Flush queue multidevice support.

Sort all jnodes by subvolume IDs in the list before its submitting.
Check subvolume ID when recovering extents to make sure that jnodes
of the same extent belong to the same subvolume.

Signed-off-by: Edward Shishkin <edward.shishkin@gmail.com>

flush_queue.c

@@ -16,6 +16,7 @@
 #include <linux/mm.h>
 #include <linux/pagemap.h>
 #include <linux/blkdev.h>
+#include <linux/list_sort.h>
 #include <linux/writeback.h>
 
 /* A flush queue object is an accumulator for keeping jnodes prepared
@@ -522,6 +523,26 @@ static void release_prepped_list(flush_queue_t *fq)
 	spin_unlock_atom(atom);
 }
 
+static int fq_compare_jnode(void* priv UNUSED_ARG,
+			    struct list_head *a, struct list_head *b)
+{
+	jnode *ja, *jb;
+
+	assert("edward-1873", a != NULL);
+	assert("edward-1874", b != NULL);
+
+	ja = jnode_by_link(a);
+	jb = jnode_by_link(b);
+
+	if (jnode_get_subvol(ja)->id < jnode_get_subvol(jb)->id)
+		return -1;
+	if (jnode_get_subvol(ja)->id > jnode_get_subvol(jb)->id)
+		return 1;
+	if (jnode_get_block(ja) < jnode_get_block(jb))
+		return -1;
+	return 1;
+}
+
 /**
  * Submit write requests for nodes on the already filled flush queue @fq.
  *
@@ -533,8 +554,6 @@ int reiser4_write_fq(flush_queue_t *fq, long *nr_submitted, int flags)
 {
 	int ret;
 	txn_atom *atom;
-	u32 mirr_id;
-	const u32 meta_id = current_volume()->vol_plug->meta_subvol_id();
 
 	while (1) {
 		atom = atom_locked_by_fq(fq);
@@ -547,18 +566,11 @@ int reiser4_write_fq(flush_queue_t *fq, long *nr_submitted, int flags)
 			break;
 		reiser4_atom_wait_event(atom);
 	}
-
 	atom->nr_running_queues++;
 	spin_unlock_atom(atom);
 
-	for_each_mirror(meta_id, mirr_id) {
-		reiser4_subvol *mirror = current_mirror(meta_id, mirr_id);
-
-		ret = write_jnode_list(ATOM_FQ_LIST(fq),
-				       fq, nr_submitted, flags, mirror);
-		if (ret)
-			break;
-	}
+	list_sort(NULL, ATOM_FQ_LIST(fq), fq_compare_jnode);
+	ret = write_jnode_list(ATOM_FQ_LIST(fq), fq, nr_submitted, flags);
 	release_prepped_list(fq);
 	return ret;
 }

jnode.h

@@ -370,8 +370,8 @@ extern struct address_space *jnode_get_mapping(const jnode * node) NONNULL;
 
 static inline reiser4_subvol *jnode_get_subvol(const jnode *node)
 {
-	assert("edard-1871", node != NULL);
-	assert("edard-1872", node->subvol != NULL);
+	assert("edward-1871", node != NULL);
+	assert("edward-1872", node->subvol != NULL);
 
 	return node->subvol;
 }

wander.c

@@ -1005,14 +1005,16 @@ static int write_jnodes_contig(jnode *first, int nr,
 }
 
 /**
+ * Submit a list of jnodes against specified subvolume @subv (it can be
+ * original subvolume, or replica).
  * This procedure recovers extents (contiguous sequences of disk block
  * numbers) in a given list of jnodes and submits write requests on this
  * per-extent basis.
  *
- * @head: the list of jnodes to write
+ * @head: the list of jnodes to submit
  */
-int write_jnode_list(struct list_head *head, flush_queue_t *fq,
-		     long *nr_submitted, int flags, reiser4_subvol *subv)
+int write_jnode_list_subv(struct list_head *head, flush_queue_t *fq,
+			  long *nr_submitted, int flags, reiser4_subvol *subv)
 {
 	int ret;
 	struct list_head *beg = head->next;
@@ -1022,6 +1024,9 @@ int write_jnode_list(struct list_head *head, flush_queue_t *fq,
 		struct list_head *cur = beg->next;
 
 		while (head != cur) {
+			assert("edward-1876",
+			       jnode_by_link(beg)->subvol ==
+			       subv);
 			assert("edward-1707",
 			       jnode_by_link(beg)->subvol ==
 			       jnode_by_link(cur)->subvol);
@@ -1044,6 +1049,67 @@ int write_jnode_list(struct list_head *head, flush_queue_t *fq,
 	return 0;
 }
 
