view Documentation/networking/ppp_generic.txt @ 897:329ea0ccb344

balloon: try harder to balloon up under memory pressure.

Currently if the balloon driver is unable to increase the guest's
reservation it assumes the failure was due to reaching its full
allocation, gives up on the ballooning operation and records the limit
it reached as the "hard limit". The driver will not try again until
the target is set again (even to the same value).

However it is possible that ballooning has in fact failed due to
memory pressure in the host and therefore it is desirable to keep
attempting to reach the target in case memory becomes available. The
most likely scenario is that some guests are ballooning down while
others are ballooning up and therefore there is temporary memory
pressure while things stabilise. You would not expect a well behaved
toolstack to ask a domain to balloon to more than its allocation nor
would you expect it to deliberately over-commit memory by setting
balloon targets which exceed the total host memory.

This patch drops the concept of a hard limit and causes the balloon
driver to retry increasing the reservation on a timer in the same
manner as when decreasing the reservation.

Also if we partially succeed in increasing the reservation
(i.e. receive less pages than we asked for) then we may as well keep
those pages rather than returning them to Xen.

Signed-off-by: Ian Campbell <ian.campbell@citrix.com>
author Keir Fraser <keir.fraser@citrix.com>
date Fri Jun 05 14:01:20 2009 +0100 (2009-06-05)
parents 831230e53067
line source
1 PPP Generic Driver and Channel Interface
2 ----------------------------------------
4 Paul Mackerras
5 paulus@samba.org
6 7 Feb 2002
8 The generic PPP driver in linux-2.4 provides an implementation of the
9 functionality which is of use in any PPP implementation, including:
11 * the network interface unit (ppp0 etc.)
12 * the interface to the networking code
13 * PPP multilink: splitting datagrams between multiple links, and
14 ordering and combining received fragments
15 * the interface to pppd, via a /dev/ppp character device
16 * packet compression and decompression
17 * TCP/IP header compression and decompression
18 * detecting network traffic for demand dialling and for idle timeouts
19 * simple packet filtering
21 For sending and receiving PPP frames, the generic PPP driver calls on
22 the services of PPP `channels'. A PPP channel encapsulates a
23 mechanism for transporting PPP frames from one machine to another. A
24 PPP channel implementation can be arbitrarily complex internally but
25 has a very simple interface with the generic PPP code: it merely has
26 to be able to send PPP frames, receive PPP frames, and optionally
27 handle ioctl requests. Currently there are PPP channel
28 implementations for asynchronous serial ports, synchronous serial
29 ports, and for PPP over ethernet.
31 This architecture makes it possible to implement PPP multilink in a
32 natural and straightforward way, by allowing more than one channel to
33 be linked to each ppp network interface unit. The generic layer is
34 responsible for splitting datagrams on transmit and recombining them
35 on receive.
38 PPP channel API
39 ---------------
41 See include/linux/ppp_channel.h for the declaration of the types and
42 functions used to communicate between the generic PPP layer and PPP
43 channels.
45 Each channel has to provide two functions to the generic PPP layer,
46 via the ppp_channel.ops pointer:
48 * start_xmit() is called by the generic layer when it has a frame to
49 send. The channel has the option of rejecting the frame for
50 flow-control reasons. In this case, start_xmit() should return 0
51 and the channel should call the ppp_output_wakeup() function at a
52 later time when it can accept frames again, and the generic layer
53 will then attempt to retransmit the rejected frame(s). If the frame
54 is accepted, the start_xmit() function should return 1.
56 * ioctl() provides an interface which can be used by a user-space
57 program to control aspects of the channel's behaviour. This
58 procedure will be called when a user-space program does an ioctl
59 system call on an instance of /dev/ppp which is bound to the
60 channel. (Usually it would only be pppd which would do this.)
62 The generic PPP layer provides seven functions to channels:
64 * ppp_register_channel() is called when a channel has been created, to
65 notify the PPP generic layer of its presence. For example, setting
66 a serial port to the PPPDISC line discipline causes the ppp_async
67 channel code to call this function.
