1 files changed, 354 insertions, 31 deletions

diff --git a/drivers/misc/habanalabs/command_submission.c b/drivers/misc/habanalabs/command_submission.c
index 409276b6374d..f82974a916c3 100644
--- a/drivers/misc/habanalabs/command_submission.c
+++ b/drivers/misc/habanalabs/command_submission.c
@@ -11,11 +11,33 @@
 #include <linux/uaccess.h>
 #include <linux/slab.h>
 
+#define HL_CS_FLAGS_SIG_WAIT	(HL_CS_FLAGS_SIGNAL | HL_CS_FLAGS_WAIT)
+
 static void job_wq_completion(struct work_struct *work);
 static long _hl_cs_wait_ioctl(struct hl_device *hdev,
 		struct hl_ctx *ctx, u64 timeout_us, u64 seq);
 static void cs_do_release(struct kref *ref);
 
+static void hl_sob_reset(struct kref *ref)
+{
+	struct hl_hw_sob *hw_sob = container_of(ref, struct hl_hw_sob,
+							kref);
+	struct hl_device *hdev = hw_sob->hdev;
+
+	hdev->asic_funcs->reset_sob(hdev, hw_sob);
+}
+
+void hl_sob_reset_error(struct kref *ref)
+{
+	struct hl_hw_sob *hw_sob = container_of(ref, struct hl_hw_sob,
+							kref);
+	struct hl_device *hdev = hw_sob->hdev;
+
+	dev_crit(hdev->dev,
+			"SOB release shouldn't be called here, q_idx: %d, sob_id: %d\n",
+			hw_sob->q_idx, hw_sob->sob_id);
+}
+
 static const char *hl_fence_get_driver_name(struct dma_fence *fence)
 {
 	return "HabanaLabs";
@@ -23,10 +45,10 @@ static const char *hl_fence_get_driver_name(struct dma_fence *fence)
 
 static const char *hl_fence_get_timeline_name(struct dma_fence *fence)
 {
-	struct hl_dma_fence *hl_fence =
-		container_of(fence, struct hl_dma_fence, base_fence);
+	struct hl_cs_compl *hl_cs_compl =
+		container_of(fence, struct hl_cs_compl, base_fence);
 
-	return dev_name(hl_fence->hdev->dev);
+	return dev_name(hl_cs_compl->hdev->dev);
 }
 
 static bool hl_fence_enable_signaling(struct dma_fence *fence)
@@ -36,17 +58,47 @@ static bool hl_fence_enable_signaling(struct dma_fence *fence)
 
 static void hl_fence_release(struct dma_fence *fence)
 {
-	struct hl_dma_fence *hl_fence =
-		container_of(fence, struct hl_dma_fence, base_fence);
+	struct hl_cs_compl *hl_cs_cmpl =
+		container_of(fence, struct hl_cs_compl, base_fence);
+	struct hl_device *hdev = hl_cs_cmpl->hdev;
+
+	if ((hl_cs_cmpl->type == CS_TYPE_SIGNAL) ||
+			(hl_cs_cmpl->type == CS_TYPE_WAIT)) {
 
-	kfree_rcu(hl_fence, base_fence.rcu);
+		dev_dbg(hdev->dev,
+			"CS 0x%llx type %d finished, sob_id: %d, sob_val: 0x%x\n",
+			hl_cs_cmpl->cs_seq,
+			hl_cs_cmpl->type,
+			hl_cs_cmpl->hw_sob->sob_id,
+			hl_cs_cmpl->sob_val);
+
+		/*
+		 * A signal CS can get completion while the corresponding wait
+		 * for signal CS is on its way to the PQ. The wait for signal CS
+		 * will get stuck if the signal CS incremented the SOB to its
+		 * max value and there are no pending (submitted) waits on this
+		 * SOB.
+		 * We do the following to void this situation:
+		 * 1. The wait for signal CS must get a ref for the signal CS as
+		 *    soon as possible in cs_ioctl_signal_wait() and put it
+		 *    before being submitted to the PQ but after it incremented
+		 *    the SOB refcnt in init_signal_wait_cs().
+		 * 2. Signal/Wait for signal CS will decrement the SOB refcnt
+		 *    here.
+		 * These two measures guarantee that the wait for signal CS will
+		 * reset the SOB upon completion rather than the signal CS and
+		 * hence the above scenario is avoided.
+		 */
+		kref_put(&hl_cs_cmpl->hw_sob->kref, hl_sob_reset);
+	}
+
+	kfree_rcu(hl_cs_cmpl, base_fence.rcu);
 }
 
