view Documentation/networking/arcnet.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 ----------------------------------------------------------------------------
2 NOTE: See also arcnet-hardware.txt in this directory for jumper-setting
3 and cabling information if you're like many of us and didn't happen to get a
4 manual with your ARCnet card.
5 ----------------------------------------------------------------------------
7 Since no one seems to listen to me otherwise, perhaps a poem will get your
8 attention:
9 This driver's getting fat and beefy,
10 But my cat is still named Fifi.
12 Hmm, I think I'm allowed to call that a poem, even though it's only two
13 lines. Hey, I'm in Computer Science, not English. Give me a break.
15 The point is: I REALLY REALLY REALLY REALLY REALLY want to hear from you if
16 you test this and get it working. Or if you don't. Or anything.
18 ARCnet 0.32 ALPHA first made it into the Linux kernel 1.1.80 - this was
19 nice, but after that even FEWER people started writing to me because they
20 didn't even have to install the patch. <sigh>
22 Come on, be a sport! Send me a success report!
24 (hey, that was even better than my original poem... this is getting bad!)
27 --------
29 --------
31 If you don't e-mail me about your success/failure soon, I may be forced to
32 start SINGING. And we don't want that, do we?
34 (You know, it might be argued that I'm pushing this point a little too much.
35 If you think so, why not flame me in a quick little e-mail? Please also
36 include the type of card(s) you're using, software, size of network, and
37 whether it's working or not.)
39 My e-mail address is: apenwarr@worldvisions.ca
42 ---------------------------------------------------------------------------
45 These are the ARCnet drivers for Linux.
48 This new release (2.91) has been put together by David Woodhouse
49 <dwmw2@cam.ac.uk>, in an attempt to tidy up the driver after adding support
50 for yet another chipset. Now the generic support has been separated from the
51 individual chipset drivers, and the source files aren't quite so packed with
52 #ifdefs! I've changed this file a bit, but kept it in the first person from
53 Avery, because I didn't want to completely rewrite it.
55 The previous release resulted from many months of on-and-off effort from me
56 (Avery Pennarun), many bug reports/fixes and suggestions from others, and in
57 particular a lot of input and coding from Tomasz Motylewski. Starting with
58 ARCnet 2.10 ALPHA, Tomasz's all-new-and-improved RFC1051 support has been
59 included and seems to be working fine!
62 Where do I discuss these drivers?
63 ---------------------------------
65 Tomasz has been so kind as to set up a new and improved mailing list.
66 Subscribe by sending a message with the BODY "subscribe linux-arcnet YOUR
67 REAL NAME" to listserv@tichy.ch.uj.edu.pl. Then, to submit messages to the
68 list, mail to linux-arcnet@tichy.ch.uj.edu.pl.
70 There are archives of the mailing list at:
71 http://tichy.ch.uj.edu.pl/lists/linux-arcnet
73 The people on linux-net@vger.kernel.org have also been known to be very
74 helpful, especially when we're talking about ALPHA Linux kernels that may or
75 may not work right in the first place.
78 Other Drivers and Info
79 ----------------------
81 You can try my ARCNET page on the World Wide Web at:
82 http://www.worldvisions.ca/~apenwarr/arcnet/
84 Also, SMC (one of the companies that makes ARCnet cards) has a WWW site you
85 might be interested in, which includes several drivers for various cards
86 including ARCnet. Try:
87 http://www.smc.com/
89 Performance Technologies makes various network software that supports
90 ARCnet:
91 http://www.perftech.com/ or ftp to ftp.perftech.com.
93 Novell makes a networking stack for DOS which includes ARCnet drivers. Try
94 FTPing to ftp.novell.com.
96 You can get the Crynwr packet driver collection (including arcether.com, the
97 one you'll want to use with ARCnet cards) from
98 oak.oakland.edu:/simtel/msdos/pktdrvr. It won't work perfectly on a 386+
99 without patches, though, and also doesn't like several cards. Fixed
100 versions are available on my WWW page, or via e-mail if you don't have WWW
101 access.
104 Installing the Driver
105 ---------------------
107 All you will need to do in order to install the driver is:
108 make config
109 (be sure to choose ARCnet in the network devices
110 and at least one chipset driver.)
111 make clean
112 make zImage
114 If you obtained this ARCnet package as an upgrade to the ARCnet driver in
115 your current kernel, you will need to first copy arcnet.c over the one in
116 the linux/drivers/net directory.
118 You will know the driver is installed properly if you get some ARCnet
119 messages when you reboot into the new Linux kernel.
121 There are four chipset options:
123 1. Standard ARCnet COM90xx chipset.
