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).
135 #include <drm/drmP.h>
136 #include <drm/i915_drm.h>
137 #include "i915_drv.h"
139 #define GEN9_LR_CONTEXT_RENDER_SIZE (22 * PAGE_SIZE)
140 #define GEN8_LR_CONTEXT_RENDER_SIZE (20 * PAGE_SIZE)
141 #define GEN8_LR_CONTEXT_OTHER_SIZE (2 * PAGE_SIZE)
143 #define RING_EXECLIST_QFULL (1 << 0x2)
144 #define RING_EXECLIST1_VALID (1 << 0x3)
145 #define RING_EXECLIST0_VALID (1 << 0x4)
146 #define RING_EXECLIST_ACTIVE_STATUS (3 << 0xE)
147 #define RING_EXECLIST1_ACTIVE (1 << 0x11)
148 #define RING_EXECLIST0_ACTIVE (1 << 0x12)
150 #define GEN8_CTX_STATUS_IDLE_ACTIVE (1 << 0)
151 #define GEN8_CTX_STATUS_PREEMPTED (1 << 1)
152 #define GEN8_CTX_STATUS_ELEMENT_SWITCH (1 << 2)
153 #define GEN8_CTX_STATUS_ACTIVE_IDLE (1 << 3)
154 #define GEN8_CTX_STATUS_COMPLETE (1 << 4)
155 #define GEN8_CTX_STATUS_LITE_RESTORE (1 << 15)
157 #define CTX_LRI_HEADER_0 0x01
158 #define CTX_CONTEXT_CONTROL 0x02
159 #define CTX_RING_HEAD 0x04
160 #define CTX_RING_TAIL 0x06
161 #define CTX_RING_BUFFER_START 0x08
162 #define CTX_RING_BUFFER_CONTROL 0x0a
163 #define CTX_BB_HEAD_U 0x0c
164 #define CTX_BB_HEAD_L 0x0e
165 #define CTX_BB_STATE 0x10
166 #define CTX_SECOND_BB_HEAD_U 0x12
167 #define CTX_SECOND_BB_HEAD_L 0x14
168 #define CTX_SECOND_BB_STATE 0x16
169 #define CTX_BB_PER_CTX_PTR 0x18
170 #define CTX_RCS_INDIRECT_CTX 0x1a
171 #define CTX_RCS_INDIRECT_CTX_OFFSET 0x1c
172 #define CTX_LRI_HEADER_1 0x21
173 #define CTX_CTX_TIMESTAMP 0x22
174 #define CTX_PDP3_UDW 0x24
175 #define CTX_PDP3_LDW 0x26
176 #define CTX_PDP2_UDW 0x28
177 #define CTX_PDP2_LDW 0x2a
178 #define CTX_PDP1_UDW 0x2c
179 #define CTX_PDP1_LDW 0x2e
180 #define CTX_PDP0_UDW 0x30
181 #define CTX_PDP0_LDW 0x32
182 #define CTX_LRI_HEADER_2 0x41
183 #define CTX_R_PWR_CLK_STATE 0x42
184 #define CTX_GPGPU_CSR_BASE_ADDRESS 0x44
186 #define GEN8_CTX_VALID (1<<0)
187 #define GEN8_CTX_FORCE_PD_RESTORE (1<<1)
188 #define GEN8_CTX_FORCE_RESTORE (1<<2)
189 #define GEN8_CTX_L3LLC_COHERENT (1<<5)
190 #define GEN8_CTX_PRIVILEGE (1<<8)
192 #define ASSIGN_CTX_PDP(ppgtt, reg_state, n) { \
193 const u64 _addr = test_bit(n, ppgtt->pdp.used_pdpes) ? \
194 ppgtt->pdp.page_directory[n]->daddr : \
195 ppgtt->scratch_pd->daddr; \
196 reg_state[CTX_PDP ## n ## _UDW+1] = upper_32_bits(_addr); \
197 reg_state[CTX_PDP ## n ## _LDW+1] = lower_32_bits(_addr); \
201 ADVANCED_CONTEXT
= 0,
206 #define GEN8_CTX_MODE_SHIFT 3
209 FAULT_AND_HALT
, /* Debug only */
211 FAULT_AND_CONTINUE
/* Unsupported */
213 #define GEN8_CTX_ID_SHIFT 32
214 #define CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT 0x17
216 static int intel_lr_context_pin(struct intel_engine_cs
*ring
,
217 struct intel_context
*ctx
);
220 * intel_sanitize_enable_execlists() - sanitize i915.enable_execlists
222 * @enable_execlists: value of i915.enable_execlists module parameter.
224 * Only certain platforms support Execlists (the prerequisites being
225 * support for Logical Ring Contexts and Aliasing PPGTT or better).
227 * Return: 1 if Execlists is supported and has to be enabled.
229 int intel_sanitize_enable_execlists(struct drm_device
*dev
, int enable_execlists
)
231 WARN_ON(i915
.enable_ppgtt
== -1);
233 if (INTEL_INFO(dev
)->gen
>= 9)
236 if (enable_execlists
== 0)
239 if (HAS_LOGICAL_RING_CONTEXTS(dev
) && USES_PPGTT(dev
) &&
240 i915
.use_mmio_flip
>= 0)
247 * intel_execlists_ctx_id() - get the Execlists Context ID
248 * @ctx_obj: Logical Ring Context backing object.
250 * Do not confuse with ctx->id! Unfortunately we have a name overload
251 * here: the old context ID we pass to userspace as a handler so that
252 * they can refer to a context, and the new context ID we pass to the
253 * ELSP so that the GPU can inform us of the context status via
256 * Return: 20-bits globally unique context ID.
258 u32
intel_execlists_ctx_id(struct drm_i915_gem_object
*ctx_obj
)
260 u32 lrca
= i915_gem_obj_ggtt_offset(ctx_obj
);
262 /* LRCA is required to be 4K aligned so the more significant 20 bits
263 * are globally unique */
267 static uint64_t execlists_ctx_descriptor(struct intel_engine_cs
*ring
,
268 struct drm_i915_gem_object
*ctx_obj
)
270 struct drm_device
*dev
= ring
->dev
;
272 uint64_t lrca
= i915_gem_obj_ggtt_offset(ctx_obj
);
274 WARN_ON(lrca
& 0xFFFFFFFF00000FFFULL
);
276 desc
= GEN8_CTX_VALID
;
277 desc
|= LEGACY_CONTEXT
<< GEN8_CTX_MODE_SHIFT
;
278 if (IS_GEN8(ctx_obj
->base
.dev
))
279 desc
|= GEN8_CTX_L3LLC_COHERENT
;
280 desc
|= GEN8_CTX_PRIVILEGE
;
282 desc
|= (u64
)intel_execlists_ctx_id(ctx_obj
) << GEN8_CTX_ID_SHIFT
;
284 /* TODO: WaDisableLiteRestore when we start using semaphore
285 * signalling between Command Streamers */
286 /* desc |= GEN8_CTX_FORCE_RESTORE; */
288 /* WaEnableForceRestoreInCtxtDescForVCS:skl */
290 INTEL_REVID(dev
) <= SKL_REVID_B0
&&
291 (ring
->id
== BCS
|| ring
->id
== VCS
||
292 ring
->id
== VECS
|| ring
->id
== VCS2
))
293 desc
|= GEN8_CTX_FORCE_RESTORE
;
298 static void execlists_elsp_write(struct intel_engine_cs
*ring
,
299 struct drm_i915_gem_object
*ctx_obj0
,
300 struct drm_i915_gem_object
*ctx_obj1
)
302 struct drm_device
*dev
= ring
->dev
;
303 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
307 /* XXX: You must always write both descriptors in the order below. */
309 temp
= execlists_ctx_descriptor(ring
, ctx_obj1
);
312 desc
[1] = (u32
)(temp
>> 32);
315 temp
= execlists_ctx_descriptor(ring
, ctx_obj0
);
316 desc
[3] = (u32
)(temp
>> 32);
319 spin_lock(&dev_priv
->uncore
.lock
);
320 intel_uncore_forcewake_get__locked(dev_priv
, FORCEWAKE_ALL
);
321 I915_WRITE_FW(RING_ELSP(ring
), desc
[1]);
322 I915_WRITE_FW(RING_ELSP(ring
), desc
[0]);
323 I915_WRITE_FW(RING_ELSP(ring
), desc
[3]);
325 /* The context is automatically loaded after the following */
326 I915_WRITE_FW(RING_ELSP(ring
), desc
[2]);
328 /* ELSP is a wo register, so use another nearby reg for posting instead */
329 POSTING_READ_FW(RING_EXECLIST_STATUS(ring
));
330 intel_uncore_forcewake_put__locked(dev_priv
, FORCEWAKE_ALL
);
331 spin_unlock(&dev_priv
->uncore
.lock
);
334 static int execlists_update_context(struct drm_i915_gem_object
*ctx_obj
,
335 struct drm_i915_gem_object
*ring_obj
,
336 struct i915_hw_ppgtt
*ppgtt
,
342 page
= i915_gem_object_get_page(ctx_obj
, 1);
343 reg_state
= kmap_atomic(page
);
345 reg_state
[CTX_RING_TAIL
+1] = tail
;
346 reg_state
[CTX_RING_BUFFER_START
+1] = i915_gem_obj_ggtt_offset(ring_obj
);
348 /* True PPGTT with dynamic page allocation: update PDP registers and
349 * point the unallocated PDPs to the scratch page
352 ASSIGN_CTX_PDP(ppgtt
, reg_state
, 3);
353 ASSIGN_CTX_PDP(ppgtt
, reg_state
, 2);
354 ASSIGN_CTX_PDP(ppgtt
, reg_state
, 1);
355 ASSIGN_CTX_PDP(ppgtt
, reg_state
, 0);
358 kunmap_atomic(reg_state
);
363 static void execlists_submit_contexts(struct intel_engine_cs
*ring
,
364 struct intel_context
*to0
, u32 tail0
,
365 struct intel_context
*to1
, u32 tail1
)
367 struct drm_i915_gem_object
*ctx_obj0
= to0
->engine
[ring
->id
].state
;
368 struct intel_ringbuffer
*ringbuf0
= to0
->engine
[ring
->id
].ringbuf
;
369 struct drm_i915_gem_object
*ctx_obj1
= NULL
;
370 struct intel_ringbuffer
*ringbuf1
= NULL
;
373 WARN_ON(!i915_gem_obj_is_pinned(ctx_obj0
));
374 WARN_ON(!i915_gem_obj_is_pinned(ringbuf0
->obj
));
376 execlists_update_context(ctx_obj0
, ringbuf0
->obj
, to0
->ppgtt
, tail0
);
379 ringbuf1
= to1
->engine
[ring
->id
].ringbuf
;
380 ctx_obj1
= to1
->engine
[ring
->id
].state
;
382 WARN_ON(!i915_gem_obj_is_pinned(ctx_obj1
));
383 WARN_ON(!i915_gem_obj_is_pinned(ringbuf1
->obj
));
385 execlists_update_context(ctx_obj1
, ringbuf1
->obj
, to1
->ppgtt
, tail1
);
388 execlists_elsp_write(ring
, ctx_obj0
, ctx_obj1
);
391 static void execlists_context_unqueue(struct intel_engine_cs
*ring
)
393 struct drm_i915_gem_request
*req0
= NULL
, *req1
= NULL
;
394 struct drm_i915_gem_request
*cursor
= NULL
, *tmp
= NULL
;
396 assert_spin_locked(&ring
->execlist_lock
);
399 * If irqs are not active generate a warning as batches that finish
400 * without the irqs may get lost and a GPU Hang may occur.
