2 * Copyright © 2014 Broadcom
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 #include <linux/module.h>
25 #include <linux/platform_device.h>
26 #include <linux/device.h>
29 #include "uapi/drm/vc4_drm.h"
32 #include "vc4_trace.h"
35 vc4_queue_hangcheck(struct drm_device *dev)
37 struct vc4_dev *vc4 = to_vc4_dev(dev);
39 mod_timer(&vc4->hangcheck.timer,
40 round_jiffies_up(jiffies + msecs_to_jiffies(100)));
43 struct vc4_hang_state {
44 struct drm_vc4_get_hang_state user_state;
47 struct drm_gem_object **bo;
51 vc4_free_hang_state(struct drm_device *dev, struct vc4_hang_state *state)
55 mutex_lock(&dev->struct_mutex);
56 for (i = 0; i < state->user_state.bo_count; i++)
57 drm_gem_object_unreference(state->bo[i]);
58 mutex_unlock(&dev->struct_mutex);
64 vc4_get_hang_state_ioctl(struct drm_device *dev, void *data,
65 struct drm_file *file_priv)
67 struct drm_vc4_get_hang_state *get_state = data;
68 struct drm_vc4_get_hang_state_bo *bo_state;
69 struct vc4_hang_state *kernel_state;
70 struct drm_vc4_get_hang_state *state;
71 struct vc4_dev *vc4 = to_vc4_dev(dev);
72 unsigned long irqflags;
76 spin_lock_irqsave(&vc4->job_lock, irqflags);
77 kernel_state = vc4->hang_state;
79 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
82 state = &kernel_state->user_state;
84 /* If the user's array isn't big enough, just return the
85 * required array size.
87 if (get_state->bo_count < state->bo_count) {
88 get_state->bo_count = state->bo_count;
89 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
93 vc4->hang_state = NULL;
94 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
96 /* Save the user's BO pointer, so we don't stomp it with the memcpy. */
97 state->bo = get_state->bo;
98 memcpy(get_state, state, sizeof(*state));
100 bo_state = kcalloc(state->bo_count, sizeof(*bo_state), GFP_KERNEL);
106 for (i = 0; i < state->bo_count; i++) {
107 struct vc4_bo *vc4_bo = to_vc4_bo(kernel_state->bo[i]);
110 ret = drm_gem_handle_create(file_priv, kernel_state->bo[i],
114 state->bo_count = i - 1;
117 bo_state[i].handle = handle;
118 bo_state[i].paddr = vc4_bo->base.paddr;
119 bo_state[i].size = vc4_bo->base.base.size;
122 if (copy_to_user((void __user *)(uintptr_t)get_state->bo,
124 state->bo_count * sizeof(*bo_state)))
131 vc4_free_hang_state(dev, kernel_state);
138 vc4_save_hang_state(struct drm_device *dev)
140 struct vc4_dev *vc4 = to_vc4_dev(dev);
141 struct drm_vc4_get_hang_state *state;
142 struct vc4_hang_state *kernel_state;
143 struct vc4_exec_info *exec;
145 unsigned long irqflags;
146 unsigned int i, unref_list_count;
148 kernel_state = kcalloc(1, sizeof(*kernel_state), GFP_KERNEL);
152 state = &kernel_state->user_state;
154 spin_lock_irqsave(&vc4->job_lock, irqflags);
155 exec = vc4_first_job(vc4);
157 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
161 unref_list_count = 0;
162 list_for_each_entry(bo, &exec->unref_list, unref_head)
165 state->bo_count = exec->bo_count + unref_list_count;
166 kernel_state->bo = kcalloc(state->bo_count, sizeof(*kernel_state->bo),
168 if (!kernel_state->bo) {
