Commit | Line | Data |
---|---|---|
54cf91dc | 1 | /* |
10be98a7 | 2 | * SPDX-License-Identifier: MIT |
54cf91dc | 3 | * |
10be98a7 | 4 | * Copyright © 2008,2010 Intel Corporation |
54cf91dc CW |
5 | */ |
6 | ||
52791eee | 7 | #include <linux/dma-resv.h> |
e9b67ec2 JN |
8 | #include <linux/highmem.h> |
9 | #include <linux/intel-iommu.h> | |
fec0445c | 10 | #include <linux/sync_file.h> |
ad778f89 CW |
11 | #include <linux/uaccess.h> |
12 | ||
cf6e7bac | 13 | #include <drm/drm_syncobj.h> |
ad778f89 | 14 | |
df0566a6 JN |
15 | #include "display/intel_frontbuffer.h" |
16 | ||
afa13085 | 17 | #include "gem/i915_gem_ioctls.h" |
10be98a7 | 18 | #include "gt/intel_context.h" |
45233ab2 | 19 | #include "gt/intel_gpu_commands.h" |
baea429d | 20 | #include "gt/intel_gt.h" |
16e87459 | 21 | #include "gt/intel_gt_buffer_pool.h" |
8f2a1057 | 22 | #include "gt/intel_gt_pm.h" |
2871ea85 | 23 | #include "gt/intel_ring.h" |
8f2a1057 | 24 | |
d3ac8d42 DCS |
25 | #include "pxp/intel_pxp.h" |
26 | ||
23d639d7 | 27 | #include "i915_cmd_parser.h" |
6da4a2c4 | 28 | #include "i915_drv.h" |
5472b3f2 | 29 | #include "i915_file_private.h" |
57822dc6 | 30 | #include "i915_gem_clflush.h" |
10be98a7 | 31 | #include "i915_gem_context.h" |
2ef97818 | 32 | #include "i915_gem_evict.h" |
6da4a2c4 | 33 | #include "i915_gem_ioctls.h" |
54cf91dc | 34 | #include "i915_trace.h" |
cda9edd0 | 35 | #include "i915_user_extensions.h" |
54cf91dc | 36 | |
7d6236bb CW |
37 | struct eb_vma { |
38 | struct i915_vma *vma; | |
39 | unsigned int flags; | |
40 | ||
41 | /** This vma's place in the execbuf reservation list */ | |
42 | struct drm_i915_gem_exec_object2 *exec; | |
43 | struct list_head bind_link; | |
44 | struct list_head reloc_link; | |
45 | ||
46 | struct hlist_node node; | |
47 | u32 handle; | |
48 | }; | |
49 | ||
ad5d95e4 DA |
50 | enum { |
51 | FORCE_CPU_RELOC = 1, | |
52 | FORCE_GTT_RELOC, | |
53 | FORCE_GPU_RELOC, | |
54 | #define DBG_FORCE_RELOC 0 /* choose one of the above! */ | |
55 | }; | |
56 | ||
bfaae47d ML |
57 | /* __EXEC_OBJECT_NO_RESERVE is BIT(31), defined in i915_vma.h */ |
58 | #define __EXEC_OBJECT_HAS_PIN BIT(30) | |
59 | #define __EXEC_OBJECT_HAS_FENCE BIT(29) | |
ed29c269 ML |
60 | #define __EXEC_OBJECT_USERPTR_INIT BIT(28) |
61 | #define __EXEC_OBJECT_NEEDS_MAP BIT(27) | |
62 | #define __EXEC_OBJECT_NEEDS_BIAS BIT(26) | |
63 | #define __EXEC_OBJECT_INTERNAL_FLAGS (~0u << 26) /* all of the above + */ | |
8ae275c2 | 64 | #define __EXEC_OBJECT_RESERVED (__EXEC_OBJECT_HAS_PIN | __EXEC_OBJECT_HAS_FENCE) |
2889caa9 CW |
65 | |
66 | #define __EXEC_HAS_RELOC BIT(31) | |
2bf541ff | 67 | #define __EXEC_ENGINE_PINNED BIT(30) |
ed29c269 ML |
68 | #define __EXEC_USERPTR_USED BIT(29) |
69 | #define __EXEC_INTERNAL_FLAGS (~0u << 29) | |
2889caa9 | 70 | #define UPDATE PIN_OFFSET_FIXED |
d23db88c CW |
71 | |
72 | #define BATCH_OFFSET_BIAS (256*1024) | |
a415d355 | 73 | |
650bc635 | 74 | #define __I915_EXEC_ILLEGAL_FLAGS \ |
08e3e21a LDM |
75 | (__I915_EXEC_UNKNOWN_FLAGS | \ |
76 | I915_EXEC_CONSTANTS_MASK | \ | |
77 | I915_EXEC_RESOURCE_STREAMER) | |
5b043f4e | 78 | |
d20ac620 CW |
79 | /* Catch emission of unexpected errors for CI! */ |
80 | #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) | |
81 | #undef EINVAL | |
82 | #define EINVAL ({ \ | |
83 | DRM_DEBUG_DRIVER("EINVAL at %s:%d\n", __func__, __LINE__); \ | |
84 | 22; \ | |
85 | }) | |
86 | #endif | |
87 | ||
2889caa9 CW |
88 | /** |
89 | * DOC: User command execution | |
90 | * | |
91 | * Userspace submits commands to be executed on the GPU as an instruction | |
92 | * stream within a GEM object we call a batchbuffer. This instructions may | |
93 | * refer to other GEM objects containing auxiliary state such as kernels, | |
94 | * samplers, render targets and even secondary batchbuffers. Userspace does | |
95 | * not know where in the GPU memory these objects reside and so before the | |
96 | * batchbuffer is passed to the GPU for execution, those addresses in the | |
97 | * batchbuffer and auxiliary objects are updated. This is known as relocation, | |
98 | * or patching. To try and avoid having to relocate each object on the next | |
99 | * execution, userspace is told the location of those objects in this pass, | |
100 | * but this remains just a hint as the kernel may choose a new location for | |
101 | * any object in the future. | |
102 | * | |
99d7e4ee KR |
103 | * At the level of talking to the hardware, submitting a batchbuffer for the |
104 | * GPU to execute is to add content to a buffer from which the HW | |
105 | * command streamer is reading. | |
106 | * | |
107 | * 1. Add a command to load the HW context. For Logical Ring Contexts, i.e. | |
108 | * Execlists, this command is not placed on the same buffer as the | |
109 | * remaining items. | |
110 | * | |
111 | * 2. Add a command to invalidate caches to the buffer. | |
112 | * | |
113 | * 3. Add a batchbuffer start command to the buffer; the start command is | |
114 | * essentially a token together with the GPU address of the batchbuffer | |
115 | * to be executed. | |
116 | * | |
117 | * 4. Add a pipeline flush to the buffer. | |
118 | * | |
119 | * 5. Add a memory write command to the buffer to record when the GPU | |
120 | * is done executing the batchbuffer. The memory write writes the | |
121 | * global sequence number of the request, ``i915_request::global_seqno``; | |
122 | * the i915 driver uses the current value in the register to determine | |
123 | * if the GPU has completed the batchbuffer. | |
124 | * | |
125 | * 6. Add a user interrupt command to the buffer. This command instructs | |
126 | * the GPU to issue an interrupt when the command, pipeline flush and | |
127 | * memory write are completed. | |
128 | * | |
129 | * 7. Inform the hardware of the additional commands added to the buffer | |
130 | * (by updating the tail pointer). | |
131 | * | |
2889caa9 CW |
132 | * Processing an execbuf ioctl is conceptually split up into a few phases. |
133 | * | |
134 | * 1. Validation - Ensure all the pointers, handles and flags are valid. | |
135 | * 2. Reservation - Assign GPU address space for every object | |
136 | * 3. Relocation - Update any addresses to point to the final locations | |
137 | * 4. Serialisation - Order the request with respect to its dependencies | |
138 | * 5. Construction - Construct a request to execute the batchbuffer | |
139 | * 6. Submission (at some point in the future execution) | |
140 | * | |
141 | * Reserving resources for the execbuf is the most complicated phase. We | |
142 | * neither want to have to migrate the object in the address space, nor do | |
143 | * we want to have to update any relocations pointing to this object. Ideally, | |
144 | * we want to leave the object where it is and for all the existing relocations | |
145 | * to match. If the object is given a new address, or if userspace thinks the | |
146 | * object is elsewhere, we have to parse all the relocation entries and update | |
147 | * the addresses. Userspace can set the I915_EXEC_NORELOC flag to hint that | |
148 | * all the target addresses in all of its objects match the value in the | |
149 | * relocation entries and that they all match the presumed offsets given by the | |
150 | * list of execbuffer objects. Using this knowledge, we know that if we haven't | |
151 | * moved any buffers, all the relocation entries are valid and we can skip | |
152 | * the update. (If userspace is wrong, the likely outcome is an impromptu GPU | |
153 | * hang.) The requirement for using I915_EXEC_NO_RELOC are: | |
154 | * | |
155 | * The addresses written in the objects must match the corresponding | |
156 | * reloc.presumed_offset which in turn must match the corresponding | |
157 | * execobject.offset. | |
158 | * | |
159 | * Any render targets written to in the batch must be flagged with | |
160 | * EXEC_OBJECT_WRITE. | |
161 | * | |
162 | * To avoid stalling, execobject.offset should match the current | |
163 | * address of that object within the active context. | |
164 | * | |
165 | * The reservation is done is multiple phases. First we try and keep any | |
166 | * object already bound in its current location - so as long as meets the | |
167 | * constraints imposed by the new execbuffer. Any object left unbound after the | |
168 | * first pass is then fitted into any available idle space. If an object does | |
169 | * not fit, all objects are removed from the reservation and the process rerun | |
170 | * after sorting the objects into a priority order (more difficult to fit | |
171 | * objects are tried first). Failing that, the entire VM is cleared and we try | |
172 | * to fit the execbuf once last time before concluding that it simply will not | |
173 | * fit. | |
174 | * | |
175 | * A small complication to all of this is that we allow userspace not only to | |
176 | * specify an alignment and a size for the object in the address space, but | |
177 | * we also allow userspace to specify the exact offset. This objects are | |
178 | * simpler to place (the location is known a priori) all we have to do is make | |
179 | * sure the space is available. | |
180 | * | |
181 | * Once all the objects are in place, patching up the buried pointers to point | |
182 | * to the final locations is a fairly simple job of walking over the relocation | |
183 | * entry arrays, looking up the right address and rewriting the value into | |
184 | * the object. Simple! ... The relocation entries are stored in user memory | |
185 | * and so to access them we have to copy them into a local buffer. That copy | |
186 | * has to avoid taking any pagefaults as they may lead back to a GEM object | |
187 | * requiring the struct_mutex (i.e. recursive deadlock). So once again we split | |
188 | * the relocation into multiple passes. First we try to do everything within an | |
189 | * atomic context (avoid the pagefaults) which requires that we never wait. If | |
190 | * we detect that we may wait, or if we need to fault, then we have to fallback | |
191 | * to a slower path. The slowpath has to drop the mutex. (Can you hear alarm | |
192 | * bells yet?) Dropping the mutex means that we lose all the state we have | |
193 | * built up so far for the execbuf and we must reset any global data. However, | |
194 | * we do leave the objects pinned in their final locations - which is a | |
195 | * potential issue for concurrent execbufs. Once we have left the mutex, we can | |
196 | * allocate and copy all the relocation entries into a large array at our | |
197 | * leisure, reacquire the mutex, reclaim all the objects and other state and | |
198 | * then proceed to update any incorrect addresses with the objects. | |
199 | * | |
200 | * As we process the relocation entries, we maintain a record of whether the | |
201 | * object is being written to. Using NORELOC, we expect userspace to provide | |
202 | * this information instead. We also check whether we can skip the relocation | |
203 | * by comparing the expected value inside the relocation entry with the target's | |
204 | * final address. If they differ, we have to map the current object and rewrite | |
205 | * the 4 or 8 byte pointer within. | |
206 | * | |
207 | * Serialising an execbuf is quite simple according to the rules of the GEM | |
208 | * ABI. Execution within each context is ordered by the order of submission. | |
209 | * Writes to any GEM object are in order of submission and are exclusive. Reads | |
210 | * from a GEM object are unordered with respect to other reads, but ordered by | |
211 | * writes. A write submitted after a read cannot occur before the read, and | |
212 | * similarly any read submitted after a write cannot occur before the write. | |
213 | * Writes are ordered between engines such that only one write occurs at any | |
214 | * time (completing any reads beforehand) - using semaphores where available | |
215 | * and CPU serialisation otherwise. Other GEM access obey the same rules, any | |
216 | * write (either via mmaps using set-domain, or via pwrite) must flush all GPU | |
217 | * reads before starting, and any read (either using set-domain or pread) must | |
218 | * flush all GPU writes before starting. (Note we only employ a barrier before, | |
219 | * we currently rely on userspace not concurrently starting a new execution | |
220 | * whilst reading or writing to an object. This may be an advantage or not | |
221 | * depending on how much you trust userspace not to shoot themselves in the | |
222 | * foot.) Serialisation may just result in the request being inserted into | |
223 | * a DAG awaiting its turn, but most simple is to wait on the CPU until | |
224 | * all dependencies are resolved. | |
225 | * | |
226 | * After all of that, is just a matter of closing the request and handing it to | |
227 | * the hardware (well, leaving it in a queue to be executed). However, we also | |
228 | * offer the ability for batchbuffers to be run with elevated privileges so | |
229 | * that they access otherwise hidden registers. (Used to adjust L3 cache etc.) | |
230 | * Before any batch is given extra privileges we first must check that it | |
231 | * contains no nefarious instructions, we check that each instruction is from | |
232 | * our whitelist and all registers are also from an allowed list. We first | |
233 | * copy the user's batchbuffer to a shadow (so that the user doesn't have | |
234 | * access to it, either by the CPU or GPU as we scan it) and then parse each | |
235 | * instruction. If everything is ok, we set a flag telling the hardware to run | |
236 | * the batchbuffer in trusted mode, otherwise the ioctl is rejected. | |
237 | */ | |
238 | ||
13149e8b LL |
239 | struct eb_fence { |
240 | struct drm_syncobj *syncobj; /* Use with ptr_mask_bits() */ | |
241 | struct dma_fence *dma_fence; | |
242 | u64 value; | |
243 | struct dma_fence_chain *chain_fence; | |
244 | }; | |
245 | ||
650bc635 | 246 | struct i915_execbuffer { |
2889caa9 CW |
247 | struct drm_i915_private *i915; /** i915 backpointer */ |
248 | struct drm_file *file; /** per-file lookup tables and limits */ | |
249 | struct drm_i915_gem_execbuffer2 *args; /** ioctl parameters */ | |
250 | struct drm_i915_gem_exec_object2 *exec; /** ioctl execobj[] */ | |
7d6236bb | 251 | struct eb_vma *vma; |
2889caa9 | 252 | |
544460c3 | 253 | struct intel_gt *gt; /* gt for the execbuf */ |
8f2a1057 CW |
254 | struct intel_context *context; /* logical state for the request */ |
255 | struct i915_gem_context *gem_context; /** caller's context */ | |
2889caa9 | 256 | |
544460c3 MB |
257 | /** our requests to build */ |
258 | struct i915_request *requests[MAX_ENGINE_INSTANCE + 1]; | |
259 | /** identity of the batch obj/vma */ | |
260 | struct eb_vma *batches[MAX_ENGINE_INSTANCE + 1]; | |
32d94048 | 261 | struct i915_vma *trampoline; /** trampoline used for chaining */ |
2889caa9 | 262 | |
544460c3 MB |
263 | /** used for excl fence in dma_resv objects when > 1 BB submitted */ |
264 | struct dma_fence *composite_fence; | |
265 | ||
2889caa9 CW |
266 | /** actual size of execobj[] as we may extend it for the cmdparser */ |
267 | unsigned int buffer_count; | |
268 | ||
544460c3 MB |
269 | /* number of batches in execbuf IOCTL */ |
270 | unsigned int num_batches; | |
271 | ||
2889caa9 CW |
272 | /** list of vma not yet bound during reservation phase */ |
273 | struct list_head unbound; | |
274 | ||
275 | /** list of vma that have execobj.relocation_count */ | |
276 | struct list_head relocs; | |
277 | ||
c43ce123 ML |
278 | struct i915_gem_ww_ctx ww; |
279 | ||
2889caa9 CW |
280 | /** |
281 | * Track the most recently used object for relocations, as we | |
282 | * frequently have to perform multiple relocations within the same | |
283 | * obj/page | |
284 | */ | |
650bc635 | 285 | struct reloc_cache { |
2889caa9 | 286 | struct drm_mm_node node; /** temporary GTT binding */ |
ad5d95e4 DA |
287 | unsigned long vaddr; /** Current kmap address */ |
288 | unsigned long page; /** Currently mapped page index */ | |
8802190f | 289 | unsigned int graphics_ver; /** Cached value of GRAPHICS_VER */ |
650bc635 | 290 | bool use_64bit_reloc : 1; |
2889caa9 CW |
291 | bool has_llc : 1; |
292 | bool has_fence : 1; | |
293 | bool needs_unfenced : 1; | |
650bc635 | 294 | } reloc_cache; |
2889caa9 CW |
295 | |
296 | u64 invalid_flags; /** Set of execobj.flags that are invalid */ | |
2889caa9 | 297 | |
544460c3 MB |
298 | /** Length of batch within object */ |
299 | u64 batch_len[MAX_ENGINE_INSTANCE + 1]; | |
2889caa9 | 300 | u32 batch_start_offset; /** Location within object of batch */ |
