287	struct intel_gt_buffer_pool_node *reloc_pool; /** relocation pool for -EDEADLK handling */
 288	struct intel_context *reloc_context;
 289
 290	u64 invalid_flags; /** Set of execobj.flags that are invalid */
 291	u32 context_flags; /** Set of execobj.flags to insert from the ctx */
 292
 293	u64 batch_len; /** Length of batch within object */
 294	u32 batch_start_offset; /** Location within object of batch */
 
 295	u32 batch_flags; /** Flags composed for emit_bb_start() */
 296	struct intel_gt_buffer_pool_node *batch_pool; /** pool node for batch buffer */
 297
 298	/**
 299	 * Indicate either the size of the hastable used to resolve
 300	 * relocation handles, or if negative that we are using a direct
 301	 * index into the execobj[].
 302	 */
 303	int lut_size;
 304	struct hlist_head *buckets; /** ht for relocation handles */
 305
 306	struct eb_fence *fences;
 307	unsigned long num_fences;
 308};
 309
 310static int eb_parse(struct i915_execbuffer *eb);
 311static struct i915_request *eb_pin_engine(struct i915_execbuffer *eb,
 312					  bool throttle);
 313static void eb_unpin_engine(struct i915_execbuffer *eb);
 314
 315static inline bool eb_use_cmdparser(const struct i915_execbuffer *eb)
 316{
 317	return intel_engine_requires_cmd_parser(eb->engine) ||
 318		(intel_engine_using_cmd_parser(eb->engine) &&
 319		 eb->args->batch_len);
 320}
 321
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 322static int eb_create(struct i915_execbuffer *eb)
 323{
 
 
 
 
 
 
 
 324	if (!(eb->args->flags & I915_EXEC_HANDLE_LUT)) {
 325		unsigned int size = 1 + ilog2(eb->buffer_count);
 326
 327		/*
 328		 * Without a 1:1 association between relocation handles and
 329		 * the execobject[] index, we instead create a hashtable.
 330		 * We size it dynamically based on available memory, starting
 331		 * first with 1:1 assocative hash and scaling back until
 332		 * the allocation succeeds.
 333		 *
 334		 * Later on we use a positive lut_size to indicate we are
 335		 * using this hashtable, and a negative value to indicate a
 336		 * direct lookup.
 337		 */
 338		do {
 339			gfp_t flags;
 340
 341			/* While we can still reduce the allocation size, don't
 342			 * raise a warning and allow the allocation to fail.
 343			 * On the last pass though, we want to try as hard
 344			 * as possible to perform the allocation and warn
 345			 * if it fails.
 346			 */
 347			flags = GFP_KERNEL;
 348			if (size > 1)
 349				flags |= __GFP_NORETRY | __GFP_NOWARN;
 350
 351			eb->buckets = kzalloc(sizeof(struct hlist_head) << size,
 352					      flags);
 353			if (eb->buckets)
 354				break;
 355		} while (--size);
 356
 357		if (unlikely(!size))
 
 358			return -ENOMEM;
 
 359
 360		eb->lut_size = size;
 361	} else {
 362		eb->lut_size = -eb->buffer_count;
 363	}
 364
 365	return 0;
 366}
 367
 368static bool
 369eb_vma_misplaced(const struct drm_i915_gem_exec_object2 *entry,
 370		 const struct i915_vma *vma,
 371		 unsigned int flags)
 372{
 373	if (vma->node.size < entry->pad_to_size)
 374		return true;
 375
 376	if (entry->alignment && !IS_ALIGNED(vma->node.start, entry->alignment))
 377		return true;
 378
 379	if (flags & EXEC_OBJECT_PINNED &&
 380	    vma->node.start != entry->offset)
 381		return true;
 382
 383	if (flags & __EXEC_OBJECT_NEEDS_BIAS &&
 384	    vma->node.start < BATCH_OFFSET_BIAS)
 385		return true;
 386
 387	if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) &&
 388	    (vma->node.start + vma->node.size + 4095) >> 32)
 389		return true;
 390
 391	if (flags & __EXEC_OBJECT_NEEDS_MAP &&
 392	    !i915_vma_is_map_and_fenceable(vma))
 393		return true;
 394
 395	return false;
 396}
 397
 398static u64 eb_pin_flags(const struct drm_i915_gem_exec_object2 *entry,
 399			unsigned int exec_flags)
 400{
 401	u64 pin_flags = 0;
 402
 403	if (exec_flags & EXEC_OBJECT_NEEDS_GTT)
 404		pin_flags |= PIN_GLOBAL;
 405
 406	/*
 407	 * Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset,
 408	 * limit address to the first 4GBs for unflagged objects.
 409	 */
 410	if (!(exec_flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS))
 411		pin_flags |= PIN_ZONE_4G;
 412
 413	if (exec_flags & __EXEC_OBJECT_NEEDS_MAP)
 414		pin_flags |= PIN_MAPPABLE;
 415
 416	if (exec_flags & EXEC_OBJECT_PINNED)
 417		pin_flags |= entry->offset | PIN_OFFSET_FIXED;
 418	else if (exec_flags & __EXEC_OBJECT_NEEDS_BIAS)
 419		pin_flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
 420
 421	return pin_flags;
 422}
 423
 424static inline int
 425eb_pin_vma(struct i915_execbuffer *eb,
 426	   const struct drm_i915_gem_exec_object2 *entry,
 427	   struct eb_vma *ev)
 428{
 429	struct i915_vma *vma = ev->vma;
 430	u64 pin_flags;
 431	int err;
 432
 433	if (vma->node.size)
 434		pin_flags = vma->node.start;
 435	else
 436		pin_flags = entry->offset & PIN_OFFSET_MASK;
 437
 438	pin_flags |= PIN_USER | PIN_NOEVICT | PIN_OFFSET_FIXED;
 439	if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_GTT))
 440		pin_flags |= PIN_GLOBAL;
 441
 442	/* Attempt to reuse the current location if available */
 443	err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, pin_flags);
 444	if (err == -EDEADLK)
 445		return err;
 446
 447	if (unlikely(err)) {
 448		if (entry->flags & EXEC_OBJECT_PINNED)
 449			return err;
 450
 451		/* Failing that pick any _free_ space if suitable */
 452		err = i915_vma_pin_ww(vma, &eb->ww,
 453					     entry->pad_to_size,
 454					     entry->alignment,
 455					     eb_pin_flags(entry, ev->flags) |
 456					     PIN_USER | PIN_NOEVICT);
 457		if (unlikely(err))
 458			return err;
 459	}
 460
 461	if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) {
 462		err = i915_vma_pin_fence(vma);
 463		if (unlikely(err)) {
 464			i915_vma_unpin(vma);
 465			return err;
 466		}
 467
 468		if (vma->fence)
 469			ev->flags |= __EXEC_OBJECT_HAS_FENCE;
 470	}
 471
 472	ev->flags |= __EXEC_OBJECT_HAS_PIN;
 473	if (eb_vma_misplaced(entry, vma, ev->flags))
 474		return -EBADSLT;
 475
 476	return 0;
 477}
 478
 479static inline void
 480eb_unreserve_vma(struct eb_vma *ev)
 481{
 482	if (!(ev->flags & __EXEC_OBJECT_HAS_PIN))
 483		return;
 484
 485	if (unlikely(ev->flags & __EXEC_OBJECT_HAS_FENCE))
 486		__i915_vma_unpin_fence(ev->vma);
 487
 488	__i915_vma_unpin(ev->vma);
 489	ev->flags &= ~__EXEC_OBJECT_RESERVED;
 490}
 491
 492static int
 493eb_validate_vma(struct i915_execbuffer *eb,
 494		struct drm_i915_gem_exec_object2 *entry,
 495		struct i915_vma *vma)
 496{
 497	/* Relocations are disallowed for all platforms after TGL-LP.  This
 498	 * also covers all platforms with local memory.
 499	 */
 500	if (entry->relocation_count &&
 501	    GRAPHICS_VER(eb->i915) >= 12 && !IS_TIGERLAKE(eb->i915))
 502		return -EINVAL;
 503
 504	if (unlikely(entry->flags & eb->invalid_flags))
 505		return -EINVAL;
 506
 507	if (unlikely(entry->alignment &&
 508		     !is_power_of_2_u64(entry->alignment)))
 509		return -EINVAL;
 510
 511	/*
 512	 * Offset can be used as input (EXEC_OBJECT_PINNED), reject
 513	 * any non-page-aligned or non-canonical addresses.
 514	 */
 515	if (unlikely(entry->flags & EXEC_OBJECT_PINNED &&
 516		     entry->offset != gen8_canonical_addr(entry->offset & I915_GTT_PAGE_MASK)))
 517		return -EINVAL;
 518
 519	/* pad_to_size was once a reserved field, so sanitize it */
 520	if (entry->flags & EXEC_OBJECT_PAD_TO_SIZE) {
 521		if (unlikely(offset_in_page(entry->pad_to_size)))
 522			return -EINVAL;
 523	} else {
 524		entry->pad_to_size = 0;
 525	}
 526	/*
 527	 * From drm_mm perspective address space is continuous,
 528	 * so from this point we're always using non-canonical
 529	 * form internally.
 530	 */
 531	entry->offset = gen8_noncanonical_addr(entry->offset);
 532
 533	if (!eb->reloc_cache.has_fence) {
 534		entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
 535	} else {
 536		if ((entry->flags & EXEC_OBJECT_NEEDS_FENCE ||
 537		     eb->reloc_cache.needs_unfenced) &&
 538		    i915_gem_object_is_tiled(vma->obj))
 539			entry->flags |= EXEC_OBJECT_NEEDS_GTT | __EXEC_OBJECT_NEEDS_MAP;
 540	}
 541
 542	if (!(entry->flags & EXEC_OBJECT_PINNED))
 543		entry->flags |= eb->context_flags;
 544
 545	return 0;
 546}
 547
 548static void
 549eb_add_vma(struct i915_execbuffer *eb,
 550	   unsigned int i, unsigned batch_idx,
 551	   struct i915_vma *vma)
 552{
 553	struct drm_i915_gem_exec_object2 *entry = &eb->exec[i];
 554	struct eb_vma *ev = &eb->vma[i];
 555
 
 
 556	ev->vma = vma;
 557	ev->exec = entry;
 558	ev->flags = entry->flags;
 559
 560	if (eb->lut_size > 0) {
 561		ev->handle = entry->handle;
 562		hlist_add_head(&ev->node,
 563			       &eb->buckets[hash_32(entry->handle,
 564						    eb->lut_size)]);
 565	}
 566
 567	if (entry->relocation_count)
 568		list_add_tail(&ev->reloc_link, &eb->relocs);
 569
 570	/*
 571	 * SNA is doing fancy tricks with compressing batch buffers, which leads
 572	 * to negative relocation deltas. Usually that works out ok since the
 573	 * relocate address is still positive, except when the batch is placed
 574	 * very low in the GTT. Ensure this doesn't happen.
 575	 *
 576	 * Note that actual hangs have only been observed on gen7, but for
 577	 * paranoia do it everywhere.
 578	 */
 579	if (i == batch_idx) {
 580		if (entry->relocation_count &&
 581		    !(ev->flags & EXEC_OBJECT_PINNED))
 582			ev->flags |= __EXEC_OBJECT_NEEDS_BIAS;
 583		if (eb->reloc_cache.has_fence)
 584			ev->flags |= EXEC_OBJECT_NEEDS_FENCE;
 585
 586		eb->batch = ev;
 587	}
 
 
 
 
 
 
 
 
 
 
 588}
 589
 590static inline int use_cpu_reloc(const struct reloc_cache *cache,
 591				const struct drm_i915_gem_object *obj)
 592{
 593	if (!i915_gem_object_has_struct_page(obj))
 594		return false;
 595
 596	if (DBG_FORCE_RELOC == FORCE_CPU_RELOC)
 597		return true;
 598
 599	if (DBG_FORCE_RELOC == FORCE_GTT_RELOC)
 600		return false;
 601
 602	return (cache->has_llc ||
 603		obj->cache_dirty ||
 604		obj->cache_level != I915_CACHE_NONE);
 605}
 606
 607static int eb_reserve_vma(struct i915_execbuffer *eb,
 608			  struct eb_vma *ev,
 609			  u64 pin_flags)
 610{
 611	struct drm_i915_gem_exec_object2 *entry = ev->exec;
 612	struct i915_vma *vma = ev->vma;
 613	int err;
 614
 615	if (drm_mm_node_allocated(&vma->node) &&
 616	    eb_vma_misplaced(entry, vma, ev->flags)) {
 617		err = i915_vma_unbind(vma);
 618		if (err)
 619			return err;
 620	}
 621
 622	err = i915_vma_pin_ww(vma, &eb->ww,
 623			   entry->pad_to_size, entry->alignment,
 624			   eb_pin_flags(entry, ev->flags) | pin_flags);
 625	if (err)
 626		return err;
 627
 628	if (entry->offset != vma->node.start) {
 629		entry->offset = vma->node.start | UPDATE;
 630		eb->args->flags |= __EXEC_HAS_RELOC;
 631	}
 632
 633	if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) {
 634		err = i915_vma_pin_fence(vma);
 635		if (unlikely(err)) {
 636			i915_vma_unpin(vma);
 637			return err;
 638		}
 639
 640		if (vma->fence)
 641			ev->flags |= __EXEC_OBJECT_HAS_FENCE;
 642	}
 643
 644	ev->flags |= __EXEC_OBJECT_HAS_PIN;
 645	GEM_BUG_ON(eb_vma_misplaced(entry, vma, ev->flags));
 646
 647	return 0;
 648}
 649
 650static int eb_reserve(struct i915_execbuffer *eb)
 651{
 652	const unsigned int count = eb->buffer_count;
 653	unsigned int pin_flags = PIN_USER | PIN_NONBLOCK;
 654	struct list_head last;
 655	struct eb_vma *ev;
 656	unsigned int i, pass;
 657	int err = 0;
 658
 659	/*
 660	 * Attempt to pin all of the buffers into the GTT.
 661	 * This is done in 3 phases:
 662	 *
 663	 * 1a. Unbind all objects that do not match the GTT constraints for
 664	 *     the execbuffer (fenceable, mappable, alignment etc).
 665	 * 1b. Increment pin count for already bound objects.
 666	 * 2.  Bind new objects.
 667	 * 3.  Decrement pin count.
 668	 *
 669	 * This avoid unnecessary unbinding of later objects in order to make
 670	 * room for the earlier objects *unless* we need to defragment.
 671	 */
 
