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 "intel_mocs.h"
141 #define GEN9_LR_CONTEXT_RENDER_SIZE (22 * PAGE_SIZE)
142 #define GEN8_LR_CONTEXT_RENDER_SIZE (20 * PAGE_SIZE)
143 #define GEN8_LR_CONTEXT_OTHER_SIZE (2 * PAGE_SIZE)
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 CTX_LRI_HEADER_0 0x01
160 #define CTX_CONTEXT_CONTROL 0x02
161 #define CTX_RING_HEAD 0x04
162 #define CTX_RING_TAIL 0x06
163 #define CTX_RING_BUFFER_START 0x08
164 #define CTX_RING_BUFFER_CONTROL 0x0a
165 #define CTX_BB_HEAD_U 0x0c
166 #define CTX_BB_HEAD_L 0x0e
167 #define CTX_BB_STATE 0x10
168 #define CTX_SECOND_BB_HEAD_U 0x12
169 #define CTX_SECOND_BB_HEAD_L 0x14
170 #define CTX_SECOND_BB_STATE 0x16
171 #define CTX_BB_PER_CTX_PTR 0x18
172 #define CTX_RCS_INDIRECT_CTX 0x1a
173 #define CTX_RCS_INDIRECT_CTX_OFFSET 0x1c
174 #define CTX_LRI_HEADER_1 0x21
175 #define CTX_CTX_TIMESTAMP 0x22
176 #define CTX_PDP3_UDW 0x24
177 #define CTX_PDP3_LDW 0x26
178 #define CTX_PDP2_UDW 0x28
179 #define CTX_PDP2_LDW 0x2a
180 #define CTX_PDP1_UDW 0x2c
181 #define CTX_PDP1_LDW 0x2e
182 #define CTX_PDP0_UDW 0x30
183 #define CTX_PDP0_LDW 0x32
184 #define CTX_LRI_HEADER_2 0x41
185 #define CTX_R_PWR_CLK_STATE 0x42
186 #define CTX_GPGPU_CSR_BASE_ADDRESS 0x44
188 #define GEN8_CTX_VALID (1<<0)
189 #define GEN8_CTX_FORCE_PD_RESTORE (1<<1)
190 #define GEN8_CTX_FORCE_RESTORE (1<<2)
191 #define GEN8_CTX_L3LLC_COHERENT (1<<5)
192 #define GEN8_CTX_PRIVILEGE (1<<8)
194 #define ASSIGN_CTX_REG(reg_state, pos, reg, val) do { \
195 (reg_state)[(pos)+0] = i915_mmio_reg_offset(reg); \
196 (reg_state)[(pos)+1] = (val); \
199 #define ASSIGN_CTX_PDP(ppgtt, reg_state, n) do { \
200 const u64 _addr = i915_page_dir_dma_addr((ppgtt), (n)); \
201 reg_state[CTX_PDP ## n ## _UDW+1] = upper_32_bits(_addr); \
202 reg_state[CTX_PDP ## n ## _LDW+1] = lower_32_bits(_addr); \
205 #define ASSIGN_CTX_PML4(ppgtt, reg_state) do { \
206 reg_state[CTX_PDP0_UDW + 1] = upper_32_bits(px_dma(&ppgtt->pml4)); \
207 reg_state[CTX_PDP0_LDW + 1] = lower_32_bits(px_dma(&ppgtt->pml4)); \
212 FAULT_AND_HALT
, /* Debug only */
214 FAULT_AND_CONTINUE
/* Unsupported */
216 #define GEN8_CTX_ID_SHIFT 32
217 #define GEN8_CTX_ID_WIDTH 21
218 #define GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT 0x17
219 #define GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT 0x26
221 /* Typical size of the average request (2 pipecontrols and a MI_BB) */
222 #define EXECLISTS_REQUEST_SIZE 64 /* bytes */
224 static int execlists_context_deferred_alloc(struct i915_gem_context
*ctx
,
225 struct intel_engine_cs
*engine
);
226 static int intel_lr_context_pin(struct i915_gem_context
*ctx
,
227 struct intel_engine_cs
*engine
);
230 * intel_sanitize_enable_execlists() - sanitize i915.enable_execlists
231 * @dev_priv: i915 device private
232 * @enable_execlists: value of i915.enable_execlists module parameter.
234 * Only certain platforms support Execlists (the prerequisites being
235 * support for Logical Ring Contexts and Aliasing PPGTT or better).
237 * Return: 1 if Execlists is supported and has to be enabled.
239 int intel_sanitize_enable_execlists(struct drm_i915_private
*dev_priv
, int enable_execlists
)
241 /* On platforms with execlist available, vGPU will only
242 * support execlist mode, no ring buffer mode.
244 if (HAS_LOGICAL_RING_CONTEXTS(dev_priv
) && intel_vgpu_active(dev_priv
))
247 if (INTEL_GEN(dev_priv
) >= 9)
250 if (enable_execlists
== 0)
253 if (HAS_LOGICAL_RING_CONTEXTS(dev_priv
) &&
254 USES_PPGTT(dev_priv
) &&
255 i915
.use_mmio_flip
>= 0)
262 logical_ring_init_platform_invariants(struct intel_engine_cs
*engine
)
264 struct drm_i915_private
*dev_priv
= engine
->i915
;
266 if (IS_GEN8(dev_priv
) || IS_GEN9(dev_priv
))
267 engine
->idle_lite_restore_wa
= ~0;
269 engine
->disable_lite_restore_wa
= (IS_SKL_REVID(dev_priv
, 0, SKL_REVID_B0
) ||
270 IS_BXT_REVID(dev_priv
, 0, BXT_REVID_A1
)) &&
271 (engine
->id
== VCS
|| engine
->id
== VCS2
);
273 engine
->ctx_desc_template
= GEN8_CTX_VALID
;
274 if (IS_GEN8(dev_priv
))
275 engine
->ctx_desc_template
|= GEN8_CTX_L3LLC_COHERENT
;
276 engine
->ctx_desc_template
|= GEN8_CTX_PRIVILEGE
;
278 /* TODO: WaDisableLiteRestore when we start using semaphore
279 * signalling between Command Streamers */
280 /* ring->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE; */
282 /* WaEnableForceRestoreInCtxtDescForVCS:skl */
283 /* WaEnableForceRestoreInCtxtDescForVCS:bxt */
284 if (engine
->disable_lite_restore_wa
)
285 engine
->ctx_desc_template
|= GEN8_CTX_FORCE_RESTORE
;
289 * intel_lr_context_descriptor_update() - calculate & cache the descriptor
290 * descriptor for a pinned context
292 * @ctx: Context to work on
293 * @engine: Engine the descriptor will be used with
295 * The context descriptor encodes various attributes of a context,
296 * including its GTT address and some flags. Because it's fairly
297 * expensive to calculate, we'll just do it once and cache the result,
298 * which remains valid until the context is unpinned.
300 * This is what a descriptor looks like, from LSB to MSB:
301 * bits 0-11: flags, GEN8_CTX_* (cached in ctx_desc_template)
302 * bits 12-31: LRCA, GTT address of (the HWSP of) this context
303 * bits 32-52: ctx ID, a globally unique tag
304 * bits 53-54: mbz, reserved for use by hardware
305 * bits 55-63: group ID, currently unused and set to 0
308 intel_lr_context_descriptor_update(struct i915_gem_context
*ctx
,
309 struct intel_engine_cs
*engine
)
311 struct intel_context
*ce
= &ctx
->engine
[engine
->id
];
314 BUILD_BUG_ON(MAX_CONTEXT_HW_ID
> (1<<GEN8_CTX_ID_WIDTH
));
316 desc
= ctx
->desc_template
; /* bits 3-4 */
317 desc
|= engine
->ctx_desc_template
; /* bits 0-11 */
318 desc
|= ce
->lrc_vma
->node
.start
+ LRC_PPHWSP_PN
* PAGE_SIZE
;
320 desc
|= (u64
)ctx
->hw_id
<< GEN8_CTX_ID_SHIFT
; /* bits 32-52 */
325 uint64_t intel_lr_context_descriptor(struct i915_gem_context
*ctx
,
326 struct intel_engine_cs
*engine
)
328 return ctx
->engine
[engine
->id
].lrc_desc
;
331 static void execlists_elsp_write(struct drm_i915_gem_request
*rq0
,
332 struct drm_i915_gem_request
*rq1
)
335 struct intel_engine_cs
*engine
= rq0
->engine
;
336 struct drm_i915_private
*dev_priv
= rq0
->i915
;
340 desc
[1] = intel_lr_context_descriptor(rq1
->ctx
, rq1
->engine
);
341 rq1
->elsp_submitted
++;
346 desc
[0] = intel_lr_context_descriptor(rq0
->ctx
, rq0
->engine
);
347 rq0
->elsp_submitted
++;
349 /* You must always write both descriptors in the order below. */
350 I915_WRITE_FW(RING_ELSP(engine
), upper_32_bits(desc
[1]));
351 I915_WRITE_FW(RING_ELSP(engine
), lower_32_bits(desc
[1]));
353 I915_WRITE_FW(RING_ELSP(engine
), upper_32_bits(desc
[0]));
354 /* The context is automatically loaded after the following */
355 I915_WRITE_FW(RING_ELSP(engine
), lower_32_bits(desc
[0]));
357 /* ELSP is a wo register, use another nearby reg for posting */
358 POSTING_READ_FW(RING_EXECLIST_STATUS_LO(engine
));
362 execlists_update_context_pdps(struct i915_hw_ppgtt
*ppgtt
, u32
*reg_state
)
364 ASSIGN_CTX_PDP(ppgtt
, reg_state
, 3);
365 ASSIGN_CTX_PDP(ppgtt
, reg_state
, 2);
366 ASSIGN_CTX_PDP(ppgtt
, reg_state
, 1);
367 ASSIGN_CTX_PDP(ppgtt
, reg_state
, 0);
370 static void execlists_update_context(struct drm_i915_gem_request
*rq
)
372 struct intel_engine_cs
*engine
= rq
->engine
;
373 struct i915_hw_ppgtt
*ppgtt
= rq
->ctx
->ppgtt
;
374 uint32_t *reg_state
= rq
->ctx
->engine
[engine
->id
].lrc_reg_state
;
376 reg_state
[CTX_RING_TAIL
+1] = rq
->tail
;
378 /* True 32b PPGTT with dynamic page allocation: update PDP
379 * registers and point the unallocated PDPs to scratch page.
380 * PML4 is allocated during ppgtt init, so this is not needed
383 if (ppgtt
&& !USES_FULL_48BIT_PPGTT(ppgtt
->base
.dev
))
384 execlists_update_context_pdps(ppgtt
, reg_state
);
387 static void execlists_submit_requests(struct drm_i915_gem_request
*rq0
,
388 struct drm_i915_gem_request
*rq1
)
390 struct drm_i915_private
*dev_priv
= rq0
->i915
;
391 unsigned int fw_domains
= rq0
->engine
->fw_domains
;
393 execlists_update_context(rq0
);
396 execlists_update_context(rq1
);
398 spin_lock_irq(&dev_priv
->uncore
.lock
);
399 intel_uncore_forcewake_get__locked(dev_priv
, fw_domains
);
401 execlists_elsp_write(rq0
, rq1
);
403 intel_uncore_forcewake_put__locked(dev_priv
, fw_domains
);
404 spin_unlock_irq(&dev_priv
->uncore
.lock
);
407 static inline void execlists_context_status_change(
408 struct drm_i915_gem_request
*rq
,
409 unsigned long status
)
412 * Only used when GVT-g is enabled now. When GVT-g is disabled,
413 * The compiler should eliminate this function as dead-code.
