Commit | Line | Data |
---|---|---|
54cf91dc | 1 | /* |
10be98a7 | 2 | * SPDX-License-Identifier: MIT |
54cf91dc | 3 | * |
10be98a7 | 4 | * Copyright © 2008,2010 Intel Corporation |
54cf91dc CW |
5 | */ |
6 | ||
daedaa33 | 7 | #include <linux/intel-iommu.h> |
52791eee | 8 | #include <linux/dma-resv.h> |
fec0445c | 9 | #include <linux/sync_file.h> |
ad778f89 CW |
10 | #include <linux/uaccess.h> |
11 | ||
cf6e7bac | 12 | #include <drm/drm_syncobj.h> |
ad778f89 | 13 | |
df0566a6 JN |
14 | #include "display/intel_frontbuffer.h" |
15 | ||
afa13085 | 16 | #include "gem/i915_gem_ioctls.h" |
10be98a7 | 17 | #include "gt/intel_context.h" |
45233ab2 | 18 | #include "gt/intel_gpu_commands.h" |
baea429d | 19 | #include "gt/intel_gt.h" |
16e87459 | 20 | #include "gt/intel_gt_buffer_pool.h" |
8f2a1057 | 21 | #include "gt/intel_gt_pm.h" |
2871ea85 | 22 | #include "gt/intel_ring.h" |
8f2a1057 | 23 | |
6da4a2c4 | 24 | #include "i915_drv.h" |
57822dc6 | 25 | #include "i915_gem_clflush.h" |
10be98a7 | 26 | #include "i915_gem_context.h" |
6da4a2c4 | 27 | #include "i915_gem_ioctls.h" |
54cf91dc | 28 | #include "i915_trace.h" |
cda9edd0 | 29 | #include "i915_user_extensions.h" |
54cf91dc | 30 | |
7d6236bb CW |
31 | struct eb_vma { |
32 | struct i915_vma *vma; | |
33 | unsigned int flags; | |
34 | ||
35 | /** This vma's place in the execbuf reservation list */ | |
36 | struct drm_i915_gem_exec_object2 *exec; | |
37 | struct list_head bind_link; | |
38 | struct list_head reloc_link; | |
39 | ||
40 | struct hlist_node node; | |
41 | u32 handle; | |
42 | }; | |
43 | ||
ad5d95e4 DA |
44 | enum { |
45 | FORCE_CPU_RELOC = 1, | |
46 | FORCE_GTT_RELOC, | |
47 | FORCE_GPU_RELOC, | |
48 | #define DBG_FORCE_RELOC 0 /* choose one of the above! */ | |
49 | }; | |
50 | ||
bfaae47d ML |
51 | /* __EXEC_OBJECT_NO_RESERVE is BIT(31), defined in i915_vma.h */ |
52 | #define __EXEC_OBJECT_HAS_PIN BIT(30) | |
53 | #define __EXEC_OBJECT_HAS_FENCE BIT(29) | |
ed29c269 ML |
54 | #define __EXEC_OBJECT_USERPTR_INIT BIT(28) |
55 | #define __EXEC_OBJECT_NEEDS_MAP BIT(27) | |
56 | #define __EXEC_OBJECT_NEEDS_BIAS BIT(26) | |
57 | #define __EXEC_OBJECT_INTERNAL_FLAGS (~0u << 26) /* all of the above + */ | |
8ae275c2 | 58 | #define __EXEC_OBJECT_RESERVED (__EXEC_OBJECT_HAS_PIN | __EXEC_OBJECT_HAS_FENCE) |
2889caa9 CW |
59 | |
60 | #define __EXEC_HAS_RELOC BIT(31) | |
2bf541ff | 61 | #define __EXEC_ENGINE_PINNED BIT(30) |
ed29c269 ML |
62 | #define __EXEC_USERPTR_USED BIT(29) |
63 | #define __EXEC_INTERNAL_FLAGS (~0u << 29) | |
2889caa9 | 64 | #define UPDATE PIN_OFFSET_FIXED |
d23db88c CW |
65 | |
66 | #define BATCH_OFFSET_BIAS (256*1024) | |
a415d355 | 67 | |
650bc635 | 68 | #define __I915_EXEC_ILLEGAL_FLAGS \ |
08e3e21a LDM |
69 | (__I915_EXEC_UNKNOWN_FLAGS | \ |
70 | I915_EXEC_CONSTANTS_MASK | \ | |
71 | I915_EXEC_RESOURCE_STREAMER) | |
5b043f4e | 72 | |
d20ac620 CW |
73 | /* Catch emission of unexpected errors for CI! */ |
74 | #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) | |
75 | #undef EINVAL | |
76 | #define EINVAL ({ \ | |
77 | DRM_DEBUG_DRIVER("EINVAL at %s:%d\n", __func__, __LINE__); \ | |
78 | 22; \ | |
79 | }) | |
80 | #endif | |
81 | ||
2889caa9 CW |
82 | /** |
83 | * DOC: User command execution | |
84 | * | |
85 | * Userspace submits commands to be executed on the GPU as an instruction | |
86 | * stream within a GEM object we call a batchbuffer. This instructions may | |
87 | * refer to other GEM objects containing auxiliary state such as kernels, | |
88 | * samplers, render targets and even secondary batchbuffers. Userspace does | |
89 | * not know where in the GPU memory these objects reside and so before the | |
90 | * batchbuffer is passed to the GPU for execution, those addresses in the | |
91 | * batchbuffer and auxiliary objects are updated. This is known as relocation, | |
92 | * or patching. To try and avoid having to relocate each object on the next | |
93 | * execution, userspace is told the location of those objects in this pass, | |
94 | * but this remains just a hint as the kernel may choose a new location for | |
95 | * any object in the future. | |
96 | * | |
99d7e4ee KR |
97 | * At the level of talking to the hardware, submitting a batchbuffer for the |
98 | * GPU to execute is to add content to a buffer from which the HW | |
99 | * command streamer is reading. | |
100 | * | |
101 | * 1. Add a command to load the HW context. For Logical Ring Contexts, i.e. | |
102 | * Execlists, this command is not placed on the same buffer as the | |
103 | * remaining items. | |
104 | * | |
105 | * 2. Add a command to invalidate caches to the buffer. | |
106 | * | |
107 | * 3. Add a batchbuffer start command to the buffer; the start command is | |
108 | * essentially a token together with the GPU address of the batchbuffer | |
109 | * to be executed. | |
110 | * | |
111 | * 4. Add a pipeline flush to the buffer. | |
112 | * | |
113 | * 5. Add a memory write command to the buffer to record when the GPU | |
114 | * is done executing the batchbuffer. The memory write writes the | |
115 | * global sequence number of the request, ``i915_request::global_seqno``; | |
116 | * the i915 driver uses the current value in the register to determine | |
117 | * if the GPU has completed the batchbuffer. | |
118 | * | |
119 | * 6. Add a user interrupt command to the buffer. This command instructs | |
120 | * the GPU to issue an interrupt when the command, pipeline flush and | |
121 | * memory write are completed. | |
122 | * | |
123 | * 7. Inform the hardware of the additional commands added to the buffer | |
124 | * (by updating the tail pointer). | |
125 | * | |
2889caa9 CW |
126 | * Processing an execbuf ioctl is conceptually split up into a few phases. |
127 | * | |
128 | * 1. Validation - Ensure all the pointers, handles and flags are valid. | |
129 | * 2. Reservation - Assign GPU address space for every object | |
130 | * 3. Relocation - Update any addresses to point to the final locations | |
131 | * 4. Serialisation - Order the request with respect to its dependencies | |
132 | * 5. Construction - Construct a request to execute the batchbuffer | |
133 | * 6. Submission (at some point in the future execution) | |
134 | * | |
135 | * Reserving resources for the execbuf is the most complicated phase. We | |
136 | * neither want to have to migrate the object in the address space, nor do | |
137 | * we want to have to update any relocations pointing to this object. Ideally, | |
138 | * we want to leave the object where it is and for all the existing relocations | |
139 | * to match. If the object is given a new address, or if userspace thinks the | |
140 | * object is elsewhere, we have to parse all the relocation entries and update | |
141 | * the addresses. Userspace can set the I915_EXEC_NORELOC flag to hint that | |
142 | * all the target addresses in all of its objects match the value in the | |
143 | * relocation entries and that they all match the presumed offsets given by the | |
144 | * list of execbuffer objects. Using this knowledge, we know that if we haven't | |
145 | * moved any buffers, all the relocation entries are valid and we can skip | |
146 | * the update. (If userspace is wrong, the likely outcome is an impromptu GPU | |
147 | * hang.) The requirement for using I915_EXEC_NO_RELOC are: | |
148 | * | |
149 | * The addresses written in the objects must match the corresponding | |
150 | * reloc.presumed_offset which in turn must match the corresponding | |
151 | * execobject.offset. | |
152 | * | |
153 | * Any render targets written to in the batch must be flagged with | |
154 | * EXEC_OBJECT_WRITE. | |
155 | * | |
156 | * To avoid stalling, execobject.offset should match the current | |
157 | * address of that object within the active context. | |
158 | * | |
159 | * The reservation is done is multiple phases. First we try and keep any | |
160 | * object already bound in its current location - so as long as meets the | |
161 | * constraints imposed by the new execbuffer. Any object left unbound after the | |
162 | * first pass is then fitted into any available idle space. If an object does | |
163 | * not fit, all objects are removed from the reservation and the process rerun | |
164 | * after sorting the objects into a priority order (more difficult to fit | |
165 | * objects are tried first). Failing that, the entire VM is cleared and we try | |
166 | * to fit the execbuf once last time before concluding that it simply will not | |
167 | * fit. | |
168 | * | |
169 | * A small complication to all of this is that we allow userspace not only to | |
170 | * specify an alignment and a size for the object in the address space, but | |
171 | * we also allow userspace to specify the exact offset. This objects are | |
172 | * simpler to place (the location is known a priori) all we have to do is make | |
173 | * sure the space is available. | |
174 | * | |
175 | * Once all the objects are in place, patching up the buried pointers to point | |
176 | * to the final locations is a fairly simple job of walking over the relocation | |
177 | * entry arrays, looking up the right address and rewriting the value into | |
178 | * the object. Simple! ... The relocation entries are stored in user memory | |
179 | * and so to access them we have to copy them into a local buffer. That copy | |
180 | * has to avoid taking any pagefaults as they may lead back to a GEM object | |
181 | * requiring the struct_mutex (i.e. recursive deadlock). So once again we split | |
182 | * the relocation into multiple passes. First we try to do everything within an | |
183 | * atomic context (avoid the pagefaults) which requires that we never wait. If | |
184 | * we detect that we may wait, or if we need to fault, then we have to fallback | |
185 | * to a slower path. The slowpath has to drop the mutex. (Can you hear alarm | |
186 | * bells yet?) Dropping the mutex means that we lose all the state we have | |
187 | * built up so far for the execbuf and we must reset any global data. However, | |
188 | * we do leave the objects pinned in their final locations - which is a | |
189 | * potential issue for concurrent execbufs. Once we have left the mutex, we can | |
190 | * allocate and copy all the relocation entries into a large array at our | |
191 | * leisure, reacquire the mutex, reclaim all the objects and other state and | |
192 | * then proceed to update any incorrect addresses with the objects. | |
193 | * | |
194 | * As we process the relocation entries, we maintain a record of whether the | |
195 | * object is being written to. Using NORELOC, we expect userspace to provide | |
196 | * this information instead. We also check whether we can skip the relocation | |
197 | * by comparing the expected value inside the relocation entry with the target's | |
198 | * final address. If they differ, we have to map the current object and rewrite | |
199 | * the 4 or 8 byte pointer within. | |
200 | * | |
201 | * Serialising an execbuf is quite simple according to the rules of the GEM | |
202 | * ABI. Execution within each context is ordered by the order of submission. | |
203 | * Writes to any GEM object are in order of submission and are exclusive. Reads | |
204 | * from a GEM object are unordered with respect to other reads, but ordered by | |
205 | * writes. A write submitted after a read cannot occur before the read, and | |
206 | * similarly any read submitted after a write cannot occur before the write. | |
207 | * Writes are ordered between engines such that only one write occurs at any | |
208 | * time (completing any reads beforehand) - using semaphores where available | |
209 | * and CPU serialisation otherwise. Other GEM access obey the same rules, any | |
210 | * write (either via mmaps using set-domain, or via pwrite) must flush all GPU | |
211 | * reads before starting, and any read (either using set-domain or pread) must | |
212 | * flush all GPU writes before starting. (Note we only employ a barrier before, | |
213 | * we currently rely on userspace not concurrently starting a new execution | |
214 | * whilst reading or writing to an object. This may be an advantage or not | |
215 | * depending on how much you trust userspace not to shoot themselves in the | |
216 | * foot.) Serialisation may just result in the request being inserted into | |
217 | * a DAG awaiting its turn, but most simple is to wait on the CPU until | |
218 | * all dependencies are resolved. | |
219 | * | |
220 | * After all of that, is just a matter of closing the request and handing it to | |
221 | * the hardware (well, leaving it in a queue to be executed). However, we also | |
222 | * offer the ability for batchbuffers to be run with elevated privileges so | |
223 | * that they access otherwise hidden registers. (Used to adjust L3 cache etc.) | |
224 | * Before any batch is given extra privileges we first must check that it | |
225 | * contains no nefarious instructions, we check that each instruction is from | |
226 | * our whitelist and all registers are also from an allowed list. We first | |
227 | * copy the user's batchbuffer to a shadow (so that the user doesn't have | |
228 | * access to it, either by the CPU or GPU as we scan it) and then parse each | |
229 | * instruction. If everything is ok, we set a flag telling the hardware to run | |
230 | * the batchbuffer in trusted mode, otherwise the ioctl is rejected. | |
231 | */ | |
232 | ||
13149e8b LL |
233 | struct eb_fence { |
234 | struct drm_syncobj *syncobj; /* Use with ptr_mask_bits() */ | |
235 | struct dma_fence *dma_fence; | |
236 | u64 value; | |
237 | struct dma_fence_chain *chain_fence; | |
238 | }; | |
239 | ||
650bc635 | 240 | struct i915_execbuffer { |
2889caa9 CW |
241 | struct drm_i915_private *i915; /** i915 backpointer */ |
242 | struct drm_file *file; /** per-file lookup tables and limits */ | |
243 | struct drm_i915_gem_execbuffer2 *args; /** ioctl parameters */ | |
244 | struct drm_i915_gem_exec_object2 *exec; /** ioctl execobj[] */ | |
7d6236bb | 245 | struct eb_vma *vma; |
2889caa9 CW |
246 | |
247 | struct intel_engine_cs *engine; /** engine to queue the request to */ | |
8f2a1057 CW |
248 | struct intel_context *context; /* logical state for the request */ |
249 | struct i915_gem_context *gem_context; /** caller's context */ | |
2889caa9 | 250 | |
e61e0f51 | 251 | struct i915_request *request; /** our request to build */ |
7d6236bb | 252 | struct eb_vma *batch; /** identity of the batch obj/vma */ |
32d94048 | 253 | struct i915_vma *trampoline; /** trampoline used for chaining */ |
2889caa9 CW |
254 | |
255 | /** actual size of execobj[] as we may extend it for the cmdparser */ | |
256 | unsigned int buffer_count; | |
257 | ||
258 | /** list of vma not yet bound during reservation phase */ | |
259 | struct list_head unbound; | |
260 | ||
261 | /** list of vma that have execobj.relocation_count */ | |
262 | struct list_head relocs; | |
263 | ||
c43ce123 ML |
264 | struct i915_gem_ww_ctx ww; |
265 | ||
2889caa9 CW |
266 | /** |
267 | * Track the most recently used object for relocations, as we | |
