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"
37 * 965+ support PIPE_CONTROL commands, which provide finer grained control
38 * over cache flushing.
41 struct drm_i915_gem_object
*obj
;
42 volatile u32
*cpu_page
;
46 static inline int ring_space(struct intel_ring_buffer
*ring
)
48 int space
= (ring
->head
& HEAD_ADDR
) - (ring
->tail
+ I915_RING_FREE_SPACE
);
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 struct pipe_control
*pc
= ring
->private;
179 u32 scratch_addr
= pc
->gtt_offset
+ 128;
183 ret
= intel_ring_begin(ring
, 6);
187 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(5));
188 intel_ring_emit(ring
, PIPE_CONTROL_CS_STALL
|
189 PIPE_CONTROL_STALL_AT_SCOREBOARD
);
190 intel_ring_emit(ring
, scratch_addr
| PIPE_CONTROL_GLOBAL_GTT
); /* address */
191 intel_ring_emit(ring
, 0); /* low dword */
192 intel_ring_emit(ring
, 0); /* high dword */
193 intel_ring_emit(ring
, MI_NOOP
);
194 intel_ring_advance(ring
);
196 ret
= intel_ring_begin(ring
, 6);
200 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(5));
201 intel_ring_emit(ring
, PIPE_CONTROL_QW_WRITE
);
202 intel_ring_emit(ring
, scratch_addr
| PIPE_CONTROL_GLOBAL_GTT
); /* address */
203 intel_ring_emit(ring
, 0);
204 intel_ring_emit(ring
, 0);
205 intel_ring_emit(ring
, MI_NOOP
);
206 intel_ring_advance(ring
);
212 gen6_render_ring_flush(struct intel_ring_buffer
*ring
,
213 u32 invalidate_domains
, u32 flush_domains
)
216 struct pipe_control
*pc
= ring
->private;
217 u32 scratch_addr
= pc
->gtt_offset
+ 128;
220 /* Force SNB workarounds for PIPE_CONTROL flushes */
221 ret
= intel_emit_post_sync_nonzero_flush(ring
);
225 /* Just flush everything. Experiments have shown that reducing the
226 * number of bits based on the write domains has little performance
230 flags
|= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH
;
231 flags
|= PIPE_CONTROL_DEPTH_CACHE_FLUSH
;
233 * Ensure that any following seqno writes only happen
234 * when the render cache is indeed flushed.
236 flags
|= PIPE_CONTROL_CS_STALL
;
238 if (invalidate_domains
) {
239 flags
|= PIPE_CONTROL_TLB_INVALIDATE
;
240 flags
|= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE
;
241 flags
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
242 flags
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
243 flags
|= PIPE_CONTROL_CONST_CACHE_INVALIDATE
;
244 flags
|= PIPE_CONTROL_STATE_CACHE_INVALIDATE
;
246 * TLB invalidate requires a post-sync write.
248 flags
|= PIPE_CONTROL_QW_WRITE
| PIPE_CONTROL_CS_STALL
;
251 ret
= intel_ring_begin(ring
, 4);
255 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(4));
256 intel_ring_emit(ring
, flags
);
257 intel_ring_emit(ring
, scratch_addr
| PIPE_CONTROL_GLOBAL_GTT
);
258 intel_ring_emit(ring
, 0);
259 intel_ring_advance(ring
);
265 gen7_render_ring_cs_stall_wa(struct intel_ring_buffer
*ring
)
269 ret
= intel_ring_begin(ring
, 4);
273 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(4));
274 intel_ring_emit(ring
, PIPE_CONTROL_CS_STALL
|
275 PIPE_CONTROL_STALL_AT_SCOREBOARD
);
276 intel_ring_emit(ring
, 0);
277 intel_ring_emit(ring
, 0);
278 intel_ring_advance(ring
);
284 gen7_render_ring_flush(struct intel_ring_buffer
*ring
,
285 u32 invalidate_domains
, u32 flush_domains
)
288 struct pipe_control
*pc
= ring
->private;
289 u32 scratch_addr
= pc
->gtt_offset
+ 128;
293 * Ensure that any following seqno writes only happen when the render
294 * cache is indeed flushed.
296 * Workaround: 4th PIPE_CONTROL command (except the ones with only
297 * read-cache invalidate bits set) must have the CS_STALL bit set. We
298 * don't try to be clever and just set it unconditionally.
300 flags
|= PIPE_CONTROL_CS_STALL
;
302 /* Just flush everything. Experiments have shown that reducing the
303 * number of bits based on the write domains has little performance
307 flags
|= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH
;
308 flags
|= PIPE_CONTROL_DEPTH_CACHE_FLUSH
;
310 if (invalidate_domains
) {
311 flags
|= PIPE_CONTROL_TLB_INVALIDATE
;
312 flags
|= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE
;
313 flags
|= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
;
314 flags
|= PIPE_CONTROL_VF_CACHE_INVALIDATE
;
315 flags
|= PIPE_CONTROL_CONST_CACHE_INVALIDATE
;
316 flags
|= PIPE_CONTROL_STATE_CACHE_INVALIDATE
;
318 * TLB invalidate requires a post-sync write.
320 flags
|= PIPE_CONTROL_QW_WRITE
;
321 flags
|= PIPE_CONTROL_GLOBAL_GTT_IVB
;
323 /* Workaround: we must issue a pipe_control with CS-stall bit
324 * set before a pipe_control command that has the state cache
325 * invalidate bit set. */
326 gen7_render_ring_cs_stall_wa(ring
);
329 ret
= intel_ring_begin(ring
, 4);
333 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(4));
334 intel_ring_emit(ring
, flags
);
335 intel_ring_emit(ring
, scratch_addr
);
336 intel_ring_emit(ring
, 0);
337 intel_ring_advance(ring
);
342 static void ring_write_tail(struct intel_ring_buffer
*ring
,
345 drm_i915_private_t
*dev_priv
= ring
->dev
->dev_private
;
346 I915_WRITE_TAIL(ring
, value
);
349 u32
intel_ring_get_active_head(struct intel_ring_buffer
*ring
)
351 drm_i915_private_t
*dev_priv
= ring
->dev
->dev_private
;
352 u32 acthd_reg
= INTEL_INFO(ring
->dev
)->gen
>= 4 ?
