1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _INTEL_RINGBUFFER_H_
3 #define _INTEL_RINGBUFFER_H_
5 #include <linux/hashtable.h>
6 #include <linux/seqlock.h>
8 #include "i915_gem_batch_pool.h"
12 #include "i915_request.h"
13 #include "i915_selftest.h"
14 #include "i915_timeline.h"
15 #include "intel_gpu_commands.h"
18 struct i915_sched_attr;
20 #define I915_CMD_HASH_ORDER 9
22 /* Early gen2 devices have a cacheline of just 32 bytes, using 64 is overkill,
23 * but keeps the logic simple. Indeed, the whole purpose of this macro is just
24 * to give some inclination as to some of the magic values used in the various
27 #define CACHELINE_BYTES 64
28 #define CACHELINE_DWORDS (CACHELINE_BYTES / sizeof(uint32_t))
30 struct intel_hw_status_page {
36 #define I915_READ_TAIL(engine) I915_READ(RING_TAIL((engine)->mmio_base))
37 #define I915_WRITE_TAIL(engine, val) I915_WRITE(RING_TAIL((engine)->mmio_base), val)
39 #define I915_READ_START(engine) I915_READ(RING_START((engine)->mmio_base))
40 #define I915_WRITE_START(engine, val) I915_WRITE(RING_START((engine)->mmio_base), val)
42 #define I915_READ_HEAD(engine) I915_READ(RING_HEAD((engine)->mmio_base))
43 #define I915_WRITE_HEAD(engine, val) I915_WRITE(RING_HEAD((engine)->mmio_base), val)
45 #define I915_READ_CTL(engine) I915_READ(RING_CTL((engine)->mmio_base))
46 #define I915_WRITE_CTL(engine, val) I915_WRITE(RING_CTL((engine)->mmio_base), val)
48 #define I915_READ_IMR(engine) I915_READ(RING_IMR((engine)->mmio_base))
49 #define I915_WRITE_IMR(engine, val) I915_WRITE(RING_IMR((engine)->mmio_base), val)
51 #define I915_READ_MODE(engine) I915_READ(RING_MI_MODE((engine)->mmio_base))
52 #define I915_WRITE_MODE(engine, val) I915_WRITE(RING_MI_MODE((engine)->mmio_base), val)
54 /* seqno size is actually only a uint32, but since we plan to use MI_FLUSH_DW to
55 * do the writes, and that must have qw aligned offsets, simply pretend it's 8b.
57 enum intel_engine_hangcheck_action {
62 ENGINE_ACTIVE_SUBUNITS,
67 static inline const char *
68 hangcheck_action_to_str(const enum intel_engine_hangcheck_action a)
75 case ENGINE_ACTIVE_SEQNO:
76 return "active seqno";
77 case ENGINE_ACTIVE_HEAD:
79 case ENGINE_ACTIVE_SUBUNITS:
80 return "active subunits";
81 case ENGINE_WAIT_KICK:
90 #define I915_MAX_SLICES 3
91 #define I915_MAX_SUBSLICES 8
93 #define instdone_slice_mask(dev_priv__) \
94 (INTEL_GEN(dev_priv__) == 7 ? \
95 1 : INTEL_INFO(dev_priv__)->sseu.slice_mask)
97 #define instdone_subslice_mask(dev_priv__) \
98 (INTEL_GEN(dev_priv__) == 7 ? \
99 1 : INTEL_INFO(dev_priv__)->sseu.subslice_mask[0])
101 #define for_each_instdone_slice_subslice(dev_priv__, slice__, subslice__) \
102 for ((slice__) = 0, (subslice__) = 0; \
103 (slice__) < I915_MAX_SLICES; \
104 (subslice__) = ((subslice__) + 1) < I915_MAX_SUBSLICES ? (subslice__) + 1 : 0, \
105 (slice__) += ((subslice__) == 0)) \
106 for_each_if((BIT(slice__) & instdone_slice_mask(dev_priv__)) && \
107 (BIT(subslice__) & instdone_subslice_mask(dev_priv__)))
109 struct intel_instdone {
111 /* The following exist only in the RCS engine */
113 u32 sampler[I915_MAX_SLICES][I915_MAX_SUBSLICES];
114 u32 row[I915_MAX_SLICES][I915_MAX_SUBSLICES];
117 struct intel_engine_hangcheck {
120 enum intel_engine_hangcheck_action action;
121 unsigned long action_timestamp;
123 struct intel_instdone instdone;
124 struct i915_request *active_request;
129 struct i915_vma *vma;
132 struct i915_timeline *timeline;
133 struct list_head request_list;
134 struct list_head active_link;
145 struct i915_gem_context;
146 struct drm_i915_reg_table;
149 * we use a single page to load ctx workarounds so all of these
150 * values are referred in terms of dwords
152 * struct i915_wa_ctx_bb:
153 * offset: specifies batch starting position, also helpful in case
154 * if we want to have multiple batches at different offsets based on
155 * some criteria. It is not a requirement at the moment but provides
156 * an option for future use.
