Whamcloud - gitweb
LU-3963 libcfs: remove proc handler wrappers
[fs/lustre-release.git] / lnet / lnet / lib-msg.c
index 4ab5fa4..b1995f2 100644 (file)
@@ -1,6 +1,4 @@
-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
  * GPL HEADER START
  *
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
@@ -28,6 +26,8 @@
 /*
  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
+ *
+ * Copyright (c) 2012, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -57,45 +57,309 @@ lnet_build_unlink_event (lnet_libmd_t *md, lnet_event_t *ev)
         EXIT;
 }
 
+/*
+ * Don't need any lock, must be called after lnet_commit_md
+ */
 void
-lnet_enq_event_locked (lnet_eq_t *eq, lnet_event_t *ev)
+lnet_build_msg_event(lnet_msg_t *msg, lnet_event_kind_t ev_type)
 {
-        lnet_event_t  *eq_slot;
+       lnet_hdr_t      *hdr = &msg->msg_hdr;
+       lnet_event_t    *ev  = &msg->msg_ev;
+
+       LASSERT(!msg->msg_routing);
+
+       ev->type = ev_type;
+
+       if (ev_type == LNET_EVENT_SEND) {
+               /* event for active message */
+               ev->target.nid    = le64_to_cpu(hdr->dest_nid);
+               ev->target.pid    = le32_to_cpu(hdr->dest_pid);
+               ev->initiator.nid = LNET_NID_ANY;
+               ev->initiator.pid = the_lnet.ln_pid;
+               ev->sender        = LNET_NID_ANY;
+
+       } else {
+               /* event for passive message */
+               ev->target.pid    = hdr->dest_pid;
+               ev->target.nid    = hdr->dest_nid;
+               ev->initiator.pid = hdr->src_pid;
+               ev->initiator.nid = hdr->src_nid;
+               ev->rlength       = hdr->payload_length;
+               ev->sender        = msg->msg_from;
+               ev->mlength       = msg->msg_wanted;
+               ev->offset        = msg->msg_offset;
+       }
+
+       switch (ev_type) {
+       default:
+               LBUG();
+
+       case LNET_EVENT_PUT: /* passive PUT */
+               ev->pt_index   = hdr->msg.put.ptl_index;
+               ev->match_bits = hdr->msg.put.match_bits;
+               ev->hdr_data   = hdr->msg.put.hdr_data;
+               return;
+
+       case LNET_EVENT_GET: /* passive GET */
+               ev->pt_index   = hdr->msg.get.ptl_index;
+               ev->match_bits = hdr->msg.get.match_bits;
+               ev->hdr_data   = 0;
+               return;
+
+       case LNET_EVENT_ACK: /* ACK */
+               ev->match_bits = hdr->msg.ack.match_bits;
+               ev->mlength    = hdr->msg.ack.mlength;
+               return;
+
+       case LNET_EVENT_REPLY: /* REPLY */
+               return;
+
+       case LNET_EVENT_SEND: /* active message */
+               if (msg->msg_type == LNET_MSG_PUT) {
+                       ev->pt_index   = le32_to_cpu(hdr->msg.put.ptl_index);
+                       ev->match_bits = le64_to_cpu(hdr->msg.put.match_bits);
+                       ev->offset     = le32_to_cpu(hdr->msg.put.offset);
+                       ev->mlength    =
+                       ev->rlength    = le32_to_cpu(hdr->payload_length);
+                       ev->hdr_data   = le64_to_cpu(hdr->msg.put.hdr_data);
+
+               } else {
+                       LASSERT(msg->msg_type == LNET_MSG_GET);
+                       ev->pt_index   = le32_to_cpu(hdr->msg.get.ptl_index);
+                       ev->match_bits = le64_to_cpu(hdr->msg.get.match_bits);
+                       ev->mlength    =
+                       ev->rlength    = le32_to_cpu(hdr->msg.get.sink_length);
+                       ev->offset     = le32_to_cpu(hdr->msg.get.src_offset);
+                       ev->hdr_data   = 0;
+               }
+               return;
+       }
+}
 