353 RING_ACTHD(ring
->mmio_base
) : ACTHD
;
355 return I915_READ(acthd_reg
);
358 static int init_ring_common(struct intel_ring_buffer
*ring
)
360 struct drm_device
*dev
= ring
->dev
;
361 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
362 struct drm_i915_gem_object
*obj
= ring
->obj
;
366 if (HAS_FORCE_WAKE(dev
))
367 gen6_gt_force_wake_get(dev_priv
);
369 /* Stop the ring if it's running. */
370 I915_WRITE_CTL(ring
, 0);
371 I915_WRITE_HEAD(ring
, 0);
372 ring
->write_tail(ring
, 0);
374 head
= I915_READ_HEAD(ring
) & HEAD_ADDR
;
376 /* G45 ring initialization fails to reset head to zero */
378 DRM_DEBUG_KMS("%s head not reset to zero "
379 "ctl %08x head %08x tail %08x start %08x\n",
382 I915_READ_HEAD(ring
),
383 I915_READ_TAIL(ring
),
384 I915_READ_START(ring
));
386 I915_WRITE_HEAD(ring
, 0);
388 if (I915_READ_HEAD(ring
) & HEAD_ADDR
) {
389 DRM_ERROR("failed to set %s head to zero "
390 "ctl %08x head %08x tail %08x start %08x\n",
393 I915_READ_HEAD(ring
),
394 I915_READ_TAIL(ring
),
395 I915_READ_START(ring
));
399 /* Initialize the ring. This must happen _after_ we've cleared the ring
400 * registers with the above sequence (the readback of the HEAD registers
401 * also enforces ordering), otherwise the hw might lose the new ring
402 * register values. */
403 I915_WRITE_START(ring
, obj
->gtt_offset
);
405 ((ring
->size
- PAGE_SIZE
) & RING_NR_PAGES
)
408 /* If the head is still not zero, the ring is dead */
409 if (wait_for((I915_READ_CTL(ring
) & RING_VALID
) != 0 &&
410 I915_READ_START(ring
) == obj
->gtt_offset
&&
411 (I915_READ_HEAD(ring
) & HEAD_ADDR
) == 0, 50)) {
412 DRM_ERROR("%s initialization failed "
413 "ctl %08x head %08x tail %08x start %08x\n",
416 I915_READ_HEAD(ring
),
417 I915_READ_TAIL(ring
),
418 I915_READ_START(ring
));
423 if (!drm_core_check_feature(ring
->dev
, DRIVER_MODESET
))
424 i915_kernel_lost_context(ring
->dev
);
426 ring
->head
= I915_READ_HEAD(ring
);
427 ring
->tail
= I915_READ_TAIL(ring
) & TAIL_ADDR
;
428 ring
->space
= ring_space(ring
);
429 ring
->last_retired_head
= -1;
433 if (HAS_FORCE_WAKE(dev
))
434 gen6_gt_force_wake_put(dev_priv
);
440 init_pipe_control(struct intel_ring_buffer
*ring
)
442 struct pipe_control
*pc
;
443 struct drm_i915_gem_object
*obj
;
449 pc
= kmalloc(sizeof(*pc
), GFP_KERNEL
);
453 obj
= i915_gem_alloc_object(ring
->dev
, 4096);
455 DRM_ERROR("Failed to allocate seqno page\n");
460 i915_gem_object_set_cache_level(obj
, I915_CACHE_LLC
);
462 ret
= i915_gem_object_pin(obj
, 4096, true, false);
466 pc
->gtt_offset
= obj
->gtt_offset
;
467 pc
->cpu_page
= kmap(sg_page(obj
->pages
->sgl
));
468 if (pc
->cpu_page
== NULL
)
471 DRM_DEBUG_DRIVER("%s pipe control offset: 0x%08x\n",
472 ring
->name
, pc
->gtt_offset
);
479 i915_gem_object_unpin(obj
);
481 drm_gem_object_unreference(&obj
->base
);
488 cleanup_pipe_control(struct intel_ring_buffer
*ring
)
490 struct pipe_control
*pc
= ring
->private;
491 struct drm_i915_gem_object
*obj
;
498 kunmap(sg_page(obj
->pages
->sgl
));
499 i915_gem_object_unpin(obj
);
500 drm_gem_object_unreference(&obj
->base
);
503 ring
->private = NULL
;
506 static int init_render_ring(struct intel_ring_buffer
*ring
)
508 struct drm_device
*dev
= ring
->dev
;
509 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
510 int ret
= init_ring_common(ring
);
512 if (INTEL_INFO(dev
)->gen
> 3)
513 I915_WRITE(MI_MODE
, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH
));
515 /* We need to disable the AsyncFlip performance optimisations in order
516 * to use MI_WAIT_FOR_EVENT within the CS. It should already be
517 * programmed to '1' on all products.
519 if (INTEL_INFO(dev
)->gen
>= 6)
520 I915_WRITE(MI_MODE
, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE
));
522 /* Required for the hardware to program scanline values for waiting */
523 if (INTEL_INFO(dev
)->gen
== 6)
525 _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_ALWAYS
));
528 I915_WRITE(GFX_MODE_GEN7
,
529 _MASKED_BIT_DISABLE(GFX_TLB_INVALIDATE_ALWAYS
) |
530 _MASKED_BIT_ENABLE(GFX_REPLAY_MODE
));
532 if (INTEL_INFO(dev
)->gen
>= 5) {
533 ret
= init_pipe_control(ring
);
539 /* From the Sandybridge PRM, volume 1 part 3, page 24:
540 * "If this bit is set, STCunit will have LRA as replacement
541 * policy. [...] This bit must be reset. LRA replacement
542 * policy is not supported."
544 I915_WRITE(CACHE_MODE_0
,
545 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB
));
547 /* This is not explicitly set for GEN6, so read the register.
548 * see intel_ring_mi_set_context() for why we care.
549 * TODO: consider explicitly setting the bit for GEN5
551 ring
->itlb_before_ctx_switch
=
552 !!(I915_READ(GFX_MODE
) & GFX_TLB_INVALIDATE_ALWAYS
);
555 if (INTEL_INFO(dev
)->gen
>= 6)
556 I915_WRITE(INSTPM
, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING
));
558 if (HAS_L3_GPU_CACHE(dev
))
559 I915_WRITE_IMR(ring
, ~GEN6_RENDER_L3_PARITY_ERROR
);
564 static void render_ring_cleanup(struct intel_ring_buffer
*ring
)
566 struct drm_device
*dev
= ring
->dev
;
571 if (HAS_BROKEN_CS_TLB(dev
))
572 drm_gem_object_unreference(to_gem_object(ring
->private));
574 cleanup_pipe_control(ring
);
578 update_mboxes(struct intel_ring_buffer
*ring
,
581 intel_ring_emit(ring
, MI_LOAD_REGISTER_IMM(1));
582 intel_ring_emit(ring
, mmio_offset
);
583 intel_ring_emit(ring
, ring
->outstanding_lazy_request
);
587 * gen6_add_request - Update the semaphore mailbox registers
589 * @ring - ring that is adding a request
590 * @seqno - return seqno stuck into the ring
592 * Update the mailbox registers in the *other* rings with the current seqno.
593 * This acts like a signal in the canonical semaphore.