69 * ppp_unregister_channel() is called when a channel is to be
70 destroyed. For example, the ppp_async channel code calls this when
71 a hangup is detected on the serial port.
73 * ppp_output_wakeup() is called by a channel when it has previously
74 rejected a call to its start_xmit function, and can now accept more
75 packets.
77 * ppp_input() is called by a channel when it has received a complete
78 PPP frame.
80 * ppp_input_error() is called by a channel when it has detected that a
81 frame has been lost or dropped (for example, because of a FCS (frame
82 check sequence) error).
84 * ppp_channel_index() returns the channel index assigned by the PPP
85 generic layer to this channel. The channel should provide some way
86 (e.g. an ioctl) to transmit this back to user-space, as user-space
87 will need it to attach an instance of /dev/ppp to this channel.
89 * ppp_unit_number() returns the unit number of the ppp network
90 interface to which this channel is connected, or -1 if the channel
91 is not connected.
93 Connecting a channel to the ppp generic layer is initiated from the
94 channel code, rather than from the generic layer. The channel is
95 expected to have some way for a user-level process to control it
96 independently of the ppp generic layer. For example, with the
97 ppp_async channel, this is provided by the file descriptor to the
98 serial port.
100 Generally a user-level process will initialize the underlying
101 communications medium and prepare it to do PPP. For example, with an
102 async tty, this can involve setting the tty speed and modes, issuing
103 modem commands, and then going through some sort of dialog with the
104 remote system to invoke PPP service there. We refer to this process
105 as `discovery'. Then the user-level process tells the medium to
106 become a PPP channel and register itself with the generic PPP layer.
107 The channel then has to report the channel number assigned to it back
108 to the user-level process. From that point, the PPP negotiation code
109 in the PPP daemon (pppd) can take over and perform the PPP
110 negotiation, accessing the channel through the /dev/ppp interface.
112 At the interface to the PPP generic layer, PPP frames are stored in
113 skbuff structures and start with the two-byte PPP protocol number.
114 The frame does *not* include the 0xff `address' byte or the 0x03
115 `control' byte that are optionally used in async PPP. Nor is there
116 any escaping of control characters, nor are there any FCS or framing
117 characters included. That is all the responsibility of the channel
118 code, if it is needed for the particular medium. That is, the skbuffs
119 presented to the start_xmit() function contain only the 2-byte
120 protocol number and the data, and the skbuffs presented to ppp_input()
121 must be in the same format.
123 The channel must provide an instance of a ppp_channel struct to
124 represent the channel. The channel is free to use the `private' field
125 however it wishes. The channel should initialize the `mtu' and
126 `hdrlen' fields before calling ppp_register_channel() and not change
127 them until after ppp_unregister_channel() returns. The `mtu' field
128 represents the maximum size of the data part of the PPP frames, that
129 is, it does not include the 2-byte protocol number.
131 If the channel needs some headroom in the skbuffs presented to it for
132 transmission (i.e., some space free in the skbuff data area before the
133 start of the PPP frame), it should set the `hdrlen' field of the
134 ppp_channel struct to the amount of headroom required. The generic
135 PPP layer will attempt to provide that much headroom but the channel
136 should still check if there is sufficient headroom and copy the skbuff
137 if there isn't.
139 On the input side, channels should ideally provide at least 2 bytes of
140 headroom in the skbuffs presented to ppp_input(). The generic PPP
141 code does not require this but will be more efficient if this is done.
144 Buffering and flow control
145 --------------------------
147 The generic PPP layer has been designed to minimize the amount of data
148 that it buffers in the transmit direction. It maintains a queue of
149 transmit packets for the PPP unit (network interface device) plus a
150 queue of transmit packets for each attached channel. Normally the
151 transmit queue for the unit will contain at most one packet; the
152 exceptions are when pppd sends packets by writing to /dev/ppp, and
153 when the core networking code calls the generic layer's start_xmit()
154 function with the queue stopped, i.e. when the generic layer has
155 called netif_stop_queue(), which only happens on a transmit timeout.
156 The start_xmit function always accepts and queues the packet which it
157 is asked to transmit.
