Re: [dpdk-dev] [RFC] af_packet: support port hotplug

[Tetsuya Mukawa <mukawa@igel.co.jp>]

On 2015/03/13 19:14, Iremonger, Bernard wrote:
>
>> -----Original Message-----
>> From: Tetsuya Mukawa [mailto:mukawa@igel.co.jp]
>> Sent: Friday, March 13, 2015 12:11 AM
>> To: Iremonger, Bernard
>> Cc: John W. Linville; dev@dpdk.org
>> Subject: Re: [dpdk-dev] [RFC] af_packet: support port hotplug
>>
>> 2015-03-13 2:05 GMT+09:00 Iremonger, Bernard <bernard.iremonger@intel.com>:
>>>
>>>> -----Original Message-----
>>>> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of John W. Linville
>>>> Sent: Tuesday, March 10, 2015 6:36 PM
>>>> To: dev@dpdk.org
>>>> Subject: [dpdk-dev] [RFC] af_packet: support port hotplug
>>>>
>>>> This patch adds finalization code to free resources allocated by the
>>>> PMD.  This is based on earlier patches for other PMDs by Tetsuya
>>>> Mukawa <mukawa@igel.co.jp>, with corrections related to data-
>>>>> name.
>>>> Signed-off-by: John W. Linville <linville@tuxdriver.com>
>>>> Cc: Tetsuya Mukawa <mukawa@igel.co.jp>
>>>> ---
>>>>  lib/librte_pmd_af_packet/rte_eth_af_packet.c | 56
>>>> ++++++++++++++++++++++++++++
>>>>  1 file changed, 56 insertions(+)
>>>>
>>>> diff --git a/lib/librte_pmd_af_packet/rte_eth_af_packet.c
>>>> b/lib/librte_pmd_af_packet/rte_eth_af_packet.c
>>>> index 80e9bdf7f819..57998ab19c96 100644
>>>> --- a/lib/librte_pmd_af_packet/rte_eth_af_packet.c
>>>> +++ b/lib/librte_pmd_af_packet/rte_eth_af_packet.c
>>>> @@ -399,6 +399,13 @@ static struct eth_dev_ops ops = {
>>>>       .stats_reset = eth_stats_reset,  };
>>>>
>>>> +static struct eth_driver rte_af_packet_pmd = {
>>>> +     .pci_drv = {
>>>> +             .name = "rte_af_packet_pmd",
>>>> +             .drv_flags = RTE_PCI_DRV_DETACHABLE,
>>>> +     },
>>>> +};
>>>> +
>>>>  /*
>>>>   * Opens an AF_PACKET socket
>>>>   */
>>>> @@ -653,6 +660,10 @@ rte_pmd_init_internals(const char *name,
>>>>       if (*eth_dev == NULL)
>>>>               goto error;
>>>>
>>>> +     /* check length of device name */
>>>> +     if ((strlen((*eth_dev)->data->name) + 1) > sizeof(data->name))
>>>> +             goto error;
>>>> +
>>>>       /*
>>>>        * now put it all together
>>>>        * - store queue data in internals, @@ -669,12 +680,14 @@
>>>> rte_pmd_init_internals(const char *name,
>>>>       data->nb_tx_queues = (uint16_t)nb_queues;
>>>>       data->dev_link = pmd_link;
>>>>       data->mac_addrs = &(*internals)->eth_addr;
>>>> +     strncpy(data->name, (*eth_dev)->data->name,
>>>> +strlen((*eth_dev)->data->name));
>>>>
>>>>       pci_dev->numa_node = numa_node;
>>>>
>>>>       (*eth_dev)->data = data;
>>>>       (*eth_dev)->dev_ops = &ops;
>>>>       (*eth_dev)->pci_dev = pci_dev;
>>>> +     (*eth_dev)->driver = &rte_af_packet_pmd;
>>>>
>>>>       return 0;
>>>>
>>>> @@ -835,10 +848,53 @@ rte_pmd_af_packet_devinit(const char *name, const char *params)
>>>>       return 0;
>>>>  }
>>>>
>>>> +static int
>>>> +rte_pmd_af_packet_devuninit(const char *name) {
>>>> +     struct rte_eth_dev *eth_dev = NULL;
>>>> +     struct pmd_internals *internals;
>>>> +     struct tpacket_req req;
>>>> +
>>>> +     unsigned q;
>>>> +
>>>> +     RTE_LOG(INFO, PMD, "Closing AF_PACKET ethdev on numa socket %u\n",
>>>> +                     rte_socket_id());
>>>> +
>>>> +     if (name == NULL)
>>>> +             return -1;
>>> Hi  Tetsuya, John,
>>>
>>> Before detaching a port, the port must be stopped and closed.
>>> The stop and close are only allowed for RTE_PROC_PRIMARY.
>>> Should there be a check for process_type here?
>>>
>>> if (rte_eal_process_type() != RTE_PROC_PRIMARY)
>>>         return -EPERM;
>>>
>>> Regards,
>>>
>>> Bernard
>>>
>> Hi Bernard,
>>
>> I agree with stop() and close() are only called by primary process, but it may not need to add like
>> above.
>> Could you please check rte_ethdev.c?
>>
>> - struct rte_eth_dev_data *rte_eth_dev_data; This array is shared between processes.
>> So we need to initialize of finalize carefully like you said.
>>
>> - struct rte_eth_dev rte_eth_devices[]
>> This array is per process.
>> And 'data' variable of this structure indicates a pointer of rte_eth_dev_data.
>>
>> All PMDs for physical NIC allocates like above when PMDs are initialized.
>> (Even when a process is secondary, initialization function of PMDs will be called) But virtual device
>> PMDs allocate rte_eth_dev_data and overwrite 'data'
>> variable of rte_eth_devices while initialization.
>>
>> As a result, primary and secondary process has their own 'rte_eth_dev_data' for a virtual device.
>> So I guess all processes need to free it not to leak memory.
>>
>> Thanks,
>> Tetsuya
>>
> Hi Tetsuya,
>
> In rte_ethdev.c   both rte_eth_dev_stop() and rte_eth_dev_close()  use the macro  PROC_PRIMARY_OR_RET().
> So for secondary processes  both functions return without doing anything. 
> Maybe this check should be added to rte_eth_dev_attach() and rte_eth_dev_detach() ?
>
> For the Physical/Virtual  Functions of the NIC  a lot of the finalization is done in the  dev->dev_ops->dev_stop() and
> dev->dev_ops->dev_close() functions. To complete the finalization the dev_uninit() function is called, this should probably do nothing for secondary processes  as the dev_stop() and dev_close() functions will not have been executed.

