Loading...
Note: File does not exist in v4.6.
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#include <drm/i915_drm.h>
14
15#include "display/intel_frontbuffer.h"
16
17#include "gem/i915_gem_ioctls.h"
18#include "gt/intel_context.h"
19#include "gt/intel_engine_pool.h"
20#include "gt/intel_gt.h"
21#include "gt/intel_gt_pm.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_trace.h"
28
29enum {
30 FORCE_CPU_RELOC = 1,
31 FORCE_GTT_RELOC,
32 FORCE_GPU_RELOC,
33#define DBG_FORCE_RELOC 0 /* choose one of the above! */
34};
35
36#define __EXEC_OBJECT_HAS_REF BIT(31)
37#define __EXEC_OBJECT_HAS_PIN BIT(30)
38#define __EXEC_OBJECT_HAS_FENCE BIT(29)
39#define __EXEC_OBJECT_NEEDS_MAP BIT(28)
40#define __EXEC_OBJECT_NEEDS_BIAS BIT(27)
41#define __EXEC_OBJECT_INTERNAL_FLAGS (~0u << 27) /* all of the above */
42#define __EXEC_OBJECT_RESERVED (__EXEC_OBJECT_HAS_PIN | __EXEC_OBJECT_HAS_FENCE)
43
44#define __EXEC_HAS_RELOC BIT(31)
45#define __EXEC_VALIDATED BIT(30)
46#define __EXEC_INTERNAL_FLAGS (~0u << 30)
47#define UPDATE PIN_OFFSET_FIXED
48
49#define BATCH_OFFSET_BIAS (256*1024)
50
51#define __I915_EXEC_ILLEGAL_FLAGS \
52 (__I915_EXEC_UNKNOWN_FLAGS | \
53 I915_EXEC_CONSTANTS_MASK | \
54 I915_EXEC_RESOURCE_STREAMER)
55
56/* Catch emission of unexpected errors for CI! */
57#if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
58#undef EINVAL
59#define EINVAL ({ \
60 DRM_DEBUG_DRIVER("EINVAL at %s:%d\n", __func__, __LINE__); \
61 22; \
62})
63#endif
64
65/**
66 * DOC: User command execution
67 *
68 * Userspace submits commands to be executed on the GPU as an instruction
69 * stream within a GEM object we call a batchbuffer. This instructions may
70 * refer to other GEM objects containing auxiliary state such as kernels,
71 * samplers, render targets and even secondary batchbuffers. Userspace does
72 * not know where in the GPU memory these objects reside and so before the
73 * batchbuffer is passed to the GPU for execution, those addresses in the
74 * batchbuffer and auxiliary objects are updated. This is known as relocation,
75 * or patching. To try and avoid having to relocate each object on the next
76 * execution, userspace is told the location of those objects in this pass,
77 * but this remains just a hint as the kernel may choose a new location for
78 * any object in the future.
79 *
80 * At the level of talking to the hardware, submitting a batchbuffer for the
81 * GPU to execute is to add content to a buffer from which the HW
82 * command streamer is reading.
83 *
84 * 1. Add a command to load the HW context. For Logical Ring Contexts, i.e.
85 * Execlists, this command is not placed on the same buffer as the
86 * remaining items.
87 *
88 * 2. Add a command to invalidate caches to the buffer.
89 *
90 * 3. Add a batchbuffer start command to the buffer; the start command is
91 * essentially a token together with the GPU address of the batchbuffer
92 * to be executed.
93 *
94 * 4. Add a pipeline flush to the buffer.
95 *
96 * 5. Add a memory write command to the buffer to record when the GPU
97 * is done executing the batchbuffer. The memory write writes the
98 * global sequence number of the request, ``i915_request::global_seqno``;
99 * the i915 driver uses the current value in the register to determine
100 * if the GPU has completed the batchbuffer.
101 *
102 * 6. Add a user interrupt command to the buffer. This command instructs
103 * the GPU to issue an interrupt when the command, pipeline flush and
104 * memory write are completed.
105 *
106 * 7. Inform the hardware of the additional commands added to the buffer
107 * (by updating the tail pointer).
108 *
109 * Processing an execbuf ioctl is conceptually split up into a few phases.
110 *
111 * 1. Validation - Ensure all the pointers, handles and flags are valid.
112 * 2. Reservation - Assign GPU address space for every object
113 * 3. Relocation - Update any addresses to point to the final locations
114 * 4. Serialisation - Order the request with respect to its dependencies
115 * 5. Construction - Construct a request to execute the batchbuffer
116 * 6. Submission (at some point in the future execution)
117 *
118 * Reserving resources for the execbuf is the most complicated phase. We
119 * neither want to have to migrate the object in the address space, nor do
120 * we want to have to update any relocations pointing to this object. Ideally,
121 * we want to leave the object where it is and for all the existing relocations
122 * to match. If the object is given a new address, or if userspace thinks the
123 * object is elsewhere, we have to parse all the relocation entries and update
124 * the addresses. Userspace can set the I915_EXEC_NORELOC flag to hint that
125 * all the target addresses in all of its objects match the value in the
126 * relocation entries and that they all match the presumed offsets given by the
127 * list of execbuffer objects. Using this knowledge, we know that if we haven't
128 * moved any buffers, all the relocation entries are valid and we can skip
129 * the update. (If userspace is wrong, the likely outcome is an impromptu GPU
130 * hang.) The requirement for using I915_EXEC_NO_RELOC are:
131 *
132 * The addresses written in the objects must match the corresponding
133 * reloc.presumed_offset which in turn must match the corresponding
134 * execobject.offset.
135 *
136 * Any render targets written to in the batch must be flagged with
137 * EXEC_OBJECT_WRITE.
138 *
139 * To avoid stalling, execobject.offset should match the current
140 * address of that object within the active context.
141 *
142 * The reservation is done is multiple phases. First we try and keep any
143 * object already bound in its current location - so as long as meets the
144 * constraints imposed by the new execbuffer. Any object left unbound after the
145 * first pass is then fitted into any available idle space. If an object does
146 * not fit, all objects are removed from the reservation and the process rerun
147 * after sorting the objects into a priority order (more difficult to fit
148 * objects are tried first). Failing that, the entire VM is cleared and we try
149 * to fit the execbuf once last time before concluding that it simply will not
150 * fit.
151 *
152 * A small complication to all of this is that we allow userspace not only to
153 * specify an alignment and a size for the object in the address space, but
154 * we also allow userspace to specify the exact offset. This objects are
155 * simpler to place (the location is known a priori) all we have to do is make
156 * sure the space is available.
157 *
158 * Once all the objects are in place, patching up the buried pointers to point
159 * to the final locations is a fairly simple job of walking over the relocation
160 * entry arrays, looking up the right address and rewriting the value into
161 * the object. Simple! ... The relocation entries are stored in user memory
162 * and so to access them we have to copy them into a local buffer. That copy
163 * has to avoid taking any pagefaults as they may lead back to a GEM object
164 * requiring the struct_mutex (i.e. recursive deadlock). So once again we split
165 * the relocation into multiple passes. First we try to do everything within an
166 * atomic context (avoid the pagefaults) which requires that we never wait. If
167 * we detect that we may wait, or if we need to fault, then we have to fallback
168 * to a slower path. The slowpath has to drop the mutex. (Can you hear alarm
169 * bells yet?) Dropping the mutex means that we lose all the state we have
170 * built up so far for the execbuf and we must reset any global data. However,
171 * we do leave the objects pinned in their final locations - which is a
172 * potential issue for concurrent execbufs. Once we have left the mutex, we can
173 * allocate and copy all the relocation entries into a large array at our
174 * leisure, reacquire the mutex, reclaim all the objects and other state and
175 * then proceed to update any incorrect addresses with the objects.
176 *
177 * As we process the relocation entries, we maintain a record of whether the
178 * object is being written to. Using NORELOC, we expect userspace to provide
179 * this information instead. We also check whether we can skip the relocation
180 * by comparing the expected value inside the relocation entry with the target's
181 * final address. If they differ, we have to map the current object and rewrite
182 * the 4 or 8 byte pointer within.
183 *
184 * Serialising an execbuf is quite simple according to the rules of the GEM
185 * ABI. Execution within each context is ordered by the order of submission.
186 * Writes to any GEM object are in order of submission and are exclusive. Reads
187 * from a GEM object are unordered with respect to other reads, but ordered by
188 * writes. A write submitted after a read cannot occur before the read, and
189 * similarly any read submitted after a write cannot occur before the write.
190 * Writes are ordered between engines such that only one write occurs at any
191 * time (completing any reads beforehand) - using semaphores where available
192 * and CPU serialisation otherwise. Other GEM access obey the same rules, any
193 * write (either via mmaps using set-domain, or via pwrite) must flush all GPU
194 * reads before starting, and any read (either using set-domain or pread) must
195 * flush all GPU writes before starting. (Note we only employ a barrier before,
196 * we currently rely on userspace not concurrently starting a new execution
197 * whilst reading or writing to an object. This may be an advantage or not
198 * depending on how much you trust userspace not to shoot themselves in the
199 * foot.) Serialisation may just result in the request being inserted into
200 * a DAG awaiting its turn, but most simple is to wait on the CPU until
201 * all dependencies are resolved.
202 *
203 * After all of that, is just a matter of closing the request and handing it to
204 * the hardware (well, leaving it in a queue to be executed). However, we also
205 * offer the ability for batchbuffers to be run with elevated privileges so
206 * that they access otherwise hidden registers. (Used to adjust L3 cache etc.)
207 * Before any batch is given extra privileges we first must check that it
208 * contains no nefarious instructions, we check that each instruction is from
209 * our whitelist and all registers are also from an allowed list. We first
210 * copy the user's batchbuffer to a shadow (so that the user doesn't have
211 * access to it, either by the CPU or GPU as we scan it) and then parse each
212 * instruction. If everything is ok, we set a flag telling the hardware to run
213 * the batchbuffer in trusted mode, otherwise the ioctl is rejected.
214 */
215
216struct i915_execbuffer {
217 struct drm_i915_private *i915; /** i915 backpointer */
218 struct drm_file *file; /** per-file lookup tables and limits */
219 struct drm_i915_gem_execbuffer2 *args; /** ioctl parameters */
220 struct drm_i915_gem_exec_object2 *exec; /** ioctl execobj[] */
221 struct i915_vma **vma;
222 unsigned int *flags;
223
224 struct intel_engine_cs *engine; /** engine to queue the request to */
225 struct intel_context *context; /* logical state for the request */
226 struct i915_gem_context *gem_context; /** caller's context */
227
228 struct i915_request *request; /** our request to build */
229 struct i915_vma *batch; /** identity of the batch obj/vma */
230
231 /** actual size of execobj[] as we may extend it for the cmdparser */
232 unsigned int buffer_count;
233
234 /** list of vma not yet bound during reservation phase */
235 struct list_head unbound;
236
237 /** list of vma that have execobj.relocation_count */
238 struct list_head relocs;
239
240 /**
241 * Track the most recently used object for relocations, as we
242 * frequently have to perform multiple relocations within the same
243 * obj/page
244 */
245 struct reloc_cache {
246 struct drm_mm_node node; /** temporary GTT binding */
247 unsigned long vaddr; /** Current kmap address */
248 unsigned long page; /** Currently mapped page index */
249 unsigned int gen; /** Cached value of INTEL_GEN */
250 bool use_64bit_reloc : 1;
251 bool has_llc : 1;
252 bool has_fence : 1;
253 bool needs_unfenced : 1;
254
255 struct i915_request *rq;
256 u32 *rq_cmd;
257 unsigned int rq_size;
258 } reloc_cache;
259
260 u64 invalid_flags; /** Set of execobj.flags that are invalid */
261 u32 context_flags; /** Set of execobj.flags to insert from the ctx */
262
263 u32 batch_start_offset; /** Location within object of batch */
264 u32 batch_len; /** Length of batch within object */
265 u32 batch_flags; /** Flags composed for emit_bb_start() */
266
267 /**
268 * Indicate either the size of the hastable used to resolve
269 * relocation handles, or if negative that we are using a direct
270 * index into the execobj[].
271 */
272 int lut_size;
273 struct hlist_head *buckets; /** ht for relocation handles */
274};
275
276#define exec_entry(EB, VMA) (&(EB)->exec[(VMA)->exec_flags - (EB)->flags])
277
278/*
279 * Used to convert any address to canonical form.
280 * Starting from gen8, some commands (e.g. STATE_BASE_ADDRESS,
281 * MI_LOAD_REGISTER_MEM and others, see Broadwell PRM Vol2a) require the
282 * addresses to be in a canonical form:
283 * "GraphicsAddress[63:48] are ignored by the HW and assumed to be in correct
284 * canonical form [63:48] == [47]."
