Skip to contentSkip to conversation

SUNRPC: Avoid digging into the ATOMIC pool

RFC PATCH In review Chuck Lever 2 participants 6 messages Mar 4, 2019

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 SUNRPC: Avoid digging into the ATOMIC pool

Page allocation requests made when the SPARSE_PAGES flag is set are allowed to fail, and are not critical. No need to spend a rare resource.

---
 net/sunrpc/socklib.c           |    2 +-
 net/sunrpc/xprtrdma/rpc_rdma.c |    2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
net/sunrpc/socklib.c +1 -1
index 7e55cfc..9faea12 100644
@@ -106,7 +106,7 @@ static size_t xdr_skb_read_and_csum_bits(struct xdr_skb_reader *desc, void *to,
 		/* ACL likes to be lazy in allocating pages - ACLs
 		 * are small by default but can get huge. */
 		if ((xdr->flags & XDRBUF_SPARSE_PAGES) && *ppage == NULL) {
-			*ppage = alloc_page(GFP_ATOMIC);
+			*ppage = alloc_page(GFP_NOWAIT | __GFP_NOWARN);
 			if (unlikely(*ppage == NULL)) {
 				if (copied == 0)
 					copied = -ENOMEM;
net/sunrpc/xprtrdma/rpc_rdma.c +1 -1
index 6c1fb27..b759b16 100644
@@ -238,7 +238,7 @@ static bool rpcrdma_results_inline(struct rpcrdma_xprt *r_xprt,
 		 */
 		if (unlikely(xdrbuf->flags & XDRBUF_SPARSE_PAGES)) {
 			if (!*ppages)
-				*ppages = alloc_page(GFP_ATOMIC);
+				*ppages = alloc_page(GFP_NOWAIT | __GFP_NOWARN);
 			if (!*ppages)
 				return -ENOBUFS;
 		}
#1 lore raw
Quoted 23 lines · Chuck Lever

Page allocation requests made when the SPARSE_PAGES flag is set are allowed to fail, and are not critical. No need to spend a rare resource.

---
 net/sunrpc/socklib.c           |    2 +-
 net/sunrpc/xprtrdma/rpc_rdma.c |    2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
net/sunrpc/socklib.c +1 -1
index 7e55cfc..9faea12 100644
@@ -106,7 +106,7 @@ static size_t xdr_skb_read_and_csum_bits(struct
 xdr_skb_reader *desc, void *to,
 		/* ACL likes to be lazy in allocating pages - ACLs
 		 * are small by default but can get huge. */
 		if ((xdr->flags & XDRBUF_SPARSE_PAGES) && *ppage ==
 NULL) {
-			*ppage = alloc_page(GFP_ATOMIC);
+			*ppage = alloc_page(GFP_NOWAIT | __GFP_NOWARN);
 			if (unlikely(*ppage == NULL)) {

net/sunrpc/socklib.c static size_t xdr_skb_read_and_csum_bits(struct xdr_skb_reader *desc, void *to, · line 112

Quoted 2 lines
net/sunrpc/socklib.c +0 -0
 				if (copied == 0)
 					copied = -ENOMEM;

Hmm... Can't we make this GFP_KERNEL? I can't see that we're holding any locks here. It does look like svc_udp_recvfrom() is wrapping the call to csum_partial_copy_to_xdr() in a local_bh_disable()/enable() pair, but is that actually needed? If so, for what?

Quoted 15 lines
net/sunrpc/xprtrdma/rpc_rdma.c +1 -1
index 6c1fb27..b759b16 100644
@@ -238,7 +238,7 @@ static bool rpcrdma_results_inline(struct
 rpcrdma_xprt *r_xprt,
 		 */
 		if (unlikely(xdrbuf->flags & XDRBUF_SPARSE_PAGES)) {
 			if (!*ppages)
-				*ppages = alloc_page(GFP_ATOMIC);
+				*ppages = alloc_page(GFP_NOWAIT |
 __GFP_NOWARN);
 			if (!*ppages)

net/sunrpc/xprtrdma/rpc_rdma.c static bool rpcrdma_results_inline(struct rpcrdma_xprt *r_xprt, · line 244

Quoted 2 lines
net/sunrpc/xprtrdma/rpc_rdma.c +0 -0
 				return -ENOBUFS;
 		}

Cheers
Trond

Signature · 3 lines
Trond Myklebust
Linux NFS client maintainer, Hammerspace
[email protected]
#2 lore raw
Quoted 25 lines · Trond Myklebust
Quoted 20 lines · Chuck Lever

Page allocation requests made when the SPARSE_PAGES flag is set are allowed to fail, and are not critical. No need to spend a rare resource.

