Skip to contentSkip to conversation

svcrdma: clean up error paths.

PATCH Applied Steve Wise 2 participants 4 messages Apr 29, 2009 – May 14, 2009 1 message not in the archive

Subsystems (3)
lore
Share & tools
lore

Steve Wise

author
#0 lore raw

PATCH 1/1 svcrdma: clean up error paths.

These fixes resolved crashes due to resource leak BUG_ON checks. The resource leaks were detected by introducing asynchronous transport errors.

---

 net/sunrpc/xprtrdma/svc_rdma_sendto.c    |    3 +++
 net/sunrpc/xprtrdma/svc_rdma_transport.c |    3 ++-
 2 files changed, 5 insertions(+), 1 deletions(-)
net/sunrpc/xprtrdma/svc_rdma_sendto.c +3 -0
index 6c26a67..8b510c5 100644
@@ -183,6 +183,7 @@ static int fast_reg_xdr(struct svcxprt_rdma *xprt,
 
  fatal_err:
 	printk("svcrdma: Error fast registering memory for xprt %p\n", xprt);
+	vec->frmr = NULL;
 	svc_rdma_put_frmr(xprt, frmr);
 	return -EIO;
 }
@@ -516,6 +517,7 @@ static int send_reply(struct svcxprt_rdma *rdma,
 		       "svcrdma: could not post a receive buffer, err=%d."
 		       "Closing transport %p.\n", ret, rdma);
 		set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
+		svc_rdma_put_frmr(rdma, vec->frmr);
 		svc_rdma_put_context(ctxt, 0);
 		return -ENOTCONN;
 	}
@@ -606,6 +608,7 @@ static int send_reply(struct svcxprt_rdma *rdma,
 	return 0;
 
  err:
+	svc_rdma_unmap_dma(ctxt);
 	svc_rdma_put_frmr(rdma, vec->frmr);
 	svc_rdma_put_context(ctxt, 1);
 	return -EIO;
net/sunrpc/xprtrdma/svc_rdma_transport.c +2 -1
index 3d810e7..4b0c2fa 100644
@@ -520,8 +520,9 @@ int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
 	svc_xprt_get(&xprt->sc_xprt);
 	ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
 	if (ret) {
-		svc_xprt_put(&xprt->sc_xprt);
+		svc_rdma_unmap_dma(ctxt);
 		svc_rdma_put_context(ctxt, 1);
+		svc_xprt_put(&xprt->sc_xprt);
 	}
 	return ret;
 

A message that is not in the archive (sent off-list or not yet received): 20090429191400.29365.36715.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org

#2 lore raw
Quoted 2 lines · Steve Wise

These fixes resolved crashes due to resource leak BUG_ON checks. The resource leaks were detected by introducing asynchronous transport errors.

Thanks, applied for 2.6.30. (And also appropriate for stable (2.6.29), I assume?)

But, could someone take a closer look at the error paths here? Questions:

- svc_rdma_post_recv() does a svc_rdma_put_context() on error-- are you sure its caller needs to as well?
- In send_reply, some of the cleanout is shared between the first return -ENOTCONN and the final err: cleanup. Could we add another err: label and share some of that cleanup?

--b.

Quoted 53 lines

---

 net/sunrpc/xprtrdma/svc_rdma_sendto.c    |    3 +++
 net/sunrpc/xprtrdma/svc_rdma_transport.c |    3 ++-
 2 files changed, 5 insertions(+), 1 deletions(-)
net/sunrpc/xprtrdma/svc_rdma_sendto.c +3 -0
index 6c26a67..8b510c5 100644
@@ -183,6 +183,7 @@ static int fast_reg_xdr(struct svcxprt_rdma *xprt,
 
