USB: serial: fix race in generic write

Fix race in generic write implementation, which could lead to
temporarily degraded throughput.

The current generic write implementation introduced by commit
27c7acf22047 ("USB: serial: reimplement generic fifo-based writes") has
always had this bug, although it's fairly hard to trigger and the
consequences are not likely to be noticed.

Specifically, a write() on one CPU while the completion handler is
running on another could result in only one of the two write urbs being
utilised to empty the remainder of the write fifo (unless there is a
second write() that doesn't race during that time).

Cc: stable <stable@vger.kernel.org> # 2.6.35
Signed-off-by: Johan Hovold <jhovold@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

diff --git a/drivers/usb/serial/generic.c b/drivers/usb/serial/generic.c
index 2b01ec8..e36b25a 100644 (file)
--- a/drivers/usb/serial/generic.c
+++ b/drivers/usb/serial/generic.c
@@ -173,16 +173,8 @@ retry:
                clear_bit_unlock(USB_SERIAL_WRITE_BUSY, &port->flags);
                return result;
        }
-       /*
-        * Try sending off another urb, unless called from completion handler
-        * (in which case there will be no free urb or no data).
-        */
-       if (mem_flags != GFP_ATOMIC)
-               goto retry;
 
-       clear_bit_unlock(USB_SERIAL_WRITE_BUSY, &port->flags);
-
-       return 0;
+       goto retry;     /* try sending off another urb */
 }
 EXPORT_SYMBOL_GPL(usb_serial_generic_write_start);