596 gen6_add_request(struct intel_ring_buffer
*ring
)
602 ret
= intel_ring_begin(ring
, 10);
606 mbox1_reg
= ring
->signal_mbox
[0];
607 mbox2_reg
= ring
->signal_mbox
[1];
609 update_mboxes(ring
, mbox1_reg
);
610 update_mboxes(ring
, mbox2_reg
);
611 intel_ring_emit(ring
, MI_STORE_DWORD_INDEX
);
612 intel_ring_emit(ring
, I915_GEM_HWS_INDEX
<< MI_STORE_DWORD_INDEX_SHIFT
);
613 intel_ring_emit(ring
, ring
->outstanding_lazy_request
);
614 intel_ring_emit(ring
, MI_USER_INTERRUPT
);
615 intel_ring_advance(ring
);
620 static inline bool i915_gem_has_seqno_wrapped(struct drm_device
*dev
,
623 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
624 return dev_priv
->last_seqno
< seqno
;
628 * intel_ring_sync - sync the waiter to the signaller on seqno
630 * @waiter - ring that is waiting
631 * @signaller - ring which has, or will signal
632 * @seqno - seqno which the waiter will block on
635 gen6_ring_sync(struct intel_ring_buffer
*waiter
,
636 struct intel_ring_buffer
*signaller
,
640 u32 dw1
= MI_SEMAPHORE_MBOX
|
641 MI_SEMAPHORE_COMPARE
|
642 MI_SEMAPHORE_REGISTER
;
644 /* Throughout all of the GEM code, seqno passed implies our current
645 * seqno is >= the last seqno executed. However for hardware the
646 * comparison is strictly greater than.
650 WARN_ON(signaller
->semaphore_register
[waiter
->id
] ==
651 MI_SEMAPHORE_SYNC_INVALID
);
653 ret
= intel_ring_begin(waiter
, 4);
657 /* If seqno wrap happened, omit the wait with no-ops */
658 if (likely(!i915_gem_has_seqno_wrapped(waiter
->dev
, seqno
))) {
659 intel_ring_emit(waiter
,
661 signaller
->semaphore_register
[waiter
->id
]);
662 intel_ring_emit(waiter
, seqno
);
663 intel_ring_emit(waiter
, 0);
664 intel_ring_emit(waiter
, MI_NOOP
);
666 intel_ring_emit(waiter
, MI_NOOP
);
667 intel_ring_emit(waiter
, MI_NOOP
);
668 intel_ring_emit(waiter
, MI_NOOP
);
669 intel_ring_emit(waiter
, MI_NOOP
);
671 intel_ring_advance(waiter
);
676 #define PIPE_CONTROL_FLUSH(ring__, addr__) \
678 intel_ring_emit(ring__, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE | \
679 PIPE_CONTROL_DEPTH_STALL); \
680 intel_ring_emit(ring__, (addr__) | PIPE_CONTROL_GLOBAL_GTT); \
681 intel_ring_emit(ring__, 0); \
682 intel_ring_emit(ring__, 0); \
686 pc_render_add_request(struct intel_ring_buffer
*ring
)
688 struct pipe_control
*pc
= ring
->private;
689 u32 scratch_addr
= pc
->gtt_offset
+ 128;
692 /* For Ironlake, MI_USER_INTERRUPT was deprecated and apparently
693 * incoherent with writes to memory, i.e. completely fubar,
694 * so we need to use PIPE_NOTIFY instead.
696 * However, we also need to workaround the qword write
697 * incoherence by flushing the 6 PIPE_NOTIFY buffers out to
698 * memory before requesting an interrupt.
700 ret
= intel_ring_begin(ring
, 32);
704 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE
|
705 PIPE_CONTROL_WRITE_FLUSH
|
706 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
);
707 intel_ring_emit(ring
, pc
->gtt_offset
| PIPE_CONTROL_GLOBAL_GTT
);
708 intel_ring_emit(ring
, ring
->outstanding_lazy_request
);
709 intel_ring_emit(ring
, 0);
710 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
711 scratch_addr
+= 128; /* write to separate cachelines */
712 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
714 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
716 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
718 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
720 PIPE_CONTROL_FLUSH(ring
, scratch_addr
);
722 intel_ring_emit(ring
, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE
|
723 PIPE_CONTROL_WRITE_FLUSH
|
724 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE
|
725 PIPE_CONTROL_NOTIFY
);
726 intel_ring_emit(ring
, pc
->gtt_offset
| PIPE_CONTROL_GLOBAL_GTT
);
727 intel_ring_emit(ring
, ring
->outstanding_lazy_request
);
728 intel_ring_emit(ring
, 0);
729 intel_ring_advance(ring
);
735 gen6_ring_get_seqno(struct intel_ring_buffer
*ring
, bool lazy_coherency
)
737 /* Workaround to force correct ordering between irq and seqno writes on
738 * ivb (and maybe also on snb) by reading from a CS register (like
739 * ACTHD) before reading the status page. */
741 intel_ring_get_active_head(ring
);
742 return intel_read_status_page(ring
, I915_GEM_HWS_INDEX
);
746 ring_get_seqno(struct intel_ring_buffer
*ring
, bool lazy_coherency
)
748 return intel_read_status_page(ring
, I915_GEM_HWS_INDEX
);
752 ring_set_seqno(struct intel_ring_buffer
*ring
, u32 seqno
)
754 intel_write_status_page(ring
, I915_GEM_HWS_INDEX
, seqno
);
758 pc_render_get_seqno(struct intel_ring_buffer
*ring
, bool lazy_coherency
)
760 struct pipe_control
*pc
= ring
->private;
761 return pc
->cpu_page
[0];
765 pc_render_set_seqno(struct intel_ring_buffer
*ring
, u32 seqno
)
767 struct pipe_control
*pc
= ring
->private;
768 pc
->cpu_page
[0] = seqno
;
772 gen5_ring_get_irq(struct intel_ring_buffer
*ring
)
774 struct drm_device
*dev
= ring
->dev
;
775 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
778 if (!dev
->irq_enabled
)
781 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
782 if (ring
->irq_refcount
++ == 0) {
783 dev_priv
->gt_irq_mask
&= ~ring
->irq_enable_mask
;
784 I915_WRITE(GTIMR
, dev_priv
->gt_irq_mask
);
787 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
793 gen5_ring_put_irq(struct intel_ring_buffer
*ring
)
795 struct drm_device
*dev
= ring
->dev
;
796 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
799 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
800 if (--ring
->irq_refcount
== 0) {
801 dev_priv
->gt_irq_mask
|= ring
->irq_enable_mask
;
802 I915_WRITE(GTIMR
, dev_priv
->gt_irq_mask
);
805 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
809 i9xx_ring_get_irq(struct intel_ring_buffer
*ring
)
811 struct drm_device
*dev
= ring
->dev
;
812 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
815 if (!dev
->irq_enabled
)
818 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
819 if (ring
->irq_refcount
++ == 0) {
820 dev_priv
->irq_mask
&= ~ring
->irq_enable_mask
;
821 I915_WRITE(IMR
, dev_priv
->irq_mask
);
824 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
830 i9xx_ring_put_irq(struct intel_ring_buffer
*ring
)
832 struct drm_device
*dev
= ring
->dev
;
833 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
836 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
837 if (--ring
->irq_refcount
== 0) {
838 dev_priv
->irq_mask
|= ring
->irq_enable_mask
;
839 I915_WRITE(IMR
, dev_priv
->irq_mask
);
842 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
846 i8xx_ring_get_irq(struct intel_ring_buffer
*ring
)
848 struct drm_device
*dev
= ring
->dev
;
849 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
852 if (!dev
->irq_enabled
)
855 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
856 if (ring
->irq_refcount
++ == 0) {
857 dev_priv
->irq_mask
&= ~ring
->irq_enable_mask
;
858 I915_WRITE16(IMR
, dev_priv
->irq_mask
);
861 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
867 i8xx_ring_put_irq(struct intel_ring_buffer
*ring
)
869 struct drm_device
*dev
= ring
->dev
;
870 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
873 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
874 if (--ring
->irq_refcount
== 0) {
875 dev_priv
->irq_mask
|= ring
->irq_enable_mask
;
876 I915_WRITE16(IMR
, dev_priv
->irq_mask
);
879 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
882 void intel_ring_setup_status_page(struct intel_ring_buffer
*ring
)
884 struct drm_device
*dev
= ring
->dev
;
885 drm_i915_private_t
*dev_priv
= ring
->dev
->dev_private
;
888 /* The ring status page addresses are no longer next to the rest of
889 * the ring registers as of gen7.