2889caa9 | 301 | u32 batch_flags; /** Flags composed for emit_bb_start() */ |
c43ce123 | 302 | struct intel_gt_buffer_pool_node *batch_pool; /** pool node for batch buffer */ |
2889caa9 CW |
303 | |
304 | /** | |
305 | * Indicate either the size of the hastable used to resolve | |
306 | * relocation handles, or if negative that we are using a direct | |
307 | * index into the execobj[]. | |
308 | */ | |
309 | int lut_size; | |
310 | struct hlist_head *buckets; /** ht for relocation handles */ | |
cda9edd0 | 311 | |
13149e8b LL |
312 | struct eb_fence *fences; |
313 | unsigned long num_fences; | |
ff20afc4 TH |
314 | #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR) |
315 | struct i915_capture_list *capture_lists[MAX_ENGINE_INSTANCE + 1]; | |
316 | #endif | |
67731b87 CW |
317 | }; |
318 | ||
8e4ba491 | 319 | static int eb_parse(struct i915_execbuffer *eb); |
544460c3 | 320 | static int eb_pin_engine(struct i915_execbuffer *eb, bool throttle); |
2bf541ff | 321 | static void eb_unpin_engine(struct i915_execbuffer *eb); |
ff20afc4 | 322 | static void eb_capture_release(struct i915_execbuffer *eb); |
8e4ba491 | 323 | |
3dbf26ed CW |
324 | static inline bool eb_use_cmdparser(const struct i915_execbuffer *eb) |
325 | { | |
544460c3 MB |
326 | return intel_engine_requires_cmd_parser(eb->context->engine) || |
327 | (intel_engine_using_cmd_parser(eb->context->engine) && | |
435e8fc0 | 328 | eb->args->batch_len); |
3dbf26ed CW |
329 | } |
330 | ||
650bc635 | 331 | static int eb_create(struct i915_execbuffer *eb) |
67731b87 | 332 | { |
2889caa9 CW |
333 | if (!(eb->args->flags & I915_EXEC_HANDLE_LUT)) { |
334 | unsigned int size = 1 + ilog2(eb->buffer_count); | |
4ff4b44c | 335 | |
2889caa9 CW |
336 | /* |
337 | * Without a 1:1 association between relocation handles and | |
338 | * the execobject[] index, we instead create a hashtable. | |
339 | * We size it dynamically based on available memory, starting | |
340 | * first with 1:1 assocative hash and scaling back until | |
341 | * the allocation succeeds. | |
342 | * | |
343 | * Later on we use a positive lut_size to indicate we are | |
344 | * using this hashtable, and a negative value to indicate a | |
345 | * direct lookup. | |
346 | */ | |
4ff4b44c | 347 | do { |
0d95c883 | 348 | gfp_t flags; |
4d470f73 CW |
349 | |
350 | /* While we can still reduce the allocation size, don't | |
351 | * raise a warning and allow the allocation to fail. | |
352 | * On the last pass though, we want to try as hard | |
353 | * as possible to perform the allocation and warn | |
354 | * if it fails. | |
355 | */ | |
0ee931c4 | 356 | flags = GFP_KERNEL; |
4d470f73 CW |
357 | if (size > 1) |
358 | flags |= __GFP_NORETRY | __GFP_NOWARN; | |
359 | ||
4ff4b44c | 360 | eb->buckets = kzalloc(sizeof(struct hlist_head) << size, |
4d470f73 | 361 | flags); |
4ff4b44c CW |
362 | if (eb->buckets) |
363 | break; | |
364 | } while (--size); | |
365 | ||
8ae275c2 | 366 | if (unlikely(!size)) |
4d470f73 | 367 | return -ENOMEM; |
eef90ccb | 368 | |
2889caa9 | 369 | eb->lut_size = size; |
650bc635 | 370 | } else { |
2889caa9 | 371 | eb->lut_size = -eb->buffer_count; |
650bc635 | 372 | } |
eef90ccb | 373 | |
650bc635 | 374 | return 0; |
67731b87 CW |
375 | } |
376 | ||
2889caa9 CW |
377 | static bool |
378 | eb_vma_misplaced(const struct drm_i915_gem_exec_object2 *entry, | |
c7c6e46f CW |
379 | const struct i915_vma *vma, |
380 | unsigned int flags) | |
2889caa9 | 381 | { |
2889caa9 CW |
382 | if (vma->node.size < entry->pad_to_size) |
383 | return true; | |
384 | ||
385 | if (entry->alignment && !IS_ALIGNED(vma->node.start, entry->alignment)) | |
386 | return true; | |
387 | ||
c7c6e46f | 388 | if (flags & EXEC_OBJECT_PINNED && |
2889caa9 CW |
389 | vma->node.start != entry->offset) |
390 | return true; | |
391 | ||
c7c6e46f | 392 | if (flags & __EXEC_OBJECT_NEEDS_BIAS && |
2889caa9 CW |
393 | vma->node.start < BATCH_OFFSET_BIAS) |
394 | return true; | |
395 | ||
c7c6e46f | 396 | if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) && |
5f22cc0b | 397 | (vma->node.start + vma->node.size + 4095) >> 32) |
2889caa9 CW |
398 | return true; |
399 | ||
1d033beb CW |
400 | if (flags & __EXEC_OBJECT_NEEDS_MAP && |
401 | !i915_vma_is_map_and_fenceable(vma)) | |
402 | return true; | |
403 | ||
2889caa9 CW |
404 | return false; |
405 | } | |
406 | ||
8a338f4b CW |
407 | static u64 eb_pin_flags(const struct drm_i915_gem_exec_object2 *entry, |
408 | unsigned int exec_flags) | |
409 | { | |
410 | u64 pin_flags = 0; | |
411 | ||
412 | if (exec_flags & EXEC_OBJECT_NEEDS_GTT) | |
413 | pin_flags |= PIN_GLOBAL; | |
414 | ||
415 | /* | |
416 | * Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset, | |
417 | * limit address to the first 4GBs for unflagged objects. | |
418 | */ | |
419 | if (!(exec_flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS)) | |
420 | pin_flags |= PIN_ZONE_4G; | |
421 | ||
422 | if (exec_flags & __EXEC_OBJECT_NEEDS_MAP) | |
423 | pin_flags |= PIN_MAPPABLE; | |
424 | ||
425 | if (exec_flags & EXEC_OBJECT_PINNED) | |
426 | pin_flags |= entry->offset | PIN_OFFSET_FIXED; | |
427 | else if (exec_flags & __EXEC_OBJECT_NEEDS_BIAS) | |
428 | pin_flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS; | |
429 | ||
430 | return pin_flags; | |
431 | } | |
432 | ||
237647f4 | 433 | static inline int |
2889caa9 | 434 | eb_pin_vma(struct i915_execbuffer *eb, |
c7c6e46f | 435 | const struct drm_i915_gem_exec_object2 *entry, |
7d6236bb | 436 | struct eb_vma *ev) |
2889caa9 | 437 | { |
7d6236bb | 438 | struct i915_vma *vma = ev->vma; |
c7c6e46f | 439 | u64 pin_flags; |
237647f4 | 440 | int err; |
2889caa9 | 441 | |
616d9cee | 442 | if (vma->node.size) |
c7c6e46f | 443 | pin_flags = vma->node.start; |
616d9cee | 444 | else |
c7c6e46f | 445 | pin_flags = entry->offset & PIN_OFFSET_MASK; |
616d9cee | 446 | |
b5cfe6f7 | 447 | pin_flags |= PIN_USER | PIN_NOEVICT | PIN_OFFSET_FIXED | PIN_VALIDATE; |
7d6236bb | 448 | if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_GTT)) |
c7c6e46f | 449 | pin_flags |= PIN_GLOBAL; |
616d9cee | 450 | |
8a338f4b | 451 | /* Attempt to reuse the current location if available */ |
237647f4 ML |
452 | err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, pin_flags); |
453 | if (err == -EDEADLK) | |
454 | return err; | |
455 | ||
456 | if (unlikely(err)) { | |
8a338f4b | 457 | if (entry->flags & EXEC_OBJECT_PINNED) |
237647f4 | 458 | return err; |
8a338f4b CW |
459 | |
460 | /* Failing that pick any _free_ space if suitable */ | |
237647f4 | 461 | err = i915_vma_pin_ww(vma, &eb->ww, |
47b08693 ML |
462 | entry->pad_to_size, |
463 | entry->alignment, | |
464 | eb_pin_flags(entry, ev->flags) | | |
b5cfe6f7 | 465 | PIN_USER | PIN_NOEVICT | PIN_VALIDATE); |
237647f4 ML |
466 | if (unlikely(err)) |
467 | return err; | |
8a338f4b | 468 | } |
2889caa9 | 469 | |
7d6236bb | 470 | if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) { |
237647f4 | 471 | err = i915_vma_pin_fence(vma); |
b5cfe6f7 | 472 | if (unlikely(err)) |
237647f4 | 473 | return err; |
2889caa9 | 474 | |
3bd40735 | 475 | if (vma->fence) |
7d6236bb | 476 | ev->flags |= __EXEC_OBJECT_HAS_FENCE; |
2889caa9 CW |
477 | } |
478 | ||
7d6236bb | 479 | ev->flags |= __EXEC_OBJECT_HAS_PIN; |
237647f4 ML |
480 | if (eb_vma_misplaced(entry, vma, ev->flags)) |
481 | return -EBADSLT; | |
482 | ||
483 | return 0; | |
2889caa9 CW |
484 | } |
485 | ||
8ae275c2 ML |
486 | static inline void |
487 | eb_unreserve_vma(struct eb_vma *ev) | |
488 | { | |
c43ce123 ML |
489 | if (unlikely(ev->flags & __EXEC_OBJECT_HAS_FENCE)) |
490 | __i915_vma_unpin_fence(ev->vma); | |
491 | ||
8ae275c2 ML |
492 | ev->flags &= ~__EXEC_OBJECT_RESERVED; |
493 | } | |
494 | ||
2889caa9 CW |
495 | static int |
496 | eb_validate_vma(struct i915_execbuffer *eb, | |
497 | struct drm_i915_gem_exec_object2 *entry, | |
498 | struct i915_vma *vma) | |
67731b87 | 499 | { |
2eb8e1a6 JE |
500 | /* Relocations are disallowed for all platforms after TGL-LP. This |
501 | * also covers all platforms with local memory. | |
502 | */ | |
503 | if (entry->relocation_count && | |
40e1956e | 504 | GRAPHICS_VER(eb->i915) >= 12 && !IS_TIGERLAKE(eb->i915)) |
2eb8e1a6 JE |
505 | return -EINVAL; |
506 | ||
2889caa9 CW |
507 | if (unlikely(entry->flags & eb->invalid_flags)) |
508 | return -EINVAL; | |
d55495b4 | 509 | |
2920516b MA |
510 | if (unlikely(entry->alignment && |
511 | !is_power_of_2_u64(entry->alignment))) | |
2889caa9 CW |
512 | return -EINVAL; |
513 | ||
514 | /* | |
515 | * Offset can be used as input (EXEC_OBJECT_PINNED), reject | |
516 | * any non-page-aligned or non-canonical addresses. | |
517 | */ | |
518 | if (unlikely(entry->flags & EXEC_OBJECT_PINNED && | |
6fc4e48f | 519 | entry->offset != gen8_canonical_addr(entry->offset & I915_GTT_PAGE_MASK))) |
2889caa9 CW |
520 | return -EINVAL; |
521 | ||
522 | /* pad_to_size was once a reserved field, so sanitize it */ | |
523 | if (entry->flags & EXEC_OBJECT_PAD_TO_SIZE) { | |
524 | if (unlikely(offset_in_page(entry->pad_to_size))) | |
525 | return -EINVAL; | |
526 | } else { | |
527 | entry->pad_to_size = 0; | |
d55495b4 | 528 | } |
2889caa9 CW |
529 | /* |
530 | * From drm_mm perspective address space is continuous, | |
531 | * so from this point we're always using non-canonical | |
532 | * form internally. | |
533 | */ | |
534 | entry->offset = gen8_noncanonical_addr(entry->offset); | |
535 | ||
c7c6e46f CW |
536 | if (!eb->reloc_cache.has_fence) { |
537 | entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE; | |
538 | } else { | |
539 | if ((entry->flags & EXEC_OBJECT_NEEDS_FENCE || | |
540 | eb->reloc_cache.needs_unfenced) && | |
541 | i915_gem_object_is_tiled(vma->obj)) | |
542 | entry->flags |= EXEC_OBJECT_NEEDS_GTT | __EXEC_OBJECT_NEEDS_MAP; | |
543 | } | |
544 | ||
2889caa9 | 545 | return 0; |
67731b87 CW |
546 | } |
547 | ||
544460c3 MB |
548 | static inline bool |
549 | is_batch_buffer(struct i915_execbuffer *eb, unsigned int buffer_idx) | |
550 | { | |
551 | return eb->args->flags & I915_EXEC_BATCH_FIRST ? | |
552 | buffer_idx < eb->num_batches : | |
553 | buffer_idx >= eb->args->buffer_count - eb->num_batches; | |
554 | } | |
555 | ||
556 | static int | |
746c8f14 | 557 | eb_add_vma(struct i915_execbuffer *eb, |
544460c3 MB |
558 | unsigned int *current_batch, |
559 | unsigned int i, | |
746c8f14 | 560 | struct i915_vma *vma) |
59bfa124 | 561 | { |
544460c3 | 562 | struct drm_i915_private *i915 = eb->i915; |
c7c6e46f | 563 | struct drm_i915_gem_exec_object2 *entry = &eb->exec[i]; |
7d6236bb | 564 | struct eb_vma *ev = &eb->vma[i]; |
2889caa9 | 565 | |
93159e12 | 566 | ev->vma = vma; |
7d6236bb CW |
567 | ev->exec = entry; |
568 | ev->flags = entry->flags; | |
569 | ||
4d470f73 | 570 | if (eb->lut_size > 0) { |
7d6236bb CW |
571 | ev->handle = entry->handle; |
572 | hlist_add_head(&ev->node, | |
2889caa9 CW |
573 | &eb->buckets[hash_32(entry->handle, |
574 | eb->lut_size)]); | |
4ff4b44c | 575 | } |
59bfa124 | 576 | |
2889caa9 | 577 | if (entry->relocation_count) |
7d6236bb | 578 | list_add_tail(&ev->reloc_link, &eb->relocs); |
2889caa9 | 579 | |
746c8f14 CW |
580 | /* |
581 | * SNA is doing fancy tricks with compressing batch buffers, which leads | |
582 | * to negative relocation deltas. Usually that works out ok since the | |
583 | * relocate address is still positive, except when the batch is placed | |
584 | * very low in the GTT. Ensure this doesn't happen. | |
585 | * | |
586 | * Note that actual hangs have only been observed on gen7, but for | |
587 | * paranoia do it everywhere. | |
588 | */ | |
544460c3 | 589 | if (is_batch_buffer(eb, i)) { |
827db9d8 | 590 | if (entry->relocation_count && |
7d6236bb CW |
591 | !(ev->flags & EXEC_OBJECT_PINNED)) |
592 | ev->flags |= __EXEC_OBJECT_NEEDS_BIAS; | |
746c8f14 | 593 | if (eb->reloc_cache.has_fence) |
7d6236bb | 594 | ev->flags |= EXEC_OBJECT_NEEDS_FENCE; |
746c8f14 | 595 | |
544460c3 MB |
596 | eb->batches[*current_batch] = ev; |
597 | ||
598 | if (unlikely(ev->flags & EXEC_OBJECT_WRITE)) { | |
599 | drm_dbg(&i915->drm, | |
600 | "Attempting to use self-modifying batch buffer\n"); | |
601 | return -EINVAL; | |
602 | } | |
603 | ||
604 | if (range_overflows_t(u64, | |
605 | eb->batch_start_offset, | |
606 | eb->args->batch_len, | |
607 | ev->vma->size)) { | |
608 | drm_dbg(&i915->drm, "Attempting to use out-of-bounds batch\n"); | |
609 | return -EINVAL; | |
610 | } | |
611 | ||
612 | if (eb->args->batch_len == 0) | |
613 | eb->batch_len[*current_batch] = ev->vma->size - | |
614 | eb->batch_start_offset; | |
615 | else | |
616 | eb->batch_len[*current_batch] = eb->args->batch_len; | |
617 | if (unlikely(eb->batch_len[*current_batch] == 0)) { /* impossible! */ | |
618 | drm_dbg(&i915->drm, "Invalid batch length\n"); | |
619 | return -EINVAL; | |
620 | } | |
621 | ||
622 | ++*current_batch; | |
746c8f14 | 623 | } |
544460c3 MB |
624 | |
625 | return 0; | |
2889caa9 CW |
626 | } |
627 | ||
ad5d95e4 DA |
628 | static inline int use_cpu_reloc(const struct reloc_cache *cache, |
629 | const struct drm_i915_gem_object *obj) | |
630 | { | |
631 | if (!i915_gem_object_has_struct_page(obj)) | |
632 | return false; | |
633 | ||
634 | if (DBG_FORCE_RELOC == FORCE_CPU_RELOC) | |
635 | return true; | |
636 | ||
637 | if (DBG_FORCE_RELOC == FORCE_GTT_RELOC) | |
638 | return false; | |
639 | ||
640 | return (cache->has_llc || | |
641 | obj->cache_dirty || | |
642 | obj->cache_level != I915_CACHE_NONE); | |
643 | } | |
644 | ||
47b08693 | 645 | static int eb_reserve_vma(struct i915_execbuffer *eb, |
7d6236bb | 646 | struct eb_vma *ev, |
2920bb94 | 647 | u64 pin_flags) |
2889caa9 | 648 | { |
7d6236bb | 649 | struct drm_i915_gem_exec_object2 *entry = ev->exec; |
7d6236bb | 650 | struct i915_vma *vma = ev->vma; |
2889caa9 CW |
651 | int err; |
652 | ||
003d8b91 CW |
653 | if (drm_mm_node_allocated(&vma->node) && |
654 | eb_vma_misplaced(entry, vma, ev->flags)) { | |
655 | err = i915_vma_unbind(vma); | |
656 | if (err) | |
657 | return err; | |
658 | } | |
659 | ||
47b08693 | 660 | err = i915_vma_pin_ww(vma, &eb->ww, |
c7c6e46f | 661 | entry->pad_to_size, entry->alignment, |
8a338f4b | 662 | eb_pin_flags(entry, ev->flags) | pin_flags); |
2889caa9 CW |
663 | if (err) |
664 | return err; | |
665 | ||
666 | if (entry->offset != vma->node.start) { | |
667 | entry->offset = vma->node.start | UPDATE; | |
668 | eb->args->flags |= __EXEC_HAS_RELOC; | |
669 | } | |
670 | ||
8a338f4b | 671 | if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) { |
3bd40735 | 672 | err = i915_vma_pin_fence(vma); |
b5cfe6f7 | 673 | if (unlikely(err)) |
2889caa9 | 674 | return err; |
2889caa9 | 675 | |
3bd40735 | 676 | if (vma->fence) |
8a338f4b | 677 | ev->flags |= __EXEC_OBJECT_HAS_FENCE; |
2889caa9 CW |
678 | } |
679 | ||
8a338f4b | 680 | ev->flags |= __EXEC_OBJECT_HAS_PIN; |
7d6236bb | 681 | GEM_BUG_ON(eb_vma_misplaced(entry, vma, ev->flags)); |
1da7b54c | 682 | |
2889caa9 CW |
683 | return 0; |
684 | } | |
685 | ||
b5cfe6f7 | 686 | static bool eb_unbind(struct i915_execbuffer *eb, bool force) |
2889caa9 CW |
687 | { |
688 | const unsigned int count = eb->buffer_count; | |
b5cfe6f7 | 689 | unsigned int i; |
2889caa9 | 690 | struct list_head last; |
b5cfe6f7 ML |
691 | bool unpinned = false; |
692 | ||
693 | /* Resort *all* the objects into priority order */ | |
694 | INIT_LIST_HEAD(&eb->unbound); | |
695 | INIT_LIST_HEAD(&last); | |
696 | ||
697 | for (i = 0; i < count; i++) { | |
698 | struct eb_vma *ev = &eb->vma[i]; | |
699 | unsigned int flags = ev->flags; | |
700 | ||
701 | if (!force && flags & EXEC_OBJECT_PINNED && | |
702 | flags & __EXEC_OBJECT_HAS_PIN) | |
703 | continue; | |
704 | ||
705 | unpinned = true; | |
706 | eb_unreserve_vma(ev); | |
707 | ||
708 | if (flags & EXEC_OBJECT_PINNED) | |
709 | /* Pinned must have their slot */ | |
710 | list_add(&ev->bind_link, &eb->unbound); | |
711 | else if (flags & __EXEC_OBJECT_NEEDS_MAP) | |
712 | /* Map require the lowest 256MiB (aperture) */ | |
713 | list_add_tail(&ev->bind_link, &eb->unbound); | |
714 | else if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS)) | |
715 | /* Prioritise 4GiB region for restricted bo */ | |
716 | list_add(&ev->bind_link, &last); | |
717 | else | |
718 | list_add_tail(&ev->bind_link, &last); | |
719 | } | |
720 | ||
721 | list_splice_tail(&last, &eb->unbound); | |
722 | return unpinned; | |
723 | } | |
724 | ||
725 | static int eb_reserve(struct i915_execbuffer *eb) | |
726 | { | |
7d6236bb | 727 | struct eb_vma *ev; |
b5cfe6f7 | 728 | unsigned int pass; |
ef398881 | 729 | int err = 0; |
