| File: | dev/pci/drm/i915/gem/i915_gem_execbuffer.c |
| Warning: | line 3056, column 3 Value stored to 'syncobj' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | /* |
| 2 | * SPDX-License-Identifier: MIT |
| 3 | * |
| 4 | * Copyright © 2008,2010 Intel Corporation |
| 5 | */ |
| 6 | |
| 7 | #include <linux/dma-resv.h> |
| 8 | #include <linux/highmem.h> |
| 9 | #include <linux/sync_file.h> |
| 10 | #include <linux/uaccess.h> |
| 11 | |
| 12 | #include <drm/drm_syncobj.h> |
| 13 | |
| 14 | #include <dev/pci/pcivar.h> |
| 15 | #include <dev/pci/agpvar.h> |
| 16 | |
| 17 | #include "display/intel_frontbuffer.h" |
| 18 | |
| 19 | #include "gem/i915_gem_ioctls.h" |
| 20 | #include "gt/intel_context.h" |
| 21 | #include "gt/intel_gpu_commands.h" |
| 22 | #include "gt/intel_gt.h" |
| 23 | #include "gt/intel_gt_buffer_pool.h" |
| 24 | #include "gt/intel_gt_pm.h" |
| 25 | #include "gt/intel_ring.h" |
| 26 | |
| 27 | #include "pxp/intel_pxp.h" |
| 28 | |
| 29 | #include "i915_cmd_parser.h" |
| 30 | #include "i915_drv.h" |
| 31 | #include "i915_file_private.h" |
| 32 | #include "i915_gem_clflush.h" |
| 33 | #include "i915_gem_context.h" |
| 34 | #include "i915_gem_evict.h" |
| 35 | #include "i915_gem_ioctls.h" |
| 36 | #include "i915_trace.h" |
| 37 | #include "i915_user_extensions.h" |
| 38 | |
| 39 | struct eb_vma { |
| 40 | struct i915_vma *vma; |
| 41 | unsigned int flags; |
| 42 | |
| 43 | /** This vma's place in the execbuf reservation list */ |
| 44 | struct drm_i915_gem_exec_object2 *exec; |
| 45 | struct list_head bind_link; |
| 46 | struct list_head reloc_link; |
| 47 | |
| 48 | struct hlist_node node; |
| 49 | u32 handle; |
| 50 | }; |
| 51 | |
| 52 | enum { |
| 53 | FORCE_CPU_RELOC = 1, |
| 54 | FORCE_GTT_RELOC, |
| 55 | FORCE_GPU_RELOC, |
| 56 | #define DBG_FORCE_RELOC0 0 /* choose one of the above! */ |
| 57 | }; |
| 58 | |
| 59 | /* __EXEC_OBJECT_NO_RESERVE is BIT(31), defined in i915_vma.h */ |
| 60 | #define __EXEC_OBJECT_HAS_PIN(1UL << (30)) BIT(30)(1UL << (30)) |
| 61 | #define __EXEC_OBJECT_HAS_FENCE(1UL << (29)) BIT(29)(1UL << (29)) |
| 62 | #define __EXEC_OBJECT_USERPTR_INIT(1UL << (28)) BIT(28)(1UL << (28)) |
| 63 | #define __EXEC_OBJECT_NEEDS_MAP(1UL << (27)) BIT(27)(1UL << (27)) |
| 64 | #define __EXEC_OBJECT_NEEDS_BIAS(1UL << (26)) BIT(26)(1UL << (26)) |
| 65 | #define __EXEC_OBJECT_INTERNAL_FLAGS(~0u << 26) (~0u << 26) /* all of the above + */ |
| 66 | #define __EXEC_OBJECT_RESERVED((1UL << (30)) | (1UL << (29))) (__EXEC_OBJECT_HAS_PIN(1UL << (30)) | __EXEC_OBJECT_HAS_FENCE(1UL << (29))) |
| 67 | |
| 68 | #define __EXEC_HAS_RELOC(1UL << (31)) BIT(31)(1UL << (31)) |
| 69 | #define __EXEC_ENGINE_PINNED(1UL << (30)) BIT(30)(1UL << (30)) |
| 70 | #define __EXEC_USERPTR_USED(1UL << (29)) BIT(29)(1UL << (29)) |
| 71 | #define __EXEC_INTERNAL_FLAGS(~0u << 29) (~0u << 29) |
| 72 | #define UPDATE(1ULL << (7)) PIN_OFFSET_FIXED(1ULL << (7)) |
| 73 | |
| 74 | #define BATCH_OFFSET_BIAS(256*1024) (256*1024) |
| 75 | |
| 76 | #define __I915_EXEC_ILLEGAL_FLAGS((-((1 << 21) << 1)) | (3<<6) | (1<<15 )) \ |
| 77 | (__I915_EXEC_UNKNOWN_FLAGS(-((1 << 21) << 1)) | \ |
| 78 | I915_EXEC_CONSTANTS_MASK(3<<6) | \ |
| 79 | I915_EXEC_RESOURCE_STREAMER(1<<15)) |
| 80 | |
| 81 | /* Catch emission of unexpected errors for CI! */ |
| 82 | #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)0 |
| 83 | #undef EINVAL22 |
| 84 | #define EINVAL22 ({ \ |
| 85 | DRM_DEBUG_DRIVER("EINVAL at %s:%d\n", __func__, __LINE__)___drm_dbg(((void *)0), DRM_UT_DRIVER, "EINVAL at %s:%d\n", __func__ , 85); \ |
| 86 | 22; \ |
| 87 | }) |
| 88 | #endif |
| 89 | |
| 90 | /** |
| 91 | * DOC: User command execution |
| 92 | * |
| 93 | * Userspace submits commands to be executed on the GPU as an instruction |
| 94 | * stream within a GEM object we call a batchbuffer. This instructions may |
| 95 | * refer to other GEM objects containing auxiliary state such as kernels, |
| 96 | * samplers, render targets and even secondary batchbuffers. Userspace does |
| 97 | * not know where in the GPU memory these objects reside and so before the |
| 98 | * batchbuffer is passed to the GPU for execution, those addresses in the |
| 99 | * batchbuffer and auxiliary objects are updated. This is known as relocation, |
| 100 | * or patching. To try and avoid having to relocate each object on the next |
| 101 | * execution, userspace is told the location of those objects in this pass, |
| 102 | * but this remains just a hint as the kernel may choose a new location for |
| 103 | * any object in the future. |
| 104 | * |
| 105 | * At the level of talking to the hardware, submitting a batchbuffer for the |
| 106 | * GPU to execute is to add content to a buffer from which the HW |
| 107 | * command streamer is reading. |
| 108 | * |
| 109 | * 1. Add a command to load the HW context. For Logical Ring Contexts, i.e. |
| 110 | * Execlists, this command is not placed on the same buffer as the |
| 111 | * remaining items. |
| 112 | * |
| 113 | * 2. Add a command to invalidate caches to the buffer. |
| 114 | * |
| 115 | * 3. Add a batchbuffer start command to the buffer; the start command is |
| 116 | * essentially a token together with the GPU address of the batchbuffer |
| 117 | * to be executed. |
| 118 | * |
| 119 | * 4. Add a pipeline flush to the buffer. |
| 120 | * |
| 121 | * 5. Add a memory write command to the buffer to record when the GPU |
| 122 | * is done executing the batchbuffer. The memory write writes the |
| 123 | * global sequence number of the request, ``i915_request::global_seqno``; |
| 124 | * the i915 driver uses the current value in the register to determine |
| 125 | * if the GPU has completed the batchbuffer. |
| 126 | * |
| 127 | * 6. Add a user interrupt command to the buffer. This command instructs |
| 128 | * the GPU to issue an interrupt when the command, pipeline flush and |
| 129 | * memory write are completed. |
| 130 | * |
| 131 | * 7. Inform the hardware of the additional commands added to the buffer |
| 132 | * (by updating the tail pointer). |
| 133 | * |
| 134 | * Processing an execbuf ioctl is conceptually split up into a few phases. |
| 135 | * |
| 136 | * 1. Validation - Ensure all the pointers, handles and flags are valid. |
| 137 | * 2. Reservation - Assign GPU address space for every object |
| 138 | * 3. Relocation - Update any addresses to point to the final locations |
| 139 | * 4. Serialisation - Order the request with respect to its dependencies |
| 140 | * 5. Construction - Construct a request to execute the batchbuffer |
| 141 | * 6. Submission (at some point in the future execution) |
| 142 | * |
| 143 | * Reserving resources for the execbuf is the most complicated phase. We |
| 144 | * neither want to have to migrate the object in the address space, nor do |
| 145 | * we want to have to update any relocations pointing to this object. Ideally, |
| 146 | * we want to leave the object where it is and for all the existing relocations |
| 147 | * to match. If the object is given a new address, or if userspace thinks the |
| 148 | * object is elsewhere, we have to parse all the relocation entries and update |
| 149 | * the addresses. Userspace can set the I915_EXEC_NORELOC flag to hint that |
| 150 | * all the target addresses in all of its objects match the value in the |
| 151 | * relocation entries and that they all match the presumed offsets given by the |
| 152 | * list of execbuffer objects. Using this knowledge, we know that if we haven't |
| 153 | * moved any buffers, all the relocation entries are valid and we can skip |
| 154 | * the update. (If userspace is wrong, the likely outcome is an impromptu GPU |
| 155 | * hang.) The requirement for using I915_EXEC_NO_RELOC are: |
| 156 | * |
| 157 | * The addresses written in the objects must match the corresponding |
| 158 | * reloc.presumed_offset which in turn must match the corresponding |
| 159 | * execobject.offset. |
| 160 | * |
| 161 | * Any render targets written to in the batch must be flagged with |
| 162 | * EXEC_OBJECT_WRITE. |
| 163 | * |
| 164 | * To avoid stalling, execobject.offset should match the current |
| 165 | * address of that object within the active context. |
| 166 | * |
| 167 | * The reservation is done is multiple phases. First we try and keep any |
| 168 | * object already bound in its current location - so as long as meets the |
| 169 | * constraints imposed by the new execbuffer. Any object left unbound after the |
| 170 | * first pass is then fitted into any available idle space. If an object does |
| 171 | * not fit, all objects are removed from the reservation and the process rerun |
| 172 | * after sorting the objects into a priority order (more difficult to fit |
| 173 | * objects are tried first). Failing that, the entire VM is cleared and we try |
| 174 | * to fit the execbuf once last time before concluding that it simply will not |
| 175 | * fit. |
| 176 | * |
| 177 | * A small complication to all of this is that we allow userspace not only to |
| 178 | * specify an alignment and a size for the object in the address space, but |
| 179 | * we also allow userspace to specify the exact offset. This objects are |
| 180 | * simpler to place (the location is known a priori) all we have to do is make |
| 181 | * sure the space is available. |
| 182 | * |
| 183 | * Once all the objects are in place, patching up the buried pointers to point |
| 184 | * to the final locations is a fairly simple job of walking over the relocation |
| 185 | * entry arrays, looking up the right address and rewriting the value into |
| 186 | * the object. Simple! ... The relocation entries are stored in user memory |
| 187 | * and so to access them we have to copy them into a local buffer. That copy |
| 188 | * has to avoid taking any pagefaults as they may lead back to a GEM object |
| 189 | * requiring the struct_mutex (i.e. recursive deadlock). So once again we split |
| 190 | * the relocation into multiple passes. First we try to do everything within an |
| 191 | * atomic context (avoid the pagefaults) which requires that we never wait. If |
| 192 | * we detect that we may wait, or if we need to fault, then we have to fallback |
| 193 | * to a slower path. The slowpath has to drop the mutex. (Can you hear alarm |
| 194 | * bells yet?) Dropping the mutex means that we lose all the state we have |
| 195 | * built up so far for the execbuf and we must reset any global data. However, |
| 196 | * we do leave the objects pinned in their final locations - which is a |
| 197 | * potential issue for concurrent execbufs. Once we have left the mutex, we can |
| 198 | * allocate and copy all the relocation entries into a large array at our |
| 199 | * leisure, reacquire the mutex, reclaim all the objects and other state and |
| 200 | * then proceed to update any incorrect addresses with the objects. |
| 201 | * |
| 202 | * As we process the relocation entries, we maintain a record of whether the |
| 203 | * object is being written to. Using NORELOC, we expect userspace to provide |
| 204 | * this information instead. We also check whether we can skip the relocation |
| 205 | * by comparing the expected value inside the relocation entry with the target's |
| 206 | * final address. If they differ, we have to map the current object and rewrite |
| 207 | * the 4 or 8 byte pointer within. |
| 208 | * |
| 209 | * Serialising an execbuf is quite simple according to the rules of the GEM |
| 210 | * ABI. Execution within each context is ordered by the order of submission. |
| 211 | * Writes to any GEM object are in order of submission and are exclusive. Reads |
| 212 | * from a GEM object are unordered with respect to other reads, but ordered by |
| 213 | * writes. A write submitted after a read cannot occur before the read, and |
| 214 | * similarly any read submitted after a write cannot occur before the write. |
| 215 | * Writes are ordered between engines such that only one write occurs at any |
| 216 | * time (completing any reads beforehand) - using semaphores where available |
| 217 | * and CPU serialisation otherwise. Other GEM access obey the same rules, any |
| 218 | * write (either via mmaps using set-domain, or via pwrite) must flush all GPU |
| 219 | * reads before starting, and any read (either using set-domain or pread) must |
| 220 | * flush all GPU writes before starting. (Note we only employ a barrier before, |
| 221 | * we currently rely on userspace not concurrently starting a new execution |
| 222 | * whilst reading or writing to an object. This may be an advantage or not |
| 223 | * depending on how much you trust userspace not to shoot themselves in the |
| 224 | * foot.) Serialisation may just result in the request being inserted into |
| 225 | * a DAG awaiting its turn, but most simple is to wait on the CPU until |
| 226 | * all dependencies are resolved. |
| 227 | * |
| 228 | * After all of that, is just a matter of closing the request and handing it to |
| 229 | * the hardware (well, leaving it in a queue to be executed). However, we also |
| 230 | * offer the ability for batchbuffers to be run with elevated privileges so |
| 231 | * that they access otherwise hidden registers. (Used to adjust L3 cache etc.) |
| 232 | * Before any batch is given extra privileges we first must check that it |
| 233 | * contains no nefarious instructions, we check that each instruction is from |
| 234 | * our whitelist and all registers are also from an allowed list. We first |
| 235 | * copy the user's batchbuffer to a shadow (so that the user doesn't have |
| 236 | * access to it, either by the CPU or GPU as we scan it) and then parse each |
| 237 | * instruction. If everything is ok, we set a flag telling the hardware to run |
| 238 | * the batchbuffer in trusted mode, otherwise the ioctl is rejected. |
| 239 | */ |
| 240 | |
| 241 | struct eb_fence { |
| 242 | struct drm_syncobj *syncobj; /* Use with ptr_mask_bits() */ |
| 243 | struct dma_fence *dma_fence; |
| 244 | u64 value; |
| 245 | struct dma_fence_chain *chain_fence; |
| 246 | }; |
| 247 | |
| 248 | struct i915_execbuffer { |
| 249 | struct drm_i915_privateinteldrm_softc *i915; /** i915 backpointer */ |
| 250 | struct drm_file *file; /** per-file lookup tables and limits */ |
| 251 | struct drm_i915_gem_execbuffer2 *args; /** ioctl parameters */ |
| 252 | struct drm_i915_gem_exec_object2 *exec; /** ioctl execobj[] */ |
| 253 | struct eb_vma *vma; |
| 254 | |
| 255 | struct intel_gt *gt; /* gt for the execbuf */ |
| 256 | struct intel_context *context; /* logical state for the request */ |
| 257 | struct i915_gem_context *gem_context; /** caller's context */ |
| 258 | |
| 259 | /** our requests to build */ |
| 260 | struct i915_request *requests[MAX_ENGINE_INSTANCE8 + 1]; |
| 261 | /** identity of the batch obj/vma */ |
| 262 | struct eb_vma *batches[MAX_ENGINE_INSTANCE8 + 1]; |
| 263 | struct i915_vma *trampoline; /** trampoline used for chaining */ |
| 264 | |
| 265 | /** used for excl fence in dma_resv objects when > 1 BB submitted */ |
| 266 | struct dma_fence *composite_fence; |
| 267 | |
| 268 | /** actual size of execobj[] as we may extend it for the cmdparser */ |
| 269 | unsigned int buffer_count; |
| 270 | |
| 271 | /* number of batches in execbuf IOCTL */ |
| 272 | unsigned int num_batches; |
| 273 | |
| 274 | /** list of vma not yet bound during reservation phase */ |
| 275 | struct list_head unbound; |
| 276 | |
| 277 | /** list of vma that have execobj.relocation_count */ |
| 278 | struct list_head relocs; |
| 279 | |
| 280 | struct i915_gem_ww_ctx ww; |
| 281 | |
| 282 | /** |
| 283 | * Track the most recently used object for relocations, as we |
| 284 | * frequently have to perform multiple relocations within the same |
| 285 | * obj/page |
| 286 | */ |
| 287 | struct reloc_cache { |
| 288 | struct drm_mm_node node; /** temporary GTT binding */ |
| 289 | unsigned long vaddr; /** Current kmap address */ |
| 290 | unsigned long page; /** Currently mapped page index */ |
| 291 | unsigned int graphics_ver; /** Cached value of GRAPHICS_VER */ |
| 292 | bool_Bool use_64bit_reloc : 1; |
| 293 | bool_Bool has_llc : 1; |
| 294 | bool_Bool has_fence : 1; |
| 295 | bool_Bool needs_unfenced : 1; |
| 296 | |
| 297 | struct agp_map *map; |
| 298 | bus_space_tag_t iot; |
| 299 | bus_space_handle_t ioh; |
| 300 | } reloc_cache; |
| 301 | |
| 302 | u64 invalid_flags; /** Set of execobj.flags that are invalid */ |
| 303 | |
| 304 | /** Length of batch within object */ |
| 305 | u64 batch_len[MAX_ENGINE_INSTANCE8 + 1]; |
| 306 | u32 batch_start_offset; /** Location within object of batch */ |
| 307 | u32 batch_flags; /** Flags composed for emit_bb_start() */ |
| 308 | struct intel_gt_buffer_pool_node *batch_pool; /** pool node for batch buffer */ |
| 309 | |
| 310 | /** |
| 311 | * Indicate either the size of the hastable used to resolve |
| 312 | * relocation handles, or if negative that we are using a direct |
| 313 | * index into the execobj[]. |
| 314 | */ |
| 315 | int lut_size; |
| 316 | struct hlist_head *buckets; /** ht for relocation handles */ |
| 317 | |
| 318 | struct eb_fence *fences; |
| 319 | unsigned long num_fences; |
| 320 | #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR)1 |
| 321 | struct i915_capture_list *capture_lists[MAX_ENGINE_INSTANCE8 + 1]; |
| 322 | #endif |
| 323 | }; |
| 324 | |
| 325 | static int eb_parse(struct i915_execbuffer *eb); |
| 326 | static int eb_pin_engine(struct i915_execbuffer *eb, bool_Bool throttle); |
| 327 | static void eb_unpin_engine(struct i915_execbuffer *eb); |
| 328 | static void eb_capture_release(struct i915_execbuffer *eb); |
| 329 | |
| 330 | static inline bool_Bool eb_use_cmdparser(const struct i915_execbuffer *eb) |
| 331 | { |
| 332 | return intel_engine_requires_cmd_parser(eb->context->engine) || |
| 333 | (intel_engine_using_cmd_parser(eb->context->engine) && |
| 334 | eb->args->batch_len); |
| 335 | } |
| 336 | |
| 337 | static int eb_create(struct i915_execbuffer *eb) |
| 338 | { |
| 339 | if (!(eb->args->flags & I915_EXEC_HANDLE_LUT(1<<12))) { |
| 340 | unsigned int size = 1 + ilog2(eb->buffer_count)((sizeof(eb->buffer_count) <= 4) ? (fls(eb->buffer_count ) - 1) : (flsl(eb->buffer_count) - 1)); |
| 341 | |
| 342 | /* |
| 343 | * Without a 1:1 association between relocation handles and |
| 344 | * the execobject[] index, we instead create a hashtable. |
| 345 | * We size it dynamically based on available memory, starting |
| 346 | * first with 1:1 assocative hash and scaling back until |
| 347 | * the allocation succeeds. |
| 348 | * |
| 349 | * Later on we use a positive lut_size to indicate we are |
| 350 | * using this hashtable, and a negative value to indicate a |
| 351 | * direct lookup. |
| 352 | */ |
| 353 | do { |
| 354 | gfp_t flags; |
| 355 | |
| 356 | /* While we can still reduce the allocation size, don't |
| 357 | * raise a warning and allow the allocation to fail. |
| 358 | * On the last pass though, we want to try as hard |
| 359 | * as possible to perform the allocation and warn |
| 360 | * if it fails. |
| 361 | */ |
| 362 | flags = GFP_KERNEL(0x0001 | 0x0004); |
| 363 | if (size > 1) |
| 364 | flags |= __GFP_NORETRY0 | __GFP_NOWARN0; |
| 365 | |
| 366 | eb->buckets = kzalloc(sizeof(struct hlist_head) << size, |
| 367 | flags); |
| 368 | if (eb->buckets) |
| 369 | break; |
| 370 | } while (--size); |
| 371 | |
| 372 | if (unlikely(!size)__builtin_expect(!!(!size), 0)) |
| 373 | return -ENOMEM12; |
| 374 | |
| 375 | eb->lut_size = size; |
| 376 | } else { |
| 377 | eb->lut_size = -eb->buffer_count; |
| 378 | } |
| 379 | |
| 380 | return 0; |
| 381 | } |
| 382 | |
| 383 | static bool_Bool |
| 384 | eb_vma_misplaced(const struct drm_i915_gem_exec_object2 *entry, |
| 385 | const struct i915_vma *vma, |
| 386 | unsigned int flags) |
| 387 | { |
| 388 | if (vma->node.size < entry->pad_to_size) |
| 389 | return true1; |
| 390 | |
| 391 | if (entry->alignment && !IS_ALIGNED(vma->node.start, entry->alignment)(((vma->node.start) & ((entry->alignment) - 1)) == 0 )) |
| 392 | return true1; |
| 393 | |
| 394 | if (flags & EXEC_OBJECT_PINNED(1<<4) && |
| 395 | vma->node.start != entry->offset) |
| 396 | return true1; |
| 397 | |
| 398 | if (flags & __EXEC_OBJECT_NEEDS_BIAS(1UL << (26)) && |
