On Tue, Oct 3, 2023 at 7:57 PM Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote: > > ok. discard that idea. Attached is a 3rd version of the same idea I argued earlier. Let normal DFS go as normal, do states_equal() on V which has 1 looping branch remain and all other explored. To achieve that when iter_next() is seen do parent->looping_states += 2; then when processing any children do parent->looping_states++; in the correct parent. Since there could be many intermediate states have to walk back parentage chain to increment correct parent. When the state reaches bpf_exit or safety, walk back the parentage chain and do looping_states--. The state is ok to use in states_equal() if looping_states==1. With this patch all existing iter tests still pass, and all Ed's special tests pass or fail as needed. Ex: loop_state_deps1 is rejected with misaligned stack, loop1 loads with success, num_iter_bug fails with bad pointer. Please review. I could be just lucky with the way tests are constructed, but I feel this is a better path to fix this issue instead of DFS/BFS combo that I have doubts about.
Attachment:
0001-iter-hack-3.patch
Description: Binary data
