Normally in the manufacture of a camera sensor the photosensitive "pixels" are formed on top of a silicon wafer, onto which several layers of circuitry are added to facilitate reading out the pixel values. This circuitry blocks some of the incident light from hitting the photosensitive areas, reducing sensitivity of the sensor (thereby requiring more amplification, which increases noise).
BSI sensors are created in the same way, but the silicon wafer is flipped over and ground down to make it thin enough for light to shine through from the other side. The readout circuitry no longer gets in the way and allows the sensor to capture up to twice as much light.
There are problems associated with this technique: mounting the circuitry that way increases cross-talk, whereby signals on different lines interfere with each other - this can cause pixels to bleed into each other.
The only commercial BSI sensors to date are very small units, cell phone and compact sizes. The technology is regarded by some as a bit of a marketing gimmick, not really producing the claimed benefits. This is principally due to:
Efficiency is more important with smaller sensors as their smaller pixels capture less light to begin with.
Gains from moving the wiring to the back are apparently greatest when the pixel sizes hit around 1.1 microns (such as the case with the 8MP iPhone sensor). For larger pixels the losses due to the wiring are not as great (as there's more space for the wires).
Having the metalisation layer on the front also causes diffraction effects which are significant as the pixels are only a couple of times the wavelength of light.
The manufacturing processes is more difficult, reducing yield, making it costly to scale up the design.
BSI sensors are mechanically much weaker due to the wafer thinning, a large BSI sensor would be very prone to breakage.