1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _INTEL_RINGBUFFER_H_
3 #define _INTEL_RINGBUFFER_H_
5 #include <linux/hashtable.h>
6 #include "i915_gem_batch_pool.h"
7 #include "i915_gem_request.h"
8 #include "i915_gem_timeline.h"
9 #include "i915_selftest.h"
13 #define I915_CMD_HASH_ORDER 9
15 /* Early gen2 devices have a cacheline of just 32 bytes, using 64 is overkill,
16 * but keeps the logic simple. Indeed, the whole purpose of this macro is just
17 * to give some inclination as to some of the magic values used in the various
20 #define CACHELINE_BYTES 64
21 #define CACHELINE_DWORDS (CACHELINE_BYTES / sizeof(uint32_t))
23 struct intel_hw_status_page {
29 #define I915_READ_TAIL(engine) I915_READ(RING_TAIL((engine)->mmio_base))
30 #define I915_WRITE_TAIL(engine, val) I915_WRITE(RING_TAIL((engine)->mmio_base), val)
32 #define I915_READ_START(engine) I915_READ(RING_START((engine)->mmio_base))
33 #define I915_WRITE_START(engine, val) I915_WRITE(RING_START((engine)->mmio_base), val)
35 #define I915_READ_HEAD(engine) I915_READ(RING_HEAD((engine)->mmio_base))
36 #define I915_WRITE_HEAD(engine, val) I915_WRITE(RING_HEAD((engine)->mmio_base), val)
38 #define I915_READ_CTL(engine) I915_READ(RING_CTL((engine)->mmio_base))
39 #define I915_WRITE_CTL(engine, val) I915_WRITE(RING_CTL((engine)->mmio_base), val)
41 #define I915_READ_IMR(engine) I915_READ(RING_IMR((engine)->mmio_base))
42 #define I915_WRITE_IMR(engine, val) I915_WRITE(RING_IMR((engine)->mmio_base), val)
44 #define I915_READ_MODE(engine) I915_READ(RING_MI_MODE((engine)->mmio_base))
45 #define I915_WRITE_MODE(engine, val) I915_WRITE(RING_MI_MODE((engine)->mmio_base), val)
47 /* seqno size is actually only a uint32, but since we plan to use MI_FLUSH_DW to
48 * do the writes, and that must have qw aligned offsets, simply pretend it's 8b.
50 #define gen8_semaphore_seqno_size sizeof(uint64_t)
51 #define GEN8_SEMAPHORE_OFFSET(__from, __to) \
52 (((__from) * I915_NUM_ENGINES + (__to)) * gen8_semaphore_seqno_size)
53 #define GEN8_SIGNAL_OFFSET(__ring, to) \
54 (dev_priv->semaphore->node.start + \
55 GEN8_SEMAPHORE_OFFSET((__ring)->id, (to)))
56 #define GEN8_WAIT_OFFSET(__ring, from) \
57 (dev_priv->semaphore->node.start + \
58 GEN8_SEMAPHORE_OFFSET(from, (__ring)->id))
60 enum intel_engine_hangcheck_action {
65 ENGINE_ACTIVE_SUBUNITS,
70 static inline const char *
71 hangcheck_action_to_str(const enum intel_engine_hangcheck_action a)
78 case ENGINE_ACTIVE_SEQNO:
79 return "active seqno";
80 case ENGINE_ACTIVE_HEAD:
82 case ENGINE_ACTIVE_SUBUNITS:
83 return "active subunits";
84 case ENGINE_WAIT_KICK:
93 #define I915_MAX_SLICES 3
94 #define I915_MAX_SUBSLICES 3
96 #define instdone_slice_mask(dev_priv__) \
97 (INTEL_GEN(dev_priv__) == 7 ? \
98 1 : INTEL_INFO(dev_priv__)->sseu.slice_mask)
100 #define instdone_subslice_mask(dev_priv__) \
101 (INTEL_GEN(dev_priv__) == 7 ? \
102 1 : INTEL_INFO(dev_priv__)->sseu.subslice_mask)
104 #define for_each_instdone_slice_subslice(dev_priv__, slice__, subslice__) \
105 for ((slice__) = 0, (subslice__) = 0; \
106 (slice__) < I915_MAX_SLICES; \
107 (subslice__) = ((subslice__) + 1) < I915_MAX_SUBSLICES ? (subslice__) + 1 : 0, \
