Samuel Thibault, le Sat 06 Sep 2014 11:39:34 +0200, a écrit : > Justus Winter, le Fri 05 Sep 2014 14:53:52 +0200, a écrit : > > there is a bug affecting vm_map. Contrary to the documentation, > > address is not ignored if anywhere is given, leading to spurious > > KERN_NO_SPACE errors. > > Uh, that's bad indeed. > > Apparently most of our source code copes with that by explicitly setting > address to 0, but some places such as pager_memcpy don't. > > I'm tempted to rather fix the behavior according to the documentation, > and go with something like this: > > diff --git a/vm/vm_user.c b/vm/vm_user.c > index f7c87cc..f741705 100644 > --- a/vm/vm_user.c > +++ b/vm/vm_user.c > @@ -336,6 +336,9 @@ kern_return_t vm_map( > if (size == 0) > return KERN_INVALID_ARGUMENT; > > + if (anywhere) > + *address = 0; > + > *address = trunc_page(*address); > size = round_page(size); > > I don't see any place in our source code which would depend on 'address' > being a hint for the map address. glibc's mmap() copes with the kernel > behavior for instance.
Actually it seems the defpager does depend on the behavior, so I have reverted the change in the debian gnumach=2:1.4+git20141109-1 package for now, which fixes swapping out to disk. I'll also upload a hurd package with vm_map fixes according to the undocumented behavior, along other packaging fixes. Samuel
