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);
283 static int gen7_ring_fbc_flush(struct intel_ring_buffer *ring, u32 value)
287 if (!ring->fbc_dirty)
290 ret = intel_ring_begin(ring, 4);
293 intel_ring_emit(ring, MI_NOOP);
294 /* WaFbcNukeOn3DBlt:ivb/hsw */
295 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
296 intel_ring_emit(ring, MSG_FBC_REND_STATE);
297 intel_ring_emit(ring, value);
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 struct pipe_control *pc = ring->private;
310 u32 scratch_addr = pc->gtt_offset + 128;
314 * Ensure that any following seqno writes only happen when the render
315 * cache is indeed flushed.
317 * Workaround: 4th PIPE_CONTROL command (except the ones with only
318 * read-cache invalidate bits set) must have the CS_STALL bit set. We
319 * don't try to be clever and just set it unconditionally.
321 flags |= PIPE_CONTROL_CS_STALL;
323 /* Just flush everything. Experiments have shown that reducing the
324 * number of bits based on the write domains has little performance
328 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
329 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
331 if (invalidate_domains) {
332 flags |= PIPE_CONTROL_TLB_INVALIDATE;
333 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
334 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
335 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
336 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
337 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
339 * TLB invalidate requires a post-sync write.
341 flags |= PIPE_CONTROL_QW_WRITE;
342 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
344 /* Workaround: we must issue a pipe_control with CS-stall bit
345 * set before a pipe_control command that has the state cache
346 * invalidate bit set. */
347 gen7_render_ring_cs_stall_wa(ring);
350 ret = intel_ring_begin(ring, 4);
354 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4));
355 intel_ring_emit(ring, flags);
356 intel_ring_emit(ring, scratch_addr);
357 intel_ring_emit(ring, 0);
358 intel_ring_advance(ring);
361 return gen7_ring_fbc_flush(ring, FBC_REND_NUKE);
366 static void ring_write_tail(struct intel_ring_buffer *ring,
369 drm_i915_private_t *dev_priv = ring->dev->dev_private;
370 I915_WRITE_TAIL(ring, value);
373 u32 intel_ring_get_active_head(struct intel_ring_buffer *ring)
375 drm_i915_private_t *dev_priv = ring->dev->dev_private;
376 u32 acthd_reg = INTEL_INFO(ring->dev)->gen >= 4 ?
377 RING_ACTHD(ring->mmio_base) : ACTHD;
379 return I915_READ(acthd_reg);
382 static int init_ring_common(struct intel_ring_buffer *ring)
384 struct drm_device *dev = ring->dev;
385 drm_i915_private_t *dev_priv = dev->dev_private;
386 struct drm_i915_gem_object *obj = ring->obj;
390 if (HAS_FORCE_WAKE(dev))
391 gen6_gt_force_wake_get(dev_priv);
393 /* Stop the ring if it's running. */
394 I915_WRITE_CTL(ring, 0);
395 I915_WRITE_HEAD(ring, 0);
396 ring->write_tail(ring, 0);
398 head = I915_READ_HEAD(ring) & HEAD_ADDR;
400 /* G45 ring initialization fails to reset head to zero */
402 DRM_DEBUG_KMS("%s head not reset to zero "
403 "ctl %08x head %08x tail %08x start %08x\n",
406 I915_READ_HEAD(ring),
407 I915_READ_TAIL(ring),
408 I915_READ_START(ring));
410 I915_WRITE_HEAD(ring, 0);
412 if (I915_READ_HEAD(ring) & HEAD_ADDR) {
413 DRM_ERROR("failed to set %s head to zero "
414 "ctl %08x head %08x tail %08x start %08x\n",
417 I915_READ_HEAD(ring),
418 I915_READ_TAIL(ring),
419 I915_READ_START(ring));
423 /* Initialize the ring. This must happen _after_ we've cleared the ring
424 * registers with the above sequence (the readback of the HEAD registers
425 * also enforces ordering), otherwise the hw might lose the new ring
426 * register values. */
427 I915_WRITE_START(ring, obj->gtt_offset);
429 ((ring->size - PAGE_SIZE) & RING_NR_PAGES)
432 /* If the head is still not zero, the ring is dead */
433 if (wait_for((I915_READ_CTL(ring) & RING_VALID) != 0 &&
434 I915_READ_START(ring) == obj->gtt_offset &&
435 (I915_READ_HEAD(ring) & HEAD_ADDR) == 0, 50)) {
436 DRM_ERROR("%s initialization failed "
437 "ctl %08x head %08x tail %08x start %08x\n",
440 I915_READ_HEAD(ring),
441 I915_READ_TAIL(ring),
442 I915_READ_START(ring));
447 if (!drm_core_check_feature(ring->dev, DRIVER_MODESET))
448 i915_kernel_lost_context(ring->dev);
450 ring->head = I915_READ_HEAD(ring);
451 ring->tail = I915_READ_TAIL(ring) & TAIL_ADDR;
452 ring->space = ring_space(ring);
453 ring->last_retired_head = -1;
456 memset(&ring->hangcheck, 0, sizeof(ring->hangcheck));
459 if (HAS_FORCE_WAKE(dev))
460 gen6_gt_force_wake_put(dev_priv);
466 init_pipe_control(struct intel_ring_buffer *ring)
468 struct pipe_control *pc;
469 struct drm_i915_gem_object *obj;
475 pc = kmalloc(sizeof(*pc), GFP_KERNEL);
479 obj = i915_gem_alloc_object(ring->dev, 4096);
481 DRM_ERROR("Failed to allocate seqno page\n");
486 i915_gem_object_set_cache_level(obj, I915_CACHE_LLC);
488 ret = i915_gem_object_pin(obj, 4096, true, false);
492 pc->gtt_offset = obj->gtt_offset;
493 pc->cpu_page = kmap(sg_page(obj->pages->sgl));
494 if (pc->cpu_page == NULL) {
499 DRM_DEBUG_DRIVER("%s pipe control offset: 0x%08x\n",
500 ring->name, pc->gtt_offset);
507 i915_gem_object_unpin(obj);
509 drm_gem_object_unreference(&obj->base);
516 cleanup_pipe_control(struct intel_ring_buffer *ring)
518 struct pipe_control *pc = ring->private;
519 struct drm_i915_gem_object *obj;
526 kunmap(sg_page(obj->pages->sgl));
527 i915_gem_object_unpin(obj);
528 drm_gem_object_unreference(&obj->base);
531 ring->private = NULL;
534 static int init_render_ring(struct intel_ring_buffer *ring)
536 struct drm_device *dev = ring->dev;
537 struct drm_i915_private *dev_priv = dev->dev_private;
538 int ret = init_ring_common(ring);
540 if (INTEL_INFO(dev)->gen > 3)
541 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH));
543 /* We need to disable the AsyncFlip performance optimisations in order
544 * to use MI_WAIT_FOR_EVENT within the CS. It should already be
545 * programmed to '1' on all products.