108 (slice__) += ((subslice__) == 0)) \
109 for_each_if((BIT(slice__) & instdone_slice_mask(dev_priv__)) && \
110 (BIT(subslice__) & instdone_subslice_mask(dev_priv__)))
112 struct intel_instdone {
114 /* The following exist only in the RCS engine */
116 u32 sampler[I915_MAX_SLICES][I915_MAX_SUBSLICES];
117 u32 row[I915_MAX_SLICES][I915_MAX_SUBSLICES];
120 struct intel_engine_hangcheck {
123 enum intel_engine_hangcheck_action action;
124 unsigned long action_timestamp;
126 struct intel_instdone instdone;
127 struct drm_i915_gem_request *active_request;
132 struct i915_vma *vma;
135 struct list_head request_list;
146 struct i915_gem_context;
147 struct drm_i915_reg_table;
150 * we use a single page to load ctx workarounds so all of these
151 * values are referred in terms of dwords
153 * struct i915_wa_ctx_bb:
154 * offset: specifies batch starting position, also helpful in case
155 * if we want to have multiple batches at different offsets based on
156 * some criteria. It is not a requirement at the moment but provides
157 * an option for future use.
158 * size: size of the batch in DWORDS
160 struct i915_ctx_workarounds {
161 struct i915_wa_ctx_bb {
164 } indirect_ctx, per_ctx;
165 struct i915_vma *vma;
168 struct drm_i915_gem_request;
171 * Engine IDs definitions.
172 * Keep instances of the same type engine together.
174 enum intel_engine_id {
179 #define _VCS(n) (VCS + (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 * @port: execlist port states
214 * For each hardware ELSP (ExecList Submission Port) we keep
215 * track of the last request and the number of times we submitted
216 * that port to hw. We then count the number of times the hw reports
217 * a context completion or preemption. As only one context can
218 * be active on hw, we limit resubmission of context to port[0]. This
219 * is called Lite Restore, of the context.
221 struct execlist_port {
223 * @request_count: combined request and submission count
225 struct drm_i915_gem_request *request_count;
226 #define EXECLIST_COUNT_BITS 2
227 #define port_request(p) ptr_mask_bits((p)->request_count, EXECLIST_COUNT_BITS)
228 #define port_count(p) ptr_unmask_bits((p)->request_count, EXECLIST_COUNT_BITS)
229 #define port_pack(rq, count) ptr_pack_bits(rq, count, EXECLIST_COUNT_BITS)
230 #define port_unpack(p, count) ptr_unpack_bits((p)->request_count, count, EXECLIST_COUNT_BITS)
231 #define port_set(p, packed) ((p)->request_count = (packed))
232 #define port_isset(p) ((p)->request_count)
233 #define port_index(p, execlists) ((p) - (execlists)->port)
236 * @context_id: context ID for port
238 GEM_DEBUG_DECL(u32 context_id);
240 #define EXECLIST_MAX_PORTS 2
241 } port[EXECLIST_MAX_PORTS];
244 * @active: is the HW active? We consider the HW as active after
245 * submitting any context for execution and until we have seen the
246 * last context completion event. After that, we do not expect any
247 * more events until we submit, and so can park the HW.
249 * As we have a small number of different sources from which we feed
250 * the HW, we track the state of each inside a single bitfield.
