2018-03-21 21:05:01

by Philip Elcan

[permalink] [raw]
Subject: [PATCH] arm64: tlbflush: avoid writing RES0 bits

Bits [47:44] of the TLBI register operand are RES0 for instructions that
require a VA, per the ARM ARM spec, so TLBI operations should avoid writing
non-zero values to these bits.

Signed-off-by: Philip Elcan <[email protected]>
---
arch/arm64/include/asm/tlbflush.h | 16 ++++++++++------
1 file changed, 10 insertions(+), 6 deletions(-)

diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
index 9e82dd7..dbd22a9 100644
--- a/arch/arm64/include/asm/tlbflush.h
+++ b/arch/arm64/include/asm/tlbflush.h
@@ -60,6 +60,9 @@
__tlbi(op, (arg) | USER_ASID_FLAG); \
} while (0)

+/* This macro masks out RES0 bits in the TLBI operand */
+#define __TLBI_VADDR(addr) (addr & ~GENMASK_ULL(47, 44))
+
/*
* TLB Management
* ==============
@@ -128,7 +131,8 @@ static inline void flush_tlb_mm(struct mm_struct *mm)
static inline void flush_tlb_page(struct vm_area_struct *vma,
unsigned long uaddr)
{
- unsigned long addr = uaddr >> 12 | (ASID(vma->vm_mm) << 48);
+ unsigned long addr = __TLBI_VADDR(uaddr >> 12) |
+ (ASID(vma->vm_mm) << 48);

dsb(ishst);
__tlbi(vale1is, addr);
@@ -154,8 +158,8 @@ static inline void __flush_tlb_range(struct vm_area_struct *vma,
return;
}

- start = asid | (start >> 12);
- end = asid | (end >> 12);
+ start = asid | __TLBI_VADDR(start >> 12);
+ end = asid | __TLBI_VADDR(end >> 12);

dsb(ishst);
for (addr = start; addr < end; addr += 1 << (PAGE_SHIFT - 12)) {
@@ -185,8 +189,8 @@ static inline void flush_tlb_kernel_range(unsigned long start, unsigned long end
return;
}

- start >>= 12;
- end >>= 12;
+ start = __TLBI_VADDR(start >> 12);
+ end = __TLBI_VADDR(end >> 12);

dsb(ishst);
for (addr = start; addr < end; addr += 1 << (PAGE_SHIFT - 12))
@@ -202,7 +206,7 @@ static inline void flush_tlb_kernel_range(unsigned long start, unsigned long end
static inline void __flush_tlb_pgtable(struct mm_struct *mm,
unsigned long uaddr)
{
- unsigned long addr = uaddr >> 12 | (ASID(mm) << 48);
+ unsigned long addr = __TLBI_VADDR(uaddr >> 12) | (ASID(mm) << 48);

__tlbi(vae1is, addr);
__tlbi_user(vae1is, addr);
--
Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.



2018-03-22 18:32:12

by Robin Murphy

[permalink] [raw]
Subject: Re: [PATCH] arm64: tlbflush: avoid writing RES0 bits

On 21/03/18 21:02, Philip Elcan wrote:
> Bits [47:44] of the TLBI register operand are RES0 for instructions that
> require a VA, per the ARM ARM spec, so TLBI operations should avoid writing
> non-zero values to these bits.

If we're going to start sanitising addresses to respect RES0 bits, then
we should probably do it properly to cope with the cases where bits
63:47 are also RES0, and others - I guess we never actually try to do
something like VAE1IS with a kernel VA, but AFAICS that would still be
busted even after this patch.

Robin.

> Signed-off-by: Philip Elcan <[email protected]>
> ---
> arch/arm64/include/asm/tlbflush.h | 16 ++++++++++------
> 1 file changed, 10 insertions(+), 6 deletions(-)
>
> diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
> index 9e82dd7..dbd22a9 100644
> --- a/arch/arm64/include/asm/tlbflush.h
> +++ b/arch/arm64/include/asm/tlbflush.h
> @@ -60,6 +60,9 @@
> __tlbi(op, (arg) | USER_ASID_FLAG); \
> } while (0)
>
> +/* This macro masks out RES0 bits in the TLBI operand */
> +#define __TLBI_VADDR(addr) (addr & ~GENMASK_ULL(47, 44))
> +
> /*
> * TLB Management
> * ==============
> @@ -128,7 +131,8 @@ static inline void flush_tlb_mm(struct mm_struct *mm)
> static inline void flush_tlb_page(struct vm_area_struct *vma,
> unsigned long uaddr)
> {
> - unsigned long addr = uaddr >> 12 | (ASID(vma->vm_mm) << 48);
> + unsigned long addr = __TLBI_VADDR(uaddr >> 12) |
> + (ASID(vma->vm_mm) << 48);
>
> dsb(ishst);
> __tlbi(vale1is, addr);
> @@ -154,8 +158,8 @@ static inline void __flush_tlb_range(struct vm_area_struct *vma,
> return;
> }
>
> - start = asid | (start >> 12);
> - end = asid | (end >> 12);
> + start = asid | __TLBI_VADDR(start >> 12);
> + end = asid | __TLBI_VADDR(end >> 12);
>
> dsb(ishst);
> for (addr = start; addr < end; addr += 1 << (PAGE_SHIFT - 12)) {
> @@ -185,8 +189,8 @@ static inline void flush_tlb_kernel_range(unsigned long start, unsigned long end
> return;
> }
>
> - start >>= 12;
> - end >>= 12;
> + start = __TLBI_VADDR(start >> 12);
> + end = __TLBI_VADDR(end >> 12);
>
> dsb(ishst);
> for (addr = start; addr < end; addr += 1 << (PAGE_SHIFT - 12))
> @@ -202,7 +206,7 @@ static inline void flush_tlb_kernel_range(unsigned long start, unsigned long end
> static inline void __flush_tlb_pgtable(struct mm_struct *mm,
> unsigned long uaddr)
> {
> - unsigned long addr = uaddr >> 12 | (ASID(mm) << 48);
> + unsigned long addr = __TLBI_VADDR(uaddr >> 12) | (ASID(mm) << 48);
>
> __tlbi(vae1is, addr);
> __tlbi_user(vae1is, addr);
>

