1 /* SPDX-License-Identifier: MIT */ 2 #ifndef _INTEL_RINGBUFFER_H_ 3 #define _INTEL_RINGBUFFER_H_ 4 5 #include <drm/drm_util.h> 6 7 #include <linux/hashtable.h> 8 #include <linux/irq_work.h> 9 #include <linux/random.h> 10 #include <linux/seqlock.h> 11 12 #include "i915_gem_batch_pool.h" 13 #include "i915_pmu.h" 14 #include "i915_reg.h" 15 #include "i915_request.h" 16 #include "i915_selftest.h" 17 #include "i915_timeline.h" 18 #include "intel_engine_types.h" 19 #include "intel_gpu_commands.h" 20 #include "intel_workarounds.h" 21 22 struct drm_printer; 23 24 /* Early gen2 devices have a cacheline of just 32 bytes, using 64 is overkill, 25 * but keeps the logic simple. Indeed, the whole purpose of this macro is just 26 * to give some inclination as to some of the magic values used in the various 27 * workarounds! 28 */ 29 #define CACHELINE_BYTES 64 30 #define CACHELINE_DWORDS (CACHELINE_BYTES / sizeof(u32)) 31 32 /* 33 * The register defines to be used with the following macros need to accept a 34 * base param, e.g: 35 * 36 * REG_FOO(base) _MMIO((base) + <relative offset>) 37 * ENGINE_READ(engine, REG_FOO); 38 * 39 * register arrays are to be defined and accessed as follows: 40 * 41 * REG_BAR(base, i) _MMIO((base) + <relative offset> + (i) * <shift>) 42 * ENGINE_READ_IDX(engine, REG_BAR, i) 43 */ 44 45 #define __ENGINE_REG_OP(op__, engine__, ...) \ 46 intel_uncore_##op__((engine__)->uncore, __VA_ARGS__) 47 48 #define __ENGINE_READ_OP(op__, engine__, reg__) \ 49 __ENGINE_REG_OP(op__, (engine__), reg__((engine__)->mmio_base)) 50 51 #define ENGINE_READ16(...) __ENGINE_READ_OP(read16, __VA_ARGS__) 52 #define ENGINE_READ(...) __ENGINE_READ_OP(read, __VA_ARGS__) 53 #define ENGINE_READ_FW(...) __ENGINE_READ_OP(read_fw, __VA_ARGS__) 54 #define ENGINE_POSTING_READ(...) __ENGINE_READ_OP(posting_read, __VA_ARGS__) 55 56 #define ENGINE_READ64(engine__, lower_reg__, upper_reg__) \ 57 __ENGINE_REG_OP(read64_2x32, (engine__), \ 58 lower_reg__((engine__)->mmio_base), \ 59 upper_reg__((engine__)->mmio_base)) 60 61 #define ENGINE_READ_IDX(engine__, reg__, idx__) \ 62 __ENGINE_REG_OP(read, (engine__), reg__((engine__)->mmio_base, (idx__))) 63 64 #define __ENGINE_WRITE_OP(op__, engine__, reg__, val__) \ 65 __ENGINE_REG_OP(op__, (engine__), reg__((engine__)->mmio_base), (val__)) 66 67 #define ENGINE_WRITE16(...) __ENGINE_WRITE_OP(write16, __VA_ARGS__) 68 #define ENGINE_WRITE(...) __ENGINE_WRITE_OP(write, __VA_ARGS__) 69 #define ENGINE_WRITE_FW(...) __ENGINE_WRITE_OP(write_fw, __VA_ARGS__) 70 71 /* seqno size is actually only a uint32, but since we plan to use MI_FLUSH_DW to 72 * do the writes, and that must have qw aligned offsets, simply pretend it's 8b. 73 */ 74 enum intel_engine_hangcheck_action { 75 ENGINE_IDLE = 0, 76 ENGINE_WAIT, 77 ENGINE_ACTIVE_SEQNO, 78 ENGINE_ACTIVE_HEAD, 79 ENGINE_ACTIVE_SUBUNITS, 80 ENGINE_WAIT_KICK, 81 ENGINE_DEAD, 82 }; 83 84 static inline const char * 85 hangcheck_action_to_str(const enum intel_engine_hangcheck_action a) 86 { 87 switch (a) { 88 case ENGINE_IDLE: 89 return "idle"; 90 case ENGINE_WAIT: 91 return "wait"; 92 case ENGINE_ACTIVE_SEQNO: 93 return "active seqno"; 94 case ENGINE_ACTIVE_HEAD: 95 return "active head"; 96 case ENGINE_ACTIVE_SUBUNITS: 97 return "active subunits"; 98 case ENGINE_WAIT_KICK: 99 return "wait kick"; 100 case ENGINE_DEAD: 101 return "dead"; 102 } 103 104 return "unknown"; 105 } 106 107 void intel_engines_set_scheduler_caps(struct drm_i915_private *i915); 108 109 static inline void 110 execlists_set_active(struct intel_engine_execlists *execlists, 111 unsigned int bit) 112 { 113 __set_bit(bit, (unsigned long *)&execlists->active); 114 } 115 116 static inline bool 117 execlists_set_active_once(struct intel_engine_execlists *execlists, 118 unsigned int bit) 119 { 120 return !