12.5 Streams and Threads

Streams can be used in multi-threaded applications in the same way they are used in single-threaded applications. But the programmer must be aware of a the possible complications. It is important to know about these also if the program one writes never use threads since the design and implementation of many stream functions is heavily influenced by the requirements added by multi-threaded programming.

The POSIX standard requires that by default the stream operations are atomic. I.e., issuing two stream operations for the same stream in two threads at the same time will cause the operations to be executed as if they were issued sequentially. The buffer operations performed while reading or writing are protected from other uses of the same stream. To do this each stream has an internal lock object which has to be (implicitly) acquired before any work can be done.

But there are situations where this is not enough and there are also situations where this is not wanted. The implicit locking is not enough if the program requires more than one stream function call to happen atomically. One example would be if an output line a program wants to generate is created by several function calls. The functions by themselves would ensure only atomicity of their own operation, but not atomicity over all the function calls. For this it is necessary to perform the stream locking in the application code.

The following example shows how the functions above can be used to generate an output line atomically even in multi-threaded applications (yes, the same job could be done with one fprintf call but it is sometimes not possible):

     FILE *fp;
     {
        ...
        flockfile (fp);
        fputs ("This is test number ", fp);
        fprintf (fp, "%d\n", test);
        funlockfile (fp)
     }

Without the explicit locking it would be possible for another thread to use the stream fp after the fputs call return and before fprintf was called with the result that the number does not follow the word `number'.

From this description it might already be clear that the locking objects in streams are no simple mutexes. Since locking the same stream twice in the same thread is allowed the locking objects must be equivalent to recursive mutexes. These mutexes keep track of the owner and the number of times the lock is acquired. The same number of funlockfile calls by the same threads is necessary to unlock the stream completely. For instance:

     void
     foo (FILE *fp)
     {
       ftrylockfile (fp);
       fputs ("in foo\n", fp);
       /* This is very wrong!!!  */
       funlockfile (fp);
     }

It is important here that the funlockfile function is only called if the ftrylockfile function succeeded in locking the stream. It is therefore always wrong to ignore the result of ftrylockfile. And it makes no sense since otherwise one would use flockfile. The result of code like that above is that either funlockfile tries to free a stream that hasn't been locked by the current thread or it frees the stream prematurely. The code should look like this:

     void
     foo (FILE *fp)
     {
       if (ftrylockfile (fp) == 0)
         {
           fputs ("in foo\n", fp);
           funlockfile (fp);
         }
     }

Now that we covered why it is necessary to have these locking it is necessary to talk about situations when locking is unwanted and what can be done. The locking operations (explicit or implicit) don't come for free. Even if a lock is not taken the cost is not zero. The operations which have to be performed require memory operations that are safe in multi-processor environments. With the many local caches involved in such systems this is quite costly. So it is best to avoid the locking completely if it is not needed – because the code in question is never used in a context where two or more threads may use a stream at a time. This can be determined most of the time for application code; for library code which can be used in many contexts one should default to be conservative and use locking.

There are two basic mechanisms to avoid locking. The first is to use the _unlocked variants of the stream operations. The POSIX standard defines quite a few of those and the GNU library adds a few more. These variants of the functions behave just like the functions with the name without the suffix except that they do not lock the stream. Using these functions is very desirable since they are potentially much faster. This is not only because the locking operation itself is avoided. More importantly, functions like putc and getc are very simple and traditionally (before the introduction of threads) were implemented as macros which are very fast if the buffer is not empty. With the addition of locking requirements these functions are no longer implemented as macros since they would would expand to too much code. But these macros are still available with the same functionality under the new names putc_unlocked and getc_unlocked. This possibly huge difference of speed also suggests the use of the _unlocked functions even if locking is required. The difference is that the locking then has to be performed in the program:

     void
     foo (FILE *fp, char *buf)
     {
       flockfile (fp);
       while (*buf != '/')
         putc_unlocked (*buf++, fp);
       funlockfile (fp);
     }

If in this example the putc function would be used and the explicit locking would be missing the putc function would have to acquire the lock in every call, potentially many times depending on when the loop terminates. Writing it the way illustrated above allows the putc_unlocked macro to be used which means no locking and direct manipulation of the buffer of the stream.

A second way to avoid locking is by using a non-standard function which was introduced in Solaris and is available in the GNU C library as well.

This function is especially useful when program code has to be used which is written without knowledge about the _unlocked functions (or if the programmer was too lazy to use them).