Quite a lot of people are complaining on confusing "Out of chain frames" reported by mps(4) and mpr(4) drivers during periods of high load. Aside of pure cosmetics, it causes request retrial, that is found to not be handled properly by some software, and possibly also reduces performance. Investigation shows that present count of chain frames is underestimated by 5 times comparing to potentially needed (2048 vs 10240).
Proposed patch removed hard limits on number of chain frames in the drivers, allocating as much of them as needed to satisfy all declared number of requests at maximum data size with worst possible fragmentation. With the default value MAXPHYS of 128KB the price of this change is about 1MB of memory per HBA, that does not sound very serious these days for server-grade HBA.
People, that for some reason want that limitation back, can still do it via the hw.mps.max_chains loader tunable, same as before.