415 if (!IS_ENABLED(CONFIG_DRM_I915_GVT
))
418 atomic_notifier_call_chain(&rq
->ctx
->status_notifier
, status
, rq
);
421 static void execlists_context_unqueue(struct intel_engine_cs
*engine
)
423 struct drm_i915_gem_request
*req0
= NULL
, *req1
= NULL
;
424 struct drm_i915_gem_request
*cursor
, *tmp
;
426 assert_spin_locked(&engine
->execlist_lock
);
429 * If irqs are not active generate a warning as batches that finish
430 * without the irqs may get lost and a GPU Hang may occur.
432 WARN_ON(!intel_irqs_enabled(engine
->i915
));
434 /* Try to read in pairs */
435 list_for_each_entry_safe(cursor
, tmp
, &engine
->execlist_queue
,
439 } else if (req0
->ctx
== cursor
->ctx
) {
440 /* Same ctx: ignore first request, as second request
441 * will update tail past first request's workload */
442 cursor
->elsp_submitted
= req0
->elsp_submitted
;
443 list_del(&req0
->execlist_link
);
444 i915_gem_request_unreference(req0
);
448 WARN_ON(req1
->elsp_submitted
);
456 execlists_context_status_change(req0
, INTEL_CONTEXT_SCHEDULE_IN
);
459 execlists_context_status_change(req1
,
460 INTEL_CONTEXT_SCHEDULE_IN
);
462 if (req0
->elsp_submitted
& engine
->idle_lite_restore_wa
) {
464 * WaIdleLiteRestore: make sure we never cause a lite restore
467 * Apply the wa NOOPS to prevent ring:HEAD == req:TAIL as we
468 * resubmit the request. See gen8_emit_request() for where we
469 * prepare the padding after the end of the request.
471 struct intel_ringbuffer
*ringbuf
;
473 ringbuf
= req0
->ctx
->engine
[engine
->id
].ringbuf
;
475 req0
->tail
&= ringbuf
->size
- 1;
478 execlists_submit_requests(req0
, req1
);
482 execlists_check_remove_request(struct intel_engine_cs
*engine
, u32 ctx_id
)
484 struct drm_i915_gem_request
*head_req
;
486 assert_spin_locked(&engine
->execlist_lock
);
488 head_req
= list_first_entry_or_null(&engine
->execlist_queue
,
489 struct drm_i915_gem_request
,
492 if (WARN_ON(!head_req
|| (head_req
->ctx_hw_id
!= ctx_id
)))
495 WARN(head_req
->elsp_submitted
== 0, "Never submitted head request\n");
497 if (--head_req
->elsp_submitted
> 0)
500 execlists_context_status_change(head_req
, INTEL_CONTEXT_SCHEDULE_OUT
);
502 list_del(&head_req
->execlist_link
);
503 i915_gem_request_unreference(head_req
);
509 get_context_status(struct intel_engine_cs
*engine
, unsigned int read_pointer
,
512 struct drm_i915_private
*dev_priv
= engine
->i915
;
515 read_pointer
%= GEN8_CSB_ENTRIES
;
517 status
= I915_READ_FW(RING_CONTEXT_STATUS_BUF_LO(engine
, read_pointer
));
519 if (status
& GEN8_CTX_STATUS_IDLE_ACTIVE
)
522 *context_id
= I915_READ_FW(RING_CONTEXT_STATUS_BUF_HI(engine
,
529 * intel_lrc_irq_handler() - handle Context Switch interrupts
530 * @data: tasklet handler passed in unsigned long
532 * Check the unread Context Status Buffers and manage the submission of new
533 * contexts to the ELSP accordingly.
535 static void intel_lrc_irq_handler(unsigned long data
)
537 struct intel_engine_cs
*engine
= (struct intel_engine_cs
*)data
;
538 struct drm_i915_private
*dev_priv
= engine
->i915
;
540 unsigned int read_pointer
, write_pointer
;
541 u32 csb
[GEN8_CSB_ENTRIES
][2];
542 unsigned int csb_read
= 0, i
;
543 unsigned int submit_contexts
= 0;
545 intel_uncore_forcewake_get(dev_priv
, engine
->fw_domains
);
547 status_pointer
= I915_READ_FW(RING_CONTEXT_STATUS_PTR(engine
));
549 read_pointer
= engine
->next_context_status_buffer
;
550 write_pointer
= GEN8_CSB_WRITE_PTR(status_pointer
);
551 if (read_pointer
> write_pointer
)
552 write_pointer
+= GEN8_CSB_ENTRIES
;
554 while (read_pointer
< write_pointer
) {
555 if (WARN_ON_ONCE(csb_read
== GEN8_CSB_ENTRIES
))
557 csb
[csb_read
][0] = get_context_status(engine
, ++read_pointer
,
562 engine
->next_context_status_buffer
= write_pointer
% GEN8_CSB_ENTRIES
;
564 /* Update the read pointer to the old write pointer. Manual ringbuffer
565 * management ftw </sarcasm> */
566 I915_WRITE_FW(RING_CONTEXT_STATUS_PTR(engine
),
567 _MASKED_FIELD(GEN8_CSB_READ_PTR_MASK
,
568 engine
->next_context_status_buffer
<< 8));
570 intel_uncore_forcewake_put(dev_priv
, engine
->fw_domains
);
572 spin_lock(&engine
->execlist_lock
);
574 for (i
= 0; i
< csb_read
; i
++) {
575 if (unlikely(csb
[i
][0] & GEN8_CTX_STATUS_PREEMPTED
)) {
576 if (csb
[i
][0] & GEN8_CTX_STATUS_LITE_RESTORE
) {
577 if (execlists_check_remove_request(engine
, csb
[i
][1]))
578 WARN(1, "Lite Restored request removed from queue\n");
580 WARN(1, "Preemption without Lite Restore\n");
583 if (csb
[i
][0] & (GEN8_CTX_STATUS_ACTIVE_IDLE
|
584 GEN8_CTX_STATUS_ELEMENT_SWITCH
))
586 execlists_check_remove_request(engine
, csb
[i
][1]);
589 if (submit_contexts
) {
590 if (!engine
->disable_lite_restore_wa
||
591 (csb
[i
][0] & GEN8_CTX_STATUS_ACTIVE_IDLE
))
592 execlists_context_unqueue(engine
);
595 spin_unlock(&engine
->execlist_lock
);
597 if (unlikely(submit_contexts
> 2))
598 DRM_ERROR("More than two context complete events?\n");
601 static void execlists_context_queue(struct drm_i915_gem_request
*request
)
603 struct intel_engine_cs
*engine
= request
->engine
;
604 struct drm_i915_gem_request
*cursor
;
605 int num_elements
= 0;
607 spin_lock_bh(&engine
->execlist_lock
);
609 list_for_each_entry(cursor
, &engine
->execlist_queue
, execlist_link
)
610 if (++num_elements
> 2)
613 if (num_elements
> 2) {
614 struct drm_i915_gem_request
*tail_req
;
616 tail_req
= list_last_entry(&engine
->execlist_queue
,
617 struct drm_i915_gem_request
,
620 if (request
->ctx
== tail_req
->ctx
) {
621 WARN(tail_req
->elsp_submitted
!= 0,
622 "More than 2 already-submitted reqs queued\n");
623 list_del(&tail_req
->execlist_link
);
624 i915_gem_request_unreference(tail_req
);
628 i915_gem_request_reference(request
);
629 list_add_tail(&request
->execlist_link
, &engine
->execlist_queue
);
630 request
->ctx_hw_id
= request
->ctx
->hw_id
;
631 if (num_elements
== 0)
632 execlists_context_unqueue(engine
);
634 spin_unlock_bh(&engine
->execlist_lock
);
637 static int logical_ring_invalidate_all_caches(struct drm_i915_gem_request
*req
)
639 struct intel_engine_cs
*engine
= req
->engine
;
640 uint32_t flush_domains
;
644 if (engine
->gpu_caches_dirty
)
645 flush_domains
= I915_GEM_GPU_DOMAINS
;
647 ret
= engine
->emit_flush(req
, I915_GEM_GPU_DOMAINS
, flush_domains
);
651 engine
->gpu_caches_dirty
= false;
655 static int execlists_move_to_gpu(struct drm_i915_gem_request
*req
,
656 struct list_head
*vmas
)
658 const unsigned other_rings
= ~intel_engine_flag(req
->engine
);
659 struct i915_vma
*vma
;
660 uint32_t flush_domains
= 0;
661 bool flush_chipset
= false;
664 list_for_each_entry(vma
, vmas
, exec_list
) {
665 struct drm_i915_gem_object
*obj
= vma
->obj
;
667 if (obj
->active
& other_rings
) {
668 ret
= i915_gem_object_sync(obj
, req
->engine
, &req
);
673 if (obj
->base
.write_domain
& I915_GEM_DOMAIN_CPU
)
674 flush_chipset
|= i915_gem_clflush_object(obj
, false);
676 flush_domains
|= obj
->base
.write_domain
;
679 if (flush_domains
& I915_GEM_DOMAIN_GTT
)
682 /* Unconditionally invalidate gpu caches and ensure that we do flush
683 * any residual writes from the previous batch.
685 return logical_ring_invalidate_all_caches(req
);
688 int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request
*request
)
690 struct intel_engine_cs
*engine
= request
->engine
;
691 struct intel_context
*ce
= &request
->ctx
->engine
[engine
->id
];
694 /* Flush enough space to reduce the likelihood of waiting after
695 * we start building the request - in which case we will just
696 * have to repeat work.
698 request
->reserved_space
+= EXECLISTS_REQUEST_SIZE
;
701 ret
= execlists_context_deferred_alloc(request
->ctx
, engine
);
706 request
->ringbuf
= ce
->ringbuf
;
708 if (i915
.enable_guc_submission
) {
710 * Check that the GuC has space for the request before
711 * going any further, as the i915_add_request() call
712 * later on mustn't fail ...
714 ret
= i915_guc_wq_check_space(request
);
719 ret
= intel_lr_context_pin(request
->ctx
, engine
);
723 ret
= intel_ring_begin(request
, 0);
727 if (!ce
->initialised
) {
728 ret
= engine
->init_context(request
);
732 ce
->initialised
= true;
735 /* Note that after this point, we have committed to using
736 * this request as it is being used to both track the
737 * state of engine initialisation and liveness of the
738 * golden renderstate above. Think twice before you try
739 * to cancel/unwind this request now.
742 request
->reserved_space
-= EXECLISTS_REQUEST_SIZE
;
746 intel_lr_context_unpin(request
->ctx
, engine
);
751 * intel_logical_ring_advance_and_submit() - advance the tail and submit the workload
752 * @request: Request to advance the logical ringbuffer of.
754 * The tail is updated in our logical ringbuffer struct, not in the actual context. What
755 * really happens during submission is that the context and current tail will be placed
756 * on a queue waiting for the ELSP to be ready to accept a new context submission. At that
757 * point, the tail *inside* the context is updated and the ELSP written to.
760 intel_logical_ring_advance_and_submit(struct drm_i915_gem_request
*request
)
762 struct intel_ringbuffer
*ringbuf
= request
->ringbuf
;
763 struct intel_engine_cs
*engine
= request
->engine
;
765 intel_logical_ring_advance(ringbuf
);
766 request
->tail
= ringbuf
->tail
;
769 * Here we add two extra NOOPs as padding to avoid
770 * lite restore of a context with HEAD==TAIL.
772 * Caller must reserve WA_TAIL_DWORDS for us!
774 intel_logical_ring_emit(ringbuf
, MI_NOOP
);
775 intel_logical_ring_emit(ringbuf
, MI_NOOP
);
776 intel_logical_ring_advance(ringbuf
);
778 if (intel_engine_stopped(engine
))
781 /* We keep the previous context alive until we retire the following
782 * request. This ensures that any the context object is still pinned
783 * for any residual writes the HW makes into it on the context switch
784 * into the next object following the breadcrumb. Otherwise, we may
785 * retire the context too early.