---

net/sunrpc/socklib.c           |    2 +-
net/sunrpc/xprtrdma/rpc_rdma.c |    2 +-
2 files changed, 2 insertions(+), 2 deletions(-)
net/sunrpc/socklib.c +0 -0
index 7e55cfc..9faea12 100644
@@ -106,7 +106,7 @@ static size_t xdr_skb_read_and_csum_bits(struct
 xdr_skb_reader *desc, void *to,
 		/* ACL likes to be lazy in allocating pages - ACLs
 		 * are small by default but can get huge. */
 		if ((xdr->flags & XDRBUF_SPARSE_PAGES) && *ppage ==
 NULL) {
Quoted 8 lines
Quoted 5 lines
net/sunrpc/socklib.c +1 -1
-			*ppage = alloc_page(GFP_ATOMIC);
+			*ppage = alloc_page(GFP_NOWAIT | __GFP_NOWARN);
 			if (unlikely(*ppage == NULL)) {
				if (copied == 0)
					copied = -ENOMEM;

Hmm... Can't we make this GFP_KERNEL? I can't see that we're holding any locks here.

We're holding the transport send lock. Since this is an order 0 allocation it's unlikely to fail even with a less aggressive request like NOWAIT.

Quoted 16 lines

It does look like svc_udp_recvfrom() is wrapping the call to csum_partial_copy_to_xdr() in a local_bh_disable()/enable() pair, but is that actually needed? If so, for what?

Quoted 12 lines
net/sunrpc/xprtrdma/rpc_rdma.c +1 -1
index 6c1fb27..b759b16 100644
@@ -238,7 +238,7 @@ static bool rpcrdma_results_inline(struct
 rpcrdma_xprt *r_xprt,
 		 */
 		if (unlikely(xdrbuf->flags & XDRBUF_SPARSE_PAGES)) {
 			if (!*ppages)
-				*ppages = alloc_page(GFP_ATOMIC);
+				*ppages = alloc_page(GFP_NOWAIT |
Quoted 12 lines
Quoted 5 lines
net/sunrpc/xprtrdma/rpc_rdma.c +0 -0
 __GFP_NOWARN);
 			if (!*ppages)
				return -ENOBUFS;
		}

Cheers
Trond

Signature · 3 lines
Trond Myklebust
Linux NFS client maintainer, Hammerspace
[email protected]

--
Chuck Lever

#3 lore raw
Quoted 36 lines · Chuck Lever
Quoted 27 lines

On Mar 4, 2019, at 11:02 AM, Trond Myklebust <
[email protected]> wrote:

Quoted 23 lines · Chuck Lever

Page allocation requests made when the SPARSE_PAGES flag is set are
allowed to fail, and are not critical. No need to spend a rare resource.

---

net/sunrpc/socklib.c           |    2 +-
net/sunrpc/xprtrdma/rpc_rdma.c |    2 +-
2 files changed, 2 insertions(+), 2 deletions(-)
net/sunrpc/socklib.c +0 -0
index 7e55cfc..9faea12 100644
@@ -106,7 +106,7 @@ static size_t

xdr_skb_read_and_csum_bits(struct
xdr_skb_reader *desc, void *to,

		/* ACL likes to be lazy in allocating pages - ACLs
		 * are small by default but can get huge. */
		if ((xdr->flags & XDRBUF_SPARSE_PAGES) && *ppage ==

NULL) {

net/sunrpc/socklib.c +0 -1
-			*ppage = alloc_page(GFP_ATOMIC);
Quoted 8 lines
Quoted 4 lines
net/sunrpc/socklib.c +1 -0
+			*ppage = alloc_page(GFP_NOWAIT | __GFP_NOWARN);
			if (unlikely(*ppage == NULL)) {
				if (copied == 0)
					copied = -ENOMEM;

Hmm... Can't we make this GFP_KERNEL? I can't see that we're holding
any locks here.

Quoted 4 lines

We're holding the transport send lock. Since this is an order 0 allocation
it's unlikely to fail even with a less aggressive request like NOWAIT.

We should only be holding the receive mutex. That prevents others from closing the socket (or starting a second receive worker) but that's all.

OTOH, the only thing using this mechanism on UDP should be the NFSv2/v3 ACL receive code, so I'm not really sure anyone cares.

