Re: [PATCH 2/2] hfsplus: add a check for hfs_bnode_find

[Viacheslav Dubeyko]

Sorry, I had some glitch during message sending. I am repeating the message sending.

> On Oct 17, 2019, at 11:52 PM, Ernesto A. FernÃndez <ernesto.mnd.fernandez@xxxxxxxxx> wrote:
> 
> On Thu, Oct 17, 2019 at 09:30:20AM +0800, Chuhong Yuan wrote:
>> On Thu, Oct 17, 2019 at 8:07 AM Ernesto A. FernÃndez
>> <ernesto.mnd.fernandez@xxxxxxxxx> wrote:
>>> 
>>> Hi,
>>> 
>>> On Wed, Oct 16, 2019 at 08:06:20PM +0800, Chuhong Yuan wrote:
>>>> hfs_brec_update_parent misses a check for hfs_bnode_find and may miss
>>>> the failure.
>>>> Add a check for it like what is done in again.
>>>> 
>>>> Signed-off-by: Chuhong Yuan <hslester96@xxxxxxxxx>
>>>> ---
>>>> fs/hfsplus/brec.c | 2 ++
>>>> 1 file changed, 2 insertions(+)
>>>> 
>>>> diff --git a/fs/hfsplus/brec.c b/fs/hfsplus/brec.c
>>>> index 1918544a7871..22bada8288c4 100644
>>>> --- a/fs/hfsplus/brec.c
>>>> +++ b/fs/hfsplus/brec.c
>>>> @@ -434,6 +434,8 @@ static int hfs_brec_update_parent(struct hfs_find_data *fd)
>>>>                      new_node->parent = tree->root;
>>>>              }
>>>>              fd->bnode = hfs_bnode_find(tree, new_node->parent);
>>>> +             if (IS_ERR(fd->bnode))
>>>> +                     return PTR_ERR(fd->bnode);
>>> 
>>> You shouldn't just return here, you still hold a reference to new_node.
>>> The call to hfs_bnode_find() after the again label seems to be making a
>>> similar mistake.
>>> 
>>> I don't think either one can actually fail though, because the parent
>>> nodes have all been read and hashed before, haven't they?
>>> 
>> 
>> I find that after hfs_bnode_findhash in hfs_bnode_find, there is a test for
>> HFS_BNODE_ERROR and may return an error. I'm not sure whether it
>> can happen here.
> 
> That would require a race between hfs_bnode_find() and hfs_bnode_create(),
> but the node has already been created.
> 

The whole function hfs_brec_update_parent() looks like the cycle. And there are several
places where PTR_ERR(node) is returned with error ([1] - [2]). So, it sounds that it needs
to follow this pattern or to rework these cases too. And, by the way, what if the node pointer
will be NULL?

Thanks,
Viacheslav Dubeyko.

[1] https://elixir.bootlin.com/linux/latest/source/fs/hfsplus/brec.c#L371
[2] https://elixir.bootlin.com/linux/latest/source/fs/hfsplus/brec.c#L402


>> 
>>>>              /* create index key and entry */
>>>>              hfs_bnode_read_key(new_node, fd->search_key, 14);
>>>>              cnid = cpu_to_be32(new_node->this);
>>>> --
>>>> 2.20.1