RE: [RFC 0/4] ethdev: rework config restore

[Konstantin Ananyev <konstantin.ananyev@huawei.com>]

> >>>>>> We have been working on optimizing the latency of calls to
> >>>>>> rte_eth_dev_start(), on ports spawned by mlx5 PMD. Most of the work
> >>>>>> requires changes in the implementation of
> >>>>>> .dev_start() PMD callback, but I also wanted to start a discussion
> >>>>>> regarding configuration restore.
> >>>>>>
> >>>>>> rte_eth_dev_start() does a few things on top of calling .dev_start() callback:
> >>>>>>
> >>>>>> - Before calling it:
> >>>>>>     - eth_dev_mac_restore() - if device supports
> >>>>>> RTE_ETH_DEV_NOLIVE_MAC_ADDR;
> >>>>>> - After calling it:
> >>>>>>     - eth_dev_mac_restore() - if device does not support
> >>>>> RTE_ETH_DEV_NOLIVE_MAC_ADDR;
> >>>>>>     - restore promiscuous config
> >>>>>>     - restore all multicast config
> >>>>>>
> >>>>>> eth_dev_mac_restore() iterates over all known MAC addresses - stored
> >>>>>> in rte_eth_dev_data.mac_addrs array - and calls
> >>>>>> .mac_addr_set() and .mac_addr_add() callbacks to apply these MAC addresses.
> >>>>>>
> >>>>>> Promiscuous config restore checks if promiscuous mode is enabled or
> >>>>>> not, and calls .promiscuous_enable() or .promiscuous_disable() callback.
> >>>>>>
> >>>>>> All multicast config restore checks if all multicast mode is enabled
> >>>>>> or not, and calls .allmulticast_enable() or .allmulticast_disable() callback.
> >>>>>>
> >>>>>> Callbacks are called directly in all of these cases, to bypass the
> >>>>>> checks for applying the same configuration, which exist in relevant APIs.
> >>>>>> Checks are bypassed to force drivers to reapply the configuration.
> >>>>>>
> >>>>>> Let's consider what happens in the following sequence of API calls.
> >>>>>>
> >>>>>> 1. rte_eth_dev_configure()
> >>>>>> 2. rte_eth_tx_queue_setup()
> >>>>>> 3. rte_eth_rx_queue_setup()
> >>>>>> 4. rte_eth_promiscuous_enable()
> >>>>>>     - Call dev->dev_ops->promiscuous_enable()
> >>>>>>     - Stores promiscuous state in dev->data->promiscuous 5.
> >>>>>> rte_eth_allmulticast_enable()
> >>>>>>     - Call dev->dev_ops->allmulticast_enable()
> >>>>>>     - Stores allmulticast state in dev->data->allmulticast 6.
> >>>>>> rte_eth_dev_start()
> >>>>>>     - Call dev->dev_ops->dev_start()
> >>>>>>     - Call dev->dev_ops->mac_addr_set() - apply default MAC address
> >>>>>>     - Call dev->dev_ops->promiscuous_enable()
> >>>>>>     - Call dev->dev_ops->allmulticast_enable()
> >>>>>>
> >>>>>> Even though all configuration is available in dev->data after step 5,
> >>>>>> library forces reapplying this configuration in step 6.
> >>>>>>
> >>>>>> In mlx5 PMD case all relevant callbacks require communication with the
> >>>>>> kernel driver, to configure the device (mlx5 PMD must create/destroy
> >>>>>> new kernel flow rules and/or change netdev config).
> >>>>>>
> >>>>>> mlx5 PMD handles applying all configuration in .dev_start(), so the
> >>>>>> following forced callbacks force additional communication with the kernel. The
> >>>>> same configuration is applied multiple times.
> >>>>>>
> >>>>>> As an optimization, mlx5 PMD could check if a given configuration was
> >>>>>> applied, but this would duplicate the functionality of the library
> >>>>>> (for example rte_eth_promiscuous_enable() does not call the driver if
> >>>>>> dev->data->promiscuous is set).
> >>>>>>