125 This is the normal ARCnet card, which you've probably got. This is the only
126 chipset driver which will autoprobe if not told where the card is.
127 It following options on the command line:
128 com90xx=[<io>[,<irq>[,<shmem>]]][,<name>] | <name>
130 If you load the chipset support as a module, the options are:
131 io=<io> irq=<irq> shmem=<shmem> device=<name>
133 To disable the autoprobe, just specify "com90xx=" on the kernel command line.
134 To specify the name alone, but allow autoprobe, just put "com90xx=<name>"
136 2. ARCnet COM20020 chipset.
138 This is the new chipset from SMC with support for promiscuous mode (packet
139 sniffing), extra diagnostic information, etc. Unfortunately, there is no
140 sensible method of autoprobing for these cards. You must specify the I/O
141 address on the kernel command line.
142 The command line options are:
143 com20020=<io>[,<irq>[,<node_ID>[,backplane[,CKP[,timeout]]]]][,name]
145 If you load the chipset support as a module, the options are:
146 io=<io> irq=<irq> node=<node_ID> backplane=<backplane> clock=<CKP>
147 timeout=<timeout> device=<name>
149 The COM20020 chipset allows you to set the node ID in software, overriding the
150 default which is still set in DIP switches on the card. If you don't have the
151 COM20020 data sheets, and you don't know what the other three options refer
152 to, then they won't interest you - forget them.
154 3. ARCnet COM90xx chipset in IO-mapped mode.
156 This will also work with the normal ARCnet cards, but doesn't use the shared
157 memory. It performs less well than the above driver, but is provided in case
158 you have a card which doesn't support shared memory, or (strangely) in case
159 you have so many ARCnet cards in your machine that you run out of shmem slots.
160 If you don't give the IO address on the kernel command line, then the driver
161 will not find the card.
162 The command line options are:
163 com90io=<io>[,<irq>][,<name>]
165 If you load the chipset support as a module, the options are:
166 io=<io> irq=<irq> device=<name>
168 4. ARCnet RIM I cards.
170 These are COM90xx chips which are _completely_ memory mapped. The support for
171 these is not tested. If you have one, please mail the author with a success
172 report. All options must be specified, except the device name.
173 Command line options:
174 arcrimi=<shmem>,<irq>,<node_ID>[,<name>]
176 If you load the chipset support as a module, the options are:
177 shmem=<shmem> irq=<irq> node=<node_ID> device=<name>
180 Loadable Module Support
181 -----------------------
183 Configure and rebuild Linux. When asked, answer 'm' to "Generic ARCnet
184 support" and to support for your ARCnet chipset if you want to use the
185 loadable module. You can also say 'y' to "Generic ARCnet support" and 'm'
186 to the chipset support if you wish.
188 make config
189 make clean
190 make zImage
191 make modules
193 If you're using a loadable module, you need to use insmod to load it, and
194 you can specify various characteristics of your card on the command
195 line. (In recent versions of the driver, autoprobing is much more reliable
196 and works as a module, so most of this is now unnecessary.)
198 For example:
199 cd /usr/src/linux/modules
200 insmod arcnet.o
201 insmod com90xx.o
202 insmod com20020.o io=0x2e0 device=eth1
205 Using the Driver
206 ----------------
208 If you build your kernel with ARCnet COM90xx support included, it should
209 probe for your card automatically when you boot. If you use a different
210 chipset driver complied into the kernel, you must give the necessary options
211 on the kernel command line, as detailed above.
213 Go read the NET-2-HOWTO and ETHERNET-HOWTO for Linux; they should be
214 available where you picked up this driver. Think of your ARCnet as a
215 souped-up (or down, as the case may be) Ethernet card.
217 By the way, be sure to change all references from "eth0" to "arc0" in the
218 HOWTOs. Remember that ARCnet isn't a "true" Ethernet, and the device name
222 Multiple Cards in One Computer
223 ------------------------------
225 Linux has pretty good support for this now, but since I've been busy, the
226 ARCnet driver has somewhat suffered in this respect. COM90xx support, if
227 compiled into the kernel, will (try to) autodetect all the installed cards.
229 If you have other cards, with support compiled into the kernel, then you can
230 just repeat the options on the kernel command line, e.g.:
231 LILO: linux com20020=0x2e0 com20020=0x380 com90io=0x260
233 If you have the chipset support built as a loadable module, then you need to
234 do something like this:
235 insmod -o arc0 com90xx
236 insmod -o arc1 com20020 io=0x2e0
237 insmod -o arc2 com90xx
238 The ARCnet drivers will now sort out their names automatically.
241 How do I get it to work with...?
242 --------------------------------
244 NFS: Should be fine linux->linux, just pretend you're using Ethernet cards.
245 oak.oakland.edu:/simtel/msdos/nfs has some nice DOS clients. There
246 is also a DOS-based NFS server called SOSS. It doesn't multitask
247 quite the way Linux does (actually, it doesn't multitask AT ALL) but
248 you never know what you might need.
250 With AmiTCP (and possibly others), you may need to set the following
251 options in your Amiga nfstab: MD 1024 MR 1024 MW 1024
252 (Thanks to Christian Gottschling <ferksy@indigo.tng.oche.de>
253 for this.)
255 Probably these refer to maximum NFS data/read/write block sizes. I
256 don't know why the defaults on the Amiga didn't work; write to me if
257 you know more.
259 DOS: If you're using the freeware arcether.com, you might want to install
260 the driver patch from my web page. It helps with PC/TCP, and also
261 can get arcether to load if it timed out too quickly during
262 initialization. In fact, if you use it on a 386+ you REALLY need
263 the patch, really.
265 Windows: See DOS :) Trumpet Winsock works fine with either the Novell or
266 Arcether client, assuming you remember to load winpkt of course.
268 LAN Manager and Windows for Workgroups: These programs use protocols that
269 are incompatible with the Internet standard. They try to pretend
270 the cards are Ethernet, and confuse everyone else on the network.
272 However, v2.00 and higher of the Linux ARCnet driver supports this
273 protocol via the 'arc0e' device. See the section on "Multiprotocol
274 Support" for more information.
276 Using the freeware Samba server and clients for Linux, you can now
277 interface quite nicely with TCP/IP-based WfWg or Lan Manager
278 networks.
280 Windows 95: Tools are included with Win95 that let you use either the LANMAN
281 style network drivers (NDIS) or Novell drivers (ODI) to handle your
282 ARCnet packets. If you use ODI, you'll need to use the 'arc0'
283 device with Linux. If you use NDIS, then try the 'arc0e' device.
284 See the "Multiprotocol Support" section below if you need arc0e,
285 you're completely insane, and/or you need to build some kind of
286 hybrid network that uses both encapsulation types.
288 OS/2: I've been told it works under Warp Connect with an ARCnet driver from
289 SMC. You need to use the 'arc0e' interface for this. If you get
290 the SMC driver to work with the TCP/IP stuff included in the
291 "normal" Warp Bonus Pack, let me know.
293 ftp.microsoft.com also has a freeware "Lan Manager for OS/2" client
294 which should use the same protocol as WfWg does. I had no luck
295 installing it under Warp, however. Please mail me with any results.
297 NetBSD/AmiTCP: These use an old version of the Internet standard ARCnet
298 protocol (RFC1051) which is compatible with the Linux driver v2.10
299 ALPHA and above using the arc0s device. (See "Multiprotocol ARCnet"
300 below.) ** Newer versions of NetBSD apparently support RFC1201.
303 Using Multiprotocol ARCnet
304 --------------------------
306 The ARCnet driver v2.10 ALPHA supports three protocols, each on its own
307 "virtual network device":
309 arc0 - RFC1201 protocol, the official Internet standard which just
310 happens to be 100% compatible with Novell's TRXNET driver.
311 Version 1.00 of the ARCnet driver supported _only_ this
312 protocol. arc0 is the fastest of the three protocols (for
313 whatever reason), and allows larger packets to be used
314 because it supports RFC1201 "packet splitting" operations.
315 Unless you have a specific need to use a different protocol,
316 I strongly suggest that you stick with this one.
318 arc0e - "Ethernet-Encapsulation" which sends packets over ARCnet
319 that are actually a lot like Ethernet packets, including the
320 6-byte hardware addresses. This protocol is compatible with
321 Microsoft's NDIS ARCnet driver, like the one in WfWg and
322 LANMAN. Because the MTU of 493 is actually smaller than the
323 one "required" by TCP/IP (576), there is a chance that some
324 network operations will not function properly. The Linux
325 TCP/IP layer can compensate in most cases, however, by
326 automatically fragmenting the TCP/IP packets to make them
327 fit. arc0e also works slightly more slowly than arc0, for
328 reasons yet to be determined. (Probably it's the smaller
329 MTU that does it.)
331 arc0s - The "[s]imple" RFC1051 protocol is the "previous" Internet
332 standard that is completely incompatible with the new
333 standard. Some software today, however, continues to
334 support the old standard (and only the old standard)
335 including NetBSD and AmiTCP. RFC1051 also does not support
336 RFC1201's packet splitting, and the MTU of 507 is still
337 smaller than the Internet "requirement," so it's quite
338 possible that you may run into problems. It's also slower
339 than RFC1201 by about 25%, for the same reason as arc0e.
