copy_*_user

copy_*_user

[Xin Tong]

I am investigating copy_from_user  and copy_to_user in linux under
i386. These two function both take a pointer with virtual address and
a pointer with physical address.

copy_from_user calls __copy_from_user_ll and copy_to_user  calls
__copy_to_user_ll. It make sense to me that __copy_to_user_ll converts
the virtual address to physical address using the current process's
page table.


unsigned long __copy_to_user_ll(void __user *to, const void *from,
unsigned long n)
{
   ...
    retval = get_user_pages(current, current->mm,
                                        (unsigned long)to, 1, 1, 0, &pg, NULL);

                        if (retval == -ENOMEM && is_global_init(current)) {
                                up_read(&current->mm->mmap_sem);
                                congestion_wait(BLK_RW_ASYNC, HZ/50);
                                goto survive;
                        }

                        if (retval != 1) {
                                up_read(&current->mm->mmap_sem);
                                break;
                        }

       maddr = kmap_atomic(pg, KM_USER0);
       memcpy(maddr + offset, from, len);
     ...
}

But it seems to be that __copy_from_user_ll  is not converted the
address at all before attempting to copy. Can someone help explain to
me why ?


Thanks

Re: copy_*_user

[Américo Wang]

On Sun, Oct 23, 2011 at 1:06 AM, Xin Tong <xerox.time.tech@gmail.com> wrote:
> I am investigating copy_from_user  and copy_to_user in linux under
> i386. These two function both take a pointer with virtual address and
> a pointer with physical address.
>
> copy_from_user calls __copy_from_user_ll and copy_to_user  calls
> __copy_to_user_ll. It make sense to me that __copy_to_user_ll converts
> the virtual address to physical address using the current process's
> page table.
>
[...]
> But it seems to be that __copy_from_user_ll  is not converted the
> address at all before attempting to copy. Can someone help explain to
> me why ?
>

You missed that __copy_to_user_ll() only does that when CONFIG_X86_WP_WORKS_OK
is not defined. And there is a comment right inside __copy_to_user_ll() said:


               /*
                 * CPU does not honor the WP bit when writing
                 * from supervisory mode, and due to preemption or SMP,
                 * the page tables can change at any time.
                 * Do it manually.      Manfred <manfred@colorfullife.com>
                 */

this is why it uses kmap_atomic()+memcpy() to copy the data.

Also, all the addresses are virtual address.

Re: copy_*_user

[Xin Tong]

I found that __copy_from_user_ll(void *to, const void __user *from,
unsigned long n) eventually calls some asm which use mov and rep on
the passed-in to & from pointers. This is in kernel mode and to & from
are virtual addresses ? are not the kernel accessing physical RAM
directly ( without pagetable ) ?


Thanks

Xin



On Sat, Oct 22, 2011 at 10:50 PM, Américo Wang <xiyou.wangcong@gmail.com> wrote:
> On Sun, Oct 23, 2011 at 1:06 AM, Xin Tong <xerox.time.tech@gmail.com> wrote:
>> I am investigating copy_from_user  and copy_to_user in linux under
>> i386. These two function both take a pointer with virtual address and
>> a pointer with physical address.
>>
>> copy_from_user calls __copy_from_user_ll and copy_to_user  calls
>> __copy_to_user_ll. It make sense to me that __copy_to_user_ll converts
>> the virtual address to physical address using the current process's
>> page table.
>>
> [...]
>> But it seems to be that __copy_from_user_ll  is not converted the
>> address at all before attempting to copy. Can someone help explain to
>> me why ?
>>
>
> You missed that __copy_to_user_ll() only does that when CONFIG_X86_WP_WORKS_OK
> is not defined. And there is a comment right inside __copy_to_user_ll() said:
>
>
>               /*
>                 * CPU does not honor the WP bit when writing
>                 * from supervisory mode, and due to preemption or SMP,
>                 * the page tables can change at any time.
>                 * Do it manually.      Manfred <manfred@colorfullife.com>
>                 */
>
> this is why it uses kmap_atomic()+memcpy() to copy the data.
>
> Also, all the addresses are virtual address.
>

Re: copy_*_user

[H. Peter Anvin]

On 10/23/2011 11:23 AM, Xin Tong wrote:
> I found that __copy_from_user_ll(void *to, const void __user *from,
> unsigned long n) eventually calls some asm which use mov and rep on
> the passed-in to & from pointers. This is in kernel mode and to & from
> are virtual addresses ? are not the kernel accessing physical RAM
> directly ( without pagetable ) ?

No.  What gave you that idea?

	-hpa

Re: copy_*_user

[Xin Tong]

I do not remember, probably I was under the impression that kernel
accesses physical memory directly from my os course in college. So ,
the kernel access memory using the page table of the current process
here?

Thanks

Xin


On Sun, Oct 23, 2011 at 6:15 AM, H. Peter Anvin <hpa@zytor.com> wrote:
> On 10/23/2011 11:23 AM, Xin Tong wrote:
>> I found that __copy_from_user_ll(void *to, const void __user *from,
>> unsigned long n) eventually calls some asm which use mov and rep on
>> the passed-in to & from pointers. This is in kernel mode and to & from
>> are virtual addresses ? are not the kernel accessing physical RAM
>> directly ( without pagetable ) ?
>
> No.  What gave you that idea?
>
>        -hpa
>
>