[fs/lustre-release.git] / lnet / klnds / gnilnd / gnilnd_conn.c

/*
 * Copyright (C) 2012 Cray, Inc.
 *
 * Copyright (c) 2014, Intel Corporation.
 *
 *   Author: Nic Henke <nic@cray.com>
 *   Author: James Shimek <jshimek@cray.com>
 *
 *   This file is part of Lustre, http://www.lustre.org.
 *
 *   Lustre is free software; you can redistribute it and/or
 *   modify it under the terms of version 2 of the GNU General Public
 *   License as published by the Free Software Foundation.
 *
 *   Lustre is distributed in the hope that it will be useful,
 *   but WITHOUT ANY WARRANTY; without even the implied warranty of
 *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *   GNU General Public License for more details.
 *
 *   You should have received a copy of the GNU General Public License
 *   along with Lustre; if not, write to the Free Software
 *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
 *
 */

#include "gnilnd.h"

void
kgnilnd_setup_smsg_attr(gni_smsg_attr_t *smsg_attr)
{
        smsg_attr->mbox_maxcredit = *kgnilnd_tunables.kgn_mbox_credits;
        smsg_attr->msg_maxsize = GNILND_MAX_MSG_SIZE;
        smsg_attr->msg_type = GNI_SMSG_TYPE_MBOX_AUTO_RETRANSMIT;
}

int
kgnilnd_map_fmablk(kgn_device_t *device, kgn_fma_memblock_t *fma_blk)
{
        gni_return_t            rrc;
        __u32                   flags = GNI_MEM_READWRITE;
        static unsigned long    reg_to;
        int                     rfto = *kgnilnd_tunables.kgn_reg_fail_timeout;

        if (fma_blk->gnm_state == GNILND_FMABLK_PHYS) {
                flags |= GNI_MEM_PHYS_CONT;
        }

        fma_blk->gnm_hold_timeout = 0;

        /* make sure we are mapping a clean block */
        LASSERTF(fma_blk->gnm_hndl.qword1 == 0UL, "fma_blk %p dirty\n", fma_blk);

        rrc = kgnilnd_mem_register(device->gnd_handle, (__u64)fma_blk->gnm_block,
                                   fma_blk->gnm_blk_size, device->gnd_rcv_fma_cqh,
                                   flags, &fma_blk->gnm_hndl);
        if (rrc != GNI_RC_SUCCESS) {
                if (rfto != GNILND_REGFAILTO_DISABLE) {
                        if (reg_to == 0) {
                                reg_to = jiffies + cfs_time_seconds(rfto);
                        } else if (time_after(jiffies, reg_to)) {
                                CERROR("FATAL:fmablk registration has failed "
                                       "for %ld seconds.\n",
                                       cfs_duration_sec(jiffies - reg_to) +
                                                rfto);
                                LBUG();
                        }
                }

                CNETERR("register fmablk failed 0x%p mbox_size %d flags %u\n",
                        fma_blk, fma_blk->gnm_mbox_size, flags);
                RETURN(-ENOMEM);
        }

        reg_to = 0;

        /* PHYS_CONT memory isn't really mapped, at least not in GART -
         *  but all mappings chew up a MDD
         */
        if (fma_blk->gnm_state != GNILND_FMABLK_PHYS) {
                atomic64_add(fma_blk->gnm_blk_size, &device->gnd_nbytes_map);
        }

        atomic_inc(&device->gnd_n_mdd);
        /* nfmablk is live (mapped) blocks */
        atomic_inc(&device->gnd_nfmablk);

        RETURN(0);
}

int
kgnilnd_alloc_fmablk(kgn_device_t *device, int use_phys)
{
        int                     rc = 0;
        int                     num_mbox;
        kgn_fma_memblock_t     *fma_blk;
        gni_smsg_attr_t         smsg_attr;
        unsigned long           fmablk_vers;

#if defined(CONFIG_CRAY_XT) && !defined(CONFIG_CRAY_COMPUTE)
        /* We allocate large blocks of memory here potentially leading
         * to memory exhaustion during massive reconnects during a network
         * outage. Limit the amount of fma blocks to use by always keeping
         * a percent of pages free initially set to 25% of total memory. */
        if (global_page_state(NR_FREE_PAGES) < kgnilnd_data.free_pages_limit) {
                LCONSOLE_INFO("Exceeding free page limit of %ld. "
                              "Free pages available %ld\n",
                              kgnilnd_data.free_pages_limit,
                              global_page_state(NR_FREE_PAGES));
                return -ENOMEM;
        }
#endif
        /* we'll use fmablk_vers and the gnd_fmablk_mutex to gate access
         * to this allocation code. Everyone will sample the version
         * before and after getting the mutex. If it has changed,
         * we'll bail out to check the lists again - this indicates that
         * some sort of change was made to the lists and it is possible
         * that there is a mailbox for us to find now. This should prevent
         * a ton of spinning in the case where there are lots of threads
         * that need a yet-to-be-allocated mailbox for a connection. */

        fmablk_vers = atomic_read(&device->gnd_fmablk_vers);
        mutex_lock(&device->gnd_fmablk_mutex);

        if (fmablk_vers != atomic_read(&device->gnd_fmablk_vers)) {
                /* version changed while we were waiting for semaphore,
                 * we'll recheck the lists assuming something nice happened */
                mutex_unlock(&device->gnd_fmablk_mutex);
                return 0;
        }

        LIBCFS_ALLOC(fma_blk, sizeof(kgn_fma_memblock_t));
        if (fma_blk == NULL) {
                CNETERR("could not allocate fma block descriptor\n");
                rc = -ENOMEM;
                GOTO(out, rc);
        }

        INIT_LIST_HEAD(&fma_blk->gnm_bufflist);

        kgnilnd_setup_smsg_attr(&smsg_attr);

        gni_smsg_buff_size_needed(&smsg_attr, &fma_blk->gnm_mbox_size);

        LASSERTF(fma_blk->gnm_mbox_size, "mbox size %d\n", fma_blk->gnm_mbox_size);

        /* gni_smsg_buff_size_needed calculates the base mailbox size and since
         * we want to hold kgn_peer_credits worth of messages in both directions,
         * we add PAYLOAD to grow the mailbox size
         */

        fma_blk->gnm_mbox_size += GNILND_MBOX_PAYLOAD;

        /* we'll only use physical during preallocate at startup -- this keeps it nice and
         * clean for runtime decisions. We'll keep the PHYS ones around until shutdown
         * as reallocating them is tough if there is memory fragmentation */

        if (use_phys) {
                fma_blk->gnm_block = kmem_cache_alloc(kgnilnd_data.kgn_mbox_cache, GFP_ATOMIC);
                if (fma_blk->gnm_block == NULL) {
                        CNETERR("could not allocate physical SMSG mailbox memory\n");
                        rc = -ENOMEM;
                        GOTO(free_desc, rc);
                }
                fma_blk->gnm_blk_size = GNILND_MBOX_SIZE;
                num_mbox = fma_blk->gnm_blk_size / fma_blk->gnm_mbox_size;

                LASSERTF(num_mbox >= 1,
                         "num_mbox %d blk_size %u mbox_size %d\n",
                          num_mbox, fma_blk->gnm_blk_size, fma_blk->gnm_mbox_size);

                fma_blk->gnm_state = GNILND_FMABLK_PHYS;

        } else {
                num_mbox = *kgnilnd_tunables.kgn_mbox_per_block;
                fma_blk->gnm_blk_size = num_mbox * fma_blk->gnm_mbox_size;

                LASSERTF(num_mbox >= 1 && num_mbox >= *kgnilnd_tunables.kgn_mbox_per_block,
                         "num_mbox %d blk_size %u mbox_size %d tunable %d\n",
                         num_mbox, fma_blk->gnm_blk_size, fma_blk->gnm_mbox_size,
                         *kgnilnd_tunables.kgn_mbox_per_block);

                fma_blk->gnm_block = kgnilnd_vzalloc(fma_blk->gnm_blk_size);
                if (fma_blk->gnm_block == NULL) {
                        CNETERR("could not allocate virtual SMSG mailbox memory, %d bytes\n", fma_blk->gnm_blk_size);
                        rc = -ENOMEM;
                        GOTO(free_desc, rc);
                }

                fma_blk->gnm_state = GNILND_FMABLK_VIRT;
        }

        /* allocate just enough space for the bits to track the mailboxes */
        CFS_ALLOC_PTR_ARRAY(fma_blk->gnm_bit_array, BITS_TO_LONGS(num_mbox));
        if (fma_blk->gnm_bit_array == NULL) {
                CNETERR("could not allocate mailbox bitmask, %lu bytes for %d mbox\n",
                       sizeof(unsigned long) * BITS_TO_LONGS(num_mbox), num_mbox);
                rc = -ENOMEM;
                GOTO(free_blk, rc);
        }
        bitmap_zero(fma_blk->gnm_bit_array, num_mbox);

        /* now that the num_mbox is set based on allocation type, get debug
         * info setup
         * */
        CFS_ALLOC_PTR_ARRAY(fma_blk->gnm_mbox_info, num_mbox);
        if (fma_blk->gnm_mbox_info == NULL) {
                CNETERR("could not allocate mailbox debug, %lu bytes for %d mbox\n",
                       sizeof(kgn_mbox_info_t) * num_mbox, num_mbox);
                rc = -ENOMEM;
                GOTO(free_bit, rc);
        }

        rc = kgnilnd_map_fmablk(device, fma_blk);
        if (rc) {
                GOTO(free_info, rc);
        }

        fma_blk->gnm_next_avail_mbox = 0;
        fma_blk->gnm_avail_mboxs = fma_blk->gnm_num_mboxs = num_mbox;

        CDEBUG(D_MALLOC, "alloc fmablk 0x%p num %d msg_maxsize %d credits %d "
                "mbox_size %d MDD %#llx.%#llx\n",
                fma_blk, num_mbox, smsg_attr.msg_maxsize, smsg_attr.mbox_maxcredit,
                fma_blk->gnm_mbox_size, fma_blk->gnm_hndl.qword1,
                fma_blk->gnm_hndl.qword2);

        /* lock Is protecting data structures, not semaphore */

        spin_lock(&device->gnd_fmablk_lock);
        list_add_tail(&fma_blk->gnm_bufflist, &device->gnd_fma_buffs);

        /* toggle under the lock so once they change the list is also
         * ready for others to traverse */
        atomic_inc(&device->gnd_fmablk_vers);

        spin_unlock(&device->gnd_fmablk_lock);

        mutex_unlock(&device->gnd_fmablk_mutex);

        return 0;

free_info:
        CFS_FREE_PTR_ARRAY(fma_blk->gnm_mbox_info, num_mbox);
free_bit:
        CFS_FREE_PTR_ARRAY(fma_blk->gnm_bit_array, BITS_TO_LONGS(num_mbox));
free_blk:
        if (fma_blk->gnm_state == GNILND_FMABLK_VIRT) {
                kgnilnd_vfree(fma_blk->gnm_block, fma_blk->gnm_blk_size);
        } else {
                kmem_cache_free(kgnilnd_data.kgn_mbox_cache, fma_blk->gnm_block);
        }
free_desc:
        LIBCFS_FREE(fma_blk, sizeof(kgn_fma_memblock_t));
out:
        mutex_unlock(&device->gnd_fmablk_mutex);
        return rc;
}

void
kgnilnd_unmap_fmablk(kgn_device_t *dev, kgn_fma_memblock_t *fma_blk)
{
        gni_return_t            rrc;

        /* if some held, set hold_timeout from conn timeouts used in this block
         * but not during shutdown, then just nuke and pave
         * During a stack reset, we need to deregister with a hold timeout
         * set so we don't use the same mdd after reset is complete */
        if ((fma_blk->gnm_held_mboxs && !kgnilnd_data.kgn_shutdown) ||
            kgnilnd_data.kgn_in_reset) {
                fma_blk->gnm_hold_timeout = GNILND_TIMEOUT2DEADMAN;
        }

        /* we are changing the state of a block, tickle version to tell
         * proc code list is stale now */
        atomic_inc(&dev->gnd_fmablk_vers);

        rrc = kgnilnd_mem_deregister(dev->gnd_handle, &fma_blk->gnm_hndl, fma_blk->gnm_hold_timeout);

        CDEBUG(rrc == GNI_RC_SUCCESS ? D_MALLOC : D_CONSOLE|D_NETERROR,
               "unmap fmablk 0x%p@%s sz %u total %d avail %d held %d mbox_size %d "
                "hold_timeout %d\n",
               fma_blk, kgnilnd_fmablk_state2str(fma_blk->gnm_state),
               fma_blk->gnm_blk_size, fma_blk->gnm_num_mboxs,
               fma_blk->gnm_avail_mboxs, fma_blk->gnm_held_mboxs,
               fma_blk->gnm_mbox_size, fma_blk->gnm_hold_timeout);

