1 // SPDX-License-Identifier: GPL-2.0+
2 /* Copyright (C) 2014-2018 Broadcom */
4 #include <linux/device.h>
5 #include <linux/dma-mapping.h>
7 #include <linux/module.h>
8 #include <linux/platform_device.h>
9 #include <linux/pm_runtime.h>
10 #include <linux/reset.h>
11 #include <linux/sched/signal.h>
12 #include <linux/uaccess.h>
14 #include <drm/drm_syncobj.h>
15 #include <uapi/drm/v3d_drm.h>
19 #include "v3d_trace.h"
22 v3d_init_core(struct v3d_dev *v3d, int core)
24 /* Set OVRTMUOUT, which means that the texture sampler uniform
25 * configuration's tmu output type field is used, instead of
26 * using the hardware default behavior based on the texture
27 * type. If you want the default behavior, you can still put
28 * "2" in the indirect texture state's output_type field.
31 V3D_CORE_WRITE(core, V3D_CTL_MISCCFG, V3D_MISCCFG_OVRTMUOUT);
33 /* Whenever we flush the L2T cache, we always want to flush
36 V3D_CORE_WRITE(core, V3D_CTL_L2TFLSTA, 0);
37 V3D_CORE_WRITE(core, V3D_CTL_L2TFLEND, ~0);
40 /* Sets invariant state for the HW. */
42 v3d_init_hw_state(struct v3d_dev *v3d)
44 v3d_init_core(v3d, 0);
48 v3d_idle_axi(struct v3d_dev *v3d, int core)
50 V3D_CORE_WRITE(core, V3D_GMP_CFG, V3D_GMP_CFG_STOP_REQ);
52 if (wait_for((V3D_CORE_READ(core, V3D_GMP_STATUS) &
53 (V3D_GMP_STATUS_RD_COUNT_MASK |
54 V3D_GMP_STATUS_WR_COUNT_MASK |
55 V3D_GMP_STATUS_CFG_BUSY)) == 0, 100)) {
56 DRM_ERROR("Failed to wait for safe GMP shutdown\n");
61 v3d_idle_gca(struct v3d_dev *v3d)
66 V3D_GCA_WRITE(V3D_GCA_SAFE_SHUTDOWN, V3D_GCA_SAFE_SHUTDOWN_EN);
68 if (wait_for((V3D_GCA_READ(V3D_GCA_SAFE_SHUTDOWN_ACK) &
69 V3D_GCA_SAFE_SHUTDOWN_ACK_ACKED) ==
70 V3D_GCA_SAFE_SHUTDOWN_ACK_ACKED, 100)) {
71 DRM_ERROR("Failed to wait for safe GCA shutdown\n");
76 v3d_reset_by_bridge(struct v3d_dev *v3d)
78 int version = V3D_BRIDGE_READ(V3D_TOP_GR_BRIDGE_REVISION);
80 if (V3D_GET_FIELD(version, V3D_TOP_GR_BRIDGE_MAJOR) == 2) {
81 V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_0,
82 V3D_TOP_GR_BRIDGE_SW_INIT_0_V3D_CLK_108_SW_INIT);
83 V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_0, 0);
85 /* GFXH-1383: The SW_INIT may cause a stray write to address 0
86 * of the unit, so reset it to its power-on value here.
