Date: Mon, 20 May 2024 22:56:23 +0800
From: Gao Xiang <hsiangkao@...ux.alibaba.com>
To: Baokun Li <libaokun@...weicloud.com>, Jeff Layton <jlayton@...nel.org>,
 netfs@...ts.linux.dev, dhowells@...hat.com
Cc: jefflexu@...ux.alibaba.com, zhujia.zj@...edance.com,
 linux-erofs@...ts.ozlabs.org, linux-fsdevel@...r.kernel.org,
 linux-kernel@...r.kernel.org, yangerkun@...wei.com, houtao1@...wei.com,
 yukuai3@...wei.com, wozizhi@...wei.com, Baokun Li <libaokun1@...wei.com>
Subject: Re: [PATCH v2 4/5] cachefiles: cyclic allocation of msg_id to avoid
 reuse

Hi Baokun,

On 2024/5/20 21:24, Baokun Li wrote:
> On 2024/5/20 20:54, Gao Xiang wrote:
>>
>>
>> On 2024/5/20 20:42, Baokun Li wrote:
>>> On 2024/5/20 18:04, Jeff Layton wrote:
>>>> On Mon, 2024-05-20 at 12:06 +0800, Baokun Li wrote:
>>>>> Hi Jeff,
>>>>>
>>>>> Thank you very much for your review!
>>>>>
>>>>> On 2024/5/19 19:11, Jeff Layton wrote:
>>>>>> On Wed, 2024-05-15 at 20:51 +0800, libaokun@...weicloud.com wrote:
>>>>>>> From: Baokun Li <libaokun1@...wei.com>
>>>>>>>
>>>>>>> Reusing the msg_id after a maliciously completed reopen request may cause
>>>>>>> a read request to remain unprocessed and result in a hung, as shown below:
>>>>>>>
>>>>>>>          t1       |      t2       |      t3
>>>>>>> -------------------------------------------------
>>>>>>> cachefiles_ondemand_select_req
>>>>>>>    cachefiles_ondemand_object_is_close(A)
>>>>>>>    cachefiles_ondemand_set_object_reopening(A)
>>>>>>>    queue_work(fscache_object_wq, &info->work)
>>>>>>>                   ondemand_object_worker
>>>>>>>                    cachefiles_ondemand_init_object(A)
>>>>>>>                     cachefiles_ondemand_send_req(OPEN)
>>>>>>>                       // get msg_id 6
>>>>>>> wait_for_completion(&req_A->done)
>>>>>>> cachefiles_ondemand_daemon_read
>>>>>>>    // read msg_id 6 req_A
>>>>>>>    cachefiles_ondemand_get_fd
>>>>>>>    copy_to_user
>>>>>>>                                   // Malicious completion msg_id 6
>>>>>>>                                   copen 6,-1
>>>>>>> cachefiles_ondemand_copen
>>>>>>> complete(&req_A->done)
>>>>>>>                                    // will not set the object to close
>>>>>>>                                    // because ondemand_id && fd is valid.
>>>>>>>
>>>>>>>                   // ondemand_object_worker() is done
>>>>>>>                   // but the object is still reopening.
>>>>>>>
>>>>>>>                                   // new open req_B
>>>>>>> cachefiles_ondemand_init_object(B)
>>>>>>> cachefiles_ondemand_send_req(OPEN)
>>>>>>>                                    // reuse msg_id 6
>>>>>>> process_open_req
>>>>>>>    copen 6,A.size
>>>>>>>    // The expected failed copen was executed successfully
>>>>>>>
>>>>>>> Expect copen to fail, and when it does, it closes fd, which sets the
>>>>>>> object to close, and then close triggers reopen again. However, due to
>>>>>>> msg_id reuse resulting in a successful copen, the anonymous fd is not
>>>>>>> closed until the daemon exits. Therefore read requests waiting for reopen
>>>>>>> to complete may trigger hung task.
>>>>>>>
>>>>>>> To avoid this issue, allocate the msg_id cyclically to avoid reusing the
>>>>>>> msg_id for a very short duration of time.
>>>>>>>
>>>>>>> Fixes: c8383054506c ("cachefiles: notify the user daemon when looking up cookie")
>>>>>>> Signed-off-by: Baokun Li <libaokun1@...wei.com>
>>>>>>> ---
>>>>>>>    fs/cachefiles/internal.h |  1 +
>>>>>>>    fs/cachefiles/ondemand.c | 20 ++++++++++++++++----
>>>>>>>    2 files changed, 17 insertions(+), 4 deletions(-)
>>>>>>>
>>>>>>> diff --git a/fs/cachefiles/internal.h b/fs/cachefiles/internal.h
>>>>>>> index 8ecd296cc1c4..9200c00f3e98 100644