402 WARN_ON(!intel_irqs_enabled(ring
->dev
->dev_private
));
404 if (list_empty(&ring
->execlist_queue
))
407 /* Try to read in pairs */
408 list_for_each_entry_safe(cursor
, tmp
, &ring
->execlist_queue
,
412 } else if (req0
->ctx
== cursor
->ctx
) {
413 /* Same ctx: ignore first request, as second request
414 * will update tail past first request's workload */
415 cursor
->elsp_submitted
= req0
->elsp_submitted
;
416 list_del(&req0
->execlist_link
);
417 list_add_tail(&req0
->execlist_link
,
418 &ring
->execlist_retired_req_list
);
426 if (IS_GEN8(ring
->dev
) || IS_GEN9(ring
->dev
)) {
428 * WaIdleLiteRestore: make sure we never cause a lite
429 * restore with HEAD==TAIL
431 if (req0
->elsp_submitted
) {
433 * Apply the wa NOOPS to prevent ring:HEAD == req:TAIL
434 * as we resubmit the request. See gen8_emit_request()
435 * for where we prepare the padding after the end of the
438 struct intel_ringbuffer
*ringbuf
;
440 ringbuf
= req0
->ctx
->engine
[ring
->id
].ringbuf
;
442 req0
->tail
&= ringbuf
->size
- 1;
446 WARN_ON(req1
&& req1
->elsp_submitted
);
448 execlists_submit_contexts(ring
, req0
->ctx
, req0
->tail
,
449 req1
? req1
->ctx
: NULL
,
450 req1
? req1
->tail
: 0);
452 req0
->elsp_submitted
++;
454 req1
->elsp_submitted
++;
457 static bool execlists_check_remove_request(struct intel_engine_cs
*ring
,
460 struct drm_i915_gem_request
*head_req
;
462 assert_spin_locked(&ring
->execlist_lock
);
464 head_req
= list_first_entry_or_null(&ring
->execlist_queue
,
465 struct drm_i915_gem_request
,
468 if (head_req
!= NULL
) {
469 struct drm_i915_gem_object
*ctx_obj
=
470 head_req
->ctx
->engine
[ring
->id
].state
;
471 if (intel_execlists_ctx_id(ctx_obj
) == request_id
) {
472 WARN(head_req
->elsp_submitted
== 0,
473 "Never submitted head request\n");
475 if (--head_req
->elsp_submitted
<= 0) {
476 list_del(&head_req
->execlist_link
);
477 list_add_tail(&head_req
->execlist_link
,
478 &ring
->execlist_retired_req_list
);
488 * intel_lrc_irq_handler() - handle Context Switch interrupts
489 * @ring: Engine Command Streamer to handle.
491 * Check the unread Context Status Buffers and manage the submission of new
492 * contexts to the ELSP accordingly.
494 void intel_lrc_irq_handler(struct intel_engine_cs
*ring
)
496 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
502 u32 submit_contexts
= 0;
504 status_pointer
= I915_READ(RING_CONTEXT_STATUS_PTR(ring
));
506 read_pointer
= ring
->next_context_status_buffer
;
507 write_pointer
= status_pointer
& 0x07;
508 if (read_pointer
> write_pointer
)
511 spin_lock(&ring
->execlist_lock
);
513 while (read_pointer
< write_pointer
) {
515 status
= I915_READ(RING_CONTEXT_STATUS_BUF(ring
) +
516 (read_pointer
% 6) * 8);
517 status_id
= I915_READ(RING_CONTEXT_STATUS_BUF(ring
) +
518 (read_pointer
% 6) * 8 + 4);
520 if (status
& GEN8_CTX_STATUS_PREEMPTED
) {
521 if (status
& GEN8_CTX_STATUS_LITE_RESTORE
) {
522 if (execlists_check_remove_request(ring
, status_id
))
523 WARN(1, "Lite Restored request removed from queue\n");
525 WARN(1, "Preemption without Lite Restore\n");
528 if ((status
& GEN8_CTX_STATUS_ACTIVE_IDLE
) ||
529 (status
& GEN8_CTX_STATUS_ELEMENT_SWITCH
)) {
530 if (execlists_check_remove_request(ring
, status_id
))
535 if (submit_contexts
!= 0)
536 execlists_context_unqueue(ring
);
538 spin_unlock(&ring
->execlist_lock
);
540 WARN(submit_contexts
> 2, "More than two context complete events?\n");
541 ring
->next_context_status_buffer
= write_pointer
% 6;
543 I915_WRITE(RING_CONTEXT_STATUS_PTR(ring
),
544 ((u32
)ring
->next_context_status_buffer
& 0x07) << 8);
547 static int execlists_context_queue(struct intel_engine_cs
*ring
,
548 struct intel_context
*to
,
550 struct drm_i915_gem_request
*request
)
552 struct drm_i915_gem_request
*cursor
;
553 int num_elements
= 0;
555 if (to
!= ring
->default_context
)
556 intel_lr_context_pin(ring
, to
);
560 * If there isn't a request associated with this submission,
561 * create one as a temporary holder.
563 request
= kzalloc(sizeof(*request
), GFP_KERNEL
);
566 request
->ring
= ring
;
568 kref_init(&request
->ref
);
569 i915_gem_context_reference(request
->ctx
);
571 i915_gem_request_reference(request
);
572 WARN_ON(to
!= request
->ctx
);
574 request
->tail
= tail
;
576 spin_lock_irq(&ring
->execlist_lock
);
578 list_for_each_entry(cursor
, &ring
->execlist_queue
, execlist_link
)
579 if (++num_elements
> 2)
582 if (num_elements
> 2) {
583 struct drm_i915_gem_request
*tail_req
;
585 tail_req
= list_last_entry(&ring
->execlist_queue
,
586 struct drm_i915_gem_request
,
589 if (to
== tail_req
->ctx
) {
590 WARN(tail_req
->elsp_submitted
!= 0,
591 "More than 2 already-submitted reqs queued\n");
592 list_del(&tail_req
->execlist_link
);
593 list_add_tail(&tail_req
->execlist_link
,
594 &ring
->execlist_retired_req_list
);
598 list_add_tail(&request
->execlist_link
, &ring
->execlist_queue
);
599 if (num_elements
== 0)
600 execlists_context_unqueue(ring
);
602 spin_unlock_irq(&ring
->execlist_lock
);
607 static int logical_ring_invalidate_all_caches(struct intel_ringbuffer
*ringbuf
,
608 struct intel_context
*ctx
)
610 struct intel_engine_cs
*ring
= ringbuf
->ring
;
611 uint32_t flush_domains
;
615 if (ring
->gpu_caches_dirty
)
616 flush_domains
= I915_GEM_GPU_DOMAINS
;
618 ret
= ring
->emit_flush(ringbuf
, ctx
,
619 I915_GEM_GPU_DOMAINS
, flush_domains
);
623 ring
->gpu_caches_dirty
= false;
627 static int execlists_move_to_gpu(struct intel_ringbuffer
*ringbuf
,
628 struct intel_context
*ctx
,
629 struct list_head
*vmas
)
631 struct intel_engine_cs
*ring
= ringbuf
->ring
;
632 const unsigned other_rings
= ~intel_ring_flag(ring
);
633 struct i915_vma
*vma
;
634 uint32_t flush_domains
= 0;
635 bool flush_chipset
= false;
638 list_for_each_entry(vma
, vmas
, exec_list
) {
639 struct drm_i915_gem_object
*obj
= vma
->obj
;
641 if (obj
->active
& other_rings
) {
642 ret
= i915_gem_object_sync(obj
, ring
);
647 if (obj
->base
.write_domain
& I915_GEM_DOMAIN_CPU
)
648 flush_chipset
|= i915_gem_clflush_object(obj
, false);
650 flush_domains
|= obj
->base
.write_domain
;
653 if (flush_domains
& I915_GEM_DOMAIN_GTT
)
656 /* Unconditionally invalidate gpu caches and ensure that we do flush
657 * any residual writes from the previous batch.
659 return logical_ring_invalidate_all_caches(ringbuf
, ctx
);
662 int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request
*request
,
663 struct intel_context
*ctx
)
667 if (ctx
!= request
->ring
->default_context
) {
668 ret
= intel_lr_context_pin(request
->ring
, ctx
);
673 request
->ringbuf
= ctx
->engine
[request
->ring
->id
].ringbuf
;
675 i915_gem_context_reference(request
->ctx
);
680 static int logical_ring_wait_for_space(struct intel_ringbuffer
*ringbuf
,
681 struct intel_context
*ctx
,
684 struct intel_engine_cs
*ring
= ringbuf
->ring
;
685 struct drm_i915_gem_request
*request
;
689 /* The whole point of reserving space is to not wait! */
690 WARN_ON(ringbuf
->reserved_in_use
);
692 if (intel_ring_space(ringbuf
) >= bytes
)
695 list_for_each_entry(request
, &ring
->request_list
, list
) {
697 * The request queue is per-engine, so can contain requests
698 * from multiple ringbuffers. Here, we must ignore any that
699 * aren't from the ringbuffer we're considering.
