LU-6261 gnilnd: Thread-safe optimizations.

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

/*
 * Copyright (C) 2004 Cluster File Systems, Inc.
 *
 * Copyright (C) 2009-2012 Cray, Inc.
 *
 *   Derived from work by Eric Barton <eric@bartonsoftware.com>
 *   Author: James Shimek <jshimek@cray.com>
 *   Author: Nic Henke <nic@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 <asm/page.h>
#include <linux/nmi.h>
#include "gnilnd.h"

/* this is useful when needed to debug wire corruption. */
static void
kgnilnd_dump_blob(int level, char *prefix, void *buf, int len) {
        __u64 *ptr;

        ptr = (__u64 *) buf;

        while (len > 0) {
                if (len >= 32) {
                        CDEBUG(level,
                               "%s 0x%p: 0x%16.16llx 0x%16.16llx 0x%16.16llx 0x%16.16llx\n",
                               prefix, ptr, *(ptr), *(ptr + 1), *(ptr + 2), *(ptr + 3));
                        ptr += 4;
                        len -= 32;
                } else if (len >= 16) {
                        CDEBUG(level,
                               "%s 0x%p: 0x%16.16llx 0x%16.16llx\n",
                               prefix, ptr, *(ptr), *(ptr + 1));
                        ptr += 2;
                        len -= 16;
                } else {
                        CDEBUG(level, "%s 0x%p: 0x%16.16llx\n",
                               prefix, ptr, *(ptr));
                        ptr++;
                        len -= 8;
                }
        }
}

static void
kgnilnd_dump_msg(int mask, kgn_msg_t *msg)
{
        CDEBUG(mask, "0x%8.8x 0x%4.4x 0x%4.4x 0x%16.16llx"
                " 0x%16.16llx 0x%8.8x 0x%4.4x 0x%4.4x 0x%8.8x\n",
                msg->gnm_magic, msg->gnm_version,
                msg->gnm_type, msg->gnm_srcnid,
                msg->gnm_connstamp, msg->gnm_seq,
                msg->gnm_cksum, msg->gnm_payload_cksum,
                msg->gnm_payload_len);
}

void
kgnilnd_schedule_device(kgn_device_t *dev)
{
        short         already_live = 0;

        /* we'll only want to wake if the scheduler thread
         * has come around and set ready to zero */
        already_live = cmpxchg(&dev->gnd_ready, GNILND_DEV_IDLE, GNILND_DEV_IRQ);

        if (!already_live) {
                wake_up_all(&dev->gnd_waitq);
        }
        return;
}

void kgnilnd_schedule_device_timer(unsigned long arg)
{
        kgn_device_t *dev = (kgn_device_t *) arg;

        kgnilnd_schedule_device(dev);
}

void
kgnilnd_device_callback(__u32 devid, __u64 arg)
{
        kgn_device_t *dev;
        int           index = (int) arg;

        if (index >= kgnilnd_data.kgn_ndevs) {
                /* use _EMERG instead of an LBUG to prevent LBUG'ing in
                 * interrupt context. */
                LCONSOLE_EMERG("callback for unknown device %d->%d\n",
                                devid, index);
                return;
        }

        dev = &kgnilnd_data.kgn_devices[index];
        /* just basic sanity */
        if (dev->gnd_id == devid) {
                kgnilnd_schedule_device(dev);
        } else {
                LCONSOLE_EMERG("callback for bad device %d devid %d\n",
                                dev->gnd_id, devid);
        }
}

/* sched_intent values:
 * < 0 : do not reschedule under any circumstances
 * == 0: reschedule if someone marked him WANTS_SCHED
 * > 0 : force a reschedule */
/* Return code 0 means it did not schedule the conn, 1
 * means it successfully scheduled the conn.
 */

int
kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent)
{
        int     conn_sched;

        /* move back to IDLE but save previous state.
         * if we see WANTS_SCHED, we'll call kgnilnd_schedule_conn and
         * let the xchg there handle any racing callers to get it
         * onto gnd_ready_conns */

        conn_sched = xchg(&conn->gnc_scheduled, GNILND_CONN_IDLE);
        LASSERTF(conn_sched == GNILND_CONN_WANTS_SCHED ||
                 conn_sched == GNILND_CONN_PROCESS,
                 "conn %p after process in bad state: %d\n",
                 conn, conn_sched);

        if (sched_intent >= 0) {
                if ((sched_intent > 0 || (conn_sched == GNILND_CONN_WANTS_SCHED))) {
                        return kgnilnd_schedule_conn_refheld(conn, 1);
                }
        }
        return 0;
}

/* Return of 0 for conn not scheduled, 1 returned if conn was scheduled or marked
 * as scheduled */

int
_kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld)
{
        kgn_device_t        *dev = conn->gnc_device;
        int                  sched;
        int                  rc;

        sched = xchg(&conn->gnc_scheduled, GNILND_CONN_WANTS_SCHED);
        /* we only care about the last person who marked want_sched since they
         * are most likely the culprit
         */
        memcpy(conn->gnc_sched_caller, caller, sizeof(conn->gnc_sched_caller));
        conn->gnc_sched_line = line;
        /* if we are IDLE, add to list - only one guy sees IDLE and "wins"
         * the chance to put it onto gnd_ready_conns.
         * otherwise, leave marked as WANTS_SCHED and the thread that "owns"
         *  the conn in process_conns will take care of moving it back to
         *  SCHED when it is done processing */

        if (sched == GNILND_CONN_IDLE) {
                /* if the conn is already scheduled, we've already requested
                 * the scheduler thread wakeup */
                if (!refheld) {
                        /* Add a reference to the conn if we are not holding a reference
                         * already from the exisiting scheduler. We now use the same
                         * reference if we need to reschedule a conn while in a scheduler
                         * thread.
                         */
                        kgnilnd_conn_addref(conn);
                }
                LASSERTF(list_empty(&conn->gnc_schedlist), "conn %p already sched state %d\n",
                         conn, sched);

                CDEBUG(D_INFO, "scheduling conn 0x%p caller %s:%d\n", conn, caller, line);

                spin_lock(&dev->gnd_lock);
                list_add_tail(&conn->gnc_schedlist, &dev->gnd_ready_conns);
                spin_unlock(&dev->gnd_lock);
                set_mb(conn->gnc_last_sched_ask, jiffies);
                rc = 1;
        } else {
                CDEBUG(D_INFO, "not scheduling conn 0x%p: %d caller %s:%d\n", conn, sched, caller, line);
                rc = 0;
        }

        /* make sure thread(s) going to process conns - but let it make
         * separate decision from conn schedule */
        kgnilnd_schedule_device(dev);
        return rc;
}

void
kgnilnd_schedule_dgram(kgn_device_t *dev)
{
        int                  wake;

        wake = xchg(&dev->gnd_dgram_ready, GNILND_DGRAM_SCHED);
        if (wake != GNILND_DGRAM_SCHED)  {
                wake_up(&dev->gnd_dgram_waitq);
        } else {
                CDEBUG(D_NETTRACE, "not waking: %d\n", wake);
        }
}

void
kgnilnd_free_tx(kgn_tx_t *tx)
{
        /* taken from kgnilnd_tx_add_state_locked */

        LASSERTF((tx->tx_list_p == NULL &&
                  tx->tx_list_state == GNILND_TX_ALLOCD) &&
                list_empty(&tx->tx_list),
                "tx %p with bad state %s (list_p %p) tx_list %s\n",
                tx, kgnilnd_tx_state2str(tx->tx_list_state), tx->tx_list_p,
                list_empty(&tx->tx_list) ? "empty" : "not empty");

        atomic_dec(&kgnilnd_data.kgn_ntx);

        /* we only allocate this if we need to */
        if (tx->tx_phys != NULL) {
                kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
                CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
                       LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
        }

        /* Only free the buffer if we used it */
        if (tx->tx_buffer_copy != NULL) {
                vfree(tx->tx_buffer_copy);
                tx->tx_buffer_copy = NULL;
                CDEBUG(D_MALLOC, "vfreed buffer2\n");
        }
#if 0
        KGNILND_POISON(tx, 0x5a, sizeof(kgn_tx_t));
#endif
        CDEBUG(D_MALLOC, "slab-freed 'tx': %lu at %p.\n", sizeof(*tx), tx);
        kmem_cache_free(kgnilnd_data.kgn_tx_cache, tx);
}

kgn_tx_t *
kgnilnd_alloc_tx (void)
{
        kgn_tx_t        *tx = NULL;

        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ALLOC_TX))
                return tx;

        tx = kmem_cache_alloc(kgnilnd_data.kgn_tx_cache, GFP_ATOMIC);
        if (tx == NULL) {
                CERROR("failed to allocate tx\n");
                return NULL;
        }
        CDEBUG(D_MALLOC, "slab-alloced 'tx': %lu at %p.\n",
               sizeof(*tx), tx);

        /* need this memset, cache alloc'd memory is not cleared */
        memset(tx, 0, sizeof(*tx));

        /* setup everything here to minimize time under the lock */
        tx->tx_buftype = GNILND_BUF_NONE;
        tx->tx_msg.gnm_type = GNILND_MSG_NONE;
        INIT_LIST_HEAD(&tx->tx_list);
        INIT_LIST_HEAD(&tx->tx_map_list);
        tx->tx_list_state = GNILND_TX_ALLOCD;

        atomic_inc(&kgnilnd_data.kgn_ntx);

        return tx;
}

/* csum_fold needs to be run on the return value before shipping over the wire */
#define _kgnilnd_cksum(seed, ptr, nob)  csum_partial(ptr, nob, seed)

/* we don't use offset as every one is passing a buffer reference that already
 * includes the offset into the base address -
 *  see kgnilnd_setup_virt_buffer and kgnilnd_setup_immediate_buffer */
static inline __u16
kgnilnd_cksum(void *ptr, size_t nob)
{
        __u16   sum;

        sum = csum_fold(_kgnilnd_cksum(0, ptr, nob));

        /* don't use magic 'no checksum' value */
        if (sum == 0)
                sum = 1;

        CDEBUG(D_INFO, "cksum 0x%x for ptr 0x%p sz %zu\n",
               sum, ptr, nob);

        return sum;
}

inline __u16
kgnilnd_cksum_kiov(unsigned int nkiov, lnet_kiov_t *kiov,
                    unsigned int offset, unsigned int nob, int dump_blob)
{
        __wsum             cksum = 0;
        __wsum             tmpck;
        __u16              retsum;
        void              *addr;
        unsigned int       fraglen;
        int                i, odd;

        LASSERT(nkiov > 0);
        LASSERT(nob > 0);

        CDEBUG(D_BUFFS, "calc cksum for kiov 0x%p nkiov %u offset %u nob %u, dump %d\n",
               kiov, nkiov, offset, nob, dump_blob);

        /* if loops changes, please change kgnilnd_setup_phys_buffer */

        while (offset >= kiov->kiov_len) {
                offset -= kiov->kiov_len;
                nkiov--;
                kiov++;
                LASSERT(nkiov > 0);
        }

        /* ignore nob here, if nob < (kiov_len - offset), kiov == 1 */
        odd = (unsigned long) (kiov[0].kiov_len - offset) & 1;

        if ((odd || *kgnilnd_tunables.kgn_vmap_cksum) && nkiov > 1) {
                struct page **pages = kgnilnd_data.kgn_cksum_map_pages[get_cpu()];

                LASSERTF(pages != NULL, "NULL pages for cpu %d map_pages 0x%p\n",
                         get_cpu(), kgnilnd_data.kgn_cksum_map_pages);

                CDEBUG(D_BUFFS, "odd %d len %u offset %u nob %u\n",
                       odd, kiov[0].kiov_len, offset, nob);

                for (i = 0; i < nkiov; i++) {
                        pages[i] = kiov[i].kiov_page;
                }

                addr = vmap(pages, nkiov, VM_MAP, PAGE_KERNEL);
                if (addr == NULL) {
                        CNETERR("Couldn't vmap %d frags on %d bytes to avoid odd length fragment in cksum\n",
                                nkiov, nob);
                        /* return zero to avoid killing tx - we'll just get warning on console
                         * when remote end sees zero checksum */
                        RETURN(0);
                }
                atomic_inc(&kgnilnd_data.kgn_nvmap_cksum);

                tmpck = _kgnilnd_cksum(0, (void *) addr + kiov[0].kiov_offset + offset, nob);
                cksum = tmpck;

                if (dump_blob) {
                        kgnilnd_dump_blob(D_BUFFS, "flat kiov RDMA payload",
                                          (void *)addr + kiov[0].kiov_offset + offset, nob);
                }
                CDEBUG(D_BUFFS, "cksum 0x%x (+0x%x) for addr 0x%p+%u len %u offset %u\n",
                       cksum, tmpck, addr, kiov[0].kiov_offset, nob, offset);
                vunmap(addr);
        } else {
                do {
                        fraglen = min(kiov->kiov_len - offset, nob);

                        /* make dang sure we don't send a bogus checksum if somehow we get
                         * an odd length fragment on anything but the last entry in a kiov  -
                         * we know from kgnilnd_setup_rdma_buffer that we can't have non
                         * PAGE_SIZE pages in the middle, so if nob < PAGE_SIZE, it is the last one */
                        LASSERTF(!(fraglen&1) || (nob < PAGE_SIZE),
                                 "odd fraglen %u on nkiov %d, nob %u kiov_len %u offset %u kiov 0x%p\n",
                                 fraglen, nkiov, nob, kiov->kiov_len, offset, kiov);

                        addr = (void *)kmap(kiov->kiov_page) + kiov->kiov_offset + offset;
                        tmpck = _kgnilnd_cksum(cksum, addr, fraglen);

                        CDEBUG(D_BUFFS,
                               "cksum 0x%x (+0x%x) for page 0x%p+%u (0x%p) len %u offset %u\n",
                               cksum, tmpck, kiov->kiov_page, kiov->kiov_offset, addr,
                               fraglen, offset);

                        cksum = tmpck;

                        if (dump_blob)
                                kgnilnd_dump_blob(D_BUFFS, "kiov cksum", addr, fraglen);

                        kunmap(kiov->kiov_page);

                        kiov++;
                        nkiov--;
                        nob -= fraglen;
                        offset = 0;

                        /* iov must not run out before end of data */
                        LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);

                } while (nob > 0);
        }

        retsum = csum_fold(cksum);

        /* don't use magic 'no checksum' value */
        if (retsum == 0)
                retsum = 1;

        CDEBUG(D_BUFFS, "retsum 0x%x from cksum 0x%x\n", retsum, cksum);

        return retsum;
}

void
kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source)
{
        msg->gnm_magic = GNILND_MSG_MAGIC;
        msg->gnm_version = GNILND_MSG_VERSION;
        msg->gnm_type = type;
        msg->gnm_payload_len = 0;
        msg->gnm_srcnid = source;
        /* gnm_connstamp gets set when FMA is sent */
        /* gnm_srcnid is set on creation via function argument
         * The right interface/net and nid is passed in when the message
         * is created.
         */
}

kgn_tx_t *
kgnilnd_new_tx_msg(int type, lnet_nid_t source)
{
        kgn_tx_t *tx = kgnilnd_alloc_tx();

        if (tx != NULL) {
                kgnilnd_init_msg(&tx->tx_msg, type, source);
        } else {
                CERROR("couldn't allocate new tx type %s!\n",
                       kgnilnd_msgtype2str(type));
        }

        return tx;
}

static void
kgnilnd_nak_rdma(kgn_conn_t *conn, int rx_type, int error, __u64 cookie, lnet_nid_t source) {
        kgn_tx_t        *tx;

        int             nak_type;

        switch (rx_type) {
        case GNILND_MSG_GET_REQ:
        case GNILND_MSG_GET_DONE:
                nak_type = GNILND_MSG_GET_NAK;
                break;
        case GNILND_MSG_PUT_REQ:
        case GNILND_MSG_PUT_ACK:
        case GNILND_MSG_PUT_DONE:
                nak_type = GNILND_MSG_PUT_NAK;
                break;
        case GNILND_MSG_PUT_REQ_REV:
        case GNILND_MSG_PUT_DONE_REV:
                nak_type = GNILND_MSG_PUT_NAK_REV;
                break;
        case GNILND_MSG_GET_REQ_REV:
        case GNILND_MSG_GET_ACK_REV:
        case GNILND_MSG_GET_DONE_REV:
                nak_type = GNILND_MSG_GET_NAK_REV;
                break;
        default:
                CERROR("invalid msg type %s (%d)\n",
                        kgnilnd_msgtype2str(rx_type), rx_type);
                LBUG();
        }
        /* only allow NAK on error and truncate to zero */
        LASSERTF(error <= 0, "error %d conn 0x%p, cookie "LPU64"\n",
                 error, conn, cookie);

        tx = kgnilnd_new_tx_msg(nak_type, source);
        if (tx == NULL) {
                CNETERR("can't get TX to NAK RDMA to %s\n",
                        libcfs_nid2str(conn->gnc_peer->gnp_nid));
                return;
        }

        tx->tx_msg.gnm_u.completion.gncm_retval = error;
        tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
        kgnilnd_queue_tx(conn, tx);
}

int
kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov, struct iovec *iov,
                               lnet_kiov_t *kiov, unsigned int offset, unsigned int nob)

