bcache: FUA fixes

Journal writes need to be marked FUA, not just REQ_FLUSH. And btree node
writes have... weird ordering requirements.

Signed-off-by: Kent Overstreet <koverstreet@google.com>

3 files changed, 35 insertions, 5 deletions

diff --git a/drivers/md/bcache/btree.c b/drivers/md/bcache/btree.c
index 09fb8a2f43da..a6ad49ac5f2b 100644
--- a/drivers/md/bcache/btree.c
+++ b/drivers/md/bcache/btree.c
@@ -328,10 +328,25 @@ static void do_btree_node_write(struct btree *b)
 
         b->bio->bi_end_io       = btree_node_write_endio;
         b->bio->bi_private      = &b->io.cl;
-        b->bio->bi_rw   = REQ_META|WRITE_SYNC;
-        b->bio->bi_size = set_blocks(i, b->c) * block_bytes(b->c);
+        b->bio->bi_rw           = REQ_META|WRITE_SYNC|REQ_FUA;
+        b->bio->bi_size         = set_blocks(i, b->c) * block_bytes(b->c);
         bch_bio_map(b->bio, i);
 
+        /*
+         * If we're appending to a leaf node, we don't technically need FUA -
+         * this write just needs to be persisted before the next journal write,
+         * which will be marked FLUSH|FUA.
+         *
+         * Similarly if we're writing a new btree root - the pointer is going to
+         * be in the next journal entry.
+         *
+         * But if we're writing a new btree node (that isn't a root) or
+         * appending to a non leaf btree node, we need either FUA or a flush
+         * when we write the parent with the new pointer. FUA is cheaper than a
+         * flush, and writes appending to leaf nodes aren't blocking anything so
+         * just make all btree node writes FUA to keep things sane.
+         */
+
         bkey_copy(&k.key, &b->key);
         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) + bset_offset(b, i));
 
@@ -2092,6 +2107,9 @@ int bch_btree_insert(struct btree_op *op, struct cache_set *c)
 void bch_btree_set_root(struct btree *b)
 {
         unsigned i;
+        struct closure cl;
+
+        closure_init_stack(&cl);
 
         trace_bcache_btree_set_root(b);
 
@@ -2107,7 +2125,8 @@ void bch_btree_set_root(struct btree *b)
         b->c->root = b;
         __bkey_put(b->c, &b->key);
 
-        bch_journal_meta(b->c, NULL);
+        bch_journal_meta(b->c, &cl);
+        closure_sync(&cl);
 }
 
 /* Cache lookup */
diff --git a/drivers/md/bcache/journal.c b/drivers/md/bcache/journal.c
index 5ca22149b749..4b250667bb7f 100644
--- a/drivers/md/bcache/journal.c
+++ b/drivers/md/bcache/journal.c
@@ -620,7 +620,7 @@ static void journal_write_unlocked(struct closure *cl)
                 bio_reset(bio);
                 bio->bi_sector  = PTR_OFFSET(k, i);
                 bio->bi_bdev    = ca->bdev;
-                bio->bi_rw      = REQ_WRITE|REQ_SYNC|REQ_META|REQ_FLUSH;
+                bio->bi_rw      = REQ_WRITE|REQ_SYNC|REQ_META|REQ_FLUSH|REQ_FUA;
                 bio->bi_size    = sectors << 9;
 
                 bio->bi_end_io  = journal_write_endio;
diff --git a/drivers/md/bcache/request.c b/drivers/md/bcache/request.c
index 17bd59704eba..bcdf1f782c3e 100644
--- a/drivers/md/bcache/request.c
+++ b/drivers/md/bcache/request.c
@@ -1035,8 +1035,19 @@ static void request_write(struct cached_dev *dc, struct search *s)
 
                 closure_bio_submit(bio, cl, s->d);
         } else {
-                s->op.cache_bio = bio;
                 bch_writeback_add(dc);
+
+                if (s->op.flush_journal) {
+                        /* Also need to send a flush to the backing device */
+                        s->op.cache_bio = bio_clone_bioset(bio, GFP_NOIO,
+                                                           dc->disk.bio_split);
+
+                        bio->bi_size = 0;
+                        bio->bi_vcnt = 0;
+                        closure_bio_submit(bio, cl, s->d);
+                } else {
+                        s->op.cache_bio = bio;
+                }
         }
 out:
         closure_call(&s->op.cl, bch_insert_data, NULL, cl);