2 * Copyright © 2014 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Ben Widawsky <ben@bwidawsk.net>
25 * Michel Thierry <michel.thierry@intel.com>
26 * Thomas Daniel <thomas.daniel@intel.com>
27 * Oscar Mateo <oscar.mateo@intel.com>
32 * DOC: Logical Rings, Logical Ring Contexts and Execlists
35 * GEN8 brings an expansion of the HW contexts: "Logical Ring Contexts".
36 * These expanded contexts enable a number of new abilities, especially
37 * "Execlists" (also implemented in this file).
39 * One of the main differences with the legacy HW contexts is that logical
40 * ring contexts incorporate many more things to the context's state, like
41 * PDPs or ringbuffer control registers:
43 * The reason why PDPs are included in the context is straightforward: as
44 * PPGTTs (per-process GTTs) are actually per-context, having the PDPs
45 * contained there mean you don't need to do a ppgtt->switch_mm yourself,
46 * instead, the GPU will do it for you on the context switch.
48 * But, what about the ringbuffer control registers (head, tail, etc..)?
49 * shouldn't we just need a set of those per engine command streamer? This is
50 * where the name "Logical Rings" starts to make sense: by virtualizing the
51 * rings, the engine cs shifts to a new "ring buffer" with every context
52 * switch. When you want to submit a workload to the GPU you: A) choose your
53 * context, B) find its appropriate virtualized ring, C) write commands to it
54 * and then, finally, D) tell the GPU to switch to that context.
56 * Instead of the legacy MI_SET_CONTEXT, the way you tell the GPU to switch
57 * to a contexts is via a context execution list, ergo "Execlists".
60 * Regarding the creation of contexts, we have:
62 * - One global default context.
63 * - One local default context for each opened fd.
64 * - One local extra context for each context create ioctl call.
66 * Now that ringbuffers belong per-context (and not per-engine, like before)
67 * and that contexts are uniquely tied to a given engine (and not reusable,
68 * like before) we need:
70 * - One ringbuffer per-engine inside each context.
71 * - One backing object per-engine inside each context.
73 * The global default context starts its life with these new objects fully
74 * allocated and populated. The local default context for each opened fd is
75 * more complex, because we don't know at creation time which engine is going
76 * to use them. To handle this, we have implemented a deferred creation of LR
79 * The local context starts its life as a hollow or blank holder, that only
80 * gets populated for a given engine once we receive an execbuffer. If later
81 * on we receive another execbuffer ioctl for the same context but a different
82 * engine, we allocate/populate a new ringbuffer and context backing object and
85 * Finally, regarding local contexts created using the ioctl call: as they are
86 * only allowed with the render ring, we can allocate & populate them right
87 * away (no need to defer anything, at least for now).
89 * Execlists implementation:
90 * Execlists are the new method by which, on gen8+ hardware, workloads are
91 * submitted for execution (as opposed to the legacy, ringbuffer-based, method).
92 * This method works as follows:
94 * When a request is committed, its commands (the BB start and any leading or
95 * trailing commands, like the seqno breadcrumbs) are placed in the ringbuffer
96 * for the appropriate context. The tail pointer in the hardware context is not
97 * updated at this time, but instead, kept by the driver in the ringbuffer
98 * structure. A structure representing this request is added to a request queue
99 * for the appropriate engine: this structure contains a copy of the context's
100 * tail after the request was written to the ring buffer and a pointer to the
103 * If the engine's request queue was empty before the request was added, the
104 * queue is processed immediately. Otherwise the queue will be processed during
105 * a context switch interrupt. In any case, elements on the queue will get sent
106 * (in pairs) to the GPU's ExecLists Submit Port (ELSP, for short) with a
107 * globally unique 20-bits submission ID.
109 * When execution of a request completes, the GPU updates the context status
110 * buffer with a context complete event and generates a context switch interrupt.
111 * During the interrupt handling, the driver examines the events in the buffer:
112 * for each context complete event, if the announced ID matches that on the head
113 * of the request queue, then that request is retired and removed from the queue.
115 * After processing, if any requests were retired and the queue is not empty
116 * then a new execution list can be submitted. The two requests at the front of
117 * the queue are next to be submitted but since a context may not occur twice in
118 * an execution list, if subsequent requests have the same ID as the first then
119 * the two requests must be combined. This is done simply by discarding requests
120 * at the head of the queue until either only one requests is left (in which case
121 * we use a NULL second context) or the first two requests have unique IDs.
123 * By always executing the first two requests in the queue the driver ensures
124 * that the GPU is kept as busy as possible. In the case where a single context
125 * completes but a second context is still executing, the request for this second
126 * context will be at the head of the queue when we remove the first one. This
127 * request will then be resubmitted along with a new request for a different context,
128 * which will cause the hardware to continue executing the second request and queue
129 * the new request (the GPU detects the condition of a context getting preempted
130 * with the same context and optimizes the context switch flow by not doing
131 * preemption, but just sampling the new tail pointer).
134 #include <linux/interrupt.h>
136 #include <drm/drmP.h>
137 #include <drm/i915_drm.h>
138 #include "i915_drv.h"
139 #include "i915_gem_render_state.h"
140 #include "i915_vgpu.h"
141 #include "intel_lrc_reg.h"
142 #include "intel_mocs.h"
143 #include "intel_workarounds.h"
145 #define RING_EXECLIST_QFULL (1 << 0x2)
146 #define RING_EXECLIST1_VALID (1 << 0x3)
147 #define RING_EXECLIST0_VALID (1 << 0x4)
148 #define RING_EXECLIST_ACTIVE_STATUS (3 << 0xE)
149 #define RING_EXECLIST1_ACTIVE (1 << 0x11)
150 #define RING_EXECLIST0_ACTIVE (1 << 0x12)
152 #define GEN8_CTX_STATUS_IDLE_ACTIVE (1 << 0)
153 #define GEN8_CTX_STATUS_PREEMPTED (1 << 1)
154 #define GEN8_CTX_STATUS_ELEMENT_SWITCH (1 << 2)
155 #define GEN8_CTX_STATUS_ACTIVE_IDLE (1 << 3)
156 #define GEN8_CTX_STATUS_COMPLETE (1 << 4)
157 #define GEN8_CTX_STATUS_LITE_RESTORE (1 << 15)
159 #define GEN8_CTX_STATUS_COMPLETED_MASK \
160 (GEN8_CTX_STATUS_COMPLETE | GEN8_CTX_STATUS_PREEMPTED)
162 /* Typical size of the average request (2 pipecontrols and a MI_BB) */
163 #define EXECLISTS_REQUEST_SIZE 64 /* bytes */
164 #define WA_TAIL_DWORDS 2
165 #define WA_TAIL_BYTES (sizeof(u32) * WA_TAIL_DWORDS)
167 static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
168 struct intel_engine_cs *engine,
169 struct intel_context *ce);
170 static void execlists_init_reg_state(u32 *reg_state,
171 struct i915_gem_context *ctx,
172 struct intel_engine_cs *engine,
173 struct intel_ring *ring);
175 static inline struct i915_priolist *to_priolist(struct rb_node *rb)
177 return rb_entry(rb, struct i915_priolist, node);
180 static inline int rq_prio(const struct i915_request *rq)
182 return rq->sched.attr.priority;
185 static inline bool need_preempt(const struct intel_engine_cs *engine,
186 const struct i915_request *last,
189 return (intel_engine_has_preemption(engine) &&
190 __execlists_need_preempt(prio, rq_prio(last)) &&
191 !i915_request_completed(last));
195 * The context descriptor encodes various attributes of a context,
196 * including its GTT address and some flags. Because it's fairly
197 * expensive to calculate, we'll just do it once and cache the result,
198 * which remains valid until the context is unpinned.
200 * This is what a descriptor looks like, from LSB to MSB::
202 * bits 0-11: flags, GEN8_CTX_* (cached in ctx->desc_template)
203 * bits 12-31: LRCA, GTT address of (the HWSP of) this context
204 * bits 32-52: ctx ID, a globally unique tag (highest bit used by GuC)
205 * bits 53-54: mbz, reserved for use by hardware
206 * bits 55-63: group ID, currently unused and set to 0
208 * Starting from Gen11, the upper dword of the descriptor has a new format:
210 * bits 32-36: reserved
211 * bits 37-47: SW context ID
212 * bits 48:53: engine instance
213 * bit 54: mbz, reserved for use by hardware
214 * bits 55-60: SW counter
215 * bits 61-63: engine class
217 * engine info, SW context ID and SW counter need to form a unique number
218 * (Context ID) per lrc.
221 intel_lr_context_descriptor_update(struct i915_gem_context *ctx,
222 struct intel_engine_cs *engine,
223 struct intel_context *ce)
227 BUILD_BUG_ON(MAX_CONTEXT_HW_ID > (BIT(GEN8_CTX_ID_WIDTH)));
228 BUILD_BUG_ON(GEN11_MAX_CONTEXT_HW_ID > (BIT(GEN11_SW_CTX_ID_WIDTH)));
230 desc = ctx->desc_template; /* bits 0-11 */
231 GEM_BUG_ON(desc & GENMASK_ULL(63, 12));
233 desc |= i915_ggtt_offset(ce->state) + LRC_HEADER_PAGES * PAGE_SIZE;
235 GEM_BUG_ON(desc & GENMASK_ULL(63, 32));
238 * The following 32bits are copied into the OA reports (dword 2).
239 * Consider updating oa_get_render_ctx_id in i915_perf.c when changing
242 if (INTEL_GEN(ctx->i915) >= 11) {
243 GEM_BUG_ON(ctx->hw_id >= BIT(GEN11_SW_CTX_ID_WIDTH));
244 desc |= (u64)ctx->hw_id << GEN11_SW_CTX_ID_SHIFT;
247 desc |= (u64)engine->instance << GEN11_ENGINE_INSTANCE_SHIFT;
250 /* TODO: decide what to do with SW counter (bits 55-60) */
252 desc |= (u64)engine->class << GEN11_ENGINE_CLASS_SHIFT;
255 GEM_BUG_ON(ctx->hw_id >= BIT(GEN8_CTX_ID_WIDTH));
256 desc |= (u64)ctx->hw_id << GEN8_CTX_ID_SHIFT; /* bits 32-52 */
262 static void unwind_wa_tail(struct i915_request *rq)
264 rq->tail = intel_ring_wrap(rq->ring, rq->wa_tail - WA_TAIL_BYTES);
265 assert_ring_tail_valid(rq->ring, rq->tail);
268 static void __unwind_incomplete_requests(struct intel_engine_cs *engine)
270 struct i915_request *rq, *rn, *active = NULL;
271 struct list_head *uninitialized_var(pl);
272 int prio = I915_PRIORITY_INVALID | I915_PRIORITY_NEWCLIENT;
274 lockdep_assert_held(&engine->timeline.lock);
276 list_for_each_entry_safe_reverse(rq, rn,
277 &engine->timeline.requests,
279 if (i915_request_completed(rq))
282 __i915_request_unsubmit(rq);
285 GEM_BUG_ON(rq->hw_context->active);
287 GEM_BUG_ON(rq_prio(rq) == I915_PRIORITY_INVALID);
288 if (rq_prio(rq) != prio) {
290 pl = i915_sched_lookup_priolist(engine, prio);
292 GEM_BUG_ON(RB_EMPTY_ROOT(&engine->execlists.queue.rb_root));
294 list_add(&rq->sched.link, pl);
300 * The active request is now effectively the start of a new client
301 * stream, so give it the equivalent small priority bump to prevent
302 * it being gazumped a second time by another peer.
