Re: [PATCH v2 03/12] cachefiles: fix slab-use-after-free in cachefiles_ondemand_get_fd()

[Baokun Li <libaokun@huaweicloud.com>]

Hi Jingbo,

Thanks for your review!

On 2024/5/20 15:24, Jingbo Xu wrote:
>
> On 5/15/24 4:45 PM, libaokun@huaweicloud.com wrote:
>> From: Baokun Li <libaokun1@huawei.com>
>>
>> We got the following issue in a fuzz test of randomly issuing the restore
>> command:
>>
>> ==================================================================
>> BUG: KASAN: slab-use-after-free in cachefiles_ondemand_daemon_read+0x609/0xab0
>> Write of size 4 at addr ffff888109164a80 by task ondemand-04-dae/4962
>>
>> CPU: 11 PID: 4962 Comm: ondemand-04-dae Not tainted 6.8.0-rc7-dirty #542
>> Call Trace:
>>   kasan_report+0x94/0xc0
>>   cachefiles_ondemand_daemon_read+0x609/0xab0
>>   vfs_read+0x169/0xb50
>>   ksys_read+0xf5/0x1e0
>>
>> Allocated by task 626:
>>   __kmalloc+0x1df/0x4b0
>>   cachefiles_ondemand_send_req+0x24d/0x690
>>   cachefiles_create_tmpfile+0x249/0xb30
>>   cachefiles_create_file+0x6f/0x140
>>   cachefiles_look_up_object+0x29c/0xa60
>>   cachefiles_lookup_cookie+0x37d/0xca0
>>   fscache_cookie_state_machine+0x43c/0x1230
>>   [...]
>>
>> Freed by task 626:
>>   kfree+0xf1/0x2c0
>>   cachefiles_ondemand_send_req+0x568/0x690
>>   cachefiles_create_tmpfile+0x249/0xb30
>>   cachefiles_create_file+0x6f/0x140
>>   cachefiles_look_up_object+0x29c/0xa60
>>   cachefiles_lookup_cookie+0x37d/0xca0
>>   fscache_cookie_state_machine+0x43c/0x1230
>>   [...]
>> ==================================================================
>>
>> Following is the process that triggers the issue:
>>
>>       mount  |   daemon_thread1    |    daemon_thread2
>> ------------------------------------------------------------
>>   cachefiles_ondemand_init_object
>>    cachefiles_ondemand_send_req
>>     REQ_A = kzalloc(sizeof(*req) + data_len)
>>     wait_for_completion(&REQ_A->done)
>>
>>              cachefiles_daemon_read
>>               cachefiles_ondemand_daemon_read
>>                REQ_A = cachefiles_ondemand_select_req
>>                cachefiles_ondemand_get_fd
>>                copy_to_user(_buffer, msg, n)
>>              process_open_req(REQ_A)
>>                                    ------ restore ------
>>                                    cachefiles_ondemand_restore
>>                                    xas_for_each(&xas, req, ULONG_MAX)
>>                                     xas_set_mark(&xas, CACHEFILES_REQ_NEW);
>>
>>                                    cachefiles_daemon_read
>>                                     cachefiles_ondemand_daemon_read
>>                                      REQ_A = cachefiles_ondemand_select_req
>>
>>               write(devfd, ("copen %u,%llu", msg->msg_id, size));
>>               cachefiles_ondemand_copen
>>                xa_erase(&cache->reqs, id)
>>                complete(&REQ_A->done)
>>     kfree(REQ_A)
>>                                      cachefiles_ondemand_get_fd(REQ_A)
>>                                       fd = get_unused_fd_flags
>>                                       file = anon_inode_getfile
>>                                       fd_install(fd, file)
>>                                       load = (void *)REQ_A->msg.data;
>>                                       load->fd = fd;
>>                                       // load UAF !!!
>>
>> This issue is caused by issuing a restore command when the daemon is still
>> alive, which results in a request being processed multiple times thus
>> triggering a UAF. So to avoid this problem, add an additional reference
>> count to cachefiles_req, which is held while waiting and reading, and then
>> released when the waiting and reading is over.
>>
>>
>> Note that since there is only one reference count for waiting, we need to
>> avoid the same request being completed multiple times, so we can only
>> complete the request if it is successfully removed from the xarray.
> Sorry the above description makes me confused.  As the same request may
> be got by different daemon threads multiple times, the introduced
> refcount mechanism can't protect it from being completed multiple times
> (which is expected).  The refcount only protects it from being freed
> multiple times.
The idea here is that because the wait only holds one reference count,
complete(&req->done) can only be called when the req has been
successfully removed from the xarry, otherwise the following UAF may
occur:

