1 # -------+---------+---------+---------+---------+---------+---------+---------+
3 The FreeBSD/sparc64 port has now changed time_t from 32-bits to 64-bits.
4 This file explains the exact steps that users should follow to update their
5 sparc64 systems for this change. People running FreeBSD on other types of
6 hardware, such as CPU's from Intel or AMD, can ignore this file. For now,
7 this change is only happening for people running FreeBSD on Sparc hardware.
9 # -------+---------+---------+---------+---------+---------+---------+---------+
10 # Copyright (c) 2004 - Garance Alistair Drosehn <gad@FreeBSD.org>.
12 # All rights reserved.
14 # Redistribution, publication, translation and use, with or without
15 # modification, in full or in part, in any form or format of this
16 # document are permitted without further permission from the author.
18 # THIS DOCUMENT IS PROVIDED BY GARANCE DROSEHN ``AS IS'' AND ANY EXPRESS
19 # OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
20 # WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
21 # DISCLAIMED. IN NO EVENT SHALL GARANCE DROSEHN BE LIABLE FOR ANY DIRECT,
22 # INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
23 # (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
24 # SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 # HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
26 # STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
27 # IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
28 # POSSIBILITY OF SUCH DAMAGE.
30 # -------+---------+---------+---------+---------+---------+---------+---------+
32 # -------+---------+---------+---------+---------+---------+---------+---------+
34 If you are in too much of a hurry to read this file, then this is not the
35 time for you to upgrade to a 64-bit time_t. Period. Stick with a system
36 using 32-bit time_t until you have plenty of time to perform an upgrade.
38 This statement is true even if you have performed a thousand system upgrades
39 in the past, and you are certain that you know everything there is to know
40 about upgrades. This upgrade *will* take you more time than previous system
41 upgrades, simply because you must recompile at least some of your ports after
42 upgrading the base system.
44 Do not start this update unless you have the extra time.
46 As of March 10th, the official value for time_t on sparc64 has changed to
47 be 64-bits. If you really must build system with 32-bit time_t's, then it
48 would be best to stick with a snapshot of current from before March 10th.
49 With src snapshots after that point, there will soon be ports which assume
50 you have a 64-bit time_t based on the value of __FreeBSD_version.
52 If you are ready to upgrade, then *READ THIS ENTIRE DOCUMENT* at least
53 once before starting the upgrade.
55 # -------+---------+---------+---------+---------+---------+---------+---------+
57 This is a major change. This change will *not* be backwards-compatible.
58 Any programs which call system-routines for handling time-values will
59 have to be recompiled after this change is made.
61 Because this change is not backwards-compatible, it is important that
62 the following steps be used when upgrading the system. "Shortcuts" that
63 have worked for EVERY SINGLE UPGRADE YOU HAVE EVER DONE IN YOUR LIFE are
64 probably irrelevant. This change is more disruptive than most of the
65 changes which are normally done on freebsd.
67 These steps are designed to minimize the chance of you running into any
68 trouble. We can not guarantee that these steps will avoid all possible
69 problems, but if you ignore these steps you are very likely to run into
70 some very painful and time-consuming headaches when upgrading.
72 Step Pre-1: Update to a recent snapshot of -current, and first build
73 that as a system with 32-bit time_t. To do this, edit
74 the file /usr/src/sys/sparc64/include/_types.h
76 typedef __int64_t __time_t; /* time()... */
77 and change '__int64_t' to '__int32_t'
78 Step Pre-2: Install that system, using whatever steps you normally
79 use, and make sure that installation seems to work okay.
80 Step Pre-3: While still running that 32-bit time_t system, it would
81 probably be a good idea to cvsup your ports tree, and
82 then upgrade portupgrade (if you use it) and upgrade any
83 shells that you use. Eg:
84 portupgrade -Rr -f ruby portupgrade
85 portupgrade -Rr -f bash
86 That way you know you have the latest versions, and you
87 will also know you have the most-recent distfiles on
90 Step Pre-4: For sparc64 machines which need DHCP:
91 The 'dhclient' in the base system is known to be unreliable
92 on a system which is upgraded to 64-bit time_t's. It may
93 work for you, but it probably will not.
