DragonFly BSD

DragonFlyBSD

DragonFlyBSD Logo with Fred, the dragonfly.
Developer	Matt Dillon
OS family	BSD
Working state	Current
Source model	Open source
Latest release	1.2 / April 8, 2005
Repository	gitweb.dragonflybsd.org/dragonfly.git
Kernel type	Hybrid
License	BSD
Official website	http://www.dragonflybsd.org

In computing, the DragonFly BSD operating system is a fork of FreeBSD. Matt Dillon, a long-time FreeBSD and Amiga developer, started work on DragonFly BSD in June 2003 and announced it on the FreeBSD mailing lists on 16 July 2003.

Dillon started DragonFly in the belief that the methods and techniques being adopted for threading and SMP in FreeBSD 5 would lead to a poorly performing system that would be very difficult to maintain. He sought to correct these suspected problems within the FreeBSD project. Due to ongoing conflicts with other developers over the implementation of his ideas in FreeBSD, and other reasons, his ability to directly change the FreeBSD code was eventually revoked. Despite this, the DragonFly BSD and FreeBSD projects still work together contributing bug fixes, driver updates and other system improvements to each other.

Intended to be "the logical continuation of the FreeBSD 4.x series", DragonFly is being developed in an entirely different direction from FreeBSD 5, including a new Light Weight Kernel Threads implementation and a light weight ports/messaging system. Many concepts planned for DragonFly were inspired by the AmigaOS.

Kernel design

Like most modern kernels, DragonFly is a hybrid, containing features of both monolithic and microkernels, attempting to make the best use of both technologies, such as the message passing capability of microkernels enabling larger portions of the OS to benefit from protected memory, as well as retaining the speed of monolithic kernels for certain critical tasks. The messaging subsystem being developed is similar to those found in microkernels such as Mach, though it is less complex by design. DragonFly's messaging subsystem has the ability to act either in a synchronous, or an asynchronous fashion, and uses this capability to attempt to achieve the best performance possible in any given situation.

There is progress being made to provide both device input/output (I/O) and virtual file system (VFS) messaging capabilities that will allow the remainder of the project goals to be met. The new infrastructure will allow many parts of the kernel to be migrated out into userland, whereby they will be more easily debugged as they will be smaller, isolated programs, instead of being small parts entwined in a larger chunk of code. The migration of select kernel code into userspace has the additional benefit of making the system more robust; if a userspace driver crashes, it will not crash the kernel.

System calls are being split into userland and kernel versions, as well as being encapsulated into messages. This will help reduce the size and complexity of the kernel by moving variants of standard system calls into a userland compatibility layer, as well as help maintain forwards and backwards compatibility between DragonFly versions. Linux and other Unix-like OS compatibility code is being migrated out similarly. Multiple instances of the 'native' userland compatibility layer created in jails could give DragonFly functionality similar to that found in User Mode Linux (UML). Unlike UML (which is essentially a port of Linux to itself as if the host kernel was a different hardware platform), DragonFly's virtualization will not require special drivers to communicate with the real hardware on the computer.

Compartmentalization

In DragonFly, threads are locked to CPUs by design, and each processor has its own LWKT scheduler. Threads are never preemptively switched from one processor to another; they are only migrated by the passing of an "Interprocessor Interrupt" (IPI) message between the CPUs involved. Interprocessor thread scheduling is also accomplished by sending asynchronous IPI messages. One advantage to this clean compartmentalization of the threading subsystem is that the processors' on-board caches in SMP systems do not contain duplicated data, allowing for higher performance by giving each processor in the system the ability to use its own cache to store different things to work on.

The LWKT subsystem is being employed to partition work among multiple kernel threads (for example in the networking code; one thread per protocol), reducing contention by removing the need to share certain resources among various kernel tasks. This thread partitioning implementation is arguably the key differentiating feature of DragonFly's design.

Protecting shared resources

In order to run safely on multiprocessor machines, access to shared resources (files, data structures etc.) must be serialized so that threads or processes do not attempt to modify the same resource at the same time. Atomic operations, spinlocks, critical sections, mutexes and serializing tokens are all possible methods that can be used to prevent concurrent access. Whereas both Linux and FreeBSD 5 employ fine-grained mutex models to achieve higher performance on multiprocessor systems, DragonFly does not. In order to prevent multiple threads from accessing or modifying a shared resource simultaneously, DragonFly employs critical sections, and serializing tokens to prevent concurrent access. Until recently, DragonFly also employed SPLs, but these were replaced with critical sections.

Much of the system's core including the LWKT subsystem, the IPI messaging subsystem and the new kernel memory allocator among other things are lockless, meaning that they work without using mutexes. Critical sections are used to protect against local interrupts and operate on a per-CPU basis, guaranteeing that a thread currently being executed will not be preempted.

