stats.c

Go to the documentation of this file.

/*
 * Copyright (c) 2015 Cisco and/or its affiliates.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at:
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */
#include <vpp/stats/stats.h>
#include <signal.h>
#include <vlib/threads.h>
#include <vnet/fib/fib_entry.h>
#include <vnet/fib/fib_table.h>
#include <vnet/dpo/load_balance.h>

#define STATS_DEBUG 0

stats_main_t stats_main;

#include <vnet/ip/ip.h>

#include <vpp/api/vpe_msg_enum.h>

#define f64_endian(a)
#define f64_print(a,b)

#define vl_typedefs             /* define message structures */
#include <vpp/api/vpe_all_api_h.h>
#undef vl_typedefs

#define vl_endianfun            /* define message structures */
#include <vpp/api/vpe_all_api_h.h>
#undef vl_endianfun

/* instantiate all the print functions we know about */
#define vl_print(handle, ...) vlib_cli_output (handle, __VA_ARGS__)
#define vl_printfun
#include <vpp/api/vpe_all_api_h.h>
#undef vl_printfun

#define foreach_stats_msg                               \
_(WANT_STATS, want_stats)                               \
_(VNET_INTERFACE_COUNTERS, vnet_interface_counters)     \
_(VNET_IP4_FIB_COUNTERS, vnet_ip4_fib_counters)         \
_(VNET_IP6_FIB_COUNTERS, vnet_ip6_fib_counters)         \
_(VNET_IP4_NBR_COUNTERS, vnet_ip4_nbr_counters)         \
_(VNET_IP6_NBR_COUNTERS, vnet_ip6_nbr_counters)

/* These constants ensure msg sizes <= 1024, aka ring allocation */
#define SIMPLE_COUNTER_BATCH_SIZE       126
#define COMBINED_COUNTER_BATCH_SIZE     63
#define IP4_FIB_COUNTER_BATCH_SIZE      48
#define IP6_FIB_COUNTER_BATCH_SIZE      30

/* 5ms */
#define STATS_RELEASE_DELAY_NS (1000 * 1000 * 5)
/*                              ns/us  us/ms        */

void
dslock (stats_main_t * sm, int release_hint, int tag)
{
  u32 thread_id;
  data_structure_lock_t *l = sm->data_structure_lock;

  if (PREDICT_FALSE (l == 0))
    return;

  thread_id = os_get_cpu_number ();
  if (l->lock && l->thread_id == thread_id)
    {
      l->count++;
      return;
    }

  if (release_hint)
    l->release_hint++;

  while (__sync_lock_test_and_set (&l->lock, 1))
    /* zzzz */ ;
  l->tag = tag;
  l->thread_id = thread_id;
  l->count = 1;
}

void
stats_dslock_with_hint (int hint, int tag)
{
  stats_main_t *sm = &stats_main;
  dslock (sm, hint, tag);
}

void
dsunlock (stats_main_t * sm)
{
  u32 thread_id;
  data_structure_lock_t *l = sm->data_structure_lock;

  if (PREDICT_FALSE (l == 0))
    return;

  thread_id = os_get_cpu_number ();
  ASSERT (l->lock && l->thread_id == thread_id);
  l->count--;
  if (l->count == 0)
    {
      l->tag = -l->tag;
      l->release_hint = 0;
      CLIB_MEMORY_BARRIER ();
      l->lock = 0;
    }
}

void
stats_dsunlock (int hint, int tag)
{
  stats_main_t *sm = &stats_main;
  dsunlock (sm);
}

static void
do_simple_interface_counters (stats_main_t * sm)
{
  vl_api_vnet_interface_counters_t *mp = 0;
  vnet_interface_main_t *im = sm->interface_main;
  api_main_t *am = sm->api_main;
  vl_shmem_hdr_t *shmem_hdr = am->shmem_hdr;
  unix_shared_memory_queue_t *q = shmem_hdr->vl_input_queue;
  vlib_simple_counter_main_t *cm;
  u32 items_this_message = 0;
  u64 v, *vp = 0;
  int i, n_counts;

  /*
   * Prevent interface registration from expanding / moving the vectors...
   * That tends never to happen, so we can hold this lock for a while.
   */
  vnet_interface_counter_lock (im);

  vec_foreach (cm, im->sw_if_counters)
  {
    n_counts = vlib_simple_counter_n_counters (cm);
    for (i = 0; i < n_counts; i++)
      {
        if (mp == 0)
          {
            items_this_message = clib_min (SIMPLE_COUNTER_BATCH_SIZE,
                                           n_counts - i);

            mp = vl_msg_api_alloc_as_if_client
              (sizeof (*mp) + items_this_message * sizeof (v));
            mp->_vl_msg_id = ntohs (VL_API_VNET_INTERFACE_COUNTERS);
            mp->vnet_counter_type = cm - im->sw_if_counters;
            mp->is_combined = 0;
            mp->first_sw_if_index = htonl (i);
            mp->count = 0;
            vp = (u64 *) mp->data;
          }
        v = vlib_get_simple_counter (cm, i);
        clib_mem_unaligned (vp, u64) = clib_host_to_net_u64 (v);
        vp++;
        mp->count++;
        if (mp->count == items_this_message)
          {
            mp->count = htonl (items_this_message);
            /* Send to the main thread... */
            vl_msg_api_send_shmem (q, (u8 *) & mp);
            mp = 0;
          }
      }
    ASSERT (mp == 0);
  }
  vnet_interface_counter_unlock (im);
}