268 | * frequently have to perform multiple relocations within the same | |
269 | * obj/page | |
270 | */ | |
650bc635 | 271 | struct reloc_cache { |
2889caa9 | 272 | struct drm_mm_node node; /** temporary GTT binding */ |
ad5d95e4 DA |
273 | unsigned long vaddr; /** Current kmap address */ |
274 | unsigned long page; /** Currently mapped page index */ | |
8802190f | 275 | unsigned int graphics_ver; /** Cached value of GRAPHICS_VER */ |
650bc635 | 276 | bool use_64bit_reloc : 1; |
2889caa9 CW |
277 | bool has_llc : 1; |
278 | bool has_fence : 1; | |
279 | bool needs_unfenced : 1; | |
650bc635 | 280 | } reloc_cache; |
2889caa9 CW |
281 | |
282 | u64 invalid_flags; /** Set of execobj.flags that are invalid */ | |
2889caa9 | 283 | |
d5e87821 | 284 | u64 batch_len; /** Length of batch within object */ |
2889caa9 | 285 | u32 batch_start_offset; /** Location within object of batch */ |
2889caa9 | 286 | u32 batch_flags; /** Flags composed for emit_bb_start() */ |
c43ce123 | 287 | struct intel_gt_buffer_pool_node *batch_pool; /** pool node for batch buffer */ |
2889caa9 CW |
288 | |
289 | /** | |
290 | * Indicate either the size of the hastable used to resolve | |
291 | * relocation handles, or if negative that we are using a direct | |
292 | * index into the execobj[]. | |
293 | */ | |
294 | int lut_size; | |
295 | struct hlist_head *buckets; /** ht for relocation handles */ | |
cda9edd0 | 296 | |
13149e8b LL |
297 | struct eb_fence *fences; |
298 | unsigned long num_fences; | |
67731b87 CW |
299 | }; |
300 | ||
8e4ba491 | 301 | static int eb_parse(struct i915_execbuffer *eb); |
2bf541ff ML |
302 | static struct i915_request *eb_pin_engine(struct i915_execbuffer *eb, |
303 | bool throttle); | |
304 | static void eb_unpin_engine(struct i915_execbuffer *eb); | |
8e4ba491 | 305 | |
3dbf26ed CW |
306 | static inline bool eb_use_cmdparser(const struct i915_execbuffer *eb) |
307 | { | |
311a50e7 | 308 | return intel_engine_requires_cmd_parser(eb->engine) || |
435e8fc0 JB |
309 | (intel_engine_using_cmd_parser(eb->engine) && |
310 | eb->args->batch_len); | |
3dbf26ed CW |
311 | } |
312 | ||
650bc635 | 313 | static int eb_create(struct i915_execbuffer *eb) |
67731b87 | 314 | { |
2889caa9 CW |
315 | if (!(eb->args->flags & I915_EXEC_HANDLE_LUT)) { |
316 | unsigned int size = 1 + ilog2(eb->buffer_count); | |
4ff4b44c | 317 | |
2889caa9 CW |
318 | /* |
319 | * Without a 1:1 association between relocation handles and | |
320 | * the execobject[] index, we instead create a hashtable. | |
321 | * We size it dynamically based on available memory, starting | |
322 | * first with 1:1 assocative hash and scaling back until | |
323 | * the allocation succeeds. | |
324 | * | |
325 | * Later on we use a positive lut_size to indicate we are | |
326 | * using this hashtable, and a negative value to indicate a | |
327 | * direct lookup. | |
328 | */ | |
4ff4b44c | 329 | do { |
0d95c883 | 330 | gfp_t flags; |
4d470f73 CW |
331 | |
332 | /* While we can still reduce the allocation size, don't | |
333 | * raise a warning and allow the allocation to fail. | |
334 | * On the last pass though, we want to try as hard | |
335 | * as possible to perform the allocation and warn | |
336 | * if it fails. | |
337 | */ | |
0ee931c4 | 338 | flags = GFP_KERNEL; |
4d470f73 CW |
339 | if (size > 1) |
340 | flags |= __GFP_NORETRY | __GFP_NOWARN; | |
341 | ||
4ff4b44c | 342 | eb->buckets = kzalloc(sizeof(struct hlist_head) << size, |
4d470f73 | 343 | flags); |
4ff4b44c CW |
344 | if (eb->buckets) |
345 | break; | |
346 | } while (--size); | |
347 | ||
8ae275c2 | 348 | if (unlikely(!size)) |
4d470f73 | 349 | return -ENOMEM; |
eef90ccb | 350 | |
2889caa9 | 351 | eb->lut_size = size; |
650bc635 | 352 | } else { |
2889caa9 | 353 | eb->lut_size = -eb->buffer_count; |
650bc635 | 354 | } |
eef90ccb | 355 | |
650bc635 | 356 | return 0; |
67731b87 CW |
357 | } |
358 | ||
2889caa9 CW |
359 | static bool |
360 | eb_vma_misplaced(const struct drm_i915_gem_exec_object2 *entry, | |
c7c6e46f CW |
361 | const struct i915_vma *vma, |
362 | unsigned int flags) | |
2889caa9 | 363 | { |
2889caa9 CW |
364 | if (vma->node.size < entry->pad_to_size) |
365 | return true; | |
366 | ||
367 | if (entry->alignment && !IS_ALIGNED(vma->node.start, entry->alignment)) | |
368 | return true; | |
369 | ||
c7c6e46f | 370 | if (flags & EXEC_OBJECT_PINNED && |
2889caa9 CW |
371 | vma->node.start != entry->offset) |
372 | return true; | |
373 | ||
c7c6e46f | 374 | if (flags & __EXEC_OBJECT_NEEDS_BIAS && |
2889caa9 CW |
375 | vma->node.start < BATCH_OFFSET_BIAS) |
376 | return true; | |
377 | ||
c7c6e46f | 378 | if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) && |
5f22cc0b | 379 | (vma->node.start + vma->node.size + 4095) >> 32) |
2889caa9 CW |
380 | return true; |
381 | ||
1d033beb CW |
382 | if (flags & __EXEC_OBJECT_NEEDS_MAP && |
383 | !i915_vma_is_map_and_fenceable(vma)) | |
384 | return true; | |
385 | ||
2889caa9 CW |
386 | return false; |
387 | } | |
388 | ||
8a338f4b CW |
389 | static u64 eb_pin_flags(const struct drm_i915_gem_exec_object2 *entry, |
390 | unsigned int exec_flags) | |
391 | { | |
392 | u64 pin_flags = 0; | |
393 | ||
394 | if (exec_flags & EXEC_OBJECT_NEEDS_GTT) | |
395 | pin_flags |= PIN_GLOBAL; | |
396 | ||
397 | /* | |
398 | * Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset, | |
399 | * limit address to the first 4GBs for unflagged objects. | |
400 | */ | |
401 | if (!(exec_flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS)) | |
402 | pin_flags |= PIN_ZONE_4G; | |
403 | ||
404 | if (exec_flags & __EXEC_OBJECT_NEEDS_MAP) | |
405 | pin_flags |= PIN_MAPPABLE; | |
406 | ||
407 | if (exec_flags & EXEC_OBJECT_PINNED) | |
408 | pin_flags |= entry->offset | PIN_OFFSET_FIXED; | |
409 | else if (exec_flags & __EXEC_OBJECT_NEEDS_BIAS) | |
410 | pin_flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS; | |
411 | ||
412 | return pin_flags; | |
413 | } | |
414 | ||
237647f4 | 415 | static inline int |
2889caa9 | 416 | eb_pin_vma(struct i915_execbuffer *eb, |
c7c6e46f | 417 | const struct drm_i915_gem_exec_object2 *entry, |
7d6236bb | 418 | struct eb_vma *ev) |
2889caa9 | 419 | { |
7d6236bb | 420 | struct i915_vma *vma = ev->vma; |
c7c6e46f | 421 | u64 pin_flags; |
237647f4 | 422 | int err; |
2889caa9 | 423 | |
616d9cee | 424 | if (vma->node.size) |
c7c6e46f | 425 | pin_flags = vma->node.start; |
616d9cee | 426 | else |
c7c6e46f | 427 | pin_flags = entry->offset & PIN_OFFSET_MASK; |
616d9cee | 428 | |
c7c6e46f | 429 | pin_flags |= PIN_USER | PIN_NOEVICT | PIN_OFFSET_FIXED; |
7d6236bb | 430 | if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_GTT)) |
c7c6e46f | 431 | pin_flags |= PIN_GLOBAL; |
616d9cee | 432 | |
8a338f4b | 433 | /* Attempt to reuse the current location if available */ |
237647f4 ML |
434 | err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, pin_flags); |
435 | if (err == -EDEADLK) | |
436 | return err; | |
437 | ||
438 | if (unlikely(err)) { | |
8a338f4b | 439 | if (entry->flags & EXEC_OBJECT_PINNED) |
237647f4 | 440 | return err; |
8a338f4b CW |
441 | |
442 | /* Failing that pick any _free_ space if suitable */ | |
237647f4 | 443 | err = i915_vma_pin_ww(vma, &eb->ww, |
47b08693 ML |
444 | entry->pad_to_size, |
445 | entry->alignment, | |
446 | eb_pin_flags(entry, ev->flags) | | |
237647f4 ML |
447 | PIN_USER | PIN_NOEVICT); |
448 | if (unlikely(err)) | |
449 | return err; | |
8a338f4b | 450 | } |
2889caa9 | 451 | |
7d6236bb | 452 | if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) { |
237647f4 ML |
453 | err = i915_vma_pin_fence(vma); |
454 | if (unlikely(err)) { | |
2889caa9 | 455 | i915_vma_unpin(vma); |
237647f4 | 456 | return err; |
2889caa9 CW |
457 | } |
458 | ||
3bd40735 | 459 | if (vma->fence) |
7d6236bb | 460 | ev->flags |= __EXEC_OBJECT_HAS_FENCE; |
2889caa9 CW |
461 | } |
462 | ||
7d6236bb | 463 | ev->flags |= __EXEC_OBJECT_HAS_PIN; |
237647f4 ML |
464 | if (eb_vma_misplaced(entry, vma, ev->flags)) |
465 | return -EBADSLT; | |
466 | ||
467 | return 0; | |
2889caa9 CW |
468 | } |
469 | ||
8ae275c2 ML |
470 | static inline void |
471 | eb_unreserve_vma(struct eb_vma *ev) | |
472 | { | |
473 | if (!(ev->flags & __EXEC_OBJECT_HAS_PIN)) | |
474 | return; | |
475 | ||
c43ce123 ML |
476 | if (unlikely(ev->flags & __EXEC_OBJECT_HAS_FENCE)) |
477 | __i915_vma_unpin_fence(ev->vma); | |
478 | ||
479 | __i915_vma_unpin(ev->vma); | |
8ae275c2 ML |
480 | ev->flags &= ~__EXEC_OBJECT_RESERVED; |
481 | } | |
482 | ||
2889caa9 CW |
483 | static int |
484 | eb_validate_vma(struct i915_execbuffer *eb, | |
485 | struct drm_i915_gem_exec_object2 *entry, | |
486 | struct i915_vma *vma) | |
67731b87 | 487 | { |
2eb8e1a6 JE |
488 | /* Relocations are disallowed for all platforms after TGL-LP. This |
489 | * also covers all platforms with local memory. | |
490 | */ | |
491 | if (entry->relocation_count && | |
40e1956e | 492 | GRAPHICS_VER(eb->i915) >= 12 && !IS_TIGERLAKE(eb->i915)) |
2eb8e1a6 JE |
493 | return -EINVAL; |
494 | ||
2889caa9 CW |
495 | if (unlikely(entry->flags & eb->invalid_flags)) |
496 | return -EINVAL; | |
d55495b4 | 497 | |
2920516b MA |
498 | if (unlikely(entry->alignment && |
499 | !is_power_of_2_u64(entry->alignment))) | |
2889caa9 CW |
500 | return -EINVAL; |
501 | ||
502 | /* | |
503 | * Offset can be used as input (EXEC_OBJECT_PINNED), reject | |
504 | * any non-page-aligned or non-canonical addresses. | |
505 | */ | |
506 | if (unlikely(entry->flags & EXEC_OBJECT_PINNED && | |
6fc4e48f | 507 | entry->offset != gen8_canonical_addr(entry->offset & I915_GTT_PAGE_MASK))) |
2889caa9 CW |
508 | return -EINVAL; |
509 | ||
510 | /* pad_to_size was once a reserved field, so sanitize it */ | |
511 | if (entry->flags & EXEC_OBJECT_PAD_TO_SIZE) { | |
512 | if (unlikely(offset_in_page(entry->pad_to_size))) | |
513 | return -EINVAL; | |
514 | } else { | |
515 | entry->pad_to_size = 0; | |
d55495b4 | 516 | } |
2889caa9 CW |
517 | /* |
518 | * From drm_mm perspective address space is continuous, | |
519 | * so from this point we're always using non-canonical | |
520 | * form internally. | |
521 | */ | |
522 | entry->offset = gen8_noncanonical_addr(entry->offset); | |
523 | ||
c7c6e46f CW |
524 | if (!eb->reloc_cache.has_fence) { |
525 | entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE; | |
526 | } else { | |
527 | if ((entry->flags & EXEC_OBJECT_NEEDS_FENCE || | |
528 | eb->reloc_cache.needs_unfenced) && | |
529 | i915_gem_object_is_tiled(vma->obj)) | |
530 | entry->flags |= EXEC_OBJECT_NEEDS_GTT | __EXEC_OBJECT_NEEDS_MAP; | |
531 | } | |
532 | ||
2889caa9 | 533 | return 0; |
67731b87 CW |
534 | } |
535 | ||
003d8b91 | 536 | static void |
746c8f14 CW |
537 | eb_add_vma(struct i915_execbuffer *eb, |
538 | unsigned int i, unsigned batch_idx, | |
539 | struct i915_vma *vma) | |
59bfa124 | 540 | { |
c7c6e46f | 541 | struct drm_i915_gem_exec_object2 *entry = &eb->exec[i]; |
7d6236bb | 542 | struct eb_vma *ev = &eb->vma[i]; |
2889caa9 | 543 | |
93159e12 | 544 | ev->vma = vma; |
7d6236bb CW |
545 | ev->exec = entry; |
546 | ev->flags = entry->flags; | |
547 | ||
4d470f73 | 548 | if (eb->lut_size > 0) { |
7d6236bb CW |
549 | ev->handle = entry->handle; |
550 | hlist_add_head(&ev->node, | |
2889caa9 CW |
551 | &eb->buckets[hash_32(entry->handle, |
552 | eb->lut_size)]); | |
4ff4b44c | 553 | } |
59bfa124 | 554 | |
2889caa9 | 555 | if (entry->relocation_count) |
7d6236bb | 556 | list_add_tail(&ev->reloc_link, &eb->relocs); |
2889caa9 | 557 | |
746c8f14 CW |
558 | /* |
559 | * SNA is doing fancy tricks with compressing batch buffers, which leads | |
560 | * to negative relocation deltas. Usually that works out ok since the | |
561 | * relocate address is still positive, except when the batch is placed | |
562 | * very low in the GTT. Ensure this doesn't happen. | |
563 | * | |
564 | * Note that actual hangs have only been observed on gen7, but for | |
565 | * paranoia do it everywhere. | |
566 | */ | |
567 | if (i == batch_idx) { | |
827db9d8 | 568 | if (entry->relocation_count && |
7d6236bb CW |
569 | !(ev->flags & EXEC_OBJECT_PINNED)) |
570 | ev->flags |= __EXEC_OBJECT_NEEDS_BIAS; | |
746c8f14 | 571 | if (eb->reloc_cache.has_fence) |
7d6236bb | 572 | ev->flags |= EXEC_OBJECT_NEEDS_FENCE; |
746c8f14 | 573 | |
7d6236bb | 574 | eb->batch = ev; |
746c8f14 | 575 | } |
2889caa9 CW |
576 | } |
577 | ||
ad5d95e4 DA |
578 | static inline int use_cpu_reloc(const struct reloc_cache *cache, |
579 | const struct drm_i915_gem_object *obj) | |
580 | { | |
581 | if (!i915_gem_object_has_struct_page(obj)) | |
582 | return false; | |
583 | ||
584 | if (DBG_FORCE_RELOC == FORCE_CPU_RELOC) | |
585 | return true; | |
586 | ||
587 | if (DBG_FORCE_RELOC == FORCE_GTT_RELOC) | |
588 | return false; | |
589 | ||
590 | return (cache->has_llc || | |
591 | obj->cache_dirty || | |
592 | obj->cache_level != I915_CACHE_NONE); | |
593 | } | |
594 | ||
47b08693 | 595 | static int eb_reserve_vma(struct i915_execbuffer *eb, |
7d6236bb | 596 | struct eb_vma *ev, |
2920bb94 | 597 | u64 pin_flags) |
2889caa9 | 598 | { |
7d6236bb | 599 | struct drm_i915_gem_exec_object2 *entry = ev->exec; |
7d6236bb | 600 | struct i915_vma *vma = ev->vma; |
2889caa9 CW |
601 | int err; |
602 | ||
003d8b91 CW |
603 | if (drm_mm_node_allocated(&vma->node) && |
604 | eb_vma_misplaced(entry, vma, ev->flags)) { | |
605 | err = i915_vma_unbind(vma); | |
606 | if (err) | |
607 | return err; | |
608 | } | |
609 | ||
47b08693 | 610 | err = i915_vma_pin_ww(vma, &eb->ww, |
c7c6e46f | 611 | entry->pad_to_size, entry->alignment, |
8a338f4b | 612 | eb_pin_flags(entry, ev->flags) | pin_flags); |
2889caa9 CW |
613 | if (err) |
614 | return err; | |
615 | ||
616 | if (entry->offset != vma->node.start) { | |
617 | entry->offset = vma->node.start | UPDATE; | |
618 | eb->args->flags |= __EXEC_HAS_RELOC; | |
619 | } | |
620 | ||
8a338f4b | 621 | if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) { |
3bd40735 | 622 | err = i915_vma_pin_fence(vma); |
2889caa9 CW |
623 | if (unlikely(err)) { |
624 | i915_vma_unpin(vma); | |
625 | return err; | |
626 | } | |
627 | ||
3bd40735 | 628 | if (vma->fence) |
8a338f4b | 629 | ev->flags |= __EXEC_OBJECT_HAS_FENCE; |
2889caa9 CW |
630 | } |
631 | ||
8a338f4b | 632 | ev->flags |= __EXEC_OBJECT_HAS_PIN; |
7d6236bb | 633 | GEM_BUG_ON(eb_vma_misplaced(entry, vma, ev->flags)); |
1da7b54c | 634 | |
2889caa9 CW |
635 | return 0; |
636 | } | |
637 | ||
638 | static int eb_reserve(struct i915_execbuffer *eb) | |
639 | { | |
640 | const unsigned int count = eb->buffer_count; | |
2920bb94 | 641 | unsigned int pin_flags = PIN_USER | PIN_NONBLOCK; |
2889caa9 | 642 | struct list_head last; |
7d6236bb | 643 | struct eb_vma *ev; |
2889caa9 | 644 | unsigned int i, pass; |
ef398881 | 645 | int err = 0; |
2889caa9 CW |
646 | |
647 | /* | |
648 | * Attempt to pin all of the buffers into the GTT. | |
649 | * This is done in 3 phases: | |
650 | * | |
651 | * 1a. Unbind all objects that do not match the GTT constraints for | |
652 | * the execbuffer (fenceable, mappable, alignment etc). | |
653 | * 1b. Increment pin count for already bound objects. | |
