Generated documentation

README.md

@@ -24,8 +24,23 @@
 &emsp;<a href="#toc4-187">zhash - expandable hash table container</a>
 &emsp;<a href="#toc4-198">zlist - singly-linked list container</a>
 &emsp;<a href="#toc4-209">zclock - millisecond clocks and delays</a>
-&emsp;<a href="#toc3-220">Development</a>
-&emsp;<a href="#toc3-230">This Document</a>
+
+**<a href="#toc2-220">Design Ideology</a>**
+&emsp;<a href="#toc3-223">The Problem with C</a>
+&emsp;<a href="#toc3-232">A Simple Class Model</a>
+&emsp;<a href="#toc3-259">Naming Style</a>
+&emsp;<a href="#toc3-268">Containers</a>
+&emsp;<a href="#toc3-280">Portability</a>
+&emsp;<a href="#toc3-306">Technical Aspects</a>
+
+**<a href="#toc2-316">Under the Hood</a>**
+&emsp;<a href="#toc3-319">Adding a New Class</a>
+&emsp;<a href="#toc3-332">Coding Style</a>
+&emsp;<a href="#toc3-351">Assertions</a>
+&emsp;<a href="#toc3-369">Documentation</a>
+&emsp;<a href="#toc3-408">Development</a>
+&emsp;<a href="#toc3-418">Porting libzapi</a>
+&emsp;<a href="#toc3-431">This Document</a>
 
 <A name="toc2-11" title="Overview" />
 ## Overview
@@ -605,7 +620,195 @@ tiny, but milliseconds are just right for slices of time we want to work
 with at the ØMQ scale. zclock doesn't give you objects to work with, we
 like the zapi class model but we're not insane. There, got it in again.
 