701 if (request
->ringbuf
!= ringbuf
)
704 /* Would completion of this request free enough space? */
705 space
= __intel_ring_space(request
->postfix
, ringbuf
->tail
,
711 if (WARN_ON(&request
->list
== &ring
->request_list
))
714 ret
= i915_wait_request(request
);
718 ringbuf
->space
= space
;
723 * intel_logical_ring_advance_and_submit() - advance the tail and submit the workload
724 * @ringbuf: Logical Ringbuffer to advance.
726 * The tail is updated in our logical ringbuffer struct, not in the actual context. What
727 * really happens during submission is that the context and current tail will be placed
728 * on a queue waiting for the ELSP to be ready to accept a new context submission. At that
729 * point, the tail *inside* the context is updated and the ELSP written to.
732 intel_logical_ring_advance_and_submit(struct intel_ringbuffer
*ringbuf
,
733 struct intel_context
*ctx
,
734 struct drm_i915_gem_request
*request
)
736 struct intel_engine_cs
*ring
= ringbuf
->ring
;
738 intel_logical_ring_advance(ringbuf
);
740 if (intel_ring_stopped(ring
))
743 execlists_context_queue(ring
, ctx
, ringbuf
->tail
, request
);
746 static int logical_ring_wrap_buffer(struct intel_ringbuffer
*ringbuf
,
747 struct intel_context
*ctx
)
749 uint32_t __iomem
*virt
;
750 int rem
= ringbuf
->size
- ringbuf
->tail
;
752 /* Can't wrap if space has already been reserved! */
753 WARN_ON(ringbuf
->reserved_in_use
);
755 if (ringbuf
->space
< rem
) {
756 int ret
= logical_ring_wait_for_space(ringbuf
, ctx
, rem
);
762 virt
= ringbuf
->virtual_start
+ ringbuf
->tail
;
765 iowrite32(MI_NOOP
, virt
++);
768 intel_ring_update_space(ringbuf
);
773 static int logical_ring_prepare(struct intel_ringbuffer
*ringbuf
,
774 struct intel_context
*ctx
, int bytes
)
779 * Add on the reserved size to the request to make sure that after
780 * the intended commands have been emitted, there is guaranteed to
781 * still be enough free space to send them to the hardware.
783 if (!ringbuf
->reserved_in_use
)
784 bytes
+= ringbuf
->reserved_size
;
786 if (unlikely(ringbuf
->tail
+ bytes
> ringbuf
->effective_size
)) {
787 ret
= logical_ring_wrap_buffer(ringbuf
, ctx
);
791 if(ringbuf
->reserved_size
) {
792 uint32_t size
= ringbuf
->reserved_size
;
794 intel_ring_reserved_space_cancel(ringbuf
);
795 intel_ring_reserved_space_reserve(ringbuf
, size
);
799 if (unlikely(ringbuf
->space
< bytes
)) {
800 ret
= logical_ring_wait_for_space(ringbuf
, ctx
, bytes
);
809 * intel_logical_ring_begin() - prepare the logical ringbuffer to accept some commands
811 * @ringbuf: Logical ringbuffer.
812 * @num_dwords: number of DWORDs that we plan to write to the ringbuffer.
814 * The ringbuffer might not be ready to accept the commands right away (maybe it needs to
815 * be wrapped, or wait a bit for the tail to be updated). This function takes care of that
816 * and also preallocates a request (every workload submission is still mediated through
817 * requests, same as it did with legacy ringbuffer submission).
819 * Return: non-zero if the ringbuffer is not ready to be written to.
821 static int intel_logical_ring_begin(struct intel_ringbuffer
*ringbuf
,
822 struct intel_context
*ctx
, int num_dwords
)
824 struct intel_engine_cs
*ring
= ringbuf
->ring
;
825 struct drm_device
*dev
= ring
->dev
;
826 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
829 ret
= i915_gem_check_wedge(&dev_priv
->gpu_error
,
830 dev_priv
->mm
.interruptible
);
834 ret
= logical_ring_prepare(ringbuf
, ctx
, num_dwords
* sizeof(uint32_t));
838 /* Preallocate the olr before touching the ring */
839 ret
= i915_gem_request_alloc(ring
, ctx
);
843 ringbuf
->space
-= num_dwords
* sizeof(uint32_t);
848 * execlists_submission() - submit a batchbuffer for execution, Execlists style
851 * @ring: Engine Command Streamer to submit to.
852 * @ctx: Context to employ for this submission.
853 * @args: execbuffer call arguments.
854 * @vmas: list of vmas.
855 * @batch_obj: the batchbuffer to submit.
856 * @exec_start: batchbuffer start virtual address pointer.
857 * @dispatch_flags: translated execbuffer call flags.
859 * This is the evil twin version of i915_gem_ringbuffer_submission. It abstracts
860 * away the submission details of the execbuffer ioctl call.
862 * Return: non-zero if the submission fails.
864 int intel_execlists_submission(struct drm_device
*dev
, struct drm_file
*file
,
865 struct intel_engine_cs
*ring
,
866 struct intel_context
*ctx
,
867 struct drm_i915_gem_execbuffer2
*args
,
868 struct list_head
*vmas
,
869 struct drm_i915_gem_object
*batch_obj
,
870 u64 exec_start
, u32 dispatch_flags
)
872 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
873 struct intel_ringbuffer
*ringbuf
= ctx
->engine
[ring
->id
].ringbuf
;
878 instp_mode
= args
->flags
& I915_EXEC_CONSTANTS_MASK
;
879 instp_mask
= I915_EXEC_CONSTANTS_MASK
;
880 switch (instp_mode
) {
881 case I915_EXEC_CONSTANTS_REL_GENERAL
:
882 case I915_EXEC_CONSTANTS_ABSOLUTE
:
883 case I915_EXEC_CONSTANTS_REL_SURFACE
:
884 if (instp_mode
!= 0 && ring
!= &dev_priv
->ring
[RCS
]) {
885 DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
889 if (instp_mode
!= dev_priv
->relative_constants_mode
) {
890 if (instp_mode
== I915_EXEC_CONSTANTS_REL_SURFACE
) {
891 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
895 /* The HW changed the meaning on this bit on gen6 */
896 instp_mask
&= ~I915_EXEC_CONSTANTS_REL_SURFACE
;
900 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode
);
904 if (args
->num_cliprects
!= 0) {
905 DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
908 if (args
->DR4
== 0xffffffff) {
909 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
913 if (args
->DR1
|| args
->DR4
|| args
->cliprects_ptr
) {
914 DRM_DEBUG("0 cliprects but dirt in cliprects fields\n");
919 if (args
->flags
& I915_EXEC_GEN7_SOL_RESET
) {
920 DRM_DEBUG("sol reset is gen7 only\n");
924 ret
= execlists_move_to_gpu(ringbuf
, ctx
, vmas
);
928 if (ring
== &dev_priv
->ring
[RCS
] &&
929 instp_mode
!= dev_priv
->relative_constants_mode
) {
930 ret
= intel_logical_ring_begin(ringbuf
, ctx
, 4);
934 intel_logical_ring_emit(ringbuf
, MI_NOOP
);
935 intel_logical_ring_emit(ringbuf
, MI_LOAD_REGISTER_IMM(1));
936 intel_logical_ring_emit(ringbuf
, INSTPM
);
937 intel_logical_ring_emit(ringbuf
, instp_mask
<< 16 | instp_mode
);
938 intel_logical_ring_advance(ringbuf
);
940 dev_priv
->relative_constants_mode
= instp_mode
;
943 ret
= ring
->emit_bb_start(ringbuf
, ctx
, exec_start
, dispatch_flags
);
947 trace_i915_gem_ring_dispatch(intel_ring_get_request(ring
), dispatch_flags
);
949 i915_gem_execbuffer_move_to_active(vmas
, ring
);
950 i915_gem_execbuffer_retire_commands(dev
, file
, ring
, batch_obj
);
955 void intel_execlists_retire_requests(struct intel_engine_cs
*ring
)
957 struct drm_i915_gem_request
*req
, *tmp
;
958 struct list_head retired_list
;
960 WARN_ON(!mutex_is_locked(&ring
->dev
->struct_mutex
));
961 if (list_empty(&ring
->execlist_retired_req_list
))
964 INIT_LIST_HEAD(&retired_list
);
965 spin_lock_irq(&ring
->execlist_lock
);
966 list_replace_init(&ring
->execlist_retired_req_list
, &retired_list
);
967 spin_unlock_irq(&ring
->execlist_lock
);
969 list_for_each_entry_safe(req
, tmp
, &retired_list
, execlist_link
) {
970 struct intel_context
*ctx
= req
->ctx
;
971 struct drm_i915_gem_object
*ctx_obj
=
972 ctx
->engine
[ring
->id
].state
;
974 if (ctx_obj
&& (ctx
!= ring
->default_context
