1 | /* |
2 | * SPDX-License-Identifier: MIT |
3 | * |
4 | * Copyright © 2008,2010 Intel Corporation |
5 | */ |
6 | |
7 | #include <linux/dma-resv.h> |
8 | #include <linux/highmem.h> |
9 | #include <linux/sync_file.h> |
10 | #include <linux/uaccess.h> |
11 | |
12 | #include <drm/drm_auth.h> |
13 | #include <drm/drm_syncobj.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_gpu_commands.h" |
20 | #include "gt/intel_gt.h" |
21 | #include "gt/intel_gt_buffer_pool.h" |
22 | #include "gt/intel_gt_pm.h" |
23 | #include "gt/intel_ring.h" |
24 | |
25 | #include "pxp/intel_pxp.h" |
26 | |
27 | #include "i915_cmd_parser.h" |
28 | #include "i915_drv.h" |
29 | #include "i915_file_private.h" |
30 | #include "i915_gem_clflush.h" |
31 | #include "i915_gem_context.h" |
32 | #include "i915_gem_evict.h" |
33 | #include "i915_gem_ioctls.h" |
34 | #include "i915_reg.h" |
35 | #include "i915_trace.h" |
36 | #include "i915_user_extensions.h" |
37 | |
38 | struct eb_vma { |
39 | struct i915_vma *vma; |
40 | unsigned int flags; |
41 | |
42 | /** This vma's place in the execbuf reservation list */ |
43 | struct drm_i915_gem_exec_object2 *exec; |
44 | struct list_head bind_link; |
45 | struct list_head reloc_link; |
46 | |
47 | struct hlist_node node; |
48 | u32 handle; |
49 | }; |
50 | |
51 | enum { |
52 | FORCE_CPU_RELOC = 1, |
53 | FORCE_GTT_RELOC, |
54 | FORCE_GPU_RELOC, |
55 | #define DBG_FORCE_RELOC 0 /* choose one of the above! */ |
56 | }; |
57 | |
58 | /* __EXEC_OBJECT_ flags > BIT(29) defined in i915_vma.h */ |
59 | #define __EXEC_OBJECT_HAS_PIN BIT(29) |
60 | #define __EXEC_OBJECT_HAS_FENCE BIT(28) |
61 | #define __EXEC_OBJECT_USERPTR_INIT BIT(27) |
62 | #define __EXEC_OBJECT_NEEDS_MAP BIT(26) |
63 | #define __EXEC_OBJECT_NEEDS_BIAS BIT(25) |
64 | #define __EXEC_OBJECT_INTERNAL_FLAGS (~0u << 25) /* all of the above + */ |
65 | #define __EXEC_OBJECT_RESERVED (__EXEC_OBJECT_HAS_PIN | __EXEC_OBJECT_HAS_FENCE) |
66 | |
67 | #define __EXEC_HAS_RELOC BIT(31) |
68 | #define __EXEC_ENGINE_PINNED BIT(30) |
69 | #define __EXEC_USERPTR_USED BIT(29) |
70 | #define __EXEC_INTERNAL_FLAGS (~0u << 29) |
71 | #define UPDATE PIN_OFFSET_FIXED |
72 | |
73 | #define BATCH_OFFSET_BIAS (256*1024) |
74 | |
75 | #define __I915_EXEC_ILLEGAL_FLAGS \ |
76 | (__I915_EXEC_UNKNOWN_FLAGS | \ |
77 | I915_EXEC_CONSTANTS_MASK | \ |
78 | I915_EXEC_RESOURCE_STREAMER) |
79 | |
80 | /* Catch emission of unexpected errors for CI! */ |
81 | #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) |
82 | #undef EINVAL |
83 | #define EINVAL ({ \ |
84 | DRM_DEBUG_DRIVER("EINVAL at %s:%d\n", __func__, __LINE__); \ |
85 | 22; \ |
86 | }) |
87 | #endif |
88 | |
89 | /** |
90 | * DOC: User command execution |
91 | * |
92 | * Userspace submits commands to be executed on the GPU as an instruction |
93 | * stream within a GEM object we call a batchbuffer. This instructions may |
94 | * refer to other GEM objects containing auxiliary state such as kernels, |
95 | * samplers, render targets and even secondary batchbuffers. Userspace does |
96 | * not know where in the GPU memory these objects reside and so before the |
97 | * batchbuffer is passed to the GPU for execution, those addresses in the |
98 | * batchbuffer and auxiliary objects are updated. This is known as relocation, |
99 | * or patching. To try and avoid having to relocate each object on the next |
100 | * execution, userspace is told the location of those objects in this pass, |
101 | * but this remains just a hint as the kernel may choose a new location for |
102 | * any object in the future. |
103 | * |
104 | * At the level of talking to the hardware, submitting a batchbuffer for the |
105 | * GPU to execute is to add content to a buffer from which the HW |
106 | * command streamer is reading. |
107 | * |
108 | * 1. Add a command to load the HW context. For Logical Ring Contexts, i.e. |
109 | * Execlists, this command is not placed on the same buffer as the |
110 | * remaining items. |
111 | * |
112 | * 2. Add a command to invalidate caches to the buffer. |
113 | * |
114 | * 3. Add a batchbuffer start command to the buffer; the start command is |
115 | * essentially a token together with the GPU address of the batchbuffer |
116 | * to be executed. |
117 | * |
118 | * 4. Add a pipeline flush to the buffer. |
119 | * |
120 | * 5. Add a memory write command to the buffer to record when the GPU |
121 | * is done executing the batchbuffer. The memory write writes the |
122 | * global sequence number of the request, ``i915_request::global_seqno``; |
123 | * the i915 driver uses the current value in the register to determine |
124 | * if the GPU has completed the batchbuffer. |
125 | * |
126 | * 6. Add a user interrupt command to the buffer. This command instructs |
127 | * the GPU to issue an interrupt when the command, pipeline flush and |
128 | * memory write are completed. |
129 | * |
130 | * 7. Inform the hardware of the additional commands added to the buffer |
131 | * (by updating the tail pointer). |
132 | * |
133 | * Processing an execbuf ioctl is conceptually split up into a few phases. |
134 | * |
135 | * 1. Validation - Ensure all the pointers, handles and flags are valid. |
136 | * 2. Reservation - Assign GPU address space for every object |
137 | * 3. Relocation - Update any addresses to point to the final locations |
138 | * 4. Serialisation - Order the request with respect to its dependencies |
139 | * 5. Construction - Construct a request to execute the batchbuffer |
140 | * 6. Submission (at some point in the future execution) |
141 | * |
142 | * Reserving resources for the execbuf is the most complicated phase. We |
143 | * neither want to have to migrate the object in the address space, nor do |
144 | * we want to have to update any relocations pointing to this object. Ideally, |
145 | * we want to leave the object where it is and for all the existing relocations |
146 | * to match. If the object is given a new address, or if userspace thinks the |
147 | * object is elsewhere, we have to parse all the relocation entries and update |
148 | * the addresses. Userspace can set the I915_EXEC_NORELOC flag to hint that |
149 | * all the target addresses in all of its objects match the value in the |
150 | * relocation entries and that they all match the presumed offsets given by the |
151 | * list of execbuffer objects. Using this knowledge, we know that if we haven't |
152 | * moved any buffers, all the relocation entries are valid and we can skip |
153 | * the update. (If userspace is wrong, the likely outcome is an impromptu GPU |
154 | * hang.) The requirement for using I915_EXEC_NO_RELOC are: |
155 | * |
156 | * The addresses written in the objects must match the corresponding |
157 | * reloc.presumed_offset which in turn must match the corresponding |
158 | * execobject.offset. |
159 | * |
160 | * Any render targets written to in the batch must be flagged with |
161 | * EXEC_OBJECT_WRITE. |
162 | * |
163 | * To avoid stalling, execobject.offset should match the current |
164 | * address of that object within the active context. |
165 | * |
166 | * The reservation is done is multiple phases. First we try and keep any |
167 | * object already bound in its current location - so as long as meets the |
168 | * constraints imposed by the new execbuffer. Any object left unbound after the |
169 | * first pass is then fitted into any available idle space. If an object does |
170 | * not fit, all objects are removed from the reservation and the process rerun |
171 | * after sorting the objects into a priority order (more difficult to fit |
172 | * objects are tried first). Failing that, the entire VM is cleared and we try |
173 | * to fit the execbuf once last time before concluding that it simply will not |
174 | * fit. |
175 | * |
176 | * A small complication to all of this is that we allow userspace not only to |
177 | * specify an alignment and a size for the object in the address space, but |
178 | * we also allow userspace to specify the exact offset. This objects are |
179 | * simpler to place (the location is known a priori) all we have to do is make |
180 | * sure the space is available. |
181 | * |
182 | * Once all the objects are in place, patching up the buried pointers to point |
183 | * to the final locations is a fairly simple job of walking over the relocation |
184 | * entry arrays, looking up the right address and rewriting the value into |
185 | * the object. Simple! ... The relocation entries are stored in user memory |
186 | * and so to access them we have to copy them into a local buffer. That copy |
187 | * has to avoid taking any pagefaults as they may lead back to a GEM object |
188 | * requiring the struct_mutex (i.e. recursive deadlock). So once again we split |
189 | * the relocation into multiple passes. First we try to do everything within an |
190 | * atomic context (avoid the pagefaults) which requires that we never wait. If |
191 | * we detect that we may wait, or if we need to fault, then we have to fallback |
192 | * to a slower path. The slowpath has to drop the mutex. (Can you hear alarm |
193 | * bells yet?) Dropping the mutex means that we lose all the state we have |
194 | * built up so far for the execbuf and we must reset any global data. However, |
195 | * we do leave the objects pinned in their final locations - which is a |
196 | * potential issue for concurrent execbufs. Once we have left the mutex, we can |
197 | * allocate and copy all the relocation entries into a large array at our |
198 | * leisure, reacquire the mutex, reclaim all the objects and other state and |
199 | * then proceed to update any incorrect addresses with the objects. |
200 | * |
201 | * As we process the relocation entries, we maintain a record of whether the |
202 | * object is being written to. Using NORELOC, we expect userspace to provide |
203 | * this information instead. We also check whether we can skip the relocation |
204 | * by comparing the expected value inside the relocation entry with the target's |
205 | * final address. If they differ, we have to map the current object and rewrite |
206 | * the 4 or 8 byte pointer within. |
207 | * |
208 | * Serialising an execbuf is quite simple according to the rules of the GEM |
209 | * ABI. Execution within each context is ordered by the order of submission. |
210 | * Writes to any GEM object are in order of submission and are exclusive. Reads |
211 | * from a GEM object are unordered with respect to other reads, but ordered by |
212 | * writes. A write submitted after a read cannot occur before the read, and |
213 | * similarly any read submitted after a write cannot occur before the write. |
214 | * Writes are ordered between engines such that only one write occurs at any |
215 | * time (completing any reads beforehand) - using semaphores where available |
216 | * and CPU serialisation otherwise. Other GEM access obey the same rules, any |
217 | * write (either via mmaps using set-domain, or via pwrite) must flush all GPU |
218 | * reads before starting, and any read (either using set-domain or pread) must |
219 | * flush all GPU writes before starting. (Note we only employ a barrier before, |
220 | * we currently rely on userspace not concurrently starting a new execution |
221 | * whilst reading or writing to an object. This may be an advantage or not |
222 | * depending on how much you trust userspace not to shoot themselves in the |
223 | * foot.) Serialisation may just result in the request being inserted into |
224 | * a DAG awaiting its turn, but most simple is to wait on the CPU until |
225 | * all dependencies are resolved. |
226 | * |
227 | * After all of that, is just a matter of closing the request and handing it to |
228 | * the hardware (well, leaving it in a queue to be executed). However, we also |
229 | * offer the ability for batchbuffers to be run with elevated privileges so |
230 | * that they access otherwise hidden registers. (Used to adjust L3 cache etc.) |
231 | * Before any batch is given extra privileges we first must check that it |
232 | * contains no nefarious instructions, we check that each instruction is from |
233 | * our whitelist and all registers are also from an allowed list. We first |
234 | * copy the user's batchbuffer to a shadow (so that the user doesn't have |
235 | * access to it, either by the CPU or GPU as we scan it) and then parse each |
236 | * instruction. If everything is ok, we set a flag telling the hardware to run |
237 | * the batchbuffer in trusted mode, otherwise the ioctl is rejected. |
238 | */ |
239 | |
240 | struct eb_fence { |
241 | struct drm_syncobj *syncobj; /* Use with ptr_mask_bits() */ |
242 | struct dma_fence *dma_fence; |
243 | u64 value; |
244 | struct dma_fence_chain *chain_fence; |
245 | }; |
246 | |
247 | struct i915_execbuffer { |
248 | struct drm_i915_private *i915; /** i915 backpointer */ |
249 | struct drm_file *file; /** per-file lookup tables and limits */ |
250 | struct drm_i915_gem_execbuffer2 *args; /** ioctl parameters */ |
251 | struct drm_i915_gem_exec_object2 *exec; /** ioctl execobj[] */ |
252 | struct eb_vma *vma; |
253 | |
254 | struct intel_gt *gt; /* gt for the execbuf */ |
255 | struct intel_context *context; /* logical state for the request */ |
256 | struct i915_gem_context *gem_context; /** caller's context */ |
257 | intel_wakeref_t wakeref; |
258 | intel_wakeref_t wakeref_gt0; |
259 | |
260 | /** our requests to build */ |
261 | struct i915_request *requests[MAX_ENGINE_INSTANCE + 1]; |
262 | /** identity of the batch obj/vma */ |
263 | struct eb_vma *batches[MAX_ENGINE_INSTANCE + 1]; |
264 | struct i915_vma *trampoline; /** trampoline used for chaining */ |
265 | |
266 | /** used for excl fence in dma_resv objects when > 1 BB submitted */ |
267 | struct dma_fence *composite_fence; |
268 | |
269 | /** actual size of execobj[] as we may extend it for the cmdparser */ |
270 | unsigned int buffer_count; |
271 | |
272 | /* number of batches in execbuf IOCTL */ |
273 | unsigned int num_batches; |
274 | |
275 | /** list of vma not yet bound during reservation phase */ |
276 | struct list_head unbound; |
277 | |
278 | /** list of vma that have execobj.relocation_count */ |
279 | struct list_head relocs; |
280 | |
281 | struct i915_gem_ww_ctx ww; |
282 | |
283 | /** |
284 | * Track the most recently used object for relocations, as we |
285 | * frequently have to perform multiple relocations within the same |
286 | * obj/page |
287 | */ |
288 | struct reloc_cache { |
289 | struct drm_mm_node node; /** temporary GTT binding */ |
290 | unsigned long vaddr; /** Current kmap address */ |
291 | unsigned long page; /** Currently mapped page index */ |
292 | unsigned int graphics_ver; /** Cached value of GRAPHICS_VER */ |
293 | bool use_64bit_reloc : 1; |
294 | bool has_llc : 1; |
295 | bool has_fence : 1; |
296 | bool needs_unfenced : 1; |
297 | } reloc_cache; |
298 | |
299 | u64 invalid_flags; /** Set of execobj.flags that are invalid */ |
300 | |
301 | /** Length of batch within object */ |
302 | u64 batch_len[MAX_ENGINE_INSTANCE + 1]; |
303 | u32 batch_start_offset; /** Location within object of batch */ |
304 | u32 batch_flags; /** Flags composed for emit_bb_start() */ |
305 | struct intel_gt_buffer_pool_node *batch_pool; /** pool node for batch buffer */ |
306 | |
307 | /** |
308 | * Indicate either the size of the hastable used to resolve |
309 | * relocation handles, or if negative that we are using a direct |
