Re: [PATCH v13 11/22] vfio iommu: Add blocking notifier to notify DMA_UNMAP

From: Kirti Wankhede
Date: Tue Nov 15 2016 - 23:17:51 EST




On 11/16/2016 9:28 AM, Alex Williamson wrote:
> On Wed, 16 Nov 2016 09:13:37 +0530
> Kirti Wankhede <kwankhede@xxxxxxxxxx> wrote:
>
>> On 11/16/2016 8:55 AM, Alex Williamson wrote:
>>> On Tue, 15 Nov 2016 20:16:12 -0700
>>> Alex Williamson <alex.williamson@xxxxxxxxxx> wrote:
>>>
>>>> On Wed, 16 Nov 2016 08:16:15 +0530
>>>> Kirti Wankhede <kwankhede@xxxxxxxxxx> wrote:
>>>>
>>>>> On 11/16/2016 3:49 AM, Alex Williamson wrote:
>>>>>> On Tue, 15 Nov 2016 20:59:54 +0530
>>>>>> Kirti Wankhede <kwankhede@xxxxxxxxxx> wrote:
>>>>>>
>>>>> ...
>>>>>
>>>>>>> @@ -854,7 +857,28 @@ static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
>>>>>>> */
>>>>>>> if (dma->task->mm != current->mm)
>>>>>>> break;
>>>>>>> +
>>>>>>> unmapped += dma->size;
>>>>>>> +
>>>>>>> + if (iommu->external_domain && !RB_EMPTY_ROOT(&dma->pfn_list)) {
>>>>>>> + struct vfio_iommu_type1_dma_unmap nb_unmap;
>>>>>>> +
>>>>>>> + nb_unmap.iova = dma->iova;
>>>>>>> + nb_unmap.size = dma->size;
>>>>>>> +
>>>>>>> + /*
>>>>>>> + * Notifier callback would call vfio_unpin_pages() which
>>>>>>> + * would acquire iommu->lock. Release lock here and
>>>>>>> + * reacquire it again.
>>>>>>> + */
>>>>>>> + mutex_unlock(&iommu->lock);
>>>>>>> + blocking_notifier_call_chain(&iommu->notifier,
>>>>>>> + VFIO_IOMMU_NOTIFY_DMA_UNMAP,
>>>>>>> + &nb_unmap);
>>>>>>> + mutex_lock(&iommu->lock);
>>>>>>> + if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
>>>>>>> + break;
>>>>>>> + }
>>>>>>
>>>>>>
>>>>>> Why exactly do we need to notify per vfio_dma rather than per unmap
>>>>>> request? If we do the latter we can send the notify first, limiting us
>>>>>> to races where a page is pinned between the notify and the locking,
>>>>>> whereas here, even our dma pointer is suspect once we re-acquire the
>>>>>> lock, we don't technically know if another unmap could have removed
>>>>>> that already. Perhaps something like this (untested):
>>>>>>
>>>>>
>>>>> There are checks to validate unmap request, like v2 check and who is
>>>>> calling unmap and is it allowed for that task to unmap. Before these
>>>>> checks its not sure that unmap region range which asked for would be
>>>>> unmapped all. Notify call should be at the place where its sure that the
>>>>> range provided to notify call is definitely going to be removed. My
>>>>> change do that.
>>>>
>>>> Ok, but that does solve the problem. What about this (untested):
>>>
>>> s/does/does not/
>>>
>>> BTW, I like how the retries here fill the gap in my previous proposal
>>> where we could still race re-pinning. We've given it an honest shot or
>>> someone is not participating if we've retried 10 times. I don't
>>> understand why the test for iommu->external_domain was there, clearly
>>> if the list is not empty, we need to notify. Thanks,
>>>
>>
>> Ok. Retry is good to give a chance to unpin all. But is it really
>> required to use BUG_ON() that would panic the host. I think WARN_ON
>> should be fine and then when container is closed or when the last group
>> is removed from the container, vfio_iommu_type1_release() is called and
>> we have a chance to unpin it all.
>
> See my comments on patch 10/22, we need to be vigilant that the vendor
> driver is participating. I don't think we should be cleaning up after
> the vendor driver on release, if we need to do that, it implies we
> already have problems in multi-mdev containers since we'll be left with
> pfn_list entries that no longer have an owner. Thanks,
>

If any vendor driver doesn't clean its pinned pages and there are
entries in pfn_list with no owner, that would be indicated by WARN_ON,
which should be fixed by that vendor driver. I still feel it shouldn't
cause host panic.
When such warning is seen with multiple mdev devices in container, it is
easy to isolate and find which vendor driver is not cleaning their
stuff, same warning would be seen with single mdev device in a
container. To isolate and find which vendor driver is culprit check with
one mdev device at a time.
Finally, we have a chance to clean all residue from
vfio_iommu_type1_release() so that vfio_iommu_type1 module doesn't leave
any leaks.

Thanks,
Kirti