selftests: net: add tcp_mmap program
authorEric Dumazet <edumazet@google.com>
Mon, 16 Apr 2018 17:33:39 +0000 (10:33 -0700)
committerDavid S. Miller <davem@davemloft.net>
Mon, 16 Apr 2018 22:26:37 +0000 (18:26 -0400)
This is a reference program showing how mmap() can be used
on TCP flows to implement receive zero copy.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tools/testing/selftests/net/Makefile
tools/testing/selftests/net/tcp_mmap.c [new file with mode: 0644]

index 785fc18..23e725f 100644 (file)
@@ -8,9 +8,11 @@ TEST_PROGS := run_netsocktests run_afpackettests test_bpf.sh netdevice.sh rtnetl
 TEST_PROGS += fib_tests.sh fib-onlink-tests.sh pmtu.sh
 TEST_GEN_FILES =  socket
 TEST_GEN_FILES += psock_fanout psock_tpacket msg_zerocopy
+TEST_GEN_FILES += tcp_mmap
 TEST_GEN_PROGS = reuseport_bpf reuseport_bpf_cpu reuseport_bpf_numa
 TEST_GEN_PROGS += reuseport_dualstack reuseaddr_conflict
 
 include ../lib.mk
 
 $(OUTPUT)/reuseport_bpf_numa: LDFLAGS += -lnuma
+$(OUTPUT)/tcp_mmap: LDFLAGS += -lpthread
diff --git a/tools/testing/selftests/net/tcp_mmap.c b/tools/testing/selftests/net/tcp_mmap.c
new file mode 100644 (file)
index 0000000..dea342f
--- /dev/null
@@ -0,0 +1,437 @@
+/*
+ * Copyright 2018 Google Inc.
+ * Author: Eric Dumazet (edumazet@google.com)
+ *
+ * Reference program demonstrating tcp mmap() usage,
+ * and SO_RCVLOWAT hints for receiver.
+ *
+ * Note : NIC with header split is needed to use mmap() on TCP :
+ * Each incoming frame must be a multiple of PAGE_SIZE bytes of TCP payload.
+ *
+ * How to use on loopback interface :
+ *
+ *  ifconfig lo mtu 61512  # 15*4096 + 40 (ipv6 header) + 32 (TCP with TS option header)
+ *  tcp_mmap -s -z &
+ *  tcp_mmap -H ::1 -z
+ *
+ *  Or leave default lo mtu, but use -M option to set TCP_MAXSEG option to (4096 + 12)
+ *      (4096 : page size on x86, 12: TCP TS option length)
+ *  tcp_mmap -s -z -M $((4096+12)) &
+ *  tcp_mmap -H ::1 -z -M $((4096+12))
+ *
+ * Note: -z option on sender uses MSG_ZEROCOPY, which forces a copy when packets go through loopback interface.
+ *       We might use sendfile() instead, but really this test program is about mmap(), for receivers ;)
+ *
+ * $ ./tcp_mmap -s &                                 # Without mmap()
+ * $ for i in {1..4}; do ./tcp_mmap -H ::1 -z ; done
+ * received 32768 MB (0 % mmap'ed) in 14.1157 s, 19.4732 Gbit
+ *   cpu usage user:0.057 sys:7.815, 240.234 usec per MB, 65531 c-switches
+ * received 32768 MB (0 % mmap'ed) in 14.6833 s, 18.7204 Gbit
+ *  cpu usage user:0.043 sys:8.103, 248.596 usec per MB, 65524 c-switches
+ * received 32768 MB (0 % mmap'ed) in 11.143 s, 24.6682 Gbit
+ *   cpu usage user:0.044 sys:6.576, 202.026 usec per MB, 65519 c-switches
+ * received 32768 MB (0 % mmap'ed) in 14.9056 s, 18.4413 Gbit
+ *   cpu usage user:0.036 sys:8.193, 251.129 usec per MB, 65530 c-switches
+ * $ kill %1   # kill tcp_mmap server
+ *
+ * $ ./tcp_mmap -s -z &                              # With mmap()
+ * $ for i in {1..4}; do ./tcp_mmap -H ::1 -z ; done
+ * received 32768 MB (99.9939 % mmap'ed) in 6.73792 s, 40.7956 Gbit
+ *   cpu usage user:0.045 sys:2.827, 87.6465 usec per MB, 65532 c-switches
+ * received 32768 MB (99.9939 % mmap'ed) in 7.26732 s, 37.8238 Gbit
+ *   cpu usage user:0.037 sys:3.087, 95.3369 usec per MB, 65532 c-switches
+ * received 32768 MB (99.9939 % mmap'ed) in 7.61661 s, 36.0893 Gbit
+ *   cpu usage user:0.046 sys:3.559, 110.016 usec per MB, 65529 c-switches
+ * received 32768 MB (99.9939 % mmap'ed) in 7.43764 s, 36.9577 Gbit
+ *   cpu usage user:0.035 sys:3.467, 106.873 usec per MB, 65530 c-switches
+ *
+ * License (GPLv2):
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE. * See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
+ */
+#define _GNU_SOURCE
+#include <pthread.h>
+#include <sys/types.h>
+#include <fcntl.h>
+#include <error.h>
+#include <sys/socket.h>
+#include <sys/mman.h>
+#include <sys/resource.h>
+#include <unistd.h>
+#include <string.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <errno.h>
+#include <time.h>
+#include <sys/time.h>
+#include <netinet/in.h>
+#include <netinet/tcp.h>
+#include <arpa/inet.h>
+#include <poll.h>
+
+#ifndef MSG_ZEROCOPY
+#define MSG_ZEROCOPY    0x4000000
