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>
34 #include "i915_trace.h"
35 #include "intel_drv.h"
38 * 965+ support PIPE_CONTROL commands, which provide finer grained control
39 * over cache flushing.
42 struct drm_i915_gem_object *obj;
43 volatile u32 *cpu_page;
47 static inline int ring_space(struct intel_ring_buffer *ring)
49 int space = (ring->head & HEAD_ADDR) - (ring->tail + 8);
55 static u32 i915_gem_get_seqno(struct drm_device *dev)
57 drm_i915_private_t *dev_priv = dev->dev_private;
60 seqno = dev_priv->next_seqno;
62 /* reserve 0 for non-seqno */
63 if (++dev_priv->next_seqno == 0)
64 dev_priv->next_seqno = 1;
70 render_ring_flush(struct intel_ring_buffer *ring,
71 u32 invalidate_domains,
74 struct drm_device *dev = ring->dev;
81 * I915_GEM_DOMAIN_RENDER is always invalidated, but is
82 * only flushed if MI_NO_WRITE_FLUSH is unset. On 965, it is
83 * also flushed at 2d versus 3d pipeline switches.
87 * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
88 * MI_READ_FLUSH is set, and is always flushed on 965.
90 * I915_GEM_DOMAIN_COMMAND may not exist?
92 * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
93 * invalidated when MI_EXE_FLUSH is set.
95 * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
96 * invalidated with every MI_FLUSH.
100 * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
101 * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
102 * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
103 * are flushed at any MI_FLUSH.
106 cmd = MI_FLUSH | MI_NO_WRITE_FLUSH;
107 if ((invalidate_domains|flush_domains) &
108 I915_GEM_DOMAIN_RENDER)
109 cmd &= ~MI_NO_WRITE_FLUSH;
110 if (INTEL_INFO(dev)->gen < 4) {
112 * On the 965, the sampler cache always gets flushed
113 * and this bit is reserved.
115 if (invalidate_domains & I915_GEM_DOMAIN_SAMPLER)
116 cmd |= MI_READ_FLUSH;
118 if (invalidate_domains & I915_GEM_DOMAIN_INSTRUCTION)
121 if (invalidate_domains & I915_GEM_DOMAIN_COMMAND &&
122 (IS_G4X(dev) || IS_GEN5(dev)))
123 cmd |= MI_INVALIDATE_ISP;
125 ret = intel_ring_begin(ring, 2);
129 intel_ring_emit(ring, cmd);
130 intel_ring_emit(ring, MI_NOOP);
131 intel_ring_advance(ring);
137 * Emits a PIPE_CONTROL with a non-zero post-sync operation, for
138 * implementing two workarounds on gen6. From section 1.4.7.1
139 * "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
141 * [DevSNB-C+{W/A}] Before any depth stall flush (including those
142 * produced by non-pipelined state commands), software needs to first
143 * send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
146 * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
147 * =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
149 * And the workaround for these two requires this workaround first:
151 * [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
152 * BEFORE the pipe-control with a post-sync op and no write-cache
155 * And this last workaround is tricky because of the requirements on
156 * that bit. From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
159 * "1 of the following must also be set:
160 * - Render Target Cache Flush Enable ([12] of DW1)
161 * - Depth Cache Flush Enable ([0] of DW1)
162 * - Stall at Pixel Scoreboard ([1] of DW1)
163 * - Depth Stall ([13] of DW1)
164 * - Post-Sync Operation ([13] of DW1)
165 * - Notify Enable ([8] of DW1)"
167 * The cache flushes require the workaround flush that triggered this
168 * one, so we can't use it. Depth stall would trigger the same.
169 * Post-sync nonzero is what triggered this second workaround, so we
170 * can't use that one either. Notify enable is IRQs, which aren't
171 * really our business. That leaves only stall at scoreboard.
174 intel_emit_post_sync_nonzero_flush(struct intel_ring_buffer *ring)
176 struct pipe_control *pc = ring->private;
177 u32 scratch_addr = pc->gtt_offset + 128;
181 ret = intel_ring_begin(ring, 6);
185 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(5));
186 intel_ring_emit(ring, PIPE_CONTROL_CS_STALL |
187 PIPE_CONTROL_STALL_AT_SCOREBOARD);
188 intel_ring_emit(ring, scratch_addr | PIPE_CONTROL_GLOBAL_GTT); /* address */
189 intel_ring_emit(ring, 0); /* low dword */
190 intel_ring_emit(ring, 0); /* high dword */
191 intel_ring_emit(ring, MI_NOOP);
192 intel_ring_advance(ring);
194 ret = intel_ring_begin(ring, 6);
198 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(5));
199 intel_ring_emit(ring, PIPE_CONTROL_QW_WRITE);
200 intel_ring_emit(ring, scratch_addr | PIPE_CONTROL_GLOBAL_GTT); /* address */
201 intel_ring_emit(ring, 0);
202 intel_ring_emit(ring, 0);
203 intel_ring_emit(ring, MI_NOOP);
204 intel_ring_advance(ring);
210 gen6_render_ring_flush(struct intel_ring_buffer *ring,
211 u32 invalidate_domains, u32 flush_domains)
214 struct pipe_control *pc = ring->private;
215 u32 scratch_addr = pc->gtt_offset + 128;
218 /* Force SNB workarounds for PIPE_CONTROL flushes */
219 intel_emit_post_sync_nonzero_flush(ring);
221 /* Just flush everything. Experiments have shown that reducing the
222 * number of bits based on the write domains has little performance
225 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
226 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
227 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
228 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
229 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
230 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
231 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
233 ret = intel_ring_begin(ring, 6);
237 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(5));
238 intel_ring_emit(ring, flags);
239 intel_ring_emit(ring, scratch_addr | PIPE_CONTROL_GLOBAL_GTT);
240 intel_ring_emit(ring, 0); /* lower dword */
241 intel_ring_emit(ring, 0); /* uppwer dword */
242 intel_ring_emit(ring, MI_NOOP);
243 intel_ring_advance(ring);
248 static void ring_write_tail(struct intel_ring_buffer *ring,
251 drm_i915_private_t *dev_priv = ring->dev->dev_private;
252 I915_WRITE_TAIL(ring, value);
255 u32 intel_ring_get_active_head(struct intel_ring_buffer *ring)
257 drm_i915_private_t *dev_priv = ring->dev->dev_private;
258 u32 acthd_reg = INTEL_INFO(ring->dev)->gen >= 4 ?
