TODO

TODO list for libguestfs
======================================================================

This list contains random ideas and musings on features we could add
to libguestfs in future.

   - RWMJ

FUSE API
--------

The API needs more test coverage, particularly lesser-used system
calls.

The big unresolved issue is UID/GID mapping between guest filesystem
IDs and the host.  It's not easy to automate this because you need
extra details about the guest itself in order to get to its
UID->username map (eg. /etc/passwd from the guest).

Haskell bindings
----------------

Complete the Haskell bindings (see discussion on haskell-cafe).

PHP bindings
------------

Add bindtests to PHP bindings.

Complete bind tests
-------------------

Complete the bind tests - must test the return values and error cases.

virt-inspector - make libvirt XML
---------------------------------

It should be possible to generate libvirt XML from virt-inspector
data, at least partially.  This would be just another output type so:

 virt-inspector --libvirt guest.img

Note that recent versions of libvirt/virt-install allow guests to be
imported, so this is not so useful any more.

Ideas for extra commands
------------------------

  General glibc / core programs:
    chgrp

  ext2 properties:
    badblocks
    debugfs
    dumpe2fs
    e2image
    e2undo
    filefrag

  SELinux:
    chcat
    restorecon
      [Wanlong Gao submitted patches for restorecon, but
       there are problems with using the restorecon binary
       from the host on the guest.  Most of the time it
       would do more harm than good.]
    setfiles

  Oddball:
    pivot_root
    fts(3) / ftw(3)

  sh-in, sh-out: run shell command with large input/output
  debug sh-in, debug sh-out: debug versions of the above

Other initrd-* commands
-----------------------

Such as:

initrd-extract
initrd-replace

virt-rescue pty
---------------

See:
http://search.cpan.org/~rgiersig/IO-Tty-1.08/Pty.pm
http://www.perlmonks.org/index.pl?node_id=582185

Note that pty requires cooperation inside the C code too (there are
two sides to a pty, and one has to be handled after the fork).

[I tried to implement this in the new C virt-rescue, but it doesn't
work.  qemu is implementing its own ptys, and they are broken.  Need
to fix qemu.]

Windows-based daemon/appliance
------------------------------

See discussion on list:
https://www.redhat.com/archives/libguestfs/2009-November/msg00165.html

virt-disk-explore
-----------------

For multi-level disk images such as live CDs:
http://rwmj.wordpress.com/2009/07/15/unpack-the-russian-doll-of-a-f11-live-cd/

It's possible with libguestfs to recursively look for anything that
might be a filesystem, mount-{,loop} it and look in those, revealing
anything in a disk image.

However this won't work easily for VM disk images in the disk image.
One would have to download those to the host and launch another
libguestfs instance.

[Not sure this is such a good idea.  See also live CD inspection idea below.]

Map filesystems to disk blocks
------------------------------

Map files/filesystems/(any other object) to the actual disk
blocks they occupy.

And vice versa.

Is it even possible?

See also contribs/visualize-alignment/

Integration with host intrusion systems
---------------------------------------

Perfect way to monitor VMs from outside the VM.  Look for file
hashes, log events, login/logout etc.

http://www.ossec.net/
http://la-samhna.de/samhain/
http://sourceforge.net/projects/aide/
http://osiris.shmoo.com/
http://sourceforge.net/projects/tripwire/

Freeze/thaw filesystems
-----------------------

Access to these ioctls:
http://git.kernel.org/linus/fcccf502540e3d7

Tips for new users in guestfish
-------------------------------

$ guestfish
Tip: You need to 'add disk.img' or 'alloc disk.img nn' to make a new image.
Type 'notips' to disable tips permanently.
><fs> add mydisk
Tip: You need to type 'run' before you can see into the disk image.
><fs> run
Tip: Use 'list-filesystems' to see what filesystems are available.
><fs> list-filesystems
/dev/vda1
Tip: Use 'mount fs /' to mount a filesystem.
><fs> mount /dev/vda1 /
Tip: Use 'll /' to view the filesystem or ...
><fs> ll /

Could we make guestfish interactive if commands are used without params?
------------------------------------------------------------------------

><fs> sparse
[[Prints man page]]
Image name? disk.img
Size of image? 10M

Better support for encrypted devices
------------------------------------

Currently LUKS support only works if the device contains volume
groups.  If it contains, eg., partitions, you cannot access them.
We would like to add:

  - Direct access to the /dev/mapper device (eg. if it contains
    anything apart from VGs).

