Send as many I/O requests to the I/O thread as possible, even though it will
still only handle one at a time.  This provides an opportunity to reduce
latency by starting one request before the previous one has been finished in
the driver.

Request handling is somewhat modernized by requesting sg pieces of a request
and handling them separately, finishing off the entire request after all the
pieces are done.

When a request queue stalls, normally because its pipe to the I/O thread is
full, it is put on the restart list.  This list is processed by starting up
the queues on it whenever there is some indication that progress might be
possible again.  Currently, this happens in the driver interrupt routine.
Some requests have been finished, so there is likely to be room in the pipe
again.

This almost doubles throughput when copying data between devices, but made no
noticable difference on anything else I tried.
Signed-off-by: Jeff Dike <jdike@linux.intel.com>
Cc: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it>
Cc: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>