        LASSERTF(rrc == GNI_RC_SUCCESS,
                "tried to double unmap or something bad, fma_blk %p (rrc %d)\n",
                fma_blk, rrc);

        if (fma_blk->gnm_hold_timeout &&
            !(kgnilnd_data.kgn_in_reset &&
              fma_blk->gnm_state == GNILND_FMABLK_PHYS)) {
                atomic_inc(&dev->gnd_n_mdd_held);
        } else {
                atomic_dec(&dev->gnd_n_mdd);
        }

        /* PHYS blocks don't get mapped */
        if (fma_blk->gnm_state != GNILND_FMABLK_PHYS) {
                atomic64_sub(fma_blk->gnm_blk_size, &dev->gnd_nbytes_map);
                fma_blk->gnm_state = GNILND_FMABLK_IDLE;
        } else if (kgnilnd_data.kgn_in_reset) {
                /* in stack reset, clear MDD handle for PHYS blocks, as we'll
                 * re-use the fma_blk after reset so we don't have to drop/allocate
                 * all of those physical blocks */
                fma_blk->gnm_hndl.qword1 = fma_blk->gnm_hndl.qword2 = 0UL;
        }

        /* Decrement here as this is the # of mapped blocks */
        atomic_dec(&dev->gnd_nfmablk);
}


/* needs lock on gnd_fmablk_lock to cover gnd_fma_buffs */
void
kgnilnd_free_fmablk_locked(kgn_device_t *dev, kgn_fma_memblock_t *fma_blk)
{
        LASSERTF(fma_blk->gnm_avail_mboxs == fma_blk->gnm_num_mboxs,
                 "fma_blk %p@%d free in bad state (%d): blk total %d avail %d held %d\n",
                 fma_blk, fma_blk->gnm_state, fma_blk->gnm_hold_timeout, fma_blk->gnm_num_mboxs,
                fma_blk->gnm_avail_mboxs, fma_blk->gnm_held_mboxs);

        atomic_inc(&dev->gnd_fmablk_vers);

        if (fma_blk->gnm_hold_timeout) {
                CDEBUG(D_MALLOC, "mdd release fmablk 0x%p sz %u avail %d held %d "
                        "mbox_size %d\n",
                        fma_blk, fma_blk->gnm_blk_size, fma_blk->gnm_avail_mboxs,
                        fma_blk->gnm_held_mboxs, fma_blk->gnm_mbox_size);

                /* We leave MDD dangling over stack reset */
                if (!kgnilnd_data.kgn_in_reset) {
                        kgnilnd_mem_mdd_release(dev->gnd_handle, &fma_blk->gnm_hndl);
                }
                /* ignoring the return code - if kgni/ghal can't find it
                 * it must be released already */
                atomic_dec(&dev->gnd_n_mdd_held);
                atomic_dec(&dev->gnd_n_mdd);
        }

        /* we cant' free the gnm_block until all the conns have released their
         * purgatory holds. While we have purgatory holds, we might check the conn
         * RX mailbox during the CLOSING process. It is possible that kgni might
         * try to look into the RX side for credits when sending the CLOSE msg too */
        CDEBUG(D_MALLOC, "fmablk %p free buffer %p mbox_size %d\n",
                fma_blk, fma_blk->gnm_block, fma_blk->gnm_mbox_size);

        if (fma_blk->gnm_state == GNILND_FMABLK_PHYS) {
                kmem_cache_free(kgnilnd_data.kgn_mbox_cache, fma_blk->gnm_block);
        } else {
                kgnilnd_vfree(fma_blk->gnm_block, fma_blk->gnm_blk_size);
        }
        fma_blk->gnm_state = GNILND_FMABLK_FREED;

        list_del(&fma_blk->gnm_bufflist);

        CFS_FREE_PTR_ARRAY(fma_blk->gnm_mbox_info, fma_blk->gnm_num_mboxs);
        CFS_FREE_PTR_ARRAY(fma_blk->gnm_bit_array,
                           BITS_TO_LONGS(fma_blk->gnm_num_mboxs));
        LIBCFS_FREE(fma_blk, sizeof(kgn_fma_memblock_t));
}

void
kgnilnd_find_free_mbox(kgn_conn_t *conn)
{
        kgn_device_t            *dev = conn->gnc_device;
        gni_smsg_attr_t         *smsg_attr = &conn->gnpr_smsg_attr;
        kgn_fma_memblock_t      *fma_blk;
        kgn_mbox_info_t         *mbox = NULL;
        int                     id;

        spin_lock(&dev->gnd_fmablk_lock);

        list_for_each_entry(fma_blk, &conn->gnc_device->gnd_fma_buffs,
                            gnm_bufflist) {
                if (fma_blk->gnm_avail_mboxs <= 0 ||
                    fma_blk->gnm_state <= GNILND_FMABLK_IDLE) {
                        continue;
                }
                /* look in bitarray for available mailbox */
                do {
                        id = find_next_zero_bit(
                                fma_blk->gnm_bit_array,
                                fma_blk->gnm_num_mboxs,
                                fma_blk->gnm_next_avail_mbox);
                      if (id == fma_blk->gnm_num_mboxs &&
                          fma_blk->gnm_next_avail_mbox != 0) {
                                /* wrap around */
                                fma_blk->gnm_next_avail_mbox = 0;
                        } else {
                                break;
                        }
                } while (1);

                LASSERTF(id < fma_blk->gnm_num_mboxs, "id %d max %d\n",
                         id, fma_blk->gnm_num_mboxs);
                set_bit(id, (volatile unsigned long *)fma_blk->gnm_bit_array);
                conn->gnc_mbox_id = id;

                fma_blk->gnm_next_avail_mbox =
                        (id == (fma_blk->gnm_num_mboxs - 1)) ? 0 : (id + 1);
                fma_blk->gnm_avail_mboxs--;
                conn->gnc_fma_blk = fma_blk;

                kgnilnd_setup_smsg_attr(smsg_attr);

                smsg_attr->msg_buffer = fma_blk->gnm_block;
                smsg_attr->mbox_offset = fma_blk->gnm_mbox_size * id;
                smsg_attr->mem_hndl = fma_blk->gnm_hndl;
                smsg_attr->buff_size = fma_blk->gnm_mbox_size;

                /* We'll set the hndl to zero for PHYS blocks unmapped during stack
                 * reset and re-use the same fma_blk after stack reset. This ensures we've
                 * properly mapped it before we use it */
                LASSERTF(fma_blk->gnm_hndl.qword1 != 0UL, "unmapped fma_blk %p, state %d\n",
                         fma_blk, fma_blk->gnm_state);

                CDEBUG(D_NET, "conn %p smsg %p fmablk %p "
                        "allocating SMSG mbox %d buf %p "
                        "offset %u hndl %#llx.%#llx\n",
                        conn, smsg_attr, fma_blk, id,
                        smsg_attr->msg_buffer, smsg_attr->mbox_offset,
                        fma_blk->gnm_hndl.qword1,
                        fma_blk->gnm_hndl.qword2);

                mbox = &fma_blk->gnm_mbox_info[id];
                mbox->mbx_create_conn_memset = jiffies;
                mbox->mbx_nallocs++;
                mbox->mbx_nallocs_total++;

                /* zero mbox to remove any old data from our last use.
                 * this better be safe, if not our purgatory timers
                 * are too short or a peer really is misbehaving */
                memset(smsg_attr->msg_buffer + smsg_attr->mbox_offset,
                       0, smsg_attr->buff_size);
                break;
        }

        spin_unlock(&dev->gnd_fmablk_lock);
}

int
kgnilnd_setup_mbox(kgn_conn_t *conn)
{
        gni_smsg_attr_t         *smsg_attr = &conn->gnpr_smsg_attr;
        int                      err = 0;

        smsg_attr->msg_buffer = NULL;
        /* Look for available mbox */
        do {
                kgnilnd_find_free_mbox(conn);

                /* nothing in the existing buffers, make a new one */
                if (smsg_attr->msg_buffer == NULL) {
                        /* for runtime allocations, we only want vmalloc */
                        err = kgnilnd_alloc_fmablk(conn->gnc_device, 0);
                        if (err) {
                                break;
                        }
                }
        } while (smsg_attr->msg_buffer == NULL);

        if (err)
                CNETERR("couldn't allocate SMSG mbox for conn %p Error: %d\n",
                        conn, err);
        return err;
}

void
kgnilnd_release_mbox(kgn_conn_t *conn, int purgatory_hold)
{
        kgn_device_t           *dev = conn->gnc_device;
        gni_smsg_attr_t        *smsg_attr = &conn->gnpr_smsg_attr;
        kgn_fma_memblock_t     *fma_blk = NULL;
        kgn_mbox_info_t        *mbox = NULL;
        int                     found = 0;
        int                     id;

        /* if we failed to setup mbox and now destroying conn */
        if (smsg_attr->msg_buffer == NULL) {
                return;
        }

        id = conn->gnc_mbox_id;

        spin_lock(&dev->gnd_fmablk_lock);
        /* make sure our conn points at a valid fma_blk
         * We use this instead of a mem block search out of smsg_attr
         * because we could have freed a block for fma_blk #1 but the fma_blk
         * is still in the list for a purgatory hold. This would induce a false
         * match if that same block gets reallocated to fma_blk #2 */
        list_for_each_entry(fma_blk, &dev->gnd_fma_buffs, gnm_bufflist) {
                if (fma_blk == conn->gnc_fma_blk) {
                        found = 1;
                        break;
                }
        }
        LASSERTF(found, "unable to find conn 0x%p with gnc_fma_blk %p "
                 "anywhere in the world\n", conn, conn->gnc_fma_blk);

        LASSERTF(id < fma_blk->gnm_num_mboxs,
                "bad id %d max %d\n",
                id, fma_blk->gnm_num_mboxs);

        /* < 0 - was held, now free it
         * == 0 - just free it
         * > 0 - hold it for now */
        if (purgatory_hold == 0) {
                CDEBUG(D_NET, "conn %p smsg %p fmablk %p freeing SMSG mbox %d "
                        "hndl %#llx.%#llx\n",
                        conn, smsg_attr, fma_blk, id,
                        fma_blk->gnm_hndl.qword1, fma_blk->gnm_hndl.qword2);
                fma_blk->gnm_avail_mboxs++;

        } else if (purgatory_hold > 0) {
                CDEBUG(D_NET, "conn %p smsg %p fmablk %p holding SMSG mbox %d "
                        "hndl %#llx.%#llx\n",
                        conn, smsg_attr, fma_blk, id,
                        fma_blk->gnm_hndl.qword1, fma_blk->gnm_hndl.qword2);

                fma_blk->gnm_held_mboxs++;
                fma_blk->gnm_max_timeout = max_t(long, fma_blk->gnm_max_timeout,
                                                 conn->gnc_timeout);
        } else {
                CDEBUG(D_NET, "conn %p smsg %p fmablk %p release SMSG mbox %d "
                        "hndl %#llx.%#llx\n",
                        conn, smsg_attr, fma_blk, id,
                        fma_blk->gnm_hndl.qword1, fma_blk->gnm_hndl.qword2);

                fma_blk->gnm_held_mboxs--;
                fma_blk->gnm_avail_mboxs++;
        }

        if (purgatory_hold <= 0) {
                /* if kgni is retransmitting, freeing the smsg block before the EP
                 * is destroyed gets messy. Bug 768295. */
                LASSERTF(conn->gnc_ephandle == NULL,
                         "can't release mbox before EP is nuked. conn 0x%p\n", conn);

                mbox = &fma_blk->gnm_mbox_info[id];
                mbox->mbx_release_from_purgatory = jiffies;

                /* clear conn gnc_fmablk if it is gone - this allows us to
                 * not worry about state so much in kgnilnd_destroy_conn
                 * and makes the guaranteed cleanup of the resources easier */
                LASSERTF(test_and_clear_bit(id, fma_blk->gnm_bit_array),
                        "conn %p bit %d already cleared in fma_blk %p\n",
                         conn, id, fma_blk);
                conn->gnc_fma_blk = NULL;
                mbox->mbx_nallocs--;
        }

        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_FMABLK_AVAIL)) {
                CERROR("LBUGs in your future: forcibly marking fma_blk %p "
                       "as mapped\n", fma_blk);
                fma_blk->gnm_state = GNILND_FMABLK_VIRT;
        }