547 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv
549 if (INTEL_INFO(dev)->gen >= 6)
550 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
552 /* Required for the hardware to program scanline values for waiting */
553 if (INTEL_INFO(dev)->gen == 6)
555 _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_ALWAYS));
558 I915_WRITE(GFX_MODE_GEN7,
559 _MASKED_BIT_DISABLE(GFX_TLB_INVALIDATE_ALWAYS) |
560 _MASKED_BIT_ENABLE(GFX_REPLAY_MODE));
562 if (INTEL_INFO(dev)->gen >= 5) {
563 ret = init_pipe_control(ring);
569 /* From the Sandybridge PRM, volume 1 part 3, page 24:
570 * "If this bit is set, STCunit will have LRA as replacement
571 * policy. [...] This bit must be reset. LRA replacement
572 * policy is not supported."
574 I915_WRITE(CACHE_MODE_0,
575 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
577 /* This is not explicitly set for GEN6, so read the register.
578 * see intel_ring_mi_set_context() for why we care.
579 * TODO: consider explicitly setting the bit for GEN5
581 ring->itlb_before_ctx_switch =
582 !!(I915_READ(GFX_MODE) & GFX_TLB_INVALIDATE_ALWAYS);
585 if (INTEL_INFO(dev)->gen >= 6)
586 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
588 if (HAS_L3_GPU_CACHE(dev))
589 I915_WRITE_IMR(ring, ~GT_RENDER_L3_PARITY_ERROR_INTERRUPT);
594 static void render_ring_cleanup(struct intel_ring_buffer *ring)
596 struct drm_device *dev = ring->dev;
601 if (HAS_BROKEN_CS_TLB(dev))
602 drm_gem_object_unreference(to_gem_object(ring->private));
604 cleanup_pipe_control(ring);
608 update_mboxes(struct intel_ring_buffer *ring,
611 /* NB: In order to be able to do semaphore MBOX updates for varying number
612 * of rings, it's easiest if we round up each individual update to a
613 * multiple of 2 (since ring updates must always be a multiple of 2)
614 * even though the actual update only requires 3 dwords.
616 #define MBOX_UPDATE_DWORDS 4
617 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
618 intel_ring_emit(ring, mmio_offset);
619 intel_ring_emit(ring, ring->outstanding_lazy_request);
620 intel_ring_emit(ring, MI_NOOP);
624 * gen6_add_request - Update the semaphore mailbox registers
626 * @ring - ring that is adding a request
627 * @seqno - return seqno stuck into the ring
629 * Update the mailbox registers in the *other* rings with the current seqno.
630 * This acts like a signal in the canonical semaphore.
633 gen6_add_request(struct intel_ring_buffer *ring)
635 struct drm_device *dev = ring->dev;
636 struct drm_i915_private *dev_priv = dev->dev_private;
637 struct intel_ring_buffer *useless;
640 ret = intel_ring_begin(ring, ((I915_NUM_RINGS-1) *
641 MBOX_UPDATE_DWORDS) +
645 #undef MBOX_UPDATE_DWORDS
647 for_each_ring(useless, dev_priv, i) {
648 u32 mbox_reg = ring->signal_mbox[i];
649 if (mbox_reg != GEN6_NOSYNC)
650 update_mboxes(ring, mbox_reg);
653 intel_ring_emit(ring, MI_STORE_DWORD_INDEX);
654 intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
655 intel_ring_emit(ring, ring->outstanding_lazy_request);
656 intel_ring_emit(ring, MI_USER_INTERRUPT);
657 intel_ring_advance(ring);
662 static inline bool i915_gem_has_seqno_wrapped(struct drm_device *dev,
665 struct drm_i915_private *dev_priv = dev->dev_private;
666 return dev_priv->last_seqno < seqno;
670 * intel_ring_sync - sync the waiter to the signaller on seqno
672 * @waiter - ring that is waiting
673 * @signaller - ring which has, or will signal
674 * @seqno - seqno which the waiter will block on
677 gen6_ring_sync(struct intel_ring_buffer *waiter,
678 struct intel_ring_buffer *signaller,
682 u32 dw1 = MI_SEMAPHORE_MBOX |
683 MI_SEMAPHORE_COMPARE |
684 MI_SEMAPHORE_REGISTER;
686 /* Throughout all of the GEM code, seqno passed implies our current
687 * seqno is >= the last seqno executed. However for hardware the
688 * comparison is strictly greater than.
692 WARN_ON(signaller->semaphore_register[waiter->id] ==
693 MI_SEMAPHORE_SYNC_INVALID);
695 ret = intel_ring_begin(waiter, 4);
699 /* If seqno wrap happened, omit the wait with no-ops */
700 if (likely(!i915_gem_has_seqno_wrapped(waiter->dev, seqno))) {
701 intel_ring_emit(waiter,
703 signaller->semaphore_register[waiter->id]);
704 intel_ring_emit(waiter, seqno);
705 intel_ring_emit(waiter, 0);
706 intel_ring_emit(waiter, MI_NOOP);
708 intel_ring_emit(waiter, MI_NOOP);
709 intel_ring_emit(waiter, MI_NOOP);
710 intel_ring_emit(waiter, MI_NOOP);
711 intel_ring_emit(waiter, MI_NOOP);
713 intel_ring_advance(waiter);
718 #define PIPE_CONTROL_FLUSH(ring__, addr__) \
720 intel_ring_emit(ring__, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE | \
721 PIPE_CONTROL_DEPTH_STALL); \
722 intel_ring_emit(ring__, (addr__) | PIPE_CONTROL_GLOBAL_GTT); \
723 intel_ring_emit(ring__, 0); \
724 intel_ring_emit(ring__, 0); \
728 pc_render_add_request(struct intel_ring_buffer *ring)
730 struct pipe_control *pc = ring->private;
731 u32 scratch_addr = pc->gtt_offset + 128;
734 /* For Ironlake, MI_USER_INTERRUPT was deprecated and apparently
735 * incoherent with writes to memory, i.e. completely fubar,
736 * so we need to use PIPE_NOTIFY instead.
738 * However, we also need to workaround the qword write
739 * incoherence by flushing the 6 PIPE_NOTIFY buffers out to
740 * memory before requesting an interrupt.
