Commit | Line | Data |
---|---|---|
54cf91dc | 1 | /* |
10be98a7 | 2 | * SPDX-License-Identifier: MIT |
54cf91dc | 3 | * |
10be98a7 | 4 | * Copyright © 2008,2010 Intel Corporation |
54cf91dc CW |
5 | */ |
6 | ||
daedaa33 | 7 | #include <linux/intel-iommu.h> |
52791eee | 8 | #include <linux/dma-resv.h> |
fec0445c | 9 | #include <linux/sync_file.h> |
ad778f89 CW |
10 | #include <linux/uaccess.h> |
11 | ||
cf6e7bac | 12 | #include <drm/drm_syncobj.h> |
ad778f89 | 13 | |
df0566a6 JN |
14 | #include "display/intel_frontbuffer.h" |
15 | ||
afa13085 | 16 | #include "gem/i915_gem_ioctls.h" |
10be98a7 | 17 | #include "gt/intel_context.h" |
45233ab2 | 18 | #include "gt/intel_gpu_commands.h" |
baea429d | 19 | #include "gt/intel_gt.h" |
16e87459 | 20 | #include "gt/intel_gt_buffer_pool.h" |
8f2a1057 | 21 | #include "gt/intel_gt_pm.h" |
2871ea85 | 22 | #include "gt/intel_ring.h" |
8f2a1057 | 23 | |
d3ac8d42 DCS |
24 | #include "pxp/intel_pxp.h" |
25 | ||
6da4a2c4 | 26 | #include "i915_drv.h" |
57822dc6 | 27 | #include "i915_gem_clflush.h" |
10be98a7 | 28 | #include "i915_gem_context.h" |
6da4a2c4 | 29 | #include "i915_gem_ioctls.h" |
54cf91dc | 30 | #include "i915_trace.h" |
cda9edd0 | 31 | #include "i915_user_extensions.h" |
ff20afc4 | 32 | #include "i915_vma_snapshot.h" |
54cf91dc | 33 | |
7d6236bb CW |
34 | struct eb_vma { |
35 | struct i915_vma *vma; | |
36 | unsigned int flags; | |
37 | ||
38 | /** This vma's place in the execbuf reservation list */ | |
39 | struct drm_i915_gem_exec_object2 *exec; | |
40 | struct list_head bind_link; | |
41 | struct list_head reloc_link; | |
42 | ||
43 | struct hlist_node node; | |
44 | u32 handle; | |
45 | }; | |
46 | ||
ad5d95e4 DA |
47 | enum { |
48 | FORCE_CPU_RELOC = 1, | |
49 | FORCE_GTT_RELOC, | |
50 | FORCE_GPU_RELOC, | |
51 | #define DBG_FORCE_RELOC 0 /* choose one of the above! */ | |
52 | }; | |
53 | ||
bfaae47d ML |
54 | /* __EXEC_OBJECT_NO_RESERVE is BIT(31), defined in i915_vma.h */ |
55 | #define __EXEC_OBJECT_HAS_PIN BIT(30) | |
56 | #define __EXEC_OBJECT_HAS_FENCE BIT(29) | |
ed29c269 ML |
57 | #define __EXEC_OBJECT_USERPTR_INIT BIT(28) |
58 | #define __EXEC_OBJECT_NEEDS_MAP BIT(27) | |
59 | #define __EXEC_OBJECT_NEEDS_BIAS BIT(26) | |
60 | #define __EXEC_OBJECT_INTERNAL_FLAGS (~0u << 26) /* all of the above + */ | |
8ae275c2 | 61 | #define __EXEC_OBJECT_RESERVED (__EXEC_OBJECT_HAS_PIN | __EXEC_OBJECT_HAS_FENCE) |
2889caa9 CW |
62 | |
63 | #define __EXEC_HAS_RELOC BIT(31) | |
2bf541ff | 64 | #define __EXEC_ENGINE_PINNED BIT(30) |
ed29c269 ML |
65 | #define __EXEC_USERPTR_USED BIT(29) |
66 | #define __EXEC_INTERNAL_FLAGS (~0u << 29) | |
2889caa9 | 67 | #define UPDATE PIN_OFFSET_FIXED |
d23db88c CW |
68 | |
69 | #define BATCH_OFFSET_BIAS (256*1024) | |
a415d355 | 70 | |
650bc635 | 71 | #define __I915_EXEC_ILLEGAL_FLAGS \ |
08e3e21a LDM |
72 | (__I915_EXEC_UNKNOWN_FLAGS | \ |
73 | I915_EXEC_CONSTANTS_MASK | \ | |
74 | I915_EXEC_RESOURCE_STREAMER) | |
5b043f4e | 75 | |
d20ac620 CW |
76 | /* Catch emission of unexpected errors for CI! */ |
77 | #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) | |
78 | #undef EINVAL | |
79 | #define EINVAL ({ \ | |
80 | DRM_DEBUG_DRIVER("EINVAL at %s:%d\n", __func__, __LINE__); \ | |
81 | 22; \ | |
82 | }) | |
83 | #endif | |
84 | ||
2889caa9 CW |
85 | /** |
86 | * DOC: User command execution | |
87 | * | |
88 | * Userspace submits commands to be executed on the GPU as an instruction | |
89 | * stream within a GEM object we call a batchbuffer. This instructions may | |
90 | * refer to other GEM objects containing auxiliary state such as kernels, | |
91 | * samplers, render targets and even secondary batchbuffers. Userspace does | |
92 | * not know where in the GPU memory these objects reside and so before the | |
93 | * batchbuffer is passed to the GPU for execution, those addresses in the | |
94 | * batchbuffer and auxiliary objects are updated. This is known as relocation, | |
95 | * or patching. To try and avoid having to relocate each object on the next | |
96 | * execution, userspace is told the location of those objects in this pass, | |
97 | * but this remains just a hint as the kernel may choose a new location for | |
98 | * any object in the future. | |
99 | * | |
99d7e4ee KR |
100 | * At the level of talking to the hardware, submitting a batchbuffer for the |
101 | * GPU to execute is to add content to a buffer from which the HW | |
102 | * command streamer is reading. | |
103 | * | |
104 | * 1. Add a command to load the HW context. For Logical Ring Contexts, i.e. | |
105 | * Execlists, this command is not placed on the same buffer as the | |
106 | * remaining items. | |
107 | * | |
108 | * 2. Add a command to invalidate caches to the buffer. | |
109 | * | |
110 | * 3. Add a batchbuffer start command to the buffer; the start command is | |
111 | * essentially a token together with the GPU address of the batchbuffer | |
112 | * to be executed. | |
113 | * | |
114 | * 4. Add a pipeline flush to the buffer. | |
115 | * | |
116 | * 5. Add a memory write command to the buffer to record when the GPU | |
117 | * is done executing the batchbuffer. The memory write writes the | |
118 | * global sequence number of the request, ``i915_request::global_seqno``; | |
119 | * the i915 driver uses the current value in the register to determine | |
120 | * if the GPU has completed the batchbuffer. | |
121 | * | |
122 | * 6. Add a user interrupt command to the buffer. This command instructs | |
123 | * the GPU to issue an interrupt when the command, pipeline flush and | |
124 | * memory write are completed. | |
125 | * | |
126 | * 7. Inform the hardware of the additional commands added to the buffer | |
127 | * (by updating the tail pointer). | |
128 | * | |
2889caa9 CW |
129 | * Processing an execbuf ioctl is conceptually split up into a few phases. |
130 | * | |
131 | * 1. Validation - Ensure all the pointers, handles and flags are valid. | |
132 | * 2. Reservation - Assign GPU address space for every object | |
133 | * 3. Relocation - Update any addresses to point to the final locations | |
134 | * 4. Serialisation - Order the request with respect to its dependencies | |
135 | * 5. Construction - Construct a request to execute the batchbuffer | |
136 | * 6. Submission (at some point in the future execution) | |
137 | * | |
138 | * Reserving resources for the execbuf is the most complicated phase. We | |
139 | * neither want to have to migrate the object in the address space, nor do | |
140 | * we want to have to update any relocations pointing to this object. Ideally, | |
141 | * we want to leave the object where it is and for all the existing relocations | |
142 | * to match. If the object is given a new address, or if userspace thinks the | |
143 | * object is elsewhere, we have to parse all the relocation entries and update | |
144 | * the addresses. Userspace can set the I915_EXEC_NORELOC flag to hint that | |
145 | * all the target addresses in all of its objects match the value in the | |
146 | * relocation entries and that they all match the presumed offsets given by the | |
147 | * list of execbuffer objects. Using this knowledge, we know that if we haven't | |
148 | * moved any buffers, all the relocation entries are valid and we can skip | |
149 | * the update. (If userspace is wrong, the likely outcome is an impromptu GPU | |
150 | * hang.) The requirement for using I915_EXEC_NO_RELOC are: | |
151 | * | |
152 | * The addresses written in the objects must match the corresponding | |
153 | * reloc.presumed_offset which in turn must match the corresponding | |
154 | * execobject.offset. | |
155 | * | |
156 | * Any render targets written to in the batch must be flagged with | |
157 | * EXEC_OBJECT_WRITE. | |
158 | * | |
159 | * To avoid stalling, execobject.offset should match the current | |
160 | * address of that object within the active context. | |
161 | * | |
162 | * The reservation is done is multiple phases. First we try and keep any | |
163 | * object already bound in its current location - so as long as meets the | |
164 | * constraints imposed by the new execbuffer. Any object left unbound after the | |
165 | * first pass is then fitted into any available idle space. If an object does | |
166 | * not fit, all objects are removed from the reservation and the process rerun | |
167 | * after sorting the objects into a priority order (more difficult to fit | |
168 | * objects are tried first). Failing that, the entire VM is cleared and we try | |
169 | * to fit the execbuf once last time before concluding that it simply will not | |
170 | * fit. | |
171 | * | |
172 | * A small complication to all of this is that we allow userspace not only to | |
173 | * specify an alignment and a size for the object in the address space, but | |
174 | * we also allow userspace to specify the exact offset. This objects are | |
175 | * simpler to place (the location is known a priori) all we have to do is make | |
176 | * sure the space is available. | |
177 | * | |
178 | * Once all the objects are in place, patching up the buried pointers to point | |
179 | * to the final locations is a fairly simple job of walking over the relocation | |
180 | * entry arrays, looking up the right address and rewriting the value into | |
181 | * the object. Simple! ... The relocation entries are stored in user memory | |
182 | * and so to access them we have to copy them into a local buffer. That copy | |
183 | * has to avoid taking any pagefaults as they may lead back to a GEM object | |
184 | * requiring the struct_mutex (i.e. recursive deadlock). So once again we split | |
185 | * the relocation into multiple passes. First we try to do everything within an | |
186 | * atomic context (avoid the pagefaults) which requires that we never wait. If | |
187 | * we detect that we may wait, or if we need to fault, then we have to fallback | |
188 | * to a slower path. The slowpath has to drop the mutex. (Can you hear alarm | |
189 | * bells yet?) Dropping the mutex means that we lose all the state we have | |
190 | * built up so far for the execbuf and we must reset any global data. However, | |
191 | * we do leave the objects pinned in their final locations - which is a | |
192 | * potential issue for concurrent execbufs. Once we have left the mutex, we can | |
193 | * allocate and copy all the relocation entries into a large array at our | |
194 | * leisure, reacquire the mutex, reclaim all the objects and other state and | |
195 | * then proceed to update any incorrect addresses with the objects. | |
196 | * | |
197 | * As we process the relocation entries, we maintain a record of whether the | |
198 | * object is being written to. Using NORELOC, we expect userspace to provide | |
199 | * this information instead. We also check whether we can skip the relocation | |
200 | * by comparing the expected value inside the relocation entry with the target's | |
201 | * final address. If they differ, we have to map the current object and rewrite | |
202 | * the 4 or 8 byte pointer within. | |
203 | * | |
204 | * Serialising an execbuf is quite simple according to the rules of the GEM | |
205 | * ABI. Execution within each context is ordered by the order of submission. | |
206 | * Writes to any GEM object are in order of submission and are exclusive. Reads | |
207 | * from a GEM object are unordered with respect to other reads, but ordered by | |
208 | * writes. A write submitted after a read cannot occur before the read, and | |
209 | * similarly any read submitted after a write cannot occur before the write. | |
210 | * Writes are ordered between engines such that only one write occurs at any | |
211 | * time (completing any reads beforehand) - using semaphores where available | |
212 | * and CPU serialisation otherwise. Other GEM access obey the same rules, any | |
213 | * write (either via mmaps using set-domain, or via pwrite) must flush all GPU | |
214 | * reads before starting, and any read (either using set-domain or pread) must | |
215 | * flush all GPU writes before starting. (Note we only employ a barrier before, | |
216 | * we currently rely on userspace not concurrently starting a new execution | |
217 | * whilst reading or writing to an object. This may be an advantage or not | |
218 | * depending on how much you trust userspace not to shoot themselves in the | |
219 | * foot.) Serialisation may just result in the request being inserted into | |
220 | * a DAG awaiting its turn, but most simple is to wait on the CPU until | |
221 | * all dependencies are resolved. | |
222 | * | |
223 | * After all of that, is just a matter of closing the request and handing it to | |
224 | * the hardware (well, leaving it in a queue to be executed). However, we also | |
225 | * offer the ability for batchbuffers to be run with elevated privileges so | |
226 | * that they access otherwise hidden registers. (Used to adjust L3 cache etc.) | |
227 | * Before any batch is given extra privileges we first must check that it | |
228 | * contains no nefarious instructions, we check that each instruction is from | |
229 | * our whitelist and all registers are also from an allowed list. We first | |
230 | * copy the user's batchbuffer to a shadow (so that the user doesn't have | |
231 | * access to it, either by the CPU or GPU as we scan it) and then parse each | |
232 | * instruction. If everything is ok, we set a flag telling the hardware to run | |
233 | * the batchbuffer in trusted mode, otherwise the ioctl is rejected. | |
234 | */ | |
235 | ||
13149e8b LL |
236 | struct eb_fence { |
237 | struct drm_syncobj *syncobj; /* Use with ptr_mask_bits() */ | |
238 | struct dma_fence *dma_fence; | |
239 | u64 value; | |
240 | struct dma_fence_chain *chain_fence; | |
241 | }; | |
242 | ||
650bc635 | 243 | struct i915_execbuffer { |
2889caa9 CW |
244 | struct drm_i915_private *i915; /** i915 backpointer */ |
245 | struct drm_file *file; /** per-file lookup tables and limits */ | |
246 | struct drm_i915_gem_execbuffer2 *args; /** ioctl parameters */ | |
247 | struct drm_i915_gem_exec_object2 *exec; /** ioctl execobj[] */ | |
7d6236bb | 248 | struct eb_vma *vma; |
2889caa9 | 249 | |
544460c3 | 250 | struct intel_gt *gt; /* gt for the execbuf */ |
8f2a1057 CW |
251 | struct intel_context *context; /* logical state for the request */ |
252 | struct i915_gem_context *gem_context; /** caller's context */ | |
2889caa9 | 253 | |
544460c3 MB |
254 | /** our requests to build */ |
255 | struct i915_request *requests[MAX_ENGINE_INSTANCE + 1]; | |
256 | /** identity of the batch obj/vma */ | |
257 | struct eb_vma *batches[MAX_ENGINE_INSTANCE + 1]; | |
32d94048 | 258 | struct i915_vma *trampoline; /** trampoline used for chaining */ |
2889caa9 | 259 | |
544460c3 MB |
260 | /** used for excl fence in dma_resv objects when > 1 BB submitted */ |
261 | struct dma_fence *composite_fence; | |
262 | ||
2889caa9 CW |
263 | /** actual size of execobj[] as we may extend it for the cmdparser */ |
264 | unsigned int buffer_count; | |
265 | ||
544460c3 MB |
266 | /* number of batches in execbuf IOCTL */ |
267 | unsigned int num_batches; | |
268 | ||
2889caa9 CW |
269 | /** list of vma not yet bound during reservation phase */ |
270 | struct list_head unbound; | |
271 | ||
272 | /** list of vma that have execobj.relocation_count */ | |
273 | struct list_head relocs; | |
274 | ||
c43ce123 ML |
275 | struct i915_gem_ww_ctx ww; |
276 | ||
2889caa9 CW |
277 | /** |
278 | * Track the most recently used object for relocations, as we | |
279 | * frequently have to perform multiple relocations within the same | |
280 | * obj/page | |
281 | */ | |
650bc635 | 282 | struct reloc_cache { |
2889caa9 | 283 | struct drm_mm_node node; /** temporary GTT binding */ |
ad5d95e4 DA |
284 | unsigned long vaddr; /** Current kmap address */ |
285 | unsigned long page; /** Currently mapped page index */ | |
8802190f | 286 | unsigned int graphics_ver; /** Cached value of GRAPHICS_VER */ |
650bc635 | 287 | bool use_64bit_reloc : 1; |
2889caa9 CW |
288 | bool has_llc : 1; |
289 | bool has_fence : 1; | |
290 | bool needs_unfenced : 1; | |
650bc635 | 291 | } reloc_cache; |
2889caa9 CW |
292 | |
293 | u64 invalid_flags; /** Set of execobj.flags that are invalid */ | |
2889caa9 | 294 | |
544460c3 MB |
295 | /** Length of batch within object */ |
296 | u64 batch_len[MAX_ENGINE_INSTANCE + 1]; | |
2889caa9 | 297 | u32 batch_start_offset; /** Location within object of batch */ |
2889caa9 | 298 | u32 batch_flags; /** Flags composed for emit_bb_start() */ |
c43ce123 | 299 | struct intel_gt_buffer_pool_node *batch_pool; /** pool node for batch buffer */ |
2889caa9 CW |
300 | |
301 | /** | |
302 | * Indicate either the size of the hastable used to resolve | |
303 | * relocation handles, or if negative that we are using a direct | |
304 | * index into the execobj[]. | |
305 | */ | |
306 | int lut_size; | |
307 | struct hlist_head *buckets; /** ht for relocation handles */ | |
cda9edd0 | 308 | |
13149e8b LL |
309 | struct eb_fence *fences; |
310 | unsigned long num_fences; | |
ff20afc4 TH |