Serializing tokens are used to prevent concurrent accesses from other CPUs and may be held simultaneously by multiple threads, ensuring that only one of those threads is running at any given time. Blocked or sleeping threads therefore do not prevent other threads from accessing the shared resource unlike a thread that is holding a mutex. Among other things, the use of serializing tokens prevents many of the situations that could result in deadlocks and priority inversions when using mutexes, as well as greatly simplifying the design and implementation of a many-step procedure that would require a shared resource among multiple threads. The serializing token code is evolving into something quite similar to the "Read-copy-update" feature now available in Linux. Unlike Linux's current RCU implementation, DragonFly's is being implemented such that only processors competing for the same token are affected rather than all processors in the computer.

Additional features

Early on in its development, DragonFly acquired a slab allocator, which replaced the aging FreeBSD 4 kernel memory allocator. The new slab allocator requires neither mutexes nor blocking operations for memory assignment tasks, and unlike the code it replaced, it is multiprocessor safe.

DragonFly uses SFBUFs (Super-Fast BUFfers) and MSFBUFs (Multi-SFBUFs). A SFBUF is used to manage ephemeral single-page mappings and cache them when appropriate. They are used for retrieving a reference to data that is held by a single VM page. This simple, yet powerful abstraction gives a broad number of abilities, such as zero-copy achieved in the sendfile(2) system call.

SFBUFs are used in numerous parts of the kernel, such as the Vnode Object Pager and the PIPE subsystems (indirectly via XIOs) for supporting high-bandwidth transfers. An SFBUF can only be used for a single VM page; MSFBUFs are used for managing ephemeral mappings of multiple-pages.

The SFBUF concept was devised by David Greenman of the FreeBSD Project when he wrote the sendfile(2) system call; it was later revised by Dr. Alan L. Cox and Matthew Dillon. MSFBUFs were designed by Hiten Pandya and Matthew Dillon.

Development and distribution

DragonFly forked from FreeBSD 4.8 and imports features and bug fixes from FreeBSD 4 and 5 where appropriate, such as ACPI and a new ATA driver framework from FreeBSD 4. As the number of DragonFly developers is currently small, with most of them focused on implementing basic functionality, device drivers are being kept mostly in sync with FreeBSD 5.x, the branch of FreeBSD where all new drivers are being written. The DragonFly developers are slowly moving toward using the "busdma" APIs, which will help to make the system easier to port to new architectures, but it is not a major focus at this time.

The BSD operating systems are actively replacing "K&R" style C code with modern, ANSI equivalents. Also like OpenBSD, DragonFly's version of the GNU Compiler Collection has an enhancement called the "Stack-Smashing Protector" (formerly known as "ProPolice") enabled by default, providing some additional protection against buffer overflow based attacks. It should be noted that as of Saturday, 23 July 2005, the kernel is no longer built with this protection by default.

Being a derivative of FreeBSD, DragonFly has inherited an easy-to-use integrated build system that can rebuild the entire base system from source with only a few commands. Like people from the other BSD projects, the DragonFly developers use a version control system called CVS to manage changes to the DragonFly source code. Unlike its parent FreeBSD, DragonFly will have both stable and unstable releases in a single source tree, due to a smaller developer base.

Like the other BSD kernels (and those of most modern operating systems), DragonFly employs a built-in kernel debugger to help the developers find kernel bugs. Furthermore, as of Wednesday, 20 October 2004, a debug kernel, which makes bug reports more useful for tracking down kernel-related problems, is installed by default, at the expense of a relatively small quantity of disk space. When a new kernel is installed, the backup copy of the previous kernel and its modules are stripped of their debugging symbols to further minimize disk space usage.

The operating system is distributed as a live CD that boots into a complete DragonFly system. It includes the base system and a complete set of manual pages, and may include source code and useful packages in future versions. The advantage of this is that with a single CD you can install the software onto a computer, use a full set of tools to repair a damaged installation, or demonstrate the capabilities of the system without installing it. Daily snapshots are available from Simon 'corecode' Schubert via FTP and HTTP for those who want to install the most recent versions of DragonFly without building from source.

Like the other free, open source BSDs, DragonFly is distributed under the terms of the modern version of the BSD license.

Releases

Version 1.0 (12 July, 2004).

DragonFly BSD 1.0 was meant to be a "technology showcase" rather than an integrated production release. It featured the new "BSD Installer," the LWKT subsystem and the associated LW ports/messaging system, a mostly MP safe networking stack, lockless memory allocator and the FreeBSD 4.x ports and packages system (which was very briefly broken following the release).

Amiga-style 'resident' application support was added which takes a snapshot of a large, dynamically linked program's virtual memory space after loading, allowing future instances of the program to start much more quickly than it otherwise would have. This replaces the prelinking capability that was being worked on earlier in the project's history, as the resident support is much more efficient. Large programs like those found in KDE with many shared libraries will benefit the most from this support.

Other features introduced in this release include variant symlinks and application checkpointing support.

Due to a serious bug in the installer, an updated 1.0A release of DragonFly was released shortly afterward.

Version 1.2 (8 April, 2005).[1]

This second release of DragonFly contained many bugfixes and new features. New to this release are things such as TCP SACK, ALTQ and PF (OpenBSDs firewall), TLS (Thread Local Storage) support, DCONS support (console over firewire), IPv6 improvements, and the rewritten namecache infrastructure, which is now distinct from the VFS code, and now capable of allowing the DragonFly developers to implement namecache based security mechanisms.

