src/vlibmemory/memory_shared.c

/*
 *------------------------------------------------------------------
 * memclnt_shared.c - API message handling, common code for both clients
 * and the vlib process itself.
 *
 *
 * Copyright (c) 2009 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 <stdio.h>
#include <stdlib.h>
#include <stddef.h>
#include <string.h>
#include <unistd.h>
#include <signal.h>

#include <vppinfra/format.h>
#include <vppinfra/byte_order.h>
#include <vppinfra/error.h>
#include <svm/queue.h>
#include <vlib/vlib.h>
#include <vlib/unix/unix.h>
#include <vlibmemory/memory_api.h>
#include <vlibmemory/vl_memory_msg_enum.h>

#define vl_typedefs
#include <vlibmemory/vl_memory_api_h.h>
#undef vl_typedefs

#define DEBUG_MESSAGE_BUFFER_OVERRUN 0

static inline void *
vl_msg_api_alloc_internal (int nbytes, int pool, int may_return_null)
{
  int i;
  msgbuf_t *rv;
  ring_alloc_t *ap;
  svm_queue_t *q;
  void *oldheap;
  vl_shmem_hdr_t *shmem_hdr;
  api_main_t *am = &api_main;

  shmem_hdr = am->shmem_hdr;

#if DEBUG_MESSAGE_BUFFER_OVERRUN > 0
  nbytes += 4;
#endif

  ASSERT (pool == 0 || vlib_get_thread_index () == 0);

  if (shmem_hdr == 0)
    {
      clib_warning ("shared memory header NULL");
      return 0;
    }

  /* account for the msgbuf_t header */
  nbytes += sizeof (msgbuf_t);

  if (shmem_hdr->vl_rings == 0)
    {
      clib_warning ("vl_rings NULL");
      ASSERT (0);
      abort ();
    }

  if (shmem_hdr->client_rings == 0)
    {
      clib_warning ("client_rings NULL");
      ASSERT (0);
      abort ();
    }

  ap = pool ? shmem_hdr->vl_rings : shmem_hdr->client_rings;
  for (i = 0; i < vec_len (ap); i++)
    {
      /* Too big? */
      if (nbytes > ap[i].size)
	{
	  continue;
	}

      q = ap[i].rp;
      if (pool == 0)
	{
	  pthread_mutex_lock (&q->mutex);
	}
      rv = (msgbuf_t *) (&q->data[0] + q->head * q->elsize);
      /*
       * Is this item still in use?
       */
      if (rv->q)
	{
	  u32 now = (u32) time (0);

	  if (PREDICT_TRUE (rv->gc_mark_timestamp == 0))
	    rv->gc_mark_timestamp = now;
	  else
	    {
	      if (now - rv->gc_mark_timestamp > 10)
		{
		  if (CLIB_DEBUG > 0)
		    {
		      u16 *msg_idp, msg_id;
		      clib_warning
			("garbage collect pool %d ring %d index %d", pool, i,
			 q->head);
		      msg_idp = (u16 *) (rv->data);
		      msg_id = clib_net_to_host_u16 (*msg_idp);
		      if (msg_id < vec_len (api_main.msg_names))
			clib_warning ("msg id %d name %s", (u32) msg_id,
				      api_main.msg_names[msg_id]);
		    }
		  shmem_hdr->garbage_collects++;
		  goto collected;
		}
	    }


	  /* yes, loser; try next larger pool */
	  ap[i].misses++;
	  if (pool == 0)
	    pthread_mutex_unlock (&q->mutex);
	  continue;
	}
    collected:

      /* OK, we have a winner */
      ap[i].hits++;
      /*
       * Remember the source queue, although we
       * don't need to know the queue to free the item.
       */
      rv->q = q;
      rv->gc_mark_timestamp = 0;
      q->head++;
      if (q->head == q->maxsize)
	q->head = 0;

      if (pool == 0)
	pthread_mutex_unlock (&q->mutex);
      goto out;
    }

  /*
   * Request too big, or head element of all size-compatible rings
   * still in use. Fall back to shared-memory malloc.
   */
  am->ring_misses++;

