usb: wusbcore: serialize access to the HWA data out endpoint

This patch serializes access to the HWA data transfer out (DTO)
endpoint.  This prevents a situation where two transfer requests being
sent concurrently to separate downstream endpoints could interleave
their transfer request and transfer data packets causing data
corruption.  The transfer processing code will now attempt to acquire
the DTO resource before sending a transfer to the HWA.  If it cannot
acquire the resource, the RPIPE that the transfer is assigned to will
be placed on a waiting list.  When the DTO resource is released, the
actor releasing the resource will serivce the RPIPEs that are waiting.

Signed-off-by: Thomas Pugliese <thomas.pugliese@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

3 files changed, 152 insertions, 21 deletions

diff --git a/drivers/usb/wusbcore/wa-hc.h b/drivers/usb/wusbcore/wa-hc.h
index b44aca3..41afaa6 100644
--- a/drivers/usb/wusbcore/wa-hc.h
+++ b/drivers/usb/wusbcore/wa-hc.h
@@ -117,6 +117,7 @@ struct wa_rpipe {
 	struct wahc *wa;
 	spinlock_t seg_lock;
 	struct list_head seg_list;
+	struct list_head list_node;
 	atomic_t segs_available;
 	u8 buffer[1];	/* For reads/writes on USB */
 };
@@ -183,7 +184,8 @@ struct wahc {
 
 	u16 rpipes;
 	unsigned long *rpipe_bm;	/* rpipe usage bitmap */
-	spinlock_t rpipe_bm_lock;	/* protect rpipe_bm */
+	struct list_head rpipe_delayed_list;	/* delayed RPIPES. */
+	spinlock_t rpipe_lock;	/* protect rpipe_bm and delayed list */
 	struct mutex rpipe_mutex;	/* assigning resources to endpoints */
 
 	/*
@@ -201,6 +203,8 @@ struct wahc {
 	void *dti_buf;
 	size_t dti_buf_size;
 
+	unsigned long dto_in_use;	/* protect dto endoint serialization. */
+
 	s32 status;			/* For reading status */
 
 	struct list_head xfer_list;
@@ -253,7 +257,8 @@ static inline void wa_nep_disarm(struct wahc *wa)
 /* RPipes */
 static inline void wa_rpipe_init(struct wahc *wa)
 {
-	spin_lock_init(&wa->rpipe_bm_lock);
+	INIT_LIST_HEAD(&wa->rpipe_delayed_list);
+	spin_lock_init(&wa->rpipe_lock);
 	mutex_init(&wa->rpipe_mutex);
 }
 
@@ -270,6 +275,7 @@ static inline void wa_init(struct wahc *wa)
 	spin_lock_init(&wa->xfer_list_lock);
 	INIT_WORK(&wa->xfer_enqueue_work, wa_urb_enqueue_run);
 	INIT_WORK(&wa->xfer_error_work, wa_process_errored_transfers_run);
+	wa->dto_in_use = 0;
 	atomic_set(&wa->xfer_id_count, 1);
 }
 
diff --git a/drivers/usb/wusbcore/wa-rpipe.c b/drivers/usb/wusbcore/wa-rpipe.c
index 554b16b..50de1d2 100644
--- a/drivers/usb/wusbcore/wa-rpipe.c
+++ b/drivers/usb/wusbcore/wa-rpipe.c
@@ -143,17 +143,18 @@ static void rpipe_init(struct wa_rpipe *rpipe)
 	kref_init(&rpipe->refcnt);
 	spin_lock_init(&rpipe->seg_lock);
 	INIT_LIST_HEAD(&rpipe->seg_list);
+	INIT_LIST_HEAD(&rpipe->list_node);
 }
 
 static unsigned rpipe_get_idx(struct wahc *wa, unsigned rpipe_idx)
 {
 	unsigned long flags;
 
-	spin_lock_irqsave(&wa->rpipe_bm_lock, flags);
+	spin_lock_irqsave(&wa->rpipe_lock, flags);
 	rpipe_idx = find_next_zero_bit(wa->rpipe_bm, wa->rpipes, rpipe_idx);
 	if (rpipe_idx < wa->rpipes)
 		set_bit(rpipe_idx, wa->rpipe_bm);
-	spin_unlock_irqrestore(&wa->rpipe_bm_lock, flags);
+	spin_unlock_irqrestore(&wa->rpipe_lock, flags);
 
 	return rpipe_idx;
 }
@@ -162,9 +163,9 @@ static void rpipe_put_idx(struct wahc *wa, unsigned rpipe_idx)
 {
 	unsigned long flags;
 