285 */
286#define GEN8_HIGH_ADDRESS_BIT 47
287static inline u64 gen8_canonical_addr(u64 address)
288{
289 return sign_extend64(address, GEN8_HIGH_ADDRESS_BIT);
290}
291
292static inline u64 gen8_noncanonical_addr(u64 address)
293{
294 return address & GENMASK_ULL(GEN8_HIGH_ADDRESS_BIT, 0);
295}
296
297static inline bool eb_use_cmdparser(const struct i915_execbuffer *eb)
298{
299 return intel_engine_requires_cmd_parser(eb->engine) ||
300 (intel_engine_using_cmd_parser(eb->engine) &&
301 eb->args->batch_len);
302}
303
304static int eb_create(struct i915_execbuffer *eb)
305{
306 if (!(eb->args->flags & I915_EXEC_HANDLE_LUT)) {
307 unsigned int size = 1 + ilog2(eb->buffer_count);
308
309 /*
310 * Without a 1:1 association between relocation handles and
311 * the execobject[] index, we instead create a hashtable.
312 * We size it dynamically based on available memory, starting
313 * first with 1:1 assocative hash and scaling back until
314 * the allocation succeeds.
315 *
316 * Later on we use a positive lut_size to indicate we are
317 * using this hashtable, and a negative value to indicate a
318 * direct lookup.
319 */
320 do {
321 gfp_t flags;
322
323 /* While we can still reduce the allocation size, don't
324 * raise a warning and allow the allocation to fail.
325 * On the last pass though, we want to try as hard
326 * as possible to perform the allocation and warn
327 * if it fails.
328 */
329 flags = GFP_KERNEL;
330 if (size > 1)
331 flags |= __GFP_NORETRY | __GFP_NOWARN;
332
333 eb->buckets = kzalloc(sizeof(struct hlist_head) << size,
334 flags);
335 if (eb->buckets)
336 break;
337 } while (--size);
338
339 if (unlikely(!size))
340 return -ENOMEM;
341
342 eb->lut_size = size;
343 } else {
344 eb->lut_size = -eb->buffer_count;
345 }
346
347 return 0;
348}
349
350static bool
351eb_vma_misplaced(const struct drm_i915_gem_exec_object2 *entry,
352 const struct i915_vma *vma,
353 unsigned int flags)
354{
355 if (vma->node.size < entry->pad_to_size)
356 return true;
357
358 if (entry->alignment && !IS_ALIGNED(vma->node.start, entry->alignment))
359 return true;
360
361 if (flags & EXEC_OBJECT_PINNED &&
362 vma->node.start != entry->offset)
363 return true;
364
365 if (flags & __EXEC_OBJECT_NEEDS_BIAS &&
366 vma->node.start < BATCH_OFFSET_BIAS)
367 return true;
368
369 if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) &&
370 (vma->node.start + vma->node.size - 1) >> 32)
371 return true;
372
373 if (flags & __EXEC_OBJECT_NEEDS_MAP &&
374 !i915_vma_is_map_and_fenceable(vma))
375 return true;
376
377 return false;
378}
379
380static inline bool
381eb_pin_vma(struct i915_execbuffer *eb,
382 const struct drm_i915_gem_exec_object2 *entry,
383 struct i915_vma *vma)
384{
385 unsigned int exec_flags = *vma->exec_flags;
386 u64 pin_flags;
387
388 if (vma->node.size)
389 pin_flags = vma->node.start;
390 else
391 pin_flags = entry->offset & PIN_OFFSET_MASK;
392
393 pin_flags |= PIN_USER | PIN_NOEVICT | PIN_OFFSET_FIXED;
394 if (unlikely(exec_flags & EXEC_OBJECT_NEEDS_GTT))
395 pin_flags |= PIN_GLOBAL;
396
397 if (unlikely(i915_vma_pin(vma, 0, 0, pin_flags)))
398 return false;
399
400 if (unlikely(exec_flags & EXEC_OBJECT_NEEDS_FENCE)) {
401 if (unlikely(i915_vma_pin_fence(vma))) {
402 i915_vma_unpin(vma);
403 return false;
404 }
405
406 if (vma->fence)
407 exec_flags |= __EXEC_OBJECT_HAS_FENCE;
408 }
409
410 *vma->exec_flags = exec_flags | __EXEC_OBJECT_HAS_PIN;
411 return !eb_vma_misplaced(entry, vma, exec_flags);
412}
413
414static inline void __eb_unreserve_vma(struct i915_vma *vma, unsigned int flags)
415{
416 GEM_BUG_ON(!(flags & __EXEC_OBJECT_HAS_PIN));
417
418 if (unlikely(flags & __EXEC_OBJECT_HAS_FENCE))
419 __i915_vma_unpin_fence(vma);
420
421 __i915_vma_unpin(vma);
422}
423
424static inline void
425eb_unreserve_vma(struct i915_vma *vma, unsigned int *flags)
426{
427 if (!(*flags & __EXEC_OBJECT_HAS_PIN))
428 return;
429
430 __eb_unreserve_vma(vma, *flags);
431 *flags &= ~__EXEC_OBJECT_RESERVED;
432}
433
434static int
435eb_validate_vma(struct i915_execbuffer *eb,
436 struct drm_i915_gem_exec_object2 *entry,
437 struct i915_vma *vma)
438{
439 if (unlikely(entry->flags & eb->invalid_flags))
440 return -EINVAL;
441
442 if (unlikely(entry->alignment && !is_power_of_2(entry->alignment)))
443 return -EINVAL;
444
445 /*
446 * Offset can be used as input (EXEC_OBJECT_PINNED), reject
447 * any non-page-aligned or non-canonical addresses.
448 */
449 if (unlikely(entry->flags & EXEC_OBJECT_PINNED &&
450 entry->offset != gen8_canonical_addr(entry->offset & I915_GTT_PAGE_MASK)))
451 return -EINVAL;
452
453 /* pad_to_size was once a reserved field, so sanitize it */
454 if (entry->flags & EXEC_OBJECT_PAD_TO_SIZE) {
455 if (unlikely(offset_in_page(entry->pad_to_size)))
456 return -EINVAL;
457 } else {
458 entry->pad_to_size = 0;
459 }
460
461 if (unlikely(vma->exec_flags)) {
462 DRM_DEBUG("Object [handle %d, index %d] appears more than once in object list\n",
463 entry->handle, (int)(entry - eb->exec));
464 return -EINVAL;
465 }
466
467 /*
468 * From drm_mm perspective address space is continuous,
469 * so from this point we're always using non-canonical
470 * form internally.
471 */
472 entry->offset = gen8_noncanonical_addr(entry->offset);
473
474 if (!eb->reloc_cache.has_fence) {
475 entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
476 } else {
477 if ((entry->flags & EXEC_OBJECT_NEEDS_FENCE ||
478 eb->reloc_cache.needs_unfenced) &&
479 i915_gem_object_is_tiled(vma->obj))
480 entry->flags |= EXEC_OBJECT_NEEDS_GTT | __EXEC_OBJECT_NEEDS_MAP;
481 }
482
483 if (!(entry->flags & EXEC_OBJECT_PINNED))
484 entry->flags |= eb->context_flags;
485
486 return 0;
487}
488
489static int
490eb_add_vma(struct i915_execbuffer *eb,
491 unsigned int i, unsigned batch_idx,
492 struct i915_vma *vma)
493{
494 struct drm_i915_gem_exec_object2 *entry = &eb->exec[i];
495 int err;
496
497 GEM_BUG_ON(i915_vma_is_closed(vma));
498
499 if (!(eb->args->flags & __EXEC_VALIDATED)) {
500 err = eb_validate_vma(eb, entry, vma);
501 if (unlikely(err))
502 return err;
503 }
504
505 if (eb->lut_size > 0) {
506 vma->exec_handle = entry->handle;
507 hlist_add_head(&vma->exec_node,
508 &eb->buckets[hash_32(entry->handle,
509 eb->lut_size)]);
510 }
511
512 if (entry->relocation_count)
513 list_add_tail(&vma->reloc_link, &eb->relocs);
514
515 /*
516 * Stash a pointer from the vma to execobj, so we can query its flags,
517 * size, alignment etc as provided by the user. Also we stash a pointer
518 * to the vma inside the execobj so that we can use a direct lookup
519 * to find the right target VMA when doing relocations.
520 */
521 eb->vma[i] = vma;
522 eb->flags[i] = entry->flags;
523 vma->exec_flags = &eb->flags[i];
524
525 /*
526 * SNA is doing fancy tricks with compressing batch buffers, which leads
527 * to negative relocation deltas. Usually that works out ok since the
528 * relocate address is still positive, except when the batch is placed
529 * very low in the GTT. Ensure this doesn't happen.
530 *
531 * Note that actual hangs have only been observed on gen7, but for
532 * paranoia do it everywhere.
533 */
534 if (i == batch_idx) {
535 if (entry->relocation_count &&
536 !(eb->flags[i] & EXEC_OBJECT_PINNED))
537 eb->flags[i] |= __EXEC_OBJECT_NEEDS_BIAS;
538 if (eb->reloc_cache.has_fence)
539 eb->flags[i] |= EXEC_OBJECT_NEEDS_FENCE;
540
541 eb->batch = vma;
542 }
543
544 err = 0;
545 if (eb_pin_vma(eb, entry, vma)) {
546 if (entry->offset != vma->node.start) {
547 entry->offset = vma->node.start | UPDATE;
548 eb->args->flags |= __EXEC_HAS_RELOC;
549 }
550 } else {
551 eb_unreserve_vma(vma, vma->exec_flags);
552
553 list_add_tail(&vma->exec_link, &eb->unbound);
554 if (drm_mm_node_allocated(&vma->node))
555 err = i915_vma_unbind(vma);
556 if (unlikely(err))
557 vma->exec_flags = NULL;
558 }
559 return err;
560}
561
562static inline int use_cpu_reloc(const struct reloc_cache *cache,
563 const struct drm_i915_gem_object *obj)
564{
565 if (!i915_gem_object_has_struct_page(obj))
566 return false;
567
568 if (DBG_FORCE_RELOC == FORCE_CPU_RELOC)
569 return true;
570
571 if (DBG_FORCE_RELOC == FORCE_GTT_RELOC)
572 return false;
573
574 return (cache->has_llc ||
575 obj->cache_dirty ||
576 obj->cache_level != I915_CACHE_NONE);
577}
578
579static int eb_reserve_vma(const struct i915_execbuffer *eb,
580 struct i915_vma *vma)
581{
582 struct drm_i915_gem_exec_object2 *entry = exec_entry(eb, vma);
583 unsigned int exec_flags = *vma->exec_flags;
584 u64 pin_flags;
585 int err;
586
587 pin_flags = PIN_USER | PIN_NONBLOCK;
588 if (exec_flags & EXEC_OBJECT_NEEDS_GTT)
589 pin_flags |= PIN_GLOBAL;
590
591 /*
592 * Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset,
593 * limit address to the first 4GBs for unflagged objects.
594 */
595 if (!(exec_flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS))
596 pin_flags |= PIN_ZONE_4G;
597
598 if (exec_flags & __EXEC_OBJECT_NEEDS_MAP)
599 pin_flags |= PIN_MAPPABLE;
600
601 if (exec_flags & EXEC_OBJECT_PINNED) {
602 pin_flags |= entry->offset | PIN_OFFSET_FIXED;
603 pin_flags &= ~PIN_NONBLOCK; /* force overlapping checks */
604 } else if (exec_flags & __EXEC_OBJECT_NEEDS_BIAS) {
605 pin_flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
606 }
607
608 err = i915_vma_pin(vma,
609 entry->pad_to_size, entry->alignment,
610 pin_flags);
611 if (err)
612 return err;
613
614 if (entry->offset != vma->node.start) {
615 entry->offset = vma->node.start | UPDATE;
616 eb->args->flags |= __EXEC_HAS_RELOC;
617 }
618
619 if (unlikely(exec_flags & EXEC_OBJECT_NEEDS_FENCE)) {
620 err = i915_vma_pin_fence(vma);
621 if (unlikely(err)) {
622 i915_vma_unpin(vma);
623 return err;
624 }
625
626 if (vma->fence)
627 exec_flags |= __EXEC_OBJECT_HAS_FENCE;
628 }
629
630 *vma->exec_flags = exec_flags | __EXEC_OBJECT_HAS_PIN;
631 GEM_BUG_ON(eb_vma_misplaced(entry, vma, exec_flags));
632
633 return 0;
634}
635
636static int eb_reserve(struct i915_execbuffer *eb)
637{
638 const unsigned int count = eb->buffer_count;
639 struct list_head last;
640 struct i915_vma *vma;
641 unsigned int i, pass;
642 int err;
643
644 /*
645 * Attempt to pin all of the buffers into the GTT.
646 * This is done in 3 phases:
647 *
648 * 1a. Unbind all objects that do not match the GTT constraints for
649 * the execbuffer (fenceable, mappable, alignment etc).
650 * 1b. Increment pin count for already bound objects.
651 * 2. Bind new objects.
652 * 3. Decrement pin count.
653 *
654 * This avoid unnecessary unbinding of later objects in order to make
655 * room for the earlier objects *unless* we need to defragment.
