svn commit: r337386 - in stable/10: share/man/man4 sys/netinet

[Jonathan T. Looney]

Author: jtl
Date: Mon Aug  6 17:46:28 2018
New Revision: 337386
URL: https://svnweb.freebsd.org/changeset/base/337386

Log:
  MFC r337384:
  
  Address concerns about CPU usage while doing TCP reassembly.
  
  Currently, the per-queue limit is a function of the receive buffer
  size and the MSS.  In certain cases (such as connections with large
  receive buffers), the per-queue segment limit can be quite large.
  Because we process segments as a linked list, large queues may not
  perform acceptably.
  
  The better long-term solution is to make the queue more efficient.
  But, in the short-term, we can provide a way for a system
  administrator to set the maximum queue size.
  
  We set the default queue limit to 100.  This is an effort to balance
  performance with a sane resource limit.  Depending on their
  environment, goals, etc., an administrator may choose to modify this
  limit in either direction.
  
  Approved by:	so
  Security:	FreeBSD-SA-18:08.tcp
  Sponsored by:	CVE-2018-6922

Modified:
  stable/10/share/man/man4/tcp.4
  stable/10/sys/netinet/tcp_reass.c
Directory Properties:
  stable/10/   (props changed)

Modified: stable/10/share/man/man4/tcp.4
==============================================================================
--- stable/10/share/man/man4/tcp.4	Mon Aug  6 17:41:53 2018	(r337385)
+++ stable/10/share/man/man4/tcp.4	Mon Aug  6 17:46:28 2018	(r337386)
@@ -436,6 +436,20 @@ no reseeding will occur.
 Reseeding should not be necessary, and will break
 .Dv TIME_WAIT
 recycling for a few minutes.
+.It Va reass.cursegments
+The current total number of segments present in all reassembly queues.
+.It Va reass.maxsegments
+The maximum limit on the total number of segments across all reassembly
+queues.
+The limit can be adjusted as a tunable.
+.It Va reass.maxqueuelen
+The maximum number of segments allowed in each reassembly queue.
+By default, the system chooses a limit based on each TCP connection's
+receive buffer size and maximum segment size (MSS).
+The actual limit applied to a session's reassembly queue will be the lower of
+the system-calculated automatic limit and the user-specified
+.Va reass.maxqueuelen
+limit.
 .It Va rexmit_min , rexmit_slop
 Adjust the retransmit timer calculation for
 .Tn TCP .

Modified: stable/10/sys/netinet/tcp_reass.c
==============================================================================
--- stable/10/sys/netinet/tcp_reass.c	Mon Aug  6 17:41:53 2018	(r337385)
+++ stable/10/sys/netinet/tcp_reass.c	Mon Aug  6 17:46:28 2018	(r337386)
@@ -96,6 +96,11 @@ SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, overflows,
 
 static uma_zone_t tcp_reass_zone;
 
+static u_int tcp_reass_maxqueuelen = 100;
+SYSCTL_UINT(_net_inet_tcp_reass, OID_AUTO, maxqueuelen, CTLFLAG_RWTUN,
+    &tcp_reass_maxqueuelen, 0,
+    "Maximum number of TCP Segments per Reassembly Queue");
+
 /* Initialize TCP reassembly queue */
 static void
 tcp_reass_zone_change(void *tag)
@@ -184,6 +189,10 @@ tcp_reass(struct tcpcb *tp, struct tcphdr *th, int *tl
 	 * socket receive buffer determines our advertised window and grows
 	 * automatically when socket buffer autotuning is enabled. Use it as the
 	 * basis for our queue limit.
+	 *
+	 * However, allow the user to specify a ceiling for the number of
+	 * segments in each queue.
+	 *
 	 * Always let the missing segment through which caused this queue.
 	 * NB: Access to the socket buffer is left intentionally unlocked as we
 	 * can tolerate stale information here.
@@ -194,7 +203,8 @@ tcp_reass(struct tcpcb *tp, struct tcphdr *th, int *tl
 	 * is understood.
 	 */
 	if ((th->th_seq != tp->rcv_nxt || !TCPS_HAVEESTABLISHED(tp->t_state)) &&
-	    tp->t_segqlen >= (so->so_rcv.sb_hiwat / tp->t_maxseg) + 1) {
+	    tp->t_segqlen >= min((so->so_rcv.sb_hiwat / tp->t_maxseg) + 1,
+	    tcp_reass_maxqueuelen)) {
 		tcp_reass_overflows++;
 		TCPSTAT_INC(tcps_rcvmemdrop);
 		m_freem(m);