[dpdk-dev] [PATCH v2] librte_ether: release memory in uninit function.

[Thomas Monjalon]

2015-07-08 09:49, Iremonger, Bernard:
> From: Tetsuya Mukawa [mailto:mukawa at igel.co.jp]
> > On 2015/07/07 19:53, Iremonger, Bernard wrote:
> > > From: Tetsuya Mukawa [mailto:mukawa at igel.co.jp]
> > >> On 2015/07/06 20:35, Qiu, Michael wrote:
> > >>> Hi, all
> > >>>
> > >>> As we has gap on the memory release action to be done in which step,
> > >>> I appreciate all your comments on this patch.
> > >>>
> > >>> Currently, the correct quit sequence for testpmd is stop() --->
> > >>> port_stop() --> port_close() --> quit(). This will lead lots of
> > >>> memory not released by default, like queues.
> > > There is also the scenario (outside of testpmd) where an application can call
> > rte_eth_dev_close() or rte_eth_dev_detach() which calls
> > rte_eth_dev_uninit() directly from an application.
> > > In these cases the memory allocated in the EAL layer should be released in
> > both rte_eth_dev_close() and rte_eth_dev_detach().
> > 
> > Hi Bernard,
> > 
> > All DPDK applications that uses hotpluggingmust call rte_eth_dev_stop() and
> > rte_eth_dev_close()APIs before detaching ports.
> > (This is described in DPDK documentation) Could we ignore applications that
> > only calls rte_eth_dev_detach()?
> > 
> > > This patch is intended to address the rte_eth_dev_detach() case only
> > (hotplug case).
> > > Perhaps there should be a separate patch to address the
> > rte_eth_dev_close() case.
> > 
> > We all needs to have consensus about how to implement finalization code in
> > close() and uninit().
> > Probably one of options will be implementing finalization code in
> > close() as much as possible.
> > 
> > Tetsuya,
> 
> Hi Tetsuya,
> Testpmd enforces the dev_stop(), dev_close() and dev_detach() sequence.
> There is no reason why an application cannot call dev_detach() directly. 
> 
> During internal review of PMD dev_uninit()  functions it was decided to ensure that this sequence was adhered to by calling dev_close() which calls dev_stop() from the dev_uninit() function.
> In the PMD hotplug patches the following calling sequence is implemented:
> dev_uninit() calls dev_close() calls dev_stop().
> At present most of the finalization code is implemented in dev_close() and dev_stop().
> The dev_uninit() functions implement what is left of the finalization code.

It appears that the API is not defined.
None of these assumptions is written in rte_ethdev.h.
Please continue the discussion by submitting a patch fixing the doxygen
comments of these functions.
How an application developper is supposed to use these functions without
having a clear explanation of their roles in doxygen?