253 #define EXECLISTS_ACTIVE_USER 0
254 #define EXECLISTS_ACTIVE_PREEMPT 1
257 * @port_mask: number of execlist ports - 1
259 unsigned int port_mask;
262 * @queue: queue of requests, in priority lists
264 struct rb_root queue;
267 * @first: leftmost level in priority @queue
269 struct rb_node *first;
272 * @fw_domains: forcewake domains for irq tasklet
274 unsigned int fw_domains;
277 * @csb_head: context status buffer head
279 unsigned int csb_head;
282 * @csb_use_mmio: access csb through mmio, instead of hwsp
287 #define INTEL_ENGINE_CS_MAX_NAME 8
289 struct intel_engine_cs {
290 struct drm_i915_private *i915;
291 char name[INTEL_ENGINE_CS_MAX_NAME];
293 enum intel_engine_id id;
304 unsigned int irq_shift;
306 struct intel_ring *buffer;
307 struct intel_timeline *timeline;
309 struct drm_i915_gem_object *default_state;
312 unsigned long irq_posted;
313 #define ENGINE_IRQ_BREADCRUMB 0
314 #define ENGINE_IRQ_EXECLIST 1
316 /* Rather than have every client wait upon all user interrupts,
317 * with the herd waking after every interrupt and each doing the
318 * heavyweight seqno dance, we delegate the task (of being the
319 * bottom-half of the user interrupt) to the first client. After
320 * every interrupt, we wake up one client, who does the heavyweight
321 * coherent seqno read and either goes back to sleep (if incomplete),
322 * or wakes up all the completed clients in parallel, before then
323 * transferring the bottom-half status to the next client in the queue.
325 * Compared to walking the entire list of waiters in a single dedicated
326 * bottom-half, we reduce the latency of the first waiter by avoiding
327 * a context switch, but incur additional coherent seqno reads when
328 * following the chain of request breadcrumbs. Since it is most likely
329 * that we have a single client waiting on each seqno, then reducing
330 * the overhead of waking that client is much preferred.
332 struct intel_breadcrumbs {
333 spinlock_t irq_lock; /* protects irq_*; irqsafe */
334 struct intel_wait *irq_wait; /* oldest waiter by retirement */
336 spinlock_t rb_lock; /* protects the rb and wraps irq_lock */
337 struct rb_root waiters; /* sorted by retirement, priority */
338 struct rb_root signals; /* sorted by retirement */
339 struct task_struct *signaler; /* used for fence signalling */
340 struct drm_i915_gem_request __rcu *first_signal;
341 struct timer_list fake_irq; /* used after a missed interrupt */
342 struct timer_list hangcheck; /* detect missed interrupts */
344 unsigned int hangcheck_interrupts;
345 unsigned int irq_enabled;
348 I915_SELFTEST_DECLARE(bool mock : 1);
352 * A pool of objects to use as shadow copies of client batch buffers
353 * when the command parser is enabled. Prevents the client from
354 * modifying the batch contents after software parsing.
356 struct i915_gem_batch_pool batch_pool;
358 struct intel_hw_status_page status_page;
359 struct i915_ctx_workarounds wa_ctx;
360 struct i915_vma *scratch;
362 u32 irq_keep_mask; /* always keep these interrupts */
363 u32 irq_enable_mask; /* bitmask to enable ring interrupt */
364 void (*irq_enable)(struct intel_engine_cs *engine);
365 void (*irq_disable)(struct intel_engine_cs *engine);
367 int (*init_hw)(struct intel_engine_cs *engine);
368 void (*reset_hw)(struct intel_engine_cs *engine,
369 struct drm_i915_gem_request *req);
371 void (*park)(struct intel_engine_cs *engine);
372 void (*unpark)(struct intel_engine_cs *engine);
374 void (*set_default_submission)(struct intel_engine_cs *engine);
376 struct intel_ring *(*context_pin)(struct intel_engine_cs *engine,
377 struct i915_gem_context *ctx);
378 void (*context_unpin)(struct intel_engine_cs *engine,
379 struct i915_gem_context *ctx);
380 int (*request_alloc)(struct drm_i915_gem_request *req);
381 int (*init_context)(struct drm_i915_gem_request *req);
383 int (*emit_flush)(struct drm_i915_gem_request *request,
385 #define EMIT_INVALIDATE BIT(0)
386 #define EMIT_FLUSH BIT(1)
387 #define EMIT_BARRIER (EMIT_INVALIDATE | EMIT_FLUSH)
388 int (*emit_bb_start)(struct drm_i915_gem_request *req,
389 u64 offset, u32 length,
390 unsigned int dispatch_flags);
391 #define I915_DISPATCH_SECURE BIT(0)
392 #define I915_DISPATCH_PINNED BIT(1)
393 #define I915_DISPATCH_RS BIT(2)
394 void (*emit_breadcrumb)(struct drm_i915_gem_request *req,
396 int emit_breadcrumb_sz;
398 /* Pass the request to the hardware queue (e.g. directly into
399 * the legacy ringbuffer or to the end of an execlist).
