| 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/hrtimer.h> | 
|---|
| 30 | #include <linux/irq_work.h> | 
|---|
| 31 | #include <linux/llist.h> | 
|---|
| 32 | #include <linux/lockdep.h> | 
|---|
| 33 |  | 
|---|
| 34 | #include <uapi/drm/i915_drm.h> | 
|---|
| 35 |  | 
|---|
| 36 | #include "gem/i915_gem_context_types.h" | 
|---|
| 37 | #include "gt/intel_context_types.h" | 
|---|
| 38 | #include "gt/intel_engine_types.h" | 
|---|
| 39 | #include "gt/intel_timeline_types.h" | 
|---|
| 40 |  | 
|---|
| 41 | #include "i915_gem.h" | 
|---|
| 42 | #include "i915_ptr_util.h" | 
|---|
| 43 | #include "i915_scheduler.h" | 
|---|
| 44 | #include "i915_selftest.h" | 
|---|
| 45 | #include "i915_sw_fence.h" | 
|---|
| 46 | #include "i915_vma_resource.h" | 
|---|
| 47 |  | 
|---|
| 48 | struct drm_file; | 
|---|
| 49 | struct drm_i915_gem_object; | 
|---|
| 50 | struct drm_printer; | 
|---|
| 51 | struct i915_deps; | 
|---|
| 52 | struct i915_request; | 
|---|
| 53 |  | 
|---|
| 54 | #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR) | 
|---|
| 55 | struct i915_capture_list { | 
|---|
| 56 | struct i915_vma_resource *vma_res; | 
|---|
| 57 | struct i915_capture_list *next; | 
|---|
| 58 | }; | 
|---|
| 59 |  | 
|---|
| 60 | void i915_request_free_capture_list(struct i915_capture_list *capture); | 
|---|
| 61 | #else | 
|---|
| 62 | #define i915_request_free_capture_list(_a) do {} while (0) | 
|---|
| 63 | #endif | 
|---|
| 64 |  | 
|---|
| 65 | #define RQ_TRACE(rq, fmt, ...) do {					\ | 
|---|
| 66 | const struct i915_request *rq__ = (rq);				\ | 
|---|
| 67 | ENGINE_TRACE(rq__->engine, "fence %llx:%lld, current %d " fmt,	\ | 
|---|
| 68 | rq__->fence.context, rq__->fence.seqno,		\ | 
|---|
| 69 | hwsp_seqno(rq__), ##__VA_ARGS__);			\ | 
|---|
| 70 | } while (0) | 
|---|
| 71 |  | 
|---|
| 72 | enum { | 
|---|
| 73 | /* | 
|---|
| 74 | * I915_FENCE_FLAG_ACTIVE - this request is currently submitted to HW. | 
|---|
| 75 | * | 
|---|
| 76 | * Set by __i915_request_submit() on handing over to HW, and cleared | 
|---|
| 77 | * by __i915_request_unsubmit() if we preempt this request. | 
|---|
| 78 | * | 
|---|
| 79 | * Finally cleared for consistency on retiring the request, when | 
|---|
| 80 | * we know the HW is no longer running this request. | 
|---|
| 81 | * | 
|---|
| 82 | * See i915_request_is_active() | 
|---|
| 83 | */ | 
|---|
| 84 | I915_FENCE_FLAG_ACTIVE = DMA_FENCE_FLAG_USER_BITS, | 
|---|
| 85 |  | 
|---|
| 86 | /* | 
|---|
| 87 | * I915_FENCE_FLAG_PQUEUE - this request is ready for execution | 
|---|
| 88 | * | 
|---|
| 89 | * Using the scheduler, when a request is ready for execution it is put | 
|---|
| 90 | * into the priority queue, and removed from that queue when transferred | 
|---|
| 91 | * to the HW runlists. We want to track its membership within the | 
|---|
| 92 | * priority queue so that we can easily check before rescheduling. | 
|---|
| 93 | * | 
|---|
| 94 | * See i915_request_in_priority_queue() | 
|---|
| 95 | */ | 
|---|
| 96 | I915_FENCE_FLAG_PQUEUE, | 
|---|
| 97 |  | 
|---|
| 98 | /* | 
|---|
| 99 | * I915_FENCE_FLAG_HOLD - this request is currently on hold | 
|---|
| 100 | * | 
|---|
| 101 | * This request has been suspended, pending an ongoing investigation. | 
|---|
| 102 | */ | 
|---|
| 103 | I915_FENCE_FLAG_HOLD, | 
|---|
| 104 |  | 
|---|
| 105 | /* | 
|---|
| 106 | * I915_FENCE_FLAG_INITIAL_BREADCRUMB - this request has the initial | 
|---|
| 107 | * breadcrumb that marks the end of semaphore waits and start of the | 
|---|
| 108 | * user payload. | 
|---|
| 109 | */ | 
|---|
| 110 | I915_FENCE_FLAG_INITIAL_BREADCRUMB, | 
|---|
| 111 |  | 
|---|
| 112 | /* | 
|---|
| 113 | * I915_FENCE_FLAG_SIGNAL - this request is currently on signal_list | 
|---|
| 114 | * | 
|---|
| 115 | * Internal bookkeeping used by the breadcrumb code to track when | 
|---|
| 116 | * a request is on the various signal_list. | 
|---|
| 117 | */ | 
|---|
| 118 | I915_FENCE_FLAG_SIGNAL, | 
|---|
| 119 |  | 
|---|
| 120 | /* | 
|---|
| 121 | * I915_FENCE_FLAG_NOPREEMPT - this request should not be preempted | 
|---|
| 122 | * | 
|---|
| 123 | * The execution of some requests should not be interrupted. This is | 
|---|
