[v2] doc: flow rule removal on port stop
diff mbox series

Message ID 20201118161520.4378-1-getelson@nvidia.com
State Accepted
Delegated to: Thomas Monjalon
Headers show
Series
  • [v2] doc: flow rule removal on port stop
Related show

Checks

Context Check Description
ci/travis-robot success Travis build: passed
ci/Intel-compilation success Compilation OK
ci/checkpatch success coding style OK

Commit Message

Gregory Etelson Nov. 18, 2020, 4:15 p.m. UTC
There is a discrepancy between RTE ETHDEV API and flow rules guide
regarding flow rules maintenance after port stop.  RTE ETHDEV API in
librte_ethdev.h declares that flow rules will not be stored in PMD
after port stop:
 >>>>> Quite start
 Please note that some configuration is not stored between calls to
 rte_eth_dev_stop()/rte_eth_dev_start(). The following configuration
 will be retained:

 - MTU
 - flow control settings
 - receive mode configuration (promiscuous mode, all-multicast mode,
   hardware checksum mode, RSS/VMDQ settings etc.)
 - VLAN filtering configuration
 - default MAC address
 - MAC addresses supplied to MAC address array
 - flow director filtering mode (but not filtering rules)
 - NIC queue statistics mappings
 <<<< Quote end

PMD cannot always correctly restore flow rules after port stop / port
start because application may alter port configuration after port stop
without PMD knowledge about undergoing changes.  Consider the
following scenario:
application configures 2 queues 0 and 1 and creates a flow rule with
'queue index 1' action. After that application stops the port and
removes queue 1.
Although PMD can implement flow rule shadow copy to be used for
restore after port start, attempt to restore flow rule from shadow
will fail in example above and PMD could not notify application about
that failure.  As the result, flow rules map in HW will differ from
what application expects.  In addition, flow rules shadow copy used
for port start restore consumes considerable amount of system memory,
especially in systems with millions of flow rules.

Signed-off-by: Gregory Etelson <getelson@nvidia.com>
Acked-by: Ori Kam <orika@nvidia.com>
---
 doc/guides/prog_guide/rte_flow.rst | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

Comments

Thomas Monjalon Nov. 22, 2020, 4:55 p.m. UTC | #1
Andrew, any comment on this v2?
(disclosure: I did not read it)


18/11/2020 17:15, Gregory Etelson:
> There is a discrepancy between RTE ETHDEV API and flow rules guide
> regarding flow rules maintenance after port stop.  RTE ETHDEV API in
> librte_ethdev.h declares that flow rules will not be stored in PMD
> after port stop:
>  >>>>> Quite start
>  Please note that some configuration is not stored between calls to
>  rte_eth_dev_stop()/rte_eth_dev_start(). The following configuration
>  will be retained:
> 
>  - MTU
>  - flow control settings
>  - receive mode configuration (promiscuous mode, all-multicast mode,
>    hardware checksum mode, RSS/VMDQ settings etc.)
>  - VLAN filtering configuration
>  - default MAC address
>  - MAC addresses supplied to MAC address array
>  - flow director filtering mode (but not filtering rules)
>  - NIC queue statistics mappings
>  <<<< Quote end
> 
> PMD cannot always correctly restore flow rules after port stop / port
> start because application may alter port configuration after port stop
> without PMD knowledge about undergoing changes.  Consider the
> following scenario:
> application configures 2 queues 0 and 1 and creates a flow rule with
> 'queue index 1' action. After that application stops the port and
> removes queue 1.
> Although PMD can implement flow rule shadow copy to be used for
> restore after port start, attempt to restore flow rule from shadow
> will fail in example above and PMD could not notify application about
> that failure.  As the result, flow rules map in HW will differ from
> what application expects.  In addition, flow rules shadow copy used
> for port start restore consumes considerable amount of system memory,
> especially in systems with millions of flow rules.
> 
> Signed-off-by: Gregory Etelson <getelson@nvidia.com>
> Acked-by: Ori Kam <orika@nvidia.com>
> ---
>  doc/guides/prog_guide/rte_flow.rst | 10 ++++++----
>  1 file changed, 6 insertions(+), 4 deletions(-)
> 
> diff --git a/doc/guides/prog_guide/rte_flow.rst b/doc/guides/prog_guide/rte_flow.rst
> index ea203e0ca4..4cff9332fa 100644
> --- a/doc/guides/prog_guide/rte_flow.rst
> +++ b/doc/guides/prog_guide/rte_flow.rst
> @@ -3229,10 +3229,12 @@ Caveats
>    temporarily replacing the burst function pointers), an appropriate error
>    code must be returned (``EBUSY``).
>  
> -- PMDs, not applications, are responsible for maintaining flow rules
> -  configuration when stopping and restarting a port or performing other
> -  actions which may affect them. They can only be destroyed explicitly by
> -  applications.
> +- Applications, not PMDs, are responsible for maintaining flow rules
> +  configuration when closing, stopping or restarting a port or performing other
> +  actions which may affect them.
> +  Applications must assume that after port close, stop or restart all flows
> +  related to that port are not valid, hardware rules are destroyed and relevant
> +  PMD resources are released.
>  
>  For devices exposing multiple ports sharing global settings affected by flow
>  rules:
>
Thomas Monjalon Nov. 24, 2020, 11:04 a.m. UTC | #2
There is also a testpmd patch pending about this behaviour:
	https://patches.dpdk.org/patch/75353/

