Annotation of embedaddon/bird/doc/prog-4.html, revision 1.1.1.1

1.1       misho       1: <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 3.2 Final//EN">
                      2: <HTML>
                      3: <HEAD>
                      4:  <META NAME="GENERATOR" CONTENT="LinuxDoc-Tools 1.0.9">
                      5:  <TITLE>BIRD Programmer's Documentation: Filters</TITLE>
                      6:  <LINK HREF="prog-5.html" REL=next>
                      7:  <LINK HREF="prog-3.html" REL=previous>
                      8:  <LINK HREF="prog.html#toc4" REL=contents>
                      9: </HEAD>
                     10: <BODY>
                     11: <A HREF="prog-5.html">Next</A>
                     12: <A HREF="prog-3.html">Previous</A>
                     13: <A HREF="prog.html#toc4">Contents</A>
                     14: <HR>
                     15: <H2><A NAME="s4">4.</A> <A HREF="prog.html#toc4">Filters</A></H2>
                     16: 
                     17: <H2><A NAME="ss4.1">4.1</A> <A HREF="prog.html#toc4.1">Filters</A>
                     18: </H2>
                     19: 
                     20: <P>
                     21: <P>You can find sources of the filter language in <CODE>filter/</CODE>
                     22: directory. File <CODE>filter/config.Y</CODE> contains filter grammar and basically translates
                     23: the source from user into a tree of <I>f_inst</I> structures. These trees are
                     24: later interpreted using code in <CODE>filter/filter.c</CODE>.
                     25: <P>A filter is represented by a tree of <I>f_inst</I> structures, one structure per
                     26: "instruction". Each <I>f_inst</I> contains <B>code</B>, <B>aux</B> value which is
                     27: usually the data type this instruction operates on and two generic
                     28: arguments (<B>a1</B>, <B>a2</B>). Some instructions contain pointer(s) to other
                     29: instructions in their (<B>a1</B>, <B>a2</B>) fields.
                     30: <P>Filters use a <I>f_val</I> structure for their data. Each <I>f_val</I>
                     31: contains type and value (types are constants prefixed with <I>T_</I>). Few
                     32: of the types are special; <I>T_RETURN</I> can be or-ed with a type to indicate
                     33: that return from a function or from the whole filter should be
                     34: forced. Important thing about <I>f_val</I>'s is that they may be copied
                     35: with a simple <CODE>=</CODE>. That's fine for all currently defined types: strings
                     36: are read-only (and therefore okay), paths are copied for each
                     37: operation (okay too).
                     38: <P>
                     39: <P><HR><H3>Function</H3>
                     40: <P><I>int</I>
                     41: <B>val_compare</B>
                     42: (<I>struct f_val</I> <B>v1</B>, <I>struct f_val</I> <B>v2</B>) --     compare two values
                     43: <P>
                     44: <H3>Arguments</H3>
                     45: <P>
                     46: <DL>
                     47: <DT><I>struct f_val</I> <B>v1</B><DD><P>first value
                     48: <DT><I>struct f_val</I> <B>v2</B><DD><P>second value
                     49: </DL>
                     50: <H3>Description</H3>
                     51: <P>Compares two values and returns -1, 0, 1 on &lt;, =, &gt; or CMP_ERROR on
                     52: error. Tree module relies on this giving consistent results so
                     53: that it can be used for building balanced trees.
                     54: 
                     55: 
                     56: <HR><H3>Function</H3>
                     57: <P><I>int</I>
                     58: <B>val_same</B>
                     59: (<I>struct f_val</I> <B>v1</B>, <I>struct f_val</I> <B>v2</B>) --     compare two values
                     60: <P>
                     61: <H3>Arguments</H3>
                     62: <P>
                     63: <DL>
                     64: <DT><I>struct f_val</I> <B>v1</B><DD><P>first value
                     65: <DT><I>struct f_val</I> <B>v2</B><DD><P>second value
                     66: </DL>
                     67: <H3>Description</H3>
                     68: <P>Compares two values and returns 1 if they are same and 0 if not.