))
975 intel_lr_context_unpin(ring
, ctx
);
976 list_del(&req
->execlist_link
);
977 i915_gem_request_unreference(req
);
981 void intel_logical_ring_stop(struct intel_engine_cs
*ring
)
983 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
986 if (!intel_ring_initialized(ring
))
989 ret
= intel_ring_idle(ring
);
990 if (ret
&& !i915_reset_in_progress(&to_i915(ring
->dev
)->gpu_error
))
991 DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",
994 /* TODO: Is this correct with Execlists enabled? */
995 I915_WRITE_MODE(ring
, _MASKED_BIT_ENABLE(STOP_RING
));
996 if (wait_for_atomic((I915_READ_MODE(ring
) & MODE_IDLE
) != 0, 1000)) {
997 DRM_ERROR("%s :timed out trying to stop ring\n", ring
->name
);
1000 I915_WRITE_MODE(ring
, _MASKED_BIT_DISABLE(STOP_RING
));
1003 int logical_ring_flush_all_caches(struct intel_ringbuffer
*ringbuf
,
1004 struct intel_context
*ctx
)
1006 struct intel_engine_cs
*ring
= ringbuf
->ring
;
1009 if (!ring
->gpu_caches_dirty
)
1012 ret
= ring
->emit_flush(ringbuf
, ctx
, 0, I915_GEM_GPU_DOMAINS
);
1016 ring
->gpu_caches_dirty
= false;
1020 static int intel_lr_context_pin(struct intel_engine_cs
*ring
,
1021 struct intel_context
*ctx
)
1023 struct drm_i915_gem_object
*ctx_obj
= ctx
->engine
[ring
->id
].state
;
1024 struct intel_ringbuffer
*ringbuf
= ctx
->engine
[ring
->id
].ringbuf
;
1027 WARN_ON(!mutex_is_locked(&ring
->dev
->struct_mutex
));
1028 if (ctx
->engine
[ring
->id
].pin_count
++ == 0) {
1029 ret
= i915_gem_obj_ggtt_pin(ctx_obj
,
1030 GEN8_LR_CONTEXT_ALIGN
, 0);
1032 goto reset_pin_count
;
1034 ret
= intel_pin_and_map_ringbuffer_obj(ring
->dev
, ringbuf
);
1042 i915_gem_object_ggtt_unpin(ctx_obj
);
1044 ctx
->engine
[ring
->id
].pin_count
= 0;
1049 void intel_lr_context_unpin(struct intel_engine_cs
*ring
,
1050 struct intel_context
*ctx
)
1052 struct drm_i915_gem_object
*ctx_obj
= ctx
->engine
[ring
->id
].state
;
1053 struct intel_ringbuffer
*ringbuf
= ctx
->engine
[ring
->id
].ringbuf
;
1056 WARN_ON(!mutex_is_locked(&ring
->dev
->struct_mutex
));
1057 if (--ctx
->engine
[ring
->id
].pin_count
== 0) {
1058 intel_unpin_ringbuffer_obj(ringbuf
);
1059 i915_gem_object_ggtt_unpin(ctx_obj
);
1064 static int intel_logical_ring_workarounds_emit(struct intel_engine_cs
*ring
,
1065 struct intel_context
*ctx
)
1068 struct intel_ringbuffer
*ringbuf
= ctx
->engine
[ring
->id
].ringbuf
;
1069 struct drm_device
*dev
= ring
->dev
;
1070 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1071 struct i915_workarounds
*w
= &dev_priv
->workarounds
;
1073 if (WARN_ON_ONCE(w
->count
== 0))
1076 ring
->gpu_caches_dirty
= true;
1077 ret
= logical_ring_flush_all_caches(ringbuf
, ctx
);
1081 ret
= intel_logical_ring_begin(ringbuf
, ctx
, w
->count
* 2 + 2);
1085 intel_logical_ring_emit(ringbuf
, MI_LOAD_REGISTER_IMM(w
->count
));
1086 for (i
= 0; i
< w
->count
; i
++) {
1087 intel_logical_ring_emit(ringbuf
, w
->reg
[i
].addr
);
1088 intel_logical_ring_emit(ringbuf
, w
->reg
[i
].value
);
1090 intel_logical_ring_emit(ringbuf
, MI_NOOP
);
1092 intel_logical_ring_advance(ringbuf
);
1094 ring
->gpu_caches_dirty
= true;
1095 ret
= logical_ring_flush_all_caches(ringbuf
, ctx
);
1102 #define wa_ctx_emit(batch, cmd) \
1104 if (WARN_ON(index >= (PAGE_SIZE / sizeof(uint32_t)))) { \
1107 batch[index++] = (cmd); \
1110 static inline uint32_t wa_ctx_start(struct i915_wa_ctx_bb
*wa_ctx
,
1112 uint32_t start_alignment
)
1114 return wa_ctx
->offset
= ALIGN(offset
, start_alignment
);
1117 static inline int wa_ctx_end(struct i915_wa_ctx_bb
*wa_ctx
,
1119 uint32_t size_alignment
)
1121 wa_ctx
->size
= offset
- wa_ctx
->offset
;
1123 WARN(wa_ctx
->size
% size_alignment
,
1124 "wa_ctx_bb failed sanity checks: size %d is not aligned to %d\n",
1125 wa_ctx
->size
, size_alignment
);
1130 * gen8_init_indirectctx_bb() - initialize indirect ctx batch with WA
1132 * @ring: only applicable for RCS
1133 * @wa_ctx: structure representing wa_ctx
1134 * offset: specifies start of the batch, should be cache-aligned. This is updated
1135 * with the offset value received as input.
1136 * size: size of the batch in DWORDS but HW expects in terms of cachelines
1137 * @batch: page in which WA are loaded
1138 * @offset: This field specifies the start of the batch, it should be
1139 * cache-aligned otherwise it is adjusted accordingly.
1140 * Typically we only have one indirect_ctx and per_ctx batch buffer which are
1141 * initialized at the beginning and shared across all contexts but this field
1142 * helps us to have multiple batches at different offsets and select them based
1143 * on a criteria. At the moment this batch always start at the beginning of the page
1144 * and at this point we don't have multiple wa_ctx batch buffers.
1146 * The number of WA applied are not known at the beginning; we use this field
1147 * to return the no of DWORDS written.
1149 * It is to be noted that this batch does not contain MI_BATCH_BUFFER_END
1150 * so it adds NOOPs as padding to make it cacheline aligned.
1151 * MI_BATCH_BUFFER_END will be added to perctx batch and both of them together
1152 * makes a complete batch buffer.
1154 * Return: non-zero if we exceed the PAGE_SIZE limit.
1157 static int gen8_init_indirectctx_bb(struct intel_engine_cs
*ring
,
1158 struct i915_wa_ctx_bb
*wa_ctx
,
1159 uint32_t *const batch
,
1162 uint32_t index
= wa_ctx_start(wa_ctx
, *offset
, CACHELINE_DWORDS
);
1164 /* WaDisableCtxRestoreArbitration:bdw,chv */
1165 wa_ctx_emit(batch
, MI_ARB_ON_OFF
| MI_ARB_DISABLE
);
1167 /* WaFlushCoherentL3CacheLinesAtContextSwitch:bdw */
1168 if (IS_BROADWELL(ring
->dev
)) {
1169 struct drm_i915_private
*dev_priv
= to_i915(ring
->dev
);
1170 uint32_t l3sqc4_flush
= (I915_READ(GEN8_L3SQCREG4
) |
1171 GEN8_LQSC_FLUSH_COHERENT_LINES
);
1173 wa_ctx_emit(batch
, MI_LOAD_REGISTER_IMM(1));
1174 wa_ctx_emit(batch
, GEN8_L3SQCREG4
);
1175 wa_ctx_emit(batch
, l3sqc4_flush
);
1177 wa_ctx_emit(batch
, GFX_OP_PIPE_CONTROL(6));
1178 wa_ctx_emit(batch
, (PIPE_CONTROL_CS_STALL
|
1179 PIPE_CONTROL_DC_FLUSH_ENABLE
));
1180 wa_ctx_emit(batch
, 0);
1181 wa_ctx_emit(batch
, 0);
1182 wa_ctx_emit(batch
, 0);
1183 wa_ctx_emit(batch
, 0);
1185 wa_ctx_emit(batch
, MI_LOAD_REGISTER_IMM(1));
1186 wa_ctx_emit(batch
, GEN8_L3SQCREG4
);
1187 wa_ctx_emit(batch
, l3sqc4_flush
& ~GEN8_LQSC_FLUSH_COHERENT_LINES
);
1190 /* Pad to end of cacheline */
1191 while (index
% CACHELINE_DWORDS
)
1192 wa_ctx_emit(batch
, MI_NOOP
);
1195 * MI_BATCH_BUFFER_END is not required in Indirect ctx BB because
1196 * execution depends on the length specified in terms of cache lines
1197 * in the register CTX_RCS_INDIRECT_CTX
1200 return wa_ctx_end(wa_ctx
, *offset
= index
, CACHELINE_DWORDS
);
1204 * gen8_init_perctx_bb() - initialize per ctx batch with WA
1206 * @ring: only applicable for RCS
1207 * @wa_ctx: structure representing wa_ctx
1208 * offset: specifies start of the batch, should be cache-aligned.
1209 * size: size of the batch in DWORDS but HW expects in terms of cachelines
1210 * @offset: This field specifies the start of this batch.
1211 * This batch is started immediately after indirect_ctx batch. Since we ensure
1212 * that indirect_ctx ends on a cacheline this batch is aligned automatically.
1214 * The number of DWORDS written are returned using this field.
1216 * This batch is terminated with MI_BATCH_BUFFER_END and so we need not add padding
1217 * to align it with cacheline as padding after MI_BATCH_BUFFER_END is redundant.