Like the first release, 1.2 still utilizes the FreeBSD ports system for third party packages, but NetBSDs "pkgsrc" now natively supports DragonFly, and is available as an option.

Dillon has stated that this will be the last release of DragonFly that employs the MP lock in common code paths.

Future directions

Supported processors

Currently, DragonFly runs on x86 (Intel and AMD) based computers, both single processor and SMP models. A port to the Athlon 64 (x86-64) has been started, but doesn't exist yet. A port to the PowerPC processor has been speculated about sometime following the eventual x86-64 port.

Package management

DragonFly currently uses FreeBSD's packaged third party software, with NetBSDs "pkgsrc" available as an option. The next version will use pkgsrc [2] as an official package management system. DragonFly is eventually going to have its own native packaging system based on "VFS environments," but this is not expected to be completed for some time, as much of the VFS thread partitioning work remains to be done, which is a prerequisite for the userland VFS (see below for more details).

Threading and messaging

Although both system calls and device I/O have been largely converted to DragonFly's threaded messaging interface, both still operate synchronously. Ultimately, the entire messaging system is to be capable of both synchronous and asynchronous operation.

Userland threading support is also a focus for upcoming releases. Currently DragonFly has only basic userland threading support that does not take advantage of multiprocessor systems (N:1 threading). Work to address this has been ongoing almost since the inception of the project, and a modern implementation is expected for version 1.4. Matt has said that ideally, an M:N implementation is preferable (N userland threads making use of a smaller number of kernel threads (M)), but the support planned for 1.4 is likely going to be a 1:1 implementation; one kernel thread for every userland thread.

Userland VFS and Journaling

Userland VFS - the ability to migrate filesystem drivers into userspace, will take a lot of work to accomplish. Some of this work is already complete, though there is still much to do. The namecache code has been extracted from and made independent of the VFS code, and converting the VFS code to DragonFly's threaded messaging interface is Matt's next major focus. This will be more difficult than converting the device I/O and system calls was, due to the fact that the VFS system inherited from FreeBSD uses a massively reentrant model.

The userland VFS system is a prerequisite of a number of desired features to be incorporated into DragonFly. Dillon envisions a new package management system based at least in part, on "VFS environments" which give the packages the environment they expect to be in, independent of the larger filesystem environment and its quirks. In addition to system call message filtering, VFS environments are also to play a role in future security mechanisms, by restricting users or processes to their own isolated environments.

A new journaling layer is being developed for DragonFly for the purpose of transparently backing up entire filesystems in real-time, securely over the Internet. What remains to be done is the ability to restore a filesystem to a previous state, as well as general stability enhancements. This differs from traditional meta-data journalling filesystems as it will work with all supported filesystems as it is implemented in the VFS layer instead of in the individual filesystem drivers, and also in that it backs up all of the data contained on a disk or partition, instead of just meta-data, allowing for the recovery of even the most damaged of installations.

While working on the journaling code, Dillon realized that the userland VFS he envisioned may be closer than he initially thought, though it is still some ways off.

SSI clustering

Ultimately, Dillon wants DragonFly to natively enable "secure anonymous system clustering over the Internet," and the light weight ports/messaging system will help to provide this capability. Security settings aside, there is technically no difference between messages created locally or on another computer over a network. Achieving this "single-system image" capability transparently will be a big job, and will take quite some time to properly implement, even with the new foundation fully in place. While some of the short term goals of the project will be completed in months, other features may take years to complete.

External links

General information

• The DragonFly BSD Project home page
• Original announcement on the FreeBSD -CURRENT mailing list
• DragonFly Packages from GoBSD.com
• DragonFly Installer
• Original Announcement

News and interviews

• The DragonFly BSD Digest
• 2003 Year-End summary
• Matt Dillon IRC Interview from http://www.slashnet.org/
• OSNews interview (13 March 2004)
• Behind DragonFly BSD interview with DFly developers that was published on O'Reilly OnLamp.com (8 July 2004)
• DragonFlyBSD 1.0A: A strong start

Technology references

• The DragonFly BSD Operating System
• The Design and Implementation of the 4.4BSD Operating System
• The Slab Allocator (Part 1) -- Bonwick94 and The Slab Allocator (Part 2) -- Bonwick2001
• Improving the FreeBSD SMP implementation (DragonFly does things differently)
• Intel's MultiProcessor Specification (PDF)
• Read-Copy Update Mutual Exclusion
• The stack-smashing protector
• The Grid Engine Project
• FreeBSD Jails

Documentation

• The DragonFly BSD handbook
• The DragonFly BSD Wiki
• Printable manual pages and supplementary documentation (Postscript and PDF)

Howtos (From The DragonFly BSD Digest)

• Installing DragonFly BSD from source (Upgrading from FreeBSD 4.x)
• How to enable profiling to help catch bugs
• Getting a kernel crash dump
• Set up a boot splash screen
• How to do a diff