    daemon_thread1    |    daemon_thread2
-------------------------------------------
cachefiles_ondemand_daemon_read
  xa_lock(&cache->reqs)
  // select req_A
  xa_unlock(&cache->reqs)
                     // restore req_A and read again
                     cachefiles_ondemand_daemon_read
                     xa_lock(&cache->reqs)
                     // select req_A
                     xa_unlock(&cache->reqs)
// goto error, erase success
xa_erase(&cache->reqs, id)
complete(&req_A->done)
// free req_A
                     // goto error, erase failed
                     complete(&req_A->done)
                     // req_A use-after-free

This is also why error requests and CLOSE requests are handled
together and why xas_load(&xas) == req is checked.
>> Fixes: e73fa11a356c ("cachefiles: add restore command to recover inflight ondemand read requests")
>> Suggested-by: Hou Tao <houtao1@huawei.com>
>> Signed-off-by: Baokun Li <libaokun1@huawei.com>
>> Reviewed-by: Jia Zhu <zhujia.zj@bytedance.com>
>> ---
>>   fs/cachefiles/internal.h |  1 +
>>   fs/cachefiles/ondemand.c | 44 ++++++++++++++++++++++------------------
>>   2 files changed, 25 insertions(+), 20 deletions(-)
>>
>> diff --git a/fs/cachefiles/internal.h b/fs/cachefiles/internal.h
>> index d33169f0018b..7745b8abc3aa 100644
>> --- a/fs/cachefiles/internal.h
>> +++ b/fs/cachefiles/internal.h
>> @@ -138,6 +138,7 @@ static inline bool cachefiles_in_ondemand_mode(struct cachefiles_cache *cache)
>>   struct cachefiles_req {
>>   	struct cachefiles_object *object;
>>   	struct completion done;
>> +	refcount_t ref;
>>   	int error;
>>   	struct cachefiles_msg msg;
>>   };
>> diff --git a/fs/cachefiles/ondemand.c b/fs/cachefiles/ondemand.c
>> index fd49728d8bae..56d12fe4bf73 100644
>> --- a/fs/cachefiles/ondemand.c
>> +++ b/fs/cachefiles/ondemand.c
>> @@ -4,6 +4,12 @@
>>   #include <linux/uio.h>
>>   #include "internal.h"
>>   
>> +static inline void cachefiles_req_put(struct cachefiles_req *req)
>> +{
>> +	if (refcount_dec_and_test(&req->ref))
>> +		kfree(req);
>> +}
>> +
>>   static int cachefiles_ondemand_fd_release(struct inode *inode,
>>   					  struct file *file)
>>   {
>> @@ -299,7 +305,6 @@ ssize_t cachefiles_ondemand_daemon_read(struct cachefiles_cache *cache,
>>   {
>>   	struct cachefiles_req *req;
>>   	struct cachefiles_msg *msg;
>> -	unsigned long id = 0;
>>   	size_t n;
>>   	int ret = 0;
>>   	XA_STATE(xas, &cache->reqs, cache->req_id_next);
>> @@ -330,41 +335,39 @@ ssize_t cachefiles_ondemand_daemon_read(struct cachefiles_cache *cache,
>>   
>>   	xas_clear_mark(&xas, CACHEFILES_REQ_NEW);
>>   	cache->req_id_next = xas.xa_index + 1;
>> +	refcount_inc(&req->ref);
>>   	xa_unlock(&cache->reqs);
>>   
>> -	id = xas.xa_index;
>> -
>>   	if (msg->opcode == CACHEFILES_OP_OPEN) {
>>   		ret = cachefiles_ondemand_get_fd(req);
>>   		if (ret) {
>>   			cachefiles_ondemand_set_object_close(req->object);
>> -			goto error;
>> +			goto out;
>>   		}
>>   	}
>>   
>> -	msg->msg_id = id;
>> +	msg->msg_id = xas.xa_index;
>>   	msg->object_id = req->object->ondemand->ondemand_id;
>>   
>>   	if (copy_to_user(_buffer, msg, n) != 0) {
>>   		ret = -EFAULT;
>>   		if (msg->opcode == CACHEFILES_OP_OPEN)
>>   			close_fd(((struct cachefiles_open *)msg->data)->fd);
>> -		goto error;
>>   	}
>> -
>> -	/* CLOSE request has no reply */
>> -	if (msg->opcode == CACHEFILES_OP_CLOSE) {
>> -		xa_erase(&cache->reqs, id);
>> -		complete(&req->done);
>> +out:
>> +	/* Remove error request and CLOSE request has no reply */
>> +	if (ret || msg->opcode == CACHEFILES_OP_CLOSE) {
>> +		xas_reset(&xas);
>> +		xas_lock(&xas);
>> +		if (xas_load(&xas) == req) {
> Just out of curiosity... How could xas_load(&xas) doesn't equal to req?

