1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _INTEL_RINGBUFFER_H_
3 #define _INTEL_RINGBUFFER_H_
5 #include <linux/hashtable.h>
7 #include "i915_gem_batch_pool.h"
8 #include "i915_gem_timeline.h"
11 #include "i915_request.h"
12 #include "i915_selftest.h"
16 #define I915_CMD_HASH_ORDER 9
18 /* Early gen2 devices have a cacheline of just 32 bytes, using 64 is overkill,
19 * but keeps the logic simple. Indeed, the whole purpose of this macro is just
20 * to give some inclination as to some of the magic values used in the various
23 #define CACHELINE_BYTES 64
24 #define CACHELINE_DWORDS (CACHELINE_BYTES / sizeof(uint32_t))
26 struct intel_hw_status_page {
32 #define I915_READ_TAIL(engine) I915_READ(RING_TAIL((engine)->mmio_base))
33 #define I915_WRITE_TAIL(engine, val) I915_WRITE(RING_TAIL((engine)->mmio_base), val)
35 #define I915_READ_START(engine) I915_READ(RING_START((engine)->mmio_base))
36 #define I915_WRITE_START(engine, val) I915_WRITE(RING_START((engine)->mmio_base), val)
38 #define I915_READ_HEAD(engine) I915_READ(RING_HEAD((engine)->mmio_base))
39 #define I915_WRITE_HEAD(engine, val) I915_WRITE(RING_HEAD((engine)->mmio_base), val)
41 #define I915_READ_CTL(engine) I915_READ(RING_CTL((engine)->mmio_base))
42 #define I915_WRITE_CTL(engine, val) I915_WRITE(RING_CTL((engine)->mmio_base), val)
44 #define I915_READ_IMR(engine) I915_READ(RING_IMR((engine)->mmio_base))
45 #define I915_WRITE_IMR(engine, val) I915_WRITE(RING_IMR((engine)->mmio_base), val)
47 #define I915_READ_MODE(engine) I915_READ(RING_MI_MODE((engine)->mmio_base))
48 #define I915_WRITE_MODE(engine, val) I915_WRITE(RING_MI_MODE((engine)->mmio_base), val)
50 /* seqno size is actually only a uint32, but since we plan to use MI_FLUSH_DW to
51 * do the writes, and that must have qw aligned offsets, simply pretend it's 8b.
53 enum intel_engine_hangcheck_action {
58 ENGINE_ACTIVE_SUBUNITS,
63 static inline const char *
64 hangcheck_action_to_str(const enum intel_engine_hangcheck_action a)
71 case ENGINE_ACTIVE_SEQNO:
72 return "active seqno";
73 case ENGINE_ACTIVE_HEAD:
75 case ENGINE_ACTIVE_SUBUNITS:
76 return "active subunits";
77 case ENGINE_WAIT_KICK:
86 #define I915_MAX_SLICES 3
87 #define I915_MAX_SUBSLICES 3
89 #define instdone_slice_mask(dev_priv__) \
90 (INTEL_GEN(dev_priv__) == 7 ? \
91 1 : INTEL_INFO(dev_priv__)->sseu.slice_mask)
93 #define instdone_subslice_mask(dev_priv__) \
94 (INTEL_GEN(dev_priv__) == 7 ? \
95 1 : INTEL_INFO(dev_priv__)->sseu.subslice_mask[0])
97 #define for_each_instdone_slice_subslice(dev_priv__, slice__, subslice__) \
98 for ((slice__) = 0, (subslice__) = 0; \
99 (slice__) < I915_MAX_SLICES; \
100 (subslice__) = ((subslice__) + 1) < I915_MAX_SUBSLICES ? (subslice__) + 1 : 0, \
101 (slice__) += ((subslice__) == 0)) \
102 for_each_if((BIT(slice__) & instdone_slice_mask(dev_priv__)) && \
103 (BIT(subslice__) & instdone_subslice_mask(dev_priv__)))
105 struct intel_instdone {
107 /* The following exist only in the RCS engine */
109 u32 sampler[I915_MAX_SLICES][I915_MAX_SUBSLICES];
110 u32 row[I915_MAX_SLICES][I915_MAX_SUBSLICES];
113 struct intel_engine_hangcheck {
116 enum intel_engine_hangcheck_action action;
117 unsigned long action_timestamp;
119 struct intel_instdone instdone;
120 struct i915_request *active_request;
125 struct i915_vma *vma;
128 struct list_head request_list;
139 struct i915_gem_context;
140 struct drm_i915_reg_table;
143 * we use a single page to load ctx workarounds so all of these
144 * values are referred in terms of dwords
146 * struct i915_wa_ctx_bb:
147 * offset: specifies batch starting position, also helpful in case
148 * if we want to have multiple batches at different offsets based on
149 * some criteria. It is not a requirement at the moment but provides
150 * an option for future use.
