1 /* 2 * Copyright © 2008-2018 Intel Corporation 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 21 * IN THE SOFTWARE. 22 * 23 */ 24 25 #ifndef I915_REQUEST_H 26 #define I915_REQUEST_H 27 28 #include <linux/dma-fence.h> 29 #include <linux/lockdep.h> 30 31 #include "gt/intel_context_types.h" 32 #include "gt/intel_engine_types.h" 33 34 #include "i915_gem.h" 35 #include "i915_scheduler.h" 36 #include "i915_selftest.h" 37 #include "i915_sw_fence.h" 38 39 #include <uapi/drm/i915_drm.h> 40 41 struct drm_file; 42 struct drm_i915_gem_object; 43 struct i915_request; 44 struct intel_timeline; 45 struct intel_timeline_cacheline; 46 47 struct i915_capture_list { 48 struct i915_capture_list *next; 49 struct i915_vma *vma; 50 }; 51 52 enum { 53 /* 54 * I915_FENCE_FLAG_ACTIVE - this request is currently submitted to HW. 55 * 56 * Set by __i915_request_submit() on handing over to HW, and cleared 57 * by __i915_request_unsubmit() if we preempt this request. 58 * 59 * Finally cleared for consistency on retiring the request, when 60 * we know the HW is no longer running this request. 61 * 62 * See i915_request_is_active() 63 */ 64 I915_FENCE_FLAG_ACTIVE = DMA_FENCE_FLAG_USER_BITS, 65 66 /* 67 * I915_FENCE_FLAG_SIGNAL - this request is currently on signal_list 68 * 69 * Internal bookkeeping used by the breadcrumb code to track when 70 * a request is on the various signal_list. 71 */ 72 I915_FENCE_FLAG_SIGNAL, 73 }; 74 75 /** 76 * Request queue structure. 77 * 78 * The request queue allows us to note sequence numbers that have been emitted 79 * and may be associated with active buffers to be retired. 80 * 81 * By keeping this list, we can avoid having to do questionable sequence 82 * number comparisons on buffer last_read|write_seqno. It also allows an 83 * emission time to be associated with the request for tracking how far ahead 84 * of the GPU the submission is. 85 * 86 * When modifying this structure be very aware that we perform a lockless 87 * RCU lookup of it that may race against reallocation of the struct 88 * from the slab freelist. We intentionally do not zero the structure on 89 * allocation so that the lookup can use the dangling pointers (and is 90 * cogniscent that those pointers may be wrong). Instead, everything that 91 * needs to be initialised must be done so explicitly. 92 * 93 * The requests are reference counted. 94 */ 95 struct i915_request { 96 struct dma_fence fence; 97 spinlock_t lock; 98 99 /** On Which ring this request was generated */ 100 struct drm_i915_private *i915; 101 102 /** 103 * Context and ring buffer related to this request 104 * Contexts are refcounted, so when this request is associated with a 105 * context, we must increment the context's refcount, to guarantee that 106 * it persists while any request is linked to it. Requests themselves 107 * are also refcounted, so the request will only be freed when the last 108 * reference to it is dismissed, and the code in 109 * i915_request_free() will then decrement the refcount on the 110 * context. 111 */ 112 struct i915_gem_context *gem_context; 113 struct intel_engine_cs *engine; 114 struct intel_context *hw_context; 115 struct intel_ring *ring; 116 struct intel_timeline __rcu *timeline; 117 struct list_head signal_link; 118 119 /* 120 * The rcu epoch of when this request was allocated. Used to judiciously 121 * apply backpressure on future allocations to ensure that under 122 * mempressure there is sufficient RCU ticks for us to reclaim our 123 * RCU protected slabs. 124 */ 125 unsigned long rcustate; 126 127 /* 128 * We pin the timeline->mutex while constructing the request to 129 * ensure that no caller accidentally drops it during construction. 130 * The timeline->mutex must be held to ensure that only this caller 131 * can use the ring and manipulate the associated timeline during 132 * construction. 