static void
do_combined_interface_counters (stats_main_t * sm)
{
  vl_api_vnet_interface_counters_t *mp = 0;
  vnet_interface_main_t *im = sm->interface_main;
  api_main_t *am = sm->api_main;
  vl_shmem_hdr_t *shmem_hdr = am->shmem_hdr;
  unix_shared_memory_queue_t *q = shmem_hdr->vl_input_queue;
  vlib_combined_counter_main_t *cm;
  u32 items_this_message = 0;
  vlib_counter_t v, *vp = 0;
  int i, n_counts;

  vnet_interface_counter_lock (im);

  vec_foreach (cm, im->combined_sw_if_counters)
  {
    n_counts = vlib_combined_counter_n_counters (cm);
    for (i = 0; i < n_counts; i++)
      {
        if (mp == 0)
          {
            items_this_message = clib_min (COMBINED_COUNTER_BATCH_SIZE,
                                           n_counts - i);

            mp = vl_msg_api_alloc_as_if_client
              (sizeof (*mp) + items_this_message * sizeof (v));
            mp->_vl_msg_id = ntohs (VL_API_VNET_INTERFACE_COUNTERS);
            mp->vnet_counter_type = cm - im->combined_sw_if_counters;
            mp->is_combined = 1;
            mp->first_sw_if_index = htonl (i);
            mp->count = 0;
            vp = (vlib_counter_t *) mp->data;
          }
        vlib_get_combined_counter (cm, i, &v);
        clib_mem_unaligned (&vp->packets, u64)
          = clib_host_to_net_u64 (v.packets);
        clib_mem_unaligned (&vp->bytes, u64) = clib_host_to_net_u64 (v.bytes);
        vp++;
        mp->count++;
        if (mp->count == items_this_message)
          {
            mp->count = htonl (items_this_message);
            /* Send to the main thread... */
            vl_msg_api_send_shmem (q, (u8 *) & mp);
            mp = 0;
          }
      }
    ASSERT (mp == 0);
  }
  vnet_interface_counter_unlock (im);
}

/* from .../vnet/vnet/ip/lookup.c. Yuck */
typedef CLIB_PACKED (struct
                     {
                     ip4_address_t address;
u32 address_length: 6;
u32 index:           26;
                     }) ip4_route_t;

static void
ip46_fib_stats_delay (stats_main_t * sm, u32 sec, u32 nsec)
{
  struct timespec _req, *req = &_req;
  struct timespec _rem, *rem = &_rem;

  req->tv_sec = sec;
  req->tv_nsec = nsec;
  while (1)
    {
      if (nanosleep (req, rem) == 0)
        break;
      *req = *rem;
      if (errno == EINTR)
        continue;
      clib_unix_warning ("nanosleep");
      break;
    }
}

/**
 * @brief The context passed when collecting adjacency counters
 */
typedef struct ip4_nbr_stats_ctx_t_
{
  /**
   * The SW IF index all these adjs belong to
   */
  u32 sw_if_index;

  /**
   * A vector of ip4 nbr counters
   */
  vl_api_ip4_nbr_counter_t *counters;
} ip4_nbr_stats_ctx_t;

static adj_walk_rc_t
ip4_nbr_stats_cb (adj_index_t ai, void *arg)
{
  vl_api_ip4_nbr_counter_t *vl_counter;
  vlib_counter_t adj_counter;
  ip4_nbr_stats_ctx_t *ctx;
  ip_adjacency_t *adj;

  ctx = arg;
  vlib_get_combined_counter (&adjacency_counters, ai, &adj_counter);

  if (0 != adj_counter.packets)
    {
      vec_add2 (ctx->counters, vl_counter, 1);
      adj = adj_get (ai);

      vl_counter->packets = clib_host_to_net_u64 (adj_counter.packets);
      vl_counter->bytes = clib_host_to_net_u64 (adj_counter.bytes);
      vl_counter->address = adj->sub_type.nbr.next_hop.ip4.as_u32;
      vl_counter->link_type = adj->ia_link;
    }
  return (ADJ_WALK_RC_CONTINUE);
}

#define MIN(x,y) (((x)<(y))?(x):(y))

static void
ip4_nbr_ship (stats_main_t * sm, ip4_nbr_stats_ctx_t * ctx)
{
  api_main_t *am = sm->api_main;
  vl_shmem_hdr_t *shmem_hdr = am->shmem_hdr;
  unix_shared_memory_queue_t *q = shmem_hdr->vl_input_queue;
  vl_api_vnet_ip4_nbr_counters_t *mp = 0;
  int first = 0;

  /*
   * If the walk context has counters, which may be left over from the last
   * suspend, then we continue from there.
   */
  while (0 != vec_len (ctx->counters))
    {
      u32 n_items = MIN (vec_len (ctx->counters),
                         IP4_FIB_COUNTER_BATCH_SIZE);
      u8 pause = 0;

      dslock (sm, 0 /* release hint */ , 1 /* tag */ );

      mp = vl_msg_api_alloc_as_if_client (sizeof (*mp) +
                                          (n_items *
                                           sizeof
                                           (vl_api_ip4_nbr_counter_t)));
      mp->_vl_msg_id = ntohs (VL_API_VNET_IP4_NBR_COUNTERS);
      mp->count = ntohl (n_items);
      mp->sw_if_index = ntohl (ctx->sw_if_index);
      mp->begin = first;
      first = 0;

      /*
       * copy the counters from the back of the context, then we can easily
       * 'erase' them by resetting the vector length.
       * The order we push the stats to the caller is not important.
       */
      clib_memcpy (mp->c,
                   &ctx->counters[vec_len (ctx->counters) - n_items],
                   n_items * sizeof (*ctx->counters));

      _vec_len (ctx->counters) = vec_len (ctx->counters) - n_items;