310 | * index into the execobj[]. |
311 | */ |
312 | int lut_size; |
313 | struct hlist_head *buckets; /** ht for relocation handles */ |
314 | |
315 | struct eb_fence *fences; |
316 | unsigned long num_fences; |
317 | #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR) |
318 | struct i915_capture_list *capture_lists[MAX_ENGINE_INSTANCE + 1]; |
319 | #endif |
320 | }; |
321 | |
322 | static int eb_parse(struct i915_execbuffer *eb); |
323 | static int eb_pin_engine(struct i915_execbuffer *eb, bool throttle); |
324 | static void eb_unpin_engine(struct i915_execbuffer *eb); |
325 | static void eb_capture_release(struct i915_execbuffer *eb); |
326 | |
327 | static bool eb_use_cmdparser(const struct i915_execbuffer *eb) |
328 | { |
329 | return intel_engine_requires_cmd_parser(engine: eb->context->engine) || |
330 | (intel_engine_using_cmd_parser(engine: eb->context->engine) && |
331 | eb->args->batch_len); |
332 | } |
333 | |
334 | static int eb_create(struct i915_execbuffer *eb) |
335 | { |
336 | if (!(eb->args->flags & I915_EXEC_HANDLE_LUT)) { |
337 | unsigned int size = 1 + ilog2(eb->buffer_count); |
338 | |
339 | /* |
340 | * Without a 1:1 association between relocation handles and |
341 | * the execobject[] index, we instead create a hashtable. |
342 | * We size it dynamically based on available memory, starting |
343 | * first with 1:1 assocative hash and scaling back until |
344 | * the allocation succeeds. |
345 | * |
346 | * Later on we use a positive lut_size to indicate we are |
347 | * using this hashtable, and a negative value to indicate a |
348 | * direct lookup. |
349 | */ |
350 | do { |
351 | gfp_t flags; |
352 | |
353 | /* While we can still reduce the allocation size, don't |
354 | * raise a warning and allow the allocation to fail. |
355 | * On the last pass though, we want to try as hard |
356 | * as possible to perform the allocation and warn |
357 | * if it fails. |
358 | */ |
359 | flags = GFP_KERNEL; |
360 | if (size > 1) |
361 | flags |= __GFP_NORETRY | __GFP_NOWARN; |
362 | |
363 | eb->buckets = kzalloc(size: sizeof(struct hlist_head) << size, |
364 | flags); |
365 | if (eb->buckets) |
366 | break; |
367 | } while (--size); |
368 | |
369 | if (unlikely(!size)) |
370 | return -ENOMEM; |
371 | |
372 | eb->lut_size = size; |
373 | } else { |
374 | eb->lut_size = -eb->buffer_count; |
375 | } |
376 | |
377 | return 0; |
378 | } |
379 | |
380 | static bool |
381 | eb_vma_misplaced(const struct drm_i915_gem_exec_object2 *entry, |
382 | const struct i915_vma *vma, |
383 | unsigned int flags) |
384 | { |
385 | const u64 start = i915_vma_offset(vma); |
386 | const u64 size = i915_vma_size(vma); |
387 | |
388 | if (size < entry->pad_to_size) |
389 | return true; |
390 | |
391 | if (entry->alignment && !IS_ALIGNED(start, entry->alignment)) |
392 | return true; |
393 | |
394 | if (flags & EXEC_OBJECT_PINNED && |
395 | start != entry->offset) |
396 | return true; |
397 | |
398 | if (flags & __EXEC_OBJECT_NEEDS_BIAS && |
399 | start < BATCH_OFFSET_BIAS) |
400 | return true; |
401 | |
402 | if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) && |
403 | (start + size + 4095) >> 32) |
404 | return true; |
405 | |
406 | if (flags & __EXEC_OBJECT_NEEDS_MAP && |
407 | !i915_vma_is_map_and_fenceable(vma)) |
408 | return true; |
409 | |
410 | return false; |
411 | } |
412 | |
413 | static u64 eb_pin_flags(const struct drm_i915_gem_exec_object2 *entry, |
414 | unsigned int exec_flags) |
415 | { |
416 | u64 pin_flags = 0; |
417 | |
418 | if (exec_flags & EXEC_OBJECT_NEEDS_GTT) |
419 | pin_flags |= PIN_GLOBAL; |
420 | |
421 | /* |
422 | * Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset, |
423 | * limit address to the first 4GBs for unflagged objects. |
424 | */ |
425 | if (!(exec_flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS)) |
426 | pin_flags |= PIN_ZONE_4G; |
427 | |
428 | if (exec_flags & __EXEC_OBJECT_NEEDS_MAP) |
429 | pin_flags |= PIN_MAPPABLE; |
430 | |
431 | if (exec_flags & EXEC_OBJECT_PINNED) |
432 | pin_flags |= entry->offset | PIN_OFFSET_FIXED; |
433 | else if (exec_flags & __EXEC_OBJECT_NEEDS_BIAS) |
434 | pin_flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS; |
435 | |
436 | return pin_flags; |
437 | } |
438 | |
439 | static int |
440 | eb_pin_vma(struct i915_execbuffer *eb, |
441 | const struct drm_i915_gem_exec_object2 *entry, |
442 | struct eb_vma *ev) |
443 | { |
444 | struct i915_vma *vma = ev->vma; |
445 | u64 pin_flags; |
446 | int err; |
447 | |
448 | if (vma->node.size) |
449 | pin_flags = __i915_vma_offset(vma); |
450 | else |
451 | pin_flags = entry->offset & PIN_OFFSET_MASK; |
452 | |
453 | pin_flags |= PIN_USER | PIN_NOEVICT | PIN_OFFSET_FIXED | PIN_VALIDATE; |
454 | if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_GTT)) |
455 | pin_flags |= PIN_GLOBAL; |
456 | |
457 | /* Attempt to reuse the current location if available */ |
458 | err = i915_vma_pin_ww(vma, ww: &eb->ww, size: 0, alignment: 0, flags: pin_flags); |
459 | if (err == -EDEADLK) |
460 | return err; |
461 | |
462 | if (unlikely(err)) { |
463 | if (entry->flags & EXEC_OBJECT_PINNED) |
464 | return err; |
465 | |
466 | /* Failing that pick any _free_ space if suitable */ |
467 | err = i915_vma_pin_ww(vma, ww: &eb->ww, |
468 | size: entry->pad_to_size, |
469 | alignment: entry->alignment, |
470 | flags: eb_pin_flags(entry, exec_flags: ev->flags) | |
471 | PIN_USER | PIN_NOEVICT | PIN_VALIDATE); |
472 | if (unlikely(err)) |
473 | return err; |
474 | } |
475 | |
476 | if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) { |
477 | err = i915_vma_pin_fence(vma); |
478 | if (unlikely(err)) |
479 | return err; |
480 | |
481 | if (vma->fence) |
482 | ev->flags |= __EXEC_OBJECT_HAS_FENCE; |
483 | } |
484 | |
485 | ev->flags |= __EXEC_OBJECT_HAS_PIN; |
486 | if (eb_vma_misplaced(entry, vma, flags: ev->flags)) |
487 | return -EBADSLT; |
488 | |
489 | return 0; |
490 | } |
491 | |
492 | static void |
493 | eb_unreserve_vma(struct eb_vma *ev) |
494 | { |
495 | if (unlikely(ev->flags & __EXEC_OBJECT_HAS_FENCE)) |
496 | __i915_vma_unpin_fence(vma: ev->vma); |
497 | |
498 | ev->flags &= ~__EXEC_OBJECT_RESERVED; |
499 | } |
500 | |
501 | static int |
502 | eb_validate_vma(struct i915_execbuffer *eb, |
503 | struct drm_i915_gem_exec_object2 *entry, |
504 | struct i915_vma *vma) |
505 | { |
506 | /* Relocations are disallowed for all platforms after TGL-LP. This |
507 | * also covers all platforms with local memory. |
508 | */ |
509 | if (entry->relocation_count && |
510 | GRAPHICS_VER(eb->i915) >= 12 && !IS_TIGERLAKE(eb->i915)) |
511 | return -EINVAL; |
512 | |
513 | if (unlikely(entry->flags & eb->invalid_flags)) |
514 | return -EINVAL; |
515 | |
516 | if (unlikely(entry->alignment && |
517 | !is_power_of_2_u64(entry->alignment))) |
518 | return -EINVAL; |
519 | |
520 | /* |
521 | * Offset can be used as input (EXEC_OBJECT_PINNED), reject |
522 | * any non-page-aligned or non-canonical addresses. |
523 | */ |
524 | if (unlikely(entry->flags & EXEC_OBJECT_PINNED && |
525 | entry->offset != gen8_canonical_addr(entry->offset & I915_GTT_PAGE_MASK))) |
526 | return -EINVAL; |
527 | |
528 | /* pad_to_size was once a reserved field, so sanitize it */ |
529 | if (entry->flags & EXEC_OBJECT_PAD_TO_SIZE) { |
530 | if (unlikely(offset_in_page(entry->pad_to_size))) |
531 | return -EINVAL; |
532 | } else { |
533 | entry->pad_to_size = 0; |
534 | } |
535 | /* |
536 | * From drm_mm perspective address space is continuous, |
537 | * so from this point we're always using non-canonical |
538 | * form internally. |
539 | */ |
540 | entry->offset = gen8_noncanonical_addr(address: entry->offset); |
541 | |
542 | if (!eb->reloc_cache.has_fence) { |
543 | entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE; |
544 | } else { |
545 | if ((entry->flags & EXEC_OBJECT_NEEDS_FENCE || |
546 | eb->reloc_cache.needs_unfenced) && |
547 | i915_gem_object_is_tiled(obj: vma->obj)) |
548 | entry->flags |= EXEC_OBJECT_NEEDS_GTT | __EXEC_OBJECT_NEEDS_MAP; |
549 | } |
550 | |
551 | return 0; |
552 | } |
553 | |
554 | static bool |
555 | is_batch_buffer(struct i915_execbuffer *eb, unsigned int buffer_idx) |
556 | { |
557 | return eb->args->flags & I915_EXEC_BATCH_FIRST ? |
558 | buffer_idx < eb->num_batches : |
559 | buffer_idx >= eb->args->buffer_count - eb->num_batches; |
560 | } |
561 | |
562 | static int |
563 | eb_add_vma(struct i915_execbuffer *eb, |
564 | unsigned int *current_batch, |
565 | unsigned int i, |
566 | struct i915_vma *vma) |
567 | { |
568 | struct drm_i915_private *i915 = eb->i915; |
569 | struct drm_i915_gem_exec_object2 *entry = &eb->exec[i]; |
570 | struct eb_vma *ev = &eb->vma[i]; |
571 | |
572 | ev->vma = vma; |
573 | ev->exec = entry; |
574 | ev->flags = entry->flags; |
575 | |
576 | if (eb->lut_size > 0) { |
577 | ev->handle = entry->handle; |
578 | hlist_add_head(n: &ev->node, |
579 | h: &eb->buckets[hash_32(val: entry->handle, |
580 | bits: eb->lut_size)]); |
581 | } |
582 | |
583 | if (entry->relocation_count) |
584 | list_add_tail(new: &ev->reloc_link, head: &eb->relocs); |
585 | |
586 | /* |
587 | * SNA is doing fancy tricks with compressing batch buffers, which leads |
588 | * to negative relocation deltas. Usually that works out ok since the |
589 | * relocate address is still positive, except when the batch is placed |
590 | * very low in the GTT. Ensure this doesn't happen. |
591 | * |
592 | * Note that actual hangs have only been observed on gen7, but for |
593 | * paranoia do it everywhere. |
594 | */ |
595 | if (is_batch_buffer(eb, buffer_idx: i)) { |
596 | if (entry->relocation_count && |
597 | !(ev->flags & EXEC_OBJECT_PINNED)) |
598 | ev->flags |= __EXEC_OBJECT_NEEDS_BIAS; |
599 | if (eb->reloc_cache.has_fence) |
600 | ev->flags |= EXEC_OBJECT_NEEDS_FENCE; |
601 | |
602 | eb->batches[*current_batch] = ev; |
603 | |
604 | if (unlikely(ev->flags & EXEC_OBJECT_WRITE)) { |
605 | drm_dbg(&i915->drm, |
606 | "Attempting to use self-modifying batch buffer\n" ); |
607 | return -EINVAL; |
608 | } |
609 | |
610 | if (range_overflows_t(u64, |
611 | eb->batch_start_offset, |
612 | eb->args->batch_len, |
613 | ev->vma->size)) { |
614 | drm_dbg(&i915->drm, "Attempting to use out-of-bounds batch\n" ); |
615 | return -EINVAL; |
616 | } |
617 | |
618 | if (eb->args->batch_len == 0) |
619 | eb->batch_len[*current_batch] = ev->vma->size - |
620 | eb->batch_start_offset; |
621 | else |
622 | eb->batch_len[*current_batch] = eb->args->batch_len; |
623 | if (unlikely(eb->batch_len[*current_batch] == 0)) { /* impossible! */ |
624 | drm_dbg(&i915->drm, "Invalid batch length\n" ); |
625 | return -EINVAL; |
626 | } |
627 | |
628 | ++*current_batch; |
629 | } |
630 | |
631 | return 0; |
632 | } |
633 | |
634 | static int use_cpu_reloc(const struct reloc_cache *cache, |
635 | const struct drm_i915_gem_object *obj) |
636 | { |
637 | if (!i915_gem_object_has_struct_page(obj)) |
638 | return false; |
639 | |
640 | if (DBG_FORCE_RELOC == FORCE_CPU_RELOC) |
641 | return true; |
642 | |
643 | if (DBG_FORCE_RELOC == FORCE_GTT_RELOC) |
644 | return false; |
645 | |
646 | /* |
647 | * For objects created by userspace through GEM_CREATE with pat_index |
648 | * set by set_pat extension, i915_gem_object_has_cache_level() always |
649 | * return true, otherwise the call would fall back to checking whether |
650 | * the object is un-cached. |
651 | */ |
652 | return (cache->has_llc || |
653 | obj->cache_dirty || |
654 | !i915_gem_object_has_cache_level(obj, lvl: I915_CACHE_NONE)); |
655 | } |
656 | |
657 | static int eb_reserve_vma(struct i915_execbuffer *eb, |
658 | struct eb_vma *ev, |
659 | u64 pin_flags) |
660 | { |
661 | struct drm_i915_gem_exec_object2 *entry = ev->exec; |
662 | struct i915_vma *vma = ev->vma; |
663 | int err; |
664 | |
665 | if (drm_mm_node_allocated(node: &vma->node) && |
666 | eb_vma_misplaced(entry, vma, flags: ev->flags)) { |
667 | err = i915_vma_unbind(vma); |
668 | if (err) |
669 | return err; |
670 | } |
671 | |
672 | err = i915_vma_pin_ww(vma, ww: &eb->ww, |
673 | size: entry->pad_to_size, alignment: entry->alignment, |
674 | flags: eb_pin_flags(entry, exec_flags: ev->flags) | pin_flags); |
675 | if (err) |
676 | return err; |
677 | |
678 | if (entry->offset != i915_vma_offset(vma)) { |
679 | entry->offset = i915_vma_offset(vma) | UPDATE; |
680 | eb->args->flags |= __EXEC_HAS_RELOC; |
681 | } |
682 | |
683 | if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) { |
684 | err = i915_vma_pin_fence(vma); |
685 | if (unlikely(err)) |
686 | return err; |
687 | |
688 | if (vma->fence) |
689 | ev->flags |= __EXEC_OBJECT_HAS_FENCE; |
690 | } |
691 | |
692 | ev->flags |= __EXEC_OBJECT_HAS_PIN; |
693 | GEM_BUG_ON(eb_vma_misplaced(entry, vma, ev->flags)); |
694 | |
695 | return 0; |
696 | } |
697 | |
698 | static bool eb_unbind(struct i915_execbuffer *eb, bool force) |
699 | { |
700 | const unsigned int count = eb->buffer_count; |
701 | unsigned int i; |
702 | struct list_head last; |
703 | bool unpinned = false; |
704 | |
705 | /* Resort *all* the objects into priority order */ |
706 | INIT_LIST_HEAD(list: &eb->unbound); |
707 | INIT_LIST_HEAD(list: &last); |
708 | |
709 | for (i = 0; i < count; i++) { |
710 | struct eb_vma *ev = &eb->vma[i]; |
711 | unsigned int flags = ev->flags; |
712 | |
713 | if (!force && flags & EXEC_OBJECT_PINNED && |
714 | flags & __EXEC_OBJECT_HAS_PIN) |
715 | continue; |
716 | |
717 | unpinned = true; |
718 | eb_unreserve_vma(ev); |
719 | |
720 | if (flags & EXEC_OBJECT_PINNED) |
721 | /* Pinned must have their slot */ |
722 | list_add(new: &ev->bind_link, head: &eb->unbound); |
723 | else if (flags & __EXEC_OBJECT_NEEDS_MAP) |
724 | /* Map require the lowest 256MiB (aperture) */ |
725 | list_add_tail(new: &ev->bind_link, head: &eb->unbound); |
726 | else if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS)) |
727 | /* Prioritise 4GiB region for restricted bo */ |
728 | list_add(new: &ev->bind_link, head: &last); |
729 | else |
730 | list_add_tail(new: &ev->bind_link, head: &last); |
731 | } |
732 | |
733 | list_splice_tail(list: &last, head: &eb->unbound); |
734 | return unpinned; |
735 | } |
736 | |
737 | static int eb_reserve(struct i915_execbuffer *eb) |
738 | { |
739 | struct eb_vma *ev; |
740 | unsigned int pass; |
741 | int err = 0; |
742 | |
743 | /* |
744 | * We have one more buffers that we couldn't bind, which could be due to |
745 | * various reasons. To resolve this we have 4 passes, with every next |
746 | * level turning the screws tighter: |
747 | * |
748 | * 0. Unbind all objects that do not match the GTT constraints for the |
749 | * execbuffer (fenceable, mappable, alignment etc). Bind all new |
750 | * objects. This avoids unnecessary unbinding of later objects in order |
751 | * to make room for the earlier objects *unless* we need to defragment. |
752 | * |
753 | * 1. Reorder the buffers, where objects with the most restrictive |
754 | * placement requirements go first (ignoring fixed location buffers for |
755 | * now). For example, objects needing the mappable aperture (the first |
756 | * 256M of GTT), should go first vs objects that can be placed just |
757 | * about anywhere. Repeat the previous pass. |
758 | * |
759 | * 2. Consider buffers that are pinned at a fixed location. Also try to |
760 | * evict the entire VM this time, leaving only objects that we were |
761 | * unable to lock. Try again to bind the buffers. (still using the new |
762 | * buffer order). |
