| 1 | /* |
| 2 | * Copyright © 2008-2015 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 | #include <linux/dma-fence-array.h> |
| 26 | #include <linux/dma-fence-chain.h> |
| 27 | #include <linux/irq_work.h> |
| 28 | #include <linux/prefetch.h> |
| 29 | #include <linux/sched.h> |
| 30 | #include <linux/sched/clock.h> |
| 31 | #include <linux/sched/signal.h> |
| 32 | |
| 33 | #include "gem/i915_gem_context.h" |
| 34 | #include "gt/intel_breadcrumbs.h" |
| 35 | #include "gt/intel_context.h" |
| 36 | #include "gt/intel_gpu_commands.h" |
| 37 | #include "gt/intel_ring.h" |
| 38 | #include "gt/intel_rps.h" |
| 39 | |
| 40 | #include "i915_active.h" |
| 41 | #include "i915_drv.h" |
| 42 | #include "i915_globals.h" |
| 43 | #include "i915_trace.h" |
| 44 | #include "intel_pm.h" |
| 45 | |
| 46 | struct execute_cb { |
| 47 | struct irq_work work; |
| 48 | struct i915_sw_fence *fence; |
| 49 | void (*hook)(struct i915_request *rq, struct dma_fence *signal); |
| 50 | struct i915_request *signal; |
| 51 | }; |
| 52 | |
| 53 | static struct i915_global_request { |
| 54 | struct i915_global base; |
| 55 | struct kmem_cache *slab_requests; |
| 56 | struct kmem_cache *slab_execute_cbs; |
| 57 | } global; |
| 58 | |
| 59 | static const char *i915_fence_get_driver_name(struct dma_fence *fence) |
| 60 | { |
| 61 | return dev_name(to_request(fence)->engine->i915->drm.dev); |
| 62 | } |
| 63 | |
| 64 | static const char *i915_fence_get_timeline_name(struct dma_fence *fence) |
| 65 | { |
| 66 | const struct i915_gem_context *ctx; |
| 67 | |
| 68 | /* |
| 69 | * The timeline struct (as part of the ppgtt underneath a context) |
| 70 | * may be freed when the request is no longer in use by the GPU. |
| 71 | * We could extend the life of a context to beyond that of all |
| 72 | * fences, possibly keeping the hw resource around indefinitely, |
| 73 | * or we just give them a false name. Since |
| 74 | * dma_fence_ops.get_timeline_name is a debug feature, the occasional |
| 75 | * lie seems justifiable. |
| 76 | */ |
| 77 | if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) |
| 78 | return "signaled"; |
| 79 | |
| 80 | ctx = i915_request_gem_context(to_request(fence)); |
| 81 | if (!ctx) |
| 82 | return "[" DRIVER_NAME "]"; |
| 83 | |
| 84 | return ctx->name; |
| 85 | } |
| 86 | |
| 87 | static bool i915_fence_signaled(struct dma_fence *fence) |
| 88 | { |
| 89 | return i915_request_completed(to_request(fence)); |
| 90 | } |
| 91 | |
| 92 | static bool i915_fence_enable_signaling(struct dma_fence *fence) |
| 93 | { |
| 94 | return i915_request_enable_breadcrumb(to_request(fence)); |
| 95 | } |
| 96 | |
| 97 | static signed long i915_fence_wait(struct dma_fence *fence, |
| 98 | bool interruptible, |
| 99 | signed long timeout) |
| 100 | { |
| 101 | return i915_request_wait(to_request(fence), |
| 102 | interruptible | I915_WAIT_PRIORITY, |
| 103 | timeout); |
| 104 | } |
| 105 | |
| 106 | struct kmem_cache *i915_request_slab_cache(void) |
| 107 | { |
| 108 | return global.slab_requests; |
| 109 | } |
| 110 | |
| 111 | static void i915_fence_release(struct dma_fence *fence) |
| 112 | { |
| 113 | struct i915_request *rq = to_request(fence); |
| 114 | |
| 115 | /* |
| 116 | * The request is put onto a RCU freelist (i.e. the address |
| 117 | * is immediately reused), mark the fences as being freed now. |
| 118 | * Otherwise the debugobjects for the fences are only marked as |
| 119 | * freed when the slab cache itself is freed, and so we would get |
| 120 | * caught trying to reuse dead objects. |
| 121 | */ |
| 122 | i915_sw_fence_fini(&rq->submit); |
| 123 | i915_sw_fence_fini(&rq->semaphore); |
| 124 | |
| 125 | /* |
| 126 | * Keep one request on each engine for reserved use under mempressure |
| 127 | * |
| 128 | * We do not hold a reference to the engine here and so have to be |
| 129 | * very careful in what rq->engine we poke. The virtual engine is |
| 130 | * referenced via the rq->context and we released that ref during |
| 131 | * i915_request_retire(), ergo we must not dereference a virtual |
| 132 | * engine here. Not that we would want to, as the only consumer of |
| 133 | * the reserved engine->request_pool is the power management parking, |
| 134 | * which must-not-fail, and that is only run on the physical engines. |
| 135 | * |
| 136 | * Since the request must have been executed to be have completed, |
| 137 | * we know that it will have been processed by the HW and will |
| 138 | * not be unsubmitted again, so rq->engine and rq->execution_mask |
| 139 | * at this point is stable. rq->execution_mask will be a single |
| 140 | * bit if the last and _only_ engine it could execution on was a |
| 141 | * physical engine, if it's multiple bits then it started on and |
| 142 | * could still be on a virtual engine. Thus if the mask is not a |
| 143 | * power-of-two we assume that rq->engine may still be a virtual |
| 144 | * engine and so a dangling invalid pointer that we cannot dereference |
| 145 | * |
| 146 | * For example, consider the flow of a bonded request through a virtual |
| 147 | * engine. The request is created with a wide engine mask (all engines |
| 148 | * that we might execute on). On processing the bond, the request mask |
| 149 | * is reduced to one or more engines. If the request is subsequently |
| 150 | * bound to a single engine, it will then be constrained to only |
| 151 | * execute on that engine and never returned to the virtual engine |
| 152 | * after timeslicing away, see __unwind_incomplete_requests(). Thus we |
| 153 | * know that if the rq->execution_mask is a single bit, rq->engine |
| 154 | * can be a physical engine with the exact corresponding mask. |
| 155 | */ |
| 156 | if (is_power_of_2(rq->execution_mask) && |
| 157 | !cmpxchg(&rq->engine->request_pool, NULL, rq)) |
| 158 | return; |
| 159 | |
| 160 | kmem_cache_free(global.slab_requests, rq); |
| 161 | } |
| 162 | |
| 163 | const struct dma_fence_ops i915_fence_ops = { |
| 164 | .get_driver_name = i915_fence_get_driver_name, |
| 165 | .get_timeline_name = i915_fence_get_timeline_name, |
| 166 | .enable_signaling = i915_fence_enable_signaling, |
| 167 | .signaled = i915_fence_signaled, |
| 168 | .wait = i915_fence_wait, |
| 169 | .release = i915_fence_release, |
| 170 | }; |
| 171 | |
| 172 | static void irq_execute_cb(struct irq_work *wrk) |
| 173 | { |
| 174 | struct execute_cb *cb = container_of(wrk, typeof(*cb), work); |
| 175 | |
| 176 | i915_sw_fence_complete(cb->fence); |
| 177 | kmem_cache_free(global.slab_execute_cbs, cb); |
| 178 | } |
| 179 | |
| 180 | static void irq_execute_cb_hook(struct irq_work *wrk) |
| 181 | { |
| 182 | struct execute_cb *cb = container_of(wrk, typeof(*cb), work); |
| 183 | |
| 184 | cb->hook(container_of(cb->fence, struct i915_request, submit), |
| 185 | &cb->signal->fence); |
| 186 | i915_request_put(cb->signal); |
| 187 | |
| 188 | irq_execute_cb(wrk); |
| 189 | } |
| 190 | |
| 191 | static __always_inline void |
| 192 | __notify_execute_cb(struct i915_request *rq, bool (*fn)(struct irq_work *wrk)) |
| 193 | { |
| 194 | struct execute_cb *cb, *cn; |
| 195 | |
| 196 | if (llist_empty(&rq->execute_cb)) |
| 197 | return; |
| 198 | |
| 199 | llist_for_each_entry_safe(cb, cn, |
| 200 | llist_del_all(&rq->execute_cb), |
| 201 | work.node.llist) |
| 202 | fn(&cb->work); |
| 203 | } |
| 204 | |
| 205 | static void __notify_execute_cb_irq(struct i915_request *rq) |
| 206 | { |
| 207 | __notify_execute_cb(rq, irq_work_queue); |
| 208 | } |
| 209 | |
| 210 | static bool irq_work_imm(struct irq_work *wrk) |
| 211 | { |
| 212 | wrk->func(wrk); |
| 213 | return false; |
| 214 | } |
| 215 | |
| 216 | static void __notify_execute_cb_imm(struct i915_request *rq) |
| 217 | { |
| 218 | __notify_execute_cb(rq, irq_work_imm); |
| 219 | } |
| 220 | |
| 221 | static void free_capture_list(struct i915_request *request) |
| 222 | { |
| 223 | struct i915_capture_list *capture; |
| 224 | |
| 225 | capture = fetch_and_zero(&request->capture_list); |
| 226 | while (capture) { |
| 227 | struct i915_capture_list *next = capture->next; |
| 228 | |
| 229 | kfree(capture); |
| 230 | capture = next; |
| 231 | } |
| 232 | } |
| 233 | |
| 234 | static void __i915_request_fill(struct i915_request *rq, u8 val) |
| 235 | { |
| 236 | void *vaddr = rq->ring->vaddr; |
| 237 | u32 head; |
| 238 | |
| 239 | head = rq->infix; |
| 240 | if (rq->postfix < head) { |
| 241 | memset(vaddr + head, val, rq->ring->size - head); |
| 242 | head = 0; |
| 243 | } |
| 244 | memset(vaddr + head, val, rq->postfix - head); |
| 245 | } |
| 246 | |
| 247 | static void remove_from_engine(struct i915_request *rq) |
| 248 | { |
| 249 | struct intel_engine_cs *engine, *locked; |
| 250 | |
| 251 | /* |
| 252 | * Virtual engines complicate acquiring the engine timeline lock, |
| 253 | * as their rq->engine pointer is not stable until under that |
| 254 | * engine lock. The simple ploy we use is to take the lock then |
| 255 | * check that the rq still belongs to the newly locked engine. |
| 256 | */ |
| 257 | locked = READ_ONCE(rq->engine); |
| 258 | spin_lock_irq(&locked->active.lock); |
| 259 | while (unlikely(locked != (engine = READ_ONCE(rq->engine)))) { |
| 260 | spin_unlock(&locked->active.lock); |
| 261 | spin_lock(&engine->active.lock); |
| 262 | locked = engine; |
| 263 | } |
| 264 | list_del_init(&rq->sched.link); |
| 265 | |
| 266 | clear_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags); |
| 267 | clear_bit(I915_FENCE_FLAG_HOLD, &rq->fence.flags); |
| 268 | |
| 269 | /* Prevent further __await_execution() registering a cb, then flush */ |
