scsi.c

讲述linux的初始化过程

 *              a consumable resource, as these are allocated into a pool
 *              when the SCSI subsystem initializes.  The preallocation is
 *              required so that in low-memory situations a disk I/O request
 *              won't cause the memory manager to try and write out a page.
 *              The request structure is generally used by ioctls and character
 *              devices.
 */
void scsi_init_cmd_from_req(Scsi_Cmnd * SCpnt, Scsi_Request * SRpnt)
{
	struct Scsi_Host *host = SCpnt->host;

	ASSERT_LOCK(&io_request_lock, 0);

	SCpnt->owner = SCSI_OWNER_MIDLEVEL;
	SRpnt->sr_command = SCpnt;

	if (!host) {
		panic("Invalid or not present host.\n");
	}

	SCpnt->cmd_len = SRpnt->sr_cmd_len;
	SCpnt->use_sg = SRpnt->sr_use_sg;

	memcpy((void *) &SCpnt->request, (const void *) &SRpnt->sr_request,
	       sizeof(SRpnt->sr_request));
	memcpy((void *) SCpnt->data_cmnd, (const void *) SRpnt->sr_cmnd, 
	       sizeof(SCpnt->data_cmnd));
	SCpnt->reset_chain = NULL;
	SCpnt->serial_number = 0;
	SCpnt->serial_number_at_timeout = 0;
	SCpnt->bufflen = SRpnt->sr_bufflen;
	SCpnt->buffer = SRpnt->sr_buffer;
	SCpnt->flags = 0;
	SCpnt->retries = 0;
	SCpnt->allowed = SRpnt->sr_allowed;
	SCpnt->done = SRpnt->sr_done;
	SCpnt->timeout_per_command = SRpnt->sr_timeout_per_command;

	SCpnt->sc_data_direction = SRpnt->sr_data_direction;

	SCpnt->sglist_len = SRpnt->sr_sglist_len;
	SCpnt->underflow = SRpnt->sr_underflow;

	SCpnt->sc_request = SRpnt;

	memcpy((void *) SCpnt->cmnd, (const void *) SRpnt->sr_cmnd, 
	       sizeof(SCpnt->cmnd));
	/* Zero the sense buffer.  Some host adapters automatically request
	 * sense on error.  0 is not a valid sense code.
	 */
	memset((void *) SCpnt->sense_buffer, 0, sizeof SCpnt->sense_buffer);
	SCpnt->request_buffer = SRpnt->sr_buffer;
	SCpnt->request_bufflen = SRpnt->sr_bufflen;
	SCpnt->old_use_sg = SCpnt->use_sg;
	if (SCpnt->cmd_len == 0)
		SCpnt->cmd_len = COMMAND_SIZE(SCpnt->cmnd[0]);
	SCpnt->old_cmd_len = SCpnt->cmd_len;
	SCpnt->sc_old_data_direction = SCpnt->sc_data_direction;
	SCpnt->old_underflow = SCpnt->underflow;

	/* Start the timer ticking.  */

	SCpnt->internal_timeout = NORMAL_TIMEOUT;
	SCpnt->abort_reason = 0;
	SCpnt->result = 0;

	SCSI_LOG_MLQUEUE(3, printk("Leaving scsi_do_cmd()\n"));
}

/*
 * Function:    scsi_do_cmd
 *
 * Purpose:     Queue a SCSI command
 *
 * Arguments:   SCpnt     - command descriptor.
 *              cmnd      - actual SCSI command to be performed.
 *              buffer    - data buffer.
 *              bufflen   - size of data buffer.
 *              done      - completion function to be run.
 *              timeout   - how long to let it run before timeout.
 *              retries   - number of retries we allow.
 *
 * Lock status: With the new queueing code, this is SMP-safe, and no locks
 *              need be held upon entry.   The old queueing code the lock was
 *              assumed to be held upon entry.
 *
 * Returns:     Nothing.
 *
 * Notes:       Prior to the new queue code, this function was not SMP-safe.
 *              Also, this function is now only used for queueing requests
 *              for things like ioctls and character device requests - this
 *              is because we essentially just inject a request into the
 *              queue for the device. Normal block device handling manipulates
 *              the queue directly.
 */
void scsi_do_cmd(Scsi_Cmnd * SCpnt, const void *cmnd,
	      void *buffer, unsigned bufflen, void (*done) (Scsi_Cmnd *),
		 int timeout, int retries)
{
	struct Scsi_Host *host = SCpnt->host;