>>>>>>> --- a/fs/cachefiles/internal.h
>>>>>>> +++ b/fs/cachefiles/internal.h
>>>>>>> @@ -128,6 +128,7 @@ struct cachefiles_cache {
>>>>>>>        unsigned long            req_id_next;
>>>>>>>        struct xarray            ondemand_ids;    /* xarray for ondemand_id allocation */
>>>>>>>        u32                ondemand_id_next;
>>>>>>> +    u32                msg_id_next;
>>>>>>>    };
>>>>>>>    static inline bool cachefiles_in_ondemand_mode(struct cachefiles_cache *cache)
>>>>>>> diff --git a/fs/cachefiles/ondemand.c b/fs/cachefiles/ondemand.c
>>>>>>> index f6440b3e7368..b10952f77472 100644
>>>>>>> --- a/fs/cachefiles/ondemand.c
>>>>>>> +++ b/fs/cachefiles/ondemand.c
>>>>>>> @@ -433,20 +433,32 @@ static int cachefiles_ondemand_send_req(struct cachefiles_object *object,
>>>>>>>            smp_mb();
>>>>>>>            if (opcode == CACHEFILES_OP_CLOSE &&
>>>>>>> - !cachefiles_ondemand_object_is_open(object)) {
>>>>>>> + !cachefiles_ondemand_object_is_open(object)) {
>>>>>>> WARN_ON_ONCE(object->ondemand->ondemand_id == 0);
>>>>>>>                xas_unlock(&xas);
>>>>>>>                ret = -EIO;
>>>>>>>                goto out;
>>>>>>>            }
>>>>>>> -        xas.xa_index = 0;
>>>>>>> +        /*
>>>>>>> +         * Cyclically find a free xas to avoid msg_id reuse that would
>>>>>>> +         * cause the daemon to successfully copen a stale msg_id.
>>>>>>> +         */
>>>>>>> +        xas.xa_index = cache->msg_id_next;
>>>>>>>            xas_find_marked(&xas, UINT_MAX, XA_FREE_MARK);
>>>>>>> +        if (xas.xa_node == XAS_RESTART) {
>>>>>>> +            xas.xa_index = 0;
>>>>>>> +            xas_find_marked(&xas, cache->msg_id_next - 1, XA_FREE_MARK);
>>>>>>> +        }
>>>>>>>            if (xas.xa_node == XAS_RESTART)
>>>>>>>                xas_set_err(&xas, -EBUSY);
>>>>>>> +
>>>>>>>            xas_store(&xas, req);
>>>>>>> -        xas_clear_mark(&xas, XA_FREE_MARK);
>>>>>>> -        xas_set_mark(&xas, CACHEFILES_REQ_NEW);
>>>>>>> +        if (xas_valid(&xas)) {
>>>>>>> +            cache->msg_id_next = xas.xa_index + 1;
>>>>>> If you have a long-standing stuck request, could this counter wrap
>>>>>> around and you still end up with reuse?
>>>>> Yes, msg_id_next is declared to be of type u32 in the hope that when
>>>>> xa_index == UINT_MAX, a wrap around occurs so that msg_id_next
>>>>> goes to zero. Limiting xa_index to no more than UINT_MAX is to avoid
>>>>> the xarry being too deep.
>>>>>
>>>>> If msg_id_next is equal to the id of a long-standing stuck request
>>>>> after the wrap-around, it is true that the reuse in the above problem
>>>>> may also occur.
>>>>>
>>>>> But I feel that a long stuck request is problematic in itself, it means
>>>>> that after we have sent 4294967295 requests, the first one has not
>>>>> been processed yet, and even if we send a million requests per
>>>>> second, this one hasn't been completed for more than an hour.
>>>>>
>>>>> We have a keep-alive process that pulls the daemon back up as
>>>>> soon as it exits, and there is a timeout mechanism for requests in
>>>>> the daemon to prevent the kernel from waiting for long periods
>>>>> of time. In other words, we should avoid the situation where
>>>>> a request is stuck for a long period of time.
>>>>>
>>>>> If you think UINT_MAX is not enough, perhaps we could raise
>>>>> the maximum value of msg_id_next to ULONG_MAX?
>>>>>> Maybe this should be using
>>>>>> ida_alloc/free instead, which would prevent that too?
>>>>>>
>>>>> The id reuse here is that the kernel has finished the open request
>>>>> req_A and freed its id_A and used it again when sending the open
>>>>> request req_B, but the daemon is still working on req_A, so the
