MkLinux

microkernel : core collection of OS-neutral abstractions (tasks, threads, ...) upon which can be built OS servers. No file systems, network protocol services etc. ==> more portable and modular

• OSF Research Institute : OSF MK, based on CMU's Mach 3.0
• 1995 : OSF MK ported to PowerMac
• OSF MK :
  • POSIX compatible FIFO and round-robin scheduling
  • CORDS framework for networking
  • configurable number of priorities (32 -> 1024)
  • counting semaphores and locks for inter-task synchronization
  • support for multiple clocks and time-out service

Steps to port OSF MK to PowerMac

Development environment and compiler tool chain : cross-compiling on x86 with Linux and HP700 with OSF/1 using GCC 2.7.1(supports PowerPC code using ELF). Debugging on a remote station (serial line) using a crossed version of GDB

• Building of a minimal kernel, allowing to run the Microkernel Performance Test Suite

Device drivers (rather platform dependent -> device drivers of other platforms could be reused using some PowerMac-dependent DMA code)

• Trade-off : debuggability and simplicity instead of performance and functionality

Linux Server architecture

SINGLE Server (the entire Linux functionality is one single Mach task). Alternative : multi-server design, with smaller specialized tasks communicating thru Mach RPC. The development team chose the former approach : maximize code reuse of the existing monolithic kernel.

MULTI-THREADED server : rely on Mach threads and user-mode cthreads library (because it offers various synchronization primitives)

• dedicated threads :
  • jiffies
  • to process replies to asynchronous device requests
  • to process incoming network packets
  • to process pager requests
  • to dispatch FAKE interrupts
  • to process system call and exception messages from user tasks
  • idle thread, to wake up timed-out tasks and tasks with pending signals

• System call redirection : communication between user tasks and the Linux Server task is a critical performance issue.

Mach 3 : uses a system call emulation facility which redirects control flow to an emulation library, that resides IN THE USER ADDRESS SPACE and communicates with the server via Mach RPC calls ==> good performance, but complex and a security issue.

INSTEAD : OSF MK extends the Mach exception mechanism : a system call from a user task raises an EXCEPTION and enters the microkernel. The microkernel sends an exception RPC to the Linux server.

• User memory access : The Linux server cannot easily access the memory of its user processes.
  • monolithic linux : segment registers

Linux server : has to use the Mach VM interface ==> PERFORMANCE ISSUE (switch to microkernel for each Server access to user task memory)

• SOLUTION : user memory areas mapped into the server's address space.

• Device Access : major/minor device numbers versus Mach device names ?
  • SOLUTION : generic emulation layer added to the Linux server, replacing the bottom half of most of the Linux device drivers

Scheduling : is a microkernel duty. The Linux "schedule" is still used to scan the task list for runnable tasks, but no actual context switching. User tasks BLOCK when running in one of the Linux server threads, using Mach synchronization primitives, provided by the cthreads library.

• Fake interrupts :

in monolithic Linux, pending signals are delivered to a user process when it returns from kernel mode to user mode. A user process can enter kernel mode when issuing a system call (causes an exception) or when interrupted by hardware.

Linux Server : can never be interrupted by hardware ! ==> SERIOUS PROBLEM : a user process looping in user-mode without doing any system calls or exceptions is virtually unkillable because signals will never be delivered to it !

SOLUTION : FAKE INTERRUPTS to allow the Linux server to regain control on a user process even if it does not cooperate. A user thread can be suspended using the Mach thread_abort_safely service.

• Jiffies emulation :
  • monolithic Linux : jiffies global variable counts clock ticks since system startup. VERY IMPORTANT AND WIDELY USED !
  • Linux Server : does not receive clock interrupts.
  • SOLUTION : a dedicated jiffies thread that is woken up by the microkernel and updates the jiffies global variable

VM emulation : page management is totally hidden by the microkernel. A task can create memory objects and map them in address space.

• monolithic Linux widely references the page tables.

Linux Server : to minimize changes, dumb emulation of the page table macros and routines. The Linux Server does not really use it (most of the time they are empty !), but it allows to compile more Linux code.

Dynamic Buffer cache : the Linux Server has no way to know that the system is short of memory. The microkernel does not report memory shortage. Classic problem of Mach servers : double paging

Linux Server on the PowerMac

It was first developped on the Intel platform. Once the microkernel was stable enough on the PowerMac, the two had to be married.

• The major part of porting the Linux Server was in the modification of the header files.

Major difference with native Linux : the Linux Server offers the full 32-bit address space, rather than the restricted 31-bit offered by native Linux

• Performance :
  • AIM III benchmark : 93 % of the performance of native Linux :)

Disk I/O performance : POOR POOR POOR ! Due to extra overhead of the Mach device interfaces, there's only one block read per disk revolution :(

• Status :
  • Linux device drivers should be included in Mach
  • ODE environment : still required for building the OSF MK