        /* we don't release or unmap PHYS blocks as part of the normal cycle --
         * those are controlled manually from startup/shutdown */
        if (fma_blk->gnm_state != GNILND_FMABLK_PHYS) {
                /* we can unmap once all are unused (held or avail)
                 * but check hold_timeout to make sure we are not trying to double
                 * unmap this buffer. If there was no hold_timeout set due to
                 * held_mboxs, we'll free the mobx here shortly and won't have to
                 * worry about catching a double free for a 'clean' fma_blk */
                if (((fma_blk->gnm_avail_mboxs + fma_blk->gnm_held_mboxs) == fma_blk->gnm_num_mboxs) &&
                    (!fma_blk->gnm_hold_timeout)) {
                        kgnilnd_unmap_fmablk(dev, fma_blk);
                }

                /* But we can only free once they are all avail */
                if (fma_blk->gnm_avail_mboxs == fma_blk->gnm_num_mboxs &&
                    fma_blk->gnm_held_mboxs == 0) {
                        /* all mailboxes are released, free fma_blk */
                        kgnilnd_free_fmablk_locked(dev, fma_blk);
                }
        }

        spin_unlock(&dev->gnd_fmablk_lock);
}

int
kgnilnd_count_phys_mbox(kgn_device_t *device)
{
        int                     i = 0;
        kgn_fma_memblock_t     *fma_blk;

        spin_lock(&device->gnd_fmablk_lock);

        list_for_each_entry(fma_blk, &device->gnd_fma_buffs, gnm_bufflist) {
                if (fma_blk->gnm_state == GNILND_FMABLK_PHYS)
                        i += fma_blk->gnm_num_mboxs;
        }
        spin_unlock(&device->gnd_fmablk_lock);

        RETURN(i);
}

int
kgnilnd_allocate_phys_fmablk(kgn_device_t *device)
{
        int     rc;

        while (kgnilnd_count_phys_mbox(device) < *kgnilnd_tunables.kgn_nphys_mbox) {

                rc = kgnilnd_alloc_fmablk(device, 1);
                if (rc) {
                        CERROR("failed phys mbox allocation, stopping at %d, rc %d\n",
                                kgnilnd_count_phys_mbox(device), rc);
                        RETURN(rc);
                }
        }
        RETURN(0);
}

int
kgnilnd_map_phys_fmablk(kgn_device_t *device)
{

        int                     rc = 0;
        kgn_fma_memblock_t     *fma_blk;

        /* use mutex to gate access to single thread, just in case */
        mutex_lock(&device->gnd_fmablk_mutex);

        spin_lock(&device->gnd_fmablk_lock);

        list_for_each_entry(fma_blk, &device->gnd_fma_buffs, gnm_bufflist) {
                if (fma_blk->gnm_state == GNILND_FMABLK_PHYS) {
                        rc = kgnilnd_map_fmablk(device, fma_blk);
                        if (rc)
                                break;
                }
        }
        spin_unlock(&device->gnd_fmablk_lock);

        mutex_unlock(&device->gnd_fmablk_mutex);

        RETURN(rc);
}

void
kgnilnd_unmap_fma_blocks(kgn_device_t *device)
{

        kgn_fma_memblock_t      *fma_blk;

        /* use mutex to gate access to single thread, just in case */
        mutex_lock(&device->gnd_fmablk_mutex);

        spin_lock(&device->gnd_fmablk_lock);

        list_for_each_entry(fma_blk, &device->gnd_fma_buffs, gnm_bufflist) {
                kgnilnd_unmap_fmablk(device, fma_blk);
        }
        spin_unlock(&device->gnd_fmablk_lock);

        mutex_unlock(&device->gnd_fmablk_mutex);
}

void
kgnilnd_free_phys_fmablk(kgn_device_t *device)
{

        kgn_fma_memblock_t      *fma_blk, *fma_blkN;

        /* use mutex to gate access to single thread, just in case */
        mutex_lock(&device->gnd_fmablk_mutex);

        spin_lock(&device->gnd_fmablk_lock);

        list_for_each_entry_safe(fma_blk, fma_blkN, &device->gnd_fma_buffs, gnm_bufflist) {
                if (fma_blk->gnm_state == GNILND_FMABLK_PHYS)
                        kgnilnd_free_fmablk_locked(device, fma_blk);
        }
        spin_unlock(&device->gnd_fmablk_lock);

        mutex_unlock(&device->gnd_fmablk_mutex);
}

/* kgnilnd dgram nid->struct managment */

static inline struct list_head *
kgnilnd_nid2dgramlist(kgn_device_t *dev, lnet_nid_t nid)
{
        unsigned int hash = ((unsigned int)nid) % *kgnilnd_tunables.kgn_peer_hash_size;

        RETURN(&dev->gnd_dgrams[hash]);
}


/* needs dev->gnd_dgram_lock held */
kgn_dgram_t *
kgnilnd_find_dgram_locked(kgn_device_t *dev, lnet_nid_t dst_nid)
{
        struct list_head *dgram_list = kgnilnd_nid2dgramlist(dev, dst_nid);
        kgn_dgram_t      *dgram;

        list_for_each_entry(dgram, dgram_list, gndg_list) {

                /* if state > POSTED, we are already handling cancel/completion */
                if ((dgram->gndg_conn_out.gncr_dstnid != dst_nid) ||
                     dgram->gndg_state > GNILND_DGRAM_POSTED)
                        continue;

                CDEBUG(D_NET, "got dgram [%p] -> %s\n",
                       dgram, libcfs_nid2str(dst_nid));
                return dgram;
        }
        return NULL;
}

int
kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid)
{
        kgn_dgram_t     *dgram;

        spin_lock(&dev->gnd_dgram_lock);
        dgram = kgnilnd_find_dgram_locked(dev, dst_nid);

        if (dgram) {
                kgnilnd_cancel_dgram_locked(dgram);
        }
        spin_unlock(&dev->gnd_dgram_lock);

        RETURN(!!(dgram == NULL));
}

int
kgnilnd_pack_connreq(kgn_connreq_t *connreq, kgn_conn_t *conn,
                     lnet_nid_t srcnid, lnet_nid_t dstnid,
                     kgn_connreq_type_t type)
{
        int err = 0;

        /* ensure we haven't violated max datagram size */
        BUILD_BUG_ON(sizeof(kgn_connreq_t) > GNI_DATAGRAM_MAXSIZE);

        /* no need to zero out, we do that when allocating dgram */
        connreq->gncr_magic     = GNILND_MSG_MAGIC;

        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PACK_SRCNID)) {
                srcnid = 0xABADBABE;
        } else if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PACK_DSTNID)) {
                dstnid = 0xDEFEC8ED;
        }

        connreq->gncr_srcnid    = srcnid;
        connreq->gncr_dstnid    = dstnid;

        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
                connreq->gncr_version = 99;
        } else {
                connreq->gncr_version   = GNILND_CONNREQ_VERSION;
        }
        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
                connreq->gncr_type = 99;
        } else {
                connreq->gncr_type      = type;
        }
        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
                connreq->gncr_peerstamp = 0;
        } else {
                connreq->gncr_peerstamp = kgnilnd_data.kgn_peerstamp;
        }
        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
                connreq->gncr_connstamp = 0;
        } else {
                connreq->gncr_connstamp = conn->gnc_my_connstamp;
        }
        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_PROTO)) {
                connreq->gncr_timeout = 0;
        } else {
                connreq->gncr_timeout   = conn->gnc_timeout;
        }

        /* the rest pack the data into the payload in other places */
        if (type == GNILND_CONNREQ_REQ) {
                kgn_gniparams_t       *req_params = &connreq->gncr_gnparams;
                req_params->gnpr_host_id = conn->gnc_device->gnd_host_id;
                req_params->gnpr_cqid = conn->gnc_cqid;

                /* allocate mailbox for this connection */
                err = kgnilnd_setup_mbox(conn);
                if (err != 0) {
                        CERROR("Failed to setup FMA mailbox (%d)\n", err);
                }
                req_params->gnpr_smsg_attr = conn->gnpr_smsg_attr;
        }

        /* XXX Nic: TBD - checksum computation */

        return err;
}

int
kgnilnd_unpack_connreq(kgn_dgram_t *dgram)
{
        kgn_connreq_t           *connreq = &dgram->gndg_conn_in;
        int                      swab, rc = 0;
        kgn_net_t               *net;

        /* the following fields must be handled in a backwards compatible
         * manner to ensure we can always send and interpret NAKs */

        if (connreq->gncr_magic != GNILND_MSG_MAGIC &&
            connreq->gncr_magic != __swab32(GNILND_MSG_MAGIC)) {
                /* Unexpected magic! */
                CERROR("Unexpected magic %08x\n",
                       connreq->gncr_magic);
                return -EBADF;
        }

        swab = (connreq->gncr_magic == __swab32(GNILND_MSG_MAGIC));
        if (swab) {
                __swab32s(&connreq->gncr_magic);
                __swab32s(&connreq->gncr_cksum);
                __swab16s(&connreq->gncr_type);
                __swab16s(&connreq->gncr_version);
                __swab32s(&connreq->gncr_timeout);
                __swab64s(&connreq->gncr_srcnid);
                __swab64s(&connreq->gncr_dstnid);
                __swab64s(&connreq->gncr_peerstamp);
                __swab64s(&connreq->gncr_connstamp);
        }

        /* Do NOT return anything but -EBADF before we munge
         * connreq->gncr_srcnid - we need that to send the nak */

        if (dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
                lnet_nid_t      incoming = connreq->gncr_srcnid;

                /* even if the incoming packet is hosed, we know who we sent
                 * the original and can set the srcnid so that we can properly
                 * look up our peer to close the loop on this connreq. We still use
                 * -EBADF to prevent a NAK - just in case there are issues with
                 * the payload coming from a random spot, etc. */
                connreq->gncr_srcnid = dgram->gndg_conn_out.gncr_dstnid;

                if (LNET_NIDADDR(dgram->gndg_conn_out.gncr_dstnid) !=
                                LNET_NIDADDR(incoming)) {
                        /* we got a datagram match for the wrong nid... */
                        CERROR("matched datagram 0x%p with srcnid %s "
                                "(%x), expecting %s (%x)\n",
                                dgram,
                                libcfs_nid2str(incoming),
                                LNET_NIDADDR(incoming),
                                libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
                                LNET_NIDADDR(dgram->gndg_conn_out.gncr_dstnid));
                        return -EBADF;
                }
        } else {
                /* if we have a wildcard datagram it should match an
                 * incoming "active" datagram that should have a fully formed
                 * srcnid and dstnid. If we couldn't unpack it, we drop as
                 * corrupted packet, otherwise we'll just verify that the dstnid
                 * matches the NID for the NET that the dgram was posted */

                /* make sure their wildcard didn't match ours, that is unpossible */
                LASSERTF(connreq->gncr_dstnid != LNET_NID_ANY,
                         "dgram 0x%p from %s, connreq 0x%p; "
                         "wildcard matched wildcard \n", dgram,
                         libcfs_nid2str(connreq->gncr_srcnid), connreq);

                rc = kgnilnd_find_net(connreq->gncr_dstnid, &net);

                if (rc == -ESHUTDOWN) {
                        CERROR("Looking up network: device is in shutdown");
                        return rc;
                } else if (rc == -ENONET) {
                        CERROR("Connection data from %s: she sent "
                        "dst_nid %s, but net lookup failed on "
                        "dgram 0x%p@%s\n",
                        libcfs_nid2str(connreq->gncr_srcnid),
                        libcfs_nid2str(connreq->gncr_dstnid),
                        dgram, kgnilnd_dgram_type2str(dgram));
                        return rc;
                }

                if (net->gnn_ni->ni_nid != connreq->gncr_dstnid) {
                        CERROR("Bad connection data from %s: she sent "
                               "dst_nid %s, but I am %s with dgram 0x%p@%s\n",
                               libcfs_nid2str(connreq->gncr_srcnid),
                               libcfs_nid2str(connreq->gncr_dstnid),
                               libcfs_nid2str(net->gnn_ni->ni_nid),
                               dgram, kgnilnd_dgram_type2str(dgram));
                        kgnilnd_net_decref(net);
                        return -EBADSLT;
                }

                /* kgnilnd_find_net takes a ref on the net it finds, You need to decref it when not needed. */
                kgnilnd_net_decref(net);
        }

        if (connreq->gncr_version != GNILND_CONNREQ_VERSION) {
                CERROR("Unexpected version %d\n", connreq->gncr_version);
                return -EPROTO;
        }