259 RING_ACTHD(ring->mmio_base) : ACTHD;
261 return I915_READ(acthd_reg);
264 static int init_ring_common(struct intel_ring_buffer *ring)
266 drm_i915_private_t *dev_priv = ring->dev->dev_private;
267 struct drm_i915_gem_object *obj = ring->obj;
270 /* Stop the ring if it's running. */
271 I915_WRITE_CTL(ring, 0);
272 I915_WRITE_HEAD(ring, 0);
273 ring->write_tail(ring, 0);
275 /* Initialize the ring. */
276 I915_WRITE_START(ring, obj->gtt_offset);
277 head = I915_READ_HEAD(ring) & HEAD_ADDR;
279 /* G45 ring initialization fails to reset head to zero */
281 DRM_DEBUG_KMS("%s head not reset to zero "
282 "ctl %08x head %08x tail %08x start %08x\n",
285 I915_READ_HEAD(ring),
286 I915_READ_TAIL(ring),
287 I915_READ_START(ring));
289 I915_WRITE_HEAD(ring, 0);
291 if (I915_READ_HEAD(ring) & HEAD_ADDR) {
292 DRM_ERROR("failed to set %s head to zero "
293 "ctl %08x head %08x tail %08x start %08x\n",
296 I915_READ_HEAD(ring),
297 I915_READ_TAIL(ring),
298 I915_READ_START(ring));
303 ((ring->size - PAGE_SIZE) & RING_NR_PAGES)
304 | RING_REPORT_64K | RING_VALID);
306 /* If the head is still not zero, the ring is dead */
307 if ((I915_READ_CTL(ring) & RING_VALID) == 0 ||
308 I915_READ_START(ring) != obj->gtt_offset ||
309 (I915_READ_HEAD(ring) & HEAD_ADDR) != 0) {
310 DRM_ERROR("%s initialization failed "
311 "ctl %08x head %08x tail %08x start %08x\n",
314 I915_READ_HEAD(ring),
315 I915_READ_TAIL(ring),
316 I915_READ_START(ring));
320 if (!drm_core_check_feature(ring->dev, DRIVER_MODESET))
321 i915_kernel_lost_context(ring->dev);
323 ring->head = I915_READ_HEAD(ring);
324 ring->tail = I915_READ_TAIL(ring) & TAIL_ADDR;
325 ring->space = ring_space(ring);
332 init_pipe_control(struct intel_ring_buffer *ring)
334 struct pipe_control *pc;
335 struct drm_i915_gem_object *obj;
341 pc = kmalloc(sizeof(*pc), GFP_KERNEL);
345 obj = i915_gem_alloc_object(ring->dev, 4096);
347 DRM_ERROR("Failed to allocate seqno page\n");
352 i915_gem_object_set_cache_level(obj, I915_CACHE_LLC);
354 ret = i915_gem_object_pin(obj, 4096, true);
358 pc->gtt_offset = obj->gtt_offset;
359 pc->cpu_page = kmap(obj->pages[0]);
360 if (pc->cpu_page == NULL)
368 i915_gem_object_unpin(obj);
370 drm_gem_object_unreference(&obj->base);
377 cleanup_pipe_control(struct intel_ring_buffer *ring)
379 struct pipe_control *pc = ring->private;
380 struct drm_i915_gem_object *obj;
386 kunmap(obj->pages[0]);
387 i915_gem_object_unpin(obj);
388 drm_gem_object_unreference(&obj->base);
391 ring->private = NULL;
394 static int init_render_ring(struct intel_ring_buffer *ring)
396 struct drm_device *dev = ring->dev;
397 struct drm_i915_private *dev_priv = dev->dev_private;
398 int ret = init_ring_common(ring);
400 if (INTEL_INFO(dev)->gen > 3) {
401 int mode = VS_TIMER_DISPATCH << 16 | VS_TIMER_DISPATCH;
402 if (IS_GEN6(dev) || IS_GEN7(dev))
403 mode |= MI_FLUSH_ENABLE << 16 | MI_FLUSH_ENABLE;
404 I915_WRITE(MI_MODE, mode);
406 I915_WRITE(GFX_MODE_GEN7,
407 GFX_MODE_DISABLE(GFX_TLB_INVALIDATE_ALWAYS) |
408 GFX_MODE_ENABLE(GFX_REPLAY_MODE));
411 if (INTEL_INFO(dev)->gen >= 5) {
412 ret = init_pipe_control(ring);
417 if (INTEL_INFO(dev)->gen >= 6) {
419 INSTPM_FORCE_ORDERING << 16 | INSTPM_FORCE_ORDERING);
425 static void render_ring_cleanup(struct intel_ring_buffer *ring)
430 cleanup_pipe_control(ring);
434 update_mboxes(struct intel_ring_buffer *ring,
438 intel_ring_emit(ring, MI_SEMAPHORE_MBOX |
439 MI_SEMAPHORE_GLOBAL_GTT |
440 MI_SEMAPHORE_REGISTER |
441 MI_SEMAPHORE_UPDATE);
442 intel_ring_emit(ring, seqno);
443 intel_ring_emit(ring, mmio_offset);
447 * gen6_add_request - Update the semaphore mailbox registers
449 * @ring - ring that is adding a request
450 * @seqno - return seqno stuck into the ring
452 * Update the mailbox registers in the *other* rings with the current seqno.