| 399 | vma->node.start < BATCH_OFFSET_BIAS(256*1024)) |
| 400 | return true1; |
| 401 | |
| 402 | if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS(1<<3)) && |
| 403 | (vma->node.start + vma->node.size + 4095) >> 32) |
| 404 | return true1; |
| 405 | |
| 406 | if (flags & __EXEC_OBJECT_NEEDS_MAP(1UL << (27)) && |
| 407 | !i915_vma_is_map_and_fenceable(vma)) |
| 408 | return true1; |
| 409 | |
| 410 | return false0; |
| 411 | } |
| 412 | |
| 413 | static u64 eb_pin_flags(const struct drm_i915_gem_exec_object2 *entry, |
| 414 | unsigned int exec_flags) |
| 415 | { |
| 416 | u64 pin_flags = 0; |
| 417 | |
| 418 | if (exec_flags & EXEC_OBJECT_NEEDS_GTT(1<<1)) |
| 419 | pin_flags |= PIN_GLOBAL(1ULL << (10)); |
| 420 | |
| 421 | /* |
| 422 | * Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset, |
| 423 | * limit address to the first 4GBs for unflagged objects. |
| 424 | */ |
| 425 | if (!(exec_flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS(1<<3))) |
| 426 | pin_flags |= PIN_ZONE_4G(1ULL << (4)); |
| 427 | |
| 428 | if (exec_flags & __EXEC_OBJECT_NEEDS_MAP(1UL << (27))) |
| 429 | pin_flags |= PIN_MAPPABLE(1ULL << (3)); |
| 430 | |
| 431 | if (exec_flags & EXEC_OBJECT_PINNED(1<<4)) |
| 432 | pin_flags |= entry->offset | PIN_OFFSET_FIXED(1ULL << (7)); |
| 433 | else if (exec_flags & __EXEC_OBJECT_NEEDS_BIAS(1UL << (26))) |
| 434 | pin_flags |= BATCH_OFFSET_BIAS(256*1024) | PIN_OFFSET_BIAS(1ULL << (6)); |
| 435 | |
| 436 | return pin_flags; |
| 437 | } |
| 438 | |
| 439 | static inline int |
| 440 | eb_pin_vma(struct i915_execbuffer *eb, |
| 441 | const struct drm_i915_gem_exec_object2 *entry, |
| 442 | struct eb_vma *ev) |
| 443 | { |
| 444 | struct i915_vma *vma = ev->vma; |
| 445 | u64 pin_flags; |
| 446 | int err; |
| 447 | |
| 448 | if (vma->node.size) |
| 449 | pin_flags = vma->node.start; |
| 450 | else |
| 451 | pin_flags = entry->offset & PIN_OFFSET_MASK-(1ULL << (12)); |
| 452 | |
| 453 | pin_flags |= PIN_USER(1ULL << (11)) | PIN_NOEVICT(1ULL << (0)) | PIN_OFFSET_FIXED(1ULL << (7)) | PIN_VALIDATE(1ULL << (8)); |
| 454 | if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_GTT)__builtin_expect(!!(ev->flags & (1<<1)), 0)) |
| 455 | pin_flags |= PIN_GLOBAL(1ULL << (10)); |
| 456 | |
| 457 | /* Attempt to reuse the current location if available */ |
| 458 | err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, pin_flags); |
| 459 | if (err == -EDEADLK11) |
| 460 | return err; |
| 461 | |
| 462 | if (unlikely(err)__builtin_expect(!!(err), 0)) { |
| 463 | if (entry->flags & EXEC_OBJECT_PINNED(1<<4)) |
| 464 | return err; |
| 465 | |
| 466 | /* Failing that pick any _free_ space if suitable */ |
| 467 | err = i915_vma_pin_ww(vma, &eb->ww, |
| 468 | entry->pad_to_size, |
| 469 | entry->alignment, |
| 470 | eb_pin_flags(entry, ev->flags) | |
| 471 | PIN_USER(1ULL << (11)) | PIN_NOEVICT(1ULL << (0)) | PIN_VALIDATE(1ULL << (8))); |
| 472 | if (unlikely(err)__builtin_expect(!!(err), 0)) |
| 473 | return err; |
| 474 | } |
| 475 | |
| 476 | if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)__builtin_expect(!!(ev->flags & (1<<0)), 0)) { |
| 477 | err = i915_vma_pin_fence(vma); |
| 478 | if (unlikely(err)__builtin_expect(!!(err), 0)) |
| 479 | return err; |
| 480 | |
| 481 | if (vma->fence) |
| 482 | ev->flags |= __EXEC_OBJECT_HAS_FENCE(1UL << (29)); |
| 483 | } |
| 484 | |
| 485 | ev->flags |= __EXEC_OBJECT_HAS_PIN(1UL << (30)); |
| 486 | if (eb_vma_misplaced(entry, vma, ev->flags)) |
| 487 | return -EBADSLT22; |
| 488 | |
| 489 | return 0; |
| 490 | } |
| 491 | |
| 492 | static inline void |
| 493 | eb_unreserve_vma(struct eb_vma *ev) |
| 494 | { |
| 495 | if (unlikely(ev->flags & __EXEC_OBJECT_HAS_FENCE)__builtin_expect(!!(ev->flags & (1UL << (29))), 0 )) |
| 496 | __i915_vma_unpin_fence(ev->vma); |
| 497 | |
| 498 | ev->flags &= ~__EXEC_OBJECT_RESERVED((1UL << (30)) | (1UL << (29))); |
| 499 | } |
| 500 | |
| 501 | static int |
| 502 | eb_validate_vma(struct i915_execbuffer *eb, |
| 503 | struct drm_i915_gem_exec_object2 *entry, |
| 504 | struct i915_vma *vma) |
| 505 | { |
| 506 | /* Relocations are disallowed for all platforms after TGL-LP. This |
| 507 | * also covers all platforms with local memory. |
| 508 | */ |
| 509 | if (entry->relocation_count && |
| 510 | GRAPHICS_VER(eb->i915)((&(eb->i915)->__runtime)->graphics.ip.ver) >= 12 && !IS_TIGERLAKE(eb->i915)IS_PLATFORM(eb->i915, INTEL_TIGERLAKE)) |
| 511 | return -EINVAL22; |
| 512 | |
| 513 | if (unlikely(entry->flags & eb->invalid_flags)__builtin_expect(!!(entry->flags & eb->invalid_flags ), 0)) |
| 514 | return -EINVAL22; |
| 515 | |
| 516 | if (unlikely(entry->alignment &&__builtin_expect(!!(entry->alignment && !is_power_of_2_u64 (entry->alignment)), 0) |
| 517 | !is_power_of_2_u64(entry->alignment))__builtin_expect(!!(entry->alignment && !is_power_of_2_u64 (entry->alignment)), 0)) |
| 518 | return -EINVAL22; |
| 519 | |
| 520 | /* |
| 521 | * Offset can be used as input (EXEC_OBJECT_PINNED), reject |
| 522 | * any non-page-aligned or non-canonical addresses. |
| 523 | */ |
| 524 | if (unlikely(entry->flags & EXEC_OBJECT_PINNED &&__builtin_expect(!!(entry->flags & (1<<4) && entry->offset != gen8_canonical_addr(entry->offset & -(1ULL << (12)))), 0) |
| 525 | entry->offset != gen8_canonical_addr(entry->offset & I915_GTT_PAGE_MASK))__builtin_expect(!!(entry->flags & (1<<4) && entry->offset != gen8_canonical_addr(entry->offset & -(1ULL << (12)))), 0)) |
| 526 | return -EINVAL22; |
| 527 | |
| 528 | /* pad_to_size was once a reserved field, so sanitize it */ |
| 529 | if (entry->flags & EXEC_OBJECT_PAD_TO_SIZE(1<<5)) { |
| 530 | if (unlikely(offset_in_page(entry->pad_to_size))__builtin_expect(!!(((vaddr_t)(entry->pad_to_size) & ( (1 << 12) - 1))), 0)) |
| 531 | return -EINVAL22; |
| 532 | } else { |
| 533 | entry->pad_to_size = 0; |
| 534 | } |
| 535 | /* |
| 536 | * From drm_mm perspective address space is continuous, |
| 537 | * so from this point we're always using non-canonical |
| 538 | * form internally. |
| 539 | */ |
| 540 | entry->offset = gen8_noncanonical_addr(entry->offset); |
| 541 | |
| 542 | if (!eb->reloc_cache.has_fence) { |
| 543 | entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE(1<<0); |
| 544 | } else { |
| 545 | if ((entry->flags & EXEC_OBJECT_NEEDS_FENCE(1<<0) || |
| 546 | eb->reloc_cache.needs_unfenced) && |
| 547 | i915_gem_object_is_tiled(vma->obj)) |
| 548 | entry->flags |= EXEC_OBJECT_NEEDS_GTT(1<<1) | __EXEC_OBJECT_NEEDS_MAP(1UL << (27)); |
| 549 | } |
| 550 | |
| 551 | return 0; |
| 552 | } |
| 553 | |
| 554 | static inline bool_Bool |
| 555 | is_batch_buffer(struct i915_execbuffer *eb, unsigned int buffer_idx) |
| 556 | { |
| 557 | return eb->args->flags & I915_EXEC_BATCH_FIRST(1<<18) ? |
| 558 | buffer_idx < eb->num_batches : |
| 559 | buffer_idx >= eb->args->buffer_count - eb->num_batches; |
| 560 | } |
| 561 | |
| 562 | static int |
| 563 | eb_add_vma(struct i915_execbuffer *eb, |
| 564 | unsigned int *current_batch, |
| 565 | unsigned int i, |
| 566 | struct i915_vma *vma) |
| 567 | { |
| 568 | struct drm_i915_privateinteldrm_softc *i915 = eb->i915; |
| 569 | struct drm_i915_gem_exec_object2 *entry = &eb->exec[i]; |
| 570 | struct eb_vma *ev = &eb->vma[i]; |
| 571 | |
| 572 | ev->vma = vma; |
| 573 | ev->exec = entry; |
| 574 | ev->flags = entry->flags; |
| 575 | |
| 576 | if (eb->lut_size > 0) { |
| 577 | ev->handle = entry->handle; |
| 578 | hlist_add_head(&ev->node, |
| 579 | &eb->buckets[hash_32(entry->handle, |
| 580 | eb->lut_size)]); |
| 581 | } |
| 582 | |
| 583 | if (entry->relocation_count) |
| 584 | list_add_tail(&ev->reloc_link, &eb->relocs); |
| 585 | |
| 586 | /* |
| 587 | * SNA is doing fancy tricks with compressing batch buffers, which leads |
| 588 | * to negative relocation deltas. Usually that works out ok since the |
| 589 | * relocate address is still positive, except when the batch is placed |
| 590 | * very low in the GTT. Ensure this doesn't happen. |
| 591 | * |
| 592 | * Note that actual hangs have only been observed on gen7, but for |
| 593 | * paranoia do it everywhere. |
| 594 | */ |
| 595 | if (is_batch_buffer(eb, i)) { |
| 596 | if (entry->relocation_count && |
| 597 | !(ev->flags & EXEC_OBJECT_PINNED(1<<4))) |
| 598 | ev->flags |= __EXEC_OBJECT_NEEDS_BIAS(1UL << (26)); |
| 599 | if (eb->reloc_cache.has_fence) |
| 600 | ev->flags |= EXEC_OBJECT_NEEDS_FENCE(1<<0); |
| 601 | |
| 602 | eb->batches[*current_batch] = ev; |
| 603 | |
| 604 | if (unlikely(ev->flags & EXEC_OBJECT_WRITE)__builtin_expect(!!(ev->flags & (1<<2)), 0)) { |
| 605 | drm_dbg(&i915->drm,__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Attempting to use self-modifying batch buffer\n" ) |
| 606 | "Attempting to use self-modifying batch buffer\n")__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Attempting to use self-modifying batch buffer\n" ); |
| 607 | return -EINVAL22; |
| 608 | } |
| 609 | |
| 610 | if (range_overflows_t(u64,({ typeof((u64)(eb->batch_start_offset)) start__ = ((u64)( eb->batch_start_offset)); typeof((u64)(eb->args->batch_len )) size__ = ((u64)(eb->args->batch_len)); typeof((u64)( ev->vma->size)) max__ = ((u64)(ev->vma->size)); ( void)(&start__ == &size__); (void)(&start__ == & max__); start__ >= max__ || size__ > max__ - start__; } ) |
| 611 | eb->batch_start_offset,({ typeof((u64)(eb->batch_start_offset)) start__ = ((u64)( eb->batch_start_offset)); typeof((u64)(eb->args->batch_len )) size__ = ((u64)(eb->args->batch_len)); typeof((u64)( ev->vma->size)) max__ = ((u64)(ev->vma->size)); ( void)(&start__ == &size__); (void)(&start__ == & max__); start__ >= max__ || size__ > max__ - start__; } ) |
| 612 | eb->args->batch_len,({ typeof((u64)(eb->batch_start_offset)) start__ = ((u64)( eb->batch_start_offset)); typeof((u64)(eb->args->batch_len )) size__ = ((u64)(eb->args->batch_len)); typeof((u64)( ev->vma->size)) max__ = ((u64)(ev->vma->size)); ( void)(&start__ == &size__); (void)(&start__ == & max__); start__ >= max__ || size__ > max__ - start__; } ) |
| 613 | ev->vma->size)({ typeof((u64)(eb->batch_start_offset)) start__ = ((u64)( eb->batch_start_offset)); typeof((u64)(eb->args->batch_len )) size__ = ((u64)(eb->args->batch_len)); typeof((u64)( ev->vma->size)) max__ = ((u64)(ev->vma->size)); ( void)(&start__ == &size__); (void)(&start__ == & max__); start__ >= max__ || size__ > max__ - start__; } )) { |
| 614 | drm_dbg(&i915->drm, "Attempting to use out-of-bounds batch\n")__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Attempting to use out-of-bounds batch\n" ); |
| 615 | return -EINVAL22; |
| 616 | } |
| 617 | |
| 618 | if (eb->args->batch_len == 0) |
| 619 | eb->batch_len[*current_batch] = ev->vma->size - |
| 620 | eb->batch_start_offset; |
| 621 | else |
| 622 | eb->batch_len[*current_batch] = eb->args->batch_len; |
| 623 | if (unlikely(eb->batch_len[*current_batch] == 0)__builtin_expect(!!(eb->batch_len[*current_batch] == 0), 0 )) { /* impossible! */ |
| 624 | drm_dbg(&i915->drm, "Invalid batch length\n")__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Invalid batch length\n" ); |
| 625 | return -EINVAL22; |
| 626 | } |
| 627 | |
| 628 | ++*current_batch; |
| 629 | } |
| 630 | |
| 631 | return 0; |
| 632 | } |
| 633 | |
| 634 | static inline int use_cpu_reloc(const struct reloc_cache *cache, |
| 635 | const struct drm_i915_gem_object *obj) |
| 636 | { |
| 637 | if (!i915_gem_object_has_struct_page(obj)) |
| 638 | return false0; |
| 639 | |
| 640 | if (DBG_FORCE_RELOC0 == FORCE_CPU_RELOC) |
| 641 | return true1; |
| 642 | |
| 643 | if (DBG_FORCE_RELOC0 == FORCE_GTT_RELOC) |
| 644 | return false0; |
| 645 | |
| 646 | return (cache->has_llc || |
| 647 | obj->cache_dirty || |
| 648 | obj->cache_level != I915_CACHE_NONE); |
| 649 | } |
| 650 | |
| 651 | static int eb_reserve_vma(struct i915_execbuffer *eb, |
| 652 | struct eb_vma *ev, |
| 653 | u64 pin_flags) |
| 654 | { |
| 655 | struct drm_i915_gem_exec_object2 *entry = ev->exec; |
| 656 | struct i915_vma *vma = ev->vma; |
| 657 | int err; |
| 658 | |
| 659 | if (drm_mm_node_allocated(&vma->node) && |
| 660 | eb_vma_misplaced(entry, vma, ev->flags)) { |
| 661 | err = i915_vma_unbind(vma); |
| 662 | if (err) |
| 663 | return err; |
| 664 | } |
| 665 | |
| 666 | err = i915_vma_pin_ww(vma, &eb->ww, |
| 667 | entry->pad_to_size, entry->alignment, |
| 668 | eb_pin_flags(entry, ev->flags) | pin_flags); |
| 669 | if (err) |
| 670 | return err; |
| 671 | |
| 672 | if (entry->offset != vma->node.start) { |
| 673 | entry->offset = vma->node.start | UPDATE(1ULL << (7)); |
| 674 | eb->args->flags |= __EXEC_HAS_RELOC(1UL << (31)); |
| 675 | } |
| 676 | |
| 677 | if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)__builtin_expect(!!(ev->flags & (1<<0)), 0)) { |
| 678 | err = i915_vma_pin_fence(vma); |
| 679 | if (unlikely(err)__builtin_expect(!!(err), 0)) |
| 680 | return err; |
| 681 | |
| 682 | if (vma->fence) |
| 683 | ev->flags |= __EXEC_OBJECT_HAS_FENCE(1UL << (29)); |
| 684 | } |
| 685 | |
| 686 | ev->flags |= __EXEC_OBJECT_HAS_PIN(1UL << (30)); |
| 687 | GEM_BUG_ON(eb_vma_misplaced(entry, vma, ev->flags))((void)0); |
| 688 | |
| 689 | return 0; |
| 690 | } |
| 691 | |
| 692 | static bool_Bool eb_unbind(struct i915_execbuffer *eb, bool_Bool force) |
| 693 | { |
| 694 | const unsigned int count = eb->buffer_count; |
| 695 | unsigned int i; |
| 696 | struct list_head last; |
| 697 | bool_Bool unpinned = false0; |
| 698 | |
| 699 | /* Resort *all* the objects into priority order */ |
| 700 | INIT_LIST_HEAD(&eb->unbound); |
| 701 | INIT_LIST_HEAD(&last); |
| 702 | |
| 703 | for (i = 0; i < count; i++) { |
| 704 | struct eb_vma *ev = &eb->vma[i]; |
| 705 | unsigned int flags = ev->flags; |
| 706 | |
| 707 | if (!force && flags & EXEC_OBJECT_PINNED(1<<4) && |
| 708 | flags & __EXEC_OBJECT_HAS_PIN(1UL << (30))) |
| 709 | continue; |
| 710 | |
| 711 | unpinned = true1; |
| 712 | eb_unreserve_vma(ev); |
| 713 | |
| 714 | if (flags & EXEC_OBJECT_PINNED(1<<4)) |
| 715 | /* Pinned must have their slot */ |
| 716 | list_add(&ev->bind_link, &eb->unbound); |
| 717 | else if (flags & __EXEC_OBJECT_NEEDS_MAP(1UL << (27))) |
| 718 | /* Map require the lowest 256MiB (aperture) */ |
| 719 | list_add_tail(&ev->bind_link, &eb->unbound); |
| 720 | else if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS(1<<3))) |
| 721 | /* Prioritise 4GiB region for restricted bo */ |
| 722 | list_add(&ev->bind_link, &last); |
| 723 | else |
| 724 | list_add_tail(&ev->bind_link, &last); |
| 725 | } |
| 726 | |
| 727 | list_splice_tail(&last, &eb->unbound); |
| 728 | return unpinned; |
| 729 | } |
| 730 | |
| 731 | static int eb_reserve(struct i915_execbuffer *eb) |
| 732 | { |
| 733 | struct eb_vma *ev; |
| 734 | unsigned int pass; |
| 735 | int err = 0; |
| 736 | bool_Bool unpinned; |
| 737 | |
| 738 | /* |
| 739 | * We have one more buffers that we couldn't bind, which could be due to |
| 740 | * various reasons. To resolve this we have 4 passes, with every next |
| 741 | * level turning the screws tighter: |
| 742 | * |
| 743 | * 0. Unbind all objects that do not match the GTT constraints for the |
| 744 | * execbuffer (fenceable, mappable, alignment etc). Bind all new |
| 745 | * objects. This avoids unnecessary unbinding of later objects in order |
| 746 | * to make room for the earlier objects *unless* we need to defragment. |
| 747 | * |
| 748 | * 1. Reorder the buffers, where objects with the most restrictive |
| 749 | * placement requirements go first (ignoring fixed location buffers for |
| 750 | * now). For example, objects needing the mappable aperture (the first |
| 751 | * 256M of GTT), should go first vs objects that can be placed just |
| 752 | * about anywhere. Repeat the previous pass. |
| 753 | * |
| 754 | * 2. Consider buffers that are pinned at a fixed location. Also try to |
| 755 | * evict the entire VM this time, leaving only objects that we were |
| 756 | * unable to lock. Try again to bind the buffers. (still using the new |
| 757 | * buffer order). |
| 758 | * |
| 759 | * 3. We likely have object lock contention for one or more stubborn |
| 760 | * objects in the VM, for which we need to evict to make forward |
| 761 | * progress (perhaps we are fighting the shrinker?). When evicting the |
| 762 | * VM this time around, anything that we can't lock we now track using |
| 763 | * the busy_bo, using the full lock (after dropping the vm->mutex to |
| 764 | * prevent deadlocks), instead of trylock. We then continue to evict the |
| 765 | * VM, this time with the stubborn object locked, which we can now |
| 766 | * hopefully unbind (if still bound in the VM). Repeat until the VM is |
| 767 | * evicted. Finally we should be able bind everything. |
| 768 | */ |
| 769 | for (pass = 0; pass <= 3; pass++) { |
| 770 | int pin_flags = PIN_USER(1ULL << (11)) | PIN_VALIDATE(1ULL << (8)); |
| 771 | |
| 772 | if (pass == 0) |
| 773 | pin_flags |= PIN_NONBLOCK(1ULL << (2)); |
| 774 | |
| 775 | if (pass >= 1) |
| 776 | unpinned = eb_unbind(eb, pass >= 2); |
| 777 | |
| 778 | if (pass == 2) { |
| 779 | err = mutex_lock_interruptible(&eb->context->vm->mutex); |
| 780 | if (!err) { |
| 781 | err = i915_gem_evict_vm(eb->context->vm, &eb->ww, NULL((void *)0)); |
| 782 | mutex_unlock(&eb->context->vm->mutex)rw_exit_write(&eb->context->vm->mutex); |
| 783 | } |
| 784 | if (err) |
| 785 | return err; |
| 786 | } |
| 787 | |
| 788 | if (pass == 3) { |
| 789 | retry: |
| 790 | err = mutex_lock_interruptible(&eb->context->vm->mutex); |
| 791 | if (!err) { |
| 792 | struct drm_i915_gem_object *busy_bo = NULL((void *)0); |
| 793 | |
| 794 | err = i915_gem_evict_vm(eb->context->vm, &eb->ww, &busy_bo); |
| 795 | mutex_unlock(&eb->context->vm->mutex)rw_exit_write(&eb->context->vm->mutex); |
| 796 | if (err && busy_bo) { |
| 797 | err = i915_gem_object_lock(busy_bo, &eb->ww); |
| 798 | i915_gem_object_put(busy_bo); |
| 799 | if (!err) |
| 800 | goto retry; |
| 801 | } |
| 802 | } |
| 803 | if (err) |
| 804 | return err; |
| 805 | } |
| 806 | |
| 807 | list_for_each_entry(ev, &eb->unbound, bind_link)for (ev = ({ const __typeof( ((__typeof(*ev) *)0)->bind_link ) *__mptr = ((&eb->unbound)->next); (__typeof(*ev) *)( (char *)__mptr - __builtin_offsetof(__typeof(*ev), bind_link ) );}); &ev->bind_link != (&eb->unbound); ev = ( { const __typeof( ((__typeof(*ev) *)0)->bind_link ) *__mptr = (ev->bind_link.next); (__typeof(*ev) *)( (char *)__mptr - __builtin_offsetof(__typeof(*ev), bind_link) );})) { |
| 808 | err = eb_reserve_vma(eb, ev, pin_flags); |
| 809 | if (err) |
| 810 | break; |
| 811 | } |
| 812 | |
| 813 | if (err != -ENOSPC28) |
| 814 | break; |
| 815 | } |
| 816 | |
| 817 | return err; |
| 818 | } |
| 819 | |
| 820 | static int eb_select_context(struct i915_execbuffer *eb) |
| 821 | { |
| 822 | struct i915_gem_context *ctx; |
| 823 | |
| 824 | ctx = i915_gem_context_lookup(eb->file->driver_priv, eb->args->rsvd1); |
| 825 | if (unlikely(IS_ERR(ctx))__builtin_expect(!!(IS_ERR(ctx)), 0)) |
| 826 | return PTR_ERR(ctx); |
| 827 | |
| 828 | eb->gem_context = ctx; |
| 829 | if (i915_gem_context_has_full_ppgtt(ctx)) |
| 830 | eb->invalid_flags |= EXEC_OBJECT_NEEDS_GTT(1<<1); |
| 831 | |
| 832 | return 0; |
| 833 | } |
| 834 | |
| 835 | static int __eb_add_lut(struct i915_execbuffer *eb, |
| 836 | u32 handle, struct i915_vma *vma) |
| 837 | { |
| 838 | struct i915_gem_context *ctx = eb->gem_context; |
| 839 | struct i915_lut_handle *lut; |
| 840 | int err; |
| 841 | |
| 842 | lut = i915_lut_handle_alloc(); |
| 843 | if (unlikely(!lut)__builtin_expect(!!(!lut), 0)) |
| 844 | return -ENOMEM12; |
| 845 | |
| 846 | i915_vma_get(vma); |
| 847 | if (!atomic_fetch_inc(&vma->open_count)__sync_fetch_and_add(&vma->open_count, 1)) |
| 848 | i915_vma_reopen(vma); |
| 849 | lut->handle = handle; |
| 850 | lut->ctx = ctx; |
| 851 | |
| 852 | /* Check that the context hasn't been closed in the meantime */ |
| 853 | err = -EINTR4; |
| 854 | if (!mutex_lock_interruptible(&ctx->lut_mutex)) { |
| 855 | if (likely(!i915_gem_context_is_closed(ctx))__builtin_expect(!!(!i915_gem_context_is_closed(ctx)), 1)) |
| 856 | err = radix_tree_insert(&ctx->handles_vma, handle, vma); |
| 857 | else |
| 858 | err = -ENOENT2; |
| 859 | if (err == 0) { /* And nor has this handle */ |
| 860 | struct drm_i915_gem_object *obj = vma->obj; |
| 861 | |
| 862 | spin_lock(&obj->lut_lock)mtx_enter(&obj->lut_lock); |
