[IPV4]: Add LC-Trie implementation notes

Signed-off-by: Robert Olsson <Robert.Olsson@data.slu.se>
Signed-off-by: David S. Miller <davem@davemloft.net>

diff --git a/Documentation/networking/fib_trie.txt b/Documentation/networking/fib_trie.txt
new file mode 100644 (file)
index 0000000..f50d0c6
--- /dev/null
+++ b/Documentation/networking/fib_trie.txt
@@ -0,0 +1,145 @@
+                       LC-trie implementation notes.
+
+Node types
+----------
+leaf 
+       An end node with data. This has a copy of the relevant key, along
+       with 'hlist' with routing table entries sorted by prefix length.
+       See struct leaf and struct leaf_info.
+
+trie node or tnode
+       An internal node, holding an array of child (leaf or tnode) pointers,
+       indexed through a subset of the key. See Level Compression.
+
+A few concepts explained
+------------------------
+Bits (tnode) 
+       The number of bits in the key segment used for indexing into the
+       child array - the "child index". See Level Compression.
+
+Pos (tnode)
+       The position (in the key) of the key segment used for indexing into
+       the child array. See Path Compression.
+
+Path Compression / skipped bits
+       Any given tnode is linked to from the child array of its parent, using
+       a segment of the key specified by the parent's "pos" and "bits" 
+       In certain cases, this tnode's own "pos" will not be immediately
+       adjacent to the parent (pos+bits), but there will be some bits
+       in the key skipped over because they represent a single path with no
+       deviations. These "skipped bits" constitute Path Compression.
+       Note that the search algorithm will simply skip over these bits when
+       searching, making it necessary to save the keys in the leaves to
+       verify that they actually do match the key we are searching for.
+
+Level Compression / child arrays
+       the trie is kept level balanced moving, under certain conditions, the
+       children of a full child (see "full_children") up one level, so that
+       instead of a pure binary tree, each internal node ("tnode") may
+       contain an arbitrarily large array of links to several children.
+       Conversely, a tnode with a mostly empty child array (see empty_children)
+       may be "halved", having some of its children moved downwards one level,
+       in order to avoid ever-increasing child arrays.
+
+empty_children
+       the number of positions in the child array of a given tnode that are
+       NULL.
+
+full_children
+       the number of children of a given tnode that aren't path compressed.
+       (in other words, they aren't NULL or leaves and their "pos" is equal
+       to this tnode's "pos"+"bits").
+
+       (The word "full" here is used more in the sense of "complete" than
+       as the opposite of "empty", which might be a tad confusing.)
+
+Comments
+---------
+
+We have tried to keep the structure of the code as close to fib_hash as 
+possible to allow verification and help up reviewing. 
+
+fib_find_node()
+       A good start for understanding this code. This function implements a
+       straightforward trie lookup.
+
+fib_insert_node()
+       Inserts a new leaf node in the trie. This is bit more complicated than
+       fib_find_node(). Inserting a new node means we might have to run the
+       level compression algorithm on part of the trie.
+
+trie_leaf_remove()
+       Looks up a key, deletes it and runs the level compression algorithm.
+
+trie_rebalance()
+       The key function for the dynamic trie after any change in the trie
+       it is run to optimize and reorganize. Tt will walk the trie upwards 
+       towards the root from a given tnode, doing a resize() at each step 
+       to implement level compression.
+
+resize()
+       Analyzes a tnode and optimizes the child array size by either inflating
+       or shrinking it repeatedly until it fullfills the criteria for optimal
+       level compression. This part follows the original paper pretty closely
+       and there may be some room for experimentation here.
+
+inflate()
+       Doubles the size of the child array within a tnode. Used by resize().
+
+halve()
+       Halves the size of the child array within a tnode - the inverse of
+       inflate(). Used by resize();
+
+fn_trie_insert(), fn_trie_delete(), fn_trie_select_default()
+       The route manipulation functions. Should conform pretty closely to the
+       corresponding functions in fib_hash.
+
+fn_trie_flush()
+       This walks the full trie (using nextleaf()) and searches for empty
+       leaves which have to be removed.
+
+fn_trie_dump()
+       Dumps the routing table ordered by prefix length. This is somewhat
+       slower than the corresponding fib_hash function, as we have to walk the
+       entire trie for each prefix length. In comparison, fib_hash is organized
+       as one "zone"/hash per prefix length.
+
+Locking
+-------
+
+fib_lock is used for an RW-lock in the same way that this is done in fib_hash.
+However, the functions are somewhat separated for other possible locking
+scenarios. It might conceivably be possible to run trie_rebalance via RCU
+to avoid read_lock in the fn_trie_lookup() function.
+
+Main lookup mechanism
+---------------------
+fn_trie_lookup() is the main lookup function.
+
+The lookup is in its simplest form just like fib_find_node(). We descend the
+trie, key segment by key segment, until we find a leaf. check_leaf() does
+the fib_semantic_match in the leaf's sorted prefix hlist.
+
+If we find a match, we are done.
+
+If we don't find a match, we enter prefix matching mode. The prefix length,
+starting out at the same as the key length, is reduced one step at a time,
+and we backtrack upwards through the trie trying to find a longest matching
+prefix. The goal is always to reach a leaf and get a positive result from the
+fib_semantic_match mechanism.
+
+Inside each tnode, the search for longest matching prefix consists of searching
+through the child array, chopping off (zeroing) the least significant "1" of
+the child index until we find a match or the child index consists of nothing but
+zeros.
+
+At this point we backtrack (t->stats.backtrack++) up the trie, continuing to
+chop off part of the key in order to find the longest matching prefix.
+
+At this point we will repeatedly descend subtries to look for a match, and there
+are some optimizations available that can provide us with "shortcuts" to avoid
+descending into dead ends. Look for "HL_OPTIMIZE" sections in the code.
+
+To alleviate any doubts about the correctness of the route selection process,
+a new netlink operation has been added. Look for NETLINK_FIB_LOOKUP, which
+gives userland access to fib_lookup().