md/raid5: close race between STRIPE_BIT_DELAY and batching.
authorNeilBrown <neilb@suse.de>
Tue, 26 May 2015 22:43:45 +0000 (08:43 +1000)
committerNeilBrown <neilb@suse.de>
Thu, 28 May 2015 01:34:40 +0000 (11:34 +1000)
When we add a write to a stripe we need to make sure the bitmap
bit is set.  While doing that the stripe is not locked so it could
be added to a batch after which further changes to STRIPE_BIT_DELAY
and ->bm_seq are ineffective.

So we need to hold off adding to a stripe until bitmap_startwrite has
completed at least once, and we need to avoid further changes to
STRIPE_BIT_DELAY once the stripe has been added to a batch.

If a bitmap_startwrite() completes after the stripe was added to a
batch, it will not have set the bit, only incremented a counter, so no
extra delay of the stripe is needed.

Reported-by: Shaohua Li <shli@kernel.org>
Signed-off-by: NeilBrown <neilb@suse.de>
drivers/md/raid5.c
drivers/md/raid5.h

index c55a68f37c72518ecafc6b9b8ba7b7bbd88f311e..42d0ea6c8597cff4a8583ae26edbfee2829762fd 100644 (file)
@@ -749,6 +749,7 @@ static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 static bool stripe_can_batch(struct stripe_head *sh)
 {
        return test_bit(STRIPE_BATCH_READY, &sh->state) &&
+               !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
                is_full_stripe_write(sh);
 }
 
@@ -2996,14 +2997,32 @@ static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
        pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
                (unsigned long long)(*bip)->bi_iter.bi_sector,
                (unsigned long long)sh->sector, dd_idx);
-       spin_unlock_irq(&sh->stripe_lock);
 
        if (conf->mddev->bitmap && firstwrite) {
+               /* Cannot hold spinlock over bitmap_startwrite,
+                * but must ensure this isn't added to a batch until
+                * we have added to the bitmap and set bm_seq.
+                * So set STRIPE_BITMAP_PENDING to prevent
+                * batching.
+                * If multiple add_stripe_bio() calls race here they
+                * much all set STRIPE_BITMAP_PENDING.  So only the first one
+                * to complete "bitmap_startwrite" gets to set
+                * STRIPE_BIT_DELAY.  This is important as once a stripe
+                * is added to a batch, STRIPE_BIT_DELAY cannot be changed
+                * any more.
+                */
+               set_bit(STRIPE_BITMAP_PENDING, &sh->state);
+               spin_unlock_irq(&sh->stripe_lock);
                bitmap_startwrite(conf->mddev->bitmap, sh->sector,
                                  STRIPE_SECTORS, 0);
-               sh->bm_seq = conf->seq_flush+1;
-               set_bit(STRIPE_BIT_DELAY, &sh->state);
+               spin_lock_irq(&sh->stripe_lock);
+               clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
+               if (!sh->batch_head) {
+                       sh->bm_seq = conf->seq_flush+1;
+                       set_bit(STRIPE_BIT_DELAY, &sh->state);
+               }
        }
+       spin_unlock_irq(&sh->stripe_lock);
 
        if (stripe_can_batch(sh))
                stripe_add_to_batch_list(conf, sh);
index 7dc0dd86074b1702276ccb51ba166a38d5d0f7e3..01cdb9f3a0c4629a96d0e9f3841e2299beb415d4 100644 (file)
@@ -337,6 +337,9 @@ enum {
        STRIPE_ON_RELEASE_LIST,
        STRIPE_BATCH_READY,
        STRIPE_BATCH_ERR,
+       STRIPE_BITMAP_PENDING,  /* Being added to bitmap, don't add
+                                * to batch yet.
+                                */
 };
 
 #define STRIPE_EXPAND_SYNC_FLAG \