Adding upstream version 6.1.76.upstream/6.1.76

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

1 files changed, 1443 insertions, 0 deletions

diff --git a/drivers/gpu/drm/vc4/vc4_gem.c b/drivers/gpu/drm/vc4/vc4_gem.c
new file mode 100644
index 000000000..628d40ff3
--- /dev/null
+++ b/drivers/gpu/drm/vc4/vc4_gem.c
@@ -0,0 +1,1443 @@
+/*
+ * Copyright © 2014 Broadcom
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <linux/module.h>
+#include <linux/platform_device.h>
+#include <linux/pm_runtime.h>
+#include <linux/device.h>
+#include <linux/io.h>
+#include <linux/sched/signal.h>
+#include <linux/dma-fence-array.h>
+
+#include <drm/drm_syncobj.h>
+
+#include "uapi/drm/vc4_drm.h"
+#include "vc4_drv.h"
+#include "vc4_regs.h"
+#include "vc4_trace.h"
+
+static void
+vc4_queue_hangcheck(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	mod_timer(&vc4->hangcheck.timer,
+		  round_jiffies_up(jiffies + msecs_to_jiffies(100)));
+}
+
+struct vc4_hang_state {
+	struct drm_vc4_get_hang_state user_state;
+
+	u32 bo_count;
+	struct drm_gem_object **bo;
+};
+
+static void
+vc4_free_hang_state(struct drm_device *dev, struct vc4_hang_state *state)
+{
+	unsigned int i;
+
+	for (i = 0; i < state->user_state.bo_count; i++)
+		drm_gem_object_put(state->bo[i]);
+
+	kfree(state);
+}
+
+int
+vc4_get_hang_state_ioctl(struct drm_device *dev, void *data,
+			 struct drm_file *file_priv)
+{
+	struct drm_vc4_get_hang_state *get_state = data;
+	struct drm_vc4_get_hang_state_bo *bo_state;
+	struct vc4_hang_state *kernel_state;
+	struct drm_vc4_get_hang_state *state;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	unsigned long irqflags;
+	u32 i;
+	int ret = 0;
+
+	if (WARN_ON_ONCE(vc4->is_vc5))
+		return -ENODEV;
+
+	if (!vc4->v3d) {
+		DRM_DEBUG("VC4_GET_HANG_STATE with no VC4 V3D probed\n");
+		return -ENODEV;
+	}
+
+	spin_lock_irqsave(&vc4->job_lock, irqflags);
+	kernel_state = vc4->hang_state;
+	if (!kernel_state) {
+		spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+		return -ENOENT;
+	}
+	state = &kernel_state->user_state;
+
+	/* If the user's array isn't big enough, just return the
+	 * required array size.
+	 */
+	if (get_state->bo_count < state->bo_count) {
+		get_state->bo_count = state->bo_count;
+		spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+		return 0;
+	}
+
+	vc4->hang_state = NULL;
+	spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+
+	/* Save the user's BO pointer, so we don't stomp it with the memcpy. */
+	state->bo = get_state->bo;
+	memcpy(get_state, state, sizeof(*state));
+
+	bo_state = kcalloc(state->bo_count, sizeof(*bo_state), GFP_KERNEL);
+	if (!bo_state) {
+		ret = -ENOMEM;
+		goto err_free;
+	}
+
+	for (i = 0; i < state->bo_count; i++) {
+		struct vc4_bo *vc4_bo = to_vc4_bo(kernel_state->bo[i]);
+		u32 handle;
+
+		ret = drm_gem_handle_create(file_priv, kernel_state->bo[i],
+					    &handle);
+
+		if (ret) {
+			state->bo_count = i;
+			goto err_delete_handle;
+		}
+		bo_state[i].handle = handle;
+		bo_state[i].paddr = vc4_bo->base.dma_addr;
+		bo_state[i].size = vc4_bo->base.base.size;
+	}
+
+	if (copy_to_user(u64_to_user_ptr(get_state->bo),
+			 bo_state,
+			 state->bo_count * sizeof(*bo_state)))
+		ret = -EFAULT;
+
+err_delete_handle:
+	if (ret) {
+		for (i = 0; i < state->bo_count; i++)
+			drm_gem_handle_delete(file_priv, bo_state[i].handle);
+	}
+
+err_free:
+	vc4_free_hang_state(dev, kernel_state);
+	kfree(bo_state);
+
+	return ret;
+}
+
+static void
+vc4_save_hang_state(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct drm_vc4_get_hang_state *state;
+	struct vc4_hang_state *kernel_state;
+	struct vc4_exec_info *exec[2];
+	struct vc4_bo *bo;
+	unsigned long irqflags;
+	unsigned int i, j, k, unref_list_count;
+
+	kernel_state = kcalloc(1, sizeof(*kernel_state), GFP_KERNEL);
+	if (!kernel_state)
+		return;
+
+	state = &kernel_state->user_state;
+
+	spin_lock_irqsave(&vc4->job_lock, irqflags);
+	exec[0] = vc4_first_bin_job(vc4);
+	exec[1] = vc4_first_render_job(vc4);
+	if (!exec[0] && !exec[1]) {
+		spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+		return;
+	}
+
+	/* Get the bos from both binner and renderer into hang state. */
+	state->bo_count = 0;
+	for (i = 0; i < 2; i++) {
+		if (!exec[i])
+			continue;
+
+		unref_list_count = 0;
+		list_for_each_entry(bo, &exec[i]->unref_list, unref_head)
+			unref_list_count++;
+		state->bo_count += exec[i]->bo_count + unref_list_count;
+	}
+
+	kernel_state->bo = kcalloc(state->bo_count,
+				   sizeof(*kernel_state->bo), GFP_ATOMIC);
+
+	if (!kernel_state->bo) {
+		spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+		return;
+	}
+
+	k = 0;
+	for (i = 0; i < 2; i++) {
+		if (!exec[i])
+			continue;
+
+		for (j = 0; j < exec[i]->bo_count; j++) {
+			bo = to_vc4_bo(&exec[i]->bo[j]->base);
+
+			/* Retain BOs just in case they were marked purgeable.