b5cfe6f7 | 730 | bool unpinned; |
2889caa9 CW |
731 | |
732 | /* | |
733 | * Attempt to pin all of the buffers into the GTT. | |
b5cfe6f7 | 734 | * This is done in 2 phases: |
2889caa9 | 735 | * |
b5cfe6f7 ML |
736 | * 1. Unbind all objects that do not match the GTT constraints for |
737 | * the execbuffer (fenceable, mappable, alignment etc). | |
738 | * 2. Bind new objects. | |
2889caa9 CW |
739 | * |
740 | * This avoid unnecessary unbinding of later objects in order to make | |
741 | * room for the earlier objects *unless* we need to defragment. | |
b5cfe6f7 ML |
742 | * |
743 | * Defragmenting is skipped if all objects are pinned at a fixed location. | |
2889caa9 | 744 | */ |
b5cfe6f7 ML |
745 | for (pass = 0; pass <= 2; pass++) { |
746 | int pin_flags = PIN_USER | PIN_VALIDATE; | |
2889caa9 | 747 | |
b5cfe6f7 ML |
748 | if (pass == 0) |
749 | pin_flags |= PIN_NONBLOCK; | |
2889caa9 | 750 | |
b5cfe6f7 ML |
751 | if (pass >= 1) |
752 | unpinned = eb_unbind(eb, pass == 2); | |
2889caa9 | 753 | |
b5cfe6f7 ML |
754 | if (pass == 2) { |
755 | err = mutex_lock_interruptible(&eb->context->vm->mutex); | |
756 | if (!err) { | |
757 | err = i915_gem_evict_vm(eb->context->vm, &eb->ww); | |
758 | mutex_unlock(&eb->context->vm->mutex); | |
759 | } | |
2889caa9 | 760 | if (err) |
c43ce123 | 761 | return err; |
b5cfe6f7 | 762 | } |
2889caa9 | 763 | |
b5cfe6f7 ML |
764 | list_for_each_entry(ev, &eb->unbound, bind_link) { |
765 | err = eb_reserve_vma(eb, ev, pin_flags); | |
766 | if (err) | |
767 | break; | |
2889caa9 | 768 | } |
2920bb94 | 769 | |
b5cfe6f7 ML |
770 | if (err != -ENOSPC) |
771 | break; | |
772 | } | |
773 | ||
774 | return err; | |
4ff4b44c | 775 | } |
59bfa124 | 776 | |
2889caa9 CW |
777 | static int eb_select_context(struct i915_execbuffer *eb) |
778 | { | |
779 | struct i915_gem_context *ctx; | |
780 | ||
781 | ctx = i915_gem_context_lookup(eb->file->driver_priv, eb->args->rsvd1); | |
046d1660 JE |
782 | if (unlikely(IS_ERR(ctx))) |
783 | return PTR_ERR(ctx); | |
2889caa9 | 784 | |
8f2a1057 | 785 | eb->gem_context = ctx; |
a82a9979 | 786 | if (i915_gem_context_has_full_ppgtt(ctx)) |
4f2c7337 | 787 | eb->invalid_flags |= EXEC_OBJECT_NEEDS_GTT; |
2889caa9 | 788 | |
2889caa9 CW |
789 | return 0; |
790 | } | |
791 | ||
93159e12 CW |
792 | static int __eb_add_lut(struct i915_execbuffer *eb, |
793 | u32 handle, struct i915_vma *vma) | |
3b96eff4 | 794 | { |
93159e12 CW |
795 | struct i915_gem_context *ctx = eb->gem_context; |
796 | struct i915_lut_handle *lut; | |
2889caa9 | 797 | int err; |
3b96eff4 | 798 | |
93159e12 CW |
799 | lut = i915_lut_handle_alloc(); |
800 | if (unlikely(!lut)) | |
801 | return -ENOMEM; | |
802 | ||
803 | i915_vma_get(vma); | |
804 | if (!atomic_fetch_inc(&vma->open_count)) | |
805 | i915_vma_reopen(vma); | |
806 | lut->handle = handle; | |
807 | lut->ctx = ctx; | |
808 | ||
809 | /* Check that the context hasn't been closed in the meantime */ | |
810 | err = -EINTR; | |
f7ce8639 | 811 | if (!mutex_lock_interruptible(&ctx->lut_mutex)) { |
e1068a9e | 812 | if (likely(!i915_gem_context_is_closed(ctx))) |
93159e12 | 813 | err = radix_tree_insert(&ctx->handles_vma, handle, vma); |
f7ce8639 CW |
814 | else |
815 | err = -ENOENT; | |
93159e12 CW |
816 | if (err == 0) { /* And nor has this handle */ |
817 | struct drm_i915_gem_object *obj = vma->obj; | |
818 | ||
096a42dd | 819 | spin_lock(&obj->lut_lock); |
93159e12 CW |
820 | if (idr_find(&eb->file->object_idr, handle) == obj) { |
821 | list_add(&lut->obj_link, &obj->lut_list); | |
822 | } else { | |
823 | radix_tree_delete(&ctx->handles_vma, handle); | |
824 | err = -ENOENT; | |
825 | } | |
096a42dd | 826 | spin_unlock(&obj->lut_lock); |
93159e12 | 827 | } |
f7ce8639 | 828 | mutex_unlock(&ctx->lut_mutex); |
93159e12 CW |
829 | } |
830 | if (unlikely(err)) | |
831 | goto err; | |
003d8b91 | 832 | |
93159e12 | 833 | return 0; |
d55495b4 | 834 | |
93159e12 | 835 | err: |
50689771 | 836 | i915_vma_close(vma); |
93159e12 CW |
837 | i915_vma_put(vma); |
838 | i915_lut_handle_free(lut); | |
839 | return err; | |
840 | } | |
746c8f14 | 841 | |
93159e12 CW |
842 | static struct i915_vma *eb_lookup_vma(struct i915_execbuffer *eb, u32 handle) |
843 | { | |
f7ce8639 CW |
844 | struct i915_address_space *vm = eb->context->vm; |
845 | ||
93159e12 CW |
846 | do { |
847 | struct drm_i915_gem_object *obj; | |
170fa29b | 848 | struct i915_vma *vma; |
93159e12 | 849 | int err; |
4ff4b44c | 850 | |
93159e12 CW |
851 | rcu_read_lock(); |
852 | vma = radix_tree_lookup(&eb->gem_context->handles_vma, handle); | |
f7ce8639 | 853 | if (likely(vma && vma->vm == vm)) |
93159e12 CW |
854 | vma = i915_vma_tryget(vma); |
855 | rcu_read_unlock(); | |
856 | if (likely(vma)) | |
857 | return vma; | |
4ff4b44c | 858 | |
170fa29b | 859 | obj = i915_gem_object_lookup(eb->file, handle); |
93159e12 CW |
860 | if (unlikely(!obj)) |
861 | return ERR_PTR(-ENOENT); | |
3b96eff4 | 862 | |
d3ac8d42 DCS |
863 | /* |
864 | * If the user has opted-in for protected-object tracking, make | |
865 | * sure the object encryption can be used. | |
866 | * We only need to do this when the object is first used with | |
867 | * this context, because the context itself will be banned when | |
868 | * the protected objects become invalid. | |
869 | */ | |
870 | if (i915_gem_context_uses_protected_content(eb->gem_context) && | |
871 | i915_gem_object_is_protected(obj)) { | |
ef6ba31d | 872 | err = intel_pxp_key_check(&vm->gt->pxp, obj, true); |
d3ac8d42 DCS |
873 | if (err) { |
874 | i915_gem_object_put(obj); | |
875 | return ERR_PTR(err); | |
876 | } | |
877 | } | |
878 | ||
f7ce8639 | 879 | vma = i915_vma_instance(obj, vm, NULL); |
772b5408 | 880 | if (IS_ERR(vma)) { |
93159e12 CW |
881 | i915_gem_object_put(obj); |
882 | return vma; | |
27173f1f BW |
883 | } |
884 | ||
93159e12 CW |
885 | err = __eb_add_lut(eb, handle, vma); |
886 | if (likely(!err)) | |
887 | return vma; | |
d1b48c1e | 888 | |
93159e12 CW |
889 | i915_gem_object_put(obj); |
890 | if (err != -EEXIST) | |
891 | return ERR_PTR(err); | |
892 | } while (1); | |
893 | } | |
4ff4b44c | 894 | |
93159e12 CW |
895 | static int eb_lookup_vmas(struct i915_execbuffer *eb) |
896 | { | |
544460c3 | 897 | unsigned int i, current_batch = 0; |
93159e12 | 898 | int err = 0; |
155ab883 | 899 | |
93159e12 | 900 | INIT_LIST_HEAD(&eb->relocs); |
93159e12 CW |
901 | |
902 | for (i = 0; i < eb->buffer_count; i++) { | |
903 | struct i915_vma *vma; | |
904 | ||
905 | vma = eb_lookup_vma(eb, eb->exec[i].handle); | |
906 | if (IS_ERR(vma)) { | |
907 | err = PTR_ERR(vma); | |
8e4ba491 | 908 | goto err; |
93159e12 | 909 | } |
d1b48c1e | 910 | |
003d8b91 | 911 | err = eb_validate_vma(eb, &eb->exec[i], vma); |
93159e12 CW |
912 | if (unlikely(err)) { |
913 | i915_vma_put(vma); | |
8e4ba491 | 914 | goto err; |
93159e12 | 915 | } |
dade2a61 | 916 | |
544460c3 MB |
917 | err = eb_add_vma(eb, ¤t_batch, i, vma); |
918 | if (err) | |
919 | return err; | |
ed29c269 ML |
920 | |
921 | if (i915_gem_object_is_userptr(vma->obj)) { | |
922 | err = i915_gem_object_userptr_submit_init(vma->obj); | |
923 | if (err) { | |
924 | if (i + 1 < eb->buffer_count) { | |
925 | /* | |
926 | * Execbuffer code expects last vma entry to be NULL, | |
927 | * since we already initialized this entry, | |
928 | * set the next value to NULL or we mess up | |
929 | * cleanup handling. | |
930 | */ | |
931 | eb->vma[i + 1].vma = NULL; | |
932 | } | |
933 | ||
934 | return err; | |
935 | } | |
936 | ||
937 | eb->vma[i].flags |= __EXEC_OBJECT_USERPTR_INIT; | |
938 | eb->args->flags |= __EXEC_USERPTR_USED; | |
939 | } | |
4ff4b44c CW |
940 | } |
941 | ||
8e4ba491 ML |
942 | return 0; |
943 | ||
944 | err: | |
7d6236bb | 945 | eb->vma[i].vma = NULL; |
2889caa9 | 946 | return err; |
3b96eff4 CW |
947 | } |
948 | ||
5cd57f67 | 949 | static int eb_lock_vmas(struct i915_execbuffer *eb) |
c43ce123 ML |
950 | { |
951 | unsigned int i; | |
952 | int err; | |
953 | ||
c43ce123 | 954 | for (i = 0; i < eb->buffer_count; i++) { |
c43ce123 ML |
955 | struct eb_vma *ev = &eb->vma[i]; |
956 | struct i915_vma *vma = ev->vma; | |
957 | ||
958 | err = i915_gem_object_lock(vma->obj, &eb->ww); | |
959 | if (err) | |
960 | return err; | |
5cd57f67 TH |
961 | } |
962 | ||
963 | return 0; | |
964 | } | |
965 | ||
966 | static int eb_validate_vmas(struct i915_execbuffer *eb) | |
967 | { | |
968 | unsigned int i; | |
969 | int err; | |
970 | ||
971 | INIT_LIST_HEAD(&eb->unbound); | |
972 | ||
973 | err = eb_lock_vmas(eb); | |
974 | if (err) | |
975 | return err; | |
976 | ||
977 | for (i = 0; i < eb->buffer_count; i++) { | |
978 | struct drm_i915_gem_exec_object2 *entry = &eb->exec[i]; | |
979 | struct eb_vma *ev = &eb->vma[i]; | |
980 | struct i915_vma *vma = ev->vma; | |
c43ce123 | 981 | |
237647f4 ML |
982 | err = eb_pin_vma(eb, entry, ev); |
983 | if (err == -EDEADLK) | |
984 | return err; | |
985 | ||
986 | if (!err) { | |
c43ce123 ML |
987 | if (entry->offset != vma->node.start) { |
988 | entry->offset = vma->node.start | UPDATE; | |
989 | eb->args->flags |= __EXEC_HAS_RELOC; | |
990 | } | |
991 | } else { | |
992 | eb_unreserve_vma(ev); | |
993 | ||
994 | list_add_tail(&ev->bind_link, &eb->unbound); | |
995 | if (drm_mm_node_allocated(&vma->node)) { | |
996 | err = i915_vma_unbind(vma); | |
997 | if (err) | |
998 | return err; | |
999 | } | |
1000 | } | |
1001 | ||
c8d4c18b CK |
1002 | err = dma_resv_reserve_fences(vma->obj->base.resv, 1); |
1003 | if (err) | |
1004 | return err; | |
bfaae47d | 1005 | |
c43ce123 ML |
1006 | GEM_BUG_ON(drm_mm_node_allocated(&vma->node) && |
1007 | eb_vma_misplaced(&eb->exec[i], vma, ev->flags)); | |
1008 | } | |
1009 | ||
1010 | if (!list_empty(&eb->unbound)) | |
1011 | return eb_reserve(eb); | |
1012 | ||
1013 | return 0; | |
1014 | } | |
1015 | ||
7d6236bb | 1016 | static struct eb_vma * |
2889caa9 | 1017 | eb_get_vma(const struct i915_execbuffer *eb, unsigned long handle) |
67731b87 | 1018 | { |
2889caa9 CW |
1019 | if (eb->lut_size < 0) { |
1020 | if (handle >= -eb->lut_size) | |
eef90ccb | 1021 | return NULL; |
7d6236bb | 1022 | return &eb->vma[handle]; |
eef90ccb CW |
1023 | } else { |
1024 | struct hlist_head *head; | |
7d6236bb | 1025 | struct eb_vma *ev; |
67731b87 | 1026 | |
2889caa9 | 1027 | head = &eb->buckets[hash_32(handle, eb->lut_size)]; |
7d6236bb CW |
1028 | hlist_for_each_entry(ev, head, node) { |
1029 | if (ev->handle == handle) | |
1030 | return ev; | |
eef90ccb CW |
1031 | } |
1032 | return NULL; | |
1033 | } | |
67731b87 CW |
1034 | } |
1035 | ||
b4b9731b | 1036 | static void eb_release_vmas(struct i915_execbuffer *eb, bool final) |
8ae275c2 ML |
1037 | { |
1038 | const unsigned int count = eb->buffer_count; | |
1039 | unsigned int i; | |
1040 | ||
1041 | for (i = 0; i < count; i++) { | |
1042 | struct eb_vma *ev = &eb->vma[i]; | |
1043 | struct i915_vma *vma = ev->vma; | |
1044 | ||
1045 | if (!vma) | |
1046 | break; | |
1047 | ||
c43ce123 | 1048 | eb_unreserve_vma(ev); |
8ae275c2 | 1049 | |
c43ce123 ML |
1050 | if (final) |
1051 | i915_vma_put(vma); | |
8ae275c2 | 1052 | } |
2bf541ff | 1053 | |
ff20afc4 | 1054 | eb_capture_release(eb); |
2bf541ff | 1055 | eb_unpin_engine(eb); |
8ae275c2 ML |
1056 | } |
1057 | ||
2889caa9 | 1058 | static void eb_destroy(const struct i915_execbuffer *eb) |
934acce3 | 1059 | { |
4d470f73 | 1060 | if (eb->lut_size > 0) |
2889caa9 | 1061 | kfree(eb->buckets); |
934acce3 MW |
1062 | } |
1063 | ||
2889caa9 | 1064 | static inline u64 |
d50415cc | 1065 | relocation_target(const struct drm_i915_gem_relocation_entry *reloc, |
2889caa9 | 1066 | const struct i915_vma *target) |
934acce3 | 1067 | { |
2889caa9 | 1068 | return gen8_canonical_addr((int)reloc->delta + target->node.start); |
934acce3 MW |
1069 | } |
1070 | ||
d50415cc CW |
1071 | static void reloc_cache_init(struct reloc_cache *cache, |
1072 | struct drm_i915_private *i915) | |
5032d871 | 1073 | { |
ad5d95e4 DA |
1074 | cache->page = -1; |
1075 | cache->vaddr = 0; | |
dfc5148f | 1076 | /* Must be a variable in the struct to allow GCC to unroll. */ |
8802190f | 1077 | cache->graphics_ver = GRAPHICS_VER(i915); |
2889caa9 | 1078 | cache->has_llc = HAS_LLC(i915); |
dfc5148f | 1079 | cache->use_64bit_reloc = HAS_64BIT_RELOC(i915); |
8802190f | 1080 | cache->has_fence = cache->graphics_ver < 4; |
7dd4f672 | 1081 | cache->needs_unfenced = INTEL_INFO(i915)->unfenced_needs_alignment; |
4ee92c71 | 1082 | cache->node.flags = 0; |
d50415cc | 1083 | } |
5032d871 | 1084 | |
20561da3 DA |
1085 | static inline void *unmask_page(unsigned long p) |
1086 | { | |
1087 | return (void *)(uintptr_t)(p & PAGE_MASK); | |
1088 | } | |
1089 | ||
1090 | static inline unsigned int unmask_flags(unsigned long p) | |
1091 | { | |
1092 | return p & ~PAGE_MASK; | |
1093 | } | |
1094 | ||
1095 | #define KMAP 0x4 /* after CLFLUSH_FLAGS */ | |
1096 | ||
1097 | static inline struct i915_ggtt *cache_to_ggtt(struct reloc_cache *cache) | |
1098 | { | |
1099 | struct drm_i915_private *i915 = | |
1100 | container_of(cache, struct i915_execbuffer, reloc_cache)->i915; | |
5c24c9d2 | 1101 | return to_gt(i915)->ggtt; |
20561da3 DA |
1102 | } |
1103 | ||
c2ea703d TH |
1104 | static void reloc_cache_unmap(struct reloc_cache *cache) |
1105 | { | |
1106 | void *vaddr; | |
1107 | ||
1108 | if (!cache->vaddr) | |
1109 | return; | |
1110 | ||
1111 | vaddr = unmask_page(cache->vaddr); | |
1112 | if (cache->vaddr & KMAP) | |
1113 | kunmap_atomic(vaddr); | |
1114 | else | |
1115 | io_mapping_unmap_atomic((void __iomem *)vaddr); | |
1116 | } | |
1117 | ||
1118 | static void reloc_cache_remap(struct reloc_cache *cache, | |
1119 | struct drm_i915_gem_object *obj) | |
1120 | { | |
1121 | void *vaddr; | |
1122 | ||
1123 | if (!cache->vaddr) | |
1124 | return; | |
1125 | ||
1126 | if (cache->vaddr & KMAP) { | |
1127 | struct page *page = i915_gem_object_get_page(obj, cache->page); | |
1128 | ||
1129 | vaddr = kmap_atomic(page); | |
1130 | cache->vaddr = unmask_flags(cache->vaddr) | | |
1131 | (unsigned long)vaddr; | |
1132 | } else { | |
1133 | struct i915_ggtt *ggtt = cache_to_ggtt(cache); | |
1134 | unsigned long offset; | |
1135 | ||
1136 | offset = cache->node.start; | |
1137 | if (!drm_mm_node_allocated(&cache->node)) | |
1138 | offset += cache->page << PAGE_SHIFT; | |
1139 | ||
1140 | cache->vaddr = (unsigned long) | |
1141 | io_mapping_map_atomic_wc(&ggtt->iomap, offset); | |
1142 | } | |
1143 | } | |
1144 | ||
c43ce123 | 1145 | static void reloc_cache_reset(struct reloc_cache *cache, struct i915_execbuffer *eb) |
ad5d95e4 DA |
1146 | { |
1147 | void *vaddr; | |
1148 | ||
1149 | if (!cache->vaddr) | |
1150 | return; | |
1151 | ||
1152 | vaddr = unmask_page(cache->vaddr); | |
1153 | if (cache->vaddr & KMAP) { | |
1af343cd ML |
1154 | struct drm_i915_gem_object *obj = |
1155 | (struct drm_i915_gem_object *)cache->node.mm; | |
ad5d95e4 DA |
1156 | if (cache->vaddr & CLFLUSH_AFTER) |
1157 | mb(); | |
1158 | ||
1159 | kunmap_atomic(vaddr); | |
1af343cd | 1160 | i915_gem_object_finish_access(obj); |
ad5d95e4 DA |
1161 | } else { |
1162 | struct i915_ggtt *ggtt = cache_to_ggtt(cache); | |
1163 | ||
1164 | intel_gt_flush_ggtt_writes(ggtt->vm.gt); | |
1165 | io_mapping_unmap_atomic((void __iomem *)vaddr); | |
1166 | ||
1167 | if (drm_mm_node_allocated(&cache->node)) { | |
1168 | ggtt->vm.clear_range(&ggtt->vm, | |
1169 | cache->node.start, | |
1170 | cache->node.size); | |
1171 | mutex_lock(&ggtt->vm.mutex); | |
1172 | drm_mm_remove_node(&cache->node); | |
1173 | mutex_unlock(&ggtt->vm.mutex); | |
1174 | } else { | |
1175 | i915_vma_unpin((struct i915_vma *)cache->node.mm); | |
1176 | } | |
1177 | } | |
1178 | ||
1179 | cache->vaddr = 0; | |
1180 | cache->page = -1; | |
1181 | } | |
1182 | ||
1183 | static void *reloc_kmap(struct drm_i915_gem_object *obj, | |
1184 | struct reloc_cache *cache, | |
102a0a90 | 1185 | unsigned long pageno) |
ad5d95e4 DA |
1186 | { |
1187 | void *vaddr; | |
102a0a90 | 1188 | struct page *page; |
ad5d95e4 DA |
1189 | |
1190 | if (cache->vaddr) { | |
1191 | kunmap_atomic(unmask_page(cache->vaddr)); | |
1192 | } else { | |
1193 | unsigned int flushes; | |
1194 | int err; | |
1195 | ||
1196 | err = i915_gem_object_prepare_write(obj, &flushes); | |
1197 | if (err) | |
1198 | return ERR_PTR(err); | |
1199 | ||
1200 | BUILD_BUG_ON(KMAP & CLFLUSH_FLAGS); | |
1201 | BUILD_BUG_ON((KMAP | CLFLUSH_FLAGS) & PAGE_MASK); | |
1202 | ||
1203 | cache->vaddr = flushes | KMAP; | |
1204 | cache->node.mm = (void *)obj; | |
1205 | if (flushes) | |
1206 | mb(); | |
1207 | } | |
1208 | ||
102a0a90 ML |
1209 | page = i915_gem_object_get_page(obj, pageno); |