311 | #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR) |
312 | struct i915_capture_list *capture_lists[MAX_ENGINE_INSTANCE + 1]; | |
313 | #endif | |
67731b87 CW |
314 | }; |
315 | ||
8e4ba491 | 316 | static int eb_parse(struct i915_execbuffer *eb); |
544460c3 | 317 | static int eb_pin_engine(struct i915_execbuffer *eb, bool throttle); |
2bf541ff | 318 | static void eb_unpin_engine(struct i915_execbuffer *eb); |
ff20afc4 | 319 | static void eb_capture_release(struct i915_execbuffer *eb); |
8e4ba491 | 320 | |
3dbf26ed CW |
321 | static inline bool eb_use_cmdparser(const struct i915_execbuffer *eb) |
322 | { | |
544460c3 MB |
323 | return intel_engine_requires_cmd_parser(eb->context->engine) || |
324 | (intel_engine_using_cmd_parser(eb->context->engine) && | |
435e8fc0 | 325 | eb->args->batch_len); |
3dbf26ed CW |
326 | } |
327 | ||
650bc635 | 328 | static int eb_create(struct i915_execbuffer *eb) |
67731b87 | 329 | { |
2889caa9 CW |
330 | if (!(eb->args->flags & I915_EXEC_HANDLE_LUT)) { |
331 | unsigned int size = 1 + ilog2(eb->buffer_count); | |
4ff4b44c | 332 | |
2889caa9 CW |
333 | /* |
334 | * Without a 1:1 association between relocation handles and | |
335 | * the execobject[] index, we instead create a hashtable. | |
336 | * We size it dynamically based on available memory, starting | |
337 | * first with 1:1 assocative hash and scaling back until | |
338 | * the allocation succeeds. | |
339 | * | |
340 | * Later on we use a positive lut_size to indicate we are | |
341 | * using this hashtable, and a negative value to indicate a | |
342 | * direct lookup. | |
343 | */ | |
4ff4b44c | 344 | do { |
0d95c883 | 345 | gfp_t flags; |
4d470f73 CW |
346 | |
347 | /* While we can still reduce the allocation size, don't | |
348 | * raise a warning and allow the allocation to fail. | |
349 | * On the last pass though, we want to try as hard | |
350 | * as possible to perform the allocation and warn | |
351 | * if it fails. | |
352 | */ | |
0ee931c4 | 353 | flags = GFP_KERNEL; |
4d470f73 CW |
354 | if (size > 1) |
355 | flags |= __GFP_NORETRY | __GFP_NOWARN; | |
356 | ||
4ff4b44c | 357 | eb->buckets = kzalloc(sizeof(struct hlist_head) << size, |
4d470f73 | 358 | flags); |
4ff4b44c CW |
359 | if (eb->buckets) |
360 | break; | |
361 | } while (--size); | |
362 | ||
8ae275c2 | 363 | if (unlikely(!size)) |
4d470f73 | 364 | return -ENOMEM; |
eef90ccb | 365 | |
2889caa9 | 366 | eb->lut_size = size; |
650bc635 | 367 | } else { |
2889caa9 | 368 | eb->lut_size = -eb->buffer_count; |
650bc635 | 369 | } |
eef90ccb | 370 | |
650bc635 | 371 | return 0; |
67731b87 CW |
372 | } |
373 | ||
2889caa9 CW |
374 | static bool |
375 | eb_vma_misplaced(const struct drm_i915_gem_exec_object2 *entry, | |
c7c6e46f CW |
376 | const struct i915_vma *vma, |
377 | unsigned int flags) | |
2889caa9 | 378 | { |
2889caa9 CW |
379 | if (vma->node.size < entry->pad_to_size) |
380 | return true; | |
381 | ||
382 | if (entry->alignment && !IS_ALIGNED(vma->node.start, entry->alignment)) | |
383 | return true; | |
384 | ||
c7c6e46f | 385 | if (flags & EXEC_OBJECT_PINNED && |
2889caa9 CW |
386 | vma->node.start != entry->offset) |
387 | return true; | |
388 | ||
c7c6e46f | 389 | if (flags & __EXEC_OBJECT_NEEDS_BIAS && |
2889caa9 CW |
390 | vma->node.start < BATCH_OFFSET_BIAS) |
391 | return true; | |
392 | ||
c7c6e46f | 393 | if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) && |
5f22cc0b | 394 | (vma->node.start + vma->node.size + 4095) >> 32) |
2889caa9 CW |
395 | return true; |
396 | ||
1d033beb CW |
397 | if (flags & __EXEC_OBJECT_NEEDS_MAP && |
398 | !i915_vma_is_map_and_fenceable(vma)) | |
399 | return true; | |
400 | ||
2889caa9 CW |
401 | return false; |
402 | } | |
403 | ||
8a338f4b CW |
404 | static u64 eb_pin_flags(const struct drm_i915_gem_exec_object2 *entry, |
405 | unsigned int exec_flags) | |
406 | { | |
407 | u64 pin_flags = 0; | |
408 | ||
409 | if (exec_flags & EXEC_OBJECT_NEEDS_GTT) | |
410 | pin_flags |= PIN_GLOBAL; | |
411 | ||
412 | /* | |
413 | * Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset, | |
414 | * limit address to the first 4GBs for unflagged objects. | |
415 | */ | |
416 | if (!(exec_flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS)) | |
417 | pin_flags |= PIN_ZONE_4G; | |
418 | ||
419 | if (exec_flags & __EXEC_OBJECT_NEEDS_MAP) | |
420 | pin_flags |= PIN_MAPPABLE; | |
421 | ||
422 | if (exec_flags & EXEC_OBJECT_PINNED) | |
423 | pin_flags |= entry->offset | PIN_OFFSET_FIXED; | |
424 | else if (exec_flags & __EXEC_OBJECT_NEEDS_BIAS) | |
425 | pin_flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS; | |
426 | ||
427 | return pin_flags; | |
428 | } | |
429 | ||
237647f4 | 430 | static inline int |
2889caa9 | 431 | eb_pin_vma(struct i915_execbuffer *eb, |
c7c6e46f | 432 | const struct drm_i915_gem_exec_object2 *entry, |
7d6236bb | 433 | struct eb_vma *ev) |
2889caa9 | 434 | { |
7d6236bb | 435 | struct i915_vma *vma = ev->vma; |
c7c6e46f | 436 | u64 pin_flags; |
237647f4 | 437 | int err; |
2889caa9 | 438 | |
616d9cee | 439 | if (vma->node.size) |
c7c6e46f | 440 | pin_flags = vma->node.start; |
616d9cee | 441 | else |
c7c6e46f | 442 | pin_flags = entry->offset & PIN_OFFSET_MASK; |
616d9cee | 443 | |
c7c6e46f | 444 | pin_flags |= PIN_USER | PIN_NOEVICT | PIN_OFFSET_FIXED; |
7d6236bb | 445 | if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_GTT)) |
c7c6e46f | 446 | pin_flags |= PIN_GLOBAL; |
616d9cee | 447 | |
8a338f4b | 448 | /* Attempt to reuse the current location if available */ |
237647f4 ML |
449 | err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, pin_flags); |
450 | if (err == -EDEADLK) | |
451 | return err; | |
452 | ||
453 | if (unlikely(err)) { | |
8a338f4b | 454 | if (entry->flags & EXEC_OBJECT_PINNED) |
237647f4 | 455 | return err; |
8a338f4b CW |
456 | |
457 | /* Failing that pick any _free_ space if suitable */ | |
237647f4 | 458 | err = i915_vma_pin_ww(vma, &eb->ww, |
47b08693 ML |
459 | entry->pad_to_size, |
460 | entry->alignment, | |
461 | eb_pin_flags(entry, ev->flags) | | |
237647f4 ML |
462 | PIN_USER | PIN_NOEVICT); |
463 | if (unlikely(err)) | |
464 | return err; | |
8a338f4b | 465 | } |
2889caa9 | 466 | |
7d6236bb | 467 | if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) { |
237647f4 ML |
468 | err = i915_vma_pin_fence(vma); |
469 | if (unlikely(err)) { | |
2889caa9 | 470 | i915_vma_unpin(vma); |
237647f4 | 471 | return err; |
2889caa9 CW |
472 | } |
473 | ||
3bd40735 | 474 | if (vma->fence) |
7d6236bb | 475 | ev->flags |= __EXEC_OBJECT_HAS_FENCE; |
2889caa9 CW |
476 | } |
477 | ||
7d6236bb | 478 | ev->flags |= __EXEC_OBJECT_HAS_PIN; |
237647f4 ML |
479 | if (eb_vma_misplaced(entry, vma, ev->flags)) |
480 | return -EBADSLT; | |
481 | ||
482 | return 0; | |
2889caa9 CW |
483 | } |
484 | ||
8ae275c2 ML |
485 | static inline void |
486 | eb_unreserve_vma(struct eb_vma *ev) | |
487 | { | |
488 | if (!(ev->flags & __EXEC_OBJECT_HAS_PIN)) | |
489 | return; | |
490 | ||
c43ce123 ML |
491 | if (unlikely(ev->flags & __EXEC_OBJECT_HAS_FENCE)) |
492 | __i915_vma_unpin_fence(ev->vma); | |
493 | ||
494 | __i915_vma_unpin(ev->vma); | |
8ae275c2 ML |
495 | ev->flags &= ~__EXEC_OBJECT_RESERVED; |
496 | } | |
497 | ||
2889caa9 CW |
498 | static int |
499 | eb_validate_vma(struct i915_execbuffer *eb, | |
500 | struct drm_i915_gem_exec_object2 *entry, | |
501 | struct i915_vma *vma) | |
67731b87 | 502 | { |
2eb8e1a6 JE |
503 | /* Relocations are disallowed for all platforms after TGL-LP. This |
504 | * also covers all platforms with local memory. | |
505 | */ | |
506 | if (entry->relocation_count && | |
40e1956e | 507 | GRAPHICS_VER(eb->i915) >= 12 && !IS_TIGERLAKE(eb->i915)) |
2eb8e1a6 JE |
508 | return -EINVAL; |
509 | ||
2889caa9 CW |
510 | if (unlikely(entry->flags & eb->invalid_flags)) |
511 | return -EINVAL; | |
d55495b4 | 512 | |
2920516b MA |
513 | if (unlikely(entry->alignment && |
514 | !is_power_of_2_u64(entry->alignment))) | |
2889caa9 CW |
515 | return -EINVAL; |
516 | ||
517 | /* | |
518 | * Offset can be used as input (EXEC_OBJECT_PINNED), reject | |
519 | * any non-page-aligned or non-canonical addresses. | |
520 | */ | |
521 | if (unlikely(entry->flags & EXEC_OBJECT_PINNED && | |
6fc4e48f | 522 | entry->offset != gen8_canonical_addr(entry->offset & I915_GTT_PAGE_MASK))) |
2889caa9 CW |
523 | return -EINVAL; |
524 | ||
525 | /* pad_to_size was once a reserved field, so sanitize it */ | |
526 | if (entry->flags & EXEC_OBJECT_PAD_TO_SIZE) { | |
527 | if (unlikely(offset_in_page(entry->pad_to_size))) | |
528 | return -EINVAL; | |
529 | } else { | |
530 | entry->pad_to_size = 0; | |
d55495b4 | 531 | } |
2889caa9 CW |
532 | /* |
533 | * From drm_mm perspective address space is continuous, | |
534 | * so from this point we're always using non-canonical | |
535 | * form internally. | |
536 | */ | |
537 | entry->offset = gen8_noncanonical_addr(entry->offset); | |
538 | ||
c7c6e46f CW |
539 | if (!eb->reloc_cache.has_fence) { |
540 | entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE; | |
541 | } else { | |
542 | if ((entry->flags & EXEC_OBJECT_NEEDS_FENCE || | |
543 | eb->reloc_cache.needs_unfenced) && | |
544 | i915_gem_object_is_tiled(vma->obj)) | |
545 | entry->flags |= EXEC_OBJECT_NEEDS_GTT | __EXEC_OBJECT_NEEDS_MAP; | |
546 | } | |
547 | ||
2889caa9 | 548 | return 0; |
67731b87 CW |
549 | } |
550 | ||
544460c3 MB |
551 | static inline bool |
552 | is_batch_buffer(struct i915_execbuffer *eb, unsigned int buffer_idx) | |
553 | { | |
554 | return eb->args->flags & I915_EXEC_BATCH_FIRST ? | |
555 | buffer_idx < eb->num_batches : | |
556 | buffer_idx >= eb->args->buffer_count - eb->num_batches; | |
557 | } | |
558 | ||
559 | static int | |
746c8f14 | 560 | eb_add_vma(struct i915_execbuffer *eb, |
544460c3 MB |
561 | unsigned int *current_batch, |
562 | unsigned int i, | |
746c8f14 | 563 | struct i915_vma *vma) |
59bfa124 | 564 | { |
544460c3 | 565 | struct drm_i915_private *i915 = eb->i915; |
c7c6e46f | 566 | struct drm_i915_gem_exec_object2 *entry = &eb->exec[i]; |
7d6236bb | 567 | struct eb_vma *ev = &eb->vma[i]; |
2889caa9 | 568 | |
93159e12 | 569 | ev->vma = vma; |
7d6236bb CW |
570 | ev->exec = entry; |
571 | ev->flags = entry->flags; | |
572 | ||
4d470f73 | 573 | if (eb->lut_size > 0) { |
7d6236bb CW |
574 | ev->handle = entry->handle; |
575 | hlist_add_head(&ev->node, | |
2889caa9 CW |
576 | &eb->buckets[hash_32(entry->handle, |
577 | eb->lut_size)]); | |
4ff4b44c | 578 | } |
59bfa124 | 579 | |
2889caa9 | 580 | if (entry->relocation_count) |
7d6236bb | 581 | list_add_tail(&ev->reloc_link, &eb->relocs); |
2889caa9 | 582 | |
746c8f14 CW |
583 | /* |
584 | * SNA is doing fancy tricks with compressing batch buffers, which leads | |
585 | * to negative relocation deltas. Usually that works out ok since the | |
586 | * relocate address is still positive, except when the batch is placed | |
587 | * very low in the GTT. Ensure this doesn't happen. | |
588 | * | |
589 | * Note that actual hangs have only been observed on gen7, but for | |
590 | * paranoia do it everywhere. | |
591 | */ | |
544460c3 | 592 | if (is_batch_buffer(eb, i)) { |
827db9d8 | 593 | if (entry->relocation_count && |
7d6236bb CW |
594 | !(ev->flags & EXEC_OBJECT_PINNED)) |
595 | ev->flags |= __EXEC_OBJECT_NEEDS_BIAS; | |
746c8f14 | 596 | if (eb->reloc_cache.has_fence) |
7d6236bb | 597 | ev->flags |= EXEC_OBJECT_NEEDS_FENCE; |
746c8f14 | 598 | |
544460c3 MB |
599 | eb->batches[*current_batch] = ev; |
600 | ||
601 | if (unlikely(ev->flags & EXEC_OBJECT_WRITE)) { | |
602 | drm_dbg(&i915->drm, | |
603 | "Attempting to use self-modifying batch buffer\n"); | |
604 | return -EINVAL; | |
605 | } | |
606 | ||
607 | if (range_overflows_t(u64, | |
608 | eb->batch_start_offset, | |
609 | eb->args->batch_len, | |
610 | ev->vma->size)) { | |
611 | drm_dbg(&i915->drm, "Attempting to use out-of-bounds batch\n"); | |
612 | return -EINVAL; | |
613 | } | |
614 | ||
615 | if (eb->args->batch_len == 0) | |
616 | eb->batch_len[*current_batch] = ev->vma->size - | |
617 | eb->batch_start_offset; | |
618 | else | |
619 | eb->batch_len[*current_batch] = eb->args->batch_len; | |
620 | if (unlikely(eb->batch_len[*current_batch] == 0)) { /* impossible! */ | |
621 | drm_dbg(&i915->drm, "Invalid batch length\n"); | |
622 | return -EINVAL; | |
623 | } | |
624 | ||
625 | ++*current_batch; | |
746c8f14 | 626 | } |
544460c3 MB |
627 | |
628 | return 0; | |
2889caa9 CW |
629 | } |
630 | ||
ad5d95e4 DA |
631 | static inline int use_cpu_reloc(const struct reloc_cache *cache, |
632 | const struct drm_i915_gem_object *obj) | |
633 | { | |
634 | if (!i915_gem_object_has_struct_page(obj)) | |
635 | return false; | |
636 | ||
637 | if (DBG_FORCE_RELOC == FORCE_CPU_RELOC) | |
638 | return true; | |
639 | ||
640 | if (DBG_FORCE_RELOC == FORCE_GTT_RELOC) | |
641 | return false; | |
642 | ||
643 | return (cache->has_llc || | |
644 | obj->cache_dirty || | |
645 | obj->cache_level != I915_CACHE_NONE); | |
646 | } | |
647 | ||
47b08693 | 648 | static int eb_reserve_vma(struct i915_execbuffer *eb, |
7d6236bb | 649 | struct eb_vma *ev, |
2920bb94 | 650 | u64 pin_flags) |
2889caa9 | 651 | { |
7d6236bb | 652 | struct drm_i915_gem_exec_object2 *entry = ev->exec; |
7d6236bb | 653 | struct i915_vma *vma = ev->vma; |
2889caa9 CW |
654 | int err; |
655 | ||
003d8b91 CW |
656 | if (drm_mm_node_allocated(&vma->node) && |
657 | eb_vma_misplaced(entry, vma, ev->flags)) { | |
658 | err = i915_vma_unbind(vma); | |
659 | if (err) | |
660 | return err; | |
661 | } | |
662 | ||
47b08693 | 663 | err = i915_vma_pin_ww(vma, &eb->ww, |
c7c6e46f | 664 | entry->pad_to_size, entry->alignment, |
8a338f4b | 665 | eb_pin_flags(entry, ev->flags) | pin_flags); |
2889caa9 CW |
666 | if (err) |
667 | return err; | |
668 | ||
669 | if (entry->offset != vma->node.start) { | |
670 | entry->offset = vma->node.start | UPDATE; | |
671 | eb->args->flags |= __EXEC_HAS_RELOC; | |
672 | } | |
673 | ||
8a338f4b | 674 | if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) { |
3bd40735 | 675 | err = i915_vma_pin_fence(vma); |
2889caa9 CW |
676 | if (unlikely(err)) { |
677 | i915_vma_unpin(vma); | |
678 | return err; | |
679 | } | |
680 | ||
3bd40735 | 681 | if (vma->fence) |
8a338f4b | 682 | ev->flags |= __EXEC_OBJECT_HAS_FENCE; |
2889caa9 CW |
683 | } |
684 | ||
8a338f4b | 685 | ev->flags |= __EXEC_OBJECT_HAS_PIN; |
7d6236bb | 686 | GEM_BUG_ON(eb_vma_misplaced(entry, vma, ev->flags)); |
1da7b54c | 687 | |
2889caa9 CW |
688 | return 0; |
689 | } | |
690 | ||
691 | static int eb_reserve(struct i915_execbuffer *eb) | |
692 | { | |
693 | const unsigned int count = eb->buffer_count; | |
2920bb94 | 694 | unsigned int pin_flags = PIN_USER | PIN_NONBLOCK; |
2889caa9 | 695 | struct list_head last; |
7d6236bb | 696 | struct eb_vma *ev; |
2889caa9 | 697 | unsigned int i, pass; |
ef398881 | 698 | int err = 0; |
2889caa9 CW |
699 | |
700 | /* | |
701 | * Attempt to pin all of the buffers into the GTT. | |
702 | * This is done in 3 phases: | |
703 | * | |
704 | * 1a. Unbind all objects that do not match the GTT constraints for | |
705 | * the execbuffer (fenceable, mappable, alignment etc). | |
706 | * 1b. Increment pin count for already bound objects. | |
707 | * 2. Bind new objects. | |
708 | * 3. Decrement pin count. | |
709 | * | |
710 | * This avoid unnecessary unbinding of later objects in order to make | |
711 | * room for the earlier objects *unless* we need to defragment. | |
712 | */ | |
2889caa9 | 713 | pass = 0; |
2889caa9 | 714 | do { |
7d6236bb CW |
715 | list_for_each_entry(ev, &eb->unbound, bind_link) { |
716 | err = eb_reserve_vma(eb, ev, pin_flags); | |
2889caa9 CW |
717 | if (err) |
718 | break; | |
719 | } | |
fd1500fc | 720 | if (err != -ENOSPC) |
c43ce123 | 721 | return err; |
2889caa9 CW |
722 | |
723 | /* Resort *all* the objects into priority order */ | |
724 | INIT_LIST_HEAD(&eb->unbound); | |
725 | INIT_LIST_HEAD(&last); | |
726 | for (i = 0; i < count; i++) { | |
7d6236bb | 727 | unsigned int flags; |
2889caa9 | 728 | |
7d6236bb CW |
729 | ev = &eb->vma[i]; |
730 | flags = ev->flags; | |
c7c6e46f CW |
731 | if (flags & EXEC_OBJECT_PINNED && |
732 | flags & __EXEC_OBJECT_HAS_PIN) | |
2889caa9 CW |
733 | continue; |
734 | ||
7d6236bb | 735 | eb_unreserve_vma(ev); |
2889caa9 | 736 | |
c7c6e46f | 737 | if (flags & EXEC_OBJECT_PINNED) |
35e882a4 | 738 | /* Pinned must have their slot */ |
7d6236bb | 739 | list_add(&ev->bind_link, &eb->unbound); |
c7c6e46f | 740 | else if (flags & __EXEC_OBJECT_NEEDS_MAP) |
35e882a4 | 741 | /* Map require the lowest 256MiB (aperture) */ |
7d6236bb | 742 | list_add_tail(&ev->bind_link, &eb->unbound); |
35e882a4 CW |
743 | else if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS)) |
744 | /* Prioritise 4GiB region for restricted bo */ | |
7d6236bb | 745 | list_add(&ev->bind_link, &last); |
2889caa9 | 746 | else |
7d6236bb | 747 | list_add_tail(&ev->bind_link, &last); |
2889caa9 CW |
748 | } |
749 | list_splice_tail(&last, &eb->unbound); | |
750 | ||
751 | switch (pass++) { | |
752 | case 0: | |
753 | break; | |
754 | ||
755 | case 1: | |
756 | /* Too fragmented, unbind everything and retry */ | |
