2 * Copyright © 2008-2010 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 * Eric Anholt <eric@anholt.net>
25 * Zou Nan hai <nanhai.zou@intel.com>
26 * Xiang Hai hao<haihao.xiang@intel.com>
32 #include <drm/i915_drm.h>
33 #include "i915_trace.h"
34 #include "intel_drv.h"
36 /* Early gen2 devices have a cacheline of just 32 bytes, using 64 is overkill,
37 * but keeps the logic simple. Indeed, the whole purpose of this macro is just
38 * to give some inclination as to some of the magic values used in the various
41 #define CACHELINE_BYTES 64
43 static inline int __ring_space(int head
, int tail
, int size
)
45 int space
= head
- (tail
+ I915_RING_FREE_SPACE
);
51 static inline int ring_space(struct intel_engine_cs
*ring
)
53 struct intel_ringbuffer
*ringbuf
= ring
->buffer
;
54 return __ring_space(ringbuf
->head
& HEAD_ADDR
, ringbuf
->tail
, ringbuf
->size
);
57 static bool intel_ring_stopped(struct intel_engine_cs
*ring
)
59 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
60 return dev_priv
->gpu_error
.stop_rings
& intel_ring_flag(ring
);
63 void __intel_ring_advance(struct intel_engine_cs
*ring
)
65 struct intel_ringbuffer
*ringbuf
= ring
->buffer
;
66 ringbuf
->tail
&= ringbuf
->size
- 1;
67 if (intel_ring_stopped(ring
))
69 ring
->write_tail(ring
, ringbuf
->tail
);
73 gen2_render_ring_flush(struct intel_engine_cs
*ring
,
74 u32 invalidate_domains
,
81 if (((invalidate_domains
|flush_domains
) & I915_GEM_DOMAIN_RENDER
) == 0)
82 cmd
|= MI_NO_WRITE_FLUSH
;
84 if (invalidate_domains
& I915_GEM_DOMAIN_SAMPLER
)
87 ret
= intel_ring_begin(ring
, 2);
91 intel_ring_emit(ring
, cmd
);
92 intel_ring_emit(ring
, MI_NOOP
);
93 intel_ring_advance(ring
);
99 gen4_render_ring_flush(struct intel_engine_cs
*ring
,
100 u32 invalidate_domains
,
103 struct drm_device
*dev
= ring
->dev
;
110 * I915_GEM_DOMAIN_RENDER is always invalidated, but is
111 * only flushed if MI_NO_WRITE_FLUSH is unset. On 965, it is
112 * also flushed at 2d versus 3d pipeline switches.
116 * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
117 * MI_READ_FLUSH is set, and is always flushed on 965.
119 * I915_GEM_DOMAIN_COMMAND may not exist?
121 * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
122 * invalidated when MI_EXE_FLUSH is set.
124 * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
125 * invalidated with every MI_FLUSH.
129 * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
130 * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
131 * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
132 * are flushed at any MI_FLUSH.
135 cmd
= MI_FLUSH
| MI_NO_WRITE_FLUSH
;
136 if ((invalidate_domains
|flush_domains
) & I915_GEM_DOMAIN_RENDER
)
137 cmd
&= ~MI_NO_WRITE_FLUSH
;
138 if (invalidate_domains
& I915_GEM_DOMAIN_INSTRUCTION
)
141 if (invalidate_domains
& I915_GEM_DOMAIN_COMMAND
&&
142 (IS_G4X(dev
) || IS_GEN5(dev
)))
143 cmd
|= MI_INVALIDATE_ISP
;
145 ret
= intel_ring_begin(ring
, 2);
149 intel_ring_emit(ring
, cmd
);
150 intel_ring_emit(ring
, MI_NOOP
);
151 intel_ring_advance(ring
);
157 * Emits a PIPE_CONTROL with a non-zero post-sync operation, for
158 * implementing two workarounds on gen6. From section 1.4.7.1
159 * "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
161 * [DevSNB-C+{W/A}] Before any depth stall flush (including those
162 * produced by non-pipelined state commands), software needs to first
163 * send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
166 * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
167 * =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
169 * And the workaround for these two requires this workaround first:
171 * [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
172 * BEFORE the pipe-control with a post-sync op and no write-cache
175 * And this last workaround is tricky because of the requirements on
176 * that bit. From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
179 * "1 of the following must also be set:
180 * - Render Target Cache Flush Enable ([12] of DW1)
181 * - Depth Cache Flush Enable ([0] of DW1)
182 * - Stall at Pixel Scoreboard ([1] of DW1)
183 * - Depth Stall ([13] of DW1)
184 * - Post-Sync Operation ([13] of DW1)
185 * - Notify Enable ([8] of DW1)"
187 * The cache flushes require the workaround flush that triggered this
188 * one, so we can't use it. Depth stall would trigger the same.
189 * Post-sync nonzero is what triggered this second workaround, so we
190 * can't use that one either. Notify enable is IRQs, which aren't
191 * really our business. That leaves only stall at scoreboard.
194 intel_emit_post_sync_nonzero_flush(struct intel_engine_cs
*ring
)
196 u32 scratch_addr
= ring
->scratch
.gtt_offset
+ 2 * CACHELINE_BYTES
;
200 ret
= intel_ring_begin(ring
, 6);
204 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(5));
205 intel_ring_emit(ring
, PIPE_CONTROL_CS_STALL
|
206 PIPE_CONTROL_STALL_AT_SCOREBOARD
);
207 intel_ring_emit(ring
, scratch_addr
| PIPE_CONTROL_GLOBAL_GTT
); /* address */
208 intel_ring_emit(ring
, 0); /* low dword */
209 intel_ring_emit(ring
, 0); /* high dword */
210 intel_ring_emit(ring
, MI_NOOP
);
211 intel_ring_advance(ring
);
213 ret
= intel_ring_begin(ring
, 6);
217 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(5));
218 intel_ring_emit(ring
, PIPE_CONTROL_QW_WRITE
);
219 intel_ring_emit(ring
, scratch_addr
| PIPE_CONTROL_GLOBAL_GTT
); /* address */
220 intel_ring_emit(ring
, 0);
221 intel_ring_emit(ring
, 0);
222 intel_ring_emit(ring
, MI_NOOP
);
223 intel_ring_advance(ring
);
229 gen6_render_ring_flush(struct intel_engine_cs
*ring
,
230 u32 invalidate_domains
, u32 flush_domains
)
233 u32 scratch_addr
= ring
->scratch
.gtt_offset
+ 2 * CACHELINE_BYTES
;
236 /* Force SNB workarounds for PIPE_CONTROL flushes */
237 ret
= intel_emit_post_sync_nonzero_flush(ring
);
241 /* Just flush everything. Experiments have shown that reducing the
242 * number of bits based on the write domains has little performance
246 flags
|= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH
;
247 flags
|= PIPE_CONTROL_DEPTH_CACHE_FLUSH
;
249 * Ensure that any following seqno writes only happen
250 * when the render cache is indeed flushed.
252 flags
|= PIPE_CONTROL_CS_STALL
;
254 if (invalidate_domains
) {
255 flags
|= PIPE_CONTROL_TLB_INVALIDATE
;
256 flags
|= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE
;
257 flags
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
258 flags
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
259 flags
|= PIPE_CONTROL_CONST_CACHE_INVALIDATE
;
260 flags
|= PIPE_CONTROL_STATE_CACHE_INVALIDATE
;
262 * TLB invalidate requires a post-sync write.
264 flags
|= PIPE_CONTROL_QW_WRITE
| PIPE_CONTROL_CS_STALL
;
267 ret
= intel_ring_begin(ring
, 4);
271 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(4));
272 intel_ring_emit(ring
, flags
);
273 intel_ring_emit(ring
, scratch_addr
| PIPE_CONTROL_GLOBAL_GTT
);
274 intel_ring_emit(ring
, 0);
275 intel_ring_advance(ring
);
281 gen7_render_ring_cs_stall_wa(struct intel_engine_cs
*ring
)
285 ret
= intel_ring_begin(ring
, 4);
289 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(4));
290 intel_ring_emit(ring
, PIPE_CONTROL_CS_STALL
|
291 PIPE_CONTROL_STALL_AT_SCOREBOARD
);
292 intel_ring_emit(ring
, 0);
293 intel_ring_emit(ring
, 0);
294 intel_ring_advance(ring
);
299 static int gen7_ring_fbc_flush(struct intel_engine_cs
*ring
, u32 value
)
303 if (!ring
->fbc_dirty
)
306 ret
= intel_ring_begin(ring
, 6);
309 /* WaFbcNukeOn3DBlt:ivb/hsw */
310 intel_ring_emit(ring
, MI_LOAD_REGISTER_IMM(1));
311 intel_ring_emit(ring
, MSG_FBC_REND_STATE
);
312 intel_ring_emit(ring
, value
);
313 intel_ring_emit(ring
, MI_STORE_REGISTER_MEM(1) | MI_SRM_LRM_GLOBAL_GTT
);
314 intel_ring_emit(ring
, MSG_FBC_REND_STATE
);
315 intel_ring_emit(ring
, ring
->scratch
.gtt_offset
+ 256);
316 intel_ring_advance(ring
);
318 ring
->fbc_dirty
= false;
323 gen7_render_ring_flush(struct intel_engine_cs
*ring
,
324 u32 invalidate_domains
, u32 flush_domains
)
327 u32 scratch_addr
= ring
->scratch
.gtt_offset
+ 2 * CACHELINE_BYTES
;
331 * Ensure that any following seqno writes only happen when the render
332 * cache is indeed flushed.
334 * Workaround: 4th PIPE_CONTROL command (except the ones with only
335 * read-cache invalidate bits set) must have the CS_STALL bit set. We
336 * don't try to be clever and just set it unconditionally.
338 flags
|= PIPE_CONTROL_CS_STALL
;
340 /* Just flush everything. Experiments have shown that reducing the
341 * number of bits based on the write domains has little performance
345 flags
|= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH
;
346 flags
|= PIPE_CONTROL_DEPTH_CACHE_FLUSH
;
348 if (invalidate_domains
) {
349 flags
|= PIPE_CONTROL_TLB_INVALIDATE
;
350 flags
|= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE
;
351 flags
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
352 flags
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
353 flags
|= PIPE_CONTROL_CONST_CACHE_INVALIDATE
;
354 flags
|= PIPE_CONTROL_STATE_CACHE_INVALIDATE
;
356 * TLB invalidate requires a post-sync write.
