Re: [PATCH] nfsd: free nfsd_file by gc after adding it to lru list

[Li Lingfeng <lilingfeng3@huawei.com>]

在 2025/1/15 23:27, Jeff Layton 写道:
> On Wed, 2025-01-15 at 10:03 -0500, Chuck Lever wrote:
>> On 1/14/25 2:39 PM, Jeff Layton wrote:
>>> On Tue, 2025-01-14 at 14:27 -0500, Jeff Layton wrote:
>>>> On Mon, 2025-01-13 at 10:59 +0800, Li Lingfeng wrote:
>>>>> In nfsd_file_put, after inserting the nfsd_file into the nfsd_file_lru
>>>>> list, gc may be triggered in another thread and immediately release this
>>>>> nfsd_file, which will lead to a UAF when accessing this nfsd_file again.
>>>>>
>>>>> All the places where unhash is done will also perform lru_remove, so there
>>>>> is no need to do lru_remove separately here. After inserting the nfsd_file
>>>>> into the nfsd_file_lru list, it can be released by relying on gc.
>>>>>
>>>>> Fixes: 4a0e73e635e3 ("NFSD: Leave open files out of the filecache LRU")
>>>>> Signed-off-by: Li Lingfeng <lilingfeng3@huawei.com>
>>>>> ---
>>>>>    fs/nfsd/filecache.c | 12 ++----------
>>>>>    1 file changed, 2 insertions(+), 10 deletions(-)
>>>>>
>>>>> diff --git a/fs/nfsd/filecache.c b/fs/nfsd/filecache.c
>>>>> index a1cdba42c4fa..37b65cb1579a 100644
>>>>> --- a/fs/nfsd/filecache.c
>>>>> +++ b/fs/nfsd/filecache.c
>>>>> @@ -372,18 +372,10 @@ nfsd_file_put(struct nfsd_file *nf)
>>>>>    		/* Try to add it to the LRU.  If that fails, decrement. */
>>>>>    		if (nfsd_file_lru_add(nf)) {
>>>>>    			/* If it's still hashed, we're done */
>>>>> -			if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
>>>>> +			if (list_lru_count(&nfsd_file_lru))
>>>>>    				nfsd_file_schedule_laundrette();
>>>>> -				return;
>>>>> -			}
>>>>>    
>>>>> -			/*
>>>>> -			 * We're racing with unhashing, so try to remove it from
>>>>> -			 * the LRU. If removal fails, then someone else already
>>>>> -			 * has our reference.
>>>>> -			 */
>>>>> -			if (!nfsd_file_lru_remove(nf))
>>>>> -				return;
>>>>> +			return;
>>>>>    		}
>>>>>    	}
>>>>>    	if (refcount_dec_and_test(&nf->nf_ref))
>>>> I think this looks OK. Filecache bugs are particularly nasty though, so
>>>> let's run this through a nice long testing cycle.
>>>>
>>>> Reviewed-by: Jeff Layton <jlayton@kernel.org>
>>> Actually, I take it back. This is problematic in another way.
>>>
>>> In this case, we're racing with another task that is unhashing the
>>> object, but we've put it on the LRU ourselves. What guarantee do we
>>> have that the unhashing and removal from the LRU didn't occur before
>>> this task called nfsd_file_lru_add()? That's why we attempt to remove
>>> it here -- we can't rely on the task that unhashed it to do so at that
>>> point.
>>>
>>> What might be best is to take and hold the rcu_read_lock() before doing
>>> the nfsd_file_lru_add, and just release it after we do these racy
>>> checks. That should make it safe to access the object.
>>>
>>> Thoughts?
>> Holding the RCU read lock will keep the dereferences safe since
>> nfsd_file objects are freed only after an RCU grace period. But will the
>> logic in nfsd_file_put() work properly on totally dead nfsd_file
>> objects? I don't see a specific failure mode there, but I'm not very
>> imaginative.
>>
>> Overall, I think RCU would help.
>>
> It should be safe to call nfsd_file_lru_add() with the rcu_read_lock()
> held. After that we're just looking at the nf_flags() and the nf_lru
> list head. On a dead file, HASHED will be clear and the
> nfsd_file_lru_remove() call will be a no-op (the list_head will be
> empty).
>
> Li Lingfeng, do you want to propose a patch for this? Unfortunately,
> your reproducer won't work after that, since you can't sleep with the
> rcu_read_lock held.
Sorry for the delay.
Of course I'm willing to do it.
Maybe I can reproduce the problem by changing msleep to mdelay? I will try.