[TUHS] Harvard and Von Neumann Architectures and Unix
jnc at mercury.lcs.mit.edu
Tue Nov 28 02:11:41 AEST 2017
> From: Doug McIlroy
> But if that had been in D space, it couldn't have been executed.
Along those lines, I was wondering about modern OS's, which I gather for
security reasons prevent execution of data, and prevent writing to code.
Programs which emit these little 'custom code fragments' (I prefer that term,
since they aren't really 'self-modifying code' - which I define as 'a program
which _changes_ _existing_ instructions) must have some way of having a chunk
of memory into which they can write, but which can also be executed.
> Where is the boundary between changing one instruction and changing them
> all? Or is this boundary a figment of imagination?
Well, the exec() call only overwrites existing instruction memory because of
the semantics of process address space in Unix - there's only one, so it has
to be over-written. An OS operating in a large segmented single-level memory
could implement an exec() as a jump....
BTW, note that although exec() in a single address-space OS is conventionally
something the OS does, this functionality _could_ be moved into the user
space, provided the right address space primitives were provided by the OS,
e.g. 'expand instruction space'. So the exec() code in user space would i)
find the executable, ii) see how much of each kind of memory it needs, iii)
get the OS to give it a block of memory/address space where the exec() code
can live while it's reading in the new code, iv) move itself there, v) use
standard read() calls to read the new image in, and then vi) jump to it.
Yes, it's probably simpler to implement it in the OS, but if one's goal is to
minimize the functionality in the kernel...
More information about the TUHS