Operating system modifications

The UNIX operating systems were originally implemented for uniprocessor machines, so a number of kernel changes were required to support multiprocessors. These changes are implemented in all SCO OpenServer 5 Release 5 kernels as well as all Open Desktop Release 3.0 systems and the SCO UNIX Release 3.2v4.2 kernel:

• Mutual exclusion functions are required to protect sections of critical code (see ``Critical code section''). On a uniprocessor system, preventing interrupts from being received guarantees that a section of critical code will finish execution, but on a multiprocessor system, other CPUs could generate interrupts even if they are disabled on the current processor.

  Multithreaded drivers must use these locking functions to protect critical code; see ``Protecting critical code''.

• These same mutual exclusion locking functions are used in other kernel subsystems to ensure that data structures are not accessed simultaneously by multiple processes. Resources that are protected in this manner include memory management data structures such as page table and page directory entries, other kernel data structures such as the proc table and the u-area, I/O ports, and software resources such as inodes.

  System calls such as read(SS) and write(SS) include locking functions around resource access such as updating an inode. In some cases, algorithms were modified to reduce the amount of time that a lock is held, which is critical for good scaling across multiple processors.

• Process scheduling algorithms have been modified to determine where a process is scheduled as well as when it is scheduled. One queue of processes that are ready to run is maintained for all processors; each CPU selects processes from this queue as if there were no other CPUs on the system. Queue integrity is maintained by protecting queue insertions and removals with the mutual exclusion functions such as lockb(D3oddi).

  Processes that require special device access, such as writing directly to video memory, are restricted to the default processor. The default processor is the CPU that can currently support a given operation. Operations that have not been multithreaded are also restricted to the default processor for the duration of the operation.

  All main memory can be accessed by all processors. Cache coherency is maintained by the hardware so that the cache is transparent to the kernel. Mutual exclusion functions such as lockb(D3oddi) and unlockb(D3oddi) are used to ensure that page tables and directories, memory free lists, and other associated structures are not modified simultaneously by different processors.