358 flags
|= PIPE_CONTROL_QW_WRITE
;
359 flags
|= PIPE_CONTROL_GLOBAL_GTT_IVB
;
361 /* Workaround: we must issue a pipe_control with CS-stall bit
362 * set before a pipe_control command that has the state cache
363 * invalidate bit set. */
364 gen7_render_ring_cs_stall_wa(ring
);
367 ret
= intel_ring_begin(ring
, 4);
371 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(4));
372 intel_ring_emit(ring
, flags
);
373 intel_ring_emit(ring
, scratch_addr
);
374 intel_ring_emit(ring
, 0);
375 intel_ring_advance(ring
);
377 if (!invalidate_domains
&& flush_domains
)
378 return gen7_ring_fbc_flush(ring
, FBC_REND_NUKE
);
384 gen8_render_ring_flush(struct intel_engine_cs
*ring
,
385 u32 invalidate_domains
, u32 flush_domains
)
388 u32 scratch_addr
= ring
->scratch
.gtt_offset
+ 2 * CACHELINE_BYTES
;
391 flags
|= PIPE_CONTROL_CS_STALL
;
394 flags
|= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH
;
395 flags
|= PIPE_CONTROL_DEPTH_CACHE_FLUSH
;
397 if (invalidate_domains
) {
398 flags
|= PIPE_CONTROL_TLB_INVALIDATE
;
399 flags
|= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE
;
400 flags
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
401 flags
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
402 flags
|= PIPE_CONTROL_CONST_CACHE_INVALIDATE
;
403 flags
|= PIPE_CONTROL_STATE_CACHE_INVALIDATE
;
404 flags
|= PIPE_CONTROL_QW_WRITE
;
405 flags
|= PIPE_CONTROL_GLOBAL_GTT_IVB
;
408 ret
= intel_ring_begin(ring
, 6);
412 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(6));
413 intel_ring_emit(ring
, flags
);
414 intel_ring_emit(ring
, scratch_addr
);
415 intel_ring_emit(ring
, 0);
416 intel_ring_emit(ring
, 0);
417 intel_ring_emit(ring
, 0);
418 intel_ring_advance(ring
);
424 static void ring_write_tail(struct intel_engine_cs
*ring
,
427 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
428 I915_WRITE_TAIL(ring
, value
);
431 u64
intel_ring_get_active_head(struct intel_engine_cs
*ring
)
433 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
436 if (INTEL_INFO(ring
->dev
)->gen
>= 8)
437 acthd
= I915_READ64_2x32(RING_ACTHD(ring
->mmio_base
),
438 RING_ACTHD_UDW(ring
->mmio_base
));
439 else if (INTEL_INFO(ring
->dev
)->gen
>= 4)
440 acthd
= I915_READ(RING_ACTHD(ring
->mmio_base
));
442 acthd
= I915_READ(ACTHD
);
447 static void ring_setup_phys_status_page(struct intel_engine_cs
*ring
)
449 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
452 addr
= dev_priv
->status_page_dmah
->busaddr
;
453 if (INTEL_INFO(ring
->dev
)->gen
>= 4)
454 addr
|= (dev_priv
->status_page_dmah
->busaddr
>> 28) & 0xf0;
455 I915_WRITE(HWS_PGA
, addr
);
458 static bool stop_ring(struct intel_engine_cs
*ring
)
460 struct drm_i915_private
*dev_priv
= to_i915(ring
->dev
);
462 if (!IS_GEN2(ring
->dev
)) {
463 I915_WRITE_MODE(ring
, _MASKED_BIT_ENABLE(STOP_RING
));
464 if (wait_for_atomic((I915_READ_MODE(ring
) & MODE_IDLE
) != 0, 1000)) {
465 DRM_ERROR("%s :timed out trying to stop ring\n", ring
->name
);
470 I915_WRITE_CTL(ring
, 0);
471 I915_WRITE_HEAD(ring
, 0);
472 ring
->write_tail(ring
, 0);
474 if (!IS_GEN2(ring
->dev
)) {
475 (void)I915_READ_CTL(ring
);
476 I915_WRITE_MODE(ring
, _MASKED_BIT_DISABLE(STOP_RING
));
479 return (I915_READ_HEAD(ring
) & HEAD_ADDR
) == 0;
482 static int init_ring_common(struct intel_engine_cs
*ring
)
484 struct drm_device
*dev
= ring
->dev
;
485 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
486 struct intel_ringbuffer
*ringbuf
= ring
->buffer
;
487 struct drm_i915_gem_object
*obj
= ringbuf
->obj
;
490 gen6_gt_force_wake_get(dev_priv
, FORCEWAKE_ALL
);
492 if (!stop_ring(ring
)) {
493 /* G45 ring initialization often fails to reset head to zero */
494 DRM_DEBUG_KMS("%s head not reset to zero "
495 "ctl %08x head %08x tail %08x start %08x\n",
498 I915_READ_HEAD(ring
),
499 I915_READ_TAIL(ring
),
500 I915_READ_START(ring
));
502 if (!stop_ring(ring
)) {
503 DRM_ERROR("failed to set %s head to zero "
504 "ctl %08x head %08x tail %08x start %08x\n",
507 I915_READ_HEAD(ring
),
508 I915_READ_TAIL(ring
),
509 I915_READ_START(ring
));
515 if (I915_NEED_GFX_HWS(dev
))
516 intel_ring_setup_status_page(ring
);
518 ring_setup_phys_status_page(ring
);
520 /* Initialize the ring. This must happen _after_ we've cleared the ring
521 * registers with the above sequence (the readback of the HEAD registers
522 * also enforces ordering), otherwise the hw might lose the new ring
523 * register values. */
524 I915_WRITE_START(ring
, i915_gem_obj_ggtt_offset(obj
));
526 ((ringbuf
->size
- PAGE_SIZE
) & RING_NR_PAGES
)
529 /* If the head is still not zero, the ring is dead */
530 if (wait_for((I915_READ_CTL(ring
) & RING_VALID
) != 0 &&
531 I915_READ_START(ring
) == i915_gem_obj_ggtt_offset(obj
) &&
532 (I915_READ_HEAD(ring
) & HEAD_ADDR
) == 0, 50)) {
533 DRM_ERROR("%s initialization failed "
534 "ctl %08x (valid? %d) head %08x tail %08x start %08x [expected %08lx]\n",
536 I915_READ_CTL(ring
), I915_READ_CTL(ring
) & RING_VALID
,
537 I915_READ_HEAD(ring
), I915_READ_TAIL(ring
),
538 I915_READ_START(ring
), (unsigned long)i915_gem_obj_ggtt_offset(obj
));
543 if (!drm_core_check_feature(ring
->dev
, DRIVER_MODESET
))
544 i915_kernel_lost_context(ring
->dev
);
546 ringbuf
->head
= I915_READ_HEAD(ring
);
547 ringbuf
->tail
= I915_READ_TAIL(ring
) & TAIL_ADDR
;
548 ringbuf
->space
= ring_space(ring
);
549 ringbuf
->last_retired_head
= -1;
552 memset(&ring
->hangcheck
, 0, sizeof(ring
->hangcheck
));
555 gen6_gt_force_wake_put(dev_priv
, FORCEWAKE_ALL
);
561 init_pipe_control(struct intel_engine_cs
*ring
)
565 if (ring
->scratch
.obj
)
568 ring
->scratch
.obj
= i915_gem_alloc_object(ring
->dev
, 4096);
569 if (ring
->scratch
.obj
== NULL
) {
570 DRM_ERROR("Failed to allocate seqno page\n");
575 ret
= i915_gem_object_set_cache_level(ring
->scratch
.obj
, I915_CACHE_LLC
);
579 ret
= i915_gem_obj_ggtt_pin(ring
->scratch
.obj
, 4096, 0);
583 ring
->scratch
.gtt_offset
= i915_gem_obj_ggtt_offset(ring
->scratch
.obj
);
584 ring
->scratch
.cpu_page
= kmap(sg_page(ring
->scratch
.obj
->pages
->sgl
));
585 if (ring
->scratch
.cpu_page
== NULL
) {
590 DRM_DEBUG_DRIVER("%s pipe control offset: 0x%08x\n",
591 ring
->name
, ring
->scratch
.gtt_offset
);
595 i915_gem_object_ggtt_unpin(ring
->scratch
.obj
);
597 drm_gem_object_unreference(&ring
->scratch
.obj
->base
);
602 static int init_render_ring(struct intel_engine_cs
*ring
)
604 struct drm_device
*dev
= ring
->dev
;
605 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
606 int ret
= init_ring_common(ring
);
608 /* WaTimedSingleVertexDispatch:cl,bw,ctg,elk,ilk,snb */
609 if (INTEL_INFO(dev
)->gen
>= 4 && INTEL_INFO(dev
)->gen
< 7)
610 I915_WRITE(MI_MODE
, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH
));
612 /* We need to disable the AsyncFlip performance optimisations in order
613 * to use MI_WAIT_FOR_EVENT within the CS. It should already be
614 * programmed to '1' on all products.
616 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv
618 if (INTEL_INFO(dev
)->gen
>= 6)
619 I915_WRITE(MI_MODE
, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE
));
621 /* Required for the hardware to program scanline values for waiting */
622 /* WaEnableFlushTlbInvalidationMode:snb */
623 if (INTEL_INFO(dev
)->gen
== 6)
625 _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT
));
627 /* WaBCSVCSTlbInvalidationMode:ivb,vlv,hsw */
629 I915_WRITE(GFX_MODE_GEN7
,
630 _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT
) |
631 _MASKED_BIT_ENABLE(GFX_REPLAY_MODE
));
633 if (INTEL_INFO(dev
)->gen
>= 5) {
634 ret
= init_pipe_control(ring
);
640 /* From the Sandybridge PRM, volume 1 part 3, page 24:
641 * "If this bit is set, STCunit will have LRA as replacement
642 * policy. [...] This bit must be reset. LRA replacement
643 * policy is not supported."
