In code-division multiple-access (CDMA) packet data networks, the throughput depends on physical-layer receiver algorithms and medium access control (MAC) layer protocols. Taking a holistic approach, we investigate the throughput of a CDMA network employing linear multiuser detection, type-I automatic retransmission request (ARQ), and packet combining. In particular, the following two models of CDMA data networks are considered. 1) Fixed-access CDMA data networks, by which we mean that the number of active users is fixed. The corresponding throughput is analyzed by using a one-dimensional Markov chain, and conditions for achieving the optimal throughput are explored. 2) Random-access CDMA data networks, in which we take into account the random arrivals/departures of users. By viewing the fixed-access network studied in the first case as a snapshot of the random-access network, we exploit the results therein to analyze the random-access network, under a processor-sharing system model. Moreover, we identify some important properties of the corresponding throughput and devise a simple recursive algorithm to find the throughput-optimal admission region. Some recent advances on large-system performance of various multiuser detection algorithms are employed in our study. The results in this paper quantitatively characterize the potential for network throughput gain by employing multiuser detection and packet combining, in both fixed-access and random-access CDMA packet data networks.