[PATCH 2/2] Staging: android: ashmem: Add support for 32bit ashmem calls in a 64bit kernel

Dan Carpenter dan.carpenter at oracle.com
Tue Dec 4 11:45:11 UTC 2012

I don't understand this, and I'm going to embarrass myself by
displaying my ignorance for all to see.  Why is this code so
different from all the other 32 bit compat code that we have in the

On Tue, Dec 04, 2012 at 10:44:14AM +0000, Serban Constantinescu wrote:
> -static int set_name(struct ashmem_area *asma, void __user *name)
> +static int set_name(struct ashmem_area *asma, userptr32_t name)

The user passes in a value which is a 32 pointer.  ashmem_ioctl()
accepts it as "unsigned long arg".  We pass it to set_name() which
truncates away the high zeros so now its a u32 (userptr32_t).  We
then cast it to (unsigned long) and then we cast it to a void

What's the point?  Why not just take"unsigned long arg" and cast it
to a pointer directly?

>  	if (unlikely(copy_from_user(asma->name + ASHMEM_NAME_PREFIX_LEN,
> -				    name, ASHMEM_NAME_LEN)))
> +				    (void *)(unsigned long)name, ASHMEM_NAME_LEN)))
>  		ret = -EFAULT;

This will introduce a Sparse complaint.  It should be:
				(void __user *)(unsigned long)name.

But actually we shouldn't need to do this casting.  Any casting
which we need to do should be done in one place instead of pushed
out to every function.

> +	switch (_IOC_NR(cmd)) {
> +		if (_IOC_SIZE(cmd) != sizeof(char[ASHMEM_NAME_LEN]))
> +			pr_err("ashmem: ASHMEM_SET_NAME transaction size differs\n");

Don't merge debug code into the kernel.  It just means people can
spam /var/log/messages.

dan carpenter

More information about the devel mailing list