Brian's Wiki: RocBSD / Todo

This page is read-only because I use it as a primitive bug tracker. Please report bugs to me via email rather than try to add them to this page.

Kernel

double check all the non-wscons stuff we had to remove from generic, maybe some can be ifdef'd out and still work
pccons
- do pc bell
- q_to_b in pccons to ndqb
i386/disksubr.c, does some funky bios CHS stuff that i don't think we need
can we remove all the bios crap passed by /boot to the kernel in machdep.c? and bios.* sysctl
remove vm_daddr vm_dsize, etc
- Also, remove i386 W^X hacks, as we don't do it in userland anymore
remove unused fields in disklabel/partition
- change magic
- remove or reset version field?
- do any macros in disklabel.h refer to fields that are gone?
vnd_secsize? isn't it always DEV_BSIZE
double check ccd.c patches
- get rid of ccd_geom totally (just secsize now)
sc_type in fd.c? used to populate disklabel geometry, used for anything else?
option PCI_CONF_MODE 1 in GENERIC on i386 (99.99% of machines are, and we don't get the right answer from the bios with the unified x86/amd64 bootloader
remove the dkcsum nonsense (i don't even totally understand what it does)
remove multiple emulations support from exec_elf, remove the whole compat subtree from git
- Remove multiple binary types alltogether? (a.out, what else?)
remove chr2blk sysctl and associated cruft all the way down (chrtoblk/blktochr functions chrtoblktbl in all the machdep code)
- Ideally unify major device numbers across archs...
VM_PSSTRINGS sysctl can go (always constant)
stack smash callback in kernel can go (int $5 now)
scroll lock stuff (ifdef soon) in pccons
pcconskbd? is it even used? remove it
remove ufs2 and ext2fs totally, no need for magic checking in sys/ufs
KERN_MSGBUF* can go, size is constant, always read through /dev/kmsg
pccons - pcinit() call doesn't need to be in sput
- BUT something needs to call pcinit very early
get rid of the "ooo... look at me, I'm storing the kernel log buf in some obscure place in memory so i can read it after a warm boot" nonsense
ditch sys/ioctl_compat.h
i386/machdep.c, can we decode the cpu model in userspace?
- Where do we actually need it?
swdevt nonsense - only 1 field in the first entry is ever used
parsedisk annd lots of nam2blk stuff can go
/usr/src/sys/arch/i386/i386/dkcsum.c can go
are {I,L}{UCLC,LCUC}, XCASE termios flags part of the standard? - if not remove them
decide what gettimeofday should do with its second arg (ignore, EINVAL?)
cleandir target in src/bsd
pass a random in an extra register to userland, use it to set __guard (in all arch's crts, and ld.so)
Can the kernel not "know" the working directory? we need the *at funcs for posix, so we could have a special filehandle (3? ehh) containing the cwd, inherited though exec, etc
remove _getlogin syscall and associated cruft
proc0 - can it exit? does it need to exist?
- Reboot when init exits? (using the exit status to decide to reboot or shutdown)?
init "wait for everything else to start" sleep stuff - go?
int cold - can it go? when does it get unset?
nfsboot needs to depend on nfsclient
hw_vendor stuff in bios.c (static now, but unneeded)
add line disciplins back in, but conditionalize on nmea | ppp | etc
- put ioctls back too, and have them return constants for non-disc case
- remove "fixes" in userland
PID_MAX only limited for backwards compat i think

Bootstrap

libsa __RocBSD__ disklabel changes, little shaky...
unify boot(8)? can pxeboot, cdboot, and boot all live in the same binary?
cdbr should load cdboot at the correct address (0x40000) right away, no need

for 0x7c00 then have cdboot move itself

merge amd64 and i386 bootloaders (start with amd64, it is more minimalist and cuts out legacy cray)
find out what the + and * next to hard disk name means in boot are they wrong? do they matter?
also hd/sd confusion (why is scsi special?)

Installer

get rid (or simplify) some of the serial console stuff in installer?
~~use mfs for /tmp by default~~ no, mfs is not what you think it is
need a story for setting timezone

Build Process

make sure ~~BSD_AUTH~~, YP, KERBEROSE, SKEY etc still have some chance of working when enabled
- Add CLFAGS+=-DXXX where in app/library makefiles, add SRCS/MAN+= xxx.c (even if xxx.c isn't in tree)
- ~~perlapi.pod (i think) looks for some t/lib files (specificed in the manifest) that we don't currently bundle~~
~~machine includes installed under /usr/include/${MACHINE}~~
- /usr/include/machine contains:
  foo.h: #ifdef __i386__ #include <i386/foo.h> #elif .. #endif
make kernel build with bsd.prog.mk like a normal program
- set CFLAGS/LDFLAGS as appropriate
- config generates a SRCS line
- config generates a defs.h which we -include defs.h in CFLAGS
- Maybe don't even install config (or libexec it?)
make cleandir again
~~beforedepend stuff for generated files is stupid... can we do better?~~ guess not
~~usr.bin/install is funky (name overlaps with install target)~~
~~makewhatis needs to be run initially (either by make install or rc.firsttime)~~
much of share/man should be moved to where it belongs
~~librpcsvc doesn't build after cleandir (INCS dependencies wrong)~~
add an uninstall target
Add a *single* ${ARCH} variable (maybe don't even predefine it in make(1)?; use $(arch -s)?), i386, amd64, powerpc. We'll have to do some magic for the macppc dir, but there shouldn't be any outside the kernel. Poison ${MACHINE} and ${MACHINE_ARCH}. Maybe it should be called MACHINE though? Either way, start with ARCH, so we catch all the places it is used.
Maybe remove arch(1) and machine(1), they're getting rather useless. Where are they used?
install fenv.h and tgmath.h
Look at cvs history for libm, some libm functions were stupidly moved to libc, move 'em back