1210 | if (!obj->mm.dirty) | |
1211 | set_page_dirty(page); | |
1212 | ||
1213 | vaddr = kmap_atomic(page); | |
ad5d95e4 | 1214 | cache->vaddr = unmask_flags(cache->vaddr) | (unsigned long)vaddr; |
102a0a90 | 1215 | cache->page = pageno; |
ad5d95e4 DA |
1216 | |
1217 | return vaddr; | |
1218 | } | |
1219 | ||
b5cfe6f7 | 1220 | static void *reloc_iomap(struct i915_vma *batch, |
47b08693 | 1221 | struct i915_execbuffer *eb, |
ad5d95e4 DA |
1222 | unsigned long page) |
1223 | { | |
b5cfe6f7 | 1224 | struct drm_i915_gem_object *obj = batch->obj; |
47b08693 | 1225 | struct reloc_cache *cache = &eb->reloc_cache; |
ad5d95e4 DA |
1226 | struct i915_ggtt *ggtt = cache_to_ggtt(cache); |
1227 | unsigned long offset; | |
1228 | void *vaddr; | |
1229 | ||
1230 | if (cache->vaddr) { | |
1231 | intel_gt_flush_ggtt_writes(ggtt->vm.gt); | |
1232 | io_mapping_unmap_atomic((void __force __iomem *) unmask_page(cache->vaddr)); | |
1233 | } else { | |
b5cfe6f7 | 1234 | struct i915_vma *vma = ERR_PTR(-ENODEV); |
ad5d95e4 DA |
1235 | int err; |
1236 | ||
1237 | if (i915_gem_object_is_tiled(obj)) | |
1238 | return ERR_PTR(-EINVAL); | |
1239 | ||
1240 | if (use_cpu_reloc(cache, obj)) | |
1241 | return NULL; | |
1242 | ||
ad5d95e4 | 1243 | err = i915_gem_object_set_to_gtt_domain(obj, true); |
ad5d95e4 DA |
1244 | if (err) |
1245 | return ERR_PTR(err); | |
1246 | ||
b5cfe6f7 ML |
1247 | /* |
1248 | * i915_gem_object_ggtt_pin_ww may attempt to remove the batch | |
1249 | * VMA from the object list because we no longer pin. | |
1250 | * | |
1251 | * Only attempt to pin the batch buffer to ggtt if the current batch | |
1252 | * is not inside ggtt, or the batch buffer is not misplaced. | |
1253 | */ | |
1254 | if (!i915_is_ggtt(batch->vm)) { | |
1255 | vma = i915_gem_object_ggtt_pin_ww(obj, &eb->ww, NULL, 0, 0, | |
1256 | PIN_MAPPABLE | | |
1257 | PIN_NONBLOCK /* NOWARN */ | | |
1258 | PIN_NOEVICT); | |
1259 | } else if (i915_vma_is_map_and_fenceable(batch)) { | |
1260 | __i915_vma_pin(batch); | |
1261 | vma = batch; | |
1262 | } | |
1263 | ||
47b08693 ML |
1264 | if (vma == ERR_PTR(-EDEADLK)) |
1265 | return vma; | |
1266 | ||
ad5d95e4 DA |
1267 | if (IS_ERR(vma)) { |
1268 | memset(&cache->node, 0, sizeof(cache->node)); | |
1269 | mutex_lock(&ggtt->vm.mutex); | |
1270 | err = drm_mm_insert_node_in_range | |
1271 | (&ggtt->vm.mm, &cache->node, | |
1272 | PAGE_SIZE, 0, I915_COLOR_UNEVICTABLE, | |
1273 | 0, ggtt->mappable_end, | |
1274 | DRM_MM_INSERT_LOW); | |
1275 | mutex_unlock(&ggtt->vm.mutex); | |
1276 | if (err) /* no inactive aperture space, use cpu reloc */ | |
1277 | return NULL; | |
1278 | } else { | |
1279 | cache->node.start = vma->node.start; | |
1280 | cache->node.mm = (void *)vma; | |
1281 | } | |
1282 | } | |
1283 | ||
1284 | offset = cache->node.start; | |
1285 | if (drm_mm_node_allocated(&cache->node)) { | |
1286 | ggtt->vm.insert_page(&ggtt->vm, | |
1287 | i915_gem_object_get_dma_address(obj, page), | |
1288 | offset, I915_CACHE_NONE, 0); | |
1289 | } else { | |
1290 | offset += page << PAGE_SHIFT; | |
1291 | } | |
1292 | ||
1293 | vaddr = (void __force *)io_mapping_map_atomic_wc(&ggtt->iomap, | |
1294 | offset); | |
1295 | cache->page = page; | |
1296 | cache->vaddr = (unsigned long)vaddr; | |
1297 | ||
1298 | return vaddr; | |
1299 | } | |
1300 | ||
b5cfe6f7 | 1301 | static void *reloc_vaddr(struct i915_vma *vma, |
47b08693 | 1302 | struct i915_execbuffer *eb, |
ad5d95e4 DA |
1303 | unsigned long page) |
1304 | { | |
47b08693 | 1305 | struct reloc_cache *cache = &eb->reloc_cache; |
ad5d95e4 DA |
1306 | void *vaddr; |
1307 | ||
1308 | if (cache->page == page) { | |
1309 | vaddr = unmask_page(cache->vaddr); | |
1310 | } else { | |
1311 | vaddr = NULL; | |
1312 | if ((cache->vaddr & KMAP) == 0) | |
b5cfe6f7 | 1313 | vaddr = reloc_iomap(vma, eb, page); |
ad5d95e4 | 1314 | if (!vaddr) |
b5cfe6f7 | 1315 | vaddr = reloc_kmap(vma->obj, cache, page); |
ad5d95e4 DA |
1316 | } |
1317 | ||
1318 | return vaddr; | |
1319 | } | |
1320 | ||
1321 | static void clflush_write32(u32 *addr, u32 value, unsigned int flushes) | |
1322 | { | |
1323 | if (unlikely(flushes & (CLFLUSH_BEFORE | CLFLUSH_AFTER))) { | |
89754df8 MC |
1324 | if (flushes & CLFLUSH_BEFORE) |
1325 | drm_clflush_virt_range(addr, sizeof(*addr)); | |
ad5d95e4 DA |
1326 | |
1327 | *addr = value; | |
1328 | ||
1329 | /* | |
1330 | * Writes to the same cacheline are serialised by the CPU | |
1331 | * (including clflush). On the write path, we only require | |
1332 | * that it hits memory in an orderly fashion and place | |
1333 | * mb barriers at the start and end of the relocation phase | |
1334 | * to ensure ordering of clflush wrt to the system. | |
1335 | */ | |
1336 | if (flushes & CLFLUSH_AFTER) | |
89754df8 | 1337 | drm_clflush_virt_range(addr, sizeof(*addr)); |
ad5d95e4 DA |
1338 | } else |
1339 | *addr = value; | |
1340 | } | |
1341 | ||
e3d29130 | 1342 | static u64 |
ad5d95e4 | 1343 | relocate_entry(struct i915_vma *vma, |
e3d29130 | 1344 | const struct drm_i915_gem_relocation_entry *reloc, |
ad5d95e4 | 1345 | struct i915_execbuffer *eb, |
e3d29130 CW |
1346 | const struct i915_vma *target) |
1347 | { | |
1348 | u64 target_addr = relocation_target(reloc, target); | |
ad5d95e4 | 1349 | u64 offset = reloc->offset; |
ce13c78f DV |
1350 | bool wide = eb->reloc_cache.use_64bit_reloc; |
1351 | void *vaddr; | |
ad5d95e4 DA |
1352 | |
1353 | repeat: | |
b5cfe6f7 | 1354 | vaddr = reloc_vaddr(vma, eb, |
ce13c78f DV |
1355 | offset >> PAGE_SHIFT); |
1356 | if (IS_ERR(vaddr)) | |
1357 | return PTR_ERR(vaddr); | |
1358 | ||
1359 | GEM_BUG_ON(!IS_ALIGNED(offset, sizeof(u32))); | |
1360 | clflush_write32(vaddr + offset_in_page(offset), | |
1361 | lower_32_bits(target_addr), | |
1362 | eb->reloc_cache.vaddr); | |
1363 | ||
1364 | if (wide) { | |
1365 | offset += sizeof(u32); | |
1366 | target_addr >>= 32; | |
1367 | wide = false; | |
1368 | goto repeat; | |
ad5d95e4 | 1369 | } |
edf4427b | 1370 | |
2889caa9 | 1371 | return target->node.start | UPDATE; |
edf4427b | 1372 | } |
edf4427b | 1373 | |
2889caa9 CW |
1374 | static u64 |
1375 | eb_relocate_entry(struct i915_execbuffer *eb, | |
7d6236bb | 1376 | struct eb_vma *ev, |
2889caa9 | 1377 | const struct drm_i915_gem_relocation_entry *reloc) |
54cf91dc | 1378 | { |
baa89ba3 | 1379 | struct drm_i915_private *i915 = eb->i915; |
7d6236bb | 1380 | struct eb_vma *target; |
2889caa9 | 1381 | int err; |
54cf91dc | 1382 | |
67731b87 | 1383 | /* we've already hold a reference to all valid objects */ |
507d977f CW |
1384 | target = eb_get_vma(eb, reloc->target_handle); |
1385 | if (unlikely(!target)) | |
54cf91dc | 1386 | return -ENOENT; |
e844b990 | 1387 | |
54cf91dc | 1388 | /* Validate that the target is in a valid r/w GPU domain */ |
b8f7ab17 | 1389 | if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) { |
baa89ba3 | 1390 | drm_dbg(&i915->drm, "reloc with multiple write domains: " |
507d977f | 1391 | "target %d offset %d " |
54cf91dc | 1392 | "read %08x write %08x", |
507d977f | 1393 | reloc->target_handle, |
54cf91dc CW |
1394 | (int) reloc->offset, |
1395 | reloc->read_domains, | |
1396 | reloc->write_domain); | |
8b78f0e5 | 1397 | return -EINVAL; |
54cf91dc | 1398 | } |
4ca4a250 DV |
1399 | if (unlikely((reloc->write_domain | reloc->read_domains) |
1400 | & ~I915_GEM_GPU_DOMAINS)) { | |
baa89ba3 | 1401 | drm_dbg(&i915->drm, "reloc with read/write non-GPU domains: " |
507d977f | 1402 | "target %d offset %d " |
54cf91dc | 1403 | "read %08x write %08x", |
507d977f | 1404 | reloc->target_handle, |
54cf91dc CW |
1405 | (int) reloc->offset, |
1406 | reloc->read_domains, | |
1407 | reloc->write_domain); | |
8b78f0e5 | 1408 | return -EINVAL; |
54cf91dc | 1409 | } |
54cf91dc | 1410 | |
2889caa9 | 1411 | if (reloc->write_domain) { |
7d6236bb | 1412 | target->flags |= EXEC_OBJECT_WRITE; |
507d977f | 1413 | |
2889caa9 CW |
1414 | /* |
1415 | * Sandybridge PPGTT errata: We need a global gtt mapping | |
1416 | * for MI and pipe_control writes because the gpu doesn't | |
1417 | * properly redirect them through the ppgtt for non_secure | |
1418 | * batchbuffers. | |
1419 | */ | |
1420 | if (reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION && | |
c2ea703d TH |
1421 | GRAPHICS_VER(eb->i915) == 6 && |
1422 | !i915_vma_is_bound(target->vma, I915_VMA_GLOBAL_BIND)) { | |
1423 | struct i915_vma *vma = target->vma; | |
1424 | ||
1425 | reloc_cache_unmap(&eb->reloc_cache); | |
1426 | mutex_lock(&vma->vm->mutex); | |
7d6236bb CW |
1427 | err = i915_vma_bind(target->vma, |
1428 | target->vma->obj->cache_level, | |
e1a4bbb6 | 1429 | PIN_GLOBAL, NULL, NULL); |
c2ea703d TH |
1430 | mutex_unlock(&vma->vm->mutex); |
1431 | reloc_cache_remap(&eb->reloc_cache, ev->vma->obj); | |
ea97c4ca | 1432 | if (err) |
2889caa9 CW |
1433 | return err; |
1434 | } | |
507d977f | 1435 | } |
54cf91dc | 1436 | |
2889caa9 CW |
1437 | /* |
1438 | * If the relocation already has the right value in it, no | |
54cf91dc CW |
1439 | * more work needs to be done. |
1440 | */ | |
ad5d95e4 DA |
1441 | if (!DBG_FORCE_RELOC && |
1442 | gen8_canonical_addr(target->vma->node.start) == reloc->presumed_offset) | |
67731b87 | 1443 | return 0; |
54cf91dc CW |
1444 | |
1445 | /* Check that the relocation address is valid... */ | |
3c94ceee | 1446 | if (unlikely(reloc->offset > |
7d6236bb | 1447 | ev->vma->size - (eb->reloc_cache.use_64bit_reloc ? 8 : 4))) { |
baa89ba3 | 1448 | drm_dbg(&i915->drm, "Relocation beyond object bounds: " |
507d977f CW |
1449 | "target %d offset %d size %d.\n", |
1450 | reloc->target_handle, | |
1451 | (int)reloc->offset, | |
7d6236bb | 1452 | (int)ev->vma->size); |
8b78f0e5 | 1453 | return -EINVAL; |
54cf91dc | 1454 | } |
b8f7ab17 | 1455 | if (unlikely(reloc->offset & 3)) { |
baa89ba3 | 1456 | drm_dbg(&i915->drm, "Relocation not 4-byte aligned: " |
507d977f CW |
1457 | "target %d offset %d.\n", |
1458 | reloc->target_handle, | |
1459 | (int)reloc->offset); | |
8b78f0e5 | 1460 | return -EINVAL; |
54cf91dc CW |
1461 | } |
1462 | ||
071750e5 CW |
1463 | /* |
1464 | * If we write into the object, we need to force the synchronisation | |
1465 | * barrier, either with an asynchronous clflush or if we executed the | |
1466 | * patching using the GPU (though that should be serialised by the | |
1467 | * timeline). To be completely sure, and since we are required to | |
1468 | * do relocations we are already stalling, disable the user's opt | |
0519bcb1 | 1469 | * out of our synchronisation. |
071750e5 | 1470 | */ |
7d6236bb | 1471 | ev->flags &= ~EXEC_OBJECT_ASYNC; |
071750e5 | 1472 | |
54cf91dc | 1473 | /* and update the user's relocation entry */ |
ad5d95e4 | 1474 | return relocate_entry(ev->vma, reloc, eb, target->vma); |
54cf91dc CW |
1475 | } |
1476 | ||
7d6236bb | 1477 | static int eb_relocate_vma(struct i915_execbuffer *eb, struct eb_vma *ev) |
54cf91dc | 1478 | { |
1d83f442 | 1479 | #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry)) |
2889caa9 | 1480 | struct drm_i915_gem_relocation_entry stack[N_RELOC(512)]; |
7d6236bb | 1481 | const struct drm_i915_gem_exec_object2 *entry = ev->exec; |
e94f7856 CW |
1482 | struct drm_i915_gem_relocation_entry __user *urelocs = |
1483 | u64_to_user_ptr(entry->relocs_ptr); | |
1484 | unsigned long remain = entry->relocation_count; | |
54cf91dc | 1485 | |
e94f7856 | 1486 | if (unlikely(remain > N_RELOC(ULONG_MAX))) |
2889caa9 | 1487 | return -EINVAL; |
ebc0808f | 1488 | |
2889caa9 CW |
1489 | /* |
1490 | * We must check that the entire relocation array is safe | |
1491 | * to read. However, if the array is not writable the user loses | |
1492 | * the updated relocation values. | |
1493 | */ | |
e94f7856 | 1494 | if (unlikely(!access_ok(urelocs, remain * sizeof(*urelocs)))) |
2889caa9 CW |
1495 | return -EFAULT; |
1496 | ||
1497 | do { | |
1498 | struct drm_i915_gem_relocation_entry *r = stack; | |
1499 | unsigned int count = | |
e94f7856 | 1500 | min_t(unsigned long, remain, ARRAY_SIZE(stack)); |
2889caa9 | 1501 | unsigned int copied; |
1d83f442 | 1502 | |
2889caa9 CW |
1503 | /* |
1504 | * This is the fast path and we cannot handle a pagefault | |
ebc0808f CW |
1505 | * whilst holding the struct mutex lest the user pass in the |
1506 | * relocations contained within a mmaped bo. For in such a case | |
1507 | * we, the page fault handler would call i915_gem_fault() and | |
1508 | * we would try to acquire the struct mutex again. Obviously | |
1509 | * this is bad and so lockdep complains vehemently. | |
1510 | */ | |
fd1500fc ML |
1511 | pagefault_disable(); |
1512 | copied = __copy_from_user_inatomic(r, urelocs, count * sizeof(r[0])); | |
1513 | pagefault_enable(); | |
ad5d95e4 DA |
1514 | if (unlikely(copied)) { |
1515 | remain = -EFAULT; | |
1516 | goto out; | |
1517 | } | |
54cf91dc | 1518 | |
2889caa9 | 1519 | remain -= count; |
1d83f442 | 1520 | do { |
7d6236bb | 1521 | u64 offset = eb_relocate_entry(eb, ev, r); |
54cf91dc | 1522 | |
2889caa9 CW |
1523 | if (likely(offset == 0)) { |
1524 | } else if ((s64)offset < 0) { | |
ad5d95e4 DA |
1525 | remain = (int)offset; |
1526 | goto out; | |
2889caa9 CW |
1527 | } else { |
1528 | /* | |
1529 | * Note that reporting an error now | |
1530 | * leaves everything in an inconsistent | |
1531 | * state as we have *already* changed | |
1532 | * the relocation value inside the | |
1533 | * object. As we have not changed the | |
1534 | * reloc.presumed_offset or will not | |
1535 | * change the execobject.offset, on the | |
1536 | * call we may not rewrite the value | |
1537 | * inside the object, leaving it | |
1538 | * dangling and causing a GPU hang. Unless | |
1539 | * userspace dynamically rebuilds the | |
1540 | * relocations on each execbuf rather than | |
1541 | * presume a static tree. | |
1542 | * | |
1543 | * We did previously check if the relocations | |
1544 | * were writable (access_ok), an error now | |
1545 | * would be a strange race with mprotect, | |
1546 | * having already demonstrated that we | |
1547 | * can read from this userspace address. | |
1548 | */ | |
1549 | offset = gen8_canonical_addr(offset & ~UPDATE); | |
97a37c91 CW |
1550 | __put_user(offset, |
1551 | &urelocs[r - stack].presumed_offset); | |
1d83f442 | 1552 | } |
2889caa9 CW |
1553 | } while (r++, --count); |
1554 | urelocs += ARRAY_SIZE(stack); | |
1555 | } while (remain); | |
ad5d95e4 | 1556 | out: |
c43ce123 | 1557 | reloc_cache_reset(&eb->reloc_cache, eb); |
ad5d95e4 | 1558 | return remain; |
54cf91dc CW |
1559 | } |
1560 | ||
fd1500fc ML |
1561 | static int |
1562 | eb_relocate_vma_slow(struct i915_execbuffer *eb, struct eb_vma *ev) | |
54cf91dc | 1563 | { |
fd1500fc ML |
1564 | const struct drm_i915_gem_exec_object2 *entry = ev->exec; |
1565 | struct drm_i915_gem_relocation_entry *relocs = | |
1566 | u64_to_ptr(typeof(*relocs), entry->relocs_ptr); | |
1567 | unsigned int i; | |
003d8b91 CW |
1568 | int err; |
1569 | ||
fd1500fc ML |
1570 | for (i = 0; i < entry->relocation_count; i++) { |
1571 | u64 offset = eb_relocate_entry(eb, ev, &relocs[i]); | |
003d8b91 | 1572 | |
fd1500fc ML |
1573 | if ((s64)offset < 0) { |
1574 | err = (int)offset; | |
1575 | goto err; | |
1576 | } | |
ef398881 | 1577 | } |
fd1500fc ML |
1578 | err = 0; |
1579 | err: | |
c43ce123 | 1580 | reloc_cache_reset(&eb->reloc_cache, eb); |
fd1500fc ML |
1581 | return err; |
1582 | } | |
2889caa9 | 1583 | |
fd1500fc ML |
1584 | static int check_relocations(const struct drm_i915_gem_exec_object2 *entry) |
1585 | { | |
1586 | const char __user *addr, *end; | |
1587 | unsigned long size; | |
1588 | char __maybe_unused c; | |
2889caa9 | 1589 | |
fd1500fc ML |
1590 | size = entry->relocation_count; |
1591 | if (size == 0) | |
1592 | return 0; | |
0e97fbb0 | 1593 | |
fd1500fc ML |
1594 | if (size > N_RELOC(ULONG_MAX)) |
1595 | return -EINVAL; | |
2889caa9 | 1596 | |
fd1500fc ML |
1597 | addr = u64_to_user_ptr(entry->relocs_ptr); |
1598 | size *= sizeof(struct drm_i915_gem_relocation_entry); | |
1599 | if (!access_ok(addr, size)) | |
1600 | return -EFAULT; | |
1601 | ||
1602 | end = addr + size; | |
1603 | for (; addr < end; addr += PAGE_SIZE) { | |
1604 | int err = __get_user(c, addr); | |
1605 | if (err) | |
1606 | return err; | |
1607 | } | |
1608 | return __get_user(c, end - 1); | |
2889caa9 CW |
1609 | } |
1610 | ||
fd1500fc | 1611 | static int eb_copy_relocations(const struct i915_execbuffer *eb) |
2889caa9 | 1612 | { |
fd1500fc | 1613 | struct drm_i915_gem_relocation_entry *relocs; |
2889caa9 CW |
1614 | const unsigned int count = eb->buffer_count; |
1615 | unsigned int i; | |
fd1500fc | 1616 | int err; |
54cf91dc | 1617 | |
