A microkernel is a minimal computer operating system kernel which, in its purest form, provides no operating-system services at all, only the mechanisms needed to implement such services, such as low-level address space management, thread management, and inter-process communication (IPC). If the hardware has a kernelmode-usermode distinction, the microkernel is the only part of the system executing in a kernel mode. The actual operating-system services are provided by "user-mode" servers. These include device drivers, protocol stacks, file systems and user-interface code.
This results in a system structure that is drastically different from the monolithic kernels of the mass market. The latter traditionally have a vertically-layered structure, where applications obtain services by performing a specific system call for each service. In contrast, a microkernel-based system features a horizontal structure, where system services are obtained by executing an IPC system call addressed to a particular server.
Microkernels are closely related to exokernels. They also have much in common with hypervisors, but the latter make no claim to minimality, and are specialized to supporting virtual machines. The L4 microkernel is frequently used as a hypervisor, which indicates that a microkernel is a possible implementation of a hypervisor. The term nanokernel is historically used to differentiate from earlier microkernels which contained actual system services, but the minimality principle used by Jochen Liedtke in the design of the L4 microkernel implies that these terms have the same meaning; microkernel is the modern terminology.
Berkeley UNIX (BSD) began the era of big kernels. In addition to operating a basic system consisting of the CPU, disks and printers, BSD started adding additional file systems, a complete TCP/IP networking system, and a number of "virtual" devices that allowed the existing programs to work invisibly over the network. This growth continued for several decades, resulting in kernels with millions of lines of source code. As a result of this growth, kernels were more prone to bugs and became increasingly difficult to maintain.
The microkernel was designed to address the increasing growth of kernels and the difficulties that came with them. In theory, the microkernel design allows for easier management of code due to its division into user-space services. This also allows for increased security and stability resulting from the reduced amount of code running in kernel mode.
For example, if a networking service crashed due to buffer overflow, only the networking service's memory would be corrupted, leaving the rest of the system still functional. On a traditional monolithic kernel, the overflow could possibly corrupt the memory of other drivers and possibly the kernel itself, which could crash the entire system.
Inter-process communication (IPC) is any mechanism which allows separate processes to communicate with each other, usually by sending messages. (Shared memory is strictly speaking also an inter-process communication mechanism, but the acronym IPC usually only refers to message passing, and it is the latter that is particularly relevant to microkernels.) This allows the operating system to be built from a number of small programs called servers, which are used by other programs on the system, invoked via IPC. Most or all support for peripheral hardware is handled in this fashion, with servers for device drivers, network protocol stacks, file systems, graphics, etc.
IPC can be synchronous or asynchronous. Asynchronous IPC is analogous to network communication: the sender dispatches a message and continues executing. The receiver checks (polls) for the availability of the message by attempting a receive, or is alerted to it via some notification mechanism. Asynchronous IPC requires that the kernel maintains buffers and queues for messages, and deals with buffer overflows; it also requires double copying of messages (sender to kernel and kernel to receiver). In synchronous IPC, the first party (sender or receiver) blocks until the other party is ready to perform the IPC. It does not require buffering or multiple copies, but the implicit rendezvous can make programming tricky. Most programmers prefer asynchronous send and synchronous receive.
First-generation microkernels typically supported synchronous as well as asynchronous IPC, and suffered from poor IPC performance. Jochen Liedtke identified design and implementation of the IPC mechanisms as the underlying reason for this poor performance. In his L4 microkernel he pioneered techniques that lead to an order-of-magnitude reduction of IPC costs. These include an IPC system call that supports a send as well as a receive operation, making all IPC synchronous, and passing as much data as possible in registers. Furthermore, Liedtke introduced the concept of the direct process switch, where during an IPC execution an (incomplete) context switch is performed from the sender directly to the receiver. If, as in L4, part or all of the message is passed in registers, this transfers the in-register part of the message without any copying at all. Furthermore, the overhead of invoking the scheduler is avoided; this is especially beneficial in the common case where IPC is used in an RPC-type fashion by a client invoking a server. Another optimization, called lazy scheduling, avoids traversing scheduling queues during IPC by leaving threads that block during IPC in the ready queue. Once the scheduler is invoked, it moves such threads to the appropriate waiting queue. As in many cases a thread gets unblocked before the next scheduler invocation, this approach saves significant work. Similar approaches have since been adopted by QNX and Minix 3.
