2 * SPDX-License-Identifier: MIT
4 * Copyright © 2008,2010 Intel Corporation
7 #include <linux/intel-iommu.h>
8 #include <linux/dma-resv.h>
9 #include <linux/sync_file.h>
10 #include <linux/uaccess.h>
12 #include <drm/drm_syncobj.h>
14 #include "display/intel_frontbuffer.h"
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"
24 #include "pxp/intel_pxp.h"
26 #include "i915_cmd_parser.h"
28 #include "i915_file_private.h"
29 #include "i915_gem_clflush.h"
30 #include "i915_gem_context.h"
31 #include "i915_gem_evict.h"
32 #include "i915_gem_ioctls.h"
33 #include "i915_trace.h"
34 #include "i915_user_extensions.h"
40 /** This vma's place in the execbuf reservation list */
41 struct drm_i915_gem_exec_object2 *exec;
42 struct list_head bind_link;
43 struct list_head reloc_link;
45 struct hlist_node node;
53 #define DBG_FORCE_RELOC 0 /* choose one of the above! */
56 /* __EXEC_OBJECT_NO_RESERVE is BIT(31), defined in i915_vma.h */
57 #define __EXEC_OBJECT_HAS_PIN BIT(30)
58 #define __EXEC_OBJECT_HAS_FENCE BIT(29)
59 #define __EXEC_OBJECT_USERPTR_INIT BIT(28)
60 #define __EXEC_OBJECT_NEEDS_MAP BIT(27)
61 #define __EXEC_OBJECT_NEEDS_BIAS BIT(26)
62 #define __EXEC_OBJECT_INTERNAL_FLAGS (~0u << 26) /* all of the above + */
63 #define __EXEC_OBJECT_RESERVED (__EXEC_OBJECT_HAS_PIN | __EXEC_OBJECT_HAS_FENCE)
65 #define __EXEC_HAS_RELOC BIT(31)
66 #define __EXEC_ENGINE_PINNED BIT(30)
67 #define __EXEC_USERPTR_USED BIT(29)
68 #define __EXEC_INTERNAL_FLAGS (~0u << 29)
69 #define UPDATE PIN_OFFSET_FIXED
71 #define BATCH_OFFSET_BIAS (256*1024)
73 #define __I915_EXEC_ILLEGAL_FLAGS \
74 (__I915_EXEC_UNKNOWN_FLAGS | \
75 I915_EXEC_CONSTANTS_MASK | \
76 I915_EXEC_RESOURCE_STREAMER)
78 /* Catch emission of unexpected errors for CI! */
79 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
82 DRM_DEBUG_DRIVER("EINVAL at %s:%d\n", __func__, __LINE__); \
88 * DOC: User command execution
90 * Userspace submits commands to be executed on the GPU as an instruction
91 * stream within a GEM object we call a batchbuffer. This instructions may
92 * refer to other GEM objects containing auxiliary state such as kernels,
93 * samplers, render targets and even secondary batchbuffers. Userspace does
94 * not know where in the GPU memory these objects reside and so before the
95 * batchbuffer is passed to the GPU for execution, those addresses in the
96 * batchbuffer and auxiliary objects are updated. This is known as relocation,
97 * or patching. To try and avoid having to relocate each object on the next
98 * execution, userspace is told the location of those objects in this pass,
99 * but this remains just a hint as the kernel may choose a new location for
100 * any object in the future.
102 * At the level of talking to the hardware, submitting a batchbuffer for the
103 * GPU to execute is to add content to a buffer from which the HW
104 * command streamer is reading.
106 * 1. Add a command to load the HW context. For Logical Ring Contexts, i.e.
107 * Execlists, this command is not placed on the same buffer as the
110 * 2. Add a command to invalidate caches to the buffer.
112 * 3. Add a batchbuffer start command to the buffer; the start command is
113 * essentially a token together with the GPU address of the batchbuffer
116 * 4. Add a pipeline flush to the buffer.
118 * 5. Add a memory write command to the buffer to record when the GPU
119 * is done executing the batchbuffer. The memory write writes the
120 * global sequence number of the request, ``i915_request::global_seqno``;
121 * the i915 driver uses the current value in the register to determine
122 * if the GPU has completed the batchbuffer.
124 * 6. Add a user interrupt command to the buffer. This command instructs
125 * the GPU to issue an interrupt when the command, pipeline flush and
126 * memory write are completed.
128 * 7. Inform the hardware of the additional commands added to the buffer
129 * (by updating the tail pointer).
131 * Processing an execbuf ioctl is conceptually split up into a few phases.
133 * 1. Validation - Ensure all the pointers, handles and flags are valid.
134 * 2. Reservation - Assign GPU address space for every object
135 * 3. Relocation - Update any addresses to point to the final locations
136 * 4. Serialisation - Order the request with respect to its dependencies
137 * 5. Construction - Construct a request to execute the batchbuffer
138 * 6. Submission (at some point in the future execution)
140 * Reserving resources for the execbuf is the most complicated phase. We
141 * neither want to have to migrate the object in the address space, nor do
142 * we want to have to update any relocations pointing to this object. Ideally,
143 * we want to leave the object where it is and for all the existing relocations
144 * to match. If the object is given a new address, or if userspace thinks the
145 * object is elsewhere, we have to parse all the relocation entries and update
146 * the addresses. Userspace can set the I915_EXEC_NORELOC flag to hint that
147 * all the target addresses in all of its objects match the value in the
148 * relocation entries and that they all match the presumed offsets given by the
149 * list of execbuffer objects. Using this knowledge, we know that if we haven't
150 * moved any buffers, all the relocation entries are valid and we can skip
151 * the update. (If userspace is wrong, the likely outcome is an impromptu GPU
152 * hang.) The requirement for using I915_EXEC_NO_RELOC are:
154 * The addresses written in the objects must match the corresponding
155 * reloc.presumed_offset which in turn must match the corresponding
158 * Any render targets written to in the batch must be flagged with
161 * To avoid stalling, execobject.offset should match the current
162 * address of that object within the active context.
164 * The reservation is done is multiple phases. First we try and keep any
165 * object already bound in its current location - so as long as meets the
166 * constraints imposed by the new execbuffer. Any object left unbound after the
167 * first pass is then fitted into any available idle space. If an object does
168 * not fit, all objects are removed from the reservation and the process rerun
169 * after sorting the objects into a priority order (more difficult to fit
170 * objects are tried first). Failing that, the entire VM is cleared and we try
171 * to fit the execbuf once last time before concluding that it simply will not
174 * A small complication to all of this is that we allow userspace not only to
175 * specify an alignment and a size for the object in the address space, but
176 * we also allow userspace to specify the exact offset. This objects are
177 * simpler to place (the location is known a priori) all we have to do is make
178 * sure the space is available.
180 * Once all the objects are in place, patching up the buried pointers to point
181 * to the final locations is a fairly simple job of walking over the relocation
182 * entry arrays, looking up the right address and rewriting the value into
183 * the object. Simple! ... The relocation entries are stored in user memory
184 * and so to access them we have to copy them into a local buffer. That copy
185 * has to avoid taking any pagefaults as they may lead back to a GEM object
186 * requiring the struct_mutex (i.e. recursive deadlock). So once again we split
187 * the relocation into multiple passes. First we try to do everything within an
188 * atomic context (avoid the pagefaults) which requires that we never wait. If
189 * we detect that we may wait, or if we need to fault, then we have to fallback
190 * to a slower path. The slowpath has to drop the mutex. (Can you hear alarm
191 * bells yet?) Dropping the mutex means that we lose all the state we have
192 * built up so far for the execbuf and we must reset any global data. However,
193 * we do leave the objects pinned in their final locations - which is a
194 * potential issue for concurrent execbufs. Once we have left the mutex, we can
195 * allocate and copy all the relocation entries into a large array at our
196 * leisure, reacquire the mutex, reclaim all the objects and other state and
197 * then proceed to update any incorrect addresses with the objects.
199 * As we process the relocation entries, we maintain a record of whether the
200 * object is being written to. Using NORELOC, we expect userspace to provide
201 * this information instead. We also check whether we can skip the relocation
202 * by comparing the expected value inside the relocation entry with the target's
203 * final address. If they differ, we have to map the current object and rewrite
204 * the 4 or 8 byte pointer within.
206 * Serialising an execbuf is quite simple according to the rules of the GEM
207 * ABI. Execution within each context is ordered by the order of submission.
208 * Writes to any GEM object are in order of submission and are exclusive. Reads
209 * from a GEM object are unordered with respect to other reads, but ordered by
210 * writes. A write submitted after a read cannot occur before the read, and
211 * similarly any read submitted after a write cannot occur before the write.
212 * Writes are ordered between engines such that only one write occurs at any
213 * time (completing any reads beforehand) - using semaphores where available
214 * and CPU serialisation otherwise. Other GEM access obey the same rules, any
215 * write (either via mmaps using set-domain, or via pwrite) must flush all GPU
216 * reads before starting, and any read (either using set-domain or pread) must
217 * flush all GPU writes before starting. (Note we only employ a barrier before,
218 * we currently rely on userspace not concurrently starting a new execution
219 * whilst reading or writing to an object. This may be an advantage or not
220 * depending on how much you trust userspace not to shoot themselves in the
221 * foot.) Serialisation may just result in the request being inserted into
222 * a DAG awaiting its turn, but most simple is to wait on the CPU until
223 * all dependencies are resolved.
225 * After all of that, is just a matter of closing the request and handing it to
226 * the hardware (well, leaving it in a queue to be executed). However, we also
227 * offer the ability for batchbuffers to be run with elevated privileges so
228 * that they access otherwise hidden registers. (Used to adjust L3 cache etc.)
229 * Before any batch is given extra privileges we first must check that it
230 * contains no nefarious instructions, we check that each instruction is from
231 * our whitelist and all registers are also from an allowed list. We first
232 * copy the user's batchbuffer to a shadow (so that the user doesn't have
233 * access to it, either by the CPU or GPU as we scan it) and then parse each
234 * instruction. If everything is ok, we set a flag telling the hardware to run
235 * the batchbuffer in trusted mode, otherwise the ioctl is rejected.
239 struct drm_syncobj *syncobj; /* Use with ptr_mask_bits() */
240 struct dma_fence *dma_fence;
242 struct dma_fence_chain *chain_fence;
245 struct i915_execbuffer {
246 struct drm_i915_private *i915; /** i915 backpointer */
247 struct drm_file *file; /** per-file lookup tables and limits */
248 struct drm_i915_gem_execbuffer2 *args; /** ioctl parameters */
249 struct drm_i915_gem_exec_object2 *exec; /** ioctl execobj[] */
252 struct intel_gt *gt; /* gt for the execbuf */
253 struct intel_context *context; /* logical state for the request */
254 struct i915_gem_context *gem_context; /** caller's context */
256 /** our requests to build */
257 struct i915_request *requests[MAX_ENGINE_INSTANCE + 1];
258 /** identity of the batch obj/vma */
259 struct eb_vma *batches[MAX_ENGINE_INSTANCE + 1];
260 struct i915_vma *trampoline; /** trampoline used for chaining */
262 /** used for excl fence in dma_resv objects when > 1 BB submitted */
263 struct dma_fence *composite_fence;
265 /** actual size of execobj[] as we may extend it for the cmdparser */
266 unsigned int buffer_count;
268 /* number of batches in execbuf IOCTL */
269 unsigned int num_batches;
271 /** list of vma not yet bound during reservation phase */
272 struct list_head unbound;
274 /** list of vma that have execobj.relocation_count */
275 struct list_head relocs;
277 struct i915_gem_ww_ctx ww;
280 * Track the most recently used object for relocations, as we
281 * frequently have to perform multiple relocations within the same
285 struct drm_mm_node node; /** temporary GTT binding */
286 unsigned long vaddr; /** Current kmap address */
287 unsigned long page; /** Currently mapped page index */
288 unsigned int graphics_ver; /** Cached value of GRAPHICS_VER */
289 bool use_64bit_reloc : 1;
292 bool needs_unfenced : 1;
295 u64 invalid_flags; /** Set of execobj.flags that are invalid */
297 /** Length of batch within object */
298 u64 batch_len[MAX_ENGINE_INSTANCE + 1];
299 u32 batch_start_offset; /** Location within object of batch */
300 u32 batch_flags; /** Flags composed for emit_bb_start() */
301 struct intel_gt_buffer_pool_node *batch_pool; /** pool node for batch buffer */
304 * Indicate either the size of the hastable used to resolve
305 * relocation handles, or if negative that we are using a direct
306 * index into the execobj[].