 static const struct dma_fence_ops hl_fence_ops = {
 	.get_driver_name = hl_fence_get_driver_name,
 	.get_timeline_name = hl_fence_get_timeline_name,
 	.enable_signaling = hl_fence_enable_signaling,
-	.wait = dma_fence_default_wait,
 	.release = hl_fence_release
 };
 
@@ -113,6 +165,7 @@ static int cs_parser(struct hl_fpriv *hpriv, struct hl_cs_job *job)
 		if (!rc) {
 			job->patched_cb = parser.patched_cb;
 			job->job_cb_size = parser.patched_cb_size;
+			job->contains_dma_pkt = parser.contains_dma_pkt;
 
 			spin_lock(&job->patched_cb->lock);
 			job->patched_cb->cs_cnt++;
@@ -259,6 +312,12 @@ static void cs_do_release(struct kref *ref)
 
 			spin_unlock(&hdev->hw_queues_mirror_lock);
 		}
+	} else if (cs->type == CS_TYPE_WAIT) {
+		/*
+		 * In case the wait for signal CS was submitted, the put occurs
+		 * in init_signal_wait_cs() right before hanging on the PQ.
+		 */
+		dma_fence_put(cs->signal_fence);
 	}
 
 	/*
@@ -312,9 +371,9 @@ static void cs_timedout(struct work_struct *work)
 }
 
 static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx,
-			struct hl_cs **cs_new)
+			enum hl_cs_type cs_type, struct hl_cs **cs_new)
 {
-	struct hl_dma_fence *fence;
+	struct hl_cs_compl *cs_cmpl;
 	struct dma_fence *other = NULL;
 	struct hl_cs *cs;
 	int rc;
@@ -326,25 +385,27 @@ static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx,
 	cs->ctx = ctx;
 	cs->submitted = false;
 	cs->completed = false;
+	cs->type = cs_type;
 	INIT_LIST_HEAD(&cs->job_list);
 	INIT_DELAYED_WORK(&cs->work_tdr, cs_timedout);
 	kref_init(&cs->refcount);
 	spin_lock_init(&cs->job_lock);
 
-	fence = kmalloc(sizeof(*fence), GFP_ATOMIC);
-	if (!fence) {
+	cs_cmpl = kmalloc(sizeof(*cs_cmpl), GFP_ATOMIC);
+	if (!cs_cmpl) {
 		rc = -ENOMEM;
 		goto free_cs;
 	}
 
-	fence->hdev = hdev;
-	spin_lock_init(&fence->lock);
-	cs->fence = &fence->base_fence;
+	cs_cmpl->hdev = hdev;
+	cs_cmpl->type = cs->type;
+	spin_lock_init(&cs_cmpl->lock);
+	cs->fence = &cs_cmpl->base_fence;
 
 	spin_lock(&ctx->cs_lock);
 
-	fence->cs_seq = ctx->cs_sequence;
-	other = ctx->cs_pending[fence->cs_seq & (HL_MAX_PENDING_CS - 1)];
+	cs_cmpl->cs_seq = ctx->cs_sequence;
+	other = ctx->cs_pending[cs_cmpl->cs_seq & (HL_MAX_PENDING_CS - 1)];
 	if ((other) && (!dma_fence_is_signaled(other))) {
 		spin_unlock(&ctx->cs_lock);
 		dev_dbg(hdev->dev,
@@ -353,16 +414,16 @@ static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx,
 		goto free_fence;
 	}
 
