net: fix psock_fanout selftest hash collision
authorWillem de Bruijn <willemb@google.com>
Tue, 19 Mar 2013 20:42:44 +0000 (20:42 +0000)
committerDavid S. Miller <davem@davemloft.net>
Wed, 20 Mar 2013 16:33:18 +0000 (12:33 -0400)
Fix flaky results with PACKET_FANOUT_HASH depending on whether the
two flows hash into the same packet socket or not.

Also adds tests for PACKET_FANOUT_LB and PACKET_FANOUT_CPU and
replaces the counting method with a packet ring.

Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tools/testing/selftests/net-afpacket/psock_fanout.c

index af9b0196b0bedc60d9c87c64caf19041e71c100a..f765f09931bda76f9906805153a3a2655572c51e 100644 (file)
  *   The test currently runs for
  *   - PACKET_FANOUT_HASH
  *   - PACKET_FANOUT_HASH with PACKET_FANOUT_FLAG_ROLLOVER
+ *   - PACKET_FANOUT_LB
+ *   - PACKET_FANOUT_CPU
  *   - PACKET_FANOUT_ROLLOVER
  *
  * Todo:
- * - datapath: PACKET_FANOUT_LB
- * - datapath: PACKET_FANOUT_CPU
  * - functionality: PACKET_FANOUT_FLAG_DEFRAG
  *
  * License (GPLv2):
  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  */
 
+#define _GNU_SOURCE            /* for sched_setaffinity */
+
 #include <arpa/inet.h>
 #include <errno.h>
+#include <fcntl.h>
 #include <linux/filter.h>
 #include <linux/if_packet.h>
 #include <net/ethernet.h>
 #include <netinet/ip.h>
 #include <netinet/udp.h>
-#include <fcntl.h>
+#include <poll.h>
+#include <sched.h>
 #include <stdint.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <sys/mman.h>
 #include <sys/socket.h>
 #include <sys/stat.h>
 #include <sys/types.h>
@@ -58,6 +63,8 @@
 
 #define DATA_LEN                       100
 #define DATA_CHAR                      'a'
+#define RING_NUM_FRAMES                        20
+#define PORT_BASE                      8000
 
 static void pair_udp_open(int fds[], uint16_t port)
 {
@@ -162,37 +169,55 @@ static int sock_fanout_open(uint16_t typeflags, int num_packets)
                return -1;
        }
 
-       val = sizeof(struct iphdr) + sizeof(struct udphdr) + DATA_LEN;
-       val *= num_packets;
-       /* hack: apparently, the above calculation is too small (TODO: fix) */
-       val *= 3;
-       if (setsockopt(fd, SOL_SOCKET, SO_RCVBUF, &val, sizeof(val))) {
-               perror("setsockopt SO_RCVBUF");
-               exit(1);
-       }
-
        sock_fanout_setfilter(fd);
        return fd;
 }
 
-static void sock_fanout_read(int fds[], const int expect[])
+static char *sock_fanout_open_ring(int fd)
 {
-       struct tpacket_stats stats;
-       socklen_t ssize;
-       int ret[2];
+       struct tpacket_req req = {
+               .tp_block_size = getpagesize(),
+               .tp_frame_size = getpagesize(),
+               .tp_block_nr   = RING_NUM_FRAMES,
+               .tp_frame_nr   = RING_NUM_FRAMES,
+       };
+       char *ring;
 
-       ssize = sizeof(stats);
-       if (getsockopt(fds[0], SOL_PACKET, PACKET_STATISTICS, &stats, &ssize)) {
-               perror("getsockopt statistics 0");
+       if (setsockopt(fd, SOL_PACKET, PACKET_RX_RING, (void *) &req,
+                      sizeof(req))) {
+               perror("packetsock ring setsockopt");
                exit(1);
        }
-       ret[0] = stats.tp_packets - stats.tp_drops;
-       ssize = sizeof(stats);
-       if (getsockopt(fds[1], SOL_PACKET, PACKET_STATISTICS, &stats, &ssize)) {
-               perror("getsockopt statistics 1");
+
+       ring = mmap(0, req.tp_block_size * req.tp_block_nr,
+                   PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+       if (!ring) {
+               fprintf(stderr, "packetsock ring mmap\n");
                exit(1);
        }
-       ret[1] = stats.tp_packets - stats.tp_drops;
+
+       return ring;
+}
+
+static int sock_fanout_read_ring(int fd, void *ring)
+{
+       struct tpacket_hdr *header = ring;
+       int count = 0;
+
+       while (header->tp_status & TP_STATUS_USER && count < RING_NUM_FRAMES) {
+               count++;
+               header = ring + (count * getpagesize());
+       }
+
+       return count;
+}
+
+static int sock_fanout_read(int fds[], char *rings[], const int expect[])
+{
+       int ret[2];
+
+       ret[0] = sock_fanout_read_ring(fds[0], rings[0]);
+       ret[1] = sock_fanout_read_ring(fds[1], rings[1]);
 
        fprintf(stderr, "info: count=%d,%d, expect=%d,%d\n",
                        ret[0], ret[1], expect[0], expect[1]);
@@ -200,8 +225,10 @@ static void sock_fanout_read(int fds[], const int expect[])
        if ((!(ret[0] == expect[0] && ret[1] == expect[1])) &&
            (!(ret[0] == expect[1] && ret[1] == expect[0]))) {
                fprintf(stderr, "ERROR: incorrect queue lengths\n");
-               exit(1);
+               return 1;
        }
+
+       return 0;
 }
 
 /* Test illegal mode + flag combination */
@@ -253,11 +280,12 @@ static void test_control_group(void)
        }
 }
 
