On Fri, Mar 17, 2017 at 11:56:21AM +0100, Gaëtan Rivet wrote:

On Thu, Mar 16, 2017 at 04:50:43PM -0400, Neil Horman wrote:

quoted

On Wed, Mar 15, 2017 at 03:25:37PM +0100, Gaëtan Rivet wrote:

quoted

On Wed, Mar 15, 2017 at 12:15:56PM +0100, Thomas Monjalon wrote:

quoted

2017-03-15 03:28, Bruce Richardson:

quoted

On Tue, Mar 14, 2017 at 03:49:47PM +0100, Gaëtan Rivet wrote:

quoted

- In the bonding, the init and configuration steps are still the
 responsibility of the application and no one else. The bonding PMD
 captures the device, re-applies its configuration upon dev_configure()
 which is actually re-applying part of the configuration already  present
within the slave eth_dev (cf rte_eth_dev_config_restore).

- In the fail-safe, the init and configuration are both the
 responsibilities of the fail-safe PMD itself, not the application
 anymore. This handling of these responsibilities in lieu of the
 application is the whole point of the "deferred hot-plug" support, of
 proposing a simple implementation to the user.

This change in responsibilities is the bulk of the fail-safe code. It
would have to be added as-is to the bonding. Verifying the correctness
of the sync of the initialization phase (acceptable states of a device
following several events registered by the fail-safe PMD) and the
configuration items between the state the application believes it is in
and the fail-safe knows it is in, is the bulk of the fail-safe code.

This function is not overlapping with that of the bonding. The reason I
did not add this whole architecture to the bonding is that when I tried
to do so, I found that I only had two possibilities:

- The current slave handling path is kept, and we only add a new one
 with additional functionalities: full init and conf handling with
 extended parsing capabilities.

- The current slave handling is scraped and replaced entirely by the new
 slave management. The old capturing of existing device is not done
 anymore.

The first solution is not acceptable, because we effectively end-up with
a maintenance nightmare by having to validate two types of slaves with
differing capabilities, differing initialization paths and differing
configuration code.  This is extremely awkward and architecturally
unsound. This is essentially the same as having the exact code of the
fail-safe as an aside in the bonding, maintening exactly the same
breadth of code while having muddier interfaces and organization.

The second solution is not acceptable, because we are bending the whole
existing bonding API to our whim. We could just as well simply rename
the fail-safe PMD as bonding, add a few grouping capabilities and call
it a day. This is not acceptable for users.

If the first solution is indeed not an option, why do you think this
second one would be unacceptable for users? If the functionality remains
the same, I don't see how it matters much for users which driver
provides it or where the code originates.

The problem with the second solution is also that bonding is not only a PMD.
It exposes its own public API that existing applications rely on, see
rte_eth_bond_*() definitions in rte_eth_bond.h.

Although bonding instances can be set up through command-line options,
target "users" are mainly applications explicitly written to use it.
This must be preserved for no other reason that it hasn't been deprecated.

I fail to see how either of your points are relevant.  The fact that the bonding
pmd exposes an api to the application has no bearing on its ability to implement
a hot plug function.

This depends on the API making sense in the context of the new
functionality.

Well, the api should always make sense in the context of any added
functionality, but it seems to me thats just another way of saying you might
need to make some modifications to the bonding api.  I'm not saying thats
necessecarily going to have to be the case, but while I'm a big proponent of ABI
stability, I would support an API change to add valid functionality if it were a
big enough feature.  Though again, I don't think thats really necessecary if you
rethink the model a little bit

This API offers to add and remove slaves to a bonding and to configure them.
In the fail-safe arch, it is not possible to add and remove slaves from the
grouping. Doing so would mean adding and removing devices from internal EAL
structures.

Ok, so you update the bonding option parser to include a fail-safe mode, which
only supports two slaves, one of which is an instance of the null-pmd and the
other is to be added at a later date in response to a hot plug event.  In the
fail safe mode, after initial option parsing, bonds operating in this mode
return an error to the application when/if it attempts to remove a slave from a
bond.  That doesn't seem so hard to me.

It is also invalid to try to configure a fail-safe slave. An application
only configures a fail-safe device, which will in turn configure its slaves.
This separation follows from the nature of a device failover.

See my previous mail, you augment the null pmd to allow storage of aribtrary
configuration strings or key/value pairs when operating as a bonded slave.  The
bond is then capable of retrieving configuration from that null instance and
pushing it to the real slave on a hot plug event.

As seen previously, the fail-safe PMD handles different responsibilities
from the bonding PMD. It is thus necessary to make different assumptions
concerning what it can and cannot do with a slave.

Only because you seem to refuse to think about the failsafe model in any other
way than the one you have implemented.

quoted

quoted

Also, trying to implement this API for the device failover function would
implies a device capture down to the devargs parsing level. This means that
a PMD could request taking over a device, messing with the internals of the
EAL: devargs list and busses lists of devices. This seems unacceptable.