787 request
->previous_context
= engine
->last_context
;
788 engine
->last_context
= request
->ctx
;
790 if (i915
.enable_guc_submission
)
791 i915_guc_submit(request
);
793 execlists_context_queue(request
);
799 * execlists_submission() - submit a batchbuffer for execution, Execlists style
800 * @params: execbuffer call parameters.
801 * @args: execbuffer call arguments.
802 * @vmas: list of vmas.
804 * This is the evil twin version of i915_gem_ringbuffer_submission. It abstracts
805 * away the submission details of the execbuffer ioctl call.
807 * Return: non-zero if the submission fails.
809 int intel_execlists_submission(struct i915_execbuffer_params
*params
,
810 struct drm_i915_gem_execbuffer2
*args
,
811 struct list_head
*vmas
)
813 struct drm_device
*dev
= params
->dev
;
814 struct intel_engine_cs
*engine
= params
->engine
;
815 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
816 struct intel_ringbuffer
*ringbuf
= params
->ctx
->engine
[engine
->id
].ringbuf
;
822 instp_mode
= args
->flags
& I915_EXEC_CONSTANTS_MASK
;
823 instp_mask
= I915_EXEC_CONSTANTS_MASK
;
824 switch (instp_mode
) {
825 case I915_EXEC_CONSTANTS_REL_GENERAL
:
826 case I915_EXEC_CONSTANTS_ABSOLUTE
:
827 case I915_EXEC_CONSTANTS_REL_SURFACE
:
828 if (instp_mode
!= 0 && engine
!= &dev_priv
->engine
[RCS
]) {
829 DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
833 if (instp_mode
!= dev_priv
->relative_constants_mode
) {
834 if (instp_mode
== I915_EXEC_CONSTANTS_REL_SURFACE
) {
835 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
839 /* The HW changed the meaning on this bit on gen6 */
840 instp_mask
&= ~I915_EXEC_CONSTANTS_REL_SURFACE
;
844 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode
);
848 if (args
->flags
& I915_EXEC_GEN7_SOL_RESET
) {
849 DRM_DEBUG("sol reset is gen7 only\n");
853 ret
= execlists_move_to_gpu(params
->request
, vmas
);
857 if (engine
== &dev_priv
->engine
[RCS
] &&
858 instp_mode
!= dev_priv
->relative_constants_mode
) {
859 ret
= intel_ring_begin(params
->request
, 4);
863 intel_logical_ring_emit(ringbuf
, MI_NOOP
);
864 intel_logical_ring_emit(ringbuf
, MI_LOAD_REGISTER_IMM(1));
865 intel_logical_ring_emit_reg(ringbuf
, INSTPM
);
866 intel_logical_ring_emit(ringbuf
, instp_mask
<< 16 | instp_mode
);
867 intel_logical_ring_advance(ringbuf
);
869 dev_priv
->relative_constants_mode
= instp_mode
;
872 exec_start
= params
->batch_obj_vm_offset
+
873 args
->batch_start_offset
;
875 ret
= engine
->emit_bb_start(params
->request
, exec_start
, params
->dispatch_flags
);
879 trace_i915_gem_ring_dispatch(params
->request
, params
->dispatch_flags
);
881 i915_gem_execbuffer_move_to_active(vmas
, params
->request
);
886 void intel_execlists_cancel_requests(struct intel_engine_cs
*engine
)
888 struct drm_i915_gem_request
*req
, *tmp
;
889 LIST_HEAD(cancel_list
);
891 WARN_ON(!mutex_is_locked(&engine
->i915
->dev
->struct_mutex
));
893 spin_lock_bh(&engine
->execlist_lock
);
894 list_replace_init(&engine
->execlist_queue
, &cancel_list
);
895 spin_unlock_bh(&engine
->execlist_lock
);
897 list_for_each_entry_safe(req
, tmp
, &cancel_list
, execlist_link
) {
898 list_del(&req
->execlist_link
);
899 i915_gem_request_unreference(req
);
903 void intel_logical_ring_stop(struct intel_engine_cs
*engine
)
905 struct drm_i915_private
*dev_priv
= engine
->i915
;
908 if (!intel_engine_initialized(engine
))
911 ret
= intel_engine_idle(engine
);
913 DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",
916 /* TODO: Is this correct with Execlists enabled? */
917 I915_WRITE_MODE(engine
, _MASKED_BIT_ENABLE(STOP_RING
));
918 if (wait_for((I915_READ_MODE(engine
) & MODE_IDLE
) != 0, 1000)) {
919 DRM_ERROR("%s :timed out trying to stop ring\n", engine
->name
);
922 I915_WRITE_MODE(engine
, _MASKED_BIT_DISABLE(STOP_RING
));
925 int logical_ring_flush_all_caches(struct drm_i915_gem_request
*req
)
927 struct intel_engine_cs
*engine
= req
->engine
;
930 if (!engine
->gpu_caches_dirty
)
933 ret
= engine
->emit_flush(req
, 0, I915_GEM_GPU_DOMAINS
);
937 engine
->gpu_caches_dirty
= false;
941 static int intel_lr_context_pin(struct i915_gem_context
*ctx
,
942 struct intel_engine_cs
*engine
)
944 struct drm_i915_private
*dev_priv
= ctx
->i915
;
945 struct intel_context
*ce
= &ctx
->engine
[engine
->id
];
950 lockdep_assert_held(&ctx
->i915
->dev
->struct_mutex
);
955 ret
= i915_gem_obj_ggtt_pin(ce
->state
, GEN8_LR_CONTEXT_ALIGN
,
956 PIN_OFFSET_BIAS
| GUC_WOPCM_TOP
);
960 vaddr
= i915_gem_object_pin_map(ce
->state
);
962 ret
= PTR_ERR(vaddr
);
966 lrc_reg_state
= vaddr
+ LRC_STATE_PN
* PAGE_SIZE
;
968 ret
= intel_pin_and_map_ringbuffer_obj(dev_priv
, ce
->ringbuf
);
972 i915_gem_context_reference(ctx
);
973 ce
->lrc_vma
= i915_gem_obj_to_ggtt(ce
->state
);
974 intel_lr_context_descriptor_update(ctx
, engine
);
976 lrc_reg_state
[CTX_RING_BUFFER_START
+1] = ce
->ringbuf
->vma
->node
.start
;
977 ce
->lrc_reg_state
= lrc_reg_state
;
978 ce
->state
->dirty
= true;
980 /* Invalidate GuC TLB. */
981 if (i915
.enable_guc_submission
)
982 I915_WRITE(GEN8_GTCR
, GEN8_GTCR_INVALIDATE
);
987 i915_gem_object_unpin_map(ce
->state
);
989 i915_gem_object_ggtt_unpin(ce
->state
);
995 void intel_lr_context_unpin(struct i915_gem_context
*ctx
,
996 struct intel_engine_cs
*engine
)
998 struct intel_context
*ce
= &ctx
->engine
[engine
->id
];
1000 lockdep_assert_held(&ctx
->i915
->dev
->struct_mutex
);
1001 GEM_BUG_ON(ce
->pin_count
== 0);
1003 if (--ce
->pin_count
)
1006 intel_unpin_ringbuffer_obj(ce
->ringbuf
);
1008 i915_gem_object_unpin_map(ce
->state
);
1009 i915_gem_object_ggtt_unpin(ce
->state
);
1013 ce
->lrc_reg_state
= NULL
;
1015 i915_gem_context_unreference(ctx
);
1018 static int intel_logical_ring_workarounds_emit(struct drm_i915_gem_request
*req
)
1021 struct intel_engine_cs
*engine
= req
->engine
;
1022 struct intel_ringbuffer
*ringbuf
= req
->ringbuf
;
1023 struct i915_workarounds
*w
= &req
->i915
->workarounds
;
1028 engine
->gpu_caches_dirty
= true;
1029 ret
= logical_ring_flush_all_caches(req
);
1033 ret
= intel_ring_begin(req
, w
->count
* 2 + 2);
1037 intel_logical_ring_emit(ringbuf
, MI_LOAD_REGISTER_IMM(w
->count
));
1038 for (i
= 0; i
< w
->count
; i
++) {
1039 intel_logical_ring_emit_reg(ringbuf
, w
->reg
[i
].addr
);
1040 intel_logical_ring_emit(ringbuf
, w
->reg
[i
].value
);
1042 intel_logical_ring_emit(ringbuf
, MI_NOOP
);
1044 intel_logical_ring_advance(ringbuf
);
1046 engine
->gpu_caches_dirty
= true;
1047 ret
= logical_ring_flush_all_caches(req
);
1054 #define wa_ctx_emit(batch, index, cmd) \
1056 int __index = (index)++; \
1057 if (WARN_ON(__index >= (PAGE_SIZE / sizeof(uint32_t)))) { \
1060 batch[__index] = (cmd); \
1063 #define wa_ctx_emit_reg(batch, index, reg) \
1064 wa_ctx_emit((batch), (index), i915_mmio_reg_offset(reg))
1067 * In this WA we need to set GEN8_L3SQCREG4[21:21] and reset it after
1068 * PIPE_CONTROL instruction. This is required for the flush to happen correctly
1069 * but there is a slight complication as this is applied in WA batch where the
1070 * values are only initialized once so we cannot take register value at the
1071 * beginning and reuse it further; hence we save its value to memory, upload a
1072 * constant value with bit21 set and then we restore it back with the saved value.
1073 * To simplify the WA, a constant value is formed by using the default value
1074 * of this register. This shouldn't be a problem because we are only modifying
1075 * it for a short period and this batch in non-premptible. We can ofcourse
1076 * use additional instructions that read the actual value of the register
1077 * at that time and set our bit of interest but it makes the WA complicated.
1079 * This WA is also required for Gen9 so extracting as a function avoids
1082 static inline int gen8_emit_flush_coherentl3_wa(struct intel_engine_cs
*engine
,
1083 uint32_t *const batch
,
1086 uint32_t l3sqc4_flush
= (0x40400000 | GEN8_LQSC_FLUSH_COHERENT_LINES
);
1089 * WaDisableLSQCROPERFforOCL:skl,kbl
1090 * This WA is implemented in skl_init_clock_gating() but since
1091 * this batch updates GEN8_L3SQCREG4 with default value we need to
1092 * set this bit here to retain the WA during flush.