169 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
173 for (i = 0; i < exec->bo_count; i++) {
174 drm_gem_object_reference(&exec->bo[i]->base);
175 kernel_state->bo[i] = &exec->bo[i]->base;
178 list_for_each_entry(bo, &exec->unref_list, unref_head) {
179 drm_gem_object_reference(&bo->base.base);
180 kernel_state->bo[i] = &bo->base.base;
184 state->start_bin = exec->ct0ca;
185 state->start_render = exec->ct1ca;
187 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
189 state->ct0ca = V3D_READ(V3D_CTNCA(0));
190 state->ct0ea = V3D_READ(V3D_CTNEA(0));
192 state->ct1ca = V3D_READ(V3D_CTNCA(1));
193 state->ct1ea = V3D_READ(V3D_CTNEA(1));
195 state->ct0cs = V3D_READ(V3D_CTNCS(0));
196 state->ct1cs = V3D_READ(V3D_CTNCS(1));
198 state->ct0ra0 = V3D_READ(V3D_CT00RA0);
199 state->ct1ra0 = V3D_READ(V3D_CT01RA0);
201 state->bpca = V3D_READ(V3D_BPCA);
202 state->bpcs = V3D_READ(V3D_BPCS);
203 state->bpoa = V3D_READ(V3D_BPOA);
204 state->bpos = V3D_READ(V3D_BPOS);
206 state->vpmbase = V3D_READ(V3D_VPMBASE);
208 state->dbge = V3D_READ(V3D_DBGE);
209 state->fdbgo = V3D_READ(V3D_FDBGO);
210 state->fdbgb = V3D_READ(V3D_FDBGB);
211 state->fdbgr = V3D_READ(V3D_FDBGR);
212 state->fdbgs = V3D_READ(V3D_FDBGS);
213 state->errstat = V3D_READ(V3D_ERRSTAT);
215 spin_lock_irqsave(&vc4->job_lock, irqflags);
216 if (vc4->hang_state) {
217 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
218 vc4_free_hang_state(dev, kernel_state);
220 vc4->hang_state = kernel_state;
221 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
226 vc4_reset(struct drm_device *dev)
228 struct vc4_dev *vc4 = to_vc4_dev(dev);
230 DRM_INFO("Resetting GPU.\n");
231 vc4_v3d_set_power(vc4, false);
232 vc4_v3d_set_power(vc4, true);
236 /* Rearm the hangcheck -- another job might have been waiting
237 * for our hung one to get kicked off, and vc4_irq_reset()
238 * would have started it.
240 vc4_queue_hangcheck(dev);
244 vc4_reset_work(struct work_struct *work)
246 struct vc4_dev *vc4 =
247 container_of(work, struct vc4_dev, hangcheck.reset_work);
249 vc4_save_hang_state(vc4->dev);
255 vc4_hangcheck_elapsed(unsigned long data)
257 struct drm_device *dev = (struct drm_device *)data;
258 struct vc4_dev *vc4 = to_vc4_dev(dev);
259 uint32_t ct0ca, ct1ca;
261 /* If idle, we can stop watching for hangs. */
262 if (list_empty(&vc4->job_list))
265 ct0ca = V3D_READ(V3D_CTNCA(0));
266 ct1ca = V3D_READ(V3D_CTNCA(1));
268 /* If we've made any progress in execution, rearm the timer
271 if (ct0ca != vc4->hangcheck.last_ct0ca ||
272 ct1ca != vc4->hangcheck.last_ct1ca) {
273 vc4->hangcheck.last_ct0ca = ct0ca;
274 vc4->hangcheck.last_ct1ca = ct1ca;
275 vc4_queue_hangcheck(dev);
279 /* We've gone too long with no progress, reset. This has to
280 * be done from a work struct, since resetting can sleep and
281 * this timer hook isn't allowed to.
283 schedule_work(&vc4->hangcheck.reset_work);
287 submit_cl(struct drm_device *dev, uint32_t thread, uint32_t start, uint32_t end)
289 struct vc4_dev *vc4 = to_vc4_dev(dev);
291 /* Set the current and end address of the control list.