+			 * This prevents the BO from being purged before
+			 * someone had a chance to dump the hang state.
+			 */
+			WARN_ON(!refcount_read(&bo->usecnt));
+			refcount_inc(&bo->usecnt);
+			drm_gem_object_get(&exec[i]->bo[j]->base);
+			kernel_state->bo[k++] = &exec[i]->bo[j]->base;
+		}
+
+		list_for_each_entry(bo, &exec[i]->unref_list, unref_head) {
+			/* No need to retain BOs coming from the ->unref_list
+			 * because they are naturally unpurgeable.
+			 */
+			drm_gem_object_get(&bo->base.base);
+			kernel_state->bo[k++] = &bo->base.base;
+		}
+	}
+
+	WARN_ON_ONCE(k != state->bo_count);
+
+	if (exec[0])
+		state->start_bin = exec[0]->ct0ca;
+	if (exec[1])
+		state->start_render = exec[1]->ct1ca;
+
+	spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+
+	state->ct0ca = V3D_READ(V3D_CTNCA(0));
+	state->ct0ea = V3D_READ(V3D_CTNEA(0));
+
+	state->ct1ca = V3D_READ(V3D_CTNCA(1));
+	state->ct1ea = V3D_READ(V3D_CTNEA(1));
+
+	state->ct0cs = V3D_READ(V3D_CTNCS(0));
+	state->ct1cs = V3D_READ(V3D_CTNCS(1));
+
+	state->ct0ra0 = V3D_READ(V3D_CT00RA0);
+	state->ct1ra0 = V3D_READ(V3D_CT01RA0);
+
+	state->bpca = V3D_READ(V3D_BPCA);
+	state->bpcs = V3D_READ(V3D_BPCS);
+	state->bpoa = V3D_READ(V3D_BPOA);
+	state->bpos = V3D_READ(V3D_BPOS);
+
+	state->vpmbase = V3D_READ(V3D_VPMBASE);
+
+	state->dbge = V3D_READ(V3D_DBGE);
+	state->fdbgo = V3D_READ(V3D_FDBGO);
+	state->fdbgb = V3D_READ(V3D_FDBGB);
+	state->fdbgr = V3D_READ(V3D_FDBGR);
+	state->fdbgs = V3D_READ(V3D_FDBGS);
+	state->errstat = V3D_READ(V3D_ERRSTAT);
+
+	/* We need to turn purgeable BOs into unpurgeable ones so that
+	 * userspace has a chance to dump the hang state before the kernel
+	 * decides to purge those BOs.
+	 * Note that BO consistency at dump time cannot be guaranteed. For
+	 * example, if the owner of these BOs decides to re-use them or mark
+	 * them purgeable again there's nothing we can do to prevent it.
+	 */
+	for (i = 0; i < kernel_state->user_state.bo_count; i++) {
+		struct vc4_bo *bo = to_vc4_bo(kernel_state->bo[i]);
+
+		if (bo->madv == __VC4_MADV_NOTSUPP)
+			continue;
+
+		mutex_lock(&bo->madv_lock);
+		if (!WARN_ON(bo->madv == __VC4_MADV_PURGED))
+			bo->madv = VC4_MADV_WILLNEED;
+		refcount_dec(&bo->usecnt);
+		mutex_unlock(&bo->madv_lock);
+	}
+
+	spin_lock_irqsave(&vc4->job_lock, irqflags);
+	if (vc4->hang_state) {
+		spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+		vc4_free_hang_state(dev, kernel_state);
+	} else {
+		vc4->hang_state = kernel_state;
+		spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+	}
+}
+
+static void
+vc4_reset(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	DRM_INFO("Resetting GPU.\n");
+
+	mutex_lock(&vc4->power_lock);
+	if (vc4->power_refcount) {
+		/* Power the device off and back on the by dropping the
+		 * reference on runtime PM.
+		 */
+		pm_runtime_put_sync_suspend(&vc4->v3d->pdev->dev);
+		pm_runtime_get_sync(&vc4->v3d->pdev->dev);
+	}
+	mutex_unlock(&vc4->power_lock);
+
+	vc4_irq_reset(dev);
+
+	/* Rearm the hangcheck -- another job might have been waiting
+	 * for our hung one to get kicked off, and vc4_irq_reset()
+	 * would have started it.
+	 */
+	vc4_queue_hangcheck(dev);
+}
+
+static void
+vc4_reset_work(struct work_struct *work)
+{
+	struct vc4_dev *vc4 =
+		container_of(work, struct vc4_dev, hangcheck.reset_work);
+
+	vc4_save_hang_state(&vc4->base);
+
+	vc4_reset(&vc4->base);
+}
+
+static void
+vc4_hangcheck_elapsed(struct timer_list *t)
+{
+	struct vc4_dev *vc4 = from_timer(vc4, t, hangcheck.timer);
+	struct drm_device *dev = &vc4->base;
+	uint32_t ct0ca, ct1ca;
+	unsigned long irqflags;
+	struct vc4_exec_info *bin_exec, *render_exec;
+
+	spin_lock_irqsave(&vc4->job_lock, irqflags);
+
+	bin_exec = vc4_first_bin_job(vc4);
+	render_exec = vc4_first_render_job(vc4);
+
+	/* If idle, we can stop watching for hangs. */
+	if (!bin_exec && !render_exec) {
+		spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+		return;
+	}
+
+	ct0ca = V3D_READ(V3D_CTNCA(0));
+	ct1ca = V3D_READ(V3D_CTNCA(1));
+
+	/* If we've made any progress in execution, rearm the timer
+	 * and wait.
+	 */
+	if ((bin_exec && ct0ca != bin_exec->last_ct0ca) ||
+	    (render_exec && ct1ca != render_exec->last_ct1ca)) {
+		if (bin_exec)
+			bin_exec->last_ct0ca = ct0ca;
+		if (render_exec)
+			render_exec->last_ct1ca = ct1ca;
+		spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+		vc4_queue_hangcheck(dev);
+		return;
+	}
+
+	spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+
+	/* We've gone too long with no progress, reset.  This has to
+	 * be done from a work struct, since resetting can sleep and
+	 * this timer hook isn't allowed to.
