53c78xx.c

GNU Mach 微内核源代码, 基于美国卡内基美隆大学的 Mach 研究项目

/*
 * PERM_OPTIONS are driver options which will be enabled for all NCR boards
 * in the system at driver initialization time.
 *
 * Don't THINK about touching these in PERM_OPTIONS : 
 *   OPTION_IO_MAPPED 
 * 	Memory mapped IO does not work under i86 Linux. 
 *
 *   OPTION_DEBUG_TEST1
 *	Test 1 does bus mastering and interrupt tests, which will help weed 
 *	out brain damaged main boards.
 *
 * These are development kernel changes.  Code for them included in this
 * driver release may or may not work.  If you turn them on, you should be 
 * running the latest copy of the development sources from
 *
 *	ftp://tsx-11.mit.edu/pub/linux/ALPHA/scsi/53c7,8xx
 *
 * and be subscribed to the [e-mail] mailing list.  To
 * subscribe, send mail to [e-mail] with 
 *
 * 	subscribe ncr53c810
 * 
 * in the text.
 *
 *
 *   OPTION_NO_ASYNC
 *	Don't negotiate for asynchronous transfers on the first command 
 *	when OPTION_ALWAYS_SYNCHRONOUS is set.  Useful for dain bramaged
 *	devices which do something bad rather than sending a MESSAGE 
 *	REJECT back to us like they should if they can't cope.
 *
 *   OPTION_SYNCHRONOUS
 *	Enable support for synchronous transfers.  Target negotiated 
 *	synchronous transfers will be responded to.  To initiate 
 *	a synchronous transfer request,  call 
 *
 *	    request_synchronous (hostno, target) 
 *
 *	from within KGDB.
 *
 *   OPTION_ALWAYS_SYNCHRONOUS
 *	Negotiate for synchronous transfers with every target after
 *	driver initialization or a SCSI bus reset.  This is a bit dangerous, 
 *	since there are some dain bramaged SCSI devices which will accept
 *	SDTR messages but keep talking asynchronously.
 *
 *   OPTION_DISCONNECT
 *	Enable support for disconnect/reconnect.  To change the 
 *	default setting on a given host adapter, call
 *
 *	    request_disconnect (hostno, allow)
 *
 *	where allow is non-zero to allow, 0 to disallow.
 * 
 *  If you really want to run 10MHz FAST SCSI-II transfers, you should 
 *  know that the NCR driver currently ignores parity information.  Most
 *  systems do 5MHz SCSI fine.  I've seen a lot that have problems faster
 *  than 8MHz.  To play it safe, we only request 5MHz transfers.
 *
 *  If you'd rather get 10MHz transfers, edit sdtr_message and change 
 *  the fourth byte from 50 to 25.
 */

#include <linux/config.h>

#ifdef CONFIG_SCSI_NCR53C7xx_sync
#ifdef CONFIG_SCSI_NCR53C7xx_DISCONNECT
#define PERM_OPTIONS (OPTION_IO_MAPPED|OPTION_DEBUG_TEST1|OPTION_DISCONNECT|\
	OPTION_SYNCHRONOUS|OPTION_ALWAYS_SYNCHRONOUS)
#else
#define PERM_OPTIONS (OPTION_IO_MAPPED|OPTION_DEBUG_TEST1|\
	OPTION_SYNCHRONOUS|OPTION_ALWAYS_SYNCHRONOUS)
#endif
#else
#ifdef CONFIG_SCSI_NCR53C7xx_DISCONNECT
#define PERM_OPTIONS (OPTION_IO_MAPPED|OPTION_DEBUG_TEST1|OPTION_DISCONNECT|\
	OPTION_SYNCHRONOUS)
#else
#define PERM_OPTIONS (OPTION_IO_MAPPED|OPTION_DEBUG_TEST1|OPTION_SYNCHRONOUS)
#endif
#endif