309 struct hlist_head *buckets; /** ht for relocation handles */
311 struct eb_fence *fences;
312 unsigned long num_fences;
313 #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR)
314 struct i915_capture_list *capture_lists[MAX_ENGINE_INSTANCE + 1];
318 static int eb_parse(struct i915_execbuffer *eb);
319 static int eb_pin_engine(struct i915_execbuffer *eb, bool throttle);
320 static void eb_unpin_engine(struct i915_execbuffer *eb);
321 static void eb_capture_release(struct i915_execbuffer *eb);
323 static inline bool eb_use_cmdparser(const struct i915_execbuffer *eb)
325 return intel_engine_requires_cmd_parser(eb->context->engine) ||
326 (intel_engine_using_cmd_parser(eb->context->engine) &&
327 eb->args->batch_len);
330 static int eb_create(struct i915_execbuffer *eb)
332 if (!(eb->args->flags & I915_EXEC_HANDLE_LUT)) {
333 unsigned int size = 1 + ilog2(eb->buffer_count);
336 * Without a 1:1 association between relocation handles and
337 * the execobject[] index, we instead create a hashtable.
338 * We size it dynamically based on available memory, starting
339 * first with 1:1 assocative hash and scaling back until
340 * the allocation succeeds.
342 * Later on we use a positive lut_size to indicate we are
343 * using this hashtable, and a negative value to indicate a
349 /* While we can still reduce the allocation size, don't
350 * raise a warning and allow the allocation to fail.
351 * On the last pass though, we want to try as hard
352 * as possible to perform the allocation and warn
357 flags |= __GFP_NORETRY | __GFP_NOWARN;
359 eb->buckets = kzalloc(sizeof(struct hlist_head) << size,
370 eb->lut_size = -eb->buffer_count;
377 eb_vma_misplaced(const struct drm_i915_gem_exec_object2 *entry,
378 const struct i915_vma *vma,
381 if (vma->node.size < entry->pad_to_size)
384 if (entry->alignment && !IS_ALIGNED(vma->node.start, entry->alignment))
387 if (flags & EXEC_OBJECT_PINNED &&
388 vma->node.start != entry->offset)
391 if (flags & __EXEC_OBJECT_NEEDS_BIAS &&
392 vma->node.start < BATCH_OFFSET_BIAS)
395 if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) &&
396 (vma->node.start + vma->node.size + 4095) >> 32)
399 if (flags & __EXEC_OBJECT_NEEDS_MAP &&
400 !i915_vma_is_map_and_fenceable(vma))
406 static u64 eb_pin_flags(const struct drm_i915_gem_exec_object2 *entry,
407 unsigned int exec_flags)
411 if (exec_flags & EXEC_OBJECT_NEEDS_GTT)
412 pin_flags |= PIN_GLOBAL;
415 * Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset,
416 * limit address to the first 4GBs for unflagged objects.
418 if (!(exec_flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS))
419 pin_flags |= PIN_ZONE_4G;
421 if (exec_flags & __EXEC_OBJECT_NEEDS_MAP)
422 pin_flags |= PIN_MAPPABLE;
424 if (exec_flags & EXEC_OBJECT_PINNED)
425 pin_flags |= entry->offset | PIN_OFFSET_FIXED;
426 else if (exec_flags & __EXEC_OBJECT_NEEDS_BIAS)
427 pin_flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
433 eb_pin_vma(struct i915_execbuffer *eb,
434 const struct drm_i915_gem_exec_object2 *entry,
437 struct i915_vma *vma = ev->vma;
442 pin_flags = vma->node.start;
444 pin_flags = entry->offset & PIN_OFFSET_MASK;
446 pin_flags |= PIN_USER | PIN_NOEVICT | PIN_OFFSET_FIXED | PIN_VALIDATE;
447 if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_GTT))
448 pin_flags |= PIN_GLOBAL;
450 /* Attempt to reuse the current location if available */
451 err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, pin_flags);
456 if (entry->flags & EXEC_OBJECT_PINNED)
459 /* Failing that pick any _free_ space if suitable */
460 err = i915_vma_pin_ww(vma, &eb->ww,
463 eb_pin_flags(entry, ev->flags) |
464 PIN_USER | PIN_NOEVICT | PIN_VALIDATE);
469 if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) {
470 err = i915_vma_pin_fence(vma);
475 ev->flags |= __EXEC_OBJECT_HAS_FENCE;
478 ev->flags |= __EXEC_OBJECT_HAS_PIN;
479 if (eb_vma_misplaced(entry, vma, ev->flags))
486 eb_unreserve_vma(struct eb_vma *ev)
488 if (unlikely(ev->flags & __EXEC_OBJECT_HAS_FENCE))
489 __i915_vma_unpin_fence(ev->vma);
491 ev->flags &= ~__EXEC_OBJECT_RESERVED;
495 eb_validate_vma(struct i915_execbuffer *eb,
496 struct drm_i915_gem_exec_object2 *entry,
497 struct i915_vma *vma)
499 /* Relocations are disallowed for all platforms after TGL-LP. This
500 * also covers all platforms with local memory.
502 if (entry->relocation_count &&
503 GRAPHICS_VER(eb->i915) >= 12 && !IS_TIGERLAKE(eb->i915))
506 if (unlikely(entry->flags & eb->invalid_flags))
509 if (unlikely(entry->alignment &&
510 !is_power_of_2_u64(entry->alignment)))
514 * Offset can be used as input (EXEC_OBJECT_PINNED), reject
515 * any non-page-aligned or non-canonical addresses.
517 if (unlikely(entry->flags & EXEC_OBJECT_PINNED &&
518 entry->offset != gen8_canonical_addr(entry->offset & I915_GTT_PAGE_MASK)))
521 /* pad_to_size was once a reserved field, so sanitize it */
522 if (entry->flags & EXEC_OBJECT_PAD_TO_SIZE) {
523 if (unlikely(offset_in_page(entry->pad_to_size)))
526 entry->pad_to_size = 0;
529 * From drm_mm perspective address space is continuous,
530 * so from this point we're always using non-canonical
533 entry->offset = gen8_noncanonical_addr(entry->offset);
535 if (!eb->reloc_cache.has_fence) {
536 entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
538 if ((entry->flags & EXEC_OBJECT_NEEDS_FENCE ||
539 eb->reloc_cache.needs_unfenced) &&
540 i915_gem_object_is_tiled(vma->obj))
541 entry->flags |= EXEC_OBJECT_NEEDS_GTT | __EXEC_OBJECT_NEEDS_MAP;
548 is_batch_buffer(struct i915_execbuffer *eb, unsigned int buffer_idx)
550 return eb->args->flags & I915_EXEC_BATCH_FIRST ?
551 buffer_idx < eb->num_batches :
552 buffer_idx >= eb->args->buffer_count - eb->num_batches;
556 eb_add_vma(struct i915_execbuffer *eb,
557 unsigned int *current_batch,
559 struct i915_vma *vma)
561 struct drm_i915_private *i915 = eb->i915;
562 struct drm_i915_gem_exec_object2 *entry = &eb->exec[i];
563 struct eb_vma *ev = &eb->vma[i];
567 ev->flags = entry->flags;
569 if (eb->lut_size > 0) {
570 ev->handle = entry->handle;
571 hlist_add_head(&ev->node,
572 &eb->buckets[hash_32(entry->handle,
576 if (entry->relocation_count)
577 list_add_tail(&ev->reloc_link, &eb->relocs);
580 * SNA is doing fancy tricks with compressing batch buffers, which leads
581 * to negative relocation deltas. Usually that works out ok since the
582 * relocate address is still positive, except when the batch is placed
583 * very low in the GTT. Ensure this doesn't happen.
585 * Note that actual hangs have only been observed on gen7, but for
586 * paranoia do it everywhere.
588 if (is_batch_buffer(eb, i)) {
589 if (entry->relocation_count &&
590 !(ev->flags & EXEC_OBJECT_PINNED))
591 ev->flags |= __EXEC_OBJECT_NEEDS_BIAS;
592 if (eb->reloc_cache.has_fence)
593 ev->flags |= EXEC_OBJECT_NEEDS_FENCE;
595 eb->batches[*current_batch] = ev;
597 if (unlikely(ev->flags & EXEC_OBJECT_WRITE)) {
599 "Attempting to use self-modifying batch buffer\n");
603 if (range_overflows_t(u64,
604 eb->batch_start_offset,
607 drm_dbg(&i915->drm, "Attempting to use out-of-bounds batch\n");
611 if (eb->args->batch_len == 0)
612 eb->batch_len[*current_batch] = ev->vma->size -
613 eb->batch_start_offset;
615 eb->batch_len[*current_batch] = eb->args->batch_len;
616 if (unlikely(eb->batch_len[*current_batch] == 0)) { /* impossible! */
617 drm_dbg(&i915->drm, "Invalid batch length\n");
627 static inline int use_cpu_reloc(const struct reloc_cache *cache,
628 const struct drm_i915_gem_object *obj)
630 if (!i915_gem_object_has_struct_page(obj))
633 if (DBG_FORCE_RELOC == FORCE_CPU_RELOC)
636 if (DBG_FORCE_RELOC == FORCE_GTT_RELOC)
639 return (cache->has_llc ||
641 obj->cache_level != I915_CACHE_NONE);
644 static int eb_reserve_vma(struct i915_execbuffer *eb,
648 struct drm_i915_gem_exec_object2 *entry = ev->exec;
649 struct i915_vma *vma = ev->vma;
652 if (drm_mm_node_allocated(&vma->node) &&
653 eb_vma_misplaced(entry, vma, ev->flags)) {
654 err = i915_vma_unbind(vma);
659 err = i915_vma_pin_ww(vma, &eb->ww,
660 entry->pad_to_size, entry->alignment,
661 eb_pin_flags(entry, ev->flags) | pin_flags);
665 if (entry->offset != vma->node.start) {
666 entry->offset = vma->node.start | UPDATE;
667 eb->args->flags |= __EXEC_HAS_RELOC;
670 if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) {
671 err = i915_vma_pin_fence(vma);
676 ev->flags |= __EXEC_OBJECT_HAS_FENCE;
679 ev->flags |= __EXEC_OBJECT_HAS_PIN;
680 GEM_BUG_ON(eb_vma_misplaced(entry, vma, ev->flags));
685 static bool eb_unbind(struct i915_execbuffer *eb, bool force)
687 const unsigned int count = eb->buffer_count;
689 struct list_head last;
690 bool unpinned = false;
692 /* Resort *all* the objects into priority order */
693 INIT_LIST_HEAD(&eb->unbound);
694 INIT_LIST_HEAD(&last);
696 for (i = 0; i < count; i++) {
697 struct eb_vma *ev = &eb->vma[i];
698 unsigned int flags = ev->flags;
700 if (!force && flags & EXEC_OBJECT_PINNED &&
701 flags & __EXEC_OBJECT_HAS_PIN)
705 eb_unreserve_vma(ev);
707 if (flags & EXEC_OBJECT_PINNED)
708 /* Pinned must have their slot */
709 list_add(&ev->bind_link, &eb->unbound);
710 else if (flags & __EXEC_OBJECT_NEEDS_MAP)
711 /* Map require the lowest 256MiB (aperture) */
712 list_add_tail(&ev->bind_link, &eb->unbound);
713 else if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS))
714 /* Prioritise 4GiB region for restricted bo */
715 list_add(&ev->bind_link, &last);
717 list_add_tail(&ev->bind_link, &last);
720 list_splice_tail(&last, &eb->unbound);
724 static int eb_reserve(struct i915_execbuffer *eb)
732 * Attempt to pin all of the buffers into the GTT.
733 * This is done in 2 phases:
735 * 1. Unbind all objects that do not match the GTT constraints for
736 * the execbuffer (fenceable, mappable, alignment etc).
737 * 2. Bind new objects.
