| 1 | /* SPDX-License-Identifier: MIT */ |
| 2 | /* |
| 3 | * Copyright © 2019 Intel Corporation |
| 4 | */ |
| 5 | |
| 6 | #ifndef __INTEL_CONTEXT_H__ |
| 7 | #define __INTEL_CONTEXT_H__ |
| 8 | |
| 9 | #include <linux/bitops.h> |
| 10 | #include <linux/lockdep.h> |
| 11 | #include <linux/types.h> |
| 12 | |
| 13 | #include "i915_active.h" |
| 14 | #include "i915_drv.h" |
| 15 | #include "intel_context_types.h" |
| 16 | #include "intel_engine_types.h" |
| 17 | #include "intel_gt_pm.h" |
| 18 | #include "intel_ring_types.h" |
| 19 | #include "intel_timeline_types.h" |
| 20 | #include "i915_trace.h" |
| 21 | |
| 22 | #define CE_TRACE(ce, fmt, ...) do { \ |
| 23 | const struct intel_context *ce__ = (ce); \ |
| 24 | ENGINE_TRACE(ce__->engine, "context:%llx " fmt, \ |
| 25 | ce__->timeline->fence_context, \ |
| 26 | ##__VA_ARGS__); \ |
| 27 | } while (0) |
| 28 | |
| 29 | #define INTEL_CONTEXT_BANNED_PREEMPT_TIMEOUT_MS (1) |
| 30 | |
| 31 | struct i915_gem_ww_ctx; |
| 32 | |
| 33 | void intel_context_init(struct intel_context *ce, |
| 34 | struct intel_engine_cs *engine); |
| 35 | void intel_context_fini(struct intel_context *ce); |
| 36 | |
| 37 | void i915_context_module_exit(void); |
| 38 | int i915_context_module_init(void); |
| 39 | |
| 40 | struct intel_context * |
| 41 | intel_context_create(struct intel_engine_cs *engine); |
| 42 | |
| 43 | int intel_context_alloc_state(struct intel_context *ce); |
| 44 | |
| 45 | void intel_context_free(struct intel_context *ce); |
| 46 | |
| 47 | int intel_context_reconfigure_sseu(struct intel_context *ce, |
| 48 | const struct intel_sseu sseu); |
| 49 | |
| 50 | #define PARENT_SCRATCH_SIZE PAGE_SIZE |
| 51 | |
| 52 | static inline bool intel_context_is_child(struct intel_context *ce) |
| 53 | { |
| 54 | return !!ce->parallel.parent; |
| 55 | } |
| 56 | |
| 57 | static inline bool intel_context_is_parent(struct intel_context *ce) |
| 58 | { |
| 59 | return !!ce->parallel.number_children; |
| 60 | } |
| 61 | |
| 62 | static inline bool intel_context_is_pinned(struct intel_context *ce); |
| 63 | |
| 64 | static inline struct intel_context * |
| 65 | intel_context_to_parent(struct intel_context *ce) |
| 66 | { |
| 67 | if (intel_context_is_child(ce)) { |
| 68 | /* |
| 69 | * The parent holds ref count to the child so it is always safe |
| 70 | * for the parent to access the child, but the child has a |
| 71 | * pointer to the parent without a ref. To ensure this is safe |
| 72 | * the child should only access the parent pointer while the |
| 73 | * parent is pinned. |
| 74 | */ |
| 75 | GEM_BUG_ON(!intel_context_is_pinned(ce->parallel.parent)); |
| 76 | |
| 77 | return ce->parallel.parent; |
| 78 | } else { |
| 79 | return ce; |
| 80 | } |
| 81 | } |
| 82 | |
| 83 | static inline bool intel_context_is_parallel(struct intel_context *ce) |
| 84 | { |
| 85 | return intel_context_is_child(ce) || intel_context_is_parent(ce); |
| 86 | } |
| 87 | |
| 88 | void intel_context_bind_parent_child(struct intel_context *parent, |
| 89 | struct intel_context *child); |
| 90 | |
| 91 | #define for_each_child(parent, ce)\ |
| 92 | list_for_each_entry(ce, &(parent)->parallel.child_list,\ |
| 93 | parallel.child_link) |
| 94 | #define for_each_child_safe(parent, ce, cn)\ |
| 95 | list_for_each_entry_safe(ce, cn, &(parent)->parallel.child_list,\ |
| 96 | parallel.child_link) |
| 97 | |
| 98 | /** |
| 99 | * intel_context_lock_pinned - Stablises the 'pinned' status of the HW context |
| 100 | * @ce: the context |
| 101 | * |
| 102 | * Acquire a lock on the pinned status of the HW context, such that the context |