/*
 * Sponsored by 
 *	iX Multiuser Multitasking Magazine
 *	Hannover, Germany
 *	[e-mail]
 *
 * Copyright 1993, 1994, 1995 Drew Eckhardt
 *      Visionary Computing 
 *      (Unix and Linux consulting and custom programming)
 *      [e-mail]
 *	[phone]
 *
 * TolerANT and SCSI SCRIPTS are registered trademarks of NCR Corporation.
 * 
 * For more information, please consult 
 *
 * NCR53C810 
 * SCSI I/O Processor
 * Programmer's Guide
 *
 * NCR 53C810
 * PCI-SCSI I/O Processor
 * Data Manual
 *
 * NCR 53C810/53C820
 * PCI-SCSI I/O Processor Design In Guide
 *
 * For literature on Symbios Logic Inc. formerly NCR, SCSI, 
 * and Communication products please call [phone] or
 * [phone]. 
 * 
 * PCI BIOS Specification Revision
 * PCI Local Bus Specification
 * PCI System Design Guide
 *
 * PCI Special Interest Group
 * M/S HF3-15A
 * 5200 N.E. Elam Young Parkway
 * Hillsboro, Oregon 97124-6497
 * [phone] 
 * [phone]
 */

/*
 * Design issues : 
 * The cumulative latency needed to propagate a read/write request 
 * through the file system, buffer cache, driver stacks, SCSI host, and 
 * SCSI device is ultimately the limiting factor in throughput once we 
 * have a sufficiently fast host adapter.
 *  
 * So, to maximize performance we want to keep the ratio of latency to data 
 * transfer time to a minimum by
 * 1.  Minimizing the total number of commands sent (typical command latency
 *	including drive and bus mastering host overhead is as high as 4.5ms)
 *	to transfer a given amount of data.  
 *
 *      This is accomplished by placing no arbitrary limit on the number
 *	of scatter/gather buffers supported, since we can transfer 1K
 *	per scatter/gather buffer without Eric's cluster patches, 
 *	4K with.  
 *
 * 2.  Minimizing the number of fatal interrupts serviced, since
 * 	fatal interrupts halt the SCSI I/O processor.  Basically,
 *	this means offloading the practical maximum amount of processing 
 *	to the SCSI chip.
 * 
 *	On the NCR53c810/820/720,  this is accomplished by using 
 *		interrupt-on-the-fly signals when commands complete, 
 *		and only handling fatal errors and SDTR / WDTR 	messages 
 *		in the host code.
 *
 *	On the NCR53c710, interrupts are generated as on the NCR53c8x0,
 *		only the lack of a interrupt-on-the-fly facility complicates
 *		things.   Also, SCSI ID registers and commands are 
 *		bit fielded rather than binary encoded.
 *		
 * 	On the NCR53c700 and NCR53c700-66, operations that are done via 
 *		indirect, table mode on the more advanced chips must be
 *	        replaced by calls through a jump table which 
 *		acts as a surrogate for the DSA.  Unfortunately, this 
 * 		will mean that we must service an interrupt for each 
 *		disconnect/reconnect.
 * 
 * 3.  Eliminating latency by pipelining operations at the different levels.
 * 	
 *	This driver allows a configurable number of commands to be enqueued
 *	for each target/lun combination (experimentally, I have discovered
 *	that two seems to work best) and will ultimately allow for 
 *	SCSI-II tagged queuing.
 * 	
 *
 * Architecture : 
 * This driver is built around a Linux queue of commands waiting to 
 * be executed, and a shared Linux/NCR array of commands to start.  Commands
 * are transfered to the array  by the run_process_issue_queue() function 
 * which is called whenever a command completes.
 *
 * As commands are completed, the interrupt routine is triggered,
 * looks for commands in the linked list of completed commands with
 * valid status, removes these commands from a list of running commands, 
 * calls the done routine, and flags their target/luns as not busy.
 *
 * Due to limitations in the intelligence of the NCR chips, certain
 * concessions are made.  In many cases, it is easier to dynamically 
 * generate/fix-up code rather than calculate on the NCR at run time.  
 * So, code is generated or fixed up for
 *
 * - Handling data transfers, using a variable number of MOVE instructions
 *	interspersed with CALL MSG_IN, WHEN MSGIN instructions.
 *
 * 	The DATAIN and DATAOUT routines	are separate, so that an incorrect
 *	direction can be trapped, and space isn't wasted. 
 *
 *	It may turn out that we're better off using some sort 
 *	of table indirect instruction in a loop with a variable
 *	sized table on the NCR53c710 and newer chips.
 *
 * - Checking for reselection (NCR53c710 and better)
 *
 * - Handling the details of SCSI context switches (NCR53c710 and better),
 *	such as reprogramming appropriate synchronous parameters, 
 *	removing the dsa structure from the NCR's queue of outstanding
 *	commands, etc.
 *
 */

/* 
 * Accommodate differences between stock 1.2.x and 1.3.x asm-i386/types.h
 * so lusers can drop in 53c7,8xx.* and get something which compiles 
 * without warnings.
 */

#if !defined(LINUX_1_2) && !defined(LINUX_1_3)
#include <linux/version.h>
#if LINUX_VERSION_CODE > 65536 + 3 * 256
#define LINUX_1_3
#else
#define LINUX_1_2
#endif
#endif

#ifdef LINUX_1_2
#define u32 bogus_u32
#define s32 bogus_s32
#include <asm/types.h>
#undef u32
#undef s32
typedef __signed__ int s32;
typedef unsigned int  u32;
#endif /* def LINUX_1_2 */

#ifdef MODULE
#include <linux/module.h>
#endif

#include <asm/dma.h>
#include <asm/io.h>
#include <asm/system.h>
#include <linux/delay.h>
#include <linux/signal.h>
#include <linux/sched.h>
#include <linux/errno.h>
#include <linux/bios32.h>
#include <linux/pci.h>
#include <linux/proc_fs.h>
#include <linux/string.h>
#include <linux/malloc.h>
#include <linux/mm.h>
#include <linux/ioport.h>
#include <linux/time.h>
#ifdef LINUX_1_2
#include "../block/blk.h"
#else
#include <linux/blk.h>
#endif
#undef current

#include "scsi.h"
#include "hosts.h"
#include "53c7,8xx.h"
#include "constants.h"
#include "sd.h"
#include <linux/stat.h>
#include <linux/stddef.h>

#ifndef LINUX_1_2
struct proc_dir_entry proc_scsi_ncr53c7xx = {
    PROC_SCSI_NCR53C7xx, 9, "ncr53c7xx",
    S_IFDIR | S_IRUGO | S_IXUGO, 2
};
#endif

static int check_address (unsigned long addr, int size);
static void dump_events (struct Scsi_Host *host, int count);
static Scsi_Cmnd * return_outstanding_commands (struct Scsi_Host *host, 
    int free, int issue);
static void hard_reset (struct Scsi_Host *host);
static void ncr_scsi_reset (struct Scsi_Host *host);
static void print_lots (struct Scsi_Host *host);
static void set_synchronous (struct Scsi_Host *host, int target, int sxfer, 
    int scntl3, int now_connected);
static int datapath_residual (struct Scsi_Host *host);
static const char * sbcl_to_phase (int sbcl);
static void print_progress (Scsi_Cmnd *cmd);
static void print_queues (struct Scsi_Host *host);
static void process_issue_queue (unsigned long flags);
static int shutdown (struct Scsi_Host *host);
static void abnormal_finished (struct NCR53c7x0_cmd *cmd, int result);
static int disable (struct Scsi_Host *host);
static int NCR53c8xx_run_tests (struct Scsi_Host *host);
static int NCR53c8xx_script_len;
static int NCR53c8xx_dsa_len;
static void NCR53c7x0_intr(int irq, void *dev_id, struct pt_regs * regs);
static int ncr_halt (struct Scsi_Host *host);
static void intr_phase_mismatch (struct Scsi_Host *host, struct NCR53c7x0_cmd 
    *cmd);
static void intr_dma (struct Scsi_Host *host, struct NCR53c7x0_cmd *cmd);
static void print_dsa (struct Scsi_Host *host, u32 *dsa,
    const char *prefix);
static int print_insn (struct Scsi_Host *host, const u32 *insn,
    const char *prefix, int kernel);

