Re: [EXT] Re: [PATCH v7 3/5] virtio-iommu: Call iommu notifier for attach/detach

[Auger Eric]

Hi Jean,

On 3/18/20 12:20 PM, Bharat Bhushan wrote:
> 
> 
>> -----Original Message-----
>> From: Jean-Philippe Brucker <address@hidden>
>> Sent: Wednesday, March 18, 2020 4:48 PM
>> To: Bharat Bhushan <address@hidden>
>> Cc: Auger Eric <address@hidden>; Peter Maydell
>> <address@hidden>; address@hidden; Tomasz Nowicki [C]
>> <address@hidden>; address@hidden; address@hidden;
>> address@hidden; address@hidden; address@hidden;
>> address@hidden; Bharat Bhushan <address@hidden>;
>> address@hidden; address@hidden
>> Subject: [EXT] Re: [PATCH v7 3/5] virtio-iommu: Call iommu notifier for
>> attach/detach
>>
>> External Email
>>
>> ----------------------------------------------------------------------
>> On Wed, Mar 18, 2020 at 03:47:44PM +0530, Bharat Bhushan wrote:
>>> Hi Jean,
>>>
>>> On Tue, Mar 17, 2020 at 9:29 PM Jean-Philippe Brucker
>>> <address@hidden> wrote:
>>>>
>>>> On Tue, Mar 17, 2020 at 02:46:55PM +0530, Bharat Bhushan wrote:
>>>>> Hi Jean,
>>>>>
>>>>> On Tue, Mar 17, 2020 at 2:23 PM Jean-Philippe Brucker
>>>>> <address@hidden> wrote:
>>>>>>
>>>>>> On Tue, Mar 17, 2020 at 12:40:39PM +0530, Bharat Bhushan wrote:
>>>>>>> Hi Jean,
>>>>>>>
>>>>>>> On Mon, Mar 16, 2020 at 3:41 PM Jean-Philippe Brucker
>>>>>>> <address@hidden> wrote:
>>>>>>>>
>>>>>>>> Hi Bharat,
>>>>>>>>
>>>>>>>> Could you Cc me on your next posting?  Unfortunately I don't
>>>>>>>> have much hardware for testing this at the moment, but I
>>>>>>>> might be able to help a little on the review.
>>>>>>>>
>>>>>>>> On Mon, Mar 16, 2020 at 02:40:00PM +0530, Bharat Bhushan wrote:
>>>>>>>>>>>>> First issue is: your guest can use 4K page and your
>>>>>>>>>>>>> host can use 64KB pages. In that case VFIO_DMA_MAP
>>>>>>>>>>>>> will fail with -EINVAL. We must devise a way to pass the host
>> settings to the VIRTIO-IOMMU device.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Even with 64KB pages, it did not work for me. I have
>>>>>>>>>>>>> obviously not the storm of VFIO_DMA_MAP failures but
>>>>>>>>>>>>> I have some, most probably due to some wrong notifications
>> somewhere. I will try to investigate on my side.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Did you test with VFIO on your side?
>>>>>>>>>>>>
>>>>>>>>>>>> I did not tried with different page sizes, only tested with 4K page
>> size.
>>>>>>>>>>>>
>>>>>>>>>>>> Yes it works, I tested with two n/w device assigned
>>>>>>>>>>>> to VM, both interfaces works
>>>>>>>>>>>>
>>>>>>>>>>>> First I will try with 64k page size.
>>>>>>>>>>>
>>>>>>>>>>> 64K page size does not work for me as well,
>>>>>>>>>>>
>>>>>>>>>>> I think we are not passing correct page_size_mask here
>>>>>>>>>>> (config.page_size_mask is set to TARGET_PAGE_MASK (
>>>>>>>>>>> which is
>>>>>>>>>>> 0xfffffffffffff000))
>>>>>>>>>> I guess you mean with guest using 4K and host using 64K.
>>>>>>>>>>>
>>>>>>>>>>> We need to set this correctly as per host page size, correct?
>>>>>>>>>> Yes that's correct. We need to put in place a control
>>>>>>>>>> path to retrieve the page settings on host through VFIO to inform the
>> virtio-iommu device.
>>>>>>>>>>
>>>>>>>>>> Besides this issue, did you try with 64kB on host and guest?
>>>>>>>>>
>>>>>>>>> I tried Followings
>>>>>>>>>   - 4k host and 4k guest  - it works with v7 version
>>>>>>>>>   - 64k host and 64k guest - it does not work with v7
>>>>>>>>>     hard-coded config.page_size_mask to 0xffffffffffff0000
>>>>>>>>> and it works
>>>>>>>>
>>>>>>>> You might get this from the iova_pgsize bitmap returned by
>>>>>>>> VFIO_IOMMU_GET_INFO. The virtio config.page_size_mask is
>>>>>>>> global so there is the usual problem of aggregating
>>>>>>>> consistent properties, but I'm guessing using the host page size as a
>> granule here is safe enough.
>>>>>>>>
>>>>>>>> If it is a problem, we can add a PROBE property for page
>>>>>>>> size mask, allowing to define per-endpoint page masks. I
>>>>>>>> have kernel patches somewhere to do just that.
>>>>>>>
>>>>>>> I do not see we need page size mask per endpoint.
>>>>>>>
>>>>>>> While I am trying to understand what "page-size-mask" guest
>>>>>>> will work with
>>>>>>>
>>>>>>> - 4K page size host and 4k page size guest
>>>>>>>   config.page_size_mask = 0xffffffffffff000 will work
>>>>>>>
>>>>>>> - 64K page size host and 64k page size guest
>>>>>>>   config.page_size_mask = 0xfffffffffff0000 will work
>>>>>>>
>>>>>>> - 64K page size host and 4k page size guest
>>>>>>>    1) config.page_size_mask = 0xffffffffffff000 will also not
>>>>>>> work as VFIO in host expect iova and size to be aligned to 64k
>>>>>>> (PAGE_SIZE in
>>>>>>> host)
>>>>>>>    2) config.page_size_mask = 0xfffffffffff0000 will not work,
>>>>>>> iova initialization (in guest) expect minimum page-size
>>>>>>> supported by h/w to be equal to 4k (PAGE_SIZE in guest)
>>>>>>>        Should we look to relax this in iova allocation code?
>>>>>>
>>>>>> Oh right, that's not great. Maybe the BUG_ON() can be removed,
>>>>>> I'll ask on the list.
>>>>>
>>>>> yes, the BUG_ON in iova_init.
So you mean in init_iova_domain()?