159 Transmit packets are dequeued from the PPP unit transmit queue and
160 then subjected to TCP/IP header compression and packet compression
161 (Deflate or BSD-Compress compression), as appropriate. After this
162 point the packets can no longer be reordered, as the decompression
163 algorithms rely on receiving compressed packets in the same order that
164 they were generated.
166 If multilink is not in use, this packet is then passed to the attached
167 channel's start_xmit() function. If the channel refuses to take
168 the packet, the generic layer saves it for later transmission. The
169 generic layer will call the channel's start_xmit() function again
170 when the channel calls ppp_output_wakeup() or when the core
171 networking code calls the generic layer's start_xmit() function
172 again. The generic layer contains no timeout and retransmission
173 logic; it relies on the core networking code for that.
175 If multilink is in use, the generic layer divides the packet into one
176 or more fragments and puts a multilink header on each fragment. It
177 decides how many fragments to use based on the length of the packet
178 and the number of channels which are potentially able to accept a
179 fragment at the moment. A channel is potentially able to accept a
180 fragment if it doesn't have any fragments currently queued up for it
181 to transmit. The channel may still refuse a fragment; in this case
182 the fragment is queued up for the channel to transmit later. This
183 scheme has the effect that more fragments are given to higher-
184 bandwidth channels. It also means that under light load, the generic
185 layer will tend to fragment large packets across all the channels,
186 thus reducing latency, while under heavy load, packets will tend to be
187 transmitted as single fragments, thus reducing the overhead of
188 fragmentation.
191 SMP safety
192 ----------
194 The PPP generic layer has been designed to be SMP-safe. Locks are
195 used around accesses to the internal data structures where necessary
196 to ensure their integrity. As part of this, the generic layer
197 requires that the channels adhere to certain requirements and in turn
198 provides certain guarantees to the channels. Essentially the channels
199 are required to provide the appropriate locking on the ppp_channel
200 structures that form the basis of the communication between the
201 channel and the generic layer. This is because the channel provides
202 the storage for the ppp_channel structure, and so the channel is
203 required to provide the guarantee that this storage exists and is
204 valid at the appropriate times.
206 The generic layer requires these guarantees from the channel:
208 * The ppp_channel object must exist from the time that
209 ppp_register_channel() is called until after the call to
210 ppp_unregister_channel() returns.
212 * No thread may be in a call to any of ppp_input(), ppp_input_error(),
213 ppp_output_wakeup(), ppp_channel_index() or ppp_unit_number() for a
214 channel at the time that ppp_unregister_channel() is called for that
215 channel.
217 * ppp_register_channel() and ppp_unregister_channel() must be called
218 from process context, not interrupt or softirq/BH context.
220 * The remaining generic layer functions may be called at softirq/BH
221 level but must not be called from a hardware interrupt handler.
223 * The generic layer may call the channel start_xmit() function at
224 softirq/BH level but will not call it at interrupt level. Thus the
225 start_xmit() function may not block.
227 * The generic layer will only call the channel ioctl() function in
228 process context.
230 The generic layer provides these guarantees to the channels:
232 * The generic layer will not call the start_xmit() function for a
233 channel while any thread is already executing in that function for
234 that channel.
236 * The generic layer will not call the ioctl() function for a channel
237 while any thread is already executing in that function for that
238 channel.
240 * By the time a call to ppp_unregister_channel() returns, no thread
241 will be executing in a call from the generic layer to that channel's
242 start_xmit() or ioctl() function, and the generic layer will not
243 call either of those functions subsequently.
246 Interface to pppd
247 -----------------
249 The PPP generic layer exports a character device interface called
250 /dev/ppp. This is used by pppd to control PPP interface units and
251 channels. Although there is only one /dev/ppp, each open instance of
252 /dev/ppp acts independently and can be attached either to a PPP unit
253 or a PPP channel. This is achieved using the file->private_data field
254 to point to a separate object for each open instance of /dev/ppp. In
255 this way an effect similar to Solaris' clone open is obtained,
256 allowing us to control an arbitrary number of PPP interfaces and
257 channels without having to fill up /dev with hundreds of device names.