656 */
657
658 pass = 0;
659 err = 0;
660 do {
661 list_for_each_entry(vma, &eb->unbound, exec_link) {
662 err = eb_reserve_vma(eb, vma);
663 if (err)
664 break;
665 }
666 if (err != -ENOSPC)
667 return err;
668
669 /* Resort *all* the objects into priority order */
670 INIT_LIST_HEAD(&eb->unbound);
671 INIT_LIST_HEAD(&last);
672 for (i = 0; i < count; i++) {
673 unsigned int flags = eb->flags[i];
674 struct i915_vma *vma = eb->vma[i];
675
676 if (flags & EXEC_OBJECT_PINNED &&
677 flags & __EXEC_OBJECT_HAS_PIN)
678 continue;
679
680 eb_unreserve_vma(vma, &eb->flags[i]);
681
682 if (flags & EXEC_OBJECT_PINNED)
683 /* Pinned must have their slot */
684 list_add(&vma->exec_link, &eb->unbound);
685 else if (flags & __EXEC_OBJECT_NEEDS_MAP)
686 /* Map require the lowest 256MiB (aperture) */
687 list_add_tail(&vma->exec_link, &eb->unbound);
688 else if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS))
689 /* Prioritise 4GiB region for restricted bo */
690 list_add(&vma->exec_link, &last);
691 else
692 list_add_tail(&vma->exec_link, &last);
693 }
694 list_splice_tail(&last, &eb->unbound);
695
696 switch (pass++) {
697 case 0:
698 break;
699
700 case 1:
701 /* Too fragmented, unbind everything and retry */
702 err = i915_gem_evict_vm(eb->context->vm);
703 if (err)
704 return err;
705 break;
706
707 default:
708 return -ENOSPC;
709 }
710 } while (1);
711}
712
713static unsigned int eb_batch_index(const struct i915_execbuffer *eb)
714{
715 if (eb->args->flags & I915_EXEC_BATCH_FIRST)
716 return 0;
717 else
718 return eb->buffer_count - 1;
719}
720
721static int eb_select_context(struct i915_execbuffer *eb)
722{
723 struct i915_gem_context *ctx;
724
725 ctx = i915_gem_context_lookup(eb->file->driver_priv, eb->args->rsvd1);
726 if (unlikely(!ctx))
727 return -ENOENT;
728
729 eb->gem_context = ctx;
730 if (ctx->vm)
731 eb->invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
732
733 eb->context_flags = 0;
734 if (test_bit(UCONTEXT_NO_ZEROMAP, &ctx->user_flags))
735 eb->context_flags |= __EXEC_OBJECT_NEEDS_BIAS;
736
737 return 0;
738}
739
740static int eb_lookup_vmas(struct i915_execbuffer *eb)
741{
742 struct radix_tree_root *handles_vma = &eb->gem_context->handles_vma;
743 struct drm_i915_gem_object *obj;
744 unsigned int i, batch;
745 int err;
746
747 if (unlikely(i915_gem_context_is_banned(eb->gem_context)))
748 return -EIO;
749
750 INIT_LIST_HEAD(&eb->relocs);
751 INIT_LIST_HEAD(&eb->unbound);
752
753 batch = eb_batch_index(eb);
754
755 mutex_lock(&eb->gem_context->mutex);
756 if (unlikely(i915_gem_context_is_closed(eb->gem_context))) {
757 err = -ENOENT;
758 goto err_ctx;
759 }
760
761 for (i = 0; i < eb->buffer_count; i++) {
762 u32 handle = eb->exec[i].handle;
763 struct i915_lut_handle *lut;
764 struct i915_vma *vma;
765
766 vma = radix_tree_lookup(handles_vma, handle);
767 if (likely(vma))
768 goto add_vma;
769
770 obj = i915_gem_object_lookup(eb->file, handle);
771 if (unlikely(!obj)) {
772 err = -ENOENT;
773 goto err_vma;
774 }
775
776 vma = i915_vma_instance(obj, eb->context->vm, NULL);
777 if (IS_ERR(vma)) {
778 err = PTR_ERR(vma);
779 goto err_obj;
780 }
781
782 lut = i915_lut_handle_alloc();
783 if (unlikely(!lut)) {
784 err = -ENOMEM;
785 goto err_obj;
786 }
787
788 err = radix_tree_insert(handles_vma, handle, vma);
789 if (unlikely(err)) {
790 i915_lut_handle_free(lut);
791 goto err_obj;
792 }
793
794 /* transfer ref to lut */
795 if (!atomic_fetch_inc(&vma->open_count))
796 i915_vma_reopen(vma);
797 lut->handle = handle;
798 lut->ctx = eb->gem_context;
799
800 i915_gem_object_lock(obj);
801 list_add(&lut->obj_link, &obj->lut_list);
802 i915_gem_object_unlock(obj);
803
804add_vma:
805 err = eb_add_vma(eb, i, batch, vma);
806 if (unlikely(err))
807 goto err_vma;
808
809 GEM_BUG_ON(vma != eb->vma[i]);
810 GEM_BUG_ON(vma->exec_flags != &eb->flags[i]);
811 GEM_BUG_ON(drm_mm_node_allocated(&vma->node) &&
812 eb_vma_misplaced(&eb->exec[i], vma, eb->flags[i]));
813 }
814
815 mutex_unlock(&eb->gem_context->mutex);
816
817 eb->args->flags |= __EXEC_VALIDATED;
818 return eb_reserve(eb);
819
820err_obj:
821 i915_gem_object_put(obj);
822err_vma:
823 eb->vma[i] = NULL;
824err_ctx:
825 mutex_unlock(&eb->gem_context->mutex);
826 return err;
827}
828
829static struct i915_vma *
830eb_get_vma(const struct i915_execbuffer *eb, unsigned long handle)
831{
832 if (eb->lut_size < 0) {
833 if (handle >= -eb->lut_size)
834 return NULL;
835 return eb->vma[handle];
836 } else {
837 struct hlist_head *head;
838 struct i915_vma *vma;
839
840 head = &eb->buckets[hash_32(handle, eb->lut_size)];
841 hlist_for_each_entry(vma, head, exec_node) {
842 if (vma->exec_handle == handle)
843 return vma;
844 }
845 return NULL;
846 }
847}
848
849static void eb_release_vmas(const struct i915_execbuffer *eb)
850{
851 const unsigned int count = eb->buffer_count;
852 unsigned int i;
853
854 for (i = 0; i < count; i++) {
855 struct i915_vma *vma = eb->vma[i];
856 unsigned int flags = eb->flags[i];
857
858 if (!vma)
859 break;
860
861 GEM_BUG_ON(vma->exec_flags != &eb->flags[i]);
862 vma->exec_flags = NULL;
863 eb->vma[i] = NULL;
864
865 if (flags & __EXEC_OBJECT_HAS_PIN)
866 __eb_unreserve_vma(vma, flags);
867
868 if (flags & __EXEC_OBJECT_HAS_REF)
869 i915_vma_put(vma);
870 }
871}
872
873static void eb_reset_vmas(const struct i915_execbuffer *eb)
874{
875 eb_release_vmas(eb);
876 if (eb->lut_size > 0)
877 memset(eb->buckets, 0,
878 sizeof(struct hlist_head) << eb->lut_size);
879}
880
881static void eb_destroy(const struct i915_execbuffer *eb)
882{
883 GEM_BUG_ON(eb->reloc_cache.rq);
884
885 if (eb->lut_size > 0)
886 kfree(eb->buckets);
887}
888
889static inline u64
890relocation_target(const struct drm_i915_gem_relocation_entry *reloc,
891 const struct i915_vma *target)
892{
893 return gen8_canonical_addr((int)reloc->delta + target->node.start);
894}
895
896static void reloc_cache_init(struct reloc_cache *cache,
897 struct drm_i915_private *i915)
898{
899 cache->page = -1;
900 cache->vaddr = 0;
901 /* Must be a variable in the struct to allow GCC to unroll. */
902 cache->gen = INTEL_GEN(i915);
903 cache->has_llc = HAS_LLC(i915);
904 cache->use_64bit_reloc = HAS_64BIT_RELOC(i915);
905 cache->has_fence = cache->gen < 4;
906 cache->needs_unfenced = INTEL_INFO(i915)->unfenced_needs_alignment;
907 cache->node.allocated = false;
908 cache->rq = NULL;
909 cache->rq_size = 0;
910}
911
912static inline void *unmask_page(unsigned long p)
913{
914 return (void *)(uintptr_t)(p & PAGE_MASK);
915}
916
917static inline unsigned int unmask_flags(unsigned long p)
918{
919 return p & ~PAGE_MASK;
920}
921
922#define KMAP 0x4 /* after CLFLUSH_FLAGS */
923
924static inline struct i915_ggtt *cache_to_ggtt(struct reloc_cache *cache)
925{
926 struct drm_i915_private *i915 =
927 container_of(cache, struct i915_execbuffer, reloc_cache)->i915;
928 return &i915->ggtt;
929}
930
931static void reloc_gpu_flush(struct reloc_cache *cache)
932{
933 GEM_BUG_ON(cache->rq_size >= cache->rq->batch->obj->base.size / sizeof(u32));
934 cache->rq_cmd[cache->rq_size] = MI_BATCH_BUFFER_END;
935
936 __i915_gem_object_flush_map(cache->rq->batch->obj, 0, cache->rq_size);
937 i915_gem_object_unpin_map(cache->rq->batch->obj);
938
939 intel_gt_chipset_flush(cache->rq->engine->gt);
940
941 i915_request_add(cache->rq);
942 cache->rq = NULL;
943}
944
945static void reloc_cache_reset(struct reloc_cache *cache)
946{
947 void *vaddr;
948
949 if (cache->rq)
950 reloc_gpu_flush(cache);
951
952 if (!cache->vaddr)
953 return;
954
955 vaddr = unmask_page(cache->vaddr);
956 if (cache->vaddr & KMAP) {
957 if (cache->vaddr & CLFLUSH_AFTER)
958 mb();
959
960 kunmap_atomic(vaddr);
961 i915_gem_object_finish_access((struct drm_i915_gem_object *)cache->node.mm);
962 } else {
963 struct i915_ggtt *ggtt = cache_to_ggtt(cache);
964
965 intel_gt_flush_ggtt_writes(ggtt->vm.gt);
966 io_mapping_unmap_atomic((void __iomem *)vaddr);
967
968 if (cache->node.allocated) {
969 ggtt->vm.clear_range(&ggtt->vm,
970 cache->node.start,
971 cache->node.size);
972 drm_mm_remove_node(&cache->node);
973 } else {
974 i915_vma_unpin((struct i915_vma *)cache->node.mm);
975 }
976 }
977
978 cache->vaddr = 0;
979 cache->page = -1;
980}
981
982static void *reloc_kmap(struct drm_i915_gem_object *obj,
983 struct reloc_cache *cache,
984 unsigned long page)
985{
986 void *vaddr;
987
988 if (cache->vaddr) {
989 kunmap_atomic(unmask_page(cache->vaddr));
990 } else {
991 unsigned int flushes;
992 int err;
993
994 err = i915_gem_object_prepare_write(obj, &flushes);
995 if (err)
996 return ERR_PTR(err);
997
998 BUILD_BUG_ON(KMAP & CLFLUSH_FLAGS);
999 BUILD_BUG_ON((KMAP | CLFLUSH_FLAGS) & PAGE_MASK);
1000
1001 cache->vaddr = flushes | KMAP;
1002 cache->node.mm = (void *)obj;
1003 if (flushes)
1004 mb();
1005 }
1006
1007 vaddr = kmap_atomic(i915_gem_object_get_dirty_page(obj, page));
1008 cache->vaddr = unmask_flags(cache->vaddr) | (unsigned long)vaddr;
1009 cache->page = page;
1010
1011 return vaddr;
1012}
1013
1014static void *reloc_iomap(struct drm_i915_gem_object *obj,
1015 struct reloc_cache *cache,
1016 unsigned long page)
1017{
1018 struct i915_ggtt *ggtt = cache_to_ggtt(cache);
1019 unsigned long offset;
1020 void *vaddr;
1021
1022 if (cache->vaddr) {
1023 intel_gt_flush_ggtt_writes(ggtt->vm.gt);
1024 io_mapping_unmap_atomic((void __force __iomem *) unmask_page(cache->vaddr));
1025 } else {
1026 struct i915_vma *vma;
1027 int err;
1028
1029 if (i915_gem_object_is_tiled(obj))
1030 return ERR_PTR(-EINVAL);
1031
1032 if (use_cpu_reloc(cache, obj))
1033 return NULL;
1034
1035 i915_gem_object_lock(obj);
1036 err = i915_gem_object_set_to_gtt_domain(obj, true);
1037 i915_gem_object_unlock(obj);
1038 if (err)
1039 return ERR_PTR(err);
1040
1041 vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
1042 PIN_MAPPABLE |
1043 PIN_NONBLOCK /* NOWARN */ |
1044 PIN_NOEVICT);
1045 if (IS_ERR(vma)) {
1046 memset(&cache->node, 0, sizeof(cache->node));
1047 err = drm_mm_insert_node_in_range
1048 (&ggtt->vm.mm, &cache->node,
1049 PAGE_SIZE, 0, I915_COLOR_UNEVICTABLE,
1050 0, ggtt->mappable_end,
1051 DRM_MM_INSERT_LOW);
1052 if (err) /* no inactive aperture space, use cpu reloc */
1053 return NULL;
1054 } else {
1055 cache->node.start = vma->node.start;
1056 cache->node.mm = (void *)vma;
1057 }
1058 }
1059
1060 offset = cache->node.start;
1061 if (cache->node.allocated) {
1062 ggtt->vm.insert_page(&ggtt->vm,
1063 i915_gem_object_get_dma_address(obj, page),
1064 offset, I915_CACHE_NONE, 0);
1065 } else {
1066 offset += page << PAGE_SHIFT;
1067 }
1068
1069 vaddr = (void __force *)io_mapping_map_atomic_wc(&ggtt->iomap,
1070 offset);
1071 cache->page = page;
1072 cache->vaddr = (unsigned long)vaddr;
1073
1074 return vaddr;
1075}
1076
1077static void *reloc_vaddr(struct drm_i915_gem_object *obj,
1078 struct reloc_cache *cache,
1079 unsigned long page)
1080{
1081 void *vaddr;
1082
1083 if (cache->page == page) {
1084 vaddr = unmask_page(cache->vaddr);
1085 } else {
1086 vaddr = NULL;
1087 if ((cache->vaddr & KMAP) == 0)
1088 vaddr = reloc_iomap(obj, cache, page);
1089 if (!vaddr)
1090 vaddr = reloc_kmap(obj, cache, page);
1091 }
1092
1093 return vaddr;
1094}
1095
1096static void clflush_write32(u32 *addr, u32 value, unsigned int flushes)
1097{
1098 if (unlikely(flushes & (CLFLUSH_BEFORE | CLFLUSH_AFTER))) {
1099 if (flushes & CLFLUSH_BEFORE) {
1100 clflushopt(addr);
1101 mb();
1102 }
1103
1104 *addr = value;
1105
1106 /*
1107 * Writes to the same cacheline are serialised by the CPU
1108 * (including clflush). On the write path, we only require
1109 * that it hits memory in an orderly fashion and place
1110 * mb barriers at the start and end of the relocation phase
1111 * to ensure ordering of clflush wrt to the system.