 
 
 
 672	pass = 0;
 673	do {
 674		list_for_each_entry(ev, &eb->unbound, bind_link) {
 675			err = eb_reserve_vma(eb, ev, pin_flags);
 676			if (err)
 677				break;
 678		}
 679		if (err != -ENOSPC)
 680			return err;
 681
 682		/* Resort *all* the objects into priority order */
 683		INIT_LIST_HEAD(&eb->unbound);
 684		INIT_LIST_HEAD(&last);
 685		for (i = 0; i < count; i++) {
 686			unsigned int flags;
 687
 688			ev = &eb->vma[i];
 689			flags = ev->flags;
 690			if (flags & EXEC_OBJECT_PINNED &&
 691			    flags & __EXEC_OBJECT_HAS_PIN)
 692				continue;
 693
 694			eb_unreserve_vma(ev);
 695
 696			if (flags & EXEC_OBJECT_PINNED)
 697				/* Pinned must have their slot */
 698				list_add(&ev->bind_link, &eb->unbound);
 699			else if (flags & __EXEC_OBJECT_NEEDS_MAP)
 700				/* Map require the lowest 256MiB (aperture) */
 701				list_add_tail(&ev->bind_link, &eb->unbound);
 702			else if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS))
 703				/* Prioritise 4GiB region for restricted bo */
 704				list_add(&ev->bind_link, &last);
 705			else
 706				list_add_tail(&ev->bind_link, &last);
 707		}
 708		list_splice_tail(&last, &eb->unbound);
 709
 
 
 
 
 
 
 
 710		switch (pass++) {
 711		case 0:
 712			break;
 713
 714		case 1:
 715			/* Too fragmented, unbind everything and retry */
 716			mutex_lock(&eb->context->vm->mutex);
 717			err = i915_gem_evict_vm(eb->context->vm);
 718			mutex_unlock(&eb->context->vm->mutex);
 719			if (err)
 720				return err;
 721			break;
 722
 723		default:
 724			return -ENOSPC;
 
 725		}
 726
 727		pin_flags = PIN_USER;
 728	} while (1);
 
 
 
 
 729}
 730
 731static unsigned int eb_batch_index(const struct i915_execbuffer *eb)
 732{
 733	if (eb->args->flags & I915_EXEC_BATCH_FIRST)
 734		return 0;
 735	else
 736		return eb->buffer_count - 1;
 737}
 738
 739static int eb_select_context(struct i915_execbuffer *eb)
 740{
 741	struct i915_gem_context *ctx;
 742
 743	ctx = i915_gem_context_lookup(eb->file->driver_priv, eb->args->rsvd1);
 744	if (unlikely(!ctx))
 745		return -ENOENT;
 746
 747	eb->gem_context = ctx;
 748	if (rcu_access_pointer(ctx->vm))
 749		eb->invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
 750
 751	eb->context_flags = 0;
 752	if (test_bit(UCONTEXT_NO_ZEROMAP, &ctx->user_flags))
 753		eb->context_flags |= __EXEC_OBJECT_NEEDS_BIAS;
 754
 755	return 0;
 756}
 757
 758static int __eb_add_lut(struct i915_execbuffer *eb,
 759			u32 handle, struct i915_vma *vma)
 760{
 761	struct i915_gem_context *ctx = eb->gem_context;
 762	struct i915_lut_handle *lut;
 763	int err;
 764
 765	lut = i915_lut_handle_alloc();
 766	if (unlikely(!lut))
 767		return -ENOMEM;
 768
 769	i915_vma_get(vma);
 770	if (!atomic_fetch_inc(&vma->open_count))
 771		i915_vma_reopen(vma);
 772	lut->handle = handle;
 773	lut->ctx = ctx;
 774
 775	/* Check that the context hasn't been closed in the meantime */
 776	err = -EINTR;
 777	if (!mutex_lock_interruptible(&ctx->lut_mutex)) {
 778		struct i915_address_space *vm = rcu_access_pointer(ctx->vm);
 779
 780		if (unlikely(vm && vma->vm != vm))
 781			err = -EAGAIN; /* user racing with ctx set-vm */
 782		else if (likely(!i915_gem_context_is_closed(ctx)))
 783			err = radix_tree_insert(&ctx->handles_vma, handle, vma);
 784		else
 785			err = -ENOENT;
 786		if (err == 0) { /* And nor has this handle */
 787			struct drm_i915_gem_object *obj = vma->obj;
 788
 789			spin_lock(&obj->lut_lock);
 790			if (idr_find(&eb->file->object_idr, handle) == obj) {
 791				list_add(&lut->obj_link, &obj->lut_list);
 792			} else {
 793				radix_tree_delete(&ctx->handles_vma, handle);
 794				err = -ENOENT;
 795			}
 796			spin_unlock(&obj->lut_lock);
 797		}
 798		mutex_unlock(&ctx->lut_mutex);
 799	}
 800	if (unlikely(err))
 801		goto err;
 802
 803	return 0;
 804
 805err:
 806	i915_vma_close(vma);
 807	i915_vma_put(vma);
 808	i915_lut_handle_free(lut);
 809	return err;
 810}
 811
 812static struct i915_vma *eb_lookup_vma(struct i915_execbuffer *eb, u32 handle)
 813{
 814	struct i915_address_space *vm = eb->context->vm;
 815
 816	do {
 817		struct drm_i915_gem_object *obj;
 818		struct i915_vma *vma;
 819		int err;
 820
 821		rcu_read_lock();
 822		vma = radix_tree_lookup(&eb->gem_context->handles_vma, handle);
 823		if (likely(vma && vma->vm == vm))
 824			vma = i915_vma_tryget(vma);
 825		rcu_read_unlock();
 826		if (likely(vma))
 827			return vma;
 828
 829		obj = i915_gem_object_lookup(eb->file, handle);
 830		if (unlikely(!obj))
 831			return ERR_PTR(-ENOENT);
 832
 833		vma = i915_vma_instance(obj, vm, NULL);
 834		if (IS_ERR(vma)) {
 835			i915_gem_object_put(obj);
 836			return vma;
 837		}
 838
 839		err = __eb_add_lut(eb, handle, vma);
 840		if (likely(!err))
 841			return vma;
 842
 843		i915_gem_object_put(obj);
 844		if (err != -EEXIST)
 845			return ERR_PTR(err);
 846	} while (1);
 847}
 848
 849static int eb_lookup_vmas(struct i915_execbuffer *eb)
 850{
 851	struct drm_i915_private *i915 = eb->i915;
 852	unsigned int batch = eb_batch_index(eb);
 853	unsigned int i;
 854	int err = 0;
 855
 856	INIT_LIST_HEAD(&eb->relocs);
 
 857
 858	for (i = 0; i < eb->buffer_count; i++) {
 859		struct i915_vma *vma;
 860
 861		vma = eb_lookup_vma(eb, eb->exec[i].handle);
 862		if (IS_ERR(vma)) {
 863			err = PTR_ERR(vma);
 864			goto err;
 865		}
 866
 867		err = eb_validate_vma(eb, &eb->exec[i], vma);
 868		if (unlikely(err)) {
 869			i915_vma_put(vma);
 870			goto err;
 871		}
 872
 873		eb_add_vma(eb, i, batch, vma);
 874
 875		if (i915_gem_object_is_userptr(vma->obj)) {
 876			err = i915_gem_object_userptr_submit_init(vma->obj);
 877			if (err) {
 878				if (i + 1 < eb->buffer_count) {
 879					/*
 880					 * Execbuffer code expects last vma entry to be NULL,
 881					 * since we already initialized this entry,
 882					 * set the next value to NULL or we mess up
 883					 * cleanup handling.
 884					 */
 885					eb->vma[i + 1].vma = NULL;
 886				}
 887
 888				return err;
 889			}
 890
 891			eb->vma[i].flags |= __EXEC_OBJECT_USERPTR_INIT;
 892			eb->args->flags |= __EXEC_USERPTR_USED;
 893		}
 894	}
 895
 896	if (unlikely(eb->batch->flags & EXEC_OBJECT_WRITE)) {
 897		drm_dbg(&i915->drm,
 898			"Attempting to use self-modifying batch buffer\n");
 899		return -EINVAL;
 900	}
 901
 902	if (range_overflows_t(u64,
 903			      eb->batch_start_offset, eb->batch_len,
 904			      eb->batch->vma->size)) {
 905		drm_dbg(&i915->drm, "Attempting to use out-of-bounds batch\n");
 906		return -EINVAL;
 907	}
 908
 909	if (eb->batch_len == 0)
 910		eb->batch_len = eb->batch->vma->size - eb->batch_start_offset;
 911	if (unlikely(eb->batch_len == 0)) { /* impossible! */
 912		drm_dbg(&i915->drm, "Invalid batch length\n");
 913		return -EINVAL;
 914	}
 915
 916	return 0;
 917
 918err:
 919	eb->vma[i].vma = NULL;
 920	return err;
 921}
 922
 923static int eb_validate_vmas(struct i915_execbuffer *eb)
 924{
 925	unsigned int i;
 926	int err;
 927
 928	INIT_LIST_HEAD(&eb->unbound);
 929
 930	for (i = 0; i < eb->buffer_count; i++) {
 931		struct drm_i915_gem_exec_object2 *entry = &eb->exec[i];
 932		struct eb_vma *ev = &eb->vma[i];
 933		struct i915_vma *vma = ev->vma;
 934
 935		err = i915_gem_object_lock(vma->obj, &eb->ww);
 936		if (err)
 937			return err;
 938
 939		err = eb_pin_vma(eb, entry, ev);
 940		if (err == -EDEADLK)
 941			return err;
 942
 943		if (!err) {
 944			if (entry->offset != vma->node.start) {
 945				entry->offset = vma->node.start | UPDATE;
 946				eb->args->flags |= __EXEC_HAS_RELOC;
 947			}
 948		} else {
 949			eb_unreserve_vma(ev);
 950
 951			list_add_tail(&ev->bind_link, &eb->unbound);
 952			if (drm_mm_node_allocated(&vma->node)) {
 953				err = i915_vma_unbind(vma);
 954				if (err)
 955					return err;
 956			}
 957		}
 958
 959		if (!(ev->flags & EXEC_OBJECT_WRITE)) {
 960			err = dma_resv_reserve_shared(vma->resv, 1);
 961			if (err)
 962				return err;
 963		}
 964
 965		GEM_BUG_ON(drm_mm_node_allocated(&vma->node) &&
 966			   eb_vma_misplaced(&eb->exec[i], vma, ev->flags));
 967	}
 968
 969	if (!list_empty(&eb->unbound))
 970		return eb_reserve(eb);
 971
 972	return 0;
 973}
 974
 975static struct eb_vma *
 976eb_get_vma(const struct i915_execbuffer *eb, unsigned long handle)
 977{
 978	if (eb->lut_size < 0) {
 979		if (handle >= -eb->lut_size)
 980			return NULL;
 981		return &eb->vma[handle];
 982	} else {
 983		struct hlist_head *head;
 984		struct eb_vma *ev;
 985
 986		head = &eb->buckets[hash_32(handle, eb->lut_size)];
 987		hlist_for_each_entry(ev, head, node) {
 988			if (ev->handle == handle)
 989				return ev;
 990		}
 991		return NULL;
 992	}
 993}
 994
 995static void eb_release_vmas(struct i915_execbuffer *eb, bool final, bool release_userptr)
 996{
 997	const unsigned int count = eb->buffer_count;
 998	unsigned int i;
 999
1000	for (i = 0; i < count; i++) {
1001		struct eb_vma *ev = &eb->vma[i];
1002		struct i915_vma *vma = ev->vma;
1003
1004		if (!vma)
1005			break;
1006
1007		eb_unreserve_vma(ev);
1008
1009		if (release_userptr && ev->flags & __EXEC_OBJECT_USERPTR_INIT) {
1010			ev->flags &= ~__EXEC_OBJECT_USERPTR_INIT;
1011			i915_gem_object_userptr_submit_fini(vma->obj);
1012		}
1013
1014		if (final)
1015			i915_vma_put(vma);
1016	}
1017
1018	eb_unpin_engine(eb);
1019}
1020
1021static void eb_destroy(const struct i915_execbuffer *eb)
1022{
1023	GEM_BUG_ON(eb->reloc_cache.rq);
1024
 
 
 
1025	if (eb->lut_size > 0)
1026		kfree(eb->buckets);
1027}
1028
1029static inline u64
1030relocation_target(const struct drm_i915_gem_relocation_entry *reloc,
1031		  const struct i915_vma *target)
1032{
1033	return gen8_canonical_addr((int)reloc->delta + target->node.start);
1034}
1035
1036static void reloc_cache_clear(struct reloc_cache *cache)
1037{
1038	cache->rq = NULL;
1039	cache->rq_cmd = NULL;
1040	cache->pool = NULL;
1041	cache->rq_size = 0;
1042}
1043
1044static void reloc_cache_init(struct reloc_cache *cache,
1045			     struct drm_i915_private *i915)
1046{
1047	cache->page = -1;
1048	cache->vaddr = 0;
1049	/* Must be a variable in the struct to allow GCC to unroll. */
1050	cache->graphics_ver = GRAPHICS_VER(i915);
1051	cache->has_llc = HAS_LLC(i915);
1052	cache->use_64bit_reloc = HAS_64BIT_RELOC(i915);
1053	cache->has_fence = cache->graphics_ver < 4;
1054	cache->needs_unfenced = INTEL_INFO(i915)->unfenced_needs_alignment;
1055	cache->node.flags = 0;
1056	reloc_cache_clear(cache);
 