In a client-server system, most communication is essentially synchronous, even if using asynchronous primitives, as the typical operation is a client invoking a server and then waiting for a reply. As it also lends itself to more efficient implementation, modern microkernels generally follow L4's lead and only provide a synchronous IPC primitive. Asynchronous IPC can be implemented on top by using helper threads. However, versions of L4 deployed in commercial products have found it necessary to add an asynchronous notification mechanism to better support asynchronous communication. This signal-like mechanism does not carry data and therefore does not require buffering by the kernel.
As synchronous IPC blocks the first party until the other is ready, unrestricted use could easily lead to deadlocks. Furthermore, a client could easily mount a denial-of-service attack on a server by sending a request and never attempting to receive the reply. Therefore synchronous IPC must provide a means to prevent indefinite blocking. Many microkernels provide timeouts on IPC calls, which limit the blocking time. In practice, choosing sensible timeout values is difficult, and systems almost inevitably use infinite timeouts for clients and zero timeouts for servers. As a consequence, the trend is towards not providing arbitrary timeouts, but only a flag which indicates that the IPC should fail immediately if the partner is not ready. This approach effectively provides a choice of the two timeout values of zero and infinity. Recent versions of L4 and MInix have gone down this path (older versions of L4 used timeouts, as does QNX).
A basic set of servers for a general-purpose microkernel includes file system servers, device driver servers, networking servers, display servers, and user interface device servers. This set of servers (drawn from QNX) provides roughly the set of services offered by a monolithic UNIX kernel. The necessary servers are started at system startup and provide services, such as file, network, and device access, to ordinary application programs. With such servers running in the environment of a user application, server development is similar to ordinary application development, rather than the build-and-boot process needed for kernel development.
Additionally, many "crashes" can be corrected for by simply stopping and restarting the server. (In a traditional system, a crash in any of the kernel-resident code would result in the entire machine crashing, forcing a reboot.) However, part of the system state is lost with the failing server, hence this approach requires applications to cope with failure. A good example is a server responsible for TCP/IP connections: If this server is restarted, applications will experience a "lost" connection, a normal occurrence in networked system. For other services, failure is less expected and may require changes to application code. For QNX, restart capability is offered as the QNX High Availability Toolkit.
In order to make all servers restartable, some microkernels have concentrated on adding various database-like techniques like transactions, replication and checkpointing in order to preserve essential state across single server restarts. An example is ChorusOS, which was targeted at high-availability applications in the telecommunications world. Chorus included features to allow any "properly written" server to be restarted at any time, with clients using those servers being paused while the server brought itself back into its original state. However, such kernel features are incompatible with the minimality principle, and are therefore not provided in modern microkernels, which instead rely on appropriate user-level protocols.
Device drivers frequently perform direct memory access (DMA), and therefore can write to arbitrary locations of physical memory, including over kernel data structures. Such drivers must therefore be trusted. It is a common misconception that this means that they must be part of the kernel. In fact, a driver is not inherently more or less trustworthy by being part of the kernel.
While running a device driver in user mode does not necessarily reduce the damage a misbehaving driver can cause, in practice it is beneficial for system stability in the presence of buggy (rather than malicious) drivers: memory-access violations by the driver code itself (as opposed to the device) may still be caught by the memory-management hardware. Furthermore, many devices are not DMA-capable, their drivers can be made untrusted by running them in user mode. Recently, an increasing number of computers feature IOMMUs, many of which can be used to restrict a device's access to physical memory. (IBM mainframes have had IO MMUs since the IBM System/360 Model 67 and System/370.) This also allows user-mode drivers to become untrusted.
User-mode drivers actually predate microkernels. The Michigan Terminal System (MTS), in 1967, supported user-space drivers, the first operating system to be designed with that capability. Historically, drivers were less of a problem, as the number of devices was small and trusted anyway, so having them in the kernel simplified the design and avoided potential performance problems. This led to the traditional driver-in-the-kernel style of UNIX, Linux, and Windows. With the proliferation of various kinds of peripherals, the amount of driver code escalated and in modern operating systems dominates the kernel in terms of code size.