| 124 | * a sensitive operation as it makes the request super important, | 
|---|
| 125 | * blocking other higher priority work. Abuse of this flag will | 
|---|
| 126 | * lead to quality of service issues. | 
|---|
| 127 | */ | 
|---|
| 128 | I915_FENCE_FLAG_NOPREEMPT, | 
|---|
| 129 |  | 
|---|
| 130 | /* | 
|---|
| 131 | * I915_FENCE_FLAG_SENTINEL - this request should be last in the queue | 
|---|
| 132 | * | 
|---|
| 133 | * A high priority sentinel request may be submitted to clear the | 
|---|
| 134 | * submission queue. As it will be the only request in-flight, upon | 
|---|
| 135 | * execution all other active requests will have been preempted and | 
|---|
| 136 | * unsubmitted. This preemptive pulse is used to re-evaluate the | 
|---|
| 137 | * in-flight requests, particularly in cases where an active context | 
|---|
| 138 | * is banned and those active requests need to be cancelled. | 
|---|
| 139 | */ | 
|---|
| 140 | I915_FENCE_FLAG_SENTINEL, | 
|---|
| 141 |  | 
|---|
| 142 | /* | 
|---|
| 143 | * I915_FENCE_FLAG_BOOST - upclock the gpu for this request | 
|---|
| 144 | * | 
|---|
| 145 | * Some requests are more important than others! In particular, a | 
|---|
| 146 | * request that the user is waiting on is typically required for | 
|---|
| 147 | * interactive latency, for which we want to minimise by upclocking | 
|---|
| 148 | * the GPU. Here we track such boost requests on a per-request basis. | 
|---|
| 149 | */ | 
|---|
| 150 | I915_FENCE_FLAG_BOOST, | 
|---|
| 151 |  | 
|---|
| 152 | /* | 
|---|
| 153 | * I915_FENCE_FLAG_SUBMIT_PARALLEL - request with a context in a | 
|---|
| 154 | * parent-child relationship (parallel submission, multi-lrc) should | 
|---|
| 155 | * trigger a submission to the GuC rather than just moving the context | 
|---|
| 156 | * tail. | 
|---|
| 157 | */ | 
|---|
| 158 | I915_FENCE_FLAG_SUBMIT_PARALLEL, | 
|---|
| 159 |  | 
|---|
| 160 | /* | 
|---|
| 161 | * I915_FENCE_FLAG_SKIP_PARALLEL - request with a context in a | 
|---|
| 162 | * parent-child relationship (parallel submission, multi-lrc) that | 
|---|
| 163 | * hit an error while generating requests in the execbuf IOCTL. | 
|---|
| 164 | * Indicates this request should be skipped as another request in | 
|---|
| 165 | * submission / relationship encountered an error. | 
|---|
| 166 | */ | 
|---|
| 167 | I915_FENCE_FLAG_SKIP_PARALLEL, | 
|---|
| 168 |  | 
|---|
| 169 | /* | 
|---|
| 170 | * I915_FENCE_FLAG_COMPOSITE - Indicates fence is part of a composite | 
|---|
| 171 | * fence (dma_fence_array) and i915 generated for parallel submission. | 
|---|
| 172 | */ | 
|---|
| 173 | I915_FENCE_FLAG_COMPOSITE, | 
|---|
| 174 | }; | 
|---|
| 175 |  | 
|---|
| 176 | /* | 
|---|
| 177 | * Request queue structure. | 
|---|
| 178 | * | 
|---|
| 179 | * The request queue allows us to note sequence numbers that have been emitted | 
|---|
| 180 | * and may be associated with active buffers to be retired. | 
|---|
| 181 | * | 
|---|
| 182 | * By keeping this list, we can avoid having to do questionable sequence | 
|---|
| 183 | * number comparisons on buffer last_read|write_seqno. It also allows an | 
|---|
| 184 | * emission time to be associated with the request for tracking how far ahead | 
|---|
| 185 | * of the GPU the submission is. | 
|---|
| 186 | * | 
|---|
| 187 | * When modifying this structure be very aware that we perform a lockless | 
|---|
| 188 | * RCU lookup of it that may race against reallocation of the struct | 
|---|
| 189 | * from the slab freelist. We intentionally do not zero the structure on | 
|---|
| 190 | * allocation so that the lookup can use the dangling pointers (and is | 
|---|
| 191 | * cognisant that those pointers may be wrong). Instead, everything that | 
|---|
| 192 | * needs to be initialised must be done so explicitly. | 
|---|
| 193 | * | 
|---|
| 194 | * The requests are reference counted. | 
|---|
| 195 | */ | 
|---|
| 196 | struct i915_request { | 
|---|
| 197 | struct dma_fence fence; | 
|---|
| 198 | spinlock_t lock; | 
|---|
| 199 |  | 
|---|
| 200 | struct drm_i915_private *i915; | 
|---|
| 201 |  | 
|---|
| 202 | /* | 
|---|
| 203 | * Context and ring buffer related to this request | 
|---|
