I have a bunch of 3750x switches that each have a 10 gig routed link back to a central 4507 (loopback = 172.30.255.255).

We carved up a /24 (of course, the /24 doesn't really exist except in our address tracking spreadsheet) into a bunch of /30's for routed WAN links and /32's for loopback addresses.

We started on the low end for /30 subnets (ie 172.30.255.0/30, 172.30.255.4/30, etc.).

We started at the high end for the /32 loopbacks (ie 172.30.255.255/32, 172.30.255.254/32, etc.)

Pretty simple, right?

Well, when I try pinging 172.30.255.255 from the access layer 3750x switches, the 3750x seems to be treating it as a broadcast ping where it lists each member that responds instead of the regular !!!!! response (this makes think something is odd with the 3750x).  Of course, only one member responds (the core).  But even the core seems to respond with the other end of the /30 instead of the actual /32 loopback (which makes me think something is odd in the core). 

I could have sworn that I've setup similar topologies without problems (ie, using 10.0.0.0/32, 10.255.255.255/32, etc as loopbacks) and as long as the mask is a /32, it should work.

Also, I can ping/ssh to that loopback if my laptop is on a directly connected subnet.  But I can't do it from any of the 3750x switches (which are also directly connected).

I've double checked for overlapping subnets, but nope.  I don't see any.  Routing looks fine.  The actual /32 is being propagated everywhere properly.

Am I running into some legacy behavior?  I thought maybe something having to do with ip classless being disabled???

Any ideas?