Display image as PS
-------------------

Display the structure of an image file as a PS.

Greater use of blkid / libblkid
-------------------------------

There are various useful functions in libblkid for listing partitions,
devices etc which we are essentially duplicating in the daemon.  It
would make more sense to just use libblkid for this.

There are some places where we call out to the 'blkid' program.  This
might be replaced by direct use of the library (if this is easier).

But it is very hard to be compatible between RHEL6 and RHEL5 when
using direct library.

Visualization
-------------

Eric Sandeen pointed out the blktrace tool which is a better way of
capturing traces than using patched qemu (see
contrib/visualize-alignment).  We would still use the same
visualization tools in conjunction with blktrace traces.

guestfish parsing
-----------------

At the moment guestfish uses an ad hoc parser which has many
shortcomings.  We should change to using a lex/yacc-based scanner and
parser (there are better parsers out there, but yacc is sufficient and
very widely available).

The scanner must deal with the case of parsing a whole command string,
eg. for a command that the user types in:

 ><fs> add-drive-opts "/tmp/foo" readonly:true

and also with parsing single words from the command line:

 guestfish add-drive-opts /tmp/foo readonly:true

Note the quotes are for scanning and don't indicate types.

We should also allow variables and expressions as part of this new
parsing code, eg:

 set roots inspect-os
 set product inspect-get-product-name %{roots[0]}

% is better than $ because of shell escaping and confusion with shell
variables.

Can we combine this with ability to set and read environment
variables?  Currently guestfish uses many environment variables like
$EDITOR without any corresponding ability to set them.

 set EDITOR /usr/bin/emacs
 echo $EDITOR  # or %{EDITOR}
 edit /etc/resolv.conf

More ntfs tools
---------------

ntfsprogs actually has a lot more useful tools than we currently
use.  Interesting ones are:

ntfscluster: display file(s) that occupy a cluster or sector

ntfsinfo: print various information about NTFS volume and files

ntfs streams: extract alternate streams from NTFS files

ntfsck: checker for NTFS filesystems

Undelete files
--------------

Two useful tools:

  - ext2undelete
  - ntfsundelete

More mkfs_opts options
----------------------

Useful options to offer:
 - Set label.
 - Set UUID.

Use /proc/self/mountinfo
------------------------

This file contains lots of interesting information about
what is mounted and where. eg:

  16 21 0:3 / /proc rw,relatime - proc /proc rw
  17 21 0:16 / /sys rw,relatime - sysfs /sys rw,seclabel
  18 23 0:5 / /dev rw,relatime - devtmpfs udev rw,seclabel,size=1906740k,nr_inodes=476685,mode=755
  26 21 253:3 / /home rw,relatime - ext4 /dev/mapper/vg-lv_home rw,seclabel,barrier=1,data=ordered

This could be used instead of current hairy code to parse the output
of the 'mount' command.  We could add new APIs to return kernel mount
options, type of filesystem at a mountpoint etc.

guestfish drive letters
-----------------------

There should be an option to mount all Windows drives as separate
paths, like C: => /c/, D: => /d/ etc.

More inspection features
------------------------

- last shutdown time
- DHCP address
- last time the software was updated
- last user who logged in
- lastlog, last, who

Integrate virt-inspector with CMDBs
-----------------------------------

Either integrate virt-inspector with Configuration Management
Databases (CMDBs) or at least check that virt-inspector produces the
right range of data so that integration would be possible.  The
standards for CMDBs come from the DMTF, see eg:

http://dmtf.org/news/pr/2009/7/dmtf-releases-cmdbf-standard-federating-configuration-management-data

Efficient way to visit all files
--------------------------------

https://rwmj.wordpress.com/2010/12/15/tip-audit-virtual-machine-for-setuid-files/#content

A naive method would look like:

  g#visit ~return_stats:true "/" (
    fun pathname stat ->
      ...
  )

However this has two disadvantages:

 - requires hand-written custom bindings in each language
 - unclear about locking, thread-safety and re-entrancy of handle g

A better way would be to have some sort of explicit "download all
filenames and stat structures", which could then be iterated over:

  let files = g#find_opts ~return_stats:true "/" in
  List.iter (
    fun pathname stat ->
      ...
  )

The problem with this is that 'files' is going to be larger than a
protocol buffer.