| 204 | * Contexts are refcounted, so when this request is associated with a | 
|---|
| 205 | * context, we must increment the context's refcount, to guarantee that | 
|---|
| 206 | * it persists while any request is linked to it. Requests themselves | 
|---|
| 207 | * are also refcounted, so the request will only be freed when the last | 
|---|
| 208 | * reference to it is dismissed, and the code in | 
|---|
| 209 | * i915_request_free() will then decrement the refcount on the | 
|---|
| 210 | * context. | 
|---|
| 211 | */ | 
|---|
| 212 | struct intel_engine_cs *engine; | 
|---|
| 213 | struct intel_context *context; | 
|---|
| 214 | struct intel_ring *ring; | 
|---|
| 215 | struct intel_timeline __rcu *timeline; | 
|---|
| 216 |  | 
|---|
| 217 | struct list_head signal_link; | 
|---|
| 218 | struct llist_node signal_node; | 
|---|
| 219 |  | 
|---|
| 220 | /* | 
|---|
| 221 | * The rcu epoch of when this request was allocated. Used to judiciously | 
|---|
| 222 | * apply backpressure on future allocations to ensure that under | 
|---|
| 223 | * mempressure there is sufficient RCU ticks for us to reclaim our | 
|---|
| 224 | * RCU protected slabs. | 
|---|
| 225 | */ | 
|---|
| 226 | unsigned long rcustate; | 
|---|
| 227 |  | 
|---|
| 228 | /* | 
|---|
| 229 | * We pin the timeline->mutex while constructing the request to | 
|---|
| 230 | * ensure that no caller accidentally drops it during construction. | 
|---|
| 231 | * The timeline->mutex must be held to ensure that only this caller | 
|---|
| 232 | * can use the ring and manipulate the associated timeline during | 
|---|
| 233 | * construction. | 
|---|
| 234 | */ | 
|---|
| 235 | struct pin_cookie cookie; | 
|---|
| 236 |  | 
|---|
| 237 | /* | 
|---|
| 238 | * Fences for the various phases in the request's lifetime. | 
|---|
| 239 | * | 
|---|
| 240 | * The submit fence is used to await upon all of the request's | 
|---|
| 241 | * dependencies. When it is signaled, the request is ready to run. | 
|---|
| 242 | * It is used by the driver to then queue the request for execution. | 
|---|
| 243 | */ | 
|---|
| 244 | struct i915_sw_fence submit; | 
|---|
| 245 | union { | 
|---|
| 246 | wait_queue_entry_t submitq; | 
|---|
| 247 | struct i915_sw_dma_fence_cb dmaq; | 
|---|
| 248 | struct i915_request_duration_cb { | 
|---|
| 249 | struct dma_fence_cb cb; | 
|---|
| 250 | ktime_t emitted; | 
|---|
| 251 | } duration; | 
|---|
| 252 | }; | 
|---|
| 253 | struct llist_head execute_cb; | 
|---|
| 254 | struct i915_sw_fence semaphore; | 
|---|
| 255 | /* | 
|---|
| 256 | * complete submit fence from an IRQ if needed for locking hierarchy | 
|---|
| 257 | * reasons. | 
|---|
| 258 | */ | 
|---|
| 259 | struct irq_work submit_work; | 
|---|
| 260 |  | 
|---|
| 261 | /* | 
|---|
| 262 | * A list of everyone we wait upon, and everyone who waits upon us. | 
|---|
| 263 | * Even though we will not be submitted to the hardware before the | 
|---|
| 264 | * submit fence is signaled (it waits for all external events as well | 
|---|
| 265 | * as our own requests), the scheduler still needs to know the | 
|---|
| 266 | * dependency tree for the lifetime of the request (from execbuf | 
|---|
| 267 | * to retirement), i.e. bidirectional dependency information for the | 
|---|
| 268 | * request not tied to individual fences. | 
|---|
| 269 | */ | 
|---|
| 270 | struct i915_sched_node sched; | 
|---|
| 271 | struct i915_dependency dep; | 
|---|
| 272 | intel_engine_mask_t execution_mask; | 
|---|
| 273 |  | 
|---|
| 274 | /* | 
|---|
| 275 | * A convenience pointer to the current breadcrumb value stored in | 
|---|
| 276 | * the HW status page (or our timeline's local equivalent). The full | 
|---|
| 277 | * path would be rq->hw_context->ring->timeline->hwsp_seqno. | 
|---|
| 278 | */ | 
|---|
| 279 | const u32 *hwsp_seqno; | 
|---|
| 280 |  | 
|---|
| 281 | /* Position in the ring of the start of the request */ | 
|---|
| 282 | u32 head; | 
|---|
| 283 |  | 
|---|
| 284 | /* Position in the ring of the start of the user packets */ | 
|---|
| 285 | u32 infix; | 
|---|
| 286 |  | 
|---|
| 287 | /* | 
|---|
| 288 | * Position in the ring of the start of the postfix. | 
|---|