763 | * |
764 | * 3. We likely have object lock contention for one or more stubborn |
765 | * objects in the VM, for which we need to evict to make forward |
766 | * progress (perhaps we are fighting the shrinker?). When evicting the |
767 | * VM this time around, anything that we can't lock we now track using |
768 | * the busy_bo, using the full lock (after dropping the vm->mutex to |
769 | * prevent deadlocks), instead of trylock. We then continue to evict the |
770 | * VM, this time with the stubborn object locked, which we can now |
771 | * hopefully unbind (if still bound in the VM). Repeat until the VM is |
772 | * evicted. Finally we should be able bind everything. |
773 | */ |
774 | for (pass = 0; pass <= 3; pass++) { |
775 | int pin_flags = PIN_USER | PIN_VALIDATE; |
776 | |
777 | if (pass == 0) |
778 | pin_flags |= PIN_NONBLOCK; |
779 | |
780 | if (pass >= 1) |
781 | eb_unbind(eb, force: pass >= 2); |
782 | |
783 | if (pass == 2) { |
784 | err = mutex_lock_interruptible(&eb->context->vm->mutex); |
785 | if (!err) { |
786 | err = i915_gem_evict_vm(vm: eb->context->vm, ww: &eb->ww, NULL); |
787 | mutex_unlock(lock: &eb->context->vm->mutex); |
788 | } |
789 | if (err) |
790 | return err; |
791 | } |
792 | |
793 | if (pass == 3) { |
794 | retry: |
795 | err = mutex_lock_interruptible(&eb->context->vm->mutex); |
796 | if (!err) { |
797 | struct drm_i915_gem_object *busy_bo = NULL; |
798 | |
799 | err = i915_gem_evict_vm(vm: eb->context->vm, ww: &eb->ww, busy_bo: &busy_bo); |
800 | mutex_unlock(lock: &eb->context->vm->mutex); |
801 | if (err && busy_bo) { |
802 | err = i915_gem_object_lock(obj: busy_bo, ww: &eb->ww); |
803 | i915_gem_object_put(obj: busy_bo); |
804 | if (!err) |
805 | goto retry; |
806 | } |
807 | } |
808 | if (err) |
809 | return err; |
810 | } |
811 | |
812 | list_for_each_entry(ev, &eb->unbound, bind_link) { |
813 | err = eb_reserve_vma(eb, ev, pin_flags); |
814 | if (err) |
815 | break; |
816 | } |
817 | |
818 | if (err != -ENOSPC) |
819 | break; |
820 | } |
821 | |
822 | return err; |
823 | } |
824 | |
825 | static int eb_select_context(struct i915_execbuffer *eb) |
826 | { |
827 | struct i915_gem_context *ctx; |
828 | |
829 | ctx = i915_gem_context_lookup(file_priv: eb->file->driver_priv, id: eb->args->rsvd1); |
830 | if (unlikely(IS_ERR(ctx))) |
831 | return PTR_ERR(ptr: ctx); |
832 | |
833 | eb->gem_context = ctx; |
834 | if (i915_gem_context_has_full_ppgtt(ctx)) |
835 | eb->invalid_flags |= EXEC_OBJECT_NEEDS_GTT; |
836 | |
837 | return 0; |
838 | } |
839 | |
840 | static int __eb_add_lut(struct i915_execbuffer *eb, |
841 | u32 handle, struct i915_vma *vma) |
842 | { |
843 | struct i915_gem_context *ctx = eb->gem_context; |
844 | struct i915_lut_handle *lut; |
845 | int err; |
846 | |
847 | lut = i915_lut_handle_alloc(); |
848 | if (unlikely(!lut)) |
849 | return -ENOMEM; |
850 | |
851 | i915_vma_get(vma); |
852 | if (!atomic_fetch_inc(v: &vma->open_count)) |
853 | i915_vma_reopen(vma); |
854 | lut->handle = handle; |
855 | lut->ctx = ctx; |
856 | |
857 | /* Check that the context hasn't been closed in the meantime */ |
858 | err = -EINTR; |
859 | if (!mutex_lock_interruptible(&ctx->lut_mutex)) { |
860 | if (likely(!i915_gem_context_is_closed(ctx))) |
861 | err = radix_tree_insert(&ctx->handles_vma, index: handle, vma); |
862 | else |
863 | err = -ENOENT; |
864 | if (err == 0) { /* And nor has this handle */ |
865 | struct drm_i915_gem_object *obj = vma->obj; |
866 | |
867 | spin_lock(lock: &obj->lut_lock); |
868 | if (idr_find(&eb->file->object_idr, id: handle) == obj) { |
869 | list_add(new: &lut->obj_link, head: &obj->lut_list); |
870 | } else { |
871 | radix_tree_delete(&ctx->handles_vma, handle); |
872 | err = -ENOENT; |
873 | } |
874 | spin_unlock(lock: &obj->lut_lock); |
875 | } |
876 | mutex_unlock(lock: &ctx->lut_mutex); |
877 | } |
878 | if (unlikely(err)) |
879 | goto err; |
880 | |
881 | return 0; |
882 | |
883 | err: |
884 | i915_vma_close(vma); |
885 | i915_vma_put(vma); |
886 | i915_lut_handle_free(lut); |
887 | return err; |
888 | } |
889 | |
890 | static struct i915_vma *eb_lookup_vma(struct i915_execbuffer *eb, u32 handle) |
891 | { |
892 | struct i915_address_space *vm = eb->context->vm; |
893 | |
894 | do { |
895 | struct drm_i915_gem_object *obj; |
896 | struct i915_vma *vma; |
897 | int err; |
898 | |
899 | rcu_read_lock(); |
900 | vma = radix_tree_lookup(&eb->gem_context->handles_vma, handle); |
901 | if (likely(vma && vma->vm == vm)) |
902 | vma = i915_vma_tryget(vma); |
903 | rcu_read_unlock(); |
904 | if (likely(vma)) |
905 | return vma; |
906 | |
907 | obj = i915_gem_object_lookup(file: eb->file, handle); |
908 | if (unlikely(!obj)) |
909 | return ERR_PTR(error: -ENOENT); |
910 | |
911 | /* |
912 | * If the user has opted-in for protected-object tracking, make |
913 | * sure the object encryption can be used. |
914 | * We only need to do this when the object is first used with |
915 | * this context, because the context itself will be banned when |
916 | * the protected objects become invalid. |
917 | */ |
918 | if (i915_gem_context_uses_protected_content(ctx: eb->gem_context) && |
919 | i915_gem_object_is_protected(obj)) { |
920 | err = intel_pxp_key_check(pxp: eb->i915->pxp, obj, assign: true); |
921 | if (err) { |
922 | i915_gem_object_put(obj); |
923 | return ERR_PTR(error: err); |
924 | } |
925 | } |
926 | |
927 | vma = i915_vma_instance(obj, vm, NULL); |
928 | if (IS_ERR(ptr: vma)) { |
929 | i915_gem_object_put(obj); |
930 | return vma; |
931 | } |
932 | |
933 | err = __eb_add_lut(eb, handle, vma); |
934 | if (likely(!err)) |
935 | return vma; |
936 | |
937 | i915_gem_object_put(obj); |
938 | if (err != -EEXIST) |
939 | return ERR_PTR(error: err); |
940 | } while (1); |
941 | } |
942 | |
943 | static int eb_lookup_vmas(struct i915_execbuffer *eb) |
944 | { |
945 | unsigned int i, current_batch = 0; |
946 | int err = 0; |
947 | |
948 | INIT_LIST_HEAD(list: &eb->relocs); |
949 | |
950 | for (i = 0; i < eb->buffer_count; i++) { |
951 | struct i915_vma *vma; |
952 | |
953 | vma = eb_lookup_vma(eb, handle: eb->exec[i].handle); |
954 | if (IS_ERR(ptr: vma)) { |
955 | err = PTR_ERR(ptr: vma); |
956 | goto err; |
957 | } |
958 | |
959 | err = eb_validate_vma(eb, entry: &eb->exec[i], vma); |
960 | if (unlikely(err)) { |
961 | i915_vma_put(vma); |
962 | goto err; |
963 | } |
964 | |
965 | err = eb_add_vma(eb, current_batch: ¤t_batch, i, vma); |
966 | if (err) |
967 | return err; |
968 | |
969 | if (i915_gem_object_is_userptr(obj: vma->obj)) { |
970 | err = i915_gem_object_userptr_submit_init(obj: vma->obj); |
971 | if (err) { |
972 | if (i + 1 < eb->buffer_count) { |
973 | /* |
974 | * Execbuffer code expects last vma entry to be NULL, |
975 | * since we already initialized this entry, |
976 | * set the next value to NULL or we mess up |
977 | * cleanup handling. |
978 | */ |
979 | eb->vma[i + 1].vma = NULL; |
980 | } |
981 | |
982 | return err; |
983 | } |
984 | |
985 | eb->vma[i].flags |= __EXEC_OBJECT_USERPTR_INIT; |
986 | eb->args->flags |= __EXEC_USERPTR_USED; |
987 | } |
988 | } |
989 | |
990 | return 0; |
991 | |
992 | err: |
993 | eb->vma[i].vma = NULL; |
994 | return err; |
995 | } |
996 | |
997 | static int eb_lock_vmas(struct i915_execbuffer *eb) |
998 | { |
999 | unsigned int i; |
1000 | int err; |
1001 | |
1002 | for (i = 0; i < eb->buffer_count; i++) { |
1003 | struct eb_vma *ev = &eb->vma[i]; |
1004 | struct i915_vma *vma = ev->vma; |
1005 | |
1006 | err = i915_gem_object_lock(obj: vma->obj, ww: &eb->ww); |
1007 | if (err) |
1008 | return err; |
1009 | } |
1010 | |
1011 | return 0; |
1012 | } |
1013 | |
1014 | static int eb_validate_vmas(struct i915_execbuffer *eb) |
1015 | { |
1016 | unsigned int i; |
1017 | int err; |
1018 | |
1019 | INIT_LIST_HEAD(list: &eb->unbound); |
1020 | |
1021 | err = eb_lock_vmas(eb); |
1022 | if (err) |
1023 | return err; |
1024 | |
1025 | for (i = 0; i < eb->buffer_count; i++) { |
1026 | struct drm_i915_gem_exec_object2 *entry = &eb->exec[i]; |
1027 | struct eb_vma *ev = &eb->vma[i]; |
1028 | struct i915_vma *vma = ev->vma; |
1029 | |
1030 | err = eb_pin_vma(eb, entry, ev); |
1031 | if (err == -EDEADLK) |
1032 | return err; |
1033 | |
1034 | if (!err) { |
1035 | if (entry->offset != i915_vma_offset(vma)) { |
1036 | entry->offset = i915_vma_offset(vma) | UPDATE; |
1037 | eb->args->flags |= __EXEC_HAS_RELOC; |
1038 | } |
1039 | } else { |
1040 | eb_unreserve_vma(ev); |
1041 | |
1042 | list_add_tail(new: &ev->bind_link, head: &eb->unbound); |
1043 | if (drm_mm_node_allocated(node: &vma->node)) { |
1044 | err = i915_vma_unbind(vma); |
1045 | if (err) |
1046 | return err; |
1047 | } |
1048 | } |
1049 | |
1050 | /* Reserve enough slots to accommodate composite fences */ |
1051 | err = dma_resv_reserve_fences(obj: vma->obj->base.resv, num_fences: eb->num_batches); |
1052 | if (err) |
1053 | return err; |
1054 | |
1055 | GEM_BUG_ON(drm_mm_node_allocated(&vma->node) && |
1056 | eb_vma_misplaced(&eb->exec[i], vma, ev->flags)); |
1057 | } |
1058 | |
1059 | if (!list_empty(head: &eb->unbound)) |
1060 | return eb_reserve(eb); |
1061 | |
1062 | return 0; |
1063 | } |
1064 | |
1065 | static struct eb_vma * |
1066 | eb_get_vma(const struct i915_execbuffer *eb, unsigned long handle) |
1067 | { |
1068 | if (eb->lut_size < 0) { |
1069 | if (handle >= -eb->lut_size) |
1070 | return NULL; |
1071 | return &eb->vma[handle]; |
1072 | } else { |
1073 | struct hlist_head *head; |
1074 | struct eb_vma *ev; |
1075 | |
1076 | head = &eb->buckets[hash_32(val: handle, bits: eb->lut_size)]; |
1077 | hlist_for_each_entry(ev, head, node) { |
1078 | if (ev->handle == handle) |
1079 | return ev; |
1080 | } |
1081 | return NULL; |
1082 | } |
1083 | } |
1084 | |
1085 | static void eb_release_vmas(struct i915_execbuffer *eb, bool final) |
1086 | { |
1087 | const unsigned int count = eb->buffer_count; |
1088 | unsigned int i; |
1089 | |
1090 | for (i = 0; i < count; i++) { |
1091 | struct eb_vma *ev = &eb->vma[i]; |
1092 | struct i915_vma *vma = ev->vma; |
1093 | |
1094 | if (!vma) |
1095 | break; |
1096 | |
1097 | eb_unreserve_vma(ev); |
1098 | |
1099 | if (final) |
1100 | i915_vma_put(vma); |
1101 | } |
1102 | |
1103 | eb_capture_release(eb); |
1104 | eb_unpin_engine(eb); |
1105 | } |
1106 | |
1107 | static void eb_destroy(const struct i915_execbuffer *eb) |
1108 | { |
1109 | if (eb->lut_size > 0) |
1110 | kfree(objp: eb->buckets); |
1111 | } |
1112 | |
1113 | static u64 |
1114 | relocation_target(const struct drm_i915_gem_relocation_entry *reloc, |
1115 | const struct i915_vma *target) |
1116 | { |
1117 | return gen8_canonical_addr(address: (int)reloc->delta + i915_vma_offset(vma: target)); |
1118 | } |
1119 | |
1120 | static void reloc_cache_init(struct reloc_cache *cache, |
1121 | struct drm_i915_private *i915) |
1122 | { |
1123 | cache->page = -1; |
1124 | cache->vaddr = 0; |
1125 | /* Must be a variable in the struct to allow GCC to unroll. */ |
1126 | cache->graphics_ver = GRAPHICS_VER(i915); |
1127 | cache->has_llc = HAS_LLC(i915); |
1128 | cache->use_64bit_reloc = HAS_64BIT_RELOC(i915); |
1129 | cache->has_fence = cache->graphics_ver < 4; |
1130 | cache->needs_unfenced = INTEL_INFO(i915)->unfenced_needs_alignment; |
1131 | cache->node.flags = 0; |
1132 | } |
1133 | |
1134 | static void *unmask_page(unsigned long p) |
1135 | { |
1136 | return (void *)(uintptr_t)(p & PAGE_MASK); |
1137 | } |
1138 | |
1139 | static unsigned int unmask_flags(unsigned long p) |
1140 | { |
1141 | return p & ~PAGE_MASK; |
1142 | } |
1143 | |
1144 | #define KMAP 0x4 /* after CLFLUSH_FLAGS */ |
1145 | |
1146 | static struct i915_ggtt *cache_to_ggtt(struct reloc_cache *cache) |
1147 | { |
1148 | struct drm_i915_private *i915 = |
1149 | container_of(cache, struct i915_execbuffer, reloc_cache)->i915; |
1150 | return to_gt(i915)->ggtt; |
1151 | } |
1152 | |
1153 | static void reloc_cache_unmap(struct reloc_cache *cache) |
1154 | { |
1155 | void *vaddr; |
1156 | |
1157 | if (!cache->vaddr) |
1158 | return; |
1159 | |
1160 | vaddr = unmask_page(p: cache->vaddr); |
1161 | if (cache->vaddr & KMAP) |
1162 | kunmap_local(vaddr); |
1163 | else |
1164 | io_mapping_unmap_atomic(vaddr: (void __iomem *)vaddr); |
1165 | } |
1166 | |
1167 | static void reloc_cache_remap(struct reloc_cache *cache, |
1168 | struct drm_i915_gem_object *obj) |
1169 | { |
1170 | void *vaddr; |
1171 | |
1172 | if (!cache->vaddr) |
1173 | return; |
1174 | |
1175 | if (cache->vaddr & KMAP) { |
1176 | struct page *page = i915_gem_object_get_page(obj, cache->page); |
1177 | |
1178 | vaddr = kmap_local_page(page); |
1179 | cache->vaddr = unmask_flags(p: cache->vaddr) | |
1180 | (unsigned long)vaddr; |
1181 | } else { |
1182 | struct i915_ggtt *ggtt = cache_to_ggtt(cache); |
1183 | unsigned long offset; |
1184 | |
1185 | offset = cache->node.start; |
1186 | if (!drm_mm_node_allocated(node: &cache->node)) |
1187 | offset += cache->page << PAGE_SHIFT; |
1188 | |
1189 | cache->vaddr = (unsigned long) |
1190 | io_mapping_map_atomic_wc(mapping: &ggtt->iomap, offset); |
1191 | } |
1192 | } |
1193 | |
1194 | static void reloc_cache_reset(struct reloc_cache *cache, struct i915_execbuffer *eb) |
1195 | { |
1196 | void *vaddr; |
1197 | |
1198 | if (!cache->vaddr) |
1199 | return; |
1200 | |
1201 | vaddr = unmask_page(p: cache->vaddr); |
1202 | if (cache->vaddr & KMAP) { |
1203 | struct drm_i915_gem_object *obj = |
1204 | (struct drm_i915_gem_object *)cache->node.mm; |
1205 | if (cache->vaddr & CLFLUSH_AFTER) |
1206 | mb(); |
1207 | |
1208 | kunmap_local(vaddr); |
1209 | i915_gem_object_finish_access(obj); |
1210 | } else { |
1211 | struct i915_ggtt *ggtt = cache_to_ggtt(cache); |
1212 | |
1213 | intel_gt_flush_ggtt_writes(gt: ggtt->vm.gt); |
1214 | io_mapping_unmap_atomic(vaddr: (void __iomem *)vaddr); |
1215 | |
1216 | if (drm_mm_node_allocated(node: &cache->node)) { |
1217 | ggtt->vm.clear_range(&ggtt->vm, |
1218 | cache->node.start, |
1219 | cache->node.size); |
1220 | mutex_lock(&ggtt->vm.mutex); |
1221 | drm_mm_remove_node(node: &cache->node); |
1222 | mutex_unlock(lock: &ggtt->vm.mutex); |
1223 | } else { |
1224 | i915_vma_unpin(vma: (struct i915_vma *)cache->node.mm); |
1225 | } |
1226 | } |
1227 | |
1228 | cache->vaddr = 0; |
1229 | cache->page = -1; |
1230 | } |
1231 | |
1232 | static void *reloc_kmap(struct drm_i915_gem_object *obj, |
1233 | struct reloc_cache *cache, |
1234 | unsigned long pageno) |
1235 | { |
1236 | void *vaddr; |
1237 | struct page *page; |
1238 | |
1239 | if (cache->vaddr) { |
1240 | kunmap_local(unmask_page(cache->vaddr)); |
1241 | } else { |
1242 | unsigned int flushes; |
1243 | int err; |
1244 | |
1245 | err = i915_gem_object_prepare_write(obj, needs_clflush: &flushes); |
1246 | if (err) |
1247 | return ERR_PTR(error: err); |
1248 | |
1249 | BUILD_BUG_ON(KMAP & CLFLUSH_FLAGS); |
1250 | BUILD_BUG_ON((KMAP | CLFLUSH_FLAGS) & PAGE_MASK); |
1251 | |
1252 | cache->vaddr = flushes | KMAP; |
1253 | cache->node.mm = (void *)obj; |
1254 | if (flushes) |
1255 | mb(); |
1256 | } |
1257 | |
1258 | page = i915_gem_object_get_page(obj, pageno); |
1259 | if (!obj->mm.dirty) |
1260 | set_page_dirty(page); |
1261 | |
1262 | vaddr = kmap_local_page(page); |
1263 | cache->vaddr = unmask_flags(p: cache->vaddr) | (unsigned long)vaddr; |
1264 | cache->page = pageno; |
1265 | |
1266 | return vaddr; |
1267 | } |
1268 | |
1269 | static void *reloc_iomap(struct i915_vma *batch, |
1270 | struct i915_execbuffer *eb, |
1271 | unsigned long page) |
1272 | { |
