On Wednesday 25 January 2017 06:07 PM, Neil Horman wrote:

On Tue, Jan 24, 2017 at 02:39:56PM +0000, Shreyansh Jain wrote:

quoted

Hello,

We are facing a peculiar problem with respect to symbol namespace in DPDK. I
think Ferruh and Thomas would have fair idea about it as they have already
reviewed and commented on it. I was hoping to get some input to take it
forward from here.

Brief Intro to DPAA2 Architecture:

This is brief about NXP's DPAA2 PMD to start with:
(A lot more information is available at [1])



                 +-------------------------------+
                 |         Application           |
                 +----.--------------------.-----+
                      |                    |
                 +----'------+       +-----'-----+
    drivers/---->|   DPIO    |       |   DPIO    |<---drivers/bus/fslmc
     bus/fslmc   +----.------+       +------.----+
                      |                     |
                 +----/-||--------------||--/----+
                 |   Queue/Buffer Manager        |<--- drivers/common/dpaa2
                 +----\-||--------------||--\----+      qbman
                      |                     |
                 +----'------+       +------'----+
    drivers/ --->|   DPNI    |       |   DPSEC   |<---drivers/cyrpto
     net/dpaa2   +----|------+       +-----|-----+     dpaa2_sec
                      |                    |
                 +----|------+      +------|-----+       +----------------+
                 | PHY H/W   |      |   SEC H/W  |      .> FSL MC BUS     |
                 +-----------+      +------------+     / +----------------+
                                                      drivers/bus/fslmc


If we consider the above layout, drivers/crypto/dpaa_sec (NXP's DPAA2 Crypto
PMD, already available on ML [2]), and drivers/net/dpaa2 (NXP's DPAA2 PMD) are
using a common code (drivers/common/dpaa2/qbman).

QBMAN (drivers/common/dpaa2/qbman) is essentially a Queue and Buffer Manager
set of APIs which allow DPIO (Data Path IO interfaces) to communicate with the
Hardware through queues (and buffers).

At the scan time, FSLMC bus is scanned and all devices (Phy or Sec) are
identified and added to a list. For each such device, appropriate I/O portals
are opened which are essentially gateway between user-space and DP* devices
using the hardware queues/buffers (qbman)

Problem:

You might have noticed that we have exposed a lot of symbols from
drivers/common and drivers/bus for drivers/net and drivers/crypto. All these
symbols are not rte_* as what has been suggested for exported symbols.

Review comments have been received for renaming these to make them rte_* or
_rte_* prefixed.

Just as a side note, these symbols are being exposed _internally_ within
drivers/* area.

There are (3) possible solutions we have:

1/ Rename all the symbols:
  - This is a difficult option for us. Renaming means breaking our linkage
    with existing code (Linux Kernel upstream candidate as well as internal
    repository).
  - Changing it means maintaining this change set internally/independently
    which is not a feasible long term solution.

2/ Merge all the libraries together:
  - In the initial RFC days, there were review comments which suggested that
    we should break the PMD into common libraries and place it in drivers/*
    parallel folders.
  - This is precisely the reason we are facing the situation.
  - Another possibility is to start duplicating the code for common. But, this
    too has a technical limitation for us as some data structures are shared
    across net and crypto and it is not possible to have multiple instances of
    those.
  - One more offshoot option could have been to keep the library external
    of the DPDK framework (external location and linked on demand basis,
    manually). We don't prefer this as this will make it difficult for any user
    to use DPAA2 easily.

3/ Finding a way to keep symbols internal to drivers/* independent of rte_*
   prefix:
   - For example, allowing symbols to be exposed limited to drivers/* area
     and not allowing them to be available across lib/* (not sure how, though!)

   <This is where I need you help - is there some suggestion or comments which
    can help us arrive to a solution?>

   My argument for this:
  - With new bus infra in place, there would be more drivers being contributed.
    It also means that there would be PMDs having their own code and symbol
    models. It would be difficult to ask all of them to mandatorily adhere
    to a naming scheme.
    This argument bodes well for lib/* because that is core (libraries) which
    should be controlled for uniformity and performance.

[1] https://www.kernel.org/doc/readme/drivers-staging-fsl-mc-README.txt
[2] http://dpdk.org/ml/archives/dev/2017-January/054251.html

So, Option 3 seems pretty easy, just use symbol aliasing.  Make all your
exported symbols static, and use the MAP_SATIC_SYMBOL macro (or use your own
inline asm if you want), to create rte_ aliases for them all, add the rte_*
variants to your version map and it should all work out.  You get to keep your
naming in tact, you can create some macro ifdeffery to make your names static or
not depending on your build environment (upstream linux kernel vs. dpdk, vs
something else), and just use the aliases when it suit you (like for dpdk naming
conventions)

Thanks. Sounds interesting, though something I will need to
investigate. I will experiment on this.

Neil