      /*
       * send to the shm q
       */
      unix_shared_memory_queue_lock (q);
      pause = unix_shared_memory_queue_is_full (q);

      vl_msg_api_send_shmem_nolock (q, (u8 *) & mp);
      unix_shared_memory_queue_unlock (q);
      dsunlock (sm);

      if (pause)
        ip46_fib_stats_delay (sm, 0 /* sec */ ,
                              STATS_RELEASE_DELAY_NS);
    }
}

static void
do_ip4_nbrs (stats_main_t * sm)
{
  vnet_main_t *vnm = vnet_get_main ();
  vnet_interface_main_t *im = &vnm->interface_main;
  vnet_sw_interface_t *si;

  ip4_nbr_stats_ctx_t ctx = {
    .sw_if_index = 0,
    .counters = NULL,
  };

  /* *INDENT-OFF* */
  pool_foreach (si, im->sw_interfaces,
  ({
    /*
     * update the interface we are now concerned with
     */
    ctx.sw_if_index = si->sw_if_index;

    /*
     * we are about to walk another interface, so we shouldn't have any pending
     * stats to export.
     */
    ASSERT(ctx.counters == NULL);

    /*
     * visit each neighbour adjacency on the interface and collect
     * its current stats.
     * Because we hold the lock the walk is synchronous, so safe to routing
     * updates. It's limited in work by the number of adjacenies on an
     * interface, which is typically not huge.
     */
    dslock (sm, 0 /* release hint */ , 1 /* tag */ );
    adj_nbr_walk (si->sw_if_index,
                  FIB_PROTOCOL_IP4,
                  ip4_nbr_stats_cb,
                  &ctx);
    dsunlock (sm);

    /*
     * if this interface has some adjacencies with counters then ship them,
     * else continue to the next interface.
     */
    if (NULL != ctx.counters)
      {
        ip4_nbr_ship(sm, &ctx);
      }
  }));
  /* *INDENT-OFF* */
}

/**
 * @brief The context passed when collecting adjacency counters
 */
typedef struct ip6_nbr_stats_ctx_t_
{
  /**
   * The SW IF index all these adjs belong to
   */
  u32 sw_if_index;

  /**
   * A vector of ip6 nbr counters
   */
  vl_api_ip6_nbr_counter_t *counters;
} ip6_nbr_stats_ctx_t;

static adj_walk_rc_t
ip6_nbr_stats_cb (adj_index_t ai,
                  void *arg)
{
  vl_api_ip6_nbr_counter_t *vl_counter;
  vlib_counter_t adj_counter;
  ip6_nbr_stats_ctx_t *ctx;
  ip_adjacency_t *adj;

  ctx = arg;
  vlib_get_combined_counter(&adjacency_counters, ai, &adj_counter);

  if (0 != adj_counter.packets)
    {
      vec_add2(ctx->counters, vl_counter, 1);
      adj = adj_get(ai);

      vl_counter->packets = clib_host_to_net_u64(adj_counter.packets);
      vl_counter->bytes   = clib_host_to_net_u64(adj_counter.bytes);
      vl_counter->address[0] = adj->sub_type.nbr.next_hop.ip6.as_u64[0];
      vl_counter->address[1] = adj->sub_type.nbr.next_hop.ip6.as_u64[1];
      vl_counter->link_type = adj->ia_link;
    }
  return (ADJ_WALK_RC_CONTINUE);
}

#define MIN(x,y) (((x)<(y))?(x):(y))

static void
ip6_nbr_ship (stats_main_t * sm,
              ip6_nbr_stats_ctx_t *ctx)
{
  api_main_t *am = sm->api_main;
  vl_shmem_hdr_t *shmem_hdr = am->shmem_hdr;
  unix_shared_memory_queue_t *q = shmem_hdr->vl_input_queue;
  vl_api_vnet_ip6_nbr_counters_t *mp = 0;
  int first = 0;

  /*
   * If the walk context has counters, which may be left over from the last
   * suspend, then we continue from there.
   */
  while (0 != vec_len(ctx->counters))
    {
      u32 n_items = MIN (vec_len (ctx->counters),
                         IP6_FIB_COUNTER_BATCH_SIZE);
      u8 pause = 0;

      dslock (sm, 0 /* release hint */ , 1 /* tag */ );

      mp = vl_msg_api_alloc_as_if_client (sizeof (*mp) +
                                          (n_items *
                                           sizeof
                                           (vl_api_ip6_nbr_counter_t)));
      mp->_vl_msg_id = ntohs (VL_API_VNET_IP6_NBR_COUNTERS);
      mp->count = ntohl (n_items);
      mp->sw_if_index = ntohl (ctx->sw_if_index);
      mp->begin = first;
      first = 0;

      /*
       * copy the counters from the back of the context, then we can easily
       * 'erase' them by resetting the vector length.
       * The order we push the stats to the caller is not important.
       */
      clib_memcpy (mp->c,
                   &ctx->counters[vec_len (ctx->counters) - n_items],
                   n_items * sizeof (*ctx->counters));

      _vec_len (ctx->counters) = vec_len (ctx->counters) - n_items;

      /*
       * send to the shm q
       */
      unix_shared_memory_queue_lock (q);
      pause = unix_shared_memory_queue_is_full (q);

      vl_msg_api_send_shmem_nolock (q, (u8 *) & mp);
      unix_shared_memory_queue_unlock (q);
      dsunlock (sm);

      if (pause)
        ip46_fib_stats_delay (sm, 0 /* sec */ ,
                              STATS_RELEASE_DELAY_NS);
    }
}

static void
do_ip6_nbrs (stats_main_t * sm)
{
  vnet_main_t *vnm = vnet_get_main ();
  vnet_interface_main_t *im = &vnm->interface_main;
  vnet_sw_interface_t *si;

  ip6_nbr_stats_ctx_t ctx = {
    .sw_if_index = 0,
    .counters = NULL,
  };

