[net-next,V1] net: adjust net_device layout for cacheline usage

Message ID 161168277983.410784.12401225493601624417.stgit@firesoul
State New
Headers show
Series
  • [net-next,V1] net: adjust net_device layout for cacheline usage
Related show

Commit Message

Jesper Dangaard Brouer Jan. 26, 2021, 5:39 p.m.
The current layout of net_device is not optimal for cacheline usage.

The member adj_list.lower linked list is split between cacheline 2 and 3.
The ifindex is placed together with stats (struct net_device_stats),
although most modern drivers don't update this stats member.

The members netdev_ops, mtu and hard_header_len are placed on three
different cachelines. These members are accessed for XDP redirect into
devmap, which were noticeably with perf tool. When not using the map
redirect variant (like TC-BPF does), then ifindex is also used, which is
placed on a separate fourth cacheline. These members are also accessed
during forwarding with regular network stack. The members priv_flags and
flags are on fast-path for network stack transmit path in __dev_queue_xmit
(currently located together with mtu cacheline).

This patch creates a read mostly cacheline, with the purpose of keeping the
above mentioned members on the same cacheline.

Some netdev_features_t members also becomes part of this cacheline, which is
on purpose, as function netif_skb_features() is on fast-path via
validate_xmit_skb().

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
---
 include/linux/netdevice.h |   53 +++++++++++++++++++++++----------------------
 1 file changed, 27 insertions(+), 26 deletions(-)

Comments

David Ahern Jan. 29, 2021, 3:51 a.m. | #1
On 1/26/21 10:39 AM, Jesper Dangaard Brouer wrote:
> The current layout of net_device is not optimal for cacheline usage.

> 

> The member adj_list.lower linked list is split between cacheline 2 and 3.

> The ifindex is placed together with stats (struct net_device_stats),

> although most modern drivers don't update this stats member.

> 

> The members netdev_ops, mtu and hard_header_len are placed on three

> different cachelines. These members are accessed for XDP redirect into

> devmap, which were noticeably with perf tool. When not using the map

> redirect variant (like TC-BPF does), then ifindex is also used, which is

> placed on a separate fourth cacheline. These members are also accessed

> during forwarding with regular network stack. The members priv_flags and

> flags are on fast-path for network stack transmit path in __dev_queue_xmit

> (currently located together with mtu cacheline).

> 

> This patch creates a read mostly cacheline, with the purpose of keeping the

> above mentioned members on the same cacheline.

> 

> Some netdev_features_t members also becomes part of this cacheline, which is

> on purpose, as function netif_skb_features() is on fast-path via

> validate_xmit_skb().


A long over due look at the organization of this struct. Do you have
performance numbers for the XDP case?
patchwork-bot+netdevbpf@kernel.org Jan. 29, 2021, 4:50 a.m. | #2
Hello:

This patch was applied to netdev/net-next.git (refs/heads/master):

On Tue, 26 Jan 2021 18:39:39 +0100 you wrote:
> The current layout of net_device is not optimal for cacheline usage.

> 

> The member adj_list.lower linked list is split between cacheline 2 and 3.

> The ifindex is placed together with stats (struct net_device_stats),

> although most modern drivers don't update this stats member.

> 

> The members netdev_ops, mtu and hard_header_len are placed on three

> different cachelines. These members are accessed for XDP redirect into

> devmap, which were noticeably with perf tool. When not using the map

> redirect variant (like TC-BPF does), then ifindex is also used, which is

> placed on a separate fourth cacheline. These members are also accessed

> during forwarding with regular network stack. The members priv_flags and

> flags are on fast-path for network stack transmit path in __dev_queue_xmit

> (currently located together with mtu cacheline).

> 

> [...]


Here is the summary with links:
  - [net-next,V1] net: adjust net_device layout for cacheline usage
    https://git.kernel.org/netdev/net-next/c/28af22c6c8df

You are awesome, thank you!
--
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html
Jesper Dangaard Brouer Jan. 29, 2021, 7:58 a.m. | #3
On Thu, 28 Jan 2021 20:51:23 -0700
David Ahern <dsahern@gmail.com> wrote:

> On 1/26/21 10:39 AM, Jesper Dangaard Brouer wrote:

> > The current layout of net_device is not optimal for cacheline usage.

> > 

> > The member adj_list.lower linked list is split between cacheline 2 and 3.

> > The ifindex is placed together with stats (struct net_device_stats),

> > although most modern drivers don't update this stats member.

> > 

> > The members netdev_ops, mtu and hard_header_len are placed on three

> > different cachelines. These members are accessed for XDP redirect into

> > devmap, which were noticeably with perf tool. When not using the map

> > redirect variant (like TC-BPF does), then ifindex is also used, which is

> > placed on a separate fourth cacheline. These members are also accessed

> > during forwarding with regular network stack. The members priv_flags and

> > flags are on fast-path for network stack transmit path in __dev_queue_xmit

> > (currently located together with mtu cacheline).

