Wednesday, Debtember 08, 1999
Exclusion
For those of you who are interested in rolling your own operating system, and if you insist on writing it all in Assembly on your typical IBM PCompatible (and the most popular type is your 32-bit multitasking operating system) please, please, do yourself a favor—avoid using CLI and STI as locking primitives. It fails miserably on SMP-based machines (which I predict will become a commoditized item, much like older 486 and early Pentium machines are now, in the next five years).
Better to use the following:
lock_for_here db 0 ; code code code mov al,1 .spin xchg al,[lock_for_here] cmp al,0 jne .spin ; critical code here ... mov byte ptr [lock_for_here],0
This actually has several advantages over using CLI and STI:
- XCHG locks the bus so that the exchange is atomic with respect to other CPUs on the bus.
- XCHG is not a protected mode instruction, unlike CLI and STI—therefore the sequence can even be used in non-priviledged code. While this may seem silly when you're writing a kernel, not every CPU has a duo user/superuser model. The 80x86 line does support four different priviledged modes.
- Interrupts can still be serviced.
- It doesn't hurt that much on a single-CPU system if it's programmed properly. Sure, the worst case is a deadlocked system, but you're careful to avoid deadlocks, right?
Better yet, wrap them up into functions. That case, if you do need to change how to lock critical sections (say if CLI/STI does prove faster for a single CPU system) you can easily change it in one location.
And it makes it just that much easier to port.