Merge branch 'cache-compile' into dev-master

Author: Joshua Reich

pyretic/core/classifier.py

@@ -0,0 +1,263 @@
+###############################################################################
+# Classifiers
+# an intermediate representation for proactive compilation.
+
+class Rule(object):
+    """
+    A rule contains a filter and the parallel composition of zero or more
+    Pyretic actions.
+    """
+
+    # Matches m should be of the match class.  Actions acts should be a list of
+    # either modify, identity, or drop policies.
+    def __init__(self,m,acts):
+        self.match = m
+        self.actions = acts
+
+    def __str__(self):
+        return str(self.match) + '\n  -> ' + str(self.actions)
+
+    def __repr__(self):
+        return str(self)
+
+    def __eq__(self, other):
+        """Based on syntactic equality of policies."""
+        return ( id(self) == id(other)
+            or ( self.match == other.match
+                 and self.actions == other.actions ) )
+
+    def __ne__(self, other):
+        """Based on syntactic equality of policies."""
+        return not (self == other)
+
+    def eval(self, in_pkt):
+        """
+        If this rule matches the packet, then return the union of the sets
+        of packets produced by the actions.  Otherwise, return None.
+        """
+        filtered_pkt = self.match.eval(in_pkt)
+        if len(filtered_pkt) == 0:
+            return None
+        rv = set()
+        for pkt in filtered_pkt:
+            for act in self.actions:
+                rv |= act.eval(pkt)
+        return rv
+
+
+class Classifier(object):
+    """
+    A classifier contains a list of rules, where the order of the list implies
+    the relative priorities of the rules.  Semantically, classifiers are
+    functions from packets to sets of packets, similar to OpenFlow flow
+    tables.
+    """
+
+    def __init__(self, new_rules=[]):
+        import types
+        if isinstance(new_rules, types.GeneratorType):
+            self.rules = [r for r in new_rules]
+        elif isinstance(new_rules,list):
+            self.rules = new_rules
+        else:
+            raise TypeError
+
+    def __len__(self):
+        return len(self.rules)
+
+    def __str__(self):
+        return '\n '.join(map(str,self.rules))
+
+    def __repr__(self):
+        return str(self)
+
+    def __eq__(self, other):
+        """Based on syntactic equality of policies."""
+        return ( id(self) == id(other)
+            or ( self.rules == other.rules ) )
+
+    def __ne__(self, other):
+        """Based on syntactic equality of policies."""
+        return not (self == other)
+
+    def __add__(self,c2):
+        from pyretic.core.language import drop
+        c1 = self
+        if c2 is None:
+            return None
+        c = Classifier([])
+        # TODO (cole): make classifiers iterable
+        for r1 in c1.rules:
+            for r2 in c2.rules:
+                intersection = r1.match.intersect(r2.match)
+                if intersection != drop:
+                    # TODO (josh) logic for detecting when sets of actions can't be combined
+                    # e.g., [modify(dstip='10.0.0.1'),fwd(1)] + [modify(srcip='10.0.0.2'),fwd(2)]
+                    actions = r1.actions + r2.actions
+                    actions = filter(lambda a: a != drop,actions)
+                    if len(actions) == 0:
+                        actions = [drop]
+                    c.rules.append(Rule(intersection, actions))
+        for r1 in c1.rules:
+            c.rules.append(r1)
+        for r2 in c2.rules:
+            c.rules.append(r2)
+        return c.optimize()
+
+    # Helper function for rshift: given a test b and an action p, return a test
+    # b' such that p >> b == b' >> p.
+    def _commute_test(self, act, pkts):
+        from pyretic.core.language import drop, identity, Controller, CountBucket, DerivedPolicy
+        while isinstance(act, DerivedPolicy):
+            act = act.policy
+        if act == identity:
+            return pkts
+        elif act == drop:
+            return drop
+        elif act == Controller or isinstance(act, CountBucket):
+            return identity
+        elif isinstance(act, modify):
+            new_match_dict = {}
+            if pkts == identity:
+                return identity
+            elif pkts == drop:
+                return drop
+            for f, v in pkts.map.iteritems():
+                if f in act.map and act.map[f] == v:
+                    continue
+                elif f in act.map and act.map[f] != v:
+                    return drop
+                else:
+                    new_match_dict[f] = v
+            if len(new_match_dict) == 0:
+                return identity
+            return match(**new_match_dict)
+        else:
+            # TODO (cole) use compile error.
+            # TODO (cole) what actions are allowable?
+            raise TypeError
