raid1.c

Linux内核源代码 为压缩文件 是<<Linux内核>>一书中的源代码

		close_sync(conf);
		sdisk = conf->mirrors + spare_disk;
		sdisk->operational = 0;
		sdisk->write_only = 0;
		break;
	/*
	 * Activate (mark read-write) the (now sync) spare disk,
	 * which means we switch it's 'raid position' (->raid_disk)
	 * with the failed disk. (only the first 'conf->nr_disks'
	 * slots are used for 'real' disks and we must preserve this
	 * property)
	 */
	case DISKOP_SPARE_ACTIVE:
		close_sync(conf);
		sdisk = conf->mirrors + spare_disk;
		fdisk = conf->mirrors + failed_disk;

		spare_desc = &sb->disks[sdisk->number];
		failed_desc = &sb->disks[fdisk->number];

		if (spare_desc != *d) {
			MD_BUG();
			err = 1;
			goto abort;
		}

		if (spare_desc->raid_disk != sdisk->raid_disk) {
			MD_BUG();
			err = 1;
			goto abort;
		}
			
		if (sdisk->raid_disk != spare_disk) {
			MD_BUG();
			err = 1;
			goto abort;
		}

		if (failed_desc->raid_disk != fdisk->raid_disk) {
			MD_BUG();
			err = 1;
			goto abort;
		}

		if (fdisk->raid_disk != failed_disk) {
			MD_BUG();
			err = 1;
			goto abort;
		}

		/*
		 * do the switch finally
		 */
		xchg_values(*spare_desc, *failed_desc);
		xchg_values(*fdisk, *sdisk);

		/*
		 * (careful, 'failed' and 'spare' are switched from now on)
		 *
		 * we want to preserve linear numbering and we want to
		 * give the proper raid_disk number to the now activated
		 * disk. (this means we switch back these values)
		 */
	
		xchg_values(spare_desc->raid_disk, failed_desc->raid_disk);
		xchg_values(sdisk->raid_disk, fdisk->raid_disk);
		xchg_values(spare_desc->number, failed_desc->number);
		xchg_values(sdisk->number, fdisk->number);

		*d = failed_desc;

		if (sdisk->dev == MKDEV(0,0))
			sdisk->used_slot = 0;
		/*
		 * this really activates the spare.
		 */
		fdisk->spare = 0;
		fdisk->write_only = 0;

		/*
		 * if we activate a spare, we definitely replace a
		 * non-operational disk slot in the 'low' area of
		 * the disk array.
		 */

		conf->working_disks++;

		break;

	case DISKOP_HOT_REMOVE_DISK:
		rdisk = conf->mirrors + removed_disk;

		if (rdisk->spare && (removed_disk < conf->raid_disks)) {
			MD_BUG();	
			err = 1;
			goto abort;
		}
		rdisk->dev = MKDEV(0,0);
		rdisk->used_slot = 0;
		conf->nr_disks--;
		break;

	case DISKOP_HOT_ADD_DISK:
		adisk = conf->mirrors + added_disk;
		added_desc = *d;

		if (added_disk != added_desc->number) {
			MD_BUG();	
			err = 1;
			goto abort;
		}

		adisk->number = added_desc->number;
		adisk->raid_disk = added_desc->raid_disk;
		adisk->dev = MKDEV(added_desc->major,added_desc->minor);

		adisk->operational = 0;
		adisk->write_only = 0;
		adisk->spare = 1;
		adisk->used_slot = 1;
		adisk->head_position = 0;
		conf->nr_disks++;

		break;

	default:
		MD_BUG();	
		err = 1;
		goto abort;
	}
abort:
	md_spin_unlock_irq(&conf->device_lock);
	if (state == DISKOP_SPARE_ACTIVE || state == DISKOP_SPARE_INACTIVE)
		/* should move to "END_REBUILD" when such exists */
		raid1_shrink_buffers(conf);

	print_raid1_conf(conf);
	return err;
}


#define IO_ERROR KERN_ALERT \
"raid1: %s: unrecoverable I/O read error for block %lu\n"