22/11/2020 17:55, Thomas Monjalon:
> Andrew, any comment on this v2?
> (disclosure: I did not read it)
> 
> 
> 18/11/2020 17:15, Gregory Etelson:
> > There is a discrepancy between RTE ETHDEV API and flow rules guide
> > regarding flow rules maintenance after port stop.  RTE ETHDEV API in
> > librte_ethdev.h declares that flow rules will not be stored in PMD
> > after port stop:
> >  >>>>> Quite start
> >  Please note that some configuration is not stored between calls to
> >  rte_eth_dev_stop()/rte_eth_dev_start(). The following configuration
> >  will be retained:
> > 
> >  - MTU
> >  - flow control settings
> >  - receive mode configuration (promiscuous mode, all-multicast mode,
> >    hardware checksum mode, RSS/VMDQ settings etc.)
> >  - VLAN filtering configuration
> >  - default MAC address
> >  - MAC addresses supplied to MAC address array
> >  - flow director filtering mode (but not filtering rules)
> >  - NIC queue statistics mappings
> >  <<<< Quote end
> > 
> > PMD cannot always correctly restore flow rules after port stop / port
> > start because application may alter port configuration after port stop
> > without PMD knowledge about undergoing changes.  Consider the
> > following scenario:
> > application configures 2 queues 0 and 1 and creates a flow rule with
> > 'queue index 1' action. After that application stops the port and
> > removes queue 1.
> > Although PMD can implement flow rule shadow copy to be used for
> > restore after port start, attempt to restore flow rule from shadow
> > will fail in example above and PMD could not notify application about
> > that failure.  As the result, flow rules map in HW will differ from
> > what application expects.  In addition, flow rules shadow copy used
> > for port start restore consumes considerable amount of system memory,
> > especially in systems with millions of flow rules.
> > 
> > Signed-off-by: Gregory Etelson <getelson@nvidia.com>
> > Acked-by: Ori Kam <orika@nvidia.com>
> > ---
> >  doc/guides/prog_guide/rte_flow.rst | 10 ++++++----
> >  1 file changed, 6 insertions(+), 4 deletions(-)
> > 
> > diff --git a/doc/guides/prog_guide/rte_flow.rst b/doc/guides/prog_guide/rte_flow.rst
> > index ea203e0ca4..4cff9332fa 100644
> > --- a/doc/guides/prog_guide/rte_flow.rst
> > +++ b/doc/guides/prog_guide/rte_flow.rst
> > @@ -3229,10 +3229,12 @@ Caveats
> >    temporarily replacing the burst function pointers), an appropriate error
> >    code must be returned (``EBUSY``).
> >  
> > -- PMDs, not applications, are responsible for maintaining flow rules
> > -  configuration when stopping and restarting a port or performing other
> > -  actions which may affect them. They can only be destroyed explicitly by
> > -  applications.
> > +- Applications, not PMDs, are responsible for maintaining flow rules
> > +  configuration when closing, stopping or restarting a port or performing other
> > +  actions which may affect them.
> > +  Applications must assume that after port close, stop or restart all flows
> > +  related to that port are not valid, hardware rules are destroyed and relevant
> > +  PMD resources are released.
> >  
> >  For devices exposing multiple ports sharing global settings affected by flow
> >  rules:
Ajit Khaparde Nov. 24, 2020, 2:41 p.m. UTC | #3
On Wed, Nov 18, 2020 at 8:15 AM Gregory Etelson <getelson@nvidia.com> wrote:
>
> There is a discrepancy between RTE ETHDEV API and flow rules guide
> regarding flow rules maintenance after port stop.  RTE ETHDEV API in
> librte_ethdev.h declares that flow rules will not be stored in PMD
> after port stop:
>  >>>>> Quite start
>  Please note that some configuration is not stored between calls to
>  rte_eth_dev_stop()/rte_eth_dev_start(). The following configuration
>  will be retained:
>
>  - MTU
>  - flow control settings
>  - receive mode configuration (promiscuous mode, all-multicast mode,
>    hardware checksum mode, RSS/VMDQ settings etc.)
>  - VLAN filtering configuration
>  - default MAC address
>  - MAC addresses supplied to MAC address array
>  - flow director filtering mode (but not filtering rules)
>  - NIC queue statistics mappings
>  <<<< Quote end
>
> PMD cannot always correctly restore flow rules after port stop / port
> start because application may alter port configuration after port stop
> without PMD knowledge about undergoing changes.  Consider the
> following scenario:
> application configures 2 queues 0 and 1 and creates a flow rule with
> 'queue index 1' action. After that application stops the port and
> removes queue 1.
> Although PMD can implement flow rule shadow copy to be used for
> restore after port start, attempt to restore flow rule from shadow
> will fail in example above and PMD could not notify application about
> that failure.  As the result, flow rules map in HW will differ from
> what application expects.  In addition, flow rules shadow copy used
> for port start restore consumes considerable amount of system memory,
> especially in systems with millions of flow rules.
>
> Signed-off-by: Gregory Etelson <getelson@nvidia.com>
> Acked-by: Ori Kam <orika@nvidia.com>
Acked-by: Ajit Khaparde <ajit.khaparde@broadcom.com>

