nilfs2: fix lockdep warnings in page operations for btree nodes

Patch series "nilfs2 lockdep warning fixes".

The first two are to resolve the lockdep warning issue, and the last one
is the accompanying cleanup and low priority.

Based on your comment, this series solves the issue by separating inode
object as needed.  Since I was worried about the impact of the object
composition changes, I tested the series carefully not to cause
regressions especially for delicate functions such like disk space
reclamation and snapshots.

This patch (of 3):

If CONFIG_LOCKDEP is enabled, nilfs2 hits lockdep warnings at
inode_to_wb() during page/folio operations for btree nodes:

  WARNING: CPU: 0 PID: 6575 at include/linux/backing-dev.h:269 inode_to_wb include/linux/backing-dev.h:269 [inline]
  WARNING: CPU: 0 PID: 6575 at include/linux/backing-dev.h:269 folio_account_dirtied mm/page-writeback.c:2460 [inline]
  WARNING: CPU: 0 PID: 6575 at include/linux/backing-dev.h:269 __folio_mark_dirty+0xa7c/0xe30 mm/page-writeback.c:2509
  Modules linked in:
  ...
  RIP: 0010:inode_to_wb include/linux/backing-dev.h:269 [inline]
  RIP: 0010:folio_account_dirtied mm/page-writeback.c:2460 [inline]
  RIP: 0010:__folio_mark_dirty+0xa7c/0xe30 mm/page-writeback.c:2509
  ...
  Call Trace:
    __set_page_dirty include/linux/pagemap.h:834 [inline]
    mark_buffer_dirty+0x4e6/0x650 fs/buffer.c:1145
    nilfs_btree_propagate_p fs/nilfs2/btree.c:1889 [inline]
    nilfs_btree_propagate+0x4ae/0xea0 fs/nilfs2/btree.c:2085
    nilfs_bmap_propagate+0x73/0x170 fs/nilfs2/bmap.c:337
    nilfs_collect_dat_data+0x45/0xd0 fs/nilfs2/segment.c:625
    nilfs_segctor_apply_buffers+0x14a/0x470 fs/nilfs2/segment.c:1009
    nilfs_segctor_scan_file+0x47a/0x700 fs/nilfs2/segment.c:1048
    nilfs_segctor_collect_blocks fs/nilfs2/segment.c:1224 [inline]
    nilfs_segctor_collect fs/nilfs2/segment.c:1494 [inline]
    nilfs_segctor_do_construct+0x14f3/0x6c60 fs/nilfs2/segment.c:2036
    nilfs_segctor_construct+0x7a7/0xb30 fs/nilfs2/segment.c:2372
    nilfs_segctor_thread_construct fs/nilfs2/segment.c:2480 [inline]
    nilfs_segctor_thread+0x3c3/0xf90 fs/nilfs2/segment.c:2563
    kthread+0x405/0x4f0 kernel/kthread.c:327
    ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:295

This is because nilfs2 uses two page caches for each inode and
inode->i_mapping never points to one of them, the btree node cache.

This causes inode_to_wb(inode) to refer to a different page cache than
the caller page/folio operations such like __folio_start_writeback(),
__folio_end_writeback(), or __folio_mark_dirty() acquired the lock.

This patch resolves the issue by allocating and using an additional
inode to hold the page cache of btree nodes.  The inode is attached
one-to-one to the traditional nilfs2 inode if it requires a block
mapping with b-tree.  This setup change is in memory only and does not
affect the disk format.

