[dpdk-users] crypto device 'crypto_aesni_mb' doesn't work in secondary process

[dpdk-users] crypto device 'crypto_aesni_mb' doesn't work in secondary process

[Dirk-Holger Lenz]

when the crypto device of type 'crypto_aesni_mb' is

created in the primary process a secondary process

crashes when writing into the encryption queue.

The dequeue function rte_cryptodev_dequeue_burst()

crashes in flush_mb_mgr() when it tries to access the

structure of function pointers qp->op_fns.

The reason is that this structure is allocated by the

primary process in its memory which is not accessible

in the secondary process (of course also the function

pointers are pointing to code of the primary process).

Re: [dpdk-users] crypto device 'crypto_aesni_mb' doesn't work in secondary process

[De Lara Guarch, Pablo]

Hi Dirk,

> -----Original Message-----
> From: users [mailto:users-bounces@dpdk.org] On Behalf Of Dirk-Holger
> Lenz
> Sent: Friday, August 4, 2017 9:14 AM
> To: users@dpdk.org
> Subject: [dpdk-users] crypto device 'crypto_aesni_mb' doesn't work in
> secondary process
> 
> when the crypto device of type 'crypto_aesni_mb' is
> 
> created in the primary process a secondary process
> 
> crashes when writing into the encryption queue.
> 
> The dequeue function rte_cryptodev_dequeue_burst()
> 
> crashes in flush_mb_mgr() when it tries to access the
> 
> structure of function pointers qp->op_fns.
> 
> The reason is that this structure is allocated by the
> 
> primary process in its memory which is not accessible
> 
> in the secondary process (of course also the function
> 
> pointers are pointing to code of the primary process).
> 

Virtual devices have to be initialized in both processes,
as they cannot be shared between them.

Thanks,
Pablo

Re: [dpdk-users] crypto device 'crypto_aesni_mb' doesn't work in secondary process

[Dirk-Holger Lenz]

Hello Pablo,

thanks for the quick answer.

I think that in general resources like pools and devices

need to be created by the primary process. That's

the reason that we implemented a process (our dpdk

process) which creates the resources on request by

applications which start and stop dynamically and which

are working as secondary processes. These applications

even don't know which type of Ethernet device or crypto

device they get, they simply ask for one of them.

For the Ethernet devices and QAT and even for the openssl

this works fine (QAT and i40evf worked in the past, we are

investigating why they are not doing for us with 17.08rc4).

I hope that it is possible to make that working for

aesni_mb as well.

Best regards

Dirk


On 08/04/2017 11:37 AM, De Lara Guarch, Pablo wrote:
> Hi Dirk,
>
>> -----Original Message-----
>> From: users [mailto:users-bounces@dpdk.org] On Behalf Of Dirk-Holger
>> Lenz
>> Sent: Friday, August 4, 2017 9:14 AM
>> To: users@dpdk.org
>> Subject: [dpdk-users] crypto device 'crypto_aesni_mb' doesn't work in
>> secondary process
>>
>> when the crypto device of type 'crypto_aesni_mb' is
>>
>> created in the primary process a secondary process
>>
>> crashes when writing into the encryption queue.
>>
>> The dequeue function rte_cryptodev_dequeue_burst()
>>
>> crashes in flush_mb_mgr() when it tries to access the
>>
>> structure of function pointers qp->op_fns.
>>
>> The reason is that this structure is allocated by the
>>
>> primary process in its memory which is not accessible
>>
>> in the secondary process (of course also the function
>>
>> pointers are pointing to code of the primary process).
>>
> Virtual devices have to be initialized in both processes,
> as they cannot be shared between them.
>
> Thanks,
> Pablo

Re: [dpdk-users] crypto device 'crypto_aesni_mb' doesn't work in secondary process

[Dirk-Holger Lenz]

Hello Pablo,

perhaps it was a bit misleading what I told, primary

and secondary processes start with the same arguments

to rte_eal_init() but the secondary processes additionally

have the secondary flag (and different settings for the cores).

Best regards

Dirk


On 08/04/2017 11:56 AM, Dirk-Holger Lenz wrote:
> Hello Pablo,
>
> thanks for the quick answer.
>
> I think that in general resources like pools and devices
>
> need to be created by the primary process. That's
>
> the reason that we implemented a process (our dpdk
>
> process) which creates the resources on request by
>
> applications which start and stop dynamically and which
>
> are working as secondary processes. These applications
>
> even don't know which type of Ethernet device or crypto
>
> device they get, they simply ask for one of them.
>
> For the Ethernet devices and QAT and even for the openssl
>
> this works fine (QAT and i40evf worked in the past, we are
>
> investigating why they are not doing for us with 17.08rc4).
>
> I hope that it is possible to make that working for
>
> aesni_mb as well.
>
> Best regards
>
> Dirk
>
>
> On 08/04/2017 11:37 AM, De Lara Guarch, Pablo wrote:
>> Hi Dirk,
>>
>>> -----Original Message-----
>>> From: users [mailto:users-bounces@dpdk.org] On Behalf Of Dirk-Holger
>>> Lenz
>>> Sent: Friday, August 4, 2017 9:14 AM
>>> To: users@dpdk.org
>>> Subject: [dpdk-users] crypto device 'crypto_aesni_mb' doesn't work in
>>> secondary process
>>>
>>> when the crypto device of type 'crypto_aesni_mb' is
>>>
>>> created in the primary process a secondary process
>>>
>>> crashes when writing into the encryption queue.
>>>
>>> The dequeue function rte_cryptodev_dequeue_burst()
>>>
>>> crashes in flush_mb_mgr() when it tries to access the
>>>
>>> structure of function pointers qp->op_fns.
>>>
>>> The reason is that this structure is allocated by the
>>>
>>> primary process in its memory which is not accessible
>>>
>>> in the secondary process (of course also the function
>>>
>>> pointers are pointing to code of the primary process).
>>>
>> Virtual devices have to be initialized in both processes,
>> as they cannot be shared between them.
>>
>> Thanks,
>> Pablo
>