+#endif
+
+#define FILE_SZ (1UL << 35)
+static int cfg_family = AF_INET6;
+static socklen_t cfg_alen = sizeof(struct sockaddr_in6);
+static int cfg_port = 8787;
+
+static int rcvbuf; /* Default: autotuning.  Can be set with -r <integer> option */
+static int sndbuf; /* Default: autotuning.  Can be set with -w <integer> option */
+static int zflg; /* zero copy option. (MSG_ZEROCOPY for sender, mmap() for receiver */
+static int xflg; /* hash received data (simple xor) (-h option) */
+static int keepflag; /* -k option: receiver shall keep all received file in memory (no munmap() calls) */
+
+static int chunk_size  = 512*1024;
+
+unsigned long htotal;
+
+static inline void prefetch(const void *x)
+{
+#if defined(__x86_64__)
+       asm volatile("prefetcht0 %P0" : : "m" (*(const char *)x));
+#endif
+}
+
+void hash_zone(void *zone, unsigned int length)
+{
+       unsigned long temp = htotal;
+
+       while (length >= 8*sizeof(long)) {
+               prefetch(zone + 384);
+               temp ^= *(unsigned long *)zone;
+               temp ^= *(unsigned long *)(zone + sizeof(long));
+               temp ^= *(unsigned long *)(zone + 2*sizeof(long));
+               temp ^= *(unsigned long *)(zone + 3*sizeof(long));
+               temp ^= *(unsigned long *)(zone + 4*sizeof(long));
+               temp ^= *(unsigned long *)(zone + 5*sizeof(long));
+               temp ^= *(unsigned long *)(zone + 6*sizeof(long));
+               temp ^= *(unsigned long *)(zone + 7*sizeof(long));
+               zone += 8*sizeof(long);
+               length -= 8*sizeof(long);
+       }
+       while (length >= 1) {
+               temp ^= *(unsigned char *)zone;
+               zone += 1;
+               length--;
+       }
+       htotal = temp;
+}
+
+void *child_thread(void *arg)
+{
+       unsigned long total_mmap = 0, total = 0;
+       unsigned long delta_usec;
+       int flags = MAP_SHARED;
+       struct timeval t0, t1;
+       char *buffer = NULL;
+       void *oaddr = NULL;
+       double throughput;
+       struct rusage ru;
+       int lu, fd;
+
+       fd = (int)(unsigned long)arg;
+
+       gettimeofday(&t0, NULL);
+
+       fcntl(fd, F_SETFL, O_NDELAY);
+       buffer = malloc(chunk_size);
+       if (!buffer) {
+               perror("malloc");
+               goto error;
+       }
+       while (1) {
+               struct pollfd pfd = { .fd = fd, .events = POLLIN, };
+               int sub;
+
+               poll(&pfd, 1, 10000);
+               if (zflg) {
+                       void *naddr;
+
+                       naddr = mmap(oaddr, chunk_size, PROT_READ, flags, fd, 0);
+                       if (naddr == (void *)-1) {
+                               if (errno == EAGAIN) {
+                                       /* That is if SO_RCVLOWAT is buggy */
+                                       usleep(1000);
+                                       continue;
+                               }
+                               if (errno == EINVAL) {
+                                       flags = MAP_SHARED;
+                                       oaddr = NULL;
+                                       goto fallback;
+                               }
+                               if (errno != EIO)
+                                       perror("mmap()");
+                               break;
+                       }
+                       total_mmap += chunk_size;
+                       if (xflg)
+                               hash_zone(naddr, chunk_size);
+                       total += chunk_size;
+                       if (!keepflag) {
+                               flags |= MAP_FIXED;
+                               oaddr = naddr;
+                       }
+                       continue;
+               }
+fallback:
+               sub = 0;
+               while (sub < chunk_size) {
+                       lu = read(fd, buffer + sub, chunk_size - sub);
+                       if (lu == 0)
+                               goto end;
+                       if (lu < 0)
+                               break;
+                       if (xflg)
+                               hash_zone(buffer + sub, lu);
+                       total += lu;
