| 1 | // SPDX-License-Identifier: MIT |
| 2 | /* |
| 3 | * Copyright © 2016-2018 Intel Corporation |
| 4 | */ |
| 5 | |
| 6 | #include <drm/drm_cache.h> |
| 7 | #include <drm/drm_print.h> |
| 8 | |
| 9 | #include "gem/i915_gem_internal.h" |
| 10 | |
| 11 | #include "i915_active.h" |
| 12 | #include "i915_drv.h" |
| 13 | #include "i915_syncmap.h" |
| 14 | #include "intel_gt.h" |
| 15 | #include "intel_ring.h" |
| 16 | #include "intel_timeline.h" |
| 17 | |
| 18 | #define TIMELINE_SEQNO_BYTES 8 |
| 19 | |
| 20 | static struct i915_vma *hwsp_alloc(struct intel_gt *gt) |
| 21 | { |
| 22 | struct drm_i915_private *i915 = gt->i915; |
| 23 | struct drm_i915_gem_object *obj; |
| 24 | struct i915_vma *vma; |
| 25 | |
| 26 | obj = i915_gem_object_create_internal(i915, PAGE_SIZE); |
| 27 | if (IS_ERR(ptr: obj)) |
| 28 | return ERR_CAST(ptr: obj); |
| 29 | |
| 30 | i915_gem_object_set_cache_coherency(obj, cache_level: I915_CACHE_LLC); |
| 31 | |
| 32 | vma = i915_vma_instance(obj, vm: >->ggtt->vm, NULL); |
| 33 | if (IS_ERR(ptr: vma)) |
| 34 | i915_gem_object_put(obj); |
| 35 | |
| 36 | return vma; |
| 37 | } |
| 38 | |
| 39 | static void __timeline_retire(struct i915_active *active) |
| 40 | { |
| 41 | struct intel_timeline *tl = |
| 42 | container_of(active, typeof(*tl), active); |
| 43 | |
| 44 | i915_vma_unpin(vma: tl->hwsp_ggtt); |
| 45 | intel_timeline_put(timeline: tl); |
| 46 | } |
| 47 | |
| 48 | static int __timeline_active(struct i915_active *active) |
| 49 | { |
| 50 | struct intel_timeline *tl = |
| 51 | container_of(active, typeof(*tl), active); |
| 52 | |
| 53 | __i915_vma_pin(vma: tl->hwsp_ggtt); |
| 54 | intel_timeline_get(timeline: tl); |
| 55 | return 0; |
| 56 | } |
| 57 | |
| 58 | I915_SELFTEST_EXPORT int |
| 59 | intel_timeline_pin_map(struct intel_timeline *timeline) |
| 60 | { |
| 61 | struct drm_i915_gem_object *obj = timeline->hwsp_ggtt->obj; |
| 62 | u32 ofs = offset_in_page(timeline->hwsp_offset); |
| 63 | void *vaddr; |
| 64 | |
| 65 | vaddr = i915_gem_object_pin_map(obj, type: I915_MAP_WB); |
| 66 | if (IS_ERR(ptr: vaddr)) |
| 67 | return PTR_ERR(ptr: vaddr); |
| 68 | |
| 69 | timeline->hwsp_map = vaddr; |
| 70 | timeline->hwsp_seqno = memset(vaddr + ofs, 0, TIMELINE_SEQNO_BYTES); |
| 71 | drm_clflush_virt_range(addr: vaddr + ofs, TIMELINE_SEQNO_BYTES); |
| 72 | |
| 73 | return 0; |
| 74 | } |
| 75 | |
| 76 | static int intel_timeline_init(struct intel_timeline *timeline, |
| 77 | struct intel_gt *gt, |
| 78 | struct i915_vma *hwsp, |
| 79 | unsigned int offset) |
| 80 | { |
| 81 | kref_init(kref: &timeline->kref); |
| 82 | atomic_set(v: &timeline->pin_count, i: 0); |
| 83 | |
| 84 | timeline->gt = gt; |
| 85 | |
| 86 | if (hwsp) { |
| 87 | timeline->hwsp_offset = offset; |
| 88 | timeline->hwsp_ggtt = i915_vma_get(vma: hwsp); |
| 89 | } else { |
| 90 | timeline->has_initial_breadcrumb = true; |
| 91 | hwsp = hwsp_alloc(gt); |