304 if (!(prio & I915_PRIORITY_NEWCLIENT)) {
305 prio |= I915_PRIORITY_NEWCLIENT;
306 list_move_tail(&active->sched.link,
307 i915_sched_lookup_priolist(engine, prio));
312 execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists)
314 struct intel_engine_cs *engine =
315 container_of(execlists, typeof(*engine), execlists);
317 __unwind_incomplete_requests(engine);
321 execlists_context_status_change(struct i915_request *rq, unsigned long status)
324 * Only used when GVT-g is enabled now. When GVT-g is disabled,
325 * The compiler should eliminate this function as dead-code.
327 if (!IS_ENABLED(CONFIG_DRM_I915_GVT))
330 atomic_notifier_call_chain(&rq->engine->context_status_notifier,
335 execlists_user_begin(struct intel_engine_execlists *execlists,
336 const struct execlist_port *port)
338 execlists_set_active_once(execlists, EXECLISTS_ACTIVE_USER);
342 execlists_user_end(struct intel_engine_execlists *execlists)
344 execlists_clear_active(execlists, EXECLISTS_ACTIVE_USER);
348 execlists_context_schedule_in(struct i915_request *rq)
350 GEM_BUG_ON(rq->hw_context->active);
352 execlists_context_status_change(rq, INTEL_CONTEXT_SCHEDULE_IN);
353 intel_engine_context_in(rq->engine);
354 rq->hw_context->active = rq->engine;
358 execlists_context_schedule_out(struct i915_request *rq, unsigned long status)
360 rq->hw_context->active = NULL;
361 intel_engine_context_out(rq->engine);
362 execlists_context_status_change(rq, status);
363 trace_i915_request_out(rq);
366 static u64 execlists_update_context(struct i915_request *rq)
368 struct intel_context *ce = rq->hw_context;
370 ce->lrc_reg_state[CTX_RING_TAIL + 1] =
371 intel_ring_set_tail(rq->ring, rq->tail);
374 * Make sure the context image is complete before we submit it to HW.
376 * Ostensibly, writes (including the WCB) should be flushed prior to
377 * an uncached write such as our mmio register access, the empirical
378 * evidence (esp. on Braswell) suggests that the WC write into memory
379 * may not be visible to the HW prior to the completion of the UC
380 * register write and that we may begin execution from the context
381 * before its image is complete leading to invalid PD chasing.
383 * Furthermore, Braswell, at least, wants a full mb to be sure that
384 * the writes are coherent in memory (visible to the GPU) prior to
385 * execution, and not just visible to other CPUs (as is the result of
392 static inline void write_desc(struct intel_engine_execlists *execlists, u64 desc, u32 port)
394 if (execlists->ctrl_reg) {
395 writel(lower_32_bits(desc), execlists->submit_reg + port * 2);
396 writel(upper_32_bits(desc), execlists->submit_reg + port * 2 + 1);
398 writel(upper_32_bits(desc), execlists->submit_reg);
399 writel(lower_32_bits(desc), execlists->submit_reg);
403 static void execlists_submit_ports(struct intel_engine_cs *engine)
405 struct intel_engine_execlists *execlists = &engine->execlists;
406 struct execlist_port *port = execlists->port;
410 * We can skip acquiring intel_runtime_pm_get() here as it was taken
411 * on our behalf by the request (see i915_gem_mark_busy()) and it will
412 * not be relinquished until the device is idle (see
413 * i915_gem_idle_work_handler()). As a precaution, we make sure
414 * that all ELSP are drained i.e. we have processed the CSB,
415 * before allowing ourselves to idle and calling intel_runtime_pm_put().
417 GEM_BUG_ON(!engine->i915->gt.awake);
420 * ELSQ note: the submit queue is not cleared after being submitted
421 * to the HW so we need to make sure we always clean it up. This is
422 * currently ensured by the fact that we always write the same number
423 * of elsq entries, keep this in mind before changing the loop below.
425 for (n = execlists_num_ports(execlists); n--; ) {
426 struct i915_request *rq;
430 rq = port_unpack(&port[n], &count);
432 GEM_BUG_ON(count > !n);
434 execlists_context_schedule_in(rq);
435 port_set(&port[n], port_pack(rq, count));
436 desc = execlists_update_context(rq);
437 GEM_DEBUG_EXEC(port[n].context_id = upper_32_bits(desc));
439 GEM_TRACE("%s in[%d]: ctx=%d.%d, global=%d (fence %llx:%d) (current %d), prio=%d\n",
441 port[n].context_id, count,
443 rq->fence.context, rq->fence.seqno,
444 intel_engine_get_seqno(engine),
451 write_desc(execlists, desc, n);
454 /* we need to manually load the submit queue */
455 if (execlists->ctrl_reg)
456 writel(EL_CTRL_LOAD, execlists->ctrl_reg);
458 execlists_clear_active(execlists, EXECLISTS_ACTIVE_HWACK);
461 static bool ctx_single_port_submission(const struct intel_context *ce)
463 return (IS_ENABLED(CONFIG_DRM_I915_GVT) &&
464 i915_gem_context_force_single_submission(ce->gem_context));
467 static bool can_merge_ctx(const struct intel_context *prev,
468 const struct intel_context *next)
473 if (ctx_single_port_submission(prev))
479 static void port_assign(struct execlist_port *port, struct i915_request *rq)
481 GEM_BUG_ON(rq == port_request(port));
483 if (port_isset(port))
484 i915_request_put(port_request(port));
486 port_set(port, port_pack(i915_request_get(rq), port_count(port)));
489 static void inject_preempt_context(struct intel_engine_cs *engine)
491 struct intel_engine_execlists *execlists = &engine->execlists;
492 struct intel_context *ce =
493 to_intel_context(engine->i915->preempt_context, engine);
496 GEM_BUG_ON(execlists->preempt_complete_status !=
497 upper_32_bits(ce->lrc_desc));
500 * Switch to our empty preempt context so
501 * the state of the GPU is known (idle).
503 GEM_TRACE("%s\n", engine->name);
504 for (n = execlists_num_ports(execlists); --n; )
505 write_desc(execlists, 0, n);
507 write_desc(execlists, ce->lrc_desc, n);
509 /* we need to manually load the submit queue */
510 if (execlists->ctrl_reg)
511 writel(EL_CTRL_LOAD, execlists->ctrl_reg);
513 execlists_clear_active(execlists, EXECLISTS_ACTIVE_HWACK);
514 execlists_set_active(execlists, EXECLISTS_ACTIVE_PREEMPT);
517 static void complete_preempt_context(struct intel_engine_execlists *execlists)
519 GEM_BUG_ON(!execlists_is_active(execlists, EXECLISTS_ACTIVE_PREEMPT));
521 if (inject_preempt_hang(execlists))
524 execlists_cancel_port_requests(execlists);
525 __unwind_incomplete_requests(container_of(execlists,
526 struct intel_engine_cs,
530 static void execlists_dequeue(struct intel_engine_cs *engine)
532 struct intel_engine_execlists * const execlists = &engine->execlists;
533 struct execlist_port *port = execlists->port;
534 const struct execlist_port * const last_port =
535 &execlists->port[execlists->port_mask];
536 struct i915_request *last = port_request(port);
541 * Hardware submission is through 2 ports. Conceptually each port
542 * has a (RING_START, RING_HEAD, RING_TAIL) tuple. RING_START is
543 * static for a context, and unique to each, so we only execute
544 * requests belonging to a single context from each ring. RING_HEAD
545 * is maintained by the CS in the context image, it marks the place
546 * where it got up to last time, and through RING_TAIL we tell the CS
547 * where we want to execute up to this time.
549 * In this list the requests are in order of execution. Consecutive
550 * requests from the same context are adjacent in the ringbuffer. We
551 * can combine these requests into a single RING_TAIL update:
553 * RING_HEAD...req1...req2
555 * since to execute req2 the CS must first execute req1.
557 * Our goal then is to point each port to the end of a consecutive
558 * sequence of requests as being the most optimal (fewest wake ups
559 * and context switches) submission.
564 * Don't resubmit or switch until all outstanding
565 * preemptions (lite-restore) are seen. Then we
566 * know the next preemption status we see corresponds
567 * to this ELSP update.
569 GEM_BUG_ON(!execlists_is_active(execlists,
570 EXECLISTS_ACTIVE_USER));
571 GEM_BUG_ON(!port_count(&port[0]));
574 * If we write to ELSP a second time before the HW has had
575 * a chance to respond to the previous write, we can confuse
576 * the HW and hit "undefined behaviour". After writing to ELSP,
577 * we must then wait until we see a context-switch event from
578 * the HW to indicate that it has had a chance to respond.
580 if (!execlists_is_active(execlists, EXECLISTS_ACTIVE_HWACK))
583 if (need_preempt(engine, last, execlists->queue_priority)) {
584 inject_preempt_context(engine);
589 * In theory, we could coalesce more requests onto
590 * the second port (the first port is active, with
591 * no preemptions pending). However, that means we
592 * then have to deal with the possible lite-restore
593 * of the second port (as we submit the ELSP, there
594 * may be a context-switch) but also we may complete
595 * the resubmission before the context-switch. Ergo,
596 * coalescing onto the second port will cause a
597 * preemption event, but we cannot predict whether
598 * that will affect port[0] or port[1].
600 * If the second port is already active, we can wait
601 * until the next context-switch before contemplating
602 * new requests. The GPU will be busy and we should be
603 * able to resubmit the new ELSP before it idles,
604 * avoiding pipeline bubbles (momentary pauses where
605 * the driver is unable to keep up the supply of new
606 * work). However, we have to double check that the
607 * priorities of the ports haven't been switch.
609 if (port_count(&port[1]))
613 * WaIdleLiteRestore:bdw,skl
614 * Apply the wa NOOPs to prevent
615 * ring:HEAD == rq:TAIL as we resubmit the
616 * request. See gen8_emit_breadcrumb() for
617 * where we prepare the padding after the
618 * end of the request.
620 last->tail = last->wa_tail;
623 while ((rb = rb_first_cached(&execlists->queue))) {
624 struct i915_priolist *p = to_priolist(rb);
625 struct i915_request *rq, *rn;
628 priolist_for_each_request_consume(rq, rn, p, i) {
630 * Can we combine this request with the current port?
631 * It has to be the same context/ringbuffer and not
632 * have any exceptions (e.g. GVT saying never to
635 * If we can combine the requests, we can execute both
636 * by updating the RING_TAIL to point to the end of the
637 * second request, and so we never need to tell the
638 * hardware about the first.
641 !can_merge_ctx(rq->hw_context, last->hw_context)) {
643 * If we are on the second port and cannot
644 * combine this request with the last, then we
647 if (port == last_port)
651 * If GVT overrides us we only ever submit
652 * port[0], leaving port[1] empty. Note that we
653 * also have to be careful that we don't queue
654 * the same context (even though a different
655 * request) to the second port.
657 if (ctx_single_port_submission(last->hw_context) ||
658 ctx_single_port_submission(rq->hw_context))
661 GEM_BUG_ON(last->hw_context == rq->hw_context);
664 port_assign(port, last);
667 GEM_BUG_ON(port_isset(port));
670 list_del_init(&rq->sched.link);
672 __i915_request_submit(rq);
673 trace_i915_request_in(rq, port_index(port, execlists));
679 rb_erase_cached(&p->node, &execlists->queue);
680 if (p->priority != I915_PRIORITY_NORMAL)
681 kmem_cache_free(engine->i915->priorities, p);
686 * Here be a bit of magic! Or sleight-of-hand, whichever you prefer.