133 */ 134 struct pin_cookie cookie; 135 136 /* 137 * Fences for the various phases in the request's lifetime. 138 * 139 * The submit fence is used to await upon all of the request's 140 * dependencies. When it is signaled, the request is ready to run. 141 * It is used by the driver to then queue the request for execution. 142 */ 143 struct i915_sw_fence submit; 144 union { 145 wait_queue_entry_t submitq; 146 struct i915_sw_dma_fence_cb dmaq; 147 }; 148 struct list_head execute_cb; 149 struct i915_sw_fence semaphore; 150 151 /* 152 * A list of everyone we wait upon, and everyone who waits upon us. 153 * Even though we will not be submitted to the hardware before the 154 * submit fence is signaled (it waits for all external events as well 155 * as our own requests), the scheduler still needs to know the 156 * dependency tree for the lifetime of the request (from execbuf 157 * to retirement), i.e. bidirectional dependency information for the 158 * request not tied to individual fences. 159 */ 160 struct i915_sched_node sched; 161 struct i915_dependency dep; 162 intel_engine_mask_t execution_mask; 163 164 /* 165 * A convenience pointer to the current breadcrumb value stored in 166 * the HW status page (or our timeline's local equivalent). The full 167 * path would be rq->hw_context->ring->timeline->hwsp_seqno. 168 */ 169 const u32 *hwsp_seqno; 170 171 /* 172 * If we need to access the timeline's seqno for this request in 173 * another request, we need to keep a read reference to this associated 174 * cacheline, so that we do not free and recycle it before the foreign 175 * observers have completed. Hence, we keep a pointer to the cacheline 176 * inside the timeline's HWSP vma, but it is only valid while this 177 * request has not completed and guarded by the timeline mutex. 178 */ 179 struct intel_timeline_cacheline *hwsp_cacheline; 180 181 /** Position in the ring of the start of the request */ 182 u32 head; 183 184 /** Position in the ring of the start of the user packets */ 185 u32 infix; 186 187 /** 188 * Position in the ring of the start of the postfix. 189 * This is required to calculate the maximum available ring space 190 * without overwriting the postfix. 191 */ 192 u32 postfix; 193 194 /** Position in the ring of the end of the whole request */ 195 u32 tail; 196 197 /** Position in the ring of the end of any workarounds after the tail */ 198 u32 wa_tail; 199 200 /** Preallocate space in the ring for the emitting the request */ 201 u32 reserved_space; 202 203 /** Batch buffer related to this request if any (used for 204 * error state dump only). 205 */ 206 struct i915_vma *batch; 207 /** 208 * Additional buffers requested by userspace to be captured upon 209 * a GPU hang. The vma/obj on this list are protected by their 210 * active reference - all objects on this list must also be 211 * on the active_list (of their final request). 212 */ 213 struct i915_capture_list *capture_list; 214 215 /** Time at which this request was emitted, in jiffies. */ 216 unsigned long emitted_jiffies; 217 218 unsigned long flags; 219 #define I915_REQUEST_WAITBOOST BIT(0) 220 #define I915_REQUEST_NOPREEMPT BIT(1) 221 222 /** timeline->request entry for this request */ 223 struct list_head link; 224 225 struct drm_i915_file_private *file_priv; 226 /** file_priv list entry for this request */ 227 struct list_head client_link; 228 229 I915_SELFTEST_DECLARE(struct { 230 struct list_head link; 231 unsigned long delay; 232 } mock;) 233 }; 234 235 #define I915_FENCE_GFP (GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN) 236 237 extern const struct dma_fence_ops i915_fence_ops; 238 239 static inline bool dma_fence_is_i915(const struct dma_fence *fence) 240 { 241 return fence->ops == &i915_fence_ops; 242 } 243 244 struct i915_request * __must_check 245 __i915_request_create(struct intel_context *ce, gfp_t gfp); 246 struct i915_request * __must_check 247 i915_request_create(struct intel_context *ce); 248 249 struct i915_request *__i915_request_commit(struct i915_request *request); 250 void __i915_request_queue(struct i915_request *rq, 251 const struct i915_sched_attr *attr); 252 253 bool i915_request_retire(struct i915_request *rq); 254 void i915_request_retire_upto(struct i915_request *rq); 255 256 static inline struct i915_request * 257 to_request(struct dma_fence *fence) 258 { 259 /* We assume that NULL fence/request are interoperable */ 260 BUILD_BUG_ON(offsetof(struct i915_request, fence) != 0); 261 GEM_BUG_ON(fence && !dma_fence_is_i915(fence)); 262 return container_of(fence, struct i915_request, fence); 263 } 264 265 static inline struct i915_request * 266 i915_request_get(struct i915_request *rq) 267 { 268 return to_request(dma_fence_get(&rq->fence)); 269 } 270 271 static inline struct i915_request * 272 i915_request_get_rcu(struct i915_request *rq) 273 { 274 return to_request(dma_fence_get_rcu(&rq->fence)); 275 } 276 277 static inline void 278 i915_request_put(struct i915_request *rq) 279 { 280 dma_fence_put(&rq->fence); 281 } 282 283 int i915_request_await_object(struct i915_request *to, 284 struct drm_i915_gem_object *obj, 285 bool write); 286 int i915_request_await_dma_fence(struct i915_request *rq, 287 struct dma_fence *fence); 288 int i915_request_await_execution(struct i915_request *rq, 289 struct dma_fence *fence, 290 void (*hook)(struct i915_request *rq, 291 struct dma_fence *signal)); 292 293 void i915_request_add(struct i915_request *rq); 294 295 bool __i915_request_submit(struct i915_request *request); 296 void i915_request_submit(struct i915_request *request); 297 298 void i915_request_skip(struct i915_request *request, int error); 299 300 void __i915_request_unsubmit(struct i915_request *request); 301 void i915_request_unsubmit(struct i915_request *request); 302 303 /* Note: part of the intel_breadcrumbs family */ 304 bool i915_request_enable_breadcrumb(struct i915_request *request); 305 void i915_request_cancel_breadcrumb(struct i915_request *request); 306 307 long i915_request_wait(struct i915_request *rq, 308 unsigned int flags, 309 long timeout) 310 __attribute__((nonnull(1))); 311 #define I915_WAIT_INTERRUPTIBLE BIT(0) 312 #define I915_WAIT_PRIORITY BIT(1) /* small priority bump for the request */ 313 #define I915_WAIT_ALL BIT(2) /* used by i915_gem_object_wait() */ 314 315 static inline bool i915_request_signaled(const struct i915_request *rq) 316 { 317 /* The request may live longer than its HWSP, so check flags first! */ 318 return test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &rq->fence.flags); 319 } 320 321 static inline bool i915_request_is_active(const struct i915_request *rq) 322 { 323 return test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags); 324 } 325 326 /** 327 * Returns true if seq1 is later than seq2. 328 */ 329 static inline bool i915_seqno_passed(u32 seq1, u32 seq2) 330 { 331 return (s32)(seq1 - seq2) >= 0; 332 } 333 334 static inline u32 __hwsp_seqno(const struct i915_request *rq) 335 { 336 return READ_ONCE(*rq->hwsp_seqno); 337 } 338 339 /** 340 * hwsp_seqno - the current breadcrumb value in the HW status page 341 * @rq: the request, to chase the relevant HW status page 342 * 343 * The emphasis in naming here is that hwsp_seqno() is not a property of the 344 * request, but an indication of the current HW state (associated with this 345 * request). Its value will change as the GPU executes more requests. 346 * 347 * Returns the current breadcrumb value in the associated HW status page (or 348 * the local timeline's equivalent) for this request. The request itself 349 * has the associated breadcrumb value of rq->fence.seqno, when the HW 350 * status page has that breadcrumb or later, this request is complete. 