Date: Tue, 05 Nov 1996 00:32:46 GMT Server: NCSA/1.5 Content-type: text/html Last-modified: Wed, 30 Aug 1995 21:21:35 GMT Content-length: 10762
EXCEPTION HANDLERS ------------------ The trouble with programmed I/O is that it both wastes CPU resources and it has potential for "incorrect" operation. What we really want: (Since most I/O devices are slow), have I/O devices signal the CPU when they have a change in status. The I/O devices tell the CPU that they are "ready." In order to do this we need: Hardware (wires) from devices to the CPU. A way for special software to be invoked when the a device signals on the wire. The modern solution bundles the software to deal with these signals (interrupts) and other situations into an EXCEPTION HANDLER. (Effectively part of the OS.) EXCEPTIONS ---------- 1. interrupts --initiated outside the instruction stream --arrive asynchronously (at no specific time) examples: I/O device status change I/O device error condition thermal override shutdown internal error detection when should the interrupt be dealt with? as soon as possible 2. traps --occur due to something in instruction stream --arrive synchronously (while instruction is executing) good test: if program was re-run, the trap would occur in precisely the same place in the code. examples: unaligned address error arithmetic overflow syscall when should the trap be dealt with? right now! The user program cannot continue until whatever caused the trap is dealt with. exception handling ------------------ the mechanism for dealing with exceptions is simple; its implementation can get complex. The implementation varies among computers (manufactures). situation: a user program is running (executing), and a device generates an interrupt request. mechanism to respond: the hardware temporarily "suspends" the user program, and instead runs code called an EXCEPTION HANDLER. After the handler is finished doing whatever it needs to, the hardware returns control to the user program. limitations of exception handler: since it is being invoked (potentially) in the middle of a user program, the handler must take extra care not to change the state of the user program. -- it can't change register values -- it can't change the stack So, how can it do anything at all? The key to this answer is that any portion of the state that it does change, it must also restore before returning to the user program. The handler often uses the system stack to temporarily store register values. WHEN to handle an interrupt -- 2 possiblilities: 1. right now! Note that this could be in the middle of an instruction. In order to do this, the hardware must be able to know where the instruction is in its execution and be able to "take up where it left off" This is very difficult to do. But, it has been done in simpler forms on a few machines. Example: arbitrary memory to memory copy 2. wait until the currently executing instruction finishes, then handle. THIS IS THE METHOD OF CHOICE. The instruction fetch/execute cycle must be expanded to 1. handle pending interrupts 2. instruction fetch 3. PC update 4. decode 5. operand load 6. operation 7. store results some terms ---------- interrupt request -- the activation of hardware somewhere that signals the initial request for an interrupt. pending interrupt -- an interrupt that hasn't been handled yet, but needs to be kernel-- the exception handler In most minds, when people think of a kernel, they think of critical portions of an operating system. The exception handler IS a critical portion of an operating system! handler -- the code of the exception handler. MIPS exception handling mechanism --------------------------------- hardware does the following: 1. sets state giving cause of exception Cause register -- coprocessor C0, register 13 32 bit register, bits 6..2 (5 bits) specify the type of the exception, called the ExcCode. Page 316 (text) gives the mapping of encodings to causes examples: 00000 (0) Interrupt 00100 (4) load from an illegal address 01000 (8) syscall instruction 01100 (12) arithmetic overflow 2. changes to kernel mode, saving the previous mode in a hardware stack (3 levels deep) The mode is saved in the Status register. coprocessor C0, register 12 defined in the processor's architecture are 2 modes, user -- the mode that user programs run under. certain instructions are not available kernel -- the operating system mode. Allows the OS to retain control over "vital" system aspects. All instructions are available. 3. disables further interrupts 4. saves current PC coprocessor C0, register 14, called Exception Program counter. Gives return address in user program. Where to return to when done handling the exception. 