-	dma_fence_init(&fence->base_fence, &hl_fence_ops, &fence->lock,
+	dma_fence_init(&cs_cmpl->base_fence, &hl_fence_ops, &cs_cmpl->lock,
 			ctx->asid, ctx->cs_sequence);
 
-	cs->sequence = fence->cs_seq;
+	cs->sequence = cs_cmpl->cs_seq;
 
-	ctx->cs_pending[fence->cs_seq & (HL_MAX_PENDING_CS - 1)] =
-							&fence->base_fence;
+	ctx->cs_pending[cs_cmpl->cs_seq & (HL_MAX_PENDING_CS - 1)] =
+							&cs_cmpl->base_fence;
 	ctx->cs_sequence++;
 
-	dma_fence_get(&fence->base_fence);
+	dma_fence_get(&cs_cmpl->base_fence);
 
 	dma_fence_put(other);
 
@@ -373,7 +434,7 @@ static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx,
 	return 0;
 
 free_fence:
-	kfree(fence);
+	kfree(cs_cmpl);
 free_cs:
 	kfree(cs);
 	return rc;
@@ -499,8 +560,8 @@ struct hl_cs_job *hl_cs_allocate_job(struct hl_device *hdev,
 	return job;
 }
 
-static int _hl_cs_ioctl(struct hl_fpriv *hpriv, void __user *chunks,
-			u32 num_chunks, u64 *cs_seq)
+static int cs_ioctl_default(struct hl_fpriv *hpriv, void __user *chunks,
+				u32 num_chunks, u64 *cs_seq)
 {
 	struct hl_device *hdev = hpriv->hdev;
 	struct hl_cs_chunk *cs_chunk_array;
@@ -538,7 +599,7 @@ static int _hl_cs_ioctl(struct hl_fpriv *hpriv, void __user *chunks,
 	/* increment refcnt for context */
 	hl_ctx_get(hdev, hpriv->ctx);
 
-	rc = allocate_cs(hdev, hpriv->ctx, &cs);
+	rc = allocate_cs(hdev, hpriv->ctx, CS_TYPE_DEFAULT, &cs);
 	if (rc) {
 		hl_ctx_put(hpriv->ctx);
 		goto free_cs_chunk_array;
@@ -652,13 +713,230 @@ out:
 	return rc;
 }
 