742 ret = intel_ring_begin(ring, 32);
746 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE |
747 PIPE_CONTROL_WRITE_FLUSH |
748 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE);
749 intel_ring_emit(ring, pc->gtt_offset | PIPE_CONTROL_GLOBAL_GTT);
750 intel_ring_emit(ring, ring->outstanding_lazy_request);
751 intel_ring_emit(ring, 0);
752 PIPE_CONTROL_FLUSH(ring, scratch_addr);
753 scratch_addr += 128; /* write to separate cachelines */
754 PIPE_CONTROL_FLUSH(ring, scratch_addr);
756 PIPE_CONTROL_FLUSH(ring, scratch_addr);
758 PIPE_CONTROL_FLUSH(ring, scratch_addr);
760 PIPE_CONTROL_FLUSH(ring, scratch_addr);
762 PIPE_CONTROL_FLUSH(ring, scratch_addr);
764 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE |
765 PIPE_CONTROL_WRITE_FLUSH |
766 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
767 PIPE_CONTROL_NOTIFY);
768 intel_ring_emit(ring, pc->gtt_offset | PIPE_CONTROL_GLOBAL_GTT);
769 intel_ring_emit(ring, ring->outstanding_lazy_request);
770 intel_ring_emit(ring, 0);
771 intel_ring_advance(ring);
777 gen6_ring_get_seqno(struct intel_ring_buffer *ring, bool lazy_coherency)
779 /* Workaround to force correct ordering between irq and seqno writes on
780 * ivb (and maybe also on snb) by reading from a CS register (like
781 * ACTHD) before reading the status page. */
783 intel_ring_get_active_head(ring);
784 return intel_read_status_page(ring, I915_GEM_HWS_INDEX);
788 ring_get_seqno(struct intel_ring_buffer *ring, bool lazy_coherency)
790 return intel_read_status_page(ring, I915_GEM_HWS_INDEX);
794 ring_set_seqno(struct intel_ring_buffer *ring, u32 seqno)
796 intel_write_status_page(ring, I915_GEM_HWS_INDEX, seqno);
800 pc_render_get_seqno(struct intel_ring_buffer *ring, bool lazy_coherency)
802 struct pipe_control *pc = ring->private;
803 return pc->cpu_page[0];
807 pc_render_set_seqno(struct intel_ring_buffer *ring, u32 seqno)
809 struct pipe_control *pc = ring->private;
810 pc->cpu_page[0] = seqno;
814 gen5_ring_get_irq(struct intel_ring_buffer *ring)
816 struct drm_device *dev = ring->dev;
817 drm_i915_private_t *dev_priv = dev->dev_private;
820 if (!dev->irq_enabled)
823 spin_lock_irqsave(&dev_priv->irq_lock, flags);
824 if (ring->irq_refcount.gt++ == 0) {
825 dev_priv->gt_irq_mask &= ~ring->irq_enable_mask;
826 I915_WRITE(GTIMR, dev_priv->gt_irq_mask);
829 spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
835 gen5_ring_put_irq(struct intel_ring_buffer *ring)
837 struct drm_device *dev = ring->dev;
838 drm_i915_private_t *dev_priv = dev->dev_private;
841 spin_lock_irqsave(&dev_priv->irq_lock, flags);
842 if (--ring->irq_refcount.gt == 0) {
843 dev_priv->gt_irq_mask |= ring->irq_enable_mask;
844 I915_WRITE(GTIMR, dev_priv->gt_irq_mask);
847 spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
851 i9xx_ring_get_irq(struct intel_ring_buffer *ring)
853 struct drm_device *dev = ring->dev;
854 drm_i915_private_t *dev_priv = dev->dev_private;
857 if (!dev->irq_enabled)
860 spin_lock_irqsave(&dev_priv->irq_lock, flags);
861 if (ring->irq_refcount.gt++ == 0) {
862 dev_priv->irq_mask &= ~ring->irq_enable_mask;
863 I915_WRITE(IMR, dev_priv->irq_mask);
866 spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
872 i9xx_ring_put_irq(struct intel_ring_buffer *ring)
874 struct drm_device *dev = ring->dev;
875 drm_i915_private_t *dev_priv = dev->dev_private;
878 spin_lock_irqsave(&dev_priv->irq_lock, flags);
879 if (--ring->irq_refcount.gt == 0) {
880 dev_priv->irq_mask |= ring->irq_enable_mask;
881 I915_WRITE(IMR, dev_priv->irq_mask);
884 spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
888 i8xx_ring_get_irq(struct intel_ring_buffer *ring)
890 struct drm_device *dev = ring->dev;
891 drm_i915_private_t *dev_priv = dev->dev_private;
894 if (!dev->irq_enabled)
897 spin_lock_irqsave(&dev_priv->irq_lock, flags);
898 if (ring->irq_refcount.gt++ == 0) {
899 dev_priv->irq_mask &= ~ring->irq_enable_mask;
900 I915_WRITE16(IMR, dev_priv->irq_mask);
903 spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
909 i8xx_ring_put_irq(struct intel_ring_buffer *ring)
911 struct drm_device *dev = ring->dev;
912 drm_i915_private_t *dev_priv = dev->dev_private;
915 spin_lock_irqsave(&dev_priv->irq_lock, flags);
916 if (--ring->irq_refcount.gt == 0) {
917 dev_priv->irq_mask |= ring->irq_enable_mask;
918 I915_WRITE16(IMR, dev_priv->irq_mask);
921 spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
924 void intel_ring_setup_status_page(struct intel_ring_buffer *ring)
926 struct drm_device *dev = ring->dev;
927 drm_i915_private_t *dev_priv = ring->dev->dev_private;
930 /* The ring status page addresses are no longer next to the rest of
931 * the ring registers as of gen7.