453 * This acts like a signal in the canonical semaphore.
456 gen6_add_request(struct intel_ring_buffer *ring,
463 ret = intel_ring_begin(ring, 10);
467 mbox1_reg = ring->signal_mbox[0];
468 mbox2_reg = ring->signal_mbox[1];
470 *seqno = i915_gem_get_seqno(ring->dev);
472 update_mboxes(ring, *seqno, mbox1_reg);
473 update_mboxes(ring, *seqno, mbox2_reg);
474 intel_ring_emit(ring, MI_STORE_DWORD_INDEX);
475 intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
476 intel_ring_emit(ring, *seqno);
477 intel_ring_emit(ring, MI_USER_INTERRUPT);
478 intel_ring_advance(ring);
484 * intel_ring_sync - sync the waiter to the signaller on seqno
486 * @waiter - ring that is waiting
487 * @signaller - ring which has, or will signal
488 * @seqno - seqno which the waiter will block on
491 intel_ring_sync(struct intel_ring_buffer *waiter,
492 struct intel_ring_buffer *signaller,
497 u32 dw1 = MI_SEMAPHORE_MBOX |
498 MI_SEMAPHORE_COMPARE |
499 MI_SEMAPHORE_REGISTER;
501 ret = intel_ring_begin(waiter, 4);
505 intel_ring_emit(waiter, dw1 | signaller->semaphore_register[ring]);
506 intel_ring_emit(waiter, seqno);
507 intel_ring_emit(waiter, 0);
508 intel_ring_emit(waiter, MI_NOOP);
509 intel_ring_advance(waiter);
514 /* VCS->RCS (RVSYNC) or BCS->RCS (RBSYNC) */
516 render_ring_sync_to(struct intel_ring_buffer *waiter,
517 struct intel_ring_buffer *signaller,
520 WARN_ON(signaller->semaphore_register[RCS] == MI_SEMAPHORE_SYNC_INVALID);
521 return intel_ring_sync(waiter,
527 /* RCS->VCS (VRSYNC) or BCS->VCS (VBSYNC) */
529 gen6_bsd_ring_sync_to(struct intel_ring_buffer *waiter,
530 struct intel_ring_buffer *signaller,
533 WARN_ON(signaller->semaphore_register[VCS] == MI_SEMAPHORE_SYNC_INVALID);
534 return intel_ring_sync(waiter,
540 /* RCS->BCS (BRSYNC) or VCS->BCS (BVSYNC) */
542 gen6_blt_ring_sync_to(struct intel_ring_buffer *waiter,
543 struct intel_ring_buffer *signaller,
546 WARN_ON(signaller->semaphore_register[BCS] == MI_SEMAPHORE_SYNC_INVALID);
547 return intel_ring_sync(waiter,
555 #define PIPE_CONTROL_FLUSH(ring__, addr__) \
557 intel_ring_emit(ring__, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE | \
558 PIPE_CONTROL_DEPTH_STALL); \
559 intel_ring_emit(ring__, (addr__) | PIPE_CONTROL_GLOBAL_GTT); \
560 intel_ring_emit(ring__, 0); \
561 intel_ring_emit(ring__, 0); \
565 pc_render_add_request(struct intel_ring_buffer *ring,
568 struct drm_device *dev = ring->dev;
569 u32 seqno = i915_gem_get_seqno(dev);
570 struct pipe_control *pc = ring->private;
571 u32 scratch_addr = pc->gtt_offset + 128;
574 /* For Ironlake, MI_USER_INTERRUPT was deprecated and apparently
575 * incoherent with writes to memory, i.e. completely fubar,
576 * so we need to use PIPE_NOTIFY instead.
578 * However, we also need to workaround the qword write
579 * incoherence by flushing the 6 PIPE_NOTIFY buffers out to
580 * memory before requesting an interrupt.