        /* XXX Nic: TBD - checksum validation */
        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_CONNREQ_DROP)) {
                return -EBADF;
        }

        if (swab && connreq->gncr_type == GNILND_CONNREQ_REQ) {
                __u64 msg_addr = (__u64) connreq->gncr_gnparams.gnpr_smsg_attr.msg_buffer;

                __swab32s(&connreq->gncr_gnparams.gnpr_host_id);
                __swab32s(&connreq->gncr_gnparams.gnpr_cqid);
                __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.buff_size);
                __swab16s(&connreq->gncr_gnparams.gnpr_smsg_attr.mbox_maxcredit);
                __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.mbox_offset);
                __swab64s(&connreq->gncr_gnparams.gnpr_smsg_attr.mem_hndl.qword1);
                __swab64s(&connreq->gncr_gnparams.gnpr_smsg_attr.mem_hndl.qword2);
                __swab64s(&msg_addr);
                __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.msg_maxsize);
                __swab32s(&connreq->gncr_gnparams.gnpr_smsg_attr.msg_type);
        } else if (swab && connreq->gncr_type == GNILND_CONNREQ_NAK) {
                __swab32s(&connreq->gncr_nakdata.gnnd_errno);
        }

        /* since we use a unique instance ID for each network, the driver
         * will take care of dropping datagrams if we don't have that network.
         */

        /* few more idiot software or configuration checks */

        switch (connreq->gncr_type) {
        case GNILND_CONNREQ_REQ:
                /* wire up EP and SMSG block - this will check the incoming data
                 * and barf a NAK back if need to */
                rc = kgnilnd_set_conn_params(dgram);
                if (rc)
                        return rc;
                break;
        case GNILND_CONNREQ_NAK:
        case GNILND_CONNREQ_CLOSE:
                break;
        default:
                CERROR("unknown connreq packet type %d\n", connreq->gncr_type);
                return -EPROTO;
        }

        if (connreq->gncr_peerstamp == 0 || connreq->gncr_connstamp == 0) {
                CERROR("Recived bad timestamps peer %llu conn %llu\n",
                connreq->gncr_peerstamp, connreq->gncr_connstamp);
                return -EPROTO;
        }

        if (connreq->gncr_timeout < GNILND_MIN_TIMEOUT) {
                CERROR("Received timeout %d < MIN %d\n",
                       connreq->gncr_timeout, GNILND_MIN_TIMEOUT);
                return -EPROTO;
        }

        return 0;
}

int
kgnilnd_alloc_dgram(kgn_dgram_t **dgramp, kgn_device_t *dev, kgn_dgram_type_t type)
{
        kgn_dgram_t         *dgram;

        dgram = kmem_cache_zalloc(kgnilnd_data.kgn_dgram_cache, GFP_ATOMIC);
        if (dgram == NULL)
                return -ENOMEM;

        INIT_LIST_HEAD(&dgram->gndg_list);
        dgram->gndg_state = GNILND_DGRAM_USED;
        dgram->gndg_type = type;
        dgram->gndg_magic = GNILND_DGRAM_MAGIC;

        atomic_inc(&dev->gnd_ndgrams);

        CDEBUG(D_MALLOC|D_NETTRACE, "slab-alloced 'dgram': %lu at %p %s ndgrams"
                " %d\n",
                sizeof(*dgram), dgram, kgnilnd_dgram_type2str(dgram),
                atomic_read(&dev->gnd_ndgrams));

        *dgramp = dgram;
        return 0;
}

/* call this on a dgram that came back from kgnilnd_ep_postdata_test_by_id
 * returns < 0 on dgram to be cleaned up
 * > 0 on dgram that isn't done yet
 * == 0 on dgram that is ok and needs connreq processing */
int
kgnilnd_process_dgram(kgn_dgram_t *dgram, gni_post_state_t post_state)
{
        int rc = 0;

        switch (post_state) {
        case GNI_POST_COMPLETED:
                /* normal state for dgrams that need actual processing */
                /* GOTO to avoid processing dgram as canceled/done */
                GOTO(process_out, rc);

        case GNI_POST_PENDING:
                /* we should only see this if we are testing a WC dgram after a
                 * cancel - it means that it needs a full cycle of waiting
                 * for kgni_sm_task to finish moving it to TERMINATED */
                LASSERTF((dgram->gndg_type == GNILND_DGRAM_WC_REQ) &&
                          (dgram->gndg_state == GNILND_DGRAM_CANCELED),
                         "POST_PENDING dgram 0x%p with bad type %d(%s) or state %d(%s)\n",
                         dgram, dgram->gndg_type, kgnilnd_dgram_type2str(dgram),
                         dgram->gndg_state, kgnilnd_dgram_state2str(dgram));

                /* positive RC as this dgram isn't done yet */
                rc = EINPROGRESS;

                /* GOTO as this isn't done yet */
                GOTO(process_out, rc);
                break;

        case GNI_POST_TERMINATED:
                /* we've called cancel and it is done or remote guy called cancel and
                 * we've receved it on a WC dgram */
#if 0
                /* we are seeing weird terminations on non WC dgrams when we have not
                 * canceled them */

                LASSERTF(dgram->gndg_state == GNILND_DGRAM_CANCELED ||
                         dgram->gndg_conn_out.gncr_dstnid == LNET_NID_ANY,
                        "dgram 0x%p with bad state %d(%s) or dst nid %s\n",
                        dgram, dgram->gndg_state, kgnilnd_dgram_state2str(dgram),
                        libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid));
#endif

                CDEBUG(D_NETTRACE, "dgram 0x%p saw %s, cleaning it up\n", dgram,
                       dgram->gndg_state == GNILND_DGRAM_CANCELED ?  "canceled" : "terminated");

                rc =  -ECANCELED;
                break;

        case GNI_POST_TIMEOUT:
                /* we could have a timeout on a wildcard dgram too - if
                 * we got the incoming request but the remote node beefed
                 * before kgni could send the match data back. We'll just error
                 * on the active case and bail out gracefully */
                if (dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
                        CNETERR("hardware timeout for connect to "
                               "%s after %lu seconds. Is node dead?\n",
                               libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
                               cfs_duration_sec(jiffies - dgram->gndg_post_time));
                }

                rc = -ETIMEDOUT;
                break;

        default:
                CERROR("dgram 0x%p with bad post_state %d\n", dgram, post_state);
                LBUG();
        }

        /* now finish cleaning up a dgram that is canceled/terminated and needs to
         * go away */

        /* If this was actively canceled, drop the count now that we are processing */
        if (dgram->gndg_state == GNILND_DGRAM_CANCELED) {
                atomic_dec(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
                /* caller responsible for gndg_list removal */
        }

process_out:

        RETURN(rc);
}

/* needs dev->gnd_dgram_lock held */
void
kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram)
{
        gni_return_t            grc;

        if (dgram->gndg_state != GNILND_DGRAM_POSTED) {
                return;
        }

        LASSERTF(dgram->gndg_conn != NULL,
                 "dgram 0x%p with NULL conn\n", dgram);

        /* C.E - WC dgrams could be canceled immediately but
         * if there was some match pending, we need to call
         * test_by_id to clear it out. If that test returns
         * POST_PENDING, it is half done and needs to go along
         * with the rest of dgrams and go through a kgni_sm_task cycle
         * and deliver a GNI_POST_TERMINATED event before they
         * are actually canceled */

        dgram->gndg_state = GNILND_DGRAM_CANCELED;

        if (dgram->gndg_conn->gnc_state >= GNILND_CONN_ESTABLISHED) {
                /* we don't need to cancel_by_id if the datagram was good */
                return;
        }

        /* let folks know there are outstanding cancels */
        atomic_inc(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
        /* leave on nid list until cancel is done for debugging fun */
        grc = kgnilnd_ep_postdata_cancel_by_id(dgram->gndg_conn->gnc_ephandle, (__u64) dgram);

        /* if we don't get success here, we have hosed up the dgram tracking
         * code and need to bail out */
        LASSERTF(grc == GNI_RC_SUCCESS,
                 "postdata_cancel returned %d for conn 0x%p to %s\n",
                 grc, dgram->gndg_conn,
                 dgram->gndg_conn->gnc_peer ?
                  libcfs_nid2str(dgram->gndg_conn->gnc_peer->gnp_nid)
                  : "<?>");

        CDEBUG(D_NETTRACE,
                "canceled dgram 0x%p conn 0x%p ephandle 0x%p\n",
                dgram, dgram->gndg_conn,
                dgram->gndg_conn->gnc_ephandle);

        if (dgram->gndg_type == GNILND_DGRAM_WC_REQ) {
                gni_post_state_t         post_state;
                int                      rc = 0;
                __u32                    remote_addr = 0, remote_id = 0;

                grc = kgnilnd_ep_postdata_test_by_id(dgram->gndg_conn->gnc_ephandle,
                                                     (__u64)dgram, &post_state,
                                                     &remote_addr, &remote_id);

                LASSERTF(grc == GNI_RC_NO_MATCH || grc == GNI_RC_SUCCESS,
                         "bad grc %d from test_by_id on dgram 0x%p\n",
                        grc, dgram);

                /* if WC was canceled immediately, we get NO_MATCH, if needs to go
                 * through full cycle, we get SUCCESS and need to parse post_state */

                CDEBUG(D_NET, "grc %d dgram 0x%p type %s post_state %d "
                        "remote_addr %u remote_id %u\n", grc, dgram,
                        kgnilnd_dgram_type2str(dgram),
                        post_state, remote_addr, remote_id);

                if (grc == GNI_RC_NO_MATCH) {
                        /* she's gone, reduce count and move along */
                        dgram->gndg_state = GNILND_DGRAM_DONE;
                        atomic_dec(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);
                        RETURN_EXIT;
                }

                rc = kgnilnd_process_dgram(dgram, post_state);

                if (rc <= 0) {
                        /* if for some weird reason we get a valid dgram back, just mark as done
                         * so we can drop it and move along.
                         * C.E - if it was completed, we'll just release the conn/mbox
                         * back into the pool and it'll get reused. That said, we should only
                         * be canceling a WC dgram on stack rest or shutdown, so that is moot */
                        dgram->gndg_state = GNILND_DGRAM_DONE;
                        atomic_dec(&dgram->gndg_conn->gnc_device->gnd_canceled_dgrams);

                        /* caller context responsible for calling kgnilnd_release_dgram() */
                } else {
                        /* still pending, let it simmer until golden brown and delicious */
                }
        }

        /* for non WC dgrams, they are still on the nid list but marked canceled waiting
         * for kgni to return their ID to us via probe - that is when we'll complete their
         * cancel processing */
}

void
kgnilnd_cleanup_dgram(kgn_dgram_t *dgram)
{
        /* release the dgram ref on conn */
        if (dgram->gndg_conn) {
                kgnilnd_conn_decref(dgram->gndg_conn);
                dgram->gndg_conn = NULL;
        }
}

void
kgnilnd_free_dgram(kgn_device_t *dev, kgn_dgram_t *dgram)
{
        LASSERTF(dgram->gndg_state == GNILND_DGRAM_USED ||
                 dgram->gndg_state == GNILND_DGRAM_DONE,
                 "dgram 0x%p with bad state %s\n",
                 dgram, kgnilnd_dgram_state2str(dgram));

        /* bit of poisoning to help detect bad driver data */
        dgram->gndg_magic = 0x6f5a6b5f;
        atomic_dec(&dev->gnd_ndgrams);

        kmem_cache_free(kgnilnd_data.kgn_dgram_cache, dgram);
        CDEBUG(D_MALLOC|D_NETTRACE, "slab-freed 'dgram': %lu at %p %s"
               " ndgrams %d\n",
               sizeof(*dgram), dgram, kgnilnd_dgram_type2str(dgram),
               atomic_read(&dev->gnd_ndgrams));
}

int
kgnilnd_post_dgram(kgn_device_t *dev, lnet_nid_t dstnid, kgn_connreq_type_t type,
                   int data_rc)
{
        int              rc = 0;
        kgn_dgram_t     *dgram = NULL;
        kgn_dgram_t     *tmpdgram;
        kgn_dgram_type_t dgtype;
        gni_return_t     grc;
        __u64            srcnid;
        ENTRY;

        switch (type) {
        case GNILND_CONNREQ_REQ:
                if (dstnid == LNET_NID_ANY)
                        dgtype = GNILND_DGRAM_WC_REQ;
                else
                        dgtype = GNILND_DGRAM_REQ;
                break;
        case GNILND_CONNREQ_NAK:
                LASSERTF(dstnid != LNET_NID_ANY, "can't NAK to LNET_NID_ANY\n");
                dgtype = GNILND_DGRAM_NAK;
                break;
        default:
                CERROR("unknown connreq type %d\n", type);
                LBUG();
        }

        rc = kgnilnd_alloc_dgram(&dgram, dev, dgtype);
        if (rc < 0) {
                rc = -ENOMEM;
                GOTO(post_failed, rc);
        }

        rc = kgnilnd_create_conn(&dgram->gndg_conn, dev);
        if (rc) {
                GOTO(post_failed, rc);
        }

        if (dgram->gndg_type == GNILND_DGRAM_WC_REQ) {
                /* clear buffer for sanity on reuse of wildcard */
                memset(&dgram->gndg_conn_in, 0, sizeof(kgn_connreq_t));
        }

        if (dstnid == LNET_NID_ANY) {
                /* set here to reset any dgram re-use */
                dgram->gndg_conn->gnc_state = GNILND_CONN_LISTEN;
        } else {
                __u32            host_id;

                rc = kgnilnd_nid_to_nicaddrs(LNET_NIDADDR(dstnid), 1, &host_id);
                if (rc <= 0) {
                        rc = -ESRCH;
                        GOTO(post_failed, rc);
                }

                dgram->gndg_conn->gnc_state = GNILND_CONN_CONNECTING;

