[dpdk-dev] [PATCH v3] ethdev: fix ethdev data alignment

[Jerin Jacob]

-----Original Message-----
> Date: Mon, 12 Feb 2018 13:44:54 +0000
> From: Matan Azrad <matan at mellanox.com>
> To: Jerin Jacob <jerin.jacob at caviumnetworks.com>, "dev at dpdk.org"
>  <dev at dpdk.org>
> CC: "ferruh.yigit at intel.com" <ferruh.yigit at intel.com>, "stable at dpdk.org"
>  <stable at dpdk.org>, Thomas Monjalon <thomas at monjalon.net>, Konstantin
>  Ananyev <konstantin.ananyev at intel.com>, Pavan Nikhilesh
>  <pbhagavatula at caviumnetworks.com>
> Subject: RE: [dpdk-dev] [PATCH v3] ethdev: fix ethdev data alignment
> 
> Hi Jerin
> 
> From: Jerin Jacob [mailto:jerin.jacob at caviumnetworks.com]
> > The struct rte_eth_dev_data is used in ethdev fastpath routines and it not
> > aligned to cache line size. This patch fixes the ethdev data alignment.
> > 
> > The alignment was broken from the "first public release" changeset where
> > ethdev data address was aligned only to the first port.
> > Remaining ports alignment was defined by the size of the struct
> > (rte_eth_dev_data). This scheme is not guaranteed to be cache line aligned
> > all the time.
> > 
> > "ethdev: add port ownership" change set introduced a
> > rte_eth_dev_shared_data container for port ownership change, This
> > resulted in rte_eth_dev->data memory for the first port also as cache
> > unaligned.
> > 
> > Added a compiler alignment attribute to make sure rte_eth_dev->data
> > always cache aligned so that CPU/compiler
> > 1) Avoid sharing the element with another cache line
> > 2) Can load/store the elements in struct rte_eth_dev_data as naturally
> > aligned.
> > 
> > Some platform like thunderX could see performance regression of 1% at
> > "ethdev: add port ownership" change set with
> > 1 port/1 queue l3fwd application and this patch fixes that regression.
> > 
> > example command:
> > sudo ./examples/l3fwd/build/l3fwd -c 0xff00 -- -p 0x1 --config="(0,0,9)"
> > 
> > Fixes: af75078fece3 ("first public release")
> > Fixes: 5b7ba31148a8 ("ethdev: add port ownership")
> 
> I don't think you need the add the 5b7ba31148a8 fix line.
> Maybe think about it in the next way:
> Is your fix can stay as a same fix before the port ownership feature addition?
> If yes, You are not fixing it.

I don't think so. 5b7ba31148a8 breaking the first port alignment case(as
mentioned in the commit log clearly).
First port alignment was correct prior to 5b7ba31148a8 change set

Do you agree?

> 
> I think that the port ownership just exposed the issue, did not initiated it.
> 
> > Cc: stable at dpdk.org
> > 
> > Cc: Matan Azrad <matan at mellanox.com>
> > Cc: Thomas Monjalon <thomas at monjalon.net>
> > Cc: Konstantin Ananyev <konstantin.ananyev at intel.com>
> > 
> > Signed-off-by: Jerin Jacob <jerin.jacob at caviumnetworks.com>
> > Signed-off-by: Pavan Nikhilesh <pbhagavatula at caviumnetworks.com>
> > ---
> > 
> > v3:
> > - added more history/data in the git log
> > - added stable at dpdk.org in Cc
> > 
> > v2:
> >  - Change the git comments based on Matan's feedback
> > https://emea01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fdpd
> > k.org%2Fdev%2Fpatchwork%2Fpatch%2F35104%2F&data=02%7C01%7Cmata
> > n%40mellanox.com%7Cd491698ead5f48626c1408d5721a8331%7C8c5c71989d
> > be4cf9a2099e499881116c%7C0%7C0%7C636540380570748030&sdata=if8cCrk3
> > U0XqWp8NE%2BEP1BjuPHX95HHN0f1yjk6krV0%3D&reserved=0
> > 
> > ---
> >  lib/librte_ether/rte_ethdev_core.h | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/lib/librte_ether/rte_ethdev_core.h
> > b/lib/librte_ether/rte_ethdev_core.h
> > index 315b31723..e5681e466 100644
> > --- a/lib/librte_ether/rte_ethdev_core.h
> > +++ b/lib/librte_ether/rte_ethdev_core.h
> > @@ -601,7 +601,7 @@ struct rte_eth_dev_data {
> >  	struct rte_vlan_filter_conf vlan_filter_conf;
> >  	/**< VLAN filter configuration. */
> >  	struct rte_eth_dev_owner owner; /**< The port owner. */ -};
> > +} __rte_cache_aligned;
> > 
> >  /**
> >   * @internal
> > --
> > 2.16.1
>