I see the BUG_ON was introduced in
0fb5fe874c42942e16c450ae05da453e13a1c09e "iommu: Make IOVA domain page
size explicit" but was it meant to remain?

Logically when we allocate buffer IOVAs for DMA accesses, later on,
shouldn't it be possible to use the actual granule set on init() and
make sure the allocated size is properly aligned.

Reading the commit msg it explicitly says that "the systems may contain
heterogeneous IOMMUs supporting differing minimum page sizes, which may
also not be common with the CPU page size".

Thanks

Eric


>>>>> I tried with removing same and it worked, but not analyzed side effects.
>>>>
>>>> It might break the assumption from device drivers that mapping a
>>>> page is safe. For example they call alloc_page() followed by
>>>> dma_map_page(). In our situation dma-iommu.c will oblige and create
>>>> one 64k mapping to one 4k physical page. As a result the endpoint
>>>> can access the neighbouring 60k of memory.
>>>>
>>>> This isn't too terrible. After all, even when the page sizes match,
>>>> device drivers can call dma_map_single() on sub-page buffers, which
>>>> will also let the endpoint access a whole page. The solution, if you
>>>> don't trust the endpoint, is to use bounce buffers.
>>>>
>>>> But I suspect it's not as simple as removing the BUG_ON(), we'll
>>>> need to go over dma-iommu.c first. And it seems like assigning
>>>> endpoints to guest userspace won't work either in this config. In
>> vfio_dma_do_map():
>>>>
>>>>         mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) -
>>>> 1;
>>>>
>>>>         WARN_ON(mask & PAGE_MASK);
>>>
>>> Yes, Agree
>>>
>>>>
>>>> If I read this correctly the WARN will trigger in a 4k guest under
>>>> 64k host, right?  So maybe we can just say that this config isn't
>>>> supported, unless it's an important use-case for virtio-iommu?
>>>
>>> I sent v8 version of patch and with that guest and host with same page
>>> size should work.
>>> While i have not yet added analyzed how to mark 4k guest and 64k host
>>> as un-supported configuration, will analyze and send patch.
>>
>> I don't think there is anything to do for QEMU, it's Linux that doesn't 
>> support the
>> configuration. We could add something like the attached patch, in the virtio-
>> iommu driver, to abort more gracefully than with a BUG_ON().
> 
> Yes agree, we need to have change in Linux side.
> 
> Thanks
> -Bharat
> 
>>
>> Thanks,
>> Jean
>