[PATCHv2 net-next 0/2] sctp: fully support memory accounting

[PATCHv2 net-next 0/2] sctp: fully support memory accounting

[Xin Long]

sctp memory accounting is added in this patchset by using
these kernel APIs on send side:

  - sk_mem_charge()
  - sk_mem_uncharge()
  - sk_wmem_schedule()
  - sk_under_memory_pressure()
  - sk_mem_reclaim()

and these on receive side:

  - sk_mem_charge()
  - sk_mem_uncharge()
  - sk_rmem_schedule()
  - sk_under_memory_pressure()
  - sk_mem_reclaim()

With sctp memory accounting, we can limit the memory allocation by
either sysctl:

  # sysctl -w net.sctp.sctp_mem="10 20 50"

or cgroup:

  # echo $((8<<14)) > \
    /sys/fs/cgroup/memory/sctp_mem/memory.kmem.tcp.limit_in_bytes

When the socket is under memory pressure, the send side will block
and wait, while the receive side will renege or drop.

v1->v2:
  - add the missing Reported/Tested/Acked/-bys.

Xin Long (2):
  sctp: implement memory accounting on tx path
  sctp: implement memory accounting on rx path

 include/net/sctp/sctp.h |  2 +-
 net/sctp/sm_statefuns.c |  6 ++++--
 net/sctp/socket.c       | 10 ++++++++--
 net/sctp/ulpevent.c     | 19 ++++++++-----------
 net/sctp/ulpqueue.c     |  3 ++-
 5 files changed, 23 insertions(+), 17 deletions(-)

-- 
2.1.0

[PATCHv2 net-next 1/2] sctp: implement memory accounting on tx path

[Xin Long]

Now when sending packets, sk_mem_charge() and sk_mem_uncharge() have been
used to set sk_forward_alloc. We just need to call sk_wmem_schedule() to
check if the allocated should be raised, and call sk_mem_reclaim() to
check if the allocated should be reduced when it's under memory pressure.

If sk_wmem_schedule() returns false, which means no memory is allowed to
allocate, it will block and wait for memory to become available.

Note different from tcp, sctp wait_for_buf happens before allocating any
skb, so memory accounting check is done with the whole msg_len before it
too.

Reported-by: Matteo Croce <mcroce@redhat.com>
Tested-by: Matteo Croce <mcroce@redhat.com>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Acked-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Xin Long <lucien.xin@gmail.com>
---
 net/sctp/socket.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/net/sctp/socket.c b/net/sctp/socket.c
index 6140471..06c6f4a 100644
--- a/net/sctp/socket.c
+++ b/net/sctp/socket.c
@@ -1913,7 +1913,10 @@ static int sctp_sendmsg_to_asoc(struct sctp_association *asoc,
 	if (sctp_wspace(asoc) < (int)msg_len)
 		sctp_prsctp_prune(asoc, sinfo, msg_len - sctp_wspace(asoc));
 
-	if (sctp_wspace(asoc) <= 0) {
+	if (sk_under_memory_pressure(sk))
+		sk_mem_reclaim(sk);
+
+	if (sctp_wspace(asoc) <= 0 || !sk_wmem_schedule(sk, msg_len)) {
 		timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 		err = sctp_wait_for_sndbuf(asoc, &timeo, msg_len);
 		if (err)
@@ -8891,7 +8894,10 @@ static int sctp_wait_for_sndbuf(struct sctp_association *asoc, long *timeo_p,
 			goto do_error;
 		if (signal_pending(current))
 			goto do_interrupted;
-		if ((int)msg_len <= sctp_wspace(asoc))
+		if (sk_under_memory_pressure(sk))
+			sk_mem_reclaim(sk);
+		if ((int)msg_len <= sctp_wspace(asoc) &&
+		    sk_wmem_schedule(sk, msg_len))
 			break;
 
 		/* Let another process have a go.  Since we are going
-- 
2.1.0

[PATCHv2 net-next 2/2] sctp: implement memory accounting on rx path

[Xin Long]

sk_forward_alloc's updating is also done on rx path, but to be consistent
we change to use sk_mem_charge() in sctp_skb_set_owner_r().

In sctp_eat_data(), it's not enough to check sctp_memory_pressure only,
which doesn't work for mem_cgroup_sockets_enabled, so we change to use
sk_under_memory_pressure().

When it's under memory pressure, sk_mem_reclaim() and sk_rmem_schedule()
should be called on both RENEGE or CHUNK DELIVERY path exit the memory
pressure status as soon as possible.

Note that sk_rmem_schedule() is using datalen to make things easy there.

Reported-by: Matteo Croce <mcroce@redhat.com>
Tested-by: Matteo Croce <mcroce@redhat.com>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Acked-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Xin Long <lucien.xin@gmail.com>
---
 include/net/sctp/sctp.h |  2 +-
 net/sctp/sm_statefuns.c |  6 ++++--
 net/sctp/ulpevent.c     | 19 ++++++++-----------
 net/sctp/ulpqueue.c     |  3 ++-
 4 files changed, 15 insertions(+), 15 deletions(-)

diff --git a/include/net/sctp/sctp.h b/include/net/sctp/sctp.h
index 1d13ec3..eefdfa5 100644
--- a/include/net/sctp/sctp.h
+++ b/include/net/sctp/sctp.h
@@ -421,7 +421,7 @@ static inline void sctp_skb_set_owner_r(struct sk_buff *skb, struct sock *sk)
 	/*
 	 * This mimics the behavior of skb_set_owner_r
 	 */
-	sk->sk_forward_alloc -= event->rmem_len;
+	sk_mem_charge(sk, event->rmem_len);
 }
 