2889caa9 | 1618 | for (i = 0; i < count; i++) { |
fd1500fc ML |
1619 | const unsigned int nreloc = eb->exec[i].relocation_count; |
1620 | struct drm_i915_gem_relocation_entry __user *urelocs; | |
1621 | unsigned long size; | |
1622 | unsigned long copied; | |
6951e589 | 1623 | |
fd1500fc ML |
1624 | if (nreloc == 0) |
1625 | continue; | |
6951e589 | 1626 | |
fd1500fc ML |
1627 | err = check_relocations(&eb->exec[i]); |
1628 | if (err) | |
1629 | goto err; | |
6951e589 | 1630 | |
fd1500fc ML |
1631 | urelocs = u64_to_user_ptr(eb->exec[i].relocs_ptr); |
1632 | size = nreloc * sizeof(*relocs); | |
6951e589 | 1633 | |
fd1500fc ML |
1634 | relocs = kvmalloc_array(size, 1, GFP_KERNEL); |
1635 | if (!relocs) { | |
1636 | err = -ENOMEM; | |
1637 | goto err; | |
6951e589 | 1638 | } |
fd1500fc ML |
1639 | |
1640 | /* copy_from_user is limited to < 4GiB */ | |
1641 | copied = 0; | |
1642 | do { | |
1643 | unsigned int len = | |
1644 | min_t(u64, BIT_ULL(31), size - copied); | |
1645 | ||
1646 | if (__copy_from_user((char *)relocs + copied, | |
1647 | (char __user *)urelocs + copied, | |
1648 | len)) | |
1649 | goto end; | |
1650 | ||
1651 | copied += len; | |
1652 | } while (copied < size); | |
1653 | ||
1654 | /* | |
1655 | * As we do not update the known relocation offsets after | |
1656 | * relocating (due to the complexities in lock handling), | |
1657 | * we need to mark them as invalid now so that we force the | |
1658 | * relocation processing next time. Just in case the target | |
1659 | * object is evicted and then rebound into its old | |
1660 | * presumed_offset before the next execbuffer - if that | |
1661 | * happened we would make the mistake of assuming that the | |
1662 | * relocations were valid. | |
1663 | */ | |
1664 | if (!user_access_begin(urelocs, size)) | |
1665 | goto end; | |
1666 | ||
1667 | for (copied = 0; copied < nreloc; copied++) | |
1668 | unsafe_put_user(-1, | |
1669 | &urelocs[copied].presumed_offset, | |
1670 | end_user); | |
1671 | user_access_end(); | |
1672 | ||
1673 | eb->exec[i].relocs_ptr = (uintptr_t)relocs; | |
1674 | } | |
1675 | ||
1676 | return 0; | |
1677 | ||
1678 | end_user: | |
1679 | user_access_end(); | |
1680 | end: | |
1681 | kvfree(relocs); | |
1682 | err = -EFAULT; | |
1683 | err: | |
1684 | while (i--) { | |
1685 | relocs = u64_to_ptr(typeof(*relocs), eb->exec[i].relocs_ptr); | |
1686 | if (eb->exec[i].relocation_count) | |
1687 | kvfree(relocs); | |
1688 | } | |
1689 | return err; | |
1690 | } | |
1691 | ||
1692 | static int eb_prefault_relocations(const struct i915_execbuffer *eb) | |
1693 | { | |
1694 | const unsigned int count = eb->buffer_count; | |
1695 | unsigned int i; | |
1696 | ||
1697 | for (i = 0; i < count; i++) { | |
1698 | int err; | |
1699 | ||
1700 | err = check_relocations(&eb->exec[i]); | |
1701 | if (err) | |
1702 | return err; | |
1703 | } | |
1704 | ||
1705 | return 0; | |
1706 | } | |
1707 | ||
ed29c269 ML |
1708 | static int eb_reinit_userptr(struct i915_execbuffer *eb) |
1709 | { | |
1710 | const unsigned int count = eb->buffer_count; | |
1711 | unsigned int i; | |
1712 | int ret; | |
1713 | ||
1714 | if (likely(!(eb->args->flags & __EXEC_USERPTR_USED))) | |
1715 | return 0; | |
1716 | ||
1717 | for (i = 0; i < count; i++) { | |
1718 | struct eb_vma *ev = &eb->vma[i]; | |
1719 | ||
1720 | if (!i915_gem_object_is_userptr(ev->vma->obj)) | |
1721 | continue; | |
1722 | ||
1723 | ret = i915_gem_object_userptr_submit_init(ev->vma->obj); | |
1724 | if (ret) | |
1725 | return ret; | |
1726 | ||
1727 | ev->flags |= __EXEC_OBJECT_USERPTR_INIT; | |
1728 | } | |
1729 | ||
1730 | return 0; | |
1731 | } | |
1732 | ||
544460c3 | 1733 | static noinline int eb_relocate_parse_slow(struct i915_execbuffer *eb) |
fd1500fc ML |
1734 | { |
1735 | bool have_copy = false; | |
1736 | struct eb_vma *ev; | |
1737 | int err = 0; | |
1738 | ||
1739 | repeat: | |
1740 | if (signal_pending(current)) { | |
1741 | err = -ERESTARTSYS; | |
1742 | goto out; | |
6951e589 | 1743 | } |
fd1500fc | 1744 | |
c43ce123 | 1745 | /* We may process another execbuffer during the unlock... */ |
b4b9731b | 1746 | eb_release_vmas(eb, false); |
c43ce123 ML |
1747 | i915_gem_ww_ctx_fini(&eb->ww); |
1748 | ||
fd1500fc ML |
1749 | /* |
1750 | * We take 3 passes through the slowpatch. | |
1751 | * | |
1752 | * 1 - we try to just prefault all the user relocation entries and | |
1753 | * then attempt to reuse the atomic pagefault disabled fast path again. | |
1754 | * | |
1755 | * 2 - we copy the user entries to a local buffer here outside of the | |
1756 | * local and allow ourselves to wait upon any rendering before | |
1757 | * relocations | |
1758 | * | |
1759 | * 3 - we already have a local copy of the relocation entries, but | |
1760 | * were interrupted (EAGAIN) whilst waiting for the objects, try again. | |
1761 | */ | |
1762 | if (!err) { | |
1763 | err = eb_prefault_relocations(eb); | |
1764 | } else if (!have_copy) { | |
1765 | err = eb_copy_relocations(eb); | |
1766 | have_copy = err == 0; | |
1767 | } else { | |
1768 | cond_resched(); | |
1769 | err = 0; | |
1770 | } | |
1771 | ||
2bf541ff | 1772 | if (!err) |
ed29c269 | 1773 | err = eb_reinit_userptr(eb); |
fd1500fc | 1774 | |
c43ce123 | 1775 | i915_gem_ww_ctx_init(&eb->ww, true); |
fd1500fc ML |
1776 | if (err) |
1777 | goto out; | |
1778 | ||
c43ce123 ML |
1779 | /* reacquire the objects */ |
1780 | repeat_validate: | |
544460c3 MB |
1781 | err = eb_pin_engine(eb, false); |
1782 | if (err) | |
2bf541ff | 1783 | goto err; |
2bf541ff | 1784 | |
c43ce123 | 1785 | err = eb_validate_vmas(eb); |
fd1500fc | 1786 | if (err) |
c43ce123 ML |
1787 | goto err; |
1788 | ||
544460c3 | 1789 | GEM_BUG_ON(!eb->batches[0]); |
fd1500fc ML |
1790 | |
1791 | list_for_each_entry(ev, &eb->relocs, reloc_link) { | |
1792 | if (!have_copy) { | |
fd1500fc | 1793 | err = eb_relocate_vma(eb, ev); |
fd1500fc ML |
1794 | if (err) |
1795 | break; | |
1796 | } else { | |
1797 | err = eb_relocate_vma_slow(eb, ev); | |
1798 | if (err) | |
1799 | break; | |
1800 | } | |
1801 | } | |
1802 | ||
c43ce123 ML |
1803 | if (err == -EDEADLK) |
1804 | goto err; | |
1805 | ||
fd1500fc ML |
1806 | if (err && !have_copy) |
1807 | goto repeat; | |
1808 | ||
1809 | if (err) | |
1810 | goto err; | |
1811 | ||
8e4ba491 ML |
1812 | /* as last step, parse the command buffer */ |
1813 | err = eb_parse(eb); | |
1814 | if (err) | |
1815 | goto err; | |
1816 | ||
fd1500fc ML |
1817 | /* |
1818 | * Leave the user relocations as are, this is the painfully slow path, | |
1819 | * and we want to avoid the complication of dropping the lock whilst | |
1820 | * having buffers reserved in the aperture and so causing spurious | |
1821 | * ENOSPC for random operations. | |
1822 | */ | |
1823 | ||
1824 | err: | |
c43ce123 | 1825 | if (err == -EDEADLK) { |
b4b9731b | 1826 | eb_release_vmas(eb, false); |
c43ce123 ML |
1827 | err = i915_gem_ww_ctx_backoff(&eb->ww); |
1828 | if (!err) | |
1829 | goto repeat_validate; | |
1830 | } | |
1831 | ||
fd1500fc ML |
1832 | if (err == -EAGAIN) |
1833 | goto repeat; | |
1834 | ||
1835 | out: | |
1836 | if (have_copy) { | |
1837 | const unsigned int count = eb->buffer_count; | |
1838 | unsigned int i; | |
1839 | ||
1840 | for (i = 0; i < count; i++) { | |
1841 | const struct drm_i915_gem_exec_object2 *entry = | |
1842 | &eb->exec[i]; | |
1843 | struct drm_i915_gem_relocation_entry *relocs; | |
1844 | ||
1845 | if (!entry->relocation_count) | |
1846 | continue; | |
1847 | ||
1848 | relocs = u64_to_ptr(typeof(*relocs), entry->relocs_ptr); | |
1849 | kvfree(relocs); | |
1850 | } | |
1851 | } | |
1852 | ||
1853 | return err; | |
1854 | } | |
1855 | ||
8e4ba491 | 1856 | static int eb_relocate_parse(struct i915_execbuffer *eb) |
54cf91dc | 1857 | { |
003d8b91 | 1858 | int err; |
2bf541ff | 1859 | bool throttle = true; |
003d8b91 | 1860 | |
c43ce123 | 1861 | retry: |
544460c3 MB |
1862 | err = eb_pin_engine(eb, throttle); |
1863 | if (err) { | |
2bf541ff ML |
1864 | if (err != -EDEADLK) |
1865 | return err; | |
1866 | ||
1867 | goto err; | |
1868 | } | |
1869 | ||
2bf541ff ML |
1870 | /* only throttle once, even if we didn't need to throttle */ |
1871 | throttle = false; | |
1872 | ||
c43ce123 ML |
1873 | err = eb_validate_vmas(eb); |
1874 | if (err == -EAGAIN) | |
1875 | goto slow; | |
1876 | else if (err) | |
1877 | goto err; | |
2889caa9 CW |
1878 | |
1879 | /* The objects are in their final locations, apply the relocations. */ | |
1880 | if (eb->args->flags & __EXEC_HAS_RELOC) { | |
7d6236bb | 1881 | struct eb_vma *ev; |
2889caa9 | 1882 | |
7d6236bb | 1883 | list_for_each_entry(ev, &eb->relocs, reloc_link) { |
7dc8f114 CW |
1884 | err = eb_relocate_vma(eb, ev); |
1885 | if (err) | |
fd1500fc | 1886 | break; |
2889caa9 | 1887 | } |
fd1500fc | 1888 | |
c43ce123 ML |
1889 | if (err == -EDEADLK) |
1890 | goto err; | |
1891 | else if (err) | |
1892 | goto slow; | |
1893 | } | |
1894 | ||
1895 | if (!err) | |
1896 | err = eb_parse(eb); | |
1897 | ||
1898 | err: | |
1899 | if (err == -EDEADLK) { | |
b4b9731b | 1900 | eb_release_vmas(eb, false); |
c43ce123 ML |
1901 | err = i915_gem_ww_ctx_backoff(&eb->ww); |
1902 | if (!err) | |
1903 | goto retry; | |
2889caa9 CW |
1904 | } |
1905 | ||
c43ce123 ML |
1906 | return err; |
1907 | ||
1908 | slow: | |
544460c3 | 1909 | err = eb_relocate_parse_slow(eb); |
c43ce123 ML |
1910 | if (err) |
1911 | /* | |
1912 | * If the user expects the execobject.offset and | |
1913 | * reloc.presumed_offset to be an exact match, | |
1914 | * as for using NO_RELOC, then we cannot update | |
1915 | * the execobject.offset until we have completed | |
1916 | * relocation. | |
1917 | */ | |
1918 | eb->args->flags &= ~__EXEC_HAS_RELOC; | |
1919 | ||
1920 | return err; | |
2889caa9 CW |
1921 | } |
1922 | ||
544460c3 MB |
1923 | /* |
1924 | * Using two helper loops for the order of which requests / batches are created | |
1925 | * and added the to backend. Requests are created in order from the parent to | |
1926 | * the last child. Requests are added in the reverse order, from the last child | |
1927 | * to parent. This is done for locking reasons as the timeline lock is acquired | |
1928 | * during request creation and released when the request is added to the | |
1929 | * backend. To make lockdep happy (see intel_context_timeline_lock) this must be | |
1930 | * the ordering. | |
1931 | */ | |
1932 | #define for_each_batch_create_order(_eb, _i) \ | |
1933 | for ((_i) = 0; (_i) < (_eb)->num_batches; ++(_i)) | |
1934 | #define for_each_batch_add_order(_eb, _i) \ | |
1935 | BUILD_BUG_ON(!typecheck(int, _i)); \ | |
1936 | for ((_i) = (_eb)->num_batches - 1; (_i) >= 0; --(_i)) | |
1937 | ||
1938 | static struct i915_request * | |
1939 | eb_find_first_request_added(struct i915_execbuffer *eb) | |
1940 | { | |
1941 | int i; | |
1942 | ||
1943 | for_each_batch_add_order(eb, i) | |
1944 | if (eb->requests[i]) | |
1945 | return eb->requests[i]; | |
1946 | ||
1947 | GEM_BUG_ON("Request not found"); | |
1948 | ||
1949 | return NULL; | |
1950 | } | |
1951 | ||
ff20afc4 TH |
1952 | #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR) |
1953 | ||
1954 | /* Stage with GFP_KERNEL allocations before we enter the signaling critical path */ | |
1955 | static void eb_capture_stage(struct i915_execbuffer *eb) | |
2889caa9 CW |
1956 | { |
1957 | const unsigned int count = eb->buffer_count; | |
ff20afc4 | 1958 | unsigned int i = count, j; |
6951e589 CW |
1959 | |
1960 | while (i--) { | |
7d6236bb CW |
1961 | struct eb_vma *ev = &eb->vma[i]; |
1962 | struct i915_vma *vma = ev->vma; | |
1963 | unsigned int flags = ev->flags; | |
03ade511 | 1964 | |
ff20afc4 TH |
1965 | if (!(flags & EXEC_OBJECT_CAPTURE)) |
1966 | continue; | |
6951e589 | 1967 | |
ff20afc4 | 1968 | for_each_batch_create_order(eb, j) { |
e61e0f51 | 1969 | struct i915_capture_list *capture; |
b0fd47ad | 1970 | |
ff20afc4 TH |
1971 | capture = kmalloc(sizeof(*capture), GFP_KERNEL); |
1972 | if (!capture) | |
1973 | continue; | |
544460c3 | 1974 | |
ff20afc4 | 1975 | capture->next = eb->capture_lists[j]; |
60dc43d1 | 1976 | capture->vma_res = i915_vma_resource_get(vma->resource); |
ff20afc4 TH |
1977 | eb->capture_lists[j] = capture; |
1978 | } | |
ff20afc4 TH |
1979 | } |
1980 | } | |
1981 | ||
1982 | /* Commit once we're in the critical path */ | |
1983 | static void eb_capture_commit(struct i915_execbuffer *eb) | |
1984 | { | |
1985 | unsigned int j; | |
1986 | ||
1987 | for_each_batch_create_order(eb, j) { | |
1988 | struct i915_request *rq = eb->requests[j]; | |
1989 | ||
1990 | if (!rq) | |
1991 | break; | |
1992 | ||
1993 | rq->capture_list = eb->capture_lists[j]; | |
1994 | eb->capture_lists[j] = NULL; | |
1995 | } | |
1996 | } | |
1997 | ||
1998 | /* | |
1999 | * Release anything that didn't get committed due to errors. | |
2000 | * The capture_list will otherwise be freed at request retire. | |
2001 | */ | |
2002 | static void eb_capture_release(struct i915_execbuffer *eb) | |
2003 | { | |
2004 | unsigned int j; | |
2005 | ||
2006 | for_each_batch_create_order(eb, j) { | |
2007 | if (eb->capture_lists[j]) { | |
2008 | i915_request_free_capture_list(eb->capture_lists[j]); | |
2009 | eb->capture_lists[j] = NULL; | |
b0fd47ad | 2010 | } |
ff20afc4 TH |
2011 | } |
2012 | } | |
2013 | ||
2014 | static void eb_capture_list_clear(struct i915_execbuffer *eb) | |
2015 | { | |
2016 | memset(eb->capture_lists, 0, sizeof(eb->capture_lists)); | |
2017 | } | |
2018 | ||
2019 | #else | |
2020 | ||
2021 | static void eb_capture_stage(struct i915_execbuffer *eb) | |
2022 | { | |
2023 | } | |
2024 | ||
2025 | static void eb_capture_commit(struct i915_execbuffer *eb) | |
2026 | { | |
2027 | } | |
2028 | ||
2029 | static void eb_capture_release(struct i915_execbuffer *eb) | |
2030 | { | |
2031 | } | |
2032 | ||
2033 | static void eb_capture_list_clear(struct i915_execbuffer *eb) | |
2034 | { | |
2035 | } | |
2036 | ||
2037 | #endif | |
2038 | ||
2039 | static int eb_move_to_gpu(struct i915_execbuffer *eb) | |
2040 | { | |
2041 | const unsigned int count = eb->buffer_count; | |
2042 | unsigned int i = count; | |
2043 | int err = 0, j; | |
2044 | ||
2045 | while (i--) { | |
2046 | struct eb_vma *ev = &eb->vma[i]; | |
2047 | struct i915_vma *vma = ev->vma; | |
2048 | unsigned int flags = ev->flags; | |
2049 | struct drm_i915_gem_object *obj = vma->obj; | |
2050 | ||
2051 | assert_vma_held(vma); | |
b0fd47ad | 2052 | |
b8f55be6 CW |
2053 | /* |
2054 | * If the GPU is not _reading_ through the CPU cache, we need | |
2055 | * to make sure that any writes (both previous GPU writes from | |
2056 | * before a change in snooping levels and normal CPU writes) | |
2057 | * caught in that cache are flushed to main memory. | |
2058 | * | |
2059 | * We want to say | |
2060 | * obj->cache_dirty && | |
2061 | * !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ) | |
2062 | * but gcc's optimiser doesn't handle that as well and emits | |
2063 | * two jumps instead of one. Maybe one day... | |
df94fd05 MA |
2064 | * |
2065 | * FIXME: There is also sync flushing in set_pages(), which | |
2066 | * serves a different purpose(some of the time at least). | |
2067 | * | |
2068 | * We should consider: | |
2069 | * | |
2070 | * 1. Rip out the async flush code. | |
2071 | * | |
2072 | * 2. Or make the sync flushing use the async clflush path | |
2073 | * using mandatory fences underneath. Currently the below | |
2074 | * async flush happens after we bind the object. | |
b8f55be6 CW |
2075 | */ |
2076 | if (unlikely(obj->cache_dirty & ~obj->cache_coherent)) { | |
0f46daa1 | 2077 | if (i915_gem_clflush_object(obj, 0)) |
c7c6e46f | 2078 | flags &= ~EXEC_OBJECT_ASYNC; |
0f46daa1 CW |
2079 | } |
2080 | ||
544460c3 | 2081 | /* We only need to await on the first request */ |
6951e589 CW |
2082 | if (err == 0 && !(flags & EXEC_OBJECT_ASYNC)) { |
2083 | err = i915_request_await_object | |
544460c3 MB |
2084 | (eb_find_first_request_added(eb), obj, |
2085 | flags & EXEC_OBJECT_WRITE); | |
6951e589 | 2086 | } |
2889caa9 | 2087 | |
544460c3 MB |
2088 | for_each_batch_add_order(eb, j) { |
2089 | if (err) | |
2090 | break; | |
2091 | if (!eb->requests[j]) | |
2092 | continue; | |
2093 | ||
2094 | err = _i915_vma_move_to_active(vma, eb->requests[j], | |
2095 | j ? NULL : | |
2096 | eb->composite_fence ? | |
2097 | eb->composite_fence : | |
2098 | &eb->requests[j]->fence, | |
2099 | flags | __EXEC_OBJECT_NO_RESERVE); | |
2100 | } | |
c59a333f | 2101 | } |
0f1dd022 | 2102 | |
ed29c269 ML |
2103 | #ifdef CONFIG_MMU_NOTIFIER |
2104 | if (!err && (eb->args->flags & __EXEC_USERPTR_USED)) { | |
