Master release schedule changes, major feature list - September 2009
Matthew Dillon
dillon at apollo.backplane.com
Wed Jul 1 19:23:52 PDT 2009
We are going to be shifting our master release schedule 2 months forward.
This is being done for several reasons. We will be releasing in
September and March instead of July and January.
This will give us a bit more time to sync up with pkgsrc quarterlies and
also moves our releases out of the dead months (high summer and
christmas) and into months where developers are more active.
So we will be releasing 2.4 in September rather then this month. There
will be a lot of stuff in 2.4, and hopefully considerbly more by the
time the release actually rolls around:
* Jordan's AMD64 work: 64 cpu bit support for UP, hopefully 64 bit
support for SMP. Probably not 32 bit emulation under 64 bits by
September though.
* Alex's work: DEVFS. I believe sufficient progress will be made in the
next two months to make it useable.
* Aggelos's NETMP work.
* Sephe's network driver work and SMP work.
* Sascha's installer improvements making a BOOT + HAMMER setup the
default installation method. Also fdisk can now be used on drives
larger then 2TB and the kernel recognizes maxed-out fdisk parameters
so disklabel64 will recognize the actual size of the disk if the fdisk
slice is maxed out.
* Additional work by me to help integrate serial number support into
DEVFS (once Alex gets a little bit further on it), and adopting
OpenBSD's /etc/devtab methodology.
The serial number work is going to become *VERY* important as HAMMER
development moves into the cluster-FS phase. At that point drives
will be able to exist anywhere on the network. Even without that
we have a real need to start addressing drives by serial number simply
due to the fact that one can trivially connect many drives to the
motherboard now via port multiplier enclosures. Counting on the
/dev/daXXX enumeration is simply not reliable.
* 2.4 will feature a complete pkgsrc rebuild using libc.so.7. The
2.3 development branch will also get this pkgsrc build in the next
few days (if not today). Thanks to Justin and Hasso.
* We will officially deploy the AHCI and the SILI driver. These drivers
are working now for most MBs but there are still a few open bug reports
with certain chipsets and devices. For people with AHCI-capable
chipsets on their motherboards, or with Sili 3132 boards (other Silicon
Image parts will be added as well), the new disk drivers feature
NCQ, hot-swap, port-multiplier, and hot-swap targets behind the
port-multiplier.
* Hasso's pkgsrc work and related improvements in libthread_xu to build
more packages out of the box.
* My HAMMER work, improving directory tree traversal times, bug fixes,
reduced stress on the machine when reblocking the disk, and so forth.
* Lots of bug fixes improving stability.
More on BOOT+HAMMER
All the BSDs have been struggling with how to boot into new filesystems,
but more to the point they've been struggling with issues of booting into
redundant drive setups (RAID arrays of various sorts).
What I have decided is that the most flexible way to deal with booting
is to adopt a Linux-like methodology where a small UFS BOOT partition
on 'a' is created containing the boot loader, kernel, and modules,
and the ROOT partition will be placed on 'd'. This also allows the
swap partition to remain on the outer cylinders of the disk in the 'b'
partition.
This gives us a very flexible basis for mounting the root partition,
including being able to more conveniently use software raid features
and more complex arrangements for the root partition and other mounts.
DragonFly will not officially support booting from BIOS-fakeraid. IMHO
my opinion is that anyone who needs *real* redundancy for the boot loader
and kernel should simply boot from a SATA SSD (SATA based Solid State
Drive). SATA SSDs have the same MTBF as the motherboard, so there's no
point creating a redundant SSD setup. One works fine. This might be
perceived as a negative but I consider it reality. One can't even
safely migrate the physical disks from a fake-raid setup to another
machine.
On the plus side, while our software RAID support is still basically just
vinum, one fringe benefit of using HAMMER is that software RAID is
actually a lot easier to implement because it does not have to be
transactional. For example, it would be ok for a mirror to become
inconsistent due to a crash/reboot (though not during the course of
normal operation). HAMMER's recovery on reboot will effectively
resynchronize the needed bits when it runs its UNDO. So while we are
a bit behind on the technology in this area HAMMER actually gives us
a considerable advantage in developing soft-raid systems that are
reliable.
All of this will eventually be combined with the serial number support
and network disk and upcoming cluster-FS work.
-Matt
Matthew Dillon
<dillon at backplane.com>
More information about the Users
mailing list