Re: [PATCH v10 01/10] powerpc/mm: Implement set_memory() routines

[Jordan Niethe <jniethe5@gmail.com>]

On Wed, Mar 31, 2021 at 10:16 PM Michael Ellerman <mpe@ellerman.id.au> wrote:
>
> Hi Jordan,
>
> A few nits below ...
>
> Jordan Niethe <jniethe5@gmail.com> writes:
> > From: Russell Currey <ruscur@russell.cc>
> >
> > The set_memory_{ro/rw/nx/x}() functions are required for STRICT_MODULE_RWX,
> > and are generally useful primitives to have.  This implementation is
> > designed to be completely generic across powerpc's many MMUs.
> >
> > It's possible that this could be optimised to be faster for specific
> > MMUs, but the focus is on having a generic and safe implementation for
> > now.
> >
> > This implementation does not handle cases where the caller is attempting
> > to change the mapping of the page it is executing from, or if another
> > CPU is concurrently using the page being altered.  These cases likely
> > shouldn't happen, but a more complex implementation with MMU-specific code
> > could safely handle them, so that is left as a TODO for now.
> >
> > On hash the linear mapping is not kept in the linux pagetable, so this
> > will not change the protection if used on that range. Currently these
> > functions are not used on the linear map so just WARN for now.
> >
> > These functions do nothing if STRICT_KERNEL_RWX is not enabled.
> >
> > Reviewed-by: Daniel Axtens <dja@axtens.net>
> > Signed-off-by: Russell Currey <ruscur@russell.cc>
> > Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
> > [jpn: -rebase on next plus "powerpc/mm/64s: Allow STRICT_KERNEL_RWX again"
> >       - WARN on hash linear map]
> > Signed-off-by: Jordan Niethe <jniethe5@gmail.com>
> > ---
> > v10: WARN if trying to change the hash linear map
> > ---
> >  arch/powerpc/Kconfig                  |  1 +
> >  arch/powerpc/include/asm/set_memory.h | 32 ++++++++++
> >  arch/powerpc/mm/Makefile              |  2 +-
> >  arch/powerpc/mm/pageattr.c            | 88 +++++++++++++++++++++++++++
> >  4 files changed, 122 insertions(+), 1 deletion(-)
> >  create mode 100644 arch/powerpc/include/asm/set_memory.h
> >  create mode 100644 arch/powerpc/mm/pageattr.c
> >
> > diff --git a/arch/powerpc/Kconfig b/arch/powerpc/Kconfig
> > index fc7f5c5933e6..4498a27ac9db 100644
> > --- a/arch/powerpc/Kconfig
> > +++ b/arch/powerpc/Kconfig
> > @@ -135,6 +135,7 @@ config PPC
> >       select ARCH_HAS_MEMBARRIER_CALLBACKS
> >       select ARCH_HAS_MEMBARRIER_SYNC_CORE
> >       select ARCH_HAS_SCALED_CPUTIME          if VIRT_CPU_ACCOUNTING_NATIVE && PPC_BOOK3S_64
> > +     select ARCH_HAS_SET_MEMORY
>
> Below you do:
>
>         if (!IS_ENABLED(CONFIG_STRICT_KERNEL_RWX))
>                 return 0;
>
> Which suggests we should instead just only select ARCH_HAS_SET_MEMORY if
> STRICT_KERNEL_RWX ?
Yeah, I'm just going to do that.
>
>
> > diff --git a/arch/powerpc/mm/Makefile b/arch/powerpc/mm/Makefile
> > index 3b4e9e4e25ea..d8a08abde1ae 100644
> > --- a/arch/powerpc/mm/Makefile
> > +++ b/arch/powerpc/mm/Makefile
> > @@ -5,7 +5,7 @@
> >
> >  ccflags-$(CONFIG_PPC64)      := $(NO_MINIMAL_TOC)
> >
> > -obj-y                                := fault.o mem.o pgtable.o mmap.o maccess.o \
> > +obj-y                                := fault.o mem.o pgtable.o mmap.o maccess.o pageattr.o \
>
> .. and then the file should only be built if ARCH_HAS_SET_MEMORY = y.
>
> >                                  init_$(BITS).o pgtable_$(BITS).o \
> >                                  pgtable-frag.o ioremap.o ioremap_$(BITS).o \
> >                                  init-common.o mmu_context.o drmem.o
> > diff --git a/arch/powerpc/mm/pageattr.c b/arch/powerpc/mm/pageattr.c
> > new file mode 100644
> > index 000000000000..9efcb01088da
> > --- /dev/null
> > +++ b/arch/powerpc/mm/pageattr.c
> > @@ -0,0 +1,88 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +
> > +/*
> > + * MMU-generic set_memory implementation for powerpc
> > + *
> > + * Copyright 2019, IBM Corporation.
>
> Should be 2019-2021.
Right.
>
> > + */
> > +
> > +#include <linux/mm.h>
> > +#include <linux/set_memory.h>
> > +
> > +#include <asm/mmu.h>
> > +#include <asm/page.h>
> > +#include <asm/pgtable.h>
> > +
> > +
> > +/*
> > + * Updates the attributes of a page in three steps:
> > + *
> > + * 1. invalidate the page table entry
> > + * 2. flush the TLB
> > + * 3. install the new entry with the updated attributes
> > + *
> > + * This is unsafe if the caller is attempting to change the mapping of the
> > + * page it is executing from, or if another CPU is concurrently using the
> > + * page being altered.
>
> Is the 2nd part of that statement true?
>
> Or, I guess maybe it is true depending on what "unsafe" means.
>
> AIUI it's unsafe to use this on the page you're executing from, and by
> unsafe we mean the kernel will potentially crash because it will lose
> the mapping for the currently executing text.
>
> Using this on a page that another CPU is accessing could be safe, if eg.
> the other CPU is reading from the page and we are just changing it from
> RW->RO.
>
> So I'm not sure they're the same type of "unsafe".