{
        kgn_msg_t       *msg = &tx->tx_msg;
        int              i;

        /* To help save on MDDs for short messages, we'll vmap a kiov to allow
         * gni_smsg_send to send that as the payload */

        LASSERT(tx->tx_buftype == GNILND_BUF_NONE);

        if (nob == 0) {
                tx->tx_buffer = NULL;
        } else if (kiov != NULL) {
                LASSERTF(niov > 0 && niov < GNILND_MAX_IMMEDIATE/PAGE_SIZE,
                         "bad niov %d\n", niov);

                while (offset >= kiov->kiov_len) {
                        offset -= kiov->kiov_len;
                        niov--;
                        kiov++;
                        LASSERT(niov > 0);
                }
                for (i = 0; i < niov; i++) {
                        /* We can't have a kiov_offset on anything but the first entry,
                         * otherwise we'll have a hole at the end of the mapping as we only map
                         * whole pages.
                         * Also, if we have a kiov_len < PAGE_SIZE but we need to map more
                         * than kiov_len, we will also have a whole at the end of that page
                         * which isn't allowed */
                        if ((kiov[i].kiov_offset != 0 && i > 0) ||
                            (kiov[i].kiov_offset + kiov[i].kiov_len != PAGE_SIZE && i < niov - 1)) {
                                CNETERR("Can't make payload contiguous in I/O VM:"
                                       "page %d, offset %u, nob %u, kiov_offset %u kiov_len %u \n",
                                       i, offset, nob, kiov->kiov_offset, kiov->kiov_len);
                                RETURN(-EINVAL);
                        }
                        tx->tx_imm_pages[i] = kiov[i].kiov_page;
                }

                /* hijack tx_phys for the later unmap */
                if (niov == 1) {
                        /* tx->phyx being equal to NULL is the signal for unmap to discern between kmap and vmap */
                        tx->tx_phys = NULL;
                        tx->tx_buffer = (void *)kmap(tx->tx_imm_pages[0]) + kiov[0].kiov_offset + offset;
                        atomic_inc(&kgnilnd_data.kgn_nkmap_short);
                        GNIDBG_TX(D_NET, tx, "kmapped page for %d bytes for kiov 0x%p, buffer 0x%p",
                                nob, kiov, tx->tx_buffer);
                } else {
                        tx->tx_phys = vmap(tx->tx_imm_pages, niov, VM_MAP, PAGE_KERNEL);
                        if (tx->tx_phys == NULL) {
                                CNETERR("Couldn't vmap %d frags on %d bytes\n", niov, nob);
                                RETURN(-ENOMEM);

                        }
                        atomic_inc(&kgnilnd_data.kgn_nvmap_short);
                        /* make sure we take into account the kiov offset as the start of the buffer */
                        tx->tx_buffer = (void *)tx->tx_phys + kiov[0].kiov_offset + offset;
                        GNIDBG_TX(D_NET, tx, "mapped %d pages for %d bytes from kiov 0x%p to 0x%p, buffer 0x%p",
                                niov, nob, kiov, tx->tx_phys, tx->tx_buffer);
                }
                tx->tx_buftype = GNILND_BUF_IMMEDIATE_KIOV;
                tx->tx_nob = nob;

        } else {
                /* For now this is almost identical to kgnilnd_setup_virt_buffer, but we
                 * could "flatten" the payload into a single contiguous buffer ready
                 * for sending direct over an FMA if we ever needed to. */

                LASSERT(niov > 0);

                while (offset >= iov->iov_len) {
                        offset -= iov->iov_len;
                        niov--;
                        iov++;
                        LASSERT(niov > 0);
                }

                if (nob > iov->iov_len - offset) {
                        CERROR("Can't handle multiple vaddr fragments\n");
                        return -EMSGSIZE;
                }

                tx->tx_buffer = (void *)(((unsigned long)iov->iov_base) + offset);

                tx->tx_buftype = GNILND_BUF_IMMEDIATE;
                tx->tx_nob = nob;
        }

        /* checksum payload early - it shouldn't be changing after lnd_send */
        if (*kgnilnd_tunables.kgn_checksum >= 2) {
                msg->gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
                if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM2)) {
                        msg->gnm_payload_cksum += 0xe00e;
                }
                if (*kgnilnd_tunables.kgn_checksum_dump > 1) {
                        kgnilnd_dump_blob(D_BUFFS, "payload checksum",
                                          tx->tx_buffer, nob);
                }
        } else {
                msg->gnm_payload_cksum = 0;
        }

        return 0;
}

int
kgnilnd_setup_virt_buffer(kgn_tx_t *tx,
                          unsigned int niov, struct iovec *iov,
                          unsigned int offset, unsigned int nob)

{
        LASSERT(nob > 0);
        LASSERT(niov > 0);
        LASSERT(tx->tx_buftype == GNILND_BUF_NONE);

        while (offset >= iov->iov_len) {
                offset -= iov->iov_len;
                niov--;
                iov++;
                LASSERT(niov > 0);
        }

        if (nob > iov->iov_len - offset) {
                CERROR("Can't handle multiple vaddr fragments\n");
                return -EMSGSIZE;
        }

        tx->tx_buftype = GNILND_BUF_VIRT_UNMAPPED;
        tx->tx_nob = nob;
        tx->tx_buffer = (void *)(((unsigned long)iov->iov_base) + offset);
        return 0;
}

int
kgnilnd_setup_phys_buffer(kgn_tx_t *tx, int nkiov, lnet_kiov_t *kiov,
                          unsigned int offset, unsigned int nob)
{
        gni_mem_segment_t *phys;
        int             rc = 0;
        unsigned int    fraglen;

        GNIDBG_TX(D_NET, tx, "niov %d kiov 0x%p offset %u nob %u", nkiov, kiov, offset, nob);

        LASSERT(nob > 0);
        LASSERT(nkiov > 0);
        LASSERT(tx->tx_buftype == GNILND_BUF_NONE);

        /* only allocate this if we are going to use it */
        tx->tx_phys = kmem_cache_alloc(kgnilnd_data.kgn_tx_phys_cache,
                                              GFP_ATOMIC);
        if (tx->tx_phys == NULL) {
                CERROR("failed to allocate tx_phys\n");
                rc = -ENOMEM;
                GOTO(error, rc);
        }

        CDEBUG(D_MALLOC, "slab-alloced 'tx->tx_phys': %lu at %p.\n",
               LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);

        /* if loops changes, please change kgnilnd_cksum_kiov
         *   and kgnilnd_setup_immediate_buffer */

        while (offset >= kiov->kiov_len) {
                offset -= kiov->kiov_len;
                nkiov--;
                kiov++;
                LASSERT(nkiov > 0);
        }

        /* at this point, kiov points to the first page that we'll actually map
         * now that we've seeked into the koiv for offset and dropped any
         * leading pages that fall entirely within the offset */
        tx->tx_buftype = GNILND_BUF_PHYS_UNMAPPED;
        tx->tx_nob = nob;

        /* kiov_offset is start of 'valid' buffer, so index offset past that */
        tx->tx_buffer = (void *)((unsigned long)(kiov->kiov_offset + offset));
        phys = tx->tx_phys;

        CDEBUG(D_NET, "tx 0x%p buffer 0x%p map start kiov 0x%p+%u niov %d offset %u\n",
               tx, tx->tx_buffer, kiov, kiov->kiov_offset, nkiov, offset);

        do {
                fraglen = min(kiov->kiov_len - offset, nob);

                /* We can't have a kiov_offset on anything but the first entry,
                 * otherwise we'll have a hole at the end of the mapping as we only map
                 * whole pages. Only the first page is allowed to have an offset -
                 * we'll add that into tx->tx_buffer and that will get used when we
                 * map in the segments (see kgnilnd_map_buffer).
                 * Also, if we have a kiov_len < PAGE_SIZE but we need to map more
                 * than kiov_len, we will also have a whole at the end of that page
                 * which isn't allowed */
                if ((phys != tx->tx_phys) &&
                    ((kiov->kiov_offset != 0) ||
                     ((kiov->kiov_len < PAGE_SIZE) && (nob > kiov->kiov_len)))) {
                        CERROR("Can't make payload contiguous in I/O VM:"
                               "page %d, offset %u, nob %u, kiov_offset %u kiov_len %u \n",
                               (int)(phys - tx->tx_phys),
                               offset, nob, kiov->kiov_offset, kiov->kiov_len);
                        rc = -EINVAL;
                        GOTO(error, rc);
                }

                if ((phys - tx->tx_phys) == LNET_MAX_IOV) {
                        CERROR ("payload too big (%d)\n", (int)(phys - tx->tx_phys));
                        rc = -EMSGSIZE;
                        GOTO(error, rc);
                }

                if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PHYS_SETUP)) {
                        rc = -EINVAL;
                        GOTO(error, rc);
                }

                CDEBUG(D_BUFFS, "page 0x%p kiov_offset %u kiov_len %u nob %u "
                               "nkiov %u offset %u\n",
                      kiov->kiov_page, kiov->kiov_offset, kiov->kiov_len, nob, nkiov, offset);

                phys->address = page_to_phys(kiov->kiov_page);
                phys++;
                kiov++;
                nkiov--;
                nob -= fraglen;
                offset = 0;

                /* iov must not run out before end of data */
                LASSERTF(nob == 0 || nkiov > 0, "nob %u nkiov %u\n", nob, nkiov);

        } while (nob > 0);

        tx->tx_phys_npages = phys - tx->tx_phys;

        return 0;

error:
        if (tx->tx_phys != NULL) {
                kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
                CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
                       sizeof(*tx->tx_phys), tx->tx_phys);
                tx->tx_phys = NULL;
        }
        return rc;
}

static inline int
kgnilnd_setup_rdma_buffer(kgn_tx_t *tx, unsigned int niov,
                          struct iovec *iov, lnet_kiov_t *kiov,
                          unsigned int offset, unsigned int nob)
{
        int     rc;

        LASSERTF((iov == NULL) != (kiov == NULL), "iov 0x%p, kiov 0x%p, tx 0x%p,"
                                                " offset %d, nob %d, niov %d\n"
                                                , iov, kiov, tx, offset, nob, niov);

        if (kiov != NULL) {
                rc = kgnilnd_setup_phys_buffer(tx, niov, kiov, offset, nob);
        } else {
                rc = kgnilnd_setup_virt_buffer(tx, niov, iov, offset, nob);
        }
        return rc;
}

/* kgnilnd_parse_lnet_rdma()
 * lntmsg - message passed in from lnet.
 * niov, kiov, offset - see lnd_t in lib-types.h for descriptions.
 * nob - actual number of bytes to in this message.
 * put_len - It is possible for PUTs to have a different length than the
 *           length stored in lntmsg->msg_len since LNET can adjust this
 *           length based on it's buffer size and offset.
 *           lnet_try_match_md() sets the mlength that we use to do the RDMA
 *           transfer.
 */
static void
kgnilnd_parse_lnet_rdma(lnet_msg_t *lntmsg, unsigned int *niov,
                        unsigned int *offset, unsigned int *nob,
                        lnet_kiov_t **kiov, int put_len)
{
        /* GETs are weird, see kgnilnd_send */
        if (lntmsg->msg_type == LNET_MSG_GET) {
                if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0) {
                        *kiov = NULL;
                } else {
                        *kiov = lntmsg->msg_md->md_iov.kiov;
                }
                *niov = lntmsg->msg_md->md_niov;
                *nob = lntmsg->msg_md->md_length;
                *offset = 0;
        } else {
                *kiov = lntmsg->msg_kiov;
                *niov = lntmsg->msg_niov;
                *nob = put_len;
                *offset = lntmsg->msg_offset;
        }
}

static inline void
kgnilnd_compute_rdma_cksum(kgn_tx_t *tx, int put_len)
{
        unsigned int     niov, offset, nob;
        lnet_kiov_t     *kiov;
        lnet_msg_t      *lntmsg = tx->tx_lntmsg[0];
        int              dump_cksum = (*kgnilnd_tunables.kgn_checksum_dump > 1);

        GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE) ||
                           (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE) ||
                           (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
                           (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
                           (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
                           (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV)),
                      "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));

        if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV) ||
            (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV)) {
                tx->tx_msg.gnm_payload_cksum = 0;
                return;
        }
        if (*kgnilnd_tunables.kgn_checksum < 3) {
                tx->tx_msg.gnm_payload_cksum = 0;
                return;
        }

        GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);

        kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov,
                                put_len);

        if (kiov != NULL) {
                tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, dump_cksum);
        } else {
                tx->tx_msg.gnm_payload_cksum = kgnilnd_cksum(tx->tx_buffer, nob);
                if (dump_cksum) {
                        kgnilnd_dump_blob(D_BUFFS, "peer RDMA payload", tx->tx_buffer, nob);
                }
        }

        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM3)) {
                tx->tx_msg.gnm_payload_cksum += 0xd00d;
        }
}

/* kgnilnd_verify_rdma_cksum()
 * tx - PUT_DONE/GET_DONE matched tx.
 * rx_cksum - received checksum to compare against.
 * put_len - see kgnilnd_parse_lnet_rdma comments.
 */
static inline int
kgnilnd_verify_rdma_cksum(kgn_tx_t *tx, __u16 rx_cksum, int put_len)
{
        int              rc = 0;
        __u16            cksum;
        unsigned int     niov, offset, nob;
        lnet_kiov_t     *kiov;
        lnet_msg_t      *lntmsg = tx->tx_lntmsg[0];
        int dump_on_err = *kgnilnd_tunables.kgn_checksum_dump;

        /* we can only match certain requests */
        GNITX_ASSERTF(tx, ((tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) ||
                           (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK) ||
                           (tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
                           (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV) ||
                           (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) ||
                           (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV)),
                      "bad type %s", kgnilnd_msgtype2str(tx->tx_msg.gnm_type));

        if ((tx->tx_msg.gnm_type == GNILND_MSG_PUT_REQ_REV) ||
            (tx->tx_msg.gnm_type == GNILND_MSG_GET_ACK_REV)) {
                return 0;
        }

        if (rx_cksum == 0)  {
                if (*kgnilnd_tunables.kgn_checksum >= 3) {
                        GNIDBG_MSG(D_WARNING, &tx->tx_msg,
                                   "no RDMA payload checksum when enabled");
                }
                return 0;
        }

        GNITX_ASSERTF(tx, lntmsg, "no LNet message!", NULL);

        kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, put_len);

        if (kiov != NULL) {
                cksum = kgnilnd_cksum_kiov(niov, kiov, offset, nob, 0);
        } else {
                cksum = kgnilnd_cksum(tx->tx_buffer, nob);
        }

        if (cksum != rx_cksum) {
                GNIDBG_MSG(D_NETERROR, &tx->tx_msg,
                           "Bad RDMA payload checksum (%x expected %x); "
                           "kiov 0x%p niov %d nob %u offset %u",
                            cksum, rx_cksum, kiov, niov, nob, offset);
                switch (dump_on_err) {
                case 2:
                        if (kiov != NULL) {
                                kgnilnd_cksum_kiov(niov, kiov, offset, nob, 1);
                        } else {
                                kgnilnd_dump_blob(D_BUFFS, "RDMA payload",
                                                  tx->tx_buffer, nob);
                        }
                        /* fall through to dump log */
                case 1:
                        libcfs_debug_dumplog();
                        break;
                default:
                        break;
                }
                rc = -ENOKEY;
                /* kgnilnd_check_fma_rx will close conn, kill tx with error */
        }
        return rc;
}

void
kgnilnd_mem_add_map_list(kgn_device_t *dev, kgn_tx_t *tx)
{
        int     bytes;