645 I915_WRITE(CACHE_MODE_0
,
646 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB
));
649 if (INTEL_INFO(dev
)->gen
>= 6)
650 I915_WRITE(INSTPM
, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING
));
653 I915_WRITE_IMR(ring
, ~GT_PARITY_ERROR(dev
));
658 static void render_ring_cleanup(struct intel_engine_cs
*ring
)
660 struct drm_device
*dev
= ring
->dev
;
662 if (ring
->scratch
.obj
== NULL
)
665 if (INTEL_INFO(dev
)->gen
>= 5) {
666 kunmap(sg_page(ring
->scratch
.obj
->pages
->sgl
));
667 i915_gem_object_ggtt_unpin(ring
->scratch
.obj
);
670 drm_gem_object_unreference(&ring
->scratch
.obj
->base
);
671 ring
->scratch
.obj
= NULL
;
674 static int gen6_signal(struct intel_engine_cs
*signaller
,
675 unsigned int num_dwords
)
677 struct drm_device
*dev
= signaller
->dev
;
678 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
679 struct intel_engine_cs
*useless
;
682 /* NB: In order to be able to do semaphore MBOX updates for varying
683 * number of rings, it's easiest if we round up each individual update
684 * to a multiple of 2 (since ring updates must always be a multiple of
685 * 2) even though the actual update only requires 3 dwords.
687 #define MBOX_UPDATE_DWORDS 4
688 if (i915_semaphore_is_enabled(dev
))
689 num_dwords
+= ((I915_NUM_RINGS
-1) * MBOX_UPDATE_DWORDS
);
691 return intel_ring_begin(signaller
, num_dwords
);
693 ret
= intel_ring_begin(signaller
, num_dwords
);
696 #undef MBOX_UPDATE_DWORDS
698 for_each_ring(useless
, dev_priv
, i
) {
699 u32 mbox_reg
= signaller
->semaphore
.mbox
.signal
[i
];
700 if (mbox_reg
!= GEN6_NOSYNC
) {
701 intel_ring_emit(signaller
, MI_LOAD_REGISTER_IMM(1));
702 intel_ring_emit(signaller
, mbox_reg
);
703 intel_ring_emit(signaller
, signaller
->outstanding_lazy_seqno
);
704 intel_ring_emit(signaller
, MI_NOOP
);
706 intel_ring_emit(signaller
, MI_NOOP
);
707 intel_ring_emit(signaller
, MI_NOOP
);
708 intel_ring_emit(signaller
, MI_NOOP
);
709 intel_ring_emit(signaller
, MI_NOOP
);
717 * gen6_add_request - Update the semaphore mailbox registers
719 * @ring - ring that is adding a request
720 * @seqno - return seqno stuck into the ring
722 * Update the mailbox registers in the *other* rings with the current seqno.
723 * This acts like a signal in the canonical semaphore.
726 gen6_add_request(struct intel_engine_cs
*ring
)
730 ret
= ring
->semaphore
.signal(ring
, 4);
734 intel_ring_emit(ring
, MI_STORE_DWORD_INDEX
);
735 intel_ring_emit(ring
, I915_GEM_HWS_INDEX
<< MI_STORE_DWORD_INDEX_SHIFT
);
736 intel_ring_emit(ring
, ring
->outstanding_lazy_seqno
);
737 intel_ring_emit(ring
, MI_USER_INTERRUPT
);
738 __intel_ring_advance(ring
);
743 static inline bool i915_gem_has_seqno_wrapped(struct drm_device
*dev
,
746 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
747 return dev_priv
->last_seqno
< seqno
;
751 * intel_ring_sync - sync the waiter to the signaller on seqno
753 * @waiter - ring that is waiting
754 * @signaller - ring which has, or will signal
755 * @seqno - seqno which the waiter will block on
758 gen6_ring_sync(struct intel_engine_cs
*waiter
,
759 struct intel_engine_cs
*signaller
,
762 u32 dw1
= MI_SEMAPHORE_MBOX
|
763 MI_SEMAPHORE_COMPARE
|
764 MI_SEMAPHORE_REGISTER
;
765 u32 wait_mbox
= signaller
->semaphore
.mbox
.wait
[waiter
->id
];
768 /* Throughout all of the GEM code, seqno passed implies our current
769 * seqno is >= the last seqno executed. However for hardware the
770 * comparison is strictly greater than.
774 WARN_ON(wait_mbox
== MI_SEMAPHORE_SYNC_INVALID
);
776 ret
= intel_ring_begin(waiter
, 4);
780 /* If seqno wrap happened, omit the wait with no-ops */
781 if (likely(!i915_gem_has_seqno_wrapped(waiter
->dev
, seqno
))) {
782 intel_ring_emit(waiter
, dw1
| wait_mbox
);
783 intel_ring_emit(waiter
, seqno
);
784 intel_ring_emit(waiter
, 0);
785 intel_ring_emit(waiter
, MI_NOOP
);
787 intel_ring_emit(waiter
, MI_NOOP
);
788 intel_ring_emit(waiter
, MI_NOOP
);
789 intel_ring_emit(waiter
, MI_NOOP
);
790 intel_ring_emit(waiter
, MI_NOOP
);
792 intel_ring_advance(waiter
);
797 #define PIPE_CONTROL_FLUSH(ring__, addr__) \
799 intel_ring_emit(ring__, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE | \
800 PIPE_CONTROL_DEPTH_STALL); \
801 intel_ring_emit(ring__, (addr__) | PIPE_CONTROL_GLOBAL_GTT); \
802 intel_ring_emit(ring__, 0); \
803 intel_ring_emit(ring__, 0); \
807 pc_render_add_request(struct intel_engine_cs
*ring
)
809 u32 scratch_addr
= ring
->scratch
.gtt_offset
+ 2 * CACHELINE_BYTES
;
812 /* For Ironlake, MI_USER_INTERRUPT was deprecated and apparently
813 * incoherent with writes to memory, i.e. completely fubar,
814 * so we need to use PIPE_NOTIFY instead.
816 * However, we also need to workaround the qword write
817 * incoherence by flushing the 6 PIPE_NOTIFY buffers out to
818 * memory before requesting an interrupt.
820 ret
= intel_ring_begin(ring
, 32);
824 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE
|
825 PIPE_CONTROL_WRITE_FLUSH
|
826 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
);
827 intel_ring_emit(ring
, ring
->scratch
.gtt_offset
| PIPE_CONTROL_GLOBAL_GTT
);
828 intel_ring_emit(ring
, ring
->outstanding_lazy_seqno
);
829 intel_ring_emit(ring
, 0);
830 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
831 scratch_addr
+= 2 * CACHELINE_BYTES
; /* write to separate cachelines */
832 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
833 scratch_addr
+= 2 * CACHELINE_BYTES
;
834 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
835 scratch_addr
+= 2 * CACHELINE_BYTES
;
836 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
837 scratch_addr
+= 2 * CACHELINE_BYTES
;
838 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
839 scratch_addr
+= 2 * CACHELINE_BYTES
;
840 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
842 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE
|
843 PIPE_CONTROL_WRITE_FLUSH
|
844 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
845 PIPE_CONTROL_NOTIFY
);
846 intel_ring_emit(ring
, ring
->scratch
.gtt_offset
| PIPE_CONTROL_GLOBAL_GTT
);
847 intel_ring_emit(ring
, ring
->outstanding_lazy_seqno
);
848 intel_ring_emit(ring
, 0);
849 __intel_ring_advance(ring
);
855 gen6_ring_get_seqno(struct intel_engine_cs
*ring
, bool lazy_coherency
)
857 /* Workaround to force correct ordering between irq and seqno writes on
858 * ivb (and maybe also on snb) by reading from a CS register (like
859 * ACTHD) before reading the status page. */
860 if (!lazy_coherency
) {
861 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
862 POSTING_READ(RING_ACTHD(ring
->mmio_base
));
865 return intel_read_status_page(ring
, I915_GEM_HWS_INDEX
);
869 ring_get_seqno(struct intel_engine_cs
*ring
, bool lazy_coherency
)
871 return intel_read_status_page(ring
, I915_GEM_HWS_INDEX
);
875 ring_set_seqno(struct intel_engine_cs
*ring
, u32 seqno
)
877 intel_write_status_page(ring
, I915_GEM_HWS_INDEX
, seqno
);
881 pc_render_get_seqno(struct intel_engine_cs
*ring
, bool lazy_coherency
)
883 return ring
->scratch
.cpu_page
[0];
887 pc_render_set_seqno(struct intel_engine_cs
*ring
, u32 seqno
)
889 ring
->scratch
.cpu_page
[0] = seqno
;
893 gen5_ring_get_irq(struct intel_engine_cs
*ring
)
895 struct drm_device
*dev
= ring
->dev
;
896 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
899 if (!dev
->irq_enabled
)
902 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
903 if (ring
->irq_refcount
++ == 0)
904 ilk_enable_gt_irq(dev_priv
, ring
->irq_enable_mask
);
905 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
911 gen5_ring_put_irq(struct intel_engine_cs
*ring
)
913 struct drm_device
*dev
= ring
->dev
;
914 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
917 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
918 if (--ring
->irq_refcount
== 0)
919 ilk_disable_gt_irq(dev_priv
, ring
->irq_enable_mask
);
920 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
924 i9xx_ring_get_irq(struct intel_engine_cs
*ring
)
926 struct drm_device
*dev
= ring
->dev
;
927 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
930 if (!dev
->irq_enabled
)
933 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
934 if (ring
->irq_refcount
++ == 0) {
935 dev_priv
->irq_mask
&= ~ring
->irq_enable_mask
;
936 I915_WRITE(IMR
, dev_priv
->irq_mask
);
939 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
945 i9xx_ring_put_irq(struct intel_engine_cs
*ring
)
947 struct drm_device
*dev
= ring
->dev
;
948 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
951 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
952 if (--ring
->irq_refcount
== 0) {
953 dev_priv
->irq_mask
|= ring
->irq_enable_mask
;
954 I915_WRITE(IMR
, dev_priv
->irq_mask
);
957 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
961 i8xx_ring_get_irq(struct intel_engine_cs
*ring
)
963 struct drm_device
*dev
= ring
->dev
;
964 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
967 if (!dev
->irq_enabled
)
970 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
971 if (ring
->irq_refcount
++ == 0) {
972 dev_priv
->irq_mask
&= ~ring
->irq_enable_mask
;
973 I915_WRITE16(IMR
, dev_priv
->irq_mask
);
976 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
982 i8xx_ring_put_irq(struct intel_engine_cs
*ring
)
984 struct drm_device
*dev
= ring
->dev
;
985 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
988 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
989 if (--ring
->irq_refcount
== 0) {
990 dev_priv
->irq_mask
|= ring
->irq_enable_mask
;
991 I915_WRITE16(IMR
, dev_priv
->irq_mask
);
994 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
997 void intel_ring_setup_status_page(struct intel_engine_cs
*ring
)
999 struct drm_device
*dev
= ring
->dev
;
1000 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
1003 /* The ring status page addresses are no longer next to the rest of
1004 * the ring registers as of gen7.
1009 mmio
= RENDER_HWS_PGA_GEN7
;
1012 mmio
= BLT_HWS_PGA_GEN7
;
1015 * VCS2 actually doesn't exist on Gen7. Only shut up
1016 * gcc switch check warning
1020 mmio
= BSD_HWS_PGA_GEN7
;
1023 mmio
= VEBOX_HWS_PGA_GEN7
;
1026 } else if (IS_GEN6(ring
->dev
)) {
1027 mmio
= RING_HWS_PGA_GEN6(ring
->mmio_base
);
1029 /* XXX: gen8 returns to sanity */
1030 mmio
= RING_HWS_PGA(ring
->mmio_base
);
1033 I915_WRITE(mmio
, (u32
)ring
->status_page
.gfx_addr
);
1037 * Flush the TLB for this page
1039 * FIXME: These two bits have disappeared on gen8, so a question
1040 * arises: do we still need this and if so how should we go about
1041 * invalidating the TLB?
1043 if (INTEL_INFO(dev
)->gen
>= 6 && INTEL_INFO(dev
)->gen
< 8) {
1044 u32 reg
= RING_INSTPM(ring
->mmio_base
);
1046 /* ring should be idle before issuing a sync flush*/
1047 WARN_ON((I915_READ_MODE(ring
) & MODE_IDLE
) == 0);
1050 _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE
|
1051 INSTPM_SYNC_FLUSH
));
1052 if (wait_for((I915_READ(reg
) & INSTPM_SYNC_FLUSH
) == 0,
1054 DRM_ERROR("%s: wait for SyncFlush to complete for TLB invalidation timed out\n",
1060 bsd_ring_flush(struct intel_engine_cs
*ring
,
1061 u32 invalidate_domains
,
1066 ret
= intel_ring_begin(ring
, 2);
1070 intel_ring_emit(ring
, MI_FLUSH
);
1071 intel_ring_emit(ring
, MI_NOOP
);
1072 intel_ring_advance(ring
);
1077 i9xx_add_request(struct intel_engine_cs
*ring
)
1081 ret
= intel_ring_begin(ring
, 4);
1085 intel_ring_emit(ring
, MI_STORE_DWORD_INDEX
);
1086 intel_ring_emit(ring
, I915_GEM_HWS_INDEX
<< MI_STORE_DWORD_INDEX_SHIFT
);
1087 intel_ring_emit(ring
, ring
->outstanding_lazy_seqno
);
1088 intel_ring_emit(ring
, MI_USER_INTERRUPT
);
1089 __intel_ring_advance(ring
);
1095 gen6_ring_get_irq(struct intel_engine_cs
*ring
)
1097 struct drm_device
*dev
= ring
->dev
;
1098 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1099 unsigned long flags
;
1101 if (!dev
->irq_enabled
)
1104 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1105 if (ring
->irq_refcount
++ == 0) {
1106 if (HAS_L3_DPF(dev
) && ring
->id
== RCS
)
1107 I915_WRITE_IMR(ring
,
1108 ~(ring
->irq_enable_mask
|
1109 GT_PARITY_ERROR(dev
)));
1111 I915_WRITE_IMR(ring
, ~ring
->irq_enable_mask
);
1112 ilk_enable_gt_irq(dev_priv
, ring
->irq_enable_mask
);
1114 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1120 gen6_ring_put_irq(struct intel_engine_cs
*ring
)
1122 struct drm_device
*dev
= ring
->dev
;
1123 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1124 unsigned long flags
;
1126 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1127 if (--ring
->irq_refcount
== 0) {
1128 if (HAS_L3_DPF(dev
) && ring
->id
== RCS
)
1129 I915_WRITE_IMR(ring
, ~GT_PARITY_ERROR(dev
));
1131 I915_WRITE_IMR(ring
, ~0);
1132 ilk_disable_gt_irq(dev_priv
, ring
->irq_enable_mask
);
1134 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1138 hsw_vebox_get_irq(struct intel_engine_cs
*ring
)
1140 struct drm_device
*dev
= ring
->dev
;
1141 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1142 unsigned long flags
;
1144 if (!dev
->irq_enabled
)
1147 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1148 if (ring
->irq_refcount
++ == 0) {
1149 I915_WRITE_IMR(ring
, ~ring
->irq_enable_mask
);
1150 snb_enable_pm_irq(dev_priv
, ring
->irq_enable_mask
);
1152 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1158 hsw_vebox_put_irq(struct intel_engine_cs
*ring
)
1160 struct drm_device
*dev
= ring
->dev
;
1161 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1162 unsigned long flags
;
1164 if (!dev
->irq_enabled
)
1167 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1168 if (--ring
->irq_refcount
== 0) {
1169 I915_WRITE_IMR(ring
, ~0);
1170 snb_disable_pm_irq(dev_priv
, ring
->irq_enable_mask
);
1172 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1176 gen8_ring_get_irq(struct intel_engine_cs
*ring
)
1178 struct drm_device
*dev
= ring
->dev
;
1179 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1180 unsigned long flags
;
1182 if (!dev
->irq_enabled
)
1185 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1186 if (ring
->irq_refcount
++ == 0) {
1187 if (HAS_L3_DPF(dev
) && ring
->id
== RCS
) {
1188 I915_WRITE_IMR(ring
,
1189 ~(ring
->irq_enable_mask
|
1190 GT_RENDER_L3_PARITY_ERROR_INTERRUPT
));
1192 I915_WRITE_IMR(ring
, ~ring
->irq_enable_mask
);
1194 POSTING_READ(RING_IMR(ring
->mmio_base
));
1196 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1202 gen8_ring_put_irq(struct intel_engine_cs
*ring
)
1204 struct drm_device
*dev
= ring
->dev
;
1205 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1206 unsigned long flags
;
1208 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1209 if (--ring
->irq_refcount
== 0) {
1210 if (HAS_L3_DPF(dev
) && ring
->id
== RCS
) {
1211 I915_WRITE_IMR(ring
,
1212 ~GT_RENDER_L3_PARITY_ERROR_INTERRUPT
);
1214 I915_WRITE_IMR(ring
, ~0);
1216 POSTING_READ(RING_IMR(ring
->mmio_base
));
1218 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1222 i965_dispatch_execbuffer(struct intel_engine_cs
*ring
,
1223 u64 offset
, u32 length
,
1228 ret
= intel_ring_begin(ring
, 2);
1232 intel_ring_emit(ring
,
1233 MI_BATCH_BUFFER_START
|
1235 (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE_I965
));
1236 intel_ring_emit(ring
, offset
);
1237 intel_ring_advance(ring
);
1242 /* Just userspace ABI convention to limit the wa batch bo to a resonable size */
1243 #define I830_BATCH_LIMIT (256*1024)
1245 i830_dispatch_execbuffer(struct intel_engine_cs
*ring
,
1246 u64 offset
, u32 len
,
1251 if (flags
& I915_DISPATCH_PINNED
) {
1252 ret
= intel_ring_begin(ring
, 4);
1256 intel_ring_emit(ring
, MI_BATCH_BUFFER
);
1257 intel_ring_emit(ring
, offset
| (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE
));
1258 intel_ring_emit(ring
, offset
+ len
- 8);
1259 intel_ring_emit(ring
, MI_NOOP
);
1260 intel_ring_advance(ring
);
1262 u32 cs_offset
= ring
->scratch
.gtt_offset
;
1264 if (len
> I830_BATCH_LIMIT
)
1267 ret
= intel_ring_begin(ring
, 9+3);
1270 /* Blit the batch (which has now all relocs applied) to the stable batch
1271 * scratch bo area (so that the CS never stumbles over its tlb
1272 * invalidation bug) ... */
1273 intel_ring_emit(ring
, XY_SRC_COPY_BLT_CMD
|
1274 XY_SRC_COPY_BLT_WRITE_ALPHA
|
1275 XY_SRC_COPY_BLT_WRITE_RGB
);
1276 intel_ring_emit(ring
, BLT_DEPTH_32
| BLT_ROP_GXCOPY
| 4096);
1277 intel_ring_emit(ring
, 0);
1278 intel_ring_emit(ring
, (DIV_ROUND_UP(len
, 4096) << 16) | 1024);
1279 intel_ring_emit(ring
, cs_offset
);
1280 intel_ring_emit(ring
, 0);
1281 intel_ring_emit(ring
, 4096);
1282 intel_ring_emit(ring
, offset
);
1283 intel_ring_emit(ring
, MI_FLUSH
);
1285 /* ... and execute it. */
1286 intel_ring_emit(ring
, MI_BATCH_BUFFER
);
1287 intel_ring_emit(ring
, cs_offset
| (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE
));
1288 intel_ring_emit(ring
, cs_offset
+ len
- 8);
1289 intel_ring_advance(ring
);
1296 i915_dispatch_execbuffer(struct intel_engine_cs
*ring
,
1297 u64 offset
, u32 len
,
1302 ret
= intel_ring_begin(ring
, 2);
1306 intel_ring_emit(ring
, MI_BATCH_BUFFER_START
| MI_BATCH_GTT
);
1307 intel_ring_emit(ring
, offset
| (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE
));
1308 intel_ring_advance(ring
);
1313 static void cleanup_status_page(struct intel_engine_cs
*ring
)
1315 struct drm_i915_gem_object
*obj
;
1317 obj
= ring
->status_page
.obj
;
1321 kunmap(sg_page(obj
->pages
->sgl
));
1322 i915_gem_object_ggtt_unpin(obj
);
1323 drm_gem_object_unreference(&obj
->base
);
1324 ring
->status_page
.obj
= NULL
;
1327 static int init_status_page(struct intel_engine_cs
*ring
)
1329 struct drm_i915_gem_object
*obj
;
1331 if ((obj
= ring
->status_page
.obj
) == NULL
) {
1334 obj
= i915_gem_alloc_object(ring
->dev
, 4096);
1336 DRM_ERROR("Failed to allocate status page\n");
1340 ret
= i915_gem_object_set_cache_level(obj
, I915_CACHE_LLC
);
1344 ret
= i915_gem_obj_ggtt_pin(obj
, 4096, 0);
1347 drm_gem_object_unreference(&obj
->base
);
1351 ring
->status_page
.obj
= obj
;
1354 ring
->status_page
.gfx_addr
= i915_gem_obj_ggtt_offset(obj
);
1355 ring
->status_page
.page_addr
= kmap(sg_page(obj
->pages
->sgl
));
1356 memset(ring
->status_page
.page_addr
, 0, PAGE_SIZE
);
1358 DRM_DEBUG_DRIVER("%s hws offset: 0x%08x\n",
1359 ring
->name
, ring
->status_page
.gfx_addr
);
1364 static int init_phys_status_page(struct intel_engine_cs
*ring
)
1366 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
1368 if (!dev_priv
->status_page_dmah
) {
1369 dev_priv
->status_page_dmah
=
1370 drm_pci_alloc(ring
->dev
, PAGE_SIZE
, PAGE_SIZE
);
1371 if (!dev_priv
->status_page_dmah
)
1375 ring
->status_page
.page_addr
= dev_priv
->status_page_dmah
->vaddr
;
1376 memset(ring
->status_page
.page_addr
, 0, PAGE_SIZE
);
1381 static int allocate_ring_buffer(struct intel_engine_cs
*ring
)
1383 struct drm_device
*dev
= ring
->dev
;
1384 struct drm_i915_private
*dev_priv
= to_i915(dev
);
1385 struct intel_ringbuffer
*ringbuf
= ring
->buffer
;
1386 struct drm_i915_gem_object
*obj
;
1389 if (intel_ring_initialized(ring
))
1394 obj
= i915_gem_object_create_stolen(dev
, ringbuf
->size
);
1396 obj
= i915_gem_alloc_object(dev
, ringbuf
->size
);
1400 ret
= i915_gem_obj_ggtt_pin(obj
, PAGE_SIZE
, PIN_MAPPABLE
);
1404 ret
= i915_gem_object_set_to_gtt_domain(obj
, true);
1408 ringbuf
->virtual_start
=
1409 ioremap_wc(dev_priv
->gtt
.mappable_base
+ i915_gem_obj_ggtt_offset(obj
),
1411 if (ringbuf
->virtual_start
== NULL
) {
1420 i915_gem_object_ggtt_unpin(obj
);
1422 drm_gem_object_unreference(&obj
->base
);
1426 static int intel_init_ring_buffer(struct drm_device
*dev
,
1427 struct intel_engine_cs
*ring
)
1429 struct intel_ringbuffer
*ringbuf
= ring
->buffer
;
1432 if (ringbuf
== NULL
) {
1433 ringbuf
= kzalloc(sizeof(*ringbuf
), GFP_KERNEL
);
1436 ring
->buffer
= ringbuf
;
1440 INIT_LIST_HEAD(&ring
->active_list
);
1441 INIT_LIST_HEAD(&ring
->request_list
);
1442 ringbuf
->size
= 32 * PAGE_SIZE
;
1443 memset(ring
->semaphore
.sync_seqno
, 0, sizeof(ring
->semaphore
.sync_seqno
));
1445 init_waitqueue_head(&ring
->irq_queue
);
1447 if (I915_NEED_GFX_HWS(dev
)) {
1448 ret
= init_status_page(ring
);
1452 BUG_ON(ring
->id
!= RCS
);
1453 ret
= init_phys_status_page(ring
);
1458 ret
= allocate_ring_buffer(ring
);
1460 DRM_ERROR("Failed to allocate ringbuffer %s: %d\n", ring
->name
, ret
);
1464 /* Workaround an erratum on the i830 which causes a hang if
1465 * the TAIL pointer points to within the last 2 cachelines
1468 ringbuf
->effective_size
= ringbuf
->size
;
1469 if (IS_I830(dev
) || IS_845G(dev
))
1470 ringbuf
->effective_size
-= 2 * CACHELINE_BYTES
;
1472 ret
= i915_cmd_parser_init_ring(ring
);
1476 ret
= ring
->init(ring
);
1484 ring
->buffer
= NULL
;
1488 void intel_cleanup_ring_buffer(struct intel_engine_cs
*ring
)
1490 struct drm_i915_private
*dev_priv
= to_i915(ring
->dev
);
1491 struct intel_ringbuffer
*ringbuf
= ring
->buffer
;
1493 if (!intel_ring_initialized(ring
))
1496 intel_stop_ring_buffer(ring
);
1497 WARN_ON(!IS_GEN2(ring
->dev
) && (I915_READ_MODE(ring
) & MODE_IDLE
) == 0);
1499 iounmap(ringbuf
->virtual_start
);
1501 i915_gem_object_ggtt_unpin(ringbuf
->obj
);
1502 drm_gem_object_unreference(&ringbuf
->obj
->base
);
1503 ringbuf
->obj
= NULL
;
1504 ring
->preallocated_lazy_request
= NULL
;
1505 ring
->outstanding_lazy_seqno
= 0;
1508 ring
->cleanup(ring
);
1510 cleanup_status_page(ring
);
1512 i915_cmd_parser_fini_ring(ring
);
1515 ring
->buffer
= NULL
;
1518 static int intel_ring_wait_request(struct intel_engine_cs
*ring
, int n
)
1520 struct intel_ringbuffer
*ringbuf
= ring
->buffer
;
1521 struct drm_i915_gem_request
*request
;
1525 if (ringbuf
->last_retired_head
!= -1) {
1526 ringbuf
->head
= ringbuf
->last_retired_head
;
1527 ringbuf
->last_retired_head
= -1;
1529 ringbuf
->space
= ring_space(ring
);
1530 if (ringbuf
->space
>= n
)
1534 list_for_each_entry(request
, &ring
->request_list
, list
) {
1535 if (__ring_space(request
->tail
, ringbuf
->tail
, ringbuf
->size
) >= n
) {
1536 seqno
= request
->seqno
;
1544 ret
= i915_wait_seqno(ring
, seqno
);
1548 i915_gem_retire_requests_ring(ring
);
1549 ringbuf
->head
= ringbuf
->last_retired_head
;
1550 ringbuf
->last_retired_head
= -1;
1552 ringbuf
->space
= ring_space(ring
);
1556 static int ring_wait_for_space(struct intel_engine_cs
*ring
, int n
)
1558 struct drm_device
*dev
= ring
->dev
;
1559 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1560 struct intel_ringbuffer
*ringbuf
= ring
->buffer
;
1564 ret
= intel_ring_wait_request(ring
, n
);
1568 /* force the tail write in case we have been skipping them */
1569 __intel_ring_advance(ring
);
1571 /* With GEM the hangcheck timer should kick us out of the loop,
1572 * leaving it early runs the risk of corrupting GEM state (due
1573 * to running on almost untested codepaths). But on resume
1574 * timers don't work yet, so prevent a complete hang in that
1575 * case by choosing an insanely large timeout. */
1576 end
= jiffies
+ 60 * HZ
;
1578 trace_i915_ring_wait_begin(ring
);
1580 ringbuf
->head
= I915_READ_HEAD(ring
);
1581 ringbuf
->space
= ring_space(ring
);
1582 if (ringbuf
->space
>= n
) {
1587 if (!drm_core_check_feature(dev
, DRIVER_MODESET
) &&
1588 dev
->primary
->master
) {
1589 struct drm_i915_master_private
*master_priv
= dev
->primary
->master
->driver_priv
;
1590 if (master_priv
->sarea_priv
)
1591 master_priv
->sarea_priv
->perf_boxes
|= I915_BOX_WAIT
;
1596 if (dev_priv
->mm
.interruptible
&& signal_pending(current
)) {
1601 ret
= i915_gem_check_wedge(&dev_priv
->gpu_error
,
1602 dev_priv
->mm
.interruptible
);
1606 if (time_after(jiffies
, end
)) {
1611 trace_i915_ring_wait_end(ring
);
1615 static int intel_wrap_ring_buffer(struct intel_engine_cs
*ring
)
1617 uint32_t __iomem
*virt
;
1618 struct intel_ringbuffer
*ringbuf
= ring
->buffer
;
1619 int rem
= ringbuf
->size
- ringbuf
->tail
;
1621 if (ringbuf
->space
< rem
) {
1622 int ret
= ring_wait_for_space(ring
, rem
);
1627 virt
= ringbuf
->virtual_start
+ ringbuf
->tail
;
1630 iowrite32(MI_NOOP
, virt
++);
1633 ringbuf
->space
= ring_space(ring
);
1638 int intel_ring_idle(struct intel_engine_cs
*ring
)
1643 /* We need to add any requests required to flush the objects and ring */
1644 if (ring
->outstanding_lazy_seqno
) {
1645 ret
= i915_add_request(ring
, NULL
);
1650 /* Wait upon the last request to be completed */
1651 if (list_empty(&ring
->request_list
))
1654 seqno
= list_entry(ring
->request_list
.prev
,
1655 struct drm_i915_gem_request
,
1658 return i915_wait_seqno(ring
, seqno
);
1662 intel_ring_alloc_seqno(struct intel_engine_cs
*ring
)
1664 if (ring
->outstanding_lazy_seqno
)
1667 if (ring
->preallocated_lazy_request
== NULL
) {
1668 struct drm_i915_gem_request
*request
;
1670 request
= kmalloc(sizeof(*request
), GFP_KERNEL
);
1671 if (request
== NULL
)
1674 ring
->preallocated_lazy_request
= request
;
1677 return i915_gem_get_seqno(ring
->dev
, &ring
->outstanding_lazy_seqno
);
1680 static int __intel_ring_prepare(struct intel_engine_cs
*ring
,
1683 struct intel_ringbuffer
*ringbuf
= ring
->buffer
;
1686 if (unlikely(ringbuf
->tail
+ bytes
> ringbuf
->effective_size
)) {
1687 ret
= intel_wrap_ring_buffer(ring
);
1692 if (unlikely(ringbuf
->space
< bytes
)) {
1693 ret
= ring_wait_for_space(ring
, bytes
);
1701 int intel_ring_begin(struct intel_engine_cs
*ring
,
1704 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
1707 ret
= i915_gem_check_wedge(&dev_priv
->gpu_error
,
1708 dev_priv
->mm
.interruptible
);
1712 ret
= __intel_ring_prepare(ring
, num_dwords
* sizeof(uint32_t));
1716 /* Preallocate the olr before touching the ring */
1717 ret
= intel_ring_alloc_seqno(ring
);
1721 ring
->buffer
->space
-= num_dwords
* sizeof(uint32_t);
1725 /* Align the ring tail to a cacheline boundary */
1726 int intel_ring_cacheline_align(struct intel_engine_cs
*ring
)
1728 int num_dwords
= (ring
->buffer
->tail
& (CACHELINE_BYTES
- 1)) / sizeof(uint32_t);
1731 if (num_dwords
== 0)
1734 num_dwords
= CACHELINE_BYTES
/ sizeof(uint32_t) - num_dwords
;
1735 ret
= intel_ring_begin(ring
, num_dwords
);
1739 while (num_dwords
--)
1740 intel_ring_emit(ring
, MI_NOOP
);
1742 intel_ring_advance(ring
);
1747 void intel_ring_init_seqno(struct intel_engine_cs
*ring
, u32 seqno
)
1749 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
1751 BUG_ON(ring
->outstanding_lazy_seqno
);
1753 if (INTEL_INFO(ring
->dev
)->gen
>= 6) {
1754 I915_WRITE(RING_SYNC_0(ring
->mmio_base
), 0);
1755 I915_WRITE(RING_SYNC_1(ring
->mmio_base
), 0);
1756 if (HAS_VEBOX(ring
->dev
))
1757 I915_WRITE(RING_SYNC_2(ring
->mmio_base
), 0);
1760 ring
->set_seqno(ring
, seqno
);
1761 ring
->hangcheck
.seqno
= seqno
;
1764 static void gen6_bsd_ring_write_tail(struct intel_engine_cs
*ring
,
1767 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
1769 /* Every tail move must follow the sequence below */
1771 /* Disable notification that the ring is IDLE. The GT
1772 * will then assume that it is busy and bring it out of rc6.
1774 I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL
,
1775 _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE
));
1777 /* Clear the context id. Here be magic! */
1778 I915_WRITE64(GEN6_BSD_RNCID
, 0x0);
1780 /* Wait for the ring not to be idle, i.e. for it to wake up. */
1781 if (wait_for((I915_READ(GEN6_BSD_SLEEP_PSMI_CONTROL
) &
1782 GEN6_BSD_SLEEP_INDICATOR
) == 0,
1784 DRM_ERROR("timed out waiting for the BSD ring to wake up\n");
1786 /* Now that the ring is fully powered up, update the tail */
1787 I915_WRITE_TAIL(ring
, value
);
1788 POSTING_READ(RING_TAIL(ring
->mmio_base
));
1790 /* Let the ring send IDLE messages to the GT again,
1791 * and so let it sleep to conserve power when idle.
1793 I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL
,
1794 _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE
));
1797 static int gen6_bsd_ring_flush(struct intel_engine_cs
*ring
,
1798 u32 invalidate
, u32 flush
)
1803 ret
= intel_ring_begin(ring
, 4);
1808 if (INTEL_INFO(ring
->dev
)->gen
>= 8)
1811 * Bspec vol 1c.5 - video engine command streamer:
1812 * "If ENABLED, all TLBs will be invalidated once the flush
1813 * operation is complete. This bit is only valid when the
1814 * Post-Sync Operation field is a value of 1h or 3h."
1816 if (invalidate
& I915_GEM_GPU_DOMAINS
)
1817 cmd
|= MI_INVALIDATE_TLB
| MI_INVALIDATE_BSD
|
1818 MI_FLUSH_DW_STORE_INDEX
| MI_FLUSH_DW_OP_STOREDW
;
1819 intel_ring_emit(ring
, cmd
);
1820 intel_ring_emit(ring
, I915_GEM_HWS_SCRATCH_ADDR
| MI_FLUSH_DW_USE_GTT
);
1821 if (INTEL_INFO(ring
->dev
)->gen
>= 8) {
1822 intel_ring_emit(ring
, 0); /* upper addr */
1823 intel_ring_emit(ring
, 0); /* value */
1825 intel_ring_emit(ring
, 0);
1826 intel_ring_emit(ring
, MI_NOOP
);
1828 intel_ring_advance(ring
);
1833 gen8_ring_dispatch_execbuffer(struct intel_engine_cs
*ring
,
1834 u64 offset
, u32 len
,
1837 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
1838 bool ppgtt
= dev_priv
->mm
.aliasing_ppgtt
!= NULL
&&
1839 !(flags
& I915_DISPATCH_SECURE
);
1842 ret
= intel_ring_begin(ring
, 4);
1846 /* FIXME(BDW): Address space and security selectors. */
1847 intel_ring_emit(ring
, MI_BATCH_BUFFER_START_GEN8
| (ppgtt
<<8));
1848 intel_ring_emit(ring
, lower_32_bits(offset
));
1849 intel_ring_emit(ring
, upper_32_bits(offset
));
1850 intel_ring_emit(ring
, MI_NOOP
);
1851 intel_ring_advance(ring
);
1857 hsw_ring_dispatch_execbuffer(struct intel_engine_cs
*ring
,
1858 u64 offset
, u32 len
,
1863 ret
= intel_ring_begin(ring
, 2);
1867 intel_ring_emit(ring
,
1868 MI_BATCH_BUFFER_START
| MI_BATCH_PPGTT_HSW
|
1869 (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE_HSW
));
1870 /* bit0-7 is the length on GEN6+ */
1871 intel_ring_emit(ring
, offset
);
1872 intel_ring_advance(ring
);
1878 gen6_ring_dispatch_execbuffer(struct intel_engine_cs
*ring
,
1879 u64 offset
, u32 len
,
1884 ret
= intel_ring_begin(ring
, 2);
1888 intel_ring_emit(ring
,
1889 MI_BATCH_BUFFER_START
|
1890 (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE_I965
));
1891 /* bit0-7 is the length on GEN6+ */
1892 intel_ring_emit(ring
, offset
);
1893 intel_ring_advance(ring
);
1898 /* Blitter support (SandyBridge+) */
1900 static int gen6_ring_flush(struct intel_engine_cs
*ring
,
1901 u32 invalidate
, u32 flush
)
1903 struct drm_device
*dev
= ring
->dev
;
1907 ret
= intel_ring_begin(ring
, 4);
1912 if (INTEL_INFO(ring
->dev
)->gen
>= 8)
1915 * Bspec vol 1c.3 - blitter engine command streamer:
1916 * "If ENABLED, all TLBs will be invalidated once the flush
1917 * operation is complete. This bit is only valid when the
1918 * Post-Sync Operation field is a value of 1h or 3h."
1920 if (invalidate
& I915_GEM_DOMAIN_RENDER
)
1921 cmd
|= MI_INVALIDATE_TLB
| MI_FLUSH_DW_STORE_INDEX
|
1922 MI_FLUSH_DW_OP_STOREDW
;
1923 intel_ring_emit(ring
, cmd
);
1924 intel_ring_emit(ring
, I915_GEM_HWS_SCRATCH_ADDR
| MI_FLUSH_DW_USE_GTT
);
1925 if (INTEL_INFO(ring
->dev
)->gen
>= 8) {
1926 intel_ring_emit(ring
, 0); /* upper addr */
1927 intel_ring_emit(ring
, 0); /* value */
1929 intel_ring_emit(ring
, 0);
1930 intel_ring_emit(ring
, MI_NOOP
);
1932 intel_ring_advance(ring
);
1934 if (IS_GEN7(dev
) && !invalidate
&& flush
)
1935 return gen7_ring_fbc_flush(ring
, FBC_REND_CACHE_CLEAN
);
1940 int intel_init_render_ring_buffer(struct drm_device
*dev
)
1942 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1943 struct intel_engine_cs
*ring
= &dev_priv
->ring
[RCS
];
1945 ring
->name
= "render ring";
1947 ring
->mmio_base
= RENDER_RING_BASE
;
1949 if (INTEL_INFO(dev
)->gen
>= 6) {
1950 ring
->add_request
= gen6_add_request
;
1951 ring
->flush
= gen7_render_ring_flush
;
1952 if (INTEL_INFO(dev
)->gen
== 6)
1953 ring
->flush
= gen6_render_ring_flush
;
1954 if (INTEL_INFO(dev
)->gen
>= 8) {
1955 ring
->flush
= gen8_render_ring_flush
;
1956 ring
->irq_get
= gen8_ring_get_irq
;
1957 ring
->irq_put
= gen8_ring_put_irq
;
1959 ring
->irq_get
= gen6_ring_get_irq
;
1960 ring
->irq_put
= gen6_ring_put_irq
;
1962 ring
->irq_enable_mask
= GT_RENDER_USER_INTERRUPT
;
1963 ring
->get_seqno
= gen6_ring_get_seqno
;
1964 ring
->set_seqno
= ring_set_seqno
;
1965 ring
->semaphore
.sync_to
= gen6_ring_sync
;
1966 ring
->semaphore
.signal
= gen6_signal
;
1968 * The current semaphore is only applied on pre-gen8 platform.
1969 * And there is no VCS2 ring on the pre-gen8 platform. So the
1970 * semaphore between RCS and VCS2 is initialized as INVALID.
1971 * Gen8 will initialize the sema between VCS2 and RCS later.