654 | * 2. Bind new objects. | |
655 | * 3. Decrement pin count. | |
656 | * | |
657 | * This avoid unnecessary unbinding of later objects in order to make | |
658 | * room for the earlier objects *unless* we need to defragment. | |
659 | */ | |
2889caa9 | 660 | pass = 0; |
2889caa9 | 661 | do { |
7d6236bb CW |
662 | list_for_each_entry(ev, &eb->unbound, bind_link) { |
663 | err = eb_reserve_vma(eb, ev, pin_flags); | |
2889caa9 CW |
664 | if (err) |
665 | break; | |
666 | } | |
fd1500fc | 667 | if (err != -ENOSPC) |
c43ce123 | 668 | return err; |
2889caa9 CW |
669 | |
670 | /* Resort *all* the objects into priority order */ | |
671 | INIT_LIST_HEAD(&eb->unbound); | |
672 | INIT_LIST_HEAD(&last); | |
673 | for (i = 0; i < count; i++) { | |
7d6236bb | 674 | unsigned int flags; |
2889caa9 | 675 | |
7d6236bb CW |
676 | ev = &eb->vma[i]; |
677 | flags = ev->flags; | |
c7c6e46f CW |
678 | if (flags & EXEC_OBJECT_PINNED && |
679 | flags & __EXEC_OBJECT_HAS_PIN) | |
2889caa9 CW |
680 | continue; |
681 | ||
7d6236bb | 682 | eb_unreserve_vma(ev); |
2889caa9 | 683 | |
c7c6e46f | 684 | if (flags & EXEC_OBJECT_PINNED) |
35e882a4 | 685 | /* Pinned must have their slot */ |
7d6236bb | 686 | list_add(&ev->bind_link, &eb->unbound); |
c7c6e46f | 687 | else if (flags & __EXEC_OBJECT_NEEDS_MAP) |
35e882a4 | 688 | /* Map require the lowest 256MiB (aperture) */ |
7d6236bb | 689 | list_add_tail(&ev->bind_link, &eb->unbound); |
35e882a4 CW |
690 | else if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS)) |
691 | /* Prioritise 4GiB region for restricted bo */ | |
7d6236bb | 692 | list_add(&ev->bind_link, &last); |
2889caa9 | 693 | else |
7d6236bb | 694 | list_add_tail(&ev->bind_link, &last); |
2889caa9 CW |
695 | } |
696 | list_splice_tail(&last, &eb->unbound); | |
697 | ||
698 | switch (pass++) { | |
699 | case 0: | |
700 | break; | |
701 | ||
702 | case 1: | |
703 | /* Too fragmented, unbind everything and retry */ | |
2850748e | 704 | mutex_lock(&eb->context->vm->mutex); |
f5d974f9 | 705 | err = i915_gem_evict_vm(eb->context->vm); |
2850748e | 706 | mutex_unlock(&eb->context->vm->mutex); |
2889caa9 | 707 | if (err) |
c43ce123 | 708 | return err; |
2889caa9 CW |
709 | break; |
710 | ||
711 | default: | |
c43ce123 | 712 | return -ENOSPC; |
2889caa9 | 713 | } |
2920bb94 CW |
714 | |
715 | pin_flags = PIN_USER; | |
2889caa9 | 716 | } while (1); |
4ff4b44c | 717 | } |
59bfa124 | 718 | |
2889caa9 CW |
719 | static unsigned int eb_batch_index(const struct i915_execbuffer *eb) |
720 | { | |
1a71cf2f CW |
721 | if (eb->args->flags & I915_EXEC_BATCH_FIRST) |
722 | return 0; | |
723 | else | |
724 | return eb->buffer_count - 1; | |
2889caa9 CW |
725 | } |
726 | ||
727 | static int eb_select_context(struct i915_execbuffer *eb) | |
728 | { | |
729 | struct i915_gem_context *ctx; | |
730 | ||
731 | ctx = i915_gem_context_lookup(eb->file->driver_priv, eb->args->rsvd1); | |
046d1660 JE |
732 | if (unlikely(IS_ERR(ctx))) |
733 | return PTR_ERR(ctx); | |
2889caa9 | 734 | |
8f2a1057 | 735 | eb->gem_context = ctx; |
a4e7ccda | 736 | if (rcu_access_pointer(ctx->vm)) |
4f2c7337 | 737 | eb->invalid_flags |= EXEC_OBJECT_NEEDS_GTT; |
2889caa9 | 738 | |
2889caa9 CW |
739 | return 0; |
740 | } | |
741 | ||
93159e12 CW |
742 | static int __eb_add_lut(struct i915_execbuffer *eb, |
743 | u32 handle, struct i915_vma *vma) | |
3b96eff4 | 744 | { |
93159e12 CW |
745 | struct i915_gem_context *ctx = eb->gem_context; |
746 | struct i915_lut_handle *lut; | |
2889caa9 | 747 | int err; |
3b96eff4 | 748 | |
93159e12 CW |
749 | lut = i915_lut_handle_alloc(); |
750 | if (unlikely(!lut)) | |
751 | return -ENOMEM; | |
752 | ||
753 | i915_vma_get(vma); | |
754 | if (!atomic_fetch_inc(&vma->open_count)) | |
755 | i915_vma_reopen(vma); | |
756 | lut->handle = handle; | |
757 | lut->ctx = ctx; | |
758 | ||
759 | /* Check that the context hasn't been closed in the meantime */ | |
760 | err = -EINTR; | |
f7ce8639 CW |
761 | if (!mutex_lock_interruptible(&ctx->lut_mutex)) { |
762 | struct i915_address_space *vm = rcu_access_pointer(ctx->vm); | |
763 | ||
764 | if (unlikely(vm && vma->vm != vm)) | |
765 | err = -EAGAIN; /* user racing with ctx set-vm */ | |
766 | else if (likely(!i915_gem_context_is_closed(ctx))) | |
93159e12 | 767 | err = radix_tree_insert(&ctx->handles_vma, handle, vma); |
f7ce8639 CW |
768 | else |
769 | err = -ENOENT; | |
93159e12 CW |
770 | if (err == 0) { /* And nor has this handle */ |
771 | struct drm_i915_gem_object *obj = vma->obj; | |
772 | ||
096a42dd | 773 | spin_lock(&obj->lut_lock); |
93159e12 CW |
774 | if (idr_find(&eb->file->object_idr, handle) == obj) { |
775 | list_add(&lut->obj_link, &obj->lut_list); | |
776 | } else { | |
777 | radix_tree_delete(&ctx->handles_vma, handle); | |
778 | err = -ENOENT; | |
779 | } | |
096a42dd | 780 | spin_unlock(&obj->lut_lock); |
93159e12 | 781 | } |
f7ce8639 | 782 | mutex_unlock(&ctx->lut_mutex); |
93159e12 CW |
783 | } |
784 | if (unlikely(err)) | |
785 | goto err; | |
003d8b91 | 786 | |
93159e12 | 787 | return 0; |
d55495b4 | 788 | |
93159e12 | 789 | err: |
50689771 | 790 | i915_vma_close(vma); |
93159e12 CW |
791 | i915_vma_put(vma); |
792 | i915_lut_handle_free(lut); | |
793 | return err; | |
794 | } | |
746c8f14 | 795 | |
93159e12 CW |
796 | static struct i915_vma *eb_lookup_vma(struct i915_execbuffer *eb, u32 handle) |
797 | { | |
f7ce8639 CW |
798 | struct i915_address_space *vm = eb->context->vm; |
799 | ||
93159e12 CW |
800 | do { |
801 | struct drm_i915_gem_object *obj; | |
170fa29b | 802 | struct i915_vma *vma; |
93159e12 | 803 | int err; |
4ff4b44c | 804 | |
93159e12 CW |
805 | rcu_read_lock(); |
806 | vma = radix_tree_lookup(&eb->gem_context->handles_vma, handle); | |
f7ce8639 | 807 | if (likely(vma && vma->vm == vm)) |
93159e12 CW |
808 | vma = i915_vma_tryget(vma); |
809 | rcu_read_unlock(); | |
810 | if (likely(vma)) | |
811 | return vma; | |
4ff4b44c | 812 | |
170fa29b | 813 | obj = i915_gem_object_lookup(eb->file, handle); |
93159e12 CW |
814 | if (unlikely(!obj)) |
815 | return ERR_PTR(-ENOENT); | |
3b96eff4 | 816 | |
f7ce8639 | 817 | vma = i915_vma_instance(obj, vm, NULL); |
772b5408 | 818 | if (IS_ERR(vma)) { |
93159e12 CW |
819 | i915_gem_object_put(obj); |
820 | return vma; | |
27173f1f BW |
821 | } |
822 | ||
93159e12 CW |
823 | err = __eb_add_lut(eb, handle, vma); |
824 | if (likely(!err)) | |
825 | return vma; | |
d1b48c1e | 826 | |
93159e12 CW |
827 | i915_gem_object_put(obj); |
828 | if (err != -EEXIST) | |
829 | return ERR_PTR(err); | |
830 | } while (1); | |
831 | } | |
4ff4b44c | 832 | |
93159e12 CW |
833 | static int eb_lookup_vmas(struct i915_execbuffer *eb) |
834 | { | |
8e4ba491 | 835 | struct drm_i915_private *i915 = eb->i915; |
93159e12 CW |
836 | unsigned int batch = eb_batch_index(eb); |
837 | unsigned int i; | |
838 | int err = 0; | |
155ab883 | 839 | |
93159e12 | 840 | INIT_LIST_HEAD(&eb->relocs); |
93159e12 CW |
841 | |
842 | for (i = 0; i < eb->buffer_count; i++) { | |
843 | struct i915_vma *vma; | |
844 | ||
845 | vma = eb_lookup_vma(eb, eb->exec[i].handle); | |
846 | if (IS_ERR(vma)) { | |
847 | err = PTR_ERR(vma); | |
8e4ba491 | 848 | goto err; |
93159e12 | 849 | } |
d1b48c1e | 850 | |
003d8b91 | 851 | err = eb_validate_vma(eb, &eb->exec[i], vma); |
93159e12 CW |
852 | if (unlikely(err)) { |
853 | i915_vma_put(vma); | |
8e4ba491 | 854 | goto err; |
93159e12 | 855 | } |
dade2a61 | 856 | |
003d8b91 | 857 | eb_add_vma(eb, i, batch, vma); |
ed29c269 ML |
858 | |
859 | if (i915_gem_object_is_userptr(vma->obj)) { | |
860 | err = i915_gem_object_userptr_submit_init(vma->obj); | |
861 | if (err) { | |
862 | if (i + 1 < eb->buffer_count) { | |
863 | /* | |
864 | * Execbuffer code expects last vma entry to be NULL, | |
865 | * since we already initialized this entry, | |
866 | * set the next value to NULL or we mess up | |
867 | * cleanup handling. | |
868 | */ | |
869 | eb->vma[i + 1].vma = NULL; | |
870 | } | |
871 | ||
872 | return err; | |
873 | } | |
874 | ||
875 | eb->vma[i].flags |= __EXEC_OBJECT_USERPTR_INIT; | |
876 | eb->args->flags |= __EXEC_USERPTR_USED; | |
877 | } | |
4ff4b44c CW |
878 | } |
879 | ||
8e4ba491 ML |
880 | if (unlikely(eb->batch->flags & EXEC_OBJECT_WRITE)) { |
881 | drm_dbg(&i915->drm, | |
882 | "Attempting to use self-modifying batch buffer\n"); | |
883 | return -EINVAL; | |
884 | } | |
885 | ||
886 | if (range_overflows_t(u64, | |
887 | eb->batch_start_offset, eb->batch_len, | |
888 | eb->batch->vma->size)) { | |
889 | drm_dbg(&i915->drm, "Attempting to use out-of-bounds batch\n"); | |
890 | return -EINVAL; | |
891 | } | |
892 | ||
893 | if (eb->batch_len == 0) | |
894 | eb->batch_len = eb->batch->vma->size - eb->batch_start_offset; | |
d5e87821 CW |
895 | if (unlikely(eb->batch_len == 0)) { /* impossible! */ |
896 | drm_dbg(&i915->drm, "Invalid batch length\n"); | |
897 | return -EINVAL; | |
898 | } | |
8e4ba491 ML |
899 | |
900 | return 0; | |
901 | ||
902 | err: | |
7d6236bb | 903 | eb->vma[i].vma = NULL; |
2889caa9 | 904 | return err; |
3b96eff4 CW |
905 | } |
906 | ||
5cd57f67 | 907 | static int eb_lock_vmas(struct i915_execbuffer *eb) |
c43ce123 ML |
908 | { |
909 | unsigned int i; | |
910 | int err; | |
911 | ||
c43ce123 | 912 | for (i = 0; i < eb->buffer_count; i++) { |
c43ce123 ML |
913 | struct eb_vma *ev = &eb->vma[i]; |
914 | struct i915_vma *vma = ev->vma; | |
915 | ||
916 | err = i915_gem_object_lock(vma->obj, &eb->ww); | |
917 | if (err) | |
918 | return err; | |
5cd57f67 TH |
919 | } |
920 | ||
921 | return 0; | |
922 | } | |
923 | ||
924 | static int eb_validate_vmas(struct i915_execbuffer *eb) | |
925 | { | |
926 | unsigned int i; | |
927 | int err; | |
928 | ||
929 | INIT_LIST_HEAD(&eb->unbound); | |
930 | ||
931 | err = eb_lock_vmas(eb); | |
932 | if (err) | |
933 | return err; | |
934 | ||
935 | for (i = 0; i < eb->buffer_count; i++) { | |
936 | struct drm_i915_gem_exec_object2 *entry = &eb->exec[i]; | |
937 | struct eb_vma *ev = &eb->vma[i]; | |
938 | struct i915_vma *vma = ev->vma; | |
c43ce123 | 939 | |
237647f4 ML |
940 | err = eb_pin_vma(eb, entry, ev); |
941 | if (err == -EDEADLK) | |
942 | return err; | |
943 | ||
944 | if (!err) { | |
c43ce123 ML |
945 | if (entry->offset != vma->node.start) { |
946 | entry->offset = vma->node.start | UPDATE; | |
947 | eb->args->flags |= __EXEC_HAS_RELOC; | |
948 | } | |
949 | } else { | |
950 | eb_unreserve_vma(ev); | |
951 | ||
952 | list_add_tail(&ev->bind_link, &eb->unbound); | |
953 | if (drm_mm_node_allocated(&vma->node)) { | |
954 | err = i915_vma_unbind(vma); | |
955 | if (err) | |
956 | return err; | |
957 | } | |
958 | } | |
959 | ||
bfaae47d ML |
960 | if (!(ev->flags & EXEC_OBJECT_WRITE)) { |
961 | err = dma_resv_reserve_shared(vma->resv, 1); | |
962 | if (err) | |
963 | return err; | |
964 | } | |
965 | ||
c43ce123 ML |
966 | GEM_BUG_ON(drm_mm_node_allocated(&vma->node) && |
967 | eb_vma_misplaced(&eb->exec[i], vma, ev->flags)); | |
968 | } | |
969 | ||
970 | if (!list_empty(&eb->unbound)) | |
971 | return eb_reserve(eb); | |
972 | ||
973 | return 0; | |
974 | } | |
975 | ||
7d6236bb | 976 | static struct eb_vma * |
2889caa9 | 977 | eb_get_vma(const struct i915_execbuffer *eb, unsigned long handle) |
67731b87 | 978 | { |
2889caa9 CW |
979 | if (eb->lut_size < 0) { |
980 | if (handle >= -eb->lut_size) | |
eef90ccb | 981 | return NULL; |
7d6236bb | 982 | return &eb->vma[handle]; |
eef90ccb CW |
983 | } else { |
984 | struct hlist_head *head; | |
7d6236bb | 985 | struct eb_vma *ev; |
67731b87 | 986 | |
2889caa9 | 987 | head = &eb->buckets[hash_32(handle, eb->lut_size)]; |
7d6236bb CW |
988 | hlist_for_each_entry(ev, head, node) { |
989 | if (ev->handle == handle) | |
990 | return ev; | |
eef90ccb CW |
991 | } |
992 | return NULL; | |
993 | } | |
67731b87 CW |
994 | } |
995 | ||
b4b9731b | 996 | static void eb_release_vmas(struct i915_execbuffer *eb, bool final) |
8ae275c2 ML |
997 | { |
998 | const unsigned int count = eb->buffer_count; | |
999 | unsigned int i; | |
1000 | ||
1001 | for (i = 0; i < count; i++) { | |
1002 | struct eb_vma *ev = &eb->vma[i]; | |
1003 | struct i915_vma *vma = ev->vma; | |
1004 | ||
1005 | if (!vma) | |
1006 | break; | |
1007 | ||
c43ce123 | 1008 | eb_unreserve_vma(ev); |
8ae275c2 | 1009 | |
c43ce123 ML |
1010 | if (final) |
1011 | i915_vma_put(vma); | |
8ae275c2 | 1012 | } |
2bf541ff ML |
1013 | |
1014 | eb_unpin_engine(eb); | |
8ae275c2 ML |
1015 | } |
1016 | ||
2889caa9 | 1017 | static void eb_destroy(const struct i915_execbuffer *eb) |
934acce3 | 1018 | { |
4d470f73 | 1019 | if (eb->lut_size > 0) |
2889caa9 | 1020 | kfree(eb->buckets); |
934acce3 MW |
1021 | } |
1022 | ||
2889caa9 | 1023 | static inline u64 |
d50415cc | 1024 | relocation_target(const struct drm_i915_gem_relocation_entry *reloc, |
2889caa9 | 1025 | const struct i915_vma *target) |
934acce3 | 1026 | { |
2889caa9 | 1027 | return gen8_canonical_addr((int)reloc->delta + target->node.start); |
934acce3 MW |
1028 | } |
1029 | ||
d50415cc CW |
1030 | static void reloc_cache_init(struct reloc_cache *cache, |
1031 | struct drm_i915_private *i915) | |
5032d871 | 1032 | { |
ad5d95e4 DA |
1033 | cache->page = -1; |
1034 | cache->vaddr = 0; | |
dfc5148f | 1035 | /* Must be a variable in the struct to allow GCC to unroll. */ |
8802190f | 1036 | cache->graphics_ver = GRAPHICS_VER(i915); |
2889caa9 | 1037 | cache->has_llc = HAS_LLC(i915); |
dfc5148f | 1038 | cache->use_64bit_reloc = HAS_64BIT_RELOC(i915); |
8802190f | 1039 | cache->has_fence = cache->graphics_ver < 4; |
7dd4f672 | 1040 | cache->needs_unfenced = INTEL_INFO(i915)->unfenced_needs_alignment; |
4ee92c71 | 1041 | cache->node.flags = 0; |
d50415cc | 1042 | } |
5032d871 | 1043 | |
20561da3 DA |
1044 | static inline void *unmask_page(unsigned long p) |
1045 | { | |
1046 | return (void *)(uintptr_t)(p & PAGE_MASK); | |
1047 | } | |
1048 | ||
1049 | static inline unsigned int unmask_flags(unsigned long p) | |
1050 | { | |
1051 | return p & ~PAGE_MASK; | |
1052 | } | |
1053 | ||
1054 | #define KMAP 0x4 /* after CLFLUSH_FLAGS */ | |
1055 | ||
1056 | static inline struct i915_ggtt *cache_to_ggtt(struct reloc_cache *cache) | |
1057 | { | |
1058 | struct drm_i915_private *i915 = | |
1059 | container_of(cache, struct i915_execbuffer, reloc_cache)->i915; | |
1060 | return &i915->ggtt; | |
1061 | } | |
1062 | ||
c43ce123 | 1063 | static void reloc_cache_reset(struct reloc_cache *cache, struct i915_execbuffer *eb) |
ad5d95e4 DA |
1064 | { |
1065 | void *vaddr; | |
1066 | ||
1067 | if (!cache->vaddr) | |
1068 | return; | |
1069 | ||
1070 | vaddr = unmask_page(cache->vaddr); | |
1071 | if (cache->vaddr & KMAP) { | |
1af343cd ML |
1072 | struct drm_i915_gem_object *obj = |
1073 | (struct drm_i915_gem_object *)cache->node.mm; | |
ad5d95e4 DA |
1074 | if (cache->vaddr & CLFLUSH_AFTER) |
1075 | mb(); | |
1076 | ||
1077 | kunmap_atomic(vaddr); | |
1af343cd | 1078 | i915_gem_object_finish_access(obj); |
ad5d95e4 DA |
1079 | } else { |
1080 | struct i915_ggtt *ggtt = cache_to_ggtt(cache); | |
1081 | ||
1082 | intel_gt_flush_ggtt_writes(ggtt->vm.gt); | |
1083 | io_mapping_unmap_atomic((void __iomem *)vaddr); | |