                     69: Comparison of values of different types is valid and returns 0.
                     70: 
                     71: 
                     72: <HR><H3>Function</H3>
                     73: <P><I>int</I>
                     74: <B>val_in_range</B>
                     75: (<I>struct f_val</I> <B>v1</B>, <I>struct f_val</I> <B>v2</B>) --     implement <CODE>~</CODE> operator
                     76: <P>
                     77: <H3>Arguments</H3>
                     78: <P>
                     79: <DL>
                     80: <DT><I>struct f_val</I> <B>v1</B><DD><P>element
                     81: <DT><I>struct f_val</I> <B>v2</B><DD><P>set
                     82: </DL>
                     83: <H3>Description</H3>
                     84: <P>Checks if <B>v1</B> is element (<CODE>~</CODE> operator) of <B>v2</B>.
                     85: 
                     86: 
                     87: <HR><H3>Function</H3>
                     88: <P><I>struct f_val</I>
                     89: <B>interpret</B>
                     90: (<I>struct f_inst *</I> <B>what</B>)
                     91: <H3>Arguments</H3>
                     92: <P>
                     93: <DL>
                     94: <DT><I>struct f_inst *</I> <B>what</B><DD><P>filter to interpret
                     95: </DL>
                     96: <H3>Description</H3>
                     97: <P>Interpret given tree of filter instructions. This is core function
                     98: of filter system and does all the hard work.
                     99: <H3>Each instruction has 4 fields</H3>
                    100: <P>code (which is instruction code),
                    101: aux (which is extension to instruction code, typically type),
                    102: arg1 and arg2 - arguments. Depending on instruction, arguments
                    103: are either integers, or pointers to instruction trees. Common
                    104: instructions like +, that have two expressions as arguments use
                    105: TWOARGS macro to get both of them evaluated.
                    106: <P><I>f_val</I> structures are copied around, so there are no problems with
                    107: memory managment.
                    108: 
                    109: 
                    110: <HR><H3>Function</H3>
                    111: <P><I>int</I>
                    112: <B>f_run</B>
                    113: (<I>struct filter *</I> <B>filter</B>, <I>struct rte **</I> <B>rte</B>, <I>struct ea_list **</I> <B>tmp_attrs</B>, <I>struct linpool *</I> <B>tmp_pool</B>, <I>int</I> <B>flags</B>) --     run a filter for a route
                    114: <P>
                    115: <H3>Arguments</H3>
                    116: <P>
                    117: <DL>
                    118: <DT><I>struct filter *</I> <B>filter</B><DD><P>filter to run
                    119: <DT><I>struct rte **</I> <B>rte</B><DD><P>route being filtered, may be modified
                    120: <DT><I>struct ea_list **</I> <B>tmp_attrs</B><DD><P>temporary attributes, prepared by caller or generated by <B>f_run()</B>
                    121: <DT><I>struct linpool *</I> <B>tmp_pool</B><DD><P>all filter allocations go from this pool
                    122: <DT><I>int</I> <B>flags</B><DD><P>flags
                    123: </DL>
                    124: <H3>Description</H3>
                    125: <P>If filter needs to modify the route, there are several
                    126: posibilities. <B>rte</B> might be read-only (with REF_COW flag), in that
                    127: case rw copy is obtained by <B>rte_cow()</B> and <B>rte</B> is replaced. If
                    128: <B>rte</B> is originally rw, it may be directly modified (and it is never
                    129: copied).
                    130: <P>The returned rte may reuse the (possibly cached, cloned) rta, or
                    131: (if rta was modificied) contains a modified uncached rta, which
                    132: uses parts allocated from <B>tmp_pool</B> and parts shared from original
                    133: rta. There is one exception - if <B>rte</B> is rw but contains a cached
                    134: rta and that is modified, rta in returned rte is also cached.
                    135: <P>Ownership of cached rtas is consistent with rte, i.e.