-	spin_lock_irqsave(&wa->rpipe_bm_lock, flags);
+	spin_lock_irqsave(&wa->rpipe_lock, flags);
 	clear_bit(rpipe_idx, wa->rpipe_bm);
-	spin_unlock_irqrestore(&wa->rpipe_bm_lock, flags);
+	spin_unlock_irqrestore(&wa->rpipe_lock, flags);
 }
 
 void rpipe_destroy(struct kref *_rpipe)
diff --git a/drivers/usb/wusbcore/wa-xfer.c b/drivers/usb/wusbcore/wa-xfer.c
index fd00e1a..f1e9a38 100644
--- a/drivers/usb/wusbcore/wa-xfer.c
+++ b/drivers/usb/wusbcore/wa-xfer.c
@@ -107,6 +107,7 @@ enum wa_seg_status {
 };
 
 static void wa_xfer_delayed_run(struct wa_rpipe *);
+static int __wa_xfer_delayed_run(struct wa_rpipe *rpipe, int *dto_waiting);
 
 /*
  * Life cycle governed by 'struct urb' (the refcount of the struct is
@@ -204,6 +205,59 @@ static void wa_xfer_put(struct wa_xfer *xfer)
 }
 
 /*
+ * Try to get exclusive access to the DTO endpoint resource.  Return true
+ * if successful.
+ */
+static inline int __wa_dto_try_get(struct wahc *wa)
+{
+	return (test_and_set_bit(0, &wa->dto_in_use) == 0);
+}
+
+/* Release the DTO endpoint resource. */
+static inline void __wa_dto_put(struct wahc *wa)
+{
+	clear_bit_unlock(0, &wa->dto_in_use);
+}
+
+/* Service RPIPEs that are waiting on the DTO resource. */
+static void wa_check_for_delayed_rpipes(struct wahc *wa)
+{
+	unsigned long flags;
+	int dto_waiting = 0;
+	struct wa_rpipe *rpipe;
+
+	spin_lock_irqsave(&wa->rpipe_lock, flags);
+	while (!list_empty(&wa->rpipe_delayed_list) && !dto_waiting) {
+		rpipe = list_first_entry(&wa->rpipe_delayed_list,
+				struct wa_rpipe, list_node);
+		__wa_xfer_delayed_run(rpipe, &dto_waiting);
+		/* remove this RPIPE from the list if it is not waiting. */
+		if (!dto_waiting) {
+			pr_debug("%s: RPIPE %d serviced and removed from delayed list.\n",
+				__func__,
+				le16_to_cpu(rpipe->descr.wRPipeIndex));
+			list_del_init(&rpipe->list_node);
+		}
+	}
+	spin_unlock_irqrestore(&wa->rpipe_lock, flags);
+}
+
+/* add this RPIPE to the end of the delayed RPIPE list. */
+static void wa_add_delayed_rpipe(struct wahc *wa, struct wa_rpipe *rpipe)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&wa->rpipe_lock, flags);
+	/* add rpipe to the list if it is not already on it. */
+	if (list_empty(&rpipe->list_node)) {
+		pr_debug("%s: adding RPIPE %d to the delayed list.\n",
+			__func__, le16_to_cpu(rpipe->descr.wRPipeIndex));
+		list_add_tail(&rpipe->list_node, &wa->rpipe_delayed_list);
+	}
+	spin_unlock_irqrestore(&wa->rpipe_lock, flags);
+}
+
+/*
  * xfer is referenced
  *
  * xfer->lock has to be unlocked
@@ -1099,9 +1153,13 @@ error_setup_sizes:
  * rpipe->seg_lock is held!
  */
 static int __wa_seg_submit(struct wa_rpipe *rpipe, struct wa_xfer *xfer,
-			   struct wa_seg *seg)
+			   struct wa_seg *seg, int *dto_done)
 {
 	int result;
+
+	/* default to done unless we encounter a multi-frame isoc segment. */
+	*dto_done = 1;
+
 	/* submit the transfer request. */
 	result = usb_submit_urb(&seg->tr_urb, GFP_ATOMIC);
 	if (result < 0) {
@@ -1142,28 +1200,34 @@ error_seg_submit:
 }
 
 /*
- * Execute more queued request segments until the maximum concurrent allowed
+ * Execute more queued request segments until the maximum concurrent allowed.
+ * Return true if the DTO resource was acquired and released.
  *
  * The ugly unlock/lock sequence on the error path is needed as the
  * xfer->lock normally nests the seg_lock and not viceversa.
- *
  */
-static void wa_xfer_delayed_run(struct wa_rpipe *rpipe)
+static int __wa_xfer_delayed_run(struct wa_rpipe *rpipe, int *dto_waiting)
 {
-	int result;
+	int result, dto_acquired = 0, dto_done = 0;
 	struct device *dev = &rpipe->wa->usb_iface->dev;
 	struct wa_seg *seg;
 	struct wa_xfer *xfer;
 	unsigned long flags;
 