1084 | ||
1085 | if (drm_mm_node_allocated(&cache->node)) { | |
1086 | ggtt->vm.clear_range(&ggtt->vm, | |
1087 | cache->node.start, | |
1088 | cache->node.size); | |
1089 | mutex_lock(&ggtt->vm.mutex); | |
1090 | drm_mm_remove_node(&cache->node); | |
1091 | mutex_unlock(&ggtt->vm.mutex); | |
1092 | } else { | |
1093 | i915_vma_unpin((struct i915_vma *)cache->node.mm); | |
1094 | } | |
1095 | } | |
1096 | ||
1097 | cache->vaddr = 0; | |
1098 | cache->page = -1; | |
1099 | } | |
1100 | ||
1101 | static void *reloc_kmap(struct drm_i915_gem_object *obj, | |
1102 | struct reloc_cache *cache, | |
102a0a90 | 1103 | unsigned long pageno) |
ad5d95e4 DA |
1104 | { |
1105 | void *vaddr; | |
102a0a90 | 1106 | struct page *page; |
ad5d95e4 DA |
1107 | |
1108 | if (cache->vaddr) { | |
1109 | kunmap_atomic(unmask_page(cache->vaddr)); | |
1110 | } else { | |
1111 | unsigned int flushes; | |
1112 | int err; | |
1113 | ||
1114 | err = i915_gem_object_prepare_write(obj, &flushes); | |
1115 | if (err) | |
1116 | return ERR_PTR(err); | |
1117 | ||
1118 | BUILD_BUG_ON(KMAP & CLFLUSH_FLAGS); | |
1119 | BUILD_BUG_ON((KMAP | CLFLUSH_FLAGS) & PAGE_MASK); | |
1120 | ||
1121 | cache->vaddr = flushes | KMAP; | |
1122 | cache->node.mm = (void *)obj; | |
1123 | if (flushes) | |
1124 | mb(); | |
1125 | } | |
1126 | ||
102a0a90 ML |
1127 | page = i915_gem_object_get_page(obj, pageno); |
1128 | if (!obj->mm.dirty) | |
1129 | set_page_dirty(page); | |
1130 | ||
1131 | vaddr = kmap_atomic(page); | |
ad5d95e4 | 1132 | cache->vaddr = unmask_flags(cache->vaddr) | (unsigned long)vaddr; |
102a0a90 | 1133 | cache->page = pageno; |
ad5d95e4 DA |
1134 | |
1135 | return vaddr; | |
1136 | } | |
1137 | ||
1138 | static void *reloc_iomap(struct drm_i915_gem_object *obj, | |
47b08693 | 1139 | struct i915_execbuffer *eb, |
ad5d95e4 DA |
1140 | unsigned long page) |
1141 | { | |
47b08693 | 1142 | struct reloc_cache *cache = &eb->reloc_cache; |
ad5d95e4 DA |
1143 | struct i915_ggtt *ggtt = cache_to_ggtt(cache); |
1144 | unsigned long offset; | |
1145 | void *vaddr; | |
1146 | ||
1147 | if (cache->vaddr) { | |
1148 | intel_gt_flush_ggtt_writes(ggtt->vm.gt); | |
1149 | io_mapping_unmap_atomic((void __force __iomem *) unmask_page(cache->vaddr)); | |
1150 | } else { | |
1151 | struct i915_vma *vma; | |
1152 | int err; | |
1153 | ||
1154 | if (i915_gem_object_is_tiled(obj)) | |
1155 | return ERR_PTR(-EINVAL); | |
1156 | ||
1157 | if (use_cpu_reloc(cache, obj)) | |
1158 | return NULL; | |
1159 | ||
ad5d95e4 | 1160 | err = i915_gem_object_set_to_gtt_domain(obj, true); |
ad5d95e4 DA |
1161 | if (err) |
1162 | return ERR_PTR(err); | |
1163 | ||
47b08693 ML |
1164 | vma = i915_gem_object_ggtt_pin_ww(obj, &eb->ww, NULL, 0, 0, |
1165 | PIN_MAPPABLE | | |
1166 | PIN_NONBLOCK /* NOWARN */ | | |
1167 | PIN_NOEVICT); | |
1168 | if (vma == ERR_PTR(-EDEADLK)) | |
1169 | return vma; | |
1170 | ||
ad5d95e4 DA |
1171 | if (IS_ERR(vma)) { |
1172 | memset(&cache->node, 0, sizeof(cache->node)); | |
1173 | mutex_lock(&ggtt->vm.mutex); | |
1174 | err = drm_mm_insert_node_in_range | |
1175 | (&ggtt->vm.mm, &cache->node, | |
1176 | PAGE_SIZE, 0, I915_COLOR_UNEVICTABLE, | |
1177 | 0, ggtt->mappable_end, | |
1178 | DRM_MM_INSERT_LOW); | |
1179 | mutex_unlock(&ggtt->vm.mutex); | |
1180 | if (err) /* no inactive aperture space, use cpu reloc */ | |
1181 | return NULL; | |
1182 | } else { | |
1183 | cache->node.start = vma->node.start; | |
1184 | cache->node.mm = (void *)vma; | |
1185 | } | |
1186 | } | |
1187 | ||
1188 | offset = cache->node.start; | |
1189 | if (drm_mm_node_allocated(&cache->node)) { | |
1190 | ggtt->vm.insert_page(&ggtt->vm, | |
1191 | i915_gem_object_get_dma_address(obj, page), | |
1192 | offset, I915_CACHE_NONE, 0); | |
1193 | } else { | |
1194 | offset += page << PAGE_SHIFT; | |
1195 | } | |
1196 | ||
1197 | vaddr = (void __force *)io_mapping_map_atomic_wc(&ggtt->iomap, | |
1198 | offset); | |
1199 | cache->page = page; | |
1200 | cache->vaddr = (unsigned long)vaddr; | |
1201 | ||
1202 | return vaddr; | |
1203 | } | |
1204 | ||
1205 | static void *reloc_vaddr(struct drm_i915_gem_object *obj, | |
47b08693 | 1206 | struct i915_execbuffer *eb, |
ad5d95e4 DA |
1207 | unsigned long page) |
1208 | { | |
47b08693 | 1209 | struct reloc_cache *cache = &eb->reloc_cache; |
ad5d95e4 DA |
1210 | void *vaddr; |
1211 | ||
1212 | if (cache->page == page) { | |
1213 | vaddr = unmask_page(cache->vaddr); | |
1214 | } else { | |
1215 | vaddr = NULL; | |
1216 | if ((cache->vaddr & KMAP) == 0) | |
47b08693 | 1217 | vaddr = reloc_iomap(obj, eb, page); |
ad5d95e4 DA |
1218 | if (!vaddr) |
1219 | vaddr = reloc_kmap(obj, cache, page); | |
1220 | } | |
1221 | ||
1222 | return vaddr; | |
1223 | } | |
1224 | ||
1225 | static void clflush_write32(u32 *addr, u32 value, unsigned int flushes) | |
1226 | { | |
1227 | if (unlikely(flushes & (CLFLUSH_BEFORE | CLFLUSH_AFTER))) { | |
1228 | if (flushes & CLFLUSH_BEFORE) { | |
1229 | clflushopt(addr); | |
1230 | mb(); | |
1231 | } | |
1232 | ||
1233 | *addr = value; | |
1234 | ||
1235 | /* | |
1236 | * Writes to the same cacheline are serialised by the CPU | |
1237 | * (including clflush). On the write path, we only require | |
1238 | * that it hits memory in an orderly fashion and place | |
1239 | * mb barriers at the start and end of the relocation phase | |
1240 | * to ensure ordering of clflush wrt to the system. | |
1241 | */ | |
1242 | if (flushes & CLFLUSH_AFTER) | |
1243 | clflushopt(addr); | |
1244 | } else | |
1245 | *addr = value; | |
1246 | } | |
1247 | ||
e3d29130 | 1248 | static u64 |
ad5d95e4 | 1249 | relocate_entry(struct i915_vma *vma, |
e3d29130 | 1250 | const struct drm_i915_gem_relocation_entry *reloc, |
ad5d95e4 | 1251 | struct i915_execbuffer *eb, |
e3d29130 CW |
1252 | const struct i915_vma *target) |
1253 | { | |
1254 | u64 target_addr = relocation_target(reloc, target); | |
ad5d95e4 | 1255 | u64 offset = reloc->offset; |
ce13c78f DV |
1256 | bool wide = eb->reloc_cache.use_64bit_reloc; |
1257 | void *vaddr; | |
ad5d95e4 DA |
1258 | |
1259 | repeat: | |
ce13c78f DV |
1260 | vaddr = reloc_vaddr(vma->obj, eb, |
1261 | offset >> PAGE_SHIFT); | |
1262 | if (IS_ERR(vaddr)) | |
1263 | return PTR_ERR(vaddr); | |
1264 | ||
1265 | GEM_BUG_ON(!IS_ALIGNED(offset, sizeof(u32))); | |
1266 | clflush_write32(vaddr + offset_in_page(offset), | |
1267 | lower_32_bits(target_addr), | |
1268 | eb->reloc_cache.vaddr); | |
1269 | ||
1270 | if (wide) { | |
1271 | offset += sizeof(u32); | |
1272 | target_addr >>= 32; | |
1273 | wide = false; | |
1274 | goto repeat; | |
ad5d95e4 | 1275 | } |
edf4427b | 1276 | |
2889caa9 | 1277 | return target->node.start | UPDATE; |
edf4427b | 1278 | } |
edf4427b | 1279 | |
2889caa9 CW |
1280 | static u64 |
1281 | eb_relocate_entry(struct i915_execbuffer *eb, | |
7d6236bb | 1282 | struct eb_vma *ev, |
2889caa9 | 1283 | const struct drm_i915_gem_relocation_entry *reloc) |
54cf91dc | 1284 | { |
baa89ba3 | 1285 | struct drm_i915_private *i915 = eb->i915; |
7d6236bb | 1286 | struct eb_vma *target; |
2889caa9 | 1287 | int err; |
54cf91dc | 1288 | |
67731b87 | 1289 | /* we've already hold a reference to all valid objects */ |
507d977f CW |
1290 | target = eb_get_vma(eb, reloc->target_handle); |
1291 | if (unlikely(!target)) | |
54cf91dc | 1292 | return -ENOENT; |
e844b990 | 1293 | |
54cf91dc | 1294 | /* Validate that the target is in a valid r/w GPU domain */ |
b8f7ab17 | 1295 | if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) { |
baa89ba3 | 1296 | drm_dbg(&i915->drm, "reloc with multiple write domains: " |
507d977f | 1297 | "target %d offset %d " |
54cf91dc | 1298 | "read %08x write %08x", |
507d977f | 1299 | reloc->target_handle, |
54cf91dc CW |
1300 | (int) reloc->offset, |
1301 | reloc->read_domains, | |
1302 | reloc->write_domain); | |
8b78f0e5 | 1303 | return -EINVAL; |
54cf91dc | 1304 | } |
4ca4a250 DV |
1305 | if (unlikely((reloc->write_domain | reloc->read_domains) |
1306 | & ~I915_GEM_GPU_DOMAINS)) { | |
baa89ba3 | 1307 | drm_dbg(&i915->drm, "reloc with read/write non-GPU domains: " |
507d977f | 1308 | "target %d offset %d " |
54cf91dc | 1309 | "read %08x write %08x", |
507d977f | 1310 | reloc->target_handle, |
54cf91dc CW |
1311 | (int) reloc->offset, |
1312 | reloc->read_domains, | |
1313 | reloc->write_domain); | |
8b78f0e5 | 1314 | return -EINVAL; |
54cf91dc | 1315 | } |
54cf91dc | 1316 | |
2889caa9 | 1317 | if (reloc->write_domain) { |
7d6236bb | 1318 | target->flags |= EXEC_OBJECT_WRITE; |
507d977f | 1319 | |
2889caa9 CW |
1320 | /* |
1321 | * Sandybridge PPGTT errata: We need a global gtt mapping | |
1322 | * for MI and pipe_control writes because the gpu doesn't | |
1323 | * properly redirect them through the ppgtt for non_secure | |
1324 | * batchbuffers. | |
1325 | */ | |
1326 | if (reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION && | |
40e1956e | 1327 | GRAPHICS_VER(eb->i915) == 6) { |
7d6236bb CW |
1328 | err = i915_vma_bind(target->vma, |
1329 | target->vma->obj->cache_level, | |
2850748e | 1330 | PIN_GLOBAL, NULL); |
ea97c4ca | 1331 | if (err) |
2889caa9 CW |
1332 | return err; |
1333 | } | |
507d977f | 1334 | } |
54cf91dc | 1335 | |
2889caa9 CW |
1336 | /* |
1337 | * If the relocation already has the right value in it, no | |
54cf91dc CW |
1338 | * more work needs to be done. |
1339 | */ | |
ad5d95e4 DA |
1340 | if (!DBG_FORCE_RELOC && |
1341 | gen8_canonical_addr(target->vma->node.start) == reloc->presumed_offset) | |
67731b87 | 1342 | return 0; |
54cf91dc CW |
1343 | |
1344 | /* Check that the relocation address is valid... */ | |
3c94ceee | 1345 | if (unlikely(reloc->offset > |
7d6236bb | 1346 | ev->vma->size - (eb->reloc_cache.use_64bit_reloc ? 8 : 4))) { |
baa89ba3 | 1347 | drm_dbg(&i915->drm, "Relocation beyond object bounds: " |
507d977f CW |
1348 | "target %d offset %d size %d.\n", |
1349 | reloc->target_handle, | |
1350 | (int)reloc->offset, | |
7d6236bb | 1351 | (int)ev->vma->size); |
8b78f0e5 | 1352 | return -EINVAL; |
54cf91dc | 1353 | } |
b8f7ab17 | 1354 | if (unlikely(reloc->offset & 3)) { |
baa89ba3 | 1355 | drm_dbg(&i915->drm, "Relocation not 4-byte aligned: " |
507d977f CW |
1356 | "target %d offset %d.\n", |
1357 | reloc->target_handle, | |
1358 | (int)reloc->offset); | |
8b78f0e5 | 1359 | return -EINVAL; |
54cf91dc CW |
1360 | } |
1361 | ||
071750e5 CW |
1362 | /* |
1363 | * If we write into the object, we need to force the synchronisation | |
1364 | * barrier, either with an asynchronous clflush or if we executed the | |
1365 | * patching using the GPU (though that should be serialised by the | |
1366 | * timeline). To be completely sure, and since we are required to | |
1367 | * do relocations we are already stalling, disable the user's opt | |
0519bcb1 | 1368 | * out of our synchronisation. |
071750e5 | 1369 | */ |
7d6236bb | 1370 | ev->flags &= ~EXEC_OBJECT_ASYNC; |
071750e5 | 1371 | |
54cf91dc | 1372 | /* and update the user's relocation entry */ |
ad5d95e4 | 1373 | return relocate_entry(ev->vma, reloc, eb, target->vma); |
54cf91dc CW |
1374 | } |
1375 | ||
7d6236bb | 1376 | static int eb_relocate_vma(struct i915_execbuffer *eb, struct eb_vma *ev) |
54cf91dc | 1377 | { |
1d83f442 | 1378 | #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry)) |
2889caa9 | 1379 | struct drm_i915_gem_relocation_entry stack[N_RELOC(512)]; |
7d6236bb | 1380 | const struct drm_i915_gem_exec_object2 *entry = ev->exec; |
e94f7856 CW |
1381 | struct drm_i915_gem_relocation_entry __user *urelocs = |
1382 | u64_to_user_ptr(entry->relocs_ptr); | |
1383 | unsigned long remain = entry->relocation_count; | |
54cf91dc | 1384 | |
e94f7856 | 1385 | if (unlikely(remain > N_RELOC(ULONG_MAX))) |
2889caa9 | 1386 | return -EINVAL; |
ebc0808f | 1387 | |
2889caa9 CW |
1388 | /* |
1389 | * We must check that the entire relocation array is safe | |
1390 | * to read. However, if the array is not writable the user loses | |
1391 | * the updated relocation values. | |
1392 | */ | |
e94f7856 | 1393 | if (unlikely(!access_ok(urelocs, remain * sizeof(*urelocs)))) |
2889caa9 CW |
1394 | return -EFAULT; |
1395 | ||
1396 | do { | |
1397 | struct drm_i915_gem_relocation_entry *r = stack; | |
1398 | unsigned int count = | |
e94f7856 | 1399 | min_t(unsigned long, remain, ARRAY_SIZE(stack)); |
2889caa9 | 1400 | unsigned int copied; |
1d83f442 | 1401 | |
2889caa9 CW |
1402 | /* |
1403 | * This is the fast path and we cannot handle a pagefault | |
ebc0808f CW |
1404 | * whilst holding the struct mutex lest the user pass in the |
1405 | * relocations contained within a mmaped bo. For in such a case | |
1406 | * we, the page fault handler would call i915_gem_fault() and | |
1407 | * we would try to acquire the struct mutex again. Obviously | |
1408 | * this is bad and so lockdep complains vehemently. | |
1409 | */ | |
fd1500fc ML |
1410 | pagefault_disable(); |
1411 | copied = __copy_from_user_inatomic(r, urelocs, count * sizeof(r[0])); | |
1412 | pagefault_enable(); | |
ad5d95e4 DA |
1413 | if (unlikely(copied)) { |
1414 | remain = -EFAULT; | |
1415 | goto out; | |
1416 | } | |
54cf91dc | 1417 | |
2889caa9 | 1418 | remain -= count; |
1d83f442 | 1419 | do { |
7d6236bb | 1420 | u64 offset = eb_relocate_entry(eb, ev, r); |
54cf91dc | 1421 | |
2889caa9 CW |
1422 | if (likely(offset == 0)) { |
1423 | } else if ((s64)offset < 0) { | |
ad5d95e4 DA |
1424 | remain = (int)offset; |
1425 | goto out; | |
2889caa9 CW |
1426 | } else { |
1427 | /* | |
1428 | * Note that reporting an error now | |
1429 | * leaves everything in an inconsistent | |
1430 | * state as we have *already* changed | |
1431 | * the relocation value inside the | |
1432 | * object. As we have not changed the | |
1433 | * reloc.presumed_offset or will not | |
1434 | * change the execobject.offset, on the | |
1435 | * call we may not rewrite the value | |
1436 | * inside the object, leaving it | |
1437 | * dangling and causing a GPU hang. Unless | |
1438 | * userspace dynamically rebuilds the | |
1439 | * relocations on each execbuf rather than | |
1440 | * presume a static tree. | |
1441 | * | |
1442 | * We did previously check if the relocations | |
1443 | * were writable (access_ok), an error now | |
1444 | * would be a strange race with mprotect, | |
1445 | * having already demonstrated that we | |
1446 | * can read from this userspace address. | |
1447 | */ | |
1448 | offset = gen8_canonical_addr(offset & ~UPDATE); | |
97a37c91 CW |
1449 | __put_user(offset, |
1450 | &urelocs[r - stack].presumed_offset); | |
1d83f442 | 1451 | } |
2889caa9 CW |
1452 | } while (r++, --count); |
1453 | urelocs += ARRAY_SIZE(stack); | |
1454 | } while (remain); | |
ad5d95e4 | 1455 | out: |
c43ce123 | 1456 | reloc_cache_reset(&eb->reloc_cache, eb); |
ad5d95e4 | 1457 | return remain; |
54cf91dc CW |
1458 | } |
1459 | ||
fd1500fc ML |
1460 | static int |
1461 | eb_relocate_vma_slow(struct i915_execbuffer *eb, struct eb_vma *ev) | |
54cf91dc | 1462 | { |
fd1500fc ML |
1463 | const struct drm_i915_gem_exec_object2 *entry = ev->exec; |
1464 | struct drm_i915_gem_relocation_entry *relocs = | |
1465 | u64_to_ptr(typeof(*relocs), entry->relocs_ptr); | |