2850748e | 757 | mutex_lock(&eb->context->vm->mutex); |
f5d974f9 | 758 | err = i915_gem_evict_vm(eb->context->vm); |
2850748e | 759 | mutex_unlock(&eb->context->vm->mutex); |
2889caa9 | 760 | if (err) |
c43ce123 | 761 | return err; |
2889caa9 CW |
762 | break; |
763 | ||
764 | default: | |
c43ce123 | 765 | return -ENOSPC; |
2889caa9 | 766 | } |
2920bb94 CW |
767 | |
768 | pin_flags = PIN_USER; | |
2889caa9 | 769 | } while (1); |
4ff4b44c | 770 | } |
59bfa124 | 771 | |
2889caa9 CW |
772 | static int eb_select_context(struct i915_execbuffer *eb) |
773 | { | |
774 | struct i915_gem_context *ctx; | |
775 | ||
776 | ctx = i915_gem_context_lookup(eb->file->driver_priv, eb->args->rsvd1); | |
046d1660 JE |
777 | if (unlikely(IS_ERR(ctx))) |
778 | return PTR_ERR(ctx); | |
2889caa9 | 779 | |
8f2a1057 | 780 | eb->gem_context = ctx; |
a82a9979 | 781 | if (i915_gem_context_has_full_ppgtt(ctx)) |
4f2c7337 | 782 | eb->invalid_flags |= EXEC_OBJECT_NEEDS_GTT; |
2889caa9 | 783 | |
2889caa9 CW |
784 | return 0; |
785 | } | |
786 | ||
93159e12 CW |
787 | static int __eb_add_lut(struct i915_execbuffer *eb, |
788 | u32 handle, struct i915_vma *vma) | |
3b96eff4 | 789 | { |
93159e12 CW |
790 | struct i915_gem_context *ctx = eb->gem_context; |
791 | struct i915_lut_handle *lut; | |
2889caa9 | 792 | int err; |
3b96eff4 | 793 | |
93159e12 CW |
794 | lut = i915_lut_handle_alloc(); |
795 | if (unlikely(!lut)) | |
796 | return -ENOMEM; | |
797 | ||
798 | i915_vma_get(vma); | |
799 | if (!atomic_fetch_inc(&vma->open_count)) | |
800 | i915_vma_reopen(vma); | |
801 | lut->handle = handle; | |
802 | lut->ctx = ctx; | |
803 | ||
804 | /* Check that the context hasn't been closed in the meantime */ | |
805 | err = -EINTR; | |
f7ce8639 | 806 | if (!mutex_lock_interruptible(&ctx->lut_mutex)) { |
e1068a9e | 807 | if (likely(!i915_gem_context_is_closed(ctx))) |
93159e12 | 808 | err = radix_tree_insert(&ctx->handles_vma, handle, vma); |
f7ce8639 CW |
809 | else |
810 | err = -ENOENT; | |
93159e12 CW |
811 | if (err == 0) { /* And nor has this handle */ |
812 | struct drm_i915_gem_object *obj = vma->obj; | |
813 | ||
096a42dd | 814 | spin_lock(&obj->lut_lock); |
93159e12 CW |
815 | if (idr_find(&eb->file->object_idr, handle) == obj) { |
816 | list_add(&lut->obj_link, &obj->lut_list); | |
817 | } else { | |
818 | radix_tree_delete(&ctx->handles_vma, handle); | |
819 | err = -ENOENT; | |
820 | } | |
096a42dd | 821 | spin_unlock(&obj->lut_lock); |
93159e12 | 822 | } |
f7ce8639 | 823 | mutex_unlock(&ctx->lut_mutex); |
93159e12 CW |
824 | } |
825 | if (unlikely(err)) | |
826 | goto err; | |
003d8b91 | 827 | |
93159e12 | 828 | return 0; |
d55495b4 | 829 | |
93159e12 | 830 | err: |
50689771 | 831 | i915_vma_close(vma); |
93159e12 CW |
832 | i915_vma_put(vma); |
833 | i915_lut_handle_free(lut); | |
834 | return err; | |
835 | } | |
746c8f14 | 836 | |
93159e12 CW |
837 | static struct i915_vma *eb_lookup_vma(struct i915_execbuffer *eb, u32 handle) |
838 | { | |
f7ce8639 CW |
839 | struct i915_address_space *vm = eb->context->vm; |
840 | ||
93159e12 CW |
841 | do { |
842 | struct drm_i915_gem_object *obj; | |
170fa29b | 843 | struct i915_vma *vma; |
93159e12 | 844 | int err; |
4ff4b44c | 845 | |
93159e12 CW |
846 | rcu_read_lock(); |
847 | vma = radix_tree_lookup(&eb->gem_context->handles_vma, handle); | |
f7ce8639 | 848 | if (likely(vma && vma->vm == vm)) |
93159e12 CW |
849 | vma = i915_vma_tryget(vma); |
850 | rcu_read_unlock(); | |
851 | if (likely(vma)) | |
852 | return vma; | |
4ff4b44c | 853 | |
170fa29b | 854 | obj = i915_gem_object_lookup(eb->file, handle); |
93159e12 CW |
855 | if (unlikely(!obj)) |
856 | return ERR_PTR(-ENOENT); | |
3b96eff4 | 857 | |
d3ac8d42 DCS |
858 | /* |
859 | * If the user has opted-in for protected-object tracking, make | |
860 | * sure the object encryption can be used. | |
861 | * We only need to do this when the object is first used with | |
862 | * this context, because the context itself will be banned when | |
863 | * the protected objects become invalid. | |
864 | */ | |
865 | if (i915_gem_context_uses_protected_content(eb->gem_context) && | |
866 | i915_gem_object_is_protected(obj)) { | |
ef6ba31d | 867 | err = intel_pxp_key_check(&vm->gt->pxp, obj, true); |
d3ac8d42 DCS |
868 | if (err) { |
869 | i915_gem_object_put(obj); | |
870 | return ERR_PTR(err); | |
871 | } | |
872 | } | |
873 | ||
f7ce8639 | 874 | vma = i915_vma_instance(obj, vm, NULL); |
772b5408 | 875 | if (IS_ERR(vma)) { |
93159e12 CW |
876 | i915_gem_object_put(obj); |
877 | return vma; | |
27173f1f BW |
878 | } |
879 | ||
93159e12 CW |
880 | err = __eb_add_lut(eb, handle, vma); |
881 | if (likely(!err)) | |
882 | return vma; | |
d1b48c1e | 883 | |
93159e12 CW |
884 | i915_gem_object_put(obj); |
885 | if (err != -EEXIST) | |
886 | return ERR_PTR(err); | |
887 | } while (1); | |
888 | } | |
4ff4b44c | 889 | |
93159e12 CW |
890 | static int eb_lookup_vmas(struct i915_execbuffer *eb) |
891 | { | |
544460c3 | 892 | unsigned int i, current_batch = 0; |
93159e12 | 893 | int err = 0; |
155ab883 | 894 | |
93159e12 | 895 | INIT_LIST_HEAD(&eb->relocs); |
93159e12 CW |
896 | |
897 | for (i = 0; i < eb->buffer_count; i++) { | |
898 | struct i915_vma *vma; | |
899 | ||
900 | vma = eb_lookup_vma(eb, eb->exec[i].handle); | |
901 | if (IS_ERR(vma)) { | |
902 | err = PTR_ERR(vma); | |
8e4ba491 | 903 | goto err; |
93159e12 | 904 | } |
d1b48c1e | 905 | |
003d8b91 | 906 | err = eb_validate_vma(eb, &eb->exec[i], vma); |
93159e12 CW |
907 | if (unlikely(err)) { |
908 | i915_vma_put(vma); | |
8e4ba491 | 909 | goto err; |
93159e12 | 910 | } |
dade2a61 | 911 | |
544460c3 MB |
912 | err = eb_add_vma(eb, ¤t_batch, i, vma); |
913 | if (err) | |
914 | return err; | |
ed29c269 ML |
915 | |
916 | if (i915_gem_object_is_userptr(vma->obj)) { | |
917 | err = i915_gem_object_userptr_submit_init(vma->obj); | |
918 | if (err) { | |
919 | if (i + 1 < eb->buffer_count) { | |
920 | /* | |
921 | * Execbuffer code expects last vma entry to be NULL, | |
922 | * since we already initialized this entry, | |
923 | * set the next value to NULL or we mess up | |
924 | * cleanup handling. | |
925 | */ | |
926 | eb->vma[i + 1].vma = NULL; | |
927 | } | |
928 | ||
929 | return err; | |
930 | } | |
931 | ||
932 | eb->vma[i].flags |= __EXEC_OBJECT_USERPTR_INIT; | |
933 | eb->args->flags |= __EXEC_USERPTR_USED; | |
934 | } | |
4ff4b44c CW |
935 | } |
936 | ||
8e4ba491 ML |
937 | return 0; |
938 | ||
939 | err: | |
7d6236bb | 940 | eb->vma[i].vma = NULL; |
2889caa9 | 941 | return err; |
3b96eff4 CW |
942 | } |
943 | ||
5cd57f67 | 944 | static int eb_lock_vmas(struct i915_execbuffer *eb) |
c43ce123 ML |
945 | { |
946 | unsigned int i; | |
947 | int err; | |
948 | ||
c43ce123 | 949 | for (i = 0; i < eb->buffer_count; i++) { |
c43ce123 ML |
950 | struct eb_vma *ev = &eb->vma[i]; |
951 | struct i915_vma *vma = ev->vma; | |
952 | ||
953 | err = i915_gem_object_lock(vma->obj, &eb->ww); | |
954 | if (err) | |
955 | return err; | |
5cd57f67 TH |
956 | } |
957 | ||
958 | return 0; | |
959 | } | |
960 | ||
961 | static int eb_validate_vmas(struct i915_execbuffer *eb) | |
962 | { | |
963 | unsigned int i; | |
964 | int err; | |
965 | ||
966 | INIT_LIST_HEAD(&eb->unbound); | |
967 | ||
968 | err = eb_lock_vmas(eb); | |
969 | if (err) | |
970 | return err; | |
971 | ||
972 | for (i = 0; i < eb->buffer_count; i++) { | |
973 | struct drm_i915_gem_exec_object2 *entry = &eb->exec[i]; | |
974 | struct eb_vma *ev = &eb->vma[i]; | |
975 | struct i915_vma *vma = ev->vma; | |
c43ce123 | 976 | |
237647f4 ML |
977 | err = eb_pin_vma(eb, entry, ev); |
978 | if (err == -EDEADLK) | |
979 | return err; | |
980 | ||
981 | if (!err) { | |
c43ce123 ML |
982 | if (entry->offset != vma->node.start) { |
983 | entry->offset = vma->node.start | UPDATE; | |
984 | eb->args->flags |= __EXEC_HAS_RELOC; | |
985 | } | |
986 | } else { | |
987 | eb_unreserve_vma(ev); | |
988 | ||
989 | list_add_tail(&ev->bind_link, &eb->unbound); | |
990 | if (drm_mm_node_allocated(&vma->node)) { | |
991 | err = i915_vma_unbind(vma); | |
992 | if (err) | |
993 | return err; | |
994 | } | |
995 | } | |
996 | ||
bfaae47d | 997 | if (!(ev->flags & EXEC_OBJECT_WRITE)) { |
95c3d275 | 998 | err = dma_resv_reserve_shared(vma->obj->base.resv, 1); |
bfaae47d ML |
999 | if (err) |
1000 | return err; | |
1001 | } | |
1002 | ||
c43ce123 ML |
1003 | GEM_BUG_ON(drm_mm_node_allocated(&vma->node) && |
1004 | eb_vma_misplaced(&eb->exec[i], vma, ev->flags)); | |
1005 | } | |
1006 | ||
1007 | if (!list_empty(&eb->unbound)) | |
1008 | return eb_reserve(eb); | |
1009 | ||
1010 | return 0; | |
1011 | } | |
1012 | ||
7d6236bb | 1013 | static struct eb_vma * |
2889caa9 | 1014 | eb_get_vma(const struct i915_execbuffer *eb, unsigned long handle) |
67731b87 | 1015 | { |
2889caa9 CW |
1016 | if (eb->lut_size < 0) { |
1017 | if (handle >= -eb->lut_size) | |
eef90ccb | 1018 | return NULL; |
7d6236bb | 1019 | return &eb->vma[handle]; |
eef90ccb CW |
1020 | } else { |
1021 | struct hlist_head *head; | |
7d6236bb | 1022 | struct eb_vma *ev; |
67731b87 | 1023 | |
2889caa9 | 1024 | head = &eb->buckets[hash_32(handle, eb->lut_size)]; |
7d6236bb CW |
1025 | hlist_for_each_entry(ev, head, node) { |
1026 | if (ev->handle == handle) | |
1027 | return ev; | |
eef90ccb CW |
1028 | } |
1029 | return NULL; | |
1030 | } | |
67731b87 CW |
1031 | } |
1032 | ||
b4b9731b | 1033 | static void eb_release_vmas(struct i915_execbuffer *eb, bool final) |
8ae275c2 ML |
1034 | { |
1035 | const unsigned int count = eb->buffer_count; | |
1036 | unsigned int i; | |
1037 | ||
1038 | for (i = 0; i < count; i++) { | |
1039 | struct eb_vma *ev = &eb->vma[i]; | |
1040 | struct i915_vma *vma = ev->vma; | |
1041 | ||
1042 | if (!vma) | |
1043 | break; | |
1044 | ||
c43ce123 | 1045 | eb_unreserve_vma(ev); |
8ae275c2 | 1046 | |
c43ce123 ML |
1047 | if (final) |
1048 | i915_vma_put(vma); | |
8ae275c2 | 1049 | } |
2bf541ff | 1050 | |
ff20afc4 | 1051 | eb_capture_release(eb); |
2bf541ff | 1052 | eb_unpin_engine(eb); |
8ae275c2 ML |
1053 | } |
1054 | ||
2889caa9 | 1055 | static void eb_destroy(const struct i915_execbuffer *eb) |
934acce3 | 1056 | { |
4d470f73 | 1057 | if (eb->lut_size > 0) |
2889caa9 | 1058 | kfree(eb->buckets); |
934acce3 MW |
1059 | } |
1060 | ||
2889caa9 | 1061 | static inline u64 |
d50415cc | 1062 | relocation_target(const struct drm_i915_gem_relocation_entry *reloc, |
2889caa9 | 1063 | const struct i915_vma *target) |
934acce3 | 1064 | { |
2889caa9 | 1065 | return gen8_canonical_addr((int)reloc->delta + target->node.start); |
934acce3 MW |
1066 | } |
1067 | ||
d50415cc CW |
1068 | static void reloc_cache_init(struct reloc_cache *cache, |
1069 | struct drm_i915_private *i915) | |
5032d871 | 1070 | { |
ad5d95e4 DA |
1071 | cache->page = -1; |
1072 | cache->vaddr = 0; | |
dfc5148f | 1073 | /* Must be a variable in the struct to allow GCC to unroll. */ |
8802190f | 1074 | cache->graphics_ver = GRAPHICS_VER(i915); |
2889caa9 | 1075 | cache->has_llc = HAS_LLC(i915); |
dfc5148f | 1076 | cache->use_64bit_reloc = HAS_64BIT_RELOC(i915); |
8802190f | 1077 | cache->has_fence = cache->graphics_ver < 4; |
7dd4f672 | 1078 | cache->needs_unfenced = INTEL_INFO(i915)->unfenced_needs_alignment; |
4ee92c71 | 1079 | cache->node.flags = 0; |
d50415cc | 1080 | } |
5032d871 | 1081 | |
20561da3 DA |
1082 | static inline void *unmask_page(unsigned long p) |
1083 | { | |
1084 | return (void *)(uintptr_t)(p & PAGE_MASK); | |
1085 | } | |
1086 | ||
1087 | static inline unsigned int unmask_flags(unsigned long p) | |
1088 | { | |
1089 | return p & ~PAGE_MASK; | |
1090 | } | |
1091 | ||
1092 | #define KMAP 0x4 /* after CLFLUSH_FLAGS */ | |
1093 | ||
1094 | static inline struct i915_ggtt *cache_to_ggtt(struct reloc_cache *cache) | |
1095 | { | |
1096 | struct drm_i915_private *i915 = | |
1097 | container_of(cache, struct i915_execbuffer, reloc_cache)->i915; | |
5c24c9d2 | 1098 | return to_gt(i915)->ggtt; |
20561da3 DA |
1099 | } |
1100 | ||
c2ea703d TH |
1101 | static void reloc_cache_unmap(struct reloc_cache *cache) |
1102 | { | |
1103 | void *vaddr; | |
1104 | ||
1105 | if (!cache->vaddr) | |
1106 | return; | |
1107 | ||
1108 | vaddr = unmask_page(cache->vaddr); | |
1109 | if (cache->vaddr & KMAP) | |
1110 | kunmap_atomic(vaddr); | |
1111 | else | |
1112 | io_mapping_unmap_atomic((void __iomem *)vaddr); | |
1113 | } | |
1114 | ||
1115 | static void reloc_cache_remap(struct reloc_cache *cache, | |
1116 | struct drm_i915_gem_object *obj) | |
1117 | { | |
1118 | void *vaddr; | |
1119 | ||
1120 | if (!cache->vaddr) | |
1121 | return; | |
1122 | ||
1123 | if (cache->vaddr & KMAP) { | |
1124 | struct page *page = i915_gem_object_get_page(obj, cache->page); | |
1125 | ||
1126 | vaddr = kmap_atomic(page); | |
1127 | cache->vaddr = unmask_flags(cache->vaddr) | | |
1128 | (unsigned long)vaddr; | |
1129 | } else { | |
1130 | struct i915_ggtt *ggtt = cache_to_ggtt(cache); | |
1131 | unsigned long offset; | |
1132 | ||
1133 | offset = cache->node.start; | |
1134 | if (!drm_mm_node_allocated(&cache->node)) | |
1135 | offset += cache->page << PAGE_SHIFT; | |
1136 | ||
1137 | cache->vaddr = (unsigned long) | |
1138 | io_mapping_map_atomic_wc(&ggtt->iomap, offset); | |
1139 | } | |
1140 | } | |
1141 | ||
c43ce123 | 1142 | static void reloc_cache_reset(struct reloc_cache *cache, struct i915_execbuffer *eb) |
ad5d95e4 DA |
1143 | { |
1144 | void *vaddr; | |
1145 | ||
1146 | if (!cache->vaddr) | |
1147 | return; | |
1148 | ||
1149 | vaddr = unmask_page(cache->vaddr); | |
1150 | if (cache->vaddr & KMAP) { | |
1af343cd ML |
1151 | struct drm_i915_gem_object *obj = |
1152 | (struct drm_i915_gem_object *)cache->node.mm; | |
ad5d95e4 DA |
1153 | if (cache->vaddr & CLFLUSH_AFTER) |
1154 | mb(); | |
1155 | ||
1156 | kunmap_atomic(vaddr); | |
1af343cd | 1157 | i915_gem_object_finish_access(obj); |
ad5d95e4 DA |
1158 | } else { |
1159 | struct i915_ggtt *ggtt = cache_to_ggtt(cache); | |
1160 | ||
1161 | intel_gt_flush_ggtt_writes(ggtt->vm.gt); | |
1162 | io_mapping_unmap_atomic((void __iomem *)vaddr); | |
1163 | ||
1164 | if (drm_mm_node_allocated(&cache->node)) { | |
1165 | ggtt->vm.clear_range(&ggtt->vm, | |
1166 | cache->node.start, | |
1167 | cache->node.size); | |
1168 | mutex_lock(&ggtt->vm.mutex); | |
1169 | drm_mm_remove_node(&cache->node); | |
1170 | mutex_unlock(&ggtt->vm.mutex); | |
1171 | } else { | |
1172 | i915_vma_unpin((struct i915_vma *)cache->node.mm); | |
1173 | } | |
1174 | } | |
1175 | ||
1176 | cache->vaddr = 0; | |
1177 | cache->page = -1; | |
1178 | } | |
1179 | ||
1180 | static void *reloc_kmap(struct drm_i915_gem_object *obj, | |
1181 | struct reloc_cache *cache, | |
102a0a90 | 1182 | unsigned long pageno) |
ad5d95e4 DA |
1183 | { |
1184 | void *vaddr; | |
102a0a90 | 1185 | struct page *page; |
ad5d95e4 DA |
1186 | |
1187 | if (cache->vaddr) { | |
1188 | kunmap_atomic(unmask_page(cache->vaddr)); | |
1189 | } else { | |
1190 | unsigned int flushes; | |
1191 | int err; | |
1192 | ||
1193 | err = i915_gem_object_prepare_write(obj, &flushes); | |
1194 | if (err) | |
1195 | return ERR_PTR(err); | |
1196 | ||
1197 | BUILD_BUG_ON(KMAP & CLFLUSH_FLAGS); | |
1198 | BUILD_BUG_ON((KMAP | CLFLUSH_FLAGS) & PAGE_MASK); | |
1199 | ||
1200 | cache->vaddr = flushes | KMAP; | |
1201 | cache->node.mm = (void *)obj; | |
1202 | if (flushes) | |
1203 | mb(); | |
1204 | } | |
1205 | ||
102a0a90 ML |
1206 | page = i915_gem_object_get_page(obj, pageno); |
1207 | if (!obj->mm.dirty) | |
1208 | set_page_dirty(page); | |
1209 | ||
1210 | vaddr = kmap_atomic(page); | |
ad5d95e4 | 1211 | cache->vaddr = unmask_flags(cache->vaddr) | (unsigned long)vaddr; |
102a0a90 | 1212 | cache->page = pageno; |
ad5d95e4 DA |
1213 | |
1214 | return vaddr; | |
1215 | } | |
1216 | ||
1217 | static void *reloc_iomap(struct drm_i915_gem_object *obj, | |
47b08693 | 1218 | struct i915_execbuffer *eb, |
ad5d95e4 DA |
1219 | unsigned long page) |
1220 | { | |
47b08693 | 1221 | struct reloc_cache *cache = &eb->reloc_cache; |
ad5d95e4 DA |
1222 | struct i915_ggtt *ggtt = cache_to_ggtt(cache); |
1223 | unsigned long offset; | |
1224 | void *vaddr; | |
1225 | ||
1226 | if (cache->vaddr) { | |
1227 | intel_gt_flush_ggtt_writes(ggtt->vm.gt); | |
1228 | io_mapping_unmap_atomic((void __force __iomem *) unmask_page(cache->vaddr)); | |
1229 | } else { | |
1230 | struct i915_vma *vma; | |
1231 | int err; | |
1232 | ||
1233 | if (i915_gem_object_is_tiled(obj)) | |
1234 | return ERR_PTR(-EINVAL); | |
1235 | ||
1236 | if (use_cpu_reloc(cache, obj)) | |
1237 | return NULL; | |
1238 | ||
ad5d95e4 | 1239 | err = i915_gem_object_set_to_gtt_domain(obj, true); |
ad5d95e4 DA |
1240 | if (err) |
1241 | return ERR_PTR(err); | |
1242 | ||
47b08693 ML |
1243 | vma = i915_gem_object_ggtt_pin_ww(obj, &eb->ww, NULL, 0, 0, |
1244 | PIN_MAPPABLE | | |
1245 | PIN_NONBLOCK /* NOWARN */ | | |
1246 | PIN_NOEVICT); | |
1247 | if (vma == ERR_PTR(-EDEADLK)) | |
1248 | return vma; | |
1249 | ||
ad5d95e4 DA |
1250 | if (IS_ERR(vma)) { |
1251 | memset(&cache->node, 0, sizeof(cache->node)); | |
1252 | mutex_lock(&ggtt->vm.mutex); | |
1253 | err = drm_mm_insert_node_in_range | |
1254 | (&ggtt->vm.mm, &cache->node, | |
