Day two of the 2011 Kernel Summit featured a changed format: the session
was open to all attendees and the room was reconfigured into a classroom
format to accommodate the additional people. Your editor led off the this
day with an whirlwind recap of the first day's discussions and decisions.
Everything said has been covered in much more detail here, so there is little
point in repeating it.
Next on the agenda was the traditional minisummit report session. Quite a
few smaller gatherings have happened in the last year, three in the day
immediately prior to the kernel summit. Reports were heard from these
- ARM subarchitecture maintainership: this was the first
gathering of most of the ARM subarchitecture maintainers ever. Grant
Likely allowed as to how ARM has not always been the most united
community within the kernel, but it has grown up a lot in the last
couple of years.
Quite a few topics were discussed at this gathering. A holy grail of
ARM development is the creation of a single binary kernel image that
can boot on any ARM CPU. That goal will probably never be completely
achieved, but progress can be made, especially on the newer
processors; even then, it will take years to accomplish. The
transition to device tree use was on the agenda, as was the
maintainership of ARM SoC subarchitectures. Contemporary hardware is
stressing the Linux device model, especially when it comes to power
management. A lot of changes will need to be made, but Grant thought
it was an encouraging development: embedded developers are
increasingly working with and improving the core kernel design. Other
topics included the CMA patch set, unifying struct clk, the
proposed pin control subsystem, and future architectures.
- BlueTooth was covered briefly by Marcel Holtmann, who noted
that much of the stack has seen little change for ten years. On the
horizon, though, are BlueTooth 3.0, which is oriented toward speed,
and BlueTooth 4.0, which, instead, is aimed at low power use. Some
initial code has been merged, more is to come.
- Video4Linux2 was presented by Mauro Carvalho Chehab. The V4L2
developers have a lot to work on, including support for
high-resolution "snapshot" modes, new digital video broadcast
standards, and migration away from the videobuf1 subsystem. Perhaps
the most interesting problem, though, is applications for complex
devices using the media controller interface. Currently only
proprietary applications know enough about any specific hardware
device to be able to configure and use it - not an entirely
satisfactory situation. There will be an attempt to move some of the
configuration and capture functionality into the libv4l library so that
ordinary applications can use the hardware. New drivers will need to
come with libv4l plugins to enable their devices.
- Filesystems, storage, and virtual memory were covered by Chris
Mason, James Bottomley, and Andrea Arcangeli, respectively. That
event was covered here when it
happened; readers interested in details can find them there.
- Avi Kivity reported on the KVM summit. Issues of interest in
that community include ports to other architectures, a new device
assignment framework that can enable the development of safe
user-space drivers, the increasing integration between KVM/QEMU and
various other kernel components, and security concerns. On that last
front, people are using KVM to confine virtual machines, but it would
be good to have more layers of defense - KVM and QEMU are a large code
base that is hard to verify on its own. So there is a lot of interest
in better sandboxing; a number of
techniques are being considered.
- James Morris gave a summary of the security summit held prior to the
Linux Plumbers Conference in September. Issues of interest there were
integrity management, kernel hardening (see LWN's report
from that discussion), a new effort to fix up and merge the grsecurity
patch set, and security module stacking.
- Thomas Gleixner reviewed the just-concluded realtime minisummit; see
LWN's report for all the details.
- Jesse Barnes gave a brief report from a gathering of graphics
developers; themes of interest there were the idea (probably not to be
implemented) of bringing some older graphics drivers into the kernel
tree, adding kernel mode setting support for overlays, and "metamodes"
that capture the workable combinations of mode settings when more than
one output is in use.
- John Linville covered the wireless minisummit. Wireless, he
said, is getting boring these days - and that is a good thing.
Developers in that area are concerned with supporting various new
protocols, handling of regulatory issues, and improving the
maintenance of the wpa_supplicant daemon.
The kernel.org administrators held a half-hour session to answer any
questions that the larger group may have about the recovery of that site.
Most of the questions had to do with key signing. H. Peter Anvin said that
a lot of the people who are complaining about the whole process have not
taken the time to really learn about how it works. That said, some people
have real problems; that is especially true of geographically remote
developers who are unable to travel to a place where they can get their
keys signed. Peter is doing his best to help those people. The
key signing to be held later in the day, he said, was most important; it
would create a large web of tightly-connected developers who could then
extend that web around the world.
How many signatures are required to make a key trustworthy? "As many as
possible." Linus added that, for emailed patches, signatures really do not
matter. The important thing for those patches is the code - maintainers
should be looking at the patch, and not the signature. There are plans to
set up an escrow service for revocation certificates, but that is far from
the first priority at this point. A question about the best expiration
time for keys got an answer of "three to five years." It was noted that
developers can create an eternal key that they keep well locked up, then
use subkeys with expiration dates for their daily business.
Rafael Wysocki's talk on regression tracking has become a kernel summit
tradition. The overall picture remains quite stable; most kernels
have 20-30 outstanding regressions one full cycle after their release. The
number of short-term regressions seems to have fallen a bit in recent
times; Rafael chalked that up to the fact that the graphics drivers have
finally stabilized somewhat.
The numbers showed that, as a whole, regressions are not being found as
quickly as they were a year or so ago. It was asked: is that good or bad?
Perhaps it is bad; we're just not as good at finding bugs as we once were.
Linus, instead, said that it's a sign that the bugs we are seeing are
becoming more subtle; they don't affect as many users, and, thus, take more
time to find. Rafael thought it may mean that we were getting more testing
of post-release kernels than before, but he wasn't sure. He was convinced
that the trend meant something, but couldn't say what.
Regression tracking has been badly hampered by the loss of the bugzilla
system when kernel.org went down. Even before then, though, the list of
complaints about that bugzilla was long. Rafael said he didn't see a whole
lot of value in restoring the bugzilla in its previous form; it would be
better to start over with an improved installation. He had a long wishlist
of things he wished would work better; the kernel.org administrators said
they would do their best, but couldn't make any promises.
Next: Shared libraries, failure handling, media
controller, kbuild, and the future of the kernel summit.
to post comments)