-        /* Allocate the next queue slot */
-        ev->sequence = eq->eq_enq_seq++;
+void
+lnet_msg_commit(lnet_msg_t *msg, int cpt)
+{
+       struct lnet_msg_container *container = the_lnet.ln_msg_containers[cpt];
+       lnet_counters_t           *counters  = the_lnet.ln_counters[cpt];
+
+       /* routed message can be committed for both receiving and sending */
+       LASSERT(!msg->msg_tx_committed);
+
+       if (msg->msg_sending) {
+               LASSERT(!msg->msg_receiving);
+
+               msg->msg_tx_cpt = cpt;
+               msg->msg_tx_committed = 1;
+               if (msg->msg_rx_committed) { /* routed message REPLY */
+                       LASSERT(msg->msg_onactivelist);
+                       return;
+               }
+       } else {
+               LASSERT(!msg->msg_sending);
+               msg->msg_rx_cpt = cpt;
+               msg->msg_rx_committed = 1;
+       }
+
+       LASSERT(!msg->msg_onactivelist);
+       msg->msg_onactivelist = 1;
+       list_add(&msg->msg_activelist, &container->msc_active);
+
+       counters->msgs_alloc++;
+       if (counters->msgs_alloc > counters->msgs_max)
+               counters->msgs_max = counters->msgs_alloc;
+}
 
-        /* size must be a power of 2 to handle sequence # overflow */
-        LASSERT (eq->eq_size != 0 &&
-                 eq->eq_size == LOWEST_BIT_SET (eq->eq_size));
-        eq_slot = eq->eq_events + (ev->sequence & (eq->eq_size - 1));
+static void
+lnet_msg_decommit_tx(lnet_msg_t *msg, int status)
+{
+       lnet_counters_t *counters;
+       lnet_event_t    *ev = &msg->msg_ev;
+
+       LASSERT(msg->msg_tx_committed);
+       if (status != 0)
+               goto out;
+
+       counters = the_lnet.ln_counters[msg->msg_tx_cpt];
+       switch (ev->type) {
+       default: /* routed message */
+               LASSERT(msg->msg_routing);
+               LASSERT(msg->msg_rx_committed);
+               LASSERT(ev->type == 0);
+
+               counters->route_length += msg->msg_len;
+               counters->route_count++;
+               goto out;
+
+       case LNET_EVENT_PUT:
+               /* should have been decommitted */
+               LASSERT(!msg->msg_rx_committed);
+               /* overwritten while sending ACK */
+               LASSERT(msg->msg_type == LNET_MSG_ACK);
+               msg->msg_type = LNET_MSG_PUT; /* fix type */
+               break;
+
+       case LNET_EVENT_SEND:
+               LASSERT(!msg->msg_rx_committed);
+               if (msg->msg_type == LNET_MSG_PUT)
+                       counters->send_length += msg->msg_len;
+               break;
+
+       case LNET_EVENT_GET:
+               LASSERT(msg->msg_rx_committed);
+               /* overwritten while sending reply, we should never be
+                * here for optimized GET */
+               LASSERT(msg->msg_type == LNET_MSG_REPLY);
+               msg->msg_type = LNET_MSG_GET; /* fix type */
+               break;
+       }
+
+       counters->send_count++;
+ out:
+       lnet_return_tx_credits_locked(msg);
+       msg->msg_tx_committed = 0;
+}
 
-        /* There is no race since both event consumers and event producers
-         * take the LNET_LOCK, so we don't screw around with memory
-         * barriers, setting the sequence number last or weird structure
-         * layout assertions. */
-        *eq_slot = *ev;
+static void
+lnet_msg_decommit_rx(lnet_msg_t *msg, int status)
+{
+       lnet_counters_t *counters;
+       lnet_event_t    *ev = &msg->msg_ev;
+
+       LASSERT(!msg->msg_tx_committed); /* decommitted or never committed */
+       LASSERT(msg->msg_rx_committed);
+
+       if (status != 0)
+               goto out;
+
+       counters = the_lnet.ln_counters[msg->msg_rx_cpt];
+       switch (ev->type) {
+       default:
+               LASSERT(ev->type == 0);
+               LASSERT(msg->msg_routing);
+               goto out;
+
+       case LNET_EVENT_ACK:
+               LASSERT(msg->msg_type == LNET_MSG_ACK);
+               break;
+
+       case LNET_EVENT_GET:
+               /* type is "REPLY" if it's an optimized GET on passive side,
+                * because optimized GET will never be committed for sending,
+                * so message type wouldn't be changed back to "GET" by
+                * lnet_msg_decommit_tx(), see details in lnet_parse_get() */
+               LASSERT(msg->msg_type == LNET_MSG_REPLY ||
+                       msg->msg_type == LNET_MSG_GET);
+               counters->send_length += msg->msg_wanted;
+               break;
+
+       case LNET_EVENT_PUT:
+               LASSERT(msg->msg_type == LNET_MSG_PUT);
+               break;
+
+       case LNET_EVENT_REPLY:
+               /* type is "GET" if it's an optimized GET on active side,
+                * see details in lnet_create_reply_msg() */
+               LASSERT(msg->msg_type == LNET_MSG_GET ||
+                       msg->msg_type == LNET_MSG_REPLY);
+               break;
+       }
+
+       counters->recv_count++;
+       if (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_REPLY)
+               counters->recv_length += msg->msg_wanted;
 