1973 ring
->semaphore
.mbox
.wait
[RCS
] = MI_SEMAPHORE_SYNC_INVALID
;
1974 ring
->semaphore
.mbox
.wait
[VCS
] = MI_SEMAPHORE_SYNC_RV
;
1975 ring
->semaphore
.mbox
.wait
[BCS
] = MI_SEMAPHORE_SYNC_RB
;
1976 ring
->semaphore
.mbox
.wait
[VECS
] = MI_SEMAPHORE_SYNC_RVE
;
1977 ring
->semaphore
.mbox
.wait
[VCS2
] = MI_SEMAPHORE_SYNC_INVALID
;
1978 ring
->semaphore
.mbox
.signal
[RCS
] = GEN6_NOSYNC
;
1979 ring
->semaphore
.mbox
.signal
[VCS
] = GEN6_VRSYNC
;
1980 ring
->semaphore
.mbox
.signal
[BCS
] = GEN6_BRSYNC
;
1981 ring
->semaphore
.mbox
.signal
[VECS
] = GEN6_VERSYNC
;
1982 ring
->semaphore
.mbox
.signal
[VCS2
] = GEN6_NOSYNC
;
1983 } else if (IS_GEN5(dev
)) {
1984 ring
->add_request
= pc_render_add_request
;
1985 ring
->flush
= gen4_render_ring_flush
;
1986 ring
->get_seqno
= pc_render_get_seqno
;
1987 ring
->set_seqno
= pc_render_set_seqno
;
1988 ring
->irq_get
= gen5_ring_get_irq
;
1989 ring
->irq_put
= gen5_ring_put_irq
;
1990 ring
->irq_enable_mask
= GT_RENDER_USER_INTERRUPT
|
1991 GT_RENDER_PIPECTL_NOTIFY_INTERRUPT
;
1993 ring
->add_request
= i9xx_add_request
;
1994 if (INTEL_INFO(dev
)->gen
< 4)
1995 ring
->flush
= gen2_render_ring_flush
;
1997 ring
->flush
= gen4_render_ring_flush
;
1998 ring
->get_seqno
= ring_get_seqno
;
1999 ring
->set_seqno
= ring_set_seqno
;
2001 ring
->irq_get
= i8xx_ring_get_irq
;
2002 ring
->irq_put
= i8xx_ring_put_irq
;
2004 ring
->irq_get
= i9xx_ring_get_irq
;
2005 ring
->irq_put
= i9xx_ring_put_irq
;
2007 ring
->irq_enable_mask
= I915_USER_INTERRUPT
;
2009 ring
->write_tail
= ring_write_tail
;
2010 if (IS_HASWELL(dev
))
2011 ring
->dispatch_execbuffer
= hsw_ring_dispatch_execbuffer
;
2012 else if (IS_GEN8(dev
))
2013 ring
->dispatch_execbuffer
= gen8_ring_dispatch_execbuffer
;
2014 else if (INTEL_INFO(dev
)->gen
>= 6)
2015 ring
->dispatch_execbuffer
= gen6_ring_dispatch_execbuffer
;
2016 else if (INTEL_INFO(dev
)->gen
>= 4)
2017 ring
->dispatch_execbuffer
= i965_dispatch_execbuffer
;
2018 else if (IS_I830(dev
) || IS_845G(dev
))
2019 ring
->dispatch_execbuffer
= i830_dispatch_execbuffer
;
2021 ring
->dispatch_execbuffer
= i915_dispatch_execbuffer
;
2022 ring
->init
= init_render_ring
;
2023 ring
->cleanup
= render_ring_cleanup
;
2025 /* Workaround batchbuffer to combat CS tlb bug. */
2026 if (HAS_BROKEN_CS_TLB(dev
)) {
2027 struct drm_i915_gem_object
*obj
;
2030 obj
= i915_gem_alloc_object(dev
, I830_BATCH_LIMIT
);
2032 DRM_ERROR("Failed to allocate batch bo\n");
2036 ret
= i915_gem_obj_ggtt_pin(obj
, 0, 0);
2038 drm_gem_object_unreference(&obj
->base
);
2039 DRM_ERROR("Failed to ping batch bo\n");
2043 ring
->scratch
.obj
= obj
;
2044 ring
->scratch
.gtt_offset
= i915_gem_obj_ggtt_offset(obj
);
2047 return intel_init_ring_buffer(dev
, ring
);
2050 int intel_render_ring_init_dri(struct drm_device
*dev
, u64 start
, u32 size
)
2052 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2053 struct intel_engine_cs
*ring
= &dev_priv
->ring
[RCS
];
2054 struct intel_ringbuffer
*ringbuf
= ring
->buffer
;
2057 if (ringbuf
== NULL
) {
2058 ringbuf
= kzalloc(sizeof(*ringbuf
), GFP_KERNEL
);
2061 ring
->buffer
= ringbuf
;
2064 ring
->name
= "render ring";
2066 ring
->mmio_base
= RENDER_RING_BASE
;
2068 if (INTEL_INFO(dev
)->gen
>= 6) {
2069 /* non-kms not supported on gen6+ */
2074 /* Note: gem is not supported on gen5/ilk without kms (the corresponding
2075 * gem_init ioctl returns with -ENODEV). Hence we do not need to set up
2076 * the special gen5 functions. */
2077 ring
->add_request
= i9xx_add_request
;
2078 if (INTEL_INFO(dev
)->gen
< 4)
2079 ring
->flush
= gen2_render_ring_flush
;
2081 ring
->flush
= gen4_render_ring_flush
;
2082 ring
->get_seqno
= ring_get_seqno
;
2083 ring
->set_seqno
= ring_set_seqno
;
2085 ring
->irq_get
= i8xx_ring_get_irq
;
2086 ring
->irq_put
= i8xx_ring_put_irq
;
2088 ring
->irq_get
= i9xx_ring_get_irq
;
2089 ring
->irq_put
= i9xx_ring_put_irq
;
2091 ring
->irq_enable_mask
= I915_USER_INTERRUPT
;
2092 ring
->write_tail
= ring_write_tail
;
2093 if (INTEL_INFO(dev
)->gen
>= 4)
2094 ring
->dispatch_execbuffer
= i965_dispatch_execbuffer
;
2095 else if (IS_I830(dev
) || IS_845G(dev
))
2096 ring
->dispatch_execbuffer
= i830_dispatch_execbuffer
;
2098 ring
->dispatch_execbuffer
= i915_dispatch_execbuffer
;
2099 ring
->init
= init_render_ring
;
2100 ring
->cleanup
= render_ring_cleanup
;
2103 INIT_LIST_HEAD(&ring
->active_list
);
2104 INIT_LIST_HEAD(&ring
->request_list
);
2106 ringbuf
->size
= size
;
2107 ringbuf
->effective_size
= ringbuf
->size
;
2108 if (IS_I830(ring
->dev
) || IS_845G(ring
->dev
))
2109 ringbuf
->effective_size
-= 2 * CACHELINE_BYTES
;
2111 ringbuf
->virtual_start
= ioremap_wc(start
, size
);
2112 if (ringbuf
->virtual_start
== NULL
) {
2113 DRM_ERROR("can not ioremap virtual address for"
2119 if (!I915_NEED_GFX_HWS(dev
)) {
2120 ret
= init_phys_status_page(ring
);
2128 iounmap(ringbuf
->virtual_start
);
2131 ring
->buffer
= NULL
;
2135 int intel_init_bsd_ring_buffer(struct drm_device
*dev
)
2137 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2138 struct intel_engine_cs
*ring
= &dev_priv
->ring
[VCS
];
2140 ring
->name
= "bsd ring";
2143 ring
->write_tail
= ring_write_tail
;
2144 if (INTEL_INFO(dev
)->gen
>= 6) {
2145 ring
->mmio_base
= GEN6_BSD_RING_BASE
;
2146 /* gen6 bsd needs a special wa for tail updates */
2148 ring
->write_tail
= gen6_bsd_ring_write_tail
;
2149 ring
->flush
= gen6_bsd_ring_flush
;
2150 ring
->add_request
= gen6_add_request
;
2151 ring
->get_seqno
= gen6_ring_get_seqno
;
2152 ring
->set_seqno
= ring_set_seqno
;
2153 if (INTEL_INFO(dev
)->gen
>= 8) {
2154 ring
->irq_enable_mask
=
2155 GT_RENDER_USER_INTERRUPT
<< GEN8_VCS1_IRQ_SHIFT
;
2156 ring
->irq_get
= gen8_ring_get_irq
;
2157 ring
->irq_put
= gen8_ring_put_irq
;
2158 ring
->dispatch_execbuffer
=
2159 gen8_ring_dispatch_execbuffer
;
2161 ring
->irq_enable_mask
= GT_BSD_USER_INTERRUPT
;
2162 ring
->irq_get
= gen6_ring_get_irq
;
2163 ring
->irq_put
= gen6_ring_put_irq
;
2164 ring
->dispatch_execbuffer
=
2165 gen6_ring_dispatch_execbuffer
;
2167 ring
->semaphore
.sync_to
= gen6_ring_sync
;
2168 ring
->semaphore
.signal
= gen6_signal
;
2170 * The current semaphore is only applied on pre-gen8 platform.
2171 * And there is no VCS2 ring on the pre-gen8 platform. So the
2172 * semaphore between VCS and VCS2 is initialized as INVALID.
2173 * Gen8 will initialize the sema between VCS2 and VCS later.
2175 ring
->semaphore
.mbox
.wait
[RCS
] = MI_SEMAPHORE_SYNC_VR
;
2176 ring
->semaphore
.mbox
.wait
[VCS
] = MI_SEMAPHORE_SYNC_INVALID
;
2177 ring
->semaphore
.mbox
.wait
[BCS
] = MI_SEMAPHORE_SYNC_VB
;
2178 ring
->semaphore
.mbox
.wait
[VECS
] = MI_SEMAPHORE_SYNC_VVE
;
2179 ring
->semaphore
.mbox
.wait
[VCS2
] = MI_SEMAPHORE_SYNC_INVALID
;
2180 ring
->semaphore
.mbox
.signal
[RCS
] = GEN6_RVSYNC
;
2181 ring
->semaphore
.mbox
.signal
[VCS
] = GEN6_NOSYNC
;
2182 ring
->semaphore
.mbox
.signal
[BCS
] = GEN6_BVSYNC
;
2183 ring
->semaphore
.mbox
.signal
[VECS
] = GEN6_VEVSYNC
;
2184 ring
->semaphore
.mbox
.signal
[VCS2
] = GEN6_NOSYNC
;
2186 ring
->mmio_base
= BSD_RING_BASE
;
2187 ring
->flush
= bsd_ring_flush
;
2188 ring
->add_request
= i9xx_add_request
;
2189 ring
->get_seqno
= ring_get_seqno
;
2190 ring
->set_seqno
= ring_set_seqno
;
2192 ring
->irq_enable_mask
= ILK_BSD_USER_INTERRUPT
;
2193 ring
->irq_get
= gen5_ring_get_irq
;
2194 ring
->irq_put
= gen5_ring_put_irq
;
2196 ring
->irq_enable_mask
= I915_BSD_USER_INTERRUPT
;
2197 ring
->irq_get
= i9xx_ring_get_irq
;
2198 ring
->irq_put
= i9xx_ring_put_irq
;
2200 ring
->dispatch_execbuffer
= i965_dispatch_execbuffer
;
2202 ring
->init
= init_ring_common
;
2204 return intel_init_ring_buffer(dev
, ring
);
2208 * Initialize the second BSD ring for Broadwell GT3.