936 mmio = RENDER_HWS_PGA_GEN7;
939 mmio = BLT_HWS_PGA_GEN7;
942 mmio = BSD_HWS_PGA_GEN7;
945 mmio = VEBOX_HWS_PGA_GEN7;
948 } else if (IS_GEN6(ring->dev)) {
949 mmio = RING_HWS_PGA_GEN6(ring->mmio_base);
951 mmio = RING_HWS_PGA(ring->mmio_base);
954 I915_WRITE(mmio, (u32)ring->status_page.gfx_addr);
959 bsd_ring_flush(struct intel_ring_buffer *ring,
960 u32 invalidate_domains,
965 ret = intel_ring_begin(ring, 2);
969 intel_ring_emit(ring, MI_FLUSH);
970 intel_ring_emit(ring, MI_NOOP);
971 intel_ring_advance(ring);
976 i9xx_add_request(struct intel_ring_buffer *ring)
980 ret = intel_ring_begin(ring, 4);
984 intel_ring_emit(ring, MI_STORE_DWORD_INDEX);
985 intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
986 intel_ring_emit(ring, ring->outstanding_lazy_request);
987 intel_ring_emit(ring, MI_USER_INTERRUPT);
988 intel_ring_advance(ring);
994 gen6_ring_get_irq(struct intel_ring_buffer *ring)
996 struct drm_device *dev = ring->dev;
997 drm_i915_private_t *dev_priv = dev->dev_private;
1000 if (!dev->irq_enabled)
1003 /* It looks like we need to prevent the gt from suspending while waiting
1004 * for an notifiy irq, otherwise irqs seem to get lost on at least the
1005 * blt/bsd rings on ivb. */
1006 gen6_gt_force_wake_get(dev_priv);
1008 spin_lock_irqsave(&dev_priv->irq_lock, flags);
1009 if (ring->irq_refcount.gt++ == 0) {
1010 if (HAS_L3_GPU_CACHE(dev) && ring->id == RCS)
1011 I915_WRITE_IMR(ring,
1012 ~(ring->irq_enable_mask |
1013 GT_RENDER_L3_PARITY_ERROR_INTERRUPT));
1015 I915_WRITE_IMR(ring, ~ring->irq_enable_mask);
1016 dev_priv->gt_irq_mask &= ~ring->irq_enable_mask;
1017 I915_WRITE(GTIMR, dev_priv->gt_irq_mask);
1018 POSTING_READ(GTIMR);
1020 spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1026 gen6_ring_put_irq(struct intel_ring_buffer *ring)
1028 struct drm_device *dev = ring->dev;
1029 drm_i915_private_t *dev_priv = dev->dev_private;
1030 unsigned long flags;
1032 spin_lock_irqsave(&dev_priv->irq_lock, flags);
1033 if (--ring->irq_refcount.gt == 0) {
1034 if (HAS_L3_GPU_CACHE(dev) && ring->id == RCS)
1035 I915_WRITE_IMR(ring,
1036 ~GT_RENDER_L3_PARITY_ERROR_INTERRUPT);
1038 I915_WRITE_IMR(ring, ~0);
1039 dev_priv->gt_irq_mask |= ring->irq_enable_mask;
1040 I915_WRITE(GTIMR, dev_priv->gt_irq_mask);
1041 POSTING_READ(GTIMR);
1043 spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1045 gen6_gt_force_wake_put(dev_priv);
1049 hsw_vebox_get_irq(struct intel_ring_buffer *ring)
1051 struct drm_device *dev = ring->dev;
1052 struct drm_i915_private *dev_priv = dev->dev_private;
1053 unsigned long flags;
1055 if (!dev->irq_enabled)
1058 spin_lock_irqsave(&dev_priv->rps.lock, flags);
1059 if (ring->irq_refcount.pm++ == 0) {
1060 u32 pm_imr = I915_READ(GEN6_PMIMR);
1061 I915_WRITE_IMR(ring, ~ring->irq_enable_mask);
1062 I915_WRITE(GEN6_PMIMR, pm_imr & ~ring->irq_enable_mask);
1063 POSTING_READ(GEN6_PMIMR);
1065 spin_unlock_irqrestore(&dev_priv->rps.lock, flags);
1071 hsw_vebox_put_irq(struct intel_ring_buffer *ring)
1073 struct drm_device *dev = ring->dev;
1074 struct drm_i915_private *dev_priv = dev->dev_private;
1075 unsigned long flags;
1077 if (!dev->irq_enabled)
1080 spin_lock_irqsave(&dev_priv->rps.lock, flags);
1081 if (--ring->irq_refcount.pm == 0) {
1082 u32 pm_imr = I915_READ(GEN6_PMIMR);
1083 I915_WRITE_IMR(ring, ~0);
1084 I915_WRITE(GEN6_PMIMR, pm_imr | ring->irq_enable_mask);
1085 POSTING_READ(GEN6_PMIMR);
1087 spin_unlock_irqrestore(&dev_priv->rps.lock, flags);
1091 i965_dispatch_execbuffer(struct intel_ring_buffer *ring,
1092 u32 offset, u32 length,
1097 ret = intel_ring_begin(ring, 2);
1101 intel_ring_emit(ring,
1102 MI_BATCH_BUFFER_START |
1104 (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE_I965));
1105 intel_ring_emit(ring, offset);
1106 intel_ring_advance(ring);
1111 /* Just userspace ABI convention to limit the wa batch bo to a resonable size */
1112 #define I830_BATCH_LIMIT (256*1024)
1114 i830_dispatch_execbuffer(struct intel_ring_buffer *ring,
1115 u32 offset, u32 len,
1120 if (flags & I915_DISPATCH_PINNED) {
1121 ret = intel_ring_begin(ring, 4);
1125 intel_ring_emit(ring, MI_BATCH_BUFFER);
1126 intel_ring_emit(ring, offset | (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE));
1127 intel_ring_emit(ring, offset + len - 8);
1128 intel_ring_emit(ring, MI_NOOP);
1129 intel_ring_advance(ring);
1131 struct drm_i915_gem_object *obj = ring->private;
1132 u32 cs_offset = obj->gtt_offset;
1134 if (len > I830_BATCH_LIMIT)
1137 ret = intel_ring_begin(ring, 9+3);
1140 /* Blit the batch (which has now all relocs applied) to the stable batch
1141 * scratch bo area (so that the CS never stumbles over its tlb
1142 * invalidation bug) ... */
1143 intel_ring_emit(ring, XY_SRC_COPY_BLT_CMD |
1144 XY_SRC_COPY_BLT_WRITE_ALPHA |
1145 XY_SRC_COPY_BLT_WRITE_RGB);
1146 intel_ring_emit(ring, BLT_DEPTH_32 | BLT_ROP_GXCOPY | 4096);
1147 intel_ring_emit(ring, 0);
1148 intel_ring_emit(ring, (DIV_ROUND_UP(len, 4096) << 16) | 1024);
1149 intel_ring_emit(ring, cs_offset);
1150 intel_ring_emit(ring, 0);
1151 intel_ring_emit(ring, 4096);
1152 intel_ring_emit(ring, offset);
1153 intel_ring_emit(ring, MI_FLUSH);
1155 /* ... and execute it. */
1156 intel_ring_emit(ring, MI_BATCH_BUFFER);