        GNITX_ASSERTF(tx, list_empty(&tx->tx_map_list),
                "already mapped!", NULL);

        spin_lock(&dev->gnd_map_lock);
        switch (tx->tx_buftype) {
        default:
                GNIDBG_TX(D_EMERG, tx,
                        "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
                spin_unlock(&dev->gnd_map_lock);
                LBUG();
                break;

        case GNILND_BUF_PHYS_MAPPED:
                bytes = tx->tx_phys_npages * PAGE_SIZE;
                dev->gnd_map_nphys++;
                dev->gnd_map_physnop += tx->tx_phys_npages;
                break;

        case GNILND_BUF_VIRT_MAPPED:
                bytes = tx->tx_nob;
                dev->gnd_map_nvirt++;
                dev->gnd_map_virtnob += tx->tx_nob;
                break;
        }

        if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
            tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
                atomic64_add(bytes, &dev->gnd_rdmaq_bytes_out);
                GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to "LPD64"",
                          bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
        }

        atomic_inc(&dev->gnd_n_mdd);
        atomic64_add(bytes, &dev->gnd_nbytes_map);

        /* clear retrans to prevent any SMSG goofiness as that code uses the same counter */
        tx->tx_retrans = 0;

        /* we only get here in the valid cases */
        list_add_tail(&tx->tx_map_list, &dev->gnd_map_list);
        dev->gnd_map_version++;
        spin_unlock(&dev->gnd_map_lock);
}

void
kgnilnd_mem_del_map_list(kgn_device_t *dev, kgn_tx_t *tx)
{
        int     bytes;

        GNITX_ASSERTF(tx, !list_empty(&tx->tx_map_list),
                "not mapped!", NULL);
        spin_lock(&dev->gnd_map_lock);

        switch (tx->tx_buftype) {
        default:
                GNIDBG_TX(D_EMERG, tx,
                        "SOFTWARE BUG: invalid mapping %d", tx->tx_buftype);
                spin_unlock(&dev->gnd_map_lock);
                LBUG();
                break;

        case GNILND_BUF_PHYS_UNMAPPED:
                bytes = tx->tx_phys_npages * PAGE_SIZE;
                dev->gnd_map_nphys--;
                dev->gnd_map_physnop -= tx->tx_phys_npages;
                break;

        case GNILND_BUF_VIRT_UNMAPPED:
                bytes = tx->tx_nob;
                dev->gnd_map_nvirt--;
                dev->gnd_map_virtnob -= tx->tx_nob;
                break;
        }

        if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
            tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
                atomic64_sub(bytes, &dev->gnd_rdmaq_bytes_out);
                LASSERTF(atomic64_read(&dev->gnd_rdmaq_bytes_out) >= 0,
                         "bytes_out negative! %ld\n", atomic64_read(&dev->gnd_rdmaq_bytes_out));
                GNIDBG_TX(D_NETTRACE, tx, "rdma -- %d to "LPD64"",
                          bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
        }

        atomic_dec(&dev->gnd_n_mdd);
        atomic64_sub(bytes, &dev->gnd_nbytes_map);

        /* we only get here in the valid cases */
        list_del_init(&tx->tx_map_list);
        dev->gnd_map_version++;
        spin_unlock(&dev->gnd_map_lock);
}

int
kgnilnd_map_buffer(kgn_tx_t *tx)
{
        kgn_conn_t       *conn = tx->tx_conn;
        kgn_device_t     *dev = conn->gnc_device;
        __u32             flags = GNI_MEM_READWRITE;
        gni_return_t      rrc;

        /* The kgnilnd_mem_register(_segments) Gemini Driver functions can
         * be called concurrently as there are internal locks that protect
         * any data structures or HW resources. We just need to ensure
         * that our concurrency doesn't result in the kgn_device_t
         * getting nuked while we are in here */

        LASSERTF(conn != NULL, "tx %p with NULL conn, someone forgot"
                " to set tx_conn before calling %s\n", tx, __FUNCTION__);

        if (unlikely(CFS_FAIL_CHECK(CFS_FAIL_GNI_MAP_TX)))
                RETURN(-ENOMEM);

        if (*kgnilnd_tunables.kgn_bte_relaxed_ordering) {
                flags |= GNI_MEM_RELAXED_PI_ORDERING;
        }

        switch (tx->tx_buftype) {
        default:
                LBUG();

        case GNILND_BUF_NONE:
        case GNILND_BUF_IMMEDIATE:
        case GNILND_BUF_IMMEDIATE_KIOV:
        case GNILND_BUF_PHYS_MAPPED:
        case GNILND_BUF_VIRT_MAPPED:
                return 0;

        case GNILND_BUF_PHYS_UNMAPPED:
                GNITX_ASSERTF(tx, tx->tx_phys != NULL, "physical buffer not there!", NULL);
                rrc = kgnilnd_mem_register_segments(dev->gnd_handle,
                        tx->tx_phys, tx->tx_phys_npages, NULL,
                        GNI_MEM_PHYS_SEGMENTS | flags,
                        &tx->tx_map_key);
                /* could race with other uses of the map counts, but this is ok
                 * - this needs to turn into a non-fatal error soon to allow
                 *  GART resource, etc starvation handling */
                if (rrc != GNI_RC_SUCCESS) {
                        GNIDBG_TX(D_NET, tx, "Can't map %d pages: dev %d "
                                "phys %u pp %u, virt %u nob "LPU64"",
                                tx->tx_phys_npages, dev->gnd_id,
                                dev->gnd_map_nphys, dev->gnd_map_physnop,
                                dev->gnd_map_nvirt, dev->gnd_map_virtnob);
                        RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
                }

                tx->tx_buftype = GNILND_BUF_PHYS_MAPPED;
                kgnilnd_mem_add_map_list(dev, tx);
                return 0;

        case GNILND_BUF_VIRT_UNMAPPED:
                rrc = kgnilnd_mem_register(dev->gnd_handle,
                        (__u64)tx->tx_buffer, tx->tx_nob,
                        NULL, flags, &tx->tx_map_key);
                if (rrc != GNI_RC_SUCCESS) {
                        GNIDBG_TX(D_NET, tx, "Can't map %u bytes: dev %d "
                                "phys %u pp %u, virt %u nob "LPU64"",
                                tx->tx_nob, dev->gnd_id,
                                dev->gnd_map_nphys, dev->gnd_map_physnop,
                                dev->gnd_map_nvirt, dev->gnd_map_virtnob);
                        RETURN(rrc == GNI_RC_ERROR_RESOURCE ? -ENOMEM : -EINVAL);
                }

                tx->tx_buftype = GNILND_BUF_VIRT_MAPPED;
                kgnilnd_mem_add_map_list(dev, tx);
                if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
                    tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
                        atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_out);
                        GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %ld\n",
                               tx->tx_nob, atomic64_read(&dev->gnd_rdmaq_bytes_out));
                }

                return 0;
        }
}

void
kgnilnd_add_purgatory_tx(kgn_tx_t *tx)
{
        kgn_conn_t              *conn = tx->tx_conn;
        kgn_mdd_purgatory_t     *gmp;

        LIBCFS_ALLOC(gmp, sizeof(*gmp));
        LASSERTF(gmp != NULL, "couldn't allocate MDD purgatory member;"
                " asserting to avoid data corruption\n");
        if (tx->tx_buffer_copy)
                gmp->gmp_map_key = tx->tx_buffer_copy_map_key;
        else
                gmp->gmp_map_key = tx->tx_map_key;

        atomic_inc(&conn->gnc_device->gnd_n_mdd_held);

        /* ensure that we don't have a blank purgatory - indicating the
         * conn is not already on purgatory lists - we'd never recover these
         * MDD if that were the case */
        GNITX_ASSERTF(tx, conn->gnc_in_purgatory,
                "conn 0x%p->%s with NULL purgatory",
                conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));

        /* link 'er up! - only place we really need to lock for
         * concurrent access */
        spin_lock(&conn->gnc_list_lock);
        list_add_tail(&gmp->gmp_list, &conn->gnc_mdd_list);
        spin_unlock(&conn->gnc_list_lock);
}

void
kgnilnd_unmap_buffer(kgn_tx_t *tx, int error)
{
        kgn_device_t     *dev;
        gni_return_t      rrc;
        int               hold_timeout = 0;

        /* code below relies on +1 relationship ... */
        CLASSERT(GNILND_BUF_PHYS_MAPPED == (GNILND_BUF_PHYS_UNMAPPED + 1));
        CLASSERT(GNILND_BUF_VIRT_MAPPED == (GNILND_BUF_VIRT_UNMAPPED + 1));

        switch (tx->tx_buftype) {
        default:
                LBUG();

        case GNILND_BUF_NONE:
        case GNILND_BUF_IMMEDIATE:
        case GNILND_BUF_PHYS_UNMAPPED:
        case GNILND_BUF_VIRT_UNMAPPED:
                break;
        case GNILND_BUF_IMMEDIATE_KIOV:
                if (tx->tx_phys != NULL) {
                        vunmap(tx->tx_phys);
                } else if (tx->tx_phys == NULL && tx->tx_buffer != NULL) {
                        kunmap(tx->tx_imm_pages[0]);
                }
                /* clear to prevent kgnilnd_free_tx from thinking
                 * this is a RDMA descriptor */
                tx->tx_phys = NULL;
                break;

        case GNILND_BUF_PHYS_MAPPED:
        case GNILND_BUF_VIRT_MAPPED:
                LASSERT(tx->tx_conn != NULL);

                dev = tx->tx_conn->gnc_device;

                /* only want to hold if we are closing conn without
                 * verified peer notification  - the theory is that
                 * a TX error can be communicated in all other cases */
                if (tx->tx_conn->gnc_state != GNILND_CONN_ESTABLISHED &&
                    kgnilnd_check_purgatory_conn(tx->tx_conn)) {
                        kgnilnd_add_purgatory_tx(tx);

                        /* The timeout we give to kgni is a deadman stop only.
                         *  we are setting high to ensure we don't have the kgni timer
                         *  fire before ours fires _and_ is handled */
                        hold_timeout = GNILND_TIMEOUT2DEADMAN;

                        GNIDBG_TX(D_NET, tx,
                                 "dev %p delaying MDD release for %dms key "LPX64"."LPX64"",
                                 tx->tx_conn->gnc_device, hold_timeout,
                                 tx->tx_map_key.qword1, tx->tx_map_key.qword2);
                }
                if (tx->tx_buffer_copy != NULL) {
                        rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_buffer_copy_map_key, hold_timeout);
                        LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
                        rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, 0);
                        LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
                } else {
                        rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, hold_timeout);
                        LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
                }

                tx->tx_buftype--;
                kgnilnd_mem_del_map_list(dev, tx);
                break;
        }
}

void
kgnilnd_tx_done(kgn_tx_t *tx, int completion)
{
        lnet_msg_t      *lntmsg0, *lntmsg1;
        int             status0, status1;
        lnet_ni_t       *ni = NULL;
        kgn_conn_t      *conn = tx->tx_conn;

        LASSERT(!in_interrupt());

        lntmsg0 = tx->tx_lntmsg[0]; tx->tx_lntmsg[0] = NULL;
        lntmsg1 = tx->tx_lntmsg[1]; tx->tx_lntmsg[1] = NULL;

        if (completion &&
            !(tx->tx_state & GNILND_TX_QUIET_ERROR) &&
            !kgnilnd_conn_clean_errno(completion)) {
                GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
                       "error %d on tx 0x%p->%s id %u/%d state %s age %ds",
                       completion, tx, conn ?
                       libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
                       tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
                       kgnilnd_tx_state2str(tx->tx_list_state),
                       cfs_duration_sec((unsigned long)jiffies - tx->tx_qtime));
        }

        /* The error codes determine if we hold onto the MDD */
        kgnilnd_unmap_buffer(tx, completion);

        /* we have to deliver a reply on lntmsg[1] for the GET, so make sure
         * we play nice with the error codes to avoid delivering a failed
         * REQUEST and then a REPLY event as well */

        /* return -EIO to lnet - it is the magic value for failed sends */
        if (tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
                status0 = 0;
                status1 = completion;
        } else {
                status0 = status1 = completion;
        }

        tx->tx_buftype = GNILND_BUF_NONE;
        tx->tx_msg.gnm_type = GNILND_MSG_NONE;

        /* lnet_finalize doesn't do anything with the *ni, so ok for us to
         * set NULL when we are a tx without a conn */
        if (conn != NULL) {
                ni = conn->gnc_peer->gnp_net->gnn_ni;

                spin_lock(&conn->gnc_tx_lock);

                LASSERTF(test_and_clear_bit(tx->tx_id.txe_idx,
                        (volatile unsigned long *)&conn->gnc_tx_bits),
                        "conn %p tx %p bit %d already cleared\n",
                        conn, tx, tx->tx_id.txe_idx);

                LASSERTF(conn->gnc_tx_ref_table[tx->tx_id.txe_idx] != NULL,
                         "msg_id %d already NULL\n", tx->tx_id.txe_idx);

                conn->gnc_tx_ref_table[tx->tx_id.txe_idx] = NULL;
                spin_unlock(&conn->gnc_tx_lock);
        }

        kgnilnd_free_tx(tx);

        /* finalize AFTER freeing lnet msgs */

        /* warning - we should hold no locks here - calling lnet_finalize
         * could free up lnet credits, resulting in a call chain back into
         * the LND via kgnilnd_send and friends */

        lnet_finalize(ni, lntmsg0, status0);

        if (lntmsg1 != NULL) {
                lnet_finalize(ni, lntmsg1, status1);
        }
}

void
kgnilnd_txlist_done(struct list_head *txlist, int error)
{
        kgn_tx_t        *tx, *txn;
        int              err_printed = 0;

        if (list_empty(txlist))
                return;

        list_for_each_entry_safe(tx, txn, txlist, tx_list) {
                /* only print the first error */
                if (err_printed)
                        tx->tx_state |= GNILND_TX_QUIET_ERROR;
                list_del_init(&tx->tx_list);
                kgnilnd_tx_done(tx, error);
                err_printed++;
        }
}
int
kgnilnd_set_tx_id(kgn_tx_t *tx, kgn_conn_t *conn)
{
        int     id;

        spin_lock(&conn->gnc_tx_lock);

        /* ID zero is NOT ALLOWED!!! */

search_again:
        id = find_next_zero_bit((unsigned long *)&conn->gnc_tx_bits,
                                 GNILND_MAX_MSG_ID, conn->gnc_next_tx);
        if (id == GNILND_MAX_MSG_ID) {
                if (conn->gnc_next_tx != 1) {
                        /* we only searched from next_tx to end and didn't find
                         * one, so search again from start */
                        conn->gnc_next_tx = 1;
                        goto search_again;
                }
                /* couldn't find one! */
                spin_unlock(&conn->gnc_tx_lock);
                return -E2BIG;
        }

        /* bump next_tx to prevent immediate reuse */
        conn->gnc_next_tx = id + 1;

        set_bit(id, (volatile unsigned long *)&conn->gnc_tx_bits);
        LASSERTF(conn->gnc_tx_ref_table[id] == NULL,
                 "tx 0x%p already at id %d\n",
                 conn->gnc_tx_ref_table[id], id);

        /* delay these until we have a valid ID - prevents bad clear of the bit
         * in kgnilnd_tx_done */
        tx->tx_conn = conn;
        tx->tx_id.txe_cqid = conn->gnc_cqid;

        tx->tx_id.txe_idx = id;
        conn->gnc_tx_ref_table[id] = tx;

        /* Using jiffies to help differentiate against TX reuse - with
         * the usual minimum of a 250HZ clock, we wrap jiffies on the same TX
         * if we are sending to the same node faster than 256000/sec.
         * To help guard against this, we OR in the tx_seq - that is 32 bits */

        tx->tx_id.txe_chips = (__u32)(jiffies | atomic_read(&conn->gnc_tx_seq));