+static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type,
+				void __user *chunks, u32 num_chunks,
+				u64 *cs_seq)
+{
+	struct hl_device *hdev = hpriv->hdev;
+	struct hl_ctx *ctx = hpriv->ctx;
+	struct hl_cs_chunk *cs_chunk_array, *chunk;
+	struct hw_queue_properties *hw_queue_prop;
+	struct dma_fence *sig_fence = NULL;
+	struct hl_cs_job *job;
+	struct hl_cs *cs;
+	struct hl_cb *cb;
+	u64 *signal_seq_arr = NULL, signal_seq;
+	u32 size_to_copy, q_idx, signal_seq_arr_len, cb_size;
+	int rc;
+
+	*cs_seq = ULLONG_MAX;
+
+	if (num_chunks > HL_MAX_JOBS_PER_CS) {
+		dev_err(hdev->dev,
+			"Number of chunks can NOT be larger than %d\n",
+			HL_MAX_JOBS_PER_CS);
+		rc = -EINVAL;
+		goto out;
+	}
+
+	cs_chunk_array = kmalloc_array(num_chunks, sizeof(*cs_chunk_array),
+					GFP_ATOMIC);
+	if (!cs_chunk_array) {
+		rc = -ENOMEM;
+		goto out;
+	}
+
+	size_to_copy = num_chunks * sizeof(struct hl_cs_chunk);
+	if (copy_from_user(cs_chunk_array, chunks, size_to_copy)) {
+		dev_err(hdev->dev, "Failed to copy cs chunk array from user\n");
+		rc = -EFAULT;
+		goto free_cs_chunk_array;
+	}
+
+	/* currently it is guaranteed to have only one chunk */
+	chunk = &cs_chunk_array[0];
+	q_idx = chunk->queue_index;
+	hw_queue_prop = &hdev->asic_prop.hw_queues_props[q_idx];
+
+	if ((q_idx >= HL_MAX_QUEUES) ||
+			(hw_queue_prop->type != QUEUE_TYPE_EXT)) {
+		dev_err(hdev->dev, "Queue index %d is invalid\n", q_idx);
+		rc = -EINVAL;
+		goto free_cs_chunk_array;
+	}
+
+	if (cs_type == CS_TYPE_WAIT) {
+		struct hl_cs_compl *sig_waitcs_cmpl;
+
+		signal_seq_arr_len = chunk->num_signal_seq_arr;
+
+		/* currently only one signal seq is supported */
+		if (signal_seq_arr_len != 1) {
+			dev_err(hdev->dev,
+				"Wait for signal CS supports only one signal CS seq\n");
+			rc = -EINVAL;
+			goto free_cs_chunk_array;
+		}
+
+		signal_seq_arr = kmalloc_array(signal_seq_arr_len,
+						sizeof(*signal_seq_arr),
+						GFP_ATOMIC);
+		if (!signal_seq_arr) {
+			rc = -ENOMEM;
+			goto free_cs_chunk_array;
+		}
+
+		size_to_copy = chunk->num_signal_seq_arr *
+				sizeof(*signal_seq_arr);
+		if (copy_from_user(signal_seq_arr,
+					u64_to_user_ptr(chunk->signal_seq_arr),
+					size_to_copy)) {
+			dev_err(hdev->dev,
+				"Failed to copy signal seq array from user\n");
+			rc = -EFAULT;
+			goto free_signal_seq_array;
+		}
+
+		/* currently it is guaranteed to have only one signal seq */
+		signal_seq = signal_seq_arr[0];
+		sig_fence = hl_ctx_get_fence(ctx, signal_seq);
+		if (IS_ERR(sig_fence)) {
+			dev_err(hdev->dev,
+				"Failed to get signal CS with seq 0x%llx\n",
+				signal_seq);
+			rc = PTR_ERR(sig_fence);
+			goto free_signal_seq_array;
+		}
+
+		if (!sig_fence) {
+			/* signal CS already finished */
+			rc = 0;
+			goto free_signal_seq_array;
+		}
+
+		sig_waitcs_cmpl =
+			container_of(sig_fence, struct hl_cs_compl, base_fence);
+
+		if (sig_waitcs_cmpl->type != CS_TYPE_SIGNAL) {
+			dev_err(hdev->dev,
+				"CS seq 0x%llx is not of a signal CS\n",
+				signal_seq);
+			dma_fence_put(sig_fence);