__test_and_set_bit(bit, (unsigned long *)&execlists->active); 121 } 122 123 static inline void 124 execlists_clear_active(struct intel_engine_execlists *execlists, 125 unsigned int bit) 126 { 127 __clear_bit(bit, (unsigned long *)&execlists->active); 128 } 129 130 static inline void 131 execlists_clear_all_active(struct intel_engine_execlists *execlists) 132 { 133 execlists->active = 0; 134 } 135 136 static inline bool 137 execlists_is_active(const struct intel_engine_execlists *execlists, 138 unsigned int bit) 139 { 140 return test_bit(bit, (unsigned long *)&execlists->active); 141 } 142 143 void execlists_user_begin(struct intel_engine_execlists *execlists, 144 const struct execlist_port *port); 145 void execlists_user_end(struct intel_engine_execlists *execlists); 146 147 void 148 execlists_cancel_port_requests(struct intel_engine_execlists * const execlists); 149 150 struct i915_request * 151 execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists); 152 153 static inline unsigned int 154 execlists_num_ports(const struct intel_engine_execlists * const execlists) 155 { 156 return execlists->port_mask + 1; 157 } 158 159 static inline struct execlist_port * 160 execlists_port_complete(struct intel_engine_execlists * const execlists, 161 struct execlist_port * const port) 162 { 163 const unsigned int m = execlists->port_mask; 164 165 GEM_BUG_ON(port_index(port, execlists) != 0); 166 GEM_BUG_ON(!execlists_is_active(execlists, EXECLISTS_ACTIVE_USER)); 167 168 memmove(port, port + 1, m * sizeof(struct execlist_port)); 169 memset(port + m, 0, sizeof(struct execlist_port)); 170 171 return port; 172 } 173 174 static inline u32 175 intel_read_status_page(const struct intel_engine_cs *engine, int reg) 176 { 177 /* Ensure that the compiler doesn't optimize away the load. */ 178 return READ_ONCE(engine->status_page.addr[reg]); 179 } 180 181 static inline void 182 intel_write_status_page(struct intel_engine_cs *engine, int reg, u32 value) 183 { 184 /* Writing into the status page should be done sparingly. Since 185 * we do when we are uncertain of the device state, we take a bit 186 * of extra paranoia to try and ensure that the HWS takes the value 187 * we give and that it doesn't end up trapped inside the CPU! 188 */ 189 if (static_cpu_has(X86_FEATURE_CLFLUSH)) { 190 mb(); 191 clflush(&engine->status_page.addr[reg]); 192 engine->status_page.addr[reg] = value; 193 clflush(&engine->status_page.addr[reg]); 194 mb(); 195 } else { 196 WRITE_ONCE(engine->status_page.addr[reg], value); 197 } 198 } 199 200 /* 201 * Reads a dword out of the status page, which is written to from the command 202 * queue by automatic updates, MI_REPORT_HEAD, MI_STORE_DATA_INDEX, or 203 * MI_STORE_DATA_IMM. 204 * 205 * The following dwords have a reserved meaning: 206 * 0x00: ISR copy, updated when an ISR bit not set in the HWSTAM changes. 207 * 0x04: ring 0 head pointer 208 * 0x05: ring 1 head pointer (915-class) 209 * 0x06: ring 2 head pointer (915-class) 210 * 0x10-0x1b: Context status DWords (GM45) 211 * 0x1f: Last written status offset. (GM45) 212 * 0x20-0x2f: Reserved (Gen6+) 213 * 214 * The area from dword 0x30 to 0x3ff is available for driver usage. 