1112 */
1113 if (flushes & CLFLUSH_AFTER)
1114 clflushopt(addr);
1115 } else
1116 *addr = value;
1117}
1118
1119static int reloc_move_to_gpu(struct i915_request *rq, struct i915_vma *vma)
1120{
1121 struct drm_i915_gem_object *obj = vma->obj;
1122 int err;
1123
1124 i915_vma_lock(vma);
1125
1126 if (obj->cache_dirty & ~obj->cache_coherent)
1127 i915_gem_clflush_object(obj, 0);
1128 obj->write_domain = 0;
1129
1130 err = i915_request_await_object(rq, vma->obj, true);
1131 if (err == 0)
1132 err = i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE);
1133
1134 i915_vma_unlock(vma);
1135
1136 return err;
1137}
1138
1139static int __reloc_gpu_alloc(struct i915_execbuffer *eb,
1140 struct i915_vma *vma,
1141 unsigned int len)
1142{
1143 struct reloc_cache *cache = &eb->reloc_cache;
1144 struct intel_engine_pool_node *pool;
1145 struct i915_request *rq;
1146 struct i915_vma *batch;
1147 u32 *cmd;
1148 int err;
1149
1150 pool = intel_engine_pool_get(&eb->engine->pool, PAGE_SIZE);
1151 if (IS_ERR(pool))
1152 return PTR_ERR(pool);
1153
1154 cmd = i915_gem_object_pin_map(pool->obj,
1155 cache->has_llc ?
1156 I915_MAP_FORCE_WB :
1157 I915_MAP_FORCE_WC);
1158 if (IS_ERR(cmd)) {
1159 err = PTR_ERR(cmd);
1160 goto out_pool;
1161 }
1162
1163 batch = i915_vma_instance(pool->obj, vma->vm, NULL);
1164 if (IS_ERR(batch)) {
1165 err = PTR_ERR(batch);
1166 goto err_unmap;
1167 }
1168
1169 err = i915_vma_pin(batch, 0, 0, PIN_USER | PIN_NONBLOCK);
1170 if (err)
1171 goto err_unmap;
1172
1173 rq = i915_request_create(eb->context);
1174 if (IS_ERR(rq)) {
1175 err = PTR_ERR(rq);
1176 goto err_unpin;
1177 }
1178
1179 err = intel_engine_pool_mark_active(pool, rq);
1180 if (err)
1181 goto err_request;
1182
1183 err = reloc_move_to_gpu(rq, vma);
1184 if (err)
1185 goto err_request;
1186
1187 err = eb->engine->emit_bb_start(rq,
1188 batch->node.start, PAGE_SIZE,
1189 cache->gen > 5 ? 0 : I915_DISPATCH_SECURE);
1190 if (err)
1191 goto skip_request;
1192
1193 i915_vma_lock(batch);
1194 err = i915_request_await_object(rq, batch->obj, false);
1195 if (err == 0)
1196 err = i915_vma_move_to_active(batch, rq, 0);
1197 i915_vma_unlock(batch);
1198 if (err)
1199 goto skip_request;
1200
1201 rq->batch = batch;
1202 i915_vma_unpin(batch);
1203
1204 cache->rq = rq;
1205 cache->rq_cmd = cmd;
1206 cache->rq_size = 0;
1207
1208 /* Return with batch mapping (cmd) still pinned */
1209 goto out_pool;
1210
1211skip_request:
1212 i915_request_skip(rq, err);
1213err_request:
1214 i915_request_add(rq);
1215err_unpin:
1216 i915_vma_unpin(batch);
1217err_unmap:
1218 i915_gem_object_unpin_map(pool->obj);
1219out_pool:
1220 intel_engine_pool_put(pool);
1221 return err;
1222}
1223
1224static u32 *reloc_gpu(struct i915_execbuffer *eb,
1225 struct i915_vma *vma,
1226 unsigned int len)
1227{
1228 struct reloc_cache *cache = &eb->reloc_cache;
1229 u32 *cmd;
1230
1231 if (cache->rq_size > PAGE_SIZE/sizeof(u32) - (len + 1))
1232 reloc_gpu_flush(cache);
1233
1234 if (unlikely(!cache->rq)) {
1235 int err;
1236
1237 /* If we need to copy for the cmdparser, we will stall anyway */
1238 if (eb_use_cmdparser(eb))
1239 return ERR_PTR(-EWOULDBLOCK);
1240
1241 if (!intel_engine_can_store_dword(eb->engine))
1242 return ERR_PTR(-ENODEV);
1243
1244 err = __reloc_gpu_alloc(eb, vma, len);
1245 if (unlikely(err))
1246 return ERR_PTR(err);
1247 }
1248
1249 cmd = cache->rq_cmd + cache->rq_size;
1250 cache->rq_size += len;
1251
1252 return cmd;
1253}
1254
1255static u64
1256relocate_entry(struct i915_vma *vma,
1257 const struct drm_i915_gem_relocation_entry *reloc,
1258 struct i915_execbuffer *eb,
1259 const struct i915_vma *target)
1260{
1261 u64 offset = reloc->offset;
1262 u64 target_offset = relocation_target(reloc, target);
1263 bool wide = eb->reloc_cache.use_64bit_reloc;
1264 void *vaddr;
1265
1266 if (!eb->reloc_cache.vaddr &&
1267 (DBG_FORCE_RELOC == FORCE_GPU_RELOC ||
1268 !dma_resv_test_signaled_rcu(vma->resv, true))) {
1269 const unsigned int gen = eb->reloc_cache.gen;
1270 unsigned int len;
1271 u32 *batch;
1272 u64 addr;
1273
1274 if (wide)
1275 len = offset & 7 ? 8 : 5;
1276 else if (gen >= 4)
1277 len = 4;
1278 else
1279 len = 3;
1280
1281 batch = reloc_gpu(eb, vma, len);
1282 if (IS_ERR(batch))
1283 goto repeat;
1284
1285 addr = gen8_canonical_addr(vma->node.start + offset);
1286 if (wide) {
1287 if (offset & 7) {
1288 *batch++ = MI_STORE_DWORD_IMM_GEN4;
1289 *batch++ = lower_32_bits(addr);
1290 *batch++ = upper_32_bits(addr);
1291 *batch++ = lower_32_bits(target_offset);
1292
1293 addr = gen8_canonical_addr(addr + 4);
1294
1295 *batch++ = MI_STORE_DWORD_IMM_GEN4;
1296 *batch++ = lower_32_bits(addr);
1297 *batch++ = upper_32_bits(addr);
1298 *batch++ = upper_32_bits(target_offset);
1299 } else {
1300 *batch++ = (MI_STORE_DWORD_IMM_GEN4 | (1 << 21)) + 1;
1301 *batch++ = lower_32_bits(addr);
1302 *batch++ = upper_32_bits(addr);
1303 *batch++ = lower_32_bits(target_offset);
1304 *batch++ = upper_32_bits(target_offset);
1305 }
1306 } else if (gen >= 6) {
1307 *batch++ = MI_STORE_DWORD_IMM_GEN4;
1308 *batch++ = 0;
1309 *batch++ = addr;
1310 *batch++ = target_offset;
1311 } else if (gen >= 4) {
1312 *batch++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
1313 *batch++ = 0;
1314 *batch++ = addr;
1315 *batch++ = target_offset;
1316 } else {
1317 *batch++ = MI_STORE_DWORD_IMM | MI_MEM_VIRTUAL;
1318 *batch++ = addr;
1319 *batch++ = target_offset;
1320 }
1321
1322 goto out;
1323 }
1324
1325repeat:
1326 vaddr = reloc_vaddr(vma->obj, &eb->reloc_cache, offset >> PAGE_SHIFT);
1327 if (IS_ERR(vaddr))
1328 return PTR_ERR(vaddr);
1329
1330 clflush_write32(vaddr + offset_in_page(offset),
1331 lower_32_bits(target_offset),
1332 eb->reloc_cache.vaddr);
1333
1334 if (wide) {
1335 offset += sizeof(u32);
1336 target_offset >>= 32;
1337 wide = false;
1338 goto repeat;
1339 }
1340
1341out:
1342 return target->node.start | UPDATE;
1343}
1344
1345static u64
1346eb_relocate_entry(struct i915_execbuffer *eb,
1347 struct i915_vma *vma,
1348 const struct drm_i915_gem_relocation_entry *reloc)
1349{
1350 struct i915_vma *target;
1351 int err;
1352
1353 /* we've already hold a reference to all valid objects */
1354 target = eb_get_vma(eb, reloc->target_handle);
1355 if (unlikely(!target))
1356 return -ENOENT;
1357
1358 /* Validate that the target is in a valid r/w GPU domain */
1359 if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
1360 DRM_DEBUG("reloc with multiple write domains: "
1361 "target %d offset %d "
1362 "read %08x write %08x",
1363 reloc->target_handle,
1364 (int) reloc->offset,
1365 reloc->read_domains,
1366 reloc->write_domain);
1367 return -EINVAL;
1368 }
1369 if (unlikely((reloc->write_domain | reloc->read_domains)
1370 & ~I915_GEM_GPU_DOMAINS)) {
1371 DRM_DEBUG("reloc with read/write non-GPU domains: "
1372 "target %d offset %d "
1373 "read %08x write %08x",
1374 reloc->target_handle,
1375 (int) reloc->offset,
1376 reloc->read_domains,
1377 reloc->write_domain);
1378 return -EINVAL;
1379 }
1380
1381 if (reloc->write_domain) {
1382 *target->exec_flags |= EXEC_OBJECT_WRITE;
1383
1384 /*
1385 * Sandybridge PPGTT errata: We need a global gtt mapping
1386 * for MI and pipe_control writes because the gpu doesn't
1387 * properly redirect them through the ppgtt for non_secure
1388 * batchbuffers.
1389 */
1390 if (reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
1391 IS_GEN(eb->i915, 6)) {
1392 err = i915_vma_bind(target, target->obj->cache_level,
1393 PIN_GLOBAL);
1394 if (WARN_ONCE(err,
1395 "Unexpected failure to bind target VMA!"))
1396 return err;
1397 }
1398 }
1399
1400 /*
1401 * If the relocation already has the right value in it, no
1402 * more work needs to be done.
1403 */
1404 if (!DBG_FORCE_RELOC &&
1405 gen8_canonical_addr(target->node.start) == reloc->presumed_offset)
1406 return 0;
1407
1408 /* Check that the relocation address is valid... */
1409 if (unlikely(reloc->offset >
1410 vma->size - (eb->reloc_cache.use_64bit_reloc ? 8 : 4))) {
1411 DRM_DEBUG("Relocation beyond object bounds: "
1412 "target %d offset %d size %d.\n",
1413 reloc->target_handle,
1414 (int)reloc->offset,
1415 (int)vma->size);
1416 return -EINVAL;
1417 }
1418 if (unlikely(reloc->offset & 3)) {
1419 DRM_DEBUG("Relocation not 4-byte aligned: "
1420 "target %d offset %d.\n",
1421 reloc->target_handle,
1422 (int)reloc->offset);
1423 return -EINVAL;
1424 }
1425
1426 /*
1427 * If we write into the object, we need to force the synchronisation
1428 * barrier, either with an asynchronous clflush or if we executed the
1429 * patching using the GPU (though that should be serialised by the
1430 * timeline). To be completely sure, and since we are required to
1431 * do relocations we are already stalling, disable the user's opt
1432 * out of our synchronisation.