+
+    # Helper function for rshift: sequentially compose actions.  a1 must be a
+    # single action.  Returns a list of actions.
+    def _sequence_actions(self, a1, as2):
+        from pyretic.core.language import drop, identity, Controller, CountBucket, DerivedPolicy
+        while isinstance(a1, DerivedPolicy):
+            a1 = a1.policy
+        # TODO: be uniform about returning copied or modified objects.
+        new_actions = []
+        if a1 == drop:
+            return [drop]
+        elif a1 == identity:
+            return as2
+        elif a1 == Controller or isinstance(a1, CountBucket):
+            return [a1]
+        elif isinstance(a1, modify):
+            for a2 in as2:
+                while isinstance(a2, DerivedPolicy):
+                    a2 = a2.policy
+                new_a1 = modify(**a1.map.copy())
+                if a2 == drop:
+                    new_actions.append(drop)
+                elif a2 == Controller or isinstance(a2, CountBucket): 
+                    new_actions.append(a2)
+                elif a2 == identity:
+                    new_actions.append(new_a1)
+                elif isinstance(a2, modify):
+                    new_a1.map.update(a2.map)
+                    new_actions.append(new_a1)
+                elif isinstance(a2, fwd):
+                    new_a1.map['outport'] = a2.outport
+                    new_actions.append(new_a1)
+                else:
+                    raise TypeError
+            return new_actions
+        else:
+            raise TypeError
+
+    # Returns a classifier.
+    def _sequence_action_classifier(self, act, c):
+        from pyretic.core.language import drop, identity
+        # TODO (cole): make classifiers easier to use w.r.t. adding/removing
+        # rules.
+        if len(c.rules) == 0:
+            return Classifier([Rule(identity, [drop])])
+        new_rules = []
+        for rule in c.rules:
+            pkts = self._commute_test(act, rule.match)
+            if pkts == identity:
+                acts = self._sequence_actions(act, rule.actions)
+                new_rules += [Rule(identity, acts)]
+                break
+            elif pkts == drop:
+                continue
+            else:
+                acts = self._sequence_actions(act, rule.actions)
+                new_rules += [Rule(pkts, acts)]
+        if new_rules == []:
+            return Classifier([Rule(identity, [drop])])
+        else:
+            return Classifier(new_rules)
+                
+    def _sequence_actions_classifier(self, acts, c):
+        from pyretic.core.language import drop, identity
+        empty_classifier = Classifier([Rule(identity, [drop])])
+        if acts == []:
+            # Treat the empty list of actions as drop.
+            return empty_classifier
+        acc = empty_classifier
+        for act in acts:
+            acc = acc + self._sequence_action_classifier(act, c)
+        return acc
+
+    def _sequence_rule_classifier(self, r, c):
+        from pyretic.core.language import drop
+        c2 = self._sequence_actions_classifier(r.actions, c)
+        for rule in c2.rules:
+            rule.match = rule.match.intersect(r.match)
+        c2.rules = [r2 for r2 in c2.rules if r2.match != drop]
+        return c2.optimize()
+
+    def __rshift__(self, c2):
+        new_rules = []
+        for rule in self.rules:
+            c3 = self._sequence_rule_classifier(rule, c2)
+            new_rules = new_rules + c3.rules
+        rv = Classifier(new_rules)
+        return rv.optimize()
+
+    def optimize(self):
+        return self.remove_shadowed_cover_single()
+
+    def remove_shadowed_exact_single(self):
+        # Eliminate every rule exactly matched by some higher priority rule
+        opt_c = Classifier([])
+        for r in self.rules:
+            if not reduce(lambda acc, new_r: acc or
+                          new_r.match == r.match,
+                          opt_c.rules,
+                          False):
+                opt_c.rules.append(r)
+        return opt_c
+
+    def remove_shadowed_cover_single(self):
+        # Eliminate every rule completely covered by some higher priority rule
+        opt_c = Classifier([])
+        for r in self.rules:
+            if not reduce(lambda acc, new_r: acc or
+                          new_r.match.covers(r.match),
+                          opt_c.rules,
+                          False):
+                opt_c.rules.append(r)
+        return opt_c
+
+    def eval(self, in_pkt):
+        """
+        Evaluate against each rule in the classifier, starting with the
+        highest priority.  Return the set of packets resulting from applying
+        the actions of the first rule that matches.
+        """
+        for rule in self.rules:
+            pkts = rule.eval(in_pkt)
+            if pkts is not None:
+                return pkts
+        raise TypeError('Classifier is not total.')