-<A name="toc3-220" title="Development" />
+<A name="toc2-220" title="Design Ideology" />
+## Design Ideology
+
+<A name="toc3-223" title="The Problem with C" />
+### The Problem with C
+
+C has the significant advantage of being a small language that, if we take a little care with formatting and naming, can be easily interchanged between developers. Every C developer will use much the same 90% of the language. Larger languages like C++ provide powerful abstractions like STL containers but at the cost of interchange. Every C++ developer will use a different 20% of the language.
+
+The huge problem with C is that any realistic application needs packages of functionality to bring the language up to the levels we expect for the 21st century. Much can be done by using external libraries but every additional library is a dependency that makes the resulting applications harder to build and port. While C itself is a highly portable language (and can be made more so by careful use of the preprocessor), most C libraries consider themselves part of the operating system, and as such do not attempt to be portable.
+
+The answer to this, as we learned from building enterprise-level C applications at iMatix from 1995-2005, is to create our own fully portable, high-quality libraries of pre-packaged functionality, in C. Doing this right solves both the requirements of richness of the language, and of portability of the final applications.
+
+<A name="toc3-232" title="A Simple Class Model" />
+### A Simple Class Model
+
+C has no standard API style. It is one thing to write a useful component, but something else to provide an API that is consistent and obvious across many components. We learned from building [OpenAMQ](http://www.openamq.org), a messaging client and server of 0.5M LoC, that a consistent model for extending C makes life for the application developer much easier.
+
+The general model is that of a class (the source package) that provides objects (in fact C structures). The application creates objects and then works with them. When done, the application destroys the object. In C, we tend to use the same name for the object as the class, when we can, and it looks like this (to take a fictitious libzapi class):
+
+    zregexp_t *regexp = zregexp_new (regexp_string);
+    if (!regexp)
+        printf ("E: invalid regular expression: %s\n", regexp_string);
+    else
+    if (zregexp_match (regexp, input_buffer))
+        printf ("I: successful match for %s\n", input buffer);
+    zregexp_destroy (&regexp);
+
+As far as the C program is concerned, the object is a reference to a structure (not a void pointer). We pass the object reference to all methods, since this is still C. We could do weird stuff like put method addresses into the structure so that we can emulate a C++ syntax but it's not worthwhile. The goal is not to emulate an OO system, it's simply to gain consistency. The constructor returns an object reference, or NULL if it fails. The destructor nullifies the class pointer, and is idempotent.
+
+What we aim at here is the simplest possible consistent syntax.
+
+No model is fully consistent, and classes can define their own rules if it helps make a better result. For example:
+
+* Some classes may not be opaque. For example, we have cases of generated serialization classes that encode and decode structures to/from binary buffers. It feels clumsy to have to use methods to access the properties of these classes.
+
+* While every class has a new method that is the formal constructor, some methods may also act as constructors. For example, a "dup" method might take one object and return a second object.
+
+* While every class has a destroy method that is the formal destructor, some methods may also act as destructors. For example, a method that sends an object may also destroy the object (so that ownership of any buffers can passed to background threads). Such methods take the same "pointer to a reference" argument as the destroy method.
+
+<A name="toc3-259" title="Naming Style" />
+### Naming Style
+
+libzapi aims for short, consistent names, following the theory that names we use most often should be shortest. Classes get one-word names, unless they are part of a family of classes in which case they may be two words, the first being the family name. Methods, similarly, get one-word names and we aim for consistency across classes (so a method that does something semantically similar in two classes will get the same name in both). So the canonical name for any method is:
+
+    zclassname_methodname
+
+And the reader can easily parse this without needing special syntax to separate the class name from the method name.
+
+<A name="toc3-268" title="Containers" />
+### Containers
+
+After a long experiment with containers, we've decided that we need exactly two containers:
+
+* A singly-linked list.
+* A hash table using text keys.
+
+These are zlist and zhash, respectively. Both store void pointers, with no attempt to manage the details of contained objects. You can use these containers to create lists of lists, hashes of lists, hashes of hashes, etc.
+
+We assume that at some point we'll need to switch to a doubly-linked list.
+
+<A name="toc3-280" title="Portability" />
+### Portability
+
+Creating a portable C application can be rewarding in terms of maintaining a single code base across many platforms, and keeping (expensive) system-specific knowledge separate from application developers. In most projects (like ØMQ core), there is no portability layer and application code does conditional compilation for all mixes of platforms. This leads to quite messy code.
+
+libzapi is a portability layer, similar to but thinner than libraries like the [Apache Portable Runtime](http://apr.apache.org) (APR).
+
+These are the places a C application is subject to arbitrary system differences:
+
+* Different compilers may offer slightly different variants of the C language, often lacking specific types or using neat non-portable names. Windows is a big culprit here. We solve this by 'patching' the language in zapi_prelude.h, e.g. defining int64_t on Windows.
+* System header files are inconsistent, i.e. you need to include different files depending on the OS type and version. We solve this by pulling in all necessary header files in zapi_prelude.h. This is a proven brute-force approach that increases recompilation times but eliminates a major source of pain.
+* System libraries are inconsistent, i.e. you need to link with different libraries depending on the OS type and version. We solve this with an external compilation tool, 'C', which detects the OS type and version (at runtime) and builds the necessary link commands.
+* System functions are inconsistent, i.e. you need to call different functions depending, again, on OS type and version. We solve this by building small abstract classes that handle specific areas of functionality, and doing conditional compilation in these.
+
+An example of the last:
+
+    #if (defined (__UNIX__))
+        pid = GetCurrentProcessId();
+    #elif (defined (__WINDOWS__))
+        pid = getpid ();
+    #else
+        pid = 0;
+    #endif
+
+libzapi uses the GNU autotools system, so non-portable code can use the macros this defines. It can also use macros defined by the zapi_prelude.h header file.
+
+<A name="toc3-306" title="Technical Aspects" />
+### Technical Aspects
+
+* *Thread safety*: the use of opaque structures is thread safe, though ØMQ applications should not share state between threads in any case.
+* *Name spaces*: we prefix class names with `z`, which ensures that all exported functions are globally safe.
+* *Library versioning*: we don't make any attempt to version the library at this stage. Classes are in our experience highly stable once they are built and tested, the only changes typically being added methods.
+* *Performance*: for critical path processing, you may want to avoid creating and destroying classes. However on modern Linux systems the heap allocator is very fast. Individual classes can choose whether or not to nullify their data on allocation.
+* *Self-testing*: every class has a `selftest` method that runs through the methods of the class. In theory, calling all selftest functions of all classes does a full unit test of the library. The `zapi_selftest` application does this.
+* *Memory management*: libzapi classes do not use any special memory management techiques to detect leaks. We've done this in the past but it makes the code relatively complex. Instead, we do memory leak testing using tools like valgrind.
+
+<A name="toc2-316" title="Under the Hood" />
+## Under the Hood
+
+<A name="toc3-319" title="Adding a New Class" />
+### Adding a New Class
+
+If you define a new libzapi class `myclass` you need to:
+
+* Write the `zmyclass.c` and `zmyclass.h` source files, in `src` and `include` respectively.
+* Add`#include <zmyclass.h>` to `include/zapi.h`.
+* Add the myclass header and test call to `src/zapi_selftest.c`.
+* Add a reference documentation to 'doc/zmyclass.txt'.
+* Add myclass to 'src/Makefile.am` and `doc/Makefile.am`.
+
+The `bin/newclass.sh` shell script will automate these steps for you.
+
+<A name="toc3-332" title="Coding Style" />
+### Coding Style
+
+In general the zctx class defines the style for the whole library. The overriding rules for coding style are consistency, clarity, and ease of maintenance. We use the C99 standard for syntax including principally:
+
+* The // comment style.
+* Variables definitions placed in or before the code that uses them.
+
+So while ANSI C code might say:
+
+    zblob_t *file_buffer;       /*  Buffer for our file */
+    ... (100 lines of code)
+    file_buffer = zblob_new ();
+    ...
+
+The style in libzapi would be:
+
+    zblob_t *file_buffer = zblob_new ();
+
+<A name="toc3-351" title="Assertions" />
+### Assertions
+
+We use assertions heavily to catch bad argument values. The libzapi classes do not attempt to validate arguments and report errors; bad arguments are treated as fatal application programming errors.
+
+We also use assertions heavily on calls to system functions that are never supposed to fail, where failure is to be treated as a fatal non-recoverable error (e.g. running out of memory).
+
+Assertion code should always take this form:
+
+    int rc = some_function (arguments);
+    assert (rc == 0);
+
+Rather than the side-effect form:
+
+    assert (some_function (arguments) == 0);
+
+Since assertions may be removed by an optimizing compiler.
+
+<A name="toc3-369" title="Documentation" />
+### Documentation
+
+Man pages are generated from the class header and source files via the doc/mkman tool, and similar functionality in the gitdown tool (http://github.com/imatix/gitdown). The header file for a class must wrap its interface as follows (example is from include/zclock.h):
+
+    //  @interface
+    //  Sleep for a number of milliseconds
+    void 
+        zclock_sleep (int msecs);
+
+    //  Return current system clock as milliseconds
+    int64_t 
+        zclock_time (void);
+
+    //  Self test of this class
+    int 
+        zclock_test (Bool verbose);
+    //  @end
+
+The source file for a class must provide documentation as follows:
+
+    /*  
+    @header
+    ...short explanation of class...
+    @discuss
+    ...longer discussion of how it works...
+    @end
+    */
+
+The source file for a class then provides the self test example as follows:
+
+    //  @selftest
+    int64_t start = zclock_time ();
+    zclock_sleep (10);
+    assert ((zclock_time () - start) >= 10);
+    //  @end
+
+The template for man pages is in doc/mkman.
+
+<A name="toc3-408" title="Development" />
 ### Development
 
