Porting Unix to the 386: Designing the Software Specification - User to Kernel Communication Primitives

User to Kernel Communication Primitives By arranging our address space as outlined, we've greatly simplified the routines that communicate between kernel and user process (now the kernel routines can directly access user space). All that is needed is a way to determine if a selected portion of user memory may be read or written before it is attempted. On some machines (such as the VAX) special instructions are available for this purpose. The 386, however, offers instructions only for use in validating segments, not pages. So we must use a different strategy.

In 386BSD, we chose to set a global variable (nofault) to a nonzero value. If a fault happens during any user/kernel communication primitive, it transfers to the address held within no fault. In this way we can catch illegal references by using the microprocessor's own address translation mechanism to find them, instead of by tedious code evaluation on every reference.

Unfortunately, one idiosyncrasy of the 386 now rears its ugly head. The designers of the 386 decided that segment attributes should be used to ultimately determine access to regions in a process, thus making their use mandatory in the system even if we don't need them. To be precise, we have page attribute bits that can be used for protection. These work as expected, unless the 386 is run in supervisor mode (as does the kernel). In this case, only the valid/invalid attribute has any effect. This nuisance or "feature" requires a bit of workaround to make the primitives complete.

<<BACK

NEXT >>