As mentioned above, the req may have been deleted or even the id

may have been reused.

>
>> +			req->error = ret;
>> +			complete(&req->done);
>> +			xas_store(&xas, NULL);
>> +		}
>> +		xas_unlock(&xas);
>>   	}
>> -
>> -	return n;
>> -
>> -error:
>> -	xa_erase(&cache->reqs, id);
>> -	req->error = ret;
>> -	complete(&req->done);
>> -	return ret;
>> +	cachefiles_req_put(req);
>> +	return ret ? ret : n;
>>   }
> This is actually a combination of a fix and a cleanup which combines the
> logic of removing error request and the CLOSE requests into one place.
> Also it relies on the cleanup made in patch 2 ("cachefiles: remove
> err_put_fd tag in cachefiles_ondemand_daemon_read()"), making it
> difficult to be atomatically back ported to the stable (as patch 2 is
> not marked as "Fixes").
>
> Thus could we make the fix first, and then make the cleanup.
I don't think that's necessary, stable automatically backports the
relevant dependency patches in case of backport patch conflicts,
and later patches modify the logic here as well.
Or add Fixes tag for patch 2?
>>   
>>   typedef int (*init_req_fn)(struct cachefiles_req *req, void *private);
>> @@ -394,6 +397,7 @@ static int cachefiles_ondemand_send_req(struct cachefiles_object *object,
>>   		goto out;
>>   	}
>>   
>> +	refcount_set(&req->ref, 1);
>>   	req->object = object;
>>   	init_completion(&req->done);
>>   	req->msg.opcode = opcode;
>> @@ -455,7 +459,7 @@ static int cachefiles_ondemand_send_req(struct cachefiles_object *object,
>>   	wake_up_all(&cache->daemon_pollwq);
>>   	wait_for_completion(&req->done);
>>   	ret = req->error;
>> -	kfree(req);
>> +	cachefiles_req_put(req);
>>   	return ret;
>>   out:
>>   	/* Reset the object to close state in error handling path.
>
> Don't we need to also convert "kfree(req)" to cachefiles_req_put(req)
> for the error path of cachefiles_ondemand_send_req()?
>
> ```
> out:
> 	/* Reset the object to close state in error handling path.
> 	 * If error occurs after creating the anonymous fd,
> 	 * cachefiles_ondemand_fd_release() will set object to close.
> 	 */
> 	if (opcode == CACHEFILES_OP_OPEN)
> 		cachefiles_ondemand_set_object_close(object);
> 	kfree(req);
> 	return ret;
> ```
When "goto out;" is called in cachefiles_ondemand_send_req(),
it means that the req is unallocated/failed to be allocated/failed to
be inserted into the xarry, and therefore the req can only be accessed
by the current function, so there is no need to consider concurrency
and reference counting.

Thanks!

-- 
With Best Regards,
Baokun Li