xref: /freebsd/share/man/man9/mod_cc.9 (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1.\"
2.\" Copyright (c) 2008-2009 Lawrence Stewart <lstewart@FreeBSD.org>
3.\" Copyright (c) 2010-2011 The FreeBSD Foundation
4.\" All rights reserved.
5.\"
6.\" Portions of this documentation were written at the Centre for Advanced
7.\" Internet Architectures, Swinburne University of Technology, Melbourne,
8.\" Australia by David Hayes and Lawrence Stewart under sponsorship from the
9.\" FreeBSD Foundation.
10.\"
11.\" Redistribution and use in source and binary forms, with or without
12.\" modification, are permitted provided that the following conditions
13.\" are met:
14.\" 1. Redistributions of source code must retain the above copyright
15.\"    notice, this list of conditions and the following disclaimer.
16.\" 2. Redistributions in binary form must reproduce the above copyright
17.\"    notice, this list of conditions and the following disclaimer in the
18.\"    documentation and/or other materials provided with the distribution.
19.\"
20.\" THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21.\" ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22.\" IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23.\" ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
24.\" ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25.\" DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26.\" OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27.\" HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28.\" LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29.\" OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30.\" SUCH DAMAGE.
31.\"
32.Dd May 13, 2021
33.Dt MOD_CC 9
34.Os
35.Sh NAME
36.Nm mod_cc ,
37.Nm DECLARE_CC_MODULE ,
38.Nm CCV
39.Nd Modular Congestion Control
40.Sh SYNOPSIS
41.In netinet/tcp.h
42.In netinet/cc/cc.h
43.In netinet/cc/cc_module.h
44.Fn DECLARE_CC_MODULE "ccname" "ccalgo"
45.Fn CCV "ccv" "what"
46.Sh DESCRIPTION
47The
48.Nm
49framework allows congestion control algorithms to be implemented as dynamically
50loadable kernel modules via the
51.Xr kld 4
52facility.
53Transport protocols can select from the list of available algorithms on a
54connection-by-connection basis, or use the system default (see
55.Xr mod_cc 4
56for more details).
57.Pp
58.Nm
59modules are identified by an
60.Xr ascii 7
61name and set of hook functions encapsulated in a
62.Vt "struct cc_algo" ,
63which has the following members:
64.Bd -literal -offset indent
65struct cc_algo {
66	char	name[TCP_CA_NAME_MAX];
67	int	(*mod_init) (void);
68	int	(*mod_destroy) (void);
69	size_t  (*cc_data_sz)(void);
70	int	(*cb_init) (struct cc_var *ccv, void *ptr);
71	void	(*cb_destroy) (struct cc_var *ccv);
72	void	(*conn_init) (struct cc_var *ccv);
73	void	(*ack_received) (struct cc_var *ccv, uint16_t type);
74	void	(*cong_signal) (struct cc_var *ccv, uint32_t type);
75	void	(*post_recovery) (struct cc_var *ccv);
76	void	(*after_idle) (struct cc_var *ccv);
77	int	(*ctl_output)(struct cc_var *, struct sockopt *, void *);
78	void    (*rttsample)(struct cc_var *, uint32_t, uint32_t, uint32_t);
79	void    (*newround)(struct cc_var *, uint32_t);
80};
81.Ed
82.Pp
83The
84.Va name
85field identifies the unique name of the algorithm, and should be no longer than
86TCP_CA_NAME_MAX-1 characters in length (the TCP_CA_NAME_MAX define lives in
87.In netinet/tcp.h
88for compatibility reasons).
89.Pp
90The
91.Va mod_init
92function is called when a new module is loaded into the system but before the
93registration process is complete.
94It should be implemented if a module needs to set up some global state prior to
95being available for use by new connections.
96Returning a non-zero value from
97.Va mod_init
98will cause the loading of the module to fail.
99.Pp
100The
101.Va mod_destroy
102function is called prior to unloading an existing module from the kernel.
103It should be implemented if a module needs to clean up any global state before
104being removed from the kernel.
105The return value is currently ignored.
106.Pp
107The
108.Va cc_data_sz
109function is called by the socket option code to get the size of
110data that the
111.Va cb_init
112function needs.
113The socket option code then preallocates the modules memory so that the
114.Va cb_init
115function will not fail (the socket option code uses M_WAITOK with
116no locks held to do this).
117.Pp
118The
119.Va cb_init
120function is called when a TCP control block
121.Vt struct tcpcb
122is created.
123It should be implemented if a module needs to allocate memory for storing
124private per-connection state.
125Returning a non-zero value from
126.Va cb_init
127will cause the connection set up to be aborted, terminating the connection as a
128result.