	ASSERT_LOCK(&io_request_lock, 0);

	SCpnt->owner = SCSI_OWNER_MIDLEVEL;

	SCSI_LOG_MLQUEUE(4,
			 {
			 int i;
			 int target = SCpnt->target;
			 printk("scsi_do_cmd (host = %d, channel = %d target = %d, "
		    "buffer =%p, bufflen = %d, done = %p, timeout = %d, "
				"retries = %d)\n"
				"command : ", host->host_no, SCpnt->channel, target, buffer,
				bufflen, done, timeout, retries);
			 for (i = 0; i < 10; ++i)
			 printk("%02x  ", ((unsigned char *) cmnd)[i]);
			 printk("\n");
			 });

	if (!host) {
		panic("Invalid or not present host.\n");
	}
	/*
	 * We must prevent reentrancy to the lowlevel host driver.  This prevents
	 * it - we enter a loop until the host we want to talk to is not busy.
	 * Race conditions are prevented, as interrupts are disabled in between the
	 * time we check for the host being not busy, and the time we mark it busy
	 * ourselves.
	 */


	/*
	 * Our own function scsi_done (which marks the host as not busy, disables
	 * the timeout counter, etc) will be called by us or by the
	 * scsi_hosts[host].queuecommand() function needs to also call
	 * the completion function for the high level driver.
	 */

	memcpy((void *) SCpnt->data_cmnd, (const void *) cmnd, 
               sizeof(SCpnt->data_cmnd));
	SCpnt->reset_chain = NULL;
	SCpnt->serial_number = 0;
	SCpnt->serial_number_at_timeout = 0;
	SCpnt->bufflen = bufflen;
	SCpnt->buffer = buffer;
	SCpnt->flags = 0;
	SCpnt->retries = 0;
	SCpnt->allowed = retries;
	SCpnt->done = done;
	SCpnt->timeout_per_command = timeout;

	memcpy((void *) SCpnt->cmnd, (const void *) cmnd, 
               sizeof(SCpnt->cmnd));
	/* Zero the sense buffer.  Some host adapters automatically request
	 * sense on error.  0 is not a valid sense code.
	 */
	memset((void *) SCpnt->sense_buffer, 0, sizeof SCpnt->sense_buffer);
	SCpnt->request_buffer = buffer;
	SCpnt->request_bufflen = bufflen;
	SCpnt->old_use_sg = SCpnt->use_sg;
	if (SCpnt->cmd_len == 0)
		SCpnt->cmd_len = COMMAND_SIZE(SCpnt->cmnd[0]);
	SCpnt->old_cmd_len = SCpnt->cmd_len;
	SCpnt->sc_old_data_direction = SCpnt->sc_data_direction;
	SCpnt->old_underflow = SCpnt->underflow;

	/* Start the timer ticking.  */

	SCpnt->internal_timeout = NORMAL_TIMEOUT;
	SCpnt->abort_reason = 0;
	SCpnt->result = 0;

	/*
	 * At this point, we merely set up the command, stick it in the normal
	 * request queue, and return.  Eventually that request will come to the
	 * top of the list, and will be dispatched.
	 */
	scsi_insert_special_cmd(SCpnt, 0);

	SCSI_LOG_MLQUEUE(3, printk("Leaving scsi_do_cmd()\n"));
}

/*
 * This function is the mid-level interrupt routine, which decides how
 *  to handle error conditions.  Each invocation of this function must
 *  do one and *only* one of the following:
 *
 *      1) Insert command in BH queue.
 *      2) Activate error handler for host.
 *
 * FIXME(eric) - I am concerned about stack overflow (still).  An
 * interrupt could come while we are processing the bottom queue,
 * which would cause another command to be stuffed onto the bottom
 * queue, and it would in turn be processed as that interrupt handler
 * is returning.  Given a sufficiently steady rate of returning
 * commands, this could cause the stack to overflow.  I am not sure
 * what is the most appropriate solution here - we should probably
 * keep a depth count, and not process any commands while we still
 * have a bottom handler active higher in the stack.
 *
 * There is currently code in the bottom half handler to monitor
 * recursion in the bottom handler and report if it ever happens.  If
 * this becomes a problem, it won't be hard to engineer something to
 * deal with it so that only the outer layer ever does any real
 * processing.  
 */
void scsi_done(Scsi_Cmnd * SCpnt)
{
	unsigned long flags;
	int tstatus;