292 * Writing the end register is what starts the job.
294 V3D_WRITE(V3D_CTNCA(thread), start);
295 V3D_WRITE(V3D_CTNEA(thread), end);
299 vc4_wait_for_seqno(struct drm_device *dev, uint64_t seqno, uint64_t timeout_ns,
302 struct vc4_dev *vc4 = to_vc4_dev(dev);
304 unsigned long timeout_expire;
307 if (vc4->finished_seqno >= seqno)
313 timeout_expire = jiffies + nsecs_to_jiffies(timeout_ns);
315 trace_vc4_wait_for_seqno_begin(dev, seqno, timeout_ns);
317 prepare_to_wait(&vc4->job_wait_queue, &wait,
318 interruptible ? TASK_INTERRUPTIBLE :
319 TASK_UNINTERRUPTIBLE);
321 if (interruptible && signal_pending(current)) {
326 if (vc4->finished_seqno >= seqno)
329 if (timeout_ns != ~0ull) {
330 if (time_after_eq(jiffies, timeout_expire)) {
334 schedule_timeout(timeout_expire - jiffies);
340 finish_wait(&vc4->job_wait_queue, &wait);
341 trace_vc4_wait_for_seqno_end(dev, seqno);
343 if (ret && ret != -ERESTARTSYS) {
344 DRM_ERROR("timeout waiting for render thread idle\n");
352 vc4_flush_caches(struct drm_device *dev)
354 struct vc4_dev *vc4 = to_vc4_dev(dev);
356 /* Flush the GPU L2 caches. These caches sit on top of system
357 * L3 (the 128kb or so shared with the CPU), and are
358 * non-allocating in the L3.
360 V3D_WRITE(V3D_L2CACTL,
363 V3D_WRITE(V3D_SLCACTL,
364 VC4_SET_FIELD(0xf, V3D_SLCACTL_T1CC) |
365 VC4_SET_FIELD(0xf, V3D_SLCACTL_T0CC) |
366 VC4_SET_FIELD(0xf, V3D_SLCACTL_UCC) |
367 VC4_SET_FIELD(0xf, V3D_SLCACTL_ICC));
370 /* Sets the registers for the next job to be actually be executed in
373 * The job_lock should be held during this.
376 vc4_submit_next_job(struct drm_device *dev)
378 struct vc4_dev *vc4 = to_vc4_dev(dev);
379 struct vc4_exec_info *exec = vc4_first_job(vc4);
384 vc4_flush_caches(dev);
386 /* Disable the binner's pre-loaded overflow memory address */
387 V3D_WRITE(V3D_BPOA, 0);
388 V3D_WRITE(V3D_BPOS, 0);
390 if (exec->ct0ca != exec->ct0ea)
391 submit_cl(dev, 0, exec->ct0ca, exec->ct0ea);
392 submit_cl(dev, 1, exec->ct1ca, exec->ct1ea);
396 vc4_update_bo_seqnos(struct vc4_exec_info *exec, uint64_t seqno)
401 for (i = 0; i < exec->bo_count; i++) {
402 bo = to_vc4_bo(&exec->bo[i]->base);
406 list_for_each_entry(bo, &exec->unref_list, unref_head) {
411 /* Queues a struct vc4_exec_info for execution. If no job is
412 * currently executing, then submits it.
414 * Unlike most GPUs, our hardware only handles one command list at a
415 * time. To queue multiple jobs at once, we'd need to edit the
416 * previous command list to have a jump to the new one at the end, and
417 * then bump the end address. That's a change for a later date,
421 vc4_queue_submit(struct drm_device *dev, struct vc4_exec_info *exec)
423 struct vc4_dev *vc4 = to_vc4_dev(dev);
425 unsigned long irqflags;
427 spin_lock_irqsave(&vc4->job_lock, irqflags);
429 seqno = ++vc4->emit_seqno;
431 vc4_update_bo_seqnos(exec, seqno);
433 list_add_tail(&exec->head, &vc4->job_list);
435 /* If no job was executing, kick ours off. Otherwise, it'll
436 * get started when the previous job's frame done interrupt
439 if (vc4_first_job(vc4) == exec) {
440 vc4_submit_next_job(dev);
441 vc4_queue_hangcheck(dev);
444 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
448 * Looks up a bunch of GEM handles for BOs and stores the array for
449 * use in the command validator that actually writes relocated
450 * addresses pointing to them.