        GNIDBG_TX(D_NET, tx, "set cookie/id/bits", NULL);

        spin_unlock(&conn->gnc_tx_lock);
        return 0;
}

static inline int
kgnilnd_tx_should_retry(kgn_conn_t *conn, kgn_tx_t *tx)
{
        int             max_retrans = *kgnilnd_tunables.kgn_max_retransmits;
        int             log_retrans;
        int             log_retrans_level;

        /* I need kgni credits to send this.  Replace tx at the head of the
         * fmaq and I'll get rescheduled when credits appear */
        tx->tx_state = 0;
        tx->tx_retrans++;
        conn->gnc_tx_retrans++;
        log_retrans = ((tx->tx_retrans < 25) || ((tx->tx_retrans % 25) == 0) ||
                        (tx->tx_retrans > (max_retrans / 2)));
        log_retrans_level = tx->tx_retrans < (max_retrans / 2) ? D_NET : D_NETERROR;

        /* Decision time - either error, warn or just retransmit */

        /* we don't care about TX timeout - it could be that the network is slower
         * or throttled. We'll keep retranmitting - so if the network is so slow
         * that we fill up our mailbox, we'll keep trying to resend that msg
         * until we exceed the max_retrans _or_ gnc_last_rx expires, indicating
         * that he hasn't send us any traffic in return */

        if (tx->tx_retrans > max_retrans) {
                /* this means we are not backing off the retransmits
                 * in a healthy manner and are likely chewing up the
                 * CPU cycles quite badly */
                GNIDBG_TOMSG(D_ERROR, &tx->tx_msg,
                        "SOFTWARE BUG: too many retransmits (%d) for tx id %x "
                        "conn 0x%p->%s\n",
                        tx->tx_retrans, tx->tx_id, conn,
                        libcfs_nid2str(conn->gnc_peer->gnp_nid));

                /* yes - double errors to help debug this condition */
                GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg, "connection dead. "
                        "unable to send to %s for %lu secs (%d tries)",
                        libcfs_nid2str(tx->tx_conn->gnc_peer->gnp_nid),
                        cfs_duration_sec(jiffies - tx->tx_cred_wait),
                        tx->tx_retrans);

                kgnilnd_close_conn(conn, -ETIMEDOUT);

                /* caller should terminate */
                RETURN(0);
        } else {
                /* some reasonable throttling of the debug message */
                if (log_retrans) {
                        unsigned long now = jiffies;
                        /* XXX Nic: Mystical TX debug here... */
                        GNIDBG_SMSG_CREDS(log_retrans_level, conn);
                        GNIDBG_TOMSG(log_retrans_level, &tx->tx_msg,
                                "NOT_DONE on conn 0x%p->%s id %x retrans %d wait %dus"
                                " last_msg %uus/%uus last_cq %uus/%uus",
                                conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
                                tx->tx_id, tx->tx_retrans,
                                jiffies_to_usecs(now - tx->tx_cred_wait),
                                jiffies_to_usecs(now - conn->gnc_last_tx),
                                jiffies_to_usecs(now - conn->gnc_last_rx),
                                jiffies_to_usecs(now - conn->gnc_last_tx_cq),
                                jiffies_to_usecs(now - conn->gnc_last_rx_cq));
                }
                /* caller should retry */
                RETURN(1);
        }
}

/* caller must be holding gnd_cq_mutex and not unlock it afterwards, as we need to drop it
 * to avoid bad ordering with state_lock */

static inline int
kgnilnd_sendmsg_nolock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
                spinlock_t *state_lock, kgn_tx_list_state_t state)
{
        kgn_conn_t      *conn = tx->tx_conn;
        kgn_msg_t       *msg = &tx->tx_msg;
        int              retry_send;
        gni_return_t     rrc;
        unsigned long    newest_last_rx, timeout;
        unsigned long    now;

        LASSERTF((msg->gnm_type == GNILND_MSG_IMMEDIATE) ?
                immediatenob <= *kgnilnd_tunables.kgn_max_immediate :
                immediatenob == 0,
                "msg 0x%p type %d wrong payload size %d\n",
                msg, msg->gnm_type, immediatenob);

        /* make sure we catch all the cases where we'd send on a dirty old mbox
         * but allow case for sending CLOSE. Since this check is within the CQ
         * mutex barrier and the close message is only sent through
         * kgnilnd_send_conn_close the last message out the door will be the
         * close message.
         */
        if (atomic_read(&conn->gnc_peer->gnp_dirty_eps) != 0 && msg->gnm_type != GNILND_MSG_CLOSE) {
                kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
                kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
                /* Return -ETIME, we are closing the connection already so we dont want to
                 * have this tx hit the wire. The tx will be killed by the calling function.
                 * Once the EP is marked dirty the close message will be the last
                 * thing to hit the wire */
                return -ETIME;
        }

        now = jiffies;
        timeout = cfs_time_seconds(conn->gnc_timeout);

        newest_last_rx = GNILND_LASTRX(conn);

        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SEND_TIMEOUT)) {
                now = now + (GNILND_TIMEOUTRX(timeout) * 2);
        }

        if (time_after_eq(now, newest_last_rx + GNILND_TIMEOUTRX(timeout))) {
                GNIDBG_CONN(D_NETERROR|D_CONSOLE, conn,
                            "Cant send to %s after timeout lapse of %lu; TO %lu\n",
                libcfs_nid2str(conn->gnc_peer->gnp_nid),
                cfs_duration_sec(now - newest_last_rx),
                cfs_duration_sec(GNILND_TIMEOUTRX(timeout)));
                kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
                kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
                return -ETIME;
        }

        GNITX_ASSERTF(tx, (conn != NULL) && (tx->tx_id.txe_idx != 0), "tx id unset!", NULL);
        /* msg->gnm_srcnid is set when the message is initialized by whatever function is
         * creating the message this allows the message to contain the correct LNET NID/NET needed
         * instead of the one that the peer/conn uses for sending the data.
         */
        msg->gnm_connstamp = conn->gnc_my_connstamp;
        msg->gnm_payload_len = immediatenob;
        kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex);
        msg->gnm_seq = atomic_read(&conn->gnc_tx_seq);

        /* always init here - kgn_checksum is a /sys module tunable
         * and can be flipped at any point, even between msg init and sending */
        msg->gnm_cksum = 0;
        if (*kgnilnd_tunables.kgn_checksum) {
                /* We must set here and not in kgnilnd_init_msg,
                 * we could resend this msg many times
                 * (NOT_DONE from gni_smsg_send below) and wouldn't pass
                 * through init_msg again */
                msg->gnm_cksum = kgnilnd_cksum(msg, sizeof(kgn_msg_t));
                if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SMSG_CKSUM1)) {
                        msg->gnm_cksum += 0xf00f;
                }
        }

        GNIDBG_TOMSG(D_NET, msg, "tx 0x%p conn 0x%p->%s sending SMSG sz %u id %x/%d [%p for %u]",
               tx, conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
               sizeof(kgn_msg_t), tx->tx_id.txe_smsg_id,
               tx->tx_id.txe_idx, immediate, immediatenob);

        if (unlikely(tx->tx_state & GNILND_TX_FAIL_SMSG)) {
                rrc = cfs_fail_val ? cfs_fail_val : GNI_RC_NOT_DONE;
        } else {
                rrc = kgnilnd_smsg_send(conn->gnc_ephandle,
                                        msg, sizeof(*msg), immediate,
                                        immediatenob,
                                        tx->tx_id.txe_smsg_id);
        }

        switch (rrc) {
        case GNI_RC_SUCCESS:
                atomic_inc(&conn->gnc_tx_seq);
                conn->gnc_last_tx = jiffies;
                /* no locking here as LIVE isn't a list */
                kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_LIVE_FMAQ, 1);

                /* this needs to be checked under lock as it might be freed from a completion
                 * event.
                 */
                if (msg->gnm_type == GNILND_MSG_NOOP) {
                        set_mb(conn->gnc_last_noop_sent, jiffies);
                }

                /* serialize with seeing CQ events for completion on this, as well as
                 * tx_seq */
                kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
                kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);

                atomic_inc(&conn->gnc_device->gnd_short_ntx);
                atomic64_add(immediatenob, &conn->gnc_device->gnd_short_txbytes);
                kgnilnd_peer_alive(conn->gnc_peer);
                GNIDBG_SMSG_CREDS(D_NET, conn);
                return 0;

        case GNI_RC_NOT_DONE:
                /* XXX Nic: We need to figure out how to track this
                 * - there are bound to be good reasons for it,
                 * but we want to know when it happens */
                kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
                kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
                /* We'll handle this error inline - makes the calling logic much more
                 * clean */

                /* If no lock, caller doesn't want us to retry */
                if (state_lock == NULL) {
                        return -EAGAIN;
                }

                retry_send = kgnilnd_tx_should_retry(conn, tx);
                if (retry_send) {
                        /* add to head of list for the state and retries */
                        spin_lock(state_lock);
                        kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, state, 0);
                        spin_unlock(state_lock);

                        /* We only reschedule for a certain number of retries, then
                         * we will wait for the CQ events indicating a release of SMSG
                         * credits */
                        if (tx->tx_retrans < (*kgnilnd_tunables.kgn_max_retransmits/4)) {
                                kgnilnd_schedule_conn(conn);
                                return 0;
                        } else {
                                /* CQ event coming in signifies either TX completed or
                                 * RX receive. Either of these *could* free up credits
                                 * in the SMSG mbox and we should try sending again */
                                GNIDBG_TX(D_NET, tx, "waiting for CQID %u event to resend",
                                         tx->tx_conn->gnc_cqid);
                                /* use +ve return code to let upper layers know they
                                 * should stop looping on sends */
                                return EAGAIN;
                        }
                } else {
                        return -EAGAIN;
                }
        default:
                /* handle bad retcode gracefully */
                kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);
                kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
                return -EIO;
        }
}

/* kgnilnd_sendmsg has hard wait on gnd_cq_mutex */
static inline int
kgnilnd_sendmsg(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
                spinlock_t *state_lock, kgn_tx_list_state_t state)
{
        kgn_device_t    *dev = tx->tx_conn->gnc_device;
        unsigned long    timestamp;
        int              rc;

        timestamp = jiffies;
        kgnilnd_gl_mutex_lock(&dev->gnd_cq_mutex);
        /* delay in jiffies - we are really concerned only with things that
         * result in a schedule() or really holding this off for long times .
         * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
        dev->gnd_mutex_delay += (long) jiffies - timestamp;

        rc = kgnilnd_sendmsg_nolock(tx, immediate, immediatenob, state_lock, state);

        RETURN(rc);
}


/* returns -EAGAIN for lock miss, anything else < 0 is hard error, >=0 for success */
static inline int
kgnilnd_sendmsg_trylock(kgn_tx_t *tx, void *immediate, unsigned int immediatenob,
                spinlock_t *state_lock, kgn_tx_list_state_t state)
{
        kgn_conn_t      *conn = tx->tx_conn;
        kgn_device_t    *dev = conn->gnc_device;
        unsigned long    timestamp;
        int              rc;

        timestamp = jiffies;

        /* technically we are doing bad things with the read_lock on the peer_conn
         * table, but we shouldn't be sleeping inside here - and we don't sleep/block
         * for the mutex. I bet lockdep is gonna flag this one though... */

        /* there are a few cases where we don't want the immediate send - like
         * when we are in the scheduler thread and it'd harm the latency of
         * getting messages up to LNet */

        /* rmb for gnd_ready */
        smp_rmb();
        if (conn->gnc_device->gnd_ready == GNILND_DEV_LOOP) {
                rc = 0;
                atomic_inc(&conn->gnc_device->gnd_fast_block);
        } else if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
               /* dont hit HW during quiesce */
                rc = 0;
        } else if (unlikely(atomic_read(&conn->gnc_peer->gnp_dirty_eps))) {
               /* dont hit HW if stale EPs and conns left to close */
                rc = 0;
        } else {
                atomic_inc(&conn->gnc_device->gnd_fast_try);
                rc = kgnilnd_gl_mutex_trylock(&conn->gnc_device->gnd_cq_mutex);
        }
        if (!rc) {
                rc = -EAGAIN;
        } else {
                /* we got the mutex and weren't blocked */

                /* delay in jiffies - we are really concerned only with things that
                 * result in a schedule() or really holding this off for long times .
                 * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
                dev->gnd_mutex_delay += (long) jiffies - timestamp;

                atomic_inc(&conn->gnc_device->gnd_fast_ok);
                tx->tx_qtime = jiffies;
                tx->tx_state = GNILND_TX_WAITING_COMPLETION;
                rc = kgnilnd_sendmsg_nolock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);
                /* _nolock unlocks the mutex for us */
        }

        RETURN(rc);
}

/* lets us know if we can push this RDMA through now */
inline int
kgnilnd_auth_rdma_bytes(kgn_device_t *dev, kgn_tx_t *tx)
{
        long    bytes_left;

        bytes_left = atomic64_sub_return(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);

        if (bytes_left < 0) {
                atomic64_add(tx->tx_nob, &dev->gnd_rdmaq_bytes_ok);
                atomic_inc(&dev->gnd_rdmaq_nstalls);
                smp_wmb();

                CDEBUG(D_NET, "no bytes to send, turning on timer for %lu\n",
                       dev->gnd_rdmaq_deadline);
                mod_timer(&dev->gnd_rdmaq_timer, dev->gnd_rdmaq_deadline);
                /* we never del this timer - at worst it schedules us.. */
                return -EAGAIN;
        } else {
                return 0;
        }
}

/* this adds a TX to the queue pending throttling authorization before
 * we allow our remote peer to launch a PUT at us */
void
kgnilnd_queue_rdma(kgn_conn_t *conn, kgn_tx_t *tx)
{
        int     rc;

        /* we cannot go into send_mapped_tx from here as we are holding locks
         * and mem registration might end up allocating memory in kgni.
         * That said, we'll push this as far as we can into the queue process */
        rc = kgnilnd_auth_rdma_bytes(conn->gnc_device, tx);

        if (rc < 0) {
                spin_lock(&conn->gnc_device->gnd_rdmaq_lock);
                kgnilnd_tx_add_state_locked(tx, NULL, conn, GNILND_TX_RDMAQ, 0);
                /* lets us know how delayed RDMA is */
                tx->tx_qtime = jiffies;
                spin_unlock(&conn->gnc_device->gnd_rdmaq_lock);
        } else {
                /* we have RDMA authorized, now it just needs a MDD and to hit the wire */
                spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
                kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 0);
                /* lets us know how delayed mapping is */
                tx->tx_qtime = jiffies;
                spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
        }

        /* make sure we wake up sched to run this */
        kgnilnd_schedule_device(tx->tx_conn->gnc_device);
}

/* push TX through state machine */
void
kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx)
{
        int            rc = 0;
        int            add_tail = 1;

        /* set the tx_id here, we delay it until we have an actual conn
         * to fiddle with
         * in some cases, the tx_id is already set to provide for things
         * like RDMA completion cookies, etc */
        if (tx->tx_id.txe_idx == 0) {
                rc = kgnilnd_set_tx_id(tx, conn);
                if (rc != 0) {
                        kgnilnd_tx_done(tx, rc);
                        return;
                }
        }

        CDEBUG(D_NET, "%s to conn %p for %s\n", kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
                conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));

        /* Only let NOOPs to be sent while fail loc is set, otherwise kill the tx.
         */
        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ONLY_NOOP) && (tx->tx_msg.gnm_type != GNILND_MSG_NOOP)) {
                kgnilnd_tx_done(tx, rc);
                return;
        }

        switch (tx->tx_msg.gnm_type) {
        case GNILND_MSG_PUT_ACK:
        case GNILND_MSG_GET_REQ:
        case GNILND_MSG_PUT_REQ_REV:
        case GNILND_MSG_GET_ACK_REV:
                /* hijacking time! If this messages will authorize our peer to
                 * send his dirty little bytes in an RDMA, we need to get permission */
                kgnilnd_queue_rdma(conn, tx);
                break;
        case GNILND_MSG_IMMEDIATE:
                /* try to send right now, can help reduce latency */
                rc = kgnilnd_sendmsg_trylock(tx, tx->tx_buffer, tx->tx_nob, &conn->gnc_list_lock, GNILND_TX_FMAQ);

                if (rc >= 0) {
                        /* it was sent, break out of switch to avoid default case of queueing */
                        break;
                }
                /* needs to queue to try again, so fall through to default case */
        case GNILND_MSG_NOOP:
                /* Just make sure this goes out first for this conn */
                add_tail = 0;
                /* fall through... */
        default:
                spin_lock(&conn->gnc_list_lock);
                kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_FMAQ, add_tail);
                tx->tx_qtime = jiffies;
                spin_unlock(&conn->gnc_list_lock);
                kgnilnd_schedule_conn(conn);
        }
}

void
kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target)
{
        kgn_peer_t      *peer;
        kgn_peer_t      *new_peer = NULL;
        kgn_conn_t      *conn = NULL;
        int              rc;
        int              node_state;