	/*
	 * We don't have to worry about this one timing out any more.
	 */
	tstatus = scsi_delete_timer(SCpnt);

	/*
	 * If we are unable to remove the timer, it means that the command
	 * has already timed out.  In this case, we have no choice but to
	 * let the timeout function run, as we have no idea where in fact
	 * that function could really be.  It might be on another processor,
	 * etc, etc.
	 */
	if (!tstatus) {
		SCpnt->done_late = 1;
		return;
	}
	/* Set the serial numbers back to zero */
	SCpnt->serial_number = 0;

	/*
	 * First, see whether this command already timed out.  If so, we ignore
	 * the response.  We treat it as if the command never finished.
	 *
	 * Since serial_number is now 0, the error handler cound detect this
	 * situation and avoid to call the the low level driver abort routine.
	 * (DB)
         *
         * FIXME(eric) - I believe that this test is now redundant, due to
         * the test of the return status of del_timer().
	 */
	if (SCpnt->state == SCSI_STATE_TIMEOUT) {
		SCSI_LOG_MLCOMPLETE(1, printk("Ignoring completion of %p due to timeout status", SCpnt));
		return;
	}
	spin_lock_irqsave(&scsi_bhqueue_lock, flags);

	SCpnt->serial_number_at_timeout = 0;
	SCpnt->state = SCSI_STATE_BHQUEUE;
	SCpnt->owner = SCSI_OWNER_BH_HANDLER;
	SCpnt->bh_next = NULL;

	/*
	 * Next, put this command in the BH queue.
	 * 
	 * We need a spinlock here, or compare and exchange if we can reorder incoming
	 * Scsi_Cmnds, as it happens pretty often scsi_done is called multiple times
	 * before bh is serviced. -jj
	 *
	 * We already have the io_request_lock here, since we are called from the
	 * interrupt handler or the error handler. (DB)
	 *
	 * This may be true at the moment, but I would like to wean all of the low
	 * level drivers away from using io_request_lock.   Technically they should
	 * all use their own locking.  I am adding a small spinlock to protect
	 * this datastructure to make it safe for that day.  (ERY)
	 */
	if (!scsi_bh_queue_head) {
		scsi_bh_queue_head = SCpnt;
		scsi_bh_queue_tail = SCpnt;
	} else {
		scsi_bh_queue_tail->bh_next = SCpnt;
		scsi_bh_queue_tail = SCpnt;
	}

	spin_unlock_irqrestore(&scsi_bhqueue_lock, flags);
	/*
	 * Mark the bottom half handler to be run.
	 */
	mark_bh(SCSI_BH);
}

/*
 * Procedure:   scsi_bottom_half_handler
 *
 * Purpose:     Called after we have finished processing interrupts, it
 *              performs post-interrupt handling for commands that may
 *              have completed.
 *
 * Notes:       This is called with all interrupts enabled.  This should reduce
 *              interrupt latency, stack depth, and reentrancy of the low-level
 *              drivers.
 *
 * The io_request_lock is required in all the routine. There was a subtle
 * race condition when scsi_done is called after a command has already
 * timed out but before the time out is processed by the error handler.
 * (DB)
 *
 * I believe I have corrected this.  We simply monitor the return status of
 * del_timer() - if this comes back as 0, it means that the timer has fired
 * and that a timeout is in progress.   I have modified scsi_done() such
 * that in this instance the command is never inserted in the bottom
 * half queue.  Thus the only time we hold the lock here is when
 * we wish to atomically remove the contents of the queue.
 */
void scsi_bottom_half_handler(void)
{
	Scsi_Cmnd *SCpnt;
	Scsi_Cmnd *SCnext;
	unsigned long flags;


	while (1 == 1) {
		spin_lock_irqsave(&scsi_bhqueue_lock, flags);
		SCpnt = scsi_bh_queue_head;
		scsi_bh_queue_head = NULL;
		spin_unlock_irqrestore(&scsi_bhqueue_lock, flags);

		if (SCpnt == NULL) {
			return;
		}
		SCnext = SCpnt->bh_next;

