Merge pull request #233 from appneta/Enhancement_#93_netmap_performance

#93 send performance upgrades as per discussions with Luggi Rizzo
appneta · Dec 26, 2015 · 6157689 · 6157689
2 parents 2c24cac + 48b2a50
commit 6157689
Show file tree

Hide file tree

Showing 11 changed files with 224 additions and 380 deletions.
diff --git a/configure.ac b/configure.ac
@@ -18,7 +18,7 @@ AC_CONFIG_MACRO_DIR([m4])
 dnl Set version info here!
 MAJOR_VERSION=4
 MINOR_VERSION=1
-MICRO_VERSION=1-beta3
+MICRO_VERSION=1-beta4
 TCPREPLAY_VERSION=$MAJOR_VERSION.$MINOR_VERSION.$MICRO_VERSION
 PACKAGE_URL=http://tcpreplay.appneta.com/
 

diff --git a/docs/CHANGELOG b/docs/CHANGELOG
@@ -1,3 +1,6 @@
+--/--/---- Version 4.1.1-beta4
+    - netmap optimizations (#93)
+
 12/19/2015 Version 4.1.1-beta3
 	- Option --unique-ip accepts --duration (#227)
 	- RFC1624 incremental checksums (#225)
@@ -38,7 +41,7 @@
     - Fix cross compiling (#140)
     - Add Quick TX module for Linux (#136)
 	- Better recovery for missing static libpcap libs (#128)
-	 
+
 09/05/2014 Version 4.0.5
     - Fix build failure with latest netmap 11 (#123)
     - Fix unable to use Zero Copy interfaces on PF_RING (#118)
@@ -70,17 +73,17 @@
 02/04/2014 Version 4.0.3
     - Fix build errors when ENABLE_VERBOSE is not set (#60)
     - Build error on host without tcpdump (#59)
-    
+
 01/17/2014 Version 4.0.2
     - Fix hangs with --mbps command (#54)
 
 01/16/2014 Version 4.0.1
     - Support for netmap version 10 API (#53)
     - Remove deprecated "absolute time" for OS X (#52)
-    - Make dosleep() rentrant to support API concurrency (#51)  
+    - Make dosleep() rentrant to support API concurrency (#51)
     - Remove compiler warnings (#50)
     - Fix slow netmap in Intel GigE "igb" server NICS (#49)
- 
+
 01/05/2014 Version 4.0.0
     - Implemented wiki page for 4.X - http://tcpreplay.appneta.com
     - Support for ARM aarch64 (appneta #43)
@@ -101,7 +104,7 @@
     - Remove unreliable RDTSC option (appneta #16)
     - Switch from BSD to GPLv3 license (appneta #5)
     - Add --unique-ip option to tcpreplay (appneta #4)
-    - Removed sleep-accel option (appneta #2) 
+    - Removed sleep-accel option (appneta #2)
     - Enhance accuracy and performance of --mbps option (appneta #2)
     - Add netmap injector (appneta #1)
     - Properly process IPv6 extension headers (#396)
@@ -211,7 +214,7 @@
     - Re-enable tcpreplay --listnics (#319)
     - Fix sendpacket always reporting using PF_PACKET, even when it doesn't (#322)
     - Fix major packet timing issue under old versions of glibc (#324)
-    
+
 05/17/2008: Version 3.3.1
     - Fix limitation of PF_PACKET only supporting Ethernet (#123)
     - Fix (again) /dev/bpf detection in FreeBSD 8.0 (#292)
@@ -246,7 +249,7 @@
 
