Re: [dpdk-dev] [PATCH v2 0/4] Link Bonding Library

[Neil Horman <nhorman@tuxdriver.com>]

On Fri, Jun 13, 2014 at 02:56:47PM +0000, Doherty, Declan wrote:
> > -----Original Message-----
> > From: Neil Horman [mailto:nhorman@tuxdriver.com]
> > Sent: Friday, June 6, 2014 3:54 PM
> > To: Doherty, Declan
> > Cc: dev@dpdk.org
> > Subject: Re: [dpdk-dev] [PATCH v2 0/4] Link Bonding Library
> > 
> > On Fri, Jun 06, 2014 at 08:23:31AM +0000, Doherty, Declan wrote:
> > > > -----Original Message-----
> > > > From: Neil Horman [mailto:nhorman@tuxdriver.com]
> > > > Sent: Thursday, May 29, 2014 12:34 PM
> > > > To: Doherty, Declan
> > > > Cc: dev@dpdk.org
> > > > Subject: Re: [dpdk-dev] [PATCH 0/4] Link Bonding Library
> > > >
> > > > On Thu, May 29, 2014 at 10:33:00AM +0000, Doherty, Declan wrote:
> > > > > -----Original Message-----
> > > > > > From: Neil Horman [mailto:nhorman@tuxdriver.com]
> > > > > > Sent: Wednesday, May 28, 2014 6:49 PM
> > > > > > To: Doherty, Declan
> > > > > > Cc: dev@dpdk.org
> > > > > > Subject: Re: [dpdk-dev] [PATCH 0/4] Link Bonding Library
> > > > > >
> > > > > > On Wed, May 28, 2014 at 04:32:00PM +0100, declan.doherty@intel.com
> > > > wrote:
> > > > > > > From: Declan Doherty <declan.doherty@intel.com>
> > > > > > >
> > > > > > > Initial release of Link Bonding Library (lib/librte_bond) with
> > > > > > > support for bonding modes :
> > > > > > >  0 - Round Robin
> > > > > > >  1 - Active Backup
> > > > > > >  2 - Balance l2 / l23 / l34
> > > > > > >  3 - Broadcast
> > > > > > >
> > > > > > Why make this a separate library?  That requires exposure of yet
> > > > > > another
> > > > API to applications.  Instead, why > not write a PMD that can enslave
> > > > other PMD's and treat them all as a single interface?  That way this
> > > > all >  > works with the existing API.
> > > > > >
> > > > > > Neil
> > > > >
> > > > > Hi Neil,
> > > > > the link bonding device is essentially a software PMD, and as such
> > > > > supports
> > > > all the standard PMD APIs, the only new APIs which the link bonding
> > > > library introduces  are for the control operations of the bonded
> > > > device which are currently unsupported by the standard PMD API.
> > > > Operations such as creating, adding/removing slaves, and configuring
> > > > the modes of operation of the device have no analogous APIs in the
> > > > current PMD API and required new ones to be created .
> > > >
> > > > Thats really only true in spirit, in the sense that this library
> > > > transmits and receives frames like a PMD does.  In practice it doesn't
> > > > work and isn't architected the same way.  You don't register the
> > > > driver using the same method as the other PMDs, which means that using
> > > > --vdev on the command line wont work for this type of device.  It also
> > > > implies that applications have to be made specifically aware of the
> > > > fact that they are using a bonded interface (i.e. they need to call
> > > > the bonding setup routines to create the bond).  I would
> > > > recommend:
> > > >
> > > > 1) Register the pmd using the PMD_DRIVER_REGISTER macro, like other
> > > > PMD's
> > > > 2) Use the kvargs library to support configuration via the --vdev
> > > > command line option, so bonds can be created administratively, rather
> > > > than just programatically
> > > > 3) Separate the command api from the PMD functionality into separate
> > > > libraries (use control mbufs to communicate configuration changes to
> > > > the pmd).  This will allow users to dynamically load the pmd
> > > > functionality (without compile or run time linking requirements), and
> > > > then optionally use the programatic interface (or not if they want to
> > > > save memory)
> > > >
> > > > Regards
> > > > Neil
> > > > -----Original Message-----
> > > > From: Neil Horman [mailto:nhorman@tuxdriver.com]
> > > > Sent: Thursday, June 5, 2014 12:04 PM
> > > > To: Doherty, Declan
> > > > Cc: dev@dpdk.org
> > > > Subject: Re: [dpdk-dev] [PATCH v2 0/4] Link Bonding Library
> > > >
> > > > On Wed, Jun 04, 2014 at 04:18:01PM +0100, declan.doherty@intel.com
> > > > wrote:
> > > > > From: Declan Doherty <declan.doherty@intel.com>
> > > > >
> > > > > v2 patch additions,
> > > > > fix for tx burst broadcast, incrementing the reference count on each
> > > > > mbuf by the number of slaves - 1 add/remove slave behavior chnange
> > > > > to fix primary slave port assignment patchcheck code fixes
> > > > >
> > > > >
> > > > >