  pthread_mutex_lock (&am->vlib_rp->mutex);
  oldheap = svm_push_data_heap (am->vlib_rp);
  if (may_return_null)
    {
      rv = clib_mem_alloc_or_null (nbytes);
      if (PREDICT_FALSE (rv == 0))
	{
	  svm_pop_heap (oldheap);
	  pthread_mutex_unlock (&am->vlib_rp->mutex);
	  return 0;
	}
    }
  else
    rv = clib_mem_alloc (nbytes);

  rv->q = 0;
  svm_pop_heap (oldheap);
  pthread_mutex_unlock (&am->vlib_rp->mutex);

out:
#if DEBUG_MESSAGE_BUFFER_OVERRUN > 0
  {
    nbytes -= 4;
    u32 *overrun;
    overrun = (u32 *) (rv->data + nbytes - sizeof (msgbuf_t));
    *overrun = 0x1badbabe;
  }
#endif
  rv->data_len = htonl (nbytes - sizeof (msgbuf_t));

  return (rv->data);
}

void *
vl_msg_api_alloc (int nbytes)
{
  int pool;
  api_main_t *am = &api_main;
  vl_shmem_hdr_t *shmem_hdr = am->shmem_hdr;

  /*
   * Clients use pool-0, vlib proc uses pool 1
   */
  pool = (am->our_pid == shmem_hdr->vl_pid);
  return vl_msg_api_alloc_internal (nbytes, pool, 0 /* may_return_null */ );
}

void *
vl_msg_api_alloc_or_null (int nbytes)
{
  int pool;
  api_main_t *am = &api_main;
  vl_shmem_hdr_t *shmem_hdr = am->shmem_hdr;

  pool = (am->our_pid == shmem_hdr->vl_pid);
  return vl_msg_api_alloc_internal (nbytes, pool, 1 /* may_return_null */ );
}

void *
vl_msg_api_alloc_as_if_client (int nbytes)
{
  return vl_msg_api_alloc_internal (nbytes, 0, 0 /* may_return_null */ );
}

void *
vl_msg_api_alloc_as_if_client_or_null (int nbytes)
{
  return vl_msg_api_alloc_internal (nbytes, 0, 1 /* may_return_null */ );
}

void *
vl_mem_api_alloc_as_if_client_w_reg (vl_api_registration_t * reg, int nbytes)
{
  api_main_t *am = &api_main;
  vl_shmem_hdr_t *save_shmem_hdr = am->shmem_hdr;
  svm_region_t *vlib_rp, *save_vlib_rp = am->vlib_rp;
  void *msg;

  vlib_rp = am->vlib_rp = reg->vlib_rp;
  am->shmem_hdr = (void *) vlib_rp->user_ctx;

  msg = vl_msg_api_alloc_internal (nbytes, 0, 0 /* may_return_null */ );

  am->shmem_hdr = save_shmem_hdr;
  am->vlib_rp = save_vlib_rp;

  return msg;
}

void
vl_msg_api_free (void *a)
{
  msgbuf_t *rv;
  void *oldheap;
  api_main_t *am = &api_main;

  rv = (msgbuf_t *) (((u8 *) a) - offsetof (msgbuf_t, data));

  /*
   * Here's the beauty of the scheme.  Only one proc/thread has
   * control of a given message buffer. To free a buffer, we just clear the
   * queue field, and leave. No locks, no hits, no errors...
   */
  if (rv->q)
    {
      rv->q = 0;
      rv->gc_mark_timestamp = 0;
#if DEBUG_MESSAGE_BUFFER_OVERRUN > 0
      {
	u32 *overrun;
	overrun = (u32 *) (rv->data + ntohl (rv->data_len));
	ASSERT (*overrun == 0x1badbabe);
      }
#endif
      return;
    }

  pthread_mutex_lock (&am->vlib_rp->mutex);
  oldheap = svm_push_data_heap (am->vlib_rp);