b4b9731b | 2105 | read_lock(&eb->i915->mm.notifier_lock); |
ed29c269 ML |
2106 | |
2107 | /* | |
2108 | * count is always at least 1, otherwise __EXEC_USERPTR_USED | |
2109 | * could not have been set | |
2110 | */ | |
2111 | for (i = 0; i < count; i++) { | |
2112 | struct eb_vma *ev = &eb->vma[i]; | |
2113 | struct drm_i915_gem_object *obj = ev->vma->obj; | |
2114 | ||
2115 | if (!i915_gem_object_is_userptr(obj)) | |
2116 | continue; | |
2117 | ||
2118 | err = i915_gem_object_userptr_submit_done(obj); | |
2119 | if (err) | |
2120 | break; | |
2121 | } | |
2122 | ||
b4b9731b | 2123 | read_unlock(&eb->i915->mm.notifier_lock); |
ed29c269 ML |
2124 | } |
2125 | #endif | |
2126 | ||
6951e589 CW |
2127 | if (unlikely(err)) |
2128 | goto err_skip; | |
2129 | ||
dcd79934 | 2130 | /* Unconditionally flush any chipset caches (for streaming writes). */ |
544460c3 | 2131 | intel_gt_chipset_flush(eb->gt); |
ff20afc4 TH |
2132 | eb_capture_commit(eb); |
2133 | ||
2113184c | 2134 | return 0; |
6951e589 CW |
2135 | |
2136 | err_skip: | |
544460c3 MB |
2137 | for_each_batch_create_order(eb, j) { |
2138 | if (!eb->requests[j]) | |
2139 | break; | |
2140 | ||
2141 | i915_request_set_error_once(eb->requests[j], err); | |
2142 | } | |
6951e589 | 2143 | return err; |
54cf91dc CW |
2144 | } |
2145 | ||
00aff3f6 | 2146 | static int i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec) |
54cf91dc | 2147 | { |
650bc635 | 2148 | if (exec->flags & __I915_EXEC_ILLEGAL_FLAGS) |
00aff3f6 | 2149 | return -EINVAL; |
ed5982e6 | 2150 | |
2f5945bc | 2151 | /* Kernel clipping was a DRI1 misfeature */ |
cda9edd0 LL |
2152 | if (!(exec->flags & (I915_EXEC_FENCE_ARRAY | |
2153 | I915_EXEC_USE_EXTENSIONS))) { | |
cf6e7bac | 2154 | if (exec->num_cliprects || exec->cliprects_ptr) |
00aff3f6 | 2155 | return -EINVAL; |
cf6e7bac | 2156 | } |
2f5945bc CW |
2157 | |
2158 | if (exec->DR4 == 0xffffffff) { | |
2159 | DRM_DEBUG("UXA submitting garbage DR4, fixing up\n"); | |
2160 | exec->DR4 = 0; | |
2161 | } | |
2162 | if (exec->DR1 || exec->DR4) | |
00aff3f6 | 2163 | return -EINVAL; |
2f5945bc CW |
2164 | |
2165 | if ((exec->batch_start_offset | exec->batch_len) & 0x7) | |
00aff3f6 | 2166 | return -EINVAL; |
2f5945bc | 2167 | |
00aff3f6 | 2168 | return 0; |
54cf91dc CW |
2169 | } |
2170 | ||
e61e0f51 | 2171 | static int i915_reset_gen7_sol_offsets(struct i915_request *rq) |
ae662d31 | 2172 | { |
73dec95e TU |
2173 | u32 *cs; |
2174 | int i; | |
ae662d31 | 2175 | |
40e1956e | 2176 | if (GRAPHICS_VER(rq->engine->i915) != 7 || rq->engine->id != RCS0) { |
5a833995 | 2177 | drm_dbg(&rq->engine->i915->drm, "sol reset is gen7/rcs only\n"); |
9d662da8 DV |
2178 | return -EINVAL; |
2179 | } | |
ae662d31 | 2180 | |
e61e0f51 | 2181 | cs = intel_ring_begin(rq, 4 * 2 + 2); |
73dec95e TU |
2182 | if (IS_ERR(cs)) |
2183 | return PTR_ERR(cs); | |
ae662d31 | 2184 | |
2889caa9 | 2185 | *cs++ = MI_LOAD_REGISTER_IMM(4); |
ae662d31 | 2186 | for (i = 0; i < 4; i++) { |
73dec95e TU |
2187 | *cs++ = i915_mmio_reg_offset(GEN7_SO_WRITE_OFFSET(i)); |
2188 | *cs++ = 0; | |
ae662d31 | 2189 | } |
2889caa9 | 2190 | *cs++ = MI_NOOP; |
e61e0f51 | 2191 | intel_ring_advance(rq, cs); |
ae662d31 EA |
2192 | |
2193 | return 0; | |
2194 | } | |
2195 | ||
4f7af194 | 2196 | static struct i915_vma * |
47b08693 ML |
2197 | shadow_batch_pin(struct i915_execbuffer *eb, |
2198 | struct drm_i915_gem_object *obj, | |
32d94048 CW |
2199 | struct i915_address_space *vm, |
2200 | unsigned int flags) | |
4f7af194 | 2201 | { |
b291ce0a | 2202 | struct i915_vma *vma; |
b291ce0a | 2203 | int err; |
4f7af194 | 2204 | |
b291ce0a CW |
2205 | vma = i915_vma_instance(obj, vm, NULL); |
2206 | if (IS_ERR(vma)) | |
2207 | return vma; | |
2208 | ||
b5cfe6f7 | 2209 | err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, flags | PIN_VALIDATE); |
b291ce0a CW |
2210 | if (err) |
2211 | return ERR_PTR(err); | |
2212 | ||
2213 | return vma; | |
4f7af194 JB |
2214 | } |
2215 | ||
47b08693 ML |
2216 | static struct i915_vma *eb_dispatch_secure(struct i915_execbuffer *eb, struct i915_vma *vma) |
2217 | { | |
2218 | /* | |
2219 | * snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure | |
2220 | * batch" bit. Hence we need to pin secure batches into the global gtt. | |
2221 | * hsw should have this fixed, but bdw mucks it up again. */ | |
2222 | if (eb->batch_flags & I915_DISPATCH_SECURE) | |
b5cfe6f7 | 2223 | return i915_gem_object_ggtt_pin_ww(vma->obj, &eb->ww, NULL, 0, 0, PIN_VALIDATE); |
47b08693 ML |
2224 | |
2225 | return NULL; | |
2226 | } | |
2227 | ||
51696691 | 2228 | static int eb_parse(struct i915_execbuffer *eb) |
71745376 | 2229 | { |
baa89ba3 | 2230 | struct drm_i915_private *i915 = eb->i915; |
c43ce123 | 2231 | struct intel_gt_buffer_pool_node *pool = eb->batch_pool; |
47b08693 | 2232 | struct i915_vma *shadow, *trampoline, *batch; |
d5e87821 | 2233 | unsigned long len; |
2889caa9 | 2234 | int err; |
71745376 | 2235 | |
47b08693 | 2236 | if (!eb_use_cmdparser(eb)) { |
544460c3 | 2237 | batch = eb_dispatch_secure(eb, eb->batches[0]->vma); |
47b08693 ML |
2238 | if (IS_ERR(batch)) |
2239 | return PTR_ERR(batch); | |
2240 | ||
2241 | goto secure_batch; | |
2242 | } | |
51696691 | 2243 | |
544460c3 MB |
2244 | if (intel_context_is_parallel(eb->context)) |
2245 | return -EINVAL; | |
2246 | ||
2247 | len = eb->batch_len[0]; | |
32d94048 CW |
2248 | if (!CMDPARSER_USES_GGTT(eb->i915)) { |
2249 | /* | |
2250 | * ppGTT backed shadow buffers must be mapped RO, to prevent | |
2251 | * post-scan tampering | |
2252 | */ | |
2253 | if (!eb->context->vm->has_read_only) { | |
baa89ba3 WK |
2254 | drm_dbg(&i915->drm, |
2255 | "Cannot prevent post-scan tampering without RO capable vm\n"); | |
32d94048 CW |
2256 | return -EINVAL; |
2257 | } | |
2258 | } else { | |
2259 | len += I915_CMD_PARSER_TRAMPOLINE_SIZE; | |
2260 | } | |
544460c3 | 2261 | if (unlikely(len < eb->batch_len[0])) /* last paranoid check of overflow */ |
d5e87821 | 2262 | return -EINVAL; |
32d94048 | 2263 | |
c43ce123 | 2264 | if (!pool) { |
544460c3 | 2265 | pool = intel_gt_get_buffer_pool(eb->gt, len, |
8f47c8c3 | 2266 | I915_MAP_WB); |
c43ce123 ML |
2267 | if (IS_ERR(pool)) |
2268 | return PTR_ERR(pool); | |
2269 | eb->batch_pool = pool; | |
2270 | } | |
71745376 | 2271 | |
c43ce123 ML |
2272 | err = i915_gem_object_lock(pool->obj, &eb->ww); |
2273 | if (err) | |
b5cfe6f7 | 2274 | return err; |
71745376 | 2275 | |
47b08693 | 2276 | shadow = shadow_batch_pin(eb, pool->obj, eb->context->vm, PIN_USER); |
b5cfe6f7 ML |
2277 | if (IS_ERR(shadow)) |
2278 | return PTR_ERR(shadow); | |
2279 | ||
c9398775 | 2280 | intel_gt_buffer_pool_mark_used(pool); |
32d94048 | 2281 | i915_gem_object_set_readonly(shadow->obj); |
57a78ca4 | 2282 | shadow->private = pool; |
32d94048 CW |
2283 | |
2284 | trampoline = NULL; | |
2285 | if (CMDPARSER_USES_GGTT(eb->i915)) { | |
2286 | trampoline = shadow; | |
2287 | ||
47b08693 | 2288 | shadow = shadow_batch_pin(eb, pool->obj, |
544460c3 | 2289 | &eb->gt->ggtt->vm, |
32d94048 | 2290 | PIN_GLOBAL); |
b5cfe6f7 ML |
2291 | if (IS_ERR(shadow)) |
2292 | return PTR_ERR(shadow); | |
2293 | ||
57a78ca4 | 2294 | shadow->private = pool; |
32d94048 CW |
2295 | |
2296 | eb->batch_flags |= I915_DISPATCH_SECURE; | |
2297 | } | |
f8c08d8f | 2298 | |
47b08693 | 2299 | batch = eb_dispatch_secure(eb, shadow); |
b5cfe6f7 ML |
2300 | if (IS_ERR(batch)) |
2301 | return PTR_ERR(batch); | |
47b08693 | 2302 | |
c8d4c18b | 2303 | err = dma_resv_reserve_fences(shadow->obj->base.resv, 1); |
93b71330 | 2304 | if (err) |
b5cfe6f7 | 2305 | return err; |
93b71330 | 2306 | |
544460c3 MB |
2307 | err = intel_engine_cmd_parser(eb->context->engine, |
2308 | eb->batches[0]->vma, | |
93b71330 | 2309 | eb->batch_start_offset, |
544460c3 | 2310 | eb->batch_len[0], |
93b71330 | 2311 | shadow, trampoline); |
32d94048 | 2312 | if (err) |
b5cfe6f7 | 2313 | return err; |
71745376 | 2314 | |
544460c3 MB |
2315 | eb->batches[0] = &eb->vma[eb->buffer_count++]; |
2316 | eb->batches[0]->vma = i915_vma_get(shadow); | |
2317 | eb->batches[0]->flags = __EXEC_OBJECT_HAS_PIN; | |
71745376 | 2318 | |
32d94048 | 2319 | eb->trampoline = trampoline; |
4f7af194 | 2320 | eb->batch_start_offset = 0; |
4f7af194 | 2321 | |
47b08693 ML |
2322 | secure_batch: |
2323 | if (batch) { | |
544460c3 MB |
2324 | if (intel_context_is_parallel(eb->context)) |
2325 | return -EINVAL; | |
2326 | ||
2327 | eb->batches[0] = &eb->vma[eb->buffer_count++]; | |
2328 | eb->batches[0]->flags = __EXEC_OBJECT_HAS_PIN; | |
2329 | eb->batches[0]->vma = i915_vma_get(batch); | |
47b08693 | 2330 | } |
51696691 | 2331 | return 0; |
71745376 | 2332 | } |
5c6c6003 | 2333 | |
544460c3 MB |
2334 | static int eb_request_submit(struct i915_execbuffer *eb, |
2335 | struct i915_request *rq, | |
2336 | struct i915_vma *batch, | |
2337 | u64 batch_len) | |
78382593 | 2338 | { |
2889caa9 | 2339 | int err; |
78382593 | 2340 | |
544460c3 MB |
2341 | if (intel_context_nopreempt(rq->context)) |
2342 | __set_bit(I915_FENCE_FLAG_NOPREEMPT, &rq->fence.flags); | |
78382593 | 2343 | |
650bc635 | 2344 | if (eb->args->flags & I915_EXEC_GEN7_SOL_RESET) { |
544460c3 | 2345 | err = i915_reset_gen7_sol_offsets(rq); |
2889caa9 CW |
2346 | if (err) |
2347 | return err; | |
78382593 OM |
2348 | } |
2349 | ||
85474441 CW |
2350 | /* |
2351 | * After we completed waiting for other engines (using HW semaphores) | |
2352 | * then we can signal that this request/batch is ready to run. This | |
2353 | * allows us to determine if the batch is still waiting on the GPU | |
2354 | * or actually running by checking the breadcrumb. | |
2355 | */ | |
544460c3 MB |
2356 | if (rq->context->engine->emit_init_breadcrumb) { |
2357 | err = rq->context->engine->emit_init_breadcrumb(rq); | |
85474441 CW |
2358 | if (err) |
2359 | return err; | |
2360 | } | |
2361 | ||
544460c3 MB |
2362 | err = rq->context->engine->emit_bb_start(rq, |
2363 | batch->node.start + | |
2364 | eb->batch_start_offset, | |
2365 | batch_len, | |
2366 | eb->batch_flags); | |
2889caa9 CW |
2367 | if (err) |
2368 | return err; | |
78382593 | 2369 | |
32d94048 | 2370 | if (eb->trampoline) { |
544460c3 | 2371 | GEM_BUG_ON(intel_context_is_parallel(rq->context)); |
32d94048 | 2372 | GEM_BUG_ON(eb->batch_start_offset); |
544460c3 MB |
2373 | err = rq->context->engine->emit_bb_start(rq, |
2374 | eb->trampoline->node.start + | |
2375 | batch_len, 0, 0); | |
32d94048 CW |
2376 | if (err) |
2377 | return err; | |
2378 | } | |
2379 | ||
2f5945bc | 2380 | return 0; |
78382593 OM |
2381 | } |
2382 | ||
544460c3 MB |
2383 | static int eb_submit(struct i915_execbuffer *eb) |
2384 | { | |
2385 | unsigned int i; | |
2386 | int err; | |
2387 | ||
2388 | err = eb_move_to_gpu(eb); | |
2389 | ||
2390 | for_each_batch_create_order(eb, i) { | |
2391 | if (!eb->requests[i]) | |
2392 | break; | |
2393 | ||
2394 | trace_i915_request_queue(eb->requests[i], eb->batch_flags); | |
2395 | if (!err) | |
2396 | err = eb_request_submit(eb, eb->requests[i], | |
2397 | eb->batches[i]->vma, | |
2398 | eb->batch_len[i]); | |
2399 | } | |
2400 | ||
2401 | return err; | |
2402 | } | |
2403 | ||
1a9c4db4 | 2404 | static int num_vcs_engines(struct drm_i915_private *i915) |
d5b2a3a4 | 2405 | { |
1a9c4db4 | 2406 | return hweight_long(VDBOX_MASK(to_gt(i915))); |
d5b2a3a4 CW |
2407 | } |
2408 | ||
204bcfef | 2409 | /* |
a8ebba75 | 2410 | * Find one BSD ring to dispatch the corresponding BSD command. |
c80ff16e | 2411 | * The engine index is returned. |
a8ebba75 | 2412 | */ |
de1add36 | 2413 | static unsigned int |
c80ff16e CW |
2414 | gen8_dispatch_bsd_engine(struct drm_i915_private *dev_priv, |
2415 | struct drm_file *file) | |
a8ebba75 | 2416 | { |
a8ebba75 ZY |
2417 | struct drm_i915_file_private *file_priv = file->driver_priv; |
2418 | ||
de1add36 | 2419 | /* Check whether the file_priv has already selected one ring. */ |
6f633402 | 2420 | if ((int)file_priv->bsd_engine < 0) |
1a07e86c CW |
2421 | file_priv->bsd_engine = |
2422 | get_random_int() % num_vcs_engines(dev_priv); | |
d23db88c | 2423 | |
c80ff16e | 2424 | return file_priv->bsd_engine; |
d23db88c CW |
2425 | } |
2426 | ||
5e2a0419 | 2427 | static const enum intel_engine_id user_ring_map[] = { |
8a68d464 CW |
2428 | [I915_EXEC_DEFAULT] = RCS0, |
2429 | [I915_EXEC_RENDER] = RCS0, | |
2430 | [I915_EXEC_BLT] = BCS0, | |
2431 | [I915_EXEC_BSD] = VCS0, | |
2432 | [I915_EXEC_VEBOX] = VECS0 | |
de1add36 TU |
2433 | }; |
2434 | ||
2bf541ff | 2435 | static struct i915_request *eb_throttle(struct i915_execbuffer *eb, struct intel_context *ce) |
e5dadff4 CW |
2436 | { |
2437 | struct intel_ring *ring = ce->ring; | |
2438 | struct intel_timeline *tl = ce->timeline; | |
2439 | struct i915_request *rq; | |
2440 | ||
2441 | /* | |
2442 | * Completely unscientific finger-in-the-air estimates for suitable | |
2443 | * maximum user request size (to avoid blocking) and then backoff. | |
2444 | */ | |
2445 | if (intel_ring_update_space(ring) >= PAGE_SIZE) | |
2446 | return NULL; | |
2447 | ||
2448 | /* | |
2449 | * Find a request that after waiting upon, there will be at least half | |
2450 | * the ring available. The hysteresis allows us to compete for the | |
2451 | * shared ring and should mean that we sleep less often prior to | |
2452 | * claiming our resources, but not so long that the ring completely | |
2453 | * drains before we can submit our next request. | |
2454 | */ | |
2455 | list_for_each_entry(rq, &tl->requests, link) { | |
2456 | if (rq->ring != ring) | |
2457 | continue; | |
2458 | ||
2459 | if (__intel_ring_space(rq->postfix, | |
2460 | ring->emit, ring->size) > ring->size / 2) | |
2461 | break; | |
2462 | } | |
2463 | if (&rq->link == &tl->requests) | |
2464 | return NULL; /* weird, we will check again later for real */ | |
2465 | ||
2466 | return i915_request_get(rq); | |
2467 | } | |
2468 | ||
544460c3 MB |
2469 | static int eb_pin_timeline(struct i915_execbuffer *eb, struct intel_context *ce, |
2470 | bool throttle) | |
e5dadff4 CW |
2471 | { |
2472 | struct intel_timeline *tl; | |
2bf541ff | 2473 | struct i915_request *rq = NULL; |
8f2a1057 | 2474 | |
a4e57f90 CW |
2475 | /* |
2476 | * Take a local wakeref for preparing to dispatch the execbuf as | |
2477 | * we expect to access the hardware fairly frequently in the | |
2478 | * process, and require the engine to be kept awake between accesses. | |
2479 | * Upon dispatch, we acquire another prolonged wakeref that we hold | |
2480 | * until the timeline is idle, which in turn releases the wakeref | |
2481 | * taken on the engine, and the parent device. | |
2482 | */ | |
e5dadff4 | 2483 | tl = intel_context_timeline_lock(ce); |
544460c3 MB |
2484 | if (IS_ERR(tl)) |
2485 | return PTR_ERR(tl); | |
a4e57f90 CW |
2486 | |
2487 | intel_context_enter(ce); | |
2bf541ff ML |
2488 | if (throttle) |
2489 | rq = eb_throttle(eb, ce); | |
e5dadff4 CW |
2490 | intel_context_timeline_unlock(tl); |
2491 | ||
544460c3 MB |
2492 | if (rq) { |
2493 | bool nonblock = eb->file->filp->f_flags & O_NONBLOCK; | |
2494 | long timeout = nonblock ? 0 : MAX_SCHEDULE_TIMEOUT; | |
2495 | ||
2496 | if (i915_request_wait(rq, I915_WAIT_INTERRUPTIBLE, | |
2497 | timeout) < 0) { | |
2498 | i915_request_put(rq); | |
2499 | ||
cb935c46 MB |
2500 | /* |
2501 | * Error path, cannot use intel_context_timeline_lock as | |
2502 | * that is user interruptable and this clean up step | |
2503 | * must be done. | |
2504 | */ | |
2505 | mutex_lock(&ce->timeline->mutex); | |
544460c3 | 2506 | intel_context_exit(ce); |
cb935c46 | 2507 | mutex_unlock(&ce->timeline->mutex); |
544460c3 MB |
2508 | |
2509 | if (nonblock) | |
2510 | return -EWOULDBLOCK; | |
2511 | else | |
2512 | return -EINTR; | |
2513 | } | |
2514 | i915_request_put(rq); | |
2515 | } | |
2516 | ||
2517 | return 0; | |
2518 | } | |
2519 | ||
2520 | static int eb_pin_engine(struct i915_execbuffer *eb, bool throttle) | |
2521 | { | |
2522 | struct intel_context *ce = eb->context, *child; | |
2523 | int err; | |
2524 | int i = 0, j = 0; | |
2525 | ||
2526 | GEM_BUG_ON(eb->args->flags & __EXEC_ENGINE_PINNED); | |
2527 | ||
2528 | if (unlikely(intel_context_is_banned(ce))) | |
2529 | return -EIO; | |
2530 | ||
2531 | /* | |
2532 | * Pinning the contexts may generate requests in order to acquire | |
2533 | * GGTT space, so do this first before we reserve a seqno for | |
2534 | * ourselves. | |
2535 | */ | |
2536 | err = intel_context_pin_ww(ce, &eb->ww); | |
2537 | if (err) | |
2538 | return err; | |
2539 | for_each_child(ce, child) { | |
2540 | err = intel_context_pin_ww(child, &eb->ww); | |
2541 | GEM_BUG_ON(err); /* perma-pinned should incr a counter */ | |
2542 | } | |