739 * This avoid unnecessary unbinding of later objects in order to make
740 * room for the earlier objects *unless* we need to defragment.
742 * Defragmenting is skipped if all objects are pinned at a fixed location.
744 for (pass = 0; pass <= 2; pass++) {
745 int pin_flags = PIN_USER | PIN_VALIDATE;
748 pin_flags |= PIN_NONBLOCK;
751 unpinned = eb_unbind(eb, pass == 2);
754 err = mutex_lock_interruptible(&eb->context->vm->mutex);
756 err = i915_gem_evict_vm(eb->context->vm, &eb->ww);
757 mutex_unlock(&eb->context->vm->mutex);
763 list_for_each_entry(ev, &eb->unbound, bind_link) {
764 err = eb_reserve_vma(eb, ev, pin_flags);
776 static int eb_select_context(struct i915_execbuffer *eb)
778 struct i915_gem_context *ctx;
780 ctx = i915_gem_context_lookup(eb->file->driver_priv, eb->args->rsvd1);
781 if (unlikely(IS_ERR(ctx)))
784 eb->gem_context = ctx;
785 if (i915_gem_context_has_full_ppgtt(ctx))
786 eb->invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
791 static int __eb_add_lut(struct i915_execbuffer *eb,
792 u32 handle, struct i915_vma *vma)
794 struct i915_gem_context *ctx = eb->gem_context;
795 struct i915_lut_handle *lut;
798 lut = i915_lut_handle_alloc();
803 if (!atomic_fetch_inc(&vma->open_count))
804 i915_vma_reopen(vma);
805 lut->handle = handle;
808 /* Check that the context hasn't been closed in the meantime */
810 if (!mutex_lock_interruptible(&ctx->lut_mutex)) {
811 if (likely(!i915_gem_context_is_closed(ctx)))
812 err = radix_tree_insert(&ctx->handles_vma, handle, vma);
815 if (err == 0) { /* And nor has this handle */
816 struct drm_i915_gem_object *obj = vma->obj;
818 spin_lock(&obj->lut_lock);
819 if (idr_find(&eb->file->object_idr, handle) == obj) {
820 list_add(&lut->obj_link, &obj->lut_list);
822 radix_tree_delete(&ctx->handles_vma, handle);
825 spin_unlock(&obj->lut_lock);
827 mutex_unlock(&ctx->lut_mutex);
837 i915_lut_handle_free(lut);
841 static struct i915_vma *eb_lookup_vma(struct i915_execbuffer *eb, u32 handle)
843 struct i915_address_space *vm = eb->context->vm;
846 struct drm_i915_gem_object *obj;
847 struct i915_vma *vma;
851 vma = radix_tree_lookup(&eb->gem_context->handles_vma, handle);
852 if (likely(vma && vma->vm == vm))
853 vma = i915_vma_tryget(vma);
858 obj = i915_gem_object_lookup(eb->file, handle);
860 return ERR_PTR(-ENOENT);
863 * If the user has opted-in for protected-object tracking, make
864 * sure the object encryption can be used.
865 * We only need to do this when the object is first used with
866 * this context, because the context itself will be banned when
867 * the protected objects become invalid.
869 if (i915_gem_context_uses_protected_content(eb->gem_context) &&
870 i915_gem_object_is_protected(obj)) {
871 err = intel_pxp_key_check(&vm->gt->pxp, obj, true);
873 i915_gem_object_put(obj);
878 vma = i915_vma_instance(obj, vm, NULL);
880 i915_gem_object_put(obj);
884 err = __eb_add_lut(eb, handle, vma);
888 i915_gem_object_put(obj);
894 static int eb_lookup_vmas(struct i915_execbuffer *eb)
896 unsigned int i, current_batch = 0;
899 INIT_LIST_HEAD(&eb->relocs);
901 for (i = 0; i < eb->buffer_count; i++) {
902 struct i915_vma *vma;
904 vma = eb_lookup_vma(eb, eb->exec[i].handle);
910 err = eb_validate_vma(eb, &eb->exec[i], vma);
916 err = eb_add_vma(eb, ¤t_batch, i, vma);
920 if (i915_gem_object_is_userptr(vma->obj)) {
921 err = i915_gem_object_userptr_submit_init(vma->obj);
923 if (i + 1 < eb->buffer_count) {
925 * Execbuffer code expects last vma entry to be NULL,
926 * since we already initialized this entry,
927 * set the next value to NULL or we mess up
930 eb->vma[i + 1].vma = NULL;
936 eb->vma[i].flags |= __EXEC_OBJECT_USERPTR_INIT;
937 eb->args->flags |= __EXEC_USERPTR_USED;
944 eb->vma[i].vma = NULL;
948 static int eb_lock_vmas(struct i915_execbuffer *eb)
953 for (i = 0; i < eb->buffer_count; i++) {
954 struct eb_vma *ev = &eb->vma[i];
955 struct i915_vma *vma = ev->vma;
957 err = i915_gem_object_lock(vma->obj, &eb->ww);
965 static int eb_validate_vmas(struct i915_execbuffer *eb)
970 INIT_LIST_HEAD(&eb->unbound);
972 err = eb_lock_vmas(eb);
976 for (i = 0; i < eb->buffer_count; i++) {
977 struct drm_i915_gem_exec_object2 *entry = &eb->exec[i];
978 struct eb_vma *ev = &eb->vma[i];
979 struct i915_vma *vma = ev->vma;
981 err = eb_pin_vma(eb, entry, ev);
986 if (entry->offset != vma->node.start) {
987 entry->offset = vma->node.start | UPDATE;
988 eb->args->flags |= __EXEC_HAS_RELOC;
991 eb_unreserve_vma(ev);
993 list_add_tail(&ev->bind_link, &eb->unbound);
994 if (drm_mm_node_allocated(&vma->node)) {
995 err = i915_vma_unbind(vma);
1001 if (!(ev->flags & EXEC_OBJECT_WRITE)) {
1002 err = dma_resv_reserve_shared(vma->obj->base.resv, 1);
1007 GEM_BUG_ON(drm_mm_node_allocated(&vma->node) &&
1008 eb_vma_misplaced(&eb->exec[i], vma, ev->flags));
1011 if (!list_empty(&eb->unbound))
1012 return eb_reserve(eb);
1017 static struct eb_vma *
1018 eb_get_vma(const struct i915_execbuffer *eb, unsigned long handle)
1020 if (eb->lut_size < 0) {
1021 if (handle >= -eb->lut_size)
1023 return &eb->vma[handle];
1025 struct hlist_head *head;
1028 head = &eb->buckets[hash_32(handle, eb->lut_size)];
1029 hlist_for_each_entry(ev, head, node) {
1030 if (ev->handle == handle)
1037 static void eb_release_vmas(struct i915_execbuffer *eb, bool final)
1039 const unsigned int count = eb->buffer_count;
1042 for (i = 0; i < count; i++) {
1043 struct eb_vma *ev = &eb->vma[i];
1044 struct i915_vma *vma = ev->vma;
1049 eb_unreserve_vma(ev);
1055 eb_capture_release(eb);
1056 eb_unpin_engine(eb);
1059 static void eb_destroy(const struct i915_execbuffer *eb)
1061 if (eb->lut_size > 0)
1066 relocation_target(const struct drm_i915_gem_relocation_entry *reloc,
1067 const struct i915_vma *target)
1069 return gen8_canonical_addr((int)reloc->delta + target->node.start);
1072 static void reloc_cache_init(struct reloc_cache *cache,
1073 struct drm_i915_private *i915)
1077 /* Must be a variable in the struct to allow GCC to unroll. */
1078 cache->graphics_ver = GRAPHICS_VER(i915);
1079 cache->has_llc = HAS_LLC(i915);
1080 cache->use_64bit_reloc = HAS_64BIT_RELOC(i915);
1081 cache->has_fence = cache->graphics_ver < 4;
1082 cache->needs_unfenced = INTEL_INFO(i915)->unfenced_needs_alignment;
1083 cache->node.flags = 0;
1086 static inline void *unmask_page(unsigned long p)
1088 return (void *)(uintptr_t)(p & PAGE_MASK);
1091 static inline unsigned int unmask_flags(unsigned long p)
1093 return p & ~PAGE_MASK;
1096 #define KMAP 0x4 /* after CLFLUSH_FLAGS */
1098 static inline struct i915_ggtt *cache_to_ggtt(struct reloc_cache *cache)
1100 struct drm_i915_private *i915 =
1101 container_of(cache, struct i915_execbuffer, reloc_cache)->i915;
1102 return to_gt(i915)->ggtt;
1105 static void reloc_cache_unmap(struct reloc_cache *cache)
1112 vaddr = unmask_page(cache->vaddr);
1113 if (cache->vaddr & KMAP)
1114 kunmap_atomic(vaddr);
1116 io_mapping_unmap_atomic((void __iomem *)vaddr);
1119 static void reloc_cache_remap(struct reloc_cache *cache,
1120 struct drm_i915_gem_object *obj)
1127 if (cache->vaddr & KMAP) {
1128 struct page *page = i915_gem_object_get_page(obj, cache->page);
1130 vaddr = kmap_atomic(page);
1131 cache->vaddr = unmask_flags(cache->vaddr) |
1132 (unsigned long)vaddr;
1134 struct i915_ggtt *ggtt = cache_to_ggtt(cache);
1135 unsigned long offset;
1137 offset = cache->node.start;
1138 if (!drm_mm_node_allocated(&cache->node))
1139 offset += cache->page << PAGE_SHIFT;
1141 cache->vaddr = (unsigned long)
1142 io_mapping_map_atomic_wc(&ggtt->iomap, offset);
1146 static void reloc_cache_reset(struct reloc_cache *cache, struct i915_execbuffer *eb)
1153 vaddr = unmask_page(cache->vaddr);
1154 if (cache->vaddr & KMAP) {
1155 struct drm_i915_gem_object *obj =
1156 (struct drm_i915_gem_object *)cache->node.mm;
1157 if (cache->vaddr & CLFLUSH_AFTER)
1160 kunmap_atomic(vaddr);
1161 i915_gem_object_finish_access(obj);
1163 struct i915_ggtt *ggtt = cache_to_ggtt(cache);
1165 intel_gt_flush_ggtt_writes(ggtt->vm.gt);
1166 io_mapping_unmap_atomic((void __iomem *)vaddr);
1168 if (drm_mm_node_allocated(&cache->node)) {
1169 ggtt->vm.clear_range(&ggtt->vm,
1172 mutex_lock(&ggtt->vm.mutex);
1173 drm_mm_remove_node(&cache->node);
1174 mutex_unlock(&ggtt->vm.mutex);
1176 i915_vma_unpin((struct i915_vma *)cache->node.mm);
1184 static void *reloc_kmap(struct drm_i915_gem_object *obj,
1185 struct reloc_cache *cache,
1186 unsigned long pageno)
1192 kunmap_atomic(unmask_page(cache->vaddr));
1194 unsigned int flushes;
1197 err = i915_gem_object_prepare_write(obj, &flushes);
1199 return ERR_PTR(err);
1201 BUILD_BUG_ON(KMAP & CLFLUSH_FLAGS);
1202 BUILD_BUG_ON((KMAP | CLFLUSH_FLAGS) & PAGE_MASK);
1204 cache->vaddr = flushes | KMAP;
1205 cache->node.mm = (void *)obj;
1210 page = i915_gem_object_get_page(obj, pageno);
1212 set_page_dirty(page);
1214 vaddr = kmap_atomic(page);
1215 cache->vaddr = unmask_flags(cache->vaddr) | (unsigned long)vaddr;
1216 cache->page = pageno;
1221 static void *reloc_iomap(struct i915_vma *batch,
1222 struct i915_execbuffer *eb,
1225 struct drm_i915_gem_object *obj = batch->obj;
1226 struct reloc_cache *cache = &eb->reloc_cache;
1227 struct i915_ggtt *ggtt = cache_to_ggtt(cache);
1228 unsigned long offset;
1232 intel_gt_flush_ggtt_writes(ggtt->vm.gt);
1233 io_mapping_unmap_atomic((void __force __iomem *) unmask_page(cache->vaddr));
1235 struct i915_vma *vma = ERR_PTR(-ENODEV);
1238 if (i915_gem_object_is_tiled(obj))
1239 return ERR_PTR(-EINVAL);
1241 if (use_cpu_reloc(cache, obj))
1244 err = i915_gem_object_set_to_gtt_domain(obj, true);
1246 return ERR_PTR(err);
1249 * i915_gem_object_ggtt_pin_ww may attempt to remove the batch
1250 * VMA from the object list because we no longer pin.
