Re: Re: [PATCH] Fix >4G i386 PAE grant table interface

[Steven Rostedt <srostedt@redhat.com>]

Jan Beulich wrote:
>>>> Keir Fraser <Keir.Fraser@cl.cam.ac.uk> 02.11.06 19:16 >>>
>> On 2/11/06 16:53, "Steven Rostedt" <srostedt@redhat.com> wrote:
>>
>>> This patch updates both the linux-2.6-sparse tree and the xen HV to use
>>> uint64_t instead of unsigned long for those particular functions.  This
>>> patch has been tested on RHEL5 Beta on a box with 12G i386.
>>>
>>> Signed-off-by: Steven Rostedt <srostedt@redhat.com>
>> Nasty bug. At least it affects only blktap. Thanks.
> 
> Looking at this it would seem to me that the second call to
> gnttab_set_unmap_op in blkltap.c is missing the
> GNTMAP_contains_pte flag, affecting auto_translated_physmap
> guests. 

It's much worst than that!

Without this patch, my patch is still broken.  In the hypervisor this 
flag is checked, and will go down the wrong path when it is missing:


int destroy_grant_host_mapping(
     u64 addr, unsigned long frame, unsigned int flags)
{
     if ( flags & GNTMAP_contains_pte )
         return destroy_grant_pte_mapping(addr, frame, current->domain);
     return destroy_grant_va_mapping(addr, frame, current);
}


You can see here that we call the va_mapping instead of the pte mapping. 
Which means that we once again truncate the top 32 bits of the address, 
since a va addr is expected to be "unsigned long" in size.

Also, this can be bad, if the va addr is really virtual, and doesn't 
match the actual offset shift (like handling high memory).

> 
> Signed-off-by: Jan Beulich <jbeulich@novell.com>
> 

please apply.

Acked-by: Steven Rostedt <srostedt@redhat.com>

-- Steve