688 * We choose queue_priority such that if we add a request of greater
689 * priority than this, we kick the submission tasklet to decide on
690 * the right order of submitting the requests to hardware. We must
691 * also be prepared to reorder requests as they are in-flight on the
692 * HW. We derive the queue_priority then as the first "hole" in
693 * the HW submission ports and if there are no available slots,
694 * the priority of the lowest executing request, i.e. last.
696 * When we do receive a higher priority request ready to run from the
697 * user, see queue_request(), the queue_priority is bumped to that
698 * request triggering preemption on the next dequeue (or subsequent
699 * interrupt for secondary ports).
701 execlists->queue_priority =
702 port != execlists->port ? rq_prio(last) : INT_MIN;
705 port_assign(port, last);
706 execlists_submit_ports(engine);
709 /* We must always keep the beast fed if we have work piled up */
710 GEM_BUG_ON(rb_first_cached(&execlists->queue) &&
711 !port_isset(execlists->port));
713 /* Re-evaluate the executing context setup after each preemptive kick */
715 execlists_user_begin(execlists, execlists->port);
717 /* If the engine is now idle, so should be the flag; and vice versa. */
718 GEM_BUG_ON(execlists_is_active(&engine->execlists,
719 EXECLISTS_ACTIVE_USER) ==
720 !port_isset(engine->execlists.port));
724 execlists_cancel_port_requests(struct intel_engine_execlists * const execlists)
726 struct execlist_port *port = execlists->port;
727 unsigned int num_ports = execlists_num_ports(execlists);
729 while (num_ports-- && port_isset(port)) {
730 struct i915_request *rq = port_request(port);
732 GEM_TRACE("%s:port%u global=%d (fence %llx:%d), (current %d)\n",
734 (unsigned int)(port - execlists->port),
736 rq->fence.context, rq->fence.seqno,
737 intel_engine_get_seqno(rq->engine));
739 GEM_BUG_ON(!execlists->active);
740 execlists_context_schedule_out(rq,
741 i915_request_completed(rq) ?
742 INTEL_CONTEXT_SCHEDULE_OUT :
743 INTEL_CONTEXT_SCHEDULE_PREEMPTED);
745 i915_request_put(rq);
747 memset(port, 0, sizeof(*port));
751 execlists_clear_all_active(execlists);
755 invalidate_csb_entries(const u32 *first, const u32 *last)
757 clflush((void *)first);
758 clflush((void *)last);
761 static void reset_csb_pointers(struct intel_engine_execlists *execlists)
763 const unsigned int reset_value = GEN8_CSB_ENTRIES - 1;
766 * After a reset, the HW starts writing into CSB entry [0]. We
767 * therefore have to set our HEAD pointer back one entry so that
768 * the *first* entry we check is entry 0. To complicate this further,
769 * as we don't wait for the first interrupt after reset, we have to
770 * fake the HW write to point back to the last entry so that our
771 * inline comparison of our cached head position against the last HW
772 * write works even before the first interrupt.
774 execlists->csb_head = reset_value;
775 WRITE_ONCE(*execlists->csb_write, reset_value);
777 invalidate_csb_entries(&execlists->csb_status[0],
778 &execlists->csb_status[GEN8_CSB_ENTRIES - 1]);
781 static void nop_submission_tasklet(unsigned long data)
783 /* The driver is wedged; don't process any more events. */
786 static void execlists_cancel_requests(struct intel_engine_cs *engine)
788 struct intel_engine_execlists * const execlists = &engine->execlists;
789 struct i915_request *rq, *rn;
793 GEM_TRACE("%s current %d\n",
794 engine->name, intel_engine_get_seqno(engine));
797 * Before we call engine->cancel_requests(), we should have exclusive
798 * access to the submission state. This is arranged for us by the
799 * caller disabling the interrupt generation, the tasklet and other
800 * threads that may then access the same state, giving us a free hand
801 * to reset state. However, we still need to let lockdep be aware that
802 * we know this state may be accessed in hardirq context, so we
803 * disable the irq around this manipulation and we want to keep
804 * the spinlock focused on its duties and not accidentally conflate
805 * coverage to the submission's irq state. (Similarly, although we
806 * shouldn't need to disable irq around the manipulation of the
807 * submission's irq state, we also wish to remind ourselves that
810 spin_lock_irqsave(&engine->timeline.lock, flags);
812 /* Cancel the requests on the HW and clear the ELSP tracker. */
813 execlists_cancel_port_requests(execlists);
814 execlists_user_end(execlists);
816 /* Mark all executing requests as skipped. */
817 list_for_each_entry(rq, &engine->timeline.requests, link) {
818 GEM_BUG_ON(!rq->global_seqno);
820 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &rq->fence.flags))
823 dma_fence_set_error(&rq->fence, -EIO);
826 /* Flush the queued requests to the timeline list (for retiring). */
827 while ((rb = rb_first_cached(&execlists->queue))) {
828 struct i915_priolist *p = to_priolist(rb);
831 priolist_for_each_request_consume(rq, rn, p, i) {
832 list_del_init(&rq->sched.link);
834 dma_fence_set_error(&rq->fence, -EIO);
835 __i915_request_submit(rq);
838 rb_erase_cached(&p->node, &execlists->queue);
839 if (p->priority != I915_PRIORITY_NORMAL)
840 kmem_cache_free(engine->i915->priorities, p);
843 intel_write_status_page(engine,
845 intel_engine_last_submit(engine));
847 /* Remaining _unready_ requests will be nop'ed when submitted */
849 execlists->queue_priority = INT_MIN;
850 execlists->queue = RB_ROOT_CACHED;
851 GEM_BUG_ON(port_isset(execlists->port));
853 GEM_BUG_ON(__tasklet_is_enabled(&execlists->tasklet));
854 execlists->tasklet.func = nop_submission_tasklet;
856 spin_unlock_irqrestore(&engine->timeline.lock, flags);
860 reset_in_progress(const struct intel_engine_execlists *execlists)
862 return unlikely(!__tasklet_is_enabled(&execlists->tasklet));
865 static void process_csb(struct intel_engine_cs *engine)
867 struct intel_engine_execlists * const execlists = &engine->execlists;
868 struct execlist_port *port = execlists->port;
869 const u32 * const buf = execlists->csb_status;
873 * Note that csb_write, csb_status may be either in HWSP or mmio.
874 * When reading from the csb_write mmio register, we have to be
875 * careful to only use the GEN8_CSB_WRITE_PTR portion, which is
876 * the low 4bits. As it happens we know the next 4bits are always
877 * zero and so we can simply masked off the low u8 of the register
878 * and treat it identically to reading from the HWSP (without having
879 * to use explicit shifting and masking, and probably bifurcating
880 * the code to handle the legacy mmio read).
882 head = execlists->csb_head;
883 tail = READ_ONCE(*execlists->csb_write);
884 GEM_TRACE("%s cs-irq head=%d, tail=%d\n", engine->name, head, tail);
885 if (unlikely(head == tail))
889 * Hopefully paired with a wmb() in HW!
891 * We must complete the read of the write pointer before any reads
892 * from the CSB, so that we do not see stale values. Without an rmb
893 * (lfence) the HW may speculatively perform the CSB[] reads *before*
894 * we perform the READ_ONCE(*csb_write).
899 struct i915_request *rq;
903 if (++head == GEN8_CSB_ENTRIES)
907 * We are flying near dragons again.
909 * We hold a reference to the request in execlist_port[]
910 * but no more than that. We are operating in softirq
911 * context and so cannot hold any mutex or sleep. That
912 * prevents us stopping the requests we are processing
913 * in port[] from being retired simultaneously (the
914 * breadcrumb will be complete before we see the
915 * context-switch). As we only hold the reference to the
916 * request, any pointer chasing underneath the request
917 * is subject to a potential use-after-free. Thus we
918 * store all of the bookkeeping within port[] as
919 * required, and avoid using unguarded pointers beneath
920 * request itself. The same applies to the atomic
924 GEM_TRACE("%s csb[%d]: status=0x%08x:0x%08x, active=0x%x\n",
926 buf[2 * head + 0], buf[2 * head + 1],
929 status = buf[2 * head];
930 if (status & (GEN8_CTX_STATUS_IDLE_ACTIVE |
931 GEN8_CTX_STATUS_PREEMPTED))
932 execlists_set_active(execlists,
933 EXECLISTS_ACTIVE_HWACK);
934 if (status & GEN8_CTX_STATUS_ACTIVE_IDLE)
935 execlists_clear_active(execlists,
936 EXECLISTS_ACTIVE_HWACK);
938 if (!(status & GEN8_CTX_STATUS_COMPLETED_MASK))
941 /* We should never get a COMPLETED | IDLE_ACTIVE! */
942 GEM_BUG_ON(status & GEN8_CTX_STATUS_IDLE_ACTIVE);
944 if (status & GEN8_CTX_STATUS_COMPLETE &&
945 buf[2*head + 1] == execlists->preempt_complete_status) {
946 GEM_TRACE("%s preempt-idle\n", engine->name);
947 complete_preempt_context(execlists);
951 if (status & GEN8_CTX_STATUS_PREEMPTED &&
952 execlists_is_active(execlists,
953 EXECLISTS_ACTIVE_PREEMPT))
956 GEM_BUG_ON(!execlists_is_active(execlists,
957 EXECLISTS_ACTIVE_USER));
959 rq = port_unpack(port, &count);
960 GEM_TRACE("%s out[0]: ctx=%d.%d, global=%d (fence %llx:%d) (current %d), prio=%d\n",
962 port->context_id, count,
963 rq ? rq->global_seqno : 0,
964 rq ? rq->fence.context : 0,
965 rq ? rq->fence.seqno : 0,
966 intel_engine_get_seqno(engine),
967 rq ? rq_prio(rq) : 0);
969 /* Check the context/desc id for this event matches */
970 GEM_DEBUG_BUG_ON(buf[2 * head + 1] != port->context_id);
972 GEM_BUG_ON(count == 0);
975 * On the final event corresponding to the
976 * submission of this context, we expect either
977 * an element-switch event or a completion
978 * event (and on completion, the active-idle
979 * marker). No more preemptions, lite-restore
982 GEM_BUG_ON(status & GEN8_CTX_STATUS_PREEMPTED);
983 GEM_BUG_ON(port_isset(&port[1]) &&
984 !(status & GEN8_CTX_STATUS_ELEMENT_SWITCH));
985 GEM_BUG_ON(!port_isset(&port[1]) &&
986 !(status & GEN8_CTX_STATUS_ACTIVE_IDLE));
989 * We rely on the hardware being strongly
990 * ordered, that the breadcrumb write is
991 * coherent (visible from the CPU) before the
992 * user interrupt and CSB is processed.
994 GEM_BUG_ON(!i915_request_completed(rq));
996 execlists_context_schedule_out(rq,
997 INTEL_CONTEXT_SCHEDULE_OUT);
998 i915_request_put(rq);
1000 GEM_TRACE("%s completed ctx=%d\n",
1001 engine->name, port->context_id);
1003 port = execlists_port_complete(execlists, port);
1004 if (port_isset(port))
1005 execlists_user_begin(execlists, port);
1007 execlists_user_end(execlists);
1009 port_set(port, port_pack(rq, count));
1011 } while (head != tail);
1013 execlists->csb_head = head;
1016 * Gen11 has proven to fail wrt global observation point between
1017 * entry and tail update, failing on the ordering and thus
1018 * we see an old entry in the context status buffer.