+                       sub += lu;
+               }
+       }
+end:
+       gettimeofday(&t1, NULL);
+       delta_usec = (t1.tv_sec - t0.tv_sec) * 1000000 + t1.tv_usec - t0.tv_usec;
+
+       throughput = 0;
+       if (delta_usec)
+               throughput = total * 8.0 / (double)delta_usec / 1000.0;
+       getrusage(RUSAGE_THREAD, &ru);
+       if (total > 1024*1024) {
+               unsigned long total_usec;
+               unsigned long mb = total >> 20;
+               total_usec = 1000000*ru.ru_utime.tv_sec + ru.ru_utime.tv_usec +
+                            1000000*ru.ru_stime.tv_sec + ru.ru_stime.tv_usec;
+               printf("received %lg MB (%lg %% mmap'ed) in %lg s, %lg Gbit\n"
+                      "  cpu usage user:%lg sys:%lg, %lg usec per MB, %lu c-switches\n",
+                               total / (1024.0 * 1024.0),
+                               100.0*total_mmap/total,
+                               (double)delta_usec / 1000000.0,
+                               throughput,
+                               (double)ru.ru_utime.tv_sec + (double)ru.ru_utime.tv_usec / 1000000.0,
+                               (double)ru.ru_stime.tv_sec + (double)ru.ru_stime.tv_usec / 1000000.0,
+                               (double)total_usec/mb,
+                               ru.ru_nvcsw);
+       }
+error:
+       free(buffer);
+       close(fd);
+       pthread_exit(0);
+}
+
+static void apply_rcvsnd_buf(int fd)
+{
+       if (rcvbuf && setsockopt(fd, SOL_SOCKET,
+                                SO_RCVBUF, &rcvbuf, sizeof(rcvbuf)) == -1) {
+               perror("setsockopt SO_RCVBUF");
+       }
+
+       if (sndbuf && setsockopt(fd, SOL_SOCKET,
+                                SO_SNDBUF, &sndbuf, sizeof(sndbuf)) == -1) {
+               perror("setsockopt SO_SNDBUF");
+       }
+}
+
+
+static void setup_sockaddr(int domain, const char *str_addr,
+                          struct sockaddr_storage *sockaddr)
+{
+       struct sockaddr_in6 *addr6 = (void *) sockaddr;
+       struct sockaddr_in *addr4 = (void *) sockaddr;
+
+       switch (domain) {
+       case PF_INET:
+               memset(addr4, 0, sizeof(*addr4));
+               addr4->sin_family = AF_INET;
+               addr4->sin_port = htons(cfg_port);
+               if (str_addr &&
+                   inet_pton(AF_INET, str_addr, &(addr4->sin_addr)) != 1)
+                       error(1, 0, "ipv4 parse error: %s", str_addr);
+               break;
+       case PF_INET6:
+               memset(addr6, 0, sizeof(*addr6));
+               addr6->sin6_family = AF_INET6;
+               addr6->sin6_port = htons(cfg_port);
+               if (str_addr &&
+                   inet_pton(AF_INET6, str_addr, &(addr6->sin6_addr)) != 1)
+                       error(1, 0, "ipv6 parse error: %s", str_addr);
+               break;
+       default:
+               error(1, 0, "illegal domain");
+       }
+}
+
+static void do_accept(int fdlisten)
+{
+       if (setsockopt(fdlisten, SOL_SOCKET, SO_RCVLOWAT,
+                      &chunk_size, sizeof(chunk_size)) == -1) {
+               perror("setsockopt SO_RCVLOWAT");
+       }
+
+       apply_rcvsnd_buf(fdlisten);
+
+       while (1) {
+               struct sockaddr_in addr;
+               socklen_t addrlen = sizeof(addr);
+               pthread_t th;
+               int fd, res;
+
+               fd = accept(fdlisten, (struct sockaddr *)&addr, &addrlen);
+               if (fd == -1) {
+                       perror("accept");
+                       continue;
+               }
+               res = pthread_create(&th, NULL, child_thread,
+                                    (void *)(unsigned long)fd);
+               if (res) {
+                       errno = res;
+                       perror("pthread_create");
+                       close(fd);
+               }
+       }
+}
+
+int main(int argc, char *argv[])
+{
+       struct sockaddr_storage listenaddr, addr;
+       unsigned int max_pacing_rate = 0;
+       unsigned long total = 0;
+       char *host = NULL;
+       int fd, c, on = 1;
+       char *buffer;
+       int sflg = 0;
+       int mss = 0;
+
+       while ((c = getopt(argc, argv, "46p:svr:w:H:zxkP:M:")) != -1) {
