Re: [PATCH v3 00/10] Rework vdev probing to use rte_bus infrastructure

-----Original Message-----
From: jblunck@gmail.com [mailto:jblunck@gmail.com] On Behalf Of Jan Blunck
Sent: Tuesday, February 28, 2017 2:49 PM
To: Shreyansh Jain <redacted>
Cc: dev <redacted>; David Marchand <redacted>; Ferruh
Yigit [off-list ref]
Subject: Re: [PATCH v3 00/10] Rework vdev probing to use rte_bus
infrastructure

On Tue, Feb 28, 2017 at 9:48 AM, Shreyansh Jain [off-list ref]
wrote:

quoted

On Monday 27 February 2017 06:39 PM, Jan Blunck wrote:

quoted

On Sat, Feb 25, 2017 at 11:28 AM, Jan Blunck [off-list ref]
wrote:

quoted

With the rte_bus infrastructure present in 17.02 it is possible to
refactor
the virtual device probing into a bus. This series also introduces the
rte_vdev_device to better keep track of devices.

This patchset depends on:
http://dpdk.org/dev/patchwork/patch/20416/
http://dpdk.org/dev/patchwork/patch/20417/

Changes since version 2:
 * implicit bus registration through rte_eal_vdrv_register()

On a second thought I don't think that this is correct though since it
opens up the possibility of racing an alternative "virtual" bus. I
don't think that this is a good thing though. I'll fix this in v4.

Thoughts?

I prefer the RTE_REGISTER* way.
The issue of duplicate bus remains whether we use the macro or the
implicit way.

If you use RTE_*, do you think that duplicate registration issue
is worth fixing?
(It would mean rte_bus_register to return error which the caller
would then need to handle).

I don't think that there is a good way to handle this if we use the
library constructors. It is better to fail the registration of the bus
and log the error via RTE_LOG(). That introduces a dependency on the
initialization of the logging system to be available before the
library constructors run (before main()).

Agree. Log dependency is the tricky part. Log subsystem is in turn
dependent on users input for level.