157 * size: size of the batch in DWORDS
159 struct i915_ctx_workarounds {
160 struct i915_wa_ctx_bb {
163 } indirect_ctx, per_ctx;
164 struct i915_vma *vma;
169 #define I915_MAX_VCS 4
170 #define I915_MAX_VECS 2
173 * Engine IDs definitions.
174 * Keep instances of the same type engine together.
176 enum intel_engine_id {
183 #define _VCS(n) (VCS + (n))
186 #define _VECS(n) (VECS + (n))
189 struct i915_priolist {
191 struct list_head requests;
196 * struct intel_engine_execlists - execlist submission queue and port state
198 * The struct intel_engine_execlists represents the combined logical state of
199 * driver and the hardware state for execlist mode of submission.
201 struct intel_engine_execlists {
203 * @tasklet: softirq tasklet for bottom handler
205 struct tasklet_struct tasklet;
208 * @default_priolist: priority list for I915_PRIORITY_NORMAL
210 struct i915_priolist default_priolist;
213 * @no_priolist: priority lists disabled
218 * @submit_reg: gen-specific execlist submission register
219 * set to the ExecList Submission Port (elsp) register pre-Gen11 and to
220 * the ExecList Submission Queue Contents register array for Gen11+
222 u32 __iomem *submit_reg;
225 * @ctrl_reg: the enhanced execlists control register, used to load the
226 * submit queue on the HW and to request preemptions to idle
228 u32 __iomem *ctrl_reg;
231 * @port: execlist port states
233 * For each hardware ELSP (ExecList Submission Port) we keep
234 * track of the last request and the number of times we submitted
235 * that port to hw. We then count the number of times the hw reports
236 * a context completion or preemption. As only one context can
237 * be active on hw, we limit resubmission of context to port[0]. This
238 * is called Lite Restore, of the context.
240 struct execlist_port {
242 * @request_count: combined request and submission count
244 struct i915_request *request_count;
245 #define EXECLIST_COUNT_BITS 2
246 #define port_request(p) ptr_mask_bits((p)->request_count, EXECLIST_COUNT_BITS)
247 #define port_count(p) ptr_unmask_bits((p)->request_count, EXECLIST_COUNT_BITS)
248 #define port_pack(rq, count) ptr_pack_bits(rq, count, EXECLIST_COUNT_BITS)
249 #define port_unpack(p, count) ptr_unpack_bits((p)->request_count, count, EXECLIST_COUNT_BITS)
250 #define port_set(p, packed) ((p)->request_count = (packed))
251 #define port_isset(p) ((p)->request_count)
252 #define port_index(p, execlists) ((p) - (execlists)->port)
255 * @context_id: context ID for port
257 GEM_DEBUG_DECL(u32 context_id);
259 #define EXECLIST_MAX_PORTS 2
260 } port[EXECLIST_MAX_PORTS];
263 * @active: is the HW active? We consider the HW as active after
264 * submitting any context for execution and until we have seen the
265 * last context completion event. After that, we do not expect any
266 * more events until we submit, and so can park the HW.
268 * As we have a small number of different sources from which we feed
269 * the HW, we track the state of each inside a single bitfield.
272 #define EXECLISTS_ACTIVE_USER 0
273 #define EXECLISTS_ACTIVE_PREEMPT 1
274 #define EXECLISTS_ACTIVE_HWACK 2
277 * @port_mask: number of execlist ports - 1
279 unsigned int port_mask;
282 * @queue_priority: Highest pending priority.
284 * When we add requests into the queue, or adjust the priority of
285 * executing requests, we compute the maximum priority of those
286 * pending requests. We can then use this value to determine if
287 * we need to preempt the executing requests to service the queue.