1057}
1058
1059static inline void *unmask_page(unsigned long p)
1060{
1061	return (void *)(uintptr_t)(p & PAGE_MASK);
1062}
1063
1064static inline unsigned int unmask_flags(unsigned long p)
1065{
1066	return p & ~PAGE_MASK;
1067}
1068
1069#define KMAP 0x4 /* after CLFLUSH_FLAGS */
1070
1071static inline struct i915_ggtt *cache_to_ggtt(struct reloc_cache *cache)
1072{
1073	struct drm_i915_private *i915 =
1074		container_of(cache, struct i915_execbuffer, reloc_cache)->i915;
1075	return &i915->ggtt;
1076}
1077
1078static void reloc_cache_put_pool(struct i915_execbuffer *eb, struct reloc_cache *cache)
 
 
1079{
1080	if (!cache->pool)
1081		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1082
1083	/*
1084	 * This is a bit nasty, normally we keep objects locked until the end
1085	 * of execbuffer, but we already submit this, and have to unlock before
1086	 * dropping the reference. Fortunately we can only hold 1 pool node at
1087	 * a time, so this should be harmless.
1088	 */
1089	i915_gem_ww_unlock_single(cache->pool->obj);
1090	intel_gt_buffer_pool_put(cache->pool);
1091	cache->pool = NULL;
1092}
1093
1094static void reloc_gpu_flush(struct i915_execbuffer *eb, struct reloc_cache *cache)
1095{
1096	struct drm_i915_gem_object *obj = cache->rq->batch->obj;
 
 
 
 
 
1097
1098	GEM_BUG_ON(cache->rq_size >= obj->base.size / sizeof(u32));
1099	cache->rq_cmd[cache->rq_size] = MI_BATCH_BUFFER_END;
1100
1101	i915_gem_object_flush_map(obj);
1102	i915_gem_object_unpin_map(obj);
1103
1104	intel_gt_chipset_flush(cache->rq->engine->gt);
 
 
 
1105
1106	i915_request_add(cache->rq);
1107	reloc_cache_put_pool(eb, cache);
1108	reloc_cache_clear(cache);
 
 
 
 
 
 
 
 
 
 
1109
1110	eb->reloc_pool = NULL;
1111}
1112
1113static void reloc_cache_reset(struct reloc_cache *cache, struct i915_execbuffer *eb)
1114{
1115	void *vaddr;
1116
1117	if (cache->rq)
1118		reloc_gpu_flush(eb, cache);
1119
1120	if (!cache->vaddr)
1121		return;
1122
1123	vaddr = unmask_page(cache->vaddr);
1124	if (cache->vaddr & KMAP) {
1125		struct drm_i915_gem_object *obj =
1126			(struct drm_i915_gem_object *)cache->node.mm;
1127		if (cache->vaddr & CLFLUSH_AFTER)
1128			mb();
1129
1130		kunmap_atomic(vaddr);
1131		i915_gem_object_finish_access(obj);
1132	} else {
1133		struct i915_ggtt *ggtt = cache_to_ggtt(cache);
1134
1135		intel_gt_flush_ggtt_writes(ggtt->vm.gt);
1136		io_mapping_unmap_atomic((void __iomem *)vaddr);
1137
1138		if (drm_mm_node_allocated(&cache->node)) {
1139			ggtt->vm.clear_range(&ggtt->vm,
1140					     cache->node.start,
1141					     cache->node.size);
1142			mutex_lock(&ggtt->vm.mutex);
1143			drm_mm_remove_node(&cache->node);
1144			mutex_unlock(&ggtt->vm.mutex);
1145		} else {
1146			i915_vma_unpin((struct i915_vma *)cache->node.mm);
1147		}
1148	}
1149
1150	cache->vaddr = 0;
1151	cache->page = -1;
1152}
1153
1154static void *reloc_kmap(struct drm_i915_gem_object *obj,
1155			struct reloc_cache *cache,
1156			unsigned long pageno)
1157{
1158	void *vaddr;
1159	struct page *page;
1160
1161	if (cache->vaddr) {
1162		kunmap_atomic(unmask_page(cache->vaddr));
1163	} else {
1164		unsigned int flushes;
1165		int err;
1166
1167		err = i915_gem_object_prepare_write(obj, &flushes);
1168		if (err)
1169			return ERR_PTR(err);
1170
1171		BUILD_BUG_ON(KMAP & CLFLUSH_FLAGS);
1172		BUILD_BUG_ON((KMAP | CLFLUSH_FLAGS) & PAGE_MASK);
1173
1174		cache->vaddr = flushes | KMAP;
1175		cache->node.mm = (void *)obj;
1176		if (flushes)
1177			mb();
1178	}
1179
1180	page = i915_gem_object_get_page(obj, pageno);
1181	if (!obj->mm.dirty)
1182		set_page_dirty(page);
1183
1184	vaddr = kmap_atomic(page);
1185	cache->vaddr = unmask_flags(cache->vaddr) | (unsigned long)vaddr;
1186	cache->page = pageno;
1187
1188	return vaddr;
1189}
1190
1191static void *reloc_iomap(struct drm_i915_gem_object *obj,
1192			 struct i915_execbuffer *eb,
1193			 unsigned long page)
1194{
1195	struct reloc_cache *cache = &eb->reloc_cache;
1196	struct i915_ggtt *ggtt = cache_to_ggtt(cache);
1197	unsigned long offset;
1198	void *vaddr;
1199
1200	if (cache->vaddr) {
1201		intel_gt_flush_ggtt_writes(ggtt->vm.gt);
1202		io_mapping_unmap_atomic((void __force __iomem *) unmask_page(cache->vaddr));
1203	} else {
1204		struct i915_vma *vma;
1205		int err;
1206
1207		if (i915_gem_object_is_tiled(obj))
1208			return ERR_PTR(-EINVAL);
1209
1210		if (use_cpu_reloc(cache, obj))
1211			return NULL;
1212
 
1213		err = i915_gem_object_set_to_gtt_domain(obj, true);
 
1214		if (err)
1215			return ERR_PTR(err);
1216
1217		vma = i915_gem_object_ggtt_pin_ww(obj, &eb->ww, NULL, 0, 0,
1218						  PIN_MAPPABLE |
1219						  PIN_NONBLOCK /* NOWARN */ |
1220						  PIN_NOEVICT);
1221		if (vma == ERR_PTR(-EDEADLK))
1222			return vma;
1223
1224		if (IS_ERR(vma)) {
1225			memset(&cache->node, 0, sizeof(cache->node));
1226			mutex_lock(&ggtt->vm.mutex);
1227			err = drm_mm_insert_node_in_range
1228				(&ggtt->vm.mm, &cache->node,
1229				 PAGE_SIZE, 0, I915_COLOR_UNEVICTABLE,
1230				 0, ggtt->mappable_end,
1231				 DRM_MM_INSERT_LOW);
1232			mutex_unlock(&ggtt->vm.mutex);
1233			if (err) /* no inactive aperture space, use cpu reloc */
1234				return NULL;
1235		} else {
1236			cache->node.start = vma->node.start;
1237			cache->node.mm = (void *)vma;
1238		}
1239	}
1240
1241	offset = cache->node.start;
1242	if (drm_mm_node_allocated(&cache->node)) {
1243		ggtt->vm.insert_page(&ggtt->vm,
1244				     i915_gem_object_get_dma_address(obj, page),
1245				     offset, I915_CACHE_NONE, 0);
1246	} else {
1247		offset += page << PAGE_SHIFT;
1248	}
1249
1250	vaddr = (void __force *)io_mapping_map_atomic_wc(&ggtt->iomap,
1251							 offset);
1252	cache->page = page;
1253	cache->vaddr = (unsigned long)vaddr;
1254
1255	return vaddr;
1256}
1257
1258static void *reloc_vaddr(struct drm_i915_gem_object *obj,
1259			 struct i915_execbuffer *eb,
1260			 unsigned long page)
1261{
1262	struct reloc_cache *cache = &eb->reloc_cache;
1263	void *vaddr;
1264
1265	if (cache->page == page) {
1266		vaddr = unmask_page(cache->vaddr);
1267	} else {
1268		vaddr = NULL;
1269		if ((cache->vaddr & KMAP) == 0)
1270			vaddr = reloc_iomap(obj, eb, page);
1271		if (!vaddr)
1272			vaddr = reloc_kmap(obj, cache, page);
1273	}
1274
1275	return vaddr;
1276}
1277
1278static void clflush_write32(u32 *addr, u32 value, unsigned int flushes)
1279{
1280	if (unlikely(flushes & (CLFLUSH_BEFORE | CLFLUSH_AFTER))) {
1281		if (flushes & CLFLUSH_BEFORE) {
1282			clflushopt(addr);
1283			mb();
1284		}
1285
1286		*addr = value;
1287
1288		/*
1289		 * Writes to the same cacheline are serialised by the CPU
1290		 * (including clflush). On the write path, we only require
1291		 * that it hits memory in an orderly fashion and place
1292		 * mb barriers at the start and end of the relocation phase
1293		 * to ensure ordering of clflush wrt to the system.
1294		 */
1295		if (flushes & CLFLUSH_AFTER)
1296			clflushopt(addr);
1297	} else
1298		*addr = value;
1299}
1300
1301static int reloc_move_to_gpu(struct i915_request *rq, struct i915_vma *vma)
1302{
1303	struct drm_i915_gem_object *obj = vma->obj;
1304	int err;
1305
1306	assert_vma_held(vma);
1307
1308	if (obj->cache_dirty & ~obj->cache_coherent)
1309		i915_gem_clflush_object(obj, 0);
1310	obj->write_domain = 0;
1311
1312	err = i915_request_await_object(rq, vma->obj, true);
1313	if (err == 0)
1314		err = i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE);
1315
 
 
1316	return err;
1317}
1318
1319static int __reloc_gpu_alloc(struct i915_execbuffer *eb,
1320			     struct intel_engine_cs *engine,
1321			     struct i915_vma *vma,
1322			     unsigned int len)
1323{
1324	struct reloc_cache *cache = &eb->reloc_cache;
1325	struct intel_gt_buffer_pool_node *pool = eb->reloc_pool;
1326	struct i915_request *rq;
1327	struct i915_vma *batch;
1328	u32 *cmd;
1329	int err;
1330
1331	if (!pool) {
1332		pool = intel_gt_get_buffer_pool(engine->gt, PAGE_SIZE,
1333						cache->has_llc ?
1334						I915_MAP_WB :
1335						I915_MAP_WC);
1336		if (IS_ERR(pool))
1337			return PTR_ERR(pool);
1338	}
1339	eb->reloc_pool = NULL;
1340
1341	err = i915_gem_object_lock(pool->obj, &eb->ww);
1342	if (err)
1343		goto err_pool;
1344
1345	cmd = i915_gem_object_pin_map(pool->obj, pool->type);
1346	if (IS_ERR(cmd)) {
1347		err = PTR_ERR(cmd);
1348		goto err_pool;
1349	}
1350	intel_gt_buffer_pool_mark_used(pool);
1351
1352	memset32(cmd, 0, pool->obj->base.size / sizeof(u32));
1353
1354	batch = i915_vma_instance(pool->obj, vma->vm, NULL);
1355	if (IS_ERR(batch)) {
1356		err = PTR_ERR(batch);
1357		goto err_unmap;
1358	}
1359
1360	err = i915_vma_pin_ww(batch, &eb->ww, 0, 0, PIN_USER | PIN_NONBLOCK);
1361	if (err)
1362		goto err_unmap;
1363
1364	if (engine == eb->context->engine) {
1365		rq = i915_request_create(eb->context);
1366	} else {
1367		struct intel_context *ce = eb->reloc_context;
1368
1369		if (!ce) {
1370			ce = intel_context_create(engine);
1371			if (IS_ERR(ce)) {
1372				err = PTR_ERR(ce);
1373				goto err_unpin;
1374			}
1375
1376			i915_vm_put(ce->vm);
1377			ce->vm = i915_vm_get(eb->context->vm);
1378			eb->reloc_context = ce;
1379		}
1380
1381		err = intel_context_pin_ww(ce, &eb->ww);
1382		if (err)
1383			goto err_unpin;
1384
1385		rq = i915_request_create(ce);
1386		intel_context_unpin(ce);
1387	}
1388	if (IS_ERR(rq)) {
1389		err = PTR_ERR(rq);
1390		goto err_unpin;
1391	}
1392
1393	err = intel_gt_buffer_pool_mark_active(pool, rq);
1394	if (err)
1395		goto err_request;
1396
1397	err = reloc_move_to_gpu(rq, vma);
1398	if (err)
1399		goto err_request;
1400
1401	err = eb->engine->emit_bb_start(rq,
1402					batch->node.start, PAGE_SIZE,
1403					cache->graphics_ver > 5 ? 0 : I915_DISPATCH_SECURE);
1404	if (err)
1405		goto skip_request;
1406
1407	assert_vma_held(batch);
1408	err = i915_request_await_object(rq, batch->obj, false);
1409	if (err == 0)
1410		err = i915_vma_move_to_active(batch, rq, 0);
 