                /* don't need to serialize, there are no CQs for the dgram
                 * EP on the kgn_net_t */
                grc = kgnilnd_ep_bind(dgram->gndg_conn->gnc_ephandle, host_id, dev->gnd_id);

                if (grc != GNI_RC_SUCCESS) {
                        rc = -ECONNABORTED;
                        GOTO(post_failed, rc);
                }

        }

        /* If we are posting wildcards post using a net of 0, otherwise we'll use the
         * net of the destination node.
         */

        if (dstnid == LNET_NID_ANY) {
                srcnid = LNET_MKNID(LNET_MKNET(GNILND, 0), dev->gnd_nid);
        } else {
                srcnid = LNET_MKNID(LNET_NIDNET(dstnid), dev->gnd_nid);
        }

        rc = kgnilnd_pack_connreq(&dgram->gndg_conn_out, dgram->gndg_conn,
                                  srcnid, dstnid, type);
        if (rc) {
                GOTO(post_failed, rc);
        }

        if (type == GNILND_CONNREQ_NAK)
                dgram->gndg_conn_out.gncr_nakdata.gnnd_errno = data_rc;

        dgram->gndg_post_time = jiffies;

        /* XXX Nic: here is where we'd add in logical network multiplexing */

        CDEBUG(D_NETTRACE, "dgram 0x%p type %s %s->%s cdm %d\n",
               dgram, kgnilnd_dgram_type2str(dgram),
               libcfs_nid2str(srcnid),
               libcfs_nid2str(dstnid), dev->gnd_id);

        /* this allocates memory, can't hold locks across */
        grc = kgnilnd_ep_postdata_w_id(dgram->gndg_conn->gnc_ephandle,
                                   &dgram->gndg_conn_out, sizeof(kgn_connreq_t),
                                   &dgram->gndg_conn_in, sizeof(kgn_connreq_t),
                                   (__u64)dgram);

        if (grc != GNI_RC_SUCCESS) {
                CNETERR("dropping failed dgram post id 0x%p type %s"
                        " reqtype %s to %s: rc %d\n",
                        dgram, kgnilnd_dgram_type2str(dgram),
                        kgnilnd_connreq_type2str(&dgram->gndg_conn_out),
                        libcfs_nid2str(dstnid), grc);
                rc = (grc == GNI_RC_ERROR_NOMEM) ? -ENOMEM : -EBADR;
                GOTO(post_failed, rc);
        }

        /* we don't need to add earlier - if someone does del_peer during post,
         * that peer will get marked as unlinked and the callers wil take care of it.
         * The dgram code is largely kgn_peer_t ignorant, so at worst, we'll just drop
         * the completed dgram later when we cant find a peer to stuff it into */

        spin_lock(&dev->gnd_dgram_lock);

        /* make sure we are not double posting targeted dgrams
         * - we can multiple post WC dgrams to help with processing speed */
        if (dstnid != LNET_NID_ANY) {
                tmpdgram = kgnilnd_find_dgram_locked(dev, dstnid);

                LASSERTF(tmpdgram == NULL,
                        "dgram 0x%p->%s already posted\n",
                         dgram, libcfs_nid2str(dstnid));
        }

        /* unmunge dstnid to help processing code cope... */
        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PACK_DSTNID)) {
                dgram->gndg_conn_out.gncr_dstnid = dstnid;
        }

        list_add_tail(&dgram->gndg_list, kgnilnd_nid2dgramlist(dev, dstnid));
        dgram->gndg_state = GNILND_DGRAM_POSTED;
        spin_unlock(&dev->gnd_dgram_lock);

post_failed:
        if (rc < 0 && dgram != NULL) {
                kgnilnd_cleanup_dgram(dgram);
                kgnilnd_free_dgram(dev, dgram);
        }

        RETURN(rc);
}

/* The shutdown flag is set from the shutdown and stack reset threads. */
void
kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram, int shutdown)
{
        /* The conns of canceled active dgrams need to be put in purgatory so
         * we don't reuse the mailbox */
        if (unlikely(dgram->gndg_state == GNILND_DGRAM_CANCELED)) {
                kgn_peer_t *peer;
                kgn_conn_t *conn = dgram->gndg_conn;
                lnet_nid_t nid = dgram->gndg_conn_out.gncr_dstnid;

                dgram->gndg_state = GNILND_DGRAM_DONE;

                /* During shutdown we've already removed the peer so we don't
                 * need to add a peer. During stack reset we don't care about
                 * MDDs since they are all released. */
                if (!shutdown) {
                        write_lock(&kgnilnd_data.kgn_peer_conn_lock);
                        peer = kgnilnd_find_peer_locked(nid);

                        if (peer != NULL) {
                                CDEBUG(D_NET, "adding peer's conn with nid %s "
                                        "to purgatory\n", libcfs_nid2str(nid));
                                kgnilnd_conn_addref(conn);
                                conn->gnc_peer = peer;
                                kgnilnd_peer_addref(peer);
                                kgnilnd_admin_addref(conn->gnc_peer->gnp_dirty_eps);
                                conn->gnc_state = GNILND_CONN_CLOSED;
                                list_add_tail(&conn->gnc_list,
                                              &peer->gnp_conns);
                                kgnilnd_add_purgatory_locked(conn,
                                                             conn->gnc_peer);
                                kgnilnd_schedule_conn(conn);
                        }
                        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                }
        }

        spin_lock(&dev->gnd_dgram_lock);
        kgnilnd_cancel_dgram_locked(dgram);
        spin_unlock(&dev->gnd_dgram_lock);

        kgnilnd_cleanup_dgram(dgram);

        /* if the dgram is 'canceled' it needs to be wait until the event
         * comes up from kgni that tells us it is safe to release */
        if (dgram->gndg_state != GNILND_DGRAM_CANCELED) {
                dgram->gndg_state = GNILND_DGRAM_DONE;

                LASSERTF(list_empty(&dgram->gndg_list), "dgram 0x%p on list\n", dgram);

                /* if it is a wildcard and we are in an appropriate state, repost
                 * the wildcard */

                if ((dgram->gndg_type == GNILND_DGRAM_WC_REQ) &&
                    (!kgnilnd_data.kgn_wc_kill && !kgnilnd_data.kgn_in_reset)) {
                        int     rerc;

                        rerc = kgnilnd_post_dgram(dev, LNET_NID_ANY, GNILND_CONNREQ_REQ, 0);
                        if (rerc != 0) {
                                /* We failed to repost the WC dgram for some reason
                                 * mark it so the repost system attempts to repost */
                                kgnilnd_admin_addref(dev->gnd_nwcdgrams);
                        }
                }

                /* always free the old dgram */
                kgnilnd_free_dgram(dev, dgram);
        }
}


int
kgnilnd_probe_for_dgram(kgn_device_t *dev, kgn_dgram_t **dgramp)
{
        kgn_dgram_t             *dgram = NULL;
        gni_post_state_t         post_state;
        gni_return_t             grc;
        int                      rc = 0;
        __u64                    readyid;
        __u32                    remote_addr = 0, remote_id = 0;
        ENTRY;

        /* Probe with the lock held. That way if we get a dgram we dont have it canceled
         * between finding the ready dgram and grabbing the lock to remove it from the
         * list. Otherwise we could be left in an inconsistent state. We own the dgram
         * once its off the list so we don't need to worry about others changing it at
         * that point. */
        spin_lock(&dev->gnd_dgram_lock);
        grc = kgnilnd_postdata_probe_by_id(dev->gnd_handle, &readyid);
        if (grc != GNI_RC_SUCCESS) {
                spin_unlock(&dev->gnd_dgram_lock);
                /* return 0 to indicate nothing happened */
                RETURN(0);
        }

        CDEBUG(D_NET, "ready %#llx on device 0x%p\n",
                readyid, dev);

        dgram = (kgn_dgram_t *)readyid;

        LASSERTF(dgram->gndg_magic == GNILND_DGRAM_MAGIC,
                 "dgram 0x%p from id %#llx with bad magic %x\n",
                 dgram, readyid, dgram->gndg_magic);

        LASSERTF(dgram->gndg_state == GNILND_DGRAM_POSTED ||
                 dgram->gndg_state == GNILND_DGRAM_CANCELED,
                 "dgram 0x%p with bad state %s\n",
                 dgram, kgnilnd_dgram_state2str(dgram));

        LASSERTF(!list_empty(&dgram->gndg_list),
                 "dgram 0x%p with bad list state %s type %s\n",
                 dgram, kgnilnd_dgram_state2str(dgram),
                 kgnilnd_dgram_type2str(dgram));

        /* now we know that the datagram structure is ok, so pull off list */
        list_del_init(&dgram->gndg_list);

        /* while we have the gnn_dgram_lock and BEFORE we call test_by_id
         * change the state from POSTED to PROCESSING to ensure that
         * nobody cancels it after we've pulled it from the wire */
        if (dgram->gndg_state == GNILND_DGRAM_POSTED) {
                dgram->gndg_state = GNILND_DGRAM_PROCESSING;
        }

        LASSERTF(dgram->gndg_conn != NULL,
                "dgram 0x%p with NULL conn\n", dgram);

        grc = kgnilnd_ep_postdata_test_by_id(dgram->gndg_conn->gnc_ephandle,
                                             (__u64)dgram, &post_state,
                                             &remote_addr, &remote_id);

        /* we now "own" this datagram */
        spin_unlock(&dev->gnd_dgram_lock);

        LASSERTF(grc != GNI_RC_NO_MATCH, "kgni lied! probe_by_id told us that"
                 " id %llu was ready\n", readyid);

        CDEBUG(D_NET, "grc %d dgram 0x%p type %s post_state %d "
                "remote_addr %u remote_id %u\n", grc, dgram,
                kgnilnd_dgram_type2str(dgram),
                post_state, remote_addr, remote_id);

        if (unlikely(grc != GNI_RC_SUCCESS)) {
                CNETERR("getting data for dgram 0x%p->%s failed rc %d. Dropping it\n",
                        dgram, libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
                        grc);
                rc = -EINVAL;
                GOTO(probe_for_out, rc);
        }

        rc = kgnilnd_process_dgram(dgram, post_state);

        /* we should never get probe finding a dgram for us and then it
         * being a WC dgram that is still in the middle of processing */
        LASSERTF(rc <= 0, "bad rc %d from process_dgram 0x%p state %d\n",
                 rc, dgram, post_state);

        if (rc == 0) {
                /* dgram is good enough for the data to be used */
                dgram->gndg_state = GNILND_DGRAM_PROCESSING;
                /* fake rc to mark that we've done something */
                rc = 1;
        } else {
                /* let kgnilnd_release_dgram take care of canceled dgrams */
                if (dgram->gndg_state != GNILND_DGRAM_CANCELED) {
                        dgram->gndg_state = GNILND_DGRAM_DONE;
                }
        }

        *dgramp = dgram;
        RETURN(rc);

probe_for_out:

        kgnilnd_release_dgram(dev, dgram, 0);
        RETURN(rc);
}

int
kgnilnd_setup_wildcard_dgram(kgn_device_t *dev)
{
        /* if kgn_wildcard is zero, return error */
        int     rc = -ENOENT, i;
        ENTRY;

        for (i = 0; i < *kgnilnd_tunables.kgn_nwildcard; i++) {
                rc = kgnilnd_post_dgram(dev, LNET_NID_ANY, GNILND_CONNREQ_REQ, 0);
                if (rc < 0) {
                        CERROR("error %d: could not post wildcard datagram # %d\n",
                                rc, i);
                        rc = -EINVAL;
                        GOTO(failed, rc);
                }
        }

failed:
        RETURN(rc);
}

int
kgnilnd_cancel_net_dgrams(kgn_net_t *net)
{
        kgn_dgram_t *dg, *dgN;
        LIST_HEAD(zombies);
        int i;
        ENTRY;