1466 | unsigned int i; | |
003d8b91 CW |
1467 | int err; |
1468 | ||
fd1500fc ML |
1469 | for (i = 0; i < entry->relocation_count; i++) { |
1470 | u64 offset = eb_relocate_entry(eb, ev, &relocs[i]); | |
003d8b91 | 1471 | |
fd1500fc ML |
1472 | if ((s64)offset < 0) { |
1473 | err = (int)offset; | |
1474 | goto err; | |
1475 | } | |
ef398881 | 1476 | } |
fd1500fc ML |
1477 | err = 0; |
1478 | err: | |
c43ce123 | 1479 | reloc_cache_reset(&eb->reloc_cache, eb); |
fd1500fc ML |
1480 | return err; |
1481 | } | |
2889caa9 | 1482 | |
fd1500fc ML |
1483 | static int check_relocations(const struct drm_i915_gem_exec_object2 *entry) |
1484 | { | |
1485 | const char __user *addr, *end; | |
1486 | unsigned long size; | |
1487 | char __maybe_unused c; | |
2889caa9 | 1488 | |
fd1500fc ML |
1489 | size = entry->relocation_count; |
1490 | if (size == 0) | |
1491 | return 0; | |
0e97fbb0 | 1492 | |
fd1500fc ML |
1493 | if (size > N_RELOC(ULONG_MAX)) |
1494 | return -EINVAL; | |
2889caa9 | 1495 | |
fd1500fc ML |
1496 | addr = u64_to_user_ptr(entry->relocs_ptr); |
1497 | size *= sizeof(struct drm_i915_gem_relocation_entry); | |
1498 | if (!access_ok(addr, size)) | |
1499 | return -EFAULT; | |
1500 | ||
1501 | end = addr + size; | |
1502 | for (; addr < end; addr += PAGE_SIZE) { | |
1503 | int err = __get_user(c, addr); | |
1504 | if (err) | |
1505 | return err; | |
1506 | } | |
1507 | return __get_user(c, end - 1); | |
2889caa9 CW |
1508 | } |
1509 | ||
fd1500fc | 1510 | static int eb_copy_relocations(const struct i915_execbuffer *eb) |
2889caa9 | 1511 | { |
fd1500fc | 1512 | struct drm_i915_gem_relocation_entry *relocs; |
2889caa9 CW |
1513 | const unsigned int count = eb->buffer_count; |
1514 | unsigned int i; | |
fd1500fc | 1515 | int err; |
54cf91dc | 1516 | |
2889caa9 | 1517 | for (i = 0; i < count; i++) { |
fd1500fc ML |
1518 | const unsigned int nreloc = eb->exec[i].relocation_count; |
1519 | struct drm_i915_gem_relocation_entry __user *urelocs; | |
1520 | unsigned long size; | |
1521 | unsigned long copied; | |
6951e589 | 1522 | |
fd1500fc ML |
1523 | if (nreloc == 0) |
1524 | continue; | |
6951e589 | 1525 | |
fd1500fc ML |
1526 | err = check_relocations(&eb->exec[i]); |
1527 | if (err) | |
1528 | goto err; | |
6951e589 | 1529 | |
fd1500fc ML |
1530 | urelocs = u64_to_user_ptr(eb->exec[i].relocs_ptr); |
1531 | size = nreloc * sizeof(*relocs); | |
6951e589 | 1532 | |
fd1500fc ML |
1533 | relocs = kvmalloc_array(size, 1, GFP_KERNEL); |
1534 | if (!relocs) { | |
1535 | err = -ENOMEM; | |
1536 | goto err; | |
6951e589 | 1537 | } |
fd1500fc ML |
1538 | |
1539 | /* copy_from_user is limited to < 4GiB */ | |
1540 | copied = 0; | |
1541 | do { | |
1542 | unsigned int len = | |
1543 | min_t(u64, BIT_ULL(31), size - copied); | |
1544 | ||
1545 | if (__copy_from_user((char *)relocs + copied, | |
1546 | (char __user *)urelocs + copied, | |
1547 | len)) | |
1548 | goto end; | |
1549 | ||
1550 | copied += len; | |
1551 | } while (copied < size); | |
1552 | ||
1553 | /* | |
1554 | * As we do not update the known relocation offsets after | |
1555 | * relocating (due to the complexities in lock handling), | |
1556 | * we need to mark them as invalid now so that we force the | |
1557 | * relocation processing next time. Just in case the target | |
1558 | * object is evicted and then rebound into its old | |
1559 | * presumed_offset before the next execbuffer - if that | |
1560 | * happened we would make the mistake of assuming that the | |
1561 | * relocations were valid. | |
1562 | */ | |
1563 | if (!user_access_begin(urelocs, size)) | |
1564 | goto end; | |
1565 | ||
1566 | for (copied = 0; copied < nreloc; copied++) | |
1567 | unsafe_put_user(-1, | |
1568 | &urelocs[copied].presumed_offset, | |
1569 | end_user); | |
1570 | user_access_end(); | |
1571 | ||
1572 | eb->exec[i].relocs_ptr = (uintptr_t)relocs; | |
1573 | } | |
1574 | ||
1575 | return 0; | |
1576 | ||
1577 | end_user: | |
1578 | user_access_end(); | |
1579 | end: | |
1580 | kvfree(relocs); | |
1581 | err = -EFAULT; | |
1582 | err: | |
1583 | while (i--) { | |
1584 | relocs = u64_to_ptr(typeof(*relocs), eb->exec[i].relocs_ptr); | |
1585 | if (eb->exec[i].relocation_count) | |
1586 | kvfree(relocs); | |
1587 | } | |
1588 | return err; | |
1589 | } | |
1590 | ||
1591 | static int eb_prefault_relocations(const struct i915_execbuffer *eb) | |
1592 | { | |
1593 | const unsigned int count = eb->buffer_count; | |
1594 | unsigned int i; | |
1595 | ||
1596 | for (i = 0; i < count; i++) { | |
1597 | int err; | |
1598 | ||
1599 | err = check_relocations(&eb->exec[i]); | |
1600 | if (err) | |
1601 | return err; | |
1602 | } | |
1603 | ||
1604 | return 0; | |
1605 | } | |
1606 | ||
ed29c269 ML |
1607 | static int eb_reinit_userptr(struct i915_execbuffer *eb) |
1608 | { | |
1609 | const unsigned int count = eb->buffer_count; | |
1610 | unsigned int i; | |
1611 | int ret; | |
1612 | ||
1613 | if (likely(!(eb->args->flags & __EXEC_USERPTR_USED))) | |
1614 | return 0; | |
1615 | ||
1616 | for (i = 0; i < count; i++) { | |
1617 | struct eb_vma *ev = &eb->vma[i]; | |
1618 | ||
1619 | if (!i915_gem_object_is_userptr(ev->vma->obj)) | |
1620 | continue; | |
1621 | ||
1622 | ret = i915_gem_object_userptr_submit_init(ev->vma->obj); | |
1623 | if (ret) | |
1624 | return ret; | |
1625 | ||
1626 | ev->flags |= __EXEC_OBJECT_USERPTR_INIT; | |
1627 | } | |
1628 | ||
1629 | return 0; | |
1630 | } | |
1631 | ||
2bf541ff ML |
1632 | static noinline int eb_relocate_parse_slow(struct i915_execbuffer *eb, |
1633 | struct i915_request *rq) | |
fd1500fc ML |
1634 | { |
1635 | bool have_copy = false; | |
1636 | struct eb_vma *ev; | |
1637 | int err = 0; | |
1638 | ||
1639 | repeat: | |
1640 | if (signal_pending(current)) { | |
1641 | err = -ERESTARTSYS; | |
1642 | goto out; | |
6951e589 | 1643 | } |
fd1500fc | 1644 | |
c43ce123 | 1645 | /* We may process another execbuffer during the unlock... */ |
b4b9731b | 1646 | eb_release_vmas(eb, false); |
c43ce123 ML |
1647 | i915_gem_ww_ctx_fini(&eb->ww); |
1648 | ||
2bf541ff ML |
1649 | if (rq) { |
1650 | /* nonblocking is always false */ | |
1651 | if (i915_request_wait(rq, I915_WAIT_INTERRUPTIBLE, | |
1652 | MAX_SCHEDULE_TIMEOUT) < 0) { | |
1653 | i915_request_put(rq); | |
1654 | rq = NULL; | |
1655 | ||
1656 | err = -EINTR; | |
1657 | goto err_relock; | |
1658 | } | |
1659 | ||
1660 | i915_request_put(rq); | |
1661 | rq = NULL; | |
1662 | } | |
1663 | ||
fd1500fc ML |
1664 | /* |
1665 | * We take 3 passes through the slowpatch. | |
1666 | * | |
1667 | * 1 - we try to just prefault all the user relocation entries and | |
1668 | * then attempt to reuse the atomic pagefault disabled fast path again. | |
1669 | * | |
1670 | * 2 - we copy the user entries to a local buffer here outside of the | |
1671 | * local and allow ourselves to wait upon any rendering before | |
1672 | * relocations | |
1673 | * | |
1674 | * 3 - we already have a local copy of the relocation entries, but | |
1675 | * were interrupted (EAGAIN) whilst waiting for the objects, try again. | |
1676 | */ | |
1677 | if (!err) { | |
1678 | err = eb_prefault_relocations(eb); | |
1679 | } else if (!have_copy) { | |
1680 | err = eb_copy_relocations(eb); | |
1681 | have_copy = err == 0; | |
1682 | } else { | |
1683 | cond_resched(); | |
1684 | err = 0; | |
1685 | } | |
1686 | ||
2bf541ff | 1687 | if (!err) |
ed29c269 | 1688 | err = eb_reinit_userptr(eb); |
fd1500fc | 1689 | |
2bf541ff | 1690 | err_relock: |
c43ce123 | 1691 | i915_gem_ww_ctx_init(&eb->ww, true); |
fd1500fc ML |
1692 | if (err) |
1693 | goto out; | |
1694 | ||
c43ce123 ML |
1695 | /* reacquire the objects */ |
1696 | repeat_validate: | |
2bf541ff ML |
1697 | rq = eb_pin_engine(eb, false); |
1698 | if (IS_ERR(rq)) { | |
1699 | err = PTR_ERR(rq); | |
47b08693 | 1700 | rq = NULL; |
2bf541ff ML |
1701 | goto err; |
1702 | } | |
1703 | ||
1704 | /* We didn't throttle, should be NULL */ | |
1705 | GEM_WARN_ON(rq); | |
1706 | ||
c43ce123 | 1707 | err = eb_validate_vmas(eb); |
fd1500fc | 1708 | if (err) |
c43ce123 ML |
1709 | goto err; |
1710 | ||
1711 | GEM_BUG_ON(!eb->batch); | |
fd1500fc ML |
1712 | |
1713 | list_for_each_entry(ev, &eb->relocs, reloc_link) { | |
1714 | if (!have_copy) { | |
fd1500fc | 1715 | err = eb_relocate_vma(eb, ev); |
fd1500fc ML |
1716 | if (err) |
1717 | break; | |
1718 | } else { | |
1719 | err = eb_relocate_vma_slow(eb, ev); | |
1720 | if (err) | |
1721 | break; | |
1722 | } | |
1723 | } | |
1724 | ||
c43ce123 ML |
1725 | if (err == -EDEADLK) |
1726 | goto err; | |
1727 | ||
fd1500fc ML |
1728 | if (err && !have_copy) |
1729 | goto repeat; | |
1730 | ||
1731 | if (err) | |
1732 | goto err; | |
1733 | ||
8e4ba491 ML |
1734 | /* as last step, parse the command buffer */ |
1735 | err = eb_parse(eb); | |
1736 | if (err) | |
1737 | goto err; | |
1738 | ||
fd1500fc ML |
1739 | /* |
1740 | * Leave the user relocations as are, this is the painfully slow path, | |
1741 | * and we want to avoid the complication of dropping the lock whilst | |
1742 | * having buffers reserved in the aperture and so causing spurious | |
1743 | * ENOSPC for random operations. | |
1744 | */ | |
1745 | ||
1746 | err: | |
c43ce123 | 1747 | if (err == -EDEADLK) { |
b4b9731b | 1748 | eb_release_vmas(eb, false); |
c43ce123 ML |
1749 | err = i915_gem_ww_ctx_backoff(&eb->ww); |
1750 | if (!err) | |
1751 | goto repeat_validate; | |
1752 | } | |
1753 | ||
fd1500fc ML |
1754 | if (err == -EAGAIN) |
1755 | goto repeat; | |
1756 | ||
1757 | out: | |
1758 | if (have_copy) { | |
1759 | const unsigned int count = eb->buffer_count; | |
1760 | unsigned int i; | |
1761 | ||
1762 | for (i = 0; i < count; i++) { | |
1763 | const struct drm_i915_gem_exec_object2 *entry = | |
1764 | &eb->exec[i]; | |
1765 | struct drm_i915_gem_relocation_entry *relocs; | |
1766 | ||
1767 | if (!entry->relocation_count) | |
1768 | continue; | |
1769 | ||
1770 | relocs = u64_to_ptr(typeof(*relocs), entry->relocs_ptr); | |
1771 | kvfree(relocs); | |
1772 | } | |
1773 | } | |
1774 | ||
2bf541ff ML |
1775 | if (rq) |
1776 | i915_request_put(rq); | |
1777 | ||
fd1500fc ML |
1778 | return err; |
1779 | } | |
1780 | ||
8e4ba491 | 1781 | static int eb_relocate_parse(struct i915_execbuffer *eb) |
54cf91dc | 1782 | { |
003d8b91 | 1783 | int err; |
2bf541ff ML |
1784 | struct i915_request *rq = NULL; |
1785 | bool throttle = true; | |
003d8b91 | 1786 | |
c43ce123 | 1787 | retry: |
2bf541ff ML |
1788 | rq = eb_pin_engine(eb, throttle); |
1789 | if (IS_ERR(rq)) { | |
1790 | err = PTR_ERR(rq); | |
1791 | rq = NULL; | |
1792 | if (err != -EDEADLK) | |
1793 | return err; | |
1794 | ||
1795 | goto err; | |
1796 | } | |
1797 | ||
1798 | if (rq) { | |
1799 | bool nonblock = eb->file->filp->f_flags & O_NONBLOCK; | |
1800 | ||
1801 | /* Need to drop all locks now for throttling, take slowpath */ | |
1802 | err = i915_request_wait(rq, I915_WAIT_INTERRUPTIBLE, 0); | |
1803 | if (err == -ETIME) { | |
1804 | if (nonblock) { | |
1805 | err = -EWOULDBLOCK; | |
1806 | i915_request_put(rq); | |
1807 | goto err; | |
1808 | } | |
1809 | goto slow; | |
1810 | } | |
1811 | i915_request_put(rq); | |
1812 | rq = NULL; | |
1813 | } | |
1814 | ||
1815 | /* only throttle once, even if we didn't need to throttle */ | |
1816 | throttle = false; | |
1817 | ||
c43ce123 ML |
1818 | err = eb_validate_vmas(eb); |
1819 | if (err == -EAGAIN) | |
1820 | goto slow; | |
1821 | else if (err) | |
1822 | goto err; | |
2889caa9 CW |
1823 | |
1824 | /* The objects are in their final locations, apply the relocations. */ | |
1825 | if (eb->args->flags & __EXEC_HAS_RELOC) { | |
7d6236bb | 1826 | struct eb_vma *ev; |
2889caa9 | 1827 | |
7d6236bb | 1828 | list_for_each_entry(ev, &eb->relocs, reloc_link) { |
7dc8f114 CW |
1829 | err = eb_relocate_vma(eb, ev); |
1830 | if (err) | |
fd1500fc | 1831 | break; |
2889caa9 | 1832 | } |
fd1500fc | 1833 | |
c43ce123 ML |
1834 | if (err == -EDEADLK) |
1835 | goto err; | |
1836 | else if (err) | |
1837 | goto slow; | |
1838 | } | |
1839 | ||
1840 | if (!err) | |
1841 | err = eb_parse(eb); | |
1842 | ||
1843 | err: | |
1844 | if (err == -EDEADLK) { | |
b4b9731b | 1845 | eb_release_vmas(eb, false); |
c43ce123 ML |
1846 | err = i915_gem_ww_ctx_backoff(&eb->ww); |
1847 | if (!err) | |
1848 | goto retry; | |
2889caa9 CW |
1849 | } |
1850 | ||
c43ce123 ML |
1851 | return err; |
1852 | ||
1853 | slow: | |
2bf541ff | 1854 | err = eb_relocate_parse_slow(eb, rq); |
c43ce123 ML |
1855 | if (err) |
1856 | /* | |
1857 | * If the user expects the execobject.offset and | |
1858 | * reloc.presumed_offset to be an exact match, | |
1859 | * as for using NO_RELOC, then we cannot update | |
1860 | * the execobject.offset until we have completed | |
1861 | * relocation. | |
1862 | */ | |
1863 | eb->args->flags &= ~__EXEC_HAS_RELOC; | |
1864 | ||
1865 | return err; | |
2889caa9 CW |
1866 | } |
1867 | ||
2889caa9 CW |
1868 | static int eb_move_to_gpu(struct i915_execbuffer *eb) |
1869 | { | |
1870 | const unsigned int count = eb->buffer_count; | |
c43ce123 | 1871 | unsigned int i = count; |
6951e589 | 1872 | int err = 0; |
6951e589 CW |
1873 | |
1874 | while (i--) { | |
7d6236bb CW |
1875 | struct eb_vma *ev = &eb->vma[i]; |
1876 | struct i915_vma *vma = ev->vma; | |
1877 | unsigned int flags = ev->flags; | |
27173f1f | 1878 | struct drm_i915_gem_object *obj = vma->obj; |
03ade511 | 1879 | |
6951e589 CW |
1880 | assert_vma_held(vma); |
1881 | ||
c7c6e46f | 1882 | if (flags & EXEC_OBJECT_CAPTURE) { |
e61e0f51 | 1883 | struct i915_capture_list *capture; |
b0fd47ad CW |
1884 | |
1885 | capture = kmalloc(sizeof(*capture), GFP_KERNEL); | |
6951e589 CW |
1886 | if (capture) { |
1887 | capture->next = eb->request->capture_list; | |
1888 | capture->vma = vma; | |
1889 | eb->request->capture_list = capture; | |
1890 | } | |
b0fd47ad CW |
1891 | } |
1892 | ||
b8f55be6 CW |
1893 | /* |
1894 | * If the GPU is not _reading_ through the CPU cache, we need | |
1895 | * to make sure that any writes (both previous GPU writes from | |
1896 | * before a change in snooping levels and normal CPU writes) | |
1897 | * caught in that cache are flushed to main memory. | |
1898 | * | |
1899 | * We want to say | |
1900 | * obj->cache_dirty && | |
1901 | * !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ) | |
1902 | * but gcc's optimiser doesn't handle that as well and emits | |
1903 | * two jumps instead of one. Maybe one day... | |
1904 | */ | |
1905 | if (unlikely(obj->cache_dirty & ~obj->cache_coherent)) { | |
0f46daa1 | 1906 | if (i915_gem_clflush_object(obj, 0)) |
c7c6e46f | 1907 | flags &= ~EXEC_OBJECT_ASYNC; |
0f46daa1 CW |
1908 | } |
1909 | ||
6951e589 CW |
1910 | if (err == 0 && !(flags & EXEC_OBJECT_ASYNC)) { |
1911 | err = i915_request_await_object | |
1912 | (eb->request, obj, flags & EXEC_OBJECT_WRITE); | |
1913 | } | |
2889caa9 | 1914 | |
6951e589 | 1915 | if (err == 0) |
bfaae47d ML |
1916 | err = i915_vma_move_to_active(vma, eb->request, |
1917 | flags | __EXEC_OBJECT_NO_RESERVE); | |
c59a333f | 1918 | } |
0f1dd022 | 1919 | |
ed29c269 ML |
1920 | #ifdef CONFIG_MMU_NOTIFIER |