1255 | PAGE_SIZE, 0, I915_COLOR_UNEVICTABLE, | |
1256 | 0, ggtt->mappable_end, | |
1257 | DRM_MM_INSERT_LOW); | |
1258 | mutex_unlock(&ggtt->vm.mutex); | |
1259 | if (err) /* no inactive aperture space, use cpu reloc */ | |
1260 | return NULL; | |
1261 | } else { | |
1262 | cache->node.start = vma->node.start; | |
1263 | cache->node.mm = (void *)vma; | |
1264 | } | |
1265 | } | |
1266 | ||
1267 | offset = cache->node.start; | |
1268 | if (drm_mm_node_allocated(&cache->node)) { | |
1269 | ggtt->vm.insert_page(&ggtt->vm, | |
1270 | i915_gem_object_get_dma_address(obj, page), | |
1271 | offset, I915_CACHE_NONE, 0); | |
1272 | } else { | |
1273 | offset += page << PAGE_SHIFT; | |
1274 | } | |
1275 | ||
1276 | vaddr = (void __force *)io_mapping_map_atomic_wc(&ggtt->iomap, | |
1277 | offset); | |
1278 | cache->page = page; | |
1279 | cache->vaddr = (unsigned long)vaddr; | |
1280 | ||
1281 | return vaddr; | |
1282 | } | |
1283 | ||
1284 | static void *reloc_vaddr(struct drm_i915_gem_object *obj, | |
47b08693 | 1285 | struct i915_execbuffer *eb, |
ad5d95e4 DA |
1286 | unsigned long page) |
1287 | { | |
47b08693 | 1288 | struct reloc_cache *cache = &eb->reloc_cache; |
ad5d95e4 DA |
1289 | void *vaddr; |
1290 | ||
1291 | if (cache->page == page) { | |
1292 | vaddr = unmask_page(cache->vaddr); | |
1293 | } else { | |
1294 | vaddr = NULL; | |
1295 | if ((cache->vaddr & KMAP) == 0) | |
47b08693 | 1296 | vaddr = reloc_iomap(obj, eb, page); |
ad5d95e4 DA |
1297 | if (!vaddr) |
1298 | vaddr = reloc_kmap(obj, cache, page); | |
1299 | } | |
1300 | ||
1301 | return vaddr; | |
1302 | } | |
1303 | ||
1304 | static void clflush_write32(u32 *addr, u32 value, unsigned int flushes) | |
1305 | { | |
1306 | if (unlikely(flushes & (CLFLUSH_BEFORE | CLFLUSH_AFTER))) { | |
1307 | if (flushes & CLFLUSH_BEFORE) { | |
1308 | clflushopt(addr); | |
1309 | mb(); | |
1310 | } | |
1311 | ||
1312 | *addr = value; | |
1313 | ||
1314 | /* | |
1315 | * Writes to the same cacheline are serialised by the CPU | |
1316 | * (including clflush). On the write path, we only require | |
1317 | * that it hits memory in an orderly fashion and place | |
1318 | * mb barriers at the start and end of the relocation phase | |
1319 | * to ensure ordering of clflush wrt to the system. | |
1320 | */ | |
1321 | if (flushes & CLFLUSH_AFTER) | |
1322 | clflushopt(addr); | |
1323 | } else | |
1324 | *addr = value; | |
1325 | } | |
1326 | ||
e3d29130 | 1327 | static u64 |
ad5d95e4 | 1328 | relocate_entry(struct i915_vma *vma, |
e3d29130 | 1329 | const struct drm_i915_gem_relocation_entry *reloc, |
ad5d95e4 | 1330 | struct i915_execbuffer *eb, |
e3d29130 CW |
1331 | const struct i915_vma *target) |
1332 | { | |
1333 | u64 target_addr = relocation_target(reloc, target); | |
ad5d95e4 | 1334 | u64 offset = reloc->offset; |
ce13c78f DV |
1335 | bool wide = eb->reloc_cache.use_64bit_reloc; |
1336 | void *vaddr; | |
ad5d95e4 DA |
1337 | |
1338 | repeat: | |
ce13c78f DV |
1339 | vaddr = reloc_vaddr(vma->obj, eb, |
1340 | offset >> PAGE_SHIFT); | |
1341 | if (IS_ERR(vaddr)) | |
1342 | return PTR_ERR(vaddr); | |
1343 | ||
1344 | GEM_BUG_ON(!IS_ALIGNED(offset, sizeof(u32))); | |
1345 | clflush_write32(vaddr + offset_in_page(offset), | |
1346 | lower_32_bits(target_addr), | |
1347 | eb->reloc_cache.vaddr); | |
1348 | ||
1349 | if (wide) { | |
1350 | offset += sizeof(u32); | |
1351 | target_addr >>= 32; | |
1352 | wide = false; | |
1353 | goto repeat; | |
ad5d95e4 | 1354 | } |
edf4427b | 1355 | |
2889caa9 | 1356 | return target->node.start | UPDATE; |
edf4427b | 1357 | } |
edf4427b | 1358 | |
2889caa9 CW |
1359 | static u64 |
1360 | eb_relocate_entry(struct i915_execbuffer *eb, | |
7d6236bb | 1361 | struct eb_vma *ev, |
2889caa9 | 1362 | const struct drm_i915_gem_relocation_entry *reloc) |
54cf91dc | 1363 | { |
baa89ba3 | 1364 | struct drm_i915_private *i915 = eb->i915; |
7d6236bb | 1365 | struct eb_vma *target; |
2889caa9 | 1366 | int err; |
54cf91dc | 1367 | |
67731b87 | 1368 | /* we've already hold a reference to all valid objects */ |
507d977f CW |
1369 | target = eb_get_vma(eb, reloc->target_handle); |
1370 | if (unlikely(!target)) | |
54cf91dc | 1371 | return -ENOENT; |
e844b990 | 1372 | |
54cf91dc | 1373 | /* Validate that the target is in a valid r/w GPU domain */ |
b8f7ab17 | 1374 | if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) { |
baa89ba3 | 1375 | drm_dbg(&i915->drm, "reloc with multiple write domains: " |
507d977f | 1376 | "target %d offset %d " |
54cf91dc | 1377 | "read %08x write %08x", |
507d977f | 1378 | reloc->target_handle, |
54cf91dc CW |
1379 | (int) reloc->offset, |
1380 | reloc->read_domains, | |
1381 | reloc->write_domain); | |
8b78f0e5 | 1382 | return -EINVAL; |
54cf91dc | 1383 | } |
4ca4a250 DV |
1384 | if (unlikely((reloc->write_domain | reloc->read_domains) |
1385 | & ~I915_GEM_GPU_DOMAINS)) { | |
baa89ba3 | 1386 | drm_dbg(&i915->drm, "reloc with read/write non-GPU domains: " |
507d977f | 1387 | "target %d offset %d " |
54cf91dc | 1388 | "read %08x write %08x", |
507d977f | 1389 | reloc->target_handle, |
54cf91dc CW |
1390 | (int) reloc->offset, |
1391 | reloc->read_domains, | |
1392 | reloc->write_domain); | |
8b78f0e5 | 1393 | return -EINVAL; |
54cf91dc | 1394 | } |
54cf91dc | 1395 | |
2889caa9 | 1396 | if (reloc->write_domain) { |
7d6236bb | 1397 | target->flags |= EXEC_OBJECT_WRITE; |
507d977f | 1398 | |
2889caa9 CW |
1399 | /* |
1400 | * Sandybridge PPGTT errata: We need a global gtt mapping | |
1401 | * for MI and pipe_control writes because the gpu doesn't | |
1402 | * properly redirect them through the ppgtt for non_secure | |
1403 | * batchbuffers. | |
1404 | */ | |
1405 | if (reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION && | |
c2ea703d TH |
1406 | GRAPHICS_VER(eb->i915) == 6 && |
1407 | !i915_vma_is_bound(target->vma, I915_VMA_GLOBAL_BIND)) { | |
1408 | struct i915_vma *vma = target->vma; | |
1409 | ||
1410 | reloc_cache_unmap(&eb->reloc_cache); | |
1411 | mutex_lock(&vma->vm->mutex); | |
7d6236bb CW |
1412 | err = i915_vma_bind(target->vma, |
1413 | target->vma->obj->cache_level, | |
e1a4bbb6 | 1414 | PIN_GLOBAL, NULL, NULL); |
c2ea703d TH |
1415 | mutex_unlock(&vma->vm->mutex); |
1416 | reloc_cache_remap(&eb->reloc_cache, ev->vma->obj); | |
ea97c4ca | 1417 | if (err) |
2889caa9 CW |
1418 | return err; |
1419 | } | |
507d977f | 1420 | } |
54cf91dc | 1421 | |
2889caa9 CW |
1422 | /* |
1423 | * If the relocation already has the right value in it, no | |
54cf91dc CW |
1424 | * more work needs to be done. |
1425 | */ | |
ad5d95e4 DA |
1426 | if (!DBG_FORCE_RELOC && |
1427 | gen8_canonical_addr(target->vma->node.start) == reloc->presumed_offset) | |
67731b87 | 1428 | return 0; |
54cf91dc CW |
1429 | |
1430 | /* Check that the relocation address is valid... */ | |
3c94ceee | 1431 | if (unlikely(reloc->offset > |
7d6236bb | 1432 | ev->vma->size - (eb->reloc_cache.use_64bit_reloc ? 8 : 4))) { |
baa89ba3 | 1433 | drm_dbg(&i915->drm, "Relocation beyond object bounds: " |
507d977f CW |
1434 | "target %d offset %d size %d.\n", |
1435 | reloc->target_handle, | |
1436 | (int)reloc->offset, | |
7d6236bb | 1437 | (int)ev->vma->size); |
8b78f0e5 | 1438 | return -EINVAL; |
54cf91dc | 1439 | } |
b8f7ab17 | 1440 | if (unlikely(reloc->offset & 3)) { |
baa89ba3 | 1441 | drm_dbg(&i915->drm, "Relocation not 4-byte aligned: " |
507d977f CW |
1442 | "target %d offset %d.\n", |
1443 | reloc->target_handle, | |
1444 | (int)reloc->offset); | |
8b78f0e5 | 1445 | return -EINVAL; |
54cf91dc CW |
1446 | } |
1447 | ||
071750e5 CW |
1448 | /* |
1449 | * If we write into the object, we need to force the synchronisation | |
1450 | * barrier, either with an asynchronous clflush or if we executed the | |
1451 | * patching using the GPU (though that should be serialised by the | |
1452 | * timeline). To be completely sure, and since we are required to | |
1453 | * do relocations we are already stalling, disable the user's opt | |
0519bcb1 | 1454 | * out of our synchronisation. |
071750e5 | 1455 | */ |
7d6236bb | 1456 | ev->flags &= ~EXEC_OBJECT_ASYNC; |
071750e5 | 1457 | |
54cf91dc | 1458 | /* and update the user's relocation entry */ |
ad5d95e4 | 1459 | return relocate_entry(ev->vma, reloc, eb, target->vma); |
54cf91dc CW |
1460 | } |
1461 | ||
7d6236bb | 1462 | static int eb_relocate_vma(struct i915_execbuffer *eb, struct eb_vma *ev) |
54cf91dc | 1463 | { |
1d83f442 | 1464 | #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry)) |
2889caa9 | 1465 | struct drm_i915_gem_relocation_entry stack[N_RELOC(512)]; |
7d6236bb | 1466 | const struct drm_i915_gem_exec_object2 *entry = ev->exec; |
e94f7856 CW |
1467 | struct drm_i915_gem_relocation_entry __user *urelocs = |
1468 | u64_to_user_ptr(entry->relocs_ptr); | |
1469 | unsigned long remain = entry->relocation_count; | |
54cf91dc | 1470 | |
e94f7856 | 1471 | if (unlikely(remain > N_RELOC(ULONG_MAX))) |
2889caa9 | 1472 | return -EINVAL; |
ebc0808f | 1473 | |
2889caa9 CW |
1474 | /* |
1475 | * We must check that the entire relocation array is safe | |
1476 | * to read. However, if the array is not writable the user loses | |
1477 | * the updated relocation values. | |
1478 | */ | |
e94f7856 | 1479 | if (unlikely(!access_ok(urelocs, remain * sizeof(*urelocs)))) |
2889caa9 CW |
1480 | return -EFAULT; |
1481 | ||
1482 | do { | |
1483 | struct drm_i915_gem_relocation_entry *r = stack; | |
1484 | unsigned int count = | |
e94f7856 | 1485 | min_t(unsigned long, remain, ARRAY_SIZE(stack)); |
2889caa9 | 1486 | unsigned int copied; |
1d83f442 | 1487 | |
2889caa9 CW |
1488 | /* |
1489 | * This is the fast path and we cannot handle a pagefault | |
ebc0808f CW |
1490 | * whilst holding the struct mutex lest the user pass in the |
1491 | * relocations contained within a mmaped bo. For in such a case | |
1492 | * we, the page fault handler would call i915_gem_fault() and | |
1493 | * we would try to acquire the struct mutex again. Obviously | |
1494 | * this is bad and so lockdep complains vehemently. | |
1495 | */ | |
fd1500fc ML |
1496 | pagefault_disable(); |
1497 | copied = __copy_from_user_inatomic(r, urelocs, count * sizeof(r[0])); | |
1498 | pagefault_enable(); | |
ad5d95e4 DA |
1499 | if (unlikely(copied)) { |
1500 | remain = -EFAULT; | |
1501 | goto out; | |
1502 | } | |
54cf91dc | 1503 | |
2889caa9 | 1504 | remain -= count; |
1d83f442 | 1505 | do { |
7d6236bb | 1506 | u64 offset = eb_relocate_entry(eb, ev, r); |
54cf91dc | 1507 | |
2889caa9 CW |
1508 | if (likely(offset == 0)) { |
1509 | } else if ((s64)offset < 0) { | |
ad5d95e4 DA |
1510 | remain = (int)offset; |
1511 | goto out; | |
2889caa9 CW |
1512 | } else { |
1513 | /* | |
1514 | * Note that reporting an error now | |
1515 | * leaves everything in an inconsistent | |
1516 | * state as we have *already* changed | |
1517 | * the relocation value inside the | |
1518 | * object. As we have not changed the | |
1519 | * reloc.presumed_offset or will not | |
1520 | * change the execobject.offset, on the | |
1521 | * call we may not rewrite the value | |
1522 | * inside the object, leaving it | |
1523 | * dangling and causing a GPU hang. Unless | |
1524 | * userspace dynamically rebuilds the | |
1525 | * relocations on each execbuf rather than | |
1526 | * presume a static tree. | |
1527 | * | |
1528 | * We did previously check if the relocations | |
1529 | * were writable (access_ok), an error now | |
1530 | * would be a strange race with mprotect, | |
1531 | * having already demonstrated that we | |
1532 | * can read from this userspace address. | |
1533 | */ | |
1534 | offset = gen8_canonical_addr(offset & ~UPDATE); | |
97a37c91 CW |
1535 | __put_user(offset, |
1536 | &urelocs[r - stack].presumed_offset); | |
1d83f442 | 1537 | } |
2889caa9 CW |
1538 | } while (r++, --count); |
1539 | urelocs += ARRAY_SIZE(stack); | |
1540 | } while (remain); | |
ad5d95e4 | 1541 | out: |
c43ce123 | 1542 | reloc_cache_reset(&eb->reloc_cache, eb); |
ad5d95e4 | 1543 | return remain; |
54cf91dc CW |
1544 | } |
1545 | ||
fd1500fc ML |
1546 | static int |
1547 | eb_relocate_vma_slow(struct i915_execbuffer *eb, struct eb_vma *ev) | |
54cf91dc | 1548 | { |
fd1500fc ML |
1549 | const struct drm_i915_gem_exec_object2 *entry = ev->exec; |
1550 | struct drm_i915_gem_relocation_entry *relocs = | |
1551 | u64_to_ptr(typeof(*relocs), entry->relocs_ptr); | |
1552 | unsigned int i; | |
003d8b91 CW |
1553 | int err; |
1554 | ||
fd1500fc ML |
1555 | for (i = 0; i < entry->relocation_count; i++) { |
1556 | u64 offset = eb_relocate_entry(eb, ev, &relocs[i]); | |
003d8b91 | 1557 | |
fd1500fc ML |
1558 | if ((s64)offset < 0) { |
1559 | err = (int)offset; | |
1560 | goto err; | |
1561 | } | |
ef398881 | 1562 | } |
fd1500fc ML |
1563 | err = 0; |
1564 | err: | |
c43ce123 | 1565 | reloc_cache_reset(&eb->reloc_cache, eb); |
fd1500fc ML |
1566 | return err; |
1567 | } | |
2889caa9 | 1568 | |
fd1500fc ML |
1569 | static int check_relocations(const struct drm_i915_gem_exec_object2 *entry) |
1570 | { | |
1571 | const char __user *addr, *end; | |
1572 | unsigned long size; | |
1573 | char __maybe_unused c; | |
2889caa9 | 1574 | |
fd1500fc ML |
1575 | size = entry->relocation_count; |
1576 | if (size == 0) | |
1577 | return 0; | |
0e97fbb0 | 1578 | |
fd1500fc ML |
1579 | if (size > N_RELOC(ULONG_MAX)) |
1580 | return -EINVAL; | |
2889caa9 | 1581 | |
fd1500fc ML |
1582 | addr = u64_to_user_ptr(entry->relocs_ptr); |
1583 | size *= sizeof(struct drm_i915_gem_relocation_entry); | |
1584 | if (!access_ok(addr, size)) | |
1585 | return -EFAULT; | |
1586 | ||
1587 | end = addr + size; | |
1588 | for (; addr < end; addr += PAGE_SIZE) { | |
1589 | int err = __get_user(c, addr); | |
1590 | if (err) | |
1591 | return err; | |
1592 | } | |
1593 | return __get_user(c, end - 1); | |
2889caa9 CW |
1594 | } |
1595 | ||
fd1500fc | 1596 | static int eb_copy_relocations(const struct i915_execbuffer *eb) |
2889caa9 | 1597 | { |
fd1500fc | 1598 | struct drm_i915_gem_relocation_entry *relocs; |
2889caa9 CW |
1599 | const unsigned int count = eb->buffer_count; |
1600 | unsigned int i; | |
fd1500fc | 1601 | int err; |
54cf91dc | 1602 | |
2889caa9 | 1603 | for (i = 0; i < count; i++) { |
fd1500fc ML |
1604 | const unsigned int nreloc = eb->exec[i].relocation_count; |
1605 | struct drm_i915_gem_relocation_entry __user *urelocs; | |
1606 | unsigned long size; | |
1607 | unsigned long copied; | |
6951e589 | 1608 | |
fd1500fc ML |
1609 | if (nreloc == 0) |
1610 | continue; | |
6951e589 | 1611 | |
fd1500fc ML |
1612 | err = check_relocations(&eb->exec[i]); |
1613 | if (err) | |
1614 | goto err; | |
6951e589 | 1615 | |
fd1500fc ML |
1616 | urelocs = u64_to_user_ptr(eb->exec[i].relocs_ptr); |
1617 | size = nreloc * sizeof(*relocs); | |
6951e589 | 1618 | |
fd1500fc ML |
1619 | relocs = kvmalloc_array(size, 1, GFP_KERNEL); |
1620 | if (!relocs) { | |
1621 | err = -ENOMEM; | |
1622 | goto err; | |
6951e589 | 1623 | } |
fd1500fc ML |
1624 | |
1625 | /* copy_from_user is limited to < 4GiB */ | |
1626 | copied = 0; | |
1627 | do { | |
1628 | unsigned int len = | |
1629 | min_t(u64, BIT_ULL(31), size - copied); | |
1630 | ||
1631 | if (__copy_from_user((char *)relocs + copied, | |
1632 | (char __user *)urelocs + copied, | |
1633 | len)) | |
1634 | goto end; | |
1635 | ||
1636 | copied += len; | |
1637 | } while (copied < size); | |
1638 | ||
1639 | /* | |
1640 | * As we do not update the known relocation offsets after | |
1641 | * relocating (due to the complexities in lock handling), | |
1642 | * we need to mark them as invalid now so that we force the | |
1643 | * relocation processing next time. Just in case the target | |
1644 | * object is evicted and then rebound into its old | |
1645 | * presumed_offset before the next execbuffer - if that | |
1646 | * happened we would make the mistake of assuming that the | |
1647 | * relocations were valid. | |
1648 | */ | |
1649 | if (!user_access_begin(urelocs, size)) | |
1650 | goto end; | |
1651 | ||
1652 | for (copied = 0; copied < nreloc; copied++) | |
1653 | unsafe_put_user(-1, | |
1654 | &urelocs[copied].presumed_offset, | |
1655 | end_user); | |
1656 | user_access_end(); | |
1657 | ||
1658 | eb->exec[i].relocs_ptr = (uintptr_t)relocs; | |
1659 | } | |
1660 | ||
1661 | return 0; | |
1662 | ||
1663 | end_user: | |
1664 | user_access_end(); | |
1665 | end: | |
1666 | kvfree(relocs); | |
1667 | err = -EFAULT; | |
1668 | err: | |
1669 | while (i--) { | |
1670 | relocs = u64_to_ptr(typeof(*relocs), eb->exec[i].relocs_ptr); | |
1671 | if (eb->exec[i].relocation_count) | |
1672 | kvfree(relocs); | |
1673 | } | |
1674 | return err; | |
1675 | } | |
1676 | ||
1677 | static int eb_prefault_relocations(const struct i915_execbuffer *eb) | |
1678 | { | |
1679 | const unsigned int count = eb->buffer_count; | |
1680 | unsigned int i; | |
1681 | ||
1682 | for (i = 0; i < count; i++) { | |
1683 | int err; | |
1684 | ||
1685 | err = check_relocations(&eb->exec[i]); | |
1686 | if (err) | |
1687 | return err; | |
1688 | } | |
1689 | ||
1690 | return 0; | |
1691 | } | |
1692 | ||
ed29c269 ML |
1693 | static int eb_reinit_userptr(struct i915_execbuffer *eb) |
1694 | { | |