        ENTRY;

        /* If I get here, I've committed to send, so I complete the tx with
         * failure on any problems */

        GNITX_ASSERTF(tx, tx->tx_conn == NULL,
                      "tx already has connection %p", tx->tx_conn);

        /* do all of the peer & conn searching in one swoop - this avoids
         * nastiness when dropping locks and needing to maintain a sane state
         * in the face of stack reset or something else nuking peers & conns */

        /* I expect to find him, so only take a read lock */
        read_lock(&kgnilnd_data.kgn_peer_conn_lock);

        peer = kgnilnd_find_peer_locked(target->nid);
        if (peer != NULL) {
                conn = kgnilnd_find_conn_locked(peer);
                /* this could be NULL during quiesce */
                if (conn != NULL)  {
                        /* Connection exists; queue message on it */
                        kgnilnd_queue_tx(conn, tx);
                        read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                        RETURN_EXIT;
                }

                /* don't create a connection if the peer is marked down */
                if (peer->gnp_down == GNILND_RCA_NODE_DOWN) {
                        read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                        rc = -ENETRESET;
                        GOTO(no_peer, rc);
                }
        }

        /* creating peer or conn; I'll need a write lock... */
        read_unlock(&kgnilnd_data.kgn_peer_conn_lock);

        CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);

        node_state = kgnilnd_get_node_state(LNET_NIDADDR(target->nid));

        /* NB - this will not block during normal operations -
         * the only writer of this is in the startup/shutdown path. */
        rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
        if (!rc) {
                rc = -ESHUTDOWN;
                GOTO(no_peer, rc);
        }

        /* ignore previous peer entirely - we cycled the lock, so we
         * will create new peer and at worst drop it if peer is still
         * in the tables */
        rc = kgnilnd_create_peer_safe(&new_peer, target->nid, net, node_state);
        if (rc != 0) {
                up_read(&kgnilnd_data.kgn_net_rw_sem);
                GOTO(no_peer, rc);
        }

        write_lock(&kgnilnd_data.kgn_peer_conn_lock);
        up_read(&kgnilnd_data.kgn_net_rw_sem);

        /* search for peer again now that we have the lock
         * if we don't find it, add our new one to the list */
        kgnilnd_add_peer_locked(target->nid, new_peer, &peer);

        /* don't create a connection if the peer is not up */
        if (peer->gnp_down != GNILND_RCA_NODE_UP) {
                write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                rc = -ENETRESET;
                GOTO(no_peer, rc);
        }

        conn = kgnilnd_find_or_create_conn_locked(peer);

        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DGRAM_DROP_TX)) {
                write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                GOTO(no_peer, rc);
        }

        if (conn != NULL) {
                /* oh hey, found a conn now... magical */
                kgnilnd_queue_tx(conn, tx);
        } else {
                /* no conn, must be trying to connect - so we queue for now */
                tx->tx_qtime = jiffies;
                kgnilnd_tx_add_state_locked(tx, peer, NULL, GNILND_TX_PEERQ, 1);
        }
        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
        RETURN_EXIT;
no_peer:
        kgnilnd_tx_done(tx, rc);
        RETURN_EXIT;
}

int
kgnilnd_rdma(kgn_tx_t *tx, int type,
            kgn_rdma_desc_t *sink, unsigned int nob, __u64 cookie)
{
        kgn_conn_t   *conn = tx->tx_conn;
        unsigned long timestamp;
        gni_post_type_t post_type;
        gni_return_t  rrc;
        int rc = 0;
        unsigned int desc_nob = nob;
        void *desc_buffer = tx->tx_buffer;
        gni_mem_handle_t desc_map_key = tx->tx_map_key;
        LASSERTF(kgnilnd_tx_mapped(tx),
                "unmapped tx %p\n", tx);
        LASSERTF(conn != NULL,
                "NULL conn on tx %p, naughty, naughty\n", tx);
        LASSERTF(nob <= sink->gnrd_nob,
                "nob %u > sink->gnrd_nob %d (%p)\n",
                nob, sink->gnrd_nob, sink);
        LASSERTF(nob <= tx->tx_nob,
                "nob %d > tx(%p)->tx_nob %d\n",
                nob, tx, tx->tx_nob);

        switch (type) {
        case GNILND_MSG_GET_DONE:
        case GNILND_MSG_PUT_DONE:
                post_type = GNI_POST_RDMA_PUT;
                break;
        case GNILND_MSG_GET_DONE_REV:
        case GNILND_MSG_PUT_DONE_REV:
                post_type = GNI_POST_RDMA_GET;
                break;
        default:
                CERROR("invalid msg type %s (%d)\n",
                        kgnilnd_msgtype2str(type), type);
                LBUG();
        }
        if (post_type == GNI_POST_RDMA_GET) {
                /* Check for remote buffer / local buffer / length alignment. All must be 4 byte
                 * aligned. If the local buffer is not aligned correctly using the copy buffer
                 * will fix that issue. If length is misaligned copy buffer will also fix the issue, we end
                 * up transferring extra bytes into the buffer but only copy the correct nob into the original
                 * buffer.  Remote offset correction is done through a combination of adjusting the offset,
                 * making sure the length and addr are aligned and copying the data into the correct location
                 * once the transfer has completed.
                 */
                if ((((__u64)((unsigned long)tx->tx_buffer)) & 3) ||
                      (sink->gnrd_addr & 3) ||
                      (nob & 3)) {

                        tx->tx_offset = ((__u64)((unsigned long)sink->gnrd_addr)) & 3;
                        if (tx->tx_offset)
                                kgnilnd_admin_addref(kgnilnd_data.kgn_rev_offset);

                        if ((nob + tx->tx_offset) & 3) {
                                desc_nob = ((nob + tx->tx_offset) + (4 - ((nob + tx->tx_offset) & 3)));
                                kgnilnd_admin_addref(kgnilnd_data.kgn_rev_length);
                        } else {
                                desc_nob = (nob + tx->tx_offset);
                        }

                        if (tx->tx_buffer_copy == NULL) {
                                /* Allocate the largest copy buffer we will need, this will prevent us from overwriting data
                                 * and require at most we allocate a few extra bytes. */
                                tx->tx_buffer_copy = vmalloc(desc_nob);

                                if (!tx->tx_buffer_copy) {
                                        /* allocation of buffer failed nak the rdma */
                                        kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
                                        kgnilnd_tx_done(tx, -EFAULT);
                                        return 0;
                                }
                                kgnilnd_admin_addref(kgnilnd_data.kgn_rev_copy_buff);
                                rc = kgnilnd_mem_register(conn->gnc_device->gnd_handle, (__u64)tx->tx_buffer_copy, desc_nob, NULL, GNI_MEM_READWRITE, &tx->tx_buffer_copy_map_key);
                                if (rc != GNI_RC_SUCCESS) {
                                        /* Registration Failed nak rdma and kill the tx. */
                                        vfree(tx->tx_buffer_copy);
                                        tx->tx_buffer_copy = NULL;
                                        kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
                                        kgnilnd_tx_done(tx, -EFAULT);
                                        return 0;
                                }
                        }
                        desc_map_key = tx->tx_buffer_copy_map_key;
                        desc_buffer = tx->tx_buffer_copy;
                }
        }

        memset(&tx->tx_rdma_desc, 0, sizeof(tx->tx_rdma_desc));
        tx->tx_rdma_desc.post_id = tx->tx_id.txe_cookie;
        tx->tx_rdma_desc.type = post_type;
        tx->tx_rdma_desc.cq_mode = GNI_CQMODE_GLOBAL_EVENT;
        tx->tx_rdma_desc.local_addr = (__u64)((unsigned long)desc_buffer);
        tx->tx_rdma_desc.local_mem_hndl = desc_map_key;
        tx->tx_rdma_desc.remote_addr = sink->gnrd_addr - tx->tx_offset;
        tx->tx_rdma_desc.remote_mem_hndl = sink->gnrd_key;
        tx->tx_rdma_desc.length = desc_nob;
        tx->tx_nob_rdma = nob;
        if (*kgnilnd_tunables.kgn_bte_dlvr_mode)
                tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_dlvr_mode;
        /* prep final completion message */
        kgnilnd_init_msg(&tx->tx_msg, type, tx->tx_msg.gnm_srcnid);
        tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
        /* send actual size RDMA'd in retval */
        tx->tx_msg.gnm_u.completion.gncm_retval = nob;

        kgnilnd_compute_rdma_cksum(tx, nob);

        if (nob == 0) {
                kgnilnd_queue_tx(conn, tx);
                return 0;
        }

        /* Don't lie (CLOSE == RDMA idle) */
        LASSERTF(!conn->gnc_close_sent, "tx %p on conn %p after close sent %d\n",
                 tx, conn, conn->gnc_close_sent);

        GNIDBG_TX(D_NET, tx, "Post RDMA type 0x%02x conn %p dlvr_mode "
                "0x%x cookie:"LPX64,
                type, conn, tx->tx_rdma_desc.dlvr_mode, cookie);

        /* set CQ dedicated for RDMA */
        tx->tx_rdma_desc.src_cq_hndl = conn->gnc_device->gnd_snd_rdma_cqh;

        timestamp = jiffies;
        kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
        kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
        /* delay in jiffies - we are really concerned only with things that
         * result in a schedule() or really holding this off for long times .
         * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
        conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;

        rrc = kgnilnd_post_rdma(conn->gnc_ephandle, &tx->tx_rdma_desc);

        if (rrc == GNI_RC_ERROR_RESOURCE) {
                kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);
                kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
                kgnilnd_unmap_buffer(tx, 0);

                if (tx->tx_buffer_copy != NULL) {
                        vfree(tx->tx_buffer_copy);
                        tx->tx_buffer_copy = NULL;
                }

                spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
                kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn,
                                            GNILND_TX_MAPQ, 0);
                spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
                kgnilnd_schedule_device(tx->tx_conn->gnc_device);
                return -EAGAIN;
        }

        spin_lock(&conn->gnc_list_lock);
        kgnilnd_tx_add_state_locked(tx, conn->gnc_peer, conn, GNILND_TX_LIVE_RDMAQ, 1);
        tx->tx_qtime = jiffies;
        spin_unlock(&conn->gnc_list_lock);
        kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);
        kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);

        /* XXX Nic: is this a place we should handle more errors for
         * robustness sake */
        LASSERT(rrc == GNI_RC_SUCCESS);
        return 0;
}

kgn_rx_t *
kgnilnd_alloc_rx(void)
{
        kgn_rx_t        *rx;

        rx = kmem_cache_alloc(kgnilnd_data.kgn_rx_cache, GFP_ATOMIC);
        if (rx == NULL) {
                CERROR("failed to allocate rx\n");
                return NULL;
        }
        CDEBUG(D_MALLOC, "slab-alloced 'rx': %lu at %p.\n",
               sizeof(*rx), rx);

        /* no memset to zero, we'll always fill all members */
        return rx;
}

/* release is to just free connection resources
 * we use this for the eager path after copying */
void
kgnilnd_release_msg(kgn_conn_t *conn)
{
        gni_return_t    rrc;
        unsigned long   timestamp;

        CDEBUG(D_NET, "consuming %p\n", conn);

        timestamp = jiffies;
        kgnilnd_gl_mutex_lock(&conn->gnc_device->gnd_cq_mutex);
        /* delay in jiffies - we are really concerned only with things that
         * result in a schedule() or really holding this off for long times .
         * NB - mutex_lock could spin for 2 jiffies before going to sleep to wait */
        conn->gnc_device->gnd_mutex_delay += (long) jiffies - timestamp;

        rrc = kgnilnd_smsg_release(conn->gnc_ephandle);
        kgnilnd_gl_mutex_unlock(&conn->gnc_device->gnd_cq_mutex);

        LASSERTF(rrc == GNI_RC_SUCCESS, "bad rrc %d\n", rrc);
        GNIDBG_SMSG_CREDS(D_NET, conn);

        return;
}

void
kgnilnd_consume_rx(kgn_rx_t *rx)
{
        kgn_conn_t      *conn = rx->grx_conn;
        kgn_msg_t       *rxmsg = rx->grx_msg;

        /* if we are eager, free the cache alloc'd msg */
        if (unlikely(rx->grx_eager)) {
                LIBCFS_FREE(rxmsg, sizeof(*rxmsg) + *kgnilnd_tunables.kgn_max_immediate);
                atomic_dec(&kgnilnd_data.kgn_neager_allocs);

                /* release ref from eager_recv */
                kgnilnd_conn_decref(conn);
        } else {
                GNIDBG_MSG(D_NET, rxmsg, "rx %p processed", rx);
                kgnilnd_release_msg(conn);
        }

        kmem_cache_free(kgnilnd_data.kgn_rx_cache, rx);
        CDEBUG(D_MALLOC, "slab-freed 'rx': %lu at %p.\n",
               sizeof(*rx), rx);

        return;
}

int
kgnilnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
{
        lnet_hdr_t       *hdr = &lntmsg->msg_hdr;
        int               type = lntmsg->msg_type;
        lnet_process_id_t target = lntmsg->msg_target;
        int               target_is_router = lntmsg->msg_target_is_router;
        int               routing = lntmsg->msg_routing;
        unsigned int      niov = lntmsg->msg_niov;
        struct iovec     *iov = lntmsg->msg_iov;
        lnet_kiov_t      *kiov = lntmsg->msg_kiov;
        unsigned int      offset = lntmsg->msg_offset;
        unsigned int      nob = lntmsg->msg_len;
        unsigned int      msg_vmflush = lntmsg->msg_vmflush;
        kgn_net_t        *net = ni->ni_data;
        kgn_tx_t         *tx;
        int               rc = 0;
        int               mpflag = 0;
        int               reverse_rdma_flag = *kgnilnd_tunables.kgn_reverse_rdma;

        /* NB 'private' is different depending on what we're sending.... */
        LASSERT(!in_interrupt());

        CDEBUG(D_NET, "sending msg type %d with %d bytes in %d frags to %s\n",
               type, nob, niov, libcfs_id2str(target));

        LASSERTF(nob == 0 || niov > 0,
                "lntmsg %p nob %d niov %d\n", lntmsg, nob, niov);
        LASSERTF(niov <= LNET_MAX_IOV,
                "lntmsg %p niov %d\n", lntmsg, niov);

        /* payload is either all vaddrs or all pages */
        LASSERTF(!(kiov != NULL && iov != NULL),
                "lntmsg %p kiov %p iov %p\n", lntmsg, kiov, iov);

        if (msg_vmflush)
                mpflag = cfs_memory_pressure_get_and_set();

        switch (type) {
        default:
                CERROR("lntmsg %p with unexpected type %d\n",
                        lntmsg, type);
                LBUG();

        case LNET_MSG_ACK:
                LASSERTF(nob == 0, "lntmsg %p nob %d\n",
                        lntmsg, nob);
                break;

        case LNET_MSG_GET:
                LASSERT(niov == 0);
                LASSERT(nob == 0);

                if (routing || target_is_router)
                        break;                  /* send IMMEDIATE */