1411	if (err)
1412		goto skip_request;
1413
1414	rq->batch = batch;
1415	i915_vma_unpin(batch);
1416
1417	cache->rq = rq;
1418	cache->rq_cmd = cmd;
1419	cache->rq_size = 0;
1420	cache->pool = pool;
1421
1422	/* Return with batch mapping (cmd) still pinned */
1423	return 0;
1424
1425skip_request:
1426	i915_request_set_error_once(rq, err);
1427err_request:
1428	i915_request_add(rq);
1429err_unpin:
1430	i915_vma_unpin(batch);
1431err_unmap:
1432	i915_gem_object_unpin_map(pool->obj);
1433err_pool:
1434	eb->reloc_pool = pool;
1435	return err;
1436}
1437
1438static bool reloc_can_use_engine(const struct intel_engine_cs *engine)
1439{
1440	return engine->class != VIDEO_DECODE_CLASS || GRAPHICS_VER(engine->i915) != 6;
1441}
1442
1443static u32 *reloc_gpu(struct i915_execbuffer *eb,
1444		      struct i915_vma *vma,
1445		      unsigned int len)
1446{
1447	struct reloc_cache *cache = &eb->reloc_cache;
1448	u32 *cmd;
1449
1450	if (cache->rq_size > PAGE_SIZE/sizeof(u32) - (len + 1))
1451		reloc_gpu_flush(eb, cache);
1452
1453	if (unlikely(!cache->rq)) {
1454		int err;
1455		struct intel_engine_cs *engine = eb->engine;
1456
1457		/* If we need to copy for the cmdparser, we will stall anyway */
1458		if (eb_use_cmdparser(eb))
1459			return ERR_PTR(-EWOULDBLOCK);
1460
1461		if (!reloc_can_use_engine(engine)) {
1462			engine = engine->gt->engine_class[COPY_ENGINE_CLASS][0];
1463			if (!engine)
1464				return ERR_PTR(-ENODEV);
1465		}
1466
1467		err = __reloc_gpu_alloc(eb, engine, vma, len);
1468		if (unlikely(err))
1469			return ERR_PTR(err);
1470	}
1471
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1472	cmd = cache->rq_cmd + cache->rq_size;
1473	cache->rq_size += len;
1474
1475	return cmd;
1476}
1477
1478static inline bool use_reloc_gpu(struct i915_vma *vma)
1479{
1480	if (DBG_FORCE_RELOC == FORCE_GPU_RELOC)
1481		return true;
1482
1483	if (DBG_FORCE_RELOC)
1484		return false;
1485
1486	return !dma_resv_test_signaled(vma->resv, true);
1487}
1488
1489static unsigned long vma_phys_addr(struct i915_vma *vma, u32 offset)
1490{
1491	struct page *page;
1492	unsigned long addr;
1493
1494	GEM_BUG_ON(vma->pages != vma->obj->mm.pages);
1495
1496	page = i915_gem_object_get_page(vma->obj, offset >> PAGE_SHIFT);
1497	addr = PFN_PHYS(page_to_pfn(page));
1498	GEM_BUG_ON(overflows_type(addr, u32)); /* expected dma32 */
1499
1500	return addr + offset_in_page(offset);
1501}
1502
1503static int __reloc_entry_gpu(struct i915_execbuffer *eb,
1504			      struct i915_vma *vma,
1505			      u64 offset,
1506			      u64 target_addr)
1507{
1508	const unsigned int ver = eb->reloc_cache.graphics_ver;
1509	unsigned int len;
1510	u32 *batch;
1511	u64 addr;
1512
1513	if (ver >= 8)
1514		len = offset & 7 ? 8 : 5;
1515	else if (ver >= 4)
1516		len = 4;
1517	else
1518		len = 3;
1519
1520	batch = reloc_gpu(eb, vma, len);
1521	if (batch == ERR_PTR(-EDEADLK))
1522		return -EDEADLK;
1523	else if (IS_ERR(batch))
1524		return false;
1525
1526	addr = gen8_canonical_addr(vma->node.start + offset);
1527	if (ver >= 8) {
1528		if (offset & 7) {
1529			*batch++ = MI_STORE_DWORD_IMM_GEN4;
1530			*batch++ = lower_32_bits(addr);
1531			*batch++ = upper_32_bits(addr);
1532			*batch++ = lower_32_bits(target_addr);
1533
1534			addr = gen8_canonical_addr(addr + 4);
1535
1536			*batch++ = MI_STORE_DWORD_IMM_GEN4;
1537			*batch++ = lower_32_bits(addr);
1538			*batch++ = upper_32_bits(addr);
1539			*batch++ = upper_32_bits(target_addr);
1540		} else {
1541			*batch++ = (MI_STORE_DWORD_IMM_GEN4 | (1 << 21)) + 1;
1542			*batch++ = lower_32_bits(addr);
1543			*batch++ = upper_32_bits(addr);
1544			*batch++ = lower_32_bits(target_addr);
1545			*batch++ = upper_32_bits(target_addr);
1546		}
1547	} else if (ver >= 6) {
1548		*batch++ = MI_STORE_DWORD_IMM_GEN4;
1549		*batch++ = 0;
1550		*batch++ = addr;
1551		*batch++ = target_addr;
1552	} else if (IS_I965G(eb->i915)) {
1553		*batch++ = MI_STORE_DWORD_IMM_GEN4;
1554		*batch++ = 0;
1555		*batch++ = vma_phys_addr(vma, offset);
1556		*batch++ = target_addr;
1557	} else if (ver >= 4) {
1558		*batch++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
1559		*batch++ = 0;
1560		*batch++ = addr;
1561		*batch++ = target_addr;
1562	} else if (ver >= 3 &&
1563		   !(IS_I915G(eb->i915) || IS_I915GM(eb->i915))) {
1564		*batch++ = MI_STORE_DWORD_IMM | MI_MEM_VIRTUAL;
1565		*batch++ = addr;
1566		*batch++ = target_addr;
1567	} else {
1568		*batch++ = MI_STORE_DWORD_IMM;
1569		*batch++ = vma_phys_addr(vma, offset);
1570		*batch++ = target_addr;
1571	}
1572
1573	return true;
1574}
1575
1576static int reloc_entry_gpu(struct i915_execbuffer *eb,
1577			    struct i915_vma *vma,
1578			    u64 offset,
1579			    u64 target_addr)
1580{
1581	if (eb->reloc_cache.vaddr)
1582		return false;
1583
1584	if (!use_reloc_gpu(vma))
1585		return false;
1586
1587	return __reloc_entry_gpu(eb, vma, offset, target_addr);
1588}
1589
1590static u64
1591relocate_entry(struct i915_vma *vma,
1592	       const struct drm_i915_gem_relocation_entry *reloc,
1593	       struct i915_execbuffer *eb,
1594	       const struct i915_vma *target)
1595{
1596	u64 target_addr = relocation_target(reloc, target);
1597	u64 offset = reloc->offset;
1598	int reloc_gpu = reloc_entry_gpu(eb, vma, offset, target_addr);
1599
1600	if (reloc_gpu < 0)
1601		return reloc_gpu;
1602
1603	if (!reloc_gpu) {
1604		bool wide = eb->reloc_cache.use_64bit_reloc;
1605		void *vaddr;
1606
1607repeat:
1608		vaddr = reloc_vaddr(vma->obj, eb,
 
1609				    offset >> PAGE_SHIFT);
1610		if (IS_ERR(vaddr))
1611			return PTR_ERR(vaddr);
1612
1613		GEM_BUG_ON(!IS_ALIGNED(offset, sizeof(u32)));
1614		clflush_write32(vaddr + offset_in_page(offset),
1615				lower_32_bits(target_addr),
1616				eb->reloc_cache.vaddr);
1617
1618		if (wide) {
1619			offset += sizeof(u32);
1620			target_addr >>= 32;
1621			wide = false;
1622			goto repeat;
1623		}
1624	}
1625
1626	return target->node.start | UPDATE;
1627}
1628
1629static u64
1630eb_relocate_entry(struct i915_execbuffer *eb,
1631		  struct eb_vma *ev,
1632		  const struct drm_i915_gem_relocation_entry *reloc)
1633{
1634	struct drm_i915_private *i915 = eb->i915;
1635	struct eb_vma *target;
1636	int err;
1637
1638	/* we've already hold a reference to all valid objects */
1639	target = eb_get_vma(eb, reloc->target_handle);
1640	if (unlikely(!target))
1641		return -ENOENT;
1642
1643	/* Validate that the target is in a valid r/w GPU domain */
1644	if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
1645		drm_dbg(&i915->drm, "reloc with multiple write domains: "
1646			  "target %d offset %d "
1647			  "read %08x write %08x",
1648			  reloc->target_handle,
1649			  (int) reloc->offset,
1650			  reloc->read_domains,
1651			  reloc->write_domain);
1652		return -EINVAL;
1653	}
1654	if (unlikely((reloc->write_domain | reloc->read_domains)
1655		     & ~I915_GEM_GPU_DOMAINS)) {
1656		drm_dbg(&i915->drm, "reloc with read/write non-GPU domains: "
1657			  "target %d offset %d "
1658			  "read %08x write %08x",
1659			  reloc->target_handle,
1660			  (int) reloc->offset,
1661			  reloc->read_domains,
1662			  reloc->write_domain);
1663		return -EINVAL;
1664	}
1665
1666	if (reloc->write_domain) {
1667		target->flags |= EXEC_OBJECT_WRITE;
1668
1669		/*
1670		 * Sandybridge PPGTT errata: We need a global gtt mapping
1671		 * for MI and pipe_control writes because the gpu doesn't
1672		 * properly redirect them through the ppgtt for non_secure
1673		 * batchbuffers.
1674		 */
1675		if (reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
1676		    GRAPHICS_VER(eb->i915) == 6) {
1677			err = i915_vma_bind(target->vma,
1678					    target->vma->obj->cache_level,
1679					    PIN_GLOBAL, NULL);
1680			if (err)
1681				return err;
1682		}
1683	}
1684
1685	/*
1686	 * If the relocation already has the right value in it, no
1687	 * more work needs to be done.
1688	 */
1689	if (!DBG_FORCE_RELOC &&
1690	    gen8_canonical_addr(target->vma->node.start) == reloc->presumed_offset)
1691		return 0;
1692
1693	/* Check that the relocation address is valid... */
1694	if (unlikely(reloc->offset >
1695		     ev->vma->size - (eb->reloc_cache.use_64bit_reloc ? 8 : 4))) {
1696		drm_dbg(&i915->drm, "Relocation beyond object bounds: "
1697			  "target %d offset %d size %d.\n",
1698			  reloc->target_handle,
1699			  (int)reloc->offset,
1700			  (int)ev->vma->size);
1701		return -EINVAL;
1702	}
1703	if (unlikely(reloc->offset & 3)) {
1704		drm_dbg(&i915->drm, "Relocation not 4-byte aligned: "
1705			  "target %d offset %d.\n",
1706			  reloc->target_handle,
1707			  (int)reloc->offset);
1708		return -EINVAL;
1709	}
1710
1711	/*
1712	 * If we write into the object, we need to force the synchronisation
1713	 * barrier, either with an asynchronous clflush or if we executed the
1714	 * patching using the GPU (though that should be serialised by the
1715	 * timeline). To be completely sure, and since we are required to
1716	 * do relocations we are already stalling, disable the user's opt
1717	 * out of our synchronisation.
1718	 */
1719	ev->flags &= ~EXEC_OBJECT_ASYNC;
1720
1721	/* and update the user's relocation entry */
1722	return relocate_entry(ev->vma, reloc, eb, target->vma);
1723}
1724
1725static int eb_relocate_vma(struct i915_execbuffer *eb, struct eb_vma *ev)
1726{
1727#define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
1728	struct drm_i915_gem_relocation_entry stack[N_RELOC(512)];
1729	const struct drm_i915_gem_exec_object2 *entry = ev->exec;
1730	struct drm_i915_gem_relocation_entry __user *urelocs =
1731		u64_to_user_ptr(entry->relocs_ptr);
1732	unsigned long remain = entry->relocation_count;
1733
1734	if (unlikely(remain > N_RELOC(ULONG_MAX)))
1735		return -EINVAL;
1736
1737	/*
1738	 * We must check that the entire relocation array is safe
1739	 * to read. However, if the array is not writable the user loses
1740	 * the updated relocation values.
1741	 */
1742	if (unlikely(!access_ok(urelocs, remain * sizeof(*urelocs))))
1743		return -EFAULT;
1744
1745	do {
1746		struct drm_i915_gem_relocation_entry *r = stack;
1747		unsigned int count =
1748			min_t(unsigned long, remain, ARRAY_SIZE(stack));
1749		unsigned int copied;
1750
1751		/*
1752		 * This is the fast path and we cannot handle a pagefault
1753		 * whilst holding the struct mutex lest the user pass in the
1754		 * relocations contained within a mmaped bo. For in such a case
1755		 * we, the page fault handler would call i915_gem_fault() and
1756		 * we would try to acquire the struct mutex again. Obviously
1757		 * this is bad and so lockdep complains vehemently.
1758		 */
1759		pagefault_disable();
1760		copied = __copy_from_user_inatomic(r, urelocs, count * sizeof(r[0]));
1761		pagefault_enable();
1762		if (unlikely(copied)) {
1763			remain = -EFAULT;
1764			goto out;
1765		}
1766
1767		remain -= count;
1768		do {
1769			u64 offset = eb_relocate_entry(eb, ev, r);
1770
1771			if (likely(offset == 0)) {
1772			} else if ((s64)offset < 0) {
1773				remain = (int)offset;
1774				goto out;
1775			} else {
1776				/*
1777				 * Note that reporting an error now
1778				 * leaves everything in an inconsistent
1779				 * state as we have *already* changed
1780				 * the relocation value inside the
1781				 * object. As we have not changed the
1782				 * reloc.presumed_offset or will not
1783				 * change the execobject.offset, on the
1784				 * call we may not rewrite the value
1785				 * inside the object, leaving it
1786				 * dangling and causing a GPU hang. Unless
1787				 * userspace dynamically rebuilds the
1788				 * relocations on each execbuf rather than
1789				 * presume a static tree.
1790				 *
1791				 * We did previously check if the relocations
1792				 * were writable (access_ok), an error now
1793				 * would be a strange race with mprotect,
1794				 * having already demonstrated that we
1795				 * can read from this userspace address.
1796				 */
1797				offset = gen8_canonical_addr(offset & ~UPDATE);
1798				__put_user(offset,
1799					   &urelocs[r - stack].presumed_offset);
1800			}
1801		} while (r++, --count);
1802		urelocs += ARRAY_SIZE(stack);
1803	} while (remain);
1804out:
1805	reloc_cache_reset(&eb->reloc_cache, eb);
1806	return remain;
1807}
1808
1809static int
1810eb_relocate_vma_slow(struct i915_execbuffer *eb, struct eb_vma *ev)
1811{
1812	const struct drm_i915_gem_exec_object2 *entry = ev->exec;
1813	struct drm_i915_gem_relocation_entry *relocs =
1814		u64_to_ptr(typeof(*relocs), entry->relocs_ptr);
1815	unsigned int i;
1816	int err;
1817
1818	for (i = 0; i < entry->relocation_count; i++) {
1819		u64 offset = eb_relocate_entry(eb, ev, &relocs[i]);
1820
1821		if ((s64)offset < 0) {
1822			err = (int)offset;
1823			goto err;
1824		}
1825	}
1826	err = 0;
1827err:
1828	reloc_cache_reset(&eb->reloc_cache, eb);
1829	return err;
1830}
1831
1832static int check_relocations(const struct drm_i915_gem_exec_object2 *entry)
1833{
1834	const char __user *addr, *end;
1835	unsigned long size;
1836	char __maybe_unused c;
1837
1838	size = entry->relocation_count;
1839	if (size == 0)
1840		return 0;
1841
1842	if (size > N_RELOC(ULONG_MAX))
1843		return -EINVAL;
1844
1845	addr = u64_to_user_ptr(entry->relocs_ptr);
1846	size *= sizeof(struct drm_i915_gem_relocation_entry);
1847	if (!access_ok(addr, size))
1848		return -EFAULT;
1849
1850	end = addr + size;
1851	for (; addr < end; addr += PAGE_SIZE) {
1852		int err = __get_user(c, addr);
1853		if (err)
1854			return err;
1855	}
1856	return __get_user(c, end - 1);
1857}
1858
1859static int eb_copy_relocations(const struct i915_execbuffer *eb)
1860{
1861	struct drm_i915_gem_relocation_entry *relocs;
1862	const unsigned int count = eb->buffer_count;
1863	unsigned int i;
1864	int err;
1865
1866	for (i = 0; i < count; i++) {
1867		const unsigned int nreloc = eb->exec[i].relocation_count;
1868		struct drm_i915_gem_relocation_entry __user *urelocs;
1869		unsigned long size;
1870		unsigned long copied;
1871
1872		if (nreloc == 0)
1873			continue;
1874
1875		err = check_relocations(&eb->exec[i]);
1876		if (err)
1877			goto err;
1878
1879		urelocs = u64_to_user_ptr(eb->exec[i].relocs_ptr);
1880		size = nreloc * sizeof(*relocs);
1881
1882		relocs = kvmalloc_array(size, 1, GFP_KERNEL);
1883		if (!relocs) {
1884			err = -ENOMEM;
1885			goto err;
1886		}
1887
1888		/* copy_from_user is limited to < 4GiB */
1889		copied = 0;
1890		do {
1891			unsigned int len =
1892				min_t(u64, BIT_ULL(31), size - copied);
1893
1894			if (__copy_from_user((char *)relocs + copied,
1895					     (char __user *)urelocs + copied,
1896					     len))
1897				goto end;
1898
1899			copied += len;
1900		} while (copied < size);
1901
1902		/*
1903		 * As we do not update the known relocation offsets after
1904		 * relocating (due to the complexities in lock handling),
1905		 * we need to mark them as invalid now so that we force the
1906		 * relocation processing next time. Just in case the target
1907		 * object is evicted and then rebound into its old
1908		 * presumed_offset before the next execbuffer - if that
1909		 * happened we would make the mistake of assuming that the
1910		 * relocations were valid.
1911		 */
1912		if (!user_access_begin(urelocs, size))
1913			goto end;
1914
1915		for (copied = 0; copied < nreloc; copied++)
1916			unsafe_put_user(-1,
1917					&urelocs[copied].presumed_offset,
1918					end_user);
1919		user_access_end();
1920
1921		eb->exec[i].relocs_ptr = (uintptr_t)relocs;
1922	}
1923
1924	return 0;
1925
1926end_user:
1927	user_access_end();
1928end:
1929	kvfree(relocs);
1930	err = -EFAULT;
1931err:
1932	while (i--) {
1933		relocs = u64_to_ptr(typeof(*relocs), eb->exec[i].relocs_ptr);
1934		if (eb->exec[i].relocation_count)
1935			kvfree(relocs);
1936	}
1937	return err;
1938}
1939
1940static int eb_prefault_relocations(const struct i915_execbuffer *eb)
1941{
1942	const unsigned int count = eb->buffer_count;
 