94 As of March 10th 2004, we have no fix for that.
95 However, the net/isc-dhcp3-client port does seem to work.
96 IF your machine needs DHCP, then you should probably install
97 that port and make sure you can get it working *before* you
98 make the change to use 64-bit time_t's.
100 Step Pre-4: For people using database-related ports:
101 Ports like PostgreSQL may change how they store data after
102 they are recompiled for 64-bTT. So, you may have to do a
103 full dump of your data while you still have a 32-bTT system,
104 and do a matching initdb/restore of your data after the port
105 has been recompiled on the upgraded 64-bTT system.
107 After you have built and installed that src-snapshot with 32-bit time_t's,
108 edit the file /usr/src/sys/sparc64/include/_types.h
110 typedef __int32_t __time_t; /* time()... */
111 and change '__int32_t' back to '__int64_t'
113 For best results, do NOT make any other changes. Do NOT cvsup the
114 source tree trying to pick up any other changes. At this point you
115 know that you have a source tree that does work for your system, so
116 stick with that source tree (except for making the above 1-line
119 At one point in my testing, I did do a 'cvsup' which just happened
120 to pull in one bad commit that broke 'make buildworld', and a second
121 bad commit that broke 'make installworld'. Believe me, you REALLY
122 REALLY do *not* want to risk problems like that!
124 I am not suggesting that you have to do two whole buildworld/
125 installworld cycles in a single day. You could easily wait a few
126 days, or even a week between them. What I am suggesting is that
127 you should not 'cvsup' your sources inbetween the two buildworlds.
129 And then follow these steps to build and install the 64-bit time_t system:
132 make cleanworld #- 2. or 'rm -Rf /usr/obj/usr/src/*'
133 make buildworld #- 3.
134 make buildkernel #- 4. Add KERNCONF if you usually do.
135 NEWSPARC_TIMETYPE=__int64_t #- 5. (Used by a safety-check done
136 export NEWSPARC_TIMETYPE #- 5a. by installkernel)
137 make installkernel #- 6. Add KERNCONF if you usually do.
140 # - - A section required for installs over NFS-mounts - - #
141 ifconfig -a #- NFS 8a. See note below.
142 shutdown now #- NFS 8b. NOT 'shutdown -r now'
143 cd /usr/src #- NFS 8c.
144 sh installworld_oldk #- NFS 8d. See note below.
145 # - - End of this section for NFS-mounts - - #
147 reboot #- 9. MUST go into single-user mode
149 For many upgrades, it is true that you can "cheat" at this point, and
150 get away without actually going into single-user mode straight from
151 the reboot. But for this upgrade, you REALLY MUST start up straight
152 into single user mode. So, reboot the machine, type a space (or
153 anything other than 'Enter') when the boot-loader is counting down.
156 boot -s #- 10. (command to boot-loader)
158 The system will ask you if you want to use /bin/sh or some other shell.
159 For this upgrade, just hit enter, even if you usually prefer like some
160 other shell instead of /bin/sh.
163 # - - A section required for installs over NFS-mounts - - #
164 PATH=/boot/kernel/bin:$PATH #- NFS 12.
165 # - - End of this section for NFS-mounts - - #
166 mount -a -t ufs #- 13.
168 # - - A section required for installs over NFS-mounts - - #
169 ifconfig hme0 inet .... #- NFS 15a. See note below.
170 mount_nfs host:srcdir /usr/src #- NFS 15b. See note below.
171 mount_nfs host:objdir /usr/obj #- NFS 15c.
172 # - - End of this section for NFS-mounts - - #
174 sh installworld_newk #- 17. Might want to add -S
176 rm -f /var/db/dhclient.leases #- 19. If this host uses DHCP
179 At this point, you should be up-and-running on a system that has 64-bit
180 values for time_t. You will have to rebuild anything which depends on
181 time_t. Later in this file is a suggested order for upgrading ports.
183 If you have a lot of ports which start up daemons or do other processing
184 at system-startup, then you might want to have this reboot also go into
185 single-user mode for upgrading all of the ports. In my case, I've always
186 done a standard reboot at this point and did not run into problems, but
187 then I only have 25 ports installed on my SPARC64 system.
189 Aside: It is slightly more reasonable to use the 'reboot' command, although
190 you may be more familar with using 'shutdown -r now'. The shutdown command
191 just turns around and executes '/sbin/reboot', and with this upgrade it is
192 best to avoid such redirection.