88 V3D_WRITE(V3D_HUB_AXICFG, V3D_HUB_AXICFG_MAX_LEN_MASK);
90 WARN_ON_ONCE(V3D_GET_FIELD(version,
91 V3D_TOP_GR_BRIDGE_MAJOR) != 7);
92 V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_1,
93 V3D_TOP_GR_BRIDGE_SW_INIT_1_V3D_CLK_108_SW_INIT);
94 V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_1, 0);
99 v3d_reset_v3d(struct v3d_dev *v3d)
102 reset_control_reset(v3d->reset);
104 v3d_reset_by_bridge(v3d);
106 v3d_init_hw_state(v3d);
110 v3d_reset(struct v3d_dev *v3d)
112 struct drm_device *dev = &v3d->drm;
114 DRM_DEV_ERROR(dev->dev, "Resetting GPU for hang.\n");
115 DRM_DEV_ERROR(dev->dev, "V3D_ERR_STAT: 0x%08x\n",
116 V3D_CORE_READ(0, V3D_ERR_STAT));
117 trace_v3d_reset_begin(dev);
119 /* XXX: only needed for safe powerdown, not reset. */
121 v3d_idle_axi(v3d, 0);
126 v3d_mmu_set_page_table(v3d);
129 trace_v3d_reset_end(dev);
133 v3d_flush_l3(struct v3d_dev *v3d)
136 u32 gca_ctrl = V3D_GCA_READ(V3D_GCA_CACHE_CTRL);
138 V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL,
139 gca_ctrl | V3D_GCA_CACHE_CTRL_FLUSH);
142 V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL,
143 gca_ctrl & ~V3D_GCA_CACHE_CTRL_FLUSH);
148 /* Invalidates the (read-only) L2C cache. This was the L2 cache for
149 * uniforms and instructions on V3D 3.2.
152 v3d_invalidate_l2c(struct v3d_dev *v3d, int core)
157 V3D_CORE_WRITE(core, V3D_CTL_L2CACTL,
162 /* Invalidates texture L2 cachelines */
164 v3d_flush_l2t(struct v3d_dev *v3d, int core)
166 /* While there is a busy bit (V3D_L2TCACTL_L2TFLS), we don't
167 * need to wait for completion before dispatching the job --
168 * L2T accesses will be stalled until the flush has completed.
169 * However, we do need to make sure we don't try to trigger a
170 * new flush while the L2_CLEAN queue is trying to
171 * synchronously clean after a job.
173 mutex_lock(&v3d->cache_clean_lock);
174 V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
175 V3D_L2TCACTL_L2TFLS |
176 V3D_SET_FIELD(V3D_L2TCACTL_FLM_FLUSH, V3D_L2TCACTL_FLM));
177 mutex_unlock(&v3d->cache_clean_lock);
180 /* Cleans texture L1 and L2 cachelines (writing back dirty data).
182 * For cleaning, which happens from the CACHE_CLEAN queue after CSD has
183 * executed, we need to make sure that the clean is done before
184 * signaling job completion. So, we synchronously wait before
185 * returning, and we make sure that L2 invalidates don't happen in the
186 * meantime to confuse our are-we-done checks.
189 v3d_clean_caches(struct v3d_dev *v3d)
191 struct drm_device *dev = &v3d->drm;
194 trace_v3d_cache_clean_begin(dev);
196 V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL, V3D_L2TCACTL_TMUWCF);
197 if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
198 V3D_L2TCACTL_L2TFLS), 100)) {
199 DRM_ERROR("Timeout waiting for L1T write combiner flush\n");
202 mutex_lock(&v3d->cache_clean_lock);
203 V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
204 V3D_L2TCACTL_L2TFLS |
205 V3D_SET_FIELD(V3D_L2TCACTL_FLM_CLEAN, V3D_L2TCACTL_FLM));
207 if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
208 V3D_L2TCACTL_L2TFLS), 100)) {
209 DRM_ERROR("Timeout waiting for L2T clean\n");
212 mutex_unlock(&v3d->cache_clean_lock);
214 trace_v3d_cache_clean_end(dev);
217 /* Invalidates the slice caches. These are read-only caches. */
219 v3d_invalidate_slices(struct v3d_dev *v3d, int core)
221 V3D_CORE_WRITE(core, V3D_CTL_SLCACTL,
222 V3D_SET_FIELD(0xf, V3D_SLCACTL_TVCCS) |
223 V3D_SET_FIELD(0xf, V3D_SLCACTL_TDCCS) |
224 V3D_SET_FIELD(0xf, V3D_SLCACTL_UCC) |
225 V3D_SET_FIELD(0xf, V3D_SLCACTL_ICC));
229 v3d_invalidate_caches(struct v3d_dev *v3d)
231 /* Invalidate the caches from the outside in. That way if
232 * another CL's concurrent use of nearby memory were to pull
233 * an invalidated cacheline back in, we wouldn't leave stale
234 * data in the inner cache.
