Re: [dpdk-dev] [PATCH v2] eal/linux: do not create user mem map repeatedly when it exists

["Burakov, Anatoly" <anatoly.burakov@intel.com>]

On 05-Aug-20 1:58 PM, wangyunjian wrote:
>> -----Original Message-----
>> From: Burakov, Anatoly [mailto:anatoly.burakov@intel.com]
>> Sent: Friday, July 31, 2020 7:55 PM
>> To: wangyunjian <wangyunjian@huawei.com>; dev@dpdk.org;
>> david.marchand@redhat.com
>> Cc: Lilijun (Jerry) <jerry.lilijun@huawei.com>; xudingke
>> <xudingke@huawei.com>; stable@dpdk.org
>> Subject: Re: [dpdk-dev] [PATCH v2] eal/linux: do not create user mem map
>> repeatedly when it exists
>>
>> On 30-Jul-20 2:16 PM, wangyunjian wrote:
>>>> -----Original Message-----
>>>> From: Burakov, Anatoly [mailto:anatoly.burakov@intel.com]
>>>> Sent: Monday, July 27, 2020 5:24 PM
>>>> To: wangyunjian <wangyunjian@huawei.com>; dev@dpdk.org;
>>>> david.marchand@redhat.com
>>>> Cc: Lilijun (Jerry) <jerry.lilijun@huawei.com>; xudingke
>>>> <xudingke@huawei.com>; stable@dpdk.org
>>>> Subject: Re: [dpdk-dev] [PATCH v2] eal/linux: do not create user mem map
>>>> repeatedly when it exists
>>>>
>>>> On 25-Jul-20 10:59 AM, wangyunjian wrote:
>>>>>> -----Original Message-----
>>>>>> From: Burakov, Anatoly [mailto:anatoly.burakov@intel.com]
>>>>>> Sent: Friday, July 24, 2020 9:25 PM
>>>>>> To: wangyunjian <wangyunjian@huawei.com>; dev@dpdk.org;
>>>>>> david.marchand@redhat.com
>>>>>> Cc: Lilijun (Jerry) <jerry.lilijun@huawei.com>; xudingke
>>>>>> <xudingke@huawei.com>; stable@dpdk.org
>>>>>> Subject: Re: [dpdk-dev] [PATCH v2] eal/linux: do not create user mem
>>>>>> map repeatedly when it exists
>>>>>>
>>>>>> On 23-Jul-20 3:48 PM, wangyunjian wrote:
>>>>>>> From: Yunjian Wang <wangyunjian@huawei.com>
>>>>>>>
>>>>>>> Currently, we will create new user mem map entry for the same memory
>>>>>>> segment, but in fact it has already been added to the user mem maps.
>>>>>>> It's not necessary to create it twice.
>>>>>>>
>>>>>>> To resolve the issue, add support to remove the same entry in the
>>>>>>> function compact_user_maps().
>>>>>>>
>>>>>>> Fixes: 0cbce3a167f1 ("vfio: skip DMA map failure if already mapped")
>>>>>>> Cc: stable@dpdk.org
>>>>>>>
>>>>>>> Signed-off-by: Yunjian Wang <wangyunjian@huawei.com>
>>>>>>> ---
>>>>>>> v2:
>>>>>>> * Remove the same entry in the function compact_user_maps()
>>>>>>> ---
>>>>>>>      lib/librte_eal/linux/eal_vfio.c | 5 +++++
>>>>>>>      1 file changed, 5 insertions(+)
>>>>>>>
>>>>>>> diff --git a/lib/librte_eal/linux/eal_vfio.c
>>>>>>> b/lib/librte_eal/linux/eal_vfio.c index abb12a354..df99307b7 100644
>>>>>>> --- a/lib/librte_eal/linux/eal_vfio.c
>>>>>>> +++ b/lib/librte_eal/linux/eal_vfio.c
>>>>>>> @@ -167,6 +167,10 @@ adjust_map(struct user_mem_map *src,
>> struct
>>>>>> user_mem_map *end,
>>>>>>>      static int
>>>>>>>      merge_map(struct user_mem_map *left, struct user_mem_map
>>>> *right)
>>>>>>>      {
>>>>>>> +	/* merge the same maps into one */
>>>>>>> +	if (memcmp(left, right, sizeof(struct user_mem_map)) == 0)
>>>>>>> +		goto out;
>>>>>>> +
>>>>>>
>>>>>> merge_map looks for adjacent maps only, but does not handle maps that
>>>>>> are wholly contained within one another ("the same map" also matches
>>>>>> this definition). wouldn't it be better to check for that instead of
>>>>>> *just* handling identical maps?
>>>>>
>>>>> What about using the initial implementation?
>>>>> We don't create new user mem map entry for the same memory segment.
>>>>
>>>> I don't like this implementation because it relies on particulars of how VFIO
>>>> mapping work without explicitly specifying them. I.e. it's prone to breaking
>>>> when changing code. That's not even mentioning that we have no
>> guarantees
>>>> on kernel behavior in that particular case being identical on all supported
>>>> platforms.
>>>>
>>>> I would honestly prefer an explicit compaction over implicit one.
>>>
>>> What about this implementation?
>>
>> Again, this works, but i feel like specializing it to just merge the
>> exact same maps is missing an opportunity to provide a more general
>> solution that merges same *and* subset maps.
> 
> Currently, the problem that I encounter is that a container has many
> devices and the application will map the same memory many times.
> The kernel driver returns EEXIST as long as there are overlapping memory
> areas. Therefore, the application needs to ensure that the memory blocks
> of the DMA do not overlap. Otherwise, it will not work normally.
> 
> Could you offer me some ideas or advise to fix it？
> 

It sounds like your approach is better if that is indeed the case.

-- 
Thanks,
Anatoly