582 ret = intel_ring_begin(ring, 32);
586 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE |
587 PIPE_CONTROL_WRITE_FLUSH |
588 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE);
589 intel_ring_emit(ring, pc->gtt_offset | PIPE_CONTROL_GLOBAL_GTT);
590 intel_ring_emit(ring, seqno);
591 intel_ring_emit(ring, 0);
592 PIPE_CONTROL_FLUSH(ring, scratch_addr);
593 scratch_addr += 128; /* write to separate cachelines */
594 PIPE_CONTROL_FLUSH(ring, scratch_addr);
596 PIPE_CONTROL_FLUSH(ring, scratch_addr);
598 PIPE_CONTROL_FLUSH(ring, scratch_addr);
600 PIPE_CONTROL_FLUSH(ring, scratch_addr);
602 PIPE_CONTROL_FLUSH(ring, scratch_addr);
603 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE |
604 PIPE_CONTROL_WRITE_FLUSH |
605 PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
606 PIPE_CONTROL_NOTIFY);
607 intel_ring_emit(ring, pc->gtt_offset | PIPE_CONTROL_GLOBAL_GTT);
608 intel_ring_emit(ring, seqno);
609 intel_ring_emit(ring, 0);
610 intel_ring_advance(ring);
617 render_ring_add_request(struct intel_ring_buffer *ring,
620 struct drm_device *dev = ring->dev;
621 u32 seqno = i915_gem_get_seqno(dev);
624 ret = intel_ring_begin(ring, 4);
628 intel_ring_emit(ring, MI_STORE_DWORD_INDEX);
629 intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
630 intel_ring_emit(ring, seqno);
631 intel_ring_emit(ring, MI_USER_INTERRUPT);
632 intel_ring_advance(ring);
639 gen6_ring_get_seqno(struct intel_ring_buffer *ring)
641 struct drm_device *dev = ring->dev;
643 /* Workaround to force correct ordering between irq and seqno writes on
644 * ivb (and maybe also on snb) by reading from a CS register (like
645 * ACTHD) before reading the status page. */
647 intel_ring_get_active_head(ring);
648 return intel_read_status_page(ring, I915_GEM_HWS_INDEX);
652 ring_get_seqno(struct intel_ring_buffer *ring)
654 return intel_read_status_page(ring, I915_GEM_HWS_INDEX);
658 pc_render_get_seqno(struct intel_ring_buffer *ring)
660 struct pipe_control *pc = ring->private;
661 return pc->cpu_page[0];
665 ironlake_enable_irq(drm_i915_private_t *dev_priv, u32 mask)
667 dev_priv->gt_irq_mask &= ~mask;
668 I915_WRITE(GTIMR, dev_priv->gt_irq_mask);
673 ironlake_disable_irq(drm_i915_private_t *dev_priv, u32 mask)
675 dev_priv->gt_irq_mask |= mask;
676 I915_WRITE(GTIMR, dev_priv->gt_irq_mask);
681 i915_enable_irq(drm_i915_private_t *dev_priv, u32 mask)
683 dev_priv->irq_mask &= ~mask;
684 I915_WRITE(IMR, dev_priv->irq_mask);
689 i915_disable_irq(drm_i915_private_t *dev_priv, u32 mask)
691 dev_priv->irq_mask |= mask;
692 I915_WRITE(IMR, dev_priv->irq_mask);
697 render_ring_get_irq(struct intel_ring_buffer *ring)
699 struct drm_device *dev = ring->dev;
700 drm_i915_private_t *dev_priv = dev->dev_private;
702 if (!dev->irq_enabled)
705 spin_lock(&ring->irq_lock);
706 if (ring->irq_refcount++ == 0) {
707 if (HAS_PCH_SPLIT(dev))
708 ironlake_enable_irq(dev_priv,
709 GT_PIPE_NOTIFY | GT_USER_INTERRUPT);
711 i915_enable_irq(dev_priv, I915_USER_INTERRUPT);
713 spin_unlock(&ring->irq_lock);
719 render_ring_put_irq(struct intel_ring_buffer *ring)
721 struct drm_device *dev = ring->dev;
722 drm_i915_private_t *dev_priv = dev->dev_private;
724 spin_lock(&ring->irq_lock);
725 if (--ring->irq_refcount == 0) {
726 if (HAS_PCH_SPLIT(dev))
727 ironlake_disable_irq(dev_priv,
731 i915_disable_irq(dev_priv, I915_USER_INTERRUPT);
733 spin_unlock(&ring->irq_lock);
736 void intel_ring_setup_status_page(struct intel_ring_buffer *ring)
738 struct drm_device *dev = ring->dev;
739 drm_i915_private_t *dev_priv = ring->dev->dev_private;
742 /* The ring status page addresses are no longer next to the rest of
743 * the ring registers as of gen7.
748 mmio = RENDER_HWS_PGA_GEN7;
751 mmio = BLT_HWS_PGA_GEN7;
754 mmio = BSD_HWS_PGA_GEN7;
757 } else if (IS_GEN6(ring->dev)) {
758 mmio = RING_HWS_PGA_GEN6(ring->mmio_base);
760 mmio = RING_HWS_PGA(ring->mmio_base);
763 I915_WRITE(mmio, (u32)ring->status_page.gfx_addr);
768 bsd_ring_flush(struct intel_ring_buffer *ring,
769 u32 invalidate_domains,
774 ret = intel_ring_begin(ring, 2);
778 intel_ring_emit(ring, MI_FLUSH);
779 intel_ring_emit(ring, MI_NOOP);
780 intel_ring_advance(ring);
785 ring_add_request(struct intel_ring_buffer *ring,
791 ret = intel_ring_begin(ring, 4);
795 seqno = i915_gem_get_seqno(ring->dev);
797 intel_ring_emit(ring, MI_STORE_DWORD_INDEX);
798 intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