237 v3d_invalidate_l2c(v3d, 0);
238 v3d_flush_l2t(v3d, 0);
239 v3d_invalidate_slices(v3d, 0);
242 /* Takes the reservation lock on all the BOs being referenced, so that
243 * at queue submit time we can update the reservations.
245 * We don't lock the RCL the tile alloc/state BOs, or overflow memory
246 * (all of which are on exec->unref_list). They're entirely private
247 * to v3d, so we don't attach dma-buf fences to them.
250 v3d_lock_bo_reservations(struct v3d_job *job,
251 struct ww_acquire_ctx *acquire_ctx)
255 ret = drm_gem_lock_reservations(job->bo, job->bo_count, acquire_ctx);
259 for (i = 0; i < job->bo_count; i++) {
260 ret = drm_gem_fence_array_add_implicit(&job->deps,
263 drm_gem_unlock_reservations(job->bo, job->bo_count,
273 * v3d_lookup_bos() - Sets up job->bo[] with the GEM objects
274 * referenced by the job.
276 * @file_priv: DRM file for this fd
277 * @job: V3D job being set up
279 * The command validator needs to reference BOs by their index within
280 * the submitted job's BO list. This does the validation of the job's
281 * BO list and reference counting for the lifetime of the job.
283 * Note that this function doesn't need to unreference the BOs on
284 * failure, because that will happen at v3d_exec_cleanup() time.
287 v3d_lookup_bos(struct drm_device *dev,
288 struct drm_file *file_priv,
297 job->bo_count = bo_count;
299 if (!job->bo_count) {
300 /* See comment on bo_index for why we have to check
303 DRM_DEBUG("Rendering requires BOs\n");
307 job->bo = kvmalloc_array(job->bo_count,
308 sizeof(struct drm_gem_cma_object *),
309 GFP_KERNEL | __GFP_ZERO);
311 DRM_DEBUG("Failed to allocate validated BO pointers\n");
315 handles = kvmalloc_array(job->bo_count, sizeof(u32), GFP_KERNEL);
318 DRM_DEBUG("Failed to allocate incoming GEM handles\n");
322 if (copy_from_user(handles,
323 (void __user *)(uintptr_t)bo_handles,
324 job->bo_count * sizeof(u32))) {
326 DRM_DEBUG("Failed to copy in GEM handles\n");
330 spin_lock(&file_priv->table_lock);
331 for (i = 0; i < job->bo_count; i++) {
332 struct drm_gem_object *bo = idr_find(&file_priv->object_idr,
335 DRM_DEBUG("Failed to look up GEM BO %d: %d\n",
338 spin_unlock(&file_priv->table_lock);
341 drm_gem_object_get(bo);
344 spin_unlock(&file_priv->table_lock);
352 v3d_job_free(struct kref *ref)
354 struct v3d_job *job = container_of(ref, struct v3d_job, refcount);
356 struct dma_fence *fence;
359 for (i = 0; i < job->bo_count; i++) {
361 drm_gem_object_put_unlocked(job->bo[i]);
365 xa_for_each(&job->deps, index, fence) {
366 dma_fence_put(fence);
368 xa_destroy(&job->deps);
370 dma_fence_put(job->irq_fence);
371 dma_fence_put(job->done_fence);
373 pm_runtime_mark_last_busy(job->v3d->dev);
374 pm_runtime_put_autosuspend(job->v3d->dev);
380 v3d_render_job_free(struct kref *ref)
382 struct v3d_render_job *job = container_of(ref, struct v3d_render_job,
384 struct v3d_bo *bo, *save;
386 list_for_each_entry_safe(bo, save, &job->unref_list, unref_head) {
387 drm_gem_object_put_unlocked(&bo->base.base);
393 void v3d_job_put(struct v3d_job *job)
395 kref_put(&job->refcount, job->free);
399 v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
400 struct drm_file *file_priv)
403 struct drm_v3d_wait_bo *args = data;
404 ktime_t start = ktime_get();
406 unsigned long timeout_jiffies =
407 nsecs_to_jiffies_timeout(args->timeout_ns);
412 ret = drm_gem_dma_resv_wait(file_priv, args->handle,
413 true, timeout_jiffies);
415 /* Decrement the user's timeout, in case we got interrupted
416 * such that the ioctl will be restarted.