1094 if (IS_SKL_REVID(engine
->i915
, 0, SKL_REVID_E0
) ||
1095 IS_KBL_REVID(engine
->i915
, 0, KBL_REVID_E0
))
1096 l3sqc4_flush
|= GEN8_LQSC_RO_PERF_DIS
;
1098 wa_ctx_emit(batch
, index
, (MI_STORE_REGISTER_MEM_GEN8
|
1099 MI_SRM_LRM_GLOBAL_GTT
));
1100 wa_ctx_emit_reg(batch
, index
, GEN8_L3SQCREG4
);
1101 wa_ctx_emit(batch
, index
, engine
->scratch
.gtt_offset
+ 256);
1102 wa_ctx_emit(batch
, index
, 0);
1104 wa_ctx_emit(batch
, index
, MI_LOAD_REGISTER_IMM(1));
1105 wa_ctx_emit_reg(batch
, index
, GEN8_L3SQCREG4
);
1106 wa_ctx_emit(batch
, index
, l3sqc4_flush
);
1108 wa_ctx_emit(batch
, index
, GFX_OP_PIPE_CONTROL(6));
1109 wa_ctx_emit(batch
, index
, (PIPE_CONTROL_CS_STALL
|
1110 PIPE_CONTROL_DC_FLUSH_ENABLE
));
1111 wa_ctx_emit(batch
, index
, 0);
1112 wa_ctx_emit(batch
, index
, 0);
1113 wa_ctx_emit(batch
, index
, 0);
1114 wa_ctx_emit(batch
, index
, 0);
1116 wa_ctx_emit(batch
, index
, (MI_LOAD_REGISTER_MEM_GEN8
|
1117 MI_SRM_LRM_GLOBAL_GTT
));
1118 wa_ctx_emit_reg(batch
, index
, GEN8_L3SQCREG4
);
1119 wa_ctx_emit(batch
, index
, engine
->scratch
.gtt_offset
+ 256);
1120 wa_ctx_emit(batch
, index
, 0);
1125 static inline uint32_t wa_ctx_start(struct i915_wa_ctx_bb
*wa_ctx
,
1127 uint32_t start_alignment
)
1129 return wa_ctx
->offset
= ALIGN(offset
, start_alignment
);
1132 static inline int wa_ctx_end(struct i915_wa_ctx_bb
*wa_ctx
,
1134 uint32_t size_alignment
)
1136 wa_ctx
->size
= offset
- wa_ctx
->offset
;
1138 WARN(wa_ctx
->size
% size_alignment
,
1139 "wa_ctx_bb failed sanity checks: size %d is not aligned to %d\n",
1140 wa_ctx
->size
, size_alignment
);
1145 * gen8_init_indirectctx_bb() - initialize indirect ctx batch with WA
1147 * @engine: only applicable for RCS
1148 * @wa_ctx: structure representing wa_ctx
1149 * offset: specifies start of the batch, should be cache-aligned. This is updated
1150 * with the offset value received as input.
1151 * size: size of the batch in DWORDS but HW expects in terms of cachelines
1152 * @batch: page in which WA are loaded
1153 * @offset: This field specifies the start of the batch, it should be
1154 * cache-aligned otherwise it is adjusted accordingly.
1155 * Typically we only have one indirect_ctx and per_ctx batch buffer which are
1156 * initialized at the beginning and shared across all contexts but this field
1157 * helps us to have multiple batches at different offsets and select them based
1158 * on a criteria. At the moment this batch always start at the beginning of the page
1159 * and at this point we don't have multiple wa_ctx batch buffers.
1161 * The number of WA applied are not known at the beginning; we use this field
1162 * to return the no of DWORDS written.
1164 * It is to be noted that this batch does not contain MI_BATCH_BUFFER_END
1165 * so it adds NOOPs as padding to make it cacheline aligned.
1166 * MI_BATCH_BUFFER_END will be added to perctx batch and both of them together
1167 * makes a complete batch buffer.
1169 * Return: non-zero if we exceed the PAGE_SIZE limit.
1172 static int gen8_init_indirectctx_bb(struct intel_engine_cs
*engine
,
1173 struct i915_wa_ctx_bb
*wa_ctx
,
1174 uint32_t *const batch
,
1177 uint32_t scratch_addr
;
1178 uint32_t index
= wa_ctx_start(wa_ctx
, *offset
, CACHELINE_DWORDS
);
1180 /* WaDisableCtxRestoreArbitration:bdw,chv */
1181 wa_ctx_emit(batch
, index
, MI_ARB_ON_OFF
| MI_ARB_DISABLE
);
1183 /* WaFlushCoherentL3CacheLinesAtContextSwitch:bdw */
1184 if (IS_BROADWELL(engine
->i915
)) {
1185 int rc
= gen8_emit_flush_coherentl3_wa(engine
, batch
, index
);
1191 /* WaClearSlmSpaceAtContextSwitch:bdw,chv */
1192 /* Actual scratch location is at 128 bytes offset */
1193 scratch_addr
= engine
->scratch
.gtt_offset
+ 2*CACHELINE_BYTES
;
1195 wa_ctx_emit(batch
, index
, GFX_OP_PIPE_CONTROL(6));
1196 wa_ctx_emit(batch
, index
, (PIPE_CONTROL_FLUSH_L3
|
1197 PIPE_CONTROL_GLOBAL_GTT_IVB
|
1198 PIPE_CONTROL_CS_STALL
|
1199 PIPE_CONTROL_QW_WRITE
));
1200 wa_ctx_emit(batch
, index
, scratch_addr
);
1201 wa_ctx_emit(batch
, index
, 0);
1202 wa_ctx_emit(batch
, index
, 0);
1203 wa_ctx_emit(batch
, index
, 0);
1205 /* Pad to end of cacheline */
1206 while (index
% CACHELINE_DWORDS
)
1207 wa_ctx_emit(batch
, index
, MI_NOOP
);
1210 * MI_BATCH_BUFFER_END is not required in Indirect ctx BB because
1211 * execution depends on the length specified in terms of cache lines
1212 * in the register CTX_RCS_INDIRECT_CTX
1215 return wa_ctx_end(wa_ctx
, *offset
= index
, CACHELINE_DWORDS
);
1219 * gen8_init_perctx_bb() - initialize per ctx batch with WA
1221 * @engine: only applicable for RCS
1222 * @wa_ctx: structure representing wa_ctx
1223 * offset: specifies start of the batch, should be cache-aligned.
1224 * size: size of the batch in DWORDS but HW expects in terms of cachelines
1225 * @batch: page in which WA are loaded
1226 * @offset: This field specifies the start of this batch.
1227 * This batch is started immediately after indirect_ctx batch. Since we ensure
1228 * that indirect_ctx ends on a cacheline this batch is aligned automatically.
1230 * The number of DWORDS written are returned using this field.
1232 * This batch is terminated with MI_BATCH_BUFFER_END and so we need not add padding
1233 * to align it with cacheline as padding after MI_BATCH_BUFFER_END is redundant.
1235 static int gen8_init_perctx_bb(struct intel_engine_cs
*engine
,
1236 struct i915_wa_ctx_bb
*wa_ctx
,
1237 uint32_t *const batch
,
1240 uint32_t index
= wa_ctx_start(wa_ctx
, *offset
, CACHELINE_DWORDS
);
1242 /* WaDisableCtxRestoreArbitration:bdw,chv */
1243 wa_ctx_emit(batch
, index
, MI_ARB_ON_OFF
| MI_ARB_ENABLE
);
1245 wa_ctx_emit(batch
, index
, MI_BATCH_BUFFER_END
);
1247 return wa_ctx_end(wa_ctx
, *offset
= index
, 1);
1250 static int gen9_init_indirectctx_bb(struct intel_engine_cs
*engine
,
1251 struct i915_wa_ctx_bb
*wa_ctx
,
1252 uint32_t *const batch
,
1256 uint32_t index
= wa_ctx_start(wa_ctx
, *offset
, CACHELINE_DWORDS
);
1258 /* WaDisableCtxRestoreArbitration:skl,bxt */
1259 if (IS_SKL_REVID(engine
->i915
, 0, SKL_REVID_D0
) ||
1260 IS_BXT_REVID(engine
->i915
, 0, BXT_REVID_A1
))
1261 wa_ctx_emit(batch
, index
, MI_ARB_ON_OFF
| MI_ARB_DISABLE
);
1263 /* WaFlushCoherentL3CacheLinesAtContextSwitch:skl,bxt */
1264 ret
= gen8_emit_flush_coherentl3_wa(engine
, batch
, index
);
1269 /* WaClearSlmSpaceAtContextSwitch:kbl */
1270 /* Actual scratch location is at 128 bytes offset */
1271 if (IS_KBL_REVID(engine
->i915
, 0, KBL_REVID_A0
)) {
1272 uint32_t scratch_addr
1273 = engine
->scratch
.gtt_offset
+ 2*CACHELINE_BYTES
;
1275 wa_ctx_emit(batch
, index
, GFX_OP_PIPE_CONTROL(6));
1276 wa_ctx_emit(batch
, index
, (PIPE_CONTROL_FLUSH_L3
|
1277 PIPE_CONTROL_GLOBAL_GTT_IVB
|
1278 PIPE_CONTROL_CS_STALL
|
1279 PIPE_CONTROL_QW_WRITE
));
1280 wa_ctx_emit(batch
, index
, scratch_addr
);
1281 wa_ctx_emit(batch
, index
, 0);
1282 wa_ctx_emit(batch
, index
, 0);
1283 wa_ctx_emit(batch
, index
, 0);
1285 /* Pad to end of cacheline */
1286 while (index
% CACHELINE_DWORDS
)
1287 wa_ctx_emit(batch
, index
, MI_NOOP
);
1289 return wa_ctx_end(wa_ctx
, *offset
= index
, CACHELINE_DWORDS
);
1292 static int gen9_init_perctx_bb(struct intel_engine_cs
*engine
,
1293 struct i915_wa_ctx_bb
*wa_ctx
,
1294 uint32_t *const batch
,
1297 uint32_t index
= wa_ctx_start(wa_ctx
, *offset
, CACHELINE_DWORDS
);
1299 /* WaSetDisablePixMaskCammingAndRhwoInCommonSliceChicken:skl,bxt */
1300 if (IS_SKL_REVID(engine
->i915
, 0, SKL_REVID_B0
) ||
1301 IS_BXT_REVID(engine
->i915
, 0, BXT_REVID_A1
)) {
1302 wa_ctx_emit(batch
, index
, MI_LOAD_REGISTER_IMM(1));
1303 wa_ctx_emit_reg(batch
, index
, GEN9_SLICE_COMMON_ECO_CHICKEN0
);
1304 wa_ctx_emit(batch
, index
,
1305 _MASKED_BIT_ENABLE(DISABLE_PIXEL_MASK_CAMMING
));
1306 wa_ctx_emit(batch
, index
, MI_NOOP
);
1309 /* WaClearTdlStateAckDirtyBits:bxt */
1310 if (IS_BXT_REVID(engine
->i915
, 0, BXT_REVID_B0
)) {
1311 wa_ctx_emit(batch
, index
, MI_LOAD_REGISTER_IMM(4));
1313 wa_ctx_emit_reg(batch
, index
, GEN8_STATE_ACK
);
1314 wa_ctx_emit(batch
, index
, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS
));
1316 wa_ctx_emit_reg(batch
, index
, GEN9_STATE_ACK_SLICE1
);
1317 wa_ctx_emit(batch
, index
, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS
));
1319 wa_ctx_emit_reg(batch
, index
, GEN9_STATE_ACK_SLICE2
);
1320 wa_ctx_emit(batch
, index
, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS
));
1322 wa_ctx_emit_reg(batch
, index
, GEN7_ROW_CHICKEN2
);
1323 /* dummy write to CS, mask bits are 0 to ensure the register is not modified */
1324 wa_ctx_emit(batch
, index
, 0x0);
1325 wa_ctx_emit(batch
, index
, MI_NOOP
);
1328 /* WaDisableCtxRestoreArbitration:skl,bxt */
1329 if (IS_SKL_REVID(engine
->i915
, 0, SKL_REVID_D0
) ||
1330 IS_BXT_REVID(engine
->i915
, 0, BXT_REVID_A1
))
1331 wa_ctx_emit(batch
, index
, MI_ARB_ON_OFF
| MI_ARB_ENABLE
);
1333 wa_ctx_emit(batch
, index
, MI_BATCH_BUFFER_END
);
1335 return wa_ctx_end(wa_ctx
, *offset
= index
, 1);
1338 static int lrc_setup_wa_ctx_obj(struct intel_engine_cs
*engine
, u32 size
)
1342 engine
->wa_ctx
.obj
= i915_gem_object_create(engine
->i915
->dev
,
1344 if (IS_ERR(engine
->wa_ctx
.obj
)) {
1345 DRM_DEBUG_DRIVER("alloc LRC WA ctx backing obj failed.\n");
1346 ret
= PTR_ERR(engine
->wa_ctx
.obj
);
1347 engine
->wa_ctx
.obj
= NULL
;
1351 ret
= i915_gem_obj_ggtt_pin(engine
->wa_ctx
.obj
, PAGE_SIZE
, 0);
1353 DRM_DEBUG_DRIVER("pin LRC WA ctx backing obj failed: %d\n",
1355 drm_gem_object_unreference(&engine
->wa_ctx
.obj
->base
);
1362 static void lrc_destroy_wa_ctx_obj(struct intel_engine_cs
*engine
)
1364 if (engine
->wa_ctx
.obj
) {
1365 i915_gem_object_ggtt_unpin(engine
->wa_ctx
.obj
);
1366 drm_gem_object_unreference(&engine
->wa_ctx
.obj
->base
);
1367 engine
->wa_ctx
.obj
= NULL
;
1371 static int intel_init_workaround_bb(struct intel_engine_cs
*engine
)
1377 struct i915_ctx_workarounds
*wa_ctx
= &engine
->wa_ctx
;
1379 WARN_ON(engine
->id
!= RCS
);
1381 /* update this when WA for higher Gen are added */
1382 if (INTEL_GEN(engine
->i915
) > 9) {
1383 DRM_ERROR("WA batch buffer is not initialized for Gen%d\n",
1384 INTEL_GEN(engine
->i915
));
1388 /* some WA perform writes to scratch page, ensure it is valid */
1389 if (engine
->scratch
.obj
== NULL
) {
1390 DRM_ERROR("scratch page not allocated for %s\n", engine
->name
);
1394 ret
= lrc_setup_wa_ctx_obj(engine
, PAGE_SIZE
);
1396 DRM_DEBUG_DRIVER("Failed to setup context WA page: %d\n", ret
);
1400 page
= i915_gem_object_get_dirty_page(wa_ctx
->obj
, 0);
1401 batch
= kmap_atomic(page
);
1404 if (IS_GEN8(engine
->i915
)) {
1405 ret
= gen8_init_indirectctx_bb(engine
,
1406 &wa_ctx
->indirect_ctx
,
1412 ret
= gen8_init_perctx_bb(engine
,
1418 } else if (IS_GEN9(engine
->i915
)) {
1419 ret
= gen9_init_indirectctx_bb(engine
,
1420 &wa_ctx
->indirect_ctx
,
1426 ret
= gen9_init_perctx_bb(engine
,
1435 kunmap_atomic(batch
);
1437 lrc_destroy_wa_ctx_obj(engine
);
1442 static void lrc_init_hws(struct intel_engine_cs
*engine
)
1444 struct drm_i915_private
*dev_priv
= engine
->i915
;
1446 I915_WRITE(RING_HWS_PGA(engine
->mmio_base
),
1447 (u32
)engine
->status_page
.gfx_addr
);
1448 POSTING_READ(RING_HWS_PGA(engine
->mmio_base
));
1451 static int gen8_init_common_ring(struct intel_engine_cs
*engine
)
1453 struct drm_i915_private
*dev_priv
= engine
->i915
;
1454 unsigned int next_context_status_buffer_hw
;
1456 lrc_init_hws(engine
);
1458 I915_WRITE_IMR(engine
,
1459 ~(engine
->irq_enable_mask
| engine
->irq_keep_mask
));
1460 I915_WRITE(RING_HWSTAM(engine
->mmio_base
), 0xffffffff);
1462 I915_WRITE(RING_MODE_GEN7(engine
),
1463 _MASKED_BIT_DISABLE(GFX_REPLAY_MODE
) |
1464 _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE
));
1465 POSTING_READ(RING_MODE_GEN7(engine
));
1468 * Instead of resetting the Context Status Buffer (CSB) read pointer to
1469 * zero, we need to read the write pointer from hardware and use its
1470 * value because "this register is power context save restored".
1471 * Effectively, these states have been observed:
1473 * | Suspend-to-idle (freeze) | Suspend-to-RAM (mem) |
1474 * BDW | CSB regs not reset | CSB regs reset |
1475 * CHT | CSB regs not reset | CSB regs not reset |
1479 next_context_status_buffer_hw
=
1480 GEN8_CSB_WRITE_PTR(I915_READ(RING_CONTEXT_STATUS_PTR(engine
)));
1483 * When the CSB registers are reset (also after power-up / gpu reset),
1484 * CSB write pointer is set to all 1's, which is not valid, use '5' in
1485 * this special case, so the first element read is CSB[0].
1487 if (next_context_status_buffer_hw
== GEN8_CSB_PTR_MASK
)
1488 next_context_status_buffer_hw
= (GEN8_CSB_ENTRIES
- 1);
1490 engine
->next_context_status_buffer
= next_context_status_buffer_hw
;
1491 DRM_DEBUG_DRIVER("Execlists enabled for %s\n", engine
->name
);
1493 intel_engine_init_hangcheck(engine
);
1495 return intel_mocs_init_engine(engine
);
1498 static int gen8_init_render_ring(struct intel_engine_cs
*engine
)
1500 struct drm_i915_private
*dev_priv
= engine
->i915
;
1503 ret
= gen8_init_common_ring(engine
);
1507 /* We need to disable the AsyncFlip performance optimisations in order
1508 * to use MI_WAIT_FOR_EVENT within the CS. It should already be
1509 * programmed to '1' on all products.
1511 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv
1513 I915_WRITE(MI_MODE
, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE
));
1515 I915_WRITE(INSTPM
, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING
));
1517 return init_workarounds_ring(engine
);
1520 static int gen9_init_render_ring(struct intel_engine_cs
*engine
)
1524 ret
= gen8_init_common_ring(engine
);
1528 return init_workarounds_ring(engine
);
1531 static int intel_logical_ring_emit_pdps(struct drm_i915_gem_request
*req
)
1533 struct i915_hw_ppgtt
*ppgtt
= req
->ctx
->ppgtt
;
1534 struct intel_engine_cs
*engine
= req
->engine
;
1535 struct intel_ringbuffer
*ringbuf
= req
->ringbuf
;
1536 const int num_lri_cmds
= GEN8_LEGACY_PDPES
* 2;
1539 ret
= intel_ring_begin(req
, num_lri_cmds
* 2 + 2);
1543 intel_logical_ring_emit(ringbuf
, MI_LOAD_REGISTER_IMM(num_lri_cmds
));
1544 for (i
= GEN8_LEGACY_PDPES
- 1; i
>= 0; i
--) {
1545 const dma_addr_t pd_daddr
= i915_page_dir_dma_addr(ppgtt
, i
);
1547 intel_logical_ring_emit_reg(ringbuf
,
1548 GEN8_RING_PDP_UDW(engine
, i
));
1549 intel_logical_ring_emit(ringbuf
, upper_32_bits(pd_daddr
));
1550 intel_logical_ring_emit_reg(ringbuf
,
1551 GEN8_RING_PDP_LDW(engine
, i
));
1552 intel_logical_ring_emit(ringbuf
, lower_32_bits(pd_daddr
));
1555 intel_logical_ring_emit(ringbuf
, MI_NOOP
);
1556 intel_logical_ring_advance(ringbuf
);
1561 static int gen8_emit_bb_start(struct drm_i915_gem_request
*req
,
1562 u64 offset
, unsigned dispatch_flags
)
1564 struct intel_ringbuffer
*ringbuf
= req
->ringbuf
;
1565 bool ppgtt
= !(dispatch_flags
& I915_DISPATCH_SECURE
);
1568 /* Don't rely in hw updating PDPs, specially in lite-restore.
1569 * Ideally, we should set Force PD Restore in ctx descriptor,
1570 * but we can't. Force Restore would be a second option, but
1571 * it is unsafe in case of lite-restore (because the ctx is
1572 * not idle). PML4 is allocated during ppgtt init so this is
1573 * not needed in 48-bit.*/
1574 if (req
->ctx
->ppgtt
&&
1575 (intel_engine_flag(req
->engine
) & req
->ctx
->ppgtt
->pd_dirty_rings
)) {
1576 if (!USES_FULL_48BIT_PPGTT(req
->i915
) &&
1577 !intel_vgpu_active(req
->i915
)) {
1578 ret
= intel_logical_ring_emit_pdps(req
);
1583 req
->ctx
->ppgtt
->pd_dirty_rings
&= ~intel_engine_flag(req
->engine
);
1586 ret
= intel_ring_begin(req
, 4);
1590 /* FIXME(BDW): Address space and security selectors. */
1591 intel_logical_ring_emit(ringbuf
, MI_BATCH_BUFFER_START_GEN8
|
1593 (dispatch_flags
& I915_DISPATCH_RS
?
1594 MI_BATCH_RESOURCE_STREAMER
: 0));
1595 intel_logical_ring_emit(ringbuf
, lower_32_bits(offset
));
1596 intel_logical_ring_emit(ringbuf
, upper_32_bits(offset
));
1597 intel_logical_ring_emit(ringbuf
, MI_NOOP
);
1598 intel_logical_ring_advance(ringbuf
);
1603 static bool gen8_logical_ring_get_irq(struct intel_engine_cs
*engine
)
1605 struct drm_i915_private
*dev_priv
= engine
->i915
;
1606 unsigned long flags
;
1608 if (WARN_ON(!intel_irqs_enabled(dev_priv
)))
1611 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1612 if (engine
->irq_refcount
++ == 0) {
1613 I915_WRITE_IMR(engine
,
1614 ~(engine
->irq_enable_mask
| engine
->irq_keep_mask
));
1615 POSTING_READ(RING_IMR(engine
->mmio_base
));
1617 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1622 static void gen8_logical_ring_put_irq(struct intel_engine_cs
*engine
)
1624 struct drm_i915_private
*dev_priv
= engine
->i915
;
1625 unsigned long flags
;
1627 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1628 if (--engine
->irq_refcount
== 0) {
1629 I915_WRITE_IMR(engine
, ~engine
->irq_keep_mask
);
1630 POSTING_READ(RING_IMR(engine
->mmio_base
));
1632 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1635 static int gen8_emit_flush(struct drm_i915_gem_request
*request
,
1636 u32 invalidate_domains
,
1639 struct intel_ringbuffer
*ringbuf
= request
->ringbuf
;
1640 struct intel_engine_cs
*engine
= ringbuf
->engine
;
1641 struct drm_i915_private
*dev_priv
= request
->i915
;
1645 ret
= intel_ring_begin(request
, 4);
1649 cmd
= MI_FLUSH_DW
+ 1;
1651 /* We always require a command barrier so that subsequent
1652 * commands, such as breadcrumb interrupts, are strictly ordered
1653 * wrt the contents of the write cache being flushed to memory
1654 * (and thus being coherent from the CPU).
1656 cmd
|= MI_FLUSH_DW_STORE_INDEX
| MI_FLUSH_DW_OP_STOREDW
;
1658 if (invalidate_domains
& I915_GEM_GPU_DOMAINS
) {
1659 cmd
|= MI_INVALIDATE_TLB
;
1660 if (engine
== &dev_priv
->engine
[VCS
])
1661 cmd
|= MI_INVALIDATE_BSD
;
1664 intel_logical_ring_emit(ringbuf
, cmd
);
1665 intel_logical_ring_emit(ringbuf
,
1666 I915_GEM_HWS_SCRATCH_ADDR
|
1667 MI_FLUSH_DW_USE_GTT
);
1668 intel_logical_ring_emit(ringbuf
, 0); /* upper addr */
1669 intel_logical_ring_emit(ringbuf
, 0); /* value */
1670 intel_logical_ring_advance(ringbuf
);
1675 static int gen8_emit_flush_render(struct drm_i915_gem_request
*request
,
1676 u32 invalidate_domains
,
1679 struct intel_ringbuffer
*ringbuf
= request
->ringbuf
;
1680 struct intel_engine_cs
*engine
= ringbuf
->engine
;
1681 u32 scratch_addr
= engine
->scratch
.gtt_offset
+ 2 * CACHELINE_BYTES
;
1682 bool vf_flush_wa
= false, dc_flush_wa
= false;
1687 flags
|= PIPE_CONTROL_CS_STALL
;
1689 if (flush_domains
) {
1690 flags
|= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH
;
1691 flags
|= PIPE_CONTROL_DEPTH_CACHE_FLUSH
;
1692 flags
|= PIPE_CONTROL_DC_FLUSH_ENABLE
;
1693 flags
|= PIPE_CONTROL_FLUSH_ENABLE
;
1696 if (invalidate_domains
) {
1697 flags
|= PIPE_CONTROL_TLB_INVALIDATE
;
1698 flags
|= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE
;
1699 flags
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
1700 flags
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
1701 flags
|= PIPE_CONTROL_CONST_CACHE_INVALIDATE
;
1702 flags
|= PIPE_CONTROL_STATE_CACHE_INVALIDATE
;
1703 flags
|= PIPE_CONTROL_QW_WRITE
;
1704 flags
|= PIPE_CONTROL_GLOBAL_GTT_IVB
;
1707 * On GEN9: before VF_CACHE_INVALIDATE we need to emit a NULL
1710 if (IS_GEN9(request
->i915
))
1713 /* WaForGAMHang:kbl */
1714 if (IS_KBL_REVID(request
->i915
, 0, KBL_REVID_B0
))
1726 ret
= intel_ring_begin(request
, len
);
1731 intel_logical_ring_emit(ringbuf
, GFX_OP_PIPE_CONTROL(6));
1732 intel_logical_ring_emit(ringbuf
, 0);
1733 intel_logical_ring_emit(ringbuf
, 0);
1734 intel_logical_ring_emit(ringbuf
, 0);
1735 intel_logical_ring_emit(ringbuf
, 0);
1736 intel_logical_ring_emit(ringbuf
, 0);
1740 intel_logical_ring_emit(ringbuf
, GFX_OP_PIPE_CONTROL(6));
1741 intel_logical_ring_emit(ringbuf
, PIPE_CONTROL_DC_FLUSH_ENABLE
);
1742 intel_logical_ring_emit(ringbuf
, 0);
1743 intel_logical_ring_emit(ringbuf
, 0);
1744 intel_logical_ring_emit(ringbuf
, 0);
1745 intel_logical_ring_emit(ringbuf
, 0);
1748 intel_logical_ring_emit(ringbuf
, GFX_OP_PIPE_CONTROL(6));
1749 intel_logical_ring_emit(ringbuf
, flags
);
1750 intel_logical_ring_emit(ringbuf
, scratch_addr
);
1751 intel_logical_ring_emit(ringbuf
, 0);
1752 intel_logical_ring_emit(ringbuf
, 0);
1753 intel_logical_ring_emit(ringbuf
, 0);
1756 intel_logical_ring_emit(ringbuf
, GFX_OP_PIPE_CONTROL(6));
1757 intel_logical_ring_emit(ringbuf
, PIPE_CONTROL_CS_STALL
);
1758 intel_logical_ring_emit(ringbuf
, 0);
1759 intel_logical_ring_emit(ringbuf
, 0);
1760 intel_logical_ring_emit(ringbuf
, 0);
1761 intel_logical_ring_emit(ringbuf
, 0);
1764 intel_logical_ring_advance(ringbuf
);
1769 static u32
gen8_get_seqno(struct intel_engine_cs
*engine
)
1771 return intel_read_status_page(engine
, I915_GEM_HWS_INDEX
);
1774 static void gen8_set_seqno(struct intel_engine_cs
*engine
, u32 seqno
)
1776 intel_write_status_page(engine
, I915_GEM_HWS_INDEX
, seqno
);
1779 static void bxt_a_seqno_barrier(struct intel_engine_cs
*engine
)
1782 * On BXT A steppings there is a HW coherency issue whereby the
1783 * MI_STORE_DATA_IMM storing the completed request's seqno
1784 * occasionally doesn't invalidate the CPU cache. Work around this by
1785 * clflushing the corresponding cacheline whenever the caller wants
1786 * the coherency to be guaranteed. Note that this cacheline is known
1787 * to be clean at this point, since we only write it in
1788 * bxt_a_set_seqno(), where we also do a clflush after the write. So
1789 * this clflush in practice becomes an invalidate operation.
1791 intel_flush_status_page(engine
, I915_GEM_HWS_INDEX
);
1794 static void bxt_a_set_seqno(struct intel_engine_cs
*engine
, u32 seqno
)
1796 intel_write_status_page(engine
, I915_GEM_HWS_INDEX
, seqno
);
1798 /* See bxt_a_get_seqno() explaining the reason for the clflush. */
1799 intel_flush_status_page(engine
, I915_GEM_HWS_INDEX
);
1803 * Reserve space for 2 NOOPs at the end of each request to be
1804 * used as a workaround for not being allowed to do lite
1805 * restore with HEAD==TAIL (WaIdleLiteRestore).
1807 #define WA_TAIL_DWORDS 2
1809 static int gen8_emit_request(struct drm_i915_gem_request
*request
)
1811 struct intel_ringbuffer
*ringbuf
= request
->ringbuf
;
1814 ret
= intel_ring_begin(request
, 6 + WA_TAIL_DWORDS
);
1818 /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */
1819 BUILD_BUG_ON(I915_GEM_HWS_INDEX_ADDR
& (1 << 5));
1821 intel_logical_ring_emit(ringbuf
,
1822 (MI_FLUSH_DW
+ 1) | MI_FLUSH_DW_OP_STOREDW
);
1823 intel_logical_ring_emit(ringbuf
,
1824 intel_hws_seqno_address(request
->engine
) |
1825 MI_FLUSH_DW_USE_GTT
);
1826 intel_logical_ring_emit(ringbuf
, 0);
1827 intel_logical_ring_emit(ringbuf
, i915_gem_request_get_seqno(request
));
1828 intel_logical_ring_emit(ringbuf
, MI_USER_INTERRUPT
);
1829 intel_logical_ring_emit(ringbuf
, MI_NOOP
);
1830 return intel_logical_ring_advance_and_submit(request
);
1833 static int gen8_emit_request_render(struct drm_i915_gem_request
*request
)
1835 struct intel_ringbuffer
*ringbuf
= request
->ringbuf
;
1838 ret
= intel_ring_begin(request
, 8 + WA_TAIL_DWORDS
);
1842 /* We're using qword write, seqno should be aligned to 8 bytes. */
1843 BUILD_BUG_ON(I915_GEM_HWS_INDEX
& 1);
1845 /* w/a for post sync ops following a GPGPU operation we
1846 * need a prior CS_STALL, which is emitted by the flush
1847 * following the batch.
1849 intel_logical_ring_emit(ringbuf
, GFX_OP_PIPE_CONTROL(6));
1850 intel_logical_ring_emit(ringbuf
,
1851 (PIPE_CONTROL_GLOBAL_GTT_IVB
|
1852 PIPE_CONTROL_CS_STALL
|
1853 PIPE_CONTROL_QW_WRITE
));
1854 intel_logical_ring_emit(ringbuf
,
1855 intel_hws_seqno_address(request
->engine
));
1856 intel_logical_ring_emit(ringbuf
, 0);
1857 intel_logical_ring_emit(ringbuf
, i915_gem_request_get_seqno(request
));
1858 /* We're thrashing one dword of HWS. */
1859 intel_logical_ring_emit(ringbuf
, 0);
1860 intel_logical_ring_emit(ringbuf
, MI_USER_INTERRUPT
);
1861 intel_logical_ring_emit(ringbuf
, MI_NOOP
);
1862 return intel_logical_ring_advance_and_submit(request
);
1865 static int intel_lr_context_render_state_init(struct drm_i915_gem_request
*req
)
1867 struct render_state so
;
1870 ret
= i915_gem_render_state_prepare(req
->engine
, &so
);
1874 if (so
.rodata
== NULL
)
1877 ret
= req
->engine
->emit_bb_start(req
, so
.ggtt_offset
,
1878 I915_DISPATCH_SECURE
);
1882 ret
= req
->engine
->emit_bb_start(req
,
1883 (so
.ggtt_offset
+ so
.aux_batch_offset
),
1884 I915_DISPATCH_SECURE
);
1888 i915_vma_move_to_active(i915_gem_obj_to_ggtt(so
.obj
), req
);
1891 i915_gem_render_state_fini(&so
);
1895 static int gen8_init_rcs_context(struct drm_i915_gem_request
*req
)
1899 ret
= intel_logical_ring_workarounds_emit(req
);
1903 ret
= intel_rcs_context_init_mocs(req
);
1905 * Failing to program the MOCS is non-fatal.The system will not
1906 * run at peak performance. So generate an error and carry on.
1909 DRM_ERROR("MOCS failed to program: expect performance issues.\n");
1911 return intel_lr_context_render_state_init(req
);
1915 * intel_logical_ring_cleanup() - deallocate the Engine Command Streamer
1917 * @engine: Engine Command Streamer.
1920 void intel_logical_ring_cleanup(struct intel_engine_cs
*engine
)
1922 struct drm_i915_private
*dev_priv
;
1924 if (!intel_engine_initialized(engine
))
1928 * Tasklet cannot be active at this point due intel_mark_active/idle
1929 * so this is just for documentation.
