Re: kmem_alloc() vs. VM_MIN_KERNEL_ADDRESS

[Maxime Villard <max%m00nbsd.net@localhost>]

Le 02/07/2019 à 16:21, J. Hannken-Illjes a écrit :
> > On 1. Jul 2019, at 20:12, Maxime Villard <max%M00nBSD.net@localhost> wrote:
> >
> > Le 01/07/2019 à 19:40, J. Hannken-Illjes a écrit :
> > > Sometimes kmem_alloc() returns an address below
> > > VM_MIN_KERNEL_ADDRESS, something like
> > >    kmem_alloc() -> 0xffffaba25800a5a8 with
> > >    VM_MIN_KERNEL_ADDRESS 0xffffad8000000000
> > > It doesn't happen after every reboot and breaks
> > > dtrace which treats the region 0.. VM_MIN_KERNEL_ADDRESS
> > > as toxic.
> > > How could we test against the lowest address the kernel
> > > may use?
> > > --
> > > J. Hannken-Illjes - hannken%eis.cs.tu-bs.de@localhost - TU Braunschweig
> >
> > Mmh, this seems to be a side effect of 1/2-KASLR which is enabled by default
> > in GENERIC. Reminder: in GENERIC, we now randomize by default every kernel
> > memory region.
> >
> > It is now possible that the direct map is below the main memory map -- it was
> > not the case before. kmem_alloc uses the direct map for certain allocations.
> >
> > First, please confirm by also dumping PMAP_DIRECT_BASE and PMAP_DIRECT_END to
> > make sure the area returned by kmem_alloc is indeed in the dmap when dtrace
> > doesn't work.
> >
> > Then, it seems that dtrace should be taught to consider as toxic everything
> > that is neither in the main map nor in the dmap. That is, only these ranges
> > are valid:
> >
> > 	VM_MIN_KERNEL_ADDRESS -> VM_MAX_KERNEL_ADDRESS
> > 	PMAP_DIRECT_BASE -> PMAP_DIRECT_END
> >
> > There is no actual "lowest" address the kernel may use; there are just two
> > separate regions that are valid. (Plus the kernel image itself, if it
> > matters.)
>
> It would work but unfortunately dtrace is a module and cannot
> use PMAP_DIRECT_BASE aka. pmap_direct_base.  These variables
> are not visible from module context and don't exist if the
> kernel is built with KASAN.

Ok.

> Solaris uses the toxic areas 0 .. _userlimit and all regions
> mapped from devices.
>
> Will the kernel ever map anything below
> VM_MIN_KERNEL_ADDRESS_DEFAULT?

No.

> Do we have a mapping that contains all device-mapped memory?

No, it's either the main map or the direct map.