151 * size: size of the batch in DWORDS
153 struct i915_ctx_workarounds {
154 struct i915_wa_ctx_bb {
157 } indirect_ctx, per_ctx;
158 struct i915_vma *vma;
163 #define I915_MAX_VCS 4
164 #define I915_MAX_VECS 2
167 * Engine IDs definitions.
168 * Keep instances of the same type engine together.
170 enum intel_engine_id {
177 #define _VCS(n) (VCS + (n))
180 #define _VECS(n) (VECS + (n))
183 struct i915_priolist {
185 struct list_head requests;
190 * struct intel_engine_execlists - execlist submission queue and port state
192 * The struct intel_engine_execlists represents the combined logical state of
193 * driver and the hardware state for execlist mode of submission.
195 struct intel_engine_execlists {
197 * @tasklet: softirq tasklet for bottom handler
199 struct tasklet_struct tasklet;
202 * @default_priolist: priority list for I915_PRIORITY_NORMAL
204 struct i915_priolist default_priolist;
207 * @no_priolist: priority lists disabled
212 * @submit_reg: gen-specific execlist submission register
213 * set to the ExecList Submission Port (elsp) register pre-Gen11 and to
214 * the ExecList Submission Queue Contents register array for Gen11+
216 u32 __iomem *submit_reg;
219 * @ctrl_reg: the enhanced execlists control register, used to load the
220 * submit queue on the HW and to request preemptions to idle
222 u32 __iomem *ctrl_reg;
225 * @port: execlist port states
227 * For each hardware ELSP (ExecList Submission Port) we keep
228 * track of the last request and the number of times we submitted
229 * that port to hw. We then count the number of times the hw reports
230 * a context completion or preemption. As only one context can
231 * be active on hw, we limit resubmission of context to port[0]. This
232 * is called Lite Restore, of the context.
234 struct execlist_port {
236 * @request_count: combined request and submission count
238 struct i915_request *request_count;
239 #define EXECLIST_COUNT_BITS 2
240 #define port_request(p) ptr_mask_bits((p)->request_count, EXECLIST_COUNT_BITS)
241 #define port_count(p) ptr_unmask_bits((p)->request_count, EXECLIST_COUNT_BITS)
242 #define port_pack(rq, count) ptr_pack_bits(rq, count, EXECLIST_COUNT_BITS)
243 #define port_unpack(p, count) ptr_unpack_bits((p)->request_count, count, EXECLIST_COUNT_BITS)
244 #define port_set(p, packed) ((p)->request_count = (packed))
245 #define port_isset(p) ((p)->request_count)
246 #define port_index(p, execlists) ((p) - (execlists)->port)
249 * @context_id: context ID for port
251 GEM_DEBUG_DECL(u32 context_id);
253 #define EXECLIST_MAX_PORTS 2
254 } port[EXECLIST_MAX_PORTS];
257 * @active: is the HW active? We consider the HW as active after
258 * submitting any context for execution and until we have seen the
259 * last context completion event. After that, we do not expect any
260 * more events until we submit, and so can park the HW.
262 * As we have a small number of different sources from which we feed
263 * the HW, we track the state of each inside a single bitfield.
266 #define EXECLISTS_ACTIVE_USER 0
267 #define EXECLISTS_ACTIVE_PREEMPT 1
268 #define EXECLISTS_ACTIVE_HWACK 2
271 * @port_mask: number of execlist ports - 1
273 unsigned int port_mask;
276 * @queue_priority: Highest pending priority.
278 * When we add requests into the queue, or adjust the priority of
279 * executing requests, we compute the maximum priority of those
280 * pending requests. We can then use this value to determine if
281 * we need to preempt the executing requests to service the queue.
