This article attempts to describe the calling conventions of syscalls on Linux/MIPS. Generally it's a less than great idea to directly access the system call interface. Libraries such as glibc, ucLibc and others are doing a great job and will hide all the possible optimizations and version dependencies from the application programmer. You have been warned.
The syscall calling conventions are similar to subroutine calls. That is the set of caller saved and callee saved registers are pretty much the same. However, arguments are passed in $a0 ...$a3 for O32 rsp. $a0 ... $a7 for N32 and N64. The result value is returned in $v0 and $a3 will be set to 0 or 1 to indicate success. The s-registers should be assumed to be saved while the contents of the $t registers should be assumed to have been lost.
A few syscalls deviate from this calling conventions, for example sigreturn which modifies the entire register file or pipe(2) which returns the first file descriptor in $v0 and the second in $v1.
|Register||use on input||use on output|
|$v0||syscall number||return value|
|$v1||—||2nd fd only for pipe(2)|
|$a0 ... $a2/$a7 except $a3||syscall arguments||returned unmodified|
|$a3||4th syscall argument||$a3 set to 0/1 for success/error|
|$t0 ... $t9||(caller saved)|
|$s0 ... $s8||(callee saved)|
And of course a syscall is invoked through a SYSCALL instruction. The kernel assumes the syscall instruction not to be in a branch delay slot, that is, it will not check for branch delay slots and do branch emulation. Even iff it was, performance considerations would make it preferable to leave the SYSCALL instruction outside the branch delay slot. On return from a syscall the program counter will be set to the instruction following the SYSCALL instruction.
Syscall restarting is a special case where $v0, $v1 and $a3 will stay unmodified. Even the program counter will stay unmodified so the same syscall will be executed again. This is something that does not matter to application programmers but may become visible in debuggers. Syscall restarting is something that is used internally by the kernel, for example when during a large read(2) syscall the kernel receives a signal.
For compatibility ABIs Linux/MIPS obviously follows whatever the native OS is doing. This happens to be similar to what Linux/MIPS is doing or from a historical perspective, Linux/MIPS is following what other, earlier MIPS UNIX implementations were doing.
Syscall number ranges
|System V Release 4 flavored syscalls||0||999|
|System V syscalls. All flavors of IRIX use this number range as well.||1000||1999|
|BSD 4.3 syscalls||2000||2999|
|Linux O32 syscalls||4000||4999|
|Linux N64 syscalls||5000||5999|
|Linux N32 syscalls||6000||6999|
For the exact syscall numbers for the three Linux ABI, please see <asm/unistd.h> of your kernel.