pping: Add perf-buffer

The XDP program pushes the calculated RTTs to userspace through the perf-buffer and the userspace program polls it to print them out Signed-off-by: Simon Sundberg <simon.sundberg@kau.se>
2024-05-06 15:54:53 +00:00 · 2021-01-07 18:14:27 +01:00
parent 8628004b6c
commit 45b138b6ab
4 changed files with 124 additions and 35 deletions
--- a/pping/Makefile
+++ b/pping/Makefile
@@ -3,6 +3,8 @@
 USER_TARGETS := pping_user
 BPF_TARGETS := pping_kern

+LDFLAGS = -pthread
+
 LIB_DIR = ../lib

 include $(LIB_DIR)/common.mk
--- a/pping/pping_kern.c
+++ b/pping/pping_kern.c
@@ -7,10 +7,11 @@
 #include <linux/ip.h>
 #include <linux/tcp.h>

+#include <string.h>
+
 #include "timestamp_map.h"

 #define MAX_TCP_OPTIONS 10
-#define BILLION 1000000000UL

 char _license[] SEC("license") = "GPL";

@@ -18,7 +19,13 @@ struct bpf_map_def SEC("maps") ts_start = {
  .type = BPF_MAP_TYPE_HASH,
  .key_size = sizeof(struct ts_key),
  .value_size = sizeof(struct ts_timestamp),
-  .max_entries = 4096,
+  .max_entries = 16384,
+};
+
+struct bpf_map_def SEC("maps") rtt_events = {
+  .type = BPF_MAP_TYPE_PERF_EVENT_ARRAY,
+  .key_size = sizeof(__u32), // CPU ID
+  .value_size = sizeof(__u32), // perf file descriptor?
 };

 static __always_inline int fill_ipv4_flow(struct ipv4_flow *flow, __u32 saddr, __u32 daddr, __u16 sport, __u16 dport)
@@ -31,6 +38,7 @@ static __always_inline int fill_ipv4_flow(struct ipv4_flow *flow, __u32 saddr, _
 }

 // Parses the TSval and TSecr values from the TCP options field - returns 0 if sucessful and -1 on failure
+// If sucessful the TSval and TSecr values will be stored at tsval and tsecr (in network byte order!)
 static __always_inline int parse_tcp_ts(struct tcphdr *tcph, void *data_end, __u32 *tsval, __u32 *tsecr)
 {
  if (tcph + 1 > data_end) // To hopefully please verifier
@@ -58,10 +66,10 @@ static __always_inline int parse_tcp_ts(struct tcphdr *tcph, void *data_end, __u
    opt_size = *(__u8 *)(pos+1); // Save value to variable so I don't have to perform any more data_end checks on option size

    if (opt == 8 && opt_size == 10) { // Option-kind is TCP timestap (yey!)
-      if (pos + opt_size > opt_end ||pos + opt_size > data_end)
+      if (pos + opt_size > opt_end || pos + opt_size > data_end)
 	return -1;
-      *tsval = bpf_ntohl(*(__u32 *)(pos + 2));
-      *tsecr = bpf_ntohl(*(__u32 *)(pos + 6));
+      *tsval = *(__u32 *)(pos + 2);
+      *tsecr = *(__u32 *)(pos + 6);
      return 0;
    }

@@ -99,7 +107,7 @@ int xdp_prog_ingress(struct xdp_md *ctx)
  if (parse_tcp_ts(tcph, data_end, &tsval, &tsecr) < 0) // No TCP timestamp
    return XDP_PASS;
  // We have a TCP-timestamp - now we can check if it's in the map
-  bpf_printk("TCP-packet with timestap. TSval: %u, TSecr: %u\n", tsval, tsecr);
+  bpf_printk("TCP-packet with timestap. TSval: %u, TSecr: %u\n", bpf_ntohl(tsval), bpf_ntohl(tsecr));
  struct ts_key key;
  fill_ipv4_flow(&(key.flow), iph->daddr, iph->saddr, tcph->dest, tcph->source); // Fill in reverse order of egress (dest <--> source)
  key.tsval = tsecr;
@@ -113,9 +121,13 @@ int xdp_prog_ingress(struct xdp_md *ctx)
  struct ts_timestamp *ts = bpf_map_lookup_elem(&ts_start, &key);
  if (ts && ts->used == 0) {
    ts->used = 1;
-    __u64 rtt = bpf_ktime_get_ns() - ts->timestamp;
-    // TODO: Push RTT + flow to userspace through perf buffer
-    bpf_printk("RTT: %llu\n", rtt);
+    //__u64 rtt = bpf_ktime_get_ns() - ts->timestamp;
+
+    struct rtt_event event = {0};
+    memcpy(&(event.flow), &(key.flow), sizeof(struct ipv4_flow));
+    event.rtt = bpf_ktime_get_ns() - ts->timestamp;
+    bpf_perf_event_output(ctx, &rtt_events, BPF_F_CURRENT_CPU, &event, sizeof(event));
+    bpf_printk("Pushed rtt event with RTT: %llu", event.rtt);
  }
  
  return XDP_PASS;
--- a/pping/pping_user.c
+++ b/pping/pping_user.c
@@ -2,9 +2,9 @@
 #include <linux/if_link.h>
 #include <bpf/bpf.h>
 #include <bpf/libbpf.h>
-#include <net/if.h>
-#include <linux/err.h> // For IS_ERR_OR_NULL macro
-#include <arpa/inet.h>
+#include <net/if.h> // For if_nametoindex
+//#include <linux/err.h> // For IS_ERR_OR_NULL macro // use libbpf_get_error instead
+#include <arpa/inet.h> // For inet_ntoa and ntohs

