xen/xenbus: Avoid a lockdep warning when adding a watch

The following lockdep warning appears during boot on a Xen dom0 system:

[   96.388794] ======================================================
[   96.388797] WARNING: possible circular locking dependency detected
[   96.388799] 6.4.0-rc5-default+ #8 Tainted: G            EL
[   96.388803] ------------------------------------------------------
[   96.388804] xenconsoled/1330 is trying to acquire lock:
[   96.388808] ffffffff82acdd10 (xs_watch_rwsem){++++}-{3:3}, at: register_xenbus_watch+0x45/0x140
[   96.388847]
               but task is already holding lock:
[   96.388849] ffff888100c92068 (&u->msgbuffer_mutex){+.+.}-{3:3}, at: xenbus_file_write+0x2c/0x600
[   96.388862]
               which lock already depends on the new lock.

[   96.388864]
               the existing dependency chain (in reverse order) is:
[   96.388866]
               -> #2 (&u->msgbuffer_mutex){+.+.}-{3:3}:
[   96.388874]        __mutex_lock+0x85/0xb30
[   96.388885]        xenbus_dev_queue_reply+0x48/0x2b0
[   96.388890]        xenbus_thread+0x1d7/0x950
[   96.388897]        kthread+0xe7/0x120
[   96.388905]        ret_from_fork+0x2c/0x50
[   96.388914]
               -> #1 (xs_response_mutex){+.+.}-{3:3}:
[   96.388923]        __mutex_lock+0x85/0xb30
[   96.388930]        xenbus_backend_ioctl+0x56/0x1c0
[   96.388935]        __x64_sys_ioctl+0x90/0xd0
[   96.388942]        do_syscall_64+0x5c/0x90
[   96.388950]        entry_SYSCALL_64_after_hwframe+0x72/0xdc
[   96.388957]
               -> #0 (xs_watch_rwsem){++++}-{3:3}:
[   96.388965]        __lock_acquire+0x1538/0x2260
[   96.388972]        lock_acquire+0xc6/0x2b0
[   96.388976]        down_read+0x2d/0x160
[   96.388983]        register_xenbus_watch+0x45/0x140
[   96.388990]        xenbus_file_write+0x53d/0x600
[   96.388994]        vfs_write+0xe4/0x490
[   96.389003]        ksys_write+0xb8/0xf0
[   96.389011]        do_syscall_64+0x5c/0x90
[   96.389017]        entry_SYSCALL_64_after_hwframe+0x72/0xdc
[   96.389023]
               other info that might help us debug this:

[   96.389025] Chain exists of:
                 xs_watch_rwsem --> xs_response_mutex --> &u->msgbuffer_mutex

[   96.413429]  Possible unsafe locking scenario:

[   96.413430]        CPU0                    CPU1
[   96.413430]        ----                    ----
[   96.413431]   lock(&u->msgbuffer_mutex);
[   96.413432]                                lock(xs_response_mutex);
[   96.413433]                                lock(&u->msgbuffer_mutex);
[   96.413434]   rlock(xs_watch_rwsem);
[   96.413436]
                *** DEADLOCK ***

[   96.413436] 1 lock held by xenconsoled/1330:
[   96.413438]  #0: ffff888100c92068 (&u->msgbuffer_mutex){+.+.}-{3:3}, at: xenbus_file_write+0x2c/0x600
[   96.413446]

An ioctl call IOCTL_XENBUS_BACKEND_SETUP (record #1 in the report)
results in calling xenbus_alloc() -> xs_suspend() which introduces
ordering xs_watch_rwsem --> xs_response_mutex. The xenbus_thread()
operation (record #2) creates xs_response_mutex --> &u->msgbuffer_mutex.
An XS_WATCH write to the xenbus file then results in a complain about
the opposite lock order &u->msgbuffer_mutex --> xs_watch_rwsem.

The dependency xs_watch_rwsem --> xs_response_mutex is spurious. Avoid
it and the warning by changing the ordering in xs_suspend(), first
acquire xs_response_mutex and then xs_watch_rwsem. Reverse also the
unlocking order in xs_suspend_cancel() for consistency, but keep
xs_resume() as is because it needs to have xs_watch_rwsem unlocked only
after exiting xs suspend and re-adding all watches.
Signed-off-by: Petr Pavlu <petr.pavlu@suse.com>
Reviewed-by: Juergen Gross <jgross@suse.com>
Link: https://lore.kernel.org/r/20230607123624.15739-1-petr.pavlu@suse.comSigned-off-by: Juergen Gross <jgross@suse.com>

drivers/xen/xenbus/xenbus_xs.c

@@ -840,8 +840,8 @@ void xs_suspend(void)
 {
 	xs_suspend_enter();
 
-	down_write(&xs_watch_rwsem);
 	mutex_lock(&xs_response_mutex);
+	down_write(&xs_watch_rwsem);
 }
 
 void xs_resume(void)
@@ -866,8 +866,8 @@ void xs_resume(void)
 
 void xs_suspend_cancel(void)
 {
-	mutex_unlock(&xs_response_mutex);
 	up_write(&xs_watch_rwsem);
+	mutex_unlock(&xs_response_mutex);
 
 	xs_suspend_exit();
 }