 /* Tests if the list has one and only one entry. */
diff --git a/net/sctp/sm_statefuns.c b/net/sctp/sm_statefuns.c
index c9ae340..7dfc34b 100644
--- a/net/sctp/sm_statefuns.c
+++ b/net/sctp/sm_statefuns.c
@@ -6412,13 +6412,15 @@ static int sctp_eat_data(const struct sctp_association *asoc,
 	 * in sctp_ulpevent_make_rcvmsg will drop the frame if we grow our
 	 * memory usage too much
 	 */
-	if (*sk->sk_prot_creator->memory_pressure) {
+	if (sk_under_memory_pressure(sk)) {
 		if (sctp_tsnmap_has_gap(map) &&
 		    (sctp_tsnmap_get_ctsn(map) + 1) == tsn) {
 			pr_debug("%s: under pressure, reneging for tsn:%u\n",
 				 __func__, tsn);
 			deliver = SCTP_CMD_RENEGE;
-		 }
+		} else {
+			sk_mem_reclaim(sk);
+		}
 	}
 
 	/*
diff --git a/net/sctp/ulpevent.c b/net/sctp/ulpevent.c
index 8cb7d98..c2a7478 100644
--- a/net/sctp/ulpevent.c
+++ b/net/sctp/ulpevent.c
@@ -634,8 +634,9 @@ struct sctp_ulpevent *sctp_ulpevent_make_rcvmsg(struct sctp_association *asoc,
 						gfp_t gfp)
 {
 	struct sctp_ulpevent *event = NULL;
-	struct sk_buff *skb;
-	size_t padding, len;
+	struct sk_buff *skb = chunk->skb;
+	struct sock *sk = asoc->base.sk;
+	size_t padding, datalen;
 	int rx_count;
 
 	/*
@@ -646,15 +647,12 @@ struct sctp_ulpevent *sctp_ulpevent_make_rcvmsg(struct sctp_association *asoc,
 	if (asoc->ep->rcvbuf_policy)
 		rx_count = atomic_read(&asoc->rmem_alloc);
 	else
-		rx_count = atomic_read(&asoc->base.sk->sk_rmem_alloc);
+		rx_count = atomic_read(&sk->sk_rmem_alloc);
 
-	if (rx_count >= asoc->base.sk->sk_rcvbuf) {
+	datalen = ntohs(chunk->chunk_hdr->length);
 
-		if ((asoc->base.sk->sk_userlocks & SOCK_RCVBUF_LOCK) ||
-		    (!sk_rmem_schedule(asoc->base.sk, chunk->skb,
-				       chunk->skb->truesize)))
-			goto fail;
-	}
+	if (rx_count >= sk->sk_rcvbuf || !sk_rmem_schedule(sk, skb, datalen))
+		goto fail;
 
 	/* Clone the original skb, sharing the data.  */
 	skb = skb_clone(chunk->skb, gfp);
@@ -681,8 +679,7 @@ struct sctp_ulpevent *sctp_ulpevent_make_rcvmsg(struct sctp_association *asoc,
 	 * The sender should never pad with more than 3 bytes.  The receiver
 	 * MUST ignore the padding bytes.
 	 */
-	len = ntohs(chunk->chunk_hdr->length);
-	padding = SCTP_PAD4(len) - len;
+	padding = SCTP_PAD4(datalen) - datalen;
 
 	/* Fixup cloned skb with just this chunks data.  */
 	skb_trim(skb, chunk->chunk_end - padding - skb->data);
diff --git a/net/sctp/ulpqueue.c b/net/sctp/ulpqueue.c
index 5dde921..770ff1f 100644
--- a/net/sctp/ulpqueue.c
+++ b/net/sctp/ulpqueue.c
@@ -1106,7 +1106,8 @@ void sctp_ulpq_renege(struct sctp_ulpq *ulpq, struct sctp_chunk *chunk,
 			freed += sctp_ulpq_renege_frags(ulpq, needed - freed);
 	}
 	/* If able to free enough room, accept this chunk. */
-	if (freed >= needed) {
+	if (sk_rmem_schedule(asoc->base.sk, chunk->skb, needed) &&
+	    freed >= needed) {
 		int retval = sctp_ulpq_tail_data(ulpq, chunk, gfp);
 		/*
 		 * Enter partial delivery if chunk has not been
-- 
2.1.0

Re: [PATCHv2 net-next 0/2] sctp: fully support memory accounting

[David Miller]

From: Xin Long <lucien.xin@gmail.com>
Date: Mon, 15 Apr 2019 17:15:05 +0800

> sctp memory accounting is added in this patchset by using
> these kernel APIs on send side:
> 
>   - sk_mem_charge()
>   - sk_mem_uncharge()
>   - sk_wmem_schedule()
>   - sk_under_memory_pressure()
>   - sk_mem_reclaim()
> 
> and these on receive side:
> 
>   - sk_mem_charge()
>   - sk_mem_uncharge()
>   - sk_rmem_schedule()
>   - sk_under_memory_pressure()
>   - sk_mem_reclaim()
> 
> With sctp memory accounting, we can limit the memory allocation by
> either sysctl:
> 
>   # sysctl -w net.sctp.sctp_mem="10 20 50"
> 
> or cgroup:
> 
>   # echo $((8<<14)) > \
>     /sys/fs/cgroup/memory/sctp_mem/memory.kmem.tcp.limit_in_bytes
> 
> When the socket is under memory pressure, the send side will block
> and wait, while the receive side will renege or drop.
> 
> v1->v2:
>   - add the missing Reported/Tested/Acked/-bys.

Series applied, thanks Xin.