| 270 | set_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags); |
| 271 | |
| 272 | spin_unlock_irq(&locked->active.lock); |
| 273 | |
| 274 | __notify_execute_cb_imm(rq); |
| 275 | } |
| 276 | |
| 277 | bool i915_request_retire(struct i915_request *rq) |
| 278 | { |
| 279 | if (!i915_request_completed(rq)) |
| 280 | return false; |
| 281 | |
| 282 | RQ_TRACE(rq, "\n"); |
| 283 | |
| 284 | GEM_BUG_ON(!i915_sw_fence_signaled(&rq->submit)); |
| 285 | trace_i915_request_retire(rq); |
| 286 | i915_request_mark_complete(rq); |
| 287 | |
| 288 | /* |
| 289 | * We know the GPU must have read the request to have |
| 290 | * sent us the seqno + interrupt, so use the position |
| 291 | * of tail of the request to update the last known position |
| 292 | * of the GPU head. |
| 293 | * |
| 294 | * Note this requires that we are always called in request |
| 295 | * completion order. |
| 296 | */ |
| 297 | GEM_BUG_ON(!list_is_first(&rq->link, |
| 298 | &i915_request_timeline(rq)->requests)); |
| 299 | if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)) |
| 300 | /* Poison before we release our space in the ring */ |
| 301 | __i915_request_fill(rq, POISON_FREE); |
| 302 | rq->ring->head = rq->postfix; |
| 303 | |
| 304 | if (!i915_request_signaled(rq)) { |
| 305 | spin_lock_irq(&rq->lock); |
| 306 | dma_fence_signal_locked(&rq->fence); |
| 307 | spin_unlock_irq(&rq->lock); |
| 308 | } |
| 309 | |
| 310 | if (test_and_set_bit(I915_FENCE_FLAG_BOOST, &rq->fence.flags)) |
| 311 | atomic_dec(&rq->engine->gt->rps.num_waiters); |
| 312 | |
| 313 | /* |
| 314 | * We only loosely track inflight requests across preemption, |
| 315 | * and so we may find ourselves attempting to retire a _completed_ |
| 316 | * request that we have removed from the HW and put back on a run |
| 317 | * queue. |
| 318 | * |
| 319 | * As we set I915_FENCE_FLAG_ACTIVE on the request, this should be |
| 320 | * after removing the breadcrumb and signaling it, so that we do not |
| 321 | * inadvertently attach the breadcrumb to a completed request. |
| 322 | */ |
| 323 | if (!list_empty(&rq->sched.link)) |
| 324 | remove_from_engine(rq); |
| 325 | GEM_BUG_ON(!llist_empty(&rq->execute_cb)); |
| 326 | |
| 327 | __list_del_entry(&rq->link); /* poison neither prev/next (RCU walks) */ |
| 328 | |
| 329 | intel_context_exit(rq->context); |
| 330 | intel_context_unpin(rq->context); |
| 331 | |
| 332 | free_capture_list(rq); |
| 333 | i915_sched_node_fini(&rq->sched); |
| 334 | i915_request_put(rq); |
| 335 | |
| 336 | return true; |
| 337 | } |
| 338 | |
| 339 | void i915_request_retire_upto(struct i915_request *rq) |
| 340 | { |
| 341 | struct intel_timeline * const tl = i915_request_timeline(rq); |
| 342 | struct i915_request *tmp; |
| 343 | |
| 344 | RQ_TRACE(rq, "\n"); |
| 345 | |
| 346 | GEM_BUG_ON(!i915_request_completed(rq)); |
| 347 | |
| 348 | do { |
| 349 | tmp = list_first_entry(&tl->requests, typeof(*tmp), link); |
| 350 | } while (i915_request_retire(tmp) && tmp != rq); |
| 351 | } |
| 352 | |
| 353 | static struct i915_request * const * |
| 354 | __engine_active(struct intel_engine_cs *engine) |
| 355 | { |
| 356 | return READ_ONCE(engine->execlists.active); |
| 357 | } |
| 358 | |
| 359 | static bool __request_in_flight(const struct i915_request *signal) |
| 360 | { |
| 361 | struct i915_request * const *port, *rq; |
| 362 | bool inflight = false; |
| 363 | |
| 364 | if (!i915_request_is_ready(signal)) |
| 365 | return false; |
| 366 | |
| 367 | /* |
| 368 | * Even if we have unwound the request, it may still be on |
| 369 | * the GPU (preempt-to-busy). If that request is inside an |
| 370 | * unpreemptible critical section, it will not be removed. Some |
| 371 | * GPU functions may even be stuck waiting for the paired request |
| 372 | * (__await_execution) to be submitted and cannot be preempted |
| 373 | * until the bond is executing. |
| 374 | * |
| 375 | * As we know that there are always preemption points between |
| 376 | * requests, we know that only the currently executing request |
| 377 | * may be still active even though we have cleared the flag. |
| 378 | * However, we can't rely on our tracking of ELSP[0] to know |
| 379 | * which request is currently active and so maybe stuck, as |
| 380 | * the tracking maybe an event behind. Instead assume that |
| 381 | * if the context is still inflight, then it is still active |
| 382 | * even if the active flag has been cleared. |
| 383 | * |
| 384 | * To further complicate matters, if there a pending promotion, the HW |
| 385 | * may either perform a context switch to the second inflight execlists, |
| 386 | * or it may switch to the pending set of execlists. In the case of the |
| 387 | * latter, it may send the ACK and we process the event copying the |
| 388 | * pending[] over top of inflight[], _overwriting_ our *active. Since |
| 389 | * this implies the HW is arbitrating and not struck in *active, we do |
| 390 | * not worry about complete accuracy, but we do require no read/write |
| 391 | * tearing of the pointer [the read of the pointer must be valid, even |
| 392 | * as the array is being overwritten, for which we require the writes |
| 393 | * to avoid tearing.] |
| 394 | * |
| 395 | * Note that the read of *execlists->active may race with the promotion |
| 396 | * of execlists->pending[] to execlists->inflight[], overwritting |
| 397 | * the value at *execlists->active. This is fine. The promotion implies |
| 398 | * that we received an ACK from the HW, and so the context is not |
| 399 | * stuck -- if we do not see ourselves in *active, the inflight status |
| 400 | * is valid. If instead we see ourselves being copied into *active, |
| 401 | * we are inflight and may signal the callback. |
| 402 | */ |
| 403 | if (!intel_context_inflight(signal->context)) |
| 404 | return false; |
| 405 | |
| 406 | rcu_read_lock(); |
| 407 | for (port = __engine_active(signal->engine); |
| 408 | (rq = READ_ONCE(*port)); /* may race with promotion of pending[] */ |
| 409 | port++) { |
| 410 | if (rq->context == signal->context) { |
| 411 | inflight = i915_seqno_passed(rq->fence.seqno, |
| 412 | signal->fence.seqno); |
| 413 | break; |
| 414 | } |
| 415 | } |
| 416 | rcu_read_unlock(); |
| 417 | |
| 418 | return inflight; |
| 419 | } |
| 420 | |
| 421 | static int |
| 422 | __await_execution(struct i915_request *rq, |
| 423 | struct i915_request *signal, |
| 424 | void (*hook)(struct i915_request *rq, |
| 425 | struct dma_fence *signal), |
| 426 | gfp_t gfp) |
| 427 | { |
| 428 | struct execute_cb *cb; |
| 429 | |
| 430 | if (i915_request_is_active(signal)) { |
| 431 | if (hook) |
| 432 | hook(rq, &signal->fence); |
| 433 | return 0; |
| 434 | } |
| 435 | |
| 436 | cb = kmem_cache_alloc(global.slab_execute_cbs, gfp); |
| 437 | if (!cb) |
| 438 | return -ENOMEM; |
| 439 | |
| 440 | cb->fence = &rq->submit; |
| 441 | i915_sw_fence_await(cb->fence); |
| 442 | init_irq_work(&cb->work, irq_execute_cb); |
| 443 | |
| 444 | if (hook) { |
| 445 | cb->hook = hook; |
| 446 | cb->signal = i915_request_get(signal); |
| 447 | cb->work.func = irq_execute_cb_hook; |
| 448 | } |
| 449 | |
| 450 | /* |
| 451 | * Register the callback first, then see if the signaler is already |
| 452 | * active. This ensures that if we race with the |
| 453 | * __notify_execute_cb from i915_request_submit() and we are not |
| 454 | * included in that list, we get a second bite of the cherry and |
| 455 | * execute it ourselves. After this point, a future |
| 456 | * i915_request_submit() will notify us. |
| 457 | * |
| 458 | * In i915_request_retire() we set the ACTIVE bit on a completed |
| 459 | * request (then flush the execute_cb). So by registering the |
| 460 | * callback first, then checking the ACTIVE bit, we serialise with |
| 461 | * the completed/retired request. |
| 462 | */ |
| 463 | if (llist_add(&cb->work.node.llist, &signal->execute_cb)) { |
| 464 | if (i915_request_is_active(signal) || |
| 465 | __request_in_flight(signal)) |
| 466 | __notify_execute_cb_imm(signal); |
| 467 | } |
| 468 | |
| 469 | return 0; |
| 470 | } |
| 471 | |
| 472 | static bool fatal_error(int error) |
| 473 | { |
| 474 | switch (error) { |
| 475 | case 0: /* not an error! */ |
| 476 | case -EAGAIN: /* innocent victim of a GT reset (__i915_request_reset) */ |
| 477 | case -ETIMEDOUT: /* waiting for Godot (timer_i915_sw_fence_wake) */ |
| 478 | return false; |
| 479 | default: |
| 480 | return true; |
| 481 | } |
| 482 | } |
| 483 | |
| 484 | void __i915_request_skip(struct i915_request *rq) |
| 485 | { |
| 486 | GEM_BUG_ON(!fatal_error(rq->fence.error)); |
| 487 | |
| 488 | if (rq->infix == rq->postfix) |
| 489 | return; |
| 490 | |
| 491 | RQ_TRACE(rq, "error: %d\n", rq->fence.error); |
| 492 | |
| 493 | /* |
| 494 | * As this request likely depends on state from the lost |
| 495 | * context, clear out all the user operations leaving the |
| 496 | * breadcrumb at the end (so we get the fence notifications). |
| 497 | */ |
| 498 | __i915_request_fill(rq, 0); |
| 499 | rq->infix = rq->postfix; |
| 500 | } |
| 501 | |
| 502 | void i915_request_set_error_once(struct i915_request *rq, int error) |
| 503 | { |
| 504 | int old; |
| 505 | |
| 506 | GEM_BUG_ON(!IS_ERR_VALUE((long)error)); |
| 507 | |
| 508 | if (i915_request_signaled(rq)) |
| 509 | return; |
| 510 | |
| 511 | old = READ_ONCE(rq->fence.error); |
| 512 | do { |
| 513 | if (fatal_error(old)) |
| 514 | return; |
| 515 | } while (!try_cmpxchg(&rq->fence.error, &old, error)); |
| 516 | } |
| 517 | |
| 518 | void i915_request_mark_eio(struct i915_request *rq) |
| 519 | { |
| 520 | if (__i915_request_is_complete(rq)) |
