Re: [RFC PATCH] arch: arm64: introduce RODATA_FULL_USE_PTE_CONT

From: Ard Biesheuvel
Date: Mon Nov 22 2021 - 04:23:26 EST


On Mon, 22 Nov 2021 at 10:17, Zhaoyang Huang <huangzhaoyang@xxxxxxxxx> wrote:
>
> On Mon, Nov 22, 2021 at 5:04 PM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote:
> >
> > On Mon, 22 Nov 2021 at 10:00, Zhaoyang Huang <huangzhaoyang@xxxxxxxxx> wrote:
> > >
> > > On Mon, Nov 22, 2021 at 4:52 PM Ard Biesheuvel <ardb@xxxxxxxxxx> wrote:
> > > >
> > > > On Mon, 22 Nov 2021 at 06:29, Huangzhaoyang <huangzhaoyang@xxxxxxxxx> wrote:
> > > > >
> > > > > From: Zhaoyang Huang <zhaoyang.huang@xxxxxxxxxx>
> > > > >
> > > > > Kernel linear mapping will be split to the smallest granularity when
> > > > > RODATA_FULL applied, which could lead to TLB pressure. Introduce a method
> > > > > to apply PTE_CONT on pte.
> > > > >
> > > > > Signed-off-by: Zhaoyang Huang <zhaoyang.huang@xxxxxxxxxx>
> > > >
> > > > How would this lead to TLB pressure, and how does the use of
> > > > contiguous mappings mitigate that? The linear mapping of the kernel is
> > > > rarely used, as all normal accesses to it go via the vmalloc region,
> > > > so in which case would TLB entries be allocated for this region in a
> > > > way that could cause a measurable performance impact?
> > > In fact, the patch is about to use PTE_CONT *OUT OF* the range of
> > > kernel text.
> >
> > OK, I had missed that, apologies.
> >
> > > Would you please have a look at the code. It apply
> > > PTE_CONT during map_mem and then clear it when load_module change the
> > > corresponding linear mapping to the area it use in vmalloc area.
> >
> > You cannot change the PTE_CONT attribute on live mappings.
> Is it an inner constraint from ASIC perspective? PTE_CONT is different
> to other attributes?

The PTE_CONT attributes on live translation entries must always be in
sync in all entries covering the same contiguous group, or you might
cause TLB conflict aborts.

> >
> > > >
> > > >
> > > > > ---
> > > > > arch/arm64/Kconfig | 9 +++++++++
> > > > > arch/arm64/mm/mmu.c | 10 ++++++++--
> > > > > arch/arm64/mm/pageattr.c | 9 +++++++++
> > > > > 3 files changed, 26 insertions(+), 2 deletions(-)
> > > > >
> > > > > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> > > > > index fee914c..3f8fbf0 100644
> > > > > --- a/arch/arm64/Kconfig
> > > > > +++ b/arch/arm64/Kconfig
> > > > > @@ -1198,6 +1198,15 @@ config RODATA_FULL_DEFAULT_ENABLED
> > > > > This requires the linear region to be mapped down to pages,
> > > > > which may adversely affect performance in some cases.
> > > > >
> > > > > +config RODATA_FULL_USE_PTE_CONT
> > > > > + bool "Apply PTE_CONT when RODATA_FULL_DEFAULT_ENABLED enabled"
> > > > > + depends on RODATA_FULL_DEFAULT_ENABLED
> > > > > + default y
> > > > > + help
> > > > > + Apply PTE_CONT on linear mapping as much as we can when
> > > > > + RODATA_FULL_DEFAULT_ENABLED enabled which could decrease the
> > > > > + impaction on performance by small pte granularity.
> > > > > +
> > > > > config ARM64_SW_TTBR0_PAN
> > > > > bool "Emulate Privileged Access Never using TTBR0_EL1 switching"
> > > > > help
> > > > > diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
> > > > > index cfd9deb..8017b17 100644
> > > > > --- a/arch/arm64/mm/mmu.c
> > > > > +++ b/arch/arm64/mm/mmu.c
> > > > > @@ -124,15 +124,21 @@ static bool pgattr_change_is_safe(u64 old, u64 new)
> > > > > * The following mapping attributes may be updated in live
> > > > > * kernel mappings without the need for break-before-make.
> > > > > */
> > > > > +#ifndef CONFIG_RODATA_FULL_USE_PTE_CONT
> > > > > pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG;
> > > > > +#else
> > > > > + pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG | PTE_CONT;
> > > > > +#endif
> >
> > What justifies this change? Why is it ok in certain cases to update
> > the PTE_CONT attribute without BBM?
> BBM?
> I am not sure if it is ok here, but I just find nobody else change
> PTE_CONT so far other than this commit.

No it is not ok. BBM == break before make, which means you need to
install invalid entries for the entire contiguous group before you can
change the PTE_CONT attributes. As other CPUs may be accessing the
same region, there is no safe way to do this, which is why we don't
permit PTE_CONT to be changed at all.


> >
> > > > >
> > > > > /* creating or taking down mappings is always safe */
> > > > > if (old == 0 || new == 0)
> > > > > return true;
> > > > >
> > > > > /* live contiguous mappings may not be manipulated at all */
> > > > > - if ((old | new) & PTE_CONT)
> > > > > +#ifndef CONFIG_RODATA_FULL_USE_PTE_CONT
> > > > > + if ((old | new) & PTE_CONT)
> > > > > return false;
> > > > > +#endif
> > > > >
> > > > > /* Transitioning from Non-Global to Global is unsafe */
> > > > > if (old & ~new & PTE_NG)
> > > > > @@ -206,7 +212,7 @@ static void alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr,
> > > > >
> > > > > /* use a contiguous mapping if the range is suitably aligned */
> > > > > if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) &&
> > > > > - (flags & NO_CONT_MAPPINGS) == 0)
> > > > > + (IS_ENABLED(CONFIG_RODATA_FULL_USE_PTE_CONT) || (flags & NO_CONT_MAPPINGS) == 0))
> > > > > __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
> > > > >
> > > > > init_pte(pmdp, addr, next, phys, __prot);
> > > > > diff --git a/arch/arm64/mm/pageattr.c b/arch/arm64/mm/pageattr.c
> > > > > index a3bacd7..88a87eb 100644
> > > > > --- a/arch/arm64/mm/pageattr.c
> > > > > +++ b/arch/arm64/mm/pageattr.c
> > > > > @@ -99,6 +99,15 @@ static int change_memory_common(unsigned long addr, int numpages,
> > > > > if (rodata_full && (pgprot_val(set_mask) == PTE_RDONLY ||
> > > > > pgprot_val(clear_mask) == PTE_RDONLY)) {
> > > > > for (i = 0; i < area->nr_pages; i++) {
> > > > > +#ifdef CONFIG_RODATA_FULL_USE_PTE_CONT
> > > > > + unsigned long cont_pte_low_bound;
> > > > > + unsigned long addr;
> > > > > +
> > > > > + addr = (u64)page_address(area->pages[i]);
> > > > > + cont_pte_low_bound = addr & CONT_PTE_MASK;
> > > > > + __change_memory_common(cont_pte_low_bound,
> > > > > + (~CONT_PTE_MASK + 1), __pgprot(0) , __pgprot(PTE_CONT));
> > > > > +#endif
> > > > > __change_memory_common((u64)page_address(area->pages[i]),
> > > > > PAGE_SIZE, set_mask, clear_mask);
> > > > > }
> > > > > --
> > > > > 1.9.1
> > > > >