Re: [PATCHv3 1/1] RDMA/rxe: Fix a dead lock problem

[Yanjun Zhu <yanjun.zhu@linux.dev>]

在 2022/4/13 22:50, Yanjun Zhu 写道:
> 
> 在 2022/4/13 8:45, Jason Gunthorpe 写道:
>> On Wed, Apr 13, 2022 at 03:42:08AM -0400, yanjun.zhu@linux.dev wrote:
>>> From: Zhu Yanjun <yanjun.zhu@linux.dev>
>>>
>>> This is a dead lock problem.
>>> The xa_lock first is acquired in this:
>>>
>>> {SOFTIRQ-ON-W} state was registered at:
>>>
>>>    lock_acquire+0x1d2/0x5a0
>>>    _raw_spin_lock+0x33/0x80
>>>    __rxe_add_to_pool+0x183/0x230 [rdma_rxe]
>>>    __ib_alloc_pd+0xf9/0x550 [ib_core]
>>>    ib_mad_init_device+0x2d9/0xd20 [ib_core]
>>>    add_client_context+0x2fa/0x450 [ib_core]
>>>    enable_device_and_get+0x1b7/0x350 [ib_core]
>>>    ib_register_device+0x757/0xaf0 [ib_core]
>>>    rxe_register_device+0x2eb/0x390 [rdma_rxe]
>>>    rxe_net_add+0x83/0xc0 [rdma_rxe]
>>>    rxe_newlink+0x76/0x90 [rdma_rxe]
>>>    nldev_newlink+0x245/0x3e0 [ib_core]
>>>    rdma_nl_rcv_msg+0x2d4/0x790 [ib_core]
>>>    rdma_nl_rcv+0x1ca/0x3f0 [ib_core]
>>>    netlink_unicast+0x43b/0x640
>>>    netlink_sendmsg+0x7eb/0xc40
>>>    sock_sendmsg+0xe0/0x110
>>>    __sys_sendto+0x1d7/0x2b0
>>>    __x64_sys_sendto+0xdd/0x1b0
>>>    do_syscall_64+0x37/0x80
>>>    entry_SYSCALL_64_after_hwframe+0x44/0xae
>>>
>>> Then xa_lock is acquired in this:
>>>
>>> {IN-SOFTIRQ-W}:
>>>
>>> Call Trace:
>>>   <TASK>
>>>    dump_stack_lvl+0x44/0x57
>>>    mark_lock.part.52.cold.79+0x3c/0x46
>>>    __lock_acquire+0x1565/0x34a0
>>>    lock_acquire+0x1d2/0x5a0
>>>    _raw_spin_lock_irqsave+0x42/0x90
>>>    rxe_pool_get_index+0x72/0x1d0 [rdma_rxe]
>>>    rxe_get_av+0x168/0x2a0 [rdma_rxe]
>>>    rxe_requester+0x75b/0x4a90 [rdma_rxe]
>>>    rxe_do_task+0x134/0x230 [rdma_rxe]
>>>    tasklet_action_common.isra.12+0x1f7/0x2d0
>>>    __do_softirq+0x1ea/0xa4c
>>>    run_ksoftirqd+0x32/0x60
>>>    smpboot_thread_fn+0x503/0x860
>>>    kthread+0x29b/0x340
>>>    ret_from_fork+0x1f/0x30
>>>   </TASK>
>>>
>>>  From the above, in the function __rxe_add_to_pool,
>>> xa_lock is acquired. Then the function __rxe_add_to_pool
>>> is interrupted by softirq. The function
>>> rxe_pool_get_index will also acquire xa_lock.
>>>
>>> Finally, the dead lock appears.
>>>
>>> [  296.806097]        CPU0
>>> [  296.808550]        ----
>>> [  296.811003]   lock(&xa->xa_lock#15);  <----- __rxe_add_to_pool
>>> [  296.814583]   <Interrupt>
>>> [  296.817209]     lock(&xa->xa_lock#15); <---- rxe_pool_get_index
>>> [  296.820961]
>>>                   *** DEADLOCK ***
>>>
>>> Fixes: 3225717f6dfa ("RDMA/rxe: Replace red-black trees by carrays")
>>> Reported-and-tested-by: Yi Zhang <yi.zhang@redhat.com>
>>> Signed-off-by: Zhu Yanjun <yanjun.zhu@linux.dev>
>>> ---
>>> V2->V3: __rxe_add_to_pool is between spin_lock and spin_unlock, so
>>>          GFP_ATOMIC is used in __rxe_add_to_pool.
>>> V1->V2: Replace GFP_KERNEL with GFP_ATOMIC
>>> ---
>>>   drivers/infiniband/sw/rxe/rxe_pool.c | 17 ++++++++++++-----
>>>   1 file changed, 12 insertions(+), 5 deletions(-)
>>>
>>> diff --git a/drivers/infiniband/sw/rxe/rxe_pool.c 
>>> b/drivers/infiniband/sw/rxe/rxe_pool.c
>>> index 87066d04ed18..b9b147df4020 100644
>>> --- a/drivers/infiniband/sw/rxe/rxe_pool.c
>>> +++ b/drivers/infiniband/sw/rxe/rxe_pool.c
>>> @@ -138,8 +138,10 @@ void *rxe_alloc(struct rxe_pool *pool)
>>>       elem->obj = obj;
>>>       kref_init(&elem->ref_cnt);
>>> -    err = xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
>>> -                  &pool->next, GFP_KERNEL);
>>> +    xa_lock_bh(&pool->xa);
>>> +    err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
>>> +                &pool->next, GFP_KERNEL);
>>> +    xa_unlock_bh(&pool->xa);
>>>       if (err)
>>>           goto err_free;
>> You can't mix bh and not bh locks, either this is an irq spinlock or
>> it is bh spinlock, pick one and also ensure that the proper xa_init
>> flag is set.
> 
> Got it. I should use irq spinlock. So XA_FLAGS_LOCK_IRQ is added in 
> xa_init flags.
> 
> So the code should be:
> 
> @@ -106,7 +106,7 @@ void rxe_pool_init(struct rxe_dev *rxe, struct 
> rxe_pool *pool,
> 
>          atomic_set(&pool->num_elem, 0);
> 
> -       xa_init_flags(&pool->xa, XA_FLAGS_ALLOC);
> +       xa_init_flags(&pool->xa, XA_FLAGS_LOCK_IRQ | XA_FLAGS_ALLOC);
>          pool->limit.min = info->min_index;
>          pool->limit.max = info->max_index;
>   }
> @@ -138,10 +138,10 @@ void *rxe_alloc(struct rxe_pool *pool)
>          elem->obj = obj;
>          kref_init(&elem->ref_cnt);
> 
> -       xa_lock_bh(&pool->xa);
> +       xa_lock_irq(&pool->xa);
>          err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem, 
> pool->limit,
>                                  &pool->next, GFP_KERNEL);
> -       xa_unlock_bh(&pool->xa);
> +       xa_unlock_irq(&pool->xa);
>          if (err)
>                  goto err_free;
> 
>>
>>> @@ -166,8 +168,10 @@ int __rxe_add_to_pool(struct rxe_pool *pool, 
>>> struct rxe_pool_elem *elem)
>>>       elem->obj = (u8 *)elem - pool->elem_offset;
>>>       kref_init(&elem->ref_cnt);
>>> -    err = xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
>>> -                  &pool->next, GFP_KERNEL);
>>> +    xa_lock_irq(&pool->xa);
>>> +    err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
>>> +                &pool->next, GFP_ATOMIC);
>>> +    xa_unlock_irq(&pool->xa);
>>>       if (err)
>>>           goto err_cnt;
>> Still no, this does almost every allocation - only AH with the
>> non-blocking flag set should use this path.
> 
> Yes. Got it.
> 
> In the following, xa_lock_irqsave/xa_unlock_irqrestore should be used.
> 
> int ib_send_cm_req(struct ib_cm_id *cm_id,
> 
>                     struct ib_cm_req_param *param)
> {
> 
> ...
> 
> spin_lock_irqsave(&cm_id_priv->lock, flags);
> 
> ...
> 
> __rxe_add_to_pool
> 
> ...
> 
> spin_unlock_irqrestore(&cm_id_priv->lock, flags);