292 * @queue: queue of requests, in priority lists
294 struct rb_root queue;
297 * @first: leftmost level in priority @queue
299 struct rb_node *first;
302 * @fw_domains: forcewake domains for irq tasklet
304 unsigned int fw_domains;
307 * @csb_head: context status buffer head
309 unsigned int csb_head;
312 * @csb_use_mmio: access csb through mmio, instead of hwsp
317 * @preempt_complete_status: expected CSB upon completing preemption
319 u32 preempt_complete_status;
322 #define INTEL_ENGINE_CS_MAX_NAME 8
324 struct intel_engine_cs {
325 struct drm_i915_private *i915;
326 char name[INTEL_ENGINE_CS_MAX_NAME];
328 enum intel_engine_id id;
340 struct intel_ring *buffer;
342 struct i915_timeline timeline;
344 struct drm_i915_gem_object *default_state;
347 unsigned long irq_posted;
348 #define ENGINE_IRQ_BREADCRUMB 0
349 #define ENGINE_IRQ_EXECLIST 1
351 /* Rather than have every client wait upon all user interrupts,
352 * with the herd waking after every interrupt and each doing the
353 * heavyweight seqno dance, we delegate the task (of being the
354 * bottom-half of the user interrupt) to the first client. After
355 * every interrupt, we wake up one client, who does the heavyweight
356 * coherent seqno read and either goes back to sleep (if incomplete),
357 * or wakes up all the completed clients in parallel, before then
358 * transferring the bottom-half status to the next client in the queue.
360 * Compared to walking the entire list of waiters in a single dedicated
361 * bottom-half, we reduce the latency of the first waiter by avoiding
362 * a context switch, but incur additional coherent seqno reads when
363 * following the chain of request breadcrumbs. Since it is most likely
364 * that we have a single client waiting on each seqno, then reducing
365 * the overhead of waking that client is much preferred.
367 struct intel_breadcrumbs {
368 spinlock_t irq_lock; /* protects irq_*; irqsafe */
369 struct intel_wait *irq_wait; /* oldest waiter by retirement */
371 spinlock_t rb_lock; /* protects the rb and wraps irq_lock */
372 struct rb_root waiters; /* sorted by retirement, priority */
373 struct list_head signals; /* sorted by retirement */
374 struct task_struct *signaler; /* used for fence signalling */
376 struct timer_list fake_irq; /* used after a missed interrupt */
377 struct timer_list hangcheck; /* detect missed interrupts */
379 unsigned int hangcheck_interrupts;
380 unsigned int irq_enabled;
383 I915_SELFTEST_DECLARE(bool mock : 1);
388 * @enable: Bitmask of enable sample events on this engine.
390 * Bits correspond to sample event types, for instance
391 * I915_SAMPLE_QUEUED is bit 0 etc.
395 * @enable_count: Reference count for the enabled samplers.
397 * Index number corresponds to the bit number from @enable.
399 unsigned int enable_count[I915_PMU_SAMPLE_BITS];
401 * @sample: Counter values for sampling events.
403 * Our internal timer stores the current counters in this field.
405 #define I915_ENGINE_SAMPLE_MAX (I915_SAMPLE_SEMA + 1)
406 struct i915_pmu_sample sample[I915_ENGINE_SAMPLE_MAX];
410 * A pool of objects to use as shadow copies of client batch buffers
411 * when the command parser is enabled. Prevents the client from
412 * modifying the batch contents after software parsing.
414 struct i915_gem_batch_pool batch_pool;
416 struct intel_hw_status_page status_page;
417 struct i915_ctx_workarounds wa_ctx;
418 struct i915_vma *scratch;
420 u32 irq_keep_mask; /* always keep these interrupts */
421 u32 irq_enable_mask; /* bitmask to enable ring interrupt */
422 void (*irq_enable)(struct intel_engine_cs *engine);
423 void (*irq_disable)(struct intel_engine_cs *engine);
425 int (*init_hw)(struct intel_engine_cs *engine);
426 void (*reset_hw)(struct intel_engine_cs *engine,
427 struct i915_request *rq);
429 void (*park)(struct intel_engine_cs *engine);
430 void (*unpark)(struct intel_engine_cs *engine);
432 void (*set_default_submission)(struct intel_engine_cs *engine);
434 struct intel_ring *(*context_pin)(struct intel_engine_cs *engine,
435 struct i915_gem_context *ctx);
436 void (*context_unpin)(struct intel_engine_cs *engine,
437 struct i915_gem_context *ctx);
438 int (*request_alloc)(struct i915_request *rq);
439 int (*init_context)(struct i915_request *rq);
441 int (*emit_flush)(struct i915_request *request, u32 mode);
442 #define EMIT_INVALIDATE BIT(0)
443 #define EMIT_FLUSH BIT(1)
444 #define EMIT_BARRIER (EMIT_INVALIDATE | EMIT_FLUSH)
445 int (*emit_bb_start)(struct i915_request *rq,
446 u64 offset, u32 length,
447 unsigned int dispatch_flags);
448 #define I915_DISPATCH_SECURE BIT(0)
449 #define I915_DISPATCH_PINNED BIT(1)
450 #define I915_DISPATCH_RS BIT(2)
451 void (*emit_breadcrumb)(struct i915_request *rq, u32 *cs);
452 int emit_breadcrumb_sz;
454 /* Pass the request to the hardware queue (e.g. directly into
455 * the legacy ringbuffer or to the end of an execlist).