1921 | if (!err && (eb->args->flags & __EXEC_USERPTR_USED)) { | |
b4b9731b | 1922 | read_lock(&eb->i915->mm.notifier_lock); |
ed29c269 ML |
1923 | |
1924 | /* | |
1925 | * count is always at least 1, otherwise __EXEC_USERPTR_USED | |
1926 | * could not have been set | |
1927 | */ | |
1928 | for (i = 0; i < count; i++) { | |
1929 | struct eb_vma *ev = &eb->vma[i]; | |
1930 | struct drm_i915_gem_object *obj = ev->vma->obj; | |
1931 | ||
1932 | if (!i915_gem_object_is_userptr(obj)) | |
1933 | continue; | |
1934 | ||
1935 | err = i915_gem_object_userptr_submit_done(obj); | |
1936 | if (err) | |
1937 | break; | |
1938 | } | |
1939 | ||
b4b9731b | 1940 | read_unlock(&eb->i915->mm.notifier_lock); |
ed29c269 ML |
1941 | } |
1942 | #endif | |
1943 | ||
6951e589 CW |
1944 | if (unlikely(err)) |
1945 | goto err_skip; | |
1946 | ||
dcd79934 | 1947 | /* Unconditionally flush any chipset caches (for streaming writes). */ |
baea429d | 1948 | intel_gt_chipset_flush(eb->engine->gt); |
2113184c | 1949 | return 0; |
6951e589 CW |
1950 | |
1951 | err_skip: | |
36e191f0 | 1952 | i915_request_set_error_once(eb->request, err); |
6951e589 | 1953 | return err; |
54cf91dc CW |
1954 | } |
1955 | ||
00aff3f6 | 1956 | static int i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec) |
54cf91dc | 1957 | { |
650bc635 | 1958 | if (exec->flags & __I915_EXEC_ILLEGAL_FLAGS) |
00aff3f6 | 1959 | return -EINVAL; |
ed5982e6 | 1960 | |
2f5945bc | 1961 | /* Kernel clipping was a DRI1 misfeature */ |
cda9edd0 LL |
1962 | if (!(exec->flags & (I915_EXEC_FENCE_ARRAY | |
1963 | I915_EXEC_USE_EXTENSIONS))) { | |
cf6e7bac | 1964 | if (exec->num_cliprects || exec->cliprects_ptr) |
00aff3f6 | 1965 | return -EINVAL; |
cf6e7bac | 1966 | } |
2f5945bc CW |
1967 | |
1968 | if (exec->DR4 == 0xffffffff) { | |
1969 | DRM_DEBUG("UXA submitting garbage DR4, fixing up\n"); | |
1970 | exec->DR4 = 0; | |
1971 | } | |
1972 | if (exec->DR1 || exec->DR4) | |
00aff3f6 | 1973 | return -EINVAL; |
2f5945bc CW |
1974 | |
1975 | if ((exec->batch_start_offset | exec->batch_len) & 0x7) | |
00aff3f6 | 1976 | return -EINVAL; |
2f5945bc | 1977 | |
00aff3f6 | 1978 | return 0; |
54cf91dc CW |
1979 | } |
1980 | ||
e61e0f51 | 1981 | static int i915_reset_gen7_sol_offsets(struct i915_request *rq) |
ae662d31 | 1982 | { |
73dec95e TU |
1983 | u32 *cs; |
1984 | int i; | |
ae662d31 | 1985 | |
40e1956e | 1986 | if (GRAPHICS_VER(rq->engine->i915) != 7 || rq->engine->id != RCS0) { |
5a833995 | 1987 | drm_dbg(&rq->engine->i915->drm, "sol reset is gen7/rcs only\n"); |
9d662da8 DV |
1988 | return -EINVAL; |
1989 | } | |
ae662d31 | 1990 | |
e61e0f51 | 1991 | cs = intel_ring_begin(rq, 4 * 2 + 2); |
73dec95e TU |
1992 | if (IS_ERR(cs)) |
1993 | return PTR_ERR(cs); | |
ae662d31 | 1994 | |
2889caa9 | 1995 | *cs++ = MI_LOAD_REGISTER_IMM(4); |
ae662d31 | 1996 | for (i = 0; i < 4; i++) { |
73dec95e TU |
1997 | *cs++ = i915_mmio_reg_offset(GEN7_SO_WRITE_OFFSET(i)); |
1998 | *cs++ = 0; | |
ae662d31 | 1999 | } |
2889caa9 | 2000 | *cs++ = MI_NOOP; |
e61e0f51 | 2001 | intel_ring_advance(rq, cs); |
ae662d31 EA |
2002 | |
2003 | return 0; | |
2004 | } | |
2005 | ||
4f7af194 | 2006 | static struct i915_vma * |
47b08693 ML |
2007 | shadow_batch_pin(struct i915_execbuffer *eb, |
2008 | struct drm_i915_gem_object *obj, | |
32d94048 CW |
2009 | struct i915_address_space *vm, |
2010 | unsigned int flags) | |
4f7af194 | 2011 | { |
b291ce0a | 2012 | struct i915_vma *vma; |
b291ce0a | 2013 | int err; |
4f7af194 | 2014 | |
b291ce0a CW |
2015 | vma = i915_vma_instance(obj, vm, NULL); |
2016 | if (IS_ERR(vma)) | |
2017 | return vma; | |
2018 | ||
47b08693 | 2019 | err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, flags); |
b291ce0a CW |
2020 | if (err) |
2021 | return ERR_PTR(err); | |
2022 | ||
2023 | return vma; | |
4f7af194 JB |
2024 | } |
2025 | ||
47b08693 ML |
2026 | static struct i915_vma *eb_dispatch_secure(struct i915_execbuffer *eb, struct i915_vma *vma) |
2027 | { | |
2028 | /* | |
2029 | * snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure | |
2030 | * batch" bit. Hence we need to pin secure batches into the global gtt. | |
2031 | * hsw should have this fixed, but bdw mucks it up again. */ | |
2032 | if (eb->batch_flags & I915_DISPATCH_SECURE) | |
2033 | return i915_gem_object_ggtt_pin_ww(vma->obj, &eb->ww, NULL, 0, 0, 0); | |
2034 | ||
2035 | return NULL; | |
2036 | } | |
2037 | ||
51696691 | 2038 | static int eb_parse(struct i915_execbuffer *eb) |
71745376 | 2039 | { |
baa89ba3 | 2040 | struct drm_i915_private *i915 = eb->i915; |
c43ce123 | 2041 | struct intel_gt_buffer_pool_node *pool = eb->batch_pool; |
47b08693 | 2042 | struct i915_vma *shadow, *trampoline, *batch; |
d5e87821 | 2043 | unsigned long len; |
2889caa9 | 2044 | int err; |
71745376 | 2045 | |
47b08693 ML |
2046 | if (!eb_use_cmdparser(eb)) { |
2047 | batch = eb_dispatch_secure(eb, eb->batch->vma); | |
2048 | if (IS_ERR(batch)) | |
2049 | return PTR_ERR(batch); | |
2050 | ||
2051 | goto secure_batch; | |
2052 | } | |
51696691 | 2053 | |
32d94048 CW |
2054 | len = eb->batch_len; |
2055 | if (!CMDPARSER_USES_GGTT(eb->i915)) { | |
2056 | /* | |
2057 | * ppGTT backed shadow buffers must be mapped RO, to prevent | |
2058 | * post-scan tampering | |
2059 | */ | |
2060 | if (!eb->context->vm->has_read_only) { | |
baa89ba3 WK |
2061 | drm_dbg(&i915->drm, |
2062 | "Cannot prevent post-scan tampering without RO capable vm\n"); | |
32d94048 CW |
2063 | return -EINVAL; |
2064 | } | |
2065 | } else { | |
2066 | len += I915_CMD_PARSER_TRAMPOLINE_SIZE; | |
2067 | } | |
d5e87821 CW |
2068 | if (unlikely(len < eb->batch_len)) /* last paranoid check of overflow */ |
2069 | return -EINVAL; | |
32d94048 | 2070 | |
c43ce123 | 2071 | if (!pool) { |
8f47c8c3 MA |
2072 | pool = intel_gt_get_buffer_pool(eb->engine->gt, len, |
2073 | I915_MAP_WB); | |
c43ce123 ML |
2074 | if (IS_ERR(pool)) |
2075 | return PTR_ERR(pool); | |
2076 | eb->batch_pool = pool; | |
2077 | } | |
71745376 | 2078 | |
c43ce123 ML |
2079 | err = i915_gem_object_lock(pool->obj, &eb->ww); |
2080 | if (err) | |
2081 | goto err; | |
71745376 | 2082 | |
47b08693 | 2083 | shadow = shadow_batch_pin(eb, pool->obj, eb->context->vm, PIN_USER); |
32d94048 CW |
2084 | if (IS_ERR(shadow)) { |
2085 | err = PTR_ERR(shadow); | |
f8c08d8f | 2086 | goto err; |
51696691 | 2087 | } |
c9398775 | 2088 | intel_gt_buffer_pool_mark_used(pool); |
32d94048 | 2089 | i915_gem_object_set_readonly(shadow->obj); |
57a78ca4 | 2090 | shadow->private = pool; |
32d94048 CW |
2091 | |
2092 | trampoline = NULL; | |
2093 | if (CMDPARSER_USES_GGTT(eb->i915)) { | |
2094 | trampoline = shadow; | |
2095 | ||
47b08693 | 2096 | shadow = shadow_batch_pin(eb, pool->obj, |
32d94048 CW |
2097 | &eb->engine->gt->ggtt->vm, |
2098 | PIN_GLOBAL); | |
2099 | if (IS_ERR(shadow)) { | |
2100 | err = PTR_ERR(shadow); | |
2101 | shadow = trampoline; | |
2102 | goto err_shadow; | |
2103 | } | |
57a78ca4 | 2104 | shadow->private = pool; |
32d94048 CW |
2105 | |
2106 | eb->batch_flags |= I915_DISPATCH_SECURE; | |
2107 | } | |
f8c08d8f | 2108 | |
47b08693 ML |
2109 | batch = eb_dispatch_secure(eb, shadow); |
2110 | if (IS_ERR(batch)) { | |
2111 | err = PTR_ERR(batch); | |
2112 | goto err_trampoline; | |
2113 | } | |
2114 | ||
93b71330 JE |
2115 | err = dma_resv_reserve_shared(shadow->resv, 1); |
2116 | if (err) | |
2117 | goto err_trampoline; | |
2118 | ||
2119 | err = intel_engine_cmd_parser(eb->engine, | |
2120 | eb->batch->vma, | |
2121 | eb->batch_start_offset, | |
2122 | eb->batch_len, | |
2123 | shadow, trampoline); | |
32d94048 | 2124 | if (err) |
47b08693 | 2125 | goto err_unpin_batch; |
71745376 | 2126 | |
7d6236bb | 2127 | eb->batch = &eb->vma[eb->buffer_count++]; |
47b08693 ML |
2128 | eb->batch->vma = i915_vma_get(shadow); |
2129 | eb->batch->flags = __EXEC_OBJECT_HAS_PIN; | |
71745376 | 2130 | |
32d94048 | 2131 | eb->trampoline = trampoline; |
4f7af194 | 2132 | eb->batch_start_offset = 0; |
4f7af194 | 2133 | |
47b08693 ML |
2134 | secure_batch: |
2135 | if (batch) { | |
2136 | eb->batch = &eb->vma[eb->buffer_count++]; | |
2137 | eb->batch->flags = __EXEC_OBJECT_HAS_PIN; | |
2138 | eb->batch->vma = i915_vma_get(batch); | |
2139 | } | |
51696691 | 2140 | return 0; |
b40d7378 | 2141 | |
47b08693 ML |
2142 | err_unpin_batch: |
2143 | if (batch) | |
2144 | i915_vma_unpin(batch); | |
32d94048 CW |
2145 | err_trampoline: |
2146 | if (trampoline) | |
2147 | i915_vma_unpin(trampoline); | |
2148 | err_shadow: | |
2149 | i915_vma_unpin(shadow); | |
b40d7378 | 2150 | err: |
51696691 | 2151 | return err; |
71745376 | 2152 | } |
5c6c6003 | 2153 | |
7d6236bb | 2154 | static int eb_submit(struct i915_execbuffer *eb, struct i915_vma *batch) |
78382593 | 2155 | { |
2889caa9 | 2156 | int err; |
78382593 | 2157 | |
9b3a8f55 CW |
2158 | if (intel_context_nopreempt(eb->context)) |
2159 | __set_bit(I915_FENCE_FLAG_NOPREEMPT, &eb->request->fence.flags); | |
2160 | ||
2889caa9 CW |
2161 | err = eb_move_to_gpu(eb); |
2162 | if (err) | |
2163 | return err; | |
78382593 | 2164 | |
650bc635 | 2165 | if (eb->args->flags & I915_EXEC_GEN7_SOL_RESET) { |
2889caa9 CW |
2166 | err = i915_reset_gen7_sol_offsets(eb->request); |
2167 | if (err) | |
2168 | return err; | |
78382593 OM |
2169 | } |
2170 | ||
85474441 CW |
2171 | /* |
2172 | * After we completed waiting for other engines (using HW semaphores) | |
2173 | * then we can signal that this request/batch is ready to run. This | |
2174 | * allows us to determine if the batch is still waiting on the GPU | |
2175 | * or actually running by checking the breadcrumb. | |
2176 | */ | |
2177 | if (eb->engine->emit_init_breadcrumb) { | |
2178 | err = eb->engine->emit_init_breadcrumb(eb->request); | |
2179 | if (err) | |
2180 | return err; | |
2181 | } | |
2182 | ||
2889caa9 | 2183 | err = eb->engine->emit_bb_start(eb->request, |
7d6236bb | 2184 | batch->node.start + |
650bc635 CW |
2185 | eb->batch_start_offset, |
2186 | eb->batch_len, | |
2889caa9 CW |
2187 | eb->batch_flags); |
2188 | if (err) | |
2189 | return err; | |
78382593 | 2190 | |
32d94048 CW |
2191 | if (eb->trampoline) { |
2192 | GEM_BUG_ON(eb->batch_start_offset); | |
2193 | err = eb->engine->emit_bb_start(eb->request, | |
2194 | eb->trampoline->node.start + | |
2195 | eb->batch_len, | |
2196 | 0, 0); | |
2197 | if (err) | |
2198 | return err; | |
2199 | } | |
2200 | ||
2f5945bc | 2201 | return 0; |
78382593 OM |
2202 | } |
2203 | ||
d5b2a3a4 CW |
2204 | static int num_vcs_engines(const struct drm_i915_private *i915) |
2205 | { | |
bb80d878 | 2206 | return hweight_long(VDBOX_MASK(&i915->gt)); |
d5b2a3a4 CW |
2207 | } |
2208 | ||
204bcfef | 2209 | /* |
a8ebba75 | 2210 | * Find one BSD ring to dispatch the corresponding BSD command. |
c80ff16e | 2211 | * The engine index is returned. |
a8ebba75 | 2212 | */ |
de1add36 | 2213 | static unsigned int |
c80ff16e CW |
2214 | gen8_dispatch_bsd_engine(struct drm_i915_private *dev_priv, |
2215 | struct drm_file *file) | |
a8ebba75 | 2216 | { |
a8ebba75 ZY |
2217 | struct drm_i915_file_private *file_priv = file->driver_priv; |
2218 | ||
de1add36 | 2219 | /* Check whether the file_priv has already selected one ring. */ |
6f633402 | 2220 | if ((int)file_priv->bsd_engine < 0) |
1a07e86c CW |
2221 | file_priv->bsd_engine = |
2222 | get_random_int() % num_vcs_engines(dev_priv); | |
d23db88c | 2223 | |
c80ff16e | 2224 | return file_priv->bsd_engine; |
d23db88c CW |
2225 | } |
2226 | ||
5e2a0419 | 2227 | static const enum intel_engine_id user_ring_map[] = { |
8a68d464 CW |
2228 | [I915_EXEC_DEFAULT] = RCS0, |
2229 | [I915_EXEC_RENDER] = RCS0, | |
2230 | [I915_EXEC_BLT] = BCS0, | |
2231 | [I915_EXEC_BSD] = VCS0, | |
2232 | [I915_EXEC_VEBOX] = VECS0 | |
de1add36 TU |
2233 | }; |
2234 | ||
2bf541ff | 2235 | static struct i915_request *eb_throttle(struct i915_execbuffer *eb, struct intel_context *ce) |
e5dadff4 CW |
2236 | { |
2237 | struct intel_ring *ring = ce->ring; | |
2238 | struct intel_timeline *tl = ce->timeline; | |
2239 | struct i915_request *rq; | |
2240 | ||
2241 | /* | |
2242 | * Completely unscientific finger-in-the-air estimates for suitable | |
2243 | * maximum user request size (to avoid blocking) and then backoff. | |
2244 | */ | |
2245 | if (intel_ring_update_space(ring) >= PAGE_SIZE) | |
2246 | return NULL; | |
2247 | ||
2248 | /* | |
2249 | * Find a request that after waiting upon, there will be at least half | |
2250 | * the ring available. The hysteresis allows us to compete for the | |
2251 | * shared ring and should mean that we sleep less often prior to | |
2252 | * claiming our resources, but not so long that the ring completely | |
2253 | * drains before we can submit our next request. | |
2254 | */ | |
2255 | list_for_each_entry(rq, &tl->requests, link) { | |
2256 | if (rq->ring != ring) | |
2257 | continue; | |
2258 | ||
2259 | if (__intel_ring_space(rq->postfix, | |
2260 | ring->emit, ring->size) > ring->size / 2) | |
2261 | break; | |
2262 | } | |
2263 | if (&rq->link == &tl->requests) | |
2264 | return NULL; /* weird, we will check again later for real */ | |
2265 | ||
2266 | return i915_request_get(rq); | |
2267 | } | |
2268 | ||
2bf541ff | 2269 | static struct i915_request *eb_pin_engine(struct i915_execbuffer *eb, bool throttle) |
e5dadff4 | 2270 | { |
2bf541ff | 2271 | struct intel_context *ce = eb->context; |
e5dadff4 | 2272 | struct intel_timeline *tl; |
2bf541ff | 2273 | struct i915_request *rq = NULL; |
e5dadff4 CW |
2274 | int err; |
2275 | ||
2bf541ff | 2276 | GEM_BUG_ON(eb->args->flags & __EXEC_ENGINE_PINNED); |
8f2a1057 | 2277 | |
9f3ccd40 | 2278 | if (unlikely(intel_context_is_banned(ce))) |
2bf541ff | 2279 | return ERR_PTR(-EIO); |
9f3ccd40 | 2280 | |
8f2a1057 CW |
2281 | /* |
2282 | * Pinning the contexts may generate requests in order to acquire | |
2283 | * GGTT space, so do this first before we reserve a seqno for | |
2284 | * ourselves. | |
2285 | */ | |
47b08693 | 2286 | err = intel_context_pin_ww(ce, &eb->ww); |
fa9f6681 | 2287 | if (err) |
2bf541ff | 2288 | return ERR_PTR(err); |
8f2a1057 | 2289 | |
a4e57f90 CW |
2290 | /* |
2291 | * Take a local wakeref for preparing to dispatch the execbuf as | |
2292 | * we expect to access the hardware fairly frequently in the | |
2293 | * process, and require the engine to be kept awake between accesses. | |
2294 | * Upon dispatch, we acquire another prolonged wakeref that we hold | |
2295 | * until the timeline is idle, which in turn releases the wakeref | |
2296 | * taken on the engine, and the parent device. | |
2297 | */ | |
e5dadff4 CW |
2298 | tl = intel_context_timeline_lock(ce); |
2299 | if (IS_ERR(tl)) { | |
2bf541ff ML |
2300 | intel_context_unpin(ce); |
2301 | return ERR_CAST(tl); | |
e5dadff4 | 2302 | } |
a4e57f90 CW |
2303 | |
2304 | intel_context_enter(ce); | |
2bf541ff ML |
2305 | if (throttle) |
2306 | rq = eb_throttle(eb, ce); | |
e5dadff4 CW |
2307 | intel_context_timeline_unlock(tl); |
2308 | ||
2bf541ff ML |
2309 | eb->args->flags |= __EXEC_ENGINE_PINNED; |
2310 | return rq; | |
8f2a1057 CW |
2311 | } |
2312 | ||
e5dadff4 | 2313 | static void eb_unpin_engine(struct i915_execbuffer *eb) |
8f2a1057 | 2314 | { |
