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");
350 vc4_flush_caches(struct drm_device *dev)
352 struct vc4_dev *vc4 = to_vc4_dev(dev);
354 /* Flush the GPU L2 caches. These caches sit on top of system
355 * L3 (the 128kb or so shared with the CPU), and are
356 * non-allocating in the L3.
358 V3D_WRITE(V3D_L2CACTL,
361 V3D_WRITE(V3D_SLCACTL,
362 VC4_SET_FIELD(0xf, V3D_SLCACTL_T1CC) |
363 VC4_SET_FIELD(0xf, V3D_SLCACTL_T0CC) |
364 VC4_SET_FIELD(0xf, V3D_SLCACTL_UCC) |
365 VC4_SET_FIELD(0xf, V3D_SLCACTL_ICC));
368 /* Sets the registers for the next job to be actually be executed in
371 * The job_lock should be held during this.
374 vc4_submit_next_job(struct drm_device *dev)
376 struct vc4_dev *vc4 = to_vc4_dev(dev);
377 struct vc4_exec_info *exec = vc4_first_job(vc4);
382 vc4_flush_caches(dev);
384 /* Disable the binner's pre-loaded overflow memory address */
385 V3D_WRITE(V3D_BPOA, 0);
386 V3D_WRITE(V3D_BPOS, 0);
388 if (exec->ct0ca != exec->ct0ea)
389 submit_cl(dev, 0, exec->ct0ca, exec->ct0ea);
390 submit_cl(dev, 1, exec->ct1ca, exec->ct1ea);
394 vc4_update_bo_seqnos(struct vc4_exec_info *exec, uint64_t seqno)
399 for (i = 0; i < exec->bo_count; i++) {
400 bo = to_vc4_bo(&exec->bo[i]->base);
404 list_for_each_entry(bo, &exec->unref_list, unref_head) {
409 /* Queues a struct vc4_exec_info for execution. If no job is
410 * currently executing, then submits it.
412 * Unlike most GPUs, our hardware only handles one command list at a
413 * time. To queue multiple jobs at once, we'd need to edit the
414 * previous command list to have a jump to the new one at the end, and
415 * then bump the end address. That's a change for a later date,
419 vc4_queue_submit(struct drm_device *dev, struct vc4_exec_info *exec)
421 struct vc4_dev *vc4 = to_vc4_dev(dev);
423 unsigned long irqflags;
425 spin_lock_irqsave(&vc4->job_lock, irqflags);
427 seqno = ++vc4->emit_seqno;
429 vc4_update_bo_seqnos(exec, seqno);
431 list_add_tail(&exec->head, &vc4->job_list);
433 /* If no job was executing, kick ours off. Otherwise, it'll
434 * get started when the previous job's frame done interrupt
437 if (vc4_first_job(vc4) == exec) {
438 vc4_submit_next_job(dev);
439 vc4_queue_hangcheck(dev);
442 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
446 * Looks up a bunch of GEM handles for BOs and stores the array for
447 * use in the command validator that actually writes relocated
448 * addresses pointing to them.
451 vc4_cl_lookup_bos(struct drm_device *dev,
452 struct drm_file *file_priv,
453 struct vc4_exec_info *exec)
455 struct drm_vc4_submit_cl *args = exec->args;
460 exec->bo_count = args->bo_handle_count;
462 if (!exec->bo_count) {
463 /* See comment on bo_index for why we have to check
466 DRM_ERROR("Rendering requires BOs to validate\n");
470 exec->bo = kcalloc(exec->bo_count, sizeof(struct drm_gem_cma_object *),
473 DRM_ERROR("Failed to allocate validated BO pointers\n");
477 handles = drm_malloc_ab(exec->bo_count, sizeof(uint32_t));
479 DRM_ERROR("Failed to allocate incoming GEM handles\n");
483 ret = copy_from_user(handles,
484 (void __user *)(uintptr_t)args->bo_handles,
485 exec->bo_count * sizeof(uint32_t));
487 DRM_ERROR("Failed to copy in GEM handles\n");
491 spin_lock(&file_priv->table_lock);
492 for (i = 0; i < exec->bo_count; i++) {
493 struct drm_gem_object *bo = idr_find(&file_priv->object_idr,
496 DRM_ERROR("Failed to look up GEM BO %d: %d\n",
499 spin_unlock(&file_priv->table_lock);
502 drm_gem_object_reference(bo);
503 exec->bo[i] = (struct drm_gem_cma_object *)bo;
505 spin_unlock(&file_priv->table_lock);
513 vc4_get_bcl(struct drm_device *dev, struct vc4_exec_info *exec)
515 struct drm_vc4_submit_cl *args = exec->args;
519 uint32_t bin_offset = 0;
520 uint32_t shader_rec_offset = roundup(bin_offset + args->bin_cl_size,
522 uint32_t uniforms_offset = shader_rec_offset + args->shader_rec_size;
523 uint32_t exec_size = uniforms_offset + args->uniforms_size;
524 uint32_t temp_size = exec_size + (sizeof(struct vc4_shader_state) *
525 args->shader_rec_count);
528 if (uniforms_offset < shader_rec_offset ||
529 exec_size < uniforms_offset ||
530 args->shader_rec_count >= (UINT_MAX /
531 sizeof(struct vc4_shader_state)) ||
532 temp_size < exec_size) {
533 DRM_ERROR("overflow in exec arguments\n");
537 /* Allocate space where we'll store the copied in user command lists
538 * and shader records.