453 vc4_cl_lookup_bos(struct drm_device *dev,
454 struct drm_file *file_priv,
455 struct vc4_exec_info *exec)
457 struct drm_vc4_submit_cl *args = exec->args;
462 exec->bo_count = args->bo_handle_count;
464 if (!exec->bo_count) {
465 /* See comment on bo_index for why we have to check
468 DRM_ERROR("Rendering requires BOs to validate\n");
472 exec->bo = kcalloc(exec->bo_count, sizeof(struct drm_gem_cma_object *),
475 DRM_ERROR("Failed to allocate validated BO pointers\n");
479 handles = drm_malloc_ab(exec->bo_count, sizeof(uint32_t));
481 DRM_ERROR("Failed to allocate incoming GEM handles\n");
485 ret = copy_from_user(handles,
486 (void __user *)(uintptr_t)args->bo_handles,
487 exec->bo_count * sizeof(uint32_t));
489 DRM_ERROR("Failed to copy in GEM handles\n");
493 spin_lock(&file_priv->table_lock);
494 for (i = 0; i < exec->bo_count; i++) {
495 struct drm_gem_object *bo = idr_find(&file_priv->object_idr,
498 DRM_ERROR("Failed to look up GEM BO %d: %d\n",
501 spin_unlock(&file_priv->table_lock);
504 drm_gem_object_reference(bo);
505 exec->bo[i] = (struct drm_gem_cma_object *)bo;
507 spin_unlock(&file_priv->table_lock);
515 vc4_get_bcl(struct drm_device *dev, struct vc4_exec_info *exec)
517 struct drm_vc4_submit_cl *args = exec->args;
521 uint32_t bin_offset = 0;
522 uint32_t shader_rec_offset = roundup(bin_offset + args->bin_cl_size,
524 uint32_t uniforms_offset = shader_rec_offset + args->shader_rec_size;
525 uint32_t exec_size = uniforms_offset + args->uniforms_size;
526 uint32_t temp_size = exec_size + (sizeof(struct vc4_shader_state) *
527 args->shader_rec_count);
530 if (uniforms_offset < shader_rec_offset ||
531 exec_size < uniforms_offset ||
532 args->shader_rec_count >= (UINT_MAX /
533 sizeof(struct vc4_shader_state)) ||
534 temp_size < exec_size) {
535 DRM_ERROR("overflow in exec arguments\n");
539 /* Allocate space where we'll store the copied in user command lists
540 * and shader records.
542 * We don't just copy directly into the BOs because we need to
543 * read the contents back for validation, and I think the
544 * bo->vaddr is uncached access.