| 103 | * can neither be bound to the GPU or unbound whilst the lock is held, i.e. |
| 104 | * intel_context_is_pinned() remains stable. |
| 105 | */ |
| 106 | static inline int intel_context_lock_pinned(struct intel_context *ce) |
| 107 | __acquires(ce->pin_mutex) |
| 108 | { |
| 109 | return mutex_lock_interruptible(&ce->pin_mutex); |
| 110 | } |
| 111 | |
| 112 | /** |
| 113 | * intel_context_is_pinned - Reports the 'pinned' status |
| 114 | * @ce: the context |
| 115 | * |
| 116 | * While in use by the GPU, the context, along with its ring and page |
| 117 | * tables is pinned into memory and the GTT. |
| 118 | * |
| 119 | * Returns: true if the context is currently pinned for use by the GPU. |
| 120 | */ |
| 121 | static inline bool |
| 122 | intel_context_is_pinned(struct intel_context *ce) |
| 123 | { |
| 124 | return atomic_read(v: &ce->pin_count); |
| 125 | } |
| 126 | |
| 127 | static inline void intel_context_cancel_request(struct intel_context *ce, |
| 128 | struct i915_request *rq) |
| 129 | { |
| 130 | GEM_BUG_ON(!ce->ops->cancel_request); |
| 131 | return ce->ops->cancel_request(ce, rq); |
| 132 | } |
| 133 | |
| 134 | /** |
| 135 | * intel_context_unlock_pinned - Releases the earlier locking of 'pinned' status |
| 136 | * @ce: the context |
| 137 | * |
| 138 | * Releases the lock earlier acquired by intel_context_unlock_pinned(). |
| 139 | */ |
| 140 | static inline void intel_context_unlock_pinned(struct intel_context *ce) |
| 141 | __releases(ce->pin_mutex) |
| 142 | { |
| 143 | mutex_unlock(lock: &ce->pin_mutex); |
| 144 | } |
| 145 | |
| 146 | int __intel_context_do_pin(struct intel_context *ce); |
| 147 | int __intel_context_do_pin_ww(struct intel_context *ce, |
| 148 | struct i915_gem_ww_ctx *ww); |
| 149 | |
| 150 | static inline bool intel_context_pin_if_active(struct intel_context *ce) |
| 151 | { |
| 152 | return atomic_inc_not_zero(v: &ce->pin_count); |
| 153 | } |
| 154 | |
| 155 | static inline int intel_context_pin(struct intel_context *ce) |
| 156 | { |
| 157 | if (likely(intel_context_pin_if_active(ce))) |
| 158 | return 0; |
| 159 | |
| 160 | return __intel_context_do_pin(ce); |
| 161 | } |
| 162 | |
| 163 | static inline int intel_context_pin_ww(struct intel_context *ce, |
| 164 | struct i915_gem_ww_ctx *ww) |
| 165 | { |
| 166 | if (likely(intel_context_pin_if_active(ce))) |
| 167 | return 0; |
| 168 | |
| 169 | return __intel_context_do_pin_ww(ce, ww); |
| 170 | } |
| 171 | |
| 172 | static inline void __intel_context_pin(struct intel_context *ce) |
| 173 | { |
| 174 | GEM_BUG_ON(!intel_context_is_pinned(ce)); |
| 175 | atomic_inc(v: &ce->pin_count); |
| 176 | } |
| 177 | |
| 178 | void __intel_context_do_unpin(struct intel_context *ce, int sub); |
| 179 | |
| 180 | static inline void intel_context_sched_disable_unpin(struct intel_context *ce) |
| 181 | { |
| 182 | __intel_context_do_unpin(ce, sub: 2); |
| 183 | } |
| 184 | |
| 185 | static inline void intel_context_unpin(struct intel_context *ce) |
| 186 | { |
| 187 | if (!ce->ops->sched_disable) { |
| 188 | __intel_context_do_unpin(ce, sub: 1); |
| 189 | } else { |
| 190 | /* |
| 191 | * Move ownership of this pin to the scheduling disable which is |
| 192 | * an async operation. When that operation completes the above |
| 193 | * intel_context_sched_disable_unpin is called potentially |
| 194 | * unpinning the context. |
| 195 | */ |
| 196 | while (!atomic_add_unless(v: &ce->pin_count, a: -1, u: 1)) { |
| 197 | if (atomic_cmpxchg(v: &ce->pin_count, old: 1, new: 2) == 1) { |
| 198 | ce->ops->sched_disable(ce); |
| 199 | break; |
| 200 | } |
| 201 | } |
| 202 | } |
| 203 | } |
| 204 | |