1157 intel_ring_emit(ring, cs_offset | (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE));
1158 intel_ring_emit(ring, cs_offset + len - 8);
1159 intel_ring_advance(ring);
1166 i915_dispatch_execbuffer(struct intel_ring_buffer *ring,
1167 u32 offset, u32 len,
1172 ret = intel_ring_begin(ring, 2);
1176 intel_ring_emit(ring, MI_BATCH_BUFFER_START | MI_BATCH_GTT);
1177 intel_ring_emit(ring, offset | (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE));
1178 intel_ring_advance(ring);
1183 static void cleanup_status_page(struct intel_ring_buffer *ring)
1185 struct drm_i915_gem_object *obj;
1187 obj = ring->status_page.obj;
1191 kunmap(sg_page(obj->pages->sgl));
1192 i915_gem_object_unpin(obj);
1193 drm_gem_object_unreference(&obj->base);
1194 ring->status_page.obj = NULL;
1197 static int init_status_page(struct intel_ring_buffer *ring)
1199 struct drm_device *dev = ring->dev;
1200 struct drm_i915_gem_object *obj;
1203 obj = i915_gem_alloc_object(dev, 4096);
1205 DRM_ERROR("Failed to allocate status page\n");
1210 i915_gem_object_set_cache_level(obj, I915_CACHE_LLC);
1212 ret = i915_gem_object_pin(obj, 4096, true, false);
1217 ring->status_page.gfx_addr = obj->gtt_offset;
1218 ring->status_page.page_addr = kmap(sg_page(obj->pages->sgl));
1219 if (ring->status_page.page_addr == NULL) {
1223 ring->status_page.obj = obj;
1224 memset(ring->status_page.page_addr, 0, PAGE_SIZE);
1226 intel_ring_setup_status_page(ring);
1227 DRM_DEBUG_DRIVER("%s hws offset: 0x%08x\n",
1228 ring->name, ring->status_page.gfx_addr);
1233 i915_gem_object_unpin(obj);
1235 drm_gem_object_unreference(&obj->base);
1240 static int init_phys_hws_pga(struct intel_ring_buffer *ring)
1242 struct drm_i915_private *dev_priv = ring->dev->dev_private;
1245 if (!dev_priv->status_page_dmah) {
1246 dev_priv->status_page_dmah =
1247 drm_pci_alloc(ring->dev, PAGE_SIZE, PAGE_SIZE);
1248 if (!dev_priv->status_page_dmah)
1252 addr = dev_priv->status_page_dmah->busaddr;
1253 if (INTEL_INFO(ring->dev)->gen >= 4)
1254 addr |= (dev_priv->status_page_dmah->busaddr >> 28) & 0xf0;
1255 I915_WRITE(HWS_PGA, addr);
1257 ring->status_page.page_addr = dev_priv->status_page_dmah->vaddr;
1258 memset(ring->status_page.page_addr, 0, PAGE_SIZE);
1263 static int intel_init_ring_buffer(struct drm_device *dev,
1264 struct intel_ring_buffer *ring)
1266 struct drm_i915_gem_object *obj;
1267 struct drm_i915_private *dev_priv = dev->dev_private;
1271 INIT_LIST_HEAD(&ring->active_list);
1272 INIT_LIST_HEAD(&ring->request_list);
1273 ring->size = 32 * PAGE_SIZE;
1274 memset(ring->sync_seqno, 0, sizeof(ring->sync_seqno));
1276 init_waitqueue_head(&ring->irq_queue);
1278 if (I915_NEED_GFX_HWS(dev)) {
1279 ret = init_status_page(ring);
1283 BUG_ON(ring->id != RCS);
1284 ret = init_phys_hws_pga(ring);
1291 obj = i915_gem_object_create_stolen(dev, ring->size);
1293 obj = i915_gem_alloc_object(dev, ring->size);
1295 DRM_ERROR("Failed to allocate ringbuffer\n");
1302 ret = i915_gem_object_pin(obj, PAGE_SIZE, true, false);
1306 ret = i915_gem_object_set_to_gtt_domain(obj, true);
1310 ring->virtual_start =
1311 ioremap_wc(dev_priv->gtt.mappable_base + obj->gtt_offset,
1313 if (ring->virtual_start == NULL) {
1314 DRM_ERROR("Failed to map ringbuffer.\n");
1319 ret = ring->init(ring);
1323 /* Workaround an erratum on the i830 which causes a hang if
1324 * the TAIL pointer points to within the last 2 cachelines
1327 ring->effective_size = ring->size;
1328 if (IS_I830(ring->dev) || IS_845G(ring->dev))
1329 ring->effective_size -= 128;
1334 iounmap(ring->virtual_start);
1336 i915_gem_object_unpin(obj);
1338 drm_gem_object_unreference(&obj->base);
1341 cleanup_status_page(ring);
1345 void intel_cleanup_ring_buffer(struct intel_ring_buffer *ring)
1347 struct drm_i915_private *dev_priv;
1350 if (ring->obj == NULL)
1353 /* Disable the ring buffer. The ring must be idle at this point */
1354 dev_priv = ring->dev->dev_private;
1355 ret = intel_ring_idle(ring);
1357 DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",
1360 I915_WRITE_CTL(ring, 0);
1362 iounmap(ring->virtual_start);
1364 i915_gem_object_unpin(ring->obj);
1365 drm_gem_object_unreference(&ring->obj->base);
1369 ring->cleanup(ring);
1371 cleanup_status_page(ring);
1374 static int intel_ring_wait_seqno(struct intel_ring_buffer *ring, u32 seqno)
1378 ret = i915_wait_seqno(ring, seqno);
1380 i915_gem_retire_requests_ring(ring);
1385 static int intel_ring_wait_request(struct intel_ring_buffer *ring, int n)
1387 struct drm_i915_gem_request *request;
1391 i915_gem_retire_requests_ring(ring);
1393 if (ring->last_retired_head != -1) {
1394 ring->head = ring->last_retired_head;
1395 ring->last_retired_head = -1;
1396 ring->space = ring_space(ring);
1397 if (ring->space >= n)
1401 list_for_each_entry(request, &ring->request_list, list) {
1404 if (request->tail == -1)
1407 space = request->tail - (ring->tail + I915_RING_FREE_SPACE);
1409 space += ring->size;
1411 seqno = request->seqno;
1415 /* Consume this request in case we need more space than
1416 * is available and so need to prevent a race between
1417 * updating last_retired_head and direct reads of
1418 * I915_RING_HEAD. It also provides a nice sanity check.