799 intel_ring_emit(ring, seqno);
800 intel_ring_emit(ring, MI_USER_INTERRUPT);
801 intel_ring_advance(ring);
808 gen6_ring_get_irq(struct intel_ring_buffer *ring, u32 gflag, u32 rflag)
810 struct drm_device *dev = ring->dev;
811 drm_i915_private_t *dev_priv = dev->dev_private;
813 if (!dev->irq_enabled)
816 /* It looks like we need to prevent the gt from suspending while waiting
817 * for an notifiy irq, otherwise irqs seem to get lost on at least the
818 * blt/bsd rings on ivb. */
820 gen6_gt_force_wake_get(dev_priv);
822 spin_lock(&ring->irq_lock);
823 if (ring->irq_refcount++ == 0) {
824 ring->irq_mask &= ~rflag;
825 I915_WRITE_IMR(ring, ring->irq_mask);
826 ironlake_enable_irq(dev_priv, gflag);
828 spin_unlock(&ring->irq_lock);
834 gen6_ring_put_irq(struct intel_ring_buffer *ring, u32 gflag, u32 rflag)
836 struct drm_device *dev = ring->dev;
837 drm_i915_private_t *dev_priv = dev->dev_private;
839 spin_lock(&ring->irq_lock);
840 if (--ring->irq_refcount == 0) {
841 ring->irq_mask |= rflag;
842 I915_WRITE_IMR(ring, ring->irq_mask);
843 ironlake_disable_irq(dev_priv, gflag);
845 spin_unlock(&ring->irq_lock);
848 gen6_gt_force_wake_put(dev_priv);
852 bsd_ring_get_irq(struct intel_ring_buffer *ring)
854 struct drm_device *dev = ring->dev;
855 drm_i915_private_t *dev_priv = dev->dev_private;
857 if (!dev->irq_enabled)
860 spin_lock(&ring->irq_lock);
861 if (ring->irq_refcount++ == 0) {
863 i915_enable_irq(dev_priv, I915_BSD_USER_INTERRUPT);
865 ironlake_enable_irq(dev_priv, GT_BSD_USER_INTERRUPT);
867 spin_unlock(&ring->irq_lock);
872 bsd_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;
877 spin_lock(&ring->irq_lock);
878 if (--ring->irq_refcount == 0) {
880 i915_disable_irq(dev_priv, I915_BSD_USER_INTERRUPT);
882 ironlake_disable_irq(dev_priv, GT_BSD_USER_INTERRUPT);
884 spin_unlock(&ring->irq_lock);
888 ring_dispatch_execbuffer(struct intel_ring_buffer *ring, u32 offset, u32 length)
892 ret = intel_ring_begin(ring, 2);
896 intel_ring_emit(ring,
897 MI_BATCH_BUFFER_START | (2 << 6) |
898 MI_BATCH_NON_SECURE_I965);
899 intel_ring_emit(ring, offset);
900 intel_ring_advance(ring);
906 render_ring_dispatch_execbuffer(struct intel_ring_buffer *ring,
909 struct drm_device *dev = ring->dev;
912 if (IS_I830(dev) || IS_845G(dev)) {
913 ret = intel_ring_begin(ring, 4);
917 intel_ring_emit(ring, MI_BATCH_BUFFER);
918 intel_ring_emit(ring, offset | MI_BATCH_NON_SECURE);
919 intel_ring_emit(ring, offset + len - 8);
920 intel_ring_emit(ring, 0);
922 ret = intel_ring_begin(ring, 2);
926 if (INTEL_INFO(dev)->gen >= 4) {
927 intel_ring_emit(ring,
928 MI_BATCH_BUFFER_START | (2 << 6) |
929 MI_BATCH_NON_SECURE_I965);
930 intel_ring_emit(ring, offset);
932 intel_ring_emit(ring,
933 MI_BATCH_BUFFER_START | (2 << 6));
934 intel_ring_emit(ring, offset | MI_BATCH_NON_SECURE);
937 intel_ring_advance(ring);
942 static void cleanup_status_page(struct intel_ring_buffer *ring)
944 drm_i915_private_t *dev_priv = ring->dev->dev_private;
945 struct drm_i915_gem_object *obj;
947 obj = ring->status_page.obj;
951 kunmap(obj->pages[0]);
952 i915_gem_object_unpin(obj);
953 drm_gem_object_unreference(&obj->base);
954 ring->status_page.obj = NULL;
956 memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map));
959 static int init_status_page(struct intel_ring_buffer *ring)
961 struct drm_device *dev = ring->dev;
962 drm_i915_private_t *dev_priv = dev->dev_private;
963 struct drm_i915_gem_object *obj;
966 obj = i915_gem_alloc_object(dev, 4096);
968 DRM_ERROR("Failed to allocate status page\n");
973 i915_gem_object_set_cache_level(obj, I915_CACHE_LLC);
975 ret = i915_gem_object_pin(obj, 4096, true);
980 ring->status_page.gfx_addr = obj->gtt_offset;
981 ring->status_page.page_addr = kmap(obj->pages[0]);
982 if (ring->status_page.page_addr == NULL) {
983 memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map));
986 ring->status_page.obj = obj;
987 memset(ring->status_page.page_addr, 0, PAGE_SIZE);
989 intel_ring_setup_status_page(ring);
990 DRM_DEBUG_DRIVER("%s hws offset: 0x%08x\n",
991 ring->name, ring->status_page.gfx_addr);
996 i915_gem_object_unpin(obj);
998 drm_gem_object_unreference(&obj->base);
1003 int intel_init_ring_buffer(struct drm_device *dev,
1004 struct intel_ring_buffer *ring)
1006 struct drm_i915_gem_object *obj;
1010 INIT_LIST_HEAD(&ring->active_list);
1011 INIT_LIST_HEAD(&ring->request_list);