418 delta_ns = ktime_to_ns(ktime_sub(ktime_get(), start));
419 if (delta_ns < args->timeout_ns)
420 args->timeout_ns -= delta_ns;
422 args->timeout_ns = 0;
424 /* Asked to wait beyond the jiffie/scheduler precision? */
425 if (ret == -ETIME && args->timeout_ns)
432 v3d_job_init(struct v3d_dev *v3d, struct drm_file *file_priv,
433 struct v3d_job *job, void (*free)(struct kref *ref),
436 struct dma_fence *in_fence = NULL;
442 ret = pm_runtime_get_sync(v3d->dev);
446 xa_init_flags(&job->deps, XA_FLAGS_ALLOC);
448 ret = drm_syncobj_find_fence(file_priv, in_sync, 0, 0, &in_fence);
452 ret = drm_gem_fence_array_add(&job->deps, in_fence);
456 kref_init(&job->refcount);
460 xa_destroy(&job->deps);
461 pm_runtime_put_autosuspend(v3d->dev);
466 v3d_push_job(struct v3d_file_priv *v3d_priv,
467 struct v3d_job *job, enum v3d_queue queue)
471 ret = drm_sched_job_init(&job->base, &v3d_priv->sched_entity[queue],
476 job->done_fence = dma_fence_get(&job->base.s_fence->finished);
478 /* put by scheduler job completion */
479 kref_get(&job->refcount);
481 drm_sched_entity_push_job(&job->base, &v3d_priv->sched_entity[queue]);
487 v3d_attach_fences_and_unlock_reservation(struct drm_file *file_priv,
489 struct ww_acquire_ctx *acquire_ctx,
491 struct dma_fence *done_fence)
493 struct drm_syncobj *sync_out;
496 for (i = 0; i < job->bo_count; i++) {
497 /* XXX: Use shared fences for read-only objects. */
498 dma_resv_add_excl_fence(job->bo[i]->resv,
502 drm_gem_unlock_reservations(job->bo, job->bo_count, acquire_ctx);
504 /* Update the return sync object for the job */
505 sync_out = drm_syncobj_find(file_priv, out_sync);
507 drm_syncobj_replace_fence(sync_out, done_fence);
508 drm_syncobj_put(sync_out);
513 * v3d_submit_cl_ioctl() - Submits a job (frame) to the V3D.
515 * @data: ioctl argument
516 * @file_priv: DRM file for this fd
518 * This is the main entrypoint for userspace to submit a 3D frame to
519 * the GPU. Userspace provides the binner command list (if
520 * applicable), and the kernel sets up the render command list to draw
521 * to the framebuffer described in the ioctl, using the command lists
522 * that the 3D engine's binner will produce.
525 v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
526 struct drm_file *file_priv)
528 struct v3d_dev *v3d = to_v3d_dev(dev);
529 struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
530 struct drm_v3d_submit_cl *args = data;
531 struct v3d_bin_job *bin = NULL;
532 struct v3d_render_job *render;
533 struct ww_acquire_ctx acquire_ctx;
536 trace_v3d_submit_cl_ioctl(&v3d->drm, args->rcl_start, args->rcl_end);
538 if (args->pad != 0) {
539 DRM_INFO("pad must be zero: %d\n", args->pad);
543 render = kcalloc(1, sizeof(*render), GFP_KERNEL);
547 render->start = args->rcl_start;
548 render->end = args->rcl_end;
549 INIT_LIST_HEAD(&render->unref_list);
551 ret = v3d_job_init(v3d, file_priv, &render->base,
552 v3d_render_job_free, args->in_sync_rcl);
558 if (args->bcl_start != args->bcl_end) {
559 bin = kcalloc(1, sizeof(*bin), GFP_KERNEL);
563 ret = v3d_job_init(v3d, file_priv, &bin->base,
564 v3d_job_free, args->in_sync_bcl);
566 v3d_job_put(&render->base);
570 bin->start = args->bcl_start;
571 bin->end = args->bcl_end;
572 bin->qma = args->qma;
573 bin->qms = args->qms;
574 bin->qts = args->qts;
575 bin->render = render;
578 ret = v3d_lookup_bos(dev, file_priv, &render->base,
579 args->bo_handles, args->bo_handle_count);
583 ret = v3d_lock_bo_reservations(&render->base, &acquire_ctx);
587 mutex_lock(&v3d->sched_lock);
589 ret = v3d_push_job(v3d_priv, &bin->base, V3D_BIN);
593 ret = drm_gem_fence_array_add(&render->base.deps,
594 dma_fence_get(bin->base.done_fence));
599 ret = v3d_push_job(v3d_priv, &render->base, V3D_RENDER);
602 mutex_unlock(&v3d->sched_lock);
604 v3d_attach_fences_and_unlock_reservation(file_priv,
608 render->base.done_fence);
611 v3d_job_put(&bin->base);
612 v3d_job_put(&render->base);
617 mutex_unlock(&v3d->sched_lock);
618 drm_gem_unlock_reservations(render->base.bo,
619 render->base.bo_count, &acquire_ctx);
622 v3d_job_put(&bin->base);
623 v3d_job_put(&render->base);
629 * v3d_submit_tfu_ioctl() - Submits a TFU (texture formatting) job to the V3D.
631 * @data: ioctl argument
632 * @file_priv: DRM file for this fd
634 * Userspace provides the register setup for the TFU, which we don't
635 * need to validate since the TFU is behind the MMU.
638 v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
639 struct drm_file *file_priv)
641 struct v3d_dev *v3d = to_v3d_dev(dev);
642 struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
643 struct drm_v3d_submit_tfu *args = data;
644 struct v3d_tfu_job *job;
645 struct ww_acquire_ctx acquire_ctx;
648 trace_v3d_submit_tfu_ioctl(&v3d->drm, args->iia);
650 job = kcalloc(1, sizeof(*job), GFP_KERNEL);
654 ret = v3d_job_init(v3d, file_priv, &job->base,
655 v3d_job_free, args->in_sync);
661 job->base.bo = kcalloc(ARRAY_SIZE(args->bo_handles),
662 sizeof(*job->base.bo), GFP_KERNEL);
664 v3d_job_put(&job->base);
670 spin_lock(&file_priv->table_lock);
671 for (job->base.bo_count = 0;
672 job->base.bo_count < ARRAY_SIZE(args->bo_handles);
673 job->base.bo_count++) {
674 struct drm_gem_object *bo;
676 if (!args->bo_handles[job->base.bo_count])
679 bo = idr_find(&file_priv->object_idr,
680 args->bo_handles[job->base.bo_count]);
682 DRM_DEBUG("Failed to look up GEM BO %d: %d\n",
684 args->bo_handles[job->base.bo_count]);
686 spin_unlock(&file_priv->table_lock);
689 drm_gem_object_get(bo);
690 job->base.bo[job->base.bo_count] = bo;
692 spin_unlock(&file_priv->table_lock);
694 ret = v3d_lock_bo_reservations(&job->base, &acquire_ctx);
698 mutex_lock(&v3d->sched_lock);
699 ret = v3d_push_job(v3d_priv, &job->base, V3D_TFU);
702 mutex_unlock(&v3d->sched_lock);
704 v3d_attach_fences_and_unlock_reservation(file_priv,
705 &job->base, &acquire_ctx,
707 job->base.done_fence);
709 v3d_job_put(&job->base);
714 mutex_unlock(&v3d->sched_lock);
715 drm_gem_unlock_reservations(job->base.bo, job->base.bo_count,
718 v3d_job_put(&job->base);
724 * v3d_submit_csd_ioctl() - Submits a CSD (texture formatting) job to the V3D.
726 * @data: ioctl argument
727 * @file_priv: DRM file for this fd
729 * Userspace provides the register setup for the CSD, which we don't
730 * need to validate since the CSD is behind the MMU.