1943	unsigned int i;
 
1944
1945	for (i = 0; i < count; i++) {
1946		int err;
1947
1948		err = check_relocations(&eb->exec[i]);
1949		if (err)
1950			return err;
1951	}
1952
1953	return 0;
1954}
1955
1956static int eb_reinit_userptr(struct i915_execbuffer *eb)
1957{
1958	const unsigned int count = eb->buffer_count;
1959	unsigned int i;
1960	int ret;
1961
1962	if (likely(!(eb->args->flags & __EXEC_USERPTR_USED)))
1963		return 0;
1964
1965	for (i = 0; i < count; i++) {
1966		struct eb_vma *ev = &eb->vma[i];
 
1967
1968		if (!i915_gem_object_is_userptr(ev->vma->obj))
1969			continue;
 
 
 
1970
1971		ret = i915_gem_object_userptr_submit_init(ev->vma->obj);
1972		if (ret)
1973			return ret;
1974
1975		ev->flags |= __EXEC_OBJECT_USERPTR_INIT;
1976	}
1977
1978	return 0;
1979}
1980
1981static noinline int eb_relocate_parse_slow(struct i915_execbuffer *eb,
1982					   struct i915_request *rq)
1983{
1984	bool have_copy = false;
1985	struct eb_vma *ev;
1986	int err = 0;
1987
1988repeat:
1989	if (signal_pending(current)) {
1990		err = -ERESTARTSYS;
1991		goto out;
1992	}
1993
1994	/* We may process another execbuffer during the unlock... */
1995	eb_release_vmas(eb, false, true);
1996	i915_gem_ww_ctx_fini(&eb->ww);
1997
1998	if (rq) {
1999		/* nonblocking is always false */
2000		if (i915_request_wait(rq, I915_WAIT_INTERRUPTIBLE,
2001				      MAX_SCHEDULE_TIMEOUT) < 0) {
2002			i915_request_put(rq);
2003			rq = NULL;
2004
2005			err = -EINTR;
2006			goto err_relock;
2007		}
2008
2009		i915_request_put(rq);
2010		rq = NULL;
2011	}
2012
2013	/*
2014	 * We take 3 passes through the slowpatch.
2015	 *
2016	 * 1 - we try to just prefault all the user relocation entries and
2017	 * then attempt to reuse the atomic pagefault disabled fast path again.
2018	 *
2019	 * 2 - we copy the user entries to a local buffer here outside of the
2020	 * local and allow ourselves to wait upon any rendering before
2021	 * relocations
2022	 *
2023	 * 3 - we already have a local copy of the relocation entries, but
2024	 * were interrupted (EAGAIN) whilst waiting for the objects, try again.
2025	 */
2026	if (!err) {
2027		err = eb_prefault_relocations(eb);
2028	} else if (!have_copy) {
2029		err = eb_copy_relocations(eb);
2030		have_copy = err == 0;
2031	} else {
2032		cond_resched();
2033		err = 0;
2034	}
2035
2036	if (!err)
2037		err = eb_reinit_userptr(eb);
2038
2039err_relock:
2040	i915_gem_ww_ctx_init(&eb->ww, true);
2041	if (err)
2042		goto out;
2043
2044	/* reacquire the objects */
2045repeat_validate:
2046	rq = eb_pin_engine(eb, false);
2047	if (IS_ERR(rq)) {
2048		err = PTR_ERR(rq);
2049		rq = NULL;
2050		goto err;
2051	}
2052
2053	/* We didn't throttle, should be NULL */
2054	GEM_WARN_ON(rq);
2055
2056	err = eb_validate_vmas(eb);
2057	if (err)
2058		goto err;
2059
2060	GEM_BUG_ON(!eb->batch);
2061
2062	list_for_each_entry(ev, &eb->relocs, reloc_link) {
2063		if (!have_copy) {
2064			pagefault_disable();
2065			err = eb_relocate_vma(eb, ev);
2066			pagefault_enable();
2067			if (err)
2068				break;
2069		} else {
2070			err = eb_relocate_vma_slow(eb, ev);
2071			if (err)
2072				break;
2073		}
2074	}
2075
2076	if (err == -EDEADLK)
2077		goto err;
2078
2079	if (err && !have_copy)
2080		goto repeat;
2081
2082	if (err)
2083		goto err;
2084
2085	/* as last step, parse the command buffer */
2086	err = eb_parse(eb);
2087	if (err)
2088		goto err;
2089
2090	/*
2091	 * Leave the user relocations as are, this is the painfully slow path,
2092	 * and we want to avoid the complication of dropping the lock whilst
2093	 * having buffers reserved in the aperture and so causing spurious
2094	 * ENOSPC for random operations.
2095	 */
2096
2097err:
2098	if (err == -EDEADLK) {
2099		eb_release_vmas(eb, false, false);
2100		err = i915_gem_ww_ctx_backoff(&eb->ww);
2101		if (!err)
2102			goto repeat_validate;
2103	}
2104
2105	if (err == -EAGAIN)
2106		goto repeat;
2107
2108out:
2109	if (have_copy) {
2110		const unsigned int count = eb->buffer_count;
2111		unsigned int i;
2112
2113		for (i = 0; i < count; i++) {
2114			const struct drm_i915_gem_exec_object2 *entry =
2115				&eb->exec[i];
2116			struct drm_i915_gem_relocation_entry *relocs;
2117
2118			if (!entry->relocation_count)
2119				continue;
2120
2121			relocs = u64_to_ptr(typeof(*relocs), entry->relocs_ptr);
2122			kvfree(relocs);
2123		}
2124	}
2125
2126	if (rq)
2127		i915_request_put(rq);
2128
2129	return err;
2130}
2131
2132static int eb_relocate_parse(struct i915_execbuffer *eb)
2133{
2134	int err;
2135	struct i915_request *rq = NULL;
2136	bool throttle = true;
2137
2138retry:
2139	rq = eb_pin_engine(eb, throttle);
2140	if (IS_ERR(rq)) {
2141		err = PTR_ERR(rq);
2142		rq = NULL;
2143		if (err != -EDEADLK)
2144			return err;
2145
2146		goto err;
2147	}
2148
2149	if (rq) {
2150		bool nonblock = eb->file->filp->f_flags & O_NONBLOCK;
2151
2152		/* Need to drop all locks now for throttling, take slowpath */
2153		err = i915_request_wait(rq, I915_WAIT_INTERRUPTIBLE, 0);
2154		if (err == -ETIME) {
2155			if (nonblock) {
2156				err = -EWOULDBLOCK;
2157				i915_request_put(rq);
2158				goto err;
2159			}
2160			goto slow;
2161		}
2162		i915_request_put(rq);
2163		rq = NULL;
2164	}
2165
2166	/* only throttle once, even if we didn't need to throttle */
2167	throttle = false;
2168
2169	err = eb_validate_vmas(eb);
2170	if (err == -EAGAIN)
2171		goto slow;
2172	else if (err)
2173		goto err;
2174
2175	/* The objects are in their final locations, apply the relocations. */
2176	if (eb->args->flags & __EXEC_HAS_RELOC) {
2177		struct eb_vma *ev;
2178
2179		list_for_each_entry(ev, &eb->relocs, reloc_link) {
2180			err = eb_relocate_vma(eb, ev);
2181			if (err)
2182				break;
2183		}
2184
2185		if (err == -EDEADLK)
2186			goto err;
2187		else if (err)
2188			goto slow;
2189	}
2190
2191	if (!err)
2192		err = eb_parse(eb);
2193
2194err:
2195	if (err == -EDEADLK) {
2196		eb_release_vmas(eb, false, false);
2197		err = i915_gem_ww_ctx_backoff(&eb->ww);
2198		if (!err)
2199			goto retry;
2200	}
2201
2202	return err;
2203
2204slow:
2205	err = eb_relocate_parse_slow(eb, rq);
2206	if (err)
2207		/*
2208		 * If the user expects the execobject.offset and
2209		 * reloc.presumed_offset to be an exact match,
2210		 * as for using NO_RELOC, then we cannot update
2211		 * the execobject.offset until we have completed
2212		 * relocation.
2213		 */
2214		eb->args->flags &= ~__EXEC_HAS_RELOC;
2215
2216	return err;
2217}
2218
2219static int eb_move_to_gpu(struct i915_execbuffer *eb)
2220{
2221	const unsigned int count = eb->buffer_count;
2222	unsigned int i = count;
2223	int err = 0;
2224
2225	while (i--) {
2226		struct eb_vma *ev = &eb->vma[i];
2227		struct i915_vma *vma = ev->vma;
2228		unsigned int flags = ev->flags;
2229		struct drm_i915_gem_object *obj = vma->obj;
2230
2231		assert_vma_held(vma);
2232
2233		if (flags & EXEC_OBJECT_CAPTURE) {
2234			struct i915_capture_list *capture;
2235
2236			capture = kmalloc(sizeof(*capture), GFP_KERNEL);
2237			if (capture) {
2238				capture->next = eb->request->capture_list;
2239				capture->vma = vma;
2240				eb->request->capture_list = capture;
2241			}
2242		}
2243
2244		/*
2245		 * If the GPU is not _reading_ through the CPU cache, we need
2246		 * to make sure that any writes (both previous GPU writes from
2247		 * before a change in snooping levels and normal CPU writes)
2248		 * caught in that cache are flushed to main memory.
2249		 *
2250		 * We want to say
2251		 *   obj->cache_dirty &&
2252		 *   !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ)
2253		 * but gcc's optimiser doesn't handle that as well and emits
2254		 * two jumps instead of one. Maybe one day...
2255		 */
2256		if (unlikely(obj->cache_dirty & ~obj->cache_coherent)) {
2257			if (i915_gem_clflush_object(obj, 0))
2258				flags &= ~EXEC_OBJECT_ASYNC;
2259		}
2260
2261		if (err == 0 && !(flags & EXEC_OBJECT_ASYNC)) {
2262			err = i915_request_await_object
2263				(eb->request, obj, flags & EXEC_OBJECT_WRITE);
2264		}
2265
2266		if (err == 0)
2267			err = i915_vma_move_to_active(vma, eb->request,
2268						      flags | __EXEC_OBJECT_NO_RESERVE);
 
