Re: [PATCH] arm64: kprobes: check the return value of set_memory_rox()

[Ryan Roberts <ryan.roberts@arm.com>]

On 04/11/2025 13:14, Ryan Roberts wrote:
> On 03/11/2025 19:45, Yang Shi wrote:
>> Since commit a166563e7ec3 ("arm64: mm: support large block mapping when
>> rodata=full"), __change_memory_common has more chance to fail due to
>> memory allocation fialure when splitting page table. So check the return
>> value of set_memory_rox(), then bail out if it fails otherwise we may have
>> RW memory mapping for kprobes insn page.
>>
>> Fixes: 195a1b7d8388 ("arm64: kprobes: call set_memory_rox() for kprobe page")
>> Signed-off-by: Yang Shi <yang@os.amperecomputing.com>
> 
> This patch looks correct so:
> 
> Reviewed-by: Ryan Roberts <ryan.roberts@arm.com>
> 
> but, I think I see an separate issue below...
> 
>> ---
>> I actually epxected 195a1b7d8388 ("arm64: kprobes: call set_memory_rox()
>> for kprobe page") can be merged in 6.17-rcX, so I just restored it to
>> before commit 10d5e97c1bf8 ("arm64: use PAGE_KERNEL_ROX directly in
>> alloc_insn_page"), however it turned out to be merged in 6.18-rc1 and it
>> is after commit a166563e7ec3 ("arm64: mm: support large block mapping when
>> rodata=full"). So I made the fix tag point to it.
>> And I don't think we need to backport this patch to pre-6.18.
>>
>>  arch/arm64/kernel/probes/kprobes.c | 5 ++++-
>>  1 file changed, 4 insertions(+), 1 deletion(-)
>>
>> diff --git a/arch/arm64/kernel/probes/kprobes.c b/arch/arm64/kernel/probes/kprobes.c
>> index 8ab6104a4883..43a0361a8bf0 100644
>> --- a/arch/arm64/kernel/probes/kprobes.c
>> +++ b/arch/arm64/kernel/probes/kprobes.c
>> @@ -49,7 +49,10 @@ void *alloc_insn_page(void)
>>  	addr = execmem_alloc(EXECMEM_KPROBES, PAGE_SIZE);
>>  	if (!addr)
>>  		return NULL;
>> -	set_memory_rox((unsigned long)addr, 1);
>> +	if (set_memory_rox((unsigned long)addr, 1)) {
> 
> How does x get cleared when freeing this memory? arm64's set_memory_x() sets
> PTE_MAYBE_GP and clears PTE_PXN. The only function that will revert that is
> set_memory_nx(). But that only gets called from module_enable_data_nx() (which I
> don't think is applicable here) and execmem_force_rw() - but only if
> CONFIG_ARCH_HAS_EXECMEM_ROX is enabled, which I don't think it is for arm64?
> 
> So I think once we flip a page executable, it will be executable forever?
> 
> Do we need to modify set_direct_map_default_noflush() to make the memory nx?
> Then vm_reset_perms() will fix it up at vfree time?


Dev just pointed this out to me. Panic over!

static int change_memory_common(unsigned long addr, int numpages,
				pgprot_t set_mask, pgprot_t clear_mask)
{
	...

	/*
	 * If we are manipulating read-only permissions, apply the same
	 * change to the linear mapping of the pages that back this VM area.
	 */
	if (rodata_full && (pgprot_val(set_mask) == PTE_RDONLY ||
			    pgprot_val(clear_mask) == PTE_RDONLY)) {
		for (i = 0; i < area->nr_pages; i++) {
			__change_memory_common(...);
		}
	}

	...
}


> 
> Thanks,
> Ryan
> 
>> +		execmem_free(addr);
>> +		return NULL;
>> +	}
>>  	return addr;
>>  }
>>  
>