| 289 | * This is required to calculate the maximum available ring space | 
|---|
| 290 | * without overwriting the postfix. | 
|---|
| 291 | */ | 
|---|
| 292 | u32 postfix; | 
|---|
| 293 |  | 
|---|
| 294 | /* Position in the ring of the end of the whole request */ | 
|---|
| 295 | u32 tail; | 
|---|
| 296 |  | 
|---|
| 297 | /* Position in the ring of the end of any workarounds after the tail */ | 
|---|
| 298 | u32 wa_tail; | 
|---|
| 299 |  | 
|---|
| 300 | /* Preallocate space in the ring for the emitting the request */ | 
|---|
| 301 | u32 reserved_space; | 
|---|
| 302 |  | 
|---|
| 303 | /* Batch buffer pointer for selftest internal use. */ | 
|---|
| 304 | I915_SELFTEST_DECLARE(struct i915_vma *batch); | 
|---|
| 305 |  | 
|---|
| 306 | struct i915_vma_resource *batch_res; | 
|---|
| 307 |  | 
|---|
| 308 | #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR) | 
|---|
| 309 | /* | 
|---|
| 310 | * Additional buffers requested by userspace to be captured upon | 
|---|
| 311 | * a GPU hang. The vma/obj on this list are protected by their | 
|---|
| 312 | * active reference - all objects on this list must also be | 
|---|
| 313 | * on the active_list (of their final request). | 
|---|
| 314 | */ | 
|---|
| 315 | struct i915_capture_list *capture_list; | 
|---|
| 316 | #endif | 
|---|
| 317 |  | 
|---|
| 318 | /* Time at which this request was emitted, in jiffies. */ | 
|---|
| 319 | unsigned long emitted_jiffies; | 
|---|
| 320 |  | 
|---|
| 321 | /* timeline->request entry for this request */ | 
|---|
| 322 | struct list_head link; | 
|---|
| 323 |  | 
|---|
| 324 | /* Watchdog support fields. */ | 
|---|
| 325 | struct i915_request_watchdog { | 
|---|
| 326 | struct llist_node link; | 
|---|
| 327 | struct hrtimer timer; | 
|---|
| 328 | } watchdog; | 
|---|
| 329 |  | 
|---|
| 330 | /* | 
|---|
| 331 | * Requests may need to be stalled when using GuC submission waiting for | 
|---|
| 332 | * certain GuC operations to complete. If that is the case, stalled | 
|---|
| 333 | * requests are added to a per context list of stalled requests. The | 
|---|
| 334 | * below list_head is the link in that list. Protected by | 
|---|
| 335 | * ce->guc_state.lock. | 
|---|
| 336 | */ | 
|---|
| 337 | struct list_head guc_fence_link; | 
|---|
| 338 |  | 
|---|
| 339 | /* | 
|---|
| 340 | * Priority level while the request is in flight. Differs | 
|---|
| 341 | * from i915 scheduler priority. See comment above | 
|---|
| 342 | * I915_SCHEDULER_CAP_STATIC_PRIORITY_MAP for details. Protected by | 
|---|
| 343 | * ce->guc_active.lock. Two special values (GUC_PRIO_INIT and | 
|---|
| 344 | * GUC_PRIO_FINI) outside the GuC priority range are used to indicate | 
|---|
| 345 | * if the priority has not been initialized yet or if no more updates | 
|---|
| 346 | * are possible because the request has completed. | 
|---|
| 347 | */ | 
|---|
| 348 | #define	GUC_PRIO_INIT	0xff | 
|---|
| 349 | #define	GUC_PRIO_FINI	0xfe | 
|---|
| 350 | u8 guc_prio; | 
|---|
| 351 |  | 
|---|
| 352 | /* | 
|---|
| 353 | * wait queue entry used to wait on the HuC load to complete | 
|---|
| 354 | */ | 
|---|
| 355 | wait_queue_entry_t hucq; | 
|---|
| 356 |  | 
|---|
| 357 | I915_SELFTEST_DECLARE(struct { | 
|---|
| 358 | struct list_head link; | 
|---|
| 359 | unsigned long delay; | 
|---|
| 360 | } mock;) | 
|---|
| 361 | }; | 
|---|
| 362 |  | 
|---|
| 363 | #define I915_FENCE_GFP (GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN) | 
|---|
| 364 |  | 
|---|
| 365 | extern const struct dma_fence_ops i915_fence_ops; | 
|---|
| 366 |  | 
|---|
| 367 | static inline bool dma_fence_is_i915(const struct dma_fence *fence) | 
|---|
| 368 | { | 
|---|
| 369 | return fence->ops == &i915_fence_ops; | 
|---|
| 370 | } | 
|---|
| 371 |  | 
|---|
| 372 | struct kmem_cache *i915_request_slab_cache(void); | 
|---|
| 373 |  | 
|---|
| 374 | struct i915_request * __must_check | 
|---|
| 375 | __i915_request_create(struct intel_context *ce, gfp_t gfp); | 
|---|
| 376 | struct i915_request * __must_check | 
|---|
| 377 | i915_request_create(struct intel_context *ce); | 
|---|
| 378 |  | 
|---|
| 379 | void __i915_request_skip(struct i915_request *rq); | 
|---|
| 380 | bool i915_request_set_error_once(struct i915_request *rq, int error); | 