                    136: if a new rte is returned, it has its own clone of cached rta
                    137: (and cached rta of read-only source rte is intact), if rte is
                    138: modified in place, old cached rta is possibly freed.
                    139: 
                    140: 
                    141: <HR><H3>Function</H3>
                    142: <P><I>int</I>
                    143: <B>filter_same</B>
                    144: (<I>struct filter *</I> <B>new</B>, <I>struct filter *</I> <B>old</B>) --     compare two filters
                    145: <P>
                    146: <H3>Arguments</H3>
                    147: <P>
                    148: <DL>
                    149: <DT><I>struct filter *</I> <B>new</B><DD><P>first filter to be compared
                    150: <DT><I>struct filter *</I> <B>old</B><DD><P>second filter to be compared, notice that this filter is
                    151: damaged while comparing.
                    152: </DL>
                    153: <H3>Description</H3>
                    154: <P>Returns 1 in case filters are same, otherwise 0. If there are
                    155: underlying bugs, it will rather say 0 on same filters than say
                    156: 1 on different.
                    157: 
                    158: 
                    159: <HR><H3>Function</H3>
                    160: <P><I>struct f_tree *</I>
                    161: <B>find_tree</B>
                    162: (<I>struct f_tree *</I> <B>t</B>, <I>struct f_val</I> <B>val</B>)
                    163: <H3>Arguments</H3>
                    164: <P>
                    165: <DL>
                    166: <DT><I>struct f_tree *</I> <B>t</B><DD><P>tree to search in
                    167: <DT><I>struct f_val</I> <B>val</B><DD><P>value to find
                    168: </DL>
                    169: <H3>Description</H3>
                    170: <P>Search for given value in the tree. I relies on fact that sorted tree is populated
                    171: by <I>f_val</I> structures (that can be compared by <B>val_compare()</B>). In each node of tree, 
                    172: either single value (then t-&gt;from==t-&gt;to) or range is present.
                    173: <P>Both set matching and <CODE><B>switch()</B> { }</CODE> construction is implemented using this function,
                    174: thus both are as fast as they can be.
                    175: 
                    176: 
                    177: <HR><H3>Function</H3>
                    178: <P><I>struct f_tree *</I>
                    179: <B>build_tree</B>
                    180: (<I>struct f_tree *</I> <B>from</B>)
                    181: <H3>Arguments</H3>
                    182: <P>
                    183: <DL>
                    184: <DT><I>struct f_tree *</I> <B>from</B><DD><P>degenerated tree (linked by <B>tree</B>-&gt;left) to be transformed into form suitable for <B>find_tree()</B>
                    185: </DL>
                    186: <H3>Description</H3>
                    187: <P>Transforms degenerated tree into balanced tree.
                    188: 
                    189: 
                    190: <HR><H3>Function</H3>
                    191: <P><I>int</I>
                    192: <B>same_tree</B>
                    193: (<I>struct f_tree *</I> <B>t1</B>, <I>struct f_tree *</I> <B>t2</B>)
                    194: <H3>Arguments</H3>
                    195: <P>
                    196: <DL>
                    197: <DT><I>struct f_tree *</I> <B>t1</B><DD><P>first tree to be compared
                    198: <DT><I>struct f_tree *</I> <B>t2</B><DD><P>second one
                    199: </DL>
                    200: <H3>Description</H3>
                    201: <P>Compares two trees and returns 1 if they are same
                    202: 
                    203: <H2><A NAME="ss4.2">4.2</A> <A HREF="prog.html#toc4.2">Trie for prefix sets</A>
                    204: </H2>
                    205: 
                    206: <P>
                    207: <P>We use a (compressed) trie to represent prefix sets. Every node
                    208: in the trie represents one prefix (<I>addr</I>/<I>plen</I>) and <I>plen</I> also
                    209: indicates the index of the bit in the address that is used to
                    210: branch at the node. If we need to represent just a set of
                    211: prefixes, it would be simple, but we have to represent a
                    212: set of prefix patterns. Each prefix pattern consists of
                    213: <I>ppaddr</I>/<I>pplen</I> and two integers: <I>low</I> and <I>high</I>, and a prefix
                    214: <I>paddr</I>/<I>plen</I> matches that pattern if the first MIN(<I>plen</I>, <I>pplen</I>)
                    215: bits of <I>paddr</I> and <I>ppaddr</I> are the same and <I>low</I> &lt;= <I>plen</I> &lt;= <I>high</I>.