1252 * Only attempt to pin the batch buffer to ggtt if the current batch
1253 * is not inside ggtt, or the batch buffer is not misplaced.
1255 if (!i915_is_ggtt(batch->vm)) {
1256 vma = i915_gem_object_ggtt_pin_ww(obj, &eb->ww, NULL, 0, 0,
1258 PIN_NONBLOCK /* NOWARN */ |
1260 } else if (i915_vma_is_map_and_fenceable(batch)) {
1261 __i915_vma_pin(batch);
1265 if (vma == ERR_PTR(-EDEADLK))
1269 memset(&cache->node, 0, sizeof(cache->node));
1270 mutex_lock(&ggtt->vm.mutex);
1271 err = drm_mm_insert_node_in_range
1272 (&ggtt->vm.mm, &cache->node,
1273 PAGE_SIZE, 0, I915_COLOR_UNEVICTABLE,
1274 0, ggtt->mappable_end,
1276 mutex_unlock(&ggtt->vm.mutex);
1277 if (err) /* no inactive aperture space, use cpu reloc */
1280 cache->node.start = vma->node.start;
1281 cache->node.mm = (void *)vma;
1285 offset = cache->node.start;
1286 if (drm_mm_node_allocated(&cache->node)) {
1287 ggtt->vm.insert_page(&ggtt->vm,
1288 i915_gem_object_get_dma_address(obj, page),
1289 offset, I915_CACHE_NONE, 0);
1291 offset += page << PAGE_SHIFT;
1294 vaddr = (void __force *)io_mapping_map_atomic_wc(&ggtt->iomap,
1297 cache->vaddr = (unsigned long)vaddr;
1302 static void *reloc_vaddr(struct i915_vma *vma,
1303 struct i915_execbuffer *eb,
1306 struct reloc_cache *cache = &eb->reloc_cache;
1309 if (cache->page == page) {
1310 vaddr = unmask_page(cache->vaddr);
1313 if ((cache->vaddr & KMAP) == 0)
1314 vaddr = reloc_iomap(vma, eb, page);
1316 vaddr = reloc_kmap(vma->obj, cache, page);
1322 static void clflush_write32(u32 *addr, u32 value, unsigned int flushes)
1324 if (unlikely(flushes & (CLFLUSH_BEFORE | CLFLUSH_AFTER))) {
1325 if (flushes & CLFLUSH_BEFORE) {
1333 * Writes to the same cacheline are serialised by the CPU
1334 * (including clflush). On the write path, we only require
1335 * that it hits memory in an orderly fashion and place
1336 * mb barriers at the start and end of the relocation phase
1337 * to ensure ordering of clflush wrt to the system.
1339 if (flushes & CLFLUSH_AFTER)
1346 relocate_entry(struct i915_vma *vma,
1347 const struct drm_i915_gem_relocation_entry *reloc,
1348 struct i915_execbuffer *eb,
1349 const struct i915_vma *target)
1351 u64 target_addr = relocation_target(reloc, target);
1352 u64 offset = reloc->offset;
1353 bool wide = eb->reloc_cache.use_64bit_reloc;
1357 vaddr = reloc_vaddr(vma, eb,
1358 offset >> PAGE_SHIFT);
1360 return PTR_ERR(vaddr);
1362 GEM_BUG_ON(!IS_ALIGNED(offset, sizeof(u32)));
1363 clflush_write32(vaddr + offset_in_page(offset),
1364 lower_32_bits(target_addr),
1365 eb->reloc_cache.vaddr);
1368 offset += sizeof(u32);
1374 return target->node.start | UPDATE;
1378 eb_relocate_entry(struct i915_execbuffer *eb,
1380 const struct drm_i915_gem_relocation_entry *reloc)
1382 struct drm_i915_private *i915 = eb->i915;
1383 struct eb_vma *target;
1386 /* we've already hold a reference to all valid objects */
1387 target = eb_get_vma(eb, reloc->target_handle);
1388 if (unlikely(!target))
1391 /* Validate that the target is in a valid r/w GPU domain */
1392 if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
1393 drm_dbg(&i915->drm, "reloc with multiple write domains: "
1394 "target %d offset %d "
1395 "read %08x write %08x",
1396 reloc->target_handle,
1397 (int) reloc->offset,
1398 reloc->read_domains,
1399 reloc->write_domain);
1402 if (unlikely((reloc->write_domain | reloc->read_domains)
1403 & ~I915_GEM_GPU_DOMAINS)) {
1404 drm_dbg(&i915->drm, "reloc with read/write non-GPU domains: "
1405 "target %d offset %d "
1406 "read %08x write %08x",
1407 reloc->target_handle,
1408 (int) reloc->offset,
1409 reloc->read_domains,
1410 reloc->write_domain);
1414 if (reloc->write_domain) {
1415 target->flags |= EXEC_OBJECT_WRITE;
1418 * Sandybridge PPGTT errata: We need a global gtt mapping
1419 * for MI and pipe_control writes because the gpu doesn't
1420 * properly redirect them through the ppgtt for non_secure
1423 if (reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
1424 GRAPHICS_VER(eb->i915) == 6 &&
1425 !i915_vma_is_bound(target->vma, I915_VMA_GLOBAL_BIND)) {
1426 struct i915_vma *vma = target->vma;
1428 reloc_cache_unmap(&eb->reloc_cache);
1429 mutex_lock(&vma->vm->mutex);
1430 err = i915_vma_bind(target->vma,
1431 target->vma->obj->cache_level,
1432 PIN_GLOBAL, NULL, NULL);
1433 mutex_unlock(&vma->vm->mutex);
1434 reloc_cache_remap(&eb->reloc_cache, ev->vma->obj);
1441 * If the relocation already has the right value in it, no
1442 * more work needs to be done.
1444 if (!DBG_FORCE_RELOC &&
1445 gen8_canonical_addr(target->vma->node.start) == reloc->presumed_offset)
1448 /* Check that the relocation address is valid... */
1449 if (unlikely(reloc->offset >
1450 ev->vma->size - (eb->reloc_cache.use_64bit_reloc ? 8 : 4))) {
1451 drm_dbg(&i915->drm, "Relocation beyond object bounds: "
1452 "target %d offset %d size %d.\n",
1453 reloc->target_handle,
1455 (int)ev->vma->size);
1458 if (unlikely(reloc->offset & 3)) {
1459 drm_dbg(&i915->drm, "Relocation not 4-byte aligned: "
1460 "target %d offset %d.\n",
1461 reloc->target_handle,
1462 (int)reloc->offset);
1467 * If we write into the object, we need to force the synchronisation
1468 * barrier, either with an asynchronous clflush or if we executed the
1469 * patching using the GPU (though that should be serialised by the
1470 * timeline). To be completely sure, and since we are required to
1471 * do relocations we are already stalling, disable the user's opt
1472 * out of our synchronisation.
1474 ev->flags &= ~EXEC_OBJECT_ASYNC;
1476 /* and update the user's relocation entry */
1477 return relocate_entry(ev->vma, reloc, eb, target->vma);
1480 static int eb_relocate_vma(struct i915_execbuffer *eb, struct eb_vma *ev)
1482 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
1483 struct drm_i915_gem_relocation_entry stack[N_RELOC(512)];
1484 const struct drm_i915_gem_exec_object2 *entry = ev->exec;
1485 struct drm_i915_gem_relocation_entry __user *urelocs =
1486 u64_to_user_ptr(entry->relocs_ptr);
1487 unsigned long remain = entry->relocation_count;
1489 if (unlikely(remain > N_RELOC(ULONG_MAX)))
1493 * We must check that the entire relocation array is safe
1494 * to read. However, if the array is not writable the user loses
1495 * the updated relocation values.
1497 if (unlikely(!access_ok(urelocs, remain * sizeof(*urelocs))))
1501 struct drm_i915_gem_relocation_entry *r = stack;
1502 unsigned int count =
1503 min_t(unsigned long, remain, ARRAY_SIZE(stack));
1504 unsigned int copied;
1507 * This is the fast path and we cannot handle a pagefault
1508 * whilst holding the struct mutex lest the user pass in the
1509 * relocations contained within a mmaped bo. For in such a case
1510 * we, the page fault handler would call i915_gem_fault() and
1511 * we would try to acquire the struct mutex again. Obviously
1512 * this is bad and so lockdep complains vehemently.
1514 pagefault_disable();
1515 copied = __copy_from_user_inatomic(r, urelocs, count * sizeof(r[0]));
1517 if (unlikely(copied)) {
1524 u64 offset = eb_relocate_entry(eb, ev, r);
1526 if (likely(offset == 0)) {
1527 } else if ((s64)offset < 0) {
1528 remain = (int)offset;
1532 * Note that reporting an error now
1533 * leaves everything in an inconsistent
1534 * state as we have *already* changed
1535 * the relocation value inside the
1536 * object. As we have not changed the
1537 * reloc.presumed_offset or will not
1538 * change the execobject.offset, on the
1539 * call we may not rewrite the value
1540 * inside the object, leaving it
1541 * dangling and causing a GPU hang. Unless
1542 * userspace dynamically rebuilds the
1543 * relocations on each execbuf rather than
1544 * presume a static tree.
1546 * We did previously check if the relocations
1547 * were writable (access_ok), an error now
1548 * would be a strange race with mprotect,
1549 * having already demonstrated that we
1550 * can read from this userspace address.
1552 offset = gen8_canonical_addr(offset & ~UPDATE);
1554 &urelocs[r - stack].presumed_offset);
1556 } while (r++, --count);
1557 urelocs += ARRAY_SIZE(stack);
1560 reloc_cache_reset(&eb->reloc_cache, eb);
1565 eb_relocate_vma_slow(struct i915_execbuffer *eb, struct eb_vma *ev)
1567 const struct drm_i915_gem_exec_object2 *entry = ev->exec;
1568 struct drm_i915_gem_relocation_entry *relocs =
1569 u64_to_ptr(typeof(*relocs), entry->relocs_ptr);
1573 for (i = 0; i < entry->relocation_count; i++) {
1574 u64 offset = eb_relocate_entry(eb, ev, &relocs[i]);
1576 if ((s64)offset < 0) {
1583 reloc_cache_reset(&eb->reloc_cache, eb);
1587 static int check_relocations(const struct drm_i915_gem_exec_object2 *entry)
1589 const char __user *addr, *end;
1591 char __maybe_unused c;
1593 size = entry->relocation_count;
1597 if (size > N_RELOC(ULONG_MAX))
1600 addr = u64_to_user_ptr(entry->relocs_ptr);
1601 size *= sizeof(struct drm_i915_gem_relocation_entry);
1602 if (!access_ok(addr, size))
1606 for (; addr < end; addr += PAGE_SIZE) {
1607 int err = __get_user(c, addr);
1611 return __get_user(c, end - 1);
1614 static int eb_copy_relocations(const struct i915_execbuffer *eb)
1616 struct drm_i915_gem_relocation_entry *relocs;
1617 const unsigned int count = eb->buffer_count;
1621 for (i = 0; i < count; i++) {
1622 const unsigned int nreloc = eb->exec[i].relocation_count;
1623 struct drm_i915_gem_relocation_entry __user *urelocs;
1625 unsigned long copied;
1630 err = check_relocations(&eb->exec[i]);
1634 urelocs = u64_to_user_ptr(eb->exec[i].relocs_ptr);
1635 size = nreloc * sizeof(*relocs);
1637 relocs = kvmalloc_array(size, 1, GFP_KERNEL);
1643 /* copy_from_user is limited to < 4GiB */
1647 min_t(u64, BIT_ULL(31), size - copied);
1649 if (__copy_from_user((char *)relocs + copied,
1650 (char __user *)urelocs + copied,
1655 } while (copied < size);
1658 * As we do not update the known relocation offsets after
1659 * relocating (due to the complexities in lock handling),
1660 * we need to mark them as invalid now so that we force the
1661 * relocation processing next time. Just in case the target
1662 * object is evicted and then rebound into its old
1663 * presumed_offset before the next execbuffer - if that
1664 * happened we would make the mistake of assuming that the
1665 * relocations were valid.
