Tulip Ethernet Card Driver Maintained by Jeff Garzik The Tulip driver is developed by Donald Becker and changed by Takashi Manabe and a cast of thousands. This driver is for the Digital "Tulip" Ethernet adapter interface. It should work with most DEC 21*4*-based chips/ethercards, as well as with work-alike chips from Lite-On (PNIC) and Macronix (MXIC) and ASIX. The author may be reached as becker@CESDIS.gsfc.nasa.gov, or C/O Center of Excellence in Space Data and Information Sciences Code 930.5, Goddard Space Flight Center, Greenbelt MD 20771 Additional information available at http://cesdis.gsfc.nasa.gov/linux/drivers/tulip.html Theory of Operation Board Compatibility =================== This device driver is designed for the DECchip "Tulip", Digital's single-chip ethernet controllers for PCI. Supported members of the family are the 21040, 21041, 21140, 21140A, 21142, and 21143. Similar work-alike chips from Lite-On, Macronics, ASIX, Compex and other listed below are also supported. These chips are used on at least 140 unique PCI board designs. The great number of chips and board designs supported is the reason for the driver size and complexity. Almost of the increasing complexity is in the board configuration and media selection code. There is very little increasing in the operational critical path length. Board-specific settings ======================= PCI bus devices are configured by the system at boot time, so no jumpers need to be set on the board. The system BIOS preferably should assign the PCI INTA signal to an otherwise unused system IRQ line. Some boards have EEPROMs tables with default media entry. The factory default is usually "autoselect". This should only be overridden when using transceiver connections without link beat e.g. 10base2 or AUI, or (rarely!) for forcing full-duplex when used with old link partners that do not do autonegotiation. Driver operation ================ Ring buffers ------------ The Tulip can use either ring buffers or lists of Tx and Rx descriptors. This driver uses statically allocated rings of Rx and Tx descriptors, set at compile time by RX/TX_RING_SIZE. This version of the driver allocates skbuffs for the Rx ring buffers at open() time and passes the skb->data field to the Tulip as receive data buffers. When an incoming frame is less than RX_COPYBREAK bytes long, a fresh skbuff is allocated and the frame is copied to the new skbuff. When the incoming frame is larger, the skbuff is passed directly up the protocol stack and replaced by a newly allocated skbuff. The RX_COPYBREAK value is chosen to trade-off the memory wasted by using a full-sized skbuff for small frames vs. the copying costs of larger frames. For small frames the copying cost is negligible (esp. considering that we are pre-loading the cache with immediately useful header information). For large frames the copying cost is non-trivial, and the larger copy might flush the cache of useful data. A subtle aspect of this choice is that the Tulip only receives into longword aligned buffers, thus the IP header at offset 14 isn't longword aligned for further processing. Copied frames are put into the new skbuff at an offset of "+2", thus copying has the beneficial effect of aligning the IP header and preloading the cache. Synchronization --------------- The driver runs as two independent, single-threaded flows of control. One is the send-packet routine, which enforces single-threaded use by the dev->tbusy flag. The other thread is the interrupt handler, which is single threaded by the hardware and other software. The send packet thread has partial control over the Tx ring and 'dev->tbusy' flag. It sets the tbusy flag whenever it's queuing a Tx packet. If the next queue slot is empty, it clears the tbusy flag when finished otherwise it sets the 'tp->tx_full' flag. The interrupt handler has exclusive control over the Rx ring and records stats from the Tx ring. (The Tx-done interrupt can't be selectively turned off, so we can't avoid the interrupt overhead by having the Tx routine reap the Tx stats.) After reaping the stats, it marks the queue entry as empty by setting the 'base' to zero. Iff the 'tp->tx_full' flag is set, it clears both the tx_full and tbusy flags. Notes ===== Thanks to Duke Kamstra of SMC for long ago providing an EtherPower board. Greg LaPolla at Linksys provided PNIC and other Linksys boards. Znyx provided a four-port card for testing. References ========== http://cesdis.gsfc.nasa.gov/linux/misc/NWay.html http://www.digital.com (search for current 21*4* datasheets and "21X4 SROM") http://www.national.com/pf/DP/DP83840A.html http://www.asix.com.tw/pmac.htm http://www.admtek.com.tw/ Errata ====== The old DEC databooks were light on details. The 21040 databook claims that CSR13, CSR14, and CSR15 should each be the last register of the set CSR12-15 written. Hmmm, now how is that possible? The DEC SROM format is very badly designed not precisely defined, leading to part of the media selection junkheap below. Some boards do not have EEPROM media tables and need to be patched up. Worse, other boards use the DEC design kit media table when it isn't correct for their board. We cannot use MII interrupts because there is no defined GPIO pin to attach them. The MII transceiver status is polled using an kernel timer. Source tree tour ================ The following is a list of files comprising the Tulip ethernet driver in drivers/net/tulip subdirectory. 21142.c - 21142-specific h/w interaction eeprom.c - EEPROM reading and parsing interrupt.c - Interrupt handler media.c - Media selection and MII support pnic.c - PNIC-specific h/w interaction timer.c - Main driver timer, and misc h/w timers tulip.h - Private driver header tulip_core.c - Driver core (a.k.a. where "everything else" goes) Version history =============== 0.9.4.2 (March 21, 2000): * Fix 21041 CSR7, CSR13/14/15 handling * Merge some PCI ids from tulip 0.91x * Merge some HAS_xxx flags and flag settings from tulip 0.91x * asm/io.h fix (submitted by many) and cleanup * s/HAS_NWAY143/HAS_NWAY/ * Cleanup 21041 mode reporting * Small code cleanups 0.9.4.1 (March 18, 2000): * Finish PCI DMA conversion (davem) * Do not netif_start_queue() at end of tulip_tx_timeout() (kuznet) * PCI DMA fix (kuznet) * eeprom.c code cleanup * Remove Xircom Tulip crud [EOF]