Re: [PATCH] cpumask 5/10 rewrite cpumask.h - single bitmap based implementation

[William Lee Irwin III <wli@holomorphy.com>]

On Sat, Jun 05, 2004 at 01:04:44AM -0700, Paul Jackson wrote:
> That looks more readable.  Thanks.
> Do you see a sensible way to pass back an odd number of u32's?  If
> NR_CPUS is say 8, then the 32 bit user code might expect to only need
> one compat_ulong_t, not two.  If NR_CPUS is 48, then it should only need
> 3, not 4.   And so forth.
> As I noted in another reply, perhaps your bitmap_to_u32_array() code
> could be modified to handle this.
> And I agree with Andrew's suggestion, that cpumask provide the
> conversion, to and from kernel memory, separately from copying the
> result to user space.

So do it in the caller, e.g..

int copy_cpus_to_user32(const u32 __user *ubuf, cpumask_t cpus)
{
	int i, ret, len = ALIGN(NR_CPUS, 32)/(32/sizeof(u32));
	u32 *ary = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
	if (!ary)
		return -ENOMEM;
	cpus_to_u32_array(ary, cpus);
	ret = copy_to_user(ubuf, ary, len);
	kfree(ary);
	return ret;
}

int copy_cpus_from_user32(cpumask_t *cpus, const u32 __user *ubuf)
{
	int i, ret, len = ALIGN(NR_CPUS, 32)/(32/sizeof(u32));
	u32 *ary = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
	if (!ary)
		return -ENOMEM;
	if (!(ret = copy_from_user(ary, ubuf, len)))
		cpus_from_u32_array(cpus, ary);
	kfree(ary);
	return ret;
}

or some such nonsense, or whatever someone can be arsed to consider a
better idea. One should note such a reformatting, as a user ABI change
in a stable series, would be unfriendly to 64-bit userspace on BE boxen.
i.e. do this only for 32-bit target userspace on 64-bit kernels + boxen.


-- wli