259 When /dev/ppp is opened, a new instance is created which is initially
260 unattached. Using an ioctl call, it can then be attached to an
261 existing unit, attached to a newly-created unit, or attached to an
262 existing channel. An instance attached to a unit can be used to send
263 and receive PPP control frames, using the read() and write() system
264 calls, along with poll() if necessary. Similarly, an instance
265 attached to a channel can be used to send and receive PPP frames on
266 that channel.
268 In multilink terms, the unit represents the bundle, while the channels
269 represent the individual physical links. Thus, a PPP frame sent by a
270 write to the unit (i.e., to an instance of /dev/ppp attached to the
271 unit) will be subject to bundle-level compression and to fragmentation
272 across the individual links (if multilink is in use). In contrast, a
273 PPP frame sent by a write to the channel will be sent as-is on that
274 channel, without any multilink header.
276 A channel is not initially attached to any unit. In this state it can
277 be used for PPP negotiation but not for the transfer of data packets.
278 It can then be connected to a PPP unit with an ioctl call, which
279 makes it available to send and receive data packets for that unit.
281 The ioctl calls which are available on an instance of /dev/ppp depend
282 on whether it is unattached, attached to a PPP interface, or attached
283 to a PPP channel. The ioctl calls which are available on an
284 unattached instance are:
286 * PPPIOCNEWUNIT creates a new PPP interface and makes this /dev/ppp
287 instance the "owner" of the interface. The argument should point to
288 an int which is the desired unit number if >= 0, or -1 to assign the
289 lowest unused unit number. Being the owner of the interface means
290 that the interface will be shut down if this instance of /dev/ppp is
291 closed.
293 * PPPIOCATTACH attaches this instance to an existing PPP interface.
294 The argument should point to an int containing the unit number.
295 This does not make this instance the owner of the PPP interface.
297 * PPPIOCATTCHAN attaches this instance to an existing PPP channel.
298 The argument should point to an int containing the channel number.
300 The ioctl calls available on an instance of /dev/ppp attached to a
301 channel are:
303 * PPPIOCDETACH detaches the instance from the channel. This ioctl is
304 deprecated since the same effect can be achieved by closing the
305 instance. In order to prevent possible races this ioctl will fail
306 with an EINVAL error if more than one file descriptor refers to this
307 instance (i.e. as a result of dup(), dup2() or fork()).
309 * PPPIOCCONNECT connects this channel to a PPP interface. The
310 argument should point to an int containing the interface unit
311 number. It will return an EINVAL error if the channel is already
312 connected to an interface, or ENXIO if the requested interface does
313 not exist.
315 * PPPIOCDISCONN disconnects this channel from the PPP interface that
316 it is connected to. It will return an EINVAL error if the channel
317 is not connected to an interface.
319 * All other ioctl commands are passed to the channel ioctl() function.
321 The ioctl calls that are available on an instance that is attached to
322 an interface unit are:
324 * PPPIOCSMRU sets the MRU (maximum receive unit) for the interface.
325 The argument should point to an int containing the new MRU value.
327 * PPPIOCSFLAGS sets flags which control the operation of the
328 interface. The argument should be a pointer to an int containing
329 the new flags value. The bits in the flags value that can be set
330 are:
331 SC_COMP_TCP enable transmit TCP header compression
332 SC_NO_TCP_CCID disable connection-id compression for
333 TCP header compression
334 SC_REJ_COMP_TCP disable receive TCP header decompression
335 SC_CCP_OPEN Compression Control Protocol (CCP) is
336 open, so inspect CCP packets
337 SC_CCP_UP CCP is up, may (de)compress packets
338 SC_LOOP_TRAFFIC send IP traffic to pppd
339 SC_MULTILINK enable PPP multilink fragmentation on
340 transmitted packets
341 SC_MP_SHORTSEQ expect short multilink sequence
342 numbers on received multilink fragments
343 SC_MP_XSHORTSEQ transmit short multilink sequence nos.
345 The values of these flags are defined in <linux/if_ppp.h>. Note
346 that the values of the SC_MULTILINK, SC_MP_SHORTSEQ and
347 SC_MP_XSHORTSEQ bits are ignored if the CONFIG_PPP_MULTILINK option
348 is not selected.