| 92 | if (IS_ERR(ptr: hwsp)) |
| 93 | return PTR_ERR(ptr: hwsp); |
| 94 | timeline->hwsp_ggtt = hwsp; |
| 95 | } |
| 96 | |
| 97 | timeline->hwsp_map = NULL; |
| 98 | timeline->hwsp_seqno = (void *)(long)timeline->hwsp_offset; |
| 99 | |
| 100 | GEM_BUG_ON(timeline->hwsp_offset >= hwsp->size); |
| 101 | |
| 102 | timeline->fence_context = dma_fence_context_alloc(num: 1); |
| 103 | |
| 104 | mutex_init(&timeline->mutex); |
| 105 | |
| 106 | INIT_ACTIVE_FENCE(&timeline->last_request); |
| 107 | INIT_LIST_HEAD(list: &timeline->requests); |
| 108 | |
| 109 | i915_syncmap_init(root: &timeline->sync); |
| 110 | i915_active_init(&timeline->active, __timeline_active, |
| 111 | __timeline_retire, 0); |
| 112 | |
| 113 | return 0; |
| 114 | } |
| 115 | |
| 116 | void intel_gt_init_timelines(struct intel_gt *gt) |
| 117 | { |
| 118 | struct intel_gt_timelines *timelines = >->timelines; |
| 119 | |
| 120 | spin_lock_init(&timelines->lock); |
| 121 | INIT_LIST_HEAD(list: &timelines->active_list); |
| 122 | } |
| 123 | |
| 124 | static void intel_timeline_fini(struct rcu_head *rcu) |
| 125 | { |
| 126 | struct intel_timeline *timeline = |
| 127 | container_of(rcu, struct intel_timeline, rcu); |
| 128 | |
| 129 | if (timeline->hwsp_map) |
| 130 | i915_gem_object_unpin_map(obj: timeline->hwsp_ggtt->obj); |
| 131 | |
| 132 | i915_vma_put(vma: timeline->hwsp_ggtt); |
| 133 | i915_active_fini(ref: &timeline->active); |
| 134 | |
| 135 | /* |
| 136 | * A small race exists between intel_gt_retire_requests_timeout and |
| 137 | * intel_timeline_exit which could result in the syncmap not getting |
| 138 | * free'd. Rather than work to hard to seal this race, simply cleanup |
| 139 | * the syncmap on fini. |
| 140 | */ |
| 141 | i915_syncmap_free(root: &timeline->sync); |
| 142 | |
| 143 | kfree(objp: timeline); |
| 144 | } |
| 145 | |
| 146 | struct intel_timeline * |
| 147 | __intel_timeline_create(struct intel_gt *gt, |
| 148 | struct i915_vma *global_hwsp, |
| 149 | unsigned int offset) |
| 150 | { |
| 151 | struct intel_timeline *timeline; |
| 152 | int err; |
| 153 | |
| 154 | timeline = kzalloc(sizeof(*timeline), GFP_KERNEL); |
| 155 | if (!timeline) |
| 156 | return ERR_PTR(error: -ENOMEM); |
| 157 | |
| 158 | err = intel_timeline_init(timeline, gt, hwsp: global_hwsp, offset); |
| 159 | if (err) { |
| 160 | kfree(objp: timeline); |
| 161 | return ERR_PTR(error: err); |
| 162 | } |
| 163 | |
| 164 | return timeline; |
| 165 | } |
| 166 | |
| 167 | struct intel_timeline * |
| 168 | intel_timeline_create_from_engine(struct intel_engine_cs *engine, |
| 169 | unsigned int offset) |
| 170 | { |
| 171 | struct i915_vma *hwsp = engine->status_page.vma; |
| 172 | struct intel_timeline *tl; |
| 173 | |
| 174 | tl = __intel_timeline_create(gt: engine->gt, global_hwsp: hwsp, offset); |
| 175 | if (IS_ERR(ptr: tl)) |
| 176 | return tl; |
| 177 | |