286 * @queue: queue of requests, in priority lists
288 struct rb_root queue;
291 * @first: leftmost level in priority @queue
293 struct rb_node *first;
296 * @fw_domains: forcewake domains for irq tasklet
298 unsigned int fw_domains;
301 * @csb_head: context status buffer head
303 unsigned int csb_head;
306 * @csb_use_mmio: access csb through mmio, instead of hwsp
311 * @preempt_complete_status: expected CSB upon completing preemption
313 u32 preempt_complete_status;
316 #define INTEL_ENGINE_CS_MAX_NAME 8
318 struct intel_engine_cs {
319 struct drm_i915_private *i915;
320 char name[INTEL_ENGINE_CS_MAX_NAME];
322 enum intel_engine_id id;
333 unsigned int irq_shift;
335 struct intel_ring *buffer;
336 struct intel_timeline *timeline;
338 struct drm_i915_gem_object *default_state;
341 unsigned long irq_posted;
342 #define ENGINE_IRQ_BREADCRUMB 0
343 #define ENGINE_IRQ_EXECLIST 1
345 /* Rather than have every client wait upon all user interrupts,
346 * with the herd waking after every interrupt and each doing the
347 * heavyweight seqno dance, we delegate the task (of being the
348 * bottom-half of the user interrupt) to the first client. After
349 * every interrupt, we wake up one client, who does the heavyweight
350 * coherent seqno read and either goes back to sleep (if incomplete),
351 * or wakes up all the completed clients in parallel, before then
352 * transferring the bottom-half status to the next client in the queue.
354 * Compared to walking the entire list of waiters in a single dedicated
355 * bottom-half, we reduce the latency of the first waiter by avoiding
356 * a context switch, but incur additional coherent seqno reads when
357 * following the chain of request breadcrumbs. Since it is most likely
358 * that we have a single client waiting on each seqno, then reducing
359 * the overhead of waking that client is much preferred.
361 struct intel_breadcrumbs {
362 spinlock_t irq_lock; /* protects irq_*; irqsafe */
363 struct intel_wait *irq_wait; /* oldest waiter by retirement */
365 spinlock_t rb_lock; /* protects the rb and wraps irq_lock */
366 struct rb_root waiters; /* sorted by retirement, priority */
367 struct list_head signals; /* sorted by retirement */
368 struct task_struct *signaler; /* used for fence signalling */
370 struct timer_list fake_irq; /* used after a missed interrupt */
371 struct timer_list hangcheck; /* detect missed interrupts */
373 unsigned int hangcheck_interrupts;
374 unsigned int irq_enabled;
377 I915_SELFTEST_DECLARE(bool mock : 1);
382 * @enable: Bitmask of enable sample events on this engine.
384 * Bits correspond to sample event types, for instance
385 * I915_SAMPLE_QUEUED is bit 0 etc.
389 * @enable_count: Reference count for the enabled samplers.
391 * Index number corresponds to the bit number from @enable.
393 unsigned int enable_count[I915_PMU_SAMPLE_BITS];
395 * @sample: Counter values for sampling events.
397 * Our internal timer stores the current counters in this field.
399 #define I915_ENGINE_SAMPLE_MAX (I915_SAMPLE_SEMA + 1)
400 struct i915_pmu_sample sample[I915_ENGINE_SAMPLE_MAX];
404 * A pool of objects to use as shadow copies of client batch buffers
405 * when the command parser is enabled. Prevents the client from
406 * modifying the batch contents after software parsing.
408 struct i915_gem_batch_pool batch_pool;
410 struct intel_hw_status_page status_page;
411 struct i915_ctx_workarounds wa_ctx;
412 struct i915_vma *scratch;
414 u32 irq_keep_mask; /* always keep these interrupts */
415 u32 irq_enable_mask; /* bitmask to enable ring interrupt */
416 void (*irq_enable)(struct intel_engine_cs *engine);
417 void (*irq_disable)(struct intel_engine_cs *engine);
419 int (*init_hw)(struct intel_engine_cs *engine);
420 void (*reset_hw)(struct intel_engine_cs *engine,
421 struct i915_request *rq);
423 void (*park)(struct intel_engine_cs *engine);
424 void (*unpark)(struct intel_engine_cs *engine);
426 void (*set_default_submission)(struct intel_engine_cs *engine);
428 struct intel_ring *(*context_pin)(struct intel_engine_cs *engine,
429 struct i915_gem_context *ctx);
430 void (*context_unpin)(struct intel_engine_cs *engine,
431 struct i915_gem_context *ctx);
432 int (*request_alloc)(struct i915_request *rq);
433 int (*init_context)(struct i915_request *rq);
435 int (*emit_flush)(struct i915_request *request, u32 mode);
436 #define EMIT_INVALIDATE BIT(0)
437 #define EMIT_FLUSH BIT(1)
438 #define EMIT_BARRIER (EMIT_INVALIDATE | EMIT_FLUSH)
439 int (*emit_bb_start)(struct i915_request *rq,
440 u64 offset, u32 length,
441 unsigned int dispatch_flags);
442 #define I915_DISPATCH_SECURE BIT(0)
443 #define I915_DISPATCH_PINNED BIT(1)
444 #define I915_DISPATCH_RS BIT(2)
445 void (*emit_breadcrumb)(struct i915_request *rq, u32 *cs);
446 int emit_breadcrumb_sz;
448 /* Pass the request to the hardware queue (e.g. directly into
449 * the legacy ringbuffer or to the end of an execlist).