1695 | const unsigned int count = eb->buffer_count; | |
1696 | unsigned int i; | |
1697 | int ret; | |
1698 | ||
1699 | if (likely(!(eb->args->flags & __EXEC_USERPTR_USED))) | |
1700 | return 0; | |
1701 | ||
1702 | for (i = 0; i < count; i++) { | |
1703 | struct eb_vma *ev = &eb->vma[i]; | |
1704 | ||
1705 | if (!i915_gem_object_is_userptr(ev->vma->obj)) | |
1706 | continue; | |
1707 | ||
1708 | ret = i915_gem_object_userptr_submit_init(ev->vma->obj); | |
1709 | if (ret) | |
1710 | return ret; | |
1711 | ||
1712 | ev->flags |= __EXEC_OBJECT_USERPTR_INIT; | |
1713 | } | |
1714 | ||
1715 | return 0; | |
1716 | } | |
1717 | ||
544460c3 | 1718 | static noinline int eb_relocate_parse_slow(struct i915_execbuffer *eb) |
fd1500fc ML |
1719 | { |
1720 | bool have_copy = false; | |
1721 | struct eb_vma *ev; | |
1722 | int err = 0; | |
1723 | ||
1724 | repeat: | |
1725 | if (signal_pending(current)) { | |
1726 | err = -ERESTARTSYS; | |
1727 | goto out; | |
6951e589 | 1728 | } |
fd1500fc | 1729 | |
c43ce123 | 1730 | /* We may process another execbuffer during the unlock... */ |
b4b9731b | 1731 | eb_release_vmas(eb, false); |
c43ce123 ML |
1732 | i915_gem_ww_ctx_fini(&eb->ww); |
1733 | ||
fd1500fc ML |
1734 | /* |
1735 | * We take 3 passes through the slowpatch. | |
1736 | * | |
1737 | * 1 - we try to just prefault all the user relocation entries and | |
1738 | * then attempt to reuse the atomic pagefault disabled fast path again. | |
1739 | * | |
1740 | * 2 - we copy the user entries to a local buffer here outside of the | |
1741 | * local and allow ourselves to wait upon any rendering before | |
1742 | * relocations | |
1743 | * | |
1744 | * 3 - we already have a local copy of the relocation entries, but | |
1745 | * were interrupted (EAGAIN) whilst waiting for the objects, try again. | |
1746 | */ | |
1747 | if (!err) { | |
1748 | err = eb_prefault_relocations(eb); | |
1749 | } else if (!have_copy) { | |
1750 | err = eb_copy_relocations(eb); | |
1751 | have_copy = err == 0; | |
1752 | } else { | |
1753 | cond_resched(); | |
1754 | err = 0; | |
1755 | } | |
1756 | ||
2bf541ff | 1757 | if (!err) |
ed29c269 | 1758 | err = eb_reinit_userptr(eb); |
fd1500fc | 1759 | |
c43ce123 | 1760 | i915_gem_ww_ctx_init(&eb->ww, true); |
fd1500fc ML |
1761 | if (err) |
1762 | goto out; | |
1763 | ||
c43ce123 ML |
1764 | /* reacquire the objects */ |
1765 | repeat_validate: | |
544460c3 MB |
1766 | err = eb_pin_engine(eb, false); |
1767 | if (err) | |
2bf541ff | 1768 | goto err; |
2bf541ff | 1769 | |
c43ce123 | 1770 | err = eb_validate_vmas(eb); |
fd1500fc | 1771 | if (err) |
c43ce123 ML |
1772 | goto err; |
1773 | ||
544460c3 | 1774 | GEM_BUG_ON(!eb->batches[0]); |
fd1500fc ML |
1775 | |
1776 | list_for_each_entry(ev, &eb->relocs, reloc_link) { | |
1777 | if (!have_copy) { | |
fd1500fc | 1778 | err = eb_relocate_vma(eb, ev); |
fd1500fc ML |
1779 | if (err) |
1780 | break; | |
1781 | } else { | |
1782 | err = eb_relocate_vma_slow(eb, ev); | |
1783 | if (err) | |
1784 | break; | |
1785 | } | |
1786 | } | |
1787 | ||
c43ce123 ML |
1788 | if (err == -EDEADLK) |
1789 | goto err; | |
1790 | ||
fd1500fc ML |
1791 | if (err && !have_copy) |
1792 | goto repeat; | |
1793 | ||
1794 | if (err) | |
1795 | goto err; | |
1796 | ||
8e4ba491 ML |
1797 | /* as last step, parse the command buffer */ |
1798 | err = eb_parse(eb); | |
1799 | if (err) | |
1800 | goto err; | |
1801 | ||
fd1500fc ML |
1802 | /* |
1803 | * Leave the user relocations as are, this is the painfully slow path, | |
1804 | * and we want to avoid the complication of dropping the lock whilst | |
1805 | * having buffers reserved in the aperture and so causing spurious | |
1806 | * ENOSPC for random operations. | |
1807 | */ | |
1808 | ||
1809 | err: | |
c43ce123 | 1810 | if (err == -EDEADLK) { |
b4b9731b | 1811 | eb_release_vmas(eb, false); |
c43ce123 ML |
1812 | err = i915_gem_ww_ctx_backoff(&eb->ww); |
1813 | if (!err) | |
1814 | goto repeat_validate; | |
1815 | } | |
1816 | ||
fd1500fc ML |
1817 | if (err == -EAGAIN) |
1818 | goto repeat; | |
1819 | ||
1820 | out: | |
1821 | if (have_copy) { | |
1822 | const unsigned int count = eb->buffer_count; | |
1823 | unsigned int i; | |
1824 | ||
1825 | for (i = 0; i < count; i++) { | |
1826 | const struct drm_i915_gem_exec_object2 *entry = | |
1827 | &eb->exec[i]; | |
1828 | struct drm_i915_gem_relocation_entry *relocs; | |
1829 | ||
1830 | if (!entry->relocation_count) | |
1831 | continue; | |
1832 | ||
1833 | relocs = u64_to_ptr(typeof(*relocs), entry->relocs_ptr); | |
1834 | kvfree(relocs); | |
1835 | } | |
1836 | } | |
1837 | ||
1838 | return err; | |
1839 | } | |
1840 | ||
8e4ba491 | 1841 | static int eb_relocate_parse(struct i915_execbuffer *eb) |
54cf91dc | 1842 | { |
003d8b91 | 1843 | int err; |
2bf541ff | 1844 | bool throttle = true; |
003d8b91 | 1845 | |
c43ce123 | 1846 | retry: |
544460c3 MB |
1847 | err = eb_pin_engine(eb, throttle); |
1848 | if (err) { | |
2bf541ff ML |
1849 | if (err != -EDEADLK) |
1850 | return err; | |
1851 | ||
1852 | goto err; | |
1853 | } | |
1854 | ||
2bf541ff ML |
1855 | /* only throttle once, even if we didn't need to throttle */ |
1856 | throttle = false; | |
1857 | ||
c43ce123 ML |
1858 | err = eb_validate_vmas(eb); |
1859 | if (err == -EAGAIN) | |
1860 | goto slow; | |
1861 | else if (err) | |
1862 | goto err; | |
2889caa9 CW |
1863 | |
1864 | /* The objects are in their final locations, apply the relocations. */ | |
1865 | if (eb->args->flags & __EXEC_HAS_RELOC) { | |
7d6236bb | 1866 | struct eb_vma *ev; |
2889caa9 | 1867 | |
7d6236bb | 1868 | list_for_each_entry(ev, &eb->relocs, reloc_link) { |
7dc8f114 CW |
1869 | err = eb_relocate_vma(eb, ev); |
1870 | if (err) | |
fd1500fc | 1871 | break; |
2889caa9 | 1872 | } |
fd1500fc | 1873 | |
c43ce123 ML |
1874 | if (err == -EDEADLK) |
1875 | goto err; | |
1876 | else if (err) | |
1877 | goto slow; | |
1878 | } | |
1879 | ||
1880 | if (!err) | |
1881 | err = eb_parse(eb); | |
1882 | ||
1883 | err: | |
1884 | if (err == -EDEADLK) { | |
b4b9731b | 1885 | eb_release_vmas(eb, false); |
c43ce123 ML |
1886 | err = i915_gem_ww_ctx_backoff(&eb->ww); |
1887 | if (!err) | |
1888 | goto retry; | |
2889caa9 CW |
1889 | } |
1890 | ||
c43ce123 ML |
1891 | return err; |
1892 | ||
1893 | slow: | |
544460c3 | 1894 | err = eb_relocate_parse_slow(eb); |
c43ce123 ML |
1895 | if (err) |
1896 | /* | |
1897 | * If the user expects the execobject.offset and | |
1898 | * reloc.presumed_offset to be an exact match, | |
1899 | * as for using NO_RELOC, then we cannot update | |
1900 | * the execobject.offset until we have completed | |
1901 | * relocation. | |
1902 | */ | |
1903 | eb->args->flags &= ~__EXEC_HAS_RELOC; | |
1904 | ||
1905 | return err; | |
2889caa9 CW |
1906 | } |
1907 | ||
544460c3 MB |
1908 | /* |
1909 | * Using two helper loops for the order of which requests / batches are created | |
1910 | * and added the to backend. Requests are created in order from the parent to | |
1911 | * the last child. Requests are added in the reverse order, from the last child | |
1912 | * to parent. This is done for locking reasons as the timeline lock is acquired | |
1913 | * during request creation and released when the request is added to the | |
1914 | * backend. To make lockdep happy (see intel_context_timeline_lock) this must be | |
1915 | * the ordering. | |
1916 | */ | |
1917 | #define for_each_batch_create_order(_eb, _i) \ | |
1918 | for ((_i) = 0; (_i) < (_eb)->num_batches; ++(_i)) | |
1919 | #define for_each_batch_add_order(_eb, _i) \ | |
1920 | BUILD_BUG_ON(!typecheck(int, _i)); \ | |
1921 | for ((_i) = (_eb)->num_batches - 1; (_i) >= 0; --(_i)) | |
1922 | ||
1923 | static struct i915_request * | |
1924 | eb_find_first_request_added(struct i915_execbuffer *eb) | |
1925 | { | |
1926 | int i; | |
1927 | ||
1928 | for_each_batch_add_order(eb, i) | |
1929 | if (eb->requests[i]) | |
1930 | return eb->requests[i]; | |
1931 | ||
1932 | GEM_BUG_ON("Request not found"); | |
1933 | ||
1934 | return NULL; | |
1935 | } | |
1936 | ||
ff20afc4 TH |
1937 | #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR) |
1938 | ||
1939 | /* Stage with GFP_KERNEL allocations before we enter the signaling critical path */ | |
1940 | static void eb_capture_stage(struct i915_execbuffer *eb) | |
2889caa9 CW |
1941 | { |
1942 | const unsigned int count = eb->buffer_count; | |
ff20afc4 TH |
1943 | unsigned int i = count, j; |
1944 | struct i915_vma_snapshot *vsnap; | |
6951e589 CW |
1945 | |
1946 | while (i--) { | |
7d6236bb CW |
1947 | struct eb_vma *ev = &eb->vma[i]; |
1948 | struct i915_vma *vma = ev->vma; | |
1949 | unsigned int flags = ev->flags; | |
03ade511 | 1950 | |
ff20afc4 TH |
1951 | if (!(flags & EXEC_OBJECT_CAPTURE)) |
1952 | continue; | |
6951e589 | 1953 | |
ff20afc4 TH |
1954 | vsnap = i915_vma_snapshot_alloc(GFP_KERNEL); |
1955 | if (!vsnap) | |
1956 | continue; | |
1957 | ||
1958 | i915_vma_snapshot_init(vsnap, vma, "user"); | |
1959 | for_each_batch_create_order(eb, j) { | |
e61e0f51 | 1960 | struct i915_capture_list *capture; |
b0fd47ad | 1961 | |
ff20afc4 TH |
1962 | capture = kmalloc(sizeof(*capture), GFP_KERNEL); |
1963 | if (!capture) | |
1964 | continue; | |
544460c3 | 1965 | |
ff20afc4 TH |
1966 | capture->next = eb->capture_lists[j]; |
1967 | capture->vma_snapshot = i915_vma_snapshot_get(vsnap); | |
1968 | eb->capture_lists[j] = capture; | |
1969 | } | |
1970 | i915_vma_snapshot_put(vsnap); | |
1971 | } | |
1972 | } | |
1973 | ||
1974 | /* Commit once we're in the critical path */ | |
1975 | static void eb_capture_commit(struct i915_execbuffer *eb) | |
1976 | { | |
1977 | unsigned int j; | |
1978 | ||
1979 | for_each_batch_create_order(eb, j) { | |
1980 | struct i915_request *rq = eb->requests[j]; | |
1981 | ||
1982 | if (!rq) | |
1983 | break; | |
1984 | ||
1985 | rq->capture_list = eb->capture_lists[j]; | |
1986 | eb->capture_lists[j] = NULL; | |
1987 | } | |
1988 | } | |
1989 | ||
1990 | /* | |
1991 | * Release anything that didn't get committed due to errors. | |
1992 | * The capture_list will otherwise be freed at request retire. | |
1993 | */ | |
1994 | static void eb_capture_release(struct i915_execbuffer *eb) | |
1995 | { | |
1996 | unsigned int j; | |
1997 | ||
1998 | for_each_batch_create_order(eb, j) { | |
1999 | if (eb->capture_lists[j]) { | |
2000 | i915_request_free_capture_list(eb->capture_lists[j]); | |
2001 | eb->capture_lists[j] = NULL; | |
b0fd47ad | 2002 | } |
ff20afc4 TH |
2003 | } |
2004 | } | |
2005 | ||
2006 | static void eb_capture_list_clear(struct i915_execbuffer *eb) | |
2007 | { | |
2008 | memset(eb->capture_lists, 0, sizeof(eb->capture_lists)); | |
2009 | } | |
2010 | ||
2011 | #else | |
2012 | ||
2013 | static void eb_capture_stage(struct i915_execbuffer *eb) | |
2014 | { | |
2015 | } | |
2016 | ||
2017 | static void eb_capture_commit(struct i915_execbuffer *eb) | |
2018 | { | |
2019 | } | |
2020 | ||
2021 | static void eb_capture_release(struct i915_execbuffer *eb) | |
2022 | { | |
2023 | } | |
2024 | ||
2025 | static void eb_capture_list_clear(struct i915_execbuffer *eb) | |
2026 | { | |
2027 | } | |
2028 | ||
2029 | #endif | |
2030 | ||
2031 | static int eb_move_to_gpu(struct i915_execbuffer *eb) | |
2032 | { | |
2033 | const unsigned int count = eb->buffer_count; | |
2034 | unsigned int i = count; | |
2035 | int err = 0, j; | |
2036 | ||
2037 | while (i--) { | |
2038 | struct eb_vma *ev = &eb->vma[i]; | |
2039 | struct i915_vma *vma = ev->vma; | |
2040 | unsigned int flags = ev->flags; | |
2041 | struct drm_i915_gem_object *obj = vma->obj; | |
2042 | ||
2043 | assert_vma_held(vma); | |
b0fd47ad | 2044 | |
b8f55be6 CW |
2045 | /* |
2046 | * If the GPU is not _reading_ through the CPU cache, we need | |
2047 | * to make sure that any writes (both previous GPU writes from | |
2048 | * before a change in snooping levels and normal CPU writes) | |
2049 | * caught in that cache are flushed to main memory. | |
2050 | * | |
2051 | * We want to say | |
2052 | * obj->cache_dirty && | |
2053 | * !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ) | |
2054 | * but gcc's optimiser doesn't handle that as well and emits | |
2055 | * two jumps instead of one. Maybe one day... | |
df94fd05 MA |
2056 | * |
2057 | * FIXME: There is also sync flushing in set_pages(), which | |
2058 | * serves a different purpose(some of the time at least). | |
2059 | * | |
2060 | * We should consider: | |
2061 | * | |
2062 | * 1. Rip out the async flush code. | |
2063 | * | |
2064 | * 2. Or make the sync flushing use the async clflush path | |
2065 | * using mandatory fences underneath. Currently the below | |
2066 | * async flush happens after we bind the object. | |
b8f55be6 CW |
2067 | */ |
2068 | if (unlikely(obj->cache_dirty & ~obj->cache_coherent)) { | |
0f46daa1 | 2069 | if (i915_gem_clflush_object(obj, 0)) |
c7c6e46f | 2070 | flags &= ~EXEC_OBJECT_ASYNC; |
0f46daa1 CW |
2071 | } |
2072 | ||
544460c3 | 2073 | /* We only need to await on the first request */ |
6951e589 CW |
2074 | if (err == 0 && !(flags & EXEC_OBJECT_ASYNC)) { |
2075 | err = i915_request_await_object | |
544460c3 MB |
2076 | (eb_find_first_request_added(eb), obj, |
2077 | flags & EXEC_OBJECT_WRITE); | |
6951e589 | 2078 | } |
2889caa9 | 2079 | |
544460c3 MB |
2080 | for_each_batch_add_order(eb, j) { |
2081 | if (err) | |
2082 | break; | |
2083 | if (!eb->requests[j]) | |
2084 | continue; | |
2085 | ||
2086 | err = _i915_vma_move_to_active(vma, eb->requests[j], | |
2087 | j ? NULL : | |
2088 | eb->composite_fence ? | |
2089 | eb->composite_fence : | |
2090 | &eb->requests[j]->fence, | |
2091 | flags | __EXEC_OBJECT_NO_RESERVE); | |
2092 | } | |
c59a333f | 2093 | } |
0f1dd022 | 2094 | |
ed29c269 ML |
2095 | #ifdef CONFIG_MMU_NOTIFIER |
2096 | if (!err && (eb->args->flags & __EXEC_USERPTR_USED)) { | |
b4b9731b | 2097 | read_lock(&eb->i915->mm.notifier_lock); |
ed29c269 ML |
2098 | |
2099 | /* | |
2100 | * count is always at least 1, otherwise __EXEC_USERPTR_USED | |
2101 | * could not have been set | |
2102 | */ | |
2103 | for (i = 0; i < count; i++) { | |
2104 | struct eb_vma *ev = &eb->vma[i]; | |
2105 | struct drm_i915_gem_object *obj = ev->vma->obj; | |
2106 | ||
2107 | if (!i915_gem_object_is_userptr(obj)) | |
2108 | continue; | |
2109 | ||
2110 | err = i915_gem_object_userptr_submit_done(obj); | |
2111 | if (err) | |
2112 | break; | |
2113 | } | |
2114 | ||
b4b9731b | 2115 | read_unlock(&eb->i915->mm.notifier_lock); |
ed29c269 ML |
2116 | } |
2117 | #endif | |
2118 | ||
6951e589 CW |
2119 | if (unlikely(err)) |
2120 | goto err_skip; | |
2121 | ||
dcd79934 | 2122 | /* Unconditionally flush any chipset caches (for streaming writes). */ |
544460c3 | 2123 | intel_gt_chipset_flush(eb->gt); |
ff20afc4 TH |
2124 | eb_capture_commit(eb); |
2125 | ||
2113184c | 2126 | return 0; |
6951e589 CW |
2127 | |
2128 | err_skip: | |
544460c3 MB |
2129 | for_each_batch_create_order(eb, j) { |
2130 | if (!eb->requests[j]) | |
2131 | break; | |
2132 | ||
2133 | i915_request_set_error_once(eb->requests[j], err); | |
2134 | } | |
6951e589 | 2135 | return err; |
54cf91dc CW |
2136 | } |
2137 | ||
00aff3f6 | 2138 | static int i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec) |
54cf91dc | 2139 | { |
650bc635 | 2140 | if (exec->flags & __I915_EXEC_ILLEGAL_FLAGS) |
00aff3f6 | 2141 | return -EINVAL; |
ed5982e6 | 2142 | |
2f5945bc | 2143 | /* Kernel clipping was a DRI1 misfeature */ |
cda9edd0 LL |
2144 | if (!(exec->flags & (I915_EXEC_FENCE_ARRAY | |
2145 | I915_EXEC_USE_EXTENSIONS))) { | |
cf6e7bac | 2146 | if (exec->num_cliprects || exec->cliprects_ptr) |
00aff3f6 | 2147 | return -EINVAL; |
cf6e7bac | 2148 | } |
2f5945bc CW |
2149 | |
2150 | if (exec->DR4 == 0xffffffff) { | |
2151 | DRM_DEBUG("UXA submitting garbage DR4, fixing up\n"); | |
2152 | exec->DR4 = 0; | |
2153 | } | |
2154 | if (exec->DR1 || exec->DR4) | |
00aff3f6 | 2155 | return -EINVAL; |
2f5945bc CW |
2156 | |
2157 | if ((exec->batch_start_offset | exec->batch_len) & 0x7) | |
00aff3f6 | 2158 | return -EINVAL; |
2f5945bc | 2159 | |
00aff3f6 | 2160 | return 0; |
54cf91dc CW |
2161 | } |
2162 | ||
e61e0f51 | 2163 | static int i915_reset_gen7_sol_offsets(struct i915_request *rq) |
ae662d31 | 2164 | { |
73dec95e TU |
2165 | u32 *cs; |
2166 | int i; | |
ae662d31 | 2167 | |
40e1956e | 2168 | if (GRAPHICS_VER(rq->engine->i915) != 7 || rq->engine->id != RCS0) { |
5a833995 | 2169 | drm_dbg(&rq->engine->i915->drm, "sol reset is gen7/rcs only\n"); |
9d662da8 DV |
2170 | return -EINVAL; |
2171 | } | |
ae662d31 | 2172 | |
e61e0f51 | 2173 | cs = intel_ring_begin(rq, 4 * 2 + 2); |
73dec95e TU |
2174 | if (IS_ERR(cs)) |
2175 | return PTR_ERR(cs); | |
ae662d31 | 2176 | |
2889caa9 | 2177 | *cs++ = MI_LOAD_REGISTER_IMM(4); |
ae662d31 | 2178 | for (i = 0; i < 4; i++) { |
73dec95e TU |
2179 | *cs++ = i915_mmio_reg_offset(GEN7_SO_WRITE_OFFSET(i)); |
2180 | *cs++ = 0; | |
ae662d31 | 2181 | } |
2889caa9 | 2182 | *cs++ = MI_NOOP; |
e61e0f51 | 2183 | intel_ring_advance(rq, cs); |
ae662d31 EA |
2184 | |
2185 | return 0; | |
2186 | } | |
2187 | ||
4f7af194 | 2188 | static struct i915_vma * |
47b08693 ML |
2189 | shadow_batch_pin(struct i915_execbuffer *eb, |