> > 

> > This patch creates a read mostly cacheline, with the purpose of keeping the

> > above mentioned members on the same cacheline.

> > 

> > Some netdev_features_t members also becomes part of this cacheline, which is

> > on purpose, as function netif_skb_features() is on fast-path via

> > validate_xmit_skb().  

> 

> A long over due look at the organization of this struct. Do you have

> performance numbers for the XDP case?


Yes, my measurements are documented here:
 https://github.com/xdp-project/xdp-project/blob/master/areas/core/xdp_redir01_net_device.org

Calc improvements of xdp_redirect_map on driver i40e:
 * (1/12115061-1/12906785)*10^9 = 5.06 ns
 * ((12906785/12115061)-1)*100  = 6.54%

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer
Jakub Kicinski Jan. 29, 2021, 7:35 p.m. | #4
On Fri, 29 Jan 2021 11:46:42 +0100 Jesper Dangaard Brouer wrote:
> > On Thu, 28 Jan 2021 20:51:23 -0700

> > David Ahern <dsahern@gmail.com> wrote:

> > > A long over due look at the organization of this struct.  

> 

> Yes, I was surprised that the cache-lines used in fast-path was this

> spread out.


I tried measuring the cache misses on struct netdevice running
relatively network-heavy production workload once but they were 
really deep in the noise. Things become much easier to optimize
with a XDP micro-benchmark, but obviously should benefit all.

> There is a comment /* Cache lines mostly used on receive path */

> but that comment no-longer start on a cacheline, so I suspect that this

> have slowly diverted over time (Eric's commit 9356b8fc07 dates back to 2005).

> 

> Patch is already applied. I expected people to would say that I also

> needed to adjust the doc-type comments.  The comments describing the

> members, seems to be ordered the same way as defined.  Should we/I keep

> that order intact?  (when moving members)


kdoc didn't complain, and as you say it's already a mess, plus it's
two screen-fulls of scrolling away... 

I think converting to inline kdoc of members would be an improvement,
if you want to sign up for that? Otherwise -EDIDNTCARE on my side :)
Eric Dumazet Jan. 29, 2021, 7:47 p.m. | #5
On 1/29/21 8:35 PM, Jakub Kicinski wrote:

> kdoc didn't complain, and as you say it's already a mess, plus it's

> two screen-fulls of scrolling away... 

> 

> I think converting to inline kdoc of members would be an improvement,

> if you want to sign up for that? Otherwise -EDIDNTCARE on my side :)

> 


What about removing this kdoc ?

kdoc for a huge structure is mostly useless...
Jakub Kicinski Jan. 29, 2021, 8:07 p.m. | #6
On Fri, 29 Jan 2021 20:47:41 +0100 Eric Dumazet wrote:
> On 1/29/21 8:35 PM, Jakub Kicinski wrote:

> 

> > kdoc didn't complain, and as you say it's already a mess, plus it's

> > two screen-fulls of scrolling away... 

> > 

> > I think converting to inline kdoc of members would be an improvement,

> > if you want to sign up for that? Otherwise -EDIDNTCARE on my side :)

> >   

> 

> What about removing this kdoc ?

> 

> kdoc for a huge structure is mostly useless...


It's definitely not useful for "us", I'd guess most seasoned developers
will just grep for uses of the field - but maybe it is useful for noobs
trying to have high-level sense of the code? 

Either way is fine by me, we can always preserve meaningful comments
inline without the kdoc decorator.
Jesper Dangaard Brouer Feb. 12, 2021, 3:49 p.m. | #7
On Fri, 29 Jan 2021 15:00:58 +0100
Jesper Dangaard Brouer <brouer@redhat.com> wrote:

> On Fri, 29 Jan 2021 14:33:02 +0100

> Eric Dumazet <eric.dumazet@gmail.com> wrote:

> 

> > On 1/26/21 6:39 PM, Jesper Dangaard Brouer wrote:  

> > > The current layout of net_device is not optimal for cacheline usage.

> > > 

[...]
> > > @@ -1877,6 +1876,23 @@ struct net_device {

[...]
> > > +

> > >  	netdev_features_t	features;

> > >  	netdev_features_t	hw_features;

> > >  	netdev_features_t	wanted_features;    

> > 

> > Probably wanted_features, hw_features are not used in fast path, only

> > in control path ?  

> 

> Yes, that was also my analysis.  I did consider moving those down, but

> I wanted to keep the first iteration simple ;-)


I've send a followup[1] to address this, thanks for pointing it out.

