Re: [PATCH v4] smb: client: Fix mount deadlock by avoiding super block iteration in DFS reconnect

[Paulo Alcantara <pc@manguebit.org>]

Wang Zhaolong <wangzhaolong@huaweicloud.com> writes:

> An AA deadlock occurs when network interruption during mount triggers
> DFS reconnection logic that calls iterate_supers_type().
>
> The detailed call process is as follows:
>
>       mount.cifs
> -------------------------
> path_mount
>   do_new_mount
>     vfs_get_tree
>       smb3_get_tree
>         cifs_smb3_do_mount
>           sget
>             alloc_super
>               down_write_nested(&s->s_umount, ..);  // Hold lock
>           cifs_root_iget
>             cifs_get_inode_info
>               smb2_query_path_info
>                 smb2_compound_op
>                   SMB2_open_init
>                     smb2_plain_req_init
>                       smb2_reconnect           // Trigger reconnection
>                         cifs_tree_connect
>                           cifs_get_dfs_tcon_super
>                             __cifs_get_super
>                               iterate_supers_type
>                                 down_read(&sb->s_umount); // Deadlock
>     do_new_mount_fc
>       up_write(&sb->s_umount);  // Release lock
>
> During mount phase, if reconnection is triggered, the foreground mount
> process may enter smb2_reconnect prior to the reconnect worker being
> scheduled, leading to a deadlock when subsequent DFS tree connect
> attempts reacquire the s_umount lock.
>
> The essential condition for triggering the issue is that the API
> iterate_supers_type() reacquires the s_umount lock. Therefore, one
> possible solution is to avoid using iterate_supers_type() and instead
> directly access the superblock through internal data structures.
>
> This patch fixes the problem by:
> - Add vfs_sb back-pointer to cifs_sb_info for direct access
> - Protect list traversal with existing tcon->sb_list_lock
> - Use atomic operations to safely manage super block references
> - Remove complex callback-based iteration in favor of simple loop
> - Rename cifs_put_tcp_super() to cifs_put_super() to avoid confusion
>
> Fixes: 3ae872de4107 ("smb: client: fix shared DFS root mounts with different prefixes")
> Signed-off-by: Wang Zhaolong <wangzhaolong@huaweicloud.com>
> ---
>
> V4:
>  - Perform a null pointer check on the return value of cifs_get_dfs_tcon_super()
>    to prevent NULL ptr dereference with DFS multiuser mount
>
> V3:
>  - Adjust the trace diagram for the super_lock_shared() section to align with
>    the latest mainline call flow. 
> V2:
>  - Adjust the trace diagram in the commit message to indicate when the lock
>    is released
>
>  fs/smb/client/cifs_fs_sb.h |  1 +
>  fs/smb/client/cifsfs.c     |  1 +
>  fs/smb/client/cifsproto.h  |  2 +-
>  fs/smb/client/dfs.c        |  4 +-
>  fs/smb/client/misc.c       | 84 ++++++++++++++------------------------
>  5 files changed, 35 insertions(+), 57 deletions(-)

Thanks for fixing the NULL ptr deref issue.

This patch still introduces a regression when reconnecting tcons
created in multiuser mounts by cifs_construct_tcon().  That is,
cifs_sb_info::prepath will not get updated in tree_connect_dfs_target()
because @cifs_sb will be NULL when calling cifs_get_dfs_tcon_super() on
non-master tcons.

Currently only master tcons will have a pointer to the superblock, which
is set in mount_setup_tlink().  You'd need to set superblock pointer to
all tcons in order to make this work.