2190 | struct drm_i915_gem_object *obj, | |
32d94048 CW |
2191 | struct i915_address_space *vm, |
2192 | unsigned int flags) | |
4f7af194 | 2193 | { |
b291ce0a | 2194 | struct i915_vma *vma; |
b291ce0a | 2195 | int err; |
4f7af194 | 2196 | |
b291ce0a CW |
2197 | vma = i915_vma_instance(obj, vm, NULL); |
2198 | if (IS_ERR(vma)) | |
2199 | return vma; | |
2200 | ||
47b08693 | 2201 | err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, flags); |
b291ce0a CW |
2202 | if (err) |
2203 | return ERR_PTR(err); | |
2204 | ||
2205 | return vma; | |
4f7af194 JB |
2206 | } |
2207 | ||
47b08693 ML |
2208 | static struct i915_vma *eb_dispatch_secure(struct i915_execbuffer *eb, struct i915_vma *vma) |
2209 | { | |
2210 | /* | |
2211 | * snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure | |
2212 | * batch" bit. Hence we need to pin secure batches into the global gtt. | |
2213 | * hsw should have this fixed, but bdw mucks it up again. */ | |
2214 | if (eb->batch_flags & I915_DISPATCH_SECURE) | |
2215 | return i915_gem_object_ggtt_pin_ww(vma->obj, &eb->ww, NULL, 0, 0, 0); | |
2216 | ||
2217 | return NULL; | |
2218 | } | |
2219 | ||
51696691 | 2220 | static int eb_parse(struct i915_execbuffer *eb) |
71745376 | 2221 | { |
baa89ba3 | 2222 | struct drm_i915_private *i915 = eb->i915; |
c43ce123 | 2223 | struct intel_gt_buffer_pool_node *pool = eb->batch_pool; |
47b08693 | 2224 | struct i915_vma *shadow, *trampoline, *batch; |
d5e87821 | 2225 | unsigned long len; |
2889caa9 | 2226 | int err; |
71745376 | 2227 | |
47b08693 | 2228 | if (!eb_use_cmdparser(eb)) { |
544460c3 | 2229 | batch = eb_dispatch_secure(eb, eb->batches[0]->vma); |
47b08693 ML |
2230 | if (IS_ERR(batch)) |
2231 | return PTR_ERR(batch); | |
2232 | ||
2233 | goto secure_batch; | |
2234 | } | |
51696691 | 2235 | |
544460c3 MB |
2236 | if (intel_context_is_parallel(eb->context)) |
2237 | return -EINVAL; | |
2238 | ||
2239 | len = eb->batch_len[0]; | |
32d94048 CW |
2240 | if (!CMDPARSER_USES_GGTT(eb->i915)) { |
2241 | /* | |
2242 | * ppGTT backed shadow buffers must be mapped RO, to prevent | |
2243 | * post-scan tampering | |
2244 | */ | |
2245 | if (!eb->context->vm->has_read_only) { | |
baa89ba3 WK |
2246 | drm_dbg(&i915->drm, |
2247 | "Cannot prevent post-scan tampering without RO capable vm\n"); | |
32d94048 CW |
2248 | return -EINVAL; |
2249 | } | |
2250 | } else { | |
2251 | len += I915_CMD_PARSER_TRAMPOLINE_SIZE; | |
2252 | } | |
544460c3 | 2253 | if (unlikely(len < eb->batch_len[0])) /* last paranoid check of overflow */ |
d5e87821 | 2254 | return -EINVAL; |
32d94048 | 2255 | |
c43ce123 | 2256 | if (!pool) { |
544460c3 | 2257 | pool = intel_gt_get_buffer_pool(eb->gt, len, |
8f47c8c3 | 2258 | I915_MAP_WB); |
c43ce123 ML |
2259 | if (IS_ERR(pool)) |
2260 | return PTR_ERR(pool); | |
2261 | eb->batch_pool = pool; | |
2262 | } | |
71745376 | 2263 | |
c43ce123 ML |
2264 | err = i915_gem_object_lock(pool->obj, &eb->ww); |
2265 | if (err) | |
2266 | goto err; | |
71745376 | 2267 | |
47b08693 | 2268 | shadow = shadow_batch_pin(eb, pool->obj, eb->context->vm, PIN_USER); |
32d94048 CW |
2269 | if (IS_ERR(shadow)) { |
2270 | err = PTR_ERR(shadow); | |
f8c08d8f | 2271 | goto err; |
51696691 | 2272 | } |
c9398775 | 2273 | intel_gt_buffer_pool_mark_used(pool); |
32d94048 | 2274 | i915_gem_object_set_readonly(shadow->obj); |
57a78ca4 | 2275 | shadow->private = pool; |
32d94048 CW |
2276 | |
2277 | trampoline = NULL; | |
2278 | if (CMDPARSER_USES_GGTT(eb->i915)) { | |
2279 | trampoline = shadow; | |
2280 | ||
47b08693 | 2281 | shadow = shadow_batch_pin(eb, pool->obj, |
544460c3 | 2282 | &eb->gt->ggtt->vm, |
32d94048 CW |
2283 | PIN_GLOBAL); |
2284 | if (IS_ERR(shadow)) { | |
2285 | err = PTR_ERR(shadow); | |
2286 | shadow = trampoline; | |
2287 | goto err_shadow; | |
2288 | } | |
57a78ca4 | 2289 | shadow->private = pool; |
32d94048 CW |
2290 | |
2291 | eb->batch_flags |= I915_DISPATCH_SECURE; | |
2292 | } | |
f8c08d8f | 2293 | |
47b08693 ML |
2294 | batch = eb_dispatch_secure(eb, shadow); |
2295 | if (IS_ERR(batch)) { | |
2296 | err = PTR_ERR(batch); | |
2297 | goto err_trampoline; | |
2298 | } | |
2299 | ||
95c3d275 | 2300 | err = dma_resv_reserve_shared(shadow->obj->base.resv, 1); |
93b71330 JE |
2301 | if (err) |
2302 | goto err_trampoline; | |
2303 | ||
544460c3 MB |
2304 | err = intel_engine_cmd_parser(eb->context->engine, |
2305 | eb->batches[0]->vma, | |
93b71330 | 2306 | eb->batch_start_offset, |
544460c3 | 2307 | eb->batch_len[0], |
93b71330 | 2308 | shadow, trampoline); |
32d94048 | 2309 | if (err) |
47b08693 | 2310 | goto err_unpin_batch; |
71745376 | 2311 | |
544460c3 MB |
2312 | eb->batches[0] = &eb->vma[eb->buffer_count++]; |
2313 | eb->batches[0]->vma = i915_vma_get(shadow); | |
2314 | eb->batches[0]->flags = __EXEC_OBJECT_HAS_PIN; | |
71745376 | 2315 | |
32d94048 | 2316 | eb->trampoline = trampoline; |
4f7af194 | 2317 | eb->batch_start_offset = 0; |
4f7af194 | 2318 | |
47b08693 ML |
2319 | secure_batch: |
2320 | if (batch) { | |
544460c3 MB |
2321 | if (intel_context_is_parallel(eb->context)) |
2322 | return -EINVAL; | |
2323 | ||
2324 | eb->batches[0] = &eb->vma[eb->buffer_count++]; | |
2325 | eb->batches[0]->flags = __EXEC_OBJECT_HAS_PIN; | |
2326 | eb->batches[0]->vma = i915_vma_get(batch); | |
47b08693 | 2327 | } |
51696691 | 2328 | return 0; |
b40d7378 | 2329 | |
47b08693 ML |
2330 | err_unpin_batch: |
2331 | if (batch) | |
2332 | i915_vma_unpin(batch); | |
32d94048 CW |
2333 | err_trampoline: |
2334 | if (trampoline) | |
2335 | i915_vma_unpin(trampoline); | |
2336 | err_shadow: | |
2337 | i915_vma_unpin(shadow); | |
b40d7378 | 2338 | err: |
51696691 | 2339 | return err; |
71745376 | 2340 | } |
5c6c6003 | 2341 | |
544460c3 MB |
2342 | static int eb_request_submit(struct i915_execbuffer *eb, |
2343 | struct i915_request *rq, | |
2344 | struct i915_vma *batch, | |
2345 | u64 batch_len) | |
78382593 | 2346 | { |
2889caa9 | 2347 | int err; |
78382593 | 2348 | |
544460c3 MB |
2349 | if (intel_context_nopreempt(rq->context)) |
2350 | __set_bit(I915_FENCE_FLAG_NOPREEMPT, &rq->fence.flags); | |
78382593 | 2351 | |
650bc635 | 2352 | if (eb->args->flags & I915_EXEC_GEN7_SOL_RESET) { |
544460c3 | 2353 | err = i915_reset_gen7_sol_offsets(rq); |
2889caa9 CW |
2354 | if (err) |
2355 | return err; | |
78382593 OM |
2356 | } |
2357 | ||
85474441 CW |
2358 | /* |
2359 | * After we completed waiting for other engines (using HW semaphores) | |
2360 | * then we can signal that this request/batch is ready to run. This | |
2361 | * allows us to determine if the batch is still waiting on the GPU | |
2362 | * or actually running by checking the breadcrumb. | |
2363 | */ | |
544460c3 MB |
2364 | if (rq->context->engine->emit_init_breadcrumb) { |
2365 | err = rq->context->engine->emit_init_breadcrumb(rq); | |
85474441 CW |
2366 | if (err) |
2367 | return err; | |
2368 | } | |
2369 | ||
544460c3 MB |
2370 | err = rq->context->engine->emit_bb_start(rq, |
2371 | batch->node.start + | |
2372 | eb->batch_start_offset, | |
2373 | batch_len, | |
2374 | eb->batch_flags); | |
2889caa9 CW |
2375 | if (err) |
2376 | return err; | |
78382593 | 2377 | |
32d94048 | 2378 | if (eb->trampoline) { |
544460c3 | 2379 | GEM_BUG_ON(intel_context_is_parallel(rq->context)); |
32d94048 | 2380 | GEM_BUG_ON(eb->batch_start_offset); |
544460c3 MB |
2381 | err = rq->context->engine->emit_bb_start(rq, |
2382 | eb->trampoline->node.start + | |
2383 | batch_len, 0, 0); | |
32d94048 CW |
2384 | if (err) |
2385 | return err; | |
2386 | } | |
2387 | ||
2f5945bc | 2388 | return 0; |
78382593 OM |
2389 | } |
2390 | ||
544460c3 MB |
2391 | static int eb_submit(struct i915_execbuffer *eb) |
2392 | { | |
2393 | unsigned int i; | |
2394 | int err; | |
2395 | ||
2396 | err = eb_move_to_gpu(eb); | |
2397 | ||
2398 | for_each_batch_create_order(eb, i) { | |
2399 | if (!eb->requests[i]) | |
2400 | break; | |
2401 | ||
2402 | trace_i915_request_queue(eb->requests[i], eb->batch_flags); | |
2403 | if (!err) | |
2404 | err = eb_request_submit(eb, eb->requests[i], | |
2405 | eb->batches[i]->vma, | |
2406 | eb->batch_len[i]); | |
2407 | } | |
2408 | ||
2409 | return err; | |
2410 | } | |
2411 | ||
1a9c4db4 | 2412 | static int num_vcs_engines(struct drm_i915_private *i915) |
d5b2a3a4 | 2413 | { |
1a9c4db4 | 2414 | return hweight_long(VDBOX_MASK(to_gt(i915))); |
d5b2a3a4 CW |
2415 | } |
2416 | ||
204bcfef | 2417 | /* |
a8ebba75 | 2418 | * Find one BSD ring to dispatch the corresponding BSD command. |
c80ff16e | 2419 | * The engine index is returned. |
a8ebba75 | 2420 | */ |
de1add36 | 2421 | static unsigned int |
c80ff16e CW |
2422 | gen8_dispatch_bsd_engine(struct drm_i915_private *dev_priv, |
2423 | struct drm_file *file) | |
a8ebba75 | 2424 | { |
a8ebba75 ZY |
2425 | struct drm_i915_file_private *file_priv = file->driver_priv; |
2426 | ||
de1add36 | 2427 | /* Check whether the file_priv has already selected one ring. */ |
6f633402 | 2428 | if ((int)file_priv->bsd_engine < 0) |
1a07e86c CW |
2429 | file_priv->bsd_engine = |
2430 | get_random_int() % num_vcs_engines(dev_priv); | |
d23db88c | 2431 | |
c80ff16e | 2432 | return file_priv->bsd_engine; |
d23db88c CW |
2433 | } |
2434 | ||
5e2a0419 | 2435 | static const enum intel_engine_id user_ring_map[] = { |
8a68d464 CW |
2436 | [I915_EXEC_DEFAULT] = RCS0, |
2437 | [I915_EXEC_RENDER] = RCS0, | |
2438 | [I915_EXEC_BLT] = BCS0, | |
2439 | [I915_EXEC_BSD] = VCS0, | |
2440 | [I915_EXEC_VEBOX] = VECS0 | |
de1add36 TU |
2441 | }; |
2442 | ||
2bf541ff | 2443 | static struct i915_request *eb_throttle(struct i915_execbuffer *eb, struct intel_context *ce) |
e5dadff4 CW |
2444 | { |
2445 | struct intel_ring *ring = ce->ring; | |
2446 | struct intel_timeline *tl = ce->timeline; | |
2447 | struct i915_request *rq; | |
2448 | ||
2449 | /* | |
2450 | * Completely unscientific finger-in-the-air estimates for suitable | |
2451 | * maximum user request size (to avoid blocking) and then backoff. | |
2452 | */ | |
2453 | if (intel_ring_update_space(ring) >= PAGE_SIZE) | |
2454 | return NULL; | |
2455 | ||
2456 | /* | |
2457 | * Find a request that after waiting upon, there will be at least half | |
2458 | * the ring available. The hysteresis allows us to compete for the | |
2459 | * shared ring and should mean that we sleep less often prior to | |
2460 | * claiming our resources, but not so long that the ring completely | |
2461 | * drains before we can submit our next request. | |
2462 | */ | |
2463 | list_for_each_entry(rq, &tl->requests, link) { | |
2464 | if (rq->ring != ring) | |
2465 | continue; | |
2466 | ||
2467 | if (__intel_ring_space(rq->postfix, | |
2468 | ring->emit, ring->size) > ring->size / 2) | |
2469 | break; | |
2470 | } | |
2471 | if (&rq->link == &tl->requests) | |
2472 | return NULL; /* weird, we will check again later for real */ | |
2473 | ||
2474 | return i915_request_get(rq); | |
2475 | } | |
2476 | ||
544460c3 MB |
2477 | static int eb_pin_timeline(struct i915_execbuffer *eb, struct intel_context *ce, |
2478 | bool throttle) | |
e5dadff4 CW |
2479 | { |
2480 | struct intel_timeline *tl; | |
2bf541ff | 2481 | struct i915_request *rq = NULL; |
8f2a1057 | 2482 | |
a4e57f90 CW |
2483 | /* |
2484 | * Take a local wakeref for preparing to dispatch the execbuf as | |
2485 | * we expect to access the hardware fairly frequently in the | |
2486 | * process, and require the engine to be kept awake between accesses. | |
2487 | * Upon dispatch, we acquire another prolonged wakeref that we hold | |
2488 | * until the timeline is idle, which in turn releases the wakeref | |
2489 | * taken on the engine, and the parent device. | |
2490 | */ | |
e5dadff4 | 2491 | tl = intel_context_timeline_lock(ce); |
544460c3 MB |
2492 | if (IS_ERR(tl)) |
2493 | return PTR_ERR(tl); | |
a4e57f90 CW |
2494 | |
2495 | intel_context_enter(ce); | |
2bf541ff ML |
2496 | if (throttle) |
2497 | rq = eb_throttle(eb, ce); | |
e5dadff4 CW |
2498 | intel_context_timeline_unlock(tl); |
2499 | ||
544460c3 MB |
2500 | if (rq) { |
2501 | bool nonblock = eb->file->filp->f_flags & O_NONBLOCK; | |
2502 | long timeout = nonblock ? 0 : MAX_SCHEDULE_TIMEOUT; | |
2503 | ||
2504 | if (i915_request_wait(rq, I915_WAIT_INTERRUPTIBLE, | |
2505 | timeout) < 0) { | |
2506 | i915_request_put(rq); | |
2507 | ||
2508 | tl = intel_context_timeline_lock(ce); | |
2509 | intel_context_exit(ce); | |
2510 | intel_context_timeline_unlock(tl); | |
2511 | ||
2512 | if (nonblock) | |
2513 | return -EWOULDBLOCK; | |
2514 | else | |
2515 | return -EINTR; | |
2516 | } | |
2517 | i915_request_put(rq); | |
2518 | } | |
2519 | ||
2520 | return 0; | |
2521 | } | |
2522 | ||
2523 | static int eb_pin_engine(struct i915_execbuffer *eb, bool throttle) | |
2524 | { | |
2525 | struct intel_context *ce = eb->context, *child; | |
2526 | int err; | |
2527 | int i = 0, j = 0; | |
2528 | ||
2529 | GEM_BUG_ON(eb->args->flags & __EXEC_ENGINE_PINNED); | |
2530 | ||
2531 | if (unlikely(intel_context_is_banned(ce))) | |
2532 | return -EIO; | |
2533 | ||
2534 | /* | |
2535 | * Pinning the contexts may generate requests in order to acquire | |
2536 | * GGTT space, so do this first before we reserve a seqno for | |
2537 | * ourselves. | |
2538 | */ | |
2539 | err = intel_context_pin_ww(ce, &eb->ww); | |
2540 | if (err) | |
2541 | return err; | |
2542 | for_each_child(ce, child) { | |
2543 | err = intel_context_pin_ww(child, &eb->ww); | |
2544 | GEM_BUG_ON(err); /* perma-pinned should incr a counter */ | |
2545 | } | |
2546 | ||
2547 | for_each_child(ce, child) { | |
2548 | err = eb_pin_timeline(eb, child, throttle); | |
2549 | if (err) | |
2550 | goto unwind; | |
2551 | ++i; | |
2552 | } | |
2553 | err = eb_pin_timeline(eb, ce, throttle); | |
2554 | if (err) | |
2555 | goto unwind; | |
2556 | ||
2bf541ff | 2557 | eb->args->flags |= __EXEC_ENGINE_PINNED; |
544460c3 MB |
2558 | return 0; |
2559 | ||
2560 | unwind: | |
2561 | for_each_child(ce, child) { | |
2562 | if (j++ < i) { | |
2563 | mutex_lock(&child->timeline->mutex); | |
2564 | intel_context_exit(child); | |
2565 | mutex_unlock(&child->timeline->mutex); | |
2566 | } | |
2567 | } | |
2568 | for_each_child(ce, child) | |
2569 | intel_context_unpin(child); | |
2570 | intel_context_unpin(ce); | |
2571 | return err; | |
8f2a1057 CW |
2572 | } |
2573 | ||
e5dadff4 | 2574 | static void eb_unpin_engine(struct i915_execbuffer *eb) |
8f2a1057 | 2575 | { |
544460c3 | 2576 | struct intel_context *ce = eb->context, *child; |
a4e57f90 | 2577 | |
2bf541ff ML |
2578 | if (!(eb->args->flags & __EXEC_ENGINE_PINNED)) |
2579 | return; | |
2580 | ||
2581 | eb->args->flags &= ~__EXEC_ENGINE_PINNED; | |
2582 | ||
544460c3 MB |
2583 | for_each_child(ce, child) { |
2584 | mutex_lock(&child->timeline->mutex); | |
2585 | intel_context_exit(child); | |
2586 | mutex_unlock(&child->timeline->mutex); | |
2587 | ||
2588 | intel_context_unpin(child); | |
2589 | } | |
2590 | ||
2591 | mutex_lock(&ce->timeline->mutex); | |
a4e57f90 | 2592 | intel_context_exit(ce); |
544460c3 | 2593 | mutex_unlock(&ce->timeline->mutex); |
a4e57f90 | 2594 | |
2850748e | 2595 | intel_context_unpin(ce); |
8f2a1057 | 2596 | } |
de1add36 | 2597 | |
5e2a0419 | 2598 | static unsigned int |
b49a7d51 | 2599 | eb_select_legacy_ring(struct i915_execbuffer *eb) |
de1add36 | 2600 | { |
8f2a1057 | 2601 | struct drm_i915_private *i915 = eb->i915; |
b49a7d51 | 2602 | struct drm_i915_gem_execbuffer2 *args = eb->args; |
de1add36 | 2603 | unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK; |
de1add36 | 2604 | |
5e2a0419 CW |
2605 | if (user_ring_id != I915_EXEC_BSD && |
2606 | (args->flags & I915_EXEC_BSD_MASK)) { | |
baa89ba3 WK |
2607 | drm_dbg(&i915->drm, |
2608 | "execbuf with non bsd ring but with invalid " | |
2609 | "bsd dispatch flags: %d\n", (int)(args->flags)); | |
5e2a0419 | 2610 | return -1; |
de1add36 TU |
2611 | } |
2612 | ||
d5b2a3a4 | 2613 | if (user_ring_id == I915_EXEC_BSD && num_vcs_engines(i915) > 1) { |
de1add36 TU |
2614 | unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK; |
2615 | ||
2616 | if (bsd_idx == I915_EXEC_BSD_DEFAULT) { | |
b49a7d51 | 2617 | bsd_idx = gen8_dispatch_bsd_engine(i915, eb->file); |
de1add36 TU |
2618 | } else if (bsd_idx >= I915_EXEC_BSD_RING1 && |
2619 | bsd_idx <= I915_EXEC_BSD_RING2) { | |
d9da6aa0 | 2620 | bsd_idx >>= I915_EXEC_BSD_SHIFT; |
de1add36 TU |
2621 | bsd_idx--; |
2622 | } else { | |
baa89ba3 WK |
2623 | drm_dbg(&i915->drm, |
2624 | "execbuf with unknown bsd ring: %u\n", | |
2625 | bsd_idx); | |
5e2a0419 | 2626 | return -1; |
de1add36 TU |
2627 | } |
2628 | ||
5e2a0419 | 2629 | return _VCS(bsd_idx); |
de1add36 TU |
2630 | } |
2631 | ||
5e2a0419 | 2632 | if (user_ring_id >= ARRAY_SIZE(user_ring_map)) { |
baa89ba3 WK |
2633 | drm_dbg(&i915->drm, "execbuf with unknown ring: %u\n", |
2634 | user_ring_id); | |
5e2a0419 | 2635 | return -1; |
de1add36 TU |
2636 | } |
2637 | ||
5e2a0419 CW |
2638 | return user_ring_map[user_ring_id]; |
2639 | } | |
2640 | ||
2641 | static int | |
2bf541ff | 2642 | eb_select_engine(struct i915_execbuffer *eb) |
5e2a0419 | 2643 | { |
544460c3 | 2644 | struct intel_context *ce, *child; |
5e2a0419 CW |
2645 | unsigned int idx; |
2646 | int err; | |
