Date:   Fri, 8 Nov 2019 14:01:10 +0100
From:   Cornelia Huck <cohuck@...hat.com>
To:     Parav Pandit <parav@...lanox.com>
Cc:     alex.williamson@...hat.com, davem@...emloft.net,
        kvm@...r.kernel.org, netdev@...r.kernel.org, saeedm@...lanox.com,
        kwankhede@...dia.com, leon@...nel.org, jiri@...lanox.com,
        linux-rdma@...r.kernel.org
Subject: Re: [PATCH net-next 11/19] vfio/mdev: Improvise mdev life cycle and
 parent removal scheme

On Thu,  7 Nov 2019 10:08:26 -0600
Parav Pandit <parav@...lanox.com> wrote:

I guess that should be s/Improvise/improve/ in $SUBJECT, no?

> mdev creation and removal sequence synchronization with parent device
> removal is improved in [1].
> 
> However such improvement using semaphore either limiting or leads to
> complex locking scheme when used across multiple subsystem such as mdev
> and devlink.
> 
> When mdev devices are used with devlink eswitch device, following
> deadlock sequence can be witnessed.
> 
> mlx5_core 0000:06:00.0: E-Switch: Disable: mode(OFFLOADS), nvfs(4), active vports(5)
> mlx5_core 0000:06:00.0: MDEV: Unregistering
> 
> WARNING: possible circular locking dependency detected
> ------------------------------------------------------
> devlink/42094 is trying to acquire lock:
> 00000000eb6fb4c7 (&parent->unreg_sem){++++}, at: mdev_unregister_device+0xf1/0x160 [mdev]
> 012but task is already holding lock:
> 00000000efcd208e (devlink_mutex){+.+.}, at: devlink_nl_pre_doit+0x1d/0x170
> 012which lock already depends on the new lock.
> 012the existing dependency chain (in reverse order) is:
> 012-> #1 (devlink_mutex){+.+.}:
>       lock_acquire+0xbd/0x1a0
>       __mutex_lock+0x84/0x8b0
>       devlink_unregister+0x17/0x60
>       mlx5_sf_unload+0x21/0x60 [mlx5_core]
>       mdev_remove+0x1e/0x40 [mdev]
>       device_release_driver_internal+0xdc/0x1a0
>       bus_remove_device+0xef/0x160
>       device_del+0x163/0x360
>       mdev_device_remove_common+0x1e/0xa0 [mdev]
>       mdev_device_remove+0x8d/0xd0 [mdev]
>       remove_store+0x71/0x90 [mdev]
>       kernfs_fop_write+0x113/0x1a0
>       vfs_write+0xad/0x1b0
>       ksys_write+0x5c/0xd0
>       do_syscall_64+0x5a/0x270
>       entry_SYSCALL_64_after_hwframe+0x49/0xbe
> 012-> #0 (&parent->unreg_sem){++++}:
>       check_prev_add+0xb0/0x810
>       __lock_acquire+0xd4b/0x1090
>       lock_acquire+0xbd/0x1a0
>       down_write+0x33/0x70
>       mdev_unregister_device+0xf1/0x160 [mdev]
>       esw_offloads_disable+0xe/0x70 [mlx5_core]
>       mlx5_eswitch_disable+0x149/0x190 [mlx5_core]
>       mlx5_devlink_eswitch_mode_set+0xd0/0x180 [mlx5_core]
>       devlink_nl_cmd_eswitch_set_doit+0x3e/0xb0
>       genl_family_rcv_msg+0x3a2/0x420
>       genl_rcv_msg+0x47/0x90
>       netlink_rcv_skb+0xc9/0x100
>       genl_rcv+0x24/0x40
>       netlink_unicast+0x179/0x220
>       netlink_sendmsg+0x2f6/0x3f0
>       sock_sendmsg+0x30/0x40
>       __sys_sendto+0xdc/0x160
>       __x64_sys_sendto+0x24/0x30
>       do_syscall_64+0x5a/0x270
>       entry_SYSCALL_64_after_hwframe+0x49/0xbe
> Possible unsafe locking scenario:
>       CPU0                    CPU1
>       ----                    ----
>  lock(devlink_mutex);
>                               lock(&parent->unreg_sem);
>                               lock(devlink_mutex);
>  lock(&parent->unreg_sem);
> 012 *** DEADLOCK ***
> 3 locks held by devlink/42094:
> 0: 0000000097a0c4aa (cb_lock){++++}, at: genl_rcv+0x15/0x40
> 1: 00000000baf61ad2 (genl_mutex){+.+.}, at: genl_rcv_msg+0x66/0x90
> 2: 00000000efcd208e (devlink_mutex){+.+.}, at: devlink_nl_pre_doit+0x1d/0x170
> 
> To summarize,
> mdev_remove()
>   read locks -> unreg_sem [ lock-A ]
>   [..]
>   devlink_unregister();
>     mutex lock devlink_mutex [ lock-B ]
> 
> devlink eswitch->switchdev-legacy mode change.
>  devlink_nl_cmd_eswitch_set_doit()
>    mutex lock devlink_mutex [ lock-B ]
>    mdev_unregister_device()
>    write locks -> unreg_sem [ lock-A]