1667 if (!user_access_begin(urelocs, size))
1670 for (copied = 0; copied < nreloc; copied++)
1672 &urelocs[copied].presumed_offset,
1676 eb->exec[i].relocs_ptr = (uintptr_t)relocs;
1688 relocs = u64_to_ptr(typeof(*relocs), eb->exec[i].relocs_ptr);
1689 if (eb->exec[i].relocation_count)
1695 static int eb_prefault_relocations(const struct i915_execbuffer *eb)
1697 const unsigned int count = eb->buffer_count;
1700 for (i = 0; i < count; i++) {
1703 err = check_relocations(&eb->exec[i]);
1711 static int eb_reinit_userptr(struct i915_execbuffer *eb)
1713 const unsigned int count = eb->buffer_count;
1717 if (likely(!(eb->args->flags & __EXEC_USERPTR_USED)))
1720 for (i = 0; i < count; i++) {
1721 struct eb_vma *ev = &eb->vma[i];
1723 if (!i915_gem_object_is_userptr(ev->vma->obj))
1726 ret = i915_gem_object_userptr_submit_init(ev->vma->obj);
1730 ev->flags |= __EXEC_OBJECT_USERPTR_INIT;
1736 static noinline int eb_relocate_parse_slow(struct i915_execbuffer *eb)
1738 bool have_copy = false;
1743 if (signal_pending(current)) {
1748 /* We may process another execbuffer during the unlock... */
1749 eb_release_vmas(eb, false);
1750 i915_gem_ww_ctx_fini(&eb->ww);
1753 * We take 3 passes through the slowpatch.
1755 * 1 - we try to just prefault all the user relocation entries and
1756 * then attempt to reuse the atomic pagefault disabled fast path again.
1758 * 2 - we copy the user entries to a local buffer here outside of the
1759 * local and allow ourselves to wait upon any rendering before
1762 * 3 - we already have a local copy of the relocation entries, but
1763 * were interrupted (EAGAIN) whilst waiting for the objects, try again.
1766 err = eb_prefault_relocations(eb);
1767 } else if (!have_copy) {
1768 err = eb_copy_relocations(eb);
1769 have_copy = err == 0;
1776 err = eb_reinit_userptr(eb);
1778 i915_gem_ww_ctx_init(&eb->ww, true);
1782 /* reacquire the objects */
1784 err = eb_pin_engine(eb, false);
1788 err = eb_validate_vmas(eb);
1792 GEM_BUG_ON(!eb->batches[0]);
1794 list_for_each_entry(ev, &eb->relocs, reloc_link) {
1796 err = eb_relocate_vma(eb, ev);
1800 err = eb_relocate_vma_slow(eb, ev);
1806 if (err == -EDEADLK)
1809 if (err && !have_copy)
1815 /* as last step, parse the command buffer */
1821 * Leave the user relocations as are, this is the painfully slow path,
1822 * and we want to avoid the complication of dropping the lock whilst
1823 * having buffers reserved in the aperture and so causing spurious
1824 * ENOSPC for random operations.
1828 if (err == -EDEADLK) {
1829 eb_release_vmas(eb, false);
1830 err = i915_gem_ww_ctx_backoff(&eb->ww);
1832 goto repeat_validate;
1840 const unsigned int count = eb->buffer_count;
1843 for (i = 0; i < count; i++) {
1844 const struct drm_i915_gem_exec_object2 *entry =
1846 struct drm_i915_gem_relocation_entry *relocs;
1848 if (!entry->relocation_count)
1851 relocs = u64_to_ptr(typeof(*relocs), entry->relocs_ptr);
1859 static int eb_relocate_parse(struct i915_execbuffer *eb)
1862 bool throttle = true;
1865 err = eb_pin_engine(eb, throttle);
1867 if (err != -EDEADLK)
1873 /* only throttle once, even if we didn't need to throttle */
1876 err = eb_validate_vmas(eb);
1882 /* The objects are in their final locations, apply the relocations. */
1883 if (eb->args->flags & __EXEC_HAS_RELOC) {
1886 list_for_each_entry(ev, &eb->relocs, reloc_link) {
1887 err = eb_relocate_vma(eb, ev);
1892 if (err == -EDEADLK)
1902 if (err == -EDEADLK) {
1903 eb_release_vmas(eb, false);
1904 err = i915_gem_ww_ctx_backoff(&eb->ww);
1912 err = eb_relocate_parse_slow(eb);
1915 * If the user expects the execobject.offset and
1916 * reloc.presumed_offset to be an exact match,
1917 * as for using NO_RELOC, then we cannot update
1918 * the execobject.offset until we have completed
1921 eb->args->flags &= ~__EXEC_HAS_RELOC;
1927 * Using two helper loops for the order of which requests / batches are created
1928 * and added the to backend. Requests are created in order from the parent to
1929 * the last child. Requests are added in the reverse order, from the last child
1930 * to parent. This is done for locking reasons as the timeline lock is acquired
1931 * during request creation and released when the request is added to the
1932 * backend. To make lockdep happy (see intel_context_timeline_lock) this must be
1935 #define for_each_batch_create_order(_eb, _i) \
1936 for ((_i) = 0; (_i) < (_eb)->num_batches; ++(_i))
1937 #define for_each_batch_add_order(_eb, _i) \
1938 BUILD_BUG_ON(!typecheck(int, _i)); \
1939 for ((_i) = (_eb)->num_batches - 1; (_i) >= 0; --(_i))
1941 static struct i915_request *
1942 eb_find_first_request_added(struct i915_execbuffer *eb)
1946 for_each_batch_add_order(eb, i)
1947 if (eb->requests[i])
1948 return eb->requests[i];
1950 GEM_BUG_ON("Request not found");
1955 #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR)
1957 /* Stage with GFP_KERNEL allocations before we enter the signaling critical path */
1958 static void eb_capture_stage(struct i915_execbuffer *eb)
1960 const unsigned int count = eb->buffer_count;
1961 unsigned int i = count, j;
1964 struct eb_vma *ev = &eb->vma[i];
1965 struct i915_vma *vma = ev->vma;
1966 unsigned int flags = ev->flags;
1968 if (!(flags & EXEC_OBJECT_CAPTURE))
1971 for_each_batch_create_order(eb, j) {
1972 struct i915_capture_list *capture;
1974 capture = kmalloc(sizeof(*capture), GFP_KERNEL);
1978 capture->next = eb->capture_lists[j];
1979 capture->vma_res = i915_vma_resource_get(vma->resource);
1980 eb->capture_lists[j] = capture;
1985 /* Commit once we're in the critical path */
1986 static void eb_capture_commit(struct i915_execbuffer *eb)
1990 for_each_batch_create_order(eb, j) {
1991 struct i915_request *rq = eb->requests[j];
1996 rq->capture_list = eb->capture_lists[j];
1997 eb->capture_lists[j] = NULL;
2002 * Release anything that didn't get committed due to errors.
2003 * The capture_list will otherwise be freed at request retire.
2005 static void eb_capture_release(struct i915_execbuffer *eb)
2009 for_each_batch_create_order(eb, j) {
2010 if (eb->capture_lists[j]) {
2011 i915_request_free_capture_list(eb->capture_lists[j]);
2012 eb->capture_lists[j] = NULL;
2017 static void eb_capture_list_clear(struct i915_execbuffer *eb)
2019 memset(eb->capture_lists, 0, sizeof(eb->capture_lists));
2024 static void eb_capture_stage(struct i915_execbuffer *eb)
2028 static void eb_capture_commit(struct i915_execbuffer *eb)
2032 static void eb_capture_release(struct i915_execbuffer *eb)
2036 static void eb_capture_list_clear(struct i915_execbuffer *eb)
2042 static int eb_move_to_gpu(struct i915_execbuffer *eb)
2044 const unsigned int count = eb->buffer_count;
2045 unsigned int i = count;
2049 struct eb_vma *ev = &eb->vma[i];
2050 struct i915_vma *vma = ev->vma;
2051 unsigned int flags = ev->flags;
2052 struct drm_i915_gem_object *obj = vma->obj;
2054 assert_vma_held(vma);
2057 * If the GPU is not _reading_ through the CPU cache, we need
2058 * to make sure that any writes (both previous GPU writes from
2059 * before a change in snooping levels and normal CPU writes)
2060 * caught in that cache are flushed to main memory.
2063 * obj->cache_dirty &&
2064 * !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ)
2065 * but gcc's optimiser doesn't handle that as well and emits
2066 * two jumps instead of one. Maybe one day...
2068 * FIXME: There is also sync flushing in set_pages(), which
2069 * serves a different purpose(some of the time at least).
2071 * We should consider:
2073 * 1. Rip out the async flush code.
2075 * 2. Or make the sync flushing use the async clflush path
2076 * using mandatory fences underneath. Currently the below
2077 * async flush happens after we bind the object.
2079 if (unlikely(obj->cache_dirty & ~obj->cache_coherent)) {
2080 if (i915_gem_clflush_object(obj, 0))
2081 flags &= ~EXEC_OBJECT_ASYNC;
2084 /* We only need to await on the first request */
2085 if (err == 0 && !(flags & EXEC_OBJECT_ASYNC)) {
2086 err = i915_request_await_object
2087 (eb_find_first_request_added(eb), obj,
2088 flags & EXEC_OBJECT_WRITE);
2091 for_each_batch_add_order(eb, j) {
2094 if (!eb->requests[j])
2097 err = _i915_vma_move_to_active(vma, eb->requests[j],
2099 eb->composite_fence ?
2100 eb->composite_fence :
2101 &eb->requests[j]->fence,
2102 flags | __EXEC_OBJECT_NO_RESERVE);
2106 #ifdef CONFIG_MMU_NOTIFIER
2107 if (!err && (eb->args->flags & __EXEC_USERPTR_USED)) {
2108 read_lock(&eb->i915->mm.notifier_lock);
2111 * count is always at least 1, otherwise __EXEC_USERPTR_USED
2112 * could not have been set
2114 for (i = 0; i < count; i++) {
2115 struct eb_vma *ev = &eb->vma[i];
2116 struct drm_i915_gem_object *obj = ev->vma->obj;
2118 if (!i915_gem_object_is_userptr(obj))
2121 err = i915_gem_object_userptr_submit_done(obj);
2126 read_unlock(&eb->i915->mm.notifier_lock);
2133 /* Unconditionally flush any chipset caches (for streaming writes). */
2134 intel_gt_chipset_flush(eb->gt);
2135 eb_capture_commit(eb);
2140 for_each_batch_create_order(eb, j) {
2141 if (!eb->requests[j])
2144 i915_request_set_error_once(eb->requests[j], err);
2149 static int i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
2151 if (exec->flags & __I915_EXEC_ILLEGAL_FLAGS)
2154 /* Kernel clipping was a DRI1 misfeature */
2155 if (!(exec->flags & (I915_EXEC_FENCE_ARRAY |
2156 I915_EXEC_USE_EXTENSIONS))) {
2157 if (exec->num_cliprects || exec->cliprects_ptr)
2161 if (exec->DR4 == 0xffffffff) {
2162 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
2165 if (exec->DR1 || exec->DR4)
2168 if ((exec->batch_start_offset | exec->batch_len) & 0x7)
2174 static int i915_reset_gen7_sol_offsets(struct i915_request *rq)
2179 if (GRAPHICS_VER(rq->engine->i915) != 7 || rq->engine->id != RCS0) {
2180 drm_dbg(&rq->engine->i915->drm, "sol reset is gen7/rcs only\n");
2184 cs = intel_ring_begin(rq, 4 * 2 + 2);
2188 *cs++ = MI_LOAD_REGISTER_IMM(4);
2189 for (i = 0; i < 4; i++) {
2190 *cs++ = i915_mmio_reg_offset(GEN7_SO_WRITE_OFFSET(i));
2194 intel_ring_advance(rq, cs);
2199 static struct i915_vma *
2200 shadow_batch_pin(struct i915_execbuffer *eb,
2201 struct drm_i915_gem_object *obj,
2202 struct i915_address_space *vm,
2205 struct i915_vma *vma;
2208 vma = i915_vma_instance(obj, vm, NULL);
2212 err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, flags | PIN_VALIDATE);
2214 return ERR_PTR(err);
2219 static struct i915_vma *eb_dispatch_secure(struct i915_execbuffer *eb, struct i915_vma *vma)
2222 * snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
2223 * batch" bit. Hence we need to pin secure batches into the global gtt.