Hi，Jason

To the function ib_send_cm_req, the call chain is as below.

ib_send_cm_req --> cm_alloc_priv_msg --> cm_alloc_msg --> rdma_create_ah 
--> _rdma_create_ah --> rxe_create_ah --> rxe_av_chk_attr 
-->__rxe_add_to_pool

As such, xa_lock_irqsave/irqrestore is selected.

Zhu Yanjun

> 
> 
> So the diff is as below:
> 
> @@ -157,6 +157,7 @@ void *rxe_alloc(struct rxe_pool *pool)
>   int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem)
>   {
>          int err;
> +       unsigned long flags;
> 
>          if (WARN_ON(pool->flags & RXE_POOL_ALLOC))
>                  return -EINVAL;
> @@ -168,10 +169,10 @@ int __rxe_add_to_pool(struct rxe_pool *pool, 
> struct rxe_pool_elem *elem)
>          elem->obj = (u8 *)elem - pool->elem_offset;
>          kref_init(&elem->ref_cnt);
> 
> -       xa_lock_irq(&pool->xa);
> +       xa_lock_irqsave(&pool->xa, flags);
>          err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem, 
> pool->limit,
>                                  &pool->next, GFP_ATOMIC);
> -       xa_unlock_irq(&pool->xa);
> +       xa_unlock_irqrestore(&pool->xa, flags);
>          if (err)
>                  goto err_cnt;
> 
> Please comment. Thanks a lot.
> 
> Zhu Yanjun
> 
>>
>> Jason