Re: [Xen-devel] [PATCH] xen/p2m: fix extra memory regions accounting

[Juergen Gross <jgross@suse.com>]

On 09/03/2015 05:01 PM, David Vrabel wrote:
> On 03/09/15 15:55, Juergen Gross wrote:
>> On 09/03/2015 04:52 PM, David Vrabel wrote:
>>> On 03/09/15 15:45, David Vrabel wrote:
>>>> On 03/09/15 15:38, Roger Pau Monné wrote:
>>>>> El 03/09/15 a les 14.25, Juergen Gross ha escrit:
>>>>>> On 09/03/2015 02:05 PM, Roger Pau Monne wrote:
>>>>>>> On systems with memory maps with ranges that don't end at page
>>>>>>> boundaries,
>>>>>>> like:
>>>>>>>
>>>>>>> [...]
>>>>>>> (XEN)  0000000000100000 - 00000000dfdf9c00 (usable)
>>>>>>> (XEN)  00000000dfdf9c00 - 00000000dfe4bc00 (ACPI NVS)
>>>>>>> [...]
>>>>>>>
>>>>>>> xen_add_extra_mem will create a protected range that ends up at
>>>>>>> 0xdfdf9c00,
>>>>>>> but the function used to check if a memory address is inside of a
>>>>>>> protected
>>>>>>> range works with pfns, which means that an attempt to map 0xdfdf9c00
>>>>>>> will be
>>>>>>> refused because the check is performed against 0xdfdf9000 instead of
>>>>>>> 0xdfdf9c00.
>>>>>>>
>>>>>>> In order to fix this, make sure that the ranges that are added to the
>>>>>>> xen_extra_mem array are aligned to page boundaries.
>>>>>>>
>>>>>>> Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
>>>>>>> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
>>>>>>> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
>>>>>>> Cc: David Vrabel <david.vrabel@citrix.com>
>>>>>>> Cc: Juergen Gross <jgross@suse.com>
>>>>>>> Cc: xen-devel@lists.xenproject.org
>>>>>>> ---
>>>>>>> AFAICT this patch needs to be backported to 3.19, 4.0, 4.1 and 4.2.
>>>>>>> ---
>>>>>>>     arch/x86/xen/setup.c | 6 ++++++
>>>>>>>     1 file changed, 6 insertions(+)
>>>>>>>
>>>>>>> diff --git a/arch/x86/xen/setup.c b/arch/x86/xen/setup.c
>>>>>>> index 55f388e..dcf5865 100644
>>>>>>> --- a/arch/x86/xen/setup.c
>>>>>>> +++ b/arch/x86/xen/setup.c
>>>>>>> @@ -68,6 +68,9 @@ static void __init xen_add_extra_mem(phys_addr_t
>>>>>>> start, phys_addr_t size)
>>>>>>>     {
>>>>>>>         int i;
>>>>>>>
>>>>>>> +    start = PAGE_ALIGN(start);
>>>>>>> +    size &= PAGE_MASK;
>>>>>>
>>>>>> This is not correct. If start wasn't page aligned and size was, you'll
>>>>>> add one additional page to xen_extra_mem.
>>>>>
>>>>> I'm not understanding this, let's put an example:
>>>>>
>>>>> start = 0x8c00
>>>>> size = 0x1000
>>>>>
>>>>> After the fixup added above this would become:
>>>>>
>>>>> start = 0x9000
>>>>> size = 0x1000
>>>>>
>>>>> So if anything, I'm adding one page less (because 0x8000 was partly
>>>>> added, and with the fixup it is not added).
>>>>
>>>> We expand the reserved (i.e., non-RAM) areas down so they're fully
>>>> covered with whole pages when we depopulate and 1:1 map them, we should
>>>> add extra memory regions that cover these same areas.
>>>
>>> Ignore this.  This was nonsense.
>>>
>>> We expand the reserved (i.e., non-RAM) areas so they're fully covered
>>> with whole pages when we depopulate and 1:1 map them, we should add the
>>> extra memory such that it does not overlap with with expanded regions.
>>> i.e., round up the start and round down the end (like Roger's patch
>>> does).
>>
>> Nearly. Roger's patch rounds up start and rounds down the size. It might
>> add non-RAM partial pages to xen_extra_mem.
>
> Yes.  You're right.

Hmm, thinking more about it, I'd prefer to change xen_extra_mem to use
pfns instead of physical addresses. This would make things much more
clear.

Roger, do you want to do the patch or should I?


Juergen