  /* *INDENT-OFF* */
  pool_foreach (si, im->sw_interfaces,
  ({
    /*
     * update the interface we are now concerned with
     */
    ctx.sw_if_index = si->sw_if_index;

    /*
     * we are about to walk another interface, so we shouldn't have any pending
     * stats to export.
     */
    ASSERT(ctx.counters == NULL);

    /*
     * visit each neighbour adjacency on the interface and collect
     * its current stats.
     * Because we hold the lock the walk is synchronous, so safe to routing
     * updates. It's limited in work by the number of adjacenies on an
     * interface, which is typically not huge.
     */
    dslock (sm, 0 /* release hint */ , 1 /* tag */ );
    adj_nbr_walk (si->sw_if_index,
                  FIB_PROTOCOL_IP6,
                  ip6_nbr_stats_cb,
                  &ctx);
    dsunlock (sm);

    /*
     * if this interface has some adjacencies with counters then ship them,
     * else continue to the next interface.
     */
    if (NULL != ctx.counters)
      {
        ip6_nbr_ship(sm, &ctx);
      }
  }));
  /* *INDENT-OFF* */
}

static void
do_ip4_fibs (stats_main_t * sm)
{
  ip4_main_t *im4 = &ip4_main;
  api_main_t *am = sm->api_main;
  vl_shmem_hdr_t *shmem_hdr = am->shmem_hdr;
  unix_shared_memory_queue_t *q = shmem_hdr->vl_input_queue;
  static ip4_route_t *routes;
  ip4_route_t *r;
  fib_table_t *fib;
  ip_lookup_main_t *lm = &im4->lookup_main;
  static uword *results;
  vl_api_vnet_ip4_fib_counters_t *mp = 0;
  u32 items_this_message;
  vl_api_ip4_fib_counter_t *ctrp = 0;
  u32 start_at_fib_index = 0;
  int i;

again:
  /* *INDENT-OFF* */
  pool_foreach (fib, im4->fibs,
  ({
    /* We may have bailed out due to control-plane activity */
    while ((fib - im4->fibs) < start_at_fib_index)
      continue;

    if (mp == 0)
      {
        items_this_message = IP4_FIB_COUNTER_BATCH_SIZE;
        mp = vl_msg_api_alloc_as_if_client
          (sizeof (*mp) +
           items_this_message * sizeof (vl_api_ip4_fib_counter_t));
        mp->_vl_msg_id = ntohs (VL_API_VNET_IP4_FIB_COUNTERS);
        mp->count = 0;
        mp->vrf_id = ntohl (fib->ft_table_id);
        ctrp = (vl_api_ip4_fib_counter_t *) mp->c;
      }
    else
      {
        /* happens if the last FIB was empty... */
        ASSERT (mp->count == 0);
        mp->vrf_id = ntohl (fib->ft_table_id);
      }

    dslock (sm, 0 /* release hint */ , 1 /* tag */ );

    vec_reset_length (routes);
    vec_reset_length (results);

    for (i = 0; i < ARRAY_LEN (fib->v4.fib_entry_by_dst_address); i++)
      {
        uword *hash = fib->v4.fib_entry_by_dst_address[i];
        hash_pair_t *p;
        ip4_route_t x;

        x.address_length = i;

        hash_foreach_pair (p, hash,
        ({
          x.address.data_u32 = p->key;
          x.index = p->value[0];

          vec_add1 (routes, x);
          if (sm->data_structure_lock->release_hint)
            {
              start_at_fib_index = fib - im4->fibs;
              dsunlock (sm);
              ip46_fib_stats_delay (sm, 0 /* sec */,
                                    STATS_RELEASE_DELAY_NS);
              mp->count = 0;
              ctrp = (vl_api_ip4_fib_counter_t *)mp->c;
              goto again;
            }
        }));
      }

    vec_foreach (r, routes)
      {
        vlib_counter_t c;

        vlib_get_combined_counter (&load_balance_main.lbm_to_counters,
                                   r->index, &c);
        /*
         * If it has actually
         * seen at least one packet, send it.
         */
        if (c.packets > 0)
          {

            /* already in net byte order */
            ctrp->address = r->address.as_u32;
            ctrp->address_length = r->address_length;
            ctrp->packets = clib_host_to_net_u64 (c.packets);
            ctrp->bytes = clib_host_to_net_u64 (c.bytes);
            mp->count++;
            ctrp++;

            if (mp->count == items_this_message)
              {
                mp->count = htonl (items_this_message);
                /*
                 * If the main thread's input queue is stuffed,
                 * drop the data structure lock (which the main thread
                 * may want), and take a pause.
                 */
                unix_shared_memory_queue_lock (q);
                if (unix_shared_memory_queue_is_full (q))
                  {
                    dsunlock (sm);
                    vl_msg_api_send_shmem_nolock (q, (u8 *) & mp);
                    unix_shared_memory_queue_unlock (q);
                    mp = 0;
                    ip46_fib_stats_delay (sm, 0 /* sec */ ,
                                          STATS_RELEASE_DELAY_NS);
                    goto again;
                  }
                vl_msg_api_send_shmem_nolock (q, (u8 *) & mp);
                unix_shared_memory_queue_unlock (q);

                items_this_message = IP4_FIB_COUNTER_BATCH_SIZE;
                mp = vl_msg_api_alloc_as_if_client
                  (sizeof (*mp) +
                   items_this_message * sizeof (vl_api_ip4_fib_counter_t));
                mp->_vl_msg_id = ntohs (VL_API_VNET_IP4_FIB_COUNTERS);
                mp->count = 0;
                mp->vrf_id = ntohl (fib->ft_table_id);
                ctrp = (vl_api_ip4_fib_counter_t *) mp->c;
              }
          }                     /* for each (mp or single) adj */
        if (sm->data_structure_lock->release_hint)
          {
            start_at_fib_index = fib - im4->fibs;
            dsunlock (sm);
            ip46_fib_stats_delay (sm, 0 /* sec */ , STATS_RELEASE_DELAY_NS);
            mp->count = 0;
            ctrp = (vl_api_ip4_fib_counter_t *) mp->c;
            goto again;
          }
      }                         /* vec_foreach (routes) */

    dsunlock (sm);

