Date:   Wed, 9 Nov 2022 08:29:28 +0100
From:   Eric Auger <eric.auger@...hat.com>
To:     Jason Wang <jasowang@...hat.com>
Cc:     "Michael S. Tsirkin" <mst@...hat.com>, eric.auger.pro@...il.com,
        kvm@...r.kernel.org, virtualization@...ts.linux-foundation.org,
        netdev@...r.kernel.org, linux-kernel@...r.kernel.org,
        peterx@...hat.com
Subject: Re: [RFC] vhost: Clear the pending messages on
 vhost_init_device_iotlb()

Hi Jason,

On 11/9/22 04:44, Jason Wang wrote:
> On Tue, Nov 8, 2022 at 6:17 PM Eric Auger <eric.auger@...hat.com> wrote:
>> Hi Michael, Jason,
>>
>> On 11/8/22 10:31, Michael S. Tsirkin wrote:
>>> On Tue, Nov 08, 2022 at 05:13:50PM +0800, Jason Wang wrote:
>>>> On Tue, Nov 8, 2022 at 4:56 PM Michael S. Tsirkin <mst@...hat.com> wrote:
>>>>> On Tue, Nov 08, 2022 at 11:09:36AM +0800, Jason Wang wrote:
>>>>>> On Tue, Nov 8, 2022 at 7:06 AM Michael S. Tsirkin <mst@...hat.com> wrote:
>>>>>>> On Mon, Nov 07, 2022 at 10:10:06PM +0100, Eric Auger wrote:
>>>>>>>> Hi Michael,
>>>>>>>> On 11/7/22 21:42, Michael S. Tsirkin wrote:
>>>>>>>>> On Mon, Nov 07, 2022 at 09:34:31PM +0100, Eric Auger wrote:
>>>>>>>>>> When the vhost iotlb is used along with a guest virtual iommu
>>>>>>>>>> and the guest gets rebooted, some MISS messages may have been
>>>>>>>>>> recorded just before the reboot and spuriously executed by
>>>>>>>>>> the virtual iommu after the reboot. Despite the device iotlb gets
>>>>>>>>>> re-initialized, the messages are not cleared. Fix that by calling
>>>>>>>>>> vhost_clear_msg() at the end of vhost_init_device_iotlb().
>>>>>>>>>>
>>>>>>>>>> Signed-off-by: Eric Auger <eric.auger@...hat.com>
>>>>>>>>>> ---
>>>>>>>>>>  drivers/vhost/vhost.c | 1 +
>>>>>>>>>>  1 file changed, 1 insertion(+)
>>>>>>>>>>
>>>>>>>>>> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
>>>>>>>>>> index 40097826cff0..422a1fdee0ca 100644
>>>>>>>>>> --- a/drivers/vhost/vhost.c
>>>>>>>>>> +++ b/drivers/vhost/vhost.c
>>>>>>>>>> @@ -1751,6 +1751,7 @@ int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
>>>>>>>>>>    }
>>>>>>>>>>
>>>>>>>>>>    vhost_iotlb_free(oiotlb);
>>>>>>>>>> +  vhost_clear_msg(d);
>>>>>>>>>>
>>>>>>>>>>    return 0;
>>>>>>>>>>  }
>>>>>>>>> Hmm.  Can't messages meanwhile get processes and affect the
>>>>>>>>> new iotlb?
>>>>>>>> Isn't the msg processing stopped at the moment this function is called
>>>>>>>> (VHOST_SET_FEATURES)?
>>>>>>>>
>>>>>>>> Thanks
>>>>>>>>
>>>>>>>> Eric
>>>>>>> It's pretty late here I'm not sure.  You tell me what prevents it.
>>>>>> So the proposed code assumes that Qemu doesn't process device IOTLB
>>>>>> before VHOST_SET_FEAETURES. Consider there's no reset in the general
>>>>>> vhost uAPI,  I wonder if it's better to move the clear to device code
>>>>>> like VHOST_NET_SET_BACKEND. So we can clear it per vq?
>>>>> Hmm this makes no sense to me. iommu sits between backend
>>>>> and frontend. Tying one to another is going to backfire.
>>>> I think we need to emulate what real devices are doing. Device should
>>>> clear the page fault message during reset, so the driver won't read
>>>> anything after reset. But we don't have a per device stop or reset
>>>> message for vhost-net. That's why the VHOST_NET_SET_BACKEND came into
>>>> my mind.
>>> That's not a reset message. Userspace can switch backends at will.
>>> I guess we could check when backend is set to -1.
>>> It's a hack but might work.
>>>
>>>>> I'm thinking more along the lines of doing everything
>>>>> under iotlb_lock.
>>>> I think the problem is we need to find a proper place to clear the
>>>> message. So I don't get how iotlb_lock can help: the message could be
>>>> still read from user space after the backend is set to NULL.
>>>>
>>>> Thanks
>>> Well I think the real problem is this.
>>>
>>> vhost_net_set_features does:
>>>
>>>         if ((features & (1ULL << VIRTIO_F_ACCESS_PLATFORM))) {
>>>                 if (vhost_init_device_iotlb(&n->dev, true))
>>>                         goto out_unlock;
>>>         }
>>>
>>>
>>> so we get a new iotlb each time features are set.
>>>
>>> But features can be changes while device is running.
>>> E.g.
>>>       VHOST_F_LOG_ALL
>>>
>>>
>>> Let's just say this hack of reusing feature bits for backend
>>> was not my brightest idea :(
>>>
>> Isn't vhost_init_device_iotlb() racy then, as d->iotlb is first updated with niotlb and later d->vqs[i]->iotlb is updated with niotlb. What does garantee this is done atomically?
>>
>> Shouldn't we hold the dev->mutex to make all the sequence atomic and
>> include vhost_clear_msg()?  Can't the vhost_clear_msg() take the dev lock?
> It depends on where we want to place the vhost_clear_msg(), e.g in
> most of the device ioctl, the dev->mutex has been held.

OK, I will double check and respin accordingly

Eric
>
> Thanks
>
>> Thanks
>>
>> Eric
>>
>>>
>>>
>>>>>
>>>>>>> BTW vhost_init_device_iotlb gets enabled parameter but ignores
>>>>>>> it, we really should drop that.
>>>>>> Yes.
>>>>>>
>>>>>>> Also, it looks like if features are set with VIRTIO_F_ACCESS_PLATFORM
>>>>>>> and then cleared, iotlb is not properly cleared - bug?
>>>>>> Not sure, old IOTLB may still work. But for safety, we need to disable
>>>>>> device IOTLB in this case.
>>>>>>
>>>>>> Thanks
>>>>>>
>>>>>>>>>> --
>>>>>>>>>> 2.37.3

Powered by blists - more mailing lists