129Note that the ptr argument passed to the function should be checked to
130see if it is non-NULL, if so it is preallocated memory that the cb_init function
131must use instead of calling malloc itself.
132.Pp
133The
134.Va cb_destroy
135function is called when a TCP control block
136.Vt struct tcpcb
137is destroyed.
138It should be implemented if a module needs to free memory allocated in
139.Va cb_init .
140.Pp
141The
142.Va conn_init
143function is called when a new connection has been established and variables are
144being initialised.
145It should be implemented to initialise congestion control algorithm variables
146for the newly established connection.
147.Pp
148The
149.Va ack_received
150function is called when a TCP acknowledgement (ACK) packet is received.
151Modules use the
152.Fa type
153argument as an input to their congestion management algorithms.
154The ACK types currently reported by the stack are CC_ACK and CC_DUPACK.
155CC_ACK indicates the received ACK acknowledges previously unacknowledged data.
156CC_DUPACK indicates the received ACK acknowledges data we have already received
157an ACK for.
158.Pp
159The
160.Va cong_signal
161function is called when a congestion event is detected by the TCP stack.
162Modules use the
163.Fa type
164argument as an input to their congestion management algorithms.
165The congestion event types currently reported by the stack are CC_ECN, CC_RTO,
166CC_RTO_ERR and CC_NDUPACK.
167CC_ECN is reported when the TCP stack receives an explicit congestion notification
168(RFC3168).
169CC_RTO is reported when the retransmission time out timer fires.
170CC_RTO_ERR is reported if the retransmission time out timer fired in error.
171CC_NDUPACK is reported if N duplicate ACKs have been received back-to-back,
172where N is the fast retransmit duplicate ack threshold (N=3 currently as per
173RFC5681).
174.Pp
175The
176.Va post_recovery
177function is called after the TCP connection has recovered from a congestion event.
178It should be implemented to adjust state as required.
179.Pp
180The
181.Va after_idle
182function is called when data transfer resumes after an idle period.
183It should be implemented to adjust state as required.
184.Pp
185The
186.Va ctl_output
187function is called when
188.Xr getsockopt 2
189or
190.Xr setsockopt 2
191is called on a
192.Xr tcp 4
193socket with the
194.Va struct sockopt
195pointer forwarded unmodified from the TCP control, and a
196.Va void *
197pointer to algorithm specific argument.
198.Pp
199The
200.Va rttsample
201function is called to pass round trip time information to the
202congestion controller.
203The additional arguments to the function include the microsecond RTT
204that is being noted, the number of times that the data being
205acknowledged was retransmitted as well as the flightsize at send.
206For transports that do not track flightsize at send, this variable
207will be the current cwnd at the time of the call.
208.Pp
209The
210.Va newround
211function is called each time a new round trip time begins.
212The montonically increasing round number is also passed to the
213congestion controller as well.
214This can be used for various purposes by the congestion controller (e.g Hystart++).
215.Pp
216Note that currently not all TCP stacks call the
217.Va rttsample
218and
219.Va newround
220function so dependency on these functions is also
221dependent upon which TCP stack is in use.
222.Pp
223The
224.Fn DECLARE_CC_MODULE
225macro provides a convenient wrapper around the
226.Xr DECLARE_MODULE 9
227macro, and is used to register a
228.Nm
229module with the
230.Nm
231framework.
232The
233.Fa ccname
234argument specifies the module's name.
235The
236.Fa ccalgo
237argument points to the module's
238.Vt struct cc_algo .
239.Pp
240.Nm
241modules must instantiate a
242.Vt struct cc_algo ,
243but are only required to set the name field, and optionally any of the function
244pointers.
245Note that if a module defines the
246.Va cb_init
247function it also must define a
248.Va cc_data_sz
249function.
250This is because when switching from one congestion control
251module to another the socket option code will preallocate memory for the
252.Va cb_init
253function.
254If no memory is allocated by the modules
255.Va cb_init
256then the
257.Va cc_data_sz
258function should return 0.
259.Pp
260The stack will skip calling any function pointer which is NULL, so there is no
261requirement to implement any of the function pointers (with the exception of
262the cb_init <-> cc_data_sz dependency noted above).
263Using the C99 designated initialiser feature to set fields is encouraged.
264.Pp
265Each function pointer which deals with congestion control state is passed a
266pointer to a
267.Vt struct cc_var ,
268which has the following members:
269.Bd -literal -offset indent
270struct cc_var {
271	void		*cc_data;
272	int		bytes_this_ack;
273	tcp_seq		curack;
274	uint32_t	flags;
275	int		type;
276	union ccv_container {
277		struct tcpcb		*tcp;
278		struct sctp_nets	*sctp;
279	} ccvc;
280	uint16_t	nsegs;
281	uint8_t		labc;
282};
283.Ed
284.Pp
285.Vt struct cc_var
286groups congestion control related variables into a single, embeddable structure
287and adds a layer of indirection to accessing transport protocol control blocks.