+	 */
+	schedule_work(&vc4->hangcheck.reset_work);
+}
+
+static void
+submit_cl(struct drm_device *dev, uint32_t thread, uint32_t start, uint32_t end)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	/* Set the current and end address of the control list.
+	 * Writing the end register is what starts the job.
+	 */
+	V3D_WRITE(V3D_CTNCA(thread), start);
+	V3D_WRITE(V3D_CTNEA(thread), end);
+}
+
+int
+vc4_wait_for_seqno(struct drm_device *dev, uint64_t seqno, uint64_t timeout_ns,
+		   bool interruptible)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	int ret = 0;
+	unsigned long timeout_expire;
+	DEFINE_WAIT(wait);
+
+	if (WARN_ON_ONCE(vc4->is_vc5))
+		return -ENODEV;
+
+	if (vc4->finished_seqno >= seqno)
+		return 0;
+
+	if (timeout_ns == 0)
+		return -ETIME;
+
+	timeout_expire = jiffies + nsecs_to_jiffies(timeout_ns);
+
+	trace_vc4_wait_for_seqno_begin(dev, seqno, timeout_ns);
+	for (;;) {
+		prepare_to_wait(&vc4->job_wait_queue, &wait,
+				interruptible ? TASK_INTERRUPTIBLE :
+				TASK_UNINTERRUPTIBLE);
+
+		if (interruptible && signal_pending(current)) {
+			ret = -ERESTARTSYS;
+			break;
+		}
+
+		if (vc4->finished_seqno >= seqno)
+			break;
+
+		if (timeout_ns != ~0ull) {
+			if (time_after_eq(jiffies, timeout_expire)) {
+				ret = -ETIME;
+				break;
+			}
+			schedule_timeout(timeout_expire - jiffies);
+		} else {
+			schedule();
+		}
+	}
+
+	finish_wait(&vc4->job_wait_queue, &wait);
+	trace_vc4_wait_for_seqno_end(dev, seqno);
+
+	return ret;
+}
+
+static void
+vc4_flush_caches(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	/* Flush the GPU L2 caches.  These caches sit on top of system
+	 * L3 (the 128kb or so shared with the CPU), and are
+	 * non-allocating in the L3.
+	 */
+	V3D_WRITE(V3D_L2CACTL,
+		  V3D_L2CACTL_L2CCLR);
+
+	V3D_WRITE(V3D_SLCACTL,
+		  VC4_SET_FIELD(0xf, V3D_SLCACTL_T1CC) |
+		  VC4_SET_FIELD(0xf, V3D_SLCACTL_T0CC) |
+		  VC4_SET_FIELD(0xf, V3D_SLCACTL_UCC) |
+		  VC4_SET_FIELD(0xf, V3D_SLCACTL_ICC));
+}
+
+static void
+vc4_flush_texture_caches(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	V3D_WRITE(V3D_L2CACTL,
+		  V3D_L2CACTL_L2CCLR);
+
+	V3D_WRITE(V3D_SLCACTL,
+		  VC4_SET_FIELD(0xf, V3D_SLCACTL_T1CC) |
+		  VC4_SET_FIELD(0xf, V3D_SLCACTL_T0CC));
+}
+
+/* Sets the registers for the next job to be actually be executed in
+ * the hardware.
+ *
+ * The job_lock should be held during this.
+ */
+void
+vc4_submit_next_bin_job(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct vc4_exec_info *exec;
+
+	if (WARN_ON_ONCE(vc4->is_vc5))
+		return;
+
+again:
+	exec = vc4_first_bin_job(vc4);
+	if (!exec)
+		return;
+
+	vc4_flush_caches(dev);
+
+	/* Only start the perfmon if it was not already started by a previous
+	 * job.
+	 */
+	if (exec->perfmon && vc4->active_perfmon != exec->perfmon)
+		vc4_perfmon_start(vc4, exec->perfmon);
+
+	/* Either put the job in the binner if it uses the binner, or
+	 * immediately move it to the to-be-rendered queue.
+	 */
+	if (exec->ct0ca != exec->ct0ea) {
+		trace_vc4_submit_cl(dev, false, exec->seqno, exec->ct0ca,
+				    exec->ct0ea);
+		submit_cl(dev, 0, exec->ct0ca, exec->ct0ea);
+	} else {
+		struct vc4_exec_info *next;
+
+		vc4_move_job_to_render(dev, exec);
+		next = vc4_first_bin_job(vc4);
+
+		/* We can't start the next bin job if the previous job had a
+		 * different perfmon instance attached to it. The same goes
+		 * if one of them had a perfmon attached to it and the other
+		 * one doesn't.
+		 */
+		if (next && next->perfmon == exec->perfmon)
+			goto again;
+	}
+}
+
+void
+vc4_submit_next_render_job(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct vc4_exec_info *exec = vc4_first_render_job(vc4);
+
+	if (!exec)
+		return;
+
+	if (WARN_ON_ONCE(vc4->is_vc5))
+		return;
+
+	/* A previous RCL may have written to one of our textures, and
+	 * our full cache flush at bin time may have occurred before
+	 * that RCL completed.  Flush the texture cache now, but not
+	 * the instructions or uniforms (since we don't write those
+	 * from an RCL).
