linux/Documentation/networking/fib_trie.txt
<<
>>
Prefs
   1                        LC-trie implementation notes.
   2
   3Node types
   4----------
   5leaf 
   6        An end node with data. This has a copy of the relevant key, along
   7        with 'hlist' with routing table entries sorted by prefix length.
   8        See struct leaf and struct leaf_info.
   9
  10trie node or tnode
  11        An internal node, holding an array of child (leaf or tnode) pointers,
  12        indexed through a subset of the key. See Level Compression.
  13
  14A few concepts explained
  15------------------------
  16Bits (tnode) 
  17        The number of bits in the key segment used for indexing into the
  18        child array - the "child index". See Level Compression.
  19
  20Pos (tnode)
  21        The position (in the key) of the key segment used for indexing into
  22        the child array. See Path Compression.
  23
  24Path Compression / skipped bits
  25        Any given tnode is linked to from the child array of its parent, using
  26        a segment of the key specified by the parent's "pos" and "bits" 
  27        In certain cases, this tnode's own "pos" will not be immediately
  28        adjacent to the parent (pos+bits), but there will be some bits
  29        in the key skipped over because they represent a single path with no
  30        deviations. These "skipped bits" constitute Path Compression.
  31        Note that the search algorithm will simply skip over these bits when
  32        searching, making it necessary to save the keys in the leaves to
  33        verify that they actually do match the key we are searching for.
  34
  35Level Compression / child arrays
  36        the trie is kept level balanced moving, under certain conditions, the
  37        children of a full child (see "full_children") up one level, so that
  38        instead of a pure binary tree, each internal node ("tnode") may
  39        contain an arbitrarily large array of links to several children.
  40        Conversely, a tnode with a mostly empty child array (see empty_children)
  41        may be "halved", having some of its children moved downwards one level,
  42        in order to avoid ever-increasing child arrays.
  43
  44empty_children
  45        the number of positions in the child array of a given tnode that are
  46        NULL.
  47
  48full_children
  49        the number of children of a given tnode that aren't path compressed.
  50        (in other words, they aren't NULL or leaves and their "pos" is equal
  51        to this tnode's "pos"+"bits").
  52
  53        (The word "full" here is used more in the sense of "complete" than
  54        as the opposite of "empty", which might be a tad confusing.)
  55
  56Comments
  57---------
  58
  59We have tried to keep the structure of the code as close to fib_hash as 
  60possible to allow verification and help up reviewing. 
  61
  62fib_find_node()
  63        A good start for understanding this code. This function implements a
  64        straightforward trie lookup.
  65
  66fib_insert_node()
  67        Inserts a new leaf node in the trie. This is bit more complicated than
  68        fib_find_node(). Inserting a new node means we might have to run the
  69        level compression algorithm on part of the trie.
  70
  71trie_leaf_remove()
  72        Looks up a key, deletes it and runs the level compression algorithm.
  73
  74trie_rebalance()
  75        The key function for the dynamic trie after any change in the trie
  76        it is run to optimize and reorganize. Tt will walk the trie upwards 
  77        towards the root from a given tnode, doing a resize() at each step 
  78        to implement level compression.
  79
  80resize()
  81        Analyzes a tnode and optimizes the child array size by either inflating
  82        or shrinking it repeatedly until it fulfills the criteria for optimal
  83        level compression. This part follows the original paper pretty closely
  84        and there may be some room for experimentation here.
  85
  86inflate()
  87        Doubles the size of the child array within a tnode. Used by resize().
  88
  89halve()
  90        Halves the size of the child array within a tnode - the inverse of
  91        inflate(). Used by resize();
  92
  93fn_trie_insert(), fn_trie_delete(), fn_trie_select_default()
  94        The route manipulation functions. Should conform pretty closely to the
  95        corresponding functions in fib_hash.
  96
  97fn_trie_flush()
  98        This walks the full trie (using nextleaf()) and searches for empty
  99        leaves which have to be removed.
 100
 101fn_trie_dump()
 102        Dumps the routing table ordered by prefix length. This is somewhat
 103        slower than the corresponding fib_hash function, as we have to walk the
 104        entire trie for each prefix length. In comparison, fib_hash is organized
 105        as one "zone"/hash per prefix length.
 106
 107Locking
 108-------
 109
 110fib_lock is used for an RW-lock in the same way that this is done in fib_hash.
 111However, the functions are somewhat separated for other possible locking
 112scenarios. It might conceivably be possible to run trie_rebalance via RCU
 113to avoid read_lock in the fn_trie_lookup() function.
 114
 115Main lookup mechanism
 116---------------------
 117fn_trie_lookup() is the main lookup function.
 118
 119The lookup is in its simplest form just like fib_find_node(). We descend the
 120trie, key segment by key segment, until we find a leaf. check_leaf() does
 121the fib_semantic_match in the leaf's sorted prefix hlist.
 122
 123If we find a match, we are done.
 124
 125If we don't find a match, we enter prefix matching mode. The prefix length,
 126starting out at the same as the key length, is reduced one step at a time,
 127and we backtrack upwards through the trie trying to find a longest matching
 128prefix. The goal is always to reach a leaf and get a positive result from the
 129fib_semantic_match mechanism.
 130
 131Inside each tnode, the search for longest matching prefix consists of searching
 132through the child array, chopping off (zeroing) the least significant "1" of
 133the child index until we find a match or the child index consists of nothing but
 134zeros.
 135
 136At this point we backtrack (t->stats.backtrack++) up the trie, continuing to
 137chop off part of the key in order to find the longest matching prefix.
 138
 139At this point we will repeatedly descend subtries to look for a match, and there
 140are some optimizations available that can provide us with "shortcuts" to avoid
 141descending into dead ends. Look for "HL_OPTIMIZE" sections in the code.
 142
 143To alleviate any doubts about the correctness of the route selection process,
 144a new netlink operation has been added. Look for NETLINK_FIB_LOOKUP, which
 145gives userland access to fib_lookup().
 146
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.