546 temp = kmalloc(temp_size, GFP_KERNEL);
548 DRM_ERROR("Failed to allocate storage for copying "
549 "in bin/render CLs.\n");
553 bin = temp + bin_offset;
554 exec->shader_rec_u = temp + shader_rec_offset;
555 exec->uniforms_u = temp + uniforms_offset;
556 exec->shader_state = temp + exec_size;
557 exec->shader_state_size = args->shader_rec_count;
559 if (copy_from_user(bin,
560 (void __user *)(uintptr_t)args->bin_cl,
561 args->bin_cl_size)) {
566 if (copy_from_user(exec->shader_rec_u,
567 (void __user *)(uintptr_t)args->shader_rec,
568 args->shader_rec_size)) {
573 if (copy_from_user(exec->uniforms_u,
574 (void __user *)(uintptr_t)args->uniforms,
575 args->uniforms_size)) {
580 bo = vc4_bo_create(dev, exec_size, true);
582 DRM_ERROR("Couldn't allocate BO for binning\n");
586 exec->exec_bo = &bo->base;
588 list_add_tail(&to_vc4_bo(&exec->exec_bo->base)->unref_head,
591 exec->ct0ca = exec->exec_bo->paddr + bin_offset;
595 exec->shader_rec_v = exec->exec_bo->vaddr + shader_rec_offset;
596 exec->shader_rec_p = exec->exec_bo->paddr + shader_rec_offset;
597 exec->shader_rec_size = args->shader_rec_size;
599 exec->uniforms_v = exec->exec_bo->vaddr + uniforms_offset;
600 exec->uniforms_p = exec->exec_bo->paddr + uniforms_offset;
601 exec->uniforms_size = args->uniforms_size;
603 ret = vc4_validate_bin_cl(dev,
604 exec->exec_bo->vaddr + bin_offset,
610 ret = vc4_validate_shader_recs(dev, exec);
618 vc4_complete_exec(struct drm_device *dev, struct vc4_exec_info *exec)
622 /* Need the struct lock for drm_gem_object_unreference(). */
623 mutex_lock(&dev->struct_mutex);
625 for (i = 0; i < exec->bo_count; i++)
626 drm_gem_object_unreference(&exec->bo[i]->base);
630 while (!list_empty(&exec->unref_list)) {
631 struct vc4_bo *bo = list_first_entry(&exec->unref_list,
632 struct vc4_bo, unref_head);
633 list_del(&bo->unref_head);
634 drm_gem_object_unreference(&bo->base.base);
636 mutex_unlock(&dev->struct_mutex);
642 vc4_job_handle_completed(struct vc4_dev *vc4)
644 unsigned long irqflags;
645 struct vc4_seqno_cb *cb, *cb_temp;
647 spin_lock_irqsave(&vc4->job_lock, irqflags);
648 while (!list_empty(&vc4->job_done_list)) {
649 struct vc4_exec_info *exec =
650 list_first_entry(&vc4->job_done_list,
651 struct vc4_exec_info, head);
652 list_del(&exec->head);
654 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
655 vc4_complete_exec(vc4->dev, exec);
656 spin_lock_irqsave(&vc4->job_lock, irqflags);
659 list_for_each_entry_safe(cb, cb_temp, &vc4->seqno_cb_list, work.entry) {
660 if (cb->seqno <= vc4->finished_seqno) {
661 list_del_init(&cb->work.entry);
662 schedule_work(&cb->work);
666 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
669 static void vc4_seqno_cb_work(struct work_struct *work)
671 struct vc4_seqno_cb *cb = container_of(work, struct vc4_seqno_cb, work);
676 int vc4_queue_seqno_cb(struct drm_device *dev,
677 struct vc4_seqno_cb *cb, uint64_t seqno,
678 void (*func)(struct vc4_seqno_cb *cb))
680 struct vc4_dev *vc4 = to_vc4_dev(dev);
682 unsigned long irqflags;
685 INIT_WORK(&cb->work, vc4_seqno_cb_work);
687 spin_lock_irqsave(&vc4->job_lock, irqflags);
688 if (seqno > vc4->finished_seqno) {
690 list_add_tail(&cb->work.entry, &vc4->seqno_cb_list);
692 schedule_work(&cb->work);
694 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