>>>>> copen id_A succeeds but operates on req_B.
>>>>>
>>>>> The id that is being used by the kernel will not be allocated here
>>>>> so it seems that ida _alloc/free does not prevent reuse either,
>>>>> could you elaborate a bit more how this works?
>>>>>
>>>> ida_alloc and free absolutely prevent reuse while the id is in use.
>>>> That's sort of the point of those functions. Basically it uses a set of
>>>> bitmaps in an xarray to track which IDs are in use, so ida_alloc only
>>>> hands out values which are not in use. See the comments over
>>>> ida_alloc_range() in lib/idr.c.
>>>>
>>> Thank you for the explanation!
>>>
>>> The logic now provides the same guarantees as ida_alloc/free.
>>> The "reused" id, indeed, is no longer in use in the kernel, but it is still
>>> in use in the userland, so a multi-threaded daemon could be handling
>>> two different requests for the same msg_id at the same time.
>>>
>>> Previously, the logic for allocating msg_ids was to start at 0 and look
>>> for a free xas.index, so it was possible for an id to be allocated to a
>>> new request just as the id was being freed.
>>>
>>> With the change to cyclic allocation, the kernel will not use the same
>>> id again until INT_MAX requests have been sent, and during the time
>>> it takes to send requests, the daemon has enough time to process
>>> requests whose ids are still in use by the daemon, but have already
>>> been freed in the kernel.
>>
>> Again, If I understand correctly, I think the main point
>> here is
>>
>> wait_for_completion(&req_A->done)
>>
>> which could hang due to some malicious deamon.  But I think it
>> should be switched to wait_for_completion_killable() instead. *
>> It's up to users to kill the mount instance if there is a
>> malicious user daemon.
>>
>> So in that case, hung task will not be triggered anymore, and
>> you don't need to care about cyclic allocation too.
>>
>> Thanks,
>> Gao Xiang
> Hi Xiang,
> 
> The problem is not as simple as you think.
> 
> If you make it killable, it just won't trigger a hung task in
> cachefiles_ondemand_send_req(), and the process waiting for the
> resource in question will also be hung.
> 
> * When the open/read request in the mount process gets stuck,
>    the sync/drop cache will trigger a hung task panic in iterate_supers()
>    as it waits for sb->umount to be unlocked.
> * After umount, anonymous fd is not closed causing a hung task panic
>    in fscache_hash_cookie() because of waiting for cookie unhash.
> * The dentry is in a loop up state, because the read request is not being
>    processed, another process looking for the same dentry is waiting for
>    the previous lookup to finish, which triggers a hung task panic in
>    d_alloc_parallel().


As for your sb->umount, d_alloc_parallel() or even i_rwsem,
which are all currently unkillable, also see some previous
threads like:

https://lore.kernel.org/linux-fsdevel/CAJfpegu6v1fRAyLvFLOPUSAhx5aAGvPGjBWv-TDQjugqjUA_hQ@mail.gmail.com/T/#u

I don't think it's the issue of on-demand cachefiles, even
NVMe or virtio-blk or networking can be stuck in
lookup, fill_sb or whatever.

Which can makes sb->umount, d_alloc_parallel() or even
i_rwsem unkillable.

> 
> Can all this be made killable?

I can understand your hung_task_panic concern but it
sounds like a workaround to me anyway.

Thanks,
Gao Xiang

> 
> Thanks,
> Baokun
>>
>>>
>>> Regards,
>>> Baokun
>>>>>>> + xas_clear_mark(&xas, XA_FREE_MARK);
>>>>>>> +            xas_set_mark(&xas, CACHEFILES_REQ_NEW);
>>>>>>> +        }
>>>>>>>            xas_unlock(&xas);
>>>>>>>        } while (xas_nomem(&xas, GFP_KERNEL));
>>>>>>>

Powered by blists - more mailing lists