-static void test_datapath(uint16_t typeflags,
-                         const int expect1[], const int expect2[])
+static int test_datapath(uint16_t typeflags, int port_off,
+                        const int expect1[], const int expect2[])
 {
        const int expect0[] = { 0, 0 };
-       int fds[2], fds_udp[2][2];
+       char *rings[2];
+       int fds[2], fds_udp[2][2], ret;
 
        fprintf(stderr, "test: datapath 0x%hx\n", typeflags);
 
@@ -267,41 +295,93 @@ static void test_datapath(uint16_t typeflags,
                fprintf(stderr, "ERROR: failed open\n");
                exit(1);
        }
-       pair_udp_open(fds_udp[0], 8000);
-       pair_udp_open(fds_udp[1], 8002);
-       sock_fanout_read(fds, expect0);
+       rings[0] = sock_fanout_open_ring(fds[0]);
+       rings[1] = sock_fanout_open_ring(fds[1]);
+       pair_udp_open(fds_udp[0], PORT_BASE);
+       pair_udp_open(fds_udp[1], PORT_BASE + port_off);
+       sock_fanout_read(fds, rings, expect0);
 
        /* Send data, but not enough to overflow a queue */
        pair_udp_send(fds_udp[0], 15);
        pair_udp_send(fds_udp[1], 5);
-       sock_fanout_read(fds, expect1);
+       ret = sock_fanout_read(fds, rings, expect1);
 
        /* Send more data, overflow the queue */
        pair_udp_send(fds_udp[0], 15);
        /* TODO: ensure consistent order between expect1 and expect2 */
-       sock_fanout_read(fds, expect2);
+       ret |= sock_fanout_read(fds, rings, expect2);
 
+       if (munmap(rings[1], RING_NUM_FRAMES * getpagesize()) ||
+           munmap(rings[0], RING_NUM_FRAMES * getpagesize())) {
+               fprintf(stderr, "close rings\n");
+               exit(1);
+       }
        if (close(fds_udp[1][1]) || close(fds_udp[1][0]) ||
            close(fds_udp[0][1]) || close(fds_udp[0][0]) ||
            close(fds[1]) || close(fds[0])) {
                fprintf(stderr, "close datapath\n");
                exit(1);
        }
+
+       return ret;
+}
+
+static int set_cpuaffinity(int cpuid)
+{
+       cpu_set_t mask;
+
+       CPU_ZERO(&mask);
+       CPU_SET(cpuid, &mask);
+       if (sched_setaffinity(0, sizeof(mask), &mask)) {
+               if (errno != EINVAL) {
+                       fprintf(stderr, "setaffinity %d\n", cpuid);
+                       exit(1);
+               }
+               return 1;
+       }
+
+       return 0;
 }
 
 int main(int argc, char **argv)
 {
-       const int expect_hash[2][2]     = { { 15, 5 }, { 5, 0 } };
-       const int expect_hash_rb[2][2]  = { { 15, 5 }, { 5, 10 } };
-       const int expect_rb[2][2]       = { { 20, 0 }, { 0, 15 } };
+       const int expect_hash[2][2]     = { { 15, 5 },  { 20, 5 } };
+       const int expect_hash_rb[2][2]  = { { 15, 5 },  { 20, 15 } };
+       const int expect_lb[2][2]       = { { 10, 10 }, { 18, 17 } };
+       const int expect_rb[2][2]       = { { 20, 0 },  { 20, 15 } };
+       const int expect_cpu0[2][2]     = { { 20, 0 },  { 20, 0 } };
+       const int expect_cpu1[2][2]     = { { 0, 20 },  { 0, 20 } };
+       int port_off = 2, tries = 5, ret;
 
        test_control_single();
        test_control_group();
 
-       test_datapath(PACKET_FANOUT_HASH, expect_hash[0], expect_hash[1]);
-       test_datapath(PACKET_FANOUT_HASH | PACKET_FANOUT_FLAG_ROLLOVER,
-                     expect_hash_rb[0], expect_hash_rb[1]);
-       test_datapath(PACKET_FANOUT_ROLLOVER, expect_rb[0], expect_rb[1]);
+       /* find a set of ports that do not collide onto the same socket */
+       ret = test_datapath(PACKET_FANOUT_HASH, port_off,
+                           expect_hash[0], expect_hash[1]);
+       while (ret && tries--) {
+               fprintf(stderr, "info: trying alternate ports (%d)\n", tries);
+               ret = test_datapath(PACKET_FANOUT_HASH, ++port_off,
+                                   expect_hash[0], expect_hash[1]);
+       }
+
+       ret |= test_datapath(PACKET_FANOUT_HASH | PACKET_FANOUT_FLAG_ROLLOVER,
+                            port_off, expect_hash_rb[0], expect_hash_rb[1]);
+       ret |= test_datapath(PACKET_FANOUT_LB,
+                            port_off, expect_lb[0], expect_lb[1]);
+       ret |= test_datapath(PACKET_FANOUT_ROLLOVER,
+                            port_off, expect_rb[0], expect_rb[1]);
+
+       set_cpuaffinity(0);
+       ret |= test_datapath(PACKET_FANOUT_CPU, port_off,
+                            expect_cpu0[0], expect_cpu0[1]);
+       if (!set_cpuaffinity(1))
+               /* TODO: test that choice alternates with previous */
+               ret |= test_datapath(PACKET_FANOUT_CPU, port_off,
+                                    expect_cpu1[0], expect_cpu1[1]);
+
+       if (ret)
+               return 1;
 
        printf("OK. All tests passed\n");
        return 0;