|---|
| 381 | struct i915_request *i915_request_mark_eio(struct i915_request *rq); | 
|---|
| 382 |  | 
|---|
| 383 | struct i915_request *__i915_request_commit(struct i915_request *request); | 
|---|
| 384 | void __i915_request_queue(struct i915_request *rq, | 
|---|
| 385 | const struct i915_sched_attr *attr); | 
|---|
| 386 | void __i915_request_queue_bh(struct i915_request *rq); | 
|---|
| 387 |  | 
|---|
| 388 | bool i915_request_retire(struct i915_request *rq); | 
|---|
| 389 | void i915_request_retire_upto(struct i915_request *rq); | 
|---|
| 390 |  | 
|---|
| 391 | static inline struct i915_request * | 
|---|
| 392 | to_request(struct dma_fence *fence) | 
|---|
| 393 | { | 
|---|
| 394 | /* We assume that NULL fence/request are interoperable */ | 
|---|
| 395 | BUILD_BUG_ON(offsetof(struct i915_request, fence) != 0); | 
|---|
| 396 | GEM_BUG_ON(fence && !dma_fence_is_i915(fence)); | 
|---|
| 397 | return container_of(fence, struct i915_request, fence); | 
|---|
| 398 | } | 
|---|
| 399 |  | 
|---|
| 400 | static inline struct i915_request * | 
|---|
| 401 | i915_request_get(struct i915_request *rq) | 
|---|
| 402 | { | 
|---|
| 403 | return to_request(fence: dma_fence_get(fence: &rq->fence)); | 
|---|
| 404 | } | 
|---|
| 405 |  | 
|---|
| 406 | static inline struct i915_request * | 
|---|
| 407 | i915_request_get_rcu(struct i915_request *rq) | 
|---|
| 408 | { | 
|---|
| 409 | return to_request(fence: dma_fence_get_rcu(fence: &rq->fence)); | 
|---|
| 410 | } | 
|---|
| 411 |  | 
|---|
| 412 | static inline void | 
|---|
| 413 | i915_request_put(struct i915_request *rq) | 
|---|
| 414 | { | 
|---|
| 415 | dma_fence_put(fence: &rq->fence); | 
|---|
| 416 | } | 
|---|
| 417 |  | 
|---|
| 418 | int i915_request_await_object(struct i915_request *to, | 
|---|
| 419 | struct drm_i915_gem_object *obj, | 
|---|
| 420 | bool write); | 
|---|
| 421 | int i915_request_await_dma_fence(struct i915_request *rq, | 
|---|
| 422 | struct dma_fence *fence); | 
|---|
| 423 | int i915_request_await_deps(struct i915_request *rq, const struct i915_deps *deps); | 
|---|
| 424 | int i915_request_await_execution(struct i915_request *rq, | 
|---|
| 425 | struct dma_fence *fence); | 
|---|
| 426 |  | 
|---|
| 427 | void i915_request_add(struct i915_request *rq); | 
|---|
| 428 |  | 
|---|
| 429 | bool __i915_request_submit(struct i915_request *request); | 
|---|
| 430 | void i915_request_submit(struct i915_request *request); | 
|---|
| 431 |  | 
|---|
| 432 | void __i915_request_unsubmit(struct i915_request *request); | 
|---|
| 433 | void i915_request_unsubmit(struct i915_request *request); | 
|---|
| 434 |  | 
|---|
| 435 | void i915_request_cancel(struct i915_request *rq, int error); | 
|---|
| 436 |  | 
|---|
| 437 | long i915_request_wait_timeout(struct i915_request *rq, | 
|---|
| 438 | unsigned int flags, | 
|---|
| 439 | long timeout) | 
|---|
| 440 | __attribute__((nonnull(1))); | 
|---|
| 441 |  | 
|---|
| 442 | long i915_request_wait(struct i915_request *rq, | 
|---|
| 443 | unsigned int flags, | 
|---|
| 444 | long timeout) | 
|---|
| 445 | __attribute__((nonnull(1))); | 
|---|
| 446 | #define I915_WAIT_INTERRUPTIBLE	BIT(0) | 
|---|
| 447 | #define I915_WAIT_PRIORITY	BIT(1) /* small priority bump for the request */ | 
|---|
| 448 | #define I915_WAIT_ALL		BIT(2) /* used by i915_gem_object_wait() */ | 
|---|
| 449 |  | 
|---|
| 450 | void i915_request_show(struct drm_printer *m, | 
|---|
| 451 | const struct i915_request *rq, | 
|---|
| 452 | const char *prefix, | 
|---|
| 453 | int indent); | 
|---|
| 454 |  | 
|---|
| 455 | static inline bool i915_request_signaled(const struct i915_request *rq) | 
|---|
| 456 | { | 
|---|
| 457 | /* The request may live longer than its HWSP, so check flags first! */ | 
|---|
| 458 | return test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &rq->fence.flags); | 
|---|
| 459 | } | 
|---|
| 460 |  | 
|---|
| 461 | static inline bool i915_request_is_active(const struct i915_request *rq) | 
|---|
| 462 | { | 
|---|
| 463 | return test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags); | 
|---|
| 464 | } | 
|---|
| 465 |  | 
|---|