+			rc = -EINVAL;
+			goto free_signal_seq_array;
+		}
+
+		if (dma_fence_is_signaled(sig_fence)) {
+			/* signal CS already finished */
+			dma_fence_put(sig_fence);
+			rc = 0;
+			goto free_signal_seq_array;
+		}
+	}
+
+	/* increment refcnt for context */
+	hl_ctx_get(hdev, ctx);
+
+	rc = allocate_cs(hdev, ctx, cs_type, &cs);
+	if (rc) {
+		if (cs_type == CS_TYPE_WAIT)
+			dma_fence_put(sig_fence);
+		hl_ctx_put(ctx);
+		goto free_signal_seq_array;
+	}
+
+	/*
+	 * Save the signal CS fence for later initialization right before
+	 * hanging the wait CS on the queue.
+	 */
+	if (cs->type == CS_TYPE_WAIT)
+		cs->signal_fence = sig_fence;
+
+	hl_debugfs_add_cs(cs);
+
+	*cs_seq = cs->sequence;
+
+	job = hl_cs_allocate_job(hdev, QUEUE_TYPE_EXT, true);
+	if (!job) {
+		dev_err(hdev->dev, "Failed to allocate a new job\n");
+		rc = -ENOMEM;
+		goto put_cs;
+	}
+
+	cb = hl_cb_kernel_create(hdev, PAGE_SIZE);
+	if (!cb) {
+		kfree(job);
+		rc = -EFAULT;
+		goto put_cs;
+	}
+
+	if (cs->type == CS_TYPE_WAIT)
+		cb_size = hdev->asic_funcs->get_wait_cb_size(hdev);
+	else
+		cb_size = hdev->asic_funcs->get_signal_cb_size(hdev);
+
+	job->id = 0;
+	job->cs = cs;
+	job->user_cb = cb;
+	job->user_cb->cs_cnt++;
+	job->user_cb_size = cb_size;
+	job->hw_queue_id = q_idx;
+
+	/*
+	 * No need in parsing, user CB is the patched CB.
+	 * We call hl_cb_destroy() out of two reasons - we don't need the CB in
+	 * the CB idr anymore and to decrement its refcount as it was
+	 * incremented inside hl_cb_kernel_create().
+	 */
+	job->patched_cb = job->user_cb;
+	job->job_cb_size = job->user_cb_size;
+	hl_cb_destroy(hdev, &hdev->kernel_cb_mgr, cb->id << PAGE_SHIFT);
+
+	cs->jobs_in_queue_cnt[job->hw_queue_id]++;
+
+	list_add_tail(&job->cs_node, &cs->job_list);
+
+	/* increment refcount as for external queues we get completion */
+	cs_get(cs);
+
+	hl_debugfs_add_job(hdev, job);
+
+	rc = hl_hw_queue_schedule_cs(cs);
+	if (rc) {
+		if (rc != -EAGAIN)
+			dev_err(hdev->dev,
+				"Failed to submit CS %d.%llu to H/W queues, error %d\n",
+				ctx->asid, cs->sequence, rc);
+		goto free_cs_object;
+	}
+
+	rc = HL_CS_STATUS_SUCCESS;
+	goto put_cs;
+
+free_cs_object:
+	cs_rollback(hdev, cs);
+	*cs_seq = ULLONG_MAX;
+	/* The path below is both for good and erroneous exits */
+put_cs:
+	/* We finished with the CS in this function, so put the ref */
+	cs_put(cs);
+free_signal_seq_array:
+	if (cs_type == CS_TYPE_WAIT)
+		kfree(signal_seq_arr);
+free_cs_chunk_array:
+	kfree(cs_chunk_array);
+out:
+	return rc;
+}
+
 int hl_cs_ioctl(struct hl_fpriv *hpriv, void *data)
 {
 	struct hl_device *hdev = hpriv->hdev;
 	union hl_cs_args *args = data;
 	struct hl_ctx *ctx = hpriv->ctx;
 	void __user *chunks_execute, *chunks_restore;
-	u32 num_chunks_execute, num_chunks_restore;
+	enum hl_cs_type cs_type;
+	u32 num_chunks_execute, num_chunks_restore, sig_wait_flags;
 	u64 cs_seq = ULONG_MAX;
 	int rc, do_ctx_switch;
 	bool need_soft_reset = false;
@@ -671,12 +949,44 @@ int hl_cs_ioctl(struct hl_fpriv *hpriv, void *data)
 		goto out;
 	}
 
