path: root/user/jkoenig.mdwn

[[!meta copyright="Copyright © 2010 Free Software Foundation, Inc."]]

[[!meta license="""[[!toggle id="license" text="GFDL 1.2+"]][[!toggleable
id="license" text="Permission is granted to copy, distribute and/or modify this
document under the terms of the GNU Free Documentation License, Version 1.2 or
any later version published by the Free Software Foundation; with no Invariant
Sections, no Front-Cover Texts, and no Back-Cover Texts.  A copy of the license
is included in the section entitled [[GNU Free Documentation
License|/fdl]]."]]"""]]

## Hurd Debian-Installer

My [proposal](http://wiki.debian.org/SummerOfCode2010/HurdDebianInstaller/JeremieKoenig)
to work on porting d-i on Hurd
as a [Google Summer of Code](http://code.google.com/soc/) student
has been accepted by the Debian project.

I will be keeping track of my progress on this page.

### Links

  * [Modified packages](http://jk.fr.eu.org/debian/unstable)
  * [Latest images](http://jk.fr.eu.org/debian/hurd-installer)
  * [Debian bugs](http://bugs.debian.org/cgi-bin/pkgreport.cgi?users=jk@jk.fr.eu.org&tag=gsoc2010)
  * [BusyBox port](http://lists.debian.org/debian-bsd/2010/05/msg00048.html)
  * [GNU Mach initrd](http://lists.gnu.org/archive/html/bug-hurd/2010-06/msg00047.html)

### Roadmap

* **mach**: initrd support
    * (./) preliminary patch posted and self-built (2010-06-12)
    * adjustments will be needed (postponed)
    * consider the alternatives discussed on bug-hurd (postponed)

* **glibc**: fix `mkdir("/")` which returned `EINVAL`
    * (./) eglibc 2.11.2-1 includes a quick fix by youpi (2010-06-15)
    * (./) more complete patch posted to bug-hurd,
      since other calls return incorrect errors under some circumstances
      (2010-06-16)
    * more work on it will be needed to make it fix the whole thing
      (postponed)

* (./) **partman** (2010-06-23)
    * (./) add hurd-i386 to
      `partman-partitioning/lib/disk-label.sh`
      (2010-06-16, commited by youpi on 2010-06-23)
    * (./) short-circuit
      `partman-basicfilesystems/init.d/kernelmodules_basicfilesystems`
      (2010-06-16)
    * (./) partman-auto recipes:
      make the default filesystem os-dependent
      when it has not been preseeded (ie. the *seen* flag is clear)
    * (./) force 4k blocks and 128 bytes inodes
    * (./) submit patches to bugs.debian.org
      ([[!debbug 586870]] and [[!debbug 586871]])
    * (./) rebuild with responsible version numbers and upload to my repository

* (./) **libparted** (2010-06-23)
    * (./) fix device paths ([[!debbug 586696]])
    * (./) fix crash on exit for part:* stores ([[!debbug 586682]])

* **hurd-udeb** (2010-06-23)
    * (./) rebuild with the hack suggested by youpi for qemu network configuration
    * (./) fix mount to accept `-o defaults`
    * cleanup, ask youpi to commit

* reloading the partition table (2010-06-25)
    * User-space part stores
        * (./) hurd-udeb now uses `part:N:device:X` for partition devices
          (2010-06-23)
        * (./) it also provides /lib/partman/commit.d/??hurd\_reloadpart,
          which basically does `settrans -ag /dev/[hs]d*`.
          (2010-06-24)
    * Kernel-based partition devices
        * (./) Mach's drivers from Linux support reloading partitions.
          With help from youpi this has been made available through a
          device\_set\_status() call.
        * (./) make libparted use it
        * Reminder:
          I should file a bug against libparted with the patch sometime.

* (./) The `/servers/exec` issue (2010-06-26)
    * Due to /servers being inexistant,
      the bootstrap ext2fs could not register the initial exec server,
      meaning that non-bootstrap filesystems used a different one
      (started from the passive translator),
      which for some reason died on shell scripts, making them stall.
    * Adding the `/servers` directory to hurd-udeb fixed it,
      as well as the /hurd/proc issue
      (failed to be run by init the first time around).
    * Reminder: report the non-bootstrap exec servers failure on scripts.

