[dpdk-dev] dpdk 16.07, issues with rte_mempool_create and rte_kni_alloc()

[dpdk-dev] dpdk 16.07, issues with rte_mempool_create and rte_kni_alloc()

[Gopakumar Choorakkot Edakkunni]

Hi all,

So 16.07 seems to have an rte_mempool_create() which can allocate items
from multiple memzones (commit d1d914eb), which is awesome and is the
reason I upgraed to 16.07 !!

But .. when I call rte_kni_alloc() with the mempool as parameter, theres a
check in there which says below

    /* KNI currently requires to have only one memory chunk */
    if (mp->nb_mem_chunks != 1)
        goto kni_fail;

And I hit that check because I am allocating a large mempool that happens
to span multiple memzones. I am perfectly fine with allocating a seperate
mempool just for KNI, except that I am unable to find any API or a flag or
some setting which says "create me a mempool that doesnt span memzones",
just so that I can use it with rte_kni_alloc()

Any suggestions will be helpful !

Rgds,
Gopa.

Re: [dpdk-dev] dpdk 16.07, issues with rte_mempool_create and rte_kni_alloc()

[Gopakumar Choorakkot Edakkunni]

Well, for my purpose I just ended up creating a seperate/smaller pool
earlier during bootup to try to guarantee its from one memseg.

But I am assuming that this KNI restriction is something thats "currently"
not fixed and is "fixable" ? Any ideas on what the summary of the reason
for this restriction is - I was gonna check if I can fix that

Rgds,
Gopa.

On Sat, Jul 30, 2016 at 10:39 PM, Gopakumar Choorakkot Edakkunni <
gopakumar.c.e@gmail.com> wrote:

> Hi all,
>
> So 16.07 seems to have an rte_mempool_create() which can allocate items
> from multiple memzones (commit d1d914eb), which is awesome and is the
> reason I upgraed to 16.07 !!
>
> But .. when I call rte_kni_alloc() with the mempool as parameter, theres a
> check in there which says below
>
>     /* KNI currently requires to have only one memory chunk */
>     if (mp->nb_mem_chunks != 1)
>         goto kni_fail;
>
> And I hit that check because I am allocating a large mempool that happens
> to span multiple memzones. I am perfectly fine with allocating a seperate
> mempool just for KNI, except that I am unable to find any API or a flag or
> some setting which says "create me a mempool that doesnt span memzones",
> just so that I can use it with rte_kni_alloc()
>
> Any suggestions will be helpful !
>
> Rgds,
> Gopa.
>
>

Re: [dpdk-dev] dpdk 16.07, issues with rte_mempool_create and rte_kni_alloc()

[Ferruh Yigit]

On 8/1/2016 10:19 PM, Gopakumar Choorakkot Edakkunni wrote:
> Well, for my purpose I just ended up creating a seperate/smaller pool
> earlier during bootup to try to guarantee its from one memseg.
> 
> But I am assuming that this KNI restriction is something thats "currently"
> not fixed and is "fixable" ? 


> Any ideas on what the summary of the reason
> for this restriction is - I was gonna check if I can fix that

KNI expects all mbufs are from a physically continuous memory. This is
because of current address translation implementation.

mbufs allocated in userspace and accessed from both user and kernel
space, so mbuf userspace virtual address needs to be converted into
kernelspace virtual address.

Currently this address translation done by first calculating an offset
between virtual addresses using first filed of mempool, later applying
same offset to all mbufs. This is why all mbufs should be in physically
continuous memory.

I think this address translation can be done in different way which can
remove the restriction, but not sure about the effect of the
performance. I will send a patch for this.

Regards,
ferruh

Re: [dpdk-dev] dpdk 16.07, issues with rte_mempool_create and rte_kni_alloc()

[Ferruh Yigit]

Hi Gopakumar,

On 8/4/2016 5:14 PM, Ferruh Yigit wrote:
> On 8/1/2016 10:19 PM, Gopakumar Choorakkot Edakkunni wrote:
>> Well, for my purpose I just ended up creating a seperate/smaller pool
>> earlier during bootup to try to guarantee its from one memseg.
>>
>> But I am assuming that this KNI restriction is something thats "currently"
>> not fixed and is "fixable" ? 
> 
> 
>> Any ideas on what the summary of the reason
>> for this restriction is - I was gonna check if I can fix that
> 
> KNI expects all mbufs are from a physically continuous memory. This is
> because of current address translation implementation.
> 
> mbufs allocated in userspace and accessed from both user and kernel
> space, so mbuf userspace virtual address needs to be converted into
> kernelspace virtual address.
> 
> Currently this address translation done by first calculating an offset
> between virtual addresses using first filed of mempool, later applying
> same offset to all mbufs. This is why all mbufs should be in physically
> continuous memory.
> 
> I think this address translation can be done in different way which can
> remove the restriction, but not sure about the effect of the
> performance. I will send a patch for this.