401 * This is called from an atomic context with irqs disabled; must
404 void (*submit_request)(struct drm_i915_gem_request *req);
406 /* Call when the priority on a request has changed and it and its
407 * dependencies may need rescheduling. Note the request itself may
408 * not be ready to run!
410 * Called under the struct_mutex.
412 void (*schedule)(struct drm_i915_gem_request *request,
416 * Cancel all requests on the hardware, or queued for execution.
417 * This should only cancel the ready requests that have been
418 * submitted to the engine (via the engine->submit_request callback).
419 * This is called when marking the device as wedged.
421 void (*cancel_requests)(struct intel_engine_cs *engine);
423 /* Some chipsets are not quite as coherent as advertised and need
424 * an expensive kick to force a true read of the up-to-date seqno.
425 * However, the up-to-date seqno is not always required and the last
426 * seen value is good enough. Note that the seqno will always be
427 * monotonic, even if not coherent.
429 void (*irq_seqno_barrier)(struct intel_engine_cs *engine);
430 void (*cleanup)(struct intel_engine_cs *engine);
432 /* GEN8 signal/wait table - never trust comments!
433 * signal to signal to signal to signal to signal to
434 * RCS VCS BCS VECS VCS2
435 * --------------------------------------------------------------------
436 * RCS | NOP (0x00) | VCS (0x08) | BCS (0x10) | VECS (0x18) | VCS2 (0x20) |
437 * |-------------------------------------------------------------------
438 * VCS | RCS (0x28) | NOP (0x30) | BCS (0x38) | VECS (0x40) | VCS2 (0x48) |
439 * |-------------------------------------------------------------------
440 * BCS | RCS (0x50) | VCS (0x58) | NOP (0x60) | VECS (0x68) | VCS2 (0x70) |
441 * |-------------------------------------------------------------------
442 * VECS | RCS (0x78) | VCS (0x80) | BCS (0x88) | NOP (0x90) | VCS2 (0x98) |
443 * |-------------------------------------------------------------------
444 * VCS2 | RCS (0xa0) | VCS (0xa8) | BCS (0xb0) | VECS (0xb8) | NOP (0xc0) |
445 * |-------------------------------------------------------------------
448 * f(x, y) := (x->id * NUM_RINGS * seqno_size) + (seqno_size * y->id)
449 * ie. transpose of g(x, y)
451 * sync from sync from sync from sync from sync from
452 * RCS VCS BCS VECS VCS2
453 * --------------------------------------------------------------------
454 * RCS | NOP (0x00) | VCS (0x28) | BCS (0x50) | VECS (0x78) | VCS2 (0xa0) |
455 * |-------------------------------------------------------------------
456 * VCS | RCS (0x08) | NOP (0x30) | BCS (0x58) | VECS (0x80) | VCS2 (0xa8) |
457 * |-------------------------------------------------------------------
458 * BCS | RCS (0x10) | VCS (0x38) | NOP (0x60) | VECS (0x88) | VCS2 (0xb0) |
459 * |-------------------------------------------------------------------
460 * VECS | RCS (0x18) | VCS (0x40) | BCS (0x68) | NOP (0x90) | VCS2 (0xb8) |
461 * |-------------------------------------------------------------------
462 * VCS2 | RCS (0x20) | VCS (0x48) | BCS (0x70) | VECS (0x98) | NOP (0xc0) |
463 * |-------------------------------------------------------------------
466 * g(x, y) := (y->id * NUM_RINGS * seqno_size) + (seqno_size * x->id)
467 * ie. transpose of f(x, y)
471 #define GEN6_SEMAPHORE_LAST VECS_HW
472 #define GEN6_NUM_SEMAPHORES (GEN6_SEMAPHORE_LAST + 1)
473 #define GEN6_SEMAPHORES_MASK GENMASK(GEN6_SEMAPHORE_LAST, 0)
475 /* our mbox written by others */
476 u32 wait[GEN6_NUM_SEMAPHORES];
477 /* mboxes this ring signals to */
478 i915_reg_t signal[GEN6_NUM_SEMAPHORES];
480 u64 signal_ggtt[I915_NUM_ENGINES];
484 int (*sync_to)(struct drm_i915_gem_request *req,
485 struct drm_i915_gem_request *signal);
486 u32 *(*signal)(struct drm_i915_gem_request *req, u32 *cs);
489 struct intel_engine_execlists execlists;
491 /* Contexts are pinned whilst they are active on the GPU. The last
492 * context executed remains active whilst the GPU is idle - the
493 * switch away and write to the context object only occurs on the
494 * next execution. Contexts are only unpinned on retirement of the
495 * following request ensuring that we can always write to the object
496 * on the context switch even after idling. Across suspend, we switch
497 * to the kernel context and trash it as the save may not happen
498 * before the hardware is powered down.