1273 | struct drm_i915_gem_object *obj = batch->obj; |
1274 | struct reloc_cache *cache = &eb->reloc_cache; |
1275 | struct i915_ggtt *ggtt = cache_to_ggtt(cache); |
1276 | unsigned long offset; |
1277 | void *vaddr; |
1278 | |
1279 | if (cache->vaddr) { |
1280 | intel_gt_flush_ggtt_writes(gt: ggtt->vm.gt); |
1281 | io_mapping_unmap_atomic(vaddr: (void __force __iomem *) unmask_page(p: cache->vaddr)); |
1282 | } else { |
1283 | struct i915_vma *vma = ERR_PTR(error: -ENODEV); |
1284 | int err; |
1285 | |
1286 | if (i915_gem_object_is_tiled(obj)) |
1287 | return ERR_PTR(error: -EINVAL); |
1288 | |
1289 | if (use_cpu_reloc(cache, obj)) |
1290 | return NULL; |
1291 | |
1292 | err = i915_gem_object_set_to_gtt_domain(obj, write: true); |
1293 | if (err) |
1294 | return ERR_PTR(error: err); |
1295 | |
1296 | /* |
1297 | * i915_gem_object_ggtt_pin_ww may attempt to remove the batch |
1298 | * VMA from the object list because we no longer pin. |
1299 | * |
1300 | * Only attempt to pin the batch buffer to ggtt if the current batch |
1301 | * is not inside ggtt, or the batch buffer is not misplaced. |
1302 | */ |
1303 | if (!i915_is_ggtt(batch->vm) || |
1304 | !i915_vma_misplaced(vma: batch, size: 0, alignment: 0, PIN_MAPPABLE)) { |
1305 | vma = i915_gem_object_ggtt_pin_ww(obj, ww: &eb->ww, NULL, size: 0, alignment: 0, |
1306 | PIN_MAPPABLE | |
1307 | PIN_NONBLOCK /* NOWARN */ | |
1308 | PIN_NOEVICT); |
1309 | } |
1310 | |
1311 | if (vma == ERR_PTR(error: -EDEADLK)) |
1312 | return vma; |
1313 | |
1314 | if (IS_ERR(ptr: vma)) { |
1315 | memset(&cache->node, 0, sizeof(cache->node)); |
1316 | mutex_lock(&ggtt->vm.mutex); |
1317 | err = drm_mm_insert_node_in_range |
1318 | (mm: &ggtt->vm.mm, node: &cache->node, |
1319 | PAGE_SIZE, alignment: 0, I915_COLOR_UNEVICTABLE, |
1320 | start: 0, end: ggtt->mappable_end, |
1321 | mode: DRM_MM_INSERT_LOW); |
1322 | mutex_unlock(lock: &ggtt->vm.mutex); |
1323 | if (err) /* no inactive aperture space, use cpu reloc */ |
1324 | return NULL; |
1325 | } else { |
1326 | cache->node.start = i915_ggtt_offset(vma); |
1327 | cache->node.mm = (void *)vma; |
1328 | } |
1329 | } |
1330 | |
1331 | offset = cache->node.start; |
1332 | if (drm_mm_node_allocated(node: &cache->node)) { |
1333 | ggtt->vm.insert_page(&ggtt->vm, |
1334 | i915_gem_object_get_dma_address(obj, page), |
1335 | offset, |
1336 | i915_gem_get_pat_index(i915: ggtt->vm.i915, |
1337 | level: I915_CACHE_NONE), |
1338 | 0); |
1339 | } else { |
1340 | offset += page << PAGE_SHIFT; |
1341 | } |
1342 | |
1343 | vaddr = (void __force *)io_mapping_map_atomic_wc(mapping: &ggtt->iomap, |
1344 | offset); |
1345 | cache->page = page; |
1346 | cache->vaddr = (unsigned long)vaddr; |
1347 | |
1348 | return vaddr; |
1349 | } |
1350 | |
1351 | static void *reloc_vaddr(struct i915_vma *vma, |
1352 | struct i915_execbuffer *eb, |
1353 | unsigned long page) |
1354 | { |
1355 | struct reloc_cache *cache = &eb->reloc_cache; |
1356 | void *vaddr; |
1357 | |
1358 | if (cache->page == page) { |
1359 | vaddr = unmask_page(p: cache->vaddr); |
1360 | } else { |
1361 | vaddr = NULL; |
1362 | if ((cache->vaddr & KMAP) == 0) |
1363 | vaddr = reloc_iomap(batch: vma, eb, page); |
1364 | if (!vaddr) |
1365 | vaddr = reloc_kmap(obj: vma->obj, cache, pageno: page); |
1366 | } |
1367 | |
1368 | return vaddr; |
1369 | } |
1370 | |
1371 | static void clflush_write32(u32 *addr, u32 value, unsigned int flushes) |
1372 | { |
1373 | if (unlikely(flushes & (CLFLUSH_BEFORE | CLFLUSH_AFTER))) { |
1374 | if (flushes & CLFLUSH_BEFORE) |
1375 | drm_clflush_virt_range(addr, length: sizeof(*addr)); |
1376 | |
1377 | *addr = value; |
1378 | |
1379 | /* |
1380 | * Writes to the same cacheline are serialised by the CPU |
1381 | * (including clflush). On the write path, we only require |
1382 | * that it hits memory in an orderly fashion and place |
1383 | * mb barriers at the start and end of the relocation phase |
1384 | * to ensure ordering of clflush wrt to the system. |
1385 | */ |
1386 | if (flushes & CLFLUSH_AFTER) |
1387 | drm_clflush_virt_range(addr, length: sizeof(*addr)); |
1388 | } else |
1389 | *addr = value; |
1390 | } |
1391 | |
1392 | static u64 |
1393 | relocate_entry(struct i915_vma *vma, |
1394 | const struct drm_i915_gem_relocation_entry *reloc, |
1395 | struct i915_execbuffer *eb, |
1396 | const struct i915_vma *target) |
1397 | { |
1398 | u64 target_addr = relocation_target(reloc, target); |
1399 | u64 offset = reloc->offset; |
1400 | bool wide = eb->reloc_cache.use_64bit_reloc; |
1401 | void *vaddr; |
1402 | |
1403 | repeat: |
1404 | vaddr = reloc_vaddr(vma, eb, |
1405 | page: offset >> PAGE_SHIFT); |
1406 | if (IS_ERR(ptr: vaddr)) |
1407 | return PTR_ERR(ptr: vaddr); |
1408 | |
1409 | GEM_BUG_ON(!IS_ALIGNED(offset, sizeof(u32))); |
1410 | clflush_write32(addr: vaddr + offset_in_page(offset), |
1411 | lower_32_bits(target_addr), |
1412 | flushes: eb->reloc_cache.vaddr); |
1413 | |
1414 | if (wide) { |
1415 | offset += sizeof(u32); |
1416 | target_addr >>= 32; |
1417 | wide = false; |
1418 | goto repeat; |
1419 | } |
1420 | |
1421 | return target->node.start | UPDATE; |
1422 | } |
1423 | |
1424 | static u64 |
1425 | eb_relocate_entry(struct i915_execbuffer *eb, |
1426 | struct eb_vma *ev, |
1427 | const struct drm_i915_gem_relocation_entry *reloc) |
1428 | { |
1429 | struct drm_i915_private *i915 = eb->i915; |
1430 | struct eb_vma *target; |
1431 | int err; |
1432 | |
1433 | /* we've already hold a reference to all valid objects */ |
1434 | target = eb_get_vma(eb, handle: reloc->target_handle); |
1435 | if (unlikely(!target)) |
1436 | return -ENOENT; |
1437 | |
1438 | /* Validate that the target is in a valid r/w GPU domain */ |
1439 | if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) { |
1440 | drm_dbg(&i915->drm, "reloc with multiple write domains: " |
1441 | "target %d offset %d " |
1442 | "read %08x write %08x\n" , |
1443 | reloc->target_handle, |
1444 | (int) reloc->offset, |
1445 | reloc->read_domains, |
1446 | reloc->write_domain); |
1447 | return -EINVAL; |
1448 | } |
1449 | if (unlikely((reloc->write_domain | reloc->read_domains) |
1450 | & ~I915_GEM_GPU_DOMAINS)) { |
1451 | drm_dbg(&i915->drm, "reloc with read/write non-GPU domains: " |
1452 | "target %d offset %d " |
1453 | "read %08x write %08x\n" , |
1454 | reloc->target_handle, |
1455 | (int) reloc->offset, |
1456 | reloc->read_domains, |
1457 | reloc->write_domain); |
1458 | return -EINVAL; |
1459 | } |
1460 | |
1461 | if (reloc->write_domain) { |
1462 | target->flags |= EXEC_OBJECT_WRITE; |
1463 | |
1464 | /* |
1465 | * Sandybridge PPGTT errata: We need a global gtt mapping |
1466 | * for MI and pipe_control writes because the gpu doesn't |
1467 | * properly redirect them through the ppgtt for non_secure |
1468 | * batchbuffers. |
1469 | */ |
1470 | if (reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION && |
1471 | GRAPHICS_VER(eb->i915) == 6 && |
1472 | !i915_vma_is_bound(vma: target->vma, I915_VMA_GLOBAL_BIND)) { |
1473 | struct i915_vma *vma = target->vma; |
1474 | |
1475 | reloc_cache_unmap(cache: &eb->reloc_cache); |
1476 | mutex_lock(&vma->vm->mutex); |
1477 | err = i915_vma_bind(vma: target->vma, |
1478 | pat_index: target->vma->obj->pat_index, |
1479 | PIN_GLOBAL, NULL, NULL); |
1480 | mutex_unlock(lock: &vma->vm->mutex); |
1481 | reloc_cache_remap(cache: &eb->reloc_cache, obj: ev->vma->obj); |
1482 | if (err) |
1483 | return err; |
1484 | } |
1485 | } |
1486 | |
1487 | /* |
1488 | * If the relocation already has the right value in it, no |
1489 | * more work needs to be done. |
1490 | */ |
1491 | if (!DBG_FORCE_RELOC && |
1492 | gen8_canonical_addr(address: i915_vma_offset(vma: target->vma)) == reloc->presumed_offset) |
1493 | return 0; |
1494 | |
1495 | /* Check that the relocation address is valid... */ |
1496 | if (unlikely(reloc->offset > |
1497 | ev->vma->size - (eb->reloc_cache.use_64bit_reloc ? 8 : 4))) { |
1498 | drm_dbg(&i915->drm, "Relocation beyond object bounds: " |
1499 | "target %d offset %d size %d.\n" , |
1500 | reloc->target_handle, |
1501 | (int)reloc->offset, |
1502 | (int)ev->vma->size); |
1503 | return -EINVAL; |
1504 | } |
1505 | if (unlikely(reloc->offset & 3)) { |
1506 | drm_dbg(&i915->drm, "Relocation not 4-byte aligned: " |
1507 | "target %d offset %d.\n" , |
1508 | reloc->target_handle, |
1509 | (int)reloc->offset); |
1510 | return -EINVAL; |
1511 | } |
1512 | |
1513 | /* |
1514 | * If we write into the object, we need to force the synchronisation |
1515 | * barrier, either with an asynchronous clflush or if we executed the |
1516 | * patching using the GPU (though that should be serialised by the |
1517 | * timeline). To be completely sure, and since we are required to |
1518 | * do relocations we are already stalling, disable the user's opt |
1519 | * out of our synchronisation. |
1520 | */ |
1521 | ev->flags &= ~EXEC_OBJECT_ASYNC; |
1522 | |
1523 | /* and update the user's relocation entry */ |
1524 | return relocate_entry(vma: ev->vma, reloc, eb, target: target->vma); |
1525 | } |
1526 | |
1527 | static int eb_relocate_vma(struct i915_execbuffer *eb, struct eb_vma *ev) |
1528 | { |
1529 | #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry)) |
1530 | struct drm_i915_gem_relocation_entry stack[N_RELOC(512)]; |
1531 | const struct drm_i915_gem_exec_object2 *entry = ev->exec; |
1532 | struct drm_i915_gem_relocation_entry __user *urelocs = |
1533 | u64_to_user_ptr(entry->relocs_ptr); |
1534 | unsigned long remain = entry->relocation_count; |
1535 | |
1536 | if (unlikely(remain > N_RELOC(ULONG_MAX))) |
1537 | return -EINVAL; |
1538 | |
1539 | /* |
1540 | * We must check that the entire relocation array is safe |
1541 | * to read. However, if the array is not writable the user loses |
1542 | * the updated relocation values. |
1543 | */ |
1544 | if (unlikely(!access_ok(urelocs, remain * sizeof(*urelocs)))) |
1545 | return -EFAULT; |
1546 | |
1547 | do { |
1548 | struct drm_i915_gem_relocation_entry *r = stack; |
1549 | unsigned int count = |
1550 | min_t(unsigned long, remain, ARRAY_SIZE(stack)); |
1551 | unsigned int copied; |
1552 | |
1553 | /* |
1554 | * This is the fast path and we cannot handle a pagefault |
1555 | * whilst holding the struct mutex lest the user pass in the |
1556 | * relocations contained within a mmaped bo. For in such a case |
1557 | * we, the page fault handler would call i915_gem_fault() and |
1558 | * we would try to acquire the struct mutex again. Obviously |
1559 | * this is bad and so lockdep complains vehemently. |
1560 | */ |
1561 | pagefault_disable(); |
1562 | copied = __copy_from_user_inatomic(to: r, from: urelocs, n: count * sizeof(r[0])); |
1563 | pagefault_enable(); |
1564 | if (unlikely(copied)) { |
1565 | remain = -EFAULT; |
1566 | goto out; |
1567 | } |
1568 | |
1569 | remain -= count; |
1570 | do { |
1571 | u64 offset = eb_relocate_entry(eb, ev, reloc: r); |
1572 | |
1573 | if (likely(offset == 0)) { |
1574 | } else if ((s64)offset < 0) { |
1575 | remain = (int)offset; |
1576 | goto out; |
1577 | } else { |
1578 | /* |
1579 | * Note that reporting an error now |
1580 | * leaves everything in an inconsistent |
1581 | * state as we have *already* changed |
1582 | * the relocation value inside the |
1583 | * object. As we have not changed the |
1584 | * reloc.presumed_offset or will not |
1585 | * change the execobject.offset, on the |
1586 | * call we may not rewrite the value |
1587 | * inside the object, leaving it |
1588 | * dangling and causing a GPU hang. Unless |
1589 | * userspace dynamically rebuilds the |
1590 | * relocations on each execbuf rather than |
1591 | * presume a static tree. |
1592 | * |
1593 | * We did previously check if the relocations |
1594 | * were writable (access_ok), an error now |
1595 | * would be a strange race with mprotect, |
1596 | * having already demonstrated that we |
1597 | * can read from this userspace address. |
1598 | */ |
1599 | offset = gen8_canonical_addr(address: offset & ~UPDATE); |
1600 | __put_user(offset, |
1601 | &urelocs[r - stack].presumed_offset); |
1602 | } |
1603 | } while (r++, --count); |
1604 | urelocs += ARRAY_SIZE(stack); |
1605 | } while (remain); |
1606 | out: |
1607 | reloc_cache_reset(cache: &eb->reloc_cache, eb); |
1608 | return remain; |
1609 | } |
1610 | |
1611 | static int |
1612 | eb_relocate_vma_slow(struct i915_execbuffer *eb, struct eb_vma *ev) |
1613 | { |
1614 | const struct drm_i915_gem_exec_object2 *entry = ev->exec; |
1615 | struct drm_i915_gem_relocation_entry *relocs = |
1616 | u64_to_ptr(typeof(*relocs), entry->relocs_ptr); |
1617 | unsigned int i; |
1618 | int err; |
1619 | |
1620 | for (i = 0; i < entry->relocation_count; i++) { |
1621 | u64 offset = eb_relocate_entry(eb, ev, reloc: &relocs[i]); |
1622 | |
1623 | if ((s64)offset < 0) { |
1624 | err = (int)offset; |
1625 | goto err; |
1626 | } |
1627 | } |
1628 | err = 0; |
1629 | err: |
1630 | reloc_cache_reset(cache: &eb->reloc_cache, eb); |
1631 | return err; |
1632 | } |
1633 | |
1634 | static int check_relocations(const struct drm_i915_gem_exec_object2 *entry) |
1635 | { |
1636 | const char __user *addr, *end; |
1637 | unsigned long size; |
1638 | char __maybe_unused c; |
1639 | |
1640 | size = entry->relocation_count; |
1641 | if (size == 0) |
1642 | return 0; |
1643 | |
1644 | if (size > N_RELOC(ULONG_MAX)) |
1645 | return -EINVAL; |
1646 | |
1647 | addr = u64_to_user_ptr(entry->relocs_ptr); |
1648 | size *= sizeof(struct drm_i915_gem_relocation_entry); |
1649 | if (!access_ok(addr, size)) |
1650 | return -EFAULT; |
1651 | |
1652 | end = addr + size; |
1653 | for (; addr < end; addr += PAGE_SIZE) { |
1654 | int err = __get_user(c, addr); |
1655 | if (err) |
1656 | return err; |
1657 | } |
1658 | return __get_user(c, end - 1); |
1659 | } |
1660 | |
1661 | static int eb_copy_relocations(const struct i915_execbuffer *eb) |
1662 | { |
1663 | struct drm_i915_gem_relocation_entry *relocs; |
1664 | const unsigned int count = eb->buffer_count; |
1665 | unsigned int i; |
1666 | int err; |
1667 | |
1668 | for (i = 0; i < count; i++) { |
1669 | const unsigned int nreloc = eb->exec[i].relocation_count; |
1670 | struct drm_i915_gem_relocation_entry __user *urelocs; |
1671 | unsigned long size; |
1672 | unsigned long copied; |
1673 | |
1674 | if (nreloc == 0) |
1675 | continue; |
1676 | |
1677 | err = check_relocations(entry: &eb->exec[i]); |
1678 | if (err) |
1679 | goto err; |
1680 | |
1681 | urelocs = u64_to_user_ptr(eb->exec[i].relocs_ptr); |
1682 | size = nreloc * sizeof(*relocs); |
1683 | |
1684 | relocs = kvmalloc_array(n: 1, size, GFP_KERNEL); |
1685 | if (!relocs) { |
1686 | err = -ENOMEM; |
1687 | goto err; |
1688 | } |
1689 | |
1690 | /* copy_from_user is limited to < 4GiB */ |
1691 | copied = 0; |
1692 | do { |
1693 | unsigned int len = |
1694 | min_t(u64, BIT_ULL(31), size - copied); |
1695 | |
1696 | if (__copy_from_user(to: (char *)relocs + copied, |
1697 | from: (char __user *)urelocs + copied, |
1698 | n: len)) |
1699 | goto end; |
1700 | |
1701 | copied += len; |
1702 | } while (copied < size); |
1703 | |
1704 | /* |
1705 | * As we do not update the known relocation offsets after |
1706 | * relocating (due to the complexities in lock handling), |
1707 | * we need to mark them as invalid now so that we force the |
1708 | * relocation processing next time. Just in case the target |
1709 | * object is evicted and then rebound into its old |
1710 | * presumed_offset before the next execbuffer - if that |