 
2269	}
 
2270
2271#ifdef CONFIG_MMU_NOTIFIER
2272	if (!err && (eb->args->flags & __EXEC_USERPTR_USED)) {
2273		spin_lock(&eb->i915->mm.notifier_lock);
2274
2275		/*
2276		 * count is always at least 1, otherwise __EXEC_USERPTR_USED
2277		 * could not have been set
2278		 */
2279		for (i = 0; i < count; i++) {
2280			struct eb_vma *ev = &eb->vma[i];
2281			struct drm_i915_gem_object *obj = ev->vma->obj;
2282
2283			if (!i915_gem_object_is_userptr(obj))
2284				continue;
2285
2286			err = i915_gem_object_userptr_submit_done(obj);
2287			if (err)
2288				break;
2289		}
2290
2291		spin_unlock(&eb->i915->mm.notifier_lock);
2292	}
2293#endif
2294
2295	if (unlikely(err))
2296		goto err_skip;
2297
2298	/* Unconditionally flush any chipset caches (for streaming writes). */
2299	intel_gt_chipset_flush(eb->engine->gt);
2300	return 0;
2301
2302err_skip:
2303	i915_request_set_error_once(eb->request, err);
2304	return err;
2305}
2306
2307static int i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
2308{
2309	if (exec->flags & __I915_EXEC_ILLEGAL_FLAGS)
2310		return -EINVAL;
2311
2312	/* Kernel clipping was a DRI1 misfeature */
2313	if (!(exec->flags & (I915_EXEC_FENCE_ARRAY |
2314			     I915_EXEC_USE_EXTENSIONS))) {
2315		if (exec->num_cliprects || exec->cliprects_ptr)
2316			return -EINVAL;
2317	}
2318
2319	if (exec->DR4 == 0xffffffff) {
2320		DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
2321		exec->DR4 = 0;
2322	}
2323	if (exec->DR1 || exec->DR4)
2324		return -EINVAL;
2325
2326	if ((exec->batch_start_offset | exec->batch_len) & 0x7)
2327		return -EINVAL;
2328
2329	return 0;
2330}
2331
2332static int i915_reset_gen7_sol_offsets(struct i915_request *rq)
2333{
2334	u32 *cs;
2335	int i;
2336
2337	if (GRAPHICS_VER(rq->engine->i915) != 7 || rq->engine->id != RCS0) {
2338		drm_dbg(&rq->engine->i915->drm, "sol reset is gen7/rcs only\n");
2339		return -EINVAL;
2340	}
2341
2342	cs = intel_ring_begin(rq, 4 * 2 + 2);
2343	if (IS_ERR(cs))
2344		return PTR_ERR(cs);
2345
2346	*cs++ = MI_LOAD_REGISTER_IMM(4);
2347	for (i = 0; i < 4; i++) {
2348		*cs++ = i915_mmio_reg_offset(GEN7_SO_WRITE_OFFSET(i));
2349		*cs++ = 0;
2350	}
2351	*cs++ = MI_NOOP;
2352	intel_ring_advance(rq, cs);
2353
2354	return 0;
2355}
2356
2357static struct i915_vma *
2358shadow_batch_pin(struct i915_execbuffer *eb,
2359		 struct drm_i915_gem_object *obj,
2360		 struct i915_address_space *vm,
2361		 unsigned int flags)
2362{
2363	struct i915_vma *vma;
2364	int err;
2365
2366	vma = i915_vma_instance(obj, vm, NULL);
2367	if (IS_ERR(vma))
2368		return vma;
2369
2370	err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, flags);
2371	if (err)
2372		return ERR_PTR(err);
2373
2374	return vma;
2375}
2376
2377static struct i915_vma *eb_dispatch_secure(struct i915_execbuffer *eb, struct i915_vma *vma)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2378{
2379	/*
2380	 * snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
2381	 * batch" bit. Hence we need to pin secure batches into the global gtt.
2382	 * hsw should have this fixed, but bdw mucks it up again. */
2383	if (eb->batch_flags & I915_DISPATCH_SECURE)
2384		return i915_gem_object_ggtt_pin_ww(vma->obj, &eb->ww, NULL, 0, 0, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2385
2386	return NULL;
 
 
 
 
 
 
2387}
2388
2389static int eb_parse(struct i915_execbuffer *eb)
2390{
2391	struct drm_i915_private *i915 = eb->i915;
2392	struct intel_gt_buffer_pool_node *pool = eb->batch_pool;
2393	struct i915_vma *shadow, *trampoline, *batch;
2394	unsigned long len;
2395	int err;
2396
2397	if (!eb_use_cmdparser(eb)) {
2398		batch = eb_dispatch_secure(eb, eb->batch->vma);
2399		if (IS_ERR(batch))
2400			return PTR_ERR(batch);
2401
2402		goto secure_batch;
2403	}
2404
2405	len = eb->batch_len;
2406	if (!CMDPARSER_USES_GGTT(eb->i915)) {
2407		/*
2408		 * ppGTT backed shadow buffers must be mapped RO, to prevent
2409		 * post-scan tampering
2410		 */
2411		if (!eb->context->vm->has_read_only) {
2412			drm_dbg(&i915->drm,
2413				"Cannot prevent post-scan tampering without RO capable vm\n");
2414			return -EINVAL;
2415		}
2416	} else {
2417		len += I915_CMD_PARSER_TRAMPOLINE_SIZE;
2418	}
2419	if (unlikely(len < eb->batch_len)) /* last paranoid check of overflow */
2420		return -EINVAL;
2421
2422	if (!pool) {
2423		pool = intel_gt_get_buffer_pool(eb->engine->gt, len,
2424						I915_MAP_WB);
2425		if (IS_ERR(pool))
2426			return PTR_ERR(pool);
2427		eb->batch_pool = pool;
2428	}
2429
2430	err = i915_gem_object_lock(pool->obj, &eb->ww);
2431	if (err)
2432		goto err;
2433
2434	shadow = shadow_batch_pin(eb, pool->obj, eb->context->vm, PIN_USER);
2435	if (IS_ERR(shadow)) {
2436		err = PTR_ERR(shadow);
2437		goto err;
2438	}
2439	intel_gt_buffer_pool_mark_used(pool);
2440	i915_gem_object_set_readonly(shadow->obj);
2441	shadow->private = pool;
2442
2443	trampoline = NULL;
2444	if (CMDPARSER_USES_GGTT(eb->i915)) {
2445		trampoline = shadow;
2446
2447		shadow = shadow_batch_pin(eb, pool->obj,
2448					  &eb->engine->gt->ggtt->vm,
2449					  PIN_GLOBAL);
2450		if (IS_ERR(shadow)) {
2451			err = PTR_ERR(shadow);
2452			shadow = trampoline;
2453			goto err_shadow;
2454		}
2455		shadow->private = pool;
2456
2457		eb->batch_flags |= I915_DISPATCH_SECURE;
2458	}
2459
2460	batch = eb_dispatch_secure(eb, shadow);
2461	if (IS_ERR(batch)) {
2462		err = PTR_ERR(batch);
2463		goto err_trampoline;
2464	}
2465
2466	err = dma_resv_reserve_shared(shadow->resv, 1);
2467	if (err)
2468		goto err_trampoline;
2469
2470	err = intel_engine_cmd_parser(eb->engine,
2471				      eb->batch->vma,
2472				      eb->batch_start_offset,
2473				      eb->batch_len,
2474				      shadow, trampoline);
2475	if (err)
2476		goto err_unpin_batch;
2477
2478	eb->batch = &eb->vma[eb->buffer_count++];
2479	eb->batch->vma = i915_vma_get(shadow);
2480	eb->batch->flags = __EXEC_OBJECT_HAS_PIN;
2481
2482	eb->trampoline = trampoline;
2483	eb->batch_start_offset = 0;
2484
2485secure_batch:
2486	if (batch) {
2487		eb->batch = &eb->vma[eb->buffer_count++];
2488		eb->batch->flags = __EXEC_OBJECT_HAS_PIN;
2489		eb->batch->vma = i915_vma_get(batch);
2490	}
2491	return 0;
2492
2493err_unpin_batch:
2494	if (batch)
2495		i915_vma_unpin(batch);
2496err_trampoline:
2497	if (trampoline)
2498		i915_vma_unpin(trampoline);
2499err_shadow:
2500	i915_vma_unpin(shadow);
2501err:
 
2502	return err;
2503}
2504
 
 
 
 
 
 
 
 
 
 
 
 
2505static int eb_submit(struct i915_execbuffer *eb, struct i915_vma *batch)
2506{
2507	int err;
2508
2509	if (intel_context_nopreempt(eb->context))
2510		__set_bit(I915_FENCE_FLAG_NOPREEMPT, &eb->request->fence.flags);
2511
2512	err = eb_move_to_gpu(eb);
2513	if (err)
2514		return err;
2515
2516	if (eb->args->flags & I915_EXEC_GEN7_SOL_RESET) {
2517		err = i915_reset_gen7_sol_offsets(eb->request);
2518		if (err)
2519			return err;
2520	}
2521
2522	/*
2523	 * After we completed waiting for other engines (using HW semaphores)
2524	 * then we can signal that this request/batch is ready to run. This
2525	 * allows us to determine if the batch is still waiting on the GPU
2526	 * or actually running by checking the breadcrumb.
2527	 */
2528	if (eb->engine->emit_init_breadcrumb) {
2529		err = eb->engine->emit_init_breadcrumb(eb->request);
2530		if (err)
2531			return err;
2532	}
2533
2534	err = eb->engine->emit_bb_start(eb->request,
2535					batch->node.start +
2536					eb->batch_start_offset,
2537					eb->batch_len,
2538					eb->batch_flags);
2539	if (err)
2540		return err;
2541
2542	if (eb->trampoline) {
2543		GEM_BUG_ON(eb->batch_start_offset);
2544		err = eb->engine->emit_bb_start(eb->request,
2545						eb->trampoline->node.start +
2546						eb->batch_len,
2547						0, 0);
2548		if (err)
2549			return err;
2550	}
2551
 
 
 
2552	return 0;
2553}
2554
2555static int num_vcs_engines(const struct drm_i915_private *i915)
2556{
2557	return hweight_long(VDBOX_MASK(&i915->gt));
2558}
2559
2560/*
2561 * Find one BSD ring to dispatch the corresponding BSD command.
2562 * The engine index is returned.
2563 */
2564static unsigned int
2565gen8_dispatch_bsd_engine(struct drm_i915_private *dev_priv,
2566			 struct drm_file *file)
2567{
2568	struct drm_i915_file_private *file_priv = file->driver_priv;
2569
2570	/* Check whether the file_priv has already selected one ring. */
2571	if ((int)file_priv->bsd_engine < 0)
2572		file_priv->bsd_engine =
2573			get_random_int() % num_vcs_engines(dev_priv);
2574
2575	return file_priv->bsd_engine;
2576}
2577
2578static const enum intel_engine_id user_ring_map[] = {
2579	[I915_EXEC_DEFAULT]	= RCS0,
2580	[I915_EXEC_RENDER]	= RCS0,
2581	[I915_EXEC_BLT]		= BCS0,
2582	[I915_EXEC_BSD]		= VCS0,
2583	[I915_EXEC_VEBOX]	= VECS0
2584};
2585
2586static struct i915_request *eb_throttle(struct i915_execbuffer *eb, struct intel_context *ce)
2587{
2588	struct intel_ring *ring = ce->ring;
2589	struct intel_timeline *tl = ce->timeline;
2590	struct i915_request *rq;
2591
2592	/*
2593	 * Completely unscientific finger-in-the-air estimates for suitable
2594	 * maximum user request size (to avoid blocking) and then backoff.
2595	 */
2596	if (intel_ring_update_space(ring) >= PAGE_SIZE)
2597		return NULL;
2598
2599	/*
2600	 * Find a request that after waiting upon, there will be at least half
2601	 * the ring available. The hysteresis allows us to compete for the
2602	 * shared ring and should mean that we sleep less often prior to
2603	 * claiming our resources, but not so long that the ring completely
2604	 * drains before we can submit our next request.
2605	 */
2606	list_for_each_entry(rq, &tl->requests, link) {
2607		if (rq->ring != ring)
2608			continue;
2609
2610		if (__intel_ring_space(rq->postfix,
2611				       ring->emit, ring->size) > ring->size / 2)
2612			break;
2613	}
2614	if (&rq->link == &tl->requests)
2615		return NULL; /* weird, we will check again later for real */
2616
2617	return i915_request_get(rq);
2618}
2619
2620static struct i915_request *eb_pin_engine(struct i915_execbuffer *eb, bool throttle)
2621{
2622	struct intel_context *ce = eb->context;
2623	struct intel_timeline *tl;
2624	struct i915_request *rq = NULL;
2625	int err;
2626
2627	GEM_BUG_ON(eb->args->flags & __EXEC_ENGINE_PINNED);
 
 
 
 
 
 
2628
2629	if (unlikely(intel_context_is_banned(ce)))
2630		return ERR_PTR(-EIO);
2631
2632	/*
2633	 * Pinning the contexts may generate requests in order to acquire
2634	 * GGTT space, so do this first before we reserve a seqno for
2635	 * ourselves.
2636	 */
2637	err = intel_context_pin_ww(ce, &eb->ww);
2638	if (err)
2639		return ERR_PTR(err);
2640
2641	/*
2642	 * Take a local wakeref for preparing to dispatch the execbuf as
2643	 * we expect to access the hardware fairly frequently in the
2644	 * process, and require the engine to be kept awake between accesses.
2645	 * Upon dispatch, we acquire another prolonged wakeref that we hold
2646	 * until the timeline is idle, which in turn releases the wakeref
2647	 * taken on the engine, and the parent device.
2648	 */
2649	tl = intel_context_timeline_lock(ce);
2650	if (IS_ERR(tl)) {
2651		intel_context_unpin(ce);
2652		return ERR_CAST(tl);
2653	}
2654
2655	intel_context_enter(ce);
2656	if (throttle)
2657		rq = eb_throttle(eb, ce);
2658	intel_context_timeline_unlock(tl);
2659
2660	eb->args->flags |= __EXEC_ENGINE_PINNED;
2661	return rq;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2662}
2663
2664static void eb_unpin_engine(struct i915_execbuffer *eb)
2665{
2666	struct intel_context *ce = eb->context;
2667	struct intel_timeline *tl = ce->timeline;
2668
2669	if (!(eb->args->flags & __EXEC_ENGINE_PINNED))
2670		return;
2671
2672	eb->args->flags &= ~__EXEC_ENGINE_PINNED;
2673
2674	mutex_lock(&tl->mutex);
2675	intel_context_exit(ce);
2676	mutex_unlock(&tl->mutex);
2677
2678	intel_context_unpin(ce);
2679}
2680
2681static unsigned int
2682eb_select_legacy_ring(struct i915_execbuffer *eb)
 