    /* Flush any data from this fib */
    if (mp->count)
      {
        mp->count = htonl (mp->count);
        vl_msg_api_send_shmem (q, (u8 *) & mp);
        mp = 0;
      }
  }));
  /* *INDENT-ON* */

  /* If e.g. the last FIB had no reportable routes, free the buffer */
  if (mp)
    vl_msg_api_free (mp);
}

typedef struct
{
  ip6_address_t address;
  u32 address_length;
  u32 index;
} ip6_route_t;

typedef struct
{
  u32 fib_index;
  ip6_route_t **routep;
  stats_main_t *sm;
} add_routes_in_fib_arg_t;

static void
add_routes_in_fib (BVT (clib_bihash_kv) * kvp, void *arg)
{
  add_routes_in_fib_arg_t *ap = arg;
  stats_main_t *sm = ap->sm;

  if (sm->data_structure_lock->release_hint)
    clib_longjmp (&sm->jmp_buf, 1);

  if (kvp->key[2] >> 32 == ap->fib_index)
    {
      ip6_address_t *addr;
      ip6_route_t *r;
      addr = (ip6_address_t *) kvp;
      vec_add2 (*ap->routep, r, 1);
      r->address = addr[0];
      r->address_length = kvp->key[2] & 0xFF;
      r->index = kvp->value;
    }
}

static void
do_ip6_fibs (stats_main_t * sm)
{
  ip6_main_t *im6 = &ip6_main;
  api_main_t *am = sm->api_main;
  vl_shmem_hdr_t *shmem_hdr = am->shmem_hdr;
  unix_shared_memory_queue_t *q = shmem_hdr->vl_input_queue;
  static ip6_route_t *routes;
  ip6_route_t *r;
  fib_table_t *fib;
  static uword *results;
  vl_api_vnet_ip6_fib_counters_t *mp = 0;
  u32 items_this_message;
  vl_api_ip6_fib_counter_t *ctrp = 0;
  u32 start_at_fib_index = 0;
  BVT (clib_bihash) * h = &im6->ip6_table[IP6_FIB_TABLE_FWDING].ip6_hash;
  add_routes_in_fib_arg_t _a, *a = &_a;

again:
  /* *INDENT-OFF* */
  pool_foreach (fib, im6->fibs,
  ({
    /* We may have bailed out due to control-plane activity */
    while ((fib - im6->fibs) < start_at_fib_index)
      continue;

    if (mp == 0)
      {
        items_this_message = IP6_FIB_COUNTER_BATCH_SIZE;
        mp = vl_msg_api_alloc_as_if_client
          (sizeof (*mp) +
           items_this_message * sizeof (vl_api_ip6_fib_counter_t));
        mp->_vl_msg_id = ntohs (VL_API_VNET_IP6_FIB_COUNTERS);
        mp->count = 0;
        mp->vrf_id = ntohl (fib->ft_table_id);
        ctrp = (vl_api_ip6_fib_counter_t *) mp->c;
      }

    dslock (sm, 0 /* release hint */ , 1 /* tag */ );

    vec_reset_length (routes);
    vec_reset_length (results);

    a->fib_index = fib - im6->fibs;
    a->routep = &routes;
    a->sm = sm;

    if (clib_setjmp (&sm->jmp_buf, 0) == 0)
      {
        start_at_fib_index = fib - im6->fibs;
        BV (clib_bihash_foreach_key_value_pair) (h, add_routes_in_fib, a);
      }
    else
      {
        dsunlock (sm);
        ip46_fib_stats_delay (sm, 0 /* sec */ ,
                              STATS_RELEASE_DELAY_NS);
        mp->count = 0;
        ctrp = (vl_api_ip6_fib_counter_t *) mp->c;
        goto again;
      }

    vec_foreach (r, routes)
    {
        vlib_counter_t c;

        vlib_get_combined_counter (&load_balance_main.lbm_to_counters,
                                   r->index, &c);
        /*
         * If it has actually
         * seen at least one packet, send it.
         */
        if (c.packets > 0)
          {
            /* already in net byte order */
            ctrp->address[0] = r->address.as_u64[0];
            ctrp->address[1] = r->address.as_u64[1];
            ctrp->address_length = (u8) r->address_length;
            ctrp->packets = clib_host_to_net_u64 (c.packets);
            ctrp->bytes = clib_host_to_net_u64 (c.bytes);
            mp->count++;
            ctrp++;

            if (mp->count == items_this_message)
              {
                mp->count = htonl (items_this_message);
                /*
                 * If the main thread's input queue is stuffed,
                 * drop the data structure lock (which the main thread
                 * may want), and take a pause.
                 */
                unix_shared_memory_queue_lock (q);
                if (unix_shared_memory_queue_is_full (q))
                  {
                    dsunlock (sm);
                    vl_msg_api_send_shmem_nolock (q, (u8 *) & mp);
                    unix_shared_memory_queue_unlock (q);
                    mp = 0;
                    ip46_fib_stats_delay (sm, 0 /* sec */ ,
                                          STATS_RELEASE_DELAY_NS);
                    goto again;
                  }
                vl_msg_api_send_shmem_nolock (q, (u8 *) & mp);
                unix_shared_memory_queue_unlock (q);