451 * This is called from an atomic context with irqs disabled; must
454 void (*submit_request)(struct i915_request *rq);
456 /* Call when the priority on a request has changed and it and its
457 * dependencies may need rescheduling. Note the request itself may
458 * not be ready to run!
460 * Called under the struct_mutex.
462 void (*schedule)(struct i915_request *request, int priority);
465 * Cancel all requests on the hardware, or queued for execution.
466 * This should only cancel the ready requests that have been
467 * submitted to the engine (via the engine->submit_request callback).
468 * This is called when marking the device as wedged.
470 void (*cancel_requests)(struct intel_engine_cs *engine);
472 /* Some chipsets are not quite as coherent as advertised and need
473 * an expensive kick to force a true read of the up-to-date seqno.
474 * However, the up-to-date seqno is not always required and the last
475 * seen value is good enough. Note that the seqno will always be
476 * monotonic, even if not coherent.
478 void (*irq_seqno_barrier)(struct intel_engine_cs *engine);
479 void (*cleanup)(struct intel_engine_cs *engine);
481 /* GEN8 signal/wait table - never trust comments!
482 * signal to signal to signal to signal to signal to
483 * RCS VCS BCS VECS VCS2
484 * --------------------------------------------------------------------
485 * RCS | NOP (0x00) | VCS (0x08) | BCS (0x10) | VECS (0x18) | VCS2 (0x20) |
486 * |-------------------------------------------------------------------
487 * VCS | RCS (0x28) | NOP (0x30) | BCS (0x38) | VECS (0x40) | VCS2 (0x48) |
488 * |-------------------------------------------------------------------
489 * BCS | RCS (0x50) | VCS (0x58) | NOP (0x60) | VECS (0x68) | VCS2 (0x70) |
490 * |-------------------------------------------------------------------
491 * VECS | RCS (0x78) | VCS (0x80) | BCS (0x88) | NOP (0x90) | VCS2 (0x98) |
492 * |-------------------------------------------------------------------
493 * VCS2 | RCS (0xa0) | VCS (0xa8) | BCS (0xb0) | VECS (0xb8) | NOP (0xc0) |
494 * |-------------------------------------------------------------------
497 * f(x, y) := (x->id * NUM_RINGS * seqno_size) + (seqno_size * y->id)
498 * ie. transpose of g(x, y)
500 * sync from sync from sync from sync from sync from
501 * RCS VCS BCS VECS VCS2
502 * --------------------------------------------------------------------
503 * RCS | NOP (0x00) | VCS (0x28) | BCS (0x50) | VECS (0x78) | VCS2 (0xa0) |
504 * |-------------------------------------------------------------------
505 * VCS | RCS (0x08) | NOP (0x30) | BCS (0x58) | VECS (0x80) | VCS2 (0xa8) |
506 * |-------------------------------------------------------------------
507 * BCS | RCS (0x10) | VCS (0x38) | NOP (0x60) | VECS (0x88) | VCS2 (0xb0) |
508 * |-------------------------------------------------------------------
509 * VECS | RCS (0x18) | VCS (0x40) | BCS (0x68) | NOP (0x90) | VCS2 (0xb8) |
510 * |-------------------------------------------------------------------
511 * VCS2 | RCS (0x20) | VCS (0x48) | BCS (0x70) | VECS (0x98) | NOP (0xc0) |
512 * |-------------------------------------------------------------------
515 * g(x, y) := (y->id * NUM_RINGS * seqno_size) + (seqno_size * x->id)
516 * ie. transpose of f(x, y)
519 #define GEN6_SEMAPHORE_LAST VECS_HW
520 #define GEN6_NUM_SEMAPHORES (GEN6_SEMAPHORE_LAST + 1)
521 #define GEN6_SEMAPHORES_MASK GENMASK(GEN6_SEMAPHORE_LAST, 0)
523 /* our mbox written by others */
524 u32 wait[GEN6_NUM_SEMAPHORES];
525 /* mboxes this ring signals to */
526 i915_reg_t signal[GEN6_NUM_SEMAPHORES];
530 int (*sync_to)(struct i915_request *rq,
531 struct i915_request *signal);
532 u32 *(*signal)(struct i915_request *rq, u32 *cs);
535 struct intel_engine_execlists execlists;
537 /* Contexts are pinned whilst they are active on the GPU. The last
538 * context executed remains active whilst the GPU is idle - the
539 * switch away and write to the context object only occurs on the
540 * next execution. Contexts are only unpinned on retirement of the
541 * following request ensuring that we can always write to the object
542 * on the context switch even after idling. Across suspend, we switch
543 * to the kernel context and trash it as the save may not happen
544 * before the hardware is powered down.
