Re: [PATCH] ceph: fix memory leak in mount error path when using test_dummy_encryption

[Xiubo Li <xiubli@redhat.com>]

On 07/11/2022 22:19, Luís Henriques wrote:
> On Mon, Nov 07, 2022 at 07:06:40PM +0800, Xiubo Li wrote:
>> On 07/11/2022 18:23, Luís Henriques wrote:
>>> On Mon, Nov 07, 2022 at 03:47:23PM +0800, Xiubo Li wrote:
>>>> On 03/11/2022 23:36, Luís Henriques wrote:
>>>>> Because ceph_init_fs_context() will never be invoced in case we get a
>>>>> mount error, destroy_mount_options() won't be releasing fscrypt resources
>>>>> with fscrypt_free_dummy_policy().  This will result in a memory leak.  Add
>>>>> an invocation to this function in the mount error path.
>>>>>
>>>>> Signed-off-by: Luís Henriques <lhenriques@suse.de>
>>>>> ---
>>>>>     fs/ceph/super.c | 1 +
>>>>>     1 file changed, 1 insertion(+)
>>>>>
>>>>> diff --git a/fs/ceph/super.c b/fs/ceph/super.c
>>>>> index 2224d44d21c0..6b9fd04b25cd 100644
>>>>> --- a/fs/ceph/super.c
>>>>> +++ b/fs/ceph/super.c
>>>>> @@ -1362,6 +1362,7 @@ static int ceph_get_tree(struct fs_context *fc)
>>>>>     	ceph_mdsc_close_sessions(fsc->mdsc);
>>>>>     	deactivate_locked_super(sb);
>>>>> +	fscrypt_free_dummy_policy(&fsc->fsc_dummy_enc_policy);
>>>> Hi Luis,
>>>>
>>>> BTW, any reason the following code won't be triggered ?
>>>>
>>>> deactivate_locked_super(sb);
>>>>
>>>>     --> fs->kill_sb(s);
>>>>
>>>>           --> ceph_kill_sb()
>>>>
>>>>                 --> kill_anon_super()
>>>>
>>>>                       --> generic_shutdown_super()
>>>>
>>>>                             --> sop->put_super()
>>>>
>>>>                                   --> ceph_put_super()
>>>>
>>>>                                         --> ceph_fscrypt_free_dummy_policy()
>>>>
>>>>                                              --> fscrypt_free_dummy_policy(
>>>>
>>> Here's what I'm seeing here:
>>>
>>>     sys_mount->path_mount->do_new_mount->vfs_get_tree->ceph_get_tree
>>>
>>> ceph_get_tree() fails due to ceph_real_mount() returning an error.  My
>>> understanding is that that, since fc->root is never set, that code path
>>> will never be triggered.  Does that make sense?
>> Okay, you are right!
>>
>> How about fixing it in ceph_real_mount() instead ?
> Sure, I can send a patch for doing that instead.  However, my opinion is
> that it makes more sense to do it, mostly because ceph_get_tree() is
> already doing clean-up work on the error path (ceph_mdsc_close_sessions()
> and deactivate_locked_super()).
>
> But let me know if you really prefer doing in ceph_read_mount() and I'll
> send v2.

IMO it'd better to do this in ceph_real_mount(), which will make the 
code to be easier to read.

Thanks!

- Xiubo

>>> An easy way to reproduce is by running fstest ceph/005 with the
>>> 'test_dummy_encryption' option.  (I'll probably need to send a patch to
>>> disable this test when this option is present.)
>> Anyway this should be fixed in kceph.
> Yes, agreed.
>
> Cheers,
> --
> Luís
>