457 * This is called from an atomic context with irqs disabled; must
460 void (*submit_request)(struct i915_request *rq);
462 /* Call when the priority on a request has changed and it and its
463 * dependencies may need rescheduling. Note the request itself may
464 * not be ready to run!
466 * Called under the struct_mutex.
468 void (*schedule)(struct i915_request *request,
469 const struct i915_sched_attr *attr);
472 * Cancel all requests on the hardware, or queued for execution.
473 * This should only cancel the ready requests that have been
474 * submitted to the engine (via the engine->submit_request callback).
475 * This is called when marking the device as wedged.
477 void (*cancel_requests)(struct intel_engine_cs *engine);
479 /* Some chipsets are not quite as coherent as advertised and need
480 * an expensive kick to force a true read of the up-to-date seqno.
481 * However, the up-to-date seqno is not always required and the last
482 * seen value is good enough. Note that the seqno will always be
483 * monotonic, even if not coherent.
485 void (*irq_seqno_barrier)(struct intel_engine_cs *engine);
486 void (*cleanup)(struct intel_engine_cs *engine);
488 /* GEN8 signal/wait table - never trust comments!
489 * signal to signal to signal to signal to signal to
490 * RCS VCS BCS VECS VCS2
491 * --------------------------------------------------------------------
492 * RCS | NOP (0x00) | VCS (0x08) | BCS (0x10) | VECS (0x18) | VCS2 (0x20) |
493 * |-------------------------------------------------------------------
494 * VCS | RCS (0x28) | NOP (0x30) | BCS (0x38) | VECS (0x40) | VCS2 (0x48) |
495 * |-------------------------------------------------------------------
496 * BCS | RCS (0x50) | VCS (0x58) | NOP (0x60) | VECS (0x68) | VCS2 (0x70) |
497 * |-------------------------------------------------------------------
498 * VECS | RCS (0x78) | VCS (0x80) | BCS (0x88) | NOP (0x90) | VCS2 (0x98) |
499 * |-------------------------------------------------------------------
500 * VCS2 | RCS (0xa0) | VCS (0xa8) | BCS (0xb0) | VECS (0xb8) | NOP (0xc0) |
501 * |-------------------------------------------------------------------
504 * f(x, y) := (x->id * NUM_RINGS * seqno_size) + (seqno_size * y->id)
505 * ie. transpose of g(x, y)
507 * sync from sync from sync from sync from sync from
508 * RCS VCS BCS VECS VCS2
509 * --------------------------------------------------------------------
510 * RCS | NOP (0x00) | VCS (0x28) | BCS (0x50) | VECS (0x78) | VCS2 (0xa0) |
511 * |-------------------------------------------------------------------
512 * VCS | RCS (0x08) | NOP (0x30) | BCS (0x58) | VECS (0x80) | VCS2 (0xa8) |
513 * |-------------------------------------------------------------------
514 * BCS | RCS (0x10) | VCS (0x38) | NOP (0x60) | VECS (0x88) | VCS2 (0xb0) |
515 * |-------------------------------------------------------------------
516 * VECS | RCS (0x18) | VCS (0x40) | BCS (0x68) | NOP (0x90) | VCS2 (0xb8) |
517 * |-------------------------------------------------------------------
518 * VCS2 | RCS (0x20) | VCS (0x48) | BCS (0x70) | VECS (0x98) | NOP (0xc0) |
519 * |-------------------------------------------------------------------
522 * g(x, y) := (y->id * NUM_RINGS * seqno_size) + (seqno_size * x->id)
523 * ie. transpose of f(x, y)
526 #define GEN6_SEMAPHORE_LAST VECS_HW
527 #define GEN6_NUM_SEMAPHORES (GEN6_SEMAPHORE_LAST + 1)
528 #define GEN6_SEMAPHORES_MASK GENMASK(GEN6_SEMAPHORE_LAST, 0)
530 /* our mbox written by others */
531 u32 wait[GEN6_NUM_SEMAPHORES];
532 /* mboxes this ring signals to */
533 i915_reg_t signal[GEN6_NUM_SEMAPHORES];
537 int (*sync_to)(struct i915_request *rq,
538 struct i915_request *signal);
539 u32 *(*signal)(struct i915_request *rq, u32 *cs);
542 struct intel_engine_execlists execlists;
544 /* Contexts are pinned whilst they are active on the GPU. The last
545 * context executed remains active whilst the GPU is idle - the
546 * switch away and write to the context object only occurs on the
547 * next execution. Contexts are only unpinned on retirement of the
548 * following request ensuring that we can always write to the object
549 * on the context switch even after idling. Across suspend, we switch
550 * to the kernel context and trash it as the save may not happen
551 * before the hardware is powered down.