  fatal_err:
 	printk("svcrdma: Error fast registering memory for xprt %p\n", xprt);
+	vec->frmr = NULL;
 	svc_rdma_put_frmr(xprt, frmr);
 	return -EIO;
 }
@@ -516,6 +517,7 @@ static int send_reply(struct svcxprt_rdma *rdma,
 		       "svcrdma: could not post a receive buffer, err=%d."
 		       "Closing transport %p.\n", ret, rdma);
 		set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
+		svc_rdma_put_frmr(rdma, vec->frmr);
 		svc_rdma_put_context(ctxt, 0);
 		return -ENOTCONN;
 	}
@@ -606,6 +608,7 @@ static int send_reply(struct svcxprt_rdma *rdma,
 	return 0;
 
  err:
+	svc_rdma_unmap_dma(ctxt);
 	svc_rdma_put_frmr(rdma, vec->frmr);
 	svc_rdma_put_context(ctxt, 1);
 	return -EIO;
net/sunrpc/xprtrdma/svc_rdma_transport.c +2 -1
index 3d810e7..4b0c2fa 100644
@@ -520,8 +520,9 @@ int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
 	svc_xprt_get(&xprt->sc_xprt);
 	ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
 	if (ret) {
-		svc_xprt_put(&xprt->sc_xprt);
+		svc_rdma_unmap_dma(ctxt);
 		svc_rdma_put_context(ctxt, 1);
+		svc_xprt_put(&xprt->sc_xprt);
 	}
 	return ret;
 
author
#3 lore raw

Hey Bruce,

Quoted 6 lines · J. Bruce Fields
Quoted 3 lines · Steve Wise

These fixes resolved crashes due to resource leak BUG_ON checks. The resource leaks were detected by introducing asynchronous transport errors.

Quoted 8 lines

Thanks, applied for 2.6.30. (And also appropriate for stable (2.6.29), I assume?)

But, could someone take a closer look at the error paths here? Questions:

- svc_rdma_post_recv() does a svc_rdma_put_context() on error-- are you sure its caller needs to as well?

The svc_rdma_put_context() call inside svc_rdma_post_recv() is for the recv context that was allocated inside that function. The caller, in this case send_reply() also does a svc_rdma_put_context(), but that is for the send context. So I think this is correct.

Quoted 4 lines

- In send_reply, some of the cleanout is shared between the first return -ENOTCONN and the final err: cleanup. Could we add another err: label and share some of that cleanup?

The only common logic I see is the svc_rdma_put_context() call that could be shared. But one case calls it with free_pages == 1 after the pages have been mapped, and the other with 0 since no pages are mapped at that point (when the call to svc_rdma_post_recv() fails). So I'm not sure its worth doing?

Steve.

#4 lore raw
Quoted 28 lines · Steve Wise

Hey Bruce,

Quoted 6 lines · J. Bruce Fields
Quoted 3 lines · Steve Wise

These fixes resolved crashes due to resource leak BUG_ON checks. The resource leaks were detected by introducing asynchronous transport errors.

Quoted 8 lines

Thanks, applied for 2.6.30. (And also appropriate for stable (2.6.29), I assume?)

But, could someone take a closer look at the error paths here? Questions:

- svc_rdma_post_recv() does a svc_rdma_put_context() on error-- are you sure its caller needs to as well?

The svc_rdma_put_context() call inside svc_rdma_post_recv() is for the recv context that was allocated inside that function. The caller, in this case send_reply() also does a svc_rdma_put_context(), but that is
for the send context. So I think this is correct.

Quoted 4 lines

- In send_reply, some of the cleanout is shared between the first return -ENOTCONN and the final err: cleanup. Could we add another err: label and share some of that cleanup?

Quoted 5 lines

The only common logic I see is the svc_rdma_put_context() call that could be shared. But one case calls it with free_pages == 1 after the pages have been mapped, and the other with 0 since no pages are mapped
at that point (when the call to svc_rdma_post_recv() fails). So I'm
not sure its worth doing?

No, I think you're probably right about both of these. Thanks for taking a look.

--b.