]> git.ipfire.org Git - thirdparty/linux.git/blob - drivers/gpu/drm/i915/gt/intel_engine_types.h
io_uring: reset -EBUSY error when io sq thread is waken up
[thirdparty/linux.git] / drivers / gpu / drm / i915 / gt / intel_engine_types.h
1 /*
2 * SPDX-License-Identifier: MIT
3 *
4 * Copyright © 2019 Intel Corporation
5 */
6
7 #ifndef __INTEL_ENGINE_TYPES__
8 #define __INTEL_ENGINE_TYPES__
9
10 #include <linux/average.h>
11 #include <linux/hashtable.h>
12 #include <linux/irq_work.h>
13 #include <linux/kref.h>
14 #include <linux/list.h>
15 #include <linux/llist.h>
16 #include <linux/rbtree.h>
17 #include <linux/timer.h>
18 #include <linux/types.h>
19 #include <linux/workqueue.h>
20
21 #include "i915_gem.h"
22 #include "i915_pmu.h"
23 #include "i915_priolist_types.h"
24 #include "i915_selftest.h"
25 #include "intel_engine_pool_types.h"
26 #include "intel_sseu.h"
27 #include "intel_timeline_types.h"
28 #include "intel_wakeref.h"
29 #include "intel_workarounds_types.h"
30
31 /* Legacy HW Engine ID */
32
33 #define RCS0_HW 0
34 #define VCS0_HW 1
35 #define BCS0_HW 2
36 #define VECS0_HW 3
37 #define VCS1_HW 4
38 #define VCS2_HW 6
39 #define VCS3_HW 7
40 #define VECS1_HW 12
41
42 /* Gen11+ HW Engine class + instance */
43 #define RENDER_CLASS 0
44 #define VIDEO_DECODE_CLASS 1
45 #define VIDEO_ENHANCEMENT_CLASS 2
46 #define COPY_ENGINE_CLASS 3
47 #define OTHER_CLASS 4
48 #define MAX_ENGINE_CLASS 4
49 #define MAX_ENGINE_INSTANCE 3
50
51 #define I915_MAX_SLICES 3
52 #define I915_MAX_SUBSLICES 8
53
54 #define I915_CMD_HASH_ORDER 9
55
56 struct dma_fence;
57 struct drm_i915_gem_object;
58 struct drm_i915_reg_table;
59 struct i915_gem_context;
60 struct i915_request;
61 struct i915_sched_attr;
62 struct intel_gt;
63 struct intel_ring;
64 struct intel_uncore;
65
66 typedef u8 intel_engine_mask_t;
67 #define ALL_ENGINES ((intel_engine_mask_t)~0ul)
68
69 struct intel_hw_status_page {
70 struct i915_vma *vma;
71 u32 *addr;
72 };
73
74 struct intel_instdone {
75 u32 instdone;
76 /* The following exist only in the RCS engine */
77 u32 slice_common;
78 u32 slice_common_extra[2];
79 u32 sampler[I915_MAX_SLICES][I915_MAX_SUBSLICES];
80 u32 row[I915_MAX_SLICES][I915_MAX_SUBSLICES];
81 };
82
83 /*
84 * we use a single page to load ctx workarounds so all of these
85 * values are referred in terms of dwords
86 *
87 * struct i915_wa_ctx_bb:
88 * offset: specifies batch starting position, also helpful in case
89 * if we want to have multiple batches at different offsets based on
90 * some criteria. It is not a requirement at the moment but provides
91 * an option for future use.
92 * size: size of the batch in DWORDS
93 */
94 struct i915_ctx_workarounds {
95 struct i915_wa_ctx_bb {
96 u32 offset;
97 u32 size;
98 } indirect_ctx, per_ctx;
99 struct i915_vma *vma;
100 };
101
102 #define I915_MAX_VCS 4
103 #define I915_MAX_VECS 2
104
105 /*
106 * Engine IDs definitions.
107 * Keep instances of the same type engine together.