1020 * Forcibly evict out entries for the next gpu csb update,
1021 * to increase the odds that we get a fresh entries with non
1022 * working hardware. The cost for doing so comes out mostly with
1023 * the wash as hardware, working or not, will need to do the
1024 * invalidation before.
1026 invalidate_csb_entries(&buf[0], &buf[GEN8_CSB_ENTRIES - 1]);
1029 static void __execlists_submission_tasklet(struct intel_engine_cs *const engine)
1031 lockdep_assert_held(&engine->timeline.lock);
1033 process_csb(engine);
1034 if (!execlists_is_active(&engine->execlists, EXECLISTS_ACTIVE_PREEMPT))
1035 execlists_dequeue(engine);
1039 * Check the unread Context Status Buffers and manage the submission of new
1040 * contexts to the ELSP accordingly.
1042 static void execlists_submission_tasklet(unsigned long data)
1044 struct intel_engine_cs * const engine = (struct intel_engine_cs *)data;
1045 unsigned long flags;
1047 GEM_TRACE("%s awake?=%d, active=%x\n",
1049 engine->i915->gt.awake,
1050 engine->execlists.active);
1052 spin_lock_irqsave(&engine->timeline.lock, flags);
1053 __execlists_submission_tasklet(engine);
1054 spin_unlock_irqrestore(&engine->timeline.lock, flags);
1057 static void queue_request(struct intel_engine_cs *engine,
1058 struct i915_sched_node *node,
1061 list_add_tail(&node->link, i915_sched_lookup_priolist(engine, prio));
1064 static void __submit_queue_imm(struct intel_engine_cs *engine)
1066 struct intel_engine_execlists * const execlists = &engine->execlists;
1068 if (reset_in_progress(execlists))
1069 return; /* defer until we restart the engine following reset */
1071 if (execlists->tasklet.func == execlists_submission_tasklet)
1072 __execlists_submission_tasklet(engine);
1074 tasklet_hi_schedule(&execlists->tasklet);
1077 static void submit_queue(struct intel_engine_cs *engine, int prio)
1079 if (prio > engine->execlists.queue_priority) {
1080 engine->execlists.queue_priority = prio;
1081 __submit_queue_imm(engine);
1085 static void execlists_submit_request(struct i915_request *request)
1087 struct intel_engine_cs *engine = request->engine;
1088 unsigned long flags;
1090 /* Will be called from irq-context when using foreign fences. */
1091 spin_lock_irqsave(&engine->timeline.lock, flags);
1093 queue_request(engine, &request->sched, rq_prio(request));
1095 GEM_BUG_ON(RB_EMPTY_ROOT(&engine->execlists.queue.rb_root));
1096 GEM_BUG_ON(list_empty(&request->sched.link));
1098 submit_queue(engine, rq_prio(request));
1100 spin_unlock_irqrestore(&engine->timeline.lock, flags);
1103 static void execlists_context_destroy(struct intel_context *ce)
1105 GEM_BUG_ON(ce->pin_count);
1110 intel_ring_free(ce->ring);
1112 GEM_BUG_ON(i915_gem_object_is_active(ce->state->obj));
1113 i915_gem_object_put(ce->state->obj);
1116 static void execlists_context_unpin(struct intel_context *ce)
1118 struct intel_engine_cs *engine;
1121 * The tasklet may still be using a pointer to our state, via an
1122 * old request. However, since we know we only unpin the context
1123 * on retirement of the following request, we know that the last
1124 * request referencing us will have had a completion CS interrupt.
1125 * If we see that it is still active, it means that the tasklet hasn't
1126 * had the chance to run yet; let it run before we teardown the
1127 * reference it may use.
1129 engine = READ_ONCE(ce->active);
1130 if (unlikely(engine)) {
1131 unsigned long flags;
1133 spin_lock_irqsave(&engine->timeline.lock, flags);
1134 process_csb(engine);
1135 spin_unlock_irqrestore(&engine->timeline.lock, flags);
1137 GEM_BUG_ON(READ_ONCE(ce->active));
1140 i915_gem_context_unpin_hw_id(ce->gem_context);
1142 intel_ring_unpin(ce->ring);
1144 ce->state->obj->pin_global--;
1145 i915_gem_object_unpin_map(ce->state->obj);
1146 i915_vma_unpin(ce->state);
1148 i915_gem_context_put(ce->gem_context);
1151 static int __context_pin(struct i915_gem_context *ctx, struct i915_vma *vma)
1157 * Clear this page out of any CPU caches for coherent swap-in/out.
1158 * We only want to do this on the first bind so that we do not stall
1159 * on an active context (which by nature is already on the GPU).
1161 if (!(vma->flags & I915_VMA_GLOBAL_BIND)) {
1162 err = i915_gem_object_set_to_wc_domain(vma->obj, true);
1167 flags = PIN_GLOBAL | PIN_HIGH;
1168 flags |= PIN_OFFSET_BIAS | i915_ggtt_pin_bias(vma);
1170 return i915_vma_pin(vma, 0, 0, flags);
1173 static struct intel_context *
1174 __execlists_context_pin(struct intel_engine_cs *engine,
1175 struct i915_gem_context *ctx,
1176 struct intel_context *ce)
1181 ret = execlists_context_deferred_alloc(ctx, engine, ce);
1184 GEM_BUG_ON(!ce->state);
1186 ret = __context_pin(ctx, ce->state);
1190 vaddr = i915_gem_object_pin_map(ce->state->obj,
1191 i915_coherent_map_type(ctx->i915) |
1193 if (IS_ERR(vaddr)) {
1194 ret = PTR_ERR(vaddr);
1198 ret = intel_ring_pin(ce->ring);
1202 ret = i915_gem_context_pin_hw_id(ctx);
1206 intel_lr_context_descriptor_update(ctx, engine, ce);
1208 GEM_BUG_ON(!intel_ring_offset_valid(ce->ring, ce->ring->head));
1210 ce->lrc_reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
1211 ce->lrc_reg_state[CTX_RING_BUFFER_START+1] =
1212 i915_ggtt_offset(ce->ring->vma);
1213 ce->lrc_reg_state[CTX_RING_HEAD + 1] = ce->ring->head;
1214 ce->lrc_reg_state[CTX_RING_TAIL + 1] = ce->ring->tail;
1216 ce->state->obj->pin_global++;
1217 i915_gem_context_get(ctx);
1221 intel_ring_unpin(ce->ring);
1223 i915_gem_object_unpin_map(ce->state->obj);
1225 __i915_vma_unpin(ce->state);
1228 return ERR_PTR(ret);
1231 static const struct intel_context_ops execlists_context_ops = {
1232 .unpin = execlists_context_unpin,
1233 .destroy = execlists_context_destroy,
1236 static struct intel_context *
1237 execlists_context_pin(struct intel_engine_cs *engine,
1238 struct i915_gem_context *ctx)
1240 struct intel_context *ce = to_intel_context(ctx, engine);
1242 lockdep_assert_held(&ctx->i915->drm.struct_mutex);
1243 GEM_BUG_ON(!ctx->ppgtt);
1245 if (likely(ce->pin_count++))
1247 GEM_BUG_ON(!ce->pin_count); /* no overflow please! */
1249 ce->ops = &execlists_context_ops;
1251 return __execlists_context_pin(engine, ctx, ce);
1254 static int emit_pdps(struct i915_request *rq)
1256 const struct intel_engine_cs * const engine = rq->engine;
1257 struct i915_hw_ppgtt * const ppgtt = rq->gem_context->ppgtt;
1261 GEM_BUG_ON(intel_vgpu_active(rq->i915));
1264 * Beware ye of the dragons, this sequence is magic!
1266 * Small changes to this sequence can cause anything from
1267 * GPU hangs to forcewake errors and machine lockups!
1270 /* Flush any residual operations from the context load */
1271 err = engine->emit_flush(rq, EMIT_FLUSH);
1275 /* Magic required to prevent forcewake errors! */
1276 err = engine->emit_flush(rq, EMIT_INVALIDATE);
1280 cs = intel_ring_begin(rq, 4 * GEN8_3LVL_PDPES + 2);
1284 /* Ensure the LRI have landed before we invalidate & continue */
1285 *cs++ = MI_LOAD_REGISTER_IMM(2 * GEN8_3LVL_PDPES) | MI_LRI_FORCE_POSTED;
1286 for (i = GEN8_3LVL_PDPES; i--; ) {
1287 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
1289 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(engine, i));
1290 *cs++ = upper_32_bits(pd_daddr);
1291 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(engine, i));
1292 *cs++ = lower_32_bits(pd_daddr);
1296 intel_ring_advance(rq, cs);
1298 /* Be doubly sure the LRI have landed before proceeding */
1299 err = engine->emit_flush(rq, EMIT_FLUSH);
1303 /* Re-invalidate the TLB for luck */
1304 return engine->emit_flush(rq, EMIT_INVALIDATE);
1307 static int execlists_request_alloc(struct i915_request *request)
1311 GEM_BUG_ON(!request->hw_context->pin_count);
1314 * Flush enough space to reduce the likelihood of waiting after
1315 * we start building the request - in which case we will just
1316 * have to repeat work.
1318 request->reserved_space += EXECLISTS_REQUEST_SIZE;
1321 * Note that after this point, we have committed to using
1322 * this request as it is being used to both track the
1323 * state of engine initialisation and liveness of the
1324 * golden renderstate above. Think twice before you try
1325 * to cancel/unwind this request now.
1328 /* Unconditionally invalidate GPU caches and TLBs. */
1329 if (i915_vm_is_48bit(&request->gem_context->ppgtt->vm))
1330 ret = request->engine->emit_flush(request, EMIT_INVALIDATE);
1332 ret = emit_pdps(request);
1336 request->reserved_space -= EXECLISTS_REQUEST_SIZE;
1341 * In this WA we need to set GEN8_L3SQCREG4[21:21] and reset it after
1342 * PIPE_CONTROL instruction. This is required for the flush to happen correctly
1343 * but there is a slight complication as this is applied in WA batch where the
1344 * values are only initialized once so we cannot take register value at the
1345 * beginning and reuse it further; hence we save its value to memory, upload a
1346 * constant value with bit21 set and then we restore it back with the saved value.
1347 * To simplify the WA, a constant value is formed by using the default value
1348 * of this register. This shouldn't be a problem because we are only modifying
1349 * it for a short period and this batch in non-premptible. We can ofcourse
1350 * use additional instructions that read the actual value of the register
1351 * at that time and set our bit of interest but it makes the WA complicated.
