hamachi.c

Linux内核源代码 为压缩文件 是<<Linux内核>>一书中的源代码

/* hamachi.c: A Packet Engines GNIC-II Gigabit Ethernet driver for Linux. */
/*
	Written 1998-2000 by Donald Becker.
	Updates 2000 by Keith Underwood.

	This software may be used and distributed according to the terms of 
	the GNU Public License (GPL), incorporated herein by reference.
	Drivers based on or derived from this code fall under the GPL and must
	retain the authorship, copyright and license notice.  This file is not
	a complete program and may only be used when the entire operating
	system is licensed under the GPL.

	The author may be reached as [e-mail], or C/O
	Scyld Computing Corporation
	410 Severn Ave., Suite 210
	Annapolis MD 21403

	This driver is for the Packet Engines GNIC-II PCI Gigabit Ethernet
	adapter.

	Support and updates available at
	http://www.scyld.com/network/hamachi.html
	or
	http://www.parl.clemson.edu/~keithu/hamachi.html

	For best viewing, set your tabs to 3.

*/

static const char *version =
"hamachi.c:v1.01 5/16/2000  Written by Donald Becker\n"
"   Some modifications by Eric kasten <[e-mail]>\n"
"   Further modifications by Keith Underwood <[e-mail]>\n"
"   Support by many others\n"
"   http://www.scyld.com/network/hamachi.html\n"
"   or\n"
"   http://www.parl.clemson.edu/~keithu/drivers/hamachi.html\n";


/* A few user-configurable values. */

static int debug = 1;		/* 1 normal messages, 0 quiet .. 7 verbose.  */
#define final_version
#define hamachi_debug debug
/* Maximum events (Rx packets, etc.) to handle at each interrupt. */
static int max_interrupt_work = 40;
static int mtu = 0;
/* Default values selected by testing on a dual processor PIII-450 */
/* These six interrupt control parameters may be set directly when loading the
 * module, or through the rx_params and tx_params variables
 */
static int max_rx_latency = 0x11;
static int max_rx_gap = 0x05;
static int min_rx_pkt = 0x18;
static int max_tx_latency = 0x00; 
static int max_tx_gap = 0x00;
static int min_tx_pkt = 0x30;

/* Set the copy breakpoint for the copy-only-tiny-frames scheme.
   -Setting to > 1518 causes all frames to be copied
	-Setting to 0 disables copies
*/
static int rx_copybreak = 0;

/* An override for the hardware detection of bus width.
	Set to 1 to force 32 bit PCI bus detection.  Set to 4 to force 64 bit.
	Add 2 to disable parity detection.
*/
static int force32 = 0;


/* Used to pass the media type, etc.
   These exist for driver interoperability.
   No media types are currently defined.
		- The lower 4 bits are reserved for the media type.
		- The next three bits may be set to one of the following:
			0x00000000 : Autodetect PCI bus
			0x00000010 : Force 32 bit PCI bus
			0x00000020 : Disable parity detection
			0x00000040 : Force 64 bit PCI bus
			Default is autodetect
		- The next bit can be used to force half-duplex.  This is a bad
		  idea since no known implementations implement half-duplex, and,
		  in general, half-duplex for gigabit ethernet is a bad idea.
			0x00000080 : Force half-duplex 
			Default is full-duplex.
		- In the original driver, the ninth bit could be used to force
		  full-duplex.  Maintain that for compatibility
		   0x00000200 : Force full-duplex
*/
#define MAX_UNITS 8				/* More are supported, limit only on options */
static int options[MAX_UNITS] = {-1, -1, -1, -1, -1, -1, -1, -1};
static int full_duplex[MAX_UNITS] = {-1, -1, -1, -1, -1, -1, -1, -1};
/* The Hamachi chipset supports 3 parameters each for Rx and Tx
 * interruput management.  Parameters will be loaded as specified into
 * the TxIntControl and RxIntControl registers.  
 *
 * The registers are arranged as follows:
 *     23 - 16   15 -  8   7    -    0
 *    _________________________________
 *   | min_pkt | max_gap | max_latency |
 *    ---------------------------------
 *   min_pkt      : The minimum number of packets processed between
 *                  interrupts. 
 *   max_gap      : The maximum inter-packet gap in units of 8.192 us
 *   max_latency  : The absolute time between interrupts in units of 8.192 us
 * 
 */
static int rx_params[MAX_UNITS] = {-1, -1, -1, -1, -1, -1, -1, -1};
static int tx_params[MAX_UNITS] = {-1, -1, -1, -1, -1, -1, -1, -1};

/* Operational parameters that are set at compile time. */

/* Keep the ring sizes a power of two for compile efficiency.
	The compiler will convert <unsigned>'%'<2^N> into a bit mask.
   Making the Tx ring too large decreases the effectiveness of channel
   bonding and packet priority.
   There are no ill effects from too-large receive rings, except for
	excessive memory usage */
/* Empirically it appears that the Tx ring needs to be a little bigger
   for these Gbit adapters or you get into an overrun condition really
   easily.  Also, things appear to work a bit better in back-to-back
   configurations if the Rx ring is 8 times the size of the Tx ring
*/
#define TX_RING_SIZE	64
#define RX_RING_SIZE	512

