APRDesign

Design of APR

The Apache Portable Run-time libraries have been designed to provide a common
interface to low level routines across any platform.  The original goal of APR
was to combine all code in Apache to one common code base.  This is not the
correct approach however, so the goal of APR has changed.  

There are places where common code is not a good thing.  For example, how to
map requests to either threads or processes should be platform specific. 
APR's place is now to combine any code that can be safely combined without
sacrificing performance.

To this end we have created a set of operations that are required for cross
platfrom development.  There may be other types that are desired and those
will be implemented in the future.  The first version of APR will focus on
what Apache 2.0 needs.  Of course, anything that is submitted will be
considered for inclusion.

This document will discuss the structure of APR, and how best to contribute
code to the effort.

APR On Windows

APR on Windows is different from APR on all other systems, because it
doesn't use autoconf. On Unix, apr_private.h (private to APR) and apr.h 
(public, used by applications that use APR) are generated by autoconf 
from acconfig.h and apr.h.in respectively. On Windows, apr_private.h 
and apr.h are created from apr_private.hw and apr.hw respectively.

!!!***  If you add code to acconfig.h or tests to configure.in or aclocal.m4,
        please give some thought to whether or not Windows needs this addition
        as well.  A general rule of thumb, is that if it is a feature macro,
        such as APR_HAS_THREADS, Windows needs it.  If the definition is going
        to be used in a public APR header file, such as apr_general.h, Windows
        needs it.
        
        The only time it is safe to add a macro or test without also adding 
        the macro to apr*.hw, is if the macro tells APR how to build.  For
        example, a test for a header file does not need to be added to Windows.
***!!!

APR Features

One of the goals of APR is to provide a common set of features across all 
platforms.  This is an admirable goal, it is also not realisitic.  We cannot
expect to be able to implement ALL features on ALL platforms.  So we are
going to do the next best thing.  Provide a common interface to ALL APR 
features on MOST platforms.

APR developers should create FEATURE MACROS for any feature that is not
available on ALL platforms.  This should be a simple definition which has
the form:

APR_HAS_FEATURE

This macro should evaluate to true if APR has this feature on this platform.
For example, Linux and Windows have mmap'ed files, and APR is providing an
interface for mmapp'ing a file.  On both Linux and Windows, APR_HAS_MMAP
should evaluate to one, and the ap_mmap_* functions should map files into
memory and return the appropriate status codes.

If your OS of choice does not have mmap'ed files, APR_HAS_MMAP should evaluate 
to zero, and all ap_mmap_* functions should not be defined.  The second step 
is a precaution that will allow us to break at compile time if a programmer 
tries to use unsupported functions. 

APR types

The base types in APR
file_io     File I/O, including pipes
lib         A portable library originally used in Apache.  This contains
            memory management, tables, and arrays.
locks       Mutex and reader/writer locks
misc        Any APR type which doesn't have any other place to belong
network_io  Network I/O
shmem       Shared Memory (Not currently implemented)   
signal      Asynchronous Signals
threadproc  Threads and Processes
time        Time 

Directory Structure

Each type has a base directory.  Inside this base directory, are
subdirectories, which contain the actual code.  These subdirectories are named
after the platforms the are compiled on.  Unix is also used as a common
directory.  If the code you are writing is POSIX based, you should look at the
code in the unix directory.  A good rule of thumb, is that if more than half
your code needs to be ifdef'ed out, and the structures required for your code
are substantively different from the POSIX code, you should create a new
directory.

Currently, the APR code is written for Unix, BeOS, Windows, and OS/2.  An
example of the directory structure is the file I/O directory:

apr
  |
   ->  file_io
          |
           -> unix            The Unix and common base code
          |
           -> win32           The Windows code
          | 
           -> os2             The OS/2 code

Obviously, BeOS does not have a directory.  This is because BeOS is currently
using the Unix directory for it's file_io.  In the near future, it will be
possible to use indiviual files from the Unix directory.