+	 */
+	vc4_flush_texture_caches(dev);
+
+	trace_vc4_submit_cl(dev, true, exec->seqno, exec->ct1ca, exec->ct1ea);
+	submit_cl(dev, 1, exec->ct1ca, exec->ct1ea);
+}
+
+void
+vc4_move_job_to_render(struct drm_device *dev, struct vc4_exec_info *exec)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	bool was_empty = list_empty(&vc4->render_job_list);
+
+	if (WARN_ON_ONCE(vc4->is_vc5))
+		return;
+
+	list_move_tail(&exec->head, &vc4->render_job_list);
+	if (was_empty)
+		vc4_submit_next_render_job(dev);
+}
+
+static void
+vc4_update_bo_seqnos(struct vc4_exec_info *exec, uint64_t seqno)
+{
+	struct vc4_bo *bo;
+	unsigned i;
+
+	for (i = 0; i < exec->bo_count; i++) {
+		bo = to_vc4_bo(&exec->bo[i]->base);
+		bo->seqno = seqno;
+
+		dma_resv_add_fence(bo->base.base.resv, exec->fence,
+				   DMA_RESV_USAGE_READ);
+	}
+
+	list_for_each_entry(bo, &exec->unref_list, unref_head) {
+		bo->seqno = seqno;
+	}
+
+	for (i = 0; i < exec->rcl_write_bo_count; i++) {
+		bo = to_vc4_bo(&exec->rcl_write_bo[i]->base);
+		bo->write_seqno = seqno;
+
+		dma_resv_add_fence(bo->base.base.resv, exec->fence,
+				   DMA_RESV_USAGE_WRITE);
+	}
+}
+
+static void
+vc4_unlock_bo_reservations(struct drm_device *dev,
+			   struct vc4_exec_info *exec,
+			   struct ww_acquire_ctx *acquire_ctx)
+{
+	int i;
+
+	for (i = 0; i < exec->bo_count; i++) {
+		struct drm_gem_object *bo = &exec->bo[i]->base;
+
+		dma_resv_unlock(bo->resv);
+	}
+
+	ww_acquire_fini(acquire_ctx);
+}
+
+/* Takes the reservation lock on all the BOs being referenced, so that
+ * at queue submit time we can update the reservations.
+ *
+ * We don't lock the RCL the tile alloc/state BOs, or overflow memory
+ * (all of which are on exec->unref_list).  They're entirely private
+ * to vc4, so we don't attach dma-buf fences to them.
+ */
+static int
+vc4_lock_bo_reservations(struct drm_device *dev,
+			 struct vc4_exec_info *exec,
+			 struct ww_acquire_ctx *acquire_ctx)
+{
+	int contended_lock = -1;
+	int i, ret;
+	struct drm_gem_object *bo;
+
+	ww_acquire_init(acquire_ctx, &reservation_ww_class);
+
+retry:
+	if (contended_lock != -1) {
+		bo = &exec->bo[contended_lock]->base;
+		ret = dma_resv_lock_slow_interruptible(bo->resv, acquire_ctx);
+		if (ret) {
+			ww_acquire_done(acquire_ctx);
+			return ret;
+		}
+	}
+
+	for (i = 0; i < exec->bo_count; i++) {
+		if (i == contended_lock)
+			continue;
+
+		bo = &exec->bo[i]->base;
+
+		ret = dma_resv_lock_interruptible(bo->resv, acquire_ctx);
+		if (ret) {
+			int j;
+
+			for (j = 0; j < i; j++) {
+				bo = &exec->bo[j]->base;
+				dma_resv_unlock(bo->resv);
+			}
+
+			if (contended_lock != -1 && contended_lock >= i) {
+				bo = &exec->bo[contended_lock]->base;
+
+				dma_resv_unlock(bo->resv);
+			}
+
+			if (ret == -EDEADLK) {
+				contended_lock = i;
+				goto retry;
+			}
+
+			ww_acquire_done(acquire_ctx);
+			return ret;
+		}
+	}
+
+	ww_acquire_done(acquire_ctx);
+
+	/* Reserve space for our shared (read-only) fence references,
+	 * before we commit the CL to the hardware.
+	 */
+	for (i = 0; i < exec->bo_count; i++) {
+		bo = &exec->bo[i]->base;
+
+		ret = dma_resv_reserve_fences(bo->resv, 1);
+		if (ret) {
+			vc4_unlock_bo_reservations(dev, exec, acquire_ctx);
+			return ret;
+		}
+	}
+
+	return 0;
+}
+
+/* Queues a struct vc4_exec_info for execution.  If no job is
+ * currently executing, then submits it.
+ *
+ * Unlike most GPUs, our hardware only handles one command list at a
+ * time.  To queue multiple jobs at once, we'd need to edit the
+ * previous command list to have a jump to the new one at the end, and
+ * then bump the end address.  That's a change for a later date,
+ * though.
+ */
+static int
+vc4_queue_submit(struct drm_device *dev, struct vc4_exec_info *exec,
+		 struct ww_acquire_ctx *acquire_ctx,
+		 struct drm_syncobj *out_sync)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct vc4_exec_info *renderjob;
+	uint64_t seqno;
+	unsigned long irqflags;
+	struct vc4_fence *fence;
+
+	fence = kzalloc(sizeof(*fence), GFP_KERNEL);
+	if (!fence)
+		return -ENOMEM;
+	fence->dev = dev;
+
+	spin_lock_irqsave(&vc4->job_lock, irqflags);
+
+	seqno = ++vc4->emit_seqno;
+	exec->seqno = seqno;
+
+	dma_fence_init(&fence->base, &vc4_fence_ops, &vc4->job_lock,
+		       vc4->dma_fence_context, exec->seqno);
+	fence->seqno = exec->seqno;
+	exec->fence = &fence->base;
+
+	if (out_sync)
+		drm_syncobj_replace_fence(out_sync, exec->fence);
+
+	vc4_update_bo_seqnos(exec, seqno);
+
+	vc4_unlock_bo_reservations(dev, exec, acquire_ctx);
+
+	list_add_tail(&exec->head, &vc4->bin_job_list);
+
+	/* If no bin job was executing and if the render job (if any) has the
+	 * same perfmon as our job attached to it (or if both jobs don't have
+	 * perfmon activated), then kick ours off.  Otherwise, it'll get
+	 * started when the previous job's flush/render done interrupt occurs.
+	 */
+	renderjob = vc4_first_render_job(vc4);
+	if (vc4_first_bin_job(vc4) == exec &&
+	    (!renderjob || renderjob->perfmon == exec->perfmon)) {
+		vc4_submit_next_bin_job(dev);
+		vc4_queue_hangcheck(dev);
+	}
+
+	spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+
+	return 0;
+}
+
+/**
+ * vc4_cl_lookup_bos() - Sets up exec->bo[] with the GEM objects
+ * referenced by the job.
