|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [Xen-devel] [v7][RFC][PATCH 08/13] xen/x86/p2m: set p2m_access_n for reserved device memory mapping
>>> On 31.10.14 at 03:50, <tiejun.chen@xxxxxxxxx> wrote:
> On 2014/10/30 17:24, Jan Beulich wrote:
>>>>> On 30.10.14 at 08:39, <tiejun.chen@xxxxxxxxx> wrote:
>>> On 2014/10/29 17:20, Jan Beulich wrote:
>>>> Getting closer. Just set a to p2m->default_access before the if(),
>>>> and overwrite it when rc == 1 inside the if(). And properly handle
>>>> the error case (just logging a message - which btw lacks a proper
>>>> XENLOG_G_* prefix - doesn't seem enough to me).
>>>
>>> Please check the follows:
>>>
>>> @@ -686,8 +686,22 @@ guest_physmap_add_entry(struct domain *d, unsigned
>>> long gfn,
>>> /* Now, actually do the two-way mapping */
>>> if ( mfn_valid(_mfn(mfn)) )
>>> {
>>> - rc = p2m_set_entry(p2m, gfn, _mfn(mfn), page_order, t,
>>> - p2m->default_access);
>>> + rc = 0;
>>> + a = p2m->default_access;
>>> + if ( !is_hardware_domain(d) )
>>> + {
>>> + rc =
>>> iommu_get_reserved_device_memory(p2m_check_reserved_device_memory,
>>> + &gfn);
>>> + /* We need to set reserved device memory as p2m_access_n. */
>>> + if ( rc == 1 )
>>> + a = p2m_access_n;
>>> + else if ( rc < 0 )
>>> + printk(XENLOG_WARNING
>>> + "Domain %d can't check reserved device memory.\n",
>>> + d->domain_id);
>>> + }
>>> +
>>> + rc = p2m_set_entry(p2m, gfn, _mfn(mfn), page_order, t, a);
>>> if ( rc )
>>> goto out; /* Failed to update p2m, bail without updating
>>> m2p.
> */
>>
>> The handling of "a" looks good now, but the error handling and
>> logging is still as broken as it was before.
>
> Do you mean I'm missing some necessary info? Like gfn and mfn, so domain
> id, gfn and mfn can show enough message.
>
> Sorry I'm poor to understand what you expect.
But I explained it already, and that explanation is still visible in
the quotes above. But to avoid any doubt, I'll repeat: "And
properly handle the error case (just logging a message - which
btw lacks a proper XENLOG_G_* prefix - doesn't seem enough
to me)."
>>>> But then again this code may change altogether if you avoid
>>>> populating the reserved regions in the first place.
>>>
>>> Are you saying this scenario?
>>>
>>> #1 Here we first set these ranges as p2m_access_n
>>> #2 We reset them as 1:1 RMRR mapping with p2m_access_rw somewhere
>>> #3 Someone may try to populate these ranges again
>>
>> No. I pointed at the fact that if you avoid populating the holes,
>> there's no need to force them to p2m_access_n. Any attempts
>> to map other than the 1:1 thing there could then simply be
>> rejected.
>
> I think any population should be rejected totally, because 1:1 mapping
> means guest can access these RMRR ranges in case of no any device
> assignment with RMRR, right? Any access to these range corrupt the real
> device usage.
Oh yes, of course I implied that the 1:1 mapping would be
permitted only for those ranges where the RMRR corresponds
to a device the guest got assigned.
Jan
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel
|
![]() |
Lists.xenproject.org is hosted with RackSpace, monitoring our |