553 struct i915_gem_context *last_retired_context;
555 /* We track the current MI_SET_CONTEXT in order to eliminate
556 * redudant context switches. This presumes that requests are not
557 * reordered! Or when they are the tracking is updated along with
558 * the emission of individual requests into the legacy command
561 struct i915_gem_context *legacy_active_context;
562 struct i915_hw_ppgtt *legacy_active_ppgtt;
564 /* status_notifier: list of callbacks for context-switch changes */
565 struct atomic_notifier_head context_status_notifier;
567 struct intel_engine_hangcheck hangcheck;
569 #define I915_ENGINE_NEEDS_CMD_PARSER BIT(0)
570 #define I915_ENGINE_SUPPORTS_STATS BIT(1)
571 #define I915_ENGINE_HAS_PREEMPTION BIT(2)
575 * Table of commands the command parser needs to know about
578 DECLARE_HASHTABLE(cmd_hash, I915_CMD_HASH_ORDER);
581 * Table of registers allowed in commands that read/write registers.
583 const struct drm_i915_reg_table *reg_tables;
587 * Returns the bitmask for the length field of the specified command.
588 * Return 0 for an unrecognized/invalid command.
590 * If the command parser finds an entry for a command in the engine's
591 * cmd_tables, it gets the command's length based on the table entry.
592 * If not, it calls this function to determine the per-engine length
593 * field encoding for the command (i.e. different opcode ranges use
594 * certain bits to encode the command length in the header).
596 u32 (*get_cmd_length_mask)(u32 cmd_header);
600 * @lock: Lock protecting the below fields.
604 * @enabled: Reference count indicating number of listeners.
606 unsigned int enabled;
608 * @active: Number of contexts currently scheduled in.
612 * @enabled_at: Timestamp when busy stats were enabled.
616 * @start: Timestamp of the last idle to active transition.
618 * Idle is defined as active == 0, active is active > 0.
622 * @total: Total time this engine was busy.
624 * Accumulated time not counting the most recent block in cases
625 * where engine is currently busy (active > 0).
632 intel_engine_needs_cmd_parser(const struct intel_engine_cs *engine)
634 return engine->flags & I915_ENGINE_NEEDS_CMD_PARSER;
638 intel_engine_supports_stats(const struct intel_engine_cs *engine)
640 return engine->flags & I915_ENGINE_SUPPORTS_STATS;
644 intel_engine_has_preemption(const struct intel_engine_cs *engine)
646 return engine->flags & I915_ENGINE_HAS_PREEMPTION;
649 static inline bool __execlists_need_preempt(int prio, int last)
651 return prio > max(0, last);
655 execlists_set_active(struct intel_engine_execlists *execlists,
658 __set_bit(bit, (unsigned long *)&execlists->active);
662 execlists_set_active_once(struct intel_engine_execlists *execlists,
665 return !__test_and_set_bit(bit, (unsigned long *)&execlists->active);
669 execlists_clear_active(struct intel_engine_execlists *execlists,
672 __clear_bit(bit, (unsigned long *)&execlists->active);
676 execlists_is_active(const struct intel_engine_execlists *execlists,
679 return test_bit(bit, (unsigned long *)&execlists->active);
682 void execlists_user_begin(struct intel_engine_execlists *execlists,
683 const struct execlist_port *port);
684 void execlists_user_end(struct intel_engine_execlists *execlists);
687 execlists_cancel_port_requests(struct intel_engine_execlists * const execlists);
690 execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists);
692 static inline unsigned int
693 execlists_num_ports(const struct intel_engine_execlists * const execlists)
695 return execlists->port_mask + 1;
698 static inline struct execlist_port *
699 execlists_port_complete(struct intel_engine_execlists * const execlists,
700 struct execlist_port * const port)
702 const unsigned int m = execlists->port_mask;
704 GEM_BUG_ON(port_index(port, execlists) != 0);
705 GEM_BUG_ON(!execlists_is_active(execlists, EXECLISTS_ACTIVE_USER));
707 memmove(port, port + 1, m * sizeof(struct execlist_port));
708 memset(port + m, 0, sizeof(struct execlist_port));
713 static inline unsigned int
714 intel_engine_flag(const struct intel_engine_cs *engine)
716 return BIT(engine->id);
720 intel_read_status_page(const struct intel_engine_cs *engine, int reg)
722 /* Ensure that the compiler doesn't optimize away the load. */
723 return READ_ONCE(engine->status_page.page_addr[reg]);
727 intel_write_status_page(struct intel_engine_cs *engine, int reg, u32 value)
729 /* Writing into the status page should be done sparingly. Since
730 * we do when we are uncertain of the device state, we take a bit
731 * of extra paranoia to try and ensure that the HWS takes the value
732 * we give and that it doesn't end up trapped inside the CPU!