1426 ret = intel_ring_wait_seqno(ring, seqno);
1430 if (WARN_ON(ring->last_retired_head == -1))
1433 ring->head = ring->last_retired_head;
1434 ring->last_retired_head = -1;
1435 ring->space = ring_space(ring);
1436 if (WARN_ON(ring->space < n))
1442 static int ring_wait_for_space(struct intel_ring_buffer *ring, int n)
1444 struct drm_device *dev = ring->dev;
1445 struct drm_i915_private *dev_priv = dev->dev_private;
1449 ret = intel_ring_wait_request(ring, n);
1453 trace_i915_ring_wait_begin(ring);
1454 /* With GEM the hangcheck timer should kick us out of the loop,
1455 * leaving it early runs the risk of corrupting GEM state (due
1456 * to running on almost untested codepaths). But on resume
1457 * timers don't work yet, so prevent a complete hang in that
1458 * case by choosing an insanely large timeout. */
1459 end = jiffies + 60 * HZ;
1462 ring->head = I915_READ_HEAD(ring);
1463 ring->space = ring_space(ring);
1464 if (ring->space >= n) {
1465 trace_i915_ring_wait_end(ring);
1469 if (dev->primary->master) {
1470 struct drm_i915_master_private *master_priv = dev->primary->master->driver_priv;
1471 if (master_priv->sarea_priv)
1472 master_priv->sarea_priv->perf_boxes |= I915_BOX_WAIT;
1477 ret = i915_gem_check_wedge(&dev_priv->gpu_error,
1478 dev_priv->mm.interruptible);
1481 } while (!time_after(jiffies, end));
1482 trace_i915_ring_wait_end(ring);
1486 static int intel_wrap_ring_buffer(struct intel_ring_buffer *ring)
1488 uint32_t __iomem *virt;
1489 int rem = ring->size - ring->tail;
1491 if (ring->space < rem) {
1492 int ret = ring_wait_for_space(ring, rem);
1497 virt = ring->virtual_start + ring->tail;
1500 iowrite32(MI_NOOP, virt++);
1503 ring->space = ring_space(ring);
1508 int intel_ring_idle(struct intel_ring_buffer *ring)
1513 /* We need to add any requests required to flush the objects and ring */
1514 if (ring->outstanding_lazy_request) {
1515 ret = i915_add_request(ring, NULL);
1520 /* Wait upon the last request to be completed */
1521 if (list_empty(&ring->request_list))
1524 seqno = list_entry(ring->request_list.prev,
1525 struct drm_i915_gem_request,
1528 return i915_wait_seqno(ring, seqno);
1532 intel_ring_alloc_seqno(struct intel_ring_buffer *ring)
1534 if (ring->outstanding_lazy_request)
1537 return i915_gem_get_seqno(ring->dev, &ring->outstanding_lazy_request);
1540 static int __intel_ring_begin(struct intel_ring_buffer *ring,
1545 if (unlikely(ring->tail + bytes > ring->effective_size)) {
1546 ret = intel_wrap_ring_buffer(ring);
1551 if (unlikely(ring->space < bytes)) {
1552 ret = ring_wait_for_space(ring, bytes);
1557 ring->space -= bytes;
1561 int intel_ring_begin(struct intel_ring_buffer *ring,
1564 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1567 ret = i915_gem_check_wedge(&dev_priv->gpu_error,
1568 dev_priv->mm.interruptible);
1572 /* Preallocate the olr before touching the ring */
1573 ret = intel_ring_alloc_seqno(ring);
1577 return __intel_ring_begin(ring, num_dwords * sizeof(uint32_t));
1580 void intel_ring_init_seqno(struct intel_ring_buffer *ring, u32 seqno)
1582 struct drm_i915_private *dev_priv = ring->dev->dev_private;
1584 BUG_ON(ring->outstanding_lazy_request);
1586 if (INTEL_INFO(ring->dev)->gen >= 6) {
1587 I915_WRITE(RING_SYNC_0(ring->mmio_base), 0);
1588 I915_WRITE(RING_SYNC_1(ring->mmio_base), 0);
1591 ring->set_seqno(ring, seqno);
1592 ring->hangcheck.seqno = seqno;
1595 void intel_ring_advance(struct intel_ring_buffer *ring)
1597 struct drm_i915_private *dev_priv = ring->dev->dev_private;
1599 ring->tail &= ring->size - 1;
1600 if (dev_priv->gpu_error.stop_rings & intel_ring_flag(ring))
1602 ring->write_tail(ring, ring->tail);
1606 static void gen6_bsd_ring_write_tail(struct intel_ring_buffer *ring,
1609 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1611 /* Every tail move must follow the sequence below */
1613 /* Disable notification that the ring is IDLE. The GT
1614 * will then assume that it is busy and bring it out of rc6.
1616 I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL,
1617 _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1619 /* Clear the context id. Here be magic! */
1620 I915_WRITE64(GEN6_BSD_RNCID, 0x0);
1622 /* Wait for the ring not to be idle, i.e. for it to wake up. */
1623 if (wait_for((I915_READ(GEN6_BSD_SLEEP_PSMI_CONTROL) &
1624 GEN6_BSD_SLEEP_INDICATOR) == 0,
1626 DRM_ERROR("timed out waiting for the BSD ring to wake up\n");
1628 /* Now that the ring is fully powered up, update the tail */
1629 I915_WRITE_TAIL(ring, value);
1630 POSTING_READ(RING_TAIL(ring->mmio_base));
1632 /* Let the ring send IDLE messages to the GT again,
1633 * and so let it sleep to conserve power when idle.
1635 I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL,
1636 _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1639 static int gen6_bsd_ring_flush(struct intel_ring_buffer *ring,
1640 u32 invalidate, u32 flush)
1645 ret = intel_ring_begin(ring, 4);
1651 * Bspec vol 1c.5 - video engine command streamer:
1652 * "If ENABLED, all TLBs will be invalidated once the flush
1653 * operation is complete. This bit is only valid when the
1654 * Post-Sync Operation field is a value of 1h or 3h."
1656 if (invalidate & I915_GEM_GPU_DOMAINS)
1657 cmd |= MI_INVALIDATE_TLB | MI_INVALIDATE_BSD |
1658 MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1659 intel_ring_emit(ring, cmd);
1660 intel_ring_emit(ring, I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT);
1661 intel_ring_emit(ring, 0);
1662 intel_ring_emit(ring, MI_NOOP);
1663 intel_ring_advance(ring);
1668 hsw_ring_dispatch_execbuffer(struct intel_ring_buffer *ring,
1669 u32 offset, u32 len,
1674 ret = intel_ring_begin(ring, 2);
1678 intel_ring_emit(ring,
1679 MI_BATCH_BUFFER_START | MI_BATCH_PPGTT_HSW |
1680 (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE_HSW));
1681 /* bit0-7 is the length on GEN6+ */
1682 intel_ring_emit(ring, offset);
1683 intel_ring_advance(ring);
1689 gen6_ring_dispatch_execbuffer(struct intel_ring_buffer *ring,
1690 u32 offset, u32 len,
1695 ret = intel_ring_begin(ring, 2);
1699 intel_ring_emit(ring,
1700 MI_BATCH_BUFFER_START |
1701 (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE_I965));
1702 /* bit0-7 is the length on GEN6+ */
1703 intel_ring_emit(ring, offset);
1704 intel_ring_advance(ring);
1709 /* Blitter support (SandyBridge+) */
1711 static int gen6_ring_flush(struct intel_ring_buffer *ring,
1712 u32 invalidate, u32 flush)
1714 struct drm_device *dev = ring->dev;
1718 ret = intel_ring_begin(ring, 4);
1724 * Bspec vol 1c.3 - blitter engine command streamer:
1725 * "If ENABLED, all TLBs will be invalidated once the flush
1726 * operation is complete. This bit is only valid when the
1727 * Post-Sync Operation field is a value of 1h or 3h."
