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

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

[Gerd Rausch]

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;
-- 
2.24.1

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

[Mark Zhang]

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?

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

[Haakon Bugge]

> 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

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

[Jason Gunthorpe]

On Thu, Jun 24, 2021 at 11:55:31AM -0700, 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(-)

I wondered for a while if it would be better to clear this in
cma_query_handler(), but it seems this is OK

Applied to for-next

Thanks,
Jason

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

[Gerd Rausch]

Hi Mark,

On 24/06/2021 22.49, Mark Zhang 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?

So the caller performs "rdma_resolve_route" which returns an immediate error,
but the expectation would be that the cm_id still points to
a valid (!= NULL) path record
(even though this error indicateed route lookup failed).

Which code-part and call-sequence would have such expectation?

I can't say I'm in love with this approach either, since the code
makes the assumption that "path_rec" (if != NULL) points to a space
that was allocated by "kmalloc" before and can be overwritten.
It's not permitted to point to something else.

But "__rdma_free" makes that assumption anyway, since it unconditionally
calls "kfree" on this pointer.

So there is precedence.

Just my 2ç,

  Gerd

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

[Gerd Rausch]

Hi Jason,

On 25/06/2021 08.03, Jason Gunthorpe wrote:
> On Thu, Jun 24, 2021 at 11:55:31AM -0700, 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(-)
> 
> I wondered for a while if it would be better to clear this in
> cma_query_handler(), but it seems this is OK
> 
> Applied to for-next
> 

Thank you, I appreciate it.

  Gerd