2647 | ||
976b55f0 | 2648 | if (i915_gem_context_user_engines(eb->gem_context)) |
b49a7d51 | 2649 | idx = eb->args->flags & I915_EXEC_RING_MASK; |
976b55f0 | 2650 | else |
b49a7d51 | 2651 | idx = eb_select_legacy_ring(eb); |
5e2a0419 CW |
2652 | |
2653 | ce = i915_gem_context_get_engine(eb->gem_context, idx); | |
2654 | if (IS_ERR(ce)) | |
2655 | return PTR_ERR(ce); | |
2656 | ||
544460c3 MB |
2657 | if (intel_context_is_parallel(ce)) { |
2658 | if (eb->buffer_count < ce->parallel.number_children + 1) { | |
2659 | intel_context_put(ce); | |
2660 | return -EINVAL; | |
2661 | } | |
2662 | if (eb->batch_start_offset || eb->args->batch_len) { | |
2663 | intel_context_put(ce); | |
2664 | return -EINVAL; | |
2665 | } | |
2666 | } | |
2667 | eb->num_batches = ce->parallel.number_children + 1; | |
2668 | ||
2669 | for_each_child(ce, child) | |
2670 | intel_context_get(child); | |
2bf541ff | 2671 | intel_gt_pm_get(ce->engine->gt); |
5e2a0419 | 2672 | |
2bf541ff ML |
2673 | if (!test_bit(CONTEXT_ALLOC_BIT, &ce->flags)) { |
2674 | err = intel_context_alloc_state(ce); | |
2675 | if (err) | |
2676 | goto err; | |
2677 | } | |
544460c3 MB |
2678 | for_each_child(ce, child) { |
2679 | if (!test_bit(CONTEXT_ALLOC_BIT, &child->flags)) { | |
2680 | err = intel_context_alloc_state(child); | |
2681 | if (err) | |
2682 | goto err; | |
2683 | } | |
2684 | } | |
2bf541ff ML |
2685 | |
2686 | /* | |
2687 | * ABI: Before userspace accesses the GPU (e.g. execbuffer), report | |
2688 | * EIO if the GPU is already wedged. | |
2689 | */ | |
2690 | err = intel_gt_terminally_wedged(ce->engine->gt); | |
2691 | if (err) | |
2692 | goto err; | |
2693 | ||
2694 | eb->context = ce; | |
544460c3 | 2695 | eb->gt = ce->engine->gt; |
2bf541ff ML |
2696 | |
2697 | /* | |
2698 | * Make sure engine pool stays alive even if we call intel_context_put | |
2699 | * during ww handling. The pool is destroyed when last pm reference | |
2700 | * is dropped, which breaks our -EDEADLK handling. | |
2701 | */ | |
2702 | return err; | |
2703 | ||
2704 | err: | |
2705 | intel_gt_pm_put(ce->engine->gt); | |
544460c3 MB |
2706 | for_each_child(ce, child) |
2707 | intel_context_put(child); | |
2bf541ff | 2708 | intel_context_put(ce); |
5e2a0419 | 2709 | return err; |
de1add36 TU |
2710 | } |
2711 | ||
2bf541ff ML |
2712 | static void |
2713 | eb_put_engine(struct i915_execbuffer *eb) | |
2714 | { | |
544460c3 MB |
2715 | struct intel_context *child; |
2716 | ||
2717 | intel_gt_pm_put(eb->gt); | |
2718 | for_each_child(eb->context, child) | |
2719 | intel_context_put(child); | |
2bf541ff ML |
2720 | intel_context_put(eb->context); |
2721 | } | |
2722 | ||
cf6e7bac | 2723 | static void |
13149e8b | 2724 | __free_fence_array(struct eb_fence *fences, unsigned int n) |
cf6e7bac | 2725 | { |
13149e8b | 2726 | while (n--) { |
cda9edd0 | 2727 | drm_syncobj_put(ptr_mask_bits(fences[n].syncobj, 2)); |
13149e8b | 2728 | dma_fence_put(fences[n].dma_fence); |
440d0f12 | 2729 | dma_fence_chain_free(fences[n].chain_fence); |
13149e8b | 2730 | } |
cf6e7bac JE |
2731 | kvfree(fences); |
2732 | } | |
2733 | ||
cda9edd0 | 2734 | static int |
13149e8b LL |
2735 | add_timeline_fence_array(struct i915_execbuffer *eb, |
2736 | const struct drm_i915_gem_execbuffer_ext_timeline_fences *timeline_fences) | |
cf6e7bac | 2737 | { |
13149e8b LL |
2738 | struct drm_i915_gem_exec_fence __user *user_fences; |
2739 | u64 __user *user_values; | |
2740 | struct eb_fence *f; | |
2741 | u64 nfences; | |
2742 | int err = 0; | |
cf6e7bac | 2743 | |
13149e8b LL |
2744 | nfences = timeline_fences->fence_count; |
2745 | if (!nfences) | |
cda9edd0 | 2746 | return 0; |
cf6e7bac | 2747 | |
d710fc16 CW |
2748 | /* Check multiplication overflow for access_ok() and kvmalloc_array() */ |
2749 | BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long)); | |
2750 | if (nfences > min_t(unsigned long, | |
13149e8b LL |
2751 | ULONG_MAX / sizeof(*user_fences), |
2752 | SIZE_MAX / sizeof(*f)) - eb->num_fences) | |
cda9edd0 | 2753 | return -EINVAL; |
cf6e7bac | 2754 | |
13149e8b LL |
2755 | user_fences = u64_to_user_ptr(timeline_fences->handles_ptr); |
2756 | if (!access_ok(user_fences, nfences * sizeof(*user_fences))) | |
2757 | return -EFAULT; | |
2758 | ||
2759 | user_values = u64_to_user_ptr(timeline_fences->values_ptr); | |
2760 | if (!access_ok(user_values, nfences * sizeof(*user_values))) | |
cda9edd0 | 2761 | return -EFAULT; |
cf6e7bac | 2762 | |
13149e8b LL |
2763 | f = krealloc(eb->fences, |
2764 | (eb->num_fences + nfences) * sizeof(*f), | |
2765 | __GFP_NOWARN | GFP_KERNEL); | |
2766 | if (!f) | |
cda9edd0 | 2767 | return -ENOMEM; |
cf6e7bac | 2768 | |
13149e8b LL |
2769 | eb->fences = f; |
2770 | f += eb->num_fences; | |
2771 | ||
2772 | BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) & | |
2773 | ~__I915_EXEC_FENCE_UNKNOWN_FLAGS); | |
2774 | ||
2775 | while (nfences--) { | |
2776 | struct drm_i915_gem_exec_fence user_fence; | |
cf6e7bac | 2777 | struct drm_syncobj *syncobj; |
13149e8b LL |
2778 | struct dma_fence *fence = NULL; |
2779 | u64 point; | |
2780 | ||
2781 | if (__copy_from_user(&user_fence, | |
2782 | user_fences++, | |
2783 | sizeof(user_fence))) | |
2784 | return -EFAULT; | |
2785 | ||
2786 | if (user_fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS) | |
2787 | return -EINVAL; | |
2788 | ||
2789 | if (__get_user(point, user_values++)) | |
2790 | return -EFAULT; | |
2791 | ||
2792 | syncobj = drm_syncobj_find(eb->file, user_fence.handle); | |
2793 | if (!syncobj) { | |
2794 | DRM_DEBUG("Invalid syncobj handle provided\n"); | |
2795 | return -ENOENT; | |
2796 | } | |
2797 | ||
2798 | fence = drm_syncobj_fence_get(syncobj); | |
cf6e7bac | 2799 | |
13149e8b LL |
2800 | if (!fence && user_fence.flags && |
2801 | !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) { | |
2802 | DRM_DEBUG("Syncobj handle has no fence\n"); | |
2803 | drm_syncobj_put(syncobj); | |
2804 | return -EINVAL; | |
cf6e7bac JE |
2805 | } |
2806 | ||
13149e8b LL |
2807 | if (fence) |
2808 | err = dma_fence_chain_find_seqno(&fence, point); | |
2809 | ||
2810 | if (err && !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) { | |
2811 | DRM_DEBUG("Syncobj handle missing requested point %llu\n", point); | |
da1ea128 | 2812 | dma_fence_put(fence); |
13149e8b LL |
2813 | drm_syncobj_put(syncobj); |
2814 | return err; | |
2815 | } | |
2816 | ||
2817 | /* | |
2818 | * A point might have been signaled already and | |
2819 | * garbage collected from the timeline. In this case | |
2820 | * just ignore the point and carry on. | |
2821 | */ | |
2822 | if (!fence && !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) { | |
2823 | drm_syncobj_put(syncobj); | |
2824 | continue; | |
2825 | } | |
2826 | ||
2827 | /* | |
2828 | * For timeline syncobjs we need to preallocate chains for | |
2829 | * later signaling. | |
2830 | */ | |
2831 | if (point != 0 && user_fence.flags & I915_EXEC_FENCE_SIGNAL) { | |
2832 | /* | |
2833 | * Waiting and signaling the same point (when point != | |
2834 | * 0) would break the timeline. | |
2835 | */ | |
2836 | if (user_fence.flags & I915_EXEC_FENCE_WAIT) { | |
2837 | DRM_DEBUG("Trying to wait & signal the same timeline point.\n"); | |
2838 | dma_fence_put(fence); | |
2839 | drm_syncobj_put(syncobj); | |
2840 | return -EINVAL; | |
2841 | } | |
2842 | ||
440d0f12 | 2843 | f->chain_fence = dma_fence_chain_alloc(); |
13149e8b LL |
2844 | if (!f->chain_fence) { |
2845 | drm_syncobj_put(syncobj); | |
2846 | dma_fence_put(fence); | |
2847 | return -ENOMEM; | |
2848 | } | |
2849 | } else { | |
2850 | f->chain_fence = NULL; | |
ebcaa1ff TU |
2851 | } |
2852 | ||
13149e8b LL |
2853 | f->syncobj = ptr_pack_bits(syncobj, user_fence.flags, 2); |
2854 | f->dma_fence = fence; | |
2855 | f->value = point; | |
2856 | f++; | |
2857 | eb->num_fences++; | |
2858 | } | |
2859 | ||
2860 | return 0; | |
2861 | } | |
2862 | ||
2863 | static int add_fence_array(struct i915_execbuffer *eb) | |
2864 | { | |
2865 | struct drm_i915_gem_execbuffer2 *args = eb->args; | |
2866 | struct drm_i915_gem_exec_fence __user *user; | |
2867 | unsigned long num_fences = args->num_cliprects; | |
2868 | struct eb_fence *f; | |
2869 | ||
2870 | if (!(args->flags & I915_EXEC_FENCE_ARRAY)) | |
2871 | return 0; | |
2872 | ||
2873 | if (!num_fences) | |
2874 | return 0; | |
2875 | ||
2876 | /* Check multiplication overflow for access_ok() and kvmalloc_array() */ | |
2877 | BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long)); | |
2878 | if (num_fences > min_t(unsigned long, | |
2879 | ULONG_MAX / sizeof(*user), | |
2880 | SIZE_MAX / sizeof(*f) - eb->num_fences)) | |
2881 | return -EINVAL; | |
2882 | ||
2883 | user = u64_to_user_ptr(args->cliprects_ptr); | |
2884 | if (!access_ok(user, num_fences * sizeof(*user))) | |
2885 | return -EFAULT; | |
2886 | ||
2887 | f = krealloc(eb->fences, | |
2888 | (eb->num_fences + num_fences) * sizeof(*f), | |
2889 | __GFP_NOWARN | GFP_KERNEL); | |
2890 | if (!f) | |
2891 | return -ENOMEM; | |
2892 | ||
2893 | eb->fences = f; | |
2894 | f += eb->num_fences; | |
2895 | while (num_fences--) { | |
2896 | struct drm_i915_gem_exec_fence user_fence; | |
2897 | struct drm_syncobj *syncobj; | |
2898 | struct dma_fence *fence = NULL; | |
2899 | ||
2900 | if (__copy_from_user(&user_fence, user++, sizeof(user_fence))) | |
2901 | return -EFAULT; | |
2902 | ||
2903 | if (user_fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS) | |
2904 | return -EINVAL; | |
2905 | ||
2906 | syncobj = drm_syncobj_find(eb->file, user_fence.handle); | |
cf6e7bac JE |
2907 | if (!syncobj) { |
2908 | DRM_DEBUG("Invalid syncobj handle provided\n"); | |
13149e8b LL |
2909 | return -ENOENT; |
2910 | } | |
2911 | ||
2912 | if (user_fence.flags & I915_EXEC_FENCE_WAIT) { | |
2913 | fence = drm_syncobj_fence_get(syncobj); | |
2914 | if (!fence) { | |
2915 | DRM_DEBUG("Syncobj handle has no fence\n"); | |
2916 | drm_syncobj_put(syncobj); | |
2917 | return -EINVAL; | |
2918 | } | |
cf6e7bac JE |
2919 | } |
2920 | ||
ebcaa1ff TU |
2921 | BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) & |
2922 | ~__I915_EXEC_FENCE_UNKNOWN_FLAGS); | |
2923 | ||
13149e8b LL |
2924 | f->syncobj = ptr_pack_bits(syncobj, user_fence.flags, 2); |
2925 | f->dma_fence = fence; | |
2926 | f->value = 0; | |
2927 | f->chain_fence = NULL; | |
2928 | f++; | |
2929 | eb->num_fences++; | |
cf6e7bac JE |
2930 | } |
2931 | ||
cda9edd0 | 2932 | return 0; |
13149e8b | 2933 | } |
cf6e7bac | 2934 | |
13149e8b LL |
2935 | static void put_fence_array(struct eb_fence *fences, int num_fences) |
2936 | { | |
2937 | if (fences) | |
2938 | __free_fence_array(fences, num_fences); | |
cf6e7bac JE |
2939 | } |
2940 | ||
2941 | static int | |
544460c3 MB |
2942 | await_fence_array(struct i915_execbuffer *eb, |
2943 | struct i915_request *rq) | |
cf6e7bac | 2944 | { |
cf6e7bac JE |
2945 | unsigned int n; |
2946 | int err; | |
2947 | ||
13149e8b | 2948 | for (n = 0; n < eb->num_fences; n++) { |
cf6e7bac | 2949 | struct drm_syncobj *syncobj; |
cf6e7bac JE |
2950 | unsigned int flags; |
2951 | ||
cda9edd0 | 2952 | syncobj = ptr_unpack_bits(eb->fences[n].syncobj, &flags, 2); |
cf6e7bac | 2953 | |
13149e8b LL |
2954 | if (!eb->fences[n].dma_fence) |
2955 | continue; | |
cf6e7bac | 2956 | |
544460c3 | 2957 | err = i915_request_await_dma_fence(rq, eb->fences[n].dma_fence); |
cf6e7bac JE |
2958 | if (err < 0) |
2959 | return err; | |
2960 | } | |
2961 | ||
2962 | return 0; | |
2963 | } | |
2964 | ||
544460c3 MB |
2965 | static void signal_fence_array(const struct i915_execbuffer *eb, |
2966 | struct dma_fence * const fence) | |
cf6e7bac | 2967 | { |
cf6e7bac JE |
2968 | unsigned int n; |
2969 | ||
13149e8b | 2970 | for (n = 0; n < eb->num_fences; n++) { |
cf6e7bac JE |
2971 | struct drm_syncobj *syncobj; |
2972 | unsigned int flags; | |
2973 | ||
cda9edd0 | 2974 | syncobj = ptr_unpack_bits(eb->fences[n].syncobj, &flags, 2); |
cf6e7bac JE |
2975 | if (!(flags & I915_EXEC_FENCE_SIGNAL)) |
2976 | continue; | |
2977 | ||
13149e8b LL |
2978 | if (eb->fences[n].chain_fence) { |
2979 | drm_syncobj_add_point(syncobj, | |
2980 | eb->fences[n].chain_fence, | |
2981 | fence, | |
2982 | eb->fences[n].value); | |
2983 | /* | |
2984 | * The chain's ownership is transferred to the | |
2985 | * timeline. | |
2986 | */ | |
2987 | eb->fences[n].chain_fence = NULL; | |
2988 | } else { | |
2989 | drm_syncobj_replace_fence(syncobj, fence); | |
2990 | } | |
cf6e7bac JE |
2991 | } |
2992 | } | |
2993 | ||
13149e8b LL |
2994 | static int |
2995 | parse_timeline_fences(struct i915_user_extension __user *ext, void *data) | |
2996 | { | |
2997 | struct i915_execbuffer *eb = data; | |
2998 | struct drm_i915_gem_execbuffer_ext_timeline_fences timeline_fences; | |
2999 | ||
3000 | if (copy_from_user(&timeline_fences, ext, sizeof(timeline_fences))) | |
3001 | return -EFAULT; | |
3002 | ||
3003 | return add_timeline_fence_array(eb, &timeline_fences); | |
3004 | } | |
3005 | ||
61231f6b CW |
3006 | static void retire_requests(struct intel_timeline *tl, struct i915_request *end) |
3007 | { | |
3008 | struct i915_request *rq, *rn; | |
3009 | ||
3010 | list_for_each_entry_safe(rq, rn, &tl->requests, link) | |
3011 | if (rq == end || !i915_request_retire(rq)) | |
3012 | break; | |
3013 | } | |
3014 | ||
544460c3 MB |
3015 | static int eb_request_add(struct i915_execbuffer *eb, struct i915_request *rq, |
3016 | int err, bool last_parallel) | |
61231f6b | 3017 | { |
61231f6b CW |
3018 | struct intel_timeline * const tl = i915_request_timeline(rq); |
3019 | struct i915_sched_attr attr = {}; | |
3020 | struct i915_request *prev; | |
3021 | ||
3022 | lockdep_assert_held(&tl->mutex); | |
3023 | lockdep_unpin_lock(&tl->mutex, rq->cookie); | |
3024 | ||
3025 | trace_i915_request_add(rq); | |
3026 | ||
3027 | prev = __i915_request_commit(rq); | |
3028 | ||
3029 | /* Check that the context wasn't destroyed before submission */ | |
207e4a71 | 3030 | if (likely(!intel_context_is_closed(eb->context))) { |
61231f6b | 3031 | attr = eb->gem_context->sched; |
61231f6b CW |
3032 | } else { |
3033 | /* Serialise with context_close via the add_to_timeline */ | |
36e191f0 CW |
3034 | i915_request_set_error_once(rq, -ENOENT); |
3035 | __i915_request_skip(rq); | |
ba38b79e | 3036 | err = -ENOENT; /* override any transient errors */ |
61231f6b CW |
3037 | } |
3038 | ||
544460c3 MB |
3039 | if (intel_context_is_parallel(eb->context)) { |
3040 | if (err) { | |
3041 | __i915_request_skip(rq); | |
3042 | set_bit(I915_FENCE_FLAG_SKIP_PARALLEL, | |
3043 | &rq->fence.flags); | |
3044 | } | |
3045 | if (last_parallel) | |
3046 | set_bit(I915_FENCE_FLAG_SUBMIT_PARALLEL, | |
3047 | &rq->fence.flags); | |
3048 | } | |
3049 | ||
61231f6b | 3050 | __i915_request_queue(rq, &attr); |
61231f6b CW |
3051 | |
3052 | /* Try to clean up the client's timeline after submitting the request */ | |
3053 | if (prev) | |
3054 | retire_requests(tl, prev); | |
3055 | ||
3056 | mutex_unlock(&tl->mutex); | |
ba38b79e CW |
3057 | |
3058 | return err; | |
61231f6b CW |
3059 | } |
3060 | ||
544460c3 MB |
3061 | static int eb_requests_add(struct i915_execbuffer *eb, int err) |
3062 | { | |
3063 | int i; | |
3064 | ||
3065 | /* | |
3066 | * We iterate in reverse order of creation to release timeline mutexes in | |
3067 | * same order. | |
3068 | */ | |
3069 | for_each_batch_add_order(eb, i) { | |
3070 | struct i915_request *rq = eb->requests[i]; | |
3071 | ||
3072 | if (!rq) | |
3073 | continue; | |
3074 | err |= eb_request_add(eb, rq, err, i == 0); | |
3075 | } | |
3076 | ||
3077 | return err; | |
3078 | } | |
3079 | ||
cda9edd0 | 3080 | static const i915_user_extension_fn execbuf_extensions[] = { |
13149e8b | 3081 | [DRM_I915_GEM_EXECBUFFER_EXT_TIMELINE_FENCES] = parse_timeline_fences, |
cda9edd0 LL |
3082 | }; |
3083 | ||
3084 | static int | |
3085 | parse_execbuf2_extensions(struct drm_i915_gem_execbuffer2 *args, | |
3086 | struct i915_execbuffer *eb) | |
3087 | { | |
cda9edd0 LL |
3088 | if (!(args->flags & I915_EXEC_USE_EXTENSIONS)) |
3089 | return 0; | |
3090 | ||
3091 | /* The execbuf2 extension mechanism reuses cliprects_ptr. So we cannot | |
3092 | * have another flag also using it at the same time. | |
3093 | */ | |
3094 | if (eb->args->flags & I915_EXEC_FENCE_ARRAY) | |
3095 | return -EINVAL; | |
3096 | ||
3097 | if (args->num_cliprects != 0) | |
3098 | return -EINVAL; | |
3099 | ||
3100 | return i915_user_extensions(u64_to_user_ptr(args->cliprects_ptr), | |
3101 | execbuf_extensions, | |
3102 | ARRAY_SIZE(execbuf_extensions), | |
3103 | eb); | |
3104 | } | |
3105 | ||
544460c3 MB |
3106 | static void eb_requests_get(struct i915_execbuffer *eb) |
3107 | { | |
3108 | unsigned int i; | |
3109 | ||
3110 | for_each_batch_create_order(eb, i) { | |
3111 | if (!eb->requests[i]) | |
3112 | break; | |
3113 | ||
3114 | i915_request_get(eb->requests[i]); | |
3115 | } | |
3116 | } | |
3117 | ||
3118 | static void eb_requests_put(struct i915_execbuffer *eb) | |
3119 | { | |
3120 | unsigned int i; | |
3121 | ||
3122 | for_each_batch_create_order(eb, i) { | |
3123 | if (!eb->requests[i]) | |
3124 | break; | |
3125 | ||
3126 | i915_request_put(eb->requests[i]); | |