 
2683{
2684	struct drm_i915_private *i915 = eb->i915;
2685	struct drm_i915_gem_execbuffer2 *args = eb->args;
2686	unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK;
2687
2688	if (user_ring_id != I915_EXEC_BSD &&
2689	    (args->flags & I915_EXEC_BSD_MASK)) {
2690		drm_dbg(&i915->drm,
2691			"execbuf with non bsd ring but with invalid "
2692			"bsd dispatch flags: %d\n", (int)(args->flags));
2693		return -1;
2694	}
2695
2696	if (user_ring_id == I915_EXEC_BSD && num_vcs_engines(i915) > 1) {
2697		unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK;
2698
2699		if (bsd_idx == I915_EXEC_BSD_DEFAULT) {
2700			bsd_idx = gen8_dispatch_bsd_engine(i915, eb->file);
2701		} else if (bsd_idx >= I915_EXEC_BSD_RING1 &&
2702			   bsd_idx <= I915_EXEC_BSD_RING2) {
2703			bsd_idx >>= I915_EXEC_BSD_SHIFT;
2704			bsd_idx--;
2705		} else {
2706			drm_dbg(&i915->drm,
2707				"execbuf with unknown bsd ring: %u\n",
2708				bsd_idx);
2709			return -1;
2710		}
2711
2712		return _VCS(bsd_idx);
2713	}
2714
2715	if (user_ring_id >= ARRAY_SIZE(user_ring_map)) {
2716		drm_dbg(&i915->drm, "execbuf with unknown ring: %u\n",
2717			user_ring_id);
2718		return -1;
2719	}
2720
2721	return user_ring_map[user_ring_id];
2722}
2723
2724static int
2725eb_select_engine(struct i915_execbuffer *eb)
 
 
2726{
2727	struct intel_context *ce;
2728	unsigned int idx;
2729	int err;
2730
2731	if (i915_gem_context_user_engines(eb->gem_context))
2732		idx = eb->args->flags & I915_EXEC_RING_MASK;
2733	else
2734		idx = eb_select_legacy_ring(eb);
2735
2736	ce = i915_gem_context_get_engine(eb->gem_context, idx);
2737	if (IS_ERR(ce))
2738		return PTR_ERR(ce);
2739
2740	intel_gt_pm_get(ce->engine->gt);
2741
2742	if (!test_bit(CONTEXT_ALLOC_BIT, &ce->flags)) {
2743		err = intel_context_alloc_state(ce);
2744		if (err)
2745			goto err;
2746	}
2747
2748	/*
2749	 * ABI: Before userspace accesses the GPU (e.g. execbuffer), report
2750	 * EIO if the GPU is already wedged.
2751	 */
2752	err = intel_gt_terminally_wedged(ce->engine->gt);
2753	if (err)
2754		goto err;
2755
2756	eb->context = ce;
2757	eb->engine = ce->engine;
2758
2759	/*
2760	 * Make sure engine pool stays alive even if we call intel_context_put
2761	 * during ww handling. The pool is destroyed when last pm reference
2762	 * is dropped, which breaks our -EDEADLK handling.
2763	 */
2764	return err;
2765
2766err:
2767	intel_gt_pm_put(ce->engine->gt);
2768	intel_context_put(ce);
2769	return err;
2770}
2771
2772static void
2773eb_put_engine(struct i915_execbuffer *eb)
2774{
2775	intel_gt_pm_put(eb->engine->gt);
2776	intel_context_put(eb->context);
2777}
2778
2779static void
2780__free_fence_array(struct eb_fence *fences, unsigned int n)
2781{
2782	while (n--) {
2783		drm_syncobj_put(ptr_mask_bits(fences[n].syncobj, 2));
2784		dma_fence_put(fences[n].dma_fence);
2785		kfree(fences[n].chain_fence);
2786	}
2787	kvfree(fences);
2788}
2789
2790static int
2791add_timeline_fence_array(struct i915_execbuffer *eb,
2792			 const struct drm_i915_gem_execbuffer_ext_timeline_fences *timeline_fences)
2793{
2794	struct drm_i915_gem_exec_fence __user *user_fences;
2795	u64 __user *user_values;
2796	struct eb_fence *f;
2797	u64 nfences;
2798	int err = 0;
2799
2800	nfences = timeline_fences->fence_count;
2801	if (!nfences)
2802		return 0;
2803
2804	/* Check multiplication overflow for access_ok() and kvmalloc_array() */
2805	BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long));
2806	if (nfences > min_t(unsigned long,
2807			    ULONG_MAX / sizeof(*user_fences),
2808			    SIZE_MAX / sizeof(*f)) - eb->num_fences)
2809		return -EINVAL;
2810
2811	user_fences = u64_to_user_ptr(timeline_fences->handles_ptr);
2812	if (!access_ok(user_fences, nfences * sizeof(*user_fences)))
2813		return -EFAULT;
2814
2815	user_values = u64_to_user_ptr(timeline_fences->values_ptr);
2816	if (!access_ok(user_values, nfences * sizeof(*user_values)))
2817		return -EFAULT;
2818
2819	f = krealloc(eb->fences,
2820		     (eb->num_fences + nfences) * sizeof(*f),
2821		     __GFP_NOWARN | GFP_KERNEL);
2822	if (!f)
2823		return -ENOMEM;
2824
2825	eb->fences = f;
2826	f += eb->num_fences;
 
 
2827
2828	BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) &
2829		     ~__I915_EXEC_FENCE_UNKNOWN_FLAGS);
2830
2831	while (nfences--) {
2832		struct drm_i915_gem_exec_fence user_fence;
2833		struct drm_syncobj *syncobj;
2834		struct dma_fence *fence = NULL;
2835		u64 point;
2836
2837		if (__copy_from_user(&user_fence,
2838				     user_fences++,
2839				     sizeof(user_fence)))
2840			return -EFAULT;
2841
2842		if (user_fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS)
2843			return -EINVAL;
2844
2845		if (__get_user(point, user_values++))
2846			return -EFAULT;
2847
2848		syncobj = drm_syncobj_find(eb->file, user_fence.handle);
2849		if (!syncobj) {
2850			DRM_DEBUG("Invalid syncobj handle provided\n");
2851			return -ENOENT;
2852		}
2853
2854		fence = drm_syncobj_fence_get(syncobj);
2855
2856		if (!fence && user_fence.flags &&
2857		    !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) {
2858			DRM_DEBUG("Syncobj handle has no fence\n");
2859			drm_syncobj_put(syncobj);
2860			return -EINVAL;
2861		}
2862
2863		if (fence)
2864			err = dma_fence_chain_find_seqno(&fence, point);
2865
2866		if (err && !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) {
2867			DRM_DEBUG("Syncobj handle missing requested point %llu\n", point);
2868			dma_fence_put(fence);
2869			drm_syncobj_put(syncobj);
2870			return err;
2871		}
2872
2873		/*
2874		 * A point might have been signaled already and
2875		 * garbage collected from the timeline. In this case
2876		 * just ignore the point and carry on.
2877		 */
2878		if (!fence && !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) {
2879			drm_syncobj_put(syncobj);
2880			continue;
2881		}
2882
2883		/*
2884		 * For timeline syncobjs we need to preallocate chains for
2885		 * later signaling.
2886		 */
2887		if (point != 0 && user_fence.flags & I915_EXEC_FENCE_SIGNAL) {
2888			/*
2889			 * Waiting and signaling the same point (when point !=
2890			 * 0) would break the timeline.
2891			 */
2892			if (user_fence.flags & I915_EXEC_FENCE_WAIT) {
2893				DRM_DEBUG("Trying to wait & signal the same timeline point.\n");
2894				dma_fence_put(fence);
2895				drm_syncobj_put(syncobj);
2896				return -EINVAL;
2897			}
2898
2899			f->chain_fence =
2900				kmalloc(sizeof(*f->chain_fence),
2901					GFP_KERNEL);
2902			if (!f->chain_fence) {
2903				drm_syncobj_put(syncobj);
2904				dma_fence_put(fence);
2905				return -ENOMEM;
2906			}
2907		} else {
2908			f->chain_fence = NULL;
2909		}
2910
2911		f->syncobj = ptr_pack_bits(syncobj, user_fence.flags, 2);
2912		f->dma_fence = fence;
2913		f->value = point;
2914		f++;
2915		eb->num_fences++;
2916	}
2917
2918	return 0;
2919}
2920
2921static int add_fence_array(struct i915_execbuffer *eb)
2922{
2923	struct drm_i915_gem_execbuffer2 *args = eb->args;
2924	struct drm_i915_gem_exec_fence __user *user;
2925	unsigned long num_fences = args->num_cliprects;
2926	struct eb_fence *f;
2927
2928	if (!(args->flags & I915_EXEC_FENCE_ARRAY))
2929		return 0;
2930
2931	if (!num_fences)
2932		return 0;
2933
2934	/* Check multiplication overflow for access_ok() and kvmalloc_array() */
2935	BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long));
2936	if (num_fences > min_t(unsigned long,
2937			       ULONG_MAX / sizeof(*user),
2938			       SIZE_MAX / sizeof(*f) - eb->num_fences))
2939		return -EINVAL;
2940
2941	user = u64_to_user_ptr(args->cliprects_ptr);
2942	if (!access_ok(user, num_fences * sizeof(*user)))
2943		return -EFAULT;
2944
2945	f = krealloc(eb->fences,
2946		     (eb->num_fences + num_fences) * sizeof(*f),
2947		     __GFP_NOWARN | GFP_KERNEL);
2948	if (!f)
2949		return -ENOMEM;
2950
2951	eb->fences = f;
2952	f += eb->num_fences;
2953	while (num_fences--) {
2954		struct drm_i915_gem_exec_fence user_fence;
2955		struct drm_syncobj *syncobj;
2956		struct dma_fence *fence = NULL;
2957
2958		if (__copy_from_user(&user_fence, user++, sizeof(user_fence)))
2959			return -EFAULT;
2960
2961		if (user_fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS)
2962			return -EINVAL;
2963
2964		syncobj = drm_syncobj_find(eb->file, user_fence.handle);
2965		if (!syncobj) {
2966			DRM_DEBUG("Invalid syncobj handle provided\n");
2967			return -ENOENT;
2968		}
2969
2970		if (user_fence.flags & I915_EXEC_FENCE_WAIT) {
2971			fence = drm_syncobj_fence_get(syncobj);
2972			if (!fence) {
2973				DRM_DEBUG("Syncobj handle has no fence\n");
2974				drm_syncobj_put(syncobj);
2975				return -EINVAL;
2976			}
2977		}
2978
2979		BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) &
2980			     ~__I915_EXEC_FENCE_UNKNOWN_FLAGS);
2981
2982		f->syncobj = ptr_pack_bits(syncobj, user_fence.flags, 2);
2983		f->dma_fence = fence;
2984		f->value = 0;
2985		f->chain_fence = NULL;
2986		f++;
2987		eb->num_fences++;
2988	}
2989
2990	return 0;
 
 
 
 
2991}
2992
2993static void put_fence_array(struct eb_fence *fences, int num_fences)
 
 
2994{
2995	if (fences)
2996		__free_fence_array(fences, num_fences);
2997}
2998
2999static int
3000await_fence_array(struct i915_execbuffer *eb)
 
3001{
 
3002	unsigned int n;
3003	int err;
3004
3005	for (n = 0; n < eb->num_fences; n++) {
3006		struct drm_syncobj *syncobj;
 
3007		unsigned int flags;
3008
3009		syncobj = ptr_unpack_bits(eb->fences[n].syncobj, &flags, 2);
 
 
3010
3011		if (!eb->fences[n].dma_fence)
3012			continue;
 
3013
3014		err = i915_request_await_dma_fence(eb->request,
3015						   eb->fences[n].dma_fence);
3016		if (err < 0)
3017			return err;
3018	}
3019
3020	return 0;
3021}
3022
3023static void signal_fence_array(const struct i915_execbuffer *eb)
 