#define REDIRECT_SECTOR KERN_ERR \
"raid1: %s: redirecting sector %lu to another mirror\n"

/*
 * This is a kernel thread which:
 *
 *	1.	Retries failed read operations on working mirrors.
 *	2.	Updates the raid superblock when problems encounter.
 *	3.	Performs writes following reads for array syncronising.
 */
static void end_sync_write(struct buffer_head *bh, int uptodate);
static void end_sync_read(struct buffer_head *bh, int uptodate);

static void raid1d (void *data)
{
	struct raid1_bh *r1_bh;
	struct buffer_head *bh;
	unsigned long flags;
	mddev_t *mddev;
	kdev_t dev;


	for (;;) {
		md_spin_lock_irqsave(&retry_list_lock, flags);
		r1_bh = raid1_retry_list;
		if (!r1_bh)
			break;
		raid1_retry_list = r1_bh->next_r1;
		md_spin_unlock_irqrestore(&retry_list_lock, flags);

		mddev = r1_bh->mddev;
		if (mddev->sb_dirty) {
			printk(KERN_INFO "dirty sb detected, updating.\n");
			mddev->sb_dirty = 0;
			md_update_sb(mddev);
		}
		bh = &r1_bh->bh_req;
		switch(r1_bh->cmd) {
		case SPECIAL:
			/* have to allocate lots of bh structures and
			 * schedule writes
			 */
			if (test_bit(R1BH_Uptodate, &r1_bh->state)) {
				int i, sum_bhs = 0;
				int disks = MD_SB_DISKS;
				struct buffer_head *bhl, *mbh;
				raid1_conf_t *conf;
				int sectors = bh->b_size >> 9;
				
				conf = mddev_to_conf(mddev);
				bhl = raid1_alloc_bh(conf, conf->raid_disks); /* don't really need this many */
				for (i = 0; i < disks ; i++) {
					if (!conf->mirrors[i].operational)
						continue;
					if (i==conf->last_used)
						/* we read from here, no need to write */
						continue;
					if (i < conf->raid_disks
					    && !conf->resync_mirrors)
						/* don't need to write this,
						 * we are just rebuilding */
						continue;
					mbh = bhl;
					if (!mbh) {
						MD_BUG();
						break;
					}
					bhl = mbh->b_next;
					mbh->b_this_page = (struct buffer_head *)1;

						
				/*
				 * prepare mirrored bh (fields ordered for max mem throughput):
				 */
					mbh->b_blocknr    = bh->b_blocknr;
					mbh->b_dev        = conf->mirrors[i].dev;
					mbh->b_rdev	  = conf->mirrors[i].dev;
					mbh->b_rsector	  = bh->b_blocknr * sectors;
					mbh->b_state      = (1<<BH_Req) | (1<<BH_Dirty) |
						(1<<BH_Mapped) | (1<<BH_Lock);
					atomic_set(&mbh->b_count, 1);
					mbh->b_size       = bh->b_size;
					mbh->b_page	  = bh->b_page;
					mbh->b_data	  = bh->b_data;
					mbh->b_list       = BUF_LOCKED;
					mbh->b_end_io     = end_sync_write;
					mbh->b_private    = r1_bh;

					mbh->b_next = r1_bh->mirror_bh_list;
					r1_bh->mirror_bh_list = mbh;

					sum_bhs++;
				}
				md_atomic_set(&r1_bh->remaining, sum_bhs);
				if (bhl) raid1_free_bh(conf, bhl);
				mbh = r1_bh->mirror_bh_list;
				while (mbh) {
					struct buffer_head *bh1 = mbh;
					mbh = mbh->b_next;
					generic_make_request(WRITE, bh1);
					md_sync_acct(bh1->b_dev, bh1->b_size/512);
				}
			} else {
				dev = bh->b_dev;
				raid1_map (mddev, &bh->b_dev, bh->b_size >> 9);
				if (bh->b_dev == dev) {
					printk (IO_ERROR, partition_name(bh->b_dev), bh->b_blocknr);
					md_done_sync(mddev, bh->b_size>>10, 0);
				} else {
					printk (REDIRECT_SECTOR,
						partition_name(bh->b_dev), bh->b_blocknr);
					bh->b_rdev = bh->b_dev;
					generic_make_request(READ, bh);
				}
			}