+	*dto_waiting = 0;
+
 	spin_lock_irqsave(&rpipe->seg_lock, flags);
 	while (atomic_read(&rpipe->segs_available) > 0
-	      && !list_empty(&rpipe->seg_list)) {
+	      && !list_empty(&rpipe->seg_list)
+	      && (dto_acquired = __wa_dto_try_get(rpipe->wa))) {
 		seg = list_first_entry(&(rpipe->seg_list), struct wa_seg,
 				 list_node);
 		list_del(&seg->list_node);
 		xfer = seg->xfer;
-		result = __wa_seg_submit(rpipe, xfer, seg);
+		result = __wa_seg_submit(rpipe, xfer, seg, &dto_done);
+		/* release the dto resource if this RPIPE is done with it. */
+		if (dto_done)
+			__wa_dto_put(rpipe->wa);
 		dev_dbg(dev, "xfer %p ID %08X#%u submitted from delayed [%d segments available] %d\n",
 			xfer, wa_xfer_id(xfer), seg->index,
 			atomic_read(&rpipe->segs_available), result);
@@ -1176,7 +1240,37 @@ static void wa_xfer_delayed_run(struct wa_rpipe *rpipe)
 			spin_lock_irqsave(&rpipe->seg_lock, flags);
 		}
 	}
+	/*
+	 * Mark this RPIPE as waiting if dto was not acquired, there are
+	 * delayed segs and no active transfers to wake us up later.
+	 */
+	if (!dto_acquired && !list_empty(&rpipe->seg_list)
+		&& (atomic_read(&rpipe->segs_available) ==
+			le16_to_cpu(rpipe->descr.wRequests)))
+		*dto_waiting = 1;
+
 	spin_unlock_irqrestore(&rpipe->seg_lock, flags);
+
+	return dto_done;
+}
+
+static void wa_xfer_delayed_run(struct wa_rpipe *rpipe)
+{
+	int dto_waiting;
+	int dto_done = __wa_xfer_delayed_run(rpipe, &dto_waiting);
+
+	/*
+	 * If this RPIPE is waiting on the DTO resource, add it to the tail of
+	 * the waiting list.
+	 * Otherwise, if the WA DTO resource was acquired and released by
+	 *  __wa_xfer_delayed_run, another RPIPE may have attempted to acquire
+	 * DTO and failed during that time.  Check the delayed list and process
+	 * any waiters.  Start searching from the next RPIPE index.
+	 */
+	if (dto_waiting)
+		wa_add_delayed_rpipe(rpipe->wa, rpipe);
+	else if (dto_done)
+		wa_check_for_delayed_rpipes(rpipe->wa);
 }
 
 /*
@@ -1188,7 +1282,7 @@ static void wa_xfer_delayed_run(struct wa_rpipe *rpipe)
  */
 static int __wa_xfer_submit(struct wa_xfer *xfer)
 {
-	int result;
+	int result, dto_acquired = 0, dto_done = 0, dto_waiting = 0;
 	struct wahc *wa = xfer->wa;
 	struct device *dev = &wa->usb_iface->dev;
 	unsigned cnt;
@@ -1207,26 +1301,56 @@ static int __wa_xfer_submit(struct wa_xfer *xfer)
 	result = 0;
 	spin_lock_irqsave(&rpipe->seg_lock, flags);
 	for (cnt = 0; cnt < xfer->segs; cnt++) {
+		int delay_seg = 1;
+
 		available = atomic_read(&rpipe->segs_available);
 		empty = list_empty(&rpipe->seg_list);
 		seg = xfer->seg[cnt];
 		dev_dbg(dev, "xfer %p ID 0x%08X#%u: available %u empty %u (%s)\n",
 			xfer, wa_xfer_id(xfer), cnt, available, empty,
 			available == 0 || !empty ? "delayed" : "submitted");
-		if (available == 0 || !empty) {
+		if (available && empty) {
+			/*
+			 * Only attempt to acquire DTO if we have a segment
+			 * to send.
+			 */
+			dto_acquired = __wa_dto_try_get(rpipe->wa);
+			if (dto_acquired) {
+				delay_seg = 0;
+				result = __wa_seg_submit(rpipe, xfer, seg,
+							&dto_done);
+				if (dto_done)
+					__wa_dto_put(rpipe->wa);
+
+				if (result < 0) {
+					__wa_xfer_abort(xfer);
+					goto error_seg_submit;
+				}
+			}
+		}
+
+		if (delay_seg) {
 			seg->status = WA_SEG_DELAYED;
 			list_add_tail(&seg->list_node, &rpipe->seg_list);
-		} else {
-			result = __wa_seg_submit(rpipe, xfer, seg);
-			if (result < 0) {
-				__wa_xfer_abort(xfer);
-				goto error_seg_submit;
-			}
 		}
 		xfer->segs_submitted++;
 	}
 error_seg_submit:
+	/*
+	 * Mark this RPIPE as waiting if dto was not acquired, there are
+	 * delayed segs and no active transfers to wake us up later.
+	 */
+	if (!dto_acquired && !list_empty(&rpipe->seg_list)
+		&& (atomic_read(&rpipe->segs_available) ==
+			le16_to_cpu(rpipe->descr.wRequests)))
+		dto_waiting = 1;
 	spin_unlock_irqrestore(&rpipe->seg_lock, flags);
+
+	if (dto_waiting)
+		wa_add_delayed_rpipe(rpipe->wa, rpipe);
+	else if (dto_done)
+		wa_check_for_delayed_rpipes(rpipe->wa);
+
 	return result;
 }