546 struct i915_gem_context *last_retired_context;
548 /* We track the current MI_SET_CONTEXT in order to eliminate
549 * redudant context switches. This presumes that requests are not
550 * reordered! Or when they are the tracking is updated along with
551 * the emission of individual requests into the legacy command
554 struct i915_gem_context *legacy_active_context;
555 struct i915_hw_ppgtt *legacy_active_ppgtt;
557 /* status_notifier: list of callbacks for context-switch changes */
558 struct atomic_notifier_head context_status_notifier;
560 struct intel_engine_hangcheck hangcheck;
562 #define I915_ENGINE_NEEDS_CMD_PARSER BIT(0)
563 #define I915_ENGINE_SUPPORTS_STATS BIT(1)
567 * Table of commands the command parser needs to know about
570 DECLARE_HASHTABLE(cmd_hash, I915_CMD_HASH_ORDER);
573 * Table of registers allowed in commands that read/write registers.
575 const struct drm_i915_reg_table *reg_tables;
579 * Returns the bitmask for the length field of the specified command.
580 * Return 0 for an unrecognized/invalid command.
582 * If the command parser finds an entry for a command in the engine's
583 * cmd_tables, it gets the command's length based on the table entry.
584 * If not, it calls this function to determine the per-engine length
585 * field encoding for the command (i.e. different opcode ranges use
586 * certain bits to encode the command length in the header).
588 u32 (*get_cmd_length_mask)(u32 cmd_header);
592 * @lock: Lock protecting the below fields.
596 * @enabled: Reference count indicating number of listeners.
598 unsigned int enabled;
600 * @active: Number of contexts currently scheduled in.
604 * @enabled_at: Timestamp when busy stats were enabled.
608 * @start: Timestamp of the last idle to active transition.
610 * Idle is defined as active == 0, active is active > 0.
614 * @total: Total time this engine was busy.
616 * Accumulated time not counting the most recent block in cases
617 * where engine is currently busy (active > 0).
623 static inline bool intel_engine_needs_cmd_parser(struct intel_engine_cs *engine)
625 return engine->flags & I915_ENGINE_NEEDS_CMD_PARSER;
628 static inline bool intel_engine_supports_stats(struct intel_engine_cs *engine)
630 return engine->flags & I915_ENGINE_SUPPORTS_STATS;
634 execlists_set_active(struct intel_engine_execlists *execlists,
637 __set_bit(bit, (unsigned long *)&execlists->active);
641 execlists_clear_active(struct intel_engine_execlists *execlists,
644 __clear_bit(bit, (unsigned long *)&execlists->active);
648 execlists_is_active(const struct intel_engine_execlists *execlists,
651 return test_bit(bit, (unsigned long *)&execlists->active);
655 execlists_cancel_port_requests(struct intel_engine_execlists * const execlists);
658 execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists);
660 static inline unsigned int
661 execlists_num_ports(const struct intel_engine_execlists * const execlists)
663 return execlists->port_mask + 1;
667 execlists_port_complete(struct intel_engine_execlists * const execlists,
668 struct execlist_port * const port)
670 const unsigned int m = execlists->port_mask;
672 GEM_BUG_ON(port_index(port, execlists) != 0);
673 GEM_BUG_ON(!execlists_is_active(execlists, EXECLISTS_ACTIVE_USER));
675 memmove(port, port + 1, m * sizeof(struct execlist_port));
676 memset(port + m, 0, sizeof(struct execlist_port));
679 static inline unsigned int
680 intel_engine_flag(const struct intel_engine_cs *engine)
682 return BIT(engine->id);
686 intel_read_status_page(const struct intel_engine_cs *engine, int reg)
688 /* Ensure that the compiler doesn't optimize away the load. */
689 return READ_ONCE(engine->status_page.page_addr[reg]);
693 intel_write_status_page(struct intel_engine_cs *engine, int reg, u32 value)
695 /* Writing into the status page should be done sparingly. Since
696 * we do when we are uncertain of the device state, we take a bit
697 * of extra paranoia to try and ensure that the HWS takes the value
698 * we give and that it doesn't end up trapped inside the CPU!
700 if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
702 clflush(&engine->status_page.page_addr[reg]);
703 engine->status_page.page_addr[reg] = value;
704 clflush(&engine->status_page.page_addr[reg]);
707 WRITE_ONCE(engine->status_page.page_addr[reg], value);
712 * Reads a dword out of the status page, which is written to from the command
713 * queue by automatic updates, MI_REPORT_HEAD, MI_STORE_DATA_INDEX, or
716 * The following dwords have a reserved meaning:
717 * 0x00: ISR copy, updated when an ISR bit not set in the HWSTAM changes.