There are a few special top level directories.  These are test, inc, include,
and libs.  Test is a directory which stores all test programs.  It is expected
that if a new type is developed, there will also be a new test program, to
help people port this new type to different platforms.  Inc is a directory for
internal header files.  This directory is likely to go away soon.  Include is
a directory which stores all required APR header files for external use.  The
distinction between internal and external header files will be made soon. 
Finally, libs is a generated directory.  When APR finishes building, it will
store it's library files in the libs directory.

Creating an APR Type

The current design of APR requires that APR types be incomplete.  It is not
possible to write flexible portable code if programs can access the internals
of APR types.  This is because different platforms are likely to define
different native types.

For this reason, each platform defines a structure in their own directories. 
Those structures are then typedef'ed in an external header file.  For example
in file_io/unix/fileio.h:

    struct ap_file_t {
        ap_context_t *cntxt;
        int filedes;
        FILE *filehand;
        ...
    }

In include/apr_file_io.h:
    typedef struct ap_file_t    ap_file_t;

This will cause a compiler error if somebody tries to access the filedes field
in this strcture.  Windows does not have a filedes field, so obviously, it is
important that programs not be able to access these.

The only exception to the incomplete type rule can be found in apr_portable.h.
This file defines the native types for each platform.  Using these types, it
is possible to extract native types for any APR type.

You may notice the ap_context_t field.  All APR types have this field.  This
type is used to allocate memory within APR.

New Function

When creating a new function, please try to adhere to these rules.

1)  Result arguments should be the first arguments.
2)  If a function needs a context, it should be the last argument.
3)  These rules are flexible, especially if it makes the code easier
    to understand because it mimics a standard function.

Documentation

Whenever a new function is added to APR, it MUST be documented.  New 
functions will not be committed unless there are docs to go along with them.
The documentation should be a comment block above the function in the header
file.

The format for the comment block is:

/**
 * Brief description of the function
 * @param parma_1_name explanation
 * @param parma_2_name explanation
 * @param parma_n_name explanation
 * @tip Any extra information people should know.
 * @deffunc function prototype if required
 */                                                                             

The last line is not strictly needed.  The parser in ScanDoc is not perfect
yet, and it can not parse prototypes that are in any form other than
	return_type program_name(type1 param1, type2 param2, ...)
This means that any function prototype that resembles:
	APR_EXPORT(ap_status_t) ap_foo(int f1, char *f2)
will need the deffunc.

For an actual example, look at any file in the include directory (ap_tables.h
hasn't been done yet).

APR Error reporting

Most APR functions should return an ap_status_t type.  The only time an
APR function does not return an ap_status_t is if it absolutly CAN NOT
fail.  Examples of this would be filling out an array when you know you are
not beyond the array's range.  If it cannot fail on your platform, but it
could conceivably fail on another platform, it should return an ap_status_t.
Unless you are sure, return an ap_status_t.  :-)

All platform return errno values unchanged.  Each platform can also have
one system error type, which can be returned after an offset is added.  
There are five types of error values in APR, each with it's own offset.

    Name			Purpose
0) 			This is 0 for all platforms and isn't really defined
 			anywhere, but it is the offset for errno values.
			(This has no name because it isn't actually defined, 
                        but completeness we are discussing it here).
1) APR_OS_START_ERROR	This is platform dependant, and is the offset at which
			APR errors start to be defined.  (Canonical error
			values are also defined in this section.  [Canonical
			error values are discussed later]).
2) APR_OS_START_STATUS	This is platform dependant, and is the offset at which
			APR status values start.
4) APR_OS_START_USEERR	This is platform dependant, and is the offset at which
			APR apps can begin to add their own error codes.
3) APR_OS_START_SYSERR	This is platform dependant, and is the offset at which
			system error values begin.

