Re: [dpdk-dev] [dpdk-stable] [PATCH 1/3] net/enic: restrict several handlers to primary process

[Ferruh Yigit <ferruh.yigit@intel.com>]

On 10/9/2019 9:48 AM, Hyong Youb Kim (hyonkim) wrote:
>> -----Original Message-----
>> From: Ferruh Yigit <ferruh.yigit@intel.com>
>> Sent: Wednesday, October 9, 2019 5:02 PM
>> To: Hyong Youb Kim (hyonkim) <hyonkim@cisco.com>
>> Cc: dev@dpdk.org; John Daley (johndale) <johndale@cisco.com>; Dirk-
>> Holger Lenz <dirk.lenz@ng4t.com>; stable@dpdk.org
>> Subject: Re: [PATCH 1/3] net/enic: restrict several handlers to primary
>> process
>>
>> On 9/6/2019 7:50 AM, Hyong Youb Kim wrote:
>>> These eth_dev_ops handlers should run only in the primary process.
>>> - filter_ctrl
>>> - reta_update
>>> - rss_hash_update
>>> - set_mc_addr_list
>>> - udp_tunnel_port_add
>>> - udp_tunnel_port_del
>>>
>>> Fixes: c2fec27b5cb0 ("net/enic: allow to change RSS settings")
>>> Fixes: 8d496995346c ("net/enic: support multicast filtering")
>>> Fixes: 8a4efd17410c ("net/enic: add handlers to add/delete vxlan port
>> number")
>>> Cc: stable@dpdk.org
>>>
>>> Reported-by: Dirk-Holger Lenz <dirk.lenz@ng4t.com>
>>> Signed-off-by: Hyong Youb Kim <hyonkim@cisco.com>
>>> Tested-by: Dirk-Holger Lenz <dirk.lenz@ng4t.com>
>>> Reviewed-by: John Daley <johndale@cisco.com>
>>> ---
>>>  drivers/net/enic/enic_ethdev.c | 12 ++++++++++++
>>>  1 file changed, 12 insertions(+)
>>>
>>> diff --git a/drivers/net/enic/enic_ethdev.c
>> b/drivers/net/enic/enic_ethdev.c
>>> index 06dc67122..85d785e62 100644
>>> --- a/drivers/net/enic/enic_ethdev.c
>>> +++ b/drivers/net/enic/enic_ethdev.c
>>> @@ -129,6 +129,8 @@ enicpmd_dev_filter_ctrl(struct rte_eth_dev *dev,
>>>  {
>>>  	int ret = 0;
>>>
>>> +	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
>>> +		return -E_RTE_SECONDARY;
>>>  	ENICPMD_FUNC_TRACE();
>>>
>>>  	switch (filter_type) {
>>
>> I remember we have similar talk with John in the past about these secondary
>> application checks in ethdev_ops.
>>
>> I would like to understand why these checks required only in enic, can you
>> please describe your use case?
>> Is there any reason secondary application can't change configuration of the
>> device, or are you updating your driver to work with specific application?
> 
> Hi,
> 
> No fundamental reasons why secondary processes cannot run these
> handlers. These checks are to make it clear that it is not safe to do
> so at the moment. It is a software limitation.
> 
> The firmware API (devcmd) we use to configure NIC settings assumes one
> user executing one command at a time. And, many of the handlers in the
> driver also assume primary process. The firmware itself has
> enough checks to prevent concurrent devcmd attempts from corrupting
> its internal state. But, host processes can get confused. For example,
> process A gets process B's results, or overwrites B's devcmd, etc.
> 
> I believe these issues are all fixable in the driver. We could use
> locks in shared memory to serialize devcmd (though risky), fix
> handlers that assume primary process, and so on. It is a to-do item for
> this driver and would require its own patch series (e.g. allow
> secondary processes to run X, Y, Z safely)..

What you have described is valid concern for all drivers, that synchronization
has been pushed to the application level.

I don't see the point of just putting protection to only one driver.

And as a alternative, what do you think about having a check in the prob for the
secondary process and assign a subset of the ethdev_ops in that case? This makes
 more clear what is supported in the secondary process, and it prevents putting
secondary process checks everywhere.