VM: improve pageout deadlock workaround

Commit 5dd4f67522ad0d49a2cecdb9b109251f546d4dd1 makes VM map entry
allocation done with VM privilege, so that a VM map isn't held locked
while physical allocations are paused, which may block the default
pager during page eviction, causing a system-wide deadlock.

First, it turns out that map entries aren't the only buffers allocated,
and second, their number can't be easily determined, which makes a
preallocation strategy very hard to implement.

This change generalizes the strategy of VM privilege increase when a
VM map is locked.

* device/ds_routines.c (io_done_thread): Use integer values instead
of booleans when setting VM privilege.
* kern/thread.c (thread_init, thread_wire): Likewise.
* vm/vm_pageout.c (vm_pageout): Likewise.
* kern/thread.h (struct thread): Turn member `vm_privilege' into an
unsigned integer.
* vm/vm_map.c (vm_map_lock): New function, where VM privilege is
temporarily increased.
(vm_map_unlock): New function, where VM privilege is decreased.
(_vm_map_entry_create): Remove VM privilege workaround from this
function.
* vm/vm_map.h (vm_map_lock, vm_map_unlock): Turn into functions.

1 files changed, 1 insertions, 1 deletions

diff --git a/device/ds_routines.c b/device/ds_routines.c
index 445e7ae1..1fabec3c 100644
--- a/device/ds_routines.c
+++ b/device/ds_routines.c
@@ -1512,7 +1512,7 @@ void io_done_thread(void)
 	/*
 	 * Set thread privileges and highest priority.
 	 */
-	current_thread()->vm_privilege = TRUE;
+	current_thread()->vm_privilege = 1;
 	stack_privilege(current_thread());
 	thread_set_own_priority(0);