[dpdk-dev] [PATCH] ethdev: fix switching domain allocation

[Ferruh Yigit]

On 1/15/2020 8:50 AM, Slava Ovsiienko wrote:
>> -----Original Message-----
>> From: Ferruh Yigit <ferruh.yigit at intel.com>
>> Sent: Tuesday, January 14, 2020 17:32
>> To: Slava Ovsiienko <viacheslavo at mellanox.com>; dev at dpdk.org
>> Cc: Thomas Monjalon <thomas at monjalon.net>; declan.doherty at intel.com;
>> stable at dpdk.org
>> Subject: Re: [PATCH] ethdev: fix switching domain allocation
>>
>> On 12/19/2019 12:47 PM, Viacheslav Ovsiienko wrote:
>>> The maximum amount of unique switching domain is supposed to be equal
>>> to RTE_MAX_ETHPORTS. The current implementation allows to allocate
>>> only RTE_MAX_ETHPORTS-1 domains.
>>>
>>> Fixes: ce9250406323 ("ethdev: add switch domain allocator")
>>> Cc: stable at dpdk.org
>>>
>>> Signed-off-by: Viacheslav Ovsiienko <viacheslavo at mellanox.com>
>>> ---
>>>  lib/librte_ethdev/rte_ethdev.c | 13 +++++++------
>>>  1 file changed, 7 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/lib/librte_ethdev/rte_ethdev.c
>>> b/lib/librte_ethdev/rte_ethdev.c index 6e9cb24..4c2312c 100644
>>> --- a/lib/librte_ethdev/rte_ethdev.c
>>> +++ b/lib/librte_ethdev/rte_ethdev.c
>>> @@ -5065,10 +5065,10 @@ enum rte_eth_switch_domain_state {
>>>  	*domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
>>>
>>>  	for (i = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID + 1;
>>> -		i < RTE_MAX_ETHPORTS; i++) {
>>> -		if (rte_eth_switch_domains[i].state ==
>>> +		i <= RTE_MAX_ETHPORTS; i++) {
>>> +		if (rte_eth_switch_domains[i - 1].state ==
>>>  			RTE_ETH_SWITCH_DOMAIN_UNUSED) {
>>> -			rte_eth_switch_domains[i].state =
>>> +			rte_eth_switch_domains[i - 1].state =
>>>  				RTE_ETH_SWITCH_DOMAIN_ALLOCATED;
>>>  			*domain_id = i;
>>
>> I would keep the indexes same but change how to set the 'domain_id' to
>> "*domain_id = i + 1;", that makes logic simpler.
> Agree.
> 
>> Would it be simpler if the invalid domain id value used as UINT16_MAX
>> instead of '0'? This enables using 'domain_id' as index and prevent this error
>> prone indexing.
> 
> My concern was not to change the existing RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID
> definition, which currently is zero. Currently, AFAIK, the switch feature is supported by mlx5
> only, other PMDs do not bother to initialize the rte_eth_dev_info-> switch_info structure
> (no one sets RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID to domain_id field for now).
> So, changing the RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID from zero might cause
> wrong switch capability reporting from PMDs.

I think PMDs shouldn't have to initialize the values that they don't use/care,
otherwise it will be very error prone. Can this be handled in the API level?

Like in 'rte_eth_dev_info_get()', after "memset(dev_info, ..)" set 'switch_info'
as INVALID before dev_ops called. PMD can overwrite this if they want, otherwise
it will stay invalid and I think this is safer.

> 
>>
>> And I think it makes sense to start the loop with "i = 0", instead of
>> 'RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID', you are walking through the
>> port list, why to involve the 'RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID'
>> here.
> I do not know why it was implemented in this way 😊
> I just was trying to introduce the minimalistic fix. I'll think how to extend my fix a bit.
> 
>>
>>>  			return 0;
>>> @@ -5082,14 +5082,15 @@ enum rte_eth_switch_domain_state {
>>> rte_eth_switch_domain_free(uint16_t domain_id)  {
>>>  	if (domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID ||
>>> -		domain_id >= RTE_MAX_ETHPORTS)
>>> +		domain_id > RTE_MAX_ETHPORTS)
>>>  		return -EINVAL;
>>>
>>> -	if (rte_eth_switch_domains[domain_id].state !=
>>> +	if (rte_eth_switch_domains[domain_id - 1].state !=
>>>  		RTE_ETH_SWITCH_DOMAIN_ALLOCATED)
>>>  		return -EINVAL;
>>>
>>> -	rte_eth_switch_domains[domain_id].state =
>> RTE_ETH_SWITCH_DOMAIN_UNUSED;
>>> +	rte_eth_switch_domains[domain_id - 1].state =
>>> +		RTE_ETH_SWITCH_DOMAIN_UNUSED;
>>>
>>>  	return 0;
>>>  }
>>>
> With best regards,
> Slava
>