Re: [PATCH 1/1] RDMA/cma: Fix rdma_resolve_route memory leak

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

> On 25 Jun 2021, at 07:49, Mark Zhang <markzhang@nvidia.com> wrote:
> 
> On 6/25/2021 2:55 AM, Gerd Rausch wrote:
>> Fix a memory leak when "rmda_resolve_route" is called
>> more than once on the same "rdma_cm_id".
>> Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com>
>> ---
>>  drivers/infiniband/core/cma.c | 3 ++-
>>  1 file changed, 2 insertions(+), 1 deletion(-)
>> diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c
>> index ab148a696c0c..4a76d5b4163e 100644
>> --- a/drivers/infiniband/core/cma.c
>> +++ b/drivers/infiniband/core/cma.c
>> @@ -2819,7 +2819,8 @@ static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
>>    	cma_init_resolve_route_work(work, id_priv);
>>  -	route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
>> +	if (!route->path_rec)
>> +		route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
>>  	if (!route->path_rec) {
>>  		ret = -ENOMEM;
>>  		goto err1;
> 
> If route->path_rec does exist (meaning this is not the first time called), then it would be freed if cma_query_ib_route() below is failed, is it good?

This may happen if rdma_resolve_route() is called after a route error event has been received. In that case, the state is set back to RDMA_CM_ADDR_RESOLVED, so it seems the API allows rdma_resolve_route() to be called again without renewing the cm_id in this case.

Another thing, if this fix is eligible, it should probably also go into cma_resolve_iboe_route() as well.


Thxs, Håkon