| 178 | /* Borrow a nearby lock; we only create these timelines during init */ |
| 179 | mutex_lock(&hwsp->vm->mutex); |
| 180 | list_add_tail(new: &tl->engine_link, head: &engine->status_page.timelines); |
| 181 | mutex_unlock(lock: &hwsp->vm->mutex); |
| 182 | |
| 183 | return tl; |
| 184 | } |
| 185 | |
| 186 | void __intel_timeline_pin(struct intel_timeline *tl) |
| 187 | { |
| 188 | GEM_BUG_ON(!atomic_read(&tl->pin_count)); |
| 189 | atomic_inc(v: &tl->pin_count); |
| 190 | } |
| 191 | |
| 192 | int intel_timeline_pin(struct intel_timeline *tl, struct i915_gem_ww_ctx *ww) |
| 193 | { |
| 194 | int err; |
| 195 | |
| 196 | if (atomic_add_unless(v: &tl->pin_count, a: 1, u: 0)) |
| 197 | return 0; |
| 198 | |
| 199 | if (!tl->hwsp_map) { |
| 200 | err = intel_timeline_pin_map(timeline: tl); |
| 201 | if (err) |
| 202 | return err; |
| 203 | } |
| 204 | |
| 205 | err = i915_ggtt_pin(vma: tl->hwsp_ggtt, ww, align: 0, PIN_HIGH); |
| 206 | if (err) |
| 207 | return err; |
| 208 | |
| 209 | tl->hwsp_offset = |
| 210 | i915_ggtt_offset(vma: tl->hwsp_ggtt) + |
| 211 | offset_in_page(tl->hwsp_offset); |
| 212 | GT_TRACE(tl->gt, "timeline:%llx using HWSP offset:%x\n" , |
| 213 | tl->fence_context, tl->hwsp_offset); |
| 214 | |
| 215 | i915_active_acquire(ref: &tl->active); |
| 216 | if (atomic_fetch_inc(v: &tl->pin_count)) { |
| 217 | i915_active_release(ref: &tl->active); |
| 218 | __i915_vma_unpin(vma: tl->hwsp_ggtt); |
| 219 | } |
| 220 | |
| 221 | return 0; |
| 222 | } |
| 223 | |
| 224 | void intel_timeline_reset_seqno(const struct intel_timeline *tl) |
| 225 | { |
| 226 | u32 *hwsp_seqno = (u32 *)tl->hwsp_seqno; |
| 227 | /* Must be pinned to be writable, and no requests in flight. */ |
| 228 | GEM_BUG_ON(!atomic_read(&tl->pin_count)); |
| 229 | |
| 230 | memset(hwsp_seqno + 1, 0, TIMELINE_SEQNO_BYTES - sizeof(*hwsp_seqno)); |
| 231 | WRITE_ONCE(*hwsp_seqno, tl->seqno); |
| 232 | drm_clflush_virt_range(addr: hwsp_seqno, TIMELINE_SEQNO_BYTES); |
| 233 | } |
| 234 | |
| 235 | void intel_timeline_enter(struct intel_timeline *tl) |
| 236 | { |
| 237 | struct intel_gt_timelines *timelines = &tl->gt->timelines; |
| 238 | |
| 239 | /* |
| 240 | * Pretend we are serialised by the timeline->mutex. |
| 241 | * |
| 242 | * While generally true, there are a few exceptions to the rule |
| 243 | * for the engine->kernel_context being used to manage power |
| 244 | * transitions. As the engine_park may be called from under any |
| 245 | * timeline, it uses the power mutex as a global serialisation |
| 246 | * lock to prevent any other request entering its timeline. |
| 247 | * |
| 248 | * The rule is generally tl->mutex, otherwise engine->wakeref.mutex. |
| 249 | * |
| 250 | * However, intel_gt_retire_request() does not know which engine |
| 251 | * it is retiring along and so cannot partake in the engine-pm |
| 252 | * barrier, and there we use the tl->active_count as a means to |
