From: Shao Mingyin nvme_fc_alloc_ctrl() initializes only the admin queue and leaves the other nvme_fc_queue entries zeroed. The initial connect path calls nvme_fc_init_io_queues() before creating the tag set, but reconnects go through nvme_fc_recreate_io_queues(), which does not initialize the queues it adds when the controller grants more I/O queues than the previous connection. blk_mq_update_nr_hw_queues() then publishes the new hardware contexts while the corresponding queue entries still have qnum == 0 and ctrl == NULL. nvme_fc_connect_io_queues() builds Create Connection commands from qnum, so the new queues are connected with queue ID 0 (the admin queue), and __nvme_fc_init_request() also classifies their requests as admin commands. Worse, nvme_fc_queue_rq() dereferences queue->ctrl, so any request dispatched to a new hardware context before the queue is initialized crashes. This includes the connect requests themselves, since connect_q belongs to the same tag set. Initialize the newly added queues before growing the tag set, matching what the initial connect path does. Existing queues are left alone so their connection state is preserved, as required by nvme_fc_recreate_io_queues(). Fixes: 3e493c00cedb ("nvme-fc: fix nulling of queue data on reconnect") Cc: stable@vger.kernel.org Signed-off-by: Shao Mingyin --- Tested on 7.3.0-next-20261030 with nvmet-fcloop in a single QEMU/TCG guest (2 vCPUs); no FC or network hardware is involved. The controller is created with one I/O queue (attr_qid_max=1), then attr_qid_max=2 deletes it and forces a reconnect that grants two queues, so nvme_fc_recreate_io_queues() grows the tag set from 1 to 2: unpatched patched nvme0: reconnect: revising io queue same count from 1 to 2 Oops: 0000 [#1] SMP NOPTI no oops, 6/6 cycles Workqueue: nvme-wq nvme_fc_connect_ctrl_work RIP: nvme_fc_queue_rq+0x21/0x190 queue_count=3 CR2: 0000000000000020 I/O from cpu1, ret=0 Kernel panic - not syncing The NULL dereference is the Create Connection for the new queue: it goes through ctrl->connect_q, which shares the tag set, and is dispatched to the new hctx whose queue->ctrl is still NULL. CR2 is 0x20, the offset nvme_check_ready() reads. The target also warns about a duplicate queue id on the same reconnect, because the new queue is connected with queue ID 0 instead of 2. drivers/nvme/host/fc.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/drivers/nvme/host/fc.c b/drivers/nvme/host/fc.c index 0b9ff50d26ff..f3603c793ede 100644 --- a/drivers/nvme/host/fc.c +++ b/drivers/nvme/host/fc.c @@ -2932,7 +2932,7 @@ nvme_fc_recreate_io_queues(struct nvme_fc_ctrl *ctrl) struct nvmf_ctrl_options *opts = ctrl->ctrl.opts; u32 prior_ioq_cnt = ctrl->ctrl.queue_count - 1; unsigned int nr_io_queues; - int ret; + int ret, i; nr_io_queues = min3(opts->nr_io_queues, num_online_cpus(), ctrl->lport->ops->max_hw_queues); @@ -2959,6 +2959,8 @@ nvme_fc_recreate_io_queues(struct nvme_fc_ctrl *ctrl) dev_info(ctrl->ctrl.device, "reconnect: revising io queue count from %d to %d\n", prior_ioq_cnt, nr_io_queues); + for (i = prior_ioq_cnt + 1; i <= nr_io_queues; i++) + nvme_fc_init_queue(ctrl, i); blk_mq_update_nr_hw_queues(&ctrl->tag_set, nr_io_queues); } -- 2.43.0