288The eventual goal is to allow a single set of
289.Nm
290modules to be shared between all congestion aware transport protocols, though
291currently only
292.Xr tcp 4
293is supported.
294.Pp
295To aid the eventual transition towards this goal, direct use of variables from
296the transport protocol's data structures is strongly discouraged.
297However, it is inevitable at the current time to require access to some of these
298variables, and so the
299.Fn CCV
300macro exists as a convenience accessor.
301The
302.Fa ccv
303argument points to the
304.Vt struct cc_var
305passed into the function by the
306.Nm
307framework.
308The
309.Fa what
310argument specifies the name of the variable to access.
311.Pp
312Apart from the
313.Va type
314and
315.Va ccv_container
316fields, the remaining fields in
317.Vt struct cc_var
318are for use by
319.Nm
320modules.
321.Pp
322The
323.Va cc_data
324field is available for algorithms requiring additional per-connection state to
325attach a dynamic memory pointer to.
326The memory should be allocated and attached in the module's
327.Va cb_init
328hook function.
329.Pp
330The
331.Va bytes_this_ack
332field specifies the number of new bytes acknowledged by the most recently
333received ACK packet.
334It is only valid in the
335.Va ack_received
336hook function.
337.Pp
338The
339.Va curack
340field specifies the sequence number of the most recently received ACK packet.
341It is only valid in the
342.Va ack_received ,
343.Va cong_signal
344and
345.Va post_recovery
346hook functions.
347.Pp
348The
349.Va flags
350field is used to pass useful information from the stack to a
351.Nm
352module.
353The CCF_ABC_SENTAWND flag is relevant in
354.Va ack_received
355and is set when appropriate byte counting (RFC3465) has counted a window's worth
356of bytes has been sent.
357It is the module's responsibility to clear the flag after it has processed the
358signal.
359The CCF_CWND_LIMITED flag is relevant in
360.Va ack_received
361and is set when the connection's ability to send data is currently constrained
362by the value of the congestion window.
363Algorithms should use the absence of this flag being set to avoid accumulating
364a large difference between the congestion window and send window.
365.Pp
366The
367.Va nsegs
368variable is used to pass in how much compression was done by the local
369LRO system.
370So for example if LRO pushed three in-order acknowledgements into
371one acknowledgement the variable would be set to three.
372.Pp
373The
374.Va labc
375variable is used in conjunction with the CCF_USE_LOCAL_ABC flag
376to override what labc variable the congestion controller will use
377for this particular acknowledgement.
378.Sh SEE ALSO
379.Xr cc_cdg 4 ,
380.Xr cc_chd 4 ,
381.Xr cc_cubic 4 ,
382.Xr cc_dctcp 4 ,
383.Xr cc_hd 4 ,
384.Xr cc_htcp 4 ,
385.Xr cc_newreno 4 ,
386.Xr cc_vegas 4 ,
387.Xr mod_cc 4 ,
388.Xr tcp 4
389.Sh ACKNOWLEDGEMENTS
390Development and testing of this software were made possible in part by grants
391from the FreeBSD Foundation and Cisco University Research Program Fund at
392Community Foundation Silicon Valley.
393.Sh FUTURE WORK
394Integrate with
395.Xr sctp 4 .
396.Sh HISTORY
397The modular Congestion Control (CC) framework first appeared in
398.Fx 9.0 .
399.Pp
400The framework was first released in 2007 by James Healy and Lawrence Stewart
401whilst working on the NewTCP research project at Swinburne University of
402Technology's Centre for Advanced Internet Architectures, Melbourne, Australia,
403which was made possible in part by a grant from the Cisco University Research
404Program Fund at Community Foundation Silicon Valley.
405More details are available at:
406.Pp
407http://caia.swin.edu.au/urp/newtcp/
408.Sh AUTHORS
409.An -nosplit
410The
411.Nm
412framework was written by
413.An Lawrence Stewart Aq Mt lstewart@FreeBSD.org ,
414.An James Healy Aq Mt jimmy@deefa.com
415and
416.An David Hayes Aq Mt david.hayes@ieee.org .
417.Pp
418This manual page was written by
419.An David Hayes Aq Mt david.hayes@ieee.org
420and
421.An Lawrence Stewart Aq Mt lstewart@FreeBSD.org .
422