 #include <stdio.h>
 #include <stdlib.h>
@@ -15,17 +15,27 @@
 #include <signal.h> // For detecting Ctrl-C
 #include <sys/resource.h> // For setting rlmit
 #include <time.h>
+#include <pthread.h>
 #include "timestamp_map.h" //key and value structs for the ts_start map

+#define BILLION 1000000000UL
 #define PPING_ELF_OBJ "pping_kern.o"
 #define XDP_PROG_SEC "pping_ingress"
 #define XDP_FLAGS XDP_FLAGS_UPDATE_IF_NOEXIST
 #define MAP_NAME "ts_start"
+#define MAP_CLEANUP_INTERVAL 1*BILLION // Clean timestamp map once per second 
+#define PERF_BUFFER_NAME "rtt_events"
+#define PERF_BUFFER_PAGES 64 // Related to the perf-buffer size?
+#define PERF_POLL_TIMEOUT_MS 100
 #define RMEMLIM 512UL << 20 /* 512 MBs */
 #define ERROR_MSG_MAX 1024
-#define BILLION 1000000000UL
 #define TIMESTAMP_LIFETIME 10*BILLION // 10 seconds

+struct map_cleanup_args {
+  int map_fd;
+  __u64 max_age_ns;
+};
+
 /* static int libbpf_print_fn(enum libbpf_print_level level, const char *format, va_list args) */
 /* { */
 /* 	return vfprintf(stderr, format, args); */
@@ -33,7 +43,7 @@

 static volatile int keep_running = 1;

-void abort_main_loop(int sig)
+void abort_program(int sig)
 {
  keep_running = 0;
 }
@@ -60,7 +70,6 @@ static int xdp_load_and_attach(int ifindex, char *obj_path, char *sec, __u32 xdp
    //.ifindex = ifindex,
    .file = obj_path,
  };
-  //attr.file = obj_path;

  err = bpf_prog_load_xattr(&attr, obj, prog_fd);
  if (err) {
@@ -95,7 +104,6 @@ static __u64 get_time_ns(clockid_t clockid)
  return (__u64)t.tv_sec * BILLION + (__u64)t.tv_nsec;
 }

-
 static int remove_old_entries_from_map(int map_fd, __u64 max_age)
 {
  int removed = 0, entries = 0;
@@ -103,7 +111,7 @@ static int remove_old_entries_from_map(int map_fd, __u64 max_age)
  struct ts_timestamp value;
  bool delete_prev = false;
  __u64 now_nsec = get_time_ns(CLOCK_MONOTONIC);
-  if (now_nsec == 0) 
+  if (now_nsec == 0)
    return -errno;

  // Cannot delete current key because then loop will reset, see https://www.bouncybouncy.net/blog/bpf_map_get_next_key-pitfalls/
@@ -126,10 +134,40 @@ static int remove_old_entries_from_map(int map_fd, __u64 max_age)
    bpf_map_delete_elem(map_fd, &prev_key);
    removed++;
  }
-  printf("Gone through %d entries and removed %d of them\n", entries, removed);
+  __u64 duration = get_time_ns(CLOCK_MONOTONIC) - now_nsec;
+  printf("Gone through %d entries and removed %d of them in %llu.%09llu\n", entries, removed, duration / BILLION, duration % BILLION);
  return removed;
 }

+static void *periodic_map_cleanup(void *args)
+{
+  struct map_cleanup_args *argp = args;
+  struct timespec interval;
+  interval.tv_sec = MAP_CLEANUP_INTERVAL / BILLION;
+  interval.tv_nsec = MAP_CLEANUP_INTERVAL % BILLION;
+  while (keep_running) {
+    remove_old_entries_from_map(argp->map_fd, argp->max_age_ns);
+    nanosleep(&interval, NULL);
+  }
+  pthread_exit(NULL);
+}
+
+static void handle_rtt_event(void *ctx, int cpu, void *data, __u32 data_size)
+{
+  const struct rtt_event *e = data;
+  struct in_addr saddr, daddr;
+  saddr.s_addr = e->flow.saddr;
+  daddr.s_addr = e->flow.daddr;
+  printf("%llu.%09llu ms %s:%d+%s:%d\n", e->rtt / BILLION, e->rtt % BILLION,
+	 inet_ntoa(saddr), ntohs(e->flow.sport),
+	 inet_ntoa(daddr), ntohs(e->flow.dport));	 
+}
+
+static void handle_missed_rtt_event(void *ctx, int cpu, __u64 lost_cnt)
+{
+  fprintf(stderr, "Lost %llu RTT events on CPU %d\n", lost_cnt, cpu);
+}
+
 int main(int argc, char *argv[])
 {
  if (argc < 2) {
@@ -137,8 +175,10 @@ int main(int argc, char *argv[])
    return EXIT_FAILURE;
  }

