I introduced this bug in 1.139 when I added the KASSERT(). Bleh. However, the code still needs to be protected at splbio(). And so I've committed.
-Matt : Hmmm. This looks like an out-an-out bug to me. The assertion is wrong. : It's scanning the backing_object and asserting that the pages in the : backing object are associated with object rather then backing_object. : : This section of code only runs when paging is in progress on a : collapseable object AND there are also idle pages in that object. : The collapse condition is probably due to an exiting process ( typical : in a buildworld ). : : ( from vm/vm_object.c ) : /* : * busy the page and move it from the backing store to the : * parent object. : */ : : vm_page_busy(p); : : KASSERT(p->object == object, ("vm_object_qcollapse(): object mismatch")); : ^^^^^^^^^^ : should be 'backing_object' : : There is also an interrupt race. Since paging can be in progress, : pages in the object can be ripped out from under it so we have to : run at splbio() in the loop. To Unsubscribe: send mail to majord...@freebsd.org with "unsubscribe freebsd-current" in the body of the message