        /* we want to cancel any outstanding dgrams - we don't want to rely
         * on del_peer_or_conn catching all of them. This helps protect us in cases
         * where we don't quite keep the peer->dgram mapping in sync due to some
         * race conditions */

        LASSERTF(net->gnn_shutdown || kgnilnd_data.kgn_in_reset,
                 "called with LND invalid state: net shutdown %d "
                 "in reset %d\n", net->gnn_shutdown,
                 kgnilnd_data.kgn_in_reset);

        spin_lock(&net->gnn_dev->gnd_dgram_lock);

        for (i = 0; i < *kgnilnd_tunables.kgn_peer_hash_size; i++) {
                list_for_each_entry_safe(dg, dgN, &net->gnn_dev->gnd_dgrams[i], gndg_list) {

                        /* skip nids not on our net or are wildcards */


                        if (dg->gndg_type == GNILND_DGRAM_WC_REQ ||
                                net->gnn_netnum != LNET_NETNUM(LNET_NIDNET(dg->gndg_conn_out.gncr_dstnid)))
                                continue;

                        kgnilnd_cancel_dgram_locked(dg);
                }
        }

        spin_unlock(&net->gnn_dev->gnd_dgram_lock);

        RETURN(0);
}

int
kgnilnd_cancel_wc_dgrams(kgn_device_t *dev)
{
        kgn_dgram_t *dg, *dgN;
        LIST_HEAD(zombies);
        ENTRY;

        /* Time to kill the outstanding WC's
         * WC's exist on net 0 only but match on any net...
         */

        LASSERTF(kgnilnd_data.kgn_in_reset || kgnilnd_data.kgn_wc_kill,
                "called with LND invalid state: WC shutdown %d "
                "in reset %d\n", kgnilnd_data.kgn_wc_kill,
                kgnilnd_data.kgn_in_reset);

        spin_lock(&dev->gnd_dgram_lock);

        do {
                dg = kgnilnd_find_dgram_locked(dev, LNET_NID_ANY);
                if (dg != NULL) {
                        LASSERTF(dg->gndg_type == GNILND_DGRAM_WC_REQ,
                                 "dgram 0x%p->%s with bad type %d (%s)\n",
                                dg, libcfs_nid2str(dg->gndg_conn_out.gncr_dstnid),
                                dg->gndg_type, kgnilnd_dgram_type2str(dg));

                        kgnilnd_cancel_dgram_locked(dg);

                        /* WC could be DONE already, check and if so add to list to be released */
                        if (dg->gndg_state == GNILND_DGRAM_DONE)
                                list_move_tail(&dg->gndg_list, &zombies);
                }
        } while (dg != NULL);

        spin_unlock(&dev->gnd_dgram_lock);

        list_for_each_entry_safe(dg, dgN, &zombies, gndg_list) {
                list_del_init(&dg->gndg_list);
                kgnilnd_release_dgram(dev, dg, 1);
        }
        RETURN(0);

}

int
kgnilnd_cancel_dgrams(kgn_device_t *dev)
{
        kgn_dgram_t *dg, *dgN;
        int i;
        ENTRY;

        /* Cancel any outstanding non wildcard datagrams regardless
         * of which net they are on as we are in base shutdown and
         * dont care about connecting anymore.
         */

        LASSERTF(kgnilnd_data.kgn_wc_kill == 1,"We didnt get called from base shutdown\n");

        spin_lock(&dev->gnd_dgram_lock);

        for (i = 0; i < (*kgnilnd_tunables.kgn_peer_hash_size -1); i++) {
                list_for_each_entry_safe(dg, dgN, &dev->gnd_dgrams[i], gndg_list) {
                        if (dg->gndg_type != GNILND_DGRAM_WC_REQ)
                                kgnilnd_cancel_dgram_locked(dg);
                }
        }

        spin_unlock(&dev->gnd_dgram_lock);

        RETURN(0);
}


void
kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev)
{
        int             i = 4;
        int             rc;
        gni_return_t    grc;
        __u64           readyid;
        kgn_dgram_t    *dgram;

        /* use do while to get at least one check run to allow
         * regression test for 762072 to hit bug if there */

        /* This function races with the dgram mover during shutdown so it is possible for
         * a dgram to be seen in kgnilnd_postdata_probe_wait_by_id but be handled in the
         * dgram mover thread instead of inside of this function.
         */

        /* This should only be called from within shutdown, baseshutdown, or stack reset.
         * there are no assertions here to verify since base_shutdown has nothing in it we can check
         * the net is gone by then.
         */

        do {
                i++;
                CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
                        "Waiting for %d canceled datagrams to clear on device %d\n",
                        atomic_read(&dev->gnd_canceled_dgrams), dev->gnd_id);

                /* check once a second */
                grc = kgnilnd_postdata_probe_wait_by_id(dev->gnd_handle,
                       250, &readyid);

                if (grc != GNI_RC_SUCCESS)
                        continue;

                CDEBUG(D_NET, "ready %#llx on device %d->0x%p\n",
                        readyid, dev->gnd_id, dev);

                rc = kgnilnd_probe_for_dgram(dev, &dgram);
                if (rc != 0) {
                        /* if we got a valid dgram or one that is now done, clean up */
                        kgnilnd_release_dgram(dev, dgram, 1);
                }
        } while (atomic_read(&dev->gnd_canceled_dgrams));
}

int
kgnilnd_start_connect(kgn_peer_t *peer)
{
        int              rc = 0;
        /* sync point for kgnilnd_del_peer_locked - do an early check to
         * catch the most common hits where del_peer is done by the
         * time we get here */
        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GNP_CONNECTING1)) {
                while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_GNP_CONNECTING1, 1)) {};
        }

        write_lock(&kgnilnd_data.kgn_peer_conn_lock);
        if (!kgnilnd_peer_active(peer) || peer->gnp_connecting != GNILND_PEER_CONNECT) {
                /* raced with peer getting unlinked */
                write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                rc = ESTALE;
                GOTO(out, rc);
        }
        peer->gnp_connecting = GNILND_PEER_POSTING;
        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);

        set_mb(peer->gnp_last_dgram_time, jiffies);
        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GNP_CONNECTING2)) {
                while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_GNP_CONNECTING2, 1)) {};
        }

        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_GNP_CONNECTING3)) {
                while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_GNP_CONNECTING3, 1)) {};
                rc = cfs_fail_val ? cfs_fail_val : -ENOMEM;
        } else {
                rc = kgnilnd_post_dgram(peer->gnp_net->gnn_dev,
                                        peer->gnp_nid, GNILND_CONNREQ_REQ, 0);
        }
        if (rc < 0) {
                set_mb(peer->gnp_last_dgram_errno, rc);
                GOTO(failed, rc);
        }

        /* while we're posting someone could have decided this peer/dgram needed to
         * die a quick death, so we check for state change and process accordingly */

        write_lock(&kgnilnd_data.kgn_peer_conn_lock);
        if (!kgnilnd_peer_active(peer) || peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH) {
                if (peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH) {
                        peer->gnp_connecting = GNILND_PEER_KILL;
                }
                write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                /* positive RC to avoid dgram cleanup - we'll have to
                 * wait for the kgni GNI_POST_TERMINATED event to
                 * finish cleaning up */
                rc = ESTALE;
                kgnilnd_find_and_cancel_dgram(peer->gnp_net->gnn_dev, peer->gnp_nid);
                GOTO(out, rc);
        }
        peer->gnp_connecting = GNILND_PEER_POSTED;
        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
        /* reaper thread will take care of any timeouts */
        CDEBUG(D_NET, "waiting for connect to finish to %s rc %d\n",
               libcfs_nid2str(peer->gnp_nid), rc);

        RETURN(rc);

failed:
        CDEBUG(D_NET, "connect to %s failed: rc %d \n",
               libcfs_nid2str(peer->gnp_nid), rc);
out:
        RETURN(rc);
}

int
kgnilnd_finish_connect(kgn_dgram_t *dgram)
{
        kgn_conn_t        *conn = dgram->gndg_conn;
        lnet_nid_t         her_nid = dgram->gndg_conn_in.gncr_srcnid;
        kgn_peer_t        *new_peer, *peer = NULL;
        kgn_tx_t          *tx;
        kgn_tx_t          *txn;
        kgn_mbox_info_t   *mbox;
        int                rc;
        int                nstale;

        /* try to find a peer that matches the nid we got in the connreq
         * kgnilnd_unpack_connreq makes sure that conn_in.gncr_srcnid is
         * HER and conn_out.gncr_srcnid is ME for both active and WC dgrams */

        /* assume this is a new peer  - it makes locking cleaner when it isn't */
        /* no holding kgn_net_rw_sem - already are at the kgnilnd_dgram_mover level */

        rc = kgnilnd_create_peer_safe(&new_peer, her_nid, NULL, GNILND_PEER_UP);
        if (rc != 0) {
                CERROR("Can't create peer for %s\n", libcfs_nid2str(her_nid));
                return rc;
        }

        write_lock(&kgnilnd_data.kgn_peer_conn_lock);

        /* this transfers ref from create_peer to the kgn_peer table */
        kgnilnd_add_peer_locked(her_nid, new_peer, &peer);

        /* if we found an existing peer, is it really ready for a new conn ? */
        if (peer != new_peer) {
                /* if this was an active connect attempt but we can't find a peer waiting for it
                 * we will dump in the trash */

                if (peer->gnp_connecting == GNILND_PEER_IDLE && dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
                        CDEBUG(D_NET, "dropping completed connreq for %s peer 0x%p->%s\n",
                               libcfs_nid2str(her_nid), peer, libcfs_nid2str(peer->gnp_nid));
                        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                        rc = ECANCELED;
                        GOTO(out, rc);
                }

                /* check to see if we can catch a connecting peer before it is
                 * removed from the connd_peers list - if not, we need to
                 * let the connreqs race and be handled by kgnilnd_conn_isdup_locked() */
                if (peer->gnp_connecting != GNILND_PEER_IDLE) {
                        spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
                        if (!list_empty(&peer->gnp_connd_list)) {
                                list_del_init(&peer->gnp_connd_list);
                                /* drop connd ref */
                                kgnilnd_peer_decref(peer);
                        }
                        spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
                        /* clear rc to make sure we don't have fake error */
                        rc = 0;
                }

                /* no matter what, we are no longer waiting to connect this peer now */
                peer->gnp_connecting = GNILND_PEER_IDLE;

                /* Refuse to duplicate an existing connection (both sides might try to
                 * connect at once).  NB we return success!  We _are_ connected so we
                 * _don't_ have any blocked txs to complete with failure. */
                rc = kgnilnd_conn_isdup_locked(peer, conn);
                if (rc != 0) {
                        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                        CDEBUG(D_NET, "Not creating duplicate connection to %s: %d\n",
                              libcfs_nid2str(her_nid), rc);
                        rc = EALREADY;
                        GOTO(out, rc);
                }
        }

        if (peer->gnp_state == GNILND_PEER_DOWN) {
                CNETERR("Received connection request from down nid %s\n",
                        libcfs_nid2str(her_nid));
        }

        peer->gnp_state = GNILND_PEER_UP;
        nstale = kgnilnd_close_stale_conns_locked(peer, conn);

        /* either way with peer (new or existing), we are ok with ref counts here as the
         * kgnilnd_add_peer_locked will use our ref on new_peer (from create_peer_safe) as the
         * ref for the peer table. */

        /* at this point, the connection request is a winner */

        /* mark 'DONE' to avoid cancel being called from release */
        dgram->gndg_state = GNILND_DGRAM_DONE;

        /* initialise timestamps before reaper looks at them */
        conn->gnc_last_rx = conn->gnc_last_rx_cq = jiffies;

        /* last_tx is initialized to jiffies - (keepalive*2) so that if the NOOP fails it will
         * immediatly send a NOOP in the reaper thread during the call to
         * kgnilnd_check_conn_timeouts_locked
         */
        conn->gnc_last_tx = jiffies - (cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout)) * 2);
        conn->gnc_state = GNILND_CONN_ESTABLISHED;

        /* save the dgram type used to establish this connection */
        conn->gnc_dgram_type = dgram->gndg_type;