351 */ 352 static inline u32 hwsp_seqno(const struct i915_request *rq) 353 { 354 u32 seqno; 355 356 rcu_read_lock(); /* the HWSP may be freed at runtime */ 357 seqno = __hwsp_seqno(rq); 358 rcu_read_unlock(); 359 360 return seqno; 361 } 362 363 static inline bool __i915_request_has_started(const struct i915_request *rq) 364 { 365 return i915_seqno_passed(hwsp_seqno(rq), rq->fence.seqno - 1); 366 } 367 368 /** 369 * i915_request_started - check if the request has begun being executed 370 * @rq: the request 371 * 372 * If the timeline is not using initial breadcrumbs, a request is 373 * considered started if the previous request on its timeline (i.e. 374 * context) has been signaled. 375 * 376 * If the timeline is using semaphores, it will also be emitting an 377 * "initial breadcrumb" after the semaphores are complete and just before 378 * it began executing the user payload. A request can therefore be active 379 * on the HW and not yet started as it is still busywaiting on its 380 * dependencies (via HW semaphores). 381 * 382 * If the request has started, its dependencies will have been signaled 383 * (either by fences or by semaphores) and it will have begun processing 384 * the user payload. 385 * 386 * However, even if a request has started, it may have been preempted and 387 * so no longer active, or it may have already completed. 388 * 389 * See also i915_request_is_active(). 390 * 391 * Returns true if the request has begun executing the user payload, or 392 * has completed: 393 */ 394 static inline bool i915_request_started(const struct i915_request *rq) 395 { 396 if (i915_request_signaled(rq)) 397 return true; 398 399 /* Remember: started but may have since been preempted! */ 400 return __i915_request_has_started(rq); 401 } 402 403 /** 404 * i915_request_is_running - check if the request may actually be executing 405 * @rq: the request 406 * 407 * Returns true if the request is currently submitted to hardware, has passed 408 * its start point (i.e. the context is setup and not busywaiting). Note that 409 * it may no longer be running by the time the function returns! 410 */ 411 static inline bool i915_request_is_running(const struct i915_request *rq) 412 { 413 if (!i915_request_is_active(rq)) 414 return false; 415 416 return __i915_request_has_started(rq); 417 } 418 419 static inline bool i915_request_completed(const struct i915_request *rq) 420 { 421 if (i915_request_signaled(rq)) 422 return true; 423 424 return i915_seqno_passed(hwsp_seqno(rq), rq->fence.seqno); 425 } 426 427 static inline void i915_request_mark_complete(struct i915_request *rq) 428 { 429 rq->hwsp_seqno = (u32 *)&rq->fence.seqno; /* decouple from HWSP */ 430 } 431 432 static inline bool i915_request_has_waitboost(const struct i915_request *rq) 433 { 434 return rq->flags & I915_REQUEST_WAITBOOST; 435 } 436 437 static inline bool i915_request_has_nopreempt(const struct i915_request *rq) 438 { 439 /* Preemption should only be disabled very rarely */ 440 return unlikely(rq->flags & I915_REQUEST_NOPREEMPT); 441 } 442 443 static inline struct intel_timeline * 444 i915_request_timeline(struct i915_request *rq) 445 { 446 /* Valid only while the request is being constructed (or retired). */ 447 return rcu_dereference_protected(rq->timeline, 448 lockdep_is_held(&rcu_access_pointer(rq->timeline)->mutex)); 449 } 450 451 static inline struct intel_timeline * 452 i915_request_active_timeline(struct i915_request *rq) 453 { 454 /* 455 * When in use during submission, we are protected by a guarantee that 456 * the context/timeline is pinned and must remain pinned until after 457 * this submission. 458 */ 459 return rcu_dereference_protected(rq->timeline, 460 lockdep_is_held(&rq->engine->active.lock)); 461 } 462 463 #endif /* I915_REQUEST_H */ 464