[PATCH] arm64: fix strlen() with CONFIG_KASAN_HW_TAGS

[Robin Murphy]

On 2021-07-12 10:00, Mark Rutland wrote:
> When the kernel is built with CONFIG_KASAN_HW_TAGS and the CPU supports
> MTE, memory accesses are checked at 16-byte granularity, and
> out-of-bounds accesses can result in tag check faults. Our current
> implementation of strlen() makes unaligned 16-byte accesses (within a
> naturally aligned 4096-byte window), and can trigger tag check faults.
> 
> This can be seen at boot time, e.g.
> 
> | BUG: KASAN: invalid-access in __pi_strlen+0x14/0x150
> | Read at addr f4ff0000c0028300 by task swapper/0/0
> | Pointer tag: [f4], memory tag: [fe]
> |
> | CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.13.0-09550-g03c2813535a2-dirty #20
> | Hardware name: linux,dummy-virt (DT)
> | Call trace:
> |  dump_backtrace+0x0/0x1b0
> |  show_stack+0x1c/0x30
> |  dump_stack_lvl+0x68/0x84
> |  print_address_description+0x7c/0x2b4
> |  kasan_report+0x138/0x38c
> |  __do_kernel_fault+0x190/0x1c4
> |  do_tag_check_fault+0x78/0x90
> |  do_mem_abort+0x44/0xb4
> |  el1_abort+0x40/0x60
> |  el1h_64_sync_handler+0xb0/0xd0
> |  el1h_64_sync+0x78/0x7c
> |  __pi_strlen+0x14/0x150
> |  __register_sysctl_table+0x7c4/0x890
> |  register_leaf_sysctl_tables+0x1a4/0x210
> |  register_leaf_sysctl_tables+0xc8/0x210
> |  __register_sysctl_paths+0x22c/0x290
> |  register_sysctl_table+0x2c/0x40
> |  sysctl_init+0x20/0x30
> |  proc_sys_init+0x3c/0x48
> |  proc_root_init+0x80/0x9c
> |  start_kernel+0x640/0x69c
> |  __primary_switched+0xc0/0xc8
> 
> To fix this, we can reduce the (strlen-internal) MIN_PAGE_SIZE to 16
> bytes when CONFIG_KASAN_HW_TAGS is selected. This will cause strlen() to
> align the base pointer downwards to a 16-byte boundary, and to discard
> the additional prefix bytes without counting them. All subsequent
> accesses will be 16-byte aligned 16-byte LDPs. While the comments say
> the body of the loop will access 32 bytes, this is performed as two
> 16-byte acceses, with the second made only if the first did not
> encounter a NUL byte, so the body of the loop will not over-read across
> a 16-byte boundary.
> 
> No other string routines are affected. The other str*() routines will
> not make any access which straddles a 16-byte boundary, and the mem*()
> routines will only make acceses which straddle a 16-byte boundary when
> which is entirely within the bounds of the relevant base and size
> arguments.

Ah, sorry for the oversight - I hadn't appreciated that in-kernel MTE 
was a thing now (and FWIW apparently the last time I refreshed Sam's 
original patches it still wasn't).

We can't use the "official" strlen-mte from Arm Optimized Routines since 
that uses ASIMD, but since I assume KASAN is still expected to have some 
performance penalty in general - even if hardware-assisted - this seems 
like a fair compromise. As for the general 16-byte assumptions, if there 
were ever an MTE4 or whatever with a smaller granule which doesn't break 
the rest of the world anyway, I think we could worry about that then.

Reviewed-by: Robin Murphy <robin.murphy at arm.com>

Thanks,
Robin.

> Fixes: 325a1de81287a3d4 (“arm64: Import updated version of Cortex Strings' strlen”)
> Signed-off-by: Mark Rutland <mark.rutland at arm.com>
> Cc: Alexander Potapenko <glider at google.com
> Cc: Andrey Konovalov <andreyknvl at gmail.com>
> Cc: Andrey Ryabinin <ryabinin.a.a at gmail.com>
> Cc: Catalin Marinas <catalin.marinas at arm.com>
> Cc: Dmitry Vyukov <dvyukov at google.com>
> Cc: Marco Elver <elver at google.com>
> Cc: Robin Murphy <robin.murphy at arm.com>
> Cc: Will Deacon <will at kernel.org>
> ---
>   arch/arm64/lib/strlen.S | 10 ++++++++++
>   1 file changed, 10 insertions(+)
> 
> Note: to build-test this, you'll also need to apply a fix from Marco [1], which
> is on its way to mainline via akpm. I've locally aplied that, and tested this
> patch with a recent build of QEMU with MTE enabled.
> 
> Mark.
> 
> diff --git a/arch/arm64/lib/strlen.S b/arch/arm64/lib/strlen.S
> index 35fbdb7d6e1a..1648790e91b3 100644
> --- a/arch/arm64/lib/strlen.S
> +++ b/arch/arm64/lib/strlen.S
> @@ -8,6 +8,7 @@
>   
>   #include <linux/linkage.h>
>   #include <asm/assembler.h>
> +#include <asm/mte-def.h>
>   
>   /* Assumptions:
>    *
> @@ -42,7 +43,16 @@
>   #define REP8_7f 0x7f7f7f7f7f7f7f7f
>   #define REP8_80 0x8080808080808080
>   
> +/*
> + * When KASAN_HW_TAGS is in use, memory is checked at MTE_GRANULE_SIZE
> + * (16-byte) granularity, and we must ensure that no access straddles this
> + * alignment boundary.
> + */
> +#ifdef CONFIG_KASAN_HW_TAGS
> +#define MIN_PAGE_SIZE MTE_GRANULE_SIZE
> +#else
>   #define MIN_PAGE_SIZE 4096
> +#endif
>   
>   	/* Since strings are short on average, we check the first 16 bytes
>   	   of the string for a NUL character.  In order to do an unaligned ldp
>