| 466 | static inline bool i915_request_in_priority_queue(const struct i915_request *rq) | 
|---|
| 467 | { | 
|---|
| 468 | return test_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags); | 
|---|
| 469 | } | 
|---|
| 470 |  | 
|---|
| 471 | static inline bool | 
|---|
| 472 | i915_request_has_initial_breadcrumb(const struct i915_request *rq) | 
|---|
| 473 | { | 
|---|
| 474 | return test_bit(I915_FENCE_FLAG_INITIAL_BREADCRUMB, &rq->fence.flags); | 
|---|
| 475 | } | 
|---|
| 476 |  | 
|---|
| 477 | /* | 
|---|
| 478 | * Returns true if seq1 is later than seq2. | 
|---|
| 479 | */ | 
|---|
| 480 | static inline bool i915_seqno_passed(u32 seq1, u32 seq2) | 
|---|
| 481 | { | 
|---|
| 482 | return (s32)(seq1 - seq2) >= 0; | 
|---|
| 483 | } | 
|---|
| 484 |  | 
|---|
| 485 | static inline u32 __hwsp_seqno(const struct i915_request *rq) | 
|---|
| 486 | { | 
|---|
| 487 | const u32 *hwsp = READ_ONCE(rq->hwsp_seqno); | 
|---|
| 488 |  | 
|---|
| 489 | return READ_ONCE(*hwsp); | 
|---|
| 490 | } | 
|---|
| 491 |  | 
|---|
| 492 | /** | 
|---|
| 493 | * hwsp_seqno - the current breadcrumb value in the HW status page | 
|---|
| 494 | * @rq: the request, to chase the relevant HW status page | 
|---|
| 495 | * | 
|---|
| 496 | * The emphasis in naming here is that hwsp_seqno() is not a property of the | 
|---|
| 497 | * request, but an indication of the current HW state (associated with this | 
|---|
| 498 | * request). Its value will change as the GPU executes more requests. | 
|---|
| 499 | * | 
|---|
| 500 | * Returns the current breadcrumb value in the associated HW status page (or | 
|---|
| 501 | * the local timeline's equivalent) for this request. The request itself | 
|---|
| 502 | * has the associated breadcrumb value of rq->fence.seqno, when the HW | 
|---|
| 503 | * status page has that breadcrumb or later, this request is complete. | 
|---|
| 504 | */ | 
|---|
| 505 | static inline u32 hwsp_seqno(const struct i915_request *rq) | 
|---|
| 506 | { | 
|---|
| 507 | u32 seqno; | 
|---|
| 508 |  | 
|---|
| 509 | rcu_read_lock(); /* the HWSP may be freed at runtime */ | 
|---|
| 510 | seqno = __hwsp_seqno(rq); | 
|---|
| 511 | rcu_read_unlock(); | 
|---|
| 512 |  | 
|---|
| 513 | return seqno; | 
|---|
| 514 | } | 
|---|
| 515 |  | 
|---|
| 516 | static inline bool __i915_request_has_started(const struct i915_request *rq) | 
|---|
| 517 | { | 
|---|
| 518 | return i915_seqno_passed(seq1: __hwsp_seqno(rq), seq2: rq->fence.seqno - 1); | 
|---|
| 519 | } | 
|---|
| 520 |  | 
|---|
| 521 | /** | 
|---|
| 522 | * i915_request_started - check if the request has begun being executed | 
|---|
| 523 | * @rq: the request | 
|---|
| 524 | * | 
|---|
| 525 | * If the timeline is not using initial breadcrumbs, a request is | 
|---|
| 526 | * considered started if the previous request on its timeline (i.e. | 
|---|
| 527 | * context) has been signaled. | 
|---|
| 528 | * | 
|---|
| 529 | * If the timeline is using semaphores, it will also be emitting an | 
|---|
| 530 | * "initial breadcrumb" after the semaphores are complete and just before | 
|---|
| 531 | * it began executing the user payload. A request can therefore be active | 
|---|
| 532 | * on the HW and not yet started as it is still busywaiting on its | 
|---|
| 533 | * dependencies (via HW semaphores). | 
|---|
| 534 | * | 
|---|
| 535 | * If the request has started, its dependencies will have been signaled | 
|---|
| 536 | * (either by fences or by semaphores) and it will have begun processing | 
|---|
| 537 | * the user payload. | 
|---|
| 538 | * | 
|---|
| 539 | * However, even if a request has started, it may have been preempted and | 
|---|
| 540 | * so no longer active, or it may have already completed. | 
|---|
| 541 | * | 
|---|
| 542 | * See also i915_request_is_active(). | 
|---|
| 543 | * | 
|---|
| 544 | * Returns true if the request has begun executing the user payload, or | 
|---|
| 545 | * has completed: | 
|---|
| 546 | */ | 
|---|
| 547 | static inline bool i915_request_started(const struct i915_request *rq) | 
|---|
| 548 | { | 
|---|
| 549 | bool result; | 
|---|
| 550 |  | 
|---|
| 551 | if (i915_request_signaled(rq)) | 
|---|
| 552 | return true; | 
|---|
| 553 |  | 
|---|
| 554 | result = true; | 