Why?  You just said yourself above that, while there is a devargs interface to
the bonding driver, there is also an api, which is the more used method to
configure bonding.  I'm not sure I agree with that, but I think its beside the
point.  Your PMD also requires configuration, and it appears necessecary that
you do so from the command line (you need to specifically ennumerate the
subdevices that you intend to provide failsafe behavior to).  I see no reason
why such a feature cant' be added to bonding, and the null pmd used as a
standin device, should the ennumerated device not yet exist).

To your argument regarding about taking over a device, I don't see how you find
that unacceptable, as it is precisely what the bonding driver does today, in the
sense that it allows an application to assign a master/slave relationship to
devices right now.  I see no reason that we can't convey the right and ability
for bonding to do that dynamically based on configuration.

No, the bonding PMD does not take over a device. It only cares about the
ether layer for its link failover. It does not care about parsing parameters
of a slave, probing devices, detaching drivers. It does not remove a device
from the pci_device_list in the EAL for example.

Again, please take a moment and think about how else your failsafe model might
be implemented in the context of bonding.  Right now, you are asserting that
your failsafe model can't be implemented in any other way becausee of the
decisions you have made.  If you change how you think of failsafe, you'll see it
can be done in other ways.

Doing so would imply exposing private internals structures from the EAL,
messing with elements reserved while doing the EAL init. This requires
controlling a priority in the device initialization order to create the
over-arching ones last (which is a hacky solution). It would wreak havoc
with the DPDK arch.

The fail-safe PMD does not rely on the EAL for handling its slaves. This is
what I explained before, when I touched upon the differing responsibilities
implied by the differences in nature between a link failover and a device
failover.

quoted

quoted

The bonding API is thus in conflict with the concept of a device failover in
the context of the current DPDK arch.

I really don't see how you get to this from your argument above.

The current DPDK arch does not expose EAL elements to be modified by PMDs,
and with good reasons. In this context, it is not possible to handle slaves
correctly for a device failover in the bonding PMD,
because the bonding PMD from the get-go expects the EAL to handle its slaves
on a device level.

quoted

quoted

quoted

quoted

Despite all the discussion, it still just doesn't make sense to me to
have more than one DPDK driver to handle failover - be it link or
device. If nothing else, it's going to be awkward to explain to users
that if they want fail-over for when a link goes down they have to use
driver A, but if they want fail-over when a NIC gets hotplugged they use
driver B, and if they want both kinds of failover - which would surely
be the expected case - they need to use both drivers. The usability is
a problem here.

Having both kind of failovers in the same PMD will always lead to the first
solution in some form or another.

It really isn't because you can model hotplug behavior as a trival form of the
failover that bonding does now (i.e. failover between a null device and a
preferred real device).

The preferred real device still has to be created / destroyed. It still
relies on EAL entry points for handling. It still puts additional
responsibilities on a PMD. Those responsibilities are expressed in sub
layers clearly defined in the fail-safe PMD. You would have to create these
sub-layers in some form in the bonding for it to be able to create a
preferred real device at some point. This additional way of handling slaves
has already been discussed as inducing a messy architecture to the bonding
PMD.

quoted

quoted

I am sure we can document all this in a way that does no cause users
confusion, with the help of community feedback such as yours.

Perhaps "net_failsafe" is a misnomer? We also thought about "net_persistent"
or "net_hotplug". Any other ideas?

It is also possible for me to remove the failover support from this series,
only providing deferred hot-plug handling at first. I could then send the
failover support as separate patches to better assert that it is a useful,
secondary feature that is essentially free to implement.

I think thats solving the wrong problem.  I've no issue with the functionality
in this patch, its really the implementation that we are all arguing against.

quoted

quoted

It seems everybody agrees on the need for the failsafe code.
We are just discussing the right place to implement it.

Gaetan, moving this code in the bonding PMD means replacing the bonding
API design by the failsafe design, right?
With the failsafe design in the bonding PMD, is it possible to keep other
bonding features?

As seen previously, the bonding API is incompatible with device failover.

Its not been seen previously, you asserted it to be so, and I certainly disagree
with that assertion.  I think others might too.

I also explained at length my assertion. I can certainly expand further if
necessary, but you need to point the elements you disagree with.

quoted

Additionally, its not really in line with this discussion, but in looking at
your hotplug detection code, I think somewhat lacking.  Currently you seem to
implement this with a timer that wakes up and checks for device existance, which
is pretty substandard in my mind.  Thats going to waste cpu cycles that might
lead to packet loss.  I'd really prefer to see you augment the eal library with
an event handling code (it can tie into udev in linux and kqueue in bsd), and
create a generic event hook, that we can use to detect device adds/removes
without having to wake up constantly to see if anything has changed.

I think it's fine. We can discuss it further once we agree on the form the
hot-plug implementation will take in the DPDK.

Well, until then, I feel like we're talking past one another at this point, and
so I'll just say this driver is a NAK for me.

Neil

quoted

quoted

Having some features enabled solely for one kind of failover, while
having
specific code paths for both, seems unecessarily complicated to me ;
following suite with my previous points about the first solution.

quoted

In case we do not have a consensus in the following days, I suggest to add
this topic in the next techboard meeting agenda.

Best regards,
-- 
Gaëtan Rivet
6WIND