718 * 0x04: ring 0 head pointer
719 * 0x05: ring 1 head pointer (915-class)
720 * 0x06: ring 2 head pointer (915-class)
721 * 0x10-0x1b: Context status DWords (GM45)
722 * 0x1f: Last written status offset. (GM45)
723 * 0x20-0x2f: Reserved (Gen6+)
725 * The area from dword 0x30 to 0x3ff is available for driver usage.
727 #define I915_GEM_HWS_INDEX 0x30
728 #define I915_GEM_HWS_INDEX_ADDR (I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT)
729 #define I915_GEM_HWS_PREEMPT_INDEX 0x32
730 #define I915_GEM_HWS_PREEMPT_ADDR (I915_GEM_HWS_PREEMPT_INDEX << MI_STORE_DWORD_INDEX_SHIFT)
731 #define I915_GEM_HWS_SCRATCH_INDEX 0x40
732 #define I915_GEM_HWS_SCRATCH_ADDR (I915_GEM_HWS_SCRATCH_INDEX << MI_STORE_DWORD_INDEX_SHIFT)
734 #define I915_HWS_CSB_BUF0_INDEX 0x10
735 #define I915_HWS_CSB_WRITE_INDEX 0x1f
736 #define CNL_HWS_CSB_WRITE_INDEX 0x2f
739 intel_engine_create_ring(struct intel_engine_cs *engine, int size);
740 int intel_ring_pin(struct intel_ring *ring,
741 struct drm_i915_private *i915,
742 unsigned int offset_bias);
743 void intel_ring_reset(struct intel_ring *ring, u32 tail);
744 unsigned int intel_ring_update_space(struct intel_ring *ring);
745 void intel_ring_unpin(struct intel_ring *ring);
746 void intel_ring_free(struct intel_ring *ring);
748 void intel_engine_stop(struct intel_engine_cs *engine);
749 void intel_engine_cleanup(struct intel_engine_cs *engine);
751 void intel_legacy_submission_resume(struct drm_i915_private *dev_priv);
753 int __must_check intel_ring_cacheline_align(struct i915_request *rq);
755 int intel_ring_wait_for_space(struct intel_ring *ring, unsigned int bytes);
756 u32 __must_check *intel_ring_begin(struct i915_request *rq, unsigned int n);
758 static inline void intel_ring_advance(struct i915_request *rq, u32 *cs)
762 * This serves as a placeholder in the code so that the reader
763 * can compare against the preceding intel_ring_begin() and
764 * check that the number of dwords emitted matches the space
765 * reserved for the command packet (i.e. the value passed to
766 * intel_ring_begin()).
768 GEM_BUG_ON((rq->ring->vaddr + rq->ring->emit) != cs);
771 static inline u32 intel_ring_wrap(const struct intel_ring *ring, u32 pos)
773 return pos & (ring->size - 1);
776 static inline u32 intel_ring_offset(const struct i915_request *rq, void *addr)
778 /* Don't write ring->size (equivalent to 0) as that hangs some GPUs. */
779 u32 offset = addr - rq->ring->vaddr;
780 GEM_BUG_ON(offset > rq->ring->size);
781 return intel_ring_wrap(rq->ring, offset);
785 assert_ring_tail_valid(const struct intel_ring *ring, unsigned int tail)
787 /* We could combine these into a single tail operation, but keeping
788 * them as seperate tests will help identify the cause should one
791 GEM_BUG_ON(!IS_ALIGNED(tail, 8));
792 GEM_BUG_ON(tail >= ring->size);
796 * Gen2 BSpec "1. Programming Environment" / 1.4.4.6
797 * Gen3 BSpec "1c Memory Interface Functions" / 2.3.4.5
798 * Gen4+ BSpec "1c Memory Interface and Command Stream" / 5.3.4.5
799 * "If the Ring Buffer Head Pointer and the Tail Pointer are on the
800 * same cacheline, the Head Pointer must not be greater than the Tail
803 * We use ring->head as the last known location of the actual RING_HEAD,
804 * it may have advanced but in the worst case it is equally the same
805 * as ring->head and so we should never program RING_TAIL to advance
806 * into the same cacheline as ring->head.
