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 static inline int ring_space(struct intel_ring_buffer
*ring
)
38 int space
= (ring
->head
& HEAD_ADDR
) - (ring
->tail
+ I915_RING_FREE_SPACE
);
44 void __intel_ring_advance(struct intel_ring_buffer
*ring
)
46 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
48 ring
->tail
&= ring
->size
- 1;
49 if (dev_priv
->gpu_error
.stop_rings
& intel_ring_flag(ring
))
51 ring
->write_tail(ring
, ring
->tail
);
55 gen2_render_ring_flush(struct intel_ring_buffer
*ring
,
56 u32 invalidate_domains
,
63 if (((invalidate_domains
|flush_domains
) & I915_GEM_DOMAIN_RENDER
) == 0)
64 cmd
|= MI_NO_WRITE_FLUSH
;
66 if (invalidate_domains
& I915_GEM_DOMAIN_SAMPLER
)
69 ret
= intel_ring_begin(ring
, 2);
73 intel_ring_emit(ring
, cmd
);
74 intel_ring_emit(ring
, MI_NOOP
);
75 intel_ring_advance(ring
);
81 gen4_render_ring_flush(struct intel_ring_buffer
*ring
,
82 u32 invalidate_domains
,
85 struct drm_device
*dev
= ring
->dev
;
92 * I915_GEM_DOMAIN_RENDER is always invalidated, but is
93 * only flushed if MI_NO_WRITE_FLUSH is unset. On 965, it is
94 * also flushed at 2d versus 3d pipeline switches.
98 * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
99 * MI_READ_FLUSH is set, and is always flushed on 965.
101 * I915_GEM_DOMAIN_COMMAND may not exist?
103 * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
104 * invalidated when MI_EXE_FLUSH is set.
106 * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
107 * invalidated with every MI_FLUSH.
111 * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
112 * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
113 * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
114 * are flushed at any MI_FLUSH.
117 cmd
= MI_FLUSH
| MI_NO_WRITE_FLUSH
;
118 if ((invalidate_domains
|flush_domains
) & I915_GEM_DOMAIN_RENDER
)
119 cmd
&= ~MI_NO_WRITE_FLUSH
;
120 if (invalidate_domains
& I915_GEM_DOMAIN_INSTRUCTION
)
123 if (invalidate_domains
& I915_GEM_DOMAIN_COMMAND
&&
124 (IS_G4X(dev
) || IS_GEN5(dev
)))
125 cmd
|= MI_INVALIDATE_ISP
;
127 ret
= intel_ring_begin(ring
, 2);
131 intel_ring_emit(ring
, cmd
);
132 intel_ring_emit(ring
, MI_NOOP
);
133 intel_ring_advance(ring
);
139 * Emits a PIPE_CONTROL with a non-zero post-sync operation, for
140 * implementing two workarounds on gen6. From section 1.4.7.1
141 * "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
143 * [DevSNB-C+{W/A}] Before any depth stall flush (including those
144 * produced by non-pipelined state commands), software needs to first
145 * send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
148 * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
149 * =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
151 * And the workaround for these two requires this workaround first:
153 * [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
154 * BEFORE the pipe-control with a post-sync op and no write-cache
157 * And this last workaround is tricky because of the requirements on
158 * that bit. From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
161 * "1 of the following must also be set:
162 * - Render Target Cache Flush Enable ([12] of DW1)
163 * - Depth Cache Flush Enable ([0] of DW1)
164 * - Stall at Pixel Scoreboard ([1] of DW1)
165 * - Depth Stall ([13] of DW1)
166 * - Post-Sync Operation ([13] of DW1)
167 * - Notify Enable ([8] of DW1)"
169 * The cache flushes require the workaround flush that triggered this
170 * one, so we can't use it. Depth stall would trigger the same.
171 * Post-sync nonzero is what triggered this second workaround, so we
172 * can't use that one either. Notify enable is IRQs, which aren't
173 * really our business. That leaves only stall at scoreboard.
176 intel_emit_post_sync_nonzero_flush(struct intel_ring_buffer
*ring
)
178 u32 scratch_addr
= ring
->scratch
.gtt_offset
+ 128;
182 ret
= intel_ring_begin(ring
, 6);
186 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(5));
187 intel_ring_emit(ring
, PIPE_CONTROL_CS_STALL
|
188 PIPE_CONTROL_STALL_AT_SCOREBOARD
);
189 intel_ring_emit(ring
, scratch_addr
| PIPE_CONTROL_GLOBAL_GTT
); /* address */
190 intel_ring_emit(ring
, 0); /* low dword */
191 intel_ring_emit(ring
, 0); /* high dword */
192 intel_ring_emit(ring
, MI_NOOP
);
193 intel_ring_advance(ring
);
195 ret
= intel_ring_begin(ring
, 6);
199 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(5));
200 intel_ring_emit(ring
, PIPE_CONTROL_QW_WRITE
);
201 intel_ring_emit(ring
, scratch_addr
| PIPE_CONTROL_GLOBAL_GTT
); /* address */
202 intel_ring_emit(ring
, 0);
203 intel_ring_emit(ring
, 0);
204 intel_ring_emit(ring
, MI_NOOP
);
205 intel_ring_advance(ring
);
211 gen6_render_ring_flush(struct intel_ring_buffer
*ring
,
212 u32 invalidate_domains
, u32 flush_domains
)
215 u32 scratch_addr
= ring
->scratch
.gtt_offset
+ 128;
218 /* Force SNB workarounds for PIPE_CONTROL flushes */
219 ret
= intel_emit_post_sync_nonzero_flush(ring
);
223 /* Just flush everything. Experiments have shown that reducing the
224 * number of bits based on the write domains has little performance
228 flags
|= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH
;
229 flags
|= PIPE_CONTROL_DEPTH_CACHE_FLUSH
;
231 * Ensure that any following seqno writes only happen
232 * when the render cache is indeed flushed.
234 flags
|= PIPE_CONTROL_CS_STALL
;
236 if (invalidate_domains
) {
237 flags
|= PIPE_CONTROL_TLB_INVALIDATE
;
238 flags
|= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE
;
239 flags
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
240 flags
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
241 flags
|= PIPE_CONTROL_CONST_CACHE_INVALIDATE
;
242 flags
|= PIPE_CONTROL_STATE_CACHE_INVALIDATE
;
244 * TLB invalidate requires a post-sync write.
246 flags
|= PIPE_CONTROL_QW_WRITE
| PIPE_CONTROL_CS_STALL
;
249 ret
= intel_ring_begin(ring
, 4);
253 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(4));
254 intel_ring_emit(ring
, flags
);
255 intel_ring_emit(ring
, scratch_addr
| PIPE_CONTROL_GLOBAL_GTT
);
256 intel_ring_emit(ring
, 0);
257 intel_ring_advance(ring
);
263 gen7_render_ring_cs_stall_wa(struct intel_ring_buffer
*ring
)
267 ret
= intel_ring_begin(ring
, 4);
271 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(4));
272 intel_ring_emit(ring
, PIPE_CONTROL_CS_STALL
|
273 PIPE_CONTROL_STALL_AT_SCOREBOARD
);
274 intel_ring_emit(ring
, 0);
275 intel_ring_emit(ring
, 0);
276 intel_ring_advance(ring
);
281 static int gen7_ring_fbc_flush(struct intel_ring_buffer
*ring
, u32 value
)
285 if (!ring
->fbc_dirty
)
288 ret
= intel_ring_begin(ring
, 6);
291 /* WaFbcNukeOn3DBlt:ivb/hsw */
292 intel_ring_emit(ring
, MI_LOAD_REGISTER_IMM(1));
293 intel_ring_emit(ring
, MSG_FBC_REND_STATE
);
294 intel_ring_emit(ring
, value
);
295 intel_ring_emit(ring
, MI_STORE_REGISTER_MEM(1) | MI_SRM_LRM_GLOBAL_GTT
);
296 intel_ring_emit(ring
, MSG_FBC_REND_STATE
);
297 intel_ring_emit(ring
, ring
->scratch
.gtt_offset
+ 256);
298 intel_ring_advance(ring
);
300 ring
->fbc_dirty
= false;
305 gen7_render_ring_flush(struct intel_ring_buffer
*ring
,
306 u32 invalidate_domains
, u32 flush_domains
)
309 u32 scratch_addr
= ring
->scratch
.gtt_offset
+ 128;
313 * Ensure that any following seqno writes only happen when the render
314 * cache is indeed flushed.
316 * Workaround: 4th PIPE_CONTROL command (except the ones with only
317 * read-cache invalidate bits set) must have the CS_STALL bit set. We
318 * don't try to be clever and just set it unconditionally.
320 flags
|= PIPE_CONTROL_CS_STALL
;
322 /* Just flush everything. Experiments have shown that reducing the
323 * number of bits based on the write domains has little performance
327 flags
|= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH
;
328 flags
|= PIPE_CONTROL_DEPTH_CACHE_FLUSH
;
330 if (invalidate_domains
) {
331 flags
|= PIPE_CONTROL_TLB_INVALIDATE
;
332 flags
|= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE
;
333 flags
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
334 flags
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
335 flags
|= PIPE_CONTROL_CONST_CACHE_INVALIDATE
;
336 flags
|= PIPE_CONTROL_STATE_CACHE_INVALIDATE
;
338 * TLB invalidate requires a post-sync write.