734 if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
736 clflush(&engine->status_page.page_addr[reg]);
737 engine->status_page.page_addr[reg] = value;
738 clflush(&engine->status_page.page_addr[reg]);
741 WRITE_ONCE(engine->status_page.page_addr[reg], value);
746 * Reads a dword out of the status page, which is written to from the command
747 * queue by automatic updates, MI_REPORT_HEAD, MI_STORE_DATA_INDEX, or
750 * The following dwords have a reserved meaning:
751 * 0x00: ISR copy, updated when an ISR bit not set in the HWSTAM changes.
752 * 0x04: ring 0 head pointer
753 * 0x05: ring 1 head pointer (915-class)
754 * 0x06: ring 2 head pointer (915-class)
755 * 0x10-0x1b: Context status DWords (GM45)
756 * 0x1f: Last written status offset. (GM45)
757 * 0x20-0x2f: Reserved (Gen6+)
759 * The area from dword 0x30 to 0x3ff is available for driver usage.
761 #define I915_GEM_HWS_INDEX 0x30
762 #define I915_GEM_HWS_INDEX_ADDR (I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT)
763 #define I915_GEM_HWS_PREEMPT_INDEX 0x32
764 #define I915_GEM_HWS_PREEMPT_ADDR (I915_GEM_HWS_PREEMPT_INDEX << MI_STORE_DWORD_INDEX_SHIFT)
765 #define I915_GEM_HWS_SCRATCH_INDEX 0x40
766 #define I915_GEM_HWS_SCRATCH_ADDR (I915_GEM_HWS_SCRATCH_INDEX << MI_STORE_DWORD_INDEX_SHIFT)
768 #define I915_HWS_CSB_BUF0_INDEX 0x10
769 #define I915_HWS_CSB_WRITE_INDEX 0x1f
770 #define CNL_HWS_CSB_WRITE_INDEX 0x2f
773 intel_engine_create_ring(struct intel_engine_cs *engine,
774 struct i915_timeline *timeline,
776 int intel_ring_pin(struct intel_ring *ring,
777 struct drm_i915_private *i915,
778 unsigned int offset_bias);
779 void intel_ring_reset(struct intel_ring *ring, u32 tail);
780 unsigned int intel_ring_update_space(struct intel_ring *ring);
781 void intel_ring_unpin(struct intel_ring *ring);
782 void intel_ring_free(struct intel_ring *ring);
784 void intel_engine_stop(struct intel_engine_cs *engine);
785 void intel_engine_cleanup(struct intel_engine_cs *engine);
787 void intel_legacy_submission_resume(struct drm_i915_private *dev_priv);
789 int __must_check intel_ring_cacheline_align(struct i915_request *rq);
791 int intel_ring_wait_for_space(struct intel_ring *ring, unsigned int bytes);
792 u32 __must_check *intel_ring_begin(struct i915_request *rq, unsigned int n);
794 static inline void intel_ring_advance(struct i915_request *rq, u32 *cs)
798 * This serves as a placeholder in the code so that the reader
799 * can compare against the preceding intel_ring_begin() and
800 * check that the number of dwords emitted matches the space
801 * reserved for the command packet (i.e. the value passed to
802 * intel_ring_begin()).
804 GEM_BUG_ON((rq->ring->vaddr + rq->ring->emit) != cs);
807 static inline u32 intel_ring_wrap(const struct intel_ring *ring, u32 pos)
809 return pos & (ring->size - 1);
812 static inline u32 intel_ring_offset(const struct i915_request *rq, void *addr)
814 /* Don't write ring->size (equivalent to 0) as that hangs some GPUs. */
815 u32 offset = addr - rq->ring->vaddr;
816 GEM_BUG_ON(offset > rq->ring->size);
817 return intel_ring_wrap(rq->ring, offset);
821 assert_ring_tail_valid(const struct intel_ring *ring, unsigned int tail)
823 /* We could combine these into a single tail operation, but keeping
824 * them as seperate tests will help identify the cause should one
827 GEM_BUG_ON(!IS_ALIGNED(tail, 8));
828 GEM_BUG_ON(tail >= ring->size);
832 * Gen2 BSpec "1. Programming Environment" / 1.4.4.6
833 * Gen3 BSpec "1c Memory Interface Functions" / 2.3.4.5
834 * Gen4+ BSpec "1c Memory Interface and Command Stream" / 5.3.4.5
835 * "If the Ring Buffer Head Pointer and the Tail Pointer are on the
836 * same cacheline, the Head Pointer must not be greater than the Tail
839 * We use ring->head as the last known location of the actual RING_HEAD,
840 * it may have advanced but in the worst case it is equally the same
841 * as ring->head and so we should never program RING_TAIL to advance
842 * into the same cacheline as ring->head.