[1] https://patchwork.kernel.org/project/netdevbpf/patch/161313782625.1008639.6000589679659428869.stgit@firesoul/
-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer
Jesper Dangaard Brouer Feb. 12, 2021, 3:57 p.m. | #8
On Fri, 29 Jan 2021 12:07:23 -0800
Jakub Kicinski <kuba@kernel.org> wrote:

> On Fri, 29 Jan 2021 20:47:41 +0100 Eric Dumazet wrote:

> > On 1/29/21 8:35 PM, Jakub Kicinski wrote:

> >   

> > > kdoc didn't complain, and as you say it's already a mess, plus it's

> > > two screen-fulls of scrolling away... 

> > > 

> > > I think converting to inline kdoc of members would be an improvement,

> > > if you want to sign up for that? Otherwise -EDIDNTCARE on my side :)

> > >     

> > 

> > What about removing this kdoc ?

> > 

> > kdoc for a huge structure is mostly useless...  

> 

> It's definitely not useful for "us", I'd guess most seasoned developers

> will just grep for uses of the field - but maybe it is useful for noobs

> trying to have high-level sense of the code? 

> 

> Either way is fine by me, we can always preserve meaningful comments

> inline without the kdoc decorator.


I agree that removing this kdoc makes sense. But as Jakub says we
should preserve meaningful comments inline.  I'll add this task to my
TODO list, but anyone feel free to do the work before I get around to it.

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer

Patch

diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index b7915484369c..2645f114de54 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -1855,7 +1855,6 @@  struct net_device {
 	unsigned long		mem_end;
 	unsigned long		mem_start;
 	unsigned long		base_addr;
-	int			irq;
 
 	/*
 	 *	Some hardware also needs these fields (state,dev_list,
@@ -1877,6 +1876,23 @@  struct net_device {
 		struct list_head lower;
 	} adj_list;
 
+	/* Read-mostly cache-line for fast-path access */
+	unsigned int		flags;
+	unsigned int		priv_flags;
+	const struct net_device_ops *netdev_ops;
+	int			ifindex;
+	unsigned short		gflags;
+	unsigned short		hard_header_len;
+
+	/* Note : dev->mtu is often read without holding a lock.
+	 * Writers usually hold RTNL.
+	 * It is recommended to use READ_ONCE() to annotate the reads,
+	 * and to use WRITE_ONCE() to annotate the writes.
+	 */
+	unsigned int		mtu;
+	unsigned short		needed_headroom;
+	unsigned short		needed_tailroom;
+
 	netdev_features_t	features;
 	netdev_features_t	hw_features;
 	netdev_features_t	wanted_features;
@@ -1885,10 +1901,15 @@  struct net_device {
 	netdev_features_t	mpls_features;
 	netdev_features_t	gso_partial_features;
 
-	int			ifindex;
+	unsigned int		min_mtu;
+	unsigned int		max_mtu;
+	unsigned short		type;
+	unsigned char		min_header_len;
+	unsigned char		name_assign_type;
+
 	int			group;
 
-	struct net_device_stats	stats;
+	struct net_device_stats	stats; /* not used by modern drivers */
 
 	atomic_long_t		rx_dropped;
 	atomic_long_t		tx_dropped;
@@ -1902,7 +1923,6 @@  struct net_device {
 	const struct iw_handler_def *wireless_handlers;
 	struct iw_public_data	*wireless_data;
 #endif
-	const struct net_device_ops *netdev_ops;
 	const struct ethtool_ops *ethtool_ops;
 #ifdef CONFIG_NET_L3_MASTER_DEV
 	const struct l3mdev_ops	*l3mdev_ops;
@@ -1921,34 +1941,12 @@  struct net_device {
 
 	const struct header_ops *header_ops;
 
-	unsigned int		flags;
-	unsigned int		priv_flags;
-
-	unsigned short		gflags;
-	unsigned short		padded;
-
 	unsigned char		operstate;
 	unsigned char		link_mode;
 
 	unsigned char		if_port;
 	unsigned char		dma;
 
-	/* Note : dev->mtu is often read without holding a lock.
-	 * Writers usually hold RTNL.
-	 * It is recommended to use READ_ONCE() to annotate the reads,
-	 * and to use WRITE_ONCE() to annotate the writes.
-	 */
-	unsigned int		mtu;
-	unsigned int		min_mtu;
-	unsigned int		max_mtu;
-	unsigned short		type;
-	unsigned short		hard_header_len;
-	unsigned char		min_header_len;
-	unsigned char		name_assign_type;
-
-	unsigned short		needed_headroom;
-	unsigned short		needed_tailroom;
-
 	/* Interface address info. */
 	unsigned char		perm_addr[MAX_ADDR_LEN];
 	unsigned char		addr_assign_type;
@@ -1959,7 +1957,10 @@  struct net_device {
 	unsigned short		neigh_priv_len;
 	unsigned short          dev_id;
 	unsigned short          dev_port;
+	unsigned short		padded;
+
 	spinlock_t		addr_list_lock;
+	int			irq;
 
 	struct netdev_hw_addr_list	uc;
 	struct netdev_hw_addr_list	mc;