2209 * It is noted that this only exists on Broadwell GT3.
2211 int intel_init_bsd2_ring_buffer(struct drm_device
*dev
)
2213 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2214 struct intel_engine_cs
*ring
= &dev_priv
->ring
[VCS2
];
2216 if ((INTEL_INFO(dev
)->gen
!= 8)) {
2217 DRM_ERROR("No dual-BSD ring on non-BDW machine\n");
2221 ring
->name
= "bds2_ring";
2224 ring
->write_tail
= ring_write_tail
;
2225 ring
->mmio_base
= GEN8_BSD2_RING_BASE
;
2226 ring
->flush
= gen6_bsd_ring_flush
;
2227 ring
->add_request
= gen6_add_request
;
2228 ring
->get_seqno
= gen6_ring_get_seqno
;
2229 ring
->set_seqno
= ring_set_seqno
;
2230 ring
->irq_enable_mask
=
2231 GT_RENDER_USER_INTERRUPT
<< GEN8_VCS2_IRQ_SHIFT
;
2232 ring
->irq_get
= gen8_ring_get_irq
;
2233 ring
->irq_put
= gen8_ring_put_irq
;
2234 ring
->dispatch_execbuffer
=
2235 gen8_ring_dispatch_execbuffer
;
2236 ring
->semaphore
.sync_to
= gen6_ring_sync
;
2237 ring
->semaphore
.signal
= gen6_signal
;
2239 * The current semaphore is only applied on the pre-gen8. And there
2240 * is no bsd2 ring on the pre-gen8. So now the semaphore_register
2241 * between VCS2 and other ring is initialized as invalid.
2242 * Gen8 will initialize the sema between VCS2 and other ring later.
2244 ring
->semaphore
.mbox
.wait
[RCS
] = MI_SEMAPHORE_SYNC_INVALID
;
2245 ring
->semaphore
.mbox
.wait
[VCS
] = MI_SEMAPHORE_SYNC_INVALID
;
2246 ring
->semaphore
.mbox
.wait
[BCS
] = MI_SEMAPHORE_SYNC_INVALID
;
2247 ring
->semaphore
.mbox
.wait
[VECS
] = MI_SEMAPHORE_SYNC_INVALID
;
2248 ring
->semaphore
.mbox
.wait
[VCS2
] = MI_SEMAPHORE_SYNC_INVALID
;
2249 ring
->semaphore
.mbox
.signal
[RCS
] = GEN6_NOSYNC
;
2250 ring
->semaphore
.mbox
.signal
[VCS
] = GEN6_NOSYNC
;
2251 ring
->semaphore
.mbox
.signal
[BCS
] = GEN6_NOSYNC
;
2252 ring
->semaphore
.mbox
.signal
[VECS
] = GEN6_NOSYNC
;
2253 ring
->semaphore
.mbox
.signal
[VCS2
] = GEN6_NOSYNC
;
2255 ring
->init
= init_ring_common
;
2257 return intel_init_ring_buffer(dev
, ring
);
2260 int intel_init_blt_ring_buffer(struct drm_device
*dev
)
2262 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2263 struct intel_engine_cs
*ring
= &dev_priv
->ring
[BCS
];
2265 ring
->name
= "blitter ring";
2268 ring
->mmio_base
= BLT_RING_BASE
;
2269 ring
->write_tail
= ring_write_tail
;
2270 ring
->flush
= gen6_ring_flush
;
2271 ring
->add_request
= gen6_add_request
;
2272 ring
->get_seqno
= gen6_ring_get_seqno
;
2273 ring
->set_seqno
= ring_set_seqno
;
2274 if (INTEL_INFO(dev
)->gen
>= 8) {
2275 ring
->irq_enable_mask
=
2276 GT_RENDER_USER_INTERRUPT
<< GEN8_BCS_IRQ_SHIFT
;
2277 ring
->irq_get
= gen8_ring_get_irq
;
2278 ring
->irq_put
= gen8_ring_put_irq
;
2279 ring
->dispatch_execbuffer
= gen8_ring_dispatch_execbuffer
;
2281 ring
->irq_enable_mask
= GT_BLT_USER_INTERRUPT
;
2282 ring
->irq_get
= gen6_ring_get_irq
;
2283 ring
->irq_put
= gen6_ring_put_irq
;
2284 ring
->dispatch_execbuffer
= gen6_ring_dispatch_execbuffer
;
2286 ring
->semaphore
.sync_to
= gen6_ring_sync
;
2287 ring
->semaphore
.signal
= gen6_signal
;
2289 * The current semaphore is only applied on pre-gen8 platform. And
2290 * there is no VCS2 ring on the pre-gen8 platform. So the semaphore
2291 * between BCS and VCS2 is initialized as INVALID.
2292 * Gen8 will initialize the sema between BCS and VCS2 later.
2294 ring
->semaphore
.mbox
.wait
[RCS
] = MI_SEMAPHORE_SYNC_BR
;
2295 ring
->semaphore
.mbox
.wait
[VCS
] = MI_SEMAPHORE_SYNC_BV
;
2296 ring
->semaphore
.mbox
.wait
[BCS
] = MI_SEMAPHORE_SYNC_INVALID
;
2297 ring
->semaphore
.mbox
.wait
[VECS
] = MI_SEMAPHORE_SYNC_BVE
;
2298 ring
->semaphore
.mbox
.wait
[VCS2
] = MI_SEMAPHORE_SYNC_INVALID
;
2299 ring
->semaphore
.mbox
.signal
[RCS
] = GEN6_RBSYNC
;
2300 ring
->semaphore
.mbox
.signal
[VCS
] = GEN6_VBSYNC
;
2301 ring
->semaphore
.mbox
.signal
[BCS
] = GEN6_NOSYNC
;
2302 ring
->semaphore
.mbox
.signal
[VECS
] = GEN6_VEBSYNC
;
2303 ring
->semaphore
.mbox
.signal
[VCS2
] = GEN6_NOSYNC
;
2304 ring
->init
= init_ring_common
;
2306 return intel_init_ring_buffer(dev
, ring
);
2309 int intel_init_vebox_ring_buffer(struct drm_device
*dev
)
2311 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2312 struct intel_engine_cs
*ring
= &dev_priv
->ring
[VECS
];
2314 ring
->name
= "video enhancement ring";
2317 ring
->mmio_base
= VEBOX_RING_BASE
;
2318 ring
->write_tail
= ring_write_tail
;
2319 ring
->flush
= gen6_ring_flush
;
2320 ring
->add_request
= gen6_add_request
;
2321 ring
->get_seqno
= gen6_ring_get_seqno
;
2322 ring
->set_seqno
= ring_set_seqno
;
2324 if (INTEL_INFO(dev
)->gen
>= 8) {
2325 ring
->irq_enable_mask
=
2326 GT_RENDER_USER_INTERRUPT
<< GEN8_VECS_IRQ_SHIFT
;
2327 ring
->irq_get
= gen8_ring_get_irq
;
2328 ring
->irq_put
= gen8_ring_put_irq
;
2329 ring
->dispatch_execbuffer
= gen8_ring_dispatch_execbuffer
;
2331 ring
->irq_enable_mask
= PM_VEBOX_USER_INTERRUPT
;
2332 ring
->irq_get
= hsw_vebox_get_irq
;
2333 ring
->irq_put
= hsw_vebox_put_irq
;
2334 ring
->dispatch_execbuffer
= gen6_ring_dispatch_execbuffer
;
2336 ring
->semaphore
.sync_to
= gen6_ring_sync
;
2337 ring
->semaphore
.signal
= gen6_signal
;
2338 ring
->semaphore
.mbox
.wait
[RCS
] = MI_SEMAPHORE_SYNC_VER
;
2339 ring
->semaphore
.mbox
.wait
[VCS
] = MI_SEMAPHORE_SYNC_VEV
;
2340 ring
->semaphore
.mbox
.wait
[BCS
] = MI_SEMAPHORE_SYNC_VEB
;
2341 ring
->semaphore
.mbox
.wait
[VECS
] = MI_SEMAPHORE_SYNC_INVALID
;
2342 ring
->semaphore
.mbox
.wait
[VCS2
] = MI_SEMAPHORE_SYNC_INVALID
;
2343 ring
->semaphore
.mbox
.signal
[RCS
] = GEN6_RVESYNC
;
2344 ring
->semaphore
.mbox
.signal
[VCS
] = GEN6_VVESYNC
;
2345 ring
->semaphore
.mbox
.signal
[BCS
] = GEN6_BVESYNC
;
2346 ring
->semaphore
.mbox
.signal
[VECS
] = GEN6_NOSYNC
;
2347 ring
->semaphore
.mbox
.signal
[VCS2
] = GEN6_NOSYNC
;
2348 ring
->init
= init_ring_common
;
2350 return intel_init_ring_buffer(dev
, ring
);
2354 intel_ring_flush_all_caches(struct intel_engine_cs
*ring
)
2358 if (!ring
->gpu_caches_dirty
)
2361 ret
= ring
->flush(ring
, 0, I915_GEM_GPU_DOMAINS
);
2365 trace_i915_gem_ring_flush(ring
, 0, I915_GEM_GPU_DOMAINS
);
2367 ring
->gpu_caches_dirty
= false;
2372 intel_ring_invalidate_all_caches(struct intel_engine_cs
*ring
)
2374 uint32_t flush_domains
;
2378 if (ring
->gpu_caches_dirty
)
2379 flush_domains
= I915_GEM_GPU_DOMAINS
;
2381 ret
= ring
->flush(ring
, I915_GEM_GPU_DOMAINS
, flush_domains
);
2385 trace_i915_gem_ring_flush(ring
, I915_GEM_GPU_DOMAINS
, flush_domains
);
2387 ring
->gpu_caches_dirty
= false;
2392 intel_stop_ring_buffer(struct intel_engine_cs
*ring
)
2396 if (!intel_ring_initialized(ring
))
2399 ret
= intel_ring_idle(ring
);
2400 if (ret
&& !i915_reset_in_progress(&to_i915(ring
->dev
)->gpu_error
))
2401 DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",