+	sig_wait_flags = args->in.cs_flags & HL_CS_FLAGS_SIG_WAIT;
+
+	if (unlikely(sig_wait_flags == HL_CS_FLAGS_SIG_WAIT)) {
+		dev_err(hdev->dev,
+			"Signal and wait CS flags are mutually exclusive, context %d\n",
+		ctx->asid);
+		rc = -EINVAL;
+		goto out;
+	}
+
+	if (unlikely((sig_wait_flags & HL_CS_FLAGS_SIG_WAIT) &&
+			(!hdev->supports_sync_stream))) {
+		dev_err(hdev->dev, "Sync stream CS is not supported\n");
+		rc = -EINVAL;
+		goto out;
+	}
+
+	if (args->in.cs_flags & HL_CS_FLAGS_SIGNAL)
+		cs_type = CS_TYPE_SIGNAL;
+	else if (args->in.cs_flags & HL_CS_FLAGS_WAIT)
+		cs_type = CS_TYPE_WAIT;
+	else
+		cs_type = CS_TYPE_DEFAULT;
+
 	chunks_execute = (void __user *) (uintptr_t) args->in.chunks_execute;
 	num_chunks_execute = args->in.num_chunks_execute;
 
-	if (!num_chunks_execute) {
+	if (cs_type == CS_TYPE_DEFAULT) {
+		if (!num_chunks_execute) {
+			dev_err(hdev->dev,
+				"Got execute CS with 0 chunks, context %d\n",
+				ctx->asid);
+			rc = -EINVAL;
+			goto out;
+		}
+	} else if (num_chunks_execute != 1) {
 		dev_err(hdev->dev,
-			"Got execute CS with 0 chunks, context %d\n",
+			"Sync stream CS mandates one chunk only, context %d\n",
 			ctx->asid);
 		rc = -EINVAL;
 		goto out;
@@ -722,7 +1032,7 @@ int hl_cs_ioctl(struct hl_fpriv *hpriv, void *data)
 			"Need to run restore phase but restore CS is empty\n");
 			rc = 0;
 		} else {
-			rc = _hl_cs_ioctl(hpriv, chunks_restore,
+			rc = cs_ioctl_default(hpriv, chunks_restore,
 						num_chunks_restore, &cs_seq);
 		}
 
@@ -764,7 +1074,12 @@ int hl_cs_ioctl(struct hl_fpriv *hpriv, void *data)
 		}
 	}
 
-	rc = _hl_cs_ioctl(hpriv, chunks_execute, num_chunks_execute, &cs_seq);
+	if (cs_type == CS_TYPE_DEFAULT)
+		rc = cs_ioctl_default(hpriv, chunks_execute, num_chunks_execute,
+					&cs_seq);
+	else
+		rc = cs_ioctl_signal_wait(hpriv, cs_type, chunks_execute,
+						num_chunks_execute, &cs_seq);
 
 out:
 	if (rc != -EAGAIN) {
@@ -796,6 +1111,10 @@ static long _hl_cs_wait_ioctl(struct hl_device *hdev,
 	fence = hl_ctx_get_fence(ctx, seq);
 	if (IS_ERR(fence)) {
 		rc = PTR_ERR(fence);
+		if (rc == -EINVAL)
+			dev_notice_ratelimited(hdev->dev,
+				"Can't wait on seq %llu because current CS is at seq %llu\n",
+				seq, ctx->cs_sequence);
 	} else if (fence) {
 		rc = dma_fence_wait_timeout(fence, true, timeout);
 		if (fence->error == -ETIMEDOUT)
@@ -803,8 +1122,12 @@ static long _hl_cs_wait_ioctl(struct hl_device *hdev,
 		else if (fence->error == -EIO)
 			rc = -EIO;
 		dma_fence_put(fence);
-	} else
+	} else {
+		dev_dbg(hdev->dev,
+			"Can't wait on seq %llu because current CS is at seq %llu (Fence is gone)\n",
+			seq, ctx->cs_sequence);
 		rc = 1;
+	}
 
 	hl_ctx_put(ctx);