500 struct i915_gem_context *last_retired_context;
502 /* We track the current MI_SET_CONTEXT in order to eliminate
503 * redudant context switches. This presumes that requests are not
504 * reordered! Or when they are the tracking is updated along with
505 * the emission of individual requests into the legacy command
508 struct i915_gem_context *legacy_active_context;
510 /* status_notifier: list of callbacks for context-switch changes */
511 struct atomic_notifier_head context_status_notifier;
513 struct intel_engine_hangcheck hangcheck;
515 bool needs_cmd_parser;
518 * Table of commands the command parser needs to know about
521 DECLARE_HASHTABLE(cmd_hash, I915_CMD_HASH_ORDER);
524 * Table of registers allowed in commands that read/write registers.
526 const struct drm_i915_reg_table *reg_tables;
530 * Returns the bitmask for the length field of the specified command.
531 * Return 0 for an unrecognized/invalid command.
533 * If the command parser finds an entry for a command in the engine's
534 * cmd_tables, it gets the command's length based on the table entry.
535 * If not, it calls this function to determine the per-engine length
536 * field encoding for the command (i.e. different opcode ranges use
537 * certain bits to encode the command length in the header).
539 u32 (*get_cmd_length_mask)(u32 cmd_header);
543 execlists_set_active(struct intel_engine_execlists *execlists,
546 __set_bit(bit, (unsigned long *)&execlists->active);
550 execlists_clear_active(struct intel_engine_execlists *execlists,
553 __clear_bit(bit, (unsigned long *)&execlists->active);
557 execlists_is_active(const struct intel_engine_execlists *execlists,
560 return test_bit(bit, (unsigned long *)&execlists->active);
564 execlists_cancel_port_requests(struct intel_engine_execlists * const execlists);
567 execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists);
569 static inline unsigned int
570 execlists_num_ports(const struct intel_engine_execlists * const execlists)
572 return execlists->port_mask + 1;
576 execlists_port_complete(struct intel_engine_execlists * const execlists,
577 struct execlist_port * const port)
579 const unsigned int m = execlists->port_mask;
581 GEM_BUG_ON(port_index(port, execlists) != 0);
582 GEM_BUG_ON(!execlists_is_active(execlists, EXECLISTS_ACTIVE_USER));
584 memmove(port, port + 1, m * sizeof(struct execlist_port));
585 memset(port + m, 0, sizeof(struct execlist_port));
588 static inline unsigned int
589 intel_engine_flag(const struct intel_engine_cs *engine)
591 return BIT(engine->id);
595 intel_read_status_page(struct intel_engine_cs *engine, int reg)
597 /* Ensure that the compiler doesn't optimize away the load. */
598 return READ_ONCE(engine->status_page.page_addr[reg]);
602 intel_write_status_page(struct intel_engine_cs *engine, int reg, u32 value)
604 /* Writing into the status page should be done sparingly. Since
605 * we do when we are uncertain of the device state, we take a bit
606 * of extra paranoia to try and ensure that the HWS takes the value
607 * we give and that it doesn't end up trapped inside the CPU!