1353 * This WA is also required for Gen9 so extracting as a function avoids
1357 gen8_emit_flush_coherentl3_wa(struct intel_engine_cs *engine, u32 *batch)
1359 /* NB no one else is allowed to scribble over scratch + 256! */
1360 *batch++ = MI_STORE_REGISTER_MEM_GEN8 | MI_SRM_LRM_GLOBAL_GTT;
1361 *batch++ = i915_mmio_reg_offset(GEN8_L3SQCREG4);
1362 *batch++ = i915_scratch_offset(engine->i915) + 256;
1365 *batch++ = MI_LOAD_REGISTER_IMM(1);
1366 *batch++ = i915_mmio_reg_offset(GEN8_L3SQCREG4);
1367 *batch++ = 0x40400000 | GEN8_LQSC_FLUSH_COHERENT_LINES;
1369 batch = gen8_emit_pipe_control(batch,
1370 PIPE_CONTROL_CS_STALL |
1371 PIPE_CONTROL_DC_FLUSH_ENABLE,
1374 *batch++ = MI_LOAD_REGISTER_MEM_GEN8 | MI_SRM_LRM_GLOBAL_GTT;
1375 *batch++ = i915_mmio_reg_offset(GEN8_L3SQCREG4);
1376 *batch++ = i915_scratch_offset(engine->i915) + 256;
1383 * Typically we only have one indirect_ctx and per_ctx batch buffer which are
1384 * initialized at the beginning and shared across all contexts but this field
1385 * helps us to have multiple batches at different offsets and select them based
1386 * on a criteria. At the moment this batch always start at the beginning of the page
1387 * and at this point we don't have multiple wa_ctx batch buffers.
1389 * The number of WA applied are not known at the beginning; we use this field
1390 * to return the no of DWORDS written.
1392 * It is to be noted that this batch does not contain MI_BATCH_BUFFER_END
1393 * so it adds NOOPs as padding to make it cacheline aligned.
1394 * MI_BATCH_BUFFER_END will be added to perctx batch and both of them together
1395 * makes a complete batch buffer.
1397 static u32 *gen8_init_indirectctx_bb(struct intel_engine_cs *engine, u32 *batch)
1399 /* WaDisableCtxRestoreArbitration:bdw,chv */
1400 *batch++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
1402 /* WaFlushCoherentL3CacheLinesAtContextSwitch:bdw */
1403 if (IS_BROADWELL(engine->i915))
1404 batch = gen8_emit_flush_coherentl3_wa(engine, batch);
1406 /* WaClearSlmSpaceAtContextSwitch:bdw,chv */
1407 /* Actual scratch location is at 128 bytes offset */
1408 batch = gen8_emit_pipe_control(batch,
1409 PIPE_CONTROL_FLUSH_L3 |
1410 PIPE_CONTROL_GLOBAL_GTT_IVB |
1411 PIPE_CONTROL_CS_STALL |
1412 PIPE_CONTROL_QW_WRITE,
1413 i915_scratch_offset(engine->i915) +
1414 2 * CACHELINE_BYTES);
1416 *batch++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
1418 /* Pad to end of cacheline */
1419 while ((unsigned long)batch % CACHELINE_BYTES)
1423 * MI_BATCH_BUFFER_END is not required in Indirect ctx BB because
1424 * execution depends on the length specified in terms of cache lines
1425 * in the register CTX_RCS_INDIRECT_CTX
1436 static u32 *emit_lri(u32 *batch, const struct lri *lri, unsigned int count)
1438 GEM_BUG_ON(!count || count > 63);
1440 *batch++ = MI_LOAD_REGISTER_IMM(count);
1442 *batch++ = i915_mmio_reg_offset(lri->reg);
1443 *batch++ = lri->value;
1444 } while (lri++, --count);
1450 static u32 *gen9_init_indirectctx_bb(struct intel_engine_cs *engine, u32 *batch)
1452 static const struct lri lri[] = {
1453 /* WaDisableGatherAtSetShaderCommonSlice:skl,bxt,kbl,glk */
1455 COMMON_SLICE_CHICKEN2,
1456 __MASKED_FIELD(GEN9_DISABLE_GATHER_AT_SET_SHADER_COMMON_SLICE,
1463 __MASKED_FIELD(FF_SLICE_CHICKEN_CL_PROVOKING_VERTEX_FIX,
1464 FF_SLICE_CHICKEN_CL_PROVOKING_VERTEX_FIX),
1470 __MASKED_FIELD(_3D_CHICKEN_SF_PROVOKING_VERTEX_FIX,
1471 _3D_CHICKEN_SF_PROVOKING_VERTEX_FIX),
1475 *batch++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
1477 /* WaFlushCoherentL3CacheLinesAtContextSwitch:skl,bxt,glk */
1478 batch = gen8_emit_flush_coherentl3_wa(engine, batch);
1480 batch = emit_lri(batch, lri, ARRAY_SIZE(lri));
1482 /* WaMediaPoolStateCmdInWABB:bxt,glk */
1483 if (HAS_POOLED_EU(engine->i915)) {
1485 * EU pool configuration is setup along with golden context
1486 * during context initialization. This value depends on
1487 * device type (2x6 or 3x6) and needs to be updated based
1488 * on which subslice is disabled especially for 2x6
1489 * devices, however it is safe to load default
1490 * configuration of 3x6 device instead of masking off
1491 * corresponding bits because HW ignores bits of a disabled
1492 * subslice and drops down to appropriate config. Please
1493 * see render_state_setup() in i915_gem_render_state.c for
1494 * possible configurations, to avoid duplication they are
1495 * not shown here again.
1497 *batch++ = GEN9_MEDIA_POOL_STATE;
1498 *batch++ = GEN9_MEDIA_POOL_ENABLE;
1499 *batch++ = 0x00777000;
1505 *batch++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
1507 /* Pad to end of cacheline */
1508 while ((unsigned long)batch % CACHELINE_BYTES)
1515 gen10_init_indirectctx_bb(struct intel_engine_cs *engine, u32 *batch)
1520 * WaPipeControlBefore3DStateSamplePattern: cnl
1522 * Ensure the engine is idle prior to programming a
1523 * 3DSTATE_SAMPLE_PATTERN during a context restore.
1525 batch = gen8_emit_pipe_control(batch,
1526 PIPE_CONTROL_CS_STALL,
1529 * WaPipeControlBefore3DStateSamplePattern says we need 4 dwords for
1530 * the PIPE_CONTROL followed by 12 dwords of 0x0, so 16 dwords in
1531 * total. However, a PIPE_CONTROL is 6 dwords long, not 4, which is
1532 * confusing. Since gen8_emit_pipe_control() already advances the
1533 * batch by 6 dwords, we advance the other 10 here, completing a
1534 * cacheline. It's not clear if the workaround requires this padding
1535 * before other commands, or if it's just the regular padding we would
1536 * already have for the workaround bb, so leave it here for now.
1538 for (i = 0; i < 10; i++)
1541 /* Pad to end of cacheline */
1542 while ((unsigned long)batch % CACHELINE_BYTES)
1548 #define CTX_WA_BB_OBJ_SIZE (PAGE_SIZE)
1550 static int lrc_setup_wa_ctx(struct intel_engine_cs *engine)
1552 struct drm_i915_gem_object *obj;
1553 struct i915_vma *vma;
1556 obj = i915_gem_object_create(engine->i915, CTX_WA_BB_OBJ_SIZE);
1558 return PTR_ERR(obj);
1560 vma = i915_vma_instance(obj, &engine->i915->ggtt.vm, NULL);
1566 err = i915_vma_pin(vma, 0, 0, PIN_GLOBAL | PIN_HIGH);
1570 engine->wa_ctx.vma = vma;
1574 i915_gem_object_put(obj);
1578 static void lrc_destroy_wa_ctx(struct intel_engine_cs *engine)
1580 i915_vma_unpin_and_release(&engine->wa_ctx.vma, 0);
1583 typedef u32 *(*wa_bb_func_t)(struct intel_engine_cs *engine, u32 *batch);
1585 static int intel_init_workaround_bb(struct intel_engine_cs *engine)
1587 struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
1588 struct i915_wa_ctx_bb *wa_bb[2] = { &wa_ctx->indirect_ctx,
1590 wa_bb_func_t wa_bb_fn[2];
1592 void *batch, *batch_ptr;
1596 if (GEM_DEBUG_WARN_ON(engine->id != RCS))
1599 switch (INTEL_GEN(engine->i915)) {
1603 wa_bb_fn[0] = gen10_init_indirectctx_bb;
1607 wa_bb_fn[0] = gen9_init_indirectctx_bb;
1611 wa_bb_fn[0] = gen8_init_indirectctx_bb;
1615 MISSING_CASE(INTEL_GEN(engine->i915));
1619 ret = lrc_setup_wa_ctx(engine);
1621 DRM_DEBUG_DRIVER("Failed to setup context WA page: %d\n", ret);
1625 page = i915_gem_object_get_dirty_page(wa_ctx->vma->obj, 0);
1626 batch = batch_ptr = kmap_atomic(page);
1629 * Emit the two workaround batch buffers, recording the offset from the
1630 * start of the workaround batch buffer object for each and their
1633 for (i = 0; i < ARRAY_SIZE(wa_bb_fn); i++) {
1634 wa_bb[i]->offset = batch_ptr - batch;
1635 if (GEM_DEBUG_WARN_ON(!IS_ALIGNED(wa_bb[i]->offset,
1636 CACHELINE_BYTES))) {
1641 batch_ptr = wa_bb_fn[i](engine, batch_ptr);
1642 wa_bb[i]->size = batch_ptr - (batch + wa_bb[i]->offset);
1645 BUG_ON(batch_ptr - batch > CTX_WA_BB_OBJ_SIZE);
1647 kunmap_atomic(batch);
1649 lrc_destroy_wa_ctx(engine);
1654 static void enable_execlists(struct intel_engine_cs *engine)
1656 struct drm_i915_private *dev_priv = engine->i915;
1658 intel_engine_set_hwsp_writemask(engine, ~0u); /* HWSTAM */
1661 * Make sure we're not enabling the new 12-deep CSB
1662 * FIFO as that requires a slightly updated handling
1663 * in the ctx switch irq. Since we're currently only
1664 * using only 2 elements of the enhanced execlists the
1665 * deeper FIFO it's not needed and it's not worth adding
1666 * more statements to the irq handler to support it.
1668 if (INTEL_GEN(dev_priv) >= 11)
1669 I915_WRITE(RING_MODE_GEN7(engine),
1670 _MASKED_BIT_DISABLE(GEN11_GFX_DISABLE_LEGACY_MODE));
1672 I915_WRITE(RING_MODE_GEN7(engine),
1673 _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE));
1675 I915_WRITE(RING_MI_MODE(engine->mmio_base),
1676 _MASKED_BIT_DISABLE(STOP_RING));
1678 I915_WRITE(RING_HWS_PGA(engine->mmio_base),
1679 engine->status_page.ggtt_offset);
1680 POSTING_READ(RING_HWS_PGA(engine->mmio_base));
1683 static bool unexpected_starting_state(struct intel_engine_cs *engine)
1685 struct drm_i915_private *dev_priv = engine->i915;
1686 bool unexpected = false;
1688 if (I915_READ(RING_MI_MODE(engine->mmio_base)) & STOP_RING) {
1689 DRM_DEBUG_DRIVER("STOP_RING still set in RING_MI_MODE\n");
1696 static int gen8_init_common_ring(struct intel_engine_cs *engine)
1698 intel_engine_apply_workarounds(engine);
1699 intel_engine_apply_whitelist(engine);
1701 intel_mocs_init_engine(engine);
1703 intel_engine_reset_breadcrumbs(engine);
1705 if (GEM_SHOW_DEBUG() && unexpected_starting_state(engine)) {
1706 struct drm_printer p = drm_debug_printer(__func__);
1708 intel_engine_dump(engine, &p, NULL);
1711 enable_execlists(engine);
1716 static struct i915_request *
1717 execlists_reset_prepare(struct intel_engine_cs *engine)
1719 struct intel_engine_execlists * const execlists = &engine->execlists;
1720 struct i915_request *request, *active;
1721 unsigned long flags;
1723 GEM_TRACE("%s: depth<-%d\n", engine->name,
1724 atomic_read(&execlists->tasklet.count));
1727 * Prevent request submission to the hardware until we have
1728 * completed the reset in i915_gem_reset_finish(). If a request
1729 * is completed by one engine, it may then queue a request
1730 * to a second via its execlists->tasklet *just* as we are
1731 * calling engine->init_hw() and also writing the ELSP.
