Re: [PATCH v2] btrfs: slience the sparse warn of rcu_string

[David Sterba <dsterba@suse.cz>]

On Mon, Sep 05, 2022 at 11:57:26AM +0800, Wang Yugui wrote:
> Hi,
> 
> > slience the sparse warn of rcu_string reported by 'make C=1'
> > 
> > warning example:
> > fs/btrfs/volumes.c:2300:41: warning: incorrect type in argument 3 (different address spaces)
> > fs/btrfs/volumes.c:2300:41:    expected char const *device_path
> > fs/btrfs/volumes.c:2300:41:    got char [noderef] __rcu *
> 
> I'm sorry that some new WARNING is triggered when 
> lockdep(CONFIG_PROVE_LOCKING/...) check is enabled.
> 
> [  112.742722] =============================
> [  112.746835] WARNING: suspicious RCU usage
> [  112.750917] 6.0.0-4.0.el7.x86_64 #1 Not tainted
> [  112.755540] -----------------------------
> [  112.759624] fs/btrfs/volumes.c:918 suspicious rcu_dereference_check() usage!
> [  112.766776] 
>                other info that might help us debug this:
> 
> [  112.774921] 
>                rcu_scheduler_active = 2, debug_locks = 1
> [  112.781566] 2 locks held by mount/1725:
> [  112.785471]  #0: ffffffffc50592a8 (uuid_mutex){+.+.}-{3:3}, at: btrfs_mount_root+0xfe/0x520 [btrfs]
> [  112.794763]  #1: ffffa1147bf068e0 (&fs_devs->device_list_mutex){+.+.}-{3:3}, at: device_list_add+0x28d/0x830 [btrfs]
> 
> 
> The fix is yet not clear,  we may need rcu_read_lock() just like
> 3181faa85bda3dc3f5e630a1846526c9caaa38e3 (cfq-iosched: fix a rcu warning)

There are probably more instances that would need the rcu lock/unlock,
eg. in btrfs_open_one_device, btrfs_dev_replace_finishing,
btrfs_destroy_dev_replace_tgtdev, btrfs_rm_device, clone_fs_devices,
device_list_add, btrfs_open_one_device where the dereference is done in
parameter list.

There's one dereference in the message helpers but not using the _in_rcu
variant:

zoned.c:
1403                 case BLK_ZONE_COND_READONLY:
1404                         btrfs_err(fs_info,
1405                 "zoned: offline/readonly zone %llu on device %s (devid %llu)",
1406                                   physical[i] >> device->zone_info->zone_size_shift,
1407                                   rcu_str_deref(device->name), device->devid);

We can keep the the patch in for-next and add the missing annotations as
they appear, updating the patch so it may take some time but we'll fix
all the warnings eventually.