> >>>>>> Question: Since all of the configuration is available before
> >>>>>> .dev_start() callback is called, why ethdev library does not expect .dev_start() to
> >>>>> take this configuration into account?
> >>>>>> In other words, why library has to reapply the configuration?
> >>>>>>
> >>>>>> I could not find any particular reason why configuration restore
> >>>>>> exists as part of the process (it was in the initial DPDK commit).
> >>>>>>
> >>>>>
> >>>>> My assumption is .dev_stop() cause these values reset in some devices, so
> >>>>> .dev_start() restores them back.
> >>>>> @Bruce or @Konstantin may remember the history.
> >>>
> >>> Yep, as I remember, at least some Intel PMDs calling hw_reset() ad dec_stop() and
> >>> even dev_start() to make sure that HW is in a clean (known) state.
> >>>
> >>>>>
> >>>>> But I agree this is device specific behavior, and can be managed by what device
> >>>>> requires.
> >>>
> >>> Probably yes.
> >>>
> >>>>>
> >>>>>> The patches included in this RFC, propose a mechanism which would help
> >>>>>> with managing which drivers rely on forceful configuration restore.
> >>>>>> Drivers could advertise if forceful configuration restore is needed
> >>>>>> through `RTE_ETH_DEV_*_FORCE_RESTORE` device flag. If this flag is
> >>>>>> set, then the driver in question requires ethdev to forcefully restore
> >>>>> configuration.
> >>>>>>
> >>>>>
> >>>>> OK to use flag for it, but not sure about using 'dev_info->dev_flags'
> >>>>> (RTE_ETH_DEV_*) for this, as this flag is shared with user and this is all dpdk
> >>>>> internal.
> >>>>>
> >>>>> What about to have a dedicated flag for it? We can have a dedicated set of flag
> >>>>> values for restore.
> >>>>
> >>>> Agreed. What do you think about the following?
> >>>
> >>> Instead of exposing that, can we probably make it transparent to the user
> >>> and probably ethdev layer too?
> >>>
> >>
> >> +1 to make it transparent to user, but not sure if we can make it
> >> transparent to ethdev layer.
> >
> > Just to be clear:
> > Let say, using example from above:
> >
> >  rte_eth_dev_start()
> >      - Call dev->dev_ops->dev_start()
> >      - Call dev->dev_ops->mac_addr_set() - apply default MAC address
> >      - Call dev->dev_ops->promiscuous_enable()
> >      - Call dev->dev_ops->allmulticast_enable()
> >
> > We probably can introduce ethdev internal function (still visible to PMDs)
> > that would do last 3 steps:
> > ethdev_replay_user_conf(...)
> >      - Call dev->dev_ops->mac_addr_set() - apply default MAC address
> >      - Call dev->dev_ops->promiscuous_enable()
> >      - Call dev->dev_ops->allmulticast_enable()
> >
> > And let PMD itself to decide does it needs to call it at dev_start() or not.
> > So it will become:
> > rte_eth_dev_start()
> >      - Call dev->dev_ops->dev_start()
> > 	-Call ethdev_replay_user_conf(.)
> >      		- Call dev->dev_ops->mac_addr_set() - apply default MAC address
> >      		- Call dev->dev_ops->promiscuous_enable()
> >      		-Call dev->dev_ops->allmulticast_enable()
> >
> > For PMDs that do need to restore user provided config
> > And
> > rte_eth_dev_start()
> >      - Call dev->dev_ops->dev_start()
> >
> > For those who do not.
> >
> 
> OK, got it what you mean.
> Pushing restore functionality to PMDs works, but this may be doing
> redundant work on each PMD.

Yes, every PMD will have to be updated with that way.
 
