Re: [PATCH 1/2] btrfs: initialize sysfs devid and device link for seed device

From: Anand Jain <anand.jain@oracle.com>
To: Josef Bacik <josef@toxicpanda.com>, linux-btrfs@vger.kernel.org
Subject: Re: [PATCH 1/2] btrfs: initialize sysfs devid and device link for seed device
Date: Sun, 23 Aug 2020 21:05:38 +0800	[thread overview]
Message-ID: <ad2a98b5-320c-4990-d1e5-9c7bc3ae71a0@oracle.com> (raw)
In-Reply-To: <779bd819-d320-39e3-0a0b-80c0c8455243@toxicpanda.com>

> 
> So now we're using the main fs_devices->devices_kobj, which is the main 
> fs_devices with fs_devices->seed being the seed fs_devices.  This is 
> fine, except when we actually mount a seed device, and in that case we 
> have fs_devices as the seed devices being used, and then if we add a 
> device we'll actually swap in the new fs_devices for the main 
> fs_devices, and we have the seed devices with the actual devices_kobj 
> that we used set in fs_devices->seed, and thus we'll leak the sysfs 
> objects for the seed devices.  Thanks,

Do you mean leaking the devinfo_kobj instead of devices_kobj? If so,
then yes, and this patch fixed it as well (I just found out).

Otherwise, no, there isn't devices_kobj leak.  We make sure only mounted
fsid has the devices_kobj initialized. So during sprouting- the
devices_kobj remains with the fs_info->fs_devices, as we move the seed
devices below the seed_devices.

static int btrfs_prepare_sprout(struct btrfs_fs_info *fs_info)
::
  list_splice_init_rcu(&fs_devices->devices, &seed_devices->devices,
  synchronize_rcu);
  list_for_each_entry(device, &seed_devices->devices, dev_list)
  device->fs_devices = seed_devices;

And during unmount, we clean up the dev links and devices_kobj.

close_ctree()
   btrfs_sysfs_remove_mounted()
   btrfs_sysfs_remove_fsid()

Anand

> Josef