Linux: Avoid unnecessary splitting of bulk transfers

* With the latest kernels it is no longer needed to always split large bulk
  transfers into multiple urbs. This patch takes advantage of this by not
  splitting when not necessary. Note that the non-split code path is in essence
  using the old split code path with an urb count which is always 1.
* This leads to more sane handling of large transfers with recent kernels,
  although our splitting code is well tested, not splitting at all still is
  a lot better :)
* When used with a recent kernel, this also fixes the problems, on XHCI attached
  devices, when a large bulk-in transfer ends with a short read in an urb other
  then the last urb.
* For more on this see: http://marc.info/?l=linux-usb&m=133797554910355

2 files changed, 58 insertions, 13 deletions

diff --git a/libusb/os/linux_usbfs.c b/libusb/os/linux_usbfs.c
index 2667b11..a8768be 100644
--- a/libusb/os/linux_usbfs.c
+++ b/libusb/os/linux_usbfs.c
@@ -1668,6 +1668,7 @@ static int submit_bulk_transfer(struct usbi_transfer *itransfer,
 	struct usbfs_urb *urbs;
 	int is_out = (transfer->endpoint & LIBUSB_ENDPOINT_DIR_MASK)
 		== LIBUSB_ENDPOINT_OUT;
+	int bulk_buffer_len, use_bulk_continuation;
 	int r;
 	int i;
 	size_t alloc_size;
@@ -1679,16 +1680,54 @@ static int submit_bulk_transfer(struct usbi_transfer *itransfer,
 			!(dpriv->caps & USBFS_CAP_ZERO_PACKET))
 		return LIBUSB_ERROR_NOT_SUPPORTED;
 
-	/* usbfs places a 16kb limit on bulk URBs. we divide up larger requests
-	 * into smaller units to meet such restriction, then fire off all the
-	 * units at once. it would be simpler if we just fired one unit at a time,
-	 * but there is a big performance gain through doing it this way. */
-	int num_urbs = transfer->length / MAX_BULK_BUFFER_LENGTH;
+	/*
+	 * Older versions of usbfs place a 16kb limit on bulk URBs. We work
+	 * around this by splitting large transfers into 16k blocks, and then
+	 * submit all urbs at once. it would be simpler to submit one urb at
+	 * a time, but there is a big performance gain doing it this way.
+	 *
+	 * Newer versions lift the 16k limit (USBFS_CAP_NO_PACKET_SIZE_LIM),
+	 * using arbritary large transfers can still be a bad idea though, as
+	 * the kernel needs to allocate physical contiguous memory for this,
+	 * which may fail for large buffers.
+	 *
+	 * The kernel solves this problem by splitting the transfer into
+	 * blocks itself when the host-controller is scatter-gather capable
+	 * (USBFS_CAP_BULK_SCATTER_GATHER), which most controllers are.
+	 *
+	 * Last, there is the issue of short-transfers when splitting, for
+	 * short split-transfers to work reliable USBFS_CAP_BULK_CONTINUATION
+	 * is needed, but this is not always available.
+	 */
+	if (dpriv->caps & USBFS_CAP_BULK_SCATTER_GATHER) {
+		/* Good! Just submit everything in one go */
+		bulk_buffer_len = transfer->length ? transfer->length : 1;
+		use_bulk_continuation = 0;
+	} else if (dpriv->caps & USBFS_CAP_BULK_CONTINUATION) {
+		/* Split the transfers and use bulk-continuation to
+		   avoid issues with short-transfers */
+		bulk_buffer_len = MAX_BULK_BUFFER_LENGTH;
+		use_bulk_continuation = 1;
+	} else if (dpriv->caps & USBFS_CAP_NO_PACKET_SIZE_LIM) {
+		/* Don't split, assume the kernel can alloc the buffer
+		   (otherwise the submit will fail with -ENOMEM) */
+		bulk_buffer_len = transfer->length ? transfer->length : 1;
+		use_bulk_continuation = 0;
+	} else {
+		/* Bad, splitting without bulk-continuation, short transfers
+		   which end before the last urb will not work reliable! */
+		/* Note we don't warn here as this is "normal" on kernels <
+		   2.6.32 and not a problem for most applications */
+		bulk_buffer_len = MAX_BULK_BUFFER_LENGTH;
+		use_bulk_continuation = 0;
+	}
+
+	int num_urbs = transfer->length / bulk_buffer_len;
 	int last_urb_partial = 0;
 
 	if (transfer->length == 0) {
 		num_urbs = 1;
-	} else if ((transfer->length % MAX_BULK_BUFFER_LENGTH) > 0) {
+	} else if ((transfer->length % bulk_buffer_len) > 0) {
 		last_urb_partial = 1;
 		num_urbs++;
 	}
@@ -1709,17 +1748,17 @@ static int submit_bulk_transfer(struct usbi_transfer *itransfer,
 		urb->usercontext = itransfer;
 		urb->type = urb_type;
 		urb->endpoint = transfer->endpoint;
-		urb->buffer = transfer->buffer + (i * MAX_BULK_BUFFER_LENGTH);
-		if ((dpriv->caps & USBFS_CAP_BULK_CONTINUATION) && !is_out)
+		urb->buffer = transfer->buffer + (i * bulk_buffer_len);
+		if (use_bulk_continuation && !is_out)
 			urb->flags = USBFS_URB_SHORT_NOT_OK;
 		if (i == num_urbs - 1 && last_urb_partial)
-			urb->buffer_length = transfer->length % MAX_BULK_BUFFER_LENGTH;
+			urb->buffer_length = transfer->length % bulk_buffer_len;
 		else if (transfer->length == 0)
 			urb->buffer_length = 0;
 		else
-			urb->buffer_length = MAX_BULK_BUFFER_LENGTH;
+			urb->buffer_length = bulk_buffer_len;
 
-		if (i > 0 && (dpriv->caps & USBFS_CAP_BULK_CONTINUATION))
+		if (i > 0 && use_bulk_continuation)
 			urb->flags |= USBFS_URB_BULK_CONTINUATION;
 
 		/* we have already checked that the flag is supported */
@@ -1806,7 +1845,13 @@ static int submit_iso_transfer(struct usbi_transfer *itransfer)
 	/* usbfs places a 32kb limit on iso URBs. we divide up larger requests
 	 * into smaller units to meet such restriction, then fire off all the
 	 * units at once. it would be simpler if we just fired one unit at a time,
-	 * but there is a big performance gain through doing it this way. */
+	 * but there is a big performance gain through doing it this way.
+	 *
+	 * Newer kernels lift the 32k limit (USBFS_CAP_NO_PACKET_SIZE_LIM),
+	 * using arbritary large transfers is still be a bad idea though, as
+	 * the kernel needs to allocate physical contiguous memory for this,
+	 * which may fail for large buffers.
+	 */
 
 	/* calculate how many URBs we need */
 	for (i = 0; i < num_packets; i++) {
diff --git a/libusb/version_nano.h b/libusb/version_nano.h
index b80cade..97588fa 100644
--- a/libusb/version_nano.h
+++ b/libusb/version_nano.h
@@ -1 +1 @@
-#define LIBUSB_NANO 10554
+#define LIBUSB_NANO 10555