Skip to contentSkip to conversation

RDMA/cxgb4: fix refcounting leak in c4iw_ref_send_wait()

PATCH In review Dan Carpenter 2 participants 3 messages Jan 24, 2022 – Jan 28, 2022

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 RDMA/cxgb4: fix refcounting leak in c4iw_ref_send_wait()

Call c4iw_put_wr_wait() if c4iw_wait_for_reply() fails. This code uses kobject so the worst impact from this bug is a DoS.

Fixes 2015f26cfade ("iw_cxgb4: add referencing to wait objects")

---
From static analysis. Not tested.

 drivers/infiniband/hw/cxgb4/iw_cxgb4.h | 17 ++++++++++++-----
 1 file changed, 12 insertions(+), 5 deletions(-)
drivers/infiniband/hw/cxgb4/iw_cxgb4.h +12 -5
index 12f33467c672..1dc0e00aba5f 100644
@@ -297,11 +297,18 @@ static inline int c4iw_ref_send_wait(struct c4iw_rdev *rdev,
 		 qpid);
 	c4iw_get_wr_wait(wr_waitp);
 	ret = c4iw_ofld_send(rdev, skb);
-	if (ret) {
-		c4iw_put_wr_wait(wr_waitp);
-		return ret;
-	}
-	return c4iw_wait_for_reply(rdev, wr_waitp, hwtid, qpid, func);
+	if (ret)
+		goto put_wait;
+
+	ret = c4iw_wait_for_reply(rdev, wr_waitp, hwtid, qpid, func);
+	if (ret)
+		goto put_wait;
+
+	return 0;
+
+put_wait:
+	c4iw_put_wr_wait(wr_waitp);
+	return ret;
 }
 
 enum db_state {
Signature · 1 line
2.20.1
maintainer
#1 lore raw
Quoted 10 lines · Dan Carpenter

Call c4iw_put_wr_wait() if c4iw_wait_for_reply() fails. This code uses kobject so the worst impact from this bug is a DoS.

Fixes 2015f26cfade ("iw_cxgb4: add referencing to wait objects")

---

Quoted 1 line

From static analysis. Not tested.

 drivers/infiniband/hw/cxgb4/iw_cxgb4.h | 17 ++++++++++++-----
 1 file changed, 12 insertions(+), 5 deletions(-)

Are you sure?

Looking at the caller alloc_srq_queue() it calls down to c4iw_ref_send_wait() then immediately exits on failure

The only caller c4iw_create_srq()

	ret = alloc_srq_queue(srq, ucontext ? &ucontext->uctx :
			&rhp->rdev.uctx, srq->wr_waitp);
	if (ret)
		goto err_free_skb;

And then

err_free_skb:

	kfree_skb(srq->destroy_skb);
err_free_srq_idx:
	c4iw_free_srq_idx(&rhp->rdev, srq->idx);
err_free_wr_wait:
	c4iw_put_wr_wait(srq->wr_waitp);

So we just double put the thing with this patch

I have no idea how this logic is supposed to work, and clearly something is buggy in here, but I can't say this is right..

Jason

#2 lore raw
Quoted 33 lines · Jason Gunthorpe
Quoted 10 lines · Dan Carpenter

Call c4iw_put_wr_wait() if c4iw_wait_for_reply() fails. This code uses kobject so the worst impact from this bug is a DoS.

Fixes 2015f26cfade ("iw_cxgb4: add referencing to wait objects")

---

Quoted 1 line

From static analysis. Not tested.

 drivers/infiniband/hw/cxgb4/iw_cxgb4.h | 17 ++++++++++++-----
 1 file changed, 12 insertions(+), 5 deletions(-)

Are you sure?

Looking at the caller alloc_srq_queue() it calls down to c4iw_ref_send_wait() then immediately exits on failure

The only caller c4iw_create_srq()

	ret = alloc_srq_queue(srq, ucontext ? &ucontext->uctx :
			&rhp->rdev.uctx, srq->wr_waitp);
	if (ret)
		goto err_free_skb;

And then

err_free_skb:

	kfree_skb(srq->destroy_skb);
err_free_srq_idx:
	c4iw_free_srq_idx(&rhp->rdev, srq->idx);
err_free_wr_wait:
	c4iw_put_wr_wait(srq->wr_waitp);
Quoted 4 lines

So we just double put the thing with this patch

I have no idea how this logic is supposed to work, and clearly something is buggy in here, but I can't say this is right..

Yeah. My patch isn't right. That refcount from my patch is supposed to be decremented in _c4iw_wake_up(). That function gets called when the firmware responds in fw6_msg(). So if we cleanup everything and the firmware sends a delayed response the it leads to a use after free.

Say the firmware never responds, then sure, that leads to a resource leak. But it's better to have a small memory leak than a use after free.

regards,
dan carpenter