1711 | * happened we would make the mistake of assuming that the |
1712 | * relocations were valid. |
1713 | */ |
1714 | if (!user_access_begin(urelocs, size)) |
1715 | goto end; |
1716 | |
1717 | for (copied = 0; copied < nreloc; copied++) |
1718 | unsafe_put_user(-1, |
1719 | &urelocs[copied].presumed_offset, |
1720 | end_user); |
1721 | user_access_end(); |
1722 | |
1723 | eb->exec[i].relocs_ptr = (uintptr_t)relocs; |
1724 | } |
1725 | |
1726 | return 0; |
1727 | |
1728 | end_user: |
1729 | user_access_end(); |
1730 | end: |
1731 | kvfree(addr: relocs); |
1732 | err = -EFAULT; |
1733 | err: |
1734 | while (i--) { |
1735 | relocs = u64_to_ptr(typeof(*relocs), eb->exec[i].relocs_ptr); |
1736 | if (eb->exec[i].relocation_count) |
1737 | kvfree(addr: relocs); |
1738 | } |
1739 | return err; |
1740 | } |
1741 | |
1742 | static int eb_prefault_relocations(const struct i915_execbuffer *eb) |
1743 | { |
1744 | const unsigned int count = eb->buffer_count; |
1745 | unsigned int i; |
1746 | |
1747 | for (i = 0; i < count; i++) { |
1748 | int err; |
1749 | |
1750 | err = check_relocations(entry: &eb->exec[i]); |
1751 | if (err) |
1752 | return err; |
1753 | } |
1754 | |
1755 | return 0; |
1756 | } |
1757 | |
1758 | static int eb_reinit_userptr(struct i915_execbuffer *eb) |
1759 | { |
1760 | const unsigned int count = eb->buffer_count; |
1761 | unsigned int i; |
1762 | int ret; |
1763 | |
1764 | if (likely(!(eb->args->flags & __EXEC_USERPTR_USED))) |
1765 | return 0; |
1766 | |
1767 | for (i = 0; i < count; i++) { |
1768 | struct eb_vma *ev = &eb->vma[i]; |
1769 | |
1770 | if (!i915_gem_object_is_userptr(obj: ev->vma->obj)) |
1771 | continue; |
1772 | |
1773 | ret = i915_gem_object_userptr_submit_init(obj: ev->vma->obj); |
1774 | if (ret) |
1775 | return ret; |
1776 | |
1777 | ev->flags |= __EXEC_OBJECT_USERPTR_INIT; |
1778 | } |
1779 | |
1780 | return 0; |
1781 | } |
1782 | |
1783 | static noinline int eb_relocate_parse_slow(struct i915_execbuffer *eb) |
1784 | { |
1785 | bool have_copy = false; |
1786 | struct eb_vma *ev; |
1787 | int err = 0; |
1788 | |
1789 | repeat: |
1790 | if (signal_pending(current)) { |
1791 | err = -ERESTARTSYS; |
1792 | goto out; |
1793 | } |
1794 | |
1795 | /* We may process another execbuffer during the unlock... */ |
1796 | eb_release_vmas(eb, final: false); |
1797 | i915_gem_ww_ctx_fini(ctx: &eb->ww); |
1798 | |
1799 | /* |
1800 | * We take 3 passes through the slowpatch. |
1801 | * |
1802 | * 1 - we try to just prefault all the user relocation entries and |
1803 | * then attempt to reuse the atomic pagefault disabled fast path again. |
1804 | * |
1805 | * 2 - we copy the user entries to a local buffer here outside of the |
1806 | * local and allow ourselves to wait upon any rendering before |
1807 | * relocations |
1808 | * |
1809 | * 3 - we already have a local copy of the relocation entries, but |
1810 | * were interrupted (EAGAIN) whilst waiting for the objects, try again. |
1811 | */ |
1812 | if (!err) { |
1813 | err = eb_prefault_relocations(eb); |
1814 | } else if (!have_copy) { |
1815 | err = eb_copy_relocations(eb); |
1816 | have_copy = err == 0; |
1817 | } else { |
1818 | cond_resched(); |
1819 | err = 0; |
1820 | } |
1821 | |
1822 | if (!err) |
1823 | err = eb_reinit_userptr(eb); |
1824 | |
1825 | i915_gem_ww_ctx_init(ctx: &eb->ww, intr: true); |
1826 | if (err) |
1827 | goto out; |
1828 | |
1829 | /* reacquire the objects */ |
1830 | repeat_validate: |
1831 | err = eb_pin_engine(eb, throttle: false); |
1832 | if (err) |
1833 | goto err; |
1834 | |
1835 | err = eb_validate_vmas(eb); |
1836 | if (err) |
1837 | goto err; |
1838 | |
1839 | GEM_BUG_ON(!eb->batches[0]); |
1840 | |
1841 | list_for_each_entry(ev, &eb->relocs, reloc_link) { |
1842 | if (!have_copy) { |
1843 | err = eb_relocate_vma(eb, ev); |
1844 | if (err) |
1845 | break; |
1846 | } else { |
1847 | err = eb_relocate_vma_slow(eb, ev); |
1848 | if (err) |
1849 | break; |
1850 | } |
1851 | } |
1852 | |
1853 | if (err == -EDEADLK) |
1854 | goto err; |
1855 | |
1856 | if (err && !have_copy) |
1857 | goto repeat; |
1858 | |
1859 | if (err) |
1860 | goto err; |
1861 | |
1862 | /* as last step, parse the command buffer */ |
1863 | err = eb_parse(eb); |
1864 | if (err) |
1865 | goto err; |
1866 | |
1867 | /* |
1868 | * Leave the user relocations as are, this is the painfully slow path, |
1869 | * and we want to avoid the complication of dropping the lock whilst |
1870 | * having buffers reserved in the aperture and so causing spurious |
1871 | * ENOSPC for random operations. |
1872 | */ |
1873 | |
1874 | err: |
1875 | if (err == -EDEADLK) { |
1876 | eb_release_vmas(eb, final: false); |
1877 | err = i915_gem_ww_ctx_backoff(ctx: &eb->ww); |
1878 | if (!err) |
1879 | goto repeat_validate; |
1880 | } |
1881 | |
1882 | if (err == -EAGAIN) |
1883 | goto repeat; |
1884 | |
1885 | out: |
1886 | if (have_copy) { |
1887 | const unsigned int count = eb->buffer_count; |
1888 | unsigned int i; |
1889 | |
1890 | for (i = 0; i < count; i++) { |
1891 | const struct drm_i915_gem_exec_object2 *entry = |
1892 | &eb->exec[i]; |
1893 | struct drm_i915_gem_relocation_entry *relocs; |
1894 | |
1895 | if (!entry->relocation_count) |
1896 | continue; |
1897 | |
1898 | relocs = u64_to_ptr(typeof(*relocs), entry->relocs_ptr); |
1899 | kvfree(addr: relocs); |
1900 | } |
1901 | } |
1902 | |
1903 | return err; |
1904 | } |
1905 | |
1906 | static int eb_relocate_parse(struct i915_execbuffer *eb) |
1907 | { |
1908 | int err; |
1909 | bool throttle = true; |
1910 | |
1911 | retry: |
1912 | err = eb_pin_engine(eb, throttle); |
1913 | if (err) { |
1914 | if (err != -EDEADLK) |
1915 | return err; |
1916 | |
1917 | goto err; |
1918 | } |
1919 | |
1920 | /* only throttle once, even if we didn't need to throttle */ |
1921 | throttle = false; |
1922 | |
1923 | err = eb_validate_vmas(eb); |
1924 | if (err == -EAGAIN) |
1925 | goto slow; |
1926 | else if (err) |
1927 | goto err; |
1928 | |
1929 | /* The objects are in their final locations, apply the relocations. */ |
1930 | if (eb->args->flags & __EXEC_HAS_RELOC) { |
1931 | struct eb_vma *ev; |
1932 | |
1933 | list_for_each_entry(ev, &eb->relocs, reloc_link) { |
1934 | err = eb_relocate_vma(eb, ev); |
1935 | if (err) |
1936 | break; |
1937 | } |
1938 | |
1939 | if (err == -EDEADLK) |
1940 | goto err; |
1941 | else if (err) |
1942 | goto slow; |
1943 | } |
1944 | |
1945 | if (!err) |
1946 | err = eb_parse(eb); |
1947 | |
1948 | err: |
1949 | if (err == -EDEADLK) { |
1950 | eb_release_vmas(eb, final: false); |
1951 | err = i915_gem_ww_ctx_backoff(ctx: &eb->ww); |
1952 | if (!err) |
1953 | goto retry; |
1954 | } |
1955 | |
1956 | return err; |
1957 | |
1958 | slow: |
1959 | err = eb_relocate_parse_slow(eb); |
1960 | if (err) |
1961 | /* |
1962 | * If the user expects the execobject.offset and |
1963 | * reloc.presumed_offset to be an exact match, |
1964 | * as for using NO_RELOC, then we cannot update |
1965 | * the execobject.offset until we have completed |
1966 | * relocation. |
1967 | */ |
1968 | eb->args->flags &= ~__EXEC_HAS_RELOC; |
1969 | |
1970 | return err; |
1971 | } |
1972 | |
1973 | /* |
1974 | * Using two helper loops for the order of which requests / batches are created |
1975 | * and added the to backend. Requests are created in order from the parent to |
1976 | * the last child. Requests are added in the reverse order, from the last child |
1977 | * to parent. This is done for locking reasons as the timeline lock is acquired |
1978 | * during request creation and released when the request is added to the |
1979 | * backend. To make lockdep happy (see intel_context_timeline_lock) this must be |
1980 | * the ordering. |
1981 | */ |
1982 | #define for_each_batch_create_order(_eb, _i) \ |
1983 | for ((_i) = 0; (_i) < (_eb)->num_batches; ++(_i)) |
1984 | #define for_each_batch_add_order(_eb, _i) \ |
1985 | BUILD_BUG_ON(!typecheck(int, _i)); \ |
1986 | for ((_i) = (_eb)->num_batches - 1; (_i) >= 0; --(_i)) |
1987 | |
1988 | static struct i915_request * |
1989 | eb_find_first_request_added(struct i915_execbuffer *eb) |
1990 | { |
1991 | int i; |
1992 | |
1993 | for_each_batch_add_order(eb, i) |
1994 | if (eb->requests[i]) |
1995 | return eb->requests[i]; |
1996 | |
1997 | GEM_BUG_ON("Request not found" ); |
1998 | |
1999 | return NULL; |
2000 | } |
2001 | |
2002 | #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR) |
2003 | |
2004 | /* Stage with GFP_KERNEL allocations before we enter the signaling critical path */ |
2005 | static int eb_capture_stage(struct i915_execbuffer *eb) |
2006 | { |
2007 | const unsigned int count = eb->buffer_count; |
2008 | unsigned int i = count, j; |
2009 | |
2010 | while (i--) { |
2011 | struct eb_vma *ev = &eb->vma[i]; |
2012 | struct i915_vma *vma = ev->vma; |
2013 | unsigned int flags = ev->flags; |
2014 | |
2015 | if (!(flags & EXEC_OBJECT_CAPTURE)) |
2016 | continue; |
2017 | |
2018 | if (i915_gem_context_is_recoverable(ctx: eb->gem_context) && |
2019 | (IS_DGFX(eb->i915) || GRAPHICS_VER_FULL(eb->i915) > IP_VER(12, 0))) |
2020 | return -EINVAL; |
2021 | |
2022 | for_each_batch_create_order(eb, j) { |
2023 | struct i915_capture_list *capture; |
2024 | |
2025 | capture = kmalloc(size: sizeof(*capture), GFP_KERNEL); |
2026 | if (!capture) |
2027 | continue; |
2028 | |
2029 | capture->next = eb->capture_lists[j]; |
2030 | capture->vma_res = i915_vma_resource_get(vma_res: vma->resource); |
2031 | eb->capture_lists[j] = capture; |
2032 | } |
2033 | } |
2034 | |
2035 | return 0; |
2036 | } |
2037 | |
2038 | /* Commit once we're in the critical path */ |
2039 | static void eb_capture_commit(struct i915_execbuffer *eb) |
2040 | { |
2041 | unsigned int j; |
2042 | |
2043 | for_each_batch_create_order(eb, j) { |
2044 | struct i915_request *rq = eb->requests[j]; |
2045 | |
2046 | if (!rq) |
2047 | break; |
2048 | |
2049 | rq->capture_list = eb->capture_lists[j]; |
2050 | eb->capture_lists[j] = NULL; |
2051 | } |
2052 | } |
2053 | |
2054 | /* |
2055 | * Release anything that didn't get committed due to errors. |
2056 | * The capture_list will otherwise be freed at request retire. |
2057 | */ |
2058 | static void eb_capture_release(struct i915_execbuffer *eb) |
2059 | { |
2060 | unsigned int j; |
2061 | |
2062 | for_each_batch_create_order(eb, j) { |
2063 | if (eb->capture_lists[j]) { |
2064 | i915_request_free_capture_list(capture: eb->capture_lists[j]); |
2065 | eb->capture_lists[j] = NULL; |
2066 | } |
2067 | } |
2068 | } |
2069 | |
2070 | static void eb_capture_list_clear(struct i915_execbuffer *eb) |
2071 | { |
2072 | memset(eb->capture_lists, 0, sizeof(eb->capture_lists)); |
2073 | } |
2074 | |
2075 | #else |
2076 | |
2077 | static int eb_capture_stage(struct i915_execbuffer *eb) |
2078 | { |
2079 | return 0; |
2080 | } |
2081 | |
2082 | static void eb_capture_commit(struct i915_execbuffer *eb) |
2083 | { |
2084 | } |
2085 | |
2086 | static void eb_capture_release(struct i915_execbuffer *eb) |
2087 | { |
2088 | } |
2089 | |
2090 | static void eb_capture_list_clear(struct i915_execbuffer *eb) |
2091 | { |
2092 | } |
2093 | |
2094 | #endif |
2095 | |
2096 | static int eb_move_to_gpu(struct i915_execbuffer *eb) |
2097 | { |
2098 | const unsigned int count = eb->buffer_count; |
2099 | unsigned int i = count; |
2100 | int err = 0, j; |
2101 | |
2102 | while (i--) { |
2103 | struct eb_vma *ev = &eb->vma[i]; |
2104 | struct i915_vma *vma = ev->vma; |
2105 | unsigned int flags = ev->flags; |
2106 | struct drm_i915_gem_object *obj = vma->obj; |
2107 | |
2108 | assert_vma_held(vma); |
2109 | |
2110 | /* |
2111 | * If the GPU is not _reading_ through the CPU cache, we need |
2112 | * to make sure that any writes (both previous GPU writes from |
2113 | * before a change in snooping levels and normal CPU writes) |
2114 | * caught in that cache are flushed to main memory. |
2115 | * |
2116 | * We want to say |
2117 | * obj->cache_dirty && |
2118 | * !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ) |
2119 | * but gcc's optimiser doesn't handle that as well and emits |
2120 | * two jumps instead of one. Maybe one day... |
2121 | * |
2122 | * FIXME: There is also sync flushing in set_pages(), which |
2123 | * serves a different purpose(some of the time at least). |
2124 | * |
2125 | * We should consider: |
2126 | * |
2127 | * 1. Rip out the async flush code. |
2128 | * |
2129 | * 2. Or make the sync flushing use the async clflush path |
2130 | * using mandatory fences underneath. Currently the below |
2131 | * async flush happens after we bind the object. |
2132 | */ |
2133 | if (unlikely(obj->cache_dirty & ~obj->cache_coherent)) { |
2134 | if (i915_gem_clflush_object(obj, flags: 0)) |
2135 | flags &= ~EXEC_OBJECT_ASYNC; |
2136 | } |
2137 | |
2138 | /* We only need to await on the first request */ |
2139 | if (err == 0 && !(flags & EXEC_OBJECT_ASYNC)) { |
2140 | err = i915_request_await_object |
2141 | (to: eb_find_first_request_added(eb), obj, |
2142 | write: flags & EXEC_OBJECT_WRITE); |
2143 | } |
2144 | |
2145 | for_each_batch_add_order(eb, j) { |
2146 | if (err) |
2147 | break; |
2148 | if (!eb->requests[j]) |
2149 | continue; |
2150 | |
2151 | err = _i915_vma_move_to_active(vma, rq: eb->requests[j], |
2152 | fence: j ? NULL : |
2153 | eb->composite_fence ? |
2154 | eb->composite_fence : |
2155 | &eb->requests[j]->fence, |
2156 | flags: flags | __EXEC_OBJECT_NO_RESERVE | |
2157 | __EXEC_OBJECT_NO_REQUEST_AWAIT); |
2158 | } |
2159 | } |
2160 | |
2161 | #ifdef CONFIG_MMU_NOTIFIER |
2162 | if (!err && (eb->args->flags & __EXEC_USERPTR_USED)) { |
2163 | for (i = 0; i < count; i++) { |
2164 | struct eb_vma *ev = &eb->vma[i]; |
2165 | struct drm_i915_gem_object *obj = ev->vma->obj; |
2166 | |
2167 | if (!i915_gem_object_is_userptr(obj)) |
2168 | continue; |
2169 | |
2170 | err = i915_gem_object_userptr_submit_done(obj); |
2171 | if (err) |
2172 | break; |
2173 | } |
2174 | } |
2175 | #endif |
2176 | |
2177 | if (unlikely(err)) |
2178 | goto err_skip; |
2179 | |
2180 | /* Unconditionally flush any chipset caches (for streaming writes). */ |
2181 | intel_gt_chipset_flush(gt: eb->gt); |
2182 | eb_capture_commit(eb); |
2183 | |
2184 | return 0; |
2185 | |
2186 | err_skip: |
2187 | for_each_batch_create_order(eb, j) { |
2188 | if (!eb->requests[j]) |
2189 | break; |
2190 | |
2191 | i915_request_set_error_once(rq: eb->requests[j], error: err); |
2192 | } |
2193 | return err; |
2194 | } |
2195 | |
2196 | static int i915_gem_check_execbuffer(struct drm_i915_private *i915, |
2197 | struct drm_i915_gem_execbuffer2 *exec) |
2198 | { |
2199 | if (exec->flags & __I915_EXEC_ILLEGAL_FLAGS) |
2200 | return -EINVAL; |
2201 | |
2202 | /* Kernel clipping was a DRI1 misfeature */ |
2203 | if (!(exec->flags & (I915_EXEC_FENCE_ARRAY | |
2204 | I915_EXEC_USE_EXTENSIONS))) { |
2205 | if (exec->num_cliprects || exec->cliprects_ptr) |
2206 | return -EINVAL; |
2207 | } |
2208 | |
2209 | if (exec->DR4 == 0xffffffff) { |
2210 | drm_dbg(&i915->drm, "UXA submitting garbage DR4, fixing up\n" ); |
2211 | exec->DR4 = 0; |
2212 | } |
2213 | if (exec->DR1 || exec->DR4) |
2214 | return -EINVAL; |
2215 | |
2216 | if ((exec->batch_start_offset | exec->batch_len) & 0x7) |
2217 | return -EINVAL; |
2218 | |
2219 | return 0; |
2220 | } |
2221 | |
2222 | static int i915_reset_gen7_sol_offsets(struct i915_request *rq) |
2223 | { |
2224 | u32 *cs; |
2225 | int i; |
2226 | |
2227 | if (GRAPHICS_VER(rq->i915) != 7 || rq->engine->id != RCS0) { |