108 */
109 enum intel_engine_id {
110 RCS0 = 0,
111 BCS0,
112 VCS0,
113 VCS1,
114 VCS2,
115 VCS3,
116 #define _VCS(n) (VCS0 + (n))
117 VECS0,
118 VECS1,
119 #define _VECS(n) (VECS0 + (n))
120 I915_NUM_ENGINES
121 #define INVALID_ENGINE ((enum intel_engine_id)-1)
122 };
123
124 /* A simple estimator for the round-trip latency of an engine */
125 DECLARE_EWMA(_engine_latency, 6, 4)
126
127 struct st_preempt_hang {
128 struct completion completion;
129 unsigned int count;
130 };
131
132 /**
133 * struct intel_engine_execlists - execlist submission queue and port state
134 *
135 * The struct intel_engine_execlists represents the combined logical state of
136 * driver and the hardware state for execlist mode of submission.
137 */
138 struct intel_engine_execlists {
139 /**
140 * @tasklet: softirq tasklet for bottom handler
141 */
142 struct tasklet_struct tasklet;
143
144 /**
145 * @timer: kick the current context if its timeslice expires
146 */
147 struct timer_list timer;
148
149 /**
150 * @preempt: reset the current context if it fails to give way
151 */
152 struct timer_list preempt;
153
154 /**
155 * @default_priolist: priority list for I915_PRIORITY_NORMAL
156 */
157 struct i915_priolist default_priolist;
158
159 /**
160 * @error_interrupt: CS Master EIR
161 *
162 * The CS generates an interrupt when it detects an error. We capture
163 * the first error interrupt, record the EIR and schedule the tasklet.
164 * In the tasklet, we process the pending CS events to ensure we have
165 * the guilty request, and then reset the engine.
166 */
167 u32 error_interrupt;
168
169 /**
170 * @no_priolist: priority lists disabled
171 */
172 bool no_priolist;
173
174 /**
175 * @submit_reg: gen-specific execlist submission register
176 * set to the ExecList Submission Port (elsp) register pre-Gen11 and to
177 * the ExecList Submission Queue Contents register array for Gen11+
178 */
179 u32 __iomem *submit_reg;
180
181 /**
182 * @ctrl_reg: the enhanced execlists control register, used to load the
183 * submit queue on the HW and to request preemptions to idle
184 */
185 u32 __iomem *ctrl_reg;
186
187 #define EXECLIST_MAX_PORTS 2
188 /**
189 * @active: the currently known context executing on HW
190 */
191 struct i915_request * const *active;
192 /**
193 * @inflight: the set of contexts submitted and acknowleged by HW
194 *
195 * The set of inflight contexts is managed by reading CS events
196 * from the HW. On a context-switch event (not preemption), we
197 * know the HW has transitioned from port0 to port1, and we
198 * advance our inflight/active tracking accordingly.
199 */
200 struct i915_request *inflight[EXECLIST_MAX_PORTS + 1 /* sentinel */];
201 /**
202 * @pending: the next set of contexts submitted to ELSP
203 *
204 * We store the array of contexts that we submit to HW (via ELSP) and
205 * promote them to the inflight array once HW has signaled the
206 * preemption or idle-to-active event.
207 */
208 struct i915_request *pending[EXECLIST_MAX_PORTS + 1];
209
210 /**
211 * @port_mask: number of execlist ports - 1
212 */
213 unsigned int port_mask;
214
215 /**
216 * @switch_priority_hint: Second context priority.
217 *
218 * We submit multiple contexts to the HW simultaneously and would
219 * like to occasionally switch between them to emulate timeslicing.
220 * To know when timeslicing is suitable, we track the priority of
221 * the context submitted second.
222 */
223 int switch_priority_hint;
224
225 /**
226 * @queue_priority_hint: Highest pending priority.
227 *
228 * When we add requests into the queue, or adjust the priority of
229 * executing requests, we compute the maximum priority of those
230 * pending requests. We can then use this value to determine if
231 * we need to preempt the executing requests to service the queue.
232 * However, since the we may have recorded the priority of an inflight
233 * request we wanted to preempt but since completed, at the time of
234 * dequeuing the priority hint may no longer may match the highest
235 * available request priority.
236 */
237 int queue_priority_hint;
238
239 /**
240 * @queue: queue of requests, in priority lists
241 */
242 struct rb_root_cached queue;
243 struct rb_root_cached virtual;
244
245 /**
246 * @csb_write: control register for Context Switch buffer
247 *
248 * Note this register may be either mmio or HWSP shadow.
249 */
250 u32 *csb_write;
251
252 /**
253 * @csb_status: status array for Context Switch buffer
254 *
255 * Note these register may be either mmio or HWSP shadow.