/*
 * Enable mii_ioctl.  Added interrupt coalescing parameter adjustment.
 * 2/19/99 Pete Wyckoff <[e-mail]>
 */
#define HAVE_PRIVATE_IOCTL

/* play with 64-bit addrlen; seems to be a teensy bit slower  --pw */
/* #define ADDRLEN 64 */

/*
 * RX_CHECKSUM turns on card-generated receive checksum generation for
 *   TCP and UDP packets.  Otherwise the upper layers do the calculation.
 * TX_CHECKSUM won't do anything too useful, even if it works.  There's no
 *   easy mechanism by which to tell the TCP/UDP stack that it need not
 *   generate checksums for this device.  But if somebody can find a way
 *   to get that to work, most of the card work is in here already.
 * 3/10/1999 Pete Wyckoff <[e-mail]>
 */
#undef  TX_CHECKSUM
#define RX_CHECKSUM

/* Operational parameters that usually are not changed. */
/* Time in jiffies before concluding the transmitter is hung. */
#define TX_TIMEOUT  (5*HZ)

#include <linux/module.h>
#include <linux/kernel.h>
#include <linux/sched.h>
#include <linux/string.h>
#include <linux/timer.h>
#include <linux/time.h>
#include <linux/ptrace.h>
#include <linux/errno.h>
#include <linux/ioport.h>
#include <linux/malloc.h>
#include <linux/interrupt.h>
#include <linux/pci.h>
#include <linux/init.h>

#include <asm/processor.h>	/* Processor type for cache alignment. */
#include <asm/bitops.h>
#include <asm/io.h>
#include <asm/unaligned.h>
#include <asm/cache.h>

#include <linux/netdevice.h>
#include <linux/etherdevice.h>
#include <linux/skbuff.h>
#include <linux/ip.h>
#include <linux/delay.h>

/* IP_MF appears to be only defined in <netinet/ip.h>, however,
   we need it for hardware checksumming support.  FYI... some of
   the definitions in <netinet/ip.h> conflict/duplicate those in
   other linux headers causing many compiler warnings.
*/
#ifndef IP_MF
  #define IP_MF 0x2000   /* IP more frags from <netinet/ip.h> */ 
#endif

/* Define IP_OFFSET to be IPOPT_OFFSET */
#ifndef IP_OFFSET
  #ifdef IPOPT_OFFSET
    #define IP_OFFSET IPOPT_OFFSET
  #else
    #define IP_OFFSET 2
  #endif
#endif

#define RUN_AT(x) (jiffies + (x))

/* Condensed bus+endian portability operations. */
#if ADDRLEN == 64
#define virt_to_desc(addr)  cpu_to_le64(virt_to_bus(addr))
#else 
#define virt_to_desc(addr)  cpu_to_le32(virt_to_bus(addr))
#define le32desc_to_virt(addr)  bus_to_virt(le32_to_cpu(addr))
#endif   


/*
				Theory of Operation

I. Board Compatibility

This device driver is designed for the Packet Engines "Hamachi"
Gigabit Ethernet chip.  The only PCA currently supported is the GNIC-II 64-bit
66Mhz PCI card.

II. Board-specific settings

No jumpers exist on the board.  The chip supports software correction of
various motherboard wiring errors, however this driver does not support
that feature.

III. Driver operation

IIIa. Ring buffers

The Hamachi uses a typical descriptor based bus-master architecture.
The descriptor list is similar to that used by the Digital Tulip.
This driver uses two statically allocated fixed-size descriptor lists
formed into rings by a branch from the final descriptor to the beginning of
the list.  The ring sizes are set at compile time by RX/TX_RING_SIZE.

This driver uses a zero-copy receive and transmit scheme similar my other
network drivers.
The driver allocates full frame size skbuffs for the Rx ring buffers at
open() time and passes the skb->data field to the Hamachi 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.  Gigabit cards are typically used on generously configured machines
and the underfilled buffers have negligible impact compared to the benefit of
a single allocation size, so the default value of zero results in never
copying packets.

IIIb/c. Transmit/Receive Structure

The Rx and Tx descriptor structure are straight-forward, with no historical
baggage that must be explained.  Unlike the awkward DBDMA structure, there
are no unused fields or option bits that had only one allowable setting.

Two details should be noted about the descriptors: The chip supports both 32
bit and 64 bit address structures, and the length field is overwritten on
the receive descriptors.  The descriptor length is set in the control word
for each channel. The development driver uses 32 bit addresses only, however
64 bit addresses may be enabled for 64 bit architectures e.g. the Alpha.

IIId. Synchronization

This driver is very similar to my other network drivers.
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 'hmp->tx_full' flag.