                    216: <P>We use a bitmask (<I>accept</I>) to represent accepted prefix lengths
                    217: at a node. As there are 33 prefix lengths (0..32 for IPv4), but
                    218: there is just one prefix of zero length in the whole trie so we
                    219: have <I>zero</I> flag in <I>f_trie</I> (indicating whether the trie accepts
                    220: prefix 0.0.0.0/0) as a special case, and <I>accept</I> bitmask
                    221: represents accepted prefix lengths from 1 to 32.
                    222: <P>There are two cases in prefix matching - a match when the length
                    223: of the prefix is smaller that the length of the prefix pattern,
                    224: (<I>plen</I> &lt; <I>pplen</I>) and otherwise. The second case is simple - we
                    225: just walk through the trie and look at every visited node
                    226: whether that prefix accepts our prefix length (<I>plen</I>). The
                    227: first case is tricky - we don't want to examine every descendant
                    228: of a final node, so (when we create the trie) we have to propagate
                    229: that information from nodes to their ascendants.
                    230: <P>Suppose that we have two masks (M1 and M2) for a node. Mask M1
                    231: represents accepted prefix lengths by just the node and mask M2
                    232: represents accepted prefix lengths by the node or any of its
                    233: descendants. Therefore M2 is a bitwise or of M1 and children's
                    234: M2 and this is a maintained invariant during trie building.
                    235: Basically, when we want to match a prefix, we walk through the trie,
                    236: check mask M1 for our prefix length and when we came to
                    237: final node, we check mask M2.
                    238: <P>There are two differences in the real implementation. First,
                    239: we use a compressed trie so there is a case that we skip our
                    240: final node (if it is not in the trie) and we came to node that
                    241: is either extension of our prefix, or completely out of path
                    242: In the first case, we also have to check M2.
                    243: <P>Second, we really need not to maintain two separate bitmasks.
                    244: Checks for mask M1 are always larger than <I>applen</I> and we need
                    245: just the first <I>pplen</I> bits of mask M2 (if trie compression
                    246: hadn't been used it would suffice to know just $applen-th bit),
                    247: so we have to store them together in <I>accept</I> mask - the first
                    248: <I>pplen</I> bits of mask M2 and then mask M1.
                    249: <P>There are four cases when we walk through a trie:
                    250: <P>- we are in NULL
                    251: - we are out of path (prefixes are inconsistent)
                    252: - we are in the wanted (final) node (node length == <I>plen</I>)
                    253: - we are beyond the end of path (node length &gt; <I>plen</I>)
                    254: - we are still on path and keep walking (node length &lt; <I>plen</I>)
                    255: <P>The walking code in <B>trie_match_prefix()</B> is structured according to
                    256: these cases.