256 */
257 u32 *csb_status;
258
259 /**
260 * @csb_size: context status buffer FIFO size
261 */
262 u8 csb_size;
263
264 /**
265 * @csb_head: context status buffer head
266 */
267 u8 csb_head;
268
269 I915_SELFTEST_DECLARE(struct st_preempt_hang preempt_hang;)
270 };
271
272 #define INTEL_ENGINE_CS_MAX_NAME 8
273
274 struct intel_engine_cs {
275 struct drm_i915_private *i915;
276 struct intel_gt *gt;
277 struct intel_uncore *uncore;
278 char name[INTEL_ENGINE_CS_MAX_NAME];
279
280 enum intel_engine_id id;
281 enum intel_engine_id legacy_idx;
282
283 unsigned int hw_id;
284 unsigned int guc_id;
285
286 intel_engine_mask_t mask;
287
288 u8 class;
289 u8 instance;
290
291 u16 uabi_class;
292 u16 uabi_instance;
293
294 u32 uabi_capabilities;
295 u32 context_size;
296 u32 mmio_base;
297
298 unsigned int context_tag;
299 #define NUM_CONTEXT_TAG roundup_pow_of_two(2 * EXECLIST_MAX_PORTS)
300
301 struct rb_node uabi_node;
302
303 struct intel_sseu sseu;
304
305 struct {
306 spinlock_t lock;
307 struct list_head requests;
308 struct list_head hold; /* ready requests, but on hold */
309 } active;
310
311 struct llist_head barrier_tasks;
312
313 struct intel_context *kernel_context; /* pinned */
314
315 intel_engine_mask_t saturated; /* submitting semaphores too late? */
316
317 struct {
318 struct delayed_work work;
319 struct i915_request *systole;
320 } heartbeat;
321
322 unsigned long serial;
323
324 unsigned long wakeref_serial;
325 struct intel_wakeref wakeref;
326 struct drm_i915_gem_object *default_state;
327 void *pinned_default_state;
328
329 struct {
330 struct intel_ring *ring;
331 struct intel_timeline *timeline;
332 } legacy;
333
334 /*
335 * We track the average duration of the idle pulse on parking the
336 * engine to keep an estimate of the how the fast the engine is
337 * under ideal conditions.
338 */
339 struct ewma__engine_latency latency;
340
341 /* Rather than have every client wait upon all user interrupts,
342 * with the herd waking after every interrupt and each doing the
343 * heavyweight seqno dance, we delegate the task (of being the
344 * bottom-half of the user interrupt) to the first client. After
345 * every interrupt, we wake up one client, who does the heavyweight
346 * coherent seqno read and either goes back to sleep (if incomplete),
347 * or wakes up all the completed clients in parallel, before then
348 * transferring the bottom-half status to the next client in the queue.
349 *
350 * Compared to walking the entire list of waiters in a single dedicated
351 * bottom-half, we reduce the latency of the first waiter by avoiding
352 * a context switch, but incur additional coherent seqno reads when
353 * following the chain of request breadcrumbs. Since it is most likely
354 * that we have a single client waiting on each seqno, then reducing
355 * the overhead of waking that client is much preferred.
356 */
357 struct intel_breadcrumbs {
358 spinlock_t irq_lock;
359 struct list_head signalers;
360
361 struct irq_work irq_work; /* for use from inside irq_lock */
362
363 unsigned int irq_enabled;
364
365 bool irq_armed;
366 } breadcrumbs;
367
368 struct intel_engine_pmu {
369 /**
370 * @enable: Bitmask of enable sample events on this engine.
371 *
372 * Bits correspond to sample event types, for instance
373 * I915_SAMPLE_QUEUED is bit 0 etc.
374 */
375 u32 enable;
376 /**
377 * @enable_count: Reference count for the enabled samplers.
378 *
379 * Index number corresponds to @enum drm_i915_pmu_engine_sample.
380 */
381 unsigned int enable_count[I915_ENGINE_SAMPLE_COUNT];
382 /**
383 * @sample: Counter values for sampling events.
384 *
385 * Our internal timer stores the current counters in this field.
386 *
387 * Index number corresponds to @enum drm_i915_pmu_engine_sample.
388 */
389 struct i915_pmu_sample sample[I915_ENGINE_SAMPLE_COUNT];
390 } pmu;
391
392 /*
393 * A pool of objects to use as shadow copies of client batch buffers
394 * when the command parser is enabled. Prevents the client from
395 * modifying the batch contents after software parsing.