> > > > This doesn't address any of the comments I made previously.
> > > > Neil
> > >
> > >
> > > Hi Neil,
> > >
> > > sorry for not replying regarding this earlier, in relation to your
> > recommendations 1 and 2, I'm currently working on the implementation of both
> > and should have an additional patch within the next couple of days to add this
> > functionality.
> > >
> > > Regarding your third recommendation, I have no problem splitting the library
> > into separate API and PMD libraries although this does seem to break convention
> > with the other non hw based pmd's, such as pmd_pcap,
> > I'm not at all sure what you mean here.  pmd_pcap registers a driver with the
> > registration mechanism, allowing for the use of the --vdev option on the command
> > line, same as all the other virtual drivers.  The physical drivers do the same
> > thing using the --whitelist and --blacklist options.  How is the bonding driver
> > following convention with pmd_pcap?
> > 
> > 
> > > so would prefer not to go down that route. Also, I don't see the ability to
> > dynamically load the pmd as justification for the use of a control mbufs interface,
> > this would require that either additional queues be created to handle control
> > messages, or that the data TX queues have the ability to identify and process
> > control mbufs, and either of these options will have a performance hit,
> > 
> > Well, you're correct in that the driver would have to identify control messages,
> > but I don't see how thats a performance hit at all.  If you create a control
> > channel queue, then all you have to introduce to the datapath is a single branch
> > to see which queue is transmitting, and your done. If control mbufs are just
> > that hurtful to performance, it seems thats an area that really needs
> > improvement then.
> > 
> > Another alternative is to add a control packet method to the rte_eth_dev
> > structure allowing you to pass control mbufs directly to a pmd out of line with
> > the datapath.  That could avoid any additional branches in the datapath, and
> > still allow you to separate user api from driver implementation.
> > 
> > 
> > > for the core handling the separate control messages or directly to the tx burst
> > performance. I am not currently aware of any functional requirements or use
> > cases for bonded devices to be used over multiple processes which is the only
> > reason I can see for using an interface based on control mbufs,
> > 
> > The reason is administrative.  Its got nothing to do with how many processes
> > want to use an interface, it has to do with an interface knowing if its using
> > the bonding driver or not (it shouldn't ever have to, but with your patch set,
> > its required to use the bonding api to create the interface and add slaves.
> > 
> > > and by providing a --vdev configuration options the user can still write their
> > applications without any knowledge of the additional link bonding programmatic
> > interface. If in the future a use case does emerge which requires multi-process
> > control of bonded devices, I think it would make sense then to add an additional
> > library to provide this functionality.
> > 
> > How exactly does that happen?  At what point does the bonding library in its
> > current form register a driver with rte_ethdev without the application first
> > calling some function in the bonding library to do so.  Because until that
> > happens, --vdev is non functional for this pmd.
> > 
> > FWIW, the concern for me here is one of packaging.  I don't want users to have
> > to pull in a library that they don't absolutely need.  the way you have this put
> > together right now, they need the entire bonding library even if they just want
> > a statically configured bond interface
> > 
> > Neil
> > 
> > 
> > >
> > > Regards
> > > Declan
> > > --------------------------------------------------------------
> 
> Hi Neil,
>  I waited to reply to this until I had v3 of the patch ready to submit. I think you mistook the intent of my last
> response, but anyway I think v3 of the patch addresses the majority of your concerns regarding the bonded device. Which
> can now be initialized using the --vdev options and supports both virtual and physical devices to be specified
> as slaves and no knowledge of the bonding API is required by the user application if the bonding ports are configure at
> runtime.
> 
Sounds great, I look forward to seeing it.

Thanks!
Neil

> Regards
>