Many, many things. Most notably:

* At least the bit-lp and bit-velle modules do no detection on loading;
  ask Simon whether this is possible to add.
* Check whether the FAN count divisors in the lm78 module are correct
  (probably not!); correct gl518sm changed-fan_div_but_did_not_update_fans
* Wouldn't it be better to display all natural voltages for the LM78 chip,
  and do *all* conversions through a user program?
* kmalloc with GFP_KERNEL can cause a context switch! Check whether this is
  safe everywhere it is used.
* Support 10-bit addresses. At this moment, they are supported nowhere, except
  in Simon Vogl's i2c modules.
* Change the i2c modules to keep the namespace clean (EXPORT_SYMBOL does not
  work for us; many static declarations need to be added; better ask Simon
  Vogl first)
* Make it SMP-safe: there are no spinlocks yet. They are needed at many
  places, probably (everywhere where global vars are accessed). This must be
  done for the i2c modules, too.
* Make lm78.c detect 'double hits', (same chip connected to both SMBus and
  ISA).
* Better lm78/lm75 detection; insmod-time paramters to set addresses 
  (difficult, because of all i2c busses!).
* Registrate which i2c addresses are used (a la ISA) for each adapter.
* Rename i2c-proc.c to i2c-user.c, and extend it with /dev entries;
  or, create a better i2c-dev.c, that understands SMBus commands.
* Write all other drivers
* How to make drivers detect thing *not*; for example, if a Winbond is 
  present, the LM75 driver should not try to access its simulated
  LM75 chips (in this case, simply disable them - but what if the LM75
  driver was loaded first?).
* Write a userland-library for SMBus/i2c access (through the /dev interface)
* Write a userland detection program for busses (clients are less important;
  but perhaps needed too).