[PATCH] vhost: fix get hpa fail from guest pages

[Maxime Coquelin]

On 11/16/21 11:08, Ding, Xuan wrote:
> Hi,
> 
>> -----Original Message-----
>> From: Xia, Chenbo <chenbo.xia at intel.com>
>> Sent: 2021年11月16日 17:01
>> To: Ding, Xuan <xuan.ding at intel.com>; Wang, YuanX <yuanx.wang at intel.com>;
>> maxime.coquelin at redhat.com
>> Cc: dev at dpdk.org; Hu, Jiayu <jiayu.hu at intel.com>; Ma, WenwuX
>> <wenwux.ma at intel.com>; He, Xingguang <xingguang.he at intel.com>; Yang,
>> YvonneX <yvonnex.yang at intel.com>
>> Subject: RE: [PATCH] vhost: fix get hpa fail from guest pages
>>
>>> -----Original Message-----
>>> From: Ding, Xuan <xuan.ding at intel.com>
>>> Sent: Tuesday, November 16, 2021 4:47 PM
>>> To: Xia, Chenbo <chenbo.xia at intel.com>; Wang, YuanX
>>> <yuanx.wang at intel.com>; maxime.coquelin at redhat.com
>>> Cc: dev at dpdk.org; Hu, Jiayu <jiayu.hu at intel.com>; Ma, WenwuX
>>> <wenwux.ma at intel.com>; He, Xingguang <xingguang.he at intel.com>; Yang,
>>> YvonneX <yvonnex.yang at intel.com>
>>> Subject: RE: [PATCH] vhost: fix get hpa fail from guest pages
>>>
>>> Hi Chenbo,
>>>
>>>> -----Original Message-----
>>>> From: Xia, Chenbo <chenbo.xia at intel.com>
>>>> Sent: 2021年11月16日 16:07
>>>> To: Xia, Chenbo <chenbo.xia at intel.com>; Wang, YuanX
>>>> <yuanx.wang at intel.com>; maxime.coquelin at redhat.com
>>>> Cc: dev at dpdk.org; Hu, Jiayu <jiayu.hu at intel.com>; Ding, Xuan
>>>> <xuan.ding at intel.com>; Ma, WenwuX <wenwux.ma at intel.com>; He,
>>>> Xingguang <xingguang.he at intel.com>; Yang, YvonneX
>>>> <yvonnex.yang at intel.com>
>>>> Subject: RE: [PATCH] vhost: fix get hpa fail from guest pages
>>>>
>>>>> -----Original Message-----
>>>>> From: Xia, Chenbo <chenbo.xia at intel.com>
>>>>> Sent: Monday, November 15, 2021 4:04 PM
>>>>> To: Wang, YuanX <yuanx.wang at intel.com>;
>> maxime.coquelin at redhat.com
>>>>> Cc: dev at dpdk.org; Hu, Jiayu <jiayu.hu at intel.com>; Ding, Xuan
>>>>> <xuan.ding at intel.com>; Ma, WenwuX <wenwux.ma at intel.com>; He,
>>>> Xingguang
>>>>> <xingguang.he at intel.com>; Yang, YvonneX <yvonnex.yang at intel.com>
>>>>> Subject: RE: [PATCH] vhost: fix get hpa fail from guest pages
>>>>>
>>>>>> -----Original Message-----
>>>>>> From: Wang, YuanX <yuanx.wang at intel.com>
>>>>>> Sent: Thursday, November 11, 2021 2:27 PM
>>>>>> To: maxime.coquelin at redhat.com; Xia, Chenbo
>>>>>> <chenbo.xia at intel.com>
>>>>>> Cc: dev at dpdk.org; Hu, Jiayu <jiayu.hu at intel.com>; Ding, Xuan
>>>>>> <xuan.ding at intel.com>; Ma, WenwuX <wenwux.ma at intel.com>; He,
>>>>>> Xingguang <xingguang.he at intel.com>; Yang, YvonneX
>>>>>> <yvonnex.yang at intel.com>; Wang,
>>>>> YuanX
>>>>>> <yuanx.wang at intel.com>
>>>>>> Subject: [PATCH] vhost: fix get hpa fail from guest pages
>>>>>>
>>>>>> When processing front-end memory regions messages, vhost saves
>>>>>> the guest/host physical address mappings to guest pages and
>>>>>> merges adjacent contiguous pages if hpa is contiguous, however
>>>>>> gpa is likely not contiguous in PA mode and merging will cause
>>>>>> the gpa range to change.
>>>>>> This patch distinguishes the case of discontinuous gpa and does
>>>>>> a range lookup on gpa when doing a binary search.
>>>>>>
>>>>>> Fixes: e246896178e("vhost: get guest/host physical address
>>>>>> mappings")
>>>>>> Fixes: 6563cf92380 ("vhost: fix async copy on multi-page
>>>>>> buffers")
>>>>>>
>>>>>> Signed-off-by: Yuan Wang <yuanx.wang at intel.com>
>>>>>> ---
>>>>>>   lib/vhost/vhost.h      | 18 ++++++++++++++++--
>>>>>>   lib/vhost/vhost_user.c | 15 +++++++++++----
>>>>>>   2 files changed, 27 insertions(+), 6 deletions(-)
>>>>>>
>>>>>> diff --git a/lib/vhost/vhost.h b/lib/vhost/vhost.h index
>>>>>> 7085e0885c..b3f0c1d07c 100644
>>>>>> --- a/lib/vhost/vhost.h
>>>>>> +++ b/lib/vhost/vhost.h
>>>>>> @@ -587,6 +587,20 @@ static __rte_always_inline int
>>>>>> guest_page_addrcmp(const void *p1,
>>>>>>   	return 0;
>>>>>>   }
>>>>>>
>>>>>> +static __rte_always_inline int guest_page_rangecmp(const void