Link: https://lkml.kernel.org/r/1647867427-30498-1-git-send-email-konishi.ryusuke@gmail.com
Link: https://lkml.kernel.org/r/1647867427-30498-2-git-send-email-konishi.ryusuke@gmail.com
Link: https://lore.kernel.org/r/YXrYvIo8YRnAOJCj@casper.infradead.org
Link: https://lore.kernel.org/r/9a20b33d-b38f-b4a2-4742-c1eb5b8e4d6c@redhat.com
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@gmail.com>
Reported-by: syzbot+0d5b462a6f07447991b3@syzkaller.appspotmail.com
Reported-by: syzbot+34ef28bb2aeb28724aa0@syzkaller.appspotmail.com
Reported-by: Hao Sun <sunhao.th@gmail.com>
Reported-by: David Hildenbrand <david@redhat.com>
Tested-by: Ryusuke Konishi <konishi.ryusuke@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 90 insertions, 14 deletions

diff --git a/fs/nilfs2/inode.c b/fs/nilfs2/inode.c
index 476a4a649f38..2557f5f5a935 100644
--- a/fs/nilfs2/inode.c
+++ b/fs/nilfs2/inode.c
@@ -29,12 +29,14 @@
  * @cno: checkpoint number
  * @root: pointer on NILFS root object (mounted checkpoint)
  * @for_gc: inode for GC flag
+ * @for_btnc: inode for B-tree node cache flag
  */
 struct nilfs_iget_args {
 	u64 ino;
 	__u64 cno;
 	struct nilfs_root *root;
-	int for_gc;
+	bool for_gc;
+	bool for_btnc;
 };
 
 static int nilfs_iget_test(struct inode *inode, void *opaque);
@@ -312,7 +314,8 @@ static int nilfs_insert_inode_locked(struct inode *inode,
 				     unsigned long ino)
 {
 	struct nilfs_iget_args args = {
-		.ino = ino, .root = root, .cno = 0, .for_gc = 0
+		.ino = ino, .root = root, .cno = 0, .for_gc = false,
+		.for_btnc = false
 	};
 
 	return insert_inode_locked4(inode, ino, nilfs_iget_test, &args);
@@ -525,6 +528,13 @@ static int nilfs_iget_test(struct inode *inode, void *opaque)
 		return 0;
 
 	ii = NILFS_I(inode);
+	if (test_bit(NILFS_I_BTNC, &ii->i_state)) {
+		if (!args->for_btnc)
+			return 0;
+	} else if (args->for_btnc) {
+		return 0;
+	}
+
 	if (!test_bit(NILFS_I_GCINODE, &ii->i_state))
 		return !args->for_gc;
 
@@ -536,15 +546,15 @@ static int nilfs_iget_set(struct inode *inode, void *opaque)
 	struct nilfs_iget_args *args = opaque;
 
 	inode->i_ino = args->ino;
-	if (args->for_gc) {
+	NILFS_I(inode)->i_cno = args->cno;
+	NILFS_I(inode)->i_root = args->root;
+	if (args->root && args->ino == NILFS_ROOT_INO)
+		nilfs_get_root(args->root);
+
+	if (args->for_gc)
 		NILFS_I(inode)->i_state = BIT(NILFS_I_GCINODE);
-		NILFS_I(inode)->i_cno = args->cno;
-		NILFS_I(inode)->i_root = NULL;
-	} else {
-		if (args->root && args->ino == NILFS_ROOT_INO)
-			nilfs_get_root(args->root);
-		NILFS_I(inode)->i_root = args->root;
-	}
+	if (args->for_btnc)
+		NILFS_I(inode)->i_state |= BIT(NILFS_I_BTNC);
 	return 0;
 }
 
@@ -552,7 +562,8 @@ struct inode *nilfs_ilookup(struct super_block *sb, struct nilfs_root *root,
 			    unsigned long ino)
 {
 	struct nilfs_iget_args args = {
-		.ino = ino, .root = root, .cno = 0, .for_gc = 0
+		.ino = ino, .root = root, .cno = 0, .for_gc = false,
+		.for_btnc = false
 	};
 
 	return ilookup5(sb, ino, nilfs_iget_test, &args);
@@ -562,7 +573,8 @@ struct inode *nilfs_iget_locked(struct super_block *sb, struct nilfs_root *root,
 				unsigned long ino)
 {
 	struct nilfs_iget_args args = {
-		.ino = ino, .root = root, .cno = 0, .for_gc = 0
+		.ino = ino, .root = root, .cno = 0, .for_gc = false,
+		.for_btnc = false
 	};
 
 	return iget5_locked(sb, ino, nilfs_iget_test, nilfs_iget_set, &args);
@@ -593,7 +605,8 @@ struct inode *nilfs_iget_for_gc(struct super_block *sb, unsigned long ino,
 				__u64 cno)
 {
 	struct nilfs_iget_args args = {
-		.ino = ino, .root = NULL, .cno = cno, .for_gc = 1
+		.ino = ino, .root = NULL, .cno = cno, .for_gc = true,
+		.for_btnc = false
 	};
 	struct inode *inode;
 	int err;
@@ -613,6 +626,68 @@ struct inode *nilfs_iget_for_gc(struct super_block *sb, unsigned long ino,
 	return inode;
 }
 
+/**
+ * nilfs_attach_btree_node_cache - attach a B-tree node cache to the inode
+ * @inode: inode object
+ *
+ * nilfs_attach_btree_node_cache() attaches a B-tree node cache to @inode,
+ * or does nothing if the inode already has it.  This function allocates
+ * an additional inode to maintain page cache of B-tree nodes one-on-one.
+ *
+ * Return Value: On success, 0 is returned. On errors, one of the following
+ * negative error code is returned.
+ *
+ * %-ENOMEM - Insufficient memory available.
+ */
+int nilfs_attach_btree_node_cache(struct inode *inode)
+{
+	struct nilfs_inode_info *ii = NILFS_I(inode);
+	struct inode *btnc_inode;
+	struct nilfs_iget_args args;
+
+	if (ii->i_assoc_inode)
+		return 0;
+
+	args.ino = inode->i_ino;
+	args.root = ii->i_root;
+	args.cno = ii->i_cno;
+	args.for_gc = test_bit(NILFS_I_GCINODE, &ii->i_state) != 0;
+	args.for_btnc = true;
+
+	btnc_inode = iget5_locked(inode->i_sb, inode->i_ino, nilfs_iget_test,
+				  nilfs_iget_set, &args);
+	if (unlikely(!btnc_inode))
+		return -ENOMEM;
+	if (btnc_inode->i_state & I_NEW) {
+		nilfs_init_btnc_inode(btnc_inode);
+		unlock_new_inode(btnc_inode);
+	}
+	NILFS_I(btnc_inode)->i_assoc_inode = inode;
+	NILFS_I(btnc_inode)->i_bmap = ii->i_bmap;
+	ii->i_assoc_inode = btnc_inode;
+
+	return 0;
+}
+
+/**
+ * nilfs_detach_btree_node_cache - detach the B-tree node cache from the inode
+ * @inode: inode object
+ *
+ * nilfs_detach_btree_node_cache() detaches the B-tree node cache and its
+ * holder inode bound to @inode, or does nothing if @inode doesn't have it.
+ */
+void nilfs_detach_btree_node_cache(struct inode *inode)
+{
+	struct nilfs_inode_info *ii = NILFS_I(inode);
+	struct inode *btnc_inode = ii->i_assoc_inode;
+
+	if (btnc_inode) {
+		NILFS_I(btnc_inode)->i_assoc_inode = NULL;
+		ii->i_assoc_inode = NULL;
+		iput(btnc_inode);
+	}
+}
+
 void nilfs_write_inode_common(struct inode *inode,
 			      struct nilfs_inode *raw_inode, int has_bmap)
 {
@@ -760,7 +835,8 @@ static void nilfs_clear_inode(struct inode *inode)
 	if (test_bit(NILFS_I_BMAP, &ii->i_state))
 		nilfs_bmap_clear(ii->i_bmap);
 
-	nilfs_btnode_cache_clear(&ii->i_btnode_cache);
+	if (!test_bit(NILFS_I_BTNC, &ii->i_state))
+		nilfs_detach_btree_node_cache(inode);
 
 	if (ii->i_root && inode->i_ino == NILFS_ROOT_INO)
 		nilfs_put_root(ii->i_root);