1219 static int gen8_init_perctx_bb(struct intel_engine_cs
*ring
,
1220 struct i915_wa_ctx_bb
*wa_ctx
,
1221 uint32_t *const batch
,
1224 uint32_t index
= wa_ctx_start(wa_ctx
, *offset
, CACHELINE_DWORDS
);
1226 /* WaDisableCtxRestoreArbitration:bdw,chv */
1227 wa_ctx_emit(batch
, MI_ARB_ON_OFF
| MI_ARB_ENABLE
);
1229 wa_ctx_emit(batch
, MI_BATCH_BUFFER_END
);
1231 return wa_ctx_end(wa_ctx
, *offset
= index
, 1);
1234 static int lrc_setup_wa_ctx_obj(struct intel_engine_cs
*ring
, u32 size
)
1238 ring
->wa_ctx
.obj
= i915_gem_alloc_object(ring
->dev
, PAGE_ALIGN(size
));
1239 if (!ring
->wa_ctx
.obj
) {
1240 DRM_DEBUG_DRIVER("alloc LRC WA ctx backing obj failed.\n");
1244 ret
= i915_gem_obj_ggtt_pin(ring
->wa_ctx
.obj
, PAGE_SIZE
, 0);
1246 DRM_DEBUG_DRIVER("pin LRC WA ctx backing obj failed: %d\n",
1248 drm_gem_object_unreference(&ring
->wa_ctx
.obj
->base
);
1255 static void lrc_destroy_wa_ctx_obj(struct intel_engine_cs
*ring
)
1257 if (ring
->wa_ctx
.obj
) {
1258 i915_gem_object_ggtt_unpin(ring
->wa_ctx
.obj
);
1259 drm_gem_object_unreference(&ring
->wa_ctx
.obj
->base
);
1260 ring
->wa_ctx
.obj
= NULL
;
1264 static int intel_init_workaround_bb(struct intel_engine_cs
*ring
)
1270 struct i915_ctx_workarounds
*wa_ctx
= &ring
->wa_ctx
;
1272 WARN_ON(ring
->id
!= RCS
);
1274 /* some WA perform writes to scratch page, ensure it is valid */
1275 if (ring
->scratch
.obj
== NULL
) {
1276 DRM_ERROR("scratch page not allocated for %s\n", ring
->name
);
1280 ret
= lrc_setup_wa_ctx_obj(ring
, PAGE_SIZE
);
1282 DRM_DEBUG_DRIVER("Failed to setup context WA page: %d\n", ret
);
1286 page
= i915_gem_object_get_page(wa_ctx
->obj
, 0);
1287 batch
= kmap_atomic(page
);
1290 if (INTEL_INFO(ring
->dev
)->gen
== 8) {
1291 ret
= gen8_init_indirectctx_bb(ring
,
1292 &wa_ctx
->indirect_ctx
,
1298 ret
= gen8_init_perctx_bb(ring
,
1305 WARN(INTEL_INFO(ring
->dev
)->gen
>= 8,
1306 "WA batch buffer is not initialized for Gen%d\n",
1307 INTEL_INFO(ring
->dev
)->gen
);
1308 lrc_destroy_wa_ctx_obj(ring
);
1312 kunmap_atomic(batch
);
1314 lrc_destroy_wa_ctx_obj(ring
);
1319 static int gen8_init_common_ring(struct intel_engine_cs
*ring
)
1321 struct drm_device
*dev
= ring
->dev
;
1322 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1324 I915_WRITE_IMR(ring
, ~(ring
->irq_enable_mask
| ring
->irq_keep_mask
));
1325 I915_WRITE(RING_HWSTAM(ring
->mmio_base
), 0xffffffff);
1327 I915_WRITE(RING_MODE_GEN7(ring
),
1328 _MASKED_BIT_DISABLE(GFX_REPLAY_MODE
) |
1329 _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE
));
1330 POSTING_READ(RING_MODE_GEN7(ring
));
1331 ring
->next_context_status_buffer
= 0;
1332 DRM_DEBUG_DRIVER("Execlists enabled for %s\n", ring
->name
);
1334 memset(&ring
->hangcheck
, 0, sizeof(ring
->hangcheck
));
1339 static int gen8_init_render_ring(struct intel_engine_cs
*ring
)
1341 struct drm_device
*dev
= ring
->dev
;
1342 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1345 ret
= gen8_init_common_ring(ring
);
1349 /* We need to disable the AsyncFlip performance optimisations in order
1350 * to use MI_WAIT_FOR_EVENT within the CS. It should already be
1351 * programmed to '1' on all products.
1353 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv
1355 I915_WRITE(MI_MODE
, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE
));
1357 I915_WRITE(INSTPM
, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING
));
1359 return init_workarounds_ring(ring
);
1362 static int gen9_init_render_ring(struct intel_engine_cs
*ring
)
1366 ret
= gen8_init_common_ring(ring
);
1370 return init_workarounds_ring(ring
);
1373 static int gen8_emit_bb_start(struct intel_ringbuffer
*ringbuf
,
1374 struct intel_context
*ctx
,
1375 u64 offset
, unsigned dispatch_flags
)
1377 bool ppgtt
= !(dispatch_flags
& I915_DISPATCH_SECURE
);
1380 ret
= intel_logical_ring_begin(ringbuf
, ctx
, 4);
1384 /* FIXME(BDW): Address space and security selectors. */
1385 intel_logical_ring_emit(ringbuf
, MI_BATCH_BUFFER_START_GEN8
| (ppgtt
<<8));
1386 intel_logical_ring_emit(ringbuf
, lower_32_bits(offset
));
1387 intel_logical_ring_emit(ringbuf
, upper_32_bits(offset
));
1388 intel_logical_ring_emit(ringbuf
, MI_NOOP
);
1389 intel_logical_ring_advance(ringbuf
);
1394 static bool gen8_logical_ring_get_irq(struct intel_engine_cs
*ring
)
1396 struct drm_device
*dev
= ring
->dev
;
1397 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1398 unsigned long flags
;
1400 if (WARN_ON(!intel_irqs_enabled(dev_priv
)))
1403 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1404 if (ring
->irq_refcount
++ == 0) {
1405 I915_WRITE_IMR(ring
, ~(ring
->irq_enable_mask
| ring
->irq_keep_mask
));
1406 POSTING_READ(RING_IMR(ring
->mmio_base
));
1408 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1413 static void gen8_logical_ring_put_irq(struct intel_engine_cs
*ring
)
1415 struct drm_device
*dev
= ring
->dev
;
1416 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1417 unsigned long flags
;
1419 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1420 if (--ring
->irq_refcount
== 0) {
1421 I915_WRITE_IMR(ring
, ~ring
->irq_keep_mask
);
1422 POSTING_READ(RING_IMR(ring
->mmio_base
));
1424 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1427 static int gen8_emit_flush(struct intel_ringbuffer
*ringbuf
,
1428 struct intel_context
*ctx
,
1429 u32 invalidate_domains
,
1432 struct intel_engine_cs
*ring
= ringbuf
->ring
;
1433 struct drm_device
*dev
= ring
->dev
;
1434 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1438 ret
= intel_logical_ring_begin(ringbuf
, ctx
, 4);
1442 cmd
= MI_FLUSH_DW
+ 1;
1444 /* We always require a command barrier so that subsequent
1445 * commands, such as breadcrumb interrupts, are strictly ordered
1446 * wrt the contents of the write cache being flushed to memory
1447 * (and thus being coherent from the CPU).
1449 cmd
|= MI_FLUSH_DW_STORE_INDEX
| MI_FLUSH_DW_OP_STOREDW
;
1451 if (invalidate_domains
& I915_GEM_GPU_DOMAINS
) {
1452 cmd
|= MI_INVALIDATE_TLB
;
1453 if (ring
== &dev_priv
->ring
[VCS
])
1454 cmd
|= MI_INVALIDATE_BSD
;
1457 intel_logical_ring_emit(ringbuf
, cmd
);
1458 intel_logical_ring_emit(ringbuf
,
1459 I915_GEM_HWS_SCRATCH_ADDR
|
1460 MI_FLUSH_DW_USE_GTT
);
1461 intel_logical_ring_emit(ringbuf
, 0); /* upper addr */
1462 intel_logical_ring_emit(ringbuf
, 0); /* value */
1463 intel_logical_ring_advance(ringbuf
);
1468 static int gen8_emit_flush_render(struct intel_ringbuffer
*ringbuf
,
1469 struct intel_context
*ctx
,
1470 u32 invalidate_domains
,
1473 struct intel_engine_cs
*ring
= ringbuf
->ring
;
1474 u32 scratch_addr
= ring
->scratch
.gtt_offset
+ 2 * CACHELINE_BYTES
;
1479 flags
|= PIPE_CONTROL_CS_STALL
;
1481 if (flush_domains
) {
1482 flags
|= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH
;
1483 flags
|= PIPE_CONTROL_DEPTH_CACHE_FLUSH
;
1486 if (invalidate_domains
) {
1487 flags
|= PIPE_CONTROL_TLB_INVALIDATE
;
1488 flags
|= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE
;
1489 flags
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
1490 flags
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
1491 flags
|= PIPE_CONTROL_CONST_CACHE_INVALIDATE
;
1492 flags
|= PIPE_CONTROL_STATE_CACHE_INVALIDATE
;
1493 flags
|= PIPE_CONTROL_QW_WRITE
;
1494 flags
|= PIPE_CONTROL_GLOBAL_GTT_IVB
;
1498 * On GEN9+ Before VF_CACHE_INVALIDATE we need to emit a NULL pipe
1501 vf_flush_wa
= INTEL_INFO(ring
->dev
)->gen
>= 9 &&
1502 flags
& PIPE_CONTROL_VF_CACHE_INVALIDATE
;
1504 ret
= intel_logical_ring_begin(ringbuf
, ctx
, vf_flush_wa
? 12 : 6);
1509 intel_logical_ring_emit(ringbuf
, GFX_OP_PIPE_CONTROL(6));
1510 intel_logical_ring_emit(ringbuf
, 0);
1511 intel_logical_ring_emit(ringbuf
, 0);
1512 intel_logical_ring_emit(ringbuf
, 0);
1513 intel_logical_ring_emit(ringbuf
, 0);
1514 intel_logical_ring_emit(ringbuf
, 0);
1517 intel_logical_ring_emit(ringbuf
, GFX_OP_PIPE_CONTROL(6));
1518 intel_logical_ring_emit(ringbuf
, flags
);
1519 intel_logical_ring_emit(ringbuf
, scratch_addr
);
1520 intel_logical_ring_emit(ringbuf
, 0);
1521 intel_logical_ring_emit(ringbuf
, 0);
1522 intel_logical_ring_emit(ringbuf
, 0);
1523 intel_logical_ring_advance(ringbuf
);
1528 static u32
gen8_get_seqno(struct intel_engine_cs
*ring
, bool lazy_coherency
)
1530 return intel_read_status_page(ring
, I915_GEM_HWS_INDEX
);
1533 static void gen8_set_seqno(struct intel_engine_cs
*ring
, u32 seqno
)
1535 intel_write_status_page(ring
, I915_GEM_HWS_INDEX
, seqno
);
1538 static int gen8_emit_request(struct intel_ringbuffer
*ringbuf
,
1539 struct drm_i915_gem_request
*request
)
1541 struct intel_engine_cs
*ring
= ringbuf
->ring
;
1546 * Reserve space for 2 NOOPs at the end of each request to be
1547 * used as a workaround for not being allowed to do lite
1548 * restore with HEAD==TAIL (WaIdleLiteRestore).