a4e57f90 | 2315 | struct intel_context *ce = eb->context; |
75d0a7f3 | 2316 | struct intel_timeline *tl = ce->timeline; |
a4e57f90 | 2317 | |
2bf541ff ML |
2318 | if (!(eb->args->flags & __EXEC_ENGINE_PINNED)) |
2319 | return; | |
2320 | ||
2321 | eb->args->flags &= ~__EXEC_ENGINE_PINNED; | |
2322 | ||
a4e57f90 CW |
2323 | mutex_lock(&tl->mutex); |
2324 | intel_context_exit(ce); | |
2325 | mutex_unlock(&tl->mutex); | |
2326 | ||
2850748e | 2327 | intel_context_unpin(ce); |
8f2a1057 | 2328 | } |
de1add36 | 2329 | |
5e2a0419 | 2330 | static unsigned int |
b49a7d51 | 2331 | eb_select_legacy_ring(struct i915_execbuffer *eb) |
de1add36 | 2332 | { |
8f2a1057 | 2333 | struct drm_i915_private *i915 = eb->i915; |
b49a7d51 | 2334 | struct drm_i915_gem_execbuffer2 *args = eb->args; |
de1add36 | 2335 | unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK; |
de1add36 | 2336 | |
5e2a0419 CW |
2337 | if (user_ring_id != I915_EXEC_BSD && |
2338 | (args->flags & I915_EXEC_BSD_MASK)) { | |
baa89ba3 WK |
2339 | drm_dbg(&i915->drm, |
2340 | "execbuf with non bsd ring but with invalid " | |
2341 | "bsd dispatch flags: %d\n", (int)(args->flags)); | |
5e2a0419 | 2342 | return -1; |
de1add36 TU |
2343 | } |
2344 | ||
d5b2a3a4 | 2345 | if (user_ring_id == I915_EXEC_BSD && num_vcs_engines(i915) > 1) { |
de1add36 TU |
2346 | unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK; |
2347 | ||
2348 | if (bsd_idx == I915_EXEC_BSD_DEFAULT) { | |
b49a7d51 | 2349 | bsd_idx = gen8_dispatch_bsd_engine(i915, eb->file); |
de1add36 TU |
2350 | } else if (bsd_idx >= I915_EXEC_BSD_RING1 && |
2351 | bsd_idx <= I915_EXEC_BSD_RING2) { | |
d9da6aa0 | 2352 | bsd_idx >>= I915_EXEC_BSD_SHIFT; |
de1add36 TU |
2353 | bsd_idx--; |
2354 | } else { | |
baa89ba3 WK |
2355 | drm_dbg(&i915->drm, |
2356 | "execbuf with unknown bsd ring: %u\n", | |
2357 | bsd_idx); | |
5e2a0419 | 2358 | return -1; |
de1add36 TU |
2359 | } |
2360 | ||
5e2a0419 | 2361 | return _VCS(bsd_idx); |
de1add36 TU |
2362 | } |
2363 | ||
5e2a0419 | 2364 | if (user_ring_id >= ARRAY_SIZE(user_ring_map)) { |
baa89ba3 WK |
2365 | drm_dbg(&i915->drm, "execbuf with unknown ring: %u\n", |
2366 | user_ring_id); | |
5e2a0419 | 2367 | return -1; |
de1add36 TU |
2368 | } |
2369 | ||
5e2a0419 CW |
2370 | return user_ring_map[user_ring_id]; |
2371 | } | |
2372 | ||
2373 | static int | |
2bf541ff | 2374 | eb_select_engine(struct i915_execbuffer *eb) |
5e2a0419 CW |
2375 | { |
2376 | struct intel_context *ce; | |
2377 | unsigned int idx; | |
2378 | int err; | |
2379 | ||
976b55f0 | 2380 | if (i915_gem_context_user_engines(eb->gem_context)) |
b49a7d51 | 2381 | idx = eb->args->flags & I915_EXEC_RING_MASK; |
976b55f0 | 2382 | else |
b49a7d51 | 2383 | idx = eb_select_legacy_ring(eb); |
5e2a0419 CW |
2384 | |
2385 | ce = i915_gem_context_get_engine(eb->gem_context, idx); | |
2386 | if (IS_ERR(ce)) | |
2387 | return PTR_ERR(ce); | |
2388 | ||
2bf541ff | 2389 | intel_gt_pm_get(ce->engine->gt); |
5e2a0419 | 2390 | |
2bf541ff ML |
2391 | if (!test_bit(CONTEXT_ALLOC_BIT, &ce->flags)) { |
2392 | err = intel_context_alloc_state(ce); | |
2393 | if (err) | |
2394 | goto err; | |
2395 | } | |
2396 | ||
2397 | /* | |
2398 | * ABI: Before userspace accesses the GPU (e.g. execbuffer), report | |
2399 | * EIO if the GPU is already wedged. | |
2400 | */ | |
2401 | err = intel_gt_terminally_wedged(ce->engine->gt); | |
2402 | if (err) | |
2403 | goto err; | |
2404 | ||
2405 | eb->context = ce; | |
2406 | eb->engine = ce->engine; | |
2407 | ||
2408 | /* | |
2409 | * Make sure engine pool stays alive even if we call intel_context_put | |
2410 | * during ww handling. The pool is destroyed when last pm reference | |
2411 | * is dropped, which breaks our -EDEADLK handling. | |
2412 | */ | |
2413 | return err; | |
2414 | ||
2415 | err: | |
2416 | intel_gt_pm_put(ce->engine->gt); | |
2417 | intel_context_put(ce); | |
5e2a0419 | 2418 | return err; |
de1add36 TU |
2419 | } |
2420 | ||
2bf541ff ML |
2421 | static void |
2422 | eb_put_engine(struct i915_execbuffer *eb) | |
2423 | { | |
2424 | intel_gt_pm_put(eb->engine->gt); | |
2425 | intel_context_put(eb->context); | |
2426 | } | |
2427 | ||
cf6e7bac | 2428 | static void |
13149e8b | 2429 | __free_fence_array(struct eb_fence *fences, unsigned int n) |
cf6e7bac | 2430 | { |
13149e8b | 2431 | while (n--) { |
cda9edd0 | 2432 | drm_syncobj_put(ptr_mask_bits(fences[n].syncobj, 2)); |
13149e8b LL |
2433 | dma_fence_put(fences[n].dma_fence); |
2434 | kfree(fences[n].chain_fence); | |
2435 | } | |
cf6e7bac JE |
2436 | kvfree(fences); |
2437 | } | |
2438 | ||
cda9edd0 | 2439 | static int |
13149e8b LL |
2440 | add_timeline_fence_array(struct i915_execbuffer *eb, |
2441 | const struct drm_i915_gem_execbuffer_ext_timeline_fences *timeline_fences) | |
cf6e7bac | 2442 | { |
13149e8b LL |
2443 | struct drm_i915_gem_exec_fence __user *user_fences; |
2444 | u64 __user *user_values; | |
2445 | struct eb_fence *f; | |
2446 | u64 nfences; | |
2447 | int err = 0; | |
cf6e7bac | 2448 | |
13149e8b LL |
2449 | nfences = timeline_fences->fence_count; |
2450 | if (!nfences) | |
cda9edd0 | 2451 | return 0; |
cf6e7bac | 2452 | |
d710fc16 CW |
2453 | /* Check multiplication overflow for access_ok() and kvmalloc_array() */ |
2454 | BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long)); | |
2455 | if (nfences > min_t(unsigned long, | |
13149e8b LL |
2456 | ULONG_MAX / sizeof(*user_fences), |
2457 | SIZE_MAX / sizeof(*f)) - eb->num_fences) | |
cda9edd0 | 2458 | return -EINVAL; |
cf6e7bac | 2459 | |
13149e8b LL |
2460 | user_fences = u64_to_user_ptr(timeline_fences->handles_ptr); |
2461 | if (!access_ok(user_fences, nfences * sizeof(*user_fences))) | |
2462 | return -EFAULT; | |
2463 | ||
2464 | user_values = u64_to_user_ptr(timeline_fences->values_ptr); | |
2465 | if (!access_ok(user_values, nfences * sizeof(*user_values))) | |
cda9edd0 | 2466 | return -EFAULT; |
cf6e7bac | 2467 | |
13149e8b LL |
2468 | f = krealloc(eb->fences, |
2469 | (eb->num_fences + nfences) * sizeof(*f), | |
2470 | __GFP_NOWARN | GFP_KERNEL); | |
2471 | if (!f) | |
cda9edd0 | 2472 | return -ENOMEM; |
cf6e7bac | 2473 | |
13149e8b LL |
2474 | eb->fences = f; |
2475 | f += eb->num_fences; | |
2476 | ||
2477 | BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) & | |
2478 | ~__I915_EXEC_FENCE_UNKNOWN_FLAGS); | |
2479 | ||
2480 | while (nfences--) { | |
2481 | struct drm_i915_gem_exec_fence user_fence; | |
cf6e7bac | 2482 | struct drm_syncobj *syncobj; |
13149e8b LL |
2483 | struct dma_fence *fence = NULL; |
2484 | u64 point; | |
2485 | ||
2486 | if (__copy_from_user(&user_fence, | |
2487 | user_fences++, | |
2488 | sizeof(user_fence))) | |
2489 | return -EFAULT; | |
2490 | ||
2491 | if (user_fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS) | |
2492 | return -EINVAL; | |
2493 | ||
2494 | if (__get_user(point, user_values++)) | |
2495 | return -EFAULT; | |
2496 | ||
2497 | syncobj = drm_syncobj_find(eb->file, user_fence.handle); | |
2498 | if (!syncobj) { | |
2499 | DRM_DEBUG("Invalid syncobj handle provided\n"); | |
2500 | return -ENOENT; | |
2501 | } | |
2502 | ||
2503 | fence = drm_syncobj_fence_get(syncobj); | |
cf6e7bac | 2504 | |
13149e8b LL |
2505 | if (!fence && user_fence.flags && |
2506 | !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) { | |
2507 | DRM_DEBUG("Syncobj handle has no fence\n"); | |
2508 | drm_syncobj_put(syncobj); | |
2509 | return -EINVAL; | |
cf6e7bac JE |
2510 | } |
2511 | ||
13149e8b LL |
2512 | if (fence) |
2513 | err = dma_fence_chain_find_seqno(&fence, point); | |
2514 | ||
2515 | if (err && !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) { | |
2516 | DRM_DEBUG("Syncobj handle missing requested point %llu\n", point); | |
da1ea128 | 2517 | dma_fence_put(fence); |
13149e8b LL |
2518 | drm_syncobj_put(syncobj); |
2519 | return err; | |
2520 | } | |
2521 | ||
2522 | /* | |
2523 | * A point might have been signaled already and | |
2524 | * garbage collected from the timeline. In this case | |
2525 | * just ignore the point and carry on. | |
2526 | */ | |
2527 | if (!fence && !(user_fence.flags & I915_EXEC_FENCE_SIGNAL)) { | |
2528 | drm_syncobj_put(syncobj); | |
2529 | continue; | |
2530 | } | |
2531 | ||
2532 | /* | |
2533 | * For timeline syncobjs we need to preallocate chains for | |
2534 | * later signaling. | |
2535 | */ | |
2536 | if (point != 0 && user_fence.flags & I915_EXEC_FENCE_SIGNAL) { | |
2537 | /* | |
2538 | * Waiting and signaling the same point (when point != | |
2539 | * 0) would break the timeline. | |
2540 | */ | |
2541 | if (user_fence.flags & I915_EXEC_FENCE_WAIT) { | |
2542 | DRM_DEBUG("Trying to wait & signal the same timeline point.\n"); | |
2543 | dma_fence_put(fence); | |
2544 | drm_syncobj_put(syncobj); | |
2545 | return -EINVAL; | |
2546 | } | |
2547 | ||
2548 | f->chain_fence = | |
2549 | kmalloc(sizeof(*f->chain_fence), | |
2550 | GFP_KERNEL); | |
2551 | if (!f->chain_fence) { | |
2552 | drm_syncobj_put(syncobj); | |
2553 | dma_fence_put(fence); | |
2554 | return -ENOMEM; | |
2555 | } | |
2556 | } else { | |
2557 | f->chain_fence = NULL; | |
ebcaa1ff TU |
2558 | } |
2559 | ||
13149e8b LL |
2560 | f->syncobj = ptr_pack_bits(syncobj, user_fence.flags, 2); |
2561 | f->dma_fence = fence; | |
2562 | f->value = point; | |
2563 | f++; | |
2564 | eb->num_fences++; | |
2565 | } | |
2566 | ||
2567 | return 0; | |
2568 | } | |
2569 | ||
2570 | static int add_fence_array(struct i915_execbuffer *eb) | |
2571 | { | |
2572 | struct drm_i915_gem_execbuffer2 *args = eb->args; | |
2573 | struct drm_i915_gem_exec_fence __user *user; | |
2574 | unsigned long num_fences = args->num_cliprects; | |
2575 | struct eb_fence *f; | |
2576 | ||
2577 | if (!(args->flags & I915_EXEC_FENCE_ARRAY)) | |
2578 | return 0; | |
2579 | ||
2580 | if (!num_fences) | |
2581 | return 0; | |
2582 | ||
2583 | /* Check multiplication overflow for access_ok() and kvmalloc_array() */ | |
2584 | BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long)); | |
2585 | if (num_fences > min_t(unsigned long, | |
2586 | ULONG_MAX / sizeof(*user), | |
2587 | SIZE_MAX / sizeof(*f) - eb->num_fences)) | |
2588 | return -EINVAL; | |
2589 | ||
2590 | user = u64_to_user_ptr(args->cliprects_ptr); | |
2591 | if (!access_ok(user, num_fences * sizeof(*user))) | |
2592 | return -EFAULT; | |
2593 | ||
2594 | f = krealloc(eb->fences, | |
2595 | (eb->num_fences + num_fences) * sizeof(*f), | |
2596 | __GFP_NOWARN | GFP_KERNEL); | |
2597 | if (!f) | |
2598 | return -ENOMEM; | |
2599 | ||
2600 | eb->fences = f; | |
2601 | f += eb->num_fences; | |
2602 | while (num_fences--) { | |
2603 | struct drm_i915_gem_exec_fence user_fence; | |
2604 | struct drm_syncobj *syncobj; | |
2605 | struct dma_fence *fence = NULL; | |
2606 | ||
2607 | if (__copy_from_user(&user_fence, user++, sizeof(user_fence))) | |
2608 | return -EFAULT; | |
2609 | ||
2610 | if (user_fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS) | |
2611 | return -EINVAL; | |
2612 | ||
2613 | syncobj = drm_syncobj_find(eb->file, user_fence.handle); | |
cf6e7bac JE |
2614 | if (!syncobj) { |
2615 | DRM_DEBUG("Invalid syncobj handle provided\n"); | |
13149e8b LL |
2616 | return -ENOENT; |
2617 | } | |
2618 | ||
2619 | if (user_fence.flags & I915_EXEC_FENCE_WAIT) { | |
2620 | fence = drm_syncobj_fence_get(syncobj); | |
2621 | if (!fence) { | |
2622 | DRM_DEBUG("Syncobj handle has no fence\n"); | |
2623 | drm_syncobj_put(syncobj); | |
2624 | return -EINVAL; | |
2625 | } | |
cf6e7bac JE |
2626 | } |
2627 | ||
ebcaa1ff TU |
2628 | BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) & |
2629 | ~__I915_EXEC_FENCE_UNKNOWN_FLAGS); | |
2630 | ||
13149e8b LL |
2631 | f->syncobj = ptr_pack_bits(syncobj, user_fence.flags, 2); |
2632 | f->dma_fence = fence; | |
2633 | f->value = 0; | |
2634 | f->chain_fence = NULL; | |
2635 | f++; | |
2636 | eb->num_fences++; | |
cf6e7bac JE |
2637 | } |
2638 | ||
cda9edd0 | 2639 | return 0; |
13149e8b | 2640 | } |
cf6e7bac | 2641 | |
13149e8b LL |
2642 | static void put_fence_array(struct eb_fence *fences, int num_fences) |
2643 | { | |
2644 | if (fences) | |
2645 | __free_fence_array(fences, num_fences); | |
cf6e7bac JE |
2646 | } |
2647 | ||
2648 | static int | |
cda9edd0 | 2649 | await_fence_array(struct i915_execbuffer *eb) |
cf6e7bac | 2650 | { |
cf6e7bac JE |
2651 | unsigned int n; |
2652 | int err; | |
2653 | ||
13149e8b | 2654 | for (n = 0; n < eb->num_fences; n++) { |
cf6e7bac | 2655 | struct drm_syncobj *syncobj; |
cf6e7bac JE |
2656 | unsigned int flags; |
2657 | ||
cda9edd0 | 2658 | syncobj = ptr_unpack_bits(eb->fences[n].syncobj, &flags, 2); |
cf6e7bac | 2659 | |
13149e8b LL |
2660 | if (!eb->fences[n].dma_fence) |
2661 | continue; | |
cf6e7bac | 2662 | |
13149e8b LL |
2663 | err = i915_request_await_dma_fence(eb->request, |
2664 | eb->fences[n].dma_fence); | |
cf6e7bac JE |
2665 | if (err < 0) |
2666 | return err; | |
2667 | } | |
2668 | ||
2669 | return 0; | |
2670 | } | |
2671 | ||
13149e8b | 2672 | static void signal_fence_array(const struct i915_execbuffer *eb) |
cf6e7bac | 2673 | { |
cf6e7bac JE |
2674 | struct dma_fence * const fence = &eb->request->fence; |
2675 | unsigned int n; | |
2676 | ||
13149e8b | 2677 | for (n = 0; n < eb->num_fences; n++) { |
cf6e7bac JE |
2678 | struct drm_syncobj *syncobj; |
2679 | unsigned int flags; | |
2680 | ||
cda9edd0 | 2681 | syncobj = ptr_unpack_bits(eb->fences[n].syncobj, &flags, 2); |
cf6e7bac JE |
2682 | if (!(flags & I915_EXEC_FENCE_SIGNAL)) |
2683 | continue; | |
2684 | ||
13149e8b LL |
2685 | if (eb->fences[n].chain_fence) { |
2686 | drm_syncobj_add_point(syncobj, | |
2687 | eb->fences[n].chain_fence, | |
2688 | fence, | |
2689 | eb->fences[n].value); | |
2690 | /* | |
2691 | * The chain's ownership is transferred to the | |
2692 | * timeline. | |
2693 | */ | |
2694 | eb->fences[n].chain_fence = NULL; | |
2695 | } else { | |
2696 | drm_syncobj_replace_fence(syncobj, fence); | |
2697 | } | |
cf6e7bac JE |
2698 | } |
2699 | } | |
2700 | ||
13149e8b LL |
2701 | static int |
2702 | parse_timeline_fences(struct i915_user_extension __user *ext, void *data) | |
2703 | { | |
2704 | struct i915_execbuffer *eb = data; | |
2705 | struct drm_i915_gem_execbuffer_ext_timeline_fences timeline_fences; | |
2706 | ||
2707 | if (copy_from_user(&timeline_fences, ext, sizeof(timeline_fences))) | |
2708 | return -EFAULT; | |
2709 | ||
2710 | return add_timeline_fence_array(eb, &timeline_fences); | |
2711 | } | |
2712 | ||
61231f6b CW |
2713 | static void retire_requests(struct intel_timeline *tl, struct i915_request *end) |
2714 | { | |
2715 | struct i915_request *rq, *rn; | |
2716 | ||
2717 | list_for_each_entry_safe(rq, rn, &tl->requests, link) | |
2718 | if (rq == end || !i915_request_retire(rq)) | |
2719 | break; | |
2720 | } | |
2721 | ||
ba38b79e | 2722 | static int eb_request_add(struct i915_execbuffer *eb, int err) |
61231f6b CW |
2723 | { |
2724 | struct i915_request *rq = eb->request; | |
2725 | struct intel_timeline * const tl = i915_request_timeline(rq); | |
2726 | struct i915_sched_attr attr = {}; | |
2727 | struct i915_request *prev; | |
2728 | ||
2729 | lockdep_assert_held(&tl->mutex); | |