This leads to thinking about changes to the protocol / generator to
make this simpler.  The proposal would be to add RBigStringList,
RBigStructList [or RBig (Ranytype ...)].  These would work like
FileOut, in that they would use file streaming to stream XDR
structures (probably written to a file on the library side).
Generated code would hide most of the implementation.

We also need to think about security issues: is it possible for the
daemon to keep sending back data forever, and if so what happens on
the library side.

[Users can now use virt-ls to solve some of these problems, but it is
not a general solution at the API level]

Interactive disk creator
------------------------

An interactive disk creator program.

Attach method for disconnected operation
----------------------------------------

http://libguestfs.org/guestfs.3.html#guestfs_set_attach_method

"Librarian" has an idea that he should be able to attach to a regular
appliance, but disconnect from it and reconnect to it later.  This
would be some sort of modified attach method (see link above).

The complexity here is that we would no longer have access to
stdin/stdout (or we'd have to direct that somewhere else).

libosinfo mappings for virt-inspector
-------------------------------------

Return libosinfo mappings from inspection API.

virt-sysprep ideas
------------------

 - other Spacewalk / RHN IDs (?)
 - Windows sysprep
   (see: https://github.com/clalancette/oz/blob/e74ce83283d468fd987583d6837b441608e5f8f0/oz/Windows.py )
 - (librarian suggests ...)
   . run external guestfish script   virt-sysprep --fish=/tmp/foo.fish
 - if drives are encrypted, then dm-crypt key should be changed
   and drives all re-encrypted
 - /etc/pki
   (Steve says ...)
   Rpm uses nss. Nss sets up its crypto database in
   /etc/pki. Depending on how long the machine ran before cloning, you
   may have picked up some certificates or things. This is an area
   that you would want to look into.
 - secure erase of inodes etc using scrub (Steve Grubb)
 - other directories that could require cleaning include:
    /var/run/*
   (thanks Marko Myllynen, James Antill)
 - remove or modify UUIDs in /etc/fstab (eg. on Ubuntu)
   (thanks Joshua Daniel Franklin)

Kazuo Moriwaka adds:

 - swap devices (both of block device and file) should be wiped. This may
   good for security purpose, and size.  I found virt-sparsify can clear
   swap partition.
 - --script is nice.  Defining default sysprep script directory
   like /usr/lib/guestfs/sysprep-scripts.d/ may be useful to integrate
   other package maintainer(or ISV)'s effort.
 - To achieve better (or crazy) coverage, a simple examination will be
   help:
     Install the same kickstart into VM twice, and diff the trees.
     Many autogenerated IDs and configs can be found :)

As well as 'virt-sysprep' there is a case for a 'virt-customize' tool
which can customize templated guests.  This would be useful within
companies/organizations that want to offer multiple guests, but all
customized with the organization logo etc.  Some ideas:

 - change the background image to some custom desktop
 - change the sign-on messages (/etc/issue.net etc)
 - Windows login script/service

Launch remote sessions over ssh
-------------------------------

We had an idea you could add a launch method that uses ssh, ie.  all
febootstrap and qemu commands happen the same as now, but prefixed by
ssh so it happens on a remote machine.

Note that proper remote support and integration with libvirt is
different from this, and people are working on that.  ssh would just
be "remote-lite".

virt-make-fs and virt-win-reg need to not be in Perl
----------------------------------------------------

Probably they should be in C or OCaml.

Integrate snap-type functionality in inspection tools
-----------------------------------------------------

Mo Morsi's "snap" program lets you describe a guest as the list of
packages (eg. RPMs) installed + changes made to those RPMs + files
added.

http://projects.morsi.org/wiki/Snap

This results in a compact description of the guest.  He even managed
to do a kind of migration of guests by simply recreating the guest
from the description on the target machine.

It would be ideal to integrate this and/or use inspection to do this.

Ongoing code cleanups
---------------------

Examine every use of 'int' in C code for signed overflow problems.

All file descriptors in the library and daemon should normally be
opened with O_CLOEXEC.  Therefore we need to examine every call to:

 - open, openat
 - creat
 - pipe (see also: pipe2)
 - dup, dup2 (see also: dup3)
 - socket, socketpair
 - accept (see also: accept4)
 - signalfd, timerfd, epoll_create

virt-sparsify enhancements
--------------------------

TMPDIR should be checked to ensure that we won't run out of space
during the conversion, since current behaviour is very bad when this
happens (it usually causes virt-sparsify to hang).  This requires
writing a small C binding to statvfs for OCaml.