>>>>>> +*p1, const
>>>>> void
>>>>>> *p2)
>>>>>> +{
>>>>>> +	const struct guest_page *page1 = (const struct guest_page *)p1;
>>>>>> +	const struct guest_page *page2 = (const struct guest_page
>>>>>> +*)p2;
>>>>>> +
>>>>>> +	if (page1->guest_phys_addr >= page2->guest_phys_addr) {
>>>>>> +		if (page1->guest_phys_addr < page2->guest_phys_addr
>> +
>>>> page2->size)
>>>>>> +			return 0;
>>>>>> +		else
>>>>>> +			return 1;
>>>>>> +	} else
>>>>>> +		return -1;
>>>>>> +}
>>>>>> +
>>>>>>   static __rte_always_inline rte_iova_t  gpa_to_first_hpa(struct
>>>>>> virtio_net *dev, uint64_t gpa,
>>>>>>   	uint64_t gpa_size, uint64_t *hpa_size) @@ -597,9 +611,9 @@
>>>>>> gpa_to_first_hpa(struct virtio_net *dev, uint64_t gpa,
>>>>>>
>>>>>>   	*hpa_size = gpa_size;
>>>>>>   	if (dev->nr_guest_pages >= VHOST_BINARY_SEARCH_THRESH) {
>>>>>> -		key.guest_phys_addr = gpa & ~(dev->guest_pages[0].size - 1);
>>>>>> +		key.guest_phys_addr = gpa;
>>>>>>   		page = bsearch(&key, dev->guest_pages, dev->nr_guest_pages,
>>>>>> -			       sizeof(struct guest_page), guest_page_addrcmp);
>>>>>> +			       sizeof(struct guest_page),
>> guest_page_rangecmp);
>>>>>>   		if (page) {
>>>>>>   			if (gpa + gpa_size <=
>>>>>>   					page->guest_phys_addr + page->size)
>>>> { diff --git
>>>>>> a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c index
>>>>>> a781346c4d..7d58fde458 100644
>>>>>> --- a/lib/vhost/vhost_user.c
>>>>>> +++ b/lib/vhost/vhost_user.c
>>>>>> @@ -999,10 +999,17 @@ add_one_guest_page(struct virtio_net *dev,
>>>>>> uint64_t guest_phys_addr,
>>>>>>   	if (dev->nr_guest_pages > 0) {
>>>>>>   		last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
>>>>>>   		/* merge if the two pages are continuous */
>>>>>> -		if (host_phys_addr == last_page->host_phys_addr +
>>>>>> -				      last_page->size) {
>>>>>> -			last_page->size += size;
>>>>>> -			return 0;
>>>>>> +		if (host_phys_addr == last_page->host_phys_addr +
>>>>>> +last_page->size)
>>>>>> {
>>>>>> +			if (rte_eal_iova_mode() == RTE_IOVA_VA) {
>>>>>> +				last_page->size += size;
>>>>>> +				return 0;
>>>>>> +			}
>>>>>
>>>>> This makes me think about a question: In IOVA_VA mode, what
>>>>> ensures HPA and GPA are both contiguous?
>>>>
>>>> When I wrote this email, I thought host_phys_addr is HPA but in VA
>>>> mode,
>>> it's
>>>> Host IOVA, aka VA in DPDK's case. So in most cases GPA and VA will
>>>> be both contiguous when the contiguous pages are all in one memory
>>>> region. But I
>>> think
>>>> we should not make such assumption as some vhost master may send
>>>> GPA- contiguous pages in different memory region, then the VA after
>>>> mmap may not be contiguous.
>>>
>>> Now we do async vfio mapping at page granularity.
>>> For 4K/2M pages, without this assumption, pages not be merged may
>>> exceed the IOMMU's capability. This limits only use 1G hugepage for DMA
>> device.
>>
>> I was not saying merging should be done, but merging with correct logic.
>>
>> That is to say, the pages that should be merged will still be merged, but make
>> the Logic to handle all cases.
> 
> Sorry, I misunderstood your point. I was to add a supplement,
> PA mode will have lots of pages to map for 4K/2M cases,
> it may exceed IOMMU's capability. But it is rational, that's what I hope
> to explain in the doc.

We are running out of time for v21.11 release. Do you agree to postpone
the fix to new release? It can be backported to v21.11 LTS later.

> Thanks,
> Xuan
> 
>>
>> Thanks,
>> Chenbo
>>
>>>
>>> Thanks,
>>> Xuan
>>>
>>>>
>>>> Maxime, What do you think?
>>>>
>>>> Thanks,
>>>> Chenbo
>>>>
>>>>>
>>>>> Maxime & Yuan, any thought?
>>>>>
>>>>> Thanks,
>>>>> Chenbo
>>>>>
>>>>>> +
>>>>>> +			if (rte_eal_iova_mode() == RTE_IOVA_PA &&
>>>>>> +				guest_phys_addr == last_page-
>>>>> guest_phys_addr +
>>>>>> last_page->size) {
>>>>>> +				last_page->size += size;
>>>>>> +				return 0;
>>>>>> +			}
>>>>>>   		}
>>>>>>   	}
>>>>>>
>>>>>> --
>>>>>> 2.25.1
>