| 205 | void intel_context_enter_engine(struct intel_context *ce); |
| 206 | void intel_context_exit_engine(struct intel_context *ce); |
| 207 | |
| 208 | static inline void intel_context_enter(struct intel_context *ce) |
| 209 | { |
| 210 | lockdep_assert_held(&ce->timeline->mutex); |
| 211 | if (ce->active_count++) |
| 212 | return; |
| 213 | |
| 214 | ce->ops->enter(ce); |
| 215 | ce->wakeref = intel_gt_pm_get(gt: ce->vm->gt); |
| 216 | } |
| 217 | |
| 218 | static inline void intel_context_mark_active(struct intel_context *ce) |
| 219 | { |
| 220 | lockdep_assert(lockdep_is_held(&ce->timeline->mutex) || |
| 221 | test_bit(CONTEXT_IS_PARKING, &ce->flags)); |
| 222 | ++ce->active_count; |
| 223 | } |
| 224 | |
| 225 | static inline void intel_context_exit(struct intel_context *ce) |
| 226 | { |
| 227 | lockdep_assert_held(&ce->timeline->mutex); |
| 228 | GEM_BUG_ON(!ce->active_count); |
| 229 | if (--ce->active_count) |
| 230 | return; |
| 231 | |
| 232 | intel_gt_pm_put_async(gt: ce->vm->gt, handle: ce->wakeref); |
| 233 | ce->ops->exit(ce); |
| 234 | } |
| 235 | |
| 236 | static inline struct intel_context *intel_context_get(struct intel_context *ce) |
| 237 | { |
| 238 | kref_get(kref: &ce->ref); |
| 239 | return ce; |
| 240 | } |
| 241 | |
| 242 | static inline void intel_context_put(struct intel_context *ce) |
| 243 | { |
| 244 | kref_put(kref: &ce->ref, release: ce->ops->destroy); |
| 245 | } |
| 246 | |
| 247 | static inline struct intel_timeline *__must_check |
| 248 | intel_context_timeline_lock(struct intel_context *ce) |
| 249 | __acquires(&ce->timeline->mutex) |
| 250 | { |
| 251 | struct intel_timeline *tl = ce->timeline; |
| 252 | int err; |
| 253 | |
| 254 | if (intel_context_is_parent(ce)) |
| 255 | err = mutex_lock_interruptible_nested(lock: &tl->mutex, subclass: 0); |
| 256 | else if (intel_context_is_child(ce)) |
| 257 | err = mutex_lock_interruptible_nested(lock: &tl->mutex, |
| 258 | subclass: ce->parallel.child_index + 1); |
| 259 | else |
| 260 | err = mutex_lock_interruptible(&tl->mutex); |
| 261 | if (err) |
| 262 | return ERR_PTR(error: err); |
| 263 | |
| 264 | return tl; |
| 265 | } |
| 266 | |
| 267 | static inline void intel_context_timeline_unlock(struct intel_timeline *tl) |
| 268 | __releases(&tl->mutex) |
| 269 | { |
| 270 | mutex_unlock(lock: &tl->mutex); |
| 271 | } |
| 272 | |
| 273 | int intel_context_prepare_remote_request(struct intel_context *ce, |
| 274 | struct i915_request *rq); |
| 275 | |
| 276 | struct i915_request *intel_context_create_request(struct intel_context *ce); |
| 277 | |
| 278 | struct i915_request *intel_context_get_active_request(struct intel_context *ce); |
| 279 | |
| 280 | static inline bool intel_context_is_barrier(const struct intel_context *ce) |
| 281 | { |
| 282 | return test_bit(CONTEXT_BARRIER_BIT, &ce->flags); |
| 283 | } |
| 284 | |
| 285 | static inline void intel_context_close(struct intel_context *ce) |
| 286 | { |
| 287 | set_bit(CONTEXT_CLOSED_BIT, addr: &ce->flags); |
| 288 | |
| 289 | if (ce->ops->close) |
| 290 | ce->ops->close(ce); |
| 291 | } |
| 292 | |
| 293 | static inline bool intel_context_is_closed(const struct intel_context *ce) |
| 294 | { |
| 295 | return test_bit(CONTEXT_CLOSED_BIT, &ce->flags); |
| 296 | } |
| 297 | |
| 298 | static inline bool intel_context_has_inflight(const struct intel_context *ce) |
| 299 | { |
| 300 | return test_bit(COPS_HAS_INFLIGHT_BIT, &ce->ops->flags); |
| 301 | } |
| 302 | |
| 303 | static inline bool intel_context_use_semaphores(const struct intel_context *ce) |
| 304 | { |
| 305 | return test_bit(CONTEXT_USE_SEMAPHORES, &ce->flags); |
| 306 | } |
| 307 | |