| 521 | return; |
| 522 | |
| 523 | GEM_BUG_ON(i915_request_signaled(rq)); |
| 524 | |
| 525 | i915_request_set_error_once(rq, -EIO); |
| 526 | i915_request_mark_complete(rq); |
| 527 | } |
| 528 | |
| 529 | bool __i915_request_submit(struct i915_request *request) |
| 530 | { |
| 531 | struct intel_engine_cs *engine = request->engine; |
| 532 | bool result = false; |
| 533 | |
| 534 | RQ_TRACE(request, "\n"); |
| 535 | |
| 536 | GEM_BUG_ON(!irqs_disabled()); |
| 537 | lockdep_assert_held(&engine->active.lock); |
| 538 | |
| 539 | /* |
| 540 | * With the advent of preempt-to-busy, we frequently encounter |
| 541 | * requests that we have unsubmitted from HW, but left running |
| 542 | * until the next ack and so have completed in the meantime. On |
| 543 | * resubmission of that completed request, we can skip |
| 544 | * updating the payload, and execlists can even skip submitting |
| 545 | * the request. |
| 546 | * |
| 547 | * We must remove the request from the caller's priority queue, |
| 548 | * and the caller must only call us when the request is in their |
| 549 | * priority queue, under the active.lock. This ensures that the |
| 550 | * request has *not* yet been retired and we can safely move |
| 551 | * the request into the engine->active.list where it will be |
| 552 | * dropped upon retiring. (Otherwise if resubmit a *retired* |
| 553 | * request, this would be a horrible use-after-free.) |
| 554 | */ |
| 555 | if (i915_request_completed(request)) |
| 556 | goto xfer; |
| 557 | |
| 558 | if (unlikely(intel_context_is_banned(request->context))) |
| 559 | i915_request_set_error_once(request, -EIO); |
| 560 | |
| 561 | if (unlikely(fatal_error(request->fence.error))) |
| 562 | __i915_request_skip(request); |
| 563 | |
| 564 | /* |
| 565 | * Are we using semaphores when the gpu is already saturated? |
| 566 | * |
| 567 | * Using semaphores incurs a cost in having the GPU poll a |
| 568 | * memory location, busywaiting for it to change. The continual |
| 569 | * memory reads can have a noticeable impact on the rest of the |
| 570 | * system with the extra bus traffic, stalling the cpu as it too |
| 571 | * tries to access memory across the bus (perf stat -e bus-cycles). |
| 572 | * |
| 573 | * If we installed a semaphore on this request and we only submit |
| 574 | * the request after the signaler completed, that indicates the |
| 575 | * system is overloaded and using semaphores at this time only |
| 576 | * increases the amount of work we are doing. If so, we disable |
| 577 | * further use of semaphores until we are idle again, whence we |
| 578 | * optimistically try again. |
| 579 | */ |
| 580 | if (request->sched.semaphores && |
| 581 | i915_sw_fence_signaled(&request->semaphore)) |
| 582 | engine->saturated |= request->sched.semaphores; |
| 583 | |
| 584 | engine->emit_fini_breadcrumb(request, |
| 585 | request->ring->vaddr + request->postfix); |
| 586 | |
| 587 | trace_i915_request_execute(request); |
| 588 | engine->serial++; |
| 589 | result = true; |
| 590 | |
| 591 | xfer: |
| 592 | if (!test_and_set_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags)) { |
| 593 | list_move_tail(&request->sched.link, &engine->active.requests); |
| 594 | clear_bit(I915_FENCE_FLAG_PQUEUE, &request->fence.flags); |
| 595 | } |
| 596 | |
| 597 | /* |
| 598 | * XXX Rollback bonded-execution on __i915_request_unsubmit()? |
| 599 | * |
| 600 | * In the future, perhaps when we have an active time-slicing scheduler, |
| 601 | * it will be interesting to unsubmit parallel execution and remove |
| 602 | * busywaits from the GPU until their master is restarted. This is |
| 603 | * quite hairy, we have to carefully rollback the fence and do a |
| 604 | * preempt-to-idle cycle on the target engine, all the while the |
| 605 | * master execute_cb may refire. |
| 606 | */ |
| 607 | __notify_execute_cb_irq(request); |
| 608 | |
| 609 | /* We may be recursing from the signal callback of another i915 fence */ |
| 610 | if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags)) |
| 611 | i915_request_enable_breadcrumb(request); |
| 612 | |
| 613 | return result; |
| 614 | } |
| 615 | |
| 616 | void i915_request_submit(struct i915_request *request) |
| 617 | { |
| 618 | struct intel_engine_cs *engine = request->engine; |
| 619 | unsigned long flags; |
| 620 | |
| 621 | /* Will be called from irq-context when using foreign fences. */ |
| 622 | spin_lock_irqsave(&engine->active.lock, flags); |
| 623 | |
| 624 | __i915_request_submit(request); |
| 625 | |
| 626 | spin_unlock_irqrestore(&engine->active.lock, flags); |
| 627 | } |
| 628 | |
| 629 | void __i915_request_unsubmit(struct i915_request *request) |
| 630 | { |
| 631 | struct intel_engine_cs *engine = request->engine; |
| 632 | |
| 633 | /* |
| 634 | * Only unwind in reverse order, required so that the per-context list |
| 635 | * is kept in seqno/ring order. |
| 636 | */ |
| 637 | RQ_TRACE(request, "\n"); |
| 638 | |
| 639 | GEM_BUG_ON(!irqs_disabled()); |
| 640 | lockdep_assert_held(&engine->active.lock); |
| 641 | |
| 642 | /* |
| 643 | * Before we remove this breadcrumb from the signal list, we have |
| 644 | * to ensure that a concurrent dma_fence_enable_signaling() does not |
| 645 | * attach itself. We first mark the request as no longer active and |
| 646 | * make sure that is visible to other cores, and then remove the |
| 647 | * breadcrumb if attached. |
| 648 | */ |
| 649 | GEM_BUG_ON(!test_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags)); |
| 650 | clear_bit_unlock(I915_FENCE_FLAG_ACTIVE, &request->fence.flags); |
| 651 | if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags)) |
| 652 | i915_request_cancel_breadcrumb(request); |
| 653 | |
| 654 | /* We've already spun, don't charge on resubmitting. */ |
| 655 | if (request->sched.semaphores && i915_request_started(request)) |
| 656 | request->sched.semaphores = 0; |
| 657 | |
| 658 | /* |
| 659 | * We don't need to wake_up any waiters on request->execute, they |
| 660 | * will get woken by any other event or us re-adding this request |
| 661 | * to the engine timeline (__i915_request_submit()). The waiters |
| 662 | * should be quite adapt at finding that the request now has a new |
| 663 | * global_seqno to the one they went to sleep on. |
| 664 | */ |
| 665 | } |
| 666 | |
| 667 | void i915_request_unsubmit(struct i915_request *request) |
| 668 | { |
| 669 | struct intel_engine_cs *engine = request->engine; |
| 670 | unsigned long flags; |
| 671 | |
| 672 | /* Will be called from irq-context when using foreign fences. */ |
| 673 | spin_lock_irqsave(&engine->active.lock, flags); |
| 674 | |
| 675 | __i915_request_unsubmit(request); |
| 676 | |
| 677 | spin_unlock_irqrestore(&engine->active.lock, flags); |
| 678 | } |
| 679 | |
| 680 | static int __i915_sw_fence_call |
| 681 | submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state) |
| 682 | { |
| 683 | struct i915_request *request = |
| 684 | container_of(fence, typeof(*request), submit); |
| 685 | |
| 686 | switch (state) { |
| 687 | case FENCE_COMPLETE: |
| 688 | trace_i915_request_submit(request); |
| 689 | |
| 690 | if (unlikely(fence->error)) |
| 691 | i915_request_set_error_once(request, fence->error); |
| 692 | |
| 693 | /* |
| 694 | * We need to serialize use of the submit_request() callback |
| 695 | * with its hotplugging performed during an emergency |
| 696 | * i915_gem_set_wedged(). We use the RCU mechanism to mark the |
| 697 | * critical section in order to force i915_gem_set_wedged() to |
| 698 | * wait until the submit_request() is completed before |
| 699 | * proceeding. |
| 700 | */ |
| 701 | rcu_read_lock(); |
| 702 | request->engine->submit_request(request); |
| 703 | rcu_read_unlock(); |
| 704 | break; |
| 705 | |
| 706 | case FENCE_FREE: |
| 707 | i915_request_put(request); |
| 708 | break; |
| 709 | } |
| 710 | |
| 711 | return NOTIFY_DONE; |
| 712 | } |
| 713 | |
| 714 | static int __i915_sw_fence_call |
| 715 | semaphore_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state) |
| 716 | { |
| 717 | struct i915_request *rq = container_of(fence, typeof(*rq), semaphore); |
| 718 | |
| 719 | switch (state) { |
| 720 | case FENCE_COMPLETE: |
| 721 | break; |
| 722 | |
| 723 | case FENCE_FREE: |
| 724 | i915_request_put(rq); |
| 725 | break; |
| 726 | } |
| 727 | |
| 728 | return NOTIFY_DONE; |
| 729 | } |
| 730 | |
| 731 | static void retire_requests(struct intel_timeline *tl) |
| 732 | { |
| 733 | struct i915_request *rq, *rn; |
| 734 | |
| 735 | list_for_each_entry_safe(rq, rn, &tl->requests, link) |
| 736 | if (!i915_request_retire(rq)) |
| 737 | break; |
| 738 | } |
| 739 | |
| 740 | static noinline struct i915_request * |
| 741 | request_alloc_slow(struct intel_timeline *tl, |
| 742 | struct i915_request **rsvd, |
| 743 | gfp_t gfp) |
| 744 | { |
| 745 | struct i915_request *rq; |
| 746 | |
| 747 | /* If we cannot wait, dip into our reserves */ |
| 748 | if (!gfpflags_allow_blocking(gfp)) { |
| 749 | rq = xchg(rsvd, NULL); |
| 750 | if (!rq) /* Use the normal failure path for one final WARN */ |
| 751 | goto out; |
| 752 | |
| 753 | return rq; |
| 754 | } |
| 755 | |
| 756 | if (list_empty(&tl->requests)) |
| 757 | goto out; |
| 758 | |
| 759 | /* Move our oldest request to the slab-cache (if not in use!) */ |
| 760 | rq = list_first_entry(&tl->requests, typeof(*rq), link); |
| 761 | i915_request_retire(rq); |
| 762 | |
| 763 | rq = kmem_cache_alloc(global.slab_requests, |
| 764 | gfp | __GFP_RETRY_MAYFAIL | __GFP_NOWARN); |
| 765 | if (rq) |
| 766 | return rq; |
| 767 | |
| 768 | /* Ratelimit ourselves to prevent oom from malicious clients */ |