                items_this_message = IP6_FIB_COUNTER_BATCH_SIZE;
                mp = vl_msg_api_alloc_as_if_client
                  (sizeof (*mp) +
                   items_this_message * sizeof (vl_api_ip6_fib_counter_t));
                mp->_vl_msg_id = ntohs (VL_API_VNET_IP6_FIB_COUNTERS);
                mp->count = 0;
                mp->vrf_id = ntohl (fib->ft_table_id);
                ctrp = (vl_api_ip6_fib_counter_t *) mp->c;
              }
          }

        if (sm->data_structure_lock->release_hint)
          {
            start_at_fib_index = fib - im6->fibs;
            dsunlock (sm);
            ip46_fib_stats_delay (sm, 0 /* sec */ , STATS_RELEASE_DELAY_NS);
            mp->count = 0;
            ctrp = (vl_api_ip6_fib_counter_t *) mp->c;
            goto again;
          }
    }                           /* vec_foreach (routes) */

    dsunlock (sm);

    /* Flush any data from this fib */
    if (mp->count)
      {
        mp->count = htonl (mp->count);
        vl_msg_api_send_shmem (q, (u8 *) & mp);
        mp = 0;
      }
  }));
  /* *INDENT-ON* */

  /* If e.g. the last FIB had no reportable routes, free the buffer */
  if (mp)
    vl_msg_api_free (mp);
}

static void
stats_thread_fn (void *arg)
{
  stats_main_t *sm = &stats_main;
  vlib_worker_thread_t *w = (vlib_worker_thread_t *) arg;
  vlib_thread_main_t *tm = vlib_get_thread_main ();

  /* stats thread wants no signals. */
  {
    sigset_t s;
    sigfillset (&s);
    pthread_sigmask (SIG_SETMASK, &s, 0);
  }

  if (vec_len (tm->thread_prefix))
    vlib_set_thread_name ((char *)
                          format (0, "%v_stats%c", tm->thread_prefix, '\0'));

  clib_mem_set_heap (w->thread_mheap);

  while (1)
    {
      /* 10 second poll interval */
      ip46_fib_stats_delay (sm, 10 /* secs */ , 0 /* nsec */ );

      if (!(sm->enable_poller))
        continue;
      do_simple_interface_counters (sm);
      do_combined_interface_counters (sm);
      do_ip4_fibs (sm);
      do_ip6_fibs (sm);
      do_ip4_nbrs (sm);
      do_ip6_nbrs (sm);
    }
}

static void
vl_api_vnet_interface_counters_t_handler (vl_api_vnet_interface_counters_t *
                                          mp)
{
  vpe_client_registration_t *reg;
  stats_main_t *sm = &stats_main;
  unix_shared_memory_queue_t *q, *q_prev = NULL;
  vl_api_vnet_interface_counters_t *mp_copy = NULL;
  u32 mp_size;

#if STATS_DEBUG > 0
  char *counter_name;
  u32 count, sw_if_index;
  int i;
#endif

  mp_size = sizeof (*mp) + (ntohl (mp->count) *
                            (mp->is_combined ? sizeof (vlib_counter_t) :
                             sizeof (u64)));

  /* *INDENT-OFF* */
  pool_foreach(reg, sm->stats_registrations,
  ({
    q = vl_api_client_index_to_input_queue (reg->client_index);
    if (q)
      {
        if (q_prev && (q_prev->cursize < q_prev->maxsize))
          {
            mp_copy = vl_msg_api_alloc_as_if_client(mp_size);
            clib_memcpy(mp_copy, mp, mp_size);
            vl_msg_api_send_shmem (q_prev, (u8 *)&mp);
            mp = mp_copy;
          }
        q_prev = q;
      }
  }));
  /* *INDENT-ON* */

#if STATS_DEBUG > 0
  count = ntohl (mp->count);
  sw_if_index = ntohl (mp->first_sw_if_index);
  if (mp->is_combined == 0)
    {
      u64 *vp, v;
      vp = (u64 *) mp->data;

      switch (mp->vnet_counter_type)
        {
        case VNET_INTERFACE_COUNTER_DROP:
          counter_name = "drop";
          break;
        case VNET_INTERFACE_COUNTER_PUNT:
          counter_name = "punt";
          break;
        case VNET_INTERFACE_COUNTER_IP4:
          counter_name = "ip4";
          break;
        case VNET_INTERFACE_COUNTER_IP6:
          counter_name = "ip6";
          break;
        case VNET_INTERFACE_COUNTER_RX_NO_BUF:
          counter_name = "rx-no-buff";
          break;
        case VNET_INTERFACE_COUNTER_RX_MISS:
          , counter_name = "rx-miss";
          break;
        case VNET_INTERFACE_COUNTER_RX_ERROR:
          , counter_name = "rx-error (fifo-full)";
          break;
        case VNET_INTERFACE_COUNTER_TX_ERROR:
          , counter_name = "tx-error (fifo-full)";
          break;
        default:
          counter_name = "bogus";
          break;
        }
      for (i = 0; i < count; i++)
        {
          v = clib_mem_unaligned (vp, u64);
          v = clib_net_to_host_u64 (v);
          vp++;
          fformat (stdout, "%U.%s %lld\n", format_vnet_sw_if_index_name,
                   sm->vnet_main, sw_if_index, counter_name, v);
          sw_if_index++;
        }
    }
  else
    {
      vlib_counter_t *vp;
      u64 packets, bytes;
      vp = (vlib_counter_t *) mp->data;