2224 * hsw should have this fixed, but bdw mucks it up again. */
2225 if (eb->batch_flags & I915_DISPATCH_SECURE)
2226 return i915_gem_object_ggtt_pin_ww(vma->obj, &eb->ww, NULL, 0, 0, PIN_VALIDATE);
2231 static int eb_parse(struct i915_execbuffer *eb)
2233 struct drm_i915_private *i915 = eb->i915;
2234 struct intel_gt_buffer_pool_node *pool = eb->batch_pool;
2235 struct i915_vma *shadow, *trampoline, *batch;
2239 if (!eb_use_cmdparser(eb)) {
2240 batch = eb_dispatch_secure(eb, eb->batches[0]->vma);
2242 return PTR_ERR(batch);
2247 if (intel_context_is_parallel(eb->context))
2250 len = eb->batch_len[0];
2251 if (!CMDPARSER_USES_GGTT(eb->i915)) {
2253 * ppGTT backed shadow buffers must be mapped RO, to prevent
2254 * post-scan tampering
2256 if (!eb->context->vm->has_read_only) {
2258 "Cannot prevent post-scan tampering without RO capable vm\n");
2262 len += I915_CMD_PARSER_TRAMPOLINE_SIZE;
2264 if (unlikely(len < eb->batch_len[0])) /* last paranoid check of overflow */
2268 pool = intel_gt_get_buffer_pool(eb->gt, len,
2271 return PTR_ERR(pool);
2272 eb->batch_pool = pool;
2275 err = i915_gem_object_lock(pool->obj, &eb->ww);
2279 shadow = shadow_batch_pin(eb, pool->obj, eb->context->vm, PIN_USER);
2281 return PTR_ERR(shadow);
2283 intel_gt_buffer_pool_mark_used(pool);
2284 i915_gem_object_set_readonly(shadow->obj);
2285 shadow->private = pool;
2288 if (CMDPARSER_USES_GGTT(eb->i915)) {
2289 trampoline = shadow;
2291 shadow = shadow_batch_pin(eb, pool->obj,
2295 return PTR_ERR(shadow);
2297 shadow->private = pool;
2299 eb->batch_flags |= I915_DISPATCH_SECURE;
2302 batch = eb_dispatch_secure(eb, shadow);
2304 return PTR_ERR(batch);
2306 err = dma_resv_reserve_shared(shadow->obj->base.resv, 1);
2310 err = intel_engine_cmd_parser(eb->context->engine,
2311 eb->batches[0]->vma,
2312 eb->batch_start_offset,
2314 shadow, trampoline);
2318 eb->batches[0] = &eb->vma[eb->buffer_count++];
2319 eb->batches[0]->vma = i915_vma_get(shadow);
2320 eb->batches[0]->flags = __EXEC_OBJECT_HAS_PIN;
2322 eb->trampoline = trampoline;
2323 eb->batch_start_offset = 0;
2327 if (intel_context_is_parallel(eb->context))
2330 eb->batches[0] = &eb->vma[eb->buffer_count++];
2331 eb->batches[0]->flags = __EXEC_OBJECT_HAS_PIN;
2332 eb->batches[0]->vma = i915_vma_get(batch);
2337 static int eb_request_submit(struct i915_execbuffer *eb,
2338 struct i915_request *rq,
2339 struct i915_vma *batch,
2344 if (intel_context_nopreempt(rq->context))
2345 __set_bit(I915_FENCE_FLAG_NOPREEMPT, &rq->fence.flags);
2347 if (eb->args->flags & I915_EXEC_GEN7_SOL_RESET) {
2348 err = i915_reset_gen7_sol_offsets(rq);
2354 * After we completed waiting for other engines (using HW semaphores)
2355 * then we can signal that this request/batch is ready to run. This
2356 * allows us to determine if the batch is still waiting on the GPU
2357 * or actually running by checking the breadcrumb.
2359 if (rq->context->engine->emit_init_breadcrumb) {
2360 err = rq->context->engine->emit_init_breadcrumb(rq);
2365 err = rq->context->engine->emit_bb_start(rq,
2367 eb->batch_start_offset,
2373 if (eb->trampoline) {
2374 GEM_BUG_ON(intel_context_is_parallel(rq->context));
2375 GEM_BUG_ON(eb->batch_start_offset);
2376 err = rq->context->engine->emit_bb_start(rq,
2377 eb->trampoline->node.start +
2386 static int eb_submit(struct i915_execbuffer *eb)
2391 err = eb_move_to_gpu(eb);
2393 for_each_batch_create_order(eb, i) {
2394 if (!eb->requests[i])
2397 trace_i915_request_queue(eb->requests[i], eb->batch_flags);
2399 err = eb_request_submit(eb, eb->requests[i],
2400 eb->batches[i]->vma,
2407 static int num_vcs_engines(struct drm_i915_private *i915)
2409 return hweight_long(VDBOX_MASK(to_gt(i915)));
2413 * Find one BSD ring to dispatch the corresponding BSD command.
2414 * The engine index is returned.
2417 gen8_dispatch_bsd_engine(struct drm_i915_private *dev_priv,
2418 struct drm_file *file)
2420 struct drm_i915_file_private *file_priv = file->driver_priv;
2422 /* Check whether the file_priv has already selected one ring. */
2423 if ((int)file_priv->bsd_engine < 0)
2424 file_priv->bsd_engine =
2425 get_random_int() % num_vcs_engines(dev_priv);
2427 return file_priv->bsd_engine;
2430 static const enum intel_engine_id user_ring_map[] = {
2431 [I915_EXEC_DEFAULT] = RCS0,
2432 [I915_EXEC_RENDER] = RCS0,
2433 [I915_EXEC_BLT] = BCS0,
2434 [I915_EXEC_BSD] = VCS0,
2435 [I915_EXEC_VEBOX] = VECS0
2438 static struct i915_request *eb_throttle(struct i915_execbuffer *eb, struct intel_context *ce)
2440 struct intel_ring *ring = ce->ring;
2441 struct intel_timeline *tl = ce->timeline;
2442 struct i915_request *rq;
2445 * Completely unscientific finger-in-the-air estimates for suitable
2446 * maximum user request size (to avoid blocking) and then backoff.
2448 if (intel_ring_update_space(ring) >= PAGE_SIZE)
2452 * Find a request that after waiting upon, there will be at least half
2453 * the ring available. The hysteresis allows us to compete for the
2454 * shared ring and should mean that we sleep less often prior to
2455 * claiming our resources, but not so long that the ring completely
2456 * drains before we can submit our next request.
2458 list_for_each_entry(rq, &tl->requests, link) {
2459 if (rq->ring != ring)
2462 if (__intel_ring_space(rq->postfix,
2463 ring->emit, ring->size) > ring->size / 2)
2466 if (&rq->link == &tl->requests)
2467 return NULL; /* weird, we will check again later for real */
2469 return i915_request_get(rq);
2472 static int eb_pin_timeline(struct i915_execbuffer *eb, struct intel_context *ce,
2475 struct intel_timeline *tl;
2476 struct i915_request *rq = NULL;
2479 * Take a local wakeref for preparing to dispatch the execbuf as
2480 * we expect to access the hardware fairly frequently in the
2481 * process, and require the engine to be kept awake between accesses.
2482 * Upon dispatch, we acquire another prolonged wakeref that we hold
2483 * until the timeline is idle, which in turn releases the wakeref
2484 * taken on the engine, and the parent device.
2486 tl = intel_context_timeline_lock(ce);
2490 intel_context_enter(ce);
2492 rq = eb_throttle(eb, ce);
2493 intel_context_timeline_unlock(tl);
2496 bool nonblock = eb->file->filp->f_flags & O_NONBLOCK;
2497 long timeout = nonblock ? 0 : MAX_SCHEDULE_TIMEOUT;
2499 if (i915_request_wait(rq, I915_WAIT_INTERRUPTIBLE,
2501 i915_request_put(rq);
2504 * Error path, cannot use intel_context_timeline_lock as
2505 * that is user interruptable and this clean up step
2508 mutex_lock(&ce->timeline->mutex);
2509 intel_context_exit(ce);
2510 mutex_unlock(&ce->timeline->mutex);
2513 return -EWOULDBLOCK;
2517 i915_request_put(rq);
2523 static int eb_pin_engine(struct i915_execbuffer *eb, bool throttle)
2525 struct intel_context *ce = eb->context, *child;
2529 GEM_BUG_ON(eb->args->flags & __EXEC_ENGINE_PINNED);
2531 if (unlikely(intel_context_is_banned(ce)))
2535 * Pinning the contexts may generate requests in order to acquire
2536 * GGTT space, so do this first before we reserve a seqno for
2539 err = intel_context_pin_ww(ce, &eb->ww);
2542 for_each_child(ce, child) {
2543 err = intel_context_pin_ww(child, &eb->ww);
2544 GEM_BUG_ON(err); /* perma-pinned should incr a counter */
2547 for_each_child(ce, child) {
2548 err = eb_pin_timeline(eb, child, throttle);
2553 err = eb_pin_timeline(eb, ce, throttle);
2557 eb->args->flags |= __EXEC_ENGINE_PINNED;
2561 for_each_child(ce, child) {
2563 mutex_lock(&child->timeline->mutex);
2564 intel_context_exit(child);
2565 mutex_unlock(&child->timeline->mutex);
2568 for_each_child(ce, child)
2569 intel_context_unpin(child);
2570 intel_context_unpin(ce);
2574 static void eb_unpin_engine(struct i915_execbuffer *eb)
2576 struct intel_context *ce = eb->context, *child;
2578 if (!(eb->args->flags & __EXEC_ENGINE_PINNED))
2581 eb->args->flags &= ~__EXEC_ENGINE_PINNED;
2583 for_each_child(ce, child) {
2584 mutex_lock(&child->timeline->mutex);
2585 intel_context_exit(child);
2586 mutex_unlock(&child->timeline->mutex);
2588 intel_context_unpin(child);
2591 mutex_lock(&ce->timeline->mutex);
2592 intel_context_exit(ce);
2593 mutex_unlock(&ce->timeline->mutex);
2595 intel_context_unpin(ce);
2599 eb_select_legacy_ring(struct i915_execbuffer *eb)
2601 struct drm_i915_private *i915 = eb->i915;
2602 struct drm_i915_gem_execbuffer2 *args = eb->args;
2603 unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK;
2605 if (user_ring_id != I915_EXEC_BSD &&
2606 (args->flags & I915_EXEC_BSD_MASK)) {
2608 "execbuf with non bsd ring but with invalid "
2609 "bsd dispatch flags: %d\n", (int)(args->flags));
2613 if (user_ring_id == I915_EXEC_BSD && num_vcs_engines(i915) > 1) {
2614 unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK;
2616 if (bsd_idx == I915_EXEC_BSD_DEFAULT) {
2617 bsd_idx = gen8_dispatch_bsd_engine(i915, eb->file);
2618 } else if (bsd_idx >= I915_EXEC_BSD_RING1 &&
2619 bsd_idx <= I915_EXEC_BSD_RING2) {
2620 bsd_idx >>= I915_EXEC_BSD_SHIFT;
2624 "execbuf with unknown bsd ring: %u\n",
2629 return _VCS(bsd_idx);
2632 if (user_ring_id >= ARRAY_SIZE(user_ring_map)) {
2633 drm_dbg(&i915->drm, "execbuf with unknown ring: %u\n",
2638 return user_ring_map[user_ring_id];
2642 eb_select_engine(struct i915_execbuffer *eb)
2644 struct intel_context *ce, *child;
2648 if (i915_gem_context_user_engines(eb->gem_context))
2649 idx = eb->args->flags & I915_EXEC_RING_MASK;
2651 idx = eb_select_legacy_ring(eb);
2653 ce = i915_gem_context_get_engine(eb->gem_context, idx);
2657 if (intel_context_is_parallel(ce)) {
2658 if (eb->buffer_count < ce->parallel.number_children + 1) {
2659 intel_context_put(ce);
2662 if (eb->batch_start_offset || eb->args->batch_len) {
2663 intel_context_put(ce);
2667 eb->num_batches = ce->parallel.number_children + 1;
2669 for_each_child(ce, child)
2670 intel_context_get(child);
2671 intel_gt_pm_get(ce->engine->gt);
2673 if (!test_bit(CONTEXT_ALLOC_BIT, &ce->flags)) {
2674 err = intel_context_alloc_state(ce);
2678 for_each_child(ce, child) {
2679 if (!test_bit(CONTEXT_ALLOC_BIT, &child->flags)) {
2680 err = intel_context_alloc_state(child);
2687 * ABI: Before userspace accesses the GPU (e.g. execbuffer), report
2688 * EIO if the GPU is already wedged.
