[PATCH v3 2/2] arm64: mm: install KPTI nG mappings with MMU enabled
Mark Rutland
mark.rutland at arm.com
Fri May 20 06:00:06 PDT 2022
Hi Ard,
On Thu, Apr 21, 2022 at 04:03:39PM +0200, Ard Biesheuvel wrote:
> In cases where we unmap the kernel while running in user space, we rely
> on ASIDs to distinguish the minimal trampoline from the full kernel
> mapping, and this means we must use non-global attributes for those
> mappings, to ensure they are scoped by ASID and will not hit in the TLB
> inadvertently.
>
> We only do this when needed, as this is generally more costly in terms
> of TLB pressure, and so we boot without these non-global attributes, and
> apply them to all existing kernel mappings once all CPUs are up and we
> know whether or not the non-global attributes are needed. At this point,
> we cannot simply unmap and remap the entire address space, so we have to
> update all existing block and page descriptors in place.
>
> Currently, we go through a lot of trouble to perform these updates with
> the MMU and caches off, to avoid violating break before make (BBM) rules
> imposed by the architecture. Since we make changes to page tables that
> are not covered by the ID map, we gain access to those descriptors by
> disabling translations altogether. This means that the stores to memory
> are issued with device attributes, and require extra care in terms of
> coherency, which is costly. We also rely on the ID map to access a
> shared flag, which requires the ID map to be executable and writable at
> the same time, which is another thing we'd prefer to avoid.
>
> So let's switch to an approach where we replace the kernel mapping with
> a minimal mapping of a few pages that can be used for the shared flag,
> as well as a minimal, ad-hoc fixmap that we can use to map each page
> table in turn as we traverse the hierarchy. This requires one PTE per
> level, and an associated page worth of VA space in the temporary
> mapping.
I really like the high-level idea!
Overall this looks good. I'm not fond of the way we synchronize between
primary/secondary CPUs, and would prefer if we could split that differently --
I've given more detail on that below (and a branch with a proprosal).
> Signed-off-by: Ard Biesheuvel <ardb at kernel.org>
> ---
> arch/arm64/include/asm/mmu.h | 4 +
> arch/arm64/kernel/cpufeature.c | 65 +++++++++++-
> arch/arm64/mm/mmu.c | 8 +-
> arch/arm64/mm/proc.S | 107 ++++++++++++--------
> 4 files changed, 134 insertions(+), 50 deletions(-)
>
> diff --git a/arch/arm64/include/asm/mmu.h b/arch/arm64/include/asm/mmu.h
> index 48f8466a4be9..b896f0ac4985 100644
> --- a/arch/arm64/include/asm/mmu.h
> +++ b/arch/arm64/include/asm/mmu.h
> @@ -68,6 +68,10 @@ extern void init_mem_pgprot(void);
> extern void create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
> unsigned long virt, phys_addr_t size,
> pgprot_t prot, bool page_mappings_only);
> +extern void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
> + unsigned long virt, phys_addr_t size,
> + pgprot_t prot,
> + phys_addr_t (*pgtable_alloc)(int), int flags);
> extern void *fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot);
> extern void mark_linear_text_alias_ro(void);
> extern bool kaslr_requires_kpti(void);
> diff --git a/arch/arm64/kernel/cpufeature.c b/arch/arm64/kernel/cpufeature.c
> index d72c4b4d389c..f0688e812e19 100644
> --- a/arch/arm64/kernel/cpufeature.c
> +++ b/arch/arm64/kernel/cpufeature.c
> @@ -1596,14 +1596,31 @@ static bool unmap_kernel_at_el0(const struct arm64_cpu_capabilities *entry,
> }
>
> #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
> +#define KPTI_NG_TEMP_VA (-(1UL << PMD_SHIFT))
> +
> +static phys_addr_t kpti_ng_temp_alloc;
> +
> +static phys_addr_t kpti_ng_pgd_alloc(int shift)
> +{
> + kpti_ng_temp_alloc -= PAGE_SIZE;
> + return kpti_ng_temp_alloc;
> +}
> +
> static void __nocfi
> kpti_install_ng_mappings(const struct arm64_cpu_capabilities *__unused)
> {
> - typedef void (kpti_remap_fn)(int, int, phys_addr_t);
> + static atomic_t flag = ATOMIC_INIT(0);
> + static pgd_t *kpti_ng_temp_pgd;
> + static u64 alloc;
> +
> + typedef void (kpti_remap_fn)(int, int, phys_addr_t, unsigned long);
> extern kpti_remap_fn idmap_kpti_install_ng_mappings;
> kpti_remap_fn *remap_fn;
>
> - int cpu = smp_processor_id();
> + int levels = CONFIG_PGTABLE_LEVELS;
> + int order = order_base_2(levels + 1);
> + int num_cpus = num_online_cpus();
> + int primary = 0;
>
> if (__this_cpu_read(this_cpu_vector) == vectors) {
> const char *v = arm64_get_bp_hardening_vector(EL1_VECTOR_KPTI);
> @@ -1619,14 +1636,54 @@ kpti_install_ng_mappings(const struct arm64_cpu_capabilities *__unused)
> if (arm64_use_ng_mappings)
> return;
>
> + // First CPU to arrive here gets the job
> + if (atomic_inc_return(&flag) == 1) {
I'm not too keen on the way this is split, because (aside from the flag), the
secondaries have no need to have anything mapped in TTBR1. If we leave the flag
in the idmap, and have the secondaries all use the reserved TTBR1 page, we can
consistently have the boot CPU do all the hard work (and we could avoid
broadcast maintenance/barriers during bulk of the rewrite, which could make it
faster).