341 The arc0s support was contributed by Tomasz Motylewski
342 and modified somewhat by me. Bugs are probably my fault.
344 You can choose not to compile arc0e and arc0s into the driver if you want -
345 this will save you a bit of memory and avoid confusion when eg. trying to
346 use the "NFS-root" stuff in recent Linux kernels.
348 The arc0e and arc0s devices are created automatically when you first
349 ifconfig the arc0 device. To actually use them, though, you need to also
350 ifconfig the other virtual devices you need. There are a number of ways you
351 can set up your network then:
354 1. Single Protocol.
356 This is the simplest way to configure your network: use just one of the
357 two available protocols. As mentioned above, it's a good idea to use
358 only arc0 unless you have a good reason (like some other software, ie.
359 WfWg, that only works with arc0e).
361 If you need only arc0, then the following commands should get you going:
362 ifconfig arc0 MY.IP.ADD.RESS
363 route add MY.IP.ADD.RESS arc0
364 route add -net SUB.NET.ADD.RESS arc0
365 [add other local routes here]
367 If you need arc0e (and only arc0e), it's a little different:
368 ifconfig arc0 MY.IP.ADD.RESS
369 ifconfig arc0e MY.IP.ADD.RESS
370 route add MY.IP.ADD.RESS arc0e
371 route add -net SUB.NET.ADD.RESS arc0e
373 arc0s works much the same way as arc0e.
376 2. More than one protocol on the same wire.
378 Now things start getting confusing. To even try it, you may need to be
379 partly crazy. Here's what *I* did. :) Note that I don't include arc0s in
380 my home network; I don't have any NetBSD or AmiTCP computers, so I only
381 use arc0s during limited testing.
383 I have three computers on my home network; two Linux boxes (which prefer
384 RFC1201 protocol, for reasons listed above), and one XT that can't run
385 Linux but runs the free Microsoft LANMAN Client instead.
387 Worse, one of the Linux computers (freedom) also has a modem and acts as
388 a router to my Internet provider. The other Linux box (insight) also has
389 its own IP address and needs to use freedom as its default gateway. The
390 XT (patience), however, does not have its own Internet IP address and so
391 I assigned it one on a "private subnet" (as defined by RFC1597).
393 To start with, take a simple network with just insight and freedom.
394 Insight needs to:
395 - talk to freedom via RFC1201 (arc0) protocol, because I like it
396 more and it's faster.
397 - use freedom as its Internet gateway.
399 That's pretty easy to do. Set up insight like this:
400 ifconfig arc0 insight
401 route add insight arc0
402 route add freedom arc0 /* I would use the subnet here (like I said
403 to to in "single protocol" above),
404 but the rest of the subnet
405 unfortunately lies across the PPP
406 link on freedom, which confuses
407 things. */
408 route add default gw freedom
410 And freedom gets configured like so:
411 ifconfig arc0 freedom
412 route add freedom arc0
413 route add insight arc0
414 /* and default gateway is configured by pppd */
416 Great, now insight talks to freedom directly on arc0, and sends packets
417 to the Internet through freedom. If you didn't know how to do the above,
418 you should probably stop reading this section now because it only gets
419 worse.
421 Now, how do I add patience into the network? It will be using LANMAN
422 Client, which means I need the arc0e device. It needs to be able to talk
423 to both insight and freedom, and also use freedom as a gateway to the
424 Internet. (Recall that patience has a "private IP address" which won't
425 work on the Internet; that's okay, I configured Linux IP masquerading on
426 freedom for this subnet).
428 So patience (necessarily; I don't have another IP number from my
429 provider) has an IP address on a different subnet than freedom and
430 insight, but needs to use freedom as an Internet gateway. Worse, most
431 DOS networking programs, including LANMAN, have braindead networking
432 schemes that rely completely on the netmask and a 'default gateway' to
433 determine how to route packets. This means that to get to freedom or
434 insight, patience WILL send through its default gateway, regardless of
435 the fact that both freedom and insight (courtesy of the arc0e device)
436 could understand a direct transmission.
438 I compensate by giving freedom an extra IP address - aliased 'gatekeeper'
439 - that is on my private subnet, the same subnet that patience is on. I
440 then define gatekeeper to be the default gateway for patience.