 libzapi is developed through a test-driven process that guarantees no memory violations or leaks in the code:
@@ -615,7 +818,20 @@ libzapi is developed through a test-driven process that guarantees no memory vio
 * Run the 'selftest' script, which uses the Valgrind memcheck tool.
 * Repeat until perfect.
 
-<A name="toc3-230" title="This Document" />
+<A name="toc3-418" title="Porting libzapi" />
+### Porting libzapi
+
+When you try libzapi on an OS that it's not been used on (ever, or for a while), you will hit code that does not compile. In some cases the patches are trivial, in other cases (usually when porting to Windows), the work needed to build equivalent functionality may be quite heavy. In any case, the benefit is that once ported, the functionality is available to all applications.
+
+Before attempting to patch code for portability, please read the `zapi_prelude.h` header file. There are several typical types of changes you may need to make to get functionality working on a specific operating system:
+
+* Defining typedefs which are missing on that specific compiler: do this in zapi_prelude.h.
+* Defining macros that rename exotic library functions to more conventional names: do this in zapi_prelude.h.
+* Reimplementing specific methods to use a non-standard API: this is typically needed on Windows. Do this in the relevant class, using #ifdefs to properly differentiate code for different platforms.
+
+The canonical 'standard operating system' for all libzapi code is Linux, gcc, POSIX.
+
+<A name="toc3-431" title="This Document" />
 ### This Document
 
 This document is originally at README.txt and is built using [gitdown](http://github.com/imatix/gitdown).