]> pilppa.com Git - linux-2.6-omap-h63xx.git/commitdiff
[PATCH] md: Only checkpoint expansion progress occasionally
authorNeilBrown <neilb@suse.de>
Mon, 27 Mar 2006 09:18:12 +0000 (01:18 -0800)
committerLinus Torvalds <torvalds@g5.osdl.org>
Mon, 27 Mar 2006 16:45:02 +0000 (08:45 -0800)
Instead of checkpointing at each stripe, only checkpoint when a new write
would overwrite uncheckpointed data.  Block any write to the uncheckpointed
area.  Arbitrarily checkpoint at least every 3Meg.

Signed-off-by: Neil Brown <neilb@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
drivers/md/raid5.c
include/linux/raid/raid5.h

index 20ae32d67e21786de09be86275583ab20819adbc..089a32604305c339b9a1fb20019fec8b15352c9d 100644 (file)
@@ -1762,8 +1762,9 @@ static int make_request(request_queue_t *q, struct bio * bi)
        for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
                DEFINE_WAIT(w);
                int disks;
-               
+
        retry:
+               prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
                if (likely(conf->expand_progress == MaxSector))
                        disks = conf->raid_disks;
                else {
@@ -1771,6 +1772,13 @@ static int make_request(request_queue_t *q, struct bio * bi)
                        disks = conf->raid_disks;
                        if (logical_sector >= conf->expand_progress)
                                disks = conf->previous_raid_disks;
+                       else {
+                               if (logical_sector >= conf->expand_lo) {
+                                       spin_unlock_irq(&conf->device_lock);
+                                       schedule();
+                                       goto retry;
+                               }
+                       }
                        spin_unlock_irq(&conf->device_lock);
                }
                new_sector = raid5_compute_sector(logical_sector, disks, disks - 1,
@@ -1779,7 +1787,6 @@ static int make_request(request_queue_t *q, struct bio * bi)
                        (unsigned long long)new_sector, 
                        (unsigned long long)logical_sector);
 
-               prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
                sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
                if (sh) {
                        if (unlikely(conf->expand_progress != MaxSector)) {
@@ -1877,6 +1884,7 @@ static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, i
                 */
                int i;
                int dd_idx;
+               sector_t writepos, safepos, gap;
 
                if (sector_nr == 0 &&
                    conf->expand_progress != 0) {
@@ -1887,15 +1895,36 @@ static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, i
                        return sector_nr;
                }
 
-               /* Cannot proceed until we've updated the superblock... */
-               wait_event(conf->wait_for_overlap,
-                          atomic_read(&conf->reshape_stripes)==0);
-               mddev->reshape_position = conf->expand_progress;
-
-               mddev->sb_dirty = 1;
-               md_wakeup_thread(mddev->thread);
-               wait_event(mddev->sb_wait, mddev->sb_dirty == 0 ||
-                       kthread_should_stop());
+               /* we update the metadata when there is more than 3Meg
+                * in the block range (that is rather arbitrary, should
+                * probably be time based) or when the data about to be
+                * copied would over-write the source of the data at
+                * the front of the range.
+                * i.e. one new_stripe forward from expand_progress new_maps
+                * to after where expand_lo old_maps to
+                */
+               writepos = conf->expand_progress +
+                       conf->chunk_size/512*(conf->raid_disks-1);
+               sector_div(writepos, conf->raid_disks-1);
+               safepos = conf->expand_lo;
+               sector_div(safepos, conf->previous_raid_disks-1);
+               gap = conf->expand_progress - conf->expand_lo;
+
+               if (writepos >= safepos ||
+                   gap > (conf->raid_disks-1)*3000*2 /*3Meg*/) {
+                       /* Cannot proceed until we've updated the superblock... */
+                       wait_event(conf->wait_for_overlap,
+                                  atomic_read(&conf->reshape_stripes)==0);
+                       mddev->reshape_position = conf->expand_progress;
+                       mddev->sb_dirty = 1;
+                       md_wakeup_thread(mddev->thread);
+                       wait_event(mddev->sb_wait, mddev->sb_dirty == 0 ||
+                                  kthread_should_stop());
+                       spin_lock_irq(&conf->device_lock);
+                       conf->expand_lo = mddev->reshape_position;
+                       spin_unlock_irq(&conf->device_lock);
+                       wake_up(&conf->wait_for_overlap);
+               }
 
                for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
                        int j;
@@ -2322,6 +2351,7 @@ static int run(mddev_t *mddev)
 
        if (conf->expand_progress != MaxSector) {
                printk("...ok start reshape thread\n");
+               conf->expand_lo = conf->expand_progress;
                atomic_set(&conf->reshape_stripes, 0);
                clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
                clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
@@ -2610,6 +2640,7 @@ static int raid5_reshape(mddev_t *mddev, int raid_disks)
        conf->previous_raid_disks = conf->raid_disks;
        conf->raid_disks = raid_disks;
        conf->expand_progress = 0;
+       conf->expand_lo = 0;
        spin_unlock_irq(&conf->device_lock);
 
        /* Add some new drives, as many as will fit.
index abcdf0d0658a57d20a4b0928dab4fe79470a54ff..914af667044f6f39eaf0a9ef51e6fa1966b44175 100644 (file)
@@ -217,6 +217,9 @@ struct raid5_private_data {
 
        /* used during an expand */
        sector_t                expand_progress;        /* MaxSector when no expand happening */
+       sector_t                expand_lo; /* from here up to expand_progress it out-of-bounds
+                                           * as we haven't flushed the metadata yet
+                                           */
        int                     previous_raid_disks;
 
        struct list_head        handle_list; /* stripes needing handling */