-  int err;
+  int err = 0, ifindex = 0;
+  bool xdp_attached = false;
  char error_msg[ERROR_MSG_MAX];
+  struct perf_buffer *pb = NULL;

  // Setup libbpf errors and debug info on callback
  //libbpf_set_print(libbpf_print_fn);
@@ -147,15 +187,15 @@ int main(int argc, char *argv[])
  err = set_rlimit(RMEMLIM);
  if (err) {
    fprintf(stderr, "Could not set rlimit to %ld bytes: %s\n", RMEMLIM, strerror(-err));
-    return EXIT_FAILURE;
+    goto cleanup;
  }

  // Get index of interface
-  int ifindex = if_nametoindex(argv[1]);
+  ifindex = if_nametoindex(argv[1]);
  if (ifindex == 0) {
    err = -errno;
    fprintf(stderr, "Could not get index of interface %s: %s\n", argv[1], strerror(-err));
-    return EXIT_FAILURE;
+    goto cleanup;
  }
  
  // Load and attach XDP program to interface
@@ -165,29 +205,57 @@ int main(int argc, char *argv[])
  err = xdp_load_and_attach(ifindex, PPING_ELF_OBJ, XDP_PROG_SEC, XDP_FLAGS, &obj, &prog_fd, error_msg);
  if (err) {
    fprintf(stderr, "%s: %s\n", error_msg, strerror(-err));
-    return EXIT_FAILURE;
+    goto cleanup;
  }
+  xdp_attached = true;
+
+  // Find map fd (to perform periodic cleanup)
  int map_fd = bpf_object__find_map_fd_by_name(obj, MAP_NAME);
  if (map_fd < 0) {
    fprintf(stderr, "Failed finding map %s in %s: %s\n", MAP_NAME, PPING_ELF_OBJ, strerror(-map_fd));
-    xdp_deatach(ifindex, XDP_FLAGS);
-    return EXIT_FAILURE;
+    goto cleanup;
+  }
+  pthread_t tid;
+  struct map_cleanup_args args = {.map_fd = map_fd, .max_age_ns = TIMESTAMP_LIFETIME};
+  err = pthread_create(&tid, NULL, periodic_map_cleanup, &args);
+  if (err) {
+    fprintf(stderr, "Failed starting thread to perform periodic map cleanup: %s\n", strerror(err));
+    goto cleanup;
+  }
+
+  // Set up perf buffer
+  struct perf_buffer_opts pb_opts;
+  pb_opts.sample_cb = handle_rtt_event;
+  pb_opts.lost_cb = handle_missed_rtt_event;
+  
+  pb = perf_buffer__new(bpf_object__find_map_fd_by_name(obj, PERF_BUFFER_NAME), PERF_BUFFER_PAGES, &pb_opts);
+  err = libbpf_get_error(pb);
+  if (err) {
+    pb = NULL;
+    fprintf(stderr, "Failed to open perf buffer %s: %s\n", PERF_BUFFER_NAME, strerror(err));
+    goto cleanup;
  }

  // Main loop
-  signal(SIGINT, abort_main_loop);
+  signal(SIGINT, abort_program);
  while(keep_running) {
-    sleep(1);
-    // TODO - print out 
-    remove_old_entries_from_map(map_fd, TIMESTAMP_LIFETIME);
+    if ((err = perf_buffer__poll(pb, PERF_POLL_TIMEOUT_MS)) < 0) {
+      if (keep_running) // Only print polling error if it wasn't caused by program termination
+	fprintf(stderr, "Error polling perf buffer: %s\n", strerror(-err));
+      break;
+    }
  }
-
-  err = xdp_deatach(ifindex, XDP_FLAGS);
-  if (err) {
-    fprintf(stderr, "Failed deatching program from ifindex %d: %s\n", ifindex, strerror(-err));
-    return EXIT_FAILURE;
+    
+ cleanup:
+  printf("Cleanup!\n");
+  perf_buffer__free(pb);
+  if (xdp_attached) {
+    err = xdp_deatach(ifindex, XDP_FLAGS);
+    if (err) {
+      fprintf(stderr, "Failed deatching program from ifindex %d: %s\n", ifindex, strerror(-err));
+    }
  }
  
-  return EXIT_SUCCESS;
+  return err != 0;
 }

--- a/pping/timestamp_map.h
+++ b/pping/timestamp_map.h
@@ -23,4 +23,11 @@ struct ts_timestamp
  // __u8 pad[7]; // Need to pad it due to compiler optimization, see "Remove struct padding with aligning members by using #pragma pack." at https://docs.cilium.io/en/v1.9/bpf/

 };
+
+struct rtt_event
+{
+  struct ipv4_flow flow;
+  __u64 rtt;
+};
+
 #endif