rds: rds_ib_device.refcount overflow
authorWengang Wang <wen.gang.wang@oracle.com>
Mon, 6 Jul 2015 06:35:11 +0000 (14:35 +0800)
committerDoug Ledford <dledford@redhat.com>
Tue, 14 Jul 2015 17:20:11 +0000 (13:20 -0400)
commit4fabb59449aa44a585b3603ffdadd4c5f4d0c033
tree5fed9bcc47497d88442ccca258cb3b0eb0b6a66a
parent0a691272509d59ea0adbaa0c420ca0a71c9d0419
rds: rds_ib_device.refcount overflow

Fixes: 3e0249f9c05c ("RDS/IB: add refcount tracking to struct rds_ib_device")

There lacks a dropping on rds_ib_device.refcount in case rds_ib_alloc_fmr
failed(mr pool running out). this lead to the refcount overflow.

A complain in line 117(see following) is seen. From vmcore:
s_ib_rdma_mr_pool_depleted is 2147485544 and rds_ibdev->refcount is -2147475448.
That is the evidence the mr pool is used up. so rds_ib_alloc_fmr is very likely
to return ERR_PTR(-EAGAIN).

115 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
116 {
117         BUG_ON(atomic_read(&rds_ibdev->refcount) <= 0);
118         if (atomic_dec_and_test(&rds_ibdev->refcount))
119                 queue_work(rds_wq, &rds_ibdev->free_work);
120 }

fix is to drop refcount when rds_ib_alloc_fmr failed.

Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>
Reviewed-by: Haggai Eran <haggaie@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
net/rds/ib_rdma.c