apr/docs/canonical_filenames.html

<HTML>
<HEAD><TITLE>APR Canonical Filenames</TITLE></HEAD>
<BODY>
<h1>APR Canonical Filename</h1>

<h2>Requirements</h2>

<p>APR porters need to address the underlying discrepancies between
file systems.  To achieve a reasonable degree of security, the
program depending upon APR needs to know that two paths may be
compared, and that a mismatch is guarenteed to reflect that the
two paths do not return the same resource</p>.

<p>The first discrepancy is in volume roots.  Unix and pure deriviates
have only one root path, "/".  Win32 and OS2 share root paths of
the form "D:/", D: is the volume designation.  However, this can
be specified as "//./D:/" as well, indicating D: volume of the
'this' machine.  Win32 and OS2 also may employ a UNC root path,
of the form "//server/share/" where share is a share-point of the
specified network server.  Finally, NetWare root paths are of the
form "server/volume:/", or the simpler "volume:/" syntax for 'this'
machine.  All these non-Unix file systems accept volume:path,
without a slash following the colon, as a path relative to the
current working directory, which APR will treat as ambigious, that
is, neither an absolute nor a relative path per se.</p>

<p>The second discrepancy is in the meaning of the 'this' directory.
In general, 'this' must be eliminated from the path where it occurs.
The syntax "path/./" and "path/" are both aliases to path.  However,
this isn't file system independent, since the double slash "//" has
a special meaning on OS2 and Win32 at the start of the path name,
and is invalid on those platforms before the "//server/share/" UNC
root path is completed.  Finally, as noted above, "//volume/" is
legal root syntax on WinNT, and perhaps others.</p>

<p>The third discrepancy is in the context of the 'parent' directory.
When "parent/path/.." occurs, the path must be unwound to "parent".
It's also critical to simply truncate leading "/../" paths to "/",
since the parent of the root is root.  This gets tricky on the
Win32 and OS2 platforms, since the ".." element is invalid before
the "//server/share/" is complete, and the "//server/share/../"
seqence is the complete UNC root "//server/share/".  In relative
paths, leading ".." elements are significant, until they are merged
with an absolute path.  The relative form must only retain the ".."
segments as leading segments, to be resolved once merged to another
relative or an absolute path.</p>

<p>The fourth discrepancy occurs with acceptance of alternate character
codes for the same element.  Path seperators are not retained within
the APR canonical forms.  The OS filesystem and APR (slashed) forms
can both be returned as strings, to be used in the proper context.
Unix, Win32 and Netware all accept slashes and backslashes as the
same path seperator symbol, although unix strictly accepts slashes.
While the APR form of the name strictly uses slashes, always consider
that there could be a platform that actually accepts slashes as a
character within a segment name.</p>

<p>The fifth and worst discrepancy plauges Win32, OS2, Netware, and some
filesystems mounted in Unix.  Case insensitivity can permit the same
file to slip through in both it's proper case and alternate cases.
Simply changing the case is insufficient for any character set beyond
ASCII, since various dilectic forms of characters suffer from one to
many or many to one translations.  An example would be u-umlaut, which
might be accepted as a single character u-umlaut, a two character
sequence u and the zero-width umlaut, the upper case form of the same,
or perhaps even a captial U alone.  This can be handled in different
ways depending on the purposes of the APR based program, but the one
requirement is that the path must be absolute in order to resolve these
ambiguities.  Methods employed include comparison of device and inode
file uniqifiers, which is a fairly fast operation, or quering the OS
for the true form of the name, which can be much slower.  Only the
acknowledgement of the file names by the OS can validate the equality
of two different cases of the same filename.</p>

<p>The sixth discrepancy, illegal or insignificant characters, is especially
significant in non-unix file systems.  Trailing periods are accepted
but never stored, therefore trailing periods must be ignored for any
form of comparison.  And all OS's have certain expectations of what
characters are illegal (or undesireable due to confusion.)</p>

<p>A final warning, canonical functions don't transform or resolve case
or character ambiguity issues until they are resolved into an absolute
path.  The relative canonical path, while useful, while useful for URL
or similar identifiers, cannot be used for testing or comparison of file
system objects.</p>

<hr>

<h2>Canonical API</h2>

Functions to manipulate the apr_canon_file_t (an opaque type) include:

<ul>
<li>Create canon_file_t (from char* path and canon_file_t parent path)
<li>Merged canon_file_t (from path and parent, both canon_file_t)
<li>Get char* path of all or some segments
<li>Get path flags of IsRelative, IsVirtualRoot, and IsAbsolute
<li>Compare two canon_file_t structures for file equality
</ul>

<p>The path is corrected to the file system case only if is in absolute
form.  The apr_canon_file_t should be preserved as long as possible and
used as the parent to create child entries to reduce the number of expensive
stat and case canonicalization calls to the OS.</p>

<p>The comparison operation provides that the APR can postpone correction
of case by simply relying upon the device and inode for equivalence.  The
stat implementation provides that two files are the same, while their
strings are not equivalent, and eliminates the need for the operating
system to return the proper form of the name.</p>

<p>In any case, returning the char* path, with a flag to request the proper
case, forces the OS calls to resolve the true names of each segment.  Where
there is a penality for this operation and the stat device and inode test
is faster, case correction is postponed until the char* result is requested.
On platforms that identify the inode, device, or proper name interchangably
with no penalities, this may occur when the name is initially processed.</p>

<hr>

<h2>Unix Example</h2>

<p>First the simplest case:</p>

<pre>
Parse Canonical Name
accepts parent path as canonical_t
        this path as string

Split this path Segments on '/'

For each of this path Segments
  If first Segment
    If this Segment is Empty ([nothing]/)
      Append this Root Segment (don't merge)
      Continue to next Segment
    Else is relative
      Append parent Segments (to merge)
      Continue with this Segment
  If Segment is '.' or empty (2 slashes)
    Discard this Segment
    Continue with next Segment
  If Segment is '..'
    If no previous Segment or previous Segment is '..'
      Append this Segment
      Continue with next Segment
    If previous Segment and previous is not Root Segment
      Discard previous Segment
    Discard this Segment
    Continue with next Segment
  Append this Relative Segment
  Continue with next Segment
</pre>

</BODY>
</HTML>