      switch (mp->vnet_counter_type)
        {
        case VNET_INTERFACE_COUNTER_RX:
          counter_name = "rx";
          break;
        case VNET_INTERFACE_COUNTER_TX:
          counter_name = "tx";
          break;
        default:
          counter_name = "bogus";
          break;
        }
      for (i = 0; i < count; i++)
        {
          packets = clib_mem_unaligned (&vp->packets, u64);
          packets = clib_net_to_host_u64 (packets);
          bytes = clib_mem_unaligned (&vp->bytes, u64);
          bytes = clib_net_to_host_u64 (bytes);
          vp++;
          fformat (stdout, "%U.%s.packets %lld\n",
                   format_vnet_sw_if_index_name,
                   sm->vnet_main, sw_if_index, counter_name, packets);
          fformat (stdout, "%U.%s.bytes %lld\n",
                   format_vnet_sw_if_index_name,
                   sm->vnet_main, sw_if_index, counter_name, bytes);
          sw_if_index++;
        }
    }
#endif
  if (q_prev && (q_prev->cursize < q_prev->maxsize))
    {
      vl_msg_api_send_shmem (q_prev, (u8 *) & mp);
    }
  else
    {
      vl_msg_api_free (mp);
    }
}

static void
vl_api_vnet_ip4_fib_counters_t_handler (vl_api_vnet_ip4_fib_counters_t * mp)
{
  vpe_client_registration_t *reg;
  stats_main_t *sm = &stats_main;
  unix_shared_memory_queue_t *q, *q_prev = NULL;
  vl_api_vnet_ip4_fib_counters_t *mp_copy = NULL;
  u32 mp_size;

  mp_size = sizeof (*mp_copy) +
    ntohl (mp->count) * sizeof (vl_api_ip4_fib_counter_t);

  /* *INDENT-OFF* */
  pool_foreach(reg, sm->stats_registrations,
  ({
    q = vl_api_client_index_to_input_queue (reg->client_index);
    if (q)
      {
        if (q_prev && (q_prev->cursize < q_prev->maxsize))
          {
            mp_copy = vl_msg_api_alloc_as_if_client(mp_size);
            clib_memcpy(mp_copy, mp, mp_size);
            vl_msg_api_send_shmem (q_prev, (u8 *)&mp);
            mp = mp_copy;
          }
        q_prev = q;
      }
  }));
  /* *INDENT-ON* */
  if (q_prev && (q_prev->cursize < q_prev->maxsize))
    {
      vl_msg_api_send_shmem (q_prev, (u8 *) & mp);
    }
  else
    {
      vl_msg_api_free (mp);
    }
}

static void
vl_api_vnet_ip4_nbr_counters_t_handler (vl_api_vnet_ip4_nbr_counters_t * mp)
{
  vpe_client_registration_t *reg;
  stats_main_t *sm = &stats_main;
  unix_shared_memory_queue_t *q, *q_prev = NULL;
  vl_api_vnet_ip4_nbr_counters_t *mp_copy = NULL;
  u32 mp_size;

  mp_size = sizeof (*mp_copy) +
    ntohl (mp->count) * sizeof (vl_api_ip4_nbr_counter_t);

  /* *INDENT-OFF* */
  pool_foreach(reg, sm->stats_registrations,
  ({
    q = vl_api_client_index_to_input_queue (reg->client_index);
    if (q)
      {
        if (q_prev && (q_prev->cursize < q_prev->maxsize))
          {
            mp_copy = vl_msg_api_alloc_as_if_client(mp_size);
            clib_memcpy(mp_copy, mp, mp_size);
            vl_msg_api_send_shmem (q_prev, (u8 *)&mp);
            mp = mp_copy;
          }
        q_prev = q;
      }
  }));
  /* *INDENT-ON* */
  if (q_prev && (q_prev->cursize < q_prev->maxsize))
    {
      vl_msg_api_send_shmem (q_prev, (u8 *) & mp);
    }
  else
    {
      vl_msg_api_free (mp);
    }
}

static void
vl_api_vnet_ip6_fib_counters_t_handler (vl_api_vnet_ip6_fib_counters_t * mp)
{
  vpe_client_registration_t *reg;
  stats_main_t *sm = &stats_main;
  unix_shared_memory_queue_t *q, *q_prev = NULL;
  vl_api_vnet_ip6_fib_counters_t *mp_copy = NULL;
  u32 mp_size;

  mp_size = sizeof (*mp_copy) +
    ntohl (mp->count) * sizeof (vl_api_ip6_fib_counter_t);

  /* *INDENT-OFF* */
  pool_foreach(reg, sm->stats_registrations,
  ({
    q = vl_api_client_index_to_input_queue (reg->client_index);
    if (q)
      {
        if (q_prev && (q_prev->cursize < q_prev->maxsize))
          {
            mp_copy = vl_msg_api_alloc_as_if_client(mp_size);
            clib_memcpy(mp_copy, mp, mp_size);
            vl_msg_api_send_shmem (q_prev, (u8 *)&mp);
            mp = mp_copy;
          }
        q_prev = q;
      }
  }));
  /* *INDENT-ON* */
  if (q_prev && (q_prev->cursize < q_prev->maxsize))
    {
      vl_msg_api_send_shmem (q_prev, (u8 *) & mp);
    }
  else
    {
      vl_msg_api_free (mp);
    }
}

static void
vl_api_vnet_ip6_nbr_counters_t_handler (vl_api_vnet_ip6_nbr_counters_t * mp)
{
  vpe_client_registration_t *reg;
  stats_main_t *sm = &stats_main;
  unix_shared_memory_queue_t *q, *q_prev = NULL;
  vl_api_vnet_ip6_nbr_counters_t *mp_copy = NULL;
  u32 mp_size;

  mp_size = sizeof (*mp_copy) +
    ntohl (mp->count) * sizeof (vl_api_ip6_nbr_counter_t);