699 /* Scheduled when any job has been completed, this walks the list of
700 * jobs that had completed and unrefs their BOs and frees their exec
704 vc4_job_done_work(struct work_struct *work)
706 struct vc4_dev *vc4 =
707 container_of(work, struct vc4_dev, job_done_work);
709 vc4_job_handle_completed(vc4);
713 vc4_wait_for_seqno_ioctl_helper(struct drm_device *dev,
715 uint64_t *timeout_ns)
717 unsigned long start = jiffies;
718 int ret = vc4_wait_for_seqno(dev, seqno, *timeout_ns, true);
720 if ((ret == -EINTR || ret == -ERESTARTSYS) && *timeout_ns != ~0ull) {
721 uint64_t delta = jiffies_to_nsecs(jiffies - start);
723 if (*timeout_ns >= delta)
724 *timeout_ns -= delta;
731 vc4_wait_seqno_ioctl(struct drm_device *dev, void *data,
732 struct drm_file *file_priv)
734 struct drm_vc4_wait_seqno *args = data;
736 return vc4_wait_for_seqno_ioctl_helper(dev, args->seqno,
741 vc4_wait_bo_ioctl(struct drm_device *dev, void *data,
742 struct drm_file *file_priv)
745 struct drm_vc4_wait_bo *args = data;
746 struct drm_gem_object *gem_obj;
749 gem_obj = drm_gem_object_lookup(dev, file_priv, args->handle);
751 DRM_ERROR("Failed to look up GEM BO %d\n", args->handle);
754 bo = to_vc4_bo(gem_obj);
756 ret = vc4_wait_for_seqno_ioctl_helper(dev, bo->seqno,
759 drm_gem_object_unreference_unlocked(gem_obj);
764 * Submits a command list to the VC4.
766 * This is what is called batchbuffer emitting on other hardware.
769 vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
770 struct drm_file *file_priv)
772 struct vc4_dev *vc4 = to_vc4_dev(dev);
773 struct drm_vc4_submit_cl *args = data;
774 struct vc4_exec_info *exec;
777 if ((args->flags & ~VC4_SUBMIT_CL_USE_CLEAR_COLOR) != 0) {
778 DRM_ERROR("Unknown flags: 0x%02x\n", args->flags);
782 exec = kcalloc(1, sizeof(*exec), GFP_KERNEL);
784 DRM_ERROR("malloc failure on exec struct\n");
789 INIT_LIST_HEAD(&exec->unref_list);
791 ret = vc4_cl_lookup_bos(dev, file_priv, exec);
795 if (exec->args->bin_cl_size != 0) {
796 ret = vc4_get_bcl(dev, exec);
804 ret = vc4_get_rcl(dev, exec);
808 /* Clear this out of the struct we'll be putting in the queue,
809 * since it's part of our stack.
813 vc4_queue_submit(dev, exec);
815 /* Return the seqno for our job. */
816 args->seqno = vc4->emit_seqno;
821 vc4_complete_exec(vc4->dev, exec);
827 vc4_gem_init(struct drm_device *dev)
829 struct vc4_dev *vc4 = to_vc4_dev(dev);
831 INIT_LIST_HEAD(&vc4->job_list);
832 INIT_LIST_HEAD(&vc4->job_done_list);
833 INIT_LIST_HEAD(&vc4->seqno_cb_list);
834 spin_lock_init(&vc4->job_lock);
836 INIT_WORK(&vc4->hangcheck.reset_work, vc4_reset_work);
837 setup_timer(&vc4->hangcheck.timer,
838 vc4_hangcheck_elapsed,
841 INIT_WORK(&vc4->job_done_work, vc4_job_done_work);
845 vc4_gem_destroy(struct drm_device *dev)
847 struct vc4_dev *vc4 = to_vc4_dev(dev);
849 /* Waiting for exec to finish would need to be done before
852 WARN_ON(vc4->emit_seqno != vc4->finished_seqno);
854 /* V3D should already have disabled its interrupt and cleared
855 * the overflow allocation registers. Now free the object.
857 if (vc4->overflow_mem) {
858 drm_gem_object_unreference_unlocked(&vc4->overflow_mem->base.base);
859 vc4->overflow_mem = NULL;
862 vc4_bo_cache_destroy(dev);
865 vc4_free_hang_state(dev, vc4->hang_state);