3c527.c

/* 3c527.c: 3Com Etherlink/MC32 driver for Linux 2.4 and 2.6.
 *
 *	(c) Copyright 1998 Red Hat Software Inc
 *	Written by Alan Cox.
 *	Further debugging by Carl Drougge.
 *      Initial SMP support by Felipe W Damasio <felipewd@terra.com.br>
 *      Heavily modified by Richard Procter <rnp@paradise.net.nz>
 *
 *	Based on skeleton.c written 1993-94 by Donald Becker and ne2.c
 *	(for the MCA stuff) written by Wim Dumon.
 *
 *	Thanks to 3Com for making this possible by providing me with the
 *	documentation.
 *
 *	This software may be used and distributed according to the terms
 *	of the GNU General Public License, incorporated herein by reference.
 *
 */

#define DRV_NAME		"3c527"
#define DRV_VERSION		"0.7-SMP"
#define DRV_RELDATE		"2003/09/21"

static const char *version =
DRV_NAME ".c:v" DRV_VERSION " " DRV_RELDATE " Richard Procter <rnp@paradise.net.nz>\n";

/**
 * DOC: Traps for the unwary
 *
 *	The diagram (Figure 1-1) and the POS summary disagree with the
 *	"Interrupt Level" section in the manual.
 *
 *	The manual contradicts itself when describing the minimum number
 *	buffers in the 'configure lists' command.
 *	My card accepts a buffer config of 4/4.
 *
 *	Setting the SAV BP bit does not save bad packets, but
 *	only enables RX on-card stats collection.
 *
 *	The documentation in places seems to miss things. In actual fact
 *	I've always eventually found everything is documented, it just
 *	requires careful study.
 *
 * DOC: Theory Of Operation
 *
 *	The 3com 3c527 is a 32bit MCA bus mastering adapter with a large
 *	amount of on board intelligence that housekeeps a somewhat dumber
 *	Intel NIC. For performance we want to keep the transmit queue deep
 *	as the card can transmit packets while fetching others from main
 *	memory by bus master DMA. Transmission and reception are driven by
 *	circular buffer queues.
 *
 *	The mailboxes can be used for controlling how the card traverses
 *	its buffer rings, but are used only for initial setup in this
 *	implementation.  The exec mailbox allows a variety of commands to
 *	be executed. Each command must complete before the next is
 *	executed. Primarily we use the exec mailbox for controlling the
 *	multicast lists.  We have to do a certain amount of interesting
 *	hoop jumping as the multicast list changes can occur in interrupt
 *	state when the card has an exec command pending. We defer such
 *	events until the command completion interrupt.
 *
 *	A copy break scheme (taken from 3c59x.c) is employed whereby
 *	received frames exceeding a configurable length are passed
 *	directly to the higher networking layers without incuring a copy,
 *	in what amounts to a time/space trade-off.
 *
 *	The card also keeps a large amount of statistical information
 *	on-board. In a perfect world, these could be used safely at no
 *	cost. However, lacking information to the contrary, processing