340 flags
|= PIPE_CONTROL_QW_WRITE
;
341 flags
|= PIPE_CONTROL_GLOBAL_GTT_IVB
;
343 /* Workaround: we must issue a pipe_control with CS-stall bit
344 * set before a pipe_control command that has the state cache
345 * invalidate bit set. */
346 gen7_render_ring_cs_stall_wa(ring
);
349 ret
= intel_ring_begin(ring
, 4);
353 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(4));
354 intel_ring_emit(ring
, flags
);
355 intel_ring_emit(ring
, scratch_addr
);
356 intel_ring_emit(ring
, 0);
357 intel_ring_advance(ring
);
359 if (!invalidate_domains
&& flush_domains
)
360 return gen7_ring_fbc_flush(ring
, FBC_REND_NUKE
);
366 gen8_render_ring_flush(struct intel_ring_buffer
*ring
,
367 u32 invalidate_domains
, u32 flush_domains
)
370 u32 scratch_addr
= ring
->scratch
.gtt_offset
+ 128;
373 flags
|= PIPE_CONTROL_CS_STALL
;
376 flags
|= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH
;
377 flags
|= PIPE_CONTROL_DEPTH_CACHE_FLUSH
;
379 if (invalidate_domains
) {
380 flags
|= PIPE_CONTROL_TLB_INVALIDATE
;
381 flags
|= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE
;
382 flags
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
383 flags
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
384 flags
|= PIPE_CONTROL_CONST_CACHE_INVALIDATE
;
385 flags
|= PIPE_CONTROL_STATE_CACHE_INVALIDATE
;
386 flags
|= PIPE_CONTROL_QW_WRITE
;
387 flags
|= PIPE_CONTROL_GLOBAL_GTT_IVB
;
390 ret
= intel_ring_begin(ring
, 6);
394 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(6));
395 intel_ring_emit(ring
, flags
);
396 intel_ring_emit(ring
, scratch_addr
);
397 intel_ring_emit(ring
, 0);
398 intel_ring_emit(ring
, 0);
399 intel_ring_emit(ring
, 0);
400 intel_ring_advance(ring
);
406 static void ring_write_tail(struct intel_ring_buffer
*ring
,
409 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
410 I915_WRITE_TAIL(ring
, value
);
413 u64
intel_ring_get_active_head(struct intel_ring_buffer
*ring
)
415 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
418 if (INTEL_INFO(ring
->dev
)->gen
>= 8)
419 acthd
= I915_READ64_2x32(RING_ACTHD(ring
->mmio_base
),
420 RING_ACTHD_UDW(ring
->mmio_base
));
421 else if (INTEL_INFO(ring
->dev
)->gen
>= 4)
422 acthd
= I915_READ(RING_ACTHD(ring
->mmio_base
));
424 acthd
= I915_READ(ACTHD
);
429 static void ring_setup_phys_status_page(struct intel_ring_buffer
*ring
)
431 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
434 addr
= dev_priv
->status_page_dmah
->busaddr
;
435 if (INTEL_INFO(ring
->dev
)->gen
>= 4)
436 addr
|= (dev_priv
->status_page_dmah
->busaddr
>> 28) & 0xf0;
437 I915_WRITE(HWS_PGA
, addr
);
440 static bool stop_ring(struct intel_ring_buffer
*ring
)
442 struct drm_i915_private
*dev_priv
= to_i915(ring
->dev
);
444 if (!IS_GEN2(ring
->dev
)) {
445 I915_WRITE_MODE(ring
, _MASKED_BIT_ENABLE(STOP_RING
));
446 if (wait_for_atomic((I915_READ_MODE(ring
) & MODE_IDLE
) != 0, 1000)) {
447 DRM_ERROR("%s :timed out trying to stop ring\n", ring
->name
);
452 I915_WRITE_CTL(ring
, 0);
453 I915_WRITE_HEAD(ring
, 0);
454 ring
->write_tail(ring
, 0);
456 if (!IS_GEN2(ring
->dev
)) {
457 (void)I915_READ_CTL(ring
);
458 I915_WRITE_MODE(ring
, _MASKED_BIT_DISABLE(STOP_RING
));
461 return (I915_READ_HEAD(ring
) & HEAD_ADDR
) == 0;
464 static int init_ring_common(struct intel_ring_buffer
*ring
)
466 struct drm_device
*dev
= ring
->dev
;
467 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
468 struct drm_i915_gem_object
*obj
= ring
->obj
;
471 gen6_gt_force_wake_get(dev_priv
, FORCEWAKE_ALL
);
473 if (!stop_ring(ring
)) {
474 /* G45 ring initialization often fails to reset head to zero */
475 DRM_DEBUG_KMS("%s head not reset to zero "
476 "ctl %08x head %08x tail %08x start %08x\n",
479 I915_READ_HEAD(ring
),
480 I915_READ_TAIL(ring
),
481 I915_READ_START(ring
));
483 if (!stop_ring(ring
)) {
484 DRM_ERROR("failed to set %s head to zero "
485 "ctl %08x head %08x tail %08x start %08x\n",
488 I915_READ_HEAD(ring
),
489 I915_READ_TAIL(ring
),
490 I915_READ_START(ring
));
496 if (I915_NEED_GFX_HWS(dev
))
497 intel_ring_setup_status_page(ring
);
499 ring_setup_phys_status_page(ring
);
501 /* Initialize the ring. This must happen _after_ we've cleared the ring
502 * registers with the above sequence (the readback of the HEAD registers
503 * also enforces ordering), otherwise the hw might lose the new ring
504 * register values. */
505 I915_WRITE_START(ring
, i915_gem_obj_ggtt_offset(obj
));
507 ((ring
->size
- PAGE_SIZE
) & RING_NR_PAGES
)
510 /* If the head is still not zero, the ring is dead */
511 if (wait_for((I915_READ_CTL(ring
) & RING_VALID
) != 0 &&
512 I915_READ_START(ring
) == i915_gem_obj_ggtt_offset(obj
) &&
513 (I915_READ_HEAD(ring
) & HEAD_ADDR
) == 0, 50)) {
514 DRM_ERROR("%s initialization failed "
515 "ctl %08x head %08x tail %08x start %08x\n",
518 I915_READ_HEAD(ring
),
519 I915_READ_TAIL(ring
),
520 I915_READ_START(ring
));
525 if (!drm_core_check_feature(ring
->dev
, DRIVER_MODESET
))
526 i915_kernel_lost_context(ring
->dev
);
528 ring
->head
= I915_READ_HEAD(ring
);
529 ring
->tail
= I915_READ_TAIL(ring
) & TAIL_ADDR
;
530 ring
->space
= ring_space(ring
);
531 ring
->last_retired_head
= -1;
534 memset(&ring
->hangcheck
, 0, sizeof(ring
->hangcheck
));
537 gen6_gt_force_wake_put(dev_priv
, FORCEWAKE_ALL
);
543 init_pipe_control(struct intel_ring_buffer
*ring
)
547 if (ring
->scratch
.obj
)
550 ring
->scratch
.obj
= i915_gem_alloc_object(ring
->dev
, 4096);
551 if (ring
->scratch
.obj
== NULL
) {
552 DRM_ERROR("Failed to allocate seqno page\n");
557 ret
= i915_gem_object_set_cache_level(ring
->scratch
.obj
, I915_CACHE_LLC
);
561 ret
= i915_gem_obj_ggtt_pin(ring
->scratch
.obj
, 4096, 0);
565 ring
->scratch
.gtt_offset
= i915_gem_obj_ggtt_offset(ring
->scratch
.obj
);
566 ring
->scratch
.cpu_page
= kmap(sg_page(ring
->scratch
.obj
->pages
->sgl
));
567 if (ring
->scratch
.cpu_page
== NULL
) {
572 DRM_DEBUG_DRIVER("%s pipe control offset: 0x%08x\n",
573 ring
->name
, ring
->scratch
.gtt_offset
);
577 i915_gem_object_ggtt_unpin(ring
->scratch
.obj
);
579 drm_gem_object_unreference(&ring
->scratch
.obj
->base
);
584 static int init_render_ring(struct intel_ring_buffer
*ring
)
586 struct drm_device
*dev
= ring
->dev
;
587 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
588 int ret
= init_ring_common(ring
);
590 /* WaTimedSingleVertexDispatch:cl,bw,ctg,elk,ilk,snb */
591 if (INTEL_INFO(dev
)->gen
>= 4 && INTEL_INFO(dev
)->gen
< 7)
592 I915_WRITE(MI_MODE
, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH
));
594 /* We need to disable the AsyncFlip performance optimisations in order
595 * to use MI_WAIT_FOR_EVENT within the CS. It should already be
596 * programmed to '1' on all products.
598 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw
600 if (INTEL_INFO(dev
)->gen
>= 6)
601 I915_WRITE(MI_MODE
, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE
));
603 /* Required for the hardware to program scanline values for waiting */
604 if (INTEL_INFO(dev
)->gen
== 6)
606 _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_ALWAYS
));
609 I915_WRITE(GFX_MODE_GEN7
,
610 _MASKED_BIT_DISABLE(GFX_TLB_INVALIDATE_ALWAYS
) |
611 _MASKED_BIT_ENABLE(GFX_REPLAY_MODE
));
613 if (INTEL_INFO(dev
)->gen
>= 5) {
614 ret
= init_pipe_control(ring
);
620 /* From the Sandybridge PRM, volume 1 part 3, page 24:
621 * "If this bit is set, STCunit will have LRA as replacement
622 * policy. [...] This bit must be reset. LRA replacement
623 * policy is not supported."
625 I915_WRITE(CACHE_MODE_0
,
626 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB
));
628 /* This is not explicitly set for GEN6, so read the register.
629 * see intel_ring_mi_set_context() for why we care.
630 * TODO: consider explicitly setting the bit for GEN5
632 ring
->itlb_before_ctx_switch
=
633 !!(I915_READ(GFX_MODE
) & GFX_TLB_INVALIDATE_ALWAYS
);
636 if (INTEL_INFO(dev
)->gen
>= 6)
637 I915_WRITE(INSTPM
, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING
));
640 I915_WRITE_IMR(ring
, ~GT_PARITY_ERROR(dev
));
645 static void render_ring_cleanup(struct intel_ring_buffer
*ring
)
647 struct drm_device
*dev
= ring
->dev
;
649 if (ring
->scratch
.obj
== NULL
)
652 if (INTEL_INFO(dev
)->gen
>= 5) {
653 kunmap(sg_page(ring
->scratch
.obj
->pages
->sgl
));
654 i915_gem_object_ggtt_unpin(ring
->scratch
.obj
);
657 drm_gem_object_unreference(&ring
->scratch
.obj
->base
);
658 ring
->scratch
.obj
= NULL
;
662 update_mboxes(struct intel_ring_buffer
*ring
,
665 /* NB: In order to be able to do semaphore MBOX updates for varying number
666 * of rings, it's easiest if we round up each individual update to a
667 * multiple of 2 (since ring updates must always be a multiple of 2)
668 * even though the actual update only requires 3 dwords.
670 #define MBOX_UPDATE_DWORDS 4
671 intel_ring_emit(ring
, MI_LOAD_REGISTER_IMM(1));
672 intel_ring_emit(ring
, mmio_offset
);
673 intel_ring_emit(ring
, ring
->outstanding_lazy_seqno
);
674 intel_ring_emit(ring
, MI_NOOP
);
678 * gen6_add_request - Update the semaphore mailbox registers
680 * @ring - ring that is adding a request
681 * @seqno - return seqno stuck into the ring
683 * Update the mailbox registers in the *other* rings with the current seqno.
684 * This acts like a signal in the canonical semaphore.
687 gen6_add_request(struct intel_ring_buffer
*ring
)
689 struct drm_device
*dev
= ring
->dev
;
690 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
691 struct intel_ring_buffer
*useless
;
692 int i
, ret
, num_dwords
= 4;
694 if (i915_semaphore_is_enabled(dev
))
695 num_dwords
+= ((I915_NUM_RINGS
-1) * MBOX_UPDATE_DWORDS
);
696 #undef MBOX_UPDATE_DWORDS
698 ret
= intel_ring_begin(ring
, num_dwords
);
702 if (i915_semaphore_is_enabled(dev
)) {
703 for_each_ring(useless
, dev_priv
, i
) {
704 u32 mbox_reg
= ring
->signal_mbox
[i
];
705 if (mbox_reg
!= GEN6_NOSYNC
)
706 update_mboxes(ring
, mbox_reg
);
710 intel_ring_emit(ring
, MI_STORE_DWORD_INDEX
);
711 intel_ring_emit(ring
, I915_GEM_HWS_INDEX
<< MI_STORE_DWORD_INDEX_SHIFT
);
712 intel_ring_emit(ring
, ring
->outstanding_lazy_seqno
);
713 intel_ring_emit(ring
, MI_USER_INTERRUPT
);
714 __intel_ring_advance(ring
);
719 static inline bool i915_gem_has_seqno_wrapped(struct drm_device
*dev
,
722 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
723 return dev_priv
->last_seqno
< seqno
;
727 * intel_ring_sync - sync the waiter to the signaller on seqno
729 * @waiter - ring that is waiting
730 * @signaller - ring which has, or will signal
731 * @seqno - seqno which the waiter will block on
734 gen6_ring_sync(struct intel_ring_buffer
*waiter
,
735 struct intel_ring_buffer
*signaller
,
739 u32 dw1
= MI_SEMAPHORE_MBOX
|
740 MI_SEMAPHORE_COMPARE
|
741 MI_SEMAPHORE_REGISTER
;
743 /* Throughout all of the GEM code, seqno passed implies our current
744 * seqno is >= the last seqno executed. However for hardware the
745 * comparison is strictly greater than.
749 WARN_ON(signaller
->semaphore_register
[waiter
->id
] ==
750 MI_SEMAPHORE_SYNC_INVALID
);
752 ret
= intel_ring_begin(waiter
, 4);
756 /* If seqno wrap happened, omit the wait with no-ops */
757 if (likely(!i915_gem_has_seqno_wrapped(waiter
->dev
, seqno
))) {
758 intel_ring_emit(waiter
,
760 signaller
->semaphore_register
[waiter
->id
]);
761 intel_ring_emit(waiter
, seqno
);
762 intel_ring_emit(waiter
, 0);
763 intel_ring_emit(waiter
, MI_NOOP
);
765 intel_ring_emit(waiter
, MI_NOOP
);
766 intel_ring_emit(waiter
, MI_NOOP
);
767 intel_ring_emit(waiter
, MI_NOOP
);
768 intel_ring_emit(waiter
, MI_NOOP
);
770 intel_ring_advance(waiter
);
775 #define PIPE_CONTROL_FLUSH(ring__, addr__) \
777 intel_ring_emit(ring__, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE | \
778 PIPE_CONTROL_DEPTH_STALL); \
779 intel_ring_emit(ring__, (addr__) | PIPE_CONTROL_GLOBAL_GTT); \
780 intel_ring_emit(ring__, 0); \
781 intel_ring_emit(ring__, 0); \
785 pc_render_add_request(struct intel_ring_buffer
*ring
)
787 u32 scratch_addr
= ring
->scratch
.gtt_offset
+ 128;
790 /* For Ironlake, MI_USER_INTERRUPT was deprecated and apparently
791 * incoherent with writes to memory, i.e. completely fubar,
792 * so we need to use PIPE_NOTIFY instead.
794 * However, we also need to workaround the qword write
795 * incoherence by flushing the 6 PIPE_NOTIFY buffers out to
796 * memory before requesting an interrupt.
798 ret
= intel_ring_begin(ring
, 32);
802 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE
|
803 PIPE_CONTROL_WRITE_FLUSH
|
804 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
);
805 intel_ring_emit(ring
, ring
->scratch
.gtt_offset
| PIPE_CONTROL_GLOBAL_GTT
);
806 intel_ring_emit(ring
, ring
->outstanding_lazy_seqno
);
807 intel_ring_emit(ring
, 0);
808 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
809 scratch_addr
+= 128; /* write to separate cachelines */
810 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
812 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
814 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
816 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
818 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
820 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE
|
821 PIPE_CONTROL_WRITE_FLUSH
|
822 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
823 PIPE_CONTROL_NOTIFY
);
824 intel_ring_emit(ring
, ring
->scratch
.gtt_offset
| PIPE_CONTROL_GLOBAL_GTT
);
825 intel_ring_emit(ring
, ring
->outstanding_lazy_seqno
);
826 intel_ring_emit(ring
, 0);
827 __intel_ring_advance(ring
);
833 gen6_ring_get_seqno(struct intel_ring_buffer
*ring
, bool lazy_coherency
)
835 /* Workaround to force correct ordering between irq and seqno writes on
836 * ivb (and maybe also on snb) by reading from a CS register (like
837 * ACTHD) before reading the status page. */
838 if (!lazy_coherency
) {
839 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
840 POSTING_READ(RING_ACTHD(ring
->mmio_base
));
843 return intel_read_status_page(ring
, I915_GEM_HWS_INDEX
);
847 ring_get_seqno(struct intel_ring_buffer
*ring
, bool lazy_coherency
)
849 return intel_read_status_page(ring
, I915_GEM_HWS_INDEX
);
853 ring_set_seqno(struct intel_ring_buffer
*ring
, u32 seqno
)
855 intel_write_status_page(ring
, I915_GEM_HWS_INDEX
, seqno
);
859 pc_render_get_seqno(struct intel_ring_buffer
*ring
, bool lazy_coherency
)
861 return ring
->scratch
.cpu_page
[0];
865 pc_render_set_seqno(struct intel_ring_buffer
*ring
, u32 seqno
)
867 ring
->scratch
.cpu_page
[0] = seqno
;
871 gen5_ring_get_irq(struct intel_ring_buffer
*ring
)
873 struct drm_device
*dev
= ring
->dev
;
874 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
877 if (!dev
->irq_enabled
)
880 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
881 if (ring
->irq_refcount
++ == 0)
882 ilk_enable_gt_irq(dev_priv
, ring
->irq_enable_mask
);
883 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
889 gen5_ring_put_irq(struct intel_ring_buffer
*ring
)
891 struct drm_device
*dev
= ring
->dev
;
892 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
895 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
896 if (--ring
->irq_refcount
== 0)
897 ilk_disable_gt_irq(dev_priv
, ring
->irq_enable_mask
);
898 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
902 i9xx_ring_get_irq(struct intel_ring_buffer
*ring
)
904 struct drm_device
*dev
= ring
->dev
;
905 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
908 if (!dev
->irq_enabled
)
911 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
912 if (ring
->irq_refcount
++ == 0) {
913 dev_priv
->irq_mask
&= ~ring
->irq_enable_mask
;
914 I915_WRITE(IMR
, dev_priv
->irq_mask
);
917 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
923 i9xx_ring_put_irq(struct intel_ring_buffer
*ring
)
925 struct drm_device
*dev
= ring
->dev
;
926 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
929 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
930 if (--ring
->irq_refcount
== 0) {
931 dev_priv
->irq_mask
|= ring
->irq_enable_mask
;
932 I915_WRITE(IMR
, dev_priv
->irq_mask
);
935 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
939 i8xx_ring_get_irq(struct intel_ring_buffer
*ring
)
941 struct drm_device
*dev
= ring
->dev
;
942 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
945 if (!dev
->irq_enabled
)
948 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
949 if (ring
->irq_refcount
++ == 0) {
950 dev_priv
->irq_mask
&= ~ring
->irq_enable_mask
;
951 I915_WRITE16(IMR
, dev_priv
->irq_mask
);
954 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
960 i8xx_ring_put_irq(struct intel_ring_buffer
*ring
)
962 struct drm_device
*dev
= ring
->dev
;
963 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
966 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
967 if (--ring
->irq_refcount
== 0) {
968 dev_priv
->irq_mask
|= ring
->irq_enable_mask
;
969 I915_WRITE16(IMR
, dev_priv
->irq_mask
);
972 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
975 void intel_ring_setup_status_page(struct intel_ring_buffer
*ring
)
977 struct drm_device
*dev
= ring
->dev
;
978 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
981 /* The ring status page addresses are no longer next to the rest of
982 * the ring registers as of gen7.
987 mmio
= RENDER_HWS_PGA_GEN7
;
990 mmio
= BLT_HWS_PGA_GEN7
;
993 mmio
= BSD_HWS_PGA_GEN7
;
996 mmio
= VEBOX_HWS_PGA_GEN7
;
999 } else if (IS_GEN6(ring
->dev
)) {
1000 mmio
= RING_HWS_PGA_GEN6(ring
->mmio_base
);
1002 /* XXX: gen8 returns to sanity */
1003 mmio
= RING_HWS_PGA(ring
->mmio_base
);
1006 I915_WRITE(mmio
, (u32
)ring
->status_page
.gfx_addr
);
1010 * Flush the TLB for this page
1012 * FIXME: These two bits have disappeared on gen8, so a question
1013 * arises: do we still need this and if so how should we go about
1014 * invalidating the TLB?
1016 if (INTEL_INFO(dev
)->gen
>= 6 && INTEL_INFO(dev
)->gen
< 8) {
1017 u32 reg
= RING_INSTPM(ring
->mmio_base
);
1019 /* ring should be idle before issuing a sync flush*/
1020 WARN_ON((I915_READ_MODE(ring
) & MODE_IDLE
) == 0);
1023 _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE
|
1024 INSTPM_SYNC_FLUSH
));
1025 if (wait_for((I915_READ(reg
) & INSTPM_SYNC_FLUSH
) == 0,
1027 DRM_ERROR("%s: wait for SyncFlush to complete for TLB invalidation timed out\n",
1033 bsd_ring_flush(struct intel_ring_buffer
*ring
,
1034 u32 invalidate_domains
,
1039 ret
= intel_ring_begin(ring
, 2);
1043 intel_ring_emit(ring
, MI_FLUSH
);
1044 intel_ring_emit(ring
, MI_NOOP
);
1045 intel_ring_advance(ring
);
1050 i9xx_add_request(struct intel_ring_buffer
*ring
)
1054 ret
= intel_ring_begin(ring
, 4);
1058 intel_ring_emit(ring
, MI_STORE_DWORD_INDEX
);
1059 intel_ring_emit(ring
, I915_GEM_HWS_INDEX
<< MI_STORE_DWORD_INDEX_SHIFT
);
1060 intel_ring_emit(ring
, ring
->outstanding_lazy_seqno
);
1061 intel_ring_emit(ring
, MI_USER_INTERRUPT
);
1062 __intel_ring_advance(ring
);
1068 gen6_ring_get_irq(struct intel_ring_buffer
*ring
)
1070 struct drm_device
*dev
= ring
->dev
;
1071 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1072 unsigned long flags
;
1074 if (!dev
->irq_enabled
)
1077 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1078 if (ring
->irq_refcount
++ == 0) {
1079 if (HAS_L3_DPF(dev
) && ring
->id
== RCS
)
1080 I915_WRITE_IMR(ring
,
1081 ~(ring
->irq_enable_mask
|
1082 GT_PARITY_ERROR(dev
)));
1084 I915_WRITE_IMR(ring
, ~ring
->irq_enable_mask
);
1085 ilk_enable_gt_irq(dev_priv
, ring
->irq_enable_mask
);
1087 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1093 gen6_ring_put_irq(struct intel_ring_buffer
*ring
)
1095 struct drm_device
*dev
= ring
->dev
;
1096 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1097 unsigned long flags
;
1099 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1100 if (--ring
->irq_refcount
== 0) {
1101 if (HAS_L3_DPF(dev
) && ring
->id
== RCS
)
1102 I915_WRITE_IMR(ring
, ~GT_PARITY_ERROR(dev
));
1104 I915_WRITE_IMR(ring
, ~0);
1105 ilk_disable_gt_irq(dev_priv
, ring
->irq_enable_mask
);
1107 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1111 hsw_vebox_get_irq(struct intel_ring_buffer
*ring
)
1113 struct drm_device
*dev
= ring
->dev
;
1114 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1115 unsigned long flags
;
1117 if (!dev
->irq_enabled
)
1120 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1121 if (ring
->irq_refcount
++ == 0) {
1122 I915_WRITE_IMR(ring
, ~ring
->irq_enable_mask
);
1123 snb_enable_pm_irq(dev_priv
, ring
->irq_enable_mask
);
1125 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1131 hsw_vebox_put_irq(struct intel_ring_buffer
*ring
)
1133 struct drm_device
*dev
= ring
->dev
;
1134 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1135 unsigned long flags
;
1137 if (!dev
->irq_enabled
)
1140 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1141 if (--ring
->irq_refcount
== 0) {
1142 I915_WRITE_IMR(ring
, ~0);
1143 snb_disable_pm_irq(dev_priv
, ring
->irq_enable_mask
);
1145 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1149 gen8_ring_get_irq(struct intel_ring_buffer
*ring
)
1151 struct drm_device
*dev
= ring
->dev
;
1152 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1153 unsigned long flags
;
1155 if (!dev
->irq_enabled
)
1158 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1159 if (ring
->irq_refcount
++ == 0) {
1160 if (HAS_L3_DPF(dev
) && ring
->id
== RCS
) {
1161 I915_WRITE_IMR(ring
,
1162 ~(ring
->irq_enable_mask
|
1163 GT_RENDER_L3_PARITY_ERROR_INTERRUPT
));
1165 I915_WRITE_IMR(ring
, ~ring
->irq_enable_mask
);
1167 POSTING_READ(RING_IMR(ring
->mmio_base
));
1169 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1175 gen8_ring_put_irq(struct intel_ring_buffer
*ring
)
1177 struct drm_device
*dev
= ring
->dev
;
1178 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1179 unsigned long flags
;
1181 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
1182 if (--ring
->irq_refcount
== 0) {
1183 if (HAS_L3_DPF(dev
) && ring
->id
== RCS
) {
1184 I915_WRITE_IMR(ring
,
1185 ~GT_RENDER_L3_PARITY_ERROR_INTERRUPT
);
1187 I915_WRITE_IMR(ring
, ~0);
1189 POSTING_READ(RING_IMR(ring
->mmio_base
));
1191 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
1195 i965_dispatch_execbuffer(struct intel_ring_buffer
*ring
,
1196 u32 offset
, u32 length
,
1201 ret
= intel_ring_begin(ring
, 2);
1205 intel_ring_emit(ring
,
1206 MI_BATCH_BUFFER_START
|
1208 (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE_I965
));
1209 intel_ring_emit(ring
, offset
);
1210 intel_ring_advance(ring
);
1215 /* Just userspace ABI convention to limit the wa batch bo to a resonable size */
1216 #define I830_BATCH_LIMIT (256*1024)
1218 i830_dispatch_execbuffer(struct intel_ring_buffer
*ring
,
1219 u32 offset
, u32 len
,
1224 if (flags
& I915_DISPATCH_PINNED
) {
1225 ret
= intel_ring_begin(ring
, 4);
1229 intel_ring_emit(ring
, MI_BATCH_BUFFER
);
1230 intel_ring_emit(ring
, offset
| (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE
));
1231 intel_ring_emit(ring
, offset
+ len
- 8);
1232 intel_ring_emit(ring
, MI_NOOP
);
1233 intel_ring_advance(ring
);
1235 u32 cs_offset
= ring
->scratch
.gtt_offset
;
1237 if (len
> I830_BATCH_LIMIT
)
1240 ret
= intel_ring_begin(ring
, 9+3);
1243 /* Blit the batch (which has now all relocs applied) to the stable batch
1244 * scratch bo area (so that the CS never stumbles over its tlb
1245 * invalidation bug) ... */
1246 intel_ring_emit(ring
, XY_SRC_COPY_BLT_CMD
|
1247 XY_SRC_COPY_BLT_WRITE_ALPHA
|
1248 XY_SRC_COPY_BLT_WRITE_RGB
);
1249 intel_ring_emit(ring
, BLT_DEPTH_32
| BLT_ROP_GXCOPY
| 4096);
1250 intel_ring_emit(ring
, 0);
1251 intel_ring_emit(ring
, (DIV_ROUND_UP(len
, 4096) << 16) | 1024);
1252 intel_ring_emit(ring
, cs_offset
);
1253 intel_ring_emit(ring
, 0);
1254 intel_ring_emit(ring
, 4096);
1255 intel_ring_emit(ring
, offset
);
1256 intel_ring_emit(ring
, MI_FLUSH
);
1258 /* ... and execute it. */
1259 intel_ring_emit(ring
, MI_BATCH_BUFFER
);
1260 intel_ring_emit(ring
, cs_offset
| (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE
));
1261 intel_ring_emit(ring
, cs_offset
+ len
- 8);
1262 intel_ring_advance(ring
);
1269 i915_dispatch_execbuffer(struct intel_ring_buffer
*ring
,
1270 u32 offset
, u32 len
,
1275 ret
= intel_ring_begin(ring
, 2);
1279 intel_ring_emit(ring
, MI_BATCH_BUFFER_START
| MI_BATCH_GTT
);
1280 intel_ring_emit(ring
, offset
| (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE
));
1281 intel_ring_advance(ring
);
1286 static void cleanup_status_page(struct intel_ring_buffer
*ring
)
1288 struct drm_i915_gem_object
*obj
;
1290 obj
= ring
->status_page
.obj
;
1294 kunmap(sg_page(obj
->pages
->sgl
));
1295 i915_gem_object_ggtt_unpin(obj
);
1296 drm_gem_object_unreference(&obj
->base
);
1297 ring
->status_page
.obj
= NULL
;
1300 static int init_status_page(struct intel_ring_buffer
*ring
)
1302 struct drm_device
*dev
= ring
->dev
;
1303 struct drm_i915_gem_object
*obj
;
1306 obj
= i915_gem_alloc_object(dev
, 4096);
1308 DRM_ERROR("Failed to allocate status page\n");
1313 ret
= i915_gem_object_set_cache_level(obj
, I915_CACHE_LLC
);
1317 ret
= i915_gem_obj_ggtt_pin(obj
, 4096, 0);
1321 ring
->status_page
.gfx_addr
= i915_gem_obj_ggtt_offset(obj
);
1322 ring
->status_page
.page_addr
= kmap(sg_page(obj
->pages
->sgl
));
1323 if (ring
->status_page
.page_addr
== NULL
) {
1327 ring
->status_page
.obj
= obj
;
1328 memset(ring
->status_page
.page_addr
, 0, PAGE_SIZE
);
1330 DRM_DEBUG_DRIVER("%s hws offset: 0x%08x\n",
1331 ring
->name
, ring
->status_page
.gfx_addr
);
1336 i915_gem_object_ggtt_unpin(obj
);
1338 drm_gem_object_unreference(&obj
->base
);
1343 static int init_phys_status_page(struct intel_ring_buffer
*ring
)
1345 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
1347 if (!dev_priv
->status_page_dmah
) {
1348 dev_priv
->status_page_dmah
=
1349 drm_pci_alloc(ring
->dev
, PAGE_SIZE
, PAGE_SIZE
);
1350 if (!dev_priv
->status_page_dmah
)
1354 ring
->status_page
.page_addr
= dev_priv
->status_page_dmah
->vaddr
;
1355 memset(ring
->status_page
.page_addr
, 0, PAGE_SIZE
);
1360 static int intel_init_ring_buffer(struct drm_device
*dev
,
1361 struct intel_ring_buffer
*ring
)
1363 struct drm_i915_gem_object
*obj
;
1364 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1368 INIT_LIST_HEAD(&ring
->active_list
);
1369 INIT_LIST_HEAD(&ring
->request_list
);
1370 ring
->size
= 32 * PAGE_SIZE
;
1371 memset(ring
->sync_seqno
, 0, sizeof(ring
->sync_seqno
));
1373 init_waitqueue_head(&ring
->irq_queue
);
1375 if (I915_NEED_GFX_HWS(dev
)) {
1376 ret
= init_status_page(ring
);
1380 BUG_ON(ring
->id
!= RCS
);
1381 ret
= init_phys_status_page(ring
);
1388 obj
= i915_gem_object_create_stolen(dev
, ring
->size
);
1390 obj
= i915_gem_alloc_object(dev
, ring
->size
);
1392 DRM_ERROR("Failed to allocate ringbuffer\n");
1399 ret
= i915_gem_obj_ggtt_pin(obj
, PAGE_SIZE
, PIN_MAPPABLE
);
1403 ret
= i915_gem_object_set_to_gtt_domain(obj
, true);
1407 ring
->virtual_start
=
1408 ioremap_wc(dev_priv
->gtt
.mappable_base
+ i915_gem_obj_ggtt_offset(obj
),
1410 if (ring
->virtual_start
== NULL
) {
1411 DRM_ERROR("Failed to map ringbuffer.\n");
1416 ret
= ring
->init(ring
);
1420 /* Workaround an erratum on the i830 which causes a hang if
1421 * the TAIL pointer points to within the last 2 cachelines
1424 ring
->effective_size
= ring
->size
;
1425 if (IS_I830(ring
->dev
) || IS_845G(ring
->dev
))
1426 ring
->effective_size
-= 128;
1428 i915_cmd_parser_init_ring(ring
);
1433 iounmap(ring
->virtual_start
);
1435 i915_gem_object_ggtt_unpin(obj
);
1437 drm_gem_object_unreference(&obj
->base
);
1440 cleanup_status_page(ring
);
1444 void intel_cleanup_ring_buffer(struct intel_ring_buffer
*ring
)
1446 struct drm_i915_private
*dev_priv
;
1449 if (ring
->obj
== NULL
)
1452 /* Disable the ring buffer. The ring must be idle at this point */
1453 dev_priv
= ring
->dev
->dev_private
;
1454 ret
= intel_ring_idle(ring
);
1455 if (ret
&& !i915_reset_in_progress(&dev_priv
->gpu_error
))
1456 DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",
1459 I915_WRITE_CTL(ring
, 0);
1461 iounmap(ring
->virtual_start
);
1463 i915_gem_object_ggtt_unpin(ring
->obj
);
1464 drm_gem_object_unreference(&ring
->obj
->base
);
1466 ring
->preallocated_lazy_request
= NULL
;
1467 ring
->outstanding_lazy_seqno
= 0;
1470 ring
->cleanup(ring
);
1472 cleanup_status_page(ring
);
1475 static int intel_ring_wait_request(struct intel_ring_buffer
*ring
, int n
)
1477 struct drm_i915_gem_request
*request
;
1478 u32 seqno
= 0, tail
;
1481 if (ring
->last_retired_head
!= -1) {
1482 ring
->head
= ring
->last_retired_head
;
1483 ring
->last_retired_head
= -1;
1485 ring
->space
= ring_space(ring
);
1486 if (ring
->space
>= n
)
1490 list_for_each_entry(request
, &ring
->request_list
, list
) {
1493 if (request
->tail
== -1)
1496 space
= request
->tail
- (ring
->tail
+ I915_RING_FREE_SPACE
);
1498 space
+= ring
->size
;
1500 seqno
= request
->seqno
;
1501 tail
= request
->tail
;
1505 /* Consume this request in case we need more space than
1506 * is available and so need to prevent a race between
1507 * updating last_retired_head and direct reads of
1508 * I915_RING_HEAD. It also provides a nice sanity check.
1516 ret
= i915_wait_seqno(ring
, seqno
);
1521 ring
->space
= ring_space(ring
);
1522 if (WARN_ON(ring
->space
< n
))
1528 static int ring_wait_for_space(struct intel_ring_buffer
*ring
, int n
)
1530 struct drm_device
*dev
= ring
->dev
;
1531 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1535 ret
= intel_ring_wait_request(ring
, n
);
1539 /* force the tail write in case we have been skipping them */
1540 __intel_ring_advance(ring
);
1542 trace_i915_ring_wait_begin(ring
);
1543 /* With GEM the hangcheck timer should kick us out of the loop,
1544 * leaving it early runs the risk of corrupting GEM state (due
1545 * to running on almost untested codepaths). But on resume
1546 * timers don't work yet, so prevent a complete hang in that
1547 * case by choosing an insanely large timeout. */
1548 end
= jiffies
+ 60 * HZ
;
1551 ring
->head
= I915_READ_HEAD(ring
);
1552 ring
->space
= ring_space(ring
);
1553 if (ring
->space
>= n
) {
1554 trace_i915_ring_wait_end(ring
);
1558 if (!drm_core_check_feature(dev
, DRIVER_MODESET
) &&
1559 dev
->primary
->master
) {
1560 struct drm_i915_master_private
*master_priv
= dev
->primary
->master
->driver_priv
;
1561 if (master_priv
->sarea_priv
)
1562 master_priv
->sarea_priv
->perf_boxes
|= I915_BOX_WAIT
;
1567 ret
= i915_gem_check_wedge(&dev_priv
->gpu_error
,
1568 dev_priv
->mm
.interruptible
);
1571 } while (!time_after(jiffies
, end
));
1572 trace_i915_ring_wait_end(ring
);
1576 static int intel_wrap_ring_buffer(struct intel_ring_buffer
*ring
)
1578 uint32_t __iomem
*virt
;
1579 int rem
= ring
->size
- ring
->tail
;
1581 if (ring
->space
< rem
) {
1582 int ret
= ring_wait_for_space(ring
, rem
);
1587 virt
= ring
->virtual_start
+ ring
->tail
;
1590 iowrite32(MI_NOOP
, virt
++);
1593 ring
->space
= ring_space(ring
);
1598 int intel_ring_idle(struct intel_ring_buffer
*ring
)
1603 /* We need to add any requests required to flush the objects and ring */
1604 if (ring
->outstanding_lazy_seqno
) {
1605 ret
= i915_add_request(ring
, NULL
);
1610 /* Wait upon the last request to be completed */
1611 if (list_empty(&ring
->request_list
))
1614 seqno
= list_entry(ring
->request_list
.prev
,
1615 struct drm_i915_gem_request
,
1618 return i915_wait_seqno(ring
, seqno
);
1622 intel_ring_alloc_seqno(struct intel_ring_buffer
*ring
)
1624 if (ring
->outstanding_lazy_seqno
)
1627 if (ring
->preallocated_lazy_request
== NULL
) {
1628 struct drm_i915_gem_request
*request
;
1630 request
= kmalloc(sizeof(*request
), GFP_KERNEL
);
1631 if (request
== NULL
)
1634 ring
->preallocated_lazy_request
= request
;
1637 return i915_gem_get_seqno(ring
->dev
, &ring
->outstanding_lazy_seqno
);
1640 static int __intel_ring_prepare(struct intel_ring_buffer
*ring
,
1645 if (unlikely(ring
->tail
+ bytes
> ring
->effective_size
)) {
1646 ret
= intel_wrap_ring_buffer(ring
);
1651 if (unlikely(ring
->space
< bytes
)) {
1652 ret
= ring_wait_for_space(ring
, bytes
);
1660 int intel_ring_begin(struct intel_ring_buffer
*ring
,
1663 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
1666 ret
= i915_gem_check_wedge(&dev_priv
->gpu_error
,
1667 dev_priv
->mm
.interruptible
);
1671 ret
= __intel_ring_prepare(ring
, num_dwords
* sizeof(uint32_t));
1675 /* Preallocate the olr before touching the ring */
1676 ret
= intel_ring_alloc_seqno(ring
);
1680 ring
->space
-= num_dwords
* sizeof(uint32_t);
1684 /* Align the ring tail to a cacheline boundary */
1685 int intel_ring_cacheline_align(struct intel_ring_buffer
*ring
)
1687 int num_dwords
= (64 - (ring
->tail
& 63)) / sizeof(uint32_t);
1690 if (num_dwords
== 0)
1693 ret
= intel_ring_begin(ring
, num_dwords
);
1697 while (num_dwords
--)
1698 intel_ring_emit(ring
, MI_NOOP
);
1700 intel_ring_advance(ring
);
1705 void intel_ring_init_seqno(struct intel_ring_buffer
*ring
, u32 seqno
)
1707 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
1709 BUG_ON(ring
->outstanding_lazy_seqno
);
1711 if (INTEL_INFO(ring
->dev
)->gen
>= 6) {
1712 I915_WRITE(RING_SYNC_0(ring
->mmio_base
), 0);
1713 I915_WRITE(RING_SYNC_1(ring
->mmio_base
), 0);
1714 if (HAS_VEBOX(ring
->dev
))
1715 I915_WRITE(RING_SYNC_2(ring
->mmio_base
), 0);
1718 ring
->set_seqno(ring
, seqno
);
1719 ring
->hangcheck
.seqno
= seqno
;
1722 static void gen6_bsd_ring_write_tail(struct intel_ring_buffer
*ring
,
1725 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
1727 /* Every tail move must follow the sequence below */
1729 /* Disable notification that the ring is IDLE. The GT
1730 * will then assume that it is busy and bring it out of rc6.
1732 I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL
,
1733 _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE
));
1735 /* Clear the context id. Here be magic! */
1736 I915_WRITE64(GEN6_BSD_RNCID
, 0x0);
1738 /* Wait for the ring not to be idle, i.e. for it to wake up. */
1739 if (wait_for((I915_READ(GEN6_BSD_SLEEP_PSMI_CONTROL
) &
1740 GEN6_BSD_SLEEP_INDICATOR
) == 0,
1742 DRM_ERROR("timed out waiting for the BSD ring to wake up\n");
1744 /* Now that the ring is fully powered up, update the tail */
1745 I915_WRITE_TAIL(ring
, value
);
1746 POSTING_READ(RING_TAIL(ring
->mmio_base
));
1748 /* Let the ring send IDLE messages to the GT again,
1749 * and so let it sleep to conserve power when idle.
1751 I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL
,
1752 _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE
));
1755 static int gen6_bsd_ring_flush(struct intel_ring_buffer
*ring
,
1756 u32 invalidate
, u32 flush
)
1761 ret
= intel_ring_begin(ring
, 4);
1766 if (INTEL_INFO(ring
->dev
)->gen
>= 8)
1769 * Bspec vol 1c.5 - video engine command streamer:
1770 * "If ENABLED, all TLBs will be invalidated once the flush
1771 * operation is complete. This bit is only valid when the
1772 * Post-Sync Operation field is a value of 1h or 3h."
1774 if (invalidate
& I915_GEM_GPU_DOMAINS
)
1775 cmd
|= MI_INVALIDATE_TLB
| MI_INVALIDATE_BSD
|
1776 MI_FLUSH_DW_STORE_INDEX
| MI_FLUSH_DW_OP_STOREDW
;
1777 intel_ring_emit(ring
, cmd
);
1778 intel_ring_emit(ring
, I915_GEM_HWS_SCRATCH_ADDR
| MI_FLUSH_DW_USE_GTT
);
1779 if (INTEL_INFO(ring
->dev
)->gen
>= 8) {
1780 intel_ring_emit(ring
, 0); /* upper addr */
1781 intel_ring_emit(ring
, 0); /* value */
1783 intel_ring_emit(ring
, 0);
1784 intel_ring_emit(ring
, MI_NOOP
);
1786 intel_ring_advance(ring
);
1791 gen8_ring_dispatch_execbuffer(struct intel_ring_buffer
*ring
,
1792 u32 offset
, u32 len
,
1795 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
1796 bool ppgtt
= dev_priv
->mm
.aliasing_ppgtt
!= NULL
&&
1797 !(flags
& I915_DISPATCH_SECURE
);
1800 ret
= intel_ring_begin(ring
, 4);
1804 /* FIXME(BDW): Address space and security selectors. */
1805 intel_ring_emit(ring
, MI_BATCH_BUFFER_START_GEN8
| (ppgtt
<<8));
1806 intel_ring_emit(ring
, offset
);
1807 intel_ring_emit(ring
, 0);
1808 intel_ring_emit(ring
, MI_NOOP
);
1809 intel_ring_advance(ring
);
1815 hsw_ring_dispatch_execbuffer(struct intel_ring_buffer
*ring
,
1816 u32 offset
, u32 len
,
1821 ret
= intel_ring_begin(ring
, 2);
1825 intel_ring_emit(ring
,
1826 MI_BATCH_BUFFER_START
| MI_BATCH_PPGTT_HSW
|
1827 (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE_HSW
));
1828 /* bit0-7 is the length on GEN6+ */
1829 intel_ring_emit(ring
, offset
);
1830 intel_ring_advance(ring
);
1836 gen6_ring_dispatch_execbuffer(struct intel_ring_buffer
*ring
,
1837 u32 offset
, u32 len
,
1842 ret
= intel_ring_begin(ring
, 2);
1846 intel_ring_emit(ring
,
1847 MI_BATCH_BUFFER_START
|
1848 (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE_I965
));
1849 /* bit0-7 is the length on GEN6+ */
1850 intel_ring_emit(ring
, offset
);
1851 intel_ring_advance(ring
);
1856 /* Blitter support (SandyBridge+) */
1858 static int gen6_ring_flush(struct intel_ring_buffer
*ring
,
1859 u32 invalidate
, u32 flush
)
1861 struct drm_device
*dev
= ring
->dev
;
1865 ret
= intel_ring_begin(ring
, 4);
1870 if (INTEL_INFO(ring
->dev
)->gen
>= 8)
1873 * Bspec vol 1c.3 - blitter engine command streamer:
1874 * "If ENABLED, all TLBs will be invalidated once the flush
1875 * operation is complete. This bit is only valid when the
1876 * Post-Sync Operation field is a value of 1h or 3h."