894 mmio
= RENDER_HWS_PGA_GEN7
;
897 mmio
= BLT_HWS_PGA_GEN7
;
900 mmio
= BSD_HWS_PGA_GEN7
;
903 } else if (IS_GEN6(ring
->dev
)) {
904 mmio
= RING_HWS_PGA_GEN6(ring
->mmio_base
);
906 mmio
= RING_HWS_PGA(ring
->mmio_base
);
909 I915_WRITE(mmio
, (u32
)ring
->status_page
.gfx_addr
);
914 bsd_ring_flush(struct intel_ring_buffer
*ring
,
915 u32 invalidate_domains
,
920 ret
= intel_ring_begin(ring
, 2);
924 intel_ring_emit(ring
, MI_FLUSH
);
925 intel_ring_emit(ring
, MI_NOOP
);
926 intel_ring_advance(ring
);
931 i9xx_add_request(struct intel_ring_buffer
*ring
)
935 ret
= intel_ring_begin(ring
, 4);
939 intel_ring_emit(ring
, MI_STORE_DWORD_INDEX
);
940 intel_ring_emit(ring
, I915_GEM_HWS_INDEX
<< MI_STORE_DWORD_INDEX_SHIFT
);
941 intel_ring_emit(ring
, ring
->outstanding_lazy_request
);
942 intel_ring_emit(ring
, MI_USER_INTERRUPT
);
943 intel_ring_advance(ring
);
949 gen6_ring_get_irq(struct intel_ring_buffer
*ring
)
951 struct drm_device
*dev
= ring
->dev
;
952 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
955 if (!dev
->irq_enabled
)
958 /* It looks like we need to prevent the gt from suspending while waiting
959 * for an notifiy irq, otherwise irqs seem to get lost on at least the
960 * blt/bsd rings on ivb. */
961 gen6_gt_force_wake_get(dev_priv
);
963 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
964 if (ring
->irq_refcount
++ == 0) {
965 if (HAS_L3_GPU_CACHE(dev
) && ring
->id
== RCS
)
966 I915_WRITE_IMR(ring
, ~(ring
->irq_enable_mask
|
967 GEN6_RENDER_L3_PARITY_ERROR
));
969 I915_WRITE_IMR(ring
, ~ring
->irq_enable_mask
);
970 dev_priv
->gt_irq_mask
&= ~ring
->irq_enable_mask
;
971 I915_WRITE(GTIMR
, dev_priv
->gt_irq_mask
);
974 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
980 gen6_ring_put_irq(struct intel_ring_buffer
*ring
)
982 struct drm_device
*dev
= ring
->dev
;
983 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
986 spin_lock_irqsave(&dev_priv
->irq_lock
, flags
);
987 if (--ring
->irq_refcount
== 0) {
988 if (HAS_L3_GPU_CACHE(dev
) && ring
->id
== RCS
)
989 I915_WRITE_IMR(ring
, ~GEN6_RENDER_L3_PARITY_ERROR
);
991 I915_WRITE_IMR(ring
, ~0);
992 dev_priv
->gt_irq_mask
|= ring
->irq_enable_mask
;
993 I915_WRITE(GTIMR
, dev_priv
->gt_irq_mask
);
996 spin_unlock_irqrestore(&dev_priv
->irq_lock
, flags
);
998 gen6_gt_force_wake_put(dev_priv
);
1002 i965_dispatch_execbuffer(struct intel_ring_buffer
*ring
,
1003 u32 offset
, u32 length
,
1008 ret
= intel_ring_begin(ring
, 2);
1012 intel_ring_emit(ring
,
1013 MI_BATCH_BUFFER_START
|
1015 (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE_I965
));
1016 intel_ring_emit(ring
, offset
);
1017 intel_ring_advance(ring
);
1022 /* Just userspace ABI convention to limit the wa batch bo to a resonable size */
1023 #define I830_BATCH_LIMIT (256*1024)
1025 i830_dispatch_execbuffer(struct intel_ring_buffer
*ring
,
1026 u32 offset
, u32 len
,
1031 if (flags
& I915_DISPATCH_PINNED
) {
1032 ret
= intel_ring_begin(ring
, 4);
1036 intel_ring_emit(ring
, MI_BATCH_BUFFER
);
1037 intel_ring_emit(ring
, offset
| (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE
));
1038 intel_ring_emit(ring
, offset
+ len
- 8);
1039 intel_ring_emit(ring
, MI_NOOP
);
1040 intel_ring_advance(ring
);
1042 struct drm_i915_gem_object
*obj
= ring
->private;
1043 u32 cs_offset
= obj
->gtt_offset
;
1045 if (len
> I830_BATCH_LIMIT
)
1048 ret
= intel_ring_begin(ring
, 9+3);
1051 /* Blit the batch (which has now all relocs applied) to the stable batch
1052 * scratch bo area (so that the CS never stumbles over its tlb
1053 * invalidation bug) ... */
1054 intel_ring_emit(ring
, XY_SRC_COPY_BLT_CMD
|
1055 XY_SRC_COPY_BLT_WRITE_ALPHA
|
1056 XY_SRC_COPY_BLT_WRITE_RGB
);
1057 intel_ring_emit(ring
, BLT_DEPTH_32
| BLT_ROP_GXCOPY
| 4096);
1058 intel_ring_emit(ring
, 0);
1059 intel_ring_emit(ring
, (DIV_ROUND_UP(len
, 4096) << 16) | 1024);
1060 intel_ring_emit(ring
, cs_offset
);
1061 intel_ring_emit(ring
, 0);
1062 intel_ring_emit(ring
, 4096);
1063 intel_ring_emit(ring
, offset
);
1064 intel_ring_emit(ring
, MI_FLUSH
);
1066 /* ... and execute it. */
1067 intel_ring_emit(ring
, MI_BATCH_BUFFER
);
1068 intel_ring_emit(ring
, cs_offset
| (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE
));
1069 intel_ring_emit(ring
, cs_offset
+ len
- 8);
1070 intel_ring_advance(ring
);
1077 i915_dispatch_execbuffer(struct intel_ring_buffer
*ring
,
1078 u32 offset
, u32 len
,
1083 ret
= intel_ring_begin(ring
, 2);
1087 intel_ring_emit(ring
, MI_BATCH_BUFFER_START
| MI_BATCH_GTT
);
1088 intel_ring_emit(ring
, offset
| (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE
));
1089 intel_ring_advance(ring
);
1094 static void cleanup_status_page(struct intel_ring_buffer
*ring
)
1096 struct drm_i915_gem_object
*obj
;
1098 obj