808 #define cacheline(a) round_down(a, CACHELINE_BYTES)
809 GEM_BUG_ON(cacheline(tail) == cacheline(ring->head) &&
814 static inline unsigned int
815 intel_ring_set_tail(struct intel_ring *ring, unsigned int tail)
817 /* Whilst writes to the tail are strictly order, there is no
818 * serialisation between readers and the writers. The tail may be
819 * read by i915_request_retire() just as it is being updated
820 * by execlists, as although the breadcrumb is complete, the context
821 * switch hasn't been seen.
823 assert_ring_tail_valid(ring, tail);
828 void intel_engine_init_global_seqno(struct intel_engine_cs *engine, u32 seqno);
830 void intel_engine_setup_common(struct intel_engine_cs *engine);
831 int intel_engine_init_common(struct intel_engine_cs *engine);
832 int intel_engine_create_scratch(struct intel_engine_cs *engine, int size);
833 void intel_engine_cleanup_common(struct intel_engine_cs *engine);
835 int intel_init_render_ring_buffer(struct intel_engine_cs *engine);
836 int intel_init_bsd_ring_buffer(struct intel_engine_cs *engine);
837 int intel_init_blt_ring_buffer(struct intel_engine_cs *engine);
838 int intel_init_vebox_ring_buffer(struct intel_engine_cs *engine);
840 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine);
841 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine);
843 static inline u32 intel_engine_get_seqno(struct intel_engine_cs *engine)
845 return intel_read_status_page(engine, I915_GEM_HWS_INDEX);
848 static inline u32 intel_engine_last_submit(struct intel_engine_cs *engine)
850 /* We are only peeking at the tail of the submit queue (and not the
851 * queue itself) in order to gain a hint as to the current active
852 * state of the engine. Callers are not expected to be taking
853 * engine->timeline->lock, nor are they expected to be concerned
854 * wtih serialising this hint with anything, so document it as
855 * a hint and nothing more.
857 return READ_ONCE(engine->timeline->seqno);
860 int init_workarounds_ring(struct intel_engine_cs *engine);
861 int intel_ring_workarounds_emit(struct i915_request *rq);
863 void intel_engine_get_instdone(struct intel_engine_cs *engine,
864 struct intel_instdone *instdone);
867 * Arbitrary size for largest possible 'add request' sequence. The code paths
868 * are complex and variable. Empirical measurement shows that the worst case
869 * is BDW at 192 bytes (6 + 6 + 36 dwords), then ILK at 136 bytes. However,
870 * we need to allocate double the largest single packet within that emission
871 * to account for tail wraparound (so 6 + 6 + 72 dwords for BDW).
873 #define MIN_SPACE_FOR_ADD_REQUEST 336
875 static inline u32 intel_hws_seqno_address(struct intel_engine_cs *engine)
877 return engine->status_page.ggtt_offset + I915_GEM_HWS_INDEX_ADDR;
880 static inline u32 intel_hws_preempt_done_address(struct intel_engine_cs *engine)
882 return engine->status_page.ggtt_offset + I915_GEM_HWS_PREEMPT_ADDR;
885 /* intel_breadcrumbs.c -- user interrupt bottom-half for waiters */
886 int intel_engine_init_breadcrumbs(struct intel_engine_cs *engine);
888 static inline void intel_wait_init(struct intel_wait *wait,
889 struct i915_request *rq)
895 static inline void intel_wait_init_for_seqno(struct intel_wait *wait, u32 seqno)
901 static inline bool intel_wait_has_seqno(const struct intel_wait *wait)
907 intel_wait_update_seqno(struct intel_wait *wait, u32 seqno)
910 return intel_wait_has_seqno(wait);
914 intel_wait_update_request(struct intel_wait *wait,
915 const struct i915_request *rq)
917 return intel_wait_update_seqno(wait, i915_request_global_seqno(rq));
921 intel_wait_check_seqno(const struct intel_wait *wait, u32 seqno)
923 return wait->seqno == seqno;
927 intel_wait_check_request(const struct intel_wait *wait,
928 const struct i915_request *rq)
930 return intel_wait_check_seqno(wait, i915_request_global_seqno(rq));
933 static inline bool intel_wait_complete(const struct intel_wait *wait)
935 return RB_EMPTY_NODE(&wait->node);
938 bool intel_engine_add_wait(struct intel_engine_cs *engine,
939 struct intel_wait *wait);
940 void intel_engine_remove_wait(struct intel_engine_cs *engine,
941 struct intel_wait *wait);
942 void intel_engine_enable_signaling(struct i915_request *request, bool wakeup);
943 void intel_engine_cancel_signaling(struct i915_request *request);
945 static inline bool intel_engine_has_waiter(const struct intel_engine_cs *engine)
947 return READ_ONCE(engine->breadcrumbs.irq_wait);
950 unsigned int intel_engine_wakeup(struct intel_engine_cs *engine);
951 #define ENGINE_WAKEUP_WAITER BIT(0)
952 #define ENGINE_WAKEUP_ASLEEP BIT(1)
954 void intel_engine_pin_breadcrumbs_irq(struct intel_engine_cs *engine);
955 void intel_engine_unpin_breadcrumbs_irq(struct intel_engine_cs *engine);
957 void __intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine);
958 void intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine);
960 void intel_engine_reset_breadcrumbs(struct intel_engine_cs *engine);
961 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine);
963 static inline u32 *gen8_emit_pipe_control(u32 *batch, u32 flags, u32 offset)
965 memset(batch, 0, 6 * sizeof(u32));
967 batch[0] = GFX_OP_PIPE_CONTROL(6);
975 gen8_emit_ggtt_write_rcs(u32 *cs, u32 value, u32 gtt_offset)
977 /* We're using qword write, offset should be aligned to 8 bytes. */
978 GEM_BUG_ON(!IS_ALIGNED(gtt_offset, 8));
980 /* w/a for post sync ops following a GPGPU operation we
981 * need a prior CS_STALL, which is emitted by the flush
982 * following the batch.
