Re: [PATCHv5 1/2] RDMA/rxe: Fix a dead lock problem

[Jason Gunthorpe <jgg@ziepe.ca>]

On Wed, Apr 20, 2022 at 09:42:23AM +0300, Leon Romanovsky wrote:
> On Sat, Apr 16, 2022 at 10:43:42PM -0400, yanjun.zhu@linux.dev wrote:
> > From: Zhu Yanjun <yanjun.zhu@linux.dev>
> > 
> > This is a dead lock problem.
> > The ah_pool 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]
> > 
> > Then ah_pool 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]
> > </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>
> > V4->V5: Commit logs are changed to avoid confusion.
> > V3->V4: xa_lock_irq locks are used.
> > 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 | 20 ++++++++++++++------
> >  1 file changed, 14 insertions(+), 6 deletions(-)
> > 
> > diff --git a/drivers/infiniband/sw/rxe/rxe_pool.c b/drivers/infiniband/sw/rxe/rxe_pool.c
> > index 87066d04ed18..f1f06dc7e64f 100644
> > +++ b/drivers/infiniband/sw/rxe/rxe_pool.c
> > @@ -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_ALLOC | XA_FLAGS_LOCK_IRQ);
> >  	pool->limit.min = info->min_index;
> >  	pool->limit.max = info->max_index;
> >  }
> > @@ -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_irq(&pool->xa);
> > +	err = __xa_alloc_cyclic(&pool->xa, &elem->index, elem, pool->limit,
> > +				&pool->next, GFP_KERNEL);
> > +	xa_unlock_irq(&pool->xa);

It should just use xa_alloc_cyclic_irq() and xa_erase_irq(). Don't
open code the lock.

> I may admit that I didn't follow your previous discussions, so maybe you
> already explained it. But why do you need xa_lock_irq() here?

The spinlock type needs to be consistent in all users.

You can only use the naked version if the spinlock is always obtained
from a process context.

You can only use bh version if the spinlock is always obtained from a
process context or bh/softirq

You can always use the irq version

What I don't understand is why IRQ and not BH? AFAIK there is no case
where rxe is called from a real IRQ, right? Or is it because you can't
nest BH under the IRQ spinlock from CM?

Jason