= ring
->status_page
.obj
;
1102 kunmap(sg_page(obj
->pages
->sgl
));
1103 i915_gem_object_unpin(obj
);
1104 drm_gem_object_unreference(&obj
->base
);
1105 ring
->status_page
.obj
= NULL
;
1108 static int init_status_page(struct intel_ring_buffer
*ring
)
1110 struct drm_device
*dev
= ring
->dev
;
1111 struct drm_i915_gem_object
*obj
;
1114 obj
= i915_gem_alloc_object(dev
, 4096);
1116 DRM_ERROR("Failed to allocate status page\n");
1121 i915_gem_object_set_cache_level(obj
, I915_CACHE_LLC
);
1123 ret
= i915_gem_object_pin(obj
, 4096, true, false);
1128 ring
->status_page
.gfx_addr
= obj
->gtt_offset
;
1129 ring
->status_page
.page_addr
= kmap(sg_page(obj
->pages
->sgl
));
1130 if (ring
->status_page
.page_addr
== NULL
) {
1134 ring
->status_page
.obj
= obj
;
1135 memset(ring
->status_page
.page_addr
, 0, PAGE_SIZE
);
1137 intel_ring_setup_status_page(ring
);
1138 DRM_DEBUG_DRIVER("%s hws offset: 0x%08x\n",
1139 ring
->name
, ring
->status_page
.gfx_addr
);
1144 i915_gem_object_unpin(obj
);
1146 drm_gem_object_unreference(&obj
->base
);
1151 static int init_phys_hws_pga(struct intel_ring_buffer
*ring
)
1153 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
1156 if (!dev_priv
->status_page_dmah
) {
1157 dev_priv
->status_page_dmah
=
1158 drm_pci_alloc(ring
->dev
, PAGE_SIZE
, PAGE_SIZE
);
1159 if (!dev_priv
->status_page_dmah
)
1163 addr
= dev_priv
->status_page_dmah
->busaddr
;
1164 if (INTEL_INFO(ring
->dev
)->gen
>= 4)
1165 addr
|= (dev_priv
->status_page_dmah
->busaddr
>> 28) & 0xf0;
1166 I915_WRITE(HWS_PGA
, addr
);
1168 ring
->status_page
.page_addr
= dev_priv
->status_page_dmah
->vaddr
;
1169 memset(ring
->status_page
.page_addr
, 0, PAGE_SIZE
);
1174 static int intel_init_ring_buffer(struct drm_device
*dev
,
1175 struct intel_ring_buffer
*ring
)
1177 struct drm_i915_gem_object
*obj
;
1178 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1182 INIT_LIST_HEAD(&ring
->active_list
);
1183 INIT_LIST_HEAD(&ring
->request_list
);
1184 ring
->size
= 32 * PAGE_SIZE
;
1185 memset(ring
->sync_seqno
, 0, sizeof(ring
->sync_seqno
));
1187 init_waitqueue_head(&ring
->irq_queue
);
1189 if (I915_NEED_GFX_HWS(dev
)) {
1190 ret
= init_status_page(ring
);
1194 BUG_ON(ring
->id
!= RCS
);
1195 ret
= init_phys_hws_pga(ring
);
1202 obj
= i915_gem_object_create_stolen(dev
, ring
->size
);
1204 obj
= i915_gem_alloc_object(dev
, ring
->size
);
1206 DRM_ERROR("Failed to allocate ringbuffer\n");
1213 ret
= i915_gem_object_pin(obj
, PAGE_SIZE
, true, false);
1217 ret
= i915_gem_object_set_to_gtt_domain(obj
, true);
1221 ring
->virtual_start
=
1222 ioremap_wc(dev_priv
->gtt
.mappable_base
+ obj
->gtt_offset
,
1224 if (ring
->virtual_start
== NULL
) {
1225 DRM_ERROR("Failed to map ringbuffer.\n");
1230 ret
= ring
->init(ring
);
1234 /* Workaround an erratum on the i830 which causes a hang if
1235 * the TAIL pointer points to within the last 2 cachelines
1238 ring
->effective_size
= ring
->size
;
1239 if (IS_I830(ring
->dev
) || IS_845G(ring
->dev
))
1240 ring
->effective_size
-= 128;
1245 iounmap(ring
->virtual_start
);
1247 i915_gem_object_unpin(obj
);
1249 drm_gem_object_unreference(&obj
->base
);
1252 cleanup_status_page(ring
);
1256 void intel_cleanup_ring_buffer(struct intel_ring_buffer
*ring
)
1258 struct drm_i915_private
*dev_priv
;
1261 if (ring
->obj
== NULL
)
1264 /* Disable the ring buffer. The ring must be idle at this point */
1265 dev_priv
= ring
->dev
->dev_private
;
1266 ret
= intel_ring_idle(ring
);
1268 DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",
1271 I915_WRITE_CTL(ring
, 0);
1273 iounmap(ring
->virtual_start
);
1275 i915_gem_object_unpin(ring
->obj
);
1276 drm_gem_object_unreference(&ring
->obj
->base
);
1280 ring
->cleanup(ring
);
1282 cleanup_status_page(ring
);
1285 static int intel_ring_wait_seqno(struct intel_ring_buffer
*ring
, u32 seqno
)
1289 ret
= i915_wait_seqno(ring
, seqno
);
1291 i915_gem_retire_requests_ring(ring
);
1296 static int intel_ring_wait_request(struct intel_ring_buffer
*ring
, int n
)
1298 struct drm_i915_gem_request
*request
;
1302 i915_gem_retire_requests_ring(ring
);
1304 if (ring
->last_retired_head
!= -1) {
1305 ring
->head
= ring
->last_retired_head
;
1306 ring
->last_retired_head
= -1;
1307 ring
->space
= ring_space(ring
);
1308 if (ring
->space
>= n
)
1312 list_for_each_entry(request
, &ring
->request_list
, list
) {
1315 if (request
->tail
== -1)
1318 space
= request
->tail
- (ring
->tail
+ I915_RING_FREE_SPACE
);
1320 space
+= ring
->size
;
1322 seqno
= request
->seqno
;
1326 /* Consume this request in case we need more space than
1327 * is available and so need to prevent a race between
1328 * updating last_retired_head and direct reads of
1329 * I915_RING_HEAD. It also provides a nice sanity check.