+ * @dev: DRM device
+ * @file_priv: DRM file for this fd
+ * @exec: V3D job being set up
+ *
+ * The command validator needs to reference BOs by their index within
+ * the submitted job's BO list.  This does the validation of the job's
+ * BO list and reference counting for the lifetime of the job.
+ */
+static int
+vc4_cl_lookup_bos(struct drm_device *dev,
+		  struct drm_file *file_priv,
+		  struct vc4_exec_info *exec)
+{
+	struct drm_vc4_submit_cl *args = exec->args;
+	uint32_t *handles;
+	int ret = 0;
+	int i;
+
+	exec->bo_count = args->bo_handle_count;
+
+	if (!exec->bo_count) {
+		/* See comment on bo_index for why we have to check
+		 * this.
+		 */
+		DRM_DEBUG("Rendering requires BOs to validate\n");
+		return -EINVAL;
+	}
+
+	exec->bo = kvmalloc_array(exec->bo_count,
+				    sizeof(struct drm_gem_dma_object *),
+				    GFP_KERNEL | __GFP_ZERO);
+	if (!exec->bo) {
+		DRM_ERROR("Failed to allocate validated BO pointers\n");
+		return -ENOMEM;
+	}
+
+	handles = kvmalloc_array(exec->bo_count, sizeof(uint32_t), GFP_KERNEL);
+	if (!handles) {
+		ret = -ENOMEM;
+		DRM_ERROR("Failed to allocate incoming GEM handles\n");
+		goto fail;
+	}
+
+	if (copy_from_user(handles, u64_to_user_ptr(args->bo_handles),
+			   exec->bo_count * sizeof(uint32_t))) {
+		ret = -EFAULT;
+		DRM_ERROR("Failed to copy in GEM handles\n");
+		goto fail;
+	}
+
+	spin_lock(&file_priv->table_lock);
+	for (i = 0; i < exec->bo_count; i++) {
+		struct drm_gem_object *bo = idr_find(&file_priv->object_idr,
+						     handles[i]);
+		if (!bo) {
+			DRM_DEBUG("Failed to look up GEM BO %d: %d\n",
+				  i, handles[i]);
+			ret = -EINVAL;
+			break;
+		}
+
+		drm_gem_object_get(bo);
+		exec->bo[i] = (struct drm_gem_dma_object *)bo;
+	}
+	spin_unlock(&file_priv->table_lock);
+
+	if (ret)
+		goto fail_put_bo;
+
+	for (i = 0; i < exec->bo_count; i++) {
+		ret = vc4_bo_inc_usecnt(to_vc4_bo(&exec->bo[i]->base));
+		if (ret)
+			goto fail_dec_usecnt;
+	}
+
+	kvfree(handles);
+	return 0;
+
+fail_dec_usecnt:
+	/* Decrease usecnt on acquired objects.
+	 * We cannot rely on  vc4_complete_exec() to release resources here,
+	 * because vc4_complete_exec() has no information about which BO has
+	 * had its ->usecnt incremented.
+	 * To make things easier we just free everything explicitly and set
+	 * exec->bo to NULL so that vc4_complete_exec() skips the 'BO release'
+	 * step.
+	 */
+	for (i-- ; i >= 0; i--)
+		vc4_bo_dec_usecnt(to_vc4_bo(&exec->bo[i]->base));
+
+fail_put_bo:
+	/* Release any reference to acquired objects. */
+	for (i = 0; i < exec->bo_count && exec->bo[i]; i++)
+		drm_gem_object_put(&exec->bo[i]->base);
+
+fail:
+	kvfree(handles);
+	kvfree(exec->bo);
+	exec->bo = NULL;
+	return ret;
+}
+
+static int
+vc4_get_bcl(struct drm_device *dev, struct vc4_exec_info *exec)
+{
+	struct drm_vc4_submit_cl *args = exec->args;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	void *temp = NULL;
+	void *bin;
+	int ret = 0;
+	uint32_t bin_offset = 0;
+	uint32_t shader_rec_offset = roundup(bin_offset + args->bin_cl_size,
+					     16);
+	uint32_t uniforms_offset = shader_rec_offset + args->shader_rec_size;
+	uint32_t exec_size = uniforms_offset + args->uniforms_size;
+	uint32_t temp_size = exec_size + (sizeof(struct vc4_shader_state) *
+					  args->shader_rec_count);
+	struct vc4_bo *bo;
+
+	if (shader_rec_offset < args->bin_cl_size ||
+	    uniforms_offset < shader_rec_offset ||
+	    exec_size < uniforms_offset ||
+	    args->shader_rec_count >= (UINT_MAX /
+					  sizeof(struct vc4_shader_state)) ||
+	    temp_size < exec_size) {
+		DRM_DEBUG("overflow in exec arguments\n");
+		ret = -EINVAL;
+		goto fail;
+	}
+
+	/* Allocate space where we'll store the copied in user command lists
+	 * and shader records.
+	 *
+	 * We don't just copy directly into the BOs because we need to
+	 * read the contents back for validation, and I think the
+	 * bo->vaddr is uncached access.
