Книга: Distributed operating systems

2.4.2. Parameter Passing

2.4.2. Parameter Passing

The function of the client stub is to take its parameters, pack them into a message, and send it to the server stub. While this sounds straightforward, it is not quite as simple as it at first appears. In this section we will look at some of the issues concerned with parameter passing in RPC systems. Packing parameters into a message is called parameter marshaling.

As the simplest possible example, consider a remote procedure, sum(i, j), that takes two integer parameters and returns their arithmetic sum. (As a practical matter, one would not normally make such a simple procedure remote due to the overhead, but as an example it will do.) The call to sum, with parameters 4 and 7, is shown in the left-hand portion of the client process in Fig. 2-19. The client stub takes its two parameters and puts them in a message as indicated. It also puts the name or number of the procedure to be called in the message because the server might support several different calls, and it has to be told which one is required.

Fig. 2-19. Computing sum(4, 7) remotely.

When the message arrives at the server, the stub examines the message to see which procedure is needed, and then makes the appropriate call. If the server also supports the remote procedures difference, product, and quotient, the server stub might have a switch statement in it, to select the procedure to be called, depending on the first field of the message. The actual call from the stub to the server looks much like the original client call, except that the parameters are variables initialized from the incoming message, rather than constants.

When the server has finished, the server stub gains control again. It takes the result, provided by the server, and packs it into a message. This message is sent back to the client stub, which unpacks it and returns the value to the client procedure (not shown in the figure).

As long as the client and server machines are identical and all the parameters and results are scalar types, such as integers, characters, and Booleans, this model works fine. However, in a large distributed system, it is common that multiple machine types are present. Each machine often has its own representation for numbers, characters, and other data items. For example, IBM mainframes use the EBCDIC character code, whereas IBM personal computers use ASCII. As a consequence, it is not possible to pass a character parameter from an IBM PC client to an IBM mainframe server using the simple scheme of Fig. 2-19: the server will interpret the character incorrectly.

Similar problems can occur with the representation of integers (ls complement versus 2s complement), and especially with floating-point numbers. In addition, an even more annoying problem exists because some machines, such as the Intel 486, number their bytes from right to left, whereas others, such as the Sun SPARC, number them the other way. The Intel format is called little endian and the sparc format is called big endian, after the politicians in Gulliver's Travels who went to war over which end of an egg to break (Cohen, 1981). As an example, consider a server with two parameters, an integer and a four-character string. Each parameter requires one 32-bit word. Figure 2-20(a) shows what the parameter portion of a message built by a client stub on an Intel 486 might look like. The first word contains the integer parameter, 5 in this case, and the second contains the string "JILL".

Fig. 2-20. (a) The original message on the 486. (b) The message after receipt on the SPARC. (c) The message after being inverted. The little numbers in boxes indicate the address of each byte.

Since messages are transferred byte for byte (actually, bit for bit) over the network, the first byte sent is the first byte to arrive. In Fig. 2-20(b) we show what the message of Fig. 2-20(a) would look like if received by a SPARC, which numbers its bytes with byte 0 at the left (high-order byte) instead of at the right (low-order byte) as do all the Intel chips. When the server stub reads the parameters at addresses 0 and 4, respectively, it will find an integer equal to 83,886,080 (5?2²⁴) and a string "JILL".

One obvious, but unfortunately incorrect, approach is to invert the bytes of each word after they are received, leading to Fig. 2-20(c). Now the integer is 5 and the string is "LLIJ". The problem here is that integers are reversed by the different byte ordering, but strings are not. Without additional information about what is a string and what is an integer, there is no way to repair the damage.

Fortunately, this information is implicitly available. Remember that the items in the message correspond to the procedure identifier and parameters. Both the client and server know what the types of the parameters are. Thus a message corresponding to a remote procedure with n parameters will have n+1 fields, one identifying the procedure and one for each of the n parameters. Once a standard has been agreed upon for representing each of the basic data types, given a parameter list and a message, it is possible to deduce which bytes belong to which parameter, and thus to solve the problem.

As a simple example, consider the procedure of Fig. 2-21 (a). It has three parameters, a character, a floating-point number, and an array of five integers. We might decide to transmit a character in the rightmost byte of a word (leaving the next 3 bytes empty), a float as a whole word, and an array as a group of words equal to the array length, preceded by a word giving the length, as shown in Fig. 2-21(b). Thus given these rules, the client stub for foobar knows that it must use the format of Fig. 2-21(b), and the server stub knows that incoming messages for foobar will have the format of Fig. 2-21(b). Having the type information for the parameters makes it possible to make any necessary conversions.