1337 ret
= intel_ring_wait_seqno(ring
, seqno
);
1341 if (WARN_ON(ring
->last_retired_head
== -1))
1344 ring
->head
= ring
->last_retired_head
;
1345 ring
->last_retired_head
= -1;
1346 ring
->space
= ring_space(ring
);
1347 if (WARN_ON(ring
->space
< n
))
1353 static int ring_wait_for_space(struct intel_ring_buffer
*ring
, int n
)
1355 struct drm_device
*dev
= ring
->dev
;
1356 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1360 ret
= intel_ring_wait_request(ring
, n
);
1364 trace_i915_ring_wait_begin(ring
);
1365 /* With GEM the hangcheck timer should kick us out of the loop,
1366 * leaving it early runs the risk of corrupting GEM state (due
1367 * to running on almost untested codepaths). But on resume
1368 * timers don't work yet, so prevent a complete hang in that
1369 * case by choosing an insanely large timeout. */
1370 end
= jiffies
+ 60 * HZ
;
1373 ring
->head
= I915_READ_HEAD(ring
);
1374 ring
->space
= ring_space(ring
);
1375 if (ring
->space
>= n
) {
1376 trace_i915_ring_wait_end(ring
);
1380 if (dev
->primary
->master
) {
1381 struct drm_i915_master_private
*master_priv
= dev
->primary
->master
->driver_priv
;
1382 if (master_priv
->sarea_priv
)
1383 master_priv
->sarea_priv
->perf_boxes
|= I915_BOX_WAIT
;
1388 ret
= i915_gem_check_wedge(&dev_priv
->gpu_error
,
1389 dev_priv
->mm
.interruptible
);
1392 } while (!time_after(jiffies
, end
));
1393 trace_i915_ring_wait_end(ring
);
1397 static int intel_wrap_ring_buffer(struct intel_ring_buffer
*ring
)
1399 uint32_t __iomem
*virt
;
1400 int rem
= ring
->size
- ring
->tail
;
1402 if (ring
->space
< rem
) {
1403 int ret
= ring_wait_for_space(ring
, rem
);
1408 virt
= ring
->virtual_start
+ ring
->tail
;
1411 iowrite32(MI_NOOP
, virt
++);
1414 ring
->space
= ring_space(ring
);
1419 int intel_ring_idle(struct intel_ring_buffer
*ring
)
1424 /* We need to add any requests required to flush the objects and ring */
1425 if (ring
->outstanding_lazy_request
) {
1426 ret
= i915_add_request(ring
, NULL
, NULL
);
1431 /* Wait upon the last request to be completed */
1432 if (list_empty(&ring
->request_list
))
1435 seqno
= list_entry(ring
->request_list
.prev
,
1436 struct drm_i915_gem_request
,
1439 return i915_wait_seqno(ring
, seqno
);
1443 intel_ring_alloc_seqno(struct intel_ring_buffer
*ring
)
1445 if (ring
->outstanding_lazy_request
)
1448 return i915_gem_get_seqno(ring
->dev
, &ring
->outstanding_lazy_request
);
1451 static int __intel_ring_begin(struct intel_ring_buffer
*ring
,
1456 if (unlikely(ring
->tail
+ bytes
> ring
->effective_size
)) {
1457 ret
= intel_wrap_ring_buffer(ring
);
1462 if (unlikely(ring
->space
< bytes
)) {
1463 ret
= ring_wait_for_space(ring
, bytes
);
1468 ring
->space
-= bytes
;
1472 int intel_ring_begin(struct intel_ring_buffer
*ring
,
1475 drm_i915_private_t
*dev_priv
= ring
->dev
->dev_private
;
1478 ret
= i915_gem_check_wedge(&dev_priv
->gpu_error
,
1479 dev_priv
->mm
.interruptible
);
1483 /* Preallocate the olr before touching the ring */
1484 ret
= intel_ring_alloc_seqno(ring
);
1488 return __intel_ring_begin(ring
, num_dwords
* sizeof(uint32_t));
1491 void intel_ring_init_seqno(struct intel_ring_buffer
*ring
, u32 seqno
)
1493 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
1495 BUG_ON(ring
->outstanding_lazy_request
);
1497 if (INTEL_INFO(ring
->dev
)->gen
>= 6) {
1498 I915_WRITE(RING_SYNC_0(ring
->mmio_base
), 0);
1499 I915_WRITE(RING_SYNC_1(ring
->mmio_base
), 0);
1502 ring
->set_seqno(ring
, seqno
);
1505 void intel_ring_advance(struct intel_ring_buffer
*ring
)
1507 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
1509 ring
->tail
&= ring
->size
- 1;
1510 if (dev_priv
->gpu_error
.stop_rings
& intel_ring_flag(ring
))
1512 ring
->write_tail(ring
, ring
->tail
);
1516 static void gen6_bsd_ring_write_tail(struct intel_ring_buffer
*ring
,
1519 drm_i915_private_t
*dev_priv
= ring
->dev
->dev_private
;
1521 /* Every tail move must follow the sequence below */
1523 /* Disable notification that the ring is IDLE. The GT
1524 * will then assume that it is busy and bring it out of rc6.
1526 I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL
,
1527 _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE
));
1529 /* Clear the context id. Here be magic! */
1530 I915_WRITE64(GEN6_BSD_RNCID
, 0x0);
1532 /* Wait for the ring not to be idle, i.e. for it to wake up. */
1533 if (wait_for((I915_READ(GEN6_BSD_SLEEP_PSMI_CONTROL
) &
1534 GEN6_BSD_SLEEP_INDICATOR
) == 0,
1536 DRM_ERROR("timed out waiting for the BSD ring to wake up\n");
1538 /* Now that the ring is fully powered up, update the tail */
1539 I915_WRITE_TAIL(ring
, value
);
1540 POSTING_READ(RING_TAIL(ring
->mmio_base
));
1542 /* Let the ring send IDLE messages to the GT again,
1543 * and so let it sleep to conserve power when idle.
1545 I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL
,
1546 _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE
));
1549 static int gen6_ring_flush(struct intel_ring_buffer
*ring
,
1550 u32 invalidate
, u32 flush
)
1555 ret
= intel_ring_begin(ring
, 4);
1561 * Bspec vol 1c.5 - video engine command streamer:
1562 * "If ENABLED, all TLBs will be invalidated once the flush
1563 * operation is complete. This bit is only valid when the
1564 * Post-Sync Operation field is a value of 1h or 3h."