1433 */
1434 *vma->exec_flags &= ~EXEC_OBJECT_ASYNC;
1435
1436 /* and update the user's relocation entry */
1437 return relocate_entry(vma, reloc, eb, target);
1438}
1439
1440static int eb_relocate_vma(struct i915_execbuffer *eb, struct i915_vma *vma)
1441{
1442#define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
1443 struct drm_i915_gem_relocation_entry stack[N_RELOC(512)];
1444 struct drm_i915_gem_relocation_entry __user *urelocs;
1445 const struct drm_i915_gem_exec_object2 *entry = exec_entry(eb, vma);
1446 unsigned int remain;
1447
1448 urelocs = u64_to_user_ptr(entry->relocs_ptr);
1449 remain = entry->relocation_count;
1450 if (unlikely(remain > N_RELOC(ULONG_MAX)))
1451 return -EINVAL;
1452
1453 /*
1454 * We must check that the entire relocation array is safe
1455 * to read. However, if the array is not writable the user loses
1456 * the updated relocation values.
1457 */
1458 if (unlikely(!access_ok(urelocs, remain*sizeof(*urelocs))))
1459 return -EFAULT;
1460
1461 do {
1462 struct drm_i915_gem_relocation_entry *r = stack;
1463 unsigned int count =
1464 min_t(unsigned int, remain, ARRAY_SIZE(stack));
1465 unsigned int copied;
1466
1467 /*
1468 * This is the fast path and we cannot handle a pagefault
1469 * whilst holding the struct mutex lest the user pass in the
1470 * relocations contained within a mmaped bo. For in such a case
1471 * we, the page fault handler would call i915_gem_fault() and
1472 * we would try to acquire the struct mutex again. Obviously
1473 * this is bad and so lockdep complains vehemently.
1474 */
1475 pagefault_disable();
1476 copied = __copy_from_user_inatomic(r, urelocs, count * sizeof(r[0]));
1477 pagefault_enable();
1478 if (unlikely(copied)) {
1479 remain = -EFAULT;
1480 goto out;
1481 }
1482
1483 remain -= count;
1484 do {
1485 u64 offset = eb_relocate_entry(eb, vma, r);
1486
1487 if (likely(offset == 0)) {
1488 } else if ((s64)offset < 0) {
1489 remain = (int)offset;
1490 goto out;
1491 } else {
1492 /*
1493 * Note that reporting an error now
1494 * leaves everything in an inconsistent
1495 * state as we have *already* changed
1496 * the relocation value inside the
1497 * object. As we have not changed the
1498 * reloc.presumed_offset or will not
1499 * change the execobject.offset, on the
1500 * call we may not rewrite the value
1501 * inside the object, leaving it
1502 * dangling and causing a GPU hang. Unless
1503 * userspace dynamically rebuilds the
1504 * relocations on each execbuf rather than
1505 * presume a static tree.
1506 *
1507 * We did previously check if the relocations
1508 * were writable (access_ok), an error now
1509 * would be a strange race with mprotect,
1510 * having already demonstrated that we
1511 * can read from this userspace address.
1512 */
1513 offset = gen8_canonical_addr(offset & ~UPDATE);
1514 if (unlikely(__put_user(offset, &urelocs[r-stack].presumed_offset))) {
1515 remain = -EFAULT;
1516 goto out;
1517 }
1518 }
1519 } while (r++, --count);
1520 urelocs += ARRAY_SIZE(stack);
1521 } while (remain);
1522out:
1523 reloc_cache_reset(&eb->reloc_cache);
1524 return remain;
1525}
1526
1527static int
1528eb_relocate_vma_slow(struct i915_execbuffer *eb, struct i915_vma *vma)
1529{
1530 const struct drm_i915_gem_exec_object2 *entry = exec_entry(eb, vma);
1531 struct drm_i915_gem_relocation_entry *relocs =
1532 u64_to_ptr(typeof(*relocs), entry->relocs_ptr);
1533 unsigned int i;
1534 int err;
1535
1536 for (i = 0; i < entry->relocation_count; i++) {
1537 u64 offset = eb_relocate_entry(eb, vma, &relocs[i]);
1538
1539 if ((s64)offset < 0) {
1540 err = (int)offset;
1541 goto err;
1542 }
1543 }
1544 err = 0;
1545err:
1546 reloc_cache_reset(&eb->reloc_cache);
1547 return err;
1548}
1549
1550static int check_relocations(const struct drm_i915_gem_exec_object2 *entry)
1551{
1552 const char __user *addr, *end;
1553 unsigned long size;
1554 char __maybe_unused c;
1555
1556 size = entry->relocation_count;
1557 if (size == 0)
1558 return 0;
1559
1560 if (size > N_RELOC(ULONG_MAX))
1561 return -EINVAL;
1562
1563 addr = u64_to_user_ptr(entry->relocs_ptr);
1564 size *= sizeof(struct drm_i915_gem_relocation_entry);
1565 if (!access_ok(addr, size))
1566 return -EFAULT;
1567
1568 end = addr + size;
1569 for (; addr < end; addr += PAGE_SIZE) {
1570 int err = __get_user(c, addr);
1571 if (err)
1572 return err;
1573 }
1574 return __get_user(c, end - 1);
1575}
1576
1577static int eb_copy_relocations(const struct i915_execbuffer *eb)
1578{
1579 struct drm_i915_gem_relocation_entry *relocs;
1580 const unsigned int count = eb->buffer_count;
1581 unsigned int i;
1582 int err;
1583
1584 for (i = 0; i < count; i++) {
1585 const unsigned int nreloc = eb->exec[i].relocation_count;
1586 struct drm_i915_gem_relocation_entry __user *urelocs;
1587 unsigned long size;
1588 unsigned long copied;
1589
1590 if (nreloc == 0)
1591 continue;
1592
1593 err = check_relocations(&eb->exec[i]);
1594 if (err)
1595 goto err;
1596
1597 urelocs = u64_to_user_ptr(eb->exec[i].relocs_ptr);
1598 size = nreloc * sizeof(*relocs);
1599
1600 relocs = kvmalloc_array(size, 1, GFP_KERNEL);
1601 if (!relocs) {
1602 err = -ENOMEM;
1603 goto err;
1604 }
1605
1606 /* copy_from_user is limited to < 4GiB */
1607 copied = 0;
1608 do {
1609 unsigned int len =
1610 min_t(u64, BIT_ULL(31), size - copied);
1611
1612 if (__copy_from_user((char *)relocs + copied,
1613 (char __user *)urelocs + copied,
1614 len))
1615 goto end;
1616
1617 copied += len;
1618 } while (copied < size);
1619
1620 /*
1621 * As we do not update the known relocation offsets after
1622 * relocating (due to the complexities in lock handling),
1623 * we need to mark them as invalid now so that we force the
1624 * relocation processing next time. Just in case the target
1625 * object is evicted and then rebound into its old
1626 * presumed_offset before the next execbuffer - if that
1627 * happened we would make the mistake of assuming that the
1628 * relocations were valid.
1629 */
1630 if (!user_access_begin(urelocs, size))
1631 goto end;
1632
1633 for (copied = 0; copied < nreloc; copied++)
1634 unsafe_put_user(-1,
1635 &urelocs[copied].presumed_offset,
1636 end_user);
1637 user_access_end();
1638
1639 eb->exec[i].relocs_ptr = (uintptr_t)relocs;
1640 }
1641
1642 return 0;
1643
1644end_user:
1645 user_access_end();
1646end:
1647 kvfree(relocs);
1648 err = -EFAULT;
1649err:
1650 while (i--) {
1651 relocs = u64_to_ptr(typeof(*relocs), eb->exec[i].relocs_ptr);
1652 if (eb->exec[i].relocation_count)
1653 kvfree(relocs);
1654 }
1655 return err;
1656}
1657
1658static int eb_prefault_relocations(const struct i915_execbuffer *eb)
1659{
1660 const unsigned int count = eb->buffer_count;
1661 unsigned int i;
1662
1663 if (unlikely(i915_modparams.prefault_disable))
1664 return 0;
1665
1666 for (i = 0; i < count; i++) {
1667 int err;
1668
1669 err = check_relocations(&eb->exec[i]);
1670 if (err)
1671 return err;
1672 }
1673
1674 return 0;
1675}
1676
1677static noinline int eb_relocate_slow(struct i915_execbuffer *eb)
1678{
1679 struct drm_device *dev = &eb->i915->drm;
1680 bool have_copy = false;
1681 struct i915_vma *vma;
1682 int err = 0;
1683
1684repeat:
1685 if (signal_pending(current)) {
1686 err = -ERESTARTSYS;
1687 goto out;
1688 }
1689
1690 /* We may process another execbuffer during the unlock... */
1691 eb_reset_vmas(eb);
1692 mutex_unlock(&dev->struct_mutex);
1693
1694 /*
1695 * We take 3 passes through the slowpatch.
1696 *
1697 * 1 - we try to just prefault all the user relocation entries and
1698 * then attempt to reuse the atomic pagefault disabled fast path again.
1699 *
1700 * 2 - we copy the user entries to a local buffer here outside of the
1701 * local and allow ourselves to wait upon any rendering before
1702 * relocations
1703 *
1704 * 3 - we already have a local copy of the relocation entries, but
1705 * were interrupted (EAGAIN) whilst waiting for the objects, try again.
1706 */
1707 if (!err) {
1708 err = eb_prefault_relocations(eb);
1709 } else if (!have_copy) {
1710 err = eb_copy_relocations(eb);
1711 have_copy = err == 0;
1712 } else {
1713 cond_resched();
1714 err = 0;
1715 }
1716 if (err) {
1717 mutex_lock(&dev->struct_mutex);
1718 goto out;
1719 }
1720
1721 /* A frequent cause for EAGAIN are currently unavailable client pages */
1722 flush_workqueue(eb->i915->mm.userptr_wq);
1723
1724 err = i915_mutex_lock_interruptible(dev);
1725 if (err) {
1726 mutex_lock(&dev->struct_mutex);
1727 goto out;
1728 }
1729
1730 /* reacquire the objects */
1731 err = eb_lookup_vmas(eb);
1732 if (err)
1733 goto err;
1734
1735 GEM_BUG_ON(!eb->batch);
1736
1737 list_for_each_entry(vma, &eb->relocs, reloc_link) {
1738 if (!have_copy) {
1739 pagefault_disable();
1740 err = eb_relocate_vma(eb, vma);
1741 pagefault_enable();
1742 if (err)
1743 goto repeat;
1744 } else {
1745 err = eb_relocate_vma_slow(eb, vma);
1746 if (err)
1747 goto err;
1748 }
1749 }
1750
1751 /*
1752 * Leave the user relocations as are, this is the painfully slow path,
1753 * and we want to avoid the complication of dropping the lock whilst
1754 * having buffers reserved in the aperture and so causing spurious
1755 * ENOSPC for random operations.
1756 */
1757
1758err:
1759 if (err == -EAGAIN)
1760 goto repeat;
1761
1762out:
1763 if (have_copy) {
1764 const unsigned int count = eb->buffer_count;
1765 unsigned int i;
1766
1767 for (i = 0; i < count; i++) {
1768 const struct drm_i915_gem_exec_object2 *entry =
1769 &eb->exec[i];
1770 struct drm_i915_gem_relocation_entry *relocs;
1771
1772 if (!entry->relocation_count)
1773 continue;
1774
1775 relocs = u64_to_ptr(typeof(*relocs), entry->relocs_ptr);
1776 kvfree(relocs);
1777 }
1778 }
1779
1780 return err;
1781}
1782
1783static int eb_relocate(struct i915_execbuffer *eb)
1784{
1785 if (eb_lookup_vmas(eb))
1786 goto slow;
1787
1788 /* The objects are in their final locations, apply the relocations. */
1789 if (eb->args->flags & __EXEC_HAS_RELOC) {
1790 struct i915_vma *vma;
1791
1792 list_for_each_entry(vma, &eb->relocs, reloc_link) {
1793 if (eb_relocate_vma(eb, vma))
1794 goto slow;
1795 }
1796 }
1797
1798 return 0;
1799
1800slow:
1801 return eb_relocate_slow(eb);
1802}
1803
1804static int eb_move_to_gpu(struct i915_execbuffer *eb)
1805{
1806 const unsigned int count = eb->buffer_count;
1807 struct ww_acquire_ctx acquire;
1808 unsigned int i;
1809 int err = 0;
1810
1811 ww_acquire_init(&acquire, &reservation_ww_class);
1812
1813 for (i = 0; i < count; i++) {
1814 struct i915_vma *vma = eb->vma[i];
1815
1816 err = ww_mutex_lock_interruptible(&vma->resv->lock, &acquire);
1817 if (!err)
1818 continue;
1819
1820 GEM_BUG_ON(err == -EALREADY); /* No duplicate vma */
1821
1822 if (err == -EDEADLK) {
1823 GEM_BUG_ON(i == 0);
1824 do {
1825 int j = i - 1;
1826
1827 ww_mutex_unlock(&eb->vma[j]->resv->lock);
1828
1829 swap(eb->flags[i], eb->flags[j]);
1830 swap(eb->vma[i], eb->vma[j]);
1831 eb->vma[i]->exec_flags = &eb->flags[i];
1832 } while (--i);
1833 GEM_BUG_ON(vma != eb->vma[0]);
1834 vma->exec_flags = &eb->flags[0];
1835
1836 err = ww_mutex_lock_slow_interruptible(&vma->resv->lock,
1837 &acquire);
1838 }
1839 if (err)
1840 break;
1841 }
1842 ww_acquire_done(&acquire);
1843
1844 while (i--) {
1845 unsigned int flags = eb->flags[i];
1846 struct i915_vma *vma = eb->vma[i];
1847 struct drm_i915_gem_object *obj = vma->obj;
1848
1849 assert_vma_held(vma);
1850
1851 if (flags & EXEC_OBJECT_CAPTURE) {
1852 struct i915_capture_list *capture;
1853
1854 capture = kmalloc(sizeof(*capture), GFP_KERNEL);
1855 if (capture) {
1856 capture->next = eb->request->capture_list;
1857 capture->vma = vma;
1858 eb->request->capture_list = capture;
1859 }
1860 }
1861
1862 /*
1863 * If the GPU is not _reading_ through the CPU cache, we need
1864 * to make sure that any writes (both previous GPU writes from
1865 * before a change in snooping levels and normal CPU writes)
1866 * caught in that cache are flushed to main memory.