-        /* Call the callback handler (if any) */
-        if (eq->eq_callback != NULL)
-                eq->eq_callback (eq_slot);
+ out:
+       lnet_return_rx_credits_locked(msg);
+       msg->msg_rx_committed = 0;
+}
 
-#ifdef __KERNEL__
-        /* Wake anyone waiting in LNetEQPoll() */
-        if (cfs_waitq_active(&the_lnet.ln_waitq))
-                cfs_waitq_broadcast(&the_lnet.ln_waitq);
-#else
-# ifndef HAVE_LIBPTHREAD
-        /* LNetEQPoll() calls into _the_ LND to wait for action */
-# else
-        /* Wake anyone waiting in LNetEQPoll() */
-        pthread_cond_broadcast(&the_lnet.ln_cond);
-# endif
-#endif
+void
+lnet_msg_decommit(lnet_msg_t *msg, int cpt, int status)
+{
+       int     cpt2 = cpt;
+
+       LASSERT(msg->msg_tx_committed || msg->msg_rx_committed);
+       LASSERT(msg->msg_onactivelist);
+
+       if (msg->msg_tx_committed) { /* always decommit for sending first */
+               LASSERT(cpt == msg->msg_tx_cpt);
+               lnet_msg_decommit_tx(msg, status);
+       }
+
+       if (msg->msg_rx_committed) {
+               /* forwarding msg committed for both receiving and sending */
+               if (cpt != msg->msg_rx_cpt) {
+                       lnet_net_unlock(cpt);
+                       cpt2 = msg->msg_rx_cpt;
+                       lnet_net_lock(cpt2);
+               }
+               lnet_msg_decommit_rx(msg, status);
+       }
+
+       list_del(&msg->msg_activelist);
+       msg->msg_onactivelist = 0;
+
+       the_lnet.ln_counters[cpt2]->msgs_alloc--;
+
+       if (cpt2 != cpt) {
+               lnet_net_unlock(cpt2);
+               lnet_net_lock(cpt);
+       }
 }
 
 void
-lnet_complete_msg_locked(lnet_msg_t *msg)
+lnet_msg_attach_md(lnet_msg_t *msg, lnet_libmd_t *md,
+                  unsigned int offset, unsigned int mlen)
+{
+       /* NB: @offset and @len are only useful for receiving */
+       /* Here, we attach the MD on lnet_msg and mark it busy and
+        * decrementing its threshold. Come what may, the lnet_msg "owns"
+        * the MD until a call to lnet_msg_detach_md or lnet_finalize()
+        * signals completion. */
+       LASSERT(!msg->msg_routing);
+
+       msg->msg_md = md;
+       if (msg->msg_receiving) { /* committed for receiving */
+               msg->msg_offset = offset;
+               msg->msg_wanted = mlen;
+       }
+
+       md->md_refcount++;
+       if (md->md_threshold != LNET_MD_THRESH_INF) {
+               LASSERT(md->md_threshold > 0);
+               md->md_threshold--;
+       }
+
+       /* build umd in event */
+       lnet_md2handle(&msg->msg_ev.md_handle, md);
+       lnet_md_deconstruct(md, &msg->msg_ev.md);
+}
+
+void
+lnet_msg_detach_md(lnet_msg_t *msg, int status)
+{
+       lnet_libmd_t    *md = msg->msg_md;
+       int             unlink;
+
+       /* Now it's safe to drop my caller's ref */
+       md->md_refcount--;
+       LASSERT(md->md_refcount >= 0);
+
+       unlink = lnet_md_unlinkable(md);
+       if (md->md_eq != NULL) {
+               msg->msg_ev.status   = status;
+               msg->msg_ev.unlinked = unlink;
+               lnet_eq_enqueue_event(md->md_eq, &msg->msg_ev);
+       }
+
+       if (unlink)
+               lnet_md_unlink(md);
+
+       msg->msg_md = NULL;
+}
+
+static int
+lnet_complete_msg_locked(lnet_msg_t *msg, int cpt)
 {
         lnet_handle_wire_t ack_wmd;
         int                rc;
@@ -106,10 +370,10 @@ lnet_complete_msg_locked(lnet_msg_t *msg)
         if (status == 0 && msg->msg_ack) {
                 /* Only send an ACK if the PUT completed successfully */
 