#if DEBUG_MESSAGE_BUFFER_OVERRUN > 0
  {
    u32 *overrun;
    overrun = (u32 *) (rv->data + ntohl (rv->data_len));
    ASSERT (*overrun == 0x1badbabe);
  }
#endif

  clib_mem_free (rv);
  svm_pop_heap (oldheap);
  pthread_mutex_unlock (&am->vlib_rp->mutex);
}

static void
vl_msg_api_free_nolock (void *a)
{
  msgbuf_t *rv;
  void *oldheap;
  api_main_t *am = &api_main;

  rv = (msgbuf_t *) (((u8 *) a) - offsetof (msgbuf_t, data));
  /*
   * Here's the beauty of the scheme.  Only one proc/thread has
   * control of a given message buffer. To free a buffer, we just clear the
   * queue field, and leave. No locks, no hits, no errors...
   */
  if (rv->q)
    {
      rv->q = 0;
      return;
    }

  oldheap = svm_push_data_heap (am->vlib_rp);
  clib_mem_free (rv);
  svm_pop_heap (oldheap);
}

void
vl_set_memory_root_path (const char *name)
{
  api_main_t *am = &api_main;

  am->root_path = name;
}

void
vl_set_memory_uid (int uid)
{
  api_main_t *am = &api_main;

  am->api_uid = uid;
}

void
vl_set_memory_gid (int gid)
{
  api_main_t *am = &api_main;

  am->api_gid = gid;
}

void
vl_set_global_memory_baseva (u64 baseva)
{
  api_main_t *am = &api_main;

  am->global_baseva = baseva;
}

void
vl_set_global_memory_size (u64 size)
{
  api_main_t *am = &api_main;

  am->global_size = size;
}

void
vl_set_api_memory_size (u64 size)
{
  api_main_t *am = &api_main;

  am->api_size = size;
}

void
vl_set_global_pvt_heap_size (u64 size)
{
  api_main_t *am = &api_main;

  am->global_pvt_heap_size = size;
}

void
vl_set_api_pvt_heap_size (u64 size)
{
  api_main_t *am = &api_main;

  am->api_pvt_heap_size = size;
}

static void
vl_api_default_mem_config (vl_shmem_hdr_t * shmem_hdr)
{
  api_main_t *am = &api_main;
  u32 vlib_input_queue_length;

  /* vlib main input queue */
  vlib_input_queue_length = 1024;
  if (am->vlib_input_queue_length)
    vlib_input_queue_length = am->vlib_input_queue_length;

  shmem_hdr->vl_input_queue =
    svm_queue_alloc_and_init (vlib_input_queue_length, sizeof (uword),
			      getpid ());

#define _(sz,n)                                                 \
    do {                                                        \
        ring_alloc_t _rp;                                       \
        _rp.rp = svm_queue_alloc_and_init ((n), (sz), 0); 	\
        _rp.size = (sz);                                        \
        _rp.nitems = n;                                         \
        _rp.hits = 0;                                           \
        _rp.misses = 0;                                         \
        vec_add1(shmem_hdr->vl_rings, _rp);                     \
    } while (0);

  foreach_vl_aring_size;
#undef _

#define _(sz,n)                                                 \
    do {                                                        \
        ring_alloc_t _rp;                                       \
        _rp.rp = svm_queue_alloc_and_init ((n), (sz), 0); 	\
        _rp.size = (sz);                                        \
        _rp.nitems = n;                                         \
        _rp.hits = 0;                                           \
        _rp.misses = 0;                                         \
        vec_add1(shmem_hdr->client_rings, _rp);                 \
    } while (0);

  foreach_clnt_aring_size;
#undef _
}

void
vl_api_mem_config (vl_shmem_hdr_t * hdr, vl_api_shm_elem_config_t * config)
{
  vl_api_shm_elem_config_t *c;
  ring_alloc_t *rp;
  u32 size;

  if (!config)
    {
      vl_api_default_mem_config (hdr);
      return;
    }

  vec_foreach (c, config)
  {
    switch (c->type)
      {
      case VL_API_QUEUE:
	hdr->vl_input_queue = svm_queue_alloc_and_init (c->count, c->size,
							getpid ());
	continue;
      case VL_API_VLIB_RING:
	vec_add2 (hdr->vl_rings, rp, 1);
	break;
      case VL_API_CLIENT_RING:
	vec_add2 (hdr->client_rings, rp, 1);
	break;
      default:
	clib_warning ("unknown config type: %d", c->type);
	continue;
      }

    size = sizeof (ring_alloc_t) + c->size;
    rp->rp = svm_queue_alloc_and_init (c->count, size, 0);
    rp->size = size;
    rp->nitems = c->count;
    rp->hits = 0;
    rp->misses = 0;
  }
}

void
vl_init_shmem (svm_region_t * vlib_rp, vl_api_shm_elem_config_t * config,
	       int is_vlib, int is_private_region)
{
  api_main_t *am = &api_main;
  vl_shmem_hdr_t *shmem_hdr = 0;
  void *oldheap;
  ASSERT (vlib_rp);