1867 *
1868 * We want to say
1869 * obj->cache_dirty &&
1870 * !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ)
1871 * but gcc's optimiser doesn't handle that as well and emits
1872 * two jumps instead of one. Maybe one day...
1873 */
1874 if (unlikely(obj->cache_dirty & ~obj->cache_coherent)) {
1875 if (i915_gem_clflush_object(obj, 0))
1876 flags &= ~EXEC_OBJECT_ASYNC;
1877 }
1878
1879 if (err == 0 && !(flags & EXEC_OBJECT_ASYNC)) {
1880 err = i915_request_await_object
1881 (eb->request, obj, flags & EXEC_OBJECT_WRITE);
1882 }
1883
1884 if (err == 0)
1885 err = i915_vma_move_to_active(vma, eb->request, flags);
1886
1887 i915_vma_unlock(vma);
1888
1889 __eb_unreserve_vma(vma, flags);
1890 vma->exec_flags = NULL;
1891
1892 if (unlikely(flags & __EXEC_OBJECT_HAS_REF))
1893 i915_vma_put(vma);
1894 }
1895 ww_acquire_fini(&acquire);
1896
1897 if (unlikely(err))
1898 goto err_skip;
1899
1900 eb->exec = NULL;
1901
1902 /* Unconditionally flush any chipset caches (for streaming writes). */
1903 intel_gt_chipset_flush(eb->engine->gt);
1904 return 0;
1905
1906err_skip:
1907 i915_request_skip(eb->request, err);
1908 return err;
1909}
1910
1911static bool i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
1912{
1913 if (exec->flags & __I915_EXEC_ILLEGAL_FLAGS)
1914 return false;
1915
1916 /* Kernel clipping was a DRI1 misfeature */
1917 if (!(exec->flags & I915_EXEC_FENCE_ARRAY)) {
1918 if (exec->num_cliprects || exec->cliprects_ptr)
1919 return false;
1920 }
1921
1922 if (exec->DR4 == 0xffffffff) {
1923 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
1924 exec->DR4 = 0;
1925 }
1926 if (exec->DR1 || exec->DR4)
1927 return false;
1928
1929 if ((exec->batch_start_offset | exec->batch_len) & 0x7)
1930 return false;
1931
1932 return true;
1933}
1934
1935static int i915_reset_gen7_sol_offsets(struct i915_request *rq)
1936{
1937 u32 *cs;
1938 int i;
1939
1940 if (!IS_GEN(rq->i915, 7) || rq->engine->id != RCS0) {
1941 DRM_DEBUG("sol reset is gen7/rcs only\n");
1942 return -EINVAL;
1943 }
1944
1945 cs = intel_ring_begin(rq, 4 * 2 + 2);
1946 if (IS_ERR(cs))
1947 return PTR_ERR(cs);
1948
1949 *cs++ = MI_LOAD_REGISTER_IMM(4);
1950 for (i = 0; i < 4; i++) {
1951 *cs++ = i915_mmio_reg_offset(GEN7_SO_WRITE_OFFSET(i));
1952 *cs++ = 0;
1953 }
1954 *cs++ = MI_NOOP;
1955 intel_ring_advance(rq, cs);
1956
1957 return 0;
1958}
1959
1960static struct i915_vma *
1961shadow_batch_pin(struct i915_execbuffer *eb, struct drm_i915_gem_object *obj)
1962{
1963 struct drm_i915_private *dev_priv = eb->i915;
1964 struct i915_vma * const vma = *eb->vma;
1965 struct i915_address_space *vm;
1966 u64 flags;
1967
1968 /*
1969 * PPGTT backed shadow buffers must be mapped RO, to prevent
1970 * post-scan tampering
1971 */
1972 if (CMDPARSER_USES_GGTT(dev_priv)) {
1973 flags = PIN_GLOBAL;
1974 vm = &dev_priv->ggtt.vm;
1975 } else if (vma->vm->has_read_only) {
1976 flags = PIN_USER;
1977 vm = vma->vm;
1978 i915_gem_object_set_readonly(obj);
1979 } else {
1980 DRM_DEBUG("Cannot prevent post-scan tampering without RO capable vm\n");
1981 return ERR_PTR(-EINVAL);
1982 }
1983
1984 return i915_gem_object_pin(obj, vm, NULL, 0, 0, flags);
1985}
1986
1987static struct i915_vma *eb_parse(struct i915_execbuffer *eb)
1988{
1989 struct intel_engine_pool_node *pool;
1990 struct i915_vma *vma;
1991 u64 batch_start;
1992 u64 shadow_batch_start;
1993 int err;
1994
1995 pool = intel_engine_pool_get(&eb->engine->pool, eb->batch_len);
1996 if (IS_ERR(pool))
1997 return ERR_CAST(pool);
1998
1999 vma = shadow_batch_pin(eb, pool->obj);
2000 if (IS_ERR(vma))
2001 goto err;
2002
2003 batch_start = gen8_canonical_addr(eb->batch->node.start) +
2004 eb->batch_start_offset;
2005
2006 shadow_batch_start = gen8_canonical_addr(vma->node.start);
2007
2008 err = intel_engine_cmd_parser(eb->gem_context,
2009 eb->engine,
2010 eb->batch->obj,
2011 batch_start,
2012 eb->batch_start_offset,
2013 eb->batch_len,
2014 pool->obj,
2015 shadow_batch_start);
2016
2017 if (err) {
2018 i915_vma_unpin(vma);
2019
2020 /*
2021 * Unsafe GGTT-backed buffers can still be submitted safely
2022 * as non-secure.
2023 * For PPGTT backing however, we have no choice but to forcibly
2024 * reject unsafe buffers
2025 */
2026 if (CMDPARSER_USES_GGTT(eb->i915) && (err == -EACCES))
2027 /* Execute original buffer non-secure */
2028 vma = NULL;
2029 else
2030 vma = ERR_PTR(err);
2031 goto err;
2032 }
2033
2034 eb->vma[eb->buffer_count] = i915_vma_get(vma);
2035 eb->flags[eb->buffer_count] =
2036 __EXEC_OBJECT_HAS_PIN | __EXEC_OBJECT_HAS_REF;
2037 vma->exec_flags = &eb->flags[eb->buffer_count];
2038 eb->buffer_count++;
2039
2040 eb->batch_start_offset = 0;
2041 eb->batch = vma;
2042
2043 if (CMDPARSER_USES_GGTT(eb->i915))
2044 eb->batch_flags |= I915_DISPATCH_SECURE;
2045
2046 /* eb->batch_len unchanged */
2047
2048 vma->private = pool;
2049 return vma;
2050
2051err:
2052 intel_engine_pool_put(pool);
2053 return vma;
2054}
2055
2056static void
2057add_to_client(struct i915_request *rq, struct drm_file *file)
2058{
2059 struct drm_i915_file_private *file_priv = file->driver_priv;
2060
2061 rq->file_priv = file_priv;
2062
2063 spin_lock(&file_priv->mm.lock);
2064 list_add_tail(&rq->client_link, &file_priv->mm.request_list);
2065 spin_unlock(&file_priv->mm.lock);
2066}
2067
2068static int eb_submit(struct i915_execbuffer *eb)
2069{
2070 int err;
2071
2072 err = eb_move_to_gpu(eb);
2073 if (err)
2074 return err;
2075
2076 if (eb->args->flags & I915_EXEC_GEN7_SOL_RESET) {
2077 err = i915_reset_gen7_sol_offsets(eb->request);
2078 if (err)
2079 return err;
2080 }
2081
2082 /*
2083 * After we completed waiting for other engines (using HW semaphores)
2084 * then we can signal that this request/batch is ready to run. This
2085 * allows us to determine if the batch is still waiting on the GPU
2086 * or actually running by checking the breadcrumb.
2087 */
2088 if (eb->engine->emit_init_breadcrumb) {
2089 err = eb->engine->emit_init_breadcrumb(eb->request);
2090 if (err)
2091 return err;
2092 }
2093
2094 err = eb->engine->emit_bb_start(eb->request,
2095 eb->batch->node.start +
2096 eb->batch_start_offset,
2097 eb->batch_len,
2098 eb->batch_flags);
2099 if (err)
2100 return err;
2101
2102 return 0;
2103}
2104
2105static int num_vcs_engines(const struct drm_i915_private *i915)
2106{
2107 return hweight64(INTEL_INFO(i915)->engine_mask &
2108 GENMASK_ULL(VCS0 + I915_MAX_VCS - 1, VCS0));
2109}
2110
2111/*
2112 * Find one BSD ring to dispatch the corresponding BSD command.
2113 * The engine index is returned.
2114 */
2115static unsigned int
2116gen8_dispatch_bsd_engine(struct drm_i915_private *dev_priv,
2117 struct drm_file *file)
2118{
2119 struct drm_i915_file_private *file_priv = file->driver_priv;
2120
2121 /* Check whether the file_priv has already selected one ring. */
2122 if ((int)file_priv->bsd_engine < 0)
2123 file_priv->bsd_engine =
2124 get_random_int() % num_vcs_engines(dev_priv);
2125
2126 return file_priv->bsd_engine;
2127}
2128
2129static const enum intel_engine_id user_ring_map[] = {
2130 [I915_EXEC_DEFAULT] = RCS0,
2131 [I915_EXEC_RENDER] = RCS0,
2132 [I915_EXEC_BLT] = BCS0,
2133 [I915_EXEC_BSD] = VCS0,
2134 [I915_EXEC_VEBOX] = VECS0
2135};
2136
2137static struct i915_request *eb_throttle(struct intel_context *ce)
2138{
2139 struct intel_ring *ring = ce->ring;
2140 struct intel_timeline *tl = ce->timeline;
2141 struct i915_request *rq;
2142
2143 /*
2144 * Completely unscientific finger-in-the-air estimates for suitable
2145 * maximum user request size (to avoid blocking) and then backoff.
2146 */
2147 if (intel_ring_update_space(ring) >= PAGE_SIZE)
2148 return NULL;
2149
2150 /*
2151 * Find a request that after waiting upon, there will be at least half
2152 * the ring available. The hysteresis allows us to compete for the
2153 * shared ring and should mean that we sleep less often prior to
2154 * claiming our resources, but not so long that the ring completely
2155 * drains before we can submit our next request.
2156 */
2157 list_for_each_entry(rq, &tl->requests, link) {
2158 if (rq->ring != ring)
2159 continue;
2160
2161 if (__intel_ring_space(rq->postfix,
2162 ring->emit, ring->size) > ring->size / 2)
2163 break;
2164 }
2165 if (&rq->link == &tl->requests)
2166 return NULL; /* weird, we will check again later for real */
2167
2168 return i915_request_get(rq);
2169}
2170
2171static int
2172__eb_pin_context(struct i915_execbuffer *eb, struct intel_context *ce)
2173{
2174 int err;
2175
2176 if (likely(atomic_inc_not_zero(&ce->pin_count)))
2177 return 0;
2178
2179 err = mutex_lock_interruptible(&eb->i915->drm.struct_mutex);
2180 if (err)
2181 return err;
2182
2183 err = __intel_context_do_pin(ce);
2184 mutex_unlock(&eb->i915->drm.struct_mutex);
2185
2186 return err;
2187}
2188
2189static void
2190__eb_unpin_context(struct i915_execbuffer *eb, struct intel_context *ce)
2191{
2192 if (likely(atomic_add_unless(&ce->pin_count, -1, 1)))
2193 return;
2194
2195 mutex_lock(&eb->i915->drm.struct_mutex);
2196 intel_context_unpin(ce);
2197 mutex_unlock(&eb->i915->drm.struct_mutex);
2198}
2199
2200static int __eb_pin_engine(struct i915_execbuffer *eb, struct intel_context *ce)
2201{
2202 struct intel_timeline *tl;
2203 struct i915_request *rq;
2204 int err;
2205
2206 /*
2207 * ABI: Before userspace accesses the GPU (e.g. execbuffer), report
2208 * EIO if the GPU is already wedged.
