Skip to contentSkip to conversation

nvme-tcp: fix NULL queue->ctrl dereference on reconnect

PATCH Superseded Christian Aurich Zanettini Martins 3 participants 3 messages Oct 28 NVMe

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 nvme-tcp: fix NULL queue->ctrl dereference on reconnect

Since commit 61b0ed3be81b ("nvme-tcp: 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 allocated. queue->ctrl is only set by nvme_tcp_alloc_queue(), so I/O submitted in between reaches nvme_tcp_queue_rq() on a queue that was never allocated and dereferences a NULL queue->ctrl:

  BUG: kernel NULL pointer dereference, address: 00000000000002ac
  RIP: 0010:nvme_tcp_queue_rq+0x33/0x3b0

Add nvme_tcp_init_queue() for the queue fields that do not depend on the connection and call it for every queue when allocating the controller. Such requests are then requeued like on any other queue that is not live. send_mutex and io_work move there as well, and the helper's mutexes are all destroyed when the controller is freed, like queue_lock already is.

Fixes 61b0ed3be81b ("nvme-tcp: parallelize I/O queue allocation and startup")

---
v2:
- Rebase on Guixin's v4 "nvme-tcp: Validate the queue state under the queue lock in the address attribute" and move its queue_lock init into the helper
- Also move send_mutex and io_work into the helper, and destroy send_mutex and pf_cache_lock next to queue_lock in nvme_tcp_free_ctrl() (Keith)
- Left in nvme_tcp_alloc_queue(): req_list and send_list, as resetting them drops entries left over from the previous connection, and the lockdep keys, which are per socket instance since commit 19bdb70c77d3 ("nvme-tcp: lockdep: use dynamic lockdep keys per socket instance")

Based on nvme-7.4 (9afc1b71d963) with "nvme-tcp: fix request PDU size when header digest is enabled" and Guixin's v4 on top. The last hunk assumes the conflict between those two in nvme_tcp_alloc_ctrl() is resolved by keeping both loop bodies and both comments. I can resend on top of nvme-7.4 once both are there.

Reproduced in QEMU with nvmet-tcp over loopback 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. Before the patch the host oopsed as above within two tries in each of three boots; with v2, 8 of 8 tries went live, with and without lockdep. 30 resets under digest I/O, with a reader looping on the address attribute, also pass and are clean with lockdep.

 drivers/nvme/host/tcp.c | 58 ++++++++++++++++++++---------------------
 1 file changed, 29 insertions(+), 29 deletions(-)
drivers/nvme/host/tcp.c +29 -29
index 4ada1ff120e2..77bc530ae0a6 100644
@@ -1559,7 +1559,6 @@ static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
 	memalloc_noio_restore(noio_flag);
 
 	kfree(queue->pdu);
-	mutex_destroy(&queue->send_mutex);
 
 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 	lockdep_unregister_key(&queue->nvme_tcp_sk_key);
@@ -1987,6 +1986,26 @@ static int nvme_tcp_start_tls(struct nvme_ctrl *nctrl,
 	return ret;
 }
 
+/*
+ * The I/O tag set is allocated, and grown on reconnect, before the queues are
+ * connected, so blk-mq can initialize requests on a queue, and dispatch to it,
+ * before nvme_tcp_alloc_queue() has run for it. Set up the fields that do not
+ * depend on the connection once, when allocating the controller, and keep them
+ * until the controller is freed.
+ */
+static void nvme_tcp_init_queue(struct nvme_tcp_ctrl *ctrl, int qid)
+{
+	struct nvme_tcp_queue *queue = &ctrl->queues[qid];
+
+	queue->ctrl = ctrl;
+	mutex_init(&queue->queue_lock);
+	mutex_init(&queue->send_mutex);
+	mutex_init(&queue->pf_cache_lock);
+	INIT_WORK(&queue->io_work, nvme_tcp_io_work);
+	queue->hdr_digest = ctrl->ctrl.opts->hdr_digest;
+	queue->data_digest = ctrl->ctrl.opts->data_digest;
+}
+
 static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid,
 				key_serial_t pskid)
 {
@@ -1995,11 +2014,8 @@ static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid,
 	int ret, rcv_pdu_size;
 	struct file *sock_file;
 
-	queue->ctrl = ctrl;
 	init_llist_head(&queue->req_list);
 	INIT_LIST_HEAD(&queue->send_list);
-	mutex_init(&queue->send_mutex);
-	INIT_WORK(&queue->io_work, nvme_tcp_io_work);
 
 	if (qid > 0)
 		queue->cmnd_capsule_len = nctrl->ioccsz * 16;
@@ -2013,14 +2029,12 @@ static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid,
 	if (ret) {
 		dev_err(nctrl->device,
 			"failed to create socket: %d\n", ret);
-		goto err_destroy_mutex;
+		return ret;
 	}
 
 	sock_file = sock_alloc_file(queue->sock, O_CLOEXEC, NULL);
-	if (IS_ERR(sock_file)) {
-		ret = PTR_ERR(sock_file);
-		goto err_destroy_mutex;
-	}
+	if (IS_ERR(sock_file))
+		return PTR_ERR(sock_file);
 
 	sk_net_refcnt_upgrade(queue->sock->sk);
 
@@ -2088,8 +2102,6 @@ static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid,
 		}
 	}
 
-	queue->data_digest = nctrl->opts->data_digest;
-
 	rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
 			nvme_tcp_hdgst_len(queue);
 	queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
@@ -2136,8 +2148,6 @@ static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid,
 	lockdep_unregister_key(&queue->nvme_tcp_sk_key);
 	lockdep_unregister_key(&queue->nvme_tcp_slock_key);
 #endif
-err_destroy_mutex:
-	mutex_destroy(&queue->send_mutex);
 	return ret;
 }
 
@@ -2838,8 +2848,11 @@ static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
 	nvmf_free_options(nctrl->opts);
 free_ctrl:
 	put_net(ctrl->net);
-	for (i = 0; i < nr_queues; i++)
+	for (i = 0; i < nr_queues; i++) {
 		mutex_destroy(&ctrl->queues[i].queue_lock);
+		mutex_destroy(&ctrl->queues[i].send_mutex);
+		mutex_destroy(&ctrl->queues[i].pf_cache_lock);
+	}
 	kfree(ctrl->queues);
 	kfree(ctrl);
 }
@@ -3238,21 +3251,8 @@ static struct nvme_tcp_ctrl *nvme_tcp_alloc_ctrl(struct device *dev,
 		goto out_free_ctrl;
 	}
 
-	/*
-	 * The I/O tag set is allocated before the I/O queues, and initializing
-	 * its requests already allocates the PDUs from the page frag caches of
-	 * the queues, sized by their header digest setting, so set up the
-	 * locks and hdr_digest here.
-	 *
-	 * The queue locks are owned by the controller, not the queues:
-	 * they outlive the per-queue connect/disconnect cycles so that
-	 * sysfs readers can never race their destruction.
-	 */
-	for (i = 0; i < ctrl->ctrl.queue_count; i++) {
-		mutex_init(&ctrl->queues[i].pf_cache_lock);
-		ctrl->queues[i].hdr_digest = opts->hdr_digest;
-		mutex_init(&ctrl->queues[i].queue_lock);
-	}
+	for (i = 0; i < ctrl->ctrl.queue_count; i++)
+		nvme_tcp_init_queue(ctrl, i);
 
 	ctrl->net = get_net(current->nsproxy->net_ns);
 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_tcp_ctrl_ops, 0);
Signature · 1 line
2.43.0
Reviewed-by on v2 1/1
#1 lore raw
Quoted 17 lines · Christian Aurich Zanettini Martins

Since commit 61b0ed3be81b ("nvme-tcp: 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 allocated. queue->ctrl is only set by nvme_tcp_alloc_queue(), so I/O submitted in between reaches nvme_tcp_queue_rq() on a queue that was never allocated and dereferences a NULL queue->ctrl:

   BUG: kernel NULL pointer dereference, address: 00000000000002ac
   RIP: 0010:nvme_tcp_queue_rq+0x33/0x3b0

Add nvme_tcp_init_queue() for the queue fields that do not depend on the connection and call it for every queue when allocating the controller. Such requests are then requeued like on any other queue that is not live. send_mutex and io_work move there as well, and the helper's mutexes are all destroyed when the controller is freed, like queue_lock already is.

Quoted 3 lines
Fixes 61b0ed3be81b ("nvme-tcp: parallelize I/O queue allocation and startup")

Looks good to me.

#2 lore raw
In reply to Christian Aurich Zanettini Martins
Quoted 138 lines · Christian Aurich Zanettini Martins

Since commit 61b0ed3be81b ("nvme-tcp: 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 allocated. queue->ctrl is only set by nvme_tcp_alloc_queue(), so I/O submitted in between reaches nvme_tcp_queue_rq() on a queue that was never allocated and dereferences a NULL queue->ctrl:

   BUG: kernel NULL pointer dereference, address: 00000000000002ac
   RIP: 0010:nvme_tcp_queue_rq+0x33/0x3b0

Add nvme_tcp_init_queue() for the queue fields that do not depend on the connection and call it for every queue when allocating the controller. Such requests are then requeued like on any other queue that is not live. send_mutex and io_work move there as well, and the helper's mutexes are all destroyed when the controller is freed, like queue_lock already is.

Fixes 61b0ed3be81b ("nvme-tcp: parallelize I/O queue allocation and startup")

---
v2:
- Rebase on Guixin's v4 "nvme-tcp: Validate the queue state under the queue lock in the address attribute" and move its queue_lock init into the helper
- Also move send_mutex and io_work into the helper, and destroy send_mutex and pf_cache_lock next to queue_lock in nvme_tcp_free_ctrl() (Keith)
- Left in nvme_tcp_alloc_queue(): req_list and send_list, as resetting them drops entries left over from the previous connection, and the lockdep keys, which are per socket instance since commit 19bdb70c77d3 ("nvme-tcp: lockdep: use dynamic lockdep keys per socket instance")

Based on nvme-7.4 (9afc1b71d963) with "nvme-tcp: fix request PDU size when header digest is enabled" and Guixin's v4 on top. The last hunk assumes the conflict between those two in nvme_tcp_alloc_ctrl() is resolved by keeping both loop bodies and both comments. I can resend on top of nvme-7.4 once both are there.

Reproduced in QEMU with nvmet-tcp over loopback 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. Before the patch the host oopsed as above within two tries in each of three boots; with v2, 8 of 8 tries went live, with and without lockdep. 30 resets under digest I/O, with a reader looping on the address attribute, also pass and are clean with lockdep.

  drivers/nvme/host/tcp.c | 58 ++++++++++++++++++++---------------------
  1 file changed, 29 insertions(+), 29 deletions(-)
drivers/nvme/host/tcp.c +23 -13
index 4ada1ff120e2..77bc530ae0a6 100644
@@ -1559,7 +1559,6 @@ static void nvme_tcp_free_queue(struct nvme_ctrl *nctrl, int qid)
  	memalloc_noio_restore(noio_flag);
 
  	kfree(queue->pdu);
-	mutex_destroy(&queue->send_mutex);
 
  #ifdef CONFIG_DEBUG_LOCK_ALLOC
  	lockdep_unregister_key(&queue->nvme_tcp_sk_key);
@@ -1987,6 +1986,26 @@ static int nvme_tcp_start_tls(struct nvme_ctrl *nctrl,
  	return ret;
  }
 