1878 if (invalidate
& I915_GEM_DOMAIN_RENDER
)
1879 cmd
|= MI_INVALIDATE_TLB
| MI_FLUSH_DW_STORE_INDEX
|
1880 MI_FLUSH_DW_OP_STOREDW
;
1881 intel_ring_emit(ring
, cmd
);
1882 intel_ring_emit(ring
, I915_GEM_HWS_SCRATCH_ADDR
| MI_FLUSH_DW_USE_GTT
);
1883 if (INTEL_INFO(ring
->dev
)->gen
>= 8) {
1884 intel_ring_emit(ring
, 0); /* upper addr */
1885 intel_ring_emit(ring
, 0); /* value */
1887 intel_ring_emit(ring
, 0);
1888 intel_ring_emit(ring
, MI_NOOP
);
1890 intel_ring_advance(ring
);
1892 if (IS_GEN7(dev
) && !invalidate
&& flush
)
1893 return gen7_ring_fbc_flush(ring
, FBC_REND_CACHE_CLEAN
);
1898 int intel_init_render_ring_buffer(struct drm_device
*dev
)
1900 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1901 struct intel_ring_buffer
*ring
= &dev_priv
->ring
[RCS
];
1903 ring
->name
= "render ring";
1905 ring
->mmio_base
= RENDER_RING_BASE
;
1907 if (INTEL_INFO(dev
)->gen
>= 6) {
1908 ring
->add_request
= gen6_add_request
;
1909 ring
->flush
= gen7_render_ring_flush
;
1910 if (INTEL_INFO(dev
)->gen
== 6)
1911 ring
->flush
= gen6_render_ring_flush
;
1912 if (INTEL_INFO(dev
)->gen
>= 8) {
1913 ring
->flush
= gen8_render_ring_flush
;
1914 ring
->irq_get
= gen8_ring_get_irq
;
1915 ring
->irq_put
= gen8_ring_put_irq
;
1917 ring
->irq_get
= gen6_ring_get_irq
;
1918 ring
->irq_put
= gen6_ring_put_irq
;
1920 ring
->irq_enable_mask
= GT_RENDER_USER_INTERRUPT
;
1921 ring
->get_seqno
= gen6_ring_get_seqno
;
1922 ring
->set_seqno
= ring_set_seqno
;
1923 ring
->sync_to
= gen6_ring_sync
;
1924 ring
->semaphore_register
[RCS
] = MI_SEMAPHORE_SYNC_INVALID
;
1925 ring
->semaphore_register
[VCS
] = MI_SEMAPHORE_SYNC_RV
;
1926 ring
->semaphore_register
[BCS
] = MI_SEMAPHORE_SYNC_RB
;
1927 ring
->semaphore_register
[VECS
] = MI_SEMAPHORE_SYNC_RVE
;
1928 ring
->signal_mbox
[RCS
] = GEN6_NOSYNC
;
1929 ring
->signal_mbox
[VCS
] = GEN6_VRSYNC
;
1930 ring
->signal_mbox
[BCS
] = GEN6_BRSYNC
;
1931 ring
->signal_mbox
[VECS
] = GEN6_VERSYNC
;
1932 } else if (IS_GEN5(dev
)) {
1933 ring
->add_request
= pc_render_add_request
;
1934 ring
->flush
= gen4_render_ring_flush
;
1935 ring
->get_seqno
= pc_render_get_seqno
;
1936 ring
->set_seqno
= pc_render_set_seqno
;
1937 ring
->irq_get
= gen5_ring_get_irq
;
1938 ring
->irq_put
= gen5_ring_put_irq
;
1939 ring
->irq_enable_mask
= GT_RENDER_USER_INTERRUPT
|
1940 GT_RENDER_PIPECTL_NOTIFY_INTERRUPT
;
1942 ring
->add_request
= i9xx_add_request
;
1943 if (INTEL_INFO(dev
)->gen
< 4)
1944 ring
->flush
= gen2_render_ring_flush
;
1946 ring
->flush
= gen4_render_ring_flush
;
1947 ring
->get_seqno
= ring_get_seqno
;
1948 ring
->set_seqno
= ring_set_seqno
;
1950 ring
->irq_get
= i8xx_ring_get_irq
;
1951 ring
->irq_put
= i8xx_ring_put_irq
;
1953 ring
->irq_get
= i9xx_ring_get_irq
;
1954 ring
->irq_put
= i9xx_ring_put_irq
;
1956 ring
->irq_enable_mask
= I915_USER_INTERRUPT
;
1958 ring
->write_tail
= ring_write_tail
;
1959 if (IS_HASWELL(dev
))
1960 ring
->dispatch_execbuffer
= hsw_ring_dispatch_execbuffer
;
1961 else if (IS_GEN8(dev
))
1962 ring
->dispatch_execbuffer
= gen8_ring_dispatch_execbuffer
;
1963 else if (INTEL_INFO(dev
)->gen
>= 6)
1964 ring
->dispatch_execbuffer
= gen6_ring_dispatch_execbuffer
;
1965 else if (INTEL_INFO(dev
)->gen
>= 4)
1966 ring
->dispatch_execbuffer
= i965_dispatch_execbuffer
;
1967 else if (IS_I830(dev
) || IS_845G(dev
))
1968 ring
->dispatch_execbuffer
= i830_dispatch_execbuffer
;
1970 ring
->dispatch_execbuffer
= i915_dispatch_execbuffer
;
1971 ring
->init
= init_render_ring
;
1972 ring
->cleanup
= render_ring_cleanup
;
1974 /* Workaround batchbuffer to combat CS tlb bug. */
1975 if (HAS_BROKEN_CS_TLB(dev
)) {
1976 struct drm_i915_gem_object
*obj
;
1979 obj
= i915_gem_alloc_object(dev
, I830_BATCH_LIMIT
);
1981 DRM_ERROR("Failed to allocate batch bo\n");
1985 ret
= i915_gem_obj_ggtt_pin(obj
, 0, 0);
1987 drm_gem_object_unreference(&obj
->base
);
1988 DRM_ERROR("Failed to ping batch bo\n");
1992 ring
->scratch
.obj
= obj
;
1993 ring
->scratch
.gtt_offset
= i915_gem_obj_ggtt_offset(obj
);
1996 return intel_init_ring_buffer(dev
, ring
);
1999 int intel_render_ring_init_dri(struct drm_device
*dev
, u64 start
, u32 size
)
2001 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2002 struct intel_ring_buffer
*ring
= &dev_priv
->ring
[RCS
];
2005 ring
->name
= "render ring";
2007 ring
->mmio_base
= RENDER_RING_BASE
;
2009 if (INTEL_INFO(dev
)->gen
>= 6) {
2010 /* non-kms not supported on gen6+ */
2014 /* Note: gem is not supported on gen5/ilk without kms (the corresponding
2015 * gem_init ioctl returns with -ENODEV). Hence we do not need to set up
2016 * the special gen5 functions. */
2017 ring
->add_request
= i9xx_add_request
;
2018 if (INTEL_INFO(dev
)->gen
< 4)
2019 ring
->flush
= gen2_render_ring_flush
;
2021 ring
->flush
= gen4_render_ring_flush
;
2022 ring
->get_seqno
= ring_get_seqno
;
2023 ring
->set_seqno
= ring_set_seqno
;
2025 ring
->irq_get
= i8xx_ring_get_irq
;
2026 ring
->irq_put
= i8xx_ring_put_irq
;
2028 ring
->irq_get
= i9xx_ring_get_irq
;
2029 ring
->irq_put
= i9xx_ring_put_irq
;
2031 ring
->irq_enable_mask
= I915_USER_INTERRUPT
;
2032 ring
->write_tail
= ring_write_tail
;
2033 if (INTEL_INFO(dev
)->gen
>= 4)
2034 ring
->dispatch_execbuffer
= i965_dispatch_execbuffer
;
2035 else if (IS_I830(dev
) || IS_845G(dev
))
2036 ring
->dispatch_execbuffer
= i830_dispatch_execbuffer
;
2038 ring
->dispatch_execbuffer
= i915_dispatch_execbuffer
;
2039 ring
->init
= init_render_ring
;
2040 ring
->cleanup
= render_ring_cleanup
;
2043 INIT_LIST_HEAD(&ring
->active_list
);
2044 INIT_LIST_HEAD(&ring
->request_list
);
2047 ring
->effective_size
= ring
->size
;
2048 if (IS_I830(ring
->dev
) || IS_845G(ring
->dev
))
2049 ring
->effective_size
-= 128;
2051 ring
->virtual_start
= ioremap_wc(start
, size
);
2052 if (ring
->virtual_start
== NULL
) {
2053 DRM_ERROR("can not ioremap virtual address for"
2058 if (!I915_NEED_GFX_HWS(dev
)) {
2059 ret
= init_phys_status_page(ring
);
