Skip to contentSkip to conversation

xprtrdma: Ensure connect worker is awoken after connect error

PATCH In review Chuck Lever 2 participants 7 messages Jun 21, 2020 – Jun 26, 2020

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 xprtrdma: Ensure connect worker is awoken after connect error

From: Dan Aloni <[email protected]>

The connect worker sleeps waiting for either successful connection establishment or an error. Commit e28ce90083f0 ("xprtrdma: kmalloc rpcrdma_ep separate from rpcrdma_xprt") mistakenly removed the wake-up in cases when connection establishment fails.

Fixes e28ce90083f0 ("xprtrdma: kmalloc rpcrdma_ep separate from rpcrdma_xprt")
[ cel: rebased on recent fixes to 5.8-rc; patch description rewritten ]
---
 net/sunrpc/xprtrdma/verbs.c |    1 +
 1 file changed, 1 insertion(+)
net/sunrpc/xprtrdma/verbs.c +1 -0
index 2198c8ec8dff..54c6809bf06e 100644
@@ -296,6 +296,7 @@ rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
 		ep->re_connect_status = -ECONNABORTED;
 disconnected:
 		rpcrdma_force_disconnect(ep);
+		wake_up_all(&ep->re_connect_wait);
 		return rpcrdma_ep_put(ep);
 	default:
 		break;
#1 lore raw

Anna, please drop this one. It appears to trigger a particularly nasty use-after-free. I'll follow up with a more complete fix soon.

(Yes, a wake-up on connect errors is indeed necessary... but the connect worker needs to be re-organized to deal properly with it).

Quoted 26 lines · Chuck Lever

From: Dan Aloni <[email protected]>

The connect worker sleeps waiting for either successful connection establishment or an error. Commit e28ce90083f0 ("xprtrdma: kmalloc rpcrdma_ep separate from rpcrdma_xprt") mistakenly removed the wake-up in cases when connection establishment fails.

Fixes e28ce90083f0 ("xprtrdma: kmalloc rpcrdma_ep separate from rpcrdma_xprt")
[ cel: rebased on recent fixes to 5.8-rc; patch description rewritten ]

---
net/sunrpc/xprtrdma/verbs.c | 1 +

1 file changed, 1 insertion(+)
net/sunrpc/xprtrdma/verbs.c +0 -0
index 2198c8ec8dff..54c6809bf06e 100644
@@ -296,6 +296,7 @@ rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
 		ep->re_connect_status = -ECONNABORTED;

disconnected:

		rpcrdma_force_disconnect(ep);
net/sunrpc/xprtrdma/verbs.c +1 -0
+		wake_up_all(&ep->re_connect_wait);
Quoted 4 lines
		return rpcrdma_ep_put(ep);
	default:
		break;

--
Chuck Lever

#2 lore raw

PATCH 1/1 xprtrdma: fix EP destruction logic

This fixes various issues found when testing disconnections and other various fault injections under a KASAN-enabled kernel.

- Swap the names of rpcrdma_ep_destroy and rpcrdma_ep_put to reflect what the functions are doing.
- In the cleanup of rpcrdma_ep_create do kfree on the EP only for the case where no one knows about it, and avoid a double free (what rpcrdma_ep_destroy did, followed by kfree). No need to set r_xprt->rx_ep to NULL because we are doing that only in the success path.
- Don't up the EP ref in RDMA_CM_EVENT_ESTABLISHED but do it sooner before rdma_connect. This is needed so that an early wake up of wait_event_interruptible in rpcrdma_xprt_connect in case of a failure does not see a freed EP.
- Still try to follow the rule that the last decref of EP performs rdma_destroy_id(id).
- Only path to disengage an EP is rpcrdma_xprt_disconnect, whether it is actually connected or not. This makes the error paths of rpcrdma_xprt_connect clearer.
- Add a mutex in rpcrdma_ep_destroy to guard against concurrent calls to rpcrdma_xprt_disconnect coming from either rpcrdma_xprt_connect or xprt_rdma_close.

---
 net/sunrpc/xprtrdma/transport.c |  2 ++
 net/sunrpc/xprtrdma/verbs.c     | 50 ++++++++++++++++++++++-----------
 net/sunrpc/xprtrdma/xprt_rdma.h |  1 +
 3 files changed, 37 insertions(+), 16 deletions(-)
net/sunrpc/xprtrdma/transport.c +2 -0
index 0c4af7f5e241..50c28be6b694 100644
@@ -350,6 +350,8 @@ xprt_setup_rdma(struct xprt_create *args)
 	xprt_rdma_format_addresses(xprt, sap);
 
 	new_xprt = rpcx_to_rdmax(xprt);
+	mutex_init(&new_xprt->rx_flush);
+
 	rc = rpcrdma_buffer_create(new_xprt);
 	if (rc) {
 		xprt_rdma_free_addresses(xprt);
net/sunrpc/xprtrdma/verbs.c +34 -16