                /* it is safe to do direct GET with out mapping buffer for RDMA as we
                 * check the eventual sink buffer here - if small enough, remote
                 * end is perfectly capable of returning data in short message -
                 * The magic is that we call lnet_parse in kgnilnd_recv with rdma_req=0
                 * for IMMEDIATE messages which will have it send a real reply instead
                 * of doing kgnilnd_recv to have the RDMA continued */
                if (lntmsg->msg_md->md_length <= *kgnilnd_tunables.kgn_max_immediate)
                       break;

                if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
                        tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ, ni->ni_nid);
                else
                        tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_REQ_REV, ni->ni_nid);

                if (tx == NULL) {
                        rc = -ENOMEM;
                        goto out;
                }
                /* slightly different options as we might actually have a GET with a
                 * MD_KIOV set but a non-NULL md_iov.iov */
                if ((lntmsg->msg_md->md_options & LNET_MD_KIOV) == 0)
                        rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
                                                      lntmsg->msg_md->md_iov.iov, NULL,
                                                      0, lntmsg->msg_md->md_length);
                else
                        rc = kgnilnd_setup_rdma_buffer(tx, lntmsg->msg_md->md_niov,
                                                      NULL, lntmsg->msg_md->md_iov.kiov,
                                                      0, lntmsg->msg_md->md_length);
                if (rc != 0) {
                        CERROR("unable to setup buffer: %d\n", rc);
                        kgnilnd_tx_done(tx, rc);
                        rc = -EIO;
                        goto out;
                }

                tx->tx_lntmsg[1] = lnet_create_reply_msg(ni, lntmsg);
                if (tx->tx_lntmsg[1] == NULL) {
                        CERROR("Can't create reply for GET to %s\n",
                               libcfs_nid2str(target.nid));
                        kgnilnd_tx_done(tx, rc);
                        rc = -EIO;
                        goto out;
                }

                tx->tx_lntmsg[0] = lntmsg;
                if ((reverse_rdma_flag & GNILND_REVERSE_GET) == 0)
                        tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;
                else
                        tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;

                /* rest of tx_msg is setup just before it is sent */
                kgnilnd_launch_tx(tx, net, &target);
                goto out;
        case LNET_MSG_REPLY:
        case LNET_MSG_PUT:
                /* to save on MDDs, we'll handle short kiov by vmap'ing
                 * and sending via SMSG */
                if (nob <= *kgnilnd_tunables.kgn_max_immediate)
                       break;

                if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
                        tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ, ni->ni_nid);
                else
                        tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_REQ_REV, ni->ni_nid);

                if (tx == NULL) {
                        rc = -ENOMEM;
                        goto out;
                }

                rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, nob);
                if (rc != 0) {
                        kgnilnd_tx_done(tx, rc);
                        rc = -EIO;
                        goto out;
                }

                tx->tx_lntmsg[0] = lntmsg;
                if ((reverse_rdma_flag & GNILND_REVERSE_PUT) == 0)
                        tx->tx_msg.gnm_u.putreq.gnprm_hdr = *hdr;
                else
                        tx->tx_msg.gnm_u.get.gngm_hdr = *hdr;

                /* rest of tx_msg is setup just before it is sent */
                kgnilnd_launch_tx(tx, net, &target);
                goto out;
        }

        /* send IMMEDIATE */

        LASSERTF(nob <= *kgnilnd_tunables.kgn_max_immediate,
                "lntmsg 0x%p too large %d\n", lntmsg, nob);

        tx = kgnilnd_new_tx_msg(GNILND_MSG_IMMEDIATE, ni->ni_nid);
        if (tx == NULL) {
                rc = -ENOMEM;
                goto out;
        }

        rc = kgnilnd_setup_immediate_buffer(tx, niov, iov, kiov, offset, nob);
        if (rc != 0) {
                kgnilnd_tx_done(tx, rc);
                goto out;
        }

        tx->tx_msg.gnm_u.immediate.gnim_hdr = *hdr;
        tx->tx_lntmsg[0] = lntmsg;
        kgnilnd_launch_tx(tx, net, &target);

out:
        /* use stored value as we could have already finalized lntmsg here from a failed launch */
        if (msg_vmflush)
                cfs_memory_pressure_restore(mpflag);
        return rc;
}

void
kgnilnd_setup_rdma(lnet_ni_t *ni, kgn_rx_t *rx, lnet_msg_t *lntmsg, int mlen)
{
        kgn_conn_t    *conn = rx->grx_conn;
        kgn_msg_t     *rxmsg = rx->grx_msg;
        unsigned int   niov = lntmsg->msg_niov;
        struct iovec  *iov = lntmsg->msg_iov;
        lnet_kiov_t   *kiov = lntmsg->msg_kiov;
        unsigned int   offset = lntmsg->msg_offset;
        unsigned int   nob = lntmsg->msg_len;
        int            done_type;
        kgn_tx_t      *tx;
        int            rc = 0;

        switch (rxmsg->gnm_type) {
        case GNILND_MSG_PUT_REQ_REV:
                done_type = GNILND_MSG_PUT_DONE_REV;
                nob = mlen;
                break;
        case GNILND_MSG_GET_REQ:
                done_type = GNILND_MSG_GET_DONE;
                break;
        default:
                CERROR("invalid msg type %s (%d)\n",
                        kgnilnd_msgtype2str(rxmsg->gnm_type),
                        rxmsg->gnm_type);
                LBUG();
        }

        tx = kgnilnd_new_tx_msg(done_type, ni->ni_nid);
        if (tx == NULL)
                goto failed_0;

        rc = kgnilnd_set_tx_id(tx, conn);
        if (rc != 0)
                goto failed_1;

        rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, nob);
        if (rc != 0)
                goto failed_1;

        tx->tx_lntmsg[0] = lntmsg;
        tx->tx_getinfo = rxmsg->gnm_u.get;

        /* we only queue from kgnilnd_recv - we might get called from other contexts
         * and we don't want to block the mutex in those cases */

        spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
        kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
        spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
        kgnilnd_schedule_device(tx->tx_conn->gnc_device);

        return;

 failed_1:
        kgnilnd_tx_done(tx, rc);
        kgnilnd_nak_rdma(conn, done_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
 failed_0:
        lnet_finalize(ni, lntmsg, rc);
}

int
kgnilnd_eager_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
                   void **new_private)
{
        kgn_rx_t        *rx = private;
        kgn_conn_t      *conn = rx->grx_conn;
        kgn_msg_t       *rxmsg = rx->grx_msg;
        kgn_msg_t       *eagermsg = NULL;
        kgn_peer_t      *peer = NULL;
        kgn_conn_t      *found_conn = NULL;

        GNIDBG_MSG(D_NET, rxmsg, "eager recv for conn %p, rxmsg %p, lntmsg %p",
                conn, rxmsg, lntmsg);

        if (rxmsg->gnm_payload_len > *kgnilnd_tunables.kgn_max_immediate) {
                GNIDBG_MSG(D_ERROR, rxmsg, "payload too large %d",
                        rxmsg->gnm_payload_len);
                return -EPROTO;
        }
        /* Grab a read lock so the connection doesnt disappear on us
         * while we look it up
         */
        read_lock(&kgnilnd_data.kgn_peer_conn_lock);

        peer = kgnilnd_find_peer_locked(rxmsg->gnm_srcnid);
        if (peer != NULL)
                found_conn = kgnilnd_find_conn_locked(peer);


        /* Verify the connection found is the same one that the message
         * is supposed to be using, if it is not output an error message
         * and return.
         */
        if (!peer || !found_conn
            || found_conn->gnc_peer_connstamp != rxmsg->gnm_connstamp) {
                read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                CERROR("Couldnt find matching peer %p or conn %p / %p\n",
                        peer, conn, found_conn);
                if (found_conn) {
                        CERROR("Unexpected connstamp "LPX64"("LPX64" expected)"
                                " from %s", rxmsg->gnm_connstamp,
                                found_conn->gnc_peer_connstamp,
                                libcfs_nid2str(peer->gnp_nid));
                }
                return -ENOTCONN;
        }

        /* add conn ref to ensure it doesn't go away until all eager
         * messages processed */
        kgnilnd_conn_addref(conn);

        /* Now that we have verified the connection is valid and added a
         * reference we can remove the read_lock on the peer_conn_lock */
        read_unlock(&kgnilnd_data.kgn_peer_conn_lock);

        /* we have no credits or buffers for this message, so copy it
         * somewhere for a later kgnilnd_recv */
        if (atomic_read(&kgnilnd_data.kgn_neager_allocs) >=
                        *kgnilnd_tunables.kgn_eager_credits) {
                CERROR("Out of eager credits to %s\n",
                        libcfs_nid2str(conn->gnc_peer->gnp_nid));
                return -ENOMEM;
        }

        atomic_inc(&kgnilnd_data.kgn_neager_allocs);

        LIBCFS_ALLOC(eagermsg, sizeof(*eagermsg) + *kgnilnd_tunables.kgn_max_immediate);
        if (eagermsg == NULL) {
                kgnilnd_conn_decref(conn);
                CERROR("couldn't allocate eager rx message for conn %p to %s\n",
                        conn, libcfs_nid2str(conn->gnc_peer->gnp_nid));
                return -ENOMEM;
        }

        /* copy msg and payload */
        memcpy(eagermsg, rxmsg, sizeof(*rxmsg) + rxmsg->gnm_payload_len);
        rx->grx_msg = eagermsg;
        rx->grx_eager = 1;

        /* stash this for lnet_finalize on cancel-on-conn-close */
        rx->grx_lntmsg = lntmsg;

        /* keep the same rx_t, it just has a new grx_msg now */
        *new_private = private;

        /* release SMSG buffer */
        kgnilnd_release_msg(conn);

        return 0;
}

int
kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
             int delayed, unsigned int niov,
             struct iovec *iov, lnet_kiov_t *kiov,
             unsigned int offset, unsigned int mlen, unsigned int rlen)
{
        kgn_rx_t    *rx = private;
        kgn_conn_t  *conn = rx->grx_conn;
        kgn_msg_t   *rxmsg = rx->grx_msg;
        kgn_tx_t    *tx;
        int          rc = 0;
        __u32        pload_cksum;
        ENTRY;

        LASSERT(!in_interrupt());
        LASSERTF(mlen <= rlen, "%d <= %d\n", mlen, rlen);
        /* Either all pages or all vaddrs */
        LASSERTF(!(kiov != NULL && iov != NULL), "kiov %p iov %p\n",
                kiov, iov);

        GNIDBG_MSG(D_NET, rxmsg, "conn %p, rxmsg %p, lntmsg %p"
                " niov=%d kiov=%p iov=%p offset=%d mlen=%d rlen=%d",
                conn, rxmsg, lntmsg,
                niov, kiov, iov, offset, mlen, rlen);

        /* we need to lock here as recv can be called from any context */
        read_lock(&kgnilnd_data.kgn_peer_conn_lock);
        if (rx->grx_eager && conn->gnc_state != GNILND_CONN_ESTABLISHED) {
                read_unlock(&kgnilnd_data.kgn_peer_conn_lock);

                /* someone closed the conn after we copied this out, nuke it */
                kgnilnd_consume_rx(rx);
                lnet_finalize(ni, lntmsg, conn->gnc_error);
                RETURN(0);
        }
        read_unlock(&kgnilnd_data.kgn_peer_conn_lock);

        switch (rxmsg->gnm_type) {
        default:
                GNIDBG_MSG(D_NETERROR, rxmsg, "conn %p, rx %p, rxmsg %p, lntmsg %p"
                " niov=%d kiov=%p iov=%p offset=%d mlen=%d rlen=%d",
                conn, rx, rxmsg, lntmsg, niov, kiov, iov, offset, mlen, rlen);
                LBUG();

        case GNILND_MSG_IMMEDIATE:
                if (mlen > rxmsg->gnm_payload_len) {
                        GNIDBG_MSG(D_ERROR, rxmsg,
                                "Immediate message from %s too big: %d > %d",
                                libcfs_nid2str(conn->gnc_peer->gnp_nid), mlen,
                                rxmsg->gnm_payload_len);
                        rc = -EINVAL;
                        kgnilnd_consume_rx(rx);
                        RETURN(rc);
                }

                /* rxmsg[1] is a pointer to the payload, sitting in the buffer
                 * right after the kgn_msg_t header - so just 'cute' way of saying
                 * rxmsg + sizeof(kgn_msg_t) */

                /* check payload checksum if sent */

                if (*kgnilnd_tunables.kgn_checksum >= 2 &&
                        !rxmsg->gnm_payload_cksum &&
                        rxmsg->gnm_payload_len != 0)
                        GNIDBG_MSG(D_WARNING, rxmsg, "no msg payload checksum when enabled");

                if (rxmsg->gnm_payload_cksum != 0) {
                        /* gnm_payload_len set in kgnilnd_sendmsg from tx->tx_nob,
                         * which is what is used to calculate the cksum on the TX side */
                        pload_cksum = kgnilnd_cksum(&rxmsg[1], rxmsg->gnm_payload_len);

                        if (rxmsg->gnm_payload_cksum != pload_cksum) {
                                GNIDBG_MSG(D_NETERROR, rxmsg,
                                           "Bad payload checksum (%x expected %x)",
                                            pload_cksum, rxmsg->gnm_payload_cksum);
                                switch (*kgnilnd_tunables.kgn_checksum_dump) {
                                case 2:
                                        kgnilnd_dump_blob(D_BUFFS, "bad payload checksum",
                                                          &rxmsg[1], rxmsg->gnm_payload_len);
                                        /* fall through to dump */
                                case 1:
                                        libcfs_debug_dumplog();
                                        break;
                                default:
                                        break;
                                }
                                rc = -ENOKEY;
                                /* checksum problems are fatal, kill the conn */
                                kgnilnd_consume_rx(rx);
                                kgnilnd_close_conn(conn, rc);
                                RETURN(rc);
                        }
                }

                if (kiov != NULL)
                        lnet_copy_flat2kiov(
                                niov, kiov, offset,
                                *kgnilnd_tunables.kgn_max_immediate,
                                &rxmsg[1], 0, mlen);
                else
                        lnet_copy_flat2iov(
                                niov, iov, offset,
                                *kgnilnd_tunables.kgn_max_immediate,
                                &rxmsg[1], 0, mlen);

                kgnilnd_consume_rx(rx);
                lnet_finalize(ni, lntmsg, 0);
                RETURN(0);

        case GNILND_MSG_PUT_REQ:
                /* LNET wants to truncate or drop transaction, sending NAK */
                if (mlen == 0) {
                        kgnilnd_consume_rx(rx);
                        lnet_finalize(ni, lntmsg, 0);

                        /* only error if lntmsg == NULL, otherwise we are just
                         * short circuiting the rdma process of 0 bytes */
                        kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
                                        lntmsg == NULL ? -ENOENT : 0,
                                        rxmsg->gnm_u.get.gngm_cookie,
                                        ni->ni_nid);
                        RETURN(0);
                }
                /* sending ACK with sink buff. info */
                tx = kgnilnd_new_tx_msg(GNILND_MSG_PUT_ACK, ni->ni_nid);
                if (tx == NULL) {
                        kgnilnd_consume_rx(rx);
                        RETURN(-ENOMEM);
                }

                rc = kgnilnd_set_tx_id(tx, conn);
                if (rc != 0) {
                        GOTO(nak_put_req, rc);
                }

                rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, mlen);
                if (rc != 0) {
                        GOTO(nak_put_req, rc);
                }

                tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
                        rxmsg->gnm_u.putreq.gnprm_cookie;
                tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
                tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
                        (__u64)((unsigned long)tx->tx_buffer);
                tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;

                tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
                tx->tx_qtime = jiffies;
                /* we only queue from kgnilnd_recv - we might get called from other contexts
                 * and we don't want to block the mutex in those cases */

                spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
                kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
                spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
                kgnilnd_schedule_device(tx->tx_conn->gnc_device);

                kgnilnd_consume_rx(rx);
                RETURN(0);

nak_put_req:
                /* make sure we send an error back when the PUT fails */
                kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
                kgnilnd_tx_done(tx, rc);
                kgnilnd_consume_rx(rx);

                /* return magic LNet network error */
                RETURN(-EIO);
        case GNILND_MSG_GET_REQ_REV:
                /* LNET wants to truncate or drop transaction, sending NAK */
                if (mlen == 0) {
                        kgnilnd_consume_rx(rx);
                        lnet_finalize(ni, lntmsg, 0);