1931 if (WARN_ON(test_bit(TASKLET_STATE_SCHED
, &engine
->irq_tasklet
.state
)))
1932 tasklet_kill(&engine
->irq_tasklet
);
1934 dev_priv
= engine
->i915
;
1936 if (engine
->buffer
) {
1937 intel_logical_ring_stop(engine
);
1938 WARN_ON((I915_READ_MODE(engine
) & MODE_IDLE
) == 0);
1941 if (engine
->cleanup
)
1942 engine
->cleanup(engine
);
1944 i915_cmd_parser_fini_ring(engine
);
1945 i915_gem_batch_pool_fini(&engine
->batch_pool
);
1947 if (engine
->status_page
.obj
) {
1948 i915_gem_object_unpin_map(engine
->status_page
.obj
);
1949 engine
->status_page
.obj
= NULL
;
1951 intel_lr_context_unpin(dev_priv
->kernel_context
, engine
);
1953 engine
->idle_lite_restore_wa
= 0;
1954 engine
->disable_lite_restore_wa
= false;
1955 engine
->ctx_desc_template
= 0;
1957 lrc_destroy_wa_ctx_obj(engine
);
1958 engine
->i915
= NULL
;
1962 logical_ring_default_vfuncs(struct intel_engine_cs
*engine
)
1964 /* Default vfuncs which can be overriden by each engine. */
1965 engine
->init_hw
= gen8_init_common_ring
;
1966 engine
->emit_request
= gen8_emit_request
;
1967 engine
->emit_flush
= gen8_emit_flush
;
1968 engine
->irq_get
= gen8_logical_ring_get_irq
;
1969 engine
->irq_put
= gen8_logical_ring_put_irq
;
1970 engine
->emit_bb_start
= gen8_emit_bb_start
;
1971 engine
->get_seqno
= gen8_get_seqno
;
1972 engine
->set_seqno
= gen8_set_seqno
;
1973 if (IS_BXT_REVID(engine
->i915
, 0, BXT_REVID_A1
)) {
1974 engine
->irq_seqno_barrier
= bxt_a_seqno_barrier
;
1975 engine
->set_seqno
= bxt_a_set_seqno
;
1980 logical_ring_default_irqs(struct intel_engine_cs
*engine
, unsigned shift
)
1982 engine
->irq_enable_mask
= GT_RENDER_USER_INTERRUPT
<< shift
;
1983 engine
->irq_keep_mask
= GT_CONTEXT_SWITCH_INTERRUPT
<< shift
;
1984 init_waitqueue_head(&engine
->irq_queue
);
1988 lrc_setup_hws(struct intel_engine_cs
*engine
,
1989 struct drm_i915_gem_object
*dctx_obj
)
1993 /* The HWSP is part of the default context object in LRC mode. */
1994 engine
->status_page
.gfx_addr
= i915_gem_obj_ggtt_offset(dctx_obj
) +
1995 LRC_PPHWSP_PN
* PAGE_SIZE
;
1996 hws
= i915_gem_object_pin_map(dctx_obj
);
1998 return PTR_ERR(hws
);
1999 engine
->status_page
.page_addr
= hws
+ LRC_PPHWSP_PN
* PAGE_SIZE
;
2000 engine
->status_page
.obj
= dctx_obj
;
2005 static const struct logical_ring_info
{
2011 } logical_rings
[] = {
2013 .name
= "render ring",
2014 .exec_id
= I915_EXEC_RENDER
,
2015 .guc_id
= GUC_RENDER_ENGINE
,
2016 .mmio_base
= RENDER_RING_BASE
,
2017 .irq_shift
= GEN8_RCS_IRQ_SHIFT
,
2020 .name
= "blitter ring",
2021 .exec_id
= I915_EXEC_BLT
,
2022 .guc_id
= GUC_BLITTER_ENGINE
,
2023 .mmio_base
= BLT_RING_BASE
,
2024 .irq_shift
= GEN8_BCS_IRQ_SHIFT
,
2028 .exec_id
= I915_EXEC_BSD
,
2029 .guc_id
= GUC_VIDEO_ENGINE
,
2030 .mmio_base
= GEN6_BSD_RING_BASE
,
2031 .irq_shift
= GEN8_VCS1_IRQ_SHIFT
,
2034 .name
= "bsd2 ring",
2035 .exec_id
= I915_EXEC_BSD
,
2036 .guc_id
= GUC_VIDEO_ENGINE2
,
2037 .mmio_base
= GEN8_BSD2_RING_BASE
,
2038 .irq_shift
= GEN8_VCS2_IRQ_SHIFT
,
2041 .name
= "video enhancement ring",
2042 .exec_id
= I915_EXEC_VEBOX
,
2043 .guc_id
= GUC_VIDEOENHANCE_ENGINE
,
2044 .mmio_base
= VEBOX_RING_BASE
,
2045 .irq_shift
= GEN8_VECS_IRQ_SHIFT
,
2049 static struct intel_engine_cs
*
2050 logical_ring_setup(struct drm_device
*dev
, enum intel_engine_id id
)
2052 const struct logical_ring_info
*info
= &logical_rings
[id
];
2053 struct drm_i915_private
*dev_priv
= to_i915(dev
);
2054 struct intel_engine_cs
*engine
= &dev_priv
->engine
[id
];
2055 enum forcewake_domains fw_domains
;
2058 engine
->name
= info
->name
;
2059 engine
->exec_id
= info
->exec_id
;
2060 engine
->guc_id
= info
->guc_id
;
2061 engine
->mmio_base
= info
->mmio_base
;
2063 engine
->i915
= dev_priv
;
2065 /* Intentionally left blank. */
2066 engine
->buffer
= NULL
;
2068 fw_domains
= intel_uncore_forcewake_for_reg(dev_priv
,
2072 fw_domains
|= intel_uncore_forcewake_for_reg(dev_priv
,
2073 RING_CONTEXT_STATUS_PTR(engine
),
2074 FW_REG_READ
| FW_REG_WRITE
);
2076 fw_domains
|= intel_uncore_forcewake_for_reg(dev_priv
,
2077 RING_CONTEXT_STATUS_BUF_BASE(engine
),
2080 engine
->fw_domains
= fw_domains
;
2082 INIT_LIST_HEAD(&engine
->active_list
);
2083 INIT_LIST_HEAD(&engine
->request_list
);
2084 INIT_LIST_HEAD(&engine
->buffers
);
2085 INIT_LIST_HEAD(&engine
->execlist_queue
);
2086 spin_lock_init(&engine
->execlist_lock
);
2088 tasklet_init(&engine
->irq_tasklet
,
2089 intel_lrc_irq_handler
, (unsigned long)engine
);
2091 logical_ring_init_platform_invariants(engine
);
2092 logical_ring_default_vfuncs(engine
);
2093 logical_ring_default_irqs(engine
, info
->irq_shift
);
2095 intel_engine_init_hangcheck(engine
);
2096 i915_gem_batch_pool_init(dev
, &engine
->batch_pool
);
2102 logical_ring_init(struct intel_engine_cs
*engine
)
2104 struct i915_gem_context
*dctx
= engine
->i915
->kernel_context
;
2107 ret
= i915_cmd_parser_init_ring(engine
);
2111 ret
= execlists_context_deferred_alloc(dctx
, engine
);
2115 /* As this is the default context, always pin it */
2116 ret
= intel_lr_context_pin(dctx
, engine
);
2118 DRM_ERROR("Failed to pin context for %s: %d\n",
2123 /* And setup the hardware status page. */
2124 ret
= lrc_setup_hws(engine
, dctx
->engine
[engine
->id
].state
);
2126 DRM_ERROR("Failed to set up hws %s: %d\n", engine
->name
, ret
);
2133 intel_logical_ring_cleanup(engine
);
2137 static int logical_render_ring_init(struct drm_device
*dev
)
2139 struct intel_engine_cs
*engine
= logical_ring_setup(dev
, RCS
);
2142 if (HAS_L3_DPF(dev
))
2143 engine
->irq_keep_mask
|= GT_RENDER_L3_PARITY_ERROR_INTERRUPT
;
2145 /* Override some for render ring. */
2146 if (INTEL_INFO(dev
)->gen
>= 9)
2147 engine
->init_hw
= gen9_init_render_ring
;
2149 engine
->init_hw
= gen8_init_render_ring
;
2150 engine
->init_context
= gen8_init_rcs_context
;
2151 engine
->cleanup
= intel_fini_pipe_control
;
2152 engine
->emit_flush
= gen8_emit_flush_render
;
2153 engine
->emit_request
= gen8_emit_request_render
;
2155 ret
= intel_init_pipe_control(engine
);
2159 ret
= intel_init_workaround_bb(engine
);
2162 * We continue even if we fail to initialize WA batch
2163 * because we only expect rare glitches but nothing
2164 * critical to prevent us from using GPU
2166 DRM_ERROR("WA batch buffer initialization failed: %d\n",
2170 ret
= logical_ring_init(engine
);
2172 lrc_destroy_wa_ctx_obj(engine
);
2178 static int logical_bsd_ring_init(struct drm_device
*dev
)
2180 struct intel_engine_cs
*engine
= logical_ring_setup(dev
, VCS
);
2182 return logical_ring_init(engine
);
2185 static int logical_bsd2_ring_init(struct drm_device
*dev
)
2187 struct intel_engine_cs
*engine
= logical_ring_setup(dev
, VCS2
);
2189 return logical_ring_init(engine
);
2192 static int logical_blt_ring_init(struct drm_device
*dev
)
2194 struct intel_engine_cs
*engine
= logical_ring_setup(dev
, BCS
);
2196 return logical_ring_init(engine
);
2199 static int logical_vebox_ring_init(struct drm_device
*dev
)
2201 struct intel_engine_cs
*engine
= logical_ring_setup(dev
, VECS
);
2203 return logical_ring_init(engine
);
2207 * intel_logical_rings_init() - allocate, populate and init the Engine Command Streamers
2210 * This function inits the engines for an Execlists submission style (the equivalent in the
2211 * legacy ringbuffer submission world would be i915_gem_init_engines). It does it only for
2212 * those engines that are present in the hardware.
2214 * Return: non-zero if the initialization failed.
2216 int intel_logical_rings_init(struct drm_device
*dev
)
2218 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2221 ret
= logical_render_ring_init(dev
);
2226 ret
= logical_bsd_ring_init(dev
);
2228 goto cleanup_render_ring
;
2232 ret
= logical_blt_ring_init(dev
);
2234 goto cleanup_bsd_ring
;
2237 if (HAS_VEBOX(dev
)) {
2238 ret
= logical_vebox_ring_init(dev
);
2240 goto cleanup_blt_ring
;
2243 if (HAS_BSD2(dev
)) {
2244 ret
= logical_bsd2_ring_init(dev
);
2246 goto cleanup_vebox_ring
;
2252 intel_logical_ring_cleanup(&dev_priv
->engine
[VECS
]);
2254 intel_logical_ring_cleanup(&dev_priv
->engine
[BCS
]);
2256 intel_logical_ring_cleanup(&dev_priv
->engine
[VCS
]);
2257 cleanup_render_ring
:
2258 intel_logical_ring_cleanup(&dev_priv
->engine
[RCS
]);
2264 make_rpcs(struct drm_i915_private
*dev_priv
)
2269 * No explicit RPCS request is needed to ensure full
2270 * slice/subslice/EU enablement prior to Gen9.
2272 if (INTEL_GEN(dev_priv
) < 9)
2276 * Starting in Gen9, render power gating can leave
2277 * slice/subslice/EU in a partially enabled state. We
2278 * must make an explicit request through RPCS for full
2281 if (INTEL_INFO(dev_priv
)->has_slice_pg
) {
2282 rpcs
|= GEN8_RPCS_S_CNT_ENABLE
;
2283 rpcs
|= INTEL_INFO(dev_priv
)->slice_total
<<
2284 GEN8_RPCS_S_CNT_SHIFT
;
2285 rpcs
|= GEN8_RPCS_ENABLE
;
2288 if (INTEL_INFO(dev_priv
)->has_subslice_pg
) {
2289 rpcs
|= GEN8_RPCS_SS_CNT_ENABLE
;
2290 rpcs
|= INTEL_INFO(dev_priv
)->subslice_per_slice
<<
2291 GEN8_RPCS_SS_CNT_SHIFT
;
2292 rpcs
|= GEN8_RPCS_ENABLE
;
2295 if (INTEL_INFO(dev_priv
)->has_eu_pg
) {
2296 rpcs
|= INTEL_INFO(dev_priv
)->eu_per_subslice
<<
2297 GEN8_RPCS_EU_MIN_SHIFT
;
2298 rpcs
|= INTEL_INFO(dev_priv
)->eu_per_subslice
<<
2299 GEN8_RPCS_EU_MAX_SHIFT
;
2300 rpcs
|= GEN8_RPCS_ENABLE
;
2306 static u32
intel_lr_indirect_ctx_offset(struct intel_engine_cs
*engine
)
2308 u32 indirect_ctx_offset
;
2310 switch (INTEL_GEN(engine
->i915
)) {
2312 MISSING_CASE(INTEL_GEN(engine
->i915
));
2315 indirect_ctx_offset
=
2316 GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT
;
2319 indirect_ctx_offset
=
2320 GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT
;
2324 return indirect_ctx_offset
;
2328 populate_lr_context(struct i915_gem_context
*ctx
,
2329 struct drm_i915_gem_object
*ctx_obj
,
2330 struct intel_engine_cs
*engine
,
2331 struct intel_ringbuffer
*ringbuf
)
2333 struct drm_i915_private
*dev_priv
= ctx
->i915
;
2334 struct i915_hw_ppgtt
*ppgtt
= ctx
->ppgtt
;
2340 ppgtt
= dev_priv
->mm
.aliasing_ppgtt
;
2342 ret
= i915_gem_object_set_to_cpu_domain(ctx_obj
, true);
2344 DRM_DEBUG_DRIVER("Could not set to CPU domain\n");
2348 vaddr
= i915_gem_object_pin_map(ctx_obj
);
2349 if (IS_ERR(vaddr
)) {
2350 ret
= PTR_ERR(vaddr
);
2351 DRM_DEBUG_DRIVER("Could not map object pages! (%d)\n", ret
);
2354 ctx_obj
->dirty
= true;
2356 /* The second page of the context object contains some fields which must
2357 * be set up prior to the first execution. */
2358 reg_state
= vaddr
+ LRC_STATE_PN
* PAGE_SIZE
;
2360 /* A context is actually a big batch buffer with several MI_LOAD_REGISTER_IMM
2361 * commands followed by (reg, value) pairs. The values we are setting here are
2362 * only for the first context restore: on a subsequent save, the GPU will
2363 * recreate this batchbuffer with new values (including all the missing
2364 * MI_LOAD_REGISTER_IMM commands that we are not initializing here). */
2365 reg_state
[CTX_LRI_HEADER_0
] =
2366 MI_LOAD_REGISTER_IMM(engine
->id
== RCS
? 14 : 11) | MI_LRI_FORCE_POSTED
;
2367 ASSIGN_CTX_REG(reg_state
, CTX_CONTEXT_CONTROL
,
2368 RING_CONTEXT_CONTROL(engine
),
2369 _MASKED_BIT_ENABLE(CTX_CTRL_INHIBIT_SYN_CTX_SWITCH
|
2370 CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT
|
2371 (HAS_RESOURCE_STREAMER(dev_priv
) ?