        /* refs are not transferred from dgram to tables, so increment to
         * take ownership */
        kgnilnd_conn_addref(conn);
        kgnilnd_peer_addref(peer);
        conn->gnc_peer = peer;
        list_add_tail(&conn->gnc_list, &peer->gnp_conns);

        kgnilnd_conn_addref(conn);               /* +1 ref for conn table */
        list_add_tail(&conn->gnc_hashlist,
                      kgnilnd_cqid2connlist(conn->gnc_cqid));
        kgnilnd_data.kgn_conn_version++;

        /* Dont send NOOP if fail_loc is set
         */
        if (!CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP)) {
                tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, peer->gnp_net->gnn_ni->ni_nid);
                if (tx == NULL) {
                        CNETERR("can't get TX to initiate NOOP to %s\n",
                                libcfs_nid2str(peer->gnp_nid));
                } else {
                        kgnilnd_queue_tx(conn, tx);
                }
        }

        /* Schedule all packets blocking for a connection */
        list_for_each_entry_safe(tx, txn, &peer->gnp_tx_queue, tx_list) {
                /* lock held here is the peer_conn lock */
                kgnilnd_tx_del_state_locked(tx, peer, NULL, GNILND_TX_ALLOCD);
                kgnilnd_queue_tx(conn, tx);
        }

        /* If this is an active connection lets mark its timestamp on the MBoX */
        if (dgram->gndg_conn_out.gncr_dstnid != LNET_NID_ANY) {
                mbox = &conn->gnc_fma_blk->gnm_mbox_info[conn->gnc_mbox_id];
                /* conn->gnc_last_rx is jiffies it better exist as it was just set */
                mbox->mbx_release_purg_active_dgram = conn->gnc_last_rx;
        }

        /* Bug 765042: wake up scheduler for a race with finish_connect and
         * complete_conn_closed with a conn in purgatory
         * since we can't use CFS_RACE due to mutex_holds in kgnilnd_process_conns,
         * we just check for set and then clear */
        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_FINISH_PURG)) {
                cfs_fail_loc = 0x0;
                /* get scheduler thread moving again */
                kgnilnd_schedule_device(conn->gnc_device);
        }

        CDEBUG(D_NET, "New conn 0x%p->%s dev %d\n",
               conn, libcfs_nid2str(her_nid), conn->gnc_device->gnd_id);

        /* make sure we reset peer reconnect interval now that we have a good conn */
        kgnilnd_peer_alive(peer);
        peer->gnp_reconnect_interval = 0;

        /* clear the unlink attribute if we dont clear it kgnilnd_del_conn_or_peer will wait
         * on the atomic forever
         */
        if (peer->gnp_pending_unlink) {
                peer->gnp_pending_unlink = 0;
                kgnilnd_admin_decref(kgnilnd_data.kgn_npending_unlink);
                CDEBUG(D_NET, "Clearing peer unlink %p\n",peer);
        }

        /* add ref to make it hang around until after we drop the lock */
        kgnilnd_conn_addref(conn);

        /* Once the peer_conn lock is dropped, the conn could actually move into
         * CLOSING->CLOSED->DONE in the scheduler thread, so hold the
         * lock until we are really done */
        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);

        /* Notify LNET that we now have a working connection to this peer.
         * This is a Cray extension to the "standard" LND behavior.
         */
        lnet_notify(peer->gnp_net->gnn_ni, peer->gnp_nid, true, true,
                    ktime_get_seconds());

        /* drop our 'hold' ref */
        kgnilnd_conn_decref(conn);

out:
        RETURN(rc);
}

void
kgnilnd_send_nak(kgn_device_t *dev, lnet_nid_t dst_nid, int error)
{
        int              rc = 0;
        ENTRY;

        LASSERTF(dst_nid != LNET_NID_ANY, "bad dst_nid %s\n", libcfs_nid2str(dst_nid));

        CDEBUG(D_NET, "NAK to %s errno %d\n", libcfs_nid2str(dst_nid), error);

        rc = kgnilnd_post_dgram(dev, dst_nid, GNILND_CONNREQ_NAK, error);

        if (rc < 0) {
                CDEBUG(D_NET, "NAK to %s failed: rc %d \n", libcfs_nid2str(dst_nid), rc);
        }
        EXIT;
}

int
kgnilnd_process_nak(kgn_dgram_t *dgram)
{
        kgn_connreq_t     *connreq = &dgram->gndg_conn_in;
        lnet_nid_t         src_nid = connreq->gncr_srcnid;
        int                errno = connreq->gncr_nakdata.gnnd_errno;
        kgn_peer_t        *peer;
        int                rc = 0;

        write_lock(&kgnilnd_data.kgn_peer_conn_lock);

        peer = kgnilnd_find_peer_locked(src_nid);
        if (peer == NULL) {
                /* we likely dropped him from bad data when we processed
                 * the original REQ */
                write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                return -EBADSLT;
        }

        /* need to check peerstamp/connstamp against the ones we find
         * to make sure we don't close new (and good?) conns that we
         * formed after this connreq failed */
        if (peer->gnp_connecting == GNILND_PEER_IDLE) {
                kgn_conn_t        conn;

                if (list_empty(&peer->gnp_conns)) {
                        /* assume already procced datagram and it barfed up
                         * on this side too */
                        CDEBUG(D_NET, "dropping NAK from %s; "
                               "peer %s is already not connected\n",
                                libcfs_nid2str(connreq->gncr_srcnid),
                                libcfs_nid2str(connreq->gncr_dstnid));
                        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                        return 0;
                }

                /* stub up a connection with the connreq XXX_stamps to allow
                 * use to use close_stale_conns_locked */
                conn.gnc_peerstamp = connreq->gncr_peerstamp;
                conn.gnc_my_connstamp = connreq->gncr_connstamp;
                conn.gnc_peer_connstamp = connreq->gncr_connstamp;
                conn.gnc_device = peer->gnp_net->gnn_dev;

                rc = kgnilnd_close_stale_conns_locked(peer, &conn);

                LCONSOLE_INFO("Received NAK from %s for %s errno %d; "
                        "closed %d connections\n",
                        libcfs_nid2str(connreq->gncr_srcnid),
                        libcfs_nid2str(connreq->gncr_dstnid), errno, rc);
        } else {
                spin_lock(&dgram->gndg_conn->gnc_device->gnd_connd_lock);

                if (list_empty(&peer->gnp_connd_list)) {
                        /* if peer isn't on waiting list, try to find one to nuke */
                        rc = kgnilnd_find_and_cancel_dgram(peer->gnp_net->gnn_dev,
                                                           peer->gnp_nid);

                        if (rc) {
                                LCONSOLE_INFO("Received NAK from %s for %s errno %d; "
                                        "canceled pending connect request\n",
                                        libcfs_nid2str(connreq->gncr_srcnid),
                                        libcfs_nid2str(connreq->gncr_dstnid), errno);
                        }

                        /* if we can't find a waiting dgram, we just drop the nak - the conn
                         * connect must have failed (didn't find conn above and clear connecting
                         * -- so nothing to do besides drop */
                } else {
                        /* peer is on list, meaning it is a new connect attempt from the one
                         * we started that generated the NAK - so just drop NAK */

                        /* use negative to prevent error message */
                        rc = -EAGAIN;
                }
                spin_unlock(&dgram->gndg_conn->gnc_device->gnd_connd_lock);
        }

        /* success! we found a peer and at least marked pending_nak */
        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);

        return rc;
}

int
kgnilnd_process_connreq(kgn_dgram_t *dgram, int *needs_nak)
{
        int                      rc;

        rc = kgnilnd_unpack_connreq(dgram);
        if (rc < 0) {
                if (rc != -EBADF) {
                        /* only NAK if we have good srcnid to use */
                        *needs_nak = 1;
                }
                goto connreq_out;
        }

        switch (dgram->gndg_conn_in.gncr_type) {
        case GNILND_CONNREQ_REQ:
                /* wire up peer & conn, send queued TX */
                rc = kgnilnd_finish_connect(dgram);

                /* don't nak when the nid is hosed */
                if ((rc < 0)) {
                        *needs_nak = 1;
                }

                break;
        case GNILND_CONNREQ_NAK:
                rc = kgnilnd_process_nak(dgram);
                /* return early to prevent reconnect bump */
                return rc;
        default:
                CERROR("unexpected connreq type %s (%d) from %s\n",
                        kgnilnd_connreq_type2str(&dgram->gndg_conn_in),
                        dgram->gndg_conn_in.gncr_type,
                        libcfs_nid2str(dgram->gndg_conn_in.gncr_srcnid));
                rc = -EINVAL;
                *needs_nak = 1;
                break;
        }

connreq_out:
        RETURN(rc);
}

int
kgnilnd_probe_and_process_dgram(kgn_device_t *dev)
{
        int                      rc;
        int                      needs_nak = 0;
        lnet_nid_t               nak_dstnid = LNET_NID_ANY;
        lnet_nid_t               orig_dstnid;
        kgn_dgram_t             *dgram = NULL;
        kgn_peer_t              *peer;
        ENTRY;

        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PAUSE_DGRAM_COMP)) {
                rc = 0;
        } else {
                rc = kgnilnd_probe_for_dgram(dev, &dgram);
        }

        if (rc == 0) {
                RETURN(0);
        } else if (rc < 0) {
                GOTO(inform_peer, rc);
        } else {
                /* rc > 1 means it did something, reset for this func  */
                rc = 0;
        }

        switch (dgram->gndg_type) {
        case GNILND_DGRAM_WC_REQ:
        case GNILND_DGRAM_REQ:
                rc = kgnilnd_process_connreq(dgram, &needs_nak);
                break;
        case GNILND_DGRAM_NAK:
                CDEBUG(D_NETTRACE, "NAK to %s done\n",
                        libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid));
                break;
        default:
                CERROR("unknown datagram type %s (%d)\n",
                       kgnilnd_dgram_type2str(dgram), dgram->gndg_type);
                break;
        }

        /* stash data to use after releasing current datagram */
        /* don't stash net - we are operating on a net already,
         * so the lock on rw_net_lock is sufficient */

        nak_dstnid = dgram->gndg_conn_in.gncr_srcnid;

inform_peer:
        LASSERTF(dgram != NULL, "dgram 0x%p rc %d needs_nak %d\n", dgram, rc, needs_nak);

        orig_dstnid = dgram->gndg_conn_out.gncr_dstnid;

        kgnilnd_release_dgram(dev, dgram, 0);

        CDEBUG(D_NET, "cleaning up dgram to %s, rc %d\n",
               libcfs_nid2str(orig_dstnid), rc);

        /* if this was a WC_REQ that matched an existing peer, it'll get marked done
         * in kgnilnd_finish_connect - if errors are from before we get to there,
         * we just drop as it is a WC_REQ - the peer CAN'T be waiting for it */
        if ((orig_dstnid != LNET_NID_ANY) && (rc < 0)) {
                /* if we have a negative rc, we want to find a peer to inform about
                 * the bad connection attempt. Sorry buddy, better luck next time! */

                write_lock(&kgnilnd_data.kgn_peer_conn_lock);
                peer = kgnilnd_find_peer_locked(orig_dstnid);

                if (peer != NULL) {
                        /* add ref to make sure he stays around past the possible unlink
                         * so we can tell LNet about him */
                        kgnilnd_peer_addref(peer);

                        /* if he still cares about the outstanding connect */
                        if (peer->gnp_connecting >= GNILND_PEER_CONNECT) {
                                /* check if he is on the connd list and remove.. */
                                spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
                                if (!list_empty(&peer->gnp_connd_list)) {
                                        list_del_init(&peer->gnp_connd_list);
                                        /* drop connd ref */
                                        kgnilnd_peer_decref(peer);
                                }
                                spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);

                                /* clear gnp_connecting so we don't have a non-connecting peer
                                 * on gnd_connd_list */
                                peer->gnp_connecting = GNILND_PEER_IDLE;

                                set_mb(peer->gnp_last_dgram_errno, rc);

                                kgnilnd_peer_increase_reconnect_locked(peer);
                        }
                }
                write_unlock(&kgnilnd_data.kgn_peer_conn_lock);

                /* now that we are outside the lock, tell Mommy */
                if (peer != NULL) {
                        kgnilnd_peer_notify(peer, rc, 0);
                        kgnilnd_peer_decref(peer);
                }
        }

        if (needs_nak) {
                kgnilnd_send_nak(dev, nak_dstnid, rc);
        }