1550 ret
= intel_logical_ring_begin(ringbuf
, request
->ctx
, 8);
1554 cmd
= MI_STORE_DWORD_IMM_GEN4
;
1555 cmd
|= MI_GLOBAL_GTT
;
1557 intel_logical_ring_emit(ringbuf
, cmd
);
1558 intel_logical_ring_emit(ringbuf
,
1559 (ring
->status_page
.gfx_addr
+
1560 (I915_GEM_HWS_INDEX
<< MI_STORE_DWORD_INDEX_SHIFT
)));
1561 intel_logical_ring_emit(ringbuf
, 0);
1562 intel_logical_ring_emit(ringbuf
,
1563 i915_gem_request_get_seqno(ring
->outstanding_lazy_request
));
1564 intel_logical_ring_emit(ringbuf
, MI_USER_INTERRUPT
);
1565 intel_logical_ring_emit(ringbuf
, MI_NOOP
);
1566 intel_logical_ring_advance_and_submit(ringbuf
, request
->ctx
, request
);
1569 * Here we add two extra NOOPs as padding to avoid
1570 * lite restore of a context with HEAD==TAIL.
1572 intel_logical_ring_emit(ringbuf
, MI_NOOP
);
1573 intel_logical_ring_emit(ringbuf
, MI_NOOP
);
1574 intel_logical_ring_advance(ringbuf
);
1579 static int intel_lr_context_render_state_init(struct intel_engine_cs
*ring
,
1580 struct intel_context
*ctx
)
1582 struct intel_ringbuffer
*ringbuf
= ctx
->engine
[ring
->id
].ringbuf
;
1583 struct render_state so
;
1584 struct drm_i915_file_private
*file_priv
= ctx
->file_priv
;
1585 struct drm_file
*file
= file_priv
? file_priv
->file
: NULL
;
1588 ret
= i915_gem_render_state_prepare(ring
, &so
);
1592 if (so
.rodata
== NULL
)
1595 ret
= ring
->emit_bb_start(ringbuf
,
1598 I915_DISPATCH_SECURE
);
1602 i915_vma_move_to_active(i915_gem_obj_to_ggtt(so
.obj
), ring
);
1604 ret
= __i915_add_request(ring
, file
, so
.obj
);
1605 /* intel_logical_ring_add_request moves object to inactive if it
1608 i915_gem_render_state_fini(&so
);
1612 static int gen8_init_rcs_context(struct intel_engine_cs
*ring
,
1613 struct intel_context
*ctx
)
1617 ret
= intel_logical_ring_workarounds_emit(ring
, ctx
);
1621 return intel_lr_context_render_state_init(ring
, ctx
);
1625 * intel_logical_ring_cleanup() - deallocate the Engine Command Streamer
1627 * @ring: Engine Command Streamer.
1630 void intel_logical_ring_cleanup(struct intel_engine_cs
*ring
)
1632 struct drm_i915_private
*dev_priv
;
1634 if (!intel_ring_initialized(ring
))
1637 dev_priv
= ring
->dev
->dev_private
;
1639 intel_logical_ring_stop(ring
);
1640 WARN_ON((I915_READ_MODE(ring
) & MODE_IDLE
) == 0);
1641 i915_gem_request_assign(&ring
->outstanding_lazy_request
, NULL
);
1644 ring
->cleanup(ring
);
1646 i915_cmd_parser_fini_ring(ring
);
1647 i915_gem_batch_pool_fini(&ring
->batch_pool
);
1649 if (ring
->status_page
.obj
) {
1650 kunmap(sg_page(ring
->status_page
.obj
->pages
->sgl
));
1651 ring
->status_page
.obj
= NULL
;
1654 lrc_destroy_wa_ctx_obj(ring
);
1657 static int logical_ring_init(struct drm_device
*dev
, struct intel_engine_cs
*ring
)
1661 /* Intentionally left blank. */
1662 ring
->buffer
= NULL
;
1665 INIT_LIST_HEAD(&ring
->active_list
);
1666 INIT_LIST_HEAD(&ring
->request_list
);
1667 i915_gem_batch_pool_init(dev
, &ring
->batch_pool
);
1668 init_waitqueue_head(&ring
->irq_queue
);
1670 INIT_LIST_HEAD(&ring
->execlist_queue
);
1671 INIT_LIST_HEAD(&ring
->execlist_retired_req_list
);
1672 spin_lock_init(&ring
->execlist_lock
);
1674 ret
= i915_cmd_parser_init_ring(ring
);
1678 ret
= intel_lr_context_deferred_create(ring
->default_context
, ring
);
1683 static int logical_render_ring_init(struct drm_device
*dev
)
1685 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1686 struct intel_engine_cs
*ring
= &dev_priv
->ring
[RCS
];
1689 ring
->name
= "render ring";
1691 ring
->mmio_base
= RENDER_RING_BASE
;
1692 ring
->irq_enable_mask
=
1693 GT_RENDER_USER_INTERRUPT
<< GEN8_RCS_IRQ_SHIFT
;
1694 ring
->irq_keep_mask
=
1695 GT_CONTEXT_SWITCH_INTERRUPT
<< GEN8_RCS_IRQ_SHIFT
;
1696 if (HAS_L3_DPF(dev
))
1697 ring
->irq_keep_mask
|= GT_RENDER_L3_PARITY_ERROR_INTERRUPT
;
1699 if (INTEL_INFO(dev
)->gen
>= 9)
1700 ring
->init_hw
= gen9_init_render_ring
;
1702 ring
->init_hw
= gen8_init_render_ring
;
1703 ring
->init_context
= gen8_init_rcs_context
;
1704 ring
->cleanup
= intel_fini_pipe_control
;
1705 ring
->get_seqno
= gen8_get_seqno
;
1706 ring
->set_seqno
= gen8_set_seqno
;
1707 ring
->emit_request
= gen8_emit_request
;
1708 ring
->emit_flush
= gen8_emit_flush_render
;
1709 ring
->irq_get
= gen8_logical_ring_get_irq
;
1710 ring
->irq_put
= gen8_logical_ring_put_irq
;
1711 ring
->emit_bb_start
= gen8_emit_bb_start
;
1715 ret
= intel_init_pipe_control(ring
);
1719 ret
= intel_init_workaround_bb(ring
);
1722 * We continue even if we fail to initialize WA batch
1723 * because we only expect rare glitches but nothing
1724 * critical to prevent us from using GPU
1726 DRM_ERROR("WA batch buffer initialization failed: %d\n",
1730 ret
= logical_ring_init(dev
, ring
);
1732 lrc_destroy_wa_ctx_obj(ring
);
1738 static int logical_bsd_ring_init(struct drm_device
*dev
)
1740 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1741 struct intel_engine_cs
*ring
= &dev_priv
->ring
[VCS
];
1743 ring
->name
= "bsd ring";
1745 ring
->mmio_base
= GEN6_BSD_RING_BASE
;
1746 ring
->irq_enable_mask
=
1747 GT_RENDER_USER_INTERRUPT
<< GEN8_VCS1_IRQ_SHIFT
;
1748 ring
->irq_keep_mask
=
1749 GT_CONTEXT_SWITCH_INTERRUPT
<< GEN8_VCS1_IRQ_SHIFT
;
1751 ring
->init_hw
= gen8_init_common_ring
;
1752 ring
->get_seqno
= gen8_get_seqno
;
1753 ring
->set_seqno
= gen8_set_seqno
;
1754 ring
->emit_request
= gen8_emit_request
;
1755 ring
->emit_flush
= gen8_emit_flush
;
1756 ring
->irq_get
= gen8_logical_ring_get_irq
;
1757 ring
->irq_put
= gen8_logical_ring_put_irq
;
1758 ring
->emit_bb_start
= gen8_emit_bb_start
;
1760 return logical_ring_init(dev
, ring
);
1763 static int logical_bsd2_ring_init(struct drm_device
*dev
)
1765 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1766 struct intel_engine_cs
*ring
= &dev_priv
->ring
[VCS2
];
1768 ring
->name
= "bds2 ring";
1770 ring
->mmio_base
= GEN8_BSD2_RING_BASE
;
1771 ring
->irq_enable_mask
=
1772 GT_RENDER_USER_INTERRUPT
<< GEN8_VCS2_IRQ_SHIFT
;
1773 ring
->irq_keep_mask
=
1774 GT_CONTEXT_SWITCH_INTERRUPT
<< GEN8_VCS2_IRQ_SHIFT
;
1776 ring
->init_hw
= gen8_init_common_ring
;
1777 ring
->get_seqno
= gen8_get_seqno
;
1778 ring
->set_seqno
= gen8_set_seqno
;
1779 ring
->emit_request
= gen8_emit_request
;
1780 ring
->emit_flush
= gen8_emit_flush
;
1781 ring
->irq_get
= gen8_logical_ring_get_irq
;
1782 ring
->irq_put
= gen8_logical_ring_put_irq
;
1783 ring
->emit_bb_start
= gen8_emit_bb_start
;
1785 return logical_ring_init(dev
, ring
);
1788 static int logical_blt_ring_init(struct drm_device
*dev
)
1790 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1791 struct intel_engine_cs
*ring
= &dev_priv
->ring
[BCS
];
1793 ring
->name
= "blitter ring";
1795 ring
->mmio_base
= BLT_RING_BASE
;
1796 ring
->irq_enable_mask
=
1797 GT_RENDER_USER_INTERRUPT
<< GEN8_BCS_IRQ_SHIFT
;
1798 ring
->irq_keep_mask
=
1799 GT_CONTEXT_SWITCH_INTERRUPT
<< GEN8_BCS_IRQ_SHIFT
;
1801 ring
->init_hw
= gen8_init_common_ring
;
1802 ring
->get_seqno
= gen8_get_seqno
;
1803 ring
->set_seqno
= gen8_set_seqno
;
1804 ring
->emit_request
= gen8_emit_request
;
1805 ring
->emit_flush
= gen8_emit_flush
;
1806 ring
->irq_get
= gen8_logical_ring_get_irq
;
1807 ring
->irq_put
= gen8_logical_ring_put_irq
;
1808 ring
->emit_bb_start
= gen8_emit_bb_start
;
1810 return logical_ring_init(dev
, ring
);
1813 static int logical_vebox_ring_init(struct drm_device
*dev
)
1815 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1816 struct intel_engine_cs
*ring
= &dev_priv
->ring
[VECS
];
1818 ring
->name
= "video enhancement ring";
1820 ring
->mmio_base
= VEBOX_RING_BASE
;
1821 ring
->irq_enable_mask
=
1822 GT_RENDER_USER_INTERRUPT
<< GEN8_VECS_IRQ_SHIFT
;
1823 ring
->irq_keep_mask
=
1824 GT_CONTEXT_SWITCH_INTERRUPT
<< GEN8_VECS_IRQ_SHIFT
;
1826 ring
->init_hw
= gen8_init_common_ring
;
1827 ring
->get_seqno
= gen8_get_seqno
;
1828 ring
->set_seqno
= gen8_set_seqno
;
1829 ring
->emit_request
= gen8_emit_request
;
1830 ring
->emit_flush
= gen8_emit_flush
;
1831 ring
->irq_get
= gen8_logical_ring_get_irq
;
1832 ring
->irq_put
= gen8_logical_ring_put_irq
;
1833 ring
->emit_bb_start
= gen8_emit_bb_start
;
1835 return logical_ring_init(dev
, ring
);
1839 * intel_logical_rings_init() - allocate, populate and init the Engine Command Streamers
1842 * This function inits the engines for an Execlists submission style (the equivalent in the
1843 * legacy ringbuffer submission world would be i915_gem_init_rings). It does it only for
1844 * those engines that are present in the hardware.