396 */
397 struct intel_engine_pool pool;
398
399 struct intel_hw_status_page status_page;
400 struct i915_ctx_workarounds wa_ctx;
401 struct i915_wa_list ctx_wa_list;
402 struct i915_wa_list wa_list;
403 struct i915_wa_list whitelist;
404
405 u32 irq_keep_mask; /* always keep these interrupts */
406 u32 irq_enable_mask; /* bitmask to enable ring interrupt */
407 void (*irq_enable)(struct intel_engine_cs *engine);
408 void (*irq_disable)(struct intel_engine_cs *engine);
409
410 int (*resume)(struct intel_engine_cs *engine);
411
412 struct {
413 void (*prepare)(struct intel_engine_cs *engine);
414
415 void (*rewind)(struct intel_engine_cs *engine, bool stalled);
416 void (*cancel)(struct intel_engine_cs *engine);
417
418 void (*finish)(struct intel_engine_cs *engine);
419 } reset;
420
421 void (*park)(struct intel_engine_cs *engine);
422 void (*unpark)(struct intel_engine_cs *engine);
423
424 void (*set_default_submission)(struct intel_engine_cs *engine);
425
426 const struct intel_context_ops *cops;
427
428 int (*request_alloc)(struct i915_request *rq);
429
430 int (*emit_flush)(struct i915_request *request, u32 mode);
431 #define EMIT_INVALIDATE BIT(0)
432 #define EMIT_FLUSH BIT(1)
433 #define EMIT_BARRIER (EMIT_INVALIDATE | EMIT_FLUSH)
434 int (*emit_bb_start)(struct i915_request *rq,
435 u64 offset, u32 length,
436 unsigned int dispatch_flags);
437 #define I915_DISPATCH_SECURE BIT(0)
438 #define I915_DISPATCH_PINNED BIT(1)
439 int (*emit_init_breadcrumb)(struct i915_request *rq);
440 u32 *(*emit_fini_breadcrumb)(struct i915_request *rq,
441 u32 *cs);
442 unsigned int emit_fini_breadcrumb_dw;
443
444 /* Pass the request to the hardware queue (e.g. directly into
445 * the legacy ringbuffer or to the end of an execlist).
446 *
447 * This is called from an atomic context with irqs disabled; must
448 * be irq safe.
449 */
450 void (*submit_request)(struct i915_request *rq);
451
452 /*
453 * Called on signaling of a SUBMIT_FENCE, passing along the signaling
454 * request down to the bonded pairs.
455 */
456 void (*bond_execute)(struct i915_request *rq,
457 struct dma_fence *signal);
458
459 /*
460 * Call when the priority on a request has changed and it and its
461 * dependencies may need rescheduling. Note the request itself may
462 * not be ready to run!
463 */
464 void (*schedule)(struct i915_request *request,
465 const struct i915_sched_attr *attr);
466
467 void (*release)(struct intel_engine_cs *engine);
468
469 struct intel_engine_execlists execlists;
470
471 /*
472 * Keep track of completed timelines on this engine for early
473 * retirement with the goal of quickly enabling powersaving as
474 * soon as the engine is idle.
475 */
476 struct intel_timeline *retire;
477 struct work_struct retire_work;
478
479 /* status_notifier: list of callbacks for context-switch changes */
480 struct atomic_notifier_head context_status_notifier;
481
482 #define I915_ENGINE_USING_CMD_PARSER BIT(0)
483 #define I915_ENGINE_SUPPORTS_STATS BIT(1)
484 #define I915_ENGINE_HAS_PREEMPTION BIT(2)
485 #define I915_ENGINE_HAS_SEMAPHORES BIT(3)
486 #define I915_ENGINE_NEEDS_BREADCRUMB_TASKLET BIT(4)
487 #define I915_ENGINE_IS_VIRTUAL BIT(5)
488 #define I915_ENGINE_HAS_RELATIVE_MMIO BIT(6)
489 #define I915_ENGINE_REQUIRES_CMD_PARSER BIT(7)
490 unsigned int flags;
491
492 /*
493 * Table of commands the command parser needs to know about
494 * for this engine.
495 */
496 DECLARE_HASHTABLE(cmd_hash, I915_CMD_HASH_ORDER);
497
498 /*
499 * Table of registers allowed in commands that read/write registers.
500 */
501 const struct drm_i915_reg_table *reg_tables;
502 int reg_table_count;
503
504 /*
505 * Returns the bitmask for the length field of the specified command.