215 */ 216 #define I915_GEM_HWS_PREEMPT 0x32 217 #define I915_GEM_HWS_PREEMPT_ADDR (I915_GEM_HWS_PREEMPT * sizeof(u32)) 218 #define I915_GEM_HWS_SEQNO 0x40 219 #define I915_GEM_HWS_SEQNO_ADDR (I915_GEM_HWS_SEQNO * sizeof(u32)) 220 #define I915_GEM_HWS_SCRATCH 0x80 221 #define I915_GEM_HWS_SCRATCH_ADDR (I915_GEM_HWS_SCRATCH * sizeof(u32)) 222 223 #define I915_HWS_CSB_BUF0_INDEX 0x10 224 #define I915_HWS_CSB_WRITE_INDEX 0x1f 225 #define CNL_HWS_CSB_WRITE_INDEX 0x2f 226 227 struct intel_ring * 228 intel_engine_create_ring(struct intel_engine_cs *engine, 229 struct i915_timeline *timeline, 230 int size); 231 int intel_ring_pin(struct intel_ring *ring); 232 void intel_ring_reset(struct intel_ring *ring, u32 tail); 233 unsigned int intel_ring_update_space(struct intel_ring *ring); 234 void intel_ring_unpin(struct intel_ring *ring); 235 void intel_ring_free(struct kref *ref); 236 237 static inline struct intel_ring *intel_ring_get(struct intel_ring *ring) 238 { 239 kref_get(&ring->ref); 240 return ring; 241 } 242 243 static inline void intel_ring_put(struct intel_ring *ring) 244 { 245 kref_put(&ring->ref, intel_ring_free); 246 } 247 248 void intel_engine_stop(struct intel_engine_cs *engine); 249 void intel_engine_cleanup(struct intel_engine_cs *engine); 250 251 int __must_check intel_ring_cacheline_align(struct i915_request *rq); 252 253 u32 __must_check *intel_ring_begin(struct i915_request *rq, unsigned int n); 254 255 static inline void intel_ring_advance(struct i915_request *rq, u32 *cs) 256 { 257 /* Dummy function. 258 * 259 * This serves as a placeholder in the code so that the reader 260 * can compare against the preceding intel_ring_begin() and 261 * check that the number of dwords emitted matches the space 262 * reserved for the command packet (i.e. the value passed to 263 * intel_ring_begin()). 264 */ 265 GEM_BUG_ON((rq->ring->vaddr + rq->ring->emit) != cs); 266 } 267 268 static inline u32 intel_ring_wrap(const struct intel_ring *ring, u32 pos) 269 { 270 return pos & (ring->size - 1); 271 } 272 273 static inline bool 274 intel_ring_offset_valid(const struct intel_ring *ring, 275 unsigned int pos) 276 { 277 if (pos & -ring->size) /* must be strictly within the ring */ 278 return false; 279 280 if (!IS_ALIGNED(pos, 8)) /* must be qword aligned */ 281 return false; 282 283 return true; 284 } 285 286 static inline u32 intel_ring_offset(const struct i915_request *rq, void *addr) 287 { 288 /* Don't write ring->size (equivalent to 0) as that hangs some GPUs. */ 289 u32 offset = addr - rq->ring->vaddr; 290 GEM_BUG_ON(offset > rq->ring->size); 291 return intel_ring_wrap(rq->ring, offset); 292 } 293 294 static inline void 295 assert_ring_tail_valid(const struct intel_ring *ring, unsigned int tail) 296 { 297 GEM_BUG_ON(!intel_ring_offset_valid(ring, tail)); 298 299 /* 300 * "Ring Buffer Use" 301 * Gen2 BSpec "1. Programming Environment" / 1.4.4.6 302 * Gen3 BSpec "1c Memory Interface Functions" / 2.3.4.5 303 * Gen4+ BSpec "1c Memory Interface and Command Stream" / 5.3.4.5 304 * "If the Ring Buffer Head Pointer and the Tail Pointer are on the 305 * same cacheline, the Head Pointer must not be greater than the Tail 306 * Pointer." 307 * 308 * We use ring->head as the last known location of the actual RING_HEAD, 309 * it may have advanced but in the worst case it is equally the same 310 * as ring->head and so we should never program RING_TAIL to advance 311 * into the same cacheline as ring->head. 312 */ 313 #define cacheline(a) round_down(a, CACHELINE_BYTES) 314 GEM_BUG_ON(cacheline(tail) == cacheline(ring->head) && 315 tail < ring->head); 316 #undef cacheline 317 } 318 319 static inline unsigned int 320 intel_ring_set_tail(struct intel_ring *ring, unsigned int tail) 321 { 322 /* Whilst writes to the tail are strictly order, there is no 323 * serialisation between readers and the writers. The tail may be 324 * read by i915_request_retire() just as it is being updated 325 * by execlists, as although the breadcrumb is complete, the context 326 * switch hasn't been seen. 327 */ 328 assert_ring_tail_valid(ring, tail); 329 ring->tail = tail; 330 return tail; 331 } 332 333 static inline unsigned int 334 __intel_ring_space(unsigned int head, unsigned int tail, unsigned int size) 335 { 336 /* 337 * "If the Ring Buffer Head Pointer and the Tail Pointer are on the 338 * same cacheline, the Head Pointer must not be greater than the Tail 339 * Pointer." 