1012 INIT_LIST_HEAD(&ring->gpu_write_list);
1014 init_waitqueue_head(&ring->irq_queue);
1015 spin_lock_init(&ring->irq_lock);
1016 ring->irq_mask = ~0;
1018 if (I915_NEED_GFX_HWS(dev)) {
1019 ret = init_status_page(ring);
1024 obj = i915_gem_alloc_object(dev, ring->size);
1026 DRM_ERROR("Failed to allocate ringbuffer\n");
1033 ret = i915_gem_object_pin(obj, PAGE_SIZE, true);
1037 ring->map.size = ring->size;
1038 ring->map.offset = dev->agp->base + obj->gtt_offset;
1040 ring->map.flags = 0;
1043 drm_core_ioremap_wc(&ring->map, dev);
1044 if (ring->map.handle == NULL) {
1045 DRM_ERROR("Failed to map ringbuffer.\n");
1050 ring->virtual_start = ring->map.handle;
1051 ret = ring->init(ring);
1055 /* Workaround an erratum on the i830 which causes a hang if
1056 * the TAIL pointer points to within the last 2 cachelines
1059 ring->effective_size = ring->size;
1060 if (IS_I830(ring->dev))
1061 ring->effective_size -= 128;
1066 drm_core_ioremapfree(&ring->map, dev);
1068 i915_gem_object_unpin(obj);
1070 drm_gem_object_unreference(&obj->base);
1073 cleanup_status_page(ring);
1077 void intel_cleanup_ring_buffer(struct intel_ring_buffer *ring)
1079 struct drm_i915_private *dev_priv;
1082 if (ring->obj == NULL)
1085 /* Disable the ring buffer. The ring must be idle at this point */
1086 dev_priv = ring->dev->dev_private;
1087 ret = intel_wait_ring_idle(ring);
1089 DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",
1092 I915_WRITE_CTL(ring, 0);
1094 drm_core_ioremapfree(&ring->map, ring->dev);
1096 i915_gem_object_unpin(ring->obj);
1097 drm_gem_object_unreference(&ring->obj->base);
1101 ring->cleanup(ring);
1103 cleanup_status_page(ring);
1106 static int intel_wrap_ring_buffer(struct intel_ring_buffer *ring)
1109 int rem = ring->size - ring->tail;
1111 if (ring->space < rem) {
1112 int ret = intel_wait_ring_buffer(ring, rem);
1117 virt = (unsigned int *)(ring->virtual_start + ring->tail);
1125 ring->space = ring_space(ring);
1130 int intel_wait_ring_buffer(struct intel_ring_buffer *ring, int n)
1132 struct drm_device *dev = ring->dev;
1133 struct drm_i915_private *dev_priv = dev->dev_private;
1137 /* If the reported head position has wrapped or hasn't advanced,
1138 * fallback to the slow and accurate path.
1140 head = intel_read_status_page(ring, 4);
1141 if (head > ring->head) {
1143 ring->space = ring_space(ring);
1144 if (ring->space >= n)
1148 trace_i915_ring_wait_begin(ring);
1149 if (drm_core_check_feature(dev, DRIVER_GEM))
1150 /* With GEM the hangcheck timer should kick us out of the loop,
1151 * leaving it early runs the risk of corrupting GEM state (due
1152 * to running on almost untested codepaths). But on resume
1153 * timers don't work yet, so prevent a complete hang in that
1154 * case by choosing an insanely large timeout. */
1155 end = jiffies + 60 * HZ;
1157 end = jiffies + 3 * HZ;
1160 ring->head = I915_READ_HEAD(ring);
1161 ring->space = ring_space(ring);
1162 if (ring->space >= n) {
1163 trace_i915_ring_wait_end(ring);
1167 if (dev->primary->master) {
1168 struct drm_i915_master_private *master_priv = dev->primary->master->driver_priv;
1169 if (master_priv->sarea_priv)
1170 master_priv->sarea_priv->perf_boxes |= I915_BOX_WAIT;
1174 if (atomic_read(&dev_priv->mm.wedged))
1176 } while (!time_after(jiffies, end));
1177 trace_i915_ring_wait_end(ring);
1181 int intel_ring_begin(struct intel_ring_buffer *ring,
1184 struct drm_i915_private *dev_priv = ring->dev->dev_private;
1185 int n = 4*num_dwords;
1188 if (unlikely(atomic_read(&dev_priv->mm.wedged)))
1191 if (unlikely(ring->tail + n > ring->effective_size)) {
1192 ret = intel_wrap_ring_buffer(ring);
1197 if (unlikely(ring->space < n)) {
1198 ret = intel_wait_ring_buffer(ring, n);
1207 void intel_ring_advance(struct intel_ring_buffer *ring)
1209 ring->tail &= ring->size - 1;
1210 ring->write_tail(ring, ring->tail);
1213 static const struct intel_ring_buffer render_ring = {
1214 .name = "render ring",
1216 .mmio_base = RENDER_RING_BASE,
1217 .size = 32 * PAGE_SIZE,
1218 .init = init_render_ring,
1219 .write_tail = ring_write_tail,
1220 .flush = render_ring_flush,
1221 .add_request = render_ring_add_request,
1222 .get_seqno = ring_get_seqno,
1223 .irq_get = render_ring_get_irq,
1224 .irq_put = render_ring_put_irq,
1225 .dispatch_execbuffer = render_ring_dispatch_execbuffer,
1226 .cleanup = render_ring_cleanup,
1227 .sync_to = render_ring_sync_to,
1228 .semaphore_register = {MI_SEMAPHORE_SYNC_INVALID,
1229 MI_SEMAPHORE_SYNC_RV,
1230 MI_SEMAPHORE_SYNC_RB},
1231 .signal_mbox = {GEN6_VRSYNC, GEN6_BRSYNC},
1234 /* ring buffer for bit-stream decoder */
1236 static const struct intel_ring_buffer bsd_ring = {
1239 .mmio_base = BSD_RING_BASE,
1240 .size = 32 * PAGE_SIZE,
1241 .init = init_ring_common,
1242 .write_tail = ring_write_tail,
1243 .flush = bsd_ring_flush,
1244 .add_request = ring_add_request,
1245 .get_seqno = ring_get_seqno,
1246 .irq_get = bsd_ring_get_irq,
1247 .irq_put = bsd_ring_put_irq,
1248 .dispatch_execbuffer = ring_dispatch_execbuffer,
1252 static void gen6_bsd_ring_write_tail(struct intel_ring_buffer *ring,
1255 drm_i915_private_t *dev_priv = ring->dev->dev_private;
1257 /* Every tail move must follow the sequence below */
1258 I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL,
1259 GEN6_BSD_SLEEP_PSMI_CONTROL_RC_ILDL_MESSAGE_MODIFY_MASK |
1260 GEN6_BSD_SLEEP_PSMI_CONTROL_RC_ILDL_MESSAGE_DISABLE);
1261 I915_WRITE(GEN6_BSD_RNCID, 0x0);
1263 if (wait_for((I915_READ(GEN6_BSD_SLEEP_PSMI_CONTROL) &
1264 GEN6_BSD_SLEEP_PSMI_CONTROL_IDLE_INDICATOR) == 0,
1266 DRM_ERROR("timed out waiting for IDLE Indicator\n");
1268 I915_WRITE_TAIL(ring, value);
1269 I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL,
1270 GEN6_BSD_SLEEP_PSMI_CONTROL_RC_ILDL_MESSAGE_MODIFY_MASK |
1271 GEN6_BSD_SLEEP_PSMI_CONTROL_RC_ILDL_MESSAGE_ENABLE);
1274 static int gen6_ring_flush(struct intel_ring_buffer *ring,
1275 u32 invalidate, u32 flush)
1280 ret = intel_ring_begin(ring, 4);
1285 if (invalidate & I915_GEM_GPU_DOMAINS)
1286 cmd |= MI_INVALIDATE_TLB | MI_INVALIDATE_BSD;
1287 intel_ring_emit(ring, cmd);
1288 intel_ring_emit(ring, 0);
1289 intel_ring_emit(ring, 0);
1290 intel_ring_emit(ring, MI_NOOP);
1291 intel_ring_advance(ring);
1296 gen6_ring_dispatch_execbuffer(struct intel_ring_buffer *ring,
1297 u32 offset, u32 len)
1301 ret = intel_ring_begin(ring, 2);
1305 intel_ring_emit(ring, MI_BATCH_BUFFER_START | MI_BATCH_NON_SECURE_I965);
1306 /* bit0-7 is the length on GEN6+ */
1307 intel_ring_emit(ring, offset);
1308 intel_ring_advance(ring);
1314 gen6_render_ring_get_irq(struct intel_ring_buffer *ring)
1316 return gen6_ring_get_irq(ring,
1318 GEN6_RENDER_USER_INTERRUPT);
1322 gen6_render_ring_put_irq(struct intel_ring_buffer *ring)
1324 return gen6_ring_put_irq(ring,
1326 GEN6_RENDER_USER_INTERRUPT);
1330 gen6_bsd_ring_get_irq(struct intel_ring_buffer *ring)
1332 return gen6_ring_get_irq(ring,
1333 GT_GEN6_BSD_USER_INTERRUPT,
1334 GEN6_BSD_USER_INTERRUPT);
1338 gen6_bsd_ring_put_irq(struct intel_ring_buffer *ring)
1340 return gen6_ring_put_irq(ring,
1341 GT_GEN6_BSD_USER_INTERRUPT,
1342 GEN6_BSD_USER_INTERRUPT);
1345 /* ring buffer for Video Codec for Gen6+ */
1346 static const struct intel_ring_buffer gen6_bsd_ring = {
1347 .name = "gen6 bsd ring",
1349 .mmio_base = GEN6_BSD_RING_BASE,
1350 .size = 32 * PAGE_SIZE,
1351 .init = init_ring_common,
1352 .write_tail = gen6_bsd_ring_write_tail,
1353 .flush = gen6_ring_flush,
1354 .add_request = gen6_add_request,
1355 .get_seqno = gen6_ring_get_seqno,
1356 .irq_get = gen6_bsd_ring_get_irq,
1357 .irq_put = gen6_bsd_ring_put_irq,
1358 .dispatch_execbuffer = gen6_ring_dispatch_execbuffer,
1359 .sync_to = gen6_bsd_ring_sync_to,
1360 .semaphore_register = {MI_SEMAPHORE_SYNC_VR,
1361 MI_SEMAPHORE_SYNC_INVALID,
1362 MI_SEMAPHORE_SYNC_VB},
1363 .signal_mbox = {GEN6_RVSYNC, GEN6_BVSYNC},
1366 /* Blitter support (SandyBridge+) */
1369 blt_ring_get_irq(struct intel_ring_buffer *ring)
1371 return gen6_ring_get_irq(ring,
1372 GT_BLT_USER_INTERRUPT,
1373 GEN6_BLITTER_USER_INTERRUPT);
1377 blt_ring_put_irq(struct intel_ring_buffer *ring)
1379 gen6_ring_put_irq(ring,
1380 GT_BLT_USER_INTERRUPT,
1381 GEN6_BLITTER_USER_INTERRUPT);
1385 /* Workaround for some stepping of SNB,
1386 * each time when BLT engine ring tail moved,
1387 * the first command in the ring to be parsed
1388 * should be MI_BATCH_BUFFER_START
1390 #define NEED_BLT_WORKAROUND(dev) \
1391 (IS_GEN6(dev) && (dev->pdev->revision < 8))
1393 static inline struct drm_i915_gem_object *
1394 to_blt_workaround(struct intel_ring_buffer *ring)
1396 return ring->private;
1399 static int blt_ring_init(struct intel_ring_buffer *ring)
1401 if (NEED_BLT_WORKAROUND(ring->dev)) {
1402 struct drm_i915_gem_object *obj;
1406 obj = i915_gem_alloc_object(ring->dev, 4096);
