[dpdk-dev] [PATCH] vhost: fix async DMA map

[Burakov, Anatoly]

On 26-Oct-21 11:58 AM, Burakov, Anatoly wrote:
> On 26-Oct-21 11:27 AM, Ding, Xuan wrote:
>> Hi,
>>
>>> -----Original Message-----
>>> From: Maxime Coquelin <maxime.coquelin at redhat.com>
>>> Sent: Tuesday, October 26, 2021 5:49 PM
>>> To: Ding, Xuan <xuan.ding at intel.com>; dev at dpdk.org;
>>> david.marchand at redhat.com; Xia, Chenbo <chenbo.xia at intel.com>
>>> Cc: Burakov, Anatoly <anatoly.burakov at intel.com>
>>> Subject: Re: [PATCH] vhost: fix async DMA map
>>>
>>>
>>>
>>> On 10/26/21 10:49, Ding, Xuan wrote:
>>>> Hi Maxime,
>>>>
>>>>> -----Original Message-----
>>>>> From: Maxime Coquelin <maxime.coquelin at redhat.com>
>>>>> Sent: Tuesday, October 26, 2021 2:53 PM
>>>>> To: Ding, Xuan <xuan.ding at intel.com>; dev at dpdk.org;
>>>>> david.marchand at redhat.com; Xia, Chenbo <chenbo.xia at intel.com>
>>>>> Cc: Burakov, Anatoly <anatoly.burakov at intel.com>
>>>>> Subject: Re: [PATCH] vhost: fix async DMA map
>>>>>
>>>>>
>>>>>
>>>>> On 10/26/21 04:07, Ding, Xuan wrote:
>>>>>> Hi Maxime,
>>>>>>
>>>>>>> -----Original Message-----
>>>>>>> From: Maxime Coquelin <maxime.coquelin at redhat.com>
>>>>>>> Sent: Tuesday, October 26, 2021 4:47 AM
>>>>>>> To: dev at dpdk.org; david.marchand at redhat.com; Xia, Chenbo
>>>>>>> <chenbo.xia at intel.com>; Ding, Xuan <xuan.ding at intel.com>
>>>>>>> Subject: Re: [PATCH] vhost: fix async DMA map
>>>>>>>
>>>>>>> Hi Xuan,
>>>>>>>
>>>>>>> On 10/25/21 22:33, Maxime Coquelin wrote:
>>>>>>>> This patch fixes possible NULL-pointer dereferencing
>>>>>>>> reported by Coverity and also fixes NUMA reallocation
>>>>>>>> of the async DMA map.
>>>>>>>>
>>>>>>>> Fixes: 7c61fa08b716 ("vhost: enable IOMMU for async vhost")
>>>>>>>>
>>>>>>>> Coverity issue: 373655
>>>>>>>>
>>>>>>>> Signed-off-by: Maxime Coquelin <maxime.coquelin at redhat.com>
>>>>>>>> ---
>>>>>>>>      lib/vhost/vhost_user.c | 45 
>>>>>>>> +++++++++++++++++++-----------------------
>>>>>>>>      1 file changed, 20 insertions(+), 25 deletions(-)
>>>>>>>>
>>>>>>>
>>>>>>> I posted this patch to fix the issue reported by Coverity and 
>>>>>>> also other
>>>>>>> issue on NUMA realloc that I found at the same time. But I wonder
>>>>>>> whether all this async_map_status is needed.
>>>>>>
>>>>>> Thanks for your fix! I can help to review and test the patch later.
>>>>>>
>>>>>> I add the async_map_status in v2 for compatibility. Some DMA device,
>>>>>> like DSA, may use kernel idxd driver only. If there is no device 
>>>>>> bound to
>>>>>> DPDK vfio and kernel vfio module is modprobed to ensure
>>>>> rte_vfio_is_enabled() is true,
>>>>>> we will unavoidably do DMA map/unmap and it will fail.
>>>>>>
>>>>>> Therefore, the dma_map_status here is used to filter this 
>>>>>> situation by
>>>>> preventing
>>>>>> unnecessary DMA unmap.
>>>>>
>>>>> Ok, then I think we can just remove the async DMA map.
>>>>>
>>>>>>>
>>>>>>> Indeed, if the only place where we DMA map is in
>>>>>>> vhost_user_mmap_region(). If it fails, the error is propagated, 
>>>>>>> the mem
>>>>>>> table are freed and NACK is replied to the master. IOW, the 
>>>>>>> device will
>>>>>>> be in an unusable state.
>>>>>>
>>>>>> I agree with you, this is the place I consider right to do DMA map
>>>>>> because we also do SW mapping here, any suggestions?
>>>>>
>>>>> No suggestion, I was just explaining that at the only place where