As a microkernel must allow building arbitrary operating-system services on top, it must provide some core functionality. At the least this includes:
A concept is tolerated inside the microkernel only if moving it outside the kernel, i.e., permitting competing implementations, would prevent the implementation of the system's required functionality.Everything else can be done in a user program, although device drivers implemented as user programs may require special privileges to access I/O hardware.
Related to the minimality principle, and equally important for microkernel design, is the separation of mechanism and policy, it is what enables the construction of arbitrary systems on top of a minimal kernel. Any policy built into the kernel cannot be overwritten at user level and therefore limits the generality of the microkernel. Policy implemented in user-level servers can be changed by replacing the servers (or letting the application choose between competing servers offering similar services).
For efficiency, most microkernels contain schedulers and manage timers, in violation of the minimality principle and the principle of policy-mechanism separation.
Start up (booting) of a microkernel-based system requires device drivers, which are not part of the kernel. Typically this means that they are packaged with the kernel in the boot image, and the kernel supports a bootstrap protocol that defines how the drivers are located and started. Some microkernels simplify this by placing some key drivers inside the kernel (in violation of the minimality principle), LynxOS and the original Minix are examples. Some even include a file system in the kernel to simplify booting.
A key component of a microkernel is a good IPC system. Since all services are performed by usermode programs, efficient means of communication between programs are essential, far more so than in monolithic kernels. The design of the IPC system makes or breaks a microkernel. To be effective, the IPC system must not only have low overhead, but also interact well with CPU scheduling.
Performance is therefore a potential issue in microkernel systems. Indeed, the experience of first-generation microkernels such as Mach and Chorus showed that systems based on them performed very poorly. However, Jochen Liedtke showed that Mach's performance problems were the result of poor design and implementation, and specifically Mach's excessive cache footprint. Liedtke demonstrated with his own L4 microkernel that through careful design and implementation, and especially by following the minimality principle, IPC costs could be reduced by more than an order of magnitude compared to Mach. L4's IPC performance is still unbeaten across a range of architectures.
While these results demonstrate that the poor performance of systems based on first-generation microkernels is not representative for second-generation kernels such as L4, this constitutes no proof that microkernel-based systems can be built with good performance. It has been shown that a monolithic Linux server ported to L4 exhibits only a few percent overhead over native Linux. However, such a single-server system exhibits few, if any, of the advantages microkernels are supposed to provide by structuring operating-system functionality into separate servers.
A number of commercial multi-server systems exist, in particular the real-time systems QNX and Integrity. No comprehensive comparison of performance relative to monolithic systems has been published for those multiserver systems. Furthermore, performance does not seem to be the overriding concern for those commercial systems, which instead emphasize simplicity for the sake of robustness. An attempt to build a high-performance multiserver operating system was the IBM Sawmill Linux project. However, this project was never completed.
It has been shown in the meantime that user-level device drivers can come close to the performance of in-kernel drivers even for such high-throughput, high-interrupt devices as Gigabit Ethernet. This seems to imply that high-performance multi-server systems are possible.
Consequently, microkernel designs have been used for systems designed for high-security applications, including KeyKOS, EROS and military systems. In fact common criteria (CC) at the highest assurance level (EAL7) has an explicit requirement that the target of evaluation be “simple”, an acknowledgment of the practical impossibility of establishing true trustworthiness for a complex system.
Recent work on microkernels is focusing on formal specifications of the kernel API, and formal proofs of security properties of the API. The first example of this is a mathematical proof of the confinement mechanisms in EROS, based on a simplified model of the EROS API. More recently, a comprehensive set of machine-checked proofs has been performed of the properties of the protection model of the seL4 version of L4.
Some projects go even further, aiming at complete formal verification, i.e. a mathematical proof that the kernel's implementation is consistent with its specification, which then provides a guarantee that the properties proved about the API actually hold for the real kernel. This degree of assurance goes beyond even CC EAL7. Such proofs are being attempted for Coyotos and seL4