[PATCH] arm64: mm: explicitly use kernel pte for ioremap_prot()

[Jinjiang Tu]

在 2026/1/24 2:36, Catalin Marinas 写道:
> On Fri, Jan 23, 2026 at 11:02:38AM +0800, Jinjiang Tu wrote:
>> Here is a syzkaller error log:
>>    [0000000020ffc000] pgd=080000010598d403, p4d=080000010598d403, pud=0800000125ddb403,
>>                       pmd=080000007833c403, pte=01608000007fcfcf
>>    Unable to handle kernel read from unreadable memory at virtual address ffff80008ea89000
>>    KASAN: probably user-memory-access in range [0x0000000475448000-0x0000000475448007]
>>    Mem abort info:
>>      ESR = 0x000000009600000f
>>      EC = 0x25: DABT (current EL), IL = 32 bits
>>      SET = 0, FnV = 0
>>      EA = 0, S1PTW = 0
>>      FSC = 0x0f: level 3 permission fault
>>    Data abort info:
>>      ISV = 0, ISS = 0x0000000f, ISS2 = 0x00000000
>>      CM = 0, WnR = 0, TnD = 0, TagAccess = 0
>>      GCS = 0, Overlay = 0, DirtyBit = 0, Xs = 0
>>    swapper pgtable: 4k pages, 48-bit VAs, pgdp=00000001244aa000
>>    [ffff80008ea89000] pgd=100000013ffff403, p4d=100000013ffff403, pud=100000013fffe403,
>> 		     pmd=100000010a453403, pte=01608000007fcfcf
>>    Internal error: Oops: 000000009600000f [#1] SMP
>>    Modules linked in: team
>>    CPU: 1 PID: 10840 Comm: syz.9.83 Kdump: loaded Tainted: G
>>    Hardware name: QEMU KVM Virtual Machine, BIOS 0.0.0 02/06/2015
>>    pstate: 20400005 (nzCv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
>>    pc : __memcpy_fromio+0x80/0xf8
>>    lr : generic_access_phys+0x20c/0x2b8
>>    sp : ffff8000a0507960
>>    x29: ffff8000a0507960 x28: 1ffff000140a0f44 x27: ffff00003833cfe0
>>    x26: 0000000000000000 x25: 0000000000001000 x24: 0010000000000001
>>    x23: ffff80008ea89000 x22: ffff00004ea63000 x21: 0000000000001000
>>    x20: ffff80008ea89000 x19: ffff00004ea62000 x18: 0000000000000000
>>    x17: 0000000000000000 x16: 0000000000000000 x15: ffff8000806f1e3c
>>    x14: ffff8000806f1d44 x13: 0000000041b58ab3 x12: ffff7000140a0f23
>>    x11: 1ffff000140a0f22 x10: ffff7000140a0f22 x9 : ffff800080579d24
>>    x8 : 0000000000000004 x7 : 0000000000000003 x6 : 0000000000000001
>>    x5 : ffff8000a0507910 x4 : ffff7000140a0f22 x3 : dfff800000000000
>>    x2 : 0000000000001000 x1 : ffff80008ea89000 x0 : ffff00004ea62000
>>    Call trace:
>>      __memcpy_fromio+0x80/0xf8
>>      generic_access_phys+0x20c/0x2b8
>>      __access_remote_vm+0x46c/0x5b8
>>      access_remote_vm+0x18/0x30
>>      environ_read+0x238/0x3e8
>>      vfs_read+0xe4/0x2b0
>>      ksys_read+0xcc/0x178
>>      __arm64_sys_read+0x4c/0x68
>>      invoke_syscall+0x68/0x1a0
>>      el0_svc_common.constprop.0+0x11c/0x150
>>      do_el0_svc+0x38/0x50
>>      el0_svc+0x50/0x258
>>      el0t_64_sync_handler+0xc0/0xc8
>>      el0t_64_sync+0x1a4/0x1a8
>>    Code: 91002339 aa1403f7 8b190276 d503201f (f94002f8)
>>
>> The local syzkaller first maps I/O address from /dev/mem to userspace,
>> overiding the stack vma with MAP_FIXED flag. As a result, when reading
>> /proc/$pid/environ, generic_access_phys() is called to access the region,
>> which triggers a PAN permission-check fault and causes a kernel access
>> fault.
>>
>> The root cause is that generic_access_phys() passes a user pte to
>> ioremap_prot(), the user pte sets PTE_USER and PTE_NG bits.  Consequently,
>> any subsequent kernel-mode access to the remapped address raises a fault.
>>
>> To fix it, similar to ioremap_prot() in x86, use _PAGE_KERNEL as template,
>> and update PTE_WRITE and PTE_ATTRINDX according to the pgprot argument.
>>
>> Fixes: 893dea9ccd08 ("arm64: Add HAVE_IOREMAP_PROT support")
>> Signed-off-by: Zeng Heng <zengheng4 at huawei.com>
>> Signed-off-by: Jinjiang Tu <tujinjiang at huawei.com>
>> ---
>>   arch/arm64/mm/ioremap.c | 6 ++++++
>>   1 file changed, 6 insertions(+)
>>
>> diff --git a/arch/arm64/mm/ioremap.c b/arch/arm64/mm/ioremap.c
>> index 10e246f11271..78a63a14465f 100644
>> --- a/arch/arm64/mm/ioremap.c
>> +++ b/arch/arm64/mm/ioremap.c
>> @@ -18,6 +18,7 @@ void __iomem *ioremap_prot(phys_addr_t phys_addr, size_t size,
>>   			   pgprot_t pgprot)
>>   {
>>   	unsigned long last_addr = phys_addr + size - 1;
>> +	unsigned long new_pgprot_val;
>>   
>>   	/* Don't allow outside PHYS_MASK */
>>   	if (last_addr & ~PHYS_MASK)
>> @@ -27,6 +28,11 @@ void __iomem *ioremap_prot(phys_addr_t phys_addr, size_t size,
>>   	if (WARN_ON(pfn_is_map_memory(__phys_to_pfn(phys_addr))))
>>   		return NULL;
>>   
>> +	new_pgprot_val = _PAGE_KERNEL & ~(PTE_WRITE | PTE_ATTRINDX_MASK);
>> +	new_pgprot_val |= (pgprot_val(pgprot) & PTE_WRITE)
>> +				| (pgprot_val(pgprot) & PTE_ATTRINDX_MASK);
>> +	pgprot = __pgprot(new_pgprot_val);
>> +
>>   	/*
>>   	 * If a hook is registered (e.g. for confidential computing
>>   	 * purposes), call that now and barf if it fails.
> generic_access_phys() is really weird - it takes the user PTE attributes
> and passes them to the kernel ioremap_prot(), hoping for the best. My
> immediate thought was to fix this in the core code. However, we also
> need to preserve the original memory type from vma->vm_page_prot and I
> don't think we have any generic macros to just preserve the attributes
> but with kernel permissions (pte_mkkernel?).

If we fix it in generic_access_phys(), we have to define a macro for all archs
that define CONFIG_HAVE_IOREMAP_PROT.

x86, loogarch, mips, sh archs only care cacheable attribute of the user prot, so
don't involve this issue. But, arc, arm64, powerpc simply copy the user pte prot
to the kernel page table.

>