Hi Bernard,

Sorry for my English.
I meant 'virtual device PMD' as PMDs like pcap or af_packet PMDs.
Not a PMDs for virtual functions on NIC.

For PMDs like a pcap and af_packet PMDs, all data structures are
allocated per processes.
(Actually I guess nothing is shared between primary and secondary
processes, because rte_eth_dev_data is overwritten by each processes.)
So we need to free per process data when detach() is called.

> For the Physical/Virtual  Functions of the NIC  the dev_init() is called for both primary and secondary processes, however a subset of the function only is executed for secondary processes.

Because of above, we probably not be able to add PROC_PRIMARY_OR_RET()
to rte_eth_dev_detach().
But I agree we should not call rte_eth_dev_detach() for secondary
process, if PMDs are like e1000 or ixgbe PMD.

To work like above, how about changing drv_flags dynamically in close()
callback?
For example, when primary process calls rte_eth_dev_close(), a callback
of PMD will be called.
(In the case of e1000 PMD, eth_em_close() is the callback.)

At that time, specify RTE_PCI_DRV_DETACHABLE flag to drv_flag in the
callback.
It means if primary process hasn't called close() yet,
rte_eth_dev_detach() will do nothing and return error.
How about doing like above?

Regards,
Tetsuya

> Regards,
>
> Bernard.
>  
>
>
>
>
>
>