1566 if (invalidate
& I915_GEM_GPU_DOMAINS
)
1567 cmd
|= MI_INVALIDATE_TLB
| MI_INVALIDATE_BSD
|
1568 MI_FLUSH_DW_STORE_INDEX
| MI_FLUSH_DW_OP_STOREDW
;
1569 intel_ring_emit(ring
, cmd
);
1570 intel_ring_emit(ring
, I915_GEM_HWS_SCRATCH_ADDR
| MI_FLUSH_DW_USE_GTT
);
1571 intel_ring_emit(ring
, 0);
1572 intel_ring_emit(ring
, MI_NOOP
);
1573 intel_ring_advance(ring
);
1578 hsw_ring_dispatch_execbuffer(struct intel_ring_buffer
*ring
,
1579 u32 offset
, u32 len
,
1584 ret
= intel_ring_begin(ring
, 2);
1588 intel_ring_emit(ring
,
1589 MI_BATCH_BUFFER_START
| MI_BATCH_PPGTT_HSW
|
1590 (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE_HSW
));
1591 /* bit0-7 is the length on GEN6+ */
1592 intel_ring_emit(ring
, offset
);
1593 intel_ring_advance(ring
);
1599 gen6_ring_dispatch_execbuffer(struct intel_ring_buffer
*ring
,
1600 u32 offset
, u32 len
,
1605 ret
= intel_ring_begin(ring
, 2);
1609 intel_ring_emit(ring
,
1610 MI_BATCH_BUFFER_START
|
1611 (flags
& I915_DISPATCH_SECURE
? 0 : MI_BATCH_NON_SECURE_I965
));
1612 /* bit0-7 is the length on GEN6+ */
1613 intel_ring_emit(ring
, offset
);
1614 intel_ring_advance(ring
);
1619 /* Blitter support (SandyBridge+) */
1621 static int blt_ring_flush(struct intel_ring_buffer
*ring
,
1622 u32 invalidate
, u32 flush
)
1627 ret
= intel_ring_begin(ring
, 4);
1633 * Bspec vol 1c.3 - blitter engine command streamer:
1634 * "If ENABLED, all TLBs will be invalidated once the flush
1635 * operation is complete. This bit is only valid when the
1636 * Post-Sync Operation field is a value of 1h or 3h."
1638 if (invalidate
& I915_GEM_DOMAIN_RENDER
)
1639 cmd
|= MI_INVALIDATE_TLB
| MI_FLUSH_DW_STORE_INDEX
|
1640 MI_FLUSH_DW_OP_STOREDW
;
1641 intel_ring_emit(ring
, cmd
);
1642 intel_ring_emit(ring
, I915_GEM_HWS_SCRATCH_ADDR
| MI_FLUSH_DW_USE_GTT
);
1643 intel_ring_emit(ring
, 0);
1644 intel_ring_emit(ring
, MI_NOOP
);
1645 intel_ring_advance(ring
);
1649 int intel_init_render_ring_buffer(struct drm_device
*dev
)
1651 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
1652 struct intel_ring_buffer
*ring
= &dev_priv
->ring
[RCS
];
1654 ring
->name
= "render ring";
1656 ring
->mmio_base
= RENDER_RING_BASE
;
1658 if (INTEL_INFO(dev
)->gen
>= 6) {
1659 ring
->add_request
= gen6_add_request
;
1660 ring
->flush
= gen7_render_ring_flush
;
1661 if (INTEL_INFO(dev
)->gen
== 6)
1662 ring
->flush
= gen6_render_ring_flush
;
1663 ring
->irq_get
= gen6_ring_get_irq
;
1664 ring
->irq_put
= gen6_ring_put_irq
;
1665 ring
->irq_enable_mask
= GT_USER_INTERRUPT
;
1666 ring
->get_seqno
= gen6_ring_get_seqno
;
1667 ring
->set_seqno
= ring_set_seqno
;
1668 ring
->sync_to
= gen6_ring_sync
;
1669 ring
->semaphore_register
[0] = MI_SEMAPHORE_SYNC_INVALID
;
1670 ring
->semaphore_register
[1] = MI_SEMAPHORE_SYNC_RV
;
1671 ring
->semaphore_register
[2] = MI_SEMAPHORE_SYNC_RB
;
1672 ring
->signal_mbox
[0] = GEN6_VRSYNC
;
1673 ring
->signal_mbox
[1] = GEN6_BRSYNC
;
1674 } else if (IS_GEN5(dev
)) {
1675 ring
->add_request
= pc_render_add_request
;
1676 ring
->flush
= gen4_render_ring_flush
;
1677 ring
->get_seqno
= pc_render_get_seqno
;
1678 ring
->set_seqno
= pc_render_set_seqno
;
1679 ring
->irq_get
= gen5_ring_get_irq
;
1680 ring
->irq_put
= gen5_ring_put_irq
;
1681 ring
->irq_enable_mask
= GT_USER_INTERRUPT
| GT_PIPE_NOTIFY
;
1683 ring
->add_request
= i9xx_add_request
;
1684 if (INTEL_INFO(dev
)->gen
< 4)
1685 ring
->flush
= gen2_render_ring_flush
;
1687 ring
->flush
= gen4_render_ring_flush
;
1688 ring
->get_seqno
= ring_get_seqno
;
1689 ring
->set_seqno
= ring_set_seqno
;
1691 ring
->irq_get
= i8xx_ring_get_irq
;
1692 ring
->irq_put
= i8xx_ring_put_irq
;
1694 ring
->irq_get
= i9xx_ring_get_irq
;
1695 ring
->irq_put
= i9xx_ring_put_irq
;
1697 ring
->irq_enable_mask
= I915_USER_INTERRUPT
;
1699 ring
->write_tail
= ring_write_tail
;
1700 if (IS_HASWELL(dev
))
1701 ring
->dispatch_execbuffer
= hsw_ring_dispatch_execbuffer
;
1702 else if (INTEL_INFO(dev
)->gen
>= 6)
1703 ring
->dispatch_execbuffer
= gen6_ring_dispatch_execbuffer
;
1704 else if (INTEL_INFO(dev
)->gen
>= 4)
1705 ring
->dispatch_execbuffer
= i965_dispatch_execbuffer
;
1706 else if (IS_I830(dev
) || IS_845G(dev
))
1707 ring
->dispatch_execbuffer
= i830_dispatch_execbuffer
;
1709 ring
->dispatch_execbuffer
= i915_dispatch_execbuffer
;
1710 ring
->init
= init_render_ring
;
1711 ring
->cleanup
= render_ring_cleanup
;
1713 /* Workaround batchbuffer to combat CS tlb bug. */
1714 if (HAS_BROKEN_CS_TLB(dev
)) {
1715 struct drm_i915_gem_object
*obj
;
1718 obj
= i915_gem_alloc_object(dev
, I830_BATCH_LIMIT
);
1720 DRM_ERROR("Failed to allocate batch bo\n");
1724 ret
= i915_gem_object_pin(obj
, 0, true, false);
1726 drm_gem_object_unreference(&obj
->base
);
1727 DRM_ERROR("Failed to ping batch bo\n");
1731 ring
->private = obj
;
1734 return intel_init_ring_buffer(dev
, ring
);
1737 int intel_render_ring_init_dri(struct drm_device
*dev
, u64 start
, u32 size
)
1739 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
1740 struct intel_ring_buffer
*ring
= &dev_priv
->ring
[RCS
];
1743 ring
->name
= "render ring";
1745 ring
->mmio_base
