[Bug 1368] inconsistency in eventdev dev_info and config structs makes some valid configs impossible

[Bruce Richardson]

On Thu, Jan 18, 2024 at 11:18:58AM +0000, bugzilla at dpdk.org wrote:
>    Bug ID [1]1368
>    Summary inconsistency in eventdev dev_info and config structs makes
>    some valid configs impossible
>    Product DPDK
>    Version unspecified
>    Hardware All
>    OS All
>    Status UNCONFIRMED
>    Severity normal
>    Priority Normal
>    Component eventdev
>    Assignee dev at dpdk.org
>    Reporter bruce.richardson at intel.com
>    Target Milestone ---
> 
> In the rte_event_dev_info struct[1], we have the max_event_queues[2] and
> max_single_link_event_port_queue_pairs[3] members. The doxygen docs on the
> latter states: "These ports and queues are not accounted for in
> max_event_ports or max_event_queues."
> 
> This implies that a device which has 8 regular queues and an extra 8
> single-link only queues, would report max_event_queues == 8, and
> max_single_link_event_port_queue_pairs == 8 on return from
> rte_event_dev_info_get() function.
> 
> Those values returned from info_get are generally to be used to guide the
> configuration using rte_event_dev_configure() API, which takes the
> rte_event_dev_config[4] struct. This has two similar fields, in
> nb_event_queues[5] and nb_single_link_event_port_queues[6]. However, a
> problem arises in that the documentation states that nb_event_queues cannot
> be greater than the previously reported max_event_queues (which by itself
> makes sense), but the documentation also states that
> nb_single_link_event_port_queues is a subset of the overall event ports and
> queues, and cannot be greater than the nb_event_queues given in the same
> config structure.
> 
> To illustrate the issue by continuing to use the same example as above,
> suppose an app wants to take that device with 8 regular queues and 8 single
> link ones, and have an app with 2 shared processing queues, e.g. for
> load-balancing packets/events among 8 cores, but also wants to use the 8
> single link queues to allow sending packets/events directly to each core
> without load balancing.  In this 2 + 8 scenario, there is no valid
> dev_config struct settings that will work:
> * making the 8 a subset of the nb_event_queues, means that nb_event_queues
>   is 10, which is greater than max_event_queues and so invalid.
> * keeping them separate, so that nb_event_queues == 2 and
>   nb_single_link_port_queues == 8 violates the constraint that the
>   single_link value cannot exceed the former nb_event_queues value.
> 
> We therefore need to adjust the constraints to make things work. Now we can
> do so, while keeping the single_link value *not included* in the
> total-count in dev_info, but have it *included* in the config struct, but
> such a setup is very confusing for the user. Therefore, I think instead we
> need to correct this by aligning the two structures - either the
> single_link queues are included in the queue/port counts in both structs,
> or they aren't included.
> 

Since I'm doing some clean-up work on rte_eventdev.h doxygen comments, I'm
happy enough to push a patch to help fix this, if we can agree on the
solution.

Of the two possibilities (make both have single-link included in count, or
make both have single-link not-included), I would suggest we go for having
them included, on the basis that that would involve an internal DPDK change
to increase the reported counts in dev_info from any drivers supporting
single link queues, but should *not* involve any changes to end
applications, which would already be specifying the values based on
nb_single_link being a subset of nb_event_queues. On the other hand,
changing semantics of the config struct fields would likely mean changes to
end-apps and so be an ABI/API break.

/Bruce