| 863 | if (idr_find(&eb->file->object_idr, handle) == obj) { |
| 864 | list_add(&lut->obj_link, &obj->lut_list); |
| 865 | } else { |
| 866 | radix_tree_delete(&ctx->handles_vma, handle); |
| 867 | err = -ENOENT2; |
| 868 | } |
| 869 | spin_unlock(&obj->lut_lock)mtx_leave(&obj->lut_lock); |
| 870 | } |
| 871 | mutex_unlock(&ctx->lut_mutex)rw_exit_write(&ctx->lut_mutex); |
| 872 | } |
| 873 | if (unlikely(err)__builtin_expect(!!(err), 0)) |
| 874 | goto err; |
| 875 | |
| 876 | return 0; |
| 877 | |
| 878 | err: |
| 879 | i915_vma_close(vma); |
| 880 | i915_vma_put(vma); |
| 881 | i915_lut_handle_free(lut); |
| 882 | return err; |
| 883 | } |
| 884 | |
| 885 | static struct i915_vma *eb_lookup_vma(struct i915_execbuffer *eb, u32 handle) |
| 886 | { |
| 887 | struct i915_address_space *vm = eb->context->vm; |
| 888 | |
| 889 | do { |
| 890 | struct drm_i915_gem_object *obj; |
| 891 | struct i915_vma *vma; |
| 892 | int err; |
| 893 | |
| 894 | rcu_read_lock(); |
| 895 | vma = radix_tree_lookup(&eb->gem_context->handles_vma, handle); |
| 896 | if (likely(vma && vma->vm == vm)__builtin_expect(!!(vma && vma->vm == vm), 1)) |
| 897 | vma = i915_vma_tryget(vma); |
| 898 | rcu_read_unlock(); |
| 899 | if (likely(vma)__builtin_expect(!!(vma), 1)) |
| 900 | return vma; |
| 901 | |
| 902 | obj = i915_gem_object_lookup(eb->file, handle); |
| 903 | if (unlikely(!obj)__builtin_expect(!!(!obj), 0)) |
| 904 | return ERR_PTR(-ENOENT2); |
| 905 | |
| 906 | /* |
| 907 | * If the user has opted-in for protected-object tracking, make |
| 908 | * sure the object encryption can be used. |
| 909 | * We only need to do this when the object is first used with |
| 910 | * this context, because the context itself will be banned when |
| 911 | * the protected objects become invalid. |
| 912 | */ |
| 913 | if (i915_gem_context_uses_protected_content(eb->gem_context) && |
| 914 | i915_gem_object_is_protected(obj)) { |
| 915 | err = intel_pxp_key_check(&vm->gt->pxp, obj, true1); |
| 916 | if (err) { |
| 917 | i915_gem_object_put(obj); |
| 918 | return ERR_PTR(err); |
| 919 | } |
| 920 | } |
| 921 | |
| 922 | vma = i915_vma_instance(obj, vm, NULL((void *)0)); |
| 923 | if (IS_ERR(vma)) { |
| 924 | i915_gem_object_put(obj); |
| 925 | return vma; |
| 926 | } |
| 927 | |
| 928 | err = __eb_add_lut(eb, handle, vma); |
| 929 | if (likely(!err)__builtin_expect(!!(!err), 1)) |
| 930 | return vma; |
| 931 | |
| 932 | i915_gem_object_put(obj); |
| 933 | if (err != -EEXIST17) |
| 934 | return ERR_PTR(err); |
| 935 | } while (1); |
| 936 | } |
| 937 | |
| 938 | static int eb_lookup_vmas(struct i915_execbuffer *eb) |
| 939 | { |
| 940 | unsigned int i, current_batch = 0; |
| 941 | int err = 0; |
| 942 | |
| 943 | INIT_LIST_HEAD(&eb->relocs); |
| 944 | |
| 945 | for (i = 0; i < eb->buffer_count; i++) { |
| 946 | struct i915_vma *vma; |
| 947 | |
| 948 | vma = eb_lookup_vma(eb, eb->exec[i].handle); |
| 949 | if (IS_ERR(vma)) { |
| 950 | err = PTR_ERR(vma); |
| 951 | goto err; |
| 952 | } |
| 953 | |
| 954 | err = eb_validate_vma(eb, &eb->exec[i], vma); |
| 955 | if (unlikely(err)__builtin_expect(!!(err), 0)) { |
| 956 | i915_vma_put(vma); |
| 957 | goto err; |
| 958 | } |
| 959 | |
| 960 | err = eb_add_vma(eb, ¤t_batch, i, vma); |
| 961 | if (err) |
| 962 | return err; |
| 963 | |
| 964 | if (i915_gem_object_is_userptr(vma->obj)) { |
| 965 | err = i915_gem_object_userptr_submit_init(vma->obj); |
| 966 | if (err) { |
| 967 | if (i + 1 < eb->buffer_count) { |
| 968 | /* |
| 969 | * Execbuffer code expects last vma entry to be NULL, |
| 970 | * since we already initialized this entry, |
| 971 | * set the next value to NULL or we mess up |
| 972 | * cleanup handling. |
| 973 | */ |
| 974 | eb->vma[i + 1].vma = NULL((void *)0); |
| 975 | } |
| 976 | |
| 977 | return err; |
| 978 | } |
| 979 | |
| 980 | eb->vma[i].flags |= __EXEC_OBJECT_USERPTR_INIT(1UL << (28)); |
| 981 | eb->args->flags |= __EXEC_USERPTR_USED(1UL << (29)); |
| 982 | } |
| 983 | } |
| 984 | |
| 985 | return 0; |
| 986 | |
| 987 | err: |
| 988 | eb->vma[i].vma = NULL((void *)0); |
| 989 | return err; |
| 990 | } |
| 991 | |
| 992 | static int eb_lock_vmas(struct i915_execbuffer *eb) |
| 993 | { |
| 994 | unsigned int i; |
| 995 | int err; |
| 996 | |
| 997 | for (i = 0; i < eb->buffer_count; i++) { |
| 998 | struct eb_vma *ev = &eb->vma[i]; |
| 999 | struct i915_vma *vma = ev->vma; |
| 1000 | |
| 1001 | err = i915_gem_object_lock(vma->obj, &eb->ww); |
| 1002 | if (err) |
| 1003 | return err; |
| 1004 | } |
| 1005 | |
| 1006 | return 0; |
| 1007 | } |
| 1008 | |
| 1009 | static int eb_validate_vmas(struct i915_execbuffer *eb) |
| 1010 | { |
| 1011 | unsigned int i; |
| 1012 | int err; |
| 1013 | |
| 1014 | INIT_LIST_HEAD(&eb->unbound); |
| 1015 | |
| 1016 | err = eb_lock_vmas(eb); |
| 1017 | if (err) |
| 1018 | return err; |
| 1019 | |
| 1020 | for (i = 0; i < eb->buffer_count; i++) { |
| 1021 | struct drm_i915_gem_exec_object2 *entry = &eb->exec[i]; |
| 1022 | struct eb_vma *ev = &eb->vma[i]; |
| 1023 | struct i915_vma *vma = ev->vma; |
| 1024 | |
| 1025 | err = eb_pin_vma(eb, entry, ev); |
| 1026 | if (err == -EDEADLK11) |
| 1027 | return err; |
| 1028 | |
| 1029 | if (!err) { |
| 1030 | if (entry->offset != vma->node.start) { |
| 1031 | entry->offset = vma->node.start | UPDATE(1ULL << (7)); |
| 1032 | eb->args->flags |= __EXEC_HAS_RELOC(1UL << (31)); |
| 1033 | } |
| 1034 | } else { |
| 1035 | eb_unreserve_vma(ev); |
| 1036 | |
| 1037 | list_add_tail(&ev->bind_link, &eb->unbound); |
| 1038 | if (drm_mm_node_allocated(&vma->node)) { |
| 1039 | err = i915_vma_unbind(vma); |
| 1040 | if (err) |
| 1041 | return err; |
| 1042 | } |
| 1043 | } |
| 1044 | |
| 1045 | /* Reserve enough slots to accommodate composite fences */ |
| 1046 | err = dma_resv_reserve_fences(vma->obj->base.resv, eb->num_batches); |
| 1047 | if (err) |
| 1048 | return err; |
| 1049 | |
| 1050 | GEM_BUG_ON(drm_mm_node_allocated(&vma->node) &&((void)0) |
| 1051 | eb_vma_misplaced(&eb->exec[i], vma, ev->flags))((void)0); |
| 1052 | } |
| 1053 | |
| 1054 | if (!list_empty(&eb->unbound)) |
| 1055 | return eb_reserve(eb); |
| 1056 | |
| 1057 | return 0; |
| 1058 | } |
| 1059 | |
| 1060 | static struct eb_vma * |
| 1061 | eb_get_vma(const struct i915_execbuffer *eb, unsigned long handle) |
| 1062 | { |
| 1063 | if (eb->lut_size < 0) { |
| 1064 | if (handle >= -eb->lut_size) |
| 1065 | return NULL((void *)0); |
| 1066 | return &eb->vma[handle]; |
| 1067 | } else { |
| 1068 | struct hlist_head *head; |
| 1069 | struct eb_vma *ev; |
| 1070 | |
| 1071 | head = &eb->buckets[hash_32(handle, eb->lut_size)]; |
| 1072 | hlist_for_each_entry(ev, head, node)for (ev = (((head)->first) ? ({ const __typeof( ((__typeof (*ev) *)0)->node ) *__mptr = ((head)->first); (__typeof (*ev) *)( (char *)__mptr - __builtin_offsetof(__typeof(*ev), node ) );}) : ((void *)0)); ev != ((void *)0); ev = (((ev)->node .next) ? ({ const __typeof( ((__typeof(*ev) *)0)->node ) * __mptr = ((ev)->node.next); (__typeof(*ev) *)( (char *)__mptr - __builtin_offsetof(__typeof(*ev), node) );}) : ((void *)0) )) { |
| 1073 | if (ev->handle == handle) |
| 1074 | return ev; |
| 1075 | } |
| 1076 | return NULL((void *)0); |
| 1077 | } |
| 1078 | } |
| 1079 | |
| 1080 | static void eb_release_vmas(struct i915_execbuffer *eb, bool_Bool final) |
| 1081 | { |
| 1082 | const unsigned int count = eb->buffer_count; |
| 1083 | unsigned int i; |
| 1084 | |
| 1085 | for (i = 0; i < count; i++) { |
| 1086 | struct eb_vma *ev = &eb->vma[i]; |
| 1087 | struct i915_vma *vma = ev->vma; |
| 1088 | |
| 1089 | if (!vma) |
| 1090 | break; |
| 1091 | |
| 1092 | eb_unreserve_vma(ev); |
| 1093 | |
| 1094 | if (final) |
| 1095 | i915_vma_put(vma); |
| 1096 | } |
| 1097 | |
| 1098 | eb_capture_release(eb); |
| 1099 | eb_unpin_engine(eb); |
| 1100 | } |
| 1101 | |
| 1102 | static void eb_destroy(const struct i915_execbuffer *eb) |
| 1103 | { |
| 1104 | if (eb->lut_size > 0) |
| 1105 | kfree(eb->buckets); |
| 1106 | } |
| 1107 | |
| 1108 | static inline u64 |
| 1109 | relocation_target(const struct drm_i915_gem_relocation_entry *reloc, |
| 1110 | const struct i915_vma *target) |
| 1111 | { |
| 1112 | return gen8_canonical_addr((int)reloc->delta + target->node.start); |
| 1113 | } |
| 1114 | |
| 1115 | static void reloc_cache_init(struct reloc_cache *cache, |
| 1116 | struct drm_i915_privateinteldrm_softc *i915) |
| 1117 | { |
| 1118 | cache->page = -1; |
| 1119 | cache->vaddr = 0; |
| 1120 | /* Must be a variable in the struct to allow GCC to unroll. */ |
| 1121 | cache->graphics_ver = GRAPHICS_VER(i915)((&(i915)->__runtime)->graphics.ip.ver); |
| 1122 | cache->has_llc = HAS_LLC(i915)((&(i915)->__info)->has_llc); |
| 1123 | cache->use_64bit_reloc = HAS_64BIT_RELOC(i915)((&(i915)->__info)->has_64bit_reloc); |
| 1124 | cache->has_fence = cache->graphics_ver < 4; |
| 1125 | cache->needs_unfenced = INTEL_INFO(i915)(&(i915)->__info)->unfenced_needs_alignment; |
| 1126 | cache->node.flags = 0; |
| 1127 | |
| 1128 | cache->map = i915->agph; |
| 1129 | cache->iot = i915->bst; |
| 1130 | } |
| 1131 | |
| 1132 | static inline void *unmask_page(unsigned long p) |
| 1133 | { |
| 1134 | return (void *)(uintptr_t)(p & LINUX_PAGE_MASK(~((1 << 12) - 1))); |
| 1135 | } |
| 1136 | |
| 1137 | static inline unsigned int unmask_flags(unsigned long p) |
| 1138 | { |
| 1139 | return p & ~LINUX_PAGE_MASK(~((1 << 12) - 1)); |
| 1140 | } |
| 1141 | |
| 1142 | #define KMAP0x4 0x4 /* after CLFLUSH_FLAGS */ |
| 1143 | |
| 1144 | static inline struct i915_ggtt *cache_to_ggtt(struct reloc_cache *cache) |
| 1145 | { |
| 1146 | struct drm_i915_privateinteldrm_softc *i915 = |
| 1147 | container_of(cache, struct i915_execbuffer, reloc_cache)({ const __typeof( ((struct i915_execbuffer *)0)->reloc_cache ) *__mptr = (cache); (struct i915_execbuffer *)( (char *)__mptr - __builtin_offsetof(struct i915_execbuffer, reloc_cache) ); })->i915; |
| 1148 | return to_gt(i915)->ggtt; |
| 1149 | } |
| 1150 | |
| 1151 | static void reloc_cache_unmap(struct reloc_cache *cache) |
| 1152 | { |
| 1153 | void *vaddr; |
| 1154 | |
| 1155 | if (!cache->vaddr) |
| 1156 | return; |
| 1157 | |
| 1158 | vaddr = unmask_page(cache->vaddr); |
| 1159 | if (cache->vaddr & KMAP0x4) |
| 1160 | kunmap_atomic(vaddr); |
| 1161 | else |
| 1162 | #ifdef __linux__ |
| 1163 | io_mapping_unmap_atomic((void __iomem *)vaddr); |
| 1164 | #else |
| 1165 | agp_unmap_atomic(cache->map, cache->ioh); |
| 1166 | #endif |
| 1167 | } |
| 1168 | |
| 1169 | static void reloc_cache_remap(struct reloc_cache *cache, |
| 1170 | struct drm_i915_gem_object *obj) |
| 1171 | { |
| 1172 | void *vaddr; |
| 1173 | |
| 1174 | if (!cache->vaddr) |
| 1175 | return; |
| 1176 | |
| 1177 | if (cache->vaddr & KMAP0x4) { |
| 1178 | struct vm_page *page = i915_gem_object_get_page(obj, cache->page); |
| 1179 | |
| 1180 | vaddr = kmap_atomic(page); |
| 1181 | cache->vaddr = unmask_flags(cache->vaddr) | |
| 1182 | (unsigned long)vaddr; |
| 1183 | } else { |
| 1184 | struct i915_ggtt *ggtt = cache_to_ggtt(cache); |
| 1185 | unsigned long offset; |
| 1186 | |
| 1187 | offset = cache->node.start; |
| 1188 | if (!drm_mm_node_allocated(&cache->node)) |
| 1189 | offset += cache->page << PAGE_SHIFT12; |
| 1190 | |
| 1191 | #ifdef __linux__ |
| 1192 | cache->vaddr = (unsigned long) |
| 1193 | io_mapping_map_atomic_wc(&ggtt->iomap, offset); |
| 1194 | #else |
| 1195 | agp_map_atomic(cache->map, offset, &cache->ioh); |
| 1196 | cache->vaddr = (unsigned long) |
| 1197 | bus_space_vaddr(cache->iot, cache->ioh)((cache->iot)->vaddr((cache->ioh))); |
| 1198 | #endif |
| 1199 | } |
| 1200 | } |
| 1201 | |
| 1202 | static void reloc_cache_reset(struct reloc_cache *cache, struct i915_execbuffer *eb) |
| 1203 | { |
| 1204 | void *vaddr; |
| 1205 | |
| 1206 | if (!cache->vaddr) |
| 1207 | return; |
| 1208 | |
| 1209 | vaddr = unmask_page(cache->vaddr); |
| 1210 | if (cache->vaddr & KMAP0x4) { |
| 1211 | struct drm_i915_gem_object *obj = |
| 1212 | (struct drm_i915_gem_object *)cache->node.mm; |
| 1213 | if (cache->vaddr & CLFLUSH_AFTER(1UL << (1))) |
| 1214 | mb()do { __asm volatile("mfence" ::: "memory"); } while (0); |
| 1215 | |
| 1216 | kunmap_atomic(vaddr); |
| 1217 | i915_gem_object_finish_access(obj); |
| 1218 | } else { |
| 1219 | struct i915_ggtt *ggtt = cache_to_ggtt(cache); |
| 1220 | |
| 1221 | intel_gt_flush_ggtt_writes(ggtt->vm.gt); |
| 1222 | #ifdef __linux__ |
| 1223 | io_mapping_unmap_atomic((void __iomem *)vaddr); |
| 1224 | #else |
| 1225 | agp_unmap_atomic(cache->map, cache->ioh); |
| 1226 | #endif |
| 1227 | |
| 1228 | if (drm_mm_node_allocated(&cache->node)) { |
| 1229 | ggtt->vm.clear_range(&ggtt->vm, |
| 1230 | cache->node.start, |
| 1231 | cache->node.size); |
| 1232 | mutex_lock(&ggtt->vm.mutex)rw_enter_write(&ggtt->vm.mutex); |
| 1233 | drm_mm_remove_node(&cache->node); |
| 1234 | mutex_unlock(&ggtt->vm.mutex)rw_exit_write(&ggtt->vm.mutex); |
| 1235 | } else { |
| 1236 | i915_vma_unpin((struct i915_vma *)cache->node.mm); |
| 1237 | } |
| 1238 | } |
| 1239 | |
| 1240 | cache->vaddr = 0; |
| 1241 | cache->page = -1; |
| 1242 | } |
| 1243 | |
| 1244 | static void *reloc_kmap(struct drm_i915_gem_object *obj, |
| 1245 | struct reloc_cache *cache, |
| 1246 | unsigned long pageno) |
| 1247 | { |
| 1248 | void *vaddr; |
| 1249 | struct vm_page *page; |
| 1250 | |
| 1251 | if (cache->vaddr) { |
| 1252 | kunmap_atomic(unmask_page(cache->vaddr)); |
| 1253 | } else { |
| 1254 | unsigned int flushes; |
| 1255 | int err; |
| 1256 | |
| 1257 | err = i915_gem_object_prepare_write(obj, &flushes); |
| 1258 | if (err) |
| 1259 | return ERR_PTR(err); |
| 1260 | |
| 1261 | BUILD_BUG_ON(KMAP & CLFLUSH_FLAGS)extern char _ctassert[(!(0x4 & ((1UL << (0)) | (1UL << (1))))) ? 1 : -1 ] __attribute__((__unused__)); |
| 1262 | BUILD_BUG_ON((KMAP | CLFLUSH_FLAGS) & LINUX_PAGE_MASK)extern char _ctassert[(!((0x4 | ((1UL << (0)) | (1UL << (1)))) & (~((1 << 12) - 1)))) ? 1 : -1 ] __attribute__ ((__unused__)); |
| 1263 | |
| 1264 | cache->vaddr = flushes | KMAP0x4; |
| 1265 | cache->node.mm = (void *)obj; |
| 1266 | if (flushes) |
| 1267 | mb()do { __asm volatile("mfence" ::: "memory"); } while (0); |
| 1268 | } |
| 1269 | |
| 1270 | page = i915_gem_object_get_page(obj, pageno); |
| 1271 | if (!obj->mm.dirty) |
| 1272 | set_page_dirty(page)x86_atomic_clearbits_u32(&page->pg_flags, 0x00000008); |
| 1273 | |
| 1274 | vaddr = kmap_atomic(page); |
| 1275 | cache->vaddr = unmask_flags(cache->vaddr) | (unsigned long)vaddr; |
| 1276 | cache->page = pageno; |
| 1277 | |
| 1278 | return vaddr; |
| 1279 | } |
| 1280 | |
| 1281 | static void *reloc_iomap(struct i915_vma *batch, |
| 1282 | struct i915_execbuffer *eb, |
| 1283 | unsigned long page) |
| 1284 | { |
| 1285 | struct drm_i915_gem_object *obj = batch->obj; |
| 1286 | struct reloc_cache *cache = &eb->reloc_cache; |
| 1287 | struct i915_ggtt *ggtt = cache_to_ggtt(cache); |
| 1288 | unsigned long offset; |
| 1289 | void *vaddr; |
| 1290 | |
| 1291 | if (cache->vaddr) { |
| 1292 | intel_gt_flush_ggtt_writes(ggtt->vm.gt); |
| 1293 | #ifdef __linux__ |
| 1294 | io_mapping_unmap_atomic((void __force __iomem *) unmask_page(cache->vaddr)); |
| 1295 | #else |
| 1296 | agp_unmap_atomic(cache->map, cache->ioh); |
| 1297 | #endif |
| 1298 | } else { |
| 1299 | struct i915_vma *vma = ERR_PTR(-ENODEV19); |
| 1300 | int err; |
| 1301 | |
| 1302 | if (i915_gem_object_is_tiled(obj)) |
| 1303 | return ERR_PTR(-EINVAL22); |
| 1304 | |
| 1305 | if (use_cpu_reloc(cache, obj)) |
| 1306 | return NULL((void *)0); |
| 1307 | |
| 1308 | err = i915_gem_object_set_to_gtt_domain(obj, true1); |
| 1309 | if (err) |
| 1310 | return ERR_PTR(err); |
| 1311 | |
| 1312 | /* |
| 1313 | * i915_gem_object_ggtt_pin_ww may attempt to remove the batch |
| 1314 | * VMA from the object list because we no longer pin. |
| 1315 | * |
| 1316 | * Only attempt to pin the batch buffer to ggtt if the current batch |
| 1317 | * is not inside ggtt, or the batch buffer is not misplaced. |
| 1318 | */ |
| 1319 | if (!i915_is_ggtt(batch->vm)((batch->vm)->is_ggtt) || |
| 1320 | !i915_vma_misplaced(batch, 0, 0, PIN_MAPPABLE(1ULL << (3)))) { |
| 1321 | vma = i915_gem_object_ggtt_pin_ww(obj, &eb->ww, NULL((void *)0), 0, 0, |
| 1322 | PIN_MAPPABLE(1ULL << (3)) | |
| 1323 | PIN_NONBLOCK(1ULL << (2)) /* NOWARN */ | |
| 1324 | PIN_NOEVICT(1ULL << (0))); |
| 1325 | } |
| 1326 | |
| 1327 | if (vma == ERR_PTR(-EDEADLK11)) |
| 1328 | return vma; |
| 1329 | |
| 1330 | if (IS_ERR(vma)) { |
| 1331 | memset(&cache->node, 0, sizeof(cache->node))__builtin_memset((&cache->node), (0), (sizeof(cache-> node))); |
| 1332 | mutex_lock(&ggtt->vm.mutex)rw_enter_write(&ggtt->vm.mutex); |
| 1333 | err = drm_mm_insert_node_in_range |
| 1334 | (&ggtt->vm.mm, &cache->node, |
| 1335 | PAGE_SIZE(1 << 12), 0, I915_COLOR_UNEVICTABLE(-1), |
| 1336 | 0, ggtt->mappable_end, |
| 1337 | DRM_MM_INSERT_LOW); |
| 1338 | mutex_unlock(&ggtt->vm.mutex)rw_exit_write(&ggtt->vm.mutex); |
| 1339 | if (err) /* no inactive aperture space, use cpu reloc */ |
| 1340 | return NULL((void *)0); |
| 1341 | } else { |
| 1342 | cache->node.start = vma->node.start; |
| 1343 | cache->node.mm = (void *)vma; |
| 1344 | } |
| 1345 | } |
| 1346 | |
| 1347 | offset = cache->node.start; |
| 1348 | if (drm_mm_node_allocated(&cache->node)) { |
| 1349 | ggtt->vm.insert_page(&ggtt->vm, |
| 1350 | i915_gem_object_get_dma_address(obj, page), |
| 1351 | offset, I915_CACHE_NONE, 0); |
| 1352 | } else { |
| 1353 | offset += page << PAGE_SHIFT12; |
| 1354 | } |
| 1355 | |
| 1356 | #ifdef __linux__ |
| 1357 | vaddr = (void __force *)io_mapping_map_atomic_wc(&ggtt->iomap, |
| 1358 | offset); |
| 1359 | #else |
| 1360 | agp_map_atomic(cache->map, offset, &cache->ioh); |
| 1361 | vaddr = bus_space_vaddr(cache->iot, cache->ioh)((cache->iot)->vaddr((cache->ioh))); |
| 1362 | #endif |
| 1363 | cache->page = page; |
| 1364 | cache->vaddr = (unsigned long)vaddr; |
| 1365 | |
| 1366 | return vaddr; |
| 1367 | } |
| 1368 | |
| 1369 | static void *reloc_vaddr(struct i915_vma *vma, |
| 1370 | struct i915_execbuffer *eb, |
| 1371 | unsigned long page) |
| 1372 | { |
| 1373 | struct reloc_cache *cache = &eb->reloc_cache; |
| 1374 | void *vaddr; |
| 1375 | |
| 1376 | if (cache->page == page) { |
| 1377 | vaddr = unmask_page(cache->vaddr); |
| 1378 | } else { |
| 1379 | vaddr = NULL((void *)0); |
| 1380 | if ((cache->vaddr & KMAP0x4) == 0) |
| 1381 | vaddr = reloc_iomap(vma, eb, page); |
| 1382 | if (!vaddr) |
| 1383 | vaddr = reloc_kmap(vma->obj, cache, page); |
| 1384 | } |
| 1385 | |
| 1386 | return vaddr; |
| 1387 | } |
| 1388 | |
| 1389 | static void clflush_write32(u32 *addr, u32 value, unsigned int flushes) |
| 1390 | { |
| 1391 | if (unlikely(flushes & (CLFLUSH_BEFORE | CLFLUSH_AFTER))__builtin_expect(!!(flushes & ((1UL << (0)) | (1UL << (1)))), 0)) { |
| 1392 | if (flushes & CLFLUSH_BEFORE(1UL << (0))) |
| 1393 | drm_clflush_virt_range(addr, sizeof(*addr)); |
| 1394 | |
| 1395 | *addr = value; |
| 1396 | |
| 1397 | /* |
| 1398 | * Writes to the same cacheline are serialised by the CPU |
| 1399 | * (including clflush). On the write path, we only require |
| 1400 | * that it hits memory in an orderly fashion and place |
| 1401 | * mb barriers at the start and end of the relocation phase |
| 1402 | * to ensure ordering of clflush wrt to the system. |
| 1403 | */ |
| 1404 | if (flushes & CLFLUSH_AFTER(1UL << (1))) |
| 1405 | drm_clflush_virt_range(addr, sizeof(*addr)); |
| 1406 | } else |
| 1407 | *addr = value; |
| 1408 | } |
| 1409 | |
| 1410 | static u64 |
| 1411 | relocate_entry(struct i915_vma *vma, |
| 1412 | const struct drm_i915_gem_relocation_entry *reloc, |
| 1413 | struct i915_execbuffer *eb, |
| 1414 | const struct i915_vma *target) |
| 1415 | { |
| 1416 | u64 target_addr = relocation_target(reloc, target); |
| 1417 | u64 offset = reloc->offset; |
| 1418 | bool_Bool wide = eb->reloc_cache.use_64bit_reloc; |
| 1419 | void *vaddr; |
| 1420 | |
| 1421 | repeat: |
| 1422 | vaddr = reloc_vaddr(vma, eb, |
| 1423 | offset >> PAGE_SHIFT12); |
| 1424 | if (IS_ERR(vaddr)) |
| 1425 | return PTR_ERR(vaddr); |
| 1426 | |
| 1427 | GEM_BUG_ON(!IS_ALIGNED(offset, sizeof(u32)))((void)0); |
| 1428 | clflush_write32(vaddr + offset_in_page(offset)((vaddr_t)(offset) & ((1 << 12) - 1)), |
| 1429 | lower_32_bits(target_addr)((u32)(target_addr)), |