442 To configure freedom (in addition to the commands above):
443 ifconfig arc0e gatekeeper
444 route add gatekeeper arc0e
445 route add patience arc0e
447 This way, freedom will send all packets for patience through arc0e,
448 giving its IP address as gatekeeper (on the private subnet). When it
449 talks to insight or the Internet, it will use its "freedom" Internet IP
450 address.
452 You will notice that we haven't configured the arc0e device on insight.
453 This would work, but is not really necessary, and would require me to
454 assign insight another special IP number from my private subnet. Since
455 both insight and patience are using freedom as their default gateway, the
456 two can already talk to each other.
458 It's quite fortunate that I set things up like this the first time (cough
459 cough) because it's really handy when I boot insight into DOS. There, it
460 runs the Novell ODI protocol stack, which only works with RFC1201 ARCnet.
461 In this mode it would be impossible for insight to communicate directly
462 with patience, since the Novell stack is incompatible with Microsoft's
463 Ethernet-Encap. Without changing any settings on freedom or patience, I
464 simply set freedom as the default gateway for insight (now in DOS,
465 remember) and all the forwarding happens "automagically" between the two
466 hosts that would normally not be able to communicate at all.
468 For those who like diagrams, I have created two "virtual subnets" on the
469 same physical ARCnet wire. You can picture it like this:
473 (registered Internet subnet) (RFC1597 private subnet)
475 (IP Masquerade)
476 /---------------\ * /---------------\
477 | | * | |
478 | +-Freedom-*-Gatekeeper-+ |
479 | | | * | |
480 \-------+-------/ | * \-------+-------/
481 | | |
482 Insight | Patience
483 (Internet)
487 It works: what now?
488 -------------------
490 Send mail describing your setup, preferably including driver version, kernel
491 version, ARCnet card model, CPU type, number of systems on your network, and
492 list of software in use to me at the following address:
493 apenwarr@worldvisions.ca
495 I do send (sometimes automated) replies to all messages I receive. My email
496 can be weird (and also usually gets forwarded all over the place along the
497 way to me), so if you don't get a reply within a reasonable time, please
498 resend.
501 It doesn't work: what now?
502 --------------------------
504 Do the same as above, but also include the output of the ifconfig and route
505 commands, as well as any pertinent log entries (ie. anything that starts
506 with "arcnet:" and has shown up since the last reboot) in your mail.
508 If you want to try fixing it yourself (I strongly recommend that you mail me
509 about the problem first, since it might already have been solved) you may
510 want to try some of the debug levels available. For heavy testing on
511 D_DURING or more, it would be a REALLY good idea to kill your klogd daemon
512 first! D_DURING displays 4-5 lines for each packet sent or received. D_TX,
513 D_RX, and D_SKB actually DISPLAY each packet as it is sent or received,
514 which is obviously quite big.
516 Starting with v2.40 ALPHA, the autoprobe routines have changed
517 significantly. In particular, they won't tell you why the card was not
518 found unless you turn on the D_INIT_REASONS debugging flag.
520 Once the driver is running, you can run the arcdump shell script (available
521 from me or in the full ARCnet package, if you have it) as root to list the
522 contents of the arcnet buffers at any time. To make any sense at all out of
523 this, you should grab the pertinent RFCs. (some are listed near the top of
524 arcnet.c). arcdump assumes your card is at 0xD0000. If it isn't, edit the
525 script.
527 Buffers 0 and 1 are used for receiving, and Buffers 2 and 3 are for sending.
528 Ping-pong buffers are implemented both ways.
530 If your debug level includes D_DURING and you did NOT define SLOW_XMIT_COPY,
531 the buffers are cleared to a constant value of 0x42 every time the card is
532 reset (which should only happen when you do an ifconfig up, or when Linux
533 decides that the driver is broken). During a transmit, unused parts of the
534 buffer will be cleared to 0x42 as well. This is to make it easier to figure
535 out which bytes are being used by a packet.
537 You can change the debug level without recompiling the kernel by typing:
538 ifconfig arc0 down metric 1xxx
539 /etc/rc.d/rc.inet1
540 where "xxx" is the debug level you want. For example, "metric 1015" would put
541 you at debug level 15. Debug level 7 is currently the default.
543 Note that the debug level is (starting with v1.90 ALPHA) a binary
544 combination of different debug flags; so debug level 7 is really 1+2+4 or
545 D_NORMAL+D_EXTRA+D_INIT. To include D_DURING, you would add 16 to this,
546 resulting in debug level 23.
548 If you don't understand that, you probably don't want to know anyway.
549 E-mail me about your problem.
552 I want to send money: what now?
553 -------------------------------
555 Go take a nap or something. You'll feel better in the morning.