194 # -------+---------+--------- Notes on the above -------+---------+---------+
196 General notes on NFS issues:
198 For this upgrade to 64-bit time_t's, the change is so disruptive that I
199 couldn't get NFS-mounts to work if I booted a "32-bit time_t system"
200 (ie: 32-bit versions of /bin, /sbin, /lib, ...) on a 64-bit kernel. So,
201 I added the installworld_oldk script. This script does two things:
202 1) Creates a mini-/bin inside /boot/kernel.
203 2) Does a minimal installworld (while still on the old kernel),
204 thus making it possible for NFS-mounts to work when you reboot.
206 The first half is a step that would be perfectly safe to do, for any
207 upgrade (including non-NFS ones), at any time. It is a generally safe
208 and interesting idea, although it really should be implemented as an
209 official target in /usr/src/Makefile to be done right.
211 The second half would USUALLY be a bad idea to do, but I think it's the
212 only way I can get this specific upgrade to work for people that install
213 from NFS-mounted directories. It is bad because you are clobbering parts
214 of your system even though (in the usual case) you would not know that
215 the new kernel actually works on your system. It also does not do a
216 full-install, so you end up booting into a system which is part old-
217 world, and part new-world. It looks like we can get away with that for
218 this upgrade, but the tactic would be too risky for "standard upgrades".
220 These instructions assume that you are already familiar with how to do
221 installations over NFS-mounted partitions. If you are not, you might
222 want to read other references, such as 'man development'.
224 Notes on step NFS 8a: ifconfig -a
226 This shows to the configuration of all your ethernet interfaces. Write
227 down the IP address and netmask of your main interface. This is
228 particularly important if the machine obtains its address via DHCP.
229 You will not be running dhclient after the reboot in step 8, so just
230 re-use the IP address that the machine is using for the present reboot.
232 Notes on step NFS 8b: shutdown now
234 This will drop you into single-user mode, without rebooting. It
235 will ask if you want to use /bin/sh for your shell. You do.
237 Notes on step NFS 8d: sh installworld_oldk
239 Note that this script only installs *part* of the new world. You will
240 still have to reboot into single-user mode and do the full installworld.
241 The installworld_oldk script will ask you if you want to build a
242 mini-/bin. For this upgrade, you should say "yes".
244 Notes on step NFS 15a:
245 On my Ultra-10, I have the 'hme0' device as my ethernet card. The output
246 of 'ifconfig -a' (from step 'NFS 7a') included the lines:
248 hme0: flags=8843<UP,BROADCAST,RUNNING,SIMPLEX,MULTICAST> mtu 1500
249 inet 192.168.1.18 netmask 0xffffffe0 broadcast 192.168.1.31
251 So for this step, I typed in the command:
252 ifconfig hme0 inet 192.168.1.18 netmask 0xffffff00
254 Notes on step NFS 15b: mount_nfs
256 At this step, you may need to specify the host as an IP address instead
257 of a hostname, because the machine will only be able to resolve hostnames
258 that are in /etc/hosts.
260 In my case, I found it easier to create a source file ahead of time
261 which included the ifconfig and mount_nfs commands that I knew I would
262 need, and then I just sourced that file after rebooting into single user
263 mode. If you made such a source file and put it in your root partition,
264 perhaps under /boot, then that file could also include all of the steps
267 Also, it is best use the 'mount_nfs' command, instead of 'mount -t nfs'.
268 If you use the 'mount' command for NFS mounts, it will turn around and
269 directly execute /sbin/mount_nfs, and that is not desirable in this case.
271 Notes on step 17: sh installworld_newk
273 This script will do some setup work, and then ask you if want it to run
274 'make installworld'. Most people should just answer "y" (yes) to that
275 prompt. You can avoid the prompt by including "-y" or "-n" on the
276 command. If you say "n" (no), then it will tell you what commands
277 you must type to do the actual installworld.
279 The script also recognizes a "-S" parameter, which causes it to use
280 symlinks instead of making copies of programs used by the installation
281 process. This option will cause less filespace to be used up in /tmp,
282 but it might be slower in some cases (especially for installs using
283 an NFS-mounted directory for /usr/obj).