2730 | lockdep_unpin_lock(&tl->mutex, rq->cookie); | |
2731 | ||
2732 | trace_i915_request_add(rq); | |
2733 | ||
2734 | prev = __i915_request_commit(rq); | |
2735 | ||
2736 | /* Check that the context wasn't destroyed before submission */ | |
207e4a71 | 2737 | if (likely(!intel_context_is_closed(eb->context))) { |
61231f6b | 2738 | attr = eb->gem_context->sched; |
61231f6b CW |
2739 | } else { |
2740 | /* Serialise with context_close via the add_to_timeline */ | |
36e191f0 CW |
2741 | i915_request_set_error_once(rq, -ENOENT); |
2742 | __i915_request_skip(rq); | |
ba38b79e | 2743 | err = -ENOENT; /* override any transient errors */ |
61231f6b CW |
2744 | } |
2745 | ||
61231f6b | 2746 | __i915_request_queue(rq, &attr); |
61231f6b CW |
2747 | |
2748 | /* Try to clean up the client's timeline after submitting the request */ | |
2749 | if (prev) | |
2750 | retire_requests(tl, prev); | |
2751 | ||
2752 | mutex_unlock(&tl->mutex); | |
ba38b79e CW |
2753 | |
2754 | return err; | |
61231f6b CW |
2755 | } |
2756 | ||
cda9edd0 | 2757 | static const i915_user_extension_fn execbuf_extensions[] = { |
13149e8b | 2758 | [DRM_I915_GEM_EXECBUFFER_EXT_TIMELINE_FENCES] = parse_timeline_fences, |
cda9edd0 LL |
2759 | }; |
2760 | ||
2761 | static int | |
2762 | parse_execbuf2_extensions(struct drm_i915_gem_execbuffer2 *args, | |
2763 | struct i915_execbuffer *eb) | |
2764 | { | |
cda9edd0 LL |
2765 | if (!(args->flags & I915_EXEC_USE_EXTENSIONS)) |
2766 | return 0; | |
2767 | ||
2768 | /* The execbuf2 extension mechanism reuses cliprects_ptr. So we cannot | |
2769 | * have another flag also using it at the same time. | |
2770 | */ | |
2771 | if (eb->args->flags & I915_EXEC_FENCE_ARRAY) | |
2772 | return -EINVAL; | |
2773 | ||
2774 | if (args->num_cliprects != 0) | |
2775 | return -EINVAL; | |
2776 | ||
2777 | return i915_user_extensions(u64_to_user_ptr(args->cliprects_ptr), | |
2778 | execbuf_extensions, | |
2779 | ARRAY_SIZE(execbuf_extensions), | |
2780 | eb); | |
2781 | } | |
2782 | ||
54cf91dc | 2783 | static int |
650bc635 | 2784 | i915_gem_do_execbuffer(struct drm_device *dev, |
54cf91dc CW |
2785 | struct drm_file *file, |
2786 | struct drm_i915_gem_execbuffer2 *args, | |
cda9edd0 | 2787 | struct drm_i915_gem_exec_object2 *exec) |
54cf91dc | 2788 | { |
44157641 | 2789 | struct drm_i915_private *i915 = to_i915(dev); |
650bc635 | 2790 | struct i915_execbuffer eb; |
fec0445c CW |
2791 | struct dma_fence *in_fence = NULL; |
2792 | struct sync_file *out_fence = NULL; | |
7d6236bb | 2793 | struct i915_vma *batch; |
fec0445c | 2794 | int out_fence_fd = -1; |
2889caa9 | 2795 | int err; |
432e58ed | 2796 | |
74c1c694 | 2797 | BUILD_BUG_ON(__EXEC_INTERNAL_FLAGS & ~__I915_EXEC_ILLEGAL_FLAGS); |
2889caa9 CW |
2798 | BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS & |
2799 | ~__EXEC_OBJECT_UNKNOWN_FLAGS); | |
54cf91dc | 2800 | |
44157641 | 2801 | eb.i915 = i915; |
650bc635 CW |
2802 | eb.file = file; |
2803 | eb.args = args; | |
ad5d95e4 | 2804 | if (DBG_FORCE_RELOC || !(args->flags & I915_EXEC_NO_RELOC)) |
2889caa9 | 2805 | args->flags |= __EXEC_HAS_RELOC; |
c7c6e46f | 2806 | |
650bc635 | 2807 | eb.exec = exec; |
8ae275c2 ML |
2808 | eb.vma = (struct eb_vma *)(exec + args->buffer_count + 1); |
2809 | eb.vma[0].vma = NULL; | |
8e02cceb | 2810 | eb.batch_pool = NULL; |
c7c6e46f | 2811 | |
2889caa9 | 2812 | eb.invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS; |
650bc635 CW |
2813 | reloc_cache_init(&eb.reloc_cache, eb.i915); |
2814 | ||
2889caa9 | 2815 | eb.buffer_count = args->buffer_count; |
650bc635 CW |
2816 | eb.batch_start_offset = args->batch_start_offset; |
2817 | eb.batch_len = args->batch_len; | |
32d94048 | 2818 | eb.trampoline = NULL; |
650bc635 | 2819 | |
cda9edd0 | 2820 | eb.fences = NULL; |
13149e8b | 2821 | eb.num_fences = 0; |
cda9edd0 | 2822 | |
2889caa9 | 2823 | eb.batch_flags = 0; |
d7d4eedd | 2824 | if (args->flags & I915_EXEC_SECURE) { |
40e1956e | 2825 | if (GRAPHICS_VER(i915) >= 11) |
44157641 JB |
2826 | return -ENODEV; |
2827 | ||
2828 | /* Return -EPERM to trigger fallback code on old binaries. */ | |
2829 | if (!HAS_SECURE_BATCHES(i915)) | |
2830 | return -EPERM; | |
2831 | ||
b3ac9f25 | 2832 | if (!drm_is_current_master(file) || !capable(CAP_SYS_ADMIN)) |
44157641 | 2833 | return -EPERM; |
d7d4eedd | 2834 | |
2889caa9 | 2835 | eb.batch_flags |= I915_DISPATCH_SECURE; |
d7d4eedd | 2836 | } |
b45305fc | 2837 | if (args->flags & I915_EXEC_IS_PINNED) |
2889caa9 | 2838 | eb.batch_flags |= I915_DISPATCH_PINNED; |
54cf91dc | 2839 | |
13149e8b LL |
2840 | err = parse_execbuf2_extensions(args, &eb); |
2841 | if (err) | |
2842 | goto err_ext; | |
2843 | ||
2844 | err = add_fence_array(&eb); | |
2845 | if (err) | |
2846 | goto err_ext; | |
2847 | ||
889333c7 CW |
2848 | #define IN_FENCES (I915_EXEC_FENCE_IN | I915_EXEC_FENCE_SUBMIT) |
2849 | if (args->flags & IN_FENCES) { | |
2850 | if ((args->flags & IN_FENCES) == IN_FENCES) | |
2851 | return -EINVAL; | |
2852 | ||
fec0445c | 2853 | in_fence = sync_file_get_fence(lower_32_bits(args->rsvd2)); |
13149e8b LL |
2854 | if (!in_fence) { |
2855 | err = -EINVAL; | |
2856 | goto err_ext; | |
2857 | } | |
fec0445c | 2858 | } |
889333c7 | 2859 | #undef IN_FENCES |
a88b6e4c | 2860 | |
fec0445c CW |
2861 | if (args->flags & I915_EXEC_FENCE_OUT) { |
2862 | out_fence_fd = get_unused_fd_flags(O_CLOEXEC); | |
2863 | if (out_fence_fd < 0) { | |
2889caa9 | 2864 | err = out_fence_fd; |
889333c7 | 2865 | goto err_in_fence; |
fec0445c CW |
2866 | } |
2867 | } | |
2868 | ||
cda9edd0 LL |
2869 | err = eb_create(&eb); |
2870 | if (err) | |
13149e8b | 2871 | goto err_out_fence; |
cda9edd0 | 2872 | |
4d470f73 | 2873 | GEM_BUG_ON(!eb.lut_size); |
2889caa9 | 2874 | |
1acfc104 CW |
2875 | err = eb_select_context(&eb); |
2876 | if (unlikely(err)) | |
2877 | goto err_destroy; | |
2878 | ||
2bf541ff | 2879 | err = eb_select_engine(&eb); |
d6f328bf | 2880 | if (unlikely(err)) |
e5dadff4 | 2881 | goto err_context; |
d6f328bf | 2882 | |
c43ce123 ML |
2883 | err = eb_lookup_vmas(&eb); |
2884 | if (err) { | |
b4b9731b | 2885 | eb_release_vmas(&eb, true); |
c43ce123 ML |
2886 | goto err_engine; |
2887 | } | |
2888 | ||
2889 | i915_gem_ww_ctx_init(&eb.ww, true); | |
2890 | ||
8e4ba491 | 2891 | err = eb_relocate_parse(&eb); |
1f727d9e | 2892 | if (err) { |
2889caa9 CW |
2893 | /* |
2894 | * If the user expects the execobject.offset and | |
2895 | * reloc.presumed_offset to be an exact match, | |
2896 | * as for using NO_RELOC, then we cannot update | |
2897 | * the execobject.offset until we have completed | |
2898 | * relocation. | |
2899 | */ | |
2900 | args->flags &= ~__EXEC_HAS_RELOC; | |
2889caa9 | 2901 | goto err_vma; |
1f727d9e | 2902 | } |
54cf91dc | 2903 | |
c43ce123 | 2904 | ww_acquire_done(&eb.ww.ctx); |
7d6236bb | 2905 | |
7d6236bb | 2906 | batch = eb.batch->vma; |
d7d4eedd | 2907 | |
0c8dac88 | 2908 | /* Allocate a request for this batch buffer nice and early. */ |
8f2a1057 | 2909 | eb.request = i915_request_create(eb.context); |
650bc635 | 2910 | if (IS_ERR(eb.request)) { |
2889caa9 | 2911 | err = PTR_ERR(eb.request); |
47b08693 | 2912 | goto err_vma; |
26827088 | 2913 | } |
0c8dac88 | 2914 | |
00dae4d3 JE |
2915 | if (unlikely(eb.gem_context->syncobj)) { |
2916 | struct dma_fence *fence; | |
2917 | ||
2918 | fence = drm_syncobj_fence_get(eb.gem_context->syncobj); | |
2919 | err = i915_request_await_dma_fence(eb.request, fence); | |
2920 | dma_fence_put(fence); | |
2921 | if (err) | |
2922 | goto err_ext; | |
2923 | } | |
2924 | ||
fec0445c | 2925 | if (in_fence) { |
889333c7 CW |
2926 | if (args->flags & I915_EXEC_FENCE_SUBMIT) |
2927 | err = i915_request_await_execution(eb.request, | |
5ac545b8 | 2928 | in_fence); |
889333c7 CW |
2929 | else |
2930 | err = i915_request_await_dma_fence(eb.request, | |
2931 | in_fence); | |
a88b6e4c CW |
2932 | if (err < 0) |
2933 | goto err_request; | |
2934 | } | |
2935 | ||
13149e8b | 2936 | if (eb.fences) { |
cda9edd0 | 2937 | err = await_fence_array(&eb); |
cf6e7bac JE |
2938 | if (err) |
2939 | goto err_request; | |
2940 | } | |
2941 | ||
fec0445c | 2942 | if (out_fence_fd != -1) { |
650bc635 | 2943 | out_fence = sync_file_create(&eb.request->fence); |
fec0445c | 2944 | if (!out_fence) { |
2889caa9 | 2945 | err = -ENOMEM; |
fec0445c CW |
2946 | goto err_request; |
2947 | } | |
2948 | } | |
2949 | ||
2889caa9 CW |
2950 | /* |
2951 | * Whilst this request exists, batch_obj will be on the | |
17f298cf CW |
2952 | * active_list, and so will hold the active reference. Only when this |
2953 | * request is retired will the the batch_obj be moved onto the | |
2954 | * inactive_list and lose its active reference. Hence we do not need | |
2955 | * to explicitly hold another reference here. | |
2956 | */ | |
7d6236bb | 2957 | eb.request->batch = batch; |
c43ce123 ML |
2958 | if (eb.batch_pool) |
2959 | intel_gt_buffer_pool_mark_active(eb.batch_pool, eb.request); | |
5f19e2bf | 2960 | |
e61e0f51 | 2961 | trace_i915_request_queue(eb.request, eb.batch_flags); |
7d6236bb | 2962 | err = eb_submit(&eb, batch); |
ed29c269 | 2963 | |
aa9b7810 | 2964 | err_request: |
e14177f1 | 2965 | i915_request_get(eb.request); |
ba38b79e | 2966 | err = eb_request_add(&eb, err); |
c8659efa | 2967 | |
13149e8b | 2968 | if (eb.fences) |
cda9edd0 | 2969 | signal_fence_array(&eb); |
cf6e7bac | 2970 | |
fec0445c | 2971 | if (out_fence) { |
2889caa9 | 2972 | if (err == 0) { |
fec0445c | 2973 | fd_install(out_fence_fd, out_fence->file); |
b6a88e4a | 2974 | args->rsvd2 &= GENMASK_ULL(31, 0); /* keep in-fence */ |
fec0445c CW |
2975 | args->rsvd2 |= (u64)out_fence_fd << 32; |
2976 | out_fence_fd = -1; | |
2977 | } else { | |
2978 | fput(out_fence->file); | |
2979 | } | |
2980 | } | |
00dae4d3 JE |
2981 | |
2982 | if (unlikely(eb.gem_context->syncobj)) { | |
2983 | drm_syncobj_replace_fence(eb.gem_context->syncobj, | |
2984 | &eb.request->fence); | |
2985 | } | |
2986 | ||
e14177f1 | 2987 | i915_request_put(eb.request); |
54cf91dc | 2988 | |
2889caa9 | 2989 | err_vma: |
b4b9731b | 2990 | eb_release_vmas(&eb, true); |
32d94048 CW |
2991 | if (eb.trampoline) |
2992 | i915_vma_unpin(eb.trampoline); | |
c43ce123 ML |
2993 | WARN_ON(err == -EDEADLK); |
2994 | i915_gem_ww_ctx_fini(&eb.ww); | |
2995 | ||
2996 | if (eb.batch_pool) | |
2997 | intel_gt_buffer_pool_put(eb.batch_pool); | |
c43ce123 | 2998 | err_engine: |
2bf541ff | 2999 | eb_put_engine(&eb); |
a4e57f90 | 3000 | err_context: |
8f2a1057 | 3001 | i915_gem_context_put(eb.gem_context); |
1acfc104 | 3002 | err_destroy: |
2889caa9 | 3003 | eb_destroy(&eb); |
4d470f73 | 3004 | err_out_fence: |
fec0445c CW |
3005 | if (out_fence_fd != -1) |
3006 | put_unused_fd(out_fence_fd); | |
4a04e371 | 3007 | err_in_fence: |
fec0445c | 3008 | dma_fence_put(in_fence); |
13149e8b LL |
3009 | err_ext: |
3010 | put_fence_array(eb.fences, eb.num_fences); | |
2889caa9 | 3011 | return err; |
54cf91dc CW |
3012 | } |
3013 | ||
d710fc16 CW |
3014 | static size_t eb_element_size(void) |
3015 | { | |
8ae275c2 | 3016 | return sizeof(struct drm_i915_gem_exec_object2) + sizeof(struct eb_vma); |
d710fc16 CW |
3017 | } |
3018 | ||
3019 | static bool check_buffer_count(size_t count) | |
3020 | { | |
3021 | const size_t sz = eb_element_size(); | |
3022 | ||
3023 | /* | |
3024 | * When using LUT_HANDLE, we impose a limit of INT_MAX for the lookup | |
3025 | * array size (see eb_create()). Otherwise, we can accept an array as | |
3026 | * large as can be addressed (though use large arrays at your peril)! | |
3027 | */ | |
3028 | ||
3029 | return !(count < 1 || count > INT_MAX || count > SIZE_MAX / sz - 1); | |
3030 | } | |
3031 | ||
54cf91dc | 3032 | int |
6a20fe7b VS |
3033 | i915_gem_execbuffer2_ioctl(struct drm_device *dev, void *data, |
3034 | struct drm_file *file) | |
54cf91dc | 3035 | { |
d0bf4582 | 3036 | struct drm_i915_private *i915 = to_i915(dev); |
54cf91dc | 3037 | struct drm_i915_gem_execbuffer2 *args = data; |
2889caa9 | 3038 | struct drm_i915_gem_exec_object2 *exec2_list; |
d710fc16 | 3039 | const size_t count = args->buffer_count; |
2889caa9 | 3040 | int err; |
54cf91dc | 3041 | |
d710fc16 | 3042 | if (!check_buffer_count(count)) { |
d0bf4582 | 3043 | drm_dbg(&i915->drm, "execbuf2 with %zd buffers\n", count); |
54cf91dc CW |
3044 | return -EINVAL; |
3045 | } | |
3046 | ||
00aff3f6 TU |
3047 | err = i915_gem_check_execbuffer(args); |
3048 | if (err) | |
3049 | return err; | |
2889caa9 | 3050 | |
47b08693 ML |
3051 | /* Allocate extra slots for use by the command parser */ |
3052 | exec2_list = kvmalloc_array(count + 2, eb_element_size(), | |
0ee931c4 | 3053 | __GFP_NOWARN | GFP_KERNEL); |
54cf91dc | 3054 | if (exec2_list == NULL) { |
d0bf4582 WK |
3055 | drm_dbg(&i915->drm, "Failed to allocate exec list for %zd buffers\n", |
3056 | count); | |
54cf91dc CW |
3057 | return -ENOMEM; |
3058 | } | |
2889caa9 CW |
3059 | if (copy_from_user(exec2_list, |
3060 | u64_to_user_ptr(args->buffers_ptr), | |
d710fc16 | 3061 | sizeof(*exec2_list) * count)) { |
d0bf4582 | 3062 | drm_dbg(&i915->drm, "copy %zd exec entries failed\n", count); |
2098105e | 3063 | kvfree(exec2_list); |
54cf91dc CW |
3064 | return -EFAULT; |
3065 | } | |
3066 | ||
cda9edd0 | 3067 | err = i915_gem_do_execbuffer(dev, file, args, exec2_list); |
2889caa9 CW |
3068 | |
3069 | /* | |
3070 | * Now that we have begun execution of the batchbuffer, we ignore | |
3071 | * any new error after this point. Also given that we have already | |
3072 | * updated the associated relocations, we try to write out the current | |
3073 | * object locations irrespective of any error. | |
3074 | */ | |
3075 | if (args->flags & __EXEC_HAS_RELOC) { | |
d593d992 | 3076 | struct drm_i915_gem_exec_object2 __user *user_exec_list = |
2889caa9 CW |
3077 | u64_to_user_ptr(args->buffers_ptr); |
3078 | unsigned int i; | |
9aab8bff | 3079 | |
2889caa9 | 3080 | /* Copy the new buffer offsets back to the user's exec list. */ |
594cc251 LT |
3081 | /* |
3082 | * Note: count * sizeof(*user_exec_list) does not overflow, | |
3083 | * because we checked 'count' in check_buffer_count(). | |
3084 | * | |
3085 | * And this range already got effectively checked earlier | |
3086 | * when we did the "copy_from_user()" above. | |
3087 | */ | |
b44f6873 CL |
3088 | if (!user_write_access_begin(user_exec_list, |
3089 | count * sizeof(*user_exec_list))) | |
8f4faed0 | 3090 | goto end; |
594cc251 | 3091 | |
9aab8bff | 3092 | for (i = 0; i < args->buffer_count; i++) { |
2889caa9 CW |
3093 | if (!(exec2_list[i].offset & UPDATE)) |
3094 | continue; | |
3095 | ||
934acce3 | 3096 | exec2_list[i].offset = |
2889caa9 CW |
3097 | gen8_canonical_addr(exec2_list[i].offset & PIN_OFFSET_MASK); |
3098 | unsafe_put_user(exec2_list[i].offset, | |
3099 | &user_exec_list[i].offset, | |
3100 | end_user); | |
54cf91dc | 3101 | } |
2889caa9 | 3102 | end_user: |
b44f6873 | 3103 | user_write_access_end(); |
8f4faed0 | 3104 | end:; |
54cf91dc CW |
3105 | } |
3106 | ||
2889caa9 | 3107 | args->flags &= ~__I915_EXEC_UNKNOWN_FLAGS; |
2098105e | 3108 | kvfree(exec2_list); |
2889caa9 | 3109 | return err; |
54cf91dc | 3110 | } |