>>>>> DMA map were done, mapping errors were properly handled and
>>> propagated.
>>>>
>>>> What about just setting async_copy to false, and allow switching to 
>>>> sync
>>> path.
>>>>
>>>>>
>>>>>>>
>>>>>>> Removing the async DMA map will simplify a lot the code, do you 
>>>>>>> agree
>>> to
>>>>>>> remove it or there is something I missed?
>>>>>>
>>>>>> See above. Indeed, it adds a lot of code. But we can't know the 
>>>>>> driver for
>>>>>> each device in vhost lib, or we can only restrict the user to bind 
>>>>>> some
>>>>> devices
>>>>>> to DPDK vfio if async logic needed.
>>>>>
>>>>> I would think we don't care if DMA unmap fails, we can just do the 
>>>>> same
>>>>> as what you do for DMA map, i.e. just ignore the error.
>>>>
>>>> Get your idea, we can do the same as DMA map, and in this way
>>> dma_map_status flag can be removed.
>>>>
>>>>>
>>>>> Thanks to this discussion, I have now more concerns on how it works. I
>>>>> think we have a problem here in case of DMA device hotplug, that 
>>>>> device
>>>>> could miss the necessary map entries from Vhost if no VFIO devices 
>>>>> were
>>>>> attached at VHST_USER_SET_MEM_TABLE time. How would you handle
>>> that
>>>>> case?
>>>>
>>>> DMA device is uncore, so I don't see the  hotplug issue here.
>>>
>>> I'm not sure what 'uncore' is, I suppose you mean your device cannot be
>>> physically added/removed to the system.
>>
>> Yes, we are at the same understanding.
>>
>>>
>>> I was not clear enough in my question. I meant that for example, the
>>> application is started and the Vhost port is created. Then, the DMA
>>> device is bound to VFIO, and probed by the DPDK application. Finally,
>>> the application register the DMA device as an async channel in Vhost.
>>>
>>> I think it will not work as the SET_MEM_TABLE will already have
>>> happened, so the guest memory won't be mapped in the VFIO container.
>>
>> Assuming the DMA device supports hotplug, this situation was not 
>> actually took into consideration, and maybe the handling should be 
>> added in app, with an event callback.
>>
>>>
>>> Do you have the same understanding?
>>>
>>>> I will have another patch containing compatibility with sync path, and
>>> async_map_status flag will be removed.
>>>> Hope to get your insights.
>>>
>>> What do you mean by compatibility with sync path?
>>
>> I mean whatever DMA map succeeds or fails, we return 0 so as not to 
>> prevent
>> SET_EM_TABLE.
>>
>>>
>>> Thanks for taking care of the async map status removal.
>>
>> Found a tricky point for removing dma_map_status flag, currently DMA 
>> unmap API will directly access the vfio_config,
>> And if without dma_map_status, we cannot get rte_errno at this moment.
> 
> This is a bug in VFIO API. We always assume VFIO type is valid, but it 
> is only valid whenever 1) VFIO support is detected at startup, and 2) we 
> assign at least one device to VFIO. If VFIO was detected but no devices 
> were assigned, the VFIO type will not be assigned, and the pointer will 
> be NULL. In this case, we should not attempt any unmaps, because there 
> are no active devices.
> 
> I'll submit a patch to fix this, thanks for catching this issue!

Hi Xuan/Maxime,

It would be great if you could test and confirm that this patch fixes 
the segfault:

http://patches.dpdk.org/project/dpdk/patch/8079312ba39435a0ac92e084cc1a3fe291008a47.1635254797.git.anatoly.burakov@intel.com/

-- 
Thanks,
Anatoly