35c05a4fbc
I've used this in a handful of RSS test applications. It is just some very simple functions to fetch the RSS configuration, query the per-bucket CPU set, and mark sockets as local to an RSS bucket. It should be sufficient for both thread-based and process-based workloads. (Yes, I wrote a manpage.) This is based on some early RSS API and wrapper API work I did whilst I was at Netflix. Thanks to Netflix for the very original work that spawned this; thanks to Peter Grehan for his feedback about RSS APIs and thanks to Jack Vogel and Navdeep Parhar for the NIC-facing side of the APIs. These fed into the simple userland API I wrote up here. Reviewed by: gallatin
154 lines
4.0 KiB
Groff
154 lines
4.0 KiB
Groff
.\" $FreeBSD$
|
|
.\"
|
|
.Dd September 29, 2016
|
|
.Dt LIBRSS 3
|
|
.Os
|
|
.Sh NAME
|
|
.Nm librss
|
|
.Nd Provide Receive-side scaling awareness to userland applications
|
|
.Sh LIBRARY
|
|
.Lb librss
|
|
.Sh SYNOPSIS
|
|
.In librss.h
|
|
.Ft struct rss_config *
|
|
.Fn rss_config_get "void"
|
|
.Ft void
|
|
.Fn rss_config_free "struct rss_config *cfg"
|
|
.Ft int
|
|
.Fn rss_config_get_bucket_count "struct rss_config *cfg"
|
|
.Ft int
|
|
.Fn rss_set_bucket_rebalance_cb "rss_bucket_rebalance_cb_t *cb" "void *cbdata"
|
|
.Ft int
|
|
.Fn rss_sock_set_bindmulti "int fd" "int af" "int val"
|
|
.Ft int
|
|
.Fn rss_sock_set_rss_bucket "int fd" "int af" "int rss_bucket"
|
|
.Ft int
|
|
.Fn rss_sock_set_recvrss "int fd" "int af" "int val"
|
|
.Sh DESCRIPTION
|
|
The
|
|
.Nm
|
|
library and the functions it provides are used for both fetching
|
|
the system RSS configuration and interacting with RSS aware
|
|
sockets.
|
|
.Pp
|
|
Applications will typically call
|
|
.Fn rss_config_get
|
|
to fetch the current RSS configuration from the system and perform
|
|
initial setup.
|
|
This typically involves spawning worker threads, one per RSS bucket,
|
|
and optionally binding them to the per-bucket CPU set.
|
|
.Pp
|
|
The
|
|
.Vt rss_config
|
|
struct is defined as:
|
|
.Bd -literal
|
|
struct rss_config {
|
|
int rss_ncpus;
|
|
int rss_nbuckets;
|
|
int rss_basecpu;
|
|
int *rss_bucket_map;
|
|
};
|
|
.Ed
|
|
.Pp
|
|
Applications will typically use the
|
|
.Fn rss_config_get_bucket_count
|
|
function to fetch the number of RSS buckets, create one thread
|
|
per RSS bucket for RSS aware work, then one RSS aware socket to receive
|
|
UDP datagrams or TCP connections
|
|
in each particular RSS bucket / thread.
|
|
.Pp
|
|
The
|
|
.Fn rss_get_bucket_cpuset
|
|
function sets the given cpuset up for the given
|
|
RSS bucket and behaviour.
|
|
Typically applications will wish to just query for
|
|
.Vt RSS_BUCKET_TYPE_KERNEL_ALL
|
|
unless they wish to potentially setup different
|
|
worker threads for transmit and receive.
|
|
.Pp
|
|
The
|
|
.Vt rss_bucket_type_t
|
|
enum is defined as:
|
|
.Bd -literal
|
|
typedef enum {
|
|
RSS_BUCKET_TYPE_NONE = 0,
|
|
RSS_BUCKET_TYPE_KERNEL_ALL = 1,
|
|
RSS_BUCKET_TYPE_KERNEL_TX = 2,
|
|
RSS_BUCKET_TYPE_KERNEL_RX = 3,
|
|
RSS_BUCKET_TYPE_MAX = 3,
|
|
} rss_bucket_type_t;
|
|
.Ed
|
|
.Pp
|
|
The rebalance callback
|
|
.Vt rss_bucket_rebalance_cb_t
|
|
is defined as:
|
|
.Bd -literal
|
|
typedef void rss_bucket_rebalance_cb_t(void *arg);
|
|
.Ed
|
|
.Pp
|
|
The
|
|
.Fn rss_set_bucket_rebalance_cb
|
|
function sets an optional callback that will be called if the kernel
|
|
rebalances RSS buckets.
|
|
This is intended as a future expansion to rebalance buckets rather than
|
|
reprogram the RSS key, so typically the only work to be performed
|
|
is to rebind worker threads to an updated cpuset.
|
|
.Pp
|
|
Once RSS setup is completed,
|
|
.Fn rss_config_free
|
|
is called to free the RSS configuration structure.
|
|
.Pp
|
|
To make a
|
|
.Vt bind
|
|
socket RSS aware, the
|
|
.Fn rss_sock_set_bindmulti
|
|
function is used to enable or disable per-RSS bucket
|
|
behaviour.
|
|
The socket filedescriptor, address family and enable flag
|
|
.Vt val
|
|
are passed in.
|
|
.Pp
|
|
If
|
|
.Vt val
|
|
is set to 1, the socket can be placed in an RSS bucket and will only accept
|
|
datagrams (for UDP) or connections (for TCP) that are received for that
|
|
RSS bucket.
|
|
If set to 0, the socket is placed in the default PCB and will see
|
|
datagrams/connections that are not initially consumed by a PCB aware
|
|
socket.
|
|
.Pp
|
|
The
|
|
.Fn rss_sock_set_rss_bucket
|
|
function configures the RSS bucket which a socket belongs in.
|
|
Note that TCP sockets created by
|
|
.Xr accept 2
|
|
will automatically be assigned to the RSS bucket.
|
|
.Pp
|
|
The
|
|
.Fn rss_sock_set_recvrss
|
|
function enables or disables receiving RSS related information
|
|
as socket options in.
|
|
.2 recvmsg
|
|
calls.
|
|
.Pp
|
|
When enabled, UDP datagrams will have a message with the
|
|
.Vt IP_RECVFLOWID
|
|
option indicating the 32-bit receive flowid as a uint32_t,
|
|
and the
|
|
.Vt IP_RECVRSSBUCKETID
|
|
option indicating the 32 bit RSS bucket id as a uint32_t.
|
|
.Sh ERRORS
|
|
The functions return either <0 or NULL as appropriate upon error.
|
|
.Sh SEE ALSO
|
|
.Xr PCBGROUP 9
|
|
.Sh HISTORY
|
|
The
|
|
.Xr librss.3
|
|
library first appeared in
|
|
.Fx 11.0 .
|
|
.Sh AUTHORS
|
|
.An Adrian Chadd Aq Mt adrian@FreeBSD.org
|
|
.Sh BUGS
|
|
There is currently no kernel mechanism to rebalance the RSS bucket to CPU
|
|
mapping, and so the callback mechanism is a no-op.
|