		for (; SCpnt; SCpnt = SCnext) {
			SCnext = SCpnt->bh_next;

			switch (scsi_decide_disposition(SCpnt)) {
			case SUCCESS:
				/*
				 * Add to BH queue.
				 */
				SCSI_LOG_MLCOMPLETE(3, printk("Command finished %d %d 0x%x\n", SCpnt->host->host_busy,
						SCpnt->host->host_failed,
							 SCpnt->result));

				scsi_finish_command(SCpnt);
				break;
			case NEEDS_RETRY:
				/*
				 * We only come in here if we want to retry a command.  The
				 * test to see whether the command should be retried should be
				 * keeping track of the number of tries, so we don't end up looping,
				 * of course.
				 */
				SCSI_LOG_MLCOMPLETE(3, printk("Command needs retry %d %d 0x%x\n", SCpnt->host->host_busy,
				SCpnt->host->host_failed, SCpnt->result));

				scsi_retry_command(SCpnt);
				break;
			case ADD_TO_MLQUEUE:
				/* 
				 * This typically happens for a QUEUE_FULL message -
				 * typically only when the queue depth is only
				 * approximate for a given device.  Adding a command
				 * to the queue for the device will prevent further commands
				 * from being sent to the device, so we shouldn't end up
				 * with tons of things being sent down that shouldn't be.
				 */
				SCSI_LOG_MLCOMPLETE(3, printk("Command rejected as device queue full, put on ml queue %p\n",
                                                              SCpnt));
				scsi_mlqueue_insert(SCpnt, SCSI_MLQUEUE_DEVICE_BUSY);
				break;
			default:
				/*
				 * Here we have a fatal error of some sort.  Turn it over to
				 * the error handler.
				 */
				SCSI_LOG_MLCOMPLETE(3, printk("Command failed %p %x active=%d busy=%d failed=%d\n",
						    SCpnt, SCpnt->result,
				  atomic_read(&SCpnt->host->host_active),
						  SCpnt->host->host_busy,
					      SCpnt->host->host_failed));

				/*
				 * Dump the sense information too.
				 */
				if ((status_byte(SCpnt->result) & CHECK_CONDITION) != 0) {
					SCSI_LOG_MLCOMPLETE(3, print_sense("bh", SCpnt));
				}
				if (SCpnt->host->eh_wait != NULL) {
					SCpnt->host->host_failed++;
					SCpnt->owner = SCSI_OWNER_ERROR_HANDLER;
					SCpnt->state = SCSI_STATE_FAILED;
					SCpnt->host->in_recovery = 1;
					/*
					 * If the host is having troubles, then look to see if this was the last
					 * command that might have failed.  If so, wake up the error handler.
					 */
					if (SCpnt->host->host_busy == SCpnt->host->host_failed) {
						SCSI_LOG_ERROR_RECOVERY(5, printk("Waking error handler thread (%d)\n",
										  atomic_read(&SCpnt->host->eh_wait->count)));
						up(SCpnt->host->eh_wait);
					}
				} else {
					/*
					 * We only get here if the error recovery thread has died.
					 */
					scsi_finish_command(SCpnt);
				}
			}
		}		/* for(; SCpnt...) */

	}			/* while(1==1) */

}

/*
 * Function:    scsi_retry_command
 *
 * Purpose:     Send a command back to the low level to be retried.
 *
 * Notes:       This command is always executed in the context of the
 *              bottom half handler, or the error handler thread. Low
 *              level drivers should not become re-entrant as a result of
 *              this.
 */
int scsi_retry_command(Scsi_Cmnd * SCpnt)
{
	memcpy((void *) SCpnt->cmnd, (void *) SCpnt->data_cmnd,
	       sizeof(SCpnt->data_cmnd));
	SCpnt->request_buffer = SCpnt->buffer;
	SCpnt->request_bufflen = SCpnt->bufflen;
	SCpnt->use_sg = SCpnt->old_use_sg;
	SCpnt->cmd_len = SCpnt->old_cmd_len;
	SCpnt->sc_data_direction = SCpnt->sc_old_data_direction;
	SCpnt->underflow = SCpnt->old_underflow;

        /*
         * Zero the sense information from the last time we tried
         * this command.
         */
	memset((void *) SCpnt->sense_buffer, 0, sizeof SCpnt->sense_buffer);

	return scsi_dispatch_cmd(SCpnt);