From: Suren Baghdasaryan surenb@google.com
commit 3e338d3c95c735dc3265a86016bb4c022ec7cadc upstream.
syzbot report [1] describes a deadlock when write operation against an ashmem fd executed at the time when ashmem is shrinking its cache results in the following lock sequence:
Possible unsafe locking scenario:
CPU0 CPU1 ---- ---- lock(fs_reclaim); lock(&sb->s_type->i_mutex_key#13); lock(fs_reclaim); lock(&sb->s_type->i_mutex_key#13);
kswapd takes fs_reclaim and then inode_lock while generic_perform_write takes inode_lock and then fs_reclaim. However ashmem does not support writing into backing shmem with a write syscall. The only way to change its content is to mmap it and operate on mapped memory. Therefore the race that lockdep is warning about is not valid. Resolve this by introducing a separate lockdep class for the backing shmem inodes.
[1]: https://lkml.kernel.org/lkml/0000000000000b5f9d059aa2037f@google.com/
Reported-by: syzbot+7a0d9d0b26efefe61780@syzkaller.appspotmail.com Signed-off-by: Suren Baghdasaryan surenb@google.com Cc: stable stable@vger.kernel.org Reviewed-by: Joel Fernandes (Google) joel@joelfernandes.org Link: https://lore.kernel.org/r/20200730192632.3088194-1-surenb@google.com Signed-off-by: Greg Kroah-Hartman gregkh@linuxfoundation.org
--- drivers/staging/android/ashmem.c | 12 ++++++++++++ 1 file changed, 12 insertions(+)
--- a/drivers/staging/android/ashmem.c +++ b/drivers/staging/android/ashmem.c @@ -100,6 +100,15 @@ static DEFINE_MUTEX(ashmem_mutex); static struct kmem_cache *ashmem_area_cachep __read_mostly; static struct kmem_cache *ashmem_range_cachep __read_mostly;
+/* + * A separate lockdep class for the backing shmem inodes to resolve the lockdep + * warning about the race between kswapd taking fs_reclaim before inode_lock + * and write syscall taking inode_lock and then fs_reclaim. + * Note that such race is impossible because ashmem does not support write + * syscalls operating on the backing shmem. + */ +static struct lock_class_key backing_shmem_inode_class; + static inline unsigned long range_size(struct ashmem_range *range) { return range->pgend - range->pgstart + 1; @@ -406,6 +415,7 @@ static int ashmem_mmap(struct file *file if (!asma->file) { char *name = ASHMEM_NAME_DEF; struct file *vmfile; + struct inode *inode;
if (asma->name[ASHMEM_NAME_PREFIX_LEN] != '\0') name = asma->name; @@ -417,6 +427,8 @@ static int ashmem_mmap(struct file *file goto out; } vmfile->f_mode |= FMODE_LSEEK; + inode = file_inode(vmfile); + lockdep_set_class(&inode->i_rwsem, &backing_shmem_inode_class); asma->file = vmfile; /* * override mmap operation of the vmfile so that it can't be