1410 ret = i915_gem_object_pin(obj, 4096, true);
1412 drm_gem_object_unreference(&obj->base);
1416 ptr = kmap(obj->pages[0]);
1417 *ptr++ = MI_BATCH_BUFFER_END;
1419 kunmap(obj->pages[0]);
1421 ret = i915_gem_object_set_to_gtt_domain(obj, false);
1423 i915_gem_object_unpin(obj);
1424 drm_gem_object_unreference(&obj->base);
1428 ring->private = obj;
1431 return init_ring_common(ring);
1434 static int blt_ring_begin(struct intel_ring_buffer *ring,
1437 if (ring->private) {
1438 int ret = intel_ring_begin(ring, num_dwords+2);
1442 intel_ring_emit(ring, MI_BATCH_BUFFER_START);
1443 intel_ring_emit(ring, to_blt_workaround(ring)->gtt_offset);
1447 return intel_ring_begin(ring, 4);
1450 static int blt_ring_flush(struct intel_ring_buffer *ring,
1451 u32 invalidate, u32 flush)
1456 ret = blt_ring_begin(ring, 4);
1461 if (invalidate & I915_GEM_DOMAIN_RENDER)
1462 cmd |= MI_INVALIDATE_TLB;
1463 intel_ring_emit(ring, cmd);
1464 intel_ring_emit(ring, 0);
1465 intel_ring_emit(ring, 0);
1466 intel_ring_emit(ring, MI_NOOP);
1467 intel_ring_advance(ring);
1471 static void blt_ring_cleanup(struct intel_ring_buffer *ring)
1476 i915_gem_object_unpin(ring->private);
1477 drm_gem_object_unreference(ring->private);
1478 ring->private = NULL;
1481 static const struct intel_ring_buffer gen6_blt_ring = {
1484 .mmio_base = BLT_RING_BASE,
1485 .size = 32 * PAGE_SIZE,
1486 .init = blt_ring_init,
1487 .write_tail = ring_write_tail,
1488 .flush = blt_ring_flush,
1489 .add_request = gen6_add_request,
1490 .get_seqno = gen6_ring_get_seqno,
1491 .irq_get = blt_ring_get_irq,
1492 .irq_put = blt_ring_put_irq,
1493 .dispatch_execbuffer = gen6_ring_dispatch_execbuffer,
1494 .cleanup = blt_ring_cleanup,
1495 .sync_to = gen6_blt_ring_sync_to,
1496 .semaphore_register = {MI_SEMAPHORE_SYNC_BR,
1497 MI_SEMAPHORE_SYNC_BV,
1498 MI_SEMAPHORE_SYNC_INVALID},
1499 .signal_mbox = {GEN6_RBSYNC, GEN6_VBSYNC},
1502 int intel_init_render_ring_buffer(struct drm_device *dev)
1504 drm_i915_private_t *dev_priv = dev->dev_private;
1505 struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
1507 *ring = render_ring;
1508 if (INTEL_INFO(dev)->gen >= 6) {
1509 ring->add_request = gen6_add_request;
1510 ring->flush = gen6_render_ring_flush;
1511 ring->irq_get = gen6_render_ring_get_irq;
1512 ring->irq_put = gen6_render_ring_put_irq;
1513 ring->get_seqno = gen6_ring_get_seqno;
1514 } else if (IS_GEN5(dev)) {
1515 ring->add_request = pc_render_add_request;
1516 ring->get_seqno = pc_render_get_seqno;
1519 if (!I915_NEED_GFX_HWS(dev)) {
1520 ring->status_page.page_addr = dev_priv->status_page_dmah->vaddr;
1521 memset(ring->status_page.page_addr, 0, PAGE_SIZE);
1524 return intel_init_ring_buffer(dev, ring);
1527 int intel_render_ring_init_dri(struct drm_device *dev, u64 start, u32 size)
1529 drm_i915_private_t *dev_priv = dev->dev_private;
1530 struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
1532 *ring = render_ring;
1533 if (INTEL_INFO(dev)->gen >= 6) {
1534 ring->add_request = gen6_add_request;
1535 ring->irq_get = gen6_render_ring_get_irq;
1536 ring->irq_put = gen6_render_ring_put_irq;
1537 } else if (IS_GEN5(dev)) {
1538 ring->add_request = pc_render_add_request;
1539 ring->get_seqno = pc_render_get_seqno;
1542 if (!I915_NEED_GFX_HWS(dev))
1543 ring->status_page.page_addr = dev_priv->status_page_dmah->vaddr;
1546 INIT_LIST_HEAD(&ring->active_list);
1547 INIT_LIST_HEAD(&ring->request_list);
1548 INIT_LIST_HEAD(&ring->gpu_write_list);
1551 ring->effective_size = ring->size;
1552 if (IS_I830(ring->dev))
1553 ring->effective_size -= 128;
1555 ring->map.offset = start;
1556 ring->map.size = size;
1558 ring->map.flags = 0;
1561 drm_core_ioremap_wc(&ring->map, dev);
1562 if (ring->map.handle == NULL) {
1563 DRM_ERROR("can not ioremap virtual address for"
1568 ring->virtual_start = (void __force __iomem *)ring->map.handle;
1572 int intel_init_bsd_ring_buffer(struct drm_device *dev)
1574 drm_i915_private_t *dev_priv = dev->dev_private;
1575 struct intel_ring_buffer *ring = &dev_priv->ring[VCS];
1577 if (IS_GEN6(dev) || IS_GEN7(dev))
1578 *ring = gen6_bsd_ring;
1582 return intel_init_ring_buffer(dev, ring);
1585 int intel_init_blt_ring_buffer(struct drm_device *dev)
1587 drm_i915_private_t *dev_priv = dev->dev_private;
1588 struct intel_ring_buffer *ring = &dev_priv->ring[BCS];
1590 *ring = gen6_blt_ring;
1592 return intel_init_ring_buffer(dev, ring);