733 v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
734 struct drm_file *file_priv)
736 struct v3d_dev *v3d = to_v3d_dev(dev);
737 struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
738 struct drm_v3d_submit_csd *args = data;
739 struct v3d_csd_job *job;
740 struct v3d_job *clean_job;
741 struct ww_acquire_ctx acquire_ctx;
744 trace_v3d_submit_csd_ioctl(&v3d->drm, args->cfg[5], args->cfg[6]);
746 if (!v3d_has_csd(v3d)) {
747 DRM_DEBUG("Attempting CSD submit on non-CSD hardware\n");
751 job = kcalloc(1, sizeof(*job), GFP_KERNEL);
755 ret = v3d_job_init(v3d, file_priv, &job->base,
756 v3d_job_free, args->in_sync);
762 clean_job = kcalloc(1, sizeof(*clean_job), GFP_KERNEL);
764 v3d_job_put(&job->base);
769 ret = v3d_job_init(v3d, file_priv, clean_job, v3d_job_free, 0);
771 v3d_job_put(&job->base);
778 ret = v3d_lookup_bos(dev, file_priv, clean_job,
779 args->bo_handles, args->bo_handle_count);
783 ret = v3d_lock_bo_reservations(clean_job, &acquire_ctx);
787 mutex_lock(&v3d->sched_lock);
788 ret = v3d_push_job(v3d_priv, &job->base, V3D_CSD);
792 ret = drm_gem_fence_array_add(&clean_job->deps,
793 dma_fence_get(job->base.done_fence));
797 ret = v3d_push_job(v3d_priv, clean_job, V3D_CACHE_CLEAN);
800 mutex_unlock(&v3d->sched_lock);
802 v3d_attach_fences_and_unlock_reservation(file_priv,
806 clean_job->done_fence);
808 v3d_job_put(&job->base);
809 v3d_job_put(clean_job);
814 mutex_unlock(&v3d->sched_lock);
815 drm_gem_unlock_reservations(clean_job->bo, clean_job->bo_count,
818 v3d_job_put(&job->base);
819 v3d_job_put(clean_job);
825 v3d_gem_init(struct drm_device *dev)
827 struct v3d_dev *v3d = to_v3d_dev(dev);
828 u32 pt_size = 4096 * 1024;
831 for (i = 0; i < V3D_MAX_QUEUES; i++)
832 v3d->queue[i].fence_context = dma_fence_context_alloc(1);
834 spin_lock_init(&v3d->mm_lock);
835 spin_lock_init(&v3d->job_lock);
836 mutex_init(&v3d->bo_lock);
837 mutex_init(&v3d->reset_lock);
838 mutex_init(&v3d->sched_lock);
839 mutex_init(&v3d->cache_clean_lock);
841 /* Note: We don't allocate address 0. Various bits of HW
842 * treat 0 as special, such as the occlusion query counters
843 * where 0 means "disabled".
845 drm_mm_init(&v3d->mm, 1, pt_size / sizeof(u32) - 1);
847 v3d->pt = dma_alloc_wc(v3d->dev, pt_size,
849 GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO);
851 drm_mm_takedown(&v3d->mm);
853 "Failed to allocate page tables. "
854 "Please ensure you have CMA enabled.\n");
858 v3d_init_hw_state(v3d);
859 v3d_mmu_set_page_table(v3d);
861 ret = v3d_sched_init(v3d);
863 drm_mm_takedown(&v3d->mm);
864 dma_free_coherent(v3d->dev, 4096 * 1024, (void *)v3d->pt,
872 v3d_gem_destroy(struct drm_device *dev)
874 struct v3d_dev *v3d = to_v3d_dev(dev);
878 /* Waiting for jobs to finish would need to be done before
881 WARN_ON(v3d->bin_job);
882 WARN_ON(v3d->render_job);
884 drm_mm_takedown(&v3d->mm);
886 dma_free_coherent(v3d->dev, 4096 * 1024, (void *)v3d->pt, v3d->pt_paddr);