Re: [PATCH rdma-rc 2/2] RDMA/core: Initialize lock when allocate a rdma_hw_stats structure

[Leon Romanovsky <leon@kernel.org>]

On Mon, Oct 25, 2021 at 11:50:43AM -0300, Jason Gunthorpe wrote:
> On Sun, Oct 24, 2021 at 09:08:21AM +0300, Leon Romanovsky wrote:
> > From: Mark Zhang <markzhang@nvidia.com>
> > 
> > Initialize the rdma_hw_stats "lock" field when do allocation, to fix the
> > warning below. Then we don't need to initialize it in sysfs, remove it.
> 
> This is a fine cleanup, but this does not describe the bug properly,
> or have the right fixes line..

I think that this Fixes line should be instead.
Fixes: 0a0800ce2a6a ("RDMA/core: Add a helper API rdma_free_hw_stats_struct")

> 
> The issue is here:
> 
> static struct rdma_counter *alloc_and_bind(struct ib_device *dev, u32 port,
> 					   struct ib_qp *qp,
> 					   enum rdma_nl_counter_mode mode)
> {
> 	counter->stats = dev->ops.counter_alloc_stats(counter);
> 	if (!counter->stats)
> 		goto err_stats;
> 
> Which does not init counter->stat's mutex.


This is exactly what Mark is doing here.

alloc_and_bind()
 -> dev->ops.counter_alloc_stats
  -> mlx5_ib_counter_alloc_stats
   -> do_alloc_stats()
    -> rdma_alloc_hw_stats_struct()
     -> mutex_init(&stats->lock); <- Mark's change.

> 
> And trim the oops reports, don't include the usless ? fns, timestamps
> or other junk.

I don't like when people "beatify" kernel reports, many times whey are
removing too much information.

Thanks

> 
> Jason