* (./) **base-installer**: (2010-06-26)
    * Work around non-existant /proc/mounts.
    * Firmlink /servers into /target after debootstrap
      to make the network available.

* (./) **grub-installer**
    * (./) add hurd support (2010-06-27)
    * /!\ grub-legacy still needs to be tested
    * submit changes as a Debian bug

**Milestone (2010-06-28):
installer kindof works, with documented manual intervention required**

* (./) Sort out the situation with dev node creation (2010-07-07):
    * Devices and servers used to be set up by debootstrap;
      the hurd package would add some missing nodes.
    * New strategy implemented in hurd and debootstrap:
        * debootstrap uses active firmlinks into the host system
          for the target system's /dev and /servers.
        * the hurd package now include a `setup-translators` script,
          which is used to register the passive translators by the installer's
          `/libexec/runsystem` and hurd's postinst script.

* **busybox**: submit upstream and to [[!debbug 323670]]
  (waiting for upstream to review)
    * (./) I have mentioned my work on the upstream mailing list,
    * (./) merge the recent changes from upstream,
      notably to the build system.
      (2010-06-23)
    * (./) ask upstream for review and merge
      (2010-06-25)
    * (./) sent as patches as requested
      (2010-07-08)
    * backport any additional changes onto the debian branch
    * hijack [[!debbug 323670]] and submit my patches

* **aptitude**:
    * Currently broken on hurd-i386:
      [[!debpkg gtest]] fails to build because of a segfault in one of the test
      cases, [[!debpkg google-mock]] and hence [[!debpkg aptitude]] are missing
      it as a build-dep.
      The older package is not installable anymore because it's linked against
      an older version of libept, which has been removed.
    * (./) I bypassed the tests and uploaded the 3 packages to my repository
      (2010-07-08)
    * The segfault will have to be sorted out. (postponed)

* (./) "Fix" the swap situation. (2010-07-08)
    * The device_close() libstore patch
      had the unfortunate effect of making swapon fail,
      since the device it activates has to be kept open.
    * add options for MAKEDEV and setup-devices
      to use the libparted stores
    * disable youpi's patch
    * make partman-basicfilesystems re-create the device
      as a kernel partition, which is needed for swapon

* (./) netcfg-static: port to hurd (2010-07-09)
    * There was some amount of hurd support already
      (namely, activating the interface by replacing the socket translator)
    * However, this code started an active translator with
      di_exec_shell_log("settrans -a ...),
      which stalled as a consequence of it capturing libdi's pipe
      as its standard output.
    * Network devices must be probed by trying to open Mach devices
      with predetermined names (currently eth%d, wl%d),
      because getifaddrs() does not seem to work on Hurd.
    * /!\ netcfg, and configuring the installed system, postponed.

* **procps** 3.2.7-11 (current hurd-i386 version) has [[!debbug 546888]]
    * (./) Submit [[!debbug 588677]] and upload the result to my repository.
      (2010-07-11)

* (./) Set up a Debian mirror with modified packages for installation
    * the [mirror](http://jk.fr.eu.org/debian/hurd-install/mirror)
      is now up and running (2010-07-06)
    * hacked the image build script to include its public key in
      debian-archive-keyring at image build time (2010-07-08)
    * Apparently debootstrap does not handle multiple versions very well.
      Fix by using dpkg-scan{package,sources} rather than apt-ftparchive
      to create index files.
      (2010-07-10)
    * Use the override files from ftp.debian.org,
      to avoid debootstrap grabbing inappropriate packages.
    * Changed them to make [[!debpkg ifupdown]],
      [[!debpkg dhcp3-client]] and [[!debpkg dhcp3-client]] priority extra,
      because they're uninstallable at the moment.
      (2010-07-12)

* (./) Put together a "jk-archive-keyring" package,
  so that the mirror is authenticated in the target system as well.
  (2010-07-12)

* Fix grub for user-space partitions (expected 2010-07-12)
    * Currently, when using user-space partitions,
      grub-probe detects the whole device rather than the partition
      as the device behind /boot/grub.
      Consequently, grub-install fails.
    * One approach would be to replace /dev/hd* by kernel devices
      for file systems as well as for swap partitions.
    * On the other hand, grub-probe will need to be fixed
      sooner or later for user-space partition support,
      and parsing the fsysopts would possibly be an option.