147 lines not on this page · raw message

net/sunrpc/xprtrdma/xprt_rdma.h +1 -0
index 0a16fdb09b2c..288645a9437f 100644
@@ -417,6 +417,7 @@ struct rpcrdma_xprt {
 	struct delayed_work	rx_connect_worker;
 	struct rpc_timeout	rx_timeout;
 	struct rpcrdma_stats	rx_stats;
+	struct mutex            rx_flush;
 };
 
 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
Signature · 1 line
2.25.4
#3 lore raw

Re: [PATCH] xprtrdma: fix EP destruction logic

Quoted 18 lines · Dan Aloni

This fixes various issues found when testing disconnections and other various fault injections under a KASAN-enabled kernel.

- Swap the names of rpcrdma_ep_destroy and rpcrdma_ep_put to reflect what the functions are doing.
- In the cleanup of rpcrdma_ep_create do kfree on the EP only for the case where no one knows about it, and avoid a double free (what rpcrdma_ep_destroy did, followed by kfree). No need to set r_xprt->rx_ep to NULL because we are doing that only in the success path.
- Don't up the EP ref in RDMA_CM_EVENT_ESTABLISHED but do it sooner before rdma_connect. This is needed so that an early wake up of wait_event_interruptible in rpcrdma_xprt_connect in case of a failure does not see a freed EP.

Quoted 8 lines

- Still try to follow the rule that the last decref of EP performs rdma_destroy_id(id).
- Only path to disengage an EP is rpcrdma_xprt_disconnect, whether it is actually connected or not. This makes the error paths of rpcrdma_xprt_connect clearer.
- Add a mutex in rpcrdma_ep_destroy to guard against concurrent calls to rpcrdma_xprt_disconnect coming from either rpcrdma_xprt_connect or xprt_rdma_close.

NAK. The RPC client provides appropriate exclusion, please let's not add more serialization that can introduce further deadlocks.

There are already patches pending in this area that make some of these changes, starting with

https://marc.info/?l=linux-nfs&m=159222724808832&w=2

Yesterday, I found the other problems you mentioned above after I was finally able to get my server to REJECT connection attempts. I'll submit patches to clean all this up once I have tested them.

Quoted 199 lines

---

net/sunrpc/xprtrdma/transport.c |  2 ++
net/sunrpc/xprtrdma/verbs.c     | 50 ++++++++++++++++++++++-----------
net/sunrpc/xprtrdma/xprt_rdma.h |  1 +
3 files changed, 37 insertions(+), 16 deletions(-)
net/sunrpc/xprtrdma/transport.c +2 -0
index 0c4af7f5e241..50c28be6b694 100644
@@ -350,6 +350,8 @@ xprt_setup_rdma(struct xprt_create *args)
 	xprt_rdma_format_addresses(xprt, sap);
 
 	new_xprt = rpcx_to_rdmax(xprt);
+	mutex_init(&new_xprt->rx_flush);
+
 	rc = rpcrdma_buffer_create(new_xprt);
 	if (rc) {
 		xprt_rdma_free_addresses(xprt);
net/sunrpc/xprtrdma/verbs.c +1 -1
index 2ae348377806..c66871bbb894 100644
@@ -84,7 +84,7 @@ static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep);
 static void rpcrdma_reps_unmap(struct rpcrdma_xprt *r_xprt);
 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
 static void rpcrdma_mrs_destroy(struct rpcrdma_xprt *r_xprt);
-static int rpcrdma_ep_destroy(struct rpcrdma_ep *ep);
+static int rpcrdma_ep_put(struct rpcrdma_ep *ep);

static struct rpcrdma_regbuf * rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,

		     gfp_t flags);
net/sunrpc/xprtrdma/verbs.c +3 -1
@@ -99,6 +99,9 @@ static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
 {
 	struct rdma_cm_id *id = r_xprt->rx_ep->re_id;
 
+	if (!id->qp)
+		return;
+
 	/* Flush Receives, then wait for deferred Reply work
 	 * to complete.
 	 */
@@ -266,7 +269,6 @@ rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
 		xprt_force_disconnect(xprt);
 		goto disconnected;
 	case RDMA_CM_EVENT_ESTABLISHED:
-		kref_get(&ep->re_kref);
 		ep->re_connect_status = 1;
 		rpcrdma_update_cm_private(ep, &event->param.conn);
 		trace_xprtrdma_inline_thresh(ep);
@@ -289,7 +291,8 @@ rpcrdma_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
 		ep->re_connect_status = -ECONNABORTED;

disconnected:

		xprt_force_disconnect(xprt);
net/sunrpc/xprtrdma/verbs.c +2 -1
-		return rpcrdma_ep_destroy(ep);
+		wake_up_all(&ep->re_connect_wait);
+		return rpcrdma_ep_put(ep);
	default:
		break;
	}
net/sunrpc/xprtrdma/verbs.c +4 -4
@@ -345,11 +348,11 @@ static struct rdma_cm_id *rpcrdma_create_id(struct rpcrdma_xprt *r_xprt,
 	return ERR_PTR(rc);
 }
 