                        /* only error if lntmsg == NULL, otherwise we are just
                         * short circuiting the rdma process of 0 bytes */
                        kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
                                        lntmsg == NULL ? -ENOENT : 0,
                                        rxmsg->gnm_u.get.gngm_cookie,
                                        ni->ni_nid);
                        RETURN(0);
                }
                /* lntmsg can be null when parsing a LNET_GET */
                if (lntmsg != NULL) {
                        /* sending ACK with sink buff. info */
                        tx = kgnilnd_new_tx_msg(GNILND_MSG_GET_ACK_REV, ni->ni_nid);
                        if (tx == NULL) {
                                kgnilnd_consume_rx(rx);
                                RETURN(-ENOMEM);
                        }

                        rc = kgnilnd_set_tx_id(tx, conn);
                        if (rc != 0)
                                GOTO(nak_get_req_rev, rc);


                        rc = kgnilnd_setup_rdma_buffer(tx, niov, iov, kiov, offset, mlen);
                        if (rc != 0)
                                GOTO(nak_get_req_rev, rc);


                        tx->tx_msg.gnm_u.putack.gnpam_src_cookie =
                                rxmsg->gnm_u.putreq.gnprm_cookie;
                        tx->tx_msg.gnm_u.putack.gnpam_dst_cookie = tx->tx_id.txe_cookie;
                        tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_addr =
                                (__u64)((unsigned long)tx->tx_buffer);
                        tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;

                        tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */

                        /* we only queue from kgnilnd_recv - we might get called from other contexts
                         * and we don't want to block the mutex in those cases */

                        spin_lock(&tx->tx_conn->gnc_device->gnd_lock);
                        kgnilnd_tx_add_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_MAPQ, 1);
                        spin_unlock(&tx->tx_conn->gnc_device->gnd_lock);
                        kgnilnd_schedule_device(tx->tx_conn->gnc_device);
                } else {
                        /* No match */
                        kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
                                        -ENOENT,
                                        rxmsg->gnm_u.get.gngm_cookie,
                                        ni->ni_nid);
                }

                kgnilnd_consume_rx(rx);
                RETURN(0);

nak_get_req_rev:
                /* make sure we send an error back when the GET fails */
                kgnilnd_nak_rdma(conn, rxmsg->gnm_type, rc, rxmsg->gnm_u.get.gngm_cookie, ni->ni_nid);
                kgnilnd_tx_done(tx, rc);
                kgnilnd_consume_rx(rx);

                /* return magic LNet network error */
                RETURN(-EIO);


        case GNILND_MSG_PUT_REQ_REV:
                /* LNET wants to truncate or drop transaction, sending NAK */
                if (mlen == 0) {
                        kgnilnd_consume_rx(rx);
                        lnet_finalize(ni, lntmsg, 0);

                        /* only error if lntmsg == NULL, otherwise we are just
                         * short circuiting the rdma process of 0 bytes */
                        kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
                                        lntmsg == NULL ? -ENOENT : 0,
                                        rxmsg->gnm_u.get.gngm_cookie,
                                        ni->ni_nid);
                        RETURN(0);
                }

                if (lntmsg != NULL) {
                        /* Matched! */
                        kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
                } else {
                        /* No match */
                        kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
                                        -ENOENT,
                                        rxmsg->gnm_u.get.gngm_cookie,
                                        ni->ni_nid);
                }
                kgnilnd_consume_rx(rx);
                RETURN(0);
        case GNILND_MSG_GET_REQ:
                if (lntmsg != NULL) {
                        /* Matched! */
                        kgnilnd_setup_rdma(ni, rx, lntmsg, mlen);
                } else {
                        /* No match */
                        kgnilnd_nak_rdma(conn, rxmsg->gnm_type,
                                        -ENOENT,
                                        rxmsg->gnm_u.get.gngm_cookie,
                                        ni->ni_nid);
                }
                kgnilnd_consume_rx(rx);
                RETURN(0);
        }
        RETURN(0);
}

/* needs write_lock on kgn_peer_conn_lock held */
int
kgnilnd_check_conn_timeouts_locked(kgn_conn_t *conn)
{
        unsigned long      timeout, keepalive;
        unsigned long      now = jiffies;
        unsigned long      newest_last_rx;
        kgn_tx_t          *tx;

        /* given that we found this conn hanging off a peer, it better damned
         * well be connected */
        LASSERTF(conn->gnc_state == GNILND_CONN_ESTABLISHED,
                 "conn 0x%p->%s with bad state%s\n", conn,
                 conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
                               : "<?>",
                 kgnilnd_conn_state2str(conn));

        CDEBUG(D_NET, "checking conn %p->%s timeout %d keepalive %d "
                      "rx_diff %lu tx_diff %lu\n",
                conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
                conn->gnc_timeout, GNILND_TO2KA(conn->gnc_timeout),
                cfs_duration_sec(now - conn->gnc_last_rx_cq),
                cfs_duration_sec(now - conn->gnc_last_tx));

        timeout = cfs_time_seconds(conn->gnc_timeout);
        keepalive = cfs_time_seconds(GNILND_TO2KA(conn->gnc_timeout));

        /* just in case our lack of RX msg processing is gumming up the works - give the
         * remove an extra chance */

        newest_last_rx = GNILND_LASTRX(conn);

        if (time_after_eq(now, newest_last_rx + timeout)) {
                uint32_t level = D_CONSOLE|D_NETERROR;

                if (conn->gnc_peer->gnp_down == GNILND_RCA_NODE_DOWN) {
                        level = D_NET;
                }
                        GNIDBG_CONN(level, conn,
                        "No gnilnd traffic received from %s for %lu "
                        "seconds, terminating connection. Is node down? ",
                        libcfs_nid2str(conn->gnc_peer->gnp_nid),
                        cfs_duration_sec(now - newest_last_rx));
                return -ETIMEDOUT;
        }

        /* we don't timeout on last_tx stalls - we are going to trust the
         * underlying network to let us know when sends are failing.
         * At worst, the peer will timeout our RX stamp and drop the connection
         * at that point. We'll then see his CLOSE or at worst his RX
         * stamp stop and drop the connection on our end */

        if (time_after_eq(now, conn->gnc_last_tx + keepalive)) {
                CDEBUG(D_NET, "sending NOOP -> %s (%p idle %lu(%lu)) "
                       "last %lu/%lu/%lu %lus/%lus/%lus\n",
                       libcfs_nid2str(conn->gnc_peer->gnp_nid), conn,
                       cfs_duration_sec(jiffies - conn->gnc_last_tx),
                       keepalive,
                       conn->gnc_last_noop_want, conn->gnc_last_noop_sent,
                       conn->gnc_last_noop_cq,
                       cfs_duration_sec(jiffies - conn->gnc_last_noop_want),
                       cfs_duration_sec(jiffies - conn->gnc_last_noop_sent),
                       cfs_duration_sec(jiffies - conn->gnc_last_noop_cq));
                set_mb(conn->gnc_last_noop_want, jiffies);
                atomic_inc(&conn->gnc_reaper_noop);
                if (CFS_FAIL_CHECK(CFS_FAIL_GNI_NOOP_SEND))
                        return 0;

                tx = kgnilnd_new_tx_msg(GNILND_MSG_NOOP, conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
                if (tx == NULL)
                        return 0;
                kgnilnd_queue_tx(conn, tx);
        }

        return 0;
}

/* needs write_lock on kgn_peer_conn_lock held */
void
kgnilnd_check_peer_timeouts_locked(kgn_peer_t *peer, struct list_head *todie,
                                    struct list_head *souls)
{
        unsigned long           timeout;
        kgn_conn_t             *conn, *connN = NULL;
        kgn_tx_t               *tx, *txN;
        int                     rc = 0;
        int                     count = 0;
        int                     reconnect;
        int                     to_reconn;
        short                   releaseconn = 0;
        unsigned long           first_rx = 0;
        int                     purgatory_conn_cnt = 0;

        CDEBUG(D_NET, "checking peer 0x%p->%s for timeouts; interval %lus\n",
                peer, libcfs_nid2str(peer->gnp_nid),
                peer->gnp_reconnect_interval);

        timeout = cfs_time_seconds(MAX(*kgnilnd_tunables.kgn_timeout,
                                       GNILND_MIN_TIMEOUT));

        conn = kgnilnd_find_conn_locked(peer);
        if (conn) {
                /* if there is a valid conn, check the queues for timeouts */
                rc = kgnilnd_check_conn_timeouts_locked(conn);
                if (rc) {
                        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_RX_CLOSE_CLOSING)) {
                                /* simulate a RX CLOSE after the timeout but before
                                 * the scheduler thread gets it */
                                conn->gnc_close_recvd = GNILND_CLOSE_INJECT1;
                                conn->gnc_peer_error = -ETIMEDOUT;
                        }
                        /* Once we mark closed, any of the scheduler threads could
                         * get it and move through before we hit the fail loc code */
                        kgnilnd_close_conn_locked(conn, rc);
                } else {
                        /* first_rx is used to decide when to release a conn from purgatory.
                         */
                        first_rx = conn->gnc_first_rx;
                }
        }

        /* now regardless of starting new conn, find tx on peer queue that
         * are old and smell bad - do this first so we don't trigger
         * reconnect on empty queue if we timeout all */
        list_for_each_entry_safe(tx, txN, &peer->gnp_tx_queue, tx_list) {
                if (time_after_eq(jiffies, tx->tx_qtime + timeout)) {
                        if (count == 0) {
                                LCONSOLE_INFO("could not send to %s due to connection"
                                       " setup failure after %lu seconds\n",
                                       libcfs_nid2str(peer->gnp_nid),
                                       cfs_duration_sec(jiffies - tx->tx_qtime));
                        }
                        kgnilnd_tx_del_state_locked(tx, peer, NULL,
                                                   GNILND_TX_ALLOCD);
                        list_add_tail(&tx->tx_list, todie);
                        count++;
                }
        }

        if (count || peer->gnp_connecting == GNILND_PEER_KILL) {
                CDEBUG(D_NET, "canceling %d tx for peer 0x%p->%s\n",
                        count, peer, libcfs_nid2str(peer->gnp_nid));
                /* if we nuked all the TX, stop peer connection attempt (if there is one..) */
                if (list_empty(&peer->gnp_tx_queue) ||
                        peer->gnp_connecting == GNILND_PEER_KILL) {
                        /* we pass down todie to use a common function - but we know there are
                         * no TX to add */
                        kgnilnd_cancel_peer_connect_locked(peer, todie);
                }
        }

        /* Don't reconnect if we are still trying to clear out old conns.
         * This prevents us sending traffic on the new mbox before ensuring we are done
         * with the old one */
        reconnect = (peer->gnp_down == GNILND_RCA_NODE_UP) &&
                    (atomic_read(&peer->gnp_dirty_eps) == 0);

        /* fast reconnect after a timeout */
        to_reconn = !conn &&
                    (peer->gnp_last_errno == -ETIMEDOUT) &&
                    *kgnilnd_tunables.kgn_fast_reconn;

        /* if we are not connected and there are tx on the gnp_tx_queue waiting
         * to be sent, we'll check the reconnect interval and fire up a new
         * connection request */

        if (reconnect &&
            (peer->gnp_connecting == GNILND_PEER_IDLE) &&
            (time_after_eq(jiffies, peer->gnp_reconnect_time)) &&
            (!list_empty(&peer->gnp_tx_queue) || to_reconn)) {

                CDEBUG(D_NET, "starting connect to %s\n",
                        libcfs_nid2str(peer->gnp_nid));
                LASSERTF(peer->gnp_connecting == GNILND_PEER_IDLE, "Peer was idle and we"
                        "have a write_lock, state issue %d\n", peer->gnp_connecting);

                peer->gnp_connecting = GNILND_PEER_CONNECT;
                kgnilnd_peer_addref(peer); /* extra ref for connd */

                spin_lock(&peer->gnp_net->gnn_dev->gnd_connd_lock);
                list_add_tail(&peer->gnp_connd_list,
                              &peer->gnp_net->gnn_dev->gnd_connd_peers);
                spin_unlock(&peer->gnp_net->gnn_dev->gnd_connd_lock);

                kgnilnd_schedule_dgram(peer->gnp_net->gnn_dev);
        }

        /* fail_loc to allow us to delay release of purgatory */
        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_PURG_REL_DELAY))
                return;

        /* This check allows us to verify that the new conn is actually being used. This allows us to
         * pull the old conns out of purgatory if they have actually seen traffic.
         * We only release a conn from purgatory during stack reset, admin command, or when a peer reconnects
         */
        if (first_rx &&
                time_after(jiffies, first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))) {
                CDEBUG(D_INFO, "We can release peer %s conn's from purgatory %lu\n",
                        libcfs_nid2str(peer->gnp_nid), first_rx + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout));
                releaseconn = 1;
        }

        list_for_each_entry_safe (conn, connN, &peer->gnp_conns, gnc_list) {
        /* check for purgatory timeouts */
                if (conn->gnc_in_purgatory) {
                        /* We cannot detach this conn from purgatory if it has not been closed so we reschedule it
                         * that way the next time we check it we can detach it from purgatory
                         */

                        if (conn->gnc_state != GNILND_CONN_DONE) {
                                /* Skip over conns that are currently not DONE. If they arent already scheduled
                                 * for completion something in the state machine is broken.
                                 */
                                continue;
                        }

                        /* We only detach a conn that is in purgatory if we have received a close message,
                         * we have a new valid connection that has successfully received data, or an admin
                         * command tells us we need to detach.
                         */

                        if (conn->gnc_close_recvd || releaseconn || conn->gnc_needs_detach) {
                                unsigned long   waiting;

                                waiting = (long) jiffies - conn->gnc_last_rx_cq;

                                /* C.E: The remote peer is expected to close the
                                 * connection (see kgnilnd_check_conn_timeouts)
                                 * via the reaper thread and nuke out the MDD and
                                 * FMA resources after conn->gnc_timeout has expired
                                 * without an FMA RX */
                                CDEBUG(D_NET, "Reconnected to %s in %lds or admin forced detach, dropping "
                                        " held resources\n",
                                        libcfs_nid2str(conn->gnc_peer->gnp_nid),
                                        cfs_duration_sec(waiting));

                                kgnilnd_detach_purgatory_locked(conn, souls);
                        } else {
                                purgatory_conn_cnt++;
                        }
                }
        }

        /* If we have too many connections in purgatory we could run out of
         * resources. Limit the number of connections to a tunable number,
         * clean up to the minimum all in one fell swoop... there are
         * situations where dvs will retry tx's and we can eat up several
         * hundread connection requests at once.
         */
        if (purgatory_conn_cnt > *kgnilnd_tunables.kgn_max_purgatory) {
                list_for_each_entry_safe(conn, connN, &peer->gnp_conns,
                                         gnc_list) {
                        if (conn->gnc_in_purgatory &&
                            conn->gnc_state == GNILND_CONN_DONE) {
                                CDEBUG(D_NET, "Dropping Held resource due to"
                                              " resource limits being hit\n");
                                kgnilnd_detach_purgatory_locked(conn, souls);

                                if (purgatory_conn_cnt-- <
                                    *kgnilnd_tunables.kgn_max_purgatory)
                                        break;
                        }
                }
        }

        return;
}

void
kgnilnd_reaper_check(int idx)
{
        struct list_head  *peers = &kgnilnd_data.kgn_peers[idx];
        struct list_head  *ctmp, *ctmpN;
        struct list_head   geriatrics;
        struct list_head   souls;

        INIT_LIST_HEAD(&geriatrics);
        INIT_LIST_HEAD(&souls);

        write_lock(&kgnilnd_data.kgn_peer_conn_lock);

        list_for_each_safe(ctmp, ctmpN, peers) {
                kgn_peer_t        *peer = NULL;

                /* don't timeout stuff if the network is mucked or shutting down */
                if (kgnilnd_check_hw_quiesce()) {
                        break;
                }
                peer = list_entry(ctmp, kgn_peer_t, gnp_list);

                kgnilnd_check_peer_timeouts_locked(peer, &geriatrics, &souls);
        }

        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);

        kgnilnd_txlist_done(&geriatrics, -EHOSTUNREACH);
        kgnilnd_release_purgatory_list(&souls);
}

void
kgnilnd_update_reaper_timeout(long timeout)
{
        LASSERT(timeout > 0);

        spin_lock(&kgnilnd_data.kgn_reaper_lock);

        if (timeout < kgnilnd_data.kgn_new_min_timeout)
                kgnilnd_data.kgn_new_min_timeout = timeout;

        spin_unlock(&kgnilnd_data.kgn_reaper_lock);
}

static void
kgnilnd_reaper_poke_with_stick(unsigned long arg)
{
        wake_up(&kgnilnd_data.kgn_reaper_waitq);
}

int
kgnilnd_reaper(void *arg)
{
        long               timeout;
        int                i;
        int                hash_index = 0;
        unsigned long      next_check_time = jiffies;
        long               current_min_timeout = MAX_SCHEDULE_TIMEOUT;
        struct timer_list  timer;
        DEFINE_WAIT(wait);

        cfs_block_allsigs();