'virt-sparsify --whitelist' option to generate skeletons (for
debugging, bug forensics, diagnosis).  The whilelist option would
specify a list of files to be *preserved*.  All other files in the
image would be replaced by equivalent files of zeroes, thus minimizing
the size of the debug image that needs to be shipped to us by the
customer.

Optimize the appliance
----------------------

Pass -cpu host.  Anything else?

[The libvirt attach-method uses 'host-model' which is basically
the same as this]

Sort out partitioning
---------------------

Ignoring some legacy APIs, we currently have a mixed selection of
'part-*' APIs, implemented using parted.  We don't like parted or
libparted very much, and would love to replace it with something else.
The part-* APIs are quirky, but not too bad and we should maintain and
extend them instead of making another set of APIs.

One option is to write "libmbr" and "libgpt" libraries that would just
do MBR and GPT respectively, and do it directly and do it well.  They
wouldn't try to abstract anything (so, unlike libparted).  We could
then reimplement the part-* APIs on top of these hopefully sensible
libraries.  This is a lot of work.

Another option is to look for tools or libraries to replace parted.
For GPT there is a fairly obvious candidate: Rod Smith's GPT fdisk
(http://www.rodsbooks.com/gdisk/).  Rod has spent a lot of time
studying GPT, and seems to know more about it than any sane man
should.  There is a command line tool designed for scripts called
'sgdisk'.  The tools are packaged for many Linux distros.  Even if
this approach works, it doesn't solve the MBR problem, so likely we'd
have to write a library for that (or perhaps go back to sfdisk but
using a very abstracted interface over sfdisk).

qemu caching
------------

(Suggested by Paolo Bonzini and Kevin Wolf)

Measure the effect of cache=none, cache=directsync,
cache=writethrough, cache=writeback.

It's doubtful that using cache=none is useful, since it disables the
host cache making read-heavy workloads slower (they rely entirely on
the smaller appliance kernel's cache).  And in libguestfs we don't
necessarily care about ongoing data integrity while writing, as long
as data is reliably written out when g.sync, g.shutdown or g.close
return.  Also in libguestfs we effectively control the whole stack, so
we can ensure write barriers happen when we want.

libvirt attach-method
---------------------

Since libguestfs 1.19.24 this mostly works.  Here are some suggested
items to work on:

 - SELinux labelling of guestfsd.sock, console.sock
   https://bugzilla.redhat.com/show_bug.cgi?id=842307
   Once this is fixed, remove <seclabel type=none> from libvirt XML

 - Check feature parity between src/launch-appliance.c and
   src/launch-libvirt.c.

 - Remote support.  (This requires work on libvirt)

virt-sparsify should use discard
--------------------------------

This requires some changes to qemu to make discard work properly
throughout the entire stack.

Reimplement some APIs to avoid protocol limits
----------------------------------------------

We should reimplement the following APIs to avoid protocol limits.
These would be changed from daemon_functions to non_daemon_functions,
with the non-daemon versions implemented using guestfs_upload and
guestfs_download (and others).  This change should be transparent from
the p.o.v of the API and ABI.

 - guestfs_readdir

hivex
-----

Add more of hivex to the API, especially for writing.

Reimplement virt-win-reg to use this API.  (This is difficult because
the Perl libraries underneath access the hivex API directly).

ruby
----

Implement blocking calls.  The API for this:

  http://www.spacevatican.org/2012/7/5/whos-afraid-of-the-big-bad-lock/

is very poorly designed and essentially impossible for us to use:

  https://bugs.ruby-lang.org/issues/5543

particularly if we also want to maintain backwards compatibility with
Ruby 1.8, and/or maintain volatile VALUEs on the stack.

ACLs and capabilities
---------------------

We need to model both filesystem ACLs and filesystem capabilities
through the API.  This is particularly important in order to be able
to implement SCAP.

ACLs can be read and written using the acl(5) library and the
functions like acl_set_file(3) etc.

Setting the ACL on a file sets the extended attribute
'system.posix_acl_access' to a binary blob.  The kernel has a whole
bunch of complex code that seems to interpret these
(linux/fs/posix_acl.c).

Filesystem capabilities can be read and written using the libcap(3)
library and functions like cap_get_file, cap_set_file.

Setting fs capabilities on a file sets the extended attribute
'security.capability' to a binary blob.  These are implemented using a
Linux Security Module (security/capability.c) and presumably by
something in exec, but I couldn't see exactly how this works.