			break;
		case READ:
		case READA:
			dev = bh->b_dev;
		
			raid1_map (mddev, &bh->b_dev, bh->b_size >> 9);
			if (bh->b_dev == dev) {
				printk (IO_ERROR, partition_name(bh->b_dev), bh->b_blocknr);
				raid1_end_bh_io(r1_bh, 0);
			} else {
				printk (REDIRECT_SECTOR,
					partition_name(bh->b_dev), bh->b_blocknr);
				bh->b_rdev = bh->b_dev;
				generic_make_request (r1_bh->cmd, bh);
			}
			break;
		}
	}
	md_spin_unlock_irqrestore(&retry_list_lock, flags);
}
#undef IO_ERROR
#undef REDIRECT_SECTOR

/*
 * Private kernel thread to reconstruct mirrors after an unclean
 * shutdown.
 */
static void raid1syncd (void *data)
{
	raid1_conf_t *conf = data;
	mddev_t *mddev = conf->mddev;

	if (!conf->resync_mirrors)
		return;
	if (conf->resync_mirrors == 2)
		return;
	down(&mddev->recovery_sem);
	if (!md_do_sync(mddev, NULL)) {
		/*
		 * Only if everything went Ok.
		 */
		conf->resync_mirrors = 0;
	}

	close_sync(conf);

	up(&mddev->recovery_sem);
	raid1_shrink_buffers(conf);
}

/*
 * perform a "sync" on one "block"
 *
 * We need to make sure that no normal I/O request - particularly write
 * requests - conflict with active sync requests.
 * This is achieved by conceptually dividing the device space into a
 * number of sections:
 *  DONE: 0 .. a-1     These blocks are in-sync
 *  ACTIVE: a.. b-1    These blocks may have active sync requests, but
 *                     no normal IO requests
 *  READY: b .. c-1    These blocks have no normal IO requests - sync
 *                     request may be happening
 *  PENDING: c .. d-1  These blocks may have IO requests, but no new
 *                     ones will be added
 *  FUTURE:  d .. end  These blocks are not to be considered yet. IO may
 *                     be happening, but not sync
 *
 * We keep a
 *   phase    which flips (0 or 1) each time d moves and
 * a count of:
 *   z =  active io requests in FUTURE since d moved - marked with
 *        current phase
 *   y =  active io requests in FUTURE before d moved, or PENDING -
 *        marked with previous phase
 *   x =  active sync requests in READY
 *   w =  active sync requests in ACTIVE
 *   v =  active io requests in DONE
 *
 * Normally, a=b=c=d=0 and z= active io requests
 *   or a=b=c=d=END and v= active io requests
 * Allowed changes to a,b,c,d:
 * A:  c==d &&  y==0 -> d+=window, y=z, z=0, phase=!phase
 * B:  y==0 -> c=d
 * C:   b=c, w+=x, x=0
 * D:  w==0 -> a=b
 * E: a==b==c==d==end -> a=b=c=d=0, z=v, v=0
 *
 * At start of sync we apply A.
 * When y reaches 0, we apply B then A then being sync requests
 * When sync point reaches c-1, we wait for y==0, and W==0, and
 * then apply apply B then A then D then C.
 * Finally, we apply E
 *
 * The sync request simply issues a "read" against a working drive
 * This is marked so that on completion the raid1d thread is woken to
 * issue suitable write requests
 */

static int raid1_sync_request (mddev_t *mddev, unsigned long block_nr)
{
	raid1_conf_t *conf = mddev_to_conf(mddev);
	struct mirror_info *mirror;
	struct raid1_bh *r1_bh;
	struct buffer_head *bh;
	int bsize;
	int disk;