| 769 | rq = list_last_entry(&tl->requests, typeof(*rq), link); |
| 770 | cond_synchronize_rcu(rq->rcustate); |
| 771 | |
| 772 | /* Retire our old requests in the hope that we free some */ |
| 773 | retire_requests(tl); |
| 774 | |
| 775 | out: |
| 776 | return kmem_cache_alloc(global.slab_requests, gfp); |
| 777 | } |
| 778 | |
| 779 | static void __i915_request_ctor(void *arg) |
| 780 | { |
| 781 | struct i915_request *rq = arg; |
| 782 | |
| 783 | spin_lock_init(&rq->lock); |
| 784 | i915_sched_node_init(&rq->sched); |
| 785 | i915_sw_fence_init(&rq->submit, submit_notify); |
| 786 | i915_sw_fence_init(&rq->semaphore, semaphore_notify); |
| 787 | |
| 788 | dma_fence_init(&rq->fence, &i915_fence_ops, &rq->lock, 0, 0); |
| 789 | |
| 790 | rq->capture_list = NULL; |
| 791 | |
| 792 | init_llist_head(&rq->execute_cb); |
| 793 | } |
| 794 | |
| 795 | struct i915_request * |
| 796 | __i915_request_create(struct intel_context *ce, gfp_t gfp) |
| 797 | { |
| 798 | struct intel_timeline *tl = ce->timeline; |
| 799 | struct i915_request *rq; |
| 800 | u32 seqno; |
| 801 | int ret; |
| 802 | |
| 803 | might_sleep_if(gfpflags_allow_blocking(gfp)); |
| 804 | |
| 805 | /* Check that the caller provided an already pinned context */ |
| 806 | __intel_context_pin(ce); |
| 807 | |
| 808 | /* |
| 809 | * Beware: Dragons be flying overhead. |
| 810 | * |
| 811 | * We use RCU to look up requests in flight. The lookups may |
| 812 | * race with the request being allocated from the slab freelist. |
| 813 | * That is the request we are writing to here, may be in the process |
| 814 | * of being read by __i915_active_request_get_rcu(). As such, |
| 815 | * we have to be very careful when overwriting the contents. During |
| 816 | * the RCU lookup, we change chase the request->engine pointer, |
| 817 | * read the request->global_seqno and increment the reference count. |
| 818 | * |
| 819 | * The reference count is incremented atomically. If it is zero, |
| 820 | * the lookup knows the request is unallocated and complete. Otherwise, |
| 821 | * it is either still in use, or has been reallocated and reset |
| 822 | * with dma_fence_init(). This increment is safe for release as we |
| 823 | * check that the request we have a reference to and matches the active |
| 824 | * request. |
| 825 | * |
| 826 | * Before we increment the refcount, we chase the request->engine |
| 827 | * pointer. We must not call kmem_cache_zalloc() or else we set |
| 828 | * that pointer to NULL and cause a crash during the lookup. If |
| 829 | * we see the request is completed (based on the value of the |
| 830 | * old engine and seqno), the lookup is complete and reports NULL. |
| 831 | * If we decide the request is not completed (new engine or seqno), |
| 832 | * then we grab a reference and double check that it is still the |
| 833 | * active request - which it won't be and restart the lookup. |
| 834 | * |
| 835 | * Do not use kmem_cache_zalloc() here! |
| 836 | */ |
| 837 | rq = kmem_cache_alloc(global.slab_requests, |
| 838 | gfp | __GFP_RETRY_MAYFAIL | __GFP_NOWARN); |
| 839 | if (unlikely(!rq)) { |
| 840 | rq = request_alloc_slow(tl, &ce->engine->request_pool, gfp); |
| 841 | if (!rq) { |
| 842 | ret = -ENOMEM; |
| 843 | goto err_unreserve; |
| 844 | } |
| 845 | } |
| 846 | |
| 847 | rq->context = ce; |
| 848 | rq->engine = ce->engine; |
| 849 | rq->ring = ce->ring; |
| 850 | rq->execution_mask = ce->engine->mask; |
| 851 | |
| 852 | kref_init(&rq->fence.refcount); |
| 853 | rq->fence.flags = 0; |
| 854 | rq->fence.error = 0; |
| 855 | INIT_LIST_HEAD(&rq->fence.cb_list); |
| 856 | |
| 857 | ret = intel_timeline_get_seqno(tl, rq, &seqno); |
| 858 | if (ret) |
| 859 | goto err_free; |
| 860 | |
| 861 | rq->fence.context = tl->fence_context; |
| 862 | rq->fence.seqno = seqno; |
| 863 | |
| 864 | RCU_INIT_POINTER(rq->timeline, tl); |
| 865 | RCU_INIT_POINTER(rq->hwsp_cacheline, tl->hwsp_cacheline); |
| 866 | rq->hwsp_seqno = tl->hwsp_seqno; |
| 867 | GEM_BUG_ON(i915_request_completed(rq)); |
| 868 | |
| 869 | rq->rcustate = get_state_synchronize_rcu(); /* acts as smp_mb() */ |
| 870 | |
| 871 | /* We bump the ref for the fence chain */ |
| 872 | i915_sw_fence_reinit(&i915_request_get(rq)->submit); |
| 873 | i915_sw_fence_reinit(&i915_request_get(rq)->semaphore); |
| 874 | |
| 875 | i915_sched_node_reinit(&rq->sched); |
| 876 | |
| 877 | /* No zalloc, everything must be cleared after use */ |
| 878 | rq->batch = NULL; |
| 879 | GEM_BUG_ON(rq->capture_list); |
| 880 | GEM_BUG_ON(!llist_empty(&rq->execute_cb)); |
| 881 | |
| 882 | /* |
| 883 | * Reserve space in the ring buffer for all the commands required to |
| 884 | * eventually emit this request. This is to guarantee that the |
| 885 | * i915_request_add() call can't fail. Note that the reserve may need |
| 886 | * to be redone if the request is not actually submitted straight |
| 887 | * away, e.g. because a GPU scheduler has deferred it. |
| 888 | * |
| 889 | * Note that due to how we add reserved_space to intel_ring_begin() |
| 890 | * we need to double our request to ensure that if we need to wrap |
| 891 | * around inside i915_request_add() there is sufficient space at |
| 892 | * the beginning of the ring as well. |
| 893 | */ |
| 894 | rq->reserved_space = |
| 895 | 2 * rq->engine->emit_fini_breadcrumb_dw * sizeof(u32); |
| 896 | |
| 897 | /* |
| 898 | * Record the position of the start of the request so that |
| 899 | * should we detect the updated seqno part-way through the |
| 900 | * GPU processing the request, we never over-estimate the |
| 901 | * position of the head. |
| 902 | */ |
| 903 | rq->head = rq->ring->emit; |
| 904 | |
| 905 | ret = rq->engine->request_alloc(rq); |
| 906 | if (ret) |
| 907 | goto err_unwind; |
| 908 | |
| 909 | rq->infix = rq->ring->emit; /* end of header; start of user payload */ |
| 910 | |
| 911 | intel_context_mark_active(ce); |
| 912 | list_add_tail_rcu(&rq->link, &tl->requests); |
| 913 | |
| 914 | return rq; |
| 915 | |
| 916 | err_unwind: |
| 917 | ce->ring->emit = rq->head; |
| 918 | |
| 919 | /* Make sure we didn't add ourselves to external state before freeing */ |
| 920 | GEM_BUG_ON(!list_empty(&rq->sched.signalers_list)); |
| 921 | GEM_BUG_ON(!list_empty(&rq->sched.waiters_list)); |
| 922 | |
| 923 | err_free: |
| 924 | kmem_cache_free(global.slab_requests, rq); |
| 925 | err_unreserve: |
| 926 | intel_context_unpin(ce); |
| 927 | return ERR_PTR(ret); |
| 928 | } |
| 929 | |
| 930 | struct i915_request * |
| 931 | i915_request_create(struct intel_context *ce) |
| 932 | { |
| 933 | struct i915_request *rq; |
| 934 | struct intel_timeline *tl; |
| 935 | |
| 936 | tl = intel_context_timeline_lock(ce); |
| 937 | if (IS_ERR(tl)) |
| 938 | return ERR_CAST(tl); |
| 939 | |
| 940 | /* Move our oldest request to the slab-cache (if not in use!) */ |
| 941 | rq = list_first_entry(&tl->requests, typeof(*rq), link); |
| 942 | if (!list_is_last(&rq->link, &tl->requests)) |
| 943 | i915_request_retire(rq); |
| 944 | |
| 945 | intel_context_enter(ce); |
| 946 | rq = __i915_request_create(ce, GFP_KERNEL); |
| 947 | intel_context_exit(ce); /* active reference transferred to request */ |
| 948 | if (IS_ERR(rq)) |
| 949 | goto err_unlock; |
| 950 | |
| 951 | /* Check that we do not interrupt ourselves with a new request */ |
| 952 | rq->cookie = lockdep_pin_lock(&tl->mutex); |
| 953 | |
| 954 | return rq; |
| 955 | |
| 956 | err_unlock: |
| 957 | intel_context_timeline_unlock(tl); |
| 958 | return rq; |
| 959 | } |
| 960 | |
| 961 | static int |
| 962 | i915_request_await_start(struct i915_request *rq, struct i915_request *signal) |
| 963 | { |
| 964 | struct dma_fence *fence; |
| 965 | int err; |
| 966 | |
| 967 | if (i915_request_timeline(rq) == rcu_access_pointer(signal->timeline)) |
| 968 | return 0; |
| 969 | |
| 970 | if (i915_request_started(signal)) |
| 971 | return 0; |
| 972 | |
| 973 | fence = NULL; |
| 974 | rcu_read_lock(); |
| 975 | spin_lock_irq(&signal->lock); |
| 976 | do { |
| 977 | struct list_head *pos = READ_ONCE(signal->link.prev); |
| 978 | struct i915_request *prev; |
| 979 | |
| 980 | /* Confirm signal has not been retired, the link is valid */ |
| 981 | if (unlikely(i915_request_started(signal))) |
| 982 | break; |
| 983 | |
| 984 | /* Is signal the earliest request on its timeline? */ |
| 985 | if (pos == &rcu_dereference(signal->timeline)->requests) |
| 986 | break; |
| 987 | |
| 988 | /* |
| 989 | * Peek at the request before us in the timeline. That |
| 990 | * request will only be valid before it is retired, so |
| 991 | * after acquiring a reference to it, confirm that it is |
| 992 | * still part of the signaler's timeline. |
| 993 | */ |
| 994 | prev = list_entry(pos, typeof(*prev), link); |
| 995 | if (!i915_request_get_rcu(prev)) |
| 996 | break; |
| 997 | |
| 998 | /* After the strong barrier, confirm prev is still attached */ |
| 999 | if (unlikely(READ_ONCE(prev->link.next) != &signal->link)) { |
| 1000 | i915_request_put(prev); |
| 1001 | break; |
| 1002 | } |
| 1003 | |
| 1004 | fence = &prev->fence; |
| 1005 | } while (0); |
| 1006 | spin_unlock_irq(&signal->lock); |
| 1007 | rcu_read_unlock(); |
| 1008 | if (!fence) |
| 1009 | return 0; |
| 1010 | |
| 1011 | err = 0; |
| 1012 | if (!intel_timeline_sync_is_later(i915_request_timeline(rq), fence)) |
| 1013 | err = i915_sw_fence_await_dma_fence(&rq->submit, |
| 1014 | fence, 0, |
| 1015 | I915_FENCE_GFP); |
| 1016 | dma_fence_put(fence); |
| 1017 | |
| 1018 | return err; |
| 1019 | } |
| 1020 | |
| 1021 | static intel_engine_mask_t |
| 1022 | already_busywaiting(struct i915_request *rq) |
