K 10
svn:author
V 7
attilio
K 8
svn:date
V 27
2013-03-14T01:53:32.428349Z
K 7
svn:log
V 1536
Move the page handling of sleeping on a page for a specific condition
under the protection of the page lock.
In detail:
1) VPO_WANTED is moved to be a per-page flag, removing it and introducing
   PG_WANTED.
2) The function vm_page_sleep_onpage() is introduced to work as an
   underlying version of vm_page_sleep() / vm_page_sleep_if_busy() to
   specify priority and timo of the sleeping operation.
3) Convert all the current sleepers on the page to not use anymore
   VM_OBJECT_SLEEP() but use vm_page_sleep().  In one case, where special
   timo, priority and return value are needed, vm_page_sleep_onpage() is
   used directly.
4) The operation results without races because the object lock is now
   dropped after the page-lock is held.  This creates a chain of trust.

Further notes:
- vm_page_sleep() and neighboor functions now always want the page-lock
  held.  The semantic of the unlocking is left unchanged.
- The burden to deal with page-locking is on the callers of the function.
  This is because soon more per-page attributes will be moved under the
  per-page lock, making the protection path a single critical section.
- The vm_page_sleep_if_busy() is temporary made an hard function.  This is
  to avoid namespace pollution by the object locking.  With more changes
  the object locking won't be necessary anymore there and it will be made
  inline again as it should be.
- vm_page_flash() and vm_page_sleep() lost the necessity to be called
  with the object lock held.

Sponsored by:	EMC / Isilon storage division

END