        RETURN(1);
}

void
kgnilnd_reaper_dgram_check(kgn_device_t *dev)
{
        kgn_dgram_t    *dgram, *tmp;
        int             i;

        spin_lock(&dev->gnd_dgram_lock);

        for (i = 0; i < (*kgnilnd_tunables.kgn_peer_hash_size - 1); i++) {
                list_for_each_entry_safe(dgram, tmp, &dev->gnd_dgrams[i], gndg_list) {
                        unsigned long            now = jiffies;
                        unsigned long            timeout;

                        /* don't timeout stuff if the network is mucked or shutting down */
                        if (kgnilnd_check_hw_quiesce()) {
                                break;
                        }

                        if ((dgram->gndg_state != GNILND_DGRAM_POSTED) ||
                            (dgram->gndg_type == GNILND_DGRAM_WC_REQ)) {
                                continue;
                        }
                        CDEBUG(D_NETTRACE, "checking dgram 0x%p type %s "
                                "state %s conn 0x%p to %s age %lus\n",
                                dgram, kgnilnd_dgram_type2str(dgram),
                                kgnilnd_dgram_state2str(dgram), dgram->gndg_conn,
                                libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
                                cfs_duration_sec(now - dgram->gndg_post_time));

                        timeout = cfs_time_seconds(*kgnilnd_tunables.kgn_timeout);

                        if (time_before(now, (dgram->gndg_post_time + timeout)))
                                continue;

                        CNETERR("%s datagram to %s timed out @ %lus dgram "
                                "0x%p state %s conn 0x%p\n",
                                kgnilnd_dgram_type2str(dgram),
                                libcfs_nid2str(dgram->gndg_conn_out.gncr_dstnid),
                                cfs_duration_sec(now - dgram->gndg_post_time),
                                dgram, kgnilnd_dgram_state2str(dgram),
                                dgram->gndg_conn);

                        kgnilnd_cancel_dgram_locked(dgram);
                }
        }
        spin_unlock(&dev->gnd_dgram_lock);
}


/* use a thread for the possibly long-blocking wait_by_id to prevent
 * stalling the global workqueues */
int
kgnilnd_dgram_waitq(void *arg)
{
        kgn_device_t     *dev = (kgn_device_t *) arg;
        char              name[16];
        gni_return_t      grc;
        __u64             readyid;
        DEFINE_WAIT(mover_done);

        snprintf(name, sizeof(name), "kgnilnd_dgn_%02d", dev->gnd_id);

        /* all gnilnd threads need to run fairly urgently */
        set_user_nice(current, *kgnilnd_tunables.kgn_nice);

        /* we dont shut down until the device shuts down ... */
        while (!kgnilnd_data.kgn_shutdown) {
                /* to quiesce or to not quiesce, that is the question */
                if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
                        KGNILND_SPIN_QUIESCE;
                }

                while (CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_PAUSE_DGRAM_COMP, 1)) {}

                /* check once a second */
                grc = kgnilnd_postdata_probe_wait_by_id(dev->gnd_handle,
                                                       1000, &readyid);

                if (grc == GNI_RC_SUCCESS) {
                        CDEBUG(D_INFO, "waking up dgram mover thread\n");
                        kgnilnd_schedule_dgram(dev);

                        /* wait for dgram thread to ping us before spinning again */
                        prepare_to_wait(&dev->gnd_dgping_waitq, &mover_done,
                                        TASK_INTERRUPTIBLE);

                        /* don't sleep if we need to quiesce */
                        if (likely(!kgnilnd_data.kgn_quiesce_trigger)) {
                                schedule();
                        }
                        finish_wait(&dev->gnd_dgping_waitq, &mover_done);
                }
        }

        kgnilnd_thread_fini();
        return 0;
}

int
kgnilnd_start_outbound_dgrams(kgn_device_t *dev, unsigned long deadline)
{
        int                      did_something = 0, rc;
        kgn_peer_t              *peer = NULL;

        spin_lock(&dev->gnd_connd_lock);

        /* Active connect - we added this in kgnilnd_launch_tx */
        while (!list_empty(&dev->gnd_connd_peers) && time_before(jiffies, deadline)) {
                peer = list_first_entry(&dev->gnd_connd_peers,
                                        kgn_peer_t, gnp_connd_list);

                /* ref for connd removed in if/else below */
               list_del_init(&peer->gnp_connd_list);

                /* gnp_connecting and membership on gnd_connd_peers should be
                 * done coherently to avoid double adding, etc */
                /* don't need kgnilnd_data.kgn_peer_conn_lock here as that is only needed
                 * to get the peer to gnp_connecting in the first place. We just need to
                 * rely on gnd_connd_lock to serialize someone pulling him from the list
                 * BEFORE clearing gnp_connecting */
                LASSERTF(peer->gnp_connecting != GNILND_PEER_IDLE, "peer 0x%p->%s not connecting\n",
                         peer, libcfs_nid2str(peer->gnp_nid));

                spin_unlock(&dev->gnd_connd_lock);

                CDEBUG(D_NET, "processing connect to %s\n",
                       libcfs_nid2str(peer->gnp_nid));

                did_something += 1;
                rc = kgnilnd_start_connect(peer);

                if (likely(rc >= 0)) {
                        /* 0 on success, positive on 'just drop peer' errors */
                        kgnilnd_peer_decref(peer);
                } else if (rc == -ENOMEM) {
                        /* if we are out of wildcards, add back to
                         * connd_list - then break out and we'll try later
                         * if other errors, we'll bail & cancel pending tx */
                        write_lock(&kgnilnd_data.kgn_peer_conn_lock);
                        if (peer->gnp_connecting == GNILND_PEER_POSTING) {
                                peer->gnp_connecting = GNILND_PEER_CONNECT;
                                spin_lock(&dev->gnd_connd_lock);
                                list_add_tail(&peer->gnp_connd_list,
                                              &dev->gnd_connd_peers);
                        } else {
                                /* connecting changed while we were posting */

                                LASSERTF(peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH, "Peer is in invalid"
                                        " state 0x%p->%s, connecting %d\n",
                                        peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
                                peer->gnp_connecting = GNILND_PEER_KILL;
                                spin_lock(&dev->gnd_connd_lock);
                                /* remove the peer ref frrom the cond list */
                                kgnilnd_peer_decref(peer);
                                /* let the system handle itself */
                        }
                        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                        /* the datagrams are a global pool,
                         * so break out of trying and hope some free
                         * up soon */
                        did_something -= 1;
                        break;
                } else {
                        /* something bad happened, you lose */
                        CNETERR("could not start connecting to %s "
                                "rc %d: Will retry until TX timeout\n",
                               libcfs_nid2str(peer->gnp_nid), rc);
                        /* It didnt post so just set connecting back to zero now.
                         * The reaper will reattempt the connection if it needs too.
                         * If the peer needs death set it so the reaper will cleanup.
                         */
                        write_lock(&kgnilnd_data.kgn_peer_conn_lock);
                        if (peer->gnp_connecting == GNILND_PEER_POSTING) {
                                peer->gnp_connecting = GNILND_PEER_IDLE;
                                kgnilnd_peer_increase_reconnect_locked(peer);
                        } else {
                                LASSERTF(peer->gnp_connecting == GNILND_PEER_NEEDS_DEATH, "Peer is in invalid"
                                        " state 0x%p->%s, connecting %d\n",
                                        peer, libcfs_nid2str(peer->gnp_nid), peer->gnp_connecting);
                                peer->gnp_connecting = GNILND_PEER_KILL;
                        }
                        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);

                        /* hold onto ref until we are really done - if it was
                         * unlinked this could result in a destroy */
                        kgnilnd_peer_decref(peer);
                }
                spin_lock(&dev->gnd_connd_lock);
        }

        spin_unlock(&dev->gnd_connd_lock);
        RETURN(did_something);
}

int
kgnilnd_repost_wc_dgrams(kgn_device_t *dev)
{
        int did_something = 0, to_repost, i;
        to_repost = atomic_read(&dev->gnd_nwcdgrams);
        ENTRY;

        for (i = 0; i < to_repost; ++i) {
                int     rerc;
                rerc = kgnilnd_post_dgram(dev, LNET_NID_ANY, GNILND_CONNREQ_REQ, 0);
                if (rerc == 0) {
                        kgnilnd_admin_decref(dev->gnd_nwcdgrams);
                        did_something += 1;
                } else {
                        CDEBUG(D_NETERROR, "error %d: dev %d could not post wildcard datagram\n",
                                rerc, dev->gnd_id);
                        break;
                }
        }

        RETURN(did_something);
}

static void
kgnilnd_dgram_poke_with_stick(unsigned long arg)
{
        int             dev_id = arg;
        kgn_device_t    *dev = &kgnilnd_data.kgn_devices[dev_id];

        wake_up(&dev->gnd_dgram_waitq);
}

/* use single thread for dgrams - should be sufficient for performance */
int
kgnilnd_dgram_mover(void *arg)
{
        kgn_device_t            *dev = (kgn_device_t *)arg;
        char                     name[16];
        int                      rc, did_something;
        unsigned long            next_purge_check = jiffies - 1;
        unsigned long            timeout;
        struct timer_list        timer;
        unsigned long            deadline = 0;
        DEFINE_WAIT(wait);

        snprintf(name, sizeof(name), "kgnilnd_dg_%02d", dev->gnd_id);

        /* all gnilnd threads need to run fairly urgently */
        set_user_nice(current, *kgnilnd_tunables.kgn_nice);

        /* we are ok not locking for these variables as the dgram waitq threads
         * will block both due to tying up net (kgn_shutdown) and the completion
         * event for the dgram_waitq (kgn_quiesce_trigger) */
        deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_dgram_timeout);
        while (!kgnilnd_data.kgn_shutdown) {
                /* Safe: kgn_shutdown only set when quiescent */

                /* race with stack reset - we want to hold off seeing any new incoming dgrams
                 * so we can force a dirty WC dgram for Bug 762072 - put right before
                 * quiesce check so that it'll go right into that and not do any
                 * dgram mucking */
                CFS_RACE(CFS_FAIL_GNI_WC_DGRAM_FREE);

                /* to quiesce or to not quiesce, that is the question */
                if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
                        KGNILND_SPIN_QUIESCE;
                }
                did_something = 0;

                CFS_RACE(CFS_FAIL_GNI_QUIESCE_RACE);

                /* process any newly completed dgrams */
                down_read(&kgnilnd_data.kgn_net_rw_sem);

                rc = kgnilnd_probe_and_process_dgram(dev);
                if (rc > 0) {
                        did_something += rc;
                }

                up_read(&kgnilnd_data.kgn_net_rw_sem);

                CFS_FAIL_TIMEOUT(CFS_FAIL_GNI_DGRAM_DEADLINE,
                        (*kgnilnd_tunables.kgn_dgram_timeout + 1));
                /* start new outbound dgrams */
                did_something += kgnilnd_start_outbound_dgrams(dev, deadline);

                /* find dead dgrams */
                if (time_after_eq(jiffies, next_purge_check)) {
                        /* these don't need to be checked that often */
                        kgnilnd_reaper_dgram_check(dev);

                        next_purge_check = (long) jiffies +
                                      cfs_time_seconds(kgnilnd_data.kgn_new_min_timeout / 4);
                }

                did_something += kgnilnd_repost_wc_dgrams(dev);

                /* careful with the jiffy wrap... */
                timeout = (long)(next_purge_check - jiffies);

                CDEBUG(D_INFO, "did %d timeout %lu next %lu jiffies %lu\n",
                       did_something, timeout, next_purge_check, jiffies);

                if ((did_something || timeout <= 0) && time_before(jiffies, deadline)) {
                        did_something = 0;
                        continue;
                }

                prepare_to_wait(&dev->gnd_dgram_waitq, &wait, TASK_INTERRUPTIBLE);

                setup_timer(&timer, kgnilnd_dgram_poke_with_stick, dev->gnd_id);
                mod_timer(&timer, (long) jiffies + timeout);

                /* last second chance for others to poke us */
                did_something += xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_IDLE);

                /* check flag variables before committing even if we
                 * did something; if we are after the deadline call
                 * schedule */
                if ((!did_something || time_after(jiffies, deadline)) &&
                    !kgnilnd_data.kgn_shutdown &&
                    !kgnilnd_data.kgn_quiesce_trigger) {
                        CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
                               timeout, cfs_duration_sec(timeout));
                        wake_up_all(&dev->gnd_dgping_waitq);
                        schedule();
                        CDEBUG(D_INFO, "awake after schedule\n");
                        deadline = jiffies + cfs_time_seconds(*kgnilnd_tunables.kgn_dgram_timeout);
                }

                del_singleshot_timer_sync(&timer);
                finish_wait(&dev->gnd_dgram_waitq, &wait);
        }

        kgnilnd_thread_fini();
        return 0;
}