Re: [PATCH] module/ksymtab: use 64-bit relative reference for target symbol

[Ard Biesheuvel]

On Fri, 24 May 2019 at 18:31, Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
>
> On Fri, May 24, 2019 at 05:55:37PM +0200, Ard Biesheuvel wrote:
> > On Fri, 24 May 2019 at 17:21, Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> > >
> > > On Thu, May 23, 2019 at 10:29:39AM +0100, Ard Biesheuvel wrote:
> > > >
> > > >
> > > > On 5/23/19 10:18 AM, Will Deacon wrote:
> > > > > On Thu, May 23, 2019 at 09:41:40AM +0100, Ard Biesheuvel wrote:
> > > > > >
> > > > > >
> > > > > > On 5/22/19 5:28 PM, Ard Biesheuvel wrote:
> > > > > > >
> > > > > > >
> > > > > > > On 5/22/19 4:02 PM, Ard Biesheuvel wrote:
> > > > > > > > The following commit
> > > > > > > >
> > > > > > > >     7290d5809571 ("module: use relative references for __ksymtab entries")
> > > > > > > >
> > > > > > > > updated the ksymtab handling of some KASLR capable architectures
> > > > > > > > so that ksymtab entries are emitted as pairs of 32-bit relative
> > > > > > > > references. This reduces the size of the entries, but more
> > > > > > > > importantly, it gets rid of statically assigned absolute
> > > > > > > > addresses, which require fixing up at boot time if the kernel
> > > > > > > > is self relocating (which takes a 24 byte RELA entry for each
> > > > > > > > member of the ksymtab struct).
> > > > > > > >
> > > > > > > > Since ksymtab entries are always part of the same module as the
> > > > > > > > symbol they export (or of the core kernel), it was assumed at the
> > > > > > > > time that a 32-bit relative reference is always sufficient to
> > > > > > > > capture the offset between a ksymtab entry and its target symbol.
> > > > > > > >
> > > > > > > > Unfortunately, this is not always true: in the case of per-CPU
> > > > > > > > variables, a per-CPU variable's base address (which usually differs
> > > > > > > > from the actual address of any of its per-CPU copies) could be at
> > > > > > > > an arbitrary offset from the ksymtab entry, and so it may be out
> > > > > > > > of range for a 32-bit relative reference.
> > > > > > > >
> > > > > >
> > > > > > (Apologies for the 3-act monologue)
> > > > >
> > > > > Exposition, development and recapitulation ;)
> > > > >
> > > > > > This turns out to be incorrect. The symbol address of per-CPU variables
> > > > > > exported by modules is always in the vicinity of __per_cpu_start, and so it
> > > > > > is simply a matter of making sure that the core kernel is in range for
> > > > > > module ksymtab entries containing 32-bit relative references.
> > > > > >
> > > > > > When running the arm64 with kaslr enabled, we currently randomize the module
> > > > > > space based on the range of ADRP/ADD instruction pairs, which have a -/+ 4
> > > > > > GB range rather than the -/+ 2 GB range of 32-bit place relative data
> > > > > > relocations. So we can fix this by simply reducing the randomization window
> > > > > > to 2 GB.
> > > > >
> > > > > Makes sense. Do you see the need for an option to disable PREL relocs
> > > > > altogether in case somebody wants the additional randomization range?
> > > > >
> > > >
> > > > No, not really. To be honest, I don't think
> > > > CONFIG_RANDOMIZE_MODULE_REGION_FULL is that useful to begin with, and the
> > > > only reason we enabled it by default at the time was to ensure that the PLT
> > > > code got some coverage after we introduced it.
> > >
> > > In code, percpu variables are accessed with absolute relocations, right?
> >
> > No, they are accessed just like ordinary symbols, so PC32 references
> > on x86 or ADRP/ADD references on arm64 are both quite common.
>
> Ah, right, now I see some PC32 percpu references.
>
> So if PC32 references are sufficient for code, why aren't they
> sufficient for ksymtab entries?  Isn't the ksymtab data address closer
> to the percpu data than the code?  Do you have an example of an out of
> range ksymtab reference?
>

Not on x86, only on arm64, which uses ADRP/ADD pairs with a -/+ 4 GB
range as opposed to the -/+ 2 GB range of PC32 and PREL32 references.
So when KASLR puts the modules far away from the kernel (but in range
for ADRP/ADD) they may be out of range for PREL32.