| 1430 | eb->reloc_cache.vaddr); |
| 1431 | |
| 1432 | if (wide) { |
| 1433 | offset += sizeof(u32); |
| 1434 | target_addr >>= 32; |
| 1435 | wide = false0; |
| 1436 | goto repeat; |
| 1437 | } |
| 1438 | |
| 1439 | return target->node.start | UPDATE(1ULL << (7)); |
| 1440 | } |
| 1441 | |
| 1442 | static u64 |
| 1443 | eb_relocate_entry(struct i915_execbuffer *eb, |
| 1444 | struct eb_vma *ev, |
| 1445 | const struct drm_i915_gem_relocation_entry *reloc) |
| 1446 | { |
| 1447 | struct drm_i915_privateinteldrm_softc *i915 = eb->i915; |
| 1448 | struct eb_vma *target; |
| 1449 | int err; |
| 1450 | |
| 1451 | /* we've already hold a reference to all valid objects */ |
| 1452 | target = eb_get_vma(eb, reloc->target_handle); |
| 1453 | if (unlikely(!target)__builtin_expect(!!(!target), 0)) |
| 1454 | return -ENOENT2; |
| 1455 | |
| 1456 | /* Validate that the target is in a valid r/w GPU domain */ |
| 1457 | if (unlikely(reloc->write_domain & (reloc->write_domain - 1))__builtin_expect(!!(reloc->write_domain & (reloc->write_domain - 1)), 0)) { |
| 1458 | drm_dbg(&i915->drm, "reloc with multiple write domains: "__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "reloc with multiple write domains: " "target %d offset %d " "read %08x write %08x", reloc->target_handle , (int) reloc->offset, reloc->read_domains, reloc->write_domain ) |
| 1459 | "target %d offset %d "__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "reloc with multiple write domains: " "target %d offset %d " "read %08x write %08x", reloc->target_handle , (int) reloc->offset, reloc->read_domains, reloc->write_domain ) |
| 1460 | "read %08x write %08x",__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "reloc with multiple write domains: " "target %d offset %d " "read %08x write %08x", reloc->target_handle , (int) reloc->offset, reloc->read_domains, reloc->write_domain ) |
| 1461 | reloc->target_handle,__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "reloc with multiple write domains: " "target %d offset %d " "read %08x write %08x", reloc->target_handle , (int) reloc->offset, reloc->read_domains, reloc->write_domain ) |
| 1462 | (int) reloc->offset,__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "reloc with multiple write domains: " "target %d offset %d " "read %08x write %08x", reloc->target_handle , (int) reloc->offset, reloc->read_domains, reloc->write_domain ) |
| 1463 | reloc->read_domains,__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "reloc with multiple write domains: " "target %d offset %d " "read %08x write %08x", reloc->target_handle , (int) reloc->offset, reloc->read_domains, reloc->write_domain ) |
| 1464 | reloc->write_domain)__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "reloc with multiple write domains: " "target %d offset %d " "read %08x write %08x", reloc->target_handle , (int) reloc->offset, reloc->read_domains, reloc->write_domain ); |
| 1465 | return -EINVAL22; |
| 1466 | } |
| 1467 | if (unlikely((reloc->write_domain | reloc->read_domains)__builtin_expect(!!((reloc->write_domain | reloc->read_domains ) & ~(0x00000002 | 0x00000004 | 0x00000008 | 0x00000010 | 0x00000020)), 0) |
| 1468 | & ~I915_GEM_GPU_DOMAINS)__builtin_expect(!!((reloc->write_domain | reloc->read_domains ) & ~(0x00000002 | 0x00000004 | 0x00000008 | 0x00000010 | 0x00000020)), 0)) { |
| 1469 | drm_dbg(&i915->drm, "reloc with read/write non-GPU domains: "__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "reloc with read/write non-GPU domains: " "target %d offset %d " "read %08x write %08x", reloc->target_handle , (int) reloc->offset, reloc->read_domains, reloc->write_domain ) |
| 1470 | "target %d offset %d "__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "reloc with read/write non-GPU domains: " "target %d offset %d " "read %08x write %08x", reloc->target_handle , (int) reloc->offset, reloc->read_domains, reloc->write_domain ) |
| 1471 | "read %08x write %08x",__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "reloc with read/write non-GPU domains: " "target %d offset %d " "read %08x write %08x", reloc->target_handle , (int) reloc->offset, reloc->read_domains, reloc->write_domain ) |
| 1472 | reloc->target_handle,__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "reloc with read/write non-GPU domains: " "target %d offset %d " "read %08x write %08x", reloc->target_handle , (int) reloc->offset, reloc->read_domains, reloc->write_domain ) |
| 1473 | (int) reloc->offset,__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "reloc with read/write non-GPU domains: " "target %d offset %d " "read %08x write %08x", reloc->target_handle , (int) reloc->offset, reloc->read_domains, reloc->write_domain ) |
| 1474 | reloc->read_domains,__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "reloc with read/write non-GPU domains: " "target %d offset %d " "read %08x write %08x", reloc->target_handle , (int) reloc->offset, reloc->read_domains, reloc->write_domain ) |
| 1475 | reloc->write_domain)__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "reloc with read/write non-GPU domains: " "target %d offset %d " "read %08x write %08x", reloc->target_handle , (int) reloc->offset, reloc->read_domains, reloc->write_domain ); |
| 1476 | return -EINVAL22; |
| 1477 | } |
| 1478 | |
| 1479 | if (reloc->write_domain) { |
| 1480 | target->flags |= EXEC_OBJECT_WRITE(1<<2); |
| 1481 | |
| 1482 | /* |
| 1483 | * Sandybridge PPGTT errata: We need a global gtt mapping |
| 1484 | * for MI and pipe_control writes because the gpu doesn't |
| 1485 | * properly redirect them through the ppgtt for non_secure |
| 1486 | * batchbuffers. |
| 1487 | */ |
| 1488 | if (reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION0x00000010 && |
| 1489 | GRAPHICS_VER(eb->i915)((&(eb->i915)->__runtime)->graphics.ip.ver) == 6 && |
| 1490 | !i915_vma_is_bound(target->vma, I915_VMA_GLOBAL_BIND((int)(1UL << (10))))) { |
| 1491 | struct i915_vma *vma = target->vma; |
| 1492 | |
| 1493 | reloc_cache_unmap(&eb->reloc_cache); |
| 1494 | mutex_lock(&vma->vm->mutex)rw_enter_write(&vma->vm->mutex); |
| 1495 | err = i915_vma_bind(target->vma, |
| 1496 | target->vma->obj->cache_level, |
| 1497 | PIN_GLOBAL(1ULL << (10)), NULL((void *)0), NULL((void *)0)); |
| 1498 | mutex_unlock(&vma->vm->mutex)rw_exit_write(&vma->vm->mutex); |
| 1499 | reloc_cache_remap(&eb->reloc_cache, ev->vma->obj); |
| 1500 | if (err) |
| 1501 | return err; |
| 1502 | } |
| 1503 | } |
| 1504 | |
| 1505 | /* |
| 1506 | * If the relocation already has the right value in it, no |
| 1507 | * more work needs to be done. |
| 1508 | */ |
| 1509 | if (!DBG_FORCE_RELOC0 && |
| 1510 | gen8_canonical_addr(target->vma->node.start) == reloc->presumed_offset) |
| 1511 | return 0; |
| 1512 | |
| 1513 | /* Check that the relocation address is valid... */ |
| 1514 | if (unlikely(reloc->offset >__builtin_expect(!!(reloc->offset > ev->vma->size - (eb->reloc_cache.use_64bit_reloc ? 8 : 4)), 0) |
| 1515 | ev->vma->size - (eb->reloc_cache.use_64bit_reloc ? 8 : 4))__builtin_expect(!!(reloc->offset > ev->vma->size - (eb->reloc_cache.use_64bit_reloc ? 8 : 4)), 0)) { |
| 1516 | drm_dbg(&i915->drm, "Relocation beyond object bounds: "__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Relocation beyond object bounds: " "target %d offset %d size %d.\n", reloc->target_handle, ( int)reloc->offset, (int)ev->vma->size) |
| 1517 | "target %d offset %d size %d.\n",__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Relocation beyond object bounds: " "target %d offset %d size %d.\n", reloc->target_handle, ( int)reloc->offset, (int)ev->vma->size) |
| 1518 | reloc->target_handle,__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Relocation beyond object bounds: " "target %d offset %d size %d.\n", reloc->target_handle, ( int)reloc->offset, (int)ev->vma->size) |
| 1519 | (int)reloc->offset,__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Relocation beyond object bounds: " "target %d offset %d size %d.\n", reloc->target_handle, ( int)reloc->offset, (int)ev->vma->size) |
| 1520 | (int)ev->vma->size)__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Relocation beyond object bounds: " "target %d offset %d size %d.\n", reloc->target_handle, ( int)reloc->offset, (int)ev->vma->size); |
| 1521 | return -EINVAL22; |
| 1522 | } |
| 1523 | if (unlikely(reloc->offset & 3)__builtin_expect(!!(reloc->offset & 3), 0)) { |
| 1524 | drm_dbg(&i915->drm, "Relocation not 4-byte aligned: "__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Relocation not 4-byte aligned: " "target %d offset %d.\n", reloc->target_handle, (int)reloc ->offset) |
| 1525 | "target %d offset %d.\n",__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Relocation not 4-byte aligned: " "target %d offset %d.\n", reloc->target_handle, (int)reloc ->offset) |
| 1526 | reloc->target_handle,__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Relocation not 4-byte aligned: " "target %d offset %d.\n", reloc->target_handle, (int)reloc ->offset) |
| 1527 | (int)reloc->offset)__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Relocation not 4-byte aligned: " "target %d offset %d.\n", reloc->target_handle, (int)reloc ->offset); |
| 1528 | return -EINVAL22; |
| 1529 | } |
| 1530 | |
| 1531 | /* |
| 1532 | * If we write into the object, we need to force the synchronisation |
| 1533 | * barrier, either with an asynchronous clflush or if we executed the |
| 1534 | * patching using the GPU (though that should be serialised by the |
| 1535 | * timeline). To be completely sure, and since we are required to |
| 1536 | * do relocations we are already stalling, disable the user's opt |
| 1537 | * out of our synchronisation. |
| 1538 | */ |
| 1539 | ev->flags &= ~EXEC_OBJECT_ASYNC(1<<6); |
| 1540 | |
| 1541 | /* and update the user's relocation entry */ |
| 1542 | return relocate_entry(ev->vma, reloc, eb, target->vma); |
| 1543 | } |
| 1544 | |
| 1545 | static int eb_relocate_vma(struct i915_execbuffer *eb, struct eb_vma *ev) |
| 1546 | { |
| 1547 | #define N_RELOC(x)((x) / sizeof(struct drm_i915_gem_relocation_entry)) ((x) / sizeof(struct drm_i915_gem_relocation_entry)) |
| 1548 | struct drm_i915_gem_relocation_entry stack[N_RELOC(512)((512) / sizeof(struct drm_i915_gem_relocation_entry))]; |
| 1549 | const struct drm_i915_gem_exec_object2 *entry = ev->exec; |
| 1550 | struct drm_i915_gem_relocation_entry __user *urelocs = |
| 1551 | u64_to_user_ptr(entry->relocs_ptr)((void *)(uintptr_t)(entry->relocs_ptr)); |
| 1552 | unsigned long remain = entry->relocation_count; |
| 1553 | |
| 1554 | if (unlikely(remain > N_RELOC(ULONG_MAX))__builtin_expect(!!(remain > ((0xffffffffffffffffUL) / sizeof (struct drm_i915_gem_relocation_entry))), 0)) |
| 1555 | return -EINVAL22; |
| 1556 | |
| 1557 | /* |
| 1558 | * We must check that the entire relocation array is safe |
| 1559 | * to read. However, if the array is not writable the user loses |
| 1560 | * the updated relocation values. |
| 1561 | */ |
| 1562 | if (unlikely(!access_ok(urelocs, remain * sizeof(*urelocs)))__builtin_expect(!!(!access_ok(urelocs, remain * sizeof(*urelocs ))), 0)) |
| 1563 | return -EFAULT14; |
| 1564 | |
| 1565 | do { |
| 1566 | struct drm_i915_gem_relocation_entry *r = stack; |
| 1567 | unsigned int count = |
| 1568 | min_t(unsigned long, remain, ARRAY_SIZE(stack))({ unsigned long __min_a = (remain); unsigned long __min_b = ( (sizeof((stack)) / sizeof((stack)[0]))); __min_a < __min_b ? __min_a : __min_b; }); |
| 1569 | unsigned int copied; |
| 1570 | |
| 1571 | /* |
| 1572 | * This is the fast path and we cannot handle a pagefault |
| 1573 | * whilst holding the struct mutex lest the user pass in the |
| 1574 | * relocations contained within a mmaped bo. For in such a case |
| 1575 | * we, the page fault handler would call i915_gem_fault() and |
| 1576 | * we would try to acquire the struct mutex again. Obviously |
| 1577 | * this is bad and so lockdep complains vehemently. |
| 1578 | */ |
| 1579 | pagefault_disable(); |
| 1580 | copied = __copy_from_user_inatomic(r, urelocs, count * sizeof(r[0])); |
| 1581 | pagefault_enable(); |
| 1582 | if (unlikely(copied)__builtin_expect(!!(copied), 0)) { |
| 1583 | remain = -EFAULT14; |
| 1584 | goto out; |
| 1585 | } |
| 1586 | |
| 1587 | remain -= count; |
| 1588 | do { |
| 1589 | u64 offset = eb_relocate_entry(eb, ev, r); |
| 1590 | |
| 1591 | if (likely(offset == 0)__builtin_expect(!!(offset == 0), 1)) { |
| 1592 | } else if ((s64)offset < 0) { |
| 1593 | remain = (int)offset; |
| 1594 | goto out; |
| 1595 | } else { |
| 1596 | /* |
| 1597 | * Note that reporting an error now |
| 1598 | * leaves everything in an inconsistent |
| 1599 | * state as we have *already* changed |
| 1600 | * the relocation value inside the |
| 1601 | * object. As we have not changed the |
| 1602 | * reloc.presumed_offset or will not |
| 1603 | * change the execobject.offset, on the |
| 1604 | * call we may not rewrite the value |
| 1605 | * inside the object, leaving it |
| 1606 | * dangling and causing a GPU hang. Unless |
| 1607 | * userspace dynamically rebuilds the |
| 1608 | * relocations on each execbuf rather than |
| 1609 | * presume a static tree. |
| 1610 | * |
| 1611 | * We did previously check if the relocations |
| 1612 | * were writable (access_ok), an error now |
| 1613 | * would be a strange race with mprotect, |
| 1614 | * having already demonstrated that we |
| 1615 | * can read from this userspace address. |
| 1616 | */ |
| 1617 | offset = gen8_canonical_addr(offset & ~UPDATE(1ULL << (7))); |
| 1618 | __put_user(offset,({ __typeof(((offset))) __tmp = ((offset)); -copyout(&(__tmp ), (&urelocs[r - stack].presumed_offset), sizeof(__tmp)); }) |
| 1619 | &urelocs[r - stack].presumed_offset)({ __typeof(((offset))) __tmp = ((offset)); -copyout(&(__tmp ), (&urelocs[r - stack].presumed_offset), sizeof(__tmp)); }); |
| 1620 | } |
| 1621 | } while (r++, --count); |
| 1622 | urelocs += ARRAY_SIZE(stack)(sizeof((stack)) / sizeof((stack)[0])); |
| 1623 | } while (remain); |
| 1624 | out: |
| 1625 | reloc_cache_reset(&eb->reloc_cache, eb); |
| 1626 | return remain; |
| 1627 | } |
| 1628 | |
| 1629 | static int |
| 1630 | eb_relocate_vma_slow(struct i915_execbuffer *eb, struct eb_vma *ev) |
| 1631 | { |
| 1632 | const struct drm_i915_gem_exec_object2 *entry = ev->exec; |
| 1633 | struct drm_i915_gem_relocation_entry *relocs = |
| 1634 | u64_to_ptr(typeof(*relocs), entry->relocs_ptr)({ 1; (typeof(*relocs) *)(uintptr_t)(entry->relocs_ptr); } ); |
| 1635 | unsigned int i; |
| 1636 | int err; |
| 1637 | |
| 1638 | for (i = 0; i < entry->relocation_count; i++) { |
| 1639 | u64 offset = eb_relocate_entry(eb, ev, &relocs[i]); |
| 1640 | |
| 1641 | if ((s64)offset < 0) { |
| 1642 | err = (int)offset; |
| 1643 | goto err; |
| 1644 | } |
| 1645 | } |
| 1646 | err = 0; |
| 1647 | err: |
| 1648 | reloc_cache_reset(&eb->reloc_cache, eb); |
| 1649 | return err; |
| 1650 | } |
| 1651 | |
| 1652 | static int check_relocations(const struct drm_i915_gem_exec_object2 *entry) |
| 1653 | { |
| 1654 | const char __user *addr, *end; |
| 1655 | unsigned long size; |
| 1656 | char __maybe_unused__attribute__((__unused__)) c; |
| 1657 | |
| 1658 | size = entry->relocation_count; |
| 1659 | if (size == 0) |
| 1660 | return 0; |
| 1661 | |
| 1662 | if (size > N_RELOC(ULONG_MAX)((0xffffffffffffffffUL) / sizeof(struct drm_i915_gem_relocation_entry ))) |
| 1663 | return -EINVAL22; |
| 1664 | |
| 1665 | addr = u64_to_user_ptr(entry->relocs_ptr)((void *)(uintptr_t)(entry->relocs_ptr)); |
| 1666 | size *= sizeof(struct drm_i915_gem_relocation_entry); |
| 1667 | if (!access_ok(addr, size)) |
| 1668 | return -EFAULT14; |
| 1669 | |
| 1670 | end = addr + size; |
| 1671 | for (; addr < end; addr += PAGE_SIZE(1 << 12)) { |
| 1672 | int err = __get_user(c, addr)-copyin((addr), &((c)), sizeof((c))); |
| 1673 | if (err) |
| 1674 | return err; |
| 1675 | } |
| 1676 | return __get_user(c, end - 1)-copyin((end - 1), &((c)), sizeof((c))); |
| 1677 | } |
| 1678 | |
| 1679 | static int eb_copy_relocations(const struct i915_execbuffer *eb) |
| 1680 | { |
| 1681 | struct drm_i915_gem_relocation_entry *relocs; |
| 1682 | const unsigned int count = eb->buffer_count; |
| 1683 | unsigned int i; |
| 1684 | int err; |
| 1685 | |
| 1686 | for (i = 0; i < count; i++) { |
| 1687 | const unsigned int nreloc = eb->exec[i].relocation_count; |
| 1688 | struct drm_i915_gem_relocation_entry __user *urelocs; |
| 1689 | unsigned long size; |
| 1690 | unsigned long copied; |
| 1691 | |
| 1692 | if (nreloc == 0) |
| 1693 | continue; |
| 1694 | |
| 1695 | err = check_relocations(&eb->exec[i]); |
| 1696 | if (err) |
| 1697 | goto err; |
| 1698 | |
| 1699 | urelocs = u64_to_user_ptr(eb->exec[i].relocs_ptr)((void *)(uintptr_t)(eb->exec[i].relocs_ptr)); |
| 1700 | size = nreloc * sizeof(*relocs); |
| 1701 | |
| 1702 | relocs = kvmalloc_array(size, 1, GFP_KERNEL(0x0001 | 0x0004)); |
| 1703 | if (!relocs) { |
| 1704 | err = -ENOMEM12; |
| 1705 | goto err; |
| 1706 | } |
| 1707 | |
| 1708 | /* copy_from_user is limited to < 4GiB */ |
| 1709 | copied = 0; |
| 1710 | do { |
| 1711 | unsigned int len = |
| 1712 | min_t(u64, BIT_ULL(31), size - copied)({ u64 __min_a = ((1ULL << (31))); u64 __min_b = (size - copied); __min_a < __min_b ? __min_a : __min_b; }); |
| 1713 | |
| 1714 | if (__copy_from_user((char *)relocs + copied, |
| 1715 | (char __user *)urelocs + copied, |
| 1716 | len)) |
| 1717 | goto end; |
| 1718 | |
| 1719 | copied += len; |
| 1720 | } while (copied < size); |
| 1721 | |
| 1722 | /* |
| 1723 | * As we do not update the known relocation offsets after |
| 1724 | * relocating (due to the complexities in lock handling), |
| 1725 | * we need to mark them as invalid now so that we force the |
| 1726 | * relocation processing next time. Just in case the target |
| 1727 | * object is evicted and then rebound into its old |
| 1728 | * presumed_offset before the next execbuffer - if that |
| 1729 | * happened we would make the mistake of assuming that the |
| 1730 | * relocations were valid. |
| 1731 | */ |
| 1732 | if (!user_access_begin(urelocs, size)access_ok(urelocs, size)) |
| 1733 | goto end; |
| 1734 | |
| 1735 | for (copied = 0; copied < nreloc; copied++) |
| 1736 | unsafe_put_user(-1,({ __typeof((-1)) __tmp = (-1); if (copyout(&(__tmp), & urelocs[copied].presumed_offset, sizeof(__tmp)) != 0) goto end_user ; }) |
| 1737 | &urelocs[copied].presumed_offset,({ __typeof((-1)) __tmp = (-1); if (copyout(&(__tmp), & urelocs[copied].presumed_offset, sizeof(__tmp)) != 0) goto end_user ; }) |
| 1738 | end_user)({ __typeof((-1)) __tmp = (-1); if (copyout(&(__tmp), & urelocs[copied].presumed_offset, sizeof(__tmp)) != 0) goto end_user ; }); |
| 1739 | user_access_end(); |
| 1740 | |
| 1741 | eb->exec[i].relocs_ptr = (uintptr_t)relocs; |
| 1742 | } |
| 1743 | |
| 1744 | return 0; |
| 1745 | |
| 1746 | end_user: |
| 1747 | user_access_end(); |
| 1748 | end: |
| 1749 | kvfree(relocs); |
| 1750 | err = -EFAULT14; |
| 1751 | err: |
| 1752 | while (i--) { |
| 1753 | relocs = u64_to_ptr(typeof(*relocs), eb->exec[i].relocs_ptr)({ 1; (typeof(*relocs) *)(uintptr_t)(eb->exec[i].relocs_ptr ); }); |
| 1754 | if (eb->exec[i].relocation_count) |
| 1755 | kvfree(relocs); |
| 1756 | } |
| 1757 | return err; |
| 1758 | } |
| 1759 | |
| 1760 | static int eb_prefault_relocations(const struct i915_execbuffer *eb) |
| 1761 | { |
| 1762 | const unsigned int count = eb->buffer_count; |
| 1763 | unsigned int i; |
| 1764 | |
| 1765 | for (i = 0; i < count; i++) { |
| 1766 | int err; |
| 1767 | |
| 1768 | err = check_relocations(&eb->exec[i]); |
| 1769 | if (err) |
| 1770 | return err; |
| 1771 | } |
| 1772 | |
| 1773 | return 0; |
| 1774 | } |
| 1775 | |
| 1776 | static int eb_reinit_userptr(struct i915_execbuffer *eb) |
| 1777 | { |
| 1778 | const unsigned int count = eb->buffer_count; |
| 1779 | unsigned int i; |
| 1780 | int ret; |
| 1781 | |
| 1782 | if (likely(!(eb->args->flags & __EXEC_USERPTR_USED))__builtin_expect(!!(!(eb->args->flags & (1UL << (29)))), 1)) |
| 1783 | return 0; |
| 1784 | |
| 1785 | for (i = 0; i < count; i++) { |
| 1786 | struct eb_vma *ev = &eb->vma[i]; |
| 1787 | |
| 1788 | if (!i915_gem_object_is_userptr(ev->vma->obj)) |
| 1789 | continue; |
| 1790 | |
| 1791 | ret = i915_gem_object_userptr_submit_init(ev->vma->obj); |