I had a go at refactoring things that way (with this rebased atop), at:
https://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git/log/?h=arm64/kpti/rework
The only downside is that we have to keep the idmap RWX, but that was already
the case (and we could rework the idmap to have separate ROX/RW code/data pages
in future).
I can post that to the list if you don't have major objections to that
approach.
> + alloc = __get_free_pages(GFP_ATOMIC | __GFP_ZERO, order);
> + kpti_ng_temp_pgd = (pgd_t *)(alloc + levels * PAGE_SIZE);
> + kpti_ng_temp_alloc = __pa(kpti_ng_temp_pgd);
> + primary = 1;
> +
> + //
> + // Create a minimal page table hierarchy that permits us to
> + // store a shared variable that secondaries will poll, and to
> + // map the swapper page tables temporarily as we traverse them.
> + //
> + // The physical pages are laid out as follows:
> + //
> + // +---------+--------+-/-------+-/------ +-\\--------+
> + // : data : PTE[] : | PMD[] : | PUD[] : || PGD[] :
> + // +---------+--------+-\-------+-\------ +-//--------+
> + // ^ ^
> + // The first two pages are mapped consecutively into this
> + // hierarchy at a PMD_SHIFT aligned virtual address, so that we
> + // have a place to store the shared variable, and so that we
> + // can manipulate the PTE level entries while the mapping is
> + // active. The first two entries cover the data page and the
> + // PTE[] page itself, the remaining entries are free to be used
> + // as a ad-hoc fixmap.
> + //
> + __create_pgd_mapping(kpti_ng_temp_pgd, __pa(alloc),
> + KPTI_NG_TEMP_VA, 2 * PAGE_SIZE,
> + PAGE_KERNEL, kpti_ng_pgd_alloc, 0);
It took me a while to realise this was relying on the order in which
__create_pgd_mapping() calls kpti_ng_pgd_alloc(). I had a look at building the
tables explicitly to make this clearer, but that didn't seem all that clearer,
so this is fine as-is.
> +
> + // Increment flag again to signal other CPUs to proceed as well
> + atomic_inc_return_release(&flag);
> + } else {
> + // Wait for the primary CPU to set up the temporary page tables
> + while (atomic_read(&flag) <= num_cpus)
> + cpu_relax();
> + }
> remap_fn = (void *)__pa_symbol(function_nocfi(idmap_kpti_install_ng_mappings));
>
> cpu_install_idmap();
> - remap_fn(cpu, num_online_cpus(), __pa_symbol(swapper_pg_dir));
> + remap_fn(!primary, num_cpus - 1, __pa(kpti_ng_temp_pgd), KPTI_NG_TEMP_VA);
> cpu_uninstall_idmap();
>
> - if (!cpu)
> + // Last CPU to leave frees the pages
> + if (atomic_dec_return(&flag) == 1) {
> + free_pages(alloc, order);
> arm64_use_ng_mappings = true;
> + }
> }
> #else
> static void
> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
> index 626ec32873c6..1c7299dfaa84 100644
> --- a/arch/arm64/mm/mmu.c
> +++ b/arch/arm64/mm/mmu.c
> @@ -360,11 +360,9 @@ static void alloc_init_pud(pgd_t *pgdp, unsigned long addr, unsigned long end,
> mutex_unlock(&fixmap_lock);
> }
>
> -static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
> - unsigned long virt, phys_addr_t size,
> - pgprot_t prot,
> - phys_addr_t (*pgtable_alloc)(int),
> - int flags)
> +void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys, unsigned long virt,
> + phys_addr_t size, pgprot_t prot,
> + phys_addr_t (*pgtable_alloc)(int), int flags)
> {
> unsigned long addr, end, next;
> pgd_t *pgdp = pgd_offset_pgd(pgdir, virt);
