new ifnet(9) KPI for FreeBSD network stack

New ifnet(9) KPI
Gleb Smirnoﬀ
glebius@FreeBSD.org
BSDCan 2015
Ottawa
June 11, 2015
Gleb Smirnoﬀ glebius@FreeBSD.org New ifnet(9) KPI June 11, 2015 1 / 24

Introduction
project wiki page and code
https://wiki.freebsd.org/projects/ifnet
svn+ssh://svn.freebsd.org/base/projects/ifnet

Introduction
struct ifnet
Glue between a driver and the stack
Used to be embedded into driver softc, now softc and
ifnet has pointers to each other

Introduction
struct ifnet
Glue between a driver and the stack
Used to be embedded into driver softc, now softc and
ifnet has pointers to each other
At attach time driver and stack fill in the structure
At run time driver changes different flags

Introduction
Why changes are needed
Extension of the struct ifnet could require
recompilation of all drivers
Editing struct ifnet could require patching all drivers

opaque ifnet
The opaque ifnet
Code snippet 1: net/if.h:
typedef struct i f n e t ∗ if_t ;
Code snippet 2: net/if_var.h:
struct i f n e t {
/∗ actual s t r u c t u r e d e f i n i t i o n ∗/
};

opaque ifnet
The opaque ifnet
Easy way: substitute any access to struct ifnet ﬁelds
with a function

opaque ifnet
The opaque ifnet
Easy way: substitute any access to struct ifnet ﬁelds
with a function
Hard way: design new KPI

new KPI what’s in struct ifnet?
what’s in struct ifnet?
driver
access
instance
speciﬁc
if_ﬂags, if_capenable, if_mtu R y
link state, baudrate RW y
counters W y
if_capabilities, if_tso* W once y/n
address lists R y
driver name, if_clone W once n
methods W once n
if_type, dlt_type, header len W once n
if_media W once * y/n

what’s else in struct ifnet?
The stuﬀ the driver shouldn’t be interested at
Layer 2 softc: if_l2com
Address family softcs: if_afdata[AF_MAX]
Bunch of other softcs: if_lagg, if_carp, if_netmap,
etc

what’s else in struct ifnet?
The stuﬀ the driver shouldn’t be interested at
Layer 2 softc: if_l2com
Address family softcs: if_afdata[AF_MAX]
Bunch of other softcs: if_lagg, if_carp, if_netmap,
etc
This all can be generalized!

new KPI allocation and attachment
new if_attach()
if_alloc() if_attach(if_attach_args *args)
sleeps, doesn’t fail (save name conﬂict)
if_attach_args is versioned
if_attach_args contains all the “W once” stuﬀ
ifdriver pointer
softc pointer
lladdr
supported media list
capabilities, TSO limits
initial values: MTU, capenable, baudrate, etc

ifdriver, ifops
struct ifdriver is static in the driver, storing all
non-instance speciﬁc stuﬀ
Interface methods in special structure ifops
Name, type, DLT, headerlength

ifmedia
Code snippet 3: net/if_media.h:
typedef int if_media_t ;
Drivers declare static/dynamic array of if_media_t,
pointed to from if_attach_args
Drivers declare if_media_change_t,
if_media_status_t in ifops
Implementation opaque and private to if_media.c

ifmedia + miibus
miibus(4) is completely ifnet(9) agnostic
mii_attach() allocates pointer to if_media_t array,
to be used later in if_attach_args

new KPI method changes
if_init is no longer a method, becomes static function
if_poll for polling(4)
if_start
if_transmit doesn’t m_freem() in case of error

new KPI runtime modified data
if_flags, if_capenable, if_mtu (R,
instance specific)
if_ioctl method is the only channel to modify the
values
stack does sanity checking
driver may refuse new value
if driver accepts value, it may cache it

if_drv_ﬂags
IFF_DRV_OACTIVE goes away together with
if_start and generic queue
IFF_DRV_RUNNING goes to driver softc, protected
by driver lock

link state, baudrate
void if_setbaudrate ( if_t , uint64_t ) ;
void if_link_state_change ( if_t , int ) ;

counters
if_inc_counter() is already in head
Driver is 100% resposible for the counters
TX counters are updated on TX completion, not on
enqueue
Use if_inc_txcounters() for TX update

traversing address lists
typedef void
ifaddr_cb_t ( void ∗ , struct sockaddr ∗ ,
struct sockaddr ∗ , struct sockaddr ∗);
typedef void
ifmaddr_cb_t ( void ∗ , struct sockaddr ∗);
void if_foreach_addr ( if_t , ifaddr_cb_t ,
void ∗);
void if_foreach_maddr ( if_t , ifmaddr_cb_t ,
void ∗);

new KPI special considerations
net80211 drivers
No ifnet layer in the driver
Task is 50% done
Will be committed to head separately
https://wiki.freebsd.org/projects/ifnet/net80211

drivers that do if_start
That’s 80-90% of all legacy drivers
if_start and struct ifqueue go away
driver opts-in for a generic queue in if_attach
stack provides: if_snd_len(), if_snd_enqueue(),
if_snd_dequeue() and if_snd_prepend()
driver’s if_transmit is a short copy&paste

lagg(4)
now lagg(4) hijacks if_transmit of an interface
it will hijack ifops
ifops can be stacked, like VOPs

ALTQ
now ALTQ works on top of ifqueue
it will hijack ifops

new KPI scope of work
scope of work
There are > 200 drivers to be converted
Only 16 has been converted so far
Conversion of typical 100Mbit driver takes 1 hour
Usually driver is reduced by 100 LOC
https://wiki.freebsd.org/projects/ifnet/progress

new KPI scope of work
open tasks
Better generic queueing/transmit code?
ALTQ
lagg(4)
VIMAGE

Conclusion
Your feedback & help is needed!
Reviewing
Criticizing, proposing better KPIs
Converting drivers
Testing

new ifnet(9) KPI for FreeBSD network stack

More Related Content

What's hot (20)

Similar to new ifnet(9) KPI for FreeBSD network stack (20)

Recently uploaded (20)

new ifnet(9) KPI for FreeBSD network stack