static void NCR53c8xx_dsa_fixup (struct NCR53c7x0_cmd *cmd);
static void NCR53c8x0_init_fixup (struct Scsi_Host *host);
static int NCR53c8x0_dstat_sir_intr (struct Scsi_Host *host, struct 
    NCR53c7x0_cmd *cmd);
static void NCR53c8x0_soft_reset (struct Scsi_Host *host);

/* INSMOD variables */
static long long perm_options = PERM_OPTIONS;
/* 14 = .5s; 15 is max; decreasing divides by two. */
static int selection_timeout = 14;
/* Size of event list (per host adapter) */
static int track_events = 0;

static struct Scsi_Host *first_host = NULL;	/* Head of list of NCR boards */
static Scsi_Host_Template *the_template = NULL;	

/*
 * KNOWN BUGS :
 * - There is some sort of conflict when the PPP driver is compiled with 
 * 	support for 16 channels?
 * 
 * - On systems which predate the 1.3.x initialization order change,
 *      the NCR driver will cause Cannot get free page messages to appear.  
 *      These are harmless, but I don't know of an easy way to avoid them.
 *
 * - With OPTION_DISCONNECT, on two systems under unknown circumstances,
 *	we get a PHASE MISMATCH with DSA set to zero (suggests that we 
 *	are occurring somewhere in the reselection code) where 
 *	DSP=some value DCMD|DBC=same value.  
 * 	
 *	Closer inspection suggests that we may be trying to execute
 *	some portion of the DSA?
 * scsi0 : handling residual transfer (+ 0 bytes from DMA FIFO)
 * scsi0 : handling residual transfer (+ 0 bytes from DMA FIFO)
 * scsi0 : no current command : unexpected phase MSGIN.
 *         DSP=0x1c46cc, DCMD|DBC=0x1c46ac, DSA=0x0
 *         DSPS=0x0, TEMP=0x1c3e70, DMODE=0x80
 * scsi0 : DSP->
 * 001c46cc : 0x001c46cc 0x00000000
 * 001c46d4 : 0x001c5ea0 0x000011f8
 *
 *	Changed the print code in the phase_mismatch handler so
 *	that we call print_lots to try to diagnose this.
 *
 */

/* 
 * Possible future direction of architecture for max performance :
 *
 * We're using a single start array for the NCR chip.  This is 
 * sub-optimal, because we cannot add a command which would conflict with 
 * an executing command to this start queue, and therefore must insert the 
 * next command for a given I/T/L combination after the first has completed;
 * incurring our interrupt latency between SCSI commands.
 *
 * To allow further pipelining of the NCR and host CPU operation, we want 
 * to set things up so that immediately on termination of a command destined 
 * for a given LUN, we get that LUN busy again.  
 * 
 * To do this, we need to add a 32 bit pointer to which is jumped to 
 * on completion of a command.  If no new command is available, this 
 * would point to the usual DSA issue queue select routine.
 *
 * If one were, it would point to a per-NCR53c7x0_cmd select routine 
 * which starts execution immediately, inserting the command at the head 
 * of the start queue if the NCR chip is selected or reselected.
 *
 * We would change so that we keep a list of outstanding commands 
 * for each unit, rather than a single running_list.  We'd insert 
 * a new command into the right running list; if the NCR didn't 
 * have something running for that yet, we'd put it in the 
 * start queue as well.  Some magic needs to happen to handle the 
 * race condition between the first command terminating before the 
 * new one is written.
 *
 * Potential for profiling : 
 * Call do_gettimeofday(struct timeval *tv) to get 800ns resolution.
 */


/*
 * TODO : 
 * 1.  To support WIDE transfers, not much needs to happen.  We
 *	should do CHMOVE instructions instead of MOVEs when
 *	we have scatter/gather segments of uneven length.  When