	spin_lock_irq(&conf->segment_lock);
	if (!block_nr) {
		/* initialize ...*/
		int buffs;
		conf->start_active = 0;
		conf->start_ready = 0;
		conf->start_pending = 0;
		conf->start_future = 0;
		conf->phase = 0;
		/* we want enough buffers to hold twice the window of 128*/
		buffs = 128 *2 / (PAGE_SIZE>>9);
		buffs = raid1_grow_buffers(conf, buffs);
		if (buffs < 2)
			goto nomem;
		
		conf->window = buffs*(PAGE_SIZE>>9)/2;
		conf->cnt_future += conf->cnt_done+conf->cnt_pending;
		conf->cnt_done = conf->cnt_pending = 0;
		if (conf->cnt_ready || conf->cnt_active)
			MD_BUG();
	}
	while ((block_nr<<1) >= conf->start_pending) {
		PRINTK("wait .. sect=%lu start_active=%d ready=%d pending=%d future=%d, cnt_done=%d active=%d ready=%d pending=%d future=%d\n",
			block_nr<<1, conf->start_active, conf->start_ready, conf->start_pending, conf->start_future,
			conf->cnt_done, conf->cnt_active, conf->cnt_ready, conf->cnt_pending, conf->cnt_future);
		wait_event_lock_irq(conf->wait_done,
					!conf->cnt_active,
					conf->segment_lock);
		wait_event_lock_irq(conf->wait_ready,
					!conf->cnt_pending,
					conf->segment_lock);
		conf->start_active = conf->start_ready;
		conf->start_ready = conf->start_pending;
		conf->start_pending = conf->start_future;
		conf->start_future = conf->start_future+conf->window;
		// Note: falling off the end is not a problem
		conf->phase = conf->phase ^1;
		conf->cnt_active = conf->cnt_ready;
		conf->cnt_ready = 0;
		conf->cnt_pending = conf->cnt_future;
		conf->cnt_future = 0;
		wake_up(&conf->wait_done);
	}
	conf->cnt_ready++;
	spin_unlock_irq(&conf->segment_lock);
		

	/* If reconstructing, and >1 working disc,
	 * could dedicate one to rebuild and others to
	 * service read requests ..
	 */
	disk = conf->last_used;
	/* make sure disk is operational */
	while (!conf->mirrors[disk].operational) {
		if (disk <= 0) disk = conf->raid_disks;
		disk--;
		if (disk == conf->last_used)
			break;
	}
	conf->last_used = disk;
	
	mirror = conf->mirrors+conf->last_used;
	
	r1_bh = raid1_alloc_buf (conf);
	r1_bh->master_bh = NULL;
	r1_bh->mddev = mddev;
	r1_bh->cmd = SPECIAL;
	bh = &r1_bh->bh_req;

	bh->b_blocknr = block_nr;
	bsize = 1024;
	while (!(bh->b_blocknr & 1) && bsize < PAGE_SIZE
			&& (bh->b_blocknr+2)*(bsize>>10) < mddev->sb->size) {
		bh->b_blocknr >>= 1;
		bsize <<= 1;
	}
	bh->b_size = bsize;
	bh->b_list = BUF_LOCKED;
	bh->b_dev = mirror->dev;
	bh->b_rdev = mirror->dev;
	bh->b_state = (1<<BH_Req) | (1<<BH_Mapped) | (1<<BH_Lock);
	if (!bh->b_page)
		BUG();
	if (!bh->b_data)
		BUG();
	if (bh->b_data != page_address(bh->b_page))
		BUG();
	bh->b_end_io = end_sync_read;
	bh->b_private = r1_bh;
	bh->b_rsector = block_nr<<1;
	init_waitqueue_head(&bh->b_wait);

	generic_make_request(READ, bh);
	md_sync_acct(bh->b_dev, bh->b_size/512);

	return (bsize >> 10);

nomem:
	raid1_shrink_buffers(conf);
	spin_unlock_irq(&conf->segment_lock);
	return -ENOMEM;
}