There's a long standing bug in blk_recount_segments(). Clustering means
physical segment colascing, not hardware segment colascing. This
basically means that we are mapping more segments than here than the bio
+ requests contains, and this causes a bug in the SCSI layer for host
adapters that have CLUSTERING enabled.

This patch makes sure that we are clustering physical segments
correctly, and correctly accounting hardware segments. Please apply.