Re: [PATCH] arm64: io: Reject present-invalid user prot in ioremap_prot()

[Catalin Marinas]

On Wed, Sep 09, 2026 at 02:04:50PM +0100, Will Deacon wrote:
> On Wed, Sep 09, 2026 at 01:50:09PM +0100, Catalin Marinas wrote:
> > On Sat, Sep 05, 2026 at 11:31:48AM +0800, Zeng Heng wrote:
> > > From: Zeng Heng <zengheng4@xxxxxxxxxx>
> > > 
> > > Mapping a stack-top page via /dev/mem with PROT_NONE and then reading
> > > that process's /proc/<pid>/cmdline triggers a spurious WARN in
> > > ioremap_prot() through generic_access_phys():
> > > 
> > >   WARNING: ./arch/arm64/include/asm/io.h:275 at generic_access_phys
> > >   Call trace:
> > >     generic_access_phys+0x1c8/0x228 (P)
> > >     __access_remote_vm+0x2b4/0x398
> > >     access_remote_vm+0x14/0x30
> > >     get_mm_cmdline+0xf8/0x2a0
> > >     proc_pid_cmdline_read+0x68/0x120
> > > 
> > > generic_access_phys() passes the full pgprot derived from the user PTE
> > > to ioremap_prot(). A PROT_NONE /dev/mem mapping is encoded as PAGE_NONE,
> > > which clears PTE_VALID and sets the software PTE_PRESENT_INVALID bit.
> > > On arm64 such an entry is still pte_present(), so follow_pfnmap_start()
> > > reports the pfn and generic_access_phys() reaches ioremap_prot().
> > > The PTE_USER assertion, which is meant to catch kernel prots being
> > > passed by mistake, then fires for a PROT_NONE user mapping
> > > that legitimately lacks PTE_USER, producing the spurious WARN.
> > > 
> > > Reject a user prot encoding a present-invalid (i.e. PROT_NONE) entry up
> > > front so that generic_access_phys() cleanly fails the access instead
> > > of warning. Note that PTE_PRESENT_INVALID aliases the PTE_NG bit and
> > > is only meaningful when PTE_VALID is clear, so both bits must be
> > > checked together.
> > > 
> > > Fixes: 8f098037139b ("arm64: io: Extract user memory type in ioremap_prot()")
> > > Signed-off-by: Zeng Heng <zengheng4@xxxxxxxxxx>
> > > ---
> > >  arch/arm64/include/asm/io.h | 4 ++++
> > >  1 file changed, 4 insertions(+)
> > > 
> > > diff --git a/arch/arm64/include/asm/io.h b/arch/arm64/include/asm/io.h
> > > index 21c8e400107c..bbfbc4682639 100644
> > > --- a/arch/arm64/include/asm/io.h
> > > +++ b/arch/arm64/include/asm/io.h
> > > @@ -272,6 +272,10 @@ static inline void __iomem *ioremap_prot(phys_addr_t phys, size_t size,
> > >  	pgprot_t prot;
> > >  	ptval_t user_prot_val = pgprot_val(user_prot);
> > >  
> > > +	if ((user_prot_val & (PTE_VALID | PTE_PRESENT_INVALID)) ==
> > > +	     PTE_PRESENT_INVALID)
> > > +		return NULL;
> > > +
> > >  	if (WARN_ON_ONCE(!(user_prot_val & PTE_USER)))
> > >  		return NULL;
> > 
> > I wonder whether we should just drop the warning and return NULL if
> > !PTE_USER && PTE_UXN. The latter check would also catch execute-only
> > mappings (Sashiko pointed out this case still trips the warning). The
> > PROT_NONE case would be covered automatically as well since PTE_USER is
> > cleared, PTE_UXN set. Maybe add a comment that that pte_protnone()
> > relies on !PTE_USER && PTE_UXN, so it's not that we avoid the PROT_NONE
> > issue by chance.
> 
> Hmm, do we want exec-only mappings to be readable via /dev/mem?

Ah, yes, got confused on how it reaches this path. It doesn't make sense
to allow exec-only to be readable. Also if we mmap(PROT_EXEC) /dev/mem,
PTE_UXN ends up set anyway via pgprot_noncached(). But it does have
PTE_VALID, so the above won't catch it.

Checking !PTE_USER should be sufficient here and return NULL. For the
warning, I think we can check PTE_NG first but it only works without
kpti.

-- 
Catalin