2209 */
2210 err = intel_gt_terminally_wedged(ce->engine->gt);
2211 if (err)
2212 return err;
2213
2214 /*
2215 * Pinning the contexts may generate requests in order to acquire
2216 * GGTT space, so do this first before we reserve a seqno for
2217 * ourselves.
2218 */
2219 err = __eb_pin_context(eb, ce);
2220 if (err)
2221 return err;
2222
2223 /*
2224 * Take a local wakeref for preparing to dispatch the execbuf as
2225 * we expect to access the hardware fairly frequently in the
2226 * process, and require the engine to be kept awake between accesses.
2227 * Upon dispatch, we acquire another prolonged wakeref that we hold
2228 * until the timeline is idle, which in turn releases the wakeref
2229 * taken on the engine, and the parent device.
2230 */
2231 tl = intel_context_timeline_lock(ce);
2232 if (IS_ERR(tl)) {
2233 err = PTR_ERR(tl);
2234 goto err_unpin;
2235 }
2236
2237 intel_context_enter(ce);
2238 rq = eb_throttle(ce);
2239
2240 intel_context_timeline_unlock(tl);
2241
2242 if (rq) {
2243 if (i915_request_wait(rq,
2244 I915_WAIT_INTERRUPTIBLE,
2245 MAX_SCHEDULE_TIMEOUT) < 0) {
2246 i915_request_put(rq);
2247 err = -EINTR;
2248 goto err_exit;
2249 }
2250
2251 i915_request_put(rq);
2252 }
2253
2254 eb->engine = ce->engine;
2255 eb->context = ce;
2256 return 0;
2257
2258err_exit:
2259 mutex_lock(&tl->mutex);
2260 intel_context_exit(ce);
2261 intel_context_timeline_unlock(tl);
2262err_unpin:
2263 __eb_unpin_context(eb, ce);
2264 return err;
2265}
2266
2267static void eb_unpin_engine(struct i915_execbuffer *eb)
2268{
2269 struct intel_context *ce = eb->context;
2270 struct intel_timeline *tl = ce->timeline;
2271
2272 mutex_lock(&tl->mutex);
2273 intel_context_exit(ce);
2274 mutex_unlock(&tl->mutex);
2275
2276 __eb_unpin_context(eb, ce);
2277}
2278
2279static unsigned int
2280eb_select_legacy_ring(struct i915_execbuffer *eb,
2281 struct drm_file *file,
2282 struct drm_i915_gem_execbuffer2 *args)
2283{
2284 struct drm_i915_private *i915 = eb->i915;
2285 unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK;
2286
2287 if (user_ring_id != I915_EXEC_BSD &&
2288 (args->flags & I915_EXEC_BSD_MASK)) {
2289 DRM_DEBUG("execbuf with non bsd ring but with invalid "
2290 "bsd dispatch flags: %d\n", (int)(args->flags));
2291 return -1;
2292 }
2293
2294 if (user_ring_id == I915_EXEC_BSD && num_vcs_engines(i915) > 1) {
2295 unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK;
2296
2297 if (bsd_idx == I915_EXEC_BSD_DEFAULT) {
2298 bsd_idx = gen8_dispatch_bsd_engine(i915, file);
2299 } else if (bsd_idx >= I915_EXEC_BSD_RING1 &&
2300 bsd_idx <= I915_EXEC_BSD_RING2) {
2301 bsd_idx >>= I915_EXEC_BSD_SHIFT;
2302 bsd_idx--;
2303 } else {
2304 DRM_DEBUG("execbuf with unknown bsd ring: %u\n",
2305 bsd_idx);
2306 return -1;
2307 }
2308
2309 return _VCS(bsd_idx);
2310 }
2311
2312 if (user_ring_id >= ARRAY_SIZE(user_ring_map)) {
2313 DRM_DEBUG("execbuf with unknown ring: %u\n", user_ring_id);
2314 return -1;
2315 }
2316
2317 return user_ring_map[user_ring_id];
2318}
2319
2320static int
2321eb_pin_engine(struct i915_execbuffer *eb,
2322 struct drm_file *file,
2323 struct drm_i915_gem_execbuffer2 *args)
2324{
2325 struct intel_context *ce;
2326 unsigned int idx;
2327 int err;
2328
2329 if (i915_gem_context_user_engines(eb->gem_context))
2330 idx = args->flags & I915_EXEC_RING_MASK;
2331 else
2332 idx = eb_select_legacy_ring(eb, file, args);
2333
2334 ce = i915_gem_context_get_engine(eb->gem_context, idx);
2335 if (IS_ERR(ce))
2336 return PTR_ERR(ce);
2337
2338 err = __eb_pin_engine(eb, ce);
2339 intel_context_put(ce);
2340
2341 return err;
2342}
2343
2344static void
2345__free_fence_array(struct drm_syncobj **fences, unsigned int n)
2346{
2347 while (n--)
2348 drm_syncobj_put(ptr_mask_bits(fences[n], 2));
2349 kvfree(fences);
2350}
2351
2352static struct drm_syncobj **
2353get_fence_array(struct drm_i915_gem_execbuffer2 *args,
2354 struct drm_file *file)
2355{
2356 const unsigned long nfences = args->num_cliprects;
2357 struct drm_i915_gem_exec_fence __user *user;
2358 struct drm_syncobj **fences;
2359 unsigned long n;
2360 int err;
2361
2362 if (!(args->flags & I915_EXEC_FENCE_ARRAY))
2363 return NULL;
2364
2365 /* Check multiplication overflow for access_ok() and kvmalloc_array() */
2366 BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long));
2367 if (nfences > min_t(unsigned long,
2368 ULONG_MAX / sizeof(*user),
2369 SIZE_MAX / sizeof(*fences)))
2370 return ERR_PTR(-EINVAL);
2371
2372 user = u64_to_user_ptr(args->cliprects_ptr);
2373 if (!access_ok(user, nfences * sizeof(*user)))
2374 return ERR_PTR(-EFAULT);
2375
2376 fences = kvmalloc_array(nfences, sizeof(*fences),
2377 __GFP_NOWARN | GFP_KERNEL);
2378 if (!fences)
2379 return ERR_PTR(-ENOMEM);
2380
2381 for (n = 0; n < nfences; n++) {
2382 struct drm_i915_gem_exec_fence fence;
2383 struct drm_syncobj *syncobj;
2384
2385 if (__copy_from_user(&fence, user++, sizeof(fence))) {
2386 err = -EFAULT;
2387 goto err;
2388 }
2389
2390 if (fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS) {
2391 err = -EINVAL;
2392 goto err;
2393 }
2394
2395 syncobj = drm_syncobj_find(file, fence.handle);
2396 if (!syncobj) {
2397 DRM_DEBUG("Invalid syncobj handle provided\n");
2398 err = -ENOENT;
2399 goto err;
2400 }
2401
2402 BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) &
2403 ~__I915_EXEC_FENCE_UNKNOWN_FLAGS);
2404
2405 fences[n] = ptr_pack_bits(syncobj, fence.flags, 2);
2406 }
2407
2408 return fences;
2409
2410err:
2411 __free_fence_array(fences, n);
2412 return ERR_PTR(err);
2413}
2414
2415static void
2416put_fence_array(struct drm_i915_gem_execbuffer2 *args,
2417 struct drm_syncobj **fences)
2418{
2419 if (fences)
2420 __free_fence_array(fences, args->num_cliprects);
2421}
2422
2423static int
2424await_fence_array(struct i915_execbuffer *eb,
2425 struct drm_syncobj **fences)
2426{
2427 const unsigned int nfences = eb->args->num_cliprects;
2428 unsigned int n;
2429 int err;
2430
2431 for (n = 0; n < nfences; n++) {
2432 struct drm_syncobj *syncobj;
2433 struct dma_fence *fence;
2434 unsigned int flags;
2435
2436 syncobj = ptr_unpack_bits(fences[n], &flags, 2);
2437 if (!(flags & I915_EXEC_FENCE_WAIT))
2438 continue;
2439
2440 fence = drm_syncobj_fence_get(syncobj);
2441 if (!fence)
2442 return -EINVAL;
2443
2444 err = i915_request_await_dma_fence(eb->request, fence);
2445 dma_fence_put(fence);
2446 if (err < 0)
2447 return err;
2448 }
2449
2450 return 0;
2451}
2452
2453static void
2454signal_fence_array(struct i915_execbuffer *eb,
2455 struct drm_syncobj **fences)
2456{
2457 const unsigned int nfences = eb->args->num_cliprects;
2458 struct dma_fence * const fence = &eb->request->fence;
2459 unsigned int n;
2460
2461 for (n = 0; n < nfences; n++) {
2462 struct drm_syncobj *syncobj;
2463 unsigned int flags;
2464
2465 syncobj = ptr_unpack_bits(fences[n], &flags, 2);
2466 if (!(flags & I915_EXEC_FENCE_SIGNAL))
2467 continue;
2468
2469 drm_syncobj_replace_fence(syncobj, fence);
2470 }
2471}
2472
2473static int
2474i915_gem_do_execbuffer(struct drm_device *dev,
2475 struct drm_file *file,
2476 struct drm_i915_gem_execbuffer2 *args,
2477 struct drm_i915_gem_exec_object2 *exec,
2478 struct drm_syncobj **fences)
2479{
2480 struct drm_i915_private *i915 = to_i915(dev);
2481 struct i915_execbuffer eb;
2482 struct dma_fence *in_fence = NULL;
2483 struct dma_fence *exec_fence = NULL;
2484 struct sync_file *out_fence = NULL;
2485 int out_fence_fd = -1;
2486 int err;
2487
2488 BUILD_BUG_ON(__EXEC_INTERNAL_FLAGS & ~__I915_EXEC_ILLEGAL_FLAGS);
2489 BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS &
2490 ~__EXEC_OBJECT_UNKNOWN_FLAGS);
2491
2492 eb.i915 = i915;
2493 eb.file = file;
2494 eb.args = args;
2495 if (DBG_FORCE_RELOC || !(args->flags & I915_EXEC_NO_RELOC))
2496 args->flags |= __EXEC_HAS_RELOC;
2497
2498 eb.exec = exec;
2499 eb.vma = (struct i915_vma **)(exec + args->buffer_count + 1);
2500 eb.vma[0] = NULL;
2501 eb.flags = (unsigned int *)(eb.vma + args->buffer_count + 1);
2502
2503 eb.invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
2504 reloc_cache_init(&eb.reloc_cache, eb.i915);
2505
2506 eb.buffer_count = args->buffer_count;
2507 eb.batch_start_offset = args->batch_start_offset;
2508 eb.batch_len = args->batch_len;
2509
2510 eb.batch_flags = 0;
2511 if (args->flags & I915_EXEC_SECURE) {
2512 if (INTEL_GEN(i915) >= 11)
2513 return -ENODEV;
2514
2515 /* Return -EPERM to trigger fallback code on old binaries. */
2516 if (!HAS_SECURE_BATCHES(i915))
2517 return -EPERM;
2518
2519 if (!drm_is_current_master(file) || !capable(CAP_SYS_ADMIN))
2520 return -EPERM;
2521
2522 eb.batch_flags |= I915_DISPATCH_SECURE;
2523 }
2524 if (args->flags & I915_EXEC_IS_PINNED)
2525 eb.batch_flags |= I915_DISPATCH_PINNED;
2526
2527 if (args->flags & I915_EXEC_FENCE_IN) {
2528 in_fence = sync_file_get_fence(lower_32_bits(args->rsvd2));
2529 if (!in_fence)
2530 return -EINVAL;
2531 }
2532
2533 if (args->flags & I915_EXEC_FENCE_SUBMIT) {
2534 if (in_fence) {
2535 err = -EINVAL;
2536 goto err_in_fence;
2537 }
2538
2539 exec_fence = sync_file_get_fence(lower_32_bits(args->rsvd2));
2540 if (!exec_fence) {
2541 err = -EINVAL;
2542 goto err_in_fence;
2543 }
2544 }
2545
2546 if (args->flags & I915_EXEC_FENCE_OUT) {
2547 out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
2548 if (out_fence_fd < 0) {
2549 err = out_fence_fd;
2550 goto err_exec_fence;
2551 }
2552 }
2553
2554 err = eb_create(&eb);
2555 if (err)
2556 goto err_out_fence;
2557
2558 GEM_BUG_ON(!eb.lut_size);
2559
2560 err = eb_select_context(&eb);
2561 if (unlikely(err))
2562 goto err_destroy;
2563
2564 err = eb_pin_engine(&eb, file, args);
2565 if (unlikely(err))
2566 goto err_context;
2567
2568 err = i915_mutex_lock_interruptible(dev);
2569 if (err)
2570 goto err_engine;
2571
2572 err = eb_relocate(&eb);
2573 if (err) {
2574 /*
2575 * If the user expects the execobject.offset and
2576 * reloc.presumed_offset to be an exact match,
2577 * as for using NO_RELOC, then we cannot update
2578 * the execobject.offset until we have completed
2579 * relocation.