+               switch (c) {
+               case '4':
+                       cfg_family = PF_INET;
+                       cfg_alen = sizeof(struct sockaddr_in);
+                       break;
+               case '6':
+                       cfg_family = PF_INET6;
+                       cfg_alen = sizeof(struct sockaddr_in6);
+                       break;
+               case 'p':
+                       cfg_port = atoi(optarg);
+                       break;
+               case 'H':
+                       host = optarg;
+                       break;
+               case 's': /* server : listen for incoming connections */
+                       sflg++;
+                       break;
+               case 'r':
+                       rcvbuf = atoi(optarg);
+                       break;
+               case 'w':
+                       sndbuf = atoi(optarg);
+                       break;
+               case 'z':
+                       zflg = 1;
+                       break;
+               case 'M':
+                       mss = atoi(optarg);
+                       break;
+               case 'x':
+                       xflg = 1;
+                       break;
+               case 'k':
+                       keepflag = 1;
+                       break;
+               case 'P':
+                       max_pacing_rate = atoi(optarg) ;
+                       break;
+               default:
+                       exit(1);
+               }
+       }
+       if (sflg) {
+               int fdlisten = socket(cfg_family, SOCK_STREAM, 0);
+
+               if (fdlisten == -1) {
+                       perror("socket");
+                       exit(1);
+               }
+               apply_rcvsnd_buf(fdlisten);
+               setsockopt(fdlisten, SOL_SOCKET, SO_REUSEADDR, &on, sizeof(on));
+
+               setup_sockaddr(cfg_family, host, &listenaddr);
+
+               if (mss &&
+                   setsockopt(fdlisten, SOL_TCP, TCP_MAXSEG, &mss, sizeof(mss)) == -1) {
+                       perror("setsockopt TCP_MAXSEG");
+                       exit(1);
+               }
+               if (bind(fdlisten, (const struct sockaddr *)&listenaddr, cfg_alen) == -1) {
+                       perror("bind");
+                       exit(1);
+               }
+               if (listen(fdlisten, 128) == -1) {
+                       perror("listen");
+                       exit(1);
+               }
+               do_accept(fdlisten);
+       }
+       buffer = mmap(NULL, chunk_size, PROT_READ | PROT_WRITE,
+                             MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+       if (buffer == (char *)-1) {
+               perror("mmap");
+               exit(1);
+       }
+
+       fd = socket(AF_INET6, SOCK_STREAM, 0);
+       if (fd == -1) {
+               perror("socket");
+               exit(1);
+       }
+       apply_rcvsnd_buf(fd);
+
+       setup_sockaddr(cfg_family, host, &addr);
+
+       if (mss &&
+           setsockopt(fd, SOL_TCP, TCP_MAXSEG, &mss, sizeof(mss)) == -1) {
+               perror("setsockopt TCP_MAXSEG");
+               exit(1);
+       }
+       if (connect(fd, (const struct sockaddr *)&addr, cfg_alen) == -1) {
+               perror("connect");
+               exit(1);
+       }
+       if (max_pacing_rate &&
+           setsockopt(fd, SOL_SOCKET, SO_MAX_PACING_RATE,
+                      &max_pacing_rate, sizeof(max_pacing_rate)) == -1)
+               perror("setsockopt SO_MAX_PACING_RATE");
+
+       if (zflg && setsockopt(fd, SOL_SOCKET, SO_ZEROCOPY,
+                              &on, sizeof(on)) == -1) {
+               perror("setsockopt SO_ZEROCOPY, (-z option disabled)");
+               zflg = 0;
+       }
+       while (total < FILE_SZ) {
+               long wr = FILE_SZ - total;
+
+               if (wr > chunk_size)
+                       wr = chunk_size;
+               /* Note : we just want to fill the pipe with 0 bytes */
+               wr = send(fd, buffer, wr, zflg ? MSG_ZEROCOPY : 0);
+               if (wr <= 0)
+                       break;
+               total += wr;
+       }
+       close(fd);
+       munmap(buffer, chunk_size);
+       return 0;
+}