| 308 | static inline void intel_context_set_use_semaphores(struct intel_context *ce) |
| 309 | { |
| 310 | set_bit(CONTEXT_USE_SEMAPHORES, addr: &ce->flags); |
| 311 | } |
| 312 | |
| 313 | static inline void intel_context_clear_use_semaphores(struct intel_context *ce) |
| 314 | { |
| 315 | clear_bit(CONTEXT_USE_SEMAPHORES, addr: &ce->flags); |
| 316 | } |
| 317 | |
| 318 | static inline bool intel_context_is_banned(const struct intel_context *ce) |
| 319 | { |
| 320 | return test_bit(CONTEXT_BANNED, &ce->flags); |
| 321 | } |
| 322 | |
| 323 | static inline bool intel_context_set_banned(struct intel_context *ce) |
| 324 | { |
| 325 | return test_and_set_bit(CONTEXT_BANNED, addr: &ce->flags); |
| 326 | } |
| 327 | |
| 328 | bool intel_context_ban(struct intel_context *ce, struct i915_request *rq); |
| 329 | |
| 330 | static inline bool intel_context_is_schedulable(const struct intel_context *ce) |
| 331 | { |
| 332 | return !test_bit(CONTEXT_EXITING, &ce->flags) && |
| 333 | !test_bit(CONTEXT_BANNED, &ce->flags); |
| 334 | } |
| 335 | |
| 336 | static inline bool intel_context_is_exiting(const struct intel_context *ce) |
| 337 | { |
| 338 | return test_bit(CONTEXT_EXITING, &ce->flags); |
| 339 | } |
| 340 | |
| 341 | static inline bool intel_context_set_exiting(struct intel_context *ce) |
| 342 | { |
| 343 | return test_and_set_bit(CONTEXT_EXITING, addr: &ce->flags); |
| 344 | } |
| 345 | |
| 346 | bool intel_context_revoke(struct intel_context *ce); |
| 347 | |
| 348 | static inline bool |
| 349 | intel_context_force_single_submission(const struct intel_context *ce) |
| 350 | { |
| 351 | return test_bit(CONTEXT_FORCE_SINGLE_SUBMISSION, &ce->flags); |
| 352 | } |
| 353 | |
| 354 | static inline void |
| 355 | intel_context_set_single_submission(struct intel_context *ce) |
| 356 | { |
| 357 | __set_bit(CONTEXT_FORCE_SINGLE_SUBMISSION, &ce->flags); |
| 358 | } |
| 359 | |
| 360 | static inline bool |
| 361 | intel_context_nopreempt(const struct intel_context *ce) |
| 362 | { |
| 363 | return test_bit(CONTEXT_NOPREEMPT, &ce->flags); |
| 364 | } |
| 365 | |
| 366 | static inline void |
| 367 | intel_context_set_nopreempt(struct intel_context *ce) |
| 368 | { |
| 369 | set_bit(CONTEXT_NOPREEMPT, addr: &ce->flags); |
| 370 | } |
| 371 | |
| 372 | static inline void |
| 373 | intel_context_clear_nopreempt(struct intel_context *ce) |
| 374 | { |
| 375 | clear_bit(CONTEXT_NOPREEMPT, addr: &ce->flags); |
| 376 | } |
| 377 | |
| 378 | #if IS_ENABLED(CONFIG_DRM_I915_REPLAY_GPU_HANGS_API) |
| 379 | static inline bool intel_context_has_own_state(const struct intel_context *ce) |
| 380 | { |
| 381 | return test_bit(CONTEXT_OWN_STATE, &ce->flags); |
| 382 | } |
| 383 | |
| 384 | static inline bool intel_context_set_own_state(struct intel_context *ce) |
| 385 | { |
| 386 | return test_and_set_bit(CONTEXT_OWN_STATE, addr: &ce->flags); |
| 387 | } |
| 388 | #else |
| 389 | static inline bool intel_context_has_own_state(const struct intel_context *ce) |
| 390 | { |
| 391 | return false; |
| 392 | } |
| 393 | |
| 394 | static inline bool intel_context_set_own_state(struct intel_context *ce) |
| 395 | { |
| 396 | return true; |
| 397 | } |
| 398 | #endif |
| 399 | |
| 400 | u64 intel_context_get_total_runtime_ns(struct intel_context *ce); |
| 401 | u64 intel_context_get_avg_runtime_ns(struct intel_context *ce); |
| 402 | |
| 403 | static inline u64 intel_context_clock(void) |
| 404 | { |
| 405 | /* As we mix CS cycles with CPU clocks, use the raw monotonic clock. */ |
| 406 | return ktime_get_raw_fast_ns(); |
| 407 | } |
| 408 | |
| 409 | #endif /* __INTEL_CONTEXT_H__ */ |
| 410 | |