* **d-i/installer/build**: (expected 2010-07-12)
    * publish the patch I use
    * sort out the changes suitable for inclusion
      and ask youpi and/or debian-boot@l.d.o to commit them

**Milestone (expected 2010-07-12):
installer works but it's still somewhat ugly and broken**

* call for testing and fix the bugs

* See if it would be possible to avoid accessing devices
  when firmlinking them into the target system.
    * Who does that?
        * find should prune them as instructed,
          and avoid a stat/readdir as a result
        * showtrans shouldn't access the translator itself
        * firmlink should access the target only when it is used

* In the same vein, permissions on underlying nodes
  are probably broken when created by hurd through `MAKEDEV -k`,
  and doing the actual chmod/chown accesses the device
    * Maybe settrans could be made to accept -o/--owner and
      -p/--perm, to set the permissions for the underlying node?

* Silence the "no kernel" warning somehow.

* Prepare a NMU for genext2fs (which is orphaned),
  and ask youpi to sponsor the upload.

* **busybox**: port 
    * fix stty/stat/ipcs on kFreeBSD,
    * generally port more stuff,
    * *ip* is needed (maybe) for network configuration,
    * *mount*, *swaponoff* can be from hurd-udeb for now,
      though the kFreeBSD people will need them

* **partman**: further adjustments
    * hide irrelevant mount options? (sync, relatime)

* Network configuration on the installed system.
  This includes porting ifupdown and isc-dhcp-client,
  which are currently uninstallable on hurd-i386.

* improve the [initrd situation](FIXME: link to bug-hurd post):
  ajust the ramdisk support in Mach,
  use tmpfs if possible,
  possibly add `module -nounzip` to grub2.
* mklibs{,-copy}:
  test library reduction,
  make it copy the ld.so -> ld.so.1 symlink.
* hurd console fonts
* better DHCP support during and after installation

**Milestone (expected 2010-07-19):
it works great and it's beautiful**

* test, fix, document
* support more types of installation images
* give a shot at the graphical installer if time permits
* integrate wireless drivers with netcfg
* see how [[zhengda]]'s work on DDE could be integrated
* etc..

### Mostly done

#### Week 1 (2010-05-24)

* genext2fs: patches submitted, [[!debbug 562999]]
  which add support for all block sizes and choosing them at runtime.
* busybox: started porting the upstream and Debian package to Hurd and FreeBSD
* rebuilding hurd-udeb from the pkg-hurd version
  and adding a ld.so link to the initrd
  fixes the exec translator crashing on startup.
  (BTW would there be a mean to detect this from the libdiskfs bootstrap code
  and report it ?)

#### Week 2 (2010-05-31 to 2010-06-06)

* *busybox*: patches [posted](http://lists.debian.org/debian-bsd/2010/05/msg00048.html).
* *libdebian-installer4*: [[!debbug 584538]]
* started working on mach initrd support
* the installation images could boot into the main-menu
  with the following changes:
      * rebuild hurd-udeb from with the latest pkg-hurd patches
      * use busybox from my osports-debian branch (see link above)
      * tweak the d-i image build scripts
      * the symlink /lib/ld.so -> ld.so.1 needs to be created somehow
        (youpi mentionned it being the job of libc0.3-udeb I think)
      * fix the poll() issue in libdebian-installer
        (patch to be submitted soon),
        also there is some hurd doxygen short-circuiting stuff
        there which does not apply any more and prevents is to build.
      * feed the initrd as a hard drive in qemu
        (with some more space added to avoid it from becoming full)