> ---
>  doc/guides/prog_guide/rte_flow.rst | 10 ++++++----
>  1 file changed, 6 insertions(+), 4 deletions(-)
>
> diff --git a/doc/guides/prog_guide/rte_flow.rst b/doc/guides/prog_guide/rte_flow.rst
> index ea203e0ca4..4cff9332fa 100644
> --- a/doc/guides/prog_guide/rte_flow.rst
> +++ b/doc/guides/prog_guide/rte_flow.rst
> @@ -3229,10 +3229,12 @@ Caveats
>    temporarily replacing the burst function pointers), an appropriate error
>    code must be returned (``EBUSY``).
>
> -- PMDs, not applications, are responsible for maintaining flow rules
> -  configuration when stopping and restarting a port or performing other
> -  actions which may affect them. They can only be destroyed explicitly by
> -  applications.
> +- Applications, not PMDs, are responsible for maintaining flow rules
> +  configuration when closing, stopping or restarting a port or performing other
> +  actions which may affect them.
> +  Applications must assume that after port close, stop or restart all flows
> +  related to that port are not valid, hardware rules are destroyed and relevant
> +  PMD resources are released.
>
>  For devices exposing multiple ports sharing global settings affected by flow
>  rules:
> --
> 2.29.2
>
Thomas Monjalon Nov. 25, 2020, 11:33 p.m. UTC | #4
24/11/2020 15:41, Ajit Khaparde:
> On Wed, Nov 18, 2020 at 8:15 AM Gregory Etelson <getelson@nvidia.com> wrote:
> >
> > There is a discrepancy between RTE ETHDEV API and flow rules guide
> > regarding flow rules maintenance after port stop.  RTE ETHDEV API in
> > librte_ethdev.h declares that flow rules will not be stored in PMD
> > after port stop:
> >  >>>>> Quite start
> >  Please note that some configuration is not stored between calls to
> >  rte_eth_dev_stop()/rte_eth_dev_start(). The following configuration
> >  will be retained:
> >
> >  - MTU
> >  - flow control settings
> >  - receive mode configuration (promiscuous mode, all-multicast mode,
> >    hardware checksum mode, RSS/VMDQ settings etc.)
> >  - VLAN filtering configuration
> >  - default MAC address
> >  - MAC addresses supplied to MAC address array
> >  - flow director filtering mode (but not filtering rules)
> >  - NIC queue statistics mappings
> >  <<<< Quote end
> >
> > PMD cannot always correctly restore flow rules after port stop / port
> > start because application may alter port configuration after port stop
> > without PMD knowledge about undergoing changes.  Consider the
> > following scenario:
> > application configures 2 queues 0 and 1 and creates a flow rule with
> > 'queue index 1' action. After that application stops the port and
> > removes queue 1.
> > Although PMD can implement flow rule shadow copy to be used for
> > restore after port start, attempt to restore flow rule from shadow
> > will fail in example above and PMD could not notify application about
> > that failure.  As the result, flow rules map in HW will differ from
> > what application expects.  In addition, flow rules shadow copy used
> > for port start restore consumes considerable amount of system memory,
> > especially in systems with millions of flow rules.
> >
> > Signed-off-by: Gregory Etelson <getelson@nvidia.com>
> > Acked-by: Ori Kam <orika@nvidia.com>
> Acked-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
> 
> > ---
> > -- PMDs, not applications, are responsible for maintaining flow rules
> > -  configuration when stopping and restarting a port or performing other
> > -  actions which may affect them. They can only be destroyed explicitly by
> > -  applications.
> > +- Applications, not PMDs, are responsible for maintaining flow rules
> > +  configuration when closing, stopping or restarting a port or performing other
> > +  actions which may affect them.
> > +  Applications must assume that after port close, stop or restart all flows
> > +  related to that port are not valid, hardware rules are destroyed and relevant
> > +  PMD resources are released.