609 if (static_cpu_has(X86_FEATURE_CLFLUSH)) {
611 clflush(&engine->status_page.page_addr[reg]);
612 engine->status_page.page_addr[reg] = value;
613 clflush(&engine->status_page.page_addr[reg]);
616 WRITE_ONCE(engine->status_page.page_addr[reg], value);
621 * Reads a dword out of the status page, which is written to from the command
622 * queue by automatic updates, MI_REPORT_HEAD, MI_STORE_DATA_INDEX, or
625 * The following dwords have a reserved meaning:
626 * 0x00: ISR copy, updated when an ISR bit not set in the HWSTAM changes.
627 * 0x04: ring 0 head pointer
628 * 0x05: ring 1 head pointer (915-class)
629 * 0x06: ring 2 head pointer (915-class)
630 * 0x10-0x1b: Context status DWords (GM45)
631 * 0x1f: Last written status offset. (GM45)
632 * 0x20-0x2f: Reserved (Gen6+)
634 * The area from dword 0x30 to 0x3ff is available for driver usage.
636 #define I915_GEM_HWS_INDEX 0x30
637 #define I915_GEM_HWS_INDEX_ADDR (I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT)
638 #define I915_GEM_HWS_PREEMPT_INDEX 0x32
639 #define I915_GEM_HWS_PREEMPT_ADDR (I915_GEM_HWS_PREEMPT_INDEX << MI_STORE_DWORD_INDEX_SHIFT)
640 #define I915_GEM_HWS_SCRATCH_INDEX 0x40
641 #define I915_GEM_HWS_SCRATCH_ADDR (I915_GEM_HWS_SCRATCH_INDEX << MI_STORE_DWORD_INDEX_SHIFT)
643 #define I915_HWS_CSB_BUF0_INDEX 0x10
644 #define I915_HWS_CSB_WRITE_INDEX 0x1f
645 #define CNL_HWS_CSB_WRITE_INDEX 0x2f
648 intel_engine_create_ring(struct intel_engine_cs *engine, int size);
649 int intel_ring_pin(struct intel_ring *ring,
650 struct drm_i915_private *i915,
651 unsigned int offset_bias);
652 void intel_ring_reset(struct intel_ring *ring, u32 tail);
653 unsigned int intel_ring_update_space(struct intel_ring *ring);
654 void intel_ring_unpin(struct intel_ring *ring);
655 void intel_ring_free(struct intel_ring *ring);
657 void intel_engine_stop(struct intel_engine_cs *engine);
658 void intel_engine_cleanup(struct intel_engine_cs *engine);
660 void intel_legacy_submission_resume(struct drm_i915_private *dev_priv);
662 int __must_check intel_ring_cacheline_align(struct drm_i915_gem_request *req);
664 int intel_ring_wait_for_space(struct intel_ring *ring, unsigned int bytes);
665 u32 __must_check *intel_ring_begin(struct drm_i915_gem_request *req,
669 intel_ring_advance(struct drm_i915_gem_request *req, u32 *cs)
673 * This serves as a placeholder in the code so that the reader
674 * can compare against the preceding intel_ring_begin() and
675 * check that the number of dwords emitted matches the space
676 * reserved for the command packet (i.e. the value passed to
677 * intel_ring_begin()).
679 GEM_BUG_ON((req->ring->vaddr + req->ring->emit) != cs);
683 intel_ring_wrap(const struct intel_ring *ring, u32 pos)
685 return pos & (ring->size - 1);
689 intel_ring_offset(const struct drm_i915_gem_request *req, void *addr)
691 /* Don't write ring->size (equivalent to 0) as that hangs some GPUs. */
692 u32 offset = addr - req->ring->vaddr;
693 GEM_BUG_ON(offset > req->ring->size);
694 return intel_ring_wrap(req->ring, offset);
698 assert_ring_tail_valid(const struct intel_ring *ring, unsigned int tail)
700 /* We could combine these into a single tail operation, but keeping
701 * them as seperate tests will help identify the cause should one
704 GEM_BUG_ON(!IS_ALIGNED(tail, 8));
705 GEM_BUG_ON(tail >= ring->size);
709 * Gen2 BSpec "1. Programming Environment" / 1.4.4.6
710 * Gen3 BSpec "1c Memory Interface Functions" / 2.3.4.5
711 * Gen4+ BSpec "1c Memory Interface and Command Stream" / 5.3.4.5
712 * "If the Ring Buffer Head Pointer and the Tail Pointer are on the
713 * same cacheline, the Head Pointer must not be greater than the Tail
716 * We use ring->head as the last known location of the actual RING_HEAD,
717 * it may have advanced but in the worst case it is equally the same
718 * as ring->head and so we should never program RING_TAIL to advance
719 * into the same cacheline as ring->head.
721 #define cacheline(a) round_down(a, CACHELINE_BYTES)
722 GEM_BUG_ON(cacheline(tail) == cacheline(ring->head) &&
727 static inline unsigned int
728 intel_ring_set_tail(struct intel_ring *ring, unsigned int tail)
730 /* Whilst writes to the tail are strictly order, there is no
731 * serialisation between readers and the writers. The tail may be
732 * read by i915_gem_request_retire() just as it is being updated
733 * by execlists, as although the breadcrumb is complete, the context
734 * switch hasn't been seen.
736 assert_ring_tail_valid(ring, tail);
741 void intel_engine_init_global_seqno(struct intel_engine_cs *engine, u32 seqno);
743 void intel_engine_setup_common(struct intel_engine_cs *engine);
744 int intel_engine_init_common(struct intel_engine_cs *engine);
745 int intel_engine_create_scratch(struct intel_engine_cs *engine, int size);
746 void intel_engine_cleanup_common(struct intel_engine_cs *engine);
748 int intel_init_render_ring_buffer(struct intel_engine_cs *engine);
749 int intel_init_bsd_ring_buffer(struct intel_engine_cs *engine);
750 int intel_init_blt_ring_buffer(struct intel_engine_cs *engine);
751 int intel_init_vebox_ring_buffer(struct intel_engine_cs *engine);
753 u64 intel_engine_get_active_head(struct intel_engine_cs *engine);
754 u64 intel_engine_get_last_batch_head(struct intel_engine_cs *engine);
756 static inline u32 intel_engine_get_seqno(struct intel_engine_cs *engine)
758 return intel_read_status_page(engine, I915_GEM_HWS_INDEX);
761 static inline u32 intel_engine_last_submit(struct intel_engine_cs *engine)
763 /* We are only peeking at the tail of the submit queue (and not the
764 * queue itself) in order to gain a hint as to the current active
765 * state of the engine. Callers are not expected to be taking
766 * engine->timeline->lock, nor are they expected to be concerned
767 * wtih serialising this hint with anything, so document it as
768 * a hint and nothing more.
770 return READ_ONCE(engine->timeline->seqno);
773 int init_workarounds_ring(struct intel_engine_cs *engine);
774 int intel_ring_workarounds_emit(struct drm_i915_gem_request *req);
776 void intel_engine_get_instdone(struct intel_engine_cs *engine,
777 struct intel_instdone *instdone);
780 * Arbitrary size for largest possible 'add request' sequence. The code paths
781 * are complex and variable. Empirical measurement shows that the worst case
782 * is BDW at 192 bytes (6 + 6 + 36 dwords), then ILK at 136 bytes. However,
783 * we need to allocate double the largest single packet within that emission
784 * to account for tail wraparound (so 6 + 6 + 72 dwords for BDW).