I think the comment was prompted by your message here:
https://lore.kernel.org/linuxppc-dev/87pnio5fva.fsf@mpe.ellerman.id.au/

So I'll rewrite the comment to separate the two cases and indicate the
2nd case only might be an issue.
>
> > + * TODO make the implementation resistant to this.
> > + *
> > + * NOTE: can be dangerous to call without STRICT_KERNEL_RWX
>
> I don't think we need that anymore?
No we don't, change_memory_attr() won't call it without STRICT_KERNEL_RWX.
>
> > + */
> > +static int change_page_attr(pte_t *ptep, unsigned long addr, void *data)
> > +{
> > +     long action = (long)data;
> > +     pte_t pte;
> > +
> > +     spin_lock(&init_mm.page_table_lock);
> > +
> > +     /* invalidate the PTE so it's safe to modify */
> > +     pte = ptep_get_and_clear(&init_mm, addr, ptep);
> > +     flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
> > +
> > +     /* modify the PTE bits as desired, then apply */
> > +     switch (action) {
> > +     case SET_MEMORY_RO:
> > +             pte = pte_wrprotect(pte);
> > +             break;
>
> So set_memory_ro() removes write, but doesn't remove execute.
>
> That doesn't match my mental model of what "set to ro" means, but I
> guess I'm wrong because the other implementations seem to do something
> similar.
Hm, looking at arm and riscv it does seem to make it just RO.
>
>
> > +     case SET_MEMORY_RW:
> > +             pte = pte_mkwrite(pte);
>
> I think we want to add pte_mkdirty() here also to avoid a fault when the
> mapping is written to.
Right.
>
> eg. pmd_mkwrite(pmd_mkdirty(pte));
>
> > +             break;
> > +     case SET_MEMORY_NX:
> > +             pte = pte_exprotect(pte);
> > +             break;
> > +     case SET_MEMORY_X:
> > +             pte = pte_mkexec(pte);
> > +             break;
> > +     default:
> > +             WARN_ON_ONCE(1);
> > +             break;
> > +     }
> > +
> > +     set_pte_at(&init_mm, addr, ptep, pte);
> > +     spin_unlock(&init_mm.page_table_lock);
> > +
> > +     return 0;
> > +}
> > +
> > +int change_memory_attr(unsigned long addr, int numpages, long action)
> > +{
> > +     unsigned long start = ALIGN_DOWN(addr, PAGE_SIZE);
> > +     unsigned long sz = numpages * PAGE_SIZE;
> > +
> > +     if (!IS_ENABLED(CONFIG_STRICT_KERNEL_RWX))
> > +             return 0;
> > +
> > +     if (numpages <= 0)
> > +             return 0;
> > +
>
> This ↓ should have a comment explaining what it's doing:
Sure.
>
> > +#ifdef CONFIG_PPC_BOOK3S_64
> > +     if (WARN_ON_ONCE(!radix_enabled() &&
> > +                  get_region_id(addr) == LINEAR_MAP_REGION_ID)) {
> > +             return -1;
> > +     }
> > +#endif
>
> Maybe:
As Chrisophe says, we can't do that because those symbols aren't
defined for !CONFIG_PPC_BOOK3S_64.
>
>         if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) &&
>             WARN_ON_ONCE(!radix_enabled() && get_region_id(addr) == LINEAR_MAP_REGION_ID)) {
>                 return -1;
>         }
>
> But then Aneesh pointed out that we should also block VMEMMAP_REGION_ID.
>
> It might be better to just check for the permitted regions.
That would probably work better.
>
>         if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled()) {
>                 int region = get_region_id(addr);
>
>                 if (WARN_ON_ONCE(region != VMALLOC_REGION_ID && region != IO_REGION_ID))
>                         return -1;
>         }
>
> > +
> > +     return apply_to_existing_page_range(&init_mm, start, sz,
> > +                                         change_page_attr, (void *)action);
> > +}
>
>
> cheers