On Tue, Aug 26, 2014 at 12:02:48PM -0700, Jarno Rajahalme wrote: > > On Aug 26, 2014, at 3:01 AM, Thomas Graf <tg...@noironetworks.com> wrote: > > > The branch is unused as size < sizeof dst->inline_values must > > always be true for inlined values. Hitting the branch would lead > > to corruption as inline_values is accessed out of bounds. > > > > Miniflows can also be dynamically allocated to have more inline space than > the default amount. classifier.c makes use of that, but it never calls > miniflow_move (even indirectly) so you are not getting an assert fail. The > motivation for right-sized miniflows in the classifier is to reduce levels of > indirection while not wasting memory. > > I should (have) extend(ed) the comment above struct miniflow definition in > lib/flow.h to describe the above. > > size is computed from the actual source miniflow instance, not from the > generic miniflow struct. If size really is larger than the default inline > space, then obviously no out of bounds access would occur. > > I?m OK changing this function to not work for dynamically allocated > miniflows, but that should be reflected in the comment above the function, > which currently reads: > > /* Initializes 'dst' with the data in 'src', destroying 'src'. > * The caller must eventually free 'dst' with miniflow_destroy(). > * 'dst' must be regularly sized miniflow, but 'src' can have > * larger than default inline values. */ > > However, it might be better to revert this patch to keep the > larger-than-default miniflows not stand out too much.
Let's revert it and extend the comment above struct miniflow. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