> diff --git a/arch/arm64/mm/proc.S b/arch/arm64/mm/proc.S
> index 5619c00f8cd4..20d726207db5 100644
> --- a/arch/arm64/mm/proc.S
> +++ b/arch/arm64/mm/proc.S
> @@ -14,6 +14,7 @@
> #include <asm/asm-offsets.h>
> #include <asm/asm_pointer_auth.h>
> #include <asm/hwcap.h>
> +#include <asm/kernel-pgtable.h>
> #include <asm/pgtable-hwdef.h>
> #include <asm/cpufeature.h>
> #include <asm/alternative.h>
> @@ -200,20 +201,21 @@ SYM_FUNC_END(idmap_cpu_replace_ttbr1)
> .popsection
>
> #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
> +
> +#define KPTI_NG_PTE_FLAGS (PTE_ATTRINDX(MT_NORMAL) | SWAPPER_PTE_FLAGS)
> +
> .pushsection ".idmap.text", "awx"
>
> .macro kpti_mk_tbl_ng, type, num_entries
> add end_\type\()p, cur_\type\()p, #\num_entries * 8
> .Ldo_\type:
> - dc cvac, cur_\()\type\()p // Ensure any existing dirty
> - dmb sy // lines are written back before
> - ldr \type, [cur_\type\()p], #8 // loading the entry
> - tbz \type, #0, .Lnext_\type // Skip invalid and
> - tbnz \type, #11, .Lnext_\type // non-global entries
> - orr \type, \type, #PTE_NG // Same bit for blocks and pages
> - str \type, [cur_\type\()p, #-8] // Update the entry and ensure
> - dmb sy // that it is visible to all
> - dc civac, cur_\()\type\()p // CPUs.
> + ldr \type, [cur_\type\()p], #8 // Load the entry
> + .ifnc \type, pte
> + tbnz \type, #11, .Lnext_\type // Skip visited entries
> + .endif
> + and valid, \type, #1
> + orr \type, \type, valid, lsl #11 // nG |= valid
> + str \type, [cur_\type\()p, #-8] // Update the entry
IMO it was clearer to skip invalid entries than to extract the valid bit value
and ORR that back into the nG position. Was that change intended as an
optimization?
> .ifnc \type, pte
> tbnz \type, #1, .Lderef_\type
> .endif
As with the last patch, I'd prfeer is we had an explicit ADD for the increment
in the `.Lnext_\type`. Even with the CMOs removed, I reckon that's clearer.
> @@ -222,19 +224,42 @@ SYM_FUNC_END(idmap_cpu_replace_ttbr1)
> b.ne .Ldo_\type
> .endm
>
> + /*
> + * Dereference the current table entry and map it into the temporary
> + * fixmap slot associated with the current level. The ad-hoc fixmap
> + * is a set of PTEs that are located above the PTEs that cover the
> + * level 3 page table and the scratch page that precedes it.
> + */
> + .macro kpti_map_pgtbl, type, level
> + str xzr, [temp_pte, #8 * (\level + 2)] // break before make
> + dsb ishst
> + add pte, flag_ptr, #PAGE_SIZE * (\level + 2)
> + lsr pte, pte, #12
> + tlbi vaae1is, pte
> + dsb ish
> + isb
> +
> + phys_to_pte pte, cur_\type\()p
> + add cur_\type\()p, flag_ptr, #PAGE_SIZE * (\level + 2)
> + orr pte, pte, pte_flags
> + str pte, [temp_pte, #8 * (\level + 2)]
> + dsb ishst
> + .endm
If we only had the boot CPU map the fixmap TTBR, we could reduce the scope of
barriers and TLB maintenance here to be nSH, and removing the broadcast
maintenance could be a big win. We'd just need a final DSB ISHST prior to
waking secondaries to ensure they see all the updates.
Mybe that's miniscule compared to keeping the MMU on, though.
Thanks,
Mark.
> +
> /*
> - * void __kpti_install_ng_mappings(int cpu, int num_cpus, phys_addr_t swapper)
> + * void __kpti_install_ng_mappings(int cpu, int num_secondaries, phys_addr_t temp_pgd,
> + * unsigned long kpti_ng_temp_va)
> *
> * Called exactly once from stop_machine context by each CPU found during boot.