 11/01/2007: Version 3.2.3
     - Fix compile issue under Linux and other OS's (#275)
-    
+
 10/31/2007: Version 3.2.2
     - Enable source MAC spoofing for OS X (#142, #151)
         * Tcpreplay now requires OS X 10.5 (Leopard)
@@ -260,15 +263,15 @@
     - Remove flowreplay code (#262)
     - Fix DLT rewrite code causing corrupted ethernet protocol type (#268)
     - Try to fix inet_aton() issue under Solaris (#260)
-    
+
 08/26/2007: Version 3.2
     - Return a more useful error message when tcpprep fails (#187)
     - Add Tomahawk test tool client/server detection algorithm (#186)
     - Improved AutoGen support (#191)
     - Improved documentation (#164, #198)
     - Added Doxygen markup (#176)
     - configure now honors --with-tcpdump flag (#192)
-    - configure now defaults to --enable-64bits which breaks backwards 
+    - configure now defaults to --enable-64bits which breaks backwards
         compatibility in certain situations.  Use --disable-64bits if this
         concerns you (#195)
     - Use safe_free() to detect bugs earlier in development (#197)
@@ -333,8 +336,8 @@
     - Add tcpbridge --verbose mode (#28)
     - Fix compile issues under HP-UX & strictly aligned systems (#141)
     - --enable-dynamic-link supports 64bit libraries (#143)
-    
-    
+
+
 03/22/2007: Version 3.0.beta13
     - Fixed detection of IPv4 packets in libtcpedit on little-endian boxes (#115, #116)
     - Fixed TCP/UDP checksum calculation on little-endian boxes (#126)
@@ -366,7 +369,7 @@
     - Fix tcpprep --include/exclude (#96)
     - Upgrade AutoGen tearoff (libopts) to latest version (5.9.0) (#97)
     - Don't do L4 checksums for non-fragment offset == 0 (#107)
-    
+
 08/07/2006: Version 3.0.beta11
     - Fix distribution to ship missing src/tcpr.h (#73)
     - Add support to tcprewrite to alter output file DLT (#74)
@@ -401,7 +404,7 @@
     - Improve tcpbridge man page
     - Massive tcprewrite fixes & cleanup (#50)
     - Much improved README document
-    - Reorganize packet editing code into a standalone module (tcpedit) 
+    - Reorganize packet editing code into a standalone module (tcpedit)
           which has been librarized. (#5)
     - Strict code cleanup (#27)
     - Fix tcpprep from generating bad cache files (#48)
@@ -503,8 +506,8 @@
       - tcprewrite
 
     - Removed Applications: (Ethereal has better utilities)
-      - capinfo 
-      - pcapmerge 
+      - capinfo
+      - pcapmerge
 
     - Merge fixes from 2.x/stable branch:
       - portmap.c endian bugs

diff --git a/src/Makefile.am b/src/Makefile.am
@@ -62,15 +62,15 @@ endif
 
 tcpreplay_edit_CFLAGS = $(LIBOPTS_CFLAGS) -I.. -Itcpedit $(LNAV_CFLAGS) @LDNETINC@ -DTCPREPLAY -DTCPREPLAY_EDIT -DHAVE_CACHEFILE_SUPPORT
 tcpreplay_edit_LDADD = ./tcpedit/libtcpedit.a ./common/libcommon.a $(LIBSTRL) @LPCAPLIB@ @LDNETLIB@ $(LIBOPTS_LDADD)
-tcpreplay_edit_SOURCES = tcpreplay_edit_opts.c send_packets.c signal_handler.c tcpreplay.c sleep.c tcpreplay_api.c replay.c
+tcpreplay_edit_SOURCES = tcpreplay_edit_opts.c send_packets.c signal_handler.c tcpreplay.c tcpreplay_api.c replay.c
 tcpreplay_edit_OBJECTS: tcpreplay_opts.h
 tcpreplay_edit_opts.h: tcpreplay_edit_opts.c
 
 tcpreplay_edit_opts.c: tcpreplay_opts.def
 	@AUTOGEN@ $(opts_list)  @NETMAPFLAGS@ -DTCPREPLAY_EDIT -b tcpreplay_edit_opts tcpreplay_opts.def
 