2690 err = intel_gt_terminally_wedged(ce->engine->gt);
2695 eb->gt = ce->engine->gt;
2698 * Make sure engine pool stays alive even if we call intel_context_put
2699 * during ww handling. The pool is destroyed when last pm reference
2700 * is dropped, which breaks our -EDEADLK handling.
2705 intel_gt_pm_put(ce->engine->gt);
2706 for_each_child(ce, child)
2707 intel_context_put(child);
2708 intel_context_put(ce);
2713 eb_put_engine(struct i915_execbuffer *eb)
2715 struct intel_context *child;
2717 intel_gt_pm_put(eb->gt);
2718 for_each_child(eb->context, child)
2719 intel_context_put(child);
2720 intel_context_put(eb->context);
2724 __free_fence_array(struct eb_fence *fences, unsigned int n)
2727 drm_syncobj_put(ptr_mask_bits(fences[n].syncobj, 2));
2728 dma_fence_put(fences[n].dma_fence);
2729 dma_fence_chain_free(fences[n].chain_fence);
2735 add_timeline_fence_array(struct i915_execbuffer *eb,
2736 const struct drm_i915_gem_execbuffer_ext_timeline_fences *timeline_fences)
2738 struct drm_i915_gem_exec_fence __user *user_fences;
2739 u64 __user *user_values;
2744 nfences = timeline_fences->fence_count;
2748 /* Check multiplication overflow for access_ok() and kvmalloc_array() */
2749 BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long));
2750 if (nfences > min_t(unsigned long,
2751 ULONG_MAX / sizeof(*user_fences),
2752 SIZE_MAX / sizeof(*f)) - eb->num_fences)
2755 user_fences = u64_to_user_ptr(timeline_fences->handles_ptr);
2756 if (!access_ok(user_fences, nfences * sizeof(*user_fences)))
2759 user_values = u64_to_user_ptr(timeline_fences->values_ptr);
2760 if (!access_ok(user_values, nfences * sizeof(*user_values)))
2763 f = krealloc(eb->fences,
2764 (eb->num_fences + nfences) * sizeof(*f),
2765 __GFP_NOWARN | GFP_KERNEL);
2770 f += eb->num_fences;
2772 BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) &
2773 ~__I915_EXEC_FENCE_UNKNOWN_FLAGS);
2776 struct drm_i915_gem_exec_fence user_fence;
2777 struct drm_syncobj *syncobj;
2778 struct dma_fence *fence = NULL;
2781 if (__copy_from_user(&user_fence,
2783 sizeof(user_fence)))
2786 if (user_fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS)
2789 if (__get_user(point, user_values++))
2792 syncobj = drm_syncobj_find(eb->file, user_fence.handle);
2794 DRM_DEBUG("Invalid syncobj handle provided\n");
2798 fence = drm_syncobj_fence_get(syncobj);
2800 if (!fence && user_fence.flags &&
2801 !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) {
2802 DRM_DEBUG("Syncobj handle has no fence\n");
2803 drm_syncobj_put(syncobj);
2808 err = dma_fence_chain_find_seqno(&fence, point);
2810 if (err && !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) {
2811 DRM_DEBUG("Syncobj handle missing requested point %llu\n", point);
2812 dma_fence_put(fence);
2813 drm_syncobj_put(syncobj);
2818 * A point might have been signaled already and
2819 * garbage collected from the timeline. In this case
2820 * just ignore the point and carry on.
2822 if (!fence && !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) {
2823 drm_syncobj_put(syncobj);
2828 * For timeline syncobjs we need to preallocate chains for
2831 if (point != 0 && user_fence.flags & I915_EXEC_FENCE_SIGNAL) {
2833 * Waiting and signaling the same point (when point !=
2834 * 0) would break the timeline.
2836 if (user_fence.flags & I915_EXEC_FENCE_WAIT) {
2837 DRM_DEBUG("Trying to wait & signal the same timeline point.\n");
2838 dma_fence_put(fence);
2839 drm_syncobj_put(syncobj);
2843 f->chain_fence = dma_fence_chain_alloc();
2844 if (!f->chain_fence) {
2845 drm_syncobj_put(syncobj);
2846 dma_fence_put(fence);
2850 f->chain_fence = NULL;
2853 f->syncobj = ptr_pack_bits(syncobj, user_fence.flags, 2);
2854 f->dma_fence = fence;
2863 static int add_fence_array(struct i915_execbuffer *eb)
2865 struct drm_i915_gem_execbuffer2 *args = eb->args;
2866 struct drm_i915_gem_exec_fence __user *user;
2867 unsigned long num_fences = args->num_cliprects;
2870 if (!(args->flags & I915_EXEC_FENCE_ARRAY))
2876 /* Check multiplication overflow for access_ok() and kvmalloc_array() */
2877 BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long));
2878 if (num_fences > min_t(unsigned long,
2879 ULONG_MAX / sizeof(*user),
2880 SIZE_MAX / sizeof(*f) - eb->num_fences))
2883 user = u64_to_user_ptr(args->cliprects_ptr);
2884 if (!access_ok(user, num_fences * sizeof(*user)))
2887 f = krealloc(eb->fences,
2888 (eb->num_fences + num_fences) * sizeof(*f),
2889 __GFP_NOWARN | GFP_KERNEL);
2894 f += eb->num_fences;
2895 while (num_fences--) {
2896 struct drm_i915_gem_exec_fence user_fence;
2897 struct drm_syncobj *syncobj;
2898 struct dma_fence *fence = NULL;
2900 if (__copy_from_user(&user_fence, user++, sizeof(user_fence)))
2903 if (user_fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS)
2906 syncobj = drm_syncobj_find(eb->file, user_fence.handle);
2908 DRM_DEBUG("Invalid syncobj handle provided\n");
2912 if (user_fence.flags & I915_EXEC_FENCE_WAIT) {
2913 fence = drm_syncobj_fence_get(syncobj);
2915 DRM_DEBUG("Syncobj handle has no fence\n");
2916 drm_syncobj_put(syncobj);
2921 BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) &
2922 ~__I915_EXEC_FENCE_UNKNOWN_FLAGS);
2924 f->syncobj = ptr_pack_bits(syncobj, user_fence.flags, 2);
2925 f->dma_fence = fence;
2927 f->chain_fence = NULL;
2935 static void put_fence_array(struct eb_fence *fences, int num_fences)
2938 __free_fence_array(fences, num_fences);
2942 await_fence_array(struct i915_execbuffer *eb,
2943 struct i915_request *rq)
2948 for (n = 0; n < eb->num_fences; n++) {
2949 struct drm_syncobj *syncobj;
2952 syncobj = ptr_unpack_bits(eb->fences[n].syncobj, &flags, 2);
2954 if (!eb->fences[n].dma_fence)
2957 err = i915_request_await_dma_fence(rq, eb->fences[n].dma_fence);
2965 static void signal_fence_array(const struct i915_execbuffer *eb,
2966 struct dma_fence * const fence)
2970 for (n = 0; n < eb->num_fences; n++) {
2971 struct drm_syncobj *syncobj;
2974 syncobj = ptr_unpack_bits(eb->fences[n].syncobj, &flags, 2);
2975 if (!(flags & I915_EXEC_FENCE_SIGNAL))
2978 if (eb->fences[n].chain_fence) {
2979 drm_syncobj_add_point(syncobj,
2980 eb->fences[n].chain_fence,
2982 eb->fences[n].value);
2984 * The chain's ownership is transferred to the
2987 eb->fences[n].chain_fence = NULL;
2989 drm_syncobj_replace_fence(syncobj, fence);
2995 parse_timeline_fences(struct i915_user_extension __user *ext, void *data)
2997 struct i915_execbuffer *eb = data;
2998 struct drm_i915_gem_execbuffer_ext_timeline_fences timeline_fences;
3000 if (copy_from_user(&timeline_fences, ext, sizeof(timeline_fences)))
3003 return add_timeline_fence_array(eb, &timeline_fences);
3006 static void retire_requests(struct intel_timeline *tl, struct i915_request *end)
3008 struct i915_request *rq, *rn;
3010 list_for_each_entry_safe(rq, rn, &tl->requests, link)
3011 if (rq == end || !i915_request_retire(rq))
3015 static int eb_request_add(struct i915_execbuffer *eb, struct i915_request *rq,
3016 int err, bool last_parallel)
3018 struct intel_timeline * const tl = i915_request_timeline(rq);
3019 struct i915_sched_attr attr = {};
3020 struct i915_request *prev;
3022 lockdep_assert_held(&tl->mutex);
3023 lockdep_unpin_lock(&tl->mutex, rq->cookie);
3025 trace_i915_request_add(rq);
3027 prev = __i915_request_commit(rq);
3029 /* Check that the context wasn't destroyed before submission */
3030 if (likely(!intel_context_is_closed(eb->context))) {
3031 attr = eb->gem_context->sched;
3033 /* Serialise with context_close via the add_to_timeline */
3034 i915_request_set_error_once(rq, -ENOENT);
3035 __i915_request_skip(rq);
3036 err = -ENOENT; /* override any transient errors */
3039 if (intel_context_is_parallel(eb->context)) {
3041 __i915_request_skip(rq);
3042 set_bit(I915_FENCE_FLAG_SKIP_PARALLEL,
3046 set_bit(I915_FENCE_FLAG_SUBMIT_PARALLEL,
3050 __i915_request_queue(rq, &attr);
3052 /* Try to clean up the client's timeline after submitting the request */
3054 retire_requests(tl, prev);
3056 mutex_unlock(&tl->mutex);
3061 static int eb_requests_add(struct i915_execbuffer *eb, int err)
3066 * We iterate in reverse order of creation to release timeline mutexes in
3069 for_each_batch_add_order(eb, i) {
3070 struct i915_request *rq = eb->requests[i];
3074 err |= eb_request_add(eb, rq, err, i == 0);
3080 static const i915_user_extension_fn execbuf_extensions[] = {
3081 [DRM_I915_GEM_EXECBUFFER_EXT_TIMELINE_FENCES] = parse_timeline_fences,
3085 parse_execbuf2_extensions(struct drm_i915_gem_execbuffer2 *args,
3086 struct i915_execbuffer *eb)
3088 if (!(args->flags & I915_EXEC_USE_EXTENSIONS))
3091 /* The execbuf2 extension mechanism reuses cliprects_ptr. So we cannot
3092 * have another flag also using it at the same time.
