gpu: host1x: Implement job tracking using DMA fences

In anticipation of removal of the intr API, implement job tracking
using DMA fences instead. The main two things about this are
making cdma_update schedule the work since fence completion can
now be called from interrupt context, and some complication in
ensuring the callback is not running when we free the fence.

Signed-off-by: Mikko Perttunen <[e-mail]>
Signed-off-by: Thierry Reding <[e-mail]>

drivers/gpu/host1x/cdma.c

···
 	host1x_hw_cdma_resume(host1x, cdma, restart_addr);
 }
 
+static void cdma_update_work(struct work_struct *work)
+{
+	struct host1x_cdma *cdma = container_of(work, struct host1x_cdma, update_work);
+
+	mutex_lock(&cdma->lock);
+	update_cdma_locked(cdma);
+	mutex_unlock(&cdma->lock);
+}
+
 /*
  * Create a cdma
  */
···
 
 	mutex_init(&cdma->lock);
 	init_completion(&cdma->complete);
+	INIT_WORK(&cdma->update_work, cdma_update_work);
 
 	INIT_LIST_HEAD(&cdma->sync_queue);
 
···
  */
 void host1x_cdma_update(struct host1x_cdma *cdma)
 {
-	mutex_lock(&cdma->lock);
-	update_cdma_locked(cdma);
-	mutex_unlock(&cdma->lock);
+	schedule_work(&cdma->update_work);
 }

drivers/gpu/host1x/cdma.h

···
 #include <linux/sched.h>
 #include <linux/completion.h>
 #include <linux/list.h>
+#include <linux/workqueue.h>
 
 struct host1x_syncpt;
 struct host1x_userctx_timeout;
···
 	struct buffer_timeout timeout;	/* channel's timeout state/wq */
 	bool running;
 	bool torndown;
+	struct work_struct update_work;
 };
 
 #define cdma_to_channel(cdma) container_of(cdma, struct host1x_channel, cdma)

drivers/gpu/host1x/hw/channel_hw.c

···
 #endif
 }
 
+static void job_complete_callback(struct dma_fence *fence, struct dma_fence_cb *cb)
+{
+	struct host1x_job *job = container_of(cb, struct host1x_job, fence_cb);
+
+	/* Schedules CDMA update. */
+	host1x_cdma_update(&job->channel->cdma);
+}
+
 static int channel_submit(struct host1x_job *job)
 {
 	struct host1x_channel *ch = job->channel;
···
 	u32 prev_max = 0;
 	u32 syncval;
 	int err;
-	struct host1x_waitlist *completed_waiter = NULL;
 	struct host1x *host = dev_get_drvdata(ch->dev->parent);
 
 	trace_host1x_channel_submit(dev_name(ch->dev),
···
 	/* get submit lock */
 	err = mutex_lock_interruptible(&ch->submitlock);
 	if (err)
-		goto error;
-
-	completed_waiter = kzalloc(sizeof(*completed_waiter), GFP_KERNEL);
-	if (!completed_waiter) {
-		mutex_unlock(&ch->submitlock);
-		err = -ENOMEM;
-		goto error;
-	}
+		return err;
 
 	host1x_channel_set_streamid(ch);
 	host1x_enable_gather_filter(ch);
···
 	err = host1x_cdma_begin(&ch->cdma, job);
 	if (err) {
 		mutex_unlock(&ch->submitlock);
-		goto error;
+		return err;
 	}
 
 	channel_program_cdma(job);
 	syncval = host1x_syncpt_read_max(sp);
+
+	/*
+	 * Create fence before submitting job to HW to avoid job completing
+	 * before the fence is set up.
+	 */
+	job->fence = host1x_fence_create(sp, syncval);
+	if (WARN(IS_ERR(job->fence), "Failed to create submit complete fence")) {
+		job->fence = NULL;
+	} else {
+		err = dma_fence_add_callback(job->fence, &job->fence_cb,
+					     job_complete_callback);
+	}
 
 	/* end CDMA submit & stash pinned hMems into sync queue */
 	host1x_cdma_end(&ch->cdma, job);
 
 	trace_host1x_channel_submitted(dev_name(ch->dev), prev_max, syncval);
 
-	/* schedule a submit complete interrupt */
-	err = host1x_intr_add_action(host, sp, syncval,
-				     HOST1X_INTR_ACTION_SUBMIT_COMPLETE, ch,
-				     completed_waiter, &job->waiter);
-	completed_waiter = NULL;
-	WARN(err, "Failed to set submit complete interrupt");
-
 	mutex_unlock(&ch->submitlock);
 
-	return 0;
+	if (err == -ENOENT)
+		host1x_cdma_update(&ch->cdma);
+	else
+		WARN(err, "Failed to set submit complete interrupt");
 
-error:
-	kfree(completed_waiter);
-	return err;
+	return 0;
 }
 
 static int host1x_channel_init(struct host1x_channel *ch, struct host1x *dev,

drivers/gpu/host1x/job.c

···
 	if (job->release)
 		job->release(job);
 
-	if (job->waiter)
-		host1x_intr_put_ref(job->syncpt->host, job->syncpt->id,
-				    job->waiter, false);
+	if (job->fence) {
+		/*
+		 * remove_callback is atomic w.r.t. fence signaling, so
+		 * after the call returns, we know that the callback is not
+		 * in execution, and the fence can be safely freed.
+		 */
+		dma_fence_remove_callback(job->fence, &job->fence_cb);
+		dma_fence_put(job->fence);
+	}
 
 	if (job->syncpt)
 		host1x_syncpt_put(job->syncpt);

include/linux/host1x.h

···
 
 #include <linux/device.h>
 #include <linux/dma-direction.h>
+#include <linux/dma-fence.h>
 #include <linux/spinlock.h>
 #include <linux/types.h>
 
···
 	u32 syncpt_incrs;
 	u32 syncpt_end;
 
-	/* Completion waiter ref */
-	void *waiter;
+	/* Completion fence for job tracking */
+	struct dma_fence *fence;
+	struct dma_fence_cb fence_cb;
 
 	/* Maximum time to wait for this job */
 	unsigned int timeout;