Fig. 2-21. (a) A procedure. (b) The corresponding message.

Even with this additional information, there are still some issues open. In particular, how should information be represented in the messages? One way is to devise a network standard or canonical form for integers, characters, booleans, floating-point numbers, and so on, and require all senders to convert their internal representation to this form while marshaling. For example, suppose that it is decided to use two's complement for integers, ASCII for characters, 0 (false) and 1 (true) for Booleans, and IEEE format for floating-point numbers, with everything stored in little endian. For any list of integers, characters, Booleans, and floating-point numbers, the exact pattern required is now deterministic down to the last bit. As a result, the server stub no longer has to worry about which byte ordering the client has because the order of the bits in the message is now fixed, independent of the client's hardware.

The problem with this method is that it is sometimes inefficient. Suppose that a big endian client is talking to a big endian server. According to the rules, the client must convert everything to little endian in the message, and the server must convert it back again when it arrives. Although this is unambiguous, it requires two conversions when in fact none were necessary. This observation gives rise to a second approach: the client uses its own native format and indicates in the first byte of the message which format this is. Thus a little endian client builds a little endian message and a big endian client builds a big endian message. As soon as a message comes in, the server stub examines the first byte to see what the client is. If it is the same as the server, no conversion is needed. Otherwise, the server stub converts everything. Although we have only discussed converting from one endian to the other, conversions between one's and two's complement, EBCDIC to ASCII, and so on, can be handled in the same way. The trick is knowing what the message layout is and what the client is. Once these are known, the rest is easy (provided that everyone can convert from everyone else's format).

Now we come to the question of where the stub procedures come from. In many RPC-based systems, they are generated automatically. As we have seen, given a specification of the server procedure and the encoding rules, the message format is uniquely determined. Thus it is possible to have a compiler read the server specification and generate a client stub that packs its parameters into the officially approved message format. Similarly, the compiler can also produce a server stub that unpacks them and calls the server. Having both stub procedures generated from a single formal specification of the server not only makes life easier for the programmers, but reduces the chance of error and makes the system transparent with respect to differences in internal representation of data items.

Finally, we come to our last and most difficult problem: How are pointers passed? The answer is: only with the greatest of difficulty, if at all. Remember that a pointer is meaningful only within the address space of the process in which it is being used. Getting back to our read example discussed earlier, if the second parameter (the address of the buffer) happens to be 1000 on the client, one cannot just pass the number 1000 to the server and expect it to work. Address 1000 on the server might be in the middle of the program text.

One solution is just to forbid pointers and reference parameters in general. However, these are so important that this solution is highly undesirable. In fact, it is not necessary either. In the read example, the client stub knows that the second parameter points to an array of characters. Suppose, for the moment, that it also knows how big the array is. One strategy then becomes apparent: copy the array into the message and send it to the server. The server stub can then call the server with a pointer to this array, even though this pointer has a different numerical value than the second parameter of read has. Changes the server makes using the pointer (e.g., storing data into it) directly affect the message buffer inside the server stub. When the server finishes, the original message can be sent back to the client stub, which then copies it back to the client. In effect, call-by-reference has been replaced by copy/restore. Although this is not always identical, it frequently is good enough.

One optimization makes this mechanism twice as efficient. If the stubs know whether the buffer is an input parameter or an output parameter to the server, one of the copies can be eliminated. If the array is input to the server (e.g., in a call to write) it need not be copied back. If it is output, it need not be sent over in the first place. The way to tell them is in the formal specification of the server procedure. Thus associated with every remote procedure is a formal specification of the procedure, written in some kind of specification language, telling what the parameters are, which are input and which are output (or both), and what their (maximum) sizes are. It is from this formal specification that the stubs are generated by a special stub compiler.

As a final comment, it is worth noting that although we can now handle pointers to simple arrays and structures, we still cannot handle the most general case of a pointer to an arbitrary data structure such as a complex graph. Some systems attempt to deal with this case by actually passing the pointer to the server stub and generating special code in the server procedure for using pointers.

Normally, a pointer is followed (dereferenced) by putting it in a register and indirecting through the register. When this special technique is used, a pointer is dereferenced by sending a message back to the client stub asking it to fetch and send the item being pointed to (reads) or store a value at the address pointed to (writes). While this method works, it is often highly inefficient. Imagine having the file server store the bytes in the buffer by sending back each one in a separate message. Still, it is better than nothing, and some systems use it.

Оглавление книги