350 * PPPIOCGFLAGS returns the value of the status/control flags for the
351 interface unit. The argument should point to an int where the ioctl
352 will store the flags value. As well as the values listed above for
353 PPPIOCSFLAGS, the following bits may be set in the returned value:
354 SC_COMP_RUN CCP compressor is running
355 SC_DECOMP_RUN CCP decompressor is running
356 SC_DC_ERROR CCP decompressor detected non-fatal error
357 SC_DC_FERROR CCP decompressor detected fatal error
359 * PPPIOCSCOMPRESS sets the parameters for packet compression or
360 decompression. The argument should point to a ppp_option_data
361 structure (defined in <linux/if_ppp.h>), which contains a
362 pointer/length pair which should describe a block of memory
363 containing a CCP option specifying a compression method and its
364 parameters. The ppp_option_data struct also contains a `transmit'
365 field. If this is 0, the ioctl will affect the receive path,
366 otherwise the transmit path.
368 * PPPIOCGUNIT returns, in the int pointed to by the argument, the unit
369 number of this interface unit.
371 * PPPIOCSDEBUG sets the debug flags for the interface to the value in
372 the int pointed to by the argument. Only the least significant bit
373 is used; if this is 1 the generic layer will print some debug
374 messages during its operation. This is only intended for debugging
375 the generic PPP layer code; it is generally not helpful for working
376 out why a PPP connection is failing.
378 * PPPIOCGDEBUG returns the debug flags for the interface in the int
379 pointed to by the argument.
381 * PPPIOCGIDLE returns the time, in seconds, since the last data
382 packets were sent and received. The argument should point to a
383 ppp_idle structure (defined in <linux/ppp_defs.h>). If the
384 CONFIG_PPP_FILTER option is enabled, the set of packets which reset
385 the transmit and receive idle timers is restricted to those which
386 pass the `active' packet filter.
388 * PPPIOCSMAXCID sets the maximum connection-ID parameter (and thus the
389 number of connection slots) for the TCP header compressor and
390 decompressor. The lower 16 bits of the int pointed to by the
391 argument specify the maximum connection-ID for the compressor. If
392 the upper 16 bits of that int are non-zero, they specify the maximum
393 connection-ID for the decompressor, otherwise the decompressor's
394 maximum connection-ID is set to 15.
396 * PPPIOCSNPMODE sets the network-protocol mode for a given network
397 protocol. The argument should point to an npioctl struct (defined
398 in <linux/if_ppp.h>). The `protocol' field gives the PPP protocol
399 number for the protocol to be affected, and the `mode' field
400 specifies what to do with packets for that protocol:
402 NPMODE_PASS normal operation, transmit and receive packets
403 NPMODE_DROP silently drop packets for this protocol
404 NPMODE_ERROR drop packets and return an error on transmit
405 NPMODE_QUEUE queue up packets for transmit, drop received
406 packets
408 At present NPMODE_ERROR and NPMODE_QUEUE have the same effect as
411 * PPPIOCGNPMODE returns the network-protocol mode for a given
412 protocol. The argument should point to an npioctl struct with the
413 `protocol' field set to the PPP protocol number for the protocol of
414 interest. On return the `mode' field will be set to the network-
415 protocol mode for that protocol.
417 * PPPIOCSPASS and PPPIOCSACTIVE set the `pass' and `active' packet
418 filters. These ioctls are only available if the CONFIG_PPP_FILTER
419 option is selected. The argument should point to a sock_fprog
420 structure (defined in <linux/filter.h>) containing the compiled BPF
421 instructions for the filter. Packets are dropped if they fail the
422 `pass' filter; otherwise, if they fail the `active' filter they are
423 passed but they do not reset the transmit or receive idle timer.
425 * PPPIOCSMRRU enables or disables multilink processing for received
426 packets and sets the multilink MRRU (maximum reconstructed receive
427 unit). The argument should point to an int containing the new MRRU
428 value. If the MRRU value is 0, processing of received multilink
429 fragments is disabled. This ioctl is only available if the
430 CONFIG_PPP_MULTILINK option is selected.
432 Last modified: 7-feb-2002