540 * We don't just copy directly into the BOs because we need to
541 * read the contents back for validation, and I think the
542 * bo->vaddr is uncached access.
544 temp = kmalloc(temp_size, GFP_KERNEL);
546 DRM_ERROR("Failed to allocate storage for copying "
547 "in bin/render CLs.\n");
551 bin = temp + bin_offset;
552 exec->shader_rec_u = temp + shader_rec_offset;
553 exec->uniforms_u = temp + uniforms_offset;
554 exec->shader_state = temp + exec_size;
555 exec->shader_state_size = args->shader_rec_count;
557 if (copy_from_user(bin,
558 (void __user *)(uintptr_t)args->bin_cl,
559 args->bin_cl_size)) {
564 if (copy_from_user(exec->shader_rec_u,
565 (void __user *)(uintptr_t)args->shader_rec,
566 args->shader_rec_size)) {
571 if (copy_from_user(exec->uniforms_u,
572 (void __user *)(uintptr_t)args->uniforms,
573 args->uniforms_size)) {
578 bo = vc4_bo_create(dev, exec_size, true);
580 DRM_ERROR("Couldn't allocate BO for binning\n");
584 exec->exec_bo = &bo->base;
586 list_add_tail(&to_vc4_bo(&exec->exec_bo->base)->unref_head,
589 exec->ct0ca = exec->exec_bo->paddr + bin_offset;
593 exec->shader_rec_v = exec->exec_bo->vaddr + shader_rec_offset;
594 exec->shader_rec_p = exec->exec_bo->paddr + shader_rec_offset;
595 exec->shader_rec_size = args->shader_rec_size;
597 exec->uniforms_v = exec->exec_bo->vaddr + uniforms_offset;
598 exec->uniforms_p = exec->exec_bo->paddr + uniforms_offset;
599 exec->uniforms_size = args->uniforms_size;
601 ret = vc4_validate_bin_cl(dev,
602 exec->exec_bo->vaddr + bin_offset,
608 ret = vc4_validate_shader_recs(dev, exec);
616 vc4_complete_exec(struct drm_device *dev, struct vc4_exec_info *exec)
620 /* Need the struct lock for drm_gem_object_unreference(). */
621 mutex_lock(&dev->struct_mutex);
623 for (i = 0; i < exec->bo_count; i++)
624 drm_gem_object_unreference(&exec->bo[i]->base);
628 while (!list_empty(&exec->unref_list)) {
629 struct vc4_bo *bo = list_first_entry(&exec->unref_list,
630 struct vc4_bo, unref_head);
631 list_del(&bo->unref_head);
632 drm_gem_object_unreference(&bo->base.base);
634 mutex_unlock(&dev->struct_mutex);
640 vc4_job_handle_completed(struct vc4_dev *vc4)
642 unsigned long irqflags;
643 struct vc4_seqno_cb *cb, *cb_temp;
645 spin_lock_irqsave(&vc4->job_lock, irqflags);
646 while (!list_empty(&vc4->job_done_list)) {
647 struct vc4_exec_info *exec =
648 list_first_entry(&vc4->job_done_list,
649 struct vc4_exec_info, head);
650 list_del(&exec->head);
652 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
653 vc4_complete_exec(vc4->dev, exec);
654 spin_lock_irqsave(&vc4->job_lock, irqflags);
657 list_for_each_entry_safe(cb, cb_temp, &vc4->seqno_cb_list, work.entry) {
658 if (cb->seqno <= vc4->finished_seqno) {
659 list_del_init(&cb->work.entry);
660 schedule_work(&cb->work);
664 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
667 static void vc4_seqno_cb_work(struct work_struct *work)
669 struct vc4_seqno_cb *cb = container_of(work, struct vc4_seqno_cb, work);
674 int vc4_queue_seqno_cb(struct drm_device *dev,
675 struct vc4_seqno_cb *cb, uint64_t seqno,
676 void (*func)(struct vc4_seqno_cb *cb))
678 struct vc4_dev *vc4 = to_vc4_dev(dev);
680 unsigned long irqflags;
683 INIT_WORK(&cb->work, vc4_seqno_cb_work);
685 spin_lock_irqsave(&vc4->job_lock, irqflags);
686 if (seqno > vc4->finished_seqno) {
688 list_add_tail(&cb->work.entry, &vc4->seqno_cb_list);