| 1792 | if (ret) |
| 1793 | return ret; |
| 1794 | |
| 1795 | ev->flags |= __EXEC_OBJECT_USERPTR_INIT(1UL << (28)); |
| 1796 | } |
| 1797 | |
| 1798 | return 0; |
| 1799 | } |
| 1800 | |
| 1801 | static noinline__attribute__((__noinline__)) int eb_relocate_parse_slow(struct i915_execbuffer *eb) |
| 1802 | { |
| 1803 | bool_Bool have_copy = false0; |
| 1804 | struct eb_vma *ev; |
| 1805 | int err = 0; |
| 1806 | |
| 1807 | repeat: |
| 1808 | if (signal_pending(current)(((({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self ))); __ci;})->ci_curproc)->p_siglist | (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof (struct cpu_info, ci_self))); __ci;})->ci_curproc)->p_p ->ps_siglist) & ~(({struct cpu_info *__ci; asm volatile ("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc)->p_sigmask)) { |
| 1809 | err = -ERESTARTSYS4; |
| 1810 | goto out; |
| 1811 | } |
| 1812 | |
| 1813 | /* We may process another execbuffer during the unlock... */ |
| 1814 | eb_release_vmas(eb, false0); |
| 1815 | i915_gem_ww_ctx_fini(&eb->ww); |
| 1816 | |
| 1817 | /* |
| 1818 | * We take 3 passes through the slowpatch. |
| 1819 | * |
| 1820 | * 1 - we try to just prefault all the user relocation entries and |
| 1821 | * then attempt to reuse the atomic pagefault disabled fast path again. |
| 1822 | * |
| 1823 | * 2 - we copy the user entries to a local buffer here outside of the |
| 1824 | * local and allow ourselves to wait upon any rendering before |
| 1825 | * relocations |
| 1826 | * |
| 1827 | * 3 - we already have a local copy of the relocation entries, but |
| 1828 | * were interrupted (EAGAIN) whilst waiting for the objects, try again. |
| 1829 | */ |
| 1830 | if (!err) { |
| 1831 | err = eb_prefault_relocations(eb); |
| 1832 | } else if (!have_copy) { |
| 1833 | err = eb_copy_relocations(eb); |
| 1834 | have_copy = err == 0; |
| 1835 | } else { |
| 1836 | cond_resched()do { if (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self ))); __ci;})->ci_schedstate.spc_schedflags & 0x0002) yield (); } while (0); |
| 1837 | err = 0; |
| 1838 | } |
| 1839 | |
| 1840 | if (!err) |
| 1841 | err = eb_reinit_userptr(eb); |
| 1842 | |
| 1843 | i915_gem_ww_ctx_init(&eb->ww, true1); |
| 1844 | if (err) |
| 1845 | goto out; |
| 1846 | |
| 1847 | /* reacquire the objects */ |
| 1848 | repeat_validate: |
| 1849 | err = eb_pin_engine(eb, false0); |
| 1850 | if (err) |
| 1851 | goto err; |
| 1852 | |
| 1853 | err = eb_validate_vmas(eb); |
| 1854 | if (err) |
| 1855 | goto err; |
| 1856 | |
| 1857 | GEM_BUG_ON(!eb->batches[0])((void)0); |
| 1858 | |
| 1859 | list_for_each_entry(ev, &eb->relocs, reloc_link)for (ev = ({ const __typeof( ((__typeof(*ev) *)0)->reloc_link ) *__mptr = ((&eb->relocs)->next); (__typeof(*ev) * )( (char *)__mptr - __builtin_offsetof(__typeof(*ev), reloc_link ) );}); &ev->reloc_link != (&eb->relocs); ev = ( { const __typeof( ((__typeof(*ev) *)0)->reloc_link ) *__mptr = (ev->reloc_link.next); (__typeof(*ev) *)( (char *)__mptr - __builtin_offsetof(__typeof(*ev), reloc_link) );})) { |
| 1860 | if (!have_copy) { |
| 1861 | err = eb_relocate_vma(eb, ev); |
| 1862 | if (err) |
| 1863 | break; |
| 1864 | } else { |
| 1865 | err = eb_relocate_vma_slow(eb, ev); |
| 1866 | if (err) |
| 1867 | break; |
| 1868 | } |
| 1869 | } |
| 1870 | |
| 1871 | if (err == -EDEADLK11) |
| 1872 | goto err; |
| 1873 | |
| 1874 | if (err && !have_copy) |
| 1875 | goto repeat; |
| 1876 | |
| 1877 | if (err) |
| 1878 | goto err; |
| 1879 | |
| 1880 | /* as last step, parse the command buffer */ |
| 1881 | err = eb_parse(eb); |
| 1882 | if (err) |
| 1883 | goto err; |
| 1884 | |
| 1885 | /* |
| 1886 | * Leave the user relocations as are, this is the painfully slow path, |
| 1887 | * and we want to avoid the complication of dropping the lock whilst |
| 1888 | * having buffers reserved in the aperture and so causing spurious |
| 1889 | * ENOSPC for random operations. |
| 1890 | */ |
| 1891 | |
| 1892 | err: |
| 1893 | if (err == -EDEADLK11) { |
| 1894 | eb_release_vmas(eb, false0); |
| 1895 | err = i915_gem_ww_ctx_backoff(&eb->ww); |
| 1896 | if (!err) |
| 1897 | goto repeat_validate; |
| 1898 | } |
| 1899 | |
| 1900 | if (err == -EAGAIN35) |
| 1901 | goto repeat; |
| 1902 | |
| 1903 | out: |
| 1904 | if (have_copy) { |
| 1905 | const unsigned int count = eb->buffer_count; |
| 1906 | unsigned int i; |
| 1907 | |
| 1908 | for (i = 0; i < count; i++) { |
| 1909 | const struct drm_i915_gem_exec_object2 *entry = |
| 1910 | &eb->exec[i]; |
| 1911 | struct drm_i915_gem_relocation_entry *relocs; |
| 1912 | |
| 1913 | if (!entry->relocation_count) |
| 1914 | continue; |
| 1915 | |
| 1916 | relocs = u64_to_ptr(typeof(*relocs), entry->relocs_ptr)({ 1; (typeof(*relocs) *)(uintptr_t)(entry->relocs_ptr); } ); |
| 1917 | kvfree(relocs); |
| 1918 | } |
| 1919 | } |
| 1920 | |
| 1921 | return err; |
| 1922 | } |
| 1923 | |
| 1924 | static int eb_relocate_parse(struct i915_execbuffer *eb) |
| 1925 | { |
| 1926 | int err; |
| 1927 | bool_Bool throttle = true1; |
| 1928 | |
| 1929 | retry: |
| 1930 | err = eb_pin_engine(eb, throttle); |
| 1931 | if (err) { |
| 1932 | if (err != -EDEADLK11) |
| 1933 | return err; |
| 1934 | |
| 1935 | goto err; |
| 1936 | } |
| 1937 | |
| 1938 | /* only throttle once, even if we didn't need to throttle */ |
| 1939 | throttle = false0; |
| 1940 | |
| 1941 | err = eb_validate_vmas(eb); |
| 1942 | if (err == -EAGAIN35) |
| 1943 | goto slow; |
| 1944 | else if (err) |
| 1945 | goto err; |
| 1946 | |
| 1947 | /* The objects are in their final locations, apply the relocations. */ |
| 1948 | if (eb->args->flags & __EXEC_HAS_RELOC(1UL << (31))) { |
| 1949 | struct eb_vma *ev; |
| 1950 | |
| 1951 | list_for_each_entry(ev, &eb->relocs, reloc_link)for (ev = ({ const __typeof( ((__typeof(*ev) *)0)->reloc_link ) *__mptr = ((&eb->relocs)->next); (__typeof(*ev) * )( (char *)__mptr - __builtin_offsetof(__typeof(*ev), reloc_link ) );}); &ev->reloc_link != (&eb->relocs); ev = ( { const __typeof( ((__typeof(*ev) *)0)->reloc_link ) *__mptr = (ev->reloc_link.next); (__typeof(*ev) *)( (char *)__mptr - __builtin_offsetof(__typeof(*ev), reloc_link) );})) { |
| 1952 | err = eb_relocate_vma(eb, ev); |
| 1953 | if (err) |
| 1954 | break; |
| 1955 | } |
| 1956 | |
| 1957 | if (err == -EDEADLK11) |
| 1958 | goto err; |
| 1959 | else if (err) |
| 1960 | goto slow; |
| 1961 | } |
| 1962 | |
| 1963 | if (!err) |
| 1964 | err = eb_parse(eb); |
| 1965 | |
| 1966 | err: |
| 1967 | if (err == -EDEADLK11) { |
| 1968 | eb_release_vmas(eb, false0); |
| 1969 | err = i915_gem_ww_ctx_backoff(&eb->ww); |
| 1970 | if (!err) |
| 1971 | goto retry; |
| 1972 | } |
| 1973 | |
| 1974 | return err; |
| 1975 | |
| 1976 | slow: |
| 1977 | err = eb_relocate_parse_slow(eb); |
| 1978 | if (err) |
| 1979 | /* |
| 1980 | * If the user expects the execobject.offset and |
| 1981 | * reloc.presumed_offset to be an exact match, |
| 1982 | * as for using NO_RELOC, then we cannot update |
| 1983 | * the execobject.offset until we have completed |
| 1984 | * relocation. |
| 1985 | */ |
| 1986 | eb->args->flags &= ~__EXEC_HAS_RELOC(1UL << (31)); |
| 1987 | |
| 1988 | return err; |
| 1989 | } |
| 1990 | |
| 1991 | /* |
| 1992 | * Using two helper loops for the order of which requests / batches are created |
| 1993 | * and added the to backend. Requests are created in order from the parent to |
| 1994 | * the last child. Requests are added in the reverse order, from the last child |
| 1995 | * to parent. This is done for locking reasons as the timeline lock is acquired |
| 1996 | * during request creation and released when the request is added to the |
| 1997 | * backend. To make lockdep happy (see intel_context_timeline_lock) this must be |
| 1998 | * the ordering. |
| 1999 | */ |
| 2000 | #define for_each_batch_create_order(_eb, _i)for ((_i) = 0; (_i) < (_eb)->num_batches; ++(_i)) \ |
| 2001 | for ((_i) = 0; (_i) < (_eb)->num_batches; ++(_i)) |
| 2002 | #define for_each_batch_add_order(_eb, _i)extern char _ctassert[(!(!1)) ? 1 : -1 ] __attribute__((__unused__ )); for ((_i) = (_eb)->num_batches - 1; (_i) >= 0; --(_i )) \ |
| 2003 | BUILD_BUG_ON(!typecheck(int, _i))extern char _ctassert[(!(!1)) ? 1 : -1 ] __attribute__((__unused__ )); \ |
| 2004 | for ((_i) = (_eb)->num_batches - 1; (_i) >= 0; --(_i)) |
| 2005 | |
| 2006 | static struct i915_request * |
| 2007 | eb_find_first_request_added(struct i915_execbuffer *eb) |
| 2008 | { |
| 2009 | int i; |
| 2010 | |
| 2011 | for_each_batch_add_order(eb, i)extern char _ctassert[(!(!1)) ? 1 : -1 ] __attribute__((__unused__ )); for ((i) = (eb)->num_batches - 1; (i) >= 0; --(i)) |
| 2012 | if (eb->requests[i]) |
| 2013 | return eb->requests[i]; |
| 2014 | |
| 2015 | GEM_BUG_ON("Request not found")((void)0); |
| 2016 | |
| 2017 | return NULL((void *)0); |
| 2018 | } |
| 2019 | |
| 2020 | #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR)1 |
| 2021 | |
| 2022 | /* Stage with GFP_KERNEL allocations before we enter the signaling critical path */ |
| 2023 | static int eb_capture_stage(struct i915_execbuffer *eb) |
| 2024 | { |
| 2025 | const unsigned int count = eb->buffer_count; |
| 2026 | unsigned int i = count, j; |
| 2027 | |
| 2028 | while (i--) { |
| 2029 | struct eb_vma *ev = &eb->vma[i]; |
| 2030 | struct i915_vma *vma = ev->vma; |
| 2031 | unsigned int flags = ev->flags; |
| 2032 | |
| 2033 | if (!(flags & EXEC_OBJECT_CAPTURE(1<<7))) |
| 2034 | continue; |
| 2035 | |
| 2036 | if (i915_gem_context_is_recoverable(eb->gem_context) && |
| 2037 | (IS_DGFX(eb->i915)((&(eb->i915)->__info)->is_dgfx) || GRAPHICS_VER_FULL(eb->i915)(((&(eb->i915)->__runtime)->graphics.ip.ver) << 8 | ((&(eb->i915)->__runtime)->graphics.ip.rel) ) > IP_VER(12, 0)((12) << 8 | (0)))) |
| 2038 | return -EINVAL22; |
| 2039 | |
| 2040 | for_each_batch_create_order(eb, j)for ((j) = 0; (j) < (eb)->num_batches; ++(j)) { |
| 2041 | struct i915_capture_list *capture; |
| 2042 | |
| 2043 | capture = kmalloc(sizeof(*capture), GFP_KERNEL(0x0001 | 0x0004)); |
| 2044 | if (!capture) |
| 2045 | continue; |
| 2046 | |
| 2047 | capture->next = eb->capture_lists[j]; |
| 2048 | capture->vma_res = i915_vma_resource_get(vma->resource); |
| 2049 | eb->capture_lists[j] = capture; |
| 2050 | } |
| 2051 | } |
| 2052 | |
| 2053 | return 0; |
| 2054 | } |
| 2055 | |
| 2056 | /* Commit once we're in the critical path */ |
| 2057 | static void eb_capture_commit(struct i915_execbuffer *eb) |
| 2058 | { |
| 2059 | unsigned int j; |
| 2060 | |
| 2061 | for_each_batch_create_order(eb, j)for ((j) = 0; (j) < (eb)->num_batches; ++(j)) { |
| 2062 | struct i915_request *rq = eb->requests[j]; |
| 2063 | |
| 2064 | if (!rq) |
| 2065 | break; |
| 2066 | |
| 2067 | rq->capture_list = eb->capture_lists[j]; |
| 2068 | eb->capture_lists[j] = NULL((void *)0); |
| 2069 | } |
| 2070 | } |
| 2071 | |
| 2072 | /* |
| 2073 | * Release anything that didn't get committed due to errors. |
| 2074 | * The capture_list will otherwise be freed at request retire. |
| 2075 | */ |
| 2076 | static void eb_capture_release(struct i915_execbuffer *eb) |
| 2077 | { |
| 2078 | unsigned int j; |
| 2079 | |
| 2080 | for_each_batch_create_order(eb, j)for ((j) = 0; (j) < (eb)->num_batches; ++(j)) { |
| 2081 | if (eb->capture_lists[j]) { |
| 2082 | i915_request_free_capture_list(eb->capture_lists[j]); |
| 2083 | eb->capture_lists[j] = NULL((void *)0); |
| 2084 | } |
| 2085 | } |
| 2086 | } |
| 2087 | |
| 2088 | static void eb_capture_list_clear(struct i915_execbuffer *eb) |
| 2089 | { |
| 2090 | memset(eb->capture_lists, 0, sizeof(eb->capture_lists))__builtin_memset((eb->capture_lists), (0), (sizeof(eb-> capture_lists))); |
| 2091 | } |
| 2092 | |
| 2093 | #else |
| 2094 | |
| 2095 | static int eb_capture_stage(struct i915_execbuffer *eb) |
| 2096 | { |
| 2097 | return 0; |
| 2098 | } |
| 2099 | |
| 2100 | static void eb_capture_commit(struct i915_execbuffer *eb) |
| 2101 | { |
| 2102 | } |
| 2103 | |
| 2104 | static void eb_capture_release(struct i915_execbuffer *eb) |
| 2105 | { |
| 2106 | } |
| 2107 | |
| 2108 | static void eb_capture_list_clear(struct i915_execbuffer *eb) |
| 2109 | { |
| 2110 | } |
| 2111 | |
| 2112 | #endif |
| 2113 | |
| 2114 | static int eb_move_to_gpu(struct i915_execbuffer *eb) |
| 2115 | { |
| 2116 | const unsigned int count = eb->buffer_count; |
| 2117 | unsigned int i = count; |
| 2118 | int err = 0, j; |
| 2119 | |
| 2120 | while (i--) { |
| 2121 | struct eb_vma *ev = &eb->vma[i]; |
| 2122 | struct i915_vma *vma = ev->vma; |
| 2123 | unsigned int flags = ev->flags; |
| 2124 | struct drm_i915_gem_object *obj = vma->obj; |
| 2125 | |
| 2126 | assert_vma_held(vma)do { (void)(&((vma)->obj->base.resv)->lock.base) ; } while(0); |
| 2127 | |
| 2128 | /* |
| 2129 | * If the GPU is not _reading_ through the CPU cache, we need |
| 2130 | * to make sure that any writes (both previous GPU writes from |
| 2131 | * before a change in snooping levels and normal CPU writes) |
| 2132 | * caught in that cache are flushed to main memory. |
| 2133 | * |
| 2134 | * We want to say |
| 2135 | * obj->cache_dirty && |
| 2136 | * !(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_READ) |
| 2137 | * but gcc's optimiser doesn't handle that as well and emits |
| 2138 | * two jumps instead of one. Maybe one day... |
| 2139 | * |
| 2140 | * FIXME: There is also sync flushing in set_pages(), which |
| 2141 | * serves a different purpose(some of the time at least). |
| 2142 | * |
| 2143 | * We should consider: |
| 2144 | * |
| 2145 | * 1. Rip out the async flush code. |
| 2146 | * |
| 2147 | * 2. Or make the sync flushing use the async clflush path |
| 2148 | * using mandatory fences underneath. Currently the below |
| 2149 | * async flush happens after we bind the object. |
| 2150 | */ |
| 2151 | if (unlikely(obj->cache_dirty & ~obj->cache_coherent)__builtin_expect(!!(obj->cache_dirty & ~obj->cache_coherent ), 0)) { |
| 2152 | if (i915_gem_clflush_object(obj, 0)) |
| 2153 | flags &= ~EXEC_OBJECT_ASYNC(1<<6); |
| 2154 | } |
| 2155 | |
| 2156 | /* We only need to await on the first request */ |
| 2157 | if (err == 0 && !(flags & EXEC_OBJECT_ASYNC(1<<6))) { |
| 2158 | err = i915_request_await_object |
| 2159 | (eb_find_first_request_added(eb), obj, |
| 2160 | flags & EXEC_OBJECT_WRITE(1<<2)); |
| 2161 | } |
| 2162 | |
| 2163 | for_each_batch_add_order(eb, j)extern char _ctassert[(!(!1)) ? 1 : -1 ] __attribute__((__unused__ )); for ((j) = (eb)->num_batches - 1; (j) >= 0; --(j)) { |
| 2164 | if (err) |
| 2165 | break; |
| 2166 | if (!eb->requests[j]) |
| 2167 | continue; |
| 2168 | |
| 2169 | err = _i915_vma_move_to_active(vma, eb->requests[j], |
| 2170 | j ? NULL((void *)0) : |
| 2171 | eb->composite_fence ? |
| 2172 | eb->composite_fence : |
| 2173 | &eb->requests[j]->fence, |
| 2174 | flags | __EXEC_OBJECT_NO_RESERVE(1UL << (31))); |
| 2175 | } |
| 2176 | } |
| 2177 | |
| 2178 | #ifdef CONFIG_MMU_NOTIFIER |
| 2179 | if (!err && (eb->args->flags & __EXEC_USERPTR_USED(1UL << (29)))) { |
| 2180 | read_lock(&eb->i915->mm.notifier_lock)mtx_enter(&eb->i915->mm.notifier_lock); |
| 2181 | |
| 2182 | /* |
| 2183 | * count is always at least 1, otherwise __EXEC_USERPTR_USED |
| 2184 | * could not have been set |
| 2185 | */ |
| 2186 | for (i = 0; i < count; i++) { |
| 2187 | struct eb_vma *ev = &eb->vma[i]; |
| 2188 | struct drm_i915_gem_object *obj = ev->vma->obj; |
| 2189 | |
| 2190 | if (!i915_gem_object_is_userptr(obj)) |
| 2191 | continue; |
| 2192 | |
| 2193 | err = i915_gem_object_userptr_submit_done(obj); |
| 2194 | if (err) |
| 2195 | break; |
| 2196 | } |
| 2197 | |
| 2198 | read_unlock(&eb->i915->mm.notifier_lock)mtx_leave(&eb->i915->mm.notifier_lock); |
| 2199 | } |
| 2200 | #endif |
| 2201 | |
| 2202 | if (unlikely(err)__builtin_expect(!!(err), 0)) |
| 2203 | goto err_skip; |
| 2204 | |
| 2205 | /* Unconditionally flush any chipset caches (for streaming writes). */ |
| 2206 | intel_gt_chipset_flush(eb->gt); |
| 2207 | eb_capture_commit(eb); |
| 2208 | |
| 2209 | return 0; |
| 2210 | |
| 2211 | err_skip: |
| 2212 | for_each_batch_create_order(eb, j)for ((j) = 0; (j) < (eb)->num_batches; ++(j)) { |
| 2213 | if (!eb->requests[j]) |
| 2214 | break; |
| 2215 | |
| 2216 | i915_request_set_error_once(eb->requests[j], err); |
| 2217 | } |
| 2218 | return err; |
| 2219 | } |
| 2220 | |
| 2221 | static int i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec) |
| 2222 | { |
| 2223 | if (exec->flags & __I915_EXEC_ILLEGAL_FLAGS((-((1 << 21) << 1)) | (3<<6) | (1<<15 ))) |
| 2224 | return -EINVAL22; |
| 2225 | |
| 2226 | /* Kernel clipping was a DRI1 misfeature */ |
| 2227 | if (!(exec->flags & (I915_EXEC_FENCE_ARRAY(1<<19) | |
| 2228 | I915_EXEC_USE_EXTENSIONS(1 << 21)))) { |
| 2229 | if (exec->num_cliprects || exec->cliprects_ptr) |
| 2230 | return -EINVAL22; |
| 2231 | } |
| 2232 | |
| 2233 | if (exec->DR4 == 0xffffffff) { |
| 2234 | DRM_DEBUG("UXA submitting garbage DR4, fixing up\n")___drm_dbg(((void *)0), DRM_UT_CORE, "UXA submitting garbage DR4, fixing up\n" ); |
| 2235 | exec->DR4 = 0; |
| 2236 | } |
| 2237 | if (exec->DR1 || exec->DR4) |
| 2238 | return -EINVAL22; |
| 2239 | |
| 2240 | if ((exec->batch_start_offset | exec->batch_len) & 0x7) |
| 2241 | return -EINVAL22; |
| 2242 | |
| 2243 | return 0; |
| 2244 | } |
| 2245 | |
| 2246 | static int i915_reset_gen7_sol_offsets(struct i915_request *rq) |
| 2247 | { |
| 2248 | u32 *cs; |
| 2249 | int i; |
| 2250 | |
| 2251 | if (GRAPHICS_VER(rq->engine->i915)((&(rq->engine->i915)->__runtime)->graphics.ip .ver) != 7 || rq->engine->id != RCS0) { |
| 2252 | drm_dbg(&rq->engine->i915->drm, "sol reset is gen7/rcs only\n")__drm_dev_dbg(((void *)0), (&rq->engine->i915->drm ) ? (&rq->engine->i915->drm)->dev : ((void *) 0), DRM_UT_DRIVER, "sol reset is gen7/rcs only\n"); |
| 2253 | return -EINVAL22; |
| 2254 | } |
| 2255 | |
| 2256 | cs = intel_ring_begin(rq, 4 * 2 + 2); |
| 2257 | if (IS_ERR(cs)) |
| 2258 | return PTR_ERR(cs); |
| 2259 | |
| 2260 | *cs++ = MI_LOAD_REGISTER_IMM(4)(((0x0) << 29) | (0x22) << 23 | (2*(4)-1)); |
| 2261 | for (i = 0; i < 4; i++) { |
| 2262 | *cs++ = i915_mmio_reg_offset(GEN7_SO_WRITE_OFFSET(i)((const i915_reg_t){ .reg = (0x5280 + (i) * 4) })); |
| 2263 | *cs++ = 0; |
| 2264 | } |
| 2265 | *cs++ = MI_NOOP(((0x0) << 29) | (0) << 23 | (0)); |
| 2266 | intel_ring_advance(rq, cs); |
| 2267 | |
| 2268 | return 0; |
| 2269 | } |
| 2270 | |
| 2271 | static struct i915_vma * |
| 2272 | shadow_batch_pin(struct i915_execbuffer *eb, |
| 2273 | struct drm_i915_gem_object *obj, |
| 2274 | struct i915_address_space *vm, |
| 2275 | unsigned int flags) |
| 2276 | { |
| 2277 | struct i915_vma *vma; |
| 2278 | int err; |
| 2279 | |
| 2280 | vma = i915_vma_instance(obj, vm, NULL((void *)0)); |
| 2281 | if (IS_ERR(vma)) |
| 2282 | return vma; |
| 2283 | |
| 2284 | err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, flags | PIN_VALIDATE(1ULL << (8))); |
| 2285 | if (err) |
| 2286 | return ERR_PTR(err); |
| 2287 | |
| 2288 | return vma; |
| 2289 | } |
| 2290 | |
| 2291 | static struct i915_vma *eb_dispatch_secure(struct i915_execbuffer *eb, struct i915_vma *vma) |
| 2292 | { |
| 2293 | /* |
| 2294 | * snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure |
| 2295 | * batch" bit. Hence we need to pin secure batches into the global gtt. |
| 2296 | * hsw should have this fixed, but bdw mucks it up again. */ |
| 2297 | if (eb->batch_flags & I915_DISPATCH_SECURE(1UL << (0))) |
| 2298 | return i915_gem_object_ggtt_pin_ww(vma->obj, &eb->ww, NULL((void *)0), 0, 0, PIN_VALIDATE(1ULL << (8))); |
| 2299 | |
| 2300 | return NULL((void *)0); |
| 2301 | } |
| 2302 | |
| 2303 | static int eb_parse(struct i915_execbuffer *eb) |
| 2304 | { |
| 2305 | struct drm_i915_privateinteldrm_softc *i915 = eb->i915; |
| 2306 | struct intel_gt_buffer_pool_node *pool = eb->batch_pool; |
| 2307 | struct i915_vma *shadow, *trampoline, *batch; |
| 2308 | unsigned long len; |
| 2309 | int err; |
| 2310 | |
| 2311 | if (!eb_use_cmdparser(eb)) { |
| 2312 | batch = eb_dispatch_secure(eb, eb->batches[0]->vma); |
| 2313 | if (IS_ERR(batch)) |
| 2314 | return PTR_ERR(batch); |
| 2315 | |
| 2316 | goto secure_batch; |
| 2317 | } |
| 2318 | |
| 2319 | if (intel_context_is_parallel(eb->context)) |
| 2320 | return -EINVAL22; |
| 2321 | |
| 2322 | len = eb->batch_len[0]; |