| 1023 | { |
| 1024 | /* |
| 1025 | * Polling a semaphore causes bus traffic, delaying other users of |
| 1026 | * both the GPU and CPU. We want to limit the impact on others, |
| 1027 | * while taking advantage of early submission to reduce GPU |
| 1028 | * latency. Therefore we restrict ourselves to not using more |
| 1029 | * than one semaphore from each source, and not using a semaphore |
| 1030 | * if we have detected the engine is saturated (i.e. would not be |
| 1031 | * submitted early and cause bus traffic reading an already passed |
| 1032 | * semaphore). |
| 1033 | * |
| 1034 | * See the are-we-too-late? check in __i915_request_submit(). |
| 1035 | */ |
| 1036 | return rq->sched.semaphores | READ_ONCE(rq->engine->saturated); |
| 1037 | } |
| 1038 | |
| 1039 | static int |
| 1040 | __emit_semaphore_wait(struct i915_request *to, |
| 1041 | struct i915_request *from, |
| 1042 | u32 seqno) |
| 1043 | { |
| 1044 | const int has_token = INTEL_GEN(to->engine->i915) >= 12; |
| 1045 | u32 hwsp_offset; |
| 1046 | int len, err; |
| 1047 | u32 *cs; |
| 1048 | |
| 1049 | GEM_BUG_ON(INTEL_GEN(to->engine->i915) < 8); |
| 1050 | GEM_BUG_ON(i915_request_has_initial_breadcrumb(to)); |
| 1051 | |
| 1052 | /* We need to pin the signaler's HWSP until we are finished reading. */ |
| 1053 | err = intel_timeline_read_hwsp(from, to, &hwsp_offset); |
| 1054 | if (err) |
| 1055 | return err; |
| 1056 | |
| 1057 | len = 4; |
| 1058 | if (has_token) |
| 1059 | len += 2; |
| 1060 | |
| 1061 | cs = intel_ring_begin(to, len); |
| 1062 | if (IS_ERR(cs)) |
| 1063 | return PTR_ERR(cs); |
| 1064 | |
| 1065 | /* |
| 1066 | * Using greater-than-or-equal here means we have to worry |
| 1067 | * about seqno wraparound. To side step that issue, we swap |
| 1068 | * the timeline HWSP upon wrapping, so that everyone listening |
| 1069 | * for the old (pre-wrap) values do not see the much smaller |
| 1070 | * (post-wrap) values than they were expecting (and so wait |
| 1071 | * forever). |
| 1072 | */ |
| 1073 | *cs++ = (MI_SEMAPHORE_WAIT | |
| 1074 | MI_SEMAPHORE_GLOBAL_GTT | |
| 1075 | MI_SEMAPHORE_POLL | |
| 1076 | MI_SEMAPHORE_SAD_GTE_SDD) + |
| 1077 | has_token; |
| 1078 | *cs++ = seqno; |
| 1079 | *cs++ = hwsp_offset; |
| 1080 | *cs++ = 0; |
| 1081 | if (has_token) { |
| 1082 | *cs++ = 0; |
| 1083 | *cs++ = MI_NOOP; |
| 1084 | } |
| 1085 | |
| 1086 | intel_ring_advance(to, cs); |
| 1087 | return 0; |
| 1088 | } |
| 1089 | |
| 1090 | static int |
| 1091 | emit_semaphore_wait(struct i915_request *to, |
| 1092 | struct i915_request *from, |
| 1093 | gfp_t gfp) |
| 1094 | { |
| 1095 | const intel_engine_mask_t mask = READ_ONCE(from->engine)->mask; |
| 1096 | struct i915_sw_fence *wait = &to->submit; |
| 1097 | |
| 1098 | if (!intel_context_use_semaphores(to->context)) |
| 1099 | goto await_fence; |
| 1100 | |
| 1101 | if (i915_request_has_initial_breadcrumb(to)) |
| 1102 | goto await_fence; |
| 1103 | |
| 1104 | if (!rcu_access_pointer(from->hwsp_cacheline)) |
| 1105 | goto await_fence; |
| 1106 | |
| 1107 | /* |
| 1108 | * If this or its dependents are waiting on an external fence |
| 1109 | * that may fail catastrophically, then we want to avoid using |
| 1110 | * sempahores as they bypass the fence signaling metadata, and we |
| 1111 | * lose the fence->error propagation. |
| 1112 | */ |
| 1113 | if (from->sched.flags & I915_SCHED_HAS_EXTERNAL_CHAIN) |
| 1114 | goto await_fence; |
| 1115 | |
| 1116 | /* Just emit the first semaphore we see as request space is limited. */ |
| 1117 | if (already_busywaiting(to) & mask) |
| 1118 | goto await_fence; |
| 1119 | |
| 1120 | if (i915_request_await_start(to, from) < 0) |
| 1121 | goto await_fence; |
| 1122 | |
| 1123 | /* Only submit our spinner after the signaler is running! */ |
| 1124 | if (__await_execution(to, from, NULL, gfp)) |
| 1125 | goto await_fence; |
| 1126 | |
| 1127 | if (__emit_semaphore_wait(to, from, from->fence.seqno)) |
| 1128 | goto await_fence; |
| 1129 | |
| 1130 | to->sched.semaphores |= mask; |
| 1131 | wait = &to->semaphore; |
| 1132 | |
| 1133 | await_fence: |
| 1134 | return i915_sw_fence_await_dma_fence(wait, |
| 1135 | &from->fence, 0, |
| 1136 | I915_FENCE_GFP); |
| 1137 | } |
| 1138 | |
| 1139 | static bool intel_timeline_sync_has_start(struct intel_timeline *tl, |
| 1140 | struct dma_fence *fence) |
| 1141 | { |
| 1142 | return __intel_timeline_sync_is_later(tl, |
| 1143 | fence->context, |
| 1144 | fence->seqno - 1); |
| 1145 | } |
| 1146 | |
| 1147 | static int intel_timeline_sync_set_start(struct intel_timeline *tl, |
| 1148 | const struct dma_fence *fence) |
| 1149 | { |
| 1150 | return __intel_timeline_sync_set(tl, fence->context, fence->seqno - 1); |
| 1151 | } |
| 1152 | |
| 1153 | static int |
| 1154 | __i915_request_await_execution(struct i915_request *to, |
| 1155 | struct i915_request *from, |
| 1156 | void (*hook)(struct i915_request *rq, |
| 1157 | struct dma_fence *signal)) |
| 1158 | { |
| 1159 | int err; |
| 1160 | |
| 1161 | GEM_BUG_ON(intel_context_is_barrier(from->context)); |
| 1162 | |
| 1163 | /* Submit both requests at the same time */ |
| 1164 | err = __await_execution(to, from, hook, I915_FENCE_GFP); |
| 1165 | if (err) |
| 1166 | return err; |
| 1167 | |
| 1168 | /* Squash repeated depenendices to the same timelines */ |
| 1169 | if (intel_timeline_sync_has_start(i915_request_timeline(to), |
| 1170 | &from->fence)) |
| 1171 | return 0; |
| 1172 | |
| 1173 | /* |
| 1174 | * Wait until the start of this request. |
| 1175 | * |
| 1176 | * The execution cb fires when we submit the request to HW. But in |
| 1177 | * many cases this may be long before the request itself is ready to |
| 1178 | * run (consider that we submit 2 requests for the same context, where |
| 1179 | * the request of interest is behind an indefinite spinner). So we hook |
| 1180 | * up to both to reduce our queues and keep the execution lag minimised |
| 1181 | * in the worst case, though we hope that the await_start is elided. |
| 1182 | */ |
| 1183 | err = i915_request_await_start(to, from); |
| 1184 | if (err < 0) |
| 1185 | return err; |
| 1186 | |
| 1187 | /* |
| 1188 | * Ensure both start together [after all semaphores in signal] |
| 1189 | * |
| 1190 | * Now that we are queued to the HW at roughly the same time (thanks |
| 1191 | * to the execute cb) and are ready to run at roughly the same time |
| 1192 | * (thanks to the await start), our signaler may still be indefinitely |
| 1193 | * delayed by waiting on a semaphore from a remote engine. If our |
| 1194 | * signaler depends on a semaphore, so indirectly do we, and we do not |
| 1195 | * want to start our payload until our signaler also starts theirs. |
| 1196 | * So we wait. |
| 1197 | * |
| 1198 | * However, there is also a second condition for which we need to wait |
| 1199 | * for the precise start of the signaler. Consider that the signaler |
| 1200 | * was submitted in a chain of requests following another context |
| 1201 | * (with just an ordinary intra-engine fence dependency between the |
| 1202 | * two). In this case the signaler is queued to HW, but not for |
| 1203 | * immediate execution, and so we must wait until it reaches the |
| 1204 | * active slot. |
| 1205 | */ |
| 1206 | if (intel_engine_has_semaphores(to->engine) && |
| 1207 | !i915_request_has_initial_breadcrumb(to)) { |
| 1208 | err = __emit_semaphore_wait(to, from, from->fence.seqno - 1); |
| 1209 | if (err < 0) |
| 1210 | return err; |
| 1211 | } |
| 1212 | |
| 1213 | /* Couple the dependency tree for PI on this exposed to->fence */ |
| 1214 | if (to->engine->schedule) { |
| 1215 | err = i915_sched_node_add_dependency(&to->sched, |
| 1216 | &from->sched, |
| 1217 | I915_DEPENDENCY_WEAK); |
| 1218 | if (err < 0) |
| 1219 | return err; |
| 1220 | } |
| 1221 | |
| 1222 | return intel_timeline_sync_set_start(i915_request_timeline(to), |
| 1223 | &from->fence); |
| 1224 | } |
| 1225 | |
| 1226 | static void mark_external(struct i915_request *rq) |
| 1227 | { |
| 1228 | /* |
| 1229 | * The downside of using semaphores is that we lose metadata passing |
| 1230 | * along the signaling chain. This is particularly nasty when we |
| 1231 | * need to pass along a fatal error such as EFAULT or EDEADLK. For |
| 1232 | * fatal errors we want to scrub the request before it is executed, |
| 1233 | * which means that we cannot preload the request onto HW and have |
| 1234 | * it wait upon a semaphore. |
| 1235 | */ |
| 1236 | rq->sched.flags |= I915_SCHED_HAS_EXTERNAL_CHAIN; |
| 1237 | } |
| 1238 | |
| 1239 | static int |
| 1240 | __i915_request_await_external(struct i915_request *rq, struct dma_fence *fence) |
| 1241 | { |
| 1242 | mark_external(rq); |
| 1243 | return i915_sw_fence_await_dma_fence(&rq->submit, fence, |
| 1244 | i915_fence_context_timeout(rq->engine->i915, |
| 1245 | fence->context), |
| 1246 | I915_FENCE_GFP); |
| 1247 | } |
| 1248 | |
| 1249 | static int |
| 1250 | i915_request_await_external(struct i915_request *rq, struct dma_fence *fence) |
| 1251 | { |
| 1252 | struct dma_fence *iter; |
| 1253 | int err = 0; |
| 1254 | |
| 1255 | if (!to_dma_fence_chain(fence)) |
| 1256 | return __i915_request_await_external(rq, fence); |
| 1257 | |
| 1258 | dma_fence_chain_for_each(iter, fence) { |
| 1259 | struct dma_fence_chain *chain = to_dma_fence_chain(iter); |
| 1260 | |
| 1261 | if (!dma_fence_is_i915(chain->fence)) { |
| 1262 | err = __i915_request_await_external(rq, iter); |
| 1263 | break; |
| 1264 | } |
| 1265 | |
| 1266 | err = i915_request_await_dma_fence(rq, chain->fence); |
| 1267 | if (err < 0) |
| 1268 | break; |
| 1269 | } |
| 1270 | |