1732 * Turning off the execlists->tasklet until the reset is over
1733 * prevents the race.
1735 __tasklet_disable_sync_once(&execlists->tasklet);
1737 spin_lock_irqsave(&engine->timeline.lock, flags);
1740 * We want to flush the pending context switches, having disabled
1741 * the tasklet above, we can assume exclusive access to the execlists.
1742 * For this allows us to catch up with an inflight preemption event,
1743 * and avoid blaming an innocent request if the stall was due to the
1744 * preemption itself.
1746 process_csb(engine);
1749 * The last active request can then be no later than the last request
1750 * now in ELSP[0]. So search backwards from there, so that if the GPU
1751 * has advanced beyond the last CSB update, it will be pardoned.
1754 request = port_request(execlists->port);
1757 * Prevent the breadcrumb from advancing before we decide
1758 * which request is currently active.
1760 intel_engine_stop_cs(engine);
1762 list_for_each_entry_from_reverse(request,
1763 &engine->timeline.requests,
1765 if (__i915_request_completed(request,
1766 request->global_seqno))
1773 spin_unlock_irqrestore(&engine->timeline.lock, flags);
1778 static void execlists_reset(struct intel_engine_cs *engine,
1779 struct i915_request *request)
1781 struct intel_engine_execlists * const execlists = &engine->execlists;
1782 unsigned long flags;
1785 GEM_TRACE("%s request global=%d, current=%d\n",
1786 engine->name, request ? request->global_seqno : 0,
1787 intel_engine_get_seqno(engine));
1789 spin_lock_irqsave(&engine->timeline.lock, flags);
1792 * Catch up with any missed context-switch interrupts.
1794 * Ideally we would just read the remaining CSB entries now that we
1795 * know the gpu is idle. However, the CSB registers are sometimes^W
1796 * often trashed across a GPU reset! Instead we have to rely on
1797 * guessing the missed context-switch events by looking at what
1798 * requests were completed.
1800 execlists_cancel_port_requests(execlists);
1802 /* Push back any incomplete requests for replay after the reset. */
1803 __unwind_incomplete_requests(engine);
1805 /* Following the reset, we need to reload the CSB read/write pointers */
1806 reset_csb_pointers(&engine->execlists);
1808 spin_unlock_irqrestore(&engine->timeline.lock, flags);
1811 * If the request was innocent, we leave the request in the ELSP
1812 * and will try to replay it on restarting. The context image may
1813 * have been corrupted by the reset, in which case we may have
1814 * to service a new GPU hang, but more likely we can continue on
1817 * If the request was guilty, we presume the context is corrupt
1818 * and have to at least restore the RING register in the context
1819 * image back to the expected values to skip over the guilty request.
1821 if (!request || request->fence.error != -EIO)
1825 * We want a simple context + ring to execute the breadcrumb update.
1826 * We cannot rely on the context being intact across the GPU hang,
1827 * so clear it and rebuild just what we need for the breadcrumb.
1828 * All pending requests for this context will be zapped, and any
1829 * future request will be after userspace has had the opportunity
1830 * to recreate its own state.
1832 regs = request->hw_context->lrc_reg_state;
1833 if (engine->pinned_default_state) {
1834 memcpy(regs, /* skip restoring the vanilla PPHWSP */
1835 engine->pinned_default_state + LRC_STATE_PN * PAGE_SIZE,
1836 engine->context_size - PAGE_SIZE);
1838 execlists_init_reg_state(regs,
1839 request->gem_context, engine, request->ring);
1841 /* Move the RING_HEAD onto the breadcrumb, past the hanging batch */
1842 regs[CTX_RING_BUFFER_START + 1] = i915_ggtt_offset(request->ring->vma);
1844 request->ring->head = intel_ring_wrap(request->ring, request->postfix);
1845 regs[CTX_RING_HEAD + 1] = request->ring->head;
1847 intel_ring_update_space(request->ring);
1849 /* Reset WaIdleLiteRestore:bdw,skl as well */
1850 unwind_wa_tail(request);
1853 static void execlists_reset_finish(struct intel_engine_cs *engine)
1855 struct intel_engine_execlists * const execlists = &engine->execlists;
1858 * After a GPU reset, we may have requests to replay. Do so now while
1859 * we still have the forcewake to be sure that the GPU is not allowed
1860 * to sleep before we restart and reload a context.
1863 if (!RB_EMPTY_ROOT(&execlists->queue.rb_root))
1864 execlists->tasklet.func(execlists->tasklet.data);
1866 tasklet_enable(&execlists->tasklet);
1867 GEM_TRACE("%s: depth->%d\n", engine->name,
1868 atomic_read(&execlists->tasklet.count));
1871 static int gen8_emit_bb_start(struct i915_request *rq,
1872 u64 offset, u32 len,
1873 const unsigned int flags)
1877 cs = intel_ring_begin(rq, 6);
1882 * WaDisableCtxRestoreArbitration:bdw,chv
1884 * We don't need to perform MI_ARB_ENABLE as often as we do (in
1885 * particular all the gen that do not need the w/a at all!), if we
1886 * took care to make sure that on every switch into this context
1887 * (both ordinary and for preemption) that arbitrartion was enabled
1888 * we would be fine. However, there doesn't seem to be a downside to
1889 * being paranoid and making sure it is set before each batch and
1890 * every context-switch.
1892 * Note that if we fail to enable arbitration before the request
1893 * is complete, then we do not see the context-switch interrupt and
1894 * the engine hangs (with RING_HEAD == RING_TAIL).
1896 * That satisfies both the GPGPU w/a and our heavy-handed paranoia.
1898 *cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
1900 /* FIXME(BDW): Address space and security selectors. */
1901 *cs++ = MI_BATCH_BUFFER_START_GEN8 |
1902 (flags & I915_DISPATCH_SECURE ? 0 : BIT(8));
1903 *cs++ = lower_32_bits(offset);
1904 *cs++ = upper_32_bits(offset);
1906 *cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
1909 intel_ring_advance(rq, cs);
1914 static void gen8_logical_ring_enable_irq(struct intel_engine_cs *engine)
1916 struct drm_i915_private *dev_priv = engine->i915;
1917 I915_WRITE_IMR(engine,
1918 ~(engine->irq_enable_mask | engine->irq_keep_mask));
1919 POSTING_READ_FW(RING_IMR(engine->mmio_base));
1922 static void gen8_logical_ring_disable_irq(struct intel_engine_cs *engine)
1924 struct drm_i915_private *dev_priv = engine->i915;
1925 I915_WRITE_IMR(engine, ~engine->irq_keep_mask);
1928 static int gen8_emit_flush(struct i915_request *request, u32 mode)
1932 cs = intel_ring_begin(request, 4);
1936 cmd = MI_FLUSH_DW + 1;
1938 /* We always require a command barrier so that subsequent
1939 * commands, such as breadcrumb interrupts, are strictly ordered
1940 * wrt the contents of the write cache being flushed to memory
1941 * (and thus being coherent from the CPU).
1943 cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1945 if (mode & EMIT_INVALIDATE) {
1946 cmd |= MI_INVALIDATE_TLB;
1947 if (request->engine->class == VIDEO_DECODE_CLASS)
1948 cmd |= MI_INVALIDATE_BSD;
1952 *cs++ = I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT;
1953 *cs++ = 0; /* upper addr */
1954 *cs++ = 0; /* value */
1955 intel_ring_advance(request, cs);
1960 static int gen8_emit_flush_render(struct i915_request *request,
1963 struct intel_engine_cs *engine = request->engine;
1965 i915_scratch_offset(engine->i915) + 2 * CACHELINE_BYTES;
1966 bool vf_flush_wa = false, dc_flush_wa = false;
1970 flags |= PIPE_CONTROL_CS_STALL;
1972 if (mode & EMIT_FLUSH) {
1973 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
1974 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
1975 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
1976 flags |= PIPE_CONTROL_FLUSH_ENABLE;
1979 if (mode & EMIT_INVALIDATE) {
1980 flags |= PIPE_CONTROL_TLB_INVALIDATE;
1981 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
1982 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1983 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
1984 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
1985 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
1986 flags |= PIPE_CONTROL_QW_WRITE;
1987 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
1990 * On GEN9: before VF_CACHE_INVALIDATE we need to emit a NULL
1993 if (IS_GEN(request->i915, 9))
1996 /* WaForGAMHang:kbl */
1997 if (IS_KBL_REVID(request->i915, 0, KBL_REVID_B0))
2009 cs = intel_ring_begin(request, len);
2014 cs = gen8_emit_pipe_control(cs, 0, 0);
2017 cs = gen8_emit_pipe_control(cs, PIPE_CONTROL_DC_FLUSH_ENABLE,
2020 cs = gen8_emit_pipe_control(cs, flags, scratch_addr);
2023 cs = gen8_emit_pipe_control(cs, PIPE_CONTROL_CS_STALL, 0);
2025 intel_ring_advance(request, cs);
2031 * Reserve space for 2 NOOPs at the end of each request to be
2032 * used as a workaround for not being allowed to do lite
2033 * restore with HEAD==TAIL (WaIdleLiteRestore).
2035 static void gen8_emit_wa_tail(struct i915_request *request, u32 *cs)
2037 /* Ensure there's always at least one preemption point per-request. */
2038 *cs++ = MI_ARB_CHECK;
2040 request->wa_tail = intel_ring_offset(request, cs);
2043 static void gen8_emit_breadcrumb(struct i915_request *request, u32 *cs)
2045 /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */
2046 BUILD_BUG_ON(I915_GEM_HWS_INDEX_ADDR & (1 << 5));
2048 cs = gen8_emit_ggtt_write(cs, request->global_seqno,
2049 intel_hws_seqno_address(request->engine));
2050 *cs++ = MI_USER_INTERRUPT;
2051 *cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
2052 request->tail = intel_ring_offset(request, cs);
2053 assert_ring_tail_valid(request->ring, request->tail);
2055 gen8_emit_wa_tail(request, cs);
2057 static const int gen8_emit_breadcrumb_sz = 6 + WA_TAIL_DWORDS;
2059 static void gen8_emit_breadcrumb_rcs(struct i915_request *request, u32 *cs)
2061 /* We're using qword write, seqno should be aligned to 8 bytes. */
2062 BUILD_BUG_ON(I915_GEM_HWS_INDEX & 1);
2064 cs = gen8_emit_ggtt_write_rcs(cs,
2065 request->global_seqno,
2066 intel_hws_seqno_address(request->engine),
2067 PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
2068 PIPE_CONTROL_DEPTH_CACHE_FLUSH |
2069 PIPE_CONTROL_DC_FLUSH_ENABLE |
2070 PIPE_CONTROL_FLUSH_ENABLE |
2071 PIPE_CONTROL_CS_STALL);
2073 *cs++ = MI_USER_INTERRUPT;
2074 *cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
2076 request->tail = intel_ring_offset(request, cs);
2077 assert_ring_tail_valid(request->ring, request->tail);
2079 gen8_emit_wa_tail(request, cs);
2081 static const int gen8_emit_breadcrumb_rcs_sz = 8 + WA_TAIL_DWORDS;
2083 static int gen8_init_rcs_context(struct i915_request *rq)
2087 ret = intel_engine_emit_ctx_wa(rq);
2091 ret = intel_rcs_context_init_mocs(rq);
2093 * Failing to program the MOCS is non-fatal.The system will not
2094 * run at peak performance. So generate an error and carry on.