+	 */
+	temp = kvmalloc_array(temp_size, 1, GFP_KERNEL);
+	if (!temp) {
+		DRM_ERROR("Failed to allocate storage for copying "
+			  "in bin/render CLs.\n");
+		ret = -ENOMEM;
+		goto fail;
+	}
+	bin = temp + bin_offset;
+	exec->shader_rec_u = temp + shader_rec_offset;
+	exec->uniforms_u = temp + uniforms_offset;
+	exec->shader_state = temp + exec_size;
+	exec->shader_state_size = args->shader_rec_count;
+
+	if (copy_from_user(bin,
+			   u64_to_user_ptr(args->bin_cl),
+			   args->bin_cl_size)) {
+		ret = -EFAULT;
+		goto fail;
+	}
+
+	if (copy_from_user(exec->shader_rec_u,
+			   u64_to_user_ptr(args->shader_rec),
+			   args->shader_rec_size)) {
+		ret = -EFAULT;
+		goto fail;
+	}
+
+	if (copy_from_user(exec->uniforms_u,
+			   u64_to_user_ptr(args->uniforms),
+			   args->uniforms_size)) {
+		ret = -EFAULT;
+		goto fail;
+	}
+
+	bo = vc4_bo_create(dev, exec_size, true, VC4_BO_TYPE_BCL);
+	if (IS_ERR(bo)) {
+		DRM_ERROR("Couldn't allocate BO for binning\n");
+		ret = PTR_ERR(bo);
+		goto fail;
+	}
+	exec->exec_bo = &bo->base;
+
+	list_add_tail(&to_vc4_bo(&exec->exec_bo->base)->unref_head,
+		      &exec->unref_list);
+
+	exec->ct0ca = exec->exec_bo->dma_addr + bin_offset;
+
+	exec->bin_u = bin;
+
+	exec->shader_rec_v = exec->exec_bo->vaddr + shader_rec_offset;
+	exec->shader_rec_p = exec->exec_bo->dma_addr + shader_rec_offset;
+	exec->shader_rec_size = args->shader_rec_size;
+
+	exec->uniforms_v = exec->exec_bo->vaddr + uniforms_offset;
+	exec->uniforms_p = exec->exec_bo->dma_addr + uniforms_offset;
+	exec->uniforms_size = args->uniforms_size;
+
+	ret = vc4_validate_bin_cl(dev,
+				  exec->exec_bo->vaddr + bin_offset,
+				  bin,
+				  exec);
+	if (ret)
+		goto fail;
+
+	ret = vc4_validate_shader_recs(dev, exec);
+	if (ret)
+		goto fail;
+
+	if (exec->found_tile_binning_mode_config_packet) {
+		ret = vc4_v3d_bin_bo_get(vc4, &exec->bin_bo_used);
+		if (ret)
+			goto fail;
+	}
+
+	/* Block waiting on any previous rendering into the CS's VBO,
+	 * IB, or textures, so that pixels are actually written by the
+	 * time we try to read them.
+	 */
+	ret = vc4_wait_for_seqno(dev, exec->bin_dep_seqno, ~0ull, true);
+
+fail:
+	kvfree(temp);
+	return ret;
+}
+
+static void
+vc4_complete_exec(struct drm_device *dev, struct vc4_exec_info *exec)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	unsigned long irqflags;
+	unsigned i;
+
+	/* If we got force-completed because of GPU reset rather than
+	 * through our IRQ handler, signal the fence now.
+	 */
+	if (exec->fence) {
+		dma_fence_signal(exec->fence);
+		dma_fence_put(exec->fence);
+	}
+
+	if (exec->bo) {
+		for (i = 0; i < exec->bo_count; i++) {
+			struct vc4_bo *bo = to_vc4_bo(&exec->bo[i]->base);
+
+			vc4_bo_dec_usecnt(bo);
+			drm_gem_object_put(&exec->bo[i]->base);
+		}
+		kvfree(exec->bo);
+	}
+
+	while (!list_empty(&exec->unref_list)) {
+		struct vc4_bo *bo = list_first_entry(&exec->unref_list,
+						     struct vc4_bo, unref_head);
+		list_del(&bo->unref_head);
+		drm_gem_object_put(&bo->base.base);
+	}
+
+	/* Free up the allocation of any bin slots we used. */
+	spin_lock_irqsave(&vc4->job_lock, irqflags);
+	vc4->bin_alloc_used &= ~exec->bin_slots;
+	spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+
+	/* Release the reference on the binner BO if needed. */
+	if (exec->bin_bo_used)
+		vc4_v3d_bin_bo_put(vc4);
+
+	/* Release the reference we had on the perf monitor. */
+	vc4_perfmon_put(exec->perfmon);
+
+	vc4_v3d_pm_put(vc4);
+
+	kfree(exec);
+}
+
+void
+vc4_job_handle_completed(struct vc4_dev *vc4)
+{
+	unsigned long irqflags;
+	struct vc4_seqno_cb *cb, *cb_temp;
+
+	if (WARN_ON_ONCE(vc4->is_vc5))
+		return;
+
+	spin_lock_irqsave(&vc4->job_lock, irqflags);
+	while (!list_empty(&vc4->job_done_list)) {
+		struct vc4_exec_info *exec =
+			list_first_entry(&vc4->job_done_list,
+					 struct vc4_exec_info, head);
+		list_del(&exec->head);
+
+		spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+		vc4_complete_exec(&vc4->base, exec);
+		spin_lock_irqsave(&vc4->job_lock, irqflags);
+	}
+
+	list_for_each_entry_safe(cb, cb_temp, &vc4->seqno_cb_list, work.entry) {
+		if (cb->seqno <= vc4->finished_seqno) {
+			list_del_init(&cb->work.entry);
+			schedule_work(&cb->work);
+		}
+	}
+
+	spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+}
+
+static void vc4_seqno_cb_work(struct work_struct *work)
+{
+	struct vc4_seqno_cb *cb = container_of(work, struct vc4_seqno_cb, work);
+
+	cb->func(cb);
+}
+
+int vc4_queue_seqno_cb(struct drm_device *dev,
+		       struct vc4_seqno_cb *cb, uint64_t seqno,
+		       void (*func)(struct vc4_seqno_cb *cb))
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	unsigned long irqflags;
+
+	if (WARN_ON_ONCE(vc4->is_vc5))
+		return -ENODEV;
+
+	cb->func = func;
+	INIT_WORK(&cb->work, vc4_seqno_cb_work);
+
+	spin_lock_irqsave(&vc4->job_lock, irqflags);
+	if (seqno > vc4->finished_seqno) {
+		cb->seqno = seqno;
+		list_add_tail(&cb->work.entry, &vc4->seqno_cb_list);
+	} else {
+		schedule_work(&cb->work);
+	}
+	spin_unlock_irqrestore(&vc4->job_lock, irqflags);
+
+	return 0;
+}
+
+/* Scheduled when any job has been completed, this walks the list of
+ * jobs that had completed and unrefs their BOs and frees their exec
+ * structs.
