[BUG] arm64: an infinite loop in generic_perform_write()

[Chen Huang]

在 2021/6/23 21:22, Mark Rutland 写道:
> On Wed, Jun 23, 2021 at 10:39:31AM +0800, Chen Huang wrote:
>> When we access a device memory in userspace, then perform an unaligned write to a file.
>> For example, we register a uio device and mmap the device, then perform an write to a
>> file, like that:
>>
>> 	device_addr = mmap(device_fd);
>> 	write(file_fd, device_addr + unaligned_num, size);
> 
> What exactly is this device, and why do you want the kernel to do a
> direct memcpy from MMIO? Why can't you copy that in userspace (where you
> have knowledge of the device), then pass the copy to a syscall?
>
I'm sorry for not describing the problem well. It's an uio device:

static struct device_driver uio_dummy_driver = {
    .name = "uio_with_name",
    .bus = &platform_bus_type,
    .probe = drv_uio_with_name_probe,
    .remove = drv_uio_with_name_remove,
};

static int drv_uio_with_name_probe(struct device *dev)
{
    uio_with_name_info.mem[0].addr = 0xa0000000;
    uio_with_name_info.mem[0].memtype = UIO_MEM_PHYS;
    uio_with_name_info.mem[0].size = 0x1000;

    if (__uio_register_device(THIS_MODULE, dev, &uio_with_name_info)) {
        printk("__uio_register_device failed\n");
        return -ENODEV;
    }
    printk("UIO init end.\n");
    return 0;
}

In userspace, I perform such operation:

 	fd = open("/tmp/test", O_RDWR | O_SYNC);
        access_address = (char *)mmap(NULL, uio_size, PROT_READ, MAP_SHARED, uio_fd, 0);
        ret = write(fd, access_address + 2, sizeof(long));

> Ignoring the lockup below, this isn't going to work in general, since
> uaccess routines do not guarantee alignment, single-copy, access sizes,
> monotonically increasing addresses, etc. Any one of those can cause a
> device to raise an external abort which may or may not be synchronous.
> 
> It does not make sense to tell the kernel to access this, since the
> kernel cannot know how to access it safely, and we can;t do that without
> knowledge of the device that we do not have.
> 
> Thanks,
> Mark.
> 
>> 	
>> We found that the infinite loop happened in generic_perform_write function:
>>
>> again:
>> 	copied = copy_page_from_iter_atomic(); //copied = 0
>> 	status = ops->write_end(); //status = 0
>> 	if (status == 0)
>> 		goto again;
>>
>> In copy_page_from_iter_atomic, the copyin() function finally call
>> __arch_copy_from_user which create an exception table entry for 'insn'.
>> Then when kernel handles the alignment_fault, it will not panic. As the
>> arm64 memory model spec said, when the address is not a multiple of the
>> element size, the access is unaligned. Unaligned accesses are allowed to
>> addresses marked as Normal, but not to Device regions. An unaligned access
>> to a Device region will trigger an exception (alignment fault).
>> 	
>> do_alignment_fault
>>     do_bad_area
>> 	__do_kernel_fault
>>            fixup_exception
>>
>> But that fixup cann't handle the unaligned copy, so the
>> copy_page_from_iter_atomic returns 0 and traps in loop.
>>
>> Reported-by: Chen Huang <chenhuang5 at huawei.com>
> .
>