On Tue, Mar 10, 2020 at 11:25:30AM +0200, Leon Romanovsky wrote:
> From: Leon Romanovsky <leonro@xxxxxxxxxxxx>
>
> >From Jason:
>
> cm_id.state is a non-atomic value that must always be read and written
> under lock, or while the thread has the only pointer to the cm_id.
>
> Critically, during MAD handling the cm_id.state is used to control when
> MAD handlers can run, and in turn what data they can touch. Without
> locking, an assignment to state can immediately allow concurrent MAD
> handlers to execute, potentially creating a mess.
>
> Several of these cases only risk load/store tearing, but create very
> confusing code. For instance changing the state from IB_CM_IDLE to
> IB_CM_LISTEN doesn't allow any MAD handlers to run in either state, but a
> superficial audit would suggest that it is not locked properly.
>
> This loose methodology has allowed two bugs to creep in. After creating an
> ID the code did not lock the state transition, apparently mistakenly
> assuming that the new ID could not be used concurrently. However, the ID
> is immediately placed in the xarray and so a carefully crafted network
> sequence could trigger races with the unlocked stores.
>
> The main solution to many of these problems is to use the xarray to create
> a two stage add - the first reserves the ID and the second publishes the
> pointer. The second stage is either omitted entirely or moved after the
> newly created ID is setup.
>
> Where it is trivial to do so other places directly put the state
> manipulation under lock, or add an assertion that it is, in fact, under
> lock.
>
> This also removes a number of places where the state is being read under
> lock, then the lock dropped, reacquired and state tested again.
>
> There remain other issues related to missing locking on cm_id data.
>
> Thanks
>
> It is based on rdma-next + rdma-rc patch c14dfddbd869
> ("RMDA/cm: Fix missing ib_cm_destroy_id() in ib_cm_insert_listen()")
>
> Jason Gunthorpe (15):
> RDMA/cm: Fix ordering of xa_alloc_cyclic() in ib_create_cm_id()
> RDMA/cm: Fix checking for allowed duplicate listens
> RDMA/cm: Remove a race freeing timewait_info
> RDMA/cm: Make the destroy_id flow more robust
> RDMA/cm: Simplify establishing a listen cm_id
> RDMA/cm: Read id.state under lock when doing pr_debug()
> RDMA/cm: Make it clear that there is no concurrency in
> cm_sidr_req_handler()
> RDMA/cm: Make it clearer how concurrency works in cm_req_handler()
> RDMA/cm: Add missing locking around id.state in cm_dup_req_handler
> RDMA/cm: Add some lockdep assertions for cm_id_priv->lock
> RDMA/cm: Allow ib_send_cm_dreq() to be done under lock
> RDMA/cm: Allow ib_send_cm_drep() to be done under lock
> RDMA/cm: Allow ib_send_cm_rej() to be done under lock
> RDMA/cm: Allow ib_send_cm_sidr_rep() to be done under lock
> RDMA/cm: Make sure the cm_id is in the IB_CM_IDLE state in destroy
Applied to for-next.
Jason