-static void rpcrdma_ep_put(struct kref *kref)
+static void rpcrdma_ep_destroy(struct kref *kref)
 {
 	struct rpcrdma_ep *ep = container_of(kref, struct rpcrdma_ep, re_kref);
 
-	if (ep->re_id->qp) {
+	if (ep->re_id && ep->re_id->qp) {
 		rdma_destroy_qp(ep->re_id);
 		ep->re_id->qp = NULL;
 	}
@@ -373,9 +376,9 @@ static void rpcrdma_ep_put(struct kref *kref)
 *     %0 if @ep still has a positive kref count, or
 *     %1 if @ep was destroyed successfully.
 */
-static int rpcrdma_ep_destroy(struct rpcrdma_ep *ep)
+static int rpcrdma_ep_put(struct rpcrdma_ep *ep)
 {
-	return kref_put(&ep->re_kref, rpcrdma_ep_put);
+	return kref_put(&ep->re_kref, rpcrdma_ep_destroy);
 }
 
 static int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
@@ -492,11 +495,12 @@ static int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt)
 	return 0;
 

out_destroy:

net/sunrpc/xprtrdma/verbs.c +1 -1
-	rpcrdma_ep_destroy(ep);
+	rpcrdma_ep_put(ep);
	rdma_destroy_id(id);
net/sunrpc/xprtrdma/verbs.c +2 -0
+	return rc;
+

out_free:

	kfree(ep);
net/sunrpc/xprtrdma/verbs.c +0 -1
-	r_xprt->rx_ep = NULL;
	return rc;
}
net/sunrpc/xprtrdma/verbs.c +1 -0
@@ -510,6 +514,7 @@ int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
 {
 	struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 	struct rpcrdma_ep *ep;
+	struct rdma_cm_id *id;
 	int rc;
 

retry:

net/sunrpc/xprtrdma/verbs.c +9 -2
@@ -518,6 +523,7 @@ int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
 	if (rc)
 		return rc;
 	ep = r_xprt->rx_ep;
+	id = ep->re_id;
 
 	ep->re_connect_status = 0;
 	xprt_clear_connected(xprt);
@@ -529,7 +535,10 @@ int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
 	if (rc)
 		goto out;
 
-	rc = rdma_connect(ep->re_id, &ep->re_remote_cma);
+	/* From this point on, CM events can discard our EP */
+	kref_get(&ep->re_kref);
+
+	rc = rdma_connect(id, &ep->re_remote_cma);
 	if (rc)
 		goto out;
 
@@ -546,14 +555,17 @@ int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
 
 	rc = rpcrdma_reqs_setup(r_xprt);
 	if (rc) {
-		rpcrdma_xprt_disconnect(r_xprt);
+		ep->re_connect_status = rc;
 		goto out;
 	}
+
 	rpcrdma_mrs_create(r_xprt);
+	trace_xprtrdma_connect(r_xprt, 0);
+	return rc;
 

out:

net/sunrpc/xprtrdma/verbs.c +2 -2
-	if (rc)
-		ep->re_connect_status = rc;
+	ep->re_connect_status = rc;
+	rpcrdma_xprt_disconnect(r_xprt);
	trace_xprtrdma_connect(r_xprt, rc);
	return rc;
}
net/sunrpc/xprtrdma/verbs.c +9 -3
@@ -570,12 +582,15 @@ int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt)
 */
 void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt)
 {
-	struct rpcrdma_ep *ep = r_xprt->rx_ep;
+	struct rpcrdma_ep *ep;
 	struct rdma_cm_id *id;
 	int rc;
 
+	mutex_lock(&r_xprt->rx_flush);
+
+	ep = r_xprt->rx_ep;
 	if (!ep)
-		return;
+		goto out;
 
 	id = ep->re_id;
 	rc = rdma_disconnect(id);
@@ -587,10 +602,13 @@ void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt)
 	rpcrdma_mrs_destroy(r_xprt);
 	rpcrdma_sendctxs_destroy(r_xprt);
 
-	if (rpcrdma_ep_destroy(ep))
+	if (rpcrdma_ep_put(ep))
 		rdma_destroy_id(id);
 
 	r_xprt->rx_ep = NULL;
+
+out:
+	mutex_unlock(&r_xprt->rx_flush);
 }
 
 /* Fixed-size circular FIFO queue. This implementation is wait-free and
net/sunrpc/xprtrdma/xprt_rdma.h +1 -0
index 0a16fdb09b2c..288645a9437f 100644
@@ -417,6 +417,7 @@ struct rpcrdma_xprt {
 	struct delayed_work	rx_connect_worker;
 	struct rpc_timeout	rx_timeout;
 	struct rpcrdma_stats	rx_stats;
+	struct mutex            rx_flush;
 };
 
 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
Signature · 1 line
2.25.4

--
Chuck Lever

#4 lore raw

Re: [PATCH] xprtrdma: fix EP destruction logic

Quoted 1 line · Chuck Lever
Quoted 1 line · Dan Aloni

[..]