2097 DRM_ERROR("MOCS failed to program: expect performance issues.\n");
2099 return i915_gem_render_state_emit(rq);
2103 * intel_logical_ring_cleanup() - deallocate the Engine Command Streamer
2104 * @engine: Engine Command Streamer.
2106 void intel_logical_ring_cleanup(struct intel_engine_cs *engine)
2108 struct drm_i915_private *dev_priv;
2111 * Tasklet cannot be active at this point due intel_mark_active/idle
2112 * so this is just for documentation.
2114 if (WARN_ON(test_bit(TASKLET_STATE_SCHED,
2115 &engine->execlists.tasklet.state)))
2116 tasklet_kill(&engine->execlists.tasklet);
2118 dev_priv = engine->i915;
2120 if (engine->buffer) {
2121 WARN_ON((I915_READ_MODE(engine) & MODE_IDLE) == 0);
2124 if (engine->cleanup)
2125 engine->cleanup(engine);
2127 intel_engine_cleanup_common(engine);
2129 lrc_destroy_wa_ctx(engine);
2131 engine->i915 = NULL;
2132 dev_priv->engine[engine->id] = NULL;
2136 void intel_execlists_set_default_submission(struct intel_engine_cs *engine)
2138 engine->submit_request = execlists_submit_request;
2139 engine->cancel_requests = execlists_cancel_requests;
2140 engine->schedule = i915_schedule;
2141 engine->execlists.tasklet.func = execlists_submission_tasklet;
2143 engine->reset.prepare = execlists_reset_prepare;
2145 engine->park = NULL;
2146 engine->unpark = NULL;
2148 engine->flags |= I915_ENGINE_SUPPORTS_STATS;
2149 if (engine->i915->preempt_context)
2150 engine->flags |= I915_ENGINE_HAS_PREEMPTION;
2152 engine->i915->caps.scheduler =
2153 I915_SCHEDULER_CAP_ENABLED |
2154 I915_SCHEDULER_CAP_PRIORITY;
2155 if (intel_engine_has_preemption(engine))
2156 engine->i915->caps.scheduler |= I915_SCHEDULER_CAP_PREEMPTION;
2160 logical_ring_default_vfuncs(struct intel_engine_cs *engine)
2162 /* Default vfuncs which can be overriden by each engine. */
2163 engine->init_hw = gen8_init_common_ring;
2165 engine->reset.prepare = execlists_reset_prepare;
2166 engine->reset.reset = execlists_reset;
2167 engine->reset.finish = execlists_reset_finish;
2169 engine->context_pin = execlists_context_pin;
2170 engine->request_alloc = execlists_request_alloc;
2172 engine->emit_flush = gen8_emit_flush;
2173 engine->emit_breadcrumb = gen8_emit_breadcrumb;
2174 engine->emit_breadcrumb_sz = gen8_emit_breadcrumb_sz;
2176 engine->set_default_submission = intel_execlists_set_default_submission;
2178 if (INTEL_GEN(engine->i915) < 11) {
2179 engine->irq_enable = gen8_logical_ring_enable_irq;
2180 engine->irq_disable = gen8_logical_ring_disable_irq;
2183 * TODO: On Gen11 interrupt masks need to be clear
2184 * to allow C6 entry. Keep interrupts enabled at
2185 * and take the hit of generating extra interrupts
2186 * until a more refined solution exists.
2189 engine->emit_bb_start = gen8_emit_bb_start;
2193 logical_ring_default_irqs(struct intel_engine_cs *engine)
2195 unsigned int shift = 0;
2197 if (INTEL_GEN(engine->i915) < 11) {
2198 const u8 irq_shifts[] = {
2199 [RCS] = GEN8_RCS_IRQ_SHIFT,
2200 [BCS] = GEN8_BCS_IRQ_SHIFT,
2201 [VCS] = GEN8_VCS1_IRQ_SHIFT,
2202 [VCS2] = GEN8_VCS2_IRQ_SHIFT,
2203 [VECS] = GEN8_VECS_IRQ_SHIFT,
2206 shift = irq_shifts[engine->id];
2209 engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT << shift;
2210 engine->irq_keep_mask = GT_CONTEXT_SWITCH_INTERRUPT << shift;
2214 logical_ring_setup(struct intel_engine_cs *engine)
2216 intel_engine_setup_common(engine);
2218 /* Intentionally left blank. */
2219 engine->buffer = NULL;
2221 tasklet_init(&engine->execlists.tasklet,
2222 execlists_submission_tasklet, (unsigned long)engine);
2224 logical_ring_default_vfuncs(engine);
2225 logical_ring_default_irqs(engine);
2228 static int logical_ring_init(struct intel_engine_cs *engine)
2230 struct drm_i915_private *i915 = engine->i915;
2231 struct intel_engine_execlists * const execlists = &engine->execlists;
2234 ret = intel_engine_init_common(engine);
2238 if (HAS_LOGICAL_RING_ELSQ(i915)) {
2239 execlists->submit_reg = i915->regs +
2240 i915_mmio_reg_offset(RING_EXECLIST_SQ_CONTENTS(engine));
2241 execlists->ctrl_reg = i915->regs +
2242 i915_mmio_reg_offset(RING_EXECLIST_CONTROL(engine));
2244 execlists->submit_reg = i915->regs +
2245 i915_mmio_reg_offset(RING_ELSP(engine));
2248 execlists->preempt_complete_status = ~0u;
2249 if (i915->preempt_context) {
2250 struct intel_context *ce =
2251 to_intel_context(i915->preempt_context, engine);
2253 execlists->preempt_complete_status =
2254 upper_32_bits(ce->lrc_desc);
2257 execlists->csb_status =
2258 &engine->status_page.page_addr[I915_HWS_CSB_BUF0_INDEX];
2260 execlists->csb_write =
2261 &engine->status_page.page_addr[intel_hws_csb_write_index(i915)];
2263 reset_csb_pointers(execlists);
2268 int logical_render_ring_init(struct intel_engine_cs *engine)
2270 struct drm_i915_private *dev_priv = engine->i915;
2273 logical_ring_setup(engine);
2275 if (HAS_L3_DPF(dev_priv))
2276 engine->irq_keep_mask |= GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
2278 /* Override some for render ring. */
2279 engine->init_context = gen8_init_rcs_context;
2280 engine->emit_flush = gen8_emit_flush_render;
2281 engine->emit_breadcrumb = gen8_emit_breadcrumb_rcs;
2282 engine->emit_breadcrumb_sz = gen8_emit_breadcrumb_rcs_sz;
2284 ret = logical_ring_init(engine);
2288 ret = intel_init_workaround_bb(engine);
2291 * We continue even if we fail to initialize WA batch
2292 * because we only expect rare glitches but nothing
2293 * critical to prevent us from using GPU
2295 DRM_ERROR("WA batch buffer initialization failed: %d\n",
2299 intel_engine_init_whitelist(engine);
2300 intel_engine_init_workarounds(engine);
2305 int logical_xcs_ring_init(struct intel_engine_cs *engine)
2307 logical_ring_setup(engine);
2309 return logical_ring_init(engine);
2313 make_rpcs(struct drm_i915_private *dev_priv)
2315 bool subslice_pg = INTEL_INFO(dev_priv)->sseu.has_subslice_pg;
2316 u8 slices = hweight8(INTEL_INFO(dev_priv)->sseu.slice_mask);
2317 u8 subslices = hweight8(INTEL_INFO(dev_priv)->sseu.subslice_mask[0]);
2321 * No explicit RPCS request is needed to ensure full
2322 * slice/subslice/EU enablement prior to Gen9.
2324 if (INTEL_GEN(dev_priv) < 9)
2328 * Since the SScount bitfield in GEN8_R_PWR_CLK_STATE is only three bits
2329 * wide and Icelake has up to eight subslices, specfial programming is
2330 * needed in order to correctly enable all subslices.
2332 * According to documentation software must consider the configuration
2333 * as 2x4x8 and hardware will translate this to 1x8x8.
2335 * Furthemore, even though SScount is three bits, maximum documented
2336 * value for it is four. From this some rules/restrictions follow:
2339 * If enabled subslice count is greater than four, two whole slices must
2340 * be enabled instead.
2343 * When more than one slice is enabled, hardware ignores the subslice
2346 * From these restrictions it follows that it is not possible to enable
2347 * a count of subslices between the SScount maximum of four restriction,
2348 * and the maximum available number on a particular SKU. Either all
2349 * subslices are enabled, or a count between one and four on the first
2352 if (IS_GEN(dev_priv, 11) && slices == 1 && subslices >= 4) {
2353 GEM_BUG_ON(subslices & 1);
2355 subslice_pg = false;
2360 * Starting in Gen9, render power gating can leave
2361 * slice/subslice/EU in a partially enabled state. We
2362 * must make an explicit request through RPCS for full
2365 if (INTEL_INFO(dev_priv)->sseu.has_slice_pg) {
2366 u32 mask, val = slices;
2368 if (INTEL_GEN(dev_priv) >= 11) {
2369 mask = GEN11_RPCS_S_CNT_MASK;
2370 val <<= GEN11_RPCS_S_CNT_SHIFT;
2372 mask = GEN8_RPCS_S_CNT_MASK;
2373 val <<= GEN8_RPCS_S_CNT_SHIFT;
2376 GEM_BUG_ON(val & ~mask);
2379 rpcs |= GEN8_RPCS_ENABLE | GEN8_RPCS_S_CNT_ENABLE | val;
2383 u32 val = subslices;
2385 val <<= GEN8_RPCS_SS_CNT_SHIFT;
2387 GEM_BUG_ON(val & ~GEN8_RPCS_SS_CNT_MASK);
2388 val &= GEN8_RPCS_SS_CNT_MASK;
2390 rpcs |= GEN8_RPCS_ENABLE | GEN8_RPCS_SS_CNT_ENABLE | val;
2393 if (INTEL_INFO(dev_priv)->sseu.has_eu_pg) {
2396 val = INTEL_INFO(dev_priv)->sseu.eu_per_subslice <<
2397 GEN8_RPCS_EU_MIN_SHIFT;
2398 GEM_BUG_ON(val & ~GEN8_RPCS_EU_MIN_MASK);
2399 val &= GEN8_RPCS_EU_MIN_MASK;
2403 val = INTEL_INFO(dev_priv)->sseu.eu_per_subslice <<
2404 GEN8_RPCS_EU_MAX_SHIFT;
2405 GEM_BUG_ON(val & ~GEN8_RPCS_EU_MAX_MASK);
2406 val &= GEN8_RPCS_EU_MAX_MASK;
2410 rpcs |= GEN8_RPCS_ENABLE;
2416 static u32 intel_lr_indirect_ctx_offset(struct intel_engine_cs *engine)
2418 u32 indirect_ctx_offset;
2420 switch (INTEL_GEN(engine->i915)) {
2422 MISSING_CASE(INTEL_GEN(engine->i915));
2425 indirect_ctx_offset =
2426 GEN11_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2429 indirect_ctx_offset =
2430 GEN10_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2433 indirect_ctx_offset =
2434 GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2437 indirect_ctx_offset =
2438 GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2442 return indirect_ctx_offset;
2445 static void execlists_init_reg_state(u32 *regs,
2446 struct i915_gem_context *ctx,
2447 struct intel_engine_cs *engine,
2448 struct intel_ring *ring)
2450 struct drm_i915_private *dev_priv = engine->i915;
2451 u32 base = engine->mmio_base;
2452 bool rcs = engine->class == RENDER_CLASS;
2454 /* A context is actually a big batch buffer with several
2455 * MI_LOAD_REGISTER_IMM commands followed by (reg, value) pairs. The
2456 * values we are setting here are only for the first context restore:
2457 * on a subsequent save, the GPU will recreate this batchbuffer with new
2458 * values (including all the missing MI_LOAD_REGISTER_IMM commands that
2459 * we are not initializing here).