= RENDER_RING_BASE
;
1747 if (INTEL_INFO(dev
)->gen
>= 6) {
1748 /* non-kms not supported on gen6+ */
1752 /* Note: gem is not supported on gen5/ilk without kms (the corresponding
1753 * gem_init ioctl returns with -ENODEV). Hence we do not need to set up
1754 * the special gen5 functions. */
1755 ring
->add_request
= i9xx_add_request
;
1756 if (INTEL_INFO(dev
)->gen
< 4)
1757 ring
->flush
= gen2_render_ring_flush
;
1759 ring
->flush
= gen4_render_ring_flush
;
1760 ring
->get_seqno
= ring_get_seqno
;
1761 ring
->set_seqno
= ring_set_seqno
;
1763 ring
->irq_get
= i8xx_ring_get_irq
;
1764 ring
->irq_put
= i8xx_ring_put_irq
;
1766 ring
->irq_get
= i9xx_ring_get_irq
;
1767 ring
->irq_put
= i9xx_ring_put_irq
;
1769 ring
->irq_enable_mask
= I915_USER_INTERRUPT
;
1770 ring
->write_tail
= ring_write_tail
;
1771 if (INTEL_INFO(dev
)->gen
>= 4)
1772 ring
->dispatch_execbuffer
= i965_dispatch_execbuffer
;
1773 else if (IS_I830(dev
) || IS_845G(dev
))
1774 ring
->dispatch_execbuffer
= i830_dispatch_execbuffer
;
1776 ring
->dispatch_execbuffer
= i915_dispatch_execbuffer
;
1777 ring
->init
= init_render_ring
;
1778 ring
->cleanup
= render_ring_cleanup
;
1781 INIT_LIST_HEAD(&ring
->active_list
);
1782 INIT_LIST_HEAD(&ring
->request_list
);
1785 ring
->effective_size
= ring
->size
;
1786 if (IS_I830(ring
->dev
) || IS_845G(ring
->dev
))
1787 ring
->effective_size
-= 128;
1789 ring
->virtual_start
= ioremap_wc(start
, size
);
1790 if (ring
->virtual_start
== NULL
) {
1791 DRM_ERROR("can not ioremap virtual address for"
1796 if (!I915_NEED_GFX_HWS(dev
)) {
1797 ret
= init_phys_hws_pga(ring
);
1805 int intel_init_bsd_ring_buffer(struct drm_device
*dev
)
1807 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
1808 struct intel_ring_buffer
*ring
= &dev_priv
->ring
[VCS
];
1810 ring
->name
= "bsd ring";
1813 ring
->write_tail
= ring_write_tail
;
1814 if (IS_GEN6(dev
) || IS_GEN7(dev
)) {
1815 ring
->mmio_base
= GEN6_BSD_RING_BASE
;
1816 /* gen6 bsd needs a special wa for tail updates */
1818 ring
->write_tail
= gen6_bsd_ring_write_tail
;
1819 ring
->flush
= gen6_ring_flush
;
1820 ring
->add_request
= gen6_add_request
;
1821 ring
->get_seqno
= gen6_ring_get_seqno
;
1822 ring
->set_seqno
= ring_set_seqno
;
1823 ring
->irq_enable_mask
= GEN6_BSD_USER_INTERRUPT
;
1824 ring
->irq_get
= gen6_ring_get_irq
;
1825 ring
->irq_put
= gen6_ring_put_irq
;
1826 ring
->dispatch_execbuffer
= gen6_ring_dispatch_execbuffer
;
1827 ring
->sync_to
= gen6_ring_sync
;
1828 ring
->semaphore_register
[0] = MI_SEMAPHORE_SYNC_VR
;
1829 ring
->semaphore_register
[1] = MI_SEMAPHORE_SYNC_INVALID
;
1830 ring
->semaphore_register
[2] = MI_SEMAPHORE_SYNC_VB
;
1831 ring
->signal_mbox
[0] = GEN6_RVSYNC
;
1832 ring
->signal_mbox
[1] = GEN6_BVSYNC
;
1834 ring
->mmio_base
= BSD_RING_BASE
;
1835 ring
->flush
= bsd_ring_flush
;
1836 ring
->add_request
= i9xx_add_request
;
1837 ring
->get_seqno
= ring_get_seqno
;
1838 ring
->set_seqno
= ring_set_seqno
;
1840 ring
->irq_enable_mask
= GT_BSD_USER_INTERRUPT
;
1841 ring
->irq_get
= gen5_ring_get_irq
;
1842 ring
->irq_put
= gen5_ring_put_irq
;
1844 ring
->irq_enable_mask
= I915_BSD_USER_INTERRUPT
;
1845 ring
->irq_get
= i9xx_ring_get_irq
;
1846 ring
->irq_put
= i9xx_ring_put_irq
;
1848 ring
->dispatch_execbuffer
= i965_dispatch_execbuffer
;
1850 ring
->init
= init_ring_common
;
1852 return intel_init_ring_buffer(dev
, ring
);
1855 int intel_init_blt_ring_buffer(struct drm_device
*dev
)
1857 drm_i915_private_t
*dev_priv
= dev
->dev_private
;
1858 struct intel_ring_buffer
*ring
= &dev_priv
->ring
[BCS
];
1860 ring
->name
= "blitter ring";
1863 ring
->mmio_base
= BLT_RING_BASE
;
1864 ring
->write_tail
= ring_write_tail
;
1865 ring
->flush
= blt_ring_flush
;
1866 ring
->add_request
= gen6_add_request
;
1867 ring
->get_seqno
= gen6_ring_get_seqno
;
1868 ring
->set_seqno
= ring_set_seqno
;
1869 ring
->irq_enable_mask
= GEN6_BLITTER_USER_INTERRUPT
;
1870 ring
->irq_get
= gen6_ring_get_irq
;
1871 ring
->irq_put
= gen6_ring_put_irq
;
1872 ring
->dispatch_execbuffer
= gen6_ring_dispatch_execbuffer
;
1873 ring
->sync_to
= gen6_ring_sync
;
1874 ring
->semaphore_register
[0] = MI_SEMAPHORE_SYNC_BR
;
1875 ring
->semaphore_register
[1] = MI_SEMAPHORE_SYNC_BV
;
1876 ring
->semaphore_register
[2] = MI_SEMAPHORE_SYNC_INVALID
;
1877 ring
->signal_mbox
[0] = GEN6_RBSYNC
;
1878 ring
->signal_mbox
[1] = GEN6_VBSYNC
;
1879 ring
->init
= init_ring_common
;
1881 return intel_init_ring_buffer(dev
, ring
);
1885 intel_ring_flush_all_caches(struct intel_ring_buffer
*ring
)
1889 if (!ring
->gpu_caches_dirty
)
1892 ret
= ring
->flush(ring
, 0, I915_GEM_GPU_DOMAINS
);
1896 trace_i915_gem_ring_flush(ring
, 0, I915_GEM_GPU_DOMAINS
);
1898 ring
->gpu_caches_dirty
= false;
1903 intel_ring_invalidate_all_caches(struct intel_ring_buffer
*ring
)
1905 uint32_t flush_domains
;
1909 if (ring
->gpu_caches_dirty
)
1910 flush_domains
= I915_GEM_GPU_DOMAINS
;
1912 ret
= ring
->flush(ring
, I915_GEM_GPU_DOMAINS
, flush_domains
);
1916 trace_i915_gem_ring_flush(ring
, I915_GEM_GPU_DOMAINS
, flush_domains
);
1918 ring
->gpu_caches_dirty
= false;