9.3 Configuring BGP  
  9.3.7 BGP continuity inside an AS  
BGP does not advertise routes learned by way of IBGP peers to other IBGP peers. If BGP did, BGP routing inside the AS would present a dangerous potential for routing loops. For IBGP routers to learn about all BGP routes inside the AS, they must connect to every other IBGP router in a full IBGP mesh. This full mesh needs to be only logical, not physical. In other words, as long as the IBGP peers can connect to each other using TCP/IP, a logical full mesh can be created even if the routers are not directly connected.

Figure illustrates one of the common mistakes that administrators make when setting BGP routing within an AS. This ISP network has three POPs (San Jose, San Francisco, and Los Angeles). Each POP has multiple non-BGP routers and a BGP border router running EBGP with other autonomous systems. The administrator has set up an IBGP connection between the San Jose border router and the San Francisco border router. Another IBGP connection has been set up between the San Francisco border router and the Los Angeles border router. In this configuration, EBGP routes learned by way of San Jose are given to San Francisco. EBGP routes learned by way of San Francisco are given to San Jose and Los Angeles. Lastly EBGP routes learned by way of Los Angeles are given to San Francisco.

However, routing in this scenario is not complete. EBGP routes learned by way of San Jose will not be given to Los Angeles, and EBGP routes learned by way of Los Angeles will not be given to San Jose. This is because the San Francisco router will not advertise IBGP routes between San Jose and Los Angeles. What is needed is an additional IBGP connection between San Jose and Los Angeles. This connection is shown as a dotted line.