+ */
+static void
+vc4_job_done_work(struct work_struct *work)
+{
+	struct vc4_dev *vc4 =
+		container_of(work, struct vc4_dev, job_done_work);
+
+	vc4_job_handle_completed(vc4);
+}
+
+static int
+vc4_wait_for_seqno_ioctl_helper(struct drm_device *dev,
+				uint64_t seqno,
+				uint64_t *timeout_ns)
+{
+	unsigned long start = jiffies;
+	int ret = vc4_wait_for_seqno(dev, seqno, *timeout_ns, true);
+
+	if ((ret == -EINTR || ret == -ERESTARTSYS) && *timeout_ns != ~0ull) {
+		uint64_t delta = jiffies_to_nsecs(jiffies - start);
+
+		if (*timeout_ns >= delta)
+			*timeout_ns -= delta;
+	}
+
+	return ret;
+}
+
+int
+vc4_wait_seqno_ioctl(struct drm_device *dev, void *data,
+		     struct drm_file *file_priv)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct drm_vc4_wait_seqno *args = data;
+
+	if (WARN_ON_ONCE(vc4->is_vc5))
+		return -ENODEV;
+
+	return vc4_wait_for_seqno_ioctl_helper(dev, args->seqno,
+					       &args->timeout_ns);
+}
+
+int
+vc4_wait_bo_ioctl(struct drm_device *dev, void *data,
+		  struct drm_file *file_priv)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	int ret;
+	struct drm_vc4_wait_bo *args = data;
+	struct drm_gem_object *gem_obj;
+	struct vc4_bo *bo;
+
+	if (WARN_ON_ONCE(vc4->is_vc5))
+		return -ENODEV;
+
+	if (args->pad != 0)
+		return -EINVAL;
+
+	gem_obj = drm_gem_object_lookup(file_priv, args->handle);
+	if (!gem_obj) {
+		DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
+		return -EINVAL;
+	}
+	bo = to_vc4_bo(gem_obj);
+
+	ret = vc4_wait_for_seqno_ioctl_helper(dev, bo->seqno,
+					      &args->timeout_ns);
+
+	drm_gem_object_put(gem_obj);
+	return ret;
+}
+
+/**
+ * vc4_submit_cl_ioctl() - Submits a job (frame) to the VC4.
+ * @dev: DRM device
+ * @data: ioctl argument
+ * @file_priv: DRM file for this fd
+ *
+ * This is the main entrypoint for userspace to submit a 3D frame to
+ * the GPU.  Userspace provides the binner command list (if
+ * applicable), and the kernel sets up the render command list to draw
+ * to the framebuffer described in the ioctl, using the command lists
+ * that the 3D engine's binner will produce.
+ */
+int
+vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
+		    struct drm_file *file_priv)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct vc4_file *vc4file = file_priv->driver_priv;
+	struct drm_vc4_submit_cl *args = data;
+	struct drm_syncobj *out_sync = NULL;
+	struct vc4_exec_info *exec;
+	struct ww_acquire_ctx acquire_ctx;
+	struct dma_fence *in_fence;
+	int ret = 0;
+
+	trace_vc4_submit_cl_ioctl(dev, args->bin_cl_size,
+				  args->shader_rec_size,
+				  args->bo_handle_count);
+
+	if (WARN_ON_ONCE(vc4->is_vc5))
+		return -ENODEV;
+
+	if (!vc4->v3d) {
+		DRM_DEBUG("VC4_SUBMIT_CL with no VC4 V3D probed\n");
+		return -ENODEV;
+	}
+
+	if ((args->flags & ~(VC4_SUBMIT_CL_USE_CLEAR_COLOR |
+			     VC4_SUBMIT_CL_FIXED_RCL_ORDER |
+			     VC4_SUBMIT_CL_RCL_ORDER_INCREASING_X |
+			     VC4_SUBMIT_CL_RCL_ORDER_INCREASING_Y)) != 0) {
+		DRM_DEBUG("Unknown flags: 0x%02x\n", args->flags);
+		return -EINVAL;
+	}
+
+	if (args->pad2 != 0) {
+		DRM_DEBUG("Invalid pad: 0x%08x\n", args->pad2);
+		return -EINVAL;
+	}
+
+	exec = kcalloc(1, sizeof(*exec), GFP_KERNEL);
+	if (!exec) {
+		DRM_ERROR("malloc failure on exec struct\n");
+		return -ENOMEM;
+	}
+	exec->dev = vc4;
+
+	ret = vc4_v3d_pm_get(vc4);
+	if (ret) {
+		kfree(exec);
+		return ret;
+	}
+
+	exec->args = args;
+	INIT_LIST_HEAD(&exec->unref_list);
+
+	ret = vc4_cl_lookup_bos(dev, file_priv, exec);
+	if (ret)
+		goto fail;
+
+	if (args->perfmonid) {
+		exec->perfmon = vc4_perfmon_find(vc4file,
+						 args->perfmonid);
+		if (!exec->perfmon) {
+			ret = -ENOENT;
+			goto fail;
+		}
+	}
+
+	if (args->in_sync) {
+		ret = drm_syncobj_find_fence(file_priv, args->in_sync,
+					     0, 0, &in_fence);
+		if (ret)
+			goto fail;
+
+		/* When the fence (or fence array) is exclusively from our
+		 * context we can skip the wait since jobs are executed in
+		 * order of their submission through this ioctl and this can
+		 * only have fences from a prior job.
