Re: [PATCHv2 2/3] powerpc/execmem: Allocate kprobes insn pages from the modules area on book3s/32
From: Mike Rapoport
Date: Mon Nov 16 2026 - 10:57:44 EST
On Mon, Nov 16, 2026 at 04:11:07PM +0100, Christophe Leroy (CS GROUP) wrote:
> Hi Mike,
>
> Le 15/11/2026 à 09:29, Mike Rapoport a écrit :
> > On Fri, Nov 13, 2026 at 06:01:37PM +0530, Ritesh Harjani (IBM) wrote:
> > > On book3s/32, a kprobe on an instruction that can't be emulated and has
> > > to be single stepped out of line, crashes the kernel. For e.g. with the
> > > following cmds...
> > > ~> echo 9 > /proc/sys/kernel/printk
> > > ~> mount -t debugfs debugfs /sys/kernel/debug
> > > ~> echo 'p:wx filldir64+0xb8' > /sys/kernel/debug/tracing/kprobe_events
> > > ~> echo 1 > /sys/kernel/debug/tracing/events/kprobes/wx/enable
> > > ~> echo 1 > /sys/kernel/debug/tracing/events/kprobes/enable
> > >
> > > ...we can see the following crash:
> > > kernel tried to execute exec-protected page (f101a008) - exploit attempt? (uid: 0)
> > > BUG: Unable to handle kernel instruction fetch
> > > Faulting instruction address: 0xc02b919c
> > > Vector: 400 (Instruction Access) at [f19a9d60]
> > > pc: c02b919c: filldir64+0xb8/0x2dc
> > > lr: c02b9148: filldir64+0x64/0x2dc
> > > sp: f19a9e20
> > > msr: 10029032
> > > current = 0xc112e900
> > > pid = 102, comm = ls
> > > enter ? for help
> > > [f19a9e20] c1a8d420 (unreliable)
> > > [f19a9e50] c02e4074 dcache_readdir+0x208/0x2a8
> > > [f19a9e90] c02b8c5c iterate_dir+0x88/0x1fc
> > > [f19a9ec0] c02b96e0 sys_getdents64+0x88/0x198
> > > [f19a9f10] c0014400 system_call_exception+0xf8/0x20c
> > > [f19a9f30] c00191ac ret_from_syscall+0x0/0x2c
> > > ---- Exception: c00 (System Call) at a7596164
> > > SP (af802310) is in userspace
> > >
> > > This happens because on book3s/32 the exec permission is per segment
> > > (256MB), not per page. The page tables have no hardware exec bit, only
> > > the segment registers have a no-execute (N) bit. mmu_mark_initmem_nx()
> > > sets N on all the kernel segments except the ones of the modules area.
> > > So nothing in the vmalloc area can be executed, even when the page is
> > > mapped with PAGE_KERNEL_ROX or PAGE_KERNEL_EXEC, because _PAGE_EXEC is
> > > a software only bit there and the segment's N bit is what the MMU
> > > enforces.
> > >
> > > Before commit 1b750c2fbf82 ("powerpc: extend execmem_params for kprobes
> > > allocations"), the kprobes insn pages came from module_alloc(). That
> > > commit moved them to the whole of the vmalloc space, on the grounds
> > > that kprobes use a breakpoint and don't need to be near the kernel
> > > text. That is fine on platforms where the exec permission is per page,
> > > as PAGE_KERNEL_ROX is then enough to make a vmalloc page executable.
> > > It is not on book3s/32, where the modules area is not only close to
> > > the kernel text but is also the only page-mapped area whose segment is
> > > left executable. So the insn pages must keep coming from there.
> > >
> > > Hence, use the modules area for EXECMEM_KPROBES on book3s/32.
> > Right, I missed that on book3s/32 does vmalloc is not executable and its
> > protection can't be changed.
> >
> > Could it be that 8xx also need a special treatment for kprobes?