-                lnet_return_credits_locked(msg);
+               lnet_msg_decommit(msg, cpt, 0);
 
-                msg->msg_ack = 0;
-                LNET_UNLOCK();
+               msg->msg_ack = 0;
+               lnet_net_unlock(cpt);
 
                 LASSERT(msg->msg_ev.type == LNET_EVENT_PUT);
                 LASSERT(!msg->msg_routing);
@@ -122,50 +386,67 @@ lnet_complete_msg_locked(lnet_msg_t *msg)
                 msg->msg_hdr.msg.ack.match_bits = msg->msg_ev.match_bits;
                 msg->msg_hdr.msg.ack.mlength = cpu_to_le32(msg->msg_ev.mlength);
 
-                rc = lnet_send(msg->msg_ev.target.nid, msg);
-
-                LNET_LOCK();
-
-                if (rc == 0)
-                        return;
-        } else if (status == 0 &&               /* OK so far */
-                   (msg->msg_routing && !msg->msg_sending)) { /* not forwarded */
-                
-                LASSERT (!msg->msg_receiving);  /* called back recv already */
-        
-                LNET_UNLOCK();
-                
-                rc = lnet_send(LNET_NID_ANY, msg);
-
-                LNET_LOCK();
-
-                if (rc == 0)
-                        return;
-        }
-
-        lnet_return_credits_locked(msg);
-
-        LASSERT (msg->msg_onactivelist);
-        msg->msg_onactivelist = 0;
-        cfs_list_del (&msg->msg_activelist);
-        the_lnet.ln_counters.msgs_alloc--;
-        lnet_msg_free(msg);
+               /* NB: we probably want to use NID of msg::msg_from as 3rd
+                * parameter (router NID) if it's routed message */
+               rc = lnet_send(msg->msg_ev.target.nid, msg, LNET_NID_ANY);
+
+               lnet_net_lock(cpt);
+               /*
+                * NB: message is committed for sending, we should return
+                * on success because LND will finalize this message later.
+                *
+                * Also, there is possibility that message is committed for
+                * sending and also failed before delivering to LND,
+                * i.e: ENOMEM, in that case we can't fall through either
+                * because CPT for sending can be different with CPT for
+                * receiving, so we should return back to lnet_finalize()
+                * to make sure we are locking the correct partition.
+                */
+               return rc;
+
+       } else if (status == 0 &&       /* OK so far */
+                  (msg->msg_routing && !msg->msg_sending)) {
+               /* not forwarded */
+               LASSERT(!msg->msg_receiving);   /* called back recv already */
+               lnet_net_unlock(cpt);
+
+               rc = lnet_send(LNET_NID_ANY, msg, LNET_NID_ANY);
+
+               lnet_net_lock(cpt);
+               /*
+                * NB: message is committed for sending, we should return
+                * on success because LND will finalize this message later.
+                *
+                * Also, there is possibility that message is committed for
+                * sending and also failed before delivering to LND,
+                * i.e: ENOMEM, in that case we can't fall through either:
+                * - The rule is message must decommit for sending first if
+                *   the it's committed for both sending and receiving
+                * - CPT for sending can be different with CPT for receiving,
+                *   so we should return back to lnet_finalize() to make
+                *   sure we are locking the correct partition.
+                */
+               return rc;
+       }
+
+       lnet_msg_decommit(msg, cpt, status);
+       lnet_msg_free_locked(msg);
+       return 0;
 }
 
-
 void
 lnet_finalize (lnet_ni_t *ni, lnet_msg_t *msg, int status)
 {
-#ifdef __KERNEL__
-        int                i;
-        int                my_slot;
-#endif
-        lnet_libmd_t      *md;
+       struct lnet_msg_container       *container;
+       int                             my_slot;
+       int                             cpt;
+       int                             rc;
+       int                             i;
 
-        LASSERT (!cfs_in_interrupt ());
+       LASSERT (!in_interrupt ());
 
