There are a number of issues with this patch:

1. It doesn't allow for different types of checker - checkers will be merged even if they are of different types.
2. So far as I can see the patch assumes that there will be only 1 non-merged checker per RS (rs->checker implies this).
3. The hash is generated only from the IP address of the real server. This therefore means that all checkers for any particular real server will be merged (this perhaps explains why point 2 above has been implemented). The hash needs to be generated from parameters of the checker itself, and the real server address is irrelevant.
4. The checker u_list is unused.
5. If RS_CHECK_MERGE_BITS is less than 16 some of the bits in addr_fold will be unused. The code should be:

hash = 0;
while (addr_fold) {
    hash ^= addr_fold & RS_CHECK_MERGE_MASK;
    addr_fold >>= RS_CHECK_MERGE_BITS;
}

6. rs_check_merge[RS_CHECK_MERGE_ENTRY_NUM] is huge (it is 8Mb on a 64-bit machine and half that on a 32-bit system) and it would need a very large number of checkers to be efficient.
7. The checker h_list and i_list only need a single list_head entry in the checker structure since a checker is moved from the h_list to the i_list when it is found to be a duplicate, and the checker can never be on both lists.

Having said the above, I like the approach. I will modify the patch in line with the following:

1. The hash will be generated from the parameters of the checker, including the type. We already have a compare function for each checker type so that can be used to compare the checkers if the hash matches.
2. Rather than use the rs_check_merge array of list_heads we will use a red-black tree.
3. Modify the checker i_list so that it is a list of all the RSs that use the checker, rather than having the first one accessed via checker->rs and the subsequent RSs via the i_list.
4. Release the duplicate checkers (i.e. those that were added to the i_list), since they contain identical information to the first one found, and simply maintain a list of RSs that the checker updates. This is similar to the way BFD_CHECK and FILE_CHECK work, where the checkers are defined outside the scope of the virtual/real servers, and the checker configuration references the separate track_file/bfd_instance.
5. We need to take account of ha_suspend.
6. Allow multiple checkers per RS.
7. Update migrate_checkers(), used when reloading.