I have sent a patch to remove KNI restriction:
http://dpdk.org/dev/patchwork/patch/15171/

Can you please test this patch with a mempool with multiple memzone?
You need to remove following check in KNI manually:
    if (mp->nb_mem_chunks != 1)
        goto kni_fail;

Thanks,
ferruh

Re: [dpdk-dev] dpdk 16.07, issues with rte_mempool_create and rte_kni_alloc()

[Ferruh Yigit]

On 8/10/2016 11:51 AM, Ferruh Yigit wrote:
> Hi Gopakumar,
> 
> On 8/4/2016 5:14 PM, Ferruh Yigit wrote:
>> On 8/1/2016 10:19 PM, Gopakumar Choorakkot Edakkunni wrote:
>>> Well, for my purpose I just ended up creating a seperate/smaller pool
>>> earlier during bootup to try to guarantee its from one memseg.
>>>
>>> But I am assuming that this KNI restriction is something thats "currently"
>>> not fixed and is "fixable" ? 
>>
>>
>>> Any ideas on what the summary of the reason
>>> for this restriction is - I was gonna check if I can fix that
>>
>> KNI expects all mbufs are from a physically continuous memory. This is
>> because of current address translation implementation.
>>
>> mbufs allocated in userspace and accessed from both user and kernel
>> space, so mbuf userspace virtual address needs to be converted into
>> kernelspace virtual address.
>>
>> Currently this address translation done by first calculating an offset
>> between virtual addresses using first filed of mempool, later applying
>> same offset to all mbufs. This is why all mbufs should be in physically
>> continuous memory.
>>
>> I think this address translation can be done in different way which can
>> remove the restriction, but not sure about the effect of the
>> performance. I will send a patch for this.
> 
> I have sent a patch to remove KNI restriction:
> http://dpdk.org/dev/patchwork/patch/15171/
> 
> Can you please test this patch with a mempool with multiple memzone?
> You need to remove following check in KNI manually:
>     if (mp->nb_mem_chunks != 1)
>         goto kni_fail;

Hi Gopakumar,

Off the list.

Any chance to test this?

Thanks,
ferruh

Re: [dpdk-dev] dpdk 16.07, issues with rte_mempool_create and rte_kni_alloc()

[Gopakumar Choorakkot Edakkunni]

Thank you Ferruh, I will give this a spin over the weekend and let you know.

Rgds,
Gopa.

On Thu, Aug 25, 2016 at 6:51 AM, Ferruh Yigit <ferruh.yigit@intel.com>
wrote:

> On 8/10/2016 11:51 AM, Ferruh Yigit wrote:
> > Hi Gopakumar,
> >
> > On 8/4/2016 5:14 PM, Ferruh Yigit wrote:
> >> On 8/1/2016 10:19 PM, Gopakumar Choorakkot Edakkunni wrote:
> >>> Well, for my purpose I just ended up creating a seperate/smaller pool
> >>> earlier during bootup to try to guarantee its from one memseg.
> >>>
> >>> But I am assuming that this KNI restriction is something thats
> "currently"
> >>> not fixed and is "fixable" ?
> >>
> >>
> >>> Any ideas on what the summary of the reason
> >>> for this restriction is - I was gonna check if I can fix that
> >>
> >> KNI expects all mbufs are from a physically continuous memory. This is
> >> because of current address translation implementation.
> >>
> >> mbufs allocated in userspace and accessed from both user and kernel
> >> space, so mbuf userspace virtual address needs to be converted into
> >> kernelspace virtual address.
> >>
> >> Currently this address translation done by first calculating an offset
> >> between virtual addresses using first filed of mempool, later applying
> >> same offset to all mbufs. This is why all mbufs should be in physically
> >> continuous memory.
> >>
> >> I think this address translation can be done in different way which can
> >> remove the restriction, but not sure about the effect of the
> >> performance. I will send a patch for this.
> >
> > I have sent a patch to remove KNI restriction:
> > http://dpdk.org/dev/patchwork/patch/15171/
> >
> > Can you please test this patch with a mempool with multiple memzone?
> > You need to remove following check in KNI manually:
> >     if (mp->nb_mem_chunks != 1)
> >         goto kni_fail;
>
> Hi Gopakumar,
>
> Off the list.
>
> Any chance to test this?
>
> Thanks,
> ferruh
>
>