| 253 | * pin the timeline in the active_list while the locks are dropped. |
| 254 | * Ergo, as that is outside of the engine-pm barrier, we need to |
| 255 | * use atomic to manipulate tl->active_count. |
| 256 | */ |
| 257 | lockdep_assert_held(&tl->mutex); |
| 258 | |
| 259 | if (atomic_add_unless(v: &tl->active_count, a: 1, u: 0)) |
| 260 | return; |
| 261 | |
| 262 | spin_lock(lock: &timelines->lock); |
| 263 | if (!atomic_fetch_inc(v: &tl->active_count)) { |
| 264 | /* |
| 265 | * The HWSP is volatile, and may have been lost while inactive, |
| 266 | * e.g. across suspend/resume. Be paranoid, and ensure that |
| 267 | * the HWSP value matches our seqno so we don't proclaim |
| 268 | * the next request as already complete. |
| 269 | */ |
| 270 | intel_timeline_reset_seqno(tl); |
| 271 | list_add_tail(new: &tl->link, head: &timelines->active_list); |
| 272 | } |
| 273 | spin_unlock(lock: &timelines->lock); |
| 274 | } |
| 275 | |
| 276 | void intel_timeline_exit(struct intel_timeline *tl) |
| 277 | { |
| 278 | struct intel_gt_timelines *timelines = &tl->gt->timelines; |
| 279 | |
| 280 | /* See intel_timeline_enter() */ |
| 281 | lockdep_assert_held(&tl->mutex); |
| 282 | |
| 283 | GEM_BUG_ON(!atomic_read(&tl->active_count)); |
| 284 | if (atomic_add_unless(v: &tl->active_count, a: -1, u: 1)) |
| 285 | return; |
| 286 | |
| 287 | spin_lock(lock: &timelines->lock); |
| 288 | if (atomic_dec_and_test(v: &tl->active_count)) |
| 289 | list_del(entry: &tl->link); |
| 290 | spin_unlock(lock: &timelines->lock); |
| 291 | |
| 292 | /* |
| 293 | * Since this timeline is idle, all bariers upon which we were waiting |
| 294 | * must also be complete and so we can discard the last used barriers |
| 295 | * without loss of information. |
| 296 | */ |
| 297 | i915_syncmap_free(root: &tl->sync); |
| 298 | } |
| 299 | |
| 300 | static u32 timeline_advance(struct intel_timeline *tl) |
| 301 | { |
| 302 | GEM_BUG_ON(!atomic_read(&tl->pin_count)); |
| 303 | GEM_BUG_ON(tl->seqno & tl->has_initial_breadcrumb); |
| 304 | |
| 305 | return tl->seqno += 1 + tl->has_initial_breadcrumb; |
| 306 | } |
| 307 | |
| 308 | static noinline int |
| 309 | __intel_timeline_get_seqno(struct intel_timeline *tl, |
| 310 | u32 *seqno) |
| 311 | { |
| 312 | u32 next_ofs = offset_in_page(tl->hwsp_offset + TIMELINE_SEQNO_BYTES); |
| 313 | |
| 314 | /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */ |
| 315 | if (TIMELINE_SEQNO_BYTES <= BIT(5) && (next_ofs & BIT(5))) |
| 316 | next_ofs = offset_in_page(next_ofs + BIT(5)); |
| 317 | |
| 318 | tl->hwsp_offset = i915_ggtt_offset(vma: tl->hwsp_ggtt) + next_ofs; |
| 319 | tl->hwsp_seqno = tl->hwsp_map + next_ofs; |
| 320 | intel_timeline_reset_seqno(tl); |
| 321 | |
| 322 | *seqno = timeline_advance(tl); |
| 323 | GEM_BUG_ON(i915_seqno_passed(*tl->hwsp_seqno, *seqno)); |
| 324 | return 0; |
| 325 | } |