2543 | ||
2544 | for_each_child(ce, child) { | |
2545 | err = eb_pin_timeline(eb, child, throttle); | |
2546 | if (err) | |
2547 | goto unwind; | |
2548 | ++i; | |
2549 | } | |
2550 | err = eb_pin_timeline(eb, ce, throttle); | |
2551 | if (err) | |
2552 | goto unwind; | |
2553 | ||
2bf541ff | 2554 | eb->args->flags |= __EXEC_ENGINE_PINNED; |
544460c3 MB |
2555 | return 0; |
2556 | ||
2557 | unwind: | |
2558 | for_each_child(ce, child) { | |
2559 | if (j++ < i) { | |
2560 | mutex_lock(&child->timeline->mutex); | |
2561 | intel_context_exit(child); | |
2562 | mutex_unlock(&child->timeline->mutex); | |
2563 | } | |
2564 | } | |
2565 | for_each_child(ce, child) | |
2566 | intel_context_unpin(child); | |
2567 | intel_context_unpin(ce); | |
2568 | return err; | |
8f2a1057 CW |
2569 | } |
2570 | ||
e5dadff4 | 2571 | static void eb_unpin_engine(struct i915_execbuffer *eb) |
8f2a1057 | 2572 | { |
544460c3 | 2573 | struct intel_context *ce = eb->context, *child; |
a4e57f90 | 2574 | |
2bf541ff ML |
2575 | if (!(eb->args->flags & __EXEC_ENGINE_PINNED)) |
2576 | return; | |
2577 | ||
2578 | eb->args->flags &= ~__EXEC_ENGINE_PINNED; | |
2579 | ||
544460c3 MB |
2580 | for_each_child(ce, child) { |
2581 | mutex_lock(&child->timeline->mutex); | |
2582 | intel_context_exit(child); | |
2583 | mutex_unlock(&child->timeline->mutex); | |
2584 | ||
2585 | intel_context_unpin(child); | |
2586 | } | |
2587 | ||
2588 | mutex_lock(&ce->timeline->mutex); | |
a4e57f90 | 2589 | intel_context_exit(ce); |
544460c3 | 2590 | mutex_unlock(&ce->timeline->mutex); |
a4e57f90 | 2591 | |
2850748e | 2592 | intel_context_unpin(ce); |
8f2a1057 | 2593 | } |
de1add36 | 2594 | |
5e2a0419 | 2595 | static unsigned int |
b49a7d51 | 2596 | eb_select_legacy_ring(struct i915_execbuffer *eb) |
de1add36 | 2597 | { |
8f2a1057 | 2598 | struct drm_i915_private *i915 = eb->i915; |
b49a7d51 | 2599 | struct drm_i915_gem_execbuffer2 *args = eb->args; |
de1add36 | 2600 | unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK; |
de1add36 | 2601 | |
5e2a0419 CW |
2602 | if (user_ring_id != I915_EXEC_BSD && |
2603 | (args->flags & I915_EXEC_BSD_MASK)) { | |
baa89ba3 WK |
2604 | drm_dbg(&i915->drm, |
2605 | "execbuf with non bsd ring but with invalid " | |
2606 | "bsd dispatch flags: %d\n", (int)(args->flags)); | |
5e2a0419 | 2607 | return -1; |
de1add36 TU |
2608 | } |
2609 | ||
d5b2a3a4 | 2610 | if (user_ring_id == I915_EXEC_BSD && num_vcs_engines(i915) > 1) { |
de1add36 TU |
2611 | unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK; |
2612 | ||
2613 | if (bsd_idx == I915_EXEC_BSD_DEFAULT) { | |
b49a7d51 | 2614 | bsd_idx = gen8_dispatch_bsd_engine(i915, eb->file); |
de1add36 TU |
2615 | } else if (bsd_idx >= I915_EXEC_BSD_RING1 && |
2616 | bsd_idx <= I915_EXEC_BSD_RING2) { | |
d9da6aa0 | 2617 | bsd_idx >>= I915_EXEC_BSD_SHIFT; |
de1add36 TU |
2618 | bsd_idx--; |
2619 | } else { | |
baa89ba3 WK |
2620 | drm_dbg(&i915->drm, |
2621 | "execbuf with unknown bsd ring: %u\n", | |
2622 | bsd_idx); | |
5e2a0419 | 2623 | return -1; |
de1add36 TU |
2624 | } |
2625 | ||
5e2a0419 | 2626 | return _VCS(bsd_idx); |
de1add36 TU |
2627 | } |
2628 | ||
5e2a0419 | 2629 | if (user_ring_id >= ARRAY_SIZE(user_ring_map)) { |
baa89ba3 WK |
2630 | drm_dbg(&i915->drm, "execbuf with unknown ring: %u\n", |
2631 | user_ring_id); | |
5e2a0419 | 2632 | return -1; |
de1add36 TU |
2633 | } |
2634 | ||
5e2a0419 CW |
2635 | return user_ring_map[user_ring_id]; |
2636 | } | |
2637 | ||
2638 | static int | |
2bf541ff | 2639 | eb_select_engine(struct i915_execbuffer *eb) |
5e2a0419 | 2640 | { |
544460c3 | 2641 | struct intel_context *ce, *child; |
5e2a0419 CW |
2642 | unsigned int idx; |
2643 | int err; | |
2644 | ||
976b55f0 | 2645 | if (i915_gem_context_user_engines(eb->gem_context)) |
b49a7d51 | 2646 | idx = eb->args->flags & I915_EXEC_RING_MASK; |
976b55f0 | 2647 | else |
b49a7d51 | 2648 | idx = eb_select_legacy_ring(eb); |
5e2a0419 CW |
2649 | |
2650 | ce = i915_gem_context_get_engine(eb->gem_context, idx); | |
2651 | if (IS_ERR(ce)) | |
2652 | return PTR_ERR(ce); | |
2653 | ||
544460c3 MB |
2654 | if (intel_context_is_parallel(ce)) { |
2655 | if (eb->buffer_count < ce->parallel.number_children + 1) { | |
2656 | intel_context_put(ce); | |
2657 | return -EINVAL; | |
2658 | } | |
2659 | if (eb->batch_start_offset || eb->args->batch_len) { | |
2660 | intel_context_put(ce); | |
2661 | return -EINVAL; | |
2662 | } | |
2663 | } | |
2664 | eb->num_batches = ce->parallel.number_children + 1; | |
2665 | ||
2666 | for_each_child(ce, child) | |
2667 | intel_context_get(child); | |
2bf541ff | 2668 | intel_gt_pm_get(ce->engine->gt); |
5e2a0419 | 2669 | |
2bf541ff ML |
2670 | if (!test_bit(CONTEXT_ALLOC_BIT, &ce->flags)) { |
2671 | err = intel_context_alloc_state(ce); | |
2672 | if (err) | |
2673 | goto err; | |
2674 | } | |
544460c3 MB |
2675 | for_each_child(ce, child) { |
2676 | if (!test_bit(CONTEXT_ALLOC_BIT, &child->flags)) { | |
2677 | err = intel_context_alloc_state(child); | |
2678 | if (err) | |
2679 | goto err; | |
2680 | } | |
2681 | } | |
2bf541ff ML |
2682 | |
2683 | /* | |
2684 | * ABI: Before userspace accesses the GPU (e.g. execbuffer), report | |
2685 | * EIO if the GPU is already wedged. | |
2686 | */ | |
2687 | err = intel_gt_terminally_wedged(ce->engine->gt); | |
2688 | if (err) | |
2689 | goto err; | |
2690 | ||
e1a7ab4f TH |
2691 | if (!i915_vm_tryget(ce->vm)) { |
2692 | err = -ENOENT; | |
2693 | goto err; | |
2694 | } | |
2695 | ||
2bf541ff | 2696 | eb->context = ce; |
544460c3 | 2697 | eb->gt = ce->engine->gt; |
2bf541ff ML |
2698 | |
2699 | /* | |
2700 | * Make sure engine pool stays alive even if we call intel_context_put | |
2701 | * during ww handling. The pool is destroyed when last pm reference | |
2702 | * is dropped, which breaks our -EDEADLK handling. | |
2703 | */ | |
2704 | return err; | |
2705 | ||
2706 | err: | |
2707 | intel_gt_pm_put(ce->engine->gt); | |
544460c3 MB |
2708 | for_each_child(ce, child) |
2709 | intel_context_put(child); | |
2bf541ff | 2710 | intel_context_put(ce); |
5e2a0419 | 2711 | return err; |
de1add36 TU |
2712 | } |
2713 | ||
2bf541ff ML |
2714 | static void |
2715 | eb_put_engine(struct i915_execbuffer *eb) | |
2716 | { | |
544460c3 MB |
2717 | struct intel_context *child; |
2718 | ||
e1a7ab4f | 2719 | i915_vm_put(eb->context->vm); |
544460c3 MB |
2720 | intel_gt_pm_put(eb->gt); |
2721 | for_each_child(eb->context, child) | |
2722 | intel_context_put(child); | |
2bf541ff ML |
2723 | intel_context_put(eb->context); |
2724 | } | |
2725 | ||
cf6e7bac | 2726 | static void |
13149e8b | 2727 | __free_fence_array(struct eb_fence *fences, unsigned int n) |
cf6e7bac | 2728 | { |
13149e8b | 2729 | while (n--) { |
cda9edd0 | 2730 | drm_syncobj_put(ptr_mask_bits(fences[n].syncobj, 2)); |
13149e8b | 2731 | dma_fence_put(fences[n].dma_fence); |
440d0f12 | 2732 | dma_fence_chain_free(fences[n].chain_fence); |
13149e8b | 2733 | } |
cf6e7bac JE |
2734 | kvfree(fences); |
2735 | } | |
2736 | ||
cda9edd0 | 2737 | static int |
13149e8b LL |
2738 | add_timeline_fence_array(struct i915_execbuffer *eb, |
2739 | const struct drm_i915_gem_execbuffer_ext_timeline_fences *timeline_fences) | |
cf6e7bac | 2740 | { |
13149e8b LL |
2741 | struct drm_i915_gem_exec_fence __user *user_fences; |
2742 | u64 __user *user_values; | |
2743 | struct eb_fence *f; | |
2744 | u64 nfences; | |
2745 | int err = 0; | |
cf6e7bac | 2746 | |
13149e8b LL |
2747 | nfences = timeline_fences->fence_count; |
2748 | if (!nfences) | |
cda9edd0 | 2749 | return 0; |
cf6e7bac | 2750 | |
d710fc16 CW |
2751 | /* Check multiplication overflow for access_ok() and kvmalloc_array() */ |
2752 | BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long)); | |
2753 | if (nfences > min_t(unsigned long, | |
13149e8b LL |
2754 | ULONG_MAX / sizeof(*user_fences), |
2755 | SIZE_MAX / sizeof(*f)) - eb->num_fences) | |
cda9edd0 | 2756 | return -EINVAL; |
cf6e7bac | 2757 | |
13149e8b LL |
2758 | user_fences = u64_to_user_ptr(timeline_fences->handles_ptr); |
2759 | if (!access_ok(user_fences, nfences * sizeof(*user_fences))) | |
2760 | return -EFAULT; | |
2761 | ||
2762 | user_values = u64_to_user_ptr(timeline_fences->values_ptr); | |
2763 | if (!access_ok(user_values, nfences * sizeof(*user_values))) | |
cda9edd0 | 2764 | return -EFAULT; |
cf6e7bac | 2765 | |
13149e8b LL |
2766 | f = krealloc(eb->fences, |
2767 | (eb->num_fences + nfences) * sizeof(*f), | |
2768 | __GFP_NOWARN | GFP_KERNEL); | |
2769 | if (!f) | |
cda9edd0 | 2770 | return -ENOMEM; |
cf6e7bac | 2771 | |
13149e8b LL |
2772 | eb->fences = f; |
2773 | f += eb->num_fences; | |
2774 | ||
2775 | BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) & | |
2776 | ~__I915_EXEC_FENCE_UNKNOWN_FLAGS); | |
2777 | ||
2778 | while (nfences--) { | |
2779 | struct drm_i915_gem_exec_fence user_fence; | |
cf6e7bac | 2780 | struct drm_syncobj *syncobj; |
13149e8b LL |
2781 | struct dma_fence *fence = NULL; |
2782 | u64 point; | |
2783 | ||
2784 | if (__copy_from_user(&user_fence, | |
2785 | user_fences++, | |
2786 | sizeof(user_fence))) | |
2787 | return -EFAULT; | |
2788 | ||
2789 | if (user_fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS) | |
2790 | return -EINVAL; | |
2791 | ||
2792 | if (__get_user(point, user_values++)) | |
2793 | return -EFAULT; | |
2794 | ||
2795 | syncobj = drm_syncobj_find(eb->file, user_fence.handle); | |
2796 | if (!syncobj) { | |
2797 | DRM_DEBUG("Invalid syncobj handle provided\n"); | |
2798 | return -ENOENT; | |
2799 | } | |
2800 | ||
2801 | fence = drm_syncobj_fence_get(syncobj); | |
cf6e7bac | 2802 | |
13149e8b LL |
2803 | if (!fence && user_fence.flags && |
2804 | !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) { | |
2805 | DRM_DEBUG("Syncobj handle has no fence\n"); | |
2806 | drm_syncobj_put(syncobj); | |
2807 | return -EINVAL; | |
cf6e7bac JE |
2808 | } |
2809 | ||
13149e8b LL |
2810 | if (fence) |
2811 | err = dma_fence_chain_find_seqno(&fence, point); | |
2812 | ||
2813 | if (err && !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) { | |
2814 | DRM_DEBUG("Syncobj handle missing requested point %llu\n", point); | |
da1ea128 | 2815 | dma_fence_put(fence); |
13149e8b LL |
2816 | drm_syncobj_put(syncobj); |
2817 | return err; | |
2818 | } | |
2819 | ||
2820 | /* | |
2821 | * A point might have been signaled already and | |
2822 | * garbage collected from the timeline. In this case | |
2823 | * just ignore the point and carry on. | |
2824 | */ | |
2825 | if (!fence && !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) { | |
2826 | drm_syncobj_put(syncobj); | |
2827 | continue; | |
2828 | } | |
2829 | ||
2830 | /* | |
2831 | * For timeline syncobjs we need to preallocate chains for | |
2832 | * later signaling. | |
2833 | */ | |
2834 | if (point != 0 && user_fence.flags & I915_EXEC_FENCE_SIGNAL) { | |
2835 | /* | |
2836 | * Waiting and signaling the same point (when point != | |
2837 | * 0) would break the timeline. | |
2838 | */ | |
2839 | if (user_fence.flags & I915_EXEC_FENCE_WAIT) { | |
2840 | DRM_DEBUG("Trying to wait & signal the same timeline point.\n"); | |
2841 | dma_fence_put(fence); | |
2842 | drm_syncobj_put(syncobj); | |
2843 | return -EINVAL; | |
2844 | } | |
2845 | ||
440d0f12 | 2846 | f->chain_fence = dma_fence_chain_alloc(); |
13149e8b LL |
2847 | if (!f->chain_fence) { |
2848 | drm_syncobj_put(syncobj); | |
2849 | dma_fence_put(fence); | |
2850 | return -ENOMEM; | |
2851 | } | |
2852 | } else { | |
2853 | f->chain_fence = NULL; | |
ebcaa1ff TU |
2854 | } |
2855 | ||
13149e8b LL |
2856 | f->syncobj = ptr_pack_bits(syncobj, user_fence.flags, 2); |
2857 | f->dma_fence = fence; | |
2858 | f->value = point; | |
2859 | f++; | |
2860 | eb->num_fences++; | |
2861 | } | |
2862 | ||
2863 | return 0; | |
2864 | } | |
2865 | ||
2866 | static int add_fence_array(struct i915_execbuffer *eb) | |
2867 | { | |
2868 | struct drm_i915_gem_execbuffer2 *args = eb->args; | |
2869 | struct drm_i915_gem_exec_fence __user *user; | |
2870 | unsigned long num_fences = args->num_cliprects; | |
2871 | struct eb_fence *f; | |
2872 | ||
2873 | if (!(args->flags & I915_EXEC_FENCE_ARRAY)) | |
2874 | return 0; | |
2875 | ||
2876 | if (!num_fences) | |
2877 | return 0; | |
2878 | ||
2879 | /* Check multiplication overflow for access_ok() and kvmalloc_array() */ | |
2880 | BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long)); | |
2881 | if (num_fences > min_t(unsigned long, | |
2882 | ULONG_MAX / sizeof(*user), | |
2883 | SIZE_MAX / sizeof(*f) - eb->num_fences)) | |
2884 | return -EINVAL; | |
2885 | ||
2886 | user = u64_to_user_ptr(args->cliprects_ptr); | |
2887 | if (!access_ok(user, num_fences * sizeof(*user))) | |
2888 | return -EFAULT; | |
2889 | ||
2890 | f = krealloc(eb->fences, | |
2891 | (eb->num_fences + num_fences) * sizeof(*f), | |
2892 | __GFP_NOWARN | GFP_KERNEL); | |
2893 | if (!f) | |
2894 | return -ENOMEM; | |
2895 | ||
2896 | eb->fences = f; | |
2897 | f += eb->num_fences; | |
2898 | while (num_fences--) { | |
2899 | struct drm_i915_gem_exec_fence user_fence; | |
2900 | struct drm_syncobj *syncobj; | |
2901 | struct dma_fence *fence = NULL; | |
2902 | ||
2903 | if (__copy_from_user(&user_fence, user++, sizeof(user_fence))) | |
2904 | return -EFAULT; | |
2905 | ||
2906 | if (user_fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS) | |
2907 | return -EINVAL; | |
2908 | ||
2909 | syncobj = drm_syncobj_find(eb->file, user_fence.handle); | |
cf6e7bac JE |
2910 | if (!syncobj) { |
2911 | DRM_DEBUG("Invalid syncobj handle provided\n"); | |
13149e8b LL |
2912 | return -ENOENT; |
2913 | } | |
2914 | ||
2915 | if (user_fence.flags & I915_EXEC_FENCE_WAIT) { | |
2916 | fence = drm_syncobj_fence_get(syncobj); | |
2917 | if (!fence) { | |
2918 | DRM_DEBUG("Syncobj handle has no fence\n"); | |
2919 | drm_syncobj_put(syncobj); | |
2920 | return -EINVAL; | |
2921 | } | |
cf6e7bac JE |
2922 | } |
2923 | ||
ebcaa1ff TU |
2924 | BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) & |
2925 | ~__I915_EXEC_FENCE_UNKNOWN_FLAGS); | |
2926 | ||
13149e8b LL |
2927 | f->syncobj = ptr_pack_bits(syncobj, user_fence.flags, 2); |
2928 | f->dma_fence = fence; | |
2929 | f->value = 0; | |
2930 | f->chain_fence = NULL; | |
2931 | f++; | |
2932 | eb->num_fences++; | |
cf6e7bac JE |
2933 | } |
2934 | ||
cda9edd0 | 2935 | return 0; |
13149e8b | 2936 | } |
cf6e7bac | 2937 | |
13149e8b LL |
2938 | static void put_fence_array(struct eb_fence *fences, int num_fences) |
2939 | { | |
2940 | if (fences) | |
2941 | __free_fence_array(fences, num_fences); | |
cf6e7bac JE |
2942 | } |
2943 | ||
2944 | static int | |
544460c3 MB |
2945 | await_fence_array(struct i915_execbuffer *eb, |
2946 | struct i915_request *rq) | |
cf6e7bac | 2947 | { |
cf6e7bac JE |
2948 | unsigned int n; |
2949 | int err; | |
2950 | ||
13149e8b | 2951 | for (n = 0; n < eb->num_fences; n++) { |
cf6e7bac | 2952 | struct drm_syncobj *syncobj; |
cf6e7bac JE |
2953 | unsigned int flags; |
2954 | ||
cda9edd0 | 2955 | syncobj = ptr_unpack_bits(eb->fences[n].syncobj, &flags, 2); |
cf6e7bac | 2956 | |
13149e8b LL |
2957 | if (!eb->fences[n].dma_fence) |
2958 | continue; | |
cf6e7bac | 2959 | |
544460c3 | 2960 | err = i915_request_await_dma_fence(rq, eb->fences[n].dma_fence); |
cf6e7bac JE |
2961 | if (err < 0) |
2962 | return err; | |
2963 | } | |
2964 | ||
2965 | return 0; | |
2966 | } | |
2967 | ||
544460c3 MB |
2968 | static void signal_fence_array(const struct i915_execbuffer *eb, |
2969 | struct dma_fence * const fence) | |
cf6e7bac | 2970 | { |
cf6e7bac JE |
2971 | unsigned int n; |
2972 | ||
13149e8b | 2973 | for (n = 0; n < eb->num_fences; n++) { |
cf6e7bac JE |
2974 | struct drm_syncobj *syncobj; |
2975 | unsigned int flags; | |
2976 | ||
cda9edd0 | 2977 | syncobj = ptr_unpack_bits(eb->fences[n].syncobj, &flags, 2); |
cf6e7bac JE |
2978 | if (!(flags & I915_EXEC_FENCE_SIGNAL)) |
2979 | continue; | |
2980 | ||
13149e8b LL |
2981 | if (eb->fences[n].chain_fence) { |
2982 | drm_syncobj_add_point(syncobj, | |
2983 | eb->fences[n].chain_fence, | |
2984 | fence, | |
2985 | eb->fences[n].value); | |
2986 | /* | |
2987 | * The chain's ownership is transferred to the | |
2988 | * timeline. | |
2989 | */ | |
2990 | eb->fences[n].chain_fence = NULL; | |
2991 | } else { | |
2992 | drm_syncobj_replace_fence(syncobj, fence); | |
2993 | } | |
cf6e7bac JE |
2994 | } |
2995 | } | |
2996 | ||
13149e8b LL |
2997 | static int |
2998 | parse_timeline_fences(struct i915_user_extension __user *ext, void *data) | |
2999 | { | |
3000 | struct i915_execbuffer *eb = data; | |
3001 | struct drm_i915_gem_execbuffer_ext_timeline_fences timeline_fences; | |