1846 * Return: non-zero if the initialization failed.
1848 int intel_logical_rings_init(struct drm_device
*dev
)
1850 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1853 ret
= logical_render_ring_init(dev
);
1858 ret
= logical_bsd_ring_init(dev
);
1860 goto cleanup_render_ring
;
1864 ret
= logical_blt_ring_init(dev
);
1866 goto cleanup_bsd_ring
;
1869 if (HAS_VEBOX(dev
)) {
1870 ret
= logical_vebox_ring_init(dev
);
1872 goto cleanup_blt_ring
;
1875 if (HAS_BSD2(dev
)) {
1876 ret
= logical_bsd2_ring_init(dev
);
1878 goto cleanup_vebox_ring
;
1881 ret
= i915_gem_set_seqno(dev
, ((u32
)~0 - 0x1000));
1883 goto cleanup_bsd2_ring
;
1888 intel_logical_ring_cleanup(&dev_priv
->ring
[VCS2
]);
1890 intel_logical_ring_cleanup(&dev_priv
->ring
[VECS
]);
1892 intel_logical_ring_cleanup(&dev_priv
->ring
[BCS
]);
1894 intel_logical_ring_cleanup(&dev_priv
->ring
[VCS
]);
1895 cleanup_render_ring
:
1896 intel_logical_ring_cleanup(&dev_priv
->ring
[RCS
]);
1902 make_rpcs(struct drm_device
*dev
)
1907 * No explicit RPCS request is needed to ensure full
1908 * slice/subslice/EU enablement prior to Gen9.
1910 if (INTEL_INFO(dev
)->gen
< 9)
1914 * Starting in Gen9, render power gating can leave
1915 * slice/subslice/EU in a partially enabled state. We
1916 * must make an explicit request through RPCS for full
1919 if (INTEL_INFO(dev
)->has_slice_pg
) {
1920 rpcs
|= GEN8_RPCS_S_CNT_ENABLE
;
1921 rpcs
|= INTEL_INFO(dev
)->slice_total
<<
1922 GEN8_RPCS_S_CNT_SHIFT
;
1923 rpcs
|= GEN8_RPCS_ENABLE
;
1926 if (INTEL_INFO(dev
)->has_subslice_pg
) {
1927 rpcs
|= GEN8_RPCS_SS_CNT_ENABLE
;
1928 rpcs
|= INTEL_INFO(dev
)->subslice_per_slice
<<
1929 GEN8_RPCS_SS_CNT_SHIFT
;
1930 rpcs
|= GEN8_RPCS_ENABLE
;
1933 if (INTEL_INFO(dev
)->has_eu_pg
) {
1934 rpcs
|= INTEL_INFO(dev
)->eu_per_subslice
<<
1935 GEN8_RPCS_EU_MIN_SHIFT
;
1936 rpcs
|= INTEL_INFO(dev
)->eu_per_subslice
<<
1937 GEN8_RPCS_EU_MAX_SHIFT
;
1938 rpcs
|= GEN8_RPCS_ENABLE
;
1945 populate_lr_context(struct intel_context
*ctx
, struct drm_i915_gem_object
*ctx_obj
,
1946 struct intel_engine_cs
*ring
, struct intel_ringbuffer
*ringbuf
)
1948 struct drm_device
*dev
= ring
->dev
;
1949 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1950 struct i915_hw_ppgtt
*ppgtt
= ctx
->ppgtt
;
1952 uint32_t *reg_state
;
1956 ppgtt
= dev_priv
->mm
.aliasing_ppgtt
;
1958 ret
= i915_gem_object_set_to_cpu_domain(ctx_obj
, true);
1960 DRM_DEBUG_DRIVER("Could not set to CPU domain\n");
1964 ret
= i915_gem_object_get_pages(ctx_obj
);
1966 DRM_DEBUG_DRIVER("Could not get object pages\n");
1970 i915_gem_object_pin_pages(ctx_obj
);
1972 /* The second page of the context object contains some fields which must
1973 * be set up prior to the first execution. */
1974 page
= i915_gem_object_get_page(ctx_obj
, 1);
1975 reg_state
= kmap_atomic(page
);
1977 /* A context is actually a big batch buffer with several MI_LOAD_REGISTER_IMM
1978 * commands followed by (reg, value) pairs. The values we are setting here are
1979 * only for the first context restore: on a subsequent save, the GPU will
1980 * recreate this batchbuffer with new values (including all the missing
1981 * MI_LOAD_REGISTER_IMM commands that we are not initializing here). */
1982 if (ring
->id
== RCS
)
1983 reg_state
[CTX_LRI_HEADER_0
] = MI_LOAD_REGISTER_IMM(14);
1985 reg_state
[CTX_LRI_HEADER_0
] = MI_LOAD_REGISTER_IMM(11);
1986 reg_state
[CTX_LRI_HEADER_0
] |= MI_LRI_FORCE_POSTED
;
1987 reg_state
[CTX_CONTEXT_CONTROL
] = RING_CONTEXT_CONTROL(ring
);
1988 reg_state
[CTX_CONTEXT_CONTROL
+1] =
1989 _MASKED_BIT_ENABLE(CTX_CTRL_INHIBIT_SYN_CTX_SWITCH
|
1990 CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT
);
1991 reg_state
[CTX_RING_HEAD
] = RING_HEAD(ring
->mmio_base
);
1992 reg_state
[CTX_RING_HEAD
+1] = 0;
1993 reg_state
[CTX_RING_TAIL
] = RING_TAIL(ring
->mmio_base
);
1994 reg_state
[CTX_RING_TAIL
+1] = 0;
1995 reg_state
[CTX_RING_BUFFER_START
] = RING_START(ring
->mmio_base
);
1996 /* Ring buffer start address is not known until the buffer is pinned.
