This is a legitimate discussion
I agree with Ken Hagan AND De Zeurkous.
I quibble with Ken Kagen only in the statement about "implementing critical security functions in userland", because we're talking about syscall wrappers here - this is kernelland.
Ken Kagen took everyone down the right road by pointing out that CALL atomically places arguments onto a privileged stack, which system calls read and act on, therefore seemingly no opportunity for exploit. (but this is actually not quite correct - see below)
What Robert Watson is mostly discussing is
a. what happens when you WEDGE something between the two?
b. Particularly when you can have concurrent processes with same access to memory where the args are?
When a userland process makes a syscall, the argument for strings/structures is merely a pointer to the string/structure in userland memory area.
While the pointer could not be changed due to the stack being used, the memory being pointed at, being in userland, certainly could.
If the execution sequence is
wrapper (reads string, accepts argument as valid, or changes them)
<WINDOW OF OPPORTUNITY>
WINDOW 1: wrapper interrupted (yields time, scheduler interrupts, etc), wrapper gets a CPU back
WINDOW 2: another process, concurrently running in another CPU, changes userland memory
SYSCALL (real one) triggered
WINDOW #1 is the single CPU window.
Window #2 is the multi CPU/multi core window.
In these windows, another process with userland access (a child process with same memory access) can replace the string/structure contents before the real SYSCALL gets triggered.
So, how do syscall wrapper application developers make sure there is no window of opportunity, particularly WINDOW #2?
As De Zeurkous says, "these problems arise from incorrect design".
The way this kind of issue is handled in a kernel is to use semaphore locking architecture, but this doesn't seem to work here, because you'd have to put semaphore structures around arbitrary userland memory used as arguments to function calls, and you'd have to trust userland apps to obey the semaphore.
A better approach would be to have the syscall wrapper application transfer data structures themselves (copy entire structures, not just pointers to structures), into protected syscall wrapper memory.
There may be even better approaches - but the point about an "incorrect design" being used to insert a wrapper between userland and kernel syscall does seem valid to me.