2018-03-26 10:05:06

by Mark Rutland

[permalink] [raw]
Subject: Re: [PATCH] arm64: tlbflush: avoid writing RES0 bits

On Wed, Mar 21, 2018 at 05:02:52PM -0400, Philip Elcan wrote:
> Bits [47:44] of the TLBI register operand are RES0 for instructions that
> require a VA, per the ARM ARM spec, so TLBI operations should avoid writing
> non-zero values to these bits.
>
> Signed-off-by: Philip Elcan <[email protected]>
> ---
> arch/arm64/include/asm/tlbflush.h | 16 ++++++++++------
> 1 file changed, 10 insertions(+), 6 deletions(-)
>
> diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
> index 9e82dd7..dbd22a9 100644
> --- a/arch/arm64/include/asm/tlbflush.h
> +++ b/arch/arm64/include/asm/tlbflush.h
> @@ -60,6 +60,9 @@
> __tlbi(op, (arg) | USER_ASID_FLAG); \
> } while (0)
>
> +/* This macro masks out RES0 bits in the TLBI operand */
> +#define __TLBI_VADDR(addr) (addr & ~GENMASK_ULL(47, 44))

If we're going to mask the address bits, it would be simpler to keep the
valid bits than to clear the invalid bits. i.e.

#define __TLBI_VADDR(addr) (addr & GENMASK_ULL(43, 0))

Maybe we want a helper that does all of the addr / asid shifting and
masking, so we do that in one place rather than spreading it across all
helpers, e.g.

#define __tlbi_addr(addr, asid) \
({ \
unsigned long __ta = (addr) >> 12; \
__ta &= GENMASK_ULL(43, 0); \
__ta |= (asid) << 48; \
__ta; \
})

Thanks,
Mark.

2018-03-27 00:53:07

by Philip Elcan

[permalink] [raw]
Subject: Re: [PATCH] arm64: tlbflush: avoid writing RES0 bits

On 3/26/2018 6:02 AM, Mark Rutland wrote:
> On Wed, Mar 21, 2018 at 05:02:52PM -0400, Philip Elcan wrote:
>> Bits [47:44] of the TLBI register operand are RES0 for instructions that
>> require a VA, per the ARM ARM spec, so TLBI operations should avoid writing
>> non-zero values to these bits.
>>
>> Signed-off-by: Philip Elcan <[email protected]>
>> ---
>> arch/arm64/include/asm/tlbflush.h | 16 ++++++++++------
>> 1 file changed, 10 insertions(+), 6 deletions(-)
>>
>> diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
>> index 9e82dd7..dbd22a9 100644
>> --- a/arch/arm64/include/asm/tlbflush.h
>> +++ b/arch/arm64/include/asm/tlbflush.h
>> @@ -60,6 +60,9 @@
>> __tlbi(op, (arg) | USER_ASID_FLAG); \
>> } while (0)
>>
>> +/* This macro masks out RES0 bits in the TLBI operand */
>> +#define __TLBI_VADDR(addr) (addr & ~GENMASK_ULL(47, 44))
>
> If we're going to mask the address bits, it would be simpler to keep the
> valid bits than to clear the invalid bits. i.e.
>
> #define __TLBI_VADDR(addr) (addr & GENMASK_ULL(43, 0))
>
> Maybe we want a helper that does all of the addr / asid shifting and
> masking, so we do that in one place rather than spreading it across all
> helpers, e.g.
>
> #define __tlbi_addr(addr, asid) \
> ({ \
> unsigned long __ta = (addr) >> 12; \
> __ta &= GENMASK_ULL(43, 0); \
> __ta |= (asid) << 48; \
> __ta; \
> })
>
> Thanks,
> Mark.
>

That makes sense and addresses Robin's comment as well. I'll send out a
v2 with your suggestion.

Thanks,
Philip

--
Qualcomm Datacenter Technologies as an affiliate of Qualcomm
Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.