The interrupt handler has exclusive control over the Rx ring and records stats
from the Tx ring.  After reaping the stats, it marks the Tx queue entry as
empty by incrementing the dirty_tx mark. Iff the 'hmp->tx_full' flag is set, it
clears both the tx_full and tbusy flags.

IV. Notes

Thanks to Kim Stearns of Packet Engines for providing a pair of GNIC-II boards.

IVb. References

Hamachi Engineering Design Specification, 5/15/97
(Note: This version was marked "Confidential".)

IVc. Errata

None noted.  

V.  Recent Changes

01/15/1999 EPK  Enlargement of the TX and RX ring sizes.  This appears 
    to help avoid some stall conditions -- this needs further research.

01/15/1999 EPK  Creation of the hamachi_tx function.  This function cleans 
    the Tx ring and is called from hamachi_start_xmit (this used to be
    called from hamachi_interrupt but it tends to delay execution of the
    interrupt handler and thus reduce bandwidth by reducing the latency
    between hamachi_rx()'s).  Notably, some modification has been made so 
    that the cleaning loop checks only to make sure that the DescOwn bit 
    isn't set in the status flag since the card is not required 
    to set the entire flag to zero after processing.

01/15/1999 EPK In the hamachi_start_tx function, the Tx ring full flag is 
    checked before attempting to add a buffer to the ring.  If the ring is full
    an attempt is made to free any dirty buffers and thus find space for
    the new buffer or the function returns non-zero which should case the
    scheduler to reschedule the buffer later.

01/15/1999 EPK Some adjustments were made to the chip intialization.  
    End-to-end flow control should now be fully active and the interrupt 
    algorithm vars have been changed.  These could probably use further tuning.

01/15/1999 EPK Added the max_{rx,tx}_latency options.  These are used to
    set the rx and tx latencies for the Hamachi interrupts. If you're having
    problems with network stalls, try setting these to higher values.
    Valid values are 0x00 through 0xff.

01/15/1999 EPK In general, the overall bandwidth has increased and 
    latencies are better (sometimes by a factor of 2).  Stalls are rare at
    this point, however there still appears to be a bug somewhere between the
    hardware and driver.  TCP checksum errors under load also appear to be
    eliminated at this point.

01/18/1999 EPK Ensured that the DescEndRing bit was being set on both the
    Rx and Tx rings.  This appears to have been affecting whether a particular
    peer-to-peer connection would hang under high load.  I believe the Rx
    rings was typically getting set correctly, but the Tx ring wasn't getting
    the DescEndRing bit set during initialization. ??? Does this mean the
    hamachi card is using the DescEndRing in processing even if a particular
    slot isn't in use -- hypothetically, the card might be searching the 
    entire Tx ring for slots with the DescOwn bit set and then processing
    them.  If the DescEndRing bit isn't set, then it might just wander off
    through memory until it hits a chunk of data with that bit set
    and then looping back.

02/09/1999 EPK Added Michel Mueller's TxDMA Interrupt and Tx-timeout 
    problem (TxCmd and RxCmd need only to be set when idle or stopped.

02/09/1999 EPK Added code to check/reset dev->tbusy in hamachi_interrupt.
    (Michel Mueller pointed out the ``permanently busy'' potential 
    problem here).

02/22/1999 EPK Added Pete Wyckoff's ioctl to control the Tx/Rx latencies. 

02/23/1999 EPK Verified that the interrupt status field bits for Tx were
    incorrectly defined and corrected (as per Michel Mueller).

02/23/1999 EPK Corrected the Tx full check to check that at least 4 slots
    were available before reseting the tbusy and tx_full flags
    (as per Michel Mueller).

03/11/1999 EPK Added Pete Wyckoff's hardware checksumming support.

12/31/1999 KDU Cleaned up assorted things and added Don's code to force
32 bit.

02/20/2000 KDU Some of the control was just plain odd.  Cleaned up the
hamachi_start_xmit() and hamachi_interrupt() code.  There is still some
re-structuring I would like to do.  

03/01/2000 KDU Experimenting with a WIDE range of interrupt mitigation
parameters on a dual P3-450 setup yielded the new default interrupt
mitigation parameters.  Tx should interrupt VERY infrequently due to
Eric's scheme.  Rx should be more often...

03/13/2000 KDU Added a patch to make the Rx Checksum code interact
nicely with non-linux machines.  

03/13/2000 KDU Experimented with some of the configuration values:  

	-It seems that enabling PCI performance commands for descriptors
	(changing RxDMACtrl and TxDMACtrl lower nibble from 5 to D) has minimal 
	performance impact for any of my tests. (ttcp, netpipe, netperf)  I will 
	leave them that way until I hear further feedback.

	-Increasing the PCI_LATENCY_TIMER to 130 
	(2 + (burst size of 128 * (0 wait states + 1))) seems to slightly
	degrade performance.  Leaving default at 64 pending further information.