  /* *INDENT-OFF* */
  pool_foreach(reg, sm->stats_registrations,
  ({
    q = vl_api_client_index_to_input_queue (reg->client_index);
    if (q)
      {
        if (q_prev && (q_prev->cursize < q_prev->maxsize))
          {
            mp_copy = vl_msg_api_alloc_as_if_client(mp_size);
            clib_memcpy(mp_copy, mp, mp_size);
            vl_msg_api_send_shmem (q_prev, (u8 *)&mp);
            mp = mp_copy;
          }
        q_prev = q;
      }
  }));
  /* *INDENT-ON* */
  if (q_prev && (q_prev->cursize < q_prev->maxsize))
    {
      vl_msg_api_send_shmem (q_prev, (u8 *) & mp);
    }
  else
    {
      vl_msg_api_free (mp);
    }
}

static void
vl_api_want_stats_t_handler (vl_api_want_stats_t * mp)
{
  stats_main_t *sm = &stats_main;
  vpe_client_registration_t *rp;
  vl_api_want_stats_reply_t *rmp;
  uword *p;
  i32 retval = 0;
  unix_shared_memory_queue_t *q;

  p = hash_get (sm->stats_registration_hash, mp->client_index);
  if (p)
    {
      if (mp->enable_disable)
        {
          clib_warning ("pid %d: already enabled...", mp->pid);
          retval = -2;
          goto reply;
        }
      else
        {
          rp = pool_elt_at_index (sm->stats_registrations, p[0]);
          pool_put (sm->stats_registrations, rp);
          hash_unset (sm->stats_registration_hash, mp->client_index);
          goto reply;
        }
    }
  if (mp->enable_disable == 0)
    {
      clib_warning ("pid %d: already disabled...", mp->pid);
      retval = -3;
      goto reply;
    }
  pool_get (sm->stats_registrations, rp);
  rp->client_index = mp->client_index;
  rp->client_pid = mp->pid;
  hash_set (sm->stats_registration_hash, rp->client_index,
            rp - sm->stats_registrations);

reply:
  if (pool_elts (sm->stats_registrations))
    sm->enable_poller = 1;
  else
    sm->enable_poller = 0;

  q = vl_api_client_index_to_input_queue (mp->client_index);

  if (!q)
    return;

  rmp = vl_msg_api_alloc (sizeof (*rmp));
  rmp->_vl_msg_id = ntohs (VL_API_WANT_STATS_REPLY);
  rmp->context = mp->context;
  rmp->retval = retval;

  vl_msg_api_send_shmem (q, (u8 *) & rmp);
}

int
stats_memclnt_delete_callback (u32 client_index)
{
  vpe_client_registration_t *rp;
  stats_main_t *sm = &stats_main;
  uword *p;

  p = hash_get (sm->stats_registration_hash, client_index);
  if (p)
    {
      rp = pool_elt_at_index (sm->stats_registrations, p[0]);
      pool_put (sm->stats_registrations, rp);
      hash_unset (sm->stats_registration_hash, client_index);
    }

  return 0;
}

#define vl_api_vnet_ip4_fib_counters_t_endian vl_noop_handler
#define vl_api_vnet_ip4_fib_counters_t_print vl_noop_handler
#define vl_api_vnet_ip6_fib_counters_t_endian vl_noop_handler
#define vl_api_vnet_ip6_fib_counters_t_print vl_noop_handler
#define vl_api_vnet_ip4_nbr_counters_t_endian vl_noop_handler
#define vl_api_vnet_ip4_nbr_counters_t_print vl_noop_handler
#define vl_api_vnet_ip6_nbr_counters_t_endian vl_noop_handler
#define vl_api_vnet_ip6_nbr_counters_t_print vl_noop_handler

static clib_error_t *
stats_init (vlib_main_t * vm)
{
  stats_main_t *sm = &stats_main;
  api_main_t *am = &api_main;
  void *vlib_worker_thread_bootstrap_fn (void *arg);

  sm->vlib_main = vm;
  sm->vnet_main = vnet_get_main ();
  sm->interface_main = &vnet_get_main ()->interface_main;
  sm->api_main = am;
  sm->stats_poll_interval_in_seconds = 10;
  sm->data_structure_lock =
    clib_mem_alloc_aligned (sizeof (data_structure_lock_t),
                            CLIB_CACHE_LINE_BYTES);
  memset (sm->data_structure_lock, 0, sizeof (*sm->data_structure_lock));

#define _(N,n)                                                  \
    vl_msg_api_set_handlers(VL_API_##N, #n,                     \
                           vl_api_##n##_t_handler,              \
                           vl_noop_handler,                     \
                           vl_api_##n##_t_endian,               \
                           vl_api_##n##_t_print,                \
                           sizeof(vl_api_##n##_t), 0 /* do NOT trace! */);
  foreach_stats_msg;
#undef _

  /* tell the msg infra not to free these messages... */
  am->message_bounce[VL_API_VNET_INTERFACE_COUNTERS] = 1;
  am->message_bounce[VL_API_VNET_IP4_FIB_COUNTERS] = 1;
  am->message_bounce[VL_API_VNET_IP6_FIB_COUNTERS] = 1;
  am->message_bounce[VL_API_VNET_IP4_NBR_COUNTERS] = 1;
  am->message_bounce[VL_API_VNET_IP6_NBR_COUNTERS] = 1;

  return 0;
}

VLIB_INIT_FUNCTION (stats_init);

/* *INDENT-OFF* */
VLIB_REGISTER_THREAD (stats_thread_reg, static) = {
  .name = "stats",
  .function = stats_thread_fn,
  .fixed_count = 1,
  .count = 1,
  .no_data_structure_clone = 1,
  .use_pthreads = 1,
};
/* *INDENT-ON* */

/*
 * fd.io coding-style-patch-verification: ON
 *
 * Local Variables:
 * eval: (c-set-style "gnu")
 * End:
 */