[PATCH 0/2] nvmet-rdma: fix response resource leak on queue teardown

[PATCH 0/2] nvmet-rdma: fix response resource leak on queue teardown

[Shin'ichiro Kawasaki]

When an nvme target with rdma transport is removed while I/Os are in
flight, a response can be posted but its send completion is never
delivered before the connection is torn down. This leaks the request
SGLs and the RDMA read/write context. The leaks are recreated by
running blktests nvme/061 with rdma transport and the siw driver [1].

[1] https://lore.kernel.org/linux-nvme/ajtk1CaN1pBreS4O@shinmob/

This series addresses the problem. The first patch is a preparation
refactoring. The second patch fixes the problem by reclaiming memory
objects on queue QP teardown.


Shin'ichiro Kawasaki (2):
  nvmet-rdma: factor out response resource cleanup
  nvmet-rdma: fix response resource leak on queue teardown

 drivers/nvme/target/rdma.c | 31 ++++++++++++++++++++++++++++---
 1 file changed, 28 insertions(+), 3 deletions(-)

-- 
2.54.0

[PATCH 1/2] nvmet-rdma: factor out response resource cleanup

[Shin'ichiro Kawasaki]

Move the RDMA read/write context teardown and the request SGL freeing
out of nvmet_rdma_release_rsp() into a new helper function
nvmet_rdma_free_rsp_resources().

This is a refactoring with no functional change, in preparation for the
following patch that uses nvmet_rdma_free_rsp_resources().

Signed-off-by: Shin'ichiro Kawasaki <shinichiro.kawasaki@wdc.com>
---
 drivers/nvme/target/rdma.c | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)

diff --git a/drivers/nvme/target/rdma.c b/drivers/nvme/target/rdma.c
index ea1185b8267e..3c34f235e542 100644
--- a/drivers/nvme/target/rdma.c
+++ b/drivers/nvme/target/rdma.c
@@ -657,18 +657,25 @@ static void nvmet_rdma_rw_ctx_destroy(struct nvmet_rdma_rsp *rsp)
 				    req->sg, req->sg_cnt, nvmet_data_dir(req));
 }
 
-static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
+static void nvmet_rdma_free_rsp_resources(struct nvmet_rdma_rsp *rsp)
 {
 	struct nvmet_rdma_queue *queue = rsp->queue;
 
-	atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
-
 	if (rsp->n_rdma)
 		nvmet_rdma_rw_ctx_destroy(rsp);
 
 	if (rsp->req.sg < rsp->cmd->inline_sg ||
 	    rsp->req.sg >= rsp->cmd->inline_sg + queue->dev->inline_page_count)
 		nvmet_req_free_sgls(&rsp->req);
+}
+
+static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
+{
+	struct nvmet_rdma_queue *queue = rsp->queue;
+
+	atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
+
+	nvmet_rdma_free_rsp_resources(rsp);
 
 	if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
 		nvmet_rdma_process_wr_wait_list(queue);
-- 
2.54.0

Re: [PATCH 1/2] nvmet-rdma: factor out response resource cleanup

[Christoph Hellwig]

Looks good:

Reviewed-by: Christoph Hellwig <hch@lst.de>

[PATCH 2/2] nvmet-rdma: fix response resource leak on queue teardown

[Shin'ichiro Kawasaki]

When an nvme target with rdma transport is removed while I/Os are in
flight, a response can be posted but its send completion is never
delivered before the connection is torn down. As a result
nvmet_rdma_send_done() and nvmet_rdma_release_rsp() are never called for
the response, and this leaks the allocated RDMA read/write context and
request SGLs.

These leaks are recreated by running blktests nvme/061 with the rdma
transport and the siw driver. Kernel kmemleak feature reports them as
follows:

  unreferenced object 0xffff88812bc490c0 (size 32):
    comm "kworker/2:1H", pid 409, jiffies 4307744490
    backtrace (crc 89afd339):
      __kmalloc_noprof+0x5f9/0x890
      sgl_alloc_order+0x7b/0x380
      nvmet_req_alloc_sgls+0x290/0x4f0 [nvmet]
      nvmet_rdma_map_sgl_keyed+0x241/0x12e0 [nvmet_rdma]
      nvmet_rdma_handle_command+0x73e/0xb80 [nvmet_rdma]
      __ib_process_cq+0x149/0x4c0 [ib_core]
      ib_cq_poll_work+0x49/0x160 [ib_core]
      process_one_work+0x8b2/0x1640
      worker_thread+0x5fd/0xfe0
      kthread+0x367/0x460
      ret_from_fork+0x655/0x9d0
      ret_from_fork_asm+0x1a/0x30

  unreferenced object 0xffff88814bd05e80 (size 64):
    comm "kworker/3:1H", pid 148, jiffies 4295195428
    backtrace (crc e35510cb):
      __kmalloc_noprof+0x5f9/0x890
      rdma_rw_ctx_init+0x333/0x1fa0 [ib_core]
      nvmet_rdma_map_sgl_keyed+0x5c8/0x12e0 [nvmet_rdma]
      nvmet_rdma_handle_command+0x73e/0xb80 [nvmet_rdma]
      __ib_process_cq+0x149/0x4c0 [ib_core]
      ib_cq_poll_work+0x49/0x160 [ib_core]
      process_one_work+0x8b2/0x1640
      worker_thread+0x5fd/0xfe0
      kthread+0x367/0x460
      ret_from_fork+0x655/0x9d0
      ret_from_fork_asm+0x1a/0x30

To avoid the memory leaks, reclaim the memory of the in-flight responses
when the queue QP is torn down. Call nvmet_rdma_free_rsp_resources()
that frees up the RDMA read/write context and the request SGLs of such
responses.

Fixes: 8f000cac6e7a ("nvmet-rdma: add a NVMe over Fabrics RDMA target driver")
Signed-off-by: Shin'ichiro Kawasaki <shinichiro.kawasaki@wdc.com>
---
 drivers/nvme/target/rdma.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)

diff --git a/drivers/nvme/target/rdma.c b/drivers/nvme/target/rdma.c
index 3c34f235e542..de5a88fbb233 100644
--- a/drivers/nvme/target/rdma.c
+++ b/drivers/nvme/target/rdma.c
@@ -1345,9 +1345,27 @@ static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
 	goto out;
 }
 
+static bool nvmet_rdma_reclaim_rsp(struct sbitmap *sb, unsigned int bitnr,
+				   void *data)
+{
+	struct nvmet_rdma_queue *queue = data;
+
+	nvmet_rdma_free_rsp_resources(&queue->rsps[bitnr]);
+
+	return true;
+}
+
 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
 {
 	ib_drain_qp(queue->qp);
+
+	/*
+	 * Reclaim resources of a response that is still in-flight when the
+	 * queue is being torn down. This happens when the connection was
+	 * forcefully disconnected while an I/O is in flight.
+	 */
+	sbitmap_for_each_set(&queue->rsp_tags, nvmet_rdma_reclaim_rsp, queue);
+
 	if (queue->cm_id)
 		rdma_destroy_id(queue->cm_id);
 	ib_destroy_qp(queue->qp);
-- 
2.54.0

Re: [PATCH 2/2] nvmet-rdma: fix response resource leak on queue teardown

[Christoph Hellwig]

Looks good;

Reviewed-by: Christoph Hellwig <hch@lst.de>