-        if (msg == NULL)
-                return;
+       if (msg == NULL)
+               return;
 #if 0
         CDEBUG(D_WARNING, "%s msg->%s Flags:%s%s%s%s%s%s%s%s%s%s%s txp %s rxp %s\n",
                lnet_msgtyp2str(msg->msg_type), libcfs_id2str(msg->msg_target),
@@ -183,74 +464,198 @@ lnet_finalize (lnet_ni_t *ni, lnet_msg_t *msg, int status)
                msg->msg_txpeer == NULL ? "<none>" : libcfs_nid2str(msg->msg_txpeer->lp_nid),
                msg->msg_rxpeer == NULL ? "<none>" : libcfs_nid2str(msg->msg_rxpeer->lp_nid));
 #endif
-        LNET_LOCK();
-
-        LASSERT (msg->msg_onactivelist);
-
         msg->msg_ev.status = status;
 
-        md = msg->msg_md;
-        if (md != NULL) {
-                int      unlink;
+       if (msg->msg_md != NULL) {
+               cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
+
+               lnet_res_lock(cpt);
+               lnet_msg_detach_md(msg, status);
+               lnet_res_unlock(cpt);
+       }
+
+ again:
+       rc = 0;
+       if (!msg->msg_tx_committed && !msg->msg_rx_committed) {
+               /* not committed to network yet */
+               LASSERT(!msg->msg_onactivelist);
+               lnet_msg_free(msg);
+               return;
+       }
+
+       /*
+        * NB: routed message can be committed for both receiving and sending,
+        * we should finalize in LIFO order and keep counters correct.
+        * (finalize sending first then finalize receiving)
+        */
+       cpt = msg->msg_tx_committed ? msg->msg_tx_cpt : msg->msg_rx_cpt;
+       lnet_net_lock(cpt);
+
+       container = the_lnet.ln_msg_containers[cpt];
+       list_add_tail(&msg->msg_list, &container->msc_finalizing);
+
+       /* Recursion breaker.  Don't complete the message here if I am (or
+        * enough other threads are) already completing messages */
 
-                /* Now it's safe to drop my caller's ref */
-                md->md_refcount--;
-                LASSERT (md->md_refcount >= 0);
+#ifdef __KERNEL__
+       my_slot = -1;
+       for (i = 0; i < container->msc_nfinalizers; i++) {
+               if (container->msc_finalizers[i] == current)
+                       break;
 
-                unlink = lnet_md_unlinkable(md);
+               if (my_slot < 0 && container->msc_finalizers[i] == NULL)
+                       my_slot = i;
+       }
 
-                msg->msg_ev.unlinked = unlink;
+       if (i < container->msc_nfinalizers || my_slot < 0) {
+               lnet_net_unlock(cpt);
+               return;
+       }
 
-                if (md->md_eq != NULL)
-                        lnet_enq_event_locked(md->md_eq, &msg->msg_ev);
+       container->msc_finalizers[my_slot] = current;
+#else
+       LASSERT(container->msc_nfinalizers == 1);
+       if (container->msc_finalizers[0] != NULL) {
+               lnet_net_unlock(cpt);
+               return;
+       }
+
+       my_slot = i = 0;
+       container->msc_finalizers[0] = (struct lnet_msg_container *)1;
+#endif
 
-                if (unlink)
-                        lnet_md_unlink(md);
+       while (!list_empty(&container->msc_finalizing)) {
+               msg = list_entry(container->msc_finalizing.next,
+                                lnet_msg_t, msg_list);
 
-                msg->msg_md = NULL;
-        }
+               list_del(&msg->msg_list);
 
-        cfs_list_add_tail (&msg->msg_list, &the_lnet.ln_finalizeq);
+               /* NB drops and regains the lnet lock if it actually does
+                * anything, so my finalizing friends can chomp along too */
+               rc = lnet_complete_msg_locked(msg, cpt);
+               if (rc != 0)
+                       break;
+       }
 
-        /* Recursion breaker.  Don't complete the message here if I am (or
-         * enough other threads are) already completing messages */
+       container->msc_finalizers[my_slot] = NULL;
+       lnet_net_unlock(cpt);
 
-#ifdef __KERNEL__
-        my_slot = -1;
-        for (i = 0; i < the_lnet.ln_nfinalizers; i++) {
-                if (the_lnet.ln_finalizers[i] == cfs_current())
-                        goto out;
-                if (my_slot < 0 && the_lnet.ln_finalizers[i] == NULL)
-                        my_slot = i;
-        }
-        if (my_slot < 0)
-                goto out;
-
-        the_lnet.ln_finalizers[my_slot] = cfs_current();
-#else
-        if (the_lnet.ln_finalizing)
-                goto out;
+       if (rc != 0)
+               goto again;
+}
+EXPORT_SYMBOL(lnet_finalize);
 