Quoted 34 lines
Quoted 16 lines

It does look like svc_udp_recvfrom() is wrapping the call to csum_partial_copy_to_xdr() in a local_bh_disable()/enable() pair, but is that actually needed? If so, for what?

Quoted 12 lines
net/sunrpc/xprtrdma/rpc_rdma.c +1 -1
index 6c1fb27..b759b16 100644
@@ -238,7 +238,7 @@ static bool rpcrdma_results_inline(struct
 rpcrdma_xprt *r_xprt,
 		 */
 		if (unlikely(xdrbuf->flags & XDRBUF_SPARSE_PAGES)) {
 			if (!*ppages)
-				*ppages = alloc_page(GFP_ATOMIC);
+				*ppages = alloc_page(GFP_NOWAIT |
Quoted 12 lines
Quoted 5 lines
net/sunrpc/xprtrdma/rpc_rdma.c +0 -0
 __GFP_NOWARN);
 			if (!*ppages)
				return -ENOBUFS;
		}

Cheers
Trond

Signature · 3 lines
Trond Myklebust
Linux NFS client maintainer, Hammerspace
[email protected]

--
Chuck Lever

Signature · 3 lines
Trond Myklebust
Linux NFS client maintainer, Hammerspace
[email protected]
#4 lore raw
Quoted 41 lines · Trond Myklebust
Quoted 36 lines · Chuck Lever
Quoted 35 lines

On Mar 4, 2019, at 11:02 AM, Trond Myklebust <
[email protected]> wrote:

Quoted 24 lines · Chuck Lever

Page allocation requests made when the SPARSE_PAGES flag is set are
allowed to fail, and are not critical. No need to spend a rare resource.

---

net/sunrpc/socklib.c           |    2 +-
net/sunrpc/xprtrdma/rpc_rdma.c |    2 +-
2 files changed, 2 insertions(+), 2 deletions(-)
net/sunrpc/socklib.c +0 -0
index 7e55cfc..9faea12 100644
@@ -106,7 +106,7 @@ static size_t

xdr_skb_read_and_csum_bits(struct
xdr_skb_reader *desc, void *to,

		/* ACL likes to be lazy in allocating pages - ACLs
		 * are small by default but can get huge. */
		if ((xdr->flags & XDRBUF_SPARSE_PAGES) && *ppage ==

NULL) {

net/sunrpc/socklib.c +1 -1
-			*ppage = alloc_page(GFP_ATOMIC);
+			*ppage = alloc_page(GFP_NOWAIT | __GFP_NOWARN);
Quoted 3 lines
			if (unlikely(*ppage == NULL)) {
				if (copied == 0)
					copied = -ENOMEM;

Hmm... Can't we make this GFP_KERNEL? I can't see that we're holding
any locks here.

Quoted 8 lines
Quoted 4 lines

We're holding the transport send lock. Since this is an order 0 allocation
it's unlikely to fail even with a less aggressive request like NOWAIT.

We should only be holding the receive mutex. That prevents others from closing the socket (or starting a second receive worker) but that's all.

OK, I guess the UDP and RPC-over-RDMA cases are different. For RDMA, this allocation is done in ->send_request.

Quoted 2 lines

OTOH, the only thing using this mechanism on UDP should be the NFSv2/v3 ACL receive code, so I'm not really sure anyone cares.

True. However, as I've said before, the right fix for this is to get rid of SPARSE_PAGES and ensure that NFS always allocates enough receive buffer space during Call encoding. But it's not clear anyone cares enough.

Will the current patch be acceptable, or do you want me to use a sharper stick when allocating?

Quoted 29 lines
Quoted 29 lines
Quoted 28 lines

It does look like svc_udp_recvfrom() is wrapping the call to csum_partial_copy_to_xdr() in a local_bh_disable()/enable() pair, but is that actually needed? If so, for what?