2372 CTX_CTRL_RS_CTX_ENABLE
: 0)));
2373 ASSIGN_CTX_REG(reg_state
, CTX_RING_HEAD
, RING_HEAD(engine
->mmio_base
),
2375 ASSIGN_CTX_REG(reg_state
, CTX_RING_TAIL
, RING_TAIL(engine
->mmio_base
),
2377 /* Ring buffer start address is not known until the buffer is pinned.
2378 * It is written to the context image in execlists_update_context()
2380 ASSIGN_CTX_REG(reg_state
, CTX_RING_BUFFER_START
,
2381 RING_START(engine
->mmio_base
), 0);
2382 ASSIGN_CTX_REG(reg_state
, CTX_RING_BUFFER_CONTROL
,
2383 RING_CTL(engine
->mmio_base
),
2384 ((ringbuf
->size
- PAGE_SIZE
) & RING_NR_PAGES
) | RING_VALID
);
2385 ASSIGN_CTX_REG(reg_state
, CTX_BB_HEAD_U
,
2386 RING_BBADDR_UDW(engine
->mmio_base
), 0);
2387 ASSIGN_CTX_REG(reg_state
, CTX_BB_HEAD_L
,
2388 RING_BBADDR(engine
->mmio_base
), 0);
2389 ASSIGN_CTX_REG(reg_state
, CTX_BB_STATE
,
2390 RING_BBSTATE(engine
->mmio_base
),
2392 ASSIGN_CTX_REG(reg_state
, CTX_SECOND_BB_HEAD_U
,
2393 RING_SBBADDR_UDW(engine
->mmio_base
), 0);
2394 ASSIGN_CTX_REG(reg_state
, CTX_SECOND_BB_HEAD_L
,
2395 RING_SBBADDR(engine
->mmio_base
), 0);
2396 ASSIGN_CTX_REG(reg_state
, CTX_SECOND_BB_STATE
,
2397 RING_SBBSTATE(engine
->mmio_base
), 0);
2398 if (engine
->id
== RCS
) {
2399 ASSIGN_CTX_REG(reg_state
, CTX_BB_PER_CTX_PTR
,
2400 RING_BB_PER_CTX_PTR(engine
->mmio_base
), 0);
2401 ASSIGN_CTX_REG(reg_state
, CTX_RCS_INDIRECT_CTX
,
2402 RING_INDIRECT_CTX(engine
->mmio_base
), 0);
2403 ASSIGN_CTX_REG(reg_state
, CTX_RCS_INDIRECT_CTX_OFFSET
,
2404 RING_INDIRECT_CTX_OFFSET(engine
->mmio_base
), 0);
2405 if (engine
->wa_ctx
.obj
) {
2406 struct i915_ctx_workarounds
*wa_ctx
= &engine
->wa_ctx
;
2407 uint32_t ggtt_offset
= i915_gem_obj_ggtt_offset(wa_ctx
->obj
);
2409 reg_state
[CTX_RCS_INDIRECT_CTX
+1] =
2410 (ggtt_offset
+ wa_ctx
->indirect_ctx
.offset
* sizeof(uint32_t)) |
2411 (wa_ctx
->indirect_ctx
.size
/ CACHELINE_DWORDS
);
2413 reg_state
[CTX_RCS_INDIRECT_CTX_OFFSET
+1] =
2414 intel_lr_indirect_ctx_offset(engine
) << 6;
2416 reg_state
[CTX_BB_PER_CTX_PTR
+1] =
2417 (ggtt_offset
+ wa_ctx
->per_ctx
.offset
* sizeof(uint32_t)) |
2421 reg_state
[CTX_LRI_HEADER_1
] = MI_LOAD_REGISTER_IMM(9) | MI_LRI_FORCE_POSTED
;
2422 ASSIGN_CTX_REG(reg_state
, CTX_CTX_TIMESTAMP
,
2423 RING_CTX_TIMESTAMP(engine
->mmio_base
), 0);
2424 /* PDP values well be assigned later if needed */
2425 ASSIGN_CTX_REG(reg_state
, CTX_PDP3_UDW
, GEN8_RING_PDP_UDW(engine
, 3),
2427 ASSIGN_CTX_REG(reg_state
, CTX_PDP3_LDW
, GEN8_RING_PDP_LDW(engine
, 3),
2429 ASSIGN_CTX_REG(reg_state
, CTX_PDP2_UDW
, GEN8_RING_PDP_UDW(engine
, 2),
2431 ASSIGN_CTX_REG(reg_state
, CTX_PDP2_LDW
, GEN8_RING_PDP_LDW(engine
, 2),
2433 ASSIGN_CTX_REG(reg_state
, CTX_PDP1_UDW
, GEN8_RING_PDP_UDW(engine
, 1),
2435 ASSIGN_CTX_REG(reg_state
, CTX_PDP1_LDW
, GEN8_RING_PDP_LDW(engine
, 1),
2437 ASSIGN_CTX_REG(reg_state
, CTX_PDP0_UDW
, GEN8_RING_PDP_UDW(engine
, 0),
2439 ASSIGN_CTX_REG(reg_state
, CTX_PDP0_LDW
, GEN8_RING_PDP_LDW(engine
, 0),
2442 if (USES_FULL_48BIT_PPGTT(ppgtt
->base
.dev
)) {
2443 /* 64b PPGTT (48bit canonical)
2444 * PDP0_DESCRIPTOR contains the base address to PML4 and
2445 * other PDP Descriptors are ignored.
2447 ASSIGN_CTX_PML4(ppgtt
, reg_state
);
2450 * PDP*_DESCRIPTOR contains the base address of space supported.
2451 * With dynamic page allocation, PDPs may not be allocated at
2452 * this point. Point the unallocated PDPs to the scratch page
2454 execlists_update_context_pdps(ppgtt
, reg_state
);
2457 if (engine
->id
== RCS
) {
2458 reg_state
[CTX_LRI_HEADER_2
] = MI_LOAD_REGISTER_IMM(1);
2459 ASSIGN_CTX_REG(reg_state
, CTX_R_PWR_CLK_STATE
, GEN8_R_PWR_CLK_STATE
,
2460 make_rpcs(dev_priv
));
2463 i915_gem_object_unpin_map(ctx_obj
);
2469 * intel_lr_context_size() - return the size of the context for an engine
2470 * @engine: which engine to find the context size for
2472 * Each engine may require a different amount of space for a context image,
2473 * so when allocating (or copying) an image, this function can be used to
2474 * find the right size for the specific engine.
2476 * Return: size (in bytes) of an engine-specific context image
2478 * Note: this size includes the HWSP, which is part of the context image
2479 * in LRC mode, but does not include the "shared data page" used with
2480 * GuC submission. The caller should account for this if using the GuC.
2482 uint32_t intel_lr_context_size(struct intel_engine_cs
*engine
)
2486 WARN_ON(INTEL_GEN(engine
->i915
) < 8);
2488 switch (engine
->id
) {
2490 if (INTEL_GEN(engine
->i915
) >= 9)
2491 ret
= GEN9_LR_CONTEXT_RENDER_SIZE
;
2493 ret
= GEN8_LR_CONTEXT_RENDER_SIZE
;
2499 ret
= GEN8_LR_CONTEXT_OTHER_SIZE
;
2507 * execlists_context_deferred_alloc() - create the LRC specific bits of a context
2508 * @ctx: LR context to create.
2509 * @engine: engine to be used with the context.
2511 * This function can be called more than once, with different engines, if we plan
2512 * to use the context with them. The context backing objects and the ringbuffers
2513 * (specially the ringbuffer backing objects) suck a lot of memory up, and that's why
2514 * the creation is a deferred call: it's better to make sure first that we need to use
2515 * a given ring with the context.
2517 * Return: non-zero on error.
2519 static int execlists_context_deferred_alloc(struct i915_gem_context
*ctx
,
2520 struct intel_engine_cs
*engine
)
2522 struct drm_i915_gem_object
*ctx_obj
;
2523 struct intel_context
*ce
= &ctx
->engine
[engine
->id
];
2524 uint32_t context_size
;
2525 struct intel_ringbuffer
*ringbuf
;
2530 context_size
= round_up(intel_lr_context_size(engine
), 4096);
2532 /* One extra page as the sharing data between driver and GuC */
2533 context_size
+= PAGE_SIZE
* LRC_PPHWSP_PN
;
2535 ctx_obj
= i915_gem_object_create(ctx
->i915
->dev
, context_size
);
2536 if (IS_ERR(ctx_obj
)) {
2537 DRM_DEBUG_DRIVER("Alloc LRC backing obj failed.\n");
2538 return PTR_ERR(ctx_obj
);
2541 ringbuf
= intel_engine_create_ringbuffer(engine
, ctx
->ring_size
);
2542 if (IS_ERR(ringbuf
)) {
2543 ret
= PTR_ERR(ringbuf
);
2544 goto error_deref_obj
;
2547 ret
= populate_lr_context(ctx
, ctx_obj
, engine
, ringbuf
);
2549 DRM_DEBUG_DRIVER("Failed to populate LRC: %d\n", ret
);
2553 ce
->ringbuf
= ringbuf
;
2554 ce
->state
= ctx_obj
;
2555 ce
->initialised
= engine
->init_context
== NULL
;
2560 intel_ringbuffer_free(ringbuf
);
2562 drm_gem_object_unreference(&ctx_obj
->base
);
2568 void intel_lr_context_reset(struct drm_i915_private
*dev_priv
,
2569 struct i915_gem_context
*ctx
)
2571 struct intel_engine_cs
*engine
;
2573 for_each_engine(engine
, dev_priv
) {
2574 struct intel_context
*ce
= &ctx
->engine
[engine
->id
];
2575 struct drm_i915_gem_object
*ctx_obj
= ce
->state
;
2577 uint32_t *reg_state
;
2582 vaddr
= i915_gem_object_pin_map(ctx_obj
);
2583 if (WARN_ON(IS_ERR(vaddr
)))
2586 reg_state
= vaddr
+ LRC_STATE_PN
* PAGE_SIZE
;
2587 ctx_obj
->dirty
= true;
2589 reg_state
[CTX_RING_HEAD
+1] = 0;
2590 reg_state
[CTX_RING_TAIL
+1] = 0;
2592 i915_gem_object_unpin_map(ctx_obj
);
2594 ce
->ringbuf
->head
= 0;
2595 ce
->ringbuf
->tail
= 0;