So, this problem starts to pop up once you hook up that devlink stuff
with the mdev stuff, and previous users of mdev just did not have a
locking scheme similar to devlink?

> 
> Hence, instead of using semaphore, such synchronization is achieved
> using srcu which is more flexible that eliminates nested locking.
> 
> SRCU based solution is already proposed before at [2].
> 
> [1] commit 5715c4dd66a3 ("vfio/mdev: Synchronize device create/remove with parent removal")
> [2] https://lore.kernel.org/patchwork/patch/1055254/

I don't quite recall the discussion there... is this a rework of a
patch you proposed before? Confused.

> 
> Signed-off-by: Parav Pandit <parav@...lanox.com>
> ---
>  drivers/vfio/mdev/mdev_core.c    | 56 +++++++++++++++++++++++---------
>  drivers/vfio/mdev/mdev_private.h |  3 +-
>  2 files changed, 43 insertions(+), 16 deletions(-)

(...)

> @@ -207,6 +207,7 @@ int mdev_register_device(struct device *dev, const struct mdev_parent_ops *ops)
>  		dev_warn(dev, "Failed to create compatibility class link\n");
>  
>  	list_add(&parent->next, &parent_list);
> +	rcu_assign_pointer(parent->self, parent);
>  	mutex_unlock(&parent_list_lock);
>  
>  	dev_info(dev, "MDEV: Registered\n");
> @@ -250,14 +251,29 @@ void mdev_unregister_device(struct device *dev)
>  	list_del(&parent->next);
>  	mutex_unlock(&parent_list_lock);
>  
> -	down_write(&parent->unreg_sem);
> +	/*
> +	 * Publish that this mdev parent is unregistering. So any new
> +	 * create/remove cannot start on this parent anymore by user.
> +	 */
> +	rcu_assign_pointer(parent->self, NULL);
> +
> +	/*
> +	 * Wait for any active create() or remove() mdev ops on the parent
> +	 * to complete.
> +	 */
> +	synchronize_srcu(&parent->unreg_srcu);
> +
> +	/*
> +	 * At this point it is confirmed that any pending user initiated
> +	 * create or remove callbacks accessing the parent are completed.
> +	 * It is safe to remove the parent now.
> +	 */

So, you're putting an srcu-handled self reference there and use that as
an indication whether the parent is unregistering?

>  
>  	class_compat_remove_link(mdev_bus_compat_class, dev, NULL);
>  
>  	device_for_each_child(dev, NULL, mdev_device_remove_cb);
>  
>  	parent_remove_sysfs_files(parent);
> -	up_write(&parent->unreg_sem);
>  
>  	mdev_put_parent(parent);
>  

Powered by blists - more mailing lists