|---|
| 555 | rcu_read_lock(); /* the HWSP may be freed at runtime */ | 
|---|
| 556 | if (likely(!i915_request_signaled(rq))) | 
|---|
| 557 | /* Remember: started but may have since been preempted! */ | 
|---|
| 558 | result = __i915_request_has_started(rq); | 
|---|
| 559 | rcu_read_unlock(); | 
|---|
| 560 |  | 
|---|
| 561 | return result; | 
|---|
| 562 | } | 
|---|
| 563 |  | 
|---|
| 564 | /** | 
|---|
| 565 | * i915_request_is_running - check if the request may actually be executing | 
|---|
| 566 | * @rq: the request | 
|---|
| 567 | * | 
|---|
| 568 | * Returns true if the request is currently submitted to hardware, has passed | 
|---|
| 569 | * its start point (i.e. the context is setup and not busywaiting). Note that | 
|---|
| 570 | * it may no longer be running by the time the function returns! | 
|---|
| 571 | */ | 
|---|
| 572 | static inline bool i915_request_is_running(const struct i915_request *rq) | 
|---|
| 573 | { | 
|---|
| 574 | bool result; | 
|---|
| 575 |  | 
|---|
| 576 | if (!i915_request_is_active(rq)) | 
|---|
| 577 | return false; | 
|---|
| 578 |  | 
|---|
| 579 | rcu_read_lock(); | 
|---|
| 580 | result = __i915_request_has_started(rq) && i915_request_is_active(rq); | 
|---|
| 581 | rcu_read_unlock(); | 
|---|
| 582 |  | 
|---|
| 583 | return result; | 
|---|
| 584 | } | 
|---|
| 585 |  | 
|---|
| 586 | /** | 
|---|
| 587 | * i915_request_is_ready - check if the request is ready for execution | 
|---|
| 588 | * @rq: the request | 
|---|
| 589 | * | 
|---|
| 590 | * Upon construction, the request is instructed to wait upon various | 
|---|
| 591 | * signals before it is ready to be executed by the HW. That is, we do | 
|---|
| 592 | * not want to start execution and read data before it is written. In practice, | 
|---|
| 593 | * this is controlled with a mixture of interrupts and semaphores. Once | 
|---|
| 594 | * the submit fence is completed, the backend scheduler will place the | 
|---|
| 595 | * request into its queue and from there submit it for execution. So we | 
|---|
| 596 | * can detect when a request is eligible for execution (and is under control | 
|---|
| 597 | * of the scheduler) by querying where it is in any of the scheduler's lists. | 
|---|
| 598 | * | 
|---|
| 599 | * Returns true if the request is ready for execution (it may be inflight), | 
|---|
| 600 | * false otherwise. | 
|---|
| 601 | */ | 
|---|
| 602 | static inline bool i915_request_is_ready(const struct i915_request *rq) | 
|---|
| 603 | { | 
|---|
| 604 | return !list_empty(head: &rq->sched.link); | 
|---|
| 605 | } | 
|---|
| 606 |  | 
|---|
| 607 | static inline bool __i915_request_is_complete(const struct i915_request *rq) | 
|---|
| 608 | { | 
|---|
| 609 | return i915_seqno_passed(seq1: __hwsp_seqno(rq), seq2: rq->fence.seqno); | 
|---|
| 610 | } | 
|---|
| 611 |  | 
|---|
| 612 | static inline bool i915_request_completed(const struct i915_request *rq) | 
|---|
| 613 | { | 
|---|
| 614 | bool result; | 
|---|
| 615 |  | 
|---|
| 616 | if (i915_request_signaled(rq)) | 
|---|
| 617 | return true; | 
|---|
| 618 |  | 
|---|
| 619 | result = true; | 
|---|
| 620 | rcu_read_lock(); /* the HWSP may be freed at runtime */ | 
|---|
| 621 | if (likely(!i915_request_signaled(rq))) | 
|---|
| 622 | result = __i915_request_is_complete(rq); | 
|---|
| 623 | rcu_read_unlock(); | 
|---|
| 624 |  | 
|---|
| 625 | return result; | 
|---|
| 626 | } | 
|---|
| 627 |  | 
|---|
| 628 | static inline void i915_request_mark_complete(struct i915_request *rq) | 
|---|
| 629 | { | 
|---|
| 630 | WRITE_ONCE(rq->hwsp_seqno, /* decouple from HWSP */ | 
|---|
| 631 | (u32 *)&rq->fence.seqno); | 
|---|
| 632 | } | 
|---|
| 633 |  | 
|---|
| 634 | static inline bool i915_request_has_waitboost(const struct i915_request *rq) | 
|---|
| 635 | { | 
|---|
| 636 | return test_bit(I915_FENCE_FLAG_BOOST, &rq->fence.flags); | 
|---|
| 637 | } | 
|---|
| 638 |  | 
|---|
| 639 | static inline bool i915_request_has_nopreempt(const struct i915_request *rq) | 
|---|
| 640 | { | 
|---|
| 641 | /* Preemption should only be disabled very rarely */ | 
|---|
| 642 | return unlikely(test_bit(I915_FENCE_FLAG_NOPREEMPT, &rq->fence.flags)); | 
|---|
| 643 | } | 