  /* $$$$ need private region config parameters */

  oldheap = svm_push_data_heap (vlib_rp);

  vec_validate (shmem_hdr, 0);
  shmem_hdr->version = VL_SHM_VERSION;
  shmem_hdr->clib_file_index = VL_API_INVALID_FI;

  /* Set up the queue and msg ring allocator */
  vl_api_mem_config (shmem_hdr, config);

  if (is_private_region == 0)
    {
      am->shmem_hdr = shmem_hdr;
      am->vlib_rp = vlib_rp;
      am->our_pid = getpid ();
      if (is_vlib)
	am->shmem_hdr->vl_pid = am->our_pid;
    }
  else
    shmem_hdr->vl_pid = am->our_pid;

  svm_pop_heap (oldheap);

  /*
   * After absolutely everything that a client might see is set up,
   * declare the shmem region valid
   */
  vlib_rp->user_ctx = shmem_hdr;

  pthread_mutex_unlock (&vlib_rp->mutex);
}

int
vl_map_shmem (const char *region_name, int is_vlib)
{
  svm_map_region_args_t _a, *a = &_a;
  svm_region_t *vlib_rp, *root_rp;
  api_main_t *am = &api_main;
  int i, rv;
  struct timespec ts, tsrem;
  char *vpe_api_region_suffix = "-vpe-api";

  clib_memset (a, 0, sizeof (*a));

  if (strstr (region_name, vpe_api_region_suffix))
    {
      u8 *root_path = format (0, "%s", region_name);
      _vec_len (root_path) = (vec_len (root_path) -
			      strlen (vpe_api_region_suffix));
      vec_terminate_c_string (root_path);
      a->root_path = (const char *) root_path;
      am->root_path = (const char *) root_path;
    }

  if (is_vlib == 0)
    {
      rv = svm_region_init_chroot (am->root_path);
      if (rv)
	return rv;
    }

  if (a->root_path != NULL)
    {
      a->name = "/vpe-api";
    }
  else
    a->name = region_name;
  a->size = am->api_size ? am->api_size : (16 << 20);
  a->flags = SVM_FLAGS_MHEAP;
  a->uid = am->api_uid;
  a->gid = am->api_gid;
  a->pvt_heap_size = am->api_pvt_heap_size;

  vlib_rp = svm_region_find_or_create (a);

  if (vlib_rp == 0)
    return (-2);

  pthread_mutex_lock (&vlib_rp->mutex);
  /* Has someone else set up the shared-memory variable table? */
  if (vlib_rp->user_ctx)
    {
      am->shmem_hdr = (void *) vlib_rp->user_ctx;
      am->our_pid = getpid ();
      if (is_vlib)
	{
	  svm_queue_t *q;
	  uword old_msg;
	  /*
	   * application restart. Reset cached pids, API message
	   * rings, list of clients; otherwise, various things
	   * fail. (e.g. queue non-empty notification)
	   */

	  /* ghosts keep the region from disappearing properly */
	  svm_client_scan_this_region_nolock (vlib_rp);
	  am->shmem_hdr->application_restarts++;
	  q = am->shmem_hdr->vl_input_queue;
	  am->shmem_hdr->vl_pid = getpid ();
	  q->consumer_pid = am->shmem_hdr->vl_pid;
	  /* Drain the input queue, freeing msgs */
	  for (i = 0; i < 10; i++)
	    {
	      if (pthread_mutex_trylock (&q->mutex) == 0)
		{
		  pthread_mutex_unlock (&q->mutex);
		  goto mutex_ok;
		}
	      ts.tv_sec = 0;
	      ts.tv_nsec = 10000 * 1000;	/* 10 ms */
	      while (nanosleep (&ts, &tsrem) < 0)
		ts = tsrem;
	    }
	  /* Mutex buggered, "fix" it */
	  clib_memset (&q->mutex, 0, sizeof (q->mutex));
	  clib_warning ("forcibly release main input queue mutex");