3002 | ||
3003 | if (copy_from_user(&timeline_fences, ext, sizeof(timeline_fences))) | |
3004 | return -EFAULT; | |
3005 | ||
3006 | return add_timeline_fence_array(eb, &timeline_fences); | |
3007 | } | |
3008 | ||
61231f6b CW |
3009 | static void retire_requests(struct intel_timeline *tl, struct i915_request *end) |
3010 | { | |
3011 | struct i915_request *rq, *rn; | |
3012 | ||
3013 | list_for_each_entry_safe(rq, rn, &tl->requests, link) | |
3014 | if (rq == end || !i915_request_retire(rq)) | |
3015 | break; | |
3016 | } | |
3017 | ||
544460c3 MB |
3018 | static int eb_request_add(struct i915_execbuffer *eb, struct i915_request *rq, |
3019 | int err, bool last_parallel) | |
61231f6b | 3020 | { |
61231f6b CW |
3021 | struct intel_timeline * const tl = i915_request_timeline(rq); |
3022 | struct i915_sched_attr attr = {}; | |
3023 | struct i915_request *prev; | |
3024 | ||
3025 | lockdep_assert_held(&tl->mutex); | |
3026 | lockdep_unpin_lock(&tl->mutex, rq->cookie); | |
3027 | ||
3028 | trace_i915_request_add(rq); | |
3029 | ||
3030 | prev = __i915_request_commit(rq); | |
3031 | ||
3032 | /* Check that the context wasn't destroyed before submission */ | |
207e4a71 | 3033 | if (likely(!intel_context_is_closed(eb->context))) { |
61231f6b | 3034 | attr = eb->gem_context->sched; |
61231f6b CW |
3035 | } else { |
3036 | /* Serialise with context_close via the add_to_timeline */ | |
36e191f0 CW |
3037 | i915_request_set_error_once(rq, -ENOENT); |
3038 | __i915_request_skip(rq); | |
ba38b79e | 3039 | err = -ENOENT; /* override any transient errors */ |
61231f6b CW |
3040 | } |
3041 | ||
544460c3 MB |
3042 | if (intel_context_is_parallel(eb->context)) { |
3043 | if (err) { | |
3044 | __i915_request_skip(rq); | |
3045 | set_bit(I915_FENCE_FLAG_SKIP_PARALLEL, | |
3046 | &rq->fence.flags); | |
3047 | } | |
3048 | if (last_parallel) | |
3049 | set_bit(I915_FENCE_FLAG_SUBMIT_PARALLEL, | |
3050 | &rq->fence.flags); | |
3051 | } | |
3052 | ||
61231f6b | 3053 | __i915_request_queue(rq, &attr); |
61231f6b CW |
3054 | |
3055 | /* Try to clean up the client's timeline after submitting the request */ | |
3056 | if (prev) | |
3057 | retire_requests(tl, prev); | |
3058 | ||
3059 | mutex_unlock(&tl->mutex); | |
ba38b79e CW |
3060 | |
3061 | return err; | |
61231f6b CW |
3062 | } |
3063 | ||
544460c3 MB |
3064 | static int eb_requests_add(struct i915_execbuffer *eb, int err) |
3065 | { | |
3066 | int i; | |
3067 | ||
3068 | /* | |
3069 | * We iterate in reverse order of creation to release timeline mutexes in | |
3070 | * same order. | |
3071 | */ | |
3072 | for_each_batch_add_order(eb, i) { | |
3073 | struct i915_request *rq = eb->requests[i]; | |
3074 | ||
3075 | if (!rq) | |
3076 | continue; | |
3077 | err |= eb_request_add(eb, rq, err, i == 0); | |
3078 | } | |
3079 | ||
3080 | return err; | |
3081 | } | |
3082 | ||
cda9edd0 | 3083 | static const i915_user_extension_fn execbuf_extensions[] = { |
13149e8b | 3084 | [DRM_I915_GEM_EXECBUFFER_EXT_TIMELINE_FENCES] = parse_timeline_fences, |
cda9edd0 LL |
3085 | }; |
3086 | ||
3087 | static int | |
3088 | parse_execbuf2_extensions(struct drm_i915_gem_execbuffer2 *args, | |
3089 | struct i915_execbuffer *eb) | |
3090 | { | |
cda9edd0 LL |
3091 | if (!(args->flags & I915_EXEC_USE_EXTENSIONS)) |
3092 | return 0; | |
3093 | ||
3094 | /* The execbuf2 extension mechanism reuses cliprects_ptr. So we cannot | |
3095 | * have another flag also using it at the same time. | |
3096 | */ | |
3097 | if (eb->args->flags & I915_EXEC_FENCE_ARRAY) | |
3098 | return -EINVAL; | |
3099 | ||
3100 | if (args->num_cliprects != 0) | |
3101 | return -EINVAL; | |
3102 | ||
3103 | return i915_user_extensions(u64_to_user_ptr(args->cliprects_ptr), | |
3104 | execbuf_extensions, | |
3105 | ARRAY_SIZE(execbuf_extensions), | |
3106 | eb); | |
3107 | } | |
3108 | ||
544460c3 MB |
3109 | static void eb_requests_get(struct i915_execbuffer *eb) |
3110 | { | |
3111 | unsigned int i; | |
3112 | ||
3113 | for_each_batch_create_order(eb, i) { | |
3114 | if (!eb->requests[i]) | |
3115 | break; | |
3116 | ||
3117 | i915_request_get(eb->requests[i]); | |
3118 | } | |
3119 | } | |
3120 | ||
3121 | static void eb_requests_put(struct i915_execbuffer *eb) | |
3122 | { | |
3123 | unsigned int i; | |
3124 | ||
3125 | for_each_batch_create_order(eb, i) { | |
3126 | if (!eb->requests[i]) | |
3127 | break; | |
3128 | ||
3129 | i915_request_put(eb->requests[i]); | |
3130 | } | |
3131 | } | |
3132 | ||
3133 | static struct sync_file * | |
3134 | eb_composite_fence_create(struct i915_execbuffer *eb, int out_fence_fd) | |
3135 | { | |
3136 | struct sync_file *out_fence = NULL; | |
3137 | struct dma_fence_array *fence_array; | |
3138 | struct dma_fence **fences; | |
3139 | unsigned int i; | |
3140 | ||
3141 | GEM_BUG_ON(!intel_context_is_parent(eb->context)); | |
3142 | ||
3143 | fences = kmalloc_array(eb->num_batches, sizeof(*fences), GFP_KERNEL); | |
3144 | if (!fences) | |
3145 | return ERR_PTR(-ENOMEM); | |
3146 | ||
7647f009 | 3147 | for_each_batch_create_order(eb, i) { |
544460c3 | 3148 | fences[i] = &eb->requests[i]->fence; |
7647f009 MB |
3149 | __set_bit(I915_FENCE_FLAG_COMPOSITE, |
3150 | &eb->requests[i]->fence.flags); | |
3151 | } | |
544460c3 MB |
3152 | |
3153 | fence_array = dma_fence_array_create(eb->num_batches, | |
3154 | fences, | |
3155 | eb->context->parallel.fence_context, | |
d46f329a | 3156 | eb->context->parallel.seqno++, |
544460c3 MB |
3157 | false); |
3158 | if (!fence_array) { | |
3159 | kfree(fences); | |
3160 | return ERR_PTR(-ENOMEM); | |
3161 | } | |
3162 | ||
3163 | /* Move ownership to the dma_fence_array created above */ | |
3164 | for_each_batch_create_order(eb, i) | |
3165 | dma_fence_get(fences[i]); | |
3166 | ||
3167 | if (out_fence_fd != -1) { | |
3168 | out_fence = sync_file_create(&fence_array->base); | |
3169 | /* sync_file now owns fence_arry, drop creation ref */ | |
3170 | dma_fence_put(&fence_array->base); | |
3171 | if (!out_fence) | |
3172 | return ERR_PTR(-ENOMEM); | |
3173 | } | |
3174 | ||
3175 | eb->composite_fence = &fence_array->base; | |
3176 | ||
3177 | return out_fence; | |
3178 | } | |
3179 | ||
3180 | static struct sync_file * | |
3181 | eb_fences_add(struct i915_execbuffer *eb, struct i915_request *rq, | |
3182 | struct dma_fence *in_fence, int out_fence_fd) | |
3183 | { | |
3184 | struct sync_file *out_fence = NULL; | |
3185 | int err; | |
3186 | ||
3187 | if (unlikely(eb->gem_context->syncobj)) { | |
3188 | struct dma_fence *fence; | |
3189 | ||
3190 | fence = drm_syncobj_fence_get(eb->gem_context->syncobj); | |
3191 | err = i915_request_await_dma_fence(rq, fence); | |
3192 | dma_fence_put(fence); | |
3193 | if (err) | |
3194 | return ERR_PTR(err); | |
3195 | } | |
3196 | ||
3197 | if (in_fence) { | |
3198 | if (eb->args->flags & I915_EXEC_FENCE_SUBMIT) | |
3199 | err = i915_request_await_execution(rq, in_fence); | |
3200 | else | |
3201 | err = i915_request_await_dma_fence(rq, in_fence); | |
3202 | if (err < 0) | |
3203 | return ERR_PTR(err); | |
3204 | } | |
3205 | ||
3206 | if (eb->fences) { | |
3207 | err = await_fence_array(eb, rq); | |
3208 | if (err) | |
3209 | return ERR_PTR(err); | |
3210 | } | |
3211 | ||
3212 | if (intel_context_is_parallel(eb->context)) { | |
3213 | out_fence = eb_composite_fence_create(eb, out_fence_fd); | |
3214 | if (IS_ERR(out_fence)) | |
3215 | return ERR_PTR(-ENOMEM); | |
3216 | } else if (out_fence_fd != -1) { | |
3217 | out_fence = sync_file_create(&rq->fence); | |
3218 | if (!out_fence) | |
3219 | return ERR_PTR(-ENOMEM); | |
3220 | } | |
3221 | ||
3222 | return out_fence; | |
3223 | } | |
3224 | ||
3225 | static struct intel_context * | |
3226 | eb_find_context(struct i915_execbuffer *eb, unsigned int context_number) | |
3227 | { | |
3228 | struct intel_context *child; | |
3229 | ||
3230 | if (likely(context_number == 0)) | |
3231 | return eb->context; | |
3232 | ||
3233 | for_each_child(eb->context, child) | |
3234 | if (!--context_number) | |
3235 | return child; | |
3236 | ||
3237 | GEM_BUG_ON("Context not found"); | |
3238 | ||
3239 | return NULL; | |
3240 | } | |
3241 | ||
3242 | static struct sync_file * | |
3243 | eb_requests_create(struct i915_execbuffer *eb, struct dma_fence *in_fence, | |
3244 | int out_fence_fd) | |
3245 | { | |
3246 | struct sync_file *out_fence = NULL; | |
3247 | unsigned int i; | |
3248 | ||
3249 | for_each_batch_create_order(eb, i) { | |
3250 | /* Allocate a request for this batch buffer nice and early. */ | |
3251 | eb->requests[i] = i915_request_create(eb_find_context(eb, i)); | |
3252 | if (IS_ERR(eb->requests[i])) { | |
86752bd6 | 3253 | out_fence = ERR_CAST(eb->requests[i]); |
544460c3 MB |
3254 | eb->requests[i] = NULL; |
3255 | return out_fence; | |
3256 | } | |
3257 | ||
3258 | /* | |
3259 | * Only the first request added (committed to backend) has to | |
3260 | * take the in fences into account as all subsequent requests | |
3261 | * will have fences inserted inbetween them. | |
3262 | */ | |
3263 | if (i + 1 == eb->num_batches) { | |
3264 | out_fence = eb_fences_add(eb, eb->requests[i], | |
3265 | in_fence, out_fence_fd); | |
3266 | if (IS_ERR(out_fence)) | |
3267 | return out_fence; | |
3268 | } | |
3269 | ||
3270 | /* | |
ff20afc4 TH |
3271 | * Not really on stack, but we don't want to call |
3272 | * kfree on the batch_snapshot when we put it, so use the | |
3273 | * _onstack interface. | |
544460c3 | 3274 | */ |
ff20afc4 | 3275 | if (eb->batches[i]->vma) |
60dc43d1 TH |
3276 | eb->requests[i]->batch_res = |
3277 | i915_vma_resource_get(eb->batches[i]->vma->resource); | |
544460c3 MB |
3278 | if (eb->batch_pool) { |
3279 | GEM_BUG_ON(intel_context_is_parallel(eb->context)); | |
3280 | intel_gt_buffer_pool_mark_active(eb->batch_pool, | |
3281 | eb->requests[i]); | |
3282 | } | |
3283 | } | |
3284 | ||
3285 | return out_fence; | |
3286 | } | |
3287 | ||
54cf91dc | 3288 | static int |
650bc635 | 3289 | i915_gem_do_execbuffer(struct drm_device *dev, |
54cf91dc CW |
3290 | struct drm_file *file, |
3291 | struct drm_i915_gem_execbuffer2 *args, | |
cda9edd0 | 3292 | struct drm_i915_gem_exec_object2 *exec) |
54cf91dc | 3293 | { |
44157641 | 3294 | struct drm_i915_private *i915 = to_i915(dev); |
650bc635 | 3295 | struct i915_execbuffer eb; |
fec0445c CW |
3296 | struct dma_fence *in_fence = NULL; |
3297 | struct sync_file *out_fence = NULL; | |
3298 | int out_fence_fd = -1; | |
2889caa9 | 3299 | int err; |
432e58ed | 3300 | |
74c1c694 | 3301 | BUILD_BUG_ON(__EXEC_INTERNAL_FLAGS & ~__I915_EXEC_ILLEGAL_FLAGS); |
2889caa9 CW |
3302 | BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS & |
3303 | ~__EXEC_OBJECT_UNKNOWN_FLAGS); | |
54cf91dc | 3304 | |
44157641 | 3305 | eb.i915 = i915; |
650bc635 CW |
3306 | eb.file = file; |
3307 | eb.args = args; | |
ad5d95e4 | 3308 | if (DBG_FORCE_RELOC || !(args->flags & I915_EXEC_NO_RELOC)) |
2889caa9 | 3309 | args->flags |= __EXEC_HAS_RELOC; |
c7c6e46f | 3310 | |
650bc635 | 3311 | eb.exec = exec; |
8ae275c2 ML |
3312 | eb.vma = (struct eb_vma *)(exec + args->buffer_count + 1); |
3313 | eb.vma[0].vma = NULL; | |
8e02cceb | 3314 | eb.batch_pool = NULL; |
c7c6e46f | 3315 | |
2889caa9 | 3316 | eb.invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS; |
650bc635 CW |
3317 | reloc_cache_init(&eb.reloc_cache, eb.i915); |
3318 | ||
2889caa9 | 3319 | eb.buffer_count = args->buffer_count; |
650bc635 | 3320 | eb.batch_start_offset = args->batch_start_offset; |
32d94048 | 3321 | eb.trampoline = NULL; |
650bc635 | 3322 | |
cda9edd0 | 3323 | eb.fences = NULL; |
13149e8b | 3324 | eb.num_fences = 0; |
cda9edd0 | 3325 | |
ff20afc4 TH |
3326 | eb_capture_list_clear(&eb); |
3327 | ||
544460c3 MB |
3328 | memset(eb.requests, 0, sizeof(struct i915_request *) * |
3329 | ARRAY_SIZE(eb.requests)); | |
3330 | eb.composite_fence = NULL; | |
3331 | ||
2889caa9 | 3332 | eb.batch_flags = 0; |
d7d4eedd | 3333 | if (args->flags & I915_EXEC_SECURE) { |
40e1956e | 3334 | if (GRAPHICS_VER(i915) >= 11) |
44157641 JB |
3335 | return -ENODEV; |
3336 | ||
3337 | /* Return -EPERM to trigger fallback code on old binaries. */ | |
3338 | if (!HAS_SECURE_BATCHES(i915)) | |
3339 | return -EPERM; | |
3340 | ||
b3ac9f25 | 3341 | if (!drm_is_current_master(file) || !capable(CAP_SYS_ADMIN)) |
44157641 | 3342 | return -EPERM; |
d7d4eedd | 3343 | |
2889caa9 | 3344 | eb.batch_flags |= I915_DISPATCH_SECURE; |
d7d4eedd | 3345 | } |
b45305fc | 3346 | if (args->flags & I915_EXEC_IS_PINNED) |
2889caa9 | 3347 | eb.batch_flags |= I915_DISPATCH_PINNED; |
54cf91dc | 3348 | |
13149e8b LL |
3349 | err = parse_execbuf2_extensions(args, &eb); |
3350 | if (err) | |
3351 | goto err_ext; | |
3352 | ||
3353 | err = add_fence_array(&eb); | |
3354 | if (err) | |
3355 | goto err_ext; | |
3356 | ||
889333c7 CW |
3357 | #define IN_FENCES (I915_EXEC_FENCE_IN | I915_EXEC_FENCE_SUBMIT) |
3358 | if (args->flags & IN_FENCES) { | |
3359 | if ((args->flags & IN_FENCES) == IN_FENCES) | |
3360 | return -EINVAL; | |
3361 | ||
fec0445c | 3362 | in_fence = sync_file_get_fence(lower_32_bits(args->rsvd2)); |
13149e8b LL |
3363 | if (!in_fence) { |
3364 | err = -EINVAL; | |
3365 | goto err_ext; | |
3366 | } | |
fec0445c | 3367 | } |
889333c7 | 3368 | #undef IN_FENCES |
a88b6e4c | 3369 | |
fec0445c CW |
3370 | if (args->flags & I915_EXEC_FENCE_OUT) { |
3371 | out_fence_fd = get_unused_fd_flags(O_CLOEXEC); | |
3372 | if (out_fence_fd < 0) { | |
2889caa9 | 3373 | err = out_fence_fd; |
889333c7 | 3374 | goto err_in_fence; |
fec0445c CW |
3375 | } |
3376 | } | |
3377 | ||
cda9edd0 LL |
3378 | err = eb_create(&eb); |
3379 | if (err) | |
13149e8b | 3380 | goto err_out_fence; |
cda9edd0 | 3381 | |
4d470f73 | 3382 | GEM_BUG_ON(!eb.lut_size); |
2889caa9 | 3383 | |
1acfc104 CW |
3384 | err = eb_select_context(&eb); |
3385 | if (unlikely(err)) | |
3386 | goto err_destroy; | |
3387 | ||
2bf541ff | 3388 | err = eb_select_engine(&eb); |
d6f328bf | 3389 | if (unlikely(err)) |
e5dadff4 | 3390 | goto err_context; |
d6f328bf | 3391 | |
c43ce123 ML |
3392 | err = eb_lookup_vmas(&eb); |
3393 | if (err) { | |
b4b9731b | 3394 | eb_release_vmas(&eb, true); |
c43ce123 ML |
3395 | goto err_engine; |
3396 | } | |
3397 | ||
3398 | i915_gem_ww_ctx_init(&eb.ww, true); | |
3399 | ||
8e4ba491 | 3400 | err = eb_relocate_parse(&eb); |
1f727d9e | 3401 | if (err) { |
2889caa9 CW |
3402 | /* |
3403 | * If the user expects the execobject.offset and | |
3404 | * reloc.presumed_offset to be an exact match, | |
3405 | * as for using NO_RELOC, then we cannot update | |
3406 | * the execobject.offset until we have completed | |
3407 | * relocation. | |
3408 | */ | |
3409 | args->flags &= ~__EXEC_HAS_RELOC; | |
2889caa9 | 3410 | goto err_vma; |
1f727d9e | 3411 | } |
54cf91dc | 3412 | |
c43ce123 | 3413 | ww_acquire_done(&eb.ww.ctx); |
ff20afc4 | 3414 | eb_capture_stage(&eb); |
7d6236bb | 3415 | |
544460c3 MB |
3416 | out_fence = eb_requests_create(&eb, in_fence, out_fence_fd); |
3417 | if (IS_ERR(out_fence)) { | |
3418 | err = PTR_ERR(out_fence); | |
9cdb54be | 3419 | out_fence = NULL; |
544460c3 | 3420 | if (eb.requests[0]) |
cf6e7bac | 3421 | goto err_request; |
544460c3 MB |
3422 | else |
3423 | goto err_vma; | |
cf6e7bac JE |
3424 | } |
3425 | ||
544460c3 | 3426 | err = eb_submit(&eb); |
ed29c269 | 3427 | |
aa9b7810 | 3428 | err_request: |
544460c3 MB |
3429 | eb_requests_get(&eb); |
3430 | err = eb_requests_add(&eb, err); | |
c8659efa | 3431 | |
13149e8b | 3432 | if (eb.fences) |
544460c3 MB |
3433 | signal_fence_array(&eb, eb.composite_fence ? |
3434 | eb.composite_fence : | |
3435 | &eb.requests[0]->fence); | |
cf6e7bac | 3436 | |
fec0445c | 3437 | if (out_fence) { |
2889caa9 | 3438 | if (err == 0) { |
fec0445c | 3439 | fd_install(out_fence_fd, out_fence->file); |
b6a88e4a | 3440 | args->rsvd2 &= GENMASK_ULL(31, 0); /* keep in-fence */ |
fec0445c CW |
3441 | args->rsvd2 |= (u64)out_fence_fd << 32; |
3442 | out_fence_fd = -1; | |
3443 | } else { | |
3444 | fput(out_fence->file); | |
3445 | } | |
3446 | } | |
00dae4d3 JE |
3447 | |
3448 | if (unlikely(eb.gem_context->syncobj)) { | |
3449 | drm_syncobj_replace_fence(eb.gem_context->syncobj, | |
544460c3 MB |
3450 | eb.composite_fence ? |
3451 | eb.composite_fence : | |
3452 | &eb.requests[0]->fence); | |
00dae4d3 JE |
3453 | } |
3454 | ||
544460c3 MB |
3455 | if (!out_fence && eb.composite_fence) |
3456 | dma_fence_put(eb.composite_fence); | |
3457 | ||
3458 | eb_requests_put(&eb); | |
54cf91dc | 3459 | |
2889caa9 | 3460 | err_vma: |
b4b9731b | 3461 | eb_release_vmas(&eb, true); |
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 | } |