| 326 | |
| 327 | int intel_timeline_get_seqno(struct intel_timeline *tl, |
| 328 | struct i915_request *rq, |
| 329 | u32 *seqno) |
| 330 | { |
| 331 | *seqno = timeline_advance(tl); |
| 332 | |
| 333 | /* Replace the HWSP on wraparound for HW semaphores */ |
| 334 | if (unlikely(!*seqno && tl->has_initial_breadcrumb)) |
| 335 | return __intel_timeline_get_seqno(tl, seqno); |
| 336 | |
| 337 | return 0; |
| 338 | } |
| 339 | |
| 340 | int intel_timeline_read_hwsp(struct i915_request *from, |
| 341 | struct i915_request *to, |
| 342 | u32 *hwsp) |
| 343 | { |
| 344 | struct intel_timeline *tl; |
| 345 | int err; |
| 346 | |
| 347 | rcu_read_lock(); |
| 348 | tl = rcu_dereference(from->timeline); |
| 349 | if (i915_request_signaled(rq: from) || |
| 350 | !i915_active_acquire_if_busy(ref: &tl->active)) |
| 351 | tl = NULL; |
| 352 | |
| 353 | if (tl) { |
| 354 | /* hwsp_offset may wraparound, so use from->hwsp_seqno */ |
| 355 | *hwsp = i915_ggtt_offset(vma: tl->hwsp_ggtt) + |
| 356 | offset_in_page(from->hwsp_seqno); |
| 357 | } |
| 358 | |
| 359 | /* ensure we wait on the right request, if not, we completed */ |
| 360 | if (tl && __i915_request_is_complete(rq: from)) { |
| 361 | i915_active_release(ref: &tl->active); |
| 362 | tl = NULL; |
| 363 | } |
| 364 | rcu_read_unlock(); |
| 365 | |
| 366 | if (!tl) |
| 367 | return 1; |
| 368 | |
| 369 | /* Can't do semaphore waits on kernel context */ |
| 370 | if (!tl->has_initial_breadcrumb) { |
| 371 | err = -EINVAL; |
| 372 | goto out; |
| 373 | } |
| 374 | |
| 375 | err = i915_active_add_request(ref: &tl->active, rq: to); |
| 376 | |
| 377 | out: |
| 378 | i915_active_release(ref: &tl->active); |
| 379 | return err; |
| 380 | } |
| 381 | |
| 382 | void intel_timeline_unpin(struct intel_timeline *tl) |
| 383 | { |
| 384 | GEM_BUG_ON(!atomic_read(&tl->pin_count)); |
| 385 | if (!atomic_dec_and_test(v: &tl->pin_count)) |
| 386 | return; |
| 387 | |
| 388 | i915_active_release(ref: &tl->active); |
| 389 | __i915_vma_unpin(vma: tl->hwsp_ggtt); |
| 390 | } |
| 391 | |
| 392 | void __intel_timeline_free(struct kref *kref) |
| 393 | { |
| 394 | struct intel_timeline *timeline = |
| 395 | container_of(kref, typeof(*timeline), kref); |
| 396 | |
| 397 | GEM_BUG_ON(atomic_read(&timeline->pin_count)); |
| 398 | GEM_BUG_ON(!list_empty(&timeline->requests)); |
| 399 | GEM_BUG_ON(timeline->retire); |
| 400 | |
| 401 | call_rcu(head: &timeline->rcu, func: intel_timeline_fini); |
| 402 | } |
| 403 | |
| 404 | void intel_gt_fini_timelines(struct intel_gt *gt) |
| 405 | { |
| 406 | struct intel_gt_timelines *timelines = >->timelines; |
| 407 | |
| 408 | GEM_BUG_ON(!list_empty(&timelines->active_list)); |
| 409 | } |
| 410 | |
| 411 | void intel_gt_show_timelines(struct intel_gt *gt, |
| 412 | struct drm_printer *m, |
| 413 | void (*show_request)(struct drm_printer *m, |
| 414 | const struct i915_request *rq, |