1729 if (invalidate & I915_GEM_DOMAIN_RENDER)
1730 cmd |= MI_INVALIDATE_TLB | MI_FLUSH_DW_STORE_INDEX |
1731 MI_FLUSH_DW_OP_STOREDW;
1732 intel_ring_emit(ring, cmd);
1733 intel_ring_emit(ring, I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT);
1734 intel_ring_emit(ring, 0);
1735 intel_ring_emit(ring, MI_NOOP);
1736 intel_ring_advance(ring);
1738 if (IS_GEN7(dev) && flush)
1739 return gen7_ring_fbc_flush(ring, FBC_REND_CACHE_CLEAN);
1744 int intel_init_render_ring_buffer(struct drm_device *dev)
1746 drm_i915_private_t *dev_priv = dev->dev_private;
1747 struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
1749 ring->name = "render ring";
1751 ring->mmio_base = RENDER_RING_BASE;
1753 if (INTEL_INFO(dev)->gen >= 6) {
1754 ring->add_request = gen6_add_request;
1755 ring->flush = gen7_render_ring_flush;
1756 if (INTEL_INFO(dev)->gen == 6)
1757 ring->flush = gen6_render_ring_flush;
1758 ring->irq_get = gen6_ring_get_irq;
1759 ring->irq_put = gen6_ring_put_irq;
1760 ring->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
1761 ring->get_seqno = gen6_ring_get_seqno;
1762 ring->set_seqno = ring_set_seqno;
1763 ring->sync_to = gen6_ring_sync;
1764 ring->semaphore_register[RCS] = MI_SEMAPHORE_SYNC_INVALID;
1765 ring->semaphore_register[VCS] = MI_SEMAPHORE_SYNC_RV;
1766 ring->semaphore_register[BCS] = MI_SEMAPHORE_SYNC_RB;
1767 ring->semaphore_register[VECS] = MI_SEMAPHORE_SYNC_RVE;
1768 ring->signal_mbox[RCS] = GEN6_NOSYNC;
1769 ring->signal_mbox[VCS] = GEN6_VRSYNC;
1770 ring->signal_mbox[BCS] = GEN6_BRSYNC;
1771 ring->signal_mbox[VECS] = GEN6_VERSYNC;
1772 } else if (IS_GEN5(dev)) {
1773 ring->add_request = pc_render_add_request;
1774 ring->flush = gen4_render_ring_flush;
1775 ring->get_seqno = pc_render_get_seqno;
1776 ring->set_seqno = pc_render_set_seqno;
1777 ring->irq_get = gen5_ring_get_irq;
1778 ring->irq_put = gen5_ring_put_irq;
1779 ring->irq_enable_mask = GT_RENDER_USER_INTERRUPT |
1780 GT_RENDER_PIPECTL_NOTIFY_INTERRUPT;
1782 ring->add_request = i9xx_add_request;
1783 if (INTEL_INFO(dev)->gen < 4)
1784 ring->flush = gen2_render_ring_flush;
1786 ring->flush = gen4_render_ring_flush;
1787 ring->get_seqno = ring_get_seqno;
1788 ring->set_seqno = ring_set_seqno;
1790 ring->irq_get = i8xx_ring_get_irq;
1791 ring->irq_put = i8xx_ring_put_irq;
1793 ring->irq_get = i9xx_ring_get_irq;
1794 ring->irq_put = i9xx_ring_put_irq;
1796 ring->irq_enable_mask = I915_USER_INTERRUPT;
1798 ring->write_tail = ring_write_tail;
1799 if (IS_HASWELL(dev))
1800 ring->dispatch_execbuffer = hsw_ring_dispatch_execbuffer;
1801 else if (INTEL_INFO(dev)->gen >= 6)
1802 ring->dispatch_execbuffer = gen6_ring_dispatch_execbuffer;
1803 else if (INTEL_INFO(dev)->gen >= 4)
1804 ring->dispatch_execbuffer = i965_dispatch_execbuffer;
1805 else if (IS_I830(dev) || IS_845G(dev))
1806 ring->dispatch_execbuffer = i830_dispatch_execbuffer;
1808 ring->dispatch_execbuffer = i915_dispatch_execbuffer;
1809 ring->init = init_render_ring;
1810 ring->cleanup = render_ring_cleanup;
1812 /* Workaround batchbuffer to combat CS tlb bug. */
1813 if (HAS_BROKEN_CS_TLB(dev)) {
1814 struct drm_i915_gem_object *obj;
1817 obj = i915_gem_alloc_object(dev, I830_BATCH_LIMIT);
1819 DRM_ERROR("Failed to allocate batch bo\n");
1823 ret = i915_gem_object_pin(obj, 0, true, false);
1825 drm_gem_object_unreference(&obj->base);
1826 DRM_ERROR("Failed to ping batch bo\n");
1830 ring->private = obj;
1833 return intel_init_ring_buffer(dev, ring);
1836 int intel_render_ring_init_dri(struct drm_device *dev, u64 start, u32 size)
1838 drm_i915_private_t *dev_priv = dev->dev_private;
1839 struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
1842 ring->name = "render ring";
1844 ring->mmio_base = RENDER_RING_BASE;
1846 if (INTEL_INFO(dev)->gen >= 6) {
1847 /* non-kms not supported on gen6+ */
1851 /* Note: gem is not supported on gen5/ilk without kms (the corresponding
1852 * gem_init ioctl returns with -ENODEV). Hence we do not need to set up
1853 * the special gen5 functions. */
1854 ring->add_request = i9xx_add_request;
1855 if (INTEL_INFO(dev)->gen < 4)
1856 ring->flush = gen2_render_ring_flush;
1858 ring->flush = gen4_render_ring_flush;
1859 ring->get_seqno = ring_get_seqno;
1860 ring->set_seqno = ring_set_seqno;
1862 ring->irq_get = i8xx_ring_get_irq;
1863 ring->irq_put = i8xx_ring_put_irq;
1865 ring->irq_get = i9xx_ring_get_irq;
1866 ring->irq_put = i9xx_ring_put_irq;
1868 ring->irq_enable_mask = I915_USER_INTERRUPT;
1869 ring->write_tail = ring_write_tail;
1870 if (INTEL_INFO(dev)->gen >= 4)
1871 ring->dispatch_execbuffer = i965_dispatch_execbuffer;
1872 else if (IS_I830(dev) || IS_845G(dev))
1873 ring->dispatch_execbuffer = i830_dispatch_execbuffer;
1875 ring->dispatch_execbuffer = i915_dispatch_execbuffer;
1876 ring->init = init_render_ring;
1877 ring->cleanup = render_ring_cleanup;
1880 INIT_LIST_HEAD(&ring->active_list);
1881 INIT_LIST_HEAD(&ring->request_list);
1884 ring->effective_size = ring->size;
1885 if (IS_I830(ring->dev) || IS_845G(ring->dev))
1886 ring->effective_size -= 128;
1888 ring->virtual_start = ioremap_wc(start, size);