|---|
| 644 |  | 
|---|
| 645 | static inline bool i915_request_has_sentinel(const struct i915_request *rq) | 
|---|
| 646 | { | 
|---|
| 647 | return unlikely(test_bit(I915_FENCE_FLAG_SENTINEL, &rq->fence.flags)); | 
|---|
| 648 | } | 
|---|
| 649 |  | 
|---|
| 650 | static inline bool i915_request_on_hold(const struct i915_request *rq) | 
|---|
| 651 | { | 
|---|
| 652 | return unlikely(test_bit(I915_FENCE_FLAG_HOLD, &rq->fence.flags)); | 
|---|
| 653 | } | 
|---|
| 654 |  | 
|---|
| 655 | static inline void i915_request_set_hold(struct i915_request *rq) | 
|---|
| 656 | { | 
|---|
| 657 | set_bit(nr: I915_FENCE_FLAG_HOLD, addr: &rq->fence.flags); | 
|---|
| 658 | } | 
|---|
| 659 |  | 
|---|
| 660 | static inline void i915_request_clear_hold(struct i915_request *rq) | 
|---|
| 661 | { | 
|---|
| 662 | clear_bit(nr: I915_FENCE_FLAG_HOLD, addr: &rq->fence.flags); | 
|---|
| 663 | } | 
|---|
| 664 |  | 
|---|
| 665 | static inline struct intel_timeline * | 
|---|
| 666 | i915_request_timeline(const struct i915_request *rq) | 
|---|
| 667 | { | 
|---|
| 668 | /* Valid only while the request is being constructed (or retired). */ | 
|---|
| 669 | return rcu_dereference_protected(rq->timeline, | 
|---|
| 670 | lockdep_is_held(&rcu_access_pointer(rq->timeline)->mutex) || | 
|---|
| 671 | test_bit(CONTEXT_IS_PARKING, &rq->context->flags)); | 
|---|
| 672 | } | 
|---|
| 673 |  | 
|---|
| 674 | static inline struct i915_gem_context * | 
|---|
| 675 | i915_request_gem_context(const struct i915_request *rq) | 
|---|
| 676 | { | 
|---|
| 677 | /* Valid only while the request is being constructed (or retired). */ | 
|---|
| 678 | return rcu_dereference_protected(rq->context->gem_context, true); | 
|---|
| 679 | } | 
|---|
| 680 |  | 
|---|
| 681 | static inline struct intel_timeline * | 
|---|
| 682 | i915_request_active_timeline(const struct i915_request *rq) | 
|---|
| 683 | { | 
|---|
| 684 | /* | 
|---|
| 685 | * When in use during submission, we are protected by a guarantee that | 
|---|
| 686 | * the context/timeline is pinned and must remain pinned until after | 
|---|
| 687 | * this submission. | 
|---|
| 688 | */ | 
|---|
| 689 | return rcu_dereference_protected(rq->timeline, | 
|---|
| 690 | lockdep_is_held(&rq->engine->sched_engine->lock)); | 
|---|
| 691 | } | 
|---|
| 692 |  | 
|---|
| 693 | static inline u32 | 
|---|
| 694 | i915_request_active_seqno(const struct i915_request *rq) | 
|---|
| 695 | { | 
|---|
| 696 | u32 hwsp_phys_base = | 
|---|
| 697 | page_mask_bits(i915_request_active_timeline(rq)->hwsp_offset); | 
|---|
| 698 | u32 hwsp_relative_offset = offset_in_page(rq->hwsp_seqno); | 
|---|
| 699 |  | 
|---|
| 700 | /* | 
|---|
| 701 | * Because of wraparound, we cannot simply take tl->hwsp_offset, | 
|---|
| 702 | * but instead use the fact that the relative for vaddr is the | 
|---|
| 703 | * offset as for hwsp_offset. Take the top bits from tl->hwsp_offset | 
|---|
| 704 | * and combine them with the relative offset in rq->hwsp_seqno. | 
|---|
| 705 | * | 
|---|
| 706 | * As rw->hwsp_seqno is rewritten when signaled, this only works | 
|---|
| 707 | * when the request isn't signaled yet, but at that point you | 
|---|
| 708 | * no longer need the offset. | 
|---|
| 709 | */ | 
|---|
| 710 |  | 
|---|
| 711 | return hwsp_phys_base + hwsp_relative_offset; | 
|---|
| 712 | } | 
|---|
| 713 |  | 
|---|
| 714 | bool | 
|---|
| 715 | i915_request_active_engine(struct i915_request *rq, | 
|---|
| 716 | struct intel_engine_cs **active); | 
|---|
| 717 |  | 
|---|
| 718 | void i915_request_notify_execute_cb_imm(struct i915_request *rq); | 
|---|
| 719 |  | 
|---|
| 720 | enum i915_request_state { | 
|---|
| 721 | I915_REQUEST_UNKNOWN = 0, | 
|---|
| 722 | I915_REQUEST_COMPLETE, | 
|---|
| 723 | I915_REQUEST_PENDING, | 
|---|
| 724 | I915_REQUEST_QUEUED, | 
|---|
| 725 | I915_REQUEST_ACTIVE, | 
|---|
| 726 | }; | 
|---|
| 727 |  | 
|---|
| 728 | enum i915_request_state i915_test_request_state(struct i915_request *rq); | 
|---|
| 729 |  | 
|---|
| 730 | void i915_request_module_exit(void); | 
|---|
| 731 | int i915_request_module_init(void); | 
|---|
| 732 |  | 
|---|
| 733 | #endif /* I915_REQUEST_H */ | 
|---|
| 734 |  | 
|---|