506 * Return 0 for an unrecognized/invalid command.
507 *
508 * If the command parser finds an entry for a command in the engine's
509 * cmd_tables, it gets the command's length based on the table entry.
510 * If not, it calls this function to determine the per-engine length
511 * field encoding for the command (i.e. different opcode ranges use
512 * certain bits to encode the command length in the header).
513 */
514 u32 (*get_cmd_length_mask)(u32 cmd_header);
515
516 struct {
517 /**
518 * @lock: Lock protecting the below fields.
519 */
520 seqlock_t lock;
521 /**
522 * @enabled: Reference count indicating number of listeners.
523 */
524 unsigned int enabled;
525 /**
526 * @active: Number of contexts currently scheduled in.
527 */
528 unsigned int active;
529 /**
530 * @enabled_at: Timestamp when busy stats were enabled.
531 */
532 ktime_t enabled_at;
533 /**
534 * @start: Timestamp of the last idle to active transition.
535 *
536 * Idle is defined as active == 0, active is active > 0.
537 */
538 ktime_t start;
539 /**
540 * @total: Total time this engine was busy.
541 *
542 * Accumulated time not counting the most recent block in cases
543 * where engine is currently busy (active > 0).
544 */
545 ktime_t total;
546 } stats;
547
548 struct {
549 unsigned long heartbeat_interval_ms;
550 unsigned long max_busywait_duration_ns;
551 unsigned long preempt_timeout_ms;
552 unsigned long stop_timeout_ms;
553 unsigned long timeslice_duration_ms;
554 } props;
555 };
556
557 static inline bool
558 intel_engine_using_cmd_parser(const struct intel_engine_cs *engine)
559 {
560 return engine->flags & I915_ENGINE_USING_CMD_PARSER;
561 }
562
563 static inline bool
564 intel_engine_requires_cmd_parser(const struct intel_engine_cs *engine)
565 {
566 return engine->flags & I915_ENGINE_REQUIRES_CMD_PARSER;
567 }
568
569 static inline bool
570 intel_engine_supports_stats(const struct intel_engine_cs *engine)
571 {
572 return engine->flags & I915_ENGINE_SUPPORTS_STATS;
573 }
574
575 static inline bool
576 intel_engine_has_preemption(const struct intel_engine_cs *engine)
577 {
578 return engine->flags & I915_ENGINE_HAS_PREEMPTION;
579 }
580
581 static inline bool
582 intel_engine_has_semaphores(const struct intel_engine_cs *engine)
583 {
584 return engine->flags & I915_ENGINE_HAS_SEMAPHORES;
585 }
586
587 static inline bool
588 intel_engine_needs_breadcrumb_tasklet(const struct intel_engine_cs *engine)
589 {
590 return engine->flags & I915_ENGINE_NEEDS_BREADCRUMB_TASKLET;
591 }
592
593 static inline bool
594 intel_engine_is_virtual(const struct intel_engine_cs *engine)
595 {
596 return engine->flags & I915_ENGINE_IS_VIRTUAL;
597 }
598
599 static inline bool
600 intel_engine_has_relative_mmio(const struct intel_engine_cs * const engine)
601 {
602 return engine->flags & I915_ENGINE_HAS_RELATIVE_MMIO;
603 }
604
605 #define instdone_has_slice(dev_priv___, sseu___, slice___) \
606 ((IS_GEN(dev_priv___, 7) ? 1 : ((sseu___)->slice_mask)) & BIT(slice___))
607
608 #define instdone_has_subslice(dev_priv__, sseu__, slice__, subslice__) \
609 (IS_GEN(dev_priv__, 7) ? (1 & BIT(subslice__)) : \
610 intel_sseu_has_subslice(sseu__, 0, subslice__))
611
612 #define for_each_instdone_slice_subslice(dev_priv_, sseu_, slice_, subslice_) \
613 for ((slice_) = 0, (subslice_) = 0; (slice_) < I915_MAX_SLICES; \
614 (subslice_) = ((subslice_) + 1) % I915_MAX_SUBSLICES, \
615 (slice_) += ((subslice_) == 0)) \
616 for_each_if((instdone_has_slice(dev_priv_, sseu_, slice_)) && \
617 (instdone_has_subslice(dev_priv_, sseu_, slice_, \
618 subslice_)))
619 #endif /* __INTEL_ENGINE_TYPES_H__ */