Quoted 6 lines
Quoted 3 lines

- Add a mutex in rpcrdma_ep_destroy to guard against concurrent calls to rpcrdma_xprt_disconnect coming from either rpcrdma_xprt_connect or xprt_rdma_close.

NAK. The RPC client provides appropriate exclusion, please let's not add more serialization that can introduce further deadlocks.

It appeared to me that this exclusion does not works well. As for my considerations, if I am not mistaken from analyzing crashes I've seen:

-> xprt_autoclose (running on xprtiod)

     -> xprt->ops->close
       -> xprt_rdma_close
         -> rpcrdma_xprt_disconnect

and:

    -> xprt_rdma_connect_worker (running on xprtiod)
      -> rpcrdma_xprt_connect
	-> rpcrdma_xprt_disconnect

I understand the rationale or at least the aim that close and connect ops should not be concurrent on the same xprt, however:

* xprt_force_disconnect, is called from various places, queues a call to xprt_autoclose to the background on xprtiod workqueue item, conditioned that !XPRT_LOCKED which is the case for connect that went to the background.
* xprt_rdma_connect also sends xprt_rdma_connect_worker as an xprtiod workqueue item, unconditionally.

So we have two work items that can run in parallel, and I don't see clear gating on this from the code.

Maybe there's a simpler fix for this. Perhaps a cancel_delayed_work_sync(&r_xprt->rx_connect_worker); is appropriate in xprt_rdma_close?

Signature · 1 line
Dan Aloni
#5 lore raw

Re: [PATCH] xprtrdma: fix EP destruction logic

Quoted 39 lines · Dan Aloni
Quoted 1 line · Chuck Lever
Quoted 1 line · Dan Aloni

[..]

Quoted 6 lines
Quoted 3 lines

- Add a mutex in rpcrdma_ep_destroy to guard against concurrent calls to rpcrdma_xprt_disconnect coming from either rpcrdma_xprt_connect or xprt_rdma_close.

NAK. The RPC client provides appropriate exclusion, please let's not add more serialization that can introduce further deadlocks.

It appeared to me that this exclusion does not works well. As for my considerations, if I am not mistaken from analyzing crashes I've seen:

-> xprt_autoclose (running on xprtiod)

    -> xprt->ops->close
      -> xprt_rdma_close
        -> rpcrdma_xprt_disconnect

and:

   -> xprt_rdma_connect_worker (running on xprtiod)
     -> rpcrdma_xprt_connect
	-> rpcrdma_xprt_disconnect

I understand the rationale or at least the aim that close and connect ops should not be concurrent on the same xprt, however:

* xprt_force_disconnect, is called from various places, queues a call to xprt_autoclose to the background on xprtiod workqueue item, conditioned that !XPRT_LOCKED which is the case for connect that went to the background.
* xprt_rdma_connect also sends xprt_rdma_connect_worker as an xprtiod workqueue item, unconditionally.

Quoted 2 lines

So we have two work items that can run in parallel, and I don't see clear gating on this from the code.

If close and connect are being called concurrently on the same xprt, then there is a bug in the generic RPC xprt code. I don't believe that to be the case here.

If xprtrdma invokes force_disconnect during connect processing, XPRT_LOCKED should be held and the close should be delayed.

Quoted 3 lines

Maybe there's a simpler fix for this. Perhaps a cancel_delayed_work_sync(&r_xprt->rx_connect_worker); is appropriate in xprt_rdma_close?

There are simpler, less deadlock-prone, and more justifiable fixes. Please stand by, I will take care of this today.

--
Chuck Lever

#6 lore raw
In reply to Chuck Lever
Quoted 5 lines · Chuck Lever

Anna, please drop this one. It appears to trigger a particularly nasty use-after-free. I'll follow up with a more complete fix soon.

(Yes, a wake-up on connect errors is indeed necessary... but the connect worker needs to be re-organized to deal properly with it).

After sending that patch I also noticed more issues with the management of the EP context. The decref inside the CM handler caused freeing of the EP while connect path still held a reference to it. A KASAN-enabled kernel revealed this easily. I've sent just now a more comprehensive patch to deal with this.

Signature · 1 line
Dan Aloni