786 #define MIN_SPACE_FOR_ADD_REQUEST 336
788 static inline u32 intel_hws_seqno_address(struct intel_engine_cs *engine)
790 return engine->status_page.ggtt_offset + I915_GEM_HWS_INDEX_ADDR;
793 static inline u32 intel_hws_preempt_done_address(struct intel_engine_cs *engine)
795 return engine->status_page.ggtt_offset + I915_GEM_HWS_PREEMPT_ADDR;
798 /* intel_breadcrumbs.c -- user interrupt bottom-half for waiters */
799 int intel_engine_init_breadcrumbs(struct intel_engine_cs *engine);
801 static inline void intel_wait_init(struct intel_wait *wait,
802 struct drm_i915_gem_request *rq)
808 static inline void intel_wait_init_for_seqno(struct intel_wait *wait, u32 seqno)
814 static inline bool intel_wait_has_seqno(const struct intel_wait *wait)
820 intel_wait_update_seqno(struct intel_wait *wait, u32 seqno)
823 return intel_wait_has_seqno(wait);
827 intel_wait_update_request(struct intel_wait *wait,
828 const struct drm_i915_gem_request *rq)
830 return intel_wait_update_seqno(wait, i915_gem_request_global_seqno(rq));
834 intel_wait_check_seqno(const struct intel_wait *wait, u32 seqno)
836 return wait->seqno == seqno;
840 intel_wait_check_request(const struct intel_wait *wait,
841 const struct drm_i915_gem_request *rq)
843 return intel_wait_check_seqno(wait, i915_gem_request_global_seqno(rq));
846 static inline bool intel_wait_complete(const struct intel_wait *wait)
848 return RB_EMPTY_NODE(&wait->node);
851 bool intel_engine_add_wait(struct intel_engine_cs *engine,
852 struct intel_wait *wait);
853 void intel_engine_remove_wait(struct intel_engine_cs *engine,
854 struct intel_wait *wait);
855 void intel_engine_enable_signaling(struct drm_i915_gem_request *request,
857 void intel_engine_cancel_signaling(struct drm_i915_gem_request *request);
859 static inline bool intel_engine_has_waiter(const struct intel_engine_cs *engine)
861 return READ_ONCE(engine->breadcrumbs.irq_wait);
864 unsigned int intel_engine_wakeup(struct intel_engine_cs *engine);
865 #define ENGINE_WAKEUP_WAITER BIT(0)
866 #define ENGINE_WAKEUP_ASLEEP BIT(1)
868 void intel_engine_pin_breadcrumbs_irq(struct intel_engine_cs *engine);
869 void intel_engine_unpin_breadcrumbs_irq(struct intel_engine_cs *engine);
871 void __intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine);
872 void intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine);
874 void intel_engine_reset_breadcrumbs(struct intel_engine_cs *engine);
875 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine);
876 bool intel_breadcrumbs_busy(struct intel_engine_cs *engine);
878 static inline u32 *gen8_emit_pipe_control(u32 *batch, u32 flags, u32 offset)
880 memset(batch, 0, 6 * sizeof(u32));
882 batch[0] = GFX_OP_PIPE_CONTROL(6);
890 gen8_emit_ggtt_write_rcs(u32 *cs, u32 value, u32 gtt_offset)
892 /* We're using qword write, offset should be aligned to 8 bytes. */
893 GEM_BUG_ON(!IS_ALIGNED(gtt_offset, 8));
895 /* w/a for post sync ops following a GPGPU operation we
896 * need a prior CS_STALL, which is emitted by the flush
897 * following the batch.
899 *cs++ = GFX_OP_PIPE_CONTROL(6);
900 *cs++ = PIPE_CONTROL_GLOBAL_GTT_IVB | PIPE_CONTROL_CS_STALL |
901 PIPE_CONTROL_QW_WRITE;
905 /* We're thrashing one dword of HWS. */
912 gen8_emit_ggtt_write(u32 *cs, u32 value, u32 gtt_offset)
914 /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */
915 GEM_BUG_ON(gtt_offset & (1 << 5));
916 /* Offset should be aligned to 8 bytes for both (QW/DW) write types */
917 GEM_BUG_ON(!IS_ALIGNED(gtt_offset, 8));
919 *cs++ = (MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW;
920 *cs++ = gtt_offset | MI_FLUSH_DW_USE_GTT;
927 bool intel_engine_is_idle(struct intel_engine_cs *engine);
928 bool intel_engines_are_idle(struct drm_i915_private *dev_priv);
930 bool intel_engine_has_kernel_context(const struct intel_engine_cs *engine);
932 void intel_engines_park(struct drm_i915_private *i915);
933 void intel_engines_unpark(struct drm_i915_private *i915);
935 void intel_engines_reset_default_submission(struct drm_i915_private *i915);
936 unsigned int intel_engines_has_context_isolation(struct drm_i915_private *i915);
938 bool intel_engine_can_store_dword(struct intel_engine_cs *engine);
940 void intel_engine_dump(struct intel_engine_cs *engine, struct drm_printer *p);
942 #endif /* _INTEL_RINGBUFFER_H_ */