md: improve the is_mddev_idle test fix
authorNeilBrown <neilb@suse.de>
Tue, 17 Jul 2007 11:06:12 +0000 (04:06 -0700)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Tue, 17 Jul 2007 17:23:15 +0000 (10:23 -0700)
Don't use 'unsigned' variable to track sync vs non-sync IO, as the only thing
we want to do with them is a signed comparison, and fix up the comment which
had become quite wrong.

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

index 3d5a6beb333b95a0d8efc29ca1fb78b4885533be..bae42331182d6d93248a0ae425fb9399c8dd6c19 100644 (file)
@@ -5091,7 +5091,7 @@ static int is_mddev_idle(mddev_t *mddev)
        mdk_rdev_t * rdev;
        struct list_head *tmp;
        int idle;
-       unsigned long curr_events;
+       long curr_events;
 
        idle = 1;
        ITERATE_RDEV(mddev,rdev,tmp) {
@@ -5099,20 +5099,29 @@ static int is_mddev_idle(mddev_t *mddev)
                curr_events = disk_stat_read(disk, sectors[0]) + 
                                disk_stat_read(disk, sectors[1]) - 
                                atomic_read(&disk->sync_io);
-               /* The difference between curr_events and last_events
-                * will be affected by any new non-sync IO (making
-                * curr_events bigger) and any difference in the amount of
-                * in-flight syncio (making current_events bigger or smaller)
-                * The amount in-flight is currently limited to
-                * 32*64K in raid1/10 and 256*PAGE_SIZE in raid5/6
-                * which is at most 4096 sectors.
-                * These numbers are fairly fragile and should be made
-                * more robust, probably by enforcing the
-                * 'window size' that md_do_sync sort-of uses.
+               /* sync IO will cause sync_io to increase before the disk_stats
+                * as sync_io is counted when a request starts, and
+                * disk_stats is counted when it completes.
+                * So resync activity will cause curr_events to be smaller than
+                * when there was no such activity.
+                * non-sync IO will cause disk_stat to increase without
+                * increasing sync_io so curr_events will (eventually)
+                * be larger than it was before.  Once it becomes
+                * substantially larger, the test below will cause
+                * the array to appear non-idle, and resync will slow
+                * down.
+                * If there is a lot of outstanding resync activity when
+                * we set last_event to curr_events, then all that activity
+                * completing might cause the array to appear non-idle
+                * and resync will be slowed down even though there might
+                * not have been non-resync activity.  This will only
+                * happen once though.  'last_events' will soon reflect
+                * the state where there is little or no outstanding
+                * resync requests, and further resync activity will
+                * always make curr_events less than last_events.
                 *
-                * Note: the following is an unsigned comparison.
                 */
-               if ((long)curr_events - (long)rdev->last_events > 4096) {
+               if (curr_events - rdev->last_events > 4096) {
                        rdev->last_events = curr_events;
                        idle = 0;
                }
index de72c49747c8b20ac6f0844bc2703d8f22f347b9..28ac632b42dd128ab79610f6455b4a8362312b9c 100644 (file)
@@ -51,7 +51,7 @@ struct mdk_rdev_s
 
        sector_t size;                  /* Device size (in blocks) */
        mddev_t *mddev;                 /* RAID array if running */
-       unsigned long last_events;      /* IO event timestamp */
+       long last_events;               /* IO event timestamp */
 
        struct block_device *bdev;      /* block device handle */