984 *cs++ = GFX_OP_PIPE_CONTROL(6);
985 *cs++ = PIPE_CONTROL_GLOBAL_GTT_IVB | PIPE_CONTROL_CS_STALL |
986 PIPE_CONTROL_QW_WRITE;
990 /* We're thrashing one dword of HWS. */
997 gen8_emit_ggtt_write(u32 *cs, u32 value, u32 gtt_offset)
999 /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */
1000 GEM_BUG_ON(gtt_offset & (1 << 5));
1001 /* Offset should be aligned to 8 bytes for both (QW/DW) write types */
1002 GEM_BUG_ON(!IS_ALIGNED(gtt_offset, 8));
1004 *cs++ = (MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW;
1005 *cs++ = gtt_offset | MI_FLUSH_DW_USE_GTT;
1012 bool intel_engine_is_idle(struct intel_engine_cs *engine);
1013 bool intel_engines_are_idle(struct drm_i915_private *dev_priv);
1015 bool intel_engine_has_kernel_context(const struct intel_engine_cs *engine);
1017 void intel_engines_park(struct drm_i915_private *i915);
1018 void intel_engines_unpark(struct drm_i915_private *i915);
1020 void intel_engines_reset_default_submission(struct drm_i915_private *i915);
1021 unsigned int intel_engines_has_context_isolation(struct drm_i915_private *i915);
1023 bool intel_engine_can_store_dword(struct intel_engine_cs *engine);
1026 void intel_engine_dump(struct intel_engine_cs *engine,
1027 struct drm_printer *m,
1028 const char *header, ...);
1030 struct intel_engine_cs *
1031 intel_engine_lookup_user(struct drm_i915_private *i915, u8 class, u8 instance);
1033 static inline void intel_engine_context_in(struct intel_engine_cs *engine)
1035 unsigned long flags;
1037 if (READ_ONCE(engine->stats.enabled) == 0)
1040 spin_lock_irqsave(&engine->stats.lock, flags);
1042 if (engine->stats.enabled > 0) {
1043 if (engine->stats.active++ == 0)
1044 engine->stats.start = ktime_get();
1045 GEM_BUG_ON(engine->stats.active == 0);
1048 spin_unlock_irqrestore(&engine->stats.lock, flags);
1051 static inline void intel_engine_context_out(struct intel_engine_cs *engine)
1053 unsigned long flags;
1055 if (READ_ONCE(engine->stats.enabled) == 0)
1058 spin_lock_irqsave(&engine->stats.lock, flags);
1060 if (engine->stats.enabled > 0) {
1063 if (engine->stats.active && --engine->stats.active == 0) {
1065 * Decrement the active context count and in case GPU
1066 * is now idle add up to the running total.
1068 last = ktime_sub(ktime_get(), engine->stats.start);
1070 engine->stats.total = ktime_add(engine->stats.total,
1072 } else if (engine->stats.active == 0) {
1074 * After turning on engine stats, context out might be
1075 * the first event in which case we account from the
1076 * time stats gathering was turned on.
1078 last = ktime_sub(ktime_get(), engine->stats.enabled_at);
1080 engine->stats.total = ktime_add(engine->stats.total,
1085 spin_unlock_irqrestore(&engine->stats.lock, flags);
1088 int intel_enable_engine_stats(struct intel_engine_cs *engine);
1089 void intel_disable_engine_stats(struct intel_engine_cs *engine);
1091 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine);
1093 #endif /* _INTEL_RINGBUFFER_H_ */