| 1271 | dma_fence_put(iter); |
| 1272 | return err; |
| 1273 | } |
| 1274 | |
| 1275 | int |
| 1276 | i915_request_await_execution(struct i915_request *rq, |
| 1277 | struct dma_fence *fence, |
| 1278 | void (*hook)(struct i915_request *rq, |
| 1279 | struct dma_fence *signal)) |
| 1280 | { |
| 1281 | struct dma_fence **child = &fence; |
| 1282 | unsigned int nchild = 1; |
| 1283 | int ret; |
| 1284 | |
| 1285 | if (dma_fence_is_array(fence)) { |
| 1286 | struct dma_fence_array *array = to_dma_fence_array(fence); |
| 1287 | |
| 1288 | /* XXX Error for signal-on-any fence arrays */ |
| 1289 | |
| 1290 | child = array->fences; |
| 1291 | nchild = array->num_fences; |
| 1292 | GEM_BUG_ON(!nchild); |
| 1293 | } |
| 1294 | |
| 1295 | do { |
| 1296 | fence = *child++; |
| 1297 | if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) { |
| 1298 | i915_sw_fence_set_error_once(&rq->submit, fence->error); |
| 1299 | continue; |
| 1300 | } |
| 1301 | |
| 1302 | if (fence->context == rq->fence.context) |
| 1303 | continue; |
| 1304 | |
| 1305 | /* |
| 1306 | * We don't squash repeated fence dependencies here as we |
| 1307 | * want to run our callback in all cases. |
| 1308 | */ |
| 1309 | |
| 1310 | if (dma_fence_is_i915(fence)) |
| 1311 | ret = __i915_request_await_execution(rq, |
| 1312 | to_request(fence), |
| 1313 | hook); |
| 1314 | else |
| 1315 | ret = i915_request_await_external(rq, fence); |
| 1316 | if (ret < 0) |
| 1317 | return ret; |
| 1318 | } while (--nchild); |
| 1319 | |
| 1320 | return 0; |
| 1321 | } |
| 1322 | |
| 1323 | static int |
| 1324 | await_request_submit(struct i915_request *to, struct i915_request *from) |
| 1325 | { |
| 1326 | /* |
| 1327 | * If we are waiting on a virtual engine, then it may be |
| 1328 | * constrained to execute on a single engine *prior* to submission. |
| 1329 | * When it is submitted, it will be first submitted to the virtual |
| 1330 | * engine and then passed to the physical engine. We cannot allow |
| 1331 | * the waiter to be submitted immediately to the physical engine |
| 1332 | * as it may then bypass the virtual request. |
| 1333 | */ |
| 1334 | if (to->engine == READ_ONCE(from->engine)) |
| 1335 | return i915_sw_fence_await_sw_fence_gfp(&to->submit, |
| 1336 | &from->submit, |
| 1337 | I915_FENCE_GFP); |
| 1338 | else |
| 1339 | return __i915_request_await_execution(to, from, NULL); |
| 1340 | } |
| 1341 | |
| 1342 | static int |
| 1343 | i915_request_await_request(struct i915_request *to, struct i915_request *from) |
| 1344 | { |
| 1345 | int ret; |
| 1346 | |
| 1347 | GEM_BUG_ON(to == from); |
| 1348 | GEM_BUG_ON(to->timeline == from->timeline); |
| 1349 | |
| 1350 | if (i915_request_completed(from)) { |
| 1351 | i915_sw_fence_set_error_once(&to->submit, from->fence.error); |
| 1352 | return 0; |
| 1353 | } |
| 1354 | |
| 1355 | if (to->engine->schedule) { |
| 1356 | ret = i915_sched_node_add_dependency(&to->sched, |
| 1357 | &from->sched, |
| 1358 | I915_DEPENDENCY_EXTERNAL); |
| 1359 | if (ret < 0) |
| 1360 | return ret; |
| 1361 | } |
| 1362 | |
| 1363 | if (is_power_of_2(to->execution_mask | READ_ONCE(from->execution_mask))) |
| 1364 | ret = await_request_submit(to, from); |
| 1365 | else |
| 1366 | ret = emit_semaphore_wait(to, from, I915_FENCE_GFP); |
| 1367 | if (ret < 0) |
| 1368 | return ret; |
| 1369 | |
| 1370 | return 0; |
| 1371 | } |
| 1372 | |
| 1373 | int |
| 1374 | i915_request_await_dma_fence(struct i915_request *rq, struct dma_fence *fence) |
| 1375 | { |
| 1376 | struct dma_fence **child = &fence; |
| 1377 | unsigned int nchild = 1; |
| 1378 | int ret; |
| 1379 | |
| 1380 | /* |
| 1381 | * Note that if the fence-array was created in signal-on-any mode, |
| 1382 | * we should *not* decompose it into its individual fences. However, |
| 1383 | * we don't currently store which mode the fence-array is operating |
| 1384 | * in. Fortunately, the only user of signal-on-any is private to |
| 1385 | * amdgpu and we should not see any incoming fence-array from |
| 1386 | * sync-file being in signal-on-any mode. |
| 1387 | */ |
| 1388 | if (dma_fence_is_array(fence)) { |
| 1389 | struct dma_fence_array *array = to_dma_fence_array(fence); |
| 1390 | |
| 1391 | child = array->fences; |
| 1392 | nchild = array->num_fences; |
| 1393 | GEM_BUG_ON(!nchild); |
| 1394 | } |
| 1395 | |
| 1396 | do { |
| 1397 | fence = *child++; |
| 1398 | if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) { |
| 1399 | i915_sw_fence_set_error_once(&rq->submit, fence->error); |
| 1400 | continue; |
| 1401 | } |
| 1402 | |
| 1403 | /* |
| 1404 | * Requests on the same timeline are explicitly ordered, along |
| 1405 | * with their dependencies, by i915_request_add() which ensures |
| 1406 | * that requests are submitted in-order through each ring. |
| 1407 | */ |
| 1408 | if (fence->context == rq->fence.context) |
| 1409 | continue; |
| 1410 | |
| 1411 | /* Squash repeated waits to the same timelines */ |
| 1412 | if (fence->context && |
| 1413 | intel_timeline_sync_is_later(i915_request_timeline(rq), |
| 1414 | fence)) |
| 1415 | continue; |
| 1416 | |
| 1417 | if (dma_fence_is_i915(fence)) |
| 1418 | ret = i915_request_await_request(rq, to_request(fence)); |
| 1419 | else |
| 1420 | ret = i915_request_await_external(rq, fence); |
| 1421 | if (ret < 0) |
| 1422 | return ret; |
| 1423 | |
| 1424 | /* Record the latest fence used against each timeline */ |
| 1425 | if (fence->context) |
| 1426 | intel_timeline_sync_set(i915_request_timeline(rq), |
| 1427 | fence); |
| 1428 | } while (--nchild); |
| 1429 | |
| 1430 | return 0; |
| 1431 | } |
| 1432 | |
| 1433 | /** |
| 1434 | * i915_request_await_object - set this request to (async) wait upon a bo |
| 1435 | * @to: request we are wishing to use |
| 1436 | * @obj: object which may be in use on another ring. |
| 1437 | * @write: whether the wait is on behalf of a writer |
| 1438 | * |
| 1439 | * This code is meant to abstract object synchronization with the GPU. |
| 1440 | * Conceptually we serialise writes between engines inside the GPU. |
| 1441 | * We only allow one engine to write into a buffer at any time, but |
| 1442 | * multiple readers. To ensure each has a coherent view of memory, we must: |
| 1443 | * |
| 1444 | * - If there is an outstanding write request to the object, the new |
| 1445 | * request must wait for it to complete (either CPU or in hw, requests |
| 1446 | * on the same ring will be naturally ordered). |
| 1447 | * |
| 1448 | * - If we are a write request (pending_write_domain is set), the new |
| 1449 | * request must wait for outstanding read requests to complete. |
| 1450 | * |
| 1451 | * Returns 0 if successful, else propagates up the lower layer error. |
| 1452 | */ |
| 1453 | int |
| 1454 | i915_request_await_object(struct i915_request *to, |
| 1455 | struct drm_i915_gem_object *obj, |
| 1456 | bool write) |
| 1457 | { |
| 1458 | struct dma_fence *excl; |
| 1459 | int ret = 0; |
| 1460 | |
| 1461 | if (write) { |
| 1462 | struct dma_fence **shared; |
| 1463 | unsigned int count, i; |
| 1464 | |
| 1465 | ret = dma_resv_get_fences_rcu(obj->base.resv, |
| 1466 | &excl, &count, &shared); |
| 1467 | if (ret) |
| 1468 | return ret; |
| 1469 | |
| 1470 | for (i = 0; i < count; i++) { |
| 1471 | ret = i915_request_await_dma_fence(to, shared[i]); |
| 1472 | if (ret) |
| 1473 | break; |
| 1474 | |
| 1475 | dma_fence_put(shared[i]); |
| 1476 | } |
| 1477 | |
| 1478 | for (; i < count; i++) |
| 1479 | dma_fence_put(shared[i]); |
| 1480 | kfree(shared); |
| 1481 | } else { |
| 1482 | excl = dma_resv_get_excl_rcu(obj->base.resv); |
| 1483 | } |
| 1484 | |
| 1485 | if (excl) { |
| 1486 | if (ret == 0) |
| 1487 | ret = i915_request_await_dma_fence(to, excl); |
| 1488 | |
| 1489 | dma_fence_put(excl); |
| 1490 | } |
| 1491 | |
| 1492 | return ret; |
| 1493 | } |
| 1494 | |
| 1495 | static struct i915_request * |
| 1496 | __i915_request_add_to_timeline(struct i915_request *rq) |
| 1497 | { |
| 1498 | struct intel_timeline *timeline = i915_request_timeline(rq); |
| 1499 | struct i915_request *prev; |
| 1500 | |
| 1501 | /* |
| 1502 | * Dependency tracking and request ordering along the timeline |
| 1503 | * is special cased so that we can eliminate redundant ordering |
| 1504 | * operations while building the request (we know that the timeline |
| 1505 | * itself is ordered, and here we guarantee it). |
| 1506 | * |
| 1507 | * As we know we will need to emit tracking along the timeline, |
| 1508 | * we embed the hooks into our request struct -- at the cost of |
| 1509 | * having to have specialised no-allocation interfaces (which will |
| 1510 | * be beneficial elsewhere). |
| 1511 | * |
| 1512 | * A second benefit to open-coding i915_request_await_request is |
| 1513 | * that we can apply a slight variant of the rules specialised |
| 1514 | * for timelines that jump between engines (such as virtual engines). |
| 1515 | * If we consider the case of virtual engine, we must emit a dma-fence |
| 1516 | * to prevent scheduling of the second request until the first is |
| 1517 | * complete (to maximise our greedy late load balancing) and this |
| 1518 | * precludes optimising to use semaphores serialisation of a single |
| 1519 | * timeline across engines. |
| 1520 | */ |
| 1521 | prev = to_request(__i915_active_fence_set(&timeline->last_request, |
| 1522 | &rq->fence)); |
| 1523 | if (prev && !i915_request_completed(prev)) { |
| 1524 | /* |
| 1525 | * The requests are supposed to be kept in order. However, |
| 1526 | * we need to be wary in case the timeline->last_request |
| 1527 | * is used as a barrier for external modification to this |