2580 */
2581 args->flags &= ~__EXEC_HAS_RELOC;
2582 goto err_vma;
2583 }
2584
2585 if (unlikely(*eb.batch->exec_flags & EXEC_OBJECT_WRITE)) {
2586 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
2587 err = -EINVAL;
2588 goto err_vma;
2589 }
2590 if (eb.batch_start_offset > eb.batch->size ||
2591 eb.batch_len > eb.batch->size - eb.batch_start_offset) {
2592 DRM_DEBUG("Attempting to use out-of-bounds batch\n");
2593 err = -EINVAL;
2594 goto err_vma;
2595 }
2596
2597 if (eb.batch_len == 0)
2598 eb.batch_len = eb.batch->size - eb.batch_start_offset;
2599
2600 if (eb_use_cmdparser(&eb)) {
2601 struct i915_vma *vma;
2602
2603 vma = eb_parse(&eb);
2604 if (IS_ERR(vma)) {
2605 err = PTR_ERR(vma);
2606 goto err_vma;
2607 }
2608 }
2609
2610 /*
2611 * snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
2612 * batch" bit. Hence we need to pin secure batches into the global gtt.
2613 * hsw should have this fixed, but bdw mucks it up again. */
2614 if (eb.batch_flags & I915_DISPATCH_SECURE) {
2615 struct i915_vma *vma;
2616
2617 /*
2618 * So on first glance it looks freaky that we pin the batch here
2619 * outside of the reservation loop. But:
2620 * - The batch is already pinned into the relevant ppgtt, so we
2621 * already have the backing storage fully allocated.
2622 * - No other BO uses the global gtt (well contexts, but meh),
2623 * so we don't really have issues with multiple objects not
2624 * fitting due to fragmentation.
2625 * So this is actually safe.
2626 */
2627 vma = i915_gem_object_ggtt_pin(eb.batch->obj, NULL, 0, 0, 0);
2628 if (IS_ERR(vma)) {
2629 err = PTR_ERR(vma);
2630 goto err_vma;
2631 }
2632
2633 eb.batch = vma;
2634 }
2635
2636 /* All GPU relocation batches must be submitted prior to the user rq */
2637 GEM_BUG_ON(eb.reloc_cache.rq);
2638
2639 /* Allocate a request for this batch buffer nice and early. */
2640 eb.request = i915_request_create(eb.context);
2641 if (IS_ERR(eb.request)) {
2642 err = PTR_ERR(eb.request);
2643 goto err_batch_unpin;
2644 }
2645
2646 if (in_fence) {
2647 err = i915_request_await_dma_fence(eb.request, in_fence);
2648 if (err < 0)
2649 goto err_request;
2650 }
2651
2652 if (exec_fence) {
2653 err = i915_request_await_execution(eb.request, exec_fence,
2654 eb.engine->bond_execute);
2655 if (err < 0)
2656 goto err_request;
2657 }
2658
2659 if (fences) {
2660 err = await_fence_array(&eb, fences);
2661 if (err)
2662 goto err_request;
2663 }
2664
2665 if (out_fence_fd != -1) {
2666 out_fence = sync_file_create(&eb.request->fence);
2667 if (!out_fence) {
2668 err = -ENOMEM;
2669 goto err_request;
2670 }
2671 }
2672
2673 /*
2674 * Whilst this request exists, batch_obj will be on the
2675 * active_list, and so will hold the active reference. Only when this
2676 * request is retired will the the batch_obj be moved onto the
2677 * inactive_list and lose its active reference. Hence we do not need
2678 * to explicitly hold another reference here.
2679 */
2680 eb.request->batch = eb.batch;
2681 if (eb.batch->private)
2682 intel_engine_pool_mark_active(eb.batch->private, eb.request);
2683
2684 trace_i915_request_queue(eb.request, eb.batch_flags);
2685 err = eb_submit(&eb);
2686err_request:
2687 add_to_client(eb.request, file);
2688 i915_request_add(eb.request);
2689
2690 if (fences)
2691 signal_fence_array(&eb, fences);
2692
2693 if (out_fence) {
2694 if (err == 0) {
2695 fd_install(out_fence_fd, out_fence->file);
2696 args->rsvd2 &= GENMASK_ULL(31, 0); /* keep in-fence */
2697 args->rsvd2 |= (u64)out_fence_fd << 32;
2698 out_fence_fd = -1;
2699 } else {
2700 fput(out_fence->file);
2701 }
2702 }
2703
2704err_batch_unpin:
2705 if (eb.batch_flags & I915_DISPATCH_SECURE)
2706 i915_vma_unpin(eb.batch);
2707 if (eb.batch->private)
2708 intel_engine_pool_put(eb.batch->private);
2709err_vma:
2710 if (eb.exec)
2711 eb_release_vmas(&eb);
2712 mutex_unlock(&dev->struct_mutex);
2713err_engine:
2714 eb_unpin_engine(&eb);
2715err_context:
2716 i915_gem_context_put(eb.gem_context);
2717err_destroy:
2718 eb_destroy(&eb);
2719err_out_fence:
2720 if (out_fence_fd != -1)
2721 put_unused_fd(out_fence_fd);
2722err_exec_fence:
2723 dma_fence_put(exec_fence);
2724err_in_fence:
2725 dma_fence_put(in_fence);
2726 return err;
2727}
2728
2729static size_t eb_element_size(void)
2730{
2731 return (sizeof(struct drm_i915_gem_exec_object2) +
2732 sizeof(struct i915_vma *) +
2733 sizeof(unsigned int));
2734}
2735
2736static bool check_buffer_count(size_t count)
2737{
2738 const size_t sz = eb_element_size();
2739
2740 /*
2741 * When using LUT_HANDLE, we impose a limit of INT_MAX for the lookup
2742 * array size (see eb_create()). Otherwise, we can accept an array as
2743 * large as can be addressed (though use large arrays at your peril)!
2744 */
2745
2746 return !(count < 1 || count > INT_MAX || count > SIZE_MAX / sz - 1);
2747}
2748
2749/*
2750 * Legacy execbuffer just creates an exec2 list from the original exec object
2751 * list array and passes it to the real function.
2752 */
2753int
2754i915_gem_execbuffer_ioctl(struct drm_device *dev, void *data,
2755 struct drm_file *file)
2756{
2757 struct drm_i915_gem_execbuffer *args = data;
2758 struct drm_i915_gem_execbuffer2 exec2;
2759 struct drm_i915_gem_exec_object *exec_list = NULL;
2760 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
2761 const size_t count = args->buffer_count;
2762 unsigned int i;
2763 int err;
2764
2765 if (!check_buffer_count(count)) {
2766 DRM_DEBUG("execbuf2 with %zd buffers\n", count);
2767 return -EINVAL;
2768 }
2769
2770 exec2.buffers_ptr = args->buffers_ptr;
2771 exec2.buffer_count = args->buffer_count;
2772 exec2.batch_start_offset = args->batch_start_offset;
2773 exec2.batch_len = args->batch_len;
2774 exec2.DR1 = args->DR1;
2775 exec2.DR4 = args->DR4;
2776 exec2.num_cliprects = args->num_cliprects;
2777 exec2.cliprects_ptr = args->cliprects_ptr;
2778 exec2.flags = I915_EXEC_RENDER;
2779 i915_execbuffer2_set_context_id(exec2, 0);
2780
2781 if (!i915_gem_check_execbuffer(&exec2))
2782 return -EINVAL;
2783
2784 /* Copy in the exec list from userland */
2785 exec_list = kvmalloc_array(count, sizeof(*exec_list),
2786 __GFP_NOWARN | GFP_KERNEL);
2787 exec2_list = kvmalloc_array(count + 1, eb_element_size(),
2788 __GFP_NOWARN | GFP_KERNEL);
2789 if (exec_list == NULL || exec2_list == NULL) {
2790 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
2791 args->buffer_count);
2792 kvfree(exec_list);
2793 kvfree(exec2_list);
2794 return -ENOMEM;
2795 }
2796 err = copy_from_user(exec_list,
2797 u64_to_user_ptr(args->buffers_ptr),
2798 sizeof(*exec_list) * count);
2799 if (err) {
2800 DRM_DEBUG("copy %d exec entries failed %d\n",
2801 args->buffer_count, err);
2802 kvfree(exec_list);
2803 kvfree(exec2_list);
2804 return -EFAULT;
2805 }
2806
2807 for (i = 0; i < args->buffer_count; i++) {
2808 exec2_list[i].handle = exec_list[i].handle;
2809 exec2_list[i].relocation_count = exec_list[i].relocation_count;
2810 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
2811 exec2_list[i].alignment = exec_list[i].alignment;
2812 exec2_list[i].offset = exec_list[i].offset;
2813 if (INTEL_GEN(to_i915(dev)) < 4)
2814 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
2815 else
2816 exec2_list[i].flags = 0;
2817 }
2818
2819 err = i915_gem_do_execbuffer(dev, file, &exec2, exec2_list, NULL);
2820 if (exec2.flags & __EXEC_HAS_RELOC) {
2821 struct drm_i915_gem_exec_object __user *user_exec_list =
2822 u64_to_user_ptr(args->buffers_ptr);
2823
2824 /* Copy the new buffer offsets back to the user's exec list. */
2825 for (i = 0; i < args->buffer_count; i++) {
2826 if (!(exec2_list[i].offset & UPDATE))
2827 continue;
2828
2829 exec2_list[i].offset =
2830 gen8_canonical_addr(exec2_list[i].offset & PIN_OFFSET_MASK);
2831 exec2_list[i].offset &= PIN_OFFSET_MASK;
2832 if (__copy_to_user(&user_exec_list[i].offset,
2833 &exec2_list[i].offset,
2834 sizeof(user_exec_list[i].offset)))
2835 break;
2836 }
2837 }
2838
2839 kvfree(exec_list);
2840 kvfree(exec2_list);
2841 return err;
2842}
2843
2844int
2845i915_gem_execbuffer2_ioctl(struct drm_device *dev, void *data,
2846 struct drm_file *file)
2847{
2848 struct drm_i915_gem_execbuffer2 *args = data;
2849 struct drm_i915_gem_exec_object2 *exec2_list;
2850 struct drm_syncobj **fences = NULL;
2851 const size_t count = args->buffer_count;
2852 int err;
2853
2854 if (!check_buffer_count(count)) {
2855 DRM_DEBUG("execbuf2 with %zd buffers\n", count);
2856 return -EINVAL;
2857 }
2858
2859 if (!i915_gem_check_execbuffer(args))
2860 return -EINVAL;
2861
2862 /* Allocate an extra slot for use by the command parser */
2863 exec2_list = kvmalloc_array(count + 1, eb_element_size(),
2864 __GFP_NOWARN | GFP_KERNEL);
2865 if (exec2_list == NULL) {
2866 DRM_DEBUG("Failed to allocate exec list for %zd buffers\n",
2867 count);
2868 return -ENOMEM;
2869 }
2870 if (copy_from_user(exec2_list,
2871 u64_to_user_ptr(args->buffers_ptr),
2872 sizeof(*exec2_list) * count)) {
2873 DRM_DEBUG("copy %zd exec entries failed\n", count);
2874 kvfree(exec2_list);
2875 return -EFAULT;
2876 }
2877
2878 if (args->flags & I915_EXEC_FENCE_ARRAY) {
2879 fences = get_fence_array(args, file);
2880 if (IS_ERR(fences)) {
2881 kvfree(exec2_list);
2882 return PTR_ERR(fences);
2883 }
2884 }
2885
2886 err = i915_gem_do_execbuffer(dev, file, args, exec2_list, fences);
2887
2888 /*
2889 * Now that we have begun execution of the batchbuffer, we ignore
2890 * any new error after this point. Also given that we have already
2891 * updated the associated relocations, we try to write out the current
2892 * object locations irrespective of any error.
2893 */
2894 if (args->flags & __EXEC_HAS_RELOC) {
2895 struct drm_i915_gem_exec_object2 __user *user_exec_list =
2896 u64_to_user_ptr(args->buffers_ptr);
2897 unsigned int i;
2898
2899 /* Copy the new buffer offsets back to the user's exec list. */
2900 /*
2901 * Note: count * sizeof(*user_exec_list) does not overflow,
2902 * because we checked 'count' in check_buffer_count().
2903 *
2904 * And this range already got effectively checked earlier
2905 * when we did the "copy_from_user()" above.
2906 */
2907 if (!user_access_begin(user_exec_list, count * sizeof(*user_exec_list)))
2908 goto end;
2909
2910 for (i = 0; i < args->buffer_count; i++) {
2911 if (!(exec2_list[i].offset & UPDATE))
2912 continue;
2913
2914 exec2_list[i].offset =
2915 gen8_canonical_addr(exec2_list[i].offset & PIN_OFFSET_MASK);
2916 unsafe_put_user(exec2_list[i].offset,
2917 &user_exec_list[i].offset,
2918 end_user);
2919 }
2920end_user:
2921 user_access_end();
2922end:;
2923 }
2924
2925 args->flags &= ~__I915_EXEC_UNKNOWN_FLAGS;
2926 put_fence_array(args, fences);
2927 kvfree(exec2_list);
2928 return err;
2929}