+/**
+ * Submit a list of jnodes.
+ * Every jnode is submitted against an origimal subvolume and all its
+ * replicas.
+ * This procedure recovers extents (contiguous sequences of disk block
+ * numbers) in a given list of jnodes and submits write requests on this
+ * per-extent basis.
+ *
+ * @head: list of jnodes to submit.
+ */
+int write_jnode_list(struct list_head *head, flush_queue_t *fq,
+		     long *nr_submitted, int flags)
+{
+	int ret;
+	struct list_head *beg = head->next;
+
+	while (head != beg) {
+		int nr = 1;
+		u32 mirr_id;
+		struct list_head *cur = beg->next;
+		reiser4_subvol *subv = jnode_get_subvol(jnode_by_link(beg));
+
+		while (head != cur) {
+			if (jnode_get_subvol(jnode_by_link(cur)) !=
+			    jnode_get_subvol(jnode_by_link(beg)))
+				break;
+			if (*jnode_get_block(jnode_by_link(cur)) !=
+			    *jnode_get_block(jnode_by_link(beg)) + nr)
+				break;
+			++nr;
+			cur = cur->next;
+		}
+		/*
+		 * submit recovered extent against original subvolume
+		 * and all its replicas
+		 */
+		for_each_mirror(subv->id, mirr_id) {
+			reiser4_subvol *mirror;
+
+			mirror = current_mirror(subv->id, mirr_id);
+
+			ret = write_jnodes_contig(jnode_by_link(beg), nr,
+					    jnode_get_block(jnode_by_link(beg)),
+					    fq, flags, mirror);
+			if (ret)
+				return ret;
+		}
+#if 0
+		notice("edward-1875",
+		       "subvol %llu: written extent (%llu, %llu)",
+		       (unsigned long long)subv->id,
+		       (unsigned long long)*jnode_get_block(jnode_by_link(beg)),
+		       (unsigned long long)nr);
+#endif
+		if (nr_submitted)
+			*nr_submitted += nr;
+		beg = cur;
+	}
+	return 0;
+}
+
 /*
  * add given wandered mapping to atom's wandered map
  */
@@ -1256,7 +1322,7 @@ static int alloc_submit_wander_records(struct commit_handle *ch,
 	/*
 	 * submit wander records
 	 */
-	ret = write_jnode_list(tx_list, fq, NULL, 0, subv);
+	ret = write_jnode_list_subv(tx_list, fq, NULL, 0, subv);
 
 	return ret;
 
@@ -1339,8 +1405,8 @@ static int play_tx_mirror(struct commit_handle *ch, reiser4_subvol *mirror)
 		return  PTR_ERR(fq);
 	spin_unlock_atom(fq->atom);
 
-	ret = write_jnode_list(&ch_sub->overwrite_set,
-			       fq, NULL, WRITEOUT_FOR_PAGE_RECLAIM, mirror);
+	ret = write_jnode_list_subv(&ch_sub->overwrite_set, fq,
+				    NULL, WRITEOUT_FOR_PAGE_RECLAIM, mirror);
 	reiser4_fq_put(fq);
 	return ret;
 }
@@ -1610,7 +1676,8 @@ static int replay_tx_subv(reiser4_subvol *subv)
 		subv = super_mirror(subv->super, orig_id, mirr_id);
 		assert("edward-1711", subvol_is_set(subv, SUBVOL_ACTIVATED));
 
-		write_jnode_list(&ch_sub->overwrite_set, NULL, NULL, 0, subv);
+		write_jnode_list_subv(&ch_sub->overwrite_set,
+				      NULL, NULL, 0, subv);
 		ret = wait_on_jnode_list(&ch_sub->overwrite_set);
 		if (ret)
 			goto error;

wander.h

@@ -120,8 +120,9 @@ extern int reiser4_journal_recover_sb_data(struct super_block *,
 extern int reiser4_init_journal_info(reiser4_subvol *subv);
 extern void reiser4_done_journal_info(reiser4_subvol *subv);
 
-extern int write_jnode_list(struct list_head *, flush_queue_t *, long *, int,
-			    reiser4_subvol *);
+extern int write_jnode_list_subv(struct list_head *, flush_queue_t *,
+				 long *, int, reiser4_subvol *);
+extern int write_jnode_list(struct list_head *, flush_queue_t *, long *, int);
 
 #endif /* __FS_REISER4_WANDER_H__ */