+/*
+ * The I/O tag set is allocated, and grown on reconnect, before the queues are
+ * connected, so blk-mq can initialize requests on a queue, and dispatch to it,
+ * before nvme_tcp_alloc_queue() has run for it. Set up the fields that do not
+ * depend on the connection once, when allocating the controller, and keep them
+ * until the controller is freed.
+ */
+static void nvme_tcp_init_queue(struct nvme_tcp_ctrl *ctrl, int qid)
+{
+	struct nvme_tcp_queue *queue = &ctrl->queues[qid];
+
+	queue->ctrl = ctrl;
+	mutex_init(&queue->queue_lock);
+	mutex_init(&queue->send_mutex);
+	mutex_init(&queue->pf_cache_lock);
+	INIT_WORK(&queue->io_work, nvme_tcp_io_work);
+	queue->hdr_digest = ctrl->ctrl.opts->hdr_digest;
+	queue->data_digest = ctrl->ctrl.opts->data_digest;
+}
+
  static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid,
  				key_serial_t pskid)
  {
@@ -1995,11 +2014,8 @@ static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid,
  	int ret, rcv_pdu_size;
  	struct file *sock_file;
 
-	queue->ctrl = ctrl;
  	init_llist_head(&queue->req_list);
  	INIT_LIST_HEAD(&queue->send_list);
-	mutex_init(&queue->send_mutex);
-	INIT_WORK(&queue->io_work, nvme_tcp_io_work);
 
  	if (qid > 0)
  		queue->cmnd_capsule_len = nctrl->ioccsz * 16;
@@ -2013,14 +2029,12 @@ static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid,
  	if (ret) {
  		dev_err(nctrl->device,
  			"failed to create socket: %d\n", ret);
-		goto err_destroy_mutex;
+		return ret;
  	}
 
  	sock_file = sock_alloc_file(queue->sock, O_CLOEXEC, NULL);
-	if (IS_ERR(sock_file)) {
-		ret = PTR_ERR(sock_file);
-		goto err_destroy_mutex;
-	}
+	if (IS_ERR(sock_file))
+		return PTR_ERR(sock_file);
 
  	sk_net_refcnt_upgrade(queue->sock->sk);
 
@@ -2088,8 +2102,6 @@ static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid,
  		}
  	}
 
-	queue->data_digest = nctrl->opts->data_digest;
-
  	rcv_pdu_size = sizeof(struct nvme_tcp_rsp_pdu) +
  			nvme_tcp_hdgst_len(queue);
  	queue->pdu = kmalloc(rcv_pdu_size, GFP_KERNEL);
@@ -2136,8 +2148,6 @@ static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid,
  	lockdep_unregister_key(&queue->nvme_tcp_sk_key);
  	lockdep_unregister_key(&queue->nvme_tcp_slock_key);
  #endif
-err_destroy_mutex:
-	mutex_destroy(&queue->send_mutex);
  	return ret;
  }
 

drivers/nvme/host/tcp.c static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl) · line 2855

Quoted 9 lines
drivers/nvme/host/tcp.c +4 -1
@@ -2838,8 +2848,11 @@ static void nvme_tcp_free_ctrl(struct nvme_ctrl *nctrl)
  	nvmf_free_options(nctrl->opts);
  free_ctrl:
  	put_net(ctrl->net);
-	for (i = 0; i < nr_queues; i++)
+	for (i = 0; i < nr_queues; i++) {
  		mutex_destroy(&ctrl->queues[i].queue_lock);
+		mutex_destroy(&ctrl->queues[i].send_mutex);
+		mutex_destroy(&ctrl->queues[i].pf_cache_lock);
+	}

Might be an idea to use a new function 'nvme_tcp_exit_queue()' here for symmetry with 'nvme_tcp_init_queue()'.

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                  Kernel Storage Architect
[email protected]                                +49 911 74053 688
SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg
HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich