[1/1] staging: android: ashmem: Fix lockdep warning for write operation
diff mbox series

Message ID 20200716024527.4009170-1-surenb@google.com
State New
Headers show
Series
  • [1/1] staging: android: ashmem: Fix lockdep warning for write operation
Related show

Commit Message

Suren Baghdasaryan July 16, 2020, 2:45 a.m. UTC
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/

Signed-off-by: Suren Baghdasaryan <surenb@google.com>
---
 drivers/staging/android/ashmem.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)

Comments

Eric Biggers July 16, 2020, 3:30 a.m. UTC | #1
On Wed, Jul 15, 2020 at 07:45:27PM -0700, Suren Baghdasaryan wrote:
> 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/
> 
> Signed-off-by: Suren Baghdasaryan <surenb@google.com>

Please add proper tags:

Reported-by: syzbot+7a0d9d0b26efefe61780@syzkaller.appspotmail.com
Fixes: ...
Cc: stable@vger.kernel.org


The Reported-by tag to use was given in the original syzbot report.

- Eric
Suren Baghdasaryan July 16, 2020, 7:30 p.m. UTC | #2
On Wed, Jul 15, 2020 at 8:30 PM Eric Biggers <ebiggers@kernel.org> wrote:
>
> On Wed, Jul 15, 2020 at 07:45:27PM -0700, Suren Baghdasaryan wrote:
> > 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/
> >
> > Signed-off-by: Suren Baghdasaryan <surenb@google.com>
>
> Please add proper tags:
>
> Reported-by: syzbot+7a0d9d0b26efefe61780@syzkaller.appspotmail.com
> Fixes: ...
> Cc: stable@vger.kernel.org
>
>
> The Reported-by tag to use was given in the original syzbot report.

Will add in v2. Thanks!

>
> - Eric
Joel Fernandes July 30, 2020, 3:23 a.m. UTC | #3
On Wed, Jul 15, 2020 at 10:45 PM Suren Baghdasaryan <surenb@google.com> wrote:
>
> 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/
>
> Signed-off-by: Suren Baghdasaryan <surenb@google.com>
> ---

Once Eric's nits are resolved:

Reviewed-by: Joel Fernandes (Google) <joel@joelfernandes.org>

Thanks.
Suren Baghdasaryan July 30, 2020, 7:15 p.m. UTC | #4
On Wed, Jul 29, 2020 at 8:24 PM Joel Fernandes <joel@joelfernandes.org> wrote:
>
> On Wed, Jul 15, 2020 at 10:45 PM Suren Baghdasaryan <surenb@google.com> wrote:
> >
> > 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/
> >
> > Signed-off-by: Suren Baghdasaryan <surenb@google.com>
> > ---
>
> Once Eric's nits are resolved:
>
> Reviewed-by: Joel Fernandes (Google) <joel@joelfernandes.org>

Thanks Joel!
I'm fixing the nits and will report the patch shortly. One note about
adding the "Fixes: " tag - this is a fix for a false positive lockdep
warning and it's unclear which patch should be quoted here (I could
not find a clear cause that started this warning). In similar
situations, for example here: https://lkml.org/lkml/2020/6/15/958
developers seem to skip that tag. So I'll do the same.

>
> Thanks.

Patch
diff mbox series

diff --git a/drivers/staging/android/ashmem.c b/drivers/staging/android/ashmem.c
index c05a214191da..10b4be1f3e78 100644
--- a/drivers/staging/android/ashmem.c
+++ b/drivers/staging/android/ashmem.c
@@ -95,6 +95,15 @@  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;
@@ -396,6 +405,7 @@  static int ashmem_mmap(struct file *file, struct vm_area_struct *vma)
 	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;
@@ -407,6 +417,8 @@  static int ashmem_mmap(struct file *file, struct vm_area_struct *vma)
 			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