1889 if (ring->virtual_start == NULL) {
1890 DRM_ERROR("can not ioremap virtual address for"
1895 if (!I915_NEED_GFX_HWS(dev)) {
1896 ret = init_phys_hws_pga(ring);
1904 int intel_init_bsd_ring_buffer(struct drm_device *dev)
1906 drm_i915_private_t *dev_priv = dev->dev_private;
1907 struct intel_ring_buffer *ring = &dev_priv->ring[VCS];
1909 ring->name = "bsd ring";
1912 ring->write_tail = ring_write_tail;
1913 if (IS_GEN6(dev) || IS_GEN7(dev)) {
1914 ring->mmio_base = GEN6_BSD_RING_BASE;
1915 /* gen6 bsd needs a special wa for tail updates */
1917 ring->write_tail = gen6_bsd_ring_write_tail;
1918 ring->flush = gen6_bsd_ring_flush;
1919 ring->add_request = gen6_add_request;
1920 ring->get_seqno = gen6_ring_get_seqno;
1921 ring->set_seqno = ring_set_seqno;
1922 ring->irq_enable_mask = GT_BSD_USER_INTERRUPT;
1923 ring->irq_get = gen6_ring_get_irq;
1924 ring->irq_put = gen6_ring_put_irq;
1925 ring->dispatch_execbuffer = gen6_ring_dispatch_execbuffer;
1926 ring->sync_to = gen6_ring_sync;
1927 ring->semaphore_register[RCS] = MI_SEMAPHORE_SYNC_VR;
1928 ring->semaphore_register[VCS] = MI_SEMAPHORE_SYNC_INVALID;
1929 ring->semaphore_register[BCS] = MI_SEMAPHORE_SYNC_VB;
1930 ring->semaphore_register[VECS] = MI_SEMAPHORE_SYNC_VVE;
1931 ring->signal_mbox[RCS] = GEN6_RVSYNC;
1932 ring->signal_mbox[VCS] = GEN6_NOSYNC;
1933 ring->signal_mbox[BCS] = GEN6_BVSYNC;
1934 ring->signal_mbox[VECS] = GEN6_VEVSYNC;
1936 ring->mmio_base = BSD_RING_BASE;
1937 ring->flush = bsd_ring_flush;
1938 ring->add_request = i9xx_add_request;
1939 ring->get_seqno = ring_get_seqno;
1940 ring->set_seqno = ring_set_seqno;
1942 ring->irq_enable_mask = ILK_BSD_USER_INTERRUPT;
1943 ring->irq_get = gen5_ring_get_irq;
1944 ring->irq_put = gen5_ring_put_irq;
1946 ring->irq_enable_mask = I915_BSD_USER_INTERRUPT;
1947 ring->irq_get = i9xx_ring_get_irq;
1948 ring->irq_put = i9xx_ring_put_irq;
1950 ring->dispatch_execbuffer = i965_dispatch_execbuffer;
1952 ring->init = init_ring_common;
1954 return intel_init_ring_buffer(dev, ring);
1957 int intel_init_blt_ring_buffer(struct drm_device *dev)
1959 drm_i915_private_t *dev_priv = dev->dev_private;
1960 struct intel_ring_buffer *ring = &dev_priv->ring[BCS];
1962 ring->name = "blitter ring";
1965 ring->mmio_base = BLT_RING_BASE;
1966 ring->write_tail = ring_write_tail;
1967 ring->flush = gen6_ring_flush;
1968 ring->add_request = gen6_add_request;
1969 ring->get_seqno = gen6_ring_get_seqno;
1970 ring->set_seqno = ring_set_seqno;
1971 ring->irq_enable_mask = GT_BLT_USER_INTERRUPT;
1972 ring->irq_get = gen6_ring_get_irq;
1973 ring->irq_put = gen6_ring_put_irq;
1974 ring->dispatch_execbuffer = gen6_ring_dispatch_execbuffer;
1975 ring->sync_to = gen6_ring_sync;
1976 ring->semaphore_register[RCS] = MI_SEMAPHORE_SYNC_BR;
1977 ring->semaphore_register[VCS] = MI_SEMAPHORE_SYNC_BV;
1978 ring->semaphore_register[BCS] = MI_SEMAPHORE_SYNC_INVALID;
1979 ring->semaphore_register[VECS] = MI_SEMAPHORE_SYNC_BVE;
1980 ring->signal_mbox[RCS] = GEN6_RBSYNC;
1981 ring->signal_mbox[VCS] = GEN6_VBSYNC;
1982 ring->signal_mbox[BCS] = GEN6_NOSYNC;
1983 ring->signal_mbox[VECS] = GEN6_VEBSYNC;
1984 ring->init = init_ring_common;
1986 return intel_init_ring_buffer(dev, ring);
1989 int intel_init_vebox_ring_buffer(struct drm_device *dev)
1991 drm_i915_private_t *dev_priv = dev->dev_private;
1992 struct intel_ring_buffer *ring = &dev_priv->ring[VECS];
1994 ring->name = "video enhancement ring";
1997 ring->mmio_base = VEBOX_RING_BASE;
1998 ring->write_tail = ring_write_tail;
1999 ring->flush = gen6_ring_flush;
2000 ring->add_request = gen6_add_request;
2001 ring->get_seqno = gen6_ring_get_seqno;
2002 ring->set_seqno = ring_set_seqno;
2003 ring->irq_enable_mask = PM_VEBOX_USER_INTERRUPT |
2004 PM_VEBOX_CS_ERROR_INTERRUPT;
2005 ring->irq_get = hsw_vebox_get_irq;
2006 ring->irq_put = hsw_vebox_put_irq;
2007 ring->dispatch_execbuffer = gen6_ring_dispatch_execbuffer;
2008 ring->sync_to = gen6_ring_sync;
2009 ring->semaphore_register[RCS] = MI_SEMAPHORE_SYNC_VER;
2010 ring->semaphore_register[VCS] = MI_SEMAPHORE_SYNC_VEV;
2011 ring->semaphore_register[BCS] = MI_SEMAPHORE_SYNC_VEB;
2012 ring->semaphore_register[VECS] = MI_SEMAPHORE_SYNC_INVALID;
2013 ring->signal_mbox[RCS] = GEN6_RVESYNC;
2014 ring->signal_mbox[VCS] = GEN6_VVESYNC;
2015 ring->signal_mbox[BCS] = GEN6_BVESYNC;
2016 ring->signal_mbox[VECS] = GEN6_NOSYNC;
2017 ring->init = init_ring_common;
2019 return intel_init_ring_buffer(dev, ring);
2023 intel_ring_flush_all_caches(struct intel_ring_buffer *ring)
2027 if (!ring->gpu_caches_dirty)
2030 ret = ring->flush(ring, 0, I915_GEM_GPU_DOMAINS);
2034 trace_i915_gem_ring_flush(ring, 0, I915_GEM_GPU_DOMAINS);
2036 ring->gpu_caches_dirty = false;
2041 intel_ring_invalidate_all_caches(struct intel_ring_buffer *ring)
2043 uint32_t flush_domains;
2047 if (ring->gpu_caches_dirty)
2048 flush_domains = I915_GEM_GPU_DOMAINS;
2050 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, flush_domains);
2054 trace_i915_gem_ring_flush(ring, I915_GEM_GPU_DOMAINS, flush_domains);
2056 ring->gpu_caches_dirty = false;