>
> Yes indeed, not for the same reasons but because 8xx ITLB-miss handler
> doesn't expect any instruction TLB miss on kernel addresses other than the
> ones from the module area:
>
> Commit 4afb0be7ccda ("powerpc/8xx: Avoid testing for kernel space in ITLB
> Miss.")
> Commit 16a71c045186 ("powerpc/8xx: Preallocate execmem page tables")
> Commit 33c527522f39 ("powerpc/8xx: Inconditionally use task PGDIR in ITLB
> misses")
>
> So at the end as MODULES_VADDR/MODULES_END only exist on book3s/32 and 8xx I
> think when they exist they should be used as follows:
>
> diff --git a/arch/powerpc/mm/mem.c b/arch/powerpc/mm/mem.c
> index b617b69452cd..856f4609c488 100644
> --- a/arch/powerpc/mm/mem.c
> +++ b/arch/powerpc/mm/mem.c
> @@ -428,8 +428,13 @@ struct execmem_info __init *execmem_arch_setup(void)
> .fallback_end = fallback_end,
> },
> [EXECMEM_KPROBES] = {
> +#ifdef MODULES_VADDR
> + .start = MODULES_VADDR,
> + .end = MODULES_END,
> +#else
> .start = VMALLOC_START,
> .end = VMALLOC_END,
> +#endif
We already have #ifdef MODULES_VADDR a few lines above, I think the
assignment of kprobes_start and kprobes_end could go there instead of if
(IS_ENABLED(CONFIG_PPC_BOOK3S_32)).
> .pgprot = kprobes_prot,
> .alignment = 1,
> },
>
>
> >
> > > Fixes: 1b750c2fbf82 ("powerpc: extend execmem_params for kprobes allocations")
> > > Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@xxxxxxxxx>
> >
> > Reviewed-by: Mike Rapoport (Microsoft) <rppt@xxxxxxxxxx>
> >
> > > ---
> > > arch/powerpc/mm/mem.c | 10 ++++++++--
> > > 1 file changed, 8 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/arch/powerpc/mm/mem.c b/arch/powerpc/mm/mem.c
> > > index 77b63b52d8f2..ae1bc7359353 100644
> > > --- a/arch/powerpc/mm/mem.c
> > > +++ b/arch/powerpc/mm/mem.c
> > > @@ -391,6 +391,7 @@ struct execmem_info __init *execmem_arch_setup(void)
> > > pgprot_t kprobes_prot = strict_kernel_rwx_enabled() ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
> > > pgprot_t prot = strict_module_rwx_enabled() ? PAGE_KERNEL : PAGE_KERNEL_EXEC;
> > > unsigned long fallback_start = 0, fallback_end = 0;
> > > + unsigned long kprobes_start = VMALLOC_START, kprobes_end = VMALLOC_END;
> > > unsigned long start, end;
> > >
> > > /*
> > > @@ -410,6 +411,11 @@ struct execmem_info __init *execmem_arch_setup(void)
> > > }
> > >
> > > end = MODULES_END;
> > > +
> > > + if (IS_ENABLED(CONFIG_PPC_BOOK3S_32)) {
> > > + kprobes_start = MODULES_VADDR;
> > > + kprobes_end = MODULES_END;
> > > + }
> > > #else
> > > start = VMALLOC_START;
> > > end = VMALLOC_END;
> > > @@ -428,8 +434,8 @@ struct execmem_info __init *execmem_arch_setup(void)
> > > .fallback_end = fallback_end,
> > > },
> > > [EXECMEM_KPROBES] = {
> > > - .start = VMALLOC_START,
> > > - .end = VMALLOC_END,
> > > + .start = kprobes_start,
> > > + .end = kprobes_end,
> > > .pgprot = kprobes_prot,
> > > .alignment = 1,
> > > },
> > > --
> > > 2.39.5
> > >
> > >
> >
>
--
Sincerely yours,
Mike.