	mutex_ok:
	  am->vlib_rp = vlib_rp;
	  while (svm_queue_sub (q, (u8 *) & old_msg, SVM_Q_NOWAIT, 0)
		 != -2 /* queue underflow */ )
	    {
	      vl_msg_api_free_nolock ((void *) old_msg);
	      am->shmem_hdr->restart_reclaims++;
	    }
	  pthread_mutex_unlock (&vlib_rp->mutex);
	  root_rp = svm_get_root_rp ();
	  ASSERT (root_rp);
	  /* Clean up the root region client list */
	  pthread_mutex_lock (&root_rp->mutex);
	  svm_client_scan_this_region_nolock (root_rp);
	  pthread_mutex_unlock (&root_rp->mutex);
	}
      else
	{
	  pthread_mutex_unlock (&vlib_rp->mutex);
	}
      am->vlib_rp = vlib_rp;
      vec_add1 (am->mapped_shmem_regions, vlib_rp);
      return 0;
    }
  /* Clients simply have to wait... */
  if (!is_vlib)
    {
      pthread_mutex_unlock (&vlib_rp->mutex);

      /* Wait up to 100 seconds... */
      for (i = 0; i < 10000; i++)
	{
	  ts.tv_sec = 0;
	  ts.tv_nsec = 10000 * 1000;	/* 10 ms */
	  while (nanosleep (&ts, &tsrem) < 0)
	    ts = tsrem;
	  if (vlib_rp->user_ctx)
	    goto ready;
	}
      /* Clean up and leave... */
      svm_region_unmap (vlib_rp);
      clib_warning ("region init fail");
      return (-2);

    ready:
      am->shmem_hdr = (void *) vlib_rp->user_ctx;
      am->our_pid = getpid ();
      am->vlib_rp = vlib_rp;
      vec_add1 (am->mapped_shmem_regions, vlib_rp);
      return 0;
    }

  /* Nope, it's our problem... */
  vl_init_shmem (vlib_rp, 0 /* default config */ , 1 /* is vlib */ ,
		 0 /* is_private_region */ );

  vec_add1 (am->mapped_shmem_regions, vlib_rp);
  return 0;
}

void
vl_register_mapped_shmem_region (svm_region_t * rp)
{
  api_main_t *am = &api_main;

  vec_add1 (am->mapped_shmem_regions, rp);
}

static void
vl_unmap_shmem_internal (u8 is_client)
{
  svm_region_t *rp;
  int i;
  api_main_t *am = &api_main;

  if (!svm_get_root_rp ())
    return;

  for (i = 0; i < vec_len (am->mapped_shmem_regions); i++)
    {
      rp = am->mapped_shmem_regions[i];
      is_client ? svm_region_unmap_client (rp) : svm_region_unmap (rp);
    }

  vec_free (am->mapped_shmem_regions);
  am->shmem_hdr = 0;

  is_client ? svm_region_exit_client () : svm_region_exit ();

  /* $$$ more careful cleanup, valgrind run... */
  vec_free (am->msg_handlers);
  vec_free (am->msg_endian_handlers);
  vec_free (am->msg_print_handlers);
}

void
vl_unmap_shmem (void)
{
  vl_unmap_shmem_internal (0);
}

void
vl_unmap_shmem_client (void)
{
  vl_unmap_shmem_internal (1);
}

void
vl_msg_api_send_shmem (svm_queue_t * q, u8 * elem)
{
  api_main_t *am = &api_main;
  uword *trace = (uword *) elem;

  if (am->tx_trace && am->tx_trace->enabled)
    vl_msg_api_trace (am, am->tx_trace, (void *) trace[0]);

  /*
   * Announce a probable binary API client bug:
   * some client's input queue is stuffed.
   * The situation may be recoverable, or not.
   */
  if (PREDICT_FALSE
      (am->vl_clients /* vpp side */  && (q->cursize == q->maxsize)))
    clib_warning ("WARNING: client input queue at %llx is stuffed...", q);
  (void) svm_queue_add (q, elem, 0 /* nowait */ );
}

int
vl_mem_api_can_send (svm_queue_t * q)
{
  return (q->cursize < q->maxsize);
}

void
vl_msg_api_send_shmem_nolock (svm_queue_t * q, u8 * elem)
{
  api_main_t *am = &api_main;
  uword *trace = (uword *) elem;

  if (am->tx_trace && am->tx_trace->enabled)
    vl_msg_api_trace (am, am->tx_trace, (void *) trace[0]);

  (void) svm_queue_add_nolock (q, elem);
}

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