The kernel is made up of many subsystems which qualify as major programs on
their own. These subsystems tend to have technical issues which are best
discussed in smaller groups of concerned developers; thus, there tend to be
several focused "minisummits" over the course of the year. The Kernel
Summit traditionally provides a slot for reports from these minisummits,
and 2010 was no exception.
Dave Miller reported on the netfilter workshop, recently held in
Seville. Many topics were covered there, including IP virtual server
(IPVS) support. Work is progressing in this area, with better support for
namespaces, DNAT support, and a new persistence engine for the SIP
protocol. Also in the works is connection synchronization support,
enabling the operation of hot backup servers.
Ingress bandwidth shaping is another area of development; it's not
something we support well now. The existing options require out-of-tree
patches. Also out of tree is the "xtables add-ons and blobs" repository
maintained by Jan Engelhardt, which contains stuff which is "obscure but
useful to somebody." Even more out of tree is the "OpenOffice.org
firewall," a means for importing iptables statistics into OpenOffice.org
which was implemented as a joke in response to silly statements by a French
There was a mention of ipset, which optimizes netfilter rules which match
sets of addresses. Ipset lacks IPv6 support and needs a user interface
rewrite; much of this work has now been completed. Ipset also uses the new
libmnl netlink library, the third such library out there. Libmnl was
created in response to deficiencies seen in libnetlink (which is too
tightly tied to iproute2) and libnl (which is too abstracted and hides
Nftables, a netfilter
replacement based on an in-kernel virtual machine, was covered here in
2009; since then, this project has gone quiet. It came up at the workshop,
though; there is interest in it because it would allow the eventual
elimination of much of the existing code and could also be used in other
parts of the stack where packet classification is done (the packet
scheduler, for example). It could become, in other words, a general packet
classification mechanism for the networking layer. There is nervousness
about adding interpreters to the kernel which nftables will have to
overcome, but that can probably be done.
Other topics mentioned included tproxy, which gained IPv6 support in
2.6.37, and Open vSwitch, a "dumb
switch" with most of the smarts in user space.
James Bottomley discussed the Storage, filesystems and memory
management minisummit held in August. That event was covered here in
detail at the time; readers are referred there for more detail.
Len Brown talked about the power management meeting, also held in
August. Len's notes on the event were published here; that will be the place to look
for a complete report.
There have been two embedded minisummits held in the last couple of
months, one in Tokyo and the other in Cambridge. There was a lot of talk
at these events about "version spread," the wide variety of kernels used by
various embedded vendors. A commitment was made, apparently, to get more
out-of-tree stuff merged upstream. Some vendors are said to be actively
interested in working closer with upstream; Google's merging of the Tegra
architecture code was given as an example.
There was apparently much complaining about the lack of support from
Several vendors have agreed to standardize on an "embedded kernel flag
version" for long-term maintenance. The enterprise vendors standardized on
2.6.32 with good effect; the embedded side wants to do the same with
2.6.35. Someday it might be nice if everybody picked the same kernel, but
that's a task for the future. There is also a renewed push to get the
yaffs2 filesystem cleaned up and merged; the recent patch posting was a bit
of a surprise to the people working on this task, though.
A wireless networking summit was held in San Francisco in September;
John Linville gave the report. The big issue in wireless appears to be WiFi Direct, a set of
protocols for ad hoc networking and resource discovery. The feature
is likely to be called "WiFi peer-to-peer" in Linux to avoid the use of the
trademarked name. It's a set of Bluetooth-like protocols which allow things like
printer sharing; someday it may eliminate the need for access points
altogether. Much of this work has been done, but the user-space interface
work is not yet there.
Mauro Carvalho Chehab reported on the Video4Linux minisummit. There
was, it seems, a lot of attendance from system-on-chip companies. That
reflects a general change in emphasis in that subsystem; it was initially
focused on desktop video, but embedded is taking over.
The Video4Linux1 API was implemented in 1999; it has been slated for
removal ever since Video4Linux2 hit the scene in 2002. In 2011, it's
actually happening: the V4L1 API will be removed in 2.6.38. There will be
user-space support for this API by way of an add-on library; the community
will also take on support for camorama, one of the few
remaining V4L1 applications.
There was talk of the new "media controller" API, meant to facilitate the
configuration of streams within SoC systems. There are no users of this
API yet, so Mauro is holding it out of the kernel for now; it may be merged
in 2.6.38. The related "mediabus" API has gone in for 2.6.37. Also merged
over the last year is the large set of out-of-tree infrared controller
drivers; these are being migrated toward the new remote controller
subsystem. A future issue is supporting the remote controller protocol
added to HDMI 1.3. Another new API is videobuf2, which is currently being
reviewed and should be merged soon.
Finally, Mauro talked about the problem of routing video output to
framebuffer devices. There is no consensus on how this should be done, and
the existing overlay code is unmaintained. There is interest in this
capability, though, so the problem is not going to go away.
There was also a brief update on the tracing minisummits presented
by Masami Hiramatsu. New work in tracing includes the kernelshark tool and
improvements to the perf probe command. Areas of ongoing
interest include the merging of uprobes and utrace, and user-space tracing
in general. Linus commented that one of the best features of perf
is the ease with which it can be used; there is nothing similar for
tracing. His hope is that tracing will move in that direction over time.
to post comments)