| 1528 | * context. |
| 1529 | */ |
| 1530 | GEM_BUG_ON(prev->context == rq->context && |
| 1531 | i915_seqno_passed(prev->fence.seqno, |
| 1532 | rq->fence.seqno)); |
| 1533 | |
| 1534 | if (is_power_of_2(READ_ONCE(prev->engine)->mask | rq->engine->mask)) |
| 1535 | i915_sw_fence_await_sw_fence(&rq->submit, |
| 1536 | &prev->submit, |
| 1537 | &rq->submitq); |
| 1538 | else |
| 1539 | __i915_sw_fence_await_dma_fence(&rq->submit, |
| 1540 | &prev->fence, |
| 1541 | &rq->dmaq); |
| 1542 | if (rq->engine->schedule) |
| 1543 | __i915_sched_node_add_dependency(&rq->sched, |
| 1544 | &prev->sched, |
| 1545 | &rq->dep, |
| 1546 | 0); |
| 1547 | } |
| 1548 | |
| 1549 | /* |
| 1550 | * Make sure that no request gazumped us - if it was allocated after |
| 1551 | * our i915_request_alloc() and called __i915_request_add() before |
| 1552 | * us, the timeline will hold its seqno which is later than ours. |
| 1553 | */ |
| 1554 | GEM_BUG_ON(timeline->seqno != rq->fence.seqno); |
| 1555 | |
| 1556 | return prev; |
| 1557 | } |
| 1558 | |
| 1559 | /* |
| 1560 | * NB: This function is not allowed to fail. Doing so would mean the the |
| 1561 | * request is not being tracked for completion but the work itself is |
| 1562 | * going to happen on the hardware. This would be a Bad Thing(tm). |
| 1563 | */ |
| 1564 | struct i915_request *__i915_request_commit(struct i915_request *rq) |
| 1565 | { |
| 1566 | struct intel_engine_cs *engine = rq->engine; |
| 1567 | struct intel_ring *ring = rq->ring; |
| 1568 | u32 *cs; |
| 1569 | |
| 1570 | RQ_TRACE(rq, "\n"); |
| 1571 | |
| 1572 | /* |
| 1573 | * To ensure that this call will not fail, space for its emissions |
| 1574 | * should already have been reserved in the ring buffer. Let the ring |
| 1575 | * know that it is time to use that space up. |
| 1576 | */ |
| 1577 | GEM_BUG_ON(rq->reserved_space > ring->space); |
| 1578 | rq->reserved_space = 0; |
| 1579 | rq->emitted_jiffies = jiffies; |
| 1580 | |
| 1581 | /* |
| 1582 | * Record the position of the start of the breadcrumb so that |
| 1583 | * should we detect the updated seqno part-way through the |
| 1584 | * GPU processing the request, we never over-estimate the |
| 1585 | * position of the ring's HEAD. |
| 1586 | */ |
| 1587 | cs = intel_ring_begin(rq, engine->emit_fini_breadcrumb_dw); |
| 1588 | GEM_BUG_ON(IS_ERR(cs)); |
| 1589 | rq->postfix = intel_ring_offset(rq, cs); |
| 1590 | |
| 1591 | return __i915_request_add_to_timeline(rq); |
| 1592 | } |
| 1593 | |
| 1594 | void __i915_request_queue_bh(struct i915_request *rq) |
| 1595 | { |
| 1596 | i915_sw_fence_commit(&rq->semaphore); |
| 1597 | i915_sw_fence_commit(&rq->submit); |
| 1598 | } |
| 1599 | |
| 1600 | void __i915_request_queue(struct i915_request *rq, |
| 1601 | const struct i915_sched_attr *attr) |
| 1602 | { |
| 1603 | /* |
| 1604 | * Let the backend know a new request has arrived that may need |
| 1605 | * to adjust the existing execution schedule due to a high priority |
| 1606 | * request - i.e. we may want to preempt the current request in order |
| 1607 | * to run a high priority dependency chain *before* we can execute this |
| 1608 | * request. |
| 1609 | * |
| 1610 | * This is called before the request is ready to run so that we can |
| 1611 | * decide whether to preempt the entire chain so that it is ready to |
| 1612 | * run at the earliest possible convenience. |
| 1613 | */ |
| 1614 | if (attr && rq->engine->schedule) |
| 1615 | rq->engine->schedule(rq, attr); |
| 1616 | |
| 1617 | local_bh_disable(); |
| 1618 | __i915_request_queue_bh(rq); |
| 1619 | local_bh_enable(); /* kick tasklets */ |
| 1620 | } |
| 1621 | |
| 1622 | void i915_request_add(struct i915_request *rq) |
| 1623 | { |
| 1624 | struct intel_timeline * const tl = i915_request_timeline(rq); |
| 1625 | struct i915_sched_attr attr = {}; |
| 1626 | struct i915_gem_context *ctx; |
| 1627 | |
| 1628 | lockdep_assert_held(&tl->mutex); |
| 1629 | lockdep_unpin_lock(&tl->mutex, rq->cookie); |
| 1630 | |
| 1631 | trace_i915_request_add(rq); |
| 1632 | __i915_request_commit(rq); |
| 1633 | |
| 1634 | /* XXX placeholder for selftests */ |
| 1635 | rcu_read_lock(); |
| 1636 | ctx = rcu_dereference(rq->context->gem_context); |
| 1637 | if (ctx) |
| 1638 | attr = ctx->sched; |
| 1639 | rcu_read_unlock(); |
| 1640 | |
| 1641 | __i915_request_queue(rq, &attr); |
| 1642 | |
| 1643 | mutex_unlock(&tl->mutex); |
| 1644 | } |
| 1645 | |
| 1646 | static unsigned long local_clock_ns(unsigned int *cpu) |
| 1647 | { |
| 1648 | unsigned long t; |
| 1649 | |
| 1650 | /* |
| 1651 | * Cheaply and approximately convert from nanoseconds to microseconds. |
| 1652 | * The result and subsequent calculations are also defined in the same |
| 1653 | * approximate microseconds units. The principal source of timing |
| 1654 | * error here is from the simple truncation. |
| 1655 | * |
| 1656 | * Note that local_clock() is only defined wrt to the current CPU; |
| 1657 | * the comparisons are no longer valid if we switch CPUs. Instead of |
| 1658 | * blocking preemption for the entire busywait, we can detect the CPU |
| 1659 | * switch and use that as indicator of system load and a reason to |
| 1660 | * stop busywaiting, see busywait_stop(). |
| 1661 | */ |
| 1662 | *cpu = get_cpu(); |
| 1663 | t = local_clock(); |
| 1664 | put_cpu(); |
| 1665 | |
| 1666 | return t; |
| 1667 | } |
| 1668 | |
| 1669 | static bool busywait_stop(unsigned long timeout, unsigned int cpu) |
| 1670 | { |
| 1671 | unsigned int this_cpu; |
| 1672 | |
| 1673 | if (time_after(local_clock_ns(&this_cpu), timeout)) |
| 1674 | return true; |
| 1675 | |
| 1676 | return this_cpu != cpu; |
| 1677 | } |
| 1678 | |
| 1679 | static bool __i915_spin_request(struct i915_request * const rq, int state) |
| 1680 | { |
| 1681 | unsigned long timeout_ns; |
| 1682 | unsigned int cpu; |
| 1683 | |
| 1684 | /* |
| 1685 | * Only wait for the request if we know it is likely to complete. |
| 1686 | * |
| 1687 | * We don't track the timestamps around requests, nor the average |
| 1688 | * request length, so we do not have a good indicator that this |
| 1689 | * request will complete within the timeout. What we do know is the |
| 1690 | * order in which requests are executed by the context and so we can |
| 1691 | * tell if the request has been started. If the request is not even |
| 1692 | * running yet, it is a fair assumption that it will not complete |
| 1693 | * within our relatively short timeout. |
| 1694 | */ |
| 1695 | if (!i915_request_is_running(rq)) |
| 1696 | return false; |
| 1697 | |
| 1698 | /* |
| 1699 | * When waiting for high frequency requests, e.g. during synchronous |
| 1700 | * rendering split between the CPU and GPU, the finite amount of time |
| 1701 | * required to set up the irq and wait upon it limits the response |
| 1702 | * rate. By busywaiting on the request completion for a short while we |
| 1703 | * can service the high frequency waits as quick as possible. However, |
| 1704 | * if it is a slow request, we want to sleep as quickly as possible. |
| 1705 | * The tradeoff between waiting and sleeping is roughly the time it |
| 1706 | * takes to sleep on a request, on the order of a microsecond. |
| 1707 | */ |
| 1708 | |
| 1709 | timeout_ns = READ_ONCE(rq->engine->props.max_busywait_duration_ns); |
| 1710 | timeout_ns += local_clock_ns(&cpu); |
| 1711 | do { |
| 1712 | if (dma_fence_is_signaled(&rq->fence)) |
| 1713 | return true; |
| 1714 | |
| 1715 | if (signal_pending_state(state, current)) |
| 1716 | break; |
| 1717 | |
| 1718 | if (busywait_stop(timeout_ns, cpu)) |
| 1719 | break; |
| 1720 | |
| 1721 | cpu_relax(); |
| 1722 | } while (!need_resched()); |
| 1723 | |
| 1724 | return false; |
| 1725 | } |
| 1726 | |
| 1727 | struct request_wait { |
| 1728 | struct dma_fence_cb cb; |
| 1729 | struct task_struct *tsk; |
| 1730 | }; |
| 1731 | |
| 1732 | static void request_wait_wake(struct dma_fence *fence, struct dma_fence_cb *cb) |
| 1733 | { |
| 1734 | struct request_wait *wait = container_of(cb, typeof(*wait), cb); |
| 1735 | |
| 1736 | wake_up_process(fetch_and_zero(&wait->tsk)); |
| 1737 | } |
| 1738 | |
| 1739 | /** |
| 1740 | * i915_request_wait - wait until execution of request has finished |
| 1741 | * @rq: the request to wait upon |
| 1742 | * @flags: how to wait |
| 1743 | * @timeout: how long to wait in jiffies |
| 1744 | * |
| 1745 | * i915_request_wait() waits for the request to be completed, for a |
| 1746 | * maximum of @timeout jiffies (with MAX_SCHEDULE_TIMEOUT implying an |
| 1747 | * unbounded wait). |
| 1748 | * |
| 1749 | * Returns the remaining time (in jiffies) if the request completed, which may |
| 1750 | * be zero or -ETIME if the request is unfinished after the timeout expires. |
| 1751 | * May return -EINTR is called with I915_WAIT_INTERRUPTIBLE and a signal is |
| 1752 | * pending before the request completes. |
| 1753 | */ |
| 1754 | long i915_request_wait(struct i915_request *rq, |
| 1755 | unsigned int flags, |
| 1756 | long timeout) |
| 1757 | { |
| 1758 | const int state = flags & I915_WAIT_INTERRUPTIBLE ? |
| 1759 | TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE; |
| 1760 | struct request_wait wait; |
| 1761 | |
| 1762 | might_sleep(); |
| 1763 | GEM_BUG_ON(timeout < 0); |
| 1764 | |
| 1765 | if (dma_fence_is_signaled(&rq->fence)) |
| 1766 | return timeout; |
| 1767 | |
| 1768 | if (!timeout) |
| 1769 | return -ETIME; |
| 1770 | |
| 1771 | trace_i915_request_wait_begin(rq, flags); |
| 1772 | |
| 1773 | /* |
| 1774 | * We must never wait on the GPU while holding a lock as we |
| 1775 | * may need to perform a GPU reset. So while we don't need to |