340 */ 341 GEM_BUG_ON(!is_power_of_2(size)); 342 return (head - tail - CACHELINE_BYTES) & (size - 1); 343 } 344 345 int intel_engines_init_mmio(struct drm_i915_private *i915); 346 int intel_engines_setup(struct drm_i915_private *i915); 347 int intel_engines_init(struct drm_i915_private *i915); 348 void intel_engines_cleanup(struct drm_i915_private *i915); 349 350 int intel_engine_init_common(struct intel_engine_cs *engine); 351 void intel_engine_cleanup_common(struct intel_engine_cs *engine); 352 353 int intel_ring_submission_setup(struct intel_engine_cs *engine); 354 int intel_ring_submission_init(struct intel_engine_cs *engine); 355 356 int intel_engine_stop_cs(struct intel_engine_cs *engine); 357 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine); 358 359 void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask); 360 361 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine); 362 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine); 363 364 void intel_engine_get_instdone(struct intel_engine_cs *engine, 365 struct intel_instdone *instdone); 366 367 void intel_engine_init_execlists(struct intel_engine_cs *engine); 368 369 void intel_engine_init_breadcrumbs(struct intel_engine_cs *engine); 370 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine); 371 372 void intel_engine_pin_breadcrumbs_irq(struct intel_engine_cs *engine); 373 void intel_engine_unpin_breadcrumbs_irq(struct intel_engine_cs *engine); 374 375 void intel_engine_signal_breadcrumbs(struct intel_engine_cs *engine); 376 void intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine); 377 378 static inline void 379 intel_engine_queue_breadcrumbs(struct intel_engine_cs *engine) 380 { 381 irq_work_queue(&engine->breadcrumbs.irq_work); 382 } 383 384 void intel_engine_breadcrumbs_irq(struct intel_engine_cs *engine); 385 386 void intel_engine_reset_breadcrumbs(struct intel_engine_cs *engine); 387 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine); 388 389 void intel_engine_print_breadcrumbs(struct intel_engine_cs *engine, 390 struct drm_printer *p); 391 392 static inline u32 *gen8_emit_pipe_control(u32 *batch, u32 flags, u32 offset) 393 { 394 memset(batch, 0, 6 * sizeof(u32)); 395 396 batch[0] = GFX_OP_PIPE_CONTROL(6); 397 batch[1] = flags; 398 batch[2] = offset; 399 400 return batch + 6; 401 } 402 403 static inline u32 * 404 gen8_emit_ggtt_write_rcs(u32 *cs, u32 value, u32 gtt_offset, u32 flags) 405 { 406 /* We're using qword write, offset should be aligned to 8 bytes. */ 407 GEM_BUG_ON(!IS_ALIGNED(gtt_offset, 8)); 408 409 /* w/a for post sync ops following a GPGPU operation we 410 * need a prior CS_STALL, which is emitted by the flush 411 * following the batch. 