[PATCH v2] jffs2: put directories f->sem on a separate lockdep class

[Helmut Grohne]

We get a lockdep warning if we read a directory (e.g. ls) and fault a
page on a regular file:

[  181.206207] mmaptest/392 is trying to acquire lock:
[  181.211066]  (&f->sem){+.+.}, at: [<c01ff7d8>] jffs2_readpage+0x30/0x5c
[  181.217663]
[  181.217663] but task is already holding lock:
[  181.223477]  (&mm->mmap_sem){++++}, at: [<c04928e8>] do_page_fault+0xa8/0x454
[  181.230596]
[  181.230596] which lock already depends on the new lock.
[  181.230596]
[  181.238755]
[  181.238755] the existing dependency chain (in reverse order) is:
[  181.246219]
[  181.246219] -> #1 (&mm->mmap_sem){++++}:
[  181.251614]        __might_fault+0x90/0xc0
[  181.255694]        filldir+0xa4/0x1f4
[  181.259334]        jffs2_readdir+0xd8/0x1d4
[  181.263499]        iterate_dir+0x78/0x128
[  181.267490]        SyS_getdents+0x8c/0x12c
[  181.271576]        ret_fast_syscall+0x0/0x28
[  181.275818]
[  181.275818] -> #0 (&f->sem){+.+.}:
[  181.280690]        lock_acquire+0xfc/0x2b0
[  181.284763]        __mutex_lock+0x8c/0xb0c
[  181.288842]        mutex_lock_nested+0x2c/0x34
[  181.293272]        jffs2_readpage+0x30/0x5c
[  181.297438]        filemap_fault+0x600/0x73c
[  181.301690]        __do_fault+0x28/0xb8
[  181.305510]        handle_mm_fault+0x854/0xec0
[  181.309937]        do_page_fault+0x384/0x454
[  181.314188]        do_DataAbort+0x4c/0xc8
[  181.318181]        __dabt_usr+0x44/0x60
[  181.321999]        0xb6dffb02

Like with inode->i_rwsem, we recognize that this can never result in a
dead lock, because we cannot page fault directory inodes and we cannot
call getdents on non-directories. Thus we need different lockdep classes
depending on the mode. This mirrors what
lockdep_annotate_inode_mutex_key does to inodes.

Link: https://www.spinics.net/lists/linux-rt-users/msg07644.html
Signed-off-by: Helmut Grohne <h.grohne at intenta.de>
Cc: Peter Zijlstra <peterz at infradead.org>
---
I have reworked only this last patch of the series to avoid sending three
extra mails. If you prefer the full series, I can resend it.

Changing lock classes after using a lock is not a common thing to do, yet
this patch does exactly that. Thus I opted for performing two extra tests
to gain more confidence in the correctness of this approach:

 * I tried adding a BUG_ON(!mutex_is_locked(&f->sem)) to before the lock
   class change to validate that the mutex on the I_NEW inode cannot be
   held. This call did not error out in my tests.

 * I tried adding a mutex_clear(&f->sem) to jffs2_do_clear_inode and
   reinitialized it before setting the mutex class to validate that it
   really isn't used in the relevant time frame. This did not cause
   lockdep to complain in my tests.

Changes in v2:
 - Initialize the lock class on every inode initialization rather than just the
   initial one in jffs2_i_init_once. (Thanks to David Woodhouse for pointing at
   this issue.)
 - Properly wrap the commit message to make checkpatch.pl fully happy.

---
 fs/jffs2/fs.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/fs/jffs2/fs.c b/fs/jffs2/fs.c
index 89a10b398d00..9bc3cd204732 100644
--- a/fs/jffs2/fs.c
+++ b/fs/jffs2/fs.c
@@ -26,6 +26,9 @@
 #include <linux/crc32.h>
 #include "nodelist.h"
 
+static struct lock_class_key jffs2_inode_info_sem_key,
+	 jffs2_inode_info_sem_dir_key;
+
 static int jffs2_flash_setup(struct jffs2_sb_info *c);
 
 int jffs2_do_setattr (struct inode *inode, struct iattr *iattr)
@@ -277,6 +280,11 @@ struct inode *jffs2_iget(struct super_block *sb, unsigned long ino)
 
 	inode->i_mode = jemode_to_cpu(latest_node.mode);
 
+	/* Mirror lock class of inode->i_rwsem, see
+	 * lockdep_annotate_inode_mutex_key.
+	 */
+	lockdep_set_class(&f->sem, S_ISDIR(inode->i_mode) ?
+		&jffs2_inode_info_sem_dir_key : &jffs2_inode_info_sem_key);
 	mutex_lock(&f->sem);
 
 	i_uid_write(inode, je16_to_cpu(latest_node.uid));
@@ -439,6 +447,11 @@ struct inode *jffs2_new_inode (struct inode *dir_i, umode_t mode, struct jffs2_r
 
 	f = JFFS2_INODE_INFO(inode);
 	jffs2_init_inode_info(f);
+	/* Mirror lock class of inode->i_rwsem, see
+	 * lockdep_annotate_inode_mutex_key.
+	 */
+	lockdep_set_class(&f->sem, S_ISDIR(inode->i_mode) ?
+		&jffs2_inode_info_sem_dir_key : &jffs2_inode_info_sem_key);
 	mutex_lock(&f->sem);
 
 	memset(ri, 0, sizeof(*ri));
-- 
2.11.0