All of these definitions can be found in apr_errno.h for all platforms.  When
an error occurs in an APR function, the function must return an error code.
If the error occurred in a system call and that system call uses errno to
report an error, then the code is returned unchanged.  For example: 

    if (open(fname, oflags, 0777) < 0)
        return errno;


The next place an error can occur is a system call that uses some error value
other than the primary error value on a platform.  This can also be handled
by APR applications.  For example:

    if (CreateFile(fname, oflags, sharemod, NULL, 
                   createflags, attributes,0) == INVALID_HANDLE_VALUE
        return (GetLAstError() + APR_OS_START_SYSERR);

These two examples implement the same function for two different platforms.
Obviously even if the underlying problem is the same on both platforms, this
will result in two different error codes being returned.  This is OKAY, and
is correct for APR.  APR relies on the fact that most of the time an error
occurs, the program logs the error and continues, it does not try to
programatically solve the problem.  This does not mean we have not provided
support for programmatically solving the problem, it just isn't the default
case.  We'll get to how this problem is solved in a little while.

If the error occurs in an APR function but it is not due to a system call,
but it is actually an APR error or just a status code from APR, then the
appropriate code should be returned.  These codes are defined in apr_errno.h
and are self explanatory.

No APR code should ever return a code between APR_OS_START_USEERR and 
APR_OS_START_SYSERR, those codes are reserved for APR applications.

To programmatically correct an error in a running application, the error codes
need to be consistent across platforms.  This should make sense.  To get
consistent error codes, APR provides a function ap_canonical_error().
This function will take as input any ap_status_t value, and return a small
subset of canonical APR error codes.  These codes will be equivalent to
Unix errno's.  Why is it a small subset?  Because we don't want to try to
convert everything in the first pass.  As more programs require that more
error codes are converted, they will be added to this function.  

Why did APR take this approach?  There are two ways to deal with error 
codes portably.

1)  return the same error code across all platforms.  2)  return platform
specific error codes and convert them when necessary.  

The problem with option number one is that it takes time to convert error 
codes to a common code, and most of the time programs want to just output 
an error string.  If we convert all errors to a common subset, we have four 
steps to output an error string:

    make syscall that fails
        convert to common error code                 step 1
        return common error code
            check for success
            call error output function               step 2
                convert back to system error         step 3
                output error string                  step 4

By keeping the errors platform specific, we can output error strings in two
steps.

    make syscall that fails
        return error code
            check for success
            call error output function               step 1
                output error string                  step 2

Less often, programs change their execution based on what error was returned.
This is no more expensive using option 2 and it is using option 1, but we
put the onus of converting the error code on the programmer themselves.
For example, using option 1:

    make syscall that fails
        convert to common error code
        return common error code
            decide execution basd on common error code

Using option 2:
    
    make syscall that fails
        return error code
            convert to common error code (using ap_canonical_error)
            decide execution based on common error code

Finally, there is one more operation on error codes.  You can get a string
that explains in human readable form what has happened.  To do this using 
APR, call ap_strerror().

On all platforms ap_strerror takes the form:

char *ap_strerror(ap_status_t err)
{
    if (err < APR_OS_START_ERRNO2)
        return (platform dependant error string generator)
    if (err < APR_OS_START_ERROR)
        return (platform dependant error string generator for
                supplemental error values)
    if (err < APR_OS_SYSERR)
        return (APR generated error or status string)
    if (err == 0)
        return "No error was found"
    else
        return "APR doesn't understand this error value"
}

Notice, this does not handle canonicalized error values well.  Those will 
return "APR doesn't understand this error value" on some platforms and
an actual error string on others.  To deal with this, just get the
string before canonicalizing your error code.

The other problem with option 1, is that it is a lossy conversion.  For
example, Windows and OS/2 have a couple hundred error codes, but POSIX errno
only defines about 50 errno values.  This means that if we convert to a
canonical error value immediately, there is no way for the programmer to
get the actual system error.