                    257: <P>
                    258: <P><HR><H3>Function</H3>
                    259: <P><I>struct f_trie *</I>
                    260: <B>f_new_trie</B>
                    261: (<I>linpool *</I> <B>lp</B>, <I>uint</I> <B>node_size</B>) --     allocates and returns a new empty trie
                    262: <P>
                    263: <H3>Arguments</H3>
                    264: <P>
                    265: <DL>
                    266: <DT><I>linpool *</I> <B>lp</B><DD><P>linear pool to allocate items from
                    267: <DT><I>uint</I> <B>node_size</B><DD><P>node size to be used (<I>f_trie_node</I> and user data)
                    268: </DL>
                    269: 
                    270: 
                    271: <HR><H3>Function</H3>
                    272: <P><I>void *</I>
                    273: <B>trie_add_prefix</B>
                    274: (<I>struct f_trie *</I> <B>t</B>, <I>ip_addr</I> <B>px</B>, <I>int</I> <B>plen</B>, <I>int</I> <B>l</B>, <I>int</I> <B>h</B>)
                    275: <H3>Arguments</H3>
                    276: <P>
                    277: <DL>
                    278: <DT><I>struct f_trie *</I> <B>t</B><DD><P>trie to add to
                    279: <DT><I>ip_addr</I> <B>px</B><DD><P>prefix address
                    280: <DT><I>int</I> <B>plen</B><DD><P>prefix length
                    281: <DT><I>int</I> <B>l</B><DD><P>prefix lower bound
                    282: <DT><I>int</I> <B>h</B><DD><P>prefix upper bound
                    283: </DL>
                    284: <H3>Description</H3>
                    285: <P>Adds prefix (prefix pattern) <B>px</B>/<B>plen</B> to trie <B>t</B>.  <B>l</B> and <B>h</B> are lower
                    286: and upper bounds on accepted prefix lengths, both inclusive.
                    287: 0 &lt;= l, h &lt;= 32 (128 for IPv6).
                    288: <P>Returns a pointer to the allocated node. The function can return a pointer to
                    289: an existing node if <B>px</B> and <B>plen</B> are the same. If px/plen == 0/0 (or ::/0),
                    290: a pointer to the root node is returned.
                    291: 
                    292: 
                    293: <HR><H3>Function</H3>
                    294: <P><I>int</I>
                    295: <B>trie_match_prefix</B>
                    296: (<I>struct f_trie *</I> <B>t</B>, <I>ip_addr</I> <B>px</B>, <I>int</I> <B>plen</B>)
                    297: <H3>Arguments</H3>
                    298: <P>
                    299: <DL>
                    300: <DT><I>struct f_trie *</I> <B>t</B><DD><P>trie
                    301: <DT><I>ip_addr</I> <B>px</B><DD><P>prefix address
                    302: <DT><I>int</I> <B>plen</B><DD><P>prefix length
                    303: </DL>
                    304: <H3>Description</H3>
                    305: <P>Tries to find a matching prefix pattern in the trie such that
                    306: prefix <B>px</B>/<B>plen</B> matches that prefix pattern. Returns 1 if there
                    307: is such prefix pattern in the trie.
                    308: 
                    309: 
                    310: <HR><H3>Function</H3>
                    311: <P><I>int</I>
                    312: <B>trie_same</B>
                    313: (<I>struct f_trie *</I> <B>t1</B>, <I>struct f_trie *</I> <B>t2</B>)
                    314: <H3>Arguments</H3>
                    315: <P>
                    316: <DL>
                    317: <DT><I>struct f_trie *</I> <B>t1</B><DD><P>first trie to be compared
                    318: <DT><I>struct f_trie *</I> <B>t2</B><DD><P>second one
                    319: </DL>
                    320: <H3>Description</H3>
                    321: <P>Compares two tries and returns 1 if they are same
                    322: 
                    323: 
                    324: <HR><H3>Function</H3>
                    325: <P><I>void</I>
                    326: <B>trie_format</B>
                    327: (<I>struct f_trie *</I> <B>t</B>, <I>buffer *</I> <B>buf</B>)
                    328: <H3>Arguments</H3>
                    329: <P>
                    330: <DL>
                    331: <DT><I>struct f_trie *</I> <B>t</B><DD><P>trie to be formatted
                    332: <DT><I>buffer *</I> <B>buf</B><DD><P>destination buffer
                    333: </DL>
                    334: <H3>Description</H3>
                    335: <P>Prints the trie to the supplied buffer.
                    336: 
                    337: <HR>
                    338: <A HREF="prog-5.html">Next</A>
                    339: <A HREF="prog-3.html">Previous</A>
                    340: <A HREF="prog.html#toc4">Contents</A>
                    341: </BODY>
                    342: </HTML>

FreeBSD-CVSweb <freebsd-cvsweb@FreeBSD.org>