2228 | drm_dbg(&rq->i915->drm, "sol reset is gen7/rcs only\n" ); |
2229 | return -EINVAL; |
2230 | } |
2231 | |
2232 | cs = intel_ring_begin(rq, num_dwords: 4 * 2 + 2); |
2233 | if (IS_ERR(ptr: cs)) |
2234 | return PTR_ERR(ptr: cs); |
2235 | |
2236 | *cs++ = MI_LOAD_REGISTER_IMM(4); |
2237 | for (i = 0; i < 4; i++) { |
2238 | *cs++ = i915_mmio_reg_offset(GEN7_SO_WRITE_OFFSET(i)); |
2239 | *cs++ = 0; |
2240 | } |
2241 | *cs++ = MI_NOOP; |
2242 | intel_ring_advance(rq, cs); |
2243 | |
2244 | return 0; |
2245 | } |
2246 | |
2247 | static struct i915_vma * |
2248 | shadow_batch_pin(struct i915_execbuffer *eb, |
2249 | struct drm_i915_gem_object *obj, |
2250 | struct i915_address_space *vm, |
2251 | unsigned int flags) |
2252 | { |
2253 | struct i915_vma *vma; |
2254 | int err; |
2255 | |
2256 | vma = i915_vma_instance(obj, vm, NULL); |
2257 | if (IS_ERR(ptr: vma)) |
2258 | return vma; |
2259 | |
2260 | err = i915_vma_pin_ww(vma, ww: &eb->ww, size: 0, alignment: 0, flags: flags | PIN_VALIDATE); |
2261 | if (err) |
2262 | return ERR_PTR(error: err); |
2263 | |
2264 | return vma; |
2265 | } |
2266 | |
2267 | static struct i915_vma *eb_dispatch_secure(struct i915_execbuffer *eb, struct i915_vma *vma) |
2268 | { |
2269 | /* |
2270 | * snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure |
2271 | * batch" bit. Hence we need to pin secure batches into the global gtt. |
2272 | * hsw should have this fixed, but bdw mucks it up again. */ |
2273 | if (eb->batch_flags & I915_DISPATCH_SECURE) |
2274 | return i915_gem_object_ggtt_pin_ww(obj: vma->obj, ww: &eb->ww, NULL, size: 0, alignment: 0, PIN_VALIDATE); |
2275 | |
2276 | return NULL; |
2277 | } |
2278 | |
2279 | static int eb_parse(struct i915_execbuffer *eb) |
2280 | { |
2281 | struct drm_i915_private *i915 = eb->i915; |
2282 | struct intel_gt_buffer_pool_node *pool = eb->batch_pool; |
2283 | struct i915_vma *shadow, *trampoline, *batch; |
2284 | unsigned long len; |
2285 | int err; |
2286 | |
2287 | if (!eb_use_cmdparser(eb)) { |
2288 | batch = eb_dispatch_secure(eb, vma: eb->batches[0]->vma); |
2289 | if (IS_ERR(ptr: batch)) |
2290 | return PTR_ERR(ptr: batch); |
2291 | |
2292 | goto secure_batch; |
2293 | } |
2294 | |
2295 | if (intel_context_is_parallel(ce: eb->context)) |
2296 | return -EINVAL; |
2297 | |
2298 | len = eb->batch_len[0]; |
2299 | if (!CMDPARSER_USES_GGTT(eb->i915)) { |
2300 | /* |
2301 | * ppGTT backed shadow buffers must be mapped RO, to prevent |
2302 | * post-scan tampering |
2303 | */ |
2304 | if (!eb->context->vm->has_read_only) { |
2305 | drm_dbg(&i915->drm, |
2306 | "Cannot prevent post-scan tampering without RO capable vm\n" ); |
2307 | return -EINVAL; |
2308 | } |
2309 | } else { |
2310 | len += I915_CMD_PARSER_TRAMPOLINE_SIZE; |
2311 | } |
2312 | if (unlikely(len < eb->batch_len[0])) /* last paranoid check of overflow */ |
2313 | return -EINVAL; |
2314 | |
2315 | if (!pool) { |
2316 | pool = intel_gt_get_buffer_pool(gt: eb->gt, size: len, |
2317 | type: I915_MAP_WB); |
2318 | if (IS_ERR(ptr: pool)) |
2319 | return PTR_ERR(ptr: pool); |
2320 | eb->batch_pool = pool; |
2321 | } |
2322 | |
2323 | err = i915_gem_object_lock(obj: pool->obj, ww: &eb->ww); |
2324 | if (err) |
2325 | return err; |
2326 | |
2327 | shadow = shadow_batch_pin(eb, obj: pool->obj, vm: eb->context->vm, PIN_USER); |
2328 | if (IS_ERR(ptr: shadow)) |
2329 | return PTR_ERR(ptr: shadow); |
2330 | |
2331 | intel_gt_buffer_pool_mark_used(node: pool); |
2332 | i915_gem_object_set_readonly(obj: shadow->obj); |
2333 | shadow->private = pool; |
2334 | |
2335 | trampoline = NULL; |
2336 | if (CMDPARSER_USES_GGTT(eb->i915)) { |
2337 | trampoline = shadow; |
2338 | |
2339 | shadow = shadow_batch_pin(eb, obj: pool->obj, |
2340 | vm: &eb->gt->ggtt->vm, |
2341 | PIN_GLOBAL); |
2342 | if (IS_ERR(ptr: shadow)) |
2343 | return PTR_ERR(ptr: shadow); |
2344 | |
2345 | shadow->private = pool; |
2346 | |
2347 | eb->batch_flags |= I915_DISPATCH_SECURE; |
2348 | } |
2349 | |
2350 | batch = eb_dispatch_secure(eb, vma: shadow); |
2351 | if (IS_ERR(ptr: batch)) |
2352 | return PTR_ERR(ptr: batch); |
2353 | |
2354 | err = dma_resv_reserve_fences(obj: shadow->obj->base.resv, num_fences: 1); |
2355 | if (err) |
2356 | return err; |
2357 | |
2358 | err = intel_engine_cmd_parser(engine: eb->context->engine, |
2359 | batch: eb->batches[0]->vma, |
2360 | batch_offset: eb->batch_start_offset, |
2361 | batch_length: eb->batch_len[0], |
2362 | shadow, trampoline); |
2363 | if (err) |
2364 | return err; |
2365 | |
2366 | eb->batches[0] = &eb->vma[eb->buffer_count++]; |
2367 | eb->batches[0]->vma = i915_vma_get(vma: shadow); |
2368 | eb->batches[0]->flags = __EXEC_OBJECT_HAS_PIN; |
2369 | |
2370 | eb->trampoline = trampoline; |
2371 | eb->batch_start_offset = 0; |
2372 | |
2373 | secure_batch: |
2374 | if (batch) { |
2375 | if (intel_context_is_parallel(ce: eb->context)) |
2376 | return -EINVAL; |
2377 | |
2378 | eb->batches[0] = &eb->vma[eb->buffer_count++]; |
2379 | eb->batches[0]->flags = __EXEC_OBJECT_HAS_PIN; |
2380 | eb->batches[0]->vma = i915_vma_get(vma: batch); |
2381 | } |
2382 | return 0; |
2383 | } |
2384 | |
2385 | static int eb_request_submit(struct i915_execbuffer *eb, |
2386 | struct i915_request *rq, |
2387 | struct i915_vma *batch, |
2388 | u64 batch_len) |
2389 | { |
2390 | int err; |
2391 | |
2392 | if (intel_context_nopreempt(ce: rq->context)) |
2393 | __set_bit(I915_FENCE_FLAG_NOPREEMPT, &rq->fence.flags); |
2394 | |
2395 | if (eb->args->flags & I915_EXEC_GEN7_SOL_RESET) { |
2396 | err = i915_reset_gen7_sol_offsets(rq); |
2397 | if (err) |
2398 | return err; |
2399 | } |
2400 | |
2401 | /* |
2402 | * After we completed waiting for other engines (using HW semaphores) |
2403 | * then we can signal that this request/batch is ready to run. This |
2404 | * allows us to determine if the batch is still waiting on the GPU |
2405 | * or actually running by checking the breadcrumb. |
2406 | */ |
2407 | if (rq->context->engine->emit_init_breadcrumb) { |
2408 | err = rq->context->engine->emit_init_breadcrumb(rq); |
2409 | if (err) |
2410 | return err; |
2411 | } |
2412 | |
2413 | err = rq->context->engine->emit_bb_start(rq, |
2414 | i915_vma_offset(vma: batch) + |
2415 | eb->batch_start_offset, |
2416 | batch_len, |
2417 | eb->batch_flags); |
2418 | if (err) |
2419 | return err; |
2420 | |
2421 | if (eb->trampoline) { |
2422 | GEM_BUG_ON(intel_context_is_parallel(rq->context)); |
2423 | GEM_BUG_ON(eb->batch_start_offset); |
2424 | err = rq->context->engine->emit_bb_start(rq, |
2425 | i915_vma_offset(vma: eb->trampoline) + |
2426 | batch_len, 0, 0); |
2427 | if (err) |
2428 | return err; |
2429 | } |
2430 | |
2431 | return 0; |
2432 | } |
2433 | |
2434 | static int eb_submit(struct i915_execbuffer *eb) |
2435 | { |
2436 | unsigned int i; |
2437 | int err; |
2438 | |
2439 | err = eb_move_to_gpu(eb); |
2440 | |
2441 | for_each_batch_create_order(eb, i) { |
2442 | if (!eb->requests[i]) |
2443 | break; |
2444 | |
2445 | trace_i915_request_queue(rq: eb->requests[i], flags: eb->batch_flags); |
2446 | if (!err) |
2447 | err = eb_request_submit(eb, rq: eb->requests[i], |
2448 | batch: eb->batches[i]->vma, |
2449 | batch_len: eb->batch_len[i]); |
2450 | } |
2451 | |
2452 | return err; |
2453 | } |
2454 | |
2455 | /* |
2456 | * Find one BSD ring to dispatch the corresponding BSD command. |
2457 | * The engine index is returned. |
2458 | */ |
2459 | static unsigned int |
2460 | gen8_dispatch_bsd_engine(struct drm_i915_private *dev_priv, |
2461 | struct drm_file *file) |
2462 | { |
2463 | struct drm_i915_file_private *file_priv = file->driver_priv; |
2464 | |
2465 | /* Check whether the file_priv has already selected one ring. */ |
2466 | if ((int)file_priv->bsd_engine < 0) |
2467 | file_priv->bsd_engine = |
2468 | get_random_u32_below(ceil: dev_priv->engine_uabi_class_count[I915_ENGINE_CLASS_VIDEO]); |
2469 | |
2470 | return file_priv->bsd_engine; |
2471 | } |
2472 | |
2473 | static const enum intel_engine_id user_ring_map[] = { |
2474 | [I915_EXEC_DEFAULT] = RCS0, |
2475 | [I915_EXEC_RENDER] = RCS0, |
2476 | [I915_EXEC_BLT] = BCS0, |
2477 | [I915_EXEC_BSD] = VCS0, |
2478 | [I915_EXEC_VEBOX] = VECS0 |
2479 | }; |
2480 | |
2481 | static struct i915_request *eb_throttle(struct i915_execbuffer *eb, struct intel_context *ce) |
2482 | { |
2483 | struct intel_ring *ring = ce->ring; |
2484 | struct intel_timeline *tl = ce->timeline; |
2485 | struct i915_request *rq; |
2486 | |
2487 | /* |
2488 | * Completely unscientific finger-in-the-air estimates for suitable |
2489 | * maximum user request size (to avoid blocking) and then backoff. |
2490 | */ |
2491 | if (intel_ring_update_space(ring) >= PAGE_SIZE) |
2492 | return NULL; |
2493 | |
2494 | /* |
2495 | * Find a request that after waiting upon, there will be at least half |
2496 | * the ring available. The hysteresis allows us to compete for the |
2497 | * shared ring and should mean that we sleep less often prior to |
2498 | * claiming our resources, but not so long that the ring completely |
2499 | * drains before we can submit our next request. |
2500 | */ |
2501 | list_for_each_entry(rq, &tl->requests, link) { |
2502 | if (rq->ring != ring) |
2503 | continue; |
2504 | |
2505 | if (__intel_ring_space(head: rq->postfix, |
2506 | tail: ring->emit, size: ring->size) > ring->size / 2) |
2507 | break; |
2508 | } |
2509 | if (&rq->link == &tl->requests) |
2510 | return NULL; /* weird, we will check again later for real */ |
2511 | |
2512 | return i915_request_get(rq); |
2513 | } |
2514 | |
2515 | static int eb_pin_timeline(struct i915_execbuffer *eb, struct intel_context *ce, |
2516 | bool throttle) |
2517 | { |
2518 | struct intel_timeline *tl; |
2519 | struct i915_request *rq = NULL; |
2520 | |
2521 | /* |
2522 | * Take a local wakeref for preparing to dispatch the execbuf as |
2523 | * we expect to access the hardware fairly frequently in the |
2524 | * process, and require the engine to be kept awake between accesses. |
2525 | * Upon dispatch, we acquire another prolonged wakeref that we hold |
2526 | * until the timeline is idle, which in turn releases the wakeref |
2527 | * taken on the engine, and the parent device. |
2528 | */ |
2529 | tl = intel_context_timeline_lock(ce); |
2530 | if (IS_ERR(ptr: tl)) |
2531 | return PTR_ERR(ptr: tl); |
2532 | |
2533 | intel_context_enter(ce); |
2534 | if (throttle) |
2535 | rq = eb_throttle(eb, ce); |
2536 | intel_context_timeline_unlock(tl); |
2537 | |
2538 | if (rq) { |
2539 | bool nonblock = eb->file->filp->f_flags & O_NONBLOCK; |
2540 | long timeout = nonblock ? 0 : MAX_SCHEDULE_TIMEOUT; |
2541 | |
2542 | if (i915_request_wait(rq, I915_WAIT_INTERRUPTIBLE, |
2543 | timeout) < 0) { |
2544 | i915_request_put(rq); |
2545 | |
2546 | /* |
2547 | * Error path, cannot use intel_context_timeline_lock as |
2548 | * that is user interruptable and this clean up step |
2549 | * must be done. |
2550 | */ |
2551 | mutex_lock(&ce->timeline->mutex); |
2552 | intel_context_exit(ce); |
2553 | mutex_unlock(lock: &ce->timeline->mutex); |
2554 | |
2555 | if (nonblock) |
2556 | return -EWOULDBLOCK; |
2557 | else |
2558 | return -EINTR; |
2559 | } |
2560 | i915_request_put(rq); |
2561 | } |
2562 | |
2563 | return 0; |
2564 | } |
2565 | |
2566 | static int eb_pin_engine(struct i915_execbuffer *eb, bool throttle) |
2567 | { |
2568 | struct intel_context *ce = eb->context, *child; |
2569 | int err; |
2570 | int i = 0, j = 0; |
2571 | |
2572 | GEM_BUG_ON(eb->args->flags & __EXEC_ENGINE_PINNED); |
2573 | |
2574 | if (unlikely(intel_context_is_banned(ce))) |
2575 | return -EIO; |
2576 | |
2577 | /* |
2578 | * Pinning the contexts may generate requests in order to acquire |
2579 | * GGTT space, so do this first before we reserve a seqno for |
2580 | * ourselves. |
2581 | */ |
2582 | err = intel_context_pin_ww(ce, ww: &eb->ww); |
2583 | if (err) |
2584 | return err; |
2585 | for_each_child(ce, child) { |
2586 | err = intel_context_pin_ww(ce: child, ww: &eb->ww); |
2587 | GEM_BUG_ON(err); /* perma-pinned should incr a counter */ |
2588 | } |
2589 | |
2590 | for_each_child(ce, child) { |
2591 | err = eb_pin_timeline(eb, ce: child, throttle); |
2592 | if (err) |
2593 | goto unwind; |
2594 | ++i; |
2595 | } |
2596 | err = eb_pin_timeline(eb, ce, throttle); |
2597 | if (err) |
2598 | goto unwind; |
2599 | |
2600 | eb->args->flags |= __EXEC_ENGINE_PINNED; |
2601 | return 0; |
2602 | |
2603 | unwind: |
2604 | for_each_child(ce, child) { |
2605 | if (j++ < i) { |
2606 | mutex_lock(&child->timeline->mutex); |
2607 | intel_context_exit(ce: child); |
2608 | mutex_unlock(lock: &child->timeline->mutex); |
2609 | } |
2610 | } |
2611 | for_each_child(ce, child) |
2612 | intel_context_unpin(ce: child); |
2613 | intel_context_unpin(ce); |
2614 | return err; |
2615 | } |
2616 | |
2617 | static void eb_unpin_engine(struct i915_execbuffer *eb) |
2618 | { |
2619 | struct intel_context *ce = eb->context, *child; |
2620 | |
2621 | if (!(eb->args->flags & __EXEC_ENGINE_PINNED)) |
2622 | return; |
2623 | |
2624 | eb->args->flags &= ~__EXEC_ENGINE_PINNED; |
2625 | |
2626 | for_each_child(ce, child) { |
2627 | mutex_lock(&child->timeline->mutex); |
2628 | intel_context_exit(ce: child); |
2629 | mutex_unlock(lock: &child->timeline->mutex); |
2630 | |
2631 | intel_context_unpin(ce: child); |
2632 | } |
2633 | |
2634 | mutex_lock(&ce->timeline->mutex); |
2635 | intel_context_exit(ce); |
2636 | mutex_unlock(lock: &ce->timeline->mutex); |
2637 | |
2638 | intel_context_unpin(ce); |
2639 | } |
2640 | |
2641 | static unsigned int |
2642 | eb_select_legacy_ring(struct i915_execbuffer *eb) |
2643 | { |
2644 | struct drm_i915_private *i915 = eb->i915; |
2645 | struct drm_i915_gem_execbuffer2 *args = eb->args; |
2646 | unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK; |
2647 | |
2648 | if (user_ring_id != I915_EXEC_BSD && |
2649 | (args->flags & I915_EXEC_BSD_MASK)) { |
2650 | drm_dbg(&i915->drm, |
2651 | "execbuf with non bsd ring but with invalid " |
2652 | "bsd dispatch flags: %d\n" , (int)(args->flags)); |
2653 | return -1; |
2654 | } |
2655 | |
2656 | if (user_ring_id == I915_EXEC_BSD && |
2657 | i915->engine_uabi_class_count[I915_ENGINE_CLASS_VIDEO] > 1) { |
2658 | unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK; |
2659 | |
2660 | if (bsd_idx == I915_EXEC_BSD_DEFAULT) { |
2661 | bsd_idx = gen8_dispatch_bsd_engine(dev_priv: i915, file: eb->file); |
2662 | } else if (bsd_idx >= I915_EXEC_BSD_RING1 && |
2663 | bsd_idx <= I915_EXEC_BSD_RING2) { |
2664 | bsd_idx >>= I915_EXEC_BSD_SHIFT; |
2665 | bsd_idx--; |
2666 | } else { |
2667 | drm_dbg(&i915->drm, |
2668 | "execbuf with unknown bsd ring: %u\n" , |
2669 | bsd_idx); |
2670 | return -1; |
2671 | } |
2672 | |
2673 | return _VCS(bsd_idx); |
2674 | } |
2675 | |
2676 | if (user_ring_id >= ARRAY_SIZE(user_ring_map)) { |
2677 | drm_dbg(&i915->drm, "execbuf with unknown ring: %u\n" , |
2678 | user_ring_id); |
2679 | return -1; |
2680 | } |
2681 | |
2682 | return user_ring_map[user_ring_id]; |
2683 | } |
2684 | |
2685 | static int |
2686 | eb_select_engine(struct i915_execbuffer *eb) |
2687 | { |
2688 | struct intel_context *ce, *child; |
2689 | struct intel_gt *gt; |
2690 | unsigned int idx; |
2691 | int err; |
2692 | |
2693 | if (i915_gem_context_user_engines(ctx: eb->gem_context)) |
2694 | idx = eb->args->flags & I915_EXEC_RING_MASK; |
2695 | else |
2696 | idx = eb_select_legacy_ring(eb); |
2697 | |
2698 | ce = i915_gem_context_get_engine(ctx: eb->gem_context, idx); |
2699 | if (IS_ERR(ptr: ce)) |