3094 if (eb->args->flags & I915_EXEC_FENCE_ARRAY)
3097 if (args->num_cliprects != 0)
3100 return i915_user_extensions(u64_to_user_ptr(args->cliprects_ptr),
3102 ARRAY_SIZE(execbuf_extensions),
3106 static void eb_requests_get(struct i915_execbuffer *eb)
3110 for_each_batch_create_order(eb, i) {
3111 if (!eb->requests[i])
3114 i915_request_get(eb->requests[i]);
3118 static void eb_requests_put(struct i915_execbuffer *eb)
3122 for_each_batch_create_order(eb, i) {
3123 if (!eb->requests[i])
3126 i915_request_put(eb->requests[i]);
3130 static struct sync_file *
3131 eb_composite_fence_create(struct i915_execbuffer *eb, int out_fence_fd)
3133 struct sync_file *out_fence = NULL;
3134 struct dma_fence_array *fence_array;
3135 struct dma_fence **fences;
3138 GEM_BUG_ON(!intel_context_is_parent(eb->context));
3140 fences = kmalloc_array(eb->num_batches, sizeof(*fences), GFP_KERNEL);
3142 return ERR_PTR(-ENOMEM);
3144 for_each_batch_create_order(eb, i) {
3145 fences[i] = &eb->requests[i]->fence;
3146 __set_bit(I915_FENCE_FLAG_COMPOSITE,
3147 &eb->requests[i]->fence.flags);
3150 fence_array = dma_fence_array_create(eb->num_batches,
3152 eb->context->parallel.fence_context,
3153 eb->context->parallel.seqno++,
3157 return ERR_PTR(-ENOMEM);
3160 /* Move ownership to the dma_fence_array created above */
3161 for_each_batch_create_order(eb, i)
3162 dma_fence_get(fences[i]);
3164 if (out_fence_fd != -1) {
3165 out_fence = sync_file_create(&fence_array->base);
3166 /* sync_file now owns fence_arry, drop creation ref */
3167 dma_fence_put(&fence_array->base);
3169 return ERR_PTR(-ENOMEM);
3172 eb->composite_fence = &fence_array->base;
3177 static struct sync_file *
3178 eb_fences_add(struct i915_execbuffer *eb, struct i915_request *rq,
3179 struct dma_fence *in_fence, int out_fence_fd)
3181 struct sync_file *out_fence = NULL;
3184 if (unlikely(eb->gem_context->syncobj)) {
3185 struct dma_fence *fence;
3187 fence = drm_syncobj_fence_get(eb->gem_context->syncobj);
3188 err = i915_request_await_dma_fence(rq, fence);
3189 dma_fence_put(fence);
3191 return ERR_PTR(err);
3195 if (eb->args->flags & I915_EXEC_FENCE_SUBMIT)
3196 err = i915_request_await_execution(rq, in_fence);
3198 err = i915_request_await_dma_fence(rq, in_fence);
3200 return ERR_PTR(err);
3204 err = await_fence_array(eb, rq);
3206 return ERR_PTR(err);
3209 if (intel_context_is_parallel(eb->context)) {
3210 out_fence = eb_composite_fence_create(eb, out_fence_fd);
3211 if (IS_ERR(out_fence))
3212 return ERR_PTR(-ENOMEM);
3213 } else if (out_fence_fd != -1) {
3214 out_fence = sync_file_create(&rq->fence);
3216 return ERR_PTR(-ENOMEM);
3222 static struct intel_context *
3223 eb_find_context(struct i915_execbuffer *eb, unsigned int context_number)
3225 struct intel_context *child;
3227 if (likely(context_number == 0))
3230 for_each_child(eb->context, child)
3231 if (!--context_number)
3234 GEM_BUG_ON("Context not found");
3239 static struct sync_file *
3240 eb_requests_create(struct i915_execbuffer *eb, struct dma_fence *in_fence,
3243 struct sync_file *out_fence = NULL;
3246 for_each_batch_create_order(eb, i) {
3247 /* Allocate a request for this batch buffer nice and early. */
3248 eb->requests[i] = i915_request_create(eb_find_context(eb, i));
3249 if (IS_ERR(eb->requests[i])) {
3250 out_fence = ERR_CAST(eb->requests[i]);
3251 eb->requests[i] = NULL;
3256 * Only the first request added (committed to backend) has to
3257 * take the in fences into account as all subsequent requests
3258 * will have fences inserted inbetween them.
3260 if (i + 1 == eb->num_batches) {
3261 out_fence = eb_fences_add(eb, eb->requests[i],
3262 in_fence, out_fence_fd);
3263 if (IS_ERR(out_fence))
3268 * Not really on stack, but we don't want to call
3269 * kfree on the batch_snapshot when we put it, so use the
3270 * _onstack interface.
3272 if (eb->batches[i]->vma)
3273 eb->requests[i]->batch_res =
3274 i915_vma_resource_get(eb->batches[i]->vma->resource);
3275 if (eb->batch_pool) {
3276 GEM_BUG_ON(intel_context_is_parallel(eb->context));
3277 intel_gt_buffer_pool_mark_active(eb->batch_pool,
3286 i915_gem_do_execbuffer(struct drm_device *dev,
3287 struct drm_file *file,
3288 struct drm_i915_gem_execbuffer2 *args,
3289 struct drm_i915_gem_exec_object2 *exec)
3291 struct drm_i915_private *i915 = to_i915(dev);
3292 struct i915_execbuffer eb;
3293 struct dma_fence *in_fence = NULL;
3294 struct sync_file *out_fence = NULL;
3295 int out_fence_fd = -1;
3298 BUILD_BUG_ON(__EXEC_INTERNAL_FLAGS & ~__I915_EXEC_ILLEGAL_FLAGS);
3299 BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS &
3300 ~__EXEC_OBJECT_UNKNOWN_FLAGS);
3305 if (DBG_FORCE_RELOC || !(args->flags & I915_EXEC_NO_RELOC))
3306 args->flags |= __EXEC_HAS_RELOC;
3309 eb.vma = (struct eb_vma *)(exec + args->buffer_count + 1);
3310 eb.vma[0].vma = NULL;
3311 eb.batch_pool = NULL;
3313 eb.invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
3314 reloc_cache_init(&eb.reloc_cache, eb.i915);
3316 eb.buffer_count = args->buffer_count;
3317 eb.batch_start_offset = args->batch_start_offset;
3318 eb.trampoline = NULL;
3323 eb_capture_list_clear(&eb);
3325 memset(eb.requests, 0, sizeof(struct i915_request *) *
3326 ARRAY_SIZE(eb.requests));
3327 eb.composite_fence = NULL;
3330 if (args->flags & I915_EXEC_SECURE) {
3331 if (GRAPHICS_VER(i915) >= 11)
3334 /* Return -EPERM to trigger fallback code on old binaries. */
3335 if (!HAS_SECURE_BATCHES(i915))
3338 if (!drm_is_current_master(file) || !capable(CAP_SYS_ADMIN))
3341 eb.batch_flags |= I915_DISPATCH_SECURE;
3343 if (args->flags & I915_EXEC_IS_PINNED)
3344 eb.batch_flags |= I915_DISPATCH_PINNED;
3346 err = parse_execbuf2_extensions(args, &eb);
3350 err = add_fence_array(&eb);
3354 #define IN_FENCES (I915_EXEC_FENCE_IN | I915_EXEC_FENCE_SUBMIT)
3355 if (args->flags & IN_FENCES) {
3356 if ((args->flags & IN_FENCES) == IN_FENCES)
3359 in_fence = sync_file_get_fence(lower_32_bits(args->rsvd2));
3367 if (args->flags & I915_EXEC_FENCE_OUT) {
3368 out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
3369 if (out_fence_fd < 0) {
3375 err = eb_create(&eb);
3379 GEM_BUG_ON(!eb.lut_size);
3381 err = eb_select_context(&eb);
3385 err = eb_select_engine(&eb);
3389 err = eb_lookup_vmas(&eb);
3391 eb_release_vmas(&eb, true);
3395 i915_gem_ww_ctx_init(&eb.ww, true);
3397 err = eb_relocate_parse(&eb);
3400 * If the user expects the execobject.offset and
3401 * reloc.presumed_offset to be an exact match,
3402 * as for using NO_RELOC, then we cannot update
3403 * the execobject.offset until we have completed
3406 args->flags &= ~__EXEC_HAS_RELOC;
3410 ww_acquire_done(&eb.ww.ctx);
3411 eb_capture_stage(&eb);
3413 out_fence = eb_requests_create(&eb, in_fence, out_fence_fd);
3414 if (IS_ERR(out_fence)) {
3415 err = PTR_ERR(out_fence);
3423 err = eb_submit(&eb);
3426 eb_requests_get(&eb);
3427 err = eb_requests_add(&eb, err);
3430 signal_fence_array(&eb, eb.composite_fence ?
3431 eb.composite_fence :
3432 &eb.requests[0]->fence);
3436 fd_install(out_fence_fd, out_fence->file);
3437 args->rsvd2 &= GENMASK_ULL(31, 0); /* keep in-fence */
3438 args->rsvd2 |= (u64)out_fence_fd << 32;
3441 fput(out_fence->file);
3445 if (unlikely(eb.gem_context->syncobj)) {
3446 drm_syncobj_replace_fence(eb.gem_context->syncobj,
3447 eb.composite_fence ?
3448 eb.composite_fence :
3449 &eb.requests[0]->fence);
3452 if (!out_fence && eb.composite_fence)
3453 dma_fence_put(eb.composite_fence);
3455 eb_requests_put(&eb);
3458 eb_release_vmas(&eb, true);
3459 WARN_ON(err == -EDEADLK);
3460 i915_gem_ww_ctx_fini(&eb.ww);
3463 intel_gt_buffer_pool_put(eb.batch_pool);
3467 i915_gem_context_put(eb.gem_context);
3471 if (out_fence_fd != -1)
3472 put_unused_fd(out_fence_fd);
3474 dma_fence_put(in_fence);
3476 put_fence_array(eb.fences, eb.num_fences);
3480 static size_t eb_element_size(void)
3482 return sizeof(struct drm_i915_gem_exec_object2) + sizeof(struct eb_vma);
3485 static bool check_buffer_count(size_t count)
3487 const size_t sz = eb_element_size();
3490 * When using LUT_HANDLE, we impose a limit of INT_MAX for the lookup
3491 * array size (see eb_create()). Otherwise, we can accept an array as
3492 * large as can be addressed (though use large arrays at your peril)!
3495 return !(count < 1 || count > INT_MAX || count > SIZE_MAX / sz - 1);
3499 i915_gem_execbuffer2_ioctl(struct drm_device *dev, void *data,
3500 struct drm_file *file)
3502 struct drm_i915_private *i915 = to_i915(dev);
3503 struct drm_i915_gem_execbuffer2 *args = data;
3504 struct drm_i915_gem_exec_object2 *exec2_list;
3505 const size_t count = args->buffer_count;
3508 if (!check_buffer_count(count)) {
3509 drm_dbg(&i915->drm, "execbuf2 with %zd buffers\n", count);
3513 err = i915_gem_check_execbuffer(args);
3517 /* Allocate extra slots for use by the command parser */
3518 exec2_list = kvmalloc_array(count + 2, eb_element_size(),
3519 __GFP_NOWARN | GFP_KERNEL);
3520 if (exec2_list == NULL) {
3521 drm_dbg(&i915->drm, "Failed to allocate exec list for %zd buffers\n",
3525 if (copy_from_user(exec2_list,
3526 u64_to_user_ptr(args->buffers_ptr),
3527 sizeof(*exec2_list) * count)) {
3528 drm_dbg(&i915->drm, "copy %zd exec entries failed\n", count);
3533 err = i915_gem_do_execbuffer(dev, file, args, exec2_list);
3536 * Now that we have begun execution of the batchbuffer, we ignore
3537 * any new error after this point. Also given that we have already
3538 * updated the associated relocations, we try to write out the current
3539 * object locations irrespective of any error.
3541 if (args->flags & __EXEC_HAS_RELOC) {
3542 struct drm_i915_gem_exec_object2 __user *user_exec_list =
3543 u64_to_user_ptr(args->buffers_ptr);
3546 /* Copy the new buffer offsets back to the user's exec list. */
3548 * Note: count * sizeof(*user_exec_list) does not overflow,
3549 * because we checked 'count' in check_buffer_count().
3551 * And this range already got effectively checked earlier
3552 * when we did the "copy_from_user()" above.
3554 if (!user_write_access_begin(user_exec_list,
3555 count * sizeof(*user_exec_list)))
3558 for (i = 0; i < args->buffer_count; i++) {
3559 if (!(exec2_list[i].offset & UPDATE))
3562 exec2_list[i].offset =
3563 gen8_canonical_addr(exec2_list[i].offset & PIN_OFFSET_MASK);
3564 unsafe_put_user(exec2_list[i].offset,
3565 &user_exec_list[i].offset,
3569 user_write_access_end();
3573 args->flags &= ~__I915_EXEC_UNKNOWN_FLAGS;