285 Both this script and the installworld_oldk script also recognize a "-M"
286 option. This option causes the script to use the absolute minimum PATH
287 setting that "should" be needed to complete an install. This option is
288 mainly just for debugging the scripts, though. If you request the
289 minimum PATH, and some important file was NOT properly copied, then the
290 installworld will immediately die at that point. This might be painful.
291 Without "-M", the same oversight would mean that you will run the wrong
292 *version* of the command, but that older version might actually work
293 perfectly fine. I did all my testing with "-M" to make sure I had
294 found all important programs, but there is probably no advantage for
295 using it for standard system upgrades. Also, if there are no important
296 files overlooked, then "-M" will not make any difference at all.
298 # -------+---------+---------+ Upgrading Ports +---------+---------+---------+
300 Similar to the recommendation for the upgrading the system, I suggest that
301 you do not 'cvsup' your local copy of the ports collection before trying to
302 rebuild everything for 64-bit time_t. For one thing, you will have a cvsup
303 compiled for 32-bTT (32-bit time_t's), and that will not work well on a
304 system which is using 64-bTT. You might find that you have to 'cvsup' for
305 some ports, but you will need to get a 64-bTT version of cvsup before you
308 One tactic to use for upgrading ports is to rebuild your already-installed
309 ports one-at-a-time. If you want to do that, and if you use portupgrade
310 to upgrade your ports, then I suggest the first thing you should do is:
312 portupgrade -Rr -f ruby portupgrade #- Ports 1.
313 Aside: if you get an error about the "ruby-rdoc" port,
314 then enter: pkg_deinstall ruby-rdoc
315 and repeat the original command.
316 portupgrade -Rr -f bash #- Ports 2.
317 If you have 'bash' installed, or include any other shells
318 which you have installed from the ports collection. If
319 your session is *using* one of these shells, then logout
320 and log back in after recompiling that shell.
321 portupgrade -Rr -f ezm3 cvsup-without-gui #- Ports 3 (maybe).
322 If you want to rebuild a 64-bit time_t version of cvsup.
323 Note: ezm3 (modula-3) needs a patch to work correctly after the
324 change to 64-BTT. That fix has not been commited to the port
325 yet [as of Mar 10th], but hopefully it will be commited soon.
327 There are pre-built packages available for ezm3 and cvsup-without-gui on
328 the new 64-bTT systems. This ezm3 package *does* include the necessarily
329 patch. These files are available on the standard ftp servers for FreeBSD.
330 If you have previous versions installed, then remove them with:
334 If you get warnings about "unable to completely remove" some
335 lib/m3 directories when deleting ezm3, then also enter:
336 rm -rf /usr/local/lib/m3
338 You can install the new packages with:
340 pkg_add ftp://ftp3.FreeBSD.org/pub/FreeBSD/ports/local-distfiles/gad/ez...
341 pkg_add ftp://ftp3.FreeBSD.org/pub/FreeBSD/ports/local-distfiles/gad/cv...
342 Replacing "ez..." with "ezm3-64btt-1.1_1.tbz" and "cv..."
343 with "cvsup-without-gui-64btt-16.1h.tbz". You can also use
344 some other standard ftp server, instead of ftp3.FreeBSD.org.
346 "Now look over all the other ports you have installed, and
347 re-compile everything that probably needs to be recompiled".
349 If you are going to do it piecemeal, the next ports to force-rebuild would
350 probably be languages like perl and python, if you have them installed.
351 After that, force-rebuild the ports like autoconf and automake, if you
352 have them installed. Or you might want to play it safe at this point,
353 and simply recompile *every* port that you have installed.
355 A different tactic to use for ports is to remove *all* ports before you
356 do the installkernel/installworld step (while you're still on a 32-bTT
357 system). Then, once you're up on the 64-bTT system, start making them
358 one-by-one. If you follow this tactic, you might want to save the output
359 of a 'pkg_info' command before you start removing ports.
361 # -------+---------+---------+---------+---------+---------+---------+---------+
363 If you run into problems when making this change, please report them to
364 the mailing list freebsd-sparc64@FreeBSD.org .
366 # -------+---------+---------+---------+---------+---------+---------+---------+
367 # Notice that the following command can be useful in some settings:
368 grep '#\- ' UPDATING.64BTT