Re: Issues around packet capture when secondary process is doing rx/tx

[Stephen Hemminger <stephen@networkplumber.org>]

On Mon, 8 Jan 2024 15:13:25 +0000
Konstantin Ananyev <konstantin.ananyev@huawei.com> wrote:

> > I have been looking at a problem reported by Sandesh
> > where packet capture does not work if rx/tx burst is done in secondary process.
> > 
> > The root cause is that existing rx/tx callback model just doesn't work
> > unless the process doing the rx/tx burst calls is the same one that
> > registered the callbacks.
> > 
> > An example sequence would be:
> > 	1. dumpcap (or pdump) as secondary tells pdump in primary to register callback
> > 	2. secondary process calls rx_burst.
> > 	3. rx_burst sees the callback but it has pointer pdump_rx which is not necessarily
> > 	   at same location in primary and secondary process.
> > 	4. indirect function call in secondary to bad location likely causes crash.  
> 
> As I remember, RX/TX callbacks were never intended to work over multiple processes.
> Right now RX/TX callbacks are private for the process, different process simply should not
> see/execute them.
> I.E. it callbacks list is part of 'struct rte_eth_dev' itself, not the rte_eth_dev.data that is shared
> between processes.
> It should be normal, wehn for the same port/queue you will end-up with different list of callbacks
> for different processes.  
> So, unless I am missing something, I don't see how we can end-up with 3) and 4) from above:
> From my understanding secondary process will never see/call primary's callbacks.
> 
> About pdump itself, it was a while when I looked at it last time, but as I remember to start it to work,
> server process has to call rte_pdump_init() which in terns register PDUMP_MP handler.
> I suppose for the secondary process to act as a 'pdump server' it needs to call rte_pdump_init() itself,
> though I am not sure such option is supported right now. 
>  
> > 
> > Some possible workarounds.
> > 	1. Keep callback list per-process: messy, but won't crash. Capture won't work
> >            without other changes. In this primary would register callback, but secondaries
> >            would not use them in rx/tx burst.
> > 
> > 	2. Replace use of rx/tx callback in pdump with change to rte_ethdev to have
> >            a capture flag. (i.e. don't use indirection).  Likely ABI problems.
> >            Basically, ignore the rx/tx callback mechanism. This is my preferred
> > 	   solution.  
> 
> It is not only the capture flag, it is also what to do with the captured packets
> (copy? If yes, then where to? examine? drop?, do something else?).
> It is probably not the best choice to add all these things into ethdev API.
> 
> > 	3. Some fix up mechanism (in EAL mp support?) to have each process fixup
> >            its callback mechanism.  
>  
> Probably the easiest way to fix that - pass to rte_pdump_enable() extra information
> that  would allow it to distinguish on what exact process (local, remote)
> we want to enable pdump functionality. Then it could act accordingly.
> 
> > 
> > 	4. Do something in pdump_init to register the callback in same process context
> > 	   (probably need callbacks to be per-process). Would mean callback is always
> >            on independent of capture being enabled.
> > 
> >         5. Get rid of indirect function call pointer, and replace it by index into
> >            a static table of callback functions. Every process would have same code
> >            (in this case pdump_rx) but at different address.  Requires all callbacks
> >            to be statically defined at build time.  
> 
> Doesn't look like a good approach - it will break many things. 
>  
> > The existing rx/tx callback is not safe id rx/tx burst is called from different process
> > than where callback is registered.  
>  
> 

Have been looking into best way to fix this, and the real answer is not to use
callbacks but instead use a flag per-queue. The natural place to put these in
rte_ethdev_driver. BUT this will mean an ABI breakage, so will have to wait for 24.11
release. Sometimes fixing a design flaw means an ABI change.