412 */ 413 *cs++ = GFX_OP_PIPE_CONTROL(6); 414 *cs++ = flags | PIPE_CONTROL_QW_WRITE | PIPE_CONTROL_GLOBAL_GTT_IVB; 415 *cs++ = gtt_offset; 416 *cs++ = 0; 417 *cs++ = value; 418 /* We're thrashing one dword of HWS. */ 419 *cs++ = 0; 420 421 return cs; 422 } 423 424 static inline u32 * 425 gen8_emit_ggtt_write(u32 *cs, u32 value, u32 gtt_offset, u32 flags) 426 { 427 /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */ 428 GEM_BUG_ON(gtt_offset & (1 << 5)); 429 /* Offset should be aligned to 8 bytes for both (QW/DW) write types */ 430 GEM_BUG_ON(!IS_ALIGNED(gtt_offset, 8)); 431 432 *cs++ = (MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW | flags; 433 *cs++ = gtt_offset | MI_FLUSH_DW_USE_GTT; 434 *cs++ = 0; 435 *cs++ = value; 436 437 return cs; 438 } 439 440 static inline void intel_engine_reset(struct intel_engine_cs *engine, 441 bool stalled) 442 { 443 if (engine->reset.reset) 444 engine->reset.reset(engine, stalled); 445 engine->serial++; /* contexts lost */ 446 } 447 448 bool intel_engine_is_idle(struct intel_engine_cs *engine); 449 bool intel_engines_are_idle(struct drm_i915_private *dev_priv); 450 451 void intel_engine_lost_context(struct intel_engine_cs *engine); 452 453 void intel_engines_reset_default_submission(struct drm_i915_private *i915); 454 unsigned int intel_engines_has_context_isolation(struct drm_i915_private *i915); 455 456 bool intel_engine_can_store_dword(struct intel_engine_cs *engine); 457 458 __printf(3, 4) 459 void intel_engine_dump(struct intel_engine_cs *engine, 460 struct drm_printer *m, 461 const char *header, ...); 462 463 struct intel_engine_cs * 464 intel_engine_lookup_user(struct drm_i915_private *i915, u8 class, u8 instance); 465 466 static inline void intel_engine_context_in(struct intel_engine_cs *engine) 467 { 468 unsigned long flags; 469 470 if (READ_ONCE(engine->stats.enabled) == 0) 471 return; 472 473 write_seqlock_irqsave(&engine->stats.lock, flags); 474 475 if (engine->stats.enabled > 0) { 476 if (engine->stats.active++ == 0) 477 engine->stats.start = ktime_get(); 478 GEM_BUG_ON(engine->stats.active == 0); 479 } 480 481 write_sequnlock_irqrestore(&engine->stats.lock, flags); 482 } 483 484 static inline void intel_engine_context_out(struct intel_engine_cs *engine) 485 { 486 unsigned long flags; 487 488 if (READ_ONCE(engine->stats.enabled) == 0) 489 return; 490 491 write_seqlock_irqsave(&engine->stats.lock, flags); 492 493 if (engine->stats.enabled > 0) { 494 ktime_t last; 495 496 if (engine->stats.active && --engine->stats.active == 0) { 497 /* 498 * Decrement the active context count and in case GPU 499 * is now idle add up to the running total. 500 */ 501 last = ktime_sub(ktime_get(), engine->stats.start); 502 503 engine->stats.total = ktime_add(engine->stats.total, 504 last); 505 } else if (engine->stats.active == 0) { 506 /* 507 * After turning on engine stats, context out might be 508 * the first event in which case we account from the 509 * time stats gathering was turned on. 510 */ 511 last = ktime_sub(ktime_get(), engine->stats.enabled_at); 512 513 engine->stats.total = ktime_add(engine->stats.total, 514 last); 515 } 516 } 517 518 write_sequnlock_irqrestore(&engine->stats.lock, flags); 519 } 520 521 int intel_enable_engine_stats(struct intel_engine_cs *engine); 522 void intel_disable_engine_stats(struct intel_engine_cs *engine); 523 524 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine); 525 526 struct i915_request * 527 intel_engine_find_active_request(struct intel_engine_cs *engine); 528 529 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 530 531 static inline bool inject_preempt_hang(struct intel_engine_execlists *execlists) 532 { 533 if (!execlists->preempt_hang.inject_hang) 534 return false; 535 536 complete(&execlists->preempt_hang.completion); 537 return true; 538 } 539 540 #else 541 542 static inline bool inject_preempt_hang(struct intel_engine_execlists *execlists) 543 { 544 return false; 545 } 546 547 #endif 548 549 #endif /* _INTEL_RINGBUFFER_H_ */ 550