844 #define cacheline(a) round_down(a, CACHELINE_BYTES)
845 GEM_BUG_ON(cacheline(tail) == cacheline(ring->head) &&
850 static inline unsigned int
851 intel_ring_set_tail(struct intel_ring *ring, unsigned int tail)
853 /* Whilst writes to the tail are strictly order, there is no
854 * serialisation between readers and the writers. The tail may be
855 * read by i915_request_retire() just as it is being updated
856 * by execlists, as although the breadcrumb is complete, the context
857 * switch hasn't been seen.
859 assert_ring_tail_valid(ring, tail);
864 void intel_engine_init_global_seqno(struct intel_engine_cs *engine, u32 seqno);
866 void intel_engine_setup_common(struct intel_engine_cs *engine);
867 int intel_engine_init_common(struct intel_engine_cs *engine);
868 int intel_engine_create_scratch(struct intel_engine_cs *engine, int size);
869 void intel_engine_cleanup_common(struct intel_engine_cs *engine);
871 int intel_init_render_ring_buffer(struct intel_engine_cs *engine);
872 int intel_init_bsd_ring_buffer(struct intel_engine_cs *engine);
873 int intel_init_blt_ring_buffer(struct intel_engine_cs *engine);
874 int intel_init_vebox_ring_buffer(struct intel_engine_cs *engine);
876 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine);
877 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine);
879 static inline u32 intel_engine_get_seqno(struct intel_engine_cs *engine)
881 return intel_read_status_page(engine, I915_GEM_HWS_INDEX);
884 static inline u32 intel_engine_last_submit(struct intel_engine_cs *engine)
886 /* We are only peeking at the tail of the submit queue (and not the
887 * queue itself) in order to gain a hint as to the current active
888 * state of the engine. Callers are not expected to be taking
889 * engine->timeline->lock, nor are they expected to be concerned
890 * wtih serialising this hint with anything, so document it as
891 * a hint and nothing more.
893 return READ_ONCE(engine->timeline.seqno);
896 void intel_engine_get_instdone(struct intel_engine_cs *engine,
897 struct intel_instdone *instdone);
900 * Arbitrary size for largest possible 'add request' sequence. The code paths
901 * are complex and variable. Empirical measurement shows that the worst case
902 * is BDW at 192 bytes (6 + 6 + 36 dwords), then ILK at 136 bytes. However,
903 * we need to allocate double the largest single packet within that emission
904 * to account for tail wraparound (so 6 + 6 + 72 dwords for BDW).
906 #define MIN_SPACE_FOR_ADD_REQUEST 336
908 static inline u32 intel_hws_seqno_address(struct intel_engine_cs *engine)
910 return engine->status_page.ggtt_offset + I915_GEM_HWS_INDEX_ADDR;
913 static inline u32 intel_hws_preempt_done_address(struct intel_engine_cs *engine)
915 return engine->status_page.ggtt_offset + I915_GEM_HWS_PREEMPT_ADDR;
918 /* intel_breadcrumbs.c -- user interrupt bottom-half for waiters */
919 int intel_engine_init_breadcrumbs(struct intel_engine_cs *engine);
921 static inline void intel_wait_init(struct intel_wait *wait,
922 struct i915_request *rq)
928 static inline void intel_wait_init_for_seqno(struct intel_wait *wait, u32 seqno)
934 static inline bool intel_wait_has_seqno(const struct intel_wait *wait)
940 intel_wait_update_seqno(struct intel_wait *wait, u32 seqno)
943 return intel_wait_has_seqno(wait);
947 intel_wait_update_request(struct intel_wait *wait,
948 const struct i915_request *rq)
950 return intel_wait_update_seqno(wait, i915_request_global_seqno(rq));
954 intel_wait_check_seqno(const struct intel_wait *wait, u32 seqno)
956 return wait->seqno == seqno;
960 intel_wait_check_request(const struct intel_wait *wait,
961 const struct i915_request *rq)
963 return intel_wait_check_seqno(wait, i915_request_global_seqno(rq));
966 static inline bool intel_wait_complete(const struct intel_wait *wait)
968 return RB_EMPTY_NODE(&wait->node);
971 bool intel_engine_add_wait(struct intel_engine_cs *engine,
972 struct intel_wait *wait);
973 void intel_engine_remove_wait(struct intel_engine_cs *engine,
974 struct intel_wait *wait);
975 bool intel_engine_enable_signaling(struct i915_request *request, bool wakeup);
976 void intel_engine_cancel_signaling(struct i915_request *request);
978 static inline bool intel_engine_has_waiter(const struct intel_engine_cs *engine)
980 return READ_ONCE(engine->breadcrumbs.irq_wait);
983 unsigned int intel_engine_wakeup(struct intel_engine_cs *engine);
984 #define ENGINE_WAKEUP_WAITER BIT(0)
985 #define ENGINE_WAKEUP_ASLEEP BIT(1)
987 void intel_engine_pin_breadcrumbs_irq(struct intel_engine_cs *engine);
988 void intel_engine_unpin_breadcrumbs_irq(struct intel_engine_cs *engine);
990 void __intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine);
991 void intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine);
993 void intel_engine_reset_breadcrumbs(struct intel_engine_cs *engine);
994 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine);
996 static inline u32 *gen8_emit_pipe_control(u32 *batch, u32 flags, u32 offset)
998 memset(batch, 0, 6 * sizeof(u32));
1000 batch[0] = GFX_OP_PIPE_CONTROL(6);
1008 gen8_emit_ggtt_write_rcs(u32 *cs, u32 value, u32 gtt_offset)
1010 /* We're using qword write, offset should be aligned to 8 bytes. */
1011 GEM_BUG_ON(!IS_ALIGNED(gtt_offset, 8));
1013 /* w/a for post sync ops following a GPGPU operation we
1014 * need a prior CS_STALL, which is emitted by the flush
1015 * following the batch.