| 415 | const char *prefix, |
| 416 | int indent)) |
| 417 | { |
| 418 | struct intel_gt_timelines *timelines = >->timelines; |
| 419 | struct intel_timeline *tl, *tn; |
| 420 | LIST_HEAD(free); |
| 421 | |
| 422 | spin_lock(lock: &timelines->lock); |
| 423 | list_for_each_entry_safe(tl, tn, &timelines->active_list, link) { |
| 424 | unsigned long count, ready, inflight; |
| 425 | struct i915_request *rq, *rn; |
| 426 | struct dma_fence *fence; |
| 427 | |
| 428 | if (!mutex_trylock(&tl->mutex)) { |
| 429 | drm_printf(p: m, f: "Timeline %llx: busy; skipping\n" , |
| 430 | tl->fence_context); |
| 431 | continue; |
| 432 | } |
| 433 | |
| 434 | intel_timeline_get(timeline: tl); |
| 435 | GEM_BUG_ON(!atomic_read(&tl->active_count)); |
| 436 | atomic_inc(v: &tl->active_count); /* pin the list element */ |
| 437 | spin_unlock(lock: &timelines->lock); |
| 438 | |
| 439 | count = 0; |
| 440 | ready = 0; |
| 441 | inflight = 0; |
| 442 | list_for_each_entry_safe(rq, rn, &tl->requests, link) { |
| 443 | if (i915_request_completed(rq)) |
| 444 | continue; |
| 445 | |
| 446 | count++; |
| 447 | if (i915_request_is_ready(rq)) |
| 448 | ready++; |
| 449 | if (i915_request_is_active(rq)) |
| 450 | inflight++; |
| 451 | } |
| 452 | |
| 453 | drm_printf(p: m, f: "Timeline %llx: { " , tl->fence_context); |
| 454 | drm_printf(p: m, f: "count: %lu, ready: %lu, inflight: %lu" , |
| 455 | count, ready, inflight); |
| 456 | drm_printf(p: m, f: ", seqno: { current: %d, last: %d }" , |
| 457 | *tl->hwsp_seqno, tl->seqno); |
| 458 | fence = i915_active_fence_get(active: &tl->last_request); |
| 459 | if (fence) { |
| 460 | drm_printf(p: m, f: ", engine: %s" , |
| 461 | to_request(fence)->engine->name); |
| 462 | dma_fence_put(fence); |
| 463 | } |
| 464 | drm_printf(p: m, f: " }\n" ); |
| 465 | |
| 466 | if (show_request) { |
| 467 | list_for_each_entry_safe(rq, rn, &tl->requests, link) |
| 468 | show_request(m, rq, "" , 2); |
| 469 | } |
| 470 | |
| 471 | mutex_unlock(lock: &tl->mutex); |
| 472 | spin_lock(lock: &timelines->lock); |
| 473 | |
| 474 | /* Resume list iteration after reacquiring spinlock */ |
| 475 | list_safe_reset_next(tl, tn, link); |
| 476 | if (atomic_dec_and_test(v: &tl->active_count)) |
| 477 | list_del(entry: &tl->link); |
| 478 | |
| 479 | /* Defer the final release to after the spinlock */ |
| 480 | if (refcount_dec_and_test(r: &tl->kref.refcount)) { |
| 481 | GEM_BUG_ON(atomic_read(&tl->active_count)); |
| 482 | list_add(new: &tl->link, head: &free); |
| 483 | } |
| 484 | } |
| 485 | spin_unlock(lock: &timelines->lock); |
| 486 | |
| 487 | list_for_each_entry_safe(tl, tn, &free, link) |
| 488 | __intel_timeline_free(kref: &tl->kref); |
| 489 | } |
| 490 | |
| 491 | #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) |
| 492 | #include "gt/selftests/mock_timeline.c" |
| 493 | #include "gt/selftest_timeline.c" |
| 494 | #endif |
| 495 | |