2461 regs[CTX_LRI_HEADER_0] = MI_LOAD_REGISTER_IMM(rcs ? 14 : 11) |
2462 MI_LRI_FORCE_POSTED;
2464 CTX_REG(regs, CTX_CONTEXT_CONTROL, RING_CONTEXT_CONTROL(engine),
2465 _MASKED_BIT_DISABLE(CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT) |
2466 _MASKED_BIT_ENABLE(CTX_CTRL_INHIBIT_SYN_CTX_SWITCH));
2467 if (INTEL_GEN(dev_priv) < 11) {
2468 regs[CTX_CONTEXT_CONTROL + 1] |=
2469 _MASKED_BIT_DISABLE(CTX_CTRL_ENGINE_CTX_SAVE_INHIBIT |
2470 CTX_CTRL_RS_CTX_ENABLE);
2472 CTX_REG(regs, CTX_RING_HEAD, RING_HEAD(base), 0);
2473 CTX_REG(regs, CTX_RING_TAIL, RING_TAIL(base), 0);
2474 CTX_REG(regs, CTX_RING_BUFFER_START, RING_START(base), 0);
2475 CTX_REG(regs, CTX_RING_BUFFER_CONTROL, RING_CTL(base),
2476 RING_CTL_SIZE(ring->size) | RING_VALID);
2477 CTX_REG(regs, CTX_BB_HEAD_U, RING_BBADDR_UDW(base), 0);
2478 CTX_REG(regs, CTX_BB_HEAD_L, RING_BBADDR(base), 0);
2479 CTX_REG(regs, CTX_BB_STATE, RING_BBSTATE(base), RING_BB_PPGTT);
2480 CTX_REG(regs, CTX_SECOND_BB_HEAD_U, RING_SBBADDR_UDW(base), 0);
2481 CTX_REG(regs, CTX_SECOND_BB_HEAD_L, RING_SBBADDR(base), 0);
2482 CTX_REG(regs, CTX_SECOND_BB_STATE, RING_SBBSTATE(base), 0);
2484 struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
2486 CTX_REG(regs, CTX_RCS_INDIRECT_CTX, RING_INDIRECT_CTX(base), 0);
2487 CTX_REG(regs, CTX_RCS_INDIRECT_CTX_OFFSET,
2488 RING_INDIRECT_CTX_OFFSET(base), 0);
2489 if (wa_ctx->indirect_ctx.size) {
2490 u32 ggtt_offset = i915_ggtt_offset(wa_ctx->vma);
2492 regs[CTX_RCS_INDIRECT_CTX + 1] =
2493 (ggtt_offset + wa_ctx->indirect_ctx.offset) |
2494 (wa_ctx->indirect_ctx.size / CACHELINE_BYTES);
2496 regs[CTX_RCS_INDIRECT_CTX_OFFSET + 1] =
2497 intel_lr_indirect_ctx_offset(engine) << 6;
2500 CTX_REG(regs, CTX_BB_PER_CTX_PTR, RING_BB_PER_CTX_PTR(base), 0);
2501 if (wa_ctx->per_ctx.size) {
2502 u32 ggtt_offset = i915_ggtt_offset(wa_ctx->vma);
2504 regs[CTX_BB_PER_CTX_PTR + 1] =
2505 (ggtt_offset + wa_ctx->per_ctx.offset) | 0x01;
2509 regs[CTX_LRI_HEADER_1] = MI_LOAD_REGISTER_IMM(9) | MI_LRI_FORCE_POSTED;
2511 CTX_REG(regs, CTX_CTX_TIMESTAMP, RING_CTX_TIMESTAMP(base), 0);
2512 /* PDP values well be assigned later if needed */
2513 CTX_REG(regs, CTX_PDP3_UDW, GEN8_RING_PDP_UDW(engine, 3), 0);
2514 CTX_REG(regs, CTX_PDP3_LDW, GEN8_RING_PDP_LDW(engine, 3), 0);
2515 CTX_REG(regs, CTX_PDP2_UDW, GEN8_RING_PDP_UDW(engine, 2), 0);
2516 CTX_REG(regs, CTX_PDP2_LDW, GEN8_RING_PDP_LDW(engine, 2), 0);
2517 CTX_REG(regs, CTX_PDP1_UDW, GEN8_RING_PDP_UDW(engine, 1), 0);
2518 CTX_REG(regs, CTX_PDP1_LDW, GEN8_RING_PDP_LDW(engine, 1), 0);
2519 CTX_REG(regs, CTX_PDP0_UDW, GEN8_RING_PDP_UDW(engine, 0), 0);
2520 CTX_REG(regs, CTX_PDP0_LDW, GEN8_RING_PDP_LDW(engine, 0), 0);
2522 if (i915_vm_is_48bit(&ctx->ppgtt->vm)) {
2523 /* 64b PPGTT (48bit canonical)
2524 * PDP0_DESCRIPTOR contains the base address to PML4 and
2525 * other PDP Descriptors are ignored.
2527 ASSIGN_CTX_PML4(ctx->ppgtt, regs);
2529 ASSIGN_CTX_PDP(ctx->ppgtt, regs, 3);
2530 ASSIGN_CTX_PDP(ctx->ppgtt, regs, 2);
2531 ASSIGN_CTX_PDP(ctx->ppgtt, regs, 1);
2532 ASSIGN_CTX_PDP(ctx->ppgtt, regs, 0);
2536 regs[CTX_LRI_HEADER_2] = MI_LOAD_REGISTER_IMM(1);
2537 CTX_REG(regs, CTX_R_PWR_CLK_STATE, GEN8_R_PWR_CLK_STATE,
2538 make_rpcs(dev_priv));
2540 i915_oa_init_reg_state(engine, ctx, regs);
2543 regs[CTX_END] = MI_BATCH_BUFFER_END;
2544 if (INTEL_GEN(dev_priv) >= 10)
2545 regs[CTX_END] |= BIT(0);
2549 populate_lr_context(struct i915_gem_context *ctx,
2550 struct drm_i915_gem_object *ctx_obj,
2551 struct intel_engine_cs *engine,
2552 struct intel_ring *ring)
2558 ret = i915_gem_object_set_to_cpu_domain(ctx_obj, true);
2560 DRM_DEBUG_DRIVER("Could not set to CPU domain\n");
2564 vaddr = i915_gem_object_pin_map(ctx_obj, I915_MAP_WB);
2565 if (IS_ERR(vaddr)) {
2566 ret = PTR_ERR(vaddr);
2567 DRM_DEBUG_DRIVER("Could not map object pages! (%d)\n", ret);
2570 ctx_obj->mm.dirty = true;
2572 if (engine->default_state) {
2574 * We only want to copy over the template context state;
2575 * skipping over the headers reserved for GuC communication,
2576 * leaving those as zero.
2578 const unsigned long start = LRC_HEADER_PAGES * PAGE_SIZE;
2581 defaults = i915_gem_object_pin_map(engine->default_state,
2583 if (IS_ERR(defaults)) {
2584 ret = PTR_ERR(defaults);
2588 memcpy(vaddr + start, defaults + start, engine->context_size);
2589 i915_gem_object_unpin_map(engine->default_state);
2592 /* The second page of the context object contains some fields which must
2593 * be set up prior to the first execution. */
2594 regs = vaddr + LRC_STATE_PN * PAGE_SIZE;
2595 execlists_init_reg_state(regs, ctx, engine, ring);
2596 if (!engine->default_state)
2597 regs[CTX_CONTEXT_CONTROL + 1] |=
2598 _MASKED_BIT_ENABLE(CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT);
2599 if (ctx == ctx->i915->preempt_context && INTEL_GEN(engine->i915) < 11)
2600 regs[CTX_CONTEXT_CONTROL + 1] |=
2601 _MASKED_BIT_ENABLE(CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT |
2602 CTX_CTRL_ENGINE_CTX_SAVE_INHIBIT);
2605 i915_gem_object_unpin_map(ctx_obj);
2609 static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
2610 struct intel_engine_cs *engine,
2611 struct intel_context *ce)
2613 struct drm_i915_gem_object *ctx_obj;
2614 struct i915_vma *vma;
2615 uint32_t context_size;
2616 struct intel_ring *ring;
2617 struct i915_timeline *timeline;
2623 context_size = round_up(engine->context_size, I915_GTT_PAGE_SIZE);
2626 * Before the actual start of the context image, we insert a few pages
2627 * for our own use and for sharing with the GuC.
2629 context_size += LRC_HEADER_PAGES * PAGE_SIZE;
2631 ctx_obj = i915_gem_object_create(ctx->i915, context_size);
2632 if (IS_ERR(ctx_obj))
2633 return PTR_ERR(ctx_obj);
2635 vma = i915_vma_instance(ctx_obj, &ctx->i915->ggtt.vm, NULL);
2638 goto error_deref_obj;
2641 timeline = i915_timeline_create(ctx->i915, ctx->name);
2642 if (IS_ERR(timeline)) {
2643 ret = PTR_ERR(timeline);
2644 goto error_deref_obj;
2647 ring = intel_engine_create_ring(engine, timeline, ctx->ring_size);
2648 i915_timeline_put(timeline);
2650 ret = PTR_ERR(ring);
2651 goto error_deref_obj;
2654 ret = populate_lr_context(ctx, ctx_obj, engine, ring);
2656 DRM_DEBUG_DRIVER("Failed to populate LRC: %d\n", ret);
2657 goto error_ring_free;
2666 intel_ring_free(ring);
2668 i915_gem_object_put(ctx_obj);
2672 void intel_lr_context_resume(struct drm_i915_private *i915)
2674 struct intel_engine_cs *engine;
2675 struct i915_gem_context *ctx;
2676 enum intel_engine_id id;
2679 * Because we emit WA_TAIL_DWORDS there may be a disparity
2680 * between our bookkeeping in ce->ring->head and ce->ring->tail and
2681 * that stored in context. As we only write new commands from
2682 * ce->ring->tail onwards, everything before that is junk. If the GPU
2683 * starts reading from its RING_HEAD from the context, it may try to
2684 * execute that junk and die.
2686 * So to avoid that we reset the context images upon resume. For
2687 * simplicity, we just zero everything out.
2689 list_for_each_entry(ctx, &i915->contexts.list, link) {
2690 for_each_engine(engine, i915, id) {
2691 struct intel_context *ce =
2692 to_intel_context(ctx, engine);
2697 intel_ring_reset(ce->ring, 0);
2699 if (ce->pin_count) { /* otherwise done in context_pin */
2700 u32 *regs = ce->lrc_reg_state;
2702 regs[CTX_RING_HEAD + 1] = ce->ring->head;
2703 regs[CTX_RING_TAIL + 1] = ce->ring->tail;
2709 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
2710 #include "selftests/intel_lrc.c"