1017 *cs++ = GFX_OP_PIPE_CONTROL(6);
1018 *cs++ = PIPE_CONTROL_GLOBAL_GTT_IVB | PIPE_CONTROL_CS_STALL |
1019 PIPE_CONTROL_QW_WRITE;
1023 /* We're thrashing one dword of HWS. */
1030 gen8_emit_ggtt_write(u32 *cs, u32 value, u32 gtt_offset)
1032 /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */
1033 GEM_BUG_ON(gtt_offset & (1 << 5));
1034 /* Offset should be aligned to 8 bytes for both (QW/DW) write types */
1035 GEM_BUG_ON(!IS_ALIGNED(gtt_offset, 8));
1037 *cs++ = (MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW;
1038 *cs++ = gtt_offset | MI_FLUSH_DW_USE_GTT;
1045 bool intel_engine_is_idle(struct intel_engine_cs *engine);
1046 bool intel_engines_are_idle(struct drm_i915_private *dev_priv);
1048 bool intel_engine_has_kernel_context(const struct intel_engine_cs *engine);
1050 void intel_engines_park(struct drm_i915_private *i915);
1051 void intel_engines_unpark(struct drm_i915_private *i915);
1053 void intel_engines_reset_default_submission(struct drm_i915_private *i915);
1054 unsigned int intel_engines_has_context_isolation(struct drm_i915_private *i915);
1056 bool intel_engine_can_store_dword(struct intel_engine_cs *engine);
1059 void intel_engine_dump(struct intel_engine_cs *engine,
1060 struct drm_printer *m,
1061 const char *header, ...);
1063 struct intel_engine_cs *
1064 intel_engine_lookup_user(struct drm_i915_private *i915, u8 class, u8 instance);
1066 static inline void intel_engine_context_in(struct intel_engine_cs *engine)
1068 unsigned long flags;
1070 if (READ_ONCE(engine->stats.enabled) == 0)
1073 write_seqlock_irqsave(&engine->stats.lock, flags);
1075 if (engine->stats.enabled > 0) {
1076 if (engine->stats.active++ == 0)
1077 engine->stats.start = ktime_get();
1078 GEM_BUG_ON(engine->stats.active == 0);
1081 write_sequnlock_irqrestore(&engine->stats.lock, flags);
1084 static inline void intel_engine_context_out(struct intel_engine_cs *engine)
1086 unsigned long flags;
1088 if (READ_ONCE(engine->stats.enabled) == 0)
1091 write_seqlock_irqsave(&engine->stats.lock, flags);
1093 if (engine->stats.enabled > 0) {
1096 if (engine->stats.active && --engine->stats.active == 0) {
1098 * Decrement the active context count and in case GPU
1099 * is now idle add up to the running total.
1101 last = ktime_sub(ktime_get(), engine->stats.start);
1103 engine->stats.total = ktime_add(engine->stats.total,
1105 } else if (engine->stats.active == 0) {
1107 * After turning on engine stats, context out might be
1108 * the first event in which case we account from the
1109 * time stats gathering was turned on.
1111 last = ktime_sub(ktime_get(), engine->stats.enabled_at);
1113 engine->stats.total = ktime_add(engine->stats.total,
1118 write_sequnlock_irqrestore(&engine->stats.lock, flags);
1121 int intel_enable_engine_stats(struct intel_engine_cs *engine);
1122 void intel_disable_engine_stats(struct intel_engine_cs *engine);
1124 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine);
1126 #endif /* _INTEL_RINGBUFFER_H_ */