[PATCH v2] nvme-rdma: fix NULL queue->ctrl dereference on reconnect

[Christian Aurich Zanettini Martins]

Since commit 2a8513091d2f ("nvme-rdma: parallelize I/O queue allocation
and startup"), a reconnect that gets more I/O queues than before grows
the tag set before the new queues are set up. queue->ctrl is only set
by nvme_rdma_setup_io_queues(), so I/O submitted in between reaches
nvme_rdma_queue_rq() on a queue that was never allocated and
dereferences a NULL queue->ctrl:

  BUG: kernel NULL pointer dereference, address: 000000000000034c
  RIP: 0010:nvme_rdma_queue_rq+0x52/0xb90

Set queue->ctrl for every queue when allocating the controller. Such
requests are then requeued like on any other queue that is not live.

Fixes: 2a8513091d2f ("nvme-rdma: parallelize I/O queue allocation and startup")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Christian Aurich Zanettini Martins <christian.aurichzm@xxxxxxxxx>
---
v2:
- Only move queue->ctrl. v1 also moved queue_lock, which needs the
  queue count in nvme_rdma_free_ctrl(), but opts can already be freed
  there when creating the controller fails (KASAN use-after-free on a
  failed connect with CONFIG_DEBUG_KOBJECT_RELEASE).
- Not carrying the Reviewed-by tags from Nilay and Christoph, as most
  of the patch changed

v1: https://lore.kernel.org/linux-nvme/20261029060521.1690223-1-christian.aurichzm@xxxxxxxxx/

Based on nvme-7.4 (89480d137bf0), also applies to mainline and v7.3.

Reproduced in QEMU with nvmet-rdma over rxe and
CONFIG_NVME_MULTIPATH=n: connect with attr_qid_max=2, keep direct reads
running on the CPUs that map to the new hw queues, then raise
attr_qid_max to 4. The window is short, so the readers run at
SCHED_FIFO with nvme-wq pinned to their CPUs. Without the patch the
host oopses within two tries in all six boots. With v2, 24 of 24 tries
go live and 30 resets under I/O pass, with and without lockdep, and
failed connects are clean under KASAN.

 drivers/nvme/host/rdma.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/drivers/nvme/host/rdma.c b/drivers/nvme/host/rdma.c
index 07216890f75c..32189fdfc0f5 100644
--- a/drivers/nvme/host/rdma.c
+++ b/drivers/nvme/host/rdma.c
@@ -735,7 +735,6 @@ static int nvme_rdma_setup_io_queues(struct nvme_rdma_ctrl *ctrl,
 	for (i = 0; i < nr_queues; i++) {
 		struct nvme_rdma_queue *queue = &ctrl->queues[first + i];
 
-		queue->ctrl = ctrl;
 		queue->queue_size = queue_size;
 
 		ctxs[i].queue = queue;
@@ -798,7 +797,6 @@ static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 	bool pi_capable = false;
 	int error;
 
-	ctrl->queues[0].ctrl = ctrl;
 	ctrl->queues[0].queue_size = NVME_AQ_DEPTH;
 	error = nvme_rdma_alloc_queue(&ctrl->queues[0]);
 	if (error)
@@ -2284,7 +2282,7 @@ static struct nvme_rdma_ctrl *nvme_rdma_alloc_ctrl(struct device *dev,
 		struct nvmf_ctrl_options *opts)
 {
 	struct nvme_rdma_ctrl *ctrl;
-	int ret;
+	int ret, i;
 
 	ctrl = kzalloc_obj(*ctrl);
 	if (!ctrl)
@@ -2343,6 +2341,9 @@ static struct nvme_rdma_ctrl *nvme_rdma_alloc_ctrl(struct device *dev,
 	if (!ctrl->queues)
 		goto out_free_ctrl;
 
+	for (i = 0; i < ctrl->ctrl.queue_count; i++)
+		ctrl->queues[i].ctrl = ctrl;
+
 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
 				0 /* no quirks, we're perfect! */);
 	if (ret)

base-commit: 89480d137bf0d943a82c64f86ae2407a15276b5e
-- 
2.43.0