cache-bypass explicitly reference pages involved in background i/o
Formerly, we assumed that any page eligible for background i/o could
be effectively pinned by lock_page. Persuant to concern expressed by
Simon that such pages might be eligible to be released by the kernel
after readpages returns, call get_page to increment the refcount on
each page before dispatching a background read (and matching put_page
when the i/o is completed).