Quoted 14 lines
net/sunrpc/xprtrdma/rpc_rdma.c +1 -1
index 6c1fb27..b759b16 100644
@@ -238,7 +238,7 @@ static bool rpcrdma_results_inline(struct
 rpcrdma_xprt *r_xprt,
 		 */
 		if (unlikely(xdrbuf->flags & XDRBUF_SPARSE_PAGES)) {
 			if (!*ppages)
-				*ppages = alloc_page(GFP_ATOMIC);
+				*ppages = alloc_page(GFP_NOWAIT |
 __GFP_NOWARN);
 			if (!*ppages)
Quoted 3 lines
				return -ENOBUFS;
		}

Cheers
Trond

Signature · 3 lines
Trond Myklebust
Linux NFS client maintainer, Hammerspace
[email protected]
Quoted 9 lines
Quoted 5 lines

--
Chuck Lever

Signature · 3 lines
Trond Myklebust
Linux NFS client maintainer, Hammerspace
[email protected]

--
Chuck Lever

#5 lore raw
Quoted 50 lines · Chuck Lever
Quoted 50 lines

On Mar 4, 2019, at 11:09 AM, Trond Myklebust <
[email protected]> wrote:

Quoted 45 lines · Chuck Lever
Quoted 32 lines

On Mar 4, 2019, at 11:02 AM, Trond Myklebust <
[email protected]> wrote:

Quoted 28 lines · Chuck Lever

Page allocation requests made when the SPARSE_PAGES flag is set
are
allowed to fail, and are not critical. No need to spend a rare
resource.

---

net/sunrpc/socklib.c           |    2 +-
net/sunrpc/xprtrdma/rpc_rdma.c |    2 +-
2 files changed, 2 insertions(+), 2 deletions(-)
net/sunrpc/socklib.c +0 -0
index 7e55cfc..9faea12 100644
@@ -106,7 +106,7 @@ static size_t

xdr_skb_read_and_csum_bits(struct
xdr_skb_reader *desc, void *to,

		/* ACL likes to be lazy in allocating pages -
ACLs
		 * are small by default but can get huge. */
		if ((xdr->flags & XDRBUF_SPARSE_PAGES) &&

*ppage ==
NULL) {

net/sunrpc/socklib.c +1 -1
-			*ppage = alloc_page(GFP_ATOMIC);
+			*ppage = alloc_page(GFP_NOWAIT |
Quoted 8 lines
Quoted 4 lines
__GFP_NOWARN);
			if (unlikely(*ppage == NULL)) {
				if (copied == 0)
					copied = -ENOMEM;

Hmm... Can't we make this GFP_KERNEL? I can't see that we're holding
any locks here.

We're holding the transport send lock. Since this is an order 0 allocation
it's unlikely to fail even with a less aggressive request like NOWAIT.

Quoted 7 lines
Quoted 4 lines

We should only be holding the receive mutex. That prevents others from
closing the socket (or starting a second receive worker) but that's all.

OK, I guess the UDP and RPC-over-RDMA cases are different. For RDMA, this allocation is done in ->send_request.

Right. My comment was only intended to apply to the UDP case.

Quoted 11 lines
Quoted 3 lines

OTOH, the only thing using this mechanism on UDP should be the NFSv2/v3
ACL receive code, so I'm not really sure anyone cares.

True. However, as I've said before, the right fix for this is to get rid of SPARSE_PAGES and ensure that NFS always allocates enough receive buffer space during Call encoding. But it's not clear anyone cares enough.

Will the current patch be acceptable, or do you want me to use a sharper stick when allocating?

It should be fine.

Quoted 49 lines
Quoted 34 lines
Quoted 34 lines
Quoted 33 lines

It does look like svc_udp_recvfrom() is wrapping the call to csum_partial_copy_to_xdr() in a local_bh_disable()/enable()
pair, but is that actually needed? If so, for what?

Quoted 15 lines
net/sunrpc/xprtrdma/rpc_rdma.c +1 -1
index 6c1fb27..b759b16 100644
@@ -238,7 +238,7 @@ static bool rpcrdma_results_inline(struct
 rpcrdma_xprt *r_xprt,
 		 */
 		if (unlikely(xdrbuf->flags &
 XDRBUF_SPARSE_PAGES)) {
 			if (!*ppages)
-				*ppages =
 alloc_page(GFP_ATOMIC);
+				*ppages = alloc_page(GFP_NOWAIT

|

Quoted 5 lines
__GFP_NOWARN);
			if (!*ppages)
				return -ENOBUFS;
		}

Cheers
Trond

Signature · 3 lines
Trond Myklebust
Linux NFS client maintainer, Hammerspace
[email protected]
Quoted 9 lines
Quoted 5 lines

--
Chuck Lever

Signature · 3 lines
Trond Myklebust
Linux NFS client maintainer, Hammerspace
[email protected]

--
Chuck Lever

Signature · 5 lines
Trond Myklebust
CTO, Hammerspace Inc
4300 El Camino Real, Suite 105
Los Altos, CA 94022
www.hammer.space