| 2323 | if (!CMDPARSER_USES_GGTT(eb->i915)(((&(eb->i915)->__runtime)->graphics.ip.ver) == 7 )) { |
| 2324 | /* |
| 2325 | * ppGTT backed shadow buffers must be mapped RO, to prevent |
| 2326 | * post-scan tampering |
| 2327 | */ |
| 2328 | if (!eb->context->vm->has_read_only) { |
| 2329 | drm_dbg(&i915->drm,__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Cannot prevent post-scan tampering without RO capable vm\n" ) |
| 2330 | "Cannot prevent post-scan tampering without RO capable vm\n")__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Cannot prevent post-scan tampering without RO capable vm\n" ); |
| 2331 | return -EINVAL22; |
| 2332 | } |
| 2333 | } else { |
| 2334 | len += I915_CMD_PARSER_TRAMPOLINE_SIZE8; |
| 2335 | } |
| 2336 | if (unlikely(len < eb->batch_len[0])__builtin_expect(!!(len < eb->batch_len[0]), 0)) /* last paranoid check of overflow */ |
| 2337 | return -EINVAL22; |
| 2338 | |
| 2339 | if (!pool) { |
| 2340 | pool = intel_gt_get_buffer_pool(eb->gt, len, |
| 2341 | I915_MAP_WB); |
| 2342 | if (IS_ERR(pool)) |
| 2343 | return PTR_ERR(pool); |
| 2344 | eb->batch_pool = pool; |
| 2345 | } |
| 2346 | |
| 2347 | err = i915_gem_object_lock(pool->obj, &eb->ww); |
| 2348 | if (err) |
| 2349 | return err; |
| 2350 | |
| 2351 | shadow = shadow_batch_pin(eb, pool->obj, eb->context->vm, PIN_USER(1ULL << (11))); |
| 2352 | if (IS_ERR(shadow)) |
| 2353 | return PTR_ERR(shadow); |
| 2354 | |
| 2355 | intel_gt_buffer_pool_mark_used(pool); |
| 2356 | i915_gem_object_set_readonly(shadow->obj); |
| 2357 | shadow->private = pool; |
| 2358 | |
| 2359 | trampoline = NULL((void *)0); |
| 2360 | if (CMDPARSER_USES_GGTT(eb->i915)(((&(eb->i915)->__runtime)->graphics.ip.ver) == 7 )) { |
| 2361 | trampoline = shadow; |
| 2362 | |
| 2363 | shadow = shadow_batch_pin(eb, pool->obj, |
| 2364 | &eb->gt->ggtt->vm, |
| 2365 | PIN_GLOBAL(1ULL << (10))); |
| 2366 | if (IS_ERR(shadow)) |
| 2367 | return PTR_ERR(shadow); |
| 2368 | |
| 2369 | shadow->private = pool; |
| 2370 | |
| 2371 | eb->batch_flags |= I915_DISPATCH_SECURE(1UL << (0)); |
| 2372 | } |
| 2373 | |
| 2374 | batch = eb_dispatch_secure(eb, shadow); |
| 2375 | if (IS_ERR(batch)) |
| 2376 | return PTR_ERR(batch); |
| 2377 | |
| 2378 | err = dma_resv_reserve_fences(shadow->obj->base.resv, 1); |
| 2379 | if (err) |
| 2380 | return err; |
| 2381 | |
| 2382 | err = intel_engine_cmd_parser(eb->context->engine, |
| 2383 | eb->batches[0]->vma, |
| 2384 | eb->batch_start_offset, |
| 2385 | eb->batch_len[0], |
| 2386 | shadow, trampoline); |
| 2387 | if (err) |
| 2388 | return err; |
| 2389 | |
| 2390 | eb->batches[0] = &eb->vma[eb->buffer_count++]; |
| 2391 | eb->batches[0]->vma = i915_vma_get(shadow); |
| 2392 | eb->batches[0]->flags = __EXEC_OBJECT_HAS_PIN(1UL << (30)); |
| 2393 | |
| 2394 | eb->trampoline = trampoline; |
| 2395 | eb->batch_start_offset = 0; |
| 2396 | |
| 2397 | secure_batch: |
| 2398 | if (batch) { |
| 2399 | if (intel_context_is_parallel(eb->context)) |
| 2400 | return -EINVAL22; |
| 2401 | |
| 2402 | eb->batches[0] = &eb->vma[eb->buffer_count++]; |
| 2403 | eb->batches[0]->flags = __EXEC_OBJECT_HAS_PIN(1UL << (30)); |
| 2404 | eb->batches[0]->vma = i915_vma_get(batch); |
| 2405 | } |
| 2406 | return 0; |
| 2407 | } |
| 2408 | |
| 2409 | static int eb_request_submit(struct i915_execbuffer *eb, |
| 2410 | struct i915_request *rq, |
| 2411 | struct i915_vma *batch, |
| 2412 | u64 batch_len) |
| 2413 | { |
| 2414 | int err; |
| 2415 | |
| 2416 | if (intel_context_nopreempt(rq->context)) |
| 2417 | __set_bit(I915_FENCE_FLAG_NOPREEMPT, &rq->fence.flags); |
| 2418 | |
| 2419 | if (eb->args->flags & I915_EXEC_GEN7_SOL_RESET(1<<8)) { |
| 2420 | err = i915_reset_gen7_sol_offsets(rq); |
| 2421 | if (err) |
| 2422 | return err; |
| 2423 | } |
| 2424 | |
| 2425 | /* |
| 2426 | * After we completed waiting for other engines (using HW semaphores) |
| 2427 | * then we can signal that this request/batch is ready to run. This |
| 2428 | * allows us to determine if the batch is still waiting on the GPU |
| 2429 | * or actually running by checking the breadcrumb. |
| 2430 | */ |
| 2431 | if (rq->context->engine->emit_init_breadcrumb) { |
| 2432 | err = rq->context->engine->emit_init_breadcrumb(rq); |
| 2433 | if (err) |
| 2434 | return err; |
| 2435 | } |
| 2436 | |
| 2437 | err = rq->context->engine->emit_bb_start(rq, |
| 2438 | batch->node.start + |
| 2439 | eb->batch_start_offset, |
| 2440 | batch_len, |
| 2441 | eb->batch_flags); |
| 2442 | if (err) |
| 2443 | return err; |
| 2444 | |
| 2445 | if (eb->trampoline) { |
| 2446 | GEM_BUG_ON(intel_context_is_parallel(rq->context))((void)0); |
| 2447 | GEM_BUG_ON(eb->batch_start_offset)((void)0); |
| 2448 | err = rq->context->engine->emit_bb_start(rq, |
| 2449 | eb->trampoline->node.start + |
| 2450 | batch_len, 0, 0); |
| 2451 | if (err) |
| 2452 | return err; |
| 2453 | } |
| 2454 | |
| 2455 | return 0; |
| 2456 | } |
| 2457 | |
| 2458 | static int eb_submit(struct i915_execbuffer *eb) |
| 2459 | { |
| 2460 | unsigned int i; |
| 2461 | int err; |
| 2462 | |
| 2463 | err = eb_move_to_gpu(eb); |
| 2464 | |
| 2465 | for_each_batch_create_order(eb, i)for ((i) = 0; (i) < (eb)->num_batches; ++(i)) { |
| 2466 | if (!eb->requests[i]) |
| 2467 | break; |
| 2468 | |
| 2469 | trace_i915_request_queue(eb->requests[i], eb->batch_flags); |
| 2470 | if (!err) |
| 2471 | err = eb_request_submit(eb, eb->requests[i], |
| 2472 | eb->batches[i]->vma, |
| 2473 | eb->batch_len[i]); |
| 2474 | } |
| 2475 | |
| 2476 | return err; |
| 2477 | } |
| 2478 | |
| 2479 | static int num_vcs_engines(struct drm_i915_privateinteldrm_softc *i915) |
| 2480 | { |
| 2481 | return hweight_long(VDBOX_MASK(to_gt(i915))({ unsigned int first__ = (VCS0); unsigned int count__ = (8); ((to_gt(i915))->info.engine_mask & (((~0UL) >> ( 64 - (first__ + count__ - 1) - 1)) & ((~0UL) << (first__ )))) >> first__; })); |
| 2482 | } |
| 2483 | |
| 2484 | /* |
| 2485 | * Find one BSD ring to dispatch the corresponding BSD command. |
| 2486 | * The engine index is returned. |
| 2487 | */ |
| 2488 | static unsigned int |
| 2489 | gen8_dispatch_bsd_engine(struct drm_i915_privateinteldrm_softc *dev_priv, |
| 2490 | struct drm_file *file) |
| 2491 | { |
| 2492 | struct drm_i915_file_private *file_priv = file->driver_priv; |
| 2493 | |
| 2494 | /* Check whether the file_priv has already selected one ring. */ |
| 2495 | if ((int)file_priv->bsd_engine < 0) |
| 2496 | file_priv->bsd_engine = |
| 2497 | prandom_u32_max(num_vcs_engines(dev_priv)); |
| 2498 | |
| 2499 | return file_priv->bsd_engine; |
| 2500 | } |
| 2501 | |
| 2502 | static const enum intel_engine_id user_ring_map[] = { |
| 2503 | [I915_EXEC_DEFAULT(0<<0)] = RCS0, |
| 2504 | [I915_EXEC_RENDER(1<<0)] = RCS0, |
| 2505 | [I915_EXEC_BLT(3<<0)] = BCS0, |
| 2506 | [I915_EXEC_BSD(2<<0)] = VCS0, |
| 2507 | [I915_EXEC_VEBOX(4<<0)] = VECS0 |
| 2508 | }; |
| 2509 | |
| 2510 | static struct i915_request *eb_throttle(struct i915_execbuffer *eb, struct intel_context *ce) |
| 2511 | { |
| 2512 | struct intel_ring *ring = ce->ring; |
| 2513 | struct intel_timeline *tl = ce->timeline; |
| 2514 | struct i915_request *rq; |
| 2515 | |
| 2516 | /* |
| 2517 | * Completely unscientific finger-in-the-air estimates for suitable |
| 2518 | * maximum user request size (to avoid blocking) and then backoff. |
| 2519 | */ |
| 2520 | if (intel_ring_update_space(ring) >= PAGE_SIZE(1 << 12)) |
| 2521 | return NULL((void *)0); |
| 2522 | |
| 2523 | /* |
| 2524 | * Find a request that after waiting upon, there will be at least half |
| 2525 | * the ring available. The hysteresis allows us to compete for the |
| 2526 | * shared ring and should mean that we sleep less often prior to |
| 2527 | * claiming our resources, but not so long that the ring completely |
| 2528 | * drains before we can submit our next request. |
| 2529 | */ |
| 2530 | list_for_each_entry(rq, &tl->requests, link)for (rq = ({ const __typeof( ((__typeof(*rq) *)0)->link ) * __mptr = ((&tl->requests)->next); (__typeof(*rq) *) ( (char *)__mptr - __builtin_offsetof(__typeof(*rq), link) ); }); &rq->link != (&tl->requests); rq = ({ const __typeof( ((__typeof(*rq) *)0)->link ) *__mptr = (rq-> link.next); (__typeof(*rq) *)( (char *)__mptr - __builtin_offsetof (__typeof(*rq), link) );})) { |
| 2531 | if (rq->ring != ring) |
| 2532 | continue; |
| 2533 | |
| 2534 | if (__intel_ring_space(rq->postfix, |
| 2535 | ring->emit, ring->size) > ring->size / 2) |
| 2536 | break; |
| 2537 | } |
| 2538 | if (&rq->link == &tl->requests) |
| 2539 | return NULL((void *)0); /* weird, we will check again later for real */ |
| 2540 | |
| 2541 | return i915_request_get(rq); |
| 2542 | } |
| 2543 | |
| 2544 | static int eb_pin_timeline(struct i915_execbuffer *eb, struct intel_context *ce, |
| 2545 | bool_Bool throttle) |
| 2546 | { |
| 2547 | struct intel_timeline *tl; |
| 2548 | struct i915_request *rq = NULL((void *)0); |
| 2549 | |
| 2550 | /* |
| 2551 | * Take a local wakeref for preparing to dispatch the execbuf as |
| 2552 | * we expect to access the hardware fairly frequently in the |
| 2553 | * process, and require the engine to be kept awake between accesses. |
| 2554 | * Upon dispatch, we acquire another prolonged wakeref that we hold |
| 2555 | * until the timeline is idle, which in turn releases the wakeref |
| 2556 | * taken on the engine, and the parent device. |
| 2557 | */ |
| 2558 | tl = intel_context_timeline_lock(ce); |
| 2559 | if (IS_ERR(tl)) |
| 2560 | return PTR_ERR(tl); |
| 2561 | |
| 2562 | intel_context_enter(ce); |
| 2563 | if (throttle) |
| 2564 | rq = eb_throttle(eb, ce); |
| 2565 | intel_context_timeline_unlock(tl); |
| 2566 | |
| 2567 | if (rq) { |
| 2568 | #ifdef __linux__ |
| 2569 | bool_Bool nonblock = eb->file->filp->f_flags & O_NONBLOCK0x0004; |
| 2570 | #else |
| 2571 | bool_Bool nonblock = eb->file->filp->f_flag & FNONBLOCK0x0004; |
| 2572 | #endif |
| 2573 | long timeout = nonblock ? 0 : MAX_SCHEDULE_TIMEOUT(0x7fffffff); |
| 2574 | |
| 2575 | if (i915_request_wait(rq, I915_WAIT_INTERRUPTIBLE(1UL << (0)), |
| 2576 | timeout) < 0) { |
| 2577 | i915_request_put(rq); |
| 2578 | |
| 2579 | /* |
| 2580 | * Error path, cannot use intel_context_timeline_lock as |
| 2581 | * that is user interruptable and this clean up step |
| 2582 | * must be done. |
| 2583 | */ |
| 2584 | mutex_lock(&ce->timeline->mutex)rw_enter_write(&ce->timeline->mutex); |
| 2585 | intel_context_exit(ce); |
| 2586 | mutex_unlock(&ce->timeline->mutex)rw_exit_write(&ce->timeline->mutex); |
| 2587 | |
| 2588 | if (nonblock) |
| 2589 | return -EWOULDBLOCK35; |
| 2590 | else |
| 2591 | return -EINTR4; |
| 2592 | } |
| 2593 | i915_request_put(rq); |
| 2594 | } |
| 2595 | |
| 2596 | return 0; |
| 2597 | } |
| 2598 | |
| 2599 | static int eb_pin_engine(struct i915_execbuffer *eb, bool_Bool throttle) |
| 2600 | { |
| 2601 | struct intel_context *ce = eb->context, *child; |
| 2602 | int err; |
| 2603 | int i = 0, j = 0; |
| 2604 | |
| 2605 | GEM_BUG_ON(eb->args->flags & __EXEC_ENGINE_PINNED)((void)0); |
| 2606 | |
| 2607 | if (unlikely(intel_context_is_banned(ce))__builtin_expect(!!(intel_context_is_banned(ce)), 0)) |
| 2608 | return -EIO5; |
| 2609 | |
| 2610 | /* |
| 2611 | * Pinning the contexts may generate requests in order to acquire |
| 2612 | * GGTT space, so do this first before we reserve a seqno for |
| 2613 | * ourselves. |
| 2614 | */ |
| 2615 | err = intel_context_pin_ww(ce, &eb->ww); |
| 2616 | if (err) |
| 2617 | return err; |
| 2618 | for_each_child(ce, child)for (child = ({ const __typeof( ((__typeof(*child) *)0)->parallel .child_link ) *__mptr = ((&(ce)->parallel.child_list)-> next); (__typeof(*child) *)( (char *)__mptr - __builtin_offsetof (__typeof(*child), parallel.child_link) );}); &child-> parallel.child_link != (&(ce)->parallel.child_list); child = ({ const __typeof( ((__typeof(*child) *)0)->parallel.child_link ) *__mptr = (child->parallel.child_link.next); (__typeof( *child) *)( (char *)__mptr - __builtin_offsetof(__typeof(*child ), parallel.child_link) );})) { |
| 2619 | err = intel_context_pin_ww(child, &eb->ww); |
| 2620 | GEM_BUG_ON(err)((void)0); /* perma-pinned should incr a counter */ |
| 2621 | } |
| 2622 | |
| 2623 | for_each_child(ce, child)for (child = ({ const __typeof( ((__typeof(*child) *)0)->parallel .child_link ) *__mptr = ((&(ce)->parallel.child_list)-> next); (__typeof(*child) *)( (char *)__mptr - __builtin_offsetof (__typeof(*child), parallel.child_link) );}); &child-> parallel.child_link != (&(ce)->parallel.child_list); child = ({ const __typeof( ((__typeof(*child) *)0)->parallel.child_link ) *__mptr = (child->parallel.child_link.next); (__typeof( *child) *)( (char *)__mptr - __builtin_offsetof(__typeof(*child ), parallel.child_link) );})) { |
| 2624 | err = eb_pin_timeline(eb, child, throttle); |
| 2625 | if (err) |
| 2626 | goto unwind; |
| 2627 | ++i; |
| 2628 | } |
| 2629 | err = eb_pin_timeline(eb, ce, throttle); |
| 2630 | if (err) |
| 2631 | goto unwind; |
| 2632 | |
| 2633 | eb->args->flags |= __EXEC_ENGINE_PINNED(1UL << (30)); |
| 2634 | return 0; |
| 2635 | |
| 2636 | unwind: |
| 2637 | for_each_child(ce, child)for (child = ({ const __typeof( ((__typeof(*child) *)0)->parallel .child_link ) *__mptr = ((&(ce)->parallel.child_list)-> next); (__typeof(*child) *)( (char *)__mptr - __builtin_offsetof (__typeof(*child), parallel.child_link) );}); &child-> parallel.child_link != (&(ce)->parallel.child_list); child = ({ const __typeof( ((__typeof(*child) *)0)->parallel.child_link ) *__mptr = (child->parallel.child_link.next); (__typeof( *child) *)( (char *)__mptr - __builtin_offsetof(__typeof(*child ), parallel.child_link) );})) { |
| 2638 | if (j++ < i) { |
| 2639 | mutex_lock(&child->timeline->mutex)rw_enter_write(&child->timeline->mutex); |
| 2640 | intel_context_exit(child); |
| 2641 | mutex_unlock(&child->timeline->mutex)rw_exit_write(&child->timeline->mutex); |
| 2642 | } |
| 2643 | } |
| 2644 | for_each_child(ce, child)for (child = ({ const __typeof( ((__typeof(*child) *)0)->parallel .child_link ) *__mptr = ((&(ce)->parallel.child_list)-> next); (__typeof(*child) *)( (char *)__mptr - __builtin_offsetof (__typeof(*child), parallel.child_link) );}); &child-> parallel.child_link != (&(ce)->parallel.child_list); child = ({ const __typeof( ((__typeof(*child) *)0)->parallel.child_link ) *__mptr = (child->parallel.child_link.next); (__typeof( *child) *)( (char *)__mptr - __builtin_offsetof(__typeof(*child ), parallel.child_link) );})) |
| 2645 | intel_context_unpin(child); |
| 2646 | intel_context_unpin(ce); |
| 2647 | return err; |
| 2648 | } |
| 2649 | |
| 2650 | static void eb_unpin_engine(struct i915_execbuffer *eb) |
| 2651 | { |
| 2652 | struct intel_context *ce = eb->context, *child; |
| 2653 | |
| 2654 | if (!(eb->args->flags & __EXEC_ENGINE_PINNED(1UL << (30)))) |
| 2655 | return; |
| 2656 | |
| 2657 | eb->args->flags &= ~__EXEC_ENGINE_PINNED(1UL << (30)); |
| 2658 | |
| 2659 | for_each_child(ce, child)for (child = ({ const __typeof( ((__typeof(*child) *)0)->parallel .child_link ) *__mptr = ((&(ce)->parallel.child_list)-> next); (__typeof(*child) *)( (char *)__mptr - __builtin_offsetof (__typeof(*child), parallel.child_link) );}); &child-> parallel.child_link != (&(ce)->parallel.child_list); child = ({ const __typeof( ((__typeof(*child) *)0)->parallel.child_link ) *__mptr = (child->parallel.child_link.next); (__typeof( *child) *)( (char *)__mptr - __builtin_offsetof(__typeof(*child ), parallel.child_link) );})) { |
| 2660 | mutex_lock(&child->timeline->mutex)rw_enter_write(&child->timeline->mutex); |
| 2661 | intel_context_exit(child); |
| 2662 | mutex_unlock(&child->timeline->mutex)rw_exit_write(&child->timeline->mutex); |
| 2663 | |
| 2664 | intel_context_unpin(child); |
| 2665 | } |
| 2666 | |
| 2667 | mutex_lock(&ce->timeline->mutex)rw_enter_write(&ce->timeline->mutex); |
| 2668 | intel_context_exit(ce); |
| 2669 | mutex_unlock(&ce->timeline->mutex)rw_exit_write(&ce->timeline->mutex); |
| 2670 | |
| 2671 | intel_context_unpin(ce); |
| 2672 | } |
| 2673 | |
| 2674 | static unsigned int |
| 2675 | eb_select_legacy_ring(struct i915_execbuffer *eb) |
| 2676 | { |
| 2677 | struct drm_i915_privateinteldrm_softc *i915 = eb->i915; |
| 2678 | struct drm_i915_gem_execbuffer2 *args = eb->args; |
| 2679 | unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK(0x3f); |
| 2680 | |
| 2681 | if (user_ring_id != I915_EXEC_BSD(2<<0) && |
| 2682 | (args->flags & I915_EXEC_BSD_MASK(3 << (13)))) { |
| 2683 | drm_dbg(&i915->drm,__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "execbuf with non bsd ring but with invalid " "bsd dispatch flags: %d\n", (int)(args->flags)) |
| 2684 | "execbuf with non bsd ring but with invalid "__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "execbuf with non bsd ring but with invalid " "bsd dispatch flags: %d\n", (int)(args->flags)) |
| 2685 | "bsd dispatch flags: %d\n", (int)(args->flags))__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "execbuf with non bsd ring but with invalid " "bsd dispatch flags: %d\n", (int)(args->flags)); |
| 2686 | return -1; |
| 2687 | } |
| 2688 | |
| 2689 | if (user_ring_id == I915_EXEC_BSD(2<<0) && num_vcs_engines(i915) > 1) { |
| 2690 | unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK(3 << (13)); |
| 2691 | |
| 2692 | if (bsd_idx == I915_EXEC_BSD_DEFAULT(0 << (13))) { |
| 2693 | bsd_idx = gen8_dispatch_bsd_engine(i915, eb->file); |
| 2694 | } else if (bsd_idx >= I915_EXEC_BSD_RING1(1 << (13)) && |
| 2695 | bsd_idx <= I915_EXEC_BSD_RING2(2 << (13))) { |
| 2696 | bsd_idx >>= I915_EXEC_BSD_SHIFT(13); |
| 2697 | bsd_idx--; |
| 2698 | } else { |
| 2699 | drm_dbg(&i915->drm,__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "execbuf with unknown bsd ring: %u\n" , bsd_idx) |
| 2700 | "execbuf with unknown bsd ring: %u\n",__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "execbuf with unknown bsd ring: %u\n" , bsd_idx) |
| 2701 | bsd_idx)__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "execbuf with unknown bsd ring: %u\n" , bsd_idx); |
| 2702 | return -1; |
| 2703 | } |
| 2704 | |
| 2705 | return _VCS(bsd_idx)(VCS0 + (bsd_idx)); |
| 2706 | } |
| 2707 | |
| 2708 | if (user_ring_id >= ARRAY_SIZE(user_ring_map)(sizeof((user_ring_map)) / sizeof((user_ring_map)[0]))) { |
| 2709 | drm_dbg(&i915->drm, "execbuf with unknown ring: %u\n",__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "execbuf with unknown ring: %u\n" , user_ring_id) |
| 2710 | user_ring_id)__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "execbuf with unknown ring: %u\n" , user_ring_id); |
| 2711 | return -1; |
| 2712 | } |
| 2713 | |
| 2714 | return user_ring_map[user_ring_id]; |
| 2715 | } |
| 2716 | |
| 2717 | static int |
| 2718 | eb_select_engine(struct i915_execbuffer *eb) |
| 2719 | { |
| 2720 | struct intel_context *ce, *child; |
| 2721 | unsigned int idx; |
| 2722 | int err; |
| 2723 | |
| 2724 | if (i915_gem_context_user_engines(eb->gem_context)) |
| 2725 | idx = eb->args->flags & I915_EXEC_RING_MASK(0x3f); |
| 2726 | else |
| 2727 | idx = eb_select_legacy_ring(eb); |
| 2728 | |
| 2729 | ce = i915_gem_context_get_engine(eb->gem_context, idx); |
| 2730 | if (IS_ERR(ce)) |
| 2731 | return PTR_ERR(ce); |
| 2732 | |
| 2733 | if (intel_context_is_parallel(ce)) { |
| 2734 | if (eb->buffer_count < ce->parallel.number_children + 1) { |
| 2735 | intel_context_put(ce); |
| 2736 | return -EINVAL22; |
| 2737 | } |
| 2738 | if (eb->batch_start_offset || eb->args->batch_len) { |
| 2739 | intel_context_put(ce); |
| 2740 | return -EINVAL22; |
| 2741 | } |
| 2742 | } |
| 2743 | eb->num_batches = ce->parallel.number_children + 1; |
| 2744 | |
| 2745 | for_each_child(ce, child)for (child = ({ const __typeof( ((__typeof(*child) *)0)->parallel .child_link ) *__mptr = ((&(ce)->parallel.child_list)-> next); (__typeof(*child) *)( (char *)__mptr - __builtin_offsetof (__typeof(*child), parallel.child_link) );}); &child-> parallel.child_link != (&(ce)->parallel.child_list); child = ({ const __typeof( ((__typeof(*child) *)0)->parallel.child_link ) *__mptr = (child->parallel.child_link.next); (__typeof( *child) *)( (char *)__mptr - __builtin_offsetof(__typeof(*child ), parallel.child_link) );})) |
| 2746 | intel_context_get(child); |
| 2747 | intel_gt_pm_get(ce->engine->gt); |
| 2748 | |
| 2749 | if (!test_bit(CONTEXT_ALLOC_BIT1, &ce->flags)) { |
| 2750 | err = intel_context_alloc_state(ce); |
| 2751 | if (err) |
| 2752 | goto err; |
| 2753 | } |
| 2754 | for_each_child(ce, child)for (child = ({ const __typeof( ((__typeof(*child) *)0)->parallel .child_link ) *__mptr = ((&(ce)->parallel.child_list)-> next); (__typeof(*child) *)( (char *)__mptr - __builtin_offsetof (__typeof(*child), parallel.child_link) );}); &child-> parallel.child_link != (&(ce)->parallel.child_list); child = ({ const __typeof( ((__typeof(*child) *)0)->parallel.child_link ) *__mptr = (child->parallel.child_link.next); (__typeof( *child) *)( (char *)__mptr - __builtin_offsetof(__typeof(*child ), parallel.child_link) );})) { |
| 2755 | if (!test_bit(CONTEXT_ALLOC_BIT1, &child->flags)) { |
| 2756 | err = intel_context_alloc_state(child); |
| 2757 | if (err) |
| 2758 | goto err; |
| 2759 | } |
| 2760 | } |
| 2761 | |
| 2762 | /* |
| 2763 | * ABI: Before userspace accesses the GPU (e.g. execbuffer), report |
| 2764 | * EIO if the GPU is already wedged. |
| 2765 | */ |
| 2766 | err = intel_gt_terminally_wedged(ce->engine->gt); |
| 2767 | if (err) |
| 2768 | goto err; |
| 2769 | |
| 2770 | if (!i915_vm_tryget(ce->vm)) { |
| 2771 | err = -ENOENT2; |
| 2772 | goto err; |
| 2773 | } |
| 2774 | |
| 2775 | eb->context = ce; |
| 2776 | eb->gt = ce->engine->gt; |
| 2777 | |
| 2778 | /* |
| 2779 | * Make sure engine pool stays alive even if we call intel_context_put |
| 2780 | * during ww handling. The pool is destroyed when last pm reference |
| 2781 | * is dropped, which breaks our -EDEADLK handling. |
| 2782 | */ |
| 2783 | return err; |
| 2784 | |
| 2785 | err: |
| 2786 | intel_gt_pm_put(ce->engine->gt); |
| 2787 | for_each_child(ce, child)for (child = ({ const __typeof( ((__typeof(*child) *)0)->parallel .child_link ) *__mptr = ((&(ce)->parallel.child_list)-> next); (__typeof(*child) *)( (char *)__mptr - __builtin_offsetof (__typeof(*child), parallel.child_link) );}); &child-> parallel.child_link != (&(ce)->parallel.child_list); child = ({ const __typeof( ((__typeof(*child) *)0)->parallel.child_link ) *__mptr = (child->parallel.child_link.next); (__typeof( *child) *)( (char *)__mptr - __builtin_offsetof(__typeof(*child ), parallel.child_link) );})) |
| 2788 | intel_context_put(child); |
| 2789 | intel_context_put(ce); |
| 2790 | return err; |
| 2791 | } |
| 2792 | |
| 2793 | static void |
| 2794 | eb_put_engine(struct i915_execbuffer *eb) |
| 2795 | { |
| 2796 | struct intel_context *child; |
| 2797 | |
| 2798 | i915_vm_put(eb->context->vm); |
| 2799 | intel_gt_pm_put(eb->gt); |
| 2800 | for_each_child(eb->context, child)for (child = ({ const __typeof( ((__typeof(*child) *)0)->parallel .child_link ) *__mptr = ((&(eb->context)->parallel. child_list)->next); (__typeof(*child) *)( (char *)__mptr - __builtin_offsetof(__typeof(*child), parallel.child_link) ); }); &child->parallel.child_link != (&(eb->context )->parallel.child_list); child = ({ const __typeof( ((__typeof (*child) *)0)->parallel.child_link ) *__mptr = (child-> parallel.child_link.next); (__typeof(*child) *)( (char *)__mptr - __builtin_offsetof(__typeof(*child), parallel.child_link) ) ;})) |
| 2801 | intel_context_put(child); |
| 2802 | intel_context_put(eb->context); |
| 2803 | } |
| 2804 | |
| 2805 | static void |
| 2806 | __free_fence_array(struct eb_fence *fences, unsigned int n) |
| 2807 | { |
| 2808 | while (n--) { |
| 2809 | drm_syncobj_put(ptr_mask_bits(fences[n].syncobj, 2)({ unsigned long __v = (unsigned long)(fences[n].syncobj); (typeof (fences[n].syncobj))(__v & -(1UL << (2))); })); |
| 2810 | dma_fence_put(fences[n].dma_fence); |
| 2811 | dma_fence_chain_free(fences[n].chain_fence); |
| 2812 | } |
| 2813 | kvfree(fences); |
| 2814 | } |
| 2815 | |
| 2816 | static int |
| 2817 | add_timeline_fence_array(struct i915_execbuffer *eb, |
| 2818 | const struct drm_i915_gem_execbuffer_ext_timeline_fences *timeline_fences) |
| 2819 | { |
| 2820 | struct drm_i915_gem_exec_fence __user *user_fences; |
| 2821 | u64 __user *user_values; |
| 2822 | struct eb_fence *f; |
| 2823 | u64 nfences; |
| 2824 | int err = 0; |
| 2825 | |
| 2826 | nfences = timeline_fences->fence_count; |
| 2827 | if (!nfences) |
| 2828 | return 0; |
| 2829 | |
| 2830 | /* Check multiplication overflow for access_ok() and kvmalloc_array() */ |
| 2831 | BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long))extern char _ctassert[(!(sizeof(size_t) > sizeof(unsigned long ))) ? 1 : -1 ] __attribute__((__unused__)); |
| 2832 | if (nfences > min_t(unsigned long,({ unsigned long __min_a = (0xffffffffffffffffUL / sizeof(*user_fences )); unsigned long __min_b = (0xffffffffffffffffUL / sizeof(*f )); __min_a < __min_b ? __min_a : __min_b; }) |
| 2833 | ULONG_MAX / sizeof(*user_fences),({ unsigned long __min_a = (0xffffffffffffffffUL / sizeof(*user_fences )); unsigned long __min_b = (0xffffffffffffffffUL / sizeof(*f )); __min_a < __min_b ? __min_a : __min_b; }) |
| 2834 | SIZE_MAX / sizeof(*f))({ unsigned long __min_a = (0xffffffffffffffffUL / sizeof(*user_fences )); unsigned long __min_b = (0xffffffffffffffffUL / sizeof(*f )); __min_a < __min_b ? __min_a : __min_b; }) - eb->num_fences) |
| 2835 | return -EINVAL22; |
| 2836 | |
| 2837 | user_fences = u64_to_user_ptr(timeline_fences->handles_ptr)((void *)(uintptr_t)(timeline_fences->handles_ptr)); |
| 2838 | if (!access_ok(user_fences, nfences * sizeof(*user_fences))) |
| 2839 | return -EFAULT14; |
| 2840 | |
| 2841 | user_values = u64_to_user_ptr(timeline_fences->values_ptr)((void *)(uintptr_t)(timeline_fences->values_ptr)); |
| 2842 | if (!access_ok(user_values, nfences * sizeof(*user_values))) |
| 2843 | return -EFAULT14; |
| 2844 | |
| 2845 | #ifdef __linux__ |
| 2846 | f = krealloc(eb->fences, |
| 2847 | (eb->num_fences + nfences) * sizeof(*f), |
| 2848 | __GFP_NOWARN0 | GFP_KERNEL(0x0001 | 0x0004)); |
| 2849 | if (!f) |
| 2850 | return -ENOMEM12; |
| 2851 | #else |
| 2852 | f = kmalloc((eb->num_fences + nfences) * sizeof(*f), |
| 2853 | __GFP_NOWARN0 | GFP_KERNEL(0x0001 | 0x0004)); |
| 2854 | if (!f) |
| 2855 | return -ENOMEM12; |
| 2856 | memcpy(f, eb->fences, eb->num_fences * sizeof(*f))__builtin_memcpy((f), (eb->fences), (eb->num_fences * sizeof (*f))); |
| 2857 | kfree(eb->fences); |
| 2858 | #endif |
| 2859 | |
| 2860 | eb->fences = f; |
| 2861 | f += eb->num_fences; |
| 2862 | |
| 2863 | #ifdef notyet |
| 2864 | BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) &extern char _ctassert[(!(~(64 - 1) & ~(-((1<<1) << 1)))) ? 1 : -1 ] __attribute__((__unused__)) |
| 2865 | ~__I915_EXEC_FENCE_UNKNOWN_FLAGS)extern char _ctassert[(!(~(64 - 1) & ~(-((1<<1) << 1)))) ? 1 : -1 ] __attribute__((__unused__)); |
| 2866 | #endif |
| 2867 | |
| 2868 | while (nfences--) { |
| 2869 | struct drm_i915_gem_exec_fence user_fence; |
| 2870 | struct drm_syncobj *syncobj; |
| 2871 | struct dma_fence *fence = NULL((void *)0); |
| 2872 | u64 point; |
| 2873 | |
| 2874 | if (__copy_from_user(&user_fence, |
| 2875 | user_fences++, |
| 2876 | sizeof(user_fence))) |
| 2877 | return -EFAULT14; |
| 2878 | |
| 2879 | if (user_fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS(-((1<<1) << 1))) |
| 2880 | return -EINVAL22; |
| 2881 | |
| 2882 | if (__get_user(point, user_values++)-copyin((user_values++), &((point)), sizeof((point)))) |
| 2883 | return -EFAULT14; |
| 2884 | |
| 2885 | syncobj = drm_syncobj_find(eb->file, user_fence.handle); |
| 2886 | if (!syncobj) { |
| 2887 | DRM_DEBUG("Invalid syncobj handle provided\n")___drm_dbg(((void *)0), DRM_UT_CORE, "Invalid syncobj handle provided\n" ); |
| 2888 | return -ENOENT2; |
| 2889 | } |
| 2890 | |
| 2891 | fence = drm_syncobj_fence_get(syncobj); |
| 2892 | |
| 2893 | if (!fence && user_fence.flags && |
| 2894 | !(user_fence.flags & I915_EXEC_FENCE_SIGNAL(1<<1))) { |
| 2895 | DRM_DEBUG("Syncobj handle has no fence\n")___drm_dbg(((void *)0), DRM_UT_CORE, "Syncobj handle has no fence\n" ); |
| 2896 | drm_syncobj_put(syncobj); |
| 2897 | return -EINVAL22; |
| 2898 | } |
| 2899 | |
| 2900 | if (fence) |
| 2901 | err = dma_fence_chain_find_seqno(&fence, point); |
| 2902 | |
| 2903 | if (err && !(user_fence.flags & I915_EXEC_FENCE_SIGNAL(1<<1))) { |
| 2904 | DRM_DEBUG("Syncobj handle missing requested point %llu\n", point)___drm_dbg(((void *)0), DRM_UT_CORE, "Syncobj handle missing requested point %llu\n" , point); |
| 2905 | dma_fence_put(fence); |
| 2906 | drm_syncobj_put(syncobj); |
| 2907 | return err; |
| 2908 | } |
| 2909 | |
| 2910 | /* |
| 2911 | * A point might have been signaled already and |
| 2912 | * garbage collected from the timeline. In this case |
| 2913 | * just ignore the point and carry on. |
| 2914 | */ |
| 2915 | if (!fence && !(user_fence.flags & I915_EXEC_FENCE_SIGNAL(1<<1))) { |
| 2916 | drm_syncobj_put(syncobj); |
| 2917 | continue; |
| 2918 | } |
| 2919 | |
| 2920 | /* |
| 2921 | * For timeline syncobjs we need to preallocate chains for |
| 2922 | * later signaling. |
| 2923 | */ |
| 2924 | if (point != 0 && user_fence.flags & I915_EXEC_FENCE_SIGNAL(1<<1)) { |
| 2925 | /* |
| 2926 | * Waiting and signaling the same point (when point != |
| 2927 | * 0) would break the timeline. |
| 2928 | */ |
| 2929 | if (user_fence.flags & I915_EXEC_FENCE_WAIT(1<<0)) { |
| 2930 | DRM_DEBUG("Trying to wait & signal the same timeline point.\n")___drm_dbg(((void *)0), DRM_UT_CORE, "Trying to wait & signal the same timeline point.\n" ); |
| 2931 | dma_fence_put(fence); |
| 2932 | drm_syncobj_put(syncobj); |
| 2933 | return -EINVAL22; |
| 2934 | } |
| 2935 | |
| 2936 | f->chain_fence = dma_fence_chain_alloc(); |
| 2937 | if (!f->chain_fence) { |
| 2938 | drm_syncobj_put(syncobj); |
| 2939 | dma_fence_put(fence); |
| 2940 | return -ENOMEM12; |
| 2941 | } |
| 2942 | } else { |
| 2943 | f->chain_fence = NULL((void *)0); |
| 2944 | } |
| 2945 | |
| 2946 | f->syncobj = ptr_pack_bits(syncobj, user_fence.flags, 2)({ unsigned long __bits = (user_fence.flags); ((void)0); ((typeof (syncobj))((unsigned long)(syncobj) | __bits)); }); |
| 2947 | f->dma_fence = fence; |
| 2948 | f->value = point; |
| 2949 | f++; |
| 2950 | eb->num_fences++; |
| 2951 | } |
| 2952 | |
| 2953 | return 0; |
| 2954 | } |
| 2955 | |
| 2956 | static int add_fence_array(struct i915_execbuffer *eb) |
| 2957 | { |
| 2958 | struct drm_i915_gem_execbuffer2 *args = eb->args; |
| 2959 | struct drm_i915_gem_exec_fence __user *user; |
| 2960 | unsigned long num_fences = args->num_cliprects; |
| 2961 | struct eb_fence *f; |
| 2962 | |
| 2963 | if (!(args->flags & I915_EXEC_FENCE_ARRAY(1<<19))) |
| 2964 | return 0; |
| 2965 | |
| 2966 | if (!num_fences) |
| 2967 | return 0; |
| 2968 | |
| 2969 | /* Check multiplication overflow for access_ok() and kvmalloc_array() */ |
| 2970 | BUILD_BUG_ON(sizeof(size_t) > sizeof(unsigned long))extern char _ctassert[(!(sizeof(size_t) > sizeof(unsigned long ))) ? 1 : -1 ] __attribute__((__unused__)); |
| 2971 | if (num_fences > min_t(unsigned long,({ unsigned long __min_a = (0xffffffffffffffffUL / sizeof(*user )); unsigned long __min_b = (0xffffffffffffffffUL / sizeof(*f ) - eb->num_fences); __min_a < __min_b ? __min_a : __min_b ; }) |
| 2972 | ULONG_MAX / sizeof(*user),({ unsigned long __min_a = (0xffffffffffffffffUL / sizeof(*user )); unsigned long __min_b = (0xffffffffffffffffUL / sizeof(*f ) - eb->num_fences); __min_a < __min_b ? __min_a : __min_b ; }) |
| 2973 | SIZE_MAX / sizeof(*f) - eb->num_fences)({ unsigned long __min_a = (0xffffffffffffffffUL / sizeof(*user )); unsigned long __min_b = (0xffffffffffffffffUL / sizeof(*f ) - eb->num_fences); __min_a < __min_b ? __min_a : __min_b ; })) |
| 2974 | return -EINVAL22; |
| 2975 | |
| 2976 | user = u64_to_user_ptr(args->cliprects_ptr)((void *)(uintptr_t)(args->cliprects_ptr)); |
| 2977 | if (!access_ok(user, num_fences * sizeof(*user))) |
| 2978 | return -EFAULT14; |
| 2979 | |
| 2980 | #ifdef __linux__ |
| 2981 | f = krealloc(eb->fences, |
| 2982 | (eb->num_fences + num_fences) * sizeof(*f), |
| 2983 | __GFP_NOWARN0 | GFP_KERNEL(0x0001 | 0x0004)); |
| 2984 | if (!f) |
| 2985 | return -ENOMEM12; |
| 2986 | #else |
| 2987 | f = kmalloc((eb->num_fences + num_fences) * sizeof(*f), |
| 2988 | __GFP_NOWARN0 | GFP_KERNEL(0x0001 | 0x0004)); |
| 2989 | if (!f) |
| 2990 | return -ENOMEM12; |
| 2991 | memcpy(f, eb->fences, eb->num_fences * sizeof(*f))__builtin_memcpy((f), (eb->fences), (eb->num_fences * sizeof (*f))); |
| 2992 | kfree(eb->fences); |
| 2993 | #endif |
| 2994 | |
| 2995 | eb->fences = f; |
| 2996 | f += eb->num_fences; |
| 2997 | while (num_fences--) { |
| 2998 | struct drm_i915_gem_exec_fence user_fence; |
| 2999 | struct drm_syncobj *syncobj; |
| 3000 | struct dma_fence *fence = NULL((void *)0); |
| 3001 | |
| 3002 | if (__copy_from_user(&user_fence, user++, sizeof(user_fence))) |
| 3003 | return -EFAULT14; |
| 3004 | |
| 3005 | if (user_fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS(-((1<<1) << 1))) |
| 3006 | return -EINVAL22; |
| 3007 | |
| 3008 | syncobj = drm_syncobj_find(eb->file, user_fence.handle); |
| 3009 | if (!syncobj) { |
| 3010 | DRM_DEBUG("Invalid syncobj handle provided\n")___drm_dbg(((void *)0), DRM_UT_CORE, "Invalid syncobj handle provided\n" ); |
| 3011 | return -ENOENT2; |
| 3012 | } |
| 3013 | |
| 3014 | if (user_fence.flags & I915_EXEC_FENCE_WAIT(1<<0)) { |
| 3015 | fence = drm_syncobj_fence_get(syncobj); |
| 3016 | if (!fence) { |
| 3017 | DRM_DEBUG("Syncobj handle has no fence\n")___drm_dbg(((void *)0), DRM_UT_CORE, "Syncobj handle has no fence\n" ); |
| 3018 | drm_syncobj_put(syncobj); |
| 3019 | return -EINVAL22; |
| 3020 | } |
| 3021 | } |
| 3022 | |
| 3023 | #ifdef notyet |
| 3024 | BUILD_BUG_ON(~(ARCH_KMALLOC_MINALIGN - 1) &extern char _ctassert[(!(~(64 - 1) & ~(-((1<<1) << 1)))) ? 1 : -1 ] __attribute__((__unused__)) |
| 3025 | ~__I915_EXEC_FENCE_UNKNOWN_FLAGS)extern char _ctassert[(!(~(64 - 1) & ~(-((1<<1) << 1)))) ? 1 : -1 ] __attribute__((__unused__)); |
| 3026 | #endif |
| 3027 | |
| 3028 | f->syncobj = ptr_pack_bits(syncobj, user_fence.flags, 2)({ unsigned long __bits = (user_fence.flags); ((void)0); ((typeof (syncobj))((unsigned long)(syncobj) | __bits)); }); |
| 3029 | f->dma_fence = fence; |
| 3030 | f->value = 0; |
| 3031 | f->chain_fence = NULL((void *)0); |
| 3032 | f++; |
| 3033 | eb->num_fences++; |
| 3034 | } |
| 3035 | |
| 3036 | return 0; |
| 3037 | } |
| 3038 | |
| 3039 | static void put_fence_array(struct eb_fence *fences, int num_fences) |
| 3040 | { |
| 3041 | if (fences) |
| 3042 | __free_fence_array(fences, num_fences); |
| 3043 | } |
| 3044 | |
| 3045 | static int |
| 3046 | await_fence_array(struct i915_execbuffer *eb, |
| 3047 | struct i915_request *rq) |
| 3048 | { |
| 3049 | unsigned int n; |
| 3050 | int err; |
| 3051 | |
| 3052 | for (n = 0; n < eb->num_fences; n++) { |
| 3053 | struct drm_syncobj *syncobj; |
| 3054 | unsigned int flags; |
| 3055 | |
| 3056 | syncobj = ptr_unpack_bits(eb->fences[n].syncobj, &flags, 2)({ unsigned long __v = (unsigned long)(eb->fences[n].syncobj ); *(&flags) = __v & ((1UL << (2)) - 1); (typeof (eb->fences[n].syncobj))(__v & -(1UL << (2))); } ); |
Value stored to 'syncobj' is never read | |
| 3057 | |
| 3058 | if (!eb->fences[n].dma_fence) |
| 3059 | continue; |
| 3060 | |
| 3061 | err = i915_request_await_dma_fence(rq, eb->fences[n].dma_fence); |
| 3062 | if (err < 0) |
| 3063 | return err; |
| 3064 | } |
| 3065 | |
| 3066 | return 0; |
| 3067 | } |
| 3068 | |
| 3069 | static void signal_fence_array(const struct i915_execbuffer *eb, |
| 3070 | struct dma_fence * const fence) |
| 3071 | { |
| 3072 | unsigned int n; |
| 3073 | |
| 3074 | for (n = 0; n < eb->num_fences; n++) { |
| 3075 | struct drm_syncobj *syncobj; |
| 3076 | unsigned int flags; |
| 3077 | |
| 3078 | syncobj = ptr_unpack_bits(eb->fences[n].syncobj, &flags, 2)({ unsigned long __v = (unsigned long)(eb->fences[n].syncobj ); *(&flags) = __v & ((1UL << (2)) - 1); (typeof (eb->fences[n].syncobj))(__v & -(1UL << (2))); } ); |
| 3079 | if (!(flags & I915_EXEC_FENCE_SIGNAL(1<<1))) |
| 3080 | continue; |
| 3081 | |
| 3082 | if (eb->fences[n].chain_fence) { |
| 3083 | drm_syncobj_add_point(syncobj, |
| 3084 | eb->fences[n].chain_fence, |
| 3085 | fence, |
| 3086 | eb->fences[n].value); |
| 3087 | /* |
| 3088 | * The chain's ownership is transferred to the |
| 3089 | * timeline. |
| 3090 | */ |
| 3091 | eb->fences[n].chain_fence = NULL((void *)0); |
| 3092 | } else { |
| 3093 | drm_syncobj_replace_fence(syncobj, fence); |
| 3094 | } |
| 3095 | } |
| 3096 | } |
| 3097 | |
| 3098 | static int |
| 3099 | parse_timeline_fences(struct i915_user_extension __user *ext, void *data) |
| 3100 | { |
| 3101 | struct i915_execbuffer *eb = data; |
| 3102 | struct drm_i915_gem_execbuffer_ext_timeline_fences timeline_fences; |
| 3103 | |
| 3104 | if (copy_from_user(&timeline_fences, ext, sizeof(timeline_fences))) |
| 3105 | return -EFAULT14; |
| 3106 | |
| 3107 | return add_timeline_fence_array(eb, &timeline_fences); |
| 3108 | } |
| 3109 | |
| 3110 | static void retire_requests(struct intel_timeline *tl, struct i915_request *end) |
| 3111 | { |
| 3112 | struct i915_request *rq, *rn; |
| 3113 | |
| 3114 | list_for_each_entry_safe(rq, rn, &tl->requests, link)for (rq = ({ const __typeof( ((__typeof(*rq) *)0)->link ) * __mptr = ((&tl->requests)->next); (__typeof(*rq) *) ( (char *)__mptr - __builtin_offsetof(__typeof(*rq), link) ); }), rn = ({ const __typeof( ((__typeof(*rq) *)0)->link ) * __mptr = (rq->link.next); (__typeof(*rq) *)( (char *)__mptr - __builtin_offsetof(__typeof(*rq), link) );}); &rq-> link != (&tl->requests); rq = rn, rn = ({ const __typeof ( ((__typeof(*rn) *)0)->link ) *__mptr = (rn->link.next ); (__typeof(*rn) *)( (char *)__mptr - __builtin_offsetof(__typeof (*rn), link) );})) |
| 3115 | if (rq == end || !i915_request_retire(rq)) |
| 3116 | break; |
| 3117 | } |
| 3118 | |
| 3119 | static int eb_request_add(struct i915_execbuffer *eb, struct i915_request *rq, |
| 3120 | int err, bool_Bool last_parallel) |
| 3121 | { |
| 3122 | struct intel_timeline * const tl = i915_request_timeline(rq); |
| 3123 | struct i915_sched_attr attr = {}; |
| 3124 | struct i915_request *prev; |
| 3125 | |
| 3126 | lockdep_assert_held(&tl->mutex)do { (void)(&tl->mutex); } while(0); |
| 3127 | lockdep_unpin_lock(&tl->mutex, rq->cookie); |
| 3128 | |
| 3129 | trace_i915_request_add(rq); |
| 3130 | |
| 3131 | prev = __i915_request_commit(rq); |
| 3132 | |
| 3133 | /* Check that the context wasn't destroyed before submission */ |
| 3134 | if (likely(!intel_context_is_closed(eb->context))__builtin_expect(!!(!intel_context_is_closed(eb->context)) , 1)) { |
| 3135 | attr = eb->gem_context->sched; |
| 3136 | } else { |
| 3137 | /* Serialise with context_close via the add_to_timeline */ |
| 3138 | i915_request_set_error_once(rq, -ENOENT2); |
| 3139 | __i915_request_skip(rq); |
| 3140 | err = -ENOENT2; /* override any transient errors */ |
| 3141 | } |
| 3142 | |
| 3143 | if (intel_context_is_parallel(eb->context)) { |
| 3144 | if (err) { |
| 3145 | __i915_request_skip(rq); |
| 3146 | set_bit(I915_FENCE_FLAG_SKIP_PARALLEL, |
| 3147 | &rq->fence.flags); |
| 3148 | } |
| 3149 | if (last_parallel) |
| 3150 | set_bit(I915_FENCE_FLAG_SUBMIT_PARALLEL, |
| 3151 | &rq->fence.flags); |
| 3152 | } |
| 3153 | |
| 3154 | __i915_request_queue(rq, &attr); |
| 3155 | |
| 3156 | /* Try to clean up the client's timeline after submitting the request */ |
| 3157 | if (prev) |
| 3158 | retire_requests(tl, prev); |
| 3159 | |
| 3160 | mutex_unlock(&tl->mutex)rw_exit_write(&tl->mutex); |
| 3161 | |
| 3162 | return err; |
| 3163 | } |
| 3164 | |
| 3165 | static int eb_requests_add(struct i915_execbuffer *eb, int err) |
| 3166 | { |
| 3167 | int i; |
| 3168 | |
| 3169 | /* |
| 3170 | * We iterate in reverse order of creation to release timeline mutexes in |
| 3171 | * same order. |
| 3172 | */ |
| 3173 | for_each_batch_add_order(eb, i)extern char _ctassert[(!(!1)) ? 1 : -1 ] __attribute__((__unused__ )); for ((i) = (eb)->num_batches - 1; (i) >= 0; --(i)) { |
| 3174 | struct i915_request *rq = eb->requests[i]; |
| 3175 | |
| 3176 | if (!rq) |
| 3177 | continue; |
| 3178 | err |= eb_request_add(eb, rq, err, i == 0); |
| 3179 | } |
| 3180 | |
| 3181 | return err; |
| 3182 | } |
| 3183 | |
| 3184 | static const i915_user_extension_fn execbuf_extensions[] = { |
| 3185 | [DRM_I915_GEM_EXECBUFFER_EXT_TIMELINE_FENCES0] = parse_timeline_fences, |