+		 */
+		if (!dma_fence_match_context(in_fence,
+					     vc4->dma_fence_context)) {
+			ret = dma_fence_wait(in_fence, true);
+			if (ret) {
+				dma_fence_put(in_fence);
+				goto fail;
+			}
+		}
+
+		dma_fence_put(in_fence);
+	}
+
+	if (exec->args->bin_cl_size != 0) {
+		ret = vc4_get_bcl(dev, exec);
+		if (ret)
+			goto fail;
+	} else {
+		exec->ct0ca = 0;
+		exec->ct0ea = 0;
+	}
+
+	ret = vc4_get_rcl(dev, exec);
+	if (ret)
+		goto fail;
+
+	ret = vc4_lock_bo_reservations(dev, exec, &acquire_ctx);
+	if (ret)
+		goto fail;
+
+	if (args->out_sync) {
+		out_sync = drm_syncobj_find(file_priv, args->out_sync);
+		if (!out_sync) {
+			ret = -EINVAL;
+			goto fail;
+		}
+
+		/* We replace the fence in out_sync in vc4_queue_submit since
+		 * the render job could execute immediately after that call.
+		 * If it finishes before our ioctl processing resumes the
+		 * render job fence could already have been freed.
+		 */
+	}
+
+	/* Clear this out of the struct we'll be putting in the queue,
+	 * since it's part of our stack.
+	 */
+	exec->args = NULL;
+
+	ret = vc4_queue_submit(dev, exec, &acquire_ctx, out_sync);
+
+	/* The syncobj isn't part of the exec data and we need to free our
+	 * reference even if job submission failed.
+	 */
+	if (out_sync)
+		drm_syncobj_put(out_sync);
+
+	if (ret)
+		goto fail;
+
+	/* Return the seqno for our job. */
+	args->seqno = vc4->emit_seqno;
+
+	return 0;
+
+fail:
+	vc4_complete_exec(&vc4->base, exec);
+
+	return ret;
+}
+
+static void vc4_gem_destroy(struct drm_device *dev, void *unused);
+int vc4_gem_init(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	int ret;
+
+	if (WARN_ON_ONCE(vc4->is_vc5))
+		return -ENODEV;
+
+	vc4->dma_fence_context = dma_fence_context_alloc(1);
+
+	INIT_LIST_HEAD(&vc4->bin_job_list);
+	INIT_LIST_HEAD(&vc4->render_job_list);
+	INIT_LIST_HEAD(&vc4->job_done_list);
+	INIT_LIST_HEAD(&vc4->seqno_cb_list);
+	spin_lock_init(&vc4->job_lock);
+
+	INIT_WORK(&vc4->hangcheck.reset_work, vc4_reset_work);
+	timer_setup(&vc4->hangcheck.timer, vc4_hangcheck_elapsed, 0);
+
+	INIT_WORK(&vc4->job_done_work, vc4_job_done_work);
+
+	ret = drmm_mutex_init(dev, &vc4->power_lock);
+	if (ret)
+		return ret;
+
+	INIT_LIST_HEAD(&vc4->purgeable.list);
+
+	ret = drmm_mutex_init(dev, &vc4->purgeable.lock);
+	if (ret)
+		return ret;
+
+	return drmm_add_action_or_reset(dev, vc4_gem_destroy, NULL);
+}
+
+static void vc4_gem_destroy(struct drm_device *dev, void *unused)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	/* Waiting for exec to finish would need to be done before
+	 * unregistering V3D.
+	 */
+	WARN_ON(vc4->emit_seqno != vc4->finished_seqno);
+
+	/* V3D should already have disabled its interrupt and cleared
+	 * the overflow allocation registers.  Now free the object.
+	 */
+	if (vc4->bin_bo) {
+		drm_gem_object_put(&vc4->bin_bo->base.base);
+		vc4->bin_bo = NULL;
+	}
+
+	if (vc4->hang_state)
+		vc4_free_hang_state(dev, vc4->hang_state);
+}
+
+int vc4_gem_madvise_ioctl(struct drm_device *dev, void *data,
+			  struct drm_file *file_priv)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct drm_vc4_gem_madvise *args = data;
+	struct drm_gem_object *gem_obj;
+	struct vc4_bo *bo;
+	int ret;
+
+	if (WARN_ON_ONCE(vc4->is_vc5))
+		return -ENODEV;
+
+	switch (args->madv) {
+	case VC4_MADV_DONTNEED:
+	case VC4_MADV_WILLNEED:
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	if (args->pad != 0)
+		return -EINVAL;
+
+	gem_obj = drm_gem_object_lookup(file_priv, args->handle);
+	if (!gem_obj) {
+		DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
+		return -ENOENT;
+	}
+
+	bo = to_vc4_bo(gem_obj);
+
+	/* Only BOs exposed to userspace can be purged. */
+	if (bo->madv == __VC4_MADV_NOTSUPP) {
+		DRM_DEBUG("madvise not supported on this BO\n");
+		ret = -EINVAL;
+		goto out_put_gem;
+	}
+
+	/* Not sure it's safe to purge imported BOs. Let's just assume it's
+	 * not until proven otherwise.
+	 */
+	if (gem_obj->import_attach) {
+		DRM_DEBUG("madvise not supported on imported BOs\n");
+		ret = -EINVAL;
+		goto out_put_gem;
+	}
+
+	mutex_lock(&bo->madv_lock);
+
+	if (args->madv == VC4_MADV_DONTNEED && bo->madv == VC4_MADV_WILLNEED &&
+	    !refcount_read(&bo->usecnt)) {
+		/* If the BO is about to be marked as purgeable, is not used
+		 * and is not already purgeable or purged, add it to the
+		 * purgeable list.
+		 */
+		vc4_bo_add_to_purgeable_pool(bo);
+	} else if (args->madv == VC4_MADV_WILLNEED &&
+		   bo->madv == VC4_MADV_DONTNEED &&
+		   !refcount_read(&bo->usecnt)) {
+		/* The BO has not been purged yet, just remove it from
+		 * the purgeable list.
+		 */
+		vc4_bo_remove_from_purgeable_pool(bo);
+	}
+
+	/* Save the purged state. */
+	args->retained = bo->madv != __VC4_MADV_PURGED;
+
+	/* Update internal madv state only if the bo was not purged. */
+	if (bo->madv != __VC4_MADV_PURGED)
+		bo->madv = args->madv;
+
+	mutex_unlock(&bo->madv_lock);
+
+	ret = 0;
+
+out_put_gem:
+	drm_gem_object_put(gem_obj);
+
+	return ret;
+}