kernel - Fix bug in uiomove() error handling, fix deadlock in read_shortcut
* Fix a recently introduced bug in uiomove() related to error handling
when iovcnt is greater than 1.
* Fix a deadlock in the vm.read_shortcut code, rare triggerable except
by vkernels.
The deadlock is due to a busied VM page being held across a uiomove().
To fix a new function uiomove_nofault() was added which causes EFAULT
to be returned for any non-trivial VM fault (any fault related to nested
objects or vnode objects).
The vop_helper_read_shortcut() procedure uses this new uiomove function
and breaks out when the case occurs, returning and falling through to
the filesystem's normal buffer-cache-based read() code.
* As an added bonus, we no longer have to unlock the VM object across
the new uiomove_nofault() call in the read shorcut code, since it is
no longer possible to deadlock against the VM object.
Reported-by: tuxillo