3127 | } | |
3128 | } | |
3129 | ||
3130 | static struct sync_file * | |
3131 | eb_composite_fence_create(struct i915_execbuffer *eb, int out_fence_fd) | |
3132 | { | |
3133 | struct sync_file *out_fence = NULL; | |
3134 | struct dma_fence_array *fence_array; | |
3135 | struct dma_fence **fences; | |
3136 | unsigned int i; | |
3137 | ||
3138 | GEM_BUG_ON(!intel_context_is_parent(eb->context)); | |
3139 | ||
3140 | fences = kmalloc_array(eb->num_batches, sizeof(*fences), GFP_KERNEL); | |
3141 | if (!fences) | |
3142 | return ERR_PTR(-ENOMEM); | |
3143 | ||
7647f009 | 3144 | for_each_batch_create_order(eb, i) { |
544460c3 | 3145 | fences[i] = &eb->requests[i]->fence; |
7647f009 MB |
3146 | __set_bit(I915_FENCE_FLAG_COMPOSITE, |
3147 | &eb->requests[i]->fence.flags); | |
3148 | } | |
544460c3 MB |
3149 | |
3150 | fence_array = dma_fence_array_create(eb->num_batches, | |
3151 | fences, | |
3152 | eb->context->parallel.fence_context, | |
62eeb9ae | 3153 | eb->context->parallel.seqno++, |
544460c3 MB |
3154 | false); |
3155 | if (!fence_array) { | |
3156 | kfree(fences); | |
3157 | return ERR_PTR(-ENOMEM); | |
3158 | } | |
3159 | ||
3160 | /* Move ownership to the dma_fence_array created above */ | |
3161 | for_each_batch_create_order(eb, i) | |
3162 | dma_fence_get(fences[i]); | |
3163 | ||
3164 | if (out_fence_fd != -1) { | |
3165 | out_fence = sync_file_create(&fence_array->base); | |
3166 | /* sync_file now owns fence_arry, drop creation ref */ | |
3167 | dma_fence_put(&fence_array->base); | |
3168 | if (!out_fence) | |
3169 | return ERR_PTR(-ENOMEM); | |
3170 | } | |
3171 | ||
3172 | eb->composite_fence = &fence_array->base; | |
3173 | ||
3174 | return out_fence; | |
3175 | } | |
3176 | ||
3177 | static struct sync_file * | |
3178 | eb_fences_add(struct i915_execbuffer *eb, struct i915_request *rq, | |
3179 | struct dma_fence *in_fence, int out_fence_fd) | |
3180 | { | |
3181 | struct sync_file *out_fence = NULL; | |
3182 | int err; | |
3183 | ||
3184 | if (unlikely(eb->gem_context->syncobj)) { | |
3185 | struct dma_fence *fence; | |
3186 | ||
3187 | fence = drm_syncobj_fence_get(eb->gem_context->syncobj); | |
3188 | err = i915_request_await_dma_fence(rq, fence); | |
3189 | dma_fence_put(fence); | |
3190 | if (err) | |
3191 | return ERR_PTR(err); | |
3192 | } | |
3193 | ||
3194 | if (in_fence) { | |
3195 | if (eb->args->flags & I915_EXEC_FENCE_SUBMIT) | |
3196 | err = i915_request_await_execution(rq, in_fence); | |
3197 | else | |
3198 | err = i915_request_await_dma_fence(rq, in_fence); | |
3199 | if (err < 0) | |
3200 | return ERR_PTR(err); | |
3201 | } | |
3202 | ||
3203 | if (eb->fences) { | |
3204 | err = await_fence_array(eb, rq); | |
3205 | if (err) | |
3206 | return ERR_PTR(err); | |
3207 | } | |
3208 | ||
3209 | if (intel_context_is_parallel(eb->context)) { | |
3210 | out_fence = eb_composite_fence_create(eb, out_fence_fd); | |
3211 | if (IS_ERR(out_fence)) | |
3212 | return ERR_PTR(-ENOMEM); | |
3213 | } else if (out_fence_fd != -1) { | |
3214 | out_fence = sync_file_create(&rq->fence); | |
3215 | if (!out_fence) | |
3216 | return ERR_PTR(-ENOMEM); | |
3217 | } | |
3218 | ||
3219 | return out_fence; | |
3220 | } | |
3221 | ||
3222 | static struct intel_context * | |
3223 | eb_find_context(struct i915_execbuffer *eb, unsigned int context_number) | |
3224 | { | |
3225 | struct intel_context *child; | |
3226 | ||
3227 | if (likely(context_number == 0)) | |
3228 | return eb->context; | |
3229 | ||
3230 | for_each_child(eb->context, child) | |
3231 | if (!--context_number) | |
3232 | return child; | |
3233 | ||
3234 | GEM_BUG_ON("Context not found"); | |
3235 | ||
3236 | return NULL; | |
3237 | } | |
3238 | ||
3239 | static struct sync_file * | |
3240 | eb_requests_create(struct i915_execbuffer *eb, struct dma_fence *in_fence, | |
3241 | int out_fence_fd) | |
3242 | { | |
3243 | struct sync_file *out_fence = NULL; | |
3244 | unsigned int i; | |
3245 | ||
3246 | for_each_batch_create_order(eb, i) { | |
3247 | /* Allocate a request for this batch buffer nice and early. */ | |
3248 | eb->requests[i] = i915_request_create(eb_find_context(eb, i)); | |
3249 | if (IS_ERR(eb->requests[i])) { | |
86752bd6 | 3250 | out_fence = ERR_CAST(eb->requests[i]); |
544460c3 MB |
3251 | eb->requests[i] = NULL; |
3252 | return out_fence; | |
3253 | } | |
3254 | ||
3255 | /* | |
3256 | * Only the first request added (committed to backend) has to | |
3257 | * take the in fences into account as all subsequent requests | |
3258 | * will have fences inserted inbetween them. | |
3259 | */ | |
3260 | if (i + 1 == eb->num_batches) { | |
3261 | out_fence = eb_fences_add(eb, eb->requests[i], | |
3262 | in_fence, out_fence_fd); | |
3263 | if (IS_ERR(out_fence)) | |
3264 | return out_fence; | |
3265 | } | |
3266 | ||
3267 | /* | |
ff20afc4 TH |
3268 | * Not really on stack, but we don't want to call |
3269 | * kfree on the batch_snapshot when we put it, so use the | |
3270 | * _onstack interface. | |
544460c3 | 3271 | */ |
ff20afc4 TH |
3272 | if (eb->batches[i]->vma) |
3273 | i915_vma_snapshot_init_onstack(&eb->requests[i]->batch_snapshot, | |
3274 | eb->batches[i]->vma, | |
3275 | "batch"); | |
544460c3 MB |
3276 | if (eb->batch_pool) { |
3277 | GEM_BUG_ON(intel_context_is_parallel(eb->context)); | |
3278 | intel_gt_buffer_pool_mark_active(eb->batch_pool, | |
3279 | eb->requests[i]); | |
3280 | } | |
3281 | } | |
3282 | ||
3283 | return out_fence; | |
3284 | } | |
3285 | ||
54cf91dc | 3286 | static int |
650bc635 | 3287 | i915_gem_do_execbuffer(struct drm_device *dev, |
54cf91dc CW |
3288 | struct drm_file *file, |
3289 | struct drm_i915_gem_execbuffer2 *args, | |
cda9edd0 | 3290 | struct drm_i915_gem_exec_object2 *exec) |
54cf91dc | 3291 | { |
44157641 | 3292 | struct drm_i915_private *i915 = to_i915(dev); |
650bc635 | 3293 | struct i915_execbuffer eb; |
fec0445c CW |
3294 | struct dma_fence *in_fence = NULL; |
3295 | struct sync_file *out_fence = NULL; | |
3296 | int out_fence_fd = -1; | |
2889caa9 | 3297 | int err; |
432e58ed | 3298 | |
74c1c694 | 3299 | BUILD_BUG_ON(__EXEC_INTERNAL_FLAGS & ~__I915_EXEC_ILLEGAL_FLAGS); |
2889caa9 CW |
3300 | BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS & |
3301 | ~__EXEC_OBJECT_UNKNOWN_FLAGS); | |
54cf91dc | 3302 | |
44157641 | 3303 | eb.i915 = i915; |
650bc635 CW |
3304 | eb.file = file; |
3305 | eb.args = args; | |
ad5d95e4 | 3306 | if (DBG_FORCE_RELOC || !(args->flags & I915_EXEC_NO_RELOC)) |
2889caa9 | 3307 | args->flags |= __EXEC_HAS_RELOC; |
c7c6e46f | 3308 | |
650bc635 | 3309 | eb.exec = exec; |
8ae275c2 ML |
3310 | eb.vma = (struct eb_vma *)(exec + args->buffer_count + 1); |
3311 | eb.vma[0].vma = NULL; | |
8e02cceb | 3312 | eb.batch_pool = NULL; |
c7c6e46f | 3313 | |
2889caa9 | 3314 | eb.invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS; |
650bc635 CW |
3315 | reloc_cache_init(&eb.reloc_cache, eb.i915); |
3316 | ||
2889caa9 | 3317 | eb.buffer_count = args->buffer_count; |
650bc635 | 3318 | eb.batch_start_offset = args->batch_start_offset; |
32d94048 | 3319 | eb.trampoline = NULL; |
650bc635 | 3320 | |
cda9edd0 | 3321 | eb.fences = NULL; |
13149e8b | 3322 | eb.num_fences = 0; |
cda9edd0 | 3323 | |
ff20afc4 TH |
3324 | eb_capture_list_clear(&eb); |
3325 | ||
544460c3 MB |
3326 | memset(eb.requests, 0, sizeof(struct i915_request *) * |
3327 | ARRAY_SIZE(eb.requests)); | |
3328 | eb.composite_fence = NULL; | |
3329 | ||
2889caa9 | 3330 | eb.batch_flags = 0; |
d7d4eedd | 3331 | if (args->flags & I915_EXEC_SECURE) { |
40e1956e | 3332 | if (GRAPHICS_VER(i915) >= 11) |
44157641 JB |
3333 | return -ENODEV; |
3334 | ||
3335 | /* Return -EPERM to trigger fallback code on old binaries. */ | |
3336 | if (!HAS_SECURE_BATCHES(i915)) | |
3337 | return -EPERM; | |
3338 | ||
b3ac9f25 | 3339 | if (!drm_is_current_master(file) || !capable(CAP_SYS_ADMIN)) |
44157641 | 3340 | return -EPERM; |
d7d4eedd | 3341 | |
2889caa9 | 3342 | eb.batch_flags |= I915_DISPATCH_SECURE; |
d7d4eedd | 3343 | } |
b45305fc | 3344 | if (args->flags & I915_EXEC_IS_PINNED) |
2889caa9 | 3345 | eb.batch_flags |= I915_DISPATCH_PINNED; |
54cf91dc | 3346 | |
13149e8b LL |
3347 | err = parse_execbuf2_extensions(args, &eb); |
3348 | if (err) | |
3349 | goto err_ext; | |
3350 | ||
3351 | err = add_fence_array(&eb); | |
3352 | if (err) | |
3353 | goto err_ext; | |
3354 | ||
889333c7 CW |
3355 | #define IN_FENCES (I915_EXEC_FENCE_IN | I915_EXEC_FENCE_SUBMIT) |
3356 | if (args->flags & IN_FENCES) { | |
3357 | if ((args->flags & IN_FENCES) == IN_FENCES) | |
3358 | return -EINVAL; | |
3359 | ||
fec0445c | 3360 | in_fence = sync_file_get_fence(lower_32_bits(args->rsvd2)); |
13149e8b LL |
3361 | if (!in_fence) { |
3362 | err = -EINVAL; | |
3363 | goto err_ext; | |
3364 | } | |
fec0445c | 3365 | } |
889333c7 | 3366 | #undef IN_FENCES |
a88b6e4c | 3367 | |
fec0445c CW |
3368 | if (args->flags & I915_EXEC_FENCE_OUT) { |
3369 | out_fence_fd = get_unused_fd_flags(O_CLOEXEC); | |
3370 | if (out_fence_fd < 0) { | |
2889caa9 | 3371 | err = out_fence_fd; |
889333c7 | 3372 | goto err_in_fence; |
fec0445c CW |
3373 | } |
3374 | } | |
3375 | ||
cda9edd0 LL |
3376 | err = eb_create(&eb); |
3377 | if (err) | |
13149e8b | 3378 | goto err_out_fence; |
cda9edd0 | 3379 | |
4d470f73 | 3380 | GEM_BUG_ON(!eb.lut_size); |
2889caa9 | 3381 | |
1acfc104 CW |
3382 | err = eb_select_context(&eb); |
3383 | if (unlikely(err)) | |
3384 | goto err_destroy; | |
3385 | ||
2bf541ff | 3386 | err = eb_select_engine(&eb); |
d6f328bf | 3387 | if (unlikely(err)) |
e5dadff4 | 3388 | goto err_context; |
d6f328bf | 3389 | |
c43ce123 ML |
3390 | err = eb_lookup_vmas(&eb); |
3391 | if (err) { | |
b4b9731b | 3392 | eb_release_vmas(&eb, true); |
c43ce123 ML |
3393 | goto err_engine; |
3394 | } | |
3395 | ||
3396 | i915_gem_ww_ctx_init(&eb.ww, true); | |
3397 | ||
8e4ba491 | 3398 | err = eb_relocate_parse(&eb); |
1f727d9e | 3399 | if (err) { |
2889caa9 CW |
3400 | /* |
3401 | * If the user expects the execobject.offset and | |
3402 | * reloc.presumed_offset to be an exact match, | |
3403 | * as for using NO_RELOC, then we cannot update | |
3404 | * the execobject.offset until we have completed | |
3405 | * relocation. | |
3406 | */ | |
3407 | args->flags &= ~__EXEC_HAS_RELOC; | |
2889caa9 | 3408 | goto err_vma; |
1f727d9e | 3409 | } |
54cf91dc | 3410 | |
c43ce123 | 3411 | ww_acquire_done(&eb.ww.ctx); |
ff20afc4 | 3412 | eb_capture_stage(&eb); |
7d6236bb | 3413 | |
544460c3 MB |
3414 | out_fence = eb_requests_create(&eb, in_fence, out_fence_fd); |
3415 | if (IS_ERR(out_fence)) { | |
3416 | err = PTR_ERR(out_fence); | |
8722ded4 | 3417 | out_fence = NULL; |
544460c3 | 3418 | if (eb.requests[0]) |
cf6e7bac | 3419 | goto err_request; |
544460c3 MB |
3420 | else |
3421 | goto err_vma; | |
cf6e7bac JE |
3422 | } |
3423 | ||
544460c3 | 3424 | err = eb_submit(&eb); |
ed29c269 | 3425 | |
aa9b7810 | 3426 | err_request: |
544460c3 MB |
3427 | eb_requests_get(&eb); |
3428 | err = eb_requests_add(&eb, err); | |
c8659efa | 3429 | |
13149e8b | 3430 | if (eb.fences) |
544460c3 MB |
3431 | signal_fence_array(&eb, eb.composite_fence ? |
3432 | eb.composite_fence : | |
3433 | &eb.requests[0]->fence); | |
cf6e7bac | 3434 | |
fec0445c | 3435 | if (out_fence) { |
2889caa9 | 3436 | if (err == 0) { |
fec0445c | 3437 | fd_install(out_fence_fd, out_fence->file); |
b6a88e4a | 3438 | args->rsvd2 &= GENMASK_ULL(31, 0); /* keep in-fence */ |
fec0445c CW |
3439 | args->rsvd2 |= (u64)out_fence_fd << 32; |
3440 | out_fence_fd = -1; | |
3441 | } else { | |
3442 | fput(out_fence->file); | |
3443 | } | |
3444 | } | |
00dae4d3 JE |
3445 | |
3446 | if (unlikely(eb.gem_context->syncobj)) { | |
3447 | drm_syncobj_replace_fence(eb.gem_context->syncobj, | |
544460c3 MB |
3448 | eb.composite_fence ? |
3449 | eb.composite_fence : | |
3450 | &eb.requests[0]->fence); | |
00dae4d3 JE |
3451 | } |
3452 | ||
544460c3 MB |
3453 | if (!out_fence && eb.composite_fence) |
3454 | dma_fence_put(eb.composite_fence); | |
3455 | ||
3456 | eb_requests_put(&eb); | |
54cf91dc | 3457 | |
2889caa9 | 3458 | err_vma: |
b4b9731b | 3459 | eb_release_vmas(&eb, true); |
32d94048 CW |
3460 | if (eb.trampoline) |
3461 | i915_vma_unpin(eb.trampoline); | |
c43ce123 ML |
3462 | WARN_ON(err == -EDEADLK); |
3463 | i915_gem_ww_ctx_fini(&eb.ww); | |
3464 | ||
3465 | if (eb.batch_pool) | |
3466 | intel_gt_buffer_pool_put(eb.batch_pool); | |
c43ce123 | 3467 | err_engine: |
2bf541ff | 3468 | eb_put_engine(&eb); |
a4e57f90 | 3469 | err_context: |
8f2a1057 | 3470 | i915_gem_context_put(eb.gem_context); |
1acfc104 | 3471 | err_destroy: |
2889caa9 | 3472 | eb_destroy(&eb); |
4d470f73 | 3473 | err_out_fence: |
fec0445c CW |
3474 | if (out_fence_fd != -1) |
3475 | put_unused_fd(out_fence_fd); | |
4a04e371 | 3476 | err_in_fence: |
fec0445c | 3477 | dma_fence_put(in_fence); |
13149e8b LL |
3478 | err_ext: |
3479 | put_fence_array(eb.fences, eb.num_fences); | |
2889caa9 | 3480 | return err; |
54cf91dc CW |
3481 | } |
3482 | ||
d710fc16 CW |
3483 | static size_t eb_element_size(void) |
3484 | { | |
8ae275c2 | 3485 | return sizeof(struct drm_i915_gem_exec_object2) + sizeof(struct eb_vma); |
d710fc16 CW |
3486 | } |
3487 | ||
3488 | static bool check_buffer_count(size_t count) | |
3489 | { | |
3490 | const size_t sz = eb_element_size(); | |
3491 | ||
3492 | /* | |
3493 | * When using LUT_HANDLE, we impose a limit of INT_MAX for the lookup | |
3494 | * array size (see eb_create()). Otherwise, we can accept an array as | |
3495 | * large as can be addressed (though use large arrays at your peril)! | |
3496 | */ | |
3497 | ||
3498 | return !(count < 1 || count > INT_MAX || count > SIZE_MAX / sz - 1); | |
3499 | } | |
3500 | ||
54cf91dc | 3501 | int |
6a20fe7b VS |
3502 | i915_gem_execbuffer2_ioctl(struct drm_device *dev, void *data, |
3503 | struct drm_file *file) | |
54cf91dc | 3504 | { |
d0bf4582 | 3505 | struct drm_i915_private *i915 = to_i915(dev); |
54cf91dc | 3506 | struct drm_i915_gem_execbuffer2 *args = data; |
2889caa9 | 3507 | struct drm_i915_gem_exec_object2 *exec2_list; |
d710fc16 | 3508 | const size_t count = args->buffer_count; |
2889caa9 | 3509 | int err; |
54cf91dc | 3510 | |
d710fc16 | 3511 | if (!check_buffer_count(count)) { |
d0bf4582 | 3512 | drm_dbg(&i915->drm, "execbuf2 with %zd buffers\n", count); |
54cf91dc CW |
3513 | return -EINVAL; |
3514 | } | |
3515 | ||
00aff3f6 TU |
3516 | err = i915_gem_check_execbuffer(args); |
3517 | if (err) | |
3518 | return err; | |
2889caa9 | 3519 | |
47b08693 ML |
3520 | /* Allocate extra slots for use by the command parser */ |
3521 | exec2_list = kvmalloc_array(count + 2, eb_element_size(), | |
0ee931c4 | 3522 | __GFP_NOWARN | GFP_KERNEL); |
54cf91dc | 3523 | if (exec2_list == NULL) { |
d0bf4582 WK |
3524 | drm_dbg(&i915->drm, "Failed to allocate exec list for %zd buffers\n", |
3525 | count); | |
54cf91dc CW |
3526 | return -ENOMEM; |
3527 | } | |
2889caa9 CW |
3528 | if (copy_from_user(exec2_list, |
3529 | u64_to_user_ptr(args->buffers_ptr), | |
d710fc16 | 3530 | sizeof(*exec2_list) * count)) { |
d0bf4582 | 3531 | drm_dbg(&i915->drm, "copy %zd exec entries failed\n", count); |
2098105e | 3532 | kvfree(exec2_list); |
54cf91dc CW |
3533 | return -EFAULT; |
3534 | } | |
3535 | ||
cda9edd0 | 3536 | err = i915_gem_do_execbuffer(dev, file, args, exec2_list); |
2889caa9 CW |
3537 | |
3538 | /* | |
3539 | * Now that we have begun execution of the batchbuffer, we ignore | |
3540 | * any new error after this point. Also given that we have already | |
3541 | * updated the associated relocations, we try to write out the current | |
3542 | * object locations irrespective of any error. | |
3543 | */ | |
3544 | if (args->flags & __EXEC_HAS_RELOC) { | |
d593d992 | 3545 | struct drm_i915_gem_exec_object2 __user *user_exec_list = |
2889caa9 CW |
3546 | u64_to_user_ptr(args->buffers_ptr); |
3547 | unsigned int i; | |
9aab8bff | 3548 | |
2889caa9 | 3549 | /* Copy the new buffer offsets back to the user's exec list. */ |
594cc251 LT |
3550 | /* |
3551 | * Note: count * sizeof(*user_exec_list) does not overflow, | |
3552 | * because we checked 'count' in check_buffer_count(). | |
3553 | * | |
3554 | * And this range already got effectively checked earlier | |
3555 | * when we did the "copy_from_user()" above. | |
3556 | */ | |
b44f6873 CL |
3557 | if (!user_write_access_begin(user_exec_list, |
3558 | count * sizeof(*user_exec_list))) | |
8f4faed0 | 3559 | goto end; |
594cc251 | 3560 | |
9aab8bff | 3561 | for (i = 0; i < args->buffer_count; i++) { |
2889caa9 CW |
3562 | if (!(exec2_list[i].offset & UPDATE)) |
3563 | continue; | |
3564 | ||
934acce3 | 3565 | exec2_list[i].offset = |
2889caa9 CW |
3566 | gen8_canonical_addr(exec2_list[i].offset & PIN_OFFSET_MASK); |
3567 | unsafe_put_user(exec2_list[i].offset, | |
3568 | &user_exec_list[i].offset, | |
3569 | end_user); | |
54cf91dc | 3570 | } |
2889caa9 | 3571 | end_user: |
b44f6873 | 3572 | user_write_access_end(); |
8f4faed0 | 3573 | end:; |
54cf91dc CW |
3574 | } |
3575 | ||
2889caa9 | 3576 | args->flags &= ~__I915_EXEC_UNKNOWN_FLAGS; |
2098105e | 3577 | kvfree(exec2_list); |
2889caa9 | 3578 | return err; |
54cf91dc | 3579 | } |