5. jumps to hardwired address 0x8000 0080. This is where the exception handler code is. Then, the code within the exception handler is run. It does the following: 1. save some registers (on system stack). The handler needs to use registers too! It may not change (clobber) register contents of the user program. So, it saves them (on stack or in memory). 2. Figure out exception type. (in ExcCode) mfc0 $k0, $13 # get Cause register andi $k0, $k0, 0x3c # Mask out all but ExcCode 3. use ExcCode in combination with a JUMP TABLE to jump to the correct location within the exception handler. 4. handle the exception (whatever it is!) 5. restore registers saved in (1). 6. atomically: (as if done in 1 step, not 3) restore previous kernel/user mode (from Status register) reenable interrupts jump back to user program (using EPC) about Jump Tables ----------------- A clever mechanism for doing something like a CASE (SWITCH) statement. A jump to one of many locations. keep a table of addresses (case1, case2, and case3): JumpTable: .word case0 .word case1 .word case2 sll $8, $8, 2 # case number shifted left 2 bits # (need a word offset into table, not byte) lw $9, JumpTable($8) # load address into $9 jr $9 # jump to address contained in $9 . . . case0: #code for case0 here . . . case1: #code for case1 here . . . case2: #code for case2 here (Note that the cases don't have to go in any specific order.) Addressing mode: label($rb) Effective address is gotten by label + ($rb) label doesn't fit into 16 bit displacement field of load/store instruction. MAL->TAL synthesis of this must be something like: la $1, label add $1, $1, $rb then use 0($1) as addressing mode in load/store instruction. some advanced topics -------------------- PRIORITIES problem: Multiple interrupt requests can arrive simultaneously. Which one should get handled first? possible solutions: FCFS -- the first one to arrive gets handled first. difficulty 1) This might allow a malicious/recalcitrant device or program to gain control of the processor. difficulty 2) There must be hardware that maintains an ordering of pending exceptions. prioritize all exceptions -- the one with the highest priority gets handled first. This is a common method for solving the problem. Priorities for various exceptions are assigned either by the manufacturer, or by a system manager through software. The priorities are normally set when a machine is booted (the OS is started up). difficulty 1) Exceptions with the same priority must still be handled in some order. Example of same priority exceptions might be all keyboard interrupts. Consider a machine with many terminals hooked up. The instruction fetch/execute cycle becomes: 1. any interrupts with a higher priority than whatever is currently running pending? 2. fetch 3. decode 4. operands 5. operation 6. result NOTE: This implies that there is some hardware notion of the priority for whatever is running (user program, keyboard interrupts, clock interrupt, etc.) What should get given the highest priority? clock? power failure? thermal shutdown? arithmetic overflow? keyboard? I/O device ready? priorities are a matter of which is most urgent, and therefore cannot wait, and how long it takes to process the interrupt. -- clock is urgent, and takes little processing, maybe only a variable increment. -- power failure is very urgent, but takes a lot or processing, because the machine will be stopped. -- overflow is urgent to the program which caused it, because it cannot continue. -- keyboard is urgent because we don't want to lose a second key press before the first is handled. (story of what can happen if interrupt priorities set wrong) REENTRANT EXCEPTION HANDLERS The best solution combines priorities with an exception handler that can itself be interrupted. There are many details to get right to make this possible. The instruction fetch/execute cycle remains the same. At the beginning of EVERY instruction (even those within the exception handler), a check is made if there are pending interrupts. Only those with higher priorities than whatever is currently running will be processed. The exception handler must be modified so that it can be interrupted. Its own state must be saved (safely). Within the handler: 1. disable interrupts momentarily in order to save important state that cannot get clobbered. (EPC, current priority level, maybe registers $26 and $27). Question: where do these things get saved? 2. re-enable interrupts for higher priorities than current level. If the priority level checking is done in hardware, then all interrupts can be re-enabled. 3. This invocation of the exception handler eventually finishes. The instruction fetch/execute cycle must be expanded to 1. handle pending interrupts 2. instruction fetch 3. PC update 4. decode 5. operand load 6. operation 7. store results