Etc

tset nonsense in etc/{skel,root}/dot*
actually could get rid of a lot of those dot.* files
where does /etc/motd show now that login is gone?
- should .profile show it?
- if it still shows it should be autogenerated (welcome to ${OSNAME})
something needs to make lastlog and wtmp
/etc/rc shouldn't generate ssh keys (does it?)
should librpcsvc install .x files in /usr/include?
parse /etc/myns like sysctl -f parser (strsep, comments, etc)
Merge new stuff from /etc/rc (baddynamic, host.random, ssh keys, etc)

Lib

~~Hide underscore symbols in shared libraries~~ no
tcgetsid is not implemented
move setsockopt to libxnet (anything else?)
pthread
- Remove pthread_switch_add_np and associated cruft
- Try to remove _thread_init or make it called as a constructor
need wcswidth and getws man page
get rid of librpcsvc?
move sysv ipc to libipc
- implement in userspace?
maybe wide characters should always use utf8? should we even bother with latin1?

Bin

can we conditionalize atrun?
warning: templates not found /usr/share/git-core/templates
-D_LIBKVM in sysctl?
roll cpp(1) into cc(1)
remove extended registers, and source vtables in dc(1)
disklabel "auto layout" need tweeking (no x11 at least)
passwd policy is too strict
warning: templates not found /usr/share/git-core/templates on git-clone
check ifconfig man page for options that can be disabled (group by interface type)
sys_auth_passwd in ssh's auth-passwd.c empty password stuff... not sure how that interacts with ssh's empty password config options
remove hostname.if files? do it the freebsd way?
Way too many superblock backups in newfs(8) by default
Our usr.bin/csh/exp.c exp2 workaround is no longer needed, fixed upstream

Share

remove some share/misc stuff?
_PATH_BOOTDIR should go (and all of /usr/mdec, do something better)
is mandoc stuff here still used?
remove some bsd.*.mk files (xconf xorg, port, etc)

Headers

remove getrrsetbyname from netdb.h
make sure all _PATH_* defines are still valid
Lots of machine/profile.h is unused, maybe it should go away (kernel profiling is likely broken now)

PowerPC

powerpc doesn't define __powerpc__
do we need svr4.h in powerpc?
longcall stuff in powerpc? Can we assume binaries are always < 32mb, and always use shortcalls?
.sdata/.sbss in powerpc, ditch it?

Global

~~conditionalize INET6 across the world~~

Distribution

simplify /dev, only minimum, maybe bump down some counts for sd*, etc
- no wd* by default
- installer should make sure root devices exist
remove resolv.conf (now, simpler, myns) from installer, dhcp too probably
osrelease stuff in installer gets it wrong (I think it pulls the, now removed, sysctl)
reboot(8) in miniroot needs getopt, maybe just ditch reboot and use kill -1 init
/root doesn't exist
set lists suck... can we do something different?

Perl

Lots of pods aren't being built
- But do we want all of them (readme's of langs/archs we don't support?)
Verify against "real" perl build again
We can ditch DynaLoader, and replace XSLoader with a stupid stub (since all XS modules are statically linked)
We can ditch ExtUtils-* (even -Constants, since they're included in the distribution now)
bsd.perl.mk still could use cleanup
- I think we can (somehow) use implied rules for pods
- The "find which directory the file is really in" nonsense is stupid
  - Either do the NetBSD ${.ALLSRC}/.MADE hack, or add a "resolve though path" modifier to make(1)
- Really should .include Makefile.inc, to get the right bindir, although the default is correct

New Features

libgmon is wrong - mcount.c doesn't match up with gcc
gdb in tree
add s/key support throughout
whatis database

Verification

double check fetch/ftp changes

Documentation

in NOKVM cases, remove command line options from usage() and man pages
gcc-local manpage? can we steal a c99 manpage from opengroup?
ensure all changes are documented in man pages
stpcpy man pages are wrong (linked to strcpy)
dprintf man page

Later

http://fedoraproject.org/wiki/Security/Features
come up with a story for nfs umount, need to do a MOUNTPROC3_UMNT call
double check security of newgrp (from netbsd)
remove all non-c99(1) options from cc? Add a -Wc, option to pass gcc specific flags?
Initialization (stdio and arc4random to name two) in libc and friends should use gcc constructors
- Maybe use a weak reference to arc4_stir to reseed on fork(), so we don't getpid() on every call
getaddrinfo locking
see if everything builds *without* objdir? should...
simplify perl's XSLoader (no need for Dynloader)
Can we make rsync use the system libmd and libz? Might have to tweak them a bit
Implement man -k and man -f
- -f is stupidly easy, use exitsing how() function to find NAME, and symlink whatis to man (plus add manpage)
- -k is a little harder... but POSIX is very lenient