| 3186 | }; |
| 3187 | |
| 3188 | static int |
| 3189 | parse_execbuf2_extensions(struct drm_i915_gem_execbuffer2 *args, |
| 3190 | struct i915_execbuffer *eb) |
| 3191 | { |
| 3192 | if (!(args->flags & I915_EXEC_USE_EXTENSIONS(1 << 21))) |
| 3193 | return 0; |
| 3194 | |
| 3195 | /* The execbuf2 extension mechanism reuses cliprects_ptr. So we cannot |
| 3196 | * have another flag also using it at the same time. |
| 3197 | */ |
| 3198 | if (eb->args->flags & I915_EXEC_FENCE_ARRAY(1<<19)) |
| 3199 | return -EINVAL22; |
| 3200 | |
| 3201 | if (args->num_cliprects != 0) |
| 3202 | return -EINVAL22; |
| 3203 | |
| 3204 | return i915_user_extensions(u64_to_user_ptr(args->cliprects_ptr)((void *)(uintptr_t)(args->cliprects_ptr)), |
| 3205 | execbuf_extensions, |
| 3206 | ARRAY_SIZE(execbuf_extensions)(sizeof((execbuf_extensions)) / sizeof((execbuf_extensions)[0 ])), |
| 3207 | eb); |
| 3208 | } |
| 3209 | |
| 3210 | static void eb_requests_get(struct i915_execbuffer *eb) |
| 3211 | { |
| 3212 | unsigned int i; |
| 3213 | |
| 3214 | for_each_batch_create_order(eb, i)for ((i) = 0; (i) < (eb)->num_batches; ++(i)) { |
| 3215 | if (!eb->requests[i]) |
| 3216 | break; |
| 3217 | |
| 3218 | i915_request_get(eb->requests[i]); |
| 3219 | } |
| 3220 | } |
| 3221 | |
| 3222 | static void eb_requests_put(struct i915_execbuffer *eb) |
| 3223 | { |
| 3224 | unsigned int i; |
| 3225 | |
| 3226 | for_each_batch_create_order(eb, i)for ((i) = 0; (i) < (eb)->num_batches; ++(i)) { |
| 3227 | if (!eb->requests[i]) |
| 3228 | break; |
| 3229 | |
| 3230 | i915_request_put(eb->requests[i]); |
| 3231 | } |
| 3232 | } |
| 3233 | |
| 3234 | static struct sync_file * |
| 3235 | eb_composite_fence_create(struct i915_execbuffer *eb, int out_fence_fd) |
| 3236 | { |
| 3237 | struct sync_file *out_fence = NULL((void *)0); |
| 3238 | struct dma_fence_array *fence_array; |
| 3239 | struct dma_fence **fences; |
| 3240 | unsigned int i; |
| 3241 | |
| 3242 | GEM_BUG_ON(!intel_context_is_parent(eb->context))((void)0); |
| 3243 | |
| 3244 | fences = kmalloc_array(eb->num_batches, sizeof(*fences), GFP_KERNEL(0x0001 | 0x0004)); |
| 3245 | if (!fences) |
| 3246 | return ERR_PTR(-ENOMEM12); |
| 3247 | |
| 3248 | for_each_batch_create_order(eb, i)for ((i) = 0; (i) < (eb)->num_batches; ++(i)) { |
| 3249 | fences[i] = &eb->requests[i]->fence; |
| 3250 | __set_bit(I915_FENCE_FLAG_COMPOSITE, |
| 3251 | &eb->requests[i]->fence.flags); |
| 3252 | } |
| 3253 | |
| 3254 | fence_array = dma_fence_array_create(eb->num_batches, |
| 3255 | fences, |
| 3256 | eb->context->parallel.fence_context, |
| 3257 | eb->context->parallel.seqno++, |
| 3258 | false0); |
| 3259 | if (!fence_array) { |
| 3260 | kfree(fences); |
| 3261 | return ERR_PTR(-ENOMEM12); |
| 3262 | } |
| 3263 | |
| 3264 | /* Move ownership to the dma_fence_array created above */ |
| 3265 | for_each_batch_create_order(eb, i)for ((i) = 0; (i) < (eb)->num_batches; ++(i)) |
| 3266 | dma_fence_get(fences[i]); |
| 3267 | |
| 3268 | if (out_fence_fd != -1) { |
| 3269 | out_fence = sync_file_create(&fence_array->base); |
| 3270 | /* sync_file now owns fence_arry, drop creation ref */ |
| 3271 | dma_fence_put(&fence_array->base); |
| 3272 | if (!out_fence) |
| 3273 | return ERR_PTR(-ENOMEM12); |
| 3274 | } |
| 3275 | |
| 3276 | eb->composite_fence = &fence_array->base; |
| 3277 | |
| 3278 | return out_fence; |
| 3279 | } |
| 3280 | |
| 3281 | static struct sync_file * |
| 3282 | eb_fences_add(struct i915_execbuffer *eb, struct i915_request *rq, |
| 3283 | struct dma_fence *in_fence, int out_fence_fd) |
| 3284 | { |
| 3285 | struct sync_file *out_fence = NULL((void *)0); |
| 3286 | int err; |
| 3287 | |
| 3288 | if (unlikely(eb->gem_context->syncobj)__builtin_expect(!!(eb->gem_context->syncobj), 0)) { |
| 3289 | struct dma_fence *fence; |
| 3290 | |
| 3291 | fence = drm_syncobj_fence_get(eb->gem_context->syncobj); |
| 3292 | err = i915_request_await_dma_fence(rq, fence); |
| 3293 | dma_fence_put(fence); |
| 3294 | if (err) |
| 3295 | return ERR_PTR(err); |
| 3296 | } |
| 3297 | |
| 3298 | if (in_fence) { |
| 3299 | if (eb->args->flags & I915_EXEC_FENCE_SUBMIT(1 << 20)) |
| 3300 | err = i915_request_await_execution(rq, in_fence); |
| 3301 | else |
| 3302 | err = i915_request_await_dma_fence(rq, in_fence); |
| 3303 | if (err < 0) |
| 3304 | return ERR_PTR(err); |
| 3305 | } |
| 3306 | |
| 3307 | if (eb->fences) { |
| 3308 | err = await_fence_array(eb, rq); |
| 3309 | if (err) |
| 3310 | return ERR_PTR(err); |
| 3311 | } |
| 3312 | |
| 3313 | if (intel_context_is_parallel(eb->context)) { |
| 3314 | out_fence = eb_composite_fence_create(eb, out_fence_fd); |
| 3315 | if (IS_ERR(out_fence)) |
| 3316 | return ERR_PTR(-ENOMEM12); |
| 3317 | } else if (out_fence_fd != -1) { |
| 3318 | out_fence = sync_file_create(&rq->fence); |
| 3319 | if (!out_fence) |
| 3320 | return ERR_PTR(-ENOMEM12); |
| 3321 | } |
| 3322 | |
| 3323 | return out_fence; |
| 3324 | } |
| 3325 | |
| 3326 | static struct intel_context * |
| 3327 | eb_find_context(struct i915_execbuffer *eb, unsigned int context_number) |
| 3328 | { |
| 3329 | struct intel_context *child; |
| 3330 | |
| 3331 | if (likely(context_number == 0)__builtin_expect(!!(context_number == 0), 1)) |
| 3332 | return eb->context; |
| 3333 | |
| 3334 | for_each_child(eb->context, child)for (child = ({ const __typeof( ((__typeof(*child) *)0)->parallel .child_link ) *__mptr = ((&(eb->context)->parallel. child_list)->next); (__typeof(*child) *)( (char *)__mptr - __builtin_offsetof(__typeof(*child), parallel.child_link) ); }); &child->parallel.child_link != (&(eb->context )->parallel.child_list); child = ({ const __typeof( ((__typeof (*child) *)0)->parallel.child_link ) *__mptr = (child-> parallel.child_link.next); (__typeof(*child) *)( (char *)__mptr - __builtin_offsetof(__typeof(*child), parallel.child_link) ) ;})) |
| 3335 | if (!--context_number) |
| 3336 | return child; |
| 3337 | |
| 3338 | GEM_BUG_ON("Context not found")((void)0); |
| 3339 | |
| 3340 | return NULL((void *)0); |
| 3341 | } |
| 3342 | |
| 3343 | static struct sync_file * |
| 3344 | eb_requests_create(struct i915_execbuffer *eb, struct dma_fence *in_fence, |
| 3345 | int out_fence_fd) |
| 3346 | { |
| 3347 | struct sync_file *out_fence = NULL((void *)0); |
| 3348 | unsigned int i; |
| 3349 | |
| 3350 | for_each_batch_create_order(eb, i)for ((i) = 0; (i) < (eb)->num_batches; ++(i)) { |
| 3351 | /* Allocate a request for this batch buffer nice and early. */ |
| 3352 | eb->requests[i] = i915_request_create(eb_find_context(eb, i)); |
| 3353 | if (IS_ERR(eb->requests[i])) { |
| 3354 | out_fence = ERR_CAST(eb->requests[i]); |
| 3355 | eb->requests[i] = NULL((void *)0); |
| 3356 | return out_fence; |
| 3357 | } |
| 3358 | |
| 3359 | /* |
| 3360 | * Only the first request added (committed to backend) has to |
| 3361 | * take the in fences into account as all subsequent requests |
| 3362 | * will have fences inserted inbetween them. |
| 3363 | */ |
| 3364 | if (i + 1 == eb->num_batches) { |
| 3365 | out_fence = eb_fences_add(eb, eb->requests[i], |
| 3366 | in_fence, out_fence_fd); |
| 3367 | if (IS_ERR(out_fence)) |
| 3368 | return out_fence; |
| 3369 | } |
| 3370 | |
| 3371 | /* |
| 3372 | * Not really on stack, but we don't want to call |
| 3373 | * kfree on the batch_snapshot when we put it, so use the |
| 3374 | * _onstack interface. |
| 3375 | */ |
| 3376 | if (eb->batches[i]->vma) |
| 3377 | eb->requests[i]->batch_res = |
| 3378 | i915_vma_resource_get(eb->batches[i]->vma->resource); |
| 3379 | if (eb->batch_pool) { |
| 3380 | GEM_BUG_ON(intel_context_is_parallel(eb->context))((void)0); |
| 3381 | intel_gt_buffer_pool_mark_active(eb->batch_pool, |
| 3382 | eb->requests[i]); |
| 3383 | } |
| 3384 | } |
| 3385 | |
| 3386 | return out_fence; |
| 3387 | } |
| 3388 | |
| 3389 | static int |
| 3390 | i915_gem_do_execbuffer(struct drm_device *dev, |
| 3391 | struct drm_file *file, |
| 3392 | struct drm_i915_gem_execbuffer2 *args, |
| 3393 | struct drm_i915_gem_exec_object2 *exec) |
| 3394 | { |
| 3395 | struct drm_i915_privateinteldrm_softc *i915 = to_i915(dev); |
| 3396 | struct i915_execbuffer eb; |
| 3397 | struct dma_fence *in_fence = NULL((void *)0); |
| 3398 | struct sync_file *out_fence = NULL((void *)0); |
| 3399 | int out_fence_fd = -1; |
| 3400 | int err; |
| 3401 | |
| 3402 | BUILD_BUG_ON(__EXEC_INTERNAL_FLAGS & ~__I915_EXEC_ILLEGAL_FLAGS)extern char _ctassert[(!((~0u << 29) & ~((-((1 << 21) << 1)) | (3<<6) | (1<<15)))) ? 1 : -1 ] __attribute__((__unused__)); |
| 3403 | BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS &extern char _ctassert[(!((~0u << 26) & ~-((1<< 7)<<1))) ? 1 : -1 ] __attribute__((__unused__)) |
| 3404 | ~__EXEC_OBJECT_UNKNOWN_FLAGS)extern char _ctassert[(!((~0u << 26) & ~-((1<< 7)<<1))) ? 1 : -1 ] __attribute__((__unused__)); |
| 3405 | |
| 3406 | eb.i915 = i915; |
| 3407 | eb.file = file; |
| 3408 | eb.args = args; |
| 3409 | if (DBG_FORCE_RELOC0 || !(args->flags & I915_EXEC_NO_RELOC(1<<11))) |
| 3410 | args->flags |= __EXEC_HAS_RELOC(1UL << (31)); |
| 3411 | |
| 3412 | eb.exec = exec; |
| 3413 | eb.vma = (struct eb_vma *)(exec + args->buffer_count + 1); |
| 3414 | eb.vma[0].vma = NULL((void *)0); |
| 3415 | eb.batch_pool = NULL((void *)0); |
| 3416 | |
| 3417 | eb.invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS-((1<<7)<<1); |
| 3418 | reloc_cache_init(&eb.reloc_cache, eb.i915); |
| 3419 | |
| 3420 | eb.buffer_count = args->buffer_count; |
| 3421 | eb.batch_start_offset = args->batch_start_offset; |
| 3422 | eb.trampoline = NULL((void *)0); |
| 3423 | |
| 3424 | eb.fences = NULL((void *)0); |
| 3425 | eb.num_fences = 0; |
| 3426 | |
| 3427 | eb_capture_list_clear(&eb); |
| 3428 | |
| 3429 | memset(eb.requests, 0, sizeof(struct i915_request *) *__builtin_memset((eb.requests), (0), (sizeof(struct i915_request *) * (sizeof((eb.requests)) / sizeof((eb.requests)[0])))) |
| 3430 | ARRAY_SIZE(eb.requests))__builtin_memset((eb.requests), (0), (sizeof(struct i915_request *) * (sizeof((eb.requests)) / sizeof((eb.requests)[0])))); |
| 3431 | eb.composite_fence = NULL((void *)0); |
| 3432 | |
| 3433 | eb.batch_flags = 0; |
| 3434 | if (args->flags & I915_EXEC_SECURE(1<<9)) { |
| 3435 | if (GRAPHICS_VER(i915)((&(i915)->__runtime)->graphics.ip.ver) >= 11) |
| 3436 | return -ENODEV19; |
| 3437 | |
| 3438 | /* Return -EPERM to trigger fallback code on old binaries. */ |
| 3439 | if (!HAS_SECURE_BATCHES(i915)(((&(i915)->__runtime)->graphics.ip.ver) < 6)) |
| 3440 | return -EPERM1; |
| 3441 | |
| 3442 | if (!drm_is_current_master(file) || !capable(CAP_SYS_ADMIN0x1)) |
| 3443 | return -EPERM1; |
| 3444 | |
| 3445 | eb.batch_flags |= I915_DISPATCH_SECURE(1UL << (0)); |
| 3446 | } |
| 3447 | if (args->flags & I915_EXEC_IS_PINNED(1<<10)) |
| 3448 | eb.batch_flags |= I915_DISPATCH_PINNED(1UL << (1)); |
| 3449 | |
| 3450 | err = parse_execbuf2_extensions(args, &eb); |
| 3451 | if (err) |
| 3452 | goto err_ext; |
| 3453 | |
| 3454 | err = add_fence_array(&eb); |
| 3455 | if (err) |
| 3456 | goto err_ext; |
| 3457 | |
| 3458 | #define IN_FENCES (I915_EXEC_FENCE_IN(1<<16) | I915_EXEC_FENCE_SUBMIT(1 << 20)) |
| 3459 | if (args->flags & IN_FENCES) { |
| 3460 | if ((args->flags & IN_FENCES) == IN_FENCES) |
| 3461 | return -EINVAL22; |
| 3462 | |
| 3463 | in_fence = sync_file_get_fence(lower_32_bits(args->rsvd2)((u32)(args->rsvd2))); |
| 3464 | if (!in_fence) { |
| 3465 | err = -EINVAL22; |
| 3466 | goto err_ext; |
| 3467 | } |
| 3468 | } |
| 3469 | #undef IN_FENCES |
| 3470 | |
| 3471 | if (args->flags & I915_EXEC_FENCE_OUT(1<<17)) { |
| 3472 | out_fence_fd = get_unused_fd_flags(O_CLOEXEC0x10000); |
| 3473 | if (out_fence_fd < 0) { |
| 3474 | err = out_fence_fd; |
| 3475 | goto err_in_fence; |
| 3476 | } |
| 3477 | } |
| 3478 | |
| 3479 | err = eb_create(&eb); |
| 3480 | if (err) |
| 3481 | goto err_out_fence; |
| 3482 | |
| 3483 | GEM_BUG_ON(!eb.lut_size)((void)0); |
| 3484 | |
| 3485 | err = eb_select_context(&eb); |
| 3486 | if (unlikely(err)__builtin_expect(!!(err), 0)) |
| 3487 | goto err_destroy; |
| 3488 | |
| 3489 | err = eb_select_engine(&eb); |
| 3490 | if (unlikely(err)__builtin_expect(!!(err), 0)) |
| 3491 | goto err_context; |
| 3492 | |
| 3493 | err = eb_lookup_vmas(&eb); |
| 3494 | if (err) { |
| 3495 | eb_release_vmas(&eb, true1); |
| 3496 | goto err_engine; |
| 3497 | } |
| 3498 | |
| 3499 | i915_gem_ww_ctx_init(&eb.ww, true1); |
| 3500 | |
| 3501 | err = eb_relocate_parse(&eb); |
| 3502 | if (err) { |
| 3503 | /* |
| 3504 | * If the user expects the execobject.offset and |
| 3505 | * reloc.presumed_offset to be an exact match, |
| 3506 | * as for using NO_RELOC, then we cannot update |
| 3507 | * the execobject.offset until we have completed |
| 3508 | * relocation. |
| 3509 | */ |
| 3510 | args->flags &= ~__EXEC_HAS_RELOC(1UL << (31)); |
| 3511 | goto err_vma; |
| 3512 | } |
| 3513 | |
| 3514 | ww_acquire_done(&eb.ww.ctx); |
| 3515 | err = eb_capture_stage(&eb); |
| 3516 | if (err) |
| 3517 | goto err_vma; |
| 3518 | |
| 3519 | out_fence = eb_requests_create(&eb, in_fence, out_fence_fd); |
| 3520 | if (IS_ERR(out_fence)) { |
| 3521 | err = PTR_ERR(out_fence); |
| 3522 | out_fence = NULL((void *)0); |
| 3523 | if (eb.requests[0]) |
| 3524 | goto err_request; |
| 3525 | else |
| 3526 | goto err_vma; |
| 3527 | } |
| 3528 | |
| 3529 | err = eb_submit(&eb); |
| 3530 | |
| 3531 | err_request: |
| 3532 | eb_requests_get(&eb); |
| 3533 | err = eb_requests_add(&eb, err); |
| 3534 | |
| 3535 | if (eb.fences) |
| 3536 | signal_fence_array(&eb, eb.composite_fence ? |
| 3537 | eb.composite_fence : |
| 3538 | &eb.requests[0]->fence); |
| 3539 | |
| 3540 | if (unlikely(eb.gem_context->syncobj)__builtin_expect(!!(eb.gem_context->syncobj), 0)) { |
| 3541 | drm_syncobj_replace_fence(eb.gem_context->syncobj, |
| 3542 | eb.composite_fence ? |
| 3543 | eb.composite_fence : |
| 3544 | &eb.requests[0]->fence); |
| 3545 | } |
| 3546 | |
| 3547 | if (out_fence) { |
| 3548 | if (err == 0) { |
| 3549 | fd_install(out_fence_fd, out_fence->file); |
| 3550 | args->rsvd2 &= GENMASK_ULL(31, 0)(((~0ULL) >> (64 - (31) - 1)) & ((~0ULL) << ( 0))); /* keep in-fence */ |
| 3551 | args->rsvd2 |= (u64)out_fence_fd << 32; |
| 3552 | out_fence_fd = -1; |
| 3553 | } else { |
| 3554 | fput(out_fence->file); |
| 3555 | } |
| 3556 | } |
| 3557 | |
| 3558 | if (!out_fence && eb.composite_fence) |
| 3559 | dma_fence_put(eb.composite_fence); |
| 3560 | |
| 3561 | eb_requests_put(&eb); |
| 3562 | |
| 3563 | err_vma: |
| 3564 | eb_release_vmas(&eb, true1); |
| 3565 | WARN_ON(err == -EDEADLK)({ int __ret = !!(err == -11); if (__ret) printf("WARNING %s failed at %s:%d\n" , "err == -11", "/usr/src/sys/dev/pci/drm/i915/gem/i915_gem_execbuffer.c" , 3565); __builtin_expect(!!(__ret), 0); }); |
| 3566 | i915_gem_ww_ctx_fini(&eb.ww); |
| 3567 | |
| 3568 | if (eb.batch_pool) |
| 3569 | intel_gt_buffer_pool_put(eb.batch_pool); |
| 3570 | err_engine: |
| 3571 | eb_put_engine(&eb); |
| 3572 | err_context: |
| 3573 | i915_gem_context_put(eb.gem_context); |
| 3574 | err_destroy: |
| 3575 | eb_destroy(&eb); |
| 3576 | err_out_fence: |
| 3577 | if (out_fence_fd != -1) |
| 3578 | put_unused_fd(out_fence_fd); |
| 3579 | err_in_fence: |
| 3580 | dma_fence_put(in_fence); |
| 3581 | err_ext: |
| 3582 | put_fence_array(eb.fences, eb.num_fences); |
| 3583 | return err; |
| 3584 | } |
| 3585 | |
| 3586 | static size_t eb_element_size(void) |
| 3587 | { |
| 3588 | return sizeof(struct drm_i915_gem_exec_object2) + sizeof(struct eb_vma); |
| 3589 | } |
| 3590 | |
| 3591 | static bool_Bool check_buffer_count(size_t count) |
| 3592 | { |
| 3593 | const size_t sz = eb_element_size(); |
| 3594 | |
| 3595 | /* |
| 3596 | * When using LUT_HANDLE, we impose a limit of INT_MAX for the lookup |
| 3597 | * array size (see eb_create()). Otherwise, we can accept an array as |
| 3598 | * large as can be addressed (though use large arrays at your peril)! |
| 3599 | */ |
| 3600 | |
| 3601 | return !(count < 1 || count > INT_MAX0x7fffffff || count > SIZE_MAX0xffffffffffffffffUL / sz - 1); |
| 3602 | } |
| 3603 | |
| 3604 | int |
| 3605 | i915_gem_execbuffer2_ioctl(struct drm_device *dev, void *data, |
| 3606 | struct drm_file *file) |
| 3607 | { |
| 3608 | struct drm_i915_privateinteldrm_softc *i915 = to_i915(dev); |
| 3609 | struct drm_i915_gem_execbuffer2 *args = data; |
| 3610 | struct drm_i915_gem_exec_object2 *exec2_list; |
| 3611 | const size_t count = args->buffer_count; |
| 3612 | int err; |
| 3613 | |
| 3614 | if (!check_buffer_count(count)) { |
| 3615 | drm_dbg(&i915->drm, "execbuf2 with %zd buffers\n", count)__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "execbuf2 with %zd buffers\n" , count); |
| 3616 | return -EINVAL22; |
| 3617 | } |
| 3618 | |
| 3619 | err = i915_gem_check_execbuffer(args); |
| 3620 | if (err) |
| 3621 | return err; |
| 3622 | |
| 3623 | /* Allocate extra slots for use by the command parser */ |
| 3624 | exec2_list = kvmalloc_array(count + 2, eb_element_size(), |
| 3625 | __GFP_NOWARN0 | GFP_KERNEL(0x0001 | 0x0004)); |
| 3626 | if (exec2_list == NULL((void *)0)) { |
| 3627 | drm_dbg(&i915->drm, "Failed to allocate exec list for %zd buffers\n",__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Failed to allocate exec list for %zd buffers\n" , count) |
| 3628 | count)__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "Failed to allocate exec list for %zd buffers\n" , count); |
| 3629 | return -ENOMEM12; |
| 3630 | } |
| 3631 | if (copy_from_user(exec2_list, |
| 3632 | u64_to_user_ptr(args->buffers_ptr)((void *)(uintptr_t)(args->buffers_ptr)), |
| 3633 | sizeof(*exec2_list) * count)) { |
| 3634 | drm_dbg(&i915->drm, "copy %zd exec entries failed\n", count)__drm_dev_dbg(((void *)0), (&i915->drm) ? (&i915-> drm)->dev : ((void *)0), DRM_UT_DRIVER, "copy %zd exec entries failed\n" , count); |
| 3635 | kvfree(exec2_list); |
| 3636 | return -EFAULT14; |
| 3637 | } |
| 3638 | |
| 3639 | err = i915_gem_do_execbuffer(dev, file, args, exec2_list); |
| 3640 | |
| 3641 | /* |
| 3642 | * Now that we have begun execution of the batchbuffer, we ignore |
| 3643 | * any new error after this point. Also given that we have already |
| 3644 | * updated the associated relocations, we try to write out the current |
| 3645 | * object locations irrespective of any error. |
| 3646 | */ |
| 3647 | if (args->flags & __EXEC_HAS_RELOC(1UL << (31))) { |
| 3648 | struct drm_i915_gem_exec_object2 __user *user_exec_list = |
| 3649 | u64_to_user_ptr(args->buffers_ptr)((void *)(uintptr_t)(args->buffers_ptr)); |
| 3650 | unsigned int i; |
| 3651 | |
| 3652 | /* Copy the new buffer offsets back to the user's exec list. */ |
| 3653 | /* |
| 3654 | * Note: count * sizeof(*user_exec_list) does not overflow, |
| 3655 | * because we checked 'count' in check_buffer_count(). |
| 3656 | * |
| 3657 | * And this range already got effectively checked earlier |
| 3658 | * when we did the "copy_from_user()" above. |
| 3659 | */ |
| 3660 | if (!user_write_access_begin(user_exec_list,access_ok(user_exec_list, count * sizeof(*user_exec_list)) |
| 3661 | count * sizeof(*user_exec_list))access_ok(user_exec_list, count * sizeof(*user_exec_list))) |
| 3662 | goto end; |
| 3663 | |
| 3664 | for (i = 0; i < args->buffer_count; i++) { |
| 3665 | if (!(exec2_list[i].offset & UPDATE(1ULL << (7)))) |
| 3666 | continue; |
| 3667 | |
| 3668 | exec2_list[i].offset = |
| 3669 | gen8_canonical_addr(exec2_list[i].offset & PIN_OFFSET_MASK-(1ULL << (12))); |
| 3670 | unsafe_put_user(exec2_list[i].offset,({ __typeof((exec2_list[i].offset)) __tmp = (exec2_list[i].offset ); if (copyout(&(__tmp), &user_exec_list[i].offset, sizeof (__tmp)) != 0) goto end_user; }) |
| 3671 | &user_exec_list[i].offset,({ __typeof((exec2_list[i].offset)) __tmp = (exec2_list[i].offset ); if (copyout(&(__tmp), &user_exec_list[i].offset, sizeof (__tmp)) != 0) goto end_user; }) |
| 3672 | end_user)({ __typeof((exec2_list[i].offset)) __tmp = (exec2_list[i].offset ); if (copyout(&(__tmp), &user_exec_list[i].offset, sizeof (__tmp)) != 0) goto end_user; }); |
| 3673 | } |
| 3674 | end_user: |
| 3675 | user_write_access_end(); |
| 3676 | end:; |
| 3677 | } |
| 3678 | |
| 3679 | args->flags &= ~__I915_EXEC_UNKNOWN_FLAGS(-((1 << 21) << 1)); |
| 3680 | kvfree(exec2_list); |
| 3681 | return err; |
| 3682 | } |