 
3024{
 
3025	struct dma_fence * const fence = &eb->request->fence;
3026	unsigned int n;
3027
3028	for (n = 0; n < eb->num_fences; n++) {
3029		struct drm_syncobj *syncobj;
3030		unsigned int flags;
3031
3032		syncobj = ptr_unpack_bits(eb->fences[n].syncobj, &flags, 2);
3033		if (!(flags & I915_EXEC_FENCE_SIGNAL))
3034			continue;
3035
3036		if (eb->fences[n].chain_fence) {
3037			drm_syncobj_add_point(syncobj,
3038					      eb->fences[n].chain_fence,
3039					      fence,
3040					      eb->fences[n].value);
3041			/*
3042			 * The chain's ownership is transferred to the
3043			 * timeline.
3044			 */
3045			eb->fences[n].chain_fence = NULL;
3046		} else {
3047			drm_syncobj_replace_fence(syncobj, fence);
3048		}
3049	}
3050}
3051
3052static int
3053parse_timeline_fences(struct i915_user_extension __user *ext, void *data)
3054{
3055	struct i915_execbuffer *eb = data;
3056	struct drm_i915_gem_execbuffer_ext_timeline_fences timeline_fences;
3057
3058	if (copy_from_user(&timeline_fences, ext, sizeof(timeline_fences)))
3059		return -EFAULT;
3060
3061	return add_timeline_fence_array(eb, &timeline_fences);
3062}
3063
3064static void retire_requests(struct intel_timeline *tl, struct i915_request *end)
3065{
3066	struct i915_request *rq, *rn;
3067
3068	list_for_each_entry_safe(rq, rn, &tl->requests, link)
3069		if (rq == end || !i915_request_retire(rq))
3070			break;
3071}
3072
3073static int eb_request_add(struct i915_execbuffer *eb, int err)
3074{
3075	struct i915_request *rq = eb->request;
3076	struct intel_timeline * const tl = i915_request_timeline(rq);
3077	struct i915_sched_attr attr = {};
3078	struct i915_request *prev;
3079
3080	lockdep_assert_held(&tl->mutex);
3081	lockdep_unpin_lock(&tl->mutex, rq->cookie);
3082
3083	trace_i915_request_add(rq);
3084
3085	prev = __i915_request_commit(rq);
3086
3087	/* Check that the context wasn't destroyed before submission */
3088	if (likely(!intel_context_is_closed(eb->context))) {
3089		attr = eb->gem_context->sched;
3090	} else {
3091		/* Serialise with context_close via the add_to_timeline */
3092		i915_request_set_error_once(rq, -ENOENT);
3093		__i915_request_skip(rq);
3094		err = -ENOENT; /* override any transient errors */
3095	}
3096
3097	__i915_request_queue(rq, &attr);
3098
3099	/* Try to clean up the client's timeline after submitting the request */
3100	if (prev)
3101		retire_requests(tl, prev);
3102
3103	mutex_unlock(&tl->mutex);
3104
3105	return err;
3106}
3107
3108static const i915_user_extension_fn execbuf_extensions[] = {
3109	[DRM_I915_GEM_EXECBUFFER_EXT_TIMELINE_FENCES] = parse_timeline_fences,
3110};
3111
3112static int
3113parse_execbuf2_extensions(struct drm_i915_gem_execbuffer2 *args,
3114			  struct i915_execbuffer *eb)
3115{
3116	if (!(args->flags & I915_EXEC_USE_EXTENSIONS))
3117		return 0;
3118
3119	/* The execbuf2 extension mechanism reuses cliprects_ptr. So we cannot
3120	 * have another flag also using it at the same time.
3121	 */
3122	if (eb->args->flags & I915_EXEC_FENCE_ARRAY)
3123		return -EINVAL;
3124
3125	if (args->num_cliprects != 0)
3126		return -EINVAL;
3127
3128	return i915_user_extensions(u64_to_user_ptr(args->cliprects_ptr),
3129				    execbuf_extensions,
3130				    ARRAY_SIZE(execbuf_extensions),
3131				    eb);
3132}
3133
3134static int
3135i915_gem_do_execbuffer(struct drm_device *dev,
3136		       struct drm_file *file,
3137		       struct drm_i915_gem_execbuffer2 *args,
3138		       struct drm_i915_gem_exec_object2 *exec)
 
3139{
3140	struct drm_i915_private *i915 = to_i915(dev);
3141	struct i915_execbuffer eb;
3142	struct dma_fence *in_fence = NULL;
3143	struct sync_file *out_fence = NULL;
3144	struct i915_vma *batch;
3145	int out_fence_fd = -1;
3146	int err;
3147
3148	BUILD_BUG_ON(__EXEC_INTERNAL_FLAGS & ~__I915_EXEC_ILLEGAL_FLAGS);
3149	BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS &
3150		     ~__EXEC_OBJECT_UNKNOWN_FLAGS);
3151
3152	eb.i915 = i915;
3153	eb.file = file;
3154	eb.args = args;
3155	if (DBG_FORCE_RELOC || !(args->flags & I915_EXEC_NO_RELOC))
3156		args->flags |= __EXEC_HAS_RELOC;
3157
3158	eb.exec = exec;
3159	eb.vma = (struct eb_vma *)(exec + args->buffer_count + 1);
3160	eb.vma[0].vma = NULL;
3161	eb.reloc_pool = eb.batch_pool = NULL;
3162	eb.reloc_context = NULL;
3163
3164	eb.invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
3165	reloc_cache_init(&eb.reloc_cache, eb.i915);
3166
3167	eb.buffer_count = args->buffer_count;
3168	eb.batch_start_offset = args->batch_start_offset;
3169	eb.batch_len = args->batch_len;
3170	eb.trampoline = NULL;
3171
3172	eb.fences = NULL;
3173	eb.num_fences = 0;
3174
3175	eb.batch_flags = 0;
3176	if (args->flags & I915_EXEC_SECURE) {
3177		if (GRAPHICS_VER(i915) >= 11)
3178			return -ENODEV;
3179
3180		/* Return -EPERM to trigger fallback code on old binaries. */
3181		if (!HAS_SECURE_BATCHES(i915))
3182			return -EPERM;
3183
3184		if (!drm_is_current_master(file) || !capable(CAP_SYS_ADMIN))
3185			return -EPERM;
3186
3187		eb.batch_flags |= I915_DISPATCH_SECURE;
3188	}
3189	if (args->flags & I915_EXEC_IS_PINNED)
3190		eb.batch_flags |= I915_DISPATCH_PINNED;
3191
3192	err = parse_execbuf2_extensions(args, &eb);
3193	if (err)
3194		goto err_ext;
3195
3196	err = add_fence_array(&eb);
3197	if (err)
3198		goto err_ext;
3199
3200#define IN_FENCES (I915_EXEC_FENCE_IN | I915_EXEC_FENCE_SUBMIT)
3201	if (args->flags & IN_FENCES) {
3202		if ((args->flags & IN_FENCES) == IN_FENCES)
3203			return -EINVAL;
3204
3205		in_fence = sync_file_get_fence(lower_32_bits(args->rsvd2));
3206		if (!in_fence) {
3207			err = -EINVAL;
3208			goto err_ext;
3209		}
3210	}
3211#undef IN_FENCES
3212
3213	if (args->flags & I915_EXEC_FENCE_OUT) {
3214		out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
3215		if (out_fence_fd < 0) {
3216			err = out_fence_fd;
3217			goto err_in_fence;
3218		}
3219	}
3220
3221	err = eb_create(&eb);
3222	if (err)
3223		goto err_out_fence;
3224
3225	GEM_BUG_ON(!eb.lut_size);
3226
3227	err = eb_select_context(&eb);
3228	if (unlikely(err))
3229		goto err_destroy;
3230
3231	err = eb_select_engine(&eb);
3232	if (unlikely(err))
3233		goto err_context;
3234
3235	err = eb_lookup_vmas(&eb);
3236	if (err) {
3237		eb_release_vmas(&eb, true, true);
3238		goto err_engine;
3239	}
3240
3241	i915_gem_ww_ctx_init(&eb.ww, true);
3242
3243	err = eb_relocate_parse(&eb);
3244	if (err) {
3245		/*
3246		 * If the user expects the execobject.offset and
3247		 * reloc.presumed_offset to be an exact match,
3248		 * as for using NO_RELOC, then we cannot update
3249		 * the execobject.offset until we have completed
3250		 * relocation.
3251		 */
3252		args->flags &= ~__EXEC_HAS_RELOC;
3253		goto err_vma;
3254	}
3255
3256	ww_acquire_done(&eb.ww.ctx);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3257
 
 
 
 
 
 
 
 
3258	batch = eb.batch->vma;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3259
3260	/* All GPU relocation batches must be submitted prior to the user rq */
3261	GEM_BUG_ON(eb.reloc_cache.rq);
3262
3263	/* Allocate a request for this batch buffer nice and early. */
3264	eb.request = i915_request_create(eb.context);
3265	if (IS_ERR(eb.request)) {
3266		err = PTR_ERR(eb.request);
3267		goto err_vma;
3268	}
3269
3270	if (in_fence) {
3271		if (args->flags & I915_EXEC_FENCE_SUBMIT)
3272			err = i915_request_await_execution(eb.request,
3273							   in_fence,
3274							   eb.engine->bond_execute);
3275		else
3276			err = i915_request_await_dma_fence(eb.request,
3277							   in_fence);
3278		if (err < 0)
3279			goto err_request;
3280	}
3281
3282	if (eb.fences) {
3283		err = await_fence_array(&eb);
3284		if (err)
3285			goto err_request;
3286	}
3287
3288	if (out_fence_fd != -1) {
3289		out_fence = sync_file_create(&eb.request->fence);
3290		if (!out_fence) {
3291			err = -ENOMEM;
3292			goto err_request;
3293		}
3294	}
3295
3296	/*
3297	 * Whilst this request exists, batch_obj will be on the
3298	 * active_list, and so will hold the active reference. Only when this
3299	 * request is retired will the the batch_obj be moved onto the
3300	 * inactive_list and lose its active reference. Hence we do not need
3301	 * to explicitly hold another reference here.
3302	 */
3303	eb.request->batch = batch;
3304	if (eb.batch_pool)
3305		intel_gt_buffer_pool_mark_active(eb.batch_pool, eb.request);
3306
3307	trace_i915_request_queue(eb.request, eb.batch_flags);
3308	err = eb_submit(&eb, batch);
3309
3310err_request:
 
3311	i915_request_get(eb.request);
3312	err = eb_request_add(&eb, err);
3313
3314	if (eb.fences)
3315		signal_fence_array(&eb);
3316
3317	if (out_fence) {
3318		if (err == 0) {
3319			fd_install(out_fence_fd, out_fence->file);
3320			args->rsvd2 &= GENMASK_ULL(31, 0); /* keep in-fence */
3321			args->rsvd2 |= (u64)out_fence_fd << 32;
3322			out_fence_fd = -1;
3323		} else {
3324			fput(out_fence->file);
3325		}
3326	}
3327	i915_request_put(eb.request);
3328
 
 
 
 
 
 
3329err_vma:
3330	eb_release_vmas(&eb, true, true);
3331	if (eb.trampoline)
3332		i915_vma_unpin(eb.trampoline);
3333	WARN_ON(err == -EDEADLK);
3334	i915_gem_ww_ctx_fini(&eb.ww);
3335
3336	if (eb.batch_pool)
3337		intel_gt_buffer_pool_put(eb.batch_pool);
3338	if (eb.reloc_pool)
3339		intel_gt_buffer_pool_put(eb.reloc_pool);
3340	if (eb.reloc_context)
3341		intel_context_put(eb.reloc_context);
3342err_engine:
3343	eb_put_engine(&eb);
3344err_context:
3345	i915_gem_context_put(eb.gem_context);
3346err_destroy:
3347	eb_destroy(&eb);
3348err_out_fence:
3349	if (out_fence_fd != -1)
3350		put_unused_fd(out_fence_fd);
3351err_in_fence:
3352	dma_fence_put(in_fence);
3353err_ext:
3354	put_fence_array(eb.fences, eb.num_fences);
3355	return err;
3356}
3357
3358static size_t eb_element_size(void)
3359{
3360	return sizeof(struct drm_i915_gem_exec_object2) + sizeof(struct eb_vma);
3361}
3362
3363static bool check_buffer_count(size_t count)
3364{
3365	const size_t sz = eb_element_size();
3366
3367	/*
3368	 * When using LUT_HANDLE, we impose a limit of INT_MAX for the lookup
3369	 * array size (see eb_create()). Otherwise, we can accept an array as
3370	 * large as can be addressed (though use large arrays at your peril)!
3371	 */
3372
3373	return !(count < 1 || count > INT_MAX || count > SIZE_MAX / sz - 1);
3374}
3375
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3376int
3377i915_gem_execbuffer2_ioctl(struct drm_device *dev, void *data,
3378			   struct drm_file *file)
3379{
3380	struct drm_i915_private *i915 = to_i915(dev);
3381	struct drm_i915_gem_execbuffer2 *args = data;
3382	struct drm_i915_gem_exec_object2 *exec2_list;
 
3383	const size_t count = args->buffer_count;
3384	int err;
3385
3386	if (!check_buffer_count(count)) {
3387		drm_dbg(&i915->drm, "execbuf2 with %zd buffers\n", count);
3388		return -EINVAL;
3389	}
3390
3391	err = i915_gem_check_execbuffer(args);
3392	if (err)
3393		return err;
3394
3395	/* Allocate extra slots for use by the command parser */
3396	exec2_list = kvmalloc_array(count + 2, eb_element_size(),
3397				    __GFP_NOWARN | GFP_KERNEL);
3398	if (exec2_list == NULL) {
3399		drm_dbg(&i915->drm, "Failed to allocate exec list for %zd buffers\n",
3400			count);
3401		return -ENOMEM;
3402	}
3403	if (copy_from_user(exec2_list,
3404			   u64_to_user_ptr(args->buffers_ptr),
3405			   sizeof(*exec2_list) * count)) {
3406		drm_dbg(&i915->drm, "copy %zd exec entries failed\n", count);
3407		kvfree(exec2_list);
3408		return -EFAULT;
3409	}
3410
3411	err = i915_gem_do_execbuffer(dev, file, args, exec2_list);
 
 
 
 
 
 
 
 
3412
3413	/*
3414	 * Now that we have begun execution of the batchbuffer, we ignore
3415	 * any new error after this point. Also given that we have already
3416	 * updated the associated relocations, we try to write out the current
3417	 * object locations irrespective of any error.
3418	 */
3419	if (args->flags & __EXEC_HAS_RELOC) {
3420		struct drm_i915_gem_exec_object2 __user *user_exec_list =
3421			u64_to_user_ptr(args->buffers_ptr);
3422		unsigned int i;
3423
3424		/* Copy the new buffer offsets back to the user's exec list. */
3425		/*
3426		 * Note: count * sizeof(*user_exec_list) does not overflow,
3427		 * because we checked 'count' in check_buffer_count().
3428		 *
3429		 * And this range already got effectively checked earlier
3430		 * when we did the "copy_from_user()" above.
3431		 */
3432		if (!user_write_access_begin(user_exec_list,
3433					     count * sizeof(*user_exec_list)))
3434			goto end;
3435
3436		for (i = 0; i < args->buffer_count; i++) {
3437			if (!(exec2_list[i].offset & UPDATE))
3438				continue;
3439
3440			exec2_list[i].offset =
3441				gen8_canonical_addr(exec2_list[i].offset & PIN_OFFSET_MASK);
3442			unsafe_put_user(exec2_list[i].offset,
3443					&user_exec_list[i].offset,
3444					end_user);
3445		}
3446end_user:
3447		user_write_access_end();
3448end:;
3449	}
3450
3451	args->flags &= ~__I915_EXEC_UNKNOWN_FLAGS;
 
3452	kvfree(exec2_list);
3453	return err;
3454}
3455
3456#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
3457#include "selftests/i915_gem_execbuffer.c"
3458#endif