 tcpreplay_CFLAGS = $(LIBOPTS_CFLAGS) -I.. $(LNAV_CFLAGS) @LDNETINC@ -DTCPREPLAY
-tcpreplay_SOURCES = tcpreplay_opts.c send_packets.c signal_handler.c tcpreplay.c sleep.c tcpreplay_api.c replay.c
+tcpreplay_SOURCES = tcpreplay_opts.c send_packets.c signal_handler.c tcpreplay.c tcpreplay_api.c replay.c
 tcpreplay_LDADD = ./common/libcommon.a $(LIBSTRL) @LPCAPLIB@ @LDNETLIB@ $(LIBOPTS_LDADD)
 tcpreplay_OBJECTS: tcpreplay_opts.h
 tcpreplay_opts.h: tcpreplay_opts.c
@@ -127,7 +127,7 @@ tcpbridge_LDADD = ./tcpedit/libtcpedit.a ./common/libcommon.a \
 if ENABLE_OSX_FRAMEWORKS
 tcpbridge_LDFLAGS = -framework CoreServices -framework Carbon
 endif
-tcpbridge_SOURCES = tcpbridge_opts.c tcpbridge.c bridge.c sleep.c
+tcpbridge_SOURCES = tcpbridge_opts.c tcpbridge.c bridge.c
 tcpbridge_OBJECTS: tcpbridge_opts.h
 tcpbridge_opts.h: tcpbridge_opts.c
 tcpbridge_opts.c: tcpbridge_opts.def tcpedit/tcpedit_opts.def

diff --git a/src/common/netmap.c b/src/common/netmap.c
@@ -420,7 +420,7 @@ sendpacket_open_netmap(const char *device, char *errbuf, void *arg) {
         }
     }
 
-    dbg(2, "Waiting 4 seconds for phy reset...");
+    dbgx(2, "Waiting %d seconds for phy reset...", ctx->options->netmap_delay);
     sleep(ctx->options->netmap_delay);
     dbg(2, "Ready!");
 
@@ -573,27 +573,19 @@ int sendpacket_send_netmap(void *p, const u_char *data, size_t len)
              */
             ioctl(sp->handle.fd, NIOCTXSYNC, NULL);
 
-            pfd.fd = sp->handle.fd;
-            pfd.events = POLLOUT;
-            pfd.revents = 0;
-            if (poll(&pfd, 1, 1000) <= 0) {
-                if (++sp->tx_timeouts == NETMAP_TX_TIMEOUT_SEC) {
-                    return -1;
+            if ((avail = nm_ring_space(txring)) == 0) {
+                pfd.fd = sp->handle.fd;
+                pfd.events = POLLOUT;
+                pfd.revents = 0;
+                if (poll(&pfd, 1, 1000) <= 0) {
+                    if (++sp->tx_timeouts == NETMAP_TX_TIMEOUT_SEC) {
+                        return -1;
+                    }
+                    return -2;
                 }
-                return -2;
             }
 
             sp->tx_timeouts = 0;
-
-            /*
-             * Do not remove this even though it looks redundant.
-             * Overall performance is increased with this restart
-             * of the TX queue.
-             *
-             * This call increases the number of available slots from
-             * 1 to all that are truly available.
-             */
-            ioctl(sp->handle.fd, NIOCTXSYNC, NULL);
         }
 
         txring = NETMAP_TXRING(sp->nm_if, sp->cur_tx_ring);

diff --git a/src/common/timer.h b/src/common/timer.h
@@ -170,16 +170,6 @@ void timesdiv(struct timespec *tvs, COUNTER div);
 
     typedef struct timeval timestamp_t;
 
-/*
- * starts a timer so we can figure out how long we have
- * been transmitting
- */
-static inline void
-get_packet_timestamp(timestamp_t *ctx)
-{
-    gettimeofday(ctx, NULL);
-}
-
 void init_timestamp(timestamp_t *ctx);