> Instead Dariusz suggests PMD to provide a flag to ehtdev to what to
> restore and common code in ethdev does the work.
> My below dedicated data struct comment is to have this flag in a new
> struct, overall like following:
> 
> rte_eth_dev_start()
>    - Call dev->dev_ops->dev_start()
>    - Call dev->dev_ops->get_restore_flags(ethdev, RTE_ETH_START, &flags)
>    - if (flags & MAC) dev->dev_ops->mac_addr_set()
>    - if (flags & PROMISC) dev->dev_ops->promiscuous_enable()
>    - ...
> 
> So PMDs only will provide what to restore with an internal API and
> common ethdev layer will restore it.
> If no restore required PMD may not implement .get_restore_flags() at all.

Makes sense to me...
Probably less disruptive would be to have a dev_ops function
that would tell what settings don't need to be replayed by ehtdev layer. 
 
> Additionally, RTE_ETH_START, RTE_ETH_RESET etc flag can be provided to
> internal API to get what to restore in different states...
> 
> >> Suggested 'internal_flag' in "struct rte_eth_dev_data" can be confusing
> >> and open to interpretation what to use it for and by time become source
> >> of defect.
> >
> > Yes, same thoughts.
> >
> >> Instead what do you think to have a separate, dedicated data struct for it?
> >
> > Hmm... not sure I understood you here...
> >
> >>
> >>> Might be we can move this restoration code into the new ethdev helper function,(ethdevd_user_config_restore()  or so)
> >>> that PMD can invoke during its dev_start() if needed?
> >>>
> >>>>
> >>>> #define RTE_ETH_DEV_INTERNAL_PROMISC_FORCE_RESTORE RTE_BIT32(0)
> >>>> #define RTE_ETH_DEV_INTERNAL_ALLMULTI_FORCE_RESTORE RTE_BIT32(1)
> >>>> #define RTE_ETH_DEV_INTERNAL_MAC_ADDR_FORCE_RESTORE RTE_BIT32(2)
> >>>>
> >>>> struct rte_eth_dev_data {
> >>>> 	/* snip */
> >>>>
> >>>> 	uint32_t dev_flags;
> >>>>
> >>>> 	/**
> >>>> 	 * Internal device capabilities, used only by ethdev library.
> >>>> 	 * Certain functionalities provided by the library might enabled/disabled,
> >>>> 	 * based on driver exposing certain capabilities.
> >>>> 	 */
> >>>> 	uint32_t internal_flags;
> >>>>
> >>>> 	/* snip */
> >>>> };
> >>>>
> >>>>> Also perhaps we have go into details what needs to be restored after 'stop' and
> >>>>> what needs to be restored after 'reset' and use similar mechanism etc...
> >>>>
> >>>> I think we should look into that.
> >>>> Any 'codification' of semantics between drivers and ethdev library is good in my opinion.
> >>>>
> >>>> At least right now, ethdev does not change any configuration in 'stop' and 'reset' from what I see.
> >>>> But that's on library side only.
> >>>>
> >>>>>> This way, if we would conclude that it makes sense for .dev_start() to
> >>>>>> handle all starting configuration aspects, we could track which drivers still rely
> >>>>> on configuration restore.
> >>>>>>
> >>>>>> Dariusz Sosnowski (4):
> >>>>>>   ethdev: rework config restore
> >>>>>>   ethdev: omit promiscuous config restore if not required
> >>>>>>   ethdev: omit all multicast config restore if not required
> >>>>>>   ethdev: omit MAC address restore if not required
> >>>>>>
> >>>>>>  lib/ethdev/rte_ethdev.c | 39 ++++++++++++++++++++++++++++++++++-----
> >>>>>>  lib/ethdev/rte_ethdev.h | 18 ++++++++++++++++++
> >>>>>>  2 files changed, 52 insertions(+), 5 deletions(-)
> >>>>>>
> >>>>>> --
> >>>>>> 2.39.5
> >>>>>>
> >>>
> >