1997 * It is written to the context image in execlists_update_context()
1999 reg_state
[CTX_RING_BUFFER_CONTROL
] = RING_CTL(ring
->mmio_base
);
2000 reg_state
[CTX_RING_BUFFER_CONTROL
+1] =
2001 ((ringbuf
->size
- PAGE_SIZE
) & RING_NR_PAGES
) | RING_VALID
;
2002 reg_state
[CTX_BB_HEAD_U
] = ring
->mmio_base
+ 0x168;
2003 reg_state
[CTX_BB_HEAD_U
+1] = 0;
2004 reg_state
[CTX_BB_HEAD_L
] = ring
->mmio_base
+ 0x140;
2005 reg_state
[CTX_BB_HEAD_L
+1] = 0;
2006 reg_state
[CTX_BB_STATE
] = ring
->mmio_base
+ 0x110;
2007 reg_state
[CTX_BB_STATE
+1] = (1<<5);
2008 reg_state
[CTX_SECOND_BB_HEAD_U
] = ring
->mmio_base
+ 0x11c;
2009 reg_state
[CTX_SECOND_BB_HEAD_U
+1] = 0;
2010 reg_state
[CTX_SECOND_BB_HEAD_L
] = ring
->mmio_base
+ 0x114;
2011 reg_state
[CTX_SECOND_BB_HEAD_L
+1] = 0;
2012 reg_state
[CTX_SECOND_BB_STATE
] = ring
->mmio_base
+ 0x118;
2013 reg_state
[CTX_SECOND_BB_STATE
+1] = 0;
2014 if (ring
->id
== RCS
) {
2015 reg_state
[CTX_BB_PER_CTX_PTR
] = ring
->mmio_base
+ 0x1c0;
2016 reg_state
[CTX_BB_PER_CTX_PTR
+1] = 0;
2017 reg_state
[CTX_RCS_INDIRECT_CTX
] = ring
->mmio_base
+ 0x1c4;
2018 reg_state
[CTX_RCS_INDIRECT_CTX
+1] = 0;
2019 reg_state
[CTX_RCS_INDIRECT_CTX_OFFSET
] = ring
->mmio_base
+ 0x1c8;
2020 reg_state
[CTX_RCS_INDIRECT_CTX_OFFSET
+1] = 0;
2021 if (ring
->wa_ctx
.obj
) {
2022 struct i915_ctx_workarounds
*wa_ctx
= &ring
->wa_ctx
;
2023 uint32_t ggtt_offset
= i915_gem_obj_ggtt_offset(wa_ctx
->obj
);
2025 reg_state
[CTX_RCS_INDIRECT_CTX
+1] =
2026 (ggtt_offset
+ wa_ctx
->indirect_ctx
.offset
* sizeof(uint32_t)) |
2027 (wa_ctx
->indirect_ctx
.size
/ CACHELINE_DWORDS
);
2029 reg_state
[CTX_RCS_INDIRECT_CTX_OFFSET
+1] =
2030 CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT
<< 6;
2032 reg_state
[CTX_BB_PER_CTX_PTR
+1] =
2033 (ggtt_offset
+ wa_ctx
->per_ctx
.offset
* sizeof(uint32_t)) |
2037 reg_state
[CTX_LRI_HEADER_1
] = MI_LOAD_REGISTER_IMM(9);
2038 reg_state
[CTX_LRI_HEADER_1
] |= MI_LRI_FORCE_POSTED
;
2039 reg_state
[CTX_CTX_TIMESTAMP
] = ring
->mmio_base
+ 0x3a8;
2040 reg_state
[CTX_CTX_TIMESTAMP
+1] = 0;
2041 reg_state
[CTX_PDP3_UDW
] = GEN8_RING_PDP_UDW(ring
, 3);
2042 reg_state
[CTX_PDP3_LDW
] = GEN8_RING_PDP_LDW(ring
, 3);
2043 reg_state
[CTX_PDP2_UDW
] = GEN8_RING_PDP_UDW(ring
, 2);
2044 reg_state
[CTX_PDP2_LDW
] = GEN8_RING_PDP_LDW(ring
, 2);
2045 reg_state
[CTX_PDP1_UDW
] = GEN8_RING_PDP_UDW(ring
, 1);
2046 reg_state
[CTX_PDP1_LDW
] = GEN8_RING_PDP_LDW(ring
, 1);
2047 reg_state
[CTX_PDP0_UDW
] = GEN8_RING_PDP_UDW(ring
, 0);
2048 reg_state
[CTX_PDP0_LDW
] = GEN8_RING_PDP_LDW(ring
, 0);
2050 /* With dynamic page allocation, PDPs may not be allocated at this point,
2051 * Point the unallocated PDPs to the scratch page
2053 ASSIGN_CTX_PDP(ppgtt
, reg_state
, 3);
2054 ASSIGN_CTX_PDP(ppgtt
, reg_state
, 2);
2055 ASSIGN_CTX_PDP(ppgtt
, reg_state
, 1);
2056 ASSIGN_CTX_PDP(ppgtt
, reg_state
, 0);
2057 if (ring
->id
== RCS
) {
2058 reg_state
[CTX_LRI_HEADER_2
] = MI_LOAD_REGISTER_IMM(1);
2059 reg_state
[CTX_R_PWR_CLK_STATE
] = GEN8_R_PWR_CLK_STATE
;
2060 reg_state
[CTX_R_PWR_CLK_STATE
+1] = make_rpcs(dev
);
2063 kunmap_atomic(reg_state
);
2066 set_page_dirty(page
);
2067 i915_gem_object_unpin_pages(ctx_obj
);
2073 * intel_lr_context_free() - free the LRC specific bits of a context
2074 * @ctx: the LR context to free.
2076 * The real context freeing is done in i915_gem_context_free: this only
2077 * takes care of the bits that are LRC related: the per-engine backing
2078 * objects and the logical ringbuffer.
2080 void intel_lr_context_free(struct intel_context
*ctx
)
2084 for (i
= 0; i
< I915_NUM_RINGS
; i
++) {
2085 struct drm_i915_gem_object
*ctx_obj
= ctx
->engine
[i
].state
;
2088 struct intel_ringbuffer
*ringbuf
=
2089 ctx
->engine
[i
].ringbuf
;
2090 struct intel_engine_cs
*ring
= ringbuf
->ring
;
2092 if (ctx
== ring
->default_context
) {
2093 intel_unpin_ringbuffer_obj(ringbuf
);
2094 i915_gem_object_ggtt_unpin(ctx_obj
);
2096 WARN_ON(ctx
->engine
[ring
->id
].pin_count
);
2097 intel_destroy_ringbuffer_obj(ringbuf
);
2099 drm_gem_object_unreference(&ctx_obj
->base
);
2104 static uint32_t get_lr_context_size(struct intel_engine_cs
*ring
)
2108 WARN_ON(INTEL_INFO(ring
->dev
)->gen
< 8);
2112 if (INTEL_INFO(ring
->dev
)->gen
>= 9)
2113 ret
= GEN9_LR_CONTEXT_RENDER_SIZE
;
2115 ret
= GEN8_LR_CONTEXT_RENDER_SIZE
;
2121 ret
= GEN8_LR_CONTEXT_OTHER_SIZE
;
2128 static void lrc_setup_hardware_status_page(struct intel_engine_cs
*ring
,
2129 struct drm_i915_gem_object
*default_ctx_obj
)
2131 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
2133 /* The status page is offset 0 from the default context object
2135 ring
->status_page
.gfx_addr
= i915_gem_obj_ggtt_offset(default_ctx_obj
);
2136 ring
->status_page
.page_addr
=
2137 kmap(sg_page(default_ctx_obj
->pages
->sgl
));
2138 ring
->status_page
.obj
= default_ctx_obj
;
2140 I915_WRITE(RING_HWS_PGA(ring
->mmio_base
),
2141 (u32
)ring
->status_page
.gfx_addr
);
2142 POSTING_READ(RING_HWS_PGA(ring
->mmio_base
));
2146 * intel_lr_context_deferred_create() - create the LRC specific bits of a context
2147 * @ctx: LR context to create.
2148 * @ring: engine to be used with the context.
2150 * This function can be called more than once, with different engines, if we plan
2151 * to use the context with them. The context backing objects and the ringbuffers
2152 * (specially the ringbuffer backing objects) suck a lot of memory up, and that's why
2153 * the creation is a deferred call: it's better to make sure first that we need to use
2154 * a given ring with the context.
2156 * Return: non-zero on error.
2158 int intel_lr_context_deferred_create(struct intel_context
*ctx
,
2159 struct intel_engine_cs
*ring
)
2161 const bool is_global_default_ctx
= (ctx
== ring
->default_context
);
2162 struct drm_device
*dev
= ring
->dev
;
2163 struct drm_i915_gem_object
*ctx_obj
;
2164 uint32_t context_size
;
2165 struct intel_ringbuffer
*ringbuf
;
2168 WARN_ON(ctx
->legacy_hw_ctx
.rcs_state
!= NULL
);
2169 WARN_ON(ctx
->engine
[ring
->id
].state
);
2171 context_size
= round_up(get_lr_context_size(ring
), 4096);
2173 ctx_obj
= i915_gem_alloc_object(dev
, context_size
);
2175 DRM_DEBUG_DRIVER("Alloc LRC backing obj failed.\n");
2179 if (is_global_default_ctx
) {
2180 ret
= i915_gem_obj_ggtt_pin(ctx_obj
, GEN8_LR_CONTEXT_ALIGN
, 0);
2182 DRM_DEBUG_DRIVER("Pin LRC backing obj failed: %d\n",
2184 drm_gem_object_unreference(&ctx_obj
->base
);
2189 ringbuf
= kzalloc(sizeof(*ringbuf
), GFP_KERNEL
);
2191 DRM_DEBUG_DRIVER("Failed to allocate ringbuffer %s\n",
2194 goto error_unpin_ctx
;
2197 ringbuf
->ring
= ring
;
2199 ringbuf
->size
= 32 * PAGE_SIZE
;
2200 ringbuf
->effective_size
= ringbuf
->size
;
2203 ringbuf
->last_retired_head
= -1;
2204 intel_ring_update_space(ringbuf
);
2206 if (ringbuf
->obj
== NULL
) {
2207 ret
= intel_alloc_ringbuffer_obj(dev
, ringbuf
);
2210 "Failed to allocate ringbuffer obj %s: %d\n",
2212 goto error_free_rbuf
;
2215 if (is_global_default_ctx
) {
2216 ret
= intel_pin_and_map_ringbuffer_obj(dev
, ringbuf
);
2219 "Failed to pin and map ringbuffer %s: %d\n",
2221 goto error_destroy_rbuf
;
2227 ret
= populate_lr_context(ctx
, ctx_obj
, ring
, ringbuf
);
2229 DRM_DEBUG_DRIVER("Failed to populate LRC: %d\n", ret
);
2233 ctx
->engine
[ring
->id
].ringbuf
= ringbuf
;
2234 ctx
->engine
[ring
->id
].state
= ctx_obj
;
2236 if (ctx
== ring
->default_context
)
2237 lrc_setup_hardware_status_page(ring
, ctx_obj
);
2238 else if (ring
->id
== RCS
&& !ctx
->rcs_initialized
) {
2239 if (ring
->init_context
) {
2240 ret
= ring
->init_context(ring
, ctx
);
2242 DRM_ERROR("ring init context: %d\n", ret
);
2243 ctx
->engine
[ring
->id
].ringbuf
= NULL
;
2244 ctx
->engine
[ring
->id
].state
= NULL
;
2249 ctx
->rcs_initialized
= true;
2255 if (is_global_default_ctx
)
2256 intel_unpin_ringbuffer_obj(ringbuf
);
2258 intel_destroy_ringbuffer_obj(ringbuf
);
2262 if (is_global_default_ctx
)
2263 i915_gem_object_ggtt_unpin(ctx_obj
);
2264 drm_gem_object_unreference(&ctx_obj
->base
);
2268 void intel_lr_context_reset(struct drm_device
*dev
,
2269 struct intel_context
*ctx
)
2271 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2272 struct intel_engine_cs
*ring
;
2275 for_each_ring(ring
, dev_priv
, i
) {
2276 struct drm_i915_gem_object
*ctx_obj
=
2277 ctx
->engine
[ring
->id
].state
;
2278 struct intel_ringbuffer
*ringbuf
=
2279 ctx
->engine
[ring
->id
].ringbuf
;
2280 uint32_t *reg_state
;
2286 if (i915_gem_object_get_pages(ctx_obj
)) {
2287 WARN(1, "Failed get_pages for context obj\n");
2290 page
= i915_gem_object_get_page(ctx_obj
, 1);
2291 reg_state
= kmap_atomic(page
);
2293 reg_state
[CTX_RING_HEAD
+1] = 0;
2294 reg_state
[CTX_RING_TAIL
+1] = 0;
2296 kunmap_atomic(reg_state
);