The large system analysis of randomly spread direct-sequence code-division multiple-access systems operating over frequency-selective fading channels is considered. Iterative multiuser detection and decoding (MUDD) based on generalized posterior mean estimation and single-user sum-product decoding is assumed to be used at the receiver. The channel state information (CSI) at the MUDD is mismatched and obtained by a linear channel estimator whose initial decisions are iteratively refined with the help of information feedback provided by the MUDD. Furthermore, a new training method by means of probability-biased signaling is proposed. The results indicate that in the large system limit and under certain threshold loads, a near single-user performance with perfect CSI can be achieved using a vanishing training overhead. It is also found that for the considered setups, the iterative linear minimum mean square error based channel estimator is near optimal for relatively slowly time-varying multipath fading channels.