Re: [PATCH rc] RDMA/cma: Ensure rdma_addr_cancel() happens before issuing more requests

[Haakon Bugge <haakon.bugge@oracle.com>]

> On 22 Sep 2021, at 10:01, Leon Romanovsky <leon@kernel.org> wrote:
> 
> On Thu, Sep 16, 2021 at 03:34:46PM -0300, Jason Gunthorpe wrote:
>> The FSM can run in a circle allowing rdma_resolve_ip() to be called twice
>> on the same id_priv. While this cannot happen without going through the
>> work, it violates the invariant that the same address resolution
>> background request cannot be active twice.
>> 
>>       CPU 1                                  CPU 2
>> 
>> rdma_resolve_addr():
>>  RDMA_CM_IDLE -> RDMA_CM_ADDR_QUERY
>>  rdma_resolve_ip(addr_handler)  #1
>> 
>> 			 process_one_req(): for #1
>>                          addr_handler():
>>                            RDMA_CM_ADDR_QUERY -> RDMA_CM_ADDR_BOUND
>>                            mutex_unlock(&id_priv->handler_mutex);
>>                            [.. handler still running ..]
>> 
>> rdma_resolve_addr():
>>  RDMA_CM_ADDR_BOUND -> RDMA_CM_ADDR_QUERY
>>  rdma_resolve_ip(addr_handler)
>>    !! two requests are now on the req_list
>> 
>> rdma_destroy_id():
>> destroy_id_handler_unlock():
>>  _destroy_id():
>>   cma_cancel_operation():
>>    rdma_addr_cancel()
>> 
>>                          // process_one_req() self removes it
>> 		          spin_lock_bh(&lock);
>>                           cancel_delayed_work(&req->work);
>> 	                   if (!list_empty(&req->list)) == true
>> 
>>      ! rdma_addr_cancel() returns after process_on_req #1 is done
>> 
>>   kfree(id_priv)
>> 
>> 			 process_one_req(): for #2
>>                          addr_handler():
>> 	                    mutex_lock(&id_priv->handler_mutex);
>>                            !! Use after free on id_priv
>> 
>> rdma_addr_cancel() expects there to be one req on the list and only
>> cancels the first one. The self-removal behavior of the work only happens
>> after the handler has returned. This yields a situations where the
>> req_list can have two reqs for the same "handle" but rdma_addr_cancel()
>> only cancels the first one.
>> 
>> The second req remains active beyond rdma_destroy_id() and will
>> use-after-free id_priv once it inevitably triggers.
>> 
>> Fix this by remembering if the id_priv has called rdma_resolve_ip() and
>> always cancel before calling it again. This ensures the req_list never
>> gets more than one item in it and doesn't cost anything in the normal flow
>> that never uses this strange error path.
>> 
>> Cc: stable@vger.kernel.org
>> Fixes: e51060f08a61 ("IB: IP address based RDMA connection manager")
>> Reported-by: syzbot+dc3dfba010d7671e05f5@syzkaller.appspotmail.com
>> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
>> ---
>> drivers/infiniband/core/cma.c      | 17 +++++++++++++++++
>> drivers/infiniband/core/cma_priv.h |  1 +
>> 2 files changed, 18 insertions(+)
>> 
>> diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c
>> index c40791baced588..751cf5ea25f296 100644
>> --- a/drivers/infiniband/core/cma.c
>> +++ b/drivers/infiniband/core/cma.c
>> @@ -1776,6 +1776,14 @@ static void cma_cancel_operation(struct rdma_id_private *id_priv,
>> {
>> 	switch (state) {
>> 	case RDMA_CM_ADDR_QUERY:
>> +		/*
>> +		 * We can avoid doing the rdma_addr_cancel() based on state,
>> +		 * only RDMA_CM_ADDR_QUERY has a work that could still execute.
>> +		 * Notice that the addr_handler work could still be exiting
>> +		 * outside this state, however due to the interaction with the
>> +		 * handler_mutex the work is guaranteed not to touch id_priv
>> +		 * during exit.
>> +		 */
>> 		rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
>> 		break;
>> 	case RDMA_CM_ROUTE_QUERY:
>> @@ -3413,6 +3421,15 @@ int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
>> 		if (dst_addr->sa_family == AF_IB) {
>> 			ret = cma_resolve_ib_addr(id_priv);
>> 		} else {
>> +			/* The FSM can return back to RDMA_CM_ADDR_BOUND after
>> +			 * rdma_resolve_ip() is called, eg through the error
>> +			 * path in addr_handler. If this happens the existing
>> +			 * request must be canceled before issuing a new one.
>> +			 */
>> +			if (id_priv->used_resolve_ip)
>> +				rdma_addr_cancel(&id->route.addr.dev_addr);
>> +			else
>> +				id_priv->used_resolve_ip = 1;
> 
> Why don't you never clear this field? If you assume that this is one lifetime
> event, can you please add a comment with an explanation "why"?

Adding to that, don't you need {READ,WRITE}_ONCE when accessing used_resolve_ip? Or will the write to it obtain global visibility because mutex_unlock(&ctx->mutex) is executed before any other context can read it?


Thxs, Håkon

> 
> Thanks
> 
>> 			ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
>> 					      &id->route.addr.dev_addr,
>> 					      timeout_ms, addr_handler,
>> diff --git a/drivers/infiniband/core/cma_priv.h b/drivers/infiniband/core/cma_priv.h
>> index 5c463da9984536..f92f101ea9818f 100644
>> --- a/drivers/infiniband/core/cma_priv.h
>> +++ b/drivers/infiniband/core/cma_priv.h
>> @@ -91,6 +91,7 @@ struct rdma_id_private {
>> 	u8			afonly;
>> 	u8			timeout;
>> 	u8			min_rnr_timer;
>> +	u8 used_resolve_ip;
>> 	enum ib_gid_type	gid_type;
>> 
>> 	/*
>> 
>> base-commit: ad17bbef3dd573da937816edc0ab84fed6a17fa6
>> -- 
>> 2.33.0