-        the_lnet.ln_finalizing = 1;
+void
+lnet_msg_container_cleanup(struct lnet_msg_container *container)
+{
+       int     count = 0;
+
+       if (container->msc_init == 0)
+               return;
+
+       while (!list_empty(&container->msc_active)) {
+               lnet_msg_t *msg = list_entry(container->msc_active.next,
+                                            lnet_msg_t, msg_activelist);
+
+               LASSERT(msg->msg_onactivelist);
+               msg->msg_onactivelist = 0;
+               list_del(&msg->msg_activelist);
+               lnet_msg_free(msg);
+               count++;
+       }
+
+       if (count > 0)
+               CERROR("%d active msg on exit\n", count);
+
+       if (container->msc_finalizers != NULL) {
+               LIBCFS_FREE(container->msc_finalizers,
+                           container->msc_nfinalizers *
+                           sizeof(*container->msc_finalizers));
+               container->msc_finalizers = NULL;
+       }
+#ifdef LNET_USE_LIB_FREELIST
+       lnet_freelist_fini(&container->msc_freelist);
 #endif
+       container->msc_init = 0;
+}
 
-        while (!cfs_list_empty(&the_lnet.ln_finalizeq)) {
-                msg = cfs_list_entry(the_lnet.ln_finalizeq.next,
-                                     lnet_msg_t, msg_list);
+int
+lnet_msg_container_setup(struct lnet_msg_container *container, int cpt)
+{
+       int     rc;
 
-                cfs_list_del(&msg->msg_list);
+       container->msc_init = 1;
 
-                /* NB drops and regains the lnet lock if it actually does
-                 * anything, so my finalizing friends can chomp along too */
-                lnet_complete_msg_locked(msg);
-        }
+       INIT_LIST_HEAD(&container->msc_active);
+       INIT_LIST_HEAD(&container->msc_finalizing);
 
-#ifdef __KERNEL__
-        the_lnet.ln_finalizers[my_slot] = NULL;
+#ifdef LNET_USE_LIB_FREELIST
+       memset(&container->msc_freelist, 0, sizeof(lnet_freelist_t));
+
+       rc = lnet_freelist_init(&container->msc_freelist,
+                               LNET_FL_MAX_MSGS, sizeof(lnet_msg_t));
+       if (rc != 0) {
+               CERROR("Failed to init freelist for message container\n");
+               lnet_msg_container_cleanup(container);
+               return rc;
+       }
 #else
-        the_lnet.ln_finalizing = 0;
+       rc = 0;
 #endif
+       /* number of CPUs */
+       container->msc_nfinalizers = cfs_cpt_weight(lnet_cpt_table(), cpt);
 
- out:
-        LNET_UNLOCK();
+       LIBCFS_CPT_ALLOC(container->msc_finalizers, lnet_cpt_table(), cpt,
+                        container->msc_nfinalizers *
+                        sizeof(*container->msc_finalizers));
+
+       if (container->msc_finalizers == NULL) {
+               CERROR("Failed to allocate message finalizers\n");
+               lnet_msg_container_cleanup(container);
+               return -ENOMEM;
+       }
+
+       return rc;
+}
+
+void
+lnet_msg_containers_destroy(void)
+{
+       struct lnet_msg_container *container;
+       int     i;
+
+       if (the_lnet.ln_msg_containers == NULL)
+               return;
+
+       cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers)
+               lnet_msg_container_cleanup(container);
+
+       cfs_percpt_free(the_lnet.ln_msg_containers);
+       the_lnet.ln_msg_containers = NULL;
+}
+
+int
+lnet_msg_containers_create(void)
+{
+       struct lnet_msg_container *container;
+       int     rc;
+       int     i;
+
+       the_lnet.ln_msg_containers = cfs_percpt_alloc(lnet_cpt_table(),
+                                                     sizeof(*container));
+
+       if (the_lnet.ln_msg_containers == NULL) {
+               CERROR("Failed to allocate cpu-partition data for network\n");
+               return -ENOMEM;
+       }
+
+       cfs_percpt_for_each(container, i, the_lnet.ln_msg_containers) {
+               rc = lnet_msg_container_setup(container, i);
+               if (rc != 0) {
+                       lnet_msg_containers_destroy();
+                       return rc;
+               }
+       }
+
+       return 0;
 }