> */
> -__idmap_kpti_flag:
> - .long 1
> SYM_FUNC_START(idmap_kpti_install_ng_mappings)
> - cpu .req w0
> + cpu .req w0 // at entry
> + pte_flags .req x0
> num_cpus .req w1
> - swapper_pa .req x2
> - swapper_ttb .req x3
> - flag_ptr .req x4
> + temp_pgd_phys .req x2 // at entry
> + temp_pte .req x2
> + flag_ptr .req x3
> + swapper_ttb .req x4
> cur_pgdp .req x5
> end_pgdp .req x6
> pgd .req x7
> @@ -245,10 +270,15 @@ SYM_FUNC_START(idmap_kpti_install_ng_mappings)
> cur_ptep .req x14
> end_ptep .req x15
> pte .req x16
> + valid .req x17
>
> mrs swapper_ttb, ttbr1_el1
> - restore_ttbr1 swapper_ttb
> - adr flag_ptr, __idmap_kpti_flag
> +
> + /* Uninstall swapper before surgery begins */
> + __idmap_cpu_set_reserved_ttbr1 x8, x9
> + offset_ttbr1 temp_pgd_phys, x8
> + msr ttbr1_el1, temp_pgd_phys
> + isb
>
> cbnz cpu, __idmap_kpti_secondary
>
> @@ -259,31 +289,24 @@ SYM_FUNC_START(idmap_kpti_install_ng_mappings)
> eor w17, w17, num_cpus
> cbnz w17, 1b
>
> - /* We need to walk swapper, so turn off the MMU. */
> - pre_disable_mmu_workaround
> - mrs x17, sctlr_el1
> - bic x17, x17, #SCTLR_ELx_M
> - msr sctlr_el1, x17
> - isb
> + mov pte_flags, #KPTI_NG_PTE_FLAGS
> +
> + /* Advance temp_pte to the fixmap page */
> + add temp_pte, flag_ptr, #PAGE_SIZE
>
> /* Everybody is enjoying the idmap, so we can rewrite swapper. */
> /* PGD */
> - mov cur_pgdp, swapper_pa
> + adrp cur_pgdp, swapper_pg_dir
> + kpti_map_pgtbl pgd, 0
> kpti_mk_tbl_ng pgd, PTRS_PER_PGD
>
> - /* Publish the updated tables and nuke all the TLBs */
> - dsb sy
> - tlbi vmalle1is
> - dsb ish
> - isb
> -
> - /* We're done: fire up the MMU again */
> - mrs x17, sctlr_el1
> - orr x17, x17, #SCTLR_ELx_M
> - set_sctlr_el1 x17
> -
> /* Set the flag to zero to indicate that we're all done */
> str wzr, [flag_ptr]
> +
> + /* We're done: fire up swapper again */
> + __idmap_cpu_set_reserved_ttbr1 x8, x9
> + msr ttbr1_el1, swapper_ttb
> + isb
> ret
>
> .Lderef_pgd:
> @@ -291,6 +314,7 @@ SYM_FUNC_START(idmap_kpti_install_ng_mappings)
> .if CONFIG_PGTABLE_LEVELS > 3
> pud .req x10
> pte_to_phys cur_pudp, pgd
> + kpti_map_pgtbl pud, 1
> kpti_mk_tbl_ng pud, PTRS_PER_PUD
> b .Lnext_pgd
> .else /* CONFIG_PGTABLE_LEVELS <= 3 */
> @@ -303,6 +327,7 @@ SYM_FUNC_START(idmap_kpti_install_ng_mappings)
> .if CONFIG_PGTABLE_LEVELS > 2
> pmd .req x13
> pte_to_phys cur_pmdp, pud
> + kpti_map_pgtbl pmd, 2
> kpti_mk_tbl_ng pmd, PTRS_PER_PMD
> b .Lnext_pud
> .else /* CONFIG_PGTABLE_LEVELS <= 2 */
> @@ -313,12 +338,14 @@ SYM_FUNC_START(idmap_kpti_install_ng_mappings)
> .Lderef_pmd:
> /* PTE */
> pte_to_phys cur_ptep, pmd
> + kpti_map_pgtbl pte, 3
> kpti_mk_tbl_ng pte, PTRS_PER_PTE
> b .Lnext_pmd
>
> .unreq cpu
> + .unreq pte_flags
> .unreq num_cpus
> - .unreq swapper_pa
> + .unreq temp_pgd_phys
> .unreq cur_pgdp
> .unreq end_pgdp
> .unreq pgd
> @@ -331,12 +358,10 @@ SYM_FUNC_START(idmap_kpti_install_ng_mappings)
> .unreq cur_ptep
> .unreq end_ptep
> .unreq pte
> + .unreq valid
>
> /* Secondary CPUs end up here */
> __idmap_kpti_secondary:
> - /* Uninstall swapper before surgery begins */
> - __idmap_cpu_set_reserved_ttbr1 x16, x17
> -
> /* Increment the flag to let the boot CPU we're ready */
> 1: ldxr w16, [flag_ptr]
> add w16, w16, #1
> @@ -350,7 +375,7 @@ __idmap_kpti_secondary:
> cbnz w16, 1b
>
> /* All done, act like nothing happened */
> - offset_ttbr1 swapper_ttb, x16
> + __idmap_cpu_set_reserved_ttbr1 x8, x9
> msr ttbr1_el1, swapper_ttb
> isb
> ret
> --
> 2.30.2
>
More information about the linux-arm-kernel
mailing list