2700 | return PTR_ERR(ptr: ce); |
2701 | |
2702 | if (intel_context_is_parallel(ce)) { |
2703 | if (eb->buffer_count < ce->parallel.number_children + 1) { |
2704 | intel_context_put(ce); |
2705 | return -EINVAL; |
2706 | } |
2707 | if (eb->batch_start_offset || eb->args->batch_len) { |
2708 | intel_context_put(ce); |
2709 | return -EINVAL; |
2710 | } |
2711 | } |
2712 | eb->num_batches = ce->parallel.number_children + 1; |
2713 | gt = ce->engine->gt; |
2714 | |
2715 | for_each_child(ce, child) |
2716 | intel_context_get(ce: child); |
2717 | eb->wakeref = intel_gt_pm_get(gt: ce->engine->gt); |
2718 | /* |
2719 | * Keep GT0 active on MTL so that i915_vma_parked() doesn't |
2720 | * free VMAs while execbuf ioctl is validating VMAs. |
2721 | */ |
2722 | if (gt->info.id) |
2723 | eb->wakeref_gt0 = intel_gt_pm_get(gt: to_gt(i915: gt->i915)); |
2724 | |
2725 | if (!test_bit(CONTEXT_ALLOC_BIT, &ce->flags)) { |
2726 | err = intel_context_alloc_state(ce); |
2727 | if (err) |
2728 | goto err; |
2729 | } |
2730 | for_each_child(ce, child) { |
2731 | if (!test_bit(CONTEXT_ALLOC_BIT, &child->flags)) { |
2732 | err = intel_context_alloc_state(ce: child); |
2733 | if (err) |
2734 | goto err; |
2735 | } |
2736 | } |
2737 | |
2738 | /* |
2739 | * ABI: Before userspace accesses the GPU (e.g. execbuffer), report |
2740 | * EIO if the GPU is already wedged. |
2741 | */ |
2742 | err = intel_gt_terminally_wedged(gt: ce->engine->gt); |
2743 | if (err) |
2744 | goto err; |
2745 | |
2746 | if (!i915_vm_tryget(vm: ce->vm)) { |
2747 | err = -ENOENT; |
2748 | goto err; |
2749 | } |
2750 | |
2751 | eb->context = ce; |
2752 | eb->gt = ce->engine->gt; |
2753 | |
2754 | /* |
2755 | * Make sure engine pool stays alive even if we call intel_context_put |
2756 | * during ww handling. The pool is destroyed when last pm reference |
2757 | * is dropped, which breaks our -EDEADLK handling. |
2758 | */ |
2759 | return err; |
2760 | |
2761 | err: |
2762 | if (gt->info.id) |
2763 | intel_gt_pm_put(gt: to_gt(i915: gt->i915), handle: eb->wakeref_gt0); |
2764 | |
2765 | intel_gt_pm_put(gt: ce->engine->gt, handle: eb->wakeref); |
2766 | for_each_child(ce, child) |
2767 | intel_context_put(ce: child); |
2768 | intel_context_put(ce); |
2769 | return err; |
2770 | } |
2771 | |
2772 | static void |
2773 | eb_put_engine(struct i915_execbuffer *eb) |
2774 | { |
2775 | struct intel_context *child; |
2776 | |
2777 | i915_vm_put(vm: eb->context->vm); |
2778 | /* |
2779 | * This works in conjunction with eb_select_engine() to prevent |
2780 | * i915_vma_parked() from interfering while execbuf validates vmas. |
2781 | */ |
2782 | if (eb->gt->info.id) |
2783 | intel_gt_pm_put(gt: to_gt(i915: eb->gt->i915), handle: eb->wakeref_gt0); |
2784 | intel_gt_pm_put(gt: eb->context->engine->gt, handle: eb->wakeref); |
2785 | for_each_child(eb->context, child) |
2786 | intel_context_put(ce: child); |
2787 | intel_context_put(ce: eb->context); |
2788 | } |
2789 | |
2790 | static void |
2791 | __free_fence_array(struct eb_fence *fences, unsigned int n) |
2792 | { |
2793 | while (n--) { |
2794 | drm_syncobj_put(ptr_mask_bits(fences[n].syncobj, 2)); |
2795 | dma_fence_put(fence: fences[n].dma_fence); |
2796 | dma_fence_chain_free(chain: fences[n].chain_fence); |
2797 | } |
2798 | kvfree(addr: fences); |
2799 | } |
2800 | |
2801 | static int |
2802 | add_timeline_fence_array(struct i915_execbuffer *eb, |
2803 | const struct drm_i915_gem_execbuffer_ext_timeline_fences *timeline_fences) |
2804 | { |
2805 | struct drm_i915_gem_exec_fence __user *user_fences; |
2806 | u64 __user *user_values; |
2807 | struct eb_fence *f; |
2808 | u64 nfences; |
2809 | int err = 0; |
2810 | |
2811 | nfences = timeline_fences->fence_count; |
2812 | if (!nfences) |
2813 | return 0; |
2814 | |
2815 | /* Check multiplication overflow for access_ok() and kvmalloc_array() */ |
2816 | BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long)); |
2817 | if (nfences > min_t(unsigned long, |
2818 | ULONG_MAX / sizeof(*user_fences), |
2819 | SIZE_MAX / sizeof(*f)) - eb->num_fences) |
2820 | return -EINVAL; |
2821 | |
2822 | user_fences = u64_to_user_ptr(timeline_fences->handles_ptr); |
2823 | if (!access_ok(user_fences, nfences * sizeof(*user_fences))) |
2824 | return -EFAULT; |
2825 | |
2826 | user_values = u64_to_user_ptr(timeline_fences->values_ptr); |
2827 | if (!access_ok(user_values, nfences * sizeof(*user_values))) |
2828 | return -EFAULT; |
2829 | |
2830 | f = krealloc(objp: eb->fences, |
2831 | new_size: (eb->num_fences + nfences) * sizeof(*f), |
2832 | __GFP_NOWARN | GFP_KERNEL); |
2833 | if (!f) |
2834 | return -ENOMEM; |
2835 | |
2836 | eb->fences = f; |
2837 | f += eb->num_fences; |
2838 | |
2839 | BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) & |
2840 | ~__I915_EXEC_FENCE_UNKNOWN_FLAGS); |
2841 | |
2842 | while (nfences--) { |
2843 | struct drm_i915_gem_exec_fence user_fence; |
2844 | struct drm_syncobj *syncobj; |
2845 | struct dma_fence *fence = NULL; |
2846 | u64 point; |
2847 | |
2848 | if (__copy_from_user(to: &user_fence, |
2849 | from: user_fences++, |
2850 | n: sizeof(user_fence))) |
2851 | return -EFAULT; |
2852 | |
2853 | if (user_fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS) |
2854 | return -EINVAL; |
2855 | |
2856 | if (__get_user(point, user_values++)) |
2857 | return -EFAULT; |
2858 | |
2859 | syncobj = drm_syncobj_find(file_private: eb->file, handle: user_fence.handle); |
2860 | if (!syncobj) { |
2861 | drm_dbg(&eb->i915->drm, |
2862 | "Invalid syncobj handle provided\n" ); |
2863 | return -ENOENT; |
2864 | } |
2865 | |
2866 | fence = drm_syncobj_fence_get(syncobj); |
2867 | |
2868 | if (!fence && user_fence.flags && |
2869 | !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) { |
2870 | drm_dbg(&eb->i915->drm, |
2871 | "Syncobj handle has no fence\n" ); |
2872 | drm_syncobj_put(obj: syncobj); |
2873 | return -EINVAL; |
2874 | } |
2875 | |
2876 | if (fence) |
2877 | err = dma_fence_chain_find_seqno(pfence: &fence, seqno: point); |
2878 | |
2879 | if (err && !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) { |
2880 | drm_dbg(&eb->i915->drm, |
2881 | "Syncobj handle missing requested point %llu\n" , |
2882 | point); |
2883 | dma_fence_put(fence); |
2884 | drm_syncobj_put(obj: syncobj); |
2885 | return err; |
2886 | } |
2887 | |
2888 | /* |
2889 | * A point might have been signaled already and |
2890 | * garbage collected from the timeline. In this case |
2891 | * just ignore the point and carry on. |
2892 | */ |
2893 | if (!fence && !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) { |
2894 | drm_syncobj_put(obj: syncobj); |
2895 | continue; |
2896 | } |
2897 | |
2898 | /* |
2899 | * For timeline syncobjs we need to preallocate chains for |
2900 | * later signaling. |
2901 | */ |
2902 | if (point != 0 && user_fence.flags & I915_EXEC_FENCE_SIGNAL) { |
2903 | /* |
2904 | * Waiting and signaling the same point (when point != |
2905 | * 0) would break the timeline. |
2906 | */ |
2907 | if (user_fence.flags & I915_EXEC_FENCE_WAIT) { |
2908 | drm_dbg(&eb->i915->drm, |
2909 | "Trying to wait & signal the same timeline point.\n" ); |
2910 | dma_fence_put(fence); |
2911 | drm_syncobj_put(obj: syncobj); |
2912 | return -EINVAL; |
2913 | } |
2914 | |
2915 | f->chain_fence = dma_fence_chain_alloc(); |
2916 | if (!f->chain_fence) { |
2917 | drm_syncobj_put(obj: syncobj); |
2918 | dma_fence_put(fence); |
2919 | return -ENOMEM; |
2920 | } |
2921 | } else { |
2922 | f->chain_fence = NULL; |
2923 | } |
2924 | |
2925 | f->syncobj = ptr_pack_bits(syncobj, user_fence.flags, 2); |
2926 | f->dma_fence = fence; |
2927 | f->value = point; |
2928 | f++; |
2929 | eb->num_fences++; |
2930 | } |
2931 | |
2932 | return 0; |
2933 | } |
2934 | |
2935 | static int add_fence_array(struct i915_execbuffer *eb) |
2936 | { |
2937 | struct drm_i915_gem_execbuffer2 *args = eb->args; |
2938 | struct drm_i915_gem_exec_fence __user *user; |
2939 | unsigned long num_fences = args->num_cliprects; |
2940 | struct eb_fence *f; |
2941 | |
2942 | if (!(args->flags & I915_EXEC_FENCE_ARRAY)) |
2943 | return 0; |
2944 | |
2945 | if (!num_fences) |
2946 | return 0; |
2947 | |
2948 | /* Check multiplication overflow for access_ok() and kvmalloc_array() */ |
2949 | BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long)); |
2950 | if (num_fences > min_t(unsigned long, |
2951 | ULONG_MAX / sizeof(*user), |
2952 | SIZE_MAX / sizeof(*f) - eb->num_fences)) |
2953 | return -EINVAL; |
2954 | |
2955 | user = u64_to_user_ptr(args->cliprects_ptr); |
2956 | if (!access_ok(user, num_fences * sizeof(*user))) |
2957 | return -EFAULT; |
2958 | |
2959 | f = krealloc(objp: eb->fences, |
2960 | new_size: (eb->num_fences + num_fences) * sizeof(*f), |
2961 | __GFP_NOWARN | GFP_KERNEL); |
2962 | if (!f) |
2963 | return -ENOMEM; |
2964 | |
2965 | eb->fences = f; |
2966 | f += eb->num_fences; |
2967 | while (num_fences--) { |
2968 | struct drm_i915_gem_exec_fence user_fence; |
2969 | struct drm_syncobj *syncobj; |
2970 | struct dma_fence *fence = NULL; |
2971 | |
2972 | if (__copy_from_user(to: &user_fence, from: user++, n: sizeof(user_fence))) |
2973 | return -EFAULT; |
2974 | |
2975 | if (user_fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS) |
2976 | return -EINVAL; |
2977 | |
2978 | syncobj = drm_syncobj_find(file_private: eb->file, handle: user_fence.handle); |
2979 | if (!syncobj) { |
2980 | drm_dbg(&eb->i915->drm, |
2981 | "Invalid syncobj handle provided\n" ); |
2982 | return -ENOENT; |
2983 | } |
2984 | |
2985 | if (user_fence.flags & I915_EXEC_FENCE_WAIT) { |
2986 | fence = drm_syncobj_fence_get(syncobj); |
2987 | if (!fence) { |
2988 | drm_dbg(&eb->i915->drm, |
2989 | "Syncobj handle has no fence\n" ); |
2990 | drm_syncobj_put(obj: syncobj); |
2991 | return -EINVAL; |
2992 | } |
2993 | } |
2994 | |
2995 | BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) & |
2996 | ~__I915_EXEC_FENCE_UNKNOWN_FLAGS); |
2997 | |
2998 | f->syncobj = ptr_pack_bits(syncobj, user_fence.flags, 2); |
2999 | f->dma_fence = fence; |
3000 | f->value = 0; |
3001 | f->chain_fence = NULL; |
3002 | f++; |
3003 | eb->num_fences++; |
3004 | } |
3005 | |
3006 | return 0; |
3007 | } |
3008 | |
3009 | static void put_fence_array(struct eb_fence *fences, int num_fences) |
3010 | { |
3011 | if (fences) |
3012 | __free_fence_array(fences, n: num_fences); |
3013 | } |
3014 | |
3015 | static int |
3016 | await_fence_array(struct i915_execbuffer *eb, |
3017 | struct i915_request *rq) |
3018 | { |
3019 | unsigned int n; |
3020 | int err; |
3021 | |
3022 | for (n = 0; n < eb->num_fences; n++) { |
3023 | if (!eb->fences[n].dma_fence) |
3024 | continue; |
3025 | |
3026 | err = i915_request_await_dma_fence(rq, fence: eb->fences[n].dma_fence); |
3027 | if (err < 0) |
3028 | return err; |
3029 | } |
3030 | |
3031 | return 0; |
3032 | } |
3033 | |
3034 | static void signal_fence_array(const struct i915_execbuffer *eb, |
3035 | struct dma_fence * const fence) |
3036 | { |
3037 | unsigned int n; |
3038 | |
3039 | for (n = 0; n < eb->num_fences; n++) { |
3040 | struct drm_syncobj *syncobj; |
3041 | unsigned int flags; |
3042 | |
3043 | syncobj = ptr_unpack_bits(eb->fences[n].syncobj, &flags, 2); |
3044 | if (!(flags & I915_EXEC_FENCE_SIGNAL)) |
3045 | continue; |
3046 | |
3047 | if (eb->fences[n].chain_fence) { |
3048 | drm_syncobj_add_point(syncobj, |
3049 | chain: eb->fences[n].chain_fence, |
3050 | fence, |
3051 | point: eb->fences[n].value); |
3052 | /* |
3053 | * The chain's ownership is transferred to the |
3054 | * timeline. |
3055 | */ |
3056 | eb->fences[n].chain_fence = NULL; |
3057 | } else { |
3058 | drm_syncobj_replace_fence(syncobj, fence); |
3059 | } |
3060 | } |
3061 | } |
3062 | |
3063 | static int |
3064 | parse_timeline_fences(struct i915_user_extension __user *ext, void *data) |
3065 | { |
3066 | struct i915_execbuffer *eb = data; |
3067 | struct drm_i915_gem_execbuffer_ext_timeline_fences timeline_fences; |
3068 | |
3069 | if (copy_from_user(to: &timeline_fences, from: ext, n: sizeof(timeline_fences))) |
3070 | return -EFAULT; |
3071 | |
3072 | return add_timeline_fence_array(eb, timeline_fences: &timeline_fences); |
3073 | } |
3074 | |
3075 | static void retire_requests(struct intel_timeline *tl, struct i915_request *end) |
3076 | { |
3077 | struct i915_request *rq, *rn; |
3078 | |
3079 | list_for_each_entry_safe(rq, rn, &tl->requests, link) |
3080 | if (rq == end || !i915_request_retire(rq)) |
3081 | break; |
3082 | } |
3083 | |
3084 | static int eb_request_add(struct i915_execbuffer *eb, struct i915_request *rq, |
3085 | int err, bool last_parallel) |
3086 | { |
3087 | struct intel_timeline * const tl = i915_request_timeline(rq); |
3088 | struct i915_sched_attr attr = {}; |
3089 | struct i915_request *prev; |
3090 | |
3091 | lockdep_assert_held(&tl->mutex); |
3092 | lockdep_unpin_lock(&tl->mutex, rq->cookie); |
3093 | |
3094 | trace_i915_request_add(rq); |
3095 | |
3096 | prev = __i915_request_commit(request: rq); |
3097 | |
3098 | /* Check that the context wasn't destroyed before submission */ |
3099 | if (likely(!intel_context_is_closed(eb->context))) { |
3100 | attr = eb->gem_context->sched; |
3101 | } else { |
3102 | /* Serialise with context_close via the add_to_timeline */ |
3103 | i915_request_set_error_once(rq, error: -ENOENT); |
3104 | __i915_request_skip(rq); |
3105 | err = -ENOENT; /* override any transient errors */ |
3106 | } |
3107 | |
3108 | if (intel_context_is_parallel(ce: eb->context)) { |
3109 | if (err) { |
3110 | __i915_request_skip(rq); |
3111 | set_bit(nr: I915_FENCE_FLAG_SKIP_PARALLEL, |
3112 | addr: &rq->fence.flags); |
3113 | } |
3114 | if (last_parallel) |
3115 | set_bit(nr: I915_FENCE_FLAG_SUBMIT_PARALLEL, |
3116 | addr: &rq->fence.flags); |
3117 | } |
3118 | |
3119 | __i915_request_queue(rq, attr: &attr); |
3120 | |
3121 | /* Try to clean up the client's timeline after submitting the request */ |
3122 | if (prev) |
3123 | retire_requests(tl, end: prev); |
3124 | |
3125 | mutex_unlock(lock: &tl->mutex); |
3126 | |
3127 | return err; |
3128 | } |
3129 | |
3130 | static int eb_requests_add(struct i915_execbuffer *eb, int err) |
3131 | { |
3132 | int i; |
3133 | |
3134 | /* |
3135 | * We iterate in reverse order of creation to release timeline mutexes in |
3136 | * same order. |
3137 | */ |
3138 | for_each_batch_add_order(eb, i) { |
3139 | struct i915_request *rq = eb->requests[i]; |
3140 | |
3141 | if (!rq) |
3142 | continue; |
3143 | err |= eb_request_add(eb, rq, err, last_parallel: i == 0); |
3144 | } |
3145 | |
3146 | return err; |
3147 | } |
3148 | |
3149 | static const i915_user_extension_fn execbuf_extensions[] = { |
3150 | [DRM_I915_GEM_EXECBUFFER_EXT_TIMELINE_FENCES] = parse_timeline_fences, |
3151 | }; |
3152 | |
3153 | static int |
3154 | parse_execbuf2_extensions(struct drm_i915_gem_execbuffer2 *args, |
3155 | struct i915_execbuffer *eb) |
3156 | { |
3157 | if (!(args->flags & I915_EXEC_USE_EXTENSIONS)) |
3158 | return 0; |
3159 | |
3160 | /* The execbuf2 extension mechanism reuses cliprects_ptr. So we cannot |
3161 | * have another flag also using it at the same time. |
3162 | */ |
3163 | if (eb->args->flags & I915_EXEC_FENCE_ARRAY) |
3164 | return -EINVAL; |
3165 | |
3166 | if (args->num_cliprects != 0) |
3167 | return -EINVAL; |
3168 | |
3169 | return i915_user_extensions(u64_to_user_ptr(args->cliprects_ptr), |
3170 | tbl: execbuf_extensions, |
3171 | ARRAY_SIZE(execbuf_extensions), |
3172 | data: eb); |
3173 | } |
3174 | |
