DPDK patches and discussions
 help / color / mirror / Atom feed
* [PATCH v1] bbdev: allow operation type enum for growth
@ 2022-06-13 18:24 Nicolas Chautru
  2022-06-13 18:24 ` Nicolas Chautru
  0 siblings, 1 reply; 7+ messages in thread
From: Nicolas Chautru @ 2022-06-13 18:24 UTC (permalink / raw)
  To: dev, thomas
  Cc: maxime.coquelin, trix, mdr, bruce.richardson, hemant.agrawal,
	david.marchand, stephen, Nicolas Chautru

Hi Thomas, 

I would like to get your view on this topic and best recommendation moving forward.

This is related to the general intent to remove using MAX value for enums. There is consensus that we should avoid this for a while notably for future-proofed ABI concerns https://patches.dpdk.org/project/dpdk/patch/20200130142003.2645765-1-ferruh.yigit@intel.com/.
But still there is arguably not yet an explicit best recommendation to handle this especially when we actualy need to expose array whose index is such an enum.
As a specific example here I am refering to RTE_BBDEV_OP_TYPE_COUNT in enum rte_bbdev_op_type which is being extended for new operation type being support in bbdev (such as https://patches.dpdk.org/project/dpdk/patch/1646956157-245769-2-git-send-email-nicolas.chautru@intel.com/ adding new FFT operation)

There is also the intent to be able to expose information for each operation type through the bbdev api such as dynamically configured queues information per such operation type https://patches.dpdk.org/project/dpdk/patch/1646785355-168133-2-git-send-email-nicolas.chautru@intel.com/

Basically we are considering best way to accomodate for this, notably based on discussions with Ray Kinsella and Bruce Richardson, to handle such a case moving forward: specifically for the example with RTE_BBDEV_OP_TYPE_COUNT and also more generally.

One possible option is captured in that patchset and is basically based on the simple principle to allow for growth and prevent ABI breakage. Ie. the last value of the enum is set with a higher value than required so that to allow insertion of new enum outside of the major ABI versions.
In that case the RTE_BBDEV_OP_TYPE_COUNT is still present and can be exposed and used while still allowing for addition thanks to the implicit padding-like room. As an alternate variant, instead of using that last enum value, that extended size could be exposed as an #define outside of the enum but would be fundamentally the same (public).

Another option would be to avoid array alltogether and use each time this a new dedicated API function (operation type enum being an input argument instead of an index to an array in an existing structure so that to get access to structure related to a given operation type enum) but that is arguably not well scalable within DPDK to use such a scheme for each enums and keep an uncluttered and clean API. In that very example that would be very odd indeed not to get this simply from info_get().

Some pros and cons, arguably the simple option in that patchset is a valid compromise option and a step in the right direction but we would like to know your view wrt best recommendation, or any other thought. 

Note: Such a change is aimed for 22.11.

Thanks and regards, 
Nic


Nicolas Chautru (1):
  bbdev: allow operation type enum for growth

 lib/bbdev/rte_bbdev.c    | 5 ++++-
 lib/bbdev/rte_bbdev_op.h | 2 +-
 2 files changed, 5 insertions(+), 2 deletions(-)

-- 
1.8.3.1


^ permalink raw reply	[flat|nested] 7+ messages in thread
end of thread, other threads:[~2022-06-23 16:13 UTC | newest]

Thread overview: 7+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-06-13 18:24 [PATCH v1] bbdev: allow operation type enum for growth Nicolas Chautru
2022-06-13 18:24 ` Nicolas Chautru
2022-06-13 19:48   ` Stephen Hemminger
2022-06-13 20:19     ` Chautru, Nicolas
2022-06-17  8:21   ` Thomas Monjalon
2022-06-17 16:12     ` Chautru, Nicolas
2022-06-23 16:09     ` Ray Kinsella

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).