Re: [PATCH v2 1/2] ARM: Fix some check warnings of tool sparse

[Russell King (Oracle)]

On Mon, Oct 10, 2022 at 06:14:56PM +0200, Ard Biesheuvel wrote:
> On Mon, 10 Oct 2022 at 18:08, Russell King (Oracle)
> <linux@xxxxxxxxxxxxxxx> wrote:
> >
> > On Mon, Oct 10, 2022 at 01:06:19PM +0200, Ard Biesheuvel wrote:
> > > On Mon, 10 Oct 2022 at 12:58, Leizhen (ThunderTown)
> > > <thunder.leizhen@xxxxxxxxxx> wrote:
> > > >
> > > >
> > > >
> > > > On 2022/10/10 18:20, Ard Biesheuvel wrote:
> > > > > On Mon, 10 Oct 2022 at 11:56, Zhen Lei <thunder.leizhen@xxxxxxxxxx> wrote:
> > > > >>
> > > > >> Fix the following warnings:
> > > > >>  warning: incorrect type in initializer (different address spaces)
> > > > >>     expected unsigned short [noderef] __user *register __p
> > > > >>     got unsigned short [usertype] *
> > > > >>  warning: cast removes address space '__user' of expression
> > > > >>
> > > > >> Signed-off-by: Zhen Lei <thunder.leizhen@xxxxxxxxxx>
> > > > >> ---
> > > > >>  arch/arm/kernel/traps.c | 10 +++++-----
> > > > >>  1 file changed, 5 insertions(+), 5 deletions(-)
> > > > >>
> > > > >> diff --git a/arch/arm/kernel/traps.c b/arch/arm/kernel/traps.c
> > > > >> index 20b2db6dcd1ced7..34aa80c09c508c1 100644
> > > > >> --- a/arch/arm/kernel/traps.c
> > > > >> +++ b/arch/arm/kernel/traps.c
> > > > >> @@ -188,9 +188,9 @@ static void dump_instr(const char *lvl, struct pt_regs *regs)
> > > > >>                         }
> > > > >>                 } else {
> > > > >>                         if (thumb)
> > > > >> -                               bad = get_user(val, &((u16 *)addr)[i]);
> > > > >> +                               bad = get_user(val, &((u16 __user *)addr)[i]);
> > > > >>                         else
> > > > >> -                               bad = get_user(val, &((u32 *)addr)[i]);
> > > > >> +                               bad = get_user(val, &((u32 __user *)addr)[i]);
> > > > >>                 }
> > > > >>
> > > > >>                 if (!bad)
> > > > >> @@ -455,15 +455,15 @@ asmlinkage void do_undefinstr(struct pt_regs *regs)
> > > > >>         if (processor_mode(regs) == SVC_MODE) {
> > > > >>  #ifdef CONFIG_THUMB2_KERNEL
> > > > >>                 if (thumb_mode(regs)) {
> > > > >> -                       instr = __mem_to_opcode_thumb16(((u16 *)pc)[0]);
> > > > >> +                       instr = __mem_to_opcode_thumb16(((__force u16 *)pc)[0]);
> > > > >
> > > > > Shouldn't this be __user as well? (and below)
> > > >
> > > > unsigned int instr;
> > > > void __user *pc;
> > > >
> > > > The __user can clear the warning, but a new warning will be generated.
> > > >
> > > > instr = __mem_to_opcode_thumb16(((u16 *)pc)[0]);
> > > >       ^new                           ^old
> > > >
> > > > arch/arm/kernel/traps.c:473:33: warning: dereference of noderef expression
> > > >
> > >
> > > This is because dereferencing a __user pointer is not permitted.
> > >
> > > So this code should be using get_kernel_nofault() here not a plain
> > > dereference of PC. So better to fix that properly instead of papering
> > > over it with a __force cast just to make sparse happy.
> >
> > Why? We won't get here unless the PC can be dereferenced. If it's not
> > able to be dereferenced, then we'd be dealing with a prefetch abort.
> >
> 
> If that is guaranteed (i.e., there is no way we might be racing with a
> module unload on another CPU or something like that), then I agree
> that dereferencing PC is fine.

If we get here for an instruction in a module that's being unloaded, we
have way bigger problems. We shouldn't be executing code in a module
being unloaded in the first place. That becomes a case of "deserves to
oops".

The more likely case would be a prefetch abort when the page is
unmapped. You'd have to try pretty hard to get an undef to race with
a module unload.

-- 
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTP is here! 40Mbps down 10Mbps up. Decent connectivity at last!