| 1776 | * serialise wait/reset with an explicit lock, we do want |
| 1777 | * lockdep to detect potential dependency cycles. |
| 1778 | */ |
| 1779 | mutex_acquire(&rq->engine->gt->reset.mutex.dep_map, 0, 0, _THIS_IP_); |
| 1780 | |
| 1781 | /* |
| 1782 | * Optimistic spin before touching IRQs. |
| 1783 | * |
| 1784 | * We may use a rather large value here to offset the penalty of |
| 1785 | * switching away from the active task. Frequently, the client will |
| 1786 | * wait upon an old swapbuffer to throttle itself to remain within a |
| 1787 | * frame of the gpu. If the client is running in lockstep with the gpu, |
| 1788 | * then it should not be waiting long at all, and a sleep now will incur |
| 1789 | * extra scheduler latency in producing the next frame. To try to |
| 1790 | * avoid adding the cost of enabling/disabling the interrupt to the |
| 1791 | * short wait, we first spin to see if the request would have completed |
| 1792 | * in the time taken to setup the interrupt. |
| 1793 | * |
| 1794 | * We need upto 5us to enable the irq, and upto 20us to hide the |
| 1795 | * scheduler latency of a context switch, ignoring the secondary |
| 1796 | * impacts from a context switch such as cache eviction. |
| 1797 | * |
| 1798 | * The scheme used for low-latency IO is called "hybrid interrupt |
| 1799 | * polling". The suggestion there is to sleep until just before you |
| 1800 | * expect to be woken by the device interrupt and then poll for its |
| 1801 | * completion. That requires having a good predictor for the request |
| 1802 | * duration, which we currently lack. |
| 1803 | */ |
| 1804 | if (IS_ACTIVE(CONFIG_DRM_I915_MAX_REQUEST_BUSYWAIT) && |
| 1805 | __i915_spin_request(rq, state)) |
| 1806 | goto out; |
| 1807 | |
| 1808 | /* |
| 1809 | * This client is about to stall waiting for the GPU. In many cases |
| 1810 | * this is undesirable and limits the throughput of the system, as |
| 1811 | * many clients cannot continue processing user input/output whilst |
| 1812 | * blocked. RPS autotuning may take tens of milliseconds to respond |
| 1813 | * to the GPU load and thus incurs additional latency for the client. |
| 1814 | * We can circumvent that by promoting the GPU frequency to maximum |
| 1815 | * before we sleep. This makes the GPU throttle up much more quickly |
| 1816 | * (good for benchmarks and user experience, e.g. window animations), |
| 1817 | * but at a cost of spending more power processing the workload |
| 1818 | * (bad for battery). |
| 1819 | */ |
| 1820 | if (flags & I915_WAIT_PRIORITY && !i915_request_started(rq)) |
| 1821 | intel_rps_boost(rq); |
| 1822 | |
| 1823 | wait.tsk = current; |
| 1824 | if (dma_fence_add_callback(&rq->fence, &wait.cb, request_wait_wake)) |
| 1825 | goto out; |
| 1826 | |
| 1827 | /* |
| 1828 | * Flush the submission tasklet, but only if it may help this request. |
| 1829 | * |
| 1830 | * We sometimes experience some latency between the HW interrupts and |
| 1831 | * tasklet execution (mostly due to ksoftirqd latency, but it can also |
| 1832 | * be due to lazy CS events), so lets run the tasklet manually if there |
| 1833 | * is a chance it may submit this request. If the request is not ready |
| 1834 | * to run, as it is waiting for other fences to be signaled, flushing |
| 1835 | * the tasklet is busy work without any advantage for this client. |
| 1836 | * |
| 1837 | * If the HW is being lazy, this is the last chance before we go to |
| 1838 | * sleep to catch any pending events. We will check periodically in |
| 1839 | * the heartbeat to flush the submission tasklets as a last resort |
| 1840 | * for unhappy HW. |
| 1841 | */ |
| 1842 | if (i915_request_is_ready(rq)) |
| 1843 | __intel_engine_flush_submission(rq->engine, false); |
| 1844 | |
| 1845 | for (;;) { |
| 1846 | set_current_state(state); |
| 1847 | |
| 1848 | if (dma_fence_is_signaled(&rq->fence)) |
| 1849 | break; |
| 1850 | |
| 1851 | if (signal_pending_state(state, current)) { |
| 1852 | timeout = -ERESTARTSYS; |
| 1853 | break; |
| 1854 | } |
| 1855 | |
| 1856 | if (!timeout) { |
| 1857 | timeout = -ETIME; |
| 1858 | break; |
| 1859 | } |
| 1860 | |
| 1861 | timeout = io_schedule_timeout(timeout); |
| 1862 | } |
| 1863 | __set_current_state(TASK_RUNNING); |
| 1864 | |
| 1865 | if (READ_ONCE(wait.tsk)) |
| 1866 | dma_fence_remove_callback(&rq->fence, &wait.cb); |
| 1867 | GEM_BUG_ON(!list_empty(&wait.cb.node)); |
| 1868 | |
| 1869 | out: |
| 1870 | mutex_release(&rq->engine->gt->reset.mutex.dep_map, _THIS_IP_); |
| 1871 | trace_i915_request_wait_end(rq); |
| 1872 | return timeout; |
| 1873 | } |
| 1874 | |
| 1875 | static int print_sched_attr(const struct i915_sched_attr *attr, |
| 1876 | char *buf, int x, int len) |
| 1877 | { |
| 1878 | if (attr->priority == I915_PRIORITY_INVALID) |
| 1879 | return x; |
| 1880 | |
| 1881 | x += snprintf(buf + x, len - x, |
| 1882 | " prio=%d", attr->priority); |
| 1883 | |
| 1884 | return x; |
| 1885 | } |
| 1886 | |
| 1887 | static char queue_status(const struct i915_request *rq) |
| 1888 | { |
| 1889 | if (i915_request_is_active(rq)) |
| 1890 | return 'E'; |
| 1891 | |
| 1892 | if (i915_request_is_ready(rq)) |
| 1893 | return intel_engine_is_virtual(rq->engine) ? 'V' : 'R'; |
| 1894 | |
| 1895 | return 'U'; |
| 1896 | } |
| 1897 | |
| 1898 | static const char *run_status(const struct i915_request *rq) |
| 1899 | { |
| 1900 | if (i915_request_completed(rq)) |
| 1901 | return "!"; |
| 1902 | |
| 1903 | if (i915_request_started(rq)) |
| 1904 | return "*"; |
| 1905 | |
| 1906 | if (!i915_sw_fence_signaled(&rq->semaphore)) |
| 1907 | return "&"; |
| 1908 | |
| 1909 | return ""; |
| 1910 | } |
| 1911 | |
| 1912 | static const char *fence_status(const struct i915_request *rq) |
| 1913 | { |
| 1914 | if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &rq->fence.flags)) |
| 1915 | return "+"; |
| 1916 | |
| 1917 | if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &rq->fence.flags)) |
| 1918 | return "-"; |
| 1919 | |
| 1920 | return ""; |
| 1921 | } |
| 1922 | |
| 1923 | void i915_request_show(struct drm_printer *m, |
| 1924 | const struct i915_request *rq, |
| 1925 | const char *prefix, |
| 1926 | int indent) |
| 1927 | { |
| 1928 | const char *name = rq->fence.ops->get_timeline_name((struct dma_fence *)&rq->fence); |
| 1929 | char buf[80] = ""; |
| 1930 | int x = 0; |
| 1931 | |
| 1932 | /* |
| 1933 | * The prefix is used to show the queue status, for which we use |
| 1934 | * the following flags: |
| 1935 | * |
| 1936 | * U [Unready] |
| 1937 | * - initial status upon being submitted by the user |
| 1938 | * |
| 1939 | * - the request is not ready for execution as it is waiting |
| 1940 | * for external fences |
| 1941 | * |
| 1942 | * R [Ready] |
| 1943 | * - all fences the request was waiting on have been signaled, |
| 1944 | * and the request is now ready for execution and will be |
| 1945 | * in a backend queue |
| 1946 | * |
| 1947 | * - a ready request may still need to wait on semaphores |
| 1948 | * [internal fences] |
| 1949 | * |
| 1950 | * V [Ready/virtual] |
| 1951 | * - same as ready, but queued over multiple backends |
| 1952 | * |
| 1953 | * E [Executing] |
| 1954 | * - the request has been transferred from the backend queue and |
| 1955 | * submitted for execution on HW |
| 1956 | * |
| 1957 | * - a completed request may still be regarded as executing, its |
| 1958 | * status may not be updated until it is retired and removed |
| 1959 | * from the lists |
| 1960 | */ |
| 1961 | |
| 1962 | x = print_sched_attr(&rq->sched.attr, buf, x, sizeof(buf)); |
| 1963 | |
| 1964 | drm_printf(m, "%s%.*s%c %llx:%lld%s%s %s @ %dms: %s\n", |
| 1965 | prefix, indent, " ", |
| 1966 | queue_status(rq), |
| 1967 | rq->fence.context, rq->fence.seqno, |
| 1968 | run_status(rq), |
| 1969 | fence_status(rq), |
| 1970 | buf, |
| 1971 | jiffies_to_msecs(jiffies - rq->emitted_jiffies), |
| 1972 | name); |
| 1973 | } |
| 1974 | |
| 1975 | #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) |
| 1976 | #include "selftests/mock_request.c" |
| 1977 | #include "selftests/i915_request.c" |
| 1978 | #endif |
| 1979 | |
| 1980 | static void i915_global_request_shrink(void) |
| 1981 | { |
| 1982 | kmem_cache_shrink(global.slab_execute_cbs); |
| 1983 | kmem_cache_shrink(global.slab_requests); |
| 1984 | } |
| 1985 | |
| 1986 | static void i915_global_request_exit(void) |
| 1987 | { |
| 1988 | kmem_cache_destroy(global.slab_execute_cbs); |
| 1989 | kmem_cache_destroy(global.slab_requests); |
| 1990 | } |
| 1991 | |
| 1992 | static struct i915_global_request global = { { |
| 1993 | .shrink = i915_global_request_shrink, |
| 1994 | .exit = i915_global_request_exit, |
| 1995 | } }; |
| 1996 | |
| 1997 | int __init i915_global_request_init(void) |
| 1998 | { |
| 1999 | global.slab_requests = |
| 2000 | kmem_cache_create("i915_request", |
| 2001 | sizeof(struct i915_request), |
| 2002 | __alignof__(struct i915_request), |
| 2003 | SLAB_HWCACHE_ALIGN | |
| 2004 | SLAB_RECLAIM_ACCOUNT | |
| 2005 | SLAB_TYPESAFE_BY_RCU, |
| 2006 | __i915_request_ctor); |
| 2007 | if (!global.slab_requests) |
| 2008 | return -ENOMEM; |
| 2009 | |
| 2010 | global.slab_execute_cbs = KMEM_CACHE(execute_cb, |
| 2011 | SLAB_HWCACHE_ALIGN | |
| 2012 | SLAB_RECLAIM_ACCOUNT | |
| 2013 | SLAB_TYPESAFE_BY_RCU); |
| 2014 | if (!global.slab_execute_cbs) |
| 2015 | goto err_requests; |
| 2016 | |
| 2017 | i915_global_register(&global.base); |
| 2018 | return 0; |
| 2019 | |
| 2020 | err_requests: |
| 2021 | kmem_cache_destroy(global.slab_requests); |
| 2022 | return -ENOMEM; |
| 2023 | } |