In short summary, this rte_flow doc change has 3 reasons:
	- consistency with ethdev API doc
	- avoid unsolvable automatic flow update after re-configuration
	- reduce memory consumption for flow rules

Andrew was asking, in previous version, how to manage reset for error recovery.
As it has been discussed in other threads, an error recovery should be notified
to the application.
We already have RTE_ETH_EVENT_INTR_RESET for VF in case of PF reset,
and a more general recovery notification mechanism is being discussed:
	https://patches.dpdk.org/patch/80094/
Then it will be possible to notify the application that the flow rules
must be restored (among other recovery measures).

For the case of port stop/close, doc update applied, thanks.

Patch
diff mbox series

diff --git a/doc/guides/prog_guide/rte_flow.rst b/doc/guides/prog_guide/rte_flow.rst
index ea203e0ca4..4cff9332fa 100644
--- a/doc/guides/prog_guide/rte_flow.rst
+++ b/doc/guides/prog_guide/rte_flow.rst
@@ -3229,10 +3229,12 @@  Caveats
   temporarily replacing the burst function pointers), an appropriate error
   code must be returned (``EBUSY``).
 
-- PMDs, not applications, are responsible for maintaining flow rules
-  configuration when stopping and restarting a port or performing other
-  actions which may affect them. They can only be destroyed explicitly by
-  applications.
+- Applications, not PMDs, are responsible for maintaining flow rules
+  configuration when closing, stopping or restarting a port or performing other
+  actions which may affect them.
+  Applications must assume that after port close, stop or restart all flows
+  related to that port are not valid, hardware rules are destroyed and relevant
+  PMD resources are released.
 
 For devices exposing multiple ports sharing global settings affected by flow
 rules: