]> err.no Git - linux-2.6/commitdiff
[PATCH] md: Don't write dirty/clean update to spares - leave them alone
authorNeilBrown <neilb@suse.de>
Mon, 26 Jun 2006 07:27:57 +0000 (00:27 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Mon, 26 Jun 2006 16:58:39 +0000 (09:58 -0700)
- record the 'event' count on each individual device (they
  might sometimes be slightly different now)
- add a new value for 'sb_dirty': '3' means that the super
  block only needs to be updated to record a clean<->dirty
  transition.
- Prefer odd event numbers for dirty states and even numbers
  for clean states
- Using all the above, don't update the superblock on
  a spare device if the update is just doing a clean-dirty
  transition.  To accomodate this, a transition from
  dirty back to clean might now decrement the events counter
  if nothing else has changed.

The net effect of this is that spare drives will not see any IO requests
during normal running of the array, so they can go to sleep if that is what
they want to do.

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/md.c
include/linux/raid/md_k.h

index 6f97817d28b905b1ada0baaf94763f1b92246604..34b6902cda467ae88f63618203c2d91af8a0fe8a 100644 (file)
@@ -1558,15 +1558,30 @@ static void md_print_devices(void)
 }
 
 
-static void sync_sbs(mddev_t * mddev)
+static void sync_sbs(mddev_t * mddev, int nospares)
 {
+       /* Update each superblock (in-memory image), but
+        * if we are allowed to, skip spares which already
+        * have the right event counter, or have one earlier
+        * (which would mean they aren't being marked as dirty
+        * with the rest of the array)
+        */
        mdk_rdev_t *rdev;
        struct list_head *tmp;
 
        ITERATE_RDEV(mddev,rdev,tmp) {
-               super_types[mddev->major_version].
-                       sync_super(mddev, rdev);
-               rdev->sb_loaded = 1;
+               if (rdev->sb_events == mddev->events ||
+                   (nospares &&
+                    rdev->raid_disk < 0 &&
+                    (rdev->sb_events&1)==0 &&
+                    rdev->sb_events+1 == mddev->events)) {
+                       /* Don't update this superblock */
+                       rdev->sb_loaded = 2;
+               } else {
+                       super_types[mddev->major_version].
+                               sync_super(mddev, rdev);
+                       rdev->sb_loaded = 1;
+               }
        }
 }
 
@@ -1576,12 +1591,42 @@ void md_update_sb(mddev_t * mddev)
        struct list_head *tmp;
        mdk_rdev_t *rdev;
        int sync_req;
+       int nospares = 0;
 
 repeat:
        spin_lock_irq(&mddev->write_lock);
        sync_req = mddev->in_sync;
        mddev->utime = get_seconds();
-       mddev->events ++;
+       if (mddev->sb_dirty == 3)
+               /* just a clean<-> dirty transition, possibly leave spares alone,
+                * though if events isn't the right even/odd, we will have to do
+                * spares after all
+                */
+               nospares = 1;
+
+       /* If this is just a dirty<->clean transition, and the array is clean
+        * and 'events' is odd, we can roll back to the previous clean state */
+       if (mddev->sb_dirty == 3
+           && (mddev->in_sync && mddev->recovery_cp == MaxSector)
+           && (mddev->events & 1))
+               mddev->events--;
+       else {
+               /* otherwise we have to go forward and ... */
+               mddev->events ++;
+               if (!mddev->in_sync || mddev->recovery_cp != MaxSector) { /* not clean */
+                       /* .. if the array isn't clean, insist on an odd 'events' */
+                       if ((mddev->events&1)==0) {
+                               mddev->events++;
+                               nospares = 0;
+                       }
+               } else {
+                       /* otherwise insist on an even 'events' (for clean states) */
+                       if ((mddev->events&1)) {
+                               mddev->events++;
+                               nospares = 0;
+                       }
+               }
+       }
 
        if (!mddev->events) {
                /*
@@ -1593,7 +1638,7 @@ repeat:
                mddev->events --;
        }
        mddev->sb_dirty = 2;
-       sync_sbs(mddev);
+       sync_sbs(mddev, nospares);
 
        /*
         * do not write anything to disk if using
@@ -1615,6 +1660,8 @@ repeat:
        ITERATE_RDEV(mddev,rdev,tmp) {
                char b[BDEVNAME_SIZE];
                dprintk(KERN_INFO "md: ");
+               if (rdev->sb_loaded != 1)
+                       continue; /* no noise on spare devices */
                if (test_bit(Faulty, &rdev->flags))
                        dprintk("(skipping faulty ");
 
@@ -1626,6 +1673,7 @@ repeat:
                        dprintk(KERN_INFO "(write) %s's sb offset: %llu\n",
                                bdevname(rdev->bdev,b),
                                (unsigned long long)rdev->sb_offset);
+                       rdev->sb_events = mddev->events;
 
                } else
                        dprintk(")\n");
@@ -1895,6 +1943,7 @@ static mdk_rdev_t *md_import_device(dev_t newdev, int super_format, int super_mi
        rdev->desc_nr = -1;
        rdev->flags = 0;
        rdev->data_offset = 0;
+       rdev->sb_events = 0;
        atomic_set(&rdev->nr_pending, 0);
        atomic_set(&rdev->read_errors, 0);
        atomic_set(&rdev->corrected_errors, 0);
@@ -4708,7 +4757,7 @@ void md_write_start(mddev_t *mddev, struct bio *bi)
                spin_lock_irq(&mddev->write_lock);
                if (mddev->in_sync) {
                        mddev->in_sync = 0;
-                       mddev->sb_dirty = 1;
+                       mddev->sb_dirty = 3;
                        md_wakeup_thread(mddev->thread);
                }
                spin_unlock_irq(&mddev->write_lock);
@@ -5055,7 +5104,7 @@ void md_check_recovery(mddev_t *mddev)
                if (mddev->safemode && !atomic_read(&mddev->writes_pending) &&
                    !mddev->in_sync && mddev->recovery_cp == MaxSector) {
                        mddev->in_sync = 1;
-                       mddev->sb_dirty = 1;
+                       mddev->sb_dirty = 3;
                }
                if (mddev->safemode == 1)
                        mddev->safemode = 0;
index 682574f3bd364fee9324b805a12e53fae76467eb..c1e0ac55bab5c37956117280af7d2c428da063a2 100644 (file)
@@ -58,6 +58,7 @@ struct mdk_rdev_s
 
        struct page     *sb_page;
        int             sb_loaded;
+       __u64           sb_events;
        sector_t        data_offset;    /* start of data in array */
        sector_t        sb_offset;
        int             sb_size;        /* bytes in the superblock */