2067 int intel_init_bsd_ring_buffer(struct drm_device
*dev
)
2069 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2070 struct intel_ring_buffer
*ring
= &dev_priv
->ring
[VCS
];
2072 ring
->name
= "bsd ring";
2075 ring
->write_tail
= ring_write_tail
;
2076 if (INTEL_INFO(dev
)->gen
>= 6) {
2077 ring
->mmio_base
= GEN6_BSD_RING_BASE
;
2078 /* gen6 bsd needs a special wa for tail updates */
2080 ring
->write_tail
= gen6_bsd_ring_write_tail
;
2081 ring
->flush
= gen6_bsd_ring_flush
;
2082 ring
->add_request
= gen6_add_request
;
2083 ring
->get_seqno
= gen6_ring_get_seqno
;
2084 ring
->set_seqno
= ring_set_seqno
;
2085 if (INTEL_INFO(dev
)->gen
>= 8) {
2086 ring
->irq_enable_mask
=
2087 GT_RENDER_USER_INTERRUPT
<< GEN8_VCS1_IRQ_SHIFT
;
2088 ring
->irq_get
= gen8_ring_get_irq
;
2089 ring
->irq_put
= gen8_ring_put_irq
;
2090 ring
->dispatch_execbuffer
=
2091 gen8_ring_dispatch_execbuffer
;
2093 ring
->irq_enable_mask
= GT_BSD_USER_INTERRUPT
;
2094 ring
->irq_get
= gen6_ring_get_irq
;
2095 ring
->irq_put
= gen6_ring_put_irq
;
2096 ring
->dispatch_execbuffer
=
2097 gen6_ring_dispatch_execbuffer
;
2099 ring
->sync_to
= gen6_ring_sync
;
2100 ring
->semaphore_register
[RCS
] = MI_SEMAPHORE_SYNC_VR
;
2101 ring
->semaphore_register
[VCS
] = MI_SEMAPHORE_SYNC_INVALID
;
2102 ring
->semaphore_register
[BCS
] = MI_SEMAPHORE_SYNC_VB
;
2103 ring
->semaphore_register
[VECS
] = MI_SEMAPHORE_SYNC_VVE
;
2104 ring
->signal_mbox
[RCS
] = GEN6_RVSYNC
;
2105 ring
->signal_mbox
[VCS
] = GEN6_NOSYNC
;
2106 ring
->signal_mbox
[BCS
] = GEN6_BVSYNC
;
2107 ring
->signal_mbox
[VECS
] = GEN6_VEVSYNC
;
2109 ring
->mmio_base
= BSD_RING_BASE
;
2110 ring
->flush
= bsd_ring_flush
;
2111 ring
->add_request
= i9xx_add_request
;
2112 ring
->get_seqno
= ring_get_seqno
;
2113 ring
->set_seqno
= ring_set_seqno
;
2115 ring
->irq_enable_mask
= ILK_BSD_USER_INTERRUPT
;
2116 ring
->irq_get
= gen5_ring_get_irq
;
2117 ring
->irq_put
= gen5_ring_put_irq
;
2119 ring
->irq_enable_mask
= I915_BSD_USER_INTERRUPT
;
2120 ring
->irq_get
= i9xx_ring_get_irq
;
2121 ring
->irq_put
= i9xx_ring_put_irq
;
2123 ring
->dispatch_execbuffer
= i965_dispatch_execbuffer
;
2125 ring
->init
= init_ring_common
;
2127 return intel_init_ring_buffer(dev
, ring
);
2130 int intel_init_blt_ring_buffer(struct drm_device
*dev
)
2132 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2133 struct intel_ring_buffer
*ring
= &dev_priv
->ring
[BCS
];
2135 ring
->name
= "blitter ring";
2138 ring
->mmio_base
= BLT_RING_BASE
;
2139 ring
->write_tail
= ring_write_tail
;
2140 ring
->flush
= gen6_ring_flush
;
2141 ring
->add_request
= gen6_add_request
;
2142 ring
->get_seqno
= gen6_ring_get_seqno
;
2143 ring
->set_seqno
= ring_set_seqno
;
2144 if (INTEL_INFO(dev
)->gen
>= 8) {
2145 ring
->irq_enable_mask
=
2146 GT_RENDER_USER_INTERRUPT
<< GEN8_BCS_IRQ_SHIFT
;
2147 ring
->irq_get
= gen8_ring_get_irq
;
2148 ring
->irq_put
= gen8_ring_put_irq
;
2149 ring
->dispatch_execbuffer
= gen8_ring_dispatch_execbuffer
;
2151 ring
->irq_enable_mask
= GT_BLT_USER_INTERRUPT
;
2152 ring
->irq_get
= gen6_ring_get_irq
;
2153 ring
->irq_put
= gen6_ring_put_irq
;
2154 ring
->dispatch_execbuffer
= gen6_ring_dispatch_execbuffer
;
2156 ring
->sync_to
= gen6_ring_sync
;
2157 ring
->semaphore_register
[RCS
] = MI_SEMAPHORE_SYNC_BR
;
2158 ring
->semaphore_register
[VCS
] = MI_SEMAPHORE_SYNC_BV
;
2159 ring
->semaphore_register
[BCS
] = MI_SEMAPHORE_SYNC_INVALID
;
2160 ring
->semaphore_register
[VECS
] = MI_SEMAPHORE_SYNC_BVE
;
2161 ring
->signal_mbox
[RCS
] = GEN6_RBSYNC
;
2162 ring
->signal_mbox
[VCS
] = GEN6_VBSYNC
;
2163 ring
->signal_mbox
[BCS
] = GEN6_NOSYNC
;
2164 ring
->signal_mbox
[VECS
] = GEN6_VEBSYNC
;
2165 ring
->init
= init_ring_common
;
2167 return intel_init_ring_buffer(dev
, ring
);
2170 int intel_init_vebox_ring_buffer(struct drm_device
*dev
)
2172 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2173 struct intel_ring_buffer
*ring
= &dev_priv
->ring
[VECS
];
2175 ring
->name
= "video enhancement ring";
2178 ring
->mmio_base
= VEBOX_RING_BASE
;
2179 ring
->write_tail
= ring_write_tail
;
2180 ring
->flush
= gen6_ring_flush
;
2181 ring
->add_request
= gen6_add_request
;
2182 ring
->get_seqno
= gen6_ring_get_seqno
;
2183 ring
->set_seqno
= ring_set_seqno
;
2185 if (INTEL_INFO(dev
)->gen
>= 8) {
2186 ring
->irq_enable_mask
=
2187 GT_RENDER_USER_INTERRUPT
<< GEN8_VECS_IRQ_SHIFT
;
2188 ring
->irq_get
= gen8_ring_get_irq
;
2189 ring
->irq_put
= gen8_ring_put_irq
;
2190 ring
->dispatch_execbuffer
= gen8_ring_dispatch_execbuffer
;
2192 ring
->irq_enable_mask
= PM_VEBOX_USER_INTERRUPT
;
2193 ring
->irq_get
= hsw_vebox_get_irq
;
2194 ring
->irq_put
= hsw_vebox_put_irq
;
2195 ring
->dispatch_execbuffer
= gen6_ring_dispatch_execbuffer
;
2197 ring
->sync_to
= gen6_ring_sync
;
2198 ring
->semaphore_register
[RCS
] = MI_SEMAPHORE_SYNC_VER
;
2199 ring
->semaphore_register
[VCS
] = MI_SEMAPHORE_SYNC_VEV
;
2200 ring
->semaphore_register
[BCS
] = MI_SEMAPHORE_SYNC_VEB
;
2201 ring
->semaphore_register
[VECS
] = MI_SEMAPHORE_SYNC_INVALID
;
2202 ring
->signal_mbox
[RCS
] = GEN6_RVESYNC
;
2203 ring
->signal_mbox
[VCS
] = GEN6_VVESYNC
;
2204 ring
->signal_mbox
[BCS
] = GEN6_BVESYNC
;
2205 ring
->signal_mbox
[VECS
] = GEN6_NOSYNC
;
2206 ring
->init
= init_ring_common
;
2208 return intel_init_ring_buffer(dev
, ring
);
2212 intel_ring_flush_all_caches(struct intel_ring_buffer
*ring
)
2216 if (!ring
->gpu_caches_dirty
)
2219 ret
= ring
->flush(ring
, 0, I915_GEM_GPU_DOMAINS
);
2223 trace_i915_gem_ring_flush(ring
, 0, I915_GEM_GPU_DOMAINS
);
2225 ring
->gpu_caches_dirty
= false;
2230 intel_ring_invalidate_all_caches(struct intel_ring_buffer
*ring
)
2232 uint32_t flush_domains
;
2236 if (ring
->gpu_caches_dirty
)
2237 flush_domains
= I915_GEM_GPU_DOMAINS
;
2239 ret
= ring
->flush(ring
, I915_GEM_GPU_DOMAINS
, flush_domains
);
2243 trace_i915_gem_ring_flush(ring
, I915_GEM_GPU_DOMAINS
, flush_domains
);
2245 ring
->gpu_caches_dirty
= false;