690 schedule_work(&cb->work);
692 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
697 /* Scheduled when any job has been completed, this walks the list of
698 * jobs that had completed and unrefs their BOs and frees their exec
702 vc4_job_done_work(struct work_struct *work)
704 struct vc4_dev *vc4 =
705 container_of(work, struct vc4_dev, job_done_work);
707 vc4_job_handle_completed(vc4);
711 vc4_wait_for_seqno_ioctl_helper(struct drm_device *dev,
713 uint64_t *timeout_ns)
715 unsigned long start = jiffies;
716 int ret = vc4_wait_for_seqno(dev, seqno, *timeout_ns, true);
718 if ((ret == -EINTR || ret == -ERESTARTSYS) && *timeout_ns != ~0ull) {
719 uint64_t delta = jiffies_to_nsecs(jiffies - start);
721 if (*timeout_ns >= delta)
722 *timeout_ns -= delta;
729 vc4_wait_seqno_ioctl(struct drm_device *dev, void *data,
730 struct drm_file *file_priv)
732 struct drm_vc4_wait_seqno *args = data;
734 return vc4_wait_for_seqno_ioctl_helper(dev, args->seqno,
739 vc4_wait_bo_ioctl(struct drm_device *dev, void *data,
740 struct drm_file *file_priv)
743 struct drm_vc4_wait_bo *args = data;
744 struct drm_gem_object *gem_obj;
750 gem_obj = drm_gem_object_lookup(dev, file_priv, args->handle);
752 DRM_ERROR("Failed to look up GEM BO %d\n", args->handle);
755 bo = to_vc4_bo(gem_obj);
757 ret = vc4_wait_for_seqno_ioctl_helper(dev, bo->seqno,
760 drm_gem_object_unreference_unlocked(gem_obj);
765 * Submits a command list to the VC4.
767 * This is what is called batchbuffer emitting on other hardware.
770 vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
771 struct drm_file *file_priv)
773 struct vc4_dev *vc4 = to_vc4_dev(dev);
774 struct drm_vc4_submit_cl *args = data;
775 struct vc4_exec_info *exec;
778 if ((args->flags & ~VC4_SUBMIT_CL_USE_CLEAR_COLOR) != 0) {
779 DRM_ERROR("Unknown flags: 0x%02x\n", args->flags);
783 exec = kcalloc(1, sizeof(*exec), GFP_KERNEL);
785 DRM_ERROR("malloc failure on exec struct\n");
790 INIT_LIST_HEAD(&exec->unref_list);
792 ret = vc4_cl_lookup_bos(dev, file_priv, exec);
796 if (exec->args->bin_cl_size != 0) {
797 ret = vc4_get_bcl(dev, exec);
805 ret = vc4_get_rcl(dev, exec);
809 /* Clear this out of the struct we'll be putting in the queue,
810 * since it's part of our stack.
814 vc4_queue_submit(dev, exec);
816 /* Return the seqno for our job. */
817 args->seqno = vc4->emit_seqno;
822 vc4_complete_exec(vc4->dev, exec);
828 vc4_gem_init(struct drm_device *dev)
830 struct vc4_dev *vc4 = to_vc4_dev(dev);
832 INIT_LIST_HEAD(&vc4->job_list);
833 INIT_LIST_HEAD(&vc4->job_done_list);
834 INIT_LIST_HEAD(&vc4->seqno_cb_list);
835 spin_lock_init(&vc4->job_lock);
837 INIT_WORK(&vc4->hangcheck.reset_work, vc4_reset_work);
838 setup_timer(&vc4->hangcheck.timer,
839 vc4_hangcheck_elapsed,
842 INIT_WORK(&vc4->job_done_work, vc4_job_done_work);
846 vc4_gem_destroy(struct drm_device *dev)
848 struct vc4_dev *vc4 = to_vc4_dev(dev);
850 /* Waiting for exec to finish would need to be done before
853 WARN_ON(vc4->emit_seqno != vc4->finished_seqno);
855 /* V3D should already have disabled its interrupt and cleared
856 * the overflow allocation registers. Now free the object.
858 if (vc4->overflow_mem) {
859 drm_gem_object_unreference_unlocked(&vc4->overflow_mem->base.base);
860 vc4->overflow_mem = NULL;
863 vc4_bo_cache_destroy(dev);
866 vc4_free_hang_state(dev, vc4->hang_state);