3175 | static void eb_requests_get(struct i915_execbuffer *eb) |
3176 | { |
3177 | unsigned int i; |
3178 | |
3179 | for_each_batch_create_order(eb, i) { |
3180 | if (!eb->requests[i]) |
3181 | break; |
3182 | |
3183 | i915_request_get(rq: eb->requests[i]); |
3184 | } |
3185 | } |
3186 | |
3187 | static void eb_requests_put(struct i915_execbuffer *eb) |
3188 | { |
3189 | unsigned int i; |
3190 | |
3191 | for_each_batch_create_order(eb, i) { |
3192 | if (!eb->requests[i]) |
3193 | break; |
3194 | |
3195 | i915_request_put(rq: eb->requests[i]); |
3196 | } |
3197 | } |
3198 | |
3199 | static struct sync_file * |
3200 | eb_composite_fence_create(struct i915_execbuffer *eb, int out_fence_fd) |
3201 | { |
3202 | struct sync_file *out_fence = NULL; |
3203 | struct dma_fence_array *fence_array; |
3204 | struct dma_fence **fences; |
3205 | unsigned int i; |
3206 | |
3207 | GEM_BUG_ON(!intel_context_is_parent(eb->context)); |
3208 | |
3209 | fences = kmalloc_array(n: eb->num_batches, size: sizeof(*fences), GFP_KERNEL); |
3210 | if (!fences) |
3211 | return ERR_PTR(error: -ENOMEM); |
3212 | |
3213 | for_each_batch_create_order(eb, i) { |
3214 | fences[i] = &eb->requests[i]->fence; |
3215 | __set_bit(I915_FENCE_FLAG_COMPOSITE, |
3216 | &eb->requests[i]->fence.flags); |
3217 | } |
3218 | |
3219 | fence_array = dma_fence_array_create(num_fences: eb->num_batches, |
3220 | fences, |
3221 | context: eb->context->parallel.fence_context, |
3222 | seqno: eb->context->parallel.seqno++, |
3223 | signal_on_any: false); |
3224 | if (!fence_array) { |
3225 | kfree(objp: fences); |
3226 | return ERR_PTR(error: -ENOMEM); |
3227 | } |
3228 | |
3229 | /* Move ownership to the dma_fence_array created above */ |
3230 | for_each_batch_create_order(eb, i) |
3231 | dma_fence_get(fence: fences[i]); |
3232 | |
3233 | if (out_fence_fd != -1) { |
3234 | out_fence = sync_file_create(fence: &fence_array->base); |
3235 | /* sync_file now owns fence_arry, drop creation ref */ |
3236 | dma_fence_put(fence: &fence_array->base); |
3237 | if (!out_fence) |
3238 | return ERR_PTR(error: -ENOMEM); |
3239 | } |
3240 | |
3241 | eb->composite_fence = &fence_array->base; |
3242 | |
3243 | return out_fence; |
3244 | } |
3245 | |
3246 | static struct sync_file * |
3247 | eb_fences_add(struct i915_execbuffer *eb, struct i915_request *rq, |
3248 | struct dma_fence *in_fence, int out_fence_fd) |
3249 | { |
3250 | struct sync_file *out_fence = NULL; |
3251 | int err; |
3252 | |
3253 | if (unlikely(eb->gem_context->syncobj)) { |
3254 | struct dma_fence *fence; |
3255 | |
3256 | fence = drm_syncobj_fence_get(syncobj: eb->gem_context->syncobj); |
3257 | err = i915_request_await_dma_fence(rq, fence); |
3258 | dma_fence_put(fence); |
3259 | if (err) |
3260 | return ERR_PTR(error: err); |
3261 | } |
3262 | |
3263 | if (in_fence) { |
3264 | if (eb->args->flags & I915_EXEC_FENCE_SUBMIT) |
3265 | err = i915_request_await_execution(rq, fence: in_fence); |
3266 | else |
3267 | err = i915_request_await_dma_fence(rq, fence: in_fence); |
3268 | if (err < 0) |
3269 | return ERR_PTR(error: err); |
3270 | } |
3271 | |
3272 | if (eb->fences) { |
3273 | err = await_fence_array(eb, rq); |
3274 | if (err) |
3275 | return ERR_PTR(error: err); |
3276 | } |
3277 | |
3278 | if (intel_context_is_parallel(ce: eb->context)) { |
3279 | out_fence = eb_composite_fence_create(eb, out_fence_fd); |
3280 | if (IS_ERR(ptr: out_fence)) |
3281 | return ERR_PTR(error: -ENOMEM); |
3282 | } else if (out_fence_fd != -1) { |
3283 | out_fence = sync_file_create(fence: &rq->fence); |
3284 | if (!out_fence) |
3285 | return ERR_PTR(error: -ENOMEM); |
3286 | } |
3287 | |
3288 | return out_fence; |
3289 | } |
3290 | |
3291 | static struct intel_context * |
3292 | eb_find_context(struct i915_execbuffer *eb, unsigned int context_number) |
3293 | { |
3294 | struct intel_context *child; |
3295 | |
3296 | if (likely(context_number == 0)) |
3297 | return eb->context; |
3298 | |
3299 | for_each_child(eb->context, child) |
3300 | if (!--context_number) |
3301 | return child; |
3302 | |
3303 | GEM_BUG_ON("Context not found" ); |
3304 | |
3305 | return NULL; |
3306 | } |
3307 | |
3308 | static struct sync_file * |
3309 | eb_requests_create(struct i915_execbuffer *eb, struct dma_fence *in_fence, |
3310 | int out_fence_fd) |
3311 | { |
3312 | struct sync_file *out_fence = NULL; |
3313 | unsigned int i; |
3314 | |
3315 | for_each_batch_create_order(eb, i) { |
3316 | /* Allocate a request for this batch buffer nice and early. */ |
3317 | eb->requests[i] = i915_request_create(ce: eb_find_context(eb, context_number: i)); |
3318 | if (IS_ERR(ptr: eb->requests[i])) { |
3319 | out_fence = ERR_CAST(ptr: eb->requests[i]); |
3320 | eb->requests[i] = NULL; |
3321 | return out_fence; |
3322 | } |
3323 | |
3324 | /* |
3325 | * Only the first request added (committed to backend) has to |
3326 | * take the in fences into account as all subsequent requests |
3327 | * will have fences inserted inbetween them. |
3328 | */ |
3329 | if (i + 1 == eb->num_batches) { |
3330 | out_fence = eb_fences_add(eb, rq: eb->requests[i], |
3331 | in_fence, out_fence_fd); |
3332 | if (IS_ERR(ptr: out_fence)) |
3333 | return out_fence; |
3334 | } |
3335 | |
3336 | /* |
3337 | * Not really on stack, but we don't want to call |
3338 | * kfree on the batch_snapshot when we put it, so use the |
3339 | * _onstack interface. |
3340 | */ |
3341 | if (eb->batches[i]->vma) |
3342 | eb->requests[i]->batch_res = |
3343 | i915_vma_resource_get(vma_res: eb->batches[i]->vma->resource); |
3344 | if (eb->batch_pool) { |
3345 | GEM_BUG_ON(intel_context_is_parallel(eb->context)); |
3346 | intel_gt_buffer_pool_mark_active(node: eb->batch_pool, |
3347 | rq: eb->requests[i]); |
3348 | } |
3349 | } |
3350 | |
3351 | return out_fence; |
3352 | } |
3353 | |
3354 | static int |
3355 | i915_gem_do_execbuffer(struct drm_device *dev, |
3356 | struct drm_file *file, |
3357 | struct drm_i915_gem_execbuffer2 *args, |
3358 | struct drm_i915_gem_exec_object2 *exec) |
3359 | { |
3360 | struct drm_i915_private *i915 = to_i915(dev); |
3361 | struct i915_execbuffer eb; |
3362 | struct dma_fence *in_fence = NULL; |
3363 | struct sync_file *out_fence = NULL; |
3364 | int out_fence_fd = -1; |
3365 | int err; |
3366 | |
3367 | BUILD_BUG_ON(__EXEC_INTERNAL_FLAGS & ~__I915_EXEC_ILLEGAL_FLAGS); |
3368 | BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS & |
3369 | ~__EXEC_OBJECT_UNKNOWN_FLAGS); |
3370 | |
3371 | eb.i915 = i915; |
3372 | eb.file = file; |
3373 | eb.args = args; |
3374 | if (DBG_FORCE_RELOC || !(args->flags & I915_EXEC_NO_RELOC)) |
3375 | args->flags |= __EXEC_HAS_RELOC; |
3376 | |
3377 | eb.exec = exec; |
3378 | eb.vma = (struct eb_vma *)(exec + args->buffer_count + 1); |
3379 | eb.vma[0].vma = NULL; |
3380 | eb.batch_pool = NULL; |
3381 | |
3382 | eb.invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS; |
3383 | reloc_cache_init(cache: &eb.reloc_cache, i915: eb.i915); |
3384 | |
3385 | eb.buffer_count = args->buffer_count; |
3386 | eb.batch_start_offset = args->batch_start_offset; |
3387 | eb.trampoline = NULL; |
3388 | |
3389 | eb.fences = NULL; |
3390 | eb.num_fences = 0; |
3391 | |
3392 | eb_capture_list_clear(eb: &eb); |
3393 | |
3394 | memset(eb.requests, 0, sizeof(struct i915_request *) * |
3395 | ARRAY_SIZE(eb.requests)); |
3396 | eb.composite_fence = NULL; |
3397 | |
3398 | eb.batch_flags = 0; |
3399 | if (args->flags & I915_EXEC_SECURE) { |
3400 | if (GRAPHICS_VER(i915) >= 11) |
3401 | return -ENODEV; |
3402 | |
3403 | /* Return -EPERM to trigger fallback code on old binaries. */ |
3404 | if (!HAS_SECURE_BATCHES(i915)) |
3405 | return -EPERM; |
3406 | |
3407 | if (!drm_is_current_master(fpriv: file) || !capable(CAP_SYS_ADMIN)) |
3408 | return -EPERM; |
3409 | |
3410 | eb.batch_flags |= I915_DISPATCH_SECURE; |
3411 | } |
3412 | if (args->flags & I915_EXEC_IS_PINNED) |
3413 | eb.batch_flags |= I915_DISPATCH_PINNED; |
3414 | |
3415 | err = parse_execbuf2_extensions(args, eb: &eb); |
3416 | if (err) |
3417 | goto err_ext; |
3418 | |
3419 | err = add_fence_array(eb: &eb); |
3420 | if (err) |
3421 | goto err_ext; |
3422 | |
3423 | #define IN_FENCES (I915_EXEC_FENCE_IN | I915_EXEC_FENCE_SUBMIT) |
3424 | if (args->flags & IN_FENCES) { |
3425 | if ((args->flags & IN_FENCES) == IN_FENCES) |
3426 | return -EINVAL; |
3427 | |
3428 | in_fence = sync_file_get_fence(lower_32_bits(args->rsvd2)); |
3429 | if (!in_fence) { |
3430 | err = -EINVAL; |
3431 | goto err_ext; |
3432 | } |
3433 | } |
3434 | #undef IN_FENCES |
3435 | |
3436 | if (args->flags & I915_EXEC_FENCE_OUT) { |
3437 | out_fence_fd = get_unused_fd_flags(O_CLOEXEC); |
3438 | if (out_fence_fd < 0) { |
3439 | err = out_fence_fd; |
3440 | goto err_in_fence; |
3441 | } |
3442 | } |
3443 | |
3444 | err = eb_create(eb: &eb); |
3445 | if (err) |
3446 | goto err_out_fence; |
3447 | |
3448 | GEM_BUG_ON(!eb.lut_size); |
3449 | |
3450 | err = eb_select_context(eb: &eb); |
3451 | if (unlikely(err)) |
3452 | goto err_destroy; |
3453 | |
3454 | err = eb_select_engine(eb: &eb); |
3455 | if (unlikely(err)) |
3456 | goto err_context; |
3457 | |
3458 | err = eb_lookup_vmas(eb: &eb); |
3459 | if (err) { |
3460 | eb_release_vmas(eb: &eb, final: true); |
3461 | goto err_engine; |
3462 | } |
3463 | |
3464 | i915_gem_ww_ctx_init(ctx: &eb.ww, intr: true); |
3465 | |
3466 | err = eb_relocate_parse(eb: &eb); |
3467 | if (err) { |
3468 | /* |
3469 | * If the user expects the execobject.offset and |
3470 | * reloc.presumed_offset to be an exact match, |
3471 | * as for using NO_RELOC, then we cannot update |
3472 | * the execobject.offset until we have completed |
3473 | * relocation. |
3474 | */ |
3475 | args->flags &= ~__EXEC_HAS_RELOC; |
3476 | goto err_vma; |
3477 | } |
3478 | |
3479 | ww_acquire_done(ctx: &eb.ww.ctx); |
3480 | err = eb_capture_stage(eb: &eb); |
3481 | if (err) |
3482 | goto err_vma; |
3483 | |
3484 | out_fence = eb_requests_create(eb: &eb, in_fence, out_fence_fd); |
3485 | if (IS_ERR(ptr: out_fence)) { |
3486 | err = PTR_ERR(ptr: out_fence); |
3487 | out_fence = NULL; |
3488 | if (eb.requests[0]) |
3489 | goto err_request; |
3490 | else |
3491 | goto err_vma; |
3492 | } |
3493 | |
3494 | err = eb_submit(eb: &eb); |
3495 | |
3496 | err_request: |
3497 | eb_requests_get(eb: &eb); |
3498 | err = eb_requests_add(eb: &eb, err); |
3499 | |
3500 | if (eb.fences) |
3501 | signal_fence_array(eb: &eb, fence: eb.composite_fence ? |
3502 | eb.composite_fence : |
3503 | &eb.requests[0]->fence); |
3504 | |
3505 | if (unlikely(eb.gem_context->syncobj)) { |
3506 | drm_syncobj_replace_fence(syncobj: eb.gem_context->syncobj, |
3507 | fence: eb.composite_fence ? |
3508 | eb.composite_fence : |
3509 | &eb.requests[0]->fence); |
3510 | } |
3511 | |
3512 | if (out_fence) { |
3513 | if (err == 0) { |
3514 | fd_install(fd: out_fence_fd, file: out_fence->file); |
3515 | args->rsvd2 &= GENMASK_ULL(31, 0); /* keep in-fence */ |
3516 | args->rsvd2 |= (u64)out_fence_fd << 32; |
3517 | out_fence_fd = -1; |
3518 | } else { |
3519 | fput(out_fence->file); |
3520 | } |
3521 | } |
3522 | |
3523 | if (!out_fence && eb.composite_fence) |
3524 | dma_fence_put(fence: eb.composite_fence); |
3525 | |
3526 | eb_requests_put(eb: &eb); |
3527 | |
3528 | err_vma: |
3529 | eb_release_vmas(eb: &eb, final: true); |
3530 | WARN_ON(err == -EDEADLK); |
3531 | i915_gem_ww_ctx_fini(ctx: &eb.ww); |
3532 | |
3533 | if (eb.batch_pool) |
3534 | intel_gt_buffer_pool_put(node: eb.batch_pool); |
3535 | err_engine: |
3536 | eb_put_engine(eb: &eb); |
3537 | err_context: |
3538 | i915_gem_context_put(ctx: eb.gem_context); |
3539 | err_destroy: |
3540 | eb_destroy(eb: &eb); |
3541 | err_out_fence: |
3542 | if (out_fence_fd != -1) |
3543 | put_unused_fd(fd: out_fence_fd); |
3544 | err_in_fence: |
3545 | dma_fence_put(fence: in_fence); |
3546 | err_ext: |
3547 | put_fence_array(fences: eb.fences, num_fences: eb.num_fences); |
3548 | return err; |
3549 | } |
3550 | |
3551 | static size_t eb_element_size(void) |
3552 | { |
3553 | return sizeof(struct drm_i915_gem_exec_object2) + sizeof(struct eb_vma); |
3554 | } |
3555 | |
3556 | static bool check_buffer_count(size_t count) |
3557 | { |
3558 | const size_t sz = eb_element_size(); |
3559 | |
3560 | /* |
3561 | * When using LUT_HANDLE, we impose a limit of INT_MAX for the lookup |
3562 | * array size (see eb_create()). Otherwise, we can accept an array as |
3563 | * large as can be addressed (though use large arrays at your peril)! |
3564 | */ |
3565 | |
3566 | return !(count < 1 || count > INT_MAX || count > SIZE_MAX / sz - 1); |
3567 | } |
3568 | |
3569 | int |
3570 | i915_gem_execbuffer2_ioctl(struct drm_device *dev, void *data, |
3571 | struct drm_file *file) |
3572 | { |
3573 | struct drm_i915_private *i915 = to_i915(dev); |
3574 | struct drm_i915_gem_execbuffer2 *args = data; |
3575 | struct drm_i915_gem_exec_object2 *exec2_list; |
3576 | const size_t count = args->buffer_count; |
3577 | int err; |
3578 | |
3579 | if (!check_buffer_count(count)) { |
3580 | drm_dbg(&i915->drm, "execbuf2 with %zd buffers\n" , count); |
3581 | return -EINVAL; |
3582 | } |
3583 | |
3584 | err = i915_gem_check_execbuffer(i915, exec: args); |
3585 | if (err) |
3586 | return err; |
3587 | |
3588 | /* Allocate extra slots for use by the command parser */ |
3589 | exec2_list = kvmalloc_array(n: count + 2, size: eb_element_size(), |
3590 | __GFP_NOWARN | GFP_KERNEL); |
3591 | if (exec2_list == NULL) { |
3592 | drm_dbg(&i915->drm, "Failed to allocate exec list for %zd buffers\n" , |
3593 | count); |
3594 | return -ENOMEM; |
3595 | } |
3596 | if (copy_from_user(to: exec2_list, |
3597 | u64_to_user_ptr(args->buffers_ptr), |
3598 | n: sizeof(*exec2_list) * count)) { |
3599 | drm_dbg(&i915->drm, "copy %zd exec entries failed\n" , count); |
3600 | kvfree(addr: exec2_list); |
3601 | return -EFAULT; |
3602 | } |
3603 | |
3604 | err = i915_gem_do_execbuffer(dev, file, args, exec: exec2_list); |
3605 | |
3606 | /* |
3607 | * Now that we have begun execution of the batchbuffer, we ignore |
3608 | * any new error after this point. Also given that we have already |
3609 | * updated the associated relocations, we try to write out the current |
3610 | * object locations irrespective of any error. |
3611 | */ |
3612 | if (args->flags & __EXEC_HAS_RELOC) { |
3613 | struct drm_i915_gem_exec_object2 __user *user_exec_list = |
3614 | u64_to_user_ptr(args->buffers_ptr); |
3615 | unsigned int i; |
3616 | |
3617 | /* Copy the new buffer offsets back to the user's exec list. */ |
3618 | /* |
3619 | * Note: count * sizeof(*user_exec_list) does not overflow, |
3620 | * because we checked 'count' in check_buffer_count(). |
3621 | * |
3622 | * And this range already got effectively checked earlier |
3623 | * when we did the "copy_from_user()" above. |
3624 | */ |
3625 | if (!user_write_access_begin(user_exec_list, |
3626 | count * sizeof(*user_exec_list))) |
3627 | goto end; |
3628 | |
3629 | for (i = 0; i < args->buffer_count; i++) { |
3630 | if (!(exec2_list[i].offset & UPDATE)) |
3631 | continue; |
3632 | |
3633 | exec2_list[i].offset = |
3634 | gen8_canonical_addr(address: exec2_list[i].offset & PIN_OFFSET_MASK); |
3635 | unsafe_put_user(exec2_list[i].offset, |
3636 | &user_exec_list[i].offset, |
3637 | end_user); |
3638 | } |
3639 | end_user: |
3640 | user_write_access_end(); |
3641 | end:; |
3642 | } |
3643 | |
3644 | args->flags &= ~__I915_EXEC_UNKNOWN_FLAGS; |
3645 | kvfree(addr: exec2_list); |
3646 | return err; |
3647 | } |
3648 | |