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

        while (!kgnilnd_data.kgn_shutdown) {
                /* I wake up every 'p' seconds to check for timeouts on some
                 * more peers.  I try to check every connection 'n' times
                 * within the global minimum of all keepalive and timeout
                 * intervals, to ensure I attend to every connection within
                 * (n+1)/n times its timeout intervals. */
                const int     p = GNILND_REAPER_THREAD_WAKE;
                const int     n = GNILND_REAPER_NCHECKS;
                int           chunk;
                /* to quiesce or to not quiesce, that is the question */
                if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
                        spin_unlock(&kgnilnd_data.kgn_reaper_lock);
                        KGNILND_SPIN_QUIESCE;
                        spin_lock(&kgnilnd_data.kgn_reaper_lock);
                }

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

                if (timeout > 0) {
                        prepare_to_wait(&kgnilnd_data.kgn_reaper_waitq, &wait,
                                        TASK_INTERRUPTIBLE);
                        spin_unlock(&kgnilnd_data.kgn_reaper_lock);
                        setup_timer(&timer, kgnilnd_reaper_poke_with_stick,
                                    next_check_time);
                        mod_timer(&timer, (long) jiffies + timeout);

                        /* check flag variables before comitting */
                        if (!kgnilnd_data.kgn_shutdown &&
                            !kgnilnd_data.kgn_quiesce_trigger) {
                                CDEBUG(D_INFO, "schedule timeout %ld (%lu sec)\n",
                                       timeout, cfs_duration_sec(timeout));
                                schedule();
                                CDEBUG(D_INFO, "awake after schedule\n");
                        }

                        del_singleshot_timer_sync(&timer);
                        spin_lock(&kgnilnd_data.kgn_reaper_lock);
                        finish_wait(&kgnilnd_data.kgn_reaper_waitq, &wait);
                        continue;
                }

                /* new_min_timeout is set from the conn timeouts and keepalive
                 * this should end up with a min timeout of
                 * GNILND_TIMEOUT2KEEPALIVE(t) or roughly LND_TIMEOUT/2 */
                if (kgnilnd_data.kgn_new_min_timeout < current_min_timeout) {
                        current_min_timeout = kgnilnd_data.kgn_new_min_timeout;
                        CDEBUG(D_NET, "Set new min timeout %ld\n",
                               current_min_timeout);
                }

                spin_unlock(&kgnilnd_data.kgn_reaper_lock);

                /* Compute how many table entries to check now so I get round
                 * the whole table fast enough given that I do this at fixed
                 * intervals of 'p' seconds) */
                chunk = *kgnilnd_tunables.kgn_peer_hash_size;
                if (kgnilnd_data.kgn_new_min_timeout > n * p)
                        chunk = (chunk * n * p) /
                                kgnilnd_data.kgn_new_min_timeout;
                if (chunk == 0)
                        chunk = 1;
                for (i = 0; i < chunk; i++) {
                        kgnilnd_reaper_check(hash_index);
                        hash_index = (hash_index + 1) %
                                *kgnilnd_tunables.kgn_peer_hash_size;
                }
                next_check_time = (long) jiffies + cfs_time_seconds(p);
                CDEBUG(D_INFO, "next check at %lu or in %d sec\n", next_check_time, p);

                spin_lock(&kgnilnd_data.kgn_reaper_lock);
        }

        spin_unlock(&kgnilnd_data.kgn_reaper_lock);

        kgnilnd_thread_fini();
        return 0;
}

int
kgnilnd_recv_bte_get(kgn_tx_t *tx) {
        unsigned niov, offset, nob;
        lnet_kiov_t     *kiov;
        lnet_msg_t *lntmsg = tx->tx_lntmsg[0];
        kgnilnd_parse_lnet_rdma(lntmsg, &niov, &offset, &nob, &kiov, tx->tx_nob_rdma);

        if (kiov != NULL) {
                lnet_copy_flat2kiov(
                        niov, kiov, offset,
                        nob,
                        tx->tx_buffer_copy + tx->tx_offset, 0, nob);
        } else {
                memcpy(tx->tx_buffer, tx->tx_buffer_copy + tx->tx_offset, nob);
        }
        return 0;
}


int
kgnilnd_check_rdma_cq(kgn_device_t *dev)
{
        gni_return_t           rrc;
        gni_post_descriptor_t *desc;
        __u64                  event_data;
        kgn_tx_ev_id_t         ev_id;
        char                   err_str[256];
        int                    should_retry, rc;
        long                   num_processed = 0;
        kgn_conn_t            *conn = NULL;
        kgn_tx_t              *tx = NULL;
        kgn_rdma_desc_t       *rdesc;
        unsigned int           rnob;
        __u64                  rcookie;

        for (;;) {
                /* make sure we don't keep looping if we need to reset */
                if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
                        return num_processed;
                }
                rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
                if (!rc) {
                        /* we didn't get the mutex, so return that there is still work
                         * to be done */
                        return 1;
                }
                if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DELAY_RDMA)) {
                        /* a bit gross - but we need a good way to test for
                         * delayed RDMA completions and the easiest way to do
                         * that is to delay the RDMA CQ events */
                        rrc = GNI_RC_NOT_DONE;
                } else {
                        rrc = kgnilnd_cq_get_event(dev->gnd_snd_rdma_cqh, &event_data);
                }

                if (rrc == GNI_RC_NOT_DONE) {
                        kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
                        CDEBUG(D_INFO, "SEND RDMA CQ %d empty processed %ld\n",
                               dev->gnd_id, num_processed);
                        return num_processed;
                }
                dev->gnd_sched_alive = jiffies;
                num_processed++;

                LASSERTF(!GNI_CQ_OVERRUN(event_data),
                        "this is bad, somehow our credits didn't protect us"
                        " from CQ overrun\n");
                LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_POST,
                        "rrc %d, GNI_CQ_GET_TYPE("LPX64") = "LPX64"\n", rrc,
                        event_data, GNI_CQ_GET_TYPE(event_data));

                rrc = kgnilnd_get_completed(dev->gnd_snd_rdma_cqh, event_data,
                                            &desc);
                kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);

                /* XXX Nic: Need better error handling here... */
                LASSERTF((rrc == GNI_RC_SUCCESS) ||
                          (rrc == GNI_RC_TRANSACTION_ERROR),
                         "rrc %d\n", rrc);

                ev_id.txe_cookie = desc->post_id;

                kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);

                if (conn == NULL || tx == NULL) {
                        /* either conn or tx was already nuked and this is a "late"
                         * completion, so drop it */
                        continue;
                }

                GNITX_ASSERTF(tx, tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
                        tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE ||
                        tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV ||
                        tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV,
                        "tx %p with type %d\n", tx, tx->tx_msg.gnm_type);

                GNIDBG_TX(D_NET, tx, "RDMA completion for %d bytes", tx->tx_nob);

                if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
                        lnet_set_reply_msg_len(NULL, tx->tx_lntmsg[1],
                                               tx->tx_msg.gnm_u.completion.gncm_retval);
                }

                rc = 0;
                if (tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV && desc->status == GNI_RC_SUCCESS) {
                        if (tx->tx_buffer_copy != NULL)
                                kgnilnd_recv_bte_get(tx);
                        rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_putinfo.gnpam_payload_cksum, tx->tx_nob_rdma);
                }

                if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE_REV && desc->status == GNI_RC_SUCCESS) {
                        if (tx->tx_buffer_copy != NULL)
                                kgnilnd_recv_bte_get(tx);
                        rc = kgnilnd_verify_rdma_cksum(tx, tx->tx_getinfo.gngm_payload_cksum, tx->tx_nob_rdma);
                }

                /* remove from rdmaq */
                kgnilnd_conn_mutex_lock(&conn->gnc_rdma_mutex);
                spin_lock(&conn->gnc_list_lock);
                kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
                spin_unlock(&conn->gnc_list_lock);
                kgnilnd_conn_mutex_unlock(&conn->gnc_rdma_mutex);

                if (likely(desc->status == GNI_RC_SUCCESS) && rc == 0) {
                        atomic_inc(&dev->gnd_rdma_ntx);
                        atomic64_add(tx->tx_nob, &dev->gnd_rdma_txbytes);
                        /* transaction succeeded, add into fmaq */
                        kgnilnd_queue_tx(conn, tx);
                        kgnilnd_peer_alive(conn->gnc_peer);

                        /* drop ref from kgnilnd_validate_tx_ev_id */
                        kgnilnd_admin_decref(conn->gnc_tx_in_use);
                        kgnilnd_conn_decref(conn);

                        continue;
                }

                /* fall through to the TRANSACTION_ERROR case */
                tx->tx_retrans++;

                /* get stringified version for log messages */
                kgnilnd_cq_error_str(event_data, &err_str, 256);
                kgnilnd_cq_error_recoverable(event_data, &should_retry);

                /* make sure we are not off in the weeds with this tx */
                if (tx->tx_retrans >
                        *kgnilnd_tunables.kgn_max_retransmits) {
                        GNIDBG_TX(D_NETERROR, tx,
                               "giving up on TX, too many retries", NULL);
                        should_retry = 0;
                }

                GNIDBG_TX(D_NETERROR, tx, "RDMA %s error (%s)",
                        should_retry ? "transient" : "unrecoverable", err_str);

                if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
                    tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
                        rdesc    = &tx->tx_putinfo.gnpam_desc;
                        rnob     = tx->tx_putinfo.gnpam_desc.gnrd_nob;
                        rcookie  = tx->tx_putinfo.gnpam_dst_cookie;
                } else {
                        rdesc    = &tx->tx_getinfo.gngm_desc;
                        rnob     = tx->tx_lntmsg[0]->msg_len;
                        rcookie  = tx->tx_getinfo.gngm_cookie;
                }

                if (should_retry) {
                        kgnilnd_rdma(tx,
                                     tx->tx_msg.gnm_type,
                                     rdesc,
                                     rnob, rcookie);
                } else {
                        kgnilnd_nak_rdma(conn,
                                         tx->tx_msg.gnm_type,
                                         -EFAULT,
                                         rcookie,
                                         tx->tx_msg.gnm_srcnid);
                        kgnilnd_tx_done(tx, -EFAULT);
                        kgnilnd_close_conn(conn, -ECOMM);
                }

                /* drop ref from kgnilnd_validate_tx_ev_id */
                kgnilnd_admin_decref(conn->gnc_tx_in_use);
                kgnilnd_conn_decref(conn);
        }
}

int
kgnilnd_check_fma_send_cq(kgn_device_t *dev)
{
        gni_return_t           rrc;
        __u64                  event_data;
        kgn_tx_ev_id_t         ev_id;
        kgn_tx_t              *tx = NULL;
        kgn_conn_t            *conn = NULL;
        int                    queued_fma, saw_reply, rc;
        long                   num_processed = 0;

        for (;;) {
                /* make sure we don't keep looping if we need to reset */
                if (unlikely(kgnilnd_data.kgn_quiesce_trigger)) {
                        return num_processed;
                }

                rc = kgnilnd_mutex_trylock(&dev->gnd_cq_mutex);
                if (!rc) {
                        /* we didn't get the mutex, so return that there is still work
                         * to be done */
                        return 1;
                }

                rrc = kgnilnd_cq_get_event(dev->gnd_snd_fma_cqh, &event_data);
                kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);

                if (rrc == GNI_RC_NOT_DONE) {
                        CDEBUG(D_INFO,
                               "SMSG send CQ %d not ready (data "LPX64") "
                               "processed %ld\n", dev->gnd_id, event_data,
                               num_processed);
                        return num_processed;
                }

                dev->gnd_sched_alive = jiffies;
                num_processed++;

                LASSERTF(!GNI_CQ_OVERRUN(event_data),
                        "this is bad, somehow our credits didn't "
                        "protect us from CQ overrun\n");
                LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_SMSG,
                        "rrc %d, GNI_CQ_GET_TYPE("LPX64") = "LPX64"\n", rrc,
                        event_data, GNI_CQ_GET_TYPE(event_data));

                /* if SMSG couldn't handle an error, time for conn to die */
                if (unlikely(rrc == GNI_RC_TRANSACTION_ERROR)) {
                        char            err_str[256];

                        /* need to take the write_lock to ensure atomicity
                         * on the conn state if we need to close it */
                        write_lock(&kgnilnd_data.kgn_peer_conn_lock);
                        conn = kgnilnd_cqid2conn_locked(GNI_CQ_GET_INST_ID(event_data));
                        if (conn == NULL) {
                                /* Conn was destroyed? */
                                CDEBUG(D_NET,
                                        "SMSG CQID lookup "LPX64" failed\n",
                                        GNI_CQ_GET_INST_ID(event_data));
                                write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
                                continue;
                        }

                        kgnilnd_cq_error_str(event_data, &err_str, 256);
                        CNETERR("SMSG send error to %s: rc %d (%s)\n",
                               libcfs_nid2str(conn->gnc_peer->gnp_nid),
                               rrc, err_str);
                        kgnilnd_close_conn_locked(conn, -ECOMM);

                        write_unlock(&kgnilnd_data.kgn_peer_conn_lock);

                        /* no need to process rest of this tx -
                         * it is getting canceled */
                        continue;
                }

                /* fall through to GNI_RC_SUCCESS case */
                ev_id.txe_smsg_id = GNI_CQ_GET_MSG_ID(event_data);

                kgnilnd_validate_tx_ev_id(&ev_id, &tx, &conn);
                if (conn == NULL || tx == NULL) {
                        /* either conn or tx was already nuked and this is a "late"
                         * completion, so drop it */
                        continue;
                }

                tx->tx_conn->gnc_last_tx_cq = jiffies;
                if (tx->tx_msg.gnm_type == GNILND_MSG_NOOP) {
                        set_mb(conn->gnc_last_noop_cq, jiffies);
                }

                /* lock tx_list_state and tx_state */
                kgnilnd_conn_mutex_lock(&conn->gnc_smsg_mutex);
                spin_lock(&tx->tx_conn->gnc_list_lock);

                GNITX_ASSERTF(tx, tx->tx_list_state == GNILND_TX_LIVE_FMAQ,
                                "state not GNILND_TX_LIVE_FMAQ", NULL);
                GNITX_ASSERTF(tx, tx->tx_state & GNILND_TX_WAITING_COMPLETION,
                        "not waiting for completion", NULL);

                GNIDBG_TX(D_NET, tx, "SMSG complete tx_state %x rc %d",
                        tx->tx_state, rrc);

                tx->tx_state &= ~GNILND_TX_WAITING_COMPLETION;

                /* This will trigger other FMA sends that were
                 * pending this completion */
                queued_fma = !list_empty(&tx->tx_conn->gnc_fmaq);

                /* we either did not expect reply or we already got it */
                saw_reply = !(tx->tx_state & GNILND_TX_WAITING_REPLY);

                spin_unlock(&tx->tx_conn->gnc_list_lock);
                kgnilnd_conn_mutex_unlock(&conn->gnc_smsg_mutex);

                if (queued_fma) {
                        CDEBUG(D_NET, "scheduling conn 0x%p->%s for fmaq\n",
                               conn,
                               libcfs_nid2str(conn->gnc_peer->gnp_nid));
                        kgnilnd_schedule_conn(conn);
                }

                /* If saw_reply is false as soon as gnc_list_lock is dropped the tx could be nuked
                 * If saw_reply is true we know that the tx is safe to use as the other thread
                 * is already finished with it.
                 */

                if